paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1911.03075 | 1 | 1911 | 2019-11-08T06:30:54 | Strongly irreducible factorization of quaternionic operators and Riesz decomposition theorem | [
"math.FA"
] | Let $T$ be a bounded quaternionic normal operator on a right quaternionic Hilbert space $\mathcal{H}$. We show that $T$ can be factorized in a strongly irreducible sense, that is, for any $\delta >0$ there exist a compact operator $K$ with $\|K\|< \delta$, a partial isometry $W$ and a strongly irreducible operator $S$ on $\mathcal{H}$ such that \begin{equation*} T = (W+K) S. \end{equation*} We illustrate our result with an example. We also prove a quaternionic version of the Riesz decomposition theorem and as a consequence, show that if the spherical spectrum of a bounded quaternionic operator (need not be normal) is disconnected by a pair of disjoint axially symmetric closed subsets, then it is strongly reducible. | math.FA | math | STRONGLY IRREDUCIBLE FACTORIZATION OF QUATERNIONIC
OPERATORS AND RIESZ DECOMPOSITION THEOREM
9
1
0
2
v
o
N
8
]
P. SANTHOSH KUMAR
Abstract. Let T be a bounded quaternionic normal operator on a right quaternionic
Hilbert space H. We show that T can be factorized in a strongly irreducible sense, that
is, for any δ > 0 there exist a compact operator K with kKk < δ, a partial isometry W and
a strongly irreducible operator S on H such that
T = (W + K)S.
We illustrate our result with an example. We also prove a quaternionic version of the Riesz
decomposition theorem and as a consequence, show that if the spherical spectrum of a
bounded quaternionic operator (need not be normal) is disconnected by a pair of disjoint
axially symmetric closed subsets, then it is strongly reducible.
.
A
F
h
t
a
m
[
1
v
5
7
0
3
0
.
1
1
9
1
:
v
i
X
r
a
1. Introduction
According to the Frobenius theorem for real division algebras, the algebra of Hamilton
quaternions [5] is the only finite dimensional associative division algebra that contains R and
C as proper real subalgebras. Similar to the case of real and complex matrices, the theory of
matrices over the real algebra of quaternions has been well developed (see [1, 6, 16, 17] and
references therein) in the literature. For example, some of the fundamental results like Schur's
canonical form and Jordan canonical form are extended to matrices with quaternion entries
[17]. In particular, the Jordan canonical form shows that every square matrix over quaternions
can be reduced under similarity to a direct sum of Jordan blocks. The Jordan canonical form
determines its complete similarity invariants and establishes the structure of quaternionic
operator on finite dimensional right quaternionic Hilbert spaces. Another significant aspect
in this direction is the diagonalization, through which the matrix is factorized interms of its
restriction to eigenspaces. In particular, as in the case of complex matrices, every quaternionic
normal matrix is diagonalizable. It is worth mentioning the result due to Weigmann [16] that
a quaternion matrix is normal if and only if its adjoint is a polynomial of given matrix with
real coefficients.
In this article, we concentrate on the class of normal operators on right
quaternionic Hilbert spaces and their factorization in a strongly irreducible sense. For this,
we adopt the notion of strong irreducibility proposed by F. Gilfeather [8] and Z. J. Jiang
[11] to the class of quaternionic operators in order to replace the notion of Jordan block for
infinite dimensional right quaternionic Hilbert spaces.
On the other hand, the study of quaternionic normal operators gained much attention in the
form of spectral theory for quantum theories and various versions of quaternionic functional
calculus (see [2, 3, 7] for details). As far as factorization of quaternionic operators concern,
the well known polar decomposition theorem shows that every bounded or densely defined
closed quaternionic operator can be decomposed as a product of a partial isometry and a
positive operator [7, 14]. Furthermore, the authors of [14] provided a necessary and sufficient
2010 Mathematics Subject Classification. Primary 47S10, 47A68; Secondary 47A15, 47B99.
Key words and phrases. Axially symmetric set, Quaternionic Hilbert space, Quaternionic normal operator,
Spherical specturm, Strongly irreducible operator, Riesz decomposition theorem.
1
2
P. SANTHOSH KUMAR
condition for any arbitrary factorization of densely defined closed quaternionic operator to be
a polar decomposition. Note that the positive operator that appear in the polar decomposition
is the modulus of the given quaternionic operator which has several reducing subsapces as
in the complex case [4]. In summary, the polar decomposition establishes the factorization
of quaternionic operator as a product of a partial isometry and an operator having several
reducing subspaces. A natural question that arises from previous observation is the following:
"whether a given quaternionic operator be decomposed as a product of a partial isometry and
an operator with no reducing subspace?" We answer this question for quaternionic normal
operators by proving a factorization in a strongly irreducible sense, by means of decomposing
the given operator as a product of a sufficiently small compact perturbation of a partial
isometry and a strongly irreducible quaternionic operator. Our result is the quaternionic
version (for normal operators) of the result proved recently in [15] by G. Tian et al.
In
which, the authors employed the properties of Cowen-Douglas operators related to complex
geometry on complex Hilbert spaces. Also see [12] for such factorization in finite dimensional
Hilbert spaces.
We organize this article in four sections. In the second section, we give some basic defi-
nitions and results that are useful for proving our assertions. In the third section, we prove
factorization of quaternionic normal operator in a strongly irreducible sense. The final section
is dedicated to Riesz decomposition theorem for quternionic operators and provide a suffient
condition for strong irreducibility.
2. Preliminaries
An Irish mathematician Sir William Rowan Hamilton [5] described this number system
known as "quaternions" as an extension of complex numbers. A quaternion is of the form:
q = q0 + q1i + q2j + q3k,
where qℓ ∈ R for ℓ = 0, 1, 2, 3 and i, j, k are the fundemental quaternion units satisfying,
(2.1)
i2 = j2 = k2 = −1 = i · j · k.
Note that the collection of all quaternions is denoted by H and it is a non-commutative
real division algebra equipped with the usual vector space operations addition and scalar
multiplication same as in the complex field C, and with the ring multiplication given by
Equation (2.1).
For every q ∈ H, the real and the imaginary part of 'q' is defined as re(q) = q0 and
im(q) = q1i + q2j + q3k respectively. Then the conjugate and the modulus of q ∈ H is given
respectively by
The set of all imaginary unit quaternions in H denoted by S and it is defined as
q = q0 − (q1i + q2j + q3k) and q =qq2
S :=(cid:8)q ∈ H : q = −q & q = 1(cid:9) =(cid:8)q ∈ H : q2 = −1(cid:9).
0 + q2
1 + q2
2 + q2
3.
However S is a 2- dimensional sphere in R4. The real subalgebra of H generated by {1, m},
where m ∈ S, is denoted by Cm := {α + mβ : α, β ∈ R} called a slice of H. It is immediate
to see that, for every q ∈ H, there is a unique mq := im(q)
im(q) ∈ S such that
q = re(q) + mqim(q) ∈ Cmq .
Consequently, H = Sm∈S
relation on H given by
STRONGLY IRREDUCIBLE FACTORIZATION
3
Cm and Cm ∩ Cn = R, for all m 6= ±n ∈ S. There is an equivalence
p ∼ q ⇔ p = s−1qs, for some s ∈ H \ {0}.
For every q ∈ H, the equivalence class of q is expressed as,
(2.2)
[q] =(cid:8)p ∈ H : re(p) = re(q) and im(p) = im(q)(cid:9).
Definition 2.1. [7] Let S be a non-empty subset of C.
(1) If S is invariant under complex conjugation, then the circularization ΩS of S is defined
by
ΩS =(cid:8)α + mβ : α, β ∈ R, α + iβ ∈ S, m ∈ S(cid:9).
(2) a subset K of H is called circular or axially symmetric, if K = ΩS, for some S ⊆ C
which is invariant under complex conjugation.
Note that by Definition 2.1 and Equation (2.2), we can express ΩS as the union of the
equivalence class of its elements,
ΩS = [z∈S
[z].
It follows that the closure of ΩS is the circularization of the closure of S. That is,
(2.3)
ΩS = ΩS.
The properties described above are useful for later sections. A more detailed discussion
about quaternions can be found in [1, 3, 7, 17]. Now we discuss some basic definitions and
existing results related to quaternionic Hilbert spaces.
Definition 2.2. [7, section 2.2] An inner product on a right H-module H is a map
h·, ·i : H × H → H
satisfy the following properties,
(1) Positivity: hx, xi ≥ 0 for all x ∈ H. In particular,
hx, xi = 0 if and only if x = 0.
(2) Right linearity: hx, yq + zi = hx, yi q + hx, zi, for all x, y, z ∈ H, q ∈ H.
(3) Quaternionic Hermiticity: hx, yi = hy, xi, for all x, y ∈ H.
The pair (H, h·, ·i) is called quaternionic pre-Hilbert space. Moreover, H is said to be
(i) quaternionic Hilbet space, if H is complete with respect to the norm induced from the
inner product h·, ·i, which is defined by
(ii) separable, if H has a countable dense subset.
Furthermore, for any subset N of H, the span of N is defined as
kxk =phx, xi, for all x ∈ H.
The orthogonal complement of a subspace M of H is,
span N :=n nXℓ=1
xℓqℓ; xℓ ∈ N , n ∈ No.
M⊥ =(cid:8)x ∈ H; hx, yi = 0, for every y ∈ M(cid:9).
hx, yi2 ≤ hx, xihy, yi, for all x, y ∈ H.
Note 2.3. The inner product h·, ·i defined on H satisfies Cauchy-Schwarz inequality:
4
P. SANTHOSH KUMAR
Definition 2.4. [7] Let (cid:0)H, h·, ·i(cid:1) be a quaternionic Hilbert space. A subset N of H with
the property that
is said to be Hilbert basis if for every x, y ∈ H, the series Pz∈N
and it holds:
hx, zihz, yi converges absolutely
hz, z′i =(cid:26) 1,
0,
if z = z′
if z 6= z′
hx, yi = Xz∈N
hx, zihz, yi.
Equivalentely, span N = H.
Remark 2.5. By [7, Proposition 2.6], every quaternionic Hilbert space H has a Hilbert basis
N . For every x ∈ H, it is uniquely decomposed as,
x = Xz∈N
zhz, xi.
Example 2.6. [13] Let µ be a lebesgue measure on [0, 1] and the set of all H-valued square
integrable µ-measurable functions on [0, 1] is defined as,
L2(cid:0)[0, 1]; H; µ(cid:1) =nf : [0, 1] → H ;
1Z
0
f (x)2 dx < ∞o.
It is a right quaternionic Hilbert space with respect to the inner product given by
hf, gi =
1Z
0
f (x)g(x) dx, for all f, g ∈ L2(cid:0)[0, 1]; H; µ(cid:1).
r
r
(x) = exp{2πmrx}, for all x ∈ [0, 1] and r ∈ Z. Then
Let us fix m ∈ S. If we define e(m)
the set N = {e(m)
: r ∈ Z} is an orthornormal system in L2(cid:0)[0, 1]; H; µ(cid:1). Further, by the
Stone-Weierstrass theorem N is a Hilbert basis in L2(cid:0)[0, 1]; H; µ(cid:1). Since N is countable, we
conclude that L2(cid:0)[0, 1]; H; µ(cid:1) is a separable quaternionic Hilbert space.
Definition 2.7. Let H be a quaternionic Hilbert space. A map T : H → H is said to be right
H- linear or quaternionic operator, if
Moreover, T is said to be bounded or continuous, if there exist a α > 0 such that
T (x + yq) = T (x) + T (y) q, for all x ∈ H.
kT xk ≤ αkxk, for all x ∈ H.
We denote the class of all bounded operators on H by B(H) and it is a real Banach algebra
with respect to the operator norm defined by
kT k = sup(cid:8)kT xk : x ∈ H, kxk ≤ 1(cid:9).
For every T ∈ B(H), by the quaternionic version of the Riesz representation theorem [7,
Theorem 2.8], there exists a unique operator denoted by T ∗ ∈ B(H), called the adjoint of T
satisfying,
hx, T yi = hT ∗x, yi, for all x, y ∈ H.
If T ∈ B(H), then the null space of T is defined by N (T ) = {x ∈ H : T x = 0} and the range
space of T is defined by R(T ) = {T x : x ∈ H}. A closed subspace M of H is said to be
invariant subspace of T , if
T (x) ∈ M,
for every x ∈ M.
STRONGLY IRREDUCIBLE FACTORIZATION
5
Furthermore, M is said to be reducing subspace of T if M is an invariant subspace of both
T and T ∗.
Definition 2.8. Let T ∈ B(H). Then T is said to be
(1) self-adjoint, if T ∗ = T ,
(2) anti self-adjoint, if T ∗ = −T ,
(3) normal if T ∗T = T T ∗,
(4) positive, if T ∗ = T and hx, T xi ≥ 0 for all x ∈ H,
(5) orthogonal projection, if T ∗ = T and T 2 = T ,
(6) partial isometry, if kT xk = kxk, for all x ∈ N (T )⊥,
(7) unitary, if T ∗T = T T ∗ = I.
Suppose that T ∈ B(H) is positive, then by [7, Theorem 2.18], there exists a unique positive
operator S ∈ B(H) such that S2 = T . Such an operator S is called the positive square root
1
2 . In fact, for every T ∈ B(H), the modulus T is defined
of T and it is dentoed by S := T
as the positive square root of T ∗T , that is, T := (T ∗T )
1
2 .
We know from the well known Cartesian decomposition that every bounded normal op-
erator on a complex Hilbert space can be decomposed uniquely as A + iB, where A, B are
bounded self-adjoint operators. There is a quaternionic analog of this result in which the role
of 'i' is replaced by an anti self-adjoint unitary operator.
Theorem 2.9. [7, Theorem 5.9] Let T ∈ B(H) be normal. Then there exists an anti self-
adjoint unitary operator J ∈ B(H) such that T J = JT , T ∗J = JT ∗ and
T =
1
2
(T + T ∗) +
1
2
JT − T ∗.
Here J is uniquely determined by T on N (T −T ∗)⊥. Moreover, the operators (T +T ∗), T −T ∗
and J commute mutually.
Definition 2.10. [3, Definition 4.8.1] For every T ∈ B(H) and q ∈ H, let us associate an
operator ∆q(T ) := T 2 − 2re(q)T + q2I. Then
(1) the spherical spectrum of T is defined as,
σS(T ) =(cid:8)q ∈ H : ∆q(T ) is not invertible in B(H)(cid:9).
So the spherical resolvent of T denoted by ρS(T ) is the complement of σS(T ). That
is,
ρS(T ) := H \ σS(T ).
(2) The spherical point spectrum of T is defined by
The spherical spectrum σS(T ) is a nonempty compact subset of H.
σpS (T ) :=(cid:8)q ∈ H : N (∆q(T )) 6= {0}(cid:9).
Now we recall the notion of slice Hlibert space associated to the given quaternionic Hilbert
space H, anti self-adjoint unitary operator J ∈ B(H) and m ∈ S.
Definition 2.11. [7, Definition 3.6] Let m ∈ S and J ∈ B(H) be an anti self-adjoint unitary
operator. These subsets HJ m
± of H associated with J and m are defined by setting
HJ m
± := {x ∈ H : J(x) = ± x · m}.
Remark 2.12. For each x ∈ H and m ∈ S, define x± := 1
2 (x ∓ Jx · m), then
(2.4)
J(x±) =
1
2(cid:0)Jx ∓ J 2x · m(cid:1) =
1
2(cid:0)Jx ± x · m(cid:1) = ± x± · m.
6
P. SANTHOSH KUMAR
It implies that x± ∈ HJ m
± is
a non-trivial closed subsets of H. In fact, we see that the inner product on H restricted to
HJ m
± . Since x 7→ ± x · m and x 7→ Jx are continuous, then HJ m
± is Cm-valued as follows: let α + mβ ∈ Cm and u, v ∈ HJ m
± for m ∈ S, then
hu, vi(α ± mβ) = αhu, vi ± βhu, v · mi
= αhu, vi ± βhu, Jvi
= αhu, vi ∓ βhJu, vi
(since J ∗ = −J)
= αhu, vi ∓ βhu · m, vi
= (α ± mβ)hu, vi
(since m = −m).
Then by being a closed subspace of H, we conclude that HJ m
Cm. These Hilbert spaces HJ m
has the following decomposition [7, Lemma 3.10]
± is a Hilbert space over the field
± are known as slice Hilbert spaces. As a Cm- Hilbert space H
(2.5)
H = HJ m
+ ⊕ HJ m
− , for any m ∈ S.
Furthermore, if N is a Hilbert basis of HJ m
that mn = −nm, is a Hilbert basis of HJ m
Hilbert basis of H.
+ , then N · n = {z · n : z ∈ N }, where n ∈ S such
− . From Equation (2.5), it follows that N is also
We denote the class of all bounded Cm- linear operators on HJ m
+ by B(HJ m
proposition develop a technique to extend Cm- linear operator on HJ m
the quaternionic operator on H.
+ ). The following
+ (for any m ∈ S) to
Proposition 2.13. [7] Let J ∈ B(H) be anti self-adjoint unitary and m ∈ S. If T ∈ B(HJ m
+ ),
then there exist a unique quaternionic operator eT ∈ B(H) such that eT (x) = T (x), for all
+ . The following additional facts hold.
x ∈ HJ m
(1) keT k = kT k.
(2) JeT = eT J.
(3) (eT )∗ = fT ∗.
(5) If T is invertible, then eT is invertible and the inverse is given by
+ ), then fST = eSeT
(4) If S ∈ B(HJ m
(eT )−1 = gT −1.
On the other hand, let V ∈ B(H), then V = eU for some U ∈ B(HJ m
Precisely, the extension eT of the operator T is defined as,
eT (x) = eT (x+ + x−) = eT (x+) + eT (x−)
= eT (x+) − eT (x− · n) · n
= T (x+) − T (x− · n) · n,
+ ) if and only if JV = V J.
for all x = x+ + x− ∈ HJ m
+ ⊕ HJ m
− .
Note 2.14. In the case of normal operator T ∈ B(H), there exists an anti self-adjoint unitary
operator J ∈ B(H) commutes with T by Theorem 2.9. Then by Proposition 2.13 there is a
complex linear operator, we denote it by T+ ∈ B(HJ i
+ ) such that T = eT+.
STRONGLY IRREDUCIBLE FACTORIZATION
7
3. Factorization in a strongly irreducible sense
One of the fundamental result in the direction of factorizing quaternionic operators is the
well known polar decomposition theorem [7, 14]. It states that if T is bounded or densely
defined closed operator on a right quaternionic Hilbert space H, then there exists a unique
partial isometry W0 ∈ B(H) satisfying
(3.1)
T = W0T and N (T ) = N (W0).
Recently, the authors of [14] obtained a necessary and sufficient condition for any arbitrary
decomposition to coincide with the polar decomposition given in Equation (3.1). We recall
the result here.
Theorem 3.1. [14, Theorem 5.13] Let T be a bounded or densely defined closed operator
defined on a right quaternionic Hilbert space H. If W ∈ B(H) is a partial isometry satisfying
T = W T , then W = W0 if and only if either N (T ) = {0} or R(T )⊥ = {0}, where W0 is the
partial isometry satisfying Equation (3.1).
In this section, firstly, we adopt the notion of strong irreducibility [8] to the class of bounded
quaternionic operators, and prove a relation between strong irreducibility and the spherical
point spectrum. Later, we prove a factorization of quaternionic normal operator in a strongly
irreducible sense, by means of replacing the partial isometry W by a desirably small compact
perturbation of W and T is replaced by strongly irreducible operator. It is a quaternionic
extension (for normal operators) of the result proved in [15].
Definition 3.2. [8] Let T ∈ B(H). Then T is said to be
(1) irreducible, if there does not exist a nontrivial orthogonal projection P ∈ B(H) (i.e.,
P 6= 0 and P 6= I) such that P T = T P . Otherwise, T is called reducible.
(2) strongly irreducible, if there does not exist a nontrivial idempotent E ∈ B(H) (i.e.,
E 6= 0 and E 6= I) such that T E = ET . Otherwise, T is called strongly reducible.
It is clear from Definition 3.2 that every strongly irreducible operator is irreducible. Similar
to the classical setup, the class of strongly irreducible quaternionic operators is closed under
similarity invariance.
In order to describe strong irreducibility or irreducibility of quaternionic normal operators,
we show that it is enough to deal with the corresponding complex linear operator defined on
slice Hilbert space.
Lemma 3.3. Let J ∈ B(H) be anti self-adjoint unitary operator and S ∈ B(HJ i
+ ), then
(1) eS is irreducible if and only if S is irreducible.
(2) eS is strongly irreducible if and only if S is strongly irreducible.
Proof. Proof of (1) : Suppose that eS is irreducible, then we show that S is irreducible. If
there is an orthogonal projection 0 6= P ∈ B(H J i
2.13, we see that
+ ) such that SP = P S, then by Proposition
Further,
(eP )∗ = fP ∗ = eP and (eP )2 = fP 2 = eP .
This shows that 0 6= eP ∈ B(H) is an othogonal projection which commutes with eS. Since eS
is irreducible, we conclude that eP = I, the identity operator on H. Thus P is an identity
eSeP =gSP =gP S = ePeS.
8
P. SANTHOSH KUMAR
operator on HJ i
+ and hence S is irreducible. Now we prove contrapositive statement. Suppose
that {0} 6= M $ H is a reducing subspace for eS. Define
+ := {x ∈ M : Jx = x · i} = M ∩ HJ i
+ .
MJ i
For x ∈ MJ i
+ , we have
JS(x) = JeS(x) = eSJ(x) = eS(x · i) = eS(x) · i = S(x) · i
JS∗(x) = JfS∗(x) = fS∗J(x) = fS∗(x · i) = fS∗(x) · i = S∗(x) · i.
+ is a reducing subspace of S.
It is enough to show that MJ i
+ is a
and
This implies that MJ i
non-trivial proper subspace of HJ i
+ .
Claim: {0} 6= MJ i
+ $ HJ i
+ .
Firstly, we assume that MJ i
+ = {0}. In this case, MJ i
x · j ∈ MJ i
+ ⊕ M J i
to the fact that M is a non trivial subspace of H. Hence MJ i
+ , where i · j = −j · i. Therefore, M = M J i
− = {0} since x ∈ MJ i
− if and only if
− = {0}. This is a contradiction
+ 6= {0}.
B(HJ i
+ = HJ i
+ $ HJ i
+ = MJ i
Secondly, assume that MJ i
+ . Let y ∈ H with y = y+ + y−, where y± ∈ HJ i
± . We
+ and so y− ∈ M. Therefore y ∈ M. This is contradiction
know that y− · j ∈ HJ i
to the fact that M is a proper subspace of H. Hence MJ i
irreducible if and only if S is irreducible.
+ . We conclude that eS is
Proof of (2) : Suppose that eS is strongly irreducible. If there is an idempotent 0 6= E ∈
+ ) such that SE = ES, then by Proposition 2.13 we have that 0 6= eE ∈ B(H) with
eE2 = eE and
Since eS is strongly irreducible, we conclude that eE = I, the identity operator on H. This
F 6= I in B(H) such that eSF = FeS. Since R(F ) is a closed subspace of H, we have
eSeE =gSE =gES = eEeS.
implies that E is the identity operator on HJ i
Conversely, assume that S is strongly irreducible.
+ and hence S is strongly irreducible.
If there is an idempotent say 0 6=
+ ⊕ (R(F )J i
HJ i
+ = R(F )J i
− )⊥, where
Furthermore, if x ∈ R(F )J i
R(F )J i
+ = {x ∈ R(F ) : J(x) = x · i} = R(F ) ∩ HJ i
+ .
+ , then
and
JS(x) = JeS(x) = eSJ(x) = eS(x · i) = eS(x) · i = S(x) · i.
JS∗(x) = JfS∗(x) = fS∗J(x) = fS∗(x · i) = fS∗(x) · i = S∗(x) · i.
This shows that S is reducible. It is a contradiction to the fact that S is irreducible. Hence
(cid:3)
Now we prove the relation between spherical point spectrum and strong irreduicibility. It
is a quaternionic analogue of the result proved by F. Gilfeather in [8].
Theorem 3.4. Let T ∈ B(H) be normal. If the spherical point spectrum of T is empty set
(cid:0)i.e., σpS (T ) = ∅(cid:1), then T is strongly irreducible.
J ∈ B(H) commuting with T such that T = fT+, where T+ ∈ B(HJ i
Proof. Since T is normal, then by Note 2.14 there is an anti self-adjoint unitary operator
+ ) is normal operator.
Now we show that the point spectrum σp(T+) of T+ is empty. Suppose that λ ∈ σp(T+), then
T+(x) = x · λ, for some x ∈ HJ i
+ \ {0}.
eS is strongly irreducible.
STRONGLY IRREDUCIBLE FACTORIZATION
9
This implies that
∆λ(T )x =(cid:0)T 2 − 2re(λ)T + λ2I(cid:1)x
= x(cid:0)λ2 − λ(λ + λ) + λ2(cid:1)
= xλ2 − 2xλre(λ) + xλ2
= 0.
Equivalently, x ∈ N (∆λ(T )) 6= {0}. This shows that
λ ∈ σp(T+) ⇔ [λ] ∈ σpS (T ).
If follows that σ(T+) = ∅(cid:0)since σpS (T ) = ∅(cid:1). Since T+ is a bounded complex normal operator
with empty point spectrum, then by [8, Theorem 2] the operator T+ is strongly irreducible.
Finally, by Lemma 3.3 we conclude that T is strongly irreducible.
(cid:3)
Now we prove that every quaternionic normal operator can be factorized in a strongly
irreducible sense.
Theorem 3.5. Let T ∈ B(H) be normal and δ > 0. Then there exist a partial isometry W ,
a compact operator K with kKk < δ and a strongly irreducible operator S in B(H) such that
T = (W + K)S.
unitary operator J ∈ B(H) commuting with T such that T = fT+, where T+ ∈ B(HJ i
Proof. Since T is normal, by Theorem 2.9 and Note 2.14, there exists an anti self-adjoint
+ ) is
a normal operator. It is clear from [15, Theorem 1.1] that for a given δ > 0, there exists
a partial isometry W+, a compact operator K+ with kK+k < δ and a strongly irreducible
operator S+ in B(HJ i
(3.2)
+ ) such that
T+ = (W+ + K+)S+.
If we define W := fW+, then we see that W is a partial isometry as follows:
x ∈ N (W )⊥, there exists x± ∈ N (W )⊥ ∩ HJ i
± such that x = x+ + x− and
kW xk2 = kW+(x+) − W+(x− · j)jk2
for every
= kW+(x+)k2 + kW+(x− · j)k2
= kx+k2 + kx−k2
= kxk2.
Now we show that the operator defined by K := eK+ is a quaternionic compact operator
+ , there is a sequence of finite rank operators
+ ) converging to K+ (uniformly) with respect to the topology induced
on H. Since K+ is a compact operator on HJ i
{Fn : n ∈ N} ⊂ B(HJ i
from the operator norm. Then by (1) of Proposition 2.13 we see that
keFn − Kk = keFn − eK+k = kFn − K+k −→ 0, as n → ∞.
converges to K uniformly. Thus the operator K is compact and its norm is given by
This implies that the sequence {eFn : n ∈ N} ⊂ B(H) of finite rank quaternionic operators
Moreover, by Lemma 3.3, the quaternionic operator defined by S := eS+ ∈ B(H) is strongly
irreducible. Now we apply quaternionic extension to bounded complex linear operator T+
and use its factorization given in Equation (3.2), we conclude that
kKk = kK+k < δ.
T = eT+ = (fW+ + eK+)eS+ = (W + K)S.
10
P. SANTHOSH KUMAR
Hence the result.
(cid:3)
We illustrate our result with the following example.
Example 3.6. Let T : L2(cid:0)[0, 1]; H; µ(cid:1) → L2(cid:0)[0, 1]; H; µ(cid:1) be defined by
(T g)(x) =
xg(x) + 1
2
xy2g(y) dy,
if 0 ≤ x ≤ 1
3 ;
xy2g(y) dy,
if
1
3 ≤ x ≤ 1,
1R0
1R0
1
2
1R0
1
2
1R0
for all g ∈ L2(cid:0)[0, 1]; H; µ(cid:1). Then the adjoint of T is given by
(T ∗g)(x) =
xg(x) + 1
2
x2yg(y) dy,
if 0 ≤ x ≤ 1
3 ;
x2yg(y) dy,
if
1
3 ≤ x ≤ 1,
for all g ∈ L2(cid:0)[0, 1]; H; µ(cid:1). Clearly, T is normal. Suppose that δ = 1
strongly irreducible sense. Define the integral operator K : L2(cid:0)[0, 1]; H; µ(cid:1) → L2(cid:0)[0, 1]; H; µ(cid:1)
2 . Now we factorize T in a
by
(Kg)(x) =
1
2
1Z
0
xy g(y)dy, for all g ∈ L2(cid:0)[0, 1]; H; µ(cid:1).
It is well known that K is a compact operator. By using Caucy-Schwarz inequality, the norm
of K is computed as,
kKgk2 =(cid:16) 1Z
0
(Kg)(x)2 dx(cid:17) 1
2
0
0
2 ≤(cid:16) 1Z
≤(cid:16) 1Z
= kgk2(cid:16) 1Z
1Z
xy2 g(y)2dydx(cid:17) 1
1Z
2(cid:16) 1Z
g(y)2 dy(cid:17) 1
x2y2dydx(cid:17) 1
1Z
0
0
0
2
0
0
xy2dydx(cid:17) 1
2
=
1
3
kgk2.
3 < 1
This shows that kKk ≤ 1
2 . We recall that the class of all bounded H-valued measur-
abale functions on [0, 1] is denoted by L∞(cid:0)[0, 1]; H; µ(cid:1). For every f ∈ L∞(cid:0)[0, 1]; H; µ(cid:1), the
multiplication operator Mf : L2(cid:0)[0, 1]; H; µ(cid:1) → L2(cid:0)[0, 1]; H; µ(cid:1) defined by
Mf (g)(x) = f (x)g(x), for all g ∈ L2(cid:0)[0, 1]; H; µ(cid:1)
is a bounded quaternionic operator with the norm kMf k = kf k∞. The adjoint of Mf is given
by M ∗
f = Mf , where f (x) = f (x) for all x ∈ [0, 1]. Let φ(x) = x, for all x ∈ [0, 1] and the
characteristic function
3 ] =(cid:26) 1,
0,
χ
[0, 1
if x ∈ [0, 1
3 ];
otherwise.
STRONGLY IRREDUCIBLE FACTORIZATION
11
Then by the direct verification, we get that
(3.3)
T =(cid:16)Mχ
+ K(cid:17)Mφ.
[0, 1
3 ]
Note that Mχ
is a partial isometry, and since Mφ is normal with σpS (Mφ) = ∅, then Mφ
is strongly irreducible by Theorem 3.4. Therefore, the factorization of T given in Equation
(3.3) is a strongly irreducible factorization.
[0, 1
3 ]
Now we contruct an example of a non-normal operator by a slight modification of the
linear operator defined in Example 3.6 and compute its strongly irreducible factorization.
Example 3.7. Let us define T : L2(cid:0)[0, 1]; H; µ(cid:1) → L2(cid:0)[0, 1]; H; µ(cid:1) by
xg(x) + j
2
yg(y)dy,
if 0 ≤ x ≤ 1
3 ;
T (g)(x) =
yg(y)dy,
if 1
3 < x ≤ 1,
xR0
j
2
xR0
for all g ∈ L2(cid:0)[0, 1]; H; µ(cid:1) and suppose that δ = 1
quaternionic non-normal operator. Let g, h ∈ L2(cid:0)[0, 1]; H; µ(cid:1). Then
2 . Firstly, we show that T is a bounded
=
(cid:10)h, T g(cid:11)
1Z
3Z
=
0
1
0
1
3Z
h(x) (T g)(x) dx
h(x)hxg(x) +
j
2
xZ
0
yg(y) dyidx +
h(x)h j
2
xZ
0
1Z
1
3
yg(y) dyidx
=
h(x)xg(x) dx +
1
3Z
xZ
h(x)h j
2
yg(y)i dydx +
1Z
xZ
0
h(x)h j
2
yg(y)i dydx.
0
1
3
By Fubini's theorem, the above integral can be written as,
0
0
(cid:10)h, T g(cid:11)
3Z
=
1
0
=
1
3Z
0
h(y) yg(y) dy +
yh(y) g(y) dy +
1
3Z
0
1
3Z
0
Thus the adjoint of T is give by
1
3Z
y
h(x)h j
2
yg(y)i dxdy +
1Z
1
3
yZ
y− 1
3
h(x)h j
2
yg(y)i dxdy
1
3Z
y
y
h(x) dxi g(y) dy +
h −j
2
1Z
1
3
y
yZ
y− 1
3
h(x) dxi g(y) dy.
(T ∗h)(y) =
1
3Ry
h(x) dx,
if 0 ≤ y ≤ 1
3 ;
h(x) dx,
if
1
3 ≤ y ≤ 1.
yg(y) − j
2 y
− j
2 y
yRy− 1
3
2
h −j
12
P. SANTHOSH KUMAR
It follows that T T ∗ 6= T ∗T . Now we show that T can be factorized in a strongly irreducible
sense. Firstly, we define K : L2(cid:0)[0, 1]; H; µ(cid:1) → L2(cid:0)[0, 1]; H; µ(cid:1) by
(Kg)(x) =
j
2
g(t) dt, for all g ∈ L2(cid:0)[0, 1]; H; µ(cid:1).
xZ
0
Our aim to show that K is a compact operator with kKk < 1
2 . Let {gn}n∈N be a sequence in
(3.4)
L2(cid:0)[0, 1]; H; µ(cid:1) with kgnk ≤ 1, for all n ∈ N. Then
gn(t) dt(cid:12)(cid:12)(cid:12) ≤
yZ
(Kgn)(x) =(cid:12)(cid:12)(cid:12)
xZ
(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)Kg(x) − Kg(y)(cid:12)(cid:12) =
g(t) dt −
xZ
(3.5)
j
2
j
2
0
0
0
for all x ∈ [0, 1] and n ∈ N. Further, by Holders inequality, we get
1
2
xZ
0
gn(x) dt ≤
1
2
,
g(t) dt(cid:12)(cid:12)(cid:12) ≤
1
2
xZ
y (cid:12)(cid:12)g(t)(cid:12)(cid:12) dt ≤
1
2
kgk2px − y.
It follows from Equations (3.4), (3.5) that the sequece {Kgn}n∈N is uniformly bounded and
equicontinuous. By Arzela-Ascoli's theorem, there is a subseqeuce {gnk } of {gn}n∈N such
that {Kgnk } converges uniformly. Thus K is a compact operator.
Now we compute the norm of K. Firstly, by applying the Fubini's theorem, we get the
adjoint of K as,
(K∗g)(x) =
−j
2
1Z
x
g(t) dt, for all g ∈ L2(cid:0)[0, 1]; H; µ(cid:1).
So the operator K∗K is given by,
(K∗Kg)(x) =
−j
2
1Z
x
(cid:16) j
2
tZ
0
g(s) ds(cid:17)dt =
1
4
1Z
x
tZ
0
g(s) dsdt
is an associated slice Hilbert space and let (K∗K)+ be the bounded complex linear operator
is a positive quaternionic compact operator. We know from [13, Corollary 2.13] that L2(cid:0)[0, 1]; C; µ(cid:1)
on L2(cid:0)[0, 1]; C; µ(cid:1) such that ^(K∗K)+ = K∗K. Then by norm of the Voterra integral operator
computed as in [9, Solution 188] and (1) of Proposition 2.13, we conclude that
kKk = kK∗Kk
1
2 = k(K∗K)+k
1
π2(cid:17) 1
2 =(cid:16) 1
2 =
1
π
<
1
2
.
and S = Mϕ, where ϕ(x) = x, for all x ∈ [0, 1]. Clearly, W is a
Let us take W := Mχ
partial isometry and since S is normal with σpS (S) = ∅, we see that S is strongly irreducible
from Theorem 3.4. Finally, we have that
[0, 1
3 ]
We pose the following question.
T = (W + K)S.
Question 3.8. Let T : D(T ) ⊆ H → H be densely defined closed right H- linear operator
(need not be normal), where D(T ) is the domain of T . Then, can T be factorized in a strongly
irreducible sense ?
STRONGLY IRREDUCIBLE FACTORIZATION
13
We expect that, by using the notion of quaternionic Cowen-Douglas operators related to
geometry of quaternionic Hilbert spaces developed in [10] and further suitable arguments,
may achieve affirmative answer to the Question 3.8.
4. Riesz Decomposition Theorem
In this section we prove Riesz decomposition theorem for bounded quaternionic operators
on right quaternionic Hilbert spaces and obtain a sufficient condition for strong irreducibility.
We recall some definitions and known results form [2, 3, 7] that are useful to establish our
result.
Definition 4.1. [2] Let U ⊆ H be an open set. Then
(1) U is said to be a slice domain or s-domain, if U is a domain in H such that U ∩ R 6= ∅
and U ∩ Cm is domain in Cm, for all m ∈ S.
(2) A real differentiable function f : U → H is said to be
(i) left s-regular, if for every m ∈ S, the function f satisfies
1
2h ∂
∂x
f (x + my) + m
∂
∂y
f (x + my)i = 0
on U ∩ Cm.
(ii) right s-regular, if for every m ∈ S, the function f satisfies
1
2h ∂
∂x
f (x + my) +
∂
∂y
f (x + my) mi = 0
on U ∩ Cm.
Note that the class of left and right s-regular functions defined on U is denoted by RL(U )
and RR(U ), respectively. One can verify that RL(U ) is a right H-module, whereas RR(U ) is
a left H-module.
The following theorem describes a quaternionic analog of the Cauchy integral formula for
s-regular functions.
Theorem 4.2. [3, Theorem 4.5.3] Let U ⊆ H be an axially symmetric s-domain such that
the boundary ∂(U ∩ Cm) is the union of a finite number of continuously differentiable Jordan
curves for every m ∈ S. Let W be an open set containing U and take dsm = −ds · m for any
m ∈ S. We have the following:
(1) If f : W → H is left s-regular function, then
Z
∂(U ∩Cm) (cid:0)q2 − 2re(s)q + s2(cid:1)−1(q − s)dsmf (s),
(4.1)
f (q) = −
1
2π
for all q ∈ U .
(4.2)
f (q) = −
1
2π
for all q ∈ U .
Z
∂(U ∩Cm)
(2) If f : W → H is right s-regular function, then
f (s)dsm(q − s)(cid:0)q2 − 2re(s)q + s2(cid:1)−1,
Moreover, the integrals appear in Equations (4.1), (4.2) does not depend on the choice of the
imaginary unit m ∈ S and on U .
14
P. SANTHOSH KUMAR
Note that the kernels −(cid:0)q2−2re(s)q+s2(cid:1)−1(q−s) and −(q−s)(cid:0)q2−2re(s)q+s2(cid:1)−1 in the
Theorem 4.2 as the limit of corresponding Cauchy kernel series
qns−1−n and
s−1−nqn,
∞Pn=0
∞Pn=0
respectively for q < s.
The quaternionic functional calculus. Let H be a right quaternionic Hilbert space and
let N be a Hilbert basis of H. It is immediate to see that the class of all bounded right H-
linear operators denoted by B(H) is a two sided quaternionic Banach module with respect to
the module actions given by
(q · T )(x) := Xz∈N
z · q hz, T xi and (T · q)(x) := Xz∈N
T (z) · q hz, xi,
for all T ∈ B(H), q ∈ H, x ∈ H. In particular, for an identity operator I ∈ B(H), we have
(q · I)(x) = Xz∈N
z · q hz, xi = (I · q)(x), for all x ∈ H, q ∈ H.
Next, we recall the notion of the spherical resolvent operator and the spherical resolvent
equation which plays a vital role in establishing quaternionic functional calculus.
Definition 4.3. [3, Definition 4.8.3] Let T ∈ B(H) and s ∈ ρS(T ). Then the left spherical
resolvent operator is defined by
L (s, T ) := −∆s(T )−1(T − sI) =
S−1
and the right spherical resolvent operator by
R (s, T ) := −(T − sI)∆s(T )−1 =
S−1
∞Xn=0
∞Xn=0
T ns−1−n,
s−1−n T n,
for kT k < s.
Theorem 4.4. [3] Let T ∈ B(H) and s ∈ ρS(T ). Then the left and the right Spherical
resolvent operator satisfies the following relations:
(4.3)
and
(4.4)
L (s, T )s − T S−1
S−1
L (s, T ) = I
sS−1
R (s, T ) − S−1
R (s, T )T = I.
Proof. We know that T commutes with ∆s(T ) and so with ∆s(T )−1. By the Definition 4.3,
we get
L (s, T )s − T S−1
S−1
L (s, T ) = −∆s(T )−1(T − sI)s + T ∆s(T )−1(T − sI)
= ∆s(T )−1h − (T − sI)s + T (T − sI)i
= ∆s(T )−1∆s(T )
= I
STRONGLY IRREDUCIBLE FACTORIZATION
and
sS−1
R (s, T ) − S−1
R (s, T )T = s(T − sI)∆s(T )−1 − (T − sI)∆s(T )−1T
=hs(T − sI) − (T − sI)Ti∆s(T )−1
= ∆s(T )∆s(T )−1
= I.
Hence the result.
15
(cid:3)
Note 4.5. Let A be a bounded linear operator on some complex Hilbert space K and λ ∈
ρ(A), the resolvent set of A. Then (λI − A) is invertible and its inverse is given by the
following power series,
(λI − A)−1 =
∞Xn=0
1
λn+1 An,
for kAk < λ.
Moreover, if λ, µ ∈ ρ(A), then we have the following relation known as resolvent equation:
(4.5)
(λI − A)−1 − (µI − A)−1 = (µ − λ)(λI − A)−1(µI − A)−1.
One of the crucial observation in estabilshing the quaternionic functional calculus is the
spherical resolvent equation whichi is a quaternionic analogue of Equation (4.5). We recall
the result here.
Theorem 4.6. [2, Theorem 3.8] Let T ∈ B(H) and let s, p ∈ ρS(T ). Then the spherical
resolvent equation is given by
R (s, T )S−1
S−1
L (p, T )
R (s, T ) − S−1
=h(cid:0)S−1
Equivalently,
L (p, T )(cid:1)p − s(cid:0)S−1
R (s, T ) − S−1
L (p, T )(cid:1)i(p2 − 2re(s)p + s2)−1.
R (s, T )S−1
S−1
L (p, T )
= (s2 − 2re(p)s + p2)−1h(cid:0)S−1
R (s, T ) − S−1
L (p, T )(cid:1)p − s(cid:0)S−1
R (s, T ) − S−1
L (p, T )(cid:1)i.
Definition 4.7. Let T ∈ B(H), W ⊆ H be open and U ⊆ H be a domain in H. Then
(1) U is said to be a T -admissible open set, if U is axially symmetric s-domain that
contains the spherical spectrum σS(T ) such that the boundary ∂(U ∩ Cm) is the union
of a finite number of continuously differentiable Jordan curves, for every m ∈ S.
(2) A function f ∈ RL(W ) is said to be locally left regular function on σS(T ), if there is
T -admissible domain U in H such that U ⊆ W .
(3) A function f ∈ RR(W ) is said to be locally right regular function on σS(T ), if there
is T -admissible domain U in H such that U ⊆ W .
The class of all locally left and locally right regular functions on σS(T ) are denoted by RL
and RR
σS (T ) respectively.
σS(T )
Finally, by using Theorem 4.2 and the quaternionic version of Hahn Banach theorem [3,
Theorem 4.1.10], the quaternionic functional calculus is defined as below.
Definition 4.8. [3, Definition 4.10.4](quaternionic functional calculus) Let T ∈ B(H) and
U ⊂ H be a T -admissible domain. Then
S−1
L (s, T )dsm f (s), for all f ∈ RL
f (T ) =
(4.6)
σS (T )
1
2π
Z
∂(U ∩Cm)
16
and
(4.7)
f (T ) =
1
2π
Z
∂(U ∩Cm)
P. SANTHOSH KUMAR
f (s)dsm S−1
R (s, T ), for all f ∈ RR
σS(T ),
where dsm = −ds · m. Note that the integrals that appear in Equations (4.6), (4.7) are
independent of the choice of imaginary unit m ∈ S and T -admissible domain U .
Riesz decomposition theorem. Before proving our result, let us discuss the adjoint of the
operator f (T ) defined as in Definition 4.8.
Remark 4.9. Let T ∈ B(H) and W be an axially symmetric open set in H. For every
f : W → H, we define f : W → H by
f (q) = f (q), for all q ∈ W.
Let f ∈ RL(W ). Then for every m ∈ S and x, y ∈ R, we see that
∂
∂x
f (x + my)+m
f (x + my)
=
∂
∂y
∂
∂x
∂
∂x
∂
∂x
= 0.
=
=
f (x − my) + m
f (x − my) − m
f (x + my) + m
∂
∂y
∂
∂y
∂
∂y
f (x − my)
f (x − my)
f (x + my)
This show that f ∈ RL(W ). Further, if we assume that f is locally left regular function that
is, f ∈ RL
σS (T ) then by Definition 4.7, there is a T -admissible domain U such that U ⊆ W .
Since σS(T ) = σS(T ∗) and by the above arguments, we conclude that f ∈ RL
σS(T ). Now we
compute the adjoint of f (T ), whenever f ∈ RL
σS (T ), as follows:
(cid:10)x, f (T )y(cid:11) =
=
=
1
2π
1
2π
1
2π
Z
Z
Z
∂(U ∩Cm)
∂(U ∩Cm)
∂(U ∩Cm)
L (s, T )yE dsm f (s)
Dx, S−1
L (s, T )∗x, yE dsm f (s)
DS−1
R (s, T ∗)x, yE dsm f (s).
DS−1
If we put s = t, then ds = dt and dsm = −dt m. Since the integration over the domain
∂(U ∩ Cm) which is symmetric about the real line, we see that dtm = dtm. Thus above
STRONGLY IRREDUCIBLE FACTORIZATION
17
integral can be modified as,
(cid:10)x, f (T )y(cid:11) =
=
1
2π
1
2π
∂(U ∩Cm)
Z
R (t, T ∗)x, yEdtmf (t)
DS−1
Z
R (t, T ∗)xE
f (t) dtmDy, S−1
Z
R (t, T ∗)xE
2πDy,
=(cid:10) f (T ∗)x, y(cid:11),
f (t) dtm S−1
∂(U ∩Cm)
∂(U ∩Cm)
1
=
σS (T ). Similarly, the result holds true
σS(T ). For further details about algebraic properties of quaternionic functional calculus,
for all x, y ∈ H. Therefore, f (T )∗ = f (T ∗) for all f ∈ RL
for RR
we refer the reader to [3, Proposition 4.11.1].
In the following lemma, we show that for any compact set in H, there is an axially sym-
metric s-domain such that its intersection with Cm is a Cauchy domain in Cm, for every
m ∈ S.
Lemma 4.10. Let K be an axially symmetric compact subset of H and W be an axially
symmetric s- domain containing K. Then there is an axially symmetric s- domain U with
the boundary ∂(U ∩ Cm) is the union of a finite number of continuously differentiable Jordan
curves (for every m ∈ S) such that K ⊂ U and U ⊂ W .
Proof. Let us fix m ∈ S. We define Km := K ∩ Cm and Wm := W ∩ Cm. Then Km is a
compact subset of the open set Wm and hence Km is separated by a positive distance from
the closed set Cm \ Wm. That is,
d(Km, Cm \ Wm) := r, for some r > 0.
Here "d" is the Eucledian metric on the slice complex plane Cm. Being a compact set, if
Km is covered by open discs of radius r
2 with center in Km, then there is a finite collection
U = {U (m)
} of open discs such that
, · · · , U (m)
, U (m)
1
2
ℓ
Km ⊆
ℓ[t=1
U (m)
t
.
ℓSt=1
It is clear that the set defined by Um :=
U (m)
t
is open and the boundary ∂Um is a finite
t
union of arcs of ∂U (m)
(circles), for t = 1, 2, · · · ℓ. This follows that U m ⊆ Wm. Finally, we
define U := ΩUm. Since the center of each disc U (m)
lies on the real line, we see that U ∩R 6= ∅
and U ∩ Cm is a domain in Cm. That is, U is an axially symmetric s-domain containing K.
Moreover, the boundary ∂(U ∩Cm) is the union of finite number of continuously differentiable
Jordan curves by the above argument. The closure of U follows from Equation (2.3) as,
t
U = ΩUm = ΩU m
⊂ ΩWm = W.
Hence the result.
(cid:3)
Corollary 4.11. Let T ∈ B(H) and let W ⊆ H be an axially symmetric s-domain contining
the spherical spectrum σS(T ). Then there is a T -admissible domain U such that U ⊆ W .
18
P. SANTHOSH KUMAR
Proof. Since σS(T ) is an axially symmetric compact subset of H and W is an axially sym-
metric s-domain containing σS(T ), then by Lemma 4.10 there exists an axially symmetric
s-domain U containing the spherical spectrum σS(T ) such that U ⊆ W . Equivalently, U is a
T -admissible domain satisfying, U ⊆ W .
(cid:3)
Theorem 4.12. Let T ∈ B(H) and let σS(T ) = σ ∪ τ , where σ and τ are disjoint nonempty
axially symmetric closed subsets of σS(T ). Then there exist a pair {Mσ, Mτ } of non-trivial
invariant subspaces of T such that
σ = σS(T Mσ ) and τ = σS(T Mτ ).
Proof. Let m ∈ S. From hypothesis, it is clear that σ ∩Cm and τ ∩Cm are disjoint non-empty
compact subsets of the Hausdorff space Cm. Then there is a pair of disjoint open sets, say
O(m)
. By axially symmetric
propoerty of σ and τ , we can write
of Cm such that σ ∩ Cm ⊂ O(m)
and τ ∩ Cm ⊂ O(m)
and O(m)
σ
σ
τ
τ
σ = Ωσ∩Cm ⊂ Ω
O(m)
σ
and τ = Ωτ ∩Cm ⊆ Ω
.
O(m)
τ
Note that Ωσ∩Cm and Ωτ ∩Cm are nonempty disjoint s-domains in H. By Lemma 4.10, there
exist a pair of axially symmetric s-domains, denote them by Uσ and Uτ , containing compact
sets σ and τ respectively. Also, the boundaries ∂(Uσ ∩ Cm) and ∂(Uτ ∩ Cm) are the union of
finite number of continuously differentiable Jordan curves satisfying,
Now we define quaternionic operators corresponding to σ and τ as follows:
U σ ⊆ Ω
O(m)
σ
and U τ ⊆ Ω
.
O(m)
τ
(4.8)
and
(4.9)
Pσ =
1
2π
Z
∂(Uσ∩Cm)
Pτ =
1
2π
Z
∂(Uτ ∩Cm)
dsm S−1
R (s, T )
dsm S−1
R (s, T ),
where dsm = −ds · m. Now we prove the theorem in four steps.
Step I: P 2
σ = Pσ and P 2
τ = Pτ .
As Uσ is axially symmetric s-domain containing the compact set σ, then by applying
σ containing σ with the
σ ∩ Cm) is the union of a finite number of continously differentiable Jordan
Lemma 4.10, there exists another axially symmetric s-domian U ′
boundary ∂(U ′
curves such that U ′
σ ⊂ Uσ. If we define χσ : σs(T ) → H by
χσ(q) =(cid:26) 1, whenever q ∈ σ
otherwise,
0,
then clearly χσ ∈ RL
σS (T ) and Pσ = χσ(T ) by the quaternionic functional calculus.
Since the integral is independent of the choice of s-domain and also using the fact that χσ is
a locally left s-regular function, we can express it by
σS(T ) ∩ RL
(4.10)
χσ(T ) =
1
2π
Z
S−1
L (p, T ) dpm = Pσ,
∂(U ′
σ∩Cm)
STRONGLY IRREDUCIBLE FACTORIZATION
19
where dpm = −dp · m. From Equations (4.8), (4.10) and Theorem 4.6, we compute P 2
follows,
σ as
Z
S−1
L (p, T ) dpm(cid:17)
∂(U ′
σ∩Cm)
R (s, T )S−1
S−1
L (p, T ) dpm
P 2
σ =(cid:16) 1
2π
∂(Uσ∩Cm)
dsm S−1
=
1
4π2
=
1
4π2
∂(U ∩Cm)
∂(U ′
dsm
dsm
Z
Z
Z
2π
R (s, T )(cid:17) ·(cid:16) 1
Z
Z
σ∩Cm)
(cid:16)S−1
s(cid:16)S−1
Z
∂(Uσ∩Cm)
1
4π2
Z
−
∂(U ′
σ∩Cm)
dsm
∂(Uσ∩Cm)
∂(U ′
σ∩Cm)
R (s, T ) − S−1
L (p, T )(cid:17)p (p2 − 2re(s)p + s2)−1dpm
L (p, T )(cid:17)(p2 − 2re(s)p + s2)−1dpm.
R (s, T ) − S−1
Now we pause our computation for a while. Let us observe the following arguments. For
every s ∈ ∂(Uσ ∩ Cm), if we define a map gs : U ′
σ → H by
gs(q) = (q2 − 2re(s)q + s2)−1, for all q ∈ U ′
σ
then
∂
∂x
gs(x + my) +
gs(x + my)m
∂
∂y
= −(cid:2)(x + my)2 − 2re(s)(x + my) + s2(cid:3)−2(cid:0)2x + 2my − 2re(s)(cid:1)
−(cid:2)(x + my)2 − 2re(s)(x + my) + s2(cid:3)−2(cid:0)2xm − 2y − 2re(s)m(cid:1)m
= −(cid:2)(x + my)2 − 2re(s)(x + my) + s2(cid:3)−2(cid:0)2x + 2my − 2re(s)(cid:1)
+(cid:2)(x + my)2 − 2re(s)(x + my) + s2(cid:3)−2(cid:0)2x + 2my − 2re(s)(cid:1)
= 0,
for every x + my ∈ U ′
Similarly, if we define hs(q) = qgs(q) for all q ∈ U ′
∂(Uσ ∩ Cm). Thus by residue theorem, we see that
σ ∩ Cm. This shows that gs ∈ RR(U ′
σ), for every s ∈ ∂(Uσ ∩ Cm).
σ), for every s ∈
σ, then hs ∈ RR(U ′
Z
gs(p)dpm = 0 ;
hs(p)dpm = 0.
∂(U ′
σ∩Cm)
∂(U ′
σ∩Cm)
This implies that
(4.11)
and
(4.12)
1
4π2
Z
∂(Uσ∩Cm)
1
4π2
Z
∂(Uσ∩Cm)
dsmS−1
R (s, T )
∂(U ′
σ∩Cm)
dsmsS−1
R (s, T )
∂(U ′
σ∩Cm)
hs(p)dpm = 0
hs(p)dpm = 0.
Moreover, by [2, Lemma 3.18], we have
(4.13)
1
2π
Z
∂(Uσ∩Cm)
dsm(cid:2)sS−1
L (p, T ) − S−1
L (p, T )p(cid:3)gs(p) = S−1
L (s, T )
Z
Z
Z
20
P. SANTHOSH KUMAR
since S−1
(4.12) and above arguments, we get that
L (s, T ) ∈ B(H). Now we resume our computation of P 2
σ . From Equations (4.11),
P 2
σ = −
1
4π2
Z
Z
dsm
S−1
L (p, T )hs(p)dpm
∂(Uσ∩Cm)
∂(U ′
σ∩Cm)
+
1
4π2
Z
∂(Uσ∩Cm)
dsms
∂(U ′
σ∩Cm)
Z
Z
σ∩Cm) (cid:2)sS−1
Z
dsm(cid:2)sS−1
∂(Uσ∩Cm)
∂(U ′
dsm
(cid:16) 1
2π
=
1
4π2
=
=
1
2π
1
2π
Z
Z
Z
∂(U ′
σ∩Cm)
∂(U ′
σ∩Cm)
∂(Uσ∩Cm)
S−1
L (p, T )gs(p)dpm
L (p, T ) − S−1
L (p, T )p(cid:3)gs(p)dpm
L (p, T )p(cid:3)gs(p)(cid:17) dpm
L (p, T ) − S−1
S−1
L (p, T ) dpm, by Equation (4.13)
= Pσ.
Further, by Remark 4.9 and Equation (4.10), the operator Pσ is self-adjoint. Similarly the
result holds true for Pτ . Therefore, Pσ and Pτ are orthogonal projections in H, we call them
as Riesz projections.
Step II: Let Mσ := R(Pσ) and Mτ = R(Pτ ). Then H = Mσ ⊕ Mτ .
Let us define U := Uσ ∪ Uτ . Then U is an axially symmetric s-domain containing σS(T )
such that ∂(U ∩ Cm) is the union of a finite number of continuously differentiable Jordan
curves. Equivalently, U is a T -admissible domain. By the quaternionic functional calculus,
we have
Pσ + Pτ =
Z
= Z
∂(Uσ∩Cm)
∂(U ∩Cm)
dm S−1
R (s, T ) +
Z
∂(Uτ ∩Cm)
dsm S−1
R (s, T )
dsm S−1
R (s, T )
= I,
where I ∈ B(H) is the identity operator. It follows that
Pσ · Pτ = Pσ(I − Pσ) = Pσ − P 2
σ = 0.
If x ∈ H, then x is uniquely expressed as, x = Pσ(x) + Pτ (x). This implies that
Step III: We prove that Mσ and Mτ are invariant subspaces of T .
H = Mσ ⊕ Mτ .
STRONGLY IRREDUCIBLE FACTORIZATION
21
It is enough to show that both Pσ and Pτ commute with T . For every x, y ∈ H, we
compute that
Z
2π
∂(U ′
S−1
σ∩Cm)
(cid:10)x, T Pσ(y)(cid:11) =(cid:10)T ∗(x), Pσ(y)(cid:11) =DT ∗x, (cid:16) 1
Z
σ∩Cm) (cid:10)T ∗x, S−1
Z
σ∩Cm) (cid:10)x, T S−1
=Dx, (cid:16) 1
L (p, T ) dpm(cid:17)yE
L (p, T )y(cid:11) dpm
L (p, T )y(cid:11) dpm
L (p, T ) dpm(cid:17)yE.
T S−1
1
2π
1
2π
Z
∂(U ′
∂(U ′
=
=
2π
∂(U ′
σ∩Cm)
This implies that
T Pσ =
=
=
∂(U ′
∂(U ′
T S−1
σ∩Cm)
Z
Z
σ∩Cm) (cid:0)S−1
Z
∂(U ′
σ∩Cm)
L (p, T )p − I(cid:1) dpm, by Equation (4.3)
S−1
L (p, T )dpm p,
since
dpm = 0.
L (p, T ) dpm, where dpm = −dp · m
Z
∂(Uσ∩Cm)
Similarly, we compute PσT as,
∂(Uσ∩Cm)
2π
=
1
2π
(cid:10)x, Pσ(T y)(cid:11) =Dx, (cid:16) 1
Z
Z
=Dx, (cid:16) 1
1
2π
=
2π
∂(Uσ∩Cm)
∂(Uσ∩Cm)
dsm S−1
Z
dsm (cid:10)x, S−1
dsm (cid:10)x, S−1
Z
R (s, T )(cid:17)(T y)E
R (s, T )T y(cid:11)
R (s, T )T y(cid:11)
R (s, T )T(cid:17)yE.
dsm S−1
∂(Uσ∩Cm)
Thus
PσT =
=
=
Z
Z
Z
∂(Uσ∩Cm)
∂(Uσ∩Cm)
∂(Uσ∩Cm)
dsm S−1
R (s, T )T, where dsm = −ds · m
dsm (cid:0)sS−1
R (s, T ) − I(cid:1), by Equation (4.4)
s dsm S−1
R (s, T ),
since
dsm = 0.
Z
∂(Uσ∩Cm)
22
P. SANTHOSH KUMAR
This shows that T Pσ = PσT . Therefore,
T Pτ = T (I − Pσ) = (I − Pσ)T = Pτ T.
As a result, we conclude that both Mσ and Mτ are invariant subspaces of T .
Step IV: Finally, we show that σ = σS(T Mσ ) and τ = σS(T Mτ ).
Suppose that q /∈ σ, then [q] /∈ σ since σ is axially symmetric. With out loss of generality,
we assume that there is an axially symmetric s-domain Uσ containing σ such that ∂(Uσ ∩ Cm)
is the union of a finite number of continuously differentiable Jordan curves for every m ∈ S.
Let us fix m ∈ S. Define the operator
Q(q)
σ :=
1
2π
Z
∂(Uσ∩Cm)
S−1
L (t, T ) dtm(cid:0)t2 − 2re(q)t + q2(cid:1)−1,
where dtm = −dt · m. We claim that Mσ is invariant subspace of Q(q)
σ . For any p ∈ Cm, if
we define a map ξp(t) =(cid:0)t2 − 2re(q)t + q2(cid:1)−1S−1
(cid:0)t2 − 2re(q)t + q2(cid:1)−1 ∈ Cm, whenever t ∈ Cm
L (p, t), for all t ∈ Uσ. It is clear that
and S−1
4.11.5]. So, by the quaternionic functional calculus, we deduce that
L (p, t) is a left s-regular function in variable t. Thus ξp ∈ RL(Uσ) by [3, Proposition
(4.14) ξp(T ) = Q(q)
σ S−1
L (p, T ) =
1
2π
Z
∂(Uσ∩Cm)
S−1
L (t, T ) dtm(cid:0)t2 − 2re(q)t + q2(cid:1)−1S−1
L (p, t).
This implies the following:
Q(q)
σ Pσ =
1
2π
Q(q)
σ S−1
L (p, T )dpm
∂(U ′
σ∩Cm)
∂(U ′
σ∩Cm)
Z
∂(Uσ∩Cm)
S−1
L (t, T ) dtm(cid:0)t2 − 2re(q)t + q2(cid:1)−1S−1
L (p, t)dpm,
Z
Z
Z
Z
=
1
4π2
=
=
1
2π
1
2π
S−1
L (t, T ) dtm(cid:0)t2 − 2re(q)t + q2(cid:1)−1(cid:16) 1
L (t, T ) dtm(cid:0)t2 − 2re(q)t + q2(cid:1)−1,
2π
S−1
∂(U ∩Cm)
∂(U ∩Cm)
by Equation (4.14)
Z
S−1
L (p, t)dpm(cid:17)
∂(U ′
σ∩Cm)
since
1
2π
Z
∂(Uσ∩Cm)
S−1
L (s, t) dsm = 1
= Q(q)
σ .
σ = Q(q)
Similarly, PσQ(q)
σ Mσ ∈ B(Mσ). Next, we show that ∆q(T )Mσ
is invertible. For this, let us define ϕq(t) = t2 − 2re(q)t + q2. Then ϕq is locally s-regular
function on σS(T ). Moreover, by following similar arguments as in Step III, we express that
σ . This means Q(q)
(4.15)
∆q(T )Mσ Pσ = ϕq(T )Pσ =
1
2π
S−1
L (p, T ) dpm(p2 − 2re(q)p + q2),
Z
∂(U ′
σ∩Cm)
STRONGLY IRREDUCIBLE FACTORIZATION
23
where dpm = −dp · m. Also we know that ϕq(p)S−1
variable p, for every t. By [3, Proposition 4.11.5], it follows that
L (t, p) is left s-regular function in the
(4.16)
∆q(T )Mσ S−1
L (t, T ) =
1
2π
L (p, T ) dpm ϕq(p)S−1
S−1
L (t, p).
Z
∂(U ′
σ∩Cm)
From Equations (4.15), (4.16) we compute ∆q(T )Mσ Q(q)
σ as follows:
∆q(T )Mσ Q(q)
σ
=
1
2π
=
1
4π2
=
=
1
2π
1
2π
Z
Z
Z
Z
∂(U ′
σ∩Cm)
∂(U ′
σ∩Cm)
∂(Uσ∩Cm)
∂(Uσ∩Cm)
∂(U ′
∆q(T )Mσ S−1
L (t, T ) dtm (cid:0)t2 − 2re(q)t + q2(cid:1)−1
σ∩Cm)
L (p, T ) dpm ϕq(p)S−1
S−1
Z
L (p, T ) dpm ϕq(p)(cid:16) 1
L (p, T ) dpm ϕq(p)(cid:0)p2 − 2re(q)p) + q2(cid:1)−1.
∂(Uσ∩Cm )
Z
2π
S−1
S−1
L (t, p) dtm(cid:0)t2 − 2re(q)t) + q2(cid:1)−1
L (t, p) dtm(cid:0)t2 − 2re(q)t) + q2(cid:1)−1(cid:17)
S−1
Since ϕq(p)(cid:0)p2 − 2re(q)p + q2(cid:1)−1 = 1, it follows that ∆q(T )Mσ Q(q)
can show that Q(q)
σ ∆q(T )Mσ = Pσ. In other words, for every x ∈ Mσ, we see that
σ = Pσ. Similarly, one
∆q(T Mσ )Q(q)
σ x = Q(q)
σ ∆p(T Mσ )x = x, for all x ∈ Mσ.
Thus ∆q(T Mσ ) is invertible and hence q ∈ ρS(T Mσ ). It follows that
(4.17)
σS(T Mσ ) ⊆ σ.
By the similar arguments, we achieve that
(4.18)
σS(T Mτ ) ⊆ τ.
Now we prove reverse inclusions. Suppose that q /∈ σS(T Mσ ) ∪ σS(T Mτ ). Then both
operators ∆q(T Mσ ) and ∆q(T Mτ ) are invertible. Hence ∆q(T ) ∈ B(H) is invertible since
H = Mσ ⊕ Mτ . Equivalently, q /∈ σS(T ). This shows that
σS(T ) ⊆ σS(T Mσ ) ∪ σS(T Mτ ) ⊆ σ ∪ τ = σS(T ).
Therefore by Equation (4.17), (4.18) and using the fact that σ and τ are disjoint, we conclude
that
σS(T Mσ ) = σ and σS(T Mτ ) = τ.
Hence the result.
(cid:3)
Corollary 4.13. Let T ∈ B(H). If the spherical spectrum σS(T ) is disconnected by a pair of
disjoint nonempty axially symmetric closed subsets, then T is strongly reducible.
Proof. From the hypothesis, assume that there is a pair {σ, τ } of disjoint nonempty axially
symmetric closed subsets of σS(T ) satisfying,
σS(T ) = σ ∪ τ.
Then by Theorem 4.12, there exist a pair of nontrivial mutually orthogonal invariant sub-
spaces Mσ and Mτ of T such that
σS(T Mσ ) = σ and σS(T Mτ ) = τ.
24
P. SANTHOSH KUMAR
Equivalently, T commutes with the corresponding projections Pσ and Pτ as shown in step III
of Theorem 4.12. This implies that T is strongly reducible.
(cid:3)
Acknowledgments. The author thanks NBHM (National Board for Higher Mathematics,
India) for financial support with ref No. 0204/66/2017/R&D-II/15350, and Indian Statistical
Institute Bangalore for providing necessary facilities to carry out this work.
The author would like to express his sincere gratitude to Prof. B.V. Rajarama Bhat for
helpful suggestions and constant support.
References
[1] S. L. Adler, Quaternionic quantum mechanics and quantum fields, Oxford University Press, New
York, 1995.
[2] D. Alpay, F. Colombo, Gantner. J and I. Sabadini et al., A new resolvent equation for the
S-functional calculus, J. Geom. Anal. 25 (2015),
[3] F. Colombo, I. Sabadini and D. C. Struppa, Noncommutative functional calculus, Progress in
Mathematics, 289, Birkhauser/Springer Basel AG, Basel, 2011.
[4] J. B. Conway, A course in functional analysis, Graduate Texts in Mathematics, 96, Springer-
Verlag, New York, 1985.
[5] W. R. Hamilton, Elements of quaternions. Part 1, reprint of the 1866 original, Cambridge Library
Collection, Cambridge University Press, Cambridge, 2009.
[6] D. Finkelstein, J.M. Jauch, S. Schiminovich and D. Speiser, Foundations of quaternion quantum
mechanics, J. Mathematical Phys. 3 (1962), 207 -- 220.
[7] R. Ghiloni, V. Moretti and A. Perotti, Continuous slice functional calculus in quaternionic Hilbert
spaces, Rev. Math. Phys. 25 (2013), no. 4, 83 pp.
[8] F. Gilfeather, Strong reducibility of operators, Indiana Univ. Math. J. 22 (1972/73), 393 -- 397.
[9] P. R. Halmos, A Hilbert space problem book, second edition, Graduate Texts in Mathematics,
19, Springer-Verlag, New York, 1982.
[10] B. Hou and G. Tian, Geometry and operator theory on quaternionic Hilbert spaces, Ann. Funct.
Anal. 6 (2015).
[11] ZL. Jiang, Topics in Operator Theory. Jilin, Changchun, China: Jilin University, 1979.
[12] J. Luo, J. Li and G. Tian, On a factorization of operators on finite dimensional Hilbert spaces,
Turkish J. Math. 40 (2016).
[13] G. Ramesh and P. Santhosh Kumar, Spectral theorem for quaternionic normal operators: Mul-
tiplication form, To appear in Bull. Sci. Math.
[14] G. Ramesh and P. Santhosh Kumar, On the polar decomposition of right linear operators in
quaternionic Hilbert spaces, J. Math. Phys. 57 (2016), no. 4, 043502, 16 pp.
[15] G. Tian, Y. Cao, Y. Ji and J. Li, On a factorization of operators as a product of an essentially
unitary operator and a strongly irreducible operator, J. Math. Anal. Appl. 429 (2015).
[16] N. A. Wiegmann, Some theorems on matrices with real quaternion elements, Canadian J. Math.
7 (1955)
[17] F. Zhang, Quaternions and matrices of quaternions, Linear Algebra Appl. 251 (1997), 21 -- 57.
Indian Statistical Institute Bangalore, Statistics and Mathematics Unit, 8th Mile, Mysore
Road, Bengaluru 560 059, India
E-mail address: [email protected]
|
1805.09548 | 2 | 1805 | 2018-06-05T07:51:02 | Brezis - Lieb spaces and an operator version of Brezis - Lieb's lemma | [
"math.FA"
] | The Brezis - Lieb spaces, in which Brezis - Lieb's lemma holds true for nets, are introduced and studied. An operator version of Brezis - Lieb's lemma is also investigated. | math.FA | math |
BREZIS - LIEB SPACES AND AN OPERATOR VERSION
OF BREZIS - LIEB'S LEMMA
E. Y. EMELYANOV1,2 AND M. MARABEH3
Abstract. The Brezis - Lieb spaces, in which Brezis - Lieb's lemma
holds true for nets, are introduced and studied. An operator version of
Brezis - Lieb's lemma is also investigated.
1. Introduction
Throughout the paper, (Ω, Σ, µ) stands for a measure space in which every
set A ∈ Σ of nonzero measure possesses a subset A0 ⊆ A, A0 ∈ Σ, such that
0 < µ(A0) < ∞. The famous Brezis - Lieb lemma [3, Thm.2] is known as
Theorem 1 [3, Thm.2], and as its corollary, Theorem 2 [3, Thm.1], and also
as Theorem 3 (cf. [12, Cor.3]), which is a corollary of Theorem 2.
Theorem 1 (Brezis - Lieb's lemma). Let j : C → C be a continuous func-
tion with j(0) = 0 such that, for every ε > 0, there exist two non-negative
continuous functions φε, ψε : C → R+ with
(1.1)
j(x + y) − j(x) 6 εφε(x) + ψε(y)
(∀x, y ∈ C).
Let gn and f be (C−valued) functions in L0(µ) such that gn
φε(gn), ψε(f ) ∈ L1(µ) for all ε > 0, n ∈ N; and let
a.e.
−−→ 0; j(f ),
ε>0,n∈N Z
sup
Ω
φε(gn(ω))dµ(ω) 6 C < ∞.
Then
(1.2)
n→∞Z
lim
Ω
j(f + gn) − (j(f ) + j(gn))dµ(ω) = 0.
For a proof of Theorem 1, see [3].
Date: June 6, 2018.
2010 Mathematics Subject Classification. Primary: 28A20, 46A19, 46B30, 46E30.
Key words and phrases. Brezis - Lieb's lemma, Banach lattice, uo-convergence, Brezis
- Lieb space.
1
2
E. Y. EMELYANOV1,2 AND M. MARABEH3
Theorem 2 (Brezis - Lieb's lemma for Lp (0 < p < ∞)). Suppose fn
and RΩ
fnpdµ 6 C < ∞ for all n and some p ∈ (0, ∞). Then
a.e.−−→ f
(1.3)
n→∞(cid:26)Z
lim
(cid:18)fnp − fn − f p(cid:19)dµ(cid:27) = Z
f pdµ.
Ω
Ω
We reproduce here the arguments from [3] since they are short and instruc-
tive. Take j(z) = φε(z) := zp and ψε(z) = Cεzp for a sufficiently large
Cε. Theorem 1 applied to gn = fn − f ensures f ∈ Lp(µ), which, in view of
(1.2), completes the proof of Theorem 2. Theorem 3 below is an immediate
corollary of Theorem 2 (cf. also [12, Cor.3]).
Theorem 3 (Brezis - Lieb's lemma for Lp (1 6 p < ∞)). Let fn
a.e.−−→ f in
Lp(µ) and kfnkp → kf kp, where kfnkp := (cid:20)RΩ
and fn ∈ fn. Then kfn − f kp → 0.
fnpdµ(cid:21)1/p
with fn ∈ Lp(µ)
Theorem 3 is a Banach lattice type result if a.e.−convergence is replaced
by uo−convergence (cf.
[9, Prop.3.1]).
It motivates us to investigate the
general class of Banach lattices, in which the statement of Theorem 3 yields.
Even more important reason for such investigation relies on the fact that
all the above versions of Brezis - Lieb's lemma in Theorems 1, 2, and 3, are
sequential due to the sequential nature of a.e.−convergence. It is worth to
mention that Corollary 1 may serves as an extension of the Brezis - Lieb
lemma (in form of Theorem 3) for nets.
In Section 2, we introduce Brezis - Lieb's spaces and their sequential version.
Then we prove Theorem 4 which gives an internal geometric characterization
of Brezis - Lieb's spaces. We also discuss possible extensions of Theorem 4
to locally solid Riesz spaces.
In Section 3, we prove Theorem 5 which can be seen as an operator version
of Theorem 1 in convergence spaces.
For the theory of vector lattices we refer to [1, 2] and for unbounded con-
vergences to [4, 5, 6, 10, 9, 8].
2. Brezis - Lieb spaces
Definition 1. A normed lattice (E, k · k) is said to be a Brezis - Lieb space
(shortly, a BL−space) (resp. σ-Brezis - Lieb space (σ-BL−space)) if, for
any net xα (resp, for any sequence xn) in X such that kxαk → kx0k (resp.
BREZIS - LIEB SPACES AND AN OPERATOR VERSION OF BREZIS - LIEB'S LEMMA3
kxnk → kx0k) and xα
(resp. kxn − x0k → 0).
uo
−→ x0 (resp. xn
uo
−→ x0) we have that kxα − x0k → 0
Trivially, any normed Brezis - Lieb space is a σ-BL-space, and any finite-
dimensional normed lattice is a BL-space. Taking into account that a.e.−con-
vergence for sequences in Lp is the same as uo−convergence [9, Prop.3.1],
Theorem 3 says exactly that Lp is a σ-BL-space for 1 6 p < ∞.
Example 1. The Banach lattice c0 is not a σ−Brezis - Lieb space. To see
this, take xn = e2n +
1
k ek and x =
1
k ek in c0. Clearly, kxk = kxnk = 1
n
Pk=1
∞
Pk=1
for all n and xn
uo
−→ x, however 1 = kx − xnk does not converge to 0.
A slight change of an infinite-dimensional BL-space may turn it into a
normed lattice which is even not a σ−BL-space.
Example 2. Let E be a Brezis - Lieb space, dim(E) = ∞. Let E1 = R⊕∞E.
uo
Take any disjoint sequence (yn)∞
−→ 0
in E [9, Cor.3.6]. Let xn = (1, yn) ∈ E1. Then kxnkE1 = sup(1, kynkE) = 1
uo
and xn = (1, yn)
−→(1, 0) =: x in E1, however kxn − xkE1 = k(0, yn)kE1 =
kynkE = 1 and so, xn does not converge to x in (E1, k · kE1). Therefore
E1 = R ⊕∞ E is not a σ−Brezis - Lieb space.
n=1 in E such that kynkE ≡ 1. Then yn
In order to characterize BL-spaces, we introduce the following definition.
Definition 2. A normed lattice (E, k · k) is said to have the Brezis - Lieb
ku0 + unk > ku0k for any
property (shortly, BL-property), whenever lim sup
disjoint normalized sequence (un)∞
n=1 in E+ and for any u0 ∈ E, u0 > 0.
n→∞
Clearly, every finite dimensional normed lattice E has the BL−property.
The Banach lattice c0 obviously does not have the BL−property. The mod-
ification of the norm in an infinite-dimensional Banach lattice E with the
BL−property, as in Example 2, turns it into a Banach lattice E1 = R ⊕∞ E
without the BL−property.
Indeed, take a disjoint normalized sequence
(yn)∞
is a disjoint normalized sequence in (E1)+ with lim sup
n=1 in E+. Let u0 = (1, 0) and un = (0, yn) for n > 1. Then (un)∞
n=0
ku0 + unk = 1. Re-
n→∞
markably, it is not a coincidence.
Theorem 4. For a σ−Dedekind complete Banach lattice E, the following
conditions are equivalent:
4
E. Y. EMELYANOV1,2 AND M. MARABEH3
(1) E is a Brezis - Lieb space;
(2) E is a σ−Brezis - Lieb space;
(3) E has the BL-property and the norm in E is order continuous.
Proof. (1) ⇒ (2) It is trivial.
(2) ⇒ (3) We show first that E has BL-property. Notice that in this part of
the proof the σ−Dedekind completeness of E will not be used. Suppose that
there exists a disjoint normalized sequence (un)∞
n=1 in E+ and a u0 > 0 in E
ku0+unk =
with lim sup
ku0+unk = ku0k. Since ku0k 6 ku0+unk, then lim
n→∞
n→∞
ku0k. Denote vn := u0 + un. By [9, Cor.3.6], un
Since E is a σ-BL-space and lim
n→∞
is impossible in view of kvn − u0k = ku0 + un − u0k = kunk = 1.
Assume that the norm in E is not order continuous. Then, by the Fremlin–
uo
−→ u0.
kvnk = ku0k, then kvn − u0k → 0, which
uo
−→ 0 and hence vn
Meyer-Nieberg theorem (see for example [2, Thm.4.14]) there exist y ∈ E+
and a disjoint sequence ek ∈ [0, y] such that kekk 6→ 0. Without lost of
generality, we may assume kekk = 1 for all k ∈ N. By the σ−Dedekind
completeness of E, for any sequence αn ∈ R+ there exist the following
vectors
(2.1)
x0 =
∞
_k=1
ek,
xn = α2ne2n +
∞
_k=1,k6=n,k6=2n
ek
(∀n ∈ N).
Now, we choose α2n > 1 in (2.1) such that kxnk = kx0k for all n ∈ N.
Clearly, xn
uo
−→ x0. Since E is a σ-BL-space then kxn − x0k → 0, violating
kxn − x0k = k(α2n − 1)e2n − enk = k(α2n − 1)e2n + enk > kenk = 1.
uo
−→ x and kxαk → kxk but kxα − xk 6→ 0. Then xα
Obtained contradiction shows that the norm in E is order continuous.
(3) ⇒ (1) If E is not a Brezis - Lieb space, then there exists a net (xα)α∈A
uo
in E such that xα
−→ x
and kxαk → kxk.
Notice that kxα − xk 6→ 0. Indeed, if kxα − xk → 0 then (xα)α∈A is
eventually in [−x, x] and then (xα)α∈A is almost order bounded. Since
uo
E is order continuous and xα
−→ x, then by [10, Pop.3.7.] kxα − xk → 0,
which is impossible. Therefore, without lost of generality, we may assume
that xα ∈ E+ and, by normalizing, also kxαk = kxk = 1 for all α.
Passing to a subnet, denoted again by xα, we may assume
(2.2)
kxα − xk > C > 0
(∀α ∈ A).
BREZIS - LIEB SPACES AND AN OPERATOR VERSION OF BREZIS - LIEB'S LEMMA5
Notice that x > (x − xα)+ = (xα − x)− uo
order continuity of the norm ensures
−→ 0, and hence (xα − x)− o
−→ 0. The
(2.3)
k(xα − x)−k → 0.
Denoting wα = (xα − x)+ and using (2.2) and (2.3), we may also assume
(2.4)
kwαk = k(xα − x)+k > C (∀α ∈ A).
In view of (2.4), we obtain
(2.5)
2 = kxαk + kxk > k(xα − x)+k = kwαk > C (∀α ∈ A).
Since wα
uo
−→(x − x)+ = 0 then, for any fixed β1, β2, ..., βn,
(2.6)
0 6 wα ∧ (wβ1 + wβ2 + ... + wβn)
o
−→ 0
(α → ∞).
uo
−→ x, then xα ∧ x
o
Since xα
−→ x. By the order
continuity of the norm, there is an increasing sequence of indices αn in A
with
uo
−→ x ∧ x = x and so xα ∧ x
(2.7)
kx − xα ∧ xk 6 2−n
(∀α > αn).
Furthermore, by (2.6), we may also suppose that
(2.8)
Since
∞
Xk=n+1
kwα ∧ (wα1 + wα2 + ... + wαn)k 6 2−n
(∀α > αn+1).
∞
Xk=1,k6=n
kwαn ∧ wαk k 6
n−1
Xk=1
kwαn ∧ (wα1 + ... + wαn−1)k+
kwαk ∧ (wα1 + ... + wαk−1)k 6 (n − 1) · 2−n+1 +
∞
∞
Xk=n+1
2−k+1 = n2−n+1,
(2.9)
the series
wαn ∧ wαk converges absolutely and hence in norm for any
Pk=1,k6=n
n ∈ N. Take
ωαn := (cid:18)wαn −
∞
Xk=1,k6=n
wαn ∧ wαk(cid:19)+
(∀n ∈ N).
(2.10)
First, we show that the sequence (ωαn)∞
ωαm ∧ ωαp = (cid:18)wαm −
wαm ∧ wαk(cid:19)+
∞
Xk=1,k6=m
Xk=1,k6=p
(wαm − wαm ∧ wαp)+ ∧ (wαp − wαp ∧ wαm)+ =
∧(cid:18)wαp −
∞
n=1 is disjoint. Let m 6= p, then
wαp ∧ wαk(cid:19)+
6
6
E. Y. EMELYANOV1,2 AND M. MARABEH3
(wαm − wαm ∧ wαp) ∧ (wαp − wαm ∧ wαp) = 0.
By (2.9),
∞
kwαn − ωαnk = (cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
wαn −(cid:18)wαn − wαn ∧
Xk=1,k6=n
wαn −(cid:18)wαn −
wαn ∧ wαk(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=1,k6=n
= (cid:13)(cid:13)(cid:13)(cid:13)
∞
∞
=
wαn ∧ wαk(cid:19)+(cid:13)(cid:13)(cid:13)(cid:13)
Xk=1,k6=n
∞
wαn ∧
6
wαn ∧ wαk(cid:13)(cid:13)(cid:13)(cid:13)
k
Xk=1,k6=n
wαn ∧ wαk k 6 n2−n+1.
(∀n ∈ N).
(2.11)
Combining (2.11) with (2.5) gives
2 > kwαnk > kωαnk > C − n2−n+1
(∀n ∈ N).
(2.12)
Passing to further increasing sequence of indices, we may assume that
kwαn k → M ∈ [C, 2]
(n → ∞).
Now
lim
n→∞(cid:13)(cid:13)(cid:13)(cid:13)
M −1x + kωαnk−1ωαn(cid:13)(cid:13)(cid:13)(cid:13)
= M −1 lim
n→∞
kx + ωαnk = [by (2.11)] =
M −1 lim
n→∞
kx + wαnk = [by (2.3)] = M −1 lim
n→∞
kx + (xαn − x)k =
M −1 lim
n→∞
kxαn k = M −1 = kM −1xk,
violating the the Brezis - Lieb property for u0 = M −1x and un = kωαnk−1ωαn,
n > 1. The obtained contradiction completes the proof.
(cid:3)
The next fact is a corollary of Theorem 4 which states a Brezis - Lieb's type
lemma for nets in Lp.
Corollary 1. Let fα
kfα − f kp → 0.
uo
−→ f in Lp(µ), (1 6 p < ∞), and kfαkp → kf kp. Then
We do not know where or not implication (2) ⇒ (3) of Theorem 4 holds true
without the assumption that the Banach lattice E is σ−Dedekind complete.
Question 1. Does every σ−Brezis - Lieb Banach lattice have order contin-
uous norm?
BREZIS - LIEB SPACES AND AN OPERATOR VERSION OF BREZIS - LIEB'S LEMMA7
In the proof of (2) ⇒ (3) σ−Dedekind completeness of E has been used
only for showing that E has order continuous norm. So, any σ−Brezis -
Lieb Banach lattice has the Brezis - Lieb property. Therefore, for answering
in positive the question of possibility to drop σ−Dedekind completeness
assumption in Theorem 4, it is sufficient to have the positive answer to the
following question.
Question 2. Does the Brezis - Lieb property imply order continuity of the
norm?
In the end of the section we discuss possible generalizations of Brezis - Lieb
spaces and Brezis - Lieb property. To avoid overloading the text, we restrict
ourselves with the case of multi-normed Brezis - Lieb lattices, postponing
the discussion of locally solid Brezis - Lieb lattices to further papers.
A multi-normed vector lattice (shortly, MNVL) E = (E, M) (see [5]):
(a) is said to be a Brezis - Lieb space if
[xα
uo
−→ x0 & m(xα) → m(x0) (∀m ∈ M)] ⇒ [xα
M−→ x0].
(b) has the Brezis - Lieb property, if for any disjoint sequence (un)∞
n=1 in
E+ such un does not converge in M to 0 and for any u0 > 0, there exists
m ∈ M such that lim sup
m(u0 + un) > m(u0).
n→∞
A σ-Brezis - Lieb MNVL is defined by replacing of nets with sequences.
By using the above definitions one can derive from Theorem 4 the following
result, whose details are left to the reader.
Corollary 2. For an MNVL E with a separating order continuous multi-
norm M, the following conditions are equivalent:
(1) E is a Brezis - Lieb space;
(2) E is a σ−Brezis - Lieb space;
(3) E has the Brezis - Lieb property.
3. Operator version of Brezis - Lieb's lemma in convergent
vector spaces
In this section, we consider both complex and real vector spaces and vector
lattices. A convergence "
c
−→" for nets in a set X is defined by the following
conditions:
(a) xα ≡ x ⇒ xα
c
−→ x, and
8
E. Y. EMELYANOV1,2 AND M. MARABEH3
c
−→ x ⇒ xβ
c
−→ x for every subnet xβ of xα.
if xα
cY−→ f (x) for every net xα in X.
(b) xα
A mapping f from a convergence set (X, cX ) into a convergence set (Y, cY )
cX−−→ x implies
is said to be cX cY −continuous (or just continuous),
f (xα)
cX−−→ x ⇒ x ∈ A. If the
A subset A of (X, cX ) is called cX−closed if A ∋ xα
set {x} is cX −closed for every x ∈ X then cX is called T1-convergence. It
is immediate to see that cX ∈ T1 iff every constant net xα ≡ x does not
cX−converge to any y 6= x.
Under convergence vector space (X, cX ) we understand a vector space X with
a convergence cX such that the linear operations in X are cX −continuous.
(E, cE ) is a convergence vector lattice if (E, cE ) is a convergence vector space
which is a vector lattice where the lattice operations are also cE−continuous.
For further references see [1, 2, 4].
Motivated by the proof of the famous Brezis - Lieb's lemma [3, Thm.2], we
present its operator version in convergent spaces.
Given a convergence complex vector space (X, cX ); two convergence complex
vector lattices (E, cE ) and (F, cF ), where F is Dedekind complete; an order
ideal E0 in E+ − E+; and a cE0 oF −continuous positive linear operator T :
E0 → F , where oF stands for the order convergence in F . Furthermore, let
J : X → E be cX cE−continuous, J(0) = 0, and, for every ε > 0, let there
exist two cX cE−continuous mappings Φε, Ψε : X → E+ with
(3.1)
J(x + y) − J x 6 εΦεx + Ψεy
(∀x, y ∈ X).
Theorem 5 (An operator version of Brezis - Lieb's lemma for nets). Let
X, E, E0, F , T : E0 → F , and J : X → E satisfy the above hypothe-
cX−−→ 0, let f ∈ X be such that
sis. Let (gα)α∈A be a net in X satisfying gα
J f , Φεgα, Ψεf ∈ E0 for all ε > 0, α ∈ A, and let some u ∈ F+ exist with
T Φεgα 6 u for all ε > 0, α ∈ A. Then
T(cid:18)J(f + gα) − (J f + J gα)(cid:19) oF−→ 0
(α → ∞).
Proof. It follows from (3.1) that
J(f + gα) − (J f + J gα) 6 J(f + gα) − J gα + J f 6 εΦεgα + Ψεf + J f ,
and hence
J(f + gα) − (J f + J gα) − εΦεgα 6 Ψεf + J f
(ε > 0, α ∈ A).
BREZIS - LIEB SPACES AND AN OPERATOR VERSION OF BREZIS - LIEB'S LEMMA9
Thus
(3.2)
0 6 wε,α := (cid:18)J(f + gα) − (J f + J gα) − εΦεgα(cid:19)+
6 Ψεf + J f
for all ε > 0 and α ∈ A. It follows from (3.2) and from cX cE−continuity of
J and Φε, that E0 ∋ wε,α
cE−→ 0 as α → ∞. Furthermore, (3.2) implies
(3.3)
J(f + gα) − (J f + J gα) 6 wε,α + εΦεgα
(ε > 0, α ∈ A).
Since T > 0 and T Φεgα 6 u, we get from (3.3)
(3.4) 0 6 T(cid:18)J(f + gα) − (J f + J gα)(cid:19) 6 T wε,α + εT Φεgα 6 T wε,α + εu
for all ε > 0 and α ∈ A. Since F is Dedekind complete and T is cE0 oF −conti-
nuous, T wε,α
oF−→ 0, and in view of (3.4)
0 6 (oF ) − lim sup
α→∞
T(cid:18)J(f + gα) − (J f + J gα)(cid:19) 6 εu
(∀ε > 0).
Then T(cid:18)J(f + gα) − (J f + J gα)(cid:19) oF−→ 0.
(cid:3)
(1) Replacing nets by sequences one can obtain a sequential version of
Theorem 5, whose details are left to the reader.
(2) In the case of F = R and X = E = L0(µ) with the almost everywhere
convergence, E0 = L1(µ), T f = R f dµ, and J : X → E given by
J f = j ◦ f , where j : C → C is continuous with j(0) = 0 such that
for every ε > 0 there exist two continuous functions φε, ψε : C → R+
satisfying
j(x + y) − j(x) 6 εφε(x) + ψε(y)
(∀x, y ∈ C),
we obtain Theorem 1, which is the classical Brezis - Lieb's lemma [3,
Thm.2], from Theorem 5, by letting Φε(f ) := φε ◦ f and Ψε(f ) :=
ψε ◦ f .
References
[1] C. D. Aliprantis and O. Burkinshaw, Locally solid Riesz spaces with ap-
plications to economics, second ed., Mathematical Surveys and Mono-
graphs, Vol. 105, American Mathematical Society, Providence, RI, 2003.
[2] C. D. Aliprantis and O. Burkinshaw, Positive operators, Springer, Dor-
drecht, 2006, Reprint of the 1985 original.
10
E. Y. EMELYANOV1,2 AND M. MARABEH3
[3] H. Brezis and E. Lieb, A relation between pointwise convergence of func-
tions and convergence of functionals, Proc. Amer. Math. Soc. 88(3),
486-490 (1983)
[4] Y. A. M. Dabboorasad and E. Y. Emelyanov, Survey on unbounded con-
vergence in the convergence vector lattices, Vladikavkaz Math. J. 20(2),
(2018)
[5] Y. A. M. Dabboorasad, E. Y. Emelyanov and M. A. A. Marabeh,
um-Topology in multi-normed vector lattices, Positivity 22(2), 653-667
(2018)
[6] Y. A. M. Dabboorasad, E. Y. Emelyanov and M. A. A.
Marabeh, uτ -Convergence in locally solid vector lattices, Positivity
https://doi.org/10.1007/s11117-018-0559-4 (2018)
[7] E. Y. Emelyanov and M. A. A. Marabeh, Two measure-free versions of
the Brezis-Lieb lemma, Vladikavkaz Math. J. 18(1), 21-25 (2016)
[8] N. Gao, D. H. Leung and F. Xanthos, Duality for unbounded order con-
vergence and applications, Positivity https://doi.org/10.1007/s11117-
017-0539-0 (2017)
[9] N. Gao, V. G. Troitsky and F. Xanthos, Uo-convergence and its appli-
cations to Ces´aro means in Banach lattices, Isr. J. Math. 220, 649-689
(2017)
[10] N. Gao and F. Xanthos, Unbounded order convergence and application
to martingales without probability, J. Math. Anal. Appl. 415, 931-947
(2014)
[11] M. Marabeh, Brezis-Lieb lemma in convergence vector lattices, Turkish
J. of Math. 42, 1436-1442 (2018)
[12] C. P. Niculescu, An overview of absolute continuity and its applications,
Internat. Ser. Numer. Math. 157, Birkhauser, Basel, 201-214 (2009)
1 Middle East Technical University, 06800 Ankara, Turkey
E-mail address: [email protected]
2 Sobolev Institute of Mathematics, 630090 Novosibirsk, Russia
E-mail address: [email protected]
3Department of Applied Mathematics, College of Sciences and Arts, Pales-
tine Technical University-Kadoorie, Tulkarem, Palestine
E-mail address: [email protected], [email protected]
|
1905.10559 | 1 | 1905 | 2019-05-25T08:54:29 | Some notes on $b$-weakly compact operators | [
"math.FA"
] | In this paper, we will study some properties of b-weakly compact operators and we will investigate their relationships to some variety of operators on the normed vector lattices. With some new conditions, we show that the modulus of an operator $T$ from Banach lattice $E$ into Dedekind complete Banach lattice $F$ exists and is $b$-weakly operator whenever $T$ is a $b$-weakly compact operator. We show that every Dunford-Pettis operator from a Banach lattice $E$ into a Banach space $X$ is b-weakly compact, and the converse holds whenever $E$ is an $AM$-space or the norm of $E^\prime$ is order continuous and $E$ has the Dunford-Pettis property. We also show that each order bounded operator from a Banach lattice into a $KB$-space admits a $b$-weakly compact modulus. | math.FA | math |
SOME NOTES ON b-WEAKLY COMPACT OPERATORS
MASOUMEH MOUSAVI AMIRI AND KAZEM HAGHNEJAD AZAR ∗
Abstract. In this paper, we will study some properties of b-weakly
compact operators and we will investigate their relationships to some
variety of operators on the normed vector lattices. With some new
conditions, we show that the modulus of an operator T from Banach
lattice E into Dedekind complete Banach lattice F exists and is b-weakly
operator whenever T is a b-weakly compact operator. We show that
every Dunford-Pettis operator from a Banach lattice E into a Banach
space X is b-weakly compact, and the converse holds whenever E is an
AM -space or the norm of E ′ is order continuous and E has the Dunford-
Pettis property. We also show that each order bounded operator from
a Banach lattice into a KB-space admits a b-weakly compact modulus.
Keywords: Banach lattice, order continuous norm, b-weakly compact
operator, Dunford-Pettis operator.
MSC(2010): Primary 46B42; Secondary 47B60.
1. Introduction
An operator T from a Banach lattice E into a Banach space X is said to be
b-weakly compact, if it maps each subset of E which is b-order bounded (i.e.
order bounded in the topological bidual E′′) into a relatively weakly compact
subset of X. Note that in [3], the class of b-weakly compact operators is
introduced by Alpay-Altin-Tonyali and several interesting characterizations
were given in [3, 5, 10, 11]. After that, a series of papers which gave different
characterizations of this class of operators were published [3, 4, 5, 6, 7, 8,
9, 10, 11]. The most beautiful property of the class of b-weakly compact
operators is that it satisfies the domination property as proved in [3]. But
one of shortcomings of this class is that the modulus of a b-weakly compact
operator need not be b-weakly compact. Note that each weakly compact
operator is b-weakly compact, but the converse is not true in general. In
[9], authors characterized Banach lattices on which each b-weakly compact
operator is weakly compact. In [11], authors proved that an operator T from
a Banach lattice E into a Banach space X is b-weakly compact if and only
if (T xn) is norm convergent for every positive increasing sequence (xn) of
the closed unit ball BE of E.
Date: Received: , Accepted: .
∗Corresponding author.
1
2
M. MOUSAVI AMIRI AND K. HAGHNEJAD AZAR
1.1. Some basic definitions. An element e > 0 in a Riesz space E is said
to be an order unit whenever for each x ∈ E there exists some λ > 0 with
x ≤ λe. For example ℓ∞ has order unit, but c0 has not. A sequence (xn)
in a vector lattice is said to be disjoint whenever xn ∧ xm = 0 holds for
n 6= m. Let E be a Riesz space. The subset E+ = {x ∈ E : x > 0} is
called the positive cone of E and the elements of E+ are called the positive
elements of E. For an operator T : E → F between two Riesz spaces we shall
say that its modulus T exists ( or that T possesses a modulus) whenever
T := T ∨ (−T ) exists in the sense that T is the supremum of the set
{−T, T } in L(E, F ). An operator T : E → F between two Riesz spaces
is called order bounded if it maps order bounded subsets of E into order
bounded subsets of F . An operator T : E → F between two Riesz spaces is
positive if T (x) ≥ 0 in F whenever x ≥ 0 in E. Note that each positive linear
mapping on a Banach lattice is continuous. It is clear that every positive
operator is order bounded, but the converse is not true in general. A Banach
lattice E has order continuous norm if kxαk → 0 for every decreasing net
(xα)α with inf α xα = 0. If E is a Banach lattice, its topological dual E′,
endowed with the dual norm and dual order is also a Banach lattice. A
Banach lattice E is said to be an AM -space if for each x, y ∈ E such that
x ∧ y = 0, we have kx + yk = max{kxk, kyk}. A Banach lattice E is said
to be KB-space whenever each increasing norm bounded sequence of E+ is
norm convergent. A subset A of a Riesz space E is called b-order bounded
in E if it is order bounded in the bidual E∼∼. An operator T : E → F
between two Banach spaces is called a Dunford-Pettis operator whenever
w−→ 0 implies T xn
k·k
−−→ 0. Recall that an operator T from a Banach lattice
xn
E into a Banach space X is said to be order weakly compact, if it maps each
order bounded subset of E into a relatively weakly compact subset of X,
i.e., T [−x, x] is relatively weakly compact in X for each x ∈ E+. Assume
that E and F are normed lattice. A positive linear operator T : E → F
is called almost interval preserving if T [0, x] is dense in [0, T x] for every
x ∈ E+. Let E be a vector lattice. A sequence {xn}∞
1 ⊂ E is called order
convergent to x as n → ∞ if there exists a sequence {yn}∞
1 such that yn ↓ 0
o1−→ x when {xn}
as n → ∞ and xn − x ≤ yn for all n. We will write xn
is order convergent to x. A sequence {xn} in a vector lattice E is strongly
o2−→ x whenever there exists a net
order convergent to x ∈ E, denoted by xn
{yβ}β∈B in E such that yβ ↓ 0 and that for every β ∈ B, there exists n0 such
that xn − x ≤ yβ for all n ≥ n0. It is clear that every order convergent
sequence is strongly order convergent, but two convergence are different, for
information see, [1]. A net (xα)α in Banach lattice E is unbounded norm
k·k
−−→ 0 for
convergent (or, un-convergent for short) to x ∈ E if xα − x ∧ u
un−→ x. This convergence
all u ∈ E+. We denote this convergence by xα
has been introduced and studied in [12, 17]. For terminology concerning
Banach lattice theory and positive operators, we refer the reader to the
excellent books of [2, 14, 16].
SOME NOTES ON b-WEAKLY COMPACT OPERATORS
3
2. Main Results
The collections of b-weakly compact operators, order weakly compact
operators, weakly compact operators and compact operators will be denoted
by Wb(E, X), Wo(E, X), W (E, X) and K(E, X), respectively, whenever
there is not confused. We have the following relationships between these
spaces:
K(E, X) ⊆ W (E, X) ⊆ Wb(E, X) ⊆ Wo(E, X).
In [3], authors show that the above inclusion may be proper. For ex-
ample, note that each weakly compact operator is a b-weakly operator,
but the converse may be false in general. Of course the identity opera-
tor I : L1[0, 1] → L1[0, 1] is a b-weakly operator, but is not weakly compact.
Now let E be a Banach lattice such that the norm of E′ is order continuous
and let X be a Banach space. Then, by [9, Theorem 2.2], it is obvious that
each b-weakly operator T : E → X is weakly compact.
The next example due to Z. L. Chen and A. W. Wickstead in [18] shows
that the order bounded b-weakly compact operators from a Banach lattice
into a Dedekind complete Banach lattice do not form a lattice, i.e., a mod-
ulus of a b-weakly compact operator need not be b-weakly compact.
Example 2.1. Let E = C[0, 1], F = l∞(Fn) where Fn = (l∞, k · kn) and
k(λk)kn = max{k(λk)k∞, n lim sup(λk)} for all (λk) ∈ l∞. Then for each
n ∈ N, Fn is a Dedekind complete AM -space, hence so is F . Define Tn :
E → Fn by Tn(f ) = (2n. RIn
k=1 ∈ Fn for all f ∈ E, where rn is the
n,th Rademacher function on [0, 1] and In = (2−n, 2−n+1).
Now define T : E → F by T (f ) = ( 1
n=1. Then T is a weakly
compact operator. So, T is a b-weakly compact operator and its modulus
T exists and T is not order weakly compact hence not b-weakly compact.
So, Wb(E, F ) is not a Riesz space.
n Tn(f ))∞
f.rkdt)∞
Alpay and Altin in [5] show that for b-weakly compact operator T from a
Banach lattice E into a Dedekind complete Banach lattice F , the modulus
of T exist and is b-weakly compact operator whenever F is AM -space with
order unit. Now in the following theorems, we show that T ∈ Wb(E, F )
whenever T ∈ Wb(E, F ) and with some new conditions, we show that
the modulus of T exists and is b-weakly operator whenever T is a b-weakly
compact operator.
Theorem 2.2. Let E and F be normed vector lattices. We have the follow-
ing assertions.
(1) If T : E → F is an order bounded operator and F is KB-space, then
T and T are b-weakly operator.
(2) If T is a b-weakly compact operator, then
T ∈ Wb(E, F )
4
Proof.
M. MOUSAVI AMIRI AND K. HAGHNEJAD AZAR
(1) Since every KB-space has order continuous norm, so F is a
Dedekind complete Banach lattice. Then, by Theorem 1.18 from [2],
T exists. Now, let (xn) be a positive increasing sequence in BE. By
Theorem 4.3 of [2], T +(xn) is a Positive increasing norm bounded
sequence in F . Since F is a KB-space, T +(xn) is norm convergent.
Thus, T + is b-weakly compact. Similarly, T − is b-weakly compact.
It follows from the identities T = T + + T − and T = T + + T − that
T and T are b-weakly compact operators, so we are done.
(2) Since 0 ≤ T −, T + ≤ T , then by using Corollary 2.9 from [3], T −
and T + are two b-weakly compact operators, so T is a b-weakly
compact operator.
(cid:3)
Theorem 2.3. Let T be an order bounded operator from Banach lattice E
into Dedekind complete Banach lattice F . If c0 dose not embed in F , then
T and T are b-weakly compact operators.
Proof. At first, assume that T is a positive operator. By Theorem 4.3 [2],
T is continuous. Thus by Proposition 1 [5], it suffices to show that (T xn)
is norm convergent to zero for each b-order bounded disjoint sequence (xn)
in E+. Now let (xn) be a b-order bounded and disjoint sequence in E+. It
follows that there is a 0 ≤ x′′ ∈ E′′ such that 0 ≤ xn ≤ x′′ for all n. Then
for each 0 ≤ x′ ∈ E′, we have
k
x′(
X
i=1
k
xn) = x′(
_
i=1
xn) ≤ x′′(x′),
for all k holds.
Hence x′(P∞
i=1 xn) < ∞ for all 0 ≤ x′ ∈ E′. Then for each 0 ≤ y′ ∈ F ′
we have Pk
i=1 y′(T xn) = Pk
i=1 T ′y′(xn) < +∞. It follows that the sequence
(sm = Pm
n=1 T xn) is weakly bounded, and by Theorem 2.5.5 [15], it is norm
bounded. Since c0 dose not embed in F , by Theorem 4.60 [2], F is a KB-
space. Since the sequence (sm) is positive, increasing and norm bounded, so
it is norm convergent, and so the sequence (Pk
n=m T xn) is norm convergent
to zero. It follows that lim kT xnk = 0. Therefore, T is a b-weakly compact
operator. Now, let T be an order bounded operator. By the identities
T = T + − T − and T = T + + T −, it follows that T and T are b-weakly
compact operators.
(cid:3)
Proposition 2.4. Let E and F be two Banach lattices. Then we have the
following assertions.
(1) If T is a positive and order weakly compact operator from E onto F ,
then F has σ-order continuous norm.
(2) If a positive b-weakly compact operator T : E → F between Banach
lattices is surjective, then the norm of F is σ-order continuous.
Proof.
(1) Assume that {yn}n ⊆ F with yn ↓ 0. Since T is surjective,
It is clear that
there is an element x1 ∈ E such that T x1 = y1.
SOME NOTES ON b-WEAKLY COMPACT OPERATORS
5
{yn}n ⊆ [0, y1] ⊆ T ([0, x1]). Since T ([0, x1]) is relatively weakly
w−→
compact, there is a subsequence {ynj }j of {yn}n such that ynj
y0 ∈ F . Since {ynj }j is a decreasing sequence, by Theorem 3.52
from [2], ynj
kynk → 0. Thus F has order continuous norm.
k·k
−−→ y0 ∈ F . It follows from yn ↓ 0 that y0 = 0. Hence
(2) Follows from (1).
(cid:3)
Now by [2, Theorem 4.11] we have the following result:
Corollary 2.5. Let E be a normed lattice with order continuous norm.
Then the norm completion of E has order continuous norm.
Proof. We prove that Theorem 4.11 (2) of [2] holds. Let 0 ≤ xn ↑≤ x holds
in E. It follows from [2, Corollary 4.10] that E is Dedekind complete. Put
y = sup xn. So, y − xn ↓ 0 in E. Therefore, ky − xnk → 0. We have
kxn − xmk ≤ kxn − yk + ky − xmk → 0,
hence {xn} is a norm Cauchy sequence.
(cid:3)
Proposition 2.6. Let E and F be two Banach lattices such that E has
order unit and F has order continuous norm. Then every order bounded
operator T : E → F is b-weakly compact.
Proof. Let A be a b-order bounded subset of E. Since E has an order unit,
A is order bounded in E. Therefore, T (A) is order bounded in F . Since F
has an order continuous norm, T (A) is a relatively weakly compact. Thus
T is b-weakly compact operator.
(cid:3)
Example 2.7. Every order bounded operator from ℓ∞ into c0 is b-weakly
compact.
Theorem 2.8. Let E and F be two Banach lattices where E has order
continuous norm. Let G be an order dense sublattice of E and let T be a
positive operator from E into F . If T ∈ Wb(G, F ), then T ∈ Wb(E, F ).
Proof. Let {xn} be a positive increasing sequence in E with supn kxnk < ∞.
Since G is order dense in E, by Theorem 1.34 from [2], we have
{y ∈ G : 0 ≤ y ≤ xn} ↑ xn,
i=1ymi and 0 ≤ T ∈ L(E, F ).
for each n. Let {ymn}∞
m=1 ⊂ G with 0 ≤ ymn ↑ xn for each n. Put
zmn = ∨n
It follows that zmn ↑m xn and
supm,n kzmnk ≤ supn kxnk < ∞. Now, if T ∈ Wb(G, F ), then {T zmn} is
norm convergent to some point y ∈ F . Now, we have the following inequal-
ities
kT xn − T zmnk ≤ kT kkxn − zmnk ≤ kT kkxn − ymnk → 0.
6
M. MOUSAVI AMIRI AND K. HAGHNEJAD AZAR
Thus by the following inequality proof holds
kT xn − yk ≤ kT xn − T zmnk + kT zmn − yk.
(cid:3)
c (E, F )) is the collection of operators T ∈ Lb(E, F ), which xn
Definition 2.9. Let E and F be two vector lattices. L(1)
L(2)
(xn
subsequence of {xn}.
c (E, F ) (resp.
o1−→ 0
o2−→ 0) whenever {xnk } is a
o2−→ 0) implies T xnk
o1−→ 0 (resp. T xnk
In [1], there are some examples which shows that two classifications of
operators L(1)
c (E, F ) and L(2)
c (E, F ) are different.
Theorem 2.10. Let E, F be two Banach lattices such that E has order
continuous norm. Then
(1) Wb(E, F )+ ⊆ L(2)
(2) If Wb(E, F ) is vector lattice and F Dedekind complete, then Wb(E, F )
c (E, F ).
is an order ideal of L(1)
c (E, F ) = L(2)
c (E, F ).
Proof.
(1) Let T be a positive b-weakly compact operator and let {xn} ⊂
E be a strongly order convergence sequence in E. Without lose of
o2−→ 0, which follows {xn} is norm conver-
generality, we set 0 ≤ xn
gent to zero. Set {xnj } as subsequence with P+∞
k=1 kxnj k < +∞.
Define ym = Pm
j=1 xnj . Then 0 ≤ ym ↑ and supm kymk < ∞.
Since T is b-weakly compact operator, {T ym} is norm convergent
to some point z ∈ F . Now by [13], page 7, it has a subsequence as
{T ymk } which is strongly order convergent to z ∈ F . Thus there is
{zβ} ⊂ F + and that for each β there exists n0 T ymk − z ≤ zβ ↓ 0
whenever k ≥ n0. If we set k ≥ k′ ≥ n0, then we have the following
inequalities
0 ≤ T xnmk
≤ T ymk − T ymk′
≤ T ymk − z + T ymk′ − z ≤ zβ + zβ ↓ 0,
which shows that T ∈ L(2)
c (E, F ) and proof immediately follows.
(2) By equality T = T + − T − and Theorem 1.7 from [1], we have
c (E, F ) = L(1)
L(2)
c (E, F ). Since Wb(E, F ) is a vector lattice, it fol-
lows from part (1) that Wb(E, F ) is a subspace of L(1)
c (E, F ). Now
proof follows from the fact that Wb(E, F ) satisfies the domination
property.
By using conditions of Theorem 2.10, we can design the following question.
(cid:3)
Question. Is Wb(E, F ) a band in L(1)
c (E, F ) = L(2)
c (E, F )?
Proposition 2.11. Let E and F be two Banach lattices such that F is a
KB-space. Then every bounded operator T : E → F is b-weakly compact.
SOME NOTES ON b-WEAKLY COMPACT OPERATORS
7
Proof. By using [5, Proposition 1], it is enough to show that {T xn}n is
norm convergent for each b-order bounded increasing sequence {xn}n in
E+. Let {xn}n be a b-order bounded increasing sequence in E+. Since F
is a KB-space, by [2, Theorem 4.63], there exists a KB-space G, a lattice
homomorphism R : E → G and an operator S : G → F such that T = S ◦ R.
Since R is a lattice homomorphism, R is a positive operator and therefore is
b-order bounded. Then R(xn) is a b-order bounded increasing sequence in
G. Since G is a KB-space, R(xn) is norm convergent in G. It follows that
S ◦ R(xn) is also norm convergent in F . Hence {T (xn)} is norm convergent
in F . This completes the proof.
(cid:3)
Example 2.12. Let E be a Banach lattice. Then every bounded operator
from E into ℓ1 is b-weakly compact.
In the following proposition, we show that each Dunford-Pettis operator
is b-weakly compact, the converse is not always true. In fact, the identity
operator of the Banach lattice ℓ2 is b-weakly compact, but it is not Dunford-
Pettis. Recall that if E is a Banach lattice and if 0 6 x′′ ∈ E′′, then the
principal ideal Ix′′ generated by x′′ ∈ E′′ under the norm k · k∞ defined by
ky′′k∞ = inf{λ > 0 :
y′′ ≤ λx′′}, y′′ ∈ Ix′′ ,
is an AM -space with unit x′′, which its closed unit ball coincides with the
order interval [−x′′, x′′].
Lemma 2.13. Let E be a Banach lattice. Then every b-order bounded
disjoint sequence in E is weakly convergent to zero.
Proof. Let {xn}n be a disjoint sequence in E such that {xn}n ⊆ [−x′′, x′′]
for some x′′ ∈ E′′. Let Y = Ix′′ ∩ E and equip Y with the order unit norm
k · k∞ generated by x′′. The space (Y, k · k∞) is an AM -space, so, Y ′ is an
AL-space and hence its norm is order continuous. Now, by [14, Theorem
2.4.14] we see that xn
(cid:3)
w
−→ 0.
Proposition 2.14. Every Dunford-Pettis operator from a Banach lattice E
into a Banach space X is b-weakly compact.
Proof. Let T be a Dunford-Pettis operator from a Banach lattice E into a
Banach space X. By [5, Proposition 1], it is enough to show that {T xn}n is
norm convergent to zero for each b-order bounded disjoint sequence {xn}n
in E+. Let {xn}n be a b-order bounded disjoint sequence in E+. As the
canonical embedding of E into E′′ is a lattice homomorphism, {xn}n is an
order bounded disjoint sequence in E′′. Thus by preceding lemma, {xn}n is
σ(E, E′) convergent to zero in E. Now, since T is Dunford-Pettis, {T xn}n
is norm convergent to zero. This completes the proof.
(cid:3)
As a consequence of [2, Theorem 5.82], [9, Theorem 2.2] and [11, Theorem
2.3], we have the following results.
8
M. MOUSAVI AMIRI AND K. HAGHNEJAD AZAR
Corollary 2.15. Let E be a Banach lattice and let X be a Banach space.
Then each b-weakly compact operator from E into X is Dunford-Pettis, if
one of the following assertions is valid:
(1) E is an AM -space.
(2) The norm of E′ is order continuous and E has the Dunford-Pettis
property (i.e. each weakly compact operator from a Banach space E
into another F is Dunford-Pettis).
For the next results we need the following lemma:
Lemma 2.16.
(1) If an operator T from a Banach space X into a Ba-
nach space Y is compact and T (X) is closed, then T (X) is finite-
dimensional. As a consequence, if T : X → Y is a surjective compact
operator between Banach spaces, then Y is finite-dimensional.
(2) If T : X → Y is a weakly compact operator between Banach spaces
and T (X) is closed, then T (X) is reflexive. As a consequence, if
T : X → Y is a surjective weakly compact operator between Banach
spaces, then Y is reflexive.
Proof.
(1) Let T : X → Y be a compact operator between Banach
spaces. Since T (X) = T (X), T (X) is a Banach space. If U denotes
the open ball of X, then T (U ) is an open set in T (X). On the other
hand, T (U ) is compact so, T (X) is locally compact and hence T (X)
is finite-dimensional.
(2) Let T : X → Y be a weakly compact operator between Banach
spaces. Since T (X) is closed, T (X) is a Banach space and from
equality T (X) = Sn∈N nT (BX), we see that T (BX) contains a closed
ball of T (X). On the other hand, T (BX) is weakly compact, so, that
closed ball is weakly compact, therefore T (X) is reflexive.
(cid:3)
Proposition 2.17. Let E be a Banach lattice and let X be a non-reflexive
Banach space. If T : E → X is a surjective b-weakly compact operator, then
the norm of E′ is not order continuous.
Proof. If the norm of E′ is order continuous then by using [9, Theorem 2.2]
and Lemma 2.16, X is reflexive which is a contradiction.
(cid:3)
Proposition 2.18. Let E, F be two Banach lattices and F ′ has order con-
tinuous norm. Suppose that T : E → F is an almost interval preserving,
injective and b-weakly compact operator which has a closed range. Then E
is reflexive.
Proof. Since T (E) is closed, T (E) is a Banach space, so, T1 : E → T (E)
1 : T (E)′ → E′
is a bijective operator between two Banach spaces. Then T ′
is bijective. Without lose generality we replace T ′ with T ′
1. On the other
hand, since T is an almost interval preserving, by Theorem 1.4.19 [14], T ′
is a lattice homomorphism, and so by Theorem 2.15 [2], T ′ and (T ′)−1
SOME NOTES ON b-WEAKLY COMPACT OPERATORS
9
are both positive operators. Since F ′ has order continuous norm, by [2,
Theorem 4.59], F ′ is a KB-space, and so T ′ is b-weakly compact operator
and the norm of E′ is order continuous. Since T is a b-weakly compact, by
[9, Theorem 2.2], T is weakly compact. So, T ′ is weakly compact. Now,
by Lemma 2.16, E′ is reflexive and then E is reflexive. This completes the
proof.
(cid:3)
Recall that a nonzero element x of a Banach lattice E is discrete if the
order ideal generated by x is equal to the subspace generated by x. The
vector lattice E is discrete if it admits a complete disjoint system of discrete
elements. For example the Banach lattice ℓ2 is discrete but L1[0, 1] is not.
Proposition 2.19. Let E be a Banach lattice and let X be a Banach space
and let T : E → X be an injective b-weakly compact operator which its range
is closed. If one the following conditions holds, then E is finite dimensional.
(1) E is an AM -space with order continuous norm.
(2) E is an AM -space and E′ is discrete.
Proof. According to the proof of Proposition 2.18, T ′ : X ′ → E′ is a surjec-
tive operator. By [11, Proposition 2.3], if one of the above conditions holds,
then T is a compact operator. Thus, T ′ is compact. Now, by Lemma 2.16,
E′ is finite dimensional. Hence, E is finite dimensional.
(cid:3)
Definition 2.20. An operator T : E → F between two normed vector
lattices is unbounded b-weakly compact if {T xn} is un-convergent for every
positive increasing sequence {xn}n in the closed unit ball BE of E.
For normed vector lattices E and F , the collection of unbounded b-weakly
compact operators will be denoted by U Wb(E, F ). If a Banach lattice F has
strong unit, by using Theorem 2.3 [17], we have Wb(E, F ) = U Wb(E, F ).
It is obvious that every b-weakly compact operator is unbounded b-weakly
compact, but the following example shows that the converse does not hold,
in general.
Example 2.21. Let Ic0 be the identity mapping from c0 into itself. Then
Ic0 is an unbounded b-weakly compact operator. But Ic0 is not a b-weakly
compact operator.
The following characterization is obvious.
Proposition 2.22. Let E and F be two normed vector lattices and let T be
an operator from E into F . Then the following are equivalent:
(1) T is unbounded b-weakly compact.
(2) {T xn} is norm convergent for each b-order bounded increasing se-
quence {xn} in E+.
It is easy to see that the class of unbounded b-weakly compact operators
satisfies the domination property.
10
M. MOUSAVI AMIRI AND K. HAGHNEJAD AZAR
Proposition 2.23. Let E and F be two normed vector lattices and let S
and T be two operators from E into F with 0 ≤ S ≤ T . If T is unbounded
b-weakly compact then S is likewise unbounded b-weakly compact.
Theorem 2.24. Let E, F be two Banach lattices and let T ∈ L(E, F ). If
for an ideal I of E the restriction T I : I → F is a surjective homomorphism
which is also a b-weakly compact operator, then T ∈ U Wb(E, F ).
Proof. Let {xn} be a positive increasing sequence in E with supn kxnk < ∞
and let x ∈ I. Then xn ∧ x ∈ I, 0 ≤ xn ∧ x ↑ and supn kxn ∧ xk ≤ kxnk < ∞.
Since T ∈ Wb(I, F ), {T (xn ∧ x)} is convergent for each x ∈ I. As T is
homomorphism and surjective, {T (xn) ∧ y} is convergent for all y ∈ F and
proof follows.
(cid:3)
Theorem 2.25. Let E and F be two Banach lattices and let T : E → F
be a surjective homomorphism. Then by one of the following conditions we
have T ∈ U Wb(E, F ).
(1) E is a KB-space.
(2) F has order continuous norm.
Proof. If E is a KB-space, then T ∈ Wb(E, F ) and proof follows.
Assume that F has order continuous norm. Let {xn} ⊂ E+ be an increasing
sequence such that sup kxnk < ∞ and let x ∈ E+. Set yn = xn ∧ x, which
follows that yn ↑≤ x and supn kynk ≤ kxk. Since T is lattice homomorphism,
T is positive, which follows T yn ↑≤ T x. By using Theorem 4.11 [2], {T yn}
is norm Cauchy, and so is norm convergence in F . On the other hand,
T (xn ∧ x) = T xn ∧ T x. Therefore, T xn ∧ y is norm convergent for each
y ∈ F .
(cid:3)
References
[1] Y. Abramovich and G. Sirotkin, On order convergence of nets, Positivity. 9 (2005),
287 -- 292.
[2] C.D. Aliprantis, O. Burkinshaw, Positive Operators, Springer, Berlin, 2006.
[3] S. Alpay, B. Altin, C. Tonyali, On property (b) of vector lattices, Positivity. 7 (2003),
135 -- 139.
[4] S. Alpay, B. Altin and C. Tonyali, A note on Riesz spaces with property-b, Czechoslo-
vak Math. J. 56 (2006), 765 -- 772.
[5] S. Alpay and B. Altin, A note on b-weakly compact operators, Positivity. 11 (2007),
575 -- 582.
[6] S. Alpay and Z. Ercan, Characterizations of Riesz spaces with b-property, Positivity.
13 (2009), 21 -- 30.
[7] B. Altin, Some properties of b-weakly compact operators, G.U. J. Science. 18 (2005),
391 -- 395,
[8] B. Altin, On b-weakly compact operators on Banach lattices, Taiwanese J. Math. 11
(2007), 143 -- 150.
[9] B. Aqzzouz, A. Elbour, On the weak compactness of b-Weakly Compact Operators,
Positivity. 14 (2010), 75 -- 81.
[10] B. Aqzzouz, A. Elbour, Some Properties of the Class of b-Weakly Compact Operators,
Complex Anal. Oper. Theory. 12 (2010), 1139 -- 1145.
SOME NOTES ON b-WEAKLY COMPACT OPERATORS
11
[11] B. Aqzzouz, M. Moussa and J. Hmichane, Some Characterizations of b-weakly com-
pact operators, Math. Reports. 62 (2010), 315 -- 324.
[12] Y. Deng, M. O,Brien and V. G. Troitsky, Unbounded norm convergence in Banach
lattices, Positivity. 21 (2017), 963 -- 974.
[13] N. Gao and F. Xanthos, Unbounded order convergence and application to martingales
without probability,J. Math. Anal. Appl. 415 (2014), 931 -- 947.
[14] P. Meyer-Nieberg, Banach lattices, Universitex. Springer, Berlin. MR1128093, 1991.
[15] R.E. Megginson, An Introduction to Banach space Theory, Springer-Verlag New
York. Inc, 1998.
[16] H. Schaefer,
lattices and positive operators, Springer-Verlag, Berlin and New York,
1974.
[17] M. Kandic, M.A.A. Marabeh and V. G. Troitsky, Unbounded norm topology in
Banach lattices, J. Math. Anal. Appl. 451 (2017), 259 -- 279.
[18] Z.L. Chen and A.W. Wickstead, Vector lattices of weakly compact operators on
Banach lattices, Amer. Math. Soc. 352 (1999), 397 -- 412.
(Masoumeh Mousavi Amiri) Department of Mathematics, University of Mo-
haghegh Ardabili, Ardabil, Iran.
E-mail address: [email protected]
(Kazem Haghnejad Azar) Department of Mathematics, University of Mohaghegh
Ardabili, Ardabil, Iran.
E-mail address: [email protected]
|
1104.2686 | 1 | 1104 | 2011-04-14T07:58:31 | Sequential Lower Semi-Continuity of Non-Local Functionals | [
"math.FA"
] | We give a necessary and sufficient condition for non-local functionals on vector-valued Lebesgue spaces to be weakly sequentially lower semi-continuous. Here a non-local functional shall have the form of a double integral of a density which depends on the function values at two different points.
The characterisation we get is essentially that the density has to be convex in one variable if we integrate over the other one with an arbitrary test function in it.
Moreover, we show that this condition is in the case of non-local functionals on real-valued Lebesgue spaces (up to some equivalence in the density) equivalent to the separate convexity of the density. | math.FA | math |
Sequential Lower Semi-Continuity of Non-Local
Functionals
Peter Elbau∗
We give a characterisation for non-local functionals
Abstract
J : Lp(X; R
n) → R ∪ {∞}, J (u) = ZX ZX
f (x, y, u(x), u(y)) dx dy
on Lebesgue spaces to be weakly sequentially lower semi-continuous. Essentially, the
requirement is that the functions
Φx,ψ : R
n → R, Φx,ψ(w) = ZX
f (x, y, w, ψ(y)) dy
are for every ψ ∈ Lp(X; R
n) for almost all x ∈ X convex.
Moreover, we show that this condition is in the case n = 1 (up to some equiva-
lence in the integrand f ) equivalent to the separate convexity of the function (w, z) 7→
f (x, y, w, z) for almost all (x, y) ∈ X ×X.
1. Introduction
The purpose of these notes is to study the properties of non-local functionals of the form
J : Lp(X; R
n) → R ∪ {∞}, J (u) =ZXZX
especially regarding the existence of minimising points.
f (x, y, u(x), u(y)) dx dy,
(1)
Such kind of functionals recently appeared in a derivative-free characterisation of the
Sobolev and the total variation seminorm [4, 10]. More precisely, it was shown that these
seminorms can be written as the limit of a sequence of non-local functionals which essentially
emerge from replacing the derivative in the seminorm by a difference quotient. As an
application of this result, it became possible to reformulate variational problems such as e.g.
the total variation regularisation for image denoising [12] by approximating the seminorm
therein with the corresponding non-local functional, thus leading to variational problems for
non-local functionals [1, 11].
As another example where such non-local variational problems arose, we mention the
variational formulation of neighbourhood filters [6].
In this case, the non-locality of the
functional was utilised to measure and, by minimising the functional, also to enforce sim-
ilarities of different regions in an image. For an overview of non-local functionals recently
introduced in image analysis, we refer to [3].
∗Johann Radon Institute for Computational and Applied Mathematics (RICAM), Linz, Austria, e-mail:
[email protected]
2
peter elbau
The main interest of this paper is the existence of minimising points of such functionals.
Following the direct method in the calculus of variations, the existence can be guaranteed by
imposing the condition that J is coercive, meaning that J (u) → ∞ whenever kukp → ∞,
and that J is sequentially lower semi-continuous with respect to the weak topology on
n) if p = ∞.
Lp(X; R
Our aim is therefore to find a good characterisation for the sequential lower semi-continuity
of a non-local functional.
n) if p ∈ (1, ∞) and with respect to the weak-star topology on L∞(X; R
Before doing so, we take in Section 3 a closer look at the conditions which we need
to ensure that the function (x, y) 7→ f (x, y, u(x), u(y)) is integrable for all functions u ∈
Lp(X; R
n). It turns out that we can reach integrability with a weaker condition than the
natural estimate of the form
f (x, y, w, z) ≤ α(x, y) + β(x)zp + β(y)wp + Cwpzp
with α ∈ L1(X × X), β ∈ L1(X), C ∈ R, unlike in the case of local functionals (i.e.
functionals of the form Jlocal : Lp(X; R
sort of bound is equivalent to the integrability, see e.g. Theorem 6.45 in [5].
n) → R, Jlocal(u) =RX flocal(x, u(x)) dx) where this
Then we turn to the sequential lower semi-continuity of non-local functionals. We show
in Section 4 that the functional J is sequentially lower semi-continuous with respect to the
strong topology if (in addition to some lower bound for the function f ) the map (w, z) 7→
f (x, y, w, z) is for almost all (x, y) ∈ X×X lower semi-continuous. Afterwards, we establish
in Section 5 the result that a non-local functional, fulfilling some regularity assumptions,
is sequentially lower semi-continuous with respect to the weak topology on Lp(X; R
n) if
p ∈ (1, ∞) and with respect to the weak-star topology on L∞(X; R
n) if p = ∞ if and only
if the function
Φx,ψ : R
n → R, Φx,ψ(w) =ZX
f (x, y, w, ψ(y)) dy
is for every ψ ∈ Lp(X; R
n) for almost all x ∈ X convex.
Finally, we discuss in Section 6 which functions f lead to the same non-local functional.
We make use of this ambiguity in the integrand to show that in the case n = 1 the condition
that Φx,ψ is for every ψ ∈ Lp(X) for almost all x ∈ X convex is (for sufficiently regular
functions f ) equivalent to the fact that there exists a function f , defining the same non-local
functional as f , such that the map (w, z) 7→ f (x, y, w, z) is for almost all (x, y) ∈ X × X
separately convex.
The first results in this direction (formulated on Sobolev instead of Lebesgue spaces)
are going back to Pablo Pedregal [9] where he gave an equivalent characterisation of the
sequential lower semi-continuity in terms of Jensen type inequalities for the integrand. In
later papers [7, 2, 8], it was further shown that at least in the homogeneous case, i.e. if the
integrand is not explicitly depending on the variables x and y, the functional is sequentially
lower semi-continuous if and only if the integrand is separately convex.
2. Definition of Non-Local Functionals
m, m, n ∈ ,
Throughout the paper, let X be a bounded, Lebesgue measurable subset of R
and p ∈ [1, ∞]. We consider X as the measure space defined by the Lebesgue measure L m
N for
on the σ-algebra of all Lebesgue measurable subsets of X. Moreover, we consider R
every N ∈ as the measure space defined by the Lebesgue measure L N on the Borel
σ-algebra of R
N .
sequential lower semi-continuity of non-local functionals
3
Let us now clarify what we mean by a non-local functional. Since we can interchange the
order of integration in (1), we may restrict our attention to functions f which are pairwise
symmetric.
Definition 1. We call a function f : X ×X ×R
n×R
n → R pairwise symmetric if
f (x, y, w, z) = f (y, x, z, w)
for all x, y ∈ X, w, z ∈ R
n.
(2)
Definition 2. Let f : X ×X ×R
(with respect to the product σ-algebra of the chosen σ-algebras on X and R
part f − fulfils that
n → R be a pairwise symmetric, measurable function
n) whose negative
n×R
ZXZX
Then we call
f −(x, y, u(x), u(y)) dx dy < ∞ for all u ∈ Lp(X; R
n).
(3)
J p
f : Lp(X; R
n) → R ∪ {∞}, J p
f (u) =ZXZX
f (x, y, u(x), u(y)) dx dy
the non-local functional on Lp(X; R
n) defined by the function f .
The pairwise symmetry of the integrand f is just introduced for convenience since a
function f and its symmetrisation f ,
f (x, y, w, z) =
1
2
( f (x, y, w, z) + f (y, x, z, w)),
would anyway define the same non-local functional.
We remark that the measurability of the function f guarantees that also the composition
f ◦ gϕ,ψ of f and the measurable function
gϕ,ψ : X ×X → X ×X ×R
n×R
n,
gϕ,ψ(x, y) = (x, y, ϕ(x), ψ(y))
is for all measurable functions ϕ, ψ : X → R
Definition 2 are well-defined.
n again measurable, so that the integrals in
In the following, we will try to characterise the functions f which fulfil the condition (3).
To formulate the results for the values p ∈ [1, ∞) and for p = ∞ in the same way, let us
introduce the function pq
n → [0, ∞], q ∈ [0, ∞], M ∈ (0, ∞), by
M : R
pq
M (w) = wq
if q ∈ [0, ∞) and
which has the nice property that
M (w) =(∞ if w > M,
if w ≤ M,
p∞
0
pp
M (ϕ(x)) dx = pp
M (kϕkp)
ZX
for every M ∈ (0, ∞) and ϕ ∈ Lp(X; R
n).
(4)
(5)
4
peter elbau
3. Integrability Conditions
We are looking for a criterion for a measurable function f : X ×X × R
fulfil the condition
n × R
n → [0, ∞) to
ZXZX
f (x, y, u(x), u(y)) dx dy < ∞ for every u ∈ Lp(X; R
n).
(6)
Since the fact that we use in the two last components of the integrand f the same function u is
only relevant when the values x and y are close to each other, we may try to consider instead
of (6) the stronger condition where we impose integrability also for different functions ϕ, ψ ∈
Lp(X; R
n) in these two components.
Proposition 3. Let f : X ×X ×R
n×R
n → [0, ∞) be a measurable function.
Then the following statements are equivalent:
i. The function f fulfils
ZXZX
f (x, y, u(x), u(y)) dx dy < ∞ for all u ∈ Lp(X; R
n).
ii. The function f fulfils
ZXZX
f (x, y, ϕ(x), ψ(y)) dx dy < ∞ for all ϕ, ψ ∈ Lp(X; R
n).
iii. There exists for every ψ ∈ Lp(X; R
n) and every M ∈ (0, ∞) a function αM,ψ ∈ L1(X)
and a constant Cψ ∈ (0, ∞) such that
f (x, y, w, ψ(y)) dy ≤ αM,ψ(x) + Cψ pp
M (w)
(7)
ZX
for almost all x ∈ X and all w ∈ R
n.
Proof. We start with the implication from (i) to (ii). Let us assume that we find two
n) such that the measurable function g : X×X → [0, ∞), g(x, y) =
functions ϕ, ψ ∈ Lp(X; R
f (x, y, ϕ(x), ψ(y)) fulfils
To prove the implication, it is enough to construct a function u ∈ Lp(X; R
n) with
ZXZX
g(x, y) dx dy = ∞.
ZXZX
f (x, y, u(x), u(y)) dx dy = ∞.
(8)
If there exists a measurable set A ⊂ X with RARAc g(x, y) dy dx = ∞, Ac = X \A, we
can simply choose u = χAϕ + χAc ψ, to get (8). Otherwise, if no such set A exists, we find
a decreasing sequence (Ak)k∈ of measurable sets Ak ⊂ X, k ∈ , such that Ak+1 ⊂ Ak,
L m(Ak+1) = 1
2
L m(Ak), and
ZAkZAk
g(x, y) dx dy = ∞ for all k ∈ .
sequential lower semi-continuity of non-local functionals
5
We next choose a subsequence (Akℓ )ℓ∈ of (Ak)k∈ such that the pairwise disjoint sets
Aℓ = Akℓ \Akℓ+1, ℓ ∈ , fulfil
Z AℓZ Aℓ
g(x, y) dx dy ≥
1
ℓ
for all
ℓ ∈ .
We further divide for every ℓ ∈ the set Aℓ× Aℓ into the measurable sets
Ej,ℓ = {(x, y) ∈ Aℓ × Aℓ : g(x, y) ∈ [j − 1, j)},
j ∈ .
Then we find for every ℓ ∈ a constant Nℓ ∈ such that
jL m(Ej,ℓ) ≥
1
2ℓ
.
Nℓ
Xj=1
(9)
(10)
Using now Lemma 20, we get for every ℓ ∈ a δℓ ∈ (0, ∞) such that the checkerboard
pattern Sδℓ ⊂ R
m defined in (44) fulfils
L 2m((Sδℓ ×Sc
δℓ) ∩ Ej,ℓ) ≥
1
8
L 2m(Ej,ℓ)
(11)
(Sδℓ ∩ Aℓ) ⊂ X and u = χSϕ + χSc ψ, where Sc = X \ S, we
(Sc
δℓ
∩ Aℓ), we have
g(x, y) dy dx ≥
∞
Xℓ=1ZSδℓ ∩ AℓZSc
δℓ
g(x, y) dy dx.
∩ Aℓ
j=1 Ej,ℓ, we get by definition (9) of the sets Ej,ℓ that
for all j ∈ [1, Nℓ] ∩ .
f (x, y, u(x), u(y)) dx dy ≥ZSZSc
Setting finally S = Fℓ∈
get (8). Indeed, since Sc ⊃Fℓ∈
ZXZX
Using then Aℓ × Aℓ ⊃FNℓ
Xℓ=1ZSδℓ ∩ AℓZSc
∩ Aℓ
∞
δℓ
g(x, y) dy dx ≥
g(x, y) dx dy
)∩Ej,ℓ
δℓ
≥
(j − 1)L 2m((Sδℓ ×Sc
δℓ) ∩ Ej,ℓ).
∞
∞
Xℓ=1
Xℓ=1
Nℓ
Xj=1ZZ(Sδℓ×Sc
Xj=1
Nℓ
Taking now the estimate (11) into account, we find with our choice (10) of Nℓ that
∞
Xℓ=1
Nℓ
Xj=1
(j − 1)L 2m((Sδℓ ×Sc
δℓ) ∩ Ej,ℓ) ≥
≥
∞
Nℓ
1
8
Xℓ=1
Xj=1
8 ∞
Xℓ=1
1
1
2ℓ
(j − 1)L 2m(Ej,ℓ)
− L m(X)2! = ∞.
This concludes the proof that (i) implies (ii). The converse direction is trivial.
Given condition (ii), we can use the result for local functionals. Indeed, we know that
Fψ(x, ϕ(x)) dx < ∞
ZX
6
peter elbau
for all ϕ ∈ Lp(X; R
n), where the function Fψ is for all ψ ∈ Lp(X; R
n) defined by
Fψ : X ×R
n → R ∪ {∞}, Fψ(x, w) =ZX
f (x, y, w, ψ(y)) dy.
Therefore, in the case p ∈ [1, ∞), there exist for every ψ ∈ Lp(X; R
and a constant Cψ ∈ (0, ∞) such that
n) a function αψ ∈ L1(X)
Fψ(x, w) ≤ αψ(x) + Cψwp
(12)
for almost all x ∈ X and all w ∈ R
every ψ ∈ L∞(X; R
n, see e.g. Theorem 6.45 in [5]. If p = ∞, we find for
n) and every constant M ∈ (0, ∞) a function αM,ψ ∈ L1(X) such that
Fψ(x, w) ≤ αM,ψ(x)
(13)
for almost all x ∈ X and all w ∈ R
function pp
form (7).
n with w ≤ M , see e.g. Theorem 6.47 in [5]. Using the
M defined in (4), the conditions (12) and (13) can be written in the condensed
The implication from (iii) to (ii) is finally found by direct calculation. From condition (7),
we get with the property (5) of the function pp
M that
ZXZX
f (x, y, ϕ(x), ψ(y)) dy dx ≤ kαM,ψk1 + Cψ pp
M (kϕkp) < ∞
for all ϕ, ψ ∈ Lp(X; R
n) if we choose in the case p = ∞ the constant M ≥ kϕk∞.
(cid:3)
We remark that condition (7) allows for non-integrable divergencies in the function
n, unlike in the study of local
(x, y) 7→ supw,z∈B f (x, y, w, z) even for bounded sets B ⊂ R
functionals where such a behaviour does not get along with the finiteness of the functional.
Since we do not want to deal with such divergencies, we give here additionally a stronger
condition on f .
n → R fulfils that there exist for every
Definition 4. If the function f : X × X × R
M ∈ (0, ∞) a constant C ∈ (0, ∞) and positive functions αM ∈ L1(X×X) and βM ∈ L1(X)
with
n × R
f (x, y, w, z) ≤ αM (x, y) + βM (x) pp
M (z) + βM (y) pp
M (w) + C pp
M (w) pp
M (z)
(14)
for almost all (x, y) ∈ X ×X and all w, z ∈ R
n, then we call f a p-bounded function.
Corollary 5. Every p-bounded, measurable function f : X ×X ×R
n×R
n → [0, ∞) fulfils
ZXZX
f (x, y, u(x), u(y)) dx dy < ∞ for all u ∈ Lp(X; R
n).
(15)
Proof. We may directly verify condition (15). Since the function f is p-bounded, we find
for every M ∈ (0, ∞) a constant C ∈ (0, ∞) and positive functions αM ∈ L1(X × X) and
βM ∈ L1(X) such that (14) holds. Then we get with the property (5) of the function pp
M
for every u ∈ Lp(X; R
n) that
ZXZX
f (x, y, u(x), u(y)) dx dy ≤ kαM k1 + 2kβM k1 pp
M (kukp) + C pp
M (kukp)2 < ∞
if we choose in the case p = ∞ the constant M greater than kuk∞.
(cid:3)
sequential lower semi-continuity of non-local functionals
7
To illustrate what kind of functions we are excluding with this stronger condition, let us
construct an example of a function f which is not p-bounded, but fulfils the integrability
condition (15).
Example 6. Let X = (0, 1) and choose n = 1. We define the function f by
f : X ×X ×R×R → [0, ∞),
f (x, y, w, z) =( 1
z
0
z ∈ [x, 1],
if
otherwise.
Then we have
ZXZX
f (x, y, ϕ(x), ψ(y)) dx dy =Zψ−1((0,1])Z ψ(y)
0
1
ψ(y)
for all functions ϕ, ψ ∈ Lp(X). In particular, f fulfils condition (15).
dx dy = L 1(cid:0)ψ−1((0, 1])(cid:1)
(16)
On the other hand, f cannot be p-bounded, since e.g.
sup
f (x, y, w, z) =
z∈[−1,1]
1
x
for all x, y ∈ X, w ∈ R,
which as a function of x ∈ X is not integrable as an estimate of the form (14) would require.
However, this construction only works if the function f depends on the variables x and y.
Otherwise, every divergency of f at finite values of w and z would lead to a non-integrable
divergency of (x, y) 7→ f (x, y, u(x), u(y)) for some function u ∈ Lp(X; R
In fact, if f
does not explicitly depend on x and y, then the p-boundedness of f is equivalent to the
integrability condition (15).
n).
Proposition 7. Let f : R
n×R
n → [0, ∞) be a measurable function. Then we have
ZXZX
f (u(x), u(y)) dx dy < ∞ for all u ∈ Lp(X; R
n)
(17)
if and only if there exists for every M ∈ (0, ∞) a constant CM ∈ (0, ∞) such that
f (w, z) ≤ CM (1 + pp
M (w))(1 + pp
M (z))
for all w, z ∈ R
n.
(18)
Proof. If f fulfils the condition (18), then (17) holds by Corollary 5.
For the other direction, we assume that for some M ∈ (0, ∞), there does not exist a
k=1 ⊂
constant CM such that condition (18) holds. Then we find a sequence (cid:0)(wk, zk)(cid:1)∞
n with
n×R
R
f (wk, zk)
≥ 22k+2
for all k ∈ .
(1 + pp
M (wk))(1 + pp
M (zk))
We choose pairwise disjoint subsets Ek, Fk ⊂ X, k ∈ , with
L m(Ek) =
L m(X)
2k+1(1 + pp
M (wk))
and L m(Fk) =
L m(X)
2k+1(1 + pp
M (zk))
.
Such subsets exist since P∞
We now define the function
measure is nonatomic, see e.g. Corollary 1.21 in [5].
k=1(L m(Ek) + L m(Fk)) ≤ L m(X) and since the Lebesgue
u : X → R
n,
u(x) =
∞
Xk=1
(wkχEk (x) + zkχFk (x)).
8
peter elbau
Then u ∈ Lp(X; R
n), since we have for p ∈ [1, ∞)
∞
∞
L m(X)
= L m(X)
2k
Xk=1
zko ≤ M.
kukp
p =
and for p = ∞
Xk=1(cid:0)wkpL m(Ek) + zkpL m(Fk)(cid:1) ≤
Moreover, we find that
kuk∞ = maxn sup
k∈
wk, sup
k∈
ZXZX
f (u(x), u(y)) dx dy ≥
=
∞
∞
Xk=1
Xk=1
Thus, f does not fulfil condition (17).
f (wk, zk)L m(Ek)L m(Fk)
L m(X)2f (wk, zk)
M (wk))(1 + pp
22k+2(1 + pp
M (zk))
= ∞.
(cid:3)
4. Strong Sequential Lower Semi-Continuity
Before we analyse the sequential lower semi-continuity of non-local functionals with respect
to the weak or to the weak-star topology on Lp(X; R
n), we shortly give a criterion for the
sequential lower semi-continuity with respect to the norm topology. To start with, let us
briefly recall the definition of sequential lower semi-continuity.
Definition 8. Let V be a topological space. Then a functional J : V → R ∪ {∞} is called
sequentially lower semi-continuous if we have for every sequence (uk)k∈ ⊂ V converging to
u ∈ V that
Similar to the case of local functionals
lim inf
k→∞
J (uk) ≥ J (u).
Jlocal : Lp(X; R
n) → R ∪ {∞}, Jlocal(u) =ZX
flocal(x, u(x)) dx,
where the sequential lower semi-continuity of the functional Jlocal is equivalent to the lower
n → R, w 7→ flocal(x, w) for almost all x ∈ X, see e.g.
semi-continuity of the function R
n → R, (w, z) 7→
Theorem 5.9 in [5], the lower semi-continuity of the function R
f (x, y, w, z) is sufficient to guarantee the sequential lower semi-continuity of the non-local
functional J p
f provided the negative part of f is additionally p-bounded. In the local case,
this kind of lower bound was already necessary for the functional Jlocal to be well-defined.
n → R and every
To simplify the notation, we define for every function f : X×X×R
n × R
n×R
(x, y) ∈ X ×X the function f(x,y) : R
n×R
n → R by
f(x,y)(w, z) = f (x, y, w, z)
for all w, z ∈ R
n.
(19)
Proposition 9. Let f : X×X×R
whose negative part is p-bounded.
n×R
n → R be a pairwise symmetric, measurable function
Then the non-local functional J p
f on Lp(X; R
lower semi-continuous with respect to the strong topology on Lp(X; R
is for almost all (x, y) ∈ X ×X lower semi-continuous.
n) defined by the function f is sequentially
n) if the function f(x,y)
sequential lower semi-continuity of non-local functionals
9
Proof. Let (uk)k∈ ⊂ Lp(X; R
subsequence (ukℓ)ℓ∈ of (uk)k∈ such that
n) be a sequence converging to u ∈ Lp(X; R
n). We choose a
holds and such that we have
lim
ℓ→∞
J p
f (ukℓ ) = lim inf
k→∞
J p
f (uk)
lim
ℓ→∞
ukℓ(x) = u(x)
for almost all x ∈ X.
To be able to apply Fatou's lemma, we use that the negative part of f is p-bounded. We
thus find for every M ∈ (0, ∞) a constant C ∈ (0, ∞) and positive functions αM ∈ L1(X×X)
and βM ∈ L1(X) such that
f (x, y, w, z) + αM (x, y) + βM (x) pp
M (z) + βM (y) pp
M (w) + C pp
M (w) pp
M (z) ≥ 0
n. We then choose the constant M in the
for almost all (x, y) ∈ X × X and all w, z ∈ R
case p = ∞ greater than supℓ∈ kukℓk∞ and find with the lower semi-continuity of f(x,y) for
almost all (x, y) ∈ X ×X that
lim
ℓ→∞ZXZX
≥ZXZX
≥ZXZX
f (x, y, ukℓ(x), ukℓ (y)) dx dy + kαM k1 + 2kβM k1 pp
M (kukp) + C pp
M (kukp)2
lim inf
ℓ→∞ (cid:0)f (x, y, ukℓ(x), ukℓ (y)) + αM (x, y) + βM (x) pp
M (ukℓ(x)) + C pp
M (ukℓ(x)) pp
+ βM (y) pp
M (ukℓ(y))
f (x, y, u(x), u(y)) dx dy + kαM k1 + 2kβM k1 pp
M (ukℓ(y))(cid:1) dx dy
M (kukp) + C pp
M (kukp)2.
Thus, lim inf k→∞ J p
continuous with respect to the strong topology on Lp(X; R
f (u), and we conclude that J p
n).
f (uk) ≥ J p
f is sequentially lower semi-
(cid:3)
We remark that without the lower bound on the function f , the lower semi-continuity
of the function f(x,y) for almost all (x, y) ∈ X × X does not imply the sequential lower
semi-continuity of the non-local functional J p
f .
Example 10. Similar to Example 6, we choose X = (0, 1) and n = 1 and define the map
f : X ×R → R,
f (x, z) =(− 1
z
0
z ∈ [x, 1],
if
otherwise.
Then the function
f : X ×X ×R×R → R,
f (x, y, w, z) = 1
2 ( f (x, z) + f (y, w))
has a negative part which is not p-bounded as was shown in Example 6, but f(x,y) is for all
x, y ∈ X lower semi-continuous.
On the other hand, we find from (16) for the non-local functional J p
f defined by the
function f that
J p
f (u) = −L 1(cid:0)u−1((0, 1])(cid:1).
So, for the sequence (uk)k∈ ⊂ Lp(X; R
get that (uk)k∈ converges uniformly to the zero function, but
n) defined by uk(x) = 1
k for all x ∈ X, k ∈ , we
lim inf
k→∞
J p
f (uk) = −1 < 0 = J p
f (0).
Thus, J p
Lp(X; R
f is not sequentially lower semi-continuous with respect to the strong topology on
n).
10
peter elbau
5. Weak Sequential Lower Semi-Continuity
After all the preparations, we are now ready to study the sequential lower semi-continuity
n) for p ∈ [1, ∞) and
of non-local functionals with respect to the weak topology on Lp(X; R
with respect to the weak-star topology on L∞(X; R
n) for p = ∞.
n → R be a pairwise symmetric, measurable function
Theorem 11. Let f : X ×X ×R
whose negative part is p-bounded. Moreover, we assume that the function f(x,y), defined
by (19), is continuous for almost all (x, y) ∈ X×X and that there exist for every M ∈ (0, ∞)
positive functions αM ∈ L1(X ×X) and βM ∈ L1(X) such that
n×R
f (x, y, w, z) ≤ αM (x, y) + βM (x) pp
M (z)
(20)
for almost all (x, y) ∈ X ×X and all w, z ∈ R
n with w ≤ M .
Then the non-local functional J p
f on Lp(X; R
lower semi-continuous with respect to the weak topology on Lp(X; R
with respect to the weak-star topology on L∞(X; R
n) defined by the function f is sequentially
n) for p ∈ [1, ∞) and
n) for p = ∞ if and only if the function
Φx,ψ : R
n → R, Φx,ψ(w) =ZX
f (x, y, w, ψ(y)) dy
(21)
is for every ψ ∈ Lp(X; R
n) for almost all x ∈ X convex.
Proof. We first show that the function Φx,ψ is for every ψ ∈ Lp(X; R
convex if J p
f is sequentially lower semi-continuous.
n) for almost all x ∈ X
So, let ψ ∈ Lp(X; R
n). To begin with, we assume that J p
f (ψ) < ∞. Since the Lebesgue
measure is nonatomic, we find for every ϑ ∈ [0, 1] a sequence (Eϑ,k)k∈ of subsets of X such
that the characteristic functions χEϑ,k ∈ L∞(X), k ∈ , converge weakly-star in L∞(X)
to the constant function ϑ, see e.g. Proposition 2.87 in [5]. Then we define for arbitrary
functions ω1, ω2 ∈ L∞(X; R
n), ϑ ∈ [0, 1], and an arbitrary measurable subset A ⊂ X the
functions uk ∈ Lp(X; R
n),
uk(x) = χA(x)ϕk(x) + χAc (x)ψ(x),
k ∈ ,
where we use the notation Ac = X \A and where the functions ϕk ∈ L∞(X; R
by
n) are given
ϕk(x) = χEϑ,k(x) ω1(x) + (1 − χEϑ,k(x)) ω2(x),
k ∈ .
The sequence (uk)k∈ thus converges weakly in Lp(X; R
n) if p ∈ [1, ∞) and weakly-star
in L∞(X; R
n) if p = ∞ to the function u ∈ Lp(X; R
n) defined by
u(x) = χA(x)¯ω(x) + χAc (x)ψ(x)
where ¯ω ∈ L∞(X; R
the sequential lower semi-continuity of J p
therefore get for all measurable sets A ⊂ X the inequality
n) is the convex combination ¯ω = ϑω1 + (1 − ϑ)ω2 of ω1 and ω2. So,
f (u). We
f implies that lim inf k→∞ J p
f (uk) ≥ J p
lim inf
k→∞ (cid:18)ZAZA
f (x, y, ϕk(x), ϕk(y)) dy dx + 2ZAZAc
≥ZAZA
f (x, y, ¯ω(x), ¯ω(y)) dy dx + 2ZAZAc
f (x, y, ϕk(x), ψ(y)) dy dx(cid:19)
f (x, y, ¯ω(x), ψ(y)) dy dx.
(22)
sequential lower semi-continuity of non-local functionals
11
To get rid of the integrals over A×A, we will now consider the limit where the measure
of A tends to zero. We remark that because of the p-boundedness of f − and the upper
bound (20) of f , there exists for every M ∈ (0, ∞) a function gM ∈ L1(X×X) such that we
have
f (x, y, ω(x), ω(y)) ≤ gM (x, y) and f (x, y, ω(x), ψ(y)) ≤ gM (x, y)
for all ω, ω ∈ L∞(X; R
particular, we have with M ≥ max{kω1k∞, kω2k∞} that
n) with kωk∞ ≤ M , kωk∞ ≤ M and almost all (x, y) ∈ X ×X. In
f (x, y, ϕk(x), ϕk(y)) ≤ max
i,j∈{1,2}
f (x, y, ωi(x), ωj(y)) ≤ gM (x, y),
f (x, y, ¯ω(x), ¯ω(y)) ≤ gM (x, y)
for almost all (x, y) ∈ X ×X and all k ∈ . To get a bound for the integrals over A×A, we
choose for every ε ∈ (0, ∞) and every measurable set E ⊂ X with positive measure a set
E′
M ⊂ E with positive measure such that
ZAZA
gM (x, y) dx dy ≤ εL m(A).
for all measurable sets A ⊂ E′
M . We then get from inequality (22) that
lim inf
k→∞ ZAZAc
f (x, y, ϕk(x), ψ(y)) dy dx + εL m(A) ≥ZAZAc
f (x, y, ¯ω(x), ψ(y)) dy dx (23)
for all measurable sets A ⊂ E′
M .
By definition of the functions ϕk, we find for all measurable sets A ⊂ E′
M and all k ∈
that
ZAZAc
f (x, y, ϕk(x), ψ(y)) dy dx =ZA
f (x, y, ω1(x), ψ(y)) dy dx
χEϑ,k(x)ZAc
+ZA
(1 − χEϑ,k (x))ZAc
f (x, y, ω2(x), ψ(y)) dy dx.
Since the functions χEϑ,k converge for k → ∞ by definition of the sets Eϑ,k weakly-star in
L∞(X) to the constant function ϑ, we further get that
lim inf
k→∞ ZAZAc
f (x, y, ϕk(x), ψ(y)) dy dx = ϑZAZAc
(f (x, y, ω1(x), ψ(y)) dy dx
+ (1 − ϑ)ZAZAc
f (x, y, ω2(x), ψ(y)) dy dx (24)
for all measurable sets A ⊂ E′
M .
By our choice of the function gM and the set E′
M we have that
ZAZA
f (x, y, ω(x), ψ(y)) dy dx ≤ εL m(A)
for all measurable sets A ⊂ E′
Therefore, we get by plugging (24) into (23) the inequality
M and all functions ω ∈ L∞(X; R
n) with kωk∞ ≤ M .
ZA(cid:0)ϑΦx,ψ(ω1(x)) + (1 − ϑ)Φx,ψ(ω2(x))(cid:1) dx + 3εL m(A) ≥ZA
Φx,ψ(¯ω(x)) dx
12
peter elbau
for all measurable sets A ⊂ E′
ω1, ω2 ∈ L∞(X; R
get after letting ε tend to zero that
M . Since this holds for every M ∈ (0, ∞) for all functions
n) with kω1k∞ ≤ M and kω2k∞ ≤ M , we can use Lemma 21 and finally
ϑΦx,ψ(w1) + (1 − ϑ)Φx,ψ(w2) ≥ Φx,ψ(ϑw1 + (1 − ϑ)w2)
for almost all x ∈ X, all ϑ ∈ [0, 1], and all w1, w2 ∈ R
for almost all x ∈ X.
n, which proves the convexity of Φx,ψ
It remains to consider the case where J p
of f − and the upper bound (20) of f ensure that J p
choose a sequence (ψk)k∈ ⊂ L∞(X; R
strongly in Lp(X; R
functions Φx,ψk, k ∈ , are for almost all x ∈ X convex. Moreover, the local functional
f (ψ) = ∞. Here we use that the p-boundedness
n). We
n) which converges pointwise almost everywhere and
n) to the function ψ. Then, by the previous result, we know that the
f (ω) < ∞ for all ω ∈ L∞(X; R
Lp(X; R
n) → R ∪ {∞},
v 7→ Φx,v(w) =ZX
f (x, y, w, v(y)) dy
is -- because of the continuity of the function f(x,y) for almost all (x, y) ∈ X × X, the p-
boundedness of f −, and the upper bound (20) of f -- for almost all x ∈ X and all w ∈ R
n
continuous with respect to the strong topology, see e.g. Corollaries 6.51 and 6.53 in [5] (the
proof works in the same way as the proof of Proposition 9). Therefore, we have for almost
all x ∈ X that
Φx,ψ(w) = lim
k→∞
Φx,ψk (w),
which shows that the function Φx,ψ is for almost all x ∈ X convex.
This concludes the proof that the sequential lower semi-continuity of J p
function ψ ∈ Lp(X; R
n) for almost all x ∈ X the convexity of the function Φx,ψ.
f implies for every
For the other direction, we assume that the function Φx,ψ is for every ψ ∈ Lp(X; R
n) if p ∈ [1, ∞) and weakly-star in L∞(X; R
for almost all x ∈ X convex. Let (uk)k∈ ⊂ Lp(X; R
Lp(X; R
In particular, (uk)k∈ is bounded in Lp(X; R
(ukℓ)ℓ∈ of (uk)k∈ generating a Young measure ν.
n)
n) be a sequence converging weakly in
n).
n) and therefore, there exists a subsequence
n) if p = ∞ to a function u ∈ Lp(X; R
I.e. we have a map ν : X → M(R
Radon measures on R
and that for all φ ∈ C0(R
satisfies for every h ∈ L1(X) the equality
n; R) denotes all signed
n, which fulfils that νx is a probability measure for almost all x ∈ X,
n φ(w) dνx(w) is measurable and
n; R), x 7→ νx, where M(R
n) the function X → R, x 7→RR
h(x)ZR
h(x)φ(ukℓ (x)) dx =ZX
n
lim
ℓ→∞ZX
φ(w) dνx(w) dx.
(25)
For a detailed introduction into the theory of Young measures, we refer to Chapter 8 in [5].
Since (ukℓ)ℓ∈ generates the Young measure ν, the sequence (ukℓ ×ukℓ)ℓ∈ in the space
n) defined by (ukℓ × ukℓ)(x, y) = (ukℓ(x), ukℓ (y)) generates the Young
Lp(X × X; R
n; R) defined by (ν ⊗ ν)(x,y) = νx ⊗ νy, where νx ⊗ νy
measure ν ⊗ ν : X ×X → M(R
denotes the product measure of νx and νy. Indeed, using the Stone -- Weierstrass theorem, it
is enough to verify that
n × R
n ×R
lim
ℓ→∞ZXZX
h(x, y)φ(ukℓ (x), ukℓ (y)) dx dy
=ZXZX
h(x, y)ZR
nZR
n
φ(w, z) dνx(w) dνy(z) dx dy
sequential lower semi-continuity of non-local functionals
13
holds for all functions h ∈ L1(X ×X), φ ∈ C0(R
and φ(w, z) = φ1(w)φ2(z), h1, h2 ∈ L1(X), φ1, φ2 ∈ C0(R
Fubini's theorem and the relation (25), see Proposition 2.3 in [9].
n × R
n) of the form h(x, y) = h1(x)h2(y)
n). But this directly follows from
Now, the p-boundedness of f − implies by the Dunford -- Pettis theorem that the functions
X ×X → [0, ∞),
(x, y) 7→ f −(x, y, ukℓ(x), ukℓ (y)),
ℓ ∈ ,
are uniformly integrable. Thus, we can apply the fundamental theorem for Young measures,
see e.g. Theorem 8.6 in [5], and find that
lim inf
ℓ→∞
J p
f (ukℓ) ≥ZXZXZR
nZR
f (x, y, w, z) dνy(z) dνx(w) dy dx.
(26)
n
Moreover, condition (20) implies that for almost all x ∈ X and all w ∈ R
n the functions
X → R,
y 7→ f (x, y, w, ukℓ(y)),
ℓ ∈ ,
are uniformly integrable. Thus, we get from the continuity of the functions f(x,y) for almost
all (x, y) ∈ X ×X again with the fundamental theorem for Young measures that
lim
ℓ→∞ZX
f (x, y, w, ukℓ(y)) dy =ZXZR
f (x, y, w, z) dνy(z) dy
n
for almost all x ∈ X and all w ∈ R
n. So for almost all x ∈ X, the function
Φx,ν : R
n → R ∪ {∞},
Φx,ν(w) =ZXZR
f (x, y, w, z) dνy(z) dy
(27)
n
is the limit of the convex functions Φx,ukℓ
, ℓ ∈ , and is therefore convex.
Using now that νx is by definition of a Young measure for almost all x ∈ X a probability
measure, we find with Jensen's inequality that
ZXZXZR
nZR
n
f (x, y, w, z) dνy(z) dνx(w) dy dx =ZXZR
≥ZX
Φx,ν(w) dνx(w) dx
n
Φx,ν(cid:0)RR
n w dνx(w)(cid:1) dx.
(28)
Since the sequence (ukℓ)ℓ∈ converges weakly in Lp(X; R
n) if p ∈ [1, ∞) and weakly-star
in L∞(X; R
n) if p = ∞ to u, relation (25) implies that
w dνx(w) = u(x)
for almost all x ∈ X.
ZR
n
Together with the pairwise symmetry of f , we then get
ZX
Φx,ν(RR
n w dνx(w)) dx =ZX
Φx,ν(u(x)) dx =ZXZR
Φy,u(z) dνy(z) dy.
(29)
n
Using now the convexity of Φy,u for almost all y ∈ X, we get again with Jensen's
inequality that
ZXZR
n
Φy,u(z) dνy(z) dy ≥ZX
Φy,u(cid:0)RR
n z dνy(z)(cid:1) dy =ZX
Φy,u(u(y)) dy = J p
f (u).
(30)
Putting together the inequalities (26), (28), (29), and (30), we have shown that
proving the sequential lower semi-continuity of J p
f .
(cid:3)
lim inf
ℓ→∞
J p
f (ukℓ) ≥ J p
f (u),
14
peter elbau
Here, the proof that the convexity of the functions Φx,ψ implies the sequential lower
semi-continuity of the functional J p
f makes only use of the upper bound (20) to show the
convexity of the function Φx,ν defined in (27). We can therefore waive this upper bound if
we guarantee the convexity of Φx,ν by imposing that the function f(x,y) is separately convex
n → R, w 7→ f(x,y)(z, w) are convex for all
(i.e. the maps R
z ∈ R
n) for almost all (x, y) ∈ X ×X, see [9].
n → R, w 7→ f(x,y)(w, z) and R
Corollary 12. Let f : X ×X ×R
whose negative part is p-bounded.
n×R
n → R be a pairwise symmetric, measurable function
Then the non-local functional J p
f on Lp(X; R
lower semi-continuous with respect to the weak topology on Lp(X; R
with respect to the weak-star topology on L∞(X; R
by (19), is for almost all (x, y) ∈ X ×X separately convex.
n) defined by the function f is sequentially
n) for p ∈ [1, ∞) and
n) for p = ∞ if the function f(x,y), defined
6. Equivalent Integrands
In this section, we will try to characterise the classes of functions which define the same
non-local functional. In particular, we are interested in finding a good representative for
each of these classes and thereby to possibly simplify the criterion of sequential lower semi-
continuity given in Theorem 11.
We will restrict our attention to rather regular integrands.
n×R
Definition 13. Let f : X×X×R
Moreover, we assume that the function
n → R be a pairwise symmetric, measurable function.
X ×X → R,
(x, y) 7→ f (x, y, 0, 0)
(31)
is integrable, that the function f(x,y), defined by (19), is continuously differentiable for
almost all (x, y) ∈ X × X, and that there exist for every M ∈ (0, ∞) positive functions
αM ∈ Lp∗
(X) ⊗ L1(X) and βM ∈ L1(X), with p∗ being the Holder conjugate of p, such that
∇wf (x, y, w, z) ≤ αM (x, y) + βM (y) pp−1
M (w)
(32)
for almost all (x, y) ∈ X ×X and all w, z ∈ R
function.
n with z ≤ M . Then we call f a p-regular
We remark that for a p-regular function f the estimate
f (x, y, w, z) ≤ f (x, y, 0, 0) +Z 1
0
w∇wf (x, y, tw, 0) dt +Z 1
0
z∇zf (x, y, w, tz) dt
implies together with the integrability of the function (31) and the bound (32) for the
derivative of f that there exist for every M ∈ (0, ∞) positive functions αM ∈ L1(X × X)
and βM ∈ L1(X) such that
f (x, y, w, z) ≤ αM (x, y) + βM (x) pp
M (z)
for almost all (x, y) ∈ X ×X and all w, z ∈ R
n with w ≤ M .
We will in the following give a characterisation of the class of functions whose corre-
sponding non-local functional constantly vanishes. If we only consider real-valued non-local
functionals, then two functions define the same non-local functional if and only if they differ
by a function of this class.
sequential lower semi-continuity of non-local functionals
15
Definition 14. We denote by N p the set of all pairwise symmetric, measurable functions
n → R whose negative part f − obeys condition (3) and for which the
f : X × X × R
non-local functional J p
n).
n) defined by f fulfils J p
f (u) = 0 for all u ∈ Lp(X; R
n × R
f on Lp(X; R
We further introduce a subset N p
0 of N p which is easier to parametrise.
Definition 15. Let N p
0 denote the set of all pairwise symmetric, measurable functions
n → R and
f : X×X×R
a symmetric function h ∈ L1(X ×X) with the properties that we find for every M ∈ (0, ∞)
a function αM ∈ L1(X) and a constant C ∈ (0, ∞) with
n → R for which there exist a measurable function g : X×X×R
n×R
g(x, y, w) dy ≤ αM (x) + C pp
M (w)
ZX
for almost all x ∈ X and all w ∈ R
n,
ZXZX
h(x, y) dx dy = 0,
g(x, y, w) dy = 0
ZX
for almost all x ∈ X and all w ∈ R
n, and
f (x, y, w, z) = g(x, y, w) + g(y, x, z) + h(x, y)
for almost all (x, y) ∈ X ×X and all w, z ∈ R
n.
Lemma 16. We have N p
0 ⊂ N p.
(33)
(34)
(35)
0 . Then there exist a measurable function g : X × X × R
Proof. Let f ∈ N p
n → R and a
symmetric function h ∈ L1(X × X) with the properties (33), (34), and (35). Because of
the integrability condition (33), we can use Fubini's theorem and find with (34) that the
non-local functional J p
f defined by the function f fulfils
J p
f (u) =ZXZX
g(x, y, u(x)) dy dx +ZXZX
g(y, x, u(y)) dx dy = 0
for every u ∈ Lp(X; R
n). Thus, f ∈ N p.
(cid:3)
If we restrict our attention to p-regular functions, then there is no difference between N p
0
and N p.
Proposition 17. Let f : X ×X × R
and only if f ∈ N p
0 .
n × R
n → R be a p-regular function. Then f ∈ N p if
Proof. Let f ∈ N p. Then the non-local functional J p
zero. So, we can take the variational derivative of J p
of the function f that
f defined by f is constantly equal to
f and get with the pairwise symmetry
0 = lim
t→0
J p
f (u + tv) − J p
f (u)
t
= 2
n
Xi=1ZXZX
∂wi f (x, y, u(x), u(y))vi(x) dx dy
(36)
for all functions u, v ∈ L∞(X; R
derivative of f to differentiate under the integral sign.
n). Here, we have used the bound (32) of the first partial
16
peter elbau
Thus, we have for all i ∈ [1, n] ∩ , almost all x ∈ X, all w ∈ R
n, and all functions
u ∈ L∞(X; R
n) that
∂wi f (x, y, w, u(y)) dy = 0.
(37)
ZX
Indeed, there would otherwise exist an i0 ∈ [1, n] ∩ , a set A ⊂ X with positive measure,
a function ϕ ∈ L∞(A; R
n), a constant δ ∈ (0, ∞), and a sign
ǫ ∈ {−1, 1} such that
n), a function u ∈ L∞(X; R
ǫZX
∂wi0
f (x, y, ϕ(x), u(y)) dy ≥ δ
for all x ∈ A.
Using the bound (32) of the function ∂wi0
f , we find a measurable subset A ⊂ A such that
Z AZ A
∂wi0
f (x, y, ϕ(x), ψ(y)) dy dx <
L m( A)
δ
2
for all functions ψ ∈ L∞(X; R
functions u, v ∈ L∞(X; R
n) by
n) with kψk∞ ≤ max{kuk∞, kϕk∞}. Defining now the
u(x) =(u(x)
ϕ(x)
if x ∈ X \ A,
if x ∈ A
and vi(x) = δi,i0 χ A(x), we get
ǫ
n
Xi=1ZXZX
∂wi f (x, y, u(x), u(y))vi(x) dx dy = ǫZ AZX
∂wi0
f (x, y, ϕ(x), u(y)) dy dx
> ǫZ AZX
∂wi0
f (x, y, ϕ(x), u(y)) dy dx − δL m( A) ≥ 0,
which is a contradiction to (36).
Now, equation (37) is only possible if the function R
constant for all i ∈ [1, n] ∩ , almost all (x, y) ∈ X ×X, and all w ∈ R
the function g : X ×X ×R
n → R by
n → R, z 7→ ∂wi f (x, y, w, z) is
n. Defining therefore
g(x, y, w) =
n
Xi=1Z 1
0
wi∂wi f (x, y, tw, 0) dt,
we find with the fundamental theorem of calculus that g(x, y, w) = f (x, y, w, z)− f (x, y, 0, z)
n. Thus, we get with the pairwise
holds for almost all (x, y) ∈ X × X and all w, z ∈ R
symmetry of f that
f (x, y, w, z) = g(x, y, w) + g(y, x, z) + f (x, y, 0, 0)
n. So, f has the form (35), where the function
for almost all (x, y) ∈ X×X and all w, z ∈ R
h ∈ L1(X×X) is defined by h(x, y) = f (x, y, 0, 0) for all x, y ∈ X. Moreover, the p-regularity
of f implies the bound (33) for g, and the conditions (34) finally follow from J p
f (0) = 0
and from (37) together with Fubini's theorem. Thus, we have shown that f ∈ N p
0 which
concludes the proof.
(cid:3)
sequential lower semi-continuity of non-local functionals
17
In the following, we will try to use this ambiguity in the integrand of a non-local functional
to find for a non-local functional J p
f , which is sequentially lower semi-continuous with respect
to the weak topology if p ∈ [1, ∞) and to weak-star topology if p = ∞, an integrand f
f such that f(x,y) is separately convex. But it seems that this is
defining the functional J p
only possible in the case where the functional is defined on a Lebesgue space of real-valued
functions, i.e. for n = 1.
Theorem 18. Let n = 1 and let f : X × X × R × R → R be a p-regular function which
additionally fulfils that the function f(x,y), defined by (19), is for almost all (x, y) ∈ X ×X
two times continuously differentiable and that there exists for every M ∈ (0, ∞) a function
αM ∈ L1(X ×X) such that
∂2
wf (x, y, w, z) ≤ αM (x, y)
(38)
for almost all (x, y) ∈ X ×X and all w, z ∈ R with w ≤ M and z ≤ M .
Then the function
Φx,ψ : R → R, Φx,ψ(w) =ZX
f (x, y, w, ψ(y)) dy
is for every function ψ ∈ Lp(X) for almost all x ∈ X convex if and only if there exist a
pairwise symmetric, measurable function f : X ×X × R× R → R and a function f0 ∈ N p
0
such that f(x,y) is for almost all (x, y) ∈ X ×X separately convex and
f (x, y, w, z) = f (x, y, w, z) + f0(x, y, w, z)
(39)
for almost all (x, y) ∈ X ×X and all w, z ∈ R.
Proof. Let us first assume that the function f is of the form (39). Then there exist by
definition of the set N p
0 a measurable function g : X×X×R → R and a symmetric function
h ∈ L1(X ×X) with the properties (33) and (34) such that
f (x, y, w, z) = f (x, y, w, z) + g(x, y, w) + g(y, x, z) + h(x, y)
for almost all (x, y) ∈ X × X and all w, z ∈ R. Thus, the function Φx,ψ fulfils for every
ψ ∈ Lp(X) for almost all x ∈ X that
Φx,ψ(w) =ZX
f (x, y, w, ψ(y)) dy +ZX
g(y, x, ψ(y)) dy +ZX
h(x, y) dy
for all w ∈ R and is therefore convex because of the separate convexity of the function f(x,y)
for almost all (x, y) ∈ X ×X.
Let us on the other hand assume that the function Φx,ψ is for every ψ ∈ Lp(X) for
almost all x ∈ X convex. Since the function f(x,y) is for almost all (x, y) ∈ X×X two times
differentiable and we have the bounds (32) and (38) for its partial derivatives, we know that
Φx,ψ is for every ψ ∈ L∞(X) for almost all x ∈ X two times differentiable and that we
can differentiate under the integral sign. The convexity of Φx,ψ therefore implies for every
ψ ∈ L∞(X) that
d2Φx,ψ
dw2
(w) =ZX
for almost all x ∈ X and all w ∈ R.
∂2
wf (x, y, w, ψ(y)) dy ≥ 0
(40)
18
peter elbau
We now define for every M ∈ (0, ∞) the measurable function
γM : X ×X ×R → R,
γM (x, y, w) = min
z∈[−M,M]
∂2
wf (x, y, w, z).
Then condition (40) implies that
γM (x, y, w) dy ≥ 0
(41)
ZX
for almost all x ∈ X, all w ∈ R, and all M ∈ (0, ∞). To prove this, we first pick for arbitrary
M ∈ (0, ∞) and ε ∈ (0, ∞) a measurable set D ⊂ X (with arbitrarily small measure) and a
λ ∈ (0, ∞) such that the function αM defined by (38) fulfils
αM (x, y) dy <
ε
3
ZA
for every measurable set A ⊂ X with L m(A) < λ and every x ∈ X \ D. By Scorza --
Dragoni's theorem, see e.g. Theorem 6.35 in [5], we can further choose a compact set E ⊂
X × X with L 2m(X × X \ E) < 1
2 λ2 such that the restricted functions γM E×[−M,M] and
∂2
wf E×[−M,M]×[−M,M] are continuous. In particular, we find a constant δ ∈ (0, ∞) such
that
∂2
wf (x, y, w, z) − ∂2
wf (x, y, w, z) <
ε
3L m(X)
for all x, x, y ∈ X and w, z ∈ [−M, M ] with (x, y) ∈ E, (x, y) ∈ E, and x− x < δ. Moreover,
we define the set
F = {x ∈ X : L m(X \Ex) ≥ λ
2 },
where Ex = {y ∈ X : (x, y) ∈ E}, x ∈ X, and remark that we have the estimate L m(F ) <
2
L 2m(X ×X \E) < λ. By Aumann's measurable selection theorem, see e.g. Theorem 6.10
λ
in [5], we finally find for every x ∈ X and every w ∈ [−M, M ] a measurable function
ψx,w : X → [−M, M ] with ∂2
wf (x, y, w, ψx,w(y)) = γM (x, y, w) for almost all y ∈ X.
Putting all this together, we get for every x ∈ X \ (D ∪ F ) and every x ∈ X \ F with
x − x < δ that
ZX
γM (x, y, w) dy ≥ZEx∩E x
∂2
wf (x, y, w, ψx,w(y)) dy −ZX\Ex∪X\E x
αM (x, y) dy
≥ZEx∩E x(cid:18)∂2
wf (x, y, w, ψx,w(y)) −
ε
3L m(X)(cid:19) dy −
ε
3
.
By Lebesgue's density theorem, we have for almost every point x ∈ X\(D ∪ F ) that the set
of points x ∈ X \ F with x − x < δ has positive measure. Because of condition (40), we
therefore find for almost every x ∈ X \(D ∪ F ) and every w ∈ [−M, M ] a point x ∈ X \F
with distance x − x < δ such that RX ∂2
wf (x, y, w, ψx,w(y)) dy ≥ 0 and thus
γM (x, y, w) dy ≥ −ε.
ZX
Letting now λ, ε, and the measure of the set D tend to zero, we get (41).
Since the map (0, ∞) → R, M 7→ γM (x, y, w) is for all x, y ∈ X and w ∈ R monotonically
decreasing, we may define γ(x, y, w) = limM→∞ γM (x, y, w). From the Lebesgue monotone
convergence theorem, we further get that
ZX
γ(x, y, w) dy = lim
M→∞ZX
γM (x, y, w) dy ≥ 0.
sequential lower semi-continuity of non-local functionals
19
With condition (38), this in particular implies for every M ∈ (0, ∞) that
ZX
γ(x, y, w) dy ≤ 2ZX
γ+(x, y, w) dy ≤ 2ZX
∂2
wf (x, y, w, 0) dy ≤ 2ZX
αM (x, y) dy
for almost all x ∈ X and all w ∈ R with w ≤ M , where γ+ denotes the positive part of
the function γ. Therefore, defining the function g : X ×X ×R → R by
g(x, y, w) =Z w
0 Z w
0 (cid:18)γ(x, y, w) −
1
L m(X)ZX
γ(x, y, w) dy(cid:19) d w d w,
Fubini's theorem implies for almost all x ∈ X and all w ∈ R that y 7→ g(x, y, w) is integrable
and fulfils
g(x, y, w) dy = 0.
(42)
Moreover, we have by construction
ZX
∂2
wg(x, y, w) ≤ γ(x, y, w) ≤ ∂2
wf (x, y, w, z)
for almost all x, y ∈ X, almost all w ∈ R, and all z ∈ R. Therefore, the function f :
X ×X ×R×R → R defined by
f (x, y, w, z) = f (x, y, w, z) − g(x, y, w) − g(y, x, z)
fulfils that f(x,y) is for almost all (x, y) ∈ X ×X separately convex.
Thus, it only remains to prove that the function g satisfies a condition of the form (33).
We find with the property (42) of g that
ZX
g(x, y, w) dy = 2ZX
for almost all x ∈ X and all w ∈ R. Using now the p-regularity of f , we find for every
M ∈ (0, ∞) positive functions αM ∈ Lp∗
(X) ⊗ L1(X) and β ∈ L1(X) such that
dy
0
∂2
g+(x, y, w) dy ≤ 2ZX(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
0 Z w
Z w
wf (x, y, w, 0) d w d w(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂wf (x, y, w, 0) − ∂wf (x, y, 0, 0) d w(cid:12)(cid:12)(cid:12)(cid:12)
(αM (x, y) + β(y) pp−1
dy
dy
M ( w)) d w(cid:12)(cid:12)(cid:12)(cid:12)
αM (x, y) dy + p∞
M (w)
ZX
0
Z w
g(x, y, w) dy ≤ 2ZX(cid:12)(cid:12)(cid:12)(cid:12)
Z w
≤ZX(cid:12)(cid:12)(cid:12)(cid:12)
g(x, y, w) dy ≤ MZX
ZX
0
for almost all x ∈ X and all w ∈ R. If p = ∞, we then immediately find
for almost all x ∈ X and all w ∈ R, and if p ∈ [1, ∞), we apply Youngs inequality to get
g(x, y, w) dy ≤
ZX
1
p∗ (cid:18)ZX
αM (x, y) dy(cid:19)p∗
+
1
p(cid:18)1 +ZX
β(y) dy(cid:19) pp
M (w)
for almost all x ∈ X and all w ∈ R.
(cid:3)
However, for functionals of vector-valued functions, this argumentation fails. We give a
counterexample to illustrate the problematic. For simplicity, we waive the symmetry of the
function f and consider only the case n = 2.
20
peter elbau
Example 19. Let X = [−1, 1], n = 2, and p ≥ 2. We choose a non-negative, convex function
a ∈ C2(R) with a(ζ) = ζ − 1 for all ζ ∈ R\(−2, 2) and define the function b ∈ C2(R) by
b(ζ) =(1 + ζ + 1
2 ζ2
(1 − ζ + 1
2 ζ2)−1
if ζ ≥ 0,
if ζ < 0.
Moreover, we define the function f : X ×X ×R
2×R
2 → R by
f (x, y, w, z) =( 1
2 (b(z1)w2
1
2 a(z1)(w2
1 + b(−z1)w2
2)
1 + w2
2) + z1w1w2
if
if
y ≥ 0,
y < 0.
Then we find for the Hessian matrix HΦx,ψ of the function Φx,ψ defined in (21) for every
x ∈ X and ψ ∈ Lp(X; R
2) the expression
Hwf (x, y, w, ψ(y)) dy
HΦx,ψ(w) =Z 1
=Z 0
−1(cid:18)a(ψ1(y))
ψ1(y)
−1
ψ1(y)
a(ψ1(y))(cid:19) dy +Z 1
0 (cid:18)b(ψ1(y))
0
0
b(−ψ1(y))(cid:19) dy.
Since a(ζ) ≥ ζ − 1 for all ζ ∈ R, we have for all ζ, ζ ∈ R that
det(cid:18)a(ζ) + b(ζ)
ζ
ζ
a(ζ) + b(−ζ)(cid:19) = a(ζ)2 + (b(ζ) + b(−ζ))a(ζ) + 1 − ζ2
≥ (a(ζ) + 1)2 − ζ2 ≥ 0.
Therefore, we find for every x ∈ X and every ψ ∈ Lp(X; R
2) that
det HΦx,ψ(w) ≥ 0
for all w ∈ R
2
and thus that the function Φx,ψ is convex.
We now want to show that there does not exist a measurable function g : X×X×R
2 → R
2 → R, w 7→ g(x, y, w) is for almost all (x, y) ∈ X×X twice continuously
such that g(x,y) : R
differentiable,
g(x, y, w) dy = 0
for almost all x ∈ X and all w ∈ R
2,
(43)
ZX
and the map f(x,y,z) : R
(x, y) ∈ X ×X and all z ∈ R
almost all (x, y) ∈ X ×[−1, 0) and all z ∈ R
2 → R, f(x,y,z)(w) = f (x, y, w, z) + g(x, y, w) is for almost all
2 convex. So, assume there exists such a function g. Then for
2, the convexity of f(x,y,z) implies that
det H f(x,y,z)(w) = det(cid:18) a(z1) + ∂2
g(x,y)(w)
z1 + ∂w1 ∂w2 g(x,y)(w)
w1
z1 + ∂w1 ∂w2 g(x,y)(w)
a(z1) + ∂2
w2 g(x,y)(w)(cid:19) ≥ 0
for all w ∈ R
have for almost all (x, y) ∈ X ×[−1, 0) and all w ∈ R
2. Using that a(ζ) = ζ − 1 for ζ ≥ 2, we find that this is only possible if we
2 that
∂2
w1
g(x,y)(w) + ∂2
w2
g(x,y)(w) ≥ 2.
sequential lower semi-continuity of non-local functionals
21
Because of the condition (43), this implies that there exists for every w ∈ R
and a set A ⊂ X×[0, 1] with positive measure such that ∂2
wi
But then for every (x, y) ∈ A and every z ∈ R
2 an i ∈ {1, 2}
g(x,y)(w) ≤ −1 for all (x, y) ∈ A.
2 with (−1)iz1 > 0, the Hessian matrix
H f(x,y,z)(w) =(cid:18)b(z1) + ∂2
∂w1 ∂w2 g(x,y)(w)
w1 g(x,y)(w)
∂w1 ∂w2 g(x,y)(w)
w2 g(x,y)(w)(cid:19)
b(−z1) + ∂2
is not positive semidefinite.
A. Some Technicalities
We give here two missing technicial details to the proofs in the previous sections.
We begin with the statement that we can cover almost one forth of every set E ⊂ X×X
by a set of the form A×Ac with some measurable set A ⊂ X. To be more precise, let us
introduce for every N ∈ the notation
QN
a (x) =
N
Yj=1
(xj − a
2 , xj + a
2 )
for the cube in R
further define for every δ ∈ (0, ∞) the checkerboard pattern Sδ ⊂ R
N with side length a ∈ (0, ∞) and center x ∈ R
m by
N . In the space R
Sδ =
{x∈Z
[
m : Pm
j=1 xj ∈2Z}
Qm
δ (xδ).
m, we
(44)
Lemma 20. Let E ⊂ X ×X be a measurable set. Then there exists for every ε ∈ (0, ∞) a
δ0 ∈ (0, ∞) such that
L 2m((Sδ ×Sc
δ) ∩ E) ≥(cid:18) 1
4
− ε(cid:19) L 2m(E)
for all
δ ∈ (0, δ0),
where Sc
δ = R
m \Sδ.
Proof. Let ε ∈ (0, ∞) be arbitrarily given. Since E is a measurable set, we can cover it with
pairwise disjoint cubes Q2m
2m, i ∈ , such that
ai (ξi), ai ∈ ∩ (0, ∞), ξi ∈
We further choose γ ∈ (0, ∞) such that
ai (ξi)\E) ≤
L 2m(E).
ε
4
ai (ξi)) ≤ εL 2m(E).
Q2m
L 2m(Fi∈
L 2m(F{i∈ : ai≤γ}Q2m
Since the set Sδ×Sc
length is an integer multiple of 2δ, we have for every δ ∈ (0, ∞), a ∈ (2δ, ∞), and ξ ∈ R
that
δ covers for every δ ∈ (0, ∞) exactly one forth of every cube whose side
2m
L 2m((Sδ ×Sc
δ) ∩ Q2m
a (ξ)) ≥
L 2m(Q2m
a (ξ)).
(a − 2δ)2m
4
=
(1 − 2δ
a )2m
4
So, with δ0 = (1 − (1 − 2ε)
1
2m ) γ
2 , we get for every δ ∈ (0, δ0) that
L 2m((Sδ ×Sc
δ) ∩ Q2m
a (ξ)) ≥(cid:18) 1
4
−
ε
2(cid:19) L 2m(Q2m
a (ξ))
22
peter elbau
for every cube Q2m
therefore,
a (ξ) with side length a ∈ (γ, ∞) and arbitrary center ξ ∈ R
2m, and
L 2m((Sδ ×Sc
δ) ∩ E) ≥ L 2m((Sδ ×Sc
ai (ξi)) −
ai (ξi)) −
L 2m(E)
ε
4
L 2m(E)
ε
4
ε
−
δ) ∩F{i∈ : ai>γ}Q2m
2(cid:19) L 2m(F{i∈ : ai>γ}Q2m
2(cid:19) (L 2m(E) − εL 2m(E)) −
− ε(cid:19) L 2m(E),
−
ε
4
≥(cid:18) 1
≥(cid:18) 1
≥(cid:18) 1
4
4
L 2m(E)
ε
4
as desired.
(cid:3)
In particular, this result shows that we can also choose for finitely many measurable sets
Ej ⊂ X ×X, j ∈ [1, N ] ∩ , N ∈ , and arbitrary ε ∈ (0, ∞) a δ ∈ (0, ∞) such that the set
Sδ ×Sc
δ fulfils
L 2m((Sδ ×Sc
δ) ∩ Ej) ≥(cid:18) 1
4
− ε(cid:19) L 2m(Ej)
for all j ∈ [1, N ] ∩ .
The second lemma slightly generalises the result that if a measurable function g : X ×
R
n → R fulfils an integral inequality of the form RA g(x, ω(x)) dx ≥ 0 for all sets A ⊂ X
n), then g(x, w) ≥ 0 for almost every x ∈ X and for all w ∈ R
and all ω ∈ L∞(X; R
n.
n → R, w 7→ g(x, w) is
Lemma 21. Let g : X×R
continuous for almost all x ∈ X, and such that there exists for every M ∈ (0, ∞) a function
αM ∈ L1(X) with g(x, w) ≤ αM (x) for almost all x ∈ X and all w ∈ [−M, M ].
n → R be a function such that the map R
If there exists for every subset E ⊂ X with positive measure and every M ∈ (0, ∞) a
measurable subset E′
M ⊂ E with positive measure such that
g(x, ω(x)) dx ≥ 0
ZA
for all measurable sets A ⊂ E′
M and all ω ∈ L∞(X; R
n) with kωk∞ ≤ M , then
g(x, w) ≥ 0
for almost all x ∈ X and all w ∈ R
n.
Proof. We define for every k ∈ and M ∈ (0, ∞) the measurable set
Ek,M =
[{w∈
n : w≤M}
{x ∈ X : g(x, w) ≤ − 1
k }.
Let us assume by contradiction that there exists a set E ⊂ X with positive measure such that
Ek,M ⊃ E
we find for every x ∈ E a value w ∈ R
has positive measure, too. We thus find some k, M ∈ with L m(Ek,M ) > 0.
n with g(x, w) < 0. Then the unionSk,M∈
Now, by assumption, there exists a set E′
k,M ⊂ Ek,M with positive measure such that
g(x, ω(x)) dx ≥ 0
(45)
ZA
sequential lower semi-continuity of non-local functionals
23
for all measurable sets A ⊂ E′
k,M and all ω ∈ L∞(X; R
n) with kωk∞ ≤ M .
On the other hand, using Aumann's measurable selection theorem, we can find a function
k,M ,
(cid:3)
n) such that kωk∞ ≤ M and g(x, ω(x)) ≤ − 1
2k for almost all x ∈ E′
ω ∈ L∞(E′
which clearly contradicts (45).
k,M ; R
References
[1] Gilles Aubert and Pierre Kornprobst. Can the nonlocal characterization of Sobolev
spaces by Bourgain et al. be useful for solving variational problems? SIAM J. Numer.
Anal., 47(2):844 -- 860, 2009.
[2] Jonathan Bevan and Pablo Pedregal. A necessary and sufficient condition for the weak
lower semicontinuity of one-dimensional non-local variational integrals. Proc. Roy. Soc.
Edinburgh Sect. A, 136(4):701 -- 708, 2006.
[3] J´erome Boulanger, Peter Elbau, Carsten Pontow, and Otmar Scherzer. Non local func-
tionals in imaging. In H.H. Bauschke, R.S. Burachik, P.L. Combettes, V. Elser, D.R.
Luke, and H. Wolkowicz, editors, Fixed-Point Algorithms for Inverse Problems in Sci-
ence, volume 49 of Springer Optimization and Its Applications. Springer, 2011.
[4] Jean Bourgain, Haım Brezis, and Petru Mironescu. Limiting embedding theorems for
W s,p when s ↑ 1 and applications. J. Anal. Math., 87:77 -- 101, 2002. Dedicated to the
memory of Thomas H. Wolff.
[5] Irene Fonseca and Giovanni Leoni. Modern methods in the calculus of variations: Lp
spaces. Springer Monographs in Mathematics. Springer, New York, 2007.
[6] Stefan Kindermann, Stanley Osher, and Peter W. Jones. Deblurring and denoising of
images by nonlocal functionals. Multiscale Model. Simul., 4(4):1091 -- 1115 (electronic),
2005.
[7] Julio Munoz. On some necessary conditions of optimality for a nonlocal variational
principle. SIAM J. Control Optim., 38(5):1521 -- 1533 (electronic), 2000.
[8] Julio Munoz. Characterisation of the weak lower semicontinuity for a type of nonlocal
integral functional: the n-dimensional scalar case. J. Math. Anal. Appl., 360(2):495 --
502, 2009.
[9] Pablo Pedregal. Nonlocal variational principles. Nonlinear Anal., 29(12):1379 -- 1392,
1997.
[10] Augusto C. Ponce. A new approach to Sobolev spaces and connections to Γ-convergence.
Calc. Var. Partial Differential Equations, 19(3):229 -- 255, 2004.
[11] Carsten Pontow and Otmar Scherzer. A derivative-free approach to total variation
regularization. Arxiv preprint arXiv:0911.1293, 2009.
[12] Leonid I. Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based
noise removal algorithms. Phys. D, 60(1-4):259 -- 268, 1992.
|
1807.07335 | 1 | 1807 | 2018-07-19T10:36:06 | Totally smooth renormings | [
"math.FA"
] | We study the problem of totally smooth renormings of Banach spaces and provide such renormings for spaces which are weakly compactly generated. We also consider renormings for $(a,B,c)$-ideals. | math.FA | math |
TOTALLY SMOOTH RENORMINGS
EVE OJA, TAURI VIIL, AND DIRK WERNER
Abstract. We study the problem of totally smooth renormings of Banach
spaces and provide such renormings for spaces which are weakly compactly
generated. We also consider renormings for (a, B, c)-ideals.
1. Introduction
Let X be a Banach space. Following Phelps [P], we say that a subspace X of
a Banach space Z has property U in Z if every functional x∗ ∈ X ∗ has a unique
norm-preserving extension z∗ ∈ Z ∗. Following Liao and Wong [LW], we say that
X is totally smooth in Z if every closed subspace Y of X has property U in Z. If
Z = X ∗∗, then we say that X has property U in its bidual or, respectively, that X
is totally smooth in its bidual. Banach spaces with property U in their biduals are
also known as Hahn-Banach smooth spaces [S].
The notion of total smoothness in the bidual was essentially considered already
in 1977 by Sullivan [S]. It was further studied in [OPV], where several geometrical
conditions equivalent to total smoothness were proved.
Let πX : X ∗∗∗ → X ∗∗∗ denote the natural projection onto the dual space X ∗. It
is known (see [O2] or, e.g., [O4, p. 21]) that X has property U in its bidual X ∗∗
if and only if X has the strong uniqueness property SU in X ∗∗, meaning that the
following condition holds: for x∗∗∗ ∈ X ∗∗∗,
kπX x∗∗∗k = kx∗∗∗k ⇒ πX x∗∗∗ = x∗∗∗.
A Banach space X is called an M -ideal in its bidual if the equality
kx∗∗∗k = kπX x∗∗∗k + kx∗∗∗ − πX x∗∗∗k
holds for every x∗∗∗ ∈ X ∗∗∗. The notion of M -ideals was introduced by Alfsen
and Effros in [AE] and has since then been studied by many authors (see, e.g., the
monograph [HWW] for results and references).
The motivation for this paper comes from the following observation in [OPV,
Remark 2.8], which is based on [HWW, Theorem III.4.6].
Observation. If X is an M -ideal in its bidual X ∗∗, then X admits an equivalent
norm under which X is totally smooth and is still an M -ideal in its bidual.
As, clearly, being an M -ideal implies property U , the natural question arises
whether the M -ideal condition in the Observation could be relaxed.
Problem 1.1. If X has property U in its bidual X ∗∗, then does X admit an
equivalent norm under which X is totally smooth in its bidual?
Date: 17.7.2018.
2010 Mathematics Subject Classification. 46B03, 46B04, 46B20.
Key words and phrases. Totally smooth Banach space, property U , (a, B, c)-ideals, renormings.
1
2
EVE OJA, TAURI VIIL, AND DIRK WERNER
Let us note at once (see Example 3.5 below) that there exist Banach spaces that
admit an equivalent norm under which they are totally smooth, but they do not
admit any equivalent norm under which they are M -ideals in their biduals.
Now, we recall that according to the Taylor -- Foguel theorem (see [T] and [Fo]),
every subspace Y of X has property U in X if and only if the dual space X ∗ is
strictly convex, i.e., its unit sphere SX ∗ contains no non-trivial line segments.
Thus, relying on the Taylor -- Foguel theorem, we can characterize totally smooth
spaces by the strict convexity of the dual space as follows.
Theorem 1.2 (see [LW]). A Banach space X is totally smooth in its bidual X ∗∗
if and only if X has property U in X ∗∗ and the dual space X ∗ is strictly convex.
Therefore, Problem 1.1 is equivalent to the following problem.
Problem 1.3. If X has property U in its bidual X ∗∗, then does X admit an
equivalent norm under which the dual space X ∗ is strictly convex and X still has
property U in its bidual?
Recall that if X is separable and has property U in X ∗∗, then X ∗ is separable (see
Theorem 2.1 below). It was observed by Sullivan in [S, p. 321] that an application of
the Kadets -- Klee renorming theorem solves Problem 1.3 (and thus also Problem 1.1)
fully and positively for separable spaces. This theorem provides a Banach space
having a separable dual with an equivalent norm whose dual norm is strictly convex
and has the property that the weak∗ topology and the norm topology coincide on
the (new) dual sphere. A proof of this renorming theorem can be found in [Di,
pp. 113 -- 117]; it first appeared in Klee's paper [Kl2] relying on work by Kadets in
[Ka].
In Section 3, using a simple Klee-type renorming [Kl1], we give a partial positive
answer to Problem 1.3, and thus to Problem 1.1, in the general case (see Theo-
rem 3.2). In particular, our results also provide an alternative proof for the separa-
ble case. In Section 4, we come back to the Observation and we show that, under
natural assumptions, its claim holds for (a, B, c)-ideals, which are a far-reaching
generalization of M -ideals, encompassing in particular u- and h-ideals.
Our notation is standard. We consider Banach spaces over the scalar field K = R
or K = C. For a Banach space X, BX is the closed unit ball and SX is the unit
sphere of X. By span(xi), we denote the closed linear span of the elements xi. For
a subspace X of Z, X ⊥ = {z∗ ∈ Z ∗ : z∗
X = 0} is the annihilator. The density
character of the space X is denoted by dens X. For a bounded linear operator T ,
T ∗ is the adjoint operator, ran T is the range, and ker T is the kernel of T .
2. Useful results
In this section, we note some useful results regarding property U and very smooth
norms.
Recall that a Banach space X is an Asplund space if every separable subspace
Y of X has a separable dual space Y ∗. The following result (implicitly in [SS,
Theorem 15]) is well known.
Theorem 2.1. A Banach space X with property U in its bidual X ∗∗ is an Asplund
space.
In addition, we will need the following known result for Asplund spaces.
TOTALLY SMOOTH RENORMINGS
3
Theorem 2.2 (see [Fa, p. 112 and Theorem 8.3.3]). For a Banach space X, the
following conditions are equivalent.
(a) X has a shrinking Markushevich basis, i.e., there are (xi)i∈I in X and
(fi)i∈I in X ∗ such that I has the cardinality dens X, and
• fi(xj ) = δij,
• span(xi) = X,
• span(fi) = X ∗.
(b) X is weakly compactly generated (WCG) and Asplund.
Recall that a Banach space X is weakly compactly generated (WCG) if X is the
closed linear span of some weakly compact subset of X. The most important result
on WCG spaces is the following Amir -- Lindenstrauss theorem.
Theorem 2.3 (see [AL] or, e.g., [Fa, Theorem 1.2.5]). A Banach space X is weakly
compactly generated if and only if there exist a set Γ 6= ∅ and an injective weak∗-
to-weak continuous linear operator from X ∗ to c0(Γ).
We will also make use of the notion of very smooth spaces. First, recall that
a Banach space X is smooth whenever for every x ∈ SX , there exists a unique
functional fx ∈ SX ∗ such that fx(x) = 1. If X is smooth, then the support mapping
x 7→ fx from SX to SX ∗ is norm-to-weak∗ continuous (see, e.g., [Di, p. 22]).
Definition 2.4 (see [DF] or, e.g., [Di, p. 31]). A smooth Banach space X is called
very smooth if the support mapping x 7→ fx from SX to SX ∗ is norm-to-weak
continuous.
It is well known that a Banach space X is smooth whenever its dual space X ∗
is strictly convex, but it need not be very smooth in general [S].
To prove that a renorming is very smooth, we will use the following result.
Lemma 2.5 (see [G] or, e.g., [HWW, Lemma III.2.14]). A Banach space X has
property U in its bidual X ∗∗ if and only if the relative weak and weak∗ topologies
on BX ∗ coincide on SX ∗ .
3. Renorming of Banach spaces with property U
In order to try to solve Problem 1.3, we follow the strategy of the proof of the
Observation. This proof has three steps.
(i) The Banach space X has a shrinking Markushevich basis (as proved by
Fabian and Godefroy [FG]).
(ii) Using the shrinking Markushevich basis, one obtains an injective weak∗-to-
weak continuous linear operator from X ∗ to c0(Γ). This allows one to equip
X with a rather standard equivalent norm · such that for eX := (X,· ),
the dual norm of eX ∗ is strictly convex (as proved already by Amir and
(iii) It can be shown that the renorming in (ii) is such that eX is still an M -ideal
Lindenstrauss [AL]).
in eX ∗∗ (as proved by Harmand and Rao in [HR]).
The step (iii) can be extended from the M -ideal case (see [HR] or [HWW, Propo-
sition III.2.11]) to the property U case by the following theorem.
4
EVE OJA, TAURI VIIL, AND DIRK WERNER
Theorem 3.1. Let X be a Banach space with property U in its bidual X ∗∗. If Y
is a Banach space and T : Y → X a weakly compact operator, then
x∗ := kx∗k + kT ∗x∗k ,
x∗ ∈ X ∗,
Moreover, if T ∗ is injective and there is a strictly convex Banach space Z such
is an equivalent dual norm on X ∗ for which eX := (X, · ) has property U in eX ∗∗.
that ran T ∗ ⊂ Z ⊂ Y ∗, then eX ∗ is strictly convex.
Proof. Since T ∗ is an adjoint operator, it is weak∗-to-weak∗ continuous, and thus
the mapping x∗ 7→ kT ∗x∗k, x∗ ∈ X ∗, is weak∗ lower semicontinuous. Therefore,
· is an equivalent dual norm by a well-known result of Klee [Kl1]; see, e.g., [Di,
p. 106].
To calculate · on the third dual X ∗∗∗, we use the following argument due to
Harmand and Rao in [HR]. By definition, the operator
S : (X ∗, · ) → X ∗ ⊕1 Y ∗,
x∗ 7→ (x∗, T ∗x∗),
is isometric; hence
is isometric too. One can easily see that S∗∗x∗∗∗ = (x∗∗∗, T ∗∗∗x∗∗∗), so
S∗∗ : (X ∗∗∗, · ) → X ∗∗∗ ⊕1 Y ∗∗∗
x∗∗∗ = kx∗∗∗k + kT ∗∗∗x∗∗∗k ,
x∗∗∗ ∈ X ∗∗∗.
By the weak compactness of T ∗, we get that ran T ∗∗∗ ⊂ Y ∗, hence πY T ∗∗∗ = T ∗∗∗.
Since πY T ∗∗∗ = T ∗∗∗πX , we conclude that T ∗∗∗ = T ∗∗∗πX .
We need to show that the natural projection πX ∈ L(X ∗∗∗) satisfies the condition
πX x∗∗∗ = x∗∗∗ ⇒ πX x∗∗∗ = x∗∗∗.
Let x∗∗∗ ∈ X ∗∗∗ be such that πX x∗∗∗ = x∗∗∗. Then
0 = πX x∗∗∗ − x∗∗∗
= kπX x∗∗∗k + kT ∗∗∗πX x∗∗∗k − kx∗∗∗k − kT ∗∗∗x∗∗∗k
= kπX x∗∗∗k − kx∗∗∗k .
Therefore, kπX x∗∗∗k = kx∗∗∗k, and thus πX x∗∗∗ = x∗∗∗ by property U of X in
X ∗∗.
Moreover, if T ∗ is injective and there is a strictly convex Banach space Z such
that ran T ∗ ⊂ Z ⊂ Y ∗, then, thanks to Klee's renorming theorem in [Kl1] (see, e.g.,
[Di, Theorem 1, p. 100]), eX ∗ is strictly convex.
Using Theorem 3.1, we can now give a partial answer to Problem 1.1.
(cid:3)
Theorem 3.2. If a WCG Banach space X has property U in its bidual X ∗∗, then
X has a shrinking Markushevich basis and X admits an equivalent very smooth
norm under which X is totally smooth in its bidual.
Proof. Since X is also Asplund (see Theorem 2.1), it has a shrinking Marku-
shevich basis (see Theorem 2.2). All we need to finish the proof is an injective
weak∗-to-weak continuous linear operator S : X ∗ → c0(Γ) (for some set Γ). Such an
operator S exists according to the Amir -- Lindenstrauss theorem (see Theorem 2.3).
However, S can be very easily constructed using our shrinking Markushevich basis.
From now, we follow the proof of [HWW, Theorem III.4.6(e)].
TOTALLY SMOOTH RENORMINGS
5
Let (xi, fi)i∈I with xi ∈ X, fi ∈ X ∗ be a shrinking Markushevich basis. Assum-
ing kxik = 1, we define an operator S : X ∗ → c0(I) by
x∗ ∈ X ∗.
x∗ 7→ (x∗(xi)) ,
It is easy to check that the operator S is well-defined, injective, and weak∗-to-weak
In particular, S is weakly compact and weak∗-to-weak∗ continuous
continuous.
when considered as an operator into c0(I)∗∗, for which we use the notation S.
Hence, S is the adjoint of a weakly compact operator T : c0(I)∗ → X.
We now equip c0(I) with Day's equivalent strictly convex norm [Da] (see, e.g.,
[Di, p. 94]). Since T satisfies the requirements in Theorem 3.1, we get an equivalent
smooth norm · on X whose dual norm is strictly convex and for which X still has
property U in its bidual. The norm · on X is, in fact, very smooth. Indeed, as was
mentioned above, the support mapping on a smooth space is always norm-to-weak∗
continuous, hence for (X,·), by Lemma 2.5, the support mapping is norm-to-weak
continuous, i.e., (X, · ) is very smooth.
(cid:3)
Since separable Banach spaces are WCG, Theorem 3.2 gives an alternative proof
to the separable case considered by Sullivan in [S] that was mentioned in the Intro-
duction.
Corollary 3.3. If a separable Banach space X has property U in its bidual X ∗∗,
then X admits an equivalent very smooth norm under which X is totally smooth in
its bidual.
Readers particularly interested in the separable case as expounded in the previ-
ous corollary should notice that the rather easy argument of [LT1, Proposition 1.f.3]
shows that a Banach space with a separable dual admits a shrinking Markushevich
basis.
In order to obtain examples of spaces having a shrinking Markushevich basis, we
can use the notion of U ∗-spaces, which is dual to property U .
Definition 3.4 (see [CN1]). A Banach space X is said to be a U ∗-space in its
bidual X ∗∗ if for every x∗∗∗ ∈ X ∗∗∗ with πX x∗∗∗ 6= 0,
kx∗∗∗ − πX x∗∗∗k < kx∗∗∗k .
In [CN1, proof of Theorem 4.4], it was observed that the proofs of [FG, Theo-
rems 1 and 3] essentially yield that every Asplund U ∗-space has a shrinking Marku-
shevich basis and is WCG. Therefore, Theorem 3.2 applies to any U ∗-space with
property U in its bidual. This, together with some help from the literature, will be
used in the next example.
Example 3.5. Let Γ be an infinite set and let 1 < p < ∞. The lp-sum lp(c0(Γ))
admits an equivalent very smooth norm under which X is totally smooth in its
bidual, but it cannot be equivalently renormed to be an M -ideal in its bidual.
Proof. It is well known (see, e.g., [HWW, Example III.1.4(a)]) that c0(Γ) is an
M -ideal in its bidual. Hence, clearly, it is a U ∗-space. This property extends to the
lp-sum, which also has property U in its bidual (see [CN1, Proposition 2.2]). We
have the desired renorming of lp(c0(Γ)) thanks to Theorem 3.2.
Assume, for the sake of contradiction, that X := (lp(c0(Γ)), · ) is an M -ideal
in X ∗∗ for some equivalent norm · on lp(c0(Γ)). Since M -ideals in their biduals
are stable by taking closed subspaces (see [HL] or, e.g., [HWW, Theorem III.1.6]),
6
EVE OJA, TAURI VIIL, AND DIRK WERNER
Y := (lp(c0), · ) is an M -ideal in Y ∗∗. This contradicts [GKS, Proposition 4.4]
stating that if a separable M -ideal Y in Y ∗∗ has a boundedly complete Schauder
decomposition (Yn)∞
n=1, then all but finitely many subspaces Yn are reflexive. In
our case, all Yn, n = 1, 2, . . . , are isomorphic to c0 and thus non-reflexive.
(cid:3)
4. Renorming of (a, B, c)-ideals
According to [GKS], a closed subspace X of a Banach space Z is said to be an
ideal in Z if there is a contractive projection P on Z ∗ such that ker P = X ⊥. In
this case, the projection P is called an ideal projection. If ran P is norming, then
the ideal is called strict. If Z = X ∗∗ and P = πX , then the ideal is called canonical.
Canonical ideals are strict, but not vice versa.
Let a, c ≥ 0 and let B ⊂ K be a compact set. If X is an ideal in Z with an ideal
projection P such that
kaz∗ + bP z∗k + ckP z∗k ≤ kz∗k
∀b ∈ B
for all z∗ in Z ∗, then X is said to be an (a, B, c)-ideal in Z. The (a, B, c)-ideals
were introduced in [O6] (see also [O5]), but got their name later in [OP]. This
approach unifies all previously studied special cases of ideals. For instance, it is
easy to see that M -ideals coincide with (1,{−1}, 1)-ideals, u-ideals coincide with
(1,{−2}, 0)-ideals, and h-ideals are the same as (1,{−(1 + λ) : λ ∈ SC}, 0)-ideals.
The notions u- and h-ideals have been deeply studied in [GKS].
In the context of (a, B, c)-ideals, the Observation (from the Introduction) triggers
the following natural question (cf. Problem 1.1).
Problem 4.1. If X is an (a, B, c)-ideal with property U in its bidual X ∗∗, then
does X admit an equivalent norm under which X is totally smooth and is still an
(a, B, c)-ideal in its bidual?
Similarly to the property U case, we prove that the renorming from step (iii)
from the proof of the Observation can be extended to canonical (a, B, c)-ideals.
Concerning the special case of M -ideals, recall that an M -ideal in its bidual is
always canonical.
Theorem 4.2. Let X be a Banach space which is a canonical (a, B, c)-ideal in its
bidual X ∗∗. If Y is a Banach space and T : Y → X is a weakly compact operator,
then
x∗ := kx∗k + kT ∗x∗k ,
x∗ ∈ X ∗,
Moreover, if T ∗ is injective and there is a strictly convex Banach space Z such
is an equivalent dual norm under which eX = (X, · ) is still a canonical (a, B, c)-
ideal in eX ∗∗.
that ran T ∗ ⊂ Z ⊂ Y ∗, then eX ∗ is strictly convex.
here we need to show that eX is a canonical (a, B, c)-ideal. This means that
Proof. The proof is essentially the same as the proof of Theorem 3.1, except that
ax∗∗∗ + bπX x∗∗∗ + cπX x∗∗∗ ≤ x∗∗∗ ∀b ∈ B
holds for all x∗∗∗ in X ∗∗∗.
Let x∗∗∗ ∈ X ∗∗∗. Then kax∗∗∗ + bπX x∗∗∗k + ckπX x∗∗∗k ≤ kx∗∗∗k for all b ∈ B.
Therefore, recalling that on X ∗∗∗, the norm · is of the form x∗∗∗ 7→ x∗∗∗ =
TOTALLY SMOOTH RENORMINGS
7
kx∗∗∗k + kT ∗∗∗x∗∗∗k and T ∗∗∗πX = T ∗∗∗, we have
ax∗∗∗ + bπX x∗∗∗ + cπX x∗∗∗
= kax∗∗∗ + bπX x∗∗∗k + kaT ∗∗∗x∗∗∗ + bT ∗∗∗πX x∗∗∗k
+ ckπX x∗∗∗k + ckT ∗∗∗πX x∗∗∗k
= kax∗∗∗ + bπX x∗∗∗k + ckπX x∗∗∗k
+ kaT ∗∗∗x∗∗∗ + bT ∗∗∗x∗∗∗k + ckT ∗∗∗x∗∗∗k
≤ kx∗∗∗k + a + bkT ∗∗∗x∗∗∗k + ckT ∗∗∗x∗∗∗k
≤ kx∗∗∗k + kT ∗∗∗x∗∗∗k = x∗∗∗,
because a + b + c ≤ 1 for every b ∈ B (this can easily be verified by considering
an arbitrary x∗ ∈ SX ∗ ).
(cid:3)
Our next result extends [FG, renorming result on p. 142 after Theorem 3] and
[HWW, Theorem III.4.6(e)] from M -ideals to strict (a, B, c)-ideals. Note that The-
orem 4.3 applies in particular to u- and h-ideals.
Theorem 4.3. Let a, c ≥ 0, and let B be a compact set of scalars. Assume that a
Banach space X is a strict (a, B, c)-ideal in X ∗∗. If X has a shrinking Markushevich
basis, then X admits an equivalent smooth norm whose dual norm is strictly convex
and under which X becomes a canonical (a, B, c)-ideal in its bidual.
The proof of Theorem 4.3 uses the following result that relies on [GK] and extends
[GKS, Proposition 5.2, (1) and (2)], where u- and h-ideals were considered.
Proposition 4.4. Let a Banach space X be a strict (a, B, c)-ideal in X ∗∗. If X
does not contain l1 isomorphically, then X is a canonical (a, B, c)-ideal in X ∗∗.
Proof. The proof follows the scheme of the proof of [GKS, Proposition 5.2, (1)
and (2)].
By assumption, there is an (a, B, c)-ideal projection P on X ∗∗∗ such that ker P =
X ⊥ and ran P is a norming subspace of X ∗∗∗. It suffices to show that ran P = X ∗,
because then P = πX (recall that ker πX = X ⊥ and ran πX = X ∗), meaning that
X is a canonical (a, B, c)-ideal in X ∗∗.
Since X does not contain l1 isomorphically, we get from [GK, Corollary 5.5] that
X ∗∗∗ contains a minimal norming subspace, which is, by definition, the intersection
of all norming subspaces of X ∗∗∗. We know that X ∗ is norming in X ∗∗∗. On
the other hand, a proper subspace U of X ∗ cannot be norming in X ∗∗∗. Indeed,
assume, for the sake of contradiction, that such a U is norming in X ∗∗∗. By the
Hahn -- Banach theorem, there is x∗∗ ∈ X ∗∗ such that x∗∗(u) = 0 for every u ∈ U ,
but x∗∗ 6= 0. Since U is norming in X ∗∗∗, we get
kx∗∗k = sup
u∈BU x∗∗(u) = 0,
which is a contradiction.
Therefore, X ∗ is the minimal norming subspace, and thus X ∗ ⊂ ran P . Since
now also ran P ⊂ X ∗ (indeed, if x∗∗∗ = x∗ + x⊥ ∈ X ∗∗∗ with x∗ ∈ X ∗, x⊥ ∈ X ⊥,
then P x∗∗∗ = P x∗ = x∗, because X ∗ ⊂ ran P ), we have ran P = X ∗, as desired. (cid:3)
Proof of Theorem 4.3. Since X has a shrinking Markushevich basis, by Theorem 2.2,
X is Asplund. An Asplund space cannot contain l1 isomorphically, and thus, by
Proposition 4.4, X is a canonical (a, B, c)-ideal.
8
EVE OJA, TAURI VIIL, AND DIRK WERNER
Now, using Theorem 4.2 (instead of Theorem 3.1) in the proof of Theorem 3.2,
this immediately yields an equivalent smooth norm · on X whose dual norm is
strictly convex and for which X is still a canonical (a, B, c)-ideal in its bidual. (cid:3)
If the latter proof is carried out under the supplementary assumption of property
U , then one obtains the following partial positive answer to Problem 4.1, which is
quite similar to Theorem 3.2.
Theorem 4.5. Let a, c ≥ 0, and let B be a compact set of scalars. Assume that a
Banach space X is a strict (a, B, c)-ideal with property U in X ∗∗. If X is WCG,
then X has a shrinking Markushevich basis and X admits an equivalent very smooth
norm under which X becomes a totally smooth canonical (a, B, c)-ideal in its bidual.
For separable spaces, we can again omit the WCG-assumption.
Corollary 4.6. Let a, c ≥ 0, and let B be a compact set of scalars. Assume that a
separable Banach space X is a strict (a, B, c)-ideal with property U in X ∗∗. Then X
admits an equivalent very smooth norm under which X is a totally smooth canonical
(a, B, c)-ideal in its bidual.
As it was recalled in Section 3, every Asplund U ∗-space has a shrinking Marku-
shevich basis. Since this property is preserved under isomorphisms, Theorem 4.5
immediately implies the following.
Corollary 4.7. Let a, c ≥ 0 and let B be a compact set of scalars. Assume that a
Banach space X is a strict (a, B, c)-ideal with property U in X ∗∗. If X is isomorphic
to a U ∗-space, then X admits an equivalent very smooth norm under which X is a
totally smooth canonical (a, B, c)-ideal in its bidual.
Similarly to Theorem 4.5, we can apply Theorem 4.3 to obtain the following
result. However, here we need to use a couple of auxiliary results from the literature.
Theorem 4.8. Let a Banach space X be a strict (a, B, c)-ideal in X ∗∗ with
max{b : b ∈ B} + c > 1. If X is isomorphic to a U ∗-space, then X admits an
equivalent smooth norm whose dual norm is strictly convex and under which X
becomes a canonical (a, B, c)-ideal in its bidual.
Proof. A U ∗-space does not contain l1 isomorphically. This fact was observed in
[CN1, Proposition 4.1] as a direct consequence of [GKS, Proposition 2.6]. Hence, X
does not contain l1 isomorphically, and therefore, by Proposition 4.4, our (a, B, c)-
ideal X is canonical. But canonical (a, B, c)-ideals with B and c as above are
Asplund spaces (see [O6, proof of Theorem 4.1] or [OZ, Lemma 4.2]). Hence, X
has a shrinking Markushevich basis, and Theorem 4.3 applies.
(cid:3)
Note that the above (a, B, c)-ideal assumption is satisfied in all important cases,
including M -, u-, and h-ideals.
Clearly, M -ideals in their biduals and, more generally, the canonical (1,{−1}, c)-
ideals with c ∈ (0, 1] are U ∗-spaces. Hence, from Theorem 4.8, we have the following
example.
Example 4.9. Let X be a canonical (1,{−1}, c)-ideal in X ∗∗ with c ∈ (0, 1]. Then
X admits an equivalent smooth norm whose dual norm is strictly convex and under
which X is a canonical (1,{−1}, c)-ideal in its bidual.
TOTALLY SMOOTH RENORMINGS
9
Remark 4.10. A particular example of a (1,{−1}, c)-ideal is provided by certain
renormings of the James space J, as shown in [CN1, Example 3.5]. Namely, for
δ > √2 the renorming Jδ of the James space in [CN1, Example 3.5] is a canonical
(1,{−1}, c)-ideal in J ∗∗
if
δ
(1 + δc)2 + (1 + c)2 + 2(δc)2
2δ2
maxn (1 + δc)2
,
δ2
o <
1
2
.
(We take this opportunity to point out a disturbing typo in [CN1] where the de-
nominator in the first item of the maximum is 2 instead of δ2.) It is easy to see
that for each c < 1/√3, there is some δ > √2 satisfying the above inequality.
The point of this remark is that the James space J cannot be renormed to be
an M -ideal in its bidual, since it is non-reflexive and its bidual is separable, being
isomorphic to J. However, a non-reflexive space that is an M -ideal in its bidual
contains a copy of c0 (see [HL] or, e.g, [HWW, Corollary III.3.7(a)]; hence the
renormed James space Jδ is a nontrivial instance of Example 4.9.
Let us conclude the paper with a couple of examples where Corollary 4.7 applies.
It is known (see [CN1, Proposition 2.2]) that the space in Example 3.5 is a
canonical u-ideal in its bidual. Using Corollary 4.7 (instead of Theorem 3.2) in
the proof of Example 3.5 allows us to strengthen this -- using the same norm as in
Example 3.5 -- as follows. Note that canonical u-ideals in their biduals could be
considered as the closest important weakenings of M -ideals in their biduals.
Example 4.11. Let Γ be an infinite set and let 1 < p < ∞. The lp-sum lp(c0(Γ))
admits an equivalent very smooth norm under which it is a totally smooth canonical
u-ideal in its bidual. But lp(c0(Γ)) cannot be equivalently renormed to be an M -
ideal in its bidual.
Our last Example 4.12 will concern a large class of spaces. We need some more
notation. For Banach spaces X and Y , we denote by L(X, Y ) the Banach space of
bounded linear operators from X to Y , and by K(X, Y ) its subspace of compact
operators. We write L(X) and K(X), respectively, if X = Y . Recall that a net
(Kα) in K(X) is a compact approximation of the identity (CAI) provided Kαx → x
αx∗ → x∗ for all x∗ ∈ X ∗, then (Kα) is called a
for all x ∈ X. If, moreover, K ∗
shrinking CAI. If X has a CAI such that the convergence is uniform on compact
subsets of X, then X is said to have the compact approximation property (CAP).
Let 1 < p < ∞. A Banach space X is said to have the upper p-property (cf.
[OW, Proposition 1.1] or [HWW, pp. 306 and 327]) if X admits a shrinking CAI
(Kα) such that
lim sup
α
sup
x,y∈BX kKαx + (y − Kαy)k ≤ (kxkp + kykp)1/p.
If X and Y are both reflexive Banach spaces, and X or Y has the CAP, then
K(X, Y )∗∗ = L(X, Y ) (see [GS, Corollary 1.3]; in the AP-case, this is a well-known
result due to Grothendieck). There is a vast literature studying the position of
K(X, Y ) in L(X, Y ) in terms of ideals (see, e.g., [HJO] for recent results and a
large set of references). We are going to use [CN2, Theorem 4.4 and Corollary 4.5],
from which one can see that K(X, Y ) is a U ∗-space with property U and also a strict
(a,{−a}, c)-ideal for all a, c > 0 such that ap + cp ≤ 1 in L(X, Y ), whenever X is
an arbitrary Banach space and Y is a Banach space having the upper p-property.
Thanks to Corollary 4.7, we have the following rather general example.
10
EVE OJA, TAURI VIIL, AND DIRK WERNER
Example 4.12. Let X and Y be reflexive Banach spaces and let 1 < p < ∞. If
Y has the upper p-property, then K(X, Y ) admits an equivalent very smooth norm
under which K(X, Y ) is a totally smooth canonical (a,{−a}, c)-ideal in its bidual
for all a, c > 0 such that ap + cp ≤ 1.
Besides the lp(Γ) spaces and the Lorentz sequence spaces d(w, p), there are
many reflexive spaces enjoying the upper p-property, as one can see from [HWW,
pp. 306, 327].
In the context of Example 4.12, we are interested in cases when
K(X, Y ) is not an M -ideal in its bidual L(X, Y ). Here the classical example, due
to Hennefeld (see [H] or, e.g., [HWW, p. 305]), is that K(d(w, p)) is not an M -ideal
in its bidual L(d(w, p)). If 1 < p ≤ q < ∞, then K(lp(Γ), d(w, q)) is an M -ideal in
L(lp(Γ), d(w, q)) by [O1] (see, e.g., [O4, p. 53 or p. 66]). However, in [O3] (see, e.g.,
[O4, p. 73 and p. 77]), it is proved that for 1 < q < p < ∞ and an infinite set Γ,
K(lp(Γ), d(w, q)) is not an M -ideal in L(lp(Γ), d(w, q)) whenever w ∈ l p
, neither
is K(d(v, p)∗, d(w, q)) in L(d(v, p)∗, d(w, q)), 1 < p, q < ∞, whenever p > (p − 1)q
and d(v, p)∗ (which is a sequence space) is contained as a linear subspace in d(w, q).
p−q
Acknowledgements. This research was partially supported by institutional re-
search funding IUT20-57 of the Estonian Ministry of Education and Research and
by the German Academic Exchange Service (DAAD).
References
[AE]
[AL]
[CN1]
[CN2]
[Da]
[Di]
[DF]
[Fa]
E. M. Alfsen and E.G. Effros, Structure in real Banach spaces. Parts I and II, Ann.
Math. 96 (1972), 98 -- 173.
D. Amir and J. Lindenstrauss, The structure of weakly compact sets in Banach spaces,
Ann. Math. 88 (1968), 35 -- 46.
J. C. Cabello and E. Nieto, On properties of M-ideals, Rocky Mount. J. Math. 28
(1998), 61 -- 93.
J. C. Cabello and E. Nieto, An ideal characterization of when a subspace of certain
Banach spaces has the metric compact approximation property, Studia Math. 129 (1998),
185 -- 196.
M. M. Day, Strict convexity and smoothness of normed spaces, Trans. Amer. Math.
Soc. 78 (1955), 516 -- 528.
J. Diestel, Geometry of Banach Spaces -- Selected Topics, Lecture Notes in Math. 485,
Springer, Berlin -- Heidelberg -- New York, 1975.
J. Diestel and B. Faires, On vector measures, Trans. Amer. Math. Soc. 198 (1974),
253 -- 271.
M. Fabian, Gâteaux Differentiability of Convex Functions and Topology: Weak Asplund
Spaces, Canadian Mathematical Society Series of Monographs and Advanced Texts, Wi-
ley, New York, 1997.
[FG] M. Fabian and G. Godefroy, The dual of every Asplund space admits a projectional
[Fo]
[G]
[GK]
resolution of identity, Studia Math. 91 (1988), 141 -- 151.
S. R. Foguel, On a theorem by A. E. Taylor, Proc. Amer. Math. Soc. 9 (1958), 325.
G. Godefroy, Points de Namioka, espaces normants, applications à la théorie
isométrique de la dualité, Israel J. Math. 38 (1981), 209 -- 220.
G. Godefroy and N. J. Kalton, The ball topology and its applications, Proceedings
of the Iowa Workshop on Banach spaces, Contemp. Math. 85 (1989), 195 -- 238.
[GKS] G. Godefroy, N. J. Kalton, and P. D. Saphar, Unconditional ideals in Banach
[GS]
[H]
spaces, Studia Math. 104 (1993), 13 -- 59.
G. Godefroy and P. D. Saphar, Duality in spaces of operators and smooth norms
on Banach spaces, Illinois J. Math. 32 (1988), 672 -- 695.
J. Hennefeld, M -ideals, HB-subspaces, and compact operators, Indiana Univ. Math.
J. 28 (1979), 927 -- 934.
[HJO] R. Haller, M. Johanson, and E. Oja, M (r, s)-ideals of compact operators, Czechoslo-
vak Math. J. 62 (2012), 673 -- 693.
TOTALLY SMOOTH RENORMINGS
11
[HL]
[HR]
P. Harmand and Å. Lima, Banach spaces which are M -ideals in their biduals, Trans.
Amer. Math. Soc. 283 (1984), 253 -- 264.
P. Harmand and T. S. S. R. K. Rao, An intersection property of balls and relations
with M -ideals, Math. Z. 197 (1988), 277 -- 290.
[HWW] P. Harmand, D. Werner, and W. Werner, M-ideals in Banach Spaces and Banach
[Ka]
[Kl1]
[Kl2]
[LT1]
[LW]
[O1]
[O2]
[O3]
[O4]
[O5]
[O6]
[OP]
Algebras, Lecture Notes in Math. 1547, Springer, Berlin, 1993.
M. I. Kadets, On weak and norm convergence, Dokl. Akad. Nauk. SSSR 122 (1958),
13 -- 16 (in Russian).
V. L. Klee, Convex bodies and periodic homeomorphisms in Hilbert space, Trans. Amer.
Math. Soc. 74 (1953), 10 -- 43.
V. L. Klee, Mappings into normed linear spaces, Fund. Math. 49 (1960), 25 -- 34.
J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces I, Springer, Berlin --
Heidelberg -- New York, 1977.
C.-J. Liao and N.-C. Wong, Smoothly embedded subspaces of a Banach space, Tai-
wanese J. Math. 14 (2010), 1629 -- 1634.
E. Oja, On the uniqueness of the norm-preserving extension of a linear functional in
the Hahn-Banach theorem, Izv. Akad. Nauk. Est. SSR 33 (1984), 424 -- 438 (in Russian).
E. Oja, Strong uniqueness of the extension of linear continuous functionals according
to the Hahn -- Banach theorem, Math. Notes 43 (1988), 134 -- 139.
E. Oja, On M -ideals of compact operators and Lorentz sequence spaces, Proc. Est. Acad.
Sci. Phys. Math. 40 (1991), 31 -- 36.
E. Oja, Extensions of Functionals and the Structure of the Space of Continuous Linear
Operators, Tartu Univ. Publ., Tartu, 1991 (in Russian).
E. Oja, Géométrie des espaces de Banach ayant des approximations de l'identité con-
tractantes, C. R. Acad. Sci. Paris Sér. I Math. 328 (1999), 1167 -- 1170.
E. Oja, Geometry of Banach spaces having shrinking approximations of the identity,
Trans. Amer. Math. Soc. 352 (2000), 2801 -- 2823.
E. Oja and M. Põldvere Norm-preserving extensions of functionals and denting points
of convex sets, Math. Z. 258 (2008), 333 -- 345.
[OPV] E. Oja, M. Põldvere, and T. Viil, On totally smooth subspaces of Banach spaces:
[OW]
[OZ]
[P]
[SS]
[S]
[T]
the Vlasov theorem revisited, Studia Math. 238 (2017), 91 -- 99.
E. Oja and D. Werner, Remarks on M -ideals of compact operators on X ⊕p X, Math.
Nachr. 152 (1991), 101 -- 111.
E. Oja and I. Zolk, On commuting approximation properties of Banach spaces, Proc.
Royal Soc. Edinb. 139A (2009), 551 -- 565.
R. R. Phelps, Uniqueness of Hahn -- Banach extensions and unique best approximation,
Trans. Amer. Math. Soc. 95 (1960), 238 -- 255.
M. A. Smith and F. Sullivan, Extremely smooth Banach spaces, in: Banach Spaces
of Analytic Functions, Proc. Conf. Kent, Ohio, 1976, J. Baker, C. Cleaver and J. Diestel
(eds.) Lecture Notes in Math. 604, Springer, Berlin 1977, 125 -- 137.
F. Sullivan, Geometrical properties determined by the higher duals of a Banach space,
Illinois J. Math. 21 (1977), 315 -- 331.
A. E. Taylor, The extension of linear functionals, Duke Math. J. 5 (1939), 538 -- 547.
(Oja) Institute of Mathematics and Statistics, University of Tartu, J. Liivi 2,
50409 Tartu, Estonia, and Estonian Academy of Sciences, Kohtu 6, 10130 Tallinn,
Estonia
E-mail address: [email protected]
(Viil) Institute of Mathematics and Statistics, University of Tartu, J. Liivi 2,
50409 Tartu, Estonia
E-mail address: [email protected]
(Werner) Department of Mathematics, Freie Universität Berlin, Arnimallee 6,
14195 Berlin, Germany
ORCID: 0000-0003-0386-9652
E-mail address: [email protected]
|
1501.01897 | 1 | 1501 | 2015-01-08T16:24:01 | Essential Spectra of Induced Operators on Subspaces and Quotients | [
"math.FA"
] | Let $X$ be a complex Banach space and let $T$ be a bounded linear operator on $X$. For any closed $T$-invariant subspace $F$ of $X$, $T$ induces operators $T_{|F}:F \longrightarrow F$ and $T/F:X/F\longrightarrow X/F$. In this note, we give a simple proof of the fact that the essential spectra of $T_{|F}$ and $T/F$ are always contained in the polynomial hull of the essential spectrum of $T$. | math.FA | math |
Essential Spectra of Induced Operators on
Subspaces and Quotients.
D.C. Moore†
Abstract. Let X be a complex Banach space and let T be a bounded linear
operator on X. For any closed T -invariant subspace F of X, T induces operators
TF : F −→ F and T /F : X/F −→ X/F .
In this note, we give a simple proof
of the fact that the essential spectra of TF and T /F are always contained in the
polynomial hull of the essential spectrum of T .
0. Introduction
Henceforth, X is a complex Banach space, L(X) is the Banach algebra of
bounded linear operators on X and K(X) is the (closed) ideal in L(X) con-
sisting of all compact operators on X. The ideal consisting of all finite-rank
elements of L(X) will be denoted using the symbol K ′(X). We adopt the
convention whereby the essential spectrum of an operator T ∈ L(X) is the set
σe(T ) = {λ ∈ C : λ − T is not a Fredholm operator}.
The essential spectral radius, re(T ) is then taken to be sup{λ : λ ∈ σe(T )}
when X is infinite dimensional and 0 when X is finite dimensional. When X
is infinite dimensional, Atkinson's theorem asserts that σe(T ) is the spectrum
of the coset T + K(X) in the Calkin algebra L(X)/K(X). In this case, σe(T )
is therefore always a non-empty compact subset of C.
Given a closed subspace F of X, iF is the inclusion operator sending F
into X, QF is the quotient operator sending X onto X/F and LF (X) is the
closed subalgebra of L(X) consisting of all T ∈ L(X) for which T F ⊆ F . It is
clear that
LF (X) = {T ∈ L(X) : QF T iF = 0}.
The latest algebra is the natural domain of two continuous unital algbera
homomorphisms. The first, which sends T ∈ LF (X) to TF ∈ L(F ), will be
denoted πr
F . The second, which sends T ∈ LF (X) to T /F ∈ L(X/F ), will
be denoted πq
F map compact
operators to compact operators.
F . It is straightforward to check that πr
F and πq
† Email: [email protected]. Telephone: 447592886332
1
It will also be convenient to set
KF (X) = LF (X) ∩ K(X)
AF (X) = LF (X)/KF (X)
, K ′
, A′
F (X) = LF (X) ∩ K ′(X)
F (X) = LF (X)/K ′
F (X)
The algebra A′
F (X) will play only the most minor role in these proceedings,
so we need not be particularly troubled by the fact that it doesn't have a
'canonical' topology (except when X is finite dimensional).
Given any complex unital algebra A and any element T ∈ A, we will write
σA(T ) = {λ ∈ C : λ − T has no inverse in A},
for the spectrum of T , suppressing the superscript A when the algebra with re-
spect to which the spectrum is being taken is unambiguous. The polynomially
convex hull of a compact subset S of C will be denoted bS.
In this paper, we are concerned with the relationships between the sets
σe(T )
, σe(TF )
, σe(T /F )
,
σAF (X)(T + KF (X))
and σA′
F (X)(T + K ′
F (X)).
There is a some precedent for investigating questions of this type. For example,
it is known that:
(a) If (U, V, W ) is a permutation of (T, TF , T /F ) then σe(U) ⊆ σe(V )∪σe(W );
(b) σA′
F (X)(T + K ′
F (X)) = σe(T ) ∪ σe(TF ).
for any Banach space X, any closed subspace F of X, and any T ∈ LF (X).
The paper, [7], of Djordjevi´c and Duggal seems to be the most comprehensive
open-access reference in this direction, and builds on the results obtained by
Barnes in [2]. Equation (b) is proved in the monograph, [3], of Barnes, Murphy,
Smyth and West. All three of the references mentioned here contain various
improvements of (a) and (b) in the presence of additional hypotheses on X, F
and T . For example, it is easy to show that:
(c) σe(T ) = σe(TF ) ∪ σe(T /F ) if F admits a T -invariant complement in X.
It is also clear that σe(T ) = σe(TF ) if F is of finite codimension in X. The
following example illustrates that the inclusion σe(TF ) ⊆ σe(T ) is too much
to hope for in the general case. I am indebted to J. R. Partington for drawing
my attention to this.
2
Example. Given any open U ⊆ R, let L2U denote the closure in L2(R) of the
subspace {u ∈ Cc(R) : u has compact support in U} (identifying Cc(R) with
its image under the obvious embedding into L2(R) in the tradition fashional).
Let T : L2(R) −→ L2(R) be the operator which satisfies
(T u)(y) = u(y − 1)
for each y ∈ R and u ∈ Cc(R). Plainly,
T (L2(0, +∞)) ⊆ L2(1, +∞) ⊆ L2(0, +∞),
so F = L2(0, +∞) is a closed T -invariant subspace of L2(R). It is also clear
that T F ⊆ L2(1, +∞) is orthogonal to L2(0, 1), so TF is not lower semi-
Fredholm. Thus 0 ∈ σe(TF ). Since T is an invertible isometry, we also have
σ(T ) ⊆ T, so σe(TF ) is certainly not contained in σe(T ).
One noticable feature of this example is that σe(TF ) is still contained in \σe(T ).
The purpose of this paper is to give an easy proof that this is actually always
true.
1. The Result: Statement and Discussion
The precise form of the result we obtain is as follows.
Theorem 1. Let X be a complex Banach space and let F be an arbitrary
closed subspace of X. Let T ∈ LF (X). Then
σe(TF ) ∪ σe(T /F ) ⊆ \σe(T )
In particular, max{re(TF ), re(T /F )} ≤ re(T ).
The proof we give is based on two simple observations, both of which occur as
exercises in the book, [1], of Abramovich and Aliprantis. These are:
(i) T ∈ LF (X) and z ∈ C \ [σ(T ) ⇒ (z − T )−1 ∈ LF (X);
(ii) T ∈ LF (X) ⇒ σ(TF ) ⊆ [σ(T ).
The details appear here for the convenience of the reader. Let L(F, X/F ) be
the Banach space of all bounded linear operators from F into X/F and define
G : C \ [σ(T ) −→ L(F, X/F )
by setting G(z) = QF ((z − T )−1)iF for each z ∈ C \ [σ(T ).
3
It is clear that G is holomorphic in C \ [σ(T ) and that assertion (i) is equivalent
to G being identically 0 on C \ [σ(T ). By the identity theorem for Banach-
space-valued functions, it suffices to show that G(z) = 0 for z > r(T ). Let
z ∈ C with z > r(T ). Then, assuming that T ∈ LF (X),
G(z) = QF lim
n→∞
nXm=0
T m
zm+1! iF = lim
n→∞
QF T miF
zm+1 = 0,
nXm=0
since T m belongs to LF (X) for each m. This clearly establishes that
σLF (X)(T ) ⊆ [σ(T )
Assertion (ii) and the fact that we also have σ(T /F ) ⊆ [σ(T ) follows immedi-
ately, since
σ(TF ) = σ(πr
and σ(T /F ) = σ(πq
F (T )) ⊆ σLF (X)(T ),
F (T )) ⊆ σLF (X)(T ).
Observation (i) has another interesting consequence (which is not noted in [1]).
Let λ be an isolated point of σ(T ) and suppose that there is a neighbourhood
V of λ such that V \ {λ} ⊆ C \ [σ(T ). By choosing r > 0 sufficiently small,
we can arrange it that γ(t) = λ + reit belongs to V for every t ∈ [0, 2π). The
spectral projection
PT (λ) =
1
2πiZγ
(z − T )−1 dz
(1)
is thus expressed as a limit of sums of elements (z − T )−1 for z ∈ C \ [σ(T ).
That PT (λ) ∈ LF (X) follows. By (ii), either λ − TF is invertible or λ occurs as
an isolated point of σ(TF ). It therefore makes sense to consider the spectral
projection of TF associated with λ. Since πr
F is an algebra homomorphism, we
have
PTF (λ) =
1
2πiZγ
(z − TF )−1 dz =
1
2πiZγ
πr
F ((z − T )−1) dz
As πr
F is a continuous linear map, this gives
PTF (λ) = πr
F(cid:18) 1
2πiZγ
(z − T )−1 dz(cid:19) = πr
F (PT (λ)) = PT (λ)F ,
with an entirely parallel conclusion available for PT /F (λ).
4
The proof of Theorem 1 also uses the following standard fact (c.f. [6], Lemma
4.3.17): if λ ∈ C is an isolated point of σ(T ) then PT (λ) has finite rank if and
only if λ − T is a Fredholm operator.
2. Proof of Theorem 1.
F or πq
Let π be either of πr
F , so that π(T ) = TF or π(T ) = T /F . It is clear that
π is a continuous unital algebra homomorphism and maps finite rank operators
to finite rank operators. Let U = C \\σe(T ). By the punctured neighbourhood
theorem ([11], Theorem 19.4), U ∩ σ(T ) consists of isolated points of σ(T ).
Thus every λ ∈ U has a neighbourhood, V , such that
V \ {λ} ⊆ C \ [σ(T ).
Let λ ∈ U and let V be a neighbourhood of λ with the property above. Since
λ /∈ σe(T ), PT (λ) is a finite rank operator. Also, as we saw in Section 1, either
λ ∈ C \ σ(T ) or is an isolated point of σ(π(T )). In the first case, λ − π(T ) is
certainly Fredholm and the proof ends here. Otherwise, the argument at the
end of Section 1 gives PT (λ) ∈ LF (X) and Pπ(T )(λ) = π(PT (λ)). Since π maps
finite rank operators, λ − π(T ) is Fredholm, so λ /∈ σe(π(T )). Having shown
that λ ∈ C \\σe(T ) ⇒ C \ σe(π(T )), when π is either of πr
F , we conclude
that
F or πq
σe(T /F ) ∪ σe(TF ) ⊆ \σe(T ),
as asserted. The inequality max{re(TF ), re(T /F )} ≤ re(T ) follows.
3. Closing Remarks
All of the ingredients of the proof have been known in isolation for many years.
In many ways, it would be quite surprising if no version of it had appeared
before. Another, similarly elementary result which does not appear to exist in
the literature is the fact that, with X, F and T as in Theorem 1,
σe(TF ) ∪ σe(T /F ) ∪ σe(T ) ⊆ σAF (X)(T + KF (X)).
This 'soft' result simply follows from the fact that πr
F and the inclusion
of LF (X) into L(X) map compact operators to compact operators. It is also
not difficult to show, in the special case where F is complemented in X, that
σle(TF ) ∪ σre(T /F ) ⊆ σe(T ), where σle and σre denote left and right essential
spectra (definitions can be found in [11], page 172). This latest observation
also appears to be missing from the literature.
F , πq
5
Acknowledgements: I am indebted to the EPSRC (Grant EP/K503101/1)
for their financial support during the preparation of this note. Some of the
material in this paper may appear the author's PhD thesis.
References
[1] Y. A. Abramovich and C. D. Aliprantis, An Invitation to Operator Theory,
AMS Graduate Studies in mathematics, Providence, 2002.
[2] B. A. Barnes, Spectral and Fredholm Theory involving the diagonal of a
bounded linear operator, Acta Sci. Math. (Szeged) 73 (2007), p237-250.
[3] B. A. Barnes, G. J. Murphy, M. R. F. Smyth and T. T. West, Riesz
and Fredholm theory in Banach algebras, Research Notes in Mathematics,
Pitman Advanced Publishing Program, London, 1982.
[4] A. Brunel and D. Revuz, Quelques applications probabilistes de la quasi-
compacit´e, Annals of the Henri Poincar´e institute, Section B, Volume 10,
No. 3, 1974, p301-337.
[5] S. R. Caradus, W. E. Pfaffenberger and B. E. Yood, Calkin Algebras and
algebras of operators on Banach spaces, Marcel Dekker, Inc., New York,
1974.
[6] E. B. Davies, Linear Operators and their Spectra, Cambridge University
Press, Cambridge, 2007.
[7] S. V. Djordjevi´c and B. P. Duggal, Spectral Properties of Linear Operators
through Invariant Subspaces, Journal of Functional Analysis, Approxima-
tion and Computation 1 (1), 2009, p19-29.
[8] S. Grabiner, Ranges of Products of Operators, Canadian Journal of Math-
ematics, Vol. XXVI, 6 (1074), p1430-1441.
[9] H. G. Heuser, Functional Analysis, translated by J. Horvath, John Wiley
and Sons, United States, 1982.
[10] K. B. Laursen and M. M. Neumann, Introduction to Local Spectral Theory,
London Mathematical Society Monographs New Series, Oxford University
Press, New York, 2000.
[11] V. Muller, Spectral Theory of Linear Operators and Spectral Systems in
Banach Algebras, Second edition, Operator Theory Advances and Appli-
cations 139, Birkhauser Verlag, Germany, 2007.
[12] A. F. Ruston, Fredholm Theory in Banach spaces, Cambridge University
Press, Cambridge, 1986.
6
|
1807.01003 | 1 | 1807 | 2018-07-03T07:44:09 | A short note on band projections in partially ordered vector spaces | [
"math.FA"
] | Consider an Archimedean partially ordered vector space $X$ with generating cone (or, more generally, a pre-Riesz space $X$). Let $P$ be a linear projection on $X$ such that both $P$ and its complementary projection $I - P$ are positive; we prove that the range of $P$ is a band. This shows that the well-known concept of band projections on vector lattices can, to a certain extent, be transferred to the framework of ordered vector spaces. | math.FA | math |
A SHORT NOTE ON BAND PROJECTIONS IN PARTIALLY
ORDERED VECTOR SPACES
JOCHEN GL UCK
Abstract. Consider an Archimedean partially ordered vector space X with
generating cone (or, more generally, a pre-Riesz space X). Let P be a linear
projection on X such that both P and its complementary projection I − P
are positive; we prove that the range of P is a band. This shows that the
well-known concept of band projections on vector lattices can, to a certain
extent, be transferred to the framework of ordered vector spaces.
1. Introduction
Consider a vector space X over the real field and let X+ be a cone in X, by
which we mean that αX+ + βX+ ⊆ X+ for all scalars α, β ∈ [0, ∞) and that
X+ ∩ (−X+) = {0}. The tuple (X, X+) is called a partially ordered vector space,
and the partial order ≤ on X referred to by this terminology is the order given by
x ≤ y if and only if y − x ∈ X+.
Partially ordered vector spaces have been present in functional analysis since
the first half of the 20th century, and a special focus has often been placed on an
important special case called vector lattices or Riesz spaces, which are partially
ordered vector spaces in which any two elements have an infimum. For an overview
over the theory of vector lattices we refer to one of the classical monographs [5],
[14], [8] and [7].
The concepts of disjointness, disjoint complements and bands are an important
building block of the theory of vector lattices. In 2006 van Gaans and Kalauch [11]
generalised those notions to partially ordered vectors spaces and, based on earlier
work by van Haandel [13], they could show that some important properties of bands
remain true on partially ordered vector spaces under very mild assumptions on the
space. This lead to a series of follow-up papers [12, 9, 2, 6, 3] where disjointness
and bands in partially ordered vector spaces were studied in more detail; moreover,
in [4] it was demonstrated that this theory can be used to generalise results about
disjointness preserving C0-semigroups on Banach lattices to the setting of ordered
Banach spaces.
In the present short note we consider a linear projection P on an ordered vector
space (X, X+) and assume that both P and I − P are positive. We call such a
projection an order projection and, under the same mild assumptions on (X, X+)
as in [11], we show that the range of P is a band (and in fact even a projection
band) in X. This opens the door for a theory of band projections on (X, X+) which
is, to a certain extent, similar to the theory of band projections on vector lattices.
In Section 2 we recall some terminology and give a brief reminder of bands in
partially ordered vector spaces. In Section 3 we discuss order projections and prove
our main result.
Date: July 4, 2018.
1
2
JOCHEN GL UCK
2. Bands in partially ordered vector spaces
Let (X, X+) be a partially ordered vector space. This space is called directed if,
for all x, y ∈ X, there exists z ∈ X such that z ≥ x and z ≥ y. The cone X+ is
called generating if X+ − X+ = X, and it is easy to see that (X, X+) is directed if
and only if X+ is generating. The partially ordered vector space (X, X+) is called
Archimedean if the inequality nx ≤ y for two vectors x, y ∈ X and all positive
integers n ∈ N := {1, 2, . . . } implies that x ≤ 0. If, for instance, X carries a norm
and X+ is closed with respect to this norm, one can readily check that (X, X+) is
Archimedean.
The concept of disjointness in (X, X+) can be defined as follows: two vectors
x, y ∈ X are called disjoint, and we denote this by x ⊥ y, if
{x + y, −x − y}u = {x − y, y − x}u;
here, the set S u := {z ∈ X : z ≥ x for all x ∈ S} denotes the set of all upper
bounds of any given set S ⊆ X. This definition of disjointness was introduced in
[11] and is motivated by the fact that two elements x and y of a vector lattice are
disjoint if and only if x + y = x − y. If ones used this fact, it is not difficult to
see that the above definition of disjointness coincides with the classical concept of
disjointness in case that (X, X+) is a vector lattice.
Some elementary properties of disjoint elements in a partially ordered vector
space (X, X+) can be found in [11, Section 1]. For instance, besides the evident
fact that x ⊥ y if and only if y ⊥ x, one also has x ⊥ x if and only if x = 0.
Moreover, note that the zero vector is disjoint to every other vector in X. We will
need the following characterisation of disjointness in case that both x and y are
positive. Recall that if two elements x, y ∈ X have a largest lower bound in X then
this largest lower bound is uniquely determined and called the infimum of x and y.
Proposition 2.1. Two vectors 0 ≤ x, y ∈ X are disjoint if and only if their
infimum exists and equals 0.
Proof. We first note that {x+y, −x−y}u = {x+y}u (since −x−y ≤ 0 ≤ x+y) and
that x + y is an upper bound of both x − y and y − x, so {x + y}u ⊆ {x − y, y − x}u.
Now assume that x and y have an infimum and that this infimum coincides with
0. Let f be an upper bound of x − y as well as of y − x. Then we have y ≥ x − f
as well as x ≥ y − f . This implies that 2y ≥ x + y − f and 2x ≥ x + y − f , so
(x + y − f )/2 is a lower bound of both x and y and thus, this vector is negative.
Hence, x + y ≤ f .
Now assume instead that {x + y}u = {x − y, y − x}u. Clearly, 0 is a lower bound
of both x and y and we have to show that it is the largest lower bound. So let
f ∈ X be another lower bound of x and y. Then x − f and y − f are both positive,
so we conclude that
x + y − f ≥ x ≥ x − y
and
x + y − f ≥ y ≥ y − x.
Hence, x + y − f is an upper bound of {x − y, y − x} and thus, by assumption, also
an upper bound of x + y. Therefore, f ≤ 0.
(cid:3)
We will also need the following simple fact which can, for instance, be derived
from Proposition 2.1 above.
Proposition 2.2. Let x1, x2 ∈ X. If 0 ≤ x1 ≤ x2 and x2 is disjoint to a positive
vector y ∈ X, then x1 is disjoint to y, too.
Now we finally come to the concept of disjoint complements. For each subset
S ⊆ X the set
S⊥ := {x ∈ X : x ⊥ y for all y ∈ S}
A SHORT NOTE ON BAND PROJECTIONS IN PARTIALLY ORDERED VECTOR SPACES 3
is called the disjoint complement of S. Inspired by the theory of vector lattices one
suspects this concept to be quite useful, but only if one can prove that S⊥ is always
a vector subspace of X. It was shown in [11, Corollary 2.2 and Section 3] that this is
true on a very large class of partially ordered vector spaces, namely on each so-called
pre-Riesz space. This concept is originally due to van Haandel [13]; for a precise
definition of pre-Riesz spaces we refer for instance to [13, Definition 1.1(viii)] or
[11, Definition 3.1]. Here, we only recall that every pre-Riesz space has generating
cone and that, conversely, every partially ordered vector space with generating cone
which is, in addition, Archimedean is a pre-Riesz space; see [11, Theorem 3.3] or
Haandel's original result in [13, Theorem 1.7(ii)] (but note that Haandel uses a
somewhat different terminology: he uses that notion integrally closed for what we
call Archimedean, and he uses the term Archimedean for another, weaker property).
Let (X, X+) be a pre-Riesz space; as mentioned above, this implies that the
disjoint complement S⊥ of any subset S of X is a vector subspace of X. A sub-
set B ⊆ X is called a band if the set B⊥⊥ := (B⊥)⊥ equals B (see [11, Defi-
nition 5.4]). For every S ⊆ X the disjoint complement S⊥ is itself a band [11,
Proposition 5.5(ii)]. Moreover, every band B ⊆ X is solid in the sense of [10,
Definition 2.1] or [11, Definition 5.1]; this was proved in [11, Proposition 5.3]. We
call a band B ⊆ X directed if and only if for all x, y ∈ B there exists z ∈ B such
that z ≥ x and z ≥ y; equivalently, the positive cone B+ := B ∩ X+ in B fulfils
B+ − B+ = B.
Inspired by the theory of vector lattices we call a band B ⊆ X a projection band
if X = B ⊕ B⊥, i.e. if X is the direct sum of B and its disjoint complement B⊥.
The following observation follows immediately from the definition of the notions
band and projection band.
Proposition 2.3. Let (X, X+) be a pre-Riesz-space and let B ⊆ X be a band.
Then B is a projection band if and only if B⊥ is a projection band.
If a projection band B ⊆ X is given, we can define a linear projection P onto
B along B⊥. This gives rise to the following terminology: a linear projection
P : X → X on a pre-Riesz space (X, X+) is called a band projection if there exists
a projection band B ⊆ X such that P is the projection onto B along B⊥; it follows
from Proposition 2.3 that a linear projection P is a band projection if and only if
I − P is a band projection. If B ⊆ X is a projection band, then there exists, of
course, only one band projection with range B, and this projection is called the
band projection onto B.
Recall that a linear operator on a partially ordered vector space (X, X+) is called
positive if it maps the cone X+ into itself. The following proposition shows, among
other things, that band projections are always positive.
Proposition 2.4. Let (X, X+) be a partially ordered vector space.
(a) Assume that a vector 0 ≤ x ∈ X can be written as x = y + z for disjoint vectors
y, z ∈ X. Then x and y are positive, too.
(b) If (X, X+) is a pre-Riesz space and P : X → X is a band projection, then P
and I − P are positive.
Proof. (a) Since x = y + z and x ≥ 0 ≥ −x = −y − z, we conclude that x is an
upper bound of {y + z, −y − z} and thus, by the disjointness of y and z, also an
upper bound of {y − z, z − y}. Hence,
y + z = x ≥ y − z
and
y + z = x ≥ z − y,
which implies z ≥ 0 and y ≥ 0.
(b) This is a consequence of (a).
(cid:3)
4
JOCHEN GL UCK
Let (X, X+) be a partially ordered vector space and let P : X → X be a linear
projection. If P is positive, then we have P X+ = P X ∩X+, so the cone P X+ on the
space P X induces the same order on P X as the order inherited from X. Moreover,
if X+ is generating in X, then P X+ is generating in P X. This observation is
employed in the proof of the following proposition and it is also used in Section 3
below.
Proposition 2.5. Let (X, X+) be a pre-Riesz space. Then every projection band
in X is directed.
Proof. If B ⊆ X is a projection band, then the band projection onto X is positive
according to Proposition 2.4(b). Since every pre-Riesz space has generating cone,
the assertion follows from the discussion right before Proposition 2.5.
(cid:3)
3. Order projections are band projections
We call a linear projection P on a partially ordered vector space (X, X+) an
order projection if both P and I − P are positive. Thus, P is an order projection if
and only if 0 ≤ P x ≤ x for all x ∈ X+. Proposition 2.4(b) above shows that every
band projection on a pre-Riesz space is an order projection. In Theorem 3.2 below
we prove that the converse implication is also true.
But first, we note the following result which is true on every partially ordered
vector space with generating cone, be it pre-Riesz or not.
Proposition 3.1. Let (X, X+) be a partially ordered vector space with generating
cone. If two order projections P and Q have the same range, then P = Q.
Proposition 3.1 was proved for ordered Banach spaces in [1, Proposition 2.1.4].
The proof for partially ordered vector spaces is virtually the same; we include it
here for the sake of completeness:
Proof of Proposition 3.1. It suffices to prove that P and Q have the same kernel.
To this end, first assume that 0 ≤ x ∈ ker P . We have 0 ≤ Qx ≤ x; as Qx is
contained in the range of P it follows that
0 ≤ Qx = P Qx ≤ P x = 0,
so indeed Qx = 0. Thus, Q maps every vector in X+ ∩ ker P to 0. Moreover, ker P
is the range of the positive projection I − P and the cone X+ is generating in X;
hence every vector in ker P can be written is a difference of two positive vectors in
ker P and thus, Q maps the entire space ker P to 0. This proves that ker P ⊆ ker Q.
By interchanging the roles of P and Q we also obtain ker Q ⊆ ker P which proves
the assertion.
(cid:3)
The following theorem is the main result of this short note.
Theorem 3.2. Let (X, X+) be a pre-Riesz space. Then a linear projection P :
X → X is an order projection if and only if it is a band projection.
Proof. If P is a band projection, then it is also an order projection by Propo-
sition 2.4(b). So assume conversely that P is an order projection.
It will be
convenient to denote the complementary projection of P by Q := I − P . Then
0 ≤ P, Q ≤ I.
We are going to prove that P X = (QX)⊥, and to this end we first show that
P X ∩ X+ = (QX ∩ X+)⊥ ∩ X+. In the following, we will tacitly use the character-
isation of disjointness of positive vectors given in Proposition 2.1.
"⊆" Let 0 ≤ x ∈ P X and let 0 ≤ y ∈ QX; we have to prove that x ⊥ y. Clearly,
0 is a lower bound of both x and y, so let f ∈ X be another lower bound of those
A SHORT NOTE ON BAND PROJECTIONS IN PARTIALLY ORDERED VECTOR SPACES 5
vectors. Then P f ≤ P y = 0 and Qf ≤ Qx = 0, so f = P f + Qf ≤ 0. Hence, x
and y have infimum 0 and are therefore disjoint.
"⊇" Let x ∈ X+ and assume that x is disjoint to every vector 0 ≤ y ∈ QX.
We have to show x ∈ P X which is equivalent to showing Qx = 0. The vector Qx
is positive and contained in QX, so it is disjoint to x. However, the vector Qx is
located between 0 and x, so it follows from Proposition 2.2 that Qx ⊥ Qx and thus,
Qx = 0.
Now we can show that actually P X = (QX)⊥:
"⊆" Let x ∈ P X and let y ∈ QX. We have to show that x and y are disjoint.
The positive cone (P X)+ := P X ∩X+ = P X+ in P X is generating in P X since X+
is generating in X. Similarly, (QX)+ := QX ∩ X+ is generating in QX. Hence, we
can find positive vectors x1, x2 ∈ (P X)+ and y1, y2 ∈ (QX)+ such that x = x1 − x2
and y = y1 − y2. The vector x1 is disjoint to both y1 and y2, so it is also disjoint
to y, and for the same reason the vector x2 is disjoint to y. Thus, x = x1 − x2 is
disjoint to y.
"⊇" Let x ∈ (QX)⊥. By the inclusion that we have proved right above, the
vector P x is also contained in (QX)⊥ and since (QX)⊥ is a vector subspace of X,
we conclude that Qx = x − P x is contained in (QX)⊥, too. In particular, Qx is
disjoint to itself, so Qx = 0. Hence, x = P x ∈ P X.
We have thus shown that P X = (QX)⊥, and now it is easy to deduce the
assertion of the theorem: by interchanging the roles of P and Q we also obtain
QX = (P X)⊥ and hence, (P X)⊥⊥ = (QX)⊥ = P X. Thus, P X is a band.
Moreover, we have X = P X ⊕ QX = P X ⊕ (P X)⊥, so P X is a projection band.
Since P is the projection onto P X along QX = (P X)⊥, we conclude that P is a
band projection.
(cid:3)
Remark 3.3. In [1, Section 2] a linear projection P on an ordered Banach space
with generating cone is called a band projection if both P and I − P are positive.
Theorem 3.2 shows that this terminology is consistent with the terminology used
in the present short note (note that an ordered Banach space in the sense of [1] has
closed cone and is thus Archimedean; hence, such a space is pre-Riesz in case that
it has a generating cone).
Let us close this note with the following consequence of Theorem 3.2.
Corollary 3.4. Let (X, X+) be a pre-Riesz space. If V ⊆ X is a vector subspace
of X and X = V + V ⊥, then V is a projection band.
Proof. First note that V ∩ V ⊥ = {0}, so the assumption implies that X = V ⊕ V ⊥.
By P : X → X we denote the projection onto V along V ⊥. Let 0 ≤ x ∈ X; since
the vectors P x and (I − P )x are disjoint, it follows from Proposition 2.4(a) that
both P x and (I − P )x are positive. Hence, P is an order projection and thus a
band projection according to Theorem 3.2. This readily implies that V = P X is a
projection band.
(cid:3)
Acknowledgement. The idea for Theorem 3.2 note arose during a very pleasant
visit of the author at the Institute of Analysis at Technische Universitat Dresden.
The author wishes to thank the members of the institute for their kind hospitality
and for many stimulating and enlightening discussions.
References
[1] Jochen Gluck and Manfred P. H. Wolff. Long–Term Analysis of Positive Operator Semigroups
via Asymptotic Domination. Preprint. Available online from arxiv.org/abs/1802.05364.
[2] Anke Kalauch, Bas Lemmens, and Onno van Gaans. Bands in partially ordered vector spaces
with order unit. Positivity, 19(3):489–511, 2015.
6
JOCHEN GL UCK
[3] Anke Kalauch and Helena Malinowski. Vector lattice covers of ideals and bands in pre-
Riesz spaces. To appear in Quaestiones Mathematicae. Preprint available online from
arxiv.org/abs/1801.07191.
[4] Anke Kalauch, Onno van Gaans, and Feng Zhang. Disjointness preserving C0-semigroups
and local operators on ordered Banach spaces. Indag. Math., New Ser., 29(2):535–547, 2018.
[5] W. A. J. Luxemburg and A. C. Zaanen. Riesz spaces. Vol. I. North-Holland Publishing
Co., Amsterdam-London; American Elsevier Publishing Co., New York, 1971. North-Holland
Mathematical Library.
[6] Helena Malinowski. Order closed ideals in pre-Riesz spaces and their relationship to bands.
Positivity, 2018. DOI: 10.1007/s11117-018-0558-5.
[7] P. Meyer-Nieberg. Banach lattices. Universitext. Springer-Verlag, Berlin, 1991.
[8] Helmut H. Schaefer. Banach lattices and positive operators. Springer-Verlag, New York-
Heidelberg, 1974. Die Grundlehren der mathematischen Wissenschaften, Band 215.
[9] O. van Gaans and A. Kalauch. Bands in pervasive pre-Riesz spaces. Oper. Matrices, 2(2):177–
191, 2008.
[10] Onno van Gaans. Seminorms on ordered vector spaces that extend to Riesz seminorms on
larger Riesz spaces. Indag. Math., New Ser., 14(1):15–30, 2003.
[11] Onno van Gaans and Anke Kalauch. Disjointness in partially ordered vector spaces. Positivity,
10(3):573–589, 2006.
[12] Onno van Gaans and Anke Kalauch. Ideals and bands in pre-Riesz spaces. Positivity,
12(4):591–611, 2008.
[13] M. van Haandel. Completions in Riesz Space Theory. PhD thesis, University of Nijmegen,
1993.
[14] A. C. Zaanen. Riesz spaces. II, volume 30 of North-Holland Mathematical Library. North-
Holland Publishing Co., Amsterdam, 1983.
E-mail address: [email protected]
|
1806.01889 | 5 | 1806 | 2019-04-14T22:53:56 | Series representations in spaces of vector-valued functions via Schauder decompositions | [
"math.FA"
] | It is a classical result that every $\mathbb{C}$-valued holomorphic function has a local power series representation. This even remains true for holomorphic functions with values in a locally complete locally convex Hausdorff space $E$ over $\mathbb{C}$. Motivated by this example we try to answer the following question. Let $E$ be a locally convex Hausdorff space over a field $\mathbb{K}$, $\mathcal{F}(\Omega)$ be a locally convex Hausdorff space of $\mathbb{K}$-valued functions on a set $\Omega$ and $\mathcal{F}(\Omega,E)$ be an $E$-valued counterpart of $\mathcal{F}(\Omega)$ (where the term $E$-valued counterpart needs clarification itself). For which spaces is it possible to lift series representations of elements of $\mathcal{F}(\Omega)$ to elements of $\mathcal{F}(\Omega,E)$? We derive sufficient conditions for the answer to be affirmative using Schauder decompositions which are applicable for many classical spaces of functions $\mathcal{F}(\Omega)$ having an equicontinuous Schauder basis. | math.FA | math |
SERIES REPRESENTATIONS IN SPACES OF VECTOR-VALUED
FUNCTIONS VIA SCHAUDER DECOMPOSITIONS
KARSTEN KRUSE
Abstract. It is a classical result that every C-valued holomorphic function
has a local power series representation. This even remains true for holomorphic
functions with values in a locally complete locally convex Hausdorff space E
over C. Motivated by this example we try to answer the following question. Let
E be a locally convex Hausdorff space over a field K, F (Ω) be a locally convex
Hausdorff space of K-valued functions on a set Ω and F (Ω, E) be an E-valued
counterpart of F (Ω) (where the term E-valued counterpart needs clarification
itself). For which spaces is it possible to lift series representations of elements
of F (Ω) to elements of F (Ω, E)? We derive sufficient conditions for the answer
to be affirmative using Schauder decompositions which are applicable for many
classical spaces of functions F (Ω) having an equicontinuous Schauder basis.
1. Introduction
The purpose of this paper is to lift series representations known from scalar-
valued functions to vector-valued functions and its underlying idea was derived from
the classical example of the (local) power series representation of a holomorphic
function. Let Dr ⊂ C be an open disc around zero with radius r > 0 and O(Dr) be
the space of holomorphic functions on Dr, i.e. the space of functions f ∶ Dr → C such
that the limit
f (1)(z) ∶= lim
h→0
h∈C,h≠0
h
f (z + h) − f (z)
,
z ∈ Dr,
(1)
exists in C. It is well-known that every f ∈ O(Dr) can be written as
f (z) = ∞Q
n=0
f (n)(0)
n!
zn,
z ∈ Dr,
where the power series on the right-hand side converges uniformly on every compact
subset of Dr and f (n)(0) is the n-th complex derivative of f at 0 which is defined
from (1) by the recursion f (0) ∶= f and f (n) ∶= (f (n−1))(1) for n ∈ N. Amazingly
by [9, 2.1 Theorem and Definition, p. 17-18] and [9, 5.2 Theorem, p. 35], this
series representation remains valid if f is a holomorphic function on Dr with values
in a locally complete locally convex Hausdorff space E over C where holomorphy
means that the limit (1) exists in E and the higher complex derivatives are defined
recursively as well. Analysing this example, we observe that O(Dr), equipped with
the topology of uniform convergence on compact subsets of Dr, is a Fréchet space,
in particular barrelled, with a Schauder basis formed by the monomials z ↦ zn.
Further, the formulas for the complex derivatives of a C-valued resp. an E-valued
function f on Dr are built up in the same way by (1).
Our goal is to derive a mechanism which uses these observations and transfers
known series representations for other spaces of scalar-valued functions to their
Date: April 16, 2019.
2010 Mathematics Subject Classification. Primary 46E40, Secondary 46A32, 46E10.
Key words and phrases. series representation, Schauder basis, Schauder decomposition, vector-
valued function, injective tensor product.
1
2
K. KRUSE
vector-valued counterparts. Let us describe the general setting. We recall from [11,
14.2, p. 292] that a sequence (fn) in a locally convex Hausdorff space F over a field
K is called a topological basis, or simply a basis, if for every f ∈ F there is a unique
sequence of coefficients (λK
n(f )) in K such that
f = ∞Q
n=1
λK
n(f )fn
(2)
n=1 λK
k ∶ F → F , P K
k (f ) ∶= ∑k
where the series converges in F . Due to the uniqueness of the coefficients the map
λK
n ∶ f ↦ λK
n (f ) is well-defined, linear and called the n-th coefficient functional asso-
ciated to (fn). Further, for each k ∈ N the map P K
n(f )fn,
is a linear projection whose range is span(f1, . . . , fn) and it is called the k-th ex-
pansion operator associated to (fn). A basis (fn) of F is called equicontinuous if
the expansion operators P K
k form an equicontinuous sequence in the linear space
L(F, F ) of continuous linear maps from F to F (see [11, 14.3, p. 296]). A basis (fn)
of F is called a Schauder basis if the coefficient functionals are continuous which,
in particular, is already fulfilled if F is a Fréchet space by [17, Corollary 28.11, p.
351]. If F is barrelled, then a Schauder basis of F is already equicontinuous and F
has the (bounded) approximation property by the uniform boundedness principle.
The starting point for our approach is equation (2) . Let F and E be locally
convex Hausdorff spaces over a field K where F has an equicontinuous Schauder
basis (fn) with associated coefficient functionals (λK
n). The expansion operators
(P K
k ) form a so-called Schauder decomposition of F (see [4, p. 77]), i.e. they are
continuous projections on F such that
min(j,k) for all j, k ∈ N,
for k ≠ j,
j = P K
k P K
k ≠ P K
k f ) converges to f for each f ∈ F .
(i) P K
(ii) P K
(iii) (P K
j
This operator theoretic definition of a Schauder decomposition is equivalent to
the usual definition in terms of closed subspaces of F given in [14, p. 377] (see
k ε idE) is a Schauder
[16, p. 219]).
decomposition of the Schwartz' ε-product F εE ∶= Le(F ′
κ, E) and each u ∈ F εE has
the series representation
In our main Theorem 3.1 we prove that (P K
u(f ′) = ∞Q
n=1
u(λK
n )f ′(fn),
f ′ ∈ F ′.
Now, suppose that F = F (Ω) is a space of K-valued functions on a set Ω with a
topology such that the point-evaluation functionals δx, x ∈ Ω, are continuous and
that there is a locally convex Hausdorff space F (Ω, E) of functions from Ω to E
such that the map
S∶ F (Ω)εE → F (Ω, E), u z→ [x ↦ u(δx)],
is a (linear topological) isomorphism. Assuming that for each n ∈ N and u ∈ F(Ω)εE
there is λE
n(S(u)) ∈ E with
λE
n(S(u)) = u(λK
n),
k ε idE) ○ S−1)k is a Schauder decomposition
(3)
we obtain in Corollary 3.6 that (S ○(P K
of F(Ω, E) and
f = lim
k→∞(S ○(P K
k ε idE) ○ S−1)k(f) = ∞Q
n=1
n(f)fn,
λE
f ∈ F(Ω, E),
which is the desired series representation in F(Ω, E). Condition (3) might seem
strange at a first glance but for example in the case of E-valued holomorphic func-
tions on Dr it guarantees that the complex derivatives at 0 appear in the Schauder
decomposition of O(Dr, E) since S(u)(n)(0) = u(δ(n)
0 ) for all u ∈ O(Dr)εE and
SERIES REPRESENTATIONS
3
0
n ∈ N0 where δ(n)
is the point-evaluation of the n-th complex derivative. We apply
our result to sequence spaces, spaces of continuously differentiable functions on a
compact interval, the space of holomorphic functions, the Schwartz space and the
space of smooth functions which are 2π-periodic in each variable.
As a byproduct of Theorem 3.1 we obtain that every element of the comple-
F is a complete space with an equicontinuous Schauder basis and E is complete.
tion F ⊗εE of the injective tensor product has a series representation as well if
Concerning series representation in F ⊗εE, little seems to be known whereas for
the completion F ⊗πE of the projective tensor product F ⊗π E of two metrisable
locally convex spaces F and E it is well-known that every f ∈ F ⊗πE has a series
representation
f = ∞Q
n=1
anfn ⊗ en
where (an) ∈ ℓ1, i.e. (an) is absolutely summable, and (fn) and (en) are null
sequences in F and E, respectively (see e.g. [10, Chap. I, §2 , n○1, Théorème 1, p.
51] or [11, 15.6.4 Corollary, p. 334]). If F and E are metrisable and one of them
is nuclear, then the isomorphy F ⊗πE ≅ F ⊗εE holds and we trivially have a series
representation of the elements of F ⊗εE as well. Other conditions on the existence
of series representations of the elements of F ⊗εE can be found in [18, Proposition
4.25, p. 88], where F and E are Banach spaces and both of them have a Schauder
basis, and in [12, Theorem 2, p. 283], where F and E are locally convex Hausdorff
spaces and both of them have an equicontinuous Schauder basis.
2. Notation and Preliminaries
We equip the spaces Rd, d ∈ N, and C with the usual Euclidean norm ⋅ . For a
subset M of a topological vector space X, we write acx(M) for the closure of the
absolutely convex hull acx(M) of M in X.
By E we always denote a non-trivial locally convex Hausdorff space, in short
lcHs, over the field K = R or C equipped with a directed fundamental system of
seminorms (pα)α∈A. If E = K, then we set (pα)α∈A ∶= { ⋅ }. We recall that for a
disk D ⊂ E, i.e. a bounded, absolutely convex set, the vector space ED ∶= ⋃n∈N nD
becomes a normed space if it is equipped with the gauge functional of D as a norm
(see [11, p. 151]). The space E is called locally complete if ED is a Banach space
for every closed disk D ⊂ E (see [11, 10.2.1 Proposition, p. 197]). For more details
on the theory of locally convex spaces see [8], [11] or [17].
By X Ω we denote the set of maps from a non-empty set Ω to a non-empty set
X, by χK the characteristic function of a subset K ⊂ Ω and by L(F, E) the space
of continuous linear operators from F to E where F and E are locally convex
Hausdorff spaces. If E = K, we just write F ′ ∶= L(F, K) for the dual space. If F
and E are (linearly topologically) isomorphic, we write F ≅ E and, if F is only
isomorphic to a subspace of E, we write F ↪E. We denote by Lt(F, E) the space
L(F, E) equipped with the locally convex topology of uniform convergence on the
absolutely convex compact subsets of F if t = κ and on the precompact (totally
bounded) subsets of F if t = γ.
The so-called ε-product of Schwartz is defined by
F εE ∶= Le(F ′
κ, E)
(4)
where L(F ′
κ, E) is equipped with the topology of uniform convergence on equicon-
tinuous subsets of F ′. This definition of the ε-product coincides with the original
one by Schwartz [22, Chap. I, §1, Définition, p. 18]. It is symmetric which means
that F εE ≅ EεF .
In the literature the definition of the ε-product is sometimes
done the other way around, i.e. EεF is defined by the right-hand side of (4) but
4
K. KRUSE
κ by F ′
due to the symmetry these definitions are equivalent and for our purpose the given
definition is more suitable. If we replace F ′
γ, we obtain Grothendieck's def-
inition of the ε-product and we remark that the two ε-products coincide if F is
quasi-complete because then F ′
κ. Jarchow uses a third, different definition of
the ε-product (see [11, 16.1, p. 344]) which coincides with the one of Schwartz if
F is complete by [11, 9.3.7 Proposition, p. 179]. However, we stick to Schwartz'
definition.
For locally convex Hausdorff spaces Fi, Ei and Ti ∈ L(Fi, Ei), i = 1, 2, we define
the ε-product T1εT2 ∈ L(F1εF2, E1εE2) of the operators T1 and T2 by
γ = F ′
(T1εT2)(u) ∶= T2 ○ u ○ T t
1,
u ∈ F1εF2,
1 ∶ E′
1 → F ′
1, e′ ↦ e′ ○ T1, is the dual map of T1. If T1 is an isomorphism
where T t
and F2 = E2, then T1ε idE2 is also an isomorphism with inverse T −1
1 ε idE2 by [22,
Chap. I, §1, Proposition 1, p. 20] (or [11, 16.2.1 Proposition, p. 347] if the Fi are
complete).
As usual we consider the tensor product F ⊗ E as an algebraic subspace of F εE
for two locally convex Hausdorff spaces F and E by means of the linear injection
Θ∶ F ⊗ E → F εE,
kQ
n=1
fn ⊗ en z→ y ↦
kQ
n=1
y(fn)en(cid:6).
Via Θ the space F ⊗ E is identified with space of operators with finite rank in F εE
and a locally convex topology is induced on F ⊗ E. We write F ⊗ε E for F ⊗ E
equipped with this topology and F ⊗εE for the completion of the injective tensor
product F ⊗ε E. By F(E) we denote the space of linear operators from E to E
with finite rank. A locally convex Hausdorff space E is said to have (Schwartz')
approximation property (AP) if the identity idE on E is contained in the closure
of F(E) in Lκ(E, E). The space E has AP if and only if E ⊗ F is dense in EεF for
every locally convex Hausdorff space (every Banach space) F by [13, Satz 10.17, p.
250]. The space E has Grothendieck's approximation property if idE is contained
in the closure of F(E) in Lγ(E, E). If E is quasi-complete, both approximation
properties coincide. For more information on the theory of ε-products and tensor
products see [7], [11] and [13].
A function f ∶ Ω → E on an open set Ω ⊂ Rd to an lcHs E is called continuously
partially differentiable (f is C 1) if for the n-th unit vector en ∈ Rd the limit
(∂en)Ef(x) ∶= lim
h→0
h∈R,h≠0
f(x + hen) − f(x)
h
exists in E for every x ∈ Ω and (∂en)Ef is continuous on Ω ((∂en)Ef is C 0) for
every 1 ≤ n ≤ d. For k ∈ N a function f is said to be k-times continuously partially
differentiable (f is Ck) if f is C 1 and all its first partial derivatives are Ck−1. A
function f is called infinitely continuously partially differentiable (f is C∞) if f is
Ck for every k ∈ N. For k ∈ N∞ ∶= N ∪{∞} the functions f ∶ Ω → E which are Ck form
a linear space which is denoted by Ck(Ω, E). For β ∈ Nd
n=1 βn ≤ k
and a function f ∶ Ω → E on an open set Ω ⊂ Rd to an lcHs E we set (∂βn)Ef ∶= f
if βn = 0, and
0 with β ∶= ∑d
(∂βn)Ef(x) ∶= (∂en)E ⋯(∂en)E
´¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¸¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¶
βn-times
f(x)
(∂β)Ef(x) =∶ (∂β1)E ⋯(∂βd)Ef(x)
if βn ≠ 0 and the right-hand side exists in E for every x ∈ Ω. Further, we define
if the right-hand side exists in E for every x ∈ Ω and set f (β) ∶= (∂β)Ef if d = 1.
SERIES REPRESENTATIONS
5
3. Schauder decomposition
Let us start with our main theorem on Schauder decompositions of ε-products.
3.1. Theorem. Let F and E be lcHs, (fn)n∈N an equicontinuous Schauder basis
of F with associated coefficient functionals (λn)n∈N and set Qn∶ F → F , Qn(f) ∶=
λn(f)fn for every n ∈ N. Then the following holds.
a) The sequence (Pk)k∈N given by Pk ∶= ∑k
n=1 Qnε idE is a Schauder decom-
b) Each u ∈ F εE has the series representation
position of F εE.
u(f ′) = ∞Q
n=1
u(λn)f ′(fn),
f ′ ∈ F ′.
c) F ⊗ E is sequentially dense in F εE.
Proof. Since (fn) is a Schauder basis of F , the sequence (∑k
n=1 Qn) converges
weakly to idF . Thus we deduce from the equicontinuity of (fn) that (∑k
n=1 Qn)
converges to idF in Lκ(F, F) by [11, Theorem 8.5.1 (b), p. 156]. For f ′ ∈ F ′ and
f ∈ F holds
m(f ′)(λn(f)fn) = f ′(λm(λn(f)fn)fm)
(Qt
○ Qt
m)(f ′)(f) = Qt
n
m(f ′)(Qn(f)) = Qt
= λm(fn)λn(f)f ′(fm) = ⎧⎪⎪⎨⎪⎪⎩
0
λn(f)f ′(fn)
, m = n,
, m ≠ n,
due to the uniqueness of the coefficient functionals (λn) (see [11, 14.2.1 Proposition,
p. 292]) and it follows for k, j ∈ N that
jQ
n=1
(
Qt
n
○
kQ
m=1
Qt
n)(f ′)(f) = min(j,k)Q
n=1
λn(f)f ′(fn) = min(j,k)Q
n=1
n(f ′)(f).
Qt
This implies that
(PkPj)(u) = u ○
jQ
n=1
Qt
n
○
kQ
m=1
n = u ○
Qt
min(j,k)Q
n=1
Qt
n = Pmin(j,k)(u)
for all u ∈ F εE.
uk,x∶ F ′ → E by uk,x(f ′) ∶= f ′(fk)x. Then uk,x ∈ F εE and
If k ≠ j, w.l.o.g. k > j, we choose x ∈ E, x ≠ 0, and define
(Pk − Pj)(uk,x) =
kQ
n=j+1
uk,x ○ Qt
n = uk,x ≠ 0
since
(uk,x ○ Qt
n)(f ′) = uk,x(λn(⋅)f ′(fn)) = λn(fk)f ′(fn)x = ⎧⎪⎪⎨⎪⎪⎩
f ′(fk)x , n = k,
, n ≠ k.
0
It remains to prove that for each u ∈ F εE
lim
k→∞
Pk(u) = u
in F εE. Let (qβ)β∈B denote the system of seminorms inducing the locally convex
topology of F . Let u ∈ F εE and α ∈ A. Due to the continuity of u there are an
absolutely convex compact set K = K(u, α) ⊂ F and C0 = C0(u, α) > 0 such that
for each f ′ ∈ F ′ we have
pα(Pk(u) − u)(f ′) = pαu(
= C0 sup
− idF ′)(f ′) ≤ C0 sup
f ∈K(
Qnf − f).
− idF ′)(f ′)(f)
f ∈Kf ′(
kQ
kQ
kQ
n=1
Qt
n
Qt
n
n=1
n=1
6
K. KRUSE
Let V be an absolutely convex 0-neighbourhood in F . As a consequence of the
equicontinuity of the polar V ○ there are C1 > 0 and β ∈ B such that
sup
f ′∈V ○
pα(Pk(u) − u)(f ′) ≤ C0C1 sup
f ∈K
qβ(
kQ
n=1
Qnf − f).
In combination with the convergence of (∑k
the convergence of (Pk(u)) to u in F εE and settles part a).
Let us turn to b) and c). Since
n=1 Qn) to idF in Lκ(F, F) this yields
Pk(u)(f ′) = u
kQ
n=1
Qt
n(f ′) = kQ
n=1
u(λn)f ′(fn)
for every f ′ ∈ F ′, we note that the range of Pk(u) is contained in span(u(λn) 1 ≤
n ≤ k) for each u ∈ F εE and k ∈ N. Hence Pk(u) has finite rank and thus belongs
to F ⊗ E implying the sequential density of F ⊗ E in F εE and the desired series
representation by part a).
(cid:3)
3.2. Remark. If F and E are complete, we have under the assumption of Theorem
3.1 that F ⊗εE ≅ F εE by c) since F εE is complete by [13, Satz 10.3, p. 234] and
F ⊗εE is the closure of F ⊗ E in F εE. Thus each element of F ⊗εE has a series
representation.
Let us apply the preceding theorem to spaces of Lebesgue integrable functions.
We consider the measure space ([0, 1], L([0, 1]), λ) of Lebesgue measurable sets
and use the notation Lp[0, 1] for the space of (equivalence classes) of Lebesgue
p-integrable functions on [0, 1]. The Haar system hn∶[0, 1] → R, n ∈ N, given by
h1(x) ∶= 1 for all x ∈ [0, 1] and
for k ∈ N0 and 1 ≤ j ≤ 2k forms a Schauder basis of Lp[0, 1] for every 1 ≤ p < ∞ and
the associated coefficient functionals are
λn(f) ∶= S
[0,1]
f(x)hn(x)dλ(x),
f ∈ Lp[0, 1], n ∈ N,
following corollary.
(see [19, Satz I, p. 317]). Because Lp[0, 1] is Banach space and thus barrelled, its
Schauder basis (hn) is equicontinuous and we directly obtain from Theorem 3.1 the
3.3. Corollary. Let E be an lcHs and 1 ≤ p < ∞. Then (∑k
a Schauder decomposition of Lp[0, 1]εE and for each u ∈ Lp[0, 1]εE holds
n=1 λn(⋅)hnε idE)k∈N is
u(f ′) = ∞Q
n=1
u(λn)f ′(fn),
f ′ ∈ Lp[0, 1]′.
Defining Lp([0, 1], E) ∶= Lp[0, 1]εE, we can read the corollary above as a state-
ment on series representations in the vector-valued version of Lp[0, 1]. However, in
many cases of spaces F(Ω) of scalar-valued functions there is a more natural way
to define the vector-valued version F(Ω, E) of F(Ω), see for example the space of
holomorphic functions from the introduction. If F(Ω)εE and F(Ω, E) are isomor-
more natural setting F(Ω, E) which motivates the following definition.
phic via the map S from the introduction, we can translate Theorem 3.1 to the
h2k+j(x) ∶=
1 ,(2j − 2)~2k+1 ≤ x < (2j − 1)~2k+1,
−1 ,(2j − 1)~2k+1 ≤ x < 2j~2k+1,
0 , else,
⎧⎪⎪⎪⎪⎨⎪⎪⎪⎪⎩
SERIES REPRESENTATIONS
7
3.4. Definition (ε-compatible). Let Ω be a set and E an lcHs. If F(Ω) ⊂ KΩ and
F(Ω, E) ⊂ EΩ are lcHs such that δx ∈ F(Ω)′ for all x ∈ Ω and
S∶ F(Ω)εE → F(Ω, E), u z→ [x ↦ u(δx)],
is an isomorphism, then we say that F(Ω) and F(Ω, E) are ε-compatible.
If we want to emphasise the dependence of S on F(Ω), we write SF (Ω). Several
sufficient conditions for S being an isomorphism are given in [15, 3.17 Theorem, p.
12].
3.5. Remark. Let Ω be a set and E an lcHs. If F(Ω) ⊂ KΩ and F(Ω, E) ⊂ EΩ are
lcHs such that δx ∈ F(Ω)′ for all x ∈ Ω and
S∶ F(Ω)εE → F(Ω, E), u z→ [x ↦ u(δx)],
is an isomorphism into, then we get by identification of isomorphic subspaces
F(Ω) ⊗ε E ⊂ F(Ω)εE ⊂ F(Ω, E)
and the embedding F(Ω) ⊗ E ↪ F(Ω, E) is given by f ⊗ e ↦ [x ↦ f(x)e].
Proof. The inclusions obviously hold and F(Ω)εE and F(Ω, E) induce the same
topology on F(Ω) ⊗ E. Further, we have
f ⊗ e
Θ
z→ [y ↦ y(f)e] S
z→ [x z→ [y ↦ y(f)e](δx)] = [x ↦ f(x)e].
3.6. Corollary. Let F(Ω) and F(Ω, E) be ε-compatible spaces, (fn)n∈N an equi-
continuous Schauder basis of F(Ω) with associated coefficient functionals (λK
n)n∈N
and let λE
n
∶ F(Ω, E) → E such that
for all n ∈ N. Set QE
Then the following holds.
u ∈ F(Ω)εE,
n(f) ∶= λE
a) The sequence (P E
n=1 QE
b) Each f ∈ F(Ω, E) has the series representation
n(S(u)) = u(λK
n),
λE
∶ F(Ω, E) → F(Ω, E), QE
∶= ∑k
k )k∈N given by P E
of F(Ω, E).
n
k
f = ∞Q
n(f)fn.
λE
c) F(Ω) ⊗ E is sequentially dense in F(Ω, E).
Proof. For each u ∈ F(Ω)εE and x ∈ Ω we note that
n=1
(cid:3)
(5)
n(f)fn for every n ∈ N.
n is a Schauder decomposition
(S ○ Pk)(u)(x) = S
= kQ
n=1
Qt
λK
kQ
n(δx) = u
kQ
n(S(u))fn(x) = (P E
n(⋅)fn(x) = kQ
○ S)(u)(x)
n=1
k
n=1
λE
which means that S ○ Pk = P E
○ S. This implies part a) and b) by Theorem 3.1 a)
since S is an isomorphism. Part c) is a direct consequence of Theorem 3.1 c) and
the isomorphy F(Ω)εE ≅ F(Ω, E).
(cid:3)
n=1
k
u(λK
n)fn(x)
In the preceding corollary we used the isomorphism S to obtain a Schauder
decomposition. On the other hand, if S is an isomorphism into which is often the
case (see [15, 3.9 Theorem, p. 9]), we can use a Schauder decomposition of F(Ω, E)
to prove the surjectivity of S.
8
K. KRUSE
3.7. Proposition. Let Ω be a set and E an lcHs. Let F(Ω) ⊂ KΩ and F(Ω, E) ⊂ EΩ
be lcHs such that δx ∈ F(Ω)′ for all x ∈ Ω and
is an isomorphism into. Let there be (fn)n∈N in F(Ω) and for every f ∈ F(Ω, E) a
sequence (λE
S∶ F(Ω)εE → F(Ω, E), u z→ [x ↦ u(δx)],
n(f))n∈N in E such that
n(f)fn,
λE
f = ∞Q
n=1
f ∈ F(Ω, E).
Then the following holds.
a) F(Ω) ⊗ E is sequentially dense in F(Ω, E).
b) If F(Ω) and E are sequentially complete, then
F(Ω, E) ≅ F(Ω)εE.
c) If F(Ω) and E are complete, then
F(Ω, E) ≅ F(Ω)εE ≅ F(Ω)⊗εE.
Proof. Let f ∈ F(Ω, E) and observe that
n(f)fn = kQ
k (f) ∶= kQ
P E
λE
n=1
n=1
fn ⊗ λE
n(f) ∈ F(Ω) ⊗ E
for every k ∈ N by Remark 3.5. Due to our assumption we have the convergence
k (f) → f in F(Ω, E). Thus F(Ω) ⊗ E is sequentially dense in F(Ω, E).
P E
Let us turn to part b). If F(Ω) and E are sequentially complete, then F(Ω)εE
is sequentially complete by [13, Satz 10.3, p. 234]. Since S is an isomorphism into
and
n(f))) = kQ
for all k, q ∈ N, k > q, we get that (Θ(∑k
n=1 fn ⊗ λE
F(Ω)εE and thus convergent. Hence we deduce that
S(Θ(
fn ⊗ λE
kQ
n=q
n=q
n(f)fn
λE
n(f)) is a Cauchy sequence in
S( lim
k→∞
Θ(
kQ
n=1
fn ⊗ λE
n(f))) = lim
k→∞
kQ
n=1(S ○ Θ)(fn ⊗ λE
n(f)) = ∞Q
n=1
λE
n(f)fn = f
which proves the surjectivity of S.
If F(Ω) and E are complete, then F(Ω)⊗εE is the closure of F(Ω) ⊗ε E in the
n(f))
complete space F(Ω)εE by [13, Satz 10.3, p. 234]. As limk→∞ Θ(∑k
n=1 fn ⊗ λE
is an element of the closure, we obtain part c).
(cid:3)
4. Applications
Sequence spaces
For our first application we recall the definition of some sequence spaces. A
matrix A ∶= (ak,j)k,j∈N of non-negative numbers is called Köthe matrix if it fulfils:
(1) ∀ k ∈ N ∃ j ∈ N ∶ ak,j > 0,
(2) ∀ k, j ∈ N ∶ ak,j ≤ ak,j+1.
For an lcHs E we define the Köthe space
λ∞(A, E) ∶= {x = (xk) ∈ E N ∀ j ∈ N, α ∈ A ∶
x j,α ∶= sup
k∈N
pα(xk)ak,j < ∞}
and the topological subspace
c0(A, E) ∶= {x = (xk) ∈ E N ∀ j ∈ N ∶
lim
k→∞
xkak,j = 0}.
SERIES REPRESENTATIONS
9
In particular, the space c0(N, E) of null-sequences in E is obtained as c0(N, E) =
c0(A, E) with ak,j ∶= 1 for all k, j ∈ N. The space of convergent sequences in E is
defined by
c(N, E) ∶= {x ∈ E N x = (xk) converges in E}
and equipped with the system of seminorms
x α ∶= sup
k∈N
pα(xk),
x ∈ c(N, E),
for α ∈ A. We define the spaces of E-valued rapidely decreasing sequences which
we need for our subsection on Fourier expansion by
pα(xk)(1 + k 2)j~2 < ∞}
s(Ω, E) ∶= {x = (xk) ∈ EΩ ∀ j ∈ N, α ∈ A ∶
with Ω = Nd, Nd
seminorms given by
0, Zd. Furthermore, we equip the space E N with the system of
x j,α ∶= sup
k∈Ω
xl,α ∶= sup
k∈N
pα(xk)χ{1,...,l}(k),
x = (xk) ∈ E N,
for l ∈ N and α ∈ A. For a non-empty set Ω we define for n ∈ Ω the n-th unit
function by
ϕn,Ω∶ Ω → K, ϕn,Ω(k) ∶= ⎧⎪⎪⎨⎪⎪⎩
1
0
, k = n,
, else,
and we simply write ϕn instead of ϕn,Ω if no confusion seems to be likely. Further,
we set ϕ∞∶ N → K, ϕ∞(k) ∶= 1, and x∞ ∶= δ∞(x) ∶= limk→∞ xk for x ∈ c(N, E).
For series representations of the elements in these sequence spaces we do not need
Corollary 3.6 due to the subsequent proposition but we can use the representation
to obtain the surjectivity of S for sequentially complete E.
4.1. Proposition. Let E be an lcHs and ℓV(Ω, E) one of the spaces c0(A, E), E N,
s(Nd, E), s(Nd
a) Then (∑n∈Ω, n ≤k δnϕn)k∈N is a Schauder decomposition of ℓV(Ω, E) and
0, E) or s(Zd, E).
x = Q
n=1(δn −δ∞)ϕn)k∈N is a Schauder decomposition of c(N, E)
b) Then (δ∞ϕ∞ +∑k
x ∈ ℓV(Ω, E).
xnϕn,
n∈Ω
and
x = x∞ϕ∞ +
∞Q
n=1(xn − x∞)ϕn,
x ∈ c(N, E).
Proof. Let us begin with a). For x = (xn) ∈ ℓV(Ω, E) let (P E
k ) be the sequence in
ℓV(Ω, E) given by P E
k is a continuous
k ≠ P E
for k ≠ j.
projection on ℓV(Ω, E), P E
Let ε > 0, α ∈ A and j ∈ N. For x ∈ c0(A, E) there is N0 ∈ N such that pα(xnan,j) < ε
for all n ≥ N0. Hence we have for x ∈ c0(A, E)
k (x) ∶= ∑ n ≤k xnϕn. It is easy to see that P E
min(k,j) for all k, j ∈ N and P E
j = P E
k P E
j
x − P E
k (x) j,α = sup
n>k
pα(xn)an,j ≤ sup
n≥N0
pα(xn)an,j ≤ ε
for all k ≥ N0. For x ∈ E N and l ∈ N we have
k (x)l,α = 0 < ε
x − P E
for all k ≥ l. For x ∈ s(Ω, E), Ω = Nd, Nd
that for all n ∈ Ω with n ≥ N1 we have
(1 + n 2)j~2
(1 + n 2)j = (1 + n 2)−j~2 < ε.
0, Zd, we notice that there is N1 ∈ N such
10
K. KRUSE
Thus we deduce for n ≥ N1
pα(xn)(1 + n 2)j~2 < εpα(xn)(1 + n 2)j ≤ ε x 2j,α
and hence
x − P E
k (x) j,α = sup
for all k ≥ N1. Therefore (P E
n >k
n ≥N1
pα(xn)(1 + n 2)j~2 ≤ ε x 2j,α
pα(xn)an,j ≤ sup
k (x)) converges to x in ℓV(Ω, E) and
x = lim
xnϕn.
P E
k (x) = Q
k→∞
n∈Ω
Now, we turn to b). For x = (xn) ∈ c(N, E) let (P E
k P E
k (x)) be the sequence in
n=1(xn − x∞)ϕn. Again, it is easy to see
j = P E
min(k,j) for all k, j ∈ N
for k ≠ j. Let ε > 0 and α ∈ A. Then there is N2 ∈ N such that
k (x) ∶= x∞ϕ∞ + ∑k
c(N, E) given by P E
that P E
and P E
pα(xn − x∞) < ε for all n ≥ N2. Thus we obtain
k is a continuous projection on c(N, E), P E
k ≠ P E
j
x − P E
k (x) α = sup
for all k ≥ N2 implying that (P E
k→∞ P E
x = lim
pα(xn − x∞) ≤ ε
pα(xn − x∞) ≤ sup
k (x)) converges to x in c(N, E) and
k (x) = x∞ϕ∞ +
∞Q
n=1(xn − x∞)ϕn.
n≥N2
n>k
(cid:3)
4.2. Theorem. Let E be a sequentially complete lcHs and ℓV(Ω, E) one of the
spaces c0(A, E), E N, s(Nd, E), s(Nd
0, E) or s(Zd, E). Then
(i) ℓV(Ω, E) ≅ ℓV(Ω)εE,
(ii) c(N, E) ≅ c(N)εE.
Proof. The map SℓV(Ω) is an isomorphism into by [15, 3.9 Theorem, p. 9] and [15,
4.13 Proposition, p. 23]. Considering c(N, E), we observe that for x ∈ c(N)
δn(x) = xn → x∞ = δ∞(x)
which implies the concergence δn → δ∞ in c(N)′
since c(N) is a Banach space. Hence we get
u(δn) = lim
u(δ∞) = lim
n→∞
n→∞
S(u)(n) = δ∞(S(u))
γ by the Banach-Steinhaus theorem
for every u ∈ c(N)εE which implies that Sc(N) is an isomorphism into by [15, 3.9
Theorem, p. 9]. From Proposition 4.1 and Proposition 3.7 we deduce our statement.
(cid:3)
Continuous and differentiable functions on a closed interval
We start with continuous functions on compact sets. We recall the following
definition from [24, p. 259]. A locally convex Hausdorff space is said to have the
metric convex compactness property (metric ccp) if the closure of the absolutely
convex hull of every metrisable compact set is compact. In particular, every se-
quentially complete space has metric ccp and this implication is strict (see [15, p.
3-4] and the references therein). Let E be an lcHs, Ω ⊂ Rd compact and denote
by C(Ω, E) ∶= C 0(Ω, E) the space of continuous functions from Ω to E. We equip
C(Ω, E) with the system of seminorms given by
f 0,α ∶= sup
x∈Ω
pα(f(x)),
f ∈ C(Ω, E),
SERIES REPRESENTATIONS
11
for α ∈ A. We want to apply our preceding results to intervals. Let −∞ < a < b < ∞
and T ∶= (tj)0≤j≤n be a partition of the interval [a, b], i.e. a = t0 < t1 < . . . < tn = b.
The hat functions hT
tj
∶[a, b] → R for the partition T are given by
hT
tj(x) ∶=
x−tj
tj −tj−1
tj+1−x
tj+1−tj
0
⎧⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎩
, tj−1 ≤ x ≤ tj,
, tj < x ≤ tj+1,
, else,
for 2 ≤ j ≤ n − 1 and
hT
a (x) ∶= ⎧⎪⎪⎨⎪⎪⎩
t1−x
t1−a
0
, a ≤ x ≤ t1,
, else,
hT
b (x) ∶= ⎧⎪⎪⎨⎪⎪⎩
x−tn−1
b−tn−1
0
, tn−1 ≤ x ≤ b,
, else.
Let T ∶= (tn)n∈N0 be a dense sequence in [a, b] with t0 = a, t1 = b and tn ≠ tm for
n ≠ m. For T n ∶= {t0, . . . , tn} there is a (unique) enumeration σ∶{0, . . . , n} → T n
such that Tn ∶= (tσ(j))0≤j≤n is a partition of [a, b]. The functions ϕT
1 ∶=
tn for n ≥ 2 are called Schauder hat functions for the sequence T
hT1
t1 and ϕT
and form a Schauder basis of C([a, b]) with associated coefficient functionals given
by λK
0 ∶= hT1
∶= hTn
t0 , ϕT
0 (f) ∶= f(t0), λK
1 (f) ∶= f(t1) and
n
λK
n+1(f) ∶= f(tn+1) −
nQ
k=0
k(f)ϕT
λK
k (tn+1),
f ∈ C([a, b]), n ≥ 1,
by [23, 2.3.5 Proposition, p. 29]. Looking at the coefficient functionals, we see that
n on
the right-hand sides even make sense if f ∈ C([a, b], E) and thus we define λE
C([a, b], E) for n ∈ N0 accordingly.
4.3. Theorem. Let E be an lcHs with metric ccp and T ∶= (tn)n∈N0 a dense sequence
in [a, b] with t0 = a, t1 = b and tn ≠ tm for n ≠ m. Then (∑k
n )k∈N0 is a
Schauder decomposition of C([a, b], E) and
n(f)ϕT
λE
n ,
f ∈ C([a, b], E).
f = ∞Q
n=0 λE
k ϕT
n=0
Proof. The spaces C([a, b]) and C([a, b], E) are ε-compatible by [15, 5.4 Example,
p. 25] if E has metric ccp. C([a, b]) is a Banach space and thus barrelled implying
n) is equicontinuous. We note that for all u ∈ C([a, b])εE
that its Schauder basis (ϕT
and x ∈ [a, b]
λE
n(S(u))(x) = u(δtn) = u(λK
n), n ∈ {0, 1},
and by induction
λE
n+1(S(u))(x) = u(δtn+1) −
k (S(u))ϕT
λE
k (tn+1) = u(δtn+1) −
nQ
k=0
= u(λK
n+1), n ≥ 1.
nQ
k=0
u(λK
k)ϕT
k (tn+1)
(cid:3)
Thus (5) is fulfilled proving our claim by Corollary 3.6.
tions vanishing at infinity given in [23, 2.7.1, p. 41-42] and [23, 2.7.4 Corollary, p.
If a = 0, b = 1 and T is the sequence of dyadic numbers given in [23, 2.1.1
n) is the so-called Faber-Schauder system. Using the
Definitions, p. 21], then (ϕT
Schauder basis and coefficient functionals of the space C0(R) of continuous func-
43] a corresponding, weaker result for the E-valued counterpart C0(R, E) holds as
well by a similar argumentation. Another corresponding result holds for the space
0,0([0, 1], E), 0 < γ < 1, of γ-Hölder continuous functions on [0, 1] with values
C[γ]
in E that vanish at zero and at infinity if one uses the Schauder basis and coef-
0,0([0, 1]) from [6, Theorem 2, p. 220] and [5, Theorem 3,
ficient functionals of C[γ]
p. 230]. The results are a bit weaker in both cases since [1, 2.4 Theorem (2), p.
12
K. KRUSE
138-139] and [15, 5.9 Example b), p. 28] only guarantee that SC0(R) and S
are isomorphisms if E is quasi-complete.
C[γ]
0,0([0,1])
Now, we turn to to spaces of continuously differentiable functions on an interval
(a, b) such that all derivatives can be continuously extended to the boundary. For
an lcHs E and k ∈ N the space Ck([a, b], E) is given by
Ck([a, b], E) ∶= {f ∈ Ck((a, b), E) (∂β)Ef continuously extendable on [a, b]
for all β ∈ N0, β ≤ k}
and equipped with the system of seminorms given by
β∈N0,β≤k
sup
x∈Ω
f α ∶=
pα(∂β)Ef(x),
for α ∈ A. From the Schauder hat functions (ϕT
n) for a dense sequence T ∶= (tn)n∈N0
in [a, b] with t0 = a, t1 = b and tn ≠ tm for n ≠ m and the associated coefficient
n we can easily get a Schauder basis for the space Ck([a, b]), k ∈ N, by
functionals λK
applying ∫ (⋅)
k-times to the series representation
f ∈ Ck([a, b], E),
a
f (k) = ∞Q
n=0
n(f (k))ϕT
λK
n ,
f ∈ Ck([a, b]),
n
t2
t1
x
tk−1
⋯
f T
S
µK
∶[a, b] → R and associated coefficient functionals µK
where we identified f (k) with its continuous extension on [a, b]. The resulting
∶ Ck([a, b]) →
Schauder basis f T
n
K, n ∈ N0, are
n (x) = 1
n!(x − a)n,
f T
n (x) =
n ≥ k,
S
S
for x ∈ [a, b] and f ∈ Ck([a, b]) (see e.g. [20, p. 586-587], [23, 2.3.7, p. 29]). Again,
the mapping rule for the coefficient functionals still makes sense if f ∈ Ck([a, b], E)
4.4. Theorem. Let E be an lcHs with metric ccp, k ∈ N, T ∶= (tn)n∈N0 a dense
sequence in [a, b] with t0 = a, t1 = b and tn ≠ tm for n ≠ m. Then (∑l
n )l∈N0
is a Schauder decomposition of Ck([a, b], E) and
n on Ck([a, b], E) for n ∈ N0 accordingly.
n(f) = f (n)(a),
n(f) = λK
ϕT
n−kdtdt1 . . . dtk−1, µK
0 ≤ n ≤ k − 1,
and so we define µE
n−k(f (k)),
n=0 µE
n f T
S
a
a
a
a
f = ∞Q
n=0
n(f)f T
µE
n ,
f ∈ Ck([a, b], E).
Proof. The spaces Ck([a, b]) and Ck([a, b], E) are ε-compatible by [15, 5.11 Exam-
ple, p. 29] if E has metric ccp. The Banach space Ck([a, b]) is barrelled giving the
for all u ∈ Ck([a, b])εE, β ∈ N0, β ≤ k, and x ∈ (a, b)
equicontinuity of its Schauder basis. Due to [15, 4.12 Proposition, p. 22] we have
(∂β)ES(u)(x) = u(δx ○(∂β)K).
[15, 4.8 Proposition, p. 22] in combination with [15, 4.9 Lemma, p. 22] applied to
Further, for every sequence (xn) in (a, b) converging to t ∈ {a, b} we obtain by
T ∶= (∂β)K
n→∞(∂β)ES(u)(xn) = u( lim
lim
n→∞
δxn
○(∂β)K).
From these observations we deduce that µE
holds. Therefore our statement is a consequence of Corollary 3.6.
n(S(u)) = u(µK
n) for all n ∈ N0, i.e. (5)
(cid:3)
SERIES REPRESENTATIONS
13
Holomorphic functions
In this short subsection we show how to get the result on power series expansion
of holomorphic functions from the introduction. Let E be an lcHs over C, z0 ∈ C,
r ∈ (0, ∞] and Ω ∶= Dr(z0) where Dr(z0) ⊂ C is an open disc around z0 with radius
r > 0. We equip O(Ω, E) with the system of seminorms given by
pα(f(z)),
for K ⊂ Ω compact and α ∈ A. We note that
f K,α ∶= sup
z∈K
f ∈ O(Ω, E),
O(Ω, E) = {f ∈ C∞(Ω, E) ∂
E
f = 0}
if E is locally complete where
E
∂
f(z) ∶= 1
2((∂e1)E + i(∂e2)E)f(z),
is the Cauchy-Riemann operator. Moreover, we set (∂ 0
f(z + h) − f(z)
C)Ef(z) ∶= lim
(∂ 1
C )Ef ∶= (∂ 1
C)E((∂n
h→0
h∈C,h≠0
and (∂n+1
that real and complex derivatives are related by
h
f ∈ C 1(Ω, E), z ∈ Ω,
C)Ef ∶= f ,
z ∈ Ω,
,
C)Ef) for n ∈ N0 and f ∈ O(Ω, E) (see (1)). We observe
z ∈ Ω,
C )Ef(z),
0.
(∂β)Ef(z) = iβ2(∂ β
for every f ∈ O(Ω, E) and β = (β1, β2) ∈ N2
4.5. Theorem. Let E be a locally complete lcHs over C, z0 ∈ C and r ∈ (0, ∞].
Then (∑k
(⋅ − z0)n)k∈N0 is a Schauder decomposition of O(Dr(z0), E)
(∂n
C)Ef(z0)
C )E f (z0)
and
(6)
(∂n
n=0
n!
(⋅ − z0)n,
f ∈ O(Dr(z0), E).
f = ∞Q
n=0
n!
Proof. The spaces O(Dr(z0)) and O(Dr(z0), E) are ε-compatible by [2, Theorem 9,
p. 232] if E is locally complete. Further, the Schauder basis ((⋅−z0)n) of O(Dr(z0))
is equicontinuous since the Fréchet space O(Dr(z0)) is barrelled. Due to [15, 4.12
Proposition, p. 22] and (6) we have for all u ∈ O(Dr(z0))εE and z ∈ Dr(z0)
C)ES(u)(z) = u(δz ○(∂n
(∂n
C)C)
which means that (5) is satisfied. Corollary 3.6 implies our statement.
(cid:3)
Fourier expansions
In this subsection we turn our attention to Fourier expansions in the space
of vector-valued rapidely decreasing functions and in the space of vector-valued
smooth, 2π-periodic functions. We start with the definition of the Pettis-integral
which we need to define the Fourier coefficients for vector-valued functions. For a
measure space (X, Σ, µ) and 1 ≤ p < ∞ let
Lp(X, µ) ∶= {f ∶ X → K measurable qp(f) ∶= S
f(x) pdµ(x) < ∞}
X
and define the quotient space of p-integrable functions by Lp(X, µ) ∶= Lp(X, µ)~{f ∈
Lp(X, µ) qp(f) = 0} which becomes a Banach space if it is equipped with the norm
fp ∶= qp(F)1~p, f = [F] ∈ Lp(X, µ). From now on we do not distinguish between
equivalence classes and their representants anymore.
14
K. KRUSE
For a measure space (X, Σ, µ) and f ∶ X → K we say that f is integrable on Λ ∈ Σ
and write f ∈ L1(Λ, µ) if χΛf ∈ L1(X, µ). Then we set
S
Λ
f(x)dµ(x) ∶= S
X
χΛ(x)f(x)dµ(x).
A function f ∶ X → E is called weakly (scalarly) measurable if the function e′ ○f ∶ X →
4.6. Definition (Pettis-integral). Let (X, Σ, µ) be a measure space and E an lcHs.
K, (e′ ○ f)(x) ∶= ⟨e′, f(x)⟩ ∶= e′(f(x)), is measurable for all e′ ∈ E ′. A weakly
measurable function is said to be weakly (scalarly) integrable if e′ ○ f ∈ L1(X, µ).
A function f ∶ X → E is called Pettis-integrable on Λ ∈ Σ if it is weakly integrable
on Λ and
∃ eΛ ∈ E ∀e′ ∈ E ′ ∶ ⟨e′, eΛ⟩ = S
⟨e′, f(x)⟩dµ(x).
Λ
In this case eΛ is unique due to E being Hausdorff and we set
f(x)dµ(x) ∶= eΛ.
S
Λ
If we consider the measure space (X, L(X), λ) of Lebesgue measurable sets for
X ⊂ Rd, we just write dx ∶= dλ(x).
4.7. Lemma. Let E be a locally complete lcHs, Ω ⊂ Rd open and f ∶ Ω → E. If f is
weakly C 1, i.e. e′ ○ f ∈ C 1(Ω) for every e′ ∈ E ′, then f is Pettis-integrable on every
compact subset of K ⊂ Ω with respect to any locally finite measure µ on Ω and
pαS
K
f(x)dµ(x) ≤ µ(K) sup
x∈K
pα(f(x)), α ∈ A.
Proof. Let K ⊂ Ω be compact and (Ω, Σ, µ) a measure space with locally finite
measure µ, i.e. Σ contains the Borel σ-algebra B(Ω) on Ω and for every x ∈ Ω
there is a neighbourhood Ux ⊂ Ω of x such that µ(Ux) < ∞. Since the map e′ ○
f is differentiable for every e′ ∈ E ′, thus Borel-measurable, and B(Ω) ⊂ Σ, it is
measurable. We deduce that e′ ○ f ∈ L1(K, µ) for every e′ ∈ E ′ because locally finite
measures are finite on compact sets. Hence the map
I ∶ E ′ → K, I(e′) ∶= S
K
⟨e′, f(x)⟩dµ(x)
is well-defined and linear. We estimate
I(e′) ≤ µ(K) sup
x∈f (K) e′(x) ≤ µ(K)
sup
x∈acx(f (K)) e′(x) ,
e′ ∈ E ′.
Due to f being weakly C 1 and [3, Proposition 2, p. 354] the absolutely convex set
acx(f(K)) is compact yielding I ∈ (E ′
which means that there is eK ∈ E such that
κ)′ ≅ E by the theorem of Mackey-Arens
⟨e′, f(x)⟩dµ(x),
⟨e′, eK⟩ = I(e′) = S
e′ ∈ E ′.
Therefore f is Pettis-integrable on K with respect to µ. For α ∈ A we set Bα ∶=
{x ∈ E pα(x) < 1} and observe that
α⟨e′,S
≤ µ(K) sup
f(x)dµ(x)⟩ = sup
αS
x∈K e′(f(x)) = µ(K) sup
e′(f(x))dµ(x)
pα(f(x))
f(x)dµ(x) = sup
pαS
e′ ∈B○
e′∈B○
sup
x∈K
K
K
K
K
e′∈B○
α
where we used [17, Proposition 22.14, p. 256] in the first and last equation to get
from pα to supe′ ∈B○
(cid:3)
and back.
α
SERIES REPRESENTATIONS
15
For an lcHs E we define the Schwartz space of E-valued rapidely decreasing
functions by
where
S(Rd, E) ∶= {f ∈ C∞(Rd, E) ∀ l ∈ N, α ∈ A ∶
f l,α < ∞}
f l,α ∶=
sup
x∈Rd
β∈Nd
0 , β ≤l
pα(∂β)Ef(x)(1 + x 2)l~2.
We recall the definition of the Hermite functions. For n ∈ N0 we set
hn∶ R → R, hn(x) ∶= (2nn!√π)−1~2(x −
dx)ne−x2~2 = (2nn!√π)−1~2Hn(x)e−x2~2,
d
with the Hermite polynomials Hn of degree n which can be computed recursively
by
H0(x) = 1
For n = (nk) ∈ Nd
and Hn+1(x) = 2xHn(x) − H ′
n(x),
0 we define the n-th Hermite function by
x ∈ R, n ∈ N0.
hnk(xk),
hn∶ Rd → R, hn(x) ∶= dM
Hnk(xk).
4.8. Proposition. Let E be a sequentially complete lcHs, f ∈ S(Rd, E) and n ∈ Nd
Then f hn is Pettis-integrable on Rd.
Proof. For k ∈ N we define the Pettis-integral
and Hn∶ Rd → R, Hn(x) ∶= dM
k=1
k=1
0.
ek ∶= S
[−k,k]d
f(x)hn(x)dx
which is a well-defined element of E by Lemma 4.7. We claim that (ek) is a
Cauchy sequence in E. First, we notice that there are j ∈ N and C > 0 such that
Hn(x) ≤ C(1 + x 2)j~2 for all x ∈ Rd as Hn is a product of polynomials in one
i=1 2ni ni!√π)−1~2 as well
variable. Now, let α ∈ A, k, m ∈ N, k > m, and set Cn ∶= (∏n
as Qk,m ∶= [−k, k]d ∖ [−m, m]d. We observe that ∏d
i=1 xi ≥ 1 for every x ∈ Qk,m
and
pα(ek − em)
= sup
e′ ∈B○
α e′(ek − em) = sup
e− x 2~2dx sup
e′ ∈B○
α
e′∈B○
≤ Cn S
Qk,m
= Cn S
Qk,m
e− x 2~2dx sup
x∈Rd
⟨e′, f(x)hn(x)⟩dx
Qk,m
sup
α S
x∈Rd e′(f(x)Hn(x))
pα(f(x)) Hn(x)
i=1
Qk,m
dM
≤ CnC f j,α S
= CnC f j,α(2 S
= 2dCnC f j,α(1 − e−k2~2)d − (1 − e−m2~2)d
xie− x 2~2dx
xe−x2~2dx)d −(2 S
[0,m]
[0,k]
xe−x2~2dx)d
proving our claim. Since E is sequentially complete, the limit y ∶= limk→∞ ek exists
in E. From e′ ○ f ∈ S(Rd) for every e′ ∈ E ′ and the dominated convergence theorem
e′ ∈ E ′,
⟨e′, y⟩ = lim
⟨e′, f(x)⟩hn(x)dx = S
k→∞⟨e′, ek⟩ = lim
⟨e′, f(x)hn(x)⟩dx,
k→∞ S
we deduce
[−k,k]d
Rd
16
K. KRUSE
which yields the Pettis-integrability of f hn on Rd with ∫Rd f(x)hn(x)dx = y.
f ∈ S(Rd, E) by
Due to the previous proposition we can define the n-th Fourier coefficient of
(cid:3)
f(n) ∶= S
Rd
f(x)hn(x)dx = S
f(x)hn(x)dx, n ∈ Nd
0,
Rd
if E is sequentially complete. We know that the map
is an isomorphism and its inverse is given by
F K
∶ S(Rd) → s(Nd
0), F K(f) ∶= f(n)n∈Nd
0) → S(Rd), (F K)−1(x) ∶= Q
0
,
n∈Nd
0
xnhn,
(F K)−1∶ s(Nd
(see e.g. [13, Satz 3.7, p. 66]). It is already known that S(Rd) and S(Rd, E) are
ε-compatible if E is quasi-complete by [21, Proposition 9, p. 108, Théorème 1, p.
111] (cf. [15, 3.22 Example, p. 14]). We improve this to sequentially complete E
and derive a Schauder decomposition of S(Rd, E) as well using our observations on
the sequence space s(Nd
0, E).
4.9. Theorem. Let E be a sequentially complete lcHs. Then the following holds.
a) If E is sequentially complete, then
F E ∶ S(Rd, E) → s(Nd
0, E), F E(f) ∶= f(n)n∈Nd
0
,
is an isomorphism, S(Rd, E) ≅ S(Rd)εE and
○(F Kε idE) ○ S −1
F E = Ss(Nd
0)
S(Rd).
b) (∑n∈Nd
0 , n ≤k
F E
n hn)k∈N is a Schauder decomposition of S(Rd, E) and
f = Q
n∈Nd
0
f(n)hn,
f ∈ S(Rd, E).
Proof. First, we show that the map F E is well-defined. Let f ∈ S(Rd, E). Then
e′ ○ f ∈ S(Rd) and
for every n ∈ Nd
which implies by [17, Mackey's theorem 23.15, p. 268] that F E(f) ∈ s(Nd
⟨e′, F E(f)n⟩ = ⟨e′, f(n)⟩ = e′ ○ f(n) = F K(e′ ○ f)n
0 and e′ ∈ E ′. Thus we have F K(e′ ○ f) ∈ s(Nd
0) for every e′ ∈ E ′
0, E) and
that F E is well-defined.
We notice that
s(Nd
s(Nd
0, E) ≅ s(Nd
0)εE ≅ S(Rd)εE ↪S(Rd, E)
where the first isomorphism is S −1
0 ) by Theorem 4.2 (i), the second is the map
(F K)−1ε idE and the third isomorphism into is the map SS(Rd) by [15, 3.9 Theorem,
p. 9] and [15, 4.12 Proposition, p. 22] since S(Rd) is barrelled. Next, we show
that SS(Rd) ○ (F K)−1ε idE ○ S −1
0) is surjective and the inverse of F E. We can
explicitely compute the composition of these maps. By the proof of Proposition 3.7
b) we get that the inverse of Ss(Nd
0, E) → s(Nd
Θ(ϕn ⊗ wn).
k→∞ Q
S −1
s(Nd
n∈Nd
0 , n ≤k
0)
s(Nd
∶ s(Nd
0 ) is given by
0)εE, w ↦ lim
0, E). Then for every x ∈ Rd and e′ ∈ E ′
0)(F K)−1t(δx) = Q
n∈Nd
0
S −1
s(Nd
Θ(ϕn ⊗ wn)(F K)−1t(δx)
Let w ∈ s(Nd
SERIES REPRESENTATIONS
17
= Q
n∈Nd
0
which gives
(F K)−1(ϕn)(x)wn = Q
n∈Nd
0
wnhn(x)
SS(Rd) ○(F K)−1ε idE ○ S −1
Let f ∈ S(Rd, E). Then w ∶= F E(f) ∈ s(Nd
e′( Q
n∈Nd
0
wnhn) = Q
wnhn(x).
n∈Nd
0
s(Nd
0)(w)(x) = Q
0, E) and
e′(f(n))hn = Q
= (F K)−1(e′ ○ f(n))n∈Nd
n∈Nd
n∈Nd
0
0
e′ ○ f(n)hn
0 = e′ ○ f
for all e′ ∈ E ′ resulting in
and
f = Q
n∈Nd
0
wnhn = Q
n∈Nd
0
F E(f)nhn
SS(Rd) ○(F K)−1ε idE ○ S −1
s(Nd
0 )(w)(x) = f(x)
for every x ∈ Rd. Thus we conclude
SS(Rd) ○(F K)−1ε idE ○ S −1
s(Nd
0) ○ F E(cid:6)(f) = f
s(Nd
yielding the surjectivity of the composition. Therefore SS(Rd) ○ (F K)−1ε idE ○
S −1
0 ) is an isomorphism with right inverse F E which implies that F E is its inverse.
In addition, the bijectivity of SS(Rd) ○(F K)−1ε idE○S −1
0) implies that
SS(Rd) ○ (F K)−1ε idE is bijective and thus SS(Rd) is surjective. Hence SS(Rd) is
0, E) via F E and
The rest of part b) follows from the isomorphy S(Rd, E) ≅ s(Nd
also an isomorphism completing the proof of part a).
0 ) and S −1
s(Nd
s(Nd
Proposition 4.1 a).
(cid:3)
Our last example of this subsection is devoted to Fourier expansions of vector-
convex Hausdorff E with the system of seminorms generated by
valued 2π-periodic smooth functions. We equip the space C∞(Rd, E) for locally
f ∈ C∞(Rd, E),
f K,l,α ∶= sup
pα((∂β)Ef(x))χK(x) = sup
pα((∂β)Ef(x)),
x∈Ω
0 , β ≤l
β∈Nd
x∈K
0 , β ≤l
β∈Nd
for K ⊂ Rd compact, l ∈ N0 and α ∈ A. By C∞
2π(Rd, E) we denote the topological
subspace of C∞(Rd, E) consisting of the functions which are 2π-periodic in each
variable. Due to Lemma 4.7 we are able to define the n-th Fourier coefficient of
f ∈ C∞
2π(Rd, E) by
f(n) ∶= (2π)−d S
[−π,π]d
f(x)e−i⟨n,x⟩dx, n ∈ Zd,
where ⟨⋅, ⋅⟩ is the usual scalar product on Rd, if E is locally complete.
4.10. Theorem. Let E be a locally complete lcHs over C. Then a Schauder decom-
position of C∞
2π(Rd, E) is given by (∑n∈Zd, n ≤k f(n)ei⟨n,⋅⟩)k∈N and
f = Q
n∈Zd
f(n)ei⟨n,⋅⟩,
f ∈ C∞
2π(Rd, E).
18
K. KRUSE
2π(Rd)′ and thus
2π(Rd)(u)(x) − SC∞
SC∞
2π(Rd, E) are ε-compatible. It follows
Proof. First, we prove that C∞
from [2, p. 228] (cf. [15, 3.23 Example, p. 15]) that SC∞(Rd)∶ C∞(Rd)εE → C∞(Rd, E)
is an isomorphism. Furthermore, for each x ∈ Rd and 1 ≤ n ≤ d we have δx = δx+2πen
in C∞
2π(Rd) and C∞
2π (Rd)(u)(x + 2πen) = u(δx − δx+2πen) = 0,
2π(Rd)εE,
2π(Rd)(u) is 2π-periodic in each variable. In addition, we observe
2π(Rd, E). An
2π(Rd)εE →
2π(Rd) is barrelled since it is a Fréchet space and thus its Schauder
implying that SC∞
that e′ ○ f is 2π-periodic in each variable for all e′ ∈ E ′ and f ∈ C∞
application of [15, 3.26 Proposition (i), p. 17] yields that SC∞
2π(Rd, E) is an isomorphism, i.e. C∞
2π(Rd, E) are ε-compatible.
C∞
basis (ei⟨n,⋅⟩) is equicontinuous. By [15, 3.17 Theorem, p. 12] the inverse of SC∞
2π(Rd) and C∞
The space C∞
2π(Rd)∶ C∞
u ∈ C∞
2π (Rd)
is given by
where J ∶ E → E ′⋆ is the canonical injection in the algebraic dual E ′⋆ of E ′ and
Rt∶ C∞
2π(Rd, E) → C∞
2π(Rd)εE, f ↦ J −1
○ Rt
f ,
Rt
f
2π(Rd)′ → E ′⋆, y z→ e′ ↦ y(e′ ○ f)(cid:6),
∶ C∞
2π(Rd)εE Then f ∶= SC∞
2π(Rd)(u) ∈ C∞
2π(Rd, E) and
for f ∈ C∞
from the Pettis-integrability of f we obtain
2π(Rd, E). Let u ∈ C∞
f(FC
Rt
n(e′ ○ f) = (2π)−d S
n)(e′) = FC
⟨e′, f(x)e−i⟨n,x⟩⟩dx
[−π,π]d
= ⟨e′,(2π)−d S
[−π,π]d
f(x)e−i⟨n,x⟩dx⟩ = ⟨e′, FE
n(f)⟩,
e′ ∈ E ′,
for all n ∈ Zd which results in
2π (Rd)(f)(FC
n) = S −1
u(FC
C∞
n) = J −1Rt
f(FC
n) = FE
n(f) = FE
n(SC∞
2π(Rd)(u))
and thus shows the validity of (5). Now, our statement follows from Corollary
3.6.
(cid:3)
Considering the coefficients in the series expansion above, we know that the map
FC
∶ C∞
2π(Rd) → s(Zd), FC(f) ∶= f(n)n∈Zd ,
is an isomorphism (see e.g. [13, Satz 1.7, p. 18]). Thus we have the following relation
if E is a locally complete Hausdorff space over C
C∞
2π(Rd, E) ≅ C∞
2π(Rd)εE ≅ s(Zd)εE ↪s(Zd, E)
where the first isomorphism is S −1
2π(Rd), the second is the map FCε idE and the third
C∞
isomorphism into is the map Ss(Zd) by [15, 3.9 Theorem, p. 9]. We can explicitely
compute the composition of these maps. With the notation from the proof above
we have for every f ∈ C∞
Ss(Zd) ○(FCε idE) ○ S −1
= Ss(Zd)FCε idES −1
= S −1
Thus the map
2π(Rd)(f)(FC)t(δn) = S −1
2π(Rd, E) and n ∈ Zd
2π(Rd)(f)(n)
2π (Rd)(f)(n) = Ss(Zd)S −1
n) = FE
2π(Rd)(f) ○(FC)t(n)
n(f) = f(n).
2π(Rd, E) → s(Zd, E), FE(f) ∶= f(n)n∈Zd ,
2π (Rd)(f)(FC
FE ∶ C∞
is well-defined and an isomorphism into if E is locally complete. If E is sequentially
complete, it is even an isomorphism to the whole space s(Zd, E) because Ss(Zd) is
surjective to the whole space then by Theorem 4.2 (i). Hence we have:
C∞
C∞
C∞
C∞
C∞
SERIES REPRESENTATIONS
19
4.11. Theorem. If E is a sequentially complete lcHs over C, then FE ∶ C∞
2π(Rd, E) →
s(Zd, E), FE(f) ∶= f(n)n∈Zd , is an isomorphism and
FE = Ss(Zd) ○(FCε idE) ○ S −1
2π(Rd).
C∞
For quasi-complete E Theorem 4.10 and Theorem 4.11 are already known by
[13, Satz 10.8, p. 239].
Acknowledgement. The author is deeply grateful to José Bonet for many helpful
suggestions honing the present paper. The main Theorem 3.1 is essentially due to
him improving a previous version of the author which became Corollary 3.6.
References
[1] K.-D. Bierstedt. Gewichtete Räume stetiger vektorwertiger Funktionen und
das injektive Tensorprodukt. II. J. Reine Angew. Math., 260:133 -- 146, 1973.
[2] J. Bonet, L. Frerick, and E. Jordá. Extension of vector-valued holomorphic
and harmonic functions. Studia Math., 183(3):225 -- 248, 2007.
[3] J. Bonet, E. Jordá, and M. Maestre. Vector-valued meromorphic functions.
Arch. Math. (Basel), 79(5):353 -- 359, 2002.
[4] J. Bonet and W. J. Ricker. Schauder decompositions and the Grothendieck and
Dunford-Pettis properties in Köthe echelon spaces of infinite order. Positivity,
11(1):77 -- 93, 2007.
[5] Z. Ciesielski. On Haar functions and on the Schauder basis of the space C⟨0,1⟩.
Bull. Acad. Pol. Serie des Sc. Math., Ast. et Ph., 7(4):227 -- 232, 1959.
[6] Z. Ciesielski. On the isomorphisms of the spaces Hα and m. Bull. Acad. Pol.
Serie des Sc. Math., Ast. et Ph., 8(4):217 -- 222, 1960.
[7] A. Defant and K. Floret. Tensor norms and operator ideals. Math. Stud. 176.
North-Holland, Amsterdam, 1993.
[8] K. Floret and J. Wloka. Einführung in die Theorie der lokalkonvexen Räume.
Lecture Notes in Math. 56. Springer, Berlin, 1968.
[9] K.-G. Grosse-Erdmann. The Borel-Okada Theorem Revisited. Habilitation.
Fernuniversität Hagen, Hagen, 1992.
[10] A. Grothendieck. Produits tensoriels topologiques et espaces nucléaires. Mem.
Amer. Math. Soc. 16. AMS, Providence, 4th edition, 1966.
[11] H. Jarchow. Locally Convex Spaces. Math. Leitfäden. Teubner, Stuttgart,
1981.
[12] H. F. Joiner II. Tensor product of Schauder bases. Math. Ann., 185(4):279 -- 284,
1970.
[13] W. Kaballo. Aufbaukurs Funktionalanalysis und Operatortheorie. Springer,
Berlin, 2014.
[14] N. J. Kalton. Schauder decompositions in locally convex spaces. Math. Proc.
Cambridge Philos. Soc., 68(2):377 -- 392, 1970.
[15] K. Kruse. Weighted vector-valued functions and the ε-product, 2017. arxiv
preprint https://arxiv.org/abs/1712.01613.
[16] H. P. Lotz. Uniform convergence of operators on L∞ and similar spaces. Math.
Z., 190(2):207 -- 220, 1985.
[17] R. Meise and D. Vogt. Introduction to Functional Analysis. Clarendon Press,
Oxford, 1997.
[18] R. A. Ryan.
Introduction to Tensor Products of Banach Spaces. Springer
Monogr. Math. Springer, Berlin, 2002.
[19] J. Schauder. Eine Eigenschaft des Haarschen Orthogonalsystems. Math. Z.,
28(1):317 -- 320, 1928.
[20] S. Schonefeld. Schauder bases in spaces of differentiable functions. Bull. Amer.
Math. Soc., 75(3):586 -- 590, 1969.
20
K. KRUSE
[21] L. Schwartz. Espaces de fonctions différentiables à valeurs vectorielles. J.
Analyse Math., 4:88 -- 148, 1955.
[22] L. Schwartz. Théorie des distributions à valeurs vectorielles. I. Ann. Inst.
Fourier (Grenoble), 7:1 -- 142, 1957.
[23] Z. Semadeni. Schauder bases in Banach spaces of continuous functions. Lecture
Notes in Math. 918. Springer, Berlin, 1982.
[24] J. Voigt. On the convex compactness property for the strong operator topology.
Note Math., XII:259 -- 269, 1992.
TU Hamburg, Institut für Mathematik, Am Schwarzenberg-Campus 3, Gebäude E,
21073 Hamburg, Germany
E-mail address: [email protected]
|
1111.4568 | 1 | 1111 | 2011-11-19T16:07:55 | Schr\"{o}dinger operators with boundary singularities: Hardy inequality, Pohozaev identity and controllability results | [
"math.FA"
] | The aim of this paper is two folded. Firstly, we study the validity of the Pohozaev-type identity for the Schr\"{o}dinger operator $$A_\la:=-\D -\frac{\la}{|x|^2}, \q \la\in \rr,$$ in the situation where the origin is located on the boundary of a smooth domain $\Omega\subset \rr^N$, $N\geq 1$. The problem we address is very much related to optimal Hardy-Poincar\'{e} inequality with boundary singularities which has been investigated in the recent past in various papers. In view of that, the proper functional framework is described and explained.
Secondly, we apply the Pohozaev identity not only to study semi-linear elliptic equations but also to derive the method of multipliers in order to study the exact boundary controllability of the wave and Schr\"{o}dinger equations corresponding to the singular operator $A_\la$. In particular, this complements and extends well known results by Vanconstenoble and Zuazua [34], who discussed the same issue in the case of interior singularity. | math.FA | math |
Schrodinger operators with boundary singularities: Hardy
inequality, Pohozaev identity and controllability results
Cristian Cazacu ∗†
September 24, 2018
Abstract
The aim of this paper is two folded. Firstly, we study the validity of the Pohozaev-type
identity for the Schrodinger operator
Aλ := −∆ −
λ
x2 ,
λ ∈ R,
in the situation where the origin is located on the boundary of a smooth domain Ω ⊂ RN ,
N ≥ 1. The problem we address is very much related to optimal Hardy-Poincar´e inequality
with boundary singularities which has been investigated in the recent past in various papers.
In view of that, the proper functional framework is described and explained.
Secondly, we apply the Pohozaev identity not only to study semi-linear elliptic equations
but also to derive the method of multipliers in order to study the exact boundary controlla-
bility of the wave and Schrodinger equations corresponding to the singular operator Aλ. In
particular, this complements and extends well known results by Vanconstenoble and Zuazua
[34], who discussed the same issue in the case of interior singularity.
Contents
1 Introduction
2 Pohozaev identity for Aλ
2.1 The case C1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 The meaning of Hλ-norm . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.3 Proofs of Theorems 2.1, 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.4 Proofs of useful lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Applications to semi-linear equations . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Brief presentation of the case C2 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1
2.3.2
Functional framework via Hardy inequality . . . . . . . . . . . . . . . . . .
The meaning of the Hλ-norm . . . . . . . . . . . . . . . . . . . . . . . . . .
2
6
6
7
7
9
11
14
17
17
18
∗BCAM - Basque Center for Applied Mathematics, Bizkaia Technology Park 500, 48160, Derio, Basque Country,
Spain
†Departamento de Matem´aticas, Universidad Aut´onoma de Madrid, Madrid 28049, Spain.
1
3 Applications to Controllability
3.1 The wave equation. Case C1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Well-posedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.2 Controllability and main results . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.3 Proofs of main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 The Schrodinger equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Open problems
5 Appendix: sharp bounds for x · ∇v(t)L2(Ω)
1
Introduction
19
19
19
20
22
24
26
26
In this paper we are dealing with the Schrodinger operator Aλ := −∆ − λ/x2, λ ∈ R, acting in
a domain where the potential 1/x2 is singular at the boundary. Our main goal consists to study
the control properties of the corresponding wave and Schrodinger equations. Moreover, we are
aimed to give necessary and sufficient conditions for the existence of non-trivial solutions to semi-
linear elliptic equations associated to Aλ. Operators like Aλ may arise in molecular physics [26],
quantum cosmology [5], combustion models [20] but also in the linearization of critical nonlinear
PDE's playing a crucial role in the asymptotic behaviour of branches of solutions in bifurcation
problems (e.g.
[8], [30]). From the mathematical view point they are interesting due to their
criticality since they are homogeneous of degree -2.
The qualitative properties of evolution problems involving the operator Aλ require either posi-
tivity or coercivity of Aλ in the sense of quadratic forms in L2. Roughly speaking, this is equivalent
to make use of Hardy-type inequalities. There is a large literature concerning the study of such
inequalities, especially in the context of interior singularities (e.g. see [36], [2], [17] and references
therein). The classical Hardy inequality is stated as follows. Assume Ω is a smooth bounded
domain in RN , N ≥ 3, containing the origin, i.e., 0 ∈ Ω; then it follows (see [22])
ZΩ ∇u2dx −
(N − 2)2
4
ZΩ
u2
x2 dx > 0,
∀u ∈ H 1
0 (Ω),
(1.1)
and the constant (N − 2)2/4 is optimal and not attained. We remind that the optimal Hardy
constant is defined by the quotient
µ(Ω) := inf
u∈C∞
0 (Ω)(cid:16)ZΩ ∇u2dx(cid:14)ZΩ
u2/x2dx(cid:17).
In this paper, we consider Ω to be a smooth subset of RN , N ≥ 1, with the origin x = 0 placed
on its boundary Γ. Hardy inequalities with an isolated singularity on the boundary were less
investigated so far. However, in the recent past some substantial work has been developed in that
direction.
It has been proved that, the best constants depends both on the local geometry near the origin
and the entire shape of the domain.
More precisely, starting with the work by Filippas, Tertikas and Tidblom [19], and continuing
with [9], [14], [15], it has been proved that, whenever Ω is a smooth domain with the origin located
on the boundary, there exists a positive constant r0 = r0(Ω, N ) > 0 such that
µ(Ω ∩ Br0(0)) =
N 2
4
.
2
(1.2)
+ which is given by the set RN
where Br0(0) denotes the N -d ball of radius r0 centered at origin. Next we recall the definition
of the upper half space RN
+ := {x = (x1, . . . , xN −1, xN ) = (x′, xN ) ∈
RN −1 × R xN > 0}. In addition, if Ω ⊂ RN
+ , N ≥ 1, the new Hardy inequality
N 2
4 ZΩ
x2 dx ∀ u ∈ H 1
ZΩ ∇u2dx ≥
u2
0 (Ω).
(1.3)
holds true and the constant N 2/4 is optimal i.e. µ(Ω) = N 2/4.
Otherwise, if Ω is a smooth domain which, up to a rotation, is not supported in RN
+ , the
constant N 2/4 is optimal, up to lower order terms in L2(Ω)-norm as shown later in inequality
(1.7). In general µ(Ω) = N 2/4 is not true for any smooth bounded domain Ω containing the origin
on the boundary (e.g. [14]).
Without losing the generality, since the operator Aλ is invariant under rotations, next we
consider Ω such that
where ν stands for the outward normal vector to Γ. Moreover, since optimal inequalities have
been obtained regardless the shape of Ω, throughout the paper we discuss two main situations of
geometries motivated by the remarks above.
x · ν = O(x2), on Γ,
(1.4)
C1. Ω is a smooth domain satisfying (1.4) and xN > 0 holds for all x ∈ Ω (i.e. Ω ⊂ RN
+ ).
C2. Ω is a smooth domain satisfying (1.4) such that xN changes sign in Ω (Ω 6⊂ RN
+ ).
Next we need to introduce the constant
RΩ = sup
x∈Ωx.
(1.5)
The following optimal Hardy-Poincar´e inequalities are valid for each one of the cases above.
If Ω fulfills the case C1, then (e.g. [9]) it holds that
∀ u ∈ C∞
0 (Ω),
ZΩ ∇u2dx ≥
N 2
4 ZΩ
u2
x2 dx +
1
4ZΩ
u2
x2 log2(RΩ/x)
dx,
(1.6)
and N 2/4 is the sharp constant.
If Ω satisfies the case C2 then (e.g.
C3 = C3(Ω, N ) > 0 such that for any u ∈ C∞
N 2
4 ZΩ
u2dx +ZΩ ∇u2dx ≥
C2ZΩ
[14]) there exist two constants C2 = C2(Ω) ∈ R and
0 (Ω) it holds
u2
x2 dx + C3ZΩ
u2
x2 log2(RΩ/x)
dx.
(1.7)
In view of these, let us now describe the content of the paper.
In Section 2, we firstly introduce the functional framework induced by the above Hardy in-
equalities. We refer to the Hilbert space Hλ defined in Subsection 2.1. Then we check the validity
of the Pohozaev identity for the Schrodinger operator Aλ in this functional setting as follows. The
domain of Aλ is defined by
D(Aλ) := {u ∈ Hλ Aλu ∈ L2(Ω)},
(1.8)
and it holds
1
2ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
dσ = −ZΩ
(x · ∇u)Aλudx −
N − 2
2
u2
Hλ,
∀ u ∈ D(Aλ),
(1.9)
3
where · Hλ denotes the norm associated to Hλ and We refer to Theorems 2.1, 2.2 for a com-
plete statement of this result. For the sake of clarity, we will mainly discuss the case C1 above.
Nevertheless, similar results could be also extended to the case C2 in a weaker functional setting
due to weaker Hardy inequalities (see Subsection 2.3).
Formally, identity (1.9) can be obtained by direct integrations. However, this is not rigorously
allowed because the lack of regularity of Aλ at the origin, otherwise the potential 1/x2 is bounded
and the standard elliptic regularity applies. In addition, we need to justify the integrability of the
boundary term in (1.9) which is no more obvious since the singularity is located on the boundary
and standard trace regularity does not applies. As we mentioned before, we give a rigorous
justification of these facts in Theorems 2.1, 2.2.
Pohozaev type identities arise in many applications and mostly when studying non-linear equa-
tions (see [13], [21], [11] and references therein).
In Section 2.2, we apply Theorem 2.2 to characterize the existence of non-trivial solutions to
a semi-linear singular elliptic PDE in star-shaped domains. We refer mainly to Theorem 2.3.
In Section 3 we present some applications of Theorem 2.2 in Controllability of conservative
systems like wave and Schrodinger equations, for which the multiplier method plays a crucial role.
In the last few decades, most of the studies in Controllability Theory and its applications to
evolution PDEs, have applied methods like Hilbert Uniqueness Method (HUM) introduced by J. L.
Lions in [27], Carleman estimates developed by Fursikov and Imanuvilov [18], microlocal analysis
due to Bardos, Lebeau and Rauch ([4], [3]), but also multiplier techniques with the pioneering
papers by Komornik and Zuazua ([24], [25], [37]). In particular, the controllability properties and
stabilization of the heat like equation corresponding to Aλ have been analyzed in [33], [12], [32] in
the case of interior singularity using tools like multiplier techniques and Carleman estimates.
Now, let us detail the problem we are interested in Section 3. For N ≥ 1 we consider a bounded
smooth domain Ω ⊂ RN where Γ denotes its boundary. Moreover, we state by Γ0 a non-empty
part of the set Γ that will be precise later.
Next we consider the Wave-like process
x2 = 0,
utt − ∆u − λ u
u(t, x) = h(t, x),
u(t, x) = 0,
u(0, x) = u0(x),
ut(0, x) = u1(x),
(t, x) ∈ (0, T ) × Ω,
(t, x) ∈ (0, T ) × Γ0,
(t, x) ∈ (0, T ) × (Γ \ Γ0),
x ∈ Ω,
x ∈ Ω.
(1.10)
′
To make the problem under consideration precise we say that the system (1.10) is exactly
λ and any target
λ , there exists a control h ∈ L2((0, T ) × Γ0) such that the solution of (1.10)
controllable from Γ0, in time T , if for any initial data (u0, u1) ∈ L2(Ω) × H
(u0, u1) ∈ L2(Ω) × H
satisfies:
′
(ut(T, x), u(T, x)) = (u1(x), u0(x))
for all x ∈ Ω.
This issue was analyzed by Vancostenoble and Zuazua [34] under the assumption that the sin-
gularity x = 0 is located in the interior of Ω. They proved well-posedness and exact controllability
of the system (1.10) for any λ ≤ λ⋆ := (N − 2)2/4 from the boundary observability region Γ0
described by
(1.11)
Γ0 := {x ∈ Γ x · ν ≥ 0}.
Roughly speaking, the authors showed in [34] that the parameter λ⋆ is critical when asking the
well-posedness and control properties of (1.10), and the results are very much related to the best
constant in the Hardy inequality with interior singularity.
In Section 3, we address the same controllability question in the case of boundary singularity.
Our main result asserts that for the same geometrical setup (1.11), we can increase the range of
4
values λ (from λ⋆ to λ(N ) := N 2/4) for which the exact boundary controllability of system (1.10)
holds. This is due to the new Hardy inequalities above.
By now classical HUM, the Controllability of system (1.10) is equivalent to so-called Observ-
ability Inequality for the adjoint system,
x2 = 0,
vtt − ∆v − λ v
v(t, x) = 0,
v(0, x) = v0(x),
vt(0, x) = v1(x),
(t, x) ∈ (0, T ) × Ω,
(t, x) ∈ (0, T ) × Γ,
x ∈ Ω,
x ∈ Ω,
(1.12)
which formally states that for any λ ≤ λ(N ) and T > 0 large enough there exists a constant
CT > 0 such that
CT(cid:16)v12
L2(Ω) +ZΩh∇v0(x)2 − λ
v2
0(x)
x2 idx(cid:17) ≤Z T
0 ZΓ0
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
dσdt,
(1.13)
holds true for v solution of (3.8). The main tool to prove (1.13) relies on the multiplier method
and compactness-uniqueness argument [27]. In view of that, Pohozaev identity provides a direct
tool to show that the solution of system (3.8) satisfies the multiplier identity which formally is
given by
1
2Z T
0 ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
dσdt =
T
2
(v12
L2(Ω) + v02
Hλ ) +ZΩ
vt(cid:0)x · ∇v +
producing a "Hidden regularity" efect for the normal derivative. We refer to Theorem 3.2 for a
rigorous statement. As a consequence, the solution of system (3.8) verifies the reverse Observability
inequality. Then identity (1.14) together with the sharp-Hardy inequality stated in Theorem 1.1
lead to Observability inequality (1.13), fact emphasized in Theorem 3.3.
N − 1
2
T
0
dx,
(1.14)
v(cid:1)(cid:12)(cid:12)(cid:12)
Theorem 1.1. Assume Ω satisfies one of the cases C1-C2 . Then, there exists a constant C =
C(Ω) ∈ R such that
ZΩ x2∇v2dx ≤ R2
ΩhZΩ ∇v2dx −
N 2
4 ZΩ
v2
x2 dxi + CZΩ
The proof of Theorem 1.1 is given in the Appendix.
v2dx ∀v ∈ C∞
0 (Ω).
(1.15)
Remark 1.1. The result of Theorem 1.1, and precisely the constant R2
Ω which appears in inequality
(1.15), helps to obtain the control time T > T0 = 2RΩ in (1.13), which is sharp from the Geometric
Control Condition considerations, see [4].
Although Theorem 1.1 is sharp for our applications to controllability, it is worth mentioning
that we are able to obtain a more general result as follows.
Theorem 1.2. Assume Ω satisfies one of the cases C1-C2. Let be ε > 0. Then, there exists a
constant Cε = C(Ω, ε) ∈ R such that
ZΩ xε∇v2dx ≤ Rε
ΩhZΩ ∇v2dx −
N 2
4 ZΩ
v2
x2 dxi + CεZΩ
v2dx ∀v ∈ C∞
0 (Ω).
(1.16)
The proof of Theorem 1.2 is omitted since it applies the same steps in the proof of Theorem 1.1.
Finally in Section 3.2 we will consider the Schrodinger-like process
x2 = 0,
iut − ∆u − λ u
u(t, x) = h(t, x),
u(t, x) = 0,
u(0, x) = u0(x),
(t, x) ∈ (0, T ) × Ω,
(t, x) ∈ (0, T ) × Γ0,
(t, x) ∈ (0, T ) × (Γ \ Γ0),
x ∈ Ω,
5
(1.17)
where the singularity is located on the boundary, and we briefly discuss the well-posedness and
controllability properties. In Section 4 we treat with some open related problems.
The main results of this paper have been announced in a short presentation in [10].
2 Pohozaev identity for Aλ
In this Section we rigorously justify the Pohozaev-type identity associated to Aλ. We will discuss
in a detail manner the case C1. The details of the case C2 are let to the reader. In the latter case
we only state the corresponding functional framework, see Subsection 2.3.
2.1 The case C1
Firstly, we introduce the functional framework which is used throughout the paper and we discuss
some of its properties.
Assume Ω ⊂ RN , N ≥ 1 is a smooth domain which satisfies the case C1 and fix λ ≤ λ(N ).
Thanks to inequality (1.6), we consider the Hardy functional
Bλ[u] =ZΩh∇u2 − λ
u2
x2idx,
(2.1)
(2.2)
which is positive and finite for all u ∈ C∞
Hλ, defined by the completion of C∞
0 (Ω) functions in the norm
u2
u ∈ C∞
Hλ = Bλ[u],
0 (Ω).
0 (Ω). For any λ ≤ λ(N ), Bλ[u] induces a Hilbert space
We point out that the space Hλ was firstly introduced by Vazquez and Zuazua [36] in the case of
interior singularity. As emphasize above, it may be extended to the case of boundary singularity.
In the subcritical case λ < λ(N ), it holds that H 1
0 (Ω) = Hλ, according to the estimates
(cid:0)1 − λ/λ(N )(cid:1)uH1
which ensure the equivalence of the norms.
0 (Ω) ≤ u2
Hλ ≤ u2
0 (Ω),
H1
∀ u ∈ C∞
0 (Ω),
The critical space Hλ(N ) turns to be slightly larger than H 1
0 (Ω). Remark that Bλ(N )[u] is finite
for any u ∈ H 1
0 (Ω), but it makes sense as an improper integral approaching the singular pole x = 0
(see the right hand side of (2.3)) for more general distributions. As happens in the case of interior
singularity (see [35]), in general the meaning of uHλ(N ) does not coincide with the improper
integral of Bλ(N )[u]. Following [35], we can construct a counterexample even in the case when the
singularity is located on the border. Indeed, we fix Ω := {x ∈ RN
x′2 + (xN − 1)2 ≤ 1} and
we consider the distribution e1 = xNx−N/2J(z0,1x) where z0,1 is the first positive zero of the
Bessel function J0. We observe that Bλ(N )[u] is finite as an improper integral approaching the
origin. On the other hand, computing we remark that
+ :
e1 − φHλ(N ) ≥ C0 > 0,
∀φ ∈ C∞
0 (Ω),
for some positive universal constant C0 > 0. This is in contradiction with the definition of
Hλ(N ) which allows the existence of a sequence φn ∈ C∞
0 (Ω) converging to e1 in Hλ(N )-norm
! Thererefore, the assumption of considering the definition of the Hλ(N )-norm as an improper
integral of Bλ(N ) is false. In other words, there are distributions u ∈ Hλ(N ) for which
u2
Hλ(N ) 6= lim
ε→0Zx≥εh∇u2 − λ(N )
u2
x2idx.
(2.3)
Next we propose an equivalent norm on Hλ, λ ≤ λ(N ), which overcomes the anomalous behavior
in (2.3) and describes perfectly the meaning of the Hλ-norm.
6
2.1.1 The meaning of Hλ-norm
For reasonable considerations that will be precise in (2.5), we introduce the functional
dx + (λ(N ) − λ)ZΩ
which is positive and finite for any u ∈ C∞
N−th canonical vector of RN . Next, we observe that, for any λ ≤ λ(N ),
x2 u −
0 (Ω) and λ ≤ λ(N ). Here, we have denoted by eN the
Bλ,1[u] =ZΩ(cid:12)(cid:12)(cid:12)∇u +
x2 dx.
(2.4)
2
u(cid:12)(cid:12)(cid:12)
eN
xN
x
N
2
u2
Bλ[u] = Bλ,1[u],
∀u ∈ C∞
0 (Ω).
(2.5)
Besides, notice that both Bλ,1[u] and Bλ[u] are norms in Hλ and they coincide on C∞
to definition (2.2) of Hλ, we conclude that the Hλ could be define as the closure of C∞
norm induced by Bλ,1[u]. Therefore, the Hλ-norm is characterized by the identification
0 (Ω). Due
0 (Ω) in the
u2
Hλ = lim
ε→0
Bε
λ,1[u],
∀u ∈ Hλ,
(2.6)
where λ ≤ λ(N ) and
Bε
λ,1[u] :=Zx≥ε(cid:12)(cid:12)(cid:12)∇u +
u(cid:12)(cid:12)(cid:12)
N
2
x
x2 u −
eN
xN
2
dx + (λ(N ) − λ)Zx≥ε
u2
x2 dx,
∀u ∈ Hλ.
Next in the paper we will understand the meaning of the norm · Hλ as in formula (2.6).
2.1.2 Main results
First of all, we note that standard elliptic estimates do not apply for Aλ to obtain enough regularity
for the normal derivative since the singularity x = 0 is located on the boundary. However, the
following trace regularity result stated in Theorem 2.1 holds true. In what follows, D(Aλ) stands
for the domain of Aλ defined in (1.8).
Next, we claim the main results of Section 2.
Theorem 2.1 (Trace regularity). Assume Ω ⊂ RN , N ≥ 1, is a bounded smooth domain satisfying
the case C1. Let us consider λ ≤ λ(N ) and u ∈ D(Aλ). Then
and moreover, there exists a positive constant C = C(Ω) > 0 such that
∂ν(cid:17)2
(cid:16) ∂u
x2 ∈ L1(Γ),
∂ν(cid:17)2
ZΓ(cid:16) ∂u
x2dσ ≤ C(u2
Hλ + Aλu2
L2(Ω)),
∀ u ∈ D(Aλ).
(2.7)
(2.8)
Moreover, we obtain the following
Theorem 2.2 (Pohozaev identity). Assume Ω ⊂ RN , N ≥ 1, is a smooth bounded domain
satisfying the case C1 and let λ ≤ λ(N ). If u ∈ D(Aλ) we claim that
1
2ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
dσ = −ZΩ
Aλu(x · ∇u)dx −
N − 2
2
u2
Hλ,
(2.9)
The proofs of Theorems 2.1, 2.2 are quite technical, so we need to apply some preliminary lem-
mas which are stated below. The proofs of Lemmas 2.1, 2.3 are postponed at the end of Subsection
2.1 while Lemma 2.2 is a consequence of an abstract approximation lemma in a forthcoming work
[1].
7
Lemma 2.1. Supppose u ∈ D(Aλ) and denote f = Aλu ∈ L2(Ω). Let us also consider θε ∈
C∞
0 (Ω), ε > 0, a family of cut-off functions such that
θε(x) = θε(x) =(cid:26) 0,
1,
x ≤ ε
x ≥ 2ε.
(2.10)
Assume ~q ∈ (C2(Ω))N is a vector field such that ~q = ν on Γ, where ν denotes the outward normal
to the boundary Γ (such an election of ~q can be always done in smooth domains, see [27], Lemma
3.1, page 29). Then we have the identity
1
∂ν(cid:17)2
2ZΓ(cid:16) ∂u
(x · ∇u)(~q · ∇u)θεdx
xiθεdx −ZΩ ∇u2(x · ~q)θεdx
N
1
+
f (x2~q · ∇uθε)dx + 2ZΩ
uxiuxjx2qj
div~qx2h∇u2 − λ
x2θεdσ = −ZΩ
Xi,j=1ZΩ
2ZΩ
−
+ZΩ x2(~q · ∇u)(∇u · ∇θε)dx.
u2
x2iθεdx −
1
2ZΩ x2~q · ∇θεh∇u2 − λ
u2
x2idx
(2.11)
Lemma 2.2. Assume f ∈ L2(Ω) and Ω ⊂ RN verifying the case C1. Let ε > 0 be small enough.
We consider the following approximation problem
(cid:26) Aλ(N )−εuε = f, x ∈ Ω
x ∈ ∂Ω.
uε = 0,
Then
uε → u
where u verifies the limit problem
strongly in Hλ(N ),
as ε → 0.
Moreover,
−∆u − λ(N )
u
x2 = f, in D′(Ω).
εZΩ
u2
ε
x2 dx → 0, as ε → 0.
(2.12)
(2.13)
Lemma 2.3. Assume Ω fulfills the case C1 and let λ ≤ λ(N ). Let f ∈ C∞(Ω). Moreover, we
assume that uλ solves the problem
(cid:26) Aλuλ = f, x ∈ Ω,
uλ ∈ Hλ.
(2.14)
Then uλ satisfies the following upper bounds: there exists r0 < RΩ small enough and there exist
constants C1, C2 > 0, independent of λ, such that
uλ(x) ≤ C1xNx−N/2+√λ(N )−λ(cid:12)(cid:12)(cid:12)
∇uλ(x) ≤ C2x−N/2+√λ(N )−λ(cid:12)(cid:12)(cid:12)
log
log
1/2
1/2
1
1
x(cid:12)(cid:12)(cid:12)
x(cid:12)(cid:12)(cid:12)
,
a.e. x ∈ Ωr0,
,
a.e. x ∈ Ωr0,
(2.15)
(2.16)
where Ωr0 := Ω ∩ Br0(0).
Notation: In order to facilitate the computations, in the sequel, we will write " & " and " . "
instead of " ≥ C" respectively " ≤ C" when we refer to universal constants C.
8
2.1.3 Proofs of Theorems 2.1, 2.2
Proof of Theorem 2.1. Following the proof of Theorem 1.1, we are able to show that,
ZΩ x∇u2dx . u2
Hλ,
∀ u ∈ Hλ.
(2.17)
From the above estimate and Cauchy-Schwartz inequality applied to identity (2.11) in Lemma 2.1
we reach to
∂ν(cid:17)2
ZΓ(cid:16) ∂u
x2θεdσ . u2
Hλ + f2
L2(Ω),
∀ u ∈ D(Aλ),
∀ ε > 0.
(2.18)
Thanks to Fatou Lemma we finish the proof of Theorem 2.1.
Proof of Theorem 2.2. We split the proof in two main steps.
Step 1. The subcritical case
Note that Hλ = H 1
0 (Ω). Let u ∈ D(Aλ) and put f := Aλu ∈ L2(Ω). By standard elliptic
estimates we note that u ∈ H 2(Ω \ Bε(0)), for any ε > 0 small enough. Moreover, the normal
derivative ∂u/∂ν ∈ L2
loc(∂Ω \ {0}). We multiply Aλu by x · ∇uθε, where θε, ε > 0, was defined in
(2.10). After integration we get
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
ZΩh∇u2 − λ
θεdσ = −ZΩ
x2iθεdx
f (x · ∇u)θεdx −
1
2ZΓ
u2
N − 2
2
(x · ∇u)(∇u · ∇θε)dx.
(2.19)
−
1
2ZΩh∇u2 − λ
u2
x2ix · ∇θεdx +ZΩ
Combining the Dominated Convergence Theorem (DCT) with Theorem 2.1, the left hand side of
(2.19) converges i.e.
ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
θεdσ →ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
dσ,
as ε → 0.
In the right hand side, we can directly pass to the limit term by term to obtain the identity (2.9)
as follows. Firstly, since x · ∇u ∈ L2(Ω) we have that
(cid:26)
f (x · ∇u)θε ≤ fx · ∇u ∈ L1(Ω),
θε → 1, a.e. asε → 0,
and by DCT we obtain
ZΩ
f (x · ∇u)θεdx →ZΩ
f (x · ∇u)dx,
as ε → 0.
Besides, from Hardy inequality and DCT we have
ZΩ ∇u2θεdx →ZΩ ∇u2dx,
ZΩ
Using the fact that ∇θε = O(1/ε) it follows that
u2
x2 θεdx →ZΩ
u2
x2 dx,
as ε → 0.
(cid:12)(cid:12)(cid:12)ZΩ ∇u2x · ∇θεdx(cid:12)(cid:12)(cid:12)
.ZB2ε\Bε ∇u2dx → 0,
9
u2
(cid:12)(cid:12)(cid:12)ZΩ
x2 x · ∇θεdx(cid:12)(cid:12)(cid:12)
(x · ∇u)(∇u · ∇θε)dx(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)ZΩ
u2
x2 dx → 0,
.ZB2ε\Bε
.ZB2ε\Bε ∇u2dx → 0,
as ε → 0. With these we conclude the solvability of Theorem 2.2 in the subcritical case λ < λ(N ).
Step 2. The critical case λ = λ(N )
As before, let us consider u ∈ D(Aλ(N )) and f := Aλ(N )u ∈ L2(Ω). Our purpose is to show
the validity of Theorem 2.2 for such u.
We proceed by approximations with subcritical values. More precisely, for ε > 0 small enough
we consider the problem
Applying Lemma 2.2 we obtain
(cid:26) Aλ(N )−εuε = f, x ∈ Ω,
uε ∈ H 1
0 (Ω).
uε → u in Hλ(N ),
εZΩ
u2
ε
x2 dx
as ε → 0,
(2.20)
(2.21)
where u solves the limit problem. According to the Pohozaev identity applied to uε we reach to
1
2ZΓ
∂ν (cid:17)2
(x · ν)(cid:16) ∂uε
dσ = −ZΩ
f (x · ∇uε)dx −
N − 2
2 (cid:16)uε2
Hλ(N )
+ εZΩ
u2
ε
x2 dx(cid:17).
(2.22)
Due to Theorem 1.1, the fact that uε → u in Hλ(N ) implies
x · ∇uε → x · ∇u in L2(Ω), as ε → 0.
Therefore, the right hand side in (2.22) converges to
−ZΩ
f (x · ∇u)dx −
N − 2
2
u2
Hλ(N )
:= H(u),
and therefore
there exists lim
ε→0
1
2ZΓ
∂ν (cid:17)2
(x · ν)(cid:16) ∂uε
dσ = H(u).
On the other hand, by standard elliptic regularity one can show that
∂uε
∂ν →
∂u
∂ν
in L2
loc(Γ \ {0}) and
∂uε
∂ν →
∂u
∂ν
a.e. on Γ.
In the sequel, we discuss two different situations for the geometry of Ω.
Case 1. Assume Ω is flat in a neighborhood of zero (i.e. x · ν = 0). Then it easily to note that
lim
ε→0ZΓ
∂ν (cid:17)2
(x · ν)(cid:16) ∂uε
dσ =ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
dσ.
In consequence, u satisfies the Pohozaev identity, by passing to the limit in (2.22).
Case 2. We assume Ω is not necessary flat at origin. We distinguish two cases when discussing
the smoothness of f .
10
The case f ∈ C∞(Ω).
Next we apply Lemma 2.3 for uε the solution of problem (2.20). and we obtain
where g = x2−N(cid:12)(cid:12)(cid:12)
log 1
x2 ≤ g, a.e. on Γ,
(x · ν)(cid:16) ∂uε
(cid:12)(cid:12)(cid:12)
x(cid:12)(cid:12)(cid:12) ∈ L1(Γ). Applying DCT we conclude
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
ε→0ZΓ
∂ν (cid:17)2(cid:12)(cid:12)(cid:12) ≤(cid:16) ∂uε
∂ν (cid:17)2
∂ν (cid:17)2
(x · ν)(cid:16) ∂uε
dσ =ZΓ
lim
dσ.
The case f ∈ L2(Ω).
We consider {fk}k≥1 ∈ C∞(Ω) such that fk → f in L2(Ω), as k → ∞.
Let us call uk the solution of Al(N )uk = fk, for all k ≥ 1. From the previous case, uk satisfies
1
2ZΓ
∂ν (cid:17)2
(x · ν)(cid:16) ∂uk
dσ = −ZΩ
fk(x · ∇uk)dx −
N − 2
2
uk2
Hλ.
(2.23)
We know that fk is a Cauchy sequence in L2(Ω), and due to
uk − ulHλ(N ) . fk − flL2(Ω) → 0, as k, l → ∞,
we deduce that {uk}k≥1 is Cauchy in Hλ(N ). Hence uk → u in Hλ(N ) and
x · ∇uk → x · ∇u in L2(Ω).
As a consequence we can pass to the limit in the right hand side of (2.23). In order to finish the
proof, next we also show we can also pass to the limit in the left hand side. Indeed, in view of
Theorem 2.1 we have
Therefore gk := ∂uk
infinity. This suffices to say that
(cid:17)2
x2dσ . uk − ul2
ZΓ(cid:16) ∂(uk − ul)
∂ν x is a Cauchy sequence in L2(Γ) and gk → g := ∂u
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
∂ν (cid:17)2
(x · ν)(cid:16) ∂uk
dσ =ZΓ
lim
k→∞ZΓ
dσ.
∂ν
Hλ + fk − flL2(Ω).
∂ν x in L2(Γ), as k goes to
Therefore we conclude the proof of Theorem 2.2.
2.1.4 Proofs of useful lemmas
Proof of Lemma 2.1. By standard elliptic estimates, we remark that u ∈ H 2
loc(Ω \ {0}). Thanks
to that, after multiplying f by x~q · ∇uθε we are allowed to integrate by parts on Ω. Firstly, we
obtain
ZΩ
∆u(x2~q · ∇uθε)dx =ZΓ
∂u
∂ν
(x2~q · ∇uθε)dσ −ZΩ ∇u · ∇(x2~q · ∇uθε)dx.
Let us now compute the boundary term above. Since u vanishes on Γ it follows that
∇u =
∂u
∂ν
ν,
on Γ,
11
(2.24)
and moreover, ~q = ν on Γ. Thanks to these we obtain
Therefore,
ZΩ
ZΓ
∂u
∂ν
∂ν(cid:17)2
(x2~q · ∇uθε)dσ =ZΓ(cid:16) ∂u
x2θεdσ.
∂ν(cid:17)2
∆u(x2~q · ∇uθε)dx =ZΓ(cid:16) ∂u
x2θεdσ −ZΩ ∇u · ∇(x2~q · ∇u)θε
−ZΩ x2(~q · ∇u)(∇u · ∇θε)dx.
Let us compute the second term in the integration above. Doing various iterations we obtain
ZΩ ∇u · ∇(x2~q · ∇u)θε = 2ZΩ
N
(x · ∇u)(~q · ∇u)θεdx +
Xi,j=1ZΩ x2qj(u2
xi)xj θεdσ
+
1
2
N
Xi,j=1ZΩ
uxiuxjx2qj
xiθεdx
(2.25)
For the last term in the integration above we get
N
1
2
Xi,j=1ZΩ x2qj(u2
xi)xj θεdσ =
−
x2θεdσ −ZΩ ∇u2(x · ~q)θεdx
1
∂ν(cid:17)2
2ZΓ(cid:16) ∂u
2ZΩ
div~qx2∇u2θεdx −
1
1
2ZΩ x2∇u2~q · ∇θεdx.
(2.26)
According to (2.25) and (2.26) we obtain
ZΩ
∆u(x2~q · ∇uθε)dx =
N
1
−
x2θεdσ − 2ZΩ
∂ν(cid:17)2
2ZΓ(cid:16) ∂u
Xi,j=1ZΩ
uxiuxjxqj
2ZΩ
div~qx2∇u2θεdx +
−ZΩ x2(~q · ∇u)(∇u · ∇θε)dx.
(x · ∇u)(~q · ∇u)θεdx
xiθεdx +ZΩ ∇u2(x · ~q)θεdx
2ZΩ x2∇u2~q · ∇θεdx
+
1
1
On the other hand, it follows that
ZΩ
u
x2 x2~q · ∇uθεdx = −
1
2ZΩ
div~qu2θεdx −
1
2ZΩ
~q · ∇θεu2dx.
From (2.27) and (2.28) we finally obtain the identity of Lemma 2.1.
Proof of Lemma 2.3. For any λ ≤ λ(N ) we fix φλ = xNx−N/2+√λ(N )−λ(cid:12)(cid:12)(cid:12)
consider the problem
1/2
log 1
x(cid:12)(cid:12)(cid:12)
(cid:26) AλUλ = f, x ∈ Ω,
Uλ ∈ Hλ.
The proof follows several steps.
(2.27)
(2.28)
. Let us also
(2.29)
Step 1. Firstly let us check the validity of the Maximum Principle:
uλ(x) ≤ Uλ(x)
a.e. in Ω.
(2.30)
12
Indeed, from the equations satisfied by Uλ, uλ we obtain
− ∆(Uλ ± uλ) − λ
(Uλ ± uλ)
x2
= f ± f ≥ 0,
∀ x ∈ Ω.
Multiplying (2.31) by the negative part (Uλ ± uλ)− we get the reverse Hardy inequality
ZΩh∇(Uλ ± uλ)−2 − λ
[(Uλ ± uλ)−]2
x2
idx ≤ 0.
(2.31)
(2.32)
From the non-attainability of the Hardy constant we necessary must have (Uλ ± uλ)− ≡ 0 in Ω.
Therefore, Uλ ± uλ ≥ 0 in Ω, fact which concludes (2.30).
Step 2. Next, we remark that there exists a positive constant C > 0, independent of λ such that
−∆φλ − λ
φλ
x2 ≥ C1,
∀x ∈ Ω.
Therefore, for C ≥ fL∞ /C1 we get
( −∆(Cφλ − Uλ) − λ (Cφλ−Uλ)
Cφλ − Uλ ≥ 0,
x2
≥ 0, ∀x ∈ Ω,
x ∈ Γ.
Therefore, applying the Maximum Principle we obtain
and the proof (2.15) is finished.
Uλ ≤ Cφλ,
∀x ∈ Ω,
λ ≤ λ(N ),
(2.33)
(2.34)
Step 3. For the estimate (2.16) we use a remark by Brezis-Marcus-Shafrir [7] as follows.
Fix x ∈ Ωr0/2 and put r = x/2. We define then uλ(y) = uλ(x + ry) where y ∈ B1(0). By
direct computations we obtain
∆uλ(y) = r2∆uλ(x + ry) = r2(cid:16) − f − λ
uλ(x + ry)
x + ry2 (cid:17)
(2.35)
= −r2f − λ
On the other hand, we remark that
x2
4x + ry2 uλ(y).
4
9 ≤ x2
x + ry2 ≤ 4,
∀ y ∈ B1(0).
By elliptic estimates it is easy to see that uλ ∈ C1(B1(0)). Applying the interpolation inequality
(see Evans [13]), we get that
∇uλ(0) . uλL∞(B1(0)) + ∆uλL∞(B1(0))
. uλL∞(B1(0)) + fL∞(Ω)
Writing ∇uλ in terms of ∇uλ we obtain
∇uλ(x) .
1
x
(uλL∞(B1(0)) + fL∞)
(2.36)
(2.37)
13
In addition, from (2.34) we have
uλL∞(B1(0)) = uλ(x + ry)L∞(B1(0))
1
log
. sup
y∈B1(0)n(xN + ryN )x + ry−N/2+√λ(N )−λ(cid:12)(cid:12)(cid:12)
. xNx−N/2+√λ(N )−λ(cid:12)(cid:12)(cid:12)
. x−(N −2)/2+√λ(N )−λ(cid:12)(cid:12)(cid:12)
x(cid:12)(cid:12)(cid:12)
x(cid:12)(cid:12)(cid:12)
1/2o
x + ry(cid:12)(cid:12)(cid:12)
+ x−(N −2)/2+√λ(N )−λ(cid:12)(cid:12)(cid:12)
x(cid:12)(cid:12)(cid:12)
log
log
1
1/2
1
1/2
log
,
1
1/2
(2.38)
which is verified for all x ∈ Ωr0 , y ∈ B1(0). From (2.37) and (2.38) we obtain the estimate (2.16)
which yield the proof of Lemma 2.3.
2.2 Applications to semi-linear equations
Pohozaev-type identities apply mostly to show non-existence results for elliptic problems. In what
follow we emphasize a direct application to a non-linear elliptic equation with boundary singular
potential. To fix the ideas, let us assume λ < λ(N ) and consider Ω ⊂ RN , N ≥ 1, a domain
satisfying the case C1. Next
α⋆ :=
stands for the critical Sobolev exponent.
Next we claim
Theorem 2.3. Let us consider the problem
N + 2
N − 2
(cid:26) −∆u − λ
u = 0,
x2 u = uα−1u, x ∈ Ω,
x ∈ Γ.
(2.39)
1. Assume λ ≤ λ(N ). If 1 < α < α⋆ the problem (2.39) has non trivial solutions in Hλ. Moreover,
N −2 the problem (2.39) has non trivial solutions in D(Aλ).
if 1 < α < N
2. (non-existence). Assume λ ≤ λ(N ) and let Ω be a smooth star-shaped domain (i.e. x · ν ≥ 0,
for all x ∈ Γ). If α ≥ α⋆ the problem does not have non trivial solutions in D(Aλ).
Proof of Theorem 2.3
Proof of 1. The existence of non trivial solutions for (2.39) reduces to study the minimization
problem
I =
inf
u∈Hλ,u6=0
Hλ
u2
uα+1
Lα+1(Ω)
.
Without losing the generality, we may consider the normalization
I =
inf
uLα+1(Ω)=1
J(u),
(2.40)
where J : Hλ → R, J(u) = u2
Hλ
and we address the question of attainability of I in (2.40).
We note that J is continuous, convex, coercive in Hλ. Let {un}n be a minimizing sequence of I,
i.e.,
By the coercivity of J we have
J(un) ց I,
unLα+1(Ω) = 1.
unHλ ≤ C,
∀n,
14
Moreover, the embedding Hλ ֒→ Lα+1(Ω) is compact for any α < α⋆ (it can be deduced combining
Theorem 1.2 and Sobolev inequality). Therefore,
(cid:26) un ⇀ u weakly in Hλ,
un → u
strongly in Lα+1(Ω).
(2.41)
Therefore, uLα+1(Ω) = 1. From the i.s.c. of the norm we have
J(un) = I,
I ≤ J(u) ≤ lim inf
n→∞
and therefore I = J(u) is attained by u, which, up to a constant, is a non-trivial solution of (2.39)
in Hλ.
If α < N/(N − 2) let us show that u ∈ D(Aλ).
Hλ ֒→ Lq(Ω), q < 2N/(N − 2), we have that uα−1u ∈ L2(Ω). In consequence, u ∈ D(Aλ).
Proof of 2. For the proof of the non-existence part we apply the Pohozaev identity in Theorem
2.2. In view of that we use the following lemma whose proof is postponed at the end of the section.
Lemma 2.4. Assume λ ≤ λ(N ) and 1 < α < ∞. Then, any solution u ∈ D(Aλ) of (2.39)
satisfies the identity
Indeed, due to the compact embedding
1
2ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
dσ =(cid:16) N
1 + α −
N − 2
2
(cid:17)ZΩ uα+1dx.
(2.42)
The case α > α⋆.
Note that x · ν ≥ 0 for all x ∈ Γ. Assuming u 6≡ 0, from Lemma 2.4 we obtain (N − 2)/2 ≤
N/(α + 1) which is equivalent to α ≤ α⋆. This is in contradiction with the hypothesis on α.
Therefore u ≡ 0 in Ω.
The case α = α⋆.
From Lemma 2.4, due to the criticality of α⋆, u must satisfy
Z∂Ω
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
dσ = 0.
We fix Ω = {x ∈ RN
+ x′2 + (xN − 1)2 ≤ 1} which is star-shaped. Therefore,
∂u
∂ν
= 0,
a.e. on Γ.
Therefore, the problem in consideration is reduced to the overdetermined system
x2 u = u
−∆u − λ
u = 0,
∂u
∂ν = 0,
4
N −2 u, x ∈ Ω,
x ∈ Γ,
x ∈ Γ.
(2.43)
Let us consider a compact subset Γ′ ⊂ Γ such that x · ν > 0 and 0 6∈ Γ′. Next, we extend Ω
with a bounded set Ω1 such that Ω1 ∩ Ω = Ø, ∂Ω1 ∩ ∂Ω = Γ′, Ω := Ω ∪ Ω1.
For ε > 0 small enough we denote the sets Ωε := Ω\ {x ∈ Ω x < ε}, Ωε := Ω\ {x ∈ Ω x < ε}.
Consider also the trivial prolongation of u to Ω
u :=(cid:26) u, x ∈ Ω,
0, Ω1.
15
(2.44)
The fact that u ∈ D(Aλ) combined with the over-determined condition in (2.43), imply that
u ∈ H 2( Ωε). Let us also show that u ∈ H 2( Ωε).
Indeed, thanks to (2.43) on Γ0 we get that
Z Ωε
where g ∈ L2( Ωε) is given by
du
∂xi
In particular we obtain that
and u verifies
gφdx,
∀φ ∈ C∞
0 ( Ωε),
∂φ
∂xj
dx = −Z Ωε
g =( ∂2u
∂xi∂xj
0,
, x ∈ Ωε,
x ∈ Ω1.
∆u =(cid:26) ∆u, x ∈ Ωε,
x ∈ Ω1.
x2 u = u
λ
0
4
−∆u = V (x)u,
x ∈ Ωε,
− ∆u −
N −2 u a.e. in Ωε
and u ≡ 0 in Ω1. In other words we can write (2.48) as
where V (x) = λ
x2 + u
N −2 . Note that V ∈ Lω( Ωε) for some ω > N/2 and u vanishes in Ω1.
4
With these we are in the hypothesis of the strong unique continuation result by Jerison and
Kenig [23]. Therefore, u ≡ 0 in Ωε and in particular u ≡ 0 in Ωε, for any ε > 0. Hence, we
conclude that u ≡ 0 in Ω. The proof of Theorem 2.3 is finished.
Proof of Lemma 2.4. Since u ∈ D(Aλ) we can apply the Pohozaev identity and we get
Next we show that
dσ =ZΩ −uα−1u(x · ∇u)dx −
N − 2
2
1
2Z∂Ω
∂ν(cid:17)2
(x · ν)(cid:16) ∂u
ZΩ uαu(x · ∇u)dx = −
N
1 + α ZΩ uα+1dx.
u2
Hλ ,
(2.49)
(2.50)
We proceed by approximation arguments. For ε > 0 small enough we consider Iε := RΩ uαu(x ·
∇u)θεdx, where θε is a cut-off function supported in Ω\Bε(0). Due to the fact that u ∈ H 2(Ω\{0})
we can integrate by parts as follows.
(2.45)
(2.46)
(2.47)
(2.48)
(2.51)
(2.52)
1
1
=
2ZΩ
2ZΩ up−1x · ∇(u2)θεdx =
Iε = −
1
2ZΩ
u2(cid:2)Nuα−1θε + x · θεuα−1 + (α − 1)x · ∇uuα−3u(cid:3)dx
2 ZΩ uα+1θεdx +
2ZΩ uα+1x · ∇θεdx −
u2div(cid:2)uα−1xθε(cid:3)dx
α − 1
Iε.
N
=
2
1
Therefore we obtain
Iε =
N
α + 1ZΩ uα+1θεdx +
1
α + 1ZΩ uα+1x · ∇θεdx.
From the equation itself it is easy to see that uα+1 ∈ L1(Ω) provided u ∈ D(Aλ). Therefore, by
the DCT we can pass to the limit as ε → 0 in (2.52) to obtain the identity (2.50). On the other
hand, multiplying (2.39) by u and integrating we obtain
Combining this with (2.50) and (2.49) we conclude (2.42).
u2
Hλ =ZΩ uα+1dx,
16
2.3 Brief presentation of the case C2
Inequalities (1.6), (1.7) can be stated in a simplified form as follows.
Assume Ω ⊂ RN is a smooth bounded domain containing the origin on the boundary. For any
l ≤ N 2/4 and any 0 < γ < 2 there exists a constant C1(γ, Ω) ≥ 0 such that
∀u ∈ H 1
0 (Ω),
ZΩ
u2
xγ dx + lZΩ
u2
x2 dx ≤ZΩ ∇u2 + C1(γ, Ω)ZΩ
u2dx.
(2.53)
2.3.1
Functional framework via Hardy inequality
Let us now define the set
C :=nC ≥ 0 s. t.
inf
u∈H1
0 (Ω)RΩ(cid:2)∇u2 − l(N )u2/x2 + Cu2(cid:3)dx
RΩ u2/xγdx
Of course, C is non empty due to inequality (2.53). Next we define
C0 = inf
C∈C
C.
Then, for any λ ≤ λ(N ) = N 2/4 we introduce the Hardy functional
u2dx,
u2
Bλ[u] :=ZΩ ∇u2dx − λZΩ
x2 dx + C0ZΩ
≥ 1 o.
(2.54)
(2.55)
(2.56)
which is positive for any u ∈ H 1
define the corresponding Hilbert space Hλ as the closure of C∞
Observe that for any λ < λ(N ) the identification Hλ = H 1
we have
0 (Ω) due to inequality (2.53) and the election of C0. Then we
0 (Ω) in the norm induced by Bλ[u].
0 (Ω) holds true. Indeed, if λ < λ(N ),
Bλ[u] ≥(cid:0)1 −
λ
λ(N )(cid:1)ZΩ ∇u2dx − C0λ
λ(N )ZΩ
u2dx.
(2.57)
On the other hand, from the definition of C0 we obtain that there exists a constant C1 = C1(γ) > 0
such that
Bλ[u] ≥ C1ZΩ
u2dx.
(2.58)
Multiplying (2.58) by C0λ/(C1λ(N )) and summing to (2.57) we get that
Bλ[u] ≥ CλZΩ ∇u2dx,
for some positive constant Cµ that converges to zero as λ tends to λ(N ).
Besides, in the critical case λ = λ(N ), Hλ is slightly larger than H 1
0 (Ω). However, using cut-off
arguments near the singularity (see e.g. [36]) we can show that
Bλ[u]λ(N ) ≥ CεuH1(Ω\Bε(0)),
where Cε is a constant going to zero as ε tends to zero.
∀u ∈ H 1
0 (Ω)
Let us define de operator Aλ := −∆ − λ/x2 + C0I and define its domain as
D(Aλ) := {u ∈ Hλ Aλu ∈ L2(Ω)}.
The norm of the operator Aλ is given by
uD(Aλ) = uL2(Ω) + AλuL2(Ω).
17
(2.59)
(2.60)
(2.61)
2.3.2 The meaning of the Hλ-norm
First of all we remark that
ZΩ ∇u2dx +ZΩ
∆Φ
Φ
u2dx =ZΩ(cid:12)(cid:12)(cid:12)∇u − ∇Φ
Φ
u(cid:12)(cid:12)(cid:12)
and any distribution satisfying Φ, 1/Φ ∈ C1(Ω \ {0}) and Φ > 0 in Ω.
Let us also consider φ(x) = φ(x) ∈ C∞(Ω) to be a cut-off function such that
2
dx,
∀ u ∈ C∞
0 (Ω),
(2.62)
φ =(cid:26) 1,
0,
x ≤ r0/2, x ∈ Ω
x ≥ r0, x ∈ Ω,
(2.63)
where r0 > 0 is aimed to be small.
Case 1.Assume the points on the boundary Γ of Ω satisfy xN > 0 in a neighborhood of the origin.
We take Φ1 = xNx−N/2 which satisfies the equation
− ∆Φ1 −
N 2
4
Φ1
x2 = 0,
a. e. in Ωr0 ,
(2.64)
where Ωr0 := Ω ∩ Br0 (0) for some r0 > 0 small enough. From (2.62) and (2.64) we obtain
ZΩr0
∇v2dx −
N 2
4 ZΩr0
v2
x2 dx =ZΩr0 (cid:12)(cid:12)∇v − ∇Φ1
Φ1
v(cid:12)(cid:12)
2
dx,
∀ v ∈ C∞
0 (Ωr0 ).
(2.65)
By a standard cut-off argument, due to (2.65) we remark that, there exist some weights ρ1, ρ2 ∈
C∞(Ω) depending on r0, supported far from origin such that
2
Φ1
u2
Bλ[u] =ZΩ(cid:12)(cid:12)(cid:12)∇(uφ) − ∇Φ1
(uφ)(cid:12)(cid:12)(cid:12)
+ (λ(N ) − λ)ZΩ
x2 dx +ZΩ
Then the meaning of · Hλ -norm is characterized by
ε→0Zx∈Ω,x>ε(cid:12)(cid:12)(cid:12)∇(uφ) − ∇Φ1
+ (λ(N ) − λ)ZΩ
x2 dx +ZΩ
Hλ = lim
u2
Φ1
u2
dx +ZΩ
ρ1∇u2dx
ρ2u2dx,
∀u ∈ C∞
0 (Ω).
(2.66)
(uφ)(cid:12)(cid:12)(cid:12)
ρ2u2dx,
2
dx +ZΩ
∀u ∈ Hλ,
ρ1∇u2dx
∀λ ≤ λ(N ).
(2.67)
Case 2. Assume the points on Γ satisfy xN ≤ 0 in a neighborhood of the origin
In this case we consider d = (x, Γ) = d(x) the function denoting the distance from a point
x ∈ Ω to Γ. We remark that close enough to origin the distribution
satisfies
Φ2 = d(x)e(1−N )d(x)x−N/2(cid:12)(cid:12)(cid:12)
4x2 Φ2 = P > 0,
−∆Φ2 −
N 2
log
1/2
,
1
x(cid:12)(cid:12)(cid:12)
∀x ∈ Ωr0
where r0 > 0 is small enough.Due to this, there exist the weights ρ1, ρ2 ∈ C∞(Ω) depending on
r0 and supported away from origin, such that the meaning of Hλ-norm is given by
u2
Hλ = lim
ε→0Zx∈Ω,x>ε(cid:12)(cid:12)(cid:12)∇(uφ) − ∇Φ2
+ (λ(N ) − λ)ZΩ
x2 +ZΩ
Φ2
u2
(uφ)(cid:12)(cid:12)(cid:12)
ρ1∇u2dx +ZΩ
2
dx +ZΩ
P
Φ2uφ2dx
ρ2u2dx,
∀u ∈ Hλ,
∀λ ≤ λ(N ).
(2.68)
18
Case 3. Assume that xN changes sign on Γ at the origin.
This case can be analyzed through Case 2 above.
Then, the Pohozaev identity and related results presented in case C1 might be extended to
case C2 by means of the weaker functional settings introduced above.
3 Applications to Controllability
In this section we study the controllability of the wave and Schrodinger equations with singularity
localized on the boundary of a smooth domain. Our motivation came through the results shown
in [34] in the context of interior singularity.
For the sake of clarity, we will discuss in a detailed manner the case C1.
3.1 The wave equation. Case C1
In the sequel, we are focused to the controllability of the wave-like system
x2 = 0,
utt − ∆u − λ u
u(t, x) = h(t, x),
u(t, x) = 0,
u(0, x) = u0(x),
ut(0, x) = u1(x),
(t, x) ∈ QT ,
(t, x) ∈ (0, T ) × Γ0,
(t, x) ∈ (0, T ) × (Γ \ Γ0),
x ∈ Ω,
x ∈ Ω.
(3.1)
(Wλ) :
where QT = (0, T )×Ω, Γ denotes the boundary of Ω and Γ0 is the boundary control region defined
in (1.11), where the control h ∈ L2((0, T ) × Γ0) is acting. We also assume λ ≤ λ(N ). In view of
the time-reversibility of the equation it is enough to consider the case where the target
(u0, u1) = (0, 0).
It is the so-called null controllability problem.
3.1.1 Well-posedness
Let us briefly discuss the well-posedness of system (3.1) in the corresponding functional setting.
Instead of (1.10) we firstly consider the more general system with non-homogeneous boundary
conditions:
x2 = 0,
utt − ∆u − λ u
u(t, x) = g(t, x),
u(0, x) = u0(x),
ut(0, x) = u1(x),
(t, x) ∈ QT ,
(t, x) ∈ ΣT ,
x ∈ Ω,
x ∈ Ω.
(3.2)
where ΣT = (0, T ) × Γ. The solution of (3.2) is defined by the transposition method (J.L. Lions
[27]):
Definition 3.1. Assume λ ≤ λ(N ). For (u0, u1) ∈ L2(Ω) × H ′
that u is a weak solution for (3.2) if
Z T
0 ZΩ
λ and g ∈ L2((0, T ) × Γ), we say
λ,Hλ −Z T
0 ZΓ
uf dx = − < u0, z′(0) >L2(Ω),L2(Ω) + < u1, z(0) >H ′
g
∂z
∂ν
∀ f ∈ D(Ω),
(3.3)
19
where < ·,· > represents the dual product between Hλ and its dual H
the non-homogeneous adjoint-backward problem
ztt − ∆z − λ z
z(t, x) = 0,
z(T, x) = z′(T, x) = 0, x ∈ Ω.
(t, x) ∈ QT ,
(t, x) ∈ ΣT ,
x2 = f,
′
λ, and z is the solution of
(3.4)
Formally, (3.3) is obtained by multiplying the system (3.4) with u and integrate on QT . Using
the Hardy inequalities above and the application of standard methods for evolution equations we
lead to the following existence result.
Theorem 3.1 (well-posedness). Assume that Ω satisfies C1. Let T > 0 be given and assume
λ and any h ∈ L2((0, T ) × Γ0) there exists a unique
λ ≤ λ(N ). For every (u0, u1) ∈ L2(Ω) × H
weak solution of (1.10) such that
′
Moreover, the solution of (1.10) satisfies
u ∈ C([0, T ]; L2(Ω)) ∩ C1([0, T ]; H
′
λ).
(u, ut)L∞(0,T ;L2(Ω)×H ′
λ) . (u0, u1)L2(Ω)×H ′
λ
+ hL2((0,T )×Γ0).
(3.5)
(3.6)
The details of the proof of Theorem 3.1 are omitted since they follow the same steps as in [34].
3.1.2 Controllability and main results
It is by now classical that controllability of (3.1) is characterized through an observability inequal-
ity for the adjoint system as follows below.
Given initial data (u0, u1) ∈ L2(Ω) × Hλ
′, a possible control h ∈ L2((0, T ) × Γ0) must satisfy
the identity
Z T
0 ZΓ0
h
∂v
∂ν
dσdt− < ut(0), v(0) >H ′
λ,Hλ
+ < u(0), vt(0) >L2(Ω),L2(Ω)= 0,
(3.7)
where v is the solution of the adjoint system
x2 = 0,
vtt − ∆v − λ v
v(t, x) = 0,
v(0, x) = v0(x),
vt(0, x) = v1(x),
(t, x) ∈ QT ,
(t, x) ∈ ΣT ,
x ∈ Ω,
x ∈ Ω.
(3.8)
The operator Aλ defined by Aλ(w1, w2) = (w2, ∆w1 + λx2w1) for all (w1, w2) ∈ D(Aλ) =
D(Aλ) × Hλ, generates the wave semigroup i.e. (Aλ, D(Aλ)) is m-dissipative in Hλ × L2(Ω). In
view of that, due to the theory of semigroups, the adjoint system is well-posed and more precisely
it holds that
Proposition 3.1 (see, e.g.[34]). (1) For any initial data (v0, v1) ∈ Hλ × L2(Ω) there exists a
unique solution of (3.8)
u ∈ C([0, T ]; Hλ) ∩ C1([0, T ]; L2(Ω)).
Moreover,
(v, vt)L∞(0,T ;Hλ×L2(Ω)) . v0Hλ + v1L2(Ω)
(3.9)
(2) For any initial data (v0, v1) ∈ D(Aλ) × Hλ there exists a unique solution of (3.8) such thar
v ∈ C([0, T ]; D(Aλ)) ∩ C1([0, T ]; Hλ) ∩ C2([0, T ]; L2(Ω)).
Moreover
(v, vt)L∞(0,T ;D(Aλ)×Hλ) . v0D(Aλ) + v1Hλ
(3.10)
20
In the sequel, we claim some "hidden regularity" effect for the system (3.8) which may not be
directly deduce from the semigroup regularity but from the equation itself.
Theorem 3.2 (Hidden regularity). Assume λ ≤ λ(N ) and v is the solution of (3.8) corre-
sponding to the initial data (v0, v1) ∈ Hλ × L2(Ω). Then v satisfies
Z T
0 ZΓ
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
dσdt .Z T
∂ν(cid:17)2
0 ZΓ(cid:16) ∂v
Moreover, v verifies the identity
x2dσdt . v02
Hλ + v12
L2(Ω).
(3.11)
2
1
T
2
dσdt =
(v02
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
2Z T
0 ZΓ
Hλ + v12
Due to Theorem 3.2 the operator (v0, v1) 7→(cid:0)R T
is a linear contin-
uous map in Hλ × L2(Ω). Let H be the completion of this norm in Hλ × L2(Ω). We consider the
functional J : H → R defined by
2Z T
0 ZΓ0
L2(Ω)) +ZΩ
0 RΓ0
(x· ν)(∂v/∂ν)2dσdt(cid:1)1/2
vt(cid:0)x · ∇v +
dσdt− < u1, v0 >H ′
+(u0, v1)L2(Ω),L2(Ω),
(x · ν)(cid:16) ∂v
∂ν(cid:17)2
J(v0, v1)(v) :=
v(cid:1)(cid:12)(cid:12)(cid:12)
N − 1
(3.13)
(3.12)
λ,Hλ
dx.
T
0
where v is the solution of (3.8) corresponding to initial data (v0, v1). Of course, < ·,· >H ′
denotes the duality product. A control h ∈ L2((0, T ) × Γ0) satisfying (3.7) could be chosen
as h = (x · ν)vmin where vmin minimizes the functional J on H among the solutions v of (3.8)
λ×L2(Ω) The existence of a minimizer of J is assured
corresponding to the initial data (u0, u1) ∈ H
by the coercivity of J, which is equivalent to the Observability inequality for the adjoint system
(3.8):
λ,Hλ
′
1
v02
Hλ + v12
L2(Ω) .Z T
0 ZΓ0
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
dσdt,
Conservation of energy.
For any λ ≤ λ(N ) and any fixed time t ≥ 0, let us define the energy associated to (3.8):
We note that our system is conservative and therefore
1
Eλ
v (t) =
Eλ
v (t) = Eλ
v (0),
2(cid:0)vt(t)2
L2(Ω) + v(t)2
Hλ(cid:1)
∀t ∈ [0, T ].
∀λ ≤ λ(N ),
(3.14)
(3.15)
Next we claim our main results which answer to the controllability question.
Theorem 3.3 (Observability inequality). For all λ ≤ λ(N ), there exists a positive constant
D1 = D1(Ω, λ, T ) such that for all T ≥ 2RΩ, and any initial data (v0, v1) ∈ Hλ × L2(Ω) the
solution of (3.8) verifies the observability inequality
El
v(0) ≤ D1Z T
0 ZΓ0
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
dσdt.
(3.16)
The proof of Theorem 3.3 relies mainly on the method of multipliers (cf.
[27]) and the so
called compactness uniqueness argument (cf.
[28]), combined with the new Hardy inequalities
above. These results guarantee the exact controllability of (1.10) when the control acts on the
part Γ0. In conclusion, we obtain
Theorem 3.4 (Controllability). Assume that Ω satisfies C1 and λ ≤ λ(N ). For any time
λ there exists h ∈ L2((0, T ) × Γ0) such
T > 2RΩ, (u0, u1) ∈ L2(Ω) × H
that the solution of (1.10) satisfies
λ and (u0, u1) ∈ L2(Ω) × H
′
′
(ut(T, x), u(T, x)) = (u1(x), u0(x))
for all x ∈ Ω.
21
3.1.3 Proofs of main results
First of all, we need to justify that the solution v of adjoint system (3.8) posses enough regularity
to guarantee the integrability of the boundary term in (3.16). The justification is not trivial
because the presence of the singularity at the boundary.
Proof of Theorem 3.2. We will proceed straightforward from Theorem 2.2.
Firstly, we consider initial data (v0, v1) in D(Aλ) = D(Aλ) × Hλ. Then, according to Propo-
sition 3.1 we have
v ∈ C([0, T ]; D(Aλ)) ∩ C1([0, T ]; Hλ) ∩ C2([0, T ]; L2(Ω)).
For a fixed time t ∈ [0, T ] we apply Theorem 2.2 for Aλv = −vtt and we obtain
ZΓ(cid:16) ∂v
∂ν
(x, t)(cid:17)2
x2dσ . v(t)2
Hλ + vtt(t)2
L2(Ω),
∀t ∈ [0, T ].
(3.17)
Integrating in time and if necessary in space we derive
Z T
0 ZΓ(cid:16) ∂v
∂ν
(x, t)(cid:17)2
x2dσdt .Z T
0
v(t)2
Hλ dt +ZΩ
v2
t (T, x)dx −ZΩ
v2
t (0, x)dx.
(3.18)
According to the conservation of energy we reach to
Z T
0 ZΓ(cid:16) ∂v
∂ν
(x, t)(cid:17)2
x2dσdt . 2Z T
0
Eλ
v (t)dt + Eλ
v (T )
= (2T + 2)Eλ
v (0)
= (T + 1)(v02
Hλ + v12
L2(Ω)).
(3.19)
Indeed, integrating in time in
Since x · ν . x2 on Γ, from above we conclude the inequality (3.11).
0
T
2
dx −
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
Next, we apply the Pohozaev identity for v(t), t ∈ [0, T ].
Theorem 2.2 for Aλv = −vtt, we get
Z T
2Z T
1
N − 2
dσdt =ZΩ
0 ZΓ
=ZΩ
2ZZQT
x · ∇(v2
=ZΩ
2 ZΩ vt(t)2
2Z T
=ZΩ
(cid:2)vt(t)2
N − 1
L2(Ω) − v(t)2
vt(x · ∇v)(cid:12)(cid:12)(cid:12)
vt(x · ∇v)(cid:12)(cid:12)(cid:12)
vt(x · ∇v)(cid:12)(cid:12)(cid:12)
vt(x · ∇v)(cid:12)(cid:12)(cid:12)
Z T
dx −
dx +
dx +
N
+
2
1
1
T
0
T
0
T
0
0
0
0
(cid:2)vt(t)2
v(t)2
Hλ dt
0
2
N − 2
N − 2
t )dxdt −
L2(Ω)dt −
Z T
Z T
L2(Ω)dt + v(t)2
Hλ(cid:3)dt
Hλ(cid:3)dt.
2
0
v(t)2
Hλ dt
v(t)2
Hλ dt
(3.20)
Multiplying the equation of (Wλ)adj by v and integrate in space, the equipartition of the energy
T
0
dx = vt(t)2
L2(Ω) − v2
Hλ ,
ZΩ
vvt(cid:12)(cid:12)(cid:12)
holds true. Due to the conservation of energy and from relations above, we obtain precisely the
identity (3.12). This yields the proof of Theorem 2.1 for initial data in the domain D(Aλ). Then, by
density arguments, one can extend the results for less regular initial data (v0, v1) ∈ Hλ×L2(Ω).
22
Proof of Theorem 3.3. In what follows we present the proof in the critical case λ = λ(N ), which
is of main interest. The subcritical case λ < λ(N ) is let to the reader.
Step 1. Firstly, from Lemma 3.2 we remark that
For a fixed time t = t0 > 0, by Cauchy-Schwartz inequality we have
vt(cid:16) N − 1
2
(cid:12)(cid:12)(cid:12)ZΩ
ZΩ
v
T
0
2
vt(
N − 1
v + x · ∇v)dx(cid:12)(cid:12)(cid:12)
2 ZΩ
v + x · ∇v(cid:17)dx(cid:12)(cid:12)(cid:12)t=t0 ≤
RΩ
2 vt2
+ (N − 1)ZΩ
RΩ
=
On the other hand it follows
+ T Eλ(N )
(0) ≤
1
2Z T
0 ZΓ0
(x · ν)(
∂v
∂ν
)2dσdt.
(3.21)
v2
t dx +
L2(Ω) +
2
1
2RΩ ZΩ(cid:16) N − 1
2 (cid:17)2
2RΩh(cid:16) N − 1
1
v + x · ∇v(cid:17)2
v2
dx
L2(Ω) + x · ∇v2
L2(Ω)
v(x · ∇v)dxi.
Applying Theorem 3.3 we deduce
Therefore we obtain
vt(cid:16) N − 1
2
(cid:12)(cid:12)(cid:12)ZΩ
1
1
L2(Ω) +
2ZΩ
v(x · ∇v)dx =
x · ∇(v2)dx = −
vt(cid:16) N − 1
1
2RΩx · ∇v2
ZΩ
2ZΩ
v + x · ∇v(cid:17)dx(cid:12)(cid:12)(cid:12)t=t0 ≤
(cid:12)(cid:12)(cid:12)ZΩ
v + x · ∇v(cid:17)dx(cid:12)(cid:12)(cid:12)t=t0 ≤ RΩEλ(N )
vt(cid:16) N − 1
v + x · ∇v(cid:17)dx(cid:12)(cid:12)(cid:12)
t=0(cid:12)(cid:12)(cid:12) ≤ 2RΩEλ(N )
2Z T
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
0 ZΓ0
(0) ≤
t=T
2
2
1
v
v
(T − 2RΩ)Eλ(N )
v
From (3.21) and (3.23) we obtain
(cid:12)(cid:12)(cid:12)ZΩ
for some constant C. Due to the conservation of the energy and taking t0 = 0 respectively t0 = T
and summing in (3.22) we get
div(x)v2dx = −
RΩ
2 vt2
L2(Ω) −
N
v2dx
2 ZΩ
2RΩ(cid:16) N 2 − 1
1
4
(cid:17)v2
L2(Ω)
(t0) − Cv(t0)2
L2(Ω),
(3.22)
(0) − C(v(0)2
L2(Ω) + v(T )2
L2(Ω)),
dσdt + C(v(0)2
L2 + v(T )2
L2).
(3.23)
(3.24)
Step 2. To get rid of the remainder term at the right hand side of (3.24) we need the following
lemma.
Lemma 3.1. There exists a positive constant C = C(T, Ω) > 0 such that
v(0)2
L2(Ω) + v(T )2
L2(Ω) ≤ CZ T
0 ZΓ0
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
dσdt
(3.25)
for all finite energy solution of (3.2).
Combining Lemma 3.1 with (3.24), the Observability inequality is finally proved.
Proof of Lemma 3.1. We apply a classical compactness-uniqueness argument. Suppose by contra-
diction that (3.25) does not hold. Then there exists a sequence (vn
1 ) of initial data such that
the corresponding solution vn verifies
0 , vn
vn(0)2
R T
0 RΓ0
L2(Ω) + vn(T )2
∂ν (cid:17)2
(x · ν)(cid:16) ∂vn
dσdt
L2(Ω)
→ ∞.
23
Normalizing we may suppose that
vn(0)2
L2(Ω) + vn(T )2
L2(Ω) = 1,
Z T
0 ZΓ0
(x · ν)(cid:16) ∂vn
∂ν (cid:17)2
dσdt → 0.
(3.26)
From (3.24) we deduce that the corresponding energy is uniformly bounded. In particular, we
deduce that vn is uniformly bounded in
C([0, T ]; Hλ(N )) ∩ C1([0, T ]; L2(Ω)).
Therefore, by extracting a subsequence
vn ⇀ v in L∞(0, T ; Hλ(N )) weakly-⋆,
(3.27)
From Theorem 2.1 we obtain
∂vn
∂ν
√x · ν ⇀
∂v
∂ν
√x · ν in L∞(0, T ; L2(Γ0)) weakly- ⋆ .
Furthermore, by lower semicontinuity,
0 ≤Z T
0 ZΓ0
Hence
and
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
Z T
0 ZΓ0
(x · ν)
∂v
∂ν
= 0,
0 ZΓ0
(x · ν)(cid:16) ∂vn
∂ν (cid:17)2
dσdt = 0.
dσdt ≤ lim inf
n→∞ Z T
∂ν(cid:17)2
(x · ν)(cid:16) ∂v
dσdt = 0,
a.e. on Γ0,
∀t ∈ [0, T ].
(3.28)
On the other hand, from compactness we deduce that
which combined with (3.26) yield to
vn → v in L∞(0, T ; L2(Ω)),
v02
L2(Ω) + v(T )2
L2(Ω) = 1.
(3.29)
To end the proof of Lemma 3.1 it suffices to observe that (3.28)-(3.29) lead to a contradiction.
Indeed, in view of (3.28) and by Holmgreen unique continuation we deduce that v ≡ 0 in Ω which
is in contradiction with (3.29).
Remark 3.1. Unique continuation results may be applied far from origin where coefficient of the
lower order term of the operator −∂tt − ∆ − λ/x2 is analytic in time (actually, it is independent
of time and bounded in space). The principal part coincides with the D'Alambertian operator, then
one can apply Homlgreen's unique continuation to get v = 0 a.e. in Ω \ B(0, ε) for any ε > 0. In
consequence, we will have v ≡ 0 in Ω, see [29].
3.2 The Schrodinger equation
In this section we consider the Schrodinger-like equation
x2 = 0,
iut − ∆u − λ u
u(t, x) = h(t, x),
u(t, x) = 0,
u(0, x) = u0(x),
(t, x) ∈ QT ,
(t, x) ∈ (0, T ) × Γ0,
(t, x) ∈ (0, T ) × (Γ \ Γ0),
x ∈ Ω,
24
(3.30)
Moreover, we assume Ω ⊂ RN , N ≥ 1, is a smooth bounded domain satisfying case C1 and
λ ≤ λ(N ) := N 2/4. For the Schrodinger equation we define the Hilbert spaces L2(Ω; C) and
H 1
0 (Ω; C) endowed with the inner products
< u, v >L2(Ω;C):= ReZΩ
u(x)v(x)dx,
∀u, v ∈ L2(Ω; C),
< u, v >H1
0 (Ω;C):= ReZΩ ∇u(x) · ∇v(x)dx,
∀u, v ∈ H 1
0 (Ω; C).
For all λ ≤ λ(N ), we also define the Hilbert space Hλ(Ω; C) as the completion of H 1
respect to the norm associated with the inner product
0 (Ω; C) with
< u, v >Hλ(Ω;C):= ReZΩ(cid:0)∇u(x) · ∇v(x) − λ
u(x)v(x)
x2
(cid:1)dx,
∀u, v ∈ H 1
0 (Ω; C).
(3.31)
The spaces L2(Ω; C), H 1
In order to simplify the notations, we will write L2(Ω), H 1
0 (Ω; C), Hλ(Ω; C) inherit the properties of the corresponding real spaces.
0 (Ω), Hλ without making confusions.
As shown for the wave equation, the system (3.30) is well posed.
Theorem 3.5 (see [34]). Let T > 0 be given and assume λ ≤ λ(N ). For every u0 ∈ H
λ and any
h ∈ L2((0, T ) × Γ0) the system (3.30) is well-posed, i.e. there exists a unique weak solution such
that
′
u ∈ C([0, T ]; H
′
λ).
Moreover, there exists constant C > 0 such that the solution of (Sλ) satisfies
uL∞(0,T ;H ′
λ) ≤ C(u0H ′
λ
+ hL2((0,T )×Γ0)).
The system (3.30) is also controllable. More precisely, the control result states as follows.
Theorem 3.6. The system (Sλ) is controllable for any λ ≤ λ(N ). More precisely, for any time
λ there exists h ∈ L2((0, T ) × Γ0) such that the solution of (Sλ)
T > 0, u0 ∈ H
satisfies
u(T, x) = u0(x)
λ and u0 ∈ H
′
′
for all x ∈ Ω.
As discussed in Subsection 3.1, the controllability is equivalent to the Observability inequality
for the solution of the adjoint system
x2 = 0,
ivt + ∆v + λ v
v(t, x) = 0,
v(0, x) = v0(x),
(t, x) ∈ QT ,
(t, x) ∈ (0, T ) × Γ,
x ∈ Ω,
(3.32)
More precisely, if v solves (3.32), then for any time T > 0, there exists a positive constant CT
such that
v02
Hλ ≤ CT Z T
0 ZΓ0
2
(x · ν)(cid:12)(cid:12)(cid:12)
∂v
∂ν(cid:12)(cid:12)(cid:12)
dσdt.
(3.33)
Observability (3.33) might be deduced directly using the multiplier identity stated in Lemma 3.2.
The proof is let to the reader since it follows the same steps in [34].
Lemma 3.2. Assume λ ≤ λ⋆ and v is the solution of (3.32) corresponding to the initial data
v0 ∈ Hλ. Then
and v satisfies the identity
x2dσdt . v02
Hλ
2
∂v
∂ν(cid:12)(cid:12)(cid:12)
Z T
0 ZΓ(cid:12)(cid:12)(cid:12)
∂ν(cid:12)(cid:12)(cid:12)
(x · ν)(cid:12)(cid:12)(cid:12)
∂v
2
1
2Z T
0 ZΓ
dσdt = Tv2
Hλ +
1
2
ImZΩ
vx · ∇vdx(cid:12)(cid:12)(cid:12)
(3.34)
t=T
t=0
.
Remark 3.2. Besides, the proof of (3.33) can be deduced from the result valid for the wave
equation. Indeed, the general theory presented in an abstract form in [31], assure the observability
of systems like z = iA0z using results available for systems of the form z = −A0z.
25
4 Open problems
1. Geometric constraints. In this paper we have shown the role of the Pohozaev identity, in the
context of boundary singularities, when studying the controllability of conservative systems like
Wave and Schrodinger equations. We proved that for any λ ≤ λ(N ) = N 2/4, the corresponding
systems are exact observable from Γ0 precised in (1.11). Our result enlarges the range of values
λ ≤ (N − 2)2/4 for which the control holds, proved firstly in [34] in the context of interior
singularities.
The geometrical assumption for Γ0 is really necessary, otherwise our proof does not work. Of
course, it is still open to be analysed the case when the central of gravity of Γ0 is centered at a
point x0 different by zero, i.e. Γ0 = {x ∈ Γ (x − x0) · ν ≥ 0}. This choice of Γ0 provide some
technical difficulties which have been also emphasized in [34]. En eventually proof in the case of
a such domain Γ0 should apply a different technique that we have used so far.
2. Multipolar singularities. The same Pohozaev identity and controllability issues could be
address for more complicated operators, like for instance L = −∆ − V (x), where V (x) denotes a
multi-particle potential. To the best of our knowledge, even if there are some important works
studying Hardy-type inequalities for multipolar potentials (see e.g. [6] et al.), an accurate analysis
is still to be done. In a forthcoming work we study two particles systems. Our goal is to analyze
the limit process when one particle collapses to the other. We apply this both in the context of
controllability and the diffusion heat processes discussing the time decay of solutions.
5 Appendix: sharp bounds for x · ∇v(t)L2(Ω)
Proof of Theorem 1.1. Without losing the generality it is enough to consider two type of geome-
tries for Ω as follows.
G1: The points on Γ satisfy xN ≥ 0 in the neighborhood of origin.
G2; The points on Γ satisfy xN < 0 in the neighborhood of origin.
In the other intermediate case (when xN changes sign at origin) the result valid for case G2
still holds true since we can prove it for test functions extended with zero up to a domain satisfying
G2.
The proof follows several steps.
Step 1. Firstly we show that Theorem 1.1 is locally true. More precisely, there exists r0 =
r0(Ω, N ) > 0 small enough, and C = C(r0) such that
ZΩr0
x2∇v2dx ≤ R2
N 2
∇v2dx −
ΩhZΩr0
0 (Ωr0 ), where Ωr0 = Ω ∩ Br0(0).
4 ZΩr0
v2
x2 dxi + C(r0)ZΩr0
holds true for any function v ∈ C∞
v2dx,
(5.1)
Next we check the validity of Step 1. For that let us consider a function φ which satisfies
−∆φ ≥
N 2
4
φ
x2 , φ > 0,
∀x ∈ Ωr0 ,
for some positive constant r0. Such a function exists for each one of the case G1-G2. Indeed, for
the case G1 we may consider φ = xNx−N/2 and for case G2 we can take
1/2
φ = d(x)e(1−N )d(x)(cid:12)(cid:12)(cid:12)
log 1
x(cid:12)(cid:12)(cid:12)
x−N/2.
Next we introduce u such that v = φu. Then we get
∇v2 = ∇φ2u2 + φ2∇u2 + 2φu∇φ · ∇u.
26
Next, integrating we get
ZΩr0
∇v2dx =ZΩr0
∇u2φ2dx −ZΩr0
∆φ
φ
v2.
(5.2)
On the other hand, we obtain
ZΩr0
x2∇v2dx =ZΩr0
x2∇φ2u2dx +ZΩr0
x2φ2∇u2dx +
1
2ZΩr0
x2∇(φ2) · ∇(u2)dx (5.3)
Next we deduce
1
2ZΩr0
x2∇(φ2) · ∇(u2)dx = −ZΩr0
2
x · ∇φ
φ
v2dx −ZΩr0
x2∇φ2u2dx −ZΩr0
∆φ
φ x2v2dx.
According to (5.3) and (5.4) we obtain
ZΩr0
x2∇v2dx =ZΩr0
x2φ2∇u2dx −ZΩr0
2x · ∇φ
φ
v2dx −ZΩr0
∆φ
φ x2v2dx.
Let us write
∆φ
φ
−
=
N 2
4x2 + P,
(5.4)
(5.5)
(5.6)
where P ≥ 0 for any x ∈ Ωr0. Then from (5.2) we have
φ2∇u2 = R2
N 2
4
ΩZΩr0
ΩZΩr0 h∇v2 −
x2φ2∇u2dx ≤ R2
ZΩr0
= R2
From above and (5.5) it follows that
∇v2dx +ZΩr0
∆φ
φ
v2i
ΩhZΩr0
x2idx − R2
v2
ΩZΩr0
P v2dx.
(5.7)
ZΩr0
x2∇v2dx ≤ R2
− 2ZΩr0
ΩZΩr0 h∇v2 −
x · ∇φ
φ
v2
N 2
4
x2idx − R2
v2dx +ZΩr0 (cid:16) N 2
ΩZΩr0
P v2dx
4x2 + P(cid:17)x2v2dx
(x2 − R2
Ω)P v2dx
= R2
ΩZΩr0 h∇v2 −
− 2ZΩr0
x · ∇φ
φ
N 2
4
v2
x2idx +ZΩr0
4 ZΩr0
v2dx.
N 2
v2dx +
(5.8)
In the case G1 for r0 small enough we have P = 0 and
holds for some positive constant C. Thanks to (5.8) we conclude the proof of Step 1 in the cases
G1.
x · ∇φ
φ (cid:12)(cid:12)(cid:12) ≤ C,
(cid:12)(cid:12)(cid:12)
∀x ∈ Ωr0,
In the case G2, for r0 small enough we have
P > 0, ∇d · x ≥ 0,
∀x ∈ Ωr0
27
Then, we remark
x · ∇φ
φ
=
x · ∇d
d
+ O(1),
and from above we also finish the proof of Step 1 in this case.
Step 2. This step consist in applying a cut-off argument to transfer the validity of inequality
(5.1) from Ωr0 to Ω. More precisely, we consider a cut-off function θ ∈ C∞
0 (Ω) such that
Then we split v ∈ C∞
0 (Ω) as follows,
θ(x) =(cid:26) 1,
0,
x ≤ r0/2,
x ≥ r0.
(5.9)
Next let us firstly prove the following lema.
v = θv + (1 − θ)v := w1 + w2.
Lemma 5.1. Let us consider a weight function ρ : C∞(Ω) → R which is bounded and non
negative. There exists C(Ω, ρ) > 0 such that the following inequality holds
ZΩ
ρ(x)∇w1 · ∇w2dx ≥ −C(Ω, ρ, r)ZΩ v2dx.
(5.10)
Proof of Lemma 5.1. From the boundary conditions, integrating by parts we have
ZΩ
ρ∇w1 · ∇w2dx =Z ρθ(1 − θ)∇v2dx +ZΩ
1
≥
ρv∇v · ∇ρ(1 − 2θ)dx −Z θ∇θ2v2dx
∞ZΩ v2dx
∞ZΩ v2dx
2ZΩr0 \Ωr0/2 ∇(v2) · ∇θ(1 − 2θ)ρdx − ρ∞Dθ2
2ZΩr0 \Ωr0 /2
div((1 − 2θ)ρ∇θ)v2dx − ρ∞Dθ2
= −
≥ −C(ρW 1,∞ ,θW 2,∞ )ZΩ v2dx.
1
(5.11)
Now we are able to finalize the proof of Step 2. Indeed, splitting v as before we get
ZΩ x2∇v2dx =ZΩr0
x2∇w12dx +ZΩ\Ωr0 /2 x2∇w22dx + 2ZΩr0 \Ωr0 /2 x2∇w1 · ∇w2dx
Applying (5.1) to w1 we obtain
ZΩ x2∇v2dx ≤ R2
ΩhZΩ ∇v2dx −
N 2
4 ZΩ
w2
1
x2 dxi + CZΩ
−ZΩr0 \Ωr0 /2
2(R2
Ω − x2)∇w1 · ∇w2dx.
v2dx−
(5.12)
Adding ρ = 2(R2
Ω − x2)) in Lemma 5.1, from (5.12) we get
ZΩ x2∇v2dx ≤ R2
On the other hand we remark that
N 2
4 ZΩr0
w2
1
x2 dxi + C(Ω, r0)ZΩ
v2dx.
(5.13)
ΩhZΩ ∇v2dx −
x2 ≥ZΩ
ZΩr0
w2
1
v2
x2 dx − C(r0)ZΩ
v2dx.
(5.14)
From (5.13) and (5.14) the conclusion of Theorem 1.1 yields choosing r0 small enough, r0 ≤
RΩ.
28
Acknoledgements The author wish to thank Enrique Zuazua and Adimurthi for useful sugges-
tions and advices.
Partially supported by the Grants MTM2008-03541 and MTM2011-29306-C02-00 of the MICINN
(Spain), project PI2010-04 of the Basque Government, the ERC Advanced Grant FP7-246775
NUMERIWAVES, the ESF Research Networking Program OPTPDE, the grant PN-II-ID-PCE-
2011-3-0075 of CNCS-UEFISCDI Romania and a doctoral fellowship from UAM (Universidad
Aut´onoma de Madrid).
References
[1] Adimurthi, private communication.
[2] N. C. Adimurthi and M. Ramaswamy, An improved Hardy-Sobolev inequality and its appli-
cation, Proc. Amer. Math. Soc. 130 (2002), no. 2, 489-505 (electronic).
[3] C. Bardos, G. Lebeau, and J Rauch, Microlocal ideas in control and stabilization, Control of
boundaries and stabilization (Clermont-Ferrand, 1988), Lecture Notes in Control and Inform.
Sci., 125, 14 -- 30, Springer, Berlin, 1989.
[4] C. Bardos, G. Lebeau, and J. Rauch, Control and stabilization for hyperbolic equations.
Mathematical and numerical aspects of wave propagation phenomena (Strasbourg, 1991),
252 -- 266, SIAM, Philadelphia, PA, 1991.
[5] H. Berestycki and M. J. Esteban, Existence and bifurcation of solutions for an elliptic degen-
erate problem, J. Differential Equations 134 (1997), no. 1, 1-25.
[6] R. Bosi, J. Dolbeault, and J. Esteban, M., Estimates for the optimal constants in multipolar
Hardy inequalities for Schrodinger and Dirac operators, Commun. Pure Appl. Anal. 7 (2008),
no. 3, 533-562.
[7] H. Brezis, M. Marcus, and I. Shafrir, Extremal functions for Hardy's inequality with weight,
J. Funct. Anal. 171 (2000), no. 1, 177-191.
[8] H. Brezis and J. L. V´azquez, Blow-up solutions of some nonlinear elliptic problems, Rev. Mat.
Univ. Complut. Madrid 10 (1997), no. 2, 443-469.
[9] C. Cazacu, On Hardy inequalities with singularities on the boundary, C. R. Acad. Sci. Paris,
Ser. I, 349, 2011, 273-277,
[10] C. Cazacu, Hardy inequality and Pohozaev identity for operators with boundary singularities:
some applications, C. R. Acad. Sci. Paris, Ser. I, 349, 2011, 1167 -- 1172.
[11] J. D´avila and I. Peral, Nonlinear elliptic problems with a singular weight on the boundary,
Calc. Var. Partial Differential Equations, 41, 2011, 3-4, 567 -- 586.
[12] S. Ervedoza, Control and stabilization properties for a singular heat equation with an inverse-
square potential, Comm. Partial Differential Equations, 33, 2008, 10-12, 1996 -- 2019.
[13] L. C. Evans, Partial differential equations, Graduate Studies in Mathematics, 19, American
Mathematical Society, Providence, RI, 1998.
[14] M. M. Fall, On the Hardy Poincar´e inequality with boundary singularities. Commun. Con-
temp. Math, to appear.
[15] M. M. Fall and R. Musina, Hardy-Poincar´e inequality with boundary singularities. Proc. Roy.
Soc. Edinburgh, to appear.
29
[16] S. Filippas, V. Maz'ya, and A. Tertikas, On a question of Brezis and Marcus, Calc. Var.
Partial Differential Equations 25 (2006), no. 4, 491-501.
[17] S. Filippas, V. G. Maz'ya, and A. Tertikas, Sharp Hardy-Sobolev inequalities, C. R. Math.
Acad. Sci. Paris 339 (2004), no. 7, 483-486.
[18] A. V. Fursikov, and O. Yu. Imanuvilov, Controllability of evolution equations, Lecture Notes
Series, 34, Seoul National University Research Institute of Mathematics Global Analysis Re-
search Center,Seoul, 1996.
[19] A. Tertikas , S. Filippas, and J. Tidblom, On the structure of Hardy-Sobolev-Maz'ya inequal-
ities, J. Eur. Math. Soc. (JEMS) 11 (2009), no. 6, 1165-1185.
[20] J. P. Garc´ıa Azorero and A. I. Peral, Hardy inequalities and some critical elliptic and parabolic
problems, J. Differential Equations 144 (1998), no. 2, 441-476.
[21] N. Ghoussoub and X. S. Kang, Hardy-Sobolev critical elliptic equations with boundary sin-
gularities, Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, 21, 2004, 6, 767 -- 793.
[22] G. H. Hardy, J. E. Littlewood, and G. P´olya, Inequalities, Cambridge Mathematical Library,
Cambridge University Press, Cambridge, 1988, Reprint of the 1952 edition.
[23] D. Jerison and C. E. Kenig, Unique continuation and absence of positive eigenvalues for
Schrodinger operators, with an appendix by E. M. Stein, Ann. of Math. (2), 121, 1985, 3,
463-494.
[24] V. Komornik, Exact controllability and stabilization, RAM: Research in Applied Mathemat-
ics, The multiplier method, Masson, Paris, 1994,
[25] V. Komornik and E. Zuazua, A direct method for the boundary stabilization of the wave
equation, J. Math. Pures Appl. (9), 69, 1990, 1, 33 -- 54.
[26] J. M. L´evy-Leblond, Electron capture by polar molecules, Phys. Rev., 153, 1, 1967, 1 -- 4.
[27] J.-L. Lions, Controlabilit´e exacte, perturbations et stabilisation de syst`emes distribu´es.
Tome 1, Recherches en Math´ematiques Appliqu´ees [Research in Applied Mathematics], 8,
Controlabilit´e exacte. [Exact controllability], With appendices by E. Zuazua, C. Bardos, G.
Lebeau and J. Rauch, Masson, Paris, 1988.
[28] E. Machtyngier, Exact controllability for the Schrodinger equation, SIAM J. Control Optim.,
32, 1994, 1, 24 -- 34.
[29] D. Tataru, Unique continuation for solutions to PDE's; between Hormander's theorem and
Holmgren's theorem, Comm. Partial Differential Equations, 20, 1995, 5-6, 855 -- 884.
[30] A. Tertikas, Critical phenomena in linear elliptic problems, J. Funct. Anal. 154 (1998), no. 1,
42-66.
[31] M. Tucsnak and G. Weiss, Observation and control for operator semigroups, Birkhauser Ad-
vanced Texts: Basler Lehrbucher. [Birkhauser Advanced Texts: Basel Textbooks], Birkhauser
Verlag, Basel, 2009, xii+483.
[32] J. Vancostenoble, Lipschitz stability in inverse source problems for singular parabolic equa-
tions, preprint 2011.
[33] J. Vancostenoble, J. and E. Zuazua, Null controllability for the heat equation with singular
inverse-square potentials, J. Funct. Anal., 254, (2008) (7), 1864-1902.
[34] J. Vancostenoble and E. Zuazua, Hardy inequalities, observability, and control for the wave
and Schrodinger equations with singular potentials, SIAM J. Math. Anal., 41, 2009, 4, 1508-
1532.
30
[35] J. L. V´azquez and N. B. Zographopoulos, Functional aspects of the Hardy inequality. Ap-
pearance of a hidden energy, http://arxiv.org/abs/1102.5661.
[36] J. L. V´azquez and E. Zuazua, The Hardy inequality and the asymptotic behaviour of the
heat equation with an inverse-square potential, J. Funct. Anal. 173 (2000), no. 1, 103-153.
[37] E. Zuazua, Exact controllability for the semilinear wave equation, J. Math. Pures Appl. (9),
69, 1990, 1, 1 -- 31.
31
|
1606.08345 | 2 | 1606 | 2017-02-08T19:32:46 | Generalized notion of amenability for a class of matrix algebras | [
"math.FA"
] | We investigate the notions of amenability and its related homological notions for a class of $I\times I$-upper triangular matrix algebra, say $UP(I,A)$, where $A$ is a Banach algebra equipped with a non-zero character. We show that $UP(I,A)$ is pseudo-contractible (amenable) if and only if $I$ is singleton and $A$ is pseudo-contractible (amenable), respectively. We also study the notions of pseudo-amenability and approximate biprojectivity of $UP(I,A)$. | math.FA | math |
GENERALIZED NOTIONS OF AMENABILITY FOR A CLASS OF MATRIX
ALGEBRAS
A. SAHAMI
Abstract. We investigate the notions of amenability and its related homological notions for a class of
I × I-upper triangular matrix algebra, say U P (I, A), where A is a Banach algebra equipped with a non-
zero character. We show that U P (I, A) is pseudo-contractible (amenable) if and only if I is singleton
and A is pseudo-contractible (amenable), respectively. We also study the notions of pseudo-amenability
and approximate biprojectivity of U P (I, A).
1. Introduction and Preliminaries
B. E. Johnson studied the class of amenable Banach algebras. Indeed a Banach algebra A is amenable
if every continuous derivation D : A → X ∗ is inner, for every Banach A-bimodule X, that is, there exists
x0 ∈ X ∗ such that
D(a) = a · x0 − x0 · a (a ∈ A).
He also showed that A is amenable if and only if there exists a bounded net (mα) in A ⊗p A such that
a · mα − mα · a → 0,
πA(mα)a → a
(a ∈ A),
where πA : A ⊗p A → A is given by πA(a ⊗ b) = ab for every a, b ∈ A, see [16]. About the same time A.
Ya. Helemskii defined the homological notions of biflatness and biprojectivity for Banach algebras. In
fact a Banach algebra A is called biflat (biprojective), if there exists a bounded A-bimodule morphism
ρ : A → (A ⊗p A)∗∗ (ρ : A → A ⊗p A) such that π∗∗
A ◦ ρ is the canonical embedding of A into A∗∗ (ρ is a
right inverse for πA), respectively see [13]. Note that a Banach algebra A is amenable if and only if A is
biflat and A has a bounded approximate identity. It is known that for a locally compact group G, L1(G)
is biflat (biprojective) if and only if G is amenable(compact), respectively. Amenability of some matrix
algebras studied by Esslamzadeh [9] and also biflatness and biprojectivity of some semigroup algebras
related to matrix algebras investigated by Ramsden in [19].
Recently approximate versions of amenability and homological properties of Banach algebras have
been under more observations. In [24] Zhang introduced the notion of approximately biprojective Banach
algebras, that is, A is approximately biprojective if there exists a net of A-bimodule morphism ρα : A →
A ⊗p A such that
πA ◦ ρα(a) → a (a ∈ A).
Author with A. Pourabbas investigated approximate biprojectivity of 2 × 2 upper triangular Banach
algebra which is a matrix algebra, also we characterized approximate biprojectivity of Segal algebras and
weighted group algebras. We show that a Segal algebra S(G) is approximately biprojective if and only if
2010 Mathematics Subject Classification. Primary 46M10 Secondary, 43A07, 43A20.
Key words and phrases. Upper triangular Banach algebra, Amenability, Left φ-amenability, Approximate biprojectivity.
1
2
A. SAHAMI
G is compact and also we show that L1(G, w) is approximately biprojective if and only if G is compact,
provided that w ≥ 1 is a continuous weight function, see [21] and [23].
Approximate amenable Banach algebras have been introduced by Ghahramani and Loy.
Indeed a
Banach algebra A is approximate amenable if for every continuous derivation D : A → X ∗, there exists
a net (xα) in X ∗ such that
D(a) = lim
α
a · xα − xα · a (a ∈ A).
Other extensions of amenability are pseudo-amenability and pseudo-contractibility. A Banach algebra A
is pseudo-amenable (pseudo-contractible) if there exists a not necessarily bounded net (mα) in A ⊗p A
such that
a · mα − mα · a → 0,
(a · mα = mα · a),
πA(mα)a → a (a ∈ A).
For more information about these new concepts the reader referred to [12], [10] and [11]. Recently in [7]
and [8] pseudo-amenability and pseudo-contractibility of certain semigroup algebras, using the properties
of matrix algebras, have been studied.
In this paper, we investigate amenability and its related homological notions for a class of matrix
algebras. We show that for a Banach algebra A with a non-zero character, I × I upper triangular
Banach algebra U P (I, A) is pseudo-contractible (amenable) if and only if I is singleton and A is pseudo-
contractible (amenable), respectively. Also we characterize whether U P (I, A) is approximate amenable,
pseudo-amenable and approximate biprojective. The paper concluded by studying amenability and ap-
proximate biprojectivity of some semigroup algebras related to a matrix algebra.
We remark some standard notations and definitions that we shall need in this paper. Let A be a
Banach algebra. Throughout this paper the character space of A is denoted by ∆(A), that is, all non-
zero multiplicative linear functionals on A. Let A be a Banach algebra. The projective tensor product
A ⊗p A is a Banach A-bimodule via the following actions
a · (b ⊗ c) = ab ⊗ c, (b ⊗ c) · a = b ⊗ ca
(a, b, c ∈ A).
Let A be a Banach algebra and I be a non-empty set. U P (I, A) is denoted for the set of all I × I
upper triangular matrices which entries come from A and
With the usual matrix operations and · as a norm, U P (I, A) becomes a Banach algebra.
(ai,j)i,j∈I = Xi,j∈I
ai,j < ∞.
2. a class of matrix algebras and generalized notions of amenability
In this section we investigate generalized notions of amenability for upper triangular Banach algebras.
We remind that a Banach algebra A with φ ∈ ∆(A) is called left(right) φ-contractible, if there exists
m ∈ A such that am = φ(a)m(ma = φ(a)m) and φ(m) = 1 for every a ∈ A, respectively. For more
information the reader referred to [18].
Theorem 2.1. Let I be a non-empty set and A be a unital Banach algebra with ∆(A) 6= ∅. U P (I, A) is
pseudo-contractible if and only if I is singleton and A is pseudo-contractible.
GENERALIZED NOTIONS OF AMENABILITY FOR A CLASS OF MATRIX ALGEBRAS
3
Proof. Let U P (I, A) be pseudo-contractible. Then U P (I, A) has a central approximate identity, say (eα).
Put Fi,j for a matrix belongs to U P (I, A) which (i, j)-th entry is eA and others are zero, where eA is
an identity of A. Thus Fi,jeα = eαFi,j for every i, j ∈ I. This equation implies that the entries on main
diagonal of eα is equal. Suppose conversely that I is infinite. Since the entries on main diagonal of eα are
equal, it implies that eα = ∞ or the main diagonal of eα is zero. In the case eα = ∞, eα does not
belong to U P (I, A) which is impossible. Otherwise if the main diagonal of eα is zero, then eαFi,i = 0.
Thus 0 = eαFi,i → Fi,i which is impossible, hence I must be finite. Suppose that I = {i1, i2, ..., in}
and φ ∈ ∆(A). Define ψ ∈ ∆(U P (I, A)) by ψ((ai,j )i,j∈I ) = φ(ain ,in ) for every (ai,j) ∈ U P (I, A). Since
U P (I, A) is pseudo-contractible, by [2, Theorem 1.1] U P (I, A) is left and right ψ-contractible. Set
J = {(ai,j) ∈ U P (I, A)ai,j = 0, for all
j 6= in}.
It is clear J is a closed ideal of U P (I, A) and ψJ 6= 0, hence by [18, Proposition 3.8] J is left and
right ψ-contractible. So there exist m1, m2 ∈ J such that jm1 = ψ(j)m1 and m2j = ψ(j)m2 and also
ψ(m1) = ψ(m2) = 1 for each j ∈ J. Set m = m1m2 ∈ J. Clearly we have
(2.1)
jm = mj = ψ(j)m, ψ(m) = ψ(m1m2) = ψ(m1)ψ(m2) = 1,
(j ∈ J).
Suppose conversely that I > 1. Set m for the matrix with n-th columns (x1, x2, ..., xn)t, where xi ∈ A
for each i ∈ {1, 2, ..., n}. Let a be an element of J which its n-th columns has the form (0, 0, ..., an)t for
an arbitrary element an ∈ A. Applying (2.1) we have
x1an = x2an = ... = xn−1an = 0, φ(an)x1 = φ(an)x2 = ... = φ(an)xn−1 = 0,
and also
anxn = xnan = φ(an)xn, φ(xn) = 1.
Pick an element an ∈ A such that φ(an) = 1. Applying (2.1) follows that x1 = x2 = ... = xn−1 = 0. Then
m becomes a matrix which its n-th columns has the form (0, 0, ..., 0, xn)t. Set b for a matrix in J which its
n-th columns has the form (b1, b2, ..., bn−1, bn)t, where bn ∈ ker φ and φ(b1) = φ(b2) = ... = φ(bn−1) = 1.
Applying (2.1) we have a1xn = 0. Taking φ on this equation we have 0 = φ(a1xn) = φ(a1)φ(xn) = 1
which is a contradiction. Therefore I must be singleton. So A is pseudo-contractible.
Converse is clear.
(cid:3)
Suppose that A is a Banach algebra and φ ∈ ∆(A). A is called (approximately) left φ-amenable if
there exists (a not necessarily) bounded net (mα) in A such that
amα − φ(a)mα → 0 φ(mα) → 1
(a ∈ A),
respectively. Right cases define similarly. For more information about these new concepts of amenability
and its related homological notions see [1], [17], [14] and [22].
Theorem 2.2. Let I be an ordered set with an smallest element. Also let A be a Banach algebra with
a left unit such that ∆(A) 6= ∅. U P (I, A) is pseudo-amenable (approximate amenable) if and only if I is
singleton and A is pseudo-amenable(approximate amenable), respectively.
4
A. SAHAMI
Proof. Here we proof the pseudo-amenable case, approximate amenability is similar. Suppose that
U P (I, A) is pseudo-amenable. Then there exists a net (mα) in U P (I, A) ⊗p U P (I, A) such that
a · mα − mα · a → 0,
πU P (I,A)(mα)a → a
(a ∈ U P (I, A)).
Let i0 be a smallest element of I. It is easy to see that ψ given by ψ(a) = φ(ai0,i0 ) is a character on
U P (I, A), for each a = (ai,j) ∈ U P (I, A). Define
T : U P (I, A) ⊗p U P (I, A) → U P (I, A)
by T (a ⊗ b) = ψ(a)b for each a, b ∈ U P (I, A). It is easy to see that T is a bounded linear map which
satisfies the following:
T (a · x) = ψ(a)T (x),
T (x · a) = T (x)a, ψ ◦ T (x) = ψ ◦ πU P (I,A)(x),
for each a, b ∈ U P (I, A) and x ∈ U P (I, A) ⊗p U P (I, A). Thus we have
ψ(a)T (mα) − T (mα)a = T (a · mα − mα · a) → 0
and ψ ◦ T (mα) = ψ ◦ πU P (I,A)(mα) → 1. Hence U P (I, A) is approximately right ψ-amenable. Using the
same arguments as in the proof of Theorem 2.1 and applying [20, Proposition 5.1] one can see that I is
singleton and A is pseudo-amenable.
Converse is clear.
(cid:3)
Let A be a Banach algebra and a ∈ A. By aεi,j we mean a matrix belongs to U P (I, A) with (i, j)-th
place is a and zero elsewhere.
Theorem 2.3. Let I be non-empty set and A be a Banach algebra such that ∆(A) 6= ∅. U P (I, A) is
amenable if and only if I is singleton and A amenable.
Proof. Let U P (I, A) be amenable. Then U P (I, A) has a bounded approximate identity, say (Eα). Let
M > 0 be a bound for (Eα). We claim that A has a bounded left approximate identity. To see this, fix
k, l ∈ I. Then for each a ∈ A, we have
0 = lim
α
(2.2)
Eαaεk,l − aεk,l = lim
α
Eα
i,jεi,j)aεk,l − aεk,l
(Xi,j
Xi
(Xi6=k
= lim
α
= lim
α
Eα
i,laεi,l − aεk,l
Eα
i,la + Eα
k,la − a.
Thus eα = Eα
k,l is a left approximate identity of A. It is easy to see that eα ≤ Eα ≤ M . So (eα) is a
bounded left approximate identity for A. We claim that I is finite. Suppose conversely that I is infinite.
Pick a ∈ A such that a = 1. Since (eα) is a bounded left approximate identity for A, then limα eαa = a,
for each a ∈ A. Thus there exists a αl,k such that α ≥ αk,l such that 1
2 < eαa. Hence for α ≥ αk,l we
have
(2.3)
1
2
< eαa ≤ eα = Eα
k,l.
GENERALIZED NOTIONS OF AMENABILITY FOR A CLASS OF MATRIX ALGEBRAS
5
Since I is infinite we can choose N ∈ N such that N > 2M. Then choose distinct k1, l1, k2, l2, ..., kN , lN
in I and α ≥ αki,li , i = 1, 2, ..., N . Using (2.3) one can see that
M <
1
2
N =
N
Xi=1
Eα
ki,li ≤ Xi,j∈I
Eα
i,j ≤ M,
which is a contradiction. So I is finite.
Applying the same method as in the proof of previous Theorem, it is easy to see that I must be
singleton, then A is amenable.
(cid:3)
3. a class of Matrix algebra and approximate biprojectivity
In this section we study approximate biprojectivity of some matrix algebra. We also investigate the
relation of approximate biprojectivity and discreteness of maximal ideal space of a Banach algebra.
Theorem 3.1. Let I be an ordered set with an smallest element. Also let A be a Banach algebra with
a right identity such that ∆(A) 6= ∅. U P (I, A) is approximately biprojective if and only if I is singleton
and A is approximately biprojective.
Proof. Let i0 be smallest element of I. Define ψ ∈ ∆(U P (I, A)) by ψ(a) = φ(ai0,i0 ), where a = (ai,j ) ∈
U P (I, A). Suppose that U P (I, A) is approximately biprojective. Since A has a right identity, by [20,
Lemma 5.2], U P (I, A) has a right approximate identity. Applying [23, Theorem 3.9], U P (I, A) is right
ψ-contractible. Using the same arguments as in the proof of the Theorem 2.1, I is singleton and A is
approximately biprojective.
Converse is clear.
(cid:3)
Remark 3.2. Let A be a Banach algebra with a left approximate identity and I be a finite set which
has at least two elements. Then U P (I, A) is never approximately biprojective. To see this, since I =
{i1, i2, ..., in} is finite then left approximate identity of A gives a left approximate identity for U P (I, A).
Define ψ ∈ ∆(U P (I, A)) by ψ(a) = φ(ain ,in ) for every a = (ai,j) ∈ U P (I, A). By [23, Theorem 3.9]
approximate biprojectivity of U P (I, A) implies that U P (I, A) is left ψ-contractible, then the rest is
similar to the proof of Theorem 2.1.
Proposition 3.3. Let A be a Banach algebra with a left approximate identity and ∆(A) be a non-empty
set. If A is approximately biprojective, then ∆(A) is discrete with respect to the w∗-topology.
Proof. Since A is an approximately biprojective Banach algebra with a left approximate identity, by [23,
Theorem 3.9] A is left φ-contractible for every φ ∈ ∆(A). Applying [4, Corollary 2.2] one can see that
∆(A) is discrete.
(cid:3)
Corollary 3.4. Let A be a Banach algebra with a left identity, φ ∈ ∆(A) and let I be a non-empty set.
If U P (I, A) is approximate biprojective, then ∆(U P (I, A)) is discrete with respect to the w∗-topology.
Proof. Note that, since φ ∈ ∆(A), ∆(U P (I, A)) is a non-empty set. Existence of left identity for A implies
that U P (I, A) has a left approximate identity, see [20, Lemma 5.2]. Applying previous Proposition one
can see that ∆(U P (I, A)) is discrete with respect to the w∗-topology.
(cid:3)
6
A. SAHAMI
Let A be a Banach algebra and φ ∈ ∆(A). A is φ-inner amenable if there exists a bounded net (aα)
in A such that
aaα − aαa → 0, φ(aα) → 1
(a ∈ A).
For more information about φ-inner amenability, see [15].
Lemma 3.5. Let A be a Banach algebra and φ ∈ ∆(A). Suppose that A has an approximate identity.
Then approximate biprojectivity of A implies that A is φ-inner amenable.
Proof. Suppose that A is approximate biprojective. Using [23, Theorem 3.9], existence of approximate
identity implies that A is left and right φ-contractible. Then there exist m1 and m2 in A such that
am1 = φ(a)m1(m2a = φ(a)m2) φ(m1) = φ(m2) = 1
(a ∈ A),
respectively. Since
one can see that
m1 = φ(m2)m1 = m2m1 = φ(m1)m2 = m2,
am1 = m1a = φ(a)m1 φ(m1) = 1, (a ∈ A).
It follows that A is φ-inner amenable.
(cid:3)
Remark 3.6. There exists a matrix algebra which is approximate biprojective but it is not φ-inner
amenable. Then the converse of previous Lemma is not always true.
To see this, let A = 0 C
0 C ! and also let a0 = 0
0
1
1 !. Define ρ : A → A ⊗p A by ρ(a) = a ⊗ a0
for every a ∈ A. It is easy to see that ρ is a bounded A-bimodule morphism and
πA ◦ ρ(a) = a,
(a ∈ A).
Then A is biprojective and it follows that A is approximate biprojective. Set φ( 0 a
b !) = b for every
0
a, b ∈ C.
conversely that A is φ-inner amenable. Then there exists a bounded net (aα) in A such that
It is easy to see that φ ∈ ∆(A). We claim that A is not φ-inner amenable. We suppose
aaα − aαa → 0, φ(aα) → 1
(a ∈ A).
It is easy to see that ab = φ(b)a for every a ∈ A. Hence we have
0 = lim
α
a0aα − aαa0 = lim φ(aα)a0 − φ(a0)aα = lim a0 − aα,
It follows that a0 = lim aα. Hence for each a ∈ A, we have
aa0 = a0a, φ(a0) = 1.
It follows that a = φ(a)a0. Thus dim A = 1 which is a contradiction.
GENERALIZED NOTIONS OF AMENABILITY FOR A CLASS OF MATRIX ALGEBRAS
7
4. Examples of semigroup algebras related to the matrix algebras
Example 4.1. Suppose that A is a Banach algebra and I is a non-empty set. Put B = U P (I, A). It
is obvious that B with matrix multiplication can be observed as a semigroup. Equip this semigroup
with the discrete topology and denote it with SB. Suppose that A has a non-zero idempotent. We claim
that ℓ1(SB) is not amenable, whenever I is an infinite set. Suppose conversely that ℓ1(SB) is amenable.
Let e be an idempotent for A. Ei,i for a matrix belongs to B which its (i, i)-th entry is e, otherwise
is 0. It is easy to see that Ei,i is an idempotent for the semigroup SB, for every i ∈ I. So the set of
idempotents of SB is infinite, whenever I is infinite. Thus by [6, Theorem 2] ℓ1(SB) is not amenable
which is contradiction.
Suppose that A is a Banach algebra with a left identity, also suppose that I is an ordered set with
smallest element. We also claim that ℓ1(SB) is never approximate biprojective. To see this suppose
conversely that ℓ1(SB) is approximately biprojective. We denote augmentation character on ℓ1(SB) by
φSB . It is easy to see that δ0 ∈ SB and φSB (δ0) = 1, where 0 is denoted for the zero matrix belongs to
SB. One can see that the center of SB, say Z(SB), is non-empty, because 0 belongs to Z(SB). Using [23,
Proposition 3.1(ii)], one can see that ℓ1(SB) is left φSB -contractible. Let i0 be an smallest element of I.
Define
J = {(ai,j) ∈ SBai,j = 0, for all
i 6= i0},
it is easy to see that J is an ideal of SB, then by [5, page 50] ℓ1(J) is a closed ideal of ℓ1(SB). Since
φSB ℓ1(J) is non-zero, ℓ1(J) is left φSB -contractible. Thus there exists m ∈ ℓ1(J) such that am = φSB (a)m
and φSB (m) = 1, for every a ∈ A. On the other hand since A has a left identity, then J has a left identity.
Thus by the same argument as in the proof of [23, Proposition 3.1(ii)] we have
m(j) = m(elj) = δjm(el) = φSB (δj)m(el) = m(el)
(j ∈ J),
where el is a left unit for J. It follows that m is a constant function belongs to ℓ1(J). Since φSB (m) = 1,
then m 6= 0 which implies that J is finite which is impossible.
References
[1] H. P. Aghababa, L. Y. Shi and Y. J. Wu; Generalized notions of character amenability Act. Math.
Sin, 29 (2013) 1329-1350.
[2] M. Alaghmandan, R. Nasr-Isfahani and M. Nemati; On φ-contractibility of the Lebesgue-Fourier
algebra of a locally compact group, Arch. Math. 95 (2010), 373-379.
[3] Y. Choi, F. Ghahramani and Y. Zhang; Approximate and pseudo-amenability of various classes of
Banach algebras, J. Func. Anal. 256 (2009) 3158-3191.
[4] M. Dashti, R. Nasr Isfahani and S. Soltani Renani; Character Amenability of Lipschitz Algebras,
Canad. Math. Bull. Vol. 57 (1), 2014 pp. 37-41.
[5] H. G. Dales, A.T.-M. Lau and D. Strauss; Banach algebras on semigroups and on their compactifi-
cations Mem. Amer. Math. Soc. 205, (2010).
[6] J. Duncan and A. L. T. Paterson; Amenability for discrete convolution semigroup algebras, Math.
Scand. 66 (1990) 141-146.
8
A. SAHAMI
[7] M. Essmaili, M. Rostami, and A. R. Medghalchi; Pseudo-contractibility a nd pseudo-amenability of
semigroup algebras, Arch. Math. 97 (2011), 167-177.
[8] M. Essmaili, M. Rostami, and A. Pourabbas; Pseudo-amenability of certain semigroup algebars,
Semigroup Forum 82 (2011), 478-484.
[9] G. H. Esslamzadeh; Double centralizer algebras of certain Banach algebras, Monatsh. Math.142
(2004), 193-203.
[10] F. Ghahramani and R. J. Loy; Generalized notions of amenability, J. Func. Anal. 208 (2004), 229-
260.
[11] F. Ghahramani, R. J. Loy and Y. Zhang; Generalized notions of amenability II, J. Func. Anal. 254
(2008), 1776-1810.
[12] F. Ghahramani, Y. Zhang; Pseudo-amenable and pseudo-contractible Banach algebras, Math. Proc.
Cambridge Philos. Soc. 142 (2007) 111-123.
[13] A. Ya. Helemskii; The homology of Banach and topological algebras, Kluwer, Academic Press,
Dordrecht, 1989.
[14] Z. Hu, M. S. Monfared and T. Traynor; On character amenable Banach algebras, Studia Math. 193
(2009) 53-78.
[15] A. Jabbari, T. Mehdi Abad and M. Zaman Abadi; On φ-inner amenable Banach algebras, Colloq.
Math. vol 122 (2011) 1-10.
[16] B. E. Johnson; Cohomology in Banach algebras, Mem. Amer. Math. Soc. 127 (1972).
[17] E. Kaniuth, A. T. Lau and J. Pym; On φ-amenability of Banach algebras, Math. Proc. Cambridge
Philos. Soc. 144 (2008) 85-96.
[18] R. Nasr Isfahani and S. Soltani Renani; Character contractibility of Banach algebras and homological
properties of Banach modules, Studia Math. 202 (2011) 205-225.
[19] P. Ramsden; Biflatness of semigroup algebras, Semigroup Forum 79, (2009) 515-530.
[20] A. Sahami; On biflatness and φ-biflatness of some Banach algebras (preprint).
[21] A. Sahami and A. Pourabbas; Approximate biprojectivity and φ-biflatness of some Banach algebras,
Colloq. Math, In press.
[22] A. Sahami and A. Pourabbas; On φ-biflat and φ-biprojective Banach algebras, Bull. Belg. Math.
Soc. Simon Stevin, 20(2013) 789-801.
[23] A. Sahami and A. Pourabbas; Approximate biprojectivity of certain semigroup algebras, Semigroup
Forum, 92(2016) 474-485.
[24] Y. Zhang; Nilpotent ideals in a class of Banach algebras, Proc. Amer. Math. Soc. 127 (11) (1999),
3237-3242.
Faculty of Basic sciences, Department of Mathematics, Ilam University, P.O.Box 69315-516, Ilam, Iran.
E-mail address: [email protected]
|
1702.06667 | 1 | 1702 | 2017-02-22T04:06:35 | Morse theory methods for quasi-linear elliptic systems of higher order | [
"math.FA",
"math.AP"
] | We develop the local Morse theory for a class of non-twice continuously differentiable functionals on Hilbert spaces, including a new generalization of the Gromoll-Meyer's splitting theorem and a weaker Marino-Prodi perturbation type result. With them some critical point theorems and famous bifurcation theorems are generalized. Then we show that these are applicable to studies of quasi-linear elliptic equations and systems of higher order given by multi-dimensional variational problems as in (1.3). | math.FA | math | Morse theory methods for quasi-linear elliptic systems
of higher order ∗
Guangcun Lu
February 22, 2017
7
1
0
2
b
e
F
2
2
]
.
A
F
h
t
a
m
[
1
v
7
6
6
6
0
.
2
0
7
1
:
v
i
X
r
a
Abstract
We develop the local Morse theory for a class of non-twice continuously differentiable
functionals on Hilbert spaces, including a new generalization of the Gromoll-Meyer's split-
ting theorem and a weaker Marino-Prodi perturbation type result. With them some critical
point theorems and famous bifurcation theorems are generalized. Then we show that these are
applicable to studies of quasi-linear elliptic equations and systems of higher order given by
multi-dimensional variational problems as in (1.3).
Contents
1 Introduction
I Abstract theories
2 Local Morse theory for a class of non-C 2 functionals
2.1 Statements of main results
. .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
2.2 Lemmas .
. . . . .
. . . . . .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
2.3 Proof of Theorem 2.1 . . . . .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
2.4 An implicit function theorem for a family of potential operators
. . . .
. . . . .
2.5 Parameterized splitting and shifting theorems
. . .
. . . . . .
. . . . .
. . . . .
2.6 Splitting and shifting theorems around critical orbits . . . . . .
. . . . .
. . . . .
2.7 Proof of Theorem 2.6 . . . . .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
2
7
8
8
12
13
15
26
34
42
∗Partially supported by the NNSF 11271044 of China.
F1. Lu: School of Mathematical Sciences, Beijing Normal University, Laboratory of Mathematics and Complex
Systems, Ministry of Education, Beijing 100875, The People's Republic of China; e-mail: [email protected]
Mathematics Subject Classification (2010): Primary 58E05, 49J52, 49J45
1
2
Guangcun Lu
3 Bifurcations for potential operators
3.1 Generalizations of a bifurcation theorem by Chow and Lauterbach . . .
. . . . .
3.2 Generalizations of Rabinowitz bifurcation theorem . . . . . .
. . . . .
. . . . .
3.3 Bifurcation for equivariant problems . . .
. . . . .
. . . . . .
. . . . .
. . . . .
II Applications to quasi-linear elliptic systems of higher order
4 Fundamental analytic properties for functionals F and F
4.1 Results and preliminaries . . .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
4.2 Proof for A) of Theorem 4.1 .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
4.3 Proof for B) of Theorem 4.1 .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
4.4 Proof for C) of Theorem 4.1 .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
4.5 Proof for D) of Theorem 4.1 .
. . . . . .
. . . . .
. . . . . .
. . . . .
. . . . .
5 (PS)- and (C)-conditions
6 Morse inequalities
7 Bifurcations for Quasi-linear elliptic systems
8 Concluding remarks
A Proof of Proposition 4.3
1 Introduction
45
46
47
56
68
68
68
72
75
81
85
87
90
92
99
99
Since Palais and Smale [53, 55, 61] generalized finite-dimensional Morse theory in [51] to nonde-
generate C 2 functionals on infinite dimensional Hilbert manifolds and used it to study multiplicity
of solutions for semilinear elliptic boundary value problems, via Gromoll and Meyer [31], Marino
and Prodi [49] and many other people's effort, such a direction has very successful developments,
see a few of nice books [3, 12, 13, 14, 50, 52, 56, 58, 70] and references therein for details. The
Morse theory for functionals on an infinite dimensional Hilbert space H have two main aspects:
Morse relations related critical groups to Betti numbers of underlying spaces (global), computa-
tion of critical groups (local). The basic tool for the latter, Gromoll-Meyer's generalized Morse
lemma (or splitting theorem) in [31], was only generalized to C 2 functionals on Hilbert spaces
([13, 50]) until author's recent work [39, 40]. Because of this, most of applications of the theory
to differential equations are restricted to semi-linear elliptic equations and Hamiltonian systems
[13, 50, 52]. Skrypnik [58] established Morse inequalities for the functional (1.8) with p = 2 and
V = W m,2
at any solution of it have no nontrivial solutions. Our new splitting lemmas in [39, 40] can be
(Ω) provided that linearizations of the corresponding Euler-Lagrange equation (1.9)
0
***
3
effectively used to study periodic solutions of Lagrangian systems on compact manifolds which
are strongly convex and has quadratic growth on the fibers, including the case of the system (1.4)
if n = 1 and Hypothesis F2,N was satisfied. Their ideas were also used to derive the desired split-
ting and shifting lemmas for the Finsler energy functional on the space of H 1-curves in [41, 43].
However, when applying these splitting lemmas to the functional in (1.8) with p = 2 we need that
the involved critical points have higher smoothness, which can only be guaranteed under more
assumptions on Lagrangian F by the regularity theory of differential equations. It is this unsatis-
factory restriction that motivates us to look for a more suitable splitting lemma which is applicable
to the functional in (1.3) under Hypothesis Fp,N with p = 2.
The following notation will be used throughout this paper. For normed linear spaces X, Y we
denote by X ∗ the dual space of X, and by L (X, Y ) the space of linear bounded operators from
X to Y . We also abbreviate L (X) := L (X, X). The open ball in a normed linear space X with
radius r and center in y ∈ X is denoted by BX (y, r) := {x ∈ X kx − ykX < r} and the cor-
responding closed ball is written as ¯BX(y, r) := {x ∈ X kx − ykX ≤ r}. The (norm)-closure
of a set S ⊂ X will be denoted by S or Cl(S). Let m and n be two positive integers, Ω ⊂ Rn a
bounded domain with boundary ∂Ω. Denote the general point of Ω by x = (x1,··· , xn) ∈ Rn and
the element of Lebesgue n-measure on Ω by dx. A multi-index is an n-tuple α = (α1,··· , αn) ∈
(N0)n, where N0 = N ∪ {0}. α := α1 + ··· + αn is called the length of α. Denote by M (k)
the number of such α of length α ≤ k, M0(k) = M (k) − M (k − 1), k = 0,··· , m, where
M (−1) = ∅. Then M (0) = M0(0) only consists of 0 = (0,··· , 0) ∈ (N0)n.
Let p ∈ [2,∞) be a real number and N ≥ 1 an integer.
Hypothesis Fp,N . For each multi-index γ as above, let
pγ ∈ (1,∞) if γ = m − n/p,
qγ = 1 if γ < m − n/p,
and qγ =
and pγ =
pγ
pγ − 1
np
n − (m − γ)p
if m − n/p < γ ≤ m,
if m − n/p ≤ γ ≤ m;
and for each two multi-indexes α, β as above, let pαβ = pβα be defined by the conditions
pαβ = 1 − 1
pα − 1
pβ
if α = β = m,
pαβ = 1 − 1
,
if m − n/p ≤ α ≤ m, β < m − n/p,
pα
pαβ = 1
pα − 1
pβ
0 < pαβ < 1 − 1
if α,β < m − n/p,
if α, β ≥ m − n/p, α + β < 2m.
For M0(k) = M (k) − M (k − 1), k = 0, 1,··· , m as above, we write ξ ∈Qm
ξ = (ξ0,··· , ξm), where ξ0 = (ξ1
k=0
RN ×M0(k) as
0,··· , ξN
α(cid:1) 1 ≤ i ≤ N
α = k
0 )T ∈ RN and
∈ RN ×M0(k)
ξk =(cid:0)ξi
for k = 1,··· , m.
4
Guangcun Lu
Denote by ξk
◦ = {ξk
α : α < m − n/p} for k = 1,··· , N . Suppose
Ω ×
mYk=0
RN ×M0(k) ∋ (x, ξ) 7→ F (x, ξ) ∈ R
is a Caratheodory function (i.e., being measurable in x for all values of ξ, and continuous in ξ for
almost all x) with the following properties:
(i) F (x, ξ) is twice continuously differentiable in ξ for almost all x, F (·, 0) ∈ L1(Ω) and
F i
α(x, ξ) :=
∂F (x, ξ)
∂ξi
α
,
i = 1,··· , N,
α ≤ m
α(·, 0) ∈ L1(Ω) if α < m − n/p, and F i
satisfy: F i
i = 1,··· , N .
(ii) There exists a continuous, positive, nondecreasing functions g1 such that for i, j = 1,··· , N ,
α,β ≤ m and the above numbers pαβ functions
α(·, 0) ∈ Lqα(Ω) if m − n/p ≤ α ≤ m,
Ω × RM (m) → R, (x, ξ) 7→ F ij
αβ(x, ξ) =
∂2F (x, ξ)
α∂ξj
∂ξi
β
satisfy:
F ij
αβ(x, ξ) ≤ g1(
NXk=1
(iii) There exists a continuous, positive, nondecreasing functions g2 such that
NXk=1 Xm−n/p≤γ≤m
ξk
◦)1 +
◦) 1 +
NXk=1
ξk
.
pαβ
ξk
γpγ
γ!p−2 NXi=1 Xα=m
(ηi
α)2
(1.1)
(1.2)
Xα=β=m
F ij
αβ(x, ξ)ηi
αηj
β ≥ g2(
NXk=1 Xγ=m
ξk
for any η = (ηij
αβ) ∈ RN ×M0(m).
Note: If m ≤ n/p the functions g1 and g2 should be understand as positive constants.
For an element of W m,p(Ω, RN ), ~u = (u1,··· , uN ) : Ω → RN , we shall denote by Dk~u
the set {Dαui
α = k, i = 1,··· , N} for each k = 1,··· , m, and form the expression
F (x, ~u(x),··· , Dm~u(x)), in which ~u(x) takes the place of ξ0, and Dαui(x) takes the place of ξi
α
respectively. Let V ⊂ W m,p(Ω, RN ) be a closed subspace containing W m,p
(Ω, RN ). Consider
0
:
variational problem
F(~u) =ZΩ
F (x, ~u,··· , Dm~u)dx,
~u ∈ V.
(1.3)
We call critical points of F generalized solutions of the boundary value problem corresponding to
the subspace V :
Xα≤m
(−1)αDαF i
α(x, ~u,··· , Dm~u) = 0,
i = 1,··· , N.
(1.4)
***
5
If N = 1, Hypothesis Fp,N can be written as the following simple version, which was first
given in [58].
Hypothesis fp. Let p, pα, qα, pαβ and Ω be as in Hypothesis Fp,N . Write ξ ∈ RM (m) as
ξ = {ξα : α ≤ m} and ξ◦ = {ξα : α < m − n/p}. Suppose that f : Ω × RM (m) → R is a
Caratheodory function with the following properties:
(i) f (x, ξ) is twice continuously differentiable in ξ for almost all x, f (·, 0) ∈ L1(Ω) and each
fα(x, ξ) := ∂f (x,ξ)
satisfy: fα(·, 0) ∈ L1(Ω) if α < m − n/p, and fα(·, 0) ∈ Lqα(Ω) if
∂ξα
m − n/p ≤ α ≤ m.
(ii) There exists a continuous, positive, nondecreasing functions g1 such that for the above numbers
pαβ functions
Ω × RM (m) → R, (x, ξ) 7→ fαβ(x, ξ) =
∂2f (x, ξ)
∂ξα∂ξβ
satisfy:
(iii) There exists a continuous, positive, nondecreasing functions g2 such that
fαβ(x, ξ) ≤ g1(ξ◦)1 + Xm−n/p≤γ≤m
fαβ(x, ξ)ηαηβ ≥ g2(ξ◦) 1 + Xγ=m
Xα=β=m
for any η ∈ RM0(m).
.
pαβ
ξγpγ
ξγ!p−2 Xα=m
(1.5)
η2
α
(1.6)
Consider the case m = 1 and n ≥ 2. Then M = n + 1 and f becomes
f : Ω × R1 × Rn → R, (x, ξ0, ξ1,··· , ξn) 7→ f (x, ξ0, ξ1,··· , ξn).
The corresponding Hypothesis f2 is: there exist constant numbers c1, c2 > 0 such that
fij(x, ξ) ≤ c1 1 +
ξk2k!2ij
, Xi,j=1
fij(x, ξ)ηiηj ≥ c2 nXi=1
i!.
η2
nXk=0
(1.7)
Here, (I) if n = 2, 20 = s ∈ (2,∞), 2i = 2, i = 1,··· , n, 2ij = 0 for i, j = 1,··· , n,
20i = 2i0 ∈ (0, 1/2−1/s) for i = 1,··· , n, and 200 ∈ (0, 1−2/s); (II) if n > 2, 20 = 2n/(n−2),
2i = 2 for i = 1,··· , n, 2ij = 0 for i, j = 1,··· , n, 2i0 = 20i ∈ (0, 1/n) for i = 1,··· , n,
200 ∈ (0, 2/n).
Under Hypothesis fp, let V ⊂ W m,p(Ω) be any closed subspace containing W m,p
(Ω). The
0
critical points of the variational problem
F(u) =ZΩ
f (x, u,··· , Dmu)dx
(1.8)
6
Guangcun Lu
in the Banach space V are called generalized solutions of the boundary value problem correspond-
ing to the subspace V :
(−1)αDαfα(x, u,··· , Dmu) = 0,
Xα≤m
where Dku(x) = {Dαu(x) :
(Ω)
(resp. W m,p(Ω)), the corresponding boundary value problem will be the Dirichlet (resp. Neumann)
problem (cf. [60, pages 6-7]). Moreover, under Hypothesis f2, if dim Ω = 2 and f ∈ C k,α for
some α ∈ (0, 1) and an integer k ≥ 3, it was proved in [60, Chapter 7, Th.4.4] that every critical
point u of F on W m,2
(Ω) sits in C k+m−1,α(Ω); in fact u is also analytic in Ω provided that f is
α = k}, k = 1,··· , m. For example, when V is W m,p
(1.9)
0
0
analytic in its arguments.
0
As stated on the pages 118-119 of [60] (see [58] for detailed arguments), under Hypothe-
sis fp the functional F in (1.8) is of class C 1; and the (derivative) mapping F ′ : W m,p
(Ω) →
[W m,p
(Ω)]∗ is Fr´echet differentiable if p > 2, but only Gateaux-differentiable if p = 2. A critical
point u of F is said to be nondegenerate if the derivative of F ′ at it, F ′′(u) : W m,p
(Ω) →
L (W m,p
(Ω)]∗) is injective. In case p = 2, if F has only nondegenerate critical
points, Skrypnik [58, Chapter 5] obtained Morse inequalities provided that F(u) → +∞ as
kukm,2 → ∞. On the other hand he also obtained
(Ω), [W m,p
0
0
0
0
Skrypnik Theorem ([60, Chap.5, Sec. 5.1, Theorem 1]).
R1 × Rn) has uniformly bounded mixed partial derivatives
∂2f (x, u, ξ)
∂2f (x, u, ξ)
fij =
,
fi0 =
∂ξi∂ξj
,
∂ξi∂u
If p = 2, m = 1 and f ∈ C 2(Ω ×
f00 =
∂2f (x, u, ξ)
∂u2
,
(and therefore f satisfies Hypothesis f2), then the functional F on W 1,2
0
derivative at zero if and only if
(Ω) has Fr´echet second
f (x, 0, ξ) =
nXi,j=1
aij(x)ξiξj +
nXi=1
bi(x)ξi + c(x).
So, generally speaking, under Hypothesis f2 the known Morse theory method cannot be effec-
tively used to study critical points of F on W m,2
(Ω) without nondegenerate conditions. A similar
0
question also appears in some optimal control problems [63].
The key of this paper is to prove a new splitting theorem (Theorem 2.2) for a class of non-
C 2 functionals on a Hilbert space under the following Hypothesis 1.1 (following the notion and
terminology in [40] without special statements). Even if for the Lagrangian systems studied in
[39], we can largely simplify the arguments therein with this new theorem. However, the theories
in [39, 40] may, sometime, provide more elaborate results as done in [44, 46, 47, 48].
Hypothesis 1.1. Let H be a Hilbert space with inner product (·,·)H and the induced norm k · k,
and let X be a dense linear subspace in H. Let V be an open neighborhood of the origin θ ∈ H,
***
7
and let L ∈ C 1(V, R) satisfy DL(θ) = 0. Assume that the gradient ∇L has a Gateaux derivative
B(u) ∈ Ls(H) at every point u ∈ V ∩ X, and that the map B : V ∩ X → Ls(H) has a
decomposition B = P + Q, where for each x ∈ V ∩ X, P (x) : H → H is a bounded linear
positive definitive operator and Q(x) : H → H is a compact linear operator with the following
properties:
(D1) All eigenfunctions of the operator B(θ) that correspond to non-positive eigenvalues belong
to X;
(D2) For any sequence {xk}k≥1 ⊂ V ∩ X with kxkk → 0 it holds that kP (xk)u − P (θ)uk → 0
for any u ∈ H;
(D3) The map Q : V ∩ X → L (H) is continuous at θ with respect to the topology induced from
H on V ∩ X;
(D4) For any sequence {xn}n≥1 ⊂ V ∩ X with kxnk → 0 (as n → ∞), there exist constants
C0 > 0 and n0 ∈ N such that
(P (xn)u, u)H ≥ C0kuk2 ∀u ∈ H, ∀n ≥ n0.
(Note: In Lemma 2.9 we shall prove that the condition (D4) is equivalent to (D4*) in [40].
Lemma 2.10 shows that this property with X = H is hereditary on closed subspaces).
Actually, we prove a more general parameterized splitting theorem, Theorem 2.18. Using it
we complete generalizations of many bifurcation theorems for potential operators in Section 3.
A weaker Marino-Prodi perturbation type result is also presented in Section 2.7. These constitute
abstract theories in Part I of this paper. Part II deals with quasi-linear elliptic systems of higher
order. In Section 4, we study fundamental analytic properties of the functional F under Hypoth-
esis Fp,N . In particular, Hypothesis F2,N assures that F satisfies Hypothesis 1.1 on any closed
subspace of W m,2(Ω, RN ) for a bounded Sobolev domain Ω ⊂ Rn. Because of these, the Morse
theory methods can be used to study the quasi-linear elliptic boundary value problem (1.4) un-
der Hypothesis F2,N as done for the semi-linear elliptic one in [12, 52]. In other sections we are
only satisfied to present some direct applications of results in Part I, for example, giving Morse
inequalities in Section 6 and some bifurcation results for quasi-linear elliptic systems in Section 7.
Further applications will be given in latter papers.
8
Part I
Abstract theories
2 Local Morse theory for a class of non-C 2 functionals
2.1 Statements of main results
Our local Morse theory mainly consist of a new splitting theorem and a Marino-Prodi perturbation
type result for a class of non-C 2 functionals.
We always assume that Hypothesis 1.1 holds without special statements. Then it implies that
∇L is of class (S)+ near θ as proved in [40, p.2966-2967]. In particular, L satisfies the (PS)
condition near θ.
For the bounded linear self-adjoint operator B(θ) on the Hilbert space H, let H = H + ⊕
H 0 ⊕ H − be the orthogonal decomposition according to the positive definite, null and negative
definite spaces of it. Denote by P ∗ the orthogonal projections onto H ∗, ∗ = +, 0,−. By [40,
Proposition B.2] the above fundamental assumptions implies that there exists a constant C0 > 0
such that each λ ∈ (−∞, C0) is either not in the spectrum σ(B(θ)) or is an isolated point of
σ(B(θ)) which is also an eigenvalue of finite multiplicity. It follows that both H 0 and H − are
finitely dimensional, and that there exists a small a0 > 0 such that [−2a0, 2a0] ∩ σ(B(θ)) at most
contains a point 0, and hence
(B(θ)u, u)H ≤ −2a0kuk2 ∀u ∈ H −. (cid:27)
(B(θ)u, u)H ≥ 2a0kuk2 ∀u ∈ H +,
(2.1)
Note that (D1) implies H − ⊕ H 0 ⊂ X. ν := dim H 0 and µ := dim H − are called the Morse
index and nullity of the critical point θ. In particular, if ν = 0 the critical point θ is said to be
nondegenerate. Without special statements, all nondegenerate critical points in this paper are in
the sense of this definition. Moreover, such a critical point must be isolated by (2.4) or (2.5)
Our first result is the following Morse-Palias Lemma. Comparing with that of [40, Remark 2.2(i)],
the smoothness of L is strengthened, but other conditions are suitably weakened.
Theorem 2.1. Under Hypothesis 1.1, if ν = 0, i.e., θ is nondegenerate, there exist a small ǫ > 0,
an open neighborhood W of θ in H and an origin-preserving homeomorphism, φ : BH +(θ, ǫ) +
BH −(θ, ǫ) → W , such that
L ◦ φ(u+ + u−) = ku+k2 − ku−k2,
∀(u+, u−) ∈ BH +(θ, ǫ) × BH −(θ, ǫ).
Moreover, if H is a closed subspace containing H −, and H + is the orthogonal complement of H −
in H, i.e., H + = H ∩ H +, then φ restricts to a homeomorphism φ : (B H +(θ, ǫ) + BH −(θ, ǫ)) →
W := W ∩ H, and L◦ φ(u+ + u−) = ku+k2−ku−k2 for all (u+, u−) ∈ B H +(θ, ǫ)× BH −(θ, ǫ).
Under the assumptions of this theorem, if X = H we can prove that ∇L is locally invertible
at θ in Theorem 2.15. Theorem 2.1 is also key for us to prove the following degenerate case.
9
Theorem 2.2 (Splitting Theorem). Let Hypothesis 1.1 hold with X = H. Suppose ν 6= 0. Then
there exist small positive numbers ǫ, r, s, a unique continuous map ϕ : BH 0(θ, ǫ) → H + ⊕ H −
satisfying ϕ(θ) = θ and
(I − P 0)∇L(z + ϕ(z)) = 0 ∀z ∈ BH 0(θ, ǫ),
(2.2)
an open neighborhood W of θ in H and an origin-preserving homeomorphism
Φ : BH 0(θ, ǫ) × (BH +(θ, r) + BH −(θ, s)) → W
of form Φ(z, u+ + u−) = z + ϕ(z) + φz(u+ + u−) with φz(u+ + u−) ∈ H + ⊕ H − such that
L ◦ Φ(z, u+ + u−) = ku+k2 − ku−k2 + L(z + ϕ(z))
for all (z, u+ + u−) ∈ BH 0(θ, ǫ) × (BH +(θ, r) + BH −(θ, s)). Moreover, ϕ is of class C 1−0, and
the homeomorphism Φ has also properties:
(a) For each z ∈ BH 0(θ, ǫ), Φ(z, θ) = z + ϕ(z), φz(u+ + u−) ∈ H − if and only if u+ = θ;
(b) The functional BH 0(θ, ǫ) ∋ z 7→ L◦(z) := L(z + ϕ(z)) is of class C 1 and
DL◦(z)v = DL(z + ϕ(z))v,
∀v ∈ H 0.
If L is of class C 2−0, so is L◦.
Since the map ϕ satisfying (2.2) is unique, as [39, 40] it is possible to prove in some cases that
ϕ and L◦ are of class C 1 and C 2, respectively.
Theorem 2.2 is a direct consequence of Theorems 2.14,2.18 and Proposition 2.17.
Under the assumptions of Theorem 2.2, we cannot assure that θ is an isolated critical point.
But, if x ∈ H is a critical point of L and very close to θ, it follows from (2.2) and (2.10) -- (2.11)
that x ∈ H 0 and satisfies ϕ(x) = θ.
Theorems 2.1,2.2 cannot be derived from those of [24]. In fact, according to the conditions (c)
and (d) in [24, Theorem 1.3] the functional L in Theorem 2.1 should satisfy:
(c') ∃ η > 0, δ > 0 such that (B(u)(u + z) − B(θ)(u + z), h) < ηku + zk · khk for all
u ∈ BH (θ, δ), z ∈ H 0 and h ∈ H \ {θ};
2 )(cid:1)> 0
(d') ∃ δ > 0 such that(cid:0)∇L(z + u+
The former implies kB(u)(u + z)− B(θ)(u + z)k ≤ ηku + zk for all u ∈ BH(θ, δ), z ∈ H 0;
2 ) ∈ BH +(θ, δ) × BH −(θ, δ) with u+
1 ) − ∇L(z + u+
2 ) + (u−
1 − u−
2 + u−
2 .
2 + u−
2 ), (u+
1 6= u+
1 − u+
1 + u−
for all (u+
1 , u−
1 ), (u+
2 , u−
1 + u−
and specially
kB(u)u − B(θ)u ≤ ηkuk ∀u ∈ BH(θ, δ),
kB(z)z − B(θ)zk ≤ ηkzk ∀z ∈ BH(θ, δ) ∩ H 0.
10
The latter implies that for some t = t(z, u+
1 , u−
1 , u+
2 , u−
2 ) ∈ (0, 1),
(cid:0)B(z + u+
2 + u−
2 + tu+ + tu−)(u+ + u−), u+ + u−(cid:1) > 0
1 − u−
2 .
with u+ = u+
1 − u+
2 and u− = u−
From these it is not hard to see that under our assumptions the conditions (c') and (d') cannot
be satisfied in general.
Let Hq(A, B; K) denote the qth relative singular homology group of a pair (A, B) of topolog-
ical spaces with coefficients in Abel group K. For each q ∈ N ∪ {0} the qth critical group (with
coefficients in K) of L at a point θ is defined by
Cq(L, θ; K) = Hq(Lc ∩ U,Lc ∩ U \ {θ}; K),
where c = L(θ), Lc = {L ≤ c} and U is a neighborhood of θ in H.
Under the assumptions of Theorem 2.1 we have Cq(L, θ; K) = δqµK as usual. For the de-
generate case, though our L◦ is only of class C 1, the proofs in [50, Theorem 8.4] and [14, Theo-
rem 5.1.17] (or [13, Theorem I.5.4]) may be slightly modify to get the following shifting theorem,
a special case of Theorem 2.19.
Theorem 2.3 (Shifting Theorem). Under the assumptions of Theorem 2.2, if θ is an isolated
critical point of L, for any Abel group K it holds that
Cq(L, θ; K) ∼= Cq−µ(L◦, θ; K) ∀q = 0, 1,··· ,
As done for C 2 functionals in [13, 14, 50, 52] some critical point theorems can be derived from
Theorem 2.3. For example, Cq(L, θ; K) is equal to δqµK (resp. δq(µ+ν)K) if θ is a local minimizer
(resp. maximizer) of L◦, and Cq(L, θ; K) = 0 for q ≤ µ and q ≥ µ + ν if θ is neither a local min-
imizer nor local maximizer of L◦. Similarly, the corresponding generalizations of Theorems 2.1,
2.1', 2.2, 2.3 and Corollary 1.3 in [13, Chapter II] can be obtained with Theorems 2.1, 2.2 and
their equivariant versions in Section 2.6. In particular, as a generalization of [13, Theorem II.1.6]
(or [14, Theorem 5.1.20]) we have
Theorem 2.4. Let Hypothesis 1.1 hold with X = H, and let θ be an isolated critical point of
mountain pass type, i.e., C1(L, θ; K) 6= 0. Suppose that ν > 0 and µ = 0 imply ν = 1. Then
Cq(L, θ; K) = δq1K.
When ν > 0 and µ = 1, C0(L◦, θ; K) 6= 0 by Theorem 2.3. We can change L◦ outside a
very small neighborhood θ ∈ BH 0(θ, ǫ) to get a C 1 functional on H 0 which is coercive (and so
satisfies the (PS)-condition). Then it follows from C0(L◦, θ; K) 6= 0 and [52, Proposition 6.95]
that θ is a local minimizer of L◦.
As a generalization of Corollary 3.1 in [13, page 102] we have also: Under the assumptions of
Theorem 2.4, if the smallest eigenvalue λ1 of B(θ) = d2L(θ) is simple whenever λ1 = 0, then
λ1 ≤ 0, and index(∇L, θ) = −1.
11
Theorem 5.1 and Corollary 5.1 in [13, page 121] are also true if "f ∈ C 2(M, R)" and "Fred-
holm operators d2f (xi)" are replaced by "f ∈ C 1(M, R) and ∇f is Gateaux differentiable" and
"under some chart around pi the functional f has a representation that satisfies Hypothesis 1.1",
respectively.
Marino and Prodi [49] studied local Morse function approximations for C 2 functionals on
Hilbert spaces. We shall generalize their result to a class of functionals satisfying the following
stronger assumption than Hypothesis 1.1.
Hypothesis 2.5. Let V be an open set of a Hilbert space H with inner product (·,·)H , and L ∈
C 1(V, R). Assume that the gradient ∇L has a Gateaux derivative B(u) ∈ Ls(H) at every point
u ∈ V , and that the map B : V → Ls(H) has a decomposition B = P + Q, where for each
u ∈ V , P (u) : H → H is a bounded linear positive definitive operator and Q(u) : H → H is a
compact linear operator with the following properties:
(i) For any u ∈ H, the map V ∋ x 7→ P (x)u ∈ H is continuous;
(ii) The map Q : V → L (H) is continuous;
(iii) P is local positive definite uniformly, i.e., each x0 ∈ V has a neighborhood U (x0) such that
for some constants C0 > 0,
(P (x)u, u)H ≥ C0kuk2,
∀u ∈ H, ∀x ∈ U (x0).
As in the proof of Theorem 4.1, under Hypothesis F2,N , we can check that the functional F in
(1.3) satisfies this hypothesis. By improving methods in [49, 13, 20] we may prove
Theorem 2.6. Under Hypothesis 2.5, suppose: (a) u0 ∈ V is a unique critical point of L, (b) the
corresponding maps ϕ and L◦ as in Theorem 2.2 near u0 are of classes C 1 and C 2, respectively,
(c) L satisfies the (PS) condition. Then for any ǫ > 0 and r > 0 such that ¯BH(u0, r) ⊂ V and
there exists a functional L ∈ C 1(V, R) with the following properties:
(i) L satisfies Hypothesis 2.5 and the (PS) condition;
(ii) supu∈V kL(i)(u) − L(i)(u)k < ǫ, i = 0, 1, 2;
(iii) L(x) = L(x) if x ∈ V and ku − u0k ≥ r;
(iv) the critical points of g, if any, are in BH(u0, r) and nondegenerate (so finitely many by the
arguments below 2.1); moreover the Morse indexes of these critical points sit in [m−, m− + n0],
where m− and n0 are the Morse index and nullity of u0, respectively.
As showed, the functionals in [39, 48] satisfy the conditions of this theorem. Marino -- Prodi's
result has many important applications in the critical point theory, see [13, 20, 30, 38] and literature
therein. With Theorem 2.6 they may be given in our framework.
Marino -- Prodi's perturbation theorem in [49] was also generalized to the equivariant case un-
der the finite (resp. compact Lie) group action by Wasserman [69] (resp. Viterbo [66]), see the
proof of Theorem 7.8 in [13, Chapter I] for full details. Similarly, we can present an equivariant
version of Theorem 2.6 for compact Lie group action, but it is omitted here.
12
Strategies of proofs for results in this section and arrangements. Under the assumptions of
Theorem 2.1, no known implicit function theorems or contraction mapping principles can be used
to get ϕ in (2.2), which is different from the case in [39, 40]. The methods in [24] provide a possible
way to construct such a ϕ. However, as mentioned above our assumptions cannot guarantee the
above conditions (c') and (d'). Fortunately, it is with Lemma 2.12 and Theorem 2.1 that we can
complete this construction.
In Section 2.2 we list some lemmas. Theorem 2.1 will be proved in Section 2.3. It is necessary
for a key implicit function theorem for a family of potential operators, Theorem 2.14, which is
proved in Section 2.4; we also give an inverse function theorem, Theorem 2.15, there. In Sec-
tion 2.5 we shall prove a parameterized splitting theorem, Theorem 2.18, and a parameterized
shifting theorem, Theorem 2.19; Theorems 2.2, 2.3 are special cases of them, respectively. The
equivariant case is considered in Section 2.6. Theorem 2.6 will be proved in Section 2.7.
2.2 Lemmas
Under Hypothesis 1.1 we have the following two lemmas as proved in [39, 40].
Lemma 2.7. There exists a function ω : V ∩ X → [0,∞) such that ω(x) → 0 as x ∈ V ∩ X and
kxk → 0, and that
(B(x)u, v)H − (B(θ)u, v)H ≤ ω(x)kuk · kvk
for any x ∈ V ∩ X, u ∈ H 0 ⊕ H − and v ∈ H.
Lemma 2.8. There exists a small neighborhood U ⊂ V of θ in H and a number a1 ∈ (0, 2a0]
such that for any x ∈ U ∩ X,
(i) (B(x)u, u)H ≥ a1kuk2 ∀u ∈ H +;
(ii) (B(x)u, v)H ≤ ω(x)kuk · kvk ∀u ∈ H +,∀v ∈ H − ⊕ H 0;
(iii) (B(x)u, u)H ≤ −a0kuk2 ∀u ∈ H −.
Lemma 2.9. Actually, (D4) is equivalent to the condition (D*) in [40], i.e.,
(D4*) There exist positive constants η0 > 0 and C ′
0 > 0 such that
(P (x)u, u) ≥ C ′
0kuk2 ∀u ∈ H, ∀x ∈ BH(θ, η0) ∩ X.
Proof. Indeed, since each P (x) is a positive definite linear operator, its spectral set is a bounded
equivalent to the statement: For any sequence {xn} ⊂ V ∩ X with kxnk → 0 (as n → ∞), there
closed subset in (0,∞). Moreover, we have σ(pP (x)) = {√λ λ ∈ σ(P (x))}. So (D4) is
holds: infn min σ(pP (xn)) > 0. Similarly, (D4*) can be equivalently expressed as: There exist
positive constants η0 > 0 such that
inf{min σ(pP (x)) x ∈ BH(θ, η0) ∩ X} > 0.
13
(D4) does not hold means that there exists a sequence {xn} ⊂ V ∩ X with kxnk → 0 (as n → ∞)
Suppose (D4) holds. Since η 7→ inf{min σ(pP (x)) x ∈ BH(θ, η) ∩ X} is non-increasing, that
such that inf n min σ(pP (xn)) → 0, which contradicts (D4).
Lemma 2.10. Suppose that Hypothesis 1.1 with X = H is satisfied. Then for any closed subspace
H ⊂ H, ( H, V , L) satisfies Hypothesis 1.1 with X = H, where V := V ∩ H and L := L V .
Proof. Clearly, L ∈ C 1( V , R) and D L(θ) = 0. Denote by Π : H → H the orthogonal projection.
Then the gradient of L at u ∈ V , ∇ L(u), is equal to Π∇L(u). It follows that ∇ L at any u ∈ V
has a Gateaux derivative B(u) = Π ◦ B(u) H ∈ Ls( H). For any u ∈ V , put P (u) = Π ◦ P (u) H
and Q(u) = Π ◦ Q(u) H . Then B = P + Q : V → Ls( H), P (u) is positive definite, and Q(u)
is a compact linear operator. It is easily checked that other conditions are satisfied.
2.3 Proof of Theorem 2.1
Take a small ǫ > 0 so that ¯BH +(θ, ǫ) ⊕ ¯BH −(θ, ǫ) is contained in the open neighborhood U in
Lemma 2.8. Let us prove the C 1 functional
¯BH +(θ, ǫ) ⊕ ¯BH −(θ, ǫ) → R, u+ + u− 7→ L(u+ + u−)
satisfies the conditions in [23, Theorem 1.1].
1 , u−
Step 1. Fix u+ ∈ ¯BH +(θ, ǫ) ∩ X and u−
Since ∇L have a Gateaux derivative B(u) ∈ Ls(H) at every point u ∈ V ∩ X, the function
2 ∈ ¯BH −(θ, ǫ) (which are contained in X by (D1)).
V → R, u 7→ (∇L(u+ + u), u−
2 − u−
1 )H
is Gateaux differentiable at every u ∈ V ∩ X. Using the mean value theorem we have t ∈ (0, 1)
such that
2 ), u−
(∇L(u+ + u−
= (cid:0)B(u+ + u−
1 + t(u−
2 − u−
≤ −a0ku−
1 k2
2 − u−
2 − u−
1 )H − (∇L(u+ + u−
1 ))(u−
1 ), u−
2 − u−
1 )H
1 ), u−
2 − u−
1(cid:1)H
2 − u−
by Lemma 2.8(iii). Note that ¯BH +(θ, ǫ) ∩ X is dense in ¯BH +(θ, ǫ) and ∇L is continuous. For all
u+ ∈ ¯BH +(θ, ǫ) and u−
2 ), u−
i ∈ ¯BH −(θ, ǫ), i = 1, 2, we deduce
2 − u−
1 ), u−
1 )H − (∇L(u+ + u−
1 )H ≤ −a0ku−
(∇L(u+ + u−
2 − u−
(2.3)
2 − u−
1 k2.
This implies the condition (ii) of [23, Theorem 1.1].
Step 2. Let u+ ∈ ¯BH +(θ, ǫ) ∩ X and u− ∈ ¯BH −(θ, ǫ) (which is contained in X by (D1)). Then
14
since DL(θ) = 0, by the mean value theorem, for some t ∈ (0, 1) we have
DL(u+ + u−)(u+ − u−)
= (∇L(u+ + u−), u+ − u−)H − (∇L(θ), u+ − u−)H
= (cid:0)B(t(u+ + u−))(u+ + u−), u+ − u−(cid:1)H
= (cid:0)B(t(u+ + u−))u+, u+(cid:1)H −(cid:0)B(t(u+ + u−))u−, u−(cid:1)H
≥ a1ku+k2 + a0ku−k2
(2.4)
by Lemma 2.8(i) and (iii). As above (2.4) also holds for all u+ ∈ ¯BH +(θ, ǫ) because ¯BH(θ, ǫ) ∩
X + is dense in ¯BH(θ, ǫ)∩ H +. Hence DL(u+ + u−)(u+ − u−) > 0 for (u+, u−) 6= (θ, θ). (This
implies θ to be an isolated critical point of L). The condition (iii) of [23, Theorem 1.1] is satisfied.
Step 3. For u+ ∈ ¯BH +(θ, ε) ∩ X, as above we have t ∈ (0, 1) such that
DL(u+)u+ = DL(u+)u+ − DL(θ)u+
= (∇L(u+), u+)H − (∇L(θ), u+)H
= (cid:0)B(tu+)u+, u+(cid:1)H
≥ a1ku+k2
(2.5)
because of Lemma 2.8(i). It follows that
DL(u+)u+ ≥ a1ku+k2 > p(ku+k) ∀u+ ∈ ¯BH +(θ, ǫ) \ {θ},
where p : (0, ε] → (0,∞) is a non-decreasing function given by p(t) = a1
2 t2. Hence the condition
(iv) of [23, Theorem 1.1] is satisfied.
For the second claim, note that (2.3) -- (2.5) hold for all u+ ∈ ¯BH +(θ, ǫ) and u−, u−
i ∈
¯BH +(θ, ǫ), i = 1, 2. Of course, they are still true for all u+ ∈ ¯B H +(θ, ǫ). Carefully checking
the proof of [23, Theorem 1.1] the conclusion is easily obtained. (Note that this claim seems un-
able to be directly derived from Lemma 2.10.)
✷
Actually, from the proof of Theorem 2.1 we may get the more general claim, which is needed
for later applications.
Theorem 2.11. Under Hypothesis 1.1, let G ∈ C 1(V, R) satisfy: i) G′(θ) = θ, ii) the gradient ∇G
has Gateaux derivative G′′(u) ∈ Ls(H) at any u ∈ V , and G′′ : V → Ls(H) are continuous at
θ. Suppose Ker(B(θ)) = {θ}, i.e., θ is a nondegenerate critical point of L. Then there exist ρ > 0,
ǫ > 0, a family of open neighborhoods of θ in H, {Wλ λ ≤ ρ} and a family of origin-preserving
homeomorphisms, φλ : BH +(θ, ǫ) + BH −(θ, ǫ) → Wλ, λ ≤ ρ, such that
(L + λG) ◦ φλ(u+ + u−) = ku+k2 − ku−k2
for all (u+, u−) ∈ BH +(θ, ǫ) × BH −(θ, ǫ). Moreover, [−ρ, ρ] × (BH +(θ, ǫ) + BH −(θ, ǫ)) ∋
(λ, u) 7→ φλ(u) ∈ H is continuous.
15
Proof. Since G′′ : V → Ls(H) are continuous at θ, as in the proof of (2.3) we may shrink ǫ > 0
and find ρ > 0 such that for all λ ∈ [−ρ, ρ], u+ ∈ ¯BH +(θ, ǫ) and u−
i ∈ ¯BH −(θ, ǫ), i = 1, 2,
λ(∇G(u+ + u−
2 ), u−
2 − u−
1 )H − λ(∇G(u+ + u−
1 ), u−
2 − u−
1 )H ≤
This and (2.3) lead to
a0
2 ku−
2 − u−
1 k2.
(∇(L + λG)(u+ + u−
1 k2.
≤ −
a0
2 ku−
2 − u−
2 ), u−
2 − u−
1 )H − (∇(L + λG)(u+ + u−
1 ), u−
2 − u−
1 )H
(2.6)
Similarly, as in the proof of (2.4) we may shrink the above ρ > 0 and ǫ > 0 so that
λDG(u+ + u−)(u+ − u−) ≤
a1
2 ku+k2 +
a0
2 ku−k2
for all λ ∈ [−ρ, ρ], u+ ∈ ¯BH +(θ, ǫ) and u− ∈ ¯BH −(θ, ǫ). This and (2.4) yield
a0
2 ku−k2
D(L + λG)(u+ + u−)(u+ − u−) ≥
a1
2 ku+k2 +
and specially D(L + λG)(u+)(u+) ≥ a1
2 ku+k2. These and (2.6) show that the conditions of [40,
Theorem A.1] are satisfied. The desired conclusions follow immediately.
2.4 An implicit function theorem for a family of potential operators
In this subsection we shall prove an implicit function theorem, Theorem 2.14, which implies the
first claim in Theorem 2.2. We also give an inverse function theorem, Theorem 2.15, though it is
not used in this paper.
Without special statements, we always assume that Hypothesis 1.1 holds in this subsection.
Take ǫ > 0, r > 0 and s > 0 so small that the closures of both
Qr,s := BH +(θ, r) ⊕ BH −(θ, s)
and BH 0(θ, ǫ) ⊕ Qr,s
are contained in the neighborhood U in Lemma 2.8. Since H 0 ⊂ X, X ∩ Qr,s is also dense in
Qr,s. Let P ⊥ = I − P 0 = P + + P −. By Lemma 2.8 we obtain a′
1ku+k2 − a′
1ku−k2 + a′
(P ⊥∇L(z + u), u+)H = (∇L(u), u+)H ≥ a′
(P ⊥∇L(z + u), u−)H = (∇L(u), u−)H ≤ −a′
0 > 0, a′
0[ω(z + u)]2ku−k2,
0[ω(z + u)]2ku+k2
1 > 0 such that
(2.7)
(2.8)
for all u ∈ Qr,s and z ∈ ¯BH 0(θ, ǫ). Since ω(z+u) → 0 as kz+uk → 0, by shrinking r > 0, s > 0
and ǫ > 0 we can require
[ω(z + u)]2 <
a′
1
2a′
0
,
∀(z, u) ∈ ¯BH 0(θ, ǫ) × Qr,s.
Then this and (2.7) -- (2.8) lead to
(P ⊥∇L(z + u), u+)H ≥ a′
(P ⊥∇L(z + u), u−)H ≤ −a′
1ku+k2 −
1ku−k2 +
a′
2 ku−k2,
1
a′
1
2 ku+k2
(2.9)
(2.10)
(2.11)
16
for all u ∈ Qr,s and z ∈ ¯BH 0(θ, ǫ), and hence
(cid:0)tP ⊥∇L(z1 + u) + (1 − t)P ⊥∇L(z2 + u), u+(cid:1)H ≥ a′
(cid:0)tP ⊥∇L(z1 + u) + (1 − t)P ⊥∇L(z2 + u), u−(cid:1)H ≤ −a′
1ku+k2 −
1ku−k2 +
(2.12)
a′
2 ku−k2,
1
a′
1
2 ku+k2 (2.13)
for all u ∈ Qr,s and zj ∈ ¯BH 0(θ, ǫ), j = 1, 2, and t ∈ [0, 1].
Lemma 2.12. If r > 0, s > 0 and ǫ > 0 are so small that (2.9) is satisfied, then
inf{ktP ⊥∇L(z1 + u) + (1 − t)P ⊥∇L(z2 + u)k (t, z1, z2, u) ∈ Ω} > 0,
where Ω = [0, 1] × ¯BH 0(θ, ǫ) × ¯BH 0(θ, ǫ) × ∂Qr,s.
Proof. Note that ∂Qr,s is union of two closed subsets of it, i.e.
∂Qr,s = [(∂BH + (θ, r)) ⊕ ¯BH −(θ, s)] ∪ [ ¯BH +(θ, r) ⊕ (∂BH −(θ, s))].
Then Ω = Λ1 ∪ Λ2, where Λ1 = [0, 1] × ¯BH 0(θ, ǫ) × ¯BH 0(θ, ǫ) × (∂BH +(θ, r)) ⊕ ¯BH −(θ, s)
and Λ2 = [0, 1] × ¯BH 0(θ, ǫ) × ¯BH 0(θ, ǫ) × BH +(θ, r) ⊕ (∂ ¯BH −(θ, s)). We firstly prove
inf{ktP ⊥∇L(z1 + u) + (1 − t)P ⊥∇L(z2 + u)k (t, z1, z2, u) ∈ Λ1} > 0.
(2.14)
By a contradiction we assume that there exist sequences {tn}n≥1 ⊂ [0, 1] and
{zn}n≥1, {z′
n}n≥1 ⊂ ¯BH 0(θ, ǫ),
{un}n≥1 ⊂ (∂BH + (θ, r)) ⊕ ¯BH −(θ, s)
such that ktnP ⊥∇L(zn + un) + (1 − tn)P ⊥∇L(z′
n + un)k → 0. Hence after removing finite
many terms we can assume
(tnP ⊥∇L(zn + un) + (1 − tn)P ⊥∇L(z′
(tnP ⊥∇L(zn + un) + (1 − tn)P ⊥∇L(z′
n ∈ ∂BH +(θ, r)) and u−
n + un), u+
n + un), u−
n )H ≤
n )H ≥ −
,
1r2
a′
4
1r2
a′
4
(2.15)
∀n ∈ N,
∀n ∈ N. (2.16)
,
a′
1
4
r2 ≥ (tnP ⊥∇L(zn + un) + (1 − tn)P ⊥∇L(z′
n ∈ ¯BH −(θ, s). So (2.15) and (2.12) lead to
1r2 −
n )H ≥ a′
n + un), u+
a′
2 ku−
n k2
1
Note that u+
and therefore
−
and hence
Moreover, from (2.13) and (2.16) we derive
a′
1r2
4 ≤ (tnP ⊥∇L(zn + un) + (1 − tn)P ⊥∇L(z′
n + un), u−
n )H ≤ −a′
1ku−
n k2 +
(2.17)
a′
1r2
2
r2
n k2 ≤
ku−
2
3
,
∀n ∈ N.
r2
n k2 ≥
ku−
4
3
,
∀n ∈ N,
which contradicts (2.17). (2.14) is proved.
Similarly, suppose that there exist sequences {tn}n≥1 ⊂ [0, 1] and
{zn}n, {z′
n}n ⊂ ¯BH 0(θ, ǫ),
{vn}n ⊂ BH +(θ, r) ⊕ (∂BH −(θ, s))
17
such that ktnP ⊥∇L(zn + vn) + (1 − tn)P ⊥∇L(z′
(tnP ⊥∇L(zn + vn) + (1 − tn)P ⊥∇L(z′
(tnP ⊥∇L(zn + vn) + (1 − tn)P ⊥∇L(z′
n ∈ BH +(θ, r) and v−
≤ (tnP ⊥∇L(zn + vn) + (1 − tn)P ⊥∇L(z′
Note that v+
1s2
a′
4
n + vn), v−
−
and so
n ∈ ∂BH −(θ, s) for all n ∈ N. Then (2.13) and (2.19) imply
a′
2 kv+
n k2
1
n )H ≤ −a′
n + vn), v−
1s2 +
n + vn)k → 0. As above we can assume
n + vn), v+
,
n )H ≤
n )H ≥ −
a′
1s2
4
1s2
a′
4
,
(2.18)
∀n ∈ N,
∀n ∈ N. (2.19)
With the same methods, (2.12) and (2.18) lead to
s2
n k2 ≤
kv+
2
3
,
∀n ∈ N.
1s2
a′
4 ≥ (tnP ⊥∇L(zn + vn) + (1 − tn)P ⊥∇L(z′
n + vn), v+
n )H ≥ a′
1kv+
n k2 −
(2.20)
a′
1
2
s2
and so
This contradicts (2.20). Hence
s2
n k2 ≥
kv+
4
3
,
∀n ∈ N.
inf{ktP ⊥∇L(z1 + u) + (1 − t)P ⊥∇L(z2 + u)k (t, z1, z2, u) ∈ Λ2} > 0.
This and (2.14) yield the desired conclusions.
Since (D4) is equivalent to (D4*), it was proved in [40, p. 2966 -- 2967] that ∇L is of class (S)+
under the conditions (S), (F), (C) and (D) in [40]. In particular, this is also true under the assump-
tions of Theorem 2.1 (without requirement H 0 = {θ}), of course the conditions of Theorem 2.1
guarantee the same claim.
In the following we always assume that r > 0, s > 0 and ǫ > 0 are as in Lemma 2.12.
Lemma 2.13. For each z ∈ BH 0(θ, ǫ), the map
fz : Qr,s ∋ u 7→ P ⊥∇L(z + u) ∈ H + ⊕ H −
is of class (S)+. Moreover, for any two points z0, z1 ∈ BH 0(θ, ǫ) the map H : [0, 1] × Qr,s →
H + ⊕ H − given by
H (t, u) = (1 − t)P ⊥∇L(z0 + u) + tP ⊥∇L(z1 + u)
is a homotopy of class (S)+ (cf. [52, Definition 4.40]).
18
Proof. By [52, Proposition 4.41] we only need to prove the first claim. Let {uj} ⊂ Qr,s weakly
converge to u ∈ H + ⊕ H −. Assume that they satisfy
lim(P ⊥∇L(z + uj), uj − u)H ≤ 0.
It suffices to prove uj → u in H + ⊕ H −. Note that uj ⇀ u in H because Qr,s ⊂ H + ⊕ H −. So
is z + uj ⇀ z + u in H. Moreover, uj − u ∈ H + ⊕ H − implies
(P ⊥∇L(z + uj), uj − u)H = (∇L(z + uj), uj − u)H
= (∇L(z + uj), (z + uj) − (z + u))H .
It follows that lim(∇L(z + uj), (z + uj) − (z + u))H ≤ 0. But ∇L is of class (S)+ near θ ∈ H,
we have z + uj → z + u and so uj → u.
Let deg denote the Browder-Skrypnik degree for demicontinuous (S)+-maps ([8, 9], [58, 59,
60]), see [52, §4.3] for a nice exposition. By Lemma 2.12 deg(f0,Qr,s, θ) is well-defined and
using the Poincar´e-Hopf theorem (cf. [19, Theorem 1.2]) we have
deg(f0,Qr,s, θ) =
∞Xq=0
(−1)qrankCq(f0, θ; G).
(2.21)
Note that LQr,s satisfies the conditions of Theorem 2.1. It follows that Cq(f0, θ; G) = δµqG,
where µ = dim H −. Hence (2.21) becomes
deg(f0,Qr,s, θ) = (−1)µ.
(2.22)
For each z ∈ BH 0(θ, ǫ), we derive from Lemma 2.12 that
inf{kfz(u)k u ∈ ∂Qr,s} > 0 and
inf{ktfz(u) + (1 − t)f0(u)k t ∈ [0, 1], u ∈ ∂Qr,s} > 0.
The former implies that deg(fz,Qr,s, θ) is well-defined, the latter and Lemma 2.13 lead to
deg(fz,Qr,s, θ) = deg(f0,Qr,s, θ) = (−1)µ
by (2.22). So there exists a point uz ∈ Qr,s such that
P ⊥∇L(z + uz) = fz(uz) = θ.
Now let us give the main result in this subsection.
(2.23)
(2.24)
Theorem 2.14 (Parameterized Implicit Function Theorem). Under the assumptions of Theorem 2.2,
suppose further that G1,··· ,Gn ∈ C 1(V, R) satisfy
(i) G′
j(θ) = θ, j = 1,··· , n;
19
(ii) for each j = 1,··· , n, the gradient ∇Gj has Gateaux derivative G′′
j (u) ∈ Ls(H) at any
u ∈ V , and G′′
j : V → Ls(H) are continuous at θ.
Then by shrinking ǫ > 0 (if necessary) we have δ > 0 and a unique continuous map
ψ : [−δ, δ]n × BH(θ, ǫ) ∩ H 0 → Qr,s ⊂ (H 0)⊥
(2.25)
such that for all (~λ, z) ∈ [−δ, δ]n × BH (θ, ǫ) ∩ H 0 with ~λ = (λ1,··· , λn), ψ(~λ, θ) = θ and
P ⊥∇L(z + ψ(~λ, z)) +
nXj=1
λjP ⊥∇Gj(z + ψ(~λ, z)) = θ,
(2.26)
where P ⊥ is as in (2.24). This ψ also satisfies
kψ(~λ, z1) − ψ(~λ, z2)k ≤ 3kz1 − z2k,
∀(~λ, z) ∈ [−δ, δ]n × BH(θ, ǫ) ∩ H 0.
(2.27)
Moreover, if G is a compact Lie group acting on H orthogonally, V , L and all Gj are G-invariant
(and hence H 0, (H 0)⊥ are G-invariant subspaces, and ∇L, ∇Gj are G-equivariant), then ψ is
equivariant on z, i.e., ψ(~λ, g · z) = g · ψ(~λ, z) for (~λ, z) ∈ [−δ, δ]n × BH(θ, ǫ) ∩ H 0 and g ∈ G.
Proof. Step 1. There exist ρ1, δ ∈ (0, 1) such that BH(θ, 2ρ1) ⊂ V and that if sequences
~λk = (λk,1,··· , λk,n) ∈ [−δ, δ]n converge to ~λ0 = (λ0,1,··· , λ0,n) ∈ [−δ, δ]n, uk ∈ BH(θ, 2ρ1)
weakly converge to u0 ∈ BH(θ, 2ρ1), and they also satisfy
lim(∇L(uk) +
nXj=1
λj∇Gj(uk), uk − u0)H ≤ 0,
(2.28)
then uk → u0. In particular, for each ~λ ∈ [−δ, δ]n, the map
BH(θ, 2ρ1) × [0, 1] → H + ⊕ H −, (t, u) 7→ P ⊥∇L(u) +
nXj=1
tλjP ⊥∇Gj(u)
is a homotopy of class (S)+ (cf. [52, Definition 4.40]).
In fact, by [40, (5.8)] we had found ρ1 > 0 and C ′
0 > 0 such that BH(θ, 2ρ1) ⊂ V and
(∇L(u), u − u′)H ≥
C ′
2 ku − u′k2 + (∇L(u′), u − u′)H
0
+ (Q(θ)(u − u′), u − u′)H
(2.29)
for any u, u′ ∈ BH(θ, 2ρ1). Similarly, for each fixed j ∈ {1,··· , n}, we have τ = τ (u, u′) ∈
(0, 1) such that
(∇Gj(u), u − u′)H = (∇Gj(u) − ∇Gj(u′), u − u′)H + (∇Gj(u′), u − u′)H
= (G′′
= ([G′′
+(G′′
j (τ u + (1 − τ )u′)(u − u′), u − u′)H + (∇Gj(u′), u − u′)H
j (τ u + (1 − τ )u′) − G′′
j (θ)(u − u′), u − u′)H ,
∀u, u′ ∈ BH(θ, 2ρ1).
j (θ)](u − u′), u − u′)H + (∇Gj(u′), u − u′)H
20
Since V ∋ v 7→ G′′
j (v) ∈ Ls(H) is continuous at θ, we may shrink ρ1 > 0 so that
j (v) − G′′
kG′′
∀v ∈ BH(θ, 2ρ1), j = 1,··· , n.
j (θ)k ≤
C ′
0
8n
,
It follows that for all u, u′ ∈ BH(θ, 2ρ1) and j = 1,··· , n,
(∇Gj(u), u − u′)H ≤
C ′
8nku − u′k2 + (∇Gj(u′), u − u′)H
0
+(G′′
j (θ)(u − u′), u − u′)H.
(2.30)
Take δ ∈ (0, 1) so that
δ
kG′′
j (θ)k <
C ′
0
8
.
nXj=1
These and (2.29) imply that for all ~λ = (λ1,··· , λn) ∈ [−δ, δ]n,
nXj=1
λj(∇Gj(u), u − u′)H
(∇L(u), u − u′)H +
C ′
4 ku − u′k2 + (∇L(u′), u − u′)H + (Q(θ)(u − u′), u − u′)H
0
nXj=1
(∇Gj(u′), u − u′)H.
≥
−
Replacing u, u′ and λj by uk, u0 and λk,j in the inequality, we derive from (2.28) that uk → u0
because (D3) implies that (∇L(u0), uk − u0)H → 0, (Q(θ)(uk − u0), uk − u0)H → 0 and
(∇Gj(u0), uk − u0)H → 0.
Note: The above proof shows that the family {L~λ := L +Pn
j=1 λjGj ~λ ∈ [−δ, δ]n} satisfies
the (PS) condition on ¯BH(θ, ε) for any ε < 2ρ1, that is, if sequences ~λk ∈ [−δ, δ]n converge
to ~λ0 ∈ [−δ, δ]n, and uk ∈ ¯BH (θ, ε) satisfies ∇L~λk
(uk) < ∞, then
{uk}k≥1 has a converging subsequence uki → u0 ∈ ¯BH(θ, ε) with ∇L~λ0
Step 2. Let r > 0, s > 0 and ǫ > 0 be as in Lemma 2.12. By shrinking them, we can assume
that ¯BH 0(θ, ǫ) × Qr,s ⊂ BH(θ, 2ρ1) and
(uk) → θ and supk L~λk
(u0) = θ.
sup{k∇L~λ(z, u)k (~λ, z, u) ∈ [−1, 1]n × ¯BH 0(θ, ǫ) ⊕ Qr,s} < ∞
(2.31)
because ∇L and ∇G1,··· ,∇Gn are all locally bounded. Then by Lemma 2.12 we may shrink
δ ∈ (0, 1) so that
inf ktP ⊥(∇L +
nXj=1
λj∇Gj)(z1 + u) + (1 − t)P ⊥(∇L +
nXj=1
λj∇Gj)(z2 + u)k > 0,
where the infimum is taken for all (t, z1, z2, u) ∈ [0, 1] × ¯BH 0(θ, ǫ) × ¯BH 0(θ, ǫ) × ∂Qr,s and
(λ1,··· , λn) ∈ [−δ, δ]n. This implies that for each (~λ, z) ∈ [−δ, δ]n × BH 0(θ, ǫ), the map
f~λ,z : Qr,s ∋ u 7→ P ⊥∇L(z + u) +
nXj=1
λjP ⊥∇Gj(z + u) ∈ H + ⊕ H −
21
has a well-defined Browder-Skrypnik degree deg(f~λ,z,Qr,s, θ) and
deg(f~λ,z,Qr,s, θ) = deg(f~0,0,Qr,s, θ) = deg(f0,Qr,s, θ) = (−1)µ,
(2.32)
where f0 is as in (2.22). Hence for each (~λ, z) ∈ [−δ, δ]n × BH 0(θ, ǫ) there exists a point u~λ,z ∈
Qr,s such that
P ⊥∇L(z + u~λ,z) +
nXj=1
λjP ⊥∇Gj(z + u~λ,z) = f~λ,z(uλ,z) = θ.
(2.33)
By shrinking the above ǫ > 0, r > 0 and s > 0 (if necessary) so that ω and a0, a1 in Lemma 2.8
can satisfy
ω(z + u) < min{a0, a1}/2,
∀(z, u) ∈ ¯BH 0(θ, ǫ) × Qr,s.
(2.34)
Step 3.
If δ > 0 is sufficiently small, then u~λ,z is a unique zero point of f~λ,z in Qr,s.
In fact, suppose that there exists another different u′
~λ,z ∈ Qr,s satisfying (2.33). We decompose
)−, and prove the conclusion in three cases:
~λ,z
~λ,z
~λ,z
~λ,z
= (u~λ,z − u′
u~λ,z − u′
• k(u~λ,z − u′
)+k > k(u~λ,z − u′
• k(u~λ,z − u′
)+k = k(u~λ,z − u′
• k(u~λ,z − u′
)+k < k(u~λ,z − u′
Let us write L~λ = L +Pn
)+ + (u~λ,z − u′
)−k,
)−k,
)−k.
~λ,z
~λ,z
~λ,z
~λ,z
~λ,z
j=1 λjGj for conveniences. Then (2.33) implies
)+)H
0 = (P ⊥∇L~λ(z + u~λ,z) − P ⊥∇L~λ(z + u′
~λ,z
), (u~λ,z − u′
= (P ⊥∇L(z + u~λ,z) − P ⊥∇L(z + u′
~λ,z
), (u~λ,z − u′
~λ,z
~λ,z
)+)H
+
nXj=1
λj(P ⊥∇Gj(z + u~λ,z) − P ⊥∇Gj(z + u′
~λ,z
), (u~λ,z − u′
~λ,z
)+)H .
(2.35)
For simplicity we write u~λ,z = uz and u′
~λ,z
= u′
z.
For the first two cases, we may use the mean value theorem to get τ ∈ (0, 1) such that
z)+)H
z), (uz − u′
z)+)H
z), (uz − u′
z)+, (uz − u′
z)−, (uz − u′
z)+)H
z)+)H
z)+)H
z)k(uz − u′
(P ⊥∇L(z + uz) − P ⊥∇L(z + u′
z), (uz − u′
z)(uz − u′
z)(uz − u′
z)(uz − u′
= (∇L(z + uz) − ∇L(z + u′
= (B(z + τ uz + (1 − τ )u′
= (B(z + τ uz + (1 − τ )u′
+ (B(z + τ uz + (1 − τ )u′
≥ a1k(uz − u′
≥ a1k(uz − u′
≥ a1k(uz − u′
a1
2 k(uz − u′
=
z)+k2 − ω(z + τ uz + (1 − t)u′
z)+k2 −
z)+k2 −
z)+k2,
a1
[k(uz − u′
4
a1
2 k(uz − u′
z)−k2 + k(uz − u′
z)+k2
z)+k2]
z)−k · k(uz − u′
z)+k
(2.36)
22
where the first inequality comes from Lemma 2.8(i)-(ii), the second one is derived from (2.34) and
the inequality 2ab ≤ a2 + b2, and the third one is because k(uz − u′
z)−k ≤ k(uz − u′
z)+k. It
follows from (2.35) -- (2.36) that
0 = (P ⊥∇L~λ(z + u~λ,z) − P ⊥∇L~λ(z + u′
~λ,z
), (u~λ,z − u′
~λ,z
)+)H
~λ,z
)+k2
≥
a1
2 k(u~λ,z − u′
nXj=1
λj(G′′
+
j (z + τ u~λ,z + (1 − τ )u′
~λ,z
)(u~λ,z − u′
~λ,z
), (u~λ,z − u′
~λ,z
)+)H.
(2.37)
By (2.30) we have a constant M > 0 such that
sup{kG′′
j (z + w)k (z, w) ∈ ¯BH 0(θ, ǫ) × Qr,s, j = 1,··· , n} < M.
(2.38)
Hence
~λ,z
λj(G′′
nXj=1
≤ nδMk(u~λ,z − u′
≤ nδM [k(u~λ,z − u′
≤ 2nδMk(u~λ,z − u′
)(u~λ,z − u′
j (z + τ u~λ,z + (1 − τ )u′
)k · k(u~λ,z − u′
)+k
)+k2 + k(u~λ,z − u′
)+k2.
~λ,z
~λ,z
~λ,z
~λ,z
~λ,z
~λ,z
), (u~λ,z − u′
~λ,z
)+)H
)−k · k(u~λ,z − u′
~λ,z
)+k]
(2.39)
Let us shrink δ > 0 in Step 2 so that δ < a1
8nM . Then this and (2.37) lead to
0 = (P ⊥∇L~λ(z + u~λ,z) − P ⊥∇L~λ(z + u′
~λ,z
), (u~λ,z − u′
~λ,z
)+)H ≥
a1
4 k(u~λ,z − u′
~λ,z
)+k2.
This contradicts (u~λ,z − u′
~λ,z
)+ 6= θ.
Similarly, for the third case, as in (2.36) we may use Lemma 2.8(ii)-(iii) to obtain
z)−)H
0 = (P ⊥∇L(z + uz) − P ⊥∇L(z + u′
z), (uz − u′
z)(uz − u′
z)(uz − u′
z)(uz − u′
= (∇L(z + uz) − ∇L(z + u′
= (B(z + tuz + (1 − t)u′
= (B(z + tuz + (1 − t)u′
+ (B(z + tuz + (1 − t)u′
≤ −a0k(uz − u′
≤ −a0k(uz − u′
≤ −a0k(uz − u′
a0
2 k(uz − u′
= −
z), (uz − u′
z)−)H
z), (uz − u′
z)−, (uz − u′
z)+, (uz − u′
z)−k2 + ω(z + tuz + (1 − t)u′
z)−k2 +
z)−k2 +
z)−k2,
z)−)H
z)−)H
z)−)H
z)k(uz − u′
z)−k2 + k(uz − u′
z)−k2
a0
[k(uz − u′
4
a0
2 k(uz − u′
z)−k · k(uz − u′
z)+k2]
z)+k
and hence
0 = (P ⊥∇L~λ(z + u~λ,z) − P ⊥∇L~λ(z + u′
~λ,z
~λ,z
)+k2
≤ −
+
a0
2 k(u~λ,z − u′
nXj=1
λj(G′′
23
), (u~λ,z − u′
~λ,z
)−)H
j (z + τ u~λ,z + (1 − τ )u′
~λ,z
)(u~λ,z − u′
~λ,z
), (u~λ,z − u′
~λ,z
)−)H.
(2.40)
As in (2.39) we may deduce
~λ,z
λj(G′′
nXj=1
≤ nδMk(u~λ,z − u′
≤ nδM [k(u~λ,z − u′
≤ 2nδMk(u~λ,z − u′
j (z + τ u~λ,z + (1 − τ )u′
)(u~λ,z − u′
)k · k(u~λ,z − u′
)−k
)−k2 + k(u~λ,z − u′
)−k2.
~λ,z
~λ,z
~λ,z
~λ,z
~λ,z
~λ,z
), (u~λ,z − u′
~λ,z
)−)H
)−k · k(u~λ,z − u′
~λ,z
)+k]
So if the above δ > 0 is also shrunk so that δ < a0
8nM , we may derive from this and (2.40) that
0 = (P ⊥∇L~λ(z + u~λ,z) − P ⊥∇L~λ(z + u′
~λ,z
), (u~λ,z − u′
~λ,z
)−)H ≤ −
a0
4 k(u~λ,z − u′
~λ,z
)−k2,
which also leads to a contradiction.
As a consequence, we have a well-defined map
ψ : [−δ, δ]n × BH 0(θ, ǫ) → Qr,s, (λ, z) 7→ u~λ,z.
(2.41)
Step 4. ψ is continuous.
Let sequence {~λk}k≥1 ∈ [−δ, δ]n and {zk}k≥1 ⊂ BH 0(θ, ǫ) converge to ~λ0 ∈ [−δ, δ]n and
z0 ∈ BH 0(θ, ǫ), respectively. We want to prove that ψ(~λk, zk) → ψ(~λ0, z0). Since {ψ(~λk, zk)}k≥1
is contained in Qr,s, we can suppose ψ(~λk, zk) ⇀ u0 ∈ Qr,s in H. Noting ψ(~λk, zk) − u0 ∈
H + ⊕ H −, by (2.33) we have
(cid:0)∇L~λk
= (cid:0)P ⊥∇L~λk
(cid:0)∇L~λk
= (cid:0)∇L~λk
(zk + ψ(~λk, zk)), ψ(~λk, zk) − u0)(cid:1)
(zk + ψ(~λk, zk)), ψ(~λk, zk) − u0)(cid:1) = 0.
(zk + ψ(~λk, zk)), (zk + ψ(~λk, zk)) − (z0 + u0)(cid:1)k
(zn + ψ(~λk, zk)), zk − z0(cid:1)
(zk + ψ(~λk, zk))k · kzk − z0k → 0.
It follows from this and (2.31) that
≤ k∇L~λk
As in the proof of Step 1, we may derive from this that zk + ψ(~λk, zk) → z0 + u0 and so
ψ(~λk, zk) → u0.
24
Moreover, (2.33) implies P ⊥∇L~λk
(zk + ψ(~λk, zk)) = 0, k = 1, 2,··· . The C 1-smoothness
(z0 + u0) = 0. By Step 3 we arrive at ψ(~λ0, z0) = u0 and hence
of L and all Gj leads to P ⊥∇L~λ0
ψ is continuous at (~λ0, z0).
Step 5. For any (~λ, zi) ∈ [−δ, δ]n × BH(θ, ǫ) ∩ H 0, i = 1, 2, by the definition of ψ, we have
P ⊥∇L(zi + ψ(~λ, zi)) +
nXj=1
λjP ⊥∇Gj(zi + ψ(~λ, zi)) = θ,
i = 1, 2,
and hence for Ξ = z1 − z2 + ψ(~λ, z1) − ψ(~λ, z2) we derive
0 = (P ⊥∇L(z1 + ψ(~λ, z1)) − P ⊥∇L(z2 + ψ(~λ, z2)), Ξ+)H
λj(P ⊥∇Gj(z1 + ψ(~λ, z1)) − P ⊥∇Gj(z2 + ψ(~λ, z2)), Ξ+)H .
(2.42)
+
nXj=1
As in the proof of (2.36) we obtain τ ∈ (0, 1) such that
(P ⊥∇L(z1 + ψ(~λ, z1)) − P ⊥∇L(z2 + ψ(~λ, z2)), Ξ+)H
= (B(τ z1 + τ ψ(~λ, z1) + (1 − τ )z2 + (1 − τ )ψ(~λ, z2))Ξ+, Ξ+)H
+ (B(τ z1 + τ ψ(~λ, z1) + (1 − τ )z2 + (1 − τ )ψ(~λ, z2))(Ξ0 + Ξ−), Ξ+)H
a1
[kΞ− + Ξ0k2 + kΞ+k2]
≥ a1kΞ+k2 −
4
3a1
a1
a1
4 kΞ−k2.
4 kΞ+k2 −
4 kΞ0k2 −
=
Let us further shrink δ > 0 in Step 3 so that δ < min{a0,a1}
16nM . As in (2.39) we may deduce
λj(P ⊥∇Gj(z1 + ψ(~λ, z1)) − P ⊥∇Gj(z2 + ψ(~λ, z2)), Ξ+)H
nXj=1
nXj=1
≤
≤ nδMkΞk · kΞ+k ≤ 2nδM [kΞk2 + kΞ+k2]
[kΞ−k2 + kΞ0k2 + 2kΞ+k2].
≤
λj(G′′
a1
8
j (τ z1 + (1 − τ )z2 + τ ψ(~λ, z1) + (1 − τ )ψ(~λ, z2))Ξ, Ξ+)H
(2.43)
(2.44)
(2.45)
This and (2.42) -- (2.43) lead to
3a1
4 kΞ+k2 −
a1
4 kΞ0k2 −
a1
4 kΞ−k2 −
a1
8
0 ≥
[kΞ−k2 + kΞ0k2 + 2kΞ+k2]
and so
0 ≥ 4kΞ+k2 − 3kΞ0k2 − 3kΞ−k2.
(2.46)
Similarly, replacing Ξ+ by Ξ− in (2.43) and (2.44) we derive
25
a0
4 kΞ+k2,
a0
4 kΞ0k2 +
(P ⊥∇L(z1 + ψ(~λ, z1)) − P ⊥∇L(z2 + ψ(~λ, z2)), Ξ−)H
3a0
4 kΞ−k2 +
nXj=1
λj(P ⊥∇Gj(z1 + ψ(~λ, z1)) − P ⊥∇Gj(z2 + ψ(~λ, z2)), Ξ−)H
[kΞ+k2 + kΞ0k2 + 2kΞ−k2].
≤ −
a0
8
≤
As above these two inequalities and the equality
0 = (P ⊥∇L(z1 + ψ(~λ, z1)) − P ⊥∇L(z2 + ψ(~λ, z2)), Ξ−)H
λj(P ⊥∇Gj(z1 + ψ(~λ, z1)) − P ⊥∇Gj(z2 + ψ(~λ, z2)), Ξ−)H
+
nXj=1
yield: 0 ≥ 4kΞ−k2 − 3kΞ0k2 − 3kΞ+k2. Combing with (2.46) we obtain
kΞ+ + Ξ−k2 = kΞ+k2 + kΞ−k2 ≤ 6kΞ0k2.
Note that Ξ9 = z1 − z2 and Ξ+ + Ξ− = ψ(~λ, z1) − ψ(~λ, z2). The desired claim is proved.
Step 6. The uniqueness of ψ implies that it is equivariant on z.
As a by-product we have also the following result though it is not used in this paper.
Theorem 2.15 (Inverse Function Theorem). If the assumptions of Theorem 2.1 hold with X = H,
then ∇L is locally invertible at θ.
Proof. We can assume that ∇L is of class (S)+ in Qr,s. Since H 0 = {θ} and ∇L = f0, we have
deg(∇L,Qr,s, θ) = (−1)µ
(2.47)
by (2.22). Moreover, := inf{k∇L(u)k u ∈ ∂Qr,s} > 0 by Lemma 2.12. For any given v ∈
BH(θ, ), let us define
H : [0, 1] × Qr,s → H, (t, u) 7→ ∇L(u) − tv.
Then kH (t, u)k = k∇L(u)−tvk ≥ k∇L(u)k−kvk ≥ −kvk > 0 for all (t, u) ∈ [0, 1]×∂Qr,s.
Assume that sequences tn → t in [0, 1], {un}n≥1 ⊂ Qr,s converges weakly to u in H, and they
satisfy lim supn→∞(H (tn, un), un − u)H ≤ 0. Then
(∇L(un), un − u)H = (H (tn, un), un − u)H + tn(v, un − u)H
leads to lim supn→∞(∇L(un), un − u)H ≤ 0. It follows that un → u in H because ∇L is of
class (S)+ in Qr,s. Hence H is a homotopy of class (S)+, and thus (2.47) gives
deg(∇L − v,Qr,s, θ) = deg(∇L,Qr,s, θ) = (−1)µ,
26
which implies ∇L(ξv) = v for some ξv ∈ Qr,s. By Step 3 in the proof of Theorem 2.15 (taking
~λ = ~0) it is easily seen that the equation ∇L(u) = v has a unique solution in Qr,s, and in
particular, ξv is unique. Then we get a map BH(θ, ) → Qr,s, v 7→ ξv to satisfy ∇L(ξv) = v
for all v ∈ BH(θ, ). We claim that this map is continuous. Arguing by contradiction, assume that
there exists a sequence vn → v in BH(θ, ), such that ξvn ⇀ ξ∗ in H and kξvn − ξvk ≥ ǫ0 for
some ǫ0 > 0 and all n = 1, 2,··· . Note that
(∇L(ξvn), ξvn − ξ∗)H = (vn, ξvn − ξ∗)H = (vn − v, ξvn − ξ∗)H + (v, ξvn − ξ∗)H → 0.
We derive that ξvn → ξ∗ in H, and so ∇L(ξvn) = vn can lead to ∇L(ξ∗) = v. The uniqueness
of solutions implies ξ∗ = ξv. This prove the claim. Hence ∇L is a homeomorphism from an open
neighborhood {ξv v ∈ BH(θ, )} of θ in Qr,s onto BH(θ, ).
Theorem 2.15 cannot be derived from the invariance of domain theorem (5.4.1) of Berger [4]
or [29, Theorem 2.5]. Recently, Ekeland proved an weaker inverse function theorem, [25, Theo-
rem 2]. Since we cannot insure that B(u) has a right-inverse L(u) which is uniformly bounded in
a neighborhood of θ, [25, Theorem 2] it cannot lead to Theorem 2.15 either.
2.5 Parameterized splitting and shifting theorems
To shorten the proof of the main theorem, we shall write parts of it into two propositions.
Proposition 2.16. Under the assumptions of Theorem 2.14, for each (~λ, z) ∈ [−δ, δ]n×BH 0(θ, ǫ),
let ψ~λ(z) = ψ(~λ, z) be given by (2.25). Then it satisfies
Lλ(z + ψλ(z)) = min{Lλ(z + u) u ∈ BH(θ, r) ∩ H +}
if H − = {θ}, and
L~λ(z + ψ~λ(z)) = min{L~λ(z + u + P −ψ~λ(z)) u ∈ BH(θ, r) ∩ H +},
L~λ(z + ψ~λ(z)) = max{L~λ(z + P +ψ~λ(z) + v) v ∈ BH(θ, s) ∩ H −}
if H − 6= {θ}.
Proof.
Case H − = {θ}. We have Qr,s = BH(θ, r) ∩ H +, and (2.26) becomes
P +∇L~λ(z + ψ~λ(z)) = 0,
∀z ∈ BH 0(θ, ǫ).
By this we may use integral mean value theorem to deduce that for each u ∈ Qr,s,
27
L~λ(z + u) − L~λ(z + ψ~λ(z))
0
0
0
0
0
(∇L(z + ψ~λ(z) + τ (u − ψ~λ(z))) − ∇L(z + ψ~λ(z)), u − ψ~λ(z))H dτ
(∇Gj(z + ψ~λ(z) + τ (u − ψ~λ(z))) − ∇Gj(z + ψ~λ(z)), u − ψ~λ(z))H dτ
(∇L~λ(z + ψ~λ(z) + τ (u − ψ~λ(z))) − ∇L~λ(z + ψ~λ(z)), u − ψ~λ(z))H dτ
(∇L~λ(z + ψ~λ(z) + τ (u − ψ~λ(z))), u − ψ~λ(z))H dτ
(P +∇L~λ(z + ψ~λ(z) + τ (u − ψ~λ(z))), u − ψ~λ(z))H dτ
(P +∇L~λ(z + ψ~λ(z) + τ (u − ψ~λ(z))) − P +∇L~λ(z + ψ~λ(z)), u − ψ~λ(z))H dτ
= Z 1
= Z 1
= Z 1
= Z 1
= Z 1
λjZ 1
nXj=1
τ dτZ 1
= Z 1
0 (cid:0)B(z + ψ~λ(z) + ρτ (u − ψ~λ(z)))(u − ψ~λ(z)), u − ψ~λ(z)(cid:1)Hdρ
τ dτZ 1
λjZ 1
nXj=1
0 (cid:0)G′′
a1
2 ku − ψ~λ(z)k2
τ dτZ 1
λjZ 1
nXj=1
j (z + ψ~λ(z) + ρτ (u − ψ~λ(z)))(u − ψ~λ(z)), u − ψ~λ(z)(cid:1)Hdρ.
0 (cid:0)G′′
j (z + ψ~λ(z) + ρτ (u − ψ~λ(z)))(u − ψ~λ(z)), u − ψ~λ(z)(cid:1)H dρ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
τ dτZ 1
0 (cid:0)G′′
≤ 2nδMku − ψ~λ(z))k2 ≤
j (z + ψ~λ(z) + ρτ (u − ψ~λ(z)))(u − ψ~λ(z)), u − ψ~λ(z)(cid:1)Hdρ
a1
4 ku − ψ~λ(z))k2.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
nXj=1
λjZ 1
0
0
0
0
≥
+
0
+
+
Here the final inequality comes from Lemma 2.8(i). For the final sum, as in (2.39) we have
These lead to
L~λ(z + u) − L~λ(z + ψ~λ(z)) ≥
a1
4 ku − ψ~λ(z)k2,
(2.48)
which implies the desired conclusion.
Case H − 6= {θ}. For each u ∈ BH(θ, r)∩ H + we have u + P −ψ~λ(z) ∈ Qr,s. As above we
can use (2.26) to derive
L~λ(z + u + P −ψ~λ(z)) − L~λ(z + ψ~λ(z)) ≥
a1
4 ku − P +ψ~λ(z)k2,
(2.49)
and therefore the second equality.
28
Finally, for each v ∈ BH(θ, r) ∩ H − we have P +ψ~λ(z) + v ∈ Qr,s. As above, using (2.26)
and Lemma 2.8(ii)-(iii) we may deduce
0
0
0
0
L~λ(z + P +ψ~λ(z) + v) − L~λ(z + ψ~λ(z))
= Z 1
(∇L~λ(z + ψ~λ(z) + t(u − P +ψ~λ(z))), v − P −ψ~λ(z))H dt
= Z 1
(∇L~λ(z + ψ~λ(z) + t(v − P −ψ~λ(z))) − ∇L~λ(z + ψ~λ(z)), v − P −ψ~λ(z))H dt
tZ 1
= Z 1
(B(z + ψ~λ(z) + τ t(v − P −ψ~λ(z)))(v − P −ψ~λ(z)), v − P −ψ~λ(z))H dtdτ
tdtZ 1
λjZ 1
nXj=1
0 (cid:0)G′′
a0
2 kv − P −ψ~λ(z)k2
tdtZ 1
λjZ 1
nXj=1
0 (cid:0)G′′
a0
4 kv − P −ψ~λ(z)k2,
j (z + ψ~λ(z) + τ t(v − P −ψ~λ(z)))(v − P −ψ~λ(z)), v − P −ψ~λ(z)(cid:1)Hdτ
j (z + ψ~λ(z) + τ t(v − P −ψ~λ(z)))(v − P −ψ~λ(z)), v − P −ψ~λ(z)(cid:1)Hdτ
≤ −
≤ −
+
+
0
0
and hence the third equality.
Proposition 2.17. Under the assumptions of Theorem 2.14, for each (~λ, z) ∈ [−δ, δ]n×BH 0(θ, ǫ),
let ψ~λ(z) = ψ(~λ, z) be given by (2.25). Then the functional L◦
: BH(θ, ǫ) ∩ H 0 → R given by
nXj=1
(z) := L~λ(z + ψ~λ(z)) = L(z + ψ(~λ, z)) +
λjGj(z + ψ(~λ, z))
(2.50)
L◦
~λ
~λ
is of class C 1, and its differential is given by
DL◦
~λ
(z)h = DL~λ(z + ψ~λ(z))h
= DL(z + ψ(~λ, z))h +
λjDGj(z + ψ(~λ, z))h,
nXj=1
~λ ∈ C 1( ¯BH (θ, ǫ) ∩ H 0) is continuous by shrinking
∀h ∈ H 0.
(2.51)
(Clearly, this implies that [−δ, δ]n ∋ ~λ 7→ L◦
ǫ > 0 since dim H 0 < ∞).
Proof. Case H − 6= {θ}. For fixed z ∈ BH (θ, ǫ) ∩ H 0, h ∈ H 0, and t ∈ R with sufficiently
small t, the last two equalities in Proposition 2.16 imply
L~λ(z + th + P +ψ~λ(z + th) + P −ϕ(z)) − L~λ(z + P +ψ~λ(z + th) + P −ψ~λ(z))
≤ L~λ(z + th + ψ~λ(z + th)) − L~λ(z + ψ~λ(z))
≤ L~λ(z + th + P +ψ~λ(z) + P −ψ~λ(z + th)) − L~λ(z + P +ψ~λ(z) + P −ψ~λ(z + th)).
(2.52)
Since L~λ is C 1 and ψ~λ is continuous we deduce,
29
lim
t→0
L~λ(z + th + P +ψ~λ(z + th) + P −ψ~λ(z)) − L~λ(z + P +ψ~λ(z + th) + P −ψ~λ(z))
t→0Z 1
DL~λ(z + sth + P +ψ~λ(z + th) + P −ψ~λ(z))hds
t
0
= lim
(2.53)
= DL~λ(z + ψ~λ(z))h.
Here the last equality follows from the Lebesgue's Dominated Convergence Theorem since
{DL~λ(z + sth + P +ψ~λ(z + th) + P −ψ~λ(z))h 0 ≤ s ≤ 1, t ≤ 1}
is bounded by the compactness of {z + sth + P +ψ~λ(z + th) + P −ψ~λ(z) 0 ≤ s ≤ 1, t ≤ 1}.
Similarly, we have
L~λ(z + th + P +ψ~λ(z) + P −ψ~λ(z + th)) − L~λ(z + P +ψ~λ(z) + P −ψ~λ(z + th))
t
lim
t→0
= DL~λ(z + ψ~λ(z))h.
Then it follows from (2.52)-(2.54) that
(2.54)
L~λ(z + th + ψ~λ(z + th)) − L~λ(z + ψ~λ(z))
t
lim
t→0
= DL~λ(z + ψ~λ(z))h.
That is, L◦
L◦
~λ
~λ
is Gateaux differentiable and DL◦
~λ
(z) = DL~λ(z + ψ~λ(z))H 0 . The latter implies that
is of class C 1 because both DL~λ and ψ~λ are continuous.
Case H − = {θ}. For fixed z ∈ BH(θ, ǫ) ∩ H 0 and h ∈ H 0, and t ∈ R with sufficiently
small t, the first equality in Proposition 2.16 implies
L~λ(z + th + ψ~λ(z + th)) − L~λ(z + ψ~λ(z + th))
≤ L~λ(z + th + ψ~λ(z + th)) − L~λ(z + ψ~λ(z))
≤ L~λ(z + th + ψ~λ(z)) − L~λ(z + ψ~λ(z)).
By the continuity of ∇L~λ and ψ~λ we have
L~λ(z + th + ψ~λ(z + th)) − L~λ(z + ψ~λ(z + th))
t
DL~λ(z + sth + ψ~λ(z + th))hds
lim
t→0
t→0Z 1
0
= lim
= DL~λ(z + ψ~λ(z))h.
(2.55)
(2.56)
(As above this follows from the Lebesgue's Dominated Convergence Theorem because {z + sth+
ψ~λ(z + th) 0 ≤ s ≤ 1, 0 ≤ t ≤ 1} is compact and thus {∇L~λ(z + sth + ψ~λ(z + th))h 0 ≤
s ≤ 1, t ≤ 1} is bounded). Similarly, we may prove
L~λ(z + th + ψ~λ(z)) − L~λ(z + ψ~λ(z))
t
lim
t→0
= DL~λ(z + ψ~λ(z))h,
(2.57)
30
and thus
L~λ(z + th + ψ~λ(z + th)) − L~λ(z + ψ~λ(z))
t
lim
t→0
= DL~λ(z + ψ~λ(z))h
by (2.55) -- (2.57). The desired claim follows immediately.
Theorem 2.18 (Parameterized Splitting Theorem). Under the assumptions of Theorem 2.14, by
shrinking δ > 0, ǫ > 0 and r > 0, s > 0, we obtain an open neighborhood W of θ in H and an
origin-preserving homeomorphism
[−δ, δ]n × BH 0(θ, ǫ) × (BH +(θ, r) + BH −(θ, s)) → [−δ, δ]n × W,
(~λ, z, u+ + u−) 7→ (~λ, Φ~λ(z, u+ + u−))
such that
L~λ ◦ Φ~λ(z, u+ + u−) = ku+k2 − ku−k2 + L~λ(z + ψ(~λ, z))
(2.58)
(2.59)
for all (~λ, z, u+ + u−) ∈ [−δ, δ]n × BH 0(θ, ǫ) × (BH +(θ, r) + BH −(θ, s)), where ψ is given by
(2.25). The functional functional L◦
: BH(θ, ǫ) ∩ H 0 → R given by (2.50) is of class C 1, and its
differential is given by (2.51). Moreover, (i) if L and Gj , j = 1,··· , n, are of class C 2−0, then so
for each ~λ ∈ [−δ, δ]n; (ii) if a compact Lie group G acts on H orthogonally, and V , L and
is L◦
G are G-invariant (and hence H 0, (H 0)⊥ are G-invariant subspaces), then for each ~λ ∈ [−δ, δ]n,
ψ(~λ,·) and Φ~λ(·,·) are G-equivariant, and L◦
(z) = L~λ(z + ψ(~λ, z)) is G-invariant.
~λ
~λ
~λ
Sometimes, for example, the corresponding conditions with [39, Theorem 1.1] or [40, Re-
is of class C 2;
mark 3.2] are also satisfied, we can prove that ψ(~λ,·) is of class C 1 and that L◦
~λ
moreover,
~λ
DL◦
d2L◦
~λ
(z)u = (∇L~λ(z + ψ(~λ, z)), u)H ,
(z)(u, v) = (L′′
~λ
(z + ψ(~λ, z))(u + Dzψ(~λ, z)u), v)H
(2.60)
(2.61)
for all z ∈ BH(θ, ǫ) ∩ H 0 and u, v ∈ H 0. In particular, since ψ(~λ, θ) = θ and Dzψ(~λ, θ) = θ,
d2L◦
~λ
(θ)(z1, z2) = (L′′
~λ
(θ)z1, z2)H = −
nXj=1
λj(G′′
j (θ)z1, z2)H ,
∀z1, z2 ∈ H 0.
(2.62)
Claim.
In this situation, if θ ∈ H is a nondegenerate critical point of L~λ then θ ∈ H 0 such a
critical point of L◦
~λ
too.
~λ
(θ)z1, u)H = (P 0L′′
In fact, suppose for some z1 ∈ H 0 that d2L◦
(θ)(z1, z2) = 0 ∀z2 ∈ H 0. (2.62) implies
(P 0L′′
(θ)z1 = θ. Moreover,
since (I − P 0)∇L~λ(z + ψ(~λ, z)) = θ for all z ∈ BH(θ, ǫ)∩ H 0. Differentiating this equality with
(z + ψ(~λ, z))(u + Dzψ(~λ, z)u) = θ for all u ∈ H 0. In particular,
respect to z we get (I − P 0)L′′
(I − P 0)L′′
(θ)z1, P 0u)H = 0 for all u ∈ H. Hence P 0L′′
(θ)z = θ for all z ∈ H 0. It follows that L′′
(θ)z1 = θ and hence z1 = θ.
~λ
~λ
~λ
~λ
~λ
~λ
Proof of Theorem 2.18. Let N = H 0, and for each ~λ ∈ [−δ, δ]n we define a map F~λ : BN (θ, ǫ)×
Qr,s → R by
F~λ(z, u) = L~λ(z + ψ(~λ, z) + u) − L~λ(z + ψ(~λ, z)).
31
(2.63)
Then D2F~λ(z, u)v = (P ⊥∇Lλ(z + ψ(~λ, z) + u), v)H for z ∈ ¯BN (θ, ǫ), u ∈ Qr,s and v ∈ N ⊥.
Moreover it holds that
F~λ(z, θ) = 0 and D2F~λ(z, θ)(v) = 0
∀v ∈ N ⊥.
(2.64)
Since BN (θ, ǫ) ⊕ Qr,s has the closure contained in the neighborhood U in Lemma 2.8, and
ψ(~λ, θ) = θ, we can shrink ν > 0, ǫ > 0, r > 0 and s > 0 so small that
z + ψ(~λ, z) + u+ + u− ∈ U
(2.65)
for all ~λ ∈ [−δ, δ]n, z ∈ ¯BN (θ, ǫ) and u+ + u− ∈ Qr,s.
Let us verify that each F~λ satisfies conditions (ii)-(iv) in [40, Theorem A.1].
1 , u−
[D2F~λ(z, u+ + u−
Step 1. For ~λ ∈ [−δ, δ]n, z ∈ ¯BN (θ, ǫ), u+ ∈ ¯BH +(θ, r) and u−
2 ) − D2F~λ(z, u+ + u−
1 )](u−
2 − u−
1 )H
2 − u−
= (∇L~λ(z + ψ~λ(z) + u+ + u−
−(∇L~λ(z + ψ~λ(z) + u+ + u−
Since ∇L~λ is Gateaux differentiable so is the function
2 ), u−
1 ), u−
1 )H .
u 7→ (∇L~λ(z + ψ~λ(z) + u+ + u), u−
2 − u−
1 )H .
By the mean value theorem we have t ∈ (0, 1) such that
2 ∈ ¯BH −(θ, ǫ), we have
2 − u−
1 )
(2.66)
j (z + ψ~λ(z) + u+ + u−
2 ), u−
1 + t(u−
+
(∇L~λ(z + ψ~λ(z) + u+ + u−
= (cid:0)B(z + ψ~λ(z) + u+ + u−
nXj=1
λj(cid:0)G′′
nXj=1
λj(cid:0)G′′
2 − u−
−a0ku−
1 k2
≤
j (z + ψ~λ(z) + u+ + u−
1 + t(u−
2 − u−
2 − u−
1 + t(u−
1 ), u−
2 − u−
2 − u−
1 )H − (∇L~λ(z + ψ~λ(z) + u+ + u−
1(cid:1)H
1 ))(u−
2 − u−
1(cid:1)H
2 − u−
2 − u−
1(cid:1)H
2 − u−
2 − u−
2 − u−
1 ))(u−
1 ))(u−
1 ), u−
1 ), u−
1 ), u−
2 − u−
1 )H
(2.67)
because of Lemma 2.8(iii). Recall that we have assumed δ < min{a0,a1}
Theorem 2.14. From this and (2.38) it follows that
8nM
in Step 3 of the proof of
j (z + ψ~λ(z) + u+ + u−
1 + t(u−
2 − u−
1 ))(u−
2 − u−
1 ), u−
1(cid:1)H
2 − u−
nXj=1
λj(cid:0)G′′
≤ nδMku−
2 − u−
1 k2 ≤
a0
8 ku−
2 − u−
1 k2.
32
This and (2.66) -- (2.67) lead to
[D2F~λ(z, u+ + u−
2 ) − D2F~λ(z, u+ + u−
1 )](u−
2 − u−
1 ) ≤ −
a0
2 ku−
2 − u−
1 k2.
This implies the condition (ii) of [40, theorem A.1].
Step 2. For ~λ ∈ [−δ, δ]n, z ∈ ¯BN (θ, ǫ), u+ ∈ ¯BH +(θ, r) and u− ∈ ¯BH −(θ, s), by (2.64) and the
mean value theorem, for some t ∈ (0, 1) we have
D2F~λ(z, u+ + u−)(u+ − u−)
+
+
= D2F~λ(z, u+ + u−)(u+ − u−) − D2F~λ(z, θ)(u+ − u−)
= (∇L~λ(z + ψ~λ(z) + u+ + u−), u+ − u−)H − (∇L~λ(z + ψ~λ(z)), u+ − u−)H
= (cid:0)B(z + ψ~λ(z) + t(u+ + u−))(u+ + u−), u+ − u−(cid:1)H
nXj=1
= (cid:0)B(z + ψ~λ(z) + t(u+ + u−))u+, u+(cid:1)H −(cid:0)B(z + ψ~λ(z) + t(u+ + u−))u−, u−(cid:1)H
nXj=1
≥ a1ku+k2 + a0ku−k2
nXj=1
+
λj(cid:0)G′′
j (z + ψ~λ(z) + t(u+ + u−))(u+ + u−), u+ − u−(cid:1)H
j (z + ψ~λ(z) + t(u+ + u−))(u+ + u−), u+ − u−(cid:1)H
λj(cid:0)G′′
j (z + ψ~λ(z) + t(u+ + u−))(u+ + u−), u+ − u−(cid:1)H
λj(cid:0)G′′
nXj=1
λj(cid:0)G′′
≤ nδMku+ + u−k · ku+ − u−k ≤
≤
j (z + ψ~λ(z) + t(u+ + u−))(u+ + u−), u+ − u−(cid:1)H
(ku+k2 + ku−k2)
a1
4 ku+k2 +
a0
4 ku−k2.
min{a0, a1}
4
(2.68)
because of Lemma 2.8(i) and (iii). As above we have
From this and (2.68) we deduce
D2F~λ(z, u+ + u−)(u+ − u−) ≥ a1
2 ku+k2 + a0
2 ku−k2.
The condition (iii) of [40, Theorem A.1] is satisfied.
Step 3. For ~λ ∈ [−δ, δ]n, z ∈ ¯BN (θ, ǫ) and u+ ∈ ¯BH +(θ, r), as above we have t ∈ (0, 1) such
that
D2F~λ(z, u+)u+ = D2F~λ(z, u+)u+ − D2F~λ(z, θ)u+
= (∇L~λ(z + ψ~λ(z) + u+), u+)H − (∇L~λ(z + ψ~λ(z)), u+)H
= (cid:0)B(z + ψ~λ(z) + tu+)u+, u+(cid:1)H +
≥ a1ku+k2 +
j (z + ψ~λ(z) + tu+)u+, u+(cid:1)H
λj(cid:0)G′′
λj(cid:0)G′′
nXj=1
nXj=1
j (z + ψ~λ(z) + tu+)u+, u+(cid:1)H
33
because of Lemma 2.8(i). Moreover, it is proved as before that
nXj=1
λj(cid:0)G′′
j (z + ψ~λ(z) + tu+)u+, u+(cid:1)H ≤
min{a0, a1}
8
ku+k2.
Hence we obtain
D2F~λ(z, u+)u+ ≥ a1ku+k2 > p(ku+k) ∀u+ ∈ ¯BH +(θ, s) \ {θ},
where p : (0, ε] → (0,∞) is a non-decreasing function given by p(t) = a1
4 t2. Namely, F~λ satisfies
the condition (iv) of [40, Theorem A.1] (the parameterized version of [23, Theoren 1.1]).
The other arguments are as before.
Step 4. The claim (i) in the part of "Moreover" follows from (2.27) directly. For the second one,
since ψ(λ,·) is G-equivariant, and Lλ is G-invariant, we derive from (2.63) that F~λ is G-invariant.
By the construction of Φ~λ(·,·) (cf. [23] and [39, Theorem A.1]), it is expressed by F~λ(z,·), one
easily sees that Φ~λ(·,·) is G-equivariant.
Theorem 2.19 (Parameterized Shifting Theorem). Suppose for some ~λ ∈ [−δ, δ]n that θ ∈ H is
an isolated critical point of L~λ (thus θ ∈ H 0 is that of L◦
). Then
✷
~λ
Cq(L~λ, θ; K) = Cq−µ(L◦
~λ
, θ; K) ∀q ∈ N ∪ {0},
(2.69)
~λ
(z) = L~λ(z + ψ(~λ, z)) = L(z + ψ(~λ, z)) +Pn
j=1 λjGj(z + ψ(~λ, z)) is as in (2.50).
are only of class C 1, the construction of the Gromoll-Meyer pair on
where L◦
Proof. Though L~λ and L◦
the pages 49-51 of [14] is also effective for them (see [16]). Hence the result can be obtained by
~λ
repeating the proof of [14, Theorem I.5.4].
Of course, with a stability theorem of critical groups the present case can also be reduced to
that of [14, Theorem I.5.4]. In fact, by Theorem 2.18 we have C∗(L~λ, θ; K) = C∗(bL~λ, θ; K),
where
bL~λ : BH 0(θ, ǫ) × (H + ⊕ H −) → R,
(z, u+ + u−) 7→ ku+k2 − ku−k2 + L◦
~λ
(z).
By a smooth cut function we can construct a C 1 functional f : H 0 → R such that f (z) = L◦
for kzk < ǫ/2, and f (z) = kzk2 for kzk ≥ 3ǫ/4. Define a functional
~λ
(z)
eL : H 0 × (H + ⊕ H −) → R, (z, u+ + u−) 7→ ku+k2 − ku−k2 + f (z).
Clearly, C∗(bL~λ, θ; K) = C∗(eL, θ; K). Since θ ∈ H 0 is an isolated critical point of L◦
ing ǫ > 0 we can assume that k∇L◦
differential topology claims that C ∞(H 0, R) is dense in C 1
ogy). Hence we can choose a function g ∈ C ∞(H 0, R) to satisfy
, by shrink-
(z)k > 0 for all z ∈ BH 0(θ, ǫ) \ {θ}. A standard result in
S(H 0, R) (equipped with strong topol-
~λ
~λ
1
2k∇f (z)k,
k∇f (z) − ∇g(z)k <
1
kf (z) − g(z)k <
2kf (z)k,
∀z ∈ BH 0(θ, ǫ) \ {θ},
∀z ∈ H 0 \ BH 0(θ, 10ǫ).
34
They imply respectively that
k∇((1 − t)f + tg)(z)k >
k(1 − t)f (z) + tg(z)k >
1
2k∇L◦(z)k > 0,
1
1
2kzk2,
2kf (z)k =
∀z ∈ BH 0(θ, ǫ/2) \ {θ},
∀z ∈ H 0 \ BH 0(θ, 10ǫ)
(2.70)
(2.71)
for all t ∈ [0, 1]. Hence each functional ft : H 0 → R, z 7→ (1 − t)f (z) + tg(z) has a unique
critical point θ in BH 0(θ, ǫ/2) by (2.70), and satisfies (PS) condition by (2.71) and finiteness of
dim H 0. It follows from the stability theorem of critical groups ([22, Theorem 5.2]) that
C∗(L◦
~λ
, θ; K) = C∗(f, θ; K) = C∗(ft, θ; K) = C∗(g, θ; K),
∀t ∈ [0, 1].
(2.72)
Since the functionals ft and H +⊕H − ∋ u++u− 7→ ku+k2−ku−k2 ∈ R satisfies (PS) condition,
so is each functional
eLt : H 0 × (H + ⊕ H −) → R, (z, u+ + u−) 7→ ku+k2 − ku−k2 + ft(z).
As above we derive from [22, Theorem 5.2] that C∗(bL, θ; K) = C∗(eL1, θ; K). But [13, Theo-
rem I.5.4] or [50, Theorem 8.4] implies C∗(eL1, θ; K) = C∗−µ(g, θ; K). Hence
C∗(L~λ, θ; K) = C∗(bL~λ, θ; K) = C∗(eL1, θ; K) = C∗−µ(g, θ; K) = C∗−µ(L◦
, θ; K).
~λ
2.6 Splitting and shifting theorems around critical orbits
Let H be a Hilbert space with inner product (·,·)H and let (H, ((·,·))) be a C 3 Hilbert-Riemannian
manifold modeled on H. Let O ⊂ H be a compact C 3 submanifold without boundary, and let
π : NO → O denote the normal bundle of it. The bundle is a C 2-Hilbert vector bundle over O,
and can be considered as a subbundle of TOH via the Riemannian metric ((·,·)). The metric ((·,·))
induces a natural C 2 orthogonal bundle projection Π : TOH → NO. For ε > 0 we denote by
NO(ε) := {(x, v) ∈ NO kvkx < ε},
the so-called normal disk bundle of radius ε. If ε > 0 is sufficiently small the exponential map exp
gives a C 2-diffeomorphism from NO(ε) onto an open neighborhood of O in H, N (O, ε).
For x ∈ O, let Ls(NOx) denote the space of those operators S ∈ L (NOx) which are self-
adjoint with respect to the inner product ((·,·))x, i.e. ((Sxu, v))x = ((u, Sxv))x for all u, v ∈ NOx.
Then we have a C 1 vector bundle Ls(NO) → O whose fiber at x ∈ O is given by Ls(NOx).
Let L : H → R be a C 1 functional. A connected C 3 submanifold O ⊂ H is called a critical
manifold of L if LO = const and DL(x)v = 0 for any x ∈ O and v ∈ TxH. If there exists a
neighborhood V of O such that V \ O contains no critical points of L we say O to be isolated.
Furthermore, we make the following
35
Hypothesis 2.20. The gradient field ∇L : H → TH is Gateaux differentiable and thus we
have a bounded linear self-adjoint operator d2L(x) ∈ Ls(TxH) for each x ∈ O; moreover,
O ∋ x 7→ d2L(x) is a continuous section of Ls(TH) → O is continuous, dim Ker(d2L(x)) =
const ∀x ∈ O, and there exists a0 > 0 such that
σ(d2L(x)) ∩ ([−2a0, 2a0] \ {0}) = ∅ ∀x ∈ O.
(2.73)
This implies that O ∋ x 7→ Bx(θx) := Πx ◦ d2L(x)N Ox = d2(L ◦ expx N Ox)(θx) is a
continuous section of Ls(NO → O, dim Ker(Bx(θx)) = const ∀x ∈ O, and
σ(Bx(θx)) ∩ ([−2a0, 2a0] \ {0}) = ∅ ∀x ∈ O.
Let χ∗ (∗ = +,−, 0) be the characteristic function of the intervals [2a0, +∞), (−2a0, a0)
and (−∞,−2a0], respectively. Then we have the orthogonal bundle projections on the normal
x (v) = χ∗(Bx(θx))v), ∗ = +,−, 0. Denote by N ∗O = P ∗NO,
bundle NO, P ∗ (defined by P ∗
∗ = +,−, 0. (Clearly, Bx(θx)(N ∗Ox) ⊂ N ∗Ox for any x ∈ O and ∗ = +,−, 0). By [13,
Lem.7.4], we have NO = N +O ⊕ N −O ⊕ N 0O. If rankN 0O = 0, the critical orbit O is called
nondegenerate.
In the following we only consider the case O is a critical orbit of a compact Lie group. The
general case can be treated as in [42]. The following assumption implies naturally Hypothesis 2.20.
Hypothesis 2.21. (i) Let G be a compact Lie group, and let H be a C 3 G-Hilbert manifold. (So
TH is a C 2 G-vector bundle, i.e. for any g ∈ G the induced action G × TH → TH given by
g · (x, v) = (g · x, dg(x)v) is a C 1 bundle map satisfying gTxH = Tg·xH ∀x ∈ H). Furthermore,
this action also preserves the Riemannian-Hilbert structure ((·,·)), i.e.
((g · u, g · v))g·x = ((u, v))x,
∀x ∈ H,
∀u, v ∈ TxH.
(In this case(cid:0)H, ((·,·))(cid:1) is said to be a C 2 G-Riemannian-Hilbert manifold).
(ii) The C 1 functional L : H → R is G-invariant, ∇L : H → TH is Gateaux differentiable,
(i.e., under any C 3 local chart the functional L has a representation that is C 1 and has a Gateaux
differentiable gradient map), and O is an isolated critical orbit which is a C 3 critical submanifold
with Morse index µO.
Since expg·x(g · v) = g · expx(v) for any g ∈ G and (x, v) ∈ TH, we have
L ◦ exp(g · x, g · v) = L(exp(g · x, g · v)) = L(g · exp(x, v)) = L(exp(x, v)).
It follows that ∇L(g · x) = g−1 · ∇L(x)g and
for any g ∈ G and (x, v) ∈ NO(ε)x, which leads to
∇(cid:0)L ◦ expN O(ε)gx(cid:1) (g · v) = g · ∇(cid:0)L ◦ expN O(ε)x(cid:1) (v)
d2(cid:0)L ◦ expN O(ε)gx(cid:1) (g · v) · g = g · d2(cid:0)L ◦ expN O(ε)x(cid:1) (v)
as bounded linear operators from NOx onto NOgx.
(2.74)
(2.75)
36
Theorem 2.22. Under Hypothesis 2.21, let for some x0 ∈ O the pair(cid:0)L ◦ expx0, BTx0 H(θ, ε)(cid:1)
(and so the pair (L ◦ expN O(ε)x0
, NO(ε)x0) by Lemma 2.10) satisfies the corresponding con-
ditions in Hypothesis 1.1 with X = H. Suppose that the critical orbit O is nondegenerate. Then
there exist ǫ > 0 and a G-equivariant homeomorphism onto an open neighborhood of the zero
section preserving fibers, Φ : N +O(ǫ) ⊕ N −O(ǫ) → NO, such that for any x ∈ O and
(u+, u−) ∈ N +O(ǫ)x × N −O(ǫ)x,
L ◦ exp◦Φ(x, u+ + u−) = ku+k2
x − ku−k2
x + LO.
(2.76)
It naturally leads to a Morse relation if L satisfies the (PS) condition and has only nondegen-
erate critical orbits between regular levels. Theorem 2.22 will be proved after the proof of the
following theorem.
Theorem 2.23. Under Hypothesis 2.21, let for some x0 ∈ O the pair (L◦exp N O(ε)x0
, NO(ε)x0 )
satisfy the corresponding conditions with Hypothesis 1.1 with X = H. Suppose that the critical
orbit O is degenerate, i.e., rankN 0O > 0. Then there exist ǫ > 0, a G-equivariant topological
bundle morphism that preserves the zero section,
h : N 0O(3ǫ) → N +O ⊕ N −O, (x, v) 7→ hx(v),
and a G-equivariant homeomorphism onto an open neighborhood of the zero section preserving
fibers, Φ : N 0O(ǫ) ⊕ N +O(ǫ) ⊕ N −O(ǫ) → NO, such that the following properties hold:
(I) The induced map, hx : N 0O(ǫ)x → TxH, satisfies
(P +
x + P −
x ) ◦ Πx∇(L ◦ expx)(v + hx(v)) = 0 ∀(x, v0) ∈ N 0O(ǫ).
(2.77)
(II) Φ has the form Φ(x, v +u+ +u−) =(cid:0)x, v +hx(v)+φ(x,v)(u+ +u−)(cid:1) with φ(x,v)(u+ +u−) ∈
(N +O ⊕ N −O)x, and satisfies
x − ku−k2
L ◦ exp◦Φ(x, v, u+ + u−) = ku+k2
x + L ◦ expx(v + hx(v))
for any x ∈ O and (v, u+, u−) ∈ N 0O(ǫ)x × N +O(ǫ)x × N −O(ǫ)x. Moreover
(II.1) Φ(x, v) = (x, v + hx(v)) ∀v ∈ N 0O(ǫ)x,
(II.2) φ(x,v)(u+ + u−) ∈ N −O if and only if u+ = θx.
(III) For each x ∈ O the function
(2.78)
N 0O(ǫ)x → R, v 7→ L◦
x(v) := L ◦ expx(v + hx(v))
(2.79)
is Gx-invariant, of class C 1, and satisfies
DL◦
x(v)v′ := (∇(L ◦ expx)(v + hx(v)), v′),
∀v′ ∈ N 0Ox.
Moreover, each hx is of class C 1−0, and hence L◦
x is of class C 2−0 if L is of class C 2−0.
37
Proof. We only outline main procedures. By the assumption and (2.75) we deduce that each
pair (L ◦ expN O(ε)x , NO(ε)x) satisfies the corresponding conditions with Hypothesis 1.1 with
X = H too, and that there exists a0 > 0 such that
By Theorem 2.2 we get a small ǫ ∈ (0, ε/3) and a continuous map
σ(cid:0)d2(cid:0)L ◦ expN O(ε)x(cid:1) (θx)(cid:1) ∩ ([−2a0, 2a0] \ {0}) = ∅,
hx0 : N 0O(3ǫ)x0 → N ±O(ε/2)x0 ,
∀x ∈ O.
(2.80)
such that hx0(g · v) = g · hx0(v), hx0(θx0) = θx0 and
(P +
x0 + P −
x0)∇(cid:16)L ◦ expN O(ε)x0(cid:17) (v + hx0(v)) = 0,
Furthermore, the function
∀v ∈ N 0O(3ǫ)x0.
L◦
x0 : N 0O(ǫ)x0 → R, v 7→ L ◦ expx0(v + hx0(v))
x0(v)u =(cid:0)∇(L ◦ expN O(ε)x0
is of class C 1, and DL◦
h : N 0O(3ǫ) → TH,
)(v + hx0(v)), u(cid:1). Define
(x, v) 7→ g · hx0(g−1 · v),
where g · x0 = x. It is continuous. Otherwise, there exists a sequence {(xj, vj)}j ⊂ N 0O(3ǫ)
which converges to a point (¯x, ¯v) ∈ N 0O(3ǫ), but h(xj , vj) 9 h(¯x, ¯v). Passing to a subsequence
we may assume that {h(xj, vj)}j has no intersection with an open neighborhood U of h(¯x, ¯v) in
TH. Let ¯g, gj ∈ G be such that ¯g · x0 = ¯x and gj · x0 = xj , j = 1, 2,··· . Then h(¯x, ¯v) =
¯g · hx0(¯g−1 · ¯v) and h(xj , vj) = gj · hx0(g−1
· vj) for each j ∈ N. Note that ¯g−1 · U is an open
neighborhood of hx0(¯g−1 · ¯v) = ¯g−1 · h(¯x, ¯v) and that {¯g−1 · h(xj, vj) = ¯g−1 · gj · hx0(g−1
· vj)}j
has no intersection with ¯g−1 · U. Since G is compact, we may assume ¯g−1 · gj → g ∈ G and
so g−1
j → (¯gg)−1 ∈ G after passing to a subsequence (if necessary). Then ¯g−1 · h(xj , vj) =
¯g−1 · gj · hx0(g−1
· vj) → g · hx0((¯gg)−1 · ¯v) = hx0(¯g−1 · ¯v). It follows that hx0(¯g−1 · ¯v) does not
belong to ¯g−1 · U. This contradicts the fact that ¯g−1 · U is an open neighborhood of hx0(¯g−1 · ¯v).
j
j
j
By the definition of h, it is clearly G-equivariant and satisfies
(P +
x + P −
x )∇(cid:0)L ◦ expN Ox(ε)(cid:1) (v + hx(v)) = 0,
Moreover, the map F : N 0O(ǫ) ⊕ N +O(ǫ) ⊕ N −O(ǫ) → R defined by
∀(x, v) ∈ N 0O(3ǫ).
(2.81)
F(x, v, u+ + u−) = Fx(v, u+ + u−)
= L ◦ expx(v + hx(v) + u+ + u−) − L ◦ expx(v + hx(v)), (2.82)
is G-invariant, and satisfies for any (x, v) ∈ N 0O(ǫ) and u ∈ N +Ox ⊕ N −Ox,
and D2Fx(v, θx)[u] = 0.
By (2.74) -- (2.75) and Lemmas 2.7, 2.8 we can immediately obtain:
Fx(v, θx) = 0
(2.83)
38
Lemma 2.24. There exist positive numbers ε1 ∈ (0, ε) and a1 ∈ (0, 2a0), and a function Ω :
NO(ε1) → [0,∞) with the property that Ω(x, v) → 0 as kvkx → 0, such that for any (x, v) ∈
NO(ε1) the following conclusions hold with Bx = d2(cid:0)L ◦ expN Ox(ε)(cid:1):
(i) ((Bx(v)u, w))x − ((Bx(θx)u, w))x ≤ Ω(x, v)kukx · kwkx for any u ∈ N 0Ox ⊕ N −Ox and
w ∈ NOx;
x for all u ∈ N +Ox;
(ii) ((Bx(v)u, u))x ≥ a1kuk2
(iii) ((Bx(v)u, w)x) ≤ Ω(x, v)kukx · kwkx for all u+ ∈ N +Ox, w ∈ N −Ox ⊕ N 0Ox;
(iv) ((Bx(v)u, u)x ≤ −a0kuk2 for all u ∈ N −Ox.
Let us choose ε2 ∈ (0, ǫ/2) so small that
(x, v0 + hx(v0) + u+ + u−) ∈ NO(ε1)
for (x, v0) ∈ N 0O(2ε2) and (x, u∗) ∈ N ∗O(2ε2), ∗ = +,−. As in the proof of [40, Lemma 3.5],
we may use [40, Lemma 2.4] to derive
Lemma 2.25. Let the constants a1 and a0 be given by Lemma 2.24(ii),(iv). For the above ε2 > 0
and each x ∈ O the restriction of the functional Fx to N 0O(2ε2)x⊕ [N +O(2ε2)x⊕ N −O(2ε2)x]
satisfies:
(i) [D2Fx(v0, u+ + u−
2 )− D2Fx(v0, u+ + u−
1 )](u−
N 0O(2ε2), (x, u+) ∈ N +O(2ε2) and (x, u−
(ii) D2Fx(v0, u+ + u−)(u+ − u−) ≥ a1ku+k2
(x, u∗) ∈ N ∗O(2ε2), ∗ = +,−;
1 ) ≤ −a1ku−
2 − u−
2 − u−
1 k2
j ) ∈ N −O(2ε2), j = 1, 2;
x + a0ku−k2
x for any (x, v0) ∈ N 0O(2ε2) and
x for any (x, v0) ∈
(iii) D2Fx(v0, u+)u+ ≥ a1ku+k2
x for any (x, v0) ∈ N 0O(2ε2) and (x, u+) ∈ N +O(2ε2).
Denote by bundle projections Π0 : N 0O(ε2) → O and
Π± : N +O ⊕ N −O → O, Π∗ : N ∗O → O, ∗ = +,−.
Let Λ = N 0O(2ε2), p : E → Λ and p∗ : E ∗ → Λ be the pullbacks of N +O ⊕ N −O and N ∗O
via Π0, ∗ = +,−. Then E = E + ⊕ E −, and for λ = (x, v0) ∈ Λ we have Eλ = N +Ox ⊕ N −Ox
and E ∗
λ = N ∗Ox, ∗ = +,−. Moreover, for each η > 0 we write
Bη(E) =(cid:8)(λ, w) λ = (x, v0) ∈ Λ & w ∈ (N +O ⊕ N −O)x(η)(cid:9) ,
¯Bη(E) =n(λ, w) λ = (x, v0) ∈ Λ & w ∈ (N +O ⊕ N −O)x(η)o .
Similarly, Bη(E ∗) and ¯Bη(E ∗) (∗ = +,−) are defined. Let J : B2ε2(E) → R be given by
J (λ, v±) = Jλ(v±) = F(x, v0, v±),
∀λ = (x, v0) ∈ Λ & ∀v± ∈ B2ε2(E)λ.
(2.84)
It is continuous, and C 1 in v±. From (2.83) and Lemma 2.25 we directly obtain
39
Lemma 2.26. The functional Jλ satisfies the conditions in Theorem A.2 of [40] (the bundle
parameterized version of [23, Theoren 1.1]). Precisely, for each λ ∈ Λ the functional Jλ :
B2ε2(E) → R satisfies:
(i) Jλ(θλ) = 0 and DJλ(θλ) = 0;
(ii) [DJλ(u+ + u−
x for any λ = (x, v0) ∈ Λ,
2 ) − DJλ(u+ + u−
1 ) ≤ −a1ku−
2 − u−
1 k2
2 − u−
u+ ∈ ¯Bε2(E +)λ and u−
j ∈ ¯Bε2(E −)λ, j = 1, 2;
1 )](u−
(iii) DJλ(λ, u+ + u−)(u+ − u−) ≥ a1ku+k2
x + a0ku−k2
x for any λ = (x, v0) ∈ Λ and u∗ ∈
¯Bε2(E ∗)λ, ∗ = +,−;
(iv) DJλ(u+)u+ ≥ a1ku+k2
x for any λ = (x, v0) ∈ Λ and u+ ∈ ¯Bε2(E +)λ.
By this we can use Theorem A.2 of [40] to get ǫ ∈ (0, ε2), an open neighborhood U of the
zero section 0E of E in B2ε2(E) and a homeomorphism
φ : Bǫ(E +) ⊕ Bǫ(E −) → U, (λ, u+ + u−) 7→ (λ, φλ(u+ + u−))
such that for all (λ, u+ + u−) ∈ Bǫ(E +) ⊕ Bǫ(E −) with λ = (x, v0) ∈ Λ,
J(φ(λ, u+ + u−)) = ku+k2
x − ku−k2
x.
(2.85)
(2.86)
λ if and only if x = θλ, and φ is a
Moreover, for each λ ∈ Λ, φλ(θλ) = θλ, φλ(x + y) ∈ E −
homoeomorphism from Bǫ(E −) onto U ∩ E −.
Note that Bǫ(E +) ⊕ Bǫ(E −) = N 0O(2ε2) ⊕ N +O(ǫ) ⊕ N −O(ǫ) and U = N 0O(2ε2) ⊕ bU ,
wherebU is an open neighborhood of the zero section of N +O⊕ N −O in N +O(2ε2)⊕ N −O(ε2).
Let W = N 0O(ǫ)⊕bU , which is an open neighborhood of the zero section of NO in N 0O(2ε2)⊕
N +O(2ε2) ⊕ N −O(ε2). By (2.85) we get a homeomorphism
φ :
N 0O(ǫ) ⊕ N +O(ǫ) ⊕ N −O(ǫ) → W,
(x, v, u+ + u−) 7→ (x, v, φ(x,v)(u+ + u−)),
and therefore a topological embedding bundle morphism that preserves the zero section,
Φ :
N 0O(ǫ) ⊕ N +O(ǫ) ⊕ N −O(ǫ) → NO,
(x, v, u+ + u−) 7→ (x, v + hx(v), φ(x,v)(u+ + u−)).
From (2.82), (2.84) and (2.86) it follows that Φ and φ satisfy
(2.87)
(2.88)
L ◦ exp◦Φ(x, v + u+ + u−) = L ◦ expx(v + hx(v) + φ(x,v)(u+ + u−))
x + L ◦ expx(v + hx(v))
x − ku−k2
= ku+k2
for all (x, v + u+, u−) ∈ N 0O(ǫ) ⊕ N +O(ǫ) ⊕ N −O(ǫ). The other conclusions easily follow
from the above arguments. Theorem 2.23 is proved.
40
Proof of Theorem 2.22.
But we need to replace the map F in (2.82) by
In the present case Lemma 2.24 also holds with N 0Ox = {θx} ∀x ∈ O.
F(x, u+ + u−) : N +O(ǫ) ⊕ N −O(ǫ) → R, (x, u+ + u−) 7→ L ◦ expx(u+ + u−).
For any x ∈ Ox, let Fx be the restriction of F to N +O(ǫ)x ⊕ N −O(ǫ)x. As in the proof of
Theorem 2.1, Lemma 2.25 is still true with N 0O(2ε2)x = {θx}. Then the desired conclusions can
be obtained by applying [40, Theorem A.2] to Λ = O and Jλ = Fx with λ = x ∈ O.
✷
Many results in [13, 50, 67, 68] also hold in our setting. Here are a few of them, which are
needed in this paper.
Corollary 2.27 (Shifting Theorem). Let K be any commutative ring. Then
i) Under the assumptions of Theorem 2.22, it holds that
C∗(L,O; Z2) ∼= C∗−µO (O; Z2),
C∗(L,O; K) ∼= C∗−µO (O; θ− ⊗ K),
(2.89)
(2.90)
where θ− is the orientation bundle of N −O.
ii) Under the assumptions of Theorem 2.23, if O has trivial normal bundle then for any commuta-
tive group K and x ∈ O,
Cq(L,O; K) ∼= ⊕q
j=0Cq−j−µO (L◦
x, θx; K) ⊗ Hj(O; K) ∀q = 0, 1,··· .
(2.91)
(Consequently, every Cq(L,O; K) is isomorphic to finite direct sum r1K⊕···⊕rsK⊕Hj(O; K),
where each ri ∈ {0, 1}, see [43, Remark 4.6]. )
As in [6, 69], (2.89) -- (2.90) follow from (2.76).
Corollary 2.28. Under the assumptions of Corollary 2.27, we have:
(i) O is a local minimum (so µO = 0) if and only if
Cq(L,O; K) ∼= δq0K ∀q ∈ Z ⇐⇒ C0(L,O; K) 6= 0.
(ii) C1(L,O; K) 6= 0 and rankN 0O = 1 then µO = 0 and
Cq(L,O; K) ∼= K ⊗ Hq−1(O; K) ∀q ∈ Z.
(2.92)
(iii) If rankN 0O = 1 in the case µO = 0, then θ is of mountain pass type (in the sense that some
(and hence any) θx with x ∈ O is a critical point of L ◦ expx on NOx of mountain pass
type) if and only if (2.92) holds;
(iv) If Ck(L,O; K) 6= 0 for k = rankN −O then for any q ∈ Z
Cq(L,O; K) ∼=(cid:26) K ⊗ K if q ≥ k,
if q < k.
0
By the Peter-Weyl theorem, the compact Lie group has a faithful representation into the real
orthogonal group O(m) (i.e., a injective Lie group homomorphism into O(m)) for some integer
41
m > 0. Hence G can be viewed as a subgroup of O(m). Let E be the Hilbert manifold consisting
of all m-orthogonal frames in the Hilbert space l2. It is a contractible space on which G acts freely.
Let BG = E/G denote the classifying space, which is a Hilbert manifold. Then E → BG is a
universal smooth principal G-bundle. Let E ×G H be the quotient of E × H by the free diagonal
action g · (p, u) = (gp, g · u). This Hilbert manifold is a fiber space on BG with fiber H.
The G-invariant functional L lifts a natural one on E × H, (p, u) 7→ L(u), and hence induces
a functional LE on E ×G H with same smoothness as L. The critical orbit O of L corresponds
to a critical manifold E ×G O of LE, and they have the same Morse indexes. Moreover, for a
G-invariant Gromoll-Meyer pair (W, W −) of O (see [67] for its existence), (E ×G W, E ×G W −)
is a Gromoll-Meyer of E ×G O.
Let c = LO and U be a G-invariant neighborhood of O with K(L) ∩ Lc ∩ U = O, where
Lc = {x ∈ H L(x) ≤ c}. For any coefficient ring K and q ∈ N ∪ {0}, the qth G critical group
of O is defined by
C ∗
G(L,O; K) = H ∗
G(Lc ∩ U, (Lc \ O) ∩ U ; K)
= H ∗(E ×G (Lc ∩ U ), E ×G ((Lc \ O) ∩ U ); K).
It is equal to H ∗
G(W, W −; K), see [13, page 76]. Moreover, if O is nondegenerate, it follows from
(2.76) and the Thom isomorphism theorem (with twisted coefficients) that
G(L,O; K) ∼= H µO−1
C ∗
G
(O; θ− ⊗ K),
(2.93)
where µO is the Morse index of O and θ− is the orientation bundle of N −O, see Theorem 7.5 on
the page 75 of [13].
More generally, the corresponding versions of Theorems 2.14,2.18 and 2.19 can also be proved.
We only write the following since it is needed in the proof of Theorem 3.20 later.
Theorem 2.29 (Parameterized Splitting Theorem around Critical Orbits). Under the assumptions
of Theorem 2.23, suppose further that G-invariant functionals Gj ∈ C 1(H, R), j = 1,··· , n,
have value zero and vanishing derivative at each point of O, and also satisfy:
(i) gradients ∇Gj have Gateaux derivatives G′′
(ii) G′′
j (u) are continuous at each point u ∈ O.
If the critical orbit O is degenerate, i.e., rankN 0O > 0, then for sufficiently small ǫ > 0, δ > 0,
j (u) at each point u near O,
(I) there exists a unique continuous map
h : [−δ, δ]n × N 0O(3ǫ) → N +O ⊕ N −O, (~λ, x, v) 7→ hx(~λ, v),
(2.94)
42
such that h(~λ,·) : N 0O(3ǫ) → N +O ⊕ N −O, (x, v) 7→ hx(~λ, v) is a G-equivariant topological
bundle morphism that preserves the zero section and satisfies
(P +
x + P −
x ) ◦ Πx∇(L~λ ◦ expx)(v + hx(~λ, v)) = 0 ∀(x, v0) ∈ N 0O(ǫ),
j=1 Gj ;
where L~λ = L +Pn
(II) there exists a continuous map
Φ : [−δ, δ]n × N 0O(ǫ) ⊕ N +O(ǫ) ⊕ N −O(ǫ) → NO
(2.95)
(2.96)
such that Φ(~λ,·) : N 0O(ǫ) ⊕ N +O(ǫ) ⊕ N −O(ǫ) → NO is a G-equivariant homeomorphism
onto an open neighborhood of the zero section preserving fibers, and that
L~λ ◦ exp◦Φ(~λ, x, v, u+ + u−) = ku+k2
x + L~λ ◦ expx(v + hx(~λ, v))
for any ~λ ∈ [−δ, δ]n, x ∈ O and (v, u+, u−) ∈ N 0O(ǫ)x × N +O(ǫ)x × N −O(ǫ)x;
(III) for each (~λ, x) ∈ [−δ, δ]n × O the functional
x − ku−k2
N 0O(ǫ)x → R, v 7→ L◦
is Gx-invariant, of class C 1, and satisfies
~λ,x
(v) := L~λ ◦ expx(v + hx(~λ, v))
DL◦
~λ,x
(v)v′ := (∇(L~λ ◦ expx)(v + hx(~λ, v)), v′),
∀v′ ∈ N 0Ox.
(2.97)
(2.98)
Moreover, if the critical orbit O is nondegenerate, i.e., rankN 0O = 0, then h does not appear, Φ
is from [−δ, δ]n × N +O(ǫ) ⊕ N −O(ǫ) to NO, and (2.97) becomes
x − ku−k2
L~λ ◦ exp◦Φ(~λ, x, u+ + u−) = ku+k2
x + L~λO
(2.99)
for any ~λ ∈ [−δ, δ]n, x ∈ O and (u+, u−) ∈ N +O(ǫ)x × N −O(ǫ)x.
2.7 Proof of Theorem 2.6
Without loss of generality we may assume θ ∈ V and u0 = θ. By the assumption we have a C 2
reduction functional L◦ : BH(θ, δ) ∩ H 0 → R such that θ is the unique critical point of it and
L◦(z) = o(kzk2). Clearly, we can shrink δ > 0 so that δ < min{r, 1}, ¯BH(θ, δ) ∩ H 0 ⊂ V and
ω in Lemma 2.8 satisfies
ω(z + ϕ(z)) <
1
2
min{a0, a1},
∀z ∈ BH(θ, δ) ∩ H 0.
(2.1)
By the uniqueness of solutions we can also require that if v ∈ BH(θ, δ) satisfies (I−P 0)∇L(v) =
0 then v = z + ϕ(z) for some z ∈ BH(θ, δ) ∩ H 0.
Take a smooth function ρ : [0,∞) → R satisfying: 0 ≤ ρ ≤ 1, ρ(t) = 1 for t ≤ δ/2, ρ(t) = 0
for t ≥ δ, and ρ′(t) < 4/δ. For b ∈ H 0 we set L◦
b(z) = L◦(z) + ρ(kzk)(b, z)H . Then
DL◦
b(z)ξ = DL(z + ϕ(z))(ξ + ϕ′(z)ξ) + ρ(kzk)(b, ξ)H
∀ξ ∈ H 0.
+ρ′(kzk)(b, z)H (z/kzk, ξ)H ,
(2.2)
Note that kDL◦(z)k ≥ ν for some ν > 0 and for all z ∈ ¯BH 0(θ, δ) \ BH 0(θ, δ/2). Suppose
kbk < ν/5. Then
43
kDL◦
b(z)k = (cid:13)(cid:13)DL◦(z) + ρ(kzk)b + (b, z)H ρ′(kzk)z/kzk(cid:13)(cid:13) ≥ ν − 5kbk > 0,
b has no critical point in ¯BH 0(θ, δ) \ BH 0(θ, δ/2). By Sard's theorem we may
and therefore L◦
take arbitrary small b such that the critical points of L◦
b , if any, are nondegenerate. Choose a C 2
function β : H → R such that β(u) = 0 for u ∈ H \ BH(θ, r), and β(u) = 1 for u ∈ BH(θ, δ).
Clearly, we can require kβ(i)(u)k ≤ M for some M > 0, i = 0, 1, 2 and for all u ∈ H. Define
(2.3)
Lb(u) = L(u) + β(u)ρ(kP 0uk)(b, P 0u)H
(2.4)
Clearly, it satisfies (iii). If b is taken to be very small, then (ii) is satisfied. Moreover, since L
satisfies the (PS) condition, kDL(u)k ≥ c for some c > 0 and for all u ∈ BH (θ, r) \ BH(θ, δ).
Hence all critical points of Lb belong to BH (θ, δ) as long as b is small enough.
Let v ∈ BH(θ, δ) be a critical point of Lb. Then
0 = L′
b(v)ξ = (∇L(v), ξ)H + ρ(kP 0vk)(b, P 0ξ)H
+ρ′(kP 0vk)(b, P 0v)H (P 0v, P 0ξ)H /kP 0vk,
∀ξ ∈ H.
(2.5)
Since ρ(kP 0vk) = 1 for kP 0vk ≤ kvk < δ, this implies (∇L(v), ξ)H = 0 for any ξ ∈ H +⊕ H −,
i.e., ((I − P 0)∇L(v), ξ)H = 0 for any ξ ∈ H. It follows that v = z + ϕ(z) for some z ∈
BH(θ, δ) ∩ H 0. [This z is nonzero. Otherwise, v = θ. But θ is not a critical point of Lb if b 6= θ].
Note that (∇L(z + ϕ(z)), ϕ′(z)ξ)H = 0 ∀ξ ∈ H 0 because ϕ′(z)ξ ∈ H + ⊕ H −. (2.5) leads to
0 = (∇L(z + ϕ(z)), ξ)H + ρ(kzk)(b, ξ)H + ρ′(kzk)(b, z)H (z, ξ)H /kzk
= (∇L(z + ϕ(z)), ξ)H + (∇L(z + ϕ(z)), ϕ′(z)ξ)H
+ρ(kzk)(b, ξ)H + ρ′(kzk)(b, z)H (z, ξ)H /kzk ∀ξ ∈ H 0,
b(z) = 0 by (2.2). That is, z is a critical point of L◦
and therefore DL◦
(2.3). It follows from (2.4) that
b , and so z ∈ BH 0(θ, δ/2) by
L′′
b (v)(ξ, η) = (L′′(v)ξ, η)H ,
∀ξ, η ∈ H.
(2.6)
Let us prove that v is a nondegenerate critical point of Lb. Suppose ξ ∈ Ker( L′′
b (v)ξ, η)H = 0 for any η ∈ H. By (2.6), we have
( L′′
b (v)). Then
L′′(v)(ξ, η) = (L′′(z + ϕ(z))ξ, η)H = 0,
∀η ∈ H.
(2.7)
Decompose ξ into ξ0 + ξ⊥, where ξ0 ∈ H 0 and ξ⊥ ∈ H + ⊕ H −. A direct computation yields
(L′′(z + ϕ(z))ξ0, η + ϕ′(z)η)H + (L′′(z + ϕ(z))ξ⊥, η + ϕ′(z)η)H = 0,
∀η ∈ H 0.
(2.8)
44
Note that (I − P 0)∇L(w + ϕ(w)) = 0 for any w ∈ BH 0(θ, δ). Hence (∇L(w + ϕ(w)), ζ)H = 0
for any w ∈ BH 0(θ, δ) and ζ ∈ H + ⊕ H −. Differentiating this equality with respect to w yields
(L′′(w + ϕ(w))(τ + ϕ′(w)τ ), ζ)H = 0,
∀τ ∈ H 0, ∀w ∈ BH 0(θ, δ), ∀ζ ∈ H + ⊕ H −.
In particular, we have (L′′(z + ϕ(z))ξ⊥, η + ϕ′(z)η)H = 0 for all η ∈ H 0. This and (2.8) yield
d2L◦(z)(ξ0, η) = (L′′(z + ϕ(z))ξ0, η + ϕ′(z)η)H = 0,
∀η ∈ H 0.
(2.9)
b(z′) = d2L◦(z′) for any z′ ∈ BH 0(θ, δ/2) by the construction of L◦
b . We obtain
b(z)(ξ, η) = 0 for all η ∈ H 0. Hence ξ0 = θ since z is a nondegenerate critical point of
Moreover, d2L◦
that d2L◦
b by the choice of b, and thus ξ = ξ⊥. By (2.6) -- (2.7) and ( L′′
L◦
b (v)ξ, η)H = 0 ∀η ∈ H, we get
(L′′(z + ϕ(z))ξ⊥, η)H = (L′′(z + ϕ(z))ξ, η)H = 0,
∀η ∈ H.
(2.10)
Hence L′′(z + ϕ(z))ξ⊥ = 0. Decompose ξ⊥ into ξ+ + ξ−, where ξ+ ∈ H + and ξ− ∈ H −. Then
L′′(z + ϕ(z))ξ+ = −L′′(z + ϕ(z))ξ−. By Lemma 2.8 and (2.1) we derive
a1kξ+k2 ≤ (L′′(z + ϕ(z))ξ+, ξ+)H = −(L′′(z + ϕ(z))ξ−, ξ+)H ≤
−a0kξ−k2 ≥ (L′′(z + ϕ(z))ξ−, ξ−)H = −(L′′(z + ϕ(z))ξ+, ξ−)H ≥ −
a0
2 kξ−k · kξ+k.
These imply kξ+k ≤ kξ−k/2 and kξ−k ≤ kξ+k/2. Hence ξ+ = ξ− = θ and so ξ⊥ = θ. This
shows that v is a nondegenerate critical point of Lb. Lemma 2.8 and (2.6) give rise to
a1
2 kξ+k · kξ−k,
L′′
b (v)(ξ, ξ) = (L′′(v)ξ, ξ)H ≥ a1kξk2,
L′′
b (v)(ξ, ξ) = (L′′(v)ξ, ξ)H ≤ −a0kξk2,
∀ξ ∈ H +,
∀ξ ∈ H −.
But H = H + ⊕ H 0 ⊕ H −, dim H − = m− and dim H 0 = n0. These show that the Morse index
of L′′
b (v) must sit in [m−, m− + n0]. (iv) is proved.
We also need to show that Lb satisfies Hypothesis 2.5 on V if b is small enough. By (2.4) we
have for all ξ, η ∈ H,
L′
b(u)ξ = L′(u)ξ + (β′(u)ξ)ρ(kP 0uk)(b, P 0u)H + β(u)ρ(kP 0uk)(b, P 0ξ)H
+β(u)ρ′(kP 0uk)(b, P 0u)H(P 0u, P 0ξ)H /kP 0uk
(2.11)
and
b (u)η, ξ)H = (L′′(u)η, ξ)H + (β′′(u)η, ξ)H ρ(kP 0uk)(b, P 0u)H
( L′′
+(β′(u)ξ)ρ(kP 0uk)(b, P 0η)H + (β′(u)ξ)(b, P 0u)H ρ′(kP 0uk)(P 0u, P 0η)H /kP 0uk
+(β′(u)η)ρ(kP 0uk)(b, P 0ξ)H + β(u)(b, P 0ξ)H ρ′(kP 0uk)(P 0u, P 0η)H /kP 0uk
+(β′(u)η)ρ′(kP 0uk)(b, P 0u)H (P 0u, P 0ξ)H /kP 0uk
+β(u)(cid:0)ρ′′(kP 0uk)(P 0u, P 0η)H /kP 0uk(cid:1)(b, P 0u)H (P 0u, P 0ξ)H/kP 0uk
+β(u)ρ′(kP 0uk)(b, P 0η)H (P 0u, P 0ξ)H /kP 0uk
+β(u)ρ′(kP 0uk)(b, P 0u)H(cid:2)(P 0η, P 0ξ)H /kP 0uk − (P 0u, P 0ξ)H (P 0u, P 0η)H /kP 0uk3(cid:3)
= (L′′(u)η, ξ)H + Υ(u, b, ξ, η).
By the constructions of β and ρ we have a constant M2 > 0 such that
Υ(u, b, ξ, η) ≤ M2kbk · kξk · kηk
45
for all u ∈ V and ξ, η ∈ H. Since we may require that the support of β can be contained a
neighborhood of θ on which (iii) of Hypothesis 2.5 holds, for sufficiently small b the positive
definite part of L′′
b , P given by ( P (u)ξ, η)H = (P (u)ξ, η)H + Υ(u, b, ξ, η), is also uniformly
positive definite on this neighborhood. Hence Lb satisfies Hypothesis 2.5.
By (2.4) and (2.11) we have positive numbers Mi, i = 0, 1, such that Lb(u)−L(u) ≤ M0kbk
b(u)ξ − L′(u)ξk ≤ M1kbk · kξk for all u ∈ V and ξ ∈ H. So (ii) and (iii) can be satisfied
and k L′
if b is small.
Finally, let us prove that Lb satisfies the (PS) condition for small b. By (ii) and (iii) in Hypoth-
esis 2.5, there exists ǫ ∈ (0, δ/2) such that for all u ∈ BH(θ, ǫ) and ξ ∈ H,
(P (u)ξ, ξ)H ≥ C0kξk2
and
kQ(u) − Q(θ)k < C0/2.
(2.12)
Since L satisfies the (PS) condition and θ is a unique critical point of L in V , we have ν0 > 0 such
that kL′(u)k ≥ ν0 for all u ∈ V \ BH(θ, ǫ). Let us choose b so small that k L′
b(u)k ≥ ν0/2 for
b(un) → 0 and supn Lb(un) < ∞, then
all u ∈ V \ BH(θ, ǫ). Then if {un}n ⊂ V satisfies L′
{un}n muse be contained in BH(θ, ǫ). It follows that ∇ Lb(un) = ∇ L(un) + b for all n. For any
two natural numbers n and m, using the mean value theorem we have τ ∈ (0, 1) such that
(∇L(un) − ∇L(um), un − um)H = (B(τ un + (1 − τ )um)(un − um), un − um)H
= (P (τ un + (1 − τ )um)(un − um), un − um)H + (Q(θ)(un − um), un − um)H
+([Q(τ un + (1 − τ )um) − Q(θ)](un − um), un − um)H
≥ C0kun − umk2 −
C0
2 kun − umk2 + (Q(θ)(un − um), un − um)H
by (2.12). Passing to a subsequence we may assume un ⇀ u0. Since Q(θ) is compact, Q(θ)un →
Q(θ)u0 and so (Q(θ)(un − um), un − um)H → 0 as n, m → ∞. Note that ∇L(un)−∇L(um) =
(∇L(un) + b) − (∇L(um) + b) → 0 as n, m → ∞. It follows from the above inequalities that
kun − umk → 0 as n, m → ∞. This implies un → u0. Theorem 2.6 is proved.
3 Bifurcations for potential operators
In this section, some previous bifurcation theorems, such as those by Rabinowitz [57], by Fadelll
and Rabinowitz [27, 28], and by Chang and Wang [15, 67, 68], by Chow and Lauterbach [18] and
by Bartsch and Clapp [1, 2], were generalized so that they can be used to study variational bifurca-
tion for the functional F in (1.8). Our methods are mainly based our Morse lemma, Theorem 2.1,
and the parameterized splitting and shifting theorems, Theorems 2.18, 2.19. The latter suggest
that multiparameter bifurcations can be studied similarly; we here give two, Theorems 3.3, 3.4.
46
3.1 Generalizations of a bifurcation theorem by Chow and Lauterbach
Let H be a real Hilbert space, I an open interval containing 0 in R, and {Bλ}λ∈I a family of
bounded linear self-adjoint operators on H such that kBλ − B0k → 0 as λ → 0. Suppose that 0
is an isolated point of the spectrum σ(B0) with n = dim Ker(B0) ∈ (0,∞), and that Ker(Bλ) =
{0} ∀ ± λ ∈ (0, ε0) for some positive number ε0 ≪ 1. By the arguments on the pages 107 and
203 in [33], for each λ ∈ (−ε0, ε0) \ {0}, Bλ has n eigenvalues near zero, and none of them is
zero. In Kato's terminology in [33, page 107], we have the so-called 0-group eig0(Bλ) consisting
of eigenvalues of Bλ which approach 0 as λ → 0. Let r(Bλ) be the number of elements in
eig0(Bλ) ∩ R− and
r+
Bλ
= lim
λ→0+
r(Bλ),
r−
Bλ
= lim
λ→0−
r(Bλ).
(3.1)
Theorem 3.1. Let U be an open neighborhood of the origin of a real Hilbert space H, and I
an open interval containing 0 in R, F ∈ C 0(I × V, R) such that L := Fλ = F(λ,·) satisfy
Hypothesis 1.1 with X = H for each λ ∈ I. Suppose that one of the following two conditions is
satisfied.
(1) For some small δ > 0, λ 7→ Fλ is continuous at λ = 0 in C 1( ¯BH (θ, δ)) topology.
(2) For some small δ > 0, λ 7→ Fλ is continuous at λ = 0 in C 0( ¯BH (θ, δ)) topology; and
(un) → θ and
(u0) =
for every sequences λn → λ0 in I and {un}n≥1 ⊂ ¯BH(θ, δ) with F ′
{Fλn (un)}n≥1 bounded, there exists a subsequence unk → u0 ∈ ¯BH(θ, δ) with F ′
0.
λn
λ0
Then
(I) If (θ, 0) is not a bifurcation point of the equation
F ′
u(λ, u) = 0,
(λ, u) ∈ I × V,
(3.2)
(i.e., (0, θ) is not in the closure of {(λ, u) ∈ I × V F ′
u(λ, u) = 0, u 6= θ}), then critical
groups C∗(Fλ, θ; K) are well-defined and have no changes as λ varies in a small neighbor-
hood of 0.
(II) (0, θ) is a bifurcation point of the equation (3.2) if the following conditions are also satisfied:
(a) Ker(d2Fλ(θ)) = {θ} for small λ 6= 0;
(b) d2Fλ(θ) → d2F0(θ) as λ → 0;
(c) 0 ∈ σ(d2F0(θ)) (and so is an isolated point of the spectrum σ(d2F0(θ)) and an eigen-
value of d2F0(θ) of the finite multiplicity by [5, Lemma 2.2]);
d2Fλ(θ) 6= r−
d2Fλ(θ).
(d) r+
47
Proof. Step 1. This is a direct consequence of the stability of critical groups. In fact, since
(θ, 0) is not a bifurcation point of the equation (3.2), we may find 0 < ε0 ≪ 1 and a small
bounded neighborhood W of θ ∈ H with W ⊂ BH(θ, δ) such that for each λ ∈ (−ε0, ε0) the
functional Fλ has a unique critical point θ sitting in W . Note that Fλ is of class (S)+. We can
assume that it satisfies the (PS) condition in W by shrinking W (if necessary). After shrinking
ε0 > 0 (if necessary), we may use the stability of critical groups (cf. [16, Theorem III.4] and [22,
Theorem 5.1]) to derive
C∗(Fλ, θ; K) = C∗(F0, θ; K),
∀λ ∈ (−ε0, ε0)
(3.3)
provided that (1) holds. If (2) is satisfied the same claim is obtained by [19, Theorem 3.6].
Step 2. By a contradiction, suppose that (θ, 0) is not a bifurcation point of the equation (3.2).
Then we have (3.3) from (I). By (a), θ is a nondegenerate critical point of Fλ. It follows from
(3.3) and Theorem 2.1 that all Fλ, 0 < λ < ε0, have the same Morse index µλ at θ ∈ H, i.e.,
(−ε0, ε0) \ {0} ∋ λ 7→ µλ is constant.
By [40, Proposition B.2], each ∈ σ(d2F0(θ)) ∩ {t ∈ R− t ≤ 0} is an isolated point in
σ(d2F0(θ)), which is also an eigenvalue of finite multiplicity. (This can also be derived from [5,
Lemma 2.2]). Since 0 ∈ σ(d2F0(θ)) by (c), we may assume σ(d2F0(θ)) ∩ {t ∈ R− t ≤ 0} =
{0, 1,··· , k}, where µi has multiplicity si for each i = 1,··· , k. As above, by this, (b) and the
arguments on the pages 107 and 203 in [33], if 0 < λ is small enough, d2Fλ(θ) has exactly si
(possible same) eigenvalues near µi, but total dimension of corresponding eigensubspaces is equal
to that of eigensubspace of i. Hence if λ ∈ (0, ε0) (resp. −λ ∈ (0, ε0)) is small enough we obtain
µλ = µ0 + r+
d2Fλ(θ)
(resp. µ−λ = µ0 + r−
d2Fλ(θ)).
These and (d) imply µλ − µ−λ = r+
the above claim that (−ε0, ε0) \ {0} ∋ λ 7→ µλ is constant.
d2Fλ(θ) − r−
d2Fλ(θ) 6= 0 for small λ ∈ (0, ε0), which contradicts
Part (II) in Theorem 3.1 is a partial generalization of a bifurcation theorem due to [18]. The
latter requires: 1) F ∈ C 2(I × V, R) (so (b) holds naturally), 2) 0 < dim Ker(d2F0(θ)) < ∞, 3)
0 is isolated in σ(d2F0(θ)), 4) (d) is satisfied. Its proof is based on center manifold theory, which
is different from ours.
3.2 Generalizations of Rabinowitz bifurcation theorem
Since the birth of the Rabinowitz bifurcation theorem [57] some generalizations and new proofs
are given, see [15, 17], [22], [32] and [67, 68], etc.
Our generalization will reduce to the following result, which may be obtained as a corollary of
[32, Theorem 2].
Theorem 3.2 ([11, Theorem 5.1]). Let X be a finite dimensional normed space, let δ > 0, λ∗ ∈ R
and for every λ ∈ [λ∗ − δ, λ∗ + δ], let φλ : B(θ, δ) → R be a function of class C 1. Assume that
48
a) the functions {(λ, u) → φλ(u)} and {(λ, u) → φ′
λ(u)} are continuous on [λ∗ − δ, λ∗ + δ] ×
B(θ, δ);
b) u = θ is a critical point of φλ∗ ; φλ has an isolated local minimum (maximum) at zero for every
λ ∈ (λ∗, λ∗+δ] and an isolated local maximum (minimum) at zero for every λ ∈ [λ∗−δ, λ∗).
Then one at least of the following assertions holds:
i) u = θ is not an isolated critical point of φλ∗ ;
ii) for every λ 6= λ∗ in a neighborhood of λ∗ there is a nontrivial critical point of φλ converging
to zero as λ → λ∗;
iii) there is a one-sided (right or left) neighborhood of λ∗ such that for every λ 6= λ∗ in the
neighborhood there are two distinct nontrivial critical points of φλ converging to zero as
λ → λ∗.
It was generalized to infinite dimension spaces in [22, Theorem 4.2].
The following is a generalization of the necessity part of Theorem 12 in [58, Chapter 4, §4.3]
(including the classical Krasnoselsi potential bifurcation theorem [37]). The sufficiency part of
Theorem 12 in [58, Chapter 4, §4.3] is contained in the case that the condition (a) in Theorem 3.5
holds.
Theorem 3.3. Let U be an open neighborhood of the origin of a real Hilbert space H. Suppose
(i) F ∈ C 1(U, R) satisfies Hypothesis 1.1 with X = H as the functional L there;
(ii) Gj ∈ C 1(U, R) satisfies G′
j(θ) = θ, j = 1,··· , n, and each gradient G′
j has the Gateaux
j (u) at any u ∈ U , which is compact linear operators and satisfy (D3) in
derivative G′′
Hypothesis 1.1 with X = H.
If (~λ∗, θ) ∈ Rn × U is a (multiparameter) bifurcation point for the equation
F ′(u) =
nXj=1
λjG′
j(u),
u ∈ U,
then ~λ∗ = (λ∗
1,··· , λ∗
n) is an eigenvalue of
(3.4)
(3.5)
F ′′(θ)v −
λjG′′
j (θ)v = 0,
v ∈ H,
nXj=1
in other words, θ is a degenerate critical point of the functional F −Pn
jG. (The solution
space of (3.5), denoted by H(~λ), is of finite dimension because it is the kernel of a Fredholm
operator).
j=1 λ∗
49
Theorem 12 in [58, Chapter 4, §4.3] also required: (a) G is weakly continuous and uniformly
differentiable in U , (b) F ′ has uniformly positive definite Frech`et derivatives and satisfies the
condition α) in [58, Chapter 3, §2.2]. If G′ is completely continuous (i.e., mapping a weakly
convergent sequence into a convergent one in norm) and has Frech´et derivative G′′(u) at u ∈ U ,
then G′′(u) ∈ L (H) is a compact linear operator (cf. [4, Remark 2.4.6]).
Let (~λ∗, θ) ∈ Rn × U be a bifurcation point for (3.4). Then we have a
sequence (~λk, uk) ∈ Rn × (U \ {θ}) such that ~λk = (λk,1,··· , λk,n) → ~λ∗, uk → θ and
Proof of Theorem 3.3.
F ′(uk) =
nXj=1
λk,jG′
j(uk),
k = 1, 2,··· .
Passing to a subsequence, if necessary, we can assume vk = uk/kukk ⇀ v∗. By the assumptions,
B = F ′′ has a decomposition P + Q as in Hypothesis 1.1 with X = H. (D4) and Lemma 2.9
imply that there exist positive constants η0 > 0 and C ′
0 > 0 such that
(P (u)h, h) ≥ C ′
0khk2 ∀h ∈ H, ∀u ∈ BH(θ, η0) ⊂ U.
Clearly, we can assume that {uk}k≥1 is contained in BH (θ, η0). Note that
1
kukk2 (F ′(uk), uk) =
nXj=1
λk,j
kukk2 (G′
j(uk), uk),
k = 1, 2,··· .
(3.6)
(3.7)
Since G′
(0, 1) such that
j(θ) = θ, j = 1,··· , n, using the Mean Value Theorem we have a sequence {tk}k≥1 ⊂
nXj=1
λk,j
kukk2 (G′
j(uk), uk) =
+
nXj=1
nXj=1
λk,j(G′′
j (tkuk)vk, vk) =
λk,j([G′′
j (tkuk) − G′′
j (θ)]vk, vk)
λk,j(G′′
j (θ)vk, vk) →
j (θ)v∗, v∗)
(3.8)
nXj=1
nXj=1
λ∗
j (G′′
because all G′′
we may use the Mean Value Theorem to yield a sequence {sk}k≥1 ⊂ (0, 1) such that
j (θ) are compact and kG′′
j (θ)k → 0 by (D3). Moreover, since F ′(θ) = θ,
j (tkuk)−G′′
1
kukk2 (F ′(uk), uk) =
=
1
kukk2 (F ′′(skuk)uk, uk)
kukk2 (P (skuk)uk, uk) +
1
1
1
kukk2 (Q(skuk)uk, uk)
≥ C ′
0 +
by (3.6). As in (3.8) we have also
kukk2 (Q(skuk)uk, uk) ∀k ∈ N
It follows from these and (3.7) that C ′
j=1 λ∗
jG′′
j (θ) − Q(θ)]v∗, v∗) and hence v∗ 6= θ.
1
kukk2 (Q(skuk)uk, uk) → (Q(θ)v∗, v∗).
0 ≤ ([Pn
to
nXj=1
and hence F ′′(θ)v∗ −Pn
j=1 λ∗
jG′′
λ∗
j (G′′
j (θ)v∗, h) = (v∗,F ′′(θ)h) ∀h ∈ H
j (θ)v∗ = 0. That is, ~λ∗ is an eigenvalue of (3.15).
✷
50
Moreover, for any h ∈ H we have
(F ′(uk), h) =
1
kukk
nXk=1
λk
kukk
and as in (3.8) we may prove that
(G′
j(uk), h),
k = 1, 2,··· ,
nXj=1
λk
kukk
(G′
j(uk), h) →
nXj=1
λ∗
j (G′′
j (θ)v∗, h),
and that for some sequence {τk}k≥1 ⊂ (0, 1), depending on {uk}k≥1 and h,
1
kukk
(F ′(uk), h) = (F ′′(τkuk)vk, h) = (vk,F ′′(τkuk)h) → (v∗,F ′′(θ)h)
(3.9)
(3.10)
because vk ⇀ v∗ and kF ′′(τkuk)h − F ′′(θ)hk → 0 by (D2) and (D3). This and (3.9) -- (3.10) lead
In the following we give two generalizations of Rabinowitz bifurcation theorem [57].
Theorem 3.4. Under the assumptions (i) -- (ii) of Theorem 3.3, suppose:
(iii) ~λ∗ is an isolated eigenvalue of (3.5);
(iv) the corresponding finite dimension reduction L◦
~λ
in Theorem 2.18 is of class C 2 for each ~λ near the origin of Rn;
with the functional L = F −Pn
j=1 λ∗
jGj as
(v) (2.62) holds with H 0 = H(~λ∗) (the solution space of (3.5) with ~λ = ~λ∗);
(vi) either dim H(~λ∗) is odd or there exists ~λ ∈ Rn \ {~0} such that the symmetric bilinear form
H(~λ∗) × H(~λ∗) ∋ (z1, z2) 7→ Q~λ(z1, z2) =
nXj=1
λj(G′′
j (θ)z1, z2)H
(3.11)
has different Morse index and coindex. (In particular, if n = 1 the latter is equivalent to the
fact that the form H(λ∗) × H(λ∗) ∋ (z1, z2) 7→ (G′′
1 (θ)z1, z2)H has different Morse index
and co-index).
Then (~λ∗, θ) ∈ Rn × U is a bifurcation point for the equation (3.4). Furthermore, if for some
~µ ∈ Rn \ {~0} the form Q~µ defined by (3.11) is either positive definite or negative one, then one of
the following alternatives occurs:
(A) (~λ∗, θ) is not an isolated solution of (3.4) in {~λ∗} × U .
(B) there exists a sequence {tk}k≥1 ⊂ R\{0} such that tk → 0 and that for each tk the equation
(3.4) with ~λ = tk~µ + ~λ∗ has infinitely many solutions converging to θ ∈ H.
51
(C) for every t in a small neighborhood of 0 ∈ R there is a nontrivial solution ut of (3.4) with
~λ = t~µ + ~λ∗ converging to θ as t → 0;
(D) there is a one-sided T neighborhood of 0 ∈ R such that for any t ∈ T \ {0}, (3.4) with
~λ = t~µ + ~λ∗ has at least two nontrivial solutions converging to zero as t → 0.
j=1 λ∗
Proof. Step 1. By the assumptions we have the conclusions of Theorem 2.18 with L = F −
Pn
jGj . Suppose that (~λ∗, θ) ∈ Rn × U is not a bifurcation point for the equation (3.4). Then
as in the proof of Theorem 3.1(I) we may find 0 < η ≪ 1 with ¯BH(θ, η) ⊂ U such that after
shrinking δ > 0 in Theorem 2.18 for each ~λ ∈ [−δ, δ]n the following claims hold:
• the functional L~λ has a unique critical point θ in ¯BH (θ, η),
• ~λ∗ = (λ∗
1,··· , λ∗
• for all ~λ ∈ [−δ, δ]n,
n) is a unique eigenvalue of (3.5) in [−δ, δ]n + ~λ∗,
nXj=1
C∗(L~λ, θ; K) = C∗(L~0, θ; K) = C∗(F −
λ∗
jGj, θ; K).
(3.12)
We may also shrink ǫ > 0, r > 0, s > 0 and W in Theorem 2.18 so that
¯BH 0(θ, ǫ) ⊕ ¯BH +(θ, r) ⊕ ¯BH −(θ, s) ⊂ BH (θ, η)
and W ⊂ BH(θ, η),
where H 0 = H(~λ∗). By (3.12) and Theorem 2.19 we have
C∗(L◦
~λ
, θ; K) = C∗(L◦
~0, θ; K),
∀~λ ∈ [−δ, δ]n.
(3.13)
(This can also be derived from the stability of critical groups as before.) For each ~λ ∈ [−δ, δ]n \
{~0}, since θ ∈ H is a nondegenerate critical point of L~λ, Claim below (2.62) tells us that θ ∈ H 0
is a nondegenerate critical point of L◦
at θ
is constant with respect to ~λ ∈ [−δ, δ]n \ {~0}.
too. Hence (3.13) implies that the Morse index of L◦
On the other hand, by (vi), if dim H(~λ∗) is odd, for every ~λ ∈ [−δ, δ]n \ {~0} the nondegener-
(θ) on H(~λ∗) must have different Morse indexes, where
~λ
~λ
ate quadratic forms d2L◦
d2L◦
(θ) : H(~λ∗) × H(~λ∗) → R given by
(θ) and d2L◦
−~λ
~λ
~λ
d2L◦
~λ
(θ)(z1, z2) = (L′′
~λ
(θ)z1, z2)H = −
nXj=1
λj(G′′
j (θ)z1, z2)H = −Q~λ(z1, z2).
−t~λ
This contradicts (3.13). Similarly, if there exists ~λ ∈ Rn \ {~0} such that the form in (3.11) has
different Morse index and coindex, then for every t > 0 with t~λ ∈ [−δ, δ]n the forms d2L◦
(θ)
t~λ
(θ) on H(~λ∗) have different Morse indexes, and hence a contradiction is obtained.
and d2L◦
Step 2. By replacing ~µ by −~µ we may assume that the form Q~µ is positive definite. Suppose that
any one of (A) -- (C) does not hold. Then there exists ǫ ∈ (0, 1) such that θ ∈ H(~λ∗) is an isolated
critical point of L◦
t~µ(θ) is negative (resp. positive)
definite for each t in (0, ǫ] (resp. [−ǫ, 0)). Then Theorem 3.2 implies that for some one-sided T
t~µ for each t ∈ [−ǫ, ǫ]. By the assumption d2L◦
52
neighborhood of 0 ∈ [−ǫ, ǫ] and any t ∈ T \ {0} the functional L◦
t~µ has two distinct nontrivial
critical points zt,1 and zt,2 converging to θ ∈ H(~λ∗). Then ut,j = zt,j + ψ(t~µ + ~λ∗, zt,j), j = 1, 2,
are two nontrivial solutions of (3.4) with ~λ = t~µ + ~λ∗, and both converge to zero as t → 0.
Note that (3.5) has no isolated eigenvalues if ∩n
j (θ)) ∩ Ker(F ′′(θ)) 6= {θ}. It is
natural to ask when ~λ∗ is an isolated eigenvalue of (3.5). For the sake of simplicity let us consider
the case n = 1. Then H(~λ∗) = Ker(F ′′(θ)) − λ∗
j=1Ker(G′′
d2L◦
~λ
(θ)(z1, z2) = −λ1(G′′
j (θ)z1, z2)H = −
λ1
λ∗
1
1 (θ)), and if λ∗
1G′′
1 6= 0 we have
(F ′′(θ)z1, z2).
In Theorem 3.4, the final condition that the form in (3.11) is either positive definite or negative
one suggests that we should require F ′′(θ) to be nondegenerate on H(~λ∗).
Suppose now that F ′′(θ) is invertible, n = 1 and write G = G1. Then (3.14) and (3.15) become
F ′(u) = λG′(u),
u ∈ U,
F ′′(θ)v − λG′′(θ)v = 0,
v ∈ H,
(3.14)
(3.15)
respectively. Moreover, 0 is not an eigenvalue of (3.15), and λ ∈ R\{0} is an eigenvalue of (3.15)
if and only if 1/λ is an eigenvalue of compact linear self-adjoint operator L := [F ′′(θ)]−1G′′(θ) ∈
Ls(H). By Riesz-Schauder theory, the spectrum of L, σ(L), contains a unique accumulation
point 0, and σ(L) \ {0} is a real countable set of eigenvalues of finite multiplicity, denoted by
{1/λn}∞
n=0Hn,
H0 = Ker(L) = Ker(G′′(θ)) and
n=1. Let Hn be the eigensubspace corresponding to 1/λn for n ∈ N. Then H = ⊕∞
Hn = Ker(I/λn − L) = Ker(F ′′(θ) − λnG′′(θ)), n = 1, 2,··· .
(3.16)
As another generalization of Rabinowitz bifurcation theorem [57] we have the following im-
provement of sufficiency of Theorem 12 in [58, Chap.4, §4.3].
Theorem 3.5. Let F,G = G1 ∈ C 1(U, R) be as in Theorem 3.3, and λ∗ be an eigenvalue of (3.15).
Suppose that the operator F ′′(θ) is invertible and also satisfies one of the following three condi-
tions: (a) positive definite, (b) negative definite, (c) each Hn in (3.16) with L = [F ′′(θ)]−1G′′(θ) is
an invariant subspace of F ′′(θ) (e.g. these are true if F ′′(θ) commutes with G′′(θ)), and F ′′(θ) is
either positive definite or negative one on Hn0 if λ∗ = λn0 . Then (λ∗, θ) ∈ R × U is a bifurcation
point for the equation (3.14) and one of the following alternatives occurs:
(i) (λ∗, θ) is not an isolated solution of (3.14) in {λ∗} × U .
(ii) there exists a sequence {κn}n≥1 ⊂ R \ {λ∗} such that κn → λ∗ and that for each κn the
equation (3.14) with λ = κn has infinitely many solutions converging to θ ∈ H.
(iii) for every λ in a small neighborhood of λ∗ there is a nontrivial solution uλ of (3.14) converg-
ing to θ as λ → λ∗;
(iv) there is a one-sided Λ neighborhood of λ∗ such that for any λ ∈ Λ \{λ∗}, (3.14) has at least
two nontrivial solutions converging to zero as λ → λ∗.
It is easily seen that the functional F in [58, §4.3, Theorem 4.3] or in [59, Chap.1, Theorem 3.4]
satisfies the conditions of this theorem for the case (a).
53
Proof of Theorem 3.5. Case 1. F ′′(θ) is either positive definite or negative one.
Clearly, we only need to consider the first case. For λ > λn and h ∈ Hn \ {θ} with n > 0,
since λnG′′(θ)h = F ′′(θ)h, we have
(F ′′(θ)h − λG′′(θ)h, h) = (1 − λ/λn)(F ′′(θ)h, h) < 0.
(3.17)
Clearly, if H0 6= {θ} (this is true if dim H = ∞), for h ∈ H0 \ {θ} it holds that
(F ′′(θ)h − λG′′(θ)h, h) = (F ′′(θ)h, h) > 0.
Let µλ denote the Morse index of Lλ := F − λG at θ. Then by (3.17) we obtain
dim Hn.
(3.18)
µλ = Xλn<λ
Assume λ∗ = λn0 for some n0 ∈ N. Then we have ε > 0 such that (λ∗ − 2ε, λ∗ + 2ε) \ {λ∗}
has no intersection with {λn}∞
n=1 since λn → ∞. By (3.18) it is easy to verify that
µλ = µλ∗,
∀λ ∈ (λ∗ − 2ε, λ∗],
µλ = µλ∗ + νλ∗,
∀λ ∈ (λ∗, λ∗ + 2ε),
(3.19)
(3.20)
where νλ∗ = dim Hn0 is the nullity of Lλ∗ at θ.
By Step 1 of proof of Theorem 2.14, we have
Claim 1. After shrinking ε > 0 we may verify that the homotopy
[λ∗ − ε, λ∗ + ε] × BH(θ, ε) → H, (λ, v) 7→ ∇Lλ(v)
is of class (S)+. So if {(κn, vn)}n≥1 ⊂ [λ∗ − ε, λ∗ + ε] × BH (θ, ε) satisfies ∇Lκn(vn) → θ and
κn → κ0, then {vn}n≥1 has a convergent subsequence in BH(θ, ε).
If (i) or (ii) holds, then (λ∗, θ) is a bifurcation point for (3.14).
Now suppose that neither (i) nor (ii) holds. Then we have
Claim 2.
θ ∈ H is an isolated critical point of Lλ for each λ ∈ [λ∗ − ε, λ∗ + ε] by shrinking
ε > 0 if necessary.
Writing H 0 = Hn0 and applying Theorem 2.18 to Lλ = Lλ∗ − (λ∗ − λ)G with λ ∈ [λ∗ −
ε, λ∗ + ε] and −G, we have δ ∈ (0, ε], ǫ > 0 and a unique continuous map
ψ : [λ∗ − δ, λ∗ + δ] × BH (θ, ǫ) ∩ H 0 → (H 0)⊥
(3.21)
such that for each λ ∈ [λ∗ − δ, λ∗ + δ], ψ(λ, θ) = θ and
P ⊥∇F(z + ψ(λ, z)) − λP ⊥∇G(z + ψ(λ, z)) = θ ∀z ∈ BH(θ, ǫ) ∩ H 0,
(3.22)
54
where P ⊥ is the orthogonal projection onto (H 0)⊥, and that the functional
BH(θ, ǫ) ∩ H 0 ∋ z 7→ L◦
λ(z) := F(z + ψ(λ, z)) − λG(z + ψ(λ, z))
(3.23)
is of class C 1, whose differential is given by
DL◦
λ(z)h = DF(z + ψ(λ, z))h − λDG(z + ψ(λ, z))h,
∀h ∈ H 0.
(3.24)
Hence the problem is reduced to finding the critical points of L◦
λ near θ ∈ H 0 for fixed λ near λ∗.
Note that Claim 2 is equivalent to the following
Claim 3.
θ ∈ H 0 is an isolated critical point of L◦
λ for each λ ∈ [λ∗ − δ, λ∗ + δ] by shrinking
δ > 0 if necessary.
Hence if θ ∈ H 0 is a local maximizer (resp. minimizer) of L◦
For a C 1 function ϕ on a neighborhood U of the origin θ ∈ RN we may always find ϕ ∈
C 1(RN , R) such that it agrees with ϕ near θ ∈ RN and is also coercive (so satisfies the (PS)-
condition). Suppose that θ is an isolated critical point of ϕ. By Proposition 6.95 and Example 6.45
λ, it must be strict.
in [52] we have
Ck(ϕ, θ; K) = δk0 ⇐⇒ θ is a local minimizer of ϕ,
Ck(ϕ, θ; K) = δkN ⇐⇒ θ is a local maximizer of ϕ, (cid:27)
(3.25)
and C0(ϕ, θ; K) = 0 = CN (ϕ, θ; K) if θ ∈ RN is neither a local maximizer nor a local minimizer
of ϕ.
Then by Theorem 2.1, (2.69) and (3.19) -- (3.20) we get that for any q ∈ N ∪ {0},
δq(µλ∗ +νλ∗ )K = Cq(Lλ, θ; K) = Cq−µλ∗ (L◦
δqµλ∗ K = Cq(Lλ, θ; K) = Cq−µλ∗ (L◦
λ, θ; K),
λ, θ; K),
∀λ ∈ (λ∗, λ∗ + δ],
∀λ ∈ [λ∗ − δ, λ∗).
It follows that
Cj(L◦
Cj(L◦
λ, θ; K) = δ(j+µλ∗ )(µλ∗ +νλ∗ )K = δjνλ∗ K,
λ, θ; K) = δ(j+µλ∗ )µλ∗
K = δj0K,
∀λ ∈ (λ∗, λ∗ + δ],
∀λ ∈ [λ∗ − δ, λ∗).
These and (3.25) imply
θ ∈ H 0 is a local minimizer of L◦
λ,
θ ∈ H 0 is a local maximizer of L◦
λ,
∀λ ∈ [λ∗ − ν, λ∗),
∀λ ∈ (λ∗, λ∗ + δ].
(3.26)
(3.27)
By (3.26) -- (3.27) and Theorem 3.2, one of the following possibilities occurs:
(I) for every λ in a small neighborhood of λ∗, L◦
λ has a nontrivial critical point converging to
θ ∈ H 0 as λ → λ∗;
(II) there is a one-sided Λ neighborhood of λ∗ such that for any λ ∈ Λ \ {λ∗}, L◦
λ has two
nontrivial critical points converging to zero as λ → λ∗.
Obviously, they lead to (iii) and (iv), respectively.
55
Case 2. Each Hn is an invariant subspace of F ′′(θ), n = 1, 2,··· . Note that Hn has an orthogonal
decomposition H +
n ) is the positive (resp. negative) definite subspace
of F ′′(θ)Hn . It is possible that H +
n = {θ}.
n (resp. H −
n = {θ} or H −
n , where H +
n ⊕ H −
As in (3.17), if H +
n 6= {θ}) and λ > λn (resp. λ < λn) we have
n 6= {θ} (resp. H −
(F ′′(θ)h − λG′′(θ)h, h) = (1 − λ/λn)(F ′′(θ)h, h) < 0
n 6= {θ}). Then the Morse index of Lλ at θ,
for h ∈ H +
n 6= {θ} (resp. h ∈ H −
Since λ∗ = λn0 , as in (3.19)-(3.20) it follows from these that
µλ = Xλn<λ
dim H +
dim H −
n .
n + Xλn>λ
µλ = µλ∗ + ν−
λ∗,
µλ = µλ∗ + ν+
λ∗,
∀λ ∈ (λ∗ − 2ε, λ∗),
∀λ ∈ (λ∗, λ∗ + 2ε),
(3.28)
(3.29)
(3.30)
(3.31)
(3.32)
(3.33)
(3.34)
λ∗ = dim H +
n0 (resp. ν−
λ∗ = dim H −
n0) is the positive (resp. negative) index of inertia of
where ν+
F ′′(θ)Hn0
.
Since F ′′(θ) is either positive definite or negative one on Hn0, we have either
µλ = µλ∗ + νλ∗,
or
µλ = µλ∗,
∀λ ∈ (λ∗ − 2ε, λ∗),
∀λ ∈ (λ∗, λ∗ + 2ε),
µλ = µλ∗ + νλ∗,
µλ = µλ∗. ∀λ ∈ (λ∗, λ∗ + 2ε),
∀λ ∈ (λ∗ − 2ε, λ∗),
We also suppose that neither (i) nor (ii) holds. Then Claim 1 and so Claim 2 holds. (3.31) --
(3.32) and (3.33) -- (3.34) lead, respectively, to
Cj(L◦
Cj(L◦
λ, θ; K) = δjνλ∗ K,
λ, θ; K) = δj0K,
and
Cj(L◦
Cj(L◦
λ, θ; K) = δjνλ∗ K,
λ, θ; K) = δj0K,
∀λ ∈ (λ∗, λ∗ + δ],
∀λ ∈ [λ∗ − δ, λ∗),
∀λ ∈ [λ∗ − δ, λ∗),
∀λ ∈ (λ∗, λ∗ + δ].
The former yields (3.26) -- (3.27) as above. Similarly, the latter and (3.25) imply
θ ∈ H 0 is a local maximizer of L◦
λ,
θ ∈ H 0 is a local minimizer of L◦
λ,
∀λ ∈ [λ∗ − δ, λ∗),
∀λ ∈ (λ∗, λ∗ + δ].
They and Theorem 3.2 show that either (c) or (d) holds.
✷
56
Remark 3.6. Suppose that (iv) is replaced by the following weaker conclusion
(iv') there is a one-sided Λ neighborhood of λ∗ such that for any λ ∈ Λ \ {λ∗}, (3.14) has a
nontrivial solution uλ converging to zero as λ → λ∗.
Then the finite dimension reduction in the proof is not needed; and the assumption "F ′′(θ) is
either positive definite or negative one on Hn0" in (c) may be replaced by the weaker condition
"F ′′(θ)Hn0
has nonzero signature".
3.3 Bifurcation for equivariant problems
Now let us generalize the above results to the equivariant case. The first is a partial generalization
of [67, Theorem 4.2] and [68, Theorem 2.5]. The proofs of the latter were based on Morse theory
ideas of [15]. Because of our theory in Section 2, the same methods may be used with some
technical improvements.
Theorem 3.7. Under the assumptions of Theorem 3.5, let G be a compact Lie group acting on
H orthogonally, and suppose that U , F and G are G-invariant. For an eigenvalue λ∗ of (3.15),
suppose that the operator F ′′(θ) is invertible and also satisfies one of the following three condi-
tions: (a) positive definite, (b) negative definite, (c) each Hn in (3.16) with L = [F ′′(θ)]−1G′′(θ)
is an invariant subspace of F ′′(θ) (e.g. these are true if F ′′(θ) commutes with G′′(θ)), and F ′′(θ)
is either positive definite or negative one on H 0 = Hn0 if λ∗ = λn0. Then
1) (λ∗, θ) ∈ R × U is a bifurcation point for the equation (3.14).
2) If dim Hn0 ≥ 2 and the unit sphere in Hn0 is not a G-orbit we must get one of the following
alternatives:
(i) (λ∗, θ) is not an isolated solution of (3.14) in {λ∗} × U ;
(ii) there exists a sequence {κn}n≥1 ⊂ R \ {λ∗} such that κn → λ∗ and that for each κn
the equation (3.14) with λ = κn has infinitely many G-orbits of solutions converging
to θ ∈ H;
(iii) for every λ in a small neighborhood of λ∗ there is a nontrivial solution uλ of (3.14)
converging to θ as λ → λ∗;
(iv) there is a one-sided Λ neighborhood of λ∗ such that for any λ ∈ Λ \ {λ∗}, (3.14) has
at least two nontrivial critical orbits converging to zero θ as λ → λ∗.
3) Suppose one of the following assumptions holds: 3.a) G = T m and Fix(G)∩ Hn0 = {θ}; 3.b)
G = T m and every orbit in Hn0 is homeomorphic to some T s for s ≥ 2; 3.c) G is a finite
group and the greatest common divisor δ of set {♯G/♯Gx x ∈ Hn0 \ {θ}} is equal to or
bigger than 2, where ♯S denotes the number of elements in a set S. Then either one of the
above (i)-(iii) or the following hold:
57
(iv)' there is a one-sided Λ neighborhood of λ∗ such that for any λ ∈ Λ \ {λ∗}, (3.14)
has at least dim Hn0 (resp. 2 dim Hn0 , δ dim Hn0) nontrivial critical orbits in the case 3.a)
(resp. 3.b), 3.c)), where every orbit is counted with its multiplicity. (The multiplicity of a
q=0 rankCq(Lλ,O) in [68,
critical orbit O of Lλ = F − λG was defined as c(O) = P∞
Definition 1.3]).
This theorem can also be viewed generalizations of [27, 28]. Even so, we still give a direct
generalization version of [27, 28] in Theorem 3.9 below (because different methods need be em-
ployed). Another different point is that the number of critical orbits in 3) is counted in a non-usual
way. After Theorem 3.9 we shall compare these two theorems.
Proof of Theorem 3.7.
1) follows from Theorem 3.5. For 2), as in the proof of Theorem 3.5,
we assume that neither (i) nor (ii) holds. Then θ ∈ H 0 is a unique critical orbit of the functional
λ(z) in (3.23) for each λ ∈ [λ∗−2δ, λ∗ +2δ] by shrinking δ > 0 and ǫ > 0
BH(θ, ǫ)∩H 0 ∋ z 7→ L◦
λ, it must be strict). Moreover, since dim H 0 < ∞,
if necessary. (Thus if θ is an extreme point of L◦
λ λ ∈ [λ∗ − 2δ, λ∗ + 2δ]} satisfies the
replacing ǫ by a slightly smaller one we can assume that {L◦
(PS) condition. That is, if zk ∈ BH(θ, ǫ)∩ H 0 and λk ∈ [λ∗ − 2δ, λ∗ + 2δ] satisfy DL◦
(zk) → 0
and supk L◦
(zk) < ∞, then {(zk, λk)}∞
k=1 has a converging subsequence.
λk
λk
λ near θ ∈ H 0 for fixed λ near λ∗.
The problem is reduced to finding the critical orbits of L◦
Firstly, we assume that F ′′(θ) is positive definite.
Since θ ∈ H 0 is an isolated critical orbit of L◦
λ∗, by (2.69) and (3.25), we have three cases:
• Cq(Lλ∗, θ; K) = δqµλ∗ K if θ ∈ H 0 is a local minimizer of L◦
λ∗;
• Cq(Lλ∗, θ; K) = δq(µλ∗ +νλ∗ )K if θ ∈ H 0 is a local maximizer of L◦
λ∗;
• Cq(Lλ∗, θ; K) = 0 for q /∈ (µλ∗, µλ∗ + νλ∗) if θ ∈ H 0 is neither a local maximizer nor a local
minimizer of L◦
λ∗.
In the third case, the stability of critical groups implies (iii) as before.
For the first two cases, as in the proofs of (3.26) -- (3.27), if θ ∈ H 0 is a local minimizer of L◦
λ∗
we can obtain
θ ∈ H 0 is a local minimizer of L◦
λ,
θ ∈ H 0 is a local maximizer of L◦
λ,
∀λ ∈ [λ∗ − 2δ, λ∗],
∀λ ∈ (λ∗, λ∗ + 2δ];
and if θ ∈ H 0 is a local maximizer of L◦
λ∗ we have
θ ∈ H 0 is a local minimizer of L◦
λ,
θ ∈ H 0 is a local maximizer of L◦
λ,
∀λ ∈ [λ∗ − 2δ, λ∗),
∀λ ∈ [λ∗, λ∗ + 2δ].
(3.35)
(3.36)
(3.37)
(3.38)
Now let us follow the proof ideas of [67, page 220] to complete the remaining arguments. But
λ is only of class C 1. Fortunately, the standard arguments (cf.
differen from the case therein our L◦
[52, Propsition 5.57]) may yield
58
Lemma 3.8. There exists a smooth map
(BH(θ, ǫ) ∩ H 0 \ {θ}) × (λ∗ − 2δ, λ∗ + 2δ) → H, (z, λ) 7→ Vλ(z)
such that for each λ ∈ (λ∗−2δ, λ∗+2δ) the map Vλ : BH(θ, ǫ)∩H 0\{θ} → H is a G-equivariant
pseudo-gradient vector field for L◦
kVλ(z)k ≤ 2kDL◦
λ, precisely for all z ∈ BH(θ, ǫ) ∩ H 0 \ {θ},
λ(z)k
λ(z), Vλ(z)i ≥
1
2kDL◦
λ(z)k2.
and
hDL◦
For completeness we are also to give the proof it, which is postponed after the proof of this
theorem.
When θ ∈ H 0 is a local minimizer of L◦
λ∗ (and so a strict local minimizer as noted above),
(3.35) and (3.36) are satisfied. Let c = L◦
λ∗(θ). We can choose ε > 0 so small that Wε = {z ∈
BH(θ, ǫ) ∩ H 0 L◦
λ∗ (z) < c + ε} is a contractible neighborhood of θ. Obverse that Wε is G-
invariant and the flow of −Vλ∗ preserves Wε. (Actually, it is not hard to construct a deformation
contraction from Wε to θ with the flow of −Vλ∗). Since θ ∈ H 0 is a unique critical orbit of the
λ λ ∈ [λ∗ − 2δ, λ∗ + 2δ]} satisfies
functional L◦
λ in Wε for each λ ∈ [λ∗ − 2δ, λ∗ + 2δ], and {L◦
the (PS) condition, by the theorem on continuous dependence of solutions of ordinary differential
equations on initial values and parameters we may deduce that there exists 0 < ¯δ < 2δ such that
the flow of −Vλ also preserves Wε for each λ ∈ [λ∗ − ¯δ, λ∗ + ¯δ]. For any λ ∈ (λ∗, λ∗ + ¯δ], since
θ ∈ H 0 is a local maximizer of L◦
λ must have a minimal critical orbit O 6= {θ}.
Note that Theorems 3.1, 3.2 also hold in our case by Corollaries 2.27, 2.28. Repeating the other
λ by (3.36), L◦
arguments in the proof of [67, page 220] we may verify that for λ ∈ (λ∗, λ∗ + δ] close to λ∗, L◦
λ
has three critical orbits.
Similarly, if (3.37) and (3.38) hold, L◦
λ has three critical orbits for each λ ∈ [λ∗ − δ, λ∗) close
to λ∗.
By considering −F ′′(θ) we get the conclusion if F ′′(θ) is negative definite.
Next, we consider the case that the condition (c) holds.
If θ ∈ H 0 is a local minimizer of L◦
λ∗, by (3.31) and (3.32) we get
θ ∈ H 0 is a local minimizer of L◦
λ,
θ ∈ H 0 is a local maximizer of L◦
λ,
∀λ ∈ [λ∗ − δ, λ∗],
∀λ ∈ (λ∗, λ∗ + δ].
If θ ∈ H 0 is a local maximizer of L◦
λ∗, by (3.33) and (3.34) we have
θ ∈ H 0 is a local maximizer of L◦
λ,
θ ∈ H 0 is a local minimizer of L◦
λ,
∀λ ∈ [λ∗ − δ, λ∗],
∀λ ∈ (λ∗, λ∗ + δ].
(3.39)
(3.40)
(3.41)
(3.42)
If θ ∈ H 0 is neither a local maximizer nor a local minimizer of L◦
λ∗, (iii) will occur by the
stability of critical groups implies as before. The proofs of the first two cases are as above.
Finally, we prove 3). As in 2) we only consider the case that F ′′(θ) is positive definite. The
other cases can be proved similarly. We assume that any one of (i)-(iii) does not occur. Then we
59
have either (3.35)-(3.36) or (3.35)-(3.36). In these two cases we obtain that θ ∈ H 0 is a local
maximizer of L◦
λ for any λ ∈ (λ∗, λ∗ + 2δ] (resp. for all λ ∈ [λ∗, λ∗ + 2δ]). Note νλ∗ = dim Hn0.
It follows (or from Theorem 2.19) that
Cq(L◦
λ, θ; K) = δqνλ∗ K,
q = 0, 1,···
for any λ ∈ (λ∗, λ∗ + 2δ] (resp. for all λ ∈ [λ∗, λ∗ + 2δ]). These show that θ is essentially the
same as nondegenerate critical point of L◦
λ with Morse index νλ∗ = dim Hn0. Obverse that the
conclusions of [68, Corollary 1.3] also hold in the present case (because X = Hn0 has finite
λ is only of class C 1. The results in the cases 3.a) and 3.b) follow as in [68].
dimension) though L◦
The case 3.c) can be derived from [68, Theorem 1.3].
✷
λ ∈ C 1(BH (θ, ǫ) ∩ H 0) is continuous by (2.51), and
Proof of Lemma 3.8. Note that λ 7→ L◦
that DL◦
λ has no any zero point in (BH (θ, ǫ) ∩ H 0 \ {θ}) × [λ∗ − 2δ, λ∗ + 2δ]. For any given
z ∈ BH(θ, ǫ) ∩ H 0 \ {θ} and λ ∈ (λ∗ − 2δ, λ∗ + 2δ) we have an open neighborhood O(z,λ) of z
in BH(θ, ǫ)∩ H 0 \{θ}, a positive number r(z,λ) with (λ− r(z,λ), λ + r(z,λ)) ⊂ (λ∗ − 2δ, λ∗ + 2δ)
and a unit vector v(z,λ) ∈ H such that for all (z′, λ′) ∈ O(z,λ) × (λ − r(z,λ), λ + r(z,λ)),
kv(z,λ)k ≤ 2kDL◦
λ′(z′)k and
hDL◦
λ′(z′), v(z,λ)i ≥
1
2kDL◦
λ′(z′)k2.
Now all above O(z,λ) × (λ − r(z,λ), λ + r(z,λ)) form an open cover Q of (BH(θ, ǫ)∩ H 0 \{θ})×
(λ∗ − 2δ, λ∗ + 2δ), and the latter admits a C ∞-unit decomposition {ηα}α∈Ξ subordinate to a
locally finite refinement {Wα}α∈Ξ of Q. Since each Wα can be contained in some open subset of
form O(z,λ) × (λ − r(z,λ), λ + r(z,λ)), we have a unit vector vα ∈ H such that
λ′(z′)k2
kvαk ≤ 2kDL◦
λ′(z′), vαi ≥
1
2kDL◦
λ′(z′)k
hDL◦
and
for all (z′, λ′) ∈ Wα. Set χ =Pα∈Ξ ηαvα. Then it is a smooth map from (BH (θ, ǫ)∩ H 0\{θ})×
(λ∗ − 2δ, λ∗ + 2δ) to H, and satisfies
λ(z)k
kχ(z, λ)k ≤ 2kDL◦
λ(z), χ(z, λ)i ≥
1
2kDL◦
λ(z)k2
hDL◦
and
for all (z, λ) ∈ (BH(θ, ǫ)∩ H 0 \{θ})× (λ∗ − 2δ, λ∗ + 2δ). Let dµ denote the right invariant Haar
measure on G. Define
(BH(θ, ǫ) ∩ H 0 \ {θ}) × (λ∗ − 2δ, λ∗ + 2δ) ∋ (z, λ) 7→ Vλ(z) =ZG
g−1χ(gz, λ)dµ ∈ H.
It is easily checked that Vλ satisfies requirements.
✷
The following is a direct generalization of Fadell -- Rabinowitz theorems [27, 28].
Theorem 3.9. Under the assumptions of Theorem 3.7, if the Lie group G is equal to Z2 or S1,
Then (λ∗, θ) ∈ R × U is a bifurcation point for the equation (3.14), and if dim Hn0 ≥ 2 and the
unit sphere in H 0 = Hn0 is not a G-orbit we must get one of the following alternatives:
60
(i) (λ∗, θ) is not an isolated solution of (3.14) in {λ∗} × U ;
(ii) there exists a sequence {κn}n≥1 ⊂ R \ {λ∗} such that κn → λ∗ and that for each κn the
equation (3.14) with λ = κn has infinitely many G-orbits of solutions converging to θ ∈ H;
(iii) there exist left and right neighborhoods Λ− and Λ+ of λ∗ in R and integers n+, n− ≥ 0,
such that n+ + n− ≥ dim H 0 = dim Hn0 and for λ ∈ Λ− \ {λ∗} (resp. λ ∈ Λ+ \ {λ∗}),
(3.14) has at least n− (resp. n+) distinct critical G-orbits different from θ, which converge
to zero θ as λ → λ∗.
Different from Theorem 3.7, if (λ∗, θ) is not an isolated solution of (3.14) in {λ∗} × U , The-
k ↑ λ∗ such that for each k ∈ N, the
k G plus those of non-trivial critical orbits
in U are at least dim H 0. Moreover, these non-trivial critical orbits converge θ as k → ∞.
orem 3.9 implies that there exist sequences λ+
numbers of non-trivial critical orbits of Lλ+
of Lλ−
Note also that the count method for critical orbits in 3) of Theorem 3.7 is not usual as the present
k ↓ λ∗ and λ−
= F − λ+
k
k
one.
Proof of Theorem 3.9. We assume that the cases (i) and (ii) do not occur. Moreover, we only
consider cases: 1) F ′′(θ) is positive definite, 2) (c) holds and F ′′(θ) is positive definite on H 0. In
these cases, by the proofs of Theorems 3.5,3.7, we obtain:
A) if θ ∈ H 0 is a local minimizer of L◦
λ∗, i.e., Cq(L◦
λ∗, θ; K) = δq0K, then
B) if θ ∈ H 0 is a local maximizer of L◦
λ∗, i.e., Cq(L◦
∀λ ∈ [λ∗ − δ, λ∗],
∀λ ∈ (λ∗, λ∗ + δ];
λ∗, θ; K) = δqνλ∗ K, then
∀λ ∈ [λ∗ − δ, λ∗),
∀λ ∈ [λ∗, λ∗ + δ];
θ ∈ H 0 is a local minimizer of L◦
λ,
θ ∈ H 0 is a local maximizer of L◦
λ,
θ ∈ H 0 is a local minimizer of L◦
λ,
θ ∈ H 0 is a local maximizer of L◦
λ,
C) if θ ∈ H 0 is neither a local maximizer of L◦
0 for q = 0, νλ∗ , then
λ∗ nor a local maximizer of it, i.e., Cq(L◦
λ∗, θ; K) =
∀λ ∈ [λ∗ − δ, λ∗),
∀λ ∈ (λ∗, λ∗ + δ].
Let us shrink ǫ > 0 in (3.23) such that θ is the only critical point of L◦
θ ∈ H 0 is a local minimizer of L◦
λ,
θ ∈ H 0 is a local maximizer of L◦
λ,
λ in BH(θ, ǫ) ∩ H 0.
Since λ 7→ L◦
λ ∈ C 1(BH (θ, ǫ) ∩ H 0) is continuous by (2.51), it is easy to see that
Rδ,ǫ := {(λ, z) ∈ [λ∗ − δ, λ∗ + δ] × (BH(θ, ǫ) ∩ H 0) DL◦
λ(z) 6= θ}
is an open subset in [λ∗ − δ, λ∗ + δ] × (BH(θ, ǫ) ∩ H 0), and
Rδ,ǫ = {(λ, z) ∈ [λ∗ − δ, λ∗ + δ] × BH(θ, ǫ) ∩ H 0 z ∈ (BH(θ, ǫ) ∩ H 0) \ K(L◦
λ)},
61
where K(L◦
smooth map, Rδ,ǫ → H 0, (λ, z) 7→ Vλ(z), such that each
λ) denotes the critical set of L◦
λ. As in the proof of Lemma 3.8 we can produce a
Vλ : BH(θ, ǫ) ∩ H 0 \ K(L◦
λ) → H 0, z 7→ Vλ(z)
is a G-equivariant C ∞ pseudo-gradient vector field for L◦
λ.
Replacing [57, (11.1)] (or [27, (2.4)]) for G = Z2, and [28, (8.19)] for G = S1 by
dϕλ
ds
(3.43)
= −Vλ(ϕλ), ϕλ(0, z) = z,
we can repeat the constructions in [57, §1] and [27, §8] to obtain:
Lemma 3.10. There is a G-invariant open neighborhood Q of θ in H 0 with compact closure Q
contained in BH(θ, ǫ)∩ H 0 such that for every λ close to λ∗, every c ∈ R and every τ1 > 0, every
G-neighborhood U of Kλ,c := K(L◦
λ(z) ≤ c} there exists an τ ∈ (0, τ1) and a G
equivariant homotopy η : [0, 1] × Q → Q with the following properties:
1◦ η(t, z) = z if z ∈ Q \ (L◦
2◦ η(t,·) is homeomorphism of Q to η(t, Q) for each t ∈ [0, 1];
3◦ η(1, Aλ,c+τ \ U ) ⊂ Aλ,c−τ ), where Aλ,d := {z ∈ Q L◦
4◦ if Kλ,c = ∅, η(1, Aλ,c+τ ) ⊂ Aλ,c−τ ).
λ)−1[c − τ1, c + τ1];
λ)∩ {z ∈ Q L◦
λ(z) ≤ d};
For ∗ = +,−, let S∗ = {z ∈ BH(θ, ǫ) ∩ H 0 ψ(s, z) ∈ BH (θ, ǫ) ∩ H 0, ∀ ∗ s > 0} and
T ∗ = S∗ ∩ ∂Q. For G = Z2 (resp. S1) let iG denote the genus in [57] (resp. the index in [28, §7]).
Lemma 3.11. Both T + and T − are G-invariant compact subset of ∂Q, and also satisfy
λ(z) z ∈ T +} > 0 and max{L◦
λ(z) z ∈ T −} < 0;
1◦ min{L◦
2◦ iZ2 (T +) + iZ2 (T −) ≥ dim H 0 and iS 1(T +) + iS 1(T −) ≥ 1
2 dim H 0.
Two inequalities in 2◦ are [27, Lemma 2.11] and [28, Theorem 8.30], respectively.
Case G = Z2. Suppose iZ2(T −) = k > 0. Let cj be defined by [27, (2.13)], but ¯Q and g(λ, v)
are replaced by Q and L◦
λ(z), respectively. We can modify the proof of (i) on the page 54 of [27]
as follows:
In the above three cases A), B) and C), for each λ ∈ [λ∗ − δ, λ∗), θ ∈ H 0 is always a
λ. Therefore for arbitrary sufficiently small ρ > 0, depending on λ,
local (strict) minimizer of L◦
L◦
λ(z) > 0 for any 0 < kzk ≤ ρ and so
c1 ≥ min
kzk=ρL◦
λ(z) > 0.
Other arguments are same. Hence we obtain: if λ ∈ [λ∗ − δ, λ∗) is close to λ∗, L◦
distinct pairs of nontrivial critical points, which also converge to θ as λ → λ∗.
λ has at least k
62
Since for every λ ∈ (λ∗, λ∗ + δ], θ ∈ H 0 is a local maximizer of L◦
get: if iZ2 (T +) = l > 0, for every λ ∈ (λ∗, λ∗ + δ] close to λ∗, L◦
nontrivial critical points converging to θ as λ → λ∗.
λ, by considering −L◦
λ we
λ has at least l distinct pairs of
These two claims together yield the desired result.
Case G = S1. Suppose iS 1(T −) = k > 0. Similarly, for cj defined by [28, (8.56)], we may
replace [28, (8.58), (8.63)] by
L◦
λ(x) ≥ min
kzk=ρL◦
λ(z) > 0,
and so
cγ+1 ≥ min
kzk=ρL◦
λ(z) > 0,
and then repeat the arguments in [28, §8] to complete the final proof. Of course, we also use the
fact that L◦
λ∗ uniformly on Q as λ → λ∗, which can be derived from (2.51).
λ → L◦
✷
Fadell -- Rabinowitz theorems in [27, 28] were also generalized to the case of arbitrary compact
Lie groups for potential operators of C 2 functionals by Bartsch and Clapp [2], Bartsch [1]. We
now give generalizations of their some results.
Fix a set A of G-spaces, a multiplicative equivariant cohomology theory h∗ and an ideal I of
the coefficient ring R = h∗(pt). Recall in [1, Definition 4.1] that the (A, h∗, I)-length of a G-space
X, (A, h∗, I)-length(X), was defined to be the smallest integer k such that there exist A1,··· , Ak
in A with the following property: For all γ ∈ h∗(X) and for all ωi ∈ I ∩ kern(h∗(pt) → h∗(Ai)),
i = 1,··· , k, the product ω1 · . . .· ωk · γ = 0 in h∗(X). Moreover, set (A, h∗, I)-length(X) = ∞
if no such k exists. Let SE denote the unit sphere in a Hilbert space E, and G be a compact Lie
group acting on E orthogonally. Denote by G the set of orbits occurring on SE. Let h∗ be any
continuous, multiplicative, equivariant cohomology theory such that kern(h∗(pt) → h∗(G/H)) is
a finitely generated ideal for all G/H ∈ G . Taking I = R = h∗(pt) the (A, h∗, I)-length of a G-
space X becomes the (G , h∗)-length ℓ(X) defined in [2]. For a bounded closed G-neighborhood
V of the origin in a G-module E it was proved in [2, Lemma 1.6] that ℓ(∂V ) = ℓ(SE). If G = Z2
(resp. S1) and h∗ = H ∗
G, ℓ becomes indexR (resp. indexC) in [28].
Hypothesis 3.12. Let G be a compact Lie group acting on H orthogonally, and U a G-invariant
open neighborhood of the origin of a real Hilbert space H. Let F,G = G1 ∈ C 1(U, R) be as in
Theorem 3.3, and G-invariant. Let λ∗ be an isolated eigenvalue of (3.15), i.e., for each λ 6= λ∗
near λ∗ the equation (3.15) has only trivial solution, and let H 0 be the corresponding eigenspace.
(Every eigenvalue of (3.15) is isolated if F ′′(θ) is invertible.) Suppose H 0 ∩ Fix(G) = {θ}.
As in the proof of Theorem 3.5, applying Theorem 2.18 to Lλ = F − λG = Lλ∗ − (λ∗ − λ)G
with λ ∈ [λ∗ − ε, λ∗ + ε] and −G, we have δ ∈ (0, ε], ǫ > 0 and a unique continuous map
ψ : [λ∗ − δ, λ∗ + δ] × (BH (θ, ǫ) ∩ H 0) → (H 0)⊥
as in (3.21), such that (3.22) and (3.23) -- (3.24) hold. But the present ψ is G-equivariant and each
L◦
λ is G-invariant.
Hypothesis 3.13. Under Hypothesis 3.12, suppose that the continuous map ψ as in (3.21) is of
class C 1 with respect to the second variable, which implies that dL◦
λ has Gateaux derivative
63
d2L◦
λ(z)(u, v) = (F ′′
− λ(G′′
λ (z + ψ(λ, z))(u + Dzψ(λ, z)u), v)H
λ(z + ψ(λ, z))(u + Dzψ(λ, z)u), v)H ∀u, v ∈ H 0
at every z ∈ BH(θ, ǫ) ∩ H 0 by (2.61), and therefore for all u, v ∈ H 0,
d2L◦
λ(θ)(u, v) = (λ∗ − λ)(G′′(θ)u, v)H = (λ∗ − λ)(P 0G′′(θ)H 0u, v)H
(3.44)
because ψ(λ, θ) = θ and Dzψ(λ, θ) = θ. Furthermore, we assume that each L◦
λ is of class C 2.
Since λ∗ is an isolated eigenvalue of (3.15), 0 is an eigenvalue of F ′′(θ) − λ0G′′(θ) with
finite multiplicity, isolated in the spectrum σ(F ′′(θ)− λ0G′′(θ)). Hence for λ near λ∗ the 0-group
eig0(F ′′(θ) − λG′′(θ)) (cf. Section 3.1) is well-defined. Let Eλ be the generalized eigenspace of
F ′′(θ) − λG′′(θ) belonging to eig0(F ′′(θ) − λG′′(θ)) ∩ R−. It is G-invariant. For λ and λ′ near
λ∗ the spaces Eλ and Eλ′ are G-isomorphic if (λ − λ∗)(λ′ − λ∗) > 0. (When the latter holds
the orthogonal eigenprojection (cf. [33, page 181] for the definition), Pλ : H → Eλ, restricts to a
G-isomorphism from Eλ′ onto Eλ.) Since θ ∈ H 0 is a nondegenerate critical point of L◦
λ for each
λ 6= λ∗ near λ∗, P 0G′′(θ)H 0 : H 0 → H 0 is an isomorphism by (3.44). Let H 0
+ and H 0
− be the
+ ⊕ H 0
positive and negative definite subspaces of P 0G′′(θ)H 0 , respectively. Then H 0 = H 0
−. Let
F −
λ (resp. F +
λ(θ))∩R+). Clearly,
λ the orthogonal complement of F −
F +
λ′ are G-isomorphic if
(λ − λ∗)(λ′ − λ∗) > 0. Hence if ℓ is the above (G , h∗)-length, for λ < λ∗ < µ close to λ∗, the
λ ) be the eigenspace belonging to σ(d2L◦
λ(θ))∩R− (resp. σ(d2L◦
λ and F −
λ in H 0, and the spaces F −
number
d := ℓ(SH 0) − min{ℓ(SF −
λ ) + ℓ(SF +
µ ), ℓ(SF +
λ ) + ℓ(SF −
µ )
is well-defined, and if ℓ(SV ) = c · dim V for every G-module V with V G = {θ} we have
d = ℓ(SF −
λ ) − ℓ(SF −
µ ) = c dim F −
λ − dim F −
µ = c dim Eλ − dim Eµ
(3.45)
(3.46)
([2]). Here the final equality comes from the fact that P 0 defines a G-isomorphism from Eλ onto
λ (cf. [2, page 353]). Moreover, it is easy to see that F −
F −
+ for λ < λ∗, and
F −
λ = H 0
− for λ > λ∗. (3.45) and (3.46), respectively, become
− and F +
+ and F +
λ = H 0
λ = H 0
λ = H 0
d = ℓ(SH 0) − 2 min{ℓ(SH 0
d = ℓ(SH 0
−), ℓ(SH 0
+) = c dim H 0
−) − ℓ(SH 0
+)),
− − dim H 0
+.
(3.47)
(3.48)
Having these we may state the following partial generalization of [2, Theorem 3.1].
Theorem 3.14. Under Hypothesis 3.13, if the number d in (3.45) or (3.47) is positive, then
(λ∗, θ) ∈ R × U is a bifurcation point for the equation (3.14) and one of the following alter-
natives occurs:
64
(i) (λ∗, θ) is not an isolated solution of (3.14) in {λ∗} × U .
(ii) there exist left and right neighborhoods Λl and Λr of λ∗ in R and integers il, ir ≥ 0 such that
il + ir ≥ d and for any λ ∈ Λl \ {λ∗} (resp. λ ∈ Λr \ {λ∗}, (3.14) has at least il (resp. ir)
distinct nontrivial solution orbits, which converge to θ in H as λ → λ∗.
Proof. Suppose that (i) does not hold. Then θ ∈ H 0 is an isolated critical point of L◦
λ∗. We
assume that θ ∈ H 0 is a unique critical point in the set Q of Lemma 3.10. Moreover the flow
λ since we have assumed L◦
ϕλ of (3.43) may be replaced by the negative gradient one χλ of L◦
to be of class C 2. Then we get a corresponding Lemma 3.10. For ∗ = +,−, let S∗ = {z ∈
BH(θ, ǫ)∩ H 0 χλ(s, z) ∈ BH (θ, ǫ)∩ H 0, ∀∗ s > 0} and T ∗ = S∗∩ ∂Q. We have the following
λ
corresponding result with the part b) of [2, Lemma 4.1].
Lemma 3.15. Both T + and T − are G-invariant compact subset of ∂Q, and also satisfy
λ(z) z ∈ T +} > 0 and max{L◦
1◦ min{L◦
2◦ T + and T − can be deformed inside Q \ {θ} into arbitrarily small neighborhoods of θ ∈ Q
λ(z) z ∈ T −} < 0;
such that L◦
λ∗ does not change sign during the deformation.
3◦ ℓ(T +) + ℓ(T −) ≥ ℓ(SH 0).
λ(T −) ⊂ R− for all λ ∈ Λ. Note that for each λ 6= λ∗ near λ∗, F +
Recall Λ = [λ∗ − δ, λ∗ + δ]. Shrinking δ > 0 (if necessary) we may assume that L◦
λ(T +) ⊂
R+ and L◦
λ is the tangent
space of the stable manifold of the negative gradient one χλ because θ ∈ H 0 is a nondegenerate
critical point of the C 2 function L◦
λ. We can complete proofs of corresponding results with [2,
Lemmas 4.2,4.3].
Clearly, even if G = S1 or Z2, Theorem 3.9 cannot be included in Theorem 3.14 and the
following two results.
Now consider generalizations of bifurcation results [1, §7.5]. Once (A, h∗, I) is understood its
(A, h∗, I)-length is written as ℓ below.
Under Hypothesis 3.12, let ϕλ be the flow of Vλ given by (3.43). Assume that Λ is equipped
with trivial G-action. By the theorem on continuous dependence of solutions of ordinary differen-
tial equations on initial values and parameters we obtain an equivariant product flow parametrized
by Λ on Λ × BH 0(θ, ǫ), (λ, z, t) → ϕ(λ, z, t) := (λ, ϕλ(z, t)). Clearly, Vλ is gradient-like with
λ. Since λ∗ is an isolated eigenvalue of (3.15), θ ∈ H 0 is an isolated critical
Lyapunov-function L◦
point of L◦
λ (and so an isolated invariant set of ϕλ) for each λ near λ∗ and λ 6= λ∗. Let ℓu(λ, θ)
(resp. ℓs(λ, θ)) be the exit-length (resp. entry-length) of θ with respect to ϕλ, see [1, §7.3]. More-
over, ℓu(λ, θ) is independent of λ ∈ Λ∩ (λ∗,∞) (resp. Λ∩ (−∞, λ∗)) close to λ∗, denoted by ℓu
(resp. ℓu
−). See [1, §7.2, §7.5] for these. By Theorems 7.10, 7.11 in [1] we immediately obtain the
+
following two theorems.
65
+ 6= ℓu
−. Then (λ∗, θ) is a bifurcation point
λ∗ , then there exits ε > 0
− so that for each λ ∈ (λ∗ − ε, λ∗) respectively
λ. Moreover, if θ ∈ H 0 is also an isolated critical point of L◦
Theorem 3.16. Under Hypothesis 3.12, suppose that ℓu
of ∇L◦
and integers dl, dr ≥ 0 with dl + dr ≥ ℓu
λ ∈ (λ∗, λ∗ + ε) at least one of the following alternatives occurs:
(i) there exist critical G-orbits Gui
λ) → L◦
λ 6= θ, L◦
λ, i ∈ Z\{0}, of L◦
λ(θ) as i → ∞. In particular, L◦
for i ≥ 1, and L◦
G-orbits. They converge to θ ∈ H 0 as λ → λ∗.
+ − ℓu
λ with ui
λ(ui
λ ) < L◦
λ(u−i
λ(ui
λ)
λ has infinitely many critical
λ(θ) < L◦
(ii) There exists an isolated invariant set Sλ ∈ BH 0(θ, ǫ) \ {θ} with ℓ(C(Sλ)) ≥ dl respectively
ℓ(C(Sλ)) ≥ dr. Moreover, Sλ converge to θ ∈ H 0 as λ → λ∗, i.e., for any neighborhood
W of θ in H 0 there exist ǫ = ǫ(W ) > 0 such that Sλ ∈ W if λ − λ∗ < ǫ.
The result is also true for G = Z/p and ℓ = ℓ0 + ℓ1 − 1 as in [1, Remark 4.14].
Theorem 3.17. Under Hypothesis 3.12, suppose that G = Z/p, p a prime, or G = S1 × Γ, Γ a
finite group, and let ℓ be any of the lengths defined in [1, 4.4, 4.14 or 4.16]. If θ ∈ H 0 is also an
isolated critical point of L◦
λ∗ , and ℓu
+ are the exit-lengths as above, then there exits ε > 0 and
integers dl, dr ≥ 0 with dl+dr ≥ ℓu
− such that the following holds: For each λ ∈ (λ∗−ε, λ∗)
respectively λ ∈ (λ∗, λ∗ + ε) there exists a compact invariant set Sλ ∈ BH 0(θ, ǫ) \ {θ} with
ℓ(C(Sλ)) ≥ dl respectively ℓ(C(Sλ)) ≥ dr.
−, ℓu
+−ℓu
Every G-critical orbit of L◦
λ produced by Theorems 3.16, 3.17 gives rise to a G-critical orbit of
Lλ, and different orbits yield different ones too. In particular, (λ∗, θ) ∈ R×U is a bifurcation point
for the equation (3.14). However, in order to understand ℓu
− we assume that Hypothesis 3.13
is satisfied. Then the flow ϕλ may be replaced by the negative gradient one χλ of L◦
λ. By the
arguments below Definition 7.1 in [1] we have ℓu
+) = ℓ(SEλ)
−) = ℓ(SEλ)) if λ > λ∗ (resp. λ < λ∗) is close to λ∗. It
(resp. ℓu
) := ℓ(SEλ∗+ρ) are independent of small ρ > 0
follows that ℓ(SEλ∗
and that
−) := ℓ(SEλ∗−ρ) and ℓ(SEλ∗
+ = ℓu(λ, θ) = ℓ(SF −
+ = ℓu(λ, θ) = ℓ(SF −
λ ) = ℓ(SH 0
λ ) = ℓ(SH 0
+ and ℓu
+
+ − ℓu
ℓu
− = ℓ(SH 0
+) − ℓ(SH 0
+ − dim Eλ∗
−) = ℓ(SEλ∗
− (resp. 1
+
) − ℓ(SEλ∗
−),
+ − dim Eλ∗
−) if G = Z/p
which may be chosen as dim Eλ∗
with a prime p (resp. G = S1 × Γ with a finite group Γ). By these, under Hypothesis 3.13,
2 dim Eλ∗
Theorem 3.16, 3.17 may be directly transformed two results about bifurcation information of
(3.14) from (λ∗, θ). We here omit them.
Remark 3.18. (i) If n = dim Ω = 1, and F,G are defined by (1.3) under Hypothesis F2,N , then
they can satisfy Hypothesis 3.13 on W m,2
(Ω, RN ); see [39, 48]. Hence the last three theorems
may be applied in this case.
0
(ii) The proof key of [1, Theorem 7.12] is to use the (local) center manifold theorem instead of
the Lyapunov-Schmidt reduction. This method was firstly used by Chow and Lauterbach [18] in
66
the non-equivariant case. Our potential operators are neither strictly Fr´echlet differentiable at θ
nor of class C 1. Hence the center manifold theorem seems unable to be used in our situation.
But, from the constructions of center manifolds by Vanderbauwhede and Iooss [64] this method is
also possible if the restrictions of our potential operators to a Banach space X continuously and
densely embedding in H are of class C 1 as in the framework of [39, 40]; see [46].
The bifurcations in the previous theorems are all from a trivial critical orbit. Finally, let us give
a result about bifurcations starting a nontrivial critical orbit.
Hypothesis 3.19. Under Hypothesis 2.21, let for some x0 ∈ O the pair (L◦expN O(ε)x0
, NO(ε)x0)
satisfy the corresponding conditions with Hypothesis 1.1 with X = H. Let G ∈ C 1(H, R) be G-
invariant, have a critical orbit O, and also satisfy:
(i) the gradient ∇G is Gateaux differentiable near O, and every derivative G′′(u) is also a compact
linear operator;
(ii) G′′ are continuous at each point u ∈ O.
The assumptions on G assure that the functional L − λG, λ ∈ R, also satisfy the conditions of
Theorems 2.22, 2.23.
Let λ∗ be an eigenvalue of
L′′(x0)v − λG′′(x0)v = 0,
v ∈ Tx0H,
and dim Ker(L′′(x0) − λ∗G′′(x0)) > dimO.
We say O to be a bifurcation G-orbit with parameter λ∗ of the equation
L′(u) = λG′(u),
u ∈ H
(3.49)
(3.50)
if for any ε > 0 and neighborhood U of O in H there exists a solution G-orbit O′ 6= O in U of
(3.50) with some λ ∈ (−ε, ε).
Note that the orthogonal complementary of Tx0O in Tx0H, NOx0, is an invariant subspace
of L′′(x0) and G′′(x0). Let L′′(x0)⊥ (resp. G′′(x0)⊥) denote the restriction self-adjoint opera-
tor of L′′(x0) (resp. G′′(x0)) from NOx0 to itself. Then L′′(x0)⊥ = d2(L ◦ expN O(ε)x0
)(θ)
and G′′(x0)⊥ = d2(G ◦ expN O(ε)x0
)(θ). Suppose that L′′(x0)⊥ is invertible, or equivalently
Ker(L′′(x0)) = Tx0O. Then 0 is not an eigenvalue of
L′′(x0)⊥v − λG′′(x0)⊥v = 0,
v ∈ NOx0,
(3.51)
and λ ∈ R \ {0} is an eigenvalue of (3.51) if and only if 1/λ is an eigenvalue of compact lin-
ear self-adjoint operator Lx0 := [L′′(x0)⊥]−1G′′(x0)⊥ ∈ Ls(NOx0). Hence σ(Lx0) \ {0} =
{1/λn}∞
x0 be the eigensub-
space corresponding to 1/λn for n ∈ N. Then NO0
n=1 ⊂ R with λn → 0, and each 1/λn has finite multiplicity. Let NOn
x0 = Ker(Lx0) = Ker(G′′(x0)⊥) and
NOn
x0 = Ker(I/λn − Lx0) = Ker(L′′(x0)⊥ − λnG′′(x0)⊥), n = 1, 2,··· ,
(3.52)
and NOx0 = ⊕∞
n=0NOn
x0 .
67
Theorem 3.20. Under Hypothesis 3.19, suppose that Ker(L′′(x0)) = Tx0O (so the operator
L′′(x0)⊥ is invertible) and λ∗ = λn0 for some n0 ∈ N. Then O is a bifurcation G-orbit with
parameter λ∗ of (3.50) if one of the following two conditions holds:
a) L′′(x0)⊥ is either positive definite or negative one, and
Cl(O; Z2) 6= Cl−νλ∗ (O; Z2)
(3.53)
x0 (is more than zero because O is a degenerate critical
for some l ∈ Z, where νλ∗ = dim NOn0
orbit of Lλ∗ by the assumption in Hypothesis 3.19);
b) each NOn
commutes with G′′(x0)⊥), and
x0 in (3.52) is an invariant subspace of L′′(x0)⊥ (e.g. these are true if L′′(x0)⊥
Cl−ν−
λ∗
(O; Z2) 6= Cl−ν+
λ∗
(O; Z2)
(3.54)
λ∗ (resp. ν−
for some l ∈ Z, where ν+
space of L′′(x0)⊥ on NOn0
x0 .
λ∗ ) is the dimension of the positive (resp. negative) definite
From the following proof it is easily seen that (3.54) may be replaced by (3.53) if we add a
condition "L′′(x0)⊥ is either positive definite or negative one on NOn0
x0 " in b).
Proof of Theorem 3.20.
some n0 ∈ N, and let νλ∗ be the nullity of Lλ∗ at O, i.e., νλ∗ = dim NOn0
Theorem 3.5 we have ε > 0 such that
Let µλ denote the Morse index of Lλ := L − λG at O, λ∗ = λn0 for
x0 . As in the proof of
(3.55)
(3.56)
(3.57)
µλ = Xλn<λ
if L′′(x0)⊥ is positive definite,
µλ = Xλn>λ
dim NOn
x0 =(cid:26) µλ∗,
µλ∗ + νλ∗,
dim NOn
x0 =(cid:26) µλ∗ + νλ∗,
µλ∗,
∀λ ∈ (λ∗ − 2ε, λ∗],
∀λ ∈ (λ∗, λ∗ + 2ε)
∀λ ∈ (λ∗ − 2ε, λ∗),
∀λ ∈ [λ∗, λ∗ + 2ε)
if L′′(x0)⊥ is negative definite, and
µλ =(cid:26) µλ∗ + ν−
λ∗,
µλ∗ + ν+
λ∗,
∀λ ∈ (λ∗ − 2ε, λ∗),
∀λ ∈ (λ∗, λ∗ + 2ε)
if b) holds.
Corresponding to Claim 1 in the proof of Theorem 3.5 we may also prove
Claim 1. After shrinking ε > 0, if {(κn, vn)}n≥1 ⊂ [λ∗ − ε, λ∗ + ε] × NO(ε) satisfies
∇Lκn(vn) → θ and κn → κ0, then {vn}n≥1 has a convergent subsequence in NO(ε).
By a contradiction, assume that O is not a bifurcation G-orbit with parameter λ∗ of (3.50).
Then we have δ ∈ (0, ε] such that for each λ ∈ [λ∗ − δ, λ∗ + δ], O is an unique critical orbit of
Lλ in NO(δ). By [14, Theorem 5.1.21] (or as in [17]) we deduce
C∗(Lλ′,O; K) = C∗(Lλ′′,O; K),
∀λ′, λ′′ ∈ [λ∗ − δ, λ∗ + δ].
(3.58)
68
Since O is a nondegenerate critical orbit of Lλ for each λ ∈ [λ∗ − δ, λ∗ + δ]\{λ∗}, as in the proof
of (2.90) we derive from (2.99) that
C∗(Lλ′,O; Z2) = C∗−µλ′ (O; Z2)
and C∗(Lλ′′,O; Z2) = C∗−µλ′′ (O; Z2)
(3.59)
for any λ′ ∈ [λ∗ − δ, λ∗) any λ′′ ∈ (λ∗, λ∗ + δ].
If L′′(x0)⊥ is positive definite, by (3.53), (3.55) and (3.59) we deduce
Cl+µλ∗ (Lλ′,O; Z2) = Cl(O; Z2) 6= Cl−νλ∗ (O; Z2) = Cl+µλ∗ (Lλ′′,O; Z2)
for any λ′ ∈ [λ∗ − δ, λ∗) any λ′′ ∈ (λ∗, λ∗ + δ]. This contradicts (3.58).
Similarly, if L′′(x0)⊥ is negative definite, we deduce
Cl+µλ∗ (Lλ′′,O; Z2) = Cl(O; Z2) 6= Cl−νλ∗ (O; Z2) = Cl+µλ∗ (Lλ′,O; Z2)
for any λ′ ∈ [λ∗ − δ, λ∗) any λ′′ ∈ (λ∗, λ∗ + δ], and also arrive at a contradiction to (3.58).
If b) holds, by (3.56), (3.59) and (3.54) we have
Cl+µλ∗ (Lλ′,O; Z2) = Cl−ν−
λ∗
(O; Z2) 6= Cl−ν+
λ∗
(O; Z2) = Cl+µλ∗ (Lλ′′,O; Z2)
for any λ′ ∈ [λ∗ − δ, λ∗) any λ′′ ∈ (λ∗, λ∗ + δ], which contradicts (3.58).
✷
Using Theorem 2.29 many results above can be generalized the case of bifurcations at a non-
trivial critical orbit.
Part II
Applications to quasi-linear elliptic systems of
higher order
4 Fundamental analytic properties for functionals F and F
4.1 Results and preliminaries
A bounded domain Ω in Rn is said to be a Sobolev domain if for each integer 0 ≤ k ≤ m − 1 the
Sobolev space embeddings hold for it, i.e.,
W m,p(Ω) ֒→ W k,q(Ω)
W m,p(Ω) ֒→֒→ W k,q(Ω)
W m,p(Ω) ֒→֒→ W k,q(Ω)
W m,p(Ω) ֒→֒→ C k,σ(Ω)
if
if
if
if
1
q ≥
1
q
>
1
p −
1
p −
1
p
m − k
> 0,
n
m − k
> 0,
n
m − k
=
q < ∞,
n
p
n
< m − (k + σ),
,
0 ≤ σ < 1,
where ֒→֒→ denotes the compact embedding. Each bounded domain Ω in Rn with suitable smooth
boundary ∂Ω is a Sobolev domain.
The key result of this section is the following theorem. Most claims of it come from the auxil-
69
iary theorem 16 in Section 3.4 of Chapter 3 in [58] or Lemma 3.2 on the page 112 of [60]. Since
only partial claims were proved with V = W m,p
proof of it.
(Ω) therein, for completeness we give a detailed
0
Theorem 4.1. Let Ω ⊂ Rn be a bounded Sobolev domain, p ∈ [2,∞) and let V be a closed
subspace of W m,p(Ω). Suppose that (i)-(ii) in Hypothesis fp hold. Then we have
A). On V the functional F in (1.8) is bounded on any bounded subset, of class C 1, and the
derivative F ′(u) of F at u is given by
hF ′(u), vi = Xα≤mZΩ
fα(x, u(x),··· , Dmu(x))Dαvdx,
∀v ∈ V.
(4.1)
Moreover, the map u → F ′(u) also maps bounded subset into bounded ones.
B). The map F ′ is of class C 1 on V if p > 2, Gateaux differentiable on V if p = 2, and for each
u ∈ V the derivative DF ′(u) ∈ L (V, V ∗) is given by
hDF ′(u)v, ϕi = Xα,β≤mZΩ
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx.
(4.2)
(In the case p = 2, equivalently, the gradient map of F , V ∋ u 7→ ∇F(u) ∈ V , given by
(∇F(u), v)m,2 = hF ′(u), vi ∀v ∈ V,
(4.3)
has a Gateaux derivative D(∇F)(u) ∈ Ls(V ) at every u ∈ V .) Moreover, DF ′ also satisfies the
following properties:
(i) For every given R > 0, {DF ′(u)kukm,p ≤ R} is bounded in Ls(V ). Consequently, when
p = 2, F is on V of class C 2−0.
(ii) For any v ∈ V , un → u0 implies DF ′(un)v → DF ′(u0)v in V ∗.
(iii) If p = 2 and f (x, ξ) is independent of all variables ξα, α = m, then V ∋ u 7→ DF ′(¯u) ∈
L (V, V ∗) is continuous, i.e., F is of class C 2, and D(∇F)(u) : V → V is completely
continuous for each u ∈ V .
In addition, if (iii) in Hypothesis fp is also satisfied, we further have
C). F ′ satisfies condition (S)+.
70
D). Suppose p = 2. For u ∈ V , let D(∇F)(u), P (u) and Q(u) be operators in L (V ) defined by
(D(∇F)(u)v, ϕ)m,2 = Xα,β≤mZΩ
(P (u)v, ϕ)m,2 = Xα=β=mZΩ
+ Xα≤m−1ZΩ
(Q(u)v, ϕ)m,2 = Xα+β<2mZΩ
− Xα≤m−1ZΩ
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx,
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx
Dαv · Dαϕdx,
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx
Dαv · Dαϕdx,
(4.4)
(4.5)
(4.6)
respectively. Then D(∇F) = P + Q, and
(i) for any v ∈ V , the map V ∋ u 7→ P (u)v ∈ W m,2(Ω) is continuous;
(ii) for every given R > 0 there exist positive constants C(R, n, m, Ω) such that
(P (u)v, v)m,2 ≥ Ckvk2
m,2
∀v ∈ V
if u ∈ W m,2(Ω) satisfies kukm,2 ≤ R;
(iii) V ∋ u 7→ Q(u) ∈ L (V ) is continuous, and Q(u) : V → V is completely continuous for
each u;
(iv) for every given R > 0 there exist positive constants Cj(R, n, m, Ω), j = 1, 2 such that
(D(∇F)(u)v, v)m,2 ≥ C1kvk2
m,2 − C2kvk2
m−1,2
∀v ∈ V
if u ∈ V satisfies kukm,2 ≤ R.
This is a special case of the following result.
Theorem 4.2. Let Ω ⊂ Rn and p ∈ [2,∞) be as in Theorem 4.1, N ≥ 1 an integer, and V a closed
subspace of W m,p(Ω, RN ). Suppose that (i)-(ii) in Hypothesis Fp,N hold. Then corresponding
conclusions to A) and B) in Theorem 4.1 are also true if the letters u, v,F therein are replaced by
~u, ~v, F, respectively, and (4.1) -- (4.2) are changed into
hF′(~u), ~vi =
F i
α(x, ~u(x),··· , Dm~u(x))Dαvidx,
∀~v ∈ V,
(4.7)
hDF′(~u)~v, ~ϕi =
F ij
αβ(x, ~u(x),··· , Dm~u(x))Dβvj · Dαϕidx.
NXi=1 Xα≤mZΩ
NXi,j=1 Xα,β≤mZΩ
Moreover, if (iii) in Hypothesis Fp,N is also satisfied, then corresponding conclusions to C) and
D) in Theorem 4.1 still remain true if the letters u, v,F therein are replaced by ~u, ~v, F, respectively,
and (4.4) -- (4.6) are changed into
71
F ij
αβ(x, ~u(x),··· , Dm~u(x))Dβvj · Dαϕidx,
F ij
αβ(x, ~u(x),··· , Dm~u(x))Dβvj · Dαϕidx
+
Dαvi · Dαϕidx,
(D(∇F)(~u)~v, ~ϕ)m,2 =
(P (~u)~v, ~ϕ)m,2 =
(Q(~u)~v, ~ϕ)m,2 =
NXi,j=1 Xα,β≤mZΩ
NXi,j=1 Xα=β=mZΩ
NXi=1 Xα≤m−1ZΩ
NXi,j=1 Xα+β<2mZΩ
NXi=1 Xα≤m−1ZΩ
−
F ij
αβ(x, ~u(x),··· , Dm~u(x))Dβvj · Dαϕidx
Dαvi · Dαϕidx.
Theorem 4.2 can be proved as that of Theorem 4.1, only more terms are added or estimated
in each step. For the sake of simplicity, we only prove Theorem 4.1. To this goal the following
preliminary results are needed.
Proposition 4.3. For the function g1 in Hypothesis, let continuous positive nondecreasing func-
tions gk : [0,∞) → R, k = 3, 4, 5, be given by
g3(t) := 1 + g1(t)[t2M (m) + t(M (m) + 1)2] + g1(t)t(M (m) + 1) + g1(t)(M (m) + 1)2,
g4(t) := g1(t)t + g1(t)
and
g5(t) := (M (m) + 1)g1(t)(t + 1).
Then (ii) in Hypothesis fp implies that for all (x, ξ),
f (x, ξ) ≤ f (x, 0) + ξ◦ Xα<m−n/p
+g3(ξ◦) 1 + Xm−n/p≤α≤m
fα(x, 0) + Xm−n/p≤α≤m
ξαpα!,
fα(x, 0)qα
(4.8)
ξγpγ!pαβ
fα(x, ξ) ≤ fα(x, 0) + g4(ξ◦) Xβ<m−n/p 1 + Xm−n/p≤γ≤m
ξγpγ!pαβ
+ g4(ξ◦) Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
fα(x, ξ) ≤ fα(x, 0) + g5(ξ◦) 1 + Xm−n/p≤γ≤m
ξβ;
(4.9)
ξγpγ!,
(4.10)
for the latter we further have
72
if α < m − n/p, and
fα(x, ξ) ≤ fα(x, 0) + g5(ξ◦) + g5(ξ◦) Xm−n/p≤γ≤m
ξγpγ!1/qα
(4.11)
if m − n/p ≤ α ≤ m.
Its proof will be given in Appendix A. The following standard result concerning the continuity
of the Nemytski operator (cf. [7, Lemma 3.2] and [60, Proposition 1.1, page 3]) will be used many
times.
Proposition 4.4. Let G be a measurable set of positive measure in RN and let f : G × RN → R
satisfy the following conditions:
(a) f (x, ξ1,··· , ξN ) is continuous in (ξ1,··· , ξN ) for almost all x ∈ G;
(b) f (x, ξ1,··· , ξN ) is measurable in x for any fixed (ξ1,··· , ξN ) ∈ RN ;
(c) there exist positive numbers C, 1 < p, p1,··· , pN < ∞ and a function g ∈ Lp(G) such that
f (x, ξ1,··· , ξN ) ≤ C
NXi=1
ξi
pi
p + g(x),
∀(x, ξ) ∈ Ω × RN .
Then the Nemytskii operator F :QN
i=1 Lpi(G) → Lp(G) defined by the formula
F (u1,··· , uN )(x) = f (x, u1(x),··· , uN (x))
is bounded (i.e. mapping bounded sets into bounded sets) and continuous.
The following basic inequalities are standard.
Lemma 4.5. (i) There exists a positive constant C only depending on p ≥ 2 such that
Z 1
0
(1 + ta + (1 − t)b)p−2dt ≥ C(1 + a + b)p−2,
∀a, b ∈ R.
(ii) (x1 + ··· + xn)q ≤ x1q + ··· + xnq for any q ∈ (0, 1) and numbers xj , j = 1,··· .
We shall prove Theorem 4.1 in four subsections. Clearly, it suffices to prove the case V =
W m,p(Ω).
4.2 Proof for A) of Theorem 4.1
Step 1. Prove the continuity of F .
By (4.8) it is easy to see that the functional F in (1.8) is well-defined on W m,p(Ω) and is
bounded on any bounded subset of W m,p(Ω).
73
Next, we prove that F is continuous at a fixed u0 ∈ W m,p(Ω). For δ > 0 let B(u0, δ) = {u ∈
W m,p(Ω)ku − u0km,p ≤ δ}. By the Sobolev embedding theorem there exist R = R(u0) > 0
such that sup{Dαu(x) : α < m − n/p, x ∈ Ω} ≤ R for all u ∈ B(u0, 1). Take a continuous
function χ : R → R such that χ(t) = t ∀t ≤ 2R, χ(t) = ±3R ∀ ± t ≥ 3R, and χ(t) ≤ 3R.
Define a function f : Ω × RM (m) → R, (x, ξ) 7→ f (x, ξ) = f (x, ξ), where ξα = χ(ξα) if
α < m − n/p, and ξα = ξα if m − n/p ≤ α ≤ m. Then 0 = 0, ξ◦ ≤ 3M (m)R. It follows
that f also satisfies the Caratheodory condition, and therefore (4.8) leads to
f (x, ξ) ≤ f (x, 0) + 3M (m)R Xα<m−n/p
+g3(3M (m)R) 1 + Xm−n/p≤α≤m
fα(x, 0) + Xm−n/p≤α≤m
ξαpα!.
fα(x, 0)qα
(4.12)
Let the functional F : W m,p(Ω) → R be defined by
F(u) =ZΩ
f (x, u(x),··· , Dmu(x))dx.
Clearly, it is equal to F on ball B(u0, 1). Hence we only need to prove that F is continuous at u0.
This can follow from (4.12) and Proposition 4.4.
Step 2. Prove the C 1-smoothness of F .
a.a. x ∈ Ω, using the intermediate value theorem and (4.10)-(4.11) we deduce
Fix u, ϕ ∈ W m,p(Ω). Let Cu,ϕ = supk<m− n
p
[kukC k + kϕkC k ]. For any t ∈ [−1, 1] \ {0} and
≤ sup
+ sup
1
t
fα(x, u(x) + ϑtϕ(x),··· , Dmu(x) + ϑtDmϕ(x)) · Dαϕ(x)
fα(x, u(x) + ϑtϕ(x),··· , Dmu(x) + ϑtDmϕ(x)) · Dαϕ(x)
(cid:12)(cid:12)(cid:12)(cid:12)
[f (x, u(x) + tϕ(x),··· , Dmu(x) + tDmϕ(x)) − f (x, u(x),··· , Dmu(x))](cid:12)(cid:12)(cid:12)(cid:12)
0≤ϑ≤1 Xα<m−n/p
0≤ϑ≤1 Xm−n/p≤α≤m
0≤ϑ≤1 Xα<m−n/p"fα(x, 0) + g5(Cu,ϕ)(cid:18)1 + Xm−n/p≤γ≤m
0≤ϑ≤1 Xm−n/p≤α≤m(cid:20)fα(x, 0) + g5(Cu,ϕ)
+g5(Cu,ϕ)(cid:18) Xm−n/p≤γ≤m
Dγu(x) + ϑtDγϕ(x)pγ(cid:19)1/qα(cid:21) · Dαϕ(x)
Dγu(x) + ϑtDγϕ(x)pγ(cid:19)#
≤ Cu,ϕ sup
+ sup
≤ Cu,ϕ Xα<m−n/p(cid:20)fα(x, 0) + g5(Cu,ϕ)
+g5(Cu,ϕ)(cid:18) Xm−n/p≤γ≤m
+ Xm−n/p≤α≤m(cid:2)fα(x, 0) + g5(Cu,ϕ)(cid:3) · Dαϕ(x)
+g5(Cu,ϕ) Xm−n/p≤α≤m(cid:18) Xm−n/p≤γ≤m
2pγ(cid:2)Dγu(x)pγ + Dγϕ(x)pγ(cid:3)(cid:19)(cid:21)
2pγ(cid:2)Dγu(x)pγ + Dγϕ(x)pγ(cid:3)(cid:19)1/qα
· Dαϕ(x).
74
It follows from the assumptions on pαβ that the right side is integrable and thus from the Lebesgue
dominated convergence theorem that the functional F is Gateaux differentiable. Moreover, the
Gateaux differential of F at u, DF(u) ∈ [W m,p(Ω)]∗, is given by
fα(x, u(x),··· , Dmu(x))Dαϕ(x)dx.
DF(u)ϕ = hDF(u), ϕi = Xα≤mZΩ
Let DαF(u) ∈ [W m,p(Ω)]∗ be defined by
hDαF(u), ϕi =ZΩ
fα(x, u(x),··· , Dmu(x))Dαϕ(x)dx.
(4.13)
Claim 1. The map DαF : W m,p(Ω) → [W m,p(Ω)]∗ is continuous.
• Case α < m − n/p.
Then kDαϕkC 0 ≤ Ckϕkm,p, ∀ϕ ∈ W m,p(Ω), where C > 0 is a constant coming from the
Sobolev embedding theorem. Fix u ∈ W m,p(Ω). For any v ∈ B(u, 1), we have
kDαF(v) − DαF(u)k = sup
kϕkm,p=1hDαF(v) − DαF(u), ϕi
≤
sup
kϕkm,p=1ZΩ fα(x, v(x),··· , Dmv(x)) − fα(x, u(x),··· , Dmu(x)) · Dαϕ(x)dx
kϕkm,p=1kDαϕkC 0ZΩ fα(x, v(x),··· , Dmv(x)) − fα(x, u(x),··· , Dmu(x))dx
≤ CZΩ fα(x, v(x),··· , Dmv(x)) − fα(x, u(x),··· , Dmu(x))dx.
sup
≤
(4.14)
Because of (4.10), by a standard method as in Step 1 we may also assume that for some constant
C ′ > 0,
Note that fα(·, 0) ∈ L1(Ω). Using Proposition 4.4 we deduce that
fα(x, ξ) ≤ fα(x, 0) + C ′(cid:18)1 + Xm−n/p≤γ≤m
L1(Ω) × Ym− n
p ≤β≤m
Yβ<n/p
ξγpγ(cid:19),
∀(x, ξ).
Lpβ (Ω) → L1(Ω), u = {uβ : β ≤ m} → fα(·, u)
is continuous, which implies the continuity of the map W m,p(Ω) ∋ u 7→ fα(·, u,··· , Dmu) ∈
L1(Ω). The latter claim and (4.14) yields that kDαF(v) − DαF(u)k → 0 as kv − ukm,p → 0.
That is, DαF is continuous in this case.
• Case m − n/p ≤ α ≤ m. Then we have
kDαF(u) − DαF(v)k = sup
kϕkm,p=1hDαF(u) − DαF(v), ϕi
≤
sup
kϕkm,p=1ZΩ fα(x, u(x),··· , Dmu(x)) − fα(x, v(x),··· , Dmv(x)) · Dαϕ(x)dx
kϕkm,p=1kDαϕkpα(cid:18)ZΩ fα(x, u(x),··· , Dmu(x)) − fα(x, v(x),··· , Dmv(x))qα dx(cid:19)1/qα
≤ C(cid:18)ZΩ fα(x, u(x),··· , Dmu(x)) − fα(x, v(x),··· , Dmv(x))qαdx(cid:19)1/qα
sup
≤
.
Since v ∈ B(u, 1), as in Step 1, by (4.11) we may also assume that for some constant C > 0,
75
ξγpγ /qα(cid:19),
∀(x, ξ).
fα(x, ξ) ≤ fα(x, 0) + C(cid:18)1 + Xm−n/p≤γ≤m
L1(Ω) × Ym− n
p ≤β≤m
Yβ<n/p
Note that fα(·, 0) ∈ Lqα(Ω) with qα = pα
pα−1 . We derive from Proposition 4.4 that
Lpβ (Ω) → Lqα(Ω), u = {uβ : β ≤ m} → fα(·, u)
is continuous. This leads to the continuity of the maps W m,p(Ω) ∋ u 7→ fα(·, u,··· , Dmu) ∈
Lqα(Ω) and hence DαF as above.
To sum up, the map DF : W m,p(Ω) → [W m,p(Ω)]∗ is continuous. As usual this implies that
F has the Fr´echet derivative F ′(u) = DF(u) at each point u ∈ W m,p(Ω) and thus is of class C 1.
Moreover, from the above proof we see that for any u ∈ W m,p(Ω),
kDαF(u)k = sup
kϕkm,p=1hDαF(u), ϕi ≤ CZΩ fα(x, u(x),··· , Dmu(x))dx
if α < m − n/p, and
kϕkm,p=1hDαF(u), ϕi ≤ C(cid:18)ZΩ fα(x, u(x),··· , Dmu(x))qαdx(cid:19)1/qα
kDαF(u)k = sup
if m − n/p ≤ α ≤ m. It follows from these and (4.10)-(4.11) that F ′ maps a bounded subset
into a bounded set.
4.3 Proof for B) of Theorem 4.1
Step 1. Prove that the right side of (4.2) determines an operator in L (W m,p(Ω), [W m,p(Ω)]∗).
By (1.5) we deduce that the right side of (4.2) satisfies
(cid:12)(cid:12)(cid:12)(cid:12) Xα,β≤mZΩ
sup
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx(cid:12)(cid:12)(cid:12)(cid:12)
g1(kukC k ) Xα,β≤mZΩ(cid:18)1 + Xm−n/p≤γ≤m
k<m−n/p
≤
Dγu(x)pγ(cid:19)pαβ
Dβv(x) · Dαϕ(x)dx.
It suffices to prove that there exists a constant C = C(u, α, β) such that
Dβv(x) · Dαϕ(x)dx ≤ Ckvkm,pkϕkm,p.
Dγu(x)pγ(cid:19)pαβ
Case α = β = m. Then pαβ = 1 − 1
Dγu(x)pγ(cid:19)pαβ
Dγu(x)pγ(cid:19)(cid:19)pαβ(cid:18)ZΩ Dβv(x)pβ(cid:19)1/pβ(cid:18)ZΩ Dαϕ(x)pα(cid:19)1/pα
ZΩ(cid:18)1 + Xm−n/p≤γ≤m
ZΩ(cid:18)1 + Xm−n/p≤γ≤m
≤ (cid:18)ZΩ(cid:18)1 + Xm−n/p≤γ≤m
pα − 1
Dβv(x) · Dαϕ(x)dx
and thus the H older inequality leads to
(4.15)
pβ
≤ Ckvkm,pkϕkm,p.
76
pα
and supx Dβv(x) ≤
C(m, n, p)kvkm,p. It follows from these that
Case m − n/p ≤ α ≤ m,
β < m − n/p. Then pαβ = 1 − 1
Dγu(x)pγ(cid:19)pαβ
ZΩ(cid:18)1 + Xm−n/p≤γ≤m
≤ C(m, n, p)kvkm,pZΩ(cid:18)1 + Xm−n/p≤γ≤m
≤ C(m, n, p)kvkm,p(cid:18)ZΩ(cid:18)1 + Xm−n/p≤γ≤m
Dβv(x) · Dαϕ(x)dx
Dγu(x)pγ(cid:19)pαβ
Dγu(x)pγ(cid:19)(cid:19)pαβ(cid:18)ZΩ Dαϕ(x)pα(cid:19)1/pα
· Dαϕ(x)dx
≤ Ckvkm,pkϕkm,p.
Case m − n/p ≤ β ≤ m, α < m − n/p. The proof is the same as the last case.
Case α,β < m − n/p. We have pαβ = 1, supx Dβv(x) ≤ C(m, n, p)kvkm,p and
supx Dαϕ(x) ≤ C(m, n, p)kϕkm,p. Hence
pαβ
Dγu(x)pγ
ZΩ1 + Xm−n/p≤γ≤m
≤ C(m, n, p)kvkm,pkϕkm,pZΩ1 + Xm−n/p≤γ≤m
≤ Ckvkm,pkϕkm,p.
Dβv(x) · Dαϕ(x)dx
Dγu(x)pγ dx
+ 1
pβ
+ 1
pα
determined by qαβ + 1
pαβ
pα − 1
= 1. Using the H older inequality we get
Case α, β ≥ m − n/p, α + β < 2m. Then 0 < pαβ < 1 − 1
ZΩ(cid:18)1 + Xm−n/p≤γ≤m
≤ Ω1/qαβ(cid:18)ZΩ(cid:18)1 + Xm−n/p≤γ≤m
Dγu(x)pγ(cid:19)pαβ
Dβv(x) · Dαϕ(x)dx
Dγu(x)pγ(cid:19)(cid:19)pαβ(cid:18)ZΩ Dβv(x)pβ(cid:19)1/pβ(cid:18)ZΩ Dαϕ(x)pα(cid:19)1/pα
. Let qαβ > 1 be
pβ
≤ Ckvkm,pkϕkm,p.
In summary, (4.15) is proved. Hence the right side of (4.2) determines an operator A(u) ∈
L (W m,p(Ω), [W m,p(Ω)]∗) by
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx.
(4.16)
hA(u)v, ϕi = Xα,β≤mZΩ
hAα(u)v, ϕi = Xβ≤mZΩ
In particular, each termPβ≤mRΩ fαβ(x, u(x),··· , Dmu(x))Dβ v · Dαϕdx determines an op-
erator Aα(u) ∈ L (W m,p(Ω), [W m,p(Ω)]∗) by
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx.
(4.17)
1
t
h
= ZΩ
0
− Xβ≤mZΩ
= Xβ≤mZΩZ 1
− Xβ≤mZΩ
= Xβ≤mZΩZ 1
= Xβ≤m
Iαβ.
0
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx
fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))Dβv · Dαϕdsdx
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx
[fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))]Dβv · Dαϕdsdx
Step 2. Prove that the map DαF : W m,p(Ω) → [W m,p(Ω)]∗ defined by (4.13) is of class C 1 for
p > 2 or p = 2 and α < m, but only Gateaux differentiable for p = 2 and α = m.
For any t ∈ [−1, 1] \ {0} and v ∈ W m,p(Ω) we derive from (4.13) and (4.17) that
77
[DαF(u + tv) − DαF(u)], ϕi − hAα(u)v, ϕi
1
[fα(x, u(x) + tv(x),··· , Dmu(x) + tDmv(x)) − fα(x, u(x),··· , Dmu(x))]Dαϕ(x)dx
t
(4.18)
pα − 1
pβ ∈ (0, 1) and we obtain
Firstly, we consider the case p > 2.
• Case α = β = m. Then pαβ = 1 − 1
Iαβ ≤ Z 1
0
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ(cid:18)ZΩ Dβvpβ(cid:19)1/pβ(cid:18)ZΩ Dαϕpα(cid:19)1/pα
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
and supx Dβv(x) ≤
pα
.
≤ Ckvkm,pkϕkm,pZ 1
0
C(m, n, p)kvkm,p. We can also deduce
• Case m − n/p ≤ α ≤ m, β < m − n/p. Then pαβ = 1 − 1
Iαβ ≤ Ckvkm,pZ 1
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
0
≤ Ckvkm,pkϕkm,pZ 1
0
−fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ(cid:18)ZΩ Dαϕpα(cid:19)1/pα
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
.
78
C(m, n, p)kϕkm,p. We can also deduce
• Case m − n/p ≤ β ≤ m, α < m − n/p. Then pβα = 1 − 1
Iαβ ≤ Ckϕkm,pZ 1
0
pβ
and supx Dαϕ(x) ≤
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))1/pβα(cid:17)pβα(cid:18)ZΩ Dβvpβ(cid:19)1/pα
−fαβ(x, u(x),··· , Dmu(x))1/pβα(cid:17)pβα
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
.
≤ Ckvkm,pkϕkm,pZ 1
0
pα − 1
pβ
. Let qαβ > 1
+ 1
pβ
+ 1
pα
= 1. Then
be determined by qαβ + 1
pαβ
dsZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
• Case α, β ≥ m − n/p, α + β < 2m. Then 0 < pαβ < 1 − 1
Z 1
≤ Ω1/qαβZ 1
−fαβ(x, u(x),··· , Dmu(x)) · Dβv · Dαϕdsdx
0
0
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ(cid:18)ZΩ Dβvpβ(cid:19)1/pβ(cid:18)ZΩ Dαϕpα(cid:19)1/pα
≤ Ckvkm,pkϕkm,pZ 1
0
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
.
• Case α < m−n/p, β < m−n/p. Then pβα = 1, supx Dβv(x) ≤ C(m, n, p)kϕkm,p
and supx Dαϕ(x) ≤ C(m, n, p)kϕkm,p. We can also deduce
Iαβ ≤ Ckvkm,pkϕkm,pZ 1
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
0
−fαβ(x, u(x),··· , Dmu(x))(cid:17).
Summarizing the above five cases, by (4.18) we obtain that for any t ∈ [−1, 1] \ {0},
k[DαF(u + tv) − DαF(u)]/t − Aα(u)vk
≤ Ckvkm,p Xβ≤mZ 1
0
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
.
(4.19)
Fix u, v ∈ W m,p(Ω). Because of (1.5), after treating as in Step 1 of Section 4.2 we assume
that for some constant C = Cu,v > 0 and all (x, ξ),
fαβ(x, ξ) ≤ C(cid:18)1 + Xm−n/p≤γ≤m
ξγpγ(cid:19)pαβ
≤ C(cid:18)1 + Xm−n/p≤γ≤m
ξγpγpαβ(cid:19).
(4.20)
79
Then we derive from Proposition 4.4 that
Yβ<n/p
L1(Ω) × Ym− n
p ≤γ≤m
Lpγ (Ω) → L1/pαβ (Ω), u = {uγ : γ ≤ m} → fαβ(·, u)
is continuous. This implies the continuity of the map
W m,p(Ω) ∋ w 7→ fαβ(·, w,··· , Dmw) ∈ L1/pαβ (Ω).
It follows from (4.19) that
t→0k[DαF(u + tv) − DαF(u)]/t − Aα(u)vk = 0,
lim
(4.21)
Namely, DαF has Gateaux derivative Aα(u) at u.
If v is allowed to varies in the ball B(u, 1) ⊂ W m,p(Ω), then we may assume that (4.20) also
holds for another constant C = Cu,1 > 0, and thus that C in (4.19) may be changed into Cu,1.
Taking t = 1 and letting kvkm,p → 0 in (4.19) we get that
kDαF(u + v) − DαF(u) − Aα(u)vk = o(kvkm,p).
(4.22)
That is, DαF has Fr´echet derivative Aα(u) at u.
Moreover, using a similar method to the above one we can prove: for any u, v ∈ W m,p(Ω),
we have C = C(m, n, p, Ω) > 0 such that
kAα(u) − Aα(v)k
≤ C Xβ≤m(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x)) − fαβ(x, v(x),··· , Dmv(x))1/pαβ(cid:17)pαβ
.
(4.23)
This shows that Aα is continuous. Hence for p > 2 we have proved:
A is continuous, and F is of class C 2 on W m,p(Ω).
pα − 1
pβ
Next, we consider the case p = 2.
If α + β < 2m the above arguments also work, and so Aα is of class C 1.
If α = β = m, then pαβ = 1 − 1
dsZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
Z 1
≤ Z 1
−fαβ(x, u(x),··· , Dmu(x)) · Dβv · Dαϕdsdx
0
0
= 0 (since pα = pβ = 2). We can only obtain
≤ Ckϕkm,2Z 1
0
−fαβ(x, u(x),··· , Dmu(x))2Dβv2(cid:17)1/2(cid:18)ZΩ Dαϕ2(cid:19)1/2
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))2Dβv2(cid:17)1/2
.
80
And therefore
≤ CZ 1
0
k[DαF(u + tv) − DαF(u)]/t − Aα(u)vk
ds(cid:16)ZΩ fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x))
−fαβ(x, u(x),··· , Dmu(x))2Dβv2(cid:17)1/2
.
Note that for fixed u, v ∈ W m,p(Ω) and all s, t ∈ [0, 1] the functions
fαβ(x, u(x) + stv(x),··· , Dmu(x) + stDmv(x)) − fαβ(x, u(x),··· , Dmu(x))2
are uniformly bounded. It follows
t→0k[DαF(u + tv) − DαF(u)]/t − Aα(u)vk = 0.
lim
Hence DαF has Gateaux derivative Aα(u) at u.
Step 3. Prove that DF ′ is bounded on any ball in W m,p(Ω). From the above arguments we
obtain C = C(m, n, p, Ω) such that
for α + β < 2m or p > 2 and α = β = m, and
≤ Ckvkm,pk · ϕkm,p(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
hAα(u)v, ϕi = (cid:12)(cid:12)(cid:12)(cid:12)ZΩ
hAα(u)v, ϕi = (cid:12)(cid:12)(cid:12)(cid:12)ZΩ
fαβ(x, u(x),··· , Dmu(x))DβvDαϕ(x)dx(cid:12)(cid:12)(cid:12)(cid:12)
fαβ(x, u(x),··· , Dmu(x))DβvDαϕ(x)dx(cid:12)(cid:12)(cid:12)(cid:12)
≤ C · ϕkm,p(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x))2Dβv2(cid:17)1/2
(4.24)
(4.25)
for p = 2 and α = β = m.
For (4.25), by (1.5) we derive
and hence
fαβ(x, u(x),··· , Dmu(x)) ≤ sup
k<m−n/2
g1(kukC k ) ∀u ∈ W m,2,
hAα(u)v, ϕi = (cid:12)(cid:12)(cid:12)(cid:12)ZΩ
fαβ(x, u(x),··· , Dmu(x))Dβvdx(cid:12)(cid:12)(cid:12)(cid:12)
g1(kukC k )(cid:16)ZΩ Dβv2(cid:17)1/2
≤ C sup
k<m−n/2
This and (4.24), (1.5) yield the desired claim.
Step 4. Prove (ii) in B). By (4.2) we derive
= Xα,β≤mZΩ
hϕ, [DF ′(uk)]∗v − [DF ′(u)]∗vi = hDF ′(uk)ϕ − DF ′(u)ϕ, vi
[fαβ(x, uk(x),··· , Dmuk(x))
−fαβ(x, u(x),··· , Dmu(x))]Dβϕ · Dαvdx.
Let uk → u in W m,p(Ω). If p > 2, with the same reasoning as in (4.24) we can derive from
this that
81
hϕ, [DF ′(uk)]∗v − [DF ′(u)]∗vi
≤ Xα,β≤mZΩ fαβ(x, uk(x),··· , Dmuk(x))
≤ Ckvkm,pkϕkm,p Xα,β≤m(cid:16)ZΩ fαβ(x, uk(x),··· , Dmuk(x))
−fαβ(x, u(x),··· , Dmu(x)) · Dβϕ · Dαvdx
Because of (1.5), as treated in Step 1 of Section 4.2 we deduce
−fαβ(x, u(x),··· , Dmu(x))1/pαβ dx(cid:17)pαβ
Xα,β≤m(cid:16)ZΩ fαβ(x, uk(x),··· , Dmuk(x)) − fαβ(x, u(x),··· , Dmu(x))1/pαβ dx(cid:17)pαβ
.
→ 0
and so k[DF ′(uk)]∗v − [DF ′(u)]∗vk → 0 as k → ∞.
For p = 2 we can use (4.25) to arrive at the same conclusion.
Step 5. The proof of (iii) in B) is the same as that of (iii) in D) later on.
4.4 Proof for C) of Theorem 4.1
Let uj ⇀ u in W m,p(Ω) and satisfy limj→∞hF ′(uj), uj − ui ≤ 0, i.e.,
limj→∞ Xα≤mZΩ
fα(x, uj(x),··· , Dmuj(x))Dα(uj − u)dx = 0.
(4.26)
By Sobolev embedding theorem we have strong convergence
Dαuj → Dαu in Lq(Ω) if q <
Dαuj → Dαu in C 0(Ω) if α < m − n/p.
np
n − p
, m − n/p ≤ α ≤ m − 1,
Since supj kujkm,p < ∞, C = sup{kukC k + kujkC k : k < m − n/p, j ∈ N} < ∞. These,
(4.10) and (4.11) lead to
Xα<m−n/pZΩ fα(x, uj(x),··· , Dmuj(x))Dα(uj − u)dx
≤ Xα<m−n/pZΩ fα(x, 0) · Dα(uj − u)dx + g5(C) Xα<m−n/pZΩ Dα(uj − u)dx
+g5(C) Xα<m−n/p Xm−n/p≤γ≤mZΩ Dγujpγ · Dα(uj − u)dx → 0
82
and
≤
Xm−n/p≤α≤m−1ZΩ fα(x, uj(x),··· , Dmuj(x))Dα(uj − u)dx
Xm−n/p≤α≤m−1ZΩ fα(x, 0) · Dα(uj − u)dx + g5(C) Xm−n/p≤α≤m−1ZΩ Dα(uj − u)dx
+g5(C) Xm−n/p≤α≤m−1ZΩ Xm−n/p≤γ≤m
Dγujpγ!1/qα
· Dα(uj − u)dx → 0.
Here the final limit is because Lemma 4.5(ii) implies that
· Dα(uj − u)dx → 0
Dγujpγ!1/qα
ZΩ Xm−n/p≤γ≤m
≤ Xm−n/p≤γ≤mZΩ Dγujpγ /qα · Dα(uj − u)dx → 0
≤ Xm−n/p≤γ≤m(cid:18)ZΩ Dγujpγ dx(cid:19)1/qα(cid:18)ZΩ Dα(uj − u)pαdx(cid:19)1/pα
→ 0.
It follows that
Xα≤m−1ZΩ fα(x, uj(x),··· , Dmuj(x))Dα(uj − u)dx → 0
as j → ∞.
(4.27)
This and (4.26) yield
Xα=mZΩ
Next, we claim
fα(x, uj(x),··· , Dmuj(x))Dα(uj − u)dx → 0 as j → ∞.
(4.28)
Xα=mZΩ fα(x, uj (x),··· , Dm−1uj(x), Dmu(x))Dα(uj − u)dx
≤ Xα=mZΩ(cid:12)(cid:12)(cid:12)[fα(x, uj(x),··· , Dm−1uj(x), Dmu(x))
−fα(x, u(x),··· , Dm−1u(x), Dmu(x))]Dα(uj − u)(cid:12)(cid:12)(cid:12)dx
+ Xα=mZΩ fα(x, u(x),··· , Dm−1u(x), Dmu(x))Dα(uj − u)dx
= I1,j + I2,j → 0.
(4.29)
Indeed, by (4.11) it is easily checked that fα(x, u(x),··· , Dm−1u(x), Dmu(x)) belongs to Lqα(Ω)
for α ≥ m − n/p. Since uj ⇀ u in W m,p(Ω) it follows from that I2,j → 0 as j → ∞.
83
By the H older inequality we have
I1,j ≤ Xα=m(cid:16)ZΩ fα(x, uj(x),··· , Dm−1uj(x), Dmu(x))
−fα(x, u(x),··· , Dm−1u(x), Dmu(x))qαdx(cid:17)1/qα(cid:16)ZΩ Dα(uj − u)pαdx(cid:17)1/pα
.
As before, using (4.11) and Proposition 4.4 it is easy to derive
(cid:16)ZΩ fα(x, uj(x),··· , Dm−1uj(x), Dmu(x))
−fα(x, u(x),··· , Dm−1u(x), Dmu(x))qαdx(cid:17)1/qα
→ 0.
Note that the sequence(cid:16)RΩ Dα(uj − u)pαdx(cid:17)1/pα
Hence (4.28) and (4.29) yield
is bounded. We get I1,j → 0.
Xα=mZΩ
[fα(x, uj (x),··· , Dm−1uj(x), Dmuj(x))
−fα(x, uj(x),··· , Dm−1uj(x), Dmu(x))]Dα(uj − u)dx → 0.
(4.30)
Note that the integrand in each term is non-negative, which may be derived from the mean value
theorem and (1.6) as seen below. This implies that for any subset E ⊂ Ω,
λj(E) :=ZE Xα=m
[fα(x, uj(x),··· , Dm−1uj(x), Dmuj(x))
−fα(x, uj (x),··· , Dm−1uj(x), Dmu(x))]Dα(uj − u)dx → 0 (4.31)
as j → ∞. We claim that for E ⊂ Ω,
meas(E)→0ZE Xα=m
lim
Dαuj(x)pdx = 0
(4.32)
uniformly with respect to j. In fact, by (4.31) we have with vj = uj − u,
λj(E) = Xα=β=mZEZ 1
0
fαβ(cid:0)x, uj(x),··· , Dm−1uj(x), Dmu(x) + sDmvj (x)(cid:1) Dβvj · Dαvjdsdx.
84
From (1.6) and Lemma 4.5(i) we deduce
0
0
Xα=β=mZEZ 1
≥ ZEZ 1
g2( Xk<m−n/p
≥ g2( Xk<m−n/p
≥ g2( Xk<m−n/p
≥ g2( Xk<m−n/p
≥ g2( Xk<m−n/p
fαβ(x, uj (x),··· , Dm−1uj(x), Dmu(x) + sDmvj(x))Dβ vj · Dαvj dsdx
kujkC k )(cid:18)1 + Xγ=m
Dγu(x) + sDγvj(x))(cid:19)p−2 Xα=m
Dαvj2dsdx
kujkC k )ZEZ 1
0 (cid:0)1 + Dγu(x) + sDγvj(x))(cid:1)p−2 Xα=m
kujkC k )ZE(cid:0)1 + Dγu(x) + Dγuj(x)(cid:1)p−2 Xα=m
kujkC k )ZE(cid:0)1 + Dγvj(x)(cid:1)p−2
kujkC k )ZE Dγvjpdx
Dγvj2dx
Dαvj2dsdx
Dαvj2dx
(4.33)
for each γ of length m. It follows
Xγ=mZE Dγvjpdx ≤
≤
M0(m)
λj(E)
g2(Pk<m−n/p kujkC k )
g2(supjPk<m−n/p kujkC k )
M0(m)
λj(E).
(4.34)
The final inequality is because uj ⇀ u and thus that supj kujkm,p < ∞, which implies that
supjPk<m−n/p kujkC k < ∞. Moreover
(cid:18) Xγ=mZE Dγujpdx(cid:19)1/p
≤(cid:18) Xγ=mZE Dγupdx(cid:19)1/p
+(cid:18) Xγ=mZE Dγvjpdx(cid:19)1/p
.
Given any ε > 0, from (4.31) and (4.34) there exist j0 ∈ N such that
(cid:18) Xγ=mZE Dγvjpdx(cid:19)1/p
< εp/2
for all j ≥ j0 and all E ⊂ Ω. Using the absolute continuity of the integral we have δ > 0 such
that for any E ⊂ Ω with mes(E) < δ,
(cid:18) Xγ=mZE Dγupdx(cid:19)1/p
These lead to
+(cid:18) Xγ=mZE Dγvjpdx(cid:19)1/p
Xγ=mZE Dγujpdx < ε
for all j ∈ N and all E ⊂ Ω with mes(E) < δ. (4.32) is proved.
< εp/2, j = 1,··· , j0.
Next, by (4.34) and (4.31), for any given σ > 0 we have
85
σpmes({Dαvj ≥ σ}) ≤ Z{Dαvj ≥σ} Dαvjpdx ≤ZΩ Dαvjpdx
≤
M0(m)
g2(supjPk<m−n/p kujkC k )
λj(Ω) → 0
as j → ∞. This means that the sequence Dαuj converges to Dαu in measure for α = m.
Combing with (4.32) we obtain that Dαuj → Dαu in Lp(Ω) for α = m. Moreover, uj ⇀ u
implies that uj → u in W m−1,p(Ω). Hence kuj − ukm,p → 0 as j → ∞.
✷
4.5 Proof for D) of Theorem 4.1
Step 1. (i) of D) can be proved as in (ii) of B).
Step 2.
Prove (ii) of D). By Sobolev embedding theorem
sup{kDγukC k : k < m − n/2, u ∈ W m,2(Ω), kukm,2 ≤ R} ∈ (0,∞).
Let C be equal to the value of g2 at this number. We derive from (1.6) that
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαvdx
(P (u)v, v)m,2 = Xα=β=mZΩ
+ Xα≤m−1ZΩ
≥ C Xα=mZΩ Dαv2dx + Xα≤m−1ZΩ
Dαv · Dαvdx
≥ min{C, 1}kvk2
m,2
for any v ∈ W m,2(Ω).
Dαv · Dαvdx
Step 3. Prove (iii) of D). As in the proof of (4.24) we can obtain C = C(m, n, p, Ω) > 0 with
p = 2 such that
([Q(u) − Q(¯u)]v, ϕ)m,2
[fαβ(x, u(x),··· , Dmu(x)) − fαβ(x, ¯u(x),··· , Dm ¯u(x))]Dβ vDαϕ(x)dx(cid:12)(cid:12)(cid:12)(cid:12)
= (cid:12)(cid:12)(cid:12)(cid:12) Xα+β<2mZΩ
≤ C Xα+β<2m(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x)) − fαβ(x, ¯u(x),··· , Dm ¯u(x))1/pαβ(cid:17)pαβ
×
×kvkm,pk · ϕkm,p.
So it follows from (1.5) and Proposition 4.4 that
kQ(u) − Q(¯u)kL (W m,2(Ω))
≤ C Xα+β<2m(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x)) − fαβ(x, ¯u(x),··· , Dm ¯u(x))1/pαβ(cid:17)pαβ
→ 0.
To prove the second claim let us decompose Q(u) into Q1(u) + Q2(u) + Q3(u), where
86
(Q1(u)v, ϕ)m,2 =
(Q2(u)v, ϕ)m,2 =
(Q3(u)v, ϕ)m,2 =
Xα≤m−1,β≤m−1ZΩ
− Xα≤m−1ZΩ
Xα=m,β≤m−1ZΩ
Xα≤m−1,β=mZΩ
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx
Dαv · Dαϕdx,
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx,
fαβ(x, u(x),··· , Dmu(x))Dβv · Dαϕdx.
Clearly, (Q2(u)v, ϕ)m,2 = (Q3(u)ϕ, v)m,2, that is, they are adjoint each other. Let vj ⇀ v in
W m,2(Ω). By the proof of (4.24) we can get C = C(m, n, p, Ω) > 0 with p = 2 such that
(Q1(u)(vj − v), ϕ)m,2
≤
fαβ(x, u(x),··· , Dmu(x))Dβ(vj − v)Dαϕ(x)dx(cid:12)(cid:12)(cid:12)(cid:12)
Xα≤m−1,β≤m−1(cid:12)(cid:12)(cid:12)(cid:12)ZΩ
+ Xα≤m−1ZΩ Dα(vj − v) · Dαϕdx
≤ Ckvj − vkm−1,2 · kϕkm−1,2 Xα≤m−1,β≤m−1(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
+kvj − vkm−1,2 · kϕkm−1,2
and hence
kQ1(u)(vj − v)km,2
≤ Ckvj − vkm−1,2 Xα≤m−1,β≤m−1(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
+kvj − vkm−1,2 → 0
because kvj − vkm−1,2 → 0 by the compactness of the embedding W m,2(Ω) ֒→ W m−1,2(Ω).
Similarly, we have
(Q2(u)(vj − v), ϕ)m,2
≤
Xα=m,β≤m−1(cid:12)(cid:12)(cid:12)(cid:12)ZΩ
≤ Ckvj − vkm−1,2 · kϕkm,2 Xα=m,β≤m−1(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
fαβ(x, u(x),··· , Dmu(x))Dβ(vj − v)Dαϕ(x)dx(cid:12)(cid:12)(cid:12)(cid:12)
and hence
kQ2(u)(vj − v)km,2
≤ Ckvj − vkm−1,2 Xα=m,β≤m−1(cid:16)ZΩ fαβ(x, u(x),··· , Dmu(x))1/pαβ(cid:17)pαβ
→ 0.
Since Q3(u) is the adjoint operator of Q2(u), it is completely continuous too.
87
Step 4. Prove (iv) of D). By the arguments in Step 3 we see: for every given R > 0 there exist
positive constants C(R, n, m, Ω) such that if u ∈ W m,2(Ω) satisfies kukm,2 ≤ R then
∀v, ϕ ∈ W m,2(Ω),
(Q1(u)v, ϕ)m,2 ≤ Ckvkm−1,2 · kϕkm−1,2
(Q2(u)v, ϕ)m,2 ≤ Ckvkm−1,2 · kϕkm,2
(Q3(u)v, ϕ)m,2 ≤ Ckvkm,2 · kϕkm−1,2
∀v, ϕ ∈ W m,2(Ω),
∀v, ϕ ∈ W m,2(Ω)
and therefore
(Q(u)v, v)m,2 ≤ 3 Ckvkm−1,2 · kvkm,2
For the constant C in (ii) of D), using the inequality ab ≤ 1
we derive with ε = C/(3 C),
∀v ∈ W m,2(Ω).
2ε a2 + ε
2 b2 for any ε > 0 and a, b ≥ 0,
(Q(u)v, v)m,2 ≤ 3 Ckvkm−1,2 · kvkm,2 ≤
for all v ∈ W m,2(Ω). Taking C1 = C/2 and C2 = 9 C 2
C
2 kvk2
m,2 +
9 C 2
2C kvk2
m−1,2
2C , this and (ii) of D) give the desired result.
5 (PS)- and (C)-conditions
A C 1 functional ϕ on a Banach-Finsler manifold M is said to satisfy the Palais-Smale condition
at the level c ∈ R ((P S)c-condition, for short) if every sequence {xj}j≥1 ⊂ X such that ϕ(xj ) →
c ∈ R and ϕ′(xj) → 0 in X ∗ has a convergent subsequence in M. When ϕ satisfies the (P S)c-
condition at every level c ∈ R we say that it satisfies the Palais-Smale condition ((P S)-condition,
for short).
When M a Banach space there is weaker condition. Call a C 1 functional ϕ on a Banach space
X to satisfy the Cerami condition at the level c ∈ R ((C)c-condition, for short) if every sequence
{xj}j≥1 ⊂ X such that ϕ(xj) → c ∈ R and (1 + kxjk)ϕ′(xj) → 0 in X ∗ has a convergent
subsequence in X. When ϕ satisfies the (C)c-condition at every level c ∈ R we say that it satisfies
the Cerami condition ((C)-condition, for short).
Actually, if a C 1 functional ϕ on a Banach space X is bounded below, Caklovic, Li and Willem
[10] showed that ϕ satisfies the (P S)-condition if and only if it is coercive. It was further proved in
[52, Proposition 5.23] that ϕ satisfies the (P S)-condition if and only if it does the (C)-condition.
Recently, it was proved in [62, Theorem 6] that if a continuous functional ϕ on X is Gateaux
differentiable and satisfies the weak Palais-Smale condition then ϕ is coercive provided {x ∈
X ϕ(x) = c} is bounded for some c ∈ R.
Theorem 5.1. Let Ω ⊂ Rn, N ∈ N, p ∈ [2,∞) and V ⊂ W m,p(Ω, RN ) be as in Theorem 4.2.
Suppose that Hypothesis Fp,N hold and that F is coercive, i.e., F(~u) → ∞ as k~ukV → ∞. Then
F satisfies the (PS)- and (C)-conditions on V . In particular, the same conclusions hold with the
functional F on V ⊂ W m,p(Ω) under the condition fp.
88
Proof. Since F is coercive, it is bounded below. By [52, Proposition 5.23] it suffices to prove that
F satisfies the (PS)-condition.
Let a sequence {~uj}j≥1 such that F(~uj) → c ∈ R and F′(~uj) → 0 as j → ∞. Since
F is coercive, the sequence {~uj}j≥1 must be bounded. Note that V is a self-reflexive Banach
space. After passing to a subsequence we may assume ~uj ⇀ ~u in V . Moreover, F′(~uj) → 0
implies limj→∞hF′(~uj), ~uj − ~ui = 0. By Theorem 4.2 (the corresponding conclusion to C) of
Theorem 4.1) we know that F′ is of class (S)+. Hence ~uj → ~u in V .
Clearly, the coercivity of F implies that it is bounded below. On the other hand, for a C 1
functional ϕ on a Banach space X which is bounded below, Li Shujie showed that it is coercive if
ϕ satisfies the (P S)-condition.
There exist some explicit conditions on F under which F is coercive on W m,p
0
(Ω, RN ), for
example, there exist some two positive constants c0, c1 such that
F (x, ξ) ≥ c0
NXi=1 Xα=m
ξi
αp − c1 ∀(x, ξ).
The coercivity requirement is too strong. In fact, the proof of Theorem 5.1 shows that under Hy-
pothesis Fp,N we only need to add some conditions so that
sup
j
F(~uj) < ∞ and F′(~uj) → 0 =⇒ sup
j k~ujkm,p < ∞.
Theorem 5.2. Under Hypothesis Fp,N , suppose that there exist κ ∈ R and Υ ∈ L1(Ω) such that
F (x, ξ) − κ
NXi=1 Xα≤m
F i
α(x, ξ)ξi
α ≥ c0
NXi=1 Xα=m
ξi
αp − c1
NXi=1
ξi
0p − Υ(x) ∀(x, ξ),
where c0 > 0 and c0 − c1Sm,p > 0 for the best constant Sm,p > 0 with
ZΩ updx ≤ Sm,pZΩ Dmupdx = Sm,p Xα=mZΩ Dαup ∀u ∈ W m,p
0
(Ω).
Then F satisfies the (PS)- and (C)-conditions on W m,p
0
(Ω, RN ).
Proof. Let {~uk}k≥1 ⊂ W m,p
and F′(~uk) → 0 (resp. (1 + k~ukk)F′(~uk) → 0). By (4.7) the latter means
(Ω, RN ) be a sequence such that F(~uk) ≤ M ∀k for some M > 0,
0
(cid:12)(cid:12)(cid:12)PN
i=1Pα≤mRΩ F i
(resp. (cid:12)(cid:12)(cid:12)PN
i=1Pα≤mRΩ F i
α(x, ~uk(x),··· , Dm~uk(x))Dαui
α(x, ~uk(x),··· , Dm~uk(x))Dαui
kdx(cid:12)(cid:12)(cid:12) ≤ εkk~ukkm,p
kdx(cid:12)(cid:12)(cid:12) ≤ εk
1+k~ukkm,p
k~ukkm,p
(5.1)
)
where εk → 0 and k~ukkm,p = kDm~ukkp as usual. By the assumption we have
89
F(~uk) − κ
NXi=1 Xα≤mZΩ
NXi=1 Xα=mZΩ Dαui
NXi=1ZΩ Dmui
kpdx − c1
kpdx − c1Sm,p
F i
α(x, ~uk(x),··· , Dm~uk(x))Dαui
kdx
Υ(x)dx
NXi=1ZΩ ui
NXi=1ZΩ Dmui
kpdx −ZΩ
kpdx −ZΩ
Υ(x)dx
≥ c0
≥ c0
and therefore
i=1RΩ Dmui
(c0 − c1Sm,p)PN
(resp. (c0 − c1Sm,p)PN
i=1RΩ Dmui
kpdx ≤RΩ Υ(x)dx + M + κεkk~ukkm,p
kpdx ≤RΩ Υ(x)dx + M + κεk ).
This implies that k~ukkm,p is bounded. Passing to a subsequence if necessary, we may assume
~uk ⇀ ~u. The remainder is the same as that of Theorem 5.1.
Theorem 5.3. Let Ω ⊂ Rn be a bounded Sobolev domain. Suppose that Hypothesis F2,N is
satisfied with the constant function g2, and that for any (x, ξ) ∈ Ω ×Qm−1
RN ×M0(k),
k=0
F (x, ξ, 0) ≤ ϕ(x) + C
NXi=1 Xα≤m−1
ξi
αr,
where ϕ ∈ L1(Ω) and 1 ≤ r < 2. Then F satisfies the (PS)- and (C)-conditions on W m,2
Proof. For any x ∈ Ω and ξ = (ξ1,··· , ξm) ∈ Qm
Qm−1
RN ×M0(k). By the mean value theorem we get
RN ×M0(k) let ξ = (ξ1,··· , ξm−1) ∈
(Ω, RN ).
k=0
k=0
0
α(x, ξ)ξi
F i
α − [F (x, ξ) − F (x, ξ, 0)] − F (x, ξ, 0)
α(x, ξ)ξi
F i
α −
α(x, ξ, tξm)ξi
F i
αdt − F (x, ξ, 0)
F i
α(x, ξ)ξi
α − F (x, ξ)
NXi=1 Xα=m
NXi=1 Xα=m
NXi=1 Xα=m
NXi,j=1 Xα=β=mZ 1
NXi=1 Xα=m
1
2
g2
=
=
=
≥
NXi=1 Xα=mZ 1
0
0
0
dtZ 1
α2 − F (x, ξ, 0).
ξi
F ij
αβ(x, ξ, (t + s − st)ξm)(1 − t)ξi
αξj
βds − F (x, ξ, 0)
90
It follows that for any ~u ∈ W m,2
0
(Ω, RN ),
1
2
g2
NXi=1 Xα=mZΩ Dαui2 ≤ZΩ
F (x, ~u,··· , Dm−1~u, 0) − F(~u) + hF′(~u), ~ui.
By the assumption and the Young inequality we derive
ϕ(x) + C
NXi=1 Xα≤m−1ZΩ Dαuir
2 Dαui2 +
2 − r
2
ε−r/(2−r)(cid:19)
ϕ(x) + C
F (x, ~u,··· , Dm−1~u, 0) ≤ZΩ
NXi=1 Xα≤m−1ZΩ(cid:18)rε
ZΩ
≤ ZΩ
≤ ZΩ
ϕ(x) + Cεk~uk2
m,2 ≤ZΩ
ϕ(x) + Cεk~uk2
m,2 + Cε−r/(2−r)
g2k~uk2
and hence
1
2
m,2 + Cε−r/(2−r) − F(u) + hF′(~u), ~ui.
Taking ε = g2
4C leads to
g2(cid:19)r/(2−r)
ϕ(x) + C(cid:18) 4C
m,2 ≤ZΩ
g2k~uk2
(Ω, RN ). Let the sequence {~uk}k≥1 ⊂ W m,2
1
4
for any ~u ∈ W m,2
(Ω, RN ) such that supk F(~uk) ≤
M for some M > 0, and F′(~uk) → 0 (resp. (1 + k~ukkm,2)F′(~uk) → 0). Then (5.2) implies
(because of hF′(~u), ~ui ≤ kF′(~u)k · k~ukm,2 ≤ (1 + k~ukm,2)kF′(~u)k) that {~uk}k≥1 is bounded in
W m,2
(Ω, RN ) and thus has a convergent subsequence as above.
− F(~u) + hF′(~u), ~ui
(5.2)
0
0
0
6 Morse inequalities
Firstly, we show that Theorem 4.2 and Theorems 2.22, 2.23 imply the generalized Morse lemma.
Theorem 6.1. Let Ω ⊂ Rn be a bounded Sobolev domain, N ∈ N, and H a closed subspace
of W m,2(Ω, RN ). Let G be a compact Lie group which acts on H in a C 3-smooth isometric
way. Suppose that Hypothesis F2,N is satisfied and that the functional F given by (1.3) is G-
invariant. Let O be an isolated critical orbit of F (always understanding as FH ). It is a compact
C 3 submanifold, whose normal bundle NO has fiber at ~u ∈ O,
NO~u = {~v ∈ H (~v, ~w)m,2 = 0 ∀ ~w ∈ T~uO ⊂ H}.
Let N +O~u, N 0O~u and N −O~u be the positive definite, null and negative definite spaces of the
bounded linear self-adjoint operator associated with the bilinear form
NO~u × NO~u ∋ (~v, ~w) 7→
NXi=1 Xα,β≤mZΩ
F ij
αβ(x, ~u(x),··· , Dm~u(x))Dβvj · Dαwidx.
91
Then dim N 0O~u and dim N −O~u are finite and independent of choice of ~u ∈ O. They are called
nullity and Morse index of O, denoted by νO and µO, respectively. Moreover, the following holds.
(i) If νO = 0 (i.e., the critical orbit O is nondegenerate), there exist ǫ > 0 and a G-equivariant
homeomorphism onto an open neighborhood of the zero section preserving fibers
Φ : N +O(ǫ) ⊕ N −O(ǫ) → NO
such that for any ~u ∈ O and (~v+, ~v−) ∈ N +O(ǫ)~u × N −O(ǫ)~u,
F ◦ E ◦ Φ(~u, ~v+ + ~v−) = k~v+k2
where E : NO → H is given by E(~u, ~v) = ~u + ~v.
m,2 − k~v−k2
m,2 + FO,
(6.1)
(ii) If νO 6= 0 there exist ǫ > 0, a G-equivariant topological bundle morphism that preserves the
zero section,
h : N 0O(3ǫ) → N +O ⊕ N −O ⊂ H, (~u, ~v) 7→ h~u(~v),
and a G-equivariant homeomorphism onto an open neighborhood of the zero section pre-
serving fibers, Φ : N 0O(ǫ)⊕N +O(ǫ)⊕N −O(ǫ) → NO, such that the following properties
hold:
(ii.1) for any ~u ∈ O and (~v0, ~v+, ~v−) ∈ N 0O(ǫ)~u × N +O(ǫ)~u × N −O(ǫ)~u,
F ◦ E ◦ Φ(~u, ~v0, ~v+ + ~v−) = k~v+k2
m,2 − k~v−k2
m,2 + F(~u + ~v0 + h~u(~v0));
(ii.2) for each ~u ∈ O the function
N 0O(ǫ)~u → R, ~v 7→ F◦
is G~u-invariant, of class C 1, and satisfies
~u(~v) := F(~u + ~v + h~u(~v))
DF◦
~u(~v) ~w := (∇F(~u + ~v + h~u(~v)), ~w),
∀ ~w ∈ N 0O~u.
(6.2)
(6.3)
Proof. Since FH satisfies Hypothesis 1.1 with X = H around each critical point by Theorem 4.2,
it follows from Lemma 2.10 that for each ~u ∈ O the restriction FN O~u satisfies Hypothesis 1.1
with X = NO~u around the origin of NO~u. Note that the exponential map on H, exp : T H =
H × H → H, is given by exp(~u, ~v) = ~u + ~v. Then Theorems 2.22, 2.23 lead to the desired
conclusions immediately.
By Corollary 2.27 and (2.93), for any commutative ring K we get
Cq(F,O; K) ∼= ⊕q
j=0Cq−j−µO (F◦
~u, θ; K) ⊗ Hj(O; K) ∀q = 0, 1, 2,···
(6.4)
(6.5)
if ~u ∈ O, νO 6= 0 and O has trivial normal bundle; and C∗(F,O; Z2) ∼= C∗−µO (O; Z2),
(O; θ− ⊗ K)
if νO = 0, where µO is the Morse index of O and θ− is the orientation bundle of N −O.
C∗(F,O; K) ∼= C∗−µO (O; θ− ⊗ K)
G(F,O; K) ∼= H µO−1
and C ∗
G
From the second equality in (6.5) and [13, Chapter I, Theorem 7.6] we immediately arrive at
92
Theorem 6.2. Under the assumptions of Theorem 6.1, Let a < b be two regular values of F and
F−1([a, b]) contains only nondegenerate critical orbits Oj with Morse indexes µj , j = 1,··· , k.
Suppose that F satisfies the (P S)c condition for each c ∈ [a, b). (For example, this is true if either
F is coercive or one of Theorems 5.2, 5.3 holds in case H = W m,2
(Ω, RN ).) Then there exists a
0
polynomial with nonnegative integral coefficients Q(t) such that
∞Xi=0
kXj=1
rankH i
G(Oj , θ−
j ⊗ K)tµj +i =
rankH i
G(Fb, Fa; K)ti + (1 + t)Q(t),
(6.6)
∞Xi=0
j is the orientation bundle of N −Oj , j = 1,··· , k. In particular, if G is trivial and each
where θ−
Oj becomes a nondegenerate critical point ~uj , then the following Morse inequalities hold:
lXj=0
(−1)l−j Nj(a, b) ≥
lXj=0
(−1)l−jβj(a, b),
l = 0, 1,··· ,
(6.7)
where for each q ∈ N ∪ {0}, Nq(a, b) = ♯{1 ≤ i ≤ k µi = q} (the number of points in {~uj}k
with Morse index q) and
j=1
βq(a, b) =
rankHq(Jb, Ja; K).
kXi=1
Furthermore, if F is coercive, has only nondegenerate critical points, and for each q ∈ {0} ∪ N
there exist only finitely many critical points with Morse index q, then the following relations hold:
qXi=0
(−1)q−iNi ≥ (−1)q, q = 0, 1, 2,··· ,
and
∞Xi=0
(−1)iNi = 1,
(6.8)
where Ni is the number of critical points of F with Morse index i.
The proof of (6.8) is standard, see the proof of [4, Corollary 6.5.10]. When H = W m,2
0
(Ω)
and F is coercive, (6.7) was first obtained by Skrypnik in [58, §5.2].
7 Bifurcations for Quasi-linear elliptic systems
Hypothesis 7.1. Let Ω ⊂ Rn be a bounded Sobolev domain, N ∈ N, and functions
RN ×M0(k) → R
RN ×M0(k) → R and G : Ω ×
F : Ω ×
m−1Yk=0
mYk=0
satisfy Hypothesis Fp,N and (i)-(ii) in Hypothesis Fp,N , respectively. Let V be a closed subspace
of W m,p(Ω, RN ) containing W m,p
(Ω, RN ).
0
We consider (generalized) bifurcation solutions of the boundary value problem corresponding
to the subspace V :
Xα≤m
(−1)αDαF i
α(x, ~u,··· , Dm~u) = λ Xα≤m−1
i = 1,··· , N.
(−1)αDαGi
α(x, ~u,··· , Dm−1~u),
(7.1)
Call ~u ∈ V a generalized solution of (7.1) if it is a critical point on V of the variational problem
F(~u) − λG(~u) =ZΩ
F (x, ~u,··· , Dm~u)dx − λZΩ
93
G(x, ~u,··· , Dm−1~u)dx.
(7.2)
As a generalization of [60, Theorem 7.2, Chapter 4], we may derive from Theorem 4.2 and
Theorem 7.1 of [60, Chapter 4]:
Theorem 7.2. Under Hypothesis 7.1, assume
(i) the functionals F and G are even, F(θ) = G(θ) = 0, G(~u) 6= 0 ∀~u ∈ V \ {θ}, and G′(~u) 6=
θ ∀~u ∈ V \ {θ};
(ii) hF′(~u), ~ui ≥ ν(k~ukm,p), where ν(t) is a continuous function and positive for t > 0;
(iii) F(~u) → +∞ as k~ukm,p → ∞.
Then for any c > 0 there exists at least a sequence {(λj , ~uj)}j ⊂ R × {~u ∈ V F(~u) = c}
satisfying (7.1).
By Theorems 3.3, 4.2 we have
Theorem 7.3. Under Hypothesis 7.1 with p = 2, let ~u0 ∈ V satisfy F′(~u0) = 0 and G′(~u0) = 0.
If (λ∗, ~u0) with certain λ∗ ∈ R is a bifurcation point for (7.1), then the linear problem
NXi,j=1 Xα,β≤m
(−1)α(cid:2)F ij
NXi,j=1 Xα,β≤m−1
αβ(x, ~u0(x),··· , Dm~u0(x))Dβvj(cid:3)
(−1)α(cid:2)Gij
αβ(x, ~u0(x),··· , Dm−1~u0(x))Dβvj(cid:3)
= λ
(7.3)
with λ = λ∗ has a nontrivial solution in V , i.e., ~u0 is a degenerate critical point of the functional
F − λ∗G on V .
Conversely, if dim Ω = 1 and ~u0 is a degenerate critical point of the functional F− λ∗G, using
Theorem 3.4 we may obtain the corresponding bifurcation results. These will be given in more
general forms, see [48].
Hypothesis 7.4. Let Hypothesis 7.1 hold with p = 2, ~u0 ∈ V satisfy F′(~u0) = 0 and G′(~u0) = 0,
and the linear problem
NXi,j=1 Xα,β≤m
(−1)α(cid:2)F ij
αβ(x, ~u0(x),··· , Dm~u0(x))Dβvj(cid:3) = 0
(7.4)
have no nontrivial solutions in V .
94
The final condition in this hypothesis means that F′′(~u) has a bounded linear inverse. Under
Hypothesis 7.4, by the arguments above Theorem 3.5, the all eigenvalues of (7.3) form a discrete
subset of R, {λj}∞
j=1, which contains no zero and satisfies λj → ∞ as j → ∞; moreover, each
λj has finite multiplicity. Let Vj be the eigensubspace of (7.3) corresponding to the eigenvalue λj,
j = 1, 2,··· . By Theorems 3.5, 4.2 we directly obtain
Theorem 7.5. Under Hypothesis 7.4, for an eigenvalue λk of (7.3) as above, assume that one of
the following three conditions holds:
(a) F′′(~u0) is positive definite, i.e., for each v ∈ V \ {θ},
NXi,j=1 Xα,β≤mZΩ
F ij
αβ(x, ~u0(x),··· , Dm~u0(x))Dβvj(x)Dαvi(x)dx > 0;
(7.5)
(b) F′′(~u0) is negative definite, i.e., for each v ∈ V \ {θ},
NXi,j=1 Xα,β≤mZΩ
F ij
αβ(x, ~u0(x),··· , Dm~u0(x))Dβvj(x)Dαvi(x)dx < 0;
(7.6)
(c) each Vj is an invariant subspace of F′′(~u0) in V , j = 1, 2,··· , and either (7.5) holds for all
v ∈ Vk \ {θ}, or (7.6) does for all v ∈ Vk \ {θ}.
Then (λk, ~u0) ∈ R × V is a bifurcation point for the problem (7.1), and one of the following
alternatives occurs:
(i) (λk, ~u0) is not an isolated solution of (7.1) in {λk} × V .
(ii) there exists a sequence {κj}j≥1 ⊂ R \ {λk} such that κj → λk and that for each κj the
problem (7.1) with λ = κj has infinitely many solutions converging to ~u0 ∈ V .
(iii) for every λ in a small neighborhood of λk there is a nontrivial solution ~uλ of (7.1) converging
to ~u0 as λ → λk;
(iv) there is a one-sided Λ neighborhood of λk such that for any λ ∈ Λ \ {λk}, (7.1) has at least
two nontrivial solutions converging to ~u0 as λ → λk.
Remark 7.6. (i) When N = 1, V = W m,2
0
(Ω), u = θ and F also satisfies
(F′(u), u)m,2 ≥ ckuk2
m,2
(7.7)
for some c > 0 and all sufficiently small kukm,2, if λ∗ is an eigenvalue of (7.3) with u = θ, it was
proved in [59, Chap.1, Theorem 3.5] that (λ, θ) is a bifurcation point of (7.1). Since F(θ) = θ, it
is clear that (7.7) implies F′′(θ) to be positive definite. Hence Theorem 7.5 contains [59, Chap.1,
Theorem 3.5] as a special example.
95
(ii) When N = 1, n ≥ 3, V = H 1
0 and
2 ξ2
0 (Ω), G(x, ξ0,··· , ξn) = 1
aij(x, ξ0)ξiξj −Z ξ0
0
1
2
nXi,j=1
F (x, ξ0,··· , ξn) =
g(x, t)dt,
(7.8)
Canino [11, Theorem 1.3] obtained a corresponding result provided that functions aij = aji, g :
Ω × R → R satisfy the following assumptions:
a.0) aij is of class C 1, and is of class C 2 in ξ0 for a.e. x ∈ Ω;
a.1) there exists C > 0 such that for a.e. x ∈ Ω, for all ξ0 ∈ R and for all i, j, k,
aij(x, ξ0) ≤ C,
Dxk aij(x, ξ0) ≤ C,
Dξ0aij(x, ξ0) ≤ C,
ξ0ξ0aij(x, ξ0) ≤ C;
a.2) there exists ν > 0 such that for a.e. x ∈ Ω, for all ξi ∈ R, i = 0,··· , n,
D2
nXi,j=1
aij(x, ξ0)ξiξj ≥ ν
ξ2
i ;
nXı=1
a.3) for a.e. x ∈ Ω, for all ξi ∈ R, i = 0,··· , n,
nXi,j=1
sDξ0aij(x, ξ0)ξiξj ≥ 0;
g) for every ξ0 ∈ R, g(x, ξ0) is measurable with respect to x, for a.e. x ∈ Ω, g(x, ξ0) is of class
C 1 with respect to s, g(x, 0) = 0; moreover, there exist b ∈ R and 0 < p < 4/(n − 2) such that,
for a.e. x ∈ Ω and all s ∈ R,
Dξ0 g(x, ξ0) ≤ b(1 + ξ0p).
It is easily checked that F in (7.8) satisfies Hypothesis F2,1. Thus [11, Theorem 1.3] is implied
in Theorem 7.5 with N = 1 and ~u = θ.
By Theorems 3.7, 3.9 we deduce
Theorem 7.7. Under Hypothesis 7.1 with p = 2, let G be a compact Lie group acting on V in a
C 3-smooth and isometric (and so orthogonal) way. Suppose that both F and G are G-invariant,
and that ~u0 ∈ Fix(G) satisfies F′(~u0) = 0 and G′(~u0) = 0. For an eigenvalue λk of (7.3)
as above, assume that one of the three conditions (a),(b) and (c) in Theorem 7.5 holds. Then
(λk, ~u0) ∈ R × V is a bifurcation point for the equation (3.14), and if dim Vk ≥ 2 and the unit
sphere in Vk is not a G-orbit we must get one of the following alternatives:
(i) (λk, ~u0) is not an isolated solution of (7.1) in {λk} × V ;
(ii) there exists a sequence {κj}j≥1 ⊂ R \ {λk} such that κj → λk and that for each κj the
problem (7.1) with λ = κj has infinitely many G-orbits of solutions converging to ~u0 ∈ V ;
96
(iii) for every λ in a small neighborhood of λk there is a nontrivial solution ~uλ of (7.1) converging
to ~u0 as λ → λk;
(iv) there is a one-sided Λ neighborhood of λk such that for any λ ∈ Λ \ {λk}, (7.1) has at least
two nontrivial critical orbits converging to ~u0 as λ → λk.
Furthermore, if the Lie group G is equal to Z2 or S1, then the above (iii)-(iv) can be replaced by
the following
(iii') there exist left and right neighborhoods Λ− and Λ+ of λk in R and integers n+, n− ≥ 0,
such that n+ + n− ≥ dim V and for λ ∈ Λ−\{λ∗} (resp. λ ∈ Λ+\{λ∗}), (7.1) has at least
n− (resp. n+) distinct critical G-orbits different from ~u0, which converge to ~u0 as λ → λk.
The corresponding claims to 3) of Theorem 3.7 can be easily written.
If ~u0 /∈ Fix(G), Theorem 3.20 may yield a result. If n = dim Ω = 1, by Theorems 3.14, 3.16,3.17
we can also obtain more results. They will be given in [46, 48].
Example
i=1(0, Ti) let C m(Qn
For 0 < Ti < ∞, i = 1,··· , n, and Ω := Qn
R/(TiZ), RN ) be the set
of all ~u ∈ C m(Rn, RN ) which are Ti-periodic with respect to the i-th variable, i = 1,··· , n.
It may be viewed as a subspace of W m,2(Ω, RN ). Denote by W m,2(Qn
R/(TiZ), RN ) the
closure of C m(Qn
(Ω, RN ) is contained in
W m,2(Qn
R/(TiZ), RN ). The compact connected Lie group G =Qn
R/(TiZ), which is iso-
morphic to Tn = Rn/Zn, acts on W m,2(Qn
R/(TiZ), RN ) via the following isometric linear
R/(TiZ), RN ) in W m,2(Ω, RN ). Clearly, W m,2
representation:
i=1
i=1
i=1
0
i=1
i=1
i=1
([t1,··· , tn] · ~u)(x1,··· , xn) = (u1(x1 + t1),··· , un(xn + tn))
(7.9)
R/(TiZ), RN ). The set of fixed points
of this action, Fix(G), consist of all constant vector functions from Ω to RN . Under Hypothesis 7.1
i=1
for [t1,··· , tn] ∈ G and ~u = (u1,··· , un) ∈ W m,2(Qn
with Ω =Qn
i=1(0, Ti) ⊂ Rn, assume also that F (x, ξ) and G(x, ξ) satisfy
F (x1,··· , xi,··· , xn, ξ) = F (x1,··· , xi,··· , xn, ξ),
G(x1,··· , xi,··· , xn, ξ) = G(x1,··· , xi,··· , xn, ξ),
k=0
Rn ×Qm
some T s, 1 ≤ s ≤ n. Clearly, if some ~u ∈ W m,2(Qn
where xi = 0 and xi = Ti, i = 1,··· , n. In other words, F and G may be viewed as functions on
RN ×M0(k) with period Ti in variables xi, i = 1,··· , n. Then the functionals F and G
are G-invariant, and every critical orbit different from points in Fix(G) must be homeomorphic to
R/(TiZ), RN ) are constant with respect
to variables xir , r = 1,··· , k < n, but not with respect to any other variable xi, then the orbit
G(~u) of ~u is homeomorphic to some T k. However, it is possible that the orbit of ~u is of T k-type
even if ~u is not constant with respect to each variable, see the proof of [65, Proposition 3.4].
i=1
97
Clearly, Hypothesis 7.4 is satisfied for each constant map ~u0 : Ω → RN . So Theorem 7.5, and
Theorem 7.7 with G = T n can be directly applied. By Theorem 3.7 we also get
i=1
Theorem 7.8. Let ~u0 : Ω → RN be a constant map. Assume that F and G satisfy the conditions of
Theorem 7.5 with V = W m,2(Qn
R/(TiZ), RN ). Then (λk, ~u0) ∈ R × V is a bifurcation point
for the equation (7.1). Moreover, if Vk (the eigensubspace of (7.3) corresponding to the eigenvalue
λk) satisfies one of the following assumptions: A) Fix(T n) ∩ Vk = {θ}, B) every orbit in Vk is
homeomorphic to some T s for s ≥ 2; then either one of the above (i)-(iii) in Theorem 7.7 with
G = T n or the following hold:
(iv)' there is a one-sided Λ neighborhood of λk such that for any λ ∈ Λ \ {λk}, the problem (7.1)
with λ = λk has at least dim Vk (resp. 2 dim Vk) nontrivial critical orbits in the case A) (resp.
B)), where every orbit is counted with its multiplicity (defined by [68, Definition 1.3]).
The theory in this paper provides necessary tools for generalizing [39, 65] to the functional
considered in the above example. They will be investigated in the latter paper.
Finally, we present a result associated with Theorems 5.4.2, 5.7.4 in [26].
Theorem 7.9. Let Ω ⊂ Rn be a bounded Sobolev domain, N ∈ N. Suppose that
Ω ×
mYk=0
RN ×M0(k) × [0, 1] ∋ (x, ξ, λ) 7→ F (x, ξ; λ) ∈ R
is differentiable with respect to λ, and satisfies the following conditions:
(i) All F (·; λ) satisfy Hypothesis F2,N uniformly with respect to λ ∈ [0, 1], i.e., the inequalities
(1.1) and (1.2) are uniformly satisfied for all λ ∈ [0, 1].
(ii) If qα = 1 for α < m − n/2, and qα = 2α/(2α − 1) for m − n/2 ≤ α ≤ m, then
sup
α≤m
sup
1≤i≤N
sup
λ ZΩ(cid:2)DλF (x, 0; λ) + DλF i
α(x, 0; λ)qα(cid:3) dx < ∞.
(iii) For all i = 1,··· , N and α ≤ m,
ξk
α(x, 0; λ)
DλF i
α(x, ξ; λ) ≤ DλF i
NXk=1
+g(
NXk=1
0) Xβ<m−n/2(cid:18)1 +
NXl=1 Xm−n/2≤β≤m(cid:18)1 +
NXk=1 Xm−n/2≤γ≤m
γ2γ(cid:19)2αβ
ξk
γ2γ(cid:19)2αβ
NXk=1 Xm−n/2≤γ≤m
ξk
ξk
0)
+g(
ξl
β;
where g : [0,∞) → R is a continuous, positive, nondecreasing function.
Let V be a closed subspace of W m,2(Ω, RN ), and for each λ ∈ [0, 1] let ~uλ be a critical point of
the functional
Fλ(~u) =ZΩ
F (x, ~u,··· , Dm~u; λ)dx
98
on V . Suppose that [0, 1] ∋ λ 7→ ~uλ ∈ V is continuous. Then one of the following alternatives
occurs:
(I) There exists certain λ0 ∈ [0, 1] such that (λ0, ~uλ0) is a bifurcation point of ∇Fλ(~u) = 0.
(II) Each ~uλ is an isolated critical point of Fλ and C∗(Fλ, ~uλ; K) = C∗(F0, ~u0; K) for all λ ∈
[0, 1]; moreover ~uλ is a local minimizer of Fλ if and only if ~u0 is a local minimizer of F0.
If DλF (·; λ) uniformly satisfy the inequalities (1.1) and (1.2) for all λ ∈ [0, 1]. Then these and
(ii) can yield (iii).
Proof. Suppose that (I) does not hold. Then each ~uλ is an isolated critical point of Fλ. Since
[0, 1] ∋ λ 7→ ~uλ ∈ V is continuous, we may find a bounded open subset O in V such that ~uλ is a
unique critical point of Fλ contained in the closure O of O. Take R > 0 such that O ⊂ BV (θ, R).
As in the proof of (4.8) we may derive from (iii) that with qα in (ii),
we have a constant C = C(m, n, N, R) > 0 such that
α(x, 0; λ)
DλF i
NXl=1 Xm−n/2≤α≤m
DλF (x, ξ; λ) ≤ DλF (x, 0; λ) +(cid:16) NXk=1
0(cid:17) NXi=1 Xα<m−n/2
ξk
0)(cid:18)1 +
NXk=1
NXi=1 Xm−n/2≤α≤m
α(x, 0; λ)qα +bg(
ξk
DλF i
+
α2α(cid:19)
ξl
for all (x, ξ, λ) and some continuous, positive, nondecreasing functionbg : [0,∞) → R. As before
Dα~u(x) x ∈ Ω(cid:27) < C for all ~u ∈ W m,2(Ω, RN ) with k~ukm,2 ≤ R.
sup(cid:26) Xα<m−n/2
Fλ1(~u) − Fλ2(~u) ≤ λ2 − λ1ZΩ
≤ λ2 − λ1(cid:20)sup
NXi=1 Xm−n/2≤α≤m
+bg(C)ZΩ(cid:18)1 +
λ DλF (x, ~u,··· , Dm~u; λ)dx
λ ZΩ DλF i
λ ZΩ DλF i
NXl=1 Xm−n/2≤α≤m
α(x, 0; λ)qα dx
Dαul2α(cid:19)dx(cid:21).
λ ZΩ DλF (x, 0; λ)dx + C
It follows that for any λi ∈ [0, 1], i = 1, 2,
This implies that [0, 1] ∋ λ 7→ Fλ is continuous in C 0( ¯BV (θ, R)). Similarly, (ii) -- (iii) yield the
continuity of the map [0, 1] ∋ λ 7→ ∇Fλ in C 0( ¯BV (θ, R), V ). Hence the map [0, 1] ∋ λ 7→ Fλ
is continuous in C 1( ¯BV (θ, R)). As in Step 1 of the proof of Theorem 3.1, the stability of critical
groups (cf. [16, Theorem III.4] and [22, Theorem 5.1]) leads to the first claim in (II).
sup
NXi=1 Xα<m−n/2
sup
α(x, 0; λ)dx
+
sup
99
For the second claim, it suffices to prove that ~u0 is a local minimizer of F0 provided ~uλ is
a local minimizer of Fλ. Since ~uλ is an isolated critical point of Fλ, by Example 1 in [13, page
33] we have Cq(Fλ, ~uλ; K) = δq0K for q = 0, 1,··· . It follows that Cq(F0, ~u0; K) = δq0K for
q = 0, 1,··· . By Theorem 2.3, this means that the Morse index of F0 at ~u0 must be zero. We can
assume ~u0 = θ after replaceing F0 by F0(~u0 + ·). So Cq(F◦
0, θ; K) = δq0K for q = 0, 1,··· .
Then θ is a local minimizer of F◦
0 by Example 4 in [13, page 43]. It follows from Theorem 2.2 (or
Theorem 6.1 with O = θ) that ~u0 = θ must be a local minimizer of F0.
8 Concluding remarks
In Section 3 we only generalize some bifurcation theorems for potential operators with the splitting
theorem obtained in this paper. Once some splitting theorems are proved, the same ideas can be
used to generalize some past bifurcation theorems. For example, we may obtain corresponding
extended versions of [2, 1] in the variational frames of [39, 40] and [5, 34]. These and applications
will be given in [46].
We here do not consider easy generalizations of the contents in Part II to a larger framework as
in [55, 61, 54] because they are developed in a more general setting as in [35, 36], see [47]. More-
over, both the theory in Part I and that of [39, 40] are applicable to one-dimensional variational
problem of higher order, see [48].
A Proof of Proposition 4.3
Recall that we have written ξ ∈ RM (m) as ξ = {ξα : α ≤ m} and denote by ξ◦ = {ξα : α <
m − n/p}. By the mean value theorem and (1.5) we get a collect of numbers {tβ ∈ (0, 1) : β ≤
m} such that
fα(x, ξ) − fα(x, 0) ≤ Xβ≤m
g1(tβξ◦) 1 + Xm−n/p≤γ≤m
≤ Xβ≤m
g1(ξ◦) 1 + Xm−n/p≤γ≤m
≤ Xβ≤m
g1(ξ◦) 1 + Xm−n/p≤γ≤m
≤ Xβ≤m
tβξγpγ!pαβ
ξγpγ!pαβ
ξγpγ!pαβ
ξβ
fαβ(x, tβξ) · ξβ
ξβ
ξβ
100
ξγpγ!pαβ
ξβ
g1(ξ◦) 1 + Xm−n/p≤γ≤m
= Xβ<m−n/p
g1(ξ◦) 1 + Xm−n/p≤γ≤m
+ Xm−n/p≤β≤m
≤ g1(ξ◦)ξ◦ Xβ<m−n/p 1 + Xm−n/p≤γ≤m
+g1(ξ◦) Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
ξγpγ!pαβ
ξγpγ!pαβ
ξγpγ!pαβ
ξβ
ξβ.
It follows that
fα(x, ξ) ≤ fα(x, 0) + g1(ξ◦)ξ◦ Xβ<m−n/p 1 + Xm−n/p≤γ≤m
ξγpγ!pαβ
+g1(ξ◦) Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
ξγpγ!pαβ
ξβ,
(A.1)
which lead to (4.9) with g4(ξ◦) := g1(ξ◦)ξ◦ + g1(ξ◦).
Suppose α < m − n/p. Then pαβ = 1 − 1/pβ = 1/qβ if m − n/p ≤ β ≤ m, and hence
the second and third terms in (A.1), respectively, becomes
ξγpγ!1/qβ
g1(ξ◦)ξ◦ Xβ<m−n/p 1 + Xm−n/p≤γ≤m
ξγpγ!,
≤ g1(ξ◦)ξ◦M (m) 1 + Xm−n/p≤γ≤m
g1(ξ◦) Xβ<m−n/p 1 + Xm−n/p≤γ≤m
ξγpγ!1/qβ
≤ g1(ξ◦) Xm−n/p≤β≤m"(cid:18)1 + Xm−n/p≤γ≤m
ξγpγ(cid:19) + ξβpβ#
≤ g1(ξ◦)(M (m) + 1) 1 + Xm−n/p≤γ≤m
ξγpγ!.
ξβ
(A.2)
(A.3)
These and (A.1)-(A.2) give rise to (4.10) with g5(ξ0) := (M (m) + 1)g1(ξ◦)(ξ◦ + 1).
Suppose m − n/p ≤ α ≤ m. Then 0 < pαβ ≤ 1 − 1/pα − 1/pβ if m − n/p ≤ β ≤ m. In
this case the second and third terms in (A.1), respectively, becomes
and
ξγpγ!pαβ
ξγpγ!1/qα
ξγpγ /qα!
g1(ξ◦)ξ◦ Xβ<m−n/p 1 + Xm−n/p≤γ≤m
≤ g1(ξ◦)ξ◦ Xβ<m−n/p 1 + Xm−n/p≤γ≤m
≤ g1(ξ◦)ξ◦M (m) 1 + Xm−n/p≤γ≤m
ξγpγ!pαβ
ξγpγ!1/qα−1/pβ
ξγpγ(cid:19)1/qα−1/pβ!pβ/(pβ −qα)
ξγpγ(cid:19)1/qα
ξγpγ!1/qα
g1(ξ◦) Xβ<m−n/p 1 + Xm−n/p≤γ≤m
≤ g1(ξ◦) Xβ<m−n/p 1 + Xm−n/p≤γ≤m
≤ g1(ξ◦) Xβ<m−n/p" (cid:18)1 + Xm−n/p≤γ≤m
≤ g1(ξ◦) Xβ<m−n/p"(cid:18)1 + Xm−n/p≤γ≤m
≤ g1(ξ◦)(M (m) + 1) 1 + Xm−n/p≤γ≤m
+ ξβpβ/qα#
ξβ
ξβ
.
These lead to (4.11).
With a similar argument to (A.1) we obtain
101
(A.4)
+ ξβpβ /qα#
(A.5)
f (x, ξ) − f (x, 0) ≤ Xα≤m
fα(x, sαξ) · ξα ≤ Xα≤m
g1(sαξ◦)sαξ◦ Xβ<m−n/p 1 + Xm−n/p≤γ≤m
+ Xα≤m
g1(sαξ◦) Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
+ Xα≤m
≤ Xα≤m
+g1(ξ◦)ξ◦ Xα≤m Xβ<m−n/p 1 + Xm−n/p≤γ≤m
+g1(ξ◦) Xα≤m Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
fα(x, 0) · ξα
ξγpγ!pαβ
ξγpγ!pαβ
= T1 + T2 + T3.
fα(x, 0) · ξα
ξα
sαξγpγ!pαβ
sαξγpγ!pαβ
sαξβ · ξα
ξα
ξβ · ξα
102
We can estimate these three terms as follows.
T1 = Xα<m−n/p
= ξ◦ Xα<m−n/p
≤ ξ◦ Xα<m−n/p
fα(x, 0) · ξα + Xm−n/p≤α≤m
fα(x, 0) + Xm−n/p≤α≤m
fα(x, 0) + Xm−n/p≤α≤m
fα(x, 0) · ξα
fα(x, 0) · ξα
fα(x, 0)qα + Xm−n/p≤α≤m
ξαpα;
ξα
ξγpγ!pαβ
ξα
ξγpγ!1/qα
ξα
ξγpγ(cid:19) + ξαpα#
ξγpγ!pαβ
T2 = g1(ξ0)ξ◦ Xα<m−n/p Xβ<m−n/p 1 + Xm−n/p≤γ≤m
+g1(ξ◦)ξ◦ Xm−n/p≤α≤m Xβ<m−n/p 1 + Xm−n/p≤γ≤m
ξγpγ!
≤ g1(ξ◦)ξ◦2M (m − n/p + 1) 1 + Xm−n/p≤γ≤m
+g1(ξ◦)ξ◦M (m − n/p + 1) Xm−n/p≤α≤m 1 + Xm−n/p≤γ≤m
≤ g1(ξ◦)ξ◦2M (m − n/p + 1) 1 + Xm−n/p≤γ≤m
+g1(ξ◦)ξ◦M (m − n/p + 1) Xm−n/p≤α≤m"(cid:18)1 + Xm−n/p≤γ≤m
ξγpγ!
≤ g1(ξ◦)ξ◦2M (m) 1 + Xm−n/p≤γ≤m
+g1(ξ◦)ξ◦(M (m) + 1)2 1 + Xm−n/p≤γ≤m
= g1(ξ◦)(cid:2)ξ◦2M (m) + ξ◦(M (m) + 1)2(cid:3) 1 + Xm−n/p≤γ≤m
ξγpγ!
ξγpγ!
ξγpγ!;
T3 = g1(ξ◦) Xα<m−n/p Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
+ g1(ξ◦) Xm−n/p≤α≤m Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
ξγpγ!pαβ
ξγpγ!pαβ
ξβ · ξα
ξβ · ξα
103
ξβ · ξα
ξβ
ξγpγ!pαβ
ξγpγ! + ξβpβ#
ξγpγ!1/qβ
≤ g1(ξ◦)ξ◦ Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
+ g1(ξ◦) Xm−n/p≤α≤m Xm−n/p≤β≤m 1 + Xm−n/p≤γ≤m
≤ g1(ξ◦)ξ◦ Xm−n/p≤β≤m" 1 + Xm−n/p≤γ≤m
+ g1(ξ◦) Xm−n/p≤α≤m Xm−n/p≤β≤m"ξβpβ + ξαpα
+ 1 + Xm−n/p≤γ≤m
≤ g1(ξ◦)ξ◦(M (m) + 1) 1 + Xm−n/p≤γ≤m
+g1(ξ◦)(M (m) + 1)2 1 + Xm−n/p≤γ≤m
ξγpγ!pαβ (1−p−1
ξγpγ!
ξγpγ!
α −p−1
β )#
because 0 < pαβ < 1 − 1
pα − 1
pβ
In summary we get
for m − n/p ≤ α ≤ m and m − n/p ≤ β ≤ m.
f (x, ξ) ≤ f (x, 0) + ξ◦ Xα<m−n/p
ξαpα
ξγpγ!
fα(x, 0) + Xm−n/p≤α≤m
fα(x, 0)qα
+ Xm−n/p≤α≤m
+ g1(ξ◦)[ξ◦2M (m) + ξ◦(M (m) + 1)2] 1 + Xm−n/p≤γ≤m
ξγpγ!
+ g1(ξ◦)ξ◦(M (m) + 1) 1 + Xm−n/p≤γ≤m
ξγpγ!
+g1(ξ◦)(M (m) + 1)2 1 + Xm−n/p≤γ≤m
fα(x, 0) + Xm−n/p≤α≤m
≤ f (x, 0) + ξ◦ Xα<m−n/p
ξαpα!,
+g3(ξ◦) 1 + Xm−n/p≤α≤m
fα(x, 0)qα
where
g3(ξ◦) = 1 + g1(ξ◦)[ξ◦2M (m) + ξ◦(M (m) + 1)2]
+g1(ξ◦)ξ◦(M (m) + 1) + g1(ξ◦)(M (m) + 1)2.
104
References
✷
[1] T. Bartsch, "Topological methods for variational problems with symmetries", Lecture Notes
in Mathematics, 1560. Springer-Verlag, Berlin, 1993.
[2] T. Bartsch, M. Clapp, Bifurcation theory for symmetric potential operators and the equiv-
ariant cup-length, Math. Z., 204(1990), 341 -- 356.
[3] T. Bartsch, A. Szulkin and M. Willem, Morse theory and nonlinear differential equa-
tions, in "Handbook of Global Analysis" Elsevier Science Ltd, (2008), 41 -- 73. (MR2389633)
[10.1016/B978-044452833-9.50003-6]
[4] M. Berger, "Nonlinearity and Functional Analysis," Acad. Press, New York-London, 1977.
(MR0488101)
[5] N. A. Bobylev and Yu. M. Burman, Morse lemmas for multi-dimensional variational prob-
lems, Nonlinear Analysis, 18 (1992), 595-604.
[6] R. Bott, Nondegenerate critical manifold, Ann. of Math. 60 (1954), 248 -- 261.
[7] F. E. Browder, Nonlinear elliptic boundary value problems. II Trans. Amer. Math. Soc. 117
(1965), 530 -- 550.
[8] F. E. Browder, Nonlinear elliptic boundary value problems and the generalized topological
degree Bull. Amer. Math. Soc. 76 (1970), 999 -- 1005.
[9] F. E. Browder, Fixed point theory and nonlinear problem, Bull. Amer. Math. Soc. (N.S), 9
(1983), 1 -- 39. (MR0699315)
[10] L. Caklovic, S.J. Li, M. Willem, A note on Palais -- Smale condition and coercivity, Differ-
ential Integral Equations, 3(1990), 799 -- 800.
[11] A. Canino, Variational bifurcation for quasilinear elliptic equations, Calc. Var., 18(2003),
269 -- 286.
[12] K. C. Chang, "Infinite Dimensional Morse Theory and its applications," Univ. de Montreal,
97, 1985. (MR0837186)
[13] K. C. Chang,
"Infinite Dimensional Morse Theory and Multiple Solution Problem,"
Birkhauser, 1993. (MR1196690)
[14] K. C. Chang, "Methods in Nonlinear Analysis," Springer Monogaphs in Mathematics,
Springer 2005. (MR2170995)
105
[15] K. C. Chang, A bifurcation theorem, J. Systems Sci. Math. Sci., 4(1984), 191-195.
[16] K. C. Chang, H. Ghoussoub, The Conley index and the critical groups via an extension of
Gromoll-Meyer theory, Topol. Methods in Nonlinear Analysis, 7(1996), 77-93.
[17] K. C. Chang, Z. Q. Wang, Notes on the bifurcation theorem, J. fixed point theory appl.,
1(2007), 195 -- 208.
[18] S. -N. Chow, R. Lauterbach, A bifurcation theorem for critical points of variational problems,
Nonlinear Anal., Theory Methods Appl., 12(1988), 51 -- 61.
[19] S. Cingolani and M. Degiovanni, On the Poincar´e-Hopf theorem for functionals defined on
Banach spaces, Adv. Nonlinear Stud., 9 (2009), 679 -- 699. (MR2560125)
[20] S. Cingolani and G. Vannella, Marino -- Prodi perturbation type results and Morse indices of
minimax critical points for a class of functionals in Banach spaces, Annali di Matematica,
186 (2007), 155 -- 183.
[21] J. N. Corvellec, Deformation techniques in metric critical point theory, Advances in Non-
linear Analysis, 2(2013), 65C89.
[22] J. N. Corvellec, A. Hantoute, Homotopical Stability of Isolated Critical Points of Continuous
Functionals, Set-Valued Analysis, 10 (2002), 143 -- 164.
[23] D. M. Duc, T. V. Hung and N. T. Khai, Morse-Palais lemma for nonsmooth func-
tionals on normed spaces, Proc. Amer. Math. Soc., 135 (2007), 921 -- 927. (MR2262891)
[10.1090/S0002-9939-06-08662-X]
[24] D. M. Duc, T. V. Hung and N. T. Khai, Critical points of non-C 2 functionals, Topological
Methods in Nonlinear Analysis, 29 (2007), 35 -- 68. (MR2308216)
[25] I. Ekeland, An inverse function theorem in Frchet spaces, Ann. Inst. H. Poincar´e Anal. Non
Lin´eaire, 28(2011), no. 1, 91 -- 105.
[26] S.V. Emelyanov, S.K. Korovin, N.A. Bobylev, A.V. Bulatov, Homotopy of extremal prob-
lems. Theory and applications. De Gruyter Series in Nonlinear Analysis and Applications,
11. Walter de Gruyter & Co., Berlin, 2007.
[27] E. Fadell, P.H. Rabinowitz, Bifurcation for odd potential operators and an alternative
topological index, J. Funct. Anal. 26(1977), 48 -- 67.
[28] E. Fadell, P.H. Rabinowitz, Generalized cohomological index theories for Lie group actions
with an application to bifurcation questions for Hamiltonian systems, Invent Math. 45(1978),
139 -- 174.
106
[29] M. Feckan, An inverse function theorem for continuous mappings, J. Math. Anal. Appl.
185(1994), no. 1, 118C128.
[30] N. Ghoussoub, "Duality and perturbation methods in critical point theory," Cambridge
University Press, 2008.
[31] D. Gromoll and W. Meyer, On differentiable functions with isolated critical points, Topology,
8 (1969), 361 -- 369. (MR0246329)
[32] A. Ioffe and E. Schwartzman, An extension of the Rabinowitz bifurcation theorem to Lips-
chitz potenzial operators in Hilbert spaces, Proc. Amer. Math. Soc., 125(1997), 2725 -- 2732.
[33] T. Kato, "Perturbation theory for linear operators", Second edition. Grundlehren der Math-
ematischen Wissenschaften, Band 132. Springer-Verlag, Berlin-New York, 1976.
[34] M. Jiang, A generalization of Morse lemma and its applications, Nonlinear Analysis, 36
(1999), 943 -- 960. (MR1684523) [10.1016/S0362-546X(97)00701-3]
[35] G. Kokarev, S. Kuksin, Quasilinear elliptic differential equations on mappings of manifolds.
I. (Russian. Russian summary), Algebra i Analiz, 15(2003), no. 4, 1 -- 60; translation in St.
Petersburg Math. J. 15(2004), no. 4, 469 -- 505.
[36] G. Kokarev, S. Kuksin, Quasilinear elliptic differential equations on mappings of manifolds.
II., Ann. Global Anal. Geom., 31(2007), no. 1, 59 -- 113.
[37] M. A. Krasnosel'skii, Topological Methods in the Theory of Nonlinear Integral Equations,
McMillan, New York, 1964.
[38] A. Lazer, S. Solimini, Nontrivial solutions of operator equations and Morse indices of
critical points of min-max type, Nonlin. Anal. TMA, 12(1988), 761-775.
[39] G. Lu, Corrigendum to "The Conley conjecture for Hamiltonian systems on the cotangent
bundle and its analogue for Lagrangian systems" [J. Funct. Anal. 256(9)(2009)2967-3034],
J. Funct. Anal., 261 (2011), 542 -- 589. (MR2502430) [10.1016/j.jfa.2009.01.001]
[40] G. Lu, The splitting lemmas for nonsmooth functionals on Hilbert spaces I, Discrete Contin.
Dyn. Syst. 33(2013), no. 7, 2939-2990.
[41] G. Lu, The splitting lemmas for nonsmooth functionals on Hilbert spaces II, Topol. Meth.
Nonlinear Anal. 44(2014), 277-335.
[42] G. Lu, The splitting lemmas for nonsmooth functionals on Hilbert spaces, arXiv:1102.2062.
[43] G. Lu, Methods of infinite dimensional Morse theory for geodesics on Finsler manifolds,
Nonlinear Anal. 113(2015), 230-282.
107
[44] G. Lu, Splitting lemmas for the Finsler energy functional on the space of H 1-curves, Proc.
London Math. Soc. 113(2016), no.3,24-76.
[45] G. Lu, Nonsmooth generalization of some critical point theorems for C 2 functionals (in
Chinese), Sci Sin Math, 46(2016), 615-638, doi:10.1360/N012015-00375.
[46] G. Lu, Splitting theorems and bifurcation, In Progress.
[47] G. Lu, Morse theory for quasi-linear elliptic equations on mappings of manifolds,
In
preparation.
[48] G. Lu, Variational methods for Lagrangian systems of higher order, In Progress.
[49] A. Marino and G. Prodi, Metodi perturbativi nella teoria di Morse, Boll. Un. Mat. Ital.,
11(1975), 1 -- 32.
[50]
J. Mawhin and M. Willem, "Critical Point Theory and Hamiltonian Systems," Applied
Mathematical Sciences 74, Springer-Verlag, New York, 1989. (MR0982267)
[51] M. Morse, "The calculus of variations in the large," American Math. Soc. Colloquium
Publications 18, Ann Arbor, Mich., 1934.
[52] D. Motreanu, V. Motreanu, N. Papageorgiou, "Topological and variational methods with
applications to nonlinear boundary value problems," Springer, New York, 2014.
[53] R. Palais, Morse theory on Hilbert manifolds, Topology, 2(1963), 299-340.
[54] R. Palais, "Foundations of global non-linear analysis," W. A. Benjamin,, 1968, 44.
[55] R. S. Palais and S. Smale, A generalized Morse theory, Bull. Amer. Math. Soc., 70(1964),
165 -- 172.
[56] K. Perera, R. P. Agarwal and Donal O'Regan, "Morse Theoretic Aspects of p-Laplacian Type
Operators," Mathematical Surveys and Monographs 161, American Mathematical Society,
Providence Rhode Island 2010. (MR2640827)
[57] P. H. Rabinowitz, A bifurcation theorem for potential operators, J. Funct. Anal., 25(1977),
412 -- 424.
[58]
I. V. Skrypnik, "Nonlinear Elliptic Equations of a Higher Order," [in Russian], Naukova
Dumka, Kiev 1973. (MR0435590)
[59] I. V. Skrypnik, Solvability and properties of solutions of nonlinear elliptic equations, J.Soviet
Math. 12(1979), 555-629.
[60]
I. V. Skrypnik, "Methods for Analysis of Nonlinear Elliptic Boundary Value Problems,"
in:Translations of Mathematical Monographs,vol.139, Providence, Rhode Island, 1994.
108
[61] S. Smale, Morse theory and a non-linear generalization of the Dirichlet problem, Ann.
Math., 80(1964), 382-396.
[62] T. Suzuki, On the relation between the weak Palais -- Smale condition and coercivity given by
Zhong, Nonlinear Analysis, 68(2008), 2471 -- 2478.
[63] S. A. Vakhrameev, Critical point theory for smooth functions on Hilbert manifolds with
singularities and its application to some optimal control problems, J. Sov. Math., 67 (1993),
2713 -- 2811. (MR1262866) [10.1007/BF01455151]
[64] A. Vanderbauwhede, G. Iooss, Center manifold theory in infinite dimensions, In: Dynamics
Reported, New Series (C.K.R.T. Jones, U. Kirchgraber, H.O. Walther eds.) Vol. 1, Springer,
Berlin 1992, 125 -- 163.
[65] G. Vannella, Morse theory applied to a T 2-equivriant problem, Topological Methods in
Nonlinear Analysis, 17 (2001), 41 -- 53.
[66] C. Viterbo,
Indice de Morse des points critiques obtenus par minimax, Ann.Inst. Henri
Poincar´e, 5 (1988), 221-225.
[67] Z. Q. Wang, Equivariant Morse theory for isolated critical orbits and its applications to
nonlinear problems, Lect. Notes in Math. No. 1306, Springer, (1988) 202-221.
[68] Z. Q. Wang, Symmetries, Morse Polynomials and Applications to Bifurcation Problems,
Acta Mathematica Sinica, New Series, 6(1990) 165-177.
[69] G. Wasserman, Equivariant differential topology, Topology, 8 (1969), 127 -- 150.
[70] W. M. Zou, M. Schechter, "Critical point theory and its applications," Springer, New York,
2006.
|
1603.07533 | 3 | 1603 | 2016-09-13T20:00:29 | Sampling measures, Muckenhoupt Hamiltonians, and triangular factorization | [
"math.FA"
] | Let $\mu$ be an even measure on the real line $\mathbb{R}$ such that $$c_1 \int_{\mathbb{R}}|f|^2\,dx \le \int_{\mathbb{R}}|f|^2\,d\mu \le c_2\int_{\mathbb{R}}|f|^2\,dx$$ for all functions $f$ in the Paley-Wiener space $\mathrm{PW}_{a}$. We prove that $\mu$ is the spectral measure for the unique Hamiltonian $\mathcal{H}=\left(w&00&\frac{1}{w}\right)$ on $[0,a]$ generated by a weight $w$ from the Muckenhoupt class $A_2[0,a]$. As a consequence of this result, we construct Krein's orthogonal entire functions with respect to $\mu$ and prove that every positive, bounded, invertible Wiener-Hopf operator on $[0,a]$ with real symbol admits triangular factorization. | math.FA | math |
SAMPLING MEASURES, MUCKENHOUPT HAMILTONIANS,
AND TRIANGULAR FACTORIZATION
R. V. BESSONOV
Abstract. Let µ be an even measure on the real line R such that
c1 ZR
f 2 dx 6 ZR
f 2 dµ 6 c2 ZR
f 2 dx
for all functions f in the Paley-Wiener space PWa. We prove that µ is the
spectral measure for the unique Hamiltonian H = (cid:16) w 0
w (cid:17) on [0, a] generated
by a weight w from the Muckenhoupt class A2[0, a]. As a consequence of this
result, we construct Krein's orthogonal entire functions with respect to µ and
prove that every positive, bounded, invertible Wiener-Hopf operator on [0, a]
with real symbol admits triangular factorization.
0 1
1. Introduction
The classical Paley-Wiener space PWa consists of entire functions of exponential
type at most a square summable on the real line, R. A measure µ on R is called
a sampling measure for the space PWa if there exist positive constants c1, c2 such
that
c1ZR f2 dx 6ZR f2 dµ 6 c2ZR f2 dx,
f ∈ PWa.
(1)
Let H be a regular Hamiltonian on [0, a], that is, H is a mapping from [0, a] to
the set of 2 × 2 non-negative matrices with real entries such that traceH is a
positive non-vanishing function in L1[0, a]. Denote by ΘH = ΘH(r, z) solution of
the following Cauchy problem:
It is known from a general theory of canonical Hamiltonian systems that for every
JX′(r) = zH(r)X(r), X : [0, a] → C2, X(0) = ( 1
z ∈ C.
measure µ satisfying (1) there exists a regular Hamiltonian H with R a
such that µ is a spectral measure for problem (2). The latter means that the
Weyl-Titchmarsh transform
√detH = a
0 ) ,
(2)
0
WH,a : X 7→
1
√π Z a
0 (cid:10)H(r)X(r), ΘH(r, ¯z)(cid:11)C2 dr,
z ∈ C,
(3)
2010 Mathematics Subject Classification. Primary 34L05, Secondary 47B35.
Key words and phrases. Canonical Hamiltonian system, Muckenhoupt weight, Inverse prob-
lem, Paley-Wiener space, Truncated Toeplitz operator, Triangular factorization.
The work is supported by RFBR grant mol_a_dk 16-31-60053 and by "Native towns", a social
investment program of PJSC "Gazprom Neft".
1
MUCKENHOUPT HAMILTONIANS
2
generated by solution ΘH of Cauchy problem (2) maps isometrically the space
L2(H, a) =nX : [0, a] → C2 : kXk2
0 (cid:10)H(r)X(r), X(r)(cid:11)C2dr < ∞o.K(H),
L2(H,a) =Z a
K(H) =nX : H(t)X(t) = 0 for almost all t ∈ [0, r]o
into the space L2(µ). A general problem in the inverse spectral theory is to translate
properties of a spectral measure µ into properties of the Hamiltonian H it generates.
Two essentially different cases of the above problem attracted much attention.
If µ is a "small perturbation" of the Lebesgue measure on R (in the sense that
the Fourier transform of µ restricted to the interval [−a, a] differs from the point
mass measure δ0 concentrated at 0 by a function in L1[−a, a]), the I. M. Gelfand --
B. M. Levitan approach [6], [12] gives a quite precise information on relation be-
tween µ and H. On the other hand, if µ is arbitrary measure on R such that
RR
1+t2 < ∞, the theory of M. G. Krein [8] (for even measures µ) and L. de Bran-
ges [4] (for all µ) implies the existence of a unique Hamiltonian H ∈ L1
loc[0,∞)
such that µ is the spectral measure for H. However, it is not known how translate
even simple properties of a Hamiltonian H (e.g., membership in Lp class for some
p > 1) to the properties of its spectral measure µ and vice versa. In this paper
we consider a "median" situation (spectral measures with sampling property (1) for
the Paley-Wiener space PWa) and use both Gelfand-Levitan and Krein-de Branges
theories.
dµ(t)
1
result of the paper.
IRI w(cid:1)·(cid:0) 1
IRI
A measure µ on R is called even if µ(S) = µ(−S) for every Borel set S ⊂ R.
A function w > 0 belongs to the Muckenhoupt class A2[0, a] if the supremum of
products (cid:0) 1
w(cid:1) over all intervals I ⊂ [0, a] is finite. Here is the main
w (cid:17)
measure for problem (2) corresponding to the unique Hamiltonian H = (cid:16) w 0
generated by a weight w ∈ A2[0, a].
The Hamiltonian H in Theorem 1 could be recovered from the spectral measure µ
Theorem 1. Let µ be an even sampling measure for PWa. Then µ is the spectral
0 1
by means of the following simple formula:
where Tµ,r is the truncated Toeplitz operator on PWr with symbol µ defined by
,
sincr =
sin rx
πx
,
r ∈ [0, a],
L2(µ)
w(r) = π
2
∂
T −1
∂r(cid:13)(cid:13)(cid:13)
µ,r sincr(cid:13)(cid:13)(cid:13)
(Tµ,rf )(z) =ZR
f (x)
sin r(x − z)
π(x − z)
dµ(x),
z ∈ C.
(4)
A nontrivial fact is that the continuous increasing function r 7→ kT −1
L2(µ)
is absolutely continuous and its derivative w/π does not vanish on a set of positive
Lebesgue measure.
In the proof of Theorem 1 we first obtain an estimate for
the "A2-norm" of w in terms of c1, c2 assuming above properties of w; then use
an approximation argument based on a description of positive truncated Toeplitz
operators on PWr and Lp-summabilty of weights w ∈ A2[0, a] for some p > 1.
µ,r sincr k2
Section 5 in [2] contains an example of a diagonal Hamiltonian H on [0, 1] such
that both H, H−1 are uniformly bounded on [0, 1], but the spectral measures of the
corresponding problem (2) fail to have sampling property. This shows that A2[0, a]
MUCKENHOUPT HAMILTONIANS
3
class does not describe canonical Hamiltonian systems generated by sampling mea-
sures for PWa.
Theorem 1 yields two results of independent interest.
Given a measure µ satisfying (1) and a number r ∈ [0, 2a], denote by (PW[0,r], µ)
the Paley-Wiener space of functions from L2(R) with Fourier spectrum in [0, r]
equipped with the inner product taken from L2(µ).
Theorem 2. Let µ be an even sampling measure for the space PWa. Then there ex-
0 f (t)Pt(z) dt
ists a family of entire functions {Pt}t∈[0,2a] such that Fµ : f 7→ 1√2π R r
is the unitary operator from L2[0, r] to (PW[0,r], µ) for every r ∈ [0, 2a].
In the case where µ is a "small perturbation" of the Lebesgue measure (see
discussion above), the functions Pr in Corollary 2 coincide with orthogonal entire
functions constructed by M. G. Krein in [10]. S. A. Denisov provides an extensive
treatment of the subject, collecting many old and new results in paper [5].
The second application of Theorem 1 concerns the classical factorization problem
for positive invertible operators. Let H be a separable Hilbert space and let B(H)
be the algebra of all bounded operators on H. Consider a complete chain N of
subspaces in H and denote by AN = {A ∈ B(H) : AE ⊂ E, E ∈ N} the nest
algebra of upper-triangular operators with respect to N . In sixties, I. C. Gohberg
and M. G. Krein proved (see Theorem 6.2 in Chapter 4 of [7]) that every positive
invertible operator T on H of the form T = I−K with K in Macaev ideal Sω admits
the triangular factorization T = A∗A, where A = I − KA is an invertible operator
on H such that KA ∈ Sω ∩ AN . Famous theorem by D. R. Larson [11] says that
every positive invertible operator T admits triangular factorization T = A∗A with
A, A−1 ∈ AN if and only if the chain N is countable. Moreover, given 0 < ε < 1,
the non-factorable operator T can be chosen so that K = I − T is a compact
operator with kKk < ε.
We consider the problem of triangular factorization for Wiener-Hopf convolution
operators. Let ψ ∈ S′ be a tempered distribution on R and let 0 < a 6 ∞. The
Wiener-Hopf operator Wψ on L2[0, a) with symbol ψ is densely defined by
syf : x 7→ f (x − y),
(Wψf )(y) =(cid:10)ψ, syf(cid:11)S ′ ,
y ∈ [0, a),
on smooth functions f with compact support in (0, a). In the case where ψ ∈ L1(R)
we have more familiar definition, Wψ : f 7→R a
0 ψ(x− y)f (x) dx. As following result
shows, Wiener-Hopf operators with real symbols are always factorable.
Theorem 3. Let 0 < a 6 ∞. Every positive, bounded, and invertible Wiener-Hopf
operator Wψ on L2[0, a) with real symbol ψ ∈ S′ admits triangular factorization:
Wψ = A∗A, where A is a bounded invertible operator such that AL2[0, r] = L2[0, r]
for every r ∈ [0, a).
Wiener-Hopf operators Wψ in Theorem 3 admit triangular factorizations in the
reverse order Wψ = AA∗ as well. Relation of absolute continuity of aforementioned
function r 7→ kT −1
L2(µ) to triangular factorization problems has been pre-
viously found in different terms by L. A. Sakhnovich, see Theorem 4.2 in [16]. On
the other hand, Theorem 3 contradicts Theorem 4.1 from another work [17] by the
same author. See discussion in Section 5.
µ,r sincr k2
MUCKENHOUPT HAMILTONIANS
4
Acknowledgement. The author is grateful to many colleagues who took a part
in discussions related to the subject of the paper, especially to Roman Romanov,
Mikhail Sodin, Pavel Zatitsky, and Dmitriy Zaporozhets.
2. Integration over simplex and the Muckenhoupt class A2
Let w be a positive function on an interval [0, a]. We associate to w the quantity
kwkA2[0,a] = sup
I⊂[0,a](cid:18) 1
IZI
w(x) dx(cid:19)·(cid:18) 1
IZI
1
w(x)
dx(cid:19) ,
where I runs over all subintervals of [0, a]. Note that k·kA2[0,a] is not a norm in the
standard sense, but we will use this convenient notation. The Muckenhoupt class
A2[0, a] consists of functions w > 0 such that kwkA2[0,a] < ∞. In this section we
present a special integral condition for a weight w to belong to the A2[0, a] class.
Let ϕ be a real-valued function on the interval [0, a]. For a real 0 < t < a and
an integer n > 1 define the mapping
n
Gϕ,n : x 7→
(−1)n+kϕ(xk),
x ∈ Kt,n,
(5)
on simplex Kt,n = {x ∈ Rn : x = (x1, . . . , xn), t > x1 > . . . > xn > 0}. Let mn
denote the usual Lebesgue measure on Rn.
Xk=1
Next proposition will be used in the proof of Theorem 1.
Proposition 2.1. Let ϕ be a function on [0, a] such that eϕ ∈ L1[0, a]. Assume
that for every r ∈ [0, a] and every integer n > 1 we have
eGϕ,n(x)dmn(x)!2
dt 6 b2,
(6)
1
an(r)Z r
0
eϕ(t) dt 6 b2;
(7)
e(−1)nϕ(t) ZKt,n
r Z r
b1 6
1
0
where b1, b2 are positive constants, and an(r) = r2n+1(2n + 1)−1(n!)−2. Then the
function w = eϕ belongs to A2[0, a] and kwkA2[0,a] 6 228(b2 + b−2
1 b2)14.
We first prove several preliminary estimates.
Lemma 2.1. Let ϕ be a function as in Proposition 2.1. Then for every r ∈ [0, a]
and b = 2(b2 + b−2
1 b2) we have
1
r Z r
0 ϕ(t) dt 6 log b,
R r
0 k(t) dt = 1 and k(r) = 0 we have R r
1
r Z r
0
0 ϕ(t)k(t) dt 6 log b.
Consequently, for every decreasing differentiable function k > 0 on [0, r] satisfying
Proof. Clearly, the first estimate in (8) follows from the second one and the
Jensen's inequality for convex function ex. Taking n = 1 in (6), we obtain
eϕ(t) dt 6 b.
(8)
3
r3 Z r
0
e−ϕ(t)(cid:18)Z t
0
2
eϕ(t1) dt1(cid:19)
dt 6 b2.
From (7) we know that 1
Using the other side estimate 1
see that
MUCKENHOUPT HAMILTONIANS
5
b−2
1 b2 >
0
1
e−ϕ(t) dt.
e−ϕ(t)t2 dt >
r Z r
t R t
0 eϕ(t1) dt1 > b1 for all t ∈ [0, r]. It follows that
r3 Z r
3
r R r
r Z r
eϕ(t) dt 6 b2 + b−2
1 b2
r/2
r/2
2
0 eϕ(t) dt 6 b2 and inequality ex 6 ex + e−x, we
for all r ∈ [0, a]. Then (8) follows from
r ∞
Xk=0
r Z r
eϕ(t) dt =
1
1
0
Ir,k ·
1
Ir,kZIr,k
eϕ(t) dt! 6 b,
where Ir,k = [2−k−1r, 2−kr]. Now if k is a function on [0, r] ⊂ [0, a] as in the
statement, we have
0 ϕ(t)k(t) dt = −Z r
Z r
= −Z r
6 − log bZ r
0 ϕ(t)Z r
k′(s)Z r
0
0
0
0
χ[t,r](s)k′(s) ds dt
χ[0,s](t)ϕ(t) dt ds
k′(s)s ds = log b.
This completes the proof.
(cid:3)
For n > 1 introduce the intervals It,n = [δnt, t], where δn = 1 − 1
n if n is even. In particular, It,n = It,n+1 for every odd n. Set
n+1 if n is odd,
and δn = 1 − 1
[ϕ]t,n = 2(−1)n+1ZKt,n
Gϕ,n(x) dmt,n(x),
where mt,n = n!
normalized so that mt,n(Kt,n) = 1.
tn · mn is the scalar multiple of the Lebesgue measure mn on Rn
Lemma 2.2. For r ∈ [0, a] and odd n > 1 we have [ϕ]δnr,n−[ϕ]δn+1r,n+1 < 6 log b,
where b is the constant from Lemma 2.1.
Proof. Arguing by induction, it is easy check that for all n > 1 and τ ∈ [0, a] we
have
[ϕ]τ,n =Z τ
0
ϕ(s)kτ,n(s) ds,
kτ,n(s) =
2n
τ n (2s − τ )n−1.
For odd (correspondingly, even) integers n the kernels kτ,n are even (correspond-
ingly, odd) functions with respect to the point τ /2. As n tends to infinity, the
kernels kτ,n tend to zero uniformly on every closed interval in (0, τ ). We also have
Z τ
0 kτ,n(s) ds = 2,
sup
s∈[τ /2,τ ]kτ,n(s) − kτ,n+1(s) 6
2
τ
.
(9)
MUCKENHOUPT HAMILTONIANS
6
Now take an odd integer n > 1 and note that δn = δn+1 = 1
we obtain
n+1 . Setting τ = δnr,
(cid:12)(cid:12)(cid:12)
[ϕ]τ,n − [ϕ]τ,n+1(cid:12)(cid:12)(cid:12)
0
ϕ(s)kτ,n(s) ds
6Z τ /2
+Z τ /2
+Z τ
τ /2 ϕ(s) · kτ,n(s) − kτ,n+1(s) ds.
2kτ,n+1 on [0, τ
2kτ,n, and k = 1
ϕ(s)kτ,n+1(s) ds
0
By Lemma 2.1 for functions ϕ, k = 1
2 ], the sum of
first two integrals is bounded from above by 4 log b. To show that the last integral
does not exceed 2 log b, use (8) and the second estimate in (9).
(cid:3)
Proof of Proposition 2.1. Take an odd integer n > 1. Since the integrand in (6)
is positive, we have
b2 >
>
1
δnr
an(r)Z r
an(r)(cid:18)Z r
1
e−ϕ(t) ZKt,n
e−ϕ(t) dt(cid:19) · ZKδn r,n
δnr
eGϕ,n(x)dmn(x)!2
dt,
eGϕ,n(x)dmn(x)!2
.
By Jensen's inequality,
eGϕ,n(x)dmn(x) >
(δnr)n
n!
2 (cid:19) .
exp(cid:18) [ϕ]δnr,n
(2n + 1)(n!)2
r2n+1
r2n
(n!)2 δ2n
n >
·
n + 1
32r
=
1
32In,r
.
For all n > 1 we have
We now see that
1
=
ZKδnr,n
n! (cid:19)2
an(r) ·(cid:18) (δnr)n
Ir,nZIr,n
Ir,n+1ZIr,n+1
1
1
Analogously, for the even integer n + 1 we have
exp(cid:16)−ϕ(t) + [ϕ]δnr,n(cid:17) dt 6 32b2.
exp(cid:16)ϕ(t) − [ϕ]δn+1r,n(cid:17) dt 6 32b2.
(10)
(11)
(12)
Recall that It,n+1 = It,n. Applying Lemma 2.2, we obtain
1
Ir,nZIr,n
exp(cid:16)ϕ(t) − [ϕ]δnr,n(cid:17) dt 6 32b2e6 log b 6 32b7,
where b is the constant from Lemma 2.1. Using inequality ex 6 ex + e−x, we get
from (10) and (11) the estimate
1
eϕ(t)−cI dt 6 64b7
for all intervals I of the form I = [(1− 1
n+1 )r, r], where r ∈ [0, a], and integer n > 1
is odd. Here cI is a constant depending on I (in fact, cI = [ϕ]δnr,n works, but from
now on the particular choice of cI plays no role). Formula (8) gives (12) with cI = 0
for intervals of the form I = [0, t].
IZI
MUCKENHOUPT HAMILTONIANS
7
Indeed, let t be the right point of J.
Next, observe that each interval J ⊂ [0, a] is contained in an interval I sat-
isfying (12) and such that I 6 2J.
If
J > t/2, take I = [0, t]. In the case J < t/2 find an odd number n > 1 such
that It,n+2 ⊂ J ⊂ It,n and take I = It,n. Fix this interval I and the corresponding
constant cI form (12). We have
(cid:18) 1
JZJ
e−ϕ dt(cid:19)
eϕ dt(cid:19)·(cid:18) 2
IZI
eϕ−cI dt(cid:19)·(cid:18) 2
e−ϕ+cI dt(cid:19) 6 (2b)14.
IZI
Since interval J is arbitrary, this shows that function w = eϕ belongs to the Mucken-
houpt class A2[0, a] and kwkA2[0,a] 6 (2b)14 = 228(b2 + b−2
(cid:3)
e−ϕ dt(cid:19) 6(cid:18) 2
IZI
6(cid:18) 2
IZI
eϕ dt(cid:19)·(cid:18) 1
JZJ
1 b2)14.
3. Proof of Theorem 1
As it was mentioned in the Introduction, we will use an approximation argument
in the proof of Theorem 1. To have a stable approximation, we need a result
describing positive truncated Toeplitz operators on PWa.
3.1. Preliminaries on truncated Toeplitz operators. Let µ > 0 be a measure
on the real line R such that kfk2
L2(R) for all functions f ∈ PW[0,a].
Define the truncated Toeplitz operator Aµ,a on PW[0,a] by the sesquilinear form
6 ckfk2
L2(µ)
(Aµ,af, g)L2(R) =ZR
f ¯g dµ,
f, g ∈ PW[0,a].
(13)
In the case where µ = u dm is absolutely continuous with respect to the Lebesgue
measure m on R and has density u, the operator Aµ,a coincides with the projec-
tion of the standard Toeplitz operator Tu on the Hardy space H 2 to the subspace
PW[0,a]. This explains the name "truncated Toeplitz" for the operator Aµ,a.
It is well-known (see, e.g., Section 6.1 in [14]) that the operator
V : h 7→
1
√π
1
z + i
h(cid:18) z − i
z + i(cid:19) ,
z ∈ C+,
(14)
maps unitarily the Hardy space H 2(D) in the open unit disk D = {ξ ∈ C : ξ < 1}
onto the Hardy space H 2 in the upper half-plane C+ = {z ∈ C : Im z > 0}.
Moreover, for every a > 0 we have V Kθa = PW[0,a], where θa = exp(cid:0)a z+1
z−1(cid:1) is
the inner function in D and Kθa is the orthogonal complement in H 2(D) to the
subspace θaH 2(D). As we will see in a moment, the truncated Toeplitz operators
defined by (13) are unitarily equivalent to truncated Toeplitz operators on the shift-
coinvariant subspace Kθa of H 2(D). See D. Sarason's paper [18] for basic properties
of truncated Toeplitz operators on general coinvariant subspaces of H 2(D).
We also will deal with the operators Tµ,a on the space PWa defined by the same
sesquilinear form
(Tµ,af, g) =ZR
f ¯g dµ,
f, g ∈ PWa.
It is easy to see that this definition agrees with formula (4). By construction, we
have Tµ,a = V −1
a Aµ,2aVa, where Va : PWa → PW[0,2a] is the unitary operator
taking a function f into eiazf .
MUCKENHOUPT HAMILTONIANS
8
Lemma 3.1. Let T be a positive bounded operator on PW[0,a] satisfying relation
(T f, f )L2(R) = (T z−i
(15)
for all functions f ∈ PW[0,a] such that f (−i) = 0. Then there exists a positive
measure µ on R such that T = Aµ,a. Similarly if T is a positive bounded operator
on PWa satisfying (15) for all f ∈ PWa such that f (−i) = 0, then T = Tµ,a for a
positive measure µ on R.
z+i f )L2(R)
z+i f, z−i
Proof. Let θa, Kθa, and V : Kθa → PW[0,a] be defined as above. Consider the
operator T = V −1T V on Kθa unitarily equivalent to the operator T on PW[0,a].
Recall that the inner product in Kθa is inherited from the space L2(T) on the unit
circle T = {ξ ∈ C :
ξ = 1}. Assumption (15) means that
( T h, h)L2(T) = ( T ξh, ξh)L2(T)
(16)
for every function h ∈ Kθa such that ξh ∈ Kθa. Indeed, (V ξh)(z) = z−i
z+i (V h)(z)
and hence V (ξh) ∈ PW[0,a] if and only if (V h)(−i) = 0. Theorem 8.1 in [18] says
that a bounded operator T on Kθa (or on any other coinvariant subspace K 2
θ of the
Hardy space H 2(D)) satisfying (16) is a truncated Toeplitz operator on Kθa. By
Theorem 2.1 in [1], for every positive bounded truncated Toeplitz operator T on
Kθa there exists a finite positive measure µ on T such that µ({1}) = 0 and
( T h, h)L2(T) =ZT h2 dµ
for all continuous functions h in Kθa. Changing variables in the last integral, we
find a positive measure µ on R such that
ZT h2 dµ =ZR f2 dµ,
f = V h.
It follows that (T f, f ) = ( T h, h)L2(T) = (Aµ,af, f )L2(R) for a dense set of functions
f in PW[0,a]. Since T is continuous, we have T = Aµ,a. The second part of the
Lemma is a direct consequence of relation Tµ,a = V −1
(cid:3)
a Aµ,2aVa.
3.2. Preliminaries on canonical Hamiltonian systems. Let H be a Hamilton-
ian on [0, a] with traceH ∈ L1[0, a]. Assume that there is no interval (r1, r2) ⊂ [0, a]
such that H(t) is a constant matrix of rank one for all points t ∈ (r1, r2). For
r ∈ [0, a] we will denote by B(H, r) the de Branges space generated by H on [0, r],
that is,
B(H, r) = WH,rL2(H, r) =nentire f : f = WH,rX, X ∈ L2(H, r)o,
where the Weyl-Titchmarsh transform WH,r is defined in (3) for a = r. The
space B(H, r) is actually the Hilbert space with respect to the inner product
(f, g)B(H,r) = (f, g)L2(µ), where µ is any spectral measure for problem (2). We refer
the reader to paper [2] for the summary of results on direct and inverse spectral
theory of canonical Hamiltonian systems and de Brange spaces of entire functions.
The readers interested in proofs or in a more detailed account may find neces-
sary information in Chapter 2 of classical book [4] by L. de Brange or its recent
exposition [15] by R. Romanov.
MUCKENHOUPT HAMILTONIANS
9
Lemma 3.2. Let µ be an even measure on R of the form µ = cm + ν, where c > 0
and ν is a finite positive measure on R with compact support. Then there exists an
infinitely smooth diagonal Hamiltonian H on [0, +∞) such that det H(r) = 1 for
all r > 0, and µ is the spectral measure for H.
Proof. The result is a kind of folklore. Since the Fourier transform of 1
c ν is
a smooth (in fact, analytic) function, one can use the classical Gelfand-Levitan
approach to find a smooth diagonal potential Q on [0, a] such that m + 1
c ν is
the spectral measure for the Dirac system JY ′ + QY = zY corresponding to the
0 ). Then rewrite system JY ′ + QY = zY as a
boundary condition Y (0) = ( 1
canonical Hamiltonian system JX′ = z HX setting X = M−1Y , H = M∗M , where
M is the matrix solution of equation JM′ = −QM , M (0) = ( 1 0
0 1 ). Observe that
det H = 1 almost everywhere on [0, a] and m+ 1
c ν is the spectral measure for system
JX′ = z HX, X(0) = ( 1
0 ). To obtain the Hamiltonian on [0, a] corresponding to
the spectral measure µ, put H =(cid:16) c 0
c(cid:17) H. Another (in a sense, equivalent) way of
Define typeB(H, r) = sup{type(f ), f ∈ B(H, r)} to be the maximal exponential
type of entire functions in de Branges space B(H, r). The following remarkable
formula of Krein [9] and de Brange (Theorem X in [3])
proving Lemma 3.2 is the application of Theorem 5.1 from [20].
0 1
(cid:3)
typeB(H, r) =Z r
0 pdetH(t) dt,
(17)
represents the maximal exponential type of functions in B(H, r) in terms of the
Hamiltonian H. Section 6 in [15] contains an elegant self-contained proof of this
result.
Lemma 3.3. Let H be a Hamiltonian on an interval [0, a] such that its spectral
measure µ satisfies (1). Assume that detH(r) = 1 for almost all r ∈ [0, a]. Then
for all r ∈ [0, a] we have B(H, r) = (PWr, µ).
Proof. Let r ∈ [0, a) and let ε > 0 be such that r ∈ [ε, a − ε). Then the
Hilbert space (PWr+ε, µ) of entire functions satisfies an axiomatic description of
de Branges spaces (Theorem 23 in [4]) and the embedding (PWr+ε, µ) ⊂ L2(µ) is
isometric. Since µ is a spectral measure for H, the embedding B(H, r) ⊂ L2(µ) is
isometric as well. Applying de Branges chain theorem (Theorem 35 in [4]), we see
that ether (PWr+ε, µ) ⊂ B(H, r) or B(H, r) ⊂ (PWr+ε, µ). Since detH = 1 almost
everywhere on [0, a], formula (17) implies the second alternative. Analogously, one
can show that (PWr−ε, µ) ⊂ B(H, r). Since this holds for every small number ε
and µ is sampling, we have B(H, r) = (PWr, µ). Finally, for r = a we have
B(H, a) = [0<r<a
B(H, r) = (PWa, µ),
where the completion is taken with respect to the norm inherited from L2(µ). (cid:3)
Let ΘH be the absolutely continuous solution of Cauchy problem (2) on [0, a],
1 )i. The reproducing kernel kB(H,r);λ at
and denote Θ+
a point λ ∈ C of the Hilbert space of entire functions B(H, r) has the the form
H = hΘH, ( 0
H = hΘH, ( 1
0 )i, Θ−
Θ+
H
(r, z)Θ−
H
kB(H,r);λ =
1
π
(r, ¯λ) − Θ−
H
z − ¯λ
(r, z)Θ+
H
(r, ¯λ)
,
z ∈ C.
(18)
MUCKENHOUPT HAMILTONIANS
10
The Paley-Wiener space PWr is the de Branges space B(H0, r) for the Hamiltonian
H0 = ( 1 0
0 1 ). The reproducing kernel of PWr at λ ∈ C will be denoted by sincr,λ:
sincr,λ =
sin r(z − ¯λ)
π(z − ¯λ)
,
z ∈ C.
Using integration by parts and equation (2), it is easy to show that for each λ ∈ C
we have
WH,rΘH(·, ¯λ) = √πkB(H,r);λ,
WH0,rΘH0(·, ¯λ) = √π sincr,λ,
where ΘH(·, ¯λ) denotes the mapping t 7→ ΘH(t, ¯λ) and ΘH0(·, ¯λ) is defined analo-
gously.
Next assertion is Lemma 4.2 in [2].
Lemma 3.4. Let µ be a sampling measure for PWa and let r ∈ [0, a]. The repro-
ducing kernel of the space (PWr, µ) at λ ∈ C equals T −1
Proof. For every function f in (PWr, µ) ⊂ PWr and every λ ∈ C we have
µ,r sincr,λ.
f (λ) = (f, sinca,λ)L2(R) = (f, T −1
µ,r sincr,λ)L2(µ),
where we used the fact that c1I 6 Tµ,a 6 c2I on PWa and hence Tµ,r is bounded
and invertible on PWr.
(cid:3)
Lemma 3.5. Let ϕ be a function on [0, a] such that eϕ ∈ L1[0, a]. Assume that a
spectral measure µ of problem (2) for the canonical Hamiltonian system generated by
H =(cid:0) eϕ 0
0 e−ϕ(cid:1) satisfies (1) for some constants c1, c2. Then function w = eϕ belongs
to the Muckenhoupt class A2[0, a] and kwkA2[0,2] 6 228c14, where c = c−1
We also have 1
1 + c2
2c−1
1 .
0 (w + 1
w ) dx 6 4c.
aR a
Proof. Let us obtain estimates (6), (7) for the function ϕ as it was suggested in
Proposition 3.2 of [2]. Take r ∈ [0, a]. Set H0 = ( 1 0
0 1 ) and consider the correspond-
ing Weyl-Titchmarsch transforms
WH0,r : L2(H0, r) → B(H0, r),
WH,r : L2(H, r) → B(H, r).
We have B(H0, r) = PWr and B(H, r) = (PWr, µ), see Lemma 3.3. Since µ
satisfies (1), the spaces PWr, (PWr, µ) coincide as sets and
1 kfk2
L2(R) 6 kT −1
L2(µ) 6 c−1
c−1
2 kfk2
µ,r fk2
L2(R)
for every function f ∈ PWr. Hence, the operator T = W−1
L2(H0, r) to L2(H, r) is correctly defined, bounded, and invertible. Moreover,
H,rT −1
µ,rWH0,r from
(19)
c−1
2 kXk2
L2(H,r) 6 c−1
L2(H0,r)
for every X ∈ L2(H0). Next, by Lemma 3.4 for each z ∈ C we have
µ,r sincr,¯z(cid:1) = ΘH(·, z).
For z = 0 and all t ∈ [0, r] we have ΘH(t, 0) = ΘH0(t, 0) = ( 1
L2(H,r) 6 c−1
0 )k2
L2(H0,r) 6 kT Xk2
H,r(cid:0)√πT −1
0 )k2
T ΘH0(·, z) = W−1
L2(H0,r) 6 k ( 1
0 ), hence
L2(H0,r).
1 kXk2
c−1
2 k ( 1
1 k ( 1
0 )k2
This relation is inequality (7) for the function ϕ and constants b1 = c−1
2 , b2 = c−1
1 .
MUCKENHOUPT HAMILTONIANS
11
Now let ∂n
from C to L2(H, r) at the point z = 0. Then T ∂n
integers n > 1. The right inequality in (19) yields
0 ΘH(·, 0) denote the derivative of order n of the mapping z 7→ ΘH(·, z)
0 ΘH0(·, 0) for all
0 ΘH(·, 0) = ∂n
(20)
(21)
(22)
∂n
0 ΘH(·, 0)k2
k∂n
From equation (2) we obtain
. . .Z tn−1
0 ΘH(t, 0) = n!Z t
0Z t1
for all t ∈ [0, r] and n > 1. Observe that
0
∂n
0
J∗H(t1)J∗H(t2) . . . J∗H(tn) ( 1
L2(H,r) 6 c−1
1 k∂n
0 ΘH0(·, 0)k2
L2(H0,r).
0 ) dtn . . . dt1,
J∗H(t1)J∗H(t2) . . . J∗H(tn) ( 1
0 ΘH0(t, 0) = J∗n(cid:0)tn
0(cid:1) ,
exp(Gϕ,n(t))(cid:17) , n is odd,
0 ) =
(cid:16)
(cid:16) (−1)
n
2 exp(Gϕ,n(t))
(cid:17) ,
n is even,
(−1)
n+3
0
0
2
where t = (t1, . . . , tn) is a point in simplex Kt,n, and Gϕ,n is defined on Kt,n by
formula (5). Substitute this representation of J∗H(t1)J∗H(t2) . . . J∗H(tn) ( 1
0 ) to
(21). Then (21), (22), and (20) give us inequality (6) for all n > 1 and all r ∈ [0, a].
It remains to use Proposition 2.1 to see that w ∈ A2[0, a] and kwkA2[0,a] 6 228c14.
The estimate 1
w ) dx 6 4c follows from Lemma 2.1.
0 (w + 1
(cid:3)
aR a
3.3. Proof of Theorem 1. Let µ be a measure on R such that estimate (1) holds
for some a > 0. Consider the truncated Toeplitz operator Tµ,a = Tµ on PWa.
We have c1I 6 Tµ 6 c2I, where I stands for the identity operator on PWa. The
operator Tµ−c1I satisfies assumptions of Lemma 3.1. Hence, there exists a measure
ν > 0 on R such that Tν = Tµ − c1I. One can suppose that ν is even (otherwise
consider the measure ν such that ν(S) = 1
2 (ν(S) + ν(−S)), and note that Tν = Tν).
Define a sequence of measures µj by µj = c1m + χjν, where m is the Lebesgue
measure on R, and χj denotes the indicator function of the interval [−j, j]. For
every j > 1 the measure µj is even and satisfies relation (1) with the same constants
c1, c2. Indeed, Tµj = c1I + Tχj ν and
c1I 6 c1I + Tχj ν 6 c1I + Tν = Tµ 6 c2I.
1
1
0
2c−1
1 + c2
0 (w+ 1
By Lemma 3.2 and Lemma 3.5, for every j there exists a smooth function wj > 0
on the interval [0, a] such that kwjkA2[0,a] 6 228c14, c = c−1
1 , and µj is
the spectral measure for the Hamiltonian Hj = (cid:16) wj 0
aR a
w )dx 6 4c for all j > 1. This allows us to use "a reverse Hölder inequality"
for weights in A2[0, a]. It says that for every C1 > 0 there exist p > 1 and C2 > 0
such that for all h ∈ A2[0, a] with khkA2[0,a] 6 C1 we have
h(x) dx(cid:19)p
wj(cid:17) on [0, a]. We also have
aZ a
that sequences {wj}j>1, (cid:8) 1
w, v, correspondingly. To simplify notations, let the sequences {wj}j>1, (cid:8) 1
h(x)p dx 6 C2(cid:18) 1
wj(cid:9)j>1 are informly bounded in Lp[0, a] for some p > 1.
wj(cid:9)j>1
themselves be weakly convergent. Let us show that v = w−1 almost everywhere
Explicit relations between C1, C2, and p can be found in [19]. From here we see
Hence we can find subsequences wjk , w−1
jk
converging weakly in Lp[0, a] to functions
aZ a
1
0
0
.
MUCKENHOUPT HAMILTONIANS
12
on the interval [0, a]. This is not always the case for arbitrary weakly convergent
sequences in Lp[0, a].
For z ∈ C denote by Θj(·, z) solution of equation (2) for the Hamiltonian Hj.
Integrating (2), we get
0 Hj(t)Θj (t, z) dt.
Then for every j > 1 and r, r′ ∈ [0, a] we have the estimates
JΘj(r, z) − ( 1
0 ) = zZ r
kΘj(r, z)kC2 6 exp(cid:18)zZ a
kΘj(r, z) − Θj(r′, z)kC2 6 z · r − r′
0 kHj(t)k dt(cid:19) ,
p (cid:18)Z a
p−1
0 kHj(t)kp · kΘj(t, z)kp
C2 dt(cid:19)
(23)
1
p
,
showing that functions Θj(·, z) are uniformly bounded and equicontinuous on [0, a].
Therefore, there is a subsequence of the sequence Θj(·, z) converging uniformly on
[0, a] to a function Θ(·, z). As before, we suppose that the sequence Θj(·, z) itself
It is clear that the limit function Θ satisfies
is uniformly convergent on [0, a].
equation (23) for the Hamiltonian H = ( w 0
0 v ). Hence, it satisfies equation (2)
for H. Fix a number r ∈ (0, a]. For every λ and z in C we have
µj ,r sincr,λ)(z) = (T −1
kB(H,r);λ(z) = lim
j→∞
Indeed, the first equality above follows from formula (18) and convergence of Θj to
Θ on [0, a] when a spectral parameter (¯λ or z) is fixed. Lemma 3.3 and Lemma 3.4
give us the second equality. Finally, using the fact that the operators Tµj ,r on PWr
tend to Tµ,r in the strong operator topology, we obtain the last equality in (24).
From (24) we see that Hilbert spaces of entire functions B(H, r), (PWr, µ) have the
same reproducing kernels. Hence B(H, r) = (PWr, µ) and formula (17) implies
kB(Hj ,r);λ(z) = lim
j→∞
µ,r sincr,λ)(z). (24)
(T −1
r =Z r
0 pdetH(t) dt,
r ∈ [0, a].
It follows that detH = 1 almost everywhere on [0, a], that is, v = w−1. Next,
from the direct spectral theory we know that the family {Θ(·, λ)}λ∈C is complete
in L2(H, a) and WH,aΘ(·, λ) = kB(H,a);λ for every λ ∈ C, where WH,a denotes the
Weyl-Titchmarsch transform associated to H. Using (24) again, we get
µ,a sinca,λ, sinca,z)L2(R)
(Θ(·, λ), Θ(·, z))L2(H,a) = πkB(H,a);λ(z) = π(T −1
= π(T −1
µ,a sinca,λ, T −1
µ,a sinca,z)L2(µ) = (WH,aΘ(·, λ),WH,aΘ(·, z))L2(µ).
Hence, the operator WH,a acts isometrically from L2(H; a) to L2(µ) and µ is a
spectral measure for H.
In particular, we can apply Lemma 3.5 to H, µ, and
conclude that the function w = eϕ is in A2[0, a] and kwkA2[0,a] 6 228c14. Uniqueness
of the Hamiltonian H follows immediately from formula (24):
Z r
0 )i dt = πkB(H,a);0(0) = πkT −1
where the right hand side is completely determined by µ, while the left hand side
determines H.
(cid:3)
w(t) dt =Z r
µ,r sincr,0 k2
0 hH(t) ( 1
0 ) , ( 1
L2(µ),
(25)
0
MUCKENHOUPT HAMILTONIANS
13
Differentiating formula (25), we obtain the following corollary.
Corollary 1. The Hamiltonian H = (cid:16) w 0
from µ by means of the following formula: w(r) = π ∂
0 1
w (cid:17) in Theorem 1 could be recovered
L2(µ), r ∈ [0, a].
µ,r sincr,0 k2
∂rkT −1
4. Proof of Theorem 2 and Theorem 3
Let us first show that Theorem 2 does not follow from a general theory of canon-
ical Hamiltonian systems. Consider the simplest case where the Hamiltonian H
coincides with the identity matrix ( 1 0
0 1 ) on [0, a]. We claim that there is no uni-
tary operator U : L2(H, a) → PW[0,2a] such that U L2(H, r) = PW[0,2r] for all
r ∈ [0, a]. Indeed, existence of such a unitary operator yields the existence of an-
other unitary operator U : L2[−a, a] → L2[0, 2a] such that U L2[−r, r] = L2[0, 2r]
for all r ∈ [0, a]. For every r1 > r2 > 0 let χ[r1,r2] denote the indicator function
of the interval [r1, r2]. Put g = U χ[0,a] and consider decomposition g = fr + hr,
where fr = U χ[0,r], hr = U χ[r,a], r ∈ [0, a]. Since U L2[−r, r] = L2[0, 2r] by our
assumption, the function fr is supported on [0, 2r]. Note also that the function hr
is orthogonal to all functions from L2[0, 2r] and hence it is supported on [2r, 2a].
From here we see that fr = χ[0,2r]g for all r ∈ [0, a]. Next, unitarity of the operator
U implies that
Z 2r
0
g(t)2 dt =Z 2r
0
fr(t)2 dt =Z a
−a χ[0,r](t)2 dt = r,
r ∈ [0, a].
It follows that g(t)2 = 1/2 for almost all t ∈ [0, 2a]. In particular, the linear span
of functions fr ∈ U L2[0, a], r ∈ [0, a], is dense in L2[0, 2a]. This contradicts to
the fact that U is a unitary operator from L2[−a, a] to L2[0, 2a]. Thus, the Weyl-
Titchmarsh transform WH,a from formula (3) can not be used to construct the
operator Fµ from Theorem 2 by means of superpositon with some simple unitary
operators like shifts, reflections, etc.
The main point that helps in proof of Theorem 2 is the fact that Hamiltonian H
generated by an even sampling measure for the Paley-Wiener space PWa must have
rank two almost everywhere on its domain of definition. It is an open question if
this is true for general (not necessarily even) sampling measures for PWa. See also
Proposition 5.1 in Section 5 for more details.
Proof of Theorem 2. Fix an even sampling measure µ and construct the Hamilto-
nians Hj , H, on [0, a] as in the proof of Theorem 1. Put ϕj = log wj and ϕ = log w,
where wj, w are the functions generating Hj, H. Recall that wj tend to w weakly
in Lp[0, a] for some p > 1 and the same is true for w−1
and w−1. Let Θj, Θ be the
solutions of system (2) generated by Hamiltonians Hj, H, correspondingly. As we
have seen, the functions Θj(·, z) = (cid:16) Θ+
j (cid:17) converge uniformly to Θ(·, z) = (cid:16) Θ+
Θ−(cid:17)
on the interval [0, a] when z ∈ C is fixed. For r ∈ [0, a], define entire functions P2r,j
j
Θ−
j
MUCKENHOUPT HAMILTONIANS
14
and P ∗2r,j by
P2r,j : z 7→ eirz(cid:18)e
P ∗2r,j : z 7→ eirz(cid:18)e
ϕj (r)
2 Θ+
ϕj (r)
2 Θ+
j (r, z) − ie−
j (r, z) + ie−
ϕj (r)
2 Θ−j (r, z)(cid:19) ,
2 Θ−j (r, z)(cid:19) ,
ϕj (r)
and let P2r, P ∗2r be defined similarly with ϕj replaced by ϕ. These functions satisfy
the Krein system of differential equations:
(P ′r,j(z) = izPr,j(z) +
P ∗r,j′ (z) =
j(r/2)
ϕ′
4
Pr,j(z),
ϕ′
j (r/2)
4
P ∗r,j(z), P0,j (z) = e
P ∗0,j (z) = e
ϕj (0)
2
ϕj (0)
2
,
,
(26)
where ϕ′j (r/2) is the value of smooth function ϕ′j at r/2. From system (26) we obtain
by integration by parts (see Lemma 9.1 in [5]) the Christoffel-Darboux formula:
Z r
0
Pt,j(z)Pt,j(λ) dt = i
P ∗r,j(z)P ∗r,j(λ) − Pr,j(z)Pr,j(λ)
z − ¯λ
.
The right hand side could be rewritten in the form
. . . = 2ei r
2 (z−¯λ) ·
Θ+
j ( r
2 , z)Θ−j ( r
2 , λ) − Θ−j ( r
2 , z)Θ+
j ( r
2 , λ)
,
z − ¯λ
which tends to 2πkr,λ(z), the scalar multiple of the reproducing kernel kr,λ at λ of
the Hilbert space ei r
2 zB(H, r
2 ) = (PW[0,r], µ), see formula (18). On the other hand,
for every pair z, λ ∈ C we have
2 (z−¯λ)(cid:16)eϕj( t
Pt,j (z)Pt,j(λ) = ei t
2 , λ) + e−ϕj( t
2 , z)Θ−j ( t
2 )Θ−j ( t
2 , λ)+
2 )Θ+
j ( t
2 , z)Θ+
2 , z)Θ−j ( t
j ( t
2 , λ) − iΘ−j ( t
+ iΘ+
j ( t
2 , z)Θ+
j ( t
2 , λ)(cid:17).
Since functions eϕj , e−ϕj converge weakly in Lp[0, a] to functions eϕ, e−ϕ, corre-
spondingly, we see that
Z r
0
Pt(z)Pt(λ) dt = lim
j→∞Z r
0
Pt,j (z)Pt,j(λ) dt = 2πkr,λ(z)
(27)
for every r ∈ [0, 2a]. Let χr be the indicator function of the interval [0, r]. Denote
by L the set of all finite linear combinations of functions t 7→ χr(t)Pt(z) on [0, 2a],
where z ∈ C and r ∈ [0, 2a]. The linear manifold L is dense in L2[0, 2a]. Indeed,
for every function g ∈ L2[0, 2a] orthogonal to L we have
0 =Z 2a
0
g(t)χr(t)Pt(0) dt =Z r
0
g(t)e
ϕ(t/2)
2
dt,
r ∈ [0, 2a],
yielding g = 0 in L2[0, 2a]. Formula (27) also shows that a nontrivial finite linear
combination of functions χr(t)Pt(z) cannot vanish almost everywhere on [0, 2a].
Consider the operator Fµ : L2[0, 2a] → (PW[0,2a], µ) densely defined on L by
Fµ : f 7→
1
√2π Z 2a
0
f (t)Pt(z) dt,
z ∈ C.
(cid:16)Fµχr1 Pt(λ),Fµχr2 Pt(z)(cid:17)L2(µ)
= 2π(kr1,λ, kr2,z)L2(µ) = 2πkr,λ(z),
=(cid:16)χr1 Pt(λ), χr2Pt(z)(cid:17)L2[0,2a]
,
where r = min(r1, r2). This shows that Fµ is an isometry on L. Since the linear
span of the set {k2a,λ, λ ∈ C} is complete in (PW[0,2a], µ), the operator Fµ is
unitary. It is also clear from the definition that Fµ maps L2[0, r] onto (PW[0,r], µ)
for every r ∈ [0, 2a].
Proof of Theorem 3. At first, consider a positive bounded invertible operator
Wψ with real symbol ψ ∈ S′ on a finite interval [0, a]. Let F denote the unitary
Fourier transform on L2(R). Take a smooth function h with support in (0, a)
and put f = F f . Consider the operator Wψ = F WψF−1 on PW[0,a]. We have
( Wψh, h)L2(R) = (cid:10) ψ,h2(cid:11)S ′ , where ψ is the Fourier transform of the tempered
distribution ψ. It follows that
(cid:3)
( Wψf, f )L2(R) = ( Wψ
z−i
z+i f, z−i
z+i f )L2(R)
on a dense subset of the set Z−i = {f ∈ PW[0,a] : f (−i) = 0}. Since Wψ is
bounded on PW[0,a], we have the last identity for all f ∈ Z−i. Hence, the operator
Wψ satisfies assumptions of Lemma 3.1 and we can find a positive Borel measure µ
on R such that
MUCKENHOUPT HAMILTONIANS
15
The operator Fµ takes the function t 7→ χr(t)Pt(λ) on [0, 2a] into √2πkr,λ, see
formula (27). Moreover, for every r1, r2 ∈ [0, 2a] we have
( Wψf, g)L2(R) =ZR
f ¯g dµ
for all f, g ∈ PW[0,a]. As in the proof of Theorem 1, we can assume that the
Indeed, since ψ is real, we have ( Wψf, f ) = ( Wψf∗, f∗) for
measure µ is even.
arbitrary f ∈ PW[0,a] and its reflection f∗ : x 7→ f (−x). By the assumption, the
operator Wψ is positive, bounded and invertible on PW[0,a]. Hence the measure
µ satisfies (1) for some c1, c2 and a/2 in place of a. By Theorem 2, there is a
unitary operator Fµ : L2[0, a] → (PW[0,a], µ) such that Fµ : L2[0, r] = (PW[0,r], µ)
for every r ∈ [0, a]. Identifying Hilbert spaces (PW[0,a], µ) and PW[0,a] as sets, we
can define the operator A = F−1
µ F on L2[0, a]. By construction, the operator A is
bounded and invertible and AL2[0, r] = L2[0, r] for every r ∈ [0, a]. We also have
(28)
dµ = (F h,F h)L2(µ) = (F−1
µ F h,F−1
µ F h)L2[0,a]
2
for all smooth functions h with support in (0, a). It follows that the operator Wψ
admits the triangular factorization Wψ = A∗A.
(Wψh, h)L2[0,a] =ZR(cid:12)(cid:12)
h(cid:12)(cid:12)
It remains to consider the case where Wψ is a positive bounded invertible Wiener-
Hopf operator on L2[0,∞) with real symbol ψ ∈ S′. It is known (see Section 4.2.7
in [14]) that in this case the Fourier transform of the distribution ψ coincides with
a function σ on R such that c1 6 σ(x) 6 c2 for some positive constants c1, c2
and almost all x ∈ R.
In particular, the measure µ = σ dm is sampling for all
Paley-Wiener spaces PW[0,r], r > 0. Since ψ is real, the function σ is even. For
every r > 1 we can use Theorem 1 and find a Hamiltonian Hr on [0, r] such that
detHr(t) = 1 for almost all t ∈ [0, r] and µ is the spectral measure for Hr. Since
the Hamiltonian H in Theorem 1 is defined uniquely, we have Hr(t) = Hr′(t) for
MUCKENHOUPT HAMILTONIANS
16
almost all t ∈ [0, min(r, r′)]. This shows that there is the Hamiltonian H on [0,∞)
such that det H = 1 almost everywhere and µ is the spectral measure for H. In
particular, we can define a family of entire functions {Pt}t>0 such that the mapping
(29)
f (t)Pt(z) dt
1
Fµ : f 7→
√2π Z r
0
µ(C+) = (f, g)L2(µ). Since c1 6 σ 6 c2 on R, the space H 2
sends unitarily the space L2[0, r] onto the space (PW[0,r], µ) for every r > 0, see
the proof of Theorem 2. Let H 2
µ(C+) be the weighted Hardy space with the inner
product (f, g)H2
µ(C+)
coincides as a set with the standard Hardy space H 2(C+) = F L2[0,∞). Define the
unitary operator Fµ from L2[0,∞) to H 2
µ(C+) by formula (29) with r = ∞ on the
dense set of compactly supported bounded functions in L2[0,∞). Then the operator
A = F−1
µ F on L2[0,∞) is bounded and invertible. Moreover, AL2[0, r] = L2[0, r]
for every r > 0, and Wψ = A∗A, see formula (28).
Remark. It can be shown that positive bounded invertible Wiener-Hopf operators
Wψ on L2[0, a) with real symbols ψ ∈ S′ admit triangular factorisation in the reverse
order, Wψ = AA∗.
In the case a = ∞ the classical Wiener-Hopf factorization
works: one can take A = F−1TϕσF , where Tϕσ is the Toeplitz operator on H 2(C+)
with analytic symbol ϕσ such that ϕσ2 = σ = F ψ. If a > 0 is finite, then we
can use Theorem 3 to find left triangular factorization Wψ = A∗ A and then put
A = Ca ACa, where Ca : f 7→ f (a − x) is the conjugate-linear isometry on L2[0, a].
Since CaWψCa = Wψ for the self-adjoint Wiener-Hopf operator Wψ on L2[0, a], and
a = I, we have Wψ = AA∗. It is also clear that the operator A is upper-triangular.
C2
(cid:3)
5. Appendix. Two results by L. A. Sakhnovich
In paper [17] L. A. Sakhnovich proved (see Theorem 4.1 and Remark 4.1 in [17])
that positive bounded invertible Wiener-Hopf operator
T : f 7→ f − µZ ∞
0
f (t)
sin π(t − x)
π(t − x)
dt,
f ∈ L2[0,∞),
0 < µ < 1,
(30)
densely defined on L2[0,∞) does not admit triangular factorization T = A∗A,
where a bounded invertible operator A on L2[0,∞) is such that AL2[0, r] = L2[0, r]
for every r > 0. Clearly, this assertion contradicts Theorem 3. Let us point out an
error in its proof.
The argument in [17] crucially uses the following claim. Let χ[−π,π] be the
indicator function of the interval [−π, π]. Formulas (4.1) − (4.4) in [17] for n = 0
and a0 = π determine the function σ′ : x 7→ 1
2π (1 − µ · χ[−π,π](x)) on R. The
function
1 + tz
Π(z) =
1
√2π
exp(cid:18) 1
2iπ Z ∞
−∞
(z − t)(1 + t2)
log σ′(t) dt(cid:19)
from formula (4.10) of [17] (see also formula (4.12) therein) is claimed to satisfy the
following relation (formula (4.18) in [17]):
However, this fact is false. Indeed, we have
lim
y→+0
Π(iy) =p1 − µ.
πi(cid:18) 1
t − z −
= −
1
1
πi
1 + tz
(z − t)(1 + t2)
t
1 + t2(cid:19)
MUCKENHOUPT HAMILTONIANS
17
and hence √2πΠ(z) is the outer function in C+ whose absolute value on R coincides
with (σ′)−1/2 almost everywhere on R. Since (σ′)−1/2 is regular (in fact, constant)
near the origin, we have
Π(iy) =
lim
y→+0
1
√2π
(σ′)−1/2(0) =
1
.
√1 − µ
We also would like to note that the last relation agrees well with the first identity
in formula (4.19) from [17].
The second part of this section concerns factorization problem for truncated
Toeplitz operators generated by general sampling measures for the space PWa
not necessarily symmetric with respect to the origin. The result is equivalent to
Theorem 4.2 in [16]. The proof below seems to be a bit more straightforward than
the original one, possibly, because we consider the one-dimensional situation.
that µ satisfies (1). The following assertions are equivalent:
Proposition 5.1. Let H be a Hamiltonian on [0, ℓ] such that R ℓ
and let µ be a spectral measure for problem (2). Set a =R ℓ
0 traceH(r) < ∞,
0 pdetH(r) dr. Assume
(a) detH > 0 almost everywhere on [0, ℓ];
(b) there exists a unitary operator Vµ : PWa → (PWa, µ) such that for every
r ∈ [0, a] we have VµPWr = (PWa, µ).
(c) there exists a bounded invertible operator A on PWa such that Tµ,a = A∗A
and for every r ∈ [0, a] we have APWr = PWr.
Given a Hamiltonian H on [0, ℓ] such that a = R ℓ
continuous from the left function ξH from [0, a] to [0, ℓ] by
0 pdetH(t) dt > 0, we define
r ∈ [0, a].
r =Z ξH(r)
0
pdetH(t) dt,
This function is continuous if and only if there are no interval (r1, r2) ⊂ [0, ℓ]
such that detH(t) = 0 for almost all t ∈ (r1, r2). The function ξH is absolutely
continuous if and only if detH(t) > 0 for almost all t ∈ [0, a], see Exercise 13 in
Chapter IX of [13].
Proof of Proposition 5.1. (a) ⇒ (b). Since detH > 0 almost everywhere on the
interval [0, ℓ], the function ξ = ξH is absolutely continuous and
ξ′(r) =
1
pdetH(ξ(r))
for almost all r ∈ [0, a]. Consider the Hamiltonian H : r 7→ ξ′(r)H(ξ(r)) on the
interval [0, a]. We have det H = 1 and Θ H
(r, z) = ΘH(ξ(r), z) on [0, a]. Changing
variable in (3), we see that B( H, r) = B(H, ξ(r)) for every r ∈ [0, a], hence µ is the
spectral measure for H. Consider the Weyl-Titchmarsh transforms generated by
Hamiltonians H and H0 = ( 1 0
0 1 ), correspondingly,
W H,a : L2( H, a) → B( H, a),
Define the operator Vµ : PWa → B( H, a) by Vµ = W H,a
H0,a, where
M H−1/2 : L2(H0, a) → L2( H, a) is the multiplication operator by H−1/2, that is,
M H−1/2 : X 7→ H−1/2X. Since M H−1/2 is unitary, the operator Vµ is unitary as
WH0,a : L2(H0, a) → PWa.
M H−1/2W−1
MUCKENHOUPT HAMILTONIANS
18
well. It is also clear that VµPWr = B( H, r) for every r ∈ [0, a]. Using Lemma 3.3,
we see that B( H, r) = (PWr, µ), as required.
(b) ⇒ (a). We will show that the function ξ = ξH is absolutely continuous. Let
χr be the indicator function of an interval [0, r]. For every r ∈ [0, a] consider the
1 ) in L2(H, ξ(r)). A straightforward
functions Xξ(r) = χξ(r) ( 1
modification of Lemma 3.3 gives B(H, ξ(r)) = (PWr, µ) for all r ∈ [0, a]. Put
0 ), Yξ(r) = χξ(r) ( 0
r = W−1
X 0
H0,aV −1
µ WH,aXξ(r),
r = W−1
Y 0
H0,aV −1
µ WH,aYξ(r).
0
0 ) , ( 1
r = χrX 0
a and Y 0
r = χrY 0
Z ξ(r)
Since Vµ is isometric and VµPWr = (PWr, µ), we have Pµ,rVµ = VµPr, where Pr,
Pµ,r are the orthogonal projections on PWa, (PWa, µ), with ranges PWr, (PWr, µ),
respectively. It follows that X 0
a . Using the fact that the
operators WH0,a, WH,a are unitary, we obtain
traceH(t) dt =Z ξ(r)
(cid:16)(cid:10)H(t) ( 1
L2(H,ℓ) + kYξ(r)k2
= kXξ(r)k2
L2(H0,a) + kY 0
rk2
= kX 0
= kχrX 0
ak2
=Z r
0 (cid:16)kX 0
κ(s) =Z s
The above equalities hold for all r ∈ [0, a]. Let us define the function κ on [0, ℓ] by
L2(H0,a) + kχrY 0
a k2
L2(H0,a),
C2(cid:17) dt.
a (t)k2
a(t)k2
0 )(cid:11) +(cid:10)H(t) ( 0
L2(H,ℓ),
r k2
L2(H0,a),
1 )(cid:11)(cid:17) dt,
traceH(t) dt,
C2 + kY 0
s ∈ [0, ℓ].
1 ) , ( 0
0
0
Then κ is an absolutely continuous function with positive derivative almost every-
where on [0, ℓ], hence the inverse mapping κ−1 is also absolutely continuous and
has positive derivative. On the other hand, formula (31) shows that κ(ξ) is an
absolutely continuous function. It follows that the superposition ξ = κ−1(κ(ξ)) is
absolutely continuous and hence detH > 0 almost everywhere on [0, ℓ].
(b) ⇒ (c). Since µ satisfies (1), the identical embedding j : PWa → (PWa, µ) is a
bounded and invertible operator. Define A = V −1
µ j. Then for all f, g in PWa we
have
(31)
(A∗Af, g) = (V −1
µ jf, V −1
µ jg)L2(R) = (jf, jg)L2(µ) =ZR
f ¯g dµ = (Tµ,af, g),
(32)
by the unitarity of the operator Vµ. It follows that Tµ,a = A∗A. By construction,
the operator A is invertible. We also have APWr = PWr for all r ∈ [0, a], hence A
is upper-triangular.
(c) ⇒ (b). Assume that Tµ,a admits a left triangular factorization Tµ,a = A∗A.
Define the operator Vµ : PWa → (PWµ, a) by Vµ = jA−1, where j is the embedding
from PWa to (PWa, µ). Then VµPWr = (PWr, µ) for every r ∈ [0, a] and
(Vµf, Vµg)L2(µ) = ((A−1)∗j∗jA−1f, g)L2(R)
= ((A∗)−1Tµ,aA−1f, g)L2(R) = (f, g)L2(R),
where we used the identity Tµ,a = j∗j, see (32). Since A and j are invertible, Vµ is
a unitary operator.
(cid:3)
MUCKENHOUPT HAMILTONIANS
19
References
[1] Anton Baranov, Roman Bessonov, and Vladimir Kapustin. Symbols of truncated Toeplitz
operators. J. Funct. Anal., 261(12):3437 -- 3456, 2011.
[2] R. V. Bessonov and R. V. Romanov. An inverse problem for weighted Paley-Wiener spaces.
preprint arXiv:1509.08117, 2015.
[3] Louis de Branges. Some Hilbert spaces of entire functions. ii. Trans. Amer. Math. Soc.,
99:118 -- 152, 1961.
[4] Louis de Branges. Hilbert spaces of entire functions. Prentice-Hall, Inc., Englewood Cliffs,
N.J., 1968.
[5] Sergey A. Denisov. Continuous analogs of polynomials orthogonal on the unit circle and Kreın
systems. IMRS Int. Math. Res. Surv., pages Art. ID 54517, 148, 2006.
[6] I. M. Gelfand and B. M. Levitan. On the determination of a differential equation from its
spectral function. American Mathematical Society, 1955.
[7] I. C. Gohberg and M. G. Krein. Theory and Applications of Volterra Operators in Hilbert
Space, volume 24 of Translations of Mathematical Monographs. American Mathematical So-
ciety, 1970.
[8] I. S. Kac and M. G. Krein. On the spectral functions of the string. Amer. Math. Soc. Transl,
103(2):19 -- 102, 1974.
[9] M. G. Krein. On a fundamental approximation problem in the theory of extrapolation and
filtration of stationary random processes. Dokl. Acad. Nauk SSSR, 94:13 -- 16, 1954.
[10] M. G. Krein. Continuous analogues of propositions on polynomials orthogonal on the unit
circle. Dokl. Akad. Nauk SSSR (N.S.), 105:637 -- 640, 1955.
[11] David R. Larson. Nest algebras and similarity transformations. Annals of Mathematics,
121(2):409 -- 427, 1985.
[12] Vladimir Aleksandrovich Marchenko. Sturm-Liouville operators and applications, volume
373. American Mathematical Soc., 2011.
[13] I. P. Natanson. Theory of functions of a real variable, volume 1. Ungar, New York, 1964.
Translated by L. F. Boron, with the editorial collaboration of and with annotations by E.
Hewitt.
[14] Nikolai K. Nikolski. Operators, functions, and systems: an easy reading. Vol. 1, volume 92
of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI,
2002. Hardy, Hankel, and Toeplitz, Translated from the French by Andreas Hartmann.
[15] Roman Romanov. Canonical systems and de Branges spaces. preprint arXiv:1408.6022, 2014.
[16] Lev A. Sakhnovich. On triangular factorization of positive operators. In Recent Advances in
Matrix and Operator Theory, pages 289 -- 308. Springer, 2007.
[17] Lev A. Sakhnovich. Effective construction of a class of positive operators in Hilbert space,
which do not admit triangular factorization. In Levy Processes, Integral Equations, Statistical
Physics: Connections and Interactions, pages 85 -- 99. Springer, 2012.
[18] Donald Sarason. Algebraic properties of truncated Toeplitz operators. Oper. Matrices,
1(4):491 -- 526, 2007.
[19] V. I. Vasyunin. The sharp constant in the reverse Hölder inequality for the Muckenhoupt
weights. Algebra i Analiz, 15(1):73 -- 117, 2003.
[20] Henrik Winkler. Operator Theory, chapter Two-Dimensional Hamiltonian Systems, pages
1 -- 22. Springer Basel, Basel, 2014.
St.Petersburg State University (7/9, Universitetskaya nab., St.Petersburg, 199034 Rus-
sia) and St.Petersburg Department of Steklov Mathematical Institute of Russian
Academy of Science (27, Fontanka, St.Petersburg, 191023 Russia)
E-mail address: [email protected]
|
1501.07409 | 1 | 1501 | 2015-01-29T10:37:59 | Continuous wavelet transform on local fields | [
"math.FA"
] | The main objective of this paper is to define the mother wavelet on local fields and study the continuous wavelet transform (CWT) and some of their basic properties. its inversion formula, the Parseval relation and associated convolution are also studied. | math.FA | math |
CONTINUOUS WAVELET TRANSFORM ON LOCAL FIELDS
ASHISH PATHAK
DEPARTMENT OF MATHEMATICS & STATISTICS
DR. HARISINGH GOUR CENTRAL UNIVERSITY
SAGAR-470003, INDIA.
Abstract. The main objective of this paper is to define the mother wavelet
on local fields and study the continuous wavelet transform (CWT) and some of
their basic properties. its inversion formula, the Parseval relation and associated
convolution are also studied.
1. Introduction
A local field means an algebraic field and a topological space with the topological
properties of locally compact, non-discrete, complete and totally disconnected,
denoted by K [7]. The additive and multiplicative groups of K are denoted by
K+ and K∗, respectively. We may choose a Haar measure dx for K+.
If α 6=
0(α ∈ K), then d(αx) is also a Haar measure. Let d(αx) = αdx and call α the
absolute value or valuation of α. Let 0 = 0. The absolute value has the following
properties:
(i) x ≥ 0 and x = 0 if and only if x = 0;
(ii) xy = xy;
(iii) x + y ≤ max(x, y).
The last one of these properties is called the ultrametric inequality. The set
Key words and phrases. Continuous Wavelet Transform , Local fields.
∗E-mail: pathak [email protected].
1
2
ASHISH PATHAK
D = {x ∈ K : x ≤ 1} is called the ring of integers in K. It is the unique maximal
compact subring of K. Define P = x ∈ K : x < 1. The set P is called the prime
ideal in K. The prime ideal in K is the unique maximal ideal in D. It is principal
and prime.
If K is a local field, then there is a nontrivial, unitary, continuous character χ on
K+ and K+ is self dual.
χ is fixed character on K+ that is trivial on D but is nontrivial on P−1 . It follows
that χ is constant on cosets of D and that if y ∈ Pk , then χy(χy(x)) = χ(xy) is
constant on cosets of P−k
Definition 1.1. The fourier transform of f ∈ L1(K) is denoted by f (ξ) and define
by the [8]
f (ξ) =ZK
f (x)χξ(x)dx =ZK
and the inverse Fourier transform by
f (x)χ(−ξx)dx, ξ ∈ K
(1.1)
f (x) =ZK
f (ξ)χx(ξ)dx x ∈ K
(1.2)
Some important properties of the Fourier transform can prove easily :
(i) f L∞(K) ≤ f L1(K).
(ii) If f ∈ L1(K), then f is uniformly continuous.
(iii) Parseval formula:If f ∈ L1(K) ∩ L2(K), then f L2(K) = f L2(K)
(iv) If the convolution of f and g is defined as
then
(f ∗ g)(t) =ZK
f (x)g(t − x)dx
F ((f ∗ g)) = F (f ).F (g)
(1.3)
(1.4)
CONTINUOUS WAVELET TRANSFORM ON LOCAL FIELDS
3
The article is divided in four sections. In section 2. we proposed the definition
of mother wavelet and define the continuous wavelet transform (CWT). In section
3. discus the some basic properties of CWT. In section 4. we prove the Plancherel
, inversion formula and define the convolution associated with CWT.
2. Continuous Wavelet transform on local fields
Similar to L2(R) [1, 3, 5], we define the wavelet on local fields and define the
continuous wavelet transform.
Definition 2.1. Admissible wavelet on local fields
The function ψ(x) ∈ L2(K) is said to be an admissible wavelet on local fields if
ψ(x) satisfies the following admissibility condition:
cψ =ZK
ψ(ξ)2
ξ
dξ < ∞
(2.1)
where ψ is the Fourier transform of ψ.
Remark 2.2. If ψ(ξ) is continuous near ξ = 0, then the existence of integral
(2.1) guarantees that , ψ(0) = 0. Since the Fourier transform of mother wavelet
ψ ∈ L1(K) ∩ L2(K) is bounded and uniformly continuous, we have
0 = ψ(0) = ZK
= ZK
ψ(x)χ0(x)dx
ψ(x)dx
This means that the integral of mother wavelet is zero:
Theorem 2.3. If ψ is a mother wavelet and φ ∈ L1(K), then the convolution
function ψ ∗ φ is a mother wavelet.
4
ASHISH PATHAK
Proof. Since
ZK
(ψ ∗ φ)(x)2dx = ZK(cid:12)(cid:12)(cid:12)ZK
≤ ZK(cid:18)ZK
≤ ZK(cid:18)ZK
= ZK
= (cid:18)ZK
φ(y)dyZKZK
φ(y)dy(cid:19)2ZK
= φ2
L1(K)ψ2
L2(K)
2
dx
ψ(x − y)φ(y)
ψ(x − y)φ(y)dy(cid:12)(cid:12)(cid:12)
ψ(x − y)φ(y)dyZK(cid:19) dx
2 φ(y)
1
1
2 dy(cid:19)2
dx
ψ(x − y)2φ(y)dydx
ψ(x)2dx
Therefore (ψ ∗ φ)(x) ∈ L2(K). Moreover
cψ∗φ = ZK
= ZK
ψ ∗ φ(ξ)2
ξ
dξ
ψ(ξ)2 φ(ξ)2
ξ
dξ
≤ φ2
L∞(K)ZK
ψ(ξ)2
ξ
dξ
This completes the proof of the theorem
(cid:3)
Definition 2.4. Continuous wavelet transform (CWT) on local fields
For ψ(x) ∈ L2(K) and a, b ∈ K, a 6= 0, we define the unitary linear operator:
U b
a : L2(K) → L2(K)
by
U b
a(ψ(x)) = ψa,b(x) =
1
1
2
a
ψ(
x − b
a
)
(2.2)
CONTINUOUS WAVELET TRANSFORM ON LOCAL FIELDS
5
ψ is called mother wavelet and ψa,b(x) are called daughter wavelets, where a is a
dilation parameter, b is a translation parameter.
The Fourier transform of ψa,b(x) is given by
ψa,b(ξ) = a
1
2 ψ(aξ)χb(ξ)
(2.3)
where ψ is the Fourier transform of ψ.
The CWT on local fields
Kψ : L2(K) → L2(K × K)
of a function f ∈ L2(K) with respect to a mother wavelet ψ is defined by
f 7→ Kψf (a, b) = (f, ψa,b)L2(K)
= ZK
= ZK
f (x)ψa,b(x)dx
f (x)
1
1
2
a
ψ(
x − b
a
)dx
(2.4)
3. Basic Properties of CWT on local fields
Before giving the fundamental properties of CWT, we list their basic properties.
Theorem 3.1. Let ψ and ϕ be to wavelets and f, g are two function belong to
L2(K), then
(1) Linearity
Kψ(ηf + ϑg)(a, b) = ηKψ(f )(a, b) + ϑKψ(g)(a, b)
(3.1)
where η and ϑ are any two scalers.
(ii)Shift property
Kψ(f (x − ς))(a, b) = Kψ(f )(a, b − ς)
(3.2)
6
ASHISH PATHAK
where ς is any scalers.
(iii) Scaling property If σ 6= 0 any scaler. The CWT of the scaled function
fσ(x) = 1
σ f ( 1
σ ) is
Kψ(fσ(x))(a, b) = Kψ(f )(
a
σ
,
b
σ
)
Kψ(f )(a, b) = Kf (ψ)(
1
a
, −
1
b
)
(iv) Symmetry
(iv) Parity
KP (ψ)(P f )(a, b) = Kψ(f )(a, −b)
where P is a parity operator define by P f (x) = f (−x).
(3.3)
(3.4)
(3.5)
Proof. The proof is the straight forward application of CWT
(cid:3)
Theorem 3.2. Show that the continuous wavelet transform can also expressed as
(Kψf )(a, b) = f ∗
where the ∗ is defined as
ψ(
x
a
)! (b)
1
pa
(3.6)
(f ∗ g)(t) =ZK
f (x)g(t − x)dx
(3.7)
Proof. From define of CWT we have
f (x)
1
1
2
a
ψ(
x − b
a
)dx
(Kψf )(a, b) = ZK
= f ∗
ψ(
x
a
)! (b)
1
pa
(3.8)
(cid:3)
CONTINUOUS WAVELET TRANSFORM ON LOCAL FIELDS
7
Theorem 3.3. if f is homogeneous function of degree n show that
(Kψf )(λa, λb) = λnλ
1
2 (Kψf )(a, b)
(3.9)
where the λ is scaler .
Proof. From define of CWT we have
(Kψf )(λa, λb) = ZK
= ZK
f (x)
1
λa
1
2
f (λx)
1
1
2
a
ψ(
ψ(
x − λb
λa
)dx
x − b
a
)λdx
= λnλ
1
2 (Kψf )(a, b)
(3.10)
(cid:3)
4. Main Properties of the CWT
This section describes important properties of the CWT, such as the Plancherel
, inversion formula and associated convolution fist, we establish the Plancherel
theorem.
Theorem 4.1. (QFT Plancherel) Let f, g ∈ L2(K). Then we have
(Kψ(f )(a, b), Kψ(g)(a, b))L2(K×K) = cψ(f, g)L2(K)
(4.1)
where cψ is given in (2.1).
8
ASHISH PATHAK
Proof. By using perseval formula for Fourier we can write the wavelet transform
as
Similarly
Kψ(f )(a, b) = ZK
f (x)
1
1
2
a
ψ(
x − b
a
)dx
= (f, ψa,b)
= ( f , ψa,b)
= ZK
f (ξ)a
1
2 ψ(aξ)χb(ξ)dξ
(4.2)
Kψ(g)(a, b) = ZK
f (ξ)a
1
2 ψ(aξ)χb(ξ)dξ
(4.3)
Now, by using above (4.2) and (4.3) we get
ZKZK
Kψ(f )(a, b)Kψ(g)(a, b)
a
dadb
a2 ZK
f (ξ) ψ(aξ)χb(ξ)dξ
g(υ) ψ(aυ)χb(υ)dυ
dadb
a ZK
f (ξ) ψ(aξ)χb(ξ)dξ
F ( f (ξ) ψ(aξ))(b)F (g(υ) ψ(aυ))(b)
f (ξ) ψ(aξ)g(ξ) ψ(aξ)
dξda
a
dadb
a2 = ZKZK
×ZK
= ZKZK
= ZKZK
= ZKZK
= ZK
= ZK
dadb
a
(4.4)
f (ξ)g(ξ)(cid:18)ZK
f (ξ)g(ξ) ZK
ψ(aξ) ψ(aξ)
ψ(aξ)2
a
da
a(cid:19) dξ
da! dξ
CONTINUOUS WAVELET TRANSFORM ON LOCAL FIELDS
9
= ZK
f (ξ)g(ξ) ZK
ψ(ω)2
ω
dz! dξ
= cψ( f , g)L2(K)
= cψ(f, g)L2(K)
Theorem 4.2. (Inversion Formula) Let f ∈ L2(K). Then we have
f (x) =
1
cψ ZKZK
Kψ(f )(a, b)ψa,b(x)
dadb
a2
where cψ is given in (2.1).
(4.5)
(cid:3)
(4.6)
Proof. Let h(x) ∈ L2(K) be any function, then by using above theorem, we have
cψ(f, g)L2(K) = ZKZK
= ZKZK
= ZKZKZK
= (cid:18)ZKZK
(Kψf )(a, b)ψa,b(x)
(Kψf )(a, b)Kψ(h)(a, b)
dadb
a2
(Kψf )(a, b)ZK
h(x)ψa,b(x)dx
dadb
a2
(Kψf )(a, b)ψa,b(x)h(x)
dadbdx
a2
dadb
a2
, h(x)(cid:19)
Hence the result follows.
If f = h
f 2
L2(K) =ZKZK
(Kψf )(a, b)2 dadb
a2
Moreover the wavelet transform is isometry from L2(K) to L2(K × K)
(cid:3)
(4.7)
10
ASHISH PATHAK
4.1. Associated convolution for CWT on local fields. Using Pathak and
Pathak techniques [5], we define the basic function D(x, y, z) , translation τx and
associated convolution # operators for CWT.
The basic function D(x, y, z) for (2.4) is define as
Kφ[D(x, y, z)](a, b) = ZK
D(x, y, z)φa,b(t)dt
= ψa,b(z) χa,b(y),
(4.8)
where ψ, φ and χ are three wavelets satisfying certain conditions (2.1).
Now, by using(4.6) we get,
D(x, y, z) = C −1
φ ZKZK
The translation τx is defined as [5]
ψa,b(z) χa,b(y) φa,b(x) a−2dadb.
(4.9)
(τxh)(y) = h∗(x, y) =ZK
φ ZKZKZK
= C −1
D(x, y, z)h(z)dz
ψa,b(z) χa,b(y) φa,b(x) h(z)a−2dadbdz.
The associated convolution is defined as
h∗(x, y)g(y)dy
(h#g)(x) = ZK
= ZKZK
= C −1
φ ZKZKZKZK
D(x, y, z) h(z) g(y)dydz
ψa,b(z) χa,b(y) φa,b(x)h(z)g(y) a−2 dadbdzdy,
(4.10)
by using the inversion formula we can write the above equation as
(h#g)(x) = C −1
(Kψh)(a, b)(Kχg)(a, b)φa,b(x)a−2dadb
φ ZKZK
= K −1
φ [(Kψh)(a, b)(Kχg)(a, b)] (x)
CONTINUOUS WAVELET TRANSFORM ON LOCAL FIELDS
11
so that
Kφ[h#g](b, a) = (Kψh)(a, b)(Kχg)(a, b)(x)
(4.11)
Acknowledgment
The work is supported by U.G.C start-up grant No:F.30-12(2014)/(BSR).
References
[1] Daubechies, Ten Lectures on Wavelets, in: CBMS/NSF Ser. Appl. Math., vol. 61, SIAM,
1992.
[2] L. Debnath, Wavelet Transforms and Their Applications, Birkhauser, Boston, 2002.
[3] C.K. Chui, An Introduction to Wavelets, Academic Press, 1992.
[4] M. Holschneider, Wavelet analysis over Abelian groups, Appl. Comput. Harmon. Anal. 2
(1995) 52-60.
[5] R. S. Pathak and Ashish Pathak . On convolution for wavelets transform ;international
journal of wavelets, multiresolution and information processing , (2008), 6(5): 739-747.
[6] D. Ramakrishnan and R. J. Valenza, Fourier Analysis on Number Fields, Graduate Texts
in Mathematics 186, Springer-Verlag, New York, 1999.
[7] H. Jiang, D. Li and N. Jin, Multiresolution analysis on local fields, J. Math. Anal. Appl.
294 (2004) 523- 532.
[8] M.H. Taibleson, Fourier Analysis on Local Fields, Princeton Univ. Press, 1975.
|
1006.3089 | 1 | 1006 | 2010-06-15T20:45:18 | Constructing non-compact operators into $c_0$ | [
"math.FA"
] | We prove that for each dense non-compact linear operator $S:X\to Y$ between Banach spaces there is a linear operator $T:Y\to c_0$ such that the operator $TS:X\to c_0$ is not compact. This generalizes the Josefson-Nissenzweig Theorem. | math.FA | math | CONSTRUCTING NON-COMPACT OPERATORS INTO c0
IRYNA BANAKH AND TARAS BANAKH
Abstract. We prove that for each dense non-compact linear operator S : X → Y between Banach
spaces there is a linear operator T : Y → c0 such that the operator T S : X → c0 is not compact. This
generalizes the Josefson-Nissenzweig Theorem.
By the Josefson-Nissenzweig Theorem [6], [7] (see also [2], [5, XII], and [3, 3.27]), for each infinite-
dimensional Banach space Y the weak∗ convergence and norm convergence in the dual Banach space
Y ∗ are distinct. This allows us to find a sequence (y∗
n)n∈ω of norm-one functionals in Y ∗ that converges
to zero in the weak∗ topology. Such functionals determine a non-compact operator T : Y → c0 that
assigns to each y ∈ Y the vanishing sequence (y∗
n(y))n∈ω ∈ c0. Thus each infinite-dimensional Banach
space Y admits a non-compact operator T : Y → c0 into the Banach space c0.
The following theorem (which is a crucial ingredient in the topological classification [1] of closed
convex sets in Fr´echet spaces) says a bit more:
Theorem 1. For any dense non-compact operator S : X → Y between Banach spaces there is an
operator T : Y → c0 such that the composition T S : X → c0 is non-compact.
By an operator we understand a linear continuous operator. An operator T : X → Y is dense if
T (X) is dense in Y .
The proof of Theorem 1 uses the famous Rosenthal ℓ1 Theorem [8] (see also [5, XI] and [2]) saying
that any bounded sequence in a Banach space X contains a subsequence which is either weakly
Cauchy or ℓ1-basic.
A sequence (xn)n∈ω in a Banach space (X, k · k) is called ℓ1-basic if there are constants 0 < c ≤
C < ∞ such that for each real sequence (αn)n∈ω ∈ ℓ1 we get
c X
n∈ω
αn ≤ (cid:13)(cid:13)(cid:13)
X
n∈ω
αnxn(cid:13)(cid:13)(cid:13)
≤ C X
αn.
n∈ω
Proof of Theorem 1. Assume that S : X → Y is a dense non-compact operator. Let (en)n∈ω be
the standard Schauder basis of the Banach space c0 and (e∗
n)n∈ω is the dual basis in the dual space
c∗
0 = ℓ1. To construct the operator T : Y → c0 with non-compact T S, we shall consider three cases.
1. First we assume that the following condition holds:
(i) there is an ℓ1-basic sequence (y∗
n)n∈ω in Y ∗ such that the sequence (S ∗y∗
n)n∈ω is ℓ1-basic and
weak∗ null in X ∗.
0
1
0
2
n
u
J
5
1
]
.
A
F
h
t
a
m
[
1
v
9
8
0
3
.
6
0
0
1
:
v
i
X
r
a
In this case we define the operator T : Y → c0 by T : y 7→ (y∗
0 → Y ∗ maps the n-th coordinate functional e∗
n(y))n∈ω. Observe that the dual
n. Consequently, the
0 onto y∗
n ∈ c∗
operator T ∗ : c∗
sequence
being ℓ1-basic, is not totally bounded in Y ∗, which implies that the dual operator (T S)∗ : c∗
is not compact. By the Schauder Theorem [4, 7.7.], the operator T S : X → c0 also is not compact.
0 → X ∗
(cid:0)S ∗y∗
n(cid:1)n∈ω = (cid:0)(T S)∗e∗
n(cid:1)n∈ω
,
2. Assume that the condition (i) does not hold but
(ii) there is an ℓ1-basic sequence (y∗
n)n∈ω in Y ∗ whose image (S ∗y∗
n)n∈ω is ℓ1-basic in X ∗.
1991 Mathematics Subject Classification. 47B07; 46B15.
Key words and phrases. Compact operator, Banach space, Josefson-Nissenzweig Theorem.
1
2
IRYNA BANAKH AND TARAS BANAKH
In this case, by [5, Exercise 3(i)] the condition (ii) combined with the negation of (i) imply the
existence of an ℓ1-basic sequence (xn)n∈ω in X whose image (Sxn)n∈ω is an ℓ1-basic sequence in Y .
Arguing as in the proof of Josefson-Nissenzweig Theorem [5, p.223], we can construct a bounded
linear operator T : Y → c0 such that T S(xn) = en ∈ c0 for all n ∈ ω. Since the operator T S is not
compact, we are done.
3. Assume that (ii) does not hold. Since the operator S is not compact, its dual S ∗ : Y ∗ → X ∗ is
not compact too, see [4, 7.7]. This means that the image S ∗(B ∗) of the closed unit ball B ∗ ⊂ Y ∗ is
not totally bounded in X ∗. Consequently, the dual ball B ∗ contains a sequence (y∗
n)n∈ω whose image
(S ∗y∗
n)n∈ω is ε-separated for some ε > 0. The latter means that kS ∗(y∗
m)k ≥ ε for all n 6= m.
n − y∗
By the Rosenthal ℓ1 Theorem, (S ∗y∗
ℓ1-basic. We lose no generality assuming that the entire sequence (S ∗y∗
or ℓ1-basic.
n)n∈ω contains a subsequence which is either weak Cauchy or
n)n∈ω is either weak Cauchy
3a. First we assume that the sequence (S ∗y∗
n)n∈ω is weak Cauchy. Then it is weak∗ Cauchy and
∞ ∈ S ∗(B ∗).
being a subset of the weakly∗ compact set S ∗(B ∗) weakly∗ converges to some point x∗
Fix any point y∗
∞. The density of the operator S : X → Y implies the
injectivity of the dual operator S ∗ : Y ∗ → X ∗. The weak∗ compactness of the closed unit ball
B ∗ ⊂ Y ∗ guarantees that S ∗B ∗ : B ∗ → X ∗ is a homeomorphic embedding for the weak∗ topologies
on B ∗ and X ∗. Now we see that the weak∗ convergence of the sequence (S ∗y∗
∞ implies
the weak∗ convergence of the sequence (y∗
∞ ∈ B ∗ with S ∗(y∗
n)n∈ω to S ∗y∗
∞) = x∗
n − y∗
∞)n∈ω to zero.
n − y∗
∞)}n∈ω is ε-separated, the operator (T S)∗ : c∗
∞)(y)(cid:1)n∈ω, is well-defined. Since the set
0 → X ∗ is not compact
Then the bounded operator T : Y → c0, T : y 7→ (cid:0)(y∗
{(T S)∗(e∗
and hence T S : X → c0 is not compact too.
n)}n∈ω = {S ∗(y∗
n − y∗
3b. Finally, assume that (S ∗y∗
lifting property of ℓ1), the sequence (y∗
the condition (ii) fails.
n)n∈ω is an ℓ1-basic sequence in X ∗. By Proposition 5.10 [4] (the
n)n∈ω is ℓ1-basic in Y ∗, which contradicts our assumption that
(cid:3)
References
[1] T. Banakh, R. Cauty, Topological classification of closed convex sets in Fr´echet spaces, preprint.
[2] E. Behrends, New proofs of Rosenthal's l1-theorem and the Josefson-Nissenzweig theorem, Bull. Polish Acad. Sci.
Math. 43:4 (1995), 283 -- 295 (1996).
[3] P. Hajek, V.S. Montesinos,J. Vanderwerff, V. Zizler, Biorthogonal systems in Banach spaces, Springer, NY, 2008.
[4] M. Fabian, P. Habala, P. Hajek, V.S. Montesinos, J. Pelant, V. Zizler, Functional analysis and infinite-dimensional
geometry, Springer, NY, 2001.
[5] J. Diestel, Sequences and series in Banach spaces, Springer, NY, 1984.
[6] B. Josefson, Weak sequential convergence in the dual of a Banach space does not imply norm convergence, Ark. Mat.
13 (1975), 79 -- 89.
[7] A. Nissenzweig, W ∗ sequential convergence, Israel J. Math. 22:3-4 (1975), 266 -- 272.
[8] H.P. Rosenthal, A characterization of Banach spaces containing ℓ1, Proc. Natl. Acad. Sci. USA 71 (1974), 2411-2413.
Department of Functional Analysis, Ya.Pidstryhach Institute for Applied Problems of Mechanics
and Mathematics, Naukova 3b, Lviv, Ukraine
E-mail address: [email protected]
Instytut Matematyki, Uniwersytet Humanistyczno-Przyrodniczy Jana Kochanowskiego, Kielce, Poland,
and Department of Mathematics, Ivan Franko National University of Lviv, Universytetska 1, 79000,
Lviv, Ukraine
E-mail address: [email protected]
|
1109.2419 | 1 | 1109 | 2011-09-12T09:55:12 | Embedding theorems for harmonic multifunctional spaces on R^n+1 | [
"math.FA"
] | We introduce and study properties of certain new multifunctional harmonic spaces in the upper halfspace.We prove several sharp embedding theorems for such multifunctional spaces,these results are new even in the case of a single function. | math.FA | math |
EMBEDDING THEOREMS FOR HARMONIC MULTIFUNCTION
SPACES ON Rn+1
+
ROMI F. SHAMOYAN AND MILOS ARSENOVI ´C†
Abstract. We introduce and study properties of certain new multi functional
harmonic spaces in the upper half space. We prove several sharp embedding
theorems for such multi functional spaces, these results are new even in the
case of a single function.
1. Introduction and auxiliary results
The theory of harmonic function spaces in the single function case is well devel-
oped by various authors during last decades, see [16], [17], [9]. The main goal of
this paper is to define, for the first time in the literature, multi functional harmonic
spaces and to establish some properties of these spaces. The proofs we found and
provided below are short modifications of proofs in the case of a single function,
but even in this special case our results are new. We believe these new interesting
objects can serve as a base for further generalizations and investigations.
Analytic analogues of theorems on multi functional spaces we obtained below
were proved in recent papers of the first author [11], [12]. We intend to expand
these results to more general harmonic function spaces based on several functions
in higher dimension. Let us note that these topics are new even in the classical case
of harmonic function spaces on the unit disk.
We set H = {(x, t) : x ∈ Rn, t > 0} ⊂ Rn+1. For z = (x, t) ∈ H we set
z = (x, −t). We denote the points in H usually by z = (x, t) or w = (y, s). The
Lebesgue measure is denoted by dm(z) = dz = dxdt or dm(w) = dw = dyds, the
Lebesgue measure of a measurable set E ⊂ H is often denoted by E. We also use
measures dmλ(z) = tλdxdt, λ ∈ R.
The space of all harmonic functions in a domain Ω is denoted by h(Ω). Weighted
harmonic Bergman spaces on H are defined, for 0 < p < ∞ and λ > −1, by
A(p, λ) = A(p, λ)(H) =(f ∈ h(H) : kf kA(p,λ) =(cid:18)ZH
f (z)pdmλ(z)(cid:19)1/p
< ∞) .
These spaces are complete metric spaces, for 1 ≤ p < ∞ they are Banach spaces.
For f ∈ h(H) and t > 0 we set Mp(f, t) = kf (·, t)kLp(Rn), 0 < p < ∞ with the
usual convention in the case p = ∞. We use harmonic mixed norm spaces
(1)
Bp,q
α =(cid:26)f ∈ h(H) : kf kq
Bp,q
α
=Z ∞
0
M q
p (f, t)tαq−1dt < ∞(cid:27) ,
† Supported by Ministry of Science, Serbia, project OI174017.
1Mathematics Subject Classification 2010 Primary 42B15, Secondary 46E35. Key words and
Phrases: multi functional spaces, harmonic functions, embedding theorems, upper half space.
1
2
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
where 0 ≤ p < ∞, 0 < q < ∞ and harmonic Triebel-Lizorkin spaces
(2)
F p,q
α =(f ∈ h(H) : kf kp
F p,q
α
=ZRn(cid:18)Z ∞
0
f (x, t)qtαq−1dt(cid:19)p/q
dx < ∞) ,
where again 0 < p ≤ ∞ and 0 < q ≤ ∞, the case q = ∞ is covered by the usual
convention. These spaces are complete metric spaces and for min(p, q) ≥ 1 they
are Banach spaces. For details on A(p, λ) spaces and more general Bp,q
α spaces see
[7]; Triebel-Lizorkin spaces were studied, in the case of analytic functions, by many
authors, see for example [15].
By X-Lp Carleson measure (or simply X Carleson measure when p is clear from
the context) of a (quasi)-normed subspace X of h(Ω) we understand those positive
Borel measures µ on Ω such that
(3)
(cid:18)ZΩ
f pdµ(cid:19)1/p
≤ Ckf kX,
f ∈ X.
Typical cases are Ω = B = {x ∈ Rn : x < 1} and Ω = H. In the case Ω is the unit
disc and X is the analytic Hardy space H p we have a classical notion of Carleson
measures on the unit disc. The cases of harmonic spaces can be found, for example,
in [4] and much earlier in [14].
One of the goals of this paper is to find complete descriptions of Carleson mea-
sures for certain new harmonic function spaces in higher dimension, and also in
the multi function case, in the latter case some restrictions on the parameters in-
volved usually appear. We note that recently several new papers appeared where
embedding theorems for analytic spaces in the unit disk were obtained and where
the classical expression RB f pdµ was modified, generalized or changed to certain
expressions of G(f, µ, p), see for example [8], [18]; see also an earlier paper [5]. In
these papers descriptions of measures for which G(f, µ, p) ≤ Ckf kX were presented.
Our sharp embedding theorems 2, 3, 4 and 8 we present below are of this character,
and the spaces we deal with are defined using the above mentioned idea. On the
other hand, our theorems 1, 5 and 6 give results modeled after (3), but in multi
functional case.
We use common convention regarding constants:
letter C denotes a constant
which can change its value from one occurrence to the next one. Given two positive
quantities A and B, we write A ≍ B if there are two constants c, C > 0 such that
cA ≤ B ≤ CA.
Many of the results of this paper rely on the following three key auxiliary results.
The first one is a concrete Whitney decomposition into cubes of the upper half space,
the second one is essentially subharmonic behavior of f p for harmonic f and for
0 < p < ∞ and the third one describes geometric properties of Whitney cubes.
Lemma 1 ([16]). There exists a collection {∆k}∞
parallel to coordinate axes such that
k=1 of closed cubes in H with sides
k=1∆k = H and diam∆k ≍ dist(∆k, ∂H).
1o. ∪∞
2o. The interiors of the cubes ∆k are pairwise disjoint.
3o. If ∆∗
the collection {∆∗
k is a cube with the same center as ∆k, but enlarged 5/4 times, then
+ , i.e. there is
k=1 forms a finitely overlapping covering of Rn+1
k}∞
a constant C = Cn such that Pk χ∆∗
k
≤ C.
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
3
Lemma 2 ([6]). Let ∆k and ∆∗
(ξk, ηk) be the center of ∆k. Then, for 0 < p < ∞ and α > 0, we have
k be the cubes from the previous lemma and let
(4)
ηαp−1
k
max
z∈∆k
f (z)p ≤
C
∆∗
kZ∆∗
k
tαp−1f (x, t)pdxdt,
f ∈ h(H),
k ≥ 1.
Lemma 3 ([16]). Let ∆k and ∆∗
be the center of the cube ∆k. Then we have:
k are as in the previous lemma, let ζk = (ξk, ηk)
(5)
(6)
(7)
mλ(∆k) ≍ ηn+1+λ
k
≍ mλ(∆∗
k),
λ ∈ R,
w − z ≍ ζ k − z,
w ∈ ∆∗
k,
z ∈ H,
t ≍ ηk,
(x, t) ∈ ∆∗
k.
Since the cubes from the above Lemmata appear quite often in this paper, we
k are the cubes from Lemma 1, centered at
fix the following notation: ∆k and ∆∗
ζk = (ξk, ηk), k ≥ 1.
We also need the following integral estimate.
Lemma 4 ([10]). If α > −1 and n + α < 2γ − 1, then
(8)
tαdz
z − w2γ ≤ Csα+n+1−2γ,
ZH
w = (y, s) ∈ H.
The following definition introduces certain multi functional spaces that appeared
in the setting of analytic functions in the unit ball in Cn in [13].
Definition 1. Let s > −1 and ~p = (p1, . . . , pm) where 0 < pi < ∞. We denote by
A(~p, s, m) the set of all m-tuples (f1, . . . , fm) of functions harmonic in H such that
(9)
(f1, . . . , fm)A(~p,s,m) =ZH
m
Yi=1
fi(z)pitsdm(z) < ∞.
If pi = p, i = 1, . . . , m, we write simply A(p, s, m).
For the first part of the following proposition see [2], for the second one see [1].
Note that the first part gives a simple estimate of the A(~p, λ, m) norm, while the
second one gives an estimate of trace in A(p, λ) norm.
Proposition 1. 1o. Let 0 < pi < ∞, −1 < si < ∞, fi ∈ A(pi, si)(H) for
i = 1, . . . , m and set λ = (m − 1)(n + 1) +Pm
Yi=1
(f1, . . . , fm)A(~p,λ,m) ≤ C
(10)
m
i=1 si. Then
kfikpi
A(pi,si).
there is a constant C > 0 such that for all f ∈ h(Hm) we have
(11)
2o. Let 0 < p < ∞ and s1, . . . , sm > −1. Set λ = (m − 1)(n + 1) +Pm
ZH
f (z, . . . , z)pdmλ(z) ≤ CZH
· · ·ZH
f (z1, . . . , zm)pdms1 (z1) . . . dmsm (zm).
j=1 sj. Then
4
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
2. Embedding theorems for multi functional spaces of harmonic
functions
In this section we give several sharp embedding theorems for multi functional
spaces of harmonic functions, necessary and sufficient conditions turn out to be
Carleson-type conditions.
The following theorem is an analogue for harmonic functions spaces of Theorem
3 from [11].
Theorem 1. Let µ a positive Boreal measure on H. Assume 0 < pi, qi < ∞,
= 1 and let α > −1. Then the following two conditions
i = 1, . . . m, satisfy Pm
on the measure µ are equivalent.
1
qi
i=1
1o. If fi, i = 1, . . . , m are functions harmonic in H, then we have
(12)
ZH
m
Yi=1
fi(z)pidµ(z) ≤ C
m
Yi=1" ∞
Xk=1 Z∆∗
k
fi(z)pitαdz!qi#1/qi
.
2o. The measure µ satisfies a Carleson type condition:
(13)
µ(∆k) ≤ C∆km(1+ α
n+1 ),
k ≥ 1.
Proof. Let us assume µ satisfies condition (13). Then we have, using Lemma 1
and Lemma 2:
µ(∆k)
∞
Xk=1
m
Yi=1
fipi
sup
∆k
fi(z)pidµ(z) ≤
fi(w)pi sαdw
fi(z)pi dµ(z) =
m
∞
Xk=1Z∆k
Yi=1
Yi=1Z∆∗
m
k
ZH
m
∞
Yi=1
Xk=1
Xk=1
∞
≤ C
µ(∆k)η−m(n+1+α)
k
≤ C
fi(w)pi sαdw.
m
Yi=1Z∆∗
k
Now one arrives at estimate (12) by applying Holder's inequality for sums with
exponents q1, . . . , qm.
Now we assume (12) holds. We fix k ∈ N and choose fi(z) = f (z) = z − ζk1−n.
Clearly
m
fi(z)pi = z − ζk−(n−1) Pm
i=1 pi .
Yi=1
Therefore (12), combined with Lemma 3, gives
µ(∆k)
(n−1) Pm
η
k
i=1 pi
≤ CZ∆k
Yi=1Z∆∗
≤ C
m
k
z − ζ k−(n−1) Pm
i=1 pi dµ(z) ≤ CZH
m
fi(z)pidµ(z)
m
Yi=1
tαdz
z − ζkpi(n−1)
∆kηα−pi(n−1)
k
,
≤ C
Yi=1
and (13) easily follows. (cid:3)
The spaces defined below were considered, in the case of analytic functions on
the unit ball in Cn, in [12].
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
5
Definition 2. Let 0 < p, q < ∞ and let µ be a positive Borel measure on H. The
space A(p, q, m, dµ) is the space of all (f1, . . . , fm) where fi ∈ h(H) for i = 1, . . . , m
such that
(14)
k(f1, . . . , fm)kq
A(p,q,m,dµ) =
∞
Xk=1 Z∆k
m
Yi=1
fi(z)pdµ(z)!q/p
< ∞.
If dµ = dms, then we write A(p, q, m, s) for the corresponding space, if m = 1 we
write A(p, q, dµ) or A(p, q, s) if dµ = dms.
Lemma 5. Let 0 < s < ∞ and β > −1. Then we have
(15)
f (z)stβdxdt ≍
ZH
∞
Xk=1
ηn+1+β
k
f s,
sup
∆k
f ∈ h(H).
We omit an easy proof based on Lemma 1 and Lemma 3.
Theorem 3.1 from [12] served as a model for the following theorem.
Theorem 2. Let 0 < p, q < ∞, 0 < s ≤ q, βi > −1 for i = 1, . . . , m and let µ be a
positive Borel measure on H. Then the following conditions are equivalent.
1o. If fi ∈ A(s, βi), i = 1, . . . , m, then
m
(16)
k(f1, . . . , fm)kA(p,q,m,dµ) ≤ C
kf kA(s,βi).
2o. The measure µ satisfies a Carleson type condition:
Pm
µ(∆k) ≤ Cη
k
i=1
p(n+1+βi)
s
,
k ≥ 1.
Proof. Assume 2o holds and choose fi ∈ A(s, βi), i = 1, . . . , m. Since 0 < s/q ≤ 1
we have, using Lemma 5
Yi=1
k(f1, . . . , fm)ks
s/q
m
Yi=1
fi(z)pdµ(z)!q/p
fiq!s/q
m
Yi=1
µ(∆k)q/p max
∆k
ηn+1+βi
k
fis
max
∆k
∞
A(p,q,m,dµ) =
Xk=1 Z∆k
≤ ∞
Xk=1
Yi=1
Xk=1
Yi=1 ∞
Xki=1
≤ C
≤ C
m
m
∞
ηn+1+βi
k
max
∆k
fis! ≤ C
m
Yi=1
kfiks
A(s,βi),
and we proved implication 2o ⇒ 1o.
Conversely, assume 1o holds and choose l ∈ N0 such that s(n + l − 1) > n + β + 1.
We use, as test functions, functions fζk,l, k ∈ N where
(17)
fw,l(z) =
∂l
∂tl
1
z − wn−1 ,
z ∈ H,
see [1] for norm and pointwise estimates related to these functions. In particular
we have
(18)
−(n−1+l)+ n+βi+1
kfζk,lkA(s,βi) ≤ Cη
k
s
,
k ≥ 1,
1 ≤ i ≤ m
6
and
(19)
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
k(fζk,l, . . . , fζk,l)kA(p,q,m,dµ) ≥(cid:18)Z∆k
fζk,l(z)mpdµ(z)(cid:19)1/p
≥ Cµ(∆k)1/pη−m(n+l−1)
.
k
The last two estimates combined with (21) give 2o. (cid:3)
The corresponding result for mixed norm spaces is the following theorem.
Theorem 3. Let 0 < p, q < ∞, 0 < s ≤ q, βi > −1 for i = 1, . . . , m, 0 < ti ≤ s
for i = 1, . . . , m and let µ be a positive Borel measure on H. Then the following
conditions are equivalent:
1o. If fi ∈ Bs,ti
(βi+1)/s, i = 1, . . . , m, then
(20)
k(f1, . . . , fm)kA(p,q,m,dµ) ≤ C
2o. If fi ∈ F s,ti
(βi+1)/s, i = 1, . . . , m, then
(21)
k(f1, . . . , fm)kA(p,q,m,dµ) ≤ C
m
Yi=1
m
Yi=1
kf kB
.
s,ti
βi+1
s
kf kF s,ti
βi+1
s
.
3o. The measure µ satisfies a Carleson type condition:
Pm
µ(∆k) ≤ Cη
k
i=1
p(n+1+βi)
s
,
k ≥ 1.
Proof. Since A(s, β) = Bs,s
dings Bs,t
implications 3o ⇒ 1o and 3o ⇒ 2o.
(β+1)/s ֒→ Bs,s
(β+1)/s, the previous theorem, combined with embed-
(β+1)/s, valid for 0 < t ≤ s, gives
(β+1)/s ֒→ Bs,s
(β+1)/s and F s,t
Now we assume 2o holds. Let us choose l ∈ N0 such that s(n + l − 1) > n + β + 1.
Set, for w = (y, σ) ∈ H, fi(z) = f (z) = fw,l(z), i = 1, . . . , m, see (17). We have,
see [1]:
(22)
kfw,lkF s,t
β+1
s
= Cσ
n
s
−(n−1+l−
β+1
s ).
Let us fix a cube ∆k. Using pointwise estimates from below for functions fw,l, see
[1], we obtain:
µ(∆k)1/pη−m(n−1+l)
k
(23)
≤ C(cid:18)Z∆k
fw,l(z)mpdµ(z)(cid:19)1/p
≤ Ck(f1, . . . , fm)kA(p,q,m,dµ),
k ≥ 1.
Pm
Now combining (23), (22) with w = ζk and (21) gives µ(∆k) ≤ Cη
k
and 3o follows.
i=1
p(n+1+βi)
s
The implication 1o ⇒ 3o can be proved using the same test functions fw,l as
α norm of these functions can be found in [1] and
above, relevant estimates of Bp,q
we leave details to the reader. (cid:3)
For w = (y, s) ∈ H we set Qw to be the cube, with sides parallel to the coordinate
axis, centered at w with side length equal to s.
Our next multi functional embedding theorem is motivated by Theorem 3.6 from
[12], which deals with the case of a single analytic function on the unit ball in Cn.
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
7
Theorem 4. Let 0 < p, q < ∞ and 0 < σi ≤ q, −1 < αi < ∞ for i = 1, . . . , m.
Let µ be a positive Borel measure on H. Then the following two conditions are
equivalent:
1o. For any m-tuple (f1, . . . , fm) of harmonic functions on H we have
(24)
k(f1, . . . , fm)kq
A(p,q,m,dµ) ≤ C
m
Yi=1ZH(cid:18)ZQw
fi(z)σi tαidz(cid:19)q/σi
dw.
2o. The measure µ satisfies the following Carleson-type condition:
(25)
m(n+1) p
µ(∆k) ≤ Cη
k
q +Pm
i=1
p(n+1+αi)
σi
,
k ≥ 1.
Proof. Let us prove 2o ⇒ 1o assuming m = 1. Let f ∈ h(H), then we have,
using Lemma 1, Lemma 2 and Lemma 3:
kf kq
A(p,q,dµ) =
=
f (z)pdµ(z)(cid:19)q/p
≤
∞
Xk=1(cid:18)max
∆k
f pµ(∆k)(cid:19)q/p
q( n+1+α
f qη
k
σ + n+1
q )
∞
∞
max
∆k
Xk=1(cid:18)Z∆k
Xk=1
Xk=1 Z∆∗
Xk=1 Z∆∗
∞
∞
k
k
≤ C
≤ C
σ + n+1
q )
q( n+1+α
η
k
f (z)σt−n−1dz!q/σ
f σdz!q/σ
ηn+1+qα/σ
k
.
We continue this estimation, using Lemma 2 and Lemma 3, and obtain
(26)
kf kq
A(p,q,dµ) ≤ C
ηn+1+qα/σ
k
∞
∞
∞
≤ C
f (w)σsαdw(cid:19) t−n−1−αdz!q/σ
f (w)σsαdw(cid:19) t−n−1dz!q/σ
f (w)σsαdw(cid:19)q/σ
f (w)σsαdw(cid:19)q/σ
Xk=1 Z∆∗
k(cid:18)ZQz
Xk=1 Z∆∗
k(cid:18)ZQz
k(cid:18)ZQz
Xk=1Z∆∗
k(cid:18)ZQz
Xk=1Z∆∗
f (w)σsαdw(cid:19)q/σ
≤ CZH(cid:18)ZQz
t−n−1dz ηn+1
≤ C
≤ C
dz,
dz
∞
k
ηn+1
k
arriving at (24) for m = 1. Note that at (26) we used Holder's inequality and at
the last step we used finite overlapping property of the cubes ∆∗
k.
Now we assume 1o, and again we restrict ourselves to the case m = 1. Let us
choose l ∈ N0 such that q(n − 1 + l) > q(n + 1 + α)/σ + n + 1. Let us fix a cube ∆k
centered at ζk = (ξk, ηk). We use a construction from [1], where interested reader
can find more details. Namely, there are constants c > 0 and δ > 0 such that for
8
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
all w ∈ H we have
(27)
Tw = cQw,
Tw = {z ∈ Qw : fw,l(z) > cz − w−(n−1+l)}.
Next, using a compactness argument, one shows that for w = (y, s) ∈ H there are
points wj = (yj, sj) ∈ H, 1 ≤ j ≤ N , such that sj ≍ s and Qw = ∪N
j=1Twj ∩ Qw.
Here N depends only on n and l. We apply this argument to w = ζk, it easily
follows that there is θk = (uk, vk) ∈ H such that µ(Tθk ∩ ∆k) ≥ N −1µ(∆k) and
vk ≍ ηk. Now we choose f = fθk,l as a test function. We have
kf kq
A(p,q,dµ) ≥(cid:18)Z∆k
f (z)pdµ(z)(cid:19)q/p
≥ Z∆k∩Tθk
f (z)pdµ(z)!q/p
(28)
≥ Cµ(∆k ∩ Tθk)q/pv−q(n−1+l)
k
≥ Cµ(∆k)η−q(n−1+l)
.
k
On the other hand, using Lemma 3 and Lemma 4, we have
ZH(cid:18)ZQw
f (z)σdmα(z)(cid:19)q/σ
dw ≤ CZH(cid:18)
sn+1+α
w − θkσ(n−1+l)(cid:19)q/σ
dw
q
σ (n+1+α)−q(n−1+l)+n+1
≤ Cv
k
q
σ (n+1+α)−q(n−1+l)+n+1
≤ Cη
k
.
This estimate, combined with (28) and (24) (with m = 1), gives desired estimate
(25) for m = 1.
The general case m > 1 can be proved along the same lines, multiple sums as
in the proof of Theorem 2 appear. Since no new ideas are involved we can leave
details to the interested reader. (cid:3)
As a preparation for our next result we state and prove the following lemma.
Lemma 6. Let α > −1, 0 < τ < ∞. Then there is a constant C such that for any
measurable function u(z) ≥ 0 on H we have
(29)
ηn+1
k Z∆∗
k
u(z)dmα(z)!τ
≤ CZ∆∗
k(cid:18)ZQw
u(z)dmα(z)(cid:19)τ
dw,
k ≥ 1.
Proof. Let us denote the left hand side by Lα and the right hand side, without
constant C, by Dα. Since Lα ≍ ηα
k D0 it suffices to consider the
special case α = 0. Let qw denote the cube centered at w = (y, s) ∈ H, with side
length equal 4s/5. Then we have ∆∗
k where
N depends only on n. Therefore (29) reduces to the following estimate:
i=1qwi for suitable w1, . . . , wN ∈ ∆∗
k L0 and Dα ≍ ηα
k = ∪N
(30)
ηn+1
k (cid:18)Zqw
d w,
w ∈ ∆∗
k.
u(z)dm(z)(cid:19)τ
k. Clearly, Rqw
k(cid:18)ZQ w
≤ CZ∆∗
u(z)dm(z) ≤ RQ w
u(z)dm(z)(cid:19)τ
Now we fix w ∈ ∆∗
u(z)dm(z) whenever qw ⊂ Q w,
and the estimate (30) follows from the simple observation, based on the sizes of qw
and Qw, that
Ek ≥ cηn+1
k
,
Ek = { w ∈ ∆∗
k : qw ⊂ Q w}. (cid:3)
In the following theorem we allow for different exponents pi.
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
9
Theorem 5. Let 0 < pi, σi < ∞, −1 < αi < ∞ for i = 1, . . . , m. Let µ be a
positive Borel measure on H. Then the following two conditions are equivalent:
1o. For any m-tuple (f1, . . . , fm) of harmonic functions on H we have
(31)
ZH
m
Yi=1
fi(z)pi dµ(z) ≤ C
m
Yi=1ZH(cid:18)ZQw
fi(z)σidmαi (z)(cid:19)pi/σi
dw.
2o. The measure µ satisfies the following Carleson-type condition:
(32)
m(n+1)+Pm
µ(∆k) ≤ Cη
k
i=1
pi(n+1+αi)
σi
,
k ≥ 1.
Proof. Let us assume (31) holds. We choose l ∈ N such that
pi(n − 1 + l) > n + 1 +
pi
σi
(n + 1 + αi)
i = 1, . . . , m.
We fix a cube ∆k and take as test functions fi = f = fθk,l, i = 1, . . . , m, where
θk = (uk, vk) is chosen as in Theorem 4. Then we have, using Lemma 4:
η−(n−1+l) Pm
i=1 pi
k
µ(∆k) ≤ CZTθk
m
fi(z)pidµ(z)
m
m
∩∆k
Yi=1
Yi=1ZH(cid:18)ZQw
Yi=1ZH(cid:18)
fi(z)pi dµ(z) ≤ CZH
Yi=1
fi(z)σidmαi (z)(cid:19)pi/σi
θk − wσi (n−1+l)(cid:19)pi/σi
sn+1+αi
dw
m
dw
≤ C
≤ C
i=1
m(n+1)+Pm
≤ Cv
k
m(n+1)+Pm
≤ Cη
k
i=1
pi(n+1+αi)
σi
−(n−1+l) Pm
i=1 pi
pi(n+1+αi)
σi
−(n−1+l) Pm
i=1 pi
which gives (32). Conversely, we assume now (32). Using finite overlapping prop-
erty of ∆∗
k we obtain
m
Yi=1ZH(cid:18)ZQw
fi(z)σidmαi (z)(cid:19)
pi
σi
dw ≥ C
∞
Xk=1
and we have also an obvious estimate
m
Yi=1Z∆∗
k(cid:18)ZQw
fi(z)σi dmαi(z)(cid:19)
pi
σi
dw
Z∆k
m
Yi=1
fi(z)pidµ(z) ≤ µ(∆k)
fi(z)pi
max
z∈∆k
m
Yi=1
m(n+1)+Pm
≤ Cη
k
i=1
pi(n+1+αi)
σi
m
Yi=1
fi(z)pi .
max
z∈∆k
Therefore it suffices to prove, for each k ≥ 1 and 1 ≤ i ≤ m:
n+1+ pi(n+1+αi)
η
k
σi
max
z∈∆k
fi(z)pi ≤ CZ∆∗
k(cid:18)ZQw
fi(z)σi dmαi (z)(cid:19)pi/σi
dw.
10
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
However, for k ≥ 1 and 1 ≤ i ≤ m, using Lemma 2 and Lemma 6 we obtain:
n+1+ pi(n+1+αi)
η
k
σi
max
z∈∆k
fi(z)pi ≤ Cηn+1
k Z∆∗
k(cid:18)ZQw
k
fi(z)σidmαi (z)!
f (z)σidmαi (z)(cid:19)
pi
σi
≤ CZ∆∗
pi
σi
dw
and the proof is completed. (cid:3)
The following theorem is motivated by Theorem A from [13].
It gives a very
general and flexible method of obtaining multi functional results starting from a
uniform estimate for a single function.
Theorem 6. Let µ be a positive Borel measure on an open proper subset G of
Rn and let (Xi, k · kXi) be a (quasi)-normed space of functions harmonic in G,
1 ≤ i ≤ k. Let d(x) = dist(x, ∂G), x ∈ G and 0 < qi < ∞, −1 < βi < ∞ for
i = 1, . . . , k. Assume
(33)
and
(34)
Then
(35)
fi(x)qi d(x)βi ≤ Ckfikqi
Xi
,
sup
x∈G
fi ∈ Xi,
i = 2, . . . , k
f1(x)q1 dµ(x) ≤ Ckf1kq1
X1
,
f1 ∈ X1.
ZG
ZG
k
Yi=1
fi(x)qi d(x)β2+···+βk dµ(x) ≤ C
kfikqi
Xi
,
fi ∈ Xi.
k
Yi=1
We refer the reader for the proof by induction to [13], where arguments readily
extend to the present situation. We also note that the above theorem can be
extended to weights more general than d(x), however it is precisely the case of d(x)
where uniform estimates of type (33) are available. For example, if G = H, then
d(z) = t and we have an estimate
f (z)pd(z)n+1+α ≤ Ckf kp
A(p,α),
f ∈ A(p, α),
z ∈ H.
Therefore, taking dµ(z) = ts1 dxdt we obtain the following corollary.
Corollary 1. If 0 < pi < ∞, −1 < si < ∞, fi ∈ A(pi, si)(H) for i = 1, . . . , m and
α > −1, then
(36)
ZH
m
Yi=1
fi(z)pit(n+1)(m−1)+Pm
i=1 si dz ≤ C
kfikpi
A(pi,si)
m
Yi=1
Using embedding Bp,q
α
֒→ Bp,p
α , 0 < q ≤ p < ∞, see [7], we deduce another
corollary.
Corollary 2. If 0 < qi ≤ pi < ∞, −1 < si < ∞ and fi ∈ Bpi,qi
then
si
for 1 ≤ i ≤ m,
(37)
ZH
m
Yi=1
fi(z)pi tn(m−1)+Pm
i=1 sipi−1dz ≤ C
m
Yi=1
kfikpi
Bpi,qi
si
.
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
11
Let Dβf denote the Riesz potential of f ∈ h(H) with respect to t variable. Using
estimate
(38) ZH
f (z)ptαp−1dz ≤ CZH
Dβ f (z)ptαp−1+βpdz = Nα,β,p(f ),
f ∈ h(H)
which is valid for 0 < p < ∞, α, β > 0, see Theorem 7 from [3], we obtain the
following corollary.
Corollary 3. Let 0 < si, pi < ∞ and fi ∈ A(pi, si) for i = 1, . . . , m. Then we have
(39)
ZH
m
Yi=1
fi(z)pit(n+1)(m−1)+Pm
i=1 sidz ≤ C
Nsi,β,pi(fi).
m
Yi=1
Similar assertions can be obtained easily also in the unit ball B.
Let H p(B), 1 < p < ∞, stand for the harmonic Hardy space on the open unit
ball.
Proposition 2. Let fi ∈ H pi(B) where 1 < pi < ∞ for i = 1, . . . , m. Then we
have
(40)
ZS
Yj=1
m
m
fj(rξ)pj dσ(ξ) (1 − r)(m−1)n ≤ C
kfjkpj
Hpj .
Yj=1
Similarly, if fi ∈ H pi(H) where 1 < pi < ∞ for i = 1, . . . , m then we have
(41)
ZRn
m
Yj=1
fj(x, t)pj dx t(m−1)n ≤ C
kfjkpj
Hpj .
m
Yj=1
The proof is by induction on m, using estimate
(42)
f (x)p(1 − x)n ≤ Ckf kp
Hp ,
f ∈ H p(B),
and an analogous one for the case of the upper half space.
Lemma 7. Let β > 0 and 1 < p < ∞. Then the operator S defined by
Sg(z) =ZH(cid:18)
tn
z − w2n(cid:19)1+β
is bounded on Lp(H, dmβn−1).
g(w)dmβn−1(w),
z ∈ H,
For a similar result for analytic functions in the unit ball see [19].
Proof. We use Schur's test, see [19] for statement and applications to related
problems. Let 1/p+1/q = 1 and set h(z) = tλ, where λ will be subject to additional
conditions. Using Lemma 4 we obtain
(43)
(44)
(45)
provided −βn < λq < βn + n. Similarly we have
ZH(cid:18)
ZH(cid:18)
tn
z − w2n(cid:19)1+β
z − w2n(cid:19)1+β
tn
sλqdmβn−1(w) ≤ Ctλq,
tλpdmβn−1(w) ≤ Ctλp,
provided −n − 2βn < λp < 0. Clearly, a parameter λ satisfying all the inequalities
exists, it suffices to choose max(−βn/q, (−n−2βn)/p) < λ < 0. Schur's test implies
that S is bounded on Lp(H, dmβn−1) for all 1 < p < ∞. (cid:3)
The following result was proved for analytic functions in [12] (Theorem 3.3).
12
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
Theorem 7. Let 0 < p < q < ∞, α > 0 and let µ be a positive Borel measure on
H. Then the following two conditions are equivalent:
1o.
(46)
2o.
(47)
ZH ZH(cid:18)
tn
z − w2n(cid:19)1+αq
dµ(z)!
q
q−p
dmαqn−1(w) < ∞.
−(αqn+n) p
q−p
η
k
µ(∆k)
q
q−p < ∞.
∞
Xk=1
Proof. Let us prove implication 1o ⇒ 2o. We have
q
q−p
dµ(z)!
dmαqn−1(w)
dµ(z)
q−p
q
tn
z − w2n(cid:19)1+αq
z − w2n(cid:19)1+αq
tn
dµ(z)!
q
q−p
dmαqn−1(w)
dmαqn−1(w)
∞
∞
tn
≥ C
∞ >ZH ZH(cid:18)
z − w2n(cid:19)1+αq
Xk=1Z∆k
Xj=1Z∆j(cid:18)
Xk=1Z∆k Z∆k(cid:18)
Xk=1
−(αqn+n) p
q−p
η
k
≥ C
≥ C
µ(∆k)
∞
∞
q
q−p ,
where in the first estimate we used Lemma 1 and in last one we used Lemma 3.
Next we prove 2o ⇒ 1o. Assume f : H → R is subharmonic. Then f q/p is also
subharmonic and we have, using Lemma 2,
f (z)dµ(z) =
ZH
≤
∞
Xk=1Z∆k
Xk=1 Z∆∗
∞
k
∞
Xk=1
f (z)dµ(z) ≤
f (z)µ(∆k)
sup
z∈∆k
f (z)q/pdmαqn−1(z)!p/q
−p(n+αqn)
q
η
k
µ(∆k).
Note that q/p and q/(q − p) are a pair of conjugate exponents, hence an application
of Holder's inequality gives
(48)
ZH
f (z)dµ(z) ≤ Ckf kLq/p(H,dmαqn−1)" ∞
Xk=1
−(n+qnα)p
q−p
η
k
µ(∆k)
q−p
q
.
q
q−p#
Note that here we used finite overlapping property of the family ∆∗
g ∈ Lq/p(H, dmαqn−1), g ≥ 0, and set
k. Let us choose
f (z) = (Sg)(z) =ZH(cid:18)
tn
z − w2n(cid:19)1+αq
g(w)sαqn−1dw.
Since, for any fixed w ∈ H, the function tn/z −w2n is subharmonic it easily follows
that f is subharmonic. Moreover, the operator S is bounded on Lq/p(H, dmαqn−1)
by Lemma 7, i.e. kSgkLq/p(H,dmαqn−1) ≤ CkgkLq/p(H,dmαqn−1).
Now we apply (48) to f = S(g) and obtain
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
13
tn
z − w2n(cid:19)1+αq
ZHZH(cid:18)
≤CkgkLq/p(H,dmαqn−1)" ∞
Xk=1
g(w)sαqn−1dwdµ(z)
−(n+qnα)p
q−p
η
k
µ(∆k)
q−p
q
.
q
q−p#
Since the last estimate is true for any non negative g ∈ Lq/p(H, dmαqn−1), 1o follows
by duality. (cid:3)
An application of the above theorem to dµ(z) = f (z)q−pdmα(z) gives a char-
acterization of A(p, q, α) spaces for 0 < p < q < ∞, α > −1, see [12].
Set rk = 2−k, Ik = [rk+1, rk) and Hk = Rn × Ik for k ∈ Z.
Definition 3. Let µ be a positive Borel measure on H and let 0 < p, q < ∞. We
define the space K p
q (µ) as the space of f ∈ h(H) such that
(49)
kf kq
q (µ) =
K p
∞
Xk=−∞(cid:18)ZHk
f (z)pdµ(z)(cid:19)q/p
< ∞.
The spaces K p
q (µ) are harmonic Herz type spaces, related classes of spaces ap-
peared in literature, for example in [10]. Clearly these spaces include weighted
Bergman spaces: K p
Note that the K p
p (ms) = A(p, s), 0 < p < ∞, s > −1.
q (µ) (quasi) norm of f is the lq norm of the sequence kf kLp(Hk,dµ),
−∞ < k < ∞. Therefore we have:
(50)
kf kK p
q2 (dµ) ≤ kf kK p
q1 (dµ),
0 < q1 ≤ q2 < ∞.
Theorem 8. Let α > −1, 0 < p ≤ q < ∞ and let µ be a positive Borel measure on
H. Then the following two conditions are equivalent.
1o. A(p, α) ֒→ K p
2o. The measure µ satisfies the following Carleson type condition:
q (µ).
(51)
µ(∆j)
∆j1+ α
n+1
≤ C,
j ≥ 1.
Proof. Sufficiency, for p = q, is proved in [1], and the general case follows from
the embedding (50).
Now we prove necessity, i.e. we assume 1o holds. We choose l ∈ N0 such that
p(n − 1 + l) > α + n + 1. Let us fix a cube ∆j and choose k ∈ Z such that
m(Hk ∩ ∆j) ≥ ∆j/2. We use a test function f = fζj ,l:
kfζj,lkq
K p
q (µ) ≥ ZHk ∩∆j
fζj,l(z)pdµ(z)!q/p
(52)
= Cµ(∆j)q/pη−q(n−1+l)
j
.
≥ Chµ(∆j)η−p(n−1+l)
j
iq/p
Next we estimate A(p, α)-norm of fζ,l using Lemma 4:
(53) kf kp
A(p,α) =ZH
fw,l(z)ptαdz ≤ CZH
tαdz
z − ζ jp(n−1+l)
≤ Cηα+n+1−p(n−1+l)
j
.
Combining our assumption kf kK p
q (µ) ≤ Ckf kA(p,α) with (52) and (53) gives 2o. (cid:3)
14
MILOS ARSENOVI ´C AND ROMI F. SHAMOYAN
Remark 1. Sharp embedding theorems of the spaces K p
q (µ)
into F p,q
α with some restriction on the indexes are also true and they follow from
Theorem 8 and here our arguments repeat arguments of proof of Theorem 3 which
was obtained from linear case (Theorem 2) and appropriate embedding results for
Bp,q
q (µ) into Bp,q
α and K p
α and F p,q
α
spaces.
References
[1] M. Arsenovi´c, R. F. Shamoyan, On embeddings, traces and multipliers in harmonic function
spaces, arXiv:1108.5343v1.
[2] M. Arsenovi´c, R. F.Shamoyan, Trace theorems in harmonic function spaces on (Rn+1
+ )m and
multipliers theorems, preprint, 2011, 10 pages.
[3] K. L. Avetisyan, Fractional integro-differentiation in harmonic mixed norm spaces on a half-
space, Comment. Math. Univ. Carolinae, Vol. 42 (2001), No. 4, 691-709.
[4] B. R. Choe, H. Koo and H. Yi, Carleson type conditions and weighted inequalities for har-
monic functions, preprint, 2010.
[5] W. S. Cohn, Generalized area operators on Hardy spaces, J. Math. Anal. Appl., 1997, 216:
112-121.
[6] A. E. Djrbashian, The classes Ap
α of harmonic functions in half-spaces and an analogue of
M. Riesz theorem, Izv. Akad. Nauk Arm. SSR, Matematika 22 (1987), no. 4, 386398 (in
Russian); English transl.: Soviet J. Contemp. Math. Anal. (Armenian Academy of Sciences)
22 (1987), no. 4, 7485.
[7] M. Djrbashian and F. Shamoian, Topics in the theory of Ap
α classes, Teubner Texte zur
Mathematik, 1988, v 105.
[8] M. Q. Gong, Z. J. Lou, Z. J. Wu, Area operators from Hp spaces to Lq spaces, Sci. China
Math., 2010, 53(2): 357-366, doi: 10.1007/s11425-010-0031-9.
[9] L. Grafakos, Classical Fourier Analysis, Graduate Texts in Mathematics 249, Springer 2008.
[10] H. Koo., K. Nam, H. Yi, Weighted harmonic Bergman kernel on half-spaces, J. Math. Soc.
Japan Vol. 58, No. 2, 2006.
[11] S. Li, R. Shamoyan, On some estimates and Carleson type measure for multifunctional
holomorphic spaces in the unit ball, Bull. Sci. math. 134 (2010) 144-154.
[12] S. Li, R. Shamoyan, On some properties of analytic spaces connected with Bergman metric
ball, Bull. Iran. Math. Soc. Vol. 34 No. 2 (2008), 121-139.
[13] S. Li, R. Shamoyan, On some extensions of theorems on atomic decompositions of Bergman
and Bloch spaces in the unit ball and related problems, Complex Variables and Elliptic Equa-
tions Vol. 54, No. 12, December 2009, 1151-1162.
[14] N. A. Shirokov, Some generalizations of the Littlewood-Paley theorem, Zap. Nauch. Sem.
LOMI 39 (1974), 162-175; J. Soviet Mat. 8 (1977), 119-129.
[15] J. M. Ortega, J. Fabrega, Holomorphic Triebel-Lizorkin Spaces, Journal of Functional Anal-
ysis 151, (1997) 177-212.
[16] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Univ.
Press, Princeton, New Jersey, 1970.
[17] E. M. Stein, G. Weiss, Introduction to Fourier analysis on Euclidean spaces, Princeton Uni-
versity Press, 1971.
[18] Z. Wu, Area operator on Bergman spaces, Science in China A (2006), 36(5), 481-507
[19] K. Zhu, Spaces of Holomorphic Functions in the Unit Ball, Springer-Verlag, New York, 2005.
Department of mathematics, University of Belgrade, Studentski Trg 16, 11000 Bel-
grade, Serbia
E-mail address: [email protected]
Bryansk University, Bryansk Russia
E-mail address: [email protected]
|
1608.06020 | 2 | 1608 | 2017-03-27T09:58:32 | Multipliers for von Neumann-Schatten Bessel sequences in separable Banach spaces | [
"math.FA"
] | In this paper we introduce the concept of von Neumann-Schatten Bessel multipliers and obtain some of their characterizations. Finally, special attention is devoted to the study of invertible Hilbert--Schmidt frame multipliers. These results are not only of interest in their own right, but also they pave the way for obtaining some new results for diagonalization of matrices in finite dimensional setting as well as for dual $g$-frames. In particular, we show that a $g$-frame is uniquely determined by the set of its $g$-frames. | math.FA | math |
MULTIPLIERS FOR VON NEUMANN-SCHATTEN BESSEL
SEQUENCES IN SEPARABLE BANACH SPACES
HOSSEIN JAVANSHIRI AND MEHDI CHOUBIN
Abstract. In this paper we introduce the concept of von Neumann-Schatten
Bessel multipliers in separable Banach spaces and obtain some of their prop-
erties. Finally, special attention is devoted to the study of invertible Hilbert-
Schmidt frame multipliers. These results are not only of interest in their own
right, but also they pave the way for obtaining some new results for diagonal-
ization of matrices in finite dimensional setting as well as for dual HS-frames.
In particular, we show that a HS-frame is uniquely determined by the set of
its dual HS-frames.
1. Introduction
Due to the fundamental works done by Feichtinger and his coauthors [14, 15],
Fourier and Gabor multipliers were formally introduced and popularized from
then on. Now the theory of Fourier and Gabor multipliers plays an important
role in theoretics and applications; For more information about the history of
this class of operators, some of their properties and their applications in scientific
disciplines and in modern life the reader can consult Section 1 of the papers [5, 26]
and the references (for examples) [6, 10, 11, 12, 19]. Balazs [3] extended the
notion of Gabor multipliers to arbitrary Hilbert spaces. In details, he considered
the operators of the form
Mm,Φ,Ψ(f ) =
(f ∈ H),
∞Xi=1
mi(cid:10)f, ψi(cid:11)φi
where Φ = {φi}∞
i=1 and Ψ = {ψi}∞
i=1 are ordinary Bessel sequences in Hilbert space
H, and m = {mi}∞
i=1 is a bounded complex scalar sequence in C. It is worthwhile
to mention that this class of operators is not only of interest for applications in
modern life, but also it is of utmost importance in different branches of linear
algebra, matrix analysis and functional analysis. For example, they are used
for the diagonalization of matrices [17, Definition 3.1], the diagonalization of
operators [4, 11, 25] and for solving approximation problems [12, 13, 19]. We
also recall that by the spectral theorem, every self-adjoint compact operator on a
Hilbert space can be represented as a multiplier using an orthonormal system. In
2010 Mathematics Subject Classification. Primary 46C50, 65F15, 42C15; Secondary 41A58,
47A58.
Key words and phrases. von Neumann-Schatten operator, Bessel sequence, Bessel multiplier,
invertible frame multiplier, g-frame, dual frame.
1
2
H. JAVANSHIRI AND M. CHOUBIN
addition to all these, multipliers generalize the frame operators, approximately
dual frames [8, 20], generalized dual frames [20, Remark 2.8(ii)], atomic systems
for subspaces [16, 21] and frames for operators [18]. Therefore, the study of Bessel
multipliers also leads us to new results concerning dual frames and local atoms,
two concepts at the core of frame theory.
Various generalization of Bessel multipliers have been introduced and stud-
ied in a series of papers recently. This paper continues these investigations. In
details, we investigate Bessel multipliers for von Neumann-Schatten Bessel se-
quences in separable Banach spaces. Here it should be noted that, just because
the study of this class of operators paves the way for obtaining some new results
for von Neumann-Schatten frame [1, 24], this inspires us to investigate this class
of operators in separable Banach spaces. Let us recall that the von Neumann-
Schatten frames in a separable Banach space was first proposed by Arefijamaal
and Sadeghi [24] to deal with all the existing frames as a united object. In fact,
the von Neumann-Schatten frame is an extension of g-frames for Hilbert spaces
[27] and p-frames for Banach spaces [9], two important generalization of ordinary
frames.
2. von Neumann-Schatten p-Bessel sequences: an overview
In this section, we give a brief overview of von Neumann-Schatten p-Bessel
sequences from [24]. Nevertheless, we shall require some facts about the theory
of von Neumann-Schatten p-class of operators. For background on this theory, we
use [23] as reference and adopt that book's notation. Moreover, our notation and
terminology are standard and, concerning frames in Hilbert respectively Banach
spaces, they are in general those of the book [7] respectively the paper [9].
1
2.1. von Neumann-Schatten p-class of operators. Let H be a separable
Hilbert space and let (B(H),k · kop) denotes the C ∗-algebra of all bounded linear
operators on H. For a compact operator A ∈ B(H),
let s1(A) ≥ s2(A) ≥
· · · ≥ 0 denote the singular values of A, that is, the eigenvalues of the positive
operator A = (A∗A)
2 , arranged in a decreasing order and repeated according
to multiplicity. For 1 ≤ p < ∞, the von Neumann-Schatten p-class Cp(H) is
i=1 sp
i (A) < ∞. For
A ∈ Cp(H), the von Neumann-Schatten p-norm of A is defined by
defined to be the set of all compact operators A for whichP∞
p =(cid:16)trAp(cid:17) 1
kAkCp(H) =(cid:16) ∞Xi=1
where tr is the trace functional which defines as tr(A) = Pe∈EhA(e), ei and E
is any orthonormal basis of H. The special case C1(H) is called the trace class
and C2(H) is called the Hilbert-Schmidt class. Recall that an operator A is in
Cp(H) if and only if Ap ∈ C1(H). In particular, kAkp
Cp(H) = kApkC1(H). It is proved
that Cp(H) is a two sided ∗-ideal of B(H) and the finite rank operators are dense
sp
i (A)(cid:17) 1
p ,
(2.1)
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
3
in (Cp(H),k · kCp(H)). Moreover, for A ∈ Cp(H), one has kAkCp(H) = kA∗kCp(H),
kAkop ≤ kAkCp(H) and if B ∈ B(H), then
and
kABkCp(H) ≤ kBkopkAkCp(H).
kBAkCp(H) ≤ kBkopkAkCp(H)
In particular, Cp(H) ⊆ Cq(H) if 1 ≤ p ≤ q < ∞. We also recall that the space
C2(H) with the inner product [T, S]tr := tr(S∗T ) is a Hilbert space.
Now, for a fixed 1 ≤ p < ∞, we define the Banach space
⊕Cp(H) =nA = {Ai}∞
i=1 : Ai ∈ Cp(H) ∀i ∈ N, and
kAkp :=(cid:16) ∞Xi=1
kAikp
Cp(H)(cid:17) 1
p < ∞o.
In particular, ⊕C2(H) is a Hilbert space with the inner product
hA,A′i :=
∞Xi=1
[Ai,A′
i]tr,
2 = hA,Ai.
and so kAk2
We conclude this subsection by recalling the notion of the tensor product of
two arbitrary elements of H which will be useful in our subsequent analysis. To
this end, suppose that x, y ∈ H and define the operator x ⊗ y on H by
(x ⊗ y) (z) = hz, yi x
(z ∈ H).
It is obvious that kx ⊗ yk = kxkkyk and the rank of x ⊗ y is one if x and y are
non-zero. Moreover,
kx ⊗ ykCp(H) = kxkkyk
and
tr(x ⊗ y) = hx, yi .
Thus x ⊗ y is in Cp(H) for all p ≥ 1. Furthermore, the following equalities are
easily verified:
(x ⊗ x′) (y ⊗ y′) = hy, x′i (x ⊗ y′) ;
(x ⊗ y)∗ = y ⊗ x;
T (x ⊗ y) = T (x) ⊗ y;
(x ⊗ y) T = x ⊗ T ∗(y),
where x′, y′ ∈ H and T ∈ B(H). Clearly, the operator x ⊗ x is a rank-one
projection if and only if hx, xi = 1, that is, x is a unit vector. Conversely, every
rank-one projection is of the form x ⊗ x for some unit vector x.
Having reached this state it remains to recall the definition and some properties
of von Neumann-Schatten p-frames for separable Banach spaces. This is the
subject matter of the next subsection.
4
H. JAVANSHIRI AND M. CHOUBIN
2.2. von Neumann-Schatten p-frames. To simplify the later discussion, we
make the following blanket assumption.
Convention. For the rest of this paper we assume that H is a Hilbert space
with orthonormal basis E = {ei}i∈I , 1 < p < ∞ and q is the conjugate exponent
to p, that is, 1/p + 1/q = 1. Moreover, the notation Cp [respectively, Cq] is used
to denote the space Cp(H) [respectively, Cq(H)] without explicit reference to the
Hilbert space H.
Recall from [24] that a countable family G = {Gi}∞
i=1 of bounded linear oper-
ators from separable Banach space X to Cp is a von Neumann-Schatten p-frame
for X with respect to H if constants 0 < AG ≤ BG < ∞ exist such that
AGkfkX ≤ ∞Xi=1
Cp! 1
kGi(f )kp
p
≤ BGkfkX
(2.2)
f 7→ {Gi(f )}∞
i=1 ,
∞Xi=1
{Ai}∞
i=1 7→
AiGi.
It is called a von Neumann-Schatten p-Bessel sequence with
for all f ∈ X .
bound BG if the second inequality holds. In [24], the authors have shown that
the von Neumann-Schatten p-frame condition is satisfied if and only if {Ai}∞
i=1 7→
P∞
i=1 AiGi is a well defined mapping from ⊕Cq onto X ∗. Motivated by this fact,
they considered the following operators:
UG : X → ⊕Cp;
(2.3)
and
TG : ⊕Cq → X ∗;
(2.4)
As usual, the operator UG is called the analysis operator, and TG is the synthesis
operator of G.
for X ∗ with respect to H if
Recall also from [24] that G is called a von Neumann-Schatten q-Riesz basis
(1) {f ∈ X : Gi(f ) = 0 ∀i ∈ N} = {0},
(2) there are positive constants AG and BG such that for any finite subset
i=1 ∈ ⊕Cq
I ⊆ N and {Ai}∞
Cq! 1
kAikq
AG Xi∈I
q
≤ kTG({Ai}∞
i=1)kX ∗ ≤ BG Xi∈I
q
Cq! 1
kAikq
.
The assumption of latter definition implies that P∞
tionally for all A = {Ai}∞
i=1 ∈ ⊕Cq and
AGkAkq ≤ kTG(A)kX ∗ ≤ BGkAkq.
(2.5)
Thus G ⊆ B(X ,Cp) is a von Neumann-Schatten q-Riesz basis for X ∗ with respect
to H if and only if the operator TG defined in (2.4) is both bounded and bounded
below. Specially, in this case the operators TG and UG are bijective.
i=1 AiGi converges uncondi-
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
5
The reader will remark that if H = C, then B(H) = Cp = C and thus ⊕Cp = ℓp.
Hence the above definitions is consistent with the corresponding definitions in the
concept of p-frames for separable Banach spaces.
We conclude this section with the following result which can be proved with a
similar argument as in the proof of [9, Corollary 2.5].
Lemma 2.1. Let X be a reflexive Banach space and let G = {Gi}∞
i=1 ⊆ B(X ,Cp)
be a von Neumann-Schatten q-Riesz basis for X ∗ with respect to H. If the q-Riesz
basis bounds of G are AG and BG, then G is a von Neumann-Schatten p-frame for
X with p-frame bounds AG and BG.
3. von Neumann-Schatten Bessel multipliers: Basic results
All over in this section X1 and X2 are separable Banach spaces and the space
Our starting point of this section is the following lemma which play a crucial
ℓr (1 ≤ r ≤ ∞) has its usual meanings.
rule in this paper.
Lemma 3.1. Let G ⊆ B(X ∗
1 ,Cp) be a von Neumann-Schatten p-Bessel sequence
with bound BG and F ⊆ B(X2,Cq) be a von Neumann-Schatten q-Bessel sequence
1 → X ∗
with bound BF . If m = {mi}∞
defined by
i=1 ∈ ℓ∞, then the operator Mm,F ,G : X ∗
2
∞Xi=1
(f ∈ X ∗
1 ),
Mm,F ,G(f ) =
miGi(f )Fi
is well-defined and kMm,F ,Gkop ≤ BF BGkmk∞.
Proof. It is easy to check that {miGi(f )}∞
hand
i=1 ∈ ⊕Cp for all f ∈ X ∗
(f ∈ X ∗
1 ).
It follows that Mm,F ,G is well defined. Moreover, we observe that
1 . On the other
nXi=1
(cid:13)(cid:13)(cid:13)
miGi(f )Fi(cid:13)(cid:13)(cid:13)op
= sup
≤ sup
i=1(cid:17)
Mm,F ,G(f ) = TF(cid:16){miGi(f )}∞
h∈X2,khk≤1n(cid:12)(cid:12)(cid:12)
tr(cid:16)miGi(f )Fi(h)(cid:17)(cid:12)(cid:12)(cid:12)o
nXi=1
h∈X2,khk≤1n nXi=1(cid:12)(cid:12)(cid:12)tr(cid:16)miGi(f )Fi(h)(cid:17)(cid:12)(cid:12)(cid:12)o
h∈X2,khk≤1n nXi=1(cid:13)(cid:13)(cid:13)miGi(f )Fi(h)(cid:13)(cid:13)(cid:13)C1o
kmiGi(f )kCpkFi(h)kCqo
h∈X2,khk≤1n nXi=1
≤ kmk∞(cid:16) nXi=1
Cp(cid:17) 1
h∈X2,khk≤1(cid:16) nXi=1
kGi(f )kp
≤ sup
≤ sup
sup
p
≤ BF BGkmk∞kfk,
Cq(cid:17) 1
kFi(h)kq
q
Mm,F ,G(f ) =
(f ∈ X ∗
1 )
1 → X ∗
miGi(f )Fi
∞Xi=1
6
H. JAVANSHIRI AND M. CHOUBIN
for all n ∈ N and f ∈ X ∗
is bounded with BF BGkmk∞.
1 . From this, we can deduced that the operator Mm,F ,G
(cid:3)
Now we are in position to introduce the main object of study of this work.
Definition 3.2. Let G ⊆ B(X ∗
1 ,Cp) be a von Neumann-Schatten p-Bessel se-
quence, and let F ⊆ B(X2,Cq) be a von Neumann-Schatten q-Bessel sequence.
Let m ∈ ℓ∞. The operator Mm,F ,G : X ∗
2 defined by
is called von Neumann-Schatten (p, q)-Bessel multiplier and the sequence m is
called its symbol.
If m is a sequence in ℓr (1 ≤ r ≤ ∞), then the mapping
i=1 7→ {miAi}∞
i=1,
Mp,m : ⊕Cp → ⊕Cp; {Ai}∞
is well-defined and bounded. Hence, the von Neumann-Schatten (p, q)-Bessel
multiplier Mm,F ,G can be written as
Mm,F ,G = TFMp,mUG.
This equality paves the way for the study of some operator properties of Mm,F ,G
in terms of the properties of Mp,m. To this end, we need the following remark
and lemma.
Remark 3.3. Recall from [23] that {en ⊗ em : n, m ∈ I} is an orthonormal basis
of C2. For the convenience of citation and a better exposition we denote by {Ek}
the orthonormal basis of C2. Hence, putting
Fi,k = {δi,jEk}∞
(cid:16)i, k ∈ N(cid:17),
(1) in the case where dimH = N, then n{Fi,k}N 2
k=1o∞
(2) in the case where H is an infinite dimensional Hilbert space, thenn{Fi,k}∞
basis for ⊕C2.
is an orthonormal
one can see that
is an orthonormal basis for ⊕C2.
j=1
i=1
k=1o∞
i=1
In what follows, the notation m is used to denote the sequence {mi}∞
mi refers to the complex conjugate of mi.
i=1, where
Lemma 3.4. The following assertions hold.
(1) If m ∈ ℓ∞, then kMp,mkop = kmk∞.
(2) M∗
(3) If dimH = N and m ∈ ℓp, then M2,m ∈ Cp(⊕C2) and
2,m = M2,m.
kM2,mkCp(⊕C2) = N 2kmkp.
hM2,mA,Bi =
=
= hA,M2,mBi.
[miAi,Bi]tr
[Ai, miBi]tr
∞Xi=1
∞Xi=1
N 2Xk=1(cid:10)A, Fi,k(cid:11)Fi,k!
mi(cid:10)A, Fi,k(cid:11)Fi,k.
=
M2,m(A) = M2,m ∞Xi=1
N 2Xk=1
∞Xi=1
(cid:16)l ∈ N ∪ {0}, lN 2 + 1 ≤ j ≤ (l + 1)N 2(cid:17),
j=1 = {m1,· · · , m1
, m2,· · · , m2
,· · ·},
}
m = { mj}∞
N 2
{z
∞Xj=1
N 2
}
{z
mj(cid:10)A, Fj(cid:11) Fj.
M2,m(A) =
(3) If A ∈ ⊕C2, then we have
Hence, if we set
Fj = Fl+1,j−lN 2
and
then we observe that
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
7
Proof. (1) That kMp,mkop ≤ kmk∞ is trivial.
In order to prove that kmk∞ ≤
kMp,mkop, suppose that j ∈ N and x ∈ H with kxk = 1. Observe that, if
A(j) = {δi,j · x ⊗ x}∞
i=1, then kA(j)kp = 1 and thus
kMp,mkop ≥ kmA(j)kp ≥ kmj · x ⊗ xkCp = mj.
It follows that kmk∞ ≤ kMp,mkop.
(2) It is suffices to note that if A,B ∈ ⊕C2, then
Therefore, M2,m is in the Schatten p-class of ⊕C2; this is because of,
∞Xj=1
mjp ≤ N 2kmkp
p.
Now, in order to prove that kM2,mkCp = N 2kmkp, first note that M2,m =
M2,m. Hence, we observe that
kM2,mkp
Cp = tr(M2,mp) =
=
N 2Xk=1(cid:10)M2,mp(Fi,k), Fi,k(cid:11)
∞Xi=1
∞Xj=1DM2, mp( Fj), FjE
8
H. JAVANSHIRI AND M. CHOUBIN
[mjpEk,Ek]tr
= N 2kmkp
p.
We have now completed the proof of the lemma.
=
∞Xj=1
N 2Xk=1
(cid:3)
By applying Lemma 3.4 with H = C, one can obtain the following improvement
of [3, Lemma 5.4(3)].
Lemma 3.5. If m ∈ ℓp, then the operator
Mm : ℓ2 → ℓ2; {ci}∞
i=1 7→ {mici}∞
i=1,
is in the Schatten p-class of ℓ2. In particular, kMmkCp(ℓ2) = kmkp.
As was mentioned in section 2, in the case where p = 2, the spaces C2 and ⊕2C2
are Hilbert. Motivated by this fact the authors of [1, 24] provided a detailed
study of the duals of a von Neumann-Schatten 2-frame for Hilbert space K with
respect to H. Let us recall from [24] that, a sequence G ⊆ B(K,C2) is said to be
a Hilbert -- Schmidt frame or simply a HS-frame for K with respect to H, if there
exist two positive numbers AG and BG such that
AGkfk2
K ≤
kGi(f )k2
C2 ≤ BGkfk2
K
∞Xi=1
Particularly, by using the Hilbert properties of the spaces, they observed that
i=1
and
TG({Ai}∞
UG(f ) = {Gi(f )}∞
where f ∈ K and {Ai}∞
i=1 ∈ ⊕C2. Moreover, they showed that the mapping
SG := TGUG is a bounded, invertible, self-adjoint and positive operator, and they
i=1 the canonical dual HS-frame of G. It is
worthwhile to mention that the HS-frames is a more general version of g-frames,
an important generalization of ordinary frames.
called the HS-frame eG := {GiS−1
G∗
i Ai,
G }∞
i=1) =
∞Xi=1
The following remark is a very useful tool in our study of HS-Bessel multiplier
when H is a finite dimensional space.
Remark 3.6. Suppose that G and F are HS-Bessel sequences for K with respect
to H and that m ∈ ℓ∞. For each f ∈ K, we observe that
Mm,F ,G(f ) = TFMp,mUG(f )
i Gi(f )
miF ∗
=
∞Xi=1
∞Xi=1 Xn,m∈I
=
mi(cid:10)f,G∗
i (en ⊗ em)(cid:11)F ∗
i (en ⊗ em).
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
9
In particular, if dim(H) = N, then
Mm,F ,G =
mi(cid:16)F ∗
∞Xi=1
NXn,m=1
i (en ⊗ em) ⊗ G∗
n,m=1oi∈N
Hence, if we set Φ = n{F ∗
i (en ⊗ em)}N 2
{z
}
, Ψ = n{G∗
}
m = {m1,· · · , m1
, m2,· · · , m2
,· · ·},
N 2
{z
and
N 2
i (en ⊗ em)(cid:17).
then Φ and Ψ are ordinary Bessel sequences and the operator Mm,F ,G is equal to
the operator M m,Φ,Ψ in the sense of Balazs [3, Definition 5.1] for ordinary Bessel
sequences. The reader will remark that in this case G and F are HS-Riesz basis
if and only if Φ and Ψ are ordinary Riesz basis, see [1, Theorem 3.3].
i (en ⊗ em)}N 2
n,m=1oi∈N
In the case where dim(H) < ∞, Remark 3.6 paves the way for obtaining some
properties of HS-Bessel multipliers from [3, 26]. In details, as an application of
this remark and Lemma 3.4, by a method similar to that of [3, Theorems 6.1 and
8.1] one can easily obtain the following generalization of those theorems. The
details are omitted.
Let F (l) = {F (l)
i }∞
i=1 be a sequence in B(K,C2) indexed by l ∈ N. We say that:
(1) The sequence F (l) converges uniformly to some sequence F ⊆ B(K,C2)
with respect to operator norm, if for i −→ ∞ we have
l {kF (l)
sup
i − Fikop} −→ 0.
(2) The sequence F (l) converges to some sequence F ⊆ B(K,C2) in ℓ2-sense
if
∀ε > 0 ∃N ∈ N such that
∀l ≥ N
(cid:16) ∞Xi=1
kF (l)
i − Fik2
Cp(cid:17)1/2
< ε.
Proposition 3.7. Suppose that G and F are HS-Bessel sequences for K with
respect to H and that dimH < ∞. Then the following assertions hold.
(1) If m ∈ c0, then Mm,F ,G is compact.
(2) If m ∈ ℓp, then Mm,F ,G is in the Schatten p-class of K and
(3) Let m(l) = {m
i }∞
kMm,F ,GkCp(K) ≤pBGBF (dimH)2kmkp.
(l)
i=1 be a sequence indexed by l converge to m in ℓp, then
kMm(l),F ,G − Mm,F ,GkCp(K) −→ 0.
(4) For the sequences F (l) ⊆ B(K,C2) indexed by l ∈ N, we can say that
(a) If m ∈ ℓ1 and the sequence F (l) is a HS-Bessel sequence converging
uniformly to F with respect to operator norm, then
kMm,F (l),G − Mm,F ,GkC1(K) −→ 0.
10
H. JAVANSHIRI AND M. CHOUBIN
(b) If m ∈ ℓ2 and the sequence F (l) converge to F in an ℓ2-sense, then
kMm,F (l),G − Mm,F ,GkC2(K) −→ 0.
in (4) apply.
(5) For HS-Bessel sequences G(l) converge to G, corresponding properties as
(6) (a) Let m(l) −→ m in ℓ1, F (l) and G(l) be HS-Bessel sequences with
bounds BF (l) and BG(l) such that supl BF (l) < ∞ and supl BG(l) < ∞.
If the sequences F (l) and G(l) converge uniformly to F respectively G
with respect to operator norm, then
kMm(l),F (l),G(l) − Mm,F ,GkC1(K) −→ 0.
(b) Let m(l) −→ m in ℓ2 and let F (l) respectively G(l) converge to F
respectively G in an ℓ2-sense, then
kMm(l),F (l),G(l) − Mm,F ,GkC2(K) −→ 0.
By another application of Remark 3.6 with the aid of [26, Theorem 5.1] we
have the following generalization of that theorem. In this proposition and in the
sequel the sequence m is called semi-normalized if 0 < inf i mi ≤ supi mi < ∞;
In this case, the notation 1/m is used to denote the sequence {1/mi}∞
i=1.
Proposition 3.8. Suppose that F is a HS -- Riesz basis for K with respect to H
and that dimH < ∞. Then the following assertions hold.
is semi -- normalized.
(1) If G is a HS-Riesz basis, then Mm,F ,G is invertible on K if and only if m
(2) If m is semi -- normalized, then Mm,F ,G is invertible on K if and only if G
is a HS -- Riesz basis.
There does not seem to be an easy way to extend Propositions 3.7 and 3.8 to
infinite dimensional case. However, we have the following result.
Proposition 3.9. Let X1 be a reflexive Banach space and G ⊆ B(X ∗
1 ,Cp) be a
von Neumann-Schatten q-Riesz basis for X1 with bounds AG and BG. Let also
F ⊆ B(X2,Cq) be a von Neumann-Schatten p-Riesz basis for X ∗
2 with bounds AF
and BF . Then for each m ∈ ℓ∞ we have
AF AGkmk∞ ≤ kMm,F ,Gkop ≤ BF BGkmk∞.
1 ,X ∗
2 ) is injective.
In particular, the operator Mm,F ,G is invertible and the mapping m 7→ Mm,F ,G
from ℓ∞ into B(X ∗
Proof. By Lemma 3.1, it will be enough to prove that we have the lower bound.
To this end, suppose that x is an arbitrary element of H with kxk = 1 and
j ∈ N. Then, A(j) = {δi,j · x ⊗ x}∞
i=1 ∈ ⊕Cp. Thus, by the surjectivity of
the operator UG, there exists an element fj ∈ X ∗
1 such that UG(fj) = A(j). In
Cp = 1. We now invoke Lemma 2.1 to conclude that
particular, P∞
i=1 kGi(fj)kp
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
11
1
1 ≤ 1
BG ≤ kfjkX ∗
p-Riesz basis for X ∗
. Also, since F ⊆ B(X2,Cq) is a Von Neumann-Schatten
AG
2 , Eq. (2.5) implies that
Now, we get
kTFMp,m(A(j))k ≥ AF(cid:13)(cid:13)Mp,m(A(j))(cid:13)(cid:13)p = AFmj.
kMm,F ,Gkop ≥ sup
j
and this completes the proof.
kTFMp,mUG(fj)k
kfjk
≥ AF AGkmk∞,
(cid:3)
4. Invertibility of HS -- frame multipliers
In this section, we turn our attention to the study of invertible HS-frame
multipliers. First let us to note that, if m is semi-normalized and Mm,F ,G is
invertible, then the HS-Bessel sequences F and G are automatically HS-frames.
We also recall that a HS-frame Gd = {Gd
i=1 is called a dual of G if TGUGd = IdK.
The main result of this section is the following theorem.
j=1 and F = {Fj}∞
Theorem 4.1. Suppose that G = {Gj}∞
j=1 are HS-frames for
K with respect to H, and that the symbol m is semi -- normalized. If Mm,F ,G is an
invertible multiplier, then there exists a unique bounded operator Γ : K → ⊕2C2
such that
(4.1)
m,F ,G = M1/m, eG,F d + Γ∗UF d,
i }∞
M−1
for all dual HS-frames F d = {F d
Proof. Define Γ : K → ⊕2C2 by
j }∞
j=1 of F .
Γ(f ) := UF (M−1
m,F ,G)∗(f ) − M2,1/mUGS−1
G (f )
(f ∈ K).
(4.2)
Then the operator Γ is bounded and
M−1
m,F ,GTF = Γ∗ + S−1
G TGM2,1/m.
Using any dual HS-frame F d of F we get
M−1
m,F ,G = S−1
G TGM2,1/mUF d + Γ∗UF d.
(4.3)
It follows that M−1
m,F ,G = M1/m, eG,F d + Γ∗UF d for all dual HS-frames F d of F .
Having reached this state it remains to show that Γ is uniquely determined. To
this end, suppose on the contrary that Eq. (4.3) are hold for two operators Γ1
and Γ2. It follows that
(Γ1 − Γ2)∗UF d = 0,
for all dual HS-frames F d of F . In particular, for each f ∈ K, we have
(Γ1 − Γ2)∗({FjS−1
F (f )}∞
j=1) = 0.
Now, let i and k be arbitrary elements in N and Fi,k = {δi,jEk}∞
duced in Remark 3.3. If for each j ∈ N, we define
j=1, which intro-
(4.4)
(4.5)
F ′
i,j,k : K → C2;
f 7→(cid:10)f, e′
1(cid:11)δi,jEk,
12
H. JAVANSHIRI AND M. CHOUBIN
where {e′
i,j,k}∞
the sequence F ′
and the HS-Bessel sequence
l}l∈L is an orthonormal basis for K. Then it is not hard to check that
j=1 is a HS-Bessel sequence for K with respect to H
i,k = {F ′
F d
i,k = {FjS−1
F + F ′
i,j,k − FjS−1
F TF UF ′
i,k}∞
j=1
is a dual HS-frame of F . Therefore, Eq. (4.4) and (4.5), implies that
i,k
(f )
i,j,k(f ) − FjS−1
0 = (Γ1 − Γ2)∗UF d
= (Γ1 − Γ2)∗(cid:16){FjS−1
F (f ) + F ′
= (Γ1 − Γ2)∗(cid:0){F ′
j=1(cid:1)
i,j,k(f )}∞
j=1(cid:1) = (Γ1 − Γ2)∗(Fi,k).
0 = (Γ1 − Γ2)∗(cid:0){F ′
1)}∞
F TF UF ′
i,j,k(e′
i,k
j=1(cid:17)
(f )}∞
for all f ∈ K. Hence, we have
This says that (Γ1 − Γ2)∗(Fi,k) = 0 for all i, k ∈ N. We now invoke Remark 3.3
to conclude that Γ1 = Γ2 and this completes the proof of the theorem.
(cid:3)
For operator Γ in Proposition 4.7 it is not hard to check that TGM2,mΓ = 0.
It follows that, if G is a HS-Riesz basis and m is semi -- normalized, then
for all dual HS-frames F d = {F d
This observation together with Proposition 3.9 give the following result.
M−1
m,F ,G = M1/m, eG,F d,
j }∞
j=1 of F .
Corollary 4.2. Let G and F be HS -- Riesz basis for K with respect to H and let
m be semi -- normalized. Then Mm,F ,G is invertible on K and
M−1
m,F ,G = M1/m, eG, eF.
The proof of Theorem 4.5 below relies on the following remark and proposition.
Remark 4.3. Following [20, Remark 2.8], we say that the HS-frame Ggd = {Ggd
i }
is a generalized dual HS-frame of G, if TGUGgd is invertible. It is noteworthy that
with an argument similar to the proof of [20, Theorem 2.1] and [1, Theorem
3.1] one can show that the generalized dual HS-frames of G are precisely the
sequences Ggd such that
Ggd
i = GiS−1
G Q + πiΨ,
(i ∈ N)
where Ψ is a bounded operator in B(K,⊕C2) such that TGΨ = 0, πi : ⊕C2 → C2
is the standard projection on the i-th component and Q is an invertible operator
in B(K). In what follows, the notation GD(G) is used to denote the set of all
generalized HS-duals of G.
In the following result and in the sequel Inv(G, m) refers to the set of all HS-
Bessel sequence F such that the operator Mm,F ,G is invertible.
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
13
Proposition 4.4. Suppose that G is a HS-frame for K with respect to H and
that m is a semi-normalized sequence. Then the mapping
Θ : Inv(G, m) → GD(G);
{Fi}∞
i=1 7→ {GiS−1
G Mm,F ,G + πiM2,mΓ}∞
i=1,
is bijective, where Γ is the unique operator in B(K,⊕2C2) which satisfies the
equality (4.1).
Proof. It obviously suffices to show that Θ is an onto map. To this end, suppose
that Ggd is a generalized HS-dual of G. Then, by Remark 4.3, we would have a
bounded operator in B(K,⊕C2) and an invertible operator Q in B(K) such that
TGΨ = 0 and
Ggd
i = GiS−1
Letting Fi : K → C2 by
G Q + πiΨ,
(i ∈ N).
Fi = (1/mi)GiS−1
G Q∗ + πi(M2,1/mΨ).
Then F is a HS-Bessel sequence and, in particular, we have Q = Mm,F ,G. Hence,
F is in Inv(G, m). On the other hand, if Γ is the unique operator which defined
by (4.2), then, for each f ∈ K, we have
Γ(f ) = UF (M−1
m,F ,G)∗(f ) − M2,1/mUGS−1
G (f )
G Q∗(M−1
m,F ,G)∗(f ) + M2,1/mΨ(f ) − M2,1/mUGS−1
G (f )
= M2,1/mUGS−1
= M2,1/mΨ(f ).
It follows that Θ(F ) = Ggd.
Theorem 4.5. Let G be a HS-frame for K with respect to H and let G′ be another
HS-frame for K with respect to H such that kTG−TG′k < √AG/2, where AG is the
lower frame bound of G. If m is semi-normalized, then there exists a one-to-one
correspondence between Inv(G, m) and Inv(G′, m).
Proof. By Proposition 4.4 above, it will be enough to prove that there exists a
one-to-one correspondence between GD(G) and GD(G′). To this end, we define
the map Λ from GD(G) into GD(G′) by
(cid:3)
Ggd 7→ {G′
iSG′TGUGgd + πiPker(TG′ )UGgd}∞
i=1,
where Pker(TG′ ) denotes the orthogonal projection of ⊕C2 onto ker(TG′). Now, with
an argument similar to the proof of Proposition 4.15 in [22] one can show that Λ
is bijective and this completes the proof.
(cid:3)
The next result characterizes another invertible HS-frame multipliers Mm,F ,G
In details, the following result is
whose inverses can be written as M1/m, eG, eF .
a generalization of a result proved by Balazs and Stoeva [5, Theorem 4.6] to
HS-frames as well as g-frames. To this end, we need the following remark.
Remark 4.6. Suppose that G and F are HS-frames for K with respect to H,
and that the symbol m is semi -- normalized.
14
H. JAVANSHIRI AND M. CHOUBIN
(1) If the HS-frames mF := {miFi}∞
i=1 and G are equivalent; that is, there
exists an invertible operator Q in B(K) such that miFi = GiQ (i ∈ N),
then Mm,F ,G is invertible and
M−1
m,F ,G = M1/m, eG,F d,
for all dual HS-frames F d of F . Indeed, on the one hand we have
Mm,F ,G = TFM2,mUG = TmF UG = Q∗SG,
and on the other hand, if the letter F Q−1 respectively (1/m)G refer to
the HS-Bessel sequence {FiQ−1}∞
i=1, then we
observe that
i=1 respectively {(1/mi)Gi}∞
M1/m, eG,F d = T eGM2,1/mUF d
G T(1/m)GUF d
G TF Q−1UF d
= S−1
= S−1
= (Q∗SG)−1.
Conversely, if Mm,F ,G is invertible and M−1
m,F ,G = M1/m, eG,F d, for all dual
HS-frames F d of F , then Theorem 4.7 implies that Γ∗UF d = 0. Now,
with an argument similar to the proof of Theorem 4.7 one can conclude
that Γ = 0. From this, by Eq. (4.2), we deduce that
(i ∈ N),
miFi = GiS−1
G Mm,G,F
and thus the HS-frames G and mF are equivalent. Hence, we can give
the interpretation below for equivalent HS-frames:
"If m is semi -- normalized, then Mm,F ,G is invertible and M−1
m,F ,G = M1/m, eG,F d,
for all dual HS-frames F d of F if and only if the HS-frames mF :=
{miFi}∞
(2) If Y denotes any one of the HS-frames G and F , then it is easy to check
that
i=1 and G are equivalent."
⊕C2 = ran(UY ) ⊕ ker(TY)
and thus Pker(TY ) + Pran(UY ) = Id⊕C2, where PX denotes the orthogonal
projection of ⊕C2 onto X. In particular, we have
Pran(UG ) = UGT eG
and
Pran(U eF ) = U eF TF .
The following proposition is now immediate.
Proposition 4.7. Suppose that G and F are HS-frames for K with respect to
H, and that there exists a non-zero constant c such that mi = c for all i ∈ N.
Then the following statements are equivalent.
(1) Mm,F ,G is invertible and M−1
(2) F and G are equivalent HS-frames.
m,F ,G = M1/m, eG, eF.
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
15
Proof. We first note that, without loss of generality, we may consider c = 1.
The necessity of the condition "F and G are equivalent HS-frames" follows from
part (1) of Remark 4.6. We prove its sufficiency. To this end, suppose that the
condition (1) is satisfied. From this, we observe that
UGT eG = UGM1/m, eG, eFMm,F ,GT eG = UGT eGU eF TF UGT eG.
We now invoke part (2) of Remark 4.6 to conclude that
Pran(UG )Pran(U eF )Pran(UG ) = Pran(UG ).
This says that ran(U eF ) ⊆ ran(UG) and thus
ran(UF ) ⊆ ran(UG).
Analogously, one can show that the reverse inclusion is also true. It follows that
ran(UF ) = ran(UG). Now, we follow the proof of [2, Lemma 2.1] to show that
there exists an invertible operator Q ∈ B(K) such that Gi = FiQ (i ∈ N). To
F }. Observe that G′ and F ′
this end, suppose that G′ = {GiS
are HS-frames with lower and upper frame bounds 1 and thus SG′ = IdK = SF ′.
Moreover, it is not hard to check that ran(UF ′) = ran(UG′) and thus
G } and F ′ = {FiS
− 1
2
− 1
2
UG′TG′ = Pran(UG′ ) = Pran(UF ′ ) = UF ′TF ′.
Hence, if we set Q = TG′UF ′, then
This says that Q is an isometry, Q∗(f ) =P∞
Q∗Q = TF ′UG′TG′UF ′ = IdK.
∗G′
(f ∈ K).
i=1 F ′
iQ(f )
∗G′
i(f ) and
i
∞Xi=1
i
f =
F ′
Now, for each f ∈ K, we have
iQ(f ) − F ′
kG′
∞Xi=1
i(f )k2
C2 =
i(f )k2
C2
C2 +
kG′
iQ(f )k2
∞Xi=1
− 2Re(cid:16)h ∞Xi=1
∞Xi=1
kF ′
iQ, fitr(cid:17)
∗G′
= kQ(f )k2 − kfk2 = 0
F ′
i
It follows that G′
i (i ∈ N). Having reached this state it remains to
prove that Q is onto or equivalently ker(Q∗) = {0}. To this end, suppose that
g ∈ ker(Q∗). Hence, UG′(g) ∈ ker(TF ′) and thus, since
iQ = F ′
ker(TF ′) = ran(UF ′)⊥ = ran(UG′)⊥ = ker(TG′),
we can deduce that UG′(g) ∈ ker(TG′). It follows that g = TG′UG′(g) = 0. We
have now completed the proof of the proposition.
(cid:3)
16
H. JAVANSHIRI AND M. CHOUBIN
We conclude this work by the following result which is of interest in its own
right. In details, as an another application of Theorem 4.7, we have the following
surprising new results about dual HS-frames as well as dual g-frames which shows
that a HS-frame [respectively, g-frame] is uniquely determined by the set of its
dual HS-frames [respectively, g-frames].
Theorem 4.8. Let G and F be HS-frames for K with respect to H. If every dual
HS-frame Gd of G is a dual HS-frame of F , then G = F .
Proof. The assumption implies that (TG − TF )UGd = 0 for all dual HS-frames Gd
of G. Hence, with an argument similar to the proof of Theorem 4.7 one can show
that TG = TF . Whence G = F .
(cid:3)
References
[1] A. Arefijamaal, G. Sadeghi, von Neumann -- Schatten Dual Frames and their Pertur-
bations, Results. Math. 69 (2015), 431 -- 441.
[2] R. Balan, Equivalence relations and distances between Hilbert frames, Proc. Amer. Math.
Soc. 127 (1999), 2353 -- 2366.
[3] P. Balazs, Basic definition and properties of Bessel multipliers, J. Math. Anal. Appl. 325
(2007), 571 -- 85.
[4] P. Balazs, Hilbert -- Schmidt operators and frames -- classification, best approximation by
multipliers and algorithms, Int. J. Wavelets Multiresolut. Inf. Process. 6 (2008), 315 -- 330.
[5] P. Balazs and D. T. Stoeva, Representation of the inverse of a frame multiplier, J.
Math. Anal. Appl. 422 (2015), 981 -- 994.
[6] J. Benedetto and G. Pfander, Frame expansions for Gabor multipliers, Appl. Comput.
Harmon. Anal. 20 (2006), 26 -- 40.
[7] O. Christensen, Introduction to Frames and Riesz Bases, Birkhauser, (2003).
[8] O. Christensen and R. S. Laugesen, Approximately dual frame pairs in Hilbert
spaces and applications to Gabor frames, Sampl. Theor. Signal Image Process. 9 (2010),
77 -- 89.
[9] O. Christensen and D. Stoeva, p -- frames in separable Banach spaces, Adv. Comput.
Math. 18 (2003), 117 -- 126.
[10] E. Cordero and F. Nicola, Remarks on Fourier multipliers and applications to the
wave equation, J. Math. Anal. Appl. 353 (2009), 583 -- 591.
[11] E. Cordero and K. Grochenig, Necessary conditions for Schatten class localization
operators, Proc. Amer. Math. Soc. 133 (2005), 3573 -- 3579
[12] E. Cordero, K. Grochenig and F. Nicola, Approximation of Fourier integral oper-
ators by Gabor multipliers, J. Fourier Anal. Appl. 18 (2012), 661 -- 684.
[13] H. G. Feichtinger, M. Hampejs and G. Kracher, Approximation of matrices by
Gabor multipliers, IEEE Signal Process. Lett. 11 (2004), 883 -- 886.
[14] H.G. Feichtinger and G. Narimani, Fourier multipliers of classical modulation spaces,
Appl. Comput. Harmon. Anal. 21 (2006), 349 -- 359.
[15] H.G. Feichtinger and K. Nowak, A First Survey of Gabor Multipliers. Advances in
Gabor analysis. Appl. Numer. Harmon. Anal., pp. 99 -- 128. Birkhauser, Boston (2003).
[16] H. G. Feichtinger and T. Werther, Atomic system for subspaces, in: L. Zayed (Ed.),
Proceedings SampTA 2001, Orlando, FL, (2001) 163 -- 165.
[17] F. Futamura, Frame diagonalization of matrices, Linear Algebra Appl. 436 (2012) 3201 --
3214.
[18] L Gavruta, Frames for operators, Appl. Comput. Harmon. Anal. 32 (2012), 139 -- 144.
VON NEUMANN-SCHATTEN BESSEL SEQUENCES
17
[19] K. Grochenig, Representation and approximation of pseudo differential operators by
sums of Gabor multipliers, Appl. Anal. 90 (2011), 385 -- 401.
[20] H. javanshiri, Some properties of approximately dual frames in Hilbert spaces, Results.
Math. 70 (2016), 475 -- 485.
[21] H. Javanshiri and A. Fattahi, Continuous atomic systems for subspaces, Mediterr. J.
Math. 13 (2016), 1871 -- 1884.
[22] G. Kutyniok, V. Paternostro and F. Philipp, The effect of perturbations of frame
sequences and fusion frames on their duals, Preprint, arXiv:1509.04160.
[23] J. R. Ringrose, Compact Non -- Self -- Adjoint Operators, Van Nostrand Reinhold Company,
1971.
[24] G. Sadeghi, A. Arefijamaal, von Neumann -- Schatten frames in separable Banach
spaces, Mediterr. J. Math. 9 (2012), 525 -- 535.
[25] R. Schatten, Norm Ideals of Completely Continuous Operators, Springer Berlin, 1960.
[26] D.T. Stoeva and P. Balazs, Invertibility of multipliers, Appl. Comput. Harmon. Anal.
33 (2012), 292 -- 299.
[27] W. C. Sun, G-frames and G-Riesz bases, J. Math. Anal. Appl. 322 (2006) 437 -- 452.
Department of Mathematics, Yazd University, P.O. Box: 89195-741, Yazd, Iran
E-mail address: [email protected]
Department of Mathematics, Velayat University, Iranshahr, Iran
E-mail address: [email protected]
|
1212.6928 | 1 | 1212 | 2012-12-31T17:09:37 | Generalized local Morrey spaces and fractional integral operators with rough kernel | [
"math.FA",
"math.AP"
] | Let $M_{\Omega,\a}$ and $I_{\Omega,\a}$ be the fractional maximal and integral operators with rough kernels, where $0 < \a < n$. In this paper, we shall study the continuity properties of $M_{\Omega,\a}$ and $I_{\Omega,\a}$ on the generalized local Morrey spaces $LM_{p,\varphi}^{{x_0}}$. The boundedness of their commutators with local Campanato functions is also obtained. | math.FA | math |
Generalized local Morrey spaces and fractional
integral operators with rough kernel
Vagif S. Guliyeva,b,1
aDepartment of Mathematics, Ahi Evran University, Kirsehir, Turkey
bInstitute of Mathematics and Mechanics of NAS of Azerbaijan, Baku
Abstract
Let MΩ,α and IΩ,α be the fractional maximal and integral operators
with rough kernels, where 0 < α < n. In this paper, we shall study the
continuity properties of MΩ,α and IΩ,α on the generalized local Morrey
{x0}
spaces LM
p,ϕ . The boundedness of their commutators with local Cam-
panato functions is also obtained.
AMS Mathematics Subject Classification:
Key words: fractional integral operator; rough kernels; generalized local Morrey
space; commutator; local Campanato space
42B20, 42B25, 42B35
1
Introduction
For x ∈ Rn and r > 0, let B(x, r) denote the open ball centered at x of radius
r and B(x, r) is the Lebesgue measure of the ball B(x, r). Let Ω ∈ Ls(Sn−1)
be homogeneous of degree zero on Rn, where Sn−1 denotes the unit sphere of Rn
(n ≥ 2) equipped with the normalized Lebesgue measure dσ and s > 1. For any
0 < α < n, then the fractional integral operator with rough kernel IΩ,α is defined
by
IΩ,αf (x) =ZRn
Ω(x − y)
x − yn−α f (y)dy
and a related fractional maximal operator with rough kernel MΩ,α is defined by
MΩ,αf (x) = sup
t>0
B(x, t)−1+ α
nZB(x,t)
Ω(x − y) f (y)dy.
1 The research of V. Guliyev was partially supported by the grant of Science Development
Foundation under the President of the Republic of Azerbaijan project EIF-2010-1(1)-40/06-1
and by the Scientific and Technological Research Council of Turkey (TUBITAK Project No:
110T695) and by the grant of 2010-Ahi Evran University Scientific Research Projects (PYO-
FEN 4001.12.18).
E-mail adresses: [email protected] (V.S. Guliyev).
1
If α = 0, then MΩ ≡ MΩ,0 is the Hardy-Littlewood maximal operator with
rough kernel. It is obvious that when Ω ≡ 1, IΩ,α is the Riesz potential Iα and
MΩ,α is the maximal operator Mα.
Theorem A Suppose that Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homogeneous of
n . If s′ < p or q < s, then
degree zero. Let 0 < α < n, 1 ≤ p < n
the operators MΩ,α and IΩ,α are bounded bounded from Lp(Rn) to Lq(Rn).
α , and 1
q = 1
p − α
Let b be a locally integrable function on Rn, then for 0 < α < n, we shall
define the commutators generated by fractional maximal and integral operators
with rough kernels and b as follows.
MΩ,b,α(f )(x) = sup
t>0
B(x, t)−1+ α
nZB(x,t)
b(x) − b(y)f (y)Ω(x − y)dy,
[b, IΩ,α]f (x) = b(x)IΩ,αf1(x) − IΩ,α(bf )(x)
=ZRn
Ω(x − y)
x − yn−α [b(x) − b(y)]f (y)dy.
Theorem B Suppose that Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homogeneous of
degree zero. Let 0 < α < n, 1 ≤ p < n
n and b ∈ BMO(Rn). If
s′ < p or q < s, then the operators MΩ,b,α and [b, IΩ,α] are bounded from Lp(Rn)
to Lq(Rn).
q = 1
p − α
α , 1
The classical Morrey spaces Mp,λ were first introduced by Morrey in [35] to
study the local behavior of solutions to second order elliptic partial differential
equations. For the boundedness of the Hardy-Littlewood maximal operator, the
fractional integral operator and the Calder´on-Zygmund singular integral opera-
tor on these spaces, we refer the readers to [1, 11, 39]. For the properties and
applications of classical Morrey spaces, see [12, 13, 22, 23] and references therein.
In the paper, we prove the boundedness of the operators IΩ,α from one general-
ized local Morrey space LM {x0}
q,ϕ2 , 1 < p < q < ∞, 1/p−1/q = α/n, and
from the space LM {x0}
q,ϕ2 , 1 < q < ∞, 1 − 1/q = α/n.
In the case b ∈ CBMOp2, we find the sufficient conditions on the pair (ϕ1, ϕ2)
which ensures the boundedness of the commutator operators [b, IΩ,α] from LM {x0}
p1,ϕ1
to LM {x0}
1,ϕ1 to the weak space W LM {x0}
q,ϕ2 , 1 < p < ∞, 1
p,ϕ1 to LM {x0}
, 1
q = 1
p − α
n , 1
q1
By A . B we mean that A ≤ CB with some positive constant C independent
of appropriate quantities. If A . B and B . A, we write A ≈ B and say that A
and B are equivalent.
p = 1
p1
+ 1
p2
= 1
p1
− α
n .
2
2 Generalized local Morrey spaces
We find it convenient to define the generalized Morrey spaces in the form as
follows.
Definition 2.1. Let ϕ(x, r) be a positive measurable function on Rn × (0, ∞)
and 1 ≤ p < ∞. We denote by Mp,ϕ ≡ Mp,ϕ(Rn) the generalized Morrey space,
the space of all functions f ∈ Lloc
p (Rn) with finite quasinorm
kf kMp,ϕ = sup
x∈Rn,r>0
ϕ(x, r)−1 B(x, r)− 1
p kf kLp(B(x,r)).
Also by W Mp,ϕ ≡ W Mp,ϕ(Rn) we denote the weak generalized Morrey space of
all functions f ∈ W Lloc
p (Rn) for which
kf kW Mp,ϕ = sup
x∈Rn,r>0
ϕ(x, r)−1 B(x, r)− 1
p kf kW Lp(B(x,r)) < ∞.
According to this definition, we recover the Morrey space Mp,λ and weak
Morrey space W Mp,λ under the choice ϕ(x, r) = r
λ−n
p :
Mp,λ = Mp,ϕ(cid:12)(cid:12)(cid:12)ϕ(x,r)=r
,
λ−n
p
W Mp,λ = W Mp,ϕ(cid:12)(cid:12)(cid:12)ϕ(x,r)=r
.
λ−n
p
Definition 2.2. Let ϕ(x, r) be a positive measurable function on Rn × (0, ∞)
and 1 ≤ p < ∞. We denote by LMp,ϕ ≡ LMp,ϕ(Rn) the generalized local Morrey
space, the space of all functions f ∈ Lloc
p (Rn) with finite quasinorm
kf kLMp,ϕ = sup
r>0
ϕ(0, r)−1 B(0, r)− 1
p kf kLp(B(0,r)).
Also by W LMp,ϕ ≡ W LMp,ϕ(Rn) we denote the weak generalized Morrey space
of all functions f ∈ W Lloc
p (Rn) for which
kf kW LMp,ϕ = sup
r>0
ϕ(0, r)−1 B(0, r)− 1
p kf kW Lp(B(0,r)) < ∞.
Definition 2.3. Let ϕ(x, r) be a positive measurable function on Rn × (0, ∞)
and 1 ≤ p < ∞. For any fixed x0 ∈ Rn we denote by LM {x0}
p,ϕ (Rn) the
generalized local Morrey space, the space of all functions f ∈ Lloc
p (Rn) with finite
quasinorm
p,ϕ ≡ LM {x0}
kf kLM
{x0}
p,ϕ
= kf (x0 + ·)kLMp,ϕ.
p,ϕ ≡ W LM {x0}
Also by W LM {x0}
space of all functions f ∈ W Lloc
p,ϕ (Rn) we denote the weak generalized Morrey
p (Rn) for which
kf kW LM
{x0}
p,ϕ
= kf (x0 + ·)kW LMp,ϕ < ∞.
3
According to this definition, we recover the local Morrey space LM {x0}
p,λ
and
weak local Morrey space W LM {x0}
p,λ under the choice ϕ(x0, r) = r
λ−n
p :
LM {x0}
p,λ = LM {x0}
W LM {x0}
p,λ = W LM {x0}
p,ϕ (cid:12)(cid:12)(cid:12)ϕ(x0,r)=r
,
λ−n
p
p,ϕ (cid:12)(cid:12)(cid:12)ϕ(x0,r)=r
.
λ−n
p
Wiener [45, 46] looked for a way to describe the behavior of a function at
the infinity. The conditions he considered are related to appropriate weighted Lq
spaces. Beurling [4] extended this idea and defined a pair of dual Banach spaces
Aq and Bq′, where 1/q + 1/q′ = 1. To be precise, Aq is a Banach algebra with
respect to the convolution, expressed as a union of certain weighted Lq spaces;
the space Bq′ is expressed as the intersection of the corresponding weighted Lq′
spaces. Feichtinger [24] observed that the space Bq can be described by
kf kBq
= sup
k≥0
2− kn
q kf χkkLq(Rn),
(2.1)
where χ0 is the characteristic function of the unit ball {x ∈ Rn : x ≤ 1}, χk is the
characteristic function of the annulus {x ∈ Rn : 2k−1 < x ≤ 2k}, k = 1, 2, . . ..
By duality, the space Aq(Rn), called Beurling algebra now, can be described by
kf kAq
=
ıXk=0
2− kn
q′ kf χkkLq(Rn).
(2.2)
Let Bq(Rn) and Aq(Rn) be the homogeneous versions of Bq(Rn) and Aq(Rn)
by taking k ∈ Z in (2.1) and (2.2) instead of k ≥ 0 there.
If λ < 0 or λ > n, then LM {x0}
p,λ (Rn) = Θ, where Θ is the set of all functions
equivalent to 0 on Rn. Note that LMp,0(Rn) = Lp(Rn) and LMp,n(Rn) = Bp(Rn).
Bp,µ = LMp,ϕ(cid:12)(cid:12)(cid:12)ϕ(0,r)=rµn
,
W Bp,µ = W LMp,ϕ(cid:12)(cid:12)(cid:12)ϕ(0,r)=rµn
.
Alvarez, Guzman-Partida and Lakey [3] in order to study the relationship
between central BMO spaces and Morrey spaces, they introduced λ-central
bounded mean oscillation spaces and central Morrey spaces Bp,µ(Rn) ≡ LMp,n+npµ(Rn),
µ ∈ [− 1
(Rn) =
Lp(Rn) and Bp,0(Rn) = Bp(Rn). Also define the weak central Morrey spaces
W Bp,µ(Rn) ≡ W LMp,n+npµ(Rn).
p or µ > 0, then Bp,µ(Rn) = Θ. Note that Bp,− 1
p , 0]. If µ < − 1
p
Inspired by this, we consider the boundedness of fractional integral operator
with rough kernel on generalized local Morrey spaces and give the central bounded
mean oscillation estimates for their commutators.
4
3
Fractional integral operator with rough ker-
nels in the spaces LM {x0}
p,ϕ
In this section we are going to use the following statement on the boundedness
of the weighted Hardy operator
wg(t) :=Z ∞
t
H ∗
g(s)w(s)ds, 0 < t < ∞,
where w is a fixed function non-negative and measurable on (0, ı).
Theorem 3.1. Let v1, v2 and w be positive almost everywhere and measurable
functions on (0, ı). The inequality
ess sup
t>0
v2(t)H ∗
wg(t) ≤ C ess sup
t>0
v1(t)g(t)
(3.1)
holds for some C > 0 for all non-negative and non-decreasing g on (0, ı) if and
only if
B := ess sup
t>0
v2(t)Z ∞
t
w(s)ds
ess sup
s<τ <∞
v1(τ )
< ı.
(3.2)
Moreover, if C ∗ is the minimal value of C in (3.1), then C ∗ = B.
Proof. Sufficiency. Assume that (3.2) holds. Whenever F , G are non-negative
functions on (0, ı) and F is non-decreasing, then
ess sup
F (t)G(t) = ess sup
F (t) ess sup
G(s),
t > 0.
(3.3)
t>0
t>0
s>t
By (3.3) we have
ess sup
t>0
v2(t)H ∗
wg(t) = ess sup
t>0
≤ ess sup
t>0
= ess sup
t>0
t
v2(t)Z ∞
v2(t)Z ∞
v2(t)Z ∞
t
t
g(s)w(s)
ess sup
s<τ <ı
ess sup
s<τ <ı
v1(τ )
v1(τ )
ds
w(s)ds
ess sup
s<τ <ı
v1(τ )
w(s)ds
ess sup
s<τ <ı
v1(τ )
ess sup
t>0
g(t) ess sup
t<τ <ı
v1(τ )
ess sup
t>0
g(t)v1(t)
≤ B ess sup
t>0
g(t)v1(t).
Necessity. Assume that the inequality (3.1) holds. The function
,
t > 0
v1(τ )
g(t) =
1
ess sup
t<τ <ı
5
is nonnegative and non-decreasing on (0, ı). Thus
B = ess sup
t>0
v2(t)Z ∞
t
w(s)ds
ess sup
s<τ <ı
v1(τ )
≤ C ess sup
t>0
v1(t)
ess sup
t<τ <ı
v1(τ )
≤ C,
hence C ∗ = B.
In [17] the following statements was proved by fractional integral operator
with rough kernels IΩ,α, containing the result in [34, 36].
Theorem 3.2. Suppose that Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homogeneous of
degree zero. Let 0 < α < n, 1 ≤ s′ < p < n
n and ϕ(x, r) satisfy
conditions
q = 1
p − α
α , 1
c−1ϕ(x, r) ≤ ϕ(x, t) ≤ c ϕ(x, r)
whenever r ≤ t ≤ 2r, where c (≥ 1) does not depend on t, r, x ∈ Rn and
Z ∞
r
tαpϕ(x, t)p dt
t
≤ C rαpϕ(x, r)p,
(3.4)
(3.5)
where C does not depend on x and r. Then the operators MΩ,α and IΩ,α are
bounded from Mp,ϕ to Mq,ϕ.
The following statements, containing results obtained in [34], [36] was proved
in [26, 28] (see also [5]-[8], [27, 29]).
Theorem 3.3. Let 0 < α < n, 1 ≤ p < n
condition
Z ∞
r
α , 1
q = 1
p − α
n and (ϕ1, ϕ2) satisfy the
tα−1ϕ1(0, t)dt ≤ C ϕ2(0, r),
(3.6)
where C does not depend on r. Then the operators Mα and Iα are bounded from
LMp,ϕ1 to LMq,ϕ2 for p > 1 and from LM1,ϕ1 to W LMq,ϕ2 for p = 1.
Lemma 3.4. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homogeneous
of degree zero. Let 0 < α < n, 1 ≤ p < n
n . Then, for p > 1 and
s′ ≤ p or q < s the inequality
α , and 1
q = 1
p − α
kIΩ,αf kLq(B(x0,r)) . r
n
q Z ı
2r
t− n
q −1kf kLp(B(x0,t))dt
holds for any ball B(x0, r) and for all f ∈ Lloc
Moreover, for p = 1 < q < s the inequality
p (Rn).
kIΩ,αf kW Lq(B(x0,r)) . r
n
q Z ı
2r
t− n
q −1kf kL1(B(x0,t))dt,
(3.7)
holds for any ball B(x0, r) and for all f ∈ L1loc.
6
p − α
Proof. Let 0 < α < n, 1 ≤ s′ ≤ p < n
ball centered at x0 and of radius r. We represent f as
α and 1
q = 1
n . Set B = B(x0, r) for the
f = f1 + f2,
f1(y) = f (y)χ2B(y),
f2(y) = f (y)χ ∁
(2B)(y),
r > 0,
(3.8)
and have
kIΩ,αf kLq(B) ≤ kIΩ,αf1kLq(B) + kIΩ,αf2kLq(B).
Since f1 ∈ Lp(Rn), IΩ,αf1 ∈ Lq(Rn) and from the boundedness of IΩ,α from
Lp(Rn) to Lq(Rn) it follows that:
kIΩ,αf1kLq(B) ≤ kIΩ,αf1kLq(Rn) ≤ Ckf1kLp(Rn) = Ckf kLp(2B),
where constant C > 0 is independent of f .
It's clear that x ∈ B, y ∈
∁
(2B) implies 1
2x0 − y ≤ x − y ≤ 3
2x0 − y. We
get
IΩ,αf2(x) ≤ 2n−αc1 Z ∁
(2B)
f (y)Ω(x − y)
x0 − yn−α
dy.
By Fubini's theorem we have
Z ∁
(2B)
f (y)Ω(x − y)
x0 − yn−α
f (y)Ω(x − y)Z ı
x0−y
dy ≈Z ∁
≈Z ı
.Z ı
(2B)
2rZ2r≤x0−y≤t
2rZB(x0,t)
dt
tn+1−α dy
dt
tn+1−α
dt
tn+1−α .
f (y)Ω(x − y)dy
f (y)Ω(x − y)dy
Applying Holder's inequality, we get
(2B)
Z ∁
.Z ı
.Z ı
2r
2r
f (y)Ω(x − y)
x0 − yn−α
dy
kf kLp(B(x0,t)) kΩ(· − y)kLs(B(x0,r)) B(x0, t)1− 1
p − 1
s
dt
tn+1−α
(3.9)
kf kLp(B(x0,t))
dt
q +1 .
n
t
Moreover, for all p ∈ [1, ı) the inequality
kIΩ,αf2kLq(B) . r
is valid. Thus
n
q Z ı
2r
kf kLp(B(x0,t))
dt
q +1 .
n
t
(3.10)
kIΩ,αf kLq(B) . kf kLp(2B) + r
7
n
q Z ı
2r
kf kLp(B(x0,t))
dt
q +1 .
n
t
On the other hand,
kf kLp(2B) ≈ r
≤ r
Thus
kf kLp(B(x0,t))
dt
q +1 .
n
t
(3.11)
dt
n
q +1
t
n
n
2r
q kf kLp(2B)Z ı
q Z ı
q Z ı
2r
n
2r
dt
q +1 .
n
t
kIΩ,αf kLq(B) . r
kf kLp(B(x0,t))
When 1 < q < s, by Fubini's theorem and the Minkowski inequality, we get
≤Z ı
2rZB(x0,t)
kIΩ,αf2kLq(B) ≤(cid:16)ZB(cid:12)(cid:12)(cid:12)Z ı
2rZB(x0,t)
s Z ı
2rZB(x0,t)
q Z ı
q Z ı
≤ r
. r
. r
q − n
2r
n
n
n
2r
kf kL1(B(x0,t))
kf kLp(B(x0,t))
f (y)Ω(x − y)dy
f (y) kΩ(· − y)kLq(B)dy
q
dt
tn+1−α(cid:12)(cid:12)(cid:12)
q(cid:17) 1
dt
tn+1−α
f (y) kΩ(· − y)kLs(B)dy
dt
tn+1−α
dt
tn+1−α
dt
q +1 .
n
t
(3.12)
Let p = 1 < q < s ≤ ı. From the weak (1, q) boundedness of IΩ,α and (3.11)
it follows that:
kIΩ,αf1kW Lq(B) ≤ kIΩ,αf1kW Lq(Rn) . kf1kL1(Rn)
= kf kL1(2B) . r
n
q Z ı
2r
kf kL1(B(x0,t))
dt
q +1 .
n
t
(3.13)
Then from (3.10) and (3.13) we get the inequality (3.7).
Theorem 3.5. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homoge-
neous of degree zero. Let 0 < α < n, 1 ≤ p < n
n , and s′ ≤ p or q < s.
Let also, the pair (ϕ1, ϕ2) satisfy the condition
q = 1
p − α
α , 1
Z ∞
r
ess inf
t<τ <∞
ϕ1(x0, τ )τ
n
p
n
q +1
t
dt ≤ C ϕ2(x0, r),
(3.14)
where C does not depend on r. Then the operators MΩ,α and IΩ,α are bounded
from LM {x0}
for p = 1.
p,ϕ1
Moreover, for p > 1
for p > 1 and from LM {x0}
1,ϕ1
to W LM {x0}
q,ϕ2
to LM {x0}
q,ϕ2
kMΩ,αf kLM
{x0}
q,ϕ2
. kIΩ,αf kLM
{x0}
q,ϕ2
. kf kLM
{x0}
p,ϕ1
8
and for p = 1
kMΩ,αf kW LM
{x0}
q,ϕ2
. kIΩ,αf kW LM
{x0}
q,ϕ2
. kf kLM
.
{x0}
1,ϕ1
Proof. By Lemma 3.4 and Theorem 3.1 with v2(r) = ϕ2(x0, r)−1, v1(r) = ϕ1(x0, r)−1r− n
and w(r) = r− n
q we have for p > 1
p
kIΩ,αf kLM
{x0}
q,ϕ2
ϕ2(x0, r)−1Z ı
r
kf kLp(B(x0,t))
dt
n
q +1
t
ϕ1(x0, r)−1 r− n
p kf kLp(B(x0,r)) = kf kLM
{x0}
p,ϕ1
. sup
r>0
. sup
r>0
and for p = 1
kIΩ,αf kW LM
{x0}
q,ϕ2
ϕ2(x0, r)−1Z ı
r
kf kL1(B(x0,t))
dt
n
q +1
t
ϕ1(x0, r)−1 r−n kf kLp(B(x0,r)) = kf kLM
.
{x0}
1,ϕ1
. sup
r>0
. sup
r>0
Corollary 3.6. Suppose that Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homogeneous of
degree zero. Let 0 < α < n, 1 ≤ p < n
n , and s′ ≤ p or q < s. Let also,
the pair (ϕ1, ϕ2) satisfy the condition
q = 1
p − α
α , 1
Z ∞
r
ess inf
t<τ <∞
ϕ1(x, τ )τ
n
p
n
q +1
t
dt ≤ C ϕ2(x, r),
where C does not depend on x and r. Then the operators MΩ,α and IΩ,α are
bounded from Mp,ϕ1 to Mq,ϕ2 for p > 1 and from M1,ϕ1 to W Mq,ϕ2 for p = 1.
Moreover, for p > 1
kMΩ,αf kMq,ϕ2 . kIΩ,αf kMq,ϕ2 . kf kMp,ϕ1 ,
and for p = 1
kMΩ,αf kW Mq,ϕ2 . kIΩ,αf kW Mq,ϕ2 . kf kM1,ϕ1 .
Corollary 3.7. Let 1 ≤ p < ∞, 0 < α < n
condition (3.14). Then the operators Mα and Iα are bounded from LM {x0}
p,ϕ1
LM {x0}
q,ϕ2
n and (ϕ1, ϕ2) satisfy
to
for p > 1 and from M {x0}
1,ϕ1 to W LM {x0}
for p = 1.
p , 1
q = 1
p − α
q,ϕ2
Remark 3.8. Note that, in the case s = ı Corollary 3.6 was proved in [29]. The
condition (3.14) in Theorem 3.5 is weaker than condition (3.6) in Theorem 3.3
(see [29]).
9
4 Commutators of fractional integral operator
with rough kernels in the spaces LM {x0}
p,ϕ
Let T be a linear operator, for a function b, we define the commutator [b, T ] by
[b, T ]f (x) = b(x) T f (x) − T (bf )(x)
erator, a well known result of Coifman, Rochberg and Weiss [14] states that the
for any suitable function f . If eT be a Calder´on-Zygmund singular integral op-
commutator [b, eT ]f = b eT f − eT (bf ) is bounded on Lp(Rn), 1 < p < ∞, if and
only if b ∈ BMO(Rn). The commutator of Calder´on-Zygmund operators plays
an important role in studying the regularity of solutions of elliptic partial differ-
ential equations of second order (see, for example, [12, 13, 22]). In [9], Chanillo
proved that the commutator [b, Iα]f = b Iαf − Iα(bf ) is bounded from Lp(Rn) to
Lq(Rn), (1 < p < q < ∞, 1
n ) if and only if b ∈ BMO(Rn).
p − α
q = 1
The definition of local Campanato space as follows.
Definition 4.1. Let 1 ≤ q < ı and 0 ≤ λ < 1
to belong to the CBMO{x0}
q,λ (Rn) (central Campanato space), if
n . A function f ∈ Lloc
q (Rn) is said
kf kCBM O{x0}
q,λ
= sup
r>0(cid:16)
f (y) − fB(x0,r)qdy(cid:17)1/q
< ∞,
1
B(x0, r)1+λq ZB(x0,r)
B(x0, r)ZB(x0,r)
1
fB(x0,r) =
f (y)dy.
where
Define
CBMO{x0}
q,λ (Rn) = {f ∈ Lloc
q (Rn) : kf kCBM O
< ı}.
{x0}
q,λ
q
q,0 (Rn). Note that, BMO(Rn) ⊂ CBMO{x0}
In [30], Lu and Yang introduced the central BMO space CBMOq(Rn) =
CBMO{0}
(Rn), 1 ≤ q < ı. The space
CBMO{x0}
(Rn) can be regarded as a local version of BMO(Rn), the space of
bounded mean oscillation, at the origin. But, they have quite different proper-
ties. The classical John-Nirenberg inequality shows that functions in BMO(Rn)
are locally exponentially integrable. This implies that, for any 1 ≤ q < ı, the
functions in BMO(Rn) can be described by means of the condition:
q
r>0(cid:16) 1
BZB
sup
f (y) − fBqdy(cid:17)1/q
< ∞,
where B denotes an arbitrary ball in Rn. However, the space CBMO{x0}
(Rn)
depends on q. If q1 < q2, then CBMO{x0}
(Rn). Therefore,
there is no analogy of the famous John-Nirenberg inequality of BMO(Rn) for the
space CBMO{x0}
(Rn)
may be quite different from that of BMO(Rn).
(Rn). One can imagine that the behavior of CBMO{x0}
(Rn) $ CBMO{x0}
q2
q1
q
q
q
10
Lemma 4.2. Let b be a function in CBMO{x0}
r1, r2 > 0. Then
q,λ (Rn), 1 ≤ q < ∞, 0 ≤ λ < 1
n and
(cid:18)
B(x0, r1)1+λq ZB(x0,r1)
1
b(y) − bB(x0,r2)qdy(cid:19) 1
q
where C > 0 is independent of b, r1 and r2.
≤ C(cid:18)1 +(cid:12)(cid:12)(cid:12) ln
r1
r2(cid:12)(cid:12)(cid:12)(cid:19) kbkCBM O
,
{x0}
q,λ
In [17] the following statement was proved for the commutators of fractional
p − α
integral operators with rough kernels, containing the result in [34, 36].
Theorem 4.3. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homo-
geneous of degree zero and b ∈ BMO(Rn). Let 0 < α < n, 1 ≤ s′ < p < n
p ,
q = 1
1
n , ϕ(x, r) which satisfies the conditions (3.4) and (3.5). Then the
operator [b, IΩ,α] is bounded from Mp,ϕ to Mq,ϕ.
Lemma 4.4. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homogeneous
p2,λ (Rn), 0 ≤ λ < 1
of degree zero. Let 0 < α < n, 1 < p < n
n ,
1
p = 1
α , b ∈ CBMO{x0}
p − α
q = 1
− α
n .
= 1
p1
n , 1
+ 1
p2
, 1
p1
Then, for s′ ≤ p or q1 < s the inequality
q1
k[b, IΩ,α]f kLq(B(x0,r)) . kbkCBM O
r
{x0}
p2,λ
n
q Z ı
2r(cid:16)1 + ln
t
r(cid:17)tnλ− n
q1
−1kf kLp1 (B(x0,t))dt
holds for any ball B(x0, r) and for all f ∈ Lloc
− α
+ 1
Proof. Let 1 < p < ∞, 0 < α < n
p2
in the proof of Lemma 3.4, we represent function f in form (3.8) and have
n , and 1
p1 (Rn).
p = 1
p − α
q = 1
= 1
p1
p , 1
, 1
p1
q1
n . As
[b, IΩ,α]f (x) =(cid:0)b(x) − bB(cid:1)IΩ,αf1(x) − IΩ,α(cid:16)(cid:0)b(·) − bB(cid:1)f1(cid:17)(x)
+(cid:0)b(x) − bB(cid:1)IΩ,αf2(x) − IΩ,α(cid:16)(cid:0)b(·) − bB(cid:1)f2(cid:17)(x)
≡ J1 + J2 + J3 + J4.
Hence we get
k[b, IΩ,α]f kLq(B) ≤ kJ1kLq(B) + kJ2kLq(B) + kJ3kLq(B) + kJ4kLq(B).
From the boundedness of [b, IΩ,α] from Lp1(Rn) to Lq1(Rn) it follows that:
kJ1kLq(B) ≤ k(cid:0)b(·) − bB(cid:1)[b, IΩ,α]f1(·)kLq(Rn)
r
n
p2
{x0}
p2,λ
≤ CkbkCBM O
+nλ kf1kLp1 (Rn)
≤ k(cid:0)b(·) − bB(cid:1)kLp2 (Rn)[b, IΩ,α]f1(·)kLq1 (Rn)
+nλ kf kLp1 (2B)Z ∞
2r (cid:16)1 + ln
q +nλ Z ∞
= CkbkCBM O
. kbkCBM O
{x0}
p2,λ
{x0}
p2,λ
+ n
q1
n
p2
2r
r
r
n
t−1− n
q1 dt
t
r(cid:17)kf kLp1 (B(x0,t))t−1− n
q1 dt.
11
For J2 we have
kJ2kLq(B) ≤ k[b, IΩ,α](cid:0)b(·) − bB(cid:1)f1kLq(Rn)
. k(b(·) − bB)f1kLp(Rn)
. kb(·) − bBkLp2 (Rn)kf1kLp1 (Rn)
. kbkCBM O
{x0}
p2,λ
. kbkCBM O
{x0}
p2,λ
r
r
n
p2
+ n
q1
+nλ kf kLp1 (2B)Z ∞
2r (cid:16)1 + ln
2r
t
n
p +nλ Z ∞
t−1− n
q1 dt
r(cid:17)kf kLp1 (B(x0,t))t−1− n
q1 dt.
For J3, it is known that x ∈ B, y ∈
∁
(2B), which implies 1
2 x0 − y ≤ x − y ≤
When s′ ≤ p, by Fubini's theorem and applying Holder inequality we have
3
2x0 − y.
IΩ,αf2(x) ≤ c0Z ∁
≈Z ∞
.Z ∞
.Z ∞
.Z ∞
2r
2r
Ω(x − y)
f (y)
x0 − yn−α dy
(2B)
2r Z2r<x0−y<t
2r ZB(x0,t)
Ω(x − y)f (y)dy t−1−n−αdt
Ω(x − y)f (y)dy t−1−n−αdt
kf kLp1 (B(x0,t)) kΩ(x − ·)kLs(B(x0,t)) B(x0, t)1− 1
p1
s t−1− n
− 1
p1
−αdt
kf kLp1 (B(x0,t)) t−1− n
q1 dt.
Hence, we get
kJ3kLq(B) = k(cid:0)b(·) − bB(cid:1)IΩ,αf2(·)kLq(Rn)
≤ k(cid:0)b(·) − bB(cid:1)kLq(Rn)Z ∞
kf kLp1 (B(x0,t)) t−1− n
q1 Z ∞
≤ k(cid:0)b(·) − bB(cid:1)kLp2 (Rn) r
q +nλ Z ∞
2r (cid:16)1 + ln
kf kLp1 (B(x0,t)) t−1− n
r(cid:17)kf kLp1 (B(x0,t))t−1− n
. kbkCBM O
q1 dt
q1 dt
{x0}
p2,λ
2r
2r
r
t
n
n
q1 dt.
12
f (y) kb(·) − bBkLp2 (B) kΩ(· − y)kLq1 (B)dy
− 1
s Z ı
2rZB(x0,t)
kf kL1(B(x0,t))
dt
tn−α+1
. kbkCBM O
{x0}
p2,λ
n
p2
r
+nλ B
1
q1
. kbkCBM O
{x0}
p2,λ
. kbkCBM O
{x0}
p2,λ
r
r
n
q +nλZ ı
q +nλZ ı
2r
n
2r(cid:16)1 + ln
t
r(cid:17)kf kLp1 (B(x0,t))
dt
+1 .
n
q1
t
f (y) kΩ(· − y)kLs(B)dy
dt
tn−α+1
(4.1)
For x ∈ B by Fubini's theorem and applying Holder inequality we have
b(y) − bB Ω(x − y)
f (y)
x − yn−α dy
(2B)
f (y)
x0 − yn−α dy
(2B)
b(y) − bB Ω(x − y)
IΩ,α(cid:16)(cid:0)b(·) − bB(cid:1)f2(cid:17)(x) .Z ∁
.Z ∁
≈Z ∞
.Z ı
2r Z2r<x0−y<t
2rZB(x0,t)
+Z ı
2r
b(y) − bB Ω(x − y) f (y)dy tα−n−1dt
b(y) − bB(x0,t)Ω(x − y) f (y)dy
bB(x0,r) − bB(x0,t)ZB(x0,t)
Ω(x − y) f (y)dy
dt
tn−α+1
dt
tn−α+1
k(b(·) − bB(x0,t))f kLp(B(x0,t)) kΩ(· − y)kLs(B(x0,t)) B(x0, t)1− 1
p − 1
s
dt
tn−α+1
bB(x0,r) − bB(x0,t)kf kLp1 (B(x0,t)) kΩ(· − y)kLs(B(x0,t)) B(x0, t)1− 1
p1
− 1
s tα−n−1dt
2r
.Z ı
+Z ı
.Z ı
2r
2r
When q1 < s, by Fubini's theorem and the Minkowski inequality, we get
2rZB(x0,t)
kJ3kLq(B) ≤(cid:16)ZB(cid:12)(cid:12)Z ı
2rZB(x0,t)
2rZB(x0,t)
≤Z ı
≤Z ı
f (y)b(x) − bBΩ(x − y)dy
f (y) k(b(·) − bB)Ω(· − y)kLq(B)dy
q
dt
tn−α+1(cid:12)(cid:12)q(cid:17) 1
dt
tn−α+1
dt
tn−α+1
kb(·) − bB(x0,t)kLp2 (B(x0,t))kf kLp1 (B(x0,t))t−1− n
q1 dt
+ kbkCBM O
. kbkCBM O
{x0}
p2 ,λ Z ı
p2,λ Z ı
2r(cid:16)1 + ln
2r(cid:16)1 + ln
{x0}
t
r(cid:17)kf kLp1 (B(x0,t)) tnλ−1− n
r(cid:17)kf kLp1 (B(x0,t)) tnλ−1− n
t
q1 dt
q1 dt.
13
Then for J4 we have
kJ4kLq(B) ≤ kIΩ,α(cid:0)b(·) − bB(cid:1)f2kLq(Rn)
q Z ∞
2r (cid:16)1 + ln
. kbkCBM O{x0}
p2 ,λ
r
n
t
r(cid:17)kf kLp1 (B(x0,t))tnλ−1− n
q1 dt.
When q1 < s, by Fubini's theorem and the Minkowski inequality, we get
2rZB(x0,t)
kIΩ,αf2kLq(B) ≤(cid:16)ZB(cid:12)(cid:12)Z ı
2rZB(x0,t)
s Z ı
2rZB(x0,t)
≤Z ı
≤ B
q − 1
1
f (y)Ω(x − y)dy
f (y) kΩ(· − y)kLq(B)dy
q
dt
tn−a+1(cid:12)(cid:12)q(cid:17) 1
dt
tn−a+1
f (y) kΩ(· − y)kLs(B)dy
dt
tn−a+1
n
. r
. r
q Z ı
q Z ı
2r
2r
n
kf kL1(B(x0,t))
dt
tn−a+1
kf kLp1 (B(x0,t))
dt
+1 .
n
q1
t
(4.2)
Now combined by all the above estimates, we end the proof of this Lemma
4.4.
The following theorem is true.
Theorem 4.5. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1) with 1 < s ≤ ∞, be a
homogeneous of degree zero. Let 0 < α < n, 1 < p < n
p2,λ (Rn),
0 ≤ λ < 1
n , 1
n . Let also, for s′ ≤ p or q1 < s
the pair (ϕ1, ϕ2) satisfy the condition
α , b ∈ CBMO{x0}
p = 1
q = 1
p − α
= 1
p1
+ 1
p2
n , 1
− α
, 1
p1
q1
Z ∞
r (cid:16)1 + ln
t
r(cid:17) ess inf
t<τ <∞
t
ϕ1(x0, τ )τ
n
p
n
q −nλ+1
dt ≤ C ϕ2(x0, r),
(4.3)
where C does not depend on r. Then, the operators MΩ,b,α and [b, IΩ,α] are
bounded from LM {x0}
p,ϕ1 to LM {x0}
q,ϕ2 . Moreover
kMΩ,b,αf kLM
{x0}
q,ϕ2
. k[b, IΩ,α]f kLM
{x0}
q,ϕ2
. kbkCBM O
kf kLM
.
{x0}
p,ϕ1
{x0}
p2 ,λ
Proof. The statement of Theorem 4.5 follows by Lemma 4.4 and Theorem 3.1 in
the same manner as in the proof of Theorem 3.5.
For the sublinear commutator of the fractional maximal operator Mb,α and
for the linear commutator of the Riesz potential [b, Iα] from Theorem 4.5 we get
the following new results.
14
Corollary 4.6. Let 0 < α < n, 1 < p < n
p = 1
1
Then, the operators Mb,α and [b, Iα] are bounded from LM {x0}
p2,λ (Rn), 0 ≤ λ < 1
n ,
n , and (ϕ1, ϕ2) satisfies the condition (4.3).
p1,ϕ1 to LM {x0}
q,ϕ2 .
α , b ∈ CBMO{x0}
q = 1
= 1
p1
+ 1
p2
p − α
n , 1
q1
− α
, 1
p1
5 Some applications
In this section, we shall apply Theorems 3.5 and 4.5 to several particular operators
such as the Marcinkiewicz operator and fractional powers of the some analytic
semigroups.
5.1 Marcinkiewicz operator
Let Sn−1 = {x ∈ Rn : x = 1} be the unit sphere in Rn equipped with the
Lebesgue measure dσ. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a
homogeneous of degree zero and satisfy the cancellation condition.
In 1958, Stein [41] defined the Marcinkiewicz integral of higher dimension µΩ
as
where
µΩ(f )(x) =(cid:18)Z ∞
FΩ,t(f )(x) =Zx−y≤t
0
FΩ,t(f )(x)2 dt
t3(cid:19)1/2
,
Ω(x − y)
x − yn−1 f (y)dy.
Since Stein's work in 1958, the continuity of Marcinkiewicz integral has been
extensively studied as a research topic and also provides useful tools in harmonic
analysis [33, 40, 42, 43].
The Marcinkiewicz operator is defined by (see [44])
where
µΩ,α(f )(x) =(cid:18)Z ∞
FΩ,α,t(f )(x) =Zx−y≤t
0
FΩ,α,t(f )(x)2 dt
t3(cid:19)1/2
,
Ω(x − y)
x − yn−1−α f (y)dy.
Note that µΩf = µΩ,0f .
Let H be the space H = {h : khk = (R ∞
clear that µΩ,α(f )(x) = kFΩ,α,t(x)k.
0
h(t)2dt/t3)1/2 < ı}. Then, it is
By Minkowski inequality and the conditions on Ω, we get
µΩ,α(f )(x) ≤ZRn
Ω(x − y)
x − yn−1−α f (y)(cid:18)Z ∞
x−y
dt
t3(cid:19)1/2
dy ≤ CIΩ,α(f )(x).
15
It is known that µΩ,α is bounded from Lp(Rn) to Lq(Rn) for p > 1, and bounded
from L1(Rn) to W Lq(Rn) for p = 1 (see [44]), then from Theorems 3.5 and 4.5
we get
Corollary 5.1. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homo-
geneous of degree zero and satisfy the cancellation condition. Let 0 < α < n,
1 ≤ p < n
n and for s′ ≤ p or q1 < s the pair (ϕ1, ϕ2) satisfy the
condition (3.14). Then µΩ,α is bounded from LM {x0}
for p > 1 and
p,ϕ1
from M {x0}
to LM {x0}
q,ϕ2
1,ϕ1 to W LM {x0}
for p = 1.
q = 1
p − α
α , 1
q,ϕ2
Corollary 5.2. Suppose that x0 ∈ Rn, Ω ∈ Ls(Sn−1), 1 < s ≤ ∞, be a homo-
geneous of degree zero and satisfy the cancellation condition. Let 0 < α < n,
1 < p < n
− α
n
and for s′ ≤ p or q1 < s the pair (ϕ1, ϕ2) satisfy the condition (3.14). Then
[a, µΩ,α] is bounded from LM {x0}
α , b ∈ CBMO{x0}
p2,λ (Rn), 0 ≤ λ < 1
p,ϕ1 to LM {x0}
q,ϕ2 .
p = 1
q = 1
p − α
= 1
p1
+ 1
p2
n , 1
q1
n , 1
p1
, 1
5.2 Fractional powers of the some analytic semigroups
The theorems of the previous sections can be applied to various operators which
are estimated from above by Riesz potentials. We give some examples.
Suppose that L is a linear operator on L2 which generates an analytic semi-
group e−tL with the kernel pt(x, y) satisfying a Gaussian upper bound, that is,
pt(x, y) ≤
c1
tn/2 e−c2
x−y2
t
(5.1)
for x, y ∈ Rn and all t > 0, where c1, c2 > 0 are independent of x, y and t.
For 0 < α < n, the fractional powers L−α/2 of the operator L are defined by
L−α/2f (x) =
1
Γ(α/2)Z ∞
0
e−tLf (x)
dt
t−α/2+1 .
Note that if L = −△ is the Laplacian on Rn, then L−α/2 is the Riesz potential
Iα. See, for example, Chapter 5 in [40].
Theorem 5.3. Let condition (5.1) be satisfied. Moreover, let 1 ≤ p < ∞,
p , 1
0 < α < n
n , (ϕ1, ϕ2) satisfy condition (3.14). Then L−α/2 is bounded
from LM {x0}
p,ϕ1 to LM {x0}
q,ϕ2
for p > 1 and from M {x0}
1,ϕ1 to W LM {x0}
for p = 1.
q = 1
p − α
q,ϕ2
Proof. Since the semigroup e−tL has the kernel pt(x, y) which satisfies condition
(5.1), it follows that
L−α/2f (x) . Iα(f )(x)
16
(see [20]). Hence by the aforementioned theorems we have
kL−α/2f kM
{x0}
q,ϕ2
. kIα(f )kM
{x0}
q,ϕ2
. kf kM
.
{x0}
p,ϕ1
Let b be a locally integrable function on Rn, the commutator of b and L−α/2
is defined as follows
[b, L−α/2]f (x) = b(x)L−α/2f (x) − L−α/2(bf )(x).
In [20] extended the result of [9] from (−∆) to the more general operator L
defined above. More precisely, they showed that when b ∈ BMO(Rn), then the
commutator operator [b, L−α/2] is bounded from Lp(Rn) to Lq(Rn) for 1 < p <
q < ∞ and 1
n . Then from Theorem 4.5 we get
p − α
q = 1
α , b ∈ CBMO{x0}
Theorem 5.4. Let condition (5.1) be satisfied. Moreover, let 0 < α < n, 1 <
− α
p < n
n ,
and (ϕ1, ϕ2) satisfies the condition (4.3). Then [b, L−α/2] is bounded from LM {x0}
p,ϕ1
to LM {x0}
q,ϕ2 .
p2,λ (Rn), 0 ≤ λ < 1
n , and 1
p = 1
= 1
p1
n , 1
q1
+ 1
p2
, 1
q = 1
p − α
p1
Property (5.1) is satisfied for large classes of differential operators (see, for
example [6]). In [6] also other examples of operators which are estimates from
above by Riesz potentials are given. In these cases Theorem 3.5 and 4.5 are also
applicable for proving boundedness of those operators and commutators from
LM {x0}
p,ϕ1 to LM {x0}
q,ϕ2 .
References
[1] D.R. Adams, A note on Riesz potentials, Duke Math. 42 (1975), 765-778.
[2] Ali Akbulut, V.S. Guliyev and R. Mustafayev, Boundedness of the max-
imal operator and singular integral operator in generalized Morrey spaces,
Mathematica Bohemica, 137 (1) 2012, 27-43.
[3] J. Alvarez, M. Guzman-Partida, J. Lakey, Spaces of bounded λ-central mean
oscillation, Morrey spaces, and λ-central Carleson measures, Collect. Math.,
51 (2000), 1-47.
[4] A. Beurling, Construction and analysis of some convolution algebras, Ann.
Inst. Fourier (Grenoble), 14 (1964), 132.
[5] V.I. Burenkov, H.V. Guliyev, V.S. Guliyev, Necessary and sufficient condi-
tions for boundedness of the fractional maximal operators in the local Morrey-
type spaces, J. Comput. Appl. Math. 208 (1) (2007), 280-301.
17
[6] V.I. Burenkov, V.S. Guliyev, Necessary and sufficient conditions for the
boundedness of the Riesz potential in local Morrey-type spaces, Potential
Anal. 30 (3) (2009), 211-249.
[7] V. Burenkov, A. Gogatishvili, V.S. Guliyev, R. Mustafayev, Boundedness of
the fractional maximal operator in local Morrey-type spaces, Complex Var.
Elliptic Equ. 55 (8-10) (2010), 739-758.
[8] V. Burenkov, A. Gogatishvili, V.S. Guliyev, R. Mustafayev, Boundedness of
the Riesz potential in local Morrey-type spaces, Potential Anal. 35 (1) (2011),
67-87.
[9] S. Chanillo, A note on commutators, Indiana Univ. Math. J. 23 (1982),
7-16.
[10] M. Carro, L. Pick, J. Soria, V.D. Stepanov, On embeddings between classical
Lorentz spaces, Math. Inequal. Appl. 4 (3) (2001), 397-428.
[11] F. Chiarenza, M. Frasca, Morrey spaces and Hardy-Littlewood maximal
function, Rend Mat. 7 (1987), 273-279.
[12] F. Chiarenza, M. Frasca, P. Longo,
Interior W 2,p-estimates for nondiver-
gence elliptic equations with discontinuous coefficients, Ricerche Mat. 40
(1991), 149-168.
[13] F. Chiarenza, M. Frasca, P. Longo, W 2,p-solvability of Dirichlet problem for
nondivergence elliptic equations with VMO coefficients, Trans. Amer. Math.
Soc. 336 (1993), 841-853.
[14] R. Coifman, R. Rochberg, G. Weiss, Factorization theorems for Hardy spaces
in several variables, Ann. of Math. 103 (2) (1976), 611-635.
[15] Y. Ding, Weak type bounds for a class of rough operators with power weights,
Proc. Amer. Math. Soc. 125 (1997), 2939-2942.
[16] Y. Ding and S. Z. Lu, Weighted norm inequalities for fractional integral
operators with rough kernel, Canad. J. Math. 50 (1998), 29-39.
[17] Y. Ding, D. Yang, Z. Zhou, Boundedness of sublinear operators and commu-
tators on Lp,ω(Rn), Yokohama Math. J. 46 (1998), 15-27.
[18] Y. Ding and S. Z. Lu, Higher order commutators for a class of rough opera-
tors, Ark. Mat. 37 (1999), 33-44.
[19] J. Duoandikoetxea, Fourier Analysis, American Mathematical Society, Prov-
idence, Rhode Island, 2000.
18
[20] X.T. Duong, L.X. Yan, On commutators of fractional integrals, Proc. Amer.
Math. Soc. 132 (12) (2004), 3549-3557.
[21] G. Di Fazio, M.A. Ragusa, Commutators and Morrey spaces, Boll. Un. Mat.
Ital. 5A (7) (1991), 323-332.
[22] G. Di Fazio, M.A. Ragusa,
Interior estimates in Morrey spaces for strong
solutions to nondivergence form equations with discontinuous coefficients, J.
Funct. Anal. 112 (1993), 241-256.
[23] G. Di Fazio, D. K. Palagachev and M. A. Ragusa, Global Morrey regular-
ity of strong solutions to the Dirichlet problem for elliptic equations with
discontinuous coefficients, J. Funct. Anal, 166 (1999), 179-196.
[24] H. Feichtinger, An elementary approach to Wieners third Tauberian theorem
on Euclidean n-space, Proceedings, Conference at Cortona 1984, Sympos.
Math., 29, Academic Press 1987.
[25] J. Garcia-Cuerva and J.L. Rubio de Francia, Weighted Norm Inequalities
and Related Topics, North-Holland Math. 16, Amsterdam, 1985.
[26] V.S. Guliyev,
Integral operators on function spaces on the homogeneous
groups and on domains in Rn. Doctor's degree dissertation, Mat. Inst.
Steklov, Moscow, 1994, 329 pp. (in Russian)
[27] V.S. Guliyev, Function spaces, Integral Operators and Two Weighted In-
equalities on Homogeneous Groups. Some Applications, Cashioglu, Baku,
1999, 332 pp. (in Russian)
[28] V.S. Guliyev, Boundedness of the maximal, potential and singular operators
in the generalized Morrey spaces, J. Inequal. Appl. 2009, Art. ID 503948, 20
pp.
[29] V.S. Guliyev, S.S. Aliyev, T. Karaman, P. S. Shukurov, Boundedness of
sublinear operators and commutators on generalized Morrey Space, Int. Eq.
Op. Theory. 71 (3) (2011), pp. 327-355.
[30] S.Z. Lu and D.C. Yang, The central BMO spaces and Littlewood-Paley op-
erators, Approx. Theory Appl. (N.S.), 11 (1995), 72-94.
[31] G. Lu, S. Lu, D. Yang, Singular integrals and commutators on homogeneous
groups, Analysis Mathematica, 28 (2002), 103-134.
[32] S.Z. Lu and Q. Wu, CBMO estimates for commutators and multilinear
singular integrals, Math. Nachr., 276 (2004), 75-88.
19
[33] S. Lu, Y. Ding, D. Yan, Singular integrals and related topics, World Scientific
Publishing, Singapore, 2006.
[34] T. Mizuhara, Boundedness of some classical operators on generalized Morrey
spaces, Harmonic Analysis (S. Igari, Editor), ICM 90 Satellite Proceedings,
Springer - Verlag, Tokyo (1991), 183-189.
[35] C.B. Morrey, On the solutions of quasi-linear elliptic partial differential
equations, Trans. Amer. Math. Soc. 43 (1938), 126-166.
[36] E. Nakai, Hardy -- Littlewood maximal operator, singular integral operators
and Riesz potentials on generalized Morrey spaces, Math. Nachr. 166 (1994),
95-103.
[37] E. Nakai, A characterization of pointwise multipliers on the Morrey spaces,
Scientiae Mathematicae 3 (2000), 445-454.
[38] B. Muckenhoupt and R. L. Wheeden, Weighted norm inequalities for singular
and fractional integrals, Trans. Amer. Math. Soc, 161 (1971), 249-258.
[39] J. Peetre, On the theory of Mp,λ, J. Funct. Anal. 4 (1969), 71-87.
[40] Stein, E.M.: Singular integrals and differentiability of functions. Princeton
University Press, Princeton, NJ, 1970.
[41] E.M. Stein, On the functions of Littlewood-Paley, Lusin, and Marcinkiewicz,
Trans. Amer. Math. Soc. 88 (1958), 430-466.
[42] E.M. Stein, Harmonic Analysis: Real Variable Methods, Orthogonality and
Oscillatory Integrals, Princeton Univ. Press, Princeton NJ, 1993.
[43] A. Torchinsky, Real Variable Methods in Harmonic Analysis, Pure and
Applied Math. 123, Academic Press, New York, 1986.
[44] A. Torchinsky and S. Wang, A note on the Marcinkiewicz integral, Colloq.
Math. 60/61 (1990), 235-243.
[45] N. Wiener, Generalized Harmonic Analysis, Acta Math., 55 (1930), 117-258.
[46] N. Wiener, Tauberian theorems, Ann. Math., 33 (1932), 1-100.
20
|
1204.2290 | 1 | 1204 | 2012-04-10T21:45:16 | Greedy Algorithms for Reduced Bases in Banach Spaces | [
"math.FA"
] | Given a Banach space X and one of its compact sets F, we consider the problem of finding a good n dimensional space X_n \subset X which can be used to approximate the elements of F. The best possible error we can achieve for such an approximation is given by the Kolmogorov width d_n(F)_X. However, finding the space which gives this performance is typically numerically intractable. Recently, a new greedy strategy for obtaining good spaces was given in the context of the reduced basis method for solving a parametric family of PDEs. The performance of this greedy algorithm was initially analyzed in A. Buffa, Y. Maday, A.T. Patera, C. Prud'homme, and G. Turinici, "A Priori convergence of the greedy algorithm for the parameterized reduced basis", M2AN Math. Model. Numer. Anal., 46(2012), 595-603 in the case X = H is a Hilbert space. The results there were significantly improved on in P. Binev, A. Cohen, W. Dahmen, R. DeVore, G. Petrova, and P. Wojtaszczyk, "Convergence rates for greedy algorithms in reduced bases Methods", SIAM J. Math. Anal., 43 (2011), 1457-1472. The purpose of the present paper is to give a new analysis of the performance of such greedy algorithms. Our analysis not only gives improved results for the Hilbert space case but can also be applied to the same greedy procedure in general Banach spaces. | math.FA | math |
Greedy Algorithms for Reduced Bases in Banach Spaces∗
Ronald DeVore, Guergana Petrova, and Przemyslaw Wojtaszczyk
June 1, 2018
Abstract
Given a Banach space X and one of its compact sets F, we consider the problem
of finding a good n dimensional space Xn ⊂ X which can be used to approximate the
elements of F. The best possible error we can achieve for such an approximation is
given by the Kolmogorov width dn(F)X . However, finding the space which gives this
performance is typically numerically intractable. Recently, a new greedy strategy for
obtaining good spaces was given in the context of the reduced basis method for solving
a parametric family of PDEs. The performance of this greedy algorithm was initially
analyzed in [2] in the case X = H is a Hilbert space. The results of [2] were significantly
improved on in [1]. The purpose of the present paper is to give a new analysis of the
performance of such greedy algorithms. Our analysis not only gives improved results
for the Hilbert space case but can also be applied to the same greedy procedure in
general Banach spaces.
Key words and phrases: greedy algorithms, convergence rates, reduced basis, general
Banach space
AMS Subject Classification: 41A46, 41A25, 46B20, 15A15
1
Introduction
Let X be a Banach space with norm k · k := k · kX, and let F be one of its compact subsets.
For notational convenience only, we shall assume that the elements f of F satisfy kfkX ≤ 1.
We consider the following greedy algorithm for generating approximation spaces for F . We
first choose a function f0 such that
kf0k = max
f ∈F kfk.
(1.1)
∗This research was supported by the Office of Naval Research Contracts ONR-N00014-08-1-1113, ONR
N00014-09-1-0107, and ONR N00014-11-1-0712; the AFOSR Contract FA95500910500; the NSF Grants
DMS-0810869, and DMS 0915231; and the EU Project POWIEW. This publication is based in part on work
supported by Award No. KUS-C1-016-04 made by King Abdullah University of Science and Technology
(KAUST)
1
Assuming {f0, . . . , fn−1} and Vn := span{f0, . . . , fn−1} have been selected, we then take
fn ∈ F such that
(1.2)
dist(f, Vn)X ,
dist(fn, Vn)Xk = max
f ∈F
and define
σn := σn(F )X := dist(fn, Vn)X := sup
f ∈F
g∈Vn kf − gk.
inf
(1.3)
This greedy algorithm was introduced, for the case X is a Hilbert space, in the reduced basis
method [5, 6] for solving a family of PDEs. Certain variants of this algorithm, known as weak
greedy algorithms, described below, are now numerically implemented with great success in
the reduced basis method. Our interest in this paper will be in the approximation properties
of this algorithm and its weak variant.
We are interested in how well the space Vn approximates the elements of F and for this
purpose we compare its performance with the best possible performance which is given by
the Kolmogorov width dn(F )X of F defined by
dn := dn(F )X := inf
dist(f, Y )X,
(1.4)
sup
f ∈F
Y
where the infimum is taken over all n dimensional subspaces Y of X. We refer the reader to
[4] for a general discussion of Kolmogorov widths. We also define
f ∈F kfk = σ0(F )X,
which corresponds to approximating by zero dimensional spaces.
d0 := d0(F )X := max
Of course, if (σn)n≥0 decays at a rate comparable to (dn)n≥0, this would mean that the
greedy selection provides essentially the best possible accuracy attainable by n-dimensional
subspaces. Various comparisons have been given between σn and dn. A first result in this
direction, in the case that X is a Hilbert space H, was given in [2] where it was proved that
(1.5)
σn(F )H ≤ Cn2ndn(F )H,
with C an absolute constant. While this is an interesting comparison, it is only useful if
dn(F )H decays to zero faster than n−12−n. Various improvements on (1.5) were given in
It was shown that if
[1], again in the Hilbert space setting. We mention two of these.
dn(F )H ≤ Cn−α, n = 1, 2, . . . , then
σn(F )H ≤ C ′
αn−α.
(1.6)
This shows that in the scale of polynomial decay the greedy algorithm performs with the
same rates as n-widths. A related result was proved for sub-exponential decay. If for some
0 < α ≤ 1, we have dn(F )H ≤ Ce−cnα, n = 1, 2, . . . , then
σn(F )H ≤ C ′
αnβ
αe−c′
,
β =
, n = 1, 2, . . . .
(1.7)
α
α + 1
In numerical implementations, the greedy algorithm is too demanding since at each it-
eration it requires finding an element from F which is at furthest distance from Vn. To cir-
cumvent this difficulty, one modifies the algorithm as follows. We fix a constant 0 < γ ≤ 1.
At the first step of the algorithm, one chooses a function f0 ∈ F such that
kf0k ≥ γσ0(F )X.
2
At the general step, if f0, . . . , fn−1 have been chosen, Vn := span{f0, . . . , fn−1}, and
σn(f )X := dist(f, Vn)X ,
is the best approximation error to f from Vn we now choose fn ∈ F such that
σn(fn)X ≥ γ max
f ∈F
σn(f )X,
(1.8)
to be the next element in the greedy selection. Note that if γ = 1, then the weak greedy
algorithm reduces to the greedy algorithm that we have introduced above.
Notice that similar to the greedy algorithm, (σn(F )X)n≥0 is also monotone decreasing.
Of course, neither the greedy algorithm or the weak greedy algorithm give a unique sequence
(fn)n≥0, nor is the sequence (σn(F )X)n≥0 unique. In all that follows, the notation reflects
any sequences which can arise in the implementation of the weak greedy selection for the
fixed value of γ.
In the present paper, we shall first prove a lemma that we use in our new analysis of the
weak greedy algorithm. This new analysis gives a significant improvement of the previous
results. We mention two of these:
The first, given in Corollary 3.3, is that
σ2n(F )H ≤
√2γ−1pdn(F )H, n = 1, 2, . . . .
(1.9)
This is the first direct comparison between (σn(F )H)n≥0 and (dn(F )H)n≥0 for the special case
X is a Hilbert space H, which guarantees a specific rate of decay for (σn(F )H)n≥0 without
any assumption of a decay rate for (dn(F )H)n≥0. Notice that, in particular, this allows one
to improve the sub-exponential results mentioned earlier (see Corollary 3.3).
The second part of our paper analyzes the performance of the greedy algorithm in a
general Banach space. We prove estimates for the decay of (σn(F )X)n≥0 similar to those in
the Hilbert space case, except that there is a loss of the order O(√n). We give examples
which show that this loss in essence cannot be removed. However, our results for a general
Banach space are still not definitive. For example, we have no result of the form (1.9) because
of the √n factor that appears in our Banach space results.
2 Main lemma
In this section, we shall prove a lemma for matrices that we employ in our analysis of weak
greedy algorithms in both Hilbert and Banach spaces.
Lemma 2.1 Let G = (gi,j) be a K × K lower triangular matrix with rows g1, . . . , gK , W
be any m dimensional subspace of RK, and P be the orthogonal projection of RK onto W .
Then
i,i ≤( 1
g2
KYi=1
ℓ2)m( 1
kP gik2
KXi=1
K − m
where k · kℓ2 is the euclidean norm of a vector in RK .
m
kgi − P gik2
ℓ2)K−m
KXi=1
,
(2.1)
3
Proof: We choose an orthonormal basis ϕ1, . . . , ϕm for the space W and complete it into
an orthonormal basis ϕ1, . . . , ϕK for RK. If we denote by Φ the K × K orthogonal matrix
whose j-th column is ϕj, then the matrix C := GΦ has entries ci,j = hgi, ϕji. We denote by
cj, the j-th column of C. It follows from the arithmetic geometric mean inequality for the
numbers {kcjk2
j=1 that
ℓ2}m
ℓ2 ≤( 1
m
kcjk2
mYj=1
mXj=1
ℓ2)m
kcjk2
=( 1
m
mXj=1
KXi=1
hgi, ϕji2)m
=( 1
m
ℓ2)m
kP gik2
KXi=1
.
(2.2)
Similarly,
KYj=m+1
ℓ2 ≤ ( 1
K − m
kcjk2
ℓ2)K−m
kcjk2
=( 1
K − m
KXi=1
kgi − P gik2
ℓ2)K−m
.(2.3)
Now, Hadamard's inequality for the matrix C and relations (2.2) and (2.3) result in
KXj=m+1
ℓ2 ≤( 1
m
(det C)2 ≤
KYj=1
kcjk2
KXi=1
kP gik2
kgi − P gik2
ℓ2)K−m
KXi=1
. (2.4)
The latter inequality and the fact that det G =
gi,i and det C = det G gives (2.1). (cid:3)
K − m
ℓ2)m( 1
KYi=1
3 A new analysis for the weak greedy algorithm in a
Hilbert space
The purpose of this section is to obtain new results for the performance of the weak greedy
algorithm in a Hilbert space that considerably improve on the analysis in [2] and [1]. This
will be accomplished by making a finer comparison between (σn(F )H)n≥0 and (dn(F )H)n≥0
than those given in [1]. We assume throughout this section that X = H is a Hilbert space
and follow the notation from [1].
Note that in general, the weak greedy algorithm does not terminate and we obtain an
infinite sequence f0, f1, f2, . . . . In order to have a consistent notation in what follows, we shall
define fm := 0, m > N, if the algorithm terminates at N, i.e. if σN (F )H = 0. By (f ∗
n)n≥0 we
denote the orthonormal system obtained from (fn)n≥0 by Gram-Schmidt orthogonalization.
It follows that the orthogonal projector Pn from H onto Vn is given by
and, in particular,
Pnf =
n−1Xi=0
hf, f ∗
i if ∗
i ,
fn = Pn+1fn =
nXj=0
an,jf ∗
j ,
an,j = hfn, f ∗
j i, j ≤ n.
4
There is no loss of generality in assuming that the infinite dimensional Hilbert space H is
ℓ2(N ∪ {0}) and that f ∗
j = ej, where ej is the vector with a one in the coordinate indexed
by j and is zero in all other coordinates, i.e. (ej)i = δj,i. We adhere to this assumption
throughout this section of the paper.
We consider the lower triangular matrix
A := (ai,j)∞
i,j=0,
ai,j := 0, j > i.
This matrix incorporates all the information about the weak greedy algorithm on F . The
following two properties characterize any lower triangular matrix A generated by such a
greedy algorithm. With the notation σn := σn(F )H, we have:
P1: The diagonal elements of A satisfy γσn ≤ an,n ≤ σn.
P2: For every m ≥ n, one hasPm
Indeed, P1 follows from
j=n a2
m,j ≤ σ2
n.
a2
n,n = kfnk2 − kPnfnk2 = kfn − Pnfnk2,
combined with the weak greedy selection property (1.8). To see P2, we note that for m ≥ n,
mXj=n
a2
m,j = kfm − Pnfmk2 ≤ max
f ∈F kf − Pnfk2 = σ2
n.
Remark 3.1 If A is any matrix satisfying P1 and P2 with (σn)n≥0 a decreasing sequence
that converges to 0, then the rows of A form a compact subset of ℓ2(N ∪{0}). If F is the set
consisting of these rows, then one of the possible realizations of the weak greedy algorithm
with constant γ will choose the rows in that order and A will be the resulting matrix.
The matrix representation A of the weak greedy algorithm was the basis of the analysis
given in [1] and will also be critical in the proof of the next theorem.
Theorem 3.2 For the weak greedy algorithm with constant γ in a Hilbert space H and for
any compact set F , we have the following inequalities between σn := σn(F )H and dn :=
dn(F )H, for any N ≥ 0, K ≥ 1, and 1 ≤ m < K,
σ2
N +i ≤ γ−2K(cid:26) K
m(cid:27)m(cid:26) K
K − m(cid:27)K−m
KYi=1
σ2m
N +1d2K−2m
m
(3.1)
Proof: We consider the K × K matrix G = (gi,j) which is formed by the rows and
columns of A with indices from {N + 1, . . . , N + K}. Each row gi is the restriction of fN +i
to the coordinates N + 1, . . . , N + K. Let Hm be the m-dimensional Kolmogorov subspace
linear space which is the restriction of Hm to the coordinates N + 1, . . . , N + K. In general,
of H for which dist(F ,Hm) = dm. Then, dist(fN +i,Hm) ≤ dm, i = 1, . . . K. Let fW be the
dim(fW ) ≤ m. Let W be an m dimensional space, W ⊂ span{eN +1, . . . , eN +K}, such that
fW ⊂ W and P and eP are the projections in RK onto W andfW , respectively. Clearly,
kP gikℓ2 ≤ kgikℓ2 ≤ σN +1,
i = 1, . . . , K,
(3.2)
5
where we have used Property P2 in the last inequality. Note that
kgi − P gikℓ2 ≤ kgi − eP gikℓ2 = dist(gi,fW ) ≤ dist(fN +i,Hm) ≤ dm,
It follows from Property P1 that
i = 1, . . . , K.
(3.3)
KYi=1
aN +i,N +i ≥ γK
σN +i.
KYi=1
(3.4)
We now apply Lemma 2.1 for this G and W , and use estimates (3.2), (3.3), and (3.4) to
derive (3.1). The proof is completed.
(cid:3)
We next record some special cases of Theorem 3.2.
Corollary 3.3 For the weak greedy algorithm with constant γ in a Hilbert space H, we have
the following:
(i) For any compact set F and n ≥ 1, we have
√2γ−1 min
In particular σ2n(F ) ≤ √2γ−1pdn(F ), n = 1, 2 . . . .
(ii) If dn(F ) ≤ C0n−α, n = 1, 2, . . . , then σn(F ) ≤ C1n−α, n = 1, 2 . . . , with C1 :=
(iii) If dn(F ) ≤ C0e−c0nα
, n = 1, 2 . . . , where
, n = 1, 2, . . . , then σn(F ) ≤ √2C0γ−1e−c1nα
σn(F ) ≤
25α+1γ−2C0.
(3.5)
1≤m<n
d
n
m (F ).
n−m
c1 = 2−1−2αc0,
Proof: (i) We take N = 0, K = n and any 1 ≤ m < n in Theorem 3.2, use the
monotonicity of (σn)n≥0 and the fact that σ0 ≤ 1 to obtain
σ2n
n ≤
nYj=1
σ2
j ≤ γ−2nn n
mom(cid:26) n
n − m(cid:27)n−m
d2n−2m
m
.
(3.6)
(ii) It follows from the monotonicity of (σn)n≥0 and (3.1) for N = K = n and any
Since x−x(1 − x)x−1 ≤ 2 for 0 < x < 1, we derive (3.5).
1 ≤ m < n that
σ2n
2n ≤
2nYj=n+1
σ2
j ≤ γ−2nn n
In the case n = 2s and m = s we have
σ4s ≤
n − m(cid:27)n−m
mom(cid:26) n
√2γ−1pσ2sds.
σ2m
n d2n−2m
m
.
(3.7)
Now we prove our claim by contradiction. Suppose it is not true and M is the first value
where σM (F ) > C1M −α. Let us first assume M = 4s. From (3.7), we have
σ4s ≤
√2γ−1pC1(2s)−αpC0s−α =p21−αC0C1γ−1s−α,
6
(3.8)
where we have used the fact that σ2s ≤ C1(2s)−α and ds ≤ C0s−α. It follows that
and therefore
C1(4s)−α < σ4s ≤p21−αC0C1γ−1s−α,
C1 < 23α+1γ−2C0 < 25α+1γ−2C0,
which is the desired contradiction. If M = 4s + q, q ∈ {1, 2, 3}, then it follows from (3.8)
and the monotonicity of (σn)n≥0 that
C12−3αs−α = C12−α(4s)−α < C1(4s + q)−α < σ4s+q ≤ σ4s ≤p21−αC0C1γ−1s−α.
From this, we obtain
C1 < 25α+1γ−2C0,
which is the desired contradiction in this case. This completes the proof of (ii).
(iii) From (i), we have
σ2n+1 ≤ σ2n ≤
from which (iii) easily follows.
√2γ−1pdn ≤p2C0γ−1e− c0
2 nα
=p2C0γ−1e−c02−1−α(2n)α
,
(3.9)
(cid:3)
Remark 3.4 Note that one can obtain a better constant c1 in (iii) if the minimum in (3.5)
is computed. Namely, this gives
√2γ−1C0 min
1≤m<n
σ2n ≤
e−c0mα (n−m)
n = √2γ−1C0e−c0nα{ max
1≤m<n(cid:16)m
n(cid:17)α(cid:16)1 −
m
n(cid:17)}.
Then, using the fact that xα(1 − x), 0 < x < 1, has a maximum at
constant.
α
α+1 results in a better
4 Bounds for the greedy algorithm in Banach spaces
We will now derive bounds for the performance of the weak greedy algorithm in a general
Banach space X.
In this section, we will use the abbreviation σn := σn(F )X and dn :=
dn(F )X. As in the Hilbert space case, we associate with the greedy procedure a lower
triangular matrix A = (ai,j)∞
i,j=0 in the following way. For each j = 0, 1, . . . , we let λj ∈ X ∗
be the linear functional of norm one that satisfies
(i) λj(Vj) = 0,
(ii) λj(fj) = dist(fj, Vj)X.
(4.1)
The existence of such a functional is a simple consequence of the Hahn-Banach theorem (see
[3, Chapt. IV, Cor.14.13]). We let A be the matrix with entries
ai,j = λj(fi).
7
From (ii) of (??) Its diagonal elements aj,j satisfy the inequality
because of the weak greedy selection property (1.8). Also, each entry ai,j satisfies
γσj ≤ aj,j = dist(fj, Vj)X = σj,
(4.2)
ai,j = λj(fi) = λj(fi − g) ≤ kλjkX ∗kfi − gk = kfi − gk,
j < i,
for every g ∈ Vj, since λj(Vj) = 0. Therefore we have
ai,j ≤ dist(fi, Vj) ≤ σj,
j < i.
(4.3)
Theorem 4.1 For the weak greedy algorithm with constant γ in a Banach space X and for
any compact set F contained in the unit ball of X, we have the following inequalities between
σn := σn(F )X and dn := dn(F )X: for any N ≥ 0, K ≥ 1, and 1 ≤ m < K,
σ2
N +i ≤ 2KK K−mγ−2K( KXi=1
KYi=1
N +i)m
σ2
d2K−2m
m
.
(4.4)
Proof: As in the proof of Theorem 3.2, we consider the K × K matrix G which is
formed by the rows and columns of A with indices from {N + 1, . . . , N + K}. Let Xm be
the Kolmogorov subspace of X for which dist(F , Xm) = dm. For each i, there is an element
hi ∈ Xm such that
kfi − hik = dist(fi, Xm)X ≤ dm,
and therefore
λj(fi) − λj(hi) = λj(fi − hi) ≤ kλjkX ∗kfi − hik ≤ dm.
(4.5)
We now consider the vectors (λN +1(h), . . . , λN +K(h)), h ∈ Xm. They span a space
W ⊂ RK of dimension ≤ m. We assume that dim(W ) = m (a slight notational adjustment
has to be made if dim(W ) < m).
It follows from (4.5) that each row gi of G can be
approximated by a vector from W in the ℓ∞ norm to accuracy dm, and therefore in the ℓ2
norm to accuracy √Kdm. Let P be the orthogonal projection of RK onto W . Hence, we
have
√Kdm,
i = 1, . . . , K.
(4.6)
It also follows from (4.3) that
kgi − P gikℓ2 ≤
and therefore
N +j)1/2
σ2
,
kP gikℓ2 ≤ kgikℓ2 ≤( iXj=1
KXi=1
kP gik2
ℓ2 ≤
KXi=1
iXj=1
σ2
N +j ≤ K
σ2
N +i.
KXi=1
(4.7)
8
Next, we apply Lemma 2.1 for this G and W and use estimates (4.2), (4.6) and (4.7) to
derive
γ2K
KYi=1
N +i ≤ (K
m
σ2
d2
σ2
K − m
KXi=1
= K K−m(cid:18)K
≤ 2KK K−m( KXi=1
N +i)m(cid:26) K 2
m(cid:27)K−m
K − m(cid:19)K−m( KXi=1
m(cid:19)m(cid:18) K
N +i)m
d2(K−m)
m
σ2
,
N +i)m
σ2
d2(K−m)
m
and the proof is completed.
(cid:3)
In analogy with Corollary 3.3, we have the following special results for the weak greedy
algorithm in a general Banach space.
Corollary 4.2 Suppose that X is a Banach space. For the weak greedy algorithm with a
constant γ, applied to a compact set F contained in the unit ball of X, the following holds
for σn := σn(F )X and dn := dn(F )X, n = 1, 2, . . . ,
(i) For any such compact set F and n ≥ 1, we have
√2γ−1 min
1≤m<n
n
σn ≤
n−m
2n ( nXi=1
2n
i) m
σ2
n−m
d
n
m .
(4.8)
In particular σ2ℓ ≤ 2γ−1√ℓdℓ, ℓ = 1, 2 . . . .
we have σn ≤ C1n−α+1/2+β, n = 1, 2 . . . , with
(ii) If for α > 0, we have dn ≤ C0n−α, n = 1, 2, . . . , then for any 0 < β < min{α, 1/2},
C1 := max(cid:26)C044α+1γ−4(cid:18) 2β + 1
2β (cid:19)α
, max
n=1,...,7{nα−β−1/2}(cid:27) .
, n = 1, 2, . . . , then σn < √2C0γ−1√ne−c1nα
,
n = 1, 2 . . . , where c1 = 2−1−2αc0. The factor √n can be deleted by reducing the constant c1.
(iii) If for α > 0, we have dn ≤ C0e−c0nα
Proof: The proofs are similar to those of Corollary 3.3 except that we use (4.4) in place
of (3.1).
(i) We take N = 0, K = n, and any 1 ≤ m < n in (4.4) and use the monotonicity of
(σn)n≥0 to obtain
σ2n
n ≤ 2nnn−mγ−2n( nXi=1
i)m
σ2
d2n−2m
m
.
(4.9)
If we take a 2n-th root of both sides, we arrive at (4.8). In particular, if n = 2ℓ and m = ℓ,
we have
√2γ−1(2ℓ)1/4{Σ2ℓ
i=1σ2
σ2ℓ ≤
i }1/4pdℓ ≤
√2γ−1(2ℓ)1/4(2ℓ)1/4pdℓ = 2γ−1pℓdℓ,
where we have used the fact that all σi ≤ 1.
9
(ii) It follows from the monotonicity of (σn)n≥0 and (4.4) for N = K = n and any
1 ≤ m < n that
σ2n ≤
Given our β, we define m =: ⌊ 2β
2β+1 n⌋ + 1 (m < n for n ≥ 2 > 2β + 1). It follows that
δ :=
.
nd(1−δ)
m ,
√2nγ−1σδ
m
n
1
n(cid:19) .
δ =
m
n ∈(cid:18) 2β
2β + 1
,
2β
2β + 1
+
(4.10)
(4.11)
We next prove (ii) by contradiction. Suppose it is not true and M is the first value where
σM > C1M −α+β+1/2. Clearly, because of the definition of C1, and the fact that σn ≤ 1, we
must have M > 7. We first consider the case M = 2n, and therefore n > 3. From (4.10) we
have
C1(2n)−α+β+1/2 < σ2n ≤
√2nγ−1C δ
1nδ(−α+β+1/2)C 1−δ
0
(δn)−α(1−δ),
where we have used the fact that σn ≤ C1n−α+β+1/2 and dm ≤ C0m−α. It follows that
C 1−δ
1 < C 1−δ
0
2α−βγ−1δ−α(1−δ)n
2β+1
2
(δ− 2β
2β+1 ),
and therefore
C1 < C02
Since for n ≥ 4 > 2(2β + 1), we have
α−β
1−δ γ− 1
1−δ δ−αn
δ− 2β
2β+1
1−δ
·
2β+1
2
.
δ <
4β + 1
2(2β + 1)
< 1,
and therefore
1
1 − δ
< 2(2β + 1).
This gives
2β + 1
2
2β+1
δ − 2β
1 − δ
·
<
(2β + 1)2
n
,
and thus n
2β+1
2
(δ− 2β
2β+1 ) < n
(2β+1)2
n < 2(2β+1)2
.
Then, for β < min{α, 1/2}
C1 < C022(α−β)(2β+1)γ−2(2β+1)(cid:18) 2β
2β (cid:19)α
< C022(2α+1)γ−4(cid:18)2β + 1
2(2β+1)2
2β + 1(cid:19)−α
< C022(4α+1)γ−4(cid:18) 2β + 1
2β (cid:19)α
,
(4.12)
which is the desired contradiction. Likewise, if M = 2n + 1 (since M > 7, we have n > 3),
for −α + β + 1/2 < 0 (which is the meaningful case),
C12−α+β+1/2(2n)−α+β+1/2 < C1(2n + 1)−α+β+1/2 < σ2n+1 ≤ σ2n
< √2nγ−1C δ
1nδ(−α+β+1/2)C 1−δ
0
(δn)−α(1−δ),
and following the same argument as above we get
C1 < C02− 1
2(1−δ) 22 α−β
1−δ γ−2(2β+1)(cid:18) 2β
2β + 1(cid:19)−α
2(2β+1)2
2β (cid:19)α
< C022(4α+1)γ−4(cid:18)2β + 1
,
10
where we have used that 2− 1
2(1−δ) < 1, and the proof is completed.
(iii) From (i), we have
σ2n+1 ≤ σ2n ≤ 2γ−1pndn ≤ 2γ−1pC0√ne− c0
from which (iii) easily follows.
2 nα
<p2C0γ−1√2n + 1e−c02−1−2α(2n+1)α
,
(cid:3)
5 Lower bounds in a Banach space
It is natural to ask whether the factor √n is necessary when proving results in a Banach
space. Here, we shall provide examples which show that a loss of this type is indeed necessary.
However, as it will be seen, there is still a small gap between what we have proved for direct
estimates and what the examples below provide.
Let us begin by considering the space X := ℓ∞(N ∪ {0}) equipped with its usual norm.
We consider a monotone decreasing sequence x0 ≥ x1 ≥ x2 ≥ · · · of positive real numbers
which converge to zero and define
fj := xjej,
j = 0, 1, . . . ,
where ej, j = 0, 1, . . . are the usual coordinate vectors in RN∪{0}. Let F := {f0, f1, . . .}.
From the monotonicity of the xj's, the greedy algorithm for F in X can choose the elements
from F in order f0, f1, . . . . Hence,
σj = σj(F )X = xj,
j ≥ 0.
We want to give an upper bound for the Kolmogorov width of F . For this, we shall use the
1 in ℓm
following result (see (7.2) of Chapter 14 in [4]) on s-widths of the unit ball bm
∞:
1 of ℓm
(5.1)
Let us now define the sequence {xj}j≥0 so that in position 2k−1 ≤ j ≤ 2k − 1 it has the
constant value 2−kα, for k = 0, 1, . . . , where α > 1/2. It follows that,
1 )X ≤ C {ln(m/s)}1/2 s−1/2,
1 ≤ s ≤ m/2.
ds(bm
σn(F )X = O(n−α), n = 1, 2, . . . .
We shall now bound the N-width of F when N = 2n+1 by constructing a good space XN
of dimension ≤ N for approximating F . The space XN will be the span of a set E of at most
N vectors. First, we place in E all of the vectors, e1, . . . , e2n. Next, for each k = 1, . . . n,
we use (5.1) to choose a basis for the space of dimension 2n−k whose vectors are supported
on [2n+k, 2n+k+1 − 1] and this space approximates each of the fj, j = 2n+k, . . . , 2n+k+1 − 1,
in X to accuracy C02−(n+k)α√k2−(n−k)/2. We place these basis vectors in E. Notice that
xj ≤ 2−2nα for j ≥ 22n. This means that for the space XN := span(E) with dimension ≤ N
we have
dN (F )X ≤ dist(F , XN )X ≤ max(cid:26)2−2nα, max
= max(cid:26)2−2nα, C02−n(α+1/2) · max
C02−(n+k)α2−(n−k)/2√k(cid:27)
2−k(α−1/2)√k(cid:27) ≤ C12−n(α+1/2), α > 1/2.
1≤k≤n
1≤k≤n
11
From the monotonicity of (dn(F )X)n≥0, we obtain that
This example shows that the factor √n which appears in (ii) of Corollary 4.2 can in
dn(F )X ≤ C2n−α−1/2, n = 1, 2, . . . .
general not be removed.
References
[1] P. Binev, A. Cohen, W. Dahmen, R. DeVore, G. Petrova, and P. Wojtaszczyk, Conver-
gence rates for greedy algorithms in reduced bases Methods, SIAM J. Math. Anal., 43
(2011), 1457–1472.
[2] A. Buffa, Y. Maday, A.T. Patera, C. Prud'homme, and G. Turinici, A Priori convergence
of the greedy algorithm for the parameterized reduced basis, M2AN Math. Model. Numer.
Anal., 46(2012), 595–603.
[3] E. Hewitt, K. Stromberg, Real and Abstract Analysis, Springer Verlag, Berlin 1969
[4] G.G. Lorentz, M. von Golitschek, and Y. Makovoz, Constructive Approximation: Ad-
vanced Problems, Springer Verlag, vol. 304, New York, 1996.
[5] Y. Maday, A.T. Patera, and G. Turinici, A priori convergence theory for reduced-basis
approximations of single-parametric elliptic partial differential equations, J. Sci. Com-
put., 17(2002), 437–446.
[6] Y. Maday, A. T. Patera, and G. Turinici, Global a priori convergence theory for reduced-
basis approximations of single-parameter symmetric coercive elliptic partial differential
equations, C. R. Acad. Sci., Paris, Ser. I, Math., 335(2002), 289–294.
Ronald DeVore, Department of Mathematics, Texas A&M University, College Station, TX,
[email protected]
Guergana Petrova, Department of Mathematics, Texas A&M University, College Station,
TX, [email protected]
Przemyslaw Wojtaszczyk, Institute of Applied Mathematics, and Interdisciplinary Centre
for Mathematical and Computational Modelling, University of Warsaw, Warsaw, Poland,
[email protected]
12
|
1202.1722 | 1 | 1202 | 2012-02-08T14:54:07 | Uniqueness of Gibbs Measure for Models With Uncountable Set of Spin Values on a Cayley Tree | [
"math.FA",
"math-ph",
"math-ph"
] | We consider models with nearest-neighbor interactions and with the set $[0,1]$ of spin values, on a Cayley tree of order $k\geq 1$.
It is known that the "splitting Gibbs measures" of the model can be described by solutions of a nonlinear integral equation. For arbitrary $k\geq 2$ we find a sufficient condition under which the integral equation has unique solution, hence under the condition the corresponding model has unique splitting Gibbs measure. | math.FA | math |
UNIQUENESS OF GIBBS MEASURE FOR MODELS WITH
UNCOUNTABLE SET OF SPIN VALUES ON A CAYLEY TREE
YU. KH. ESHKABILOV, F. H. HAYDAROV, U. A. ROZIKOV
Abstract. We consider models with nearest-neighbor interactions and with the set
[0, 1] of spin values, on a Cayley tree of order k ≥ 1. It is known that the "splitting
Gibbs measures" of the model can be described by solutions of a nonlinear integral
equation. For arbitrary k ≥ 2 we find a sufficient condition under which the integral
equation has unique solution, hence under the condition the corresponding model has
unique splitting Gibbs measure.
Mathematics Subject Classifications (2010). 82B05, 82B20 (primary); 60K35
(secondary)
Key words. Cayley tree, configuration, Gibbs measures, uniqueness.
1. Introduction
In this paper we consider models (Hamiltonians) with a nearest neighbor interaction
and uncountably many spin values on a Cayley tree.
One of the central problems in the theory of Gibbs measures is to describe infinite-
volume (or limiting) Gibbs measures corresponding to a given Hamiltonian. The exis-
tence of such measures for a wide class of Hamiltonians was established in the ground-
breaking work of Dobrushin (see, e.g. [18]). However, a complete analysis of the set of
limiting Gibbs measures for a specific Hamiltonian is often a difficult problem.
There are several papers devoted to models on Cayley trees, see for example [1]-
[6], [8], [9], [12], [14]- [16], [19], [20], [22]. All these works devoted to models with a finite
set of spin values. These models have the following common property: The existence
of finitely many translation-invariant and uncountable numbers of the non-translation-
invariant extreme Gibbs measures. Also for several models (see, for example, [5,8,15,16])
it were proved that there exist three periodic Gibbs measures (which are invariant with
respect to normal subgroups of finite index of the group representation of Cayley tree)
and there are uncountable number of non-periodic Gibbs measures.
In [7] the Potts model with a countable set of spin values on a Cayley tree is considered
and it was showed that the set of translation-invariant splitting Gibbs measures of the
model contains at most one point, independently on parameters of the Potts model with
countable set of spin values on Cayley tree. This is a crucial difference from the models
with a finite set of spin values, since the last ones may have more than one translation-
invariant Gibbs measures.
How "rich" is the set of translation-invariant Gibbs measures for models with an
uncountable spin values? In [17] models with nearest-neighbor interactions and with the
1
2
YU. KH. ESHKABILOV, F. H. HAYDAROV, U. A. ROZIKOV
set [0, 1] of spin values, on a Cayley tree of order k ≥ 1 are considered and we reduced
the problem of describing the "splitting Gibbs measures" of the model to the description
of the solutions of some nonlinear integral equation. For k = 1 we showed that the
integral equation has a unique solution. In case k ≥ 2 some models (with the set [0, 1]
of spin values) which have a unique splitting Gibbs measure are constructed. In this
paper we continue this investigations and give a sufficient condition on Hamiltonian of
the model with an uncountable set of spin values under which the model has unique
translation-invariant splitting Gibbs measure. But we have not any example of model
(with uncountable spin values) with more than one translation-invariant Gibbs measure.
So this is still an open problem to find such a model.
2. Preliminaries
A Cayley tree Gk = (V, L) of order k ≥ 1 is an infinite homogeneous tree (see [1]), i.e.,
a graph without cycles, with exactly k + 1 edges incident to each vertices. Here V is the
set of vertices and L that of edges (arcs).
Consider models where the spin takes values in the set [0, 1], and is assigned to the
vertexes of the tree. For A ⊂ V a configuration σA on A is an arbitrary function
σA : A → [0, 1]. Denote ΩA = [0, 1]A the set of all configurations on A. A configuration
σ on V is then defined as a function x ∈ V 7→ σ(x) ∈ [0, 1]; the set of all configurations
is [0, 1]V . The (formal) Hamiltonian of the model is :
H(σ) = −J Xhx,yi∈L
ξσ(x)σ(y),
(2.1)
where J ∈ R \ {0} and ξ : (u, v) ∈ [0, 1]2 → ξuv ∈ R is a given bounded, measurable
function. As usually, hx, yi stands for nearest neighbor vertices.
Let λ be the Lebesgue measure on [0, 1]. On the set of all configurations on A the
a priori measure λA is introduced as the Afold product of the measure λ. Here and
further on A denotes the cardinality of A. We consider a standard sigma-algebra B
of subsets of Ω = [0, 1]V generated by the measurable cylinder subsets. A probability
measure µ on (Ω,B) is called a Gibbs measure (with Hamiltonian H) if it satisfies the
DLR equation, namely for any n = 1, 2, . . . and σn ∈ ΩVn:
and β = 1
a 'ball' on the tree, of radius l = 1, 2, . . ., centered at a fixed vertex x0 (an origin):
T , T > 0 is temperature. Here and below, Wl stands for a 'sphere' and Vl for
Wl = {x ∈ V : d(x, x0) = l}, Vl = {x ∈ V : d(x, x0) ≤ l};
where ν Vn
ωWn+1
is the conditional Gibbs density
ν Vn
ωWn+1
(σn) =
µ(cid:16)nσ ∈ Ω : σ(cid:12)(cid:12)Vn
(σn),
ωWn+1
µ(dω)ν Vn
= σno(cid:17) =ZΩ
Zn(cid:16)ω(cid:12)(cid:12)Wn+1(cid:17) exp (cid:16)−βH(cid:16)σn ω(cid:12)(cid:12)Wn+1(cid:17)(cid:17) ,
1
ON GIBBS MEASURE OF A MODEL
3
and
Ln = {hx, yi ∈ L : x, y ∈ Vn};
distance d(x, y), x, y ∈ V , is the length of (i.e. the number of edges in) the shortest path
connecting x with y. ΩVn is the set of configurations in Vn (and ΩWn that in Wn; see
denote the restrictions of configurations σ, ω ∈ Ω
to Vn and Wn+1, respectively. Next, σn : x ∈ Vn 7→ σn(x) is a configuration in Vn and
below). Furthermore, σ(cid:12)(cid:12)Vn
H(cid:16)σn ω(cid:12)(cid:12)Wn+1(cid:17) is defined as the sum H (σn) + U(cid:16)σn, ω(cid:12)(cid:12)Wn+1(cid:17) where
and ω(cid:12)(cid:12)Wn+1
H (σn) = −J Xhx,yi∈Ln
ξσn(x)σn(y),
U(cid:16)σn, ω(cid:12)(cid:12)Wn+1(cid:17) = −J
Xhx,yi: x∈Vn,y∈Wn+1
ξσn(x)ω(y).
Finally, Zn(cid:16)ω(cid:12)(cid:12)Wn+1(cid:17) stands for the partition function in Vn, with the boundary condition
ω(cid:12)(cid:12)Wn+1
:
Zn(cid:16)ω(cid:12)(cid:12)Wn+1(cid:17) =ZΩVn
exp (cid:16)−βH(cid:16)eσn ω(cid:12)(cid:12)Wn+1(cid:17)(cid:17) λVn(deσn).
Due to the nearest-neighbor character of the interaction, the Gibbs measure possesses
a natural Markov property: for given a configuration ωn on Wn, random configurations in
Vn−1 (i.e., 'inside' Wn) and in V \Vn+1 (i.e., 'outside' Wn) are conditionally independent.
We use a standard definition of a translation-invariant measure (see, e.g., [18]). The
main object of study in this paper are translation-invariant Gibbs measures for the model
(2.1) on Cayley tree. In [17] this problem of description of such measures was reduced to
the description of the solutions of a nonlinear integral equation. For finite and countable
sets of spin values this argument is well known (see, e.g. [2]- [7], [14], [19], [20], [22]).
Write x < y if the path from x0 to y goes through x. Call vertex y a direct successor
of x if y > x and x, y are nearest neighbors. Denote by S(x) the set of direct successors
of x. Observe that any vertex x 6= x0 has k direct successors and x0 has k + 1.
Let h : x ∈ V 7→ hx = (ht,x, t ∈ [0, 1]) ∈ R[0,1] be mapping of x ∈ V \ {x0} with
ht,x < C where C is a constant which does not depend on t. Given n = 1, 2, . . ., consider
the probability distribution µ(n) on ΩVn defined by
µ(n)(σn) = Z −1
n exp −βH(σn) + Xx∈Wn
exp −βH(eσn) + Xx∈Wn
hσ(x),x! ,
heσ(x),x! λVn(deσn).
(2.2)
(2.3)
Here, as before, σn : x ∈ Vn 7→ σ(x) and Zn is the corresponding partition function:
Zn =ZΩVn
4
YU. KH. ESHKABILOV, F. H. HAYDAROV, U. A. ROZIKOV
The probability distributions µ(n) are compatible if for any n ≥ 1 and σn−1 ∈ ΩVn−1:
(2.4)
µ(n)(σn−1 ∨ ωn)λWn(d(ωn)) = µ(n−1)(σn−1).
ZΩWn
Here σn−1 ∨ ωn ∈ ΩVn is the concatenation of σn−1 and ωn. In this case there exists
a unique measure µ on ΩV such that, for any n and σn ∈ ΩVn, µ(cid:18)(cid:26)σ(cid:12)(cid:12)(cid:12)Vn
= σn(cid:27)(cid:19) =
µ(n)(σn).
Definition 2.1. The measure µ is called splitting Gibbs measure corresponding to Hamil-
tonian (2.1) and function x 7→ hx, x 6= x0.
The following statement describes conditions on hx guaranteeing compatibility of the
corresponding distributions µ(n)(σn).
Proposition 2.2. [17] The probability distributions µ(n)(σn), n = 1, 2, . . ., in (2.2) are
compatible iff for any x ∈ V \ {x0} the following equation holds:
0 exp(J βξtu)f (u, y)du
0 exp(J βξ0u)f (u, y)du
(2.5)
.
f (t, x) = Yy∈S(x)R 1
R 1
Here, and below f (t, x) = exp(ht,x − h0,x), t ∈ [0, 1] and du = λ(du) is the Lebesgue
measure.
From Proposition 2.2 it follows that for any h = {hx ∈ R[0,1], x ∈ V } satisfying
(2.5) there exists a unique Gibbs measure µ and vice versa. However, the analysis of
solutions to (2.5) is not easy. This difficulty depends on the given function ξ. In the
next sections we will give a condition on such function under which the corresponding
integral equation has unique solution.
3. Uniqueness of translational - invariant solution of (2.5)
In this section we consider ξtu as a continuous function and we are going to fund
a condition on ξtu under which the equation (2.5) has unique solution in the class of
translational-invariant functions f (t, x), i.e f (t, x) = f (t), for any x ∈ V . For such
functions equation (2.5) can be written as
0 K(t, u)f (u)du
0 K(0, u)f (u)du!k
f (t) = R 1
R 1
where K(t, u) = exp(J βξtu) > 0, f (t) > 0, t, u ∈ [0, 1].
We put
,
(3.1)
We are interested to positive continuous solutions to (3.1), i.e. such that
C +[0, 1] = {f ∈ C[0, 1] : f (x) ≥ 0}.
0 [0, 1] = {f ∈ C[0, 1] : f (x) ≥ 0} \ {θ ≡ 0}.
f ∈ C +
Note that equation (3.1) is not linear for any k ≥ 1.
ON GIBBS MEASURE OF A MODEL
5
Define the linear operator W : C[0, 1] → C[0, 1] by
K(t, u)f (u)du
0
(W f )(t) =Z 1
ω(f ) ≡ (W f )(0) =Z 1
0
and defined the linear functional ω : C[0, 1] → R by
K(0, u)f (u)du.
Then equation (3.1) can be written as
where
f (t) = (Akf )(t) = ((Bf )(t))k,
(Bf )(t) =
(W f )(t)
(W f )(0)
, f ∈ C +
0 [0, 1], k ≥ 1.
(3.2)
(3.3)
(3.4)
(3.5)
3.1. Existence of solutions to the nonlinear equation (3.4). In [17] for k = 1 we
have proved that the equation (3.4) has unique solution for arbitrary K(·,·) ∈ C +[0, 1]2
and f (·) ∈ C +[0, 1]. But for k ≥ 2 the uniqueness is not proved yet. Denote
Fk =(f ∈ C +[0, 1] : f (t) ≥(cid:18) m
M0(cid:19)k) , k ∈ N,
where
m = min
t,u∈[0,1]
K(t, u), M0 = max
u∈[0,1]
K(0, u).
It is easy to see that Fk is a closed and convex subset of C[0, 1]. Moreover this set is
invariant with respect to operator Ak, i.e. Ak(Fk) ⊂ Fk.
Proposition 3.2. The operator Ak is continuous on Fk for any k ≥ 2.
Proof. For arbitrary C > 0 we denote
F0 =(cid:8)f ∈ C +[0, 1] : f (t) ≥ C, ∀t ∈ [0, 1](cid:9) .
It is obvious that the operator A1 is continuous on the set F0 (see Lemma 2 in [17]).
Let f ∈ Fk be an arbitrary element and {fn} ⊂ Fk such that limn→∞ fn = f . Since
the operator A1 is continuous we have limn→∞ A1fn = A1f . Consequently, there exists
C1 > 0 such that kA1fnk ≤ C1 for n ∈ N. Moreover we have
, t ∈ [0, 1],
(A1f )(t) ≤ C2 =
M
m0
where
We have
M = max
t,u∈[0,1]
K(t, u), m0 = min
u∈[0,1]
K(0, u).
Akfn − Akf = (Bfn)k − (Bf )k = qk,n(t)(A1fn − A1f ),
(3.6)
6
where
Consequently,
Hence
YU. KH. ESHKABILOV, F. H. HAYDAROV, U. A. ROZIKOV
qk,n(t) =
k−1Xj=0
(A1fn)k−j−1(t)(A1f )j(t) > 0, t ∈ [0, 1].
qk,n(t) ≤ C =
k−1Xj=0
(C1)k−j−1(C2)j , t ∈ [0, 1].
kAkfn − Akfk ≤ CkA1fn − A1fk, n ∈ N.
Since A1 is a continuous from the last inequality it follows that Ak is continuous on
Fk.
(cid:3)
Denote
k =(f ∈ C +[0, 1] :(cid:18) m
M0(cid:19)k
F 0
≤ f (t) ≤(cid:18) M
m0(cid:19)k) .
0 [0, 1] is a solution of the equation Akf = f , then
Proposition 3.3. Let k ≥ 2. If f ∈ C +
f ∈ F 0
k .
Proof. Straightforward.
Proposition 3.4. Let k ≥ 2. The set Ak(F 0
Proof. By Arzel´a-Askoli's theorem (see [21], ch.III,§3) it suffices to prove that the set of
functions Ak(F 0
k ) is equi-continuous and there exists γ > 0 such that
k ) is relatively compact in C[0, 1].
(cid:3)
h(t) ≤ γ, ∀t ∈ [0, 1] and ∀h ∈ Ak(F 0
k ).
Let h ∈ Ak(F 0
k ) be an arbitrary function, we have
m0(cid:19)k
0 < h(t) ≤(cid:18) M
k such that h = Akf .
k ) is equi-continuous. For arbitrary t, t′ ∈ [0, 1] we have
and there exists a function f ∈ F 0
Now we shall prove that Ak(F 0
(h = Akf )
h(t) − h(t′) = (A1f )k(t) − (A1f )k(t′) =
k−1Xj=0
(A1f )k−j−1(t)(A1f )j(t′)(A1f )(t) − (A1f )(t′) ≤
k(cid:18) M
m0(cid:19)k−1
k(cid:18) M
m0(cid:19)2k−1
ω(f )Z 1
0 K(t, u) − K(t′, u)f (u)du ≤
ω(f )Z 1
0 K(t, u) − K(t′, u)du,
1
1
where ω(f ) is defined in (3.3).
ON GIBBS MEASURE OF A MODEL
7
We have
Consequently,
h(t) − h(t′) ≤
M0(cid:19)k
ω(f ) ≥ m0 ·(cid:18) m
, f ∈ F 0
k .
m0(cid:19)2k−1Z 1
m0(cid:18) M0
m(cid:19)k(cid:18) M
k
0 K(t, u) − K(t′, u)du.
Since the kernel K(t, u) is uniformly continuous on [0, 1]2, we conclude that Ak(F 0
k )
(cid:3)
also is equi-continuous.
By Propositions 3.2-3.4 and Schauder's theorem (see [13], p.20) one gets the following
Theorem 3.5. The equation Akf = f has at least one solution in C +
of all solutions of the equation is a subset in F 0
k .
3.6. The Hammerstein's nonlinear equation. For every k ∈ N we consider an inte-
gral operator Hk acting in C +[0, 1] as follows:
0 [0, 1] and the set
(Hkf )(t) =Z 1
0
K(t, u)f k(u)du.
If k ≥ 2 then the operator Hk is a nonlinear operator which is called Hammerstein's
operator of order k. Moreover the linear operator equation H1f = f has a unique
positive solution f in C[0, 1] (see [10], p.80).
For a nonlinear homogeneous operator A it is known that if there is one positive eigen-
function of the operator A then the number of the positive eigenfunctions is continuum
(see [10], p.186).
Denote
Lemma 3.7. The equation
M0 =(cid:8)f ∈ C +[0, 1] : f (0) = 1(cid:9) .
has a strongly positive solution iff the equation
Akf = f, k ≥ 2
Hkf = λf, k ≥ 2
(3.7)
(3.8)
has a strongly positive solution in M0.
Proof. Necessariness. Let f0 ∈ C +
From this equality we get
(W f0)(t) = ω(f0) kpf0(t).
(Hkh)(t) = λ0h(t),
0 [0, 1] be a solution of the equation (3.7). We have
where h(t) = kpf0(t) and λ0 = ω(f0) > 0.
It is easy to see that h ∈ M0 and h(t) is an eigenfunction of the Hammerstein's
operator Hk, corresponding the positive eigenvalue λ0.
8
YU. KH. ESHKABILOV, F. H. HAYDAROV, U. A. ROZIKOV
Sufficiency. Let k ≥ 2 and h ∈ M0 be an eigenfunction of the Hammerstein's operator.
Then there is a number λ0 > 0 such that Hkh = λ0h. From h(0) = 1 we get λ0 =
(Hkh)(0) = ω(hk). Then
h(t) =
Hkh
ω(hk)
.
From this equality we get Akf0 = f0 with f0 = hk ∈ C +
proof.
Theorem 3.8. If k ≥ 2 then every number λ > 0 is an eigenvalue of the Hammerstein's
operator Hk.
Proof. By Theorem 3.5 and Lemma 3.7 there exist λ0 > 0 and f0 ∈ M0 such that
0 [0, 1]. This completes the
(cid:3)
Hkf0 = λ0f0.
Take λ ∈ (0, +∞), λ 6= λ0. Define function h0(t) ∈ C +
0 [0, 1] by
Then
f0(t),
λ0
h0(t) = k−1r λ
Hkh0 = Hk k−1r λ
λ0
t ∈ [0, 1].
f0! = λh0.
This completes the proof.
(cid:3)
Denote
K =(cid:26)f ∈ C +[0, 1] : M · min
M(cid:19) 1
M ·(cid:18) 1
Pk =(f ∈ C[0, 1] :
m
k−1
t∈[0,1]
f (t) ≥ m · max
t∈[0,1]
f (t)(cid:27) ,
k−1) , k ≥ 2.
m(cid:19) 1
M
m ·(cid:18) 1
≤ f (t) ≤
Proposition 3.9. Let k ≥ 2.
a) The following holds
b) If a function f0 ∈ C +
0 [0, 1] is a solution of the equation
Hk(C +[0, 1]) ⊂ K.
Hkf = f
(3.9)
then f0 ∈ Pk.
Proof. a) Let h ∈ Hk(C +[0, 1]) be an arbitrary function. Then there exists a function
f ∈ C +[0, 1] such that h = Hkf . Since h is continuous on [0, 1], there are t1, t2 ∈ [0, 1]
such that
hmin = min
t∈[0,1]
hmax = max
t∈[0,1]
h(t) = h(t1) = (Hkf )(t1),
h(t) = h(t2) = (Hkf )(t2).
By the property a) we have
Then we obtain
Also we have
Then fmin ≥ mf k
min, i.e.
Hence be the property a) we get
kfk ≥(cid:18) 1
M(cid:19) 1
k−1
.
f (t) ≥ fmin = min
t∈[0,1]
f (t) ≥
m
M kfk.
.
k−1
m
f (t) ≥
M (cid:18) 1
M(cid:19) 1
f (t) = (Hkf )(t) ≥ mZ 1
f k(u)du ≥ mf k
m(cid:19) 1
fmin ≤(cid:18) 1
k−1
0
.
min.
f (t) ≤ fmax ≤
M
m
fmin ≤
k−1
M
m (cid:18) 1
m(cid:19) 1
.
ON GIBBS MEASURE OF A MODEL
9
Hence
i.e. h ∈ K.
Consequently,
b) Let f ∈ C +
f k(u)du ≥ mZ 1
hmin ≥ mZ 1
0 [0, 1] be a solution of the equation (3.9). Then we have kfk ≤ Mkfkk.
f k(u)du =
K(t2, u)
hmax,
m
M
M
0
0
Thus we have f ∈ Pk.
3.10. The uniqueness of fixed point of the operators Ak and Hk. Now we shall
prove that Akf = f and Hkf = f have a unique solution in C +
Lemma 3.11. Assume function f ∈ C[0, 1] changes its sign on [0, 1]. Then for every
a ∈ R the following inequality holds
0 [0, 1].
(cid:3)
1
kfak ≥
where fa = fa(t) = f (t) − a, t ∈ [0, 1].
Proof. By conditions of lemma there are t1, t2 ∈ [0, 1] such that
fmin = f (t1) < 0, fmax = f (t2) > 0.
n + 1kfk, n ∈ N,
In case a = 0 the proof is obvious. We assume a > 0
a) Let fmin ≥ fmax. Then kfk = fmin = f (t1). Hence
kfak = max{f (t1) − a,f (t2) − a} = f (t1) − a > f (t1) = kfk ≥
1
n + 1kfk, n ∈ N.
10
YU. KH. ESHKABILOV, F. H. HAYDAROV, U. A. ROZIKOV
b) Let fmin < fmax and 1
Consequently,
2kfk ≥ a. Then kfk = fmax = f (t2) and kfk − a ≥ a > 0.
kfak = max{f (t1) − a,f (t2) − a} ≥ f (t2) − a = kfk − a ≥
2kfk < a. Then kfk = f (t2) and
1
2kfk ≥
c) Let fmin < fmax and 1
kfak = max{f (t1) − a,f (t2) − a} ≥ f (t1) − a > a >
1
n + 1kfk, n ∈ N.
1
2kfk ≥
1
n + 1kfk, n ∈ N.
Thus for a > 0 the proof is completed. For a < 0 we put ga(t) = g(t) − a′ with
g(t) = −f (t) and a′ = −a > 0. Then
kfak = kgak ≥
n + 1kfk, n ∈ N.
n + 1kgk =
1
1
This completes the proof.
Theorem 3.12. Let k ≥ 2. If the kernel K(t, u) satisfies the condition
(cid:3)
(3.10)
(cid:18) M
m(cid:19)k
M(cid:17)k
−(cid:16) m
<
1
k
,
then the operator Hk has a unique fixed point in C +
0 [0, 1].
max
0 [0, 1] and f2 ∈ C +
Proof. By Theorem 3.8 the Hammerstein's equation Hkf = f has at least one solution.
Assume that there are two solutions f1 ∈ C +
0 [0, 1], i.e Hkfi = fi,
i = 1, 2. Denote f (t) = f1(t) − f2(t). Then by Theorem 46.6 of [11] the function f (t)
changes its sign on [0, 1]. From Lemma 3.11 we get
t∈[0,1](cid:12)(cid:12)(cid:12)(cid:12)f (t) −
(γ1 + γ2)Z 1
M(cid:17)k
γ1 =(cid:16) m
f (t) =Z 1
f (s)ds(cid:12)(cid:12)(cid:12)(cid:12) ≥
, γ2 =(cid:18) M
m(cid:19)k
By a mean value Theorem we have
K(t, u)kξk−1(u)f (u)du,
1
2kfk,
where
k
2
0
0
.
here ξ ∈ C +[0, 1] and
min{f1(t), f2(t)} ≤ ξ(t) ≤ max{f1(t), f2(t)}, t ∈ [0, 1].
By Proposition 3.9 we have ξ ∈ Pk, i.e.
Hence
k−1
m
M
≤ ξ(t) ≤
M (cid:18) 1
M(cid:19) 1
γ1 ≤ K(t, u)ξk−1(u) ≤ γ2, t, u ∈ [0, 1].
m (cid:18) 1
m(cid:19) 1
k−1
, t ∈ [0, 1].
ON GIBBS MEASURE OF A MODEL
11
Therefore
Then
γ1 + γ2
(cid:12)(cid:12)(cid:12)(cid:12)k · K(t, u)ξk−1(u) −
(cid:12)(cid:12)(cid:12)(cid:12)f (t) −
(γ1 + γ2)Z 1
(cid:12)(cid:12)(cid:12)(cid:12) ≤
f (u)du(cid:12)(cid:12)(cid:12)(cid:12) ≤
k
2
k
2
2
0
γ2 − γ1
2
.
(γ2 − γ1)kfk.
(3.11)
Assume the kernel K(t, u) satisfies the condition (3.10). Then k(γ2 − γ1) < 1 and the
inequality (3.11) contradicts to Lemma 3.11. This completes the proof.
Theorem 3.13. Let k ≥ 2. If the kernel K(t, u) satisfies the condition (3.10), then for
every λ > 0 the Hammerstein's equation Hkf = λf has unique solution in C +
(cid:3)
0 [0, 1].
Proof. Clearly the equation Hkf = λf is equivalent to the following equation
Z 1
0
Kλ(t, u)f k(u)du = f (t),
(3.12)
0 [0, 1].
λ K(t, u). The kernel Kλ(t, u) satisfies the condition (3.10) with m = m
λ
λ . Consequently, by Theorem 3.12 it follows that the equation (3.12) has
(cid:3)
where Kλ(t, u) = 1
and M = M
unique solution in C +
Theorem 3.14. Let k ≥ 2. If the kernel K(t, u) satisfies the condition (3.10), then the
equation Akf = f has unique solution in C +
Proof. Assume there are two solutions f1, f2 ∈ C +[0, 1], f1 6= f2, i.e. Akfi = fi, i = 1, 2.
By Lemma 3.7 the functions hi(t) = kpfi(t), t ∈ [0, 1] are solutions of the Hammerstein's
equation, i.e.
0 [0, 1].
Hkhi = λihi,
i = 1, 2,
where λi = ω(fi) > 0 and hi ∈ M0. On the other hand Theorem 3.13 implies that
λ1 6= λ2. Let h0(t) ∈ C +[0, 1] be a fixed point of the Hammerstein's operator Hk. Then
by Theorems 3.8 and 3.13 we get
Consequently,
hi = k−1pλih0(t), i = 1, 2.
= γk, with γ = k−1r λ1
λ2
f1(t)
f2(t)
.
Using this equality we obtain
f1(t) = (Akf1)(t) = Ak(γkf2) = Akf2(t) = f2(t).
This completes the proof.
(cid:3)
Consider the following Hamiltonian
H(σ) = −J Xhx,yi∈L
ξσ(x)σ(y) = − Xhx,yi∈L
ln K(σ(x), σ(y)),
(3.13)
where J ∈ R \ {0} and K(t, u) satisfies the condition (3.10). Then as a corollary of
Proposition 2.2 and Theorem 3.14 we get the following
12
YU. KH. ESHKABILOV, F. H. HAYDAROV, U. A. ROZIKOV
Theorem 3.15. Let k ≥ 2. If the function K(t, u) of the Hamiltonian (3.13) satisfies the
condition (3.10), then the model (3.13) has unique translational invariant Gibbs measure.
Example. It is easy to see that the condition (3.10) is satisfied iff
(3.14)
M
m ≤ ηk =
Consider the following function
ks 1 + √4k2 + 1
nXj=1
For this function we have m = a, M =Pm
mXi=1
K(t, u) =
cij tiuj + a, cij ≥ 0, a > 0.
i=1Pn
2k
, k ≥ 2.
a) If 1
b) If 1
satisfied.
j=1 cij ≤ ηk − 1 then for function (3.14) the condition (3.10) is satisfied.
j=1 cij > ηk − 1 then for function (3.14) the condition (3.10) is not
Remark. Is there a kernel K(t, u) > 0 of the equation (3.1) when the equation has
i=1Pn
aPm
aPm
i=1Pn
j=1 cij + a. The following is obvious
more than one solutions? This is still open problem.
Acknowledgements
UAR thanks the TWAS Research Grant: 09-009 RG/Maths/As-I; UNESCO FR:
3240230333. He also thanks the Department of Algebra, University of Santiago de Com-
postela, Spain, for providing financial support to his visit to the Department.
References
[1] Baxter, R.J.: Exactly Solved Models in Statistical Mechanics (Academic, London, 1982).
[2] Bleher, P.M. and Ganikhodjaev N.N.: On pure phases of the Ising model on the Bethe lattice.
Theor. Probab. Appl. 35 (1990), 216-227.
[3] Bleher, P.M., Ruiz, J. and Zagrebnov V.A.: On the purity of the limiting Gibbs state for the Ising
model on the Bethe lattice. Journ. Statist. Phys. 79 (1995), 473-482.
[4] Ganikhodjaev, N.N.: On pure phases of the ferromagnet Potts with three states on the Bethe lattice
of order two. Theor. Math. Phys. 85 (1990), 163 -- 175.
[5] Ganikhodjaev, N.N. and Rozikov, U.A. Description of periodic extreme Gibbs measures of some
lattice model on the Cayley tree. Theor. and Math. Phys. 111 (1997), 480-486.
[6] Ganikhodjaev, N.N. and Rozikov, U.A.: On disordered phase in the ferromagnetic Potts model on
the Bethe lattice. Osaka J. Math. 37 (2000), 373-383.
[7] Ganikhodjaev, N.N. and Rozikov, U.A. : The Potts model with countable set of spin values on a
Cayley Tree. Letters Math. Phys. 75 (2006), 99-109.
[8] Ganikhodjaev, N.N. and Rozikov, U.A. On Ising model with four competing interactions on Cayley
tree. Math. Phys. Anal. Geom. 12 (2009), 141-156.
[9] Kotecky, R. and Shlosman, S.B.: First-order phase transition in large entropy lattice models. Com-
mun. Math. Phys. 83 (1982), 493-515.
[10] Krasnosel'ski, M.A.: Positive solutions of opertor equations. (Gos. Izd. Moscow, 1969) (Russian)
[11] Krasnosel'ski, M.A. and Zabrejko P.P.: Geometric methods of nonlinear analysis (Nauka. Moscow,
1975) (Russian)
[12] Mossel, E.: Survey: information flow on trees. Graphs, morphisms and statistical physics, 155-170,
DIMACS Ser. Discrete Math. Theor. Comput. Sci. 63. (AMS Providence, RI, 2004)
ON GIBBS MEASURE OF A MODEL
13
[13] Nirenberg, L.: Topics in nonlinear functional analysis (AMS, Courant Lec. Notes in Math, 6, N.Y.
2001).
[14] Preston, C.: Gibbs states on countable sets (Cambridge University Press, London 1974).
[15] Rozikov, U.A. Partition structures of the Cayley tree and applications for describing periodic Gibbs
distributions. Theor. and Math. Phys. 112 (1997), 929-933.
[16] Rozikov, U.A. Description of periodic Gibbs measures of the Ising model on the Cayley tree. Russian
Math. Surv. 56 (2001), 172-173.
[17] Rozikov, U.A. and Eshkabilov, Yu.Kh.: On models with uncountable set of spin values on a Cayley
tree: Integral equations. Math. Phys. Anal. Geom. 13 (2010), 275-286.
[18] Sinai,Ya.G.: Theory of phase transitions: Rigorous Results (Pergamon, Oxford, 1982).
[19] Spitzer, F.: Markov random fields on an infinite tree, Ann. Prob. 3 (1975), 387 -- 398.
[20] Suhov, Y.M. and Rozikov, U.A.: A hard - core model on a Cayley tree: an example of a loss network,
Queueing Syst. 46 (2004), 197 -- 212.
[21] Yosida, K.: Functional analysis (Springer-Verlag, 1965)
[22] Zachary, S.: Countable state space Markov random fields and Markov chains on trees, Ann. Prob.
11 (1983), 894 -- 903.
Yu. Kh. Eshkabilov, National University of Uzbekistan, Tashkent, Uzbekistan.
E-mail address: [email protected]
F. H. Haydarov, National University of Uzbekistan, Tashkent, Uzbekistan.
U. A. Rozikov, Institute of mathematics and information technologies, Tashkent, Uzbek-
istan.
E-mail address: [email protected]
|
1311.2209 | 1 | 1311 | 2013-11-09T20:37:57 | Spectral measures associated with the factorization of the Lebesgue measure on a set via convolution | [
"math.FA"
] | Let $Q$ be a fundamental domain of some full-rank lattice in ${\Bbb R}^d$ and let $\mu$ and $\nu$ be two positive Borel measures on ${\Bbb R}^d$ such that the convolution $\mu\ast\nu$ is a multiple of $\chi_Q$. We consider the problem as to whether or not both measures must be spectral (i.e. each of their respective associated $L^2$ space admits an orthogonal basis of exponentials) and we show that this is the case when $Q = [0,1]^d$. This theorem yields a large class of examples of spectral measures which are either absolutely continuous, singularly continuous or purely discrete spectral measures. In addition, we propose a generalized Fuglede's conjecture for spectral measures on ${\Bbb R}^1$ and we show that it implies the classical Fuglede's conjecture on ${\Bbb R}^1$. | math.FA | math |
SPECTRAL MEASURES ASSOCIATED WITH THE
FACTORIZATION OF THE LEBESGUE MEASURE ON A SET
VIA CONVOLUTION
JEAN-PIERRE GABARDO AND CHUN-KIT LAI
Abstract. Let Q be a fundamental domain of some full-rank lattice in Rd and let
µ and ν be two positive Borel measures on Rd such that the convolution µ ∗ ν is a
multiple of χQ. We consider the problem as to whether or not both measures must
be spectral (i.e. each of their respective associated L2 space admits an orthogonal
basis of exponentials) and we show that this is the case when Q = [0, 1]d. This
theorem yields a large class of examples of spectral measures which are either
absolutely continuous, singularly continuous or purely discrete spectral measures.
In addition, we propose a generalized Fuglede's conjecture for spectral measures
on R1 and we show that it implies the classical Fuglede's conjecture on R1.
1. Introduction
Let µ be a compactly supported Borel probability measure on Rd. We say that
µ is a spectral measure if there exists a countable set Λ ⊂ Rd called spectrum such
that E(Λ) := {e2πihλ,xi : λ ∈ Λ} is an orthonormal basis for L2(µ).
If Ω ⊂ Rd
is measurable with finite positive Lebesgue measure and dµ(x) = χΩ(x)dx is a
spectral measure, then we say that Ω is a spectral set. Spectral sets were first intro-
duced by Fuglede ([Fu]) and have a very delicate and mysterious relationship with
translational tiling because of the spectral set conjecture (known also as Fuglede's
conjecture) proposed by Fuglede.
Conjecture (Fuglede's Conjecture): A bounded measurable set Ω on Rd of pos-
itive Lebesgue measure is a spectral set if and only if Ω is a translational tile.
We say that Ω is a translational tile if there exists a discrete set J such that
St∈J (Ω + t) = Rd, and the Lebesgue measure of (Ω + t) ∩ (Ω + t′) is zero for
any distinct t and t′ in J . Although this conjecture was eventually disproved in
dimension d ≥ 3 ([T, KM1, KM2]), most of the known examples of spectral sets
are constructed from translational tiles. An important class of examples of spectral
sets constructed in [PW] consists of sets of the form A + [0, 1] tiling [0, N] for some
2010 Mathematics Subject Classification. Primary 42B05, 42A85, 28A25.
Key words and phrases. Convolutions, Fuglede's conjecture, Lebesgue measures, Spectral mea-
sures, Spectra.
The first named author was supported by an NSERC grant.
1
N, where A ⊂ Z. In fact, in this case, the corresponding equally weighted discrete
measure on A is a spectral measure.
The first singular spectral measure was constructed by Jorgensen and Pedersen
[JP]. They showed that the standard Cantor measures are spectral measures if the
contraction is 1
2n , while there are at most two orthogonal exponentials when the
1
contraction is
2n+1. Following this discovery, more spectral self-similar/self-affine
measures were also found ([S], [ LaW], [DJ]).
In these investigations, the tiling
conditions on the digit sets play an important role. An interesting question arises
naturally:
Question: What kind of measures are spectral measures and how are they related
to translational tilings?
This question seems to be out of reach using our current knowledge. In this paper,
we aim to describe a unifying framework bridging the gap between singular spectral
measures and spectral sets. Let us introduce some simple notations. Denote by L
the Lebesgue measure in Rd and by LE the normalized Lebesgue measure restricted
to the measurable set E (i.e. LE(F ) = L(E ∩ F )/L(E)). For a finite set A, we
denote by A the cardinality of A and by δA the measurePa∈A δa, where δa is the
Dirac mass at a. We also write A ⊕ B = C if every element in C can be uniquely
expressed as a sum a + b with a ∈ A and b ∈ B. We now make some observations
about specific examples of spectral measures known in the literature.
(1) According to [PW], if A ⊂ Z and the set Ω = A + [0, 1) tiles [0, N), then Ω is
a spectral set. We can thus find a set B such that A ⊕ B = {0, 1, ..., N − 1}. This
means that(cid:16) 1
BδB(cid:17) ∗ LΩ = L[0,N ].
(2) Let µ be the standard one-fourth Cantor (probability) measure defined by the
self-similar identity
µ(·) =
1
2
µ(4·) +
1
2
µ(4 · −2).
It is known that µ is a spectral measure [JP]. At the same time, we observe that if
we define ν to be the one-fourth Cantor measure obeying the equation
ν(·) =
1
2
ν(4·) +
1
2
ν(4 · −1),
then µ ∗ ν = L[0,1]. This can be seen directly by computing the Fourier transform of
both measures.
In fact, we may view the operation of convolution with a positive measure as
certain kind of generalized translation. The above examples suggest the following
question. Let Q be a fundamental domain of some full-rank lattice on Rd.
2
F (Q): Any positive Borel measures µ and ν such that µ ∗ ν = LQ are spectral
measures.
Unfortunately, we cannot expect the above statement to be true for all Q.
In
fact, if µ = LE with E is the translational tile without a spectrum constructed
in [KM1], then µ ∗ ν = LQ for some fundamental domain Q as seen directly from
the construction of this counterexample. However, in order to understand which
measures are spectral, it is useful to know to what extent the statement F (Q) is
true for some specific Q. Our first main result unifies the examples of discrete
spectral measures, spectral sets and the singular spectral measures given in (1) and
(2) above.
Theorem 1.1. For any d ≥ 1, the statement F ([0, 1]d) is true. Moreover, for any
positive Borel measures µ and ν such that µ ∗ ν = L[0,1]d, we can find spectra Λµ and
Λν for µ and ν respectively satisfying the property that
Λµ ⊕ Λν = Z.
We now give a brief explanation of the proof of Theorem 1.1. We first focus on R1
where the proof involves two main steps. The first step is a complete characterization
of the Borel probability measures µ and ν satisfying the identity µ ∗ ν = L[0,1]. This
characterization is actually a known result in probability due to Lewis [Le].
In
particular, Lewis proved that only two cases could occur: either one measure is
absolutely continuous and the other one is purely discrete or they are both singular.
To prove our theorem, we will express the measures µ and ν as weak limits of
convolutions of some discrete measures using the result of Lewis (See Section 2).
The second step is to construct spectra for µ and ν. This is done by observing that
the discrete measures obtained at each level are spectral measures. We then show
that the spectral property carries over by passing to the weak limit. This argument
is a generalization of the proof in [DHL] (See Section 3). After the dimension one
case is established, we characterize the Borel probability measures µ and ν satisfying
µ ∗ ν = L[0,1]d as Cartesian products of one-dimensional Borel probability measures
σi and τi, i = 1, ..., d, on R1 satisfying σi ∗ τi = L[0,1] and also prove the spectral
property for those (See Section 5).
It is very unclear whether F (Q) is true if Q is not a hypercube. We will focus
our attention on R1 in which Fuglede's conjecture remains open. We propose the
following generalized Fuglede's conjecture for spectral measures on R1 and it is direct
to see that a full generality of F (Q) on R1 will imply one direction of this generalized
conjecture.
Conjecture (Generalized Fuglede's Conjecture): A compactly supported Borel
probability measure µ on R1 is spectral if and only if there exists a Borel probability
measure ν and a fundamental domain Q of some lattice on R1 such that µ ∗ ν = LQ.
3
This is an open conjecture on R1 and we will prove that it extends the classical
Fuglede's conjecture.
Theorem 1.2. The generalized Fuglede's conjecture implies Fuglede's conjecture on
R1.
Let us make some remarks on the classical Fuglede's conjecture on R1. There
is some evidence that the conjecture may be true on R1. In particular, the known
fact that all tiling sets of a tile and all spectra of a spectral set are periodic offers
some credibility to the conjecture [LW1, IK]. Moreover, some algebraic conditions,
if satisfied, are sufficient to settle the conjecture on R, although these conditions are
not easy to check [DL2].
As our focus is the one-dimensional case, we organize our paper as follows: In
Section 2, we describe the factorization of the Lebesgue measure on [0, 1] given by
Lewis and, for the reader's convenience, we provide a somewhat different proof of
the factorization theorem that avoids some of the complications of the original ones
stemming from the use of probabilistic tools. We then prove the spectral property
in Section 3 and discuss the generalized Fuglede's conjecture on R1 in Section 4.
We will finally prove Theorem 1.1 in higher dimension in Section 5. As this piece
of work offers us several new directions for further research, we end this paper with
some remarks and open question in Section 6.
Note: During the preparation of the manuscript, we were made aware that Pro-
fessor Xinggang He and his student [AH] discovered independently a new class of
one-dimensional spectral measures obtained via a Moran construction of fractals.
These one-dimensional spectral measures turn out to coincide exactly with those we
consider in this paper.
2. Factorization of Lebesgue measures
Let L[0,1] be the Lebesgue measure supported on [0, 1] and let µ and ν be two Borel
probability measures supported on [0, 1]. We say that (µ, ν) is a complementary pair
of measures with respect to L[0,1] if
µ ∗ ν = L[0,1].
Let N = {Nk}∞
associate with N the discrete measures
k=1 be a sequence of positive integers greater than or equal to 2. We
νk =
1
Nk
Nk−1Xj=0
δ
j
N1 ···Nk
,
k ≥ 1.
4
(2.1)
For a given Borel set E, recall that LE is the normalized Lebesgue measure supported
on E. We now observe that the Lebesgue measure supported on [0, 1] admits a
natural decomposition as convolution products.
L[0,1) =ν1 ∗ (L[0, 1
N1
])
=ν1 ∗ ν2 ∗ (L[0,
])
1
N1N2
= · · ·
=ν1 ∗ ν2 ∗ · · · ∗ νk ∗ (L[0,
1
N1 ···Nk
]).
The sequence of measures ν1 ∗ ν2 ∗ · · · ∗ νk converges weakly to L[0,1]. Therefore, one
can write the Lebesgue measure as an infinite convolution of discrete measures.
L[0,1] = ν1 ∗ ν2 ∗ · · · .
(2.2)
Given a set N as above, we will consider two types of factorization (Type I and
Type II) of L[0,1] as the convolution of two measures obtained from the infinite
factorization obtained in (2.2).
Type I. There exists a finite positive integer k such that we have either
µN = ν1 ∗ ν3 ∗ ... ∗ ν2k−1 and νN = ν2 ∗ ν4 ∗ ... ∗ ν2k ∗ (L[0,
1
N1N2 ···N2k
])
or
µN = ν1 ∗ ν3 ∗ ... ∗ ν2k−1 ∗ (L[0,
1
N1N2 ···N2k
]) and νN = ν2 ∗ ν4 ∗ ... ∗ ν2k.
Type II
µN = ν1 ∗ ν3 ∗ · · · ∗ ν2k−1 ∗ · · ·
νN = ν2 ∗ ν4 ∗ · · · ∗ ν2k · · ·
(2.3)
(2.4)
Remark 2.1. The reader might want to construct more general decompositions ob-
tained by choosing other factorizations of (2.2), but note that if convolution product
of two consecutive factors of (2.2) belong to the same factor in the factorization, say
νk and νk+1, then we have
νk ∗ νk+1 =
1
NkNk+1
NkNk+1Xj=0
δj/N1N2...(NkNk+1)
and we would then be able to write the given convolution product as one of type I
or type II associated with a different N .
5
Note in both cases that µN ∗ νN = L[0,1] by (2.2). Therefore, they are µN and
νN form a complementary pair with respect to L[0,1].
In the case of the Type I
decomposition, one is purely discrete and one is absolutely continuous while in the
Type II decomposition, both factors are singularly continuous measures. We say
that a complementary pair (µ, ν) is natural if we can find a sequence N of positive
integers such that (µ, ν) = (µN , νN ).
Theorem 2.2. If µ and ν are positive Borel probability measures supported on [0, 1]
and µ ∗ ν = L[0,1], then µ and ν are natural complementary pair.
This theorem is essentially due to Lewis [Le] who considered the problem in prob-
ability consisting in characterizing the type of the distributions of pairs of indepen-
dent random variables X and Y whose sum X + Y is a uniform random variable on
[−π, π]. For the reader's convenience, we will give here another proof based on his
ideas as his result is not widely known. Moreover, the proof we give here is more
analytical in flavor and avoids some of the complications arising in the original proof
from the use of probability tools. The main important step of the proof is to show
that if two probablity measures µ and ν satisfy µ ∗ ν = L[0,1], then one of them, say
µ, must be "1/N periodic" in the sense that µ =(cid:16)1/NPN −1
integer N ≥ 2 and µ1 ∗ ν = L[0,1/N ]. This is done by analyzing the structure of the
zeros of the Fourier transform of µ and ν (Lemma 2.5).
j=0 δj/N(cid:17) ∗ µ1 for some
We now define the (complex) Fourier transform of a compactly supported proba-
bility measure µ by the formula
bµ(ξ) =Z e−2πiξxdµ(x), ξ ∈ C.
We will consider convolution products yielding the Lebesgue measure supported on
[−1/2, 1/2] instead of [0, 1] to exploit some symmetric properties of the solutions (as
explained below). Note that µ ∗ ν = L[−1/2,1/2] is equivalent to
.
(2.5)
sin πξ
πξ
bµ(ξ)bν(ξ) = \L[−1/2,1/2](ξ) =
Z(bµ) = {ξ ∈ C :bµ(ξ) = 0}
The zero set of the Fourier transformbµ in the complex plane will be denoted by
Since ((δx ∗ µ) ∗ (δ−x ∗ ν) = L[−1/2,1/2] for any real numbers x, we may assume the
smallest closed interval containing the support of µ is given by [−a, a]. Denote by
supp µ the closed support of µ. Given a probability measure ρ, we also define the
measure ρ to be the measure satisfying ρ(B) = ρ(−B) for any Borel set B ⊂ R.
Lemma 2.3. Let µ and ν be two probability measures such that µ ∗ ν = L[−1/2,1/2]
and assume that the smallest closed interval containing supp µ is of the form [−a, a],
6
a > 0. Then we have
Z \ {0} = Z(bµ) ∪ Z(bν) (as a disjoint union).
Moreover, the smallest closed interval containing supp ν is given by [−b, b] where
b = 1/2 − a and both µ and ν have symmetric distributions around the origin (i.e.
µ = µ and ν = ν).
(2.6)
Proof. It is well-known that bµ is a non-zero entire analytic function, so its zero set
is a discrete set in the complex plane. Furthermore, since the zeros of \χ[−1/2,1/2] are
simple, (2.6) follows from (2.5). Let [c, b] be the smallest closed interval containing
the support of ν. Then a + b = 1/2 and −a + c = −1/2 showing that c = −b and
b = 1/2 − a.
Finally, note that, since µ is a positive measure, Z(cid:0)(µ)b(cid:1) = Z(bµ). Therefore,
Z(cid:0)(µ)b(cid:1) ∪ Z(bν) = Z \ {0}. Consider the tempered distribution ρ := µ ∗ ν ∗ δZ. Then
bρ = (µ)b·bν · δZ = δ0. Hence, ρ is the Lebesgue measure on R and the restriction of ρ
to the interval [−1/2, 1/2] is µ ∗ ν. This shows that µ ∗ ν = L[−1/2,1/2], which means
that µ ∗ ν = µ ∗ ν. Taking Fourier transform, we obtain µ = µ. The proof of the
symmetry of ν is similar.
(cid:3)
Note that Lewis used the Hadamard factorization theorem to prove the symmetry
property of µ and ν in Lemma 2.3. The ideas of the following two lemmas are due
to Lewis and form the crucial parts of the argument.
Lemma 2.4. Let r ≥ 1 be the smallest positive zero of bµ. Then
≤ a ≤
≤ b ≤
1
4r
1
2r
1
2r
1
4r
and
1
2
1
2
−
−
.
Proof. We just need to prove the lower estimates for both a and b as the upper ones
will follow from these and the fact that a + b = 1/2. Since r is a zero ofbµ, then −r
is also a zero and we must haveR cos(2πrx)dµ(x) = 0. This implies that 2πra ≥ π
4r . In particular, the claim is true for r = 1.
and thus a ≥ 1
2
For the upper bound, we consider the following functions for different r.
2 )Qk−1
cos(2πx),
cos(2πx) − cos(2π2x),
cos( πrx
(cos( π(r−1)x
h(x) :=
1..., r − 1. Hence R h(x)dν(x) = 0 as 1, · · · , r − 1 are zeros of bν. By checking the
By expanding h(x), we see that h(x) is a linear combination of cos(2πkx), for k =
sign of each factor, we see that if 2πx ≤ π(r − 1)/r, then h(x) ≥ 0.
j=1(cos(2πx) − cos 2(2j−1)π
) − cos( π(r+1)x
Consider the case where r > 2 is even. We have either 2πb ≥ π(r − 1)/r (i.e.
b ≥ 1/2 − 1/2r) or ν is supported on the atoms ±(1/r), · · · , ±(r − 3)/r. However,
r = 2;
r = 3;
r > 2, r = 2k;
j=1(cos(2πx) − cos (2π)(2j)
)Qk−2
),
2
2
r
r
), r > 2, r = 2k − 1.
7
ν cannot be supported on those atoms sincebν would be a polynomial in cos(2πx/r)
of degree at most r − 3, but there are r − 1 zeros forbν, a contradiction. Therefore,
we must have b ≥ 1/2 − 1/(2r). The proof for the other cases follows from a similar
argument.
(cid:3)
Lemma 2.5. Let N > 0 be a positive integer and let µ and ν be two probability
measures on R such that µ ∗ ν = L[0,1/N ] with neither bµ norbν being identically one.
Suppose that N ∈ Z(bν) and let Nr with r > 1 be the smallest positive zero of bµ.
Then
Proof. By rescaling the measures by a factor of N, it is easy to see that it suffices to
consider the case N = 1. By translating the measure (i.e. µ∗(δ−1/2 ∗ν) = L[−1/2,1/2]),
it suffices to prove the lemma for the case µ ∗ ν = L[−1/2,1/2], where µ = µ and ν = ν.
Z(bµ) ⊂ NrZ.
Let ρ(E) = ν({0})δ0(E) + 2ν(E ∩ (0, 1/2]) and ρ(E) = ρ(−E) for E Borel. Then,
the fact that ν(E) = ν(−E) implies that ρ + ρ = 2ν. Therefore,
µ ∗ ρ + µ ∗ ρ = 2L[−1/2,1/2].
(2.7)
This implies, in particular, that µ ∗ ρ is absolutely continuous with respect to the
Lebesgue measure and we can let g(x) ≥ 0 be its density. Then g(−x) is the density
of (µ ∗ ρ) = µ ∗ ρ. By (2.7),
g(x) + g(−x) = 2, a.e.
As supp (µ ∗ ρ) (and hence supp g(−x)) is contained in [−1/2, a], g(x) = 2 on
[a, 1/2]. We may therefore write
g = 2χ[a,1/2] + gχ[−a,a] =2χ[a,1/2] + gχ[−a,0] + (2 − g(−x))χ[0,a]
=2χ[0,1/2] + (gχ[−a,0] − g(−x)χ[0,a]).
Note that 2χ[0,1/2] is the density of the measure L[0,1/2]. Taking Fourier transform,
we have
g(−x) sin(2πξx)dx
(2.8)
bµ(ξ)bρ(ξ) =bg(ξ) = \L[0,1/2](ξ) + 2iZ a
Suppose that r is even. Asbµ(r) = 0, we must have
0
g(−x) sin(2πrx)dx = 0.
Z a
0
Since a ≤ 1/2r by Lemma 2.4, we have sin(2πrx) ≥ 0 on [0, a] and thus g(−x) = 0
there. Thus, (2.8) implies that
Hence, Z(bµ) ⊂ 2Z.
bµ(ξ)bρ(ξ) = \L[0,1/2](ξ).
8
(2.9)
Writing r = 2nm where m is odd, we deduce from the above argument that
Z(bµ) ⊂ 2Z. Consider the measure µ1(E) = µ(E/2) and ρ1(E) = ρ(E/2) we have
bµ1(ξ) = bµ(2ξ) and bρ1(ξ) = bρ(2ξ). By (2.9), we have bµ1(ξ)bρ1(ξ) = [L[0,1](ξ) (i.e.
µ1 ∗ (δ−1/2 ∗ ρ1) = L[−1/2,1/2]). Moreover, Z(bµ1) = 1
2Z(bµ). In this case, the smallest
positive zero of bµ1 will be 2n−1m. Therefore, repeating the above argument, we have
Z(bµ) ⊂ 2nZ and the proof will be finished if we can prove our claim if r is odd.
Suppose now that r is odd. We consider the measures ν1(E) = ν(E ∩ [−a, b]) and
ν2(E) = ν(E ∩ [−b, −a)) (Here, it is more convenient not to normalize ν1 and ν2 as
probability measures). We have then ν = ν1 + ν2 and L[−1/2,1/2] = µ ∗ ν1 + µ ∗ ν2.
Let g1 and g2 be the density of µ ∗ ν1 and µ ∗ ν2 respectively. The above implies that
g1(x) + g2(x) = 1 a.e. on [−1/2, 1/2].
Note that the supp g1 is contained in [−2a, 1/2] and supp g2 is contained in [−1/2, 0].
It follows that g1 = 1 almost everywhere on [0, 1/2]. We may therefore write
g1 = χ[0,1/2] + g1χ[−2a,0].
Taking Fourier transforms and noting that bg1(ξ) =bµ(ξ)bν1(ξ), we obtain
As bµ(r) = 0, by substituting ξ = r and equating the imaginary parts, we have
bµ(ξ)bν1(ξ) = \χ[0,1/2](ξ) +Z 2a
g1(−x) sin(2πrx)dx.
g1(−x)e2πiξxdx.
0
(2.10)
1
πr
=Z 2a
0
By Lemma 2.4, 2a ≥ 1/2r and therefore,
1
πr
0
=Z 1/2r
≤Z 1/2r
≤Z 1/2r
0
0
g1(−x) sin(2πrx)dx +Z 2a
1/2r
g1(−x) sin(2πrx)dx
g1(−x) sin(2πrx)dx (as sin(2πrx) ≤ 0 on [1/2r, 2a])
sin(2πrx)dx =
1
πr
.
(as g1(−x) ≤ 1)
1/2r g1(−x) sin(2πrx)dx = 0,
which implies that g1(−x) = 0 on [1/2r, 2a]. Considering the real part of the equa-
Hence, we must have g1(−x) = 1 on [0, 1/2r] and R 2a
tion (2.10) and noting thatbµ(ξ) is real-valued (as µ = µ), we have
Since Z(bµ) ⊂ Z, the previous equation shows that in fact Z(bµ) ⊂ rZ, completing
2πξ(cid:18)sin πξ + sin
bµ(ξ)Re (bν2(ξ)) =
+Z 1/2r
r (cid:19) .
cos(2πξx)dx =
sin πξ
2πξ
the proof.
πξ
(cid:3)
1
0
9
Proof of Theorem 2.2. Let (µ, ν) be a complementary pair with respect to L[0,1].
Consider the periodization of the measure µ defined by µp = µ ∗ δZ. Its distribu-
We may assume thatbν(1) 6= 0 and we let N1 > 1 be the smallest positive zero ofbν.
We have Z(bν) ⊂ N1Z by Lemma 2.5. As the zero sets of bµ and bν are disjoint (see
(2.6)), the set {k ∈ Z :bµ(k) 6= 0} is contained in N1Z.
bµp =bµ · δZ =bµ · δN1Z
Hence, µp is indeed 1/N1-periodic. It follows immediately that
tional Fourier transform (as a tempered distribution) is given by
µ = ν1 ∗ α1
and ν ∗ α1 = L[0,1/N1]
(2.11)
N1PN1−1
where ν1 = 1
j=0 δj/N1 and α1(E) = N1µ(E ∩ [0, 1/N1]) for any Borel set E.
The case where α1 is the Dirac measure at the origin immediately yields a type I
µ = ν1 ∗ α1,
decomposition. Otherwise, we apply Lemma 2.5 on the pair (ν, α1). Sincebν(N1) = 0,
we have cα1(N1) 6= 0 and we can let N2 ne the smallest positive integer such that
cα1(N1N2) = 0. By Lemma 2.5, we have Z(cα1) ⊂ N1N2Z. We obtain
ν = ν2 ∗ α2 α1 ∗ α2 = L[0,1/N1N2]
where ν2 = 1
j=0 δj/N1N2. The case where α2 is a Dirac measure at the origin
yelds again a type I decomposition. Otherwise, we continue this inductive process
and define recursively the probability measures αk, k ≥ 1. If αk = δ0 for some k, the
process stops and we have arrived at a type I decomposition. If αk 6= δ0 for all k,
we have then expressed both measures µ and ν at the infinite convolution products
µ = ν1 ∗ ν3 ∗ . . . ,
ν = ν2 ∗ ν4 ∗ . . . , which yields a type II decomposition.
N2PN2−1
✷
Theorem 2.2 also gives us a new proof of classification of the set A and B such
that A ⊕ B = {0, ..., n − 1} which was proved in [Lo] and [PW] using a theorem of
De Bruijn.
Corollary 2.6. Let En = {0, 1, · · · , n − 1} and let A and B be two finite set of
integers such that A ⊕ B = {0, ..., n − 1}. Suppose that 1 ∈ A. Then there exist
integers N1, ..., N2k such that N1...N2k = n and
A = EN0 ⊕ N0N1EN2 ⊕ ... ⊕ N0N1...N2k−1E2k
B = N0EN1 ⊕ N0N1N2EN3 ⊕ ... ⊕ N0N1...N2k−2E2k−1.
Proof. As A ⊕ B = {0, ..., n − 1}, we have
n A(cid:19) ∗(cid:18) 1
By Theorem 2.2, the measures µ =(cid:16) 1
(cid:18) 1
A
B
δ 1
n B(cid:19) ∗ L[0,1/n] = L[0,1].
n A(cid:17) and ν = (cid:16) 1
Bδ 1
δ 1
A δ 1
10
ural complementary pair. As one of them is discrete and the other is absolutely
n B(cid:17) ∗ L[0,1/n] are nat-
continuous, they correspond to a type I decomposition. Since 1 ∈ A, we have thus
1/n ∈ 1
nA. By comparing the support of the measures, we obtain the existence of
integers N ′
1, N ′
2... such that
1
n
A =
1
N ′
1
EN ′
1 ⊕
1
2N ′
1N ′
3
N ′
EN ′
3 ⊕ ... ⊕
1
2...N ′
N ′
1N ′
2k−1
EN ′
2k−1
.
1
n
1...N ′
and n = N ′
1
1N ′
N ′
2
B =
EN ′
2 ⊕
N ′
2k. Letting Nr = N ′
EN ′
1
2N ′
1N ′
2k−r, we obtain the desired factorization.
1
2...N ′
1N ′
2k
4 ⊕ ... ⊕
3N ′
4
EN ′
N ′
2k
(cid:3)
3. The spectral property
In this section, we show that all measures appearing in natural complementary
pairs are spectral measures. Recall that a Borel probability measure µ is called a
spectral measure with associated spectrum Λ if the collection of exponentials E(Λ) =
{e2πiλx}λ∈Λ forms an orthonormal basis for L2(µ). It is easy to see that E(Λ) is an
orthonormal set in L2(µ) if and only if
By a well-known result in [JP], Λ is a spectrum of µ if and only if
(3.1)
Λ − Λ ⊂ Z(bµ) ∪ {0}.
Q(ξ) :=Xλ∈Λ
bµ(ξ + λ)2 ≡ 1.
In fact, if E(Λ) is an orthonormal set, Q(ξ) ≤ 1 and Q is an entire function of
exponential type ([JP], see also [DHL]). Let N = {Ni}∞
i=1 be a collection of positive
integers and consider the Type I and II decomposition as in the previous section.
Let
µ(k) = ν1 ∗ ν3 ∗ · · · ∗ ν2k−1, ν(k) = ν2 ∗ ν4 ∗ · · · ∗ ν2k
and for a given N , we let A1 = {0, .., N1 − 1} and An = N1 · · · Nn−1 · {0, .., Nn − 1}
for n ≥ 2. We start with a simple observation.
Proposition 3.1. Each νn is a spectral measure with spectrum An. For all k ≥ 1,
µ(k) is a spectral measure with spectrum given by
Λk =
kMj=1
A2j−1
(3.2)
In particular, the type I natural complementary pair µN and νN defined in the pre-
vious section are spectral measures.
11
j=0 δj/Nn is a spectral measure
j=0 δj/(N1···Nn) is a spectral
Proof. It is immediate to see that the measure 1
with spectrum {0, .., Nn − 1}. Therefore, νn = 1
measure with spectrum N1 · · · Nn−1 · {0, .., Nn − 1} = An.
NnPNn−1
NnPNn−1
Note that Z(bνn) = N1N2...NnZ \ N1N2...Nn−1Z and
[ν2j−1(ξ).
dµ(k)(ξ) =
kYj=1
For notational convenience, we define N0 = 1. Taking distinct λ1, λ2 ∈ Λk and
j=1 rℓ,jN1N2...N2j−2, for ℓ = 1, 2, we have
writing λℓ =Pk
λ1 − λ2 =
(r1,j − r2,j)N1N2 · · · N2j−2 =
kXj=1
sjN1N2 · · · N2j−2,
kXj=J
where J is the first index such that r1,j 6= r2,j and −(N2J−1 − 1) ≤ sJ ≤ N2J−1 − 1.
so [ν2J−1(λ1 − λ2) = [ν2J−1(N1 · · · N2J−2sJ ) = 0. Therefore, dµ(k)(λ1 − λ2) = 0. This
proves the orthogonality of E(Λk) in L2(µ(k)). As L2(µ(k)) is a finite dimensional
vector space of dimension N1N3 · · · N2k−1 = card (E(Λ)), the collection E(Λ) must
be complete in L2(µ(k)).
To prove the last statement, we just consider the case where µN = µ(k) and
νN = ν(k) ∗ (L[0,
]) and
νN = ν2 ∗ ν4 ∗ ... ∗ ν2k is similar. It is easily seen, as before, that ν(k) is also a discrete
spectral measure with spectrum
]), as the case µN = ν1 ∗ ν3 ∗ ... ∗ ν2k−1 ∗ (L[0,
N1N2 ···N2k
1
N1 ···N2k
1
A2j.
kMj=1
fΛk =
α has N1N2...N2kZ as a spectrum. It follows that
Moreover, dν(k) is N1...N2k-periodic. Let α denote the measure L[0,
Xλ∈ fΛk+N1···N2kZ
cνN (ξ + λ)2 = Xλ∈ fΛk,m∈Z
=Xλ∈ fΛk
dν(k)(ξ + λ + N1...N2km)2bα(ξ + λ + N1...N2km)2
dν(k)(ξ + λ)2 ·Xm∈Z
bα(ξ + λ + N1...N2km)2 ≡ 1.
1
N1N2 ···N2k
]. Then
Hence, νN a spectral measure with spectrum fΛk + N1 · · · N2kZ.
It remains to deal with the spectral property for complementary pairs µN and νN
of type II. Since these two measures have essentially the same form, we will discuss
(cid:3)
12
only the case µ := µN . Note that the measure µ will be the weak limit of the
measures µ(k) and
bµ(ξ) =
∞Yj=1
[ν2j−1(ξ) =dµ(k)(ξ) ·
N2j−1PN2j−1−1
r=0
δ
∞Yj=k+1
r
N1 ···N2j−1
Here we recall that ν2j−1 = 1
given by
[ν2j−1(ξ).
(3.3)
and its Fourier transform is
[ν2j−1(ξ) = e−πi(N2j−1−1)ξ/(N1···N2j−1)
sin(πξ/(N1 · · · N2j−2))
N2j−1 sin(πξ/(N1 · · · N2j−1))
.
(3.4)
Let
Λµ =
∞Mj=1
A2j−1 =
Λk
∞[k=1
(Only finite sums of elements of A2j−1, j ≥ 1, appear in Λµ). The exponentials
{e2πiλx}λ∈Λµ are mutually orthogonal in L2(µ) by Proposition 3.1. Our goal is verify
(3.1). To do this, we note that, as Q is an entire function, we just need to show that
Q(ξ) ≡ 1 on a neighborhood of 0. Let
Qk(ξ) = Xλ∈Λk
bµ(ξ + λ)2.
Now, we fix two positive integers n and p. By (3.3) and the fact that {Λk}k≥1 is an
increasing sequence of sets,
Qn+p(ξ) =Qn(ξ) + Xλ∈Λn+p\Λn
=Qn(ξ) + Xλ∈Λn+p\Λn
bµ(ξ + λ)2
\µ(n+p)(ξ + λ)2 ·(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
We need the following proposition which provides a crucial estimate for the last term
in the previous expression in order to establish the spectral property.
∞Yj=n+p+1
2
[ν2j−1(ξ + λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
.
(3.5)
Proposition 3.2. There exists c > 0 such that
inf
k≥1
2
inf
λ∈Λk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∞Yj=k+1
[ν2j−1(ξ + λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for all ξ < 1/2, where Λk is given in (3.2).
≥ c
Proof. Let λ ∈ Λk and xk,λ = ξ+λ
N1N2···Nk
. We first note that, by (3.4),
13
2
[ν2j−1(ξ + λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∞Yj=k+1
Writing λ =Pk
=
=
∞Yj=k+1
∞Yj=k+1
sin2(π(ξ + λ)/(N1 · · · N2j−2))
N 2
2j−1 sin2((π(ξ + λ))/(N1 · · · N2j−1))
sin2(πx2j−2,λ)
2j−1 sin2(πx2j−1,λ)
N 2
.
(3.6)
that λ ≤ N1 · · · N2k−1 − 1. Hence, we have
j=1 rjN1N2...N2j−2 with 0 ≤ rj ≤ N2j−1 − 1, we see immediately
λ
N1 · · · N2k
≤
N1 · · · N2k−1 − 1
N1 · · · N2k
≤
1
N2k
≤
1
2
.
Therefore, for all ξ < 1/2, we have
C := sup
k≥1
sup
λ∈Λk
x2k,λ = sup
k≥1
sup
λ∈Λk
ξ + λ
N1 · · · N2k
<
3
4
as all Nj ≥ 2. Note that Nkxk,λ = xk−1,λ and using two elementary inequalities
sin x ≤ x and sin x ≥ x − x3
3! , we have the following estimation for the product in
(3.6),
∞Yj=k+1
sin2(πx2j−2,λ)
2j−1 sin2(πx2j−1,λ)
N 2
≥
=
≥
=
∞Yj=k+1(cid:18)1 −
∞Yj=k+1 1 −
∞Yj=k+1 1 −
∞Yj=1 1 −
π2
π2
3π2
x2
x2k,λ
π2
6
2j−2,λ(cid:19)2
N2k+1...N2j−2(cid:19)2!2
6 (cid:18)
22(j−k)−2(cid:19)2!2
6 (cid:18)
22j−2(cid:19)2!2
32 (cid:18) 1
:= c.
C
As P∞
complete.
j=1 1/22j−2 < ∞ and all factors are positive, c > 0 and hence the proof is
(cid:3)
Proof of Theorem 1.1 on R1. In view of Theorem 2.2, we just need to show that all
natural complementary pairs are spectral measures. Let N be a sequence of positive
integers greater than or equal to 2. If the pair is of Type I, then Proposition 3.1
shows that both factors are spectral measures.
It remains to consider the Type II case. Let µN and νN be defined in (2.3) and
(2.4). As mentioned before, we only need to prove that µ = µN is a spectral measure.
14
Let c be the positive number determined in Proposition 3.2. By Proposition 3.1 and
(3.1), we have
Using this fact and Proposition 3.2, we obtain from (3.5) that
\µ(n+p)(ξ + λ)2 = 1 − Xλ∈Λn
Xλ∈Λn+p\Λn
Qn+p(ξ) ≥ Qn(ξ) + c · 1 − Xλ∈Λn
\µ(n+p)(ξ + λ)2.
\µ(n+p)(ξ + λ)2! .
Fixing n and letting p go to infinity, it follows that
Q(ξ) ≥ Qn(ξ) + c(1 − Xλ∈Λn
bµ(ξ + λ)2) = Qn(ξ) + c(1 − Qn(ξ)).
Finally, taking n to infinity, we obtain that c(1 −Q(ξ)) ≤ 0. But c > 0 and Q(ξ) ≤ 1
because {e2πiλx}λ∈Λ is an orthogonal set in L2(µ). This show that Q(ξ) = 1 for
ξ ≤ 1/2 and thus for all ξ ∈ R by analyticity, completing the proof.
We now establish the tiling property of the spectra. Suppose that we are given
a type I decomposition. Then Proposition 3.1 implies that µN and νN have the
following spectra:
Λµ =
kMj=1
A2j−1, Λν =
k−1Mj=1
A2j ⊕ N1 · · · N2k−1Z.
It can be seen immediately that Λµ ⊕ Λν = {0, 1, · · · , N2k−1 − 1} ⊕ N2k−1Z = Z.
Suppose now the decomposition is of type II. Note that the complementary mea-
sures have the following spectra using the above notations.
Λµ =
∞Mj=1
A2j−1, Λν =
A2j
∞Mj=1
Note that −Λν is also spectrum of ν. We now claim that Λµ ⊕ (−Λν) = Z. Observe
that
A1 ⊕ (−A2) = {−N1N2 + N1, .., N1 − 1}.
A1 ⊕ (−A2) ⊕ A3 = {−N1N2 + N1, .., N1N2N3 − N1N2 + N1 − 1}.
Inductively, the sets A1 ⊕ (−A2) ⊕ ... ⊕ (−1)k−1Ak cover an increasing sequence of
consecutive integers. showing that Λµ ⊕ (−Λν) = Z. This proves our claim.
(cid:3)
15
4. Generalized Fuglede's conjecture
In this section, we will formulate a generalization of Fuglede's conjecture and
prove that it implies the original one. Recall the conjecture we are interested in:
Conjecture (Generalized Fuglede's Conjecture): A compactly supported Borel
probability measure µ on R1 is spectral if and only if there exists a Borel probability
measure ν and a fundamental domain Q of some lattice on R1 such that µ ∗ ν = LQ.
We first prove the following proposition.
Proposition 4.1. Let Ω and Q be bounded measurable sets of positive Lebesgue
measure on R1. Suppose that LΩ ∗ ν = LQ, for some Borel probability measure ν.
Then
ν =
NXk=1
1
N
δak , Q =
(Ω + ak)
N[k=1
and L((Ω + ak) ∩ (Ω + aℓ)) = 0 for all k 6= ℓ.
Proof. We first note that LΩ ∗ν = LQ if and only if (LΩ ∗δy)∗(ν ∗δx ∗δ−y) = (LQ ∗δx)
for any real numbers x and y. Therefore, there is no loss of generality to assume
that the smallest closed intervals containing Ω and Q are respectively [0, a] and [0, b].
As Q = supp (LΩ ∗ ν) = Ω + supp ν, The support of ν has to be contained in the
non-negative part of the real line.
Let ǫ > 0 and consider the interval Eǫ = [0, ǫ). Let ηǫ ∈ Eǫ be a Lebesgue point
of χQ. Then, using LΩ ∗ ν = LQ,
1
L(Q)
L (Q ∩ [ηǫ, ηǫ + h)) =
=
1
L(Ω)Z ηǫ+h
L(Ω)Z ηǫ+h
1
0
0
L (Ω ∩ ([ηǫ, ηǫ + h) − y)) dν(y)
L ((Ω + y) ∩ [ηǫ, ηǫ + h)) dν(y),
since Ω and supp ν are contained in [0, ∞). This implies that
L(Ω)
L(Q)
L(Q ∩ [ηǫ, ηǫ + h)) ≤ L([ηǫ, ηǫ + h))ν([0, ηǫ + h)) = hν([0, ηǫ + h)).
Since ηǫ is a Lebesgue point of χQ, we have limh→0
taking h → 0, we deduce that L(Ω)
the inequality
= 1. Therefore, by
L(Q) ≤ ν([0, ηǫ]). Letting ǫ approach zero, we obtain
h
L(Q∩[ηǫ,ηǫ+h))
L(Ω)
L(Q)
≤ ν({0}).
(4.1)
Since L(Ω) > 0, ν has an atom at 0 and we can write
ν = p0δ0 + (1 − p0)ν1, p0 = ν({0}) and ν1({0}) = 0.
(4.2)
16
The equation LΩ ∗ ν = LQ can thus be rewritten as
Since the left hand side of (4.3) is still a positive measure, this implies that
(1 − p0)LΩ ∗ ν1 = LQ − p0LΩ.
(4.3)
L(Ω)
L(Q)
Combining it with (4.1), we conclude that p0 = L(Ω)
0 ≤ (LQ − p0LΩ)(Ω) ≤
− p0.
L(Q) and, using (4.3), we obtain
LΩ ∗ ν1 = LQ\Ω
If p0 = 1, then Q = Ω and ν = δ0, so we are done.
If not, we then repeat
the argument with Q replaced by Q \ Ω. We can find Ω + a1 ⊂ Q \ Ω such that
p1 := ν1({a1}) > 0 and ν1 = p1δa1 + (1 − p1)ν2. Moreover, p1 = L(Ω)/L(Q \ Ω). By
(4.2),
ν =
L(Ω)
L(Q)
(δ0 + δa1) + (1 − p1)ν2.
The theorem will be proved if p1 = 1. Otherwise, we continue this process to obtain
a maximal number N of measure disjoint translates of Ω, Ω + a1,..,Ω + aN −1 such
k=0 (Ω + ak). Since L(Ω) > 0 and L(Q) ≥ NL(Ω), N is the largest
integer such that L(Q) ≥ NL(Ω). We can then write
that Q ⊃ SN −1
ν =
L(Ω)
L(Q)(cid:0)δ0 + ... + δaN −1(cid:1) + (1 − pN −1)νN .
If pN −1 < 1, we could iterate this process to obtain one more disjoint translate of Ω
contained in Q, which is certainly impossible by this choice of N. Hence, pN −1 = 1.
As ν is a probability measure, we must have L(Ω)/L(Q) = 1/N. Therefore, the
proposition is proved.
(cid:3)
Theorem 4.2. The validity of generalized Fuglede's conjecture implies that of the
original Fuglede's conjecture on R1.
Proof. Suppose that Ω is a bounded spectral set, then LΩ is a spectral measure.
By the generalized Fuglede's conjecture, we can find a probability measure ν and a
fundamental domain Q of some lattice Γ such that
LΩ ∗ ν = LQ.
By Proposition 4.1, ν is a purely discrete measure that can be written as ν = 1
for some finite discrete subset A and
#AδA
As Q is a fundamental domain Q of the lattice Γ, Ω is a translational tile with tiling
set given by A + Γ.
Q = [a∈A
(Ω + a).
17
Conversely, suppose that Ω is a bounded translational tile with tiling set J . By
the result of Lagarias and Wang [LW1], all tiling sets on R1 are periodic. This implies
that we can find a finite set A ⊂ R and a lattice Γ such that J = A + Γ. This
means that the set Q = Ω + A is a fundamental domain of Γ. Letting ν = 1
#AδA,
LΩ ∗ ν = LQ. By the generalized Fuglede's conjecture, LΩ is a spectral measure and
Ω is a spectral set.
(cid:3)
5. The Higher Dimensional Case
Let µ1,...,µd be Borel probability measures on R1. The Cartesian product of these
measures is the unique Borel probability measure µ1 ⊗ ... ⊗ µd on Rd such that
(µ1 ⊗ ... ⊗ µd)(E1 × ... × Ed) =
µi(Ei),
dYi=1
for any Borel sets Ei, 1 ≤ i ≤ d, on R1. In this section, we characterize the measures
µ and ν on Rd which are solutions of the equation
as Cartesian products of the measures satisfying the corresponding one-dimensional
equation.
µ ∗ ν = L[0,1]d.
(5.1)
Theorem 5.1. Let µ and ν be compactly supported probability measures on Rd.
Then µ and ν are solutions to (5.1) if and only if there exists compactly supported
Borel probability measures {σi}d
i=1 on R1 such that
i=1 and {τi}d
µ = σ1 ⊗ ... ⊗ σd, ν = τ1 ⊗ ... ⊗ τd
(5.2)
and σi ∗ τi = L[0,1] for all i = 1, ..., d.
Note that the sufficiency part of the theorem follows by a direct computation.We
only need to establish the necessity part of the theorem. Denote by P the orthogonal
projection of the first coordinate on Rd and Q the orthogonal projection of the
corresponding orthogonal complement. If µ is a positive Borel measure on Rd, we
denote by µP −1 the positive Borel measure on R1 defined by µP −1(E) = µ(P −1(E))
for any Borel set E ⊂ R and the measure µQ−1 is similarly defined. We will need
the following lemmas.
Lemma 5.2. Let µ and ν be two probability measures on Rd. Then
(µ ∗ ν)P −1 = (µP −1) ∗ (νP −1), and (µ ∗ ν)Q−1 = (µQ−1) ∗ (νQ−1).
In particular, if µ and ν are two Borel probability measures satisfying (5.1), then we
have
(µP −1) ∗ (νP −1) = L[0,1) and (µQ−1) ∗ (νQ−1) = L[0,1]d−1.
18
Proof. The proof follows easily from the fact that
(µP −1)b(ξ) =bµ(ξ, 0, ..., 0), and (µQ−1)b(ξ2, ..., ξd) =bµ(0, ξ2, ..., ξd).
(cid:3)
Lemma 5.3. Let ν be a Borel probability measure on Rd. Then, there is at most
one probability measure µ on Rd satisfying µ ∗ ν = L[0,1]d.
Proof. If µ is as above, we have
(ξ), ξ ∈ Rd.
bµ(ξ)bν(ξ) =(cid:0)L[0,1]d(cid:1)b
Therefore, bµ(ξ) is thus determined on the set
Since F = Rd and bµ is continuous (as µ is compactly supported), bµ and thus µ is
F = {ξ ∈ Rd : ξi 6∈ Zd, i = 1, ..., d},
completely determined by (5.3).
(5.3)
(cid:3)
The previous lemma is also valid if [0, 1]d is replaced by a d-dimensional rectan-
gular box. Now, we proceed to the proof of Theorem 5.1.
Proof of Theorem 5.1. We prove the necessity part of the theorem by induction on
the dimension. The statement is proved when d = 1 in Theorem 2.2. Assuming that
the statement is true for d − 1, we now establish it on Rd.
Let µ and ν be two Borel probability measures satisfying µ∗ν = L[0,1]d. By Lemma
5.2 and Theorem 2.2 (see also equation (2.11)), we can find an integer N1 ≥ 2 such
that µP −1 and νP −1 can be decomposed (after possibly interchanging these two
measures) as
µP −1 = ν1 ∗ α1, and α1 ∗ (νP −1) = L[0,1/N1]
(5.4)
j=0 δj/N1 and α1(E) = N1(µP −1)(E ∩ [0, 1/N1)) for any Borel
where ν1 = 1/N1PN1−1
set E. Let CN1 be the d-dimensional rectangular box h0, 1
[0, 1]d \ CN1 =h 1
, 1i × [0, 1]d−1 and
N1
µ (CN1) = µP −1(cid:18)(cid:20)0,
1
N1(cid:19)(cid:19) =
1
N1
.
N1(cid:17) × [0, 1]d−1. Then
Hence, we can define two Borel probability measures on Rd, ρ1 and eρ1, satisfying
N1
N1 − 1
µ(cid:0)E ∩(cid:0)[0, 1]d \ CN1(cid:1)(cid:1)
ρ1(E) = N1µ (E ∩ CN1) , eρ1(E) =
19
for any Borel sets E. Then µ = 1
N1
ρ1 + (1 − 1
N1
supp ν ⊂ [0, 1]d, we have ν ∗eρ = 0 on the rectangular box CN1. Hence,
ρ1 ∗ ν = N1(µ ∗ ν) = LCN1
on CN1.
)eρ1. Since supp eρ ⊂ [0, 1]d \ CN1 and
We can thus write ρ1 ∗ ν = LCN1
as ρ1 ∗ ν and LCN1
are probability measures. Hence,
+ η where η is a positive measure. However, η = 0
(ν1 ⊗ δ0d−1) ∗ ρ1 ∗ ν = 1
N1
δ(j/N1,0...,0)! ∗ ρ1 ∗ ν = L[0,1]d
N1−1Xj=0
where 0d−1 = (0, ..., 0) ∈ Rd−1. By Lemma 5.3, we have that
µ = (ν1 ⊗ δ0d−1) ∗ ρ1, and ρ1 ∗ ν = L[0,1/N1]×[0,1]d−1
(5.5)
Furthermore, ρ1P −1 = α1 where α1 is defined in (5.4).
We now consider two cases depending on whether µP −1 and νP −1 correspond to
a type I or type II decomposition (as defined in Section 2).
Case 1 (Type I decomposition): Using the notations introduced in Section 2,
we have then, without loss of generality, that
µP −1 = ν1 ∗ ...ν2k−1, νP −1 = ν2 ∗ ...ν2k ∗ L[0,
1
N1 ...N2k
].
By the previous steps, the identities in (5.5) hold. A similar argument, shows the
existence of a probability measure ρ2 such that
ν = (ν2 ⊗ δ0d−1) ∗ ρ2 and ρ1 ∗ ρ2 = L[0,
]×[0,1]d−1.
1
N1N2
Continuing this procedure 2k-times, we deduce the existence of probability measures
ρ2k−1 and ρ2k such that
and
µ = ((ν1 ∗ ν3 ∗ ... ∗ ν2k−1) ⊗ δ0d−1) ∗ ρ2k−1
ν = ((ν2 ∗ ν4 ∗ ... ∗ ν2k) ⊗ δ0d−1) ∗ ρ2k
ρ2k−1 ∗ ρ2k = L[0,1/N1N2...N2k]×[0,1]d−1.
(5.6)
(5.7)
(5.8)
By (5.6) and Lemma 5.2, µP −1 = ν1∗...ν2k−1∗ρ2k−1P −1, showing that ρ2k−1P −1 = δ0.
Hence, we can write ρ2k−1 = δ0 ⊗ σ for some positive measure σ on Rd−1. Using
(5.8) and Lemma 5.2 again, we obtain that σ ∗ (ρ2kQ−1) = L[0,1]d−1. Hence,
ρ2k−1 ∗ ρ2k =L[0,1/N1N2...N2k] ⊗ L[0,1]d−1
=L[0,1/N1N2...N2k] ⊗ (σ ∗ (ρ2kQ−1))
=(δ0 ⊗ σ) ∗ (L[0,1/N1N2...N2k] ⊗ (ρ2kQ−1))
=ρ2k−1 ∗ (L[0,1/N1N2...N2k] ⊗ (ρ2kQ−1)).
20
Lemma 5.3 shows that ρ2k = L[0,1/N1N2...N2k] ⊗ (ρ2kQ−1) and (5.7) implies that
ν = νP −1 ⊗ ρ2kQ−1. Finally, applying the induction hypothesis to the identity
σ ∗ (ρ2kQ−1) = L[0,1]d−1, we can write σ = σ2 ⊗ ... ⊗ σd and ρ2k−1Q−1 = τ2 ⊗ ... ⊗ τd
with σi ∗ τi = L[0,1] and Theorem 5.1 for dimension d follows.
Case 2 (Type II decomposition). In this case, we can without loss of generality
assume that
µP −1 = ν1 ∗ ν3 ∗ ..., νP −1 = ν2 ∗ ν4 ∗ ...
and we still have (5.6), (5.7) and (5.8) for all k = 1, 2, .... with ρnP −1 6= δ0 for
any integer n. As ρn are all probability measures, we can assume, by passing to
subsequences if necessary, that the sequences {ρ2k−1} and {ρ2k} converge weakly to
some probability measures that we denote by σ and τ , respectively. From (5.8), it is
immediate to see that the supports of σ and τ are both contained in {0} × [0, 1]d−1.
We can write σ = δ0 ⊗ σ′ and τ = δ0 ⊗ τ ′. By passing to weak limit in (5.6) and
(5.7), we have
µ = (µP −1 ⊗ δ0d−1) ∗ (δ0 ⊗ σ′), ν = (νP −1 ⊗ δ0d−1) ∗ (δ0 ⊗ τ ′).
(5.9)
As µ ∗ ν = L[0,1]d and (µP −1
1 ⊗ δ0d−1) ∗ (νP −1
1 ⊗ δ0d−1) = L[0,1] ⊗ δ0d−1, we have
σ′ ∗ τ ′ = L[0,1]d−1,
The conclusion follows immediately by (5.9) using the induction hypothesis.
(cid:3)
Proof of Theorem 1.1 on Rd. The proof follows from the result on R1. By Theorem
5.1, we can write µ = σ1 ⊗ ... ⊗ σd and ν = τ1 ⊗ ... ⊗ τd with σi ∗ τi = L[0,1].
Therefore, our conclusion on R1 implies that σi and τi are spectral measures on R1
with spectrum Λσi and Λτi respectively. Moreover, they satisfies Λσi ⊕ Λτi = Z. Now
we define
Λµ =
dOi=1
Λσi, Λν =
Λτi,
dOi=1
i=1 Ai := {(a1, ..., ad) : ai ∈ Ai} for sets Ai ⊂ R1. We claim that Λµ is a
spectrum for µ (the proof that Λν is a spectrum for ν is similar).
where Nd
Note thatbµ(ξ) =Qd
Xλ∈Λµ
i=1bσi(ξi). From this, it follows easily that
bσi(ξi + λi)2 = 1.
bµ(ξ + λ)2 =
Xλi∈Λσi
dYi=1
Hence, Λµ is a spectrum for µ. That the tiling property of the spectra (i.e. Λµ⊕Λν =
Zd) follows immediately from the tiling property of Λσi and Λτi.
(cid:3)
21
6. Remarks and Open questions
As indicated in the introduction, the statement F (Q) is false in general. Nonethe-
less, this statement suggests many related questions that may help us understand
the relationship among convolutions, translational tilings and spectral measures.
Motivated by the generalized Fuglede's conjecture, one of the main questions we
would like to ask is:
(Q1): For which Q is the statement F (Q) true?
This question seems to be hard if we go beyond cubes as the methods of this paper
would be difficult to extend. An easier, but still interesting question concerns the
decomposition of the Lebesgue measure on sets as convolution product of singular
measures:
(Q2): For what kind of measurable (resp. spectral) sets Q can LQ be decomposed
into the convolution of two singularly continuous (resp. spectral) measure ?
One natural type of such sets will be the self-affine tiles [LW2]. These tiles can be
described as infinite convolution product of discrete measures and can therefore be
decomposed into two singular measures using methods similar to those in Section 2.
Fourier frames and exponential Riesz bases are natural generalization of expo-
nential orthonormal bases. It has been an interesting question to produce singular
measures with Fourier frames but not exponential orthonormal bases. By now we
only know we can produce such measures by considering measures which are abso-
lutely continuous with respect to a spectral measure with density bounded above
and away from 0 or convolving a spectral measure with some discrete measures
[HLL, DL1]. These methods are rather restrictive. As absolutely continuous (w.r.t.
Lebesgue) measures with Fourier frames were completely classified in [Lai], we ask
(Q3): Can we produce new singular measures admitting Fourier frames by decom-
posing an absolutely continuous (w.r.t. Lebesgue) measures with Fourier frames?
Conversely, is it true that all measures admitting Fourier frames are constructed in
this way?
Given a spectral measure µ, another important issue is to classify its spectrum.
This question has been studied for Lebesgue measures and some Cantor measures
in [LRW, DHS, DHL]. However, there is no satisfactory answer when the measure
is singular. The tiling statement of Theorem 1.1, suggests a possible answer.
(Q4): Let µ and ν be a natural complementary pair of L[0,1]. Let also Λµ be a
spectrum for L2(µ), does there exist a spectrum Λν for L2(ν) such that Λµ ⊕Λν = Z?
It is not difficult to prove that (Q4) actually holds for type I decompositions.
The remaining challenge is to answer the question for type II decompositions.
22
References
[AH] L.-X An, and X.-G He , A class of spectral Moran measures, preprint
[DHL] X.-R. Dai, X.-G He and C.-K Lai, Spectral property of Cantor measures with consecutive
digits, Adv. Math., 242 (2013), 187-208.
[DHS] D. Dutkay, D. Han and Q. Sun, On spectra of a Cantor measure, Adv. Math., 221
(2009), 251-276.
[DJ] D. Dutkay and P. Jorgensen, Fourier frequencies in affine iterated function systems,
J. Funct. Anal., 247 (2007), 110-137.
[DL1] D. Dutkay and C.-K. Lai, Uniformity of measures with Fourier
frames,
http://arxiv.org/abs/1202.6028, 2012.
[DL2] D. Dutkay and C.-K. Lai, Some reductions of the spectral set conjecture to integers, to
appear in Math Proc Cambridge Phil Soc.
[Fu] B. Fuglede, Commuting self-adjoint partial differential operators and a group theoretic
problem, J. Funct. Anal., 16 (1974), 101-121.
[JP] P. Jorgensen and S. Pedersen, Dense analytic subspaces in fractal L2 spaces, J. Anal.
Math., 75 (1998), 185-228.
[HLL] X.-G. He, C.-K. Lai and K.-S. Lau, Exponential spectra in L2(µ), Appl. Comput. Har-
mon. Anal., 34 (2013), 327-338.
[IK] A. Iosevich and M. Kolountzakis, Periodicity of the spectrum in dimension one. ,
http://arxiv.org/abs/1108.5689, 2012.
[KM1] M. Kolountzakis and M. Matolcsi, Tiles with no spectra, Forum Math., 18 (2006),
519-528.
[KM2] M. Kolountzakis and M. Matolcsi, Complex Hadamard matrices and the Spectral Set
Conjecture, Collect. Math., Extra (2006), 281-291.
[LW1] J. Lagarias and Y. Wang, Tiling the line by translates of one tile, Invent. Math. 124
(1996), 341 - 365.
[LW2] J. Lagarias and Y. Wang, Self-Affine tiles in Rn, Adv. Math. 121 (1996), 21 - 49.
[ LaW] I. Laba and Y. Wang, On spectral Cantor measures, J. Funct. Anal., 193 (2002), 409-420.
[Le] T. Lewis, The factorization of the Rectangular Distribution, J. Applied Probab., 4 (1967),
529-542.
[Lo] C. Long, Addition theorems for sets of integers, Pacific J. of Math., 23 (1967), 107-112.
[LRW] J. Lagarias, J. Reeds and Y. Wang, Orthonormal bases of exponentials for the n-cubes,
Duke Math. J., 103 (2000), 25-37.
[Lai] C.-K. Lai, On Fourier frame of absolutely continuous measures, J. Funct. Anal., 261 (2011),
2877-2889.
[PW] S. Pedersen and Y. Wang, Universal spectra, universal tiling sets and the spectral set
[S]
[T]
conjecture, Math Scand., 88 (2001), 246-256.
R. Strichartz, Mock Fourier series and transforms associated with certain Cantor mea-
sures, J. Anal. Math., 81 (2000), 209-238.
T. Tao, Fuglede's conjecture is false in 5 or higher dimensions, Math. Res. Lett., 11 (2004),
251-258.
E-mail address: [email protected]
E-mail address: [email protected]
Department of Mathematics and Statistics, McMaster University, Hamilton, On-
tario, L8S 4K1, Canada
23
|
1109.3805 | 1 | 1109 | 2011-09-17T17:21:09 | On the existence of universal series by trigonometric system | [
"math.FA"
] | In this paper we prove the following: let $\omega(t)$ be a continuous function, increasing in $[0,\infty)$ and $\omega(+0)=0$. Then there exists a series of the form$\sum_{k=-\infty}^\infty C_ke^{ikx}$ with $\sum_{k=-\infty}^\infty C^2_k \omega(|C_k|)<\infty$, $C_{-k}=\bar{C}_k$, with the following property: for each $\epsilon>0$ a weighted function $\mu(x), 0<\mu(x) \le1,| \{x\in[0,2\pi]: \mu(x)\not =1 \}| <\epsilon $ can be constructed, so that the series is universal in the weighted space $L_\mu^1[0,2\pi]$ with respect to rearrangements. | math.FA | math |
On the existence of universal series by
trigonometric system
S.A.Episkoposian
e-mail: [email protected]
In this paper we prove the following:
let ω(t) be a continuous function,
increasing in [0,∞) and ω(+0) = 0. Then there exists a series of the form
∞
∞
Xk=−∞
Ckeikx with
C2
k ω(Ck) < ∞, C−k = Ck,
Xk=−∞
with the following property: for each ε > 0 a weighted function µ(x), 0 < µ(x) ≤
1,{x ∈ [0, 2π] : µ(x) 6= 1} < ε can be constructed, so that the series is universal
in the weighted space L1
µ[0, 2π] with respect to rearrangements.
§1. INTRODUCTION
In 1932 F. Riesz(see [1], p. 655) proved that there exists a function f0(x) ∈
L1[0, 2π] so that its Fourier series with respect to the trigonometric system
does not converge in L1[0, 2π]. Consequently, there exist functions in the space
L1[0, 2π] that cannot be represented by trigonometric series in the metric of L1.
Let µ(x) be a measurable on [0, 2π] function with 0 < µ(x) ≤ 1, x ∈ [0, 2π]
µ[0, 2π] be a space of measurable functions f (x), x ∈ [0, 2π] with
and let L1
Z 2π
0
f (x)µ(x)dx < ∞.
L1
µ[0, 2π], such that for every function f (x) in the space L1
K.Kazarian and R.Zink in [2] proved that there exist a weighted space
µ[0, 2π] one can find
k=−∞ Ckeikx that converges to f (x) in the metric of
a trigonometric series P∞
µ[0, 2π].
L1
Moreover using other construction of weighted function µ(x) M.Grigorian
proved the following result [3]
-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -
The author was supported in part by Grant- 08-83 from the Government of
Armenia
AMS Classification 2000 Primary 42A20 .
1
Theorem 1. There exists a trigonometric series of the form
∞
Xk=−∞
Ckeikx with
∞
Xk=−∞
Ckq < ∞, ∀q > 2 C−k = Ck
(1.1)
with the following property: such that for any number ǫ > 0 a weighted function
µ(x), 0 < µ(x) ≤ 1,{x ∈ [0, 2π] : µ(x) 6= 1} < ǫ can be constructed, so that
the series (1.1) is universal in L1
µ[0, 2π] with respect to subseries (see Definition
3).
Now we present the definitions of universal series:
Definition 1. A functional series
∞
Xk=1
fk(x), fk(x) ∈ L1
µ[0, 2π]
(1.2)
µ[0, 2π] with respect to rearrange-
µ[0, 2π] the members of (1.2) can be rear-
∞
fσ(k)(x) converges to the function f (x)
in the metric L1
µ[0, 2π], i.e.
is said to be universal in weighted spaces L1
ments, if for any function f (x) ∈ L1
Xk=1
ranged so that the obtained series
fσ(k)(x) − f (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
n→∞Z 2π
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xk=1
lim
n
0
· µ(x)dx = 0.
Definition 2. The series (1.2) is said to be universal in weighted spaces
µ[0, 2π] in the usual sense, if for any function f (x) ∈ L1
L1
µ[0, 2π] there exists a
growing sequence of natural numbers nk such that the sequence of partial sums
with numbers nk of the series (1.2) converges to the function f (x) in the metric
L1
µ[0, 2π].
Definition 3. The series (1.2) is said to be universal in weighted spaces
µ[0, 2π] concerning subseries, if for any function f (x) ∈ L1
L1
µ[0, 2π] it is possible
to choose a partial series
fnk (x) from (1.2), which converges to the f (x) in
∞
the metric L1
µ[0, 2π].
Xk=1
The above mentioned definitions are given not in the most general form and
only in the generality, in which they will be applied in the present paper.
system universal in weighted L1
In this paper we consider a question on existence of series by trigonometric
µ[0, 2π] spaces with respect to rearrangements.
Note, that many papers are devoted (see [3]- [10]) to the question on existence
of various types of universal series in the sense of convergence almost everywhere
and on a measure.
Here we will give those results which directly concern to the Theorems,
proved in this paper.
The first usual universal in the sense of convergence almost everywhere
trigonometric series were constructed by D.E.Menshov [4] and V.Ya.Kozlov [5].
2
The series of the form
1
2
+
∞
Xk=1
akcoskx + bksinkx
(1.3)
was constructed just by them such that for any measurable on [0, 2π] function
f (x) there exists the growing sequence of natural numbers nk such that the
series (1.3) having the sequence of partial sums with numbers nk converges to
f (x) almost everywhere on [0, 2π].
convergence almost everywhere by convergence in the metric L1
Note here, that in this result, when f (x) ∈ L1
This result was distributed by A.A.Talalian on arbitrary orthonormal com-
n=1 - the
plete systems (see [6]). He also established (see [7]), that if {φn(x)}∞
normalized basis of space Lp
[0,1], p > 1, then there exists a series of the form
[0,2π], it is impossible to replace
[0,2π].
∞
Xk=1
akφk(x), ak → 0.
(1.4)
which has property:
for any measurable function f (x) the members of series
(1.4) can be rearranged so that the again received series converge on a measure
on [0,1] to f (x).
W. Orlicz [8] observed the fact that there exist functional series that are
universal with respect to rearrangements in the sense of a.e. convergence in the
class of a.e. finite measurable functions.
It is also useful to note that even Riemann proved that every convergent
numerical series which is not absolutely convergent is universal with respect to
rearrangements in the class of all real numbers.
In [9] it is proved the following result:
Theorem 2. Let {βk}∞
k=0 be a sequence of positive numbers with lim
k→∞
0. There exists a trigonometric series of the form
∞
Xk=−∞
Ckeikx with
∞
Xk=−∞
Ck βk < ∞, C−k = Ck
βk =
(1.5)
with the following property: such that for any number ǫ > 0 a weighted function
µ(x), 0 < µ(x) ≤ 1,{x ∈ [0, 2π] : µ(x) 6= 1} < ǫ can be constructed, so that
the series (1.5) is universal in L1
µ[0, 2π] with respect to rearrangements ( in the
usual sense).
In this paper we prove the following results.
Theorem 3. Let ω(t) be a continuous function, increasing in [0,∞) and
ω(+0) = 0. Then there exists a series of the form
∞
Xk=−∞
Ckeikx with
∞
Xk=−∞
C2
k ω(Ck) < ∞, C−k = Ck,
(1.6)
3
with the following property: for each ε > 0 a weighted function µ(x), 0 < µ(x) ≤
1,{x ∈ [0, 2π] : µ(x) 6= 1} < ε can be constructed, so that the series (1.6) is
universal in the weighted space L1
µ[0, 2π] with respect to rearrangements.
Analogous of this Theorem for Walsh system was proved by author in [10].
Remark. Using the proofs of Theorem 3 we can construct the series of
µ[0, 2π] with respect
the form (1.6) which are universal in the weighted space L1
simultaneously to rearrangements as well as to subseries.
The author thanks Professor M.G.Grigorian for his attention to this paper.
§2. BASIC LEMMA
Lemma . Let ω(t) be a continuous function,
ω(+0) = 0. Then for any given numbers 0 < ε < 1
function
increasing in [0,∞) and
2 , N0 > 2 and a step
f (x) =
(2.1)
where ∆s is an interval of the form ∆(i)
a measurable set E ⊂ [0, 2π] and a polynomial P (x) of the form
2m(cid:3), 1 ≤ i ≤ 2m, there exists
i
P (x) = XN0≤k<N
Ckeikx
which satisfy the conditions:
q
Xs=1
γs · χ∆s(x),
m = (cid:2) i−1
2m ,
E > 2π − ε,
ZE P (x) − f (x)dx < ε,
Ck2 · ω(Ck) < ε, C−k = C k
Ckeikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dx
< ε +Ze f (x)dx,
(1)
(2)
(3)
(4)
XN0≤k<N
Ze(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
N0≤m<N
XN0≤k≤m
max
for every measurable subset e of E.
number η with
η <
Proof of Lemma . Let 0 < ǫ < 1
2 be an arbitrary number. For any positive
ǫ2
4 ·(cid:20)Z 2π
0
−1
,
f 2(x)dx(cid:21)
(2.2)
4
ǫ
2
, δ}, s = 1, 2, ..., q.
4
max
1≤s≤q
1,
ǫ γs ·p∆s < min{
g(x) =
g1(t)e−iktdt(cid:12)(cid:12)(cid:12)(cid:12)
Z 2π
1 − 2
ε ,
<
0
1
2π (cid:12)(cid:12)(cid:12)(cid:12)
2 , 3ε·π
if x ∈ [0, 2π] \(cid:2) ε·π
2 (cid:3) .
if x ∈(cid:2) ε·π
2 , 3ε·π
2 (cid:3) ;
ε
16 · √N0
,
k < N0,
Set
where
(2.3)
(2.4)
(2.5)
(2.6)
(2.7)
(2.8)
(2.9)
by definition of function ω(t), there exists a positive number δ < ǫ so that for
any t, 0 < t < δ we have
Without restriction of generality, we assume that
ω(t) < ω(δ) < η.
we choose natural numbers ν1 and N1 so large that the following inequalities be
satisfied:
g1(x) = γ1 · g(ν1 · x) · χ∆1(x).
(By χE(x) we denote the characteristic function of the set E.)
We put
By (2.5), (2.7) and (2.8) we have
E1 = {x ∈ ∆s : gs(x) = γs},
E1 > 2π · (1 − ǫ) · ∆1; g1(x) = 0, x /∈ ∆1,
Z 2π
0
g2
1(x)dx <
2
ǫ · γ12 · ∆1.
k=−∞ is complete in L2[0, 2π], we can
(2.10)
Since the trigonometric system {eikx}∞
choose a natural number N1 > N0 so large that
Z 2π
0
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where
dx ≤
ε
8
,
(2.11)
X0≤k<N1
C(1)
k =
C(1)
k eikx − g1(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2π Z 2π
1
0
g1(t)e−iktdt.
Hence by (2.6) and (2.7) we obtain
Z 2π
0
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XN0≤k<N1
C(1)
k eikx − g1(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dx ≤
ε
8
+
X0≤k<N0
k 2
C(1)
1
2
<
ε
4
.
5
Now assume that the numbers ν1 < ν2 < ...νs−1, N1 < N2 < ... < Ns−1,
functions g1(x), g2(x), ..., gs−1(x) and the sets E1, E2, ...., Es−1 are defined.
We take sufficiently large natural numbers νs > νs−1 and Ns > Ns−1 to
satisfy
Z 2π
0
1
2π (cid:12)(cid:12)(cid:12)(cid:12)
where
Set
<
gs(t)e−iktdt(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z 2π
X0≤k<N1
0
ε
,
1 ≤ s ≤ q,
dx ≤
C(s)
16 ·pNs−1
k eikx − gs(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2π Z 2π
Es = {x ∈ ∆s : gs(x) = γs},
k =
C(s)
1
0
k < Ns−1,
(2.12)
ε
4s−1 ,
(2.13)
gs(t)e−iktdt.
(2.14)
(2.15)
gs(x) = γs · g(νs · x) · χ∆s(x),
Using the above arguments (see (2.9)-(2.11)), we conclude that the function
gs(x) and the set Es satisfy the conditions:
(2.16)
(2.17)
(2.18)
Es > 2π · (1 − ǫ) · ∆s; gs(x) = 0, x /∈ ∆s,
g2
s (x)dx <
2
ǫ · γs2 · ∆s.
0
Z 2π
XNs−1≤k<Ns
Z 2π
0
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
C(s)
k eikx − g1(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dx <
ε
2s+1 .
Thus, by induction we can define natural numbers ν1 < ν2 < ...νq, N1 <
N2 < ... < Nq, functions g1(x), g2(x), ..., gq(x) and sets E1, E2, ...., Eq such that
conditions (2.16)- (2.18) are satisfied for all s, 1 ≤ s ≤ q.
We define a set E and a polynomial P (x) as follows:
E =
P (x) = XN0≤k<N
Ckeikx =
where
Es,
q
[s=1
Xs=1
q
XNs−1≤k<Ns
C(s)
k eikx
,
(2.19)
(2.20)
6
Ck = C(s)
k
f or Ns−1 ≤ k < Ns, s = 1, 2, ..., q, C−k = C k N = Nq − 1.
(2.21)
By Bessel's inequality and (2.5), (2.14) for all s ∈ [1, q] we get
1
2
XNs−1≤k<Ns
k 2
C(s)
≤(cid:20)Z 2π
o
g2
s(x)dx(cid:21)
1
2
≤
2
√ε · γs ·p∆s, s = 1, 2, ..., q.
(2.22)
From (2.5), (2.14) and (2.15) it follows that
Taking relations (2.1), (2.5), (2.12), (2.14), (2.18) - (2.21) we obtain
E > 2π − ε.
ZE P (x) − f (x)dx ≤
q
Xs=1
C(s)
k eikx − gs(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XNs−1≤k<Ns
ZE(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
√ǫ · γs ·p∆s(cid:21) < δ.
1≤s≤q(cid:20) 2
Ck ≤ max
By (2.4), (2.21) and (2.22) for any k ∈ [N0, N ] we have
From this and (2.3) we get
ω(Ck) < ω(δ) < η, ∀ k ∈ [N0, N ].
Hence by (2.1), (2.2), (2.4) and (2.22) we obtain
dx
< ε
f 2(x)dx(cid:21) < ǫ.
XN0≤k<N
Ck2·ω(Ck) < η·
q
Xs=1
XNs−1≤k<Ns
k 2
C(s)
< η·
4
ǫ·(cid:20)Z 2π
0
That is, the statements 1) - 3) of Lemma are satisfied. Now we will check
the fulfillment of statement 4) of Lemma 2.
Let N0 ≤ m < N , then for some s0, 1 ≤ s0 ≤ q,
will have (see (2.13) and (2.14))
(Ns0 ≤ m < Ns0+1) we
Ckeikx =
XN0≤k≤m
s0
Xs=1
XNs−1≤k<Ns
k eikx
C(s)
+ XNs0−1≤k≤m
C(s0+1)
k
eikx.
Hence and from (2.1), (2.4), (2.5), (2.11) and (2.12) for any measurable set
e ⊂ E we obtain
7
s0
+
<
C(s)
≤
s0
dx ≤
Ze(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XN0≤k≤m
Ze(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XNs−1≤k<Ns
Xs=1
Xs=1Ze gs(x)dx +Ze(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2s+1 +Ze f (x)dx +
Xs=1
Ckeikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
k eikx − gs(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XNs0−1≤k≤m
dx
+
eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
√ε · γs0+1 ·p∆s0+1 <
C(s0+1)
k
s0
ε
dx <
2
<Ze f (x)dx + ε.
Lemma is proved.
§3.PROOF OF THEOREM 3
Let ω(t) be a continuous function, increasing in [0,∞) and ω(+0) = 0 and
let
{fn(x)}∞
n=1, x ∈ [0, 2π]
(3.1)
be a sequence of all step functions, values and constancy interval endpoints
of which are rational numbers.Applying Lemma 2 consecutively, we can find a
sequence {Es}∞
s=1 of sets and a sequence of polynomials
Ps(x) = XNs−1≤k<Ns
C(s)
k eikx,
1 = N0 < N1 < ... < Ns < ...., s = 1, 2, ....,
which satisfy the conditions:
Ps(x) = fs(x), x ∈ Es
Es > 1 − 2−2(s+1), Es ⊂ [0, 2π],
k ) < 2−2s, C(s)
C(s)
XNs−1≤k<Ns(cid:12)(cid:12)(cid:12)
k (cid:12)(cid:12)(cid:12) · ω(C(s)
−k = C
8
(3.2)
(3.3)
(3.4)
(3.5)
(s)
k
dx
< 2−2(s+1) +Ze fs(x)dx,
max
XNs−1≤k<p
for every measurable subset e of Es.
Denote
Ze(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Ns−1≤p<Ns
Xk=−∞
for Ns−1 ≤ k < Ns, s = 1, 2, ....
eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XNs−1≤k<Ns
Xs=1
where Ck = C(s)
k
Let ε be an arbitrary positive number. Setting
Ckeikx =
∞
∞
C(s)
k eikx
,
Es, n = 1, 2, ....;
∞
Ωn =
∞
\s=n
\s=n0
Ωn = Ωn0[ ∞
[n=n0+1
Ωn \ Ωn−1! .
E = Ωn0 =
Es, n0 = [log1/2 ε] + 1;
B =
∞
[n=n0
(3.6)
(3.7)
(3.8)
It is clear (see (3.4)) that B = 2π and E > 2π − ε.
We define a function µ(x) in the following way:
µ(x) =(cid:26) 1 f or x ∈ E ∪ ([0, 2π] \ B);
µn f or x ∈ Ωn \ Ωn−1, n ≥ n0 + 1,
(3.9)
where
µn ="22n ·
hs#−1
n
Ys=1
where
; hs = fs(x)C + max
Ns−1≤p<Ns k XNs−1≤k<p
C(s)
k eikxkC +1,
(3.10)
g(x)C = max
g(x) is a continuous function on [0, 2π].
From (3.5),(3.7)-(3.10) we obtain
(A) -- µ(x) is a measurable function and
x∈[0,2π]g(x),
0 < µ(x) ≤ 1,
{x ∈ [0, 2π] : µ(x) 6= 1} < ε.
9
∞
Xk=1
(B) --
Ck2 · ω(Ck) < ∞.
Hence, obviously we have
lim
k→∞
Ck = 0.
(3.11)
It follows from (3.8)-(3.10) that for all s ≥ n0 and p ∈ [Ns−1, Ns)
=
µ(x)dx =
∞
C(s)
Z[0,2π]\Ωs(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XNs−1≤k<p
ZΩn\Ωn−1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xn=s+1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2−2n
Z 2π
XNs−1≤k<p
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XNs−1≤k<p
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
s dx
<
C(s)
C(s)
h−1
0
µndx
≤
≤
∞
Xn=s+1
1
3
2−2s.
(3.12)
By (3.3), (3.8)-(3.10) for all s ≥ n0 we have
Z 1
0 Ps(x) − fs(x) µ(x)dx =ZΩs Ps(x) − fs(x) µ(x)dx+
∞
+Z[0,2π]\Ωs Ps(x) − fs(x) µ(x)dx = 2−2(s+1)+
Xn=s+1"ZΩn\Ωn−1 Ps(x) − fs(x) µndx# ≤ 2−2(s+1)+
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
fs(x) +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2−2s
s dx
0
Z 2π
h−1
<
XNs−1≤k<Ns
< 2−2(s+1) +
2−2s < 2−2s.
C(s)
1
3
+
∞
≤
Xn=s+1
(3.13)
Taking relations (3.6), (3.8)- (3.10) and (3.12) into account we obtain that
for all p ∈ [Ns−1, Ns) ands ≥ n0 + 1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z 2π
XNs−1≤k<p
=ZΩs(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
XNs−1≤k<p
0
C(s)
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
C(s)
10
µ(x)dx =
µ(x)dx+
<
s
<
s
C(s)
C(s)
µ(x)dx <
XNs−1≤k<p
XNs−1≤k<p
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+Z[0,2π]\Ωs(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
k eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ZΩn\Ωn−1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dx
Xn=n0+1
· µn +
Xn=n0+1 2−2(s+1) +ZΩn\Ωn−1 fs(x)dx! µn +
µn +ZΩs fs(x)µ(x)dx +
Xn=n0+1
= 2−2(s+1) ·
<Z 2π
fs(x)µ(x)dx + 2−2s.
µ[0, 2π] , i. e.R 2π
1
3
1
3
0
s
1
3
2−2s <
2−2s =
2−2s <
Let f (x) ∈ L1
It is easy to see that we can choose a function fν1 (x) from the sequence (3.1)
f (x)µ(x)dx < ∞ .
0
such that
Z 2π
0
Hence, we have
f (x) − fν1(x) µ(x)dx < 2−2, ν1 > n0 + 1.
(3.15)
f (x)µ(x)dx.
From (2.1), (A), (3.13) and (3.15) we obtain with m1 = 1
fν1 (x) µ(x)dx < 2−2 +Z 2π
Z 2π
0
0
(3.14)
(3.16)
(3.17)
Z 2π
0
+Z 2π
0
0
f (x) − fν1(x) µ(x)dx+
(cid:12)(cid:12)f (x) −(cid:2)Pν1 (x) + Cm1 eim1x(cid:3)(cid:12)(cid:12) µ(x)dx ≤
≤Z 2π
+Z 2π
(cid:12)(cid:12)Cm1 eim1x(cid:12)(cid:12) µ(x)dx < 2 · 2−2 + 2π · Cm1 .
fν1 (x) − Pν1 (x) µ(x)dx+
0
Assume that numbers ν1 < ν2 < ... < νq−1; m1 < m2 < ... < mq−1 are
chosen in such a way that the following condition is satisfied:
11
We choose a function fνq (x) from the sequence (3.1) such that
j
0
q−1
µ(x)dx <
f (x) −
Z 2π
Xs=1(cid:2)Pνs (x) + Cms eimsx(cid:3)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
< 2 · 2−2j + 2π ·(cid:12)(cid:12)Cmj(cid:12)(cid:12) , 1 ≤ j ≤ q − 1.
Xs=1(cid:2)Pνs (x) + Cms eimsx(cid:3)! − fnq (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)fνq (x)(cid:12)(cid:12) µ(x)dx < 2−2q + 2 · 2−2(q−1) + 2π ·(cid:12)(cid:12)Cmq−1(cid:12)(cid:12) =
µ(x)dx < 2−2q,
Z 2π
0
f (x) −
Z 2π
0
where
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
νq > νq−1; νq > mq−1
This with (3.18) imply
By (3.13), (3.14) and (3.20) we obtain
= 9 · 2−2q + 2π ·(cid:12)(cid:12)Cmq−1(cid:12)(cid:12) .
Z 2π
(cid:12)(cid:12)fνq (x) − Pνq (x)(cid:12)(cid:12) µ(x)dx < 2−2νq ,
Pνq (x) = XNνq −1≤k<Nνq
C(νq)
eikx.
k
0
(3.18)
(3.19)
(3.20)
(3.21)
(3.23)
µ(x)dx < 10 · 2−2q + 2π ·(cid:12)(cid:12)Cmq−1(cid:12)(cid:12) . (3.22)
Denote
Nνq −1≤p<N νqZ 2π
max
0
p
k
C(νq )
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xk=Nνq −1
eikx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
mq = minnn ∈ N : n /∈nn{k}Nνs −1
From (2.1), (A), (3.19) and (3.21) we have
s=1oo .
q
s=1 ∪ {ms}q−1
k=Nνs−1oq
Xs=1(cid:2)Pνs (x) + Cms eimsx(cid:3)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xs=1(cid:2)Pνs (x) + Cms eimsx(cid:3)! − fνq (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+Z 2π
(cid:12)(cid:12)fνq (x) − Pνq (x)(cid:12)(cid:12) µ(x)dx+
q−1
0
0
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z 2π
f (x) −
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f (x) −
µ(x)dx ≤
µ(x)dx+
≤Z 2π
0
12
+Z 2π
0
(cid:12)(cid:12)Cmq eimqx(cid:12)(cid:12) µ(x)dx < 2 · 2−2q + 2π ·(cid:12)(cid:12)Cmq(cid:12)(cid:12) .
(3.24)
Thus, by induction we on q can choose from series (3.7) a sequence of members
Cmq eimqx, q = 1, 2, ...,
and a sequence of polynomials
C(νq )
eikx, Nnq−1 > Nnq−1 , q = 1, 2, ....
Pνq (x) = XNνq −1≤k<Nνq
such that conditions (3.22) - (3.24) are satisfied for all q ≥ 1.
Taking account the choice of Pνq (x) and Cmq eimqx (see (3.23) and (3.25))
(3.25)
k
we conclude that the series
∞
XNνq −1≤k<Nνq
eikx + Cmq eiqx
k
C(νq)
is obtained from the series (3.7) by rearrangement of members. Denote this
Xq=1
series by P Cσ(k)eiσ(k)x.
It follows from (3.11), (3.22) and (3.24) that the series P Cσ(k)eiσ(k)x con-
verges to the function f (x) in the metric L1
versal with respect to rearrangements (see Definition 1).
µ[0, 2π], i.e. the series (3.7) is uni-
This completes the proof of Theorem 4.
REFERENCES
[1] N.K.Bary, Trigonometric series, Nauka, Moscow,1961; English trans.
Pergamon Press, Oxford, 1964.
in
[2] K.S.Kazarian, R.Zink ,Some ramifications of a theorem of Boas and Pollard
concerning the completion of a set of functions in L2, Trans.Amer.Math.Soc.,
v.349, n.11, p. 4367-4383.
[3] M. G. Grigorian "On the representation of functions by orthogonal series in
weighted Lp spaces, Studia. Math. 134(3)1999, 211-237.
[4] D.E.Menshov, On the partial summs of trigonometric series, Mat. Sb.
20(1947), 197-238 [in russian].
[5] V.Ya.Kozlov, On the complete systems of orthogonal functions, Mat. Sb.
26(1950), 351-364 [in russian].
13
[6] A.A.Talalian, On the convergence almost everywhere the sumbsequence of
partial summs of general orthogonal series,Izv. Ak. Nauk Arm. SSR ser. Math.
10(1957), 17-34 [in russian].
[7] A.A.Talalian, On the universal series with respect to rearrangements, Izv.
AN. SSSR ser. Math. 24(1960), 567-604 [in russian].
[8] W.Orlicz , Uber die unabhangig von der Anordnung fast uberall kniwergenten
Reihen, Bull. de l'Academie Polonaise des Sciences, 81 (1927), p. 117-125.
[9] S.A.Episkoposian, M.G.Grigorian,"On universal trigonometric series in weighted
spaces L1
µ[0, 2π] " , East Journal on Approximations, 1999,v.5 , n.4, 483-492.
[10] S.A. Episkoposian , "On the existence of universal series by Walsh system
", Izvestiya Natsionalnoi Akademii Nauk Armenii, English trans. in: Journal of
Contemporary Mathematical Analysis, 2003, v. 38 , n.4, p.25-40.
Department of Physics,
State University of Yerevan,
Alex Manukian 1, 375049 Yerevan, Armenia
e-mail: [email protected]
14
|
1705.02186 | 3 | 1705 | 2018-01-10T01:41:50 | An extension of Jensen's operator inequality and its application to Young inequality | [
"math.FA"
] | Jensen's operator inequality for convexifiable functions is obtained. This result contains classical Jensen's operator inequality as a particular case. As a consequence, a new refinement and a reverse of Young's inequality is given. | math.FA | math | AN EXTENSION OF JENSEN'S OPERATOR INEQUALITY AND ITS
APPLICATION TO YOUNG INEQUALITY
HAMID REZA MORADI, SHIGERU FURUICHI, FLAVIA-CORINA MITROI-SYMEONIDIS AND RAZIEH
NASERI
Abstract. Jensen's operator inequality for convexifiable functions is obtained. This result
contains classical Jensen's operator inequality as a particular case. As a consequence, a new
refinement and a reverse of Young's inequality are given.
8
1
0
2
n
a
J
0
1
]
.
A
F
h
t
a
m
[
3
v
6
8
1
2
0
.
5
0
7
1
:
v
i
X
r
a
1. Introduction and Preliminaries
In this article, H will denote a Hilbert space, and the term "operator" we shall mean endor-
morphism of H. The following result that provides an operator version for the Jensen inequality
is due to Mond and Pecari´c [12]:
Theorem 1.1. (Jensen's operator inequality for convex functions). Let A ∈ B (H) be a self-
adjoint operator with Sp (A) ⊆ [m, M] for some scalars m < M. If f (t) is a convex function
on [m, M], then
(1.1)
f (hAx, xi) ≤ hf (A) x, xi ,
for every unit vector x ∈ H.
Over the years, various extensions and generalizations of (1.1) have been obtained in the
literature, e.g., [6, 7, 13]. For this background we refer to any expository text such as [5].
The aim of this paper is to find an inequality which contains (1.1) as a special case. Our
result also allows to obtain a refinement and a reverse for the scalar Young inequality. More
precisely, it will be shown that for two non-negative numbers a, b we have
K r (h, 2) exp (cid:18) v (1 − v)
2
−
r
4(cid:19)(cid:18) a − b
D (cid:19)2! ≤
a∇vb
a♯vb
≤ K R (h, 2) exp (cid:18) v (1 − v)
2
R
4(cid:19)(cid:18) a − b
D (cid:19)2! ,
−
2010 Mathematics Subject Classification. Primary 47A63, 26A51. Secondary 26D15, 47A64, 46L05.
Key words and phrases. Convexifiable functions, Jensen's inequality, Young inequality, operator inequality.
1
2
An extension of Jensen's operator inequality and its application to Young inequality
where r = min {v, 1 − v}, R = max{v, 1 − v}, D = max{a, b} and K(h, 2) = (h+1)2
Kantorovich constant with h = b
a.
4h
is the
To make the text more self-contained we give a brief overview of convexifiable functions.
Given a continuous f : I → R defined on the compact interval I ⊂ R, consider a function
ϕ : I × R → R defined by ϕ (x, α) = f (x) − 1
If ϕ (x, α) is a convex function on I
for some α = α∗, then ϕ (x, α) is called a convexification of f and α∗ a convexifier on I. A
function f is convexifiable if it has a convexification. It is noted in [15, Corollary 2.9] that if the
continuously differentiable function f has Lipschitz derivative (i.e., f ′ (x) − f ′ (y) ≤ Lx − y
for any x, y ∈ I and some constant L), then α = −L is a convexifier of f .
The following fact concerning convexifiable functions plays an important role in our discussion
2 αx2.
(see [15, Corollary 2.8]):
(P)
If f is twice continuously differentiable, then α = min
t∈I
f ′′ (t) is a convexifier of f .
The reader may consult [16] for additional information about this topic. For all other notions
used in the paper, we refer the reader to the monograph [5].
2. Main Results
After the above preparation, we are ready to prove the analogue of (1.1) for non-convex
functions.
Theorem 2.1. (Jensen's operator inequality for non-convex functions). Let f be a continuous
convexifiable function on the interval I and α a convexifier of f . Then
(2.1)
f (hAx, xi) ≤ hf (A) x, xi −
1
2
α(cid:0)(cid:10)A2x, x(cid:11) − hAx, xi2(cid:1) ,
for every self-adjoint operator A with Sp (A) ⊆ I and every unit vector x ∈ H.
Proof. The idea of proof evolves from the approach in [17]. Let gα : I → R with gα (x) =
ϕ (x, α). According to the assumption, gα (x) is convex. Therefore
for every unit vector x ∈ H. This expression is equivalent to the desired inequality (2.1). (cid:3)
gα (hAx, xi) ≤ hgα (A) x, xi ,
A few remarks concerning Theorem 2.1 are in order.
Remark 2.1.
(a) Using the fact that for a convex function f one can choose the convexifier α = 0, one
recovers the inequality (1.1).
H.R. Moradi, S. Furuichi, F.-C. Mitroi-Symeonidis & R. Naseri
3
(b) For continuously differentiable function f with Lipschitz derivative and Lipschitz con-
stant L, we have
f (hAx, xi) ≤ hf (A) x, xi +
1
2
L(cid:0)(cid:10)A2x, x(cid:11) − hAx, xi2(cid:1) .
An important special case of Theorem 2.1, which refines inequality (1.1) can be explicitly
stated using the property (P).
Remark 2.2. Let f : I → R be a twice continuously differentiable strictly convex function and
α = min
t∈I
f ′′ (t). Then
(2.2)
f (hAx, xi) ≤ hf (A) x, xi −
1
2
α(cid:0)(cid:10)A2x, x(cid:11) − hAx, xi2(cid:1) ≤ hf (A) x, xi ,
for every positive operator A with Sp (A) ⊆ I and every unit vector x ∈ H.
The inequality (2.2) is obtained in the paper [13, Theorem 3.3] (where this result was derived
for the strongly convex functions) with a different technique (see also [4]).
The proof of the following corollary is adapted from the one of [5, Theorem 1.3], but we put
a sketch of the proof for the reader.
n
(2.3)
f n
Xi=1
Corollary 2.1. Let f be a continuous convexifiable function on the interval I and α a con-
vexifier. Let A1, . . . , An be self-adjoint operators on H with Sp (Ai) ⊆ I for 1 ≤ i ≤ n and
i=1 kxik2 = 1. Then
x1, . . . , xn ∈ H be such that Pn
hAixi, xii! ≤
1
Xi=1
2
Proof. In fact, x :=
x1
...
xn
"diagonal" operator on Hn
is a unit vector in the Hilbert space Hn.
hAixi, xii!2
.
i xi, xi(cid:11) − n
Xi=1
hf (Ai) xi, xii −
n
α
Xi=1 (cid:10)A2
If we introduce the
· · ·
0
...
. . .
· · · An
then, obviously, Sp (A) ⊆ I, kxk = 1, hf (A) x, xi =Pn
hA2x, xi = Pn
A :=
for A and x.
i=1 hA2
A1
...
0
,
i=1 hf (Ai) xi, xii, hAx, xi =Pn
i xi, xii. Hence, to complete the proof, it is enough to apply Theorem 2.1
i=1 hAixi, xii,
(cid:3)
Corollary 2.1 leads us to the following result. The argument depends on an idea of [1,
Corollary 1].
4
An extension of Jensen's operator inequality and its application to Young inequality
i=1 pi = 1. Then
Corollary 2.2. Let f be a continuous convexifiable function on the interval I and α a con-
vexifier. Let A1, . . . , An be self-adjoint operators on H with Sp (Ai) ⊆ I for 1 ≤ i ≤ n and let
p1, . . . , pn be positive scalars such that Pn
hpiAix, xi!2
(2.4) f n
hpiAix, xi! ≤
Xi=1
,
for every unit vector x ∈ H.
i=1 kxik2 = 1
Proof. Suppose that x ∈ H is a unit vector. Putting xi = √pix ∈ H so that Pn
α
Xi=1 (cid:10)piA2
and applying Corollary 2.1 we obtain the desired result (2.4).
i x, x(cid:11) − n
Xi=1
hpif (Ai) x, xi −
Xi=1
1
2
(cid:3)
n
n
The clear advantage of our approach over the Jensen operator inequality is shown in the fol-
(2.5)
i=1 pi = 1, then
lowing example. Before proceeding we recall the following multiple operator version of Jensen's
inequality [1, Corollary 1]: Let f : [m, M] ⊆ R → R be a convex function and Ai be self-adjoint
operators with Sp (Ai) ⊆ [m, M], i = 1, . . . , n for some scalars m < M. If pi ≥ 0, i = 1, . . . , n
with Pn
f n
Xi=1
for every x ∈ H with kxk = 1.
Example 2.1. We use the same idea from [17, Illustration 1]. Let f (t) = sin t (0 ≤ t ≤ 2π),
0 2π!
f ′′ (t) = −1, n = 2, p1 = p, p2 = 1 − p, H = R2, A1 = 2π 0
α = min
0≤t≤2π
1!. After simple calculations (thanks to the continuous functional calculus), from
hpiAix, xi! ≤
0!, A2 = 0
and x = 0
hpif (Ai) x, xi,
Xi=1
0
0
n
(2.4) we infer that
(2.6)
and (2.5) implies
sin (2π (1 − p)) ≤ 2π2p (1 − p) ,
0 ≤ p ≤ 1
(2.7)
0 ≤ p ≤ 1.
Not so surprisingly, the inequality (2.7) can break down when 1
applicable here). However, the new upper bound in (2.6) holds.
sin (2π (1 − p)) ≤ 0,
2 ≤ p ≤ 1 (i.e., (2.5) is not
The weighted version of [17, Theorem 3] follows from Corollary 2.2, i.e.,
(2.8)
f n
Xi=1
piti! ≤
n
Xi=1
pif (ti) −
1
2
n
α
Xi=1
pit2
i − n
Xi=1
piti!2
,
H.R. Moradi, S. Furuichi, F.-C. Mitroi-Symeonidis & R. Naseri
5
where ti ∈ I and Pn
(2.9)
i=1 pi = 1. For the case n = 2, the inequality (2.8) reduces to
f ((1 − v) t1 + vt2) ≤ (1 − v) f (t1) + vf (t2) −
v (1 − v)
2
α(t1 − t2)2,
where 0 ≤ v ≤ 1. In particular
(2.10)
f(cid:18) t1 + t2
2 (cid:19) ≤
f (t1) + f (t2)
2
1
8
−
α(t1 − t2)2.
It is notable that Theorem 2.1 is equivalent to the inequality (2.8). The following provides a
refinement of the arithmetic-geometric mean inequality.
Proposition 2.1. For each a, b > 0 and 0 ≤ v ≤ 1, we have
√ab ≤ Hv (a, b) −
d
(2.11)
8(cid:16)(1 − 2v)(cid:16)log
where d = min{a, b} and Hv (a, b) = a1−v bv +b1−vav
Proof. Assume that f is a twice differentiable convex function such that α ≤ f ′′ where α ∈ R.
Under these conditions, it follows that
is the Heinz mean.
8(cid:16)log
≤
≤
2
,
a
b(cid:17)2
a
b(cid:17)(cid:17)2
d
a + b
2 −
a + b
2
f(cid:18) a + b
2 (cid:19) = f(cid:18)(1 − v) a + vb + (1 − v) b + va
(cid:19)
f ((1 − v) a + vb) + f ((1 − v) b + va)
2
1
8
−
α((a − b) (1 − 2v))2
(by (2.10))
≤
≤
≤
f (a) + f (b)
2
f (a) + f (b)
2
1
8
−
,
2
α(a − b)2
(by (2.9))
for α ≥ 0. Now taking f (t) = et with t ∈ I = [a, b] in the above inequalities, we deduce the
desired inequality (2.11).
(cid:3)
Remark 2.3. As Bhatia pointed out in [2], the Heinz means interpolate between the geometric
mean and the arithmetic mean, i.e.,
(2.12)
√ab ≤ Hv (a, b) ≤
a + b
2
.
Of course, the first inequality in (2.11) yields an improvement of (2.12). The inequalities in
(2.11) also sharpens up the following inequality which is due to Dragomir (see [3, Remark 1]):
d
8(cid:16)log
a
b(cid:17)2
a + b
2 −
≤
√ab.
6
An extension of Jensen's operator inequality and its application to Young inequality
Studying about the arithmetic-geometric mean inequality, we cannot avoid mentioning its
cousin, the Young inequality. The following inequalities provides a multiplicative type refine-
ment and reverse of the Young's inequality:
(2.13)
K r (h, 2) ≤
(1 − v) a + vb
a1−vbv
≤ K R (h, 2) ,
where 0 ≤ v ≤ 1, r = min{v, 1 − v}, R = max{v, 1 − v} and K(h, 2) = (h+1)2
4h with h = b
a .
[18, Corollary 3], while the second one was given by
The first one was proved by Zuo et al.
Liao et al. [9, Corollary 2.2].
Our aim in the following is to establish a refinement for the inequalities in (2.13). The crucial
role for our purposes will play the following facts:
If f is a convex function on the fixed closed interval I, then
(2.14)
nλ( n
Xi=1
1
n
f (xi) − f n
Xi=1
1
n
xi!) ≤
n
Xi=1
pif (xi) − f n
Xi=1
pixi! ,
n
(2.15)
Xi=1
pif (xi) − f n
Xi=1
where p1, . . . , pn ≥ 0 with Pn
i=1 pi = 1, λ = min{p1, . . . , pn}, µ = max{p1, . . . , pn}. Notice
[10, Theorem 1, P.717], while the second
pixi! ≤ nµ( n
Xi=1
f (xi) − f n
Xi=1
that the first inequality goes back to Pecari´c et al.
xi!) ,
1
n
1
n
one was obtained by Mitroi in [11, Corollary 3.1].
Now we come to the announced theorem. In order to simplify the notations, we put a♯vb =
a1−vbv and a∇vb = (1 − v) a + vb.
Theorem 2.2. Let a, b > 0 and 0 ≤ v ≤ 1. Then
(2.16)
D (cid:19)2! ≤
K r (h, 2) exp (cid:18) v (1 − v)
4(cid:19)(cid:18) a − b
−
2
r
a∇vb
a♯vb
≤ K R (h, 2) exp (cid:18) v (1 − v)
2
4(cid:19)(cid:18) a − b
where r = min{v, 1 − v}, R = max{v, 1 − v}, D = max{a, b} and K(h, 2) = (h+1)2
h = b
a .
−
R
D (cid:19)2! ,
4h with
H.R. Moradi, S. Furuichi, F.-C. Mitroi-Symeonidis & R. Naseri
7
Proof. Employing the inequality (2.14) for the twice differentiable convex function f with α ≤
f ′′, we have
n
n
n
n
f (xi) − f 1
nλ( 1
Xi=1
Xi=1
2
nλ
i − 1
Xi=1
Xi=1
1
n
≤
x2
α
n
n
n
n
xi!) −
pif (xi) + f n
Xi=1
Xi=1
xi!2
−
i − n
Xi=1
Xi=1
pix2
n
pixi!
pixi!2
.
Here we set n = 2, x1 = a, x2 = b, p1 = 1 − v, p2 = v, λ = r and f (x) = − log x with I = [a, b]
(so α = min
D2 ). Thus we deduce the first inequality in (2.16). The second inequality
x∈I
(cid:3)
in (2.16) is also obtained similarly by using the inequality (2.15).
f ′′ (x) = 1
Remark 2.4.
(a) Since v(1−v)
2 − r
Therefore the first inequality in (2.16) provides an improvement for the first inequality
4 ≥ 0 for each 0 ≤ v ≤ 1, we have exp(cid:16)(cid:16) v(1−v)
2 − r
4(cid:17)(cid:0) a−b
D (cid:1)2(cid:17) ≥ 1.
in (2.13).
(b) Since v(1−v)
2 − R
4 ≤ 0 for each 0 ≤ v ≤ 1, we get exp(cid:16)(cid:16) v(1−v)
2 − R
fore the second inequality in (2.16) provides an improvement for the second inequality
4(cid:17)(cid:0) a−b
D (cid:1)2(cid:17) ≤ 1. There-
in (2.13).
Proposition 2.2. Under the same assumptions in Theorem 2.2, we have
(h + 1)2
4h
≥ exp 1
4(cid:18) a − b
D (cid:19)2! .
Proof. We prove the case a ≤ b, then h ≥ 1. We set f1(h) ≡ 2 log(h+1)−log h−2 log 2− 1
It is quite easy to see that f ′
a ≥ b, (then 0 < h ≤ 1), we also set f2(h) ≡ 2 log(h + 1) − log h − 2 log 2 − 1
direct calculation f ′
follows.
.
2h3(h+1) ≥ 0, so that f1(h) ≥ f1(1) = 0. For the case
4(h − 1)2. By
2h(h+1) ≤ 0, so that f2(h) ≥ f2(1) = 0. Thus the statement
2(h) = − (h−1)2(h+2)
1(h) = (2h+1)(h−1)2
(cid:3)
h2
4
(h−1)2
Remark 2.5. Dragomir obtained a refinement and reverse of Young's inequality in [3, Theorem
3] as:
(2.17)
exp v (1 − v)
2
D (cid:19)2! ≤
(cid:18) a − b
a♯vb ≤ exp v (1 − v)
a∇vb
2
(cid:18) a − b
d (cid:19)2! ,
where d = min{a, b}. From the following facts (a) and (b), we claim that our inequalities are
non-trivial results.
(a) From Proposition 2.2, our lower bound in (2.16) is tighter than the one in (2.17).
8
An extension of Jensen's operator inequality and its application to Young inequality
(b) Numerical computations show that there is no ordering between the right hand side in
(2.16) and the one in the second inequality of (2.17) shown in [3, Theorem 3]. For
example, if we take a = 2, b = 1 and v = 0.1, then
K R (h, 2) exp (cid:18) v (1 − v)
2
−
R
4(cid:19)(cid:18) a − b
D (cid:19)2! − exp v(1 − v)
2
(cid:18) a − b
d (cid:19)2! ≃ 0.0168761,
whereas it approximately equals −0.0436069 when a = 2, b = 1 and v = 0.3.
We give a further remark in relation to comparisons with other inequalities.
Remark 2.6. The following refined Young inequality and its reverse are known
(2.18)
K r′
(√t, 2)tv + r(1 −
√t)2 ≤ (1 − v) + vt ≤ K R′
(√t, 2)tv + r(1 −
√t)2,
where t > 0, r′ = min{2r, 1 − 2r} and R′ = max{2r, 1 − 2r}. The first and second inequality
were given in [14, Lemma 2.1] and in [9, Theorem 2.1], respectively.
Numerical computations show that there is no ordering between our inequalities (2.16) and
a with a ≥ b in (2.16)),
the above ones. Actually, if we take v = 0.45 and t = 0.1 (we set t = b
then
K R′
(√t, 2)tv + r(1 −
√t)2 − tvK R(h, 2) exp(cid:18)(cid:18) v(1 − v)
2
−
R
4(cid:19) (1 − t)2(cid:19) ≃ 0.0363059,
while it equals approximately −0.0860004 when v = 0.9 and t = 0.1.
Similarly, when v = 0.45 and t = 0.1 we get
K r′
(√t, 2)tv + r(1 −
√t)2 − tvK r(h, 2) exp(cid:18)(cid:18) v(1 − v)
2
−
r
4(cid:19) (1 − t)2(cid:19) ≃ −0.0126828,
while it equals approximately 0.037896 when v = 0.9 and t = 0.1.
Obviously, in the inequality (2.13), we cannot replace K r (h, 2) by K R (h, 2), or vice versa.
In this regard, we have the following theorem. The proof is almost the same as that of Theorem
2.2 (it is enough to use the convexity of the function gβ (x) = β
f ′′ (x)).
2 x2−f (x) where β = max
x∈I
Theorem 2.3. Let all the assumptions of Theorem 2.2 hold except that d = min{a, b}. Then
K R (h, 2) exp (cid:18) v (1 − v)
d (cid:19)2! ≤
4(cid:19)(cid:18) a − b
−
R
2
a∇vb
a♯vb
≤ K r (h, 2) exp (cid:18) v (1 − v)
2
r
4(cid:19)(cid:18) a − b
d (cid:19)2! .
−
We end this paper by presenting the operator inequalities based on Theorems 2.2 and 2.3,
thanks to the Kubo-Ando theory [8].
H.R. Moradi, S. Furuichi, F.-C. Mitroi-Symeonidis & R. Naseri
9
Corollary 2.3. Let A, B be two positive invertible operators and positive real numbers m, m′,
M, M ′ that satisfy one of the following conditions:
(i) 0 < m′I ≤ A ≤ mI < M I ≤ B ≤ M ′I.
(ii) 0 < m′I ≤ B ≤ mI < M I ≤ A ≤ M ′I.
4(cid:19)(cid:18) 1 − h
where r = min{v, 1 − v}, R = max{v, 1 − v} and K(h, 2) = (h+1)2
−
2
r
h (cid:19)2! A♯vB,
4h with h = M
m and h′ = M ′
m′ .
Proof. On account of (2.16), we have
r
4(cid:19)(cid:18) 1 − x
max{1, x}(cid:19)2!) T v
4(cid:19)(cid:18) 1 − x
for the positive operator T such that hI ≤ T ≤ h′I. Setting T = A− 1
−
R
2
In the first case we have I < hI = M
max{1, x}(cid:19)2!) T v,
2 BA− 1
2 .
Then
(2.19)
and
(2.20)
r
4(cid:19)(cid:18)1 − h
h (cid:19)2! A♯vB
R
4(cid:19)(cid:18) 1 − h′
h′ (cid:19)2! A♯vB
−
R
4(cid:19)(cid:18) 1 − h′
h′ (cid:19)2! A♯vB
2
−
K r (h, 2) exp (cid:18) v (1 − v)
≤ A∇vB
≤ K R (h′, 2) exp (cid:18) v (1 − v)
2
2
−
K R (h, 2) exp (cid:18) v (1 − v)
≤ A∇vB
≤ K r (h′, 2) exp (cid:18) v (1 − v)
2
min
h≤x≤h′ (K r (x, 2) exp (cid:18) v (1 − v)
≤ (1 − v) I + vT
h≤x≤h′ (K R (x, 2) exp (cid:18) v (1 − v)
≤ max
−
min
m I ≤ A− 1
1≤h≤x≤h′ (K r (x, 2) exp (cid:18) v (1 − v)
≤ (1 − v) I + vA− 1
1≤h≤x≤h′ (K R (x, 2) exp (cid:18) v (1 − v)
≤ max
2 BA− 1
−
2
2
2
2 BA− 1
m′ I = h′I, which implies that
2 ≤ M ′
4(cid:19)(cid:18) 1 − x
r
x (cid:19)2!)(cid:16)A− 1
2 BA− 1
2(cid:17)v
−
R
4(cid:19)(cid:18) 1 − x
x (cid:19)2!)(cid:16)A− 1
2 BA− 1
.
2(cid:17)v
10
An extension of Jensen's operator inequality and its application to Young inequality
We can write (2.20) in the form
r
4(cid:19)(cid:18)1 − h
h (cid:19)2!(cid:16)A− 1
2 BA− 1
2(cid:17)v
2
−
K r (h, 2) exp (cid:18) v (1 − v)
≤ (1 − v) I + vA− 1
≤ K R (h′, 2) exp (cid:18) v (1 − v)
2 BA− 1
2
2
−
R
4(cid:19)(cid:18) 1 − h′
h′ (cid:19)2!(cid:16)A− 1
2 BA− 1
.
2(cid:17)v
Finally, multiplying both sides of the previous inequality by A
1
2 we get the desired result (2.19).
The proof of other cases is similar, we omit the details.
(cid:3)
Acknowledgement
The authors thank anonymous referees for giving valuable comments and suggestions to
improve our manuscript. The author (S.F.) was partially supported by JSPS KAKENHI Grant
Number 16K05257.
References
[1] Agarwal, R. P., Dragomir, S. S.: A survey of Jensen type inequalities for functions of selfadjoint operators
in Hilbert spaces. Comput. Math. Appl., 59(12), 3785 -- 3812 (2010)
[2] Bhatia, R.: Interpolating the arithmetic-geometric mean inequality and its operator version. Linear Algebra
Appl., 413, 355 -- 363 (2006)
[3] Dragomir, S. S.: On new refinements and reverse of Young's operator inequality. arXiv:1510.01314v1
[4] Dragomir, S. S.: Some Jensen's type inequalities for twice differentiable functions of selfadjoint operators
in Hilbert spaces. Filomat., 23(3) 211 -- 222 (2009)
[5] Furuta, T., Mi´ci´c Hot, J., Pecari´c, J. Seo, Y.: Mond-Pecari´c method in operator inequalities. Monographs
in Inequalities 1, Element, Zagreb, 2005
[6] Horv´ath, L., Khan, K. A., Pecari´c, J.: Cyclic refinements of the different versions of operator Jensen's
inequality. Electron. J. Linear Algebra., 31(1), 125 -- 133 (2016)
[7] Kian, M.: Operator Jensen inequality for superquadratic functions. Linear Algebra Appl., 456, 82 -- 87 (2014)
[8] Kubo, F., Ando, T.: Means of positive linear operators. Math. Ann., 246(3), 205 -- 224 (1980)
[9] Liao, W., Wu, J., Zhao, J.: New versions of reverse Young and Heinz mean inequalities with the Kantorovich
constant. Taiwanese J. Math., 19(2), 467 -- 479 (2015)
[10] Mitrinovi´c, D. S., Pecari´c, J., Fink, A. M.: Classical and new inequalities in analysis. Kluwer Academic
Publishers, Dordrecht/Boston/London, 1993
[11] Mitroi, F. C.: About the precision in Jensen-Steffensen inequality. An. Univ. Craiova Ser. Mat. Inform.,
37(4), 73 -- 84 (2010)
[12] Mond, B., Pecari´c, J.: Convex inequalities in Hilbert space. Houston J. Math., 19(3), 405 -- 420 (1993)
[13] Moradi, H. R., Omidvar, M. E., Adil Khan, M., Nikodem, K.: Around Jensen's inequality for strongly
convex functions. Aequationes Math., (2017). https://doi.org/10.1007/s00010-017-0496-5
H.R. Moradi, S. Furuichi, F.-C. Mitroi-Symeonidis & R. Naseri
11
[14] Wu, J., Zhao, J.: Operator inequalities and reverse inequalities related to the Kittaneh-Manasrah inequal-
ities. Linear Multilinear Algebra., 62, 884 -- 894 (2014)
[15] Zlobec, S.: Characterization of convexifiable functions. Optimization., 55(3), 251 -- 261 (2006)
[16] Zlobec, S.: Convexifiable functions in integral calculus. Glasnik matematicki., 40(2), 241 -- 247 (2005)
[17] Zlobec, S.: Jensen's inequality for nonconvex functions. Math. Commun., 9(2), 119 -- 124 (2004)
[18] Zuo, H., Shi, G., Fujii, M.: Refined Young inequality with Kantorovich constant. J. Math. Inequal., 5(4),
551 -- 556 (2011)
(H.R. Moradi) Young Researchers and Elite Club, Mashhad Branch, Islamic Azad University, Mashhad, Iran.
E-mail address: [email protected]
(S. Furuichi) Department of Information Science, College of Humanities and Sciences, Nihon University, 3-25-40, Sakurajyousui,
Setagaya-ku, Tokyo, 156-8550, Japan.
E-mail address: [email protected]
(F.-C. Mitroi-Symeonidis) Department of Mathematical Methods and Models, Faculty of Applied Sciences, University Politehnica
of Bucharest, Romania.
E-mail address: [email protected]
(R. Naseri) Department of Mathematics, Payame Noor University, P.O. Box 19395-3697, Tehran, Iran.
E-mail address: [email protected]
|
1401.1394 | 2 | 1401 | 2015-03-13T17:17:07 | Curvature invariant on noncommutative polyballs | [
"math.FA",
"math.CV",
"math.OA"
] | In this paper we develop a theory of curvature (resp. multiplicity) invariant for tensor products of full Fock spaces and also for tensor products of symmetric Fock spaces. This is an attempt to find a more general framework for these invariants and extend some of the results obtained by Arveson for the symmetric Fock space, by the author and Kribs for the full Fock space, and by Fang for the Hardy space over the polydisc. To prove the existence of the curvature and its basic properties in these settings requires a new approach based on noncommutative Berezin transforms and multivariable operator theory on polyballs and varieties, as well as summability results for completely positive maps. The results are presented in the more general setting of regular polyballs. | math.FA | math |
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
GELU POPESCU
s (Hn
s (Hn1 )⊗· · ·⊗F 2
Abstract. In this paper we develop a theory of curvature (resp. multiplicity) invariant for tensor
products of full Fock spaces F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hn
k ) and also for tensor products of symmetric Fock
spaces F 2
k ). This is an attempt to find a more general framework for these invariants
and extend some of the results obtained by Arveson for the symmetric Fock space, by the author and
Kribs for the full Fock space, and by Fang for the Hardy space H 2(Dk) over the polydisc. To prove
the existence of the curvature and its basic properties in these settings requires a new approach based
on noncommutative Berezin transforms and multivariable operator theory on polyballs and varieties, as
well as summability results for completely positive maps. The results are presented in the more general
setting of regular polyballs.
Contents
Introduction
1. Curvature invariant on noncommutative polyballs
2. The curvature operator and classification
3. Invariant subspaces and multiplicity invariant
4. Stability, continuity, and multiplicative properties
5. Commutative polyballs and curvature invariant
References
Introduction
In [3] and [4], Arveson introduced and studied a notion of curvature for finite rank contractive
Hilbert modules over C[z1, . . . , zn], which is basically a numerical invariant for commuting n-tuples
T := (T1, . . . , Tn) in the unit ball
[B(H)n]−
n ≥ 0} ,
1 := {(X1, . . . , Xn) ∈ B(H)n : I − X1X ∗
1 − · · · − XnX ∗
1 − · · · − TnT ∗
with rank ∆T < ∞, where ∆T := I − T1T ∗
n and B(H) is the algebra of bounded linear
operators on a Hilbert space H. Subsequently, the author [21], [22] and, independently, Kribs [15] defined
and studied a notion of curvature for arbitrary elements in [B(H)n]−
1 and, in particular, for the full
Fock space F 2(Hn) with n generators. Some of these results were extended by Muhly and Solel [16]
to a class of completely positive maps on semifinite factors. The theory of Arveson's curvature on the
symmetric Fock space F 2
s (Hn) with n generators was significantly expanded due to the work by Greene,
Richter, and Sundberg [12], Fang [9], and Gleason, Richter, and Sundberg [13]. Englis remarked in [8]
that using Arveson's ideas one can extend the notion of curvature to complete Nevanlinna-Pick kernels.
The extension of Arveson's theory to holomorphic spaces with non Nevanlinna-Pick kernels was first
considered by Fang [11] who was able to show that the main results about the curvature invariant on
the symmetric Fock space carry over, with different proofs and using commutative algebra techniques,
to the Hardy space H 2(Dk) over the polydisc, in spite of its extremely complicated lattice of invariant
subspaces (see [29]). Inspired by some results on the invariant subspaces of the Dirichlet shift obtained
Date: November 7, 2014, revised version.
2000 Mathematics Subject Classification. Primary: 46L52; 32A70; Secondary: 47A13; 47A15.
Key words and phrases. Noncommutative polyball; Curvature invariant; Multiplicity invariant; Berezin transform; Char-
acteristic function; Fock space; Creation operators, Invariant subspaces.
Research supported in part by an NSF grant.
1
2
GELU POPESCU
by Richter [28], the theory of curvature invariant was extended to the Dirichlet space by Fang [10]. In
the noncommutative setting, a notion of curvature invariant for noncommutative domains generated by
positive regular free polynomials was considered in [23].
The goal of the present paper is to develop a theory of curvature invariant for the regular polyball
Bn(H), which will be introduced below. In particular, our results allow one to formulate a theory of
curvature invariant and multiplicity invariant for the tensor product of full Fock spaces F 2(Hn1 ) ⊗ · · · ⊗
F 2(Hnk ) and also for the tensor product of symmetric Fock spaces F 2
s (Hnk ). To prove
the existence of the curvature and its basic properties in these settings requires a new approach based
on noncommutative Berezin transforms and multivariable operator theory on polyballs and varieties (see
[20], [24], [25], and [26]), and also certain summability results for completely positive maps which are
trace contractive. In particular, we obtain new proofs for the existence of the curvature on the full Fock
space F 2(Hn), the Hardy space H 2(Dk) (which corresponds to n1 = · · · = nk = 1), and the symmetric
Fock space F 2
s (Hn). We expect our paper to be a step ahead towards finding a set of numerical (or
operatorial) unitary invariants which completely classify the elements of large classes of noncommutative
domains in B(H)m which admit universal operator models.
s (Hn1 ) ⊗ · · · ⊗ F 2
The results of the present paper play a crucial role in [27], where we introduce and study the Euler
characteristic associated with the elements of polyballs, and obtain an analogue of Arveson's version of
Gauss-Bonnet-Chern theorem from Riemannian geometry, which connects the curvature to the Euler
In particular, we prove that if M is an invari-
characteristic of some associated algebraic modules.
ant subspace of F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ), ni ≥ 2, which is graded (generated by multi-homogeneous
polynomials), then the curvature and the Euler characteristic of the orthocomplement of M coincide.
To present our results, we need some notation and preliminaries. Throughout this paper, we denote
by B(H)n1 ×c · · · ×c B(H)nk , where ni ∈ N := {1, 2, . . .}, the set of all tuples X := (X1, . . . , Xk) in
B(H)n1 × · · · × B(H)nk with the property that the entries of Xs := (Xs,1, . . . , Xs,ns) are commuting
with the entries of Xt := (Xt,1, . . . , Xt,nt) for any s, t ∈ {1, . . . , k}, s 6= t. Note that the operators
Xs,1, . . . , Xs,ns are not necessarily commuting. Let n := (n1, . . . , nk) and define the regular polyball
where the defect mapping ∆p
Bn(H) := {X ∈ B(H)n1 ×c · · · ×c B(H)nk : ∆p
X : B(H) → B(H) is defined by
X := (id − ΦX1 )p1 ◦ · · · ◦ (id − ΦXk )pk ,
∆p
and ΦXi : B(H) → B(H) is the completely positive linear map defined by ΦXi (Y ) :=Pni
j=1 Xi,jY X ∗
i,j
for Y ∈ B(H). We use the convention that (id − Φfi,Xi)0 = id. For information on completely bounded
(resp. positive) maps we refer the reader to [17].
Let Hni be an ni-dimensional complex Hilbert space with orthonormal basis ei
1, . . . , ei
ni. We consider
the full Fock space of Hni defined by
X(I) ≥ 0 for 0 ≤ p ≤ (1, . . . , 1)} ,
p := (p1, . . . , pk) ∈ Zk
+,
F 2(Hni ) := C1 ⊕Mp≥1
H ⊗p
ni
,
1, . . . , gi
ni and the identity gi
:= 1 ∈ C. The length of α ∈ F+
is the (Hilbert) tensor product of p copies of Hni . Let F+
j1 ⊗ · · · ⊗ ei
where H ⊗p
ni be the unital free semigroup
ni
on ni generators gi
ni and
ei
jp , where
gi
0
j1, . . . , jp ∈ {1, . . . , ni}. We define the left creation operator Si,j acting on the Fock space F 2(Hni ) by
, and the operator Si,j acting on the tensor product F 2(Hn1 )⊗· · ·⊗F 2(Hnk )
setting Si,jei
by setting
ni is defined by α := 0 if α = gi
0 and α := p if α = gi
j α, α ∈ F+
0. Set ei
α := ei
α := ei
gi
jp if α = gi
j1 · · · gi
jp ∈ F+
j1 · · · gi
ni
Si,j := I ⊗ · · · ⊗ I
⊗Si,j ⊗ I ⊗ · · · ⊗ I
,
i − 1 times
k − i times
{z
}
{z
}
where i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}. Let T = (T1, . . . , Tk) ∈ Bn(H) with Ti := (Ti,1, . . . , Ti,ni).
We use the notation Ti,αi := Ti,j1 · · · Ti,jp if αi = gi
:= I. The noncommutative
Berezin kernel associated with any element T in the noncommutative polyball Bn(H) is the operator
ni and Ti,gi
jp ∈ F+
j1 · · · gi
0
KT : H → F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ ∆T(I)(H)
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
3
defined by
KTh := Xβi∈F+
ni ,i=1,...,k
β1 ⊗ · · · ⊗ ek
e1
βk ⊗ ∆T(I)1/2T ∗
1,β1 · · · T ∗
k,βk
h,
where the defect of T is defined by ∆T(I) := (id − ΦT1) ◦ · · · ◦ (id − ΦTk )(I). A very important property
of the Berezin kernel is that
KTT ∗
i,j = (S∗
i,j ⊗ I)KT.
limqi→∞ Φqi
Ti
This was used in [25] to prove that T ∈ B(H)n1 × · · · × B(H)nk is a pure element in the regular polyball
Bn(H), i.e.
(I) = 0 in the weak operator topology, if and only if there is a Hilbert space
K and a subspace M ⊂ F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ K invariant under each operator Si,j such that
i,j ⊗ I)M⊥ under an appropriate idetification of H with M⊥. The k-tuple S := (S1, . . . , Sk),
i,j = (S∗
T ∗
where Si := (Si,1, . . . , Si,ni ), is an element in the regular polyball Bn(⊗k
i=1F 2(Hni )) and plays the role
of universal model for the abstract polyball Bn := {Bn(H) : H is a Hilbert space}. For more results
concerning noncommutative Berezin transforms and multivariable operator theory on noncommutative
balls and polydomains, we refer the reader to [20], [24], [25], and [26].
In Section 1, we introduce the curvature of any element T ∈ Bn(H) with trace class defect, i.e.
trace [∆T(I)] < ∞, by setting
curv (T) := lim
m→∞
1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
tracehK∗
T(P (1)
q1 ⊗ · · · ⊗ P (k)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)KTi
qk i
,
where KT is the Berezin kernel of T and P (i)
is the orthogonal projection of the full Fock space F 2(Hni )
qi
αi with αi ∈ F+
onto the span of all vectors ei
ni and αi = qi. The curvature curv(T) is a unitary invariant
for T that measures how far T is from being "free", i.e. a multiple of the universal model S. We prove
several summability results for completely positive maps which are trace contractive. These are used
together with the theory of Berezin transforms on noncommutative polyballs to prove the existence of
the curvature curv(T) and established several asymptotic formulas for the curvature invariant which are
very useful later on. It is also shown that the curvature of T := (T1, . . . , Tk) can be expressed in terms
of the associated completely positive maps ΦT1 , . . . , ΦTk . In particular, we prove that
curv (T) =
lim
(q1,...,qk)∈Zk
+
1
nq1
1 · · · nqk
k
trace(cid:2)Φq1
T1
◦ · · · ◦ Φqk
Tk
(∆T(I))(cid:3) ,
which implies the inequalities 0 ≤ curv (T) ≤ trace [∆T(I)] ≤ rank [∆T(I)].
In Section 2, we introduce the curvature operator ∆S⊗I H (KTK∗
T)(N⊗IH) associated with each element
T in the polyball Bn(H), which can be seen as a normalized "differential" of the Berezin transform. We
show that if T has characteristic function and finite rank, i.e. rank ∆T(I) < ∞, then the curvature
operator associated with T is trace class and
curv(T) = trace [∆S⊗IH (KTK∗
T)(N ⊗ IH)] .
This is used to obtain an index type result for the curvature, namely,
curv(T) = rank [∆T(I)] − trace [ΘT(PC ⊗ I)Θ∗
T(N ⊗ IH)] ,
where ΘT is the characteristic function of T. As a consequence of these results, we show that the
curvature invariant can be used to detect the elements T ∈ B(H)n1 × · · · × B(H)nk which are unitarily
equivalent to S ⊗ IK for some Hilbert space with dim K < ∞. These are precisely the pure elements T
in the regular polyball Bn(H) such that rank ∆T(I) is finite, ∆S⊗I (I − KTK∗
T) ≥ 0, and
curv(T) = rank [∆T(I)].
In this case, the Berezin kernel KT is a unitary operator and Ti,j = K∗
T(Si,j ⊗ I∆T(I)(H))KT.
We say that M is an invariant subspace of the tensor product F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ H or that
M is invariant under S ⊗ IH if it is invariant under each operator Si,j ⊗ IH. In Section 2, we show that
the curvature invariant completely classifies the finite rank Beurling type invariant subspaces of S ⊗ IH
4
GELU POPESCU
which do not contain reducing subspaces. In particular, the curvature invariant classifies the finite rank
Beurling type invariant subspaces of F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) (see Theorem 2.9).
Given an invariant subspace M of the tensor product F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ E, where E is a finite
dimensional Hilbert space, we introduce its multiplicity, in Section 3, by setting
m(M) := lim
m→∞
1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
tracehPM(P (1)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IE )i
qk i
.
Analogously, we define m(M⊥) by using PM⊥ instead of PM. The multiplicity measures the size of the
subspace. We prove that the multiplicity of M exists and provide several asymptotic formulas for it and
also an important connection with the curvature invariant, namely,
m(M) = dim E − curv(M),
where M is the compression of S ⊗ IE to the orthocomplement of M.
Unlike the case of the symmetric Fock space and the Hardy space over the polydisc when the curvature
and the multiplicity turn out to be nonnegative integers (see [12] and [11]), for the tensor product of full
Fock spaces, we prove that they can be any nonnegative real number, as long as at least one of the Fock
spaces has at least 2 generators (see Corollary 3.9). Moreover, we show that if n = (n1, . . . , nk) ∈ Nk
is such that ni ≥ 2 and nj ≥ 2 and i 6= j, then, for each t ∈ (0, 1), there exists an uncountable family
{T (ω)(t)}ω∈Ω of non-isomorphic pure elements of rank one defect in the regular polyball such that
curv(T (ω)(t)) = t,
for all ω ∈ Ω.
We also show that the curvature invariant detects the inner sequences of multipliers of ⊗k
when it takes the extremal value zero.
i=1F 2(Hni )
In Section 4, we consider several properties concerning the stability, continuity, and multiplicative
properties for the curvature and multiplicity. In particular, we show that the multiplicity invariant is
lower semi-continuous, i.e., if M and Mm are invariant subspaces of ⊗k
i=1F 2(Hni ) ⊗ E with dim E < ∞
and WOT- limm→∞ PMm = PM, then
lim inf
m→∞
m(Mm) ≥ m(M).
In Section 5, we introduce a curvature invariant associated with the elements of the commutative
polyball Bc
n(H) with the property that they have finite rank defects and constrained characteristic
functions. Since the case n1 = · · · = nk = 1 is considered in the previous sections, we assume,
throughout this section, that at least one ni ≥ 2. The commutative polyball Bc
n(H) is the set of all
X = (X1, . . . , Xk) ∈ Bn(H) with Xi = (Xi,1, . . . , Xi,ni), where the entries Xi,j are commuting opera-
tors. According to [26], the universal model associated with the abstract commutative polyball Bc
n is the
k-tuple B := (B1, . . . , Bk) with Bi = (Bi,1, . . . , Bi,ni ), where the operator Bi,j is acting on the tensor
product of symmetric Fock spaces F 2
s (Hnk ) and is defined by setting
s (Hn1 ) ⊗ · · · ⊗ F 2
Bi,j := I ⊗ · · · ⊗ I
⊗Bi,j ⊗ I ⊗ · · · ⊗ I
,
i − 1 times
k − i times
{z
}
{z
}
where Bi,j is the compression of the left creation operator Si,j to the symmetric Fock space F 2
s (Hni ). For
basic results concerning constrained Berezin transforms and multivariable model theory on the commu-
tative polyball Bc
n(H) we refer the reader to [26]. All the results of this section are under the assumption
that the elements T ∈ Bc
n(H) have characteristic functions.
Given an element T in the commutative polyball Bc
n(H) with the property that it has finite rank
defect and constrained characteristic function, we introduce its curvature by setting
curvc(T) := lim
m→∞
1
k (cid:19) Xq1≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
traceheK∗
tracehQ(1)
T(Q(1)
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IH)eKTi
qki
,
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
5
is the orthogonal projection of the symmetric Fock space F 2
where Q(i)
s (Hni ) onto its subspace of homo-
qi
geneous polynomials of degree qi. In spite of many similarities with the noncommutative case, the proof
of the existence of curvc(T) is quite different from the one for curv(T) and depends on the theory of
characteristic functions. We obtain commutative analogues of Theorem 2.6, Corollary 1.4, Theorem 2.7,
and Theorem 3.1. In particular, we mention that the curvature of T := (T1, . . . , Tk) can be expressed in
terms of the associated completely positive maps ΦT1 , . . . , ΦTk by the asymptotic formula
curvc(T) = n1! · · · nk!
lim
q1→∞
· · ·
lim
qk→∞
traceh(id − Φq1+1
T1
) ◦ · · · ◦ (id − Φqk+1
1 · · · qnk
qn1
Tk
k
)(I)i
.
When k = 1, we recover Arveson's asymptotic formula for the curvature.
Given a Beurling type invariant subspace M of the tensor product F 2
s (Hn1)⊗· · · ⊗ F 2
s (Hnk )⊗E, where
E is a finite dimensional Hilbert space, we introduce its multiplicity by setting
mc(M) := lim
m→∞
1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
tracehPM(Q(1)
tracehQ(1)
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IE )i
qk i
.
We show that the multiplicity invariant exists for Beurling type invariant subspace. It remains an open
problem whether the multiplicity invariant exists for arbitrary invariant subspace of the tensor product
s (Hn1 ) ⊗ · · · ⊗ F 2
F 2
s (Hnk ) ⊗ E. This is true for the polydisc, when n1 = · · · = nk = 1, and for the
symmetric Fock space, when k = 1. We remark that there are commutative analogues of all the results
from Section 4, concerning the stability, continuity, and multiplicative properties of the curvature and
multiplicity invariants.
Regarding the results of Section 5, when k ≥ 2 and at least one ni ≥ 2, an important problem remains
open. Can one drop the condition that the elements of the commutative polyball have constrained
characteristic functions ? In case of a positive answer, the multiplicity invariant would exist for any
invariant subspace of F 2
s (Hnk )
can be seen as a reproducing kernel Hilbert space of holomorphic functions on Bn1 × · · · × Bnk (see [26]),
where Bn := {z ∈ Cn : z < 1}. Under this identification, one can obtain an integral type formula for
the curvature, similar to the one introduced by Arveson [3]. Having in mind the results on the curvature
invariant on the symmetric Fock space (see [3], [4], [12], [9], [13]), significant problems still remain open
in the setting of Section 5.
s (Hnk ) ⊗ E. We should mention that F 2
s (Hn1 ) ⊗ · · · ⊗ F 2
s (Hn1 ) ⊗ · · · ⊗ F 2
Finally, we remark that one can re-formulate the results of the paper in terms of Hilbert modules
[7] over the complex semigroup algebra C[F+
nk ] generated by the direct product of the free
semigroups F+
nk . In this setting, the Hilbert module associated with the universal model S acting
n1
on the tensor product F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) plays the role of rank-one free module in the algebraic
theory [14]. The commutative case can be re-formulated in a similar manner.
n1 × · · · × F+
, . . . , F+
1. Curvature invariant on noncommutative polyballs
In this section we introduce the curvature invariant associated with the elements of the polyball Bn(H)
which have trace class defects. We prove several summability results for completely positive maps which
are trace contractive. These are used together with the theory of Berezin transforms on noncommutative
polyballs to prove the existence of the curvature and establish several asymptotic formulas which will
be very useful in the coming sections. It is also shown that the curvature of T := (T1, . . . , Tk) can be
expressed in terms of the associated completely positive maps ΦT1 , . . . , ΦTk .
Given two k-tuples q = (q1, . . . , qk) and p = (p1, . . . , pk) in Zk
+, we consider the partial order q ≤ p
defined by qi ≤ pi for any i ∈ {1, . . . , k}. We consider Zk
+ as a directed set with respect to this partial
order. Denote by T (H) the ideal of trace class operators on the Hilbert space H and let T +(H) be its
positive cone.
6
GELU POPESCU
Lemma 1.1. Let φ1, . . . , φk be commuting positive linear maps on B(H) such that φi(T +(H)) ⊂ T +(H)
and
for any X ∈ T +(H) and i ∈ {1, . . . , k}. Then the limit
trace [φi(X)] ≤ trace (X)
lim
(q1,...,qk)∈Zk
+
trace [φq1
1 ◦ · · · ◦ φqk
k (X)]
exists and is equal to
lim
m→∞
for any X ∈ T +(H).
1
k − 1 (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k − 1
q1+···+qk =m
trace [φq1
1 ◦ · · · ◦ φqk
k (X)]
Proof. Due to the hypotheses, if Y ∈ T +(H) and qi ∈ Z+ := {0, 1, . . .}, then φqi
0 ≤ trace [φqi+1
and
i (Y )]. Consequently, limqi→∞ trace [φqi
i (Y ) ∈ T +(H) and
i (Y )] exists. Assume that 1 ≤ p < k
(Y )] ≤ trace [φqi
i
(1.1)
lim
q1→∞
· · ·
lim
qp→∞
for any Y ∈ T +(H). Using the fact that φ1, . . . , φk are commuting and
qp+1+1
p+1
(φq1
1 ◦ · · · ◦ φqp
qp+1
p+1 (φq1
1 ◦ · · · ◦ φqp
tracehφ
tracehφq1
we can apply relation (1.1) when Y is equal to φ
0 ≤ lim
qp→∞
· · ·
lim
q1→∞
1 ◦ · · · ◦ φqp
p (φ
qp+1+1
p+1
for any qp+1 ∈ Z+. Consequently, limqp+1→∞ · · · limq1→∞ trace(cid:2)φq1
T +(H). Since the sequence {trace [φq1
indices q1, . . . , qk, it is clear that
k (X)]}q1,...,qk
1 ◦ · · · ◦ φqk
p (Y )(cid:3)
exists
1 ◦ · · · ◦ φqp
trace(cid:2)φq1
p (X))i ≤ trace(cid:2)φ
(X))i ≤ lim
qp+1
p+1 (X) or φ
qp→∞
p (X))(cid:3) ,
qp+1+1
p+1
(X) and deduce that
· · ·
lim
q1→∞
trace(cid:2)φq1
qp+1
1 ◦ · · · ◦ φ
p
qp+1
p (φ
1 ◦ · · · ◦ φqp
p+1 (X))(cid:3)
(X)(cid:3) exists for any X ∈
is decreasing with respect to each of the
lim
q1→∞
· · ·
lim
qk→∞
trace [φq1
1 ◦ · · · ◦ φqk
k (X)] =
lim
(q1,...,qk)∈Zk
+
=
inf
(q1,...,qk)∈Zk
+
trace [φq1
1 ◦ · · · ◦ φqk
k (X)]
trace [φq1
1 ◦ · · · ◦ φqk
k (X)]
and the order of the iterated limits does not matter. Set L(X) := inf (q1,...,qk)∈Zk
Given ǫ > 0 let N0 ∈ N be such that trace [φq1
Consider the following sets:
k (X)].
k (X)] − L(X) < ǫ for any q1 ≥ N0, . . . qk ≥ N0.
1 ◦ · · · ◦ φqk
+
trace [φq1
1 ◦ · · · ◦ φqk
Note that A1, . . . , Ak, BN0 are disjoint sets and
· · ·
A1 :=(cid:8)(q1, . . . , qk) ∈ Zk
A2 :=(cid:8)(q1, . . . , qk) ∈ Zk
Ak :=(cid:8)(q1, . . . , qk) ∈ Zk
BN0 :=(cid:8)(q1, . . . , qk) ∈ Zk
(cid:8)(q1, . . . , qk) ∈ Zk
card ((cid:8)(q1, . . . , qk) ∈ Zk
+ : q1 + · · · + qk = m, q1 < N0(cid:9)
+ : q1 + · · · + qk = m, q1 ≥ N0, q2 < N0(cid:9)
+ : q1 + · · · + qk = m, q1 ≥ N0, , . . . , qk−1 ≥ N0, qk < N0(cid:9)
+ : q1 + · · · + qk = m, q1 ≥ N0, , . . . , qk ≥ N0(cid:9) .
Ai! ∪ BN0 .
+ : q1 + · · · + qk = m(cid:9) = k[i=1
+ : q1 + · · · + qk = m(cid:9)) =(cid:18)m + k − 1
k − 1 (cid:19)
Regarding the cardinality of these sets, a close look reveals that
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
7
and card (Ak) ≤ · · · ≤ card (A1). Moreover, we have
card (A1) =
N0−1Xj=0 (cid:18)m − j + k − 2
k − 2
(cid:19) ≤ N0(cid:18)m + k − 2
k − 2 (cid:19) .
Consequently, using the fact that trace [φi(X)] ≤ trace (X) for any X ∈ T +(H) and i ∈ {1, . . . , k}, we
obtain
X(q1,...,qk)∈∪k
i=1Ai
trace [φq1
1 ◦ · · · ◦ φqk
k (X)] =
≤
trace [φq1
1 ◦ · · · ◦ φqk
kXi=1 X(q1,...,qk)∈Ai
card (Ai)trace (X) ≤ kN0(cid:18)m + k − 2
kXi=1
k (X)]
k − 2 (cid:19) trace (X).
Now, using these inequalities and that trace [φq1
we have
1 ◦ · · · ◦ φqk
k (X)] − L(X) < ǫ for any q1 ≥ N0, . . . qk ≥ N0,
k (X)] − L(X)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1 ◦ · · · ◦ φqk
k (X)]
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
trace [φq1
1 ◦ · · · ◦ φqk
1
1
≤
q1 +···+qk =m
trace [φq1
k − 1 (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k − 1
k − 1 (cid:19) X(q1,...,qk)∈∪k
(cid:18)m + k − 1
k − 1 (cid:19) X(q1,...,qk)∈BN0
(cid:18)m + k − 1
kN0(trace (X) + L(X))(cid:18)m + k − 2
k − 2 (cid:19)
(cid:18)m + k − 1
k − 1 (cid:19)
trace [φq1
i=1Ai
≤
+
1
Since limm→∞
k − 2
m + k − 2
m + k − 1
k − 1
1 ◦ · · · ◦ φqk
k (X)] − L(X) +
1
k − 1 (cid:19) L(X)card(cid:0)∪k
(cid:18)m + k − 1
i=1Ai(cid:1)
card (BN0)
+
(cid:18)m + k − 1
k − 1 (cid:19) ǫ.
= 0 and card (BN0 ) ≤(cid:18)m + k − 1
k − 1 (cid:19), one can easily complete the proof. (cid:3)
We remark that Lemma 1.1 implies lim(q1,...,qk)∈Zk
k (X)] ≤ trace (X) for any X ∈
T +(H). Using the classical Stolz-Ces`aro convergence theorem and Lemma 1.1, one can easily prove the
following result.
+
trace [φq1
1 ◦ · · · ◦ φqk
Lemma 1.2. Under the conditions of Lemma 1.1, the limit
exists and is equal to
lim
m→∞
lim
q1→∞
· · ·
1
q1 +···+qk ≤m
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
qkXsk=0
q1Xs1=0
lim
qk→∞
1
qk
1
q1
· · ·
trace [φq1
1 ◦ · · · ◦ φqk
k (X)]
trace [φq1
1 ◦ · · · ◦ φqk
k (X)]
for any X ∈ T +(H). Moreover, these limits are equal to those from Lemma 1.1.
8
GELU POPESCU
We remark that if we take X = ∆φ(I) := (id − φ1) ◦ · · · ◦ (id − φk)(I) and assume that it is a positive
trace class operator, then we can show that the limits in Lemma 1.2 are equal to
) ◦ · · · ◦ (id − φqk+1
traceh(id − φq1+1
1
q1 · · · qk
k
)(I)i
.
curv(φ) := lim
q1→∞
· · ·
lim
qk→∞
This can be seen as a curvature invariant associated with the k-tuple of positive maps φ = (φ1, . . . , φk)
satisfying the conditions of Lemma 1.1 and such that the defect ∆φ(I) is a positive trace class operator.
In this case, we have 0 ≤ curv(φ) ≤ trace [∆φ(I)].
Let T = (T1, . . . , Tk) be in the regular polyball Bn(H). If the defect of T,
∆T(I) := (id − ΦT1) ◦ · · · ◦ (id − ΦTk )(I),
is a trace class operator, we say that T has trace class defect. When ∆T(I) has finite rank, we say
that T has finite rank and write rank (T) := rank [∆T(I)]. We introduce the curvature of any element
T ∈ Bn(H) with trace class defect by setting
(1.2)
curv(T) := lim
m→∞
1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
tracehK∗
T(P (1)
q1 ⊗ · · · ⊗ P (k)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)KTi
qk i
,
where KT is the Berezin kernel of T and P (i)
is the orthogonal projection of the full Fock space F 2(Hni )
qi
αi with α ∈ F+
onto the span of all vectors ei
ni and αi = qi. In what follows, we show that curv(T) exists
and established several asymptotic formulas for the curvature invariant in terms of the Berezin kernel.
For each qi ∈ {0, 1, . . .} and i ∈ {1, . . . , k}, let P≤(q1,...,qk) be the orthogonal projection of ⊗k
i=1F 2(Hni )
onto the the span of all vectors of the form e1
α1 ⊗ · · · ⊗ ek
αk , where αi ∈ F+
ni, αi ≤ qi.
Theorem 1.3. If T = (T1, . . . , Tk) is in the regular polyball Bn(H) and has trace class defect, then the
curvature of T exists and satisfies the asymptotic formulas
curv(T) =
lim
(q1,...,qk)∈Zk
+
1
T(P (1)
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
tracehK∗
tracehP (1)
k − 1 (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k − 1
trace(cid:2)K∗
qk ⊗ IH)KTi
qk i
tracehK∗
tracehP (1)
T(P≤(q1,...,qk) ⊗ IH)KT(cid:3)
trace(cid:2)P≤(q1,...,qk)(cid:3)
T(P (1)
q1+···+qk =m
qk→∞
lim
.
· · ·
= lim
m→∞
= lim
q1→∞
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)KTi
qk i
Proof. Let S := (S1, . . . , Sk) ∈ Bn(⊗k
i=1F 2(Hni )) with Si := (Si,1, . . . , Si,ni ) be the universal model
associated with the abstract polyball Bn. It is easy to see that S := (S1, . . . , Sk) is a pure k-tuple and
(id − ΦS1) ◦ · · · ◦ (id − ΦSk )(I) = PC, where PC is the orthogonal projection from ⊗k
i=1F 2(Hni ) onto
C1 ⊂ ⊗k
i=1F 2(Hni ), where C1 is identified with C1 ⊗ · · · ⊗ C1. First, we prove that
(1.3)
for any q1, . . . , qk ∈ Z+. Taking into account that KTT ∗
j ∈ {1, . . . , ni}, we deduce that
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)KT = Φq1
T1
T(P (1)
K∗
◦ · · · ◦ Φqk
Tk
(∆T(I))
i,j = (S∗
i,j ⊗ I)KT for any i ∈ {1, . . . , k} and
K∗
T(cid:8)[Φq1
S1
◦ · · · ◦ Φqk
Sk
= Φq1
T1
◦ (id − ΦSk ) ◦ · · · ◦ (id − ΦS1 )(I)] ⊗ IH(cid:9) KT
◦ (id − ΦT1 ) ◦ · · · ◦ (id − ΦTk )(K∗
◦ · · · ◦ Φqk
Tk
TKT).
(1.4)
Since
K∗
TKT = lim
qk→∞
. . .
lim
q1→∞
(id − Φqk
Tk
) ◦ · · · ◦ (id − Φq1
T1
)(I),
where the limits are in the weak operator topology, we can prove that
(1.5)
∆T(K∗
TKT) = ∆T(I).
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
9
Indeed, note that that {(id − Φqk
Tk
operators and
) ◦ · · · ◦ (id − Φq1
T1
)(I)}q=(q1,...,qk)∈Zk
+
is an increasing sequence of positive
(id − Φqk
Tk
) ◦ · · · ◦ (id − Φq1
T1
)(I) =
Φsk
Tk
◦ · · ·
Φs1
T1
◦ (id − ΦTk ) ◦ · · · ◦ (id − ΦT1 )(I).
qk−1Xsk=0
q1−1Xs1=0
Since ΦT1 , . . . , ΦTk are commuting WOT-continuous completely positive linear maps and limqi→∞ Φqi
Ti
exists in the weak operator topology for each i ∈ {1, . . . , k}, we have
(I)
(id − ΦT1)(K∗
TKT) = lim
qk→∞
. . .
lim
q1→∞
(id − Φqk
Tk
) ◦ · · · ◦ (id − Φq1
T1
= lim
qk→∞
. . .
lim
q2→∞
(id − Φqk
Tk
= lim
qk→∞
. . .
lim
q2→∞
(id − Φqk
Tk
) ◦ · · · ◦ (id − Φq2
T2
) ◦ · · · ◦ (id − Φq2
T2
) ◦ (id − ΦT1)(I)
)(cid:20) lim
q1→∞
(id − Φq1
T1
) ◦ (id − ΦT1)(I).
) ◦ (id − ΦT1 )(I)(cid:21)
Applying now id − ΦT2, a similar reasoning leads to
(id − ΦT2 ) ◦ (id − ΦT1)(K∗
. . .
TKT) = lim
qk→∞
lim
q3→∞
(id − Φqk
Tk
) ◦ · · · ◦ (id − Φq3
T3
) ◦ (id − ΦT1 ) ◦ (id − ΦT2 )(I).
Continuing this process, we obtain relation (1.5). Now, using relations (1.4) and (1.5), we have
K∗
T(P (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)KT = K∗
= K∗
= Φq1
T1
= Φq1
T1
S1
Th(Φq1
T(cid:2)Φq1
S1
− Φq1+1
S1
) ◦ · · · ◦ (Φqk
Sk
− Φqk+1
Sk
◦ · · · ◦ Φqk
Sk
◦ (id − ΦSk ) ◦ · · · ◦ (id − ΦS1)(I) ⊗ IH(cid:3) KT
◦ (id − ΦT1 ) ◦ · · · ◦ (id − ΦTk )(K∗
(∆T(I)).
TKT)
◦ · · · ◦ Φqk
Tk
◦ · · · ◦ Φqk
Tk
)(I) ⊗ IHi KT
This proves relation (1.3). Now, note that for each i ∈ {1, . . . , k} and X ∈ T +(H), ΦTi (X) is a positive
trace class operator and
trace [ΦTi (X)] =
trace [Ti,jXT ∗
i,j] = trace
niXj=1
T ∗
i,jTi,jX
trace (X) ≤ nitrace (X).
niXj=1
≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
niXj=1
T ∗
i,jTi,j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Applying Lemma 1.1 when φi := 1
ni
ΦTi , i ∈ {1, . . . , k}, and X = ∆T(I), we deduce that the limit
exists, which proves the existence of the curvature invariant defined by relation (1.2). Setting
an application of Stolz-Ces`aro convergence theorem to the sequence {xq1 }∞
q1=0 implies
L := lim
q1→∞
· · ·
lim
qk→∞
1
nq1
1 · · · nqk
k
xq1 := lim
q2→∞
· · ·
lim
qk→∞
1
nq1
1 · · · nqk
k
lim
q1→∞
1
1 + n1 + · · · + nq1
1
q1Xs1=0
lim
q2→∞
· · ·
lim
qk→∞
nq2
2 · · · nqk
k
Similarly, setting yq2 := limq3→∞ · · · limqk→∞
that
1
nq2
qk
2 ···n
k
lim
q2→∞
1
1 + n2 + · · · + nq2
2
q2Xs2=0
lim
q3→∞
· · ·
lim
qk→∞
3 · · · nqk
nq3
k
1
1
◦ · · · ◦ Φqk
Tk
◦ · · · ◦ Φqk
Tk
(∆T(I))(cid:3)
(∆T(I))(cid:3) ,
◦ Φq2
T2
◦ · · · ◦ Φqk
Tk
◦ Φq2
T2
◦ · · · ◦ Φqk
Tk
T1
T1
T1
trace(cid:2)Φq1
trace(cid:2)Φq1
trace(cid:2)Φs1
trace(cid:2)Φs1
trace(cid:2)Φs1
T1
T1
◦ Φs2
T2
◦ Φq3
T3
· · · ◦ Φqk
Tk
xq1 = L.
q1→∞
(∆T(I))(cid:3) = lim
(∆T(I))(cid:3), we deduce
(∆T(I))(cid:3) = lim
q2→∞
yq2 .
10
GELU POPESCU
Continuing this process and putting together all these relations, we obtain
◦ · · · ◦ Φsk
Tk
sk=0 Φs1
T1
i=1(1 + ni + · · · + nqi
i )
) ◦ · · · ◦ (id − Φqk+1
Tk
i=1(1 + ni + · · · + nqi
i )
(∆T(I))(cid:3)
)(I)i
.
L = lim
q1→∞
· · ·
lim
qk→∞
= lim
q1→∞
· · ·
lim
qk→∞
curv(T) = lim
q1→∞
· · ·
= lim
q1→∞
· · ·
lim
qk→∞
lim
qk→∞
T1
trace(cid:2)Pq1
s1=0 · · ·Pqk
Qk
traceh(id − Φq1+1
Qk
trace(cid:2)Pq1
trace(cid:2)K∗
Consequently, using relation (1.3), we obtain
(∆T(I))(cid:3)
◦ · · · ◦ Φsk
Tk
i=1(1 + ni + · · · + nqi
i )
sk=0 Φs1
T1
s1=0 · · ·Pqk
Qk
T(P≤(q1,...,qk) ⊗ IH)KT(cid:3)
trace(cid:2)P≤(q1,...,qk)(cid:3)
,
(cid:3)
where the order of the iterated limits is irrelevant. The proof is complete.
If m ∈ Z+, we denote by P≤m the orthogonal projection of ⊗k
i=1F 2(Hni ) onto the the span of all
ni and α1 + · · · + αk ≤ m. In the particular case of the regular
vectors eα1 ⊗ · · · ⊗ eαk , where αi ∈ F+
polydisk, i.e. n1 = · · · = nk = 1, we have tracehP (1)
q1 ⊗ · · · ⊗ P (k)
qk i = 1 and, therefore,
curv(T) = lim
m→∞
trace [K∗
T(P≤m ⊗ IH)KT]
trace P≤m
.
Theorem 1.3 implies several other asymptotic formulas for the curvature in the regular polydisc. Using
Theorem 1.3 and Lemma 1.2 when φi := 1
ΦTi , i ∈ {1, . . . , k}, and X = ∆T(I), we obtain the following
ni
Corollary 1.4. Let T = (T1, . . . , Tk) be an element in the regular polyball Bn(H), n = (n1, . . . , nk) ∈ Nk,
with trace class defect and let ΦT1 , . . . , ΦTk be the associated completely positive maps. Then the curvature
of T satisfies the asymptotic formulas
curv(T) = lim
m→∞
= lim
m→∞
= lim
q1→∞
1
1
1
q1 +···+qk ≤m
k (cid:19) Xq1≥0,...,qk ≥0
(cid:18)m + k
k − 1 (cid:19) Xq1≥0,...,qk ≥0
(cid:18)m + k − 1
traceh(id − Φq1+1
Qk
trace(cid:2)Φq1
nq1
1 · · · nqk
q1 +···+qk =m
qk→∞
lim
T1
T1
1
k
· · ·
◦ · · · ◦ Φqk
Tk
nq1
1 · · · nqk
k
T1
trace(cid:2)Φq1
trace(cid:2)Φq1
k
T1
1
nq1
1 · · · nqk
) ◦ · · · ◦ (id − Φqk+1
(∆T(I))(cid:3)
(∆T(I))(cid:3)
◦ · · · ◦ Φqk
Tk
i=1(1 + ni + · · · + nqi
i )
◦ · · · ◦ Φqk
Tk
Tk
)(I)i
(∆T(I))(cid:3) .
=
lim
(q1,...,qk)∈Zk
+
Remark 1.5. A closer look at the proof of Thoorem 1.3 reveals that, if T = (T1, . . . , Tk) is an element
in B(H)n1 ×c · · · ×c B(H)nk such that ∆T(I) is a trace class positive operator, the limits in Corollary 1.4
exist and can be used as the definition of the curvature of T , extending the one on the regular polyball.
Corollary 1.6. If T ∈ Bn(H) has trace class defect, then
0 ≤ curv(T) ≤ trace [∆T(I)] ≤ rank [∆T(I)].
Corollary 1.7. If T ∈ Bn(H) and T′ ∈ Bn(H′) have trace class defects, then T ⊕ T′ ∈ Bn(H ⊕ H′) has
trace class defect and
If, in addition, dim H′ < ∞, then curv(T ⊕ T′) = curv(T).
curv(T ⊕ T′) = curv(T) + curv(T′).
More properties for the curvature will be presented in the coming sections.
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
11
2. The curvature operator and classification
In this section, we show that if T ∈ Bn(H) has characteristic function and finite rank, then its
curvature is the trace of the curvature operator, to be introduced, and satisfies an index type formula
which is used to obtain certain classification results. In particular, we show that the curvature invariant
classifies the finite rank Beurling type invariant subspaces of the tensor product of full Fock spaces
F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ).
Lemma 2.1. Let φ1, . . . , φk be positive linear maps on B(H) such that each φi is pure., i.e. φn
i (I) → 0
weakly as n → ∞. Then ∆φ := (id − φ1) ◦ · · · ◦ (id − φk) is a one-to-one map and each X ∈ B(H) has
the representation
where the iterated series converge in the weak operator topology. If, in addition, ∆φ(X) ≥ 0, then
X =
∞Xsk=0
sk−1
k−1 · · ·
φsk
φ
∞Xsk−1=0
k
X = X(s1,...,sk)∈Zk
+
∞Xs1=0
φs1
1 (∆φ(X)) · · ·! ,
φsk
k · · · φs1
1 (∆φ(X)).
Proof. We use the notation ∆(p1,...,pk)
φ
:= (id − φ1)p1 ◦ · · · ◦ (id − φk)pk for pi ∈ {0, 1}. Note that
(2.1)
1 (∆φ(X)) = ∆(0,1,...,1)
φs1
φ
(X) − φq1+1
1
(∆(0,1,...,1)
φ
(X)).
q1Xs1=0
If Z ∈ B(H) is a positive operator and x, y ∈ H, the Cauchy-Schwarz inequality implies
hφqi
i (Z)x, yi ≤ kZk hφqi
i (I)x, xi1/2 hφqi
i (I)y, yi1/2
.
i (I) → 0 weakly as qi → ∞, we deduce that hφqi
Since φqi
i (Z)x, yi → 0 as qi → ∞. Taking into account
that any bounded linear operator is a linear combination of positive operators, we conclude that the
convergence above holds for any Z ∈ B(H). Passing to the limit in relation (2.1) as qi → ∞, we obtain
(X)
(X)) = ∆(0,0,1...,1)
1 (∆φ(X)) = ∆(0,1,...,1)
2 (∆(0,1,...,1)
s1=0 φs1
s2=0 φs2
φ
φ
φ
P∞
and, continuing this process,
Putting together these relations, we deduce the first equality of the lemma. Now, one can see that ∆φ is
one-to-one. If we assume that ∆φ(X) ≥ 0, then the multi-sequence
φ
φ
(X)) = ∆(0,...,0)
k (∆(0,...,0,1)
φsk
(X). Similarly, we obtain P∞
∞Xsk=0
k
qk−1Xsk−1=0
k−1 · · ·
q1Xs1=0
φsk−1
φsk
φs1
qkXsk=0
(X) = X.
1 (∆φ(X)) · · ·!
is increasing with respect to each indexes q1, . . . , qk. The last part of the lemma follows.
(cid:3)
A simple consequence of Lemma 2.1 is the following result which will be used later.
Corollary 2.2. If T = (T1, . . . , Tk) is a pure element in the regular polyball Bn(H), then the defect map
∆T is one-to-one and any X ∈ B(H) has the following Taylor type representation around its defect
where the iterated series converge in the weak operator topology. If, in addition, ∆T(X) ≥ 0, then
X =
Φs2
Φs1
T1 ∞Xs2=0
∞Xs1=0
X = X(s1,...,sk)∈Zk
+
T2 · · ·
∞Xsk=0
(∆T(X)) · · ·!! ,
Φsk
Tk
Φs1
T1
◦ · · · ◦ Φsk
Tk
(∆T(X)).
12
GELU POPESCU
For each q = (q1, . . . , qk) ∈ Zk
+, define the operator N≤q acting on F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) by setting
Lemma 2.3. Let Y be a bounded operator acting on ⊗k
i=1F 2(Hni ) ⊗ H with dim(H) < ∞. Then
(2.2)
N≤q := X0≤si≤qi
i∈{1,...,k}
1
tracehP (1)
qk ⊗ IH)Yi
qk i
traceh(P (1)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
s1 ⊗ · · · ⊗ P (k)
sk i P (1)
s1 ⊗ · · · ⊗ P (k)
sk
.
= trace [∆S⊗IH(Y )(N≤q ⊗ IH)] ,
where S is the universal model associated with the abstract polyball Bn.
Proof. Let S := (S1, . . . , Sk) with Si := (Si,1, . . . , Si,ni ). It is easy to see that S := (S1, . . . , Sk) is a pure
k-tuple in the noncommutative polyball Bn(⊗k
i=1F 2(Hni )). Applying Corollary 2.2 to S ⊗ IH, we have
Y =
∞Xs1=0
Φs1
S1⊗IH ∞Xs2=0
S2⊗IH · · ·
Φs2
∞Xsk=0
Φsk
Sk⊗IH
(∆S⊗IH(Y )) · · ·!! ,
where the iterated series converge in the weak operator topology. Denoting Pq := P (1)
q = (q1, . . . , qk) ∈ Zk
+, the relation above implies
q1 ⊗ · · · ⊗ P (k)
qk
for
(Pq ⊗ IH) Y = (Pq ⊗ IH) q1Xs1=0
Φs1
S1⊗IH q2Xs2=0
S2⊗IH · · ·
Φs2
qkXsk=0
Φsk
Sk⊗IH
(∆S⊗IH (Y )) · · ·!!! .
Consequently, we have
trace [(Pq ⊗ IH) Y ]
1,α1 ⊗ IH)
◦ · · · ◦ Φsk
Sk⊗IH
(∆S⊗IH (Y ))(cid:3)
=
=
=
· · ·
· · ·
· · ·
q1Xs1=0
q1Xs1=0
q1Xs1=0
qkXsk=0
qkXsk=0
qkXsk=0
S1⊗IH
trace(cid:2)(Pq ⊗ IH) Φs1
trace(Pq ⊗ IH) Xα1∈F+
trace∆S⊗IH (Y ) Xα1 ∈F+
,···αk ∈F+
nk
α1=s1··· ,αk =sk
n1
(S1,α1 · · · Sk,αk ⊗ IH)∆S⊗IH (Y )(S∗
k,αk · · · S∗
(S∗
k,αk · · · S∗
1,α1
,···αk ∈F+
nk
α1=s1··· ,αk =sk
n1
PqS1,α1 · · · Sk,αk ) ⊗ IH .
Note that, for each i ∈ {1, . . . , k} and αi ∈ F+
ni with αi = si ≤ qi, we have
S∗
i,αi (I ⊗ · · · ⊗ P (i)
qi ⊗ · · · ⊗ I) = (I ⊗ · · · ⊗ P (i)
qi−si ⊗ · · · ⊗ I)S∗
i,αi
.
Consequently,
Xα1∈F+
,···αk ∈F+
nk
α1=s1··· ,αk =sk
n1
S∗
k,αk · · · S∗
1,α1
PqS1,α1 · · · Sk,αk
= P (1)
q1−s1 ⊗ · · · ⊗ P (k)
qk−sk Xα1 ∈F+
,···αk ∈F+
nk
α1=s1··· ,αk =sk
n1
= ns1
1 · · · nsk
k P (1)
q1−s1 ⊗ · · · ⊗ P (k)
qk−sk
.
S∗
k,αk · · · S∗
1,α1
S1,α1 · · · Sk,αk
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
13
Using the relations above, we obtain
trace [(Pq ⊗ IH) Y ]
1 · · · nsk
q1−s1 ⊗ · · · ⊗ P (k)
k (cid:16)P (1)
k (cid:16)P (1)
=
· · ·
· · ·
qkXsk=0
traceh∆S⊗IH(Y )ns1
q1Xs1=0
= trace"∆S⊗IH (Y )
q1Xs1=0
qkXsk=0
= trace(∆S⊗IH(Y )" q1Xs1=0
= trace(∆S⊗IH(Y )" q1Xs1=0
= trace∆S⊗IH (Y ) X0≤si≤qi
1
ns1
1
P (1)
i∈{1,...,k}
1 P (1)
ns1
trace [Ps]
qk−sk ⊗ IH(cid:17)i
qk−sk ⊗ IH(cid:17)#
qk−sk! ⊗ IH#) .
ns1
1 · · · nsk
q1−s1 ⊗ · · · ⊗ P (k)
k P (k)
nsk
q1−s1! ⊗ · · · ⊗ qkXsk=0
s1 ! ⊗ · · · ⊗ qkXsk=0
Ps ⊗ IH = trace [∆S⊗IH(Y )(N≤q ⊗ IH)] .
sk ! ⊗ IH#)
1
nsk
k
P (k)
1
(cid:3)
Hence, we deduce that
trace [(Pq ⊗ IH) Y ]
trace [Pq]
This completes the proof.
In spite of the fact that the operator ∆S⊗IH(KTK∗
and Lemma 2.3, in the particular case when Y = KTK∗
T) is not positive in general, using Theorem 1.3
T, we obtain the following result.
Corollary 2.4. If T is in the regular polyball Bn(H) and has finite rank, then
curv(T) = lim
q∈Zk
+
trace [∆S⊗IH(KTK∗
T)(N≤q ⊗ IH)] ,
where KT is the Berezin kernel of T.
Define the bounded linear operator N acting on F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) by setting
N := Xq=(q1,...,qk)∈Zk
+
1
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
qk i P (1)
q1 ⊗ · · · ⊗ P (k)
qk
and note that N is not a trace class operator.
Theorem 2.5. Let Y be a bounded operator acting on ⊗k
i=1F 2(Hni ) ⊗ H such that dim H < ∞ and
∆S⊗I (Y ) := (id − ΦS1⊗I ) ◦ · · · ◦ (id − ΦSk⊗I )(Y ) ≥ 0.
Then ∆S⊗I (Y )(N ⊗ IH) is a trace class operator and
trace [∆S⊗IH(Y )(N ⊗ IH)] = lim
q1→∞
· · ·
lim
qk→∞
traceh(P (1)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)Yi
qk i
.
Proof. First, note that, under the given conditions, Y should be a positive operator and
0 ≤
traceh(P (1)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)Yi
qk i
≤ kY k dim H
for any qi ≥ 0 and i ∈ {1, . . . , k}. Consequently, due to Lemma 2.3, we have
0 ≤ trace [∆S⊗IH(Y )(N≤q ⊗ IH)] ≤ kY k dim H.
14
GELU POPESCU
Since {N≤q}q∈Zk
∆S⊗IH(Y ) ≥ 0, we deduce that
+
is an increasing multi-sequence of positive operators convergent to N and we have
trace [∆S⊗IH (Y )(N ⊗ IH)] = lim
q∈Zk
+
trace [∆S⊗IH(Y )(N≤q ⊗ IH)] = sup
q∈Zk
+
trace [∆S⊗IH(Y )(N≤q ⊗ IH)] .
Therefore, ∆S⊗IH(Y )(N ⊗ IH) is a trace class operator. Using again Lemma 2.3, we can complete the
proof.
(cid:3)
Let S := (S1, . . . , Sk) be the universal model associated to the abstract noncommutative polyball Bn.
i=1F 2(Hni ) ⊗ K is called multi-analytic with respect to S if
An operator M : ⊗k
M (Si,j ⊗ IH) = (Si,j ⊗ IK)M for any i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}. In case M is a partial isometry,
we call it inner multi-analytic operator. We introduce the curvature operator ∆S⊗IH (KTK∗
T)(N ⊗ IH)
associated with each element T in the polyball Bn(H).
i=1F 2(Hni ) ⊗ H → ⊗k
An element T ∈ Bn(H) is said to have characteristic function if there is a multi-analytic operator
i=1F 2(Hni ) ⊗ ∆T(I)(H) with respect the universal model S such that
T = I . The characteristic function is essentially unique if we restrict it to its support.
ΘT : ⊗k
KTK∗
We proved in [25] that T has characteristic function if and only if ∆S⊗I(I − KTK∗
i=1F 2(Hni ) ⊗ E → ⊗k
T + ΘTΘ∗
T) ≥ 0.
Theorem 2.6. If T ∈ Bn(H) has finite rank and characteristic function, then the curvature operator of
T is trace class and
curv(T) = trace [∆S⊗IH (KTK∗
T)(N ⊗ IH)] = rank ∆T(I) − trace [ΘT(PC ⊗ I)Θ∗
T(N ⊗ IH)] ,
where ΘT is the characteristic function of T.
Proof. Applying Theorem 1.3 and Theorem 2.5, we obtain
curv(T) =
lim
(q1,...,qk)∈Zk
+
traceh(P (1)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
qk i
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IH)KTK∗
Ti
= rank ∆T(I) − trace [∆S⊗I (ΘTΘ∗
= rank ∆T(I) − trace [ΘT(PC ⊗ I)Θ∗
T(N ⊗ IH)] .
T)(N ⊗ IH)] = trace [∆S⊗I (KTK∗
T)(N ⊗ IH)]
The latter equality is due to the fact that ΘT is a multi-analytic operator and ∆S⊗I (I) = PC ⊗ I. The
proof is complete.
(cid:3)
limqi→∞ Φqi
Ti
We remark that T is pure, i.e.
(I) = 0 in the weak operator topology for each i ∈
{1, . . . , k}, if and only if KT is an isometry. We also mention that the set of all pure elements of the
polyball Bn(H) coincides with the pure elements X = (X1, . . . , Xk) ∈ B(H)n1 ×c · · · ×c B(H)nk such that
∆X(I) ≥ 0. In what follows, we show that the curvature invariant can be used to detect the elements
T ∈ B(H)n1 × · · · × B(H)nk which are unitarily equivalent to S ⊗ IK for some finite dimensional Hilbert
space K.
Theorem 2.7. If T = (T1, . . . , Tk) ∈ B(H)n1 ×· · ·×B(H)nk , then the following statements are equivalent:
(i) T is unitarily equivalent to S ⊗ IK for some finite dimensional Hilbert space K;
(ii) T is a pure finite rank element in the regular polyball Bn(H) such that ∆S⊗I (I − KTK∗
T) ≥ 0
and
curv(T) = rank [∆T(I)].
In this case, the Berezin kernel KT is a unitary operator and
Ti,j = K∗
T(Si,j ⊗ I∆T(I)(H))KT,
i ∈ {1, . . . , k}, j ∈ {1, . . . , ni}.
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
15
Proof. Assume that item (i) holds. Then there is unitary operator U : H → ⊗k
i=1F 2(Hni ) ⊗ K such that
Ti,j = U ∗(Si,j ⊗ IK)U for all i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}. Consequently, using Corollary 1.4, we
have
traceh(id − Φq1+1
) ◦ · · · ◦ (id − Φqk+1
Sk⊗IK
S1⊗IK
i=1(1 + ni + · · · + nqi
i )
)(I)i
curv(T) = lim
q1→∞
· · ·
lim
qk→∞
= curv(S ⊗ IK) = dim K.
Qk
On the other hand, rank [∆T(I)] = rank [∆S⊗IK(I)] = dim K. Therefore, curv(T) = rank [∆T(I)]. The
fact that T is pure is obvious. On the other hand, one can easily see that the Berezin kernel of S ⊗ IK
is a unitary operator and, consequently, so is the Berezin kernel of T. This completes the proof of the
implication (i) → (ii).
Assume that item (ii) holds. Since ∆S⊗I (I − KTK∗
T) ≥ 0, the k-tuple T has characteristic function
i=1F 2(Hni ) ⊗ ∆T(I)(H) which is a multi-analytic operator with respect
T = I . The support of ΘT is the smallest reducing
i=1F 2(Hni ) ⊗ E under each operator Si,j, containing the co-invariant subspace
T + ΘTΘ∗
i=1F 2(Hni) ⊗ E → ⊗k
ΘT : ⊗k
to the universal model S and KTK∗
subspace supp (ΘT) ⊂ ⊗k
M := Θ∗
T(⊗k
i=1F 2(Hni ) ⊗ ∆T(I)(H)). Due to Theorem 5.1 from [25], we deduce that
supp (ΘT) =
(S(α) ⊗ IH)(M) = ⊗k
i=1F 2(Hni ) ⊗ L,
_
(α)∈F+
n1 ×···×F+
nk
where L := (PC ⊗ IH)Θ∗(⊗k
that ΘT is a partial isometry. Due to the fact that Si,j are isometries, the initial space of ΘT, i.e.,
i=1F 2(Hni) ⊗ ∆T(I)(H)). Since T is pure, Theorem 6.3 from [25] implies
Θ∗
T(⊗k
i=1F 2(Hni ) ⊗ ∆T(I)(H)) = {x ∈ ⊗k
i=1F 2(Hni ) ⊗ E : kΦTxk = kxk}
i=1F 2(Hni )) ⊗ E under each operator Si,j, containing the co-invariant subspace Θ∗
is reducing under each Si,j. Since the support of ΘT is the smallest reducing subspace supp (ΘT) ⊂
i=1F 2(Hni ) ⊗
⊗k
∆T(I)(H)), we must have supp (ΘT) = Θ∗
i=1F 2(Hni ) ⊗ ∆T(I)(H)). Note that Φ := ΘTsupp (ΘT)
is an isometric multi-analytic operator and ΦΦ∗ = ΘTΘ∗
T. Due to Theorem 2.6, we have
T(⊗k
T(⊗k
curv(T) = rank [∆T(I)] − trace [∆S⊗I (ΘTΘ∗
T)(N ⊗ IH)]
rank [∆T(I)] − trace [Φ(PC ⊗ IL)Φ∗(N ⊗ IH)] .
q1 ⊗ · · · ⊗ P (k)
Since curv(T) = rank ∆T(I), we deduce that trace [Φ(PC ⊗ IL)Φ∗(N ⊗ IH)] = 0. Taking into account
that the trace is faithful, we obtain Φ(PC ⊗ IL)Φ∗(N ⊗ IH) = 0. The latter relation implies the equality
Φ(PC ⊗ IL)Φ∗(P (1)
+. Hence, we have Φ(PC ⊗ IL)Φ∗ = 0.
Since Φ is an isometry, we have Φ(PC ⊗ IL) = 0. Taking into account that Φ is a multi-analytic operator
with respect to S, we deduce that Φ = 0 and, consequently ΘT = 0. Now relation KTK∗
T = I
shows that KT is a co-isometry. On the other hand, since T is pure, the Berezin kernel associated
with T is an isometry. Therefore KT is a unitary operator. Since Ti,j = K∗
T(Si,j ⊗ I∆T(I)(H))KT for
i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}, the proof is complete.
(cid:3)
qk ⊗ IH) = 0 for any (q1, . . . , qk) ∈ Zk
T + ΘTΘ∗
Let S := (S1, . . . , Sk), where Si := (Si,1, . . . , Si,ni ), be the universal model for the abstract polyball
Bn, and let H be a Hilbert space. We say that M is an invariant subspace of ⊗k
i=1F 2(Hni ) ⊗ H or
that M is invariant under S ⊗ IH if it is invariant under each operator Si,j ⊗ IH for i ∈ {1, . . . , k} and
j ∈ {1, . . . , nj}.
Definition 2.8. Given two invariant subspaces M and N under S ⊗ IH, we say that they are unitarily
equivalent if there is a unitary operator U : M → N such that U (Si,j ⊗ IH)M = (Si,j ⊗ IH)N U for any
i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}.
According to Theorem 5.1 from [25], if a subspace M ⊂ ⊗k
i=1F 2(Hni ) ⊗ H is co-invariant under each
operator Si,j ⊗ IH, then
span(cid:8)(S1,β1 · · · Sk,βk ⊗ IE ) M : β1 ∈ F+
n1
, . . . , βk ∈ F+
i=1F 2(Hni ) ⊗ E,
nk(cid:9) = ⊗k
16
GELU POPESCU
where E := (PC ⊗ IH)(M). Consequently, a subspace R ⊆ ⊗k
if and only if there exists a subspace G ⊆ H such that R = ⊗k
i=1F 2(Hni ) ⊗ H is reducing under S ⊗ IH
i=1F 2(Hni ) ⊗ G.
We recall that Beurling [5] obtain a characterization of the invariant subspaces of the Hardy space
H 2(D) in terms of inner functions. This result was extended to the full Fock space F 2(Hn) in [18] and [19].
On the other hand, it is well known [29] that the lattice of the invariant subspaces for the Hardy space
H 2(Dk) is very complicated and contains many invariant subspaces which are not of Beurling type. The
i=1F 2(Hni ). Following the classical
same complicated situation occurs in the case of the tensor product ⊗k
case, we say that M is a Beurling type invariant subspace for S ⊗ IH if there is an inner multi-analytic
i=1F 2(Hni) ⊗ H with respect to S, where E is a Hilbert space, such
operator Ψ : ⊗k
i=1F 2(Hni ) ⊗ E → ⊗k
i=1F 2(Hni ) ⊗ E(cid:3). In this case, Ψ can be chosen to be isometric. In [25], we proved that
M is a Beurling type invariant subspace for S ⊗ IH if and only if
that M = Ψ(cid:2)⊗k
(id − ΦS1⊗IH ) ◦ · · · ◦ (id − ΦSk⊗IH )(PM) ≥ 0,
where PM is the orthogonal projection on M. If M is a Beurling type invariant subspace of S ⊗ IH,
then (S ⊗ IH)M := ((S1 ⊗ IH)M, . . . , (Sk ⊗ IH)M) is in the polyball Bn(M), where (Si ⊗ IH)M :=
((Si,1 ⊗ IH)M, . . . , (Si,ni ⊗ IH)M). We say that M has finite rank if (S ⊗ IH)M has finite rank.
The next result shows that the curvature invariant completely classifies the finite rank Beurling type
In particular, the curvature
invariant subspaces of S ⊗ IH which do not contain reducing subspaces.
invariant classifies the finite rank Beurling type invariant subspaces of F 2(Hn1) ⊗ · · · ⊗ F 2(Hnk ).
Theorem 2.9. Let M and N be finite rank Beurling type invariant subspaces of ⊗k
do not contain reducing subspaces of S ⊗ IH. Then M and N are unitarily equivalent if and only if
i=1F 2(Hni ) ⊗ H which
curv((S ⊗ IH)M) = curv((S ⊗ IH)N ).
Proof. If M is a Beurling type invariant subspaces of ⊗k
L and an isometric multi-analytic operator Ψ : ⊗k
Ψ[⊗k
i=1F 2(Hni) ⊗ L]. Since PM = ΨΨ∗, we have
i=1F 2(Hni ) ⊗ H, then there is a Hilbert space
i=1F 2(Hni ) ⊗ H such that M =
i=1F 2(Hni ) ⊗ L → ⊗k
∆(S⊗I)M(IM) =(cid:0)id − Φ(S1⊗I)M(cid:1) ◦ · · · ◦(cid:0)id − Φ(Sk⊗I)M(cid:1) (PM)
= Ψ (id − ΦS1⊗I ) ◦ · · · ◦ (id − ΦSk⊗I ) (I)Ψ∗M = Ψ(PC ⊗ IL)Ψ∗M.
Let {ℓω}ω∈Ω be an orthonormal basis for L. Note that {vω := Ψ(1 ⊗ ℓω) : ω ∈ Ω} is an orthonormal set
and
{Ψ(e1
β1 ⊗ · · · ⊗ ek
βk ⊗ ℓω) : βi ∈ F+
ni
, i ∈ {1, . . . , k}, ω ∈ Ω}
is an orthonormal basis for M. Note also that Ψ(PC ⊗ IL)Ψ∗(M) coincides with the closure of the range
of the defect operator ∆(S⊗I)M(IM) and also to the closed linear span of {vω := Ψ(1 ⊗ ℓω) : ω ∈ Ω}.
Consequently, rank [(S ⊗ I)M] = card Ω = dim L. Now, assume that rank (S ⊗ I)M) = p = dim L. Since
Ψ is a multi-analytic operator and PM = ΨΨ∗, we have
Hence, we deduce that
traceh(cid:16)id − Φq1+1
(S1⊗I)M(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
= traceh(cid:16)id − Φq1+1
S1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
(Sk⊗I)M(cid:17) (IM)i = traceh(cid:16)id − Φq1+1
Sk (cid:17) (I)i dim L.
Using Corollary 1.4 and the fact that curv(S) = 1, we deduce that curv((S ⊗ I)M) = dim L = p =
rank [(S ⊗ I)M]. Now, if M and N are finite rank Beurling type invariant subspaces of ⊗k
i=1F 2(Hni ) ⊗ H
and (S ⊗ I)M is unitarily equivalent to (S ⊗ I)N , then it is clear that curv((S ⊗ I)M) = curv((S ⊗ I)N ).
To prove the converse, assume that the later equality holds. As we saw above, we must have rank ((S ⊗
I)M) = rank ((S ⊗ I)N ). Consequently, the defect spaces associated with (S ⊗ I)M and (S ⊗ I)N have
the same dimension. Using the Wold decomposition from [25], we conclude that (S ⊗ I)M is unitarily
equivalent to (S ⊗ I)N . The proof is complete.
(cid:3)
(cid:16)id − Φq1+1
(S1⊗I)M(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
(Sk⊗I)M(cid:17) (IM) =(cid:16)id − Φq1+1
= Ψ(cid:16)id − Φq1+1
S1⊗I(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
S1⊗IL(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Sk⊗I(cid:17) (PM)
Sk⊗IL(cid:17) (I)Ψ∗M.
S1⊗IL(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Sk⊗IL(cid:17) (I)i
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
17
3. Invariant subspaces and multiplicity invariant
In this section, we prove the existence of the multiplicity of any invariant subspace of the tensor
product F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ E, where E is a finite dimensional Hilbert space. We provide several
asymptotic formulas for the multiplicity and an important connection with the curvature invariant. We
show that the range of both the curvature and the multiplicity invariant is the interval [0, ∞). We prove
that there is an uncountable family of non-unitarily equivalent pure elements in the noncommutative ball
with rank one and the same prescribed curvature. We also show that the curvature invariant detects the
inner sequences of multipliers of ⊗k
i=1F 2(Hni) when it takes the extremal value zero.
Let M be an invariant subspace of the tensor product F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ E, where E is a
finite dimensional Hilbert space. We introduce the multiplicity of M by setting
q1 ⊗ · · · ⊗ P (k)
1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
tracehPM(P (1)
tracehP (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IE )i
qk i
.
m(M) := lim
m→∞
Similarly, we define m(M⊥) by replacing PM with PM⊥. Note that m(M) + m(M⊥) = dim E.
Theorem 3.1. Let M be an invariant subspace of the tensor product F 2(Hn1 )⊗· · ·⊗ F 2(Hnk )⊗E, where
E is a finite dimensional Hilbert space. Then the multiplicity of M exists and satisfies the equations
m(M) =
lim
(q1,...,qk)∈Zk
+
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
tracehPM(P (1)
tracehP (1)
k − 1 (cid:19) Xq1≥0,...,qk ≥0
(cid:18)m + k − 1
q1 +···+qk =m
1
qk ⊗ IE )i
qk i
tracehPM(P (1)
tracehP (1)
= lim
q1→∞
· · ·
lim
qk→∞
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IE )i
qk i
= lim
m→∞
= dim E − curv(M),
trace(cid:2)PM(P≤(q1,...,qk) ⊗ IE )(cid:3)
trace(cid:2)P≤(q1,...,qk)(cid:3)
where M := (M1, . . . , Mk) with Mi := (Mi,1, . . . , Mi,ni ) and Mi,j := PM⊥ (Si,j ⊗ IE )M⊥ .
r,s ⊗ IE )M⊥ and deduce that
∆p
M(IM⊥ ) = PM⊥ ∆p
Proof. Since M is an invariant subspace under each operator Si,j ⊗ IE for i ∈ {1, . . . , k}, j ∈ {1, . . . , ni},
we have M ∗
i,jM ∗
r,s = (S∗
i,j ⊗ IE )(S∗
(3.1)
for any p = (p1, . . . , pk) with pi ∈ {0, 1}. Therefore, M is in the polyball Bn(M⊥) and has finite rank.
Since M⊥ is invariant under S∗
◦ · · · ◦ Φqk
Mk
i,j ⊗ IE and using relation (3.1), we obtain
(∆M(IM⊥ ))] = trace [PM⊥Φq1
S⊗I (I)M⊥ ≥ 0
trace [Φq1
M1
= trace [PM⊥(P (1)
S1⊗I ◦ · · · ◦ Φqk
q1 ⊗ · · · ⊗ P (k)
Sk⊗I (∆S⊗I (I))M⊥ ]
qk ⊗ IE )]
for any qi ∈ Z+. Therefore, due to Theorem 1.3 and Corollary 1.4, curv(M) exists and one can easily
complete the proof.
(cid:3)
Using Theorem 3.1 and Lemma 2.3, we deduce the following result.
Corollary 3.2. Let M be an invariant subspace of tensor product F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ E, where
E is a finite dimensional Hilbert space. Then the multiplicity of M satisfies the formula
m(M) = lim
q∈Zk
+
trace [∆S⊗IE (PM)(N≤q ⊗ IE )] ,
where N≤q is defined by relation (2.2).
Corollary 3.3. If M is an invariant subspace for the Hardy space H 2(Dk) of the polydisc, then
exists, where P≤m is the orthogonal projection onto the polynomials of degree ≤ m.
m(M) := lim
m→∞
trace [PM(P≤m ⊗ IE )]
trace [P≤m]
18
GELU POPESCU
The existence of the limit in Corollary 3.3 was first proved by Fang (see [11]) using different methods.
According to Theorem 3.1, we have several other asymptotic formulas for the multiplicity.
Theorem 3.4. Given a function κ : N → N and n(i) ∈ Nκ(i) for i ∈ {1, . . . , p}, let S(n(i)) and S(n(1),...,n(p))
be the universal models of the polyballs Bn(i) and B(n(1),...,n(p)), respectively. For each i ∈ {1, . . . , p},
assume that
(i) Ei is a finite dimensional Hilbert space with dim Ei = qi ∈ N;
(ii) Mi is an invariant subspace under S(n(i)) ⊗ IEi ;
(iii) ci := curv(cid:16)PM⊥
(iv) c := curv(cid:16)PM⊥(cid:16)S(n(1),...,n(p)) ⊗ IE1⊗···⊗Ep(cid:17) M⊥(cid:17),
i (cid:16)S(n(i)) ⊗ IEi(cid:17) M⊥
i (cid:17);
where, under the appropriate identification, M := M1 ⊗ · · · ⊗ Mp is viewed as an invariant subspace for
S(n(1),...,n(p)) ⊗ IE1⊗···⊗Ep . Then the curvature invariant satisfies the equation
c = q1 · · · qp −
(qi − ci)
pYi=1
and the multiplicity invariant satisfies the equation m(M1 ⊗ · · · ⊗ Mp) =Qp
Proof. For each i ∈ {1, . . . , p} let n(i) := (n(i)
1 , . . . , n(i)
i=1 m(Mi).
full Fock space with n(i)
j generators, and denote by P
the span of all homogeneous polynomials of F 2(H
n(i)
j
κ(i)) ∈ Nκ(i). If j ∈ {1, . . . , κ(i)}, let F 2(H (i)
the orthogonal projection of F 2(H
(n(i)
j )
q(i)
) of degree equal to q(i)
j
j ∈ Z+. Set
nj ) be the
) onto
n(i)
j
Q1 := P
(n(1)
1 )
q(1)
1
⊗ · · · ⊗ P
κ(1))
(n(1)
q(1)
κ(1)
,
· · · , Qp := P
(n(p)
1 )
q(p)
1
⊗ · · · ⊗ P
κ(p))
(n(p)
q(p)
κ(p)
and let U be the unitary operator which provides the canonical identification of the Hilbert tensor product
⊗p
Note that
nκ(i) )io ⊗ (E1 ⊗ · · · ⊗ Ep).
nκ(i) ) ⊗ Eii with n⊗p
i=1hF 2(H (i)
i=1hF 2(H (i)
n1 ) ⊗ · · · ⊗ F 2(H (i)
n1 ) ⊗ · · · ⊗ F 2(H (i)
U [(Q1 ⊗ IE1 ) ⊗ · · · ⊗ (Qp ⊗ IEp )] = [Q1 ⊗ · · · ⊗ Qp ⊗ IE1⊗···⊗Ep]U
and U (M1 ⊗ · · · ⊗ Mp) is an invariant subspace under S(n(1),...,n(p)) ⊗ IE1⊗···⊗Ep . Using Theorem 3.1, we
obtain
m(cid:0)⊗k
i=1Mi(cid:1) = lim
= lim
= lim
trace [Q1 ⊗ · · · ⊗ Qp]
trace(cid:2)PU(M1⊗···⊗Mp)(cid:0)Q1 ⊗ · · · ⊗ Qp ⊗ IE1⊗···⊗Ep(cid:1)(cid:3)
trace(cid:2)U ∗PU(M1⊗···⊗Mp)U U ∗(cid:0)Q1 ⊗ · · · ⊗ Qp ⊗ IE1⊗···⊗Ep(cid:1) U(cid:3)
trace(cid:8)PM1⊗···⊗Mp(cid:2)(Q1 ⊗ IE1) ⊗ · · · ⊗(cid:0)Qp ⊗ IEp(cid:1)(cid:3)(cid:9)
trace [Q1 ⊗ · · · ⊗ Qp]
trace [Q1 ⊗ · · · ⊗ Qp]
1 , . . . , q(1)
κ(1), . . . , q(p)
1 , . . . , q(p)
κ(p)) ∈ Zκ(1)+···+κ(p)
+
. Note that the latter limit
where the limit is taken over (q(1)
is equal to the product
lim
1 ,...,q(1)
κ(1))∈Z
(q(1)
κ(1)
+
trace [PM1 (Q1 ⊗ IE1 )]
trace [Q1]
· · ·
lim
1 ,...,q(p)
κ(p))∈Z
(q(p)
κ(p)
+
i ∈ {1, . . . , p}, set
which, due to Theorem 3.1, is equal toQp
i (cid:16)S(n(i)) ⊗ IEi(cid:17) M⊥
M(i) := PM⊥
i
i=1 m (Mi). Therefore, m (⊗p
i=1 m (Mi) . For each
and M := PM⊥(cid:16)S(n(1),...,n(p)) ⊗ IE1⊗···⊗Ep(cid:17) M⊥ .
Since, due to Thorem 3.1, curv(M) = dim(E1 ⊗ · · · ⊗ Ep) − m(M) and curv(M(i)) = dim(Ei) − m(Mi),
we deduce the corresponding identity for the curvature. The proof is complete.
(cid:3)
trace [Qp]
trace(cid:2)PMp(cid:0)Qp ⊗ IEp(cid:1)(cid:3)
i=1Mi) =Qp
,
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
19
Note that, in particular, if M ⊂ H 2(Dn) ⊗ Cr and N ⊂ H 2(Dp) ⊗ Cq are invariant subspaces,
then so is M ⊗ N ⊂ H 2(Dn+p) ⊗ Crq and the multiplicity invariant has the multiplicative property
m(M ⊗ N ) = m(M)m(N ).
We recall that a Beurling type characterization of the invariant subspaces of the full Fock space was
obtained in [18].
Corollary 3.5. For each i ∈ {1, . . . , k}, let Mi be an invariant subspace of F 2(Hni). Then the tensor
product M := ⊗k
i=1Mi is an invariant subspace of ⊗k
i=1F 2(Hni ) and
curv(PM⊥ SM⊥ ) = 1 −
kYi=1(cid:16)1 − curvi(PM⊥
i
SiM⊥
i
)(cid:17)
and m(cid:0)⊗k
i=1Mi(cid:1) =
kYi=1
mi (Mi) ,
where curvi and mi are the curvature and the multiplicity with respect to F 2(Hni ), respectively.
i=1F 2(Hni ) ⊗ E such that they do not contain
Theorem 3.6. Let M, N be invariant subspaces of ⊗k
nontrivial reducing subspaces for the universal model S ⊗ IE . If PM⊥(S ⊗ IE )M⊥ is unitarily equivalent
to PN ⊥ (S ⊗ IE )N ⊥ , then there is a unitary operator U ∈ B(E) such that
(I ⊗ U )PM = PN (I ⊗ U ).
Conversely, if M, N are Beurling type invariant subspaces and there is a unitary operator U ∈ B(E)
such that (I ⊗ U )PM = PN (I ⊗ U ), then PM⊥ (S ⊗ IE )M⊥ is unitarily equivalent to PN ⊥(S ⊗ IE )N ⊥ .
Proof. Let M, N be invariant subspaces of ⊗k
reducing subspaces for S ⊗ I. Denote S(α) := S1,α1 · · · Sk,αk if (α) := (α1, . . . , αk) is in F+
and note that the Cuntz-Toeplitz algebra [6] generated by the shifts Si,j satisfies the relation
i=1F 2(Hni ) ⊗ E such that they do not contain nontrivial
n1 × · · · × F+
nk ,
C∗(Si,j) = span{S(α)S∗
(β) : (α), (β) ∈ F+
n1 × · · · × F+
nk }.
(β)) := PM⊥ S(α)S∗
(β)M⊥ and Ψ2(S(α)S∗
nk , respectively. Consider the ∗-representations π1 : C∗(Si,j ) → B(⊗k
Let Ψ1 : C∗(Si,j ) → B(M⊥) and Ψ2 : C∗(Si,j ) → B(N ⊥) be the unital completely positive linear maps
defined by Ψ1(S(α)S∗
(β)N ⊥ for any (α), (β) ∈
i=1F 2(Hni ) ⊗ E) and
n1 × · · · × F+
F+
π2 : C∗(Si,j) → B(⊗k
i=1F 2(Hni) ⊗ E) defined by π1(X) := X ⊗ IE and π2(X) := X ⊗ IE for any X ∈
C∗(Si,j), respectively. Note that Ψs(X) = V ∗
i=1F 2(Hni ) ⊗ E
and V2 : N ⊥ → ⊗k
i=1F 2(Hni) ⊗ E are the injection maps. Note that if M is an invariant subspace of
i=1F 2(Hni ) ⊗ H, then M does not contain nontrivial reducing subspaces for S ⊗ IH if and only if M⊥
⊗k
is a cyclic subspace for S ⊗ IH, i.e.
s πs(X)Vs for s = 1, 2, where V1 : M⊥ → ⊗k
(β)) := PN ⊥ S(α)S∗
span(cid:8)(S1,β1 · · · Sk,βk ⊗ IE ) M⊥ : β1 ∈ F+
n1
nk(cid:9) = ⊗k
, . . . , βk ∈ F+
i=1F 2(Hni ) ⊗ H.
Consequently, π1 and π2 are minimal Stinespring dilations of the unital completely positive linear maps
Ψ1 and Ψ2, respectively. Now, assume that there is a unitary operator Z : M⊥ → N ⊥ such that
ZPM⊥(Si,j ⊗ IE )M⊥ = PN ⊥ (Si,j ⊗ IE )N ⊥ Z for any i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}. It is easy to
see that ZΨ1(X) = Ψ2(X)Z for any X ∈ C∗(Si,j). Now, using standard arguments concerning the
uniqueness of minimal dilations of completely positive maps of C∗-algebras (see [1]), we deduce that
i=1F 2(Hni ) ⊗ E) such that W π1(X) = π2(X)W for any
there is a unique unitary operator W ∈ B(⊗k
X ∈ C∗(Si,j), and W V1 = V2Z. Consequently, W V1V ∗
2 , which is equivalent to
W PM = PN W . Since C∗(Si,j) is irreducible (see Lemma 5.5 from [25]) we must have W = I ⊗ U where
U ∈ B(E) is a unitary operator.
1 W ∗ = V2U U ∗V ∗
2 = V2V ∗
Now, we prove the converse. Assume that U ∈ B(E) is a unitary operator such that (I ⊗ U )PM =
PN (I ⊗ U ), where M, N are Beurling type invariant subspaces of ⊗k
i=1F 2(Hni ) ⊗ E such that they do not
contain nontrivial reducing subspaces for the universal model S ⊗ IE . Therefore, we can find isometric
multi-analytic operators ψs : ⊗k
1 = PM and
ψ2ψ∗
1 = ψψ∗. Since ψ1, ψ
are multi-analytic operators and PC = (id − ΦS1) · · · (id − ΦSk )(I), we obtain
2 = PN . Consequently, setting ψ := (I ⊗ U ∗)ψ2, the relations above imply ψ1ψ∗
i=1F 2(Hni ) ⊗ E, s = 1, 2, such that ψ1ψ∗
i=1F 2(Hni ) ⊗ Es → ⊗k
k(PC ⊗ IE1 )ψ∗
1 xk = k(PC ⊗ IE2 )ψ∗xk,
x ∈ ⊗k
i=1F 2(Hni ) ⊗ E.
20
GELU POPESCU
1 x) := (PC⊗IE2)ψ∗x for x ∈ ⊗k
Define the operator A ∈ B(E1, E2) by setting A((PC ⊗IE1 )ψ∗
i=1F 2(Hni)⊗E.
Since ψ1 is an isometric multi-analytic operator and M has no nontrivial reducing subspaces for the
universal model S ⊗ IE , we must have E1 = (PC ⊗ IE1 )ψ∗
i=1F 2(Hni ) ⊗ E). A similar result holds for
E2. Therefore, the operator A is unitary and ψ(PC ⊗ IE2 ) = ψ1(PC ⊗ IE1 )A∗. Hence, ψAC⊗E1 = ψ1C⊗E1
and, due to the analyticity of ψ and ψ1, we have ψ(I ⊗ A) = ψ1. Since ψ := (I ⊗ U ∗)ψ2, we deduce that
(I ⊗ U )ψ1 = ψ2(I ⊗ A). Now, note that the unitary operator I ⊗ U takes M onto N and M⊥ onto N ⊥.
Since (S∗
i,j ⊗ IE , we
deduce that (S∗
i,j ⊗ IE ) and M⊥, N ⊥ are invariant subspaces under S∗
i,j ⊗ IE )N ⊥ , where Λ := (I ⊗ U )M⊥. Consequently,
i,j ⊗ IE )(I ⊗ U ∗) = (I ⊗ U ∗)(S∗
i,j ⊗ IE )M⊥ Λ∗ = Λ∗(S∗
1(⊗k
ΛPM⊥(Si,j ⊗ IE )M⊥ = PN ⊥(Si,j ⊗ IE )N ⊥Λ
for any i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}. This completes the proof.
(cid:3)
We remark that when M, N are nontrivial invariant subspaces of ⊗k
that PM⊥ SM⊥ is unitarily equivalent to PN ⊥ SN ⊥ if and only M = N .
i=1F 2(Hni), Theorem 3.6 shows
Theorem 3.7. Let n = (n1, . . . , nk) ∈ Nk be such that ni ≥ 2 and nj ≥ 2 for some i, j ∈ {1, . . . , k},
i 6= j. Then, for each t ∈ (0, 1), there exists an uncountable family {T (ω)(t)}ω∈Ω of pure elements in the
regular polyball with the following properties:
(i) T (ω)(t) in not unitarily equivalent to T (σ)(t) for any ω, σ ∈ Ω, ω 6= σ.
(ii) rank [T (ω)(t)] = 1 and curv [T (ω)(t)] = t for all ω ∈ Ω.
Proof. Assume that ni ≥ 2. If t ∈ [0, 1), there exists a subsequence of natural numbers {kp}N
k1 < k2 < · · · , where N ∈ N or N = ∞, and dp ∈ {1, 2, . . . , ni − 1}, such that 1 − t =PN
ni is the unital free semigroup on ni generators gi
ni and the identity gi
1, . . . , gi
dp
kp
n
i
p=1
0. Define the following
p=1, 1 ≤
. Recall
ni )kp−1 , (gi
2)kp−kp−1(gi
ni )kp−1 , . . . , (gi
dp)kp−kp−1(gi
p = 2, 3, . . . , N.
ni )kp−1o ,
is an invariant subspace of F 2(Hni ). If kp ≤ qi < kp+1, then we have
that F+
subsets of F+
ni :
J i
J 1
1 :=(cid:8)(gi
p :=n(gi
It is clear that
(3.2)
1)k1 , . . . , (gi
1)kp−kp−1(g1
d1)k1(cid:9) ,
tracehPMi(t)⊥ P (i)
qi i
qi i
tracehP (i)
= 1 −
= 1 −
βiE
qi
1
, ei
ei
βi
p=1J i
p
F 2(Hni ) ⊗ ei
β
Mi(t) := Mβ∈∪N
tracehPMi(t)P (i)
qi i
qi i
tracehP (i)
niDPMi(t)P (i)
qi i Xβi∈F+
tracehP (i)
βiE
DP (i)
= 1 − d1
qi i
tracehPMi(t)P (i)
qi i
tracehP (i)
qi PMi(t)P (i)
+ · · · + dpn
d1nqi−k1
qi−kp
i
qi ei
βi
= 1 −
nqi
i
nqi
i
, ei
i
nk1
i
= 1 − Xβi∈F+
ni
= 1 −
kPMi(t)ei
βik2
1
nqi
ni ,βi=qi
i Xβi∈F+
i ! .
dp
kp
+ · · · +
n
= 1 −
NXp=1
dp
kp
1
n
= t.
Hence and using Theorem 3.1, we deduce that
curvi[PMi(t)⊥ SiMi(t)⊥] = 1 − lim
qi→∞
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
21
Fix t ∈ (0, 1), assume that n1, n2 ≥ 2, and let ω ∈ (1 − t, 1). Set
N1(ω) := M1(1 − ω)
and define
and N2(ω, t) := M2(cid:18)1 −
1 − t
ω (cid:19)
M(ω) := N1(ω) ⊗ N2(ω, t) ⊗ F 2(Hn3 ) ⊗ · · · ⊗ F 2(Hnk ),
which is an invariant subspace under Si,j for any i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}. Defining T (ω)(t) :=
PM(ω) ⊥ SM(ω) ⊥ and using Corollary 3.5, we deduce that
curv [T (ω)(t)] = 1 −(cid:2)1 − curv1(cid:0)PN1(ω)⊥ S1N1(ω)⊥(cid:1)(cid:3)(cid:2)1 − curv2(cid:0)PN2(ω,t)⊥ S2N2(ω,t)⊥(cid:1)(cid:3)
1 − t
= 1 − [1 − (1 − ω)](cid:20)1 −(cid:18)1 −
ω (cid:19)(cid:21) = t.
Since the curvature is a unitary invariant and curv2[PN2(ω,t)⊥ S2N2(ω,t)⊥] = 1 − 1−t
ω , we deduce that
PN2(ω,t)⊥S2N2(ω,t)⊥ is not unitary equivalent to PN2(σ,t)⊥S2N2(σ,t)⊥ if ω, σ ∈ (1−t, 1) and ω 6= σ. Due to
Theorem 3.6, when E = C, we deduce that N2(ω, t) 6= N2(σ, t), which implies M(ω)(t) 6= M(σ)(t). Using
again Theorem 3.6, we conclude that T (ω)(t) in not unitarily equivalent to T (σ)(t) for any ω, σ ∈ (1− t, 1),
ω 6= σ. The fact that rank [T (ω)(t)] = 1 is obvious. The proof is complete.
(cid:3)
We remark that Theorem 3.7 shows, in particular, that the curvature is not a complete invariant.
Corollary 3.8. Let n = (n1, . . . , nk) ∈ Nk be such that ni ≥ 2 and nj ≥ 2 for some i, j ∈ {1, . . . , k},
i 6= j. For each s ∈ (0, 1), there is an uncountable family of distinct invariant subspaces {N (ω)}ω∈Ω of
F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) such that m(N (ω)) = s for ω ∈ Ω.
Proof. A closer look at the proof of Theorem 3.7 and using Corollary 3.5, reveals that mi(Mi(t)) = 1 − t
for any t ∈ [0, 1), where Mi(t) is defined by relation (3.2). Fix s ∈ (0, 1), assume that n1, n2 ≥ 2, and let
ω ∈ (s, 1). We define the subspace
N (ω) = M1(1 − ω) ⊗ M2(cid:16)1 −
s
ω(cid:17) ⊗ F 2(Hn3 ) · · · ⊗ F 2(Hnk ),
where M1(1 − ω) ⊂ F 2(Hn1 ) and M2(cid:0)1 − s
(3.2). Using again Corollary 3.5, we obtain m(N (ω)) = s for all ω ∈ (s, 1). As in the proof of Theorem
3.7, one can show that {N (ω)}ω∈(s,1) are distinct invariant subspaces.
(cid:3)
ω(cid:1) ⊂ F 2(Hn2 ) are invariant subspaces defined by relation
We remark that in the particular case when n1 = · · · = nk = 1, Fang proved in [11] that the multiplicity
of any invariant subspace of H(Dk) is always a non-negative integer. When at least one ni is ≥ 2, we
have the following result.
Corollary 3.9. Let n = (n1, . . . , nk) ∈ Nk be such that there exists at least one number ni ≥ 2 and let
t ∈ [0, 1]. Then there exists a pure element T in the polyball Bn(H) such that rank (T) = 1 and
Moreover, for each s ∈ [0, 1] there is an invariant subspace N of F 2(Hn1 )⊗· · ·⊗F 2(Hnk ) with multiplicity
curv(T) = t.
m(N ) = s.
Proof. If t = 1, we have curv(S) = 1. We can assume that n1 ≥ 2. If t ∈ [0, 1), consider the subspace
M(t) := M1(t) ⊗ F 2(Hn2 ) ⊗ · · · ⊗ F 2(Hnk ), where M1(t) is defined by relation (3.2). As in the proof of
Theorem 3.7, one can show that curv(PM(t)⊥ SM(t)⊥) = t. This implies m(M(t)) = 1 − t and completes
the proof.
(cid:3)
Using Corollary 3.9 and tensoring with the identity on Cm, one can see that for any t ∈ [0, m], there
exists a pure element T in the polyball with rank (T) = m and curv(T) = t.
We remark that due to Theorem 3.1, if M is a proper invariant subspace of ⊗k
i=1F 2(Hni ) of finite
codimension, then
curv(PM⊥ SM⊥) = 0
and m(M) = 1.
However, we have the following result.
22
GELU POPESCU
Proposition 3.10. If ni ≥ 2 for at least one i ∈ {1, . . . , k}, then there exist invariant subspaces M of
⊗k
i=1F 2(Hni ) of infinite codimension such that curv(PM⊥ SM⊥) = 0.
Proof. We can assume that n1 ≥ 2. Let M1 be the invariant subspace of F 2(Hn1 ) such that M⊥
closed span of the vectors (e1
1)s, where s = 0, 1, . . . . Note that
1 is the
curv1(PM⊥
1
SM⊥
1
) = 1 − lim
q1→∞
1
nq1
1 Xβ∈F+
n1 ,β=q1
kP (1)
M1
e1
βk2 = 1 − lim
q1→∞
nq1
1 − 1
nq1
1
= 0.
For each i ∈ {2, . . . , k}, let Mi be a proper invariant subspace of ⊗k
SM⊥
curv(PM⊥
where M = ⊗k
i=1F 2(Hni) of finite codimension. Then
) = 0 for i ∈ {2, . . . , k} and, using Corollary 3.5, we deduce that curv(PM⊥ SM⊥ ) = 0,
(cid:3)
i=1Mi. The proof is complete.
i
i
We remark that if T is an element of the regular polyball Bn(H) and M is an invariant subspace
under T, then TM is not necessarily in the polyball. On the other hand, if T has finite rank, TM could
have infinite rank. Indeed, if at least one ni ≥ 2, then there are invariant subspaces M of ⊗k
i=1F 2(Hni )
of infinite codimension such that such that rank (SM) = ∞. To see this, assume that n1 ≥ 2, take
M1 := ⊕∞
if ni ≥ 2 for some i ∈ {1, . . . , k}, then there are Beurling type invariant subspaces M of ⊗k
of infinite codimension with rank (SM) = curv(SM) = 1.
M := [F 2(Hn1 ) ⊗ e1
g1)p(cid:3) and M := M1 ⊗ F 2(Hn2 ) ⊗ · · · ⊗ F 2(Hnk ). We also mention that
i=1F 2(Hni )
Indeed, assume that n1 ≥ 2 and take
p=1(cid:2)F 2(Hn1 ) ⊗ e1
1] ⊗ F 2(Hn2 ) ⊗ · · · ⊗ F 2(Hnk ).
g2 ⊗ (e1
Lemma 3.11. If T ∈ Bn(H) and M is an invariant subspace under T, then TM ∈ Bn(M) if and only
if
for any pi ∈ {0, 1}. If, in addition, T is pure, then the condition above is equivalent to
(id − ΦT1)p1 ◦ · · · ◦ (id − ΦTk )pk (PM) ≥ 0
(id − ΦT1 ) ◦ · · · ◦ (id − ΦTk ) (PM) ≥ 0.
In particular, if M ⊂ ⊗k
(S ⊗ I)M is in the polyball if and only if M is a Beurling type invariant subspace for S ⊗ I.
k=1F 2(Hni ) ⊗ K is an invariant subspace under the universal model S ⊗ I, then
Proof. Since M is an invariant subspace under T, we have
(cid:0)id − ΦT1M(cid:1)p1 ◦ · · · ◦(cid:0)id − ΦTkM(cid:1)pk (IM) = [(id − ΦT1 )p1 ◦ · · · ◦ (id − ΦTk )pk (PM)]M.
Since the direct implication is obvious, we prove the converse. Assume that
(cid:0)id − ΦT1M(cid:1)p1 ◦ · · · ◦(cid:0)id − ΦTkM(cid:1)pk (IM) ≥ 0.
Let h ∈ H and consider the orthogonal decomposition h = x + y with x ∈ M and y ∈ M⊥. Using the
fact that M⊥ is invariant subspace under each operator T ∗
i,j, we have
h(id − ΦT1 )p1 ◦ · · · ◦ (id − ΦTk )pk (PM)(x + y), x + yi
= h(id − ΦT1 )p1 ◦ · · · ◦ (id − ΦTk )pk (PM)x, x + yi + h(id − ΦT1 )p1 ◦ · · · ◦ (id − ΦTk )pk (PM)y, x + yi
= h(id − ΦT1 )p1 ◦ · · · ◦ (id − ΦTk )pk (PM)x, xi + kyk2 ≥ 0.
Consequently, (id − ΦT1 )p1 ◦ · · · ◦ (id − ΦTk )pk (PM) ≥ 0. If, in addition, T is pure and (id − ΦT1 ) ◦ · · · ◦
(id − ΦTk ) (PM) ≥ 0, then, since ΦT1 is positive linear map, we deduce that
Φm
T1 (id − ΦT2) ◦ · · · ◦ (id − ΦTk ) (PM) ≤ (id − ΦT2 ) ◦ · · · ◦ (id − ΦTk ) (PM)
for any m ∈ N. Taking into account that Φm
T1 (I) → 0 as m → ∞, we obtain (id − ΦT2 ) ◦ · · · ◦
(id − ΦTk ) (PM) ≥ 0. Similarly, one can show that (id − ΦT1 )p1 ◦ · · · ◦ (id − ΦTk )pk (PM) ≥ 0 for any
pi ∈ {0, 1}. Now, using Theorem 5.2 from [25], the last part of this lemma follows.
(cid:3)
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
23
Proposition 3.12. Let M be a proper invariant subspace of ⊗k
curv(PM⊥ SM⊥) = 0 if and only if there is an inner sequence {ψs}∞
⊗k
s=1 ψsψ∗
s where the convergence is in the strong operator topology, and
i=1F 2(Hni). Then SM ∈ Bn(M) and
s=1 for M, i.e., ψs are multipliers of
i=1F 2(Hni ), PM =P∞
lim
(q1,...,qk)∈Zk
+
1
1 · · · nqk
nq1
k
Xαi=qi,i∈{1,...,k}
∞Xs=1
kψs(e1
α1 ⊗ · · · ⊗ ek
αk k2 = 1.
Proof. According to Lemma 3.11, condition SM ∈ Bn(M) holds if and only if M is a Beurling type
i=1F 2(Hni )
invariant subspace. Therefore, there exist multi-analytic operators ψp : ⊗k
i=1F 2(Hni ) → ⊗k
p=1 ψpψ∗
p. Due to Theorem 3.1, condition curv(PM⊥ SM⊥ ) = 0 is equivalent to
such that PM =P∞
lim
(q1,...,qk)∈Zk
+
traceh(cid:16)P∞
p=1 ψpψ∗
tracehP (1)
p(cid:17) (P (1)
q1 ⊗ · · · ⊗ P (k)
qk i
q1 ⊗ · · · ⊗ P (k)
qk )i
= 1.
(cid:3)
Now, one can easily complete the proof.
i=1F 2(Hni )⊗E such that it does not contain nontrivial
Lemma 3.13. Let M be an invariant subspace of ⊗k
reducing subspaces for the universal model S ⊗ IE , and let T := PM⊥ (S ⊗ IE)M⊥ . Then there is a unitary
operator Z : ∆T(I)(H) → E such that (I ⊗ Z)KT = V , where KT is the Berezin kernel associated with
T and V is the injection of M⊥ into ⊗k
i=1F 2(Hni ) ⊗ E.
Moreover, T has characteristic function if and only if M is a Beurling type invariant subspace.
Proof. Note that M⊥ is a cyclic subspace for S ⊗ IE . Therefore, S ⊗ IE is a minimal isometric dilation
of T := PM⊥ (S ⊗ IE )M⊥ . On the other hand, according to [25] (see Theorem 5.6 and its proof) if
KT : H → F 2(Hn1 ) ⊗ · · · ⊗ F 2(Hnk ) ⊗ ∆T(I)(H)
is the noncommutative Berezin kernel, then the subspace Kf ,TH is co-invariant under each operator
Si,j ⊗ I∆TH for any i ∈ {1, . . . , k}, j ∈ {1, . . . , ni} and the dilation provided by the relation
T(α) = K∗
T(S(α) ⊗ I∆T(I)(H))KT,
(α) ∈ F+
n1 × · · · × F+
nk
,
is minimal. Moreover, since the Cuntz-Toeplitz algebra [6] satisfies the relation
nk },
C∗(Si,j) = span{S(α)S∗
(β) : (α), (β) ∈ F+
n1 × · · · × F+
the two dilations mentioned above coincide up to an isomorphism. Consequently, there is a unitary
operator Z : ∆T(I)(H) → E such that (I ⊗ Z)KT = V , where KT is the Berezin kernel associated with
T = (I ⊗ Z ∗)PM⊥(I ⊗ Z) and
T and V is the injection of M⊥ into ⊗k
i=1F 2(Hni ) ⊗ E. Note that KTK∗
∆S⊗IE (I − KTK∗
T) = (I ⊗ Z ∗)∆S⊗IE (PM)(I ⊗ Z).
As in the proof of Lemma 3.11, M is a Beurling type invariant subspace if and only if ∆S⊗IE (PM) ≥ 0.
Using the identity above, we deduce that ∆S(I − KTK∗
T) ≥ 0, which, due to Theorem 6.2 from [25], is
equivalent to T having characteristic function. The proof is complete.
(cid:3)
Proposition 3.14. The following statements hold.
(i) If M is a proper Beurling type invariant subspace of ⊗k
0 ≤ curv(PM⊥ SM⊥) < 1
and
i=1F 2(Hni ), then
0 < m(M) ≤ 1.
(ii) If Mi 6= {0}, i ∈ {1, . . . , k}, is an invariant subspace of F 2(Hni ), then M := ⊗k
i=1Mi is an
invariant subspace of ⊗k
i=1F 2(Hni ) and curv(PM⊥ SM⊥ ) < 1.
Proof. To prove item (i), note that Lemma 3.13 shows that there is a unitary operator Z : ∆T(I)(M⊥) →
C such that (I ⊗ Z)KT = V , where KT is the Berezin kernel associated with T := PM⊥ SM⊥ and V is
the injection of M⊥ into ⊗k
T = (I ⊗ Z ∗)PM⊥ (I ⊗ Z) and
i=1F 2(Hni ). Hence, we deduce that KTK∗
∆S⊗IC(I − KTK∗
T) = (I ⊗ Z ∗)∆S(PM)(I ⊗ Z).
Since M is a Beurling type invariant subspace, we have ∆S(PM) ≥ 0, which implies ∆S(I − KTK∗
T) ≥ 0.
If we assume that curv(T) = 1, then curv(T) = rank (T) and, due to Theorem 2.7, T is unitarily
24
GELU POPESCU
equivalent to S. Consequently, M⊥ is an invariant subspace for Si,j and, therefore, reducing for Si,j.
Since the C∗-algebra C∗(Si,j ) is irreducible, we get a contradiction. Now, we prove part (ii). Since any
invariant subspace of F 2(Hni) is of Beurling type (see [18]), we can apply part (i), when k = 1, and
deduce that curvi(PM⊥
(cid:3)
) < 1. Now, using Proposition 3.5, one can complete the proof.
SiM⊥
i
i
We remark that Proposition 3.14 implies that PM⊥ SM⊥ is not unitarily equivalent to the universal
It remains an open question whether Proposition 3.14 part (i) is true for arbitrary proper
model S.
invariant subspace of ⊗k
i=1F 2(Hni ), which is the case when k = 1 (see [21]).
4. Stability, continuity, and multiplicative properties
In this section, we provide several properties concerning the stability, continuity, and multiplicative
properties for the curvature and the multiplicity invariants.
Theorem 4.1. Let T ∈ Bn(H) have finite rank and let M be an invariant subspace under T such that
TM ∈ Bn(M) and dim M⊥ < ∞. Then TM has finite rank and
curv(T) − curv(TM) ≤ dim M⊥
(ni − 1).
kYi=1
Proof. Note that rank (TM) = rank ∆T(PM). Since ∆T(PM) = ∆T(IH) − ∆T(PM⊥), we deduce that
rank (TM) < ∞. Since M is an invariant subspace under T, we have
Using Corollary 1.4, we deduce that curv(T) = curv(TM) if ni = 1 for at least one i ∈ {1, . . . , k}. When
ni ≥ 2, we obtain the desired inequality. The proof is complete.
(cid:3)
Theorem 4.2. Let T ∈ Bn(H) have finite rank and let M be a co-invariant subspace under T with
dim M⊥ < ∞. Then PMTM has finite rank and
curv(T) = curv(PMTM).
Proof. Set A := (A1, . . . , Ak) and Ai := (Ai,1, . . . , Ai,ni ), where Ai,j := PMTi,jM for i ∈ {1, . . . , k} and
j ∈ {1, . . . , ni}. Since Φqi
Ai
(IM) = PMΦqi
Ti
(IH)M, we have
traceh(cid:16)id − Φq1+1
≤ tracehPM(cid:16)id − Φq1+1
≤ traceh(cid:16)id − Φq1+1
Ak (cid:17) (IM)i
A1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Tk (cid:17) (IH)i + tracehPM⊥(cid:16)id − Φq1+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Ak (cid:17) (IM)i + tracehPM⊥(cid:16)id − Φq1+1
A1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Tk (cid:17) (IH)i
Tk (cid:17) (IH)i .
Note also that
traceh(cid:16)id − Φq1+1
T1M(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
= traceh(cid:16)id − Φq1+1
≤ traceh(cid:16)id − Φq1+1
traceh(cid:16)id − Φq1+1
traceh(cid:16)id − Φq1+1
TkM(cid:17) (IM)i
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Tk (cid:17) (PM)i
Tk (cid:17) (IH)i + traceh(cid:16)id − Φq1+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Tk (cid:17) (PM⊥ )i ≤ (1 + nq1+1
Tk (cid:17) (IH)i
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
T1M(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
traceh(cid:16)id − Φq1+1
Qk
i=1(1 + ni + · · · + nqi
i )
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
) · · · (1 + nqk+1
−
k
1
Consequently, we have
)trace [PM⊥ ].
trace [PM⊥ ].
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(1 + nq1+1
i=1(1 + ni + · · · + nqi
i )
) · · · (1 + nqk+1
)
i=1(1 + ni + · · · + nqi
i )
k
1
≤
Qk
Qk
Tk (cid:17) (PM⊥ )i .
TkM(cid:17) (IM)i
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
25
Consequently, we have
On the other hand, since(cid:16)id − Φq1+1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
tracehPM⊥(cid:16)id − Φq1+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
traceh(cid:16)id − Φq1+1
i=1(1 + ni + · · · + nqi
i )
≤
i=1(1 + ni + · · · + nqi
i )
1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Tk (cid:17) (IH)i
−
dim M⊥.
Qk
Qk
Tk (cid:17) (IH) ≤ I, we deduce that
Tk (cid:17) (IH)i ≤ dim M⊥.
A1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
traceh(cid:16)id − Φq1+1
Qk
i=1(1 + ni + · · · + nqi
i )
Ak (cid:17) (IM)i
Using Corollary 1.4, we deduce that curv(T) = curv(A). The proof is complete.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:3)
We denote by A the set of all k-tuples φ = (φ1, . . . , φk) of commuting positive linear maps on B(H)
such that φi(T +(H)) ⊂ T +(H) and trace [φi(X)] ≤ trace (X) for any X ∈ T +(H) and i ∈ {1, . . . , k}.
Lemma 4.3. Let φ = (φ1, . . . , φk) and φ(m) = (φ1,m, . . . , φk,m), m ∈ N, be in A and let X and Xm be
in T +(H). If
lim
m→∞
1,m ◦ · · · ◦ φqk
1 ◦ · · · ◦ φqk
k (X)]
for each q = (q1, . . . , qk) ∈ Zk
m→∞ ( lim
lim sup
q∈Zk
+
+, then
tracehφq1
tracehφq1
k,m(Xm)i = trace [φq1
k,m(Xm)i) ≤ lim
q∈Zk
+
1,m ◦ · · · ◦ φqk
trace [φq1
1 ◦ · · · ◦ φqk
k (X)] .
Proof. For each q = (q1, . . . , qk) ∈ Zk
+, m ∈ N, let
xq := trace [φq1
1 ◦ · · · ◦ φqk
k (X)]
and
x(m)
q
:= tracehφq1
1,m ◦ · · · ◦ φqk
k,m(Xm)i .
Ti
qk ⊗ IK)KT(m )i
Ti
and {x(m)
+
According to Lemma 1.1, {xq}q∈Zk
are decreasing multi-sequences with respect to each of
x(m)
q . We prove that lim supm→∞ x(m) ≤
the indices q1, . . . , qk. Set x := limq∈Zk
x, by contradiction. Passing to a subsequence we can assume that there is ǫ > 0 such that x(m) − x ≥
2ǫ > 0 for any m ∈ N. Since x := limq∈Zk
xq, we can choose N ≥ 1 such that xq − x < ǫ for any
q = (q1, . . . , qk) ∈ Zk
xq and x(m) := limq∈Zk
q ≥ x(m) and
q }q∈Zk
+ with qi ≥ N . Fix such a q and note that x(m)
x(m)
q − xq ≥ (x(m)
q − x(m)) + (x(m) − x) − xq − x ≥ 2ǫ − ǫ = ǫ.
+
+
+
+
This contradicts the hypothesis that limm→∞ x(m)
q = xq. The proof is complete.
(cid:3)
Theorem 4.4. Let T and {T(m)}m∈N be elements in the polyball Bn(H) such that they have finite ranks
which are uniformly bounded and WOT- limm→∞ KT(m) K∗
T(m) = KTK∗
T. Then
lim sup
m→∞
curv(T(m)) ≤ curv(T).
Proof. Due to the hypothesis, we can assume that KT(m) and KT are taking values in ⊗k
where K is a Hilbert space with dim K < ∞. Consequently, we have
i=1F 2(Hni ) ⊗ K,
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IK)KT(m ) K∗
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IK)KTK∗
T(m )i = traceh(P (1)
for any q = (q1, . . . , qk) ∈ Zk
+. Due to Theorem 1.3 and its proof, we have
lim
m→∞
traceh(P (1)
trace(cid:20)Φq1
T (m)
1
lim
m→∞
◦ · · · ◦ Φqk
T (m)
k
(∆T(m )(I))(cid:21) = lim
m→∞
tracehK∗
= traceh(P (1)
= trace(cid:2)Φq1
T1
T(m )(P (1)
q1 ⊗ · · · ⊗ P (k)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IK)KTK∗
◦ · · · ◦ Φqk
Tk
(∆T(I))(cid:3)
we obtain
lim sup
m→∞
trace(cid:20)Φq1
T (m)
1
lim
q∈Zk
+
◦ · · · ◦ Φqk
T (m)
k
nq1
1 · · · nqk
k
(∆T(m ) (I))(cid:21)
k (cid:19) , Xm = ∆T(m ) (I),
trace(cid:2)Φq1
≤ lim
q∈Zk
+
T1
◦ · · · ◦ Φqk
Tk
nq1
1 · · · nqk
k
(∆T(I))(cid:3)
.
26
GELU POPESCU
for any q = (q1, . . . , qk) ∈ Zk
+. Applying Lemma 4.3 when
φ =(cid:18) 1
Φ(m) =(cid:18) 1
n1
ΦT1 , . . . ,
1
nk
n1
ΦTk(cid:19) , X = ∆T(I),
ΦT (m)
1
, . . . ,
ΦT (m)
1
nk
Using again Theorem 1.3, we conclude that lim supm→∞ curv(T(m)) ≤ curv(T). The proof is complete.
(cid:3)
The next result shows that the multiplicity invariant is lower semi-continuous.
Theorem 4.5. Let M and Mm be invariant subspaces of ⊗k
WOT- limm→∞ PMm = PM, then
i=1F 2(Hni ) ⊗ E with dim E < ∞.
If
lim inf
m→∞
m(Mm) ≥ m(M).
Proof. Let M := (M1, . . . , Mk) with Mi := (Mi,1, . . . , Mi,ni ) and Mi,j := PM⊥ (Si,j ⊗ IE )M⊥ . Similarly,
we define M(m) := (M (m)
). Due to Theorem 3.1, we have
, . . . , M (m)
1
k
(4.1)
m(M) = dim E − curv(M)
and m(Mm) = dim E − curv(M(m)).
As in the proof of the same theorem, we have
trace [Φq1
M1
◦ · · · ◦ Φqk
Mk
(∆M(IM⊥ ))] = trace [PM⊥ (P (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IE )]
and a similar relation associated with M(m) holds. Since WOT- limm→∞ PMm = PM, we deduce that
lim
m→∞
trace [PMm (P (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IE )] = trace [PM(P (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IE )].
= nq1
1 · · · nqk
k dim E − trace [PM(P (1)
q1 ⊗ · · · ⊗ P (k)
qk ⊗ IE )]
Consequently, we obtain
trace [Φq1
lim
m→∞
M (m)
1
◦ · · · ◦ Φqk
M (m)
k
(∆M(m)(IM⊥
m
))] = trace [Φq1
M1
◦ · · · ◦ Φqk
Mk
(∆M(IM⊥ ))]
As in the proof of Theorem 4.4, one can use Lemma 4.3 to show that lim supm→∞ curv(M(m)) ≤ curv(M).
Now, relation (4.1) implies lim inf m→∞ m(Mm) ≥ m(M). The proof is complete.
(cid:3)
The next result shows that the curvature is upper semi-continuous.
Theorem 4.6. Let T and {T(m)}m∈N be elements in the polyball Bn(H) such that they have finite ranks
which are uniformly bounded. If T(m) → T, as m → ∞, in the norm topology, then
lim sup
m→∞
curv(T(m)) ≤ curv(T).
Proof. As in the proof of Theorem 4.4, due to Lemma 4.3, it is enough to show that limm→∞ trace (A(m)
trace (Aq) for each q = (q1, . . . , qk) ∈ Zk
Φq1
T1
(∆T(I)). Taking into account that A(m)
q ) =
(∆T(m )(I)) and Aq :=
and Aq have finite rank, we have
+, where A(m)
◦ · · · ◦ Φqk
Tk
◦ · · · ◦ Φqk
:= Φq1
T (m)
T (m)
q
q
k
1
(4.2)
q − Aqkrank [A(m)
q − Aq].
Since T(m) → T, as m → ∞, in the norm topology, it is easy to see that kA(m)
On the other hand, note that there is M > 0 such that rank [∆T(m )(I)] ≤ M for any m ∈ N and
q − Aqk → 0 as m → ∞.
rank [A(m)
q − Aq] ≤ nq1
1 · · · nqk
k (rank [∆T(m )(I)] + rank [∆T(I)]) .
(cid:12)(cid:12)(cid:12)trace [A(m)
q − Aq](cid:12)(cid:12)(cid:12) ≤ kA(m)
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
27
Consequently, relation (4.2) implies trace (A(m)
to the one of Theorem 4.4. The proof is complete.
q ) → trace (Aq) as m → ∞. The rest of the proof is similar
(cid:3)
Given a function κ : N → N and n(i) := (n(i)
polyball Bn(i) (Hi), where Hi is a Hilbert space. Let X(i) ∈ Bn(i) (Hi) with X(i) := (X (i)
X (i)
r
κ(i)) ∈ Nκ(i) for i ∈ {1, . . . , p}, we consider the
κ(i)) and
for r ∈ {1, . . . , κ(i)}. If X := (X(1), . . . , X(p)) ∈ Bn(1) (H1) × · · · ×
) ∈ B(Hi)n(i)
1 , . . . , X (i)
1 , . . . , n(i)
:= (X (i)
r,1, . . . , X (i)
r
r,n(i)
r
Bn(p)(Hp), we define the ampliation eX by setting eX := (eX(1), . . . ,eX(p)), where eX(i) := (eX (i)
and eX (i)
r,1, . . . , eX (i)
:= (eX (i)
) for r ∈ {1, . . . , κ(i)}, and
r,s := IH1 ⊗ · · · ⊗ IHi−1 ⊗ X (i)
r,s ⊗ IHi+1 ⊗ IHp
r,n(i)
r
r
κ(i))
1 , . . . , eX (i)
for all i ∈ {1, . . . , p}, r ∈ {1, . . . , κ(i)}, and s ∈ {1, . . . , n(i)
r }.
eX (i)
Theorem 4.7. Let X := (X(1), . . . , X(p)) ∈ Bn(1) (H1) × · · · × Bn(p)(Hp) be such that each X(i) has trace
class defect. Then the ampliation eX is in the regular polyball B(n(1),...,n(p))(H1 ⊗ · · · ⊗ Hp), has trace class
defect, and
curv (eX) =
pYi=1
curv (X(i)).
Proof. Note that, for each m(i) ∈ Zκ(i)
+ with 0 ≤ m(i) ≤ (1, . . . , 1) and i ∈ {1, . . . , p}, we have
∆(m(1),...,m(p))
(IH1⊗···⊗Hp) = ∆m(1)
X(1) (IH1 ) ⊗ · · · ⊗ ∆m(p)
X(p) (IHp ) ≥ 0,
eX
which shows that eX is in the regular polyball B(n(1),...,n(p))(H1 ⊗ · · · ⊗ Hp) and
trace(cid:2)∆eX(IH1⊗···⊗Hp )(cid:3) = trace [∆X(1) (IH1 )] · · · trace(cid:2)∆X(p) (IHp )(cid:3) < ∞.
+ for i ∈ {1, . . . , p}. According to Corollary 1.4, we have
Let q(i) := (q(i)
1 , . . . , q(i)
κ(i)) ∈ Zκ(i)
curv (eX)
=
lim
q(1)∈Zκ(1)
+ ,...,q(p)∈Zκ(p)
+
trace(cid:20)Φq(1)
1
X (1)
1
trace(cid:20)Φq(1)
1
X (1)
1
◦ · · · ◦ Φ
q(1)
κ(1)
X (1)
κ(1)
(n(1)
1 )q(1)
1
· · · (n(1)
κ(1))q(1)
κ(1)
= lim
q(1)∈Z
κ(1)
+
◦ · · · ◦ Φ
q(1)
κ(1)
X (1)
κ(1)
1 )q(1)
h(n(1)
(∆X(1) (IH1 ))(cid:21)
1
(∆X(1) (IH1 )) ⊗ · · · ⊗ Φq(p)
1
X (p)
1 )q(p)
κ(1)i · · ·h(n(p)
κ(1))q(1)
1
1
· · · (n(1)
◦ · · · ◦ Φ
· · · (n(p)
trace(cid:20)Φq(p)
1
X (p)
1
q(p)
κ(p)
X (p)
κ(p))q(p)
q(p)
κ(p)
X (p)
κ(p)(cid:0)∆X(p) (IHp )(cid:1)(cid:21)
κ(p)i
κ(p)(cid:0)∆X(p)(IHp )(cid:1)(cid:21)
· · · (n(p)
κ(p))q(p)
κ(p)
◦ · · · ◦ Φ
(n(p)
1 )q(p)
1
· · ·
lim
q(p)∈Z
κ(p)
+
=
curv (X(i)).
pYi=1
The proof is complete.
(cid:3)
Proposition 4.8. Let n(i) := (n(i)
for
i ∈ {1, . . . , p}. Then each X(i) is unitarily equivalent to S(n(i)) ⊗ IEi for some Hilbert space Ei with
dim Ei < ∞, if and only if the following conditions are satisfied.
κ(i)) ∈ Nκ(i) and X(i) ∈ B(Hi)n(i)
1 × · · · × B(Hi)n(i)
1 , . . . , n(i)
κ(i)
rank.
(i) The ampliation eX := (eX(1), . . . ,eX(p)) ∈ B(n(1),...,n(p))(H1 ⊗ · · · ⊗ Hp) is a pure element with finite
(ii) Either eX or each eX(i), i ∈ {1, . . . , p}, has characteristic function.
(iii) curv (eX) = rank [∆eX(I)].
28
GELU POPESCU
In this case, the Berezin kernel KeX is a unitary operator which is unitarily equivalent to the tensor
product KX(1) ⊗ · · · ⊗ KX(p) , and
for all i ∈ {1, . . . , p}, r ∈ {1, . . . , κ(i)}, and s ∈ {1, . . . , n(i)
polyball Bn(i) and KX(i) is the associated Berezin kernel of X(i).
X (i)
r,s = K∗
X(i) (S(n(i))
r,s ⊗ I∆
X(i) (I)(Hi))KX(i)
r }, where S(n(i)) is the universal model of the
Proof. The direct implication follows due to Theorem 2.7. We prove the converse. Note that since
eX := (eX(1), . . . ,eX(p)) ∈ B(n(1),...,n(p))(H1 ⊗ · · · ⊗ Hp) is a pure element with finite rank, each element
X(i) is a pure element in the polyball Bn(i) (Hi) and has finite rank. The Berezin kernel of eX,
r=1 F 2(H
n(p)
r
r
KeX : H1 ⊗ · · · ⊗ Hp →(cid:16)⊗κ(1)
can be identified, up to a unitary equivalence, with KX(1) ⊗ · · ·⊗ KX(p) which is acting from H1 ⊗ · · ·⊗ Hp
to
n(1)
r=1 F 2(H
)(cid:17) ⊗ · · · ⊗(cid:16)⊗κ(p)
)(cid:17) ⊗ ∆X(1) (I)(H1)i ⊗ · · · ⊗h(cid:16)⊗κ(p)
)(cid:17) ⊗ ∆eX(I)(H1 ⊗ · · · ⊗ Hp),
)(cid:17) ⊗ ∆X(p) (I)(Hp)i .
Assume that eX has characteristic function. Since curv (eX) = rank ∆eX(I), Theorem 2.7 implies that KeX
is a unitary operator. Consequently, each Berezin kernel KX(i) is a unitary operator and
h(cid:16)⊗κ(1)
r=1 F 2(H
n(1)
r
r=1 F 2(H
n(p)
r
X (i)
r,s = K∗
X(i) (S(n(i))
r,s ⊗ I∆
for all i ∈ {1, . . . , p}, r ∈ {1, . . . , κ(i)}, and s ∈ {1, . . . , n(i)
polyball Bn(i) . Now, assume that each eX(i), i ∈ {1, . . . , p}, has characteristic function. Note that, due
to Theorem 4.7 relation curv (eX) = rank [∆eX(I)] is equivalent
curv (X(i)) =
1 ≤
rank [∆X(i) ].
X(i) (I)(Hi))KX(i)
r }, where S(n(i)) is the universal model of the
pYi=1
pYi=1
Since 1 ≤ curv (X(i)) ≤ rank [∆X(i) ], we deduce that curv (X(i)) = rank [∆X(i) ]. Applying Theorem 2.7
to X(i) ∈ Bn(i) (Hi) for each i ∈ {1, . . . , p}, one can complete the proof.
(cid:3)
Corollary 4.9. Let T1, . . . , Tp be contractions on a Hilbert space H. Then each of them is unitarily
equivalent to a unilateral shift of finite multiplicity if and only if the ampliation
eT := (T1 ⊗ I ⊗ · · · ⊗ I, I ⊗ T2 ⊗ I ⊗ · · · ⊗ I, . . . , I ⊗ · · · ⊗ I ⊗ Tp)
is a pure element of the regular polydisc B(1,...,1)(H⊗· · ·⊗H), has finite rank, and curv (eT ) = rank [∆eT (I)].
5. Commutative polyballs and curvature invariant
In this section, we introduce the curvature invariant associated with the elements of the commutative
n(H) with the property that they have finite rank defects and characteristic functions. There
polyball Bc
are commutative analogues of most of the results from the previous sections.
Since the case n1 = · · · = nk = 1 was considered in the previous sections, we assume, throughout this
section, that the k-tuple n = (n1, . . . , nk) has at least one ni ≥ 2. The commutative polyball Bc
n(H) is the
set of all X = (X1, . . . , Xk) ∈ Bn(H) with Xi = (Xi,1, . . . , Xi,ni ), where the entries Xi,j are commuting
operators. According to [26], the universal model associated with the abstract commutative polyball Bc
n
is the k-tuple B := (B1, . . . , Bk) with Bi = (Bi,1, . . . , Bi,ni), where the operator Bi,j is acting on the
tensor Hilbert space F 2
s (Hnk ) and is defined by setting
s (Hn1 ) ⊗ · · · ⊗ F 2
Bi,j := I ⊗ · · · ⊗ I
⊗Bi,j ⊗ I ⊗ · · · ⊗ I
,
i ∈ {1, . . . , k}, j ∈ {1, . . . , ni},
i − 1 times
k − i times
{z
}
{z
}
where Bi,j := PF 2
We recall that F 2
transforms and model theory on the commutative polyball Bc
s (Hni ) ⊂ F 2(Hni ) is the symmetric Fock space on ni generators.
s (Hni )Si,jF 2
s (Hni) is coinvariant under each operator Si,j. For basic results concerning Berezin
n(H) we refer the reader to [26].
s (Hni ) and F 2
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
29
For each q = (q1, . . . , qk) ∈ Zk
i=1F 2
s (Hni ) by
+, define the operator eN≤q on the tensor product ⊗k
1
s1 ⊗ · · · ⊗ Q(k)
sk
,
eN≤q := X0≤si≤qi
i∈{1,...,k}
tracehQ(1)
s1 ⊗ · · · ⊗ Q(k)
ski Q(1)
:= PF 2
where Q(i)
s (Hni )P (i)
si
Section 1. Note that Q(i)
si
subspace of homogeneous polynomials of degree si and
s (Hni ) for i ∈ {1, . . . , k} and the orthogonal projection P (i)
si
is the orthogonal projection of the symmetric Fock space F 2
s (Hni ) onto its
is defined in
si F 2
trace [Q(i)
si ] =
(si + 1) · · · (si + ni − 1)
(ni − 1)!
=(cid:18)si + ni − 1
ni − 1 (cid:19) .
Lemma 5.1. Let Y be a bounded operator on ⊗k
i=1F 2
s (Hni ) ⊗ H and dim(H) < ∞. Then
traceh(Q(1)
tracehQ(1)
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IH)Yi
qki
= traceh∆B⊗IH (Y )(eN≤q ⊗ IH)i ,
where B is the universal model associated with the abstract commutative polyball Bc
n.
Proof. Note that B := (B1, . . . , Bk) is a pure k-tuple in the polyball Bc
2.2, we have
n(⊗k
i=1F 2
s (Hni )). Due to Corollary
Y =
∞Xs1=0
Φs1
B1⊗IH ∞Xs2=0
B2⊗IH · · ·
Φs2
∞Xsk=0
Φsk
Bk⊗IH
(∆B⊗IH (Y )) · · ·!! ,
where the iterated series converge in the weak operator topology. Setting Qq := Q(1)
q = (q1, . . . , qk) ∈ Zk
+, we deduce that
q1 ⊗ · · · ⊗ Q(k)
qk
for
(Qq ⊗ IH) Y = (Qq ⊗ IH) q1Xs1=0
Φs1
B1⊗IH q2Xs2=0
B2⊗IH · · ·
Φs2
qkXsk=0
Φsk
Bk⊗IH
(∆B⊗IH (Y )) · · ·!!! .
Hence, trace [(Qq ⊗ IH) Y ] is equal to
◦ · · · ◦ Φsk
Bk⊗IH
(∆B⊗IH (Y ))(cid:3)
=
=
=
· · ·
· · ·
· · ·
q1Xs1=0
q1Xs1=0
q1Xs1=0
qkXsk=0
qkXsk=0
qkXsk=0
B1⊗IH
trace(cid:2)(Qq ⊗ IH) Φs1
trace(Qq ⊗ IH) Xα1 ∈F+
trace∆B⊗IH (Y ) Xα1 ∈F+
n1
,···αk ∈F+
nk
α1=s1··· ,αk =sk
,···αk ∈F+
nk
α1=s1··· ,αk =sk
n1
1,α1 ⊗ IH)
(B1,α1 · · · Bk,αk ⊗ IH)∆B⊗IH (Y )(B∗
k,αk · · · B∗
(B∗
k,αk · · · B∗
1,α1
QqB1,α1 · · · Bk,αk ) ⊗ IH .
Since the symmetric Fock space F 2
s (Hni ) is coinvariant under each operator Si,j, one can see that
B∗
i,αi (I ⊗ · · · ⊗ Q(i)
qi ⊗ · · · ⊗ I) = (I ⊗ · · · ⊗ Q(i)
qi−si ⊗ · · · ⊗ I)B∗
i,αi
30
GELU POPESCU
for any i ∈ {1, . . . , k} and αi ∈ F+
ni with αi = si ≤ qi. This can be used to deduce that
Xα1∈F+
,···αk ∈F+
nk
α1=s1··· ,αk =sk
n1
B∗
k,αk · · · B∗
1,α1
QqB1,α1 · · · Bk,αk
= Q(1)
q1−s1 ⊗ · · · ⊗ Q(k)
qk−sk Xα1∈F+
,···αk ∈F+
nk
α1=s1··· ,αk =sk
n1
B∗
k,αk · · · B∗
1,α1
B1,α1 · · · Bk,αk
=
trace [Q(1)
q1 ]
trace [Q(1)
q1−s1 ]
Q(1)
q1−s1 ⊗ · · · ⊗
trace [Q(k)
qk ]
trace [Q(k)
qk−sk ]
Q(k)
qk−sk
.
Putting together the relations above, we obtain
trace [(Qq ⊗ IH) Y ]
=
· · ·
qkXsk=0
q1Xs1=0
= trace(∆B⊗IH (Y ) q1Xs1=0
trace(∆B⊗IH (Y )" trace [Q(1)
trace [Q(1)
q1 ]
q1−s1]
Q(1)
q1−s1 ⊗ · · · ⊗
trace [Q(k)
qk ]
trace [Q(k)
qk−sk ]
qk−sk ⊗ IH#)
Q(k)
trace [Q(1)
q1 ]
trace [Q(1)
s1 ]
Q(1)
s1! ⊗ · · · ⊗ qkXsk=0
trace [Q(k)
qk ]
trace [Q(k)
sk ]
sk! ⊗ IH) .
Q(k)
Consequently, we have
trace [(Qq ⊗ IH) Y ]
trace [Qq]
1
1
= trace(∆B⊗IH (Y )" q1Xs1=0
= trace∆B⊗IH (Y ) X0≤si≤qi
= traceh∆B⊗IH(Y )(eN≤q ⊗ IH)i ,
i∈{1,...,k}
trace [Q(1)
s1 ]
trace [Qs]
1
trace [Q(k)
sk ]
sk! ⊗ IH#)
Q(k)
Q(1)
s1! ⊗ · · · ⊗ qkXsk=0
Qs ⊗ IH
which completes the proof.
(cid:3)
.
Define the bounded linear operator eN on F 2
eN := Xq=(q1,...,qk)∈Zk
+
1
tracehQ(1)
s (Hn1 ) ⊗ · · · ⊗ F 2
s (Hnk ) by setting
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk
qki Q(1)
Theorem 5.2. Let Y be a bounded operator acting on ⊗k
i=1F 2
s (Hni ) ⊗ H and dim H < ∞. If
∆B⊗IH (Y ) := (id − ΦB1⊗IH ) ◦ · · · ◦ (id − ΦBk⊗IH )(Y ) ≥ 0,
then ∆B⊗IH(Y )(eN ⊗ IH) is a trace class operator and
traceh∆B⊗IH (Y )(eN ⊗ IH)i =
lim
q=(q1,...,qk)∈Zk
+
traceh(Q(1)
tracehQ(1)
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IH)Yi
qki
.
Proof. First, note that Lemma 5.1 implies
0 ≤ traceh∆B⊗IH(Y )(eN≤q ⊗ IH)i =
traceh(Q(1)
tracehQ(1)
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IH)Yi
qk i
≤ kY k dim H
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
31
is an increasing multi-sequence of positive operators
for any qi ≥ 0 and i ∈ {1, . . . , k}. Since {eN≤q}q∈Zk
convergent to eN and ∆B⊗IH (Y ) ≥ 0 we deduce that
traceh∆B⊗IH (Y )(eN ⊗ IH)i = lim
+
q∈Zk
+
traceh∆B⊗IH (Y )(eN≤q ⊗ IH)i
traceh∆B⊗IH (Y )(eN≤q ⊗ IH)i .
= sup
q∈Zk
+
The proof is complete.
Given an element T in the commutative polyball Bc
n(H) with the property that it has finite rank and
characteristic function, we introduce its curvature by setting
curvc(T) := lim
m→∞
1
k (cid:19) Xq1≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
traceheK∗
tracehQ(1)
T(Q(1)
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IH)eKTi
qki
.
(cid:3)
i=1F 2
i=1F 2
is complete.
n(H) is the bounded
s (Hni ) ⊗ ∆T(I)(H) defined by
s (Hni ) ⊗ I∆T(I)(H)(cid:17) KT,
Therefore, ∆B⊗IH (Y )(eN ⊗ IH) is a trace class operator and the equality in the theorem holds. The proof
operator eKT : H → ⊗k
We recall from [26] that the constrained Berezin kernel associated with T ∈ Bc
where KT is the noncommutative Berezin kernel associated with T ∈ Bn(H). One can easily see that
the range of KT is in ⊗k
and j ∈ {1, . . . , ni}. An element T ∈ Bc
eKT :=(cid:16)P⊗k
i,j ⊗ I)eKT. for any i ∈ {1, . . . , k}
s (Hni ) ⊗ ∆T(I)(H) and eKTT ∗
there is a multi-analytic operator eΘT : ⊗k
the universal model B, i.e. eΘT(Bi,j ⊗ I) = (Bi,j ⊗ I)eΘT for any i ∈ {1, . . . , k} and j ∈ {1, . . . , ni},
such that eKTeK∗
∆B⊗I(I −eKTeK∗
T)(eN ⊗ IH) the curvature
In the commutative setting, we call the bounded operator ∆B⊗IH (eKTeK∗
T = I . The characteristic function is essentially unique if we restrict it to
its support. We proved in [26] that an element T has constrained characteristic function if and only if
n(H) has characteristic function.
n(H) is said to have constrained characteristic function if
i=1F 2
s (Hni ) ⊗ ∆T(I)(H) with respect
s (Hni ) ⊗ E → ⊗k
T) ≥ 0. We remark that if k = 1, then any element in Bc
T + eΘTeΘ∗
operator associated with T ∈ Bc
n(H) has finite rank and characteristic function, then the curvature operator
i,j = (B∗
i=1F 2
i=1F 2
n(H).
Theorem 5.3. If T ∈ Bc
associated with T is trace class and
traceh∆B⊗IH (eKTeK∗
T)(eN ⊗ IH)i =
lim
(q1,...,qk)∈Zk
+
q1 ⊗ · · · ⊗ Q(k)
traceh(Q(1)
tracehQ(1)
q1 ⊗ · · · ⊗ Q(k)
Ti
qk ⊗ IH)eKTeK∗
qki
T(eN ⊗ IH)i ,
= rank [∆T(I)] − traceheΘT(PC ⊗ I)eΘ∗
strained characteristic function is a multi-analytic operator with respect to B, we obtain
T and taking into account that the con-
where eΘT is the constrained characteristic function of T.
Proof. Applying Theorem 5.2 when Y = I − eKTeK∗
Ti
qk ⊗ IH)eKTeK∗
traceh(Q(1)
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
(q1,...,qk)∈Zk
+
lim
tracehQ(1)
qki
T = eΘTeΘ∗
T)(eN ⊗ IH)i
= traceh∆B⊗I (eKTeK∗
= rank [∆T(I)] − traceh∆B⊗I (eΘTeΘ∗
= rank [∆T(I)] − traceheΘT(PC ⊗ I)eΘ∗
T)(eN ⊗ IH)i
T(eN ⊗ IH)i .
(cid:3)
32
GELU POPESCU
Theorem 5.4. If T ∈ Bc
exists and satisfies the asymptotic formulas
n(H) has finite rank and characteristic function, then the curvature curvc(T)
curvc(T) = lim
m→∞
= lim
m→∞
1
1
T1
q1 +···+qk ≤m
◦ · · · ◦ Φqk
Tk
(∆T(I))(cid:3)
trace(cid:2)Φq1
k (cid:19) Xq1 ≥0,...,qk ≥0
qi(cid:17)
(cid:18)m + k
i=1 trace(cid:16)Q(i)
Qk
trace(cid:2)Φq1
(∆T(I))(cid:3)
k − 1 (cid:19) Xq1 ≥0,...,qk ≥0
qi(cid:17)
(cid:18)m + k − 1
i=1 trace(cid:16)Q(i)
Qk
(∆T(I))(cid:3)
trace(cid:2)Φq1
qi(cid:17)
i=1 trace(cid:16)Q(i)
Qk
traceh(id − Φq1+1
) ◦ · · · ◦ (id − Φqk+1
qn1
1 · · · qnk
◦ · · · ◦ Φqk
Tk
◦ · · · ◦ Φqk
Tk
lim
q1→∞
q1 +···+qk =m
qk→∞
lim
· · ·
Tk
T1
T1
T1
k
=
lim
(q1,...,qk)∈Zk
+
= n1! · · · nk!
)(I)i
.
Proof. Since the range of the Berezin kernel KT is in the Hilbert space ⊗k
i=1F 2
s (Hni) ⊗ ∆T(I)(H) and
eKT :=(cid:16)P⊗k
(5.1)
i=1F 2
T(Q(1)
q1 ⊗ · · · ⊗ Q(k)
s (Hni ) ⊗ I∆T(I)(H)(cid:17) KT, relation (1.3) implies
qk ⊗ IH)eKT = Φq1
traceh(Q(1)
eK∗
xq =
lim
lim
T1
(q1,...,qk)∈Zk
+
(q1,...,qk)∈Zk
+
for any q1, . . . , qk ∈ Z+. Hence, and due to Theorem 5.3, we deduce that
◦ · · · ◦ Φqk
Tk
(∆T(I))
q1 ⊗ · · · ⊗ Q(k)
Ti
qk ⊗ IH)eKTeK∗
,
q1 ⊗ · · · ⊗ Q(k)
tracehQ(1)
qki
where xq :=
(∆T(I))i
tracehΦq1
i=1 trace(cid:16)Q(i)
qi (cid:17)
Qk
+. If X ∈ T +(H), we have
q = (q1, . . . , qk) ∈ Zk
◦···◦Φ
qk
Tk
T1
for q = (q1, . . . , qk) ∈ Zk
+. Let T := (T1, . . . , Tk) ∈ Bc
n(H) and
i,αi = trace Xαi∈F+
ni ,αi=qi
T ∗
i,αi
Ti,αi X
trace (X) ≤ trace [Q(i)
qi ] trace (X).
trace [Φqi
Ti
(X)] = trace Xαi∈F+
≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xαi∈F+
ni ,αi=qi
Ti,αi XT ∗
ni ,αi=qi
i,αi
T ∗
Ti,αi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(∆T(I))(cid:3)
trace(cid:2)Φq1
qi(cid:17)
i=1 trace(cid:16)Q(i)
Qk
◦ · · · ◦ Φqk
Tk
T1
The latter equality was proved by Arveson in [2]. Applying the inequality above repeatedly, we obtain
(5.2)
≤ trace [∆T(I)] .
Due to Theorem 5.2 and Theorem 5.3, the multi-sequence {xq}q=(q1,...,qk)∈Zk
is decreasing with respect
xq exists. Given ǫ > 0, let N0 ∈ N be such that
to each of the indices q1, . . . , qk, and L := limq∈Zk
xq − L < ǫ for any q1 ≥ N0, . . . qk ≥ N0. Consider the sets A1, . . . , Ak, BN0 defined in the proof of
Lemma 1.1. Using relation (5.2), we have xq ≤ trace [∆T(I)] for any q ∈ Zk
+. Hence, we deduce that
+
+
Xq=(q1,...,qk)∈∪k
i=1Ai
xq ≤
kXi=1
card (Ai)trace [∆T(I)] ≤ kN0(cid:18)m + k − 2
k − 2 (cid:19) trace [∆T(I)].
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
33
Now, as in the proof of Lemma 1.1, we obtain that
xq.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
1
< ǫ
q1 +···+qk =m
xq = lim
m→∞
xq − L(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
k − 1 (cid:19) Xq1≥0,...,qk ≥0
(cid:18)m + k − 1
k − 1 (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k − 1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
traceheK∗
q1 ⊗ · · · ⊗ Q(k)
T(Q(1)
lim
m→∞
xq = L.
q1 +···+qk ≤m
q1+···+qk =m
1
q1 ⊗ · · · ⊗ Q(k)
tracehQ(1)
qk ⊗ IH)eKTi
qk i
= L,
for m big enough, which shows that
L := lim
q∈Zk
+
Using Stolz-Ces`aro convergence theorem, we deduce that
Now, using relation (5.1), we conclude
lim
m→∞
1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
q1+···+qk ≤m
and, consequently, the curvature curvc(T) exists. Now, we prove the last equality in the theorem. Since
L = limq1→∞ · · · limqk→∞ xq and setting yq1 := limq2→∞ · · · limqk→∞ xq, an application of Stolz-Ces`aro
convergence theorem to the sequence {yq1 }∞
q1=0 implies
= lim
q1→∞
yq1 = L.
= lim
q2→∞
zq2 .
Similarly, setting zq2 := limq3→∞ · · · limqk→∞
T1
Continuing this process and putting together these results, we obtain
lim
q1→∞
1
s1i
s1=0 tracehQ(1)
Pq1
lim
q2→∞
1
s2i
s2=0 tracehQ(2)
Pq2
L = lim
q1→∞
Note thatPqi
si=0 tracehQ(i)
L = n1! · · · nk!
The proof is complete.
, we deduce that
T1
T1
qk
Tk
· · ·
· · ·
lim
lim
qk→∞
◦···◦Φ
qk→∞
◦ Φs2
T2
◦ Φq2
T2
lim
q2→∞
lim
q3→∞
◦ · · · ◦ Φqk
Tk
◦ · · · ◦ Φqk
Tk
q2Xs2=0
q1Xs1=0
(∆T(I))i
tracehΦs1
◦Φq2
T2
i=2 trace(cid:16)Q(i)
qi (cid:17)
Qk
◦ Φq3
T3
trace(cid:2)Φs1
(∆T(I))(cid:3)
qi(cid:17)
i=2 trace(cid:16)Q(i)
Qk
trace(cid:2)Φs1
traceh(id − Φq1+1
i=1(cid:16)Pqi
Qk
sii = (qi+1)(qi+2)···(qi+ni)
traceh(id − Φq1+1
qi(cid:17)
i=3 trace(cid:16)Q(i)
Qk
sii(cid:17)
si=0 tracehQ(i)
) ◦ · · · ◦ (id − Φqk+1
1 · · · qnk
qn1
) ◦ · · · ◦ (id − Φqk+1
and, consequently,
lim
q1→∞
qk→∞
qk→∞
· · ·
lim
· · ·
lim
ni!
Tk
T1
Tk
T1
k
(∆T(I))(cid:3)
)(I)i
.
)(I)i
.
(cid:3)
We remark that, due to relation (5.1), all the asymptotic formulas in Theorem 5.4 can be written in
terms of the constrained Berezin kernel eKT.
Theorem 5.5. If T = (T1, . . . , Tk) ∈ B(H)n1 ×· · ·×B(H)nk , then the following statements are equivalent:
(i) T is unitarily equivalent to B ⊗ IK for some finite dimensional Hilbert space K;
(ii) T is a pure finite rank element in the polyball Bc
curvc(T) = rank [∆T(I)].
n(H) such that ∆B⊗I (I − eKTeK∗
T) ≥ 0, and
34
GELU POPESCU
In this case, the constrained Berezin kernel eKT is a unitary operator and
i ∈ {1, . . . , k}, j ∈ {1, . . . , ni}.
Ti,j = eK∗
T(Bi,j ⊗ I∆T(I)(H))eKT,
Proof. The proof of the implication (i) =⇒ (ii) is similar to that of Theorem 2.7, but uses Theorem 5.4.
(5.3)
(α)∈F+
i=1F 2
i=1F 2
i=1F 2
i=1F 2
n1 ×···×F+
nk
s (Hni ) ⊗ L,
s (Hni) ⊗ E → ⊗k
(B(α) ⊗ IH)(M) = ⊗k
s (Hni) ⊗ ∆T(I)(H)). Due to Theorem 5.3, we have
T +eΘTeΘ∗
_
T) ≥ 0, the tuple T has characteristic function
s (Hni ) ⊗ ∆T(I)(H) which is a multi-analytic operator with respect to
T = I . Since T is pure, eKT is an isometry and, consequently,
Assume that item (ii) holds. Since ∆B⊗I (I − eKTeK∗
eΘT : ⊗k
the universal model B and eKTeK∗
eΘT is a partial isometry. According to [26], the support of eΘT satisfies the relation
supp (eΘT) =
where L := (PC ⊗ IH)eΘ∗(⊗k
T(eN ⊗ IH)i .
curvc(T) = rank [∆T(I)] − traceheΘT(PC ⊗ IL)eΘ∗
Since curvc(T) = rank [∆T(I)], we deduce that traceheΘT(PC ⊗ IL)eΘ∗
T(eN ⊗ IH)i = 0. Since the trace
T(eN ⊗ IH) = 0, which implies eΘT(PC ⊗ IL)eΘ∗
is faithful, we obtain eΘT(PC ⊗ IL)eΘ∗
+. Therefore, eΘT(PC ⊗ IL)eΘ∗
fact that eΘT is a multi-analytic operator with respect to B, we deduce that eΘT = 0. Using relation
eKTeK∗
T + eΘTeΘ∗
T = I , we deduce that that eKT is a co-isometry. Therefore, eKT is a unitary operator.
T(Bi,j ⊗ I∆T(I)(H))eKT for i ∈ {1, . . . , k} and j ∈ {1, . . . , ni}, the proof is complete. (cid:3)
Since Ti,j = eK∗
qk ⊗
T = 0. Due to relation 5.3 and the
n(H) have finite rank and let M be an invariant subspace under T such that
n(M). If T and TM have characteristic functions and dim M⊥ < ∞, then TM has finite
IH) = 0 for any (q1, . . . , qk) ∈ Zk
q1 ⊗ · · · ⊗ Q(k)
T(Q(1)
Theorem 5.6. Let T ∈ Bc
TM ∈ Bc
rank and curvc(T) = curvc(TM).
Proof. Note that rank [TM] = rank [∆T(PM)]. Since ∆T(PM) = ∆T(IH) − ∆T(PM⊥ ), we deduce that
rank [TM] < ∞. Since M is an invariant subspace under T, we have
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Tk (cid:17) (PM⊥ )i .
+, one can easily see that
+. On the other hand, it is easy to see that the dimension of the range of the
traceh(cid:16)id − Φq1+1
T1M(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
= traceh(cid:16)id − Φq1+1
≤ traceh(cid:16)id − Φq1+1
Taking into account that(cid:13)(cid:13)Φp1
operator(cid:16)id + Φq1+1
for any (q1, . . . , qk) ∈ Zk
T1
· · · Φpk
Tk
TkM(cid:17) (IM)i
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
(cid:13)(cid:13)(cid:13)(cid:16)id + Φq1+1
T1 (cid:17) ◦ · · · ◦(cid:16)id + Φqk+1
(cid:16)1 + trace (Q(1)
traceh(cid:16)id − Φq1+1
Tk (cid:17) (PM)i
Tk (cid:17) (IH)i + traceh(cid:16)id − Φq1+1
(PM⊥ )(cid:13)(cid:13) ≤ 1 for any (p1, . . . , pk) ∈ Zk
Tk (cid:17) (PM⊥ )(cid:13)(cid:13)(cid:13) ≤ 2k
T1 (cid:17) ◦ · · · ◦(cid:16)id + Φqk+1
Tk (cid:17) (PM⊥ ) is less than or equal to
qk+1)(cid:17) dim M⊥.
q1+1)(cid:17) · · ·(cid:16)1 + trace (Q(k)
Tk (cid:17) (PM⊥ )i
T1 (cid:17) ◦ · · · ◦(cid:16)id + Φqk+1
q1+1)(cid:17) · · ·(cid:16)1 + trace (Q(k)
≤ traceh(cid:16)id + Φq1+1
≤ 2k(cid:16)1 + trace (Q(1)
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
Tk (cid:17) (PM⊥ )i
qk+1)(cid:17) trace [PM⊥ ].
Using the fact that if A is a finite rank positive operator, then trace (A) ≤ kAkrank [A], we deduce that
CURVATURE INVARIANT ON NONCOMMUTATIVE POLYBALLS
35
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Consequently, we have
traceh(cid:16)id − Φq1+1
1 · · · qnk
qn1
Tk (cid:17) (IH)i
T1 (cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
2k(cid:16)1 + trace (Q(1)
traceh(cid:16)id − Φq1+1
qk+1)(cid:17)
q1+1)(cid:17) · · ·(cid:16)1 + trace (Q(k)
trace [PM⊥ ].
−
k
qn1
1 · · · qnk
k
≤
T1M(cid:17) ◦ · · · ◦(cid:16)id − Φqk+1
TkM(cid:17) (IM)i
1 · · · qnk
qn1
k
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
qi+1) = (qi+2)(qi+3)···(qi+ni)
Since trace (Q(i)
= 0. Passing to the limit as q1 →
∞, . . . , qk → ∞ in the inequality above and using Theorem 5.4, we deduce that curvc(T) − curvc(TM) =
0. The proof is complete.
(cid:3)
, we have limqi→∞
(ni−1)!
ni
q
i
trace (Q(i)
qi +1)
Combining Corollary 2.6 from [26] with Lemma 3.11, we deduce that if M ⊂ ⊗k
s (Hni ) ⊗ K is an
invariant subspace under the universal model B ⊗ IK, then (B ⊗ IK)M is in the commutative polyball if
and only if M is a Beurling type invariant subspace for B ⊗ IK.
i=1F 2
The proof of the next lemma is similar to that of Lemma 3.13. The only difference is that we need to
use Theorem 2.7. We omit the proof.
Lemma 5.7. Let M be an invariant subspace of ⊗k
s (Hni)⊗E such that it does not contain nontrivial
reducing subspaces for the universal model B⊗IE , and let T := PM⊥ (B⊗IE)M⊥ . Then there is a unitary
i=1F 2
T and V is the injection of M⊥ into ⊗k
operator Z : ∆T(I)(H) → E such that (I ⊗ Z)eKT = V , where eKT is the Berezin kernel associated with
Moreover, T has characteristic function if and only if M is a Beurling type invariant subspace.
s (Hni ) ⊗ E.
i=1F 2
Given a Beurling type invariant subspace M of the tensor product F 2
s (Hn1)⊗· · · ⊗ F 2
s (Hnk )⊗E, where
E is a finite dimensional Hilbert space, we introduce its multiplicity by setting
mc(M) := lim
m→∞
1
k (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k
q1 +···+qk ≤m
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IE )i
tracehPM(Q(1)
qi(cid:17)
i=1 trace(cid:16)Q(i)
Qk
.
The multiplicity measures the size of the subspace M. Note that if M = F 2
s (Hnk ) ⊗ E,
then mc(M) = dim E.
In what follows, we show that the multiplicity invariant exists. However, it
remains an open problem whether the multiplicity invariant exists for arbitrary invariant subspaces of
the tensor product F 2
s (Hnk ) ⊗ E. This is true for the polydisc, when n1 = · · · nk = 1,
and the symmetric Fock space, when k = 1.
s (Hn1) ⊗ · · · ⊗ F 2
s (Hn1 ) ⊗ · · · ⊗ F 2
The proof of the next result is similar to that of Theorem 3.1, but uses Lemma 5.7 and Theorem 5.4.
We shall omit it.
Theorem 5.8. Let M be a Beurling type invariant subspace of F 2
s (Hnk ) ⊗ E, where E
is a finite dimensional Hilbert space. Then the the multiplicity mc(M) exists and satisfies the equations
s (Hn1 ) ⊗ · · · ⊗ F 2
mc(M) =
lim
(q1,...,qk)∈Zk
+
· · ·
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IE )i
tracehPM(Q(1)
qki
tracehQ(1)
trace(cid:2)PM(Q≤(q1,...,qk) ⊗ IE )(cid:3)
trace(cid:2)Q≤(q1,...,qk)(cid:3)
tracehPM(Q(1)
k − 1 (cid:19) Xq1 ≥0,...,qk ≥0
(cid:18)m + k − 1
tracehQ(1)
q1+···+qk =m
qk→∞
lim
1
= lim
q1→∞
= lim
m→∞
= dim E − curvc(M),
q1 ⊗ · · · ⊗ Q(k)
q1 ⊗ · · · ⊗ Q(k)
qk ⊗ IE )i
qki
where M := (M1, . . . , Mk) with Mi := (Mi,1, . . . , Mi,ni ) and Mi,j := PM⊥ (Bi,j ⊗ IE )M⊥ .
36
GELU POPESCU
We remark that there are commutative analogues of all the results from Section 4, concerning the
continuity and multiplicative properties of the curvature and multiplicity invariants. Since the proofs are
very similar we will omit them. We only mention the following result concerning the lower semi-continuity
of the multiplicity invariant.
Theorem 5.9. Let M and Mm be Beurling type invariant subspaces in ⊗k
If WOT- limm→∞ PMm = PM, then
i=1F 2
s (Hni )⊗E with dim E < ∞.
lim inf
m→∞
mc(Mm) ≥ mc(M).
References
[1] W.B. Arveson, Subalgebras of C ∗-algebras, Acta.Math. 123 (1969), 141 -- 224.
[2] W.B. Arveson, Subalgebras of C ∗-algebras III: Multivariable operator theory, Acta Math. 181 (1998), 159 -- 228.
[3] W.B. Arveson, The curvature invariant of a Hilbert module over C[z1, . . . , zn], J. Reine Angew. Math. 522 (2000),
173 -- 236.
[4] W.B. Arveson, The Dirac operator of a commuting d-tuple, J. Funct. Anal. 189 (2002), 53 -- 79.
[5] A. Beurling, On two problems concerning linear transformations in Hilbert space, Acta Math. 81 (1948), 239 -- 251.
[6] J. Cuntz, Simple C ∗ -- algebras generated by isometries, Commun.Math.Phys. 57 (1977), 173 -- 185.
[7] R.G. Douglas and V.I. Paulsen, Hilbert modules over function algebras, Pitman Research Notes in Mathematics
Series, vol.217, 1989.
[8] M. Englis, Operator models and Arveson's curvature invariant, Topological algebras, their applications, and related
topics, 171 -- 183, Banach Center Publ., 67, Polish Acad. Sci., Warsaw, 2005.
[9] X. Fang, Hilbert polynomials and Arveson's curvature invariant, J. Funct. Anal. 198 (2003), no. 2, 445 -- 464.
[10] X. Fang, Invariant subspaces of the Dirichlet space and commutative algebra, J. Reine Angew. Math. 569 (2004),
189 -- 211.
[11] X. Fang, Additive invariants on the Hardy space over the polydisc, J. Funct. Anal. 253 (2007), no. 1, 359 -- 372.
[12] D. Greene, S. Richter, and C Sundberg, The structure of inner multipliers on spaces with complete Nevanlinna-Pick
kernels, J. Funct. Anal. 194 (2002), 311 -- 331.
[13] J. Gleason, S. Richter, and C Sundberg, On the index of invariant subspaces in spaces of analytic functions of
several complex variables, J. Reine Angew. Math. 587 (2005), 49 -- 76.
[14] I. Kaplansky, Commutative rings, Allyn and Bacon, Boston, 1970.
[15] D.W. Kribs, The curvature invariant of a non-commuting n-tuple, Integral Equations Operator Theory 41 (2001),
426 -- 454.
[16] P.S. Muhly and B. Solel, The curvature and index of completely positive maps, Proc. London Math. Soc. (3) 87
(2003).
[17] V.I. Paulsen, Completely Bounded Maps and Dilations, Pitman Research Notes in Mathematics, Vol.146, New York,
1986.
[18] G. Popescu, Characteristic functions for infinite sequences of noncommuting operators, J. Operator Theory 22 (1989),
51 -- 71.
[19] G. Popescu, Multi-analytic operators on Fock spaces, Math. Ann. 303 (1995), 31 -- 46.
[20] G. Popescu, Poisson transforms on some C ∗-algebras generated by isometries, J. Funct. Anal. 161 (1999), 27 -- 61.
[21] G. Popescu, Curvature invariant for Hilbert modules over free semigroup algebras, Adv. Math. 158 (2001), 264 -- 309.
[22] G. Popescu, Operator theory on noncommutative varieties, Indiana Univ. Math. J. 55 (2) (2006), 389 -- 442.
[23] G. Popescu, Operator theory on noncommutative domains, Mem. Amer. Math. Soc. 205 (2010), no. 964, vi+124 pp.
[24] G. Popescu, Free holomorphic automorphisms of the unit ball of B(H)n, J. Reine Angew. Math., 638 (2010), 119-168.
[25] G. Popescu, Berezin transforms on noncommutative polydomains, Trans. Amer. Math. Soc., to appear.
[26] G. Popescu, Berezin transforms on noncommutative varieties in polydomains, J. Funct. Anal. 265 (2013), no. 10,
2500-2552.
[27] G. Popescu, Euler characteristic on noncommutative polyballs, J. Reine Angew. Math., to appear.
[28] S. Richter, Invariant subspaces of the Dirichlet shift, J. Reine Angew. Math. 386 (1988), 205 -- 220.
[29] W. Rudin, Function theory in polydiscs, W. A. Benjamin, Inc., New York-Amsterdam 1969 vii+188 pp.
Department of Mathematics, The University of Texas at San Antonio, San Antonio, TX 78249, USA
E-mail address: [email protected]
|
1104.5358 | 1 | 1104 | 2011-04-28T11:30:49 | Optimal solutions to matrix-valued Nehari problems and related limit theorems | [
"math.FA"
] | In a 1990 paper Helton and Young showed that under certain conditions the optimal solution of the Nehari problem corresponding to a finite rank Hankel operator with scalar entries can be efficiently approximated by certain functions defined in terms of finite dimensional restrictions of the Hankel operator. In this paper it is shown that these approximants appear as optimal solutions to restricted Nehari problems. The latter problems can be solved using relaxed commutant lifting theory. This observation is used to extent the Helton and Young approximation result to a matrix-valued setting. As in the Helton and Young paper the rate of convergence depends on the choice of the initial space in the approximation scheme. | math.FA | math |
Optimal solutions to matrix-valued Nehari
problems and related limit theorems
A.E. Frazho, S. ter Horst and M.A. Kaashoek
Abstract
In a 1990 paper Helton and Young showed that under certain condi-
tions the optimal solution of the Nehari problem corresponding to a finite
rank Hankel operator with scalar entries can be efficiently approximated
by certain functions defined in terms of finite dimensional restrictions of
the Hankel operator. In this paper it is shown that these approximants
appear as optimal solutions to restricted Nehari problems. The latter
problems can be solved using relaxed commutant lifting theory. This ob-
servation is used to extent the Helton and Young approximation result to a
matrix-valued setting. As in the Helton and Young paper the rate of con-
vergence depends on the choice of the initial space in the approximation
scheme.
1 Introduction
Since the 1980s, the Nehari problem played an important role in system and con-
trol theory, in particular, in the H ∞-control solutions to sensitivity minimiza-
tion and robust stabilization, cf., [9]. In system and control theory the Nehari
problem appears mostly as a distance problem: Given G in L∞, determine the
distance of G to H ∞, that is, find the quantity d := inf{kG − F k∞ F ∈ H ∞}
and, if possible, find an F ∈ H ∞ for which this infimum is attained. Here all
functions are complex-valued functions on the unit circle T. It is well-known
that the solution to this problem is determined by the Hankel operator H which
maps H 2 into K 2 = L2 ⊖ H 2 according to the rule Hf = P−(Gf ), where P− is
the orthogonal projection of L2 onto K 2. Note that H is uniquely determined
by the Fourier coefficients of G with negative index. Its operator norm deter-
mines the minimal distance.
In fact, d = kHk and the infimum is attained.
Furthermore, if H has a maximizing vector ϕ, that is, ϕ is a non-zero function
in H 2 such that kHϕk = kHk kϕk, then the AAK theory [1, 2] (see also [18])
tells us that the best approximation bG of G in H ∞ is unique and is given by
(Hϕ)(eit)
(1.1)
bG(eit) = G(eit) −
ϕ(eit)
a.e.
By now the connection between the Nehari problem and Hankel operators is
well established, also for matrix-valued and operator-valued functions, and has
1
been put into the larger setting of metric constrained interpolation problems,
see, for example, the books [6, Chapter IX], [13, Chapter XXXV], [7, Chapter
I], [17, Chapter 5] and [3, Chapter 7], and the references therein.
The present paper is inspired by Helton-Young [14]. Note that formula (1.1)
and the maximizing vector ϕ, may be hard to compute, especially if H has
large or infinite rank. Therefore, to approximate the optimal solution (1.1),
Helton-Young [14] replaces H by the restriction H = HH 2⊖znqH 2 to arrive at
,
a.e.
( Heϕ)(eit)
eϕ(eit)
eG(eit) = G(eit) −
H. Note that a maximizing vector eϕ of
(1.2)
H always exists,
a maximizing vector of
since rank H ≤ n + deg q, irrespectively of the rank of H being finite, or not.
as an approximant of bG. Here n is a positive integer, q is a polynomial and eϕ is
In [14] it is shown that eG is a computationally efficient approximation of
the optimal solution bG when the zeros of the polynomial q are close to the
a simple singular value of H, then kbG − eGk∞ converges to 0 as n → ∞. This
poles of G in the open unit disk D that are close to the unit circle T. To be
more precise, it is shown that if G is rational, i.e., rank H < ∞, and kHk is
convergence is proportional to rn if the poles of G in D are within the disc
Dr = {z ∈ C z < r}, and the rate of convergence can be improved by an
appropriate choice of the polynomial q.
It is well-known that the Nehari problem fits in the commutant lifting frame-
work, and that the solution formula (1.1) follows as a corollary of the commutant
lifting theorem. We shall see that the same holds true for formula (1.2) provided
one uses the relaxed commutant lifting framework of [8]; cf., Corollary 2.5 in
[8].
To make the connection with relaxed commutant lifting more precise, define
Rn to be the orthogonal projection of H 2 onto H 2 ⊖ zn−1qH 2, and put Qn =
SRn, where S is the forward shift on H 2. Then the operators Rn and Qn both
map H 2 into H 2 ⊖ znqH 2, and the restriction operator Hn := HH 2⊖znqH 2
satisfies the intertwining relation V−HnRn = HnQn. Here V− is the compression
of the forward shift V on L2 to K 2. Given this intertwining relation, the relaxed
commutant lifting theorem [8, Theorem 1.1] tells us that there exists an operator
Bn from H 2 ⊖ znqH 2 into L2 such that
P−Bn = Hn,
V BnRn = BQn,
kBnk = kHnk.
(1.3)
The second identity in (1.3) implies (see Lemma 2.2 below) that for a solution
Bn to (1.3) there exists a unique function Φn ∈ L2 such that the action of Bn
is given by
(Bnh)(eit) = Φn(eit)h(eit) a.e.
(h ∈ H 2 ⊖ znqH 2).
(1.4)
Furthermore, since Hn has finite rank, there exists only one solution Bn to (1.3)
(see Proposition 2.3 below), and if ψn = eϕ is a maximizing vector of Hn, then
2
this unique solution is given by (1.4) with Φn equal to
Φn(eit) =
(Hnψn)(eit)
ψn(eit)
=
( Heϕ)(eit)
eϕ(eit)
,
a.e..
(1.5)
problem.
Thus G − eG appears as an optimal solution to a relaxed commutant lifting
This observation together with the relaxed commutant lifting theory devel-
oped in the last decade, enabled us to extent the Helton-Young convergence
result for optimal solutions in [14] to a matrix-valued setting, that is, to derive
an analogous convergence result for optimal solutions to matrix-valued Nehari
problems; see Theorem 3.1 below. A complication in this endeavor is that for-
mula (1.1) generalizes to the vector-valued case, but not to the matrix-valued
case. Furthermore, in the matrix-valued case there is in general no unique so-
lution. We overcome the latter complication by only considering the central
solutions which satisfy an additional maximum entropy-like condition. On the
way we also derive explicit state space formulas for optimal solutions to the
classical and restricted Nehari problem assuming that the Hankel operator is
of finite rank and satisfies an appropriate condition on the space spanned by
its maximizing vectors. These state space formulas play an essential role in the
proof of the convergence theorem.
This paper consists of 6 sections including the present introduction.
In
Section 2, which has a preliminary character, we introduce a restricted version
of the matrix-valued Nehari problem, and use relaxed commutant lifting theory
to show that it always has an optimal solution. Furthermore, again using relaxed
commutant lifting theory, we derive a formula for the (unique) central optimal
solution.
In Section 4
the formula for the (unique) central optimal solution derived in Section 2 is
developed further, and in Section 5 this formula is specified for the classical
Nehari problem. Using these formulas Section 6 presents the proof of the main
convergence theorem.
In Section 3 we state our main convergence result.
Notation and terminology. We conclude this introduction with a few words
about notation and terminology. Given p, q in N, the set of positive integers, we
write L2
q×p for the space of all q×p-matrices with entries in L2, the Lebesgue
space of square integrable functions on the unit circle. Analogously, we write
H 2
q×p for the space of all q×p-matrices with entries in the classical Hardy space
H 2, and K 2
q×p stands for the space of all q×p-matrices with entries in the
space K 2 = L2 ⊖ H 2, the orthogonal compliment of H 2 in L2. Note that each
F ∈ L2
q×p
and F− ∈ K 2
q×p. We shall refer to F+ as the analytic part of F and to F−
as its co-analytic part. When there is only one column we simply write L2
p,
H 2
p are
Hilbert spaces and K 2
q×p stands for the space of all q×p-
matrices whose entries are essentially bounded on the unit circle with respect to
the Lebesque measure, and H ∞
q×p stands for the space of all q×p-matrices whose
entries are analytic and uniformly bounded on the open unit disc D. Note that
q×p can be written uniquely as a sum F = F+ + F− with F+ ∈ H 2
p×1 and K 2
p . Finally, L∞
p×1. Note that L2
p×1, H 2
p ⊖ H 2
p, H 2
p and K 2
p and K 2
p instead of L2
p = L2
3
q×p belongs to L2
each F ∈ L∞
analytic part F− of F are well defined. These functions belong to L2
it may happen that neither F+ nor F− belong to L∞
need the following embedding and projection operators:
q×p and hence the analytic part F+ and the co-
q×p and
q×p. In the sequel we shall
E : Cp → H 2
p , Eu(λ) = u (z ∈ D);
1
2πZ 2π
0
Π : K 2
q → Cq, Πf =
e−itf (eit) dt.
(1.6)
(1.7)
q×p, and H : H 2
p → K 2
Throughout G ∈ L∞
the co-analytic part of G, that is, Hf = P−(Gf ) for each f ∈ H 2
the orthogonal projection of L2
forward shift on H 2
L2
q.
q is the Hankel operator defined by
p . Here P− is
q . Note that V−H = HS, where S is the
q of the forward shift V on
p and V− is the compression to K 2
q onto K 2
Finally, we associate with the Hankel operator H two auxiliary operators
involving the closure of its range, i.e., the space X = Im H, as follows:
Z : X → X ,
W : H 2
p → X ,
Z = V−X ,
W f = Hf
(f ∈ H 2
p ).
(1.8)
(1.9)
Note that X := Im H is a V−-invariant subspace of K 2
q . Hence Z is a well-
defined contraction. Furthermore, if rank H is finite, then the spectral radius
rspec(Z) is strictly less than one and the co-analytic part G− of G is the rational
matrix function given by
G−(λ) = (ΠX )(λI − Z)−1W E.
In system theory the right hand side of the above identity is known as the
restricted backward shift realization of G−; see, for example, [5, Section 7.1].
This realization is minimal, and hence the eigenvalues of Z coincide with the
poles of G− in D.
In particular, rspec(Z) < 1. Since V−H = HS, we have
ZW = W S. Furthermore, Ker H ∗ = K 2
q ⊖ X .
2 Restricted Nehari problems and relaxed com-
mutant lifting
In this section we introduce a restricted version of the Nehari problem, and
we prove that it is equivalent to a certain relaxed commutant lifting problem.
Throughout M is a subspace of H 2
p such that
S∗M ⊂ M,
Ker S∗ ⊂ M.
(2.1)
With M we associate operators RM and QM acting on H 2
into M. By definition RM is the orthogonal projection of H 2
QM = SRM.
p , both mapping H 2
p
p onto S∗M and
4
We begin by introducing the notion of an M-norm. We say that Φ ∈ L2
q×p
q for each h ∈ M and the map h 7→ Φh is a
has a finite M-norm if Φh ∈ L2
bounded linear operator, and in that case we define
kΦkM = sup{kΦhkL2
q h ∈ M,
khkH 2
p ≤ 1}.
If M is finite dimensional, then each Φ ∈ L2
more, Φ ∈ L∞
kΦkM ≤ kΦk∞, with equality if M = H 2
M-norm and G ∈ L∞
q×p has a finite M-norm. Further-
q×p has a finite M-norm for every choice of M, and in this case
q×p has a finite
q×p imply G − Φ has a finite M-norm.
p . Note that Φ ∈ L2
We are now ready to formulate the M-restricted Nehari problem. Given
p , we define the optimal M-restricted Nehari
G ∈ L∞
problem to be the problem of determining the quantity
q×p and a subspace M of H 2
dM := inf{kG − F kM F ∈ H 2
q×p and F has a finite M-norm},
(2.2)
and, if possible, to find a function F ∈ H 2
q×p of finite M-norm at which the
infimum is attained. In this case, a function F attaining the infimum is called
an optimal solution. The suboptimal variant of the problem allows the norm
kG−F kM to be larger than the infimum. When M = H 2
p , the problem coincides
with the classical matrix-valued Nehari problem in L∞
q×p. In [15, 16] the case
where M = H 2
p , with k ∈ N, was considered.
p ⊖ SkH 2
Proposition 2.1. Let G ∈ L∞
p satisfying
the conditions in (2.1). Then the M-restricted Nehari problem has an optimal
solution and the quantity dM in (2.2) is equal to γM := kHMk, where H :
H 2
q is the Hankel operator defined by the co-analytic part of G.
q×p, and let M be a subspace of H 2
p → K 2
We shall derive the above result as a corollary to the relaxed commutant
lifting theorem [8, Theorem 1.1], in a way similar to the way one proves the
Nehari theorem using the classical commutant lifting theorem (see, for example,
[6, Section II.3]). For this purpose we need the following notion. We say that
an operator B from M into L2
q×p if the action of B is
given by
q is defined by a Φ ∈ L2
(Bh)(eit) = Φ(eit)h(eit) a.e.
(h ∈ M).
(2.3)
In that case, Φ has a finite M-norm, and kΦkM = kBk. When (2.3) holds we
refer to Φ as the defining function of B. The following lemma characterizes
operators B from M into L2
q×p in terms of an
intertwining relation.
q defined by a function Φ ∈ L2
Lemma 2.2. Let M be a subspace of H 2
p satisfying (2.1), and let B be a bounded
operator from M into L2
q×p if and only if
B satisfies the intertwining relation V BRM = BQM. In that case, Φ(·)u =
BEu(·) for any u ∈ Cp and kBk = kΦkM
q. Then B is defined by a Φ ∈ L2
Proof. This result follows by a modification of the proof of Lemma 3.2 in [11].
We omit the details.
5
Proof of Proposition 2.1. Put γM = kHMk. Recall that the Hankel opera-
tor H satifies the intertwining relation V−H = HS. This implies V−HMRM =
HMQM. Here RM and QM are the operators defined in the first paragraph
of the present section. Since Q∗
MRM and V is an isometric lifting
of V−, the quintet
MQM = R∗
{HM, V−, V, RM, QM, γM}
(2.4)
is a lifting data set in the sense of Section 1 in [8]. Thus Theorem 1.1 in [8]
guarantees the existence of an operator B from M into L2
q with the properties
P−B = HM,
V BRM = BQM,
kBk = γM.
(2.5)
By Lemma 2.2 the second equality in (2.5) tells us there exists a Φ ∈ L2
q×p
defining B, that is, the action of B is given by (2.3). As Φ(·)u = BEu(·), the
first identity in (2.5) shows that G− = Φ−, and hence F := G − Φ ∈ H 2
q×p.
Furthermore,
kG − F kM = kΦkM = kBk = γM,
because of the third identity in (2.5). Thus the quantity dM in (2.2) is less than
or equal to γM.
It remains to prove that dM ≥ γM. In order to do this, let F ∈ H 2
q×p and
have a finite M-norm. Put Φ = G − F . Then Φ has a finite M-norm. Let B be
q×p, we have G− = Φ−,
the operator from M into L2
and hence the first identity in (2.5) holds with B in place of B. It follows that
q defined by Φ. Since F ∈ H 2
kG − F kM = k ΦkM = k Bk ≥ kHMk = γM.
This completes the proof.
In the scalar case, or more generally in the case when p = 1, the optimal
solution is unique. Moreover this unique solution is given by a formula analogous
to (1.2); cf., [1]. This is the contents of the next proposition which is proved
in much the same way as the corresponding result for the Nehari problem. We
omit the details.
Proposition 2.3. Assume p = 1, that is, G ∈ L∞
q and M a subspace of H 2
satisfying (2.1). Assume that HM has a maximizing vector ψ ∈ M. Then
there exists only one optimal solution F to the M-restricted Nehari problem
(2.5), and this solution is given by
F (eit) = G(eit) −
(Hψ)(eit)
ψ(eit)
a.e.
(2.6)
In general, if p > 1 the optimal solution is not unique. To deal with this
non-uniqueness, we shall single out a particular optimal solution.
First note that the proof of Proposition 2.1 shows that there is a one-to-one
correspondence between the optimal solutions of the M-restricted Nehari prob-
lem of G and all interpolants for HM with respect to the lifting data set (2.4),
6
that is, all operators B from M into L2
is given by
q satisfying (2.5). This correspondence
B 7→ F = G − Φ, where Φ is the defining function of B.
(2.7)
Next we use that the relaxed commutant lifting theory tells us that among
all interpolants for HM with respect to the lifting data set (2.4) there is a
particular one, which is called the central interpolant for HM with respect to
the lifting data set (2.4); see [8, Section 4]. This central interpolant is uniquely
determined by a maximum entropy principle (see [8, Section 8]) and given by
an explicit formula using the operators appearing in the lifting data set.
Using the correspondence (2.7) we say that an optimal solution F of the
M-restricted Nehari problem of G is the central optimal solution whenever
Φ := G − F is the defining function of the central interpolant B for HM with
respect to the lifting data set (2.4). Furthermore, using the formula given in
[8, Section 4] for the central interpolant the correspondence (2.7) allows us to
derive a formula for the central optimal solution. To state this formula we need
to make some preparations.
As before γM = kHMk. Note that kHPMSk ≤ kHPMk = kHMk, where
PM is the orthogonal projection of H 2(Cp) on M. This allows us to define the
following defect operators acting on H 2(Cp)
DM = (γ2
D◦
M = (γ2
MI − PMH ∗HPM)1/2 on H 2(Cp),
MI − S∗PMH ∗HPMS)1/2 on H 2(Cp).
For later purposes we note that S∗D2
MS = D◦2
M. Next define
(2.8)
(2.9)
(2.10)
ω2(cid:21) : H 2
p →(cid:20) Cq
p(cid:21) ,
ω =(cid:20)ω1
DMRM(cid:21)
ω(DMQM) =(cid:20)ΠHRM
H 2
and ωKer Q∗
MDM = 0.
(2.11)
From the relaxed commutant lifting theory we know that ω is a well defined
partial isometry with initial space F = Im DMQM. Furthermore, the forward
shift operator V on L2
q is the Sz.-Nagy-Schaffer isometric lifting of V−. Then as
a consequence of [8, Theorem 4.3] and the above analysis we obtain the following
result.
q×p, and let M be a subspace of H 2
Proposition 2.4. Let G ∈ L∞
p satisfying the
conditions in (2.1). Then the central optimal solution FM to the M-restricted
Nehari problem is given by FM = G − ΦM, where ΦM ∈ L2
q×p has finite M-
norm, the co-analytic part of ΦM is equal to G−, and the analytic part ΦM,+
of ΦM is given by
ΦM,+(λ) = ω1(I − λω2)−1DME.
(2.12)
Here E is defined by (1.6), and ω1 and ω2 are defined by (2.10) and (2.11).
7
It is this central optimal solution FM we shall be working with. From
Corollary 4.4 in [8] (see also [10, Theorem 1.1]) we know that F = Im DMQM =
DM implies that the central solution of (2.5) is the only optimal solution to the
M-restricted Nehari problem. The latter fact will play a role in Section 4.
3 Statement of the main convergence result
Let G ∈ L∞
q×p, and let H be the Hankel operator defined by the co-analytic part
of G. In our main approximation result we shall assume that the following two
conditions are satisfied:
(C1) H has finite rank,
(C2) none of the maximizing vectors of H belongs SH 2
by the maximizing vectors of H has dimension p.
p , and the space spanned
Note that (C1) is equivalent to G being the sum of a rational matrix function
with all its poles in D and a matrix-valued H ∞ function.
In the scalar case the second part of (C2) implies the first part. To see this
let p = q = 1, and assume that the space spanned by the maximizing vectors
of H is one dimensional. Let Sv be a maximizing vector of H. Since S is an
isometry and V−H = HS, we have v 6= 0 and
kHkkvk = kHkkSvk = kHSvk = kV−Hvk ≤ kHvk ≤ kHkkvk.
Thus the inequalities are equalities, and v is a maximizing vector of H. As
the the space spanned by the maximizing vectors of H is assumed to be one
dimensional, v must be a scalar multiple of Sv, which can only happen when
v = 0, which contradicts v 6= 0. Thus the first part of (C2) is fulfilled. Next
observe that for p = q = 1 the statement "the space spanned by the maximizing
vectors of H has dimension one" is just equivalent to the requirement that kHk
is a simple singular value of H, which is precisely the condition used in Theorem
2 of the Helton-Young paper [14].
As we shall see in Section 5 the two conditions (C1) and (C2) guarantee that
the solution to the optimal Nehari problem is unique.
For our approximation scheme we fix a finite dimensional subspace M0 of
H 2
p invariant under S∗, and we define recursively
Mk = Ker S∗ ⊕ SMk−1,
k ∈ N.
(3.1)
Since M0 is invariant under S∗, the space M⊥
Beurling-Lax theorem tells us that M⊥
0 = ΘH 2
taken to be inner. Using this representation one checks that Mk = H 2
0 is invariant under S, and the
ℓ , where Θ ∈ H ∞
p×ℓ and can be
p ⊖zkΘH 2
ℓ
p .
for each k ∈ N. It follows that M0 ⊂ M1 ⊂ M2 ⊂ · · · and Wk≥0 Mk = H 2
Furthermore,
S∗Mk ⊂ Mk
and Ker S∗ ⊂ Mk,
k ∈ N.
(3.2)
Note that the spaces Mk = H 2 ⊖ zkqH 2, k = 1, 2, . . ., appearing in [14] satisfy
(3.1) with M0 = H 2 ⊖ qH 2.
8
Theorem 3.1. Let G ∈ L∞
q×p. Assume that conditions (C1) and (C2) are
satisfied, and let the sequence of subspaces {Mk}k∈N be defined by (3.1) with
M0 a finite dimensional S∗-invariant subspace of H 2
p . Let F be the unique
optimal solution to the Nehari problem for G, and for each k ∈ N+ let Fk be the
central optimal solution to the Mk-restricted Nehari problem. Then G − F is
a rational function in H ∞
q×p, and for k ∈ N+ sufficiently large, the same holds
true for G − Fk. Furthermore, kFk − F k∞ → 0 for k → ∞. More precisely, if
all the poles of G inside D are within the disk Dr = {λ λ < r}, for r < 1, then
there exists a number L > 0 such that kFk − F k∞ < Lrk for k large enough.
Improving the rate of convergence is one of the main issues in [14], where it
is shown that for the case when the poles of G inside D are close to the unit
circle, that is, r close to 1, convergence with M0 = {0} may occur at a slow rate.
In [14] it is also shown how to choose (in the scalar case) a scalar polynomial
q so that the choice M0 = H 2 ⊖ qH 2 increases the rate of convergence.
In
fact, if the roots of q coincide with the poles of G in Dr\D
0, then starting with
0 ) rather than O(rk). In Section
M0 = H 2 ⊖qH 2 the convergence is of order O(rk
6 we shall see that Theorem 3.1 remains true if r < 1 is larger than the spectral
radius of the operator V−HM⊥
, and thus again the convergence rate can be
improved by an appropriate choice of M0. To give a trivial example: when M0
is chosen in such a way that it includes Im H ∗, all the central optimal solutions
Fk in Theorem 3.1 coincide with the unique optimal solution solution F to the
Nehari problem.
0
4 The central optimal solution revisited
As before G ∈ L∞
q×p and H is the Hankel operator defined by the co-analytic
part of G. Furthermore, M is a subspace of H 2
p satisfying (2.1). In this section
we assume that kHPMSk < γM = kHPMk. In other words, we assume that
the defect operator D◦
M defined by (2.9) is invertible. This additional condition
allows us to simplify the formula for the central optimal solution to the M-
restricted Nehari problem presented in Proposition 2.4. We shall prove the
following theorem.
Theorem 4.1. Let G ∈ L∞
Assume the defect operator D◦
q×p, and let M be a subspace of H 2
p satisfying (2.1).
M defined by (2.9) is invertible, and put
ΛM = D◦−2
M S∗D2
M.
(4.1)
Then rspec(ΛM) ≤ 1, and the central optimal solution FM to the M-restricted
Nehari problem is given by FM = G − ΦM, where ΦM ∈ L2
q×p has finite M-
norm, the co-analytic part of ΦM is equal to G−, and the analytic part of ΦM
is given by
ΦM,+(λ) = ΠH(I − λΛM)−1ΛME = NM(λ)MM(λ)−1
(λ ∈ D),
(4.2)
where
NM(λ) = ΠH(I − λS∗)−1ΛME, MM(λ) = I − λE∗(I − λS∗)−1ΛME. (4.3)
9
In particular, M (λ) is invertible for each λ ∈ D.
The formulas in the above theorem for the central optimal solution are in-
spired by the formulas for the central suboptimal solution in Sections IV.3 and
IV.4 of [7].
We first prove two lemmas. In what follows PM and RM are the orthogonal
projections of H 2
p onto M and S∗M, respectively, and QM = SRM.
Lemma 4.2. Let M be a subspace of H 2
p satisfying (2.1). Then
RM = S∗PMS, RMS∗ = S∗PM, QM = PMS.
(4.4)
Proof. Note that
(S∗PMS)2 = S∗PMSS∗PMS = S∗PMS − S∗PM(I − SS∗)PMS.
Since I − SS∗ is the orthogonal projection onto Ker S∗, the second part of (2.1)
implies that PM(I − SS∗) = I − SS∗. Thus (S∗PMS)2 = S∗PMS, and hence
S∗PMS is an orthogonal projection. The range of this orthogonal projection is
S∗M, and therefore the first identity in (4.4) is proved.
Using this first identity and PM(I − SS∗) = I − SS∗ we see that
RMS∗ = S∗PMSS∗ = S∗PM − S∗PM(I − SS∗) = S∗PM.
Thus the second identity in (4.4) also holds. Finally,
QM = SRM = (RMS∗)∗ = (S∗PM)∗ = PMS.
Thus (4.4) is proved.
Lemma 4.3. Let G ∈ L∞
Assume the defect operator D◦
F of the operator DMQM is closed and the orthogonal projection of H 2
F is given by
p satisfying (2.1).
M defined by (2.9) is invertible. Then the range
p onto
q×p, and let M be a subspace of H 2
PF = DMQMD◦−2
M Q∗
MDM.
Proof. We begin with two identities:
DMPM = PMDM, D◦
MRM = RMD◦
M.
(4.5)
(4.6)
Since PM is an orthogonal projection, the first equality in (4.6) follows directly
from the definition of DM in (2.8). To prove the second, we use the second
identity in (4.4). Taking adjoints and using the fact that RM and PM are
orthogonal projections, we see that PMS = SRM. It follows that D◦
M is also
given by
D◦
M = (γ2
MI − RMS∗H ∗HSRM)1/2.
(4.7)
From this formula for D◦
M the second identity in (4.6) is clear.
10
Now assume that D◦
M is invertible, and let P be the operator defined by
the right hand side of (4.5). Clearly, P is selfadjoint. Let us prove that P is a
projection. Using the second equality in (4.6) we have
MD2
P 2 = DMQMD◦−2
= DMQMD◦−2
= DMQMRMD◦−2
M Q∗
M (RMS∗D2
M (S∗D2
M Q∗
MDM
M Q∗
M RMQ∗
MSRM)D◦−2
MS)D◦−2
MQMD◦−2
MDM
MDM.
Observe that QMRM = SR2
that
M = SRM = QM. Since D◦2
M = S∗D2
MS, it follows
P 2 = DMQMD◦−2
M Q∗
MDM = P.
Thus P is an orthogonal projection. This implies that DMQM has a closed
range, and PF = P .
Proof of Theorem 4.1. Our starting point is formula (2.12). Recall that ω1
and ω2 are zero on Ker Q∗
MDM. From Lemma 4.3 we know that DMQM has
a closed range. It follows that ω1 = ω1PF and ω2 = ω2PF , where PF is the
orthogonal projection of H 2
p onto F = Im DMQM. Using the formula for PF
given by (4.5), the second intertwining relation in (4.6), the identities in (4.4)
and the definition of ω in (2.10), (2.10) we compute
ω1DM = ω1PF DM = ω1DMQMD◦−2
= ΠHRMD◦−2
= ΠHD◦−2
M S∗RMD2
M RMS∗D2
M
M Q∗
MD2
M = ΠHD◦−2
M = ΠHΛMPM,
M RMS∗D2
M
and
ω2DM = ω2PF DM = ω2DMQMD◦−2
= DMRMD◦−2
M Q∗
MD2
M Q∗
MD2
M
M = DMΛMPM.
Furthermore, using the intertwing relations in (4.6) and the second identity in
(4.4) we see that RMΛM = ΛMPM. In particular, ΛM leaves M invariant.
Let us now prove that rspec(ΛM) ≤ 1. Note that
rspec(ω2) = rspec(ω2PF ) = rspec(DMRMD◦−2
M S∗DM)
= rspec(RMD◦−2
M S∗D2
M) = rspec(RMΛM) = rspec(ΛMPM).
Thus rspec(ΛMPM) ≤ 1, because ω2 is contractive. Since ΛM leaves M invari-
p = M ⊕ M⊥ the
ant, we see that relative to the orthogonal decomposition H 2
operator ΛM decomposes as
ΛM =(cid:20)PMΛMPM
0
(I − PM)ΛM(I − PM)(cid:21) .
⋆
(4.8)
11
Note that (I −PM)(I −RM) = (I −PM). Using the latter identity, the formulas
(2.8) and (4.7), and the intertwining relations in (4.6), we obtain
(I − PM)ΛM(I − PM) = (I − PM)(I − RM)ΛM(I − PM)
= (I − PM)(I − RM)S∗(I − PM)
= (I − PM)S∗(I − PM).
Thus (I−PM)ΛM(I−PM) is a contraction. Hence rspec((I−PM)ΛM(I−PM) ≤
1. But then (4.8) shows that rspec(ΛM) ≤ 1.
Next, using that ΛMM ⊂ S∗M ⊂ M and Im E = Ker S∗ ⊂ M, we obtain
for each λ ∈ D that
ΦM,+(λ) = ω1(I − λω2)−1DME = ω1DM(I − λΛMPM)−1E
= ΠHΛMPM(I − λΛMPM)−1E = ΠH(I − λΛMPM)−1ΛMPME
= ΠH(I − λΛM)−1ΛME,
which gives formula (4.2).
Finally, to see that (4.3) holds, note that ΛMS = I. Hence ΛM is a left
inverse of S. Since E is an isometry with Im E = Ker S∗, we have ΛM =
S∗ + ΛMEE∗. Therefore, for each λ ∈ D,
ΦM,+(λ) = ΠH(I − λΛM)−1ΛME = ΠH(I − λS∗ − λΛMEE∗)−1ΛME
= ΠH(I − λ(I − λS∗)−1ΛMEE∗)−1(I − λS∗)−1ΛME
= ΠH(I − λS∗)−1ΛME(I − λE∗(I − λS∗)−1ΛME)−1
= N (λ)M (λ)−1.
In particular, M (λ) is invertible.
Remark. From RMΛM = ΛMPM we see that ΛM leaves M invariant. Thus,
if M in Theorem 4.1 is finite dimensional, then ΦM,+ in (4.2) is a rational
function in H 2
p×q, and hence ΦM,+ is a rational p×q matrix function which has
no pole in the closed unit disk.
Next we present a criterion in terms of maximizing vectors under which
Theorem 4.1 applies.
Proposition 4.4. Assume rank HPM is finite. Then D◦
only if none of the maximizing vectors of HPM belongs to SH 2
p .
M is invertible if and
M. Thus we have to show that invertibility of D◦
p is a maximizing vector of HPM if and only if 0 6=
M is equivalent to
Proof. A vector h ∈ H 2
h ∈ D⊥
D⊥
M ∩ SH 2
p = {0}.
Assume D⊥
M ∩ SH 2
p 6= {0}. Thus, using the definition of a maximizing
vector, there exists Sv with v 6= 0 such that kHPMSvk = γMkSvk. Since S is
an isometry we see that kHPMSvk = γMkvk. It follows that v is in the kernel
of D◦
M is not invertible.
M, and hence D◦
12
Conversely, assume that D⊥
p = {0}. Note that rank (HPMS) is also
finite. Hence HPMS has a maximizing vector, say v. We may assume that
kvk = 1. By our assumption the vector Sv is not a maximizing vector of HPM.
Hence
M ∩ SH 2
kHPMSk = kHPMSkkvk = kHPMSvk < kHPMkkSvk = γMkSvk = γM.
Therefore D◦2
Consequently, D◦
M = γ2
M is invertible.
MI−S∗PMH ∗HPMS is positive definite, and thus invertible.
For later purposes we mention the following. It is straightforward to prove
MI − HPMSS∗PMH ∗ is
M is invertible if and only if the operator γ2
that D◦
invertible, and in that case we have
ΛMPMH ∗ = RMH ∗V ∗
−(γ2
MI − HPMSS∗PMH ∗)−1×
×(γ2
MI − HPMH ∗),
ΛME = −RMH ∗V ∗
−(γ2
MI − HPMSS∗PMH ∗)−1HE.
(4.9)
(4.10)
These formulas can be simplified further using the operators Z and W associated
to the Hankel operator H which have been introduced at the end of Section 1,
q ⊖ X = Ker H ∗, the
see (1.8) and (1.9). Recall that X = Im H. Since K 2
MI − HPMSS∗PMH ∗ and
space X is a reducing subspace for the operators γ2
MI − HPMH ∗. Furthermore,
γ2
∆M := (γ2
ΞM := (γ2
MI − HPMSS∗PMH ∗)X = γ2
MI − HPMH ∗)X = γ2
MIX − W PMW ∗.
MIX − ZW RMW ∗Z ∗,
(4.11)
(4.12)
Note that ∆M is invertible if and only if D◦
operators, (4.9) and (4.10) can be written as
M is invertible. Using the above
ΛMPMW ∗ = RMW ∗Z ∗∆−1
M ΞM, ΛME = −RMW ∗Z ∗∆−1
M W E.
(4.13)
Corollary 4.5. Let G ∈ L∞
p satisfying (2.1).
Assume the operator ∆M defined by (4.11) is invertible. Then the defect opera-
tor D◦
M defined by (2.9) is invertible, and the functions NM and MM appearing
in (4.3) are also given by
q×p, and let M be a subspace of H 2
NM(λ) = NM,1(λ) + NM,2(λ),
NM,1(λ) = −ΠHW ∗(I − λZ ∗)−1Z ∗∆−1
NM,2(λ) = ΠH(I − λS∗)−1(I − RM)W ∗Z ∗∆−1
M W E
M W E.
and
MM(λ) = MM,1(λ) + MM,2(λ),
MM,1(λ) = I + λE∗W ∗(I − λZ ∗)−1Z ∗∆−1
MM,2(λ) = −λE∗(I − λS∗)−1(I − RM)W ∗Z ∗∆−1
M W E
M W E.
13
(4.14)
(4.15)
(4.16)
(4.17)
(4.18)
(4.19)
Furthermore, if rspec(Z ∗∆−1
and
M ΞM) < 1, then MM,1(λ) is invertible for λ ≤ 1
MM,1(λ)−1 = I − λE∗W ∗(I − λZ ∗∆−1
M ΞM)−1Z ∗∆−1
M W E,
λ ≤ 1.
(4.20)
Proof. For operators A and B the invertibility of I + AB is equivalent to the
invertibility of I + BA. Using this fact it is clear that the invertibility of D◦
M
follows form the invertibility of ∆M. Hence we can apply Theorem 4.1. Writing
RM as I − (I − RM) and using (4.13) we see that (4.14) holds with NM,2 being
given by (4.16) and with
NM,1(λ) = −ΠH(I − λS∗)−1W ∗Z ∗∆−1
M W E.
(4.21)
The intertwining relation W S = ZW yields (I − λS∗)−1W ∗ = W ∗(I − λZ ∗)−1.
Using the latter identity in (4.21) yields (4.15). In a similar way one proves the
identities (4.17)-(4.19).
To complete the proof assume rspec(Z ∗∆−1
M ΞM) < 1. Then the inversion
formula for MM,1(λ) follows from the standard inversion formula from [4, The-
orem 2.2.1], where we note that the state operator in the inversion formula
equals
Z ∗ − Z ∗∆−1
M W EE∗W ∗ = Z ∗∆−1
= Z ∗∆−1
= Z ∗∆−1
= Z ∗∆−1
M (γ2
M (γ2
M (γ2
M (γ2
MI − ZW RMW ∗Z ∗ − W EE∗W ∗)
MI − W (SRMS∗ + EE∗)W ∗)
MI − W (SS∗PM + EE∗PM)W ∗)
MI − W PMW ∗) = Z ∗∆−1
M ΞM,
as claimed. Here we used the second identity in (4.4), and the fact that PME =
E, because Im E = Ker S∗ ⊂ M.
5 The special case where M = H 2
p
Throughout this section M = H 2
p -restricted
Nehari problem, which is just the usual Nehari problem. Since M = H 2
p , we
will surpress the index M in our notation, and just write D, D◦, D, D◦, Λ, etc.
instead of DM, D◦
p , that is, we are dealing with the H 2
M, ΛM, etc. In particular,
M, DM, D◦
γ = kHk, D = (γ2I − H ∗H)1/2, D◦ = (γ2I − S∗H ∗HS)1/2.
(5.1)
We shall assume (cf., the first paragraph of Section 3) that the following two
conditions are satisfied
(C1) H has finite rank,
(C2) none of the maximizing vectors of H belongs SH 2
by the maximizing vectors of H has dimension p.
p , and the space spanned
14
Note that the space spanned by the maximizing vectors of H is equal to Ker D =
D⊥, where D is the closure of the range of D. As H 2
p = Ker S∗ ⊕ SH 2, we see
that
(C2) ⇐⇒ H 2
p = Ker D +SH 2
p ⇐⇒ H 2
p = Ker S∗ +D.
(5.2)
Here + means direct sum, not necessarily orthogonal direct sum.
Let Z and W be the operators defined by (1.8) and (1.9), respectively, and
∆ = γ2IX − ZW W ∗Z ∗, Ξ = γ2IX − W W ∗.
(5.3)
We shall prove the following theorem.
Theorem 5.1. Let G ∈ L2
q×p, and assume that the Hankel operator H asso-
ciated with the co-analytic part of G satisfies conditions (C1) and (C2). Then
the operator ∆ defined by the first identity in (5.3) is invertible and the Nehari
problem associated with G has a unique optimal solution F ∈ H ∞
q×p. Moreover,
this unique solution is given by F = G+ − Φ+, where G+ is the analytic part of
G and Φ+ is the rational q×p matrix-valued H ∞ function given by
Φ+(λ) = N (λ)M (λ)−1, where
N (λ) = −ΠHW ∗(IX − λZ ∗)−1Z ∗∆−1W E,
M (λ) = I + λE∗W ∗(IX − λZ ∗)−1Z ∗∆−1W E,
Furthermore, rspec(Z ∗∆−1Ξ) < 1, and the inverse of M (λ) is given by
M (λ)−1 = I − λE∗W ∗(IX − λZ ∗∆−1Ξ)−1Z ∗∆−1W E.
Here Ξ is the operator defined by the second identity in (5.3).
The fact that condition (C2) implies uniqueness of the optimal solution
follows from [2]; cf., Theorem 7.5 (2) in [3]. It will be convenient first to prove
the following lemma.
Lemma 5.2. Assume H is compact and (C2) is satisfied. Then the following
holds.
(i) The operator D◦ is invertible, and the range of DS is closed and is equal
to D. In particular, the optimal solution to the Nehari problem is unique.
(ii) The subspace Ker D = D⊥ of H 2
p is cyclic for S.
(iii) The operators ω2 = DD◦−2S∗D and Λ = D◦−2S∗D are well-defined and
strongly stable.
Proof. We split the proof into three parts according to the three items.
Part 1. We prove (i). Since H is compact, the selfadjoint operator D has closed
range and a finite dimensional null space. Thus D is a Fredholm operator of
index zero. See [12, Section XI.1] for the definitions of these notions. Note
15
S is a Fredholm operator of index p. Thus DS is also a Fredholm operator.
p = DSH 2
In particular, the range of DS is closed, and hence F := DSH 2
p .
Moreover,
ind(DS) = ind(D) + ind(S) = −p.
Here ind denotes the index of a Fredholm operator, and we used the fact ([12,
Theorem XI.3.2.]) that the index of a product of two Fredholm operators is
the sum of the indices of the factors. On the other hand, since Ker D ∩ SH 2
p
consists of the zero vector only, we see that Ker DS = {0}, and hence, using
the definition of the index, we have p = codim DSH 2
p = D
and, by the third part of (5.2), we have codim D = p Thus F = D. The latter
implies that the central solution of (2.5) is the only optimal solution of the
Nehari problem; see the remark made at the end of Section 2.
p . But DSH 2
p ⊂ DH 2
Finally, Ker DS = {0} and DS has closed range, yields D◦2 = S∗D2S is
invertible. This completes the proof of (i).
Part 2. We prove (ii). We begin with a remark. From (i) we know that that
D◦ is invertible. Thus the operators ω2 = DD◦−2S∗D and Λ = D◦−2S∗D2 are
well defined. Clearly, ω2D = DΛ, and hence ωk
2 D = DΛk for k = 0, 1, 2, . . .. It
follows that
Λk+1 = D◦−2S∗D2Λk = D◦−2S∗Dωk
2 ,
k = 0, 1, 2, . . . .
(5.4)
Since ω2 is a contraction, we conclude that supk≥0 kΛkk < ∞.
p = W∞
to W∞
k=0 SkD⊥. Take h ∈ H 2
Our aim is to prove that H 2
p perpendicular
k=0 SkD⊥. The latter is equivalent to S∗kh being perpendicular to D⊥ for
k = 0, 1, 2, . . ., that is, S∗kh ∈ D for k = 0, 1, 2, . . .. Recall that the range of
D is closed, because H is compact. Thus for each k = 0, 1, 2, . . . the vector
S∗kh = Dhk for some hk ∈ D. Thus S∗k+1h = S∗Dhk. Since D◦ is invertible,
p tells us that P := DSD◦−2S∗D is
Lemma 4.3 specified for the case M = H 2
the orthogonal projection of H 2
p onto F = D = Im D. Thus for k = 0, 1, 2, . . .
we have
S∗kh = Dhk = DP hk = D2SD◦−2S∗Dhk = D2SD◦−2S∗k+1h = Λ∗S∗k+1h,
p . This proves (ii).
and by induction h = Λ∗kS∗kh. Since limk→0 kS∗k+1hk = 0, and supk≥0 kΛkk <
k=0 SkD⊥ =
∞, it follows that khk = 0. Hence h = 0, and we can conclude thatW∞
H 2
Part 3. We prove (iii). We already know that ω2 and Λ are well defined. We first
prove that ω2 is strongly stable, that is, limk→∞ ωk
p . Note
2 DSk = D for k = 0, 1, 2, . . .. Since D⊥ = Ker D, we
that ω2DS = D. Hence ωk
have for any nonnegative integers k, l that ωk+l
2SD⊥ = 0. In other
words, the kernel of ωk
p . According to
p ⊖ Xk, and since
ν=0 SνD⊥. Let v ∈ H 2
p . Thus PYk v → 0, with Yk = H 2
2 includes Xk :=Wk
2 v = 0 for any v ∈ H 2
2 DSkD⊥ = ωl
(ii), we haveW∞
ν=0 SνD⊥ = H 2
ω2 is contractive, we find that
kωk
2 vk = kωk
2 PYk vk ≤ kPYk vk → 0.
16
Thus ω2 is strongly stable, as claimed, and the fact that Λ is strongly stable
follows immediately from (5.4).
Proof of Theorem 5.1. From Lemma 5.2 (i) we know that D◦ is invertible,
and the optimal solution is unique. Since the invertibility of D◦ implies the
invertibility of ∆, we can apply Theorem 4.1 and Corollary 4.5 with M = H 2
p
to get the desired formula for Φ+. Note that RH 2
p = I, and hence in this case
the functions appearing in (4.16) and (4.19) are identically zero.
Put T = Z ∗∆Ξ. Next we show that rspec(T ) < 1. By specifying the first
identity in (4.13) we see that ΛW ∗ = W ∗T , and thus ΛkW ∗ = W ∗T k for
each k ∈ N. Since Λ is strongly stable (by Lemma 5.2 (iii)), we arrive at
limk→∞ W ∗T kx = 0. The fact that H has finite rank, implies that the range
of H is closed, and hence W is surjective. But then (W W ∗)−1W is a left
inverse of W ∗, and T kx = (W W ∗)−1W T kx → 0 if k → ∞. Thus T is strongly
stable. Since the underlying space X is finite dimensional, we conclude that
rspec(Z ∗∆Ξ) = rspec(T ) < 1.
Finally, since rspec(Z ∗∆Ξ) = rspec(T ) < 1, the invertibility of M (λ) for kλk
and the formula for its inverse follow by specifying the final part of Corollary
4.5 for the case when M = H 2
p .
6 Convergence of central optimal solutions
Throughout G ∈ L∞
q×p and H is the Hankel operator defined by the co-analytic
part of G. We assume that conditions (C1) and (C2) formulated in the first
paragraph of Section 3 are satisfied. Furthermore, M0 is a finite dimensional
S∗-invariant subspace of H 2
p , and M0, M1, M2, . . . is a sequence of subspaces
of H 2
p defined recursively by (3.1). We set Pk = PMk . From the remarks made
in the paragraph preceding Theorem 3.1 one sees that
I − Pk = Sk(I − P0)S∗k, S∗Pk = Pk−1S∗ PkE = E.
(k ∈ N).
(6.1)
Here E is the embedding operator defined by (1.6).
In this section we will proof Theorem 3.1. In fact we will show that with an
appropriate choice of the initial space M0 convergence occurs at an ever faster
rate than stated in Theorem 3.1. We start with a lemma that will be of help
when proving the increased rate of convergence.
Lemma 6.1. Let Z and W be the operators defined by (1.8) and (1.9), re-
spectively, and put X0 = W M⊥
0 ⊂ X . Then X0 is Z-invariant of X = Im W ,
and rspec(Z0) ≤ rspec(Z). Furthermore, let the operators Z0 : X0 → X0 and
W0 : H 2
p → X0 be defined by Z0 = ZX0 and W0 = ΠX0 W , where ΠX0 is the
orthogonal projection of X onto X0. Then
Z kW (I − P0) = Π∗
X0 Z k
0 W0(I − P0),
k = 0, 1, 2, . . . .
(6.2)
17
Proof. Since ZW = W S and M⊥
0 is invariant under S, we see that X0 is
invariant under Z, and thus rspec(Z0) ≤ rspec(Z). From the definition of Z0 and
W0 we see that ZΠ∗
X0W0(I − P0) = W (I − P0). Thus
X0 Z0 and Π∗
X0 = Π∗
Z kW (I − P0) = Z kΠ∗
X0 W0(I − P0) = Π∗
X0 Z k
0 W0(I − P0),
k = 0, 1, 2, . . .
This proves (6.2).
Assume 0 < r < 1 such that the poles of G inside D are in the open disc
Dr. As mentioned in the introduction, the poles of G inside D coincide with the
eigenvalues of Z. Thus rspec(Z) < r. By Lemma 6.1, rspec(Z0) ≤ rspec(Z) < r.
In what follows we fix 0 < r0 < 1 such that rspec(Z0) < r0 < r. We will
show that the convergence of the central optimal solutions Fk in Theorem 3.1
is proportional to rk
0 .
For simplicity, we will adapt the notation of Section 5, and write γ, ∆, N and
M instead of γH 2
. Futhermore, we use the abbreviated
notation Pk, γk, Λk, Ξk, and ∆k for the operators PMk , γMk , ΛMk , ΞMk , and
∆Mk appearing in Section 4 for M = Mk.
and MH 2
, ∆H 2
, NH 2
p
p
p
p
As a first step towards the proof of our convergence result we prove the
following lemma.
Lemma 6.2. Assume conditions (C1) and (C2) are satisfied. Then ∆k →r2
and for k ∈ N large enough ∆k is invertible, and ∆−1
∆−1.
0
k →r2
0
∆,
Proof. We begin with a few remarks. Recall that for M in (2.1) the operator
p onto S∗M; see the first
RM is defined to be the orthogonal projection of H 2
paragraph of Section 2. For M = Mk we have S∗Mk = Mk−1 by (3.1),
and thus M = Mk implies RMk = Pk−1. It follows that the operator ∆k is
kIX − ZW Pk−1W ∗Z ∗; c.f., the second part of (4.11). From
given by ∆k = γ2
the invertibility of D◦
we obtain that ∆k is invertible as well; see the first
paragraph of the proof Corollary 4.5. The identities in (4.13) for M = Mk now
take the form
Mk
ΛkPkW ∗ = Pk−1W ∗Z ∗∆−1
k Ξk, ΛkE = −Pk−1W ∗Z ∗∆−1
k W E.
(6.3)
Observe that γ2
k = kHPkk2 = kPkH ∗k2 = rspec(HPkH ∗) = kHPkH ∗k. By a
similar computation γ2 = kHH ∗k. Thus, using (6.1) and (6.2),
γ2 − γ2
k = kHH ∗k − kHPkH ∗k ≤ kHH ∗ − HPkH ∗k
0 W0(I − P0)W ∗
0 k = kHk2 kZ k
= kHSk(I − P0)S∗kH ∗k = kZ k
≤ kZ k
0 k kHk k(I − P0)k kH ∗kkZ ∗k
0 Z ∗k
0 k
0 k2.
It follows that γ2
k →r2
0
γ2. Next, again by (6.1) and (6.2), we obtain
∆k = γ2
= γ2
= ∆ + (γ2
kI − ZW Pk−1W ∗Z ∗
kI − ZW W ∗Z ∗ + ZW Sk−1(I − P0)S∗k−1W ∗Z ∗
0 PX0 .
0 W0(I − P0)W0Z ∗k
k − γ2)I + PX0 Z k−1
18
Clearly the second and third summand converge to zero proportional to r2k
0 ,
and thus we may conclude that ∆k →r2
∆.
0
Since ∆ is invertible by Theorem 5.1. The result of the previous paragraph
k k < L for some L > 0
implies that for k large enough ∆k is invertible and k∆−1
independent of k. Consequently ∆−1
∆−1.
k →r2
0
Proof of Theorem 3.1 (with rk
0 -convergence). We split the proof into four
parts. Throughout k ∈ N is assumed to be large enough so that ∆k is invertible;
see Lemma 6.2.
Part 1. Let N and M be as in Theorem 5.1. Put
Nk,1(λ) = −ΠHW ∗(I − λZ ∗)−1Z ∗∆−1
Mk,1(λ) = I + λE∗W ∗(I − λZ ∗)−1Z ∗∆−1
k W E,
k W E.
(6.4)
(6.5)
Since the only dependence on k in Nk,1 and Mk,1 occurs in the form of ∆k, it
follows from Lemma 6.2 that
Mk,1 →r2
0
M and Nk,1 →r2
0
N.
Part 2. From Corollary 4.5 we know that
Nk(λ) = Nk,1(λ) + Nk,2(λ), Nk,2(λ) = ΠHΓk(λ)Z ∗∆−1
Mk(λ) = Mk,1(λ) + Mk,2(λ), Mk,2(λ) = −λE∗Γk(λ)Z ∗∆−1
k W E,
k W E.
(6.6)
(6.7)
(6.8)
Here Γk(λ) = (I − λS∗)−1(I − Pk−1)W ∗. In this part we show that Mk,2 →r0 0.
Using the first identity in (6.1), the intertwining relation ZW = W S, and
(6.2) we see that
Γk(λ) = (I − λS∗)−1Sk−1(I − P0)S∗k−1W ∗
0 Z ∗k−1
= (I − λS∗)−1Sk−1(I − P0)W ∗
0
ΠX0 .
Next we use that
(I − λS∗)−1Sk−1 =
k−2Xj=0
λjSk−1−j + λk−1(I − λS∗)−1.
Thus Γk(λ) = Γk,1(λ) + Γk,2(λ), where
Γk,1(λ) =(cid:16) k−2Xj=0
λj Sk−1−j(cid:17)(I − P0)W ∗
0 Z ∗k−1
0
ΠX0 ,
Γk,2(λ) = λk−1(I − λS∗)−1(I − P0)W ∗
0 Z ∗k−1
0
ΠX0 .
Now recall that M0 is S∗-invariant, and write S0 = P0SP0 = P0S. The fact
that M0 is finite dimensional implies rspec(S0) < 1. The computation
(I − λS∗)−1(I − P0)W ∗
0 = (I − λS∗)−1W ∗
= W ∗
0 (I − λZ ∗
0 − (I − λS∗)−1P0W ∗
0
0 )−1P0W ∗
0 ,
0 )−1 − (I − λS∗
19
shows that (I −λS∗)−1(I −P0)W ∗
r0 < 1, we conclude that Γk,2 →r0 0.
0 is uniformly bounded on D. Since rspec(Z0) <
Next observe that E∗(cid:0)Pk−2
each k ∈ N. We conclude that
j=0 λjSk−1−j(cid:1) = 0, and thus E∗Γk,1(λ) = 0 for
Mk,2(λ) = −λE∗Γk,2(λ)Z ∗∆−1
k W E.
But then Γk,2 →r0 0 implies that the same holds true for Mk,2, that is, Mk,2 →r0
0. Indeed, this follows from the above identity and the fact that the sequence
∆−1
Part 3. In this part we show that Nk,2 →r0 0. To do this we first observe that
is uniformly bounded.
k
ΠHSk−1−j = ΠV k−1−j
−
H = ΠV k−1−j
−
PX W = ΠPX Z k−1−jW.
Post-multiplying this identity with I − P0 and using (6.2) yields
ΠHSk−1−j(I − P0) = ΠPX0 Z k−1−j
0
W0(I − P0).
It follows that
Nk,2(λ) =(cid:16) k−2Xj=0
λj ΠPX0 Z k−1−j
0
(cid:17)W0(I − P0)W ∗
0 Z ∗k−1
0
+
+ ΠHΓk,2(λ)Z ∗∆−1
k W E.
(6.9)
From the previous part of the proof we know that Γk,2 →r0 0, and by Lemma
6.2 the sequence ∆−1
is uniformly bounded. It follows that the second term in
the right hand side of (6.9) converges to zero with a rate proportional to rk
0 .
Note that for λ ∈ D we have
k
k
k−2Xj=0
λj ΠPX0 Z k−1−j
0
k ≤
k−2Xj=0
kZ0kk−1−j ≤
∞Xj=1
kZ j
0k ≤
L0r0
1 − r0
.
0
0 . We conclude that Nk,2 →r0 0.
Since rspec(Z0) < r0 < 1, we also have kZ ∗k−1
k →r0 0. It follows that the first
term in the right hand side of (6.9) converges to zero with a rate proportional
to rk
Part 4. To complete the proof, it remains to show that M −1
k (λ) →r0 M −1(λ)
uniformly on D. By similar computations as in the proof of Lemma 6.2, it
Z ∗∆−1Ξ. By Theorem 5.1 we
follows that Ξk →r2
have rspec(Z ∗∆−1Ξ) < 1. Thus for k large enough also rspec(Z ∗∆−1
k Ξk) < 1,
and Mk,1(λ) is invertible on D. From the fact that Mk,1 →r2
M , we see that
M −1
k,1 and M −1 indicating here the functions on D with
values Mk,1(λ)−1 and M (λ)−1 for each λ ∈ D. In particular, the functions M −1
k,1
are uniformly bounded on D by a constant independent of k, which implies
Ξ. Hence Z ∗∆−1
M −1, with M −1
k Ξk →r2
k,1 →r2
0
0
0
0
I + M −1
k,1 Mk,2 →r0 I,
(I + M −1
k,1 Mk,2)−1 →r0 I.
20
As a consequence
M −1
k = (Mk,1 + Mk,2)−1 = (I + M −1
k,1 Mk,2)−1M −1
k,1 →r0 I · M −1 = M −1,
which completes the proof.
rather than rk
Concluding remarks
Note that the functions Mk,1 and Nk,1 given by (6.4) and (6.5) converge with a
rate proportional to r2k
0 ; cf., (6.6). Consequently the same holds
0
k,1. Thus a much faster convergence may be achieved when Nk,1M −1
true for M −1
k,1
are used instead of NkM −1
k . However, for the inverse of Mk,1 to exist on D we
need k to be large enough to guarantee rspec(Z ∗∆kΞk) < 1, and it is at present
not clear how large k should be.
For the scalar case condition (C2) is rather natural.
Indeed (see the second
paragraph of Section 3) for the scalar case condition (C2) is equivalent to the
requirement that the largest singular value of the Hankel operator is simple. The
latter condition also appears in model reduction problems. In the matrix-valued
case (C2) seems rather special. We expect that a version of Theorem 3.1 can
be proved by only using the first part of (C2), that is, by assuming that none of
the maximizing vectors of the Hankel operator belongs to SH 2
p ; cf., Proposition
4.4. However, note that in that case the optimal solution of the Nehari problem
may not be unique.
Computational examples show that it may happen that the approximations
of the optimal solution to the Nehari problem considered in this paper oscillate
to the optimal solution when the initial space M0 = {0}. Although the rate of
convergence can be improved considerably by choosing a different initial space
M0, the same examples show that the approximations still oscillate in much the
same way as before to the optimal solution. This suggests that approximating
the optimal solution may not be practical in some problems. In this case, one
may have to adjust these approximating optimal solutions. We plan to return
to this phenomenon in a later paper.
Acknowledgement. The authors thank Joe Ball for mentioning the Helton-
Young paper [14] to the second author.
References
[1] Adamjan, V. M.; Arov, D. Z.; Kreın, M. G. Infinite Hankel matrices and
generalized problems of Carath´eodory-Fej´er and F. Riesz. Functional Ana-
lyis and Applications 2 (1968), 1 -- 18 [English translation].
[2] V.M. Adamjan, D.Z. Arov, M.G. Krein, Infinite block Hankel matrices and
their connection with the interpolation problem. Amer. Math. Soc. Transl.
(2) 111 (1978), 133 -- 156 [Russian original 1971].
[3] D.Z. Arov and H. Dym, J-contractive matrix valued functions and related
topics, Cambridge University Press, 2008.
21
[4] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran, Factorization of
matrix and operator functions: the state space method, OT 178, Birkhauser
Verlag, Basel, 2008.
[5] M.J. Corless and A.E. Frazho, Linear systems and control, Marcel Dekker,
Inc., New York, 2003.
[6] C. Foias and A. E. Frazho, The Commutant Lifting Approach to Interpo-
lation Problems, OT 44, Birkhauser-Verlag, Basel, 1990.
[7] C. Foias, A. E. Frazho, I. Gohberg and M.A. Kaashoek, Metric Constrained
Interpolation, Commutant Lifting and Systems, OT 100, Birkhauser-
Verlag, Basel, 1998.
[8] C. Foias, A.E. Frazho, and M.A. Kaashoek, Relaxation of metric con-
strained interpolation and a new lifting theorem, Integral Equations Op-
erator Theory 42 (2002), 253 -- 310.
[9] B.A. Francis, A Course in H∞ Control Theory, Lecture Notes in Control
and Information Sciences 88, Springer, Berlin, 1987.
[10] A.E. Frazho, S. ter Horst, and M.A. Kaashoek, All solutions to the relaxed
commutant lifting problem, Acta Sci. Math. (Szeged) 72 (2006), 299 -- 318.
[11] A.E. Frazho, S. ter Horst, and M.A. Kaashoek, Relaxed commutant lifting:
an equivalent version and a new application, in: Recent Advances in Op-
erator Theory and Applications, OT 187, Birkhauser Verlag, Basel, 2009,
pp. 157 -- 168.
[12] I. Gohberg, S. Goldberg and M.A. Kaashoek,Classes of Linear Operators,
Volume I, OT 49, Birkhauser Verlag, Basel, 1990.
[13] I. Gohberg, S. Goldberg and M.A. Kaashoek,Classes of Linear Operators,
Volume II, OT 63, Birkhauser Verlag, Basel, 1993.
[14] J.W. Helton, and N.J. Young, Approximation of Hankel operators: trun-
cation error in an H ∞ design method, in: Signal Processing, Part II, IMA
Vol. Math. Appl., 23 Springer, New York, 1990, pp. 115 -- 137.
[15] S. ter Horst, Relaxed commutant lifting and Nehari interpolation, Ph.D.
dissertation, VU University, Amsterdam, 2007.
[16] S. ter Horst, Relaxed commutant lifting and a relaxed Nehari problem:
Redheffer state space formulas, Math. Nachr. 282 (2009), 1753 -- 1769.
[17] V.V. Peller, Hankel Operators and their Applications, Springer Monographs
in Mathematics, Springer 2003.
[18] D. Sarason, Generalized interpolation in H ∞, Trans. Amer. Math. Soc.
127 (1967), 179 -- 203.
22
|
1002.0454 | 1 | 1002 | 2010-02-02T11:00:43 | On Stochastic generalized functions | [
"math.FA",
"math.PR"
] | We introduced a new algebra of stochastic generalized functions which contains to the space of stochastic distributions G, [25]. As an application, we prove existence and uniqueness of the solution of a stochastic Cauchy problem involving singularities. | math.FA | math |
ON STOCHASTIC GENERALIZED FUNCTIONS
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
Abstract. We introduced a new algebra of stochastic generalized
functions which contains to the space of stochastic distributions G ∗,
[25]. As an application, we prove existence and uniqueness of the
solution of a stochastic Cauchy problem involving singularities.
1. Introduction
The algebras of generalized functions were introduced by J. F. Colombeau
[8] and has been studied and developed by many authors (see [7], [8],
[12] and [22] and references). These algebras of generalized functions
contain the Schwartz distributions, thus one can deal with its multipli-
cation and other types of nonlinearities. Many applications have been
carried out in various fields of mathematics such as partial differen-
tial equations, Lie analysis, local and microlocal analysis, probability
theory, differential geometry. In particular, the algebras of generalized
functions found applications in non-linear partial differential equations
([7], [8], [12] and [22]), where the classical distributional methods are
limited by the impossibility to consistently define an intrinsic product
of distributions [27].
In recent years stochastic partial differential equations have been stud-
ied actively. Often, solutions to such equations do not exist in the usual
sense, rather they are generalized functions. It is possible consider the
solutions as generalized functions in the space-time variables or as gen-
eralized stochastic functions which are point-wise defined in space and
time. In this work we take the latter approach, more precisely we con-
sider stochastic process in algebras of generalized functions, following
to F. Russo and their coauthors (see for instance [2], [3] ,[23] and [26].
Key words and phrases. White noise, Wick product, Generalized functions,
Colombeau Algebras, parabolic equation.
2000 Mathematics Subject Classification. Primary: 46F30, 60G20 ; Secondary:
60B99.
Research partially supported by FAPESP 02/10246-2, 2007/54740-4 and CNPQ
302704/2008-6.
1
2
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
We considerer the following stochastic Cauchy problem,
(1)
(cid:26) ut
u(0, x) = f (x)
= Lu + uW (t, x)
where the product is taken in the usual form, L is an uniformly elliptic
partial differential operator, f ∈ C 2
b (Rl) and W is a space-time white
noise [14]. There are a great interest into solve the problem (1) (see
for instance [10], [15], [19], [20], [17], [21], [24] and [30]). We observe
that the properties of the solutions (Ito-Stratonovich) of (1) depend
crucially on the noise type and its dimension (see [17] and [30]).
The aim of this article is solve the stochastic Cauchy problem (1) with-
out restriction on the dimension of the white noise W . Our approach
transfer the generalized nature of the solutions to the stochastic com-
ponent (see [15] for more details).
It is thus necessary to introduce
random variables with values in algebras of generalized functions. The
fundamental idea is to introduce a new algebra of generalized functions
of Colombeau type in the context of the Hida stochastic distributions.
This algebra of generalized functions contains the stochastic distribu-
tions G∗ introduced by J. Potthoff and M. Timpel in [25]. In this con-
text and under general assumptions we show existence and uniqueness
for the stochastic Cauchy problem (1). Furthermore, others aspects of
these generalized functions are studied.
The plan of exposition is as follows: in section 2, we provide some basic
concepts on classical white noise theory, these are, Wiener-Ito chaos ex-
pansions, stochastic distributions and Wick products. In section 3, we
introduce three definitions of products of stochastic distributions and
we study some properties. In section 4, we give a new algebra of sto-
chastic generalized functions G, this algebra contains to the stochastic
distributions G∗ ( see [25]) and extends the product in the test space
G. Moreover, we study its elementary properties and we show that the
symmetric product of stochastic distributions , defined in section 3, is
associated with the product in G. Finally, in section 5, we study a
stochastic Cauchy problem with space-time depend noise white in the
context of these generalized functions.
2. White Noise Theory
2.1. Preliminaries. Let S(Rd) be the space of infinitely differentiable
functions from Rd to R (C) which together with all its derivatives are
of rapidly decreasing. We consider S(Rd) with the topology given by
the family of seminorms
kfkα,β = sup{xαDβf (x) : x ∈ Rd}
ON STOCHASTIC GENERALIZED FUNCTIONS
3
where α, β ∈ Nd
0. It follows that S(Rd) is a nuclear space. The Schwartz
space of tempered distributions is the dual space S′(Rd).
We shall summarize the definitions and basic properties of Hermite
polynomials and Hermite functions. The j-th Hermite polynomial, de-
noted by hj, is defined to be
hj(x) = (−1)ne
x2
2
dj
dxj e− x2
2
(2)
for j ∈ N0.
We observe that
(3)
hj(x) = 2− j
2
[j/2]
Xk=0
(−1)kn!(√2x)j−2k
k!(j − 2k)!
.
The j-th Hermite function ηj is given by
ηj(x) = (√2πn!)− 1
2 e− 1
4 x2
hj(x)
(4)
for j ∈ N0.
useful.
The following properties of the Hermite functions will be extremely
• {ηj : j ∈ N0} is an orthonormal basis of L2(R),
• sup{ηj(x) : x ∈ R} = O(j− 1
4 ),
• The ηj are eigenfunctions of the number operator N + 1 =
1
2(− d2
dx2 + x2 + 1),
(N + 1)ηj = (j + 1)ηj
for all j ∈ N0.
The α-th Hermite function ηα : Rd → R (α ∈ Nd
0) is given by
ηα(x) =
d
Yi=1
ηαi(xi).
It follows that {ηα : α ∈ Nd
0} is an orthonormal basis of L2(Rd).
Let us denote by (N + 1)d = (N + 1)⊗d, the differential operator
(N + 1)⊗d =
1
2
(−
∂2
∂x2
d
+ x2
d + 1) · · ·
1
2
(−
∂2
∂x2
1
+ x2
1 + 1).
It is easy to check that for α, β ∈ Nd
0,
(N + 1)β
d (ηα) = (α + 1)βηα
where (α + 1)β = Πd
i=1(1 + αi)βi.
4
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
The topology of S(Rd) has an alternative description in terms of the
family of norms { · β : β ∈ Nd
d (ϕ)2
0}. The norm · β is given by
0 = Xα∈Nd
(α + 1)2β < ϕ, ηα >2,
0
ϕ2
β := (N + 1)β
where < ϕ, ηα >= R ϕ(x)ηα(x)dx are the Fourier-Hermite coefficients
of the expansion of ϕ.
It follows easily that the seminorm families {·β : β ∈ Nd
0} and {k·kα,β :
α, β ∈ Nd
The Hermite representation theorem for S(Rd) (S′(Rd)) states a topo-
logical isomorphism from S(Rd) (S′(Rd)) onto the space of sequences
sd (s′d).
Let sd be the space of sequences
0} are equivalent on S(Rd).
sd = {(aβ) ∈ ℓ2(Nd)) : Xβ
(β + 1)2α aβ 2< ∞, for all α ∈ Nd
0}.
The space s is a locally convex space with the sequence of norms
(aβ)α = (Xβ
(β + 1)2α aα 2)
1
2
The topological dual space to sd, denoted by s′d, is given by
s′d = {(bβ) : for some (C, α) ∈ R × Nd
0, bβ ≤ C(β + 1)α for all β},
and the natural pairing of elements from sd and s′d, denoted by h·,·i, is
given by
h(bβ), (aβ)i = Xβ
bβaβ
for (bβ) ∈ s′d and (aβ) ∈ sd.
It is clear that s′d is an algebra with the pointwise operations:
(bβ) + (b′β) = (bβ + b′β)
(bβ) · (b′β) = (bβb′β),
and sd is an ideal of s′d.
The relation between sd (s′d) and S(Rd) (S′(Rd)) is induced by the
Hermite functions, via the Fourier-Hermite coefficients.
Theorem 1. N-representation theorem for S(Rd) and S′(Rd):
a) Let h : S(Rd) → sd be the application
h(ϕ) = (< ϕ, ηβ >).
Then h is a topological isomorphism. Moreover,
kh(ϕ)kβ = kϕkβ
ON STOCHASTIC GENERALIZED FUNCTIONS
5
for all ϕ ∈ S(Rd).
b) Let H : S′(Rd) → s′d be the application H(T ) = (T (ηβ)). Then H is
a topological isomorphism. Moreover, if T ∈ S′(Rd) we have that
T = Xβ
T (ηβ)ηβ
in the weak sense and for all ϕ ∈ S(Rd),
T (ϕ) = hH(T ), h(ϕ)i.
Proof. See for instance, M. Reed and B. Simon [28] pp. 143.
(cid:3)
We say that the sequences h(ϕ) and H(T ) are the Hermite coefficients
of the tempered function ϕ and the tempered distribution T , respec-
tively.
2.2. White noise and chaos expansions. In this subsection we
briefly recall some of the basic concepts and results of White noise anal-
ysis. Our presentation and notation will follow [14] and [15]. According
to the Bochner-Minlos theorem, there exits an unique probability µ on
Bd the Borel σ-field of S′(Rd) such that for each φ ∈ S(Rd),
(5)
ei(ω,φ) dµ(ω) = e− 1
L2(Rd).
2φ2
ZS ′(Rd)
The probability space (S′(Rd),Bd, µ) is then called the White noise
probability space. We will denote by (L2) the space L2(S′(Rd),B, µ).
Let {αj : j ∈ N0} be an enumeration of the set of multi-indices {α : α ∈
Nd
0}. We can assume that this enumeration has the following property:
If k < l then αk
1 + ... + αk
d (see [15] pag. 19 for details).
It is clear that {ηα : α ∈ Nd
0} can be enumerated by {ηj : j ∈ N0}. We
denote by J the set of sequences of nonnegative integers with compact
support.
The α-th Hermite polynomial Hα : S′(Rd) → R (α ∈ J ) is given by
1 + ... + αl
d ≤ αl
Hα(ω) =
hαj (< ω, ηj >).
∞
Yj=1
It follows that {Hα : α ∈ J } is an orthogonal basis of (L2) and kHαk2 =
α!. The Wiener-Ito chaos expansion theorem says that if f ∈ (L2), then
there exists an unique sequence of functions fn ∈ L2(Rd) ⊗n such that
f (ω) =
In(fn)(ω)
∞
Xn=0
6
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
where In : L2(Rd) ⊗n → (L2) is the multiple Wiener integral of order n.
Moreover, we have the isometry
E[f 2] =
∞
Xn=0
n!fn2
0.
where we denote by · 0 the norm · L2(Rd) ⊗n.
Let f ∈ S(R), from (5) for d = 1 we have that < ·, f > is a Gaussian
random variable with expectation 0 and variance f0. We observe that
for t ∈ R,
(6)
is a Brownian motion. Moreover, if f ∈ L2(R) we have that
B(t) :=< ·, 1[0,t] >
< ·, f >= I1(f ).
2.3. Stochastic distributions. The space of stochastic distributions
(also called of space of regular Hida distributions) G∗ has been intro-
duced by J. Potthoff and M. Timpel in [25]. Their development was
motivated by stochastic differential equations of Wick type and its ap-
plications, see for instance [1], [5], [6] and [13].
Let N be the Ornstein-Uhlenbeck operator on (L2), we recall that
N(In(f )) = nIn(f )
for all n ∈ N0 and f ∈ L2(Rd) ⊗n.
Let us denote by G√ (p ∈ N0) the (L2)-domain of exp(pN). It is clear
that G√ with the norm
∞
Xn=0
0
0 =
e2pnn!Fn2
p := kexp(pN)Fk2
kFk2
is a Hilbert space.
Let G be the projective limit of the family {G√, p ≥ 0} and G∗ be its
dual. It follows that G∗ is the inductive limit of the family {G−√, p ≥ 0},
where G−√ is the dual of G√. In this way we obtained the Gelfand triple
We observe that the duality in this Gelfand triple is given by
G ⊂ (L2) ⊂ G∗.
≪ F, f ≫=
∞
Xn=0
n! < Fn, fn >
where f = P∞n=0 In(fn) ∈ G and F = P∞n=0 In(Fn) ∈ G∗.
ON STOCHASTIC GENERALIZED FUNCTIONS
7
Example 1. The Brownian motion B(t) =< ·, 1[0,t] > is an element of
G. J. Potthof and M. Timpel [25] showed that for x ∈ R the Donsker
delta function δx ◦ B(t) belong to G∗.
Let F ∈ G∗ with Wiener-Ito expansion P∞i=0 In(Fn). The S-transform
of F ∈ G∗, denoted by S(F ), is the function from S(Rd) to C given by
S(F )(ϕ) =
< Fn, ϕ⊗n > .
∞
Xn=0
The Wick product F ⋄ G of F and G in G∗, is given by S−1(S(F )S(G)).
This product is well defined and continuous from G∗×G∗ to G∗, see [6],
[16] and [25].
We observe that the Wick product is associative, commutative and G
and G∗ are algebras respect to the Wick product. An easy computation
shows that if F = P∞i=0 Ii(Fi) and G = P∞i=0 Ii(Gi) belong to G∗ then
the Wiener-Ito coefficients of the Wick product F ⋄ G = P∞i=0 Ii(Hi)
verify that
Hi = Xi=m+n
Fn ⊗Gm.
The Wick exponential of f ∈ L2(Rd), denoted by e:<·,f >, is the element
of (L2) defined by
(7)
∞
Xm=0
1
m!
Im(f⊗m
) = e<·,f >− 1
2f2
0.
We observe the following useful identity
(8)
ke:<·,f >k2
p =
∞
Xm=0
e2pm
m! f2m
0 = ee2pf2
0.
The space G is an algebra for the pointwise multiplication and this
multiplication is continuous from G ×G into G, see Theorem 2.5 of [25]
(d = 1) and [6] for the general case. In particular, for all m ∈ N0 there
exists r, s ∈ N0 and a constant Cm > 0 such that
(9)
for all ϕ, ψ ∈ G.
Thus we can define the product of a distribution by a test function.
Definition 1. Let ϕ ∈ G and F ∈ G∗. We define ϕF ∈ G∗ by
kϕψkm ≤ Cmkϕkrkψks
≪ ϕF, ψ ≫=≪ F, ϕψ ≫
for all ψ ∈ G.
8
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
It is clear from (9) that ϕF is well defined.
Let ϕ = P∞n=0 In(ϕn) ∈ G and F = P∞m=0 Im(Fm) ∈ G∗. Then ϕF =
P∞j=0 Ij(Hj) where
Hj = Xj=m+n
k (cid:19) ϕn+k ⊗kFm+k.
Xk=0
k! (cid:18)n + k
k (cid:19) (cid:18)m + k
(10)
m∧n
Here ϕn ⊗kFm is the symmetrization of ϕn ⊗k Fm. See for instance [16]
pag. 70.
3. Distributional products via approximation in chaos
In this section, strongly inspired in [9], we define three new products
of stochastic distributions. A possible approach to define a product of
a pair of distributions is approximate one of them by test functions,
multiply this approximation by the other distribution, and pass to a
In the case of the sequential approach (see [4] pp 242, [22])
limit.
the approximation is done by convolution with δ-sequences,
in this
work we propose take the approximation given by the Wiener-Ito chaos
expansion of the distribution.
In order to define the chaos approximation we consider the projections
Πm : G∗ → G (m ∈ N0). These projections are defined by
ΠmF =
m
Xn=0
In(Fn),
where F = P∞n=0 In(Fn) ∈ G∗.
Proposition 1. a) Let f = P∞n=0 In(fn) ∈ G. Then
Πmf = f.
lim
m→∞
b) Let F = P∞n=0 In(Fn) ∈ G∗. Then
ΠmF = F.
lim
m→∞
Proof. a) Let p > 0 and ε > 0. Since f ∈ G, there exists m0 ∈ N0 such
that
∞
Xn=m+1
n!e2pn fn2
0 < ε2
for m ≥ m0.
ON STOCHASTIC GENERALIZED FUNCTIONS
9
Then
kΠmf − fk2
p =
∞
Xn=m+1
n!e2pn fn2
0
< ε2
for m ≥ m0. We conclude that limm→∞ Πmf = f .
b) Let ε > 0. Since F ∈ G∗ there exists q ∈ N0 such that
kFk2
−q =
n!e−2qn Fn2
0 < ∞.
∞
Xn=0
Then there exists m0 ∈ N0 such that
∞
Xn=m+1
n!e−2qn Fn2
0 < ε2.
for m ≥ m0.
Thus
kΠmF − Fk2
−q =
∞
Xn=m+1
n!e−2qn Fn2
0
< ε2,
for m ≥ m0.
Now, we introduce the following products of stochastic distributions.
(cid:3)
Definition 2. Let F and G be stochastic distributions. The products
F •r G, F •l G and F • G are, by definition,
(11)
F (ΠmG)
F •r G = lim
m→∞
(12)
(13)
F •l G = lim
m→∞
(ΠmF )G
F • G = lim
m→∞
(ΠmF )(ΠmG).
where the limits are taking in the weak sense.
Proposition 2. The products (11), (12) and (13) extend the product
of elements of G and G∗.
Proof. It follows from the bi-continuity of the product of elements of G
and G∗.
(cid:3)
10
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
We recall that two elements F = P∞n=0 In(Fn) and G = P∞n=0 In(Gn) of
G∗ are strongly independent (see F. Benth and J. Potthoff [5]), if there
exists two intervals IF and IG whose intersection has Lebesgue measure
zero such that supp(Fn) ⊆ I n
Proposition 3. Let F and G be strongly independent in G∗. Then
F and supp(Gn) ⊆ I n
G for all n ∈ N.
F • G = F ⋄ G.
Proof. Let F = P∞n=0 In(Fn) and G = P∞k=0 Ik(Gk) be strongly inde-
pendent in G∗. It is clear that limm→∞(ΠmF )(ΠmG) is equal to
lim
m→∞
Xn=0
Xk=0
n∧k
Xj=0
m
m
k! (cid:18)n
j(cid:19)(cid:18)k
j(cid:19)In+k−2j(Fn ⊗jGk).
Since F and G are strongly independent in G∗,
n∧k
Xj=0
k! (cid:18)n
j(cid:19)(cid:18)k
j(cid:19)In+k−2j(Fn ⊗jGk) = In+k(Fn ⊗Gk).
Combining the above equalities we conclude that
F • G = lim
m→∞
(ΠmF )(ΠmG)
m
m
Xk=0
In+k(Fn ⊗Gk)
Xn=0
= lim
m→∞
= F ⋄ G.
(cid:3)
4. Stochastic generalized functions
In order to introduce an algebra of Colombeau type such that include
the stochastic distributions we consider G N0 the space of sequences of
G. It is clear that G N0 has the structure of an associative, commutative
algebra with the natural operations:
(ϕn) + (ψn) = (ϕn + ψn)
(14)
(15)
(16)
where (ϕn) and (ψn) are in G and a ∈ R.
We introduce the following sequence spaces. Let e be given by
a(ϕn) = (aϕn)
(ϕn) · (ψn) = (ϕnψn)
e = {(an) ∈ ℓ2 : lim
n→∞
enman = 0, for all m ∈ N0}.
ON STOCHASTIC GENERALIZED FUNCTIONS
11
We consider e equipped with the sequence of norms (m ∈ N0)
(an)m = (
e2nm an 2)
1
2
∞
Xn=0
or with the equivalent sequence of norms
(an) m,∞= sup
n
enman.
It is clear that e is a locally convex space.
Let us denote by e′ the topological dual space to e (see [29]). It follows
that e′ is given by
e′ = {(bn) : for some (C, m) ∈ R × N0, bn ≤ Ce2nm for all n}.
The natural pairing of elements from e and e′, denoted by h·,·i, is given
by
h(bn), (an)i =
∞
Xn=0
bnan
for (bn) ∈ e′ and (an) ∈ e.
We observe that e′ is an algebra with the pointwise operations and e
is an ideal.
Definition 3. Let
Ge = {(ϕm) ∈ G N0
: ∀ p ∈ N0 (kϕmkp) ∈ e },
and
Ge′ = {(ϕm) ∈ G N0
: ∀ p ∈ N0 (kϕmkp) ∈ e′ }.
Lemma 1. Ge′ is a subalgebra of G N0 and Ge is an ideal of Ge′.
Proof. Let (ϕn), (ψn) ∈ Ge′ and m ∈ N0. From (9) it follows that there
exists r, s ∈ N0 and a constant Cm > 0 such that
kϕnψnkm ≤ Cmkϕnkrkψnks.
By definition, there exists constants D, E > 0 and p, q ∈ N0 such that
kϕnkr ≤ Denp
kψnks ≤ Eenq.
Combining these inequalities, we obtain
kϕnψnkm ≤ CmDEen(p+q).
This proves that (kϕnψnkm) ∈ e′, thus (ϕn) · (ψn) ∈ Ge′.
12
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
It remains to prove that Ge is an ideal of Ge′. Let (ϕn) ∈ Ge′, (ψn) ∈ Ge
and m ∈ N0. By definitions, we have that for each r ∈ N0 there exists
a constant D > 0 and p ∈ N0 such that
and for all s, l ∈ N0,
kϕnkr ≤ Denp
(kψnks)2
l =
∞
Xn=0
e2lpkψnk2
s < ∞.
Combining the inequality (9) with the above equations we obtain
(kϕnψnkm)2
l =
∞
Xn=0
≤ C 2
m
≤ C 2
< ∞.
Xn=0
mD2
m
e2nlkϕnψnk2
∞
rkψnk2
s
e2nlkϕnk2
∞
Xn=0
e2n(l+p)kψnk2
s
We have proved that (kϕnψnkm) ∈ e, for all m ∈ N0, this is (ϕn)·(ψn) ∈
Ge.
Proposition 4. Let F ∈ G∗. Then (Fm) ∈ Ge′, where Fm = ΠmF .
Proof. Since F = P∞n=0 In(Fn) ∈ G∗, there exists q ∈ N0 such that
(cid:3)
kFk2
−q =
Then
∞
Xn=0
n! e−2qn Fn2
0 < ∞.
m
kFmk2
p =
Xn=0
n! e2pn Fn2
≤ e(2p+2q)m kFk2
−q
0
for all p ∈ N0. This proves that (Fm) ∈ Ge′.
Definition 4. The algebra of stochastic generalized functions is defined
as
(cid:3)
G = Ge′/Ge.
The elements of G are called stochastic generalized functions.
Let (ϕm) ∈ Ge′ we will denote by [ϕm] the equivalent class (ϕm) + Ge.
ON STOCHASTIC GENERALIZED FUNCTIONS
13
Proposition 5. Let ι : G∗ → G be the application
ι(F ) = [Fm].
Then ι is a linear embedding. Moreover,we have that for all ϕ ∈ G,
ι(ϕ) = [ϕ].
Proof. It is clear from the above Proposition, that ι is a well defined
linear application. Let F ∈ G∗ such that ι(F ) = 0. As (Fm) ∈ Ge we
have limm→∞ kFmkp = 0, for all p ∈ N0. This is the sequence (Fm)
converge weakly to 0, it follows that F = 0.
It remains to prove that ι(ϕ) = [ϕ], for all ϕ ∈ G. Let ϕ = P∞n=0 In(fn) ∈
G and p ∈ N0. Then
e2amkϕ − Πmϕk2
p = e2am
∞
0
n! e2pn fn2
Xn=m+1
n! e2n(p+a) fn2
0,
∞
Xn=m+1
≤
for all a ∈ N0.
Since ϕ ∈ G, we have
∞
Xn=m+1
n! e2n(p+a) fn2
0 < ∞.
Combining these facts, we have that
lim
m→∞
e2amkϕ − Πmϕk2
p = 0
for all a ∈ N0. This shows that (ϕ − Πmϕ) ∈ Ge, which completes the
proof.
(cid:3)
Corollary 1. Let ϕ, ψ ∈ G. Then
Proof. We first observe that
ι(ϕψ) = ι(ϕ) · ι(ψ).
(ϕψ) − (ϕm) · (ψm) = (ϕ) · (ψ − ψm) + (ϕ − ϕm) · (ψ).
Applying Proposition 5 and Lemma 1 we obtain (ϕψ) − (ϕm) · (ψm) ∈
Ge. Therefore ι(ϕψ) = ι(ϕ) · ι(ψ).
Now, we give some examples of stochastic generalized functions.
(cid:3)
14
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
Example 2. The white noise generalized function is defined by
Wx = [Wm(x)] = [I1(
m
Xj=0
ηj(x)ηj(·))].
For d = 1 we have the following identity
√n + 1
x − t (cid:16)ηn+1(x)ηn(t) − ηn+1(t)ηn(x)(cid:17).
(17)
Thus
n
ηj(t)ηj(x) =
Xj=0
Wx = [Z √n + 1
x − t (cid:16)ηn+1(x)ηn(t) − ηn+1(t)ηn(x)(cid:17) dB(t)].
Example 3. The element W 2
x . From the above example we have
W 2
x = [Wm(x)2] = [I1(
m
Xj=0
ηj(x)ηj(·)) I1(
m
Xj=0
ηj(x)ηj(·))].
From the formula (10) we get
W 2
x = [Wm(x) ⋄ Wm(x) +
m
Xj=0
η2
j (x)].
Example 4. Let x ∈ R and t > 0, the Donsker delta is the stochastic
generalized function defined by
m
ι(δx(Bt)) = [
Xn=0
1
n!
t−nHn,t(x)Hn,t(Bt)],
where Hn,t is the n-th Hermite polynomial with variance t (see for in-
stance [14]).
Definition 5. Let [ϕm] and [ψm] be stochastic generalized functions.
We say that [ϕm] and [ψm] are associated, denoted by [ϕm] ≈ [ψm], if
for all ϕ ∈ G we have that
lim
m→∞ ≪ ϕm − ψm, ϕ ≫= 0.
We observe that the relation ≈ is well defined as an equivalence relation
on G.
Lemma 2. Let F and G be in G∗. Then:
(1) If ι(F ) ≈ ι(G) then F = G.
(2) If there exists F • G then ι(F ) · ι(G) ≈ ι(F • G).
(3) If ψ ∈ G and F ∈ G∗ then ι(ψ)ι(F ) ≈ ψF .
ON STOCHASTIC GENERALIZED FUNCTIONS
15
Proof. (1) Combining Proposition 1 and ι(F ) ≈ ι(G) we conclude that,
≪ F − G, ϕ ≫= lim
m→∞ ≪ (ΠmF − ΠmG) ϕ ≫= 0
for all ϕ ∈ G. Thus F = G.
The proofs of (2) and (3) are straightforward.
(cid:3)
Remark 1. We observe that our construction of stochastic general-
ized functions also can be done for others space of test functions and
distributions spaces, for instance Hida and Kondratiev spaces.
We introduce the generalized sequences ring g with the purpose of to
define the expectation of generalized functions of G.
Definition 6. The generalized sequences ring, denoted by g, is define
to be
(18)
Let (bn) ∈ e′ we will use [bn] by denoted the equivalent class (bn) + e.
It is clear from the definition that e′ is a ring but is not a field. In
g = e′/e.
fact, e′ has zero divisors.
Lemma 3. a) Let ι0 : R → g be the application
ι0(a) = [a].
Then ι0 is an embedding.
b) G is a g-module with the natural operations.
Proof. The proof is straightforward.
(cid:3)
We define the notion of association in g.
Definition 7. Let [an] and [bn] be in g. We says that [an] and [bn] are
associated, denoted by [an] ≈ [bn], if
lim
n→∞
(an − bn) = 0.
We observe that the relation ≈ is well defined equivalence relation on
g.
Proposition 6. Let [ϕm] = [ψm] ∈ G. Then [E(ϕm)] = [E(ψm)] ∈ g.
Proof. Let [ϕm] ∈ G. By definition, (ϕm) ∈ Ge′. Thus for all p ∈ N0
we have that (kϕmkp) ∈ e′. It follows that for all p ∈ N0 there exists
given C > 0 and a ∈ N0 such that
kϕmk2
p ≤ C 2e2am
for all m ∈ N0.
16
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
By definition of k · kp, we have that E(ϕm)2 ≤ kϕmk2
p. Thus
E(ϕm) ≤ Ceam
for all m ∈ N0. This implies that [E(ϕm)] ∈ g.
It remains to prove the independence of the representative. By
definition, (ϕm − ψm) ∈ Ge. Thus for all p ∈ N0 we have that
(kϕm − ψmkp) ∈ e. It follows that given a, p ∈ N0 and ε > 0 there
exists m0 ∈ N0 such that
e2amkϕm − ψmk2
p < ε2
for all m ≥ m0.
By the definition of k·kp, we have that E(ϕm)−E(ψm)2 ≤ kϕm−ψmk2
p.
Thus
eamE(ϕm) − E(ψm) < ε
for all m ≥ m0. We conclude that [E(ϕm)] = [E(ψm)].
Definition 8. Let F = [ϕm] be a generalized function in G. The
expectation of F , denoted by E(F ), is defined to be the generalized
number
(cid:3)
E(F ) = [E(ϕm].
Proposition 7. Let f ∈ L1(µ) and F = [ϕm] ∈ G such that F ≈ f .
Then
ι0(E(f )) = E(F ).
Proof. The proof is straightforward from the definitions.
(cid:3)
5. A Stochastic Cauchy Problem
In this section we study the stochastic Cauchy problem
(19)
(cid:26) ut = Lu + uW (t, x)
u0 = f
where L is an uniformly elliptic partial differential operator, f ∈ C 2
b (Rl)
and W is the white noise parametric generalized function defined below
in Example 5.
In order to solve (19) we introduce the algebra Gα(D) of parametric
generalized functions. We proceed in a similar way to the construction
of the algebra G, the details are left to the reader.
Let D a domain in Rd and α ∈ Nd
0. We denote by C α(D,G) the set of
functions f : D × S′(Rd) → R (C) such that f (x,·) ∈ G for all x ∈ D
and f (·, ω) ∈ C α
b (D) for almost ω ∈ S′(Rd).
ON STOCHASTIC GENERALIZED FUNCTIONS
17
Definition 9. Let Gα
e′(D) be the set of (fm) ∈ (C α(D,G))N0 such that
x kDβfm(x,·)kp) ∈ e′
(sup
0 with β ≤ α and p ∈ N0.
for all β ∈ Nd
We observe that Gα
0 ≤ γ ≤ α the partial differential operator Dγ transform Gα
Gα−γ
Definition 10. Let Gα
that
e′(D) is a subalgebra of (C α(D,G))N0 and for γ ∈ Nd
0,
e′(D) into
e (D) be the set of (fm) ∈ (C α(D,G))N0 such
x kDβfm(x,·)kp) ∈ e
(sup
(D).
e′
0 with β ≤ α and p ∈ N0.
e (D) is an ideal of Gα
for all β ∈ Nd
It is clear that Gα
Definition 11. We define the algebra of parametric generalized func-
tions Gα(D) as
e′(D).
Gα
e′(D)/Gα
e (D).
The elements of Gα(D) are called parametric generalized functions.
In the case that α = (0, .., 0) we denote Gα(D) by G(D). It is clear
that G ⊂ Gα(D) and C α
Example 5. Let D be a domain in Rd and α ∈ Nd
parametric generalized function is defined by
b (D) ⊂ Gα(D).
0. The white noise
Wx = (Wm(x)) = (I1(
m
Xj=0
ηj(x)ηj(·))),
for x ∈ D.
In fact, we observe that there exists constants Cα and aα such that for
all 0 ≤ β ≤ α,
kDβ
xWm(x)k2
p = kDβ
= kI1(
x ηj(x)ηj(·))k2
Dβ
0
0
m
x Wm(x)k2
Xj=0
Dβ
x ηj(x)ηj(·)2
0
=
m
Xj=0
Xj=0
Dβ
m
=
xηj(x)2
≤ Cα(m + 1)aα.
18
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
Definition 12. We says that u = [um] ∈ G(1,2,...,2)([0, T ] × Rl) is a
generalized solution of the Cauchy problem (19) if ut = Lu + uW (t, x)
in G([0, T ] × Rl) and u0 = f in G(2,..,2)(Rl).
We observe that [um] ∈ Gα(D) implies that [um(0,·)] is a well defined
element of G(2,..,2)(Rl).
Theorem 2. There exists an unique generalized solution u = [um] in
Gα(D) for the Cauchy problem (19).
Proof. Let d = 1 + l, α = (1, 2, ..., 2) ∈ N1+l
(T > 0). We consider that L have the form
and D = [0, T ] × Rl
0
Lf =
1
2
l
Xi,j=1
l
bi
∂
∂xi
f
∂2
aij
f +
∂xi∂xj
Xi=1
b (Rl) for all i and j.
where aij and bj belong to C 2
We split the proof in three steps:
(a) We solve the family of parabolic problems (m ∈ N0)
(20)
um = Lum + umWm(t, x)
d
dt
with u(x, 0) = f (x).
(b) We check that the nets of solutions (um) belongs to Gα
(c) We check that if (um) and (vm) are two nets of solutions of (19),
then [um] = [vm].
In order to prove (a), we use that the Feymann-Kac formula give a
representation of the solution of (20) (see for instance [11]). We observe
that there exist σ = (σij) ∈ C 2(Rl; Rl×l) such that
e′(D).
aij =
l
Xk=1
σikσjk.
We consider the following stochastic differential equation,
(21)
dX = σ(X) dB + b(X) dt,
where B is an l-dimensional Brownian motion in an auxiliar probability
space, σ = (σij) and b = (bi).
The solution of (21) with X(0) = x ∈ Rl is denoted by X(t, x). Ap-
plying the Feynman-Kac formula to um we have
(22)
um(t, x) = E(f (X(t, x)) eR t
0 Wm(t−s,X(s,x)) ds),
where E denotes the expectation in the auxiliar probability space.
It remains to prove b), this is (um) belongs to Gα
e′(D).
ON STOCHASTIC GENERALIZED FUNCTIONS
19
We claim that for each p ∈ N0 there exists constants C and a such that
d
dt
um(t, x)kp, sup
(t,x) kum(t, x)kp, sup
sup
(t,x) k
where 0 ≤ β ≤ (2, ..., 2).
In fact, by the definition of Wick exponential (7),
(23)
um(t, x) = E(f (X(t, x)) e
j=0 ηj ((t−s),X(s,x))ηj (·) ds2
2 R t
0 Pm
1
(t,x) kDβum(t, x)kp ≤ Ceam
0 e:R t
0 Wm(t−s,X(s,x)) ds).
From the uniform boundedness and orthonormality of the Hermite
functions, we have that there exist a positive constant a depending
of T such that
(24)
Z t
0
m
Xj=0
ηj((t − s), X(s, x))ηj(·)ds2
0 ≤ am.
Combining (23), (24) and (8), we obtain that sup(t,x) kum(t, x)k2
bounded by
p is
(sup
x∈Rl f (x))2eam.
dtum(t, x)kp ≤ Ceam, applying the Ito
In order to show that sup(t,x) k d
formula in (22), we obtain that
um(t, x) = E({f (X(t, x))(Wm(0, X(t, x)) + Z t
d
dt
0
+ Lfm(X(t, x))} eR t
0 Wm(t−s,X(s,x)) ds).
We need dominate the following terms:
d
dt
Wm(t − s, X(s, x)) ds)
0 Wm(t−s,X(s,x)) dsk2
p,
0 Wm(t−s,X(s,x)) dsk2
p.
a) A1 = kE(f (X(t, x))Wm(0, X(t, x))eR t
b) A2 = kE(f (X(t, x))R t
c) A3 = kE(Lf (X(t, x))eR t
dt Wm(t−s, X(s, x)) ds eR t
0 Wm(t−s,X(s,x)) dsk2
p.
0 Wm(t−s,X(s,x)) dsk2
p,
d
0
1
0 Pm
2 R t
a) We observe that A1 is equal to
0Wm(0, X(t, x))e:R t
kE(f (X(t, x))e
By the boundeness of f and the inequality (24) we have that
0 Wm(t−s,X(s,x)) dsk2
p).
A1 ≤ CeamE(kWm(0, X(t, x))e:R t
j=0 ηj ((t−s),X(s,x))ηj (·) ds2
From (9) there exists Cp > 0 and r, r′ ∈ N0 such that
0 Wm(t−s,X(s,x)) dsk2
kWm(0, X(t, x))e:R t
p
is lower or equal to
CpkWm(0, X(t, x))k2
rke:R t
0 Wm(t−s,X(s,x)) dsk2
r′ ≤ Ceam.
20
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
Thus A1 ≤ Ceam.
b) and c). The estimative for A2 is obtained in a similar way and the
estimative for A3 is analogous to the estimative for kumk2
p. We conclude
that
d
dt
k
um(t, x)kp ≤ Ceam.
It remains to prove that sup(t,x)∈K kDβum(t, x)kp ≤ Ceam for 0 ≤ β ≤
(2, ..., 2).
We first observe that Dium (i = 1, .., l) is equal to
l
Xj=1
E({
Dif (X(t, x))DiXj(t, x)+f (X(t, x))
l
Xj=1 Z t
0
DiWm(t−s, X(t, x))
×DiXj(s, x)) ds}eR t
0 Wm(t−s,X(s,x)) ds).
It is clear that DiXj satisfy the following system of linear stochastic
equations
DiXj = δij +Z t
0
l
(Dkσjh)(X(s, x))DiXk(s, x) dBh(s)+
Xk,h=1
(25)
+Z t
0
l
Xk=1
(Dkbj)(X(s, x))DiXk(s, x) ds.
From the equation (25) and Gronwall lemma we get that
E(
l
Xi=1
DiXj(t, x)2)
is locally uniformly bounded in t and x.
We need dominate the terms
d) B1 = kE(Pl
e) B2 = kE(f (X(t, x))(Pl
eR t
0 Wm(t−s,X(s,x)) dsk2
p.
j=1R t
d) We observe that B1 is equal to
j=1 Dif (X(t, x))DiXj(t, x)eR t
0 Wm(t−s,X(s,x)) ds)k2
p,
0 DiWm(t−s, X(t, x)))DiXj(s, x)) ds)
l
Xj=1
kE(
Dif (X(t, x))DiXj(t, x)e:R t
0 Wm(t−s,X(s,x)) dse
1
2 R t
0 Pm
j=0 ηj ((t−s),X(s,x))ηj (·) ds2
0)k2
p.
ON STOCHASTIC GENERALIZED FUNCTIONS
21
By the boundeness of the derivatives of f and the inequality (24) we
have that
B2 ≤ CeamE(
l
Xi=1
DiXj(t, x)2ke:R t
0 Wm(t−s,X(s,x)) dsk2
p) ≤ Ceam.
e) The estimative for B2 is obtained in an analogous way, using (9).
Therefore,
(t,x) kDium(t, x)kp ≤ Ceam.
sup
The estimative for kDiDjumkp is done in a similar way. So we conclude
that [um] ∈ Gα(D).
We considerer the uniqueness. Suppose that [um] and [vm] are two
generalized solutions of (19), we denote by dm to um − vm. By the
definition, we have that dm satisfy
(26)
(cid:26) d
dt dm
um(0, x) = N0,m(x)
= Ldm + dmWm(t, x) + Nm(t, x)
with Nm(t, x) ∈ Gα
Kac formula to dm we obtain that
e (D) and N0,m ∈ Gα
e (Rl). Applying the Feynman-
dm(t, x) = E(N0,m(X(t, x)) eR t
0 Wm(t−s,X(s,x)) ds
+Z t
0
Nm(s, x) eR s
0 Wm(s−u,X(u,x)) du ds).
It is straightforward using similar techniques that
(t,x) kdm(t, x)kp, sup
sup
(t,x) k
d
dt
dm(t, x)kp, sup
(t,x) kDβdm(t, x)kp ∈ e
for each p ∈ N0 and 0 ≤ β ≤ (2, ..., 2). We conclude that the solution
of (19) is unique.
(cid:3)
We end the paper with the following remarks.
Remark 2. Wick solution versus generalized solution: Let us consider
the following Cauchy problem,
(27)
(cid:26) vt = Lv + v ⋄ W (t, x)
v0 = f
where L is an uniformly elliptic partial differential operator, f ∈ C 2
and W is the white noise parametric generalized function.
b (Rl)
22
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
We observe that the proof of existence and uniqueness given in [24]
extends to the following Cauchy problem,
(28)
(cid:26) d
dt vm
vm(·, 0) = f
= Lvm + vm ⋄ Wm(t, x)
where W (t, x) = [Wm(t, x)] is the white noise parametric generalized
function.
The sequence of solutions (vm) converges in the Hida distribution space
to the solution of equation (27) (see [17]).
We have that the S-transformation of the solution of equation (28) is
S(vm(t, x))(h) = E(f (X(t, x)) eR t
0 hm(t−s,X(s,x)) ds)
and the S-transformation of the generalized solution u = [um] of (19)
is
S(um(t, x))(h) = E(f (X(t, x)) eR t
This implies that the generalized solution u = [um] is not associated to
any stochastic distribution for d > 1.
0 hm(t−s,X(s,x)) ds)e R t
0 Wm(. ,t−s,X(s,x)) ds0.
Remark 3. We can show in a similar way that there exists an unique
generalized solution of the stochastic Cauchy problem (19) with initial
data f ∈ G(2,..,2)(Rl).
References
[1] K. Aase , B. Oksendal, N. Privault and J. Uboe: A white noise generalization
of the Clark-Haussmann-Ocone theorem with application to mathematical
finance. Finance Stoch. 4, 465-496, 2000.
[2] S. Albeverio, Z. Haba and F. Russo: A two-space dimensional semilinear
heat equation perturbed by (Gaussian) white noise. Probab. Theory Related
Fields 121, 3, 319-366, 2001.
[3] S. Albeverio, Z. Haba and F. Russo: On non-linear two-space-dimensional
wave equation perturbed by space-time white noise. Israel Math. Conf. Proc.
10, 1-25, 1996.
[4] P. Antosik, J. Mikusi´nski and R. Sikorski: Theory of Distributions. The
sequential approach. Elsevier Scientific Publishing Company, 1973.
[5] F. Benth and J. Potthoff: On the martingale property for generalized sto-
chastic process. Stochastics Stochastics Rep. 58 , 349-367, 1996.
[6] F. Benth and T. Theting: Some regularity results for the stochastic pressure
equation of Wick-type. Stochastic Anal. Appl. 20, 6, 1191-1223, 2002.
[7] H. A. Biagioni: A nonlinear theory of generalized functions. Lect. Notes
Math. 1421. Springer-Verlag, Berlin, 1990.
[8] J. F. Colombeau: Elementary introduction to new generalized functions.
Noth-Holland Math Studies 113, Noth Holland, Amsterdam, 1985.
[9] P. Catuogno, S. Molina and C. Olivera: On Hermite representation of dis-
tributions and products. Integral Transforms and Special Functions 18, 4,
233-243, 2007.
ON STOCHASTIC GENERALIZED FUNCTIONS
23
[10] P. Chow: Generalized solution of some parabolic equations with a random
drift. Appl. Math. Optim, 20, 1, 1-17, 1989.
[11] M. Freidlin: Functional integration and partial differential equations. Prince-
ton University Press, 1985.
[12] M. Grosser , M. Kunzinger and M. Oberguggenberger and R. Steinbauer:
Geometric theory of generalized function with applications to general rela-
tivity. Kluwer Academic Publishers, Dordrecht, 2001.
[13] M. Grothaus, Y.G. Kondratiev and L. Streit: Regular generalized functions
in Gaussian analysis. Infinite Dimensional Anal. Quantum Prob. 2, 1, 1-25.
1999.
[14] T. Hida, H. Kuo, J. Potthoff and L. Streit: White noise. An infinite-
dimensional calculus. Kluwer, 1993.
[15] H. Holden, B. Oksendal, J. Uboe and T. Zhang: Stochastic partial differ-
ential equations. A modeling, white noise functional approach. Birkhuser
Boston, 1996.
[16] Z. Huang and J. Yan: Introduction to infinite dimensional stochastic analy-
sis. Kluwer, 1999.
[17] T. Lindstrom, B. Oksendal, J. Uboe and T. Zhang: Stability properties of
stochastic partial differential equations. Stochastic Anal. Appl. 13, 2, 177-
204, 1995.
[18] H. Kuo: White noise distribution theory. CRC, 1996.
[19] H. Kunita: Generalized solution of a stochastic partial differential equation.
J. Theoretical Probability 7, 2, 279-308, 1994.
[20] S. Lototsky and B. Rozovskii: Wiener chaos solutions of linear stochastic
evolution equations. Ann. Probab. 34, 2, 638-662, 2006.
[21] D. Nualart and M. Zakai: Generalized Brownian functionals and the solution
to a stochastic partial differential equation. J. Funct. Anal. 84, 2, 279-296,
1989.
[22] M. Oberguggenberger: Multiplication of distributions and applications to
partial differential equations. Pitman Research Notes in Math. Series 259.
Ed. Longman Science and Technology, 1993.
[23] M. Oberguggenberger and F. Russo: Nonlinear SPDEs: Colombeau solu-
tions and pathwise limits. Stochastic analysis and related topics, VI (Geilo,
1996), 319-332, Progr. Probab., 42, Birkhauser Boston, Boston, MA, 1998.
[24] J. Potthoff, G. Vage, and H. Watanabe: Generalized solutions of linear
parabolic stochastic partial differential equations. Appl. Math. Optim, 38, 1,
95-107, 1998.
[25] J. Potthoff and M. Timpel, On a dual pair of spaces of smooth and general-
ized random variables. Potential Anal. 4, 6, 637-654, 1995.
[26] F. Russo, Colombeau generalized functions and stochastic analysis, Edit. A.l.
Cardoso, M. de Faria, J Potthoff, R. Seneor, L. Streit, Stochastic analysis
and applications in physics , NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci.,
Kluwer Acad. Publ., Dordrecht, 449(1994) 329-249.
[27] L. Schwartz: Th´eorie des distributions. Hermann, Paris, 1966.
[28] B. Simon and M. Reed: Methods of modern mathematical physics. vol. 1.
Academic Press, 1980.
[29] M. Valdivia: Topics in locally convex spaces. North-Holland Mathematics
Studies, 67. Notas de Matemtica 85. North-Holland Publishing Co., 1982.
24
PEDRO CATUOGNO AND CHRISTIAN OLIVERA
[30] J. B. Walsh: An introduction to stochastic partial differential equations, In:
Ecole dEte de Probabilites de Saint Flour XIV, Lecture Notes in Mathemat-
ics 1180 (1986) 265-438.
Departamento de Matem´atica, Universidade Estadual de Campinas,,
13.081-970 - Campinas - SP, Brazil.
E-mail address: [email protected] ; [email protected]
|
1503.08912 | 3 | 1503 | 2015-09-05T15:14:21 | Modulus of supporting convexity and supporting smoothness | [
"math.FA"
] | We introduce the moduli of the supporting convexity and the supporting smoothness of a Banach space, which characterize the deviation of the unit sphere from an arbitrary supporting hyperplane. We show that the modulus of supporting smoothness, the Bana{\'s} modulus, and the modulus of smoothness are all equivalent at zero, the modulus of supporting convexity is equivalent at zero to the modulus of convexity. We prove a Day--Nordlander type result for these moduli. | math.FA | math | MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING
SMOOTHNESS
G.M. IVANOV∗
Abstract. We introduce the moduli of the supporting convexity and the supporting smooth-
ness of a Banach space, which characterize the deviation of the unit sphere from an arbitrary
supporting hyperplane. We show that the modulus of supporting smoothness, the Bana´s mod-
ulus, and the modulus of smoothness are all equivalent at zero, the modulus of supporting
convexity is equivalent at zero to the modulus of convexity. We prove a Day -- Nordlander type
result for these moduli.
5
1
0
2
1. Introduction
p
e
S
5
]
The properties of a Banach space are completely determined by its unit ball. The geometry
of the unit ball of a Banach space X may be described, for instance, using the properties of
some moduli attached to X. (For example, the moduli of convexity, of smoothness, Milman's
moduli, etc.) The aim of this paper is to introduce and explore some new type of moduli, which
characterize the deviation of the unit sphere from an arbitrary supporting hyperplane.
In the sequel we shall need some additional notation. Let X be a real Banach space. For
a set A ⊂ X by ∂A, int A we denote the boundary and the interior of A. We use hp, xi to
denote the value of a functional p ∈ X ∗ at a vector x ∈ X. For R > 0 and c ∈ X we denote by
BR(c) the closed ball with center c and radius R, by B∗
R(c) we denote the ball in the conjugate
space. By definition, put J1(x) = {p ∈ ∂B∗
1(o) : hp, xi = kxk}. For convenience, the length of
segment ab is denoted by kabk , i.e., kabk = ka − bk .
We say that y is quasiorthogonal to the vector x ∈ X \ {o} and write yqx if there exists a
functional p ∈ J1(x) such that hp, yi = 0. Note that the following conditions are equivalent:
-- y is quasiorthogonal to x
-- for any λ ∈ R the vector x + λy lies in the supporting hyperplane to the ball Bkxk(o) at x;
-- for any λ ∈ R the following inequality holds kx + λyk > kxk ;
-- x is orthogonal to y in the sense of Birkhoff -- James ([6], Ch. 2, §1).
.
A
F
h
t
a
m
[
3
v
2
1
9
8
0
.
3
0
5
1
:
v
i
X
r
a
Let
and
δX(ε) = inf (cid:26)1 − kx + yk
2
: x, y ∈ B1(0), kx − yk ≥ ε(cid:27)
ρX(τ ) = sup(cid:26)kx + yk
2
+ kx − yk
2
− 1 : kxk = 1,kyk = τ(cid:27) .
The functions δX(·) : [0, 2] → [0, 1] and ρX(·) : R+ → R+ are referred to as the moduli of
convexity and smoothness of X respectively.
∗Supported by the Russian Foundation for Basic Research, grant 13-01-00295.
1
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
2
Let f and g be two non-negative functions, each one defined on a segment [0, ε]. We shall
consider f and g as equivalent at zero, denoted by f (t) ≍ g(t) as t → 0, if there exist positive
constants a, b, c, d, e such that af (bt) 6 g(t) 6 cf (dt) for t ∈ [0, e].
The rest of this paper is organized as follows. In Section 2 we prove several simple technical
lemmas, in Section 3 we introduce the definitions of the modulus of supporting convexity
and the modulus of supporting smoothness and consider their basic properties, in Section 4
we show these modulus are equivalent at zero to the modulus of convexity and smoothness
respectively, in Section 5 we prove that the moduli of smoothness, of supporting smoothness
and the modulus of Bana´s are all equivalent at zero, and, finally, in Section 6 we prove some
estimates for these moduli concerning the maximal value of the Lipschitz constant for the metric
projection operator onto a hyperplane.
The author is grateful to professor G.E. Ivanov for constant attention to this work.
2. Technical results
In this section we prove several simple technical results.
The proof of the next lemma is trivial.
Lemma 1. Suppose the set B1(o) \ int Br(o1) is nonempty. Then it is arcwise connected.
Lemma 2. Let X2 be a two-dimensional Banach space. Suppose a, b, c, d ∈ ∂B1(o) and the
segments ab, cd intersect in point x. Then the following inequality holds
min{kcxk ,kxdk} 6 max{kaxk ,kxbk}.
Proof.
Assume the converse. Then for some ε > 0 we get min{kcxk ,kxdk} > max{kaxk ,kxbk}+ε = r.
Since the segment ab belongs to int Br(x) and separates it into two parts, then we cannot
connect points c, d in B1(o) \ int Br(x). This contradicts Lemma 1. The lemma is proved.
(cid:3)
Lemma 3. Let x, y ∈ X, x 6= o, p ∈ ∂B∗
(1)
1(o) such that hp, xi = kxk . Then
kx + yk 6 kxk + hp, yi + 2 kxk ρX(cid:18)kyk
kxk(cid:19) .
Proof.
By definition of the modulus of smoothness, we get
1
2 (cid:18)kx + yk
kxk
+ kx − yk
kxk (cid:19) − 1 6 ρX(cid:18)kyk
kxk(cid:19) .
Multiplying both sides by 2 kxk , after some transformations we obtain:
kxk(cid:19) 6
kx + yk 6 2 kxk − kx − yk + 2 kxk ρX(cid:18)kyk
2 kxk + hp, y − xi + 2 kxk ρX(cid:18)kyk
kxk(cid:19) = kxk + hp, yi + 2 kxk ρX(cid:18)kyk
kxk(cid:19) .
(cid:3)
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
3
Lemma 4. For any vectors x, y ∈ X \ {o} the following inequality is true
Proof.
Using the triangle inequality, we get
x
kxk −
y
kyk(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)
(cid:3)
= (cid:13)(cid:13)(cid:13)(cid:13)
(cid:18) x
kxk −
6
y
1
kxk(cid:19) +(cid:18) y
kxk −
kxk kx − yk + kyk(cid:12)(cid:12)(cid:12)(cid:12)
x
kxk −
(cid:13)(cid:13)(cid:13)(cid:13)
.
y
6
kxk
2 kx − yk
kyk(cid:13)(cid:13)(cid:13)(cid:13)
kyk(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
6 (cid:13)(cid:13)(cid:13)(cid:13)
(cid:18) x
kxk −
kyk(cid:12)(cid:12)(cid:12)(cid:12)
kxk −
6
1
1
y
y
kxk(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)(cid:13)
(cid:18) y
kxk −
y
kyk(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
6
2 kx − yk
.
kxk
3. Definitions and basic properties
Let x, y ∈ ∂B1(o) be such that yqx. By definition, put
λX(x, y, r) = min{λ ∈ R : kx + ry − λxk = 1}
for any r ∈ [0, 1]. Denote
λ−
X(x, y, r) = min{λX(x, y, r), λX(x,−y, r)}; λ+
X(x, y, r) = max{λX(x, y, r), λX(x,−y, r)}.
Definition 1. For any r ∈ [0, 1] and x ∈ ∂B1(o) we define the modulus of local supporting
convexity as
and respectively, the modulus of local supporting smoothness as
λ−
X (x, r) = inf λ−
X (x, y, t),
λ+
X(x, r) = sup λ+
X(x, y, t),
where we choose (y, t) such that kyk = 1, yqx, 0 6 t 6 r to minimize (maximize) λ−
(λ+
X (x, r)
X(x, r)).
It is clear that λ−
X(x, r) 6 λ+
X(x, r) 6 1.
Definition 2. For any r ∈ [0, 1] we define the modulus of supporting convexity as
X(x, t),
and respectively, the modulus of supporting smoothness as
X (x, t),
λ+
X(r) = sup λ+
λ−
X (r) = inf λ−
where we choose (x, t) such that x ∈ B1(o), 0 6 t 6 r to minimize (maximize) λ−
X(r)).
Let us explain the geometrical meaning of the moduli of supporting convexity and of sup-
porting smoothness. Fix y, x ∈ ∂B1(o) such that yqx. Consider the plane L = Lin{y, x}.
We use (a1, a2) to denote the vector a = a1y + a2x in this plane. The coordinate line
ℓ = {(a1, a2)a1 ∈ R, a2 = 0} is a tangent to the unit "circle" S = L ∩ ∂B1(x). By the
convexity of the ball, there is a convex function f : [−1, 1] → R such that for a1 ∈ [−1, 1] the
point (a1, f (a1)) belongs to the lower semicircle of S (see Fig. 1). Hence for a1 ∈ [−1, 1] the
functions λ−
X(a1) are the lower and upper bounds to the f (a1) respectively, i.e.
the following inequalities hold λ−
X(a1) and λ+
X(r) (λ+
X(a1) 6 f (a1) 6 λ+
X(a1) .
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
PSfrag replacements
4
S
(0, 1)
f (a1)
λ+
X (a1)
(0, 0)
(−1, 0)
λ−
X (a1)
ℓ
(1, 0)
Figure 1. Geometrical meaning of the λ+
X(r) , λ−
X(r) .
Lemma 5. Let X be an arbitrary Banach space, then:
X (0) = λ−
(i) λ+
(ii) for any r ∈ [0, 1] the following inequality holds: 0 6 λ−
(iii) for any 0 < r1 < r2 < 1 we have
X(0) = 0;
X(r) 6 λ+
X(r) 6 r;
(2)
(3)
r2
r1
λ−
X(r1) 6 λ−
λ−
X(r2) − λ−
X(r1) 6
X (r2) ,
r2 − r1
1 − r1
;
(iv) the modulus of supporting convexity is an increasing, continuous function on [0, 1) and
(v) the modulus of supporting smoothness is a strictly increasing, convex and continuous func-
moreover it is a strictly increasing function on the set {r ∈ [0, 1] : λ−
tion on [0, 1] and furthermore λ+
X (r) > 0};
X(1) = 1.
Proof.
Let us introduce some notation. Fix x, y ∈ ∂B1(o) such that yqx, and real numbers r1, r2 such
that 0 < r1 < r2 < 1. Let z = x + y, zi = x + riy where i = 1, 2. Let y1, y2 ∈ ∂B1(o) such
that yizi k ox and the intersection of the segment yizi and the ball B1(o) is the point yi where
i = 1, 2. (see Fig. 2). By construction kyizik = λX(x, y, ri) where i = 1, 2. The reader will
PSfrag replacements
y
y2
y1
z2
z1
z
o
x
Figure 2. Illustration for Lemma 5.
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
5
have no difficulty in showing that it is enough to prove all the assertions of this Lemma for
λX(x, y, r). Now, let us prove the Lemma.
X(0) = 0.
(1) By the definitions, we have λ+
(2) The first two inequalities of assertion (ii) are trivial.
X(0) = λ−
By similarity, we have
λX(x, y, r) 6 r. Indeed, y1z1 k zy and y1z1 ⊂ △xyz. Taking the supremum we get
assertion (ii).
(3) Taking into account that B1(o) is convex, we get y1z1 ⊂ xy2z2. By construc-
tion we have that y1z1 k z2y2. By the similarity, we get ky2z2k > r2
r1 ky1z1k , i.e.
r2
λX(x, y, r1) 6 λX(x, y, r2). Taking the infimum in λX(x, y, r2), we complete the proof
r1
of inequality (2).
By the convexity of the unit ball, we obtain that segment y2z2 lies in trapezoid y1z1zy.
By construction y2z2 k y1z1 k yz. By similarity, we get
r2 − r1
1 − r1
ky2z2k − ky1z1k 6 (1 − ky1z1k)
.
r2 − r1
1 − r1
X(r1) , we have ky2z2k − λ−
6
Taking the infimum in ky1z1k → λ−
(3).
X(r1) 6 r2−r1
1−r1
. This yields
(4) Assertion (iv) is the direct consequence of assertion (iii).
(5) The function λ+
X(·) is the supremum of the convex functions, therefore it's convex.
X(·) is a convex bounded function and λX (x, y, r) is continuous in r, we obtain
X(·) is continuous on [0, 1]. We will prove that λ+
X(r) > 0 on (0, 1] in Lemma 7
X (0) = 0 and convexity of the modulus of supporting
X(r) ≤ r
X(1) = 1 is the consequence of inequality
Since λ+
that λ+
below. By this and the equality λ+
smoothness, we get that it is a strictly increasing function. The inequality λ+
was proved in assertion (ii). The equality λ+
(11) at r = 1, which will be proved below.
(cid:3)
From Lemma 5 we have that in the definitions of the moduli of the supporting smoothness
and supporting convexity one may choose t = r.
Remark 1. Since any two plane central sections of the unit ball in a Hilbert space H are equal,
we have
λ+
H (r) = λ−
H (r) = δH (2r) = 1 − √1 − r2.
4. Comparison of supporting moduli with the moduli of convexity and
smoothness
Theorem 1. Let X be an arbitrary Banach space. Then λ−
r ∈ [0; 1] :
(4)
δX (r) 6 λ−
X(r) 6 δX(2r) .
X(ε) ≍ δX(ε) as ε → 0 and for any
Proof.
1) By the definition of the modulus of supporting convexity for any ε > 0 there ex-
ists a parallelogram xyzd such that x, z ∈ ∂B1(o), the point d lies in the segment xo
and kxyk = r, xyqox, kyzk 6 λ−
X(r) + ε. Therefore kodk = 1 − kyzk , consequently
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
6
X(r) + ε. Passing to the limit as ε → 0, we obtain the left-
δX(r) = δX(kzdk) 6 kyzk 6 λ−
hand side of chain (4).
2) Let us prove the right-hand side of chain (4).
Fix r ∈ (0, 1) (if r = 0 or r = 1 the inequality is trivial). By the definition of the modulus
of supporting convexity for any ε > 0 there exist points aε, bε on the unit sphere such that
kaεbεk > 2r and for the point cε = aε+bε
(5)
the following inequality holds:
2
1 − kocεk 6 δX (2r) + ε.
Let the ray ocε intersect the unit sphere in a point x. Denote by l1 the supporting line to the
unit sphere such that l1 lies in the plane oaεbε and x ∈ l1. Let l2 be a line such that l1 k l2 and
cε ∈ l2. Denote by f, g the points of intersections of ∂B1(o) and the line l2. From Lemma 2 it
follows that kf − cεk > r or kg − cεk > r. Without loss of generality, put kg − cεk > r. Let l3
be a line such that l3 k ocε and g ∈ lε. By definition, we put y = l3 ∩ l1 (see Fig. 3). Then
PSfrag replacements
l2
l1
bε
f
o
cε
x
l3
g
aε
y
Figure 3. Illustration for Lemma 1.
δX(2r) + ε > kcεxk > λ−
X(cid:18)x,
y − x
ky − xk
,ky − xk(cid:19) > λ−
X(cid:18)x,
y − x
ky − xk
, r(cid:19) > λ−
X(r) ,
i.e., δX(2r) + ε > λ−
(cid:3)
X (r) . Passing to the limit as ε → 0, we complete the proof.
Lemma 6. Let r ∈ [0, 1
(6)
2 ]. Then
λ+
X(r) 6 ρX(2r) .
X (r) . Since λ+
Proof.
Denote λ = λ+
2 . Let µ ∈ (0, λ). By the
Definitions 1, 2 there exist x, y ∈ ∂B1(o) such that yqx and λX (x, y, r) = µ′ ∈ (µ, λ), and
consequently kx + ry − µ′xk = 1. Since yqx there exists p ∈ J1(x) = J1(x − µ′x) such that
hp, yi = 0.
X(r) 6 r for any r ∈ [0, 1], then λ 6 1
Using Lemma 3, we get
1 = kx + ry − µ′xk 6 kx − µ′xk + hp, ryi + 2(1 − µ′)ρX(cid:18) r
1 − µ′(cid:19) =
= 1 − µ′ + 2(1 − µ′)ρX(cid:18) r
1 − µ′(cid:19) .
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
7
To complete the proof, it suffices to note that µ′ < 1
is a convex function.
(cid:3)
2 , ρX(0) = 0 and the modulus of smoothness
Lemma 7. Let r ∈ [0, 1]. Then
(7)
ρX(cid:16) r
2(cid:17) 6 λ+
X(r) .
Proof.
Taking into account the definition of the modulus of smoothness,
it follows that for any
2(cid:3) and ε ∈ [0, ρX(τ )) there exist x and y such that the following inequality is true
τ ∈ (cid:2)0, 1
(8)
Without loss of generality, we can assume that kx + τ yk > kx − τ yk (hence kx + τ yk > 1).
Denote u = x+τ y
kx + τ yk + kx − τ yk − 2 > 2(ρX(τ ) − ε).
kx−τ yk . By Lemma 4, we obtain
kx+τ yk , v = x−τ y
(9)
By the triangle inequality, we get
ku − vk 6
4τ
kx + τ yk
;
ku + vk 6
2 kxk
kx + τ yk
1
+ kx − τ yk(cid:12)(cid:12)(cid:12)(cid:12)
1
kx + τ yk −
1
kx − τ yk(cid:12)(cid:12)(cid:12)(cid:12)
(kx + τ yk + kx − τ yk − 2).
= 2 −
kx + τ yk
Now, by inequality (8), we have that
(10)
ku + vk 6 2 −
2(ρX(τ ) − ε)
kx + τ yk
.
Let us consider the plane ouv. By ω denote a point lying on the smallest arc uv of
the unit circle such that the supporting line to the unit ball at ω is parallel to uv. Ob-
, i.e.
ku−vk , ku−vk
. Combining this with inequalities (10), we get
2 (cid:17) > 1 − ku+vk
or λX (cid:16)ω,− u−v
ku−vk , ku−vk
2
2 (cid:17) > 1 − ku+vk
2
viously, either λX (cid:16)ω, u−v
X(cid:16) ku−vk
2 (cid:17) > 1 − ku+vk
λ+
2
2(ρX(τ ) − ε)
kx + τ yk
6 2λ+
X(cid:18)ku − vk
2
(cid:19) .
Now, by inequality (9), we obtain
2
kx + τ yk
X(cid:18)
(ρX(τ ) − ε) 6 2λ+
2τ
kx + τ yk(cid:19) 6
2
kx + τ yk
λ+
X (2τ ) .
Multiplying both sides by kx+τ yk
(cid:3)
2
and passing to the limit as ε → 0, we obtain (7).
Remark 2. By Lemma 7 and the properties of the modulus of smoothness, it follows that
λ+
X(r) > 0 for all r > 0.
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
8
By Lemmas 6, 7 and the properties of the modulus of smoothness we have the following
result.
Theorem 2. Let X be an arbitrary Banach space. Then λ+
r ∈ [0, 1
2]:
X(τ ) ≍ ρX(τ ) as τ → 0 and for any
ρX(cid:16) r
2(cid:17) 6 λ+
X (r) 6 ρX(2r) .
5. Comparison with the Bana´s modulus
In the paper [1] J. Bana´s defined and studied some new modulus of smoothness. Namely, he
defined
2
δ+
ε ∈ [0, 2].
: x, y ∈ B1(o), kx − yk 6 ε(cid:27) ,
X (ε) = sup(cid:26)1 − kx + yk
The function δ+
X(·) is called the Bana´s modulus. In the papers [1, 2, 3, 4] several interesting
results concerning this modulus were obtained. Particulary, in [1], J. Bana´s proved that a space
is uniformly smooth iff δ+
ε → 0 as ε → 0. However, from the definition a space is uniformly
smooth if and only if ρX(ε)
ε → 0 as ε → 0. This leads to the question: are the modulus of
smoothness and the modulus of Bana´s equivalent at zero? It is easy to check that there exist
positive constant a, b such that δ+
X(t) 6 aρX(bt) , but the lower estimate of the modulus of
Bana´s in terms of the modulus of smoothness is unknown. In the next theorem we prove that
the modulus of Bana´s and the modulus of supporting smoothness are equivalent at zero, so
Theorem 2 answers the above question.
X(ε)
Theorem 3. Let X be an arbitrary Banach space. Then δ+
following inequalities hold:
X(ε) ≍ δX (ε) as ε → 0 and the
(11)
(12)
X(r)
δ+
X (2r) ≤ λ+
X(r) ≤ 2δ+
λ+
X (3r)
∀r ∈ [0, 1] ;
∀r ∈ (cid:20)0,
3(cid:21) .
2
Proof.
1) First we shall prove inequality (11) for r ∈ [0, 1).
Let a, b be points of the unit sphere such that ka − bk 6 2r. By X2 denote the plane aob.
There exists a point y2 of the unit sphere of the plane X2 such that the supporting line l2 to
the unit ball at this point is parallel to ab. By definition, put y1 = oy2∩ ab. There exists a point
a2 in the projection of the point a on l2 such that the segments y1y2, aa2 are equal in length and
parallel. The point b2 is defined in the same way, such that y1y2 and bb2 are parallel (see Fig.
6). Without loss of generality we assume that ky2a2k 6 r < 1. Since the modulus of supporting
smoothness is an increasing function, we have ky1y2k = kaa2k 6 λ+
X(y2, r) .
Taking the supremum, we obtain inequality (11).
X(y2,ky2a2k) 6 λ+
Taking into account that the modulus of Bana´s is a continuous and increasing function, we
obtain inequality (11) for r = 1.
2) Let us prove inequality (12).
By the definition of modulus of supporting smoothness for any ε ∈ (0, λ+
-- a point x ∈ ∂B1(o);
-- a line ℓ1 supporting to the unit ball at point x;
X(r)) there exist
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
9
-- a point y on ℓ1 and a point z ∈ ∂B1(o) such that
kxyk = r, kyzk > 0, zy k ox and λ+(cid:16)x, xy
Let ℓ2 be a line parallel to ℓ1 such that z ∈ ℓ2. Let z, z1 be points of the intersections of line
ℓ2 and ∂B1(o). By y1 denote the projections of z1 on ℓ1 such that z1y1 k ox (see Fig. 4).
kxyk , r(cid:17) = kyzk > λ+
X(r) − ε > 0.
PSfrag replacements
z
y
o
e
f
x
z1
d
ℓ2
ℓ1
y1
Figure 4. Illustration for the second part of Theorem 3.
xy
kxyk
We shall prove that kzz1k > 2r. In the converse case, kxy1k < r. Note that if we fix
x, y ∈ ∂B1(o) such that yqx, then the function λ+(x, y,·) is strictly increasing on the set of its
positive values. Since xy1 and xy lie on the same line and by to the definition of λ+, we obtain
λ+(cid:18)x,
, r(cid:19) = kyzk = ky1z1k 6 λ+(cid:18)x,
, r(cid:19).
Contradiction. Consequently kzz1k > 2r.
By definition, put e = ox ∩ zz1. By the continuity reasons there exists a point d on the
arc z1x of the unit sphere such that for the point f = zd ∩ ox the following equality holds
kd − fk = kf − zk . Since ℓ1 is a supporting line to the unit sphere, we have kxfk > kyzk
2 . Note
that kdzk 6 2(kzek + kefk) 6 3r. Combining the last two inequalities, we get
,kxy1k(cid:19) < λ+(cid:18)x,
, r(cid:19) = λ+(cid:18)x,
xy1
kxy1k
xy1
kxy1k
xy
kxyk
2
Passing to the limit as ε → 0, we obtain inequality (12).
δ+
X (3r) > kxfk >
(cid:3)
λ+
X(r) − ε
.
From Theorems 2 and 3 we have the following corollary.
Corollary 1. Let X be an arbitrary Banach space, then δ+
following inequalities hold:
X(ε) ≍ ρX(ε) as ε → 0 and the
1
2
ρX(cid:16) r
6(cid:17) 6 δ+
X(r) 6 ρX(r) ,
r ∈ (cid:20)0,
1
2(cid:21) .
The Day-Nordlander theorem (see [7]) asserts that δX(ε) 6 δH (ε) for ε ∈ [0, 2], where H
denotes an arbitrary Hilbert space. On the other hand, repeating the arguments from the paper
[7] we can show that for any Banach space the following estimate is true δ+
X(r) for
H(ε) 6 δ+
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
10
ε ∈ [0, 2]. From this and Theorems 2, 3 we obtain a Day -- Nordlander type result for the moduli
of supporting convexity and supporting smoothness:
Corollary 2. Let X be an arbitrary Banach space. Then
λ−
X(r) 6 λ−
H (r) = 1 − √1 − r2 = λ+
H(r) 6 λ+
X (r)
∀r ∈ [0, 1].
If at least one of these inequalities turns into equality, then X is a Hilbert space.
6. Estimates for Lipschitz constant for the metric projection onto a
hyperplane
Let us introduce the following characteristic of a space:
ξX = sup
kxk=1,
kyk=1
p∈J1(y)kx − hp, xiyk.
sup
Note that if y ∈ ∂B1(0), p ∈ J1(y), then the vector (x − hp, xiy) is a metric projection of
x onto the hyperplane Hp = {x ∈ X : hp, xi = 0}. So, ξX = supy∈B1(o) supp∈J1(x) ξp
X, where
ξp
X is half of diameter of a unit ball's projection onto the hyperplane Hp. Therefore, ξX is the
maximal value of the Lipschitz constant for the metric projection operator onto a hyperplane.
Obviously, ξX ≤ 2 and ξH = 1 for a Hilbert space H.
Theorem 4. For any Banach space X the following inequality is true:
(13)
1
1 − λ−
X(cid:16) 1−λ−
X(1)
2
6 ξX 6
(cid:17)
1
1 − λ+
X(cid:16) 1−λ−
X(1)
2
.
(cid:17)
Proof.
First let us introduce some notation. Let x0 be an arbitrary point on the unit sphere. Let l be
a supporting line to the unit ball at the point x0. Define l2 as the line such that the following
conditions hold:
a) l2 k ox0;
b) l2 ∩ l 6= ∅, by definition, put x2 = l2 ∩ l;
c) l a is supporting line to the unit ball at some point y2;
d) ky2x2k 6 1.
Let x1 be a point on segment x0x2 such that kx0x1k = 1, let l1 be a line such that x1 ∈ l1 and
l1 k ox0. By definition, put y1 as the intersection point of line l1 and the segment oy2. Let b be
a point on ∂B1(o) such that the segment ob is parallel to x0x1. By construction, we have that
x0x1bo is a parallelogram, therefore b ∈ l1 and y1 ∈ x1b. Let a be the intersection point of the
line l1 and the unit sphere such that a ∈ x1y1.
From the intercept theorem, we have kx0x2k
koy2k = kx0x1k
koy1k . Therefore
(14)
kx0x2k =
1
=
1
.
koy1k
1 − ky1y2k
Since x0x1bo is a parallelogram, we get kx1bk = kox0k = 1. By construction we have that
kx0x1k = 1. Therefore,
(15)
kabk 6 1 − λ−
X (1) .
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
11
Define a2 as the projection of the point a on l2 such that aa2 k oy2. In the same way we define
the point b2. Then the segments y1y2, aa2 и bb2 are parallel and equal in length (as parallel
segments between two parallel lines). By the definition of the modulus of supporting convexity
and by inequality (15), we obtain
(16)
ky1y2k 6 λ+
X(min{ka2y2k ,ky2b2k}) 6 λ+
X(cid:18)kabk
2 (cid:19) 6 λ+
X(cid:18) 1 − λ−
X (1)
2
(cid:19) .
Combining this and equality (14), we finally prove the right-hand side of inequality (13).
Let ε be an arbitrary positive real number. Note that we could choose a point x0 such that
kx1ak 6 λ−
X(1) + ε, i.e. kabk > 1 − λ−
ky1y2k > λ−
X (max{ka2y2k ,ky2b2k}) > λ−
X(1) − ε. Like in (16), we obtain
X(cid:18)kabk
2 (cid:19) > λ−
X(cid:18)1 − λ−
X(1) − ε
2
(cid:19) .
Passing to limit as ε → 0 and using inequality (14), we prove the left-hand side of inequality
(13).
(cid:3)
PSfrag replacements
Remark 3. The estimate (13) is reached in case of a Hilbert space. The right-hand side of
inequality (13) does not exceed 2, i.e. this estimate is not trivial.
Conjecture 1. The right-hand side of
Lp, p ∈ (1; +∞).
inequality (13) becomes an equality in case of
l2
l1
b
b2
y2
a2
x2
y1
a
x1
o
x0
l
Figure 5. Illustration for Theorem 4.
In the following lemma we obtain a lower estimate of the modulus of supporting smoothness
by the inverse function to the modulus of convexity.
Lemma 8. For any r ∈ [0, 1] the following inequalities hold:
r
(17)
δ−1
δ−1
r
1
2
1 −
X (cid:16)1 −
2(cid:17) 6 1 −
1
2
X (cid:18)1 −
ξX(cid:19) 6 λ+
X (r) .
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
12
Proof.
The left-hand side of inequality (17) is a straightforward consequence of the inequality ξX ≤ 2.
Let us prove the right-hand side of inequality (17). In case of r = 0 it is trivial. Let x0 be an
arbitrary point on the unit sphere. Define Hx as a supporting hyperplane to the unit ball at
the point x0. Let x1 be a point of the supporting hyperplane Hx such that kx0x1k = r. Denote
the ray {ox0 + αx0x1 : α > 0} as ℓ. Let l1, l2 be the lines parallel to ox0 such that
a) l2 is a supporting line to the unit ball at the point y2 and l2 ∩ ℓ = x2;
b) l1 intersects the ray ℓ at x1 and intersects the unit sphere at points a, b.
Let y1 = oy2 ∩ ab (see Fig. 6). By the definition of λ+
l2
X(r) and since the unit ball is centrally
PSfrag replacements
l1
b2
b
y2
y1
a2
ℓ
x2
a
x1
o
x0
Figure 6. Illustration for Lemma 8 and for the first part of Theorem 3.
symmetric, we get kabk > 2(1 − λ+
(18)
X(r)). Obviously, ky1y2k > δX (kabk) . Consequently, we have
Using the intercept theorem, we obtain
(19)
ky1y2k = ky1y2k
koy2k
By inequalities (18) and (19), we have
δX(cid:0)2(1 − λ+
= kx1x2k
kx0x2k
X(r))(cid:1) 6 δX (kabk) 6 ky1y2k .
= kx0x2k − kx0x1k
= 1 −
kx0x2k
r
kx0x2k
6 1 −
r
ξX
.
(cid:3)
δX(cid:0)2(1 − λ+
X (r))(cid:1) 6 1 −
r
ξX
It is easy to check that in a Hilbert space H the following equality holds
Substituting this in inequality (17) and since ξH = 1, we obtain
δ−1
H (τ ) = 2p1 − (1 − τ )2.
δH(2r) = 1 −
1
2
δ−1
H (1 − r) 6 λ+
H(r).
According to (1), we have that if X is a Hilbert space, then the right hand estimate in
inequality (17) is reached.
MODULUS OF SUPPORTING CONVEXITY AND SUPPORTING SMOOTHNESS
13
References
[1] Bana´s J. On moduli of smoothness of Banach spaces // Bull. Pol. Acad. Sci., Math. 1986. Vol. 34. P. 287 --
293.
[2] Bana´s J., Fraczek K. Deformation of Banach spaces // Comment. Math. Univ. Carolinae. 1993. Vol. 34.
P. 47 -- 53.
[3] Bana´s J., Hajnosz A., Wedrychowicz S. On convexity and smoothness of Banach space // Commentationes
Mathematicae Universitatis Carolinae. 1990. Vol. 31, no. 3. P. 445 -- 452.
[4] Bana´s J., Rzepka B. Functions related to convexity and smoothness of normed spaces // Rendiconti del
Circolo Matematico di Palermo. 1997. Vol. 46, no. 3. P. 395 -- 424.
[5] Baronti M., Papini P. Convexity, smoothness and moduli // Nonlinear Analysis: Theory, Methods &
Applications. 2009. Vol. 70, no. 6. P. 2457 -- 2465.
[6] Diestel J. Geometry of Banach Spaces - Selected Topics. Springer-Verlag Berlin Heidelberg, 1975. Vol. 485.
[7] Nordlander G. The modulus of convexity in normed linear spaces // Arkiv for Matematik. 1960. Vol. 4,
no 1. P. 15 -- 17.
Department of Higher Mathematics, Moscow Institute of Physics and Technology, Insti-
tutskii pereulok 9, Dolgoprudny, Moscow region, 141700, Russia
National Research University Higher School of Economics, School of Applied Mathematics
and Information Science, Bolshoi Trekhsvyatitelskiy 3, Moscow, 109028, Russia
E-mail address: [email protected]
|
1904.11129 | 1 | 1904 | 2019-04-25T02:34:09 | Committee spaces and the random column-row property | [
"math.FA"
] | A committee space is a Hilbert space of power series, perhaps in several or noncommuting variables, such that $\|z^\alpha\|\|z^\beta\| \geq \|z^{\alpha+\beta}\|.$ Such a space satisfies the true column-row property when ever the map transposing a column multiplier to a row multiplier is contractive. We describe a model for random multipliers and show that such random multipliers satisfy the true column-row property. We also show that the column-row property holds asymptotically for large random Nevanlinna-Pick interpolation problems where the nodes are chosen identically and independently. These results suggest that finding a violation of the true column-row property for the Drury-Arveson space via na\"ive random search is unlikely. | math.FA | math | COMMITTEE SPACES AND THE RANDOM
COLUMN-ROW PROPERTY
J. E. PASCOE
βk ≥
αkkz
Abstract. A committee space is a Hilbert space of power series,
perhaps in several or noncommuting variables, such that kz
α+βk. Such a space satisfies the true column-row property when
kz
ever the map transposing a column multiplier to a row multiplier is
contractive. We describe a model for random multipliers and show
that such random multipliers satisfy the true column-row property.
We also show that the column-row property holds asymptotically
for large random Nevanlinna-Pick interpolation problems where
the nodes are chosen identically and independently. These results
suggest that finding a violation of the true column-row property
for the Drury-Arveson space via naıve random search is unlikely.
9
1
0
2
r
p
A
5
2
]
.
A
F
h
t
a
m
[
1
v
9
2
1
1
1
.
4
0
9
1
:
v
i
X
r
a
1. Introduction
Let H be a Hilbert space of formal power series such that polynomi-
als are dense. (Either commutative or not, and perhaps in several or
infinitely many variables.) We say H is a committee space whenever
(1) hzα, zβi = 0 if α 6= β,
(2) Committee inequality: kzαkkzβk ≥ kzα+βk.
One can verify that the following spaces are committee spaces:
• Hardy space. The space of power series in one variable such
that each monomial has norm one.
• The Drury-Arveson space. The Hilbert space of commuting
power series in d variables such that
hzα, zαi =
where (cid:0)α
(α1, . . . αd),
α(cid:1) is the multi-nomial coefficient. That is,
if α =
1
,
(cid:0)α
α(cid:1)
α (cid:19) =
(cid:18)α
(α1 + . . . + αd)!
α1! . . . αd!
,
Date: April 26, 2019.
2010 Mathematics Subject Classification. 46E22, 47B32.
1
2
J. E. PASCOE
which is number of ways to divide α people into committees
of size αi. The committee inequality follows from the fact that
α + β (cid:19) ≥ (cid:18)α
(cid:18)α + β
β (cid:19)
α (cid:19)(cid:18)β
which in turn corresponds to the fact there are less ways to
divide people into committees when constraints are placed on
the composition of those committees.
• Fock space. The space of noncommutative power series such
that each monomial has norm one.
• Dirichlet space. The space of power series in one variable such
that the hzn, zni = n + 1.
We denote the monomial shifts on H by Mzγ . Note that Mzγ M ∗
zγ and
zγ Mzγ are diagonal and with respect to the basis zα, and therefore
is bounded by the committee inequality. The
M ∗
kMzγ k = supα
multipliers Mf of H are denoted as Mult H.
kzγ+αk
kzαk
fik ≤
fik ≤ CkPi M ∗
We say H satisfies the true column-row property if kPi MfiM ∗
kPi M ∗
kPi MfiM ∗
fiMfik for any sequence of multipliers. We say H satisfies
the column-row property if there is a constant C > 0 such that
fiMfik for any sequence of multipliers. The
column-row property is important in interpolation theory [8, 2, 3, 7].
So far as the author knows, there are no known commutative com-
plete Nevanlinna-Pick spaces for which the true column-row property
fails, and, in general, complete Nevanlinna-Pick spaces are committee
spaces. The column row property fails for the Fock space in two or
more variables [4].
The goal of this manuscript is to understand when a random sequence
of multipliers Mf1, Mf2, . . . satifies the column-row property. That is,
when is
k lim
n→∞
1
n
n
Xi=1
MfiM ∗
fik ≤ k lim
n→∞
1
n
n
M ∗
fiMfik.
Xi=1
Here a normalization is taken to guarantee convergence. We note a
slight subtlety here, the limits must be evaluated in the weak operator
topology as opposed to in norm. If our method of sampling secretly
sampled from a space with finite dimension, all limits would reduce to
bona fide norm limits.
2. A model for sampling random multipliers
Let vγ be a sequence of random variables indexed by multi-indices γ
such that
(1) E(vγ) = 0,
THE RANDOM COLUMN-ROW PROPERTY
3
(2) E(vγ2) = wγ < ∞,
(3) E(vγvγ ′) = 0 if γ 6= γ′,
(4) There is constant C > 0 such that the function fv = P vγzγ
satisfies kMfv k ≤ C almost surely.
We call such a sequence vγ a random multiplier model. We will
compute
Rv = E(MfvM ∗
fv ), Cv = E(M ∗
fv Mfv).
Note that, in the weak operator topology, almost surely,
Rv = lim
n→∞
Cv = lim
n→∞
1
n
1
n
n
Xi=1
Xi=1
n
MfiM ∗
fi,
M ∗
fiMfi
where fi is a sequence of functions sampled from the random multiplier
model.
3. The row column norm holds for random multipliers
We will now prove the following theorem.
Theorem 3.1. Let H be a committee space. Let vγ be a random mul-
tiplier model. Then,
kRvk ≤ kCvk.
Proof.
Lemma 3.2. Let α, γ be multi-indices. If there is a β such that α =
γ + β, then,
kM ∗
zγ zαk =
kzαk2
kzβk
.
Otherwise,
kM ∗
zγ zαk = 0.
Proof. Note that if
hM ∗
zγ zα, zβi = hzα, Mzγ zβi = hzα, zγ+βi
is to be non-zero, then α = γ + β. Moreover, β is unique. Now,
kM ∗
zγ zαk = hM ∗
zγ zα,
zβ
kzβk
i
= hzα,
zα
kzβk
i
=
kzαk2
kzβk
.
4
J. E. PASCOE
(cid:3)
Lemma 3.3.
Proof. Note,
kRvk ≤ sup
α Xγ+β=α
kzγk2wγ.
Rv = E(MfvM ∗
fv )
vγvγ ′Mzγ M ∗
zγ′ )
= E(Xγ,γ ′
= Xγ
wγMzγ M ∗
zγ .
So, Rv must be a diagonal operator with respect to the monomial basis.
Therefore, we may compute the norm as follows,
kRvk = sup
α
hRvzα, zαi
kzαk2
zγ zα, zαi
= sup
α
= sup
= sup
kzαk2
kzαk4wγ
hPγ wγMzγ M ∗
α Xγ+β=α
α Xγ+β=α
kzβk2
kzβ+γk2wγ
kzβk2kzαk2
≤ Xγ+β=α
kzγk2wγ,
where the last line holds by the committee inequality.
(cid:3)
Lemma 3.4.
Proof. Note,
kCvk = Xγ
kzγk2wγ.
Cv = E(M ∗
= E(Xγ,γ ′
= Xγ
fvMfv )
vγvγ ′M ∗
zγ Mzγ′ )
wγM ∗
zγ Mzγ .
THE RANDOM COLUMN-ROW PROPERTY
5
So, Cv must be a diagonal operator with respect to the monomial basis.
Therefore, we may compute the norm as follows,
kCvk = sup
α
hCvzα, zαi
kzαk2
zγ Mzγ zα, zαi
kzαk2
kzγ+αk2wγ
hPγ wγM ∗
α Xγ
kzγk2wγ,
kzαk2
= sup
α
= sup
= Xγ
where the last equality follows from the committee inequality, as that
implies the maximum is attained when α is the trivial multi-index. (cid:3)
Now, we see that
kRvk ≤ sup
α Xγ+β=α
kzγkwγ ≤ Xγ
kzγkwγ ≤ kCvk,
which proves Theorem 3.1.
(cid:3)
Note that, formally, we could have assumed there was constant C >
0 such that the function fv = P vγzγ satisfies kfvkH ≤ C almost
surely, and, algebraically, everything would have still worked. This is
reminiscent of the theorem of Cochran-Shapiro-Ullrich [6], that, given
a function in the Dirichlet space, randomly multiplying each coefficient
by ±1 gives a multiplier of the Dirichlet space.
3.1. The truncated shift case. We note that the above calculations
hold true, with mild but insightful cosmetic differences, for the restric-
tion of the shifts to monomials of degree less than n. We see that, in
fact,
kR(n)
v k ≤ max
α≤n Xγ+β=α
kzγkwγ ≤ Xγ≤n
kzγkwγ ≤ kC (n)
v k.
When we are working in more than one variable, one can see that, given
a multi-index α of degree less than or equal to n, the number of γ such
that γ + β = α is very small compared to the set of all multi-indices
γ of degree less than n. Thus, heuristically, one expects that kRvk will
be much smaller than kCvk, although detailed estimates will depend
intricately on the parameters wγ. We interpret this as explanation of
the fact that numerical experiments to test the column row property
have not produced counterexamples.
6
J. E. PASCOE
4. The random basis lemma
We say a H-valued random variable h is a random vector if P (h ⊥
g) < 1 for all g ∈ H.
The goal of this section is to prove the following lemma, which we
will need for technical reasons later. The content is essentially that
an infinite sequence of random vectors is a (perhaps overdetermined)
basis.
Lemma 4.1. If h1, h2, . . . is a sequence independent identically dis-
tributed of random vectors, then, almost surely, the closed span of the
hi is equal to H.
Proof. First we will need a lemma.
Lemma 4.2. Let h be a random vector. There is a countable subset A
of H such that the closed span of the elements of A is equal to H and
for every point a ∈ A, P (h ∈ U) > 0 for any neighborhood U of a.
Proof. For any subset A such that for every point a ∈ A, P (h ∈ U) > 0
for any neighborhood U of a, and the closed span of the elements of A
is not equal to H, we will show that we can grow A by a single element
which not in closed span of the elements of A. We can only do this
a countable number of times because the Hilbert space dimension of
H is countable. (Otherwise, via Gram-Schmidt, we could construct an
uncountable orthonormal set by transfinite induction.)
Choose g such that g ⊥ a for all a ∈ A. Now, P (h ⊥ g) < 1. So there
must be a point b such that P (h ∈ U) > 0 for every neighborhood of
b and b is not perpendicular to g, therefore, b is not in the span of the
elements of A.
(cid:3)
Suppose h1, h2, . . . is a sequence independent identically distributed
of random vectors. Let A be as in Lemma 1. Index A a a sequence
an. Let Bm,n be a ball of radius 1/m centered at an Almost surely,
the sequence hi must visit Bm,n infinitely often, as P (hi ∈ Bm,n) > 0.
Therefore A is a subset of the closure of the values of the sequence.
(We have essentially the fact that a random function f : N → N2 is
surjective with infinite multiplicity.)
(cid:3)
5. Sampling random Nevanlinna-Pick problems
Given a committee space H, the natural domain of the multiplier
algebra of H is the set of all tuples of matrices (x1, x2, . . .) such that the
map taking Mzi 7→ xi is a completely contractive homomorphism. In
the commutative case, where H might also be interpreted as a space of
complex analytic functions, the most important and familiar points are
THE RANDOM COLUMN-ROW PROPERTY
7
the scalars, which are, in principle, the domain where it makes sense
to evaluate all functions in Mult H.
Given X = (x(1), . . . , x(m)) a sequence of points in the natural do-
main of the space of multipliers, and compatiable target values y(k)
ij ,
the Nevanlinna-Pick problem asks when there are fij such that
fij(x(k)) = y(k)
ij and the block multiplier [Mfij ]i,j has norm less than or
equal to 1. Let PX be the projection onto {hh(x(k)) = 0 for all k}⊥.
Note that, as {hh(x(k)) = 0 for all k} is an invariant subspace for the
shifts, its orthogonal complement is invariant for the adjoints, that is,
M ∗
f PX = PX M ∗
f PX, PXMf = PXMf PX.
ij
A neccessary condition for the Nevanlinna-Pick problem to be solvable
is for there to be functions gij(x(k)) = y(k)
such that [PX Mgij PX]i,j has
norm less than 1. (In fact, the norm of this block (operator) matrix is
independent of the choice of gij, when they exist.) Moreover, whenever
we are working in a complete Nevanlinna-Pick space the condition is
also sufficient. See [1] and [5] for a theory of commutative and noncom-
mutative complete Nevanlinna-Pick spaces respectively. One way to
test for the potential failure of the column-row property in a complete
Nevanlinna-Pick space would be to choose a random Nevanlinna-Pick
problem and then show that when we interpret the target data as a row,
the problem is not solvable, but when the target data is interpreted as
a column it is. Define,
RX
v = E(PXMfv PXM ∗
v k and kC X
fvPX ), C X
v = E(PXM ∗
fv PXMfvPX ).
We interpret kRX
a Nevanlinna-Pick probem such that yk
random multipliers when H is a complete Nevanlinna-Pick space.
v k as the minimum norm of a solutions to
i = fi(x(k)) where the fi are
Let x be a random variable taking values in the natural domain
of the space of multipliers. We say x is a random point whenever
P (h(x) = 0) < 1 for every h ∈ H.
Lemma 5.1. Let H be a committee space. Let x(1), x(2), . . . be an infi-
nite sequence of identically distributed independent random points, and
let Xn = (x(1), . . . , x(n)). The sequence PXn → I in the strong operator
topology.
Proof. Note that the sequence PXn must converge monotonically to
some projection in the strong operator topology, so it suffices to show
that its range is all of H.
Define a random vector h by choosing random point x and then
choosing a random unit vector in the finite dimensional space {ηη(x(k)) =
8
J. E. PASCOE
0}⊥. Note h is a random vector because P (η(x) = 0) < 1 by definition
of random point and therefore P (h ⊥ g) < 1.
Note that since the closed span of h(k) is almost surely all of H by
Lemma 4.1, then we are done as each h(k) is in the range of PXn for all
n ≥ k.
(cid:3)
As a corollary of our main result, we see that random Nevanlinna-
Pick problems satisfy the column-row property.
Corollary 5.2. Let H be a committee space. Let x(1), x(2), . . . be an
infinite sequence of identically distributed independent random points,
and let Xn = (x(1), . . . , x(n)). Let v be a random multiplier model. Then,
almost surely,
kRvk = lim
n→∞
kRXn
v k ≤ lim
n→∞
kC Xn
v k = kCvk.
Proof. By Theorem 3.1, it is enough to show the limits converge to
the appropriate values. Recall that PXn → I in the strong operator
topology almost surely by Lemma 5.1. Moreover, note the sequence
PXn is increasing.
Note,
RXn
v = E(PXnMfv PXnM ∗
fv PXn)
= E(PXnMfv M ∗
= PXnE(Mfv M ∗
fvPXn)
fv)PXn.
converges in the strong operator topology to Rv, and
Therefore, RXn
each kRXn
v
v k ≤ kRvk.
Now consider,
Therefore,
C Xn
v = E(PXnM ∗
fv PXnMfv PXn)
= E(M ∗
fvPXnMfv).
Cv − C Xn
v = E(M ∗
fv (I − PXn)Mfv)
is positive semi-definite. So, we see that kC Xn
the proof of Lemma 3.4,
v k ≤ kCvk. Recall, from
kCvk = hCv1, 1i = X kzγk2wγ.
So, it is sufficient to show that
lim
n→∞
hC Xn
v 1, 1i = Xγ
kzγk2wγ.
THE RANDOM COLUMN-ROW PROPERTY
9
Calculating,
Now consider,
C Xn
v = E(M ∗
= E(Xγ,γ ′
= Xγ
fv PXnMfv).
vγvγ ′M ∗
zγ PXnMzγ′ )
wγM ∗
zγ PXnMzγ .
hC Xn
v 1, 1i = hXγ
= Xγ
= Xγ
wγM ∗
zγ PXnMzγ 1, 1i
wγhPXnzγ, PXnzγi
kPXzγk2wγ.
As n → ∞ this converges monotonically to Pγ kzγk2wγ, since PXn is
increasing and PXn → I in the strong operator topology.
(cid:3)
5.1. Some conclusions on potential experiments. Suppose one
were looking for a counterexample to the claim some space, for exam-
ple the Drury-Arveson space, satisfied the true column row property.
The obvious thing to try is to take a random Nevanlinna-Pick prob-
lem. However, our result Theorem 5.2 says that if we choose many
interpolation nodes and a long column, we are doomed. Therefore, it
is advisable to choose the least number of interpolation nodes possible
and the shortest conceivable column length. (For example, 2.) Further-
more, if we are forced to choose a lot of nodes, it would be best not to
choose them uniformly.
Experiments performed on the two variable Fock space, performed
by Augat, Jury, and the present author, which are described in [4], gave
fairly poor results on random data. For examples arising from a single
2 by 2 matrix interpolation node with a column of length 2, choos-
ing elements at random yielded a column-row constant of only about
1.0043. At the time, it was thought that "bigger is better," however our
Corollary 5.2 says that is not the case. Later, hand-crafted examples
gave showed that the column-row property fails for any constant.
References
[1] J. Agler and J.E. McCarthy. Pick Interpolation and Hilbert Function Spaces.
American Mathematical Society, Providence, 2002.
10
J. E. PASCOE
[2] Alexandru Aleman, Michael Hartz, John E McCarthy, and Stefan Richter. In-
terpolating Sequences in Spaces with the Complete Pick Property. Int. Math.
Res. Not., 10 2017.
[3] Alexandru Aleman, Michael Hartz, John E McCarthy, and Stefan Richter. Weak
products of complete Pick spaces. Indiana Univ. Math. Jour., to appear., 2018.
[4] Meric Augat, Michael Jury, and James Eldred Pascoe. Effective non-
commutative Nevanlinna-Pick interpolation on the row ball and failure of the
column-row property for the Fock space. in preparation., 2019.
[5] Joseph A. Ball, Gregory Marx, and Victor Vinnikov. Noncommutative reproduc-
ing kernel Hilbert spaces. Journal of Functional Analysis, 271(7):1844 -- 1920,
2016.
[6] W. George Cochran, Joel H. Shapiro, and David C. Ullrich. Random Dirich-
let Functions: Multipliers and Smoothness. Canadian Journal of Mathematics,
45(2):255268, 1993.
[7] Michael T. Jury and Robert T.W. Martin. Factorization in weak products of
complete Pick spaces. Bulletin of the London Mathematical Society, 51(2):223 --
229, 2019.
[8] Tavan T. Trent. A corona theorem for multipliers on Dirichlet space. Integral
Equations and Operator Theory, 49(1):123 -- 139, May 2004.
|
1902.10766 | 1 | 1902 | 2019-02-27T20:26:52 | Boundedness of weighted iterated Hardy-type operators involving suprema from weighted Lebesgue spaces into weighted Ces\`{a}ro function spaces | [
"math.FA"
] | In this paper the boundedness of the weighted iterated Hardy-type operators $T_{u,b}$ and $T_{u,b}^*$ involving suprema from weighted Lebesgue space $L_p(v)$ into weighted Ces\`{a}ro function spaces ${\operatorname{Ces}}_{q}(w,a)$ are characterized. These results allow us to obtain the characterization of the boundedness of the supremal operator $R_u$ from $L^p(v)$ into ${\operatorname{Ces}}_{q}(w,a)$ on the cone of monotone non-increasing functions. For the convenience of the reader, we formulate the statement on the boundedness of the weighted Hardy operator $P_{u,b }$ from $L^p(v)$ into ${\operatorname{Ces}}_{q}(w,a)$ on the cone of monotone non-increasing functions. Under additional condition on $u$ and $b$, we are able to characterize the boundedness of weighted iterated Hardy-type operator $T_{u,b}$ involving suprema from $L^p(v)$ into ${\operatorname{Ces}}_q(w,a)$ on the cone of monotone non-increasing functions. At the end of the paper, as an application of obtained results, we calculate the norm of the fractional maximal function $M_{\gamma}$ from $\Lambda^p(v)$ into $\Gamma^q(w)$. | math.FA | math |
BOUNDEDNESS OF WEIGHTED ITERATED HARDY-TYPE OPERATORS INVOLVING SUPREMA
FROM WEIGHTED LEBESGUE SPACES INTO WEIGHTED CES `ARO FUNCTION SPACES
R.CH. MUSTAFAYEV AND N. BILGIC¸ LI
Abstract. In this paper the boundedness of the weighted iterated Hardy-type operators Tu,b and T ∗
u,b involving suprema
from weighted Lebesgue space Lp(v) into weighted Ces`aro function spaces Cesq(w, a) are characterized. These results
allow us to obtain the characterization of the boundedness of the supremal operator Ru from Lp(v) into Cesq(w, a)
on the cone of monotone non-increasing functions. For the convenience of the reader, we formulate the statement
on the boundedness of the weighted Hardy operator Pu,b from Lp(v) into Cesq(w, a) on the cone of monotone non-
increasing functions. Under additional condition on u and b, we are able to characterize the boundedness of weighted
iterated Hardy-type operator Tu,b involving suprema from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing
functions. At the end of the paper, as an application of obtained results, we calculate the norm of the fractional maximal
function Mγ from Λp(v) into Γq(w).
Many Banach spaces which play an important role in functional analysis and its applications are obtained in a
special way: the norms of these spaces are generated by positive sublinear operators and by Lp-norms.
In connection with Hardy and Copson operators
1. Introduction
(P f )(x) :=
1
xZ x
0
f (t) dt
and
(Q f )(x) :=Z ∞
x
f (t)
t
dt,
(x > 0),
the classical Ces`aro function space
Ces(p) :=(cid:26) f : k f kCes(p) :=(cid:18)Z ∞
0
xZ x
(cid:18) 1
0
f (t) dt(cid:19) p
dx(cid:19)
and the classical Copson function space
Cop(p) :=(cid:26) f : k f kCop(p) :=(cid:18)Z ∞
0
(cid:18)Z ∞
x
f (t)
t
dt(cid:19)p
dx(cid:19)
1
p
1
p
< ∞(cid:27),
< ∞(cid:27),
where 1 < p ≤ ∞, with the usual modifications if p = ∞, are of interest.
The classical Ces`aro function spaces Ces(p) have been introduced in 1970 by Shiue [48]. These spaces have
been defined analogously to the Ces`aro sequence spaces that appeared two years earlier in [40] when the Dutch
Mathematical Society posted a problem to find a representation of their dual spaces. In 1971 Leibowitz proved that
ces1 = {0} and for 1 < q < p ≤ ∞, ℓp and cesq sequence spaces are proper subspaces of cesp [32]. The problem posted
[40] was resolved by Jagers [28] in 1974 who gave an explicit isometric description of the dual of Ces`aro sequence
space. In [51], Sy, Zhang and Lee gave a description of dual spaces of Ces(p) spaces based on Jagers' result. In
1996 different, isomorphic description due to Bennett appeared in [4]. In [4, Theorem 21.1] Bennett observes that
the classical Ces`aro function space and the classical Copson function space coincide for p > 1. He also derives
estimates for the norms of the corresponding inclusion operators. The same result, with different estimates, is
due to Boas [7], who in fact obtained the integral analogue of the Askey-Boas Theorem [6, Lemma 6.18] and [1].
These results generalized in [27] using the blocking technique. In [2] they investigated dual spaces for Ces(p) for
1 < p < ∞. Their description can be viewed as being analogous to one given for sequence spaces in [4]. For a
long time, Ces`aro function spaces have not attracted a lot of attention contrary to their sequence counterparts. In
fact there is quite rich literature concerning different topics studied in Ces`aro sequence spaces as for instance in
[11 -- 15]. However, recently in a series of papers, Astashkin and Maligranda started to study the structure of Ces`aro
2010 Mathematics Subject Classification. 46E30, 26D10, 42B25, 42B35.
Key words and phrases. weighted iterated Hardy operators involving suprema, Ces`aro function spaces, fractional maximal functions,
classical Lorentz spaces.
1
function spaces. Among others, in [2] they investigated dual spaces for Ces(p) for 1 < p < ∞. Their description can
be viewed as being analogous to one given for sequence spaces in [4] (For more detailed information about history
of classical Ces`aro spaces see recent survey paper [3]).
Throughout the paper we assume that I := (a, b) ⊆ (0, ∞). By M(I) we denote the set of all measurable functions
on I. The symbol M+(I) stands for the collection of all f ∈ M(I) which are non-negative on I, while M+,↓(I) is
used to denote the subset of those functions which are non-increasing on I, respectively. A weight is a function
v ∈ M+(0, ∞) such that 0 < V(x) < ∞ for all x ∈ (0, ∞), where
2
The family of all weight functions (also called just weights) on (0, ∞) is given by W(0, ∞).
For p ∈ (0, ∞] and w ∈ M+(I), we define the functional k · kp,w,I on M(I) by
V(x) :=Z x
0
v(t) dt.
If, in addition, w ∈ W(I), then the weighted Lebesgue space Lp(w, I) is given by
k f kp,w,I :=
(cid:16)RI
f (x)pw(x) dx(cid:17)1/p
ess supI f (x)w(x)
if
if
p < ∞
p = ∞.
Lp(w, I) = { f ∈ M(I) : k f kp,w,I < ∞}
and it is equipped with the quasi-norm k · kp,w,I.
When I = (0, ∞), we write Lp(w) instead of Lp(w, (0, ∞)).
We adopt the following usual conventions.
Convention 1.1. We adopt the following conventions:
• Throughout the paper we put 0 · ∞ = 0, ∞/∞ = 0 and 0/0 = 0.
• If p ∈ [1, +∞], we define p′ by 1/p + 1/p′ = 1.
• If 0 < q < p < ∞, we define r by 1/r = 1/q − 1/p.
• If I = (a, b) ⊆ R and g is monotone function on I, then by g(a) and g(b) we mean the limits limx→a+ g(x)
and limx→b− g(x), respectively.
Throughout the paper, we always denote by c and C a positive constant, which is independent of main parameters
but it may vary from line to line. However a constant with subscript or superscript such as c1 does not change in
different occurrences. By a . b, (b & a) we mean that a ≤ λb, where λ > 0 depends on inessential parameters. If
a . b and b . a, we write a ≈ b and say that a and b are equivalent.
Unless a special remark is made, the differential element dx is omitted when the integrals under consideration
are the Lebesgue integrals.
The weighted Ces`aro and Copson function spaces are defined as follows:
Definition 1.2. Let 0 < p ≤ ∞, u ∈ M+(0, ∞) and v ∈ W(0, ∞). The weighted Ces`aro and Copson spaces are defined
by
Cesp(u, v) : =(cid:26) f ∈ M+(0, ∞) : k f kCes p(u,v) :=(cid:13)(cid:13)(cid:13)
Copp(u, v) : =(cid:26) f ∈ M+(0, ∞) : k f kCop p(u,v) :=(cid:13)(cid:13)(cid:13)
k f k1,v,(0,·)(cid:13)(cid:13)(cid:13)p,u,(0,∞)
< ∞(cid:27),
k f k1,v,(·,∞)(cid:13)(cid:13)(cid:13)p,u,(0,∞)
< ∞(cid:27),
and
respectively.
When v ≡ 1 on (0, ∞), we simply write Cesp(u) and Copp(u) instead of Cesp(u, v) and Copp(u, v), respectively.
Recall that Cesp(u, v) and Copp(u, v) are contained in the scale of weighted Ces`aro and Copson function spaces
Cesp,q(u, v) and Copp,q(u, v) defined in [22]. Obviously, Ces(p) = Cesp(x−1) and Cop(p) = Copp(x−1). In [29],
Kami´nska and Kubiak computed the dual norm of the Ces`aro function space Cesp(u), generated by 1 < p < ∞ and
an arbitrary positive weight u. A description presented in [29] resembles the approach of Jagers [28] for sequence
spaces.
Let u ∈ W(0, ∞) ∩ C(0, ∞), b ∈ W(0, ∞) and B(t) := R t
0
t ∈ (0, ∞). The weighted iterated Hardy-type operators involving suprema Tu,b and T ∗
b(s) ds. Assume that b is a weight such that b(t) > 0
u,b are defined at
for a.e.
g ∈ M+(0, ∞) by
3
(Tu,bg)(t) := sup
t≤τ
(T ∗
u,bg)(t) := sup
t≤τ
u(τ)
B(τ)Z τ
B(τ)Z ∞
u(τ)
0
τ
g(y)b(y) dy,
t ∈ (0, ∞),
g(y)b(y) dy,
t ∈ (0, ∞).
Such operators have been found indispensable in the search for optimal pairs of rearrangement-invariant norms
for which a Sobolev-type inequality holds (cf. [30]). They constitute a very useful tool for characterization of
the associate norm of an operator-induced norm, which naturally appears as an optimal domain norm in a Sobolev
embedding (cf. [38], [39]). Supremum operators are also very useful in limiting interpolation theory as can be seen
from their appearance for example in [18], [17], [16], [45]. Recall that Tu,b successfully controls non-increasing
rearrangements of wide range of maximal functions (see, for instance, [34] and references therein).
It was shown in [23] that for every h ∈ M+(0, ∞) and t ∈ (0, ∞)
where
Moreover, if the condition
(Tu,bh)(t) = (T ¯u,bh)(t),
¯u(t) := B(t) sup
t≤τ
u(τ)
B(τ)
,
t ∈ (0, ∞).
holds, then for all f ∈ M+,↓(0, ∞),
sup
0<t<∞
u(t)
B(t)Z t
0
b(τ)
u(τ)
dτ < ∞.
(Tu,b f )(t) ≈ (Ru f )(t) + (P ¯u,b f )(t),
t ∈ (0, ∞),
(1.1)
(1.2)
where the supremal operator Ru and the weighted Hardy operator Pu,b are defined for h ∈ M+(0, ∞) and t ∈ (0, ∞)
by
(Ruh)(t) : = sup
t≤τ
u(τ)h(τ),
(Pu,bh)(t) : =
u(t)
B(t)Z t
0
h(τ)b(τ) dτ,
respectively.
Recall that the boundedness of Ru from Lp(v) into Lq(w) on the cone of monotone non-increasing functions, that
is, the validity of the inequality
kRu f kLq(w) ≤ C k f kLp(v),
f ∈ M+,↓(0, ∞)
(1.3)
was completely characterized in [23] in the case 0 < p ≤ q < ∞. In the case 0 < q < p < ∞, [23] provides solution
when u is equivalent to a non-decreasing function on (0, ∞). The complete solution of inequality (1.3) using a
certain reduction method was presented in [21]. Another solution of (1.3) was obtained in [31].
Note that inequality
kPu,b( f )kq,w,(0,∞) ≤ ck f kp,v,(0,∞),
f ∈ M+,↓(0, ∞)
(1.4)
was considered by many authors and there exist several characterizations of this inequality (see, papers [5, 8, 9, 19,
20, 26]).
The complete characterizations of inequality
kTu,b f kq,w,(0,∞) ≤ Ck f kp,v,(0,∞),
f ∈ M+,↓(0, ∞)
(1.5)
for 0 < q ≤ ∞, 0 < p ≤ ∞ were given in [21] and [34]. Inequality (1.5) was characterized in [23, Theorem 3.5] under
condition (1.1). Note that the case when 0 < p ≤ 1 < q < ∞ was not considered in [23]. It is also worth to mention
that in the case when 1 < p < ∞, 0 < q < p < ∞, q , 1 [23, Theorem 3.5] contains only discrete condition. In [25]
the new reduction theorem was obtained when 0 < p ≤ 1, and this technique allowed to characterize inequality (1.5)
4
when b ≡ 1, and in the case when 0 < q < p ≤ 1, [25] contains only discrete condition. Using the results in [41 -- 44],
another characterization of (1.5) was obtained in [50] and [46].
In this paper we investigate the boundedness of Tu,b and T ∗
u,b from the weighted Lebesgue spaces Lp(v) into the
weighted Ces`aro spaces Cesq(w, a), when 1 < p, q < ∞ (see, Theorems 3.1 and 3.3). These results allow us to obtain
the characterization of the boundedness of Ru from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing
functions (see, Theorem 4.1). For the convenience of the reader, we formulate the statement on the boundedness
of Pu,b from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing functions (see, Theorem 5.1). In view
of (1.2), we are able to characterize the boundedness of Tu,b from Lp(v) into Cesq(w, a) on the cone of monotone
non-increasing functions (see, Theorem 6.1). At the end of the paper, as an application of obtained results, we
calculate the norm of the fractional maximal function Mγ from Λp(v) into Γq(w).
The paper is organized as follows. We start with formulations of "an integration by parts" formula in Section 2.
The boundedness results for Tu,b and T ∗
u,b from Lp(v) into Cesq(w, a) are presented in Section 3. The characteriza-
tions of the boundedness of Ru, Pu,b and Tu,b from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing
functions are given in Sections 4, 5 and 6, respectively. Finally, the obtained in previous sections results are applied
to calculate the norm of the operator Mγ : Λp(v) → Γq(w) in Section 7.
2. "An integration by parts" formula
We recall the following "an integration by parts" formula. For the convenience of the reader we give the proof
here (cf. [49, Lemma, p. 176]).
non-negative non-increasing right-continuous function on (0, ∞). Then
Theorem 2.1. Let α > 0. Let g be a non-negative function on (0, ∞) such that 0 < R t
f (t)] dt < ∞ ⇐⇒ A2 :=Z(0,∞)(cid:18)Z t
A1 := Z ∞
g(t)[ f (t) − lim
t→+∞
(cid:18)Z t
g(cid:19)α
0
0
0
0
g(cid:19)α+1
d[− f (t)] < ∞.
g < ∞, t > 0 and let f be a
Moreover, A1 ≈ A2.
Proof. Assume at first that limt→+∞ f (t) = 0. Let
Then
Since
we have that
Z x
0 (cid:18)Z t
0
g(cid:19)α
Integrating by parts, we get that
A1 =Z ∞
0
(cid:18)Z t
0
g(cid:19)α
g(t) f (t) dt < ∞.
Z x
0 (cid:18)Z t
0
g(cid:19)α
g(t) f (t) dt → 0,
as
x → 0 + .
g(t) f (t) dt ≥ f (x) Z x
0 (cid:18)Z t
0
g(cid:19)α
g(t) dt ≈ f (x)(cid:18)Z x
0
g(cid:19)α+1
,
x > 0,
f (x)(cid:18)Z x
0
g(cid:19)α+1
→ 0,
as
x → 0 + .
0
∞
A2 =Z(0,∞)(cid:18)Z t
d[− f (t)] = − f (t)(cid:18)Z t
g(cid:19)α+1
g(cid:19)α+1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+Z(0,∞)
+ (α + 1)Z ∞
f (t)(cid:18)Z t
f (t)(cid:18)Z t
g(cid:19)α+1
g(cid:19)α+1
≤ (α + 1)Z ∞
(cid:18)Z t
g(cid:19)α
g(t) f (t) dt = (α + 1)A1.
− lim
t→+∞
= lim
t→0+
0
0
0
0
0
0
0
Thus
A2 . A1.
g(cid:19)α+1
f (t) d(cid:18)Z t
(cid:18)Z t
g(cid:19)α
0
0
g(t) f (t) dt
Now assume that
Then
Since
Z[x,∞)(cid:18)Z t
0
g(cid:19)α+1
we obtain that
5
A2 :=Z(0,∞)(cid:18)Z t
0
g(cid:19)α+1
d[− f (t)] < ∞.
Z[x,∞)(cid:18)Z t
0
g(cid:19)α+1
d[− f (t)] → 0,
as
x → +∞.
0
d[− f (t)] ≥(cid:18)Z x
=(cid:18)Z x
f (x)(cid:18)Z x
0
0
d[− f (t)]
g(cid:19)α+1Z[x,∞)
g(cid:19)α+1
[ f (x) − lim
x→+∞
f (x)] = f (x)(cid:18)Z x
0
g(cid:19)α+1
,
x > 0,
g(cid:19)α+1
→ 0,
as
x → +∞.
Thus, integrating by parts, we get that
g(cid:19)α+1
0
0
0
0
∞
A1 =Z ∞
g(t) f (t) dt ≈Z ∞
(cid:18)Z t
f (t) d(cid:18)Z t
g(cid:19)α
(cid:18)Z t
= f (t)(cid:18)Z t
g(cid:19)α+1
g(cid:19)α+1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f (t)(cid:18)Z t
f (t)(cid:18)Z t
g(cid:19)α+1
(cid:18)Z t
g(cid:19)α+1
≤Z ∞
+Z ∞
d[− f (t)] = A2.
g(cid:19)α+1
− lim
t→0+
= lim
t→∞
0
0
0
0
0
0
0
0
d[− f (t)]
+Z ∞
0
(cid:18)Z t
0
g(cid:19)α+1
d[− f (t)]
Hence
We have shown that if limx→+∞ f (x) = 0, then
A1 . A2.
and
A1 < ∞ ⇐⇒ A2 < ∞,
A1 ≈ A2.
Now assume that limx→+∞ f (x) > 0. Then, applying previous statement to the function f (x) − limx→+∞ f (x), we
arrive at
0
The proof is completed.
Z ∞
(cid:18)Z t
0
g(cid:19)α
g(t)[ f (t) − lim
x→+∞
f (x)] dt ≈Z(0,∞)(cid:18)Z t
0
g(cid:19)α+1
d[− f (t)].
(cid:3)
Remark 2.2. Note that if f ∈ M+,↓(0, ∞) is such that limx→+∞ f (x) > 0, then
Z ∞
0
(cid:18)Z t
0
g(cid:19)α
g(t) f (t) dt < ∞ =⇒ Z ∞
0
g(x) dx < ∞.
Indeed: for each x ∈ (0, ∞)
∞ >Z ∞
0
(cid:18)Z t
0
g(cid:19)α
holds. Thus
lim
x→+∞
f (x) ·(cid:18)Z x
0
g(cid:19)α+1
≤ f (x)(cid:18)Z x
0
g(cid:19)α+1
≤Z ∞
0
g(t) f (t) dt ≥Z x
0 (cid:18)Z t
≥ f (x) Z x
g(cid:19)α
0 (cid:18)Z t
0
0
g(cid:19)α
g(t) f (t) dt
g(cid:19)α+1
g(t) dt ≈ f (x)(cid:18)Z x
(cid:18)Z t
g(cid:19)α
0
0
g(t) f (t) dt < ∞.
Hence
Therefore
lim
x→+∞
6
g(cid:19)α+1
< ∞.
0
f (x) ·(cid:18)Z ∞
Z ∞
0
g < ∞.
non-negative non-increasing right-continuous function on (0, ∞). Then
Corollary 2.3. Let α > 0. Let g be a non-negative function on (0, ∞) such that 0 < R t
f (x) ·(cid:18)Z ∞
g(t) f (t) dt ≈Z(0,∞)(cid:18)Z t
d[− f (t)] + lim
x→+∞
g(cid:19)α+1
(cid:18)Z t
Z ∞
g(cid:19)α
0
0
0
0
0
g(cid:19)α+1
.
g < ∞, t > 0 and let f be a
Proof. If limx→+∞ f (x) = 0, then the statement follows by Theorem 2.1. If limx→+∞ f (x) > 0, then by Remark 2.2,
we know that
Z ∞
(cid:18)Z t
g(cid:19)α
g(t) f (t) dt < ∞ =⇒ Z ∞
0
g(x) dx < ∞.
0
Therefore, by Theorem 2.1, we get that
0
Z ∞
0
(cid:18)Z t
0
g(cid:19)α
g(t) f (t) dt =Z ∞
0
(cid:18)Z t
g(cid:19)α
≈Z(0,∞)(cid:18)Z t
g(cid:19)α+1
0
0
g(t)[ f (t) − lim
x→+∞
f (x)] dt + lim
x→+∞
d[− f (t)] + lim
x→+∞
f (x) ·(cid:18)Z ∞
0
(cid:18)Z t
0
g(cid:19)α
g(t) dt
f (x) ·Z ∞
g(cid:19)α+1
0
.
The proof is completed.
non-negative non-decreasing left-continuous function on (0, ∞). Then
Theorem 2.4. Let α > 0. Let g be a non-negative function on (0, ∞) such that 0 < R ∞
g(t)[ f (t) − f (0+)] dt < ∞ ⇐⇒ B2 := Z(0,∞)(cid:18)Z ∞
B1 := Z ∞
(cid:18)Z ∞
g(cid:19)α
t
t
t
g(cid:19)α+1
d[ f (t)] < ∞.
0
Moreover, B1 ≈ B2.
(cid:3)
g < ∞, t > 0 and let f be a
Proof. Assume at first that f (0+) = 0. Let
Then
Since
we have that
0
B1 := Z ∞
(cid:18)Z ∞
g(cid:19)α
t
Z ∞
x
(cid:18)Z ∞
t
g(cid:19)α
g(t) f (t) dt < ∞.
g(t) f (t) dt → 0,
as
x → ∞.
Z ∞
x
(cid:18)Z ∞
t
g(cid:19)α
g(t) f (t) dt ≥ f (x) Z ∞
x
(cid:18)Z ∞
t
g(cid:19)α
g(t) dt ≈ f (x)(cid:18)Z ∞
x
g(cid:19)α+1
,
x > 0,
f (x)(cid:18)Z ∞
x
g(cid:19)α+1
→ 0,
as
x → ∞.
Hence, integrating by parts, we get that
0
∞
g(cid:19)α+1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
−Z(0,∞)
+ (α + 1)Z ∞
g(cid:19)α+1
f (t) d(cid:18)Z ∞
g(cid:19)α+1
(cid:18)Z ∞
g(cid:19)α
0
t
t
g(t) f (t) dt
t
B2 =Z(0,∞)(cid:18)Z ∞
f (t)(cid:18)Z ∞
≤ (α + 1)Z ∞
= lim
t→∞
t
0
t
− lim
t→0+
d[ f (t)] = f (t)(cid:18)Z ∞
f (t)(cid:18)Z ∞
g(cid:19)α+1
g(cid:19)α+1
(cid:18)Z ∞
g(cid:19)α
B2 := Z(0,∞)(cid:18)Z ∞
t
t
t
g(t) f (t) dt = (α + 1)B1.
g(cid:19)α+1
d[ f (t)] < ∞.
Now assume that
Then
Since
we obtain that
Z(0,x](cid:18)Z ∞
t
g(cid:19)α+1
Z(0,x](cid:18)Z ∞
t
g(cid:19)α+1
d[ f (t)] → 0,
as
x → 0 + .
d[ f (t)]
g(cid:19)α+1Z(0,x]
g(cid:19)α+1
x
d[ f (t)] ≥(cid:18)Z ∞
=(cid:18)Z ∞
f (x)(cid:18)Z ∞
x
x
g(cid:19)α+1
→ 0,
as
x → 0 + .
[ f (x) − f (0+)] = f (x)(cid:18)Z ∞
x
g(cid:19)α+1
,
x > 0,
Thus, integrating by parts, we get that
g(cid:19)α+1(cid:21)
t
0
0
∞
B1 =Z ∞
(cid:18)Z ∞
g(t) f (t) dt ≈Z ∞
f (t) d(cid:20) −(cid:18)Z ∞
g(cid:19)α
= − f (t)(cid:18)Z ∞
(cid:18)Z ∞
g(cid:19)α+1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
g(cid:19)α+1
f (t)(cid:18)Z ∞
f (t)(cid:18)Z ∞
g(cid:19)α+1
(cid:18)Z ∞
g(cid:19)α+1
≤Z ∞
+Z ∞
g(cid:19)α+1
d[ f (t)] = B2.
= lim
t→0+
− lim
t→∞
d[ f (t)]
0
0
0
t
t
t
t
t
t
+Z ∞
0
(cid:18)Z ∞
t
g(cid:19)α+1
d[ f (t)]
We have shown that if f (0+) = 0, then
and
B1 < ∞ ⇐⇒ B2 < ∞,
B1 ≈ B2.
Now assume that f (0+) > 0. Then, applying previous statement to the function f (x) − f (0+), we arrive at
Z ∞
0
(cid:18)Z ∞
t
g(cid:19)α
g(t)[ f (t) − f (0+)] dt ≈Z(0,∞)(cid:18)Z ∞
t
g(cid:19)α+1
d[ f (t)].
The proof is completed.
Remark 2.5. Note that if f is a non-negative non-decreasing function on (0, ∞) such that f (0+) > 0, then
Z ∞
0
(cid:18)Z ∞
t
g(cid:19)α
g(t) f (t) dt < ∞ =⇒ Z ∞
0
g(x) dx < ∞.
7
(cid:3)
Indeed: for each x ∈ (0, ∞)
∞ >Z ∞
0
(cid:18)Z ∞
t
g(cid:19)α
g(t) f (t) dt ≥Z ∞
x
(cid:18)Z ∞
≥ f (x) Z ∞
t
x
g(t) f (t) dt
g(cid:19)α
(cid:18)Z ∞
≤Z ∞
t
0
g(cid:19)α
(cid:18)Z ∞
t
g(t) dt ≈ f (x)(cid:18)Z ∞
x
g(cid:19)α+1
g(cid:19)α
g(t) f (t) dt < ∞.
f (0+)(cid:18)Z ∞
x
g(cid:19)α+1
holds. Thus
Hence
Therefore
≤ f (x)(cid:18)Z ∞
x
g(cid:19)α+1
f (0+)(cid:18)Z ∞
Z ∞
0
0
g < ∞.
g(cid:19)α+1
< ∞.
8
g < ∞, t > 0 and let f be a
g(cid:19)α+1
.
non-negative non-decreasing left-continuous function on (0, ∞). Then
Corollary 2.6. Let α > 0. Let g be a non-negative function on (0, ∞) such that 0 < R ∞
d[ f (t)] + f (0+)(cid:18)Z ∞
g(t) f (t) dt ≈Z(0,∞)(cid:18)Z ∞
(cid:18)Z ∞
g(cid:19)α+1
Z ∞
g(cid:19)α
0
0
t
t
t
Proof. If f (0+) = 0, then the statement follows by Theorem 2.4. If f (0+) > 0, then by Remark 2.5, we know that
Z ∞
0
(cid:18)Z ∞
t
g(cid:19)α
g(t) f (t) dt < ∞ =⇒ Z ∞
0
g(x) dx < ∞.
Therefore, by Theorem 2.4, we get that
Z ∞
0
(cid:18)Z ∞
t
g(cid:19)α
g(t) f (t) dt =Z ∞
0
(cid:18)Z ∞
g(t)[ f (t) − f (0+)] dt + f (0+) Z ∞
g(cid:19)α
≈Z(0,∞)(cid:18)Z ∞
g(cid:19)α+1
g(cid:19)α+1
d[ f (t)] + f (0+)(cid:18)Z ∞
0
0
.
t
t
(cid:18)Z ∞
t
g(cid:19)α
g(t) dt
The proof is completed.
(cid:3)
3. The boundedness of Tu,b and T ∗
u,b
from Lp(v) into Cesq(w, a)
In this section we give solutions of the following two inequalities
(cid:18)Z ∞
0
(cid:18)Z x
0 (cid:18) sup
t≤τ
u(τ)
B(τ)Z τ
0
and
(cid:18)Z ∞
0
(cid:18)Z x
0 (cid:18) sup
t≤τ
u(τ)
B(τ)Z ∞
τ
h(y)b(y) dy(cid:19) a(t) dt(cid:19)q
w(x) dx(cid:19)
h(y)b(y) dy(cid:19) a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
1
q
≤ C(cid:18)Z ∞
0
h(s)pv(s) ds(cid:19)
≤ C(cid:18)Z ∞
0
h(s)pv(s) ds(cid:19)
1
p
,
1
p
,
h ∈ M+(0, ∞)
(3.1)
h ∈ M+(0, ∞),
(3.2)
where 1 < p ≤ q < ∞ and a, u, v, w ∈ W(0, ∞). Using the duality argument, we reduce the problem to the bound-
edness for the dual of integral Volterra operator with a kernel satisfying Oinarovs condition and weighted Stieltjes
operator.
Note that the characterization of inequalities
and
(cid:18)Z ∞
0
(cid:18)Z ∞
x
(cid:18) sup
t≤τ
(cid:18)Z ∞
0
(cid:18)Z ∞
x
(cid:18) sup
t≤τ
u(τ)
B(τ)Z τ
0
u(τ)
B(τ)Z ∞
τ
h(y)b(y) dy(cid:19) a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
≤ C(cid:18)Z ∞
0
h(s)pv(s) ds(cid:19)
1
p
,
h ∈ M+(0, ∞)
(3.3)
h(y)b(y) dy(cid:19) a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
≤ C(cid:18)Z ∞
0
h(s)pv(s) ds(cid:19)
1
p
,
h ∈ M+(0, ∞)
(3.4)
can be reduced to the solutions of (3.1) and (3.2).
Recall that, if F is a non-negative non-decreasing function on (0, ∞), then
likewise, when F is a non-negative non-increasing function on (0, ∞), then
ess sup
t∈(0,∞)
F(t)G(t) = ess sup
t∈(0,∞)
F(t) ess sup
τ∈(t,∞)
G(τ),
ess sup
t∈(0,∞)
F(t)G(t) = ess sup
t∈(0,∞)
F(t) ess sup
τ∈(0,t)
G(τ)
(see, for instance, [24, p. 85]).
We need the following notations:
A(t) :=R t
0
a(s)ds,
U(t) :=R t
0
u(s)ds,
W(t) :=R t
0
w(s)ds.
(3.5)
(3.6)
Theorem 3.1. Let 1 < p, q < ∞. Assume that u ∈ W(0, ∞) ∩ C(0, ∞) and a, v, w ∈ W(0, ∞). Moreover, assume that
9
0 <Z x
0
v(t)1−p′
dt < ∞
for all
x > 0.
(i) If p ≤ q, then
sup
h≥0
(ii) If q < p, then
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ u(τ)R τ
0
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
≈ sup
1
p
0
0
0
+ sup
v(x)1−p′
h(s)pv(s) ds(cid:19)
(cid:18)R ∞
t∈(0,∞)(cid:18)Z t
v(x)1−p′(cid:18)Z t
x (cid:18) sup
t∈(0,∞)(cid:18)Z t
dx(cid:19)
x∈(0,∞)(cid:18)Z[x,∞)
d(cid:18) − sup
x∈(0,∞)(cid:18)Z(0,x]
A(t)p′
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
+ sup
+ sup
t≤τ
0
1
s≤τ
u(τ)(cid:19)a(s) ds(cid:19) p′
p′ (cid:18)Z ∞
(cid:18)Z y
t (cid:18) sup
u(τ)p′(cid:18)Z τ
v(s)1−p′
s≤τ
0
t
t≤τ
d(cid:18) − sup
t→∞(cid:18) sup
u(τ)p′(cid:18)Z τ
u(τ)(cid:18)Z τ
lim
t≤τ
1
q
0
0
1
1
q
t
1
1
q
p′ (cid:18)Z ∞
dx(cid:19)
w(y) dy(cid:19)
u(τ)(cid:19)a(s) ds(cid:19)q
w(y) dy(cid:19)
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)
v(s)1−p′
p′(cid:19);
ds(cid:19)
v(s)1−p′
0
x
1
1
1
q
1
q
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ u(τ)R τ
0
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
sup
h≥0
1
p
0
0
0
0
r
q′
h(s)pv(s) ds(cid:19)
(cid:18)R ∞
≈ (cid:18)Z ∞
(cid:18)Z t
v(x)1−p′
+(cid:18)Z ∞
(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
+(cid:18)Z ∞
(cid:18)Z(0,x]
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
dx(cid:19)
v(x)1−p′(cid:18)Z t
d(cid:18) −(cid:18) sup
A(t)p′
t≤τ
0
0
0
0
r
q
r
t
r
p
s≤y
s≤y
v(t)1−p′(cid:18)Z ∞
x (cid:18) sup
u(τ)p′(cid:18)Z τ
w(z) dz(cid:19)
(cid:18)Z z
u(y)(cid:19)a(s) ds(cid:19)q
t (cid:18) sup
dt(cid:19)
u(y)(cid:19)a(s) ds(cid:19) p′
p′(cid:18)Z ∞
w(s) ds(cid:19)
w(t) dt(cid:19)
dx(cid:19)
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)(cid:19)
v(s)1−p′
p′(cid:19).
ds(cid:19)
u(τ)p′(cid:18)Z τ
u(τ)(cid:18)Z τ
d(cid:18) −(cid:18) sup
t→∞(cid:18) sup
v(s)1−p′
v(s)1−p′
lim
t≤τ
t≤τ
1
q
r
p
0
0
0
x
z
1
r
r
0
r
p
1
r
1
r
1
r
A(x)qw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
Proof. Assume that 1 < p ≤ q < ∞. By duality, using Fubini's Theorem, and interchanging the suprema, we get
that
(cid:18)R ∞
0 (cid:18)R x
0 (cid:18) supt≤τ u(τ)R τ
0
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
sup
h≥0
1
p
(cid:18)R ∞
0
h(s)pv(s) ds(cid:19)
= sup
h≥0
1
(cid:18)R ∞
0
h(s)pv(s) ds(cid:19)
sup
g≥0
1
p
0 (cid:18)R x
R ∞
0 (cid:18) supt≤τ u(τ)R τ
0
h(y) dy(cid:19)a(t) dt(cid:19)g(x) dx
(cid:18)R ∞
0
g(x)q′ w(x)1−q′ dx(cid:19)
1
q′
= sup
g≥0
1
g(x)q′ w(x)1−q′ dx(cid:19)
sup
h≥0
1
q′
(cid:18)R ∞
0
Applying [23, Theorems 4.4], on using (3.5), we arrive at
sup
h≥0
where
0
0 (cid:18) supt≤τ u(τ)R τ
R ∞
(cid:18)R ∞
0
t
h(y) dy(cid:19)(cid:18)R ∞
h(s)pv(s) ds(cid:19)
10
g(x) dx(cid:19)a(t) dt
1
p
.
0
0 (cid:18) supt≤τ u(τ)R τ
R ∞
(cid:18)R ∞
0
t
h(y) dy(cid:19)(cid:18)R ∞
h(s)pv(s) ds(cid:19)
g(x) dx(cid:19)a(t) dt
1
p
≈ D + E,
0
D := (cid:18)Z ∞
E := (cid:18)Z ∞
0
t
s≤τ
(cid:18)Z ∞
(cid:18) sup
0 (cid:18)Z ∞
(cid:18)Z t
s
s
u(τ)(cid:19)(cid:18)Z ∞
g(x) dx(cid:19)a(s) ds(cid:19)
p′
g(x) dx(cid:19)a(s) ds(cid:19)
t≤τ
p (cid:18) sup
u(τ)p′(cid:18)Z τ
0
u(τ)(cid:19)(cid:18)Z t
0
v(s)1−p′
v(s)1−p′
ds(cid:19)(cid:19)(cid:18)Z ∞
t
t
ds(cid:19)(cid:18)Z ∞
g(x) dx(cid:19)a(t) dt(cid:19)
1
p′
.
g(x) dx(cid:19)a(t) dt(cid:19)
1
p′
,
p′
p (cid:18) sup
t≤τ
Integrating by parts (applying Corollary 2.6), on using Fubini's Theorem, we arrive at
0
D ≈(cid:18)Z ∞
=(cid:18)Z ∞
0
t
(cid:18)Z ∞
(cid:18)Z ∞
t
s≤τ
(cid:18) sup
g(x)Z x
u(τ)(cid:19)(cid:18)Z ∞
(cid:18) sup
g(x) dx(cid:19)a(s) ds(cid:19) p′
u(τ)(cid:19)a(s) ds dx(cid:19) p′
s≤τ
s
t
v(t)1−p′
1
p′
dt(cid:19)
dt(cid:19)
1
p′
.
v(t)1−p′
Similarly, integrating by parts (applying Corollary 2.3), on using Fubini's Theorem, we get at
s
0
0
0
0
t≤τ
v(s)1−p′
E ≈(cid:18)Z ∞
u(τ)p′(cid:18)Z τ
(cid:18) sup
g(x) dx(cid:19)a(s) ds(cid:19) p′
≈(cid:18)Z ∞
0 (cid:18)Z ∞
(cid:18)Z t
+(cid:18)Z ∞
(cid:18)Z ∞
g(x) dx(cid:19)a(s) ds(cid:19) lim
g(x)A(x) dx(cid:19) p′
≈(cid:18)Z ∞
(cid:18)Z t
d(cid:18) −(cid:18) sup
g(x) dx(cid:19) p′
+(cid:18)Z ∞
(cid:18)Z ∞
A(t)p′
+(cid:18)Z ∞
g(x)A(x) dx(cid:19) lim
t→∞(cid:18) sup
t≤τ
t≤τ
0
0
0
0
s
t
sup
g≥0
D
1
q′
(cid:18)R ∞
0
gq′ w1−q′(cid:19)
= sup
g≥0
≈ sup
0
0
(cid:18)R ∞
t∈(0,∞)(cid:18)Z t
t∈(0,∞)(cid:18)Z t
gq′ w1−q′(cid:19)
v(x)1−p′ (cid:18)Z t
x (cid:18) sup
dx(cid:19)
v(x)1−p′
+ sup
s≤y
0
0
1
(i) Let p ≤ q. By [35, Theorem 1.1], we obtain that
0 (cid:18)Z ∞
ds(cid:19)(cid:19) d(cid:18)Z t
d(cid:18) −(cid:18) sup
g(x) dx(cid:19)a(s) ds(cid:19) p′
u(τ)p′(cid:18)Z τ
v(s)1−p′
(cid:19)
ds(cid:19)(cid:19)(cid:19)(cid:19)
t≤τ
s
0
1
p′
1
p′
t→∞(cid:18) sup
t≤τ
u(τ)p′(cid:18)Z τ
0
u(τ)p′(cid:18)Z τ
0
v(s)1−p′
t≤τ
d(cid:18) −(cid:18) sup
u(τ)(cid:18)Z τ
0
u(τ)p′(cid:18)Z τ
ds(cid:19)
v(s)1−p′
0
v(s)1−p′
ds(cid:19)(cid:19)
ds(cid:19)(cid:19)(cid:19)(cid:19)
v(s)1−p′
1
p′
1
p′
ds(cid:19)(cid:19)(cid:19)(cid:19)
1
p′(cid:19) := E1 + E2 + E3.
1
q′ (cid:18)Z ∞
1
(cid:18)Z ∞
t
g(x)Z x
t
(cid:18) sup
s≤y
u(y)(cid:19)a(s) ds dx(cid:19) p′
v(t)1−p′
1
p′
dt(cid:19)
u(y)(cid:19)a(s) ds(cid:19) p′
(cid:18)Z z
t (cid:18) sup
s≤y
t
p′ (cid:18)Z ∞
1
1
q
t
p′ (cid:18)Z ∞
dx(cid:19)
u(y)(cid:19)a(s) ds(cid:19)q
w(y) dy(cid:19)
w(z) dz(cid:19)
1
q
.
(3.7)
By [33, Theorem 1, p. 40 and Theorem 3, p. 44], respectively, we have that
11
sup
g≥0
E1
1
q′
(cid:18)R ∞
0
gq′ w1−q′(cid:19)
(cid:18)R ∞
0 (cid:18)R t
0
= sup
g≥0
g(x)A(x) dx(cid:19) p′
≈ sup
x∈(0,∞)(cid:18)Z[x,∞)
d(cid:18) −(cid:18) sup
t≤τ
0
d(cid:18) −(cid:18) supt≤τ u(τ)p′(cid:18)R τ
(cid:18)R ∞
u(τ)p′(cid:18)Z τ
gq′ w1−q′(cid:19)
v(s)1−p′
ds(cid:19)(cid:19)(cid:19)(cid:19)
1
q′
0
0
v(s)1−p′
1
p′
ds(cid:19)(cid:19)(cid:19)(cid:19)
1
p′ (cid:18)Z x
0
1
q
A(y)qw(y) dy(cid:19)
and
sup
g≥0
E2
1
q′
(cid:18)R ∞
0
gq′ w1−q′(cid:19)
0 (cid:18)R ∞
(cid:18)R ∞
t
= sup
g≥0
g(x) dx(cid:19) p′
A(t)p′
≈ sup
x∈(0,∞)(cid:18)Z(0,x]
A(t)p′
By duality, we have that
d(cid:18) −(cid:18) sup
t≤τ
0
1
q′
d(cid:18) −(cid:18) supt≤τ u(τ)p′(cid:18)R τ
gq′ w1−q′(cid:19)
(cid:18)R ∞
u(τ)p′(cid:18)Z τ
v(s)1−p′
0
0
v(s)1−p′
1
p′
ds(cid:19)(cid:19)(cid:19)(cid:19)
ds(cid:19)(cid:19)(cid:19)(cid:19)
1
p′ (cid:18)Z ∞
x
1
q
.
w(y) dy(cid:19)
E3
sup
g≥0
1
q′
0
0
g(x)A(x) dx
= sup
g≥0
gq′ w1−q′(cid:19)
(cid:18)R ∞
R ∞
gq′ w1−q′(cid:19)
(cid:18)R ∞
=(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
1
q′
0
0
· lim
t→∞(cid:18) sup
t≤τ
1
q
lim
t→∞(cid:18) sup
t≤τ
u(τ)(cid:18)Z τ
0
v(s)1−p′
ds(cid:19)
1
p′(cid:19)
u(τ)(cid:18)Z τ
0
v(s)1−p′
ds(cid:19)
1
p′(cid:19).
Thus, we get that
sup
g≥0
≈ sup
E
1
q′
0
(cid:18)R ∞
gq′ w1−q′(cid:19)
x∈(0,∞)(cid:18)Z[x,∞)
x∈(0,∞)(cid:18)Z(0,x]
+(cid:18)Z ∞
+ sup
0
t≤τ
d(cid:18) − sup
A(t)p′
v(s)1−p′
0
u(τ)p′(cid:18)Z τ
d(cid:18) − sup
t→∞(cid:18) sup
lim
t≤τ
t≤τ
1
q
0
u(τ)p′(cid:18)Z τ
u(τ)(cid:18)Z τ
0
1
ds(cid:19)(cid:19)(cid:19)
v(s)1−p′
1
q
1
0
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
ds(cid:19)(cid:19)(cid:19)
p′(cid:19).
ds(cid:19)
x
1
w(y) dy(cid:19)
v(s)1−p′
A(y)qw(y) dy(cid:19)
1
q
(3.8)
Combining (3.7) and (3.8), we arrive at
12
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ u(τ)R τ
0
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
sup
h≥0
≈ sup
1
p
0
0
0
+ sup
v(x)1−p′
(cid:18)R ∞
h(s)pv(s) ds(cid:19)
t∈(0,∞)(cid:18)Z t
v(x)1−p′(cid:18)Z t
x (cid:18) sup
t∈(0,∞)(cid:18)Z t
dx(cid:19)
x∈(0,∞)(cid:18)Z[x,∞)
d(cid:18) − sup
x∈(0,∞)(cid:18)Z(0,x]
A(t)p′
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
+ sup
+ sup
t≤τ
1
q
0
1
s≤τ
u(τ)(cid:19)a(s) ds(cid:19) p′
p′ (cid:18)Z ∞
(cid:18)Z y
t (cid:18) sup
u(τ)p′(cid:18)Z τ
v(s)1−p′
s≤τ
0
t
t≤τ
d(cid:18) − sup
t→∞(cid:18) sup
u(τ)p′(cid:18)Z τ
u(τ)(cid:18)Z τ
lim
t≤τ
0
0
1
1
q
t
1
1
q
p′ (cid:18)Z ∞
dx(cid:19)
w(y) dy(cid:19)
u(τ)(cid:19)a(s) ds(cid:19)q
w(y) dy(cid:19)
p′ (cid:18)Z x
ds(cid:19)(cid:19)(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
ds(cid:19)(cid:19)(cid:19)
w(y) dy(cid:19)
v(s)1−p′
p′(cid:19).
ds(cid:19)
v(s)1−p′
0
x
1
1
1
q
1
q
(ii) Let now q < p. By [35, Theorem 1.2], we obtain that
sup
g≥0
D
1
q′
(cid:18)R ∞
0
gq′ w1−q′(cid:19)
0 (cid:18)R ∞
(cid:18)R ∞
t
= sup
g≥0
t (cid:18) sups≤y u(y)(cid:19)a(s) ds dx(cid:19) p′
v(t)1−p′
1
p′
dt(cid:19)
g(x)R x
r
q′
0
≈ (cid:18)Z ∞
(cid:18)Z t
v(x)1−p′(cid:19)
+(cid:18)Z ∞
(cid:18)Z t
v(x)1−p′(cid:18)Z t
0
0
0
1
q′
0
gq′ w1−q′(cid:19)
(cid:18)R ∞
v(t)1−p′(cid:18)Z ∞
(cid:18)Z z
t (cid:18) sup
u(y)(cid:19)a(s) ds(cid:19) p′
x (cid:18) sup
s≤y
s≤y
t
u(y)(cid:19)a(s) ds(cid:19)q
p′(cid:18)Z ∞
dx(cid:19)
w(z) dz(cid:19)
w(s) ds(cid:19)
r
p
t
r
1
r
r
q
dt(cid:19)
w(t) dt(cid:19)
1
r
.
(3.9)
By [33, Theorem 2, p. 48], we have that
sup
g≥0
E1
1
q′
(cid:18)R ∞
0
gq′ w1−q′(cid:19)
0 (cid:18)R t
(cid:18)R ∞
0
= sup
g≥0
g(x)A(x) dx(cid:19) p′
and
≈(cid:18)Z ∞
0
(cid:18)Z[x,∞)
d(cid:18) −(cid:18) sup
t≤τ
sup
g≥0
E2
1
q′
(cid:18)R ∞
0
gq′ w1−q′(cid:19)
v(s)1−p′
1
p′
ds(cid:19)(cid:19)(cid:19)(cid:19)
0
d(cid:18) −(cid:18) supt≤τ u(τ)p′(cid:18)R τ
(cid:18)R ∞
u(τ)p′(cid:18)Z τ
gq′ w1−q′(cid:19)
v(s)1−p′
ds(cid:19)(cid:19)(cid:19)(cid:19)
1
q′
0
0
r
p′ (cid:18)Z x
0
A(y)qw(y) dy(cid:19)
r
p
1
r
A(x)qw(x) dx(cid:19)
0 (cid:18)R ∞
(cid:18)R ∞
t
g(x) dx(cid:19) p′
= sup
g≥0
A(t)p′
≈(cid:18)Z ∞
0
(cid:18)Z(0,x]
A(t)p′
Consequently, we arrive at
d(cid:18) −(cid:18) sup
t≤τ
v(s)1−p′
1
p′
ds(cid:19)(cid:19)(cid:19)(cid:19)
13
0
1
q′
d(cid:18) −(cid:18) supt≤τ u(τ)p′(cid:18)R τ
gq′ w1−q′(cid:19)
(cid:18)R ∞
u(τ)p′(cid:18)Z τ
v(s)1−p′
0
0
ds(cid:19)(cid:19)(cid:19)(cid:19)
r
p′ (cid:18)Z ∞
x
r
p
w(y) dy(cid:19)
w(x) dx(cid:19)
1
r
.
E
sup
g≥0
1
q′
0
0
(cid:18)R ∞
≈ (cid:18)Z ∞
+(cid:18)Z ∞
+(cid:18)Z ∞
gq′ w1−q′(cid:19)
(cid:18)Z[x,∞)
(cid:18)Z(0,x]
A(y)qw(y) dy(cid:19)
d(cid:18) −(cid:18) sup
A(t)p′
t≤τ
0
0
v(s)1−p′
0
u(τ)p′(cid:18)Z τ
d(cid:18) −(cid:18) sup
t→∞(cid:18) sup
lim
t≤τ
t≤τ
1
q
u(τ)p′(cid:18)Z τ
u(τ)(cid:18)Z τ
0
0
ds(cid:19)(cid:19)(cid:19)(cid:19)
v(s)1−p′
v(s)1−p′
r
p
r
r
0
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
ds(cid:19)(cid:19)(cid:19)(cid:19)
p′(cid:19).
ds(cid:19)
x
1
w(y) dy(cid:19)
A(x)qw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
r
p
1
r
(3.10)
Combining (3.9) and (3.10), we arrive at
sup
h≥0
(cid:18)R ∞
0 (cid:18)R x
0 (cid:18) supt≤τ u(τ)R τ
0
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
1
p
0
0
0
0
r
q′
h(s)pv(s) ds(cid:19)
(cid:18)R ∞
≈ (cid:18)Z ∞
(cid:18)Z t
v(x)1−p′(cid:19)
+(cid:18)Z ∞
(cid:18)Z t
v(x)1−p′(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
d(cid:18) −(cid:18) sup
+(cid:18)Z ∞
(cid:18)Z(0,x]
A(t)p′
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
t≤τ
0
0
0
0
t
s≤y
(cid:18)Z z
t (cid:18) sup
u(y)(cid:19)a(s) ds(cid:19) p′
v(t)1−p′(cid:18)Z ∞
x (cid:18) sup
u(τ)p′(cid:18)Z τ
v(s)1−p′
s≤y
0
t≤τ
d(cid:18) −(cid:18) sup
t→∞(cid:18) sup
u(τ)p′(cid:18)Z τ
u(τ)(cid:18)Z τ
lim
t≤τ
1
q
0
0
r
q
1
r
r
r
t
r
p
u(y)(cid:19)a(s) ds(cid:19)q
dt(cid:19)
w(z) dz(cid:19)
p′(cid:18)Z ∞
w(s) ds(cid:19)
w(t) dt(cid:19)
dx(cid:19)
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)(cid:19)
v(s)1−p′
p′(cid:19).
ds(cid:19)
v(s)1−p′
r
p
0
x
1
r
r
p
1
r
1
r
A(x)qw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
The proof is completed.
(cid:3)
Theorem 3.2. Let 1 < p, q < ∞ and b ∈ W(0, ∞) be such that b(t) > 0 for a.e.
W(0, ∞) ∩ C(0, ∞) and a, v, w ∈ W(0, ∞). Moreover, assume that
t ∈ (0, ∞). Assume that u ∈
(i) If p ≤ q, then
0 <Z x
0
v(t)1−p′
dt < ∞
for all
x > 0.
(cid:18)R ∞
0 (cid:18)R x
0 (cid:18) supt≤τ
sup
h≥0
0
u(τ)
B(τ)R τ
(cid:18)R ∞
t∈(0,∞)(cid:18)Z t
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
b(x)p′
v(x)1−p′(cid:18)Z t
1
p
0
0
x (cid:18) sup
s≤τ
≈ sup
1
q
w(x) dx(cid:19)
u(τ)
B(τ)(cid:19)a(s) ds(cid:19) p′
dx(cid:19)
1
p′ (cid:18)Z ∞
t
1
q
w(y) dy(cid:19)
14
b(x)p′
v(x)1−p′
0
+ sup
+ sup
t∈(0,∞)(cid:18)Z t
x∈(0,∞)(cid:18)Z[x,∞)
x∈(0,∞)(cid:18)Z(0,x]
+(cid:18)Z ∞
+ sup
0
d(cid:18) − sup
A(t)p′
1
t
p′ (cid:18)Z ∞
dx(cid:19)
B(τ)(cid:19)p′
(cid:18)Z τ
t≤τ (cid:18) u(τ)
B(τ)(cid:19)p′
(cid:18)Z τ
t≤τ (cid:18) u(τ)
d(cid:18) − sup
B(τ)(cid:18)Z τ
t→∞(cid:18) sup
u(τ)
lim
t≤τ
1
q
0
0
0
A(y)qw(y) dy(cid:19)
(cid:18)Z y
t (cid:18) sup
v(s)1−p′
b(s)p′
s≤τ
b(s)p′
1
q
1
u(τ)
B(τ)(cid:19)a(s) ds(cid:19)q
w(y) dy(cid:19)
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
ds(cid:19)(cid:19)(cid:19)
v(s)1−p′
p′(cid:19);
ds(cid:19)
0
x
1
1
1
q
1
q
b(s)p′
v(s)1−p′
(ii) If q < p, then
(cid:18)R ∞
0 (cid:18)R x
0 (cid:18) supt≤τ
sup
h≥0
1
q
w(x) dx(cid:19)
0
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
b(x)p′
v(x)1−p′
r
q′
1
p
0
0
0
u(τ)
B(τ)R τ
(cid:18)R ∞
≈ (cid:18)Z ∞
(cid:18)Z t
+(cid:18)Z ∞
(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
+(cid:18)Z ∞
(cid:18)Z(0,x]
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
b(x)p′
0
0
0
0
0
r
q
1
r
dt(cid:19)
r
t
r
p
s≤y
s≤y
u(y)
u(y)
b(t)p′
v(t)1−p′(cid:18)Z ∞
(cid:18)Z z
B(y)(cid:19)a(s) ds(cid:19)q
dx(cid:19)
t (cid:18) sup
B(y)(cid:19)a(s) ds(cid:19) p′
p′(cid:18)Z ∞
v(x)1−p′(cid:18)Z t
w(s) ds(cid:19)
dx(cid:19)
x (cid:18) sup
B(τ)(cid:19)p′
(cid:18)Z τ
t≤τ (cid:18) u(τ)
d(cid:18) −(cid:18) sup
ds(cid:19)(cid:19)(cid:19)(cid:19)
v(s)1−p′
B(τ)(cid:19)p′
(cid:18)Z τ
t≤τ (cid:18) u(τ)
d(cid:18) −(cid:18) sup
v(s)1−p′
A(t)p′
B(τ)(cid:18)Z τ
t→∞(cid:18) sup
p′ (cid:18)Z x
ds(cid:19)(cid:19)(cid:19)(cid:19)
p′(cid:19).
ds(cid:19)
w(z) dz(cid:19)
w(t) dt(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
v(s)1−p′
b(s)p′
b(s)p′
b(s)p′
u(τ)
lim
t≤τ
1
q
0
0
0
0
x
z
1
r
r
1
r
1
r
r
p
r
p
A(x)qw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
Proof. The statement follows by Theorem 3.1 at once if we note that
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ
sup
h≥0
u(τ)
B(τ)R τ
(cid:18)R ∞
0 (cid:18)R x
(cid:18)R ∞
0
1
q
w(x) dx(cid:19)
0
1
p
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
0 (cid:18) supt≤τ
(cid:18)R ∞
B(τ)R τ
h(s)pb(s)−pv(s) ds(cid:19)
u(τ)
0
0
h(y) dy(cid:19)a(t) dt(cid:19)q
1
p
= sup
h≥0
w(x) dx(cid:19)
1
q
.
(cid:3)
Theorem 3.3. Let 1 < p, q < ∞ and b ∈ W(0, ∞) be such that b(t) > 0 for a.e.
W(0, ∞) ∩ C(0, ∞) and a, v, w ∈ W(0, ∞). Moreover, assume that
t ∈ (0, ∞). Assume that u ∈
Denote by
0 <Z ∞
x
v(t)1−p′
dt < ∞
for all
x > 0.
ψ(x) := (cid:18)Z ∞
x
b(t)p′
v1−p′
(t) dt(cid:19)− p′
p′+1
b(x)p′
v1−p′
(x)
and
(i) If p ≤ q, then
Ψ(x) := (cid:18)Z ∞
x
b(t)p′
v1−p′
1
p′+1
.
(t) dt(cid:19)
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ
sup
h≥0
1
q
w(x) dx(cid:19)
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
Ψ(x)−p′
1
p
15
1
q
≈ sup
τ
0
0
0
u(τ)
+ sup
B(τ)R ∞
(cid:18)R ∞
t∈(0,∞)(cid:18)Z t
t∈(0,∞)(cid:18)Z t
x∈(0,∞)(cid:18)Z[x,∞)
x∈(0,∞)(cid:18)Z(0,x]
+ (cid:18)Z ∞
+ (cid:18)Z ∞
+ sup
+ sup
0
0
1
q
w(y) dy(cid:19)
w(y) dy(cid:19)
Ψ(x)−p′
d(cid:18) − sup
A(t)p′
1
t
t
1
s≤τ
s≤τ
u(τ)
u(τ)
B(τ)
B(τ)
x (cid:18) sup
Ψ(τ)2(cid:19)a(s) ds(cid:19) p′
ψ(x)(cid:18)Z t
p′ (cid:18)Z ∞
(cid:18)Z y
ψ(x) dx(cid:19)
t (cid:18) sup
B(τ)(cid:19)p′
Ψ(τ)2p′(cid:18)Z τ
t≤τ (cid:18) u(τ)
B(τ)(cid:19)p′
t≤τ (cid:18) u(τ)
d(cid:18) − sup
t→∞(cid:18) sup
(cid:18)Z x
0 (cid:18) sup
Ψ(τ)2p′(cid:18)Z τ
Ψ(τ)2(cid:18)Z τ
Ψ(τ)2(cid:19)a(t) dt(cid:19)q
p′ (cid:18)Z ∞
dx(cid:19)
Ψ(τ)2(cid:19)a(s) ds(cid:19)q
p′ (cid:18)Z x
ψ(s) ds(cid:19)(cid:19)(cid:19)
ψ(s) ds(cid:19)(cid:19)(cid:19)
Ψ(s)−p′
p′(cid:19)
ψ(s) ds(cid:19)
w(x) dx(cid:19)
Ψ(s)−p′
Ψ(s)−p′
B(τ)
B(τ)
u(τ)
u(τ)
lim
t≤τ
t≤τ
1
q
1
q
;
0
0
0
0
0
1
1
1
A(y)qw(y) dy(cid:19)
ψ(s) ds(cid:19)− 1
p(cid:18)Z ∞
1
q
1
q
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
x
(ii) If q < p, then
(cid:18)R ∞
0 (cid:18)R x
0 (cid:18) supt≤τ
sup
h≥0
1
q
τ
0
0
0
0
1
p
r
q′
u(τ)
Ψ(t)−p′
Ψ(x)−p′
Ψ(x)−p′
w(x) dx(cid:19)
B(τ)R ∞
(cid:18)R ∞
ψ(t)(cid:18)Z ∞
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
ψ(x) dx(cid:19)
ψ(x)(cid:18)Z t
x (cid:18) sup
B(τ)(cid:19)p′
Ψ(τ)2p′(cid:18)Z τ
t≤τ (cid:18) u(τ)
B(τ)(cid:19)p′
t≤τ (cid:18) u(τ)
d(cid:18) −(cid:18) sup
t→∞(cid:18) sup
(cid:18)Z x
0 (cid:18) sup
≈ (cid:18)Z ∞
(cid:18)Z t
+(cid:18)Z ∞
(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
+(cid:18)Z ∞
(cid:18)Z(0,x]
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
ψ(s) ds(cid:19)− 1
p(cid:18)Z ∞
+ (cid:18)Z ∞
d(cid:18) −(cid:18) sup
A(t)p′
B(τ)
B(τ)
B(y)
u(τ)
u(τ)
u(y)
lim
s≤y
t≤τ
t≤τ
1
q
0
0
0
0
0
0
0
t
(cid:18)Z z
t (cid:18) sup
Ψ(y)2(cid:19)a(s) ds(cid:19) p′
s≤y
Ψ(s)−p′
r
q
1
r
dt(cid:19)
r
r
p
1
r
u(y)
B(y)
Ψ(y)2(cid:19)a(s) ds(cid:19)q
p′(cid:18)Z ∞
w(s) ds(cid:19)
dx(cid:19)
p′ (cid:18)Z x
ψ(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
ψ(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
Ψ(s)−p′
w(z) dz(cid:19)
w(t) dt(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
0
z
x
r
r
1
r
r
p
r
p
A(x)qw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
0
Ψ(τ)2p′(cid:18)Z τ
Ψ(τ)2(cid:18)Z τ
Ψ(τ)2(cid:19)a(t) dt(cid:19)q
0
Ψ(s)−p′
ψ(s) ds(cid:19)
w(x) dx(cid:19)
1
q
1
p′(cid:19)
.
Proof. By [20, Corollary 3.5], we have that
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ
sup
h≥0
τ
u(τ)
B(τ)R ∞
(cid:18)R ∞
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
1
p
0
1
q
w(x) dx(cid:19)
16
1
q
w(x) dx(cid:19)
= sup
h≥0
≈ sup
h≥0
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18)R x
(cid:18)R ∞
0
0 (cid:18) supt≤τ
(cid:18)R ∞
0 (cid:18) supt≤τ
(cid:18)R ∞
0 (cid:18) supt≤τ
0
+ (cid:18)R ∞
0 (cid:18)R x
τ
u(τ)
h(y) dy(cid:19)a(t) dt(cid:19)q
B(τ)R ∞
h(s)pb(s)−pv(s) ds(cid:19)
1
p
u(τ)
B(τ)
0
1
p
h(y) dy(cid:19)a(t) dt(cid:19)q
Ψ(τ)2R τ
h(s)pΨ(s)pψ(s)1−p ds(cid:19)
Ψ(τ)2(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
ψ(s) ds(cid:19)
1
q
.
1
p
u(τ)
B(τ)
(cid:18)R ∞
0
1
q
w(x) dx(cid:19)
(i) Let p ≤ q. By Theorem 3.1, (i), we get that
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ
sup
h≥0
1
q
w(x) dx(cid:19)
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
Ψ(x)−p′
1
p
≈ sup
τ
0
0
0
u(τ)
+ sup
B(τ)R ∞
(cid:18)R ∞
t∈(0,∞)(cid:18)Z t
t∈(0,∞)(cid:18)Z t
x∈(0,∞)(cid:18)Z[x,∞)
x∈(0,∞)(cid:18)Z(0,x]
+ (cid:18)Z ∞
+ (cid:18)Z ∞
+ sup
+ sup
0
0
1
q
w(y) dy(cid:19)
w(y) dy(cid:19)
1
q
Ψ(x)−p′
d(cid:18) − sup
A(t)p′
1
t
t
1
s≤τ
s≤τ
u(τ)
u(τ)
B(τ)
B(τ)
x (cid:18) sup
Ψ(τ)2(cid:19)a(s) ds(cid:19) p′
ψ(x)(cid:18)Z t
p′ (cid:18)Z ∞
(cid:18)Z y
t (cid:18) sup
ψ(x) dx(cid:19)
B(τ)(cid:19)p′
Ψ(τ)2p′(cid:18)Z τ
t≤τ (cid:18) u(τ)
B(τ)(cid:19)p′
t≤τ (cid:18) u(τ)
d(cid:18) − sup
t→∞(cid:18) sup
(cid:18)Z x
0 (cid:18) sup
Ψ(τ)2p′(cid:18)Z τ
Ψ(τ)2(cid:18)Z τ
Ψ(τ)2(cid:19)a(t) dt(cid:19)q
p′ (cid:18)Z ∞
dx(cid:19)
Ψ(τ)2(cid:19)a(s) ds(cid:19)q
p′ (cid:18)Z x
ψ(s) ds(cid:19)(cid:19)(cid:19)
ψ(s) ds(cid:19)(cid:19)(cid:19)
Ψ(s)−p′
p′(cid:19)
ψ(s) ds(cid:19)
w(x) dx(cid:19)
Ψ(s)−p′
Ψ(s)−p′
B(τ)
B(τ)
u(τ)
u(τ)
lim
t≤τ
t≤τ
1
q
1
q
;
0
0
0
0
1
1
A(y)qw(y) dy(cid:19)
ψ(s) ds(cid:19)− 1
p(cid:18)Z ∞
0
1
q
1
q
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
x
1
(ii) Let q < p. By Theorem 3.1, (ii), we obtain that
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ
sup
h≥0
u(τ)
B(τ)R ∞
(cid:18)R ∞
Ψ(x)−p′
0
1
q
τ
1
p
r
q′
Ψ(t)−p′
w(x) dx(cid:19)
ψ(t)(cid:18)Z ∞
h(y)b(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pv(s) ds(cid:19)
ψ(x) dx(cid:19)
ψ(x)(cid:18)Z t
x (cid:18) sup
B(τ)(cid:19)p′
Ψ(τ)2p′(cid:18)Z τ
t≤τ (cid:18) u(τ)
B(τ)(cid:19)p′
t≤τ (cid:18) u(τ)
d(cid:18) −(cid:18) sup
d(cid:18) −(cid:18) sup
A(t)p′
B(y)
u(y)
s≤y
0
t
Ψ(x)−p′
0
0
0
≈ (cid:18)Z ∞
(cid:18)Z t
+(cid:18)Z ∞
(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
+(cid:18)Z ∞
(cid:18)Z(0,x]
0
0
0
(cid:18)Z z
t (cid:18) sup
Ψ(y)2(cid:19)a(s) ds(cid:19) p′
s≤y
Ψ(s)−p′
Ψ(τ)2p′(cid:18)Z τ
0
r
q
1
r
dt(cid:19)
r
r
p
1
r
u(y)
B(y)
Ψ(y)2(cid:19)a(s) ds(cid:19)q
p′(cid:18)Z ∞
w(s) ds(cid:19)
dx(cid:19)
p′ (cid:18)Z x
ψ(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
ψ(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
Ψ(s)−p′
w(z) dz(cid:19)
w(t) dt(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
0
z
x
r
r
1
r
r
p
r
p
A(x)qw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
0
+(cid:18)Z ∞
+ (cid:18)Z ∞
0
1
q
lim
A(y)qw(y) dy(cid:19)
ψ(s) ds(cid:19)− 1
p(cid:18)Z ∞
t→∞(cid:18) sup
(cid:18)Z x
0 (cid:18) sup
t≤τ
t≤τ
0
u(τ)
B(τ)
Ψ(s)−p′
Ψ(τ)2(cid:18)Z τ
Ψ(τ)2(cid:19)a(t) dt(cid:19)q
0
ψ(s) ds(cid:19)
w(x) dx(cid:19)
1
q
u(τ)
B(τ)
The proof is completed.
1
p′(cid:19)
.
17
(cid:3)
4. The boundedness of Ru from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing functions
In this section we characterize the boundedness of Ru from Lp(v) into Cesq(w, a) on the cone of monotone
non-increasing functions.
Theorem 4.1. Let 1 < p, q < ∞. Assume that u ∈ W(0, ∞) ∩ C(0, ∞) and a, v, w ∈ W(0, ∞).
(i) If p ≤ q, then
sup
f ∈M+,↓(0,∞)
1
q
w(x) dx(cid:19)
≈ sup
1
p
0
0
0
+ sup
V(x)p′
0 (cid:18)R x
(cid:18)R ∞
(Ru f )(t)a(t) dt(cid:19)q
(cid:18)R ∞
f (s)pv(s) ds(cid:19)
t∈(0,∞)(cid:18)Z t
v(x)(cid:18)Z t
t∈(0,∞)(cid:18)Z t
v(x) dx(cid:19)
x∈(0,∞)(cid:18)Z[x,∞)
d(cid:18) − sup
x∈(0,∞)(cid:18)Z(0,x]
A(t)p′
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
V(x)p′
+ sup
+ sup
t≤τ
0
0
x (cid:18) sup
s≤τ
u(τ)V(τ)−2(cid:19)a(s) ds(cid:19) p′
1
t
p′ (cid:18)Z ∞
(cid:18)Z y
t (cid:18) sup
V(τ)−2p′(cid:18)Z τ
s≤τ
0
u(τ)p′
u(τ)p′
t≤τ
d(cid:18) − sup
t→∞(cid:18) sup
lim
t≤τ
1
q
V(τ)−2p′(cid:18)Z τ
u(τ)V(τ)−2(cid:18)Z τ
0
0
1
1
q
t
1
1
q
dx(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
u(τ)V(τ)−2(cid:19)a(s) ds(cid:19)q
w(y) dy(cid:19)
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
V(s)p′
p′ (cid:18)Z ∞
v(s) ds(cid:19)(cid:19)(cid:19)
w(y) dy(cid:19)
p′(cid:19);
v(s) ds(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)
V(s)p′
V(s)p′
0
x
1
1
1
q
1
q
(ii) If q < p, then
sup
f ∈M+,↓(0,∞)
1
q
w(x) dx(cid:19)
0
0
0
0
1
p
r
q′
0 (cid:18)R x
(cid:18)R ∞
(Ru f )(t)a(t) dt(cid:19)q
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
≈ (cid:18)Z ∞
(cid:18)Z t
v(x) dx(cid:19)
V(x)p′
+(cid:18)Z ∞
(cid:18)Z t
v(x)(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
d(cid:18) −(cid:18) sup
+(cid:18)Z ∞
(cid:18)Z(0,x]
A(t)p′
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
V(x)p′
t≤τ
1
q
0
0
0
0
0
V(t)p′
s≤y
x (cid:18) sup
u(τ)p′
t
s≤y
v(t)(cid:18)Z ∞
(cid:18)Z z
t (cid:18) sup
u(y)V(y)−2(cid:19)a(s) ds(cid:19) p′
V(τ)−2p′(cid:18)Z τ
V(s)p′
0
t≤τ
d(cid:18) −(cid:18) sup
t→∞(cid:18) sup
lim
t≤τ
u(τ)p′
V(τ)−2p′(cid:18)Z τ
u(τ)V(τ)−2(cid:18)Z τ
0
0
V(s)p′
r
q
1
r
r
r
z
r
p
1
r
p′(cid:18)Z ∞
u(y)V(y)−2(cid:19)a(s) ds(cid:19)q
w(s) ds(cid:19)
dx(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
V(s)p′
dt(cid:19)
w(z) dz(cid:19)
w(t) dt(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
p′ (cid:18)Z x
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
p′(cid:19).
v(s) ds(cid:19)
r
p
r
p
0
x
1
r
1
r
xqw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
Proof. By [26, Theorem 3.2] (cf. [20, Theorem 2.3]), we get that
sup
f ∈M+,↓(0,∞)
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ u(τ) f (τ)(cid:19)a(t) dt(cid:19)q
1
q
w(x) dx(cid:19)
1
p
0
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
0 (cid:18) supt≤τ u(τ)V(τ)−2R τ
0 (cid:18)R x
(cid:18)R ∞
0
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
.
1
p
≈ sup
h≥0
By Theorem 3.1, we have that
(i) if p ≤ q, then
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18) supt≤τ u(τ)V(τ)−2R τ
0
sup
h≥0
(cid:18)R ∞
0
h(s)pV(s)−pv(s)1−p ds(cid:19)
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
≈ sup
1
p
0
0
0
+ sup
V(x)p′
V(x)p′
h(s)pV(s)−pv(s)1−p ds(cid:19)
v(x)(cid:18)Z t
v(x) dx(cid:19)
d(cid:18) − sup
A(t)p′
(cid:18)R ∞
t∈(0,∞)(cid:18)Z t
t∈(0,∞)(cid:18)Z t
x∈(0,∞)(cid:18)Z[x,∞)
x∈(0,∞)(cid:18)Z(0,x]
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
+ sup
+ sup
t≤τ
0
x (cid:18) sup
s≤τ
u(τ)V(τ)−2(cid:19)a(s) ds(cid:19) p′
1
t
p′ (cid:18)Z ∞
(cid:18)Z y
t (cid:18) sup
V(τ)−2p′(cid:18)Z τ
s≤τ
0
u(τ)p′
u(τ)p′
t≤τ
d(cid:18) − sup
t→∞(cid:18) sup
lim
t≤τ
1
q
V(τ)−2p′(cid:18)Z τ
u(τ)V(τ)−2(cid:18)Z τ
0
0
1
1
q
t
1
1
q
dx(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
u(τ)V(τ)−2(cid:19)a(s) ds(cid:19)q
w(y) dy(cid:19)
p′ (cid:18)Z x
A(y)qw(y) dy(cid:19)
V(s)p′
p′ (cid:18)Z ∞
w(y) dy(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)
p′(cid:19).
v(s) ds(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)
V(s)p′
V(s)p′
0
x
1
1
1
q
1
q
(ii) if q < p, then
(cid:18)R ∞
0 (cid:18)R x
0 (cid:18) supt≤τ u(τ)V(τ)−2R τ
0
h(y) dy(cid:19)a(t) dt(cid:19)q
w(x) dx(cid:19)
1
q
sup
h≥0
1
p
h(s)pV(s)−pv(s)1−p ds(cid:19)
V(t)p′
V(x)p′
r
q′
0
0
0
0
V(x)p′
(cid:18)R ∞
≈ (cid:18)Z ∞
(cid:18)Z t
+(cid:18)Z ∞
(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
+(cid:18)Z ∞
(cid:18)Z(0,x]
+(cid:18)Z ∞
A(y)qw(y) dy(cid:19)
v(x) dx(cid:19)
v(x)(cid:18)Z t
d(cid:18) −(cid:18) sup
A(t)p′
t≤τ
0
0
0
0
t≤τ
d(cid:18) −(cid:18) sup
t→∞(cid:18) sup
lim
t≤τ
1
q
s≤y
x (cid:18) sup
u(τ)p′
t
s≤y
v(t)(cid:18)Z ∞
(cid:18)Z z
t (cid:18) sup
u(y)V(y)−2(cid:19)a(s) ds(cid:19) p′
V(τ)−2p′(cid:18)Z τ
V(s)p′
0
u(τ)p′
V(τ)−2p′(cid:18)Z τ
u(τ)V(τ)−2(cid:18)Z τ
0
0
V(s)p′
r
q
1
r
r
r
z
r
p
1
r
p′(cid:18)Z ∞
u(y)V(y)−2(cid:19)a(s) ds(cid:19)q
w(s) ds(cid:19)
dx(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
V(s)p′
dt(cid:19)
w(z) dz(cid:19)
w(t) dt(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
p′ (cid:18)Z x
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
p′(cid:19).
v(s) ds(cid:19)
r
p
r
p
0
x
1
r
A(x)qw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
18
1
r
(cid:3)
5. The boundedness of Pu,b from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing functions
19
In this section we characterize the boundedness of weighted Hardy operator Pu,b from Lp(v) into Cesq(w, a) on
the cone of monotone non-increasing functions.
Theorem 5.1. Let 1 < p, q < ∞ and b ∈ W(0, ∞) be such that b(t) > 0 for a.e.
W(0, ∞) ∩ C(0, ∞) and a, v, w ∈ W(0, ∞).
t ∈ (0, ∞). Assume that u ∈
(i) If p ≤ q, then
sup
f ∈M+,↓(0,∞)
≈ sup
1
q
1
0
0
0
1
p
+ sup
w(x) dx(cid:19)
0 (cid:18)R x
(cid:18)R ∞
(Pu,b f )(t)a(t) dt(cid:19)q
(cid:18)R ∞
f (s)pv(s) ds(cid:19)
x∈(0,∞)(cid:18)Z x
0 (cid:18)Z t
a(y)u(y) dy(cid:19)q
w(t) dt(cid:19)
x∈(0,∞)(cid:18)Z ∞
q(cid:18)Z x
0 (cid:18)Z s
w(t) dt(cid:19)
x∈(0,∞)(cid:18)Z ∞
(cid:18)Z t
u(τ) dτ(cid:19)q
x∈(0,∞)(cid:18)Z ∞
q(cid:18)Z x
0 (cid:18)Z x
w(t) dt(cid:19)
+(cid:18)Z ∞
p (cid:18)Z ∞
(cid:18)Z x
v(s) ds(cid:19)− 1
+ sup
+ sup
B(τ)
a(τ)
0
0
0
0
x
x
x
x
s
1
1
p′
v(s) ds(cid:19)
V(s)−p′
1
1
x
q(cid:18)Z ∞
a(y)u(y) dy(cid:19) p′
q(cid:18)Z x
w(t) dt(cid:19)
u(τ) dτ(cid:19) p′
w(x) dx(cid:19)
B(τ)
a(τ)
1
q
;
a(t)u(t) dt(cid:19)q
V(s)−p′
1
p′
1
p′
v(s) ds(cid:19)
v(s) ds(cid:19)
v(s) ds(cid:19)
1
p′
V(s)(cid:19)p′
0 (cid:18) B(s)
V(s)(cid:19)p′
(cid:18) B(s)
(ii) If q < p, then
sup
f ∈M+,↓(0,∞)
1
q
w(x) dx(cid:19)
0
0
0
0
1
p
0 (cid:18)R x
(cid:18)R ∞
(Pu,b f )(t)a(t) dt(cid:19)q
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
≈(cid:18)Z ∞
(cid:18)Z x
0 (cid:18)Z t
a(y)u(y) dy(cid:19)q
+(cid:18)Z ∞
(cid:18)Z ∞
p (cid:18)Z x
w(t) dt(cid:19)
+(cid:18)Z ∞
(cid:18)Z ∞
p (cid:18)Z x
w(t) dt(cid:19)
+(cid:18)Z ∞
(cid:18)Z ∞
w(t)(cid:18)Z t
v(s) ds(cid:19)− 1
p (cid:18)Z ∞
+(cid:18)Z ∞
B(τ)
a(τ)
0
0
0
x
x
x
x
0
0
r
r
1
r
w(x) dx(cid:19)
r
p′ (cid:18)Z x
0
r
v(z) dz(cid:19)
x
0
a(τ)
V(z)−p′
V(z)−p′
p (cid:18)Z ∞
w(t) dt(cid:19)
a(y)u(y) dy(cid:19) p′
0 (cid:18)Z z
V(z)(cid:19)p′
u(τ) dτ(cid:19) p′
0 (cid:18)Z x
(cid:18) B(z)
V(s)(cid:19)p′
q (cid:18)Z x
u(τ) dτ(cid:19)q
0 (cid:18) B(s)
dt(cid:19)
a(y)u(y) dy(cid:19)q
w(x) dx(cid:19)
B(τ)
1
q
.
z
r
r
p′
v(z) dz(cid:19)
v(z) dz(cid:19)
v(s) ds(cid:19)
(cid:18)Z x
0
1
r
a(y)u(y) dy(cid:19)q
w(x) dx(cid:19)
w(x) dx(cid:19)
V(x)(cid:19)p′
q′ (cid:18) B(x)
r
p′
1
r
r
1
r
v(x) dx(cid:19)
Proof. By [26, Theorem 3.1], using Fubini's Theorem, we get that
sup
f ∈M+,↓(0,∞)
0 (cid:18)R x
(cid:18)R ∞
0
(Pu,b f )(t)a(t) dt(cid:19)q
(cid:18)R ∞
f (s)pv(s) ds(cid:19)
0
1
q
w(x) dx(cid:19)
1
p
1
q
20
w(x) dx(cid:19)
1
q
w(x) dx(cid:19)
+ (cid:18)R ∞
0 (cid:18)R x
0
a(t)u(t) dt(cid:19)q
v(s) ds(cid:19)
(cid:18)R ∞
0
1
q
w(x) dx(cid:19)
1
p
=
sup
f ∈M+,↓(0,∞)
≈ sup
h≥0
≈ sup
h≥0
t
τ
0
0
0
0
1
q
1
p
1
p
1
p
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18)R x
(cid:18)R ∞
0 (cid:18)R x
(cid:18)R ∞
f (τ)b(τ) dτ(cid:19)a(t) dt(cid:19)q
0 (cid:18) u(t)
B(t)R t
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
h(y) dy(cid:19)b(τ) dτ(cid:19)a(t) dt(cid:19)q
h(s)pV(s)pv(s)1−p ds(cid:19)
w(x) dx(cid:19)
0 (cid:18) u(t)
0 (cid:18)R ∞
B(t)R t
(cid:18)R ∞
h(y) dy(cid:19)a(t)u(t) dt(cid:19)q
0 (cid:18)R ∞
h(s)pV(s)pv(s)1−p ds(cid:19)
(cid:18)R ∞
0 (cid:18) u(t)
0 (cid:18)R x
(cid:18)R ∞
B(t)R t
(cid:18)R ∞
a(t)u(t) dt(cid:19)q
h(y) dy(cid:19)q(cid:18)R x
0 (cid:18)R ∞
(cid:18)R ∞
(cid:18)R ∞
h(s)pV(s)pv(s)1−p ds(cid:19)
B(t) u(t) dt(cid:19) dy(cid:19)q
h(y)(cid:18)R x
0 (cid:18)R x
(cid:18)R ∞
(cid:18)R ∞
h(s)pB(s)−pV(s)pv(s)1−p ds(cid:19)
h(y)B(y) dy(cid:19)a(t) dt(cid:19)q
h(s)pV(s)pv(s)1−p ds(cid:19)
w(x) dx(cid:19)
w(x) dx(cid:19)
a(t)
1
p
1
p
1
p
0
0
0
0
0
0
y
x
1
q
+ sup
h≥0
≈ sup
h≥0
+ sup
h≥0
1
q
w(x) dx(cid:19)
1
q
0
0
0
0
1
p
+ (cid:18)R ∞
0 (cid:18)R x
w(x) dx(cid:19)
a(t)u(t) dt(cid:19)q
v(s) ds(cid:19)
(cid:18)R ∞
a(t)u(t) dt(cid:19) dy(cid:19)q
h(y)(cid:18)R y
0 (cid:18)R x
(cid:18)R ∞
(cid:18)R ∞
h(s)pV(s)pv(s)1−p ds(cid:19)
a(t)u(t) dt(cid:19)q
(cid:18)R ∞
v(s) ds(cid:19)
w(x) dx(cid:19)
0 (cid:18)R x
1
p
1
q
0
0
0
.
1
q
+ sup
h≥0
+ (cid:18)R ∞
1
q
w(x) dx(cid:19)
1
p
(i) Let p ≤ q. Using the characterizations of weighted Hardy-type inequalities (see, for instance, [37, Section
1]), by [35, Theorem 1.1], we obtain that
sup
f ∈M+,↓(0,∞)
≈ sup
1
q
1
0
0
0
1
p
+ sup
w(x) dx(cid:19)
0 (cid:18)R x
(cid:18)R ∞
(Pu,b f )(t)a(t) dt(cid:19)q
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
x∈(0,∞)(cid:18)Z x
0 (cid:18)Z t
a(y)u(y) dy(cid:19)q
w(t) dt(cid:19)
x∈(0,∞)(cid:18)Z ∞
q(cid:18)Z x
0 (cid:18)Z s
w(t) dt(cid:19)
x∈(0,∞)(cid:18)Z ∞
(cid:18)Z t
u(τ) dτ(cid:19)q
x∈(0,∞)(cid:18)Z ∞
q(cid:18)Z x
0 (cid:18)Z x
w(t) dt(cid:19)
p (cid:18)Z ∞
+(cid:18)Z ∞
(cid:18)Z x
v(s) ds(cid:19)− 1
+ sup
+ sup
B(τ)
a(τ)
0
x
x
x
x
s
1
0
0
0
V(s)−p′
1
1
x
q(cid:18)Z ∞
a(y)u(y) dy(cid:19) p′
q(cid:18)Z x
w(t) dt(cid:19)
u(τ) dτ(cid:19) p′
w(x) dx(cid:19)
B(τ)
a(τ)
1
q
.
a(t)u(t) dt(cid:19)q
1
p′
v(s) ds(cid:19)
V(s)−p′
1
p′
1
p′
v(s) ds(cid:19)
v(s) ds(cid:19)
v(s) ds(cid:19)
1
p′
V(s)(cid:19)p′
0 (cid:18) B(s)
V(s)(cid:19)p′
(cid:18) B(s)
(ii) Let now q < p. Using the characterizations of weighted Hardy-type inequalities (see, for instance, [37,
Section 1]), by [35, Theorem 1.2], we obtain that
21
sup
f ∈M+,↓(0,∞)
1
q
w(x) dx(cid:19)
0
0
0
0
1
p
0 (cid:18)R x
(cid:18)R ∞
(Pu,b f )(t)a(t) dt(cid:19)q
(cid:18)R ∞
f (s)pv(s) ds(cid:19)
≈(cid:18)Z ∞
(cid:18)Z x
0 (cid:18)Z t
a(y)u(y) dy(cid:19)q
+(cid:18)Z ∞
(cid:18)Z ∞
p (cid:18)Z x
w(t) dt(cid:19)
+(cid:18)Z ∞
(cid:18)Z ∞
p (cid:18)Z x
w(t) dt(cid:19)
+(cid:18)Z ∞
(cid:18)Z ∞
w(t)(cid:18)Z t
v(s) ds(cid:19)− 1
+(cid:18)Z ∞
p (cid:18)Z ∞
a(τ)
B(τ)
0
0
0
0
0
x
x
x
x
r
r
1
r
w(x) dx(cid:19)
r
p′ (cid:18)Z x
0
r
v(z) dz(cid:19)
x
0
a(τ)
V(z)−p′
V(z)−p′
p (cid:18)Z ∞
w(t) dt(cid:19)
a(y)u(y) dy(cid:19) p′
0 (cid:18)Z z
V(z)(cid:19)p′
u(τ) dτ(cid:19) p′
0 (cid:18)Z x
(cid:18) B(z)
V(s)(cid:19)p′
q (cid:18)Z x
u(τ) dτ(cid:19)q
0 (cid:18) B(s)
dt(cid:19)
a(y)u(y) dy(cid:19)q
w(x) dx(cid:19)
B(τ)
1
q
.
z
r
r
p′
v(z) dz(cid:19)
v(z) dz(cid:19)
v(s) ds(cid:19)
(cid:18)Z x
0
1
r
a(y)u(y) dy(cid:19)q
w(x) dx(cid:19)
w(x) dx(cid:19)
V(x)(cid:19)p′
q′ (cid:18) B(x)
r
p′
1
r
r
1
r
v(x) dx(cid:19)
The proof is completed.
(cid:3)
6. The boundedness of Tu,b from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing functions
In this section we combine the results from previous two sections to present the characterization of the bound-
edness of Tu,b from Lp(v) into Cesq(w, a) on the cone of monotone non-increasing functions.
Theorem 6.1. Let 1 < p, q < ∞ and b ∈ W(0, ∞) be such that b(t) > 0 for a.e.
W(0, ∞) ∩ C(0, ∞) and a, v, w ∈ W(0, ∞). Moreover, assume that condition (1.1) holds.
t ∈ (0, ∞). Assume that u ∈
(i) If p ≤ q, then
sup
f ∈M+,↓(0,∞)
1
q
w(x) dx(cid:19)
≈ sup
0
0
0
0
1
p
+ sup
V(x)p′
0 (cid:18)R x
(cid:18)R ∞
(Tu,b f )(t)a(t) dt(cid:19)q
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
t∈(0,∞)(cid:18)Z t
v(x)(cid:18)Z t
V(x)p′
t∈(0,∞)(cid:18)Z t
v(x) dx(cid:19)
x∈(0,∞)(cid:18)Z[x,∞)
d(cid:18) − sup
x∈(0,∞)(cid:18)Z(0,x]
A(t)p′
+ (cid:18)Z ∞
x∈(0,∞)(cid:18)Z x
A(y)qw(y) dy(cid:19)
0 (cid:18)Z t
+ sup
+ sup
+ sup
lim
t≤τ
0
0
x (cid:18) sup
s≤τ
1
u(τ)V(τ)−2(cid:19)a(s) ds(cid:19) p′
p′ (cid:18)Z ∞
(cid:18)Z y
t (cid:18) sup
V(τ)−2p′(cid:18)Z τ
s≤τ
0
t
u(τ)p′
0
1
q
t≤τ
u(τ)p′
d(cid:18) − sup
t→∞(cid:18) sup
a(y)¯u(y) dy(cid:19)q
V(τ)−2p′(cid:18)Z τ
u(τ)V(τ)−2(cid:18)Z τ
q(cid:18)Z ∞
w(t) dt(cid:19)
t≤τ
x
1
1
q
1
q
w(y) dy(cid:19)
w(y) dy(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
x
1
q
1
q
1
t
1
dx(cid:19)
p′ (cid:18)Z ∞
u(τ)V(τ)−2(cid:19)a(s) ds(cid:19)q
p′ (cid:18)Z x
V(s)p′
v(s) ds(cid:19)(cid:19)(cid:19)
p′(cid:19)
v(s) ds(cid:19)
v(s) ds(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)
V(s)p′
V(s)−p′
V(s)p′
1
p′
0
0
1
1
22
1
r
V(s)−p′
x
+ sup
+ sup
x∈(0,∞)(cid:18)Z ∞
x∈(0,∞)(cid:18)Z ∞
x∈(0,∞)(cid:18)Z ∞
+ (cid:18)Z ∞
+ sup
x
x
0
1
x
0
a(τ)
B(τ)
w(t) dt(cid:19)
(cid:18)Z t
w(t) dt(cid:19)
p (cid:18)Z ∞
q(cid:18)Z x
0 (cid:18)Z s
¯u(τ) dτ(cid:19)q
q(cid:18)Z x
0 (cid:18)Z x
(cid:18)Z x
s
0
0
1
1
a(y)¯u(y) dy(cid:19) p′
q(cid:18)Z x
w(t) dt(cid:19)
¯u(τ) dτ(cid:19) p′
w(x) dx(cid:19)
B(τ)
a(τ)
a(t)¯u(t) dt(cid:19)q
v(s) ds(cid:19)− 1
V(s)(cid:19)p′
0 (cid:18) B(s)
V(s)(cid:19)p′
(cid:18) B(s)
1
q
;
1
p′
1
p′
v(s) ds(cid:19)
v(s) ds(cid:19)
v(s) ds(cid:19)
1
p′
(ii) If q < p, then
sup
f ∈M+,↓(0,∞)
1
q
w(x) dx(cid:19)
0
0
0
0
0
0
0
0
1
p
r
q′
t≤τ
V(x)p′
0 (cid:18)R x
(cid:18)R ∞
(Tu,b f )(t)a(t) dt(cid:19)q
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
≈ (cid:18)Z ∞
(cid:18)Z t
v(x) dx(cid:19)
V(x)p′
+ (cid:18)Z ∞
(cid:18)Z t
v(x)(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
d(cid:18) −(cid:18) sup
+ (cid:18)Z ∞
(cid:18)Z(0,x]
A(t)p′
+ (cid:18)Z ∞
A(y)qw(y) dy(cid:19)
+ (cid:18)Z ∞
(cid:18)Z x
0 (cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z ∞
w(t) dt(cid:19)
+ (cid:18)Z ∞
(cid:18)Z ∞
w(t) dt(cid:19)
+ (cid:18)Z ∞
(cid:18)Z ∞
w(t)(cid:18)Z t
v(s) ds(cid:19)− 1
+ (cid:18)Z ∞
p (cid:18)Z ∞
B(τ)
lim
0
0
0
0
0
0
0
x
x
x
r
V(t)p′
s≤y
x (cid:18) sup
u(τ)p′
t
s≤y
v(t)(cid:18)Z ∞
(cid:18)Z z
t (cid:18) sup
u(y)V(y)−2(cid:19)a(s) ds(cid:19) p′
V(τ)−2p′(cid:18)Z τ
V(s)p′
0
r
r
x
0
0
1
q
t≤τ
t≤τ
u(τ)p′
V(s)p′
V(z)−p′
V(τ)−2p′(cid:18)Z τ
d(cid:18) −(cid:18) sup
u(τ)V(τ)−2(cid:18)Z τ
t→∞(cid:18) sup
p (cid:18)Z ∞
a(y)¯u(y) dy(cid:19)q
w(t) dt(cid:19)
a(y)¯u(y) dy(cid:19) p′
p (cid:18)Z x
0 (cid:18)Z z
V(z)(cid:19)p′
¯u(τ) dτ(cid:19) p′
p (cid:18)Z x
0 (cid:18)Z x
(cid:18) B(z)
V(s)(cid:19)p′
q (cid:18)Z x
¯u(τ) dτ(cid:19)q
0 (cid:18) B(s)
dt(cid:19)
a(y)¯u(y) dy(cid:19)q
w(x) dx(cid:19)
(cid:18)Z x
V(z)−p′
B(τ)
a(τ)
a(τ)
1
q
0
0
0
.
x
z
r
r
q
1
r
r
p
1
r
w(z) dz(cid:19)
dt(cid:19)
w(t) dt(cid:19)
A(y)qw(y) dy(cid:19)
p′ (cid:18)Z ∞
w(y) dy(cid:19)
r
p
r
p
xqw(x) dx(cid:19)
w(x) dx(cid:19)
1
r
a(y)u(y) dy(cid:19)q
1
r
w(x) dx(cid:19)
u(y)V(y)−2(cid:19)a(s) ds(cid:19)q
w(s) ds(cid:19)
dx(cid:19)
p′(cid:18)Z ∞
z
r
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
V(s)p′
r
r
r
1
x
0
p′ (cid:18)Z x
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
p′(cid:19)
v(s) ds(cid:19)
p′ (cid:18)Z x
v(z) dz(cid:19)
v(z) dz(cid:19)
w(x) dx(cid:19)
v(z) dz(cid:19)
w(x) dx(cid:19)
V(x)(cid:19)p′
q′ (cid:18) B(x)
v(s) ds(cid:19)
r
p′
r
p′
1
r
0
r
1
r
1
r
v(x) dx(cid:19)
Proof. By (1.2), we have that
sup
f ∈M+,↓(0,∞)
(cid:18)R ∞
0 (cid:18)R x
0
(Tu,b f )(t)a(t) dt(cid:19)q
(cid:18)R ∞
f (s)pv(s) ds(cid:19)
0
1
q
w(x) dx(cid:19)
1
p
≈
sup
f ∈M+,↓(0,∞)
1
q
w(x) dx(cid:19)
(cid:18)R ∞
0 (cid:18)R x
0
0
1
p
(Ru f )(t)a(t) dt(cid:19)q
(cid:18)R ∞
f (s)pv(s) ds(cid:19)
(P ¯u,b f )(t)a(t) dt(cid:19)q
(cid:18)R ∞
0 (cid:18)R x
f (s)pv(s) ds(cid:19)
(cid:18)R ∞
0
0
1
q
.
w(x) dx(cid:19)
1
p
+
sup
f ∈M+,↓(0,∞)
It remains to apply Theorems 4.1 and 5.1.
23
(cid:3)
7. The boundedness of Mγ from Λp(v) into Γq(w)
Suppose that f is a measurable a.e. finite function on Rn. Then its non-increasing rearrangement f ∗ is given by
and let f ∗∗ denotes the Hardy-Littlewood maximal function of f ∗, i.e.
f ∗(t) = inf{λ > 0 : {x ∈ Rn : f (x) > λ} ≤ t},
t ∈ (0, ∞),
f ∗∗(t) :=
1
t Z t
0
f ∗(τ) dτ,
t > 0.
Quite many familiar function spaces can be defined by using the non-increasing rearrangement of a function. One
of the most important classes of such spaces are the so-called classical Lorentz spaces.
Let p ∈ (0, ∞) and w ∈ W(0, ∞). Then the classical Lorentz spaces Λp(w) and Γp(w) consist of all measurable
functions f on Rn for which k f kΛp(w) := k f ∗kp,w,(0,∞) < ∞ and k f kΓp(w) := k f ∗∗kp,w,(0,∞) < ∞, respectively. For more
information about the Lorentz Λ and Γ spaces see e.g. [9] and the references therein.
The fractional maximal operator, Mγ, γ ∈ (0, n), is defined at a locally integrable function f on Rn by
(Mγ f )(x) := sup
Q∋x
Qγ/n−1ZQ
f (y) dy,
x ∈ Rn.
It was shown in [10, Theorem 1.1] that
(Mγ f )∗(t) . sup
τ>t
τγ/n−1Z τ
0
f ∗(y) dy . (Mγ f )∗(t)
(7.1)
for every locally integrable function f on Rn and t ∈ (0, ∞), where f (x) := f ∗(ωnxn) and ωn is the volume of the
unit ball in Rn.
The characterization of the boundedness of Mγ between classical Lorentz spaces Λp(v) and Λq(w) was obtained
in [10] for the particular case when 1 < p ≤ q < ∞ and in [36, Theorem 2.10] in the case of more general operators
and for extended range of p and q (For the characteriation of the boundedness of more general fractional maximal
functions between Λp(v) and Λq(w), see [34], and the references therein).
As an application of obtained results, we calculate the norm of the fractional maximal function Mγ from Λp(v)
into Γq(w).
Theorem 7.1. Let 1 < p, q < ∞ and 0 < γ < n. Assume that v, w ∈ W(0, ∞).
(i) If p ≤ q, then
kMγkΛp(v)→Γq(w)
≈ sup
x (cid:18) sup
s≤τ
0
0
+ sup
+ sup
V(x)p′
V(x)p′
t∈(0,∞)(cid:18)Z t
v(x)(cid:18)Z t
t∈(0,∞)(cid:18)Z t
v(x) dx(cid:19)
x∈(0,∞)(cid:18)Z[x,∞)
d(cid:18) − sup
x∈(0,∞)(cid:18)Z(0,x]
d(cid:18) − sup
t p′
+(cid:18)Z ∞
t→∞(cid:18) sup
w(y) dy(cid:19)
x∈(0,∞)(cid:18)Z x
n qw(t) dt(cid:19)
+ sup
+ sup
lim
t≤τ
t≤τ
t≤τ
τ
1
q
t
0
0
γ
1
γ
n p′
τ
τ
τ
t
1
γ
τ
s≤τ
n V(τ)−2(cid:19) ds(cid:19)p′
p′ (cid:18)Z ∞
(cid:18)Z y
t (cid:18) sup
V(τ)−2p′(cid:18)Z τ
V(τ)−2p′(cid:18)Z τ
n V(τ)−2(cid:18)Z τ
n p′
0
0
0
γ
γ
1
t
γ
p′ (cid:18)Z ∞
dx(cid:19)
n V(τ)−2(cid:19) ds(cid:19)q
v(s) ds(cid:19)(cid:19)(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)
p′(cid:19)
v(s) ds(cid:19)
1
V(s)p′
V(s)p′
V(s)p′
1
q
1
q
y−qw(y) dy(cid:19)
y−qw(y) dy(cid:19)
p′ (cid:18)Z x
w(y) dy(cid:19)
p′ (cid:18)Z ∞
y−qw(y) dy(cid:19)
1
q
0
x
1
1
1
q
q(cid:18)Z ∞
x
V(s)−p′
1
p′
v(s) ds(cid:19)
+1)p′
V(t)−p′
x
+ sup
+ sup
x∈(0,∞)(cid:18)Z ∞
x∈(0,∞)(cid:18)Z ∞
x∈(0,∞)(cid:18)Z ∞
+(cid:18)Z ∞
+ sup
x
x
0
1
n
1
γ
0
t( γ
γ
n − x
y−qw(y) dy(cid:19)
n(cid:19)q
(cid:18)t
t−qw(t) dt(cid:19)
p (cid:18)Z ∞
q(cid:18)Z x
t−qw(t) dt(cid:19)
q(cid:18)Z x
0 (cid:18)x
n qw(x) dx(cid:19)
x
0
γ
1
v(s) ds(cid:19)− 1
sp′
0
q(cid:18)Z x
n(cid:19)p′
γ
γ
n − s
1
q
;
1
p′
v(t) dt(cid:19)
V(s)−p′
sp′
V(s)−p′
1
p′
v(s) ds(cid:19)
v(s) ds(cid:19)
1
p′
(ii) If q < p, then
kMγkΛp(v)→Γq(w)
r
q′
0
0
0
0
0
0
1
q
τ
t≤τ
t≤τ
V(x)p′
V(x)p′
≈ (cid:18)Z ∞
(cid:18)Z t
v(x) dx(cid:19)
+(cid:18)Z ∞
(cid:18)Z t
v(x)(cid:18)Z t
+ (cid:18)Z ∞
(cid:18)Z[x,∞)
d(cid:18) −(cid:18) sup
+(cid:18)Z ∞
(cid:18)Z(0,x]
d(cid:18) −(cid:18) sup
t p′
+(cid:18)Z ∞
w(y) dy(cid:19)
t→∞(cid:18) sup
+(cid:18)Z ∞
(cid:18)Z x
n qw(t) dt(cid:19)
+(cid:18)Z ∞
(cid:18)Z ∞
t−qw(t) dt(cid:19)
+(cid:18)Z ∞
(cid:18)Z ∞
t−qw(t) dt(cid:19)
+(cid:18)Z ∞
(cid:18)Z ∞
n(cid:19)q
(cid:18)t
v(s) ds(cid:19)− 1
p (cid:18)Z ∞
+(cid:18)Z ∞
γ
n − x
lim
t≤τ
x
t
0
0
0
0
0
0
x
x
x
γ
γ
γ
0
0
r
r
r
x
0
z( γ
p (cid:18)Z ∞
p (cid:18)Z x
p (cid:18)Z x
0 (cid:18)x
t−qw(t) dt(cid:19)
n qw(x) dx(cid:19)
γ
γ
n − z
n(cid:19)p′
q (cid:18)Z x
0
r
1
q
.
z−qw(z) dz(cid:19)
r
p
r
q
1
r
1
r
dt(cid:19)
t−qw(t) dt(cid:19)
w(x) dx(cid:19)
r
p
1
r
w(y) dy(cid:19)
y−qw(y) dy(cid:19)
r
p
1
r
x−qw(x) dx(cid:19)
V(t)p′
x (cid:18) sup
s≤y
γ
n p′
γ
n p′
τ
γ
n p′
τ
r
s−qw(s) ds(cid:19)
p′ (cid:18)Z x
p′ (cid:18)Z ∞
0
x
r
t
γ
γ
z
0
y
y
s≤y
v(t)(cid:18)Z ∞
(cid:18)Z z
n V(y)−2(cid:19) ds(cid:19)q
t (cid:18) sup
n V(y)−2(cid:19) ds(cid:19) p′
p′(cid:18)Z ∞
dx(cid:19)
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
V(s)p′
v(s) ds(cid:19)(cid:19)(cid:19)(cid:19)
p′(cid:19)
v(s) ds(cid:19)
n qw(x) dx(cid:19)
V(τ)−2p′(cid:18)Z τ
V(τ)−2p′(cid:18)Z τ
V(τ)−2(cid:18)Z τ
v(z) dz(cid:19)
V(z)−p′
V(z)−p′
V(s)p′
V(s)p′
+1)p′
r
p′
x
1
r
0
0
γ
1
n
r
r
p′
1
r
x−qw(x) dx(cid:19)
v(z) dz(cid:19)
V(z)−p′
zp′
sp′
V(s)−p′
r
p′
v(z) dz(cid:19)
v(s) ds(cid:19)
r
q′
1
r
x−qw(x) dx(cid:19)
V(x)−p′
xp′
1
r
v(x) dx(cid:19)
24
(cid:3)
Proof. From inequalities (7.1), we have that
kMγkΛp(v)→Γq(w) ≈
sup
f ∈M+,↓(0,∞)
0 (cid:18)R x
(cid:18)R ∞
0
1
q
x−qw(x) dx(cid:19)
(Tu,b f )(t) dt(cid:19)q
(cid:18)R ∞
f (s)pv(s) ds(cid:19)
0
1
p
with u(τ) = τγ/n and b ≡ 1. Note that
sup
0<t<∞
in this case. So, it remains to apply Theorem 6.1.
u(t)
B(t)Z t
0
b(τ)
u(τ)
dτ < ∞
[1] R. Askey and R. P. Boas Jr., Some integrability theorems for power series with positive coefficients, Mathematical Essays Dedicated to
A. J. Macintyre, Ohio Univ. Press, Athens, Ohio, 1970, pp. 23 -- 32. MR0277956 (43 #3689)
References
25
[2] S. V. Astashkin and L. Maligranda, Structure of Ces`aro function spaces, Indag. Math. (N.S.) 20 (2009), no. 3, 329 -- 379, DOI
10.1016/S0019-3577(10)00002-9. MR2639977 (2011c:46056)
[3]
[4] G. Bennett, Factorizing the classical inequalities, Mem. Amer. Math. Soc. 120 (1996), no. 576, viii+130, DOI 10.1090/memo/0576.
, Structure of Ces`aro function spaces: a survey, Banach Center Publ. 102 (2014), 13 -- 40.
MR1317938 (96h:26020)
[5] G. Bennett and K.- G. Grosse-Erdmann, Weighted Hardy inequalities for decreasing sequences and functions, Math. Ann. 334 (2006),
no. 3, 489 -- 531, DOI 10.1007/s00208-005-0678-7. MR2207873 (2006m:26038)
[6] R. P. Boas Jr., Integrability theorems for trigonometric transforms, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 38,
Springer-Verlag New York Inc., New York, 1967. MR0219973 (36 #3043)
[7]
[8] M. Carro, A. Gogatishvili, J. Martin, and L. Pick, Weighted inequalities involving two Hardy operators with applications to embeddings
, Some integral inequalities related to Hardy's inequality, J. Analyse Math. 23 (1970), 53 -- 63. MR0274685 (43 #447)
of function spaces, J. Operator Theory 59 (2008), no. 2, 309 -- 332. MR2411048 (2009f:26024)
[9] M. Carro, L. Pick, J. Soria, and V. D. Stepanov, On embeddings between classical Lorentz spaces, Math. Inequal. Appl. 4 (2001),
no. 3, 397 -- 428, DOI 10.7153/mia-04-37. MR1841071 (2002d:46026)
[10] A. Cianchi, R. Kerman, B. Opic, and L. Pick, A sharp rearrangement inequality for the fractional maximal operator, Studia Math. 138
(2000), no. 3, 277 -- 284. MR1758860 (2001h:42029)
[11] S. Chen, Y. Cui, H. Hudzik, and B. Sims, Geometric properties related to fixed point theory in some Banach function lattices, Handbook
of metric fixed point theory, Kluwer Acad. Publ., Dordrecht, 2001, pp. 339 -- 389. MR1904283 (2003f:46031)
[12] Y. Cui and R. Płuciennik, Local uniform nonsquareness in Ces`aro sequence spaces, Comment. Math. Prace Mat. 37 (1997), 47 -- 58.
MR1608225 (99b:46025)
[13] Y. Cui and H. Hudzik, Some geometric properties related to fixed point theory in Ces`aro spaces, Collect. Math. 50 (1999), no. 3,
277 -- 288. MR1744077 (2001f:46033)
[14]
, Packing constant for Cesaro sequence spaces, Proceedings of the Third World Congress of Nonlinear Analysts, Part 4 (Cata-
nia, 2000), 2001, pp. 2695 -- 2702, DOI 10.1016/S0362-546X(01)00389-3. MR1972393 (2004c:46033)
[15] Y. Cui, H. Hudzik, and Y. Li, On the Garcia-Falset coefficient in some Banach sequence spaces, Function spaces (Pozna´n, 1998),
Lecture Notes in Pure and Appl. Math., vol. 213, Dekker, New York, 2000, pp. 141 -- 148. MR1772119 (2001h:46009)
[16] M. Cwikel and E. Pustylnik, Weak type interpolation near "endpoint" spaces, J. Funct. Anal. 171 (2000), no. 2, 235 -- 277, DOI
10.1006/jfan.1999.3502. MR1745635 (2001b:46118)
[17] R. Ya. Doktorskii, Reiterative relations of the real interpolation method, Dokl. Akad. Nauk SSSR 321 (1991), no. 2, 241 -- 245
(Russian); English transl., Soviet Math. Dokl. 44 (1992), no. 3, 665 -- 669. MR1153547 (93b:46143)
[18] W. D. Evans and B. Opic, Real interpolation with logarithmic functors and reiteration, Canad. J. Math. 52 (2000), no. 5, 920 -- 960,
DOI 10.4153/CJM-2000-039-2. MR1782334 (2001i:46115)
[19] A. Gogatishvili, M. Johansson, C. A. Okpoti, and L.-E. Persson, Characterisation of embeddings in Lorentz spaces, Bull. Austral.
Math. Soc. 76 (2007), no. 1, 69 -- 92, DOI 10.1017/S0004972700039484. MR2343440 (2008j:46017)
[20] A. Gogatishvili and R.Ch. Mustafayev, Weighted iterated Hardy-type inequalities, Math. Inequal. Appl. 20 (2017), no. 3, 683 -- 728.
MR3653914
[21]
[22] A. Gogatishvili, R. Mustafayev, and T. Unver, Embeddings between weighted Copson and Ces`aro function spaces, Czechoslovak
, Iterated Hardy-type inequalities involving suprema, Math. Inequal. Appl. 20 (2017), no. 4, 901 -- 927. MR3711402
Math. J. 67(142) (2017), no. 4, 1105 -- 1132. MR3736022
[23] A. Gogatishvili, B. Opic, and L. Pick, Weighted inequalities for Hardy-type operators involving suprema, Collect. Math. 57 (2006),
no. 3, 227 -- 255.
[24] A. Gogatishvili and L. Pick, Embeddings and duality theorems for weak classical Lorentz spaces, Canad. Math. Bull. 49 (2006), no. 1,
82 -- 95, DOI 10.4153/CMB-2006-008-3. MR2198721
[25]
, A reduction theorem for supremum operators, J. Comput. Appl. Math. 208 (2007), no. 1, 270 -- 279. MR2347749
(2009a:26013)
[26] A. Gogatishvili and V. D. Stepanov, Reduction theorems for weighted integral inequalities on the cone of monotone functions, Uspekhi
Mat. Nauk 68 (2013), no. 4(412), 3 -- 68, DOI 10.1070/rm2013v068n04abeh004849 (Russian, with Russian summary); English transl.,
Russian Math. Surveys 68 (2013), no. 4, 597 -- 664. MR3154814
[27] K.-G. Grosse-Erdmann, The blocking technique, weighted mean operators and Hardy's inequality, Lecture Notes in Mathematics,
vol. 1679, Springer-Verlag, Berlin, 1998. MR1611898 (99d:26024)
[28] A. A. Jagers, A note on Ces`aro sequence spaces, Nieuw Arch. Wisk. (3) 22 (1974), 113 -- 124. MR0348444 (50 #942)
[29] A. Kami´nska and D. Kubiak, On the dual of Ces`aro function space, Nonlinear Anal. 75 (2012), no. 5, 2760 -- 2773, DOI
10.1016/j.na.2011.11.019. MR2878472 (2012m:46034)
[30] R. Kerman and L. Pick, Optimal Sobolev imbeddings, Forum Math. 18 (2006), no. 4, 535 -- 570, DOI 10.1515/FORUM.2006.028.
MR2254384 (2007g:46052)
[31] M. Krepela, Integral conditions for Hardy-type operators involving suprema, Collect. Math. 68 (2017), no. 1, 21 -- 50, DOI
10.1007/s13348-016-0170-6. MR3591463
[32] G.M. Leibowitz, A note on the Ces`aro sequence spaces, Tamkang J. Math. 2 (1971), 151 -- 157.
[33] V. G. Maz'ja, Sobolev spaces, Springer Series in Soviet Mathematics, Springer-Verlag, Berlin, 1985. Translated from the Russian by
T. O. Shaposhnikova. MR817985
[34] R. Ch. Mustafayev and N. Bilgic¸li, Generalized fractional maximal functions in Lorentz spaces Λ, J. Math. Inequal. 12 (2018), no. 3,
827 -- 851, DOI 10.7153/jmi-2018-12-62. MR3857365
26
[35] R. Oinarov, Two-sided estimates for the norm of some classes of integral operators, Trudy Mat. Inst. Steklov. 204 (1993), no. Issled.
po Teor. Differ. Funktsii Mnogikh Peremen. i ee Prilozh. 16, 240 -- 250 (Russian); English transl., Proc. Steklov Inst. Math. 3 (204)
(1994), 205 -- 214. MR1320028
[36] B. Opic, On boundedness of fractional maximal operators between classical Lorentz spaces, Function spaces, differential operators
and nonlinear analysis (Pudasjarvi, 1999), Acad. Sci. Czech Repub., Prague, 2000, pp. 187 -- 196. MR1755309 (2001g:42043)
[37] B. Opic and A. Kufner, Hardy-type inequalities, Pitman Research Notes in Mathematics Series, vol. 219, Longman Scientific &
Technical, Harlow, 1990. MR1069756 (92b:26028)
[38] L. Pick, Supremum operators and optimal Sobolev inequalities, Function spaces, differential operators and nonlinear analysis
(Pudasjarvi, 1999), Acad. Sci. Czech Repub., Prague, 2000, pp. 207 -- 219. MR1755311 (2000m:46075)
[39]
, Optimal Sobolev embeddings -- old and new, Function spaces, interpolation theory and related topics (Lund, 2000), de Gruyter,
Berlin, 2002, pp. 403 -- 411. MR1943297 (2003j:46054)
[40] Programma van Jaarlijkse Prijsvragen (Annual Problem Section), Nieuw Arch. Wiskd. 16 (1968), 47 -- 51.
[41] D. V. Prokhorov and V. D. Stepanov, On weighted Hardy inequalities in mixed norms, Proc. Steklov Inst. Math. 283 (2013), 149 -- 164.
, Weighted estimates for a class of sublinear operators, Dokl. Akad. Nauk 453 (2013), no. 5, 486 -- 488 (Russian); English
[42]
transl., Dokl. Math. 88 (2013), no. 3, 721 -- 723. MR3203323
[43]
, Estimates for a class of sublinear integral operators, Dokl. Akad. Nauk 456 (2014), no. 6, 645 -- 649 (Russian); English transl.,
Dokl. Math. 89 (2014), no. 3, 372 -- 377. MR3287911
[44] D. V. Prokhorov, On the boundedness of a class of sublinear integral operators, Dokl. Akad. Nauk 92 (2015), no. 2, 602 -- 605 (Russian).
[45] E. Pustylnik, Optimal interpolation in spaces of Lorentz-Zygmund type, J. Anal. Math. 79 (1999), 113 -- 157, DOI 10.1007/BF02788238.
MR1749309 (2001a:46028)
[46] G. `E. Shambilova, Weighted inequalities for a class of quasilinear integral operators on the cone of monotone functions, Sibirsk. Mat.
Zh. 55 (2014), no. 4, 912 -- 936 (Russian, with Russian summary); English transl., Sib. Math. J. 55 (2014), no. 4, 745 -- 767. MR3242605
[47] J.-S. Shiue, On the Ces`aro sequence spaces, Tamkang J. Math. 1 (1970), no. 1, 19 -- 25.
[48]
[49] V. D. Stepanov, The weighted Hardy's inequality for nonincreasing functions, Trans. Amer. Math. Soc. 338 (1993), no. 1, 173 -- 186,
, A note on Ces`aro function space, Tamkang J. Math. 1 (1970), no. 2, 91 -- 95. MR0276751 (43 #2491)
DOI 10.2307/2154450. MR1097171
[50] V. D. Stepanov and G. `E. Shambilova, Weight boundedness of a class of quasilinear operators on the cone of monotone functions,
Dokl. Math. 90 (2014), no. 2, 569 -- 572.
[51] P. W. Sy, W. Y. Zhang, and P. Y. Lee, The dual of Ces`aro function spaces, Glas. Mat. Ser. III 22(42) (1987), no. 1, 103 -- 112 (English,
with Serbo-Croatian summary). MR940098 (89g:46059)
Rza Mustafayev, Department of Mathematics, Faculty of Science, Karamanoglu Mehmetbey University, Karaman, 70100, Turkey
E-mail address: [email protected]
Nevin Bilgic¸ li, Department of Mathematics, Faculty of Science and Arts, Kirikkale University, 71450 Yahsihan, Kirikkale, Turkey
E-mail address: [email protected]
|
1711.06042 | 3 | 1711 | 2019-05-16T09:58:06 | Norms of truncated Toeplitz operators and numerical radii of restricted shifts | [
"math.FA",
"math.CV"
] | This paper gives a new approach to the calculation of the numerical radius of a restricted shift operator by linking it to the norm of a truncated Toeplitz operator (TTO), which can be be calculated by various methods. Further results on the norm of a TTO are derived, and a conjecture on the existence of continuous symbols for compact TTO is resolved. | math.FA | math |
Norms of truncated Toeplitz operators and
numerical radii of restricted shifts
Pamela Gorkin and Jonathan R. Partington
May 17, 2019
Abstract
This paper gives a new approach to the calculation of the numerical
radius of a restricted shift operator by linking it to the norm of a
truncated Toeplitz operator (TTO), which can be calculated by various
methods. Further results on the norm of a TTO are derived, and a
conjecture on the existence of continuous symbols for compact TTO is
resolved.
Keywords: restricted shift, truncated Toeplitz operator, numerical radius,
Hankel operator, Blaschke product
MSC (2010): 47B35, 47A12, 30H10.
1
Introduction
Given a bounded operator T on a Hilbert space the numerical range of T is
defined by
W (T ) = {hT x, xi : kxk = 1}.
For operators on spaces of dimension 2 the elliptical range theorem [28, 34]
tells us that the numerical range of T is elliptical with foci at the eigenvalues
a, b of T and minor axis of length
(cid:0)tr(T ⋆T ) − a2 − b2(cid:1)1/2
.
For operators on spaces of larger dimension, it is usually difficult to describe
the numerical range explicitly, though it can be done in many special cases.
The numerical range plays an important role in applications such as the sta-
bility of linear systems (see, e.g., [2, 13]), in part because it always contains
1
the eigenvalues of the operator T . Even computing the numerical radius of
T , denoted w(T ) and defined by
w(T ) = sup{ω : ω ∈ W (T )},
is difficult but useful; for example, it gives us bounds on the norm of the
operator:
1
2kAk ≤ w(A) ≤ kAk,
and more general estimates for analytic functions of the operator. The au-
thors of this paper were motivated to study the numerical radius partly
because of its possible connections to the following:
Crouzeix conjecture. For every polynomial p and matrix A, we have
kp(A)k ≤ 2 max{p(z) : z ∈ W (A)}.
This question was first stated in 2004 [8] and important recent progress has
been made [9], but at the time of this writing the question remains open.
In this paper we focus on a particular class of operators known as com-
pressions of the shift operator and consider the numerical radius of the
operators in this class, where little is known. Much more is known about
the geometry of the numerical range of these operators and we refer the
reader to [10, 23, 24, 25, 26, 35]. Though the work in these papers may also
be used to solve some of the problems mentioned here, our goal is to add
other methods that may be useful.
We turn to the class of operators of interest. Let L2 denote the Lebesgue
space on the circle T = ∂D, H 2 denote the classical Hardy space on the
open unit disk D, and S : H 2 → H 2 defined by (Sf )(z) = zf (z) the shift
operator. Let P− denote the orthogonal projection of L2 onto L2 ⊖ H 2. An
inner function is a bounded analytic function with radial limits of modulus
1 almost everywhere. Beurling's theorem tells us that the closed, nontrivial,
invariant subspaces for S are precisely those of the form uH 2, with u a
nonconstant inner function. Thus, the nontrivial invariant subspaces for the
adjoint of S, denoted S∗, are precisely the ones of the form Ku = H 2⊖ uH 2.
In this paper, we consider compressions of the shift operator: Let u be an
inner function and Su : Ku → Ku be defined by Su = PKuSKu, where PKu
is the projection of L2 onto Ku. On H 2 we have
PKu(f ) = f − uP+(uf ) = uP−(uf ).
2
If the symbol u of the compression of the shift is a discontinuous inner
function, then w(Su) = 1. Thus, as we study compressions of the shift
operator, we focus on continuous inner functions; that is, we consider finite
Blaschke products as the symbol when we try to determine the numerical
radius of Su.
The first main result in this paper allows us to compute the numerical
radius of SB when B is a finite Blaschke product and all zeros are real.
This gives precise formulas for w(SB) for Blaschke products of degree up
to 4 and an algorithm for computing w(SB) when B has degree greater
than 4. The proof is accomplished by using an algorithm developed by
Foias and Tannenbaum ([14], see also [15] and Section 3) together with a
result that connects the numerical range to the norms of certain closely
related truncated Toeplitz operators. Then, via interpolation, we are able
to produce an algorithm that involves computing the zeros of a special class
of polynomials. In theory, then, the numerical radius is obtained via these
zeros.
Note that every finite matrix A that defines a contraction with spectrum
in the unit disk satisfying rank(I − A∗A) = 1 is unitarily equvalent to an
operator SB for B a finite Blaschke product (see [25]), so that our results
apply in (formally) more general situations.
In Section 4, we discuss more general truncated Toeplitz operators with
symbol ϕ ∈ L2, or Au
ϕ, defined on the dense subset Ku ∩ L∞ of Ku by
Au
ϕ(f ) = PKu(ϕf ), where u is an inner function.
Work in the earlier sections focuses on the study of the distance from (1+az)
to BH ∞ where B is a finite Blaschke product, while the results of later
sections consider the distance to the Sarason algebra, H ∞ + C(T); that
is, we consider kg + B(H ∞ + C(T))k where g ∈ H ∞ + C(T) and B is an
interpolating Blaschke product. We obtain asymptotic distance estimates
for a function f ∈ H ∞ + C to B(H ∞ + C(T) when B is an interpolating
Blaschke product, with more precise results in the case when the Blaschke
product is a thin interpolating Blaschke product. Then we generalize a result
of Ahern and Clark on compactness of truncated Toeplitz operators and
conclude the paper with an example that provides an answer to a question
stated in [7], constructing a compact truncated Toeplitz operator with no
continuous symbol.
3
2 Notation
For B a finite Blaschke product with zeros aj ∈ D, the corresponding repro-
ducing kernels
kaj (z) =
1
1 − ajz
lie in KB, and in fact form a basis for KB when the zeros of B are distinct.
When the zeros are distinct, we obtain an orthonormal basis by applying
the Gram -- Schmidt process to the reproducing kernels. This is called the
Takenaka -- Malmquist basis. In order for the matrix to be upper triangular,
we find the matrix representation with respect to this basis in the reverse
order and we note that we also will need to reorder the zeros of the Blaschke
product. This can also be adjusted for non-distinct zeros and does not affect
the matrix representation we give. In any event, the matrix representing the
compression SB when B is a finite Blaschke product with zeros (aj) can be
written as
aj
(cid:0)Qj−1
k=i+1(−ak)(cid:1)p1 − ai2p1 − aj2
0
if i = j,
if i < j,
if i > j.
(1)
aij =
For ϕ ∈ L∞, let Mϕ : L2 → L2 denote the multiplication operator defined by
Mϕf = ϕf . The Toeplitz operator Tϕ : H 2 → H 2 is defined by Tϕ = P+Mϕ
and the Hankel operator Hϕ : H 2 → L2 ⊖ H 2 is Hϕ = P−Mϕ. It should be
noted that Hϕ = Hψ if and only if ϕ − ψ ∈ H ∞. The Hankel operator Hϕ
is compact if and only if ϕ is in the algebra
H ∞ + C(T) := {f + c : f ∈ H ∞, c ∈ C(T)},
where C(T) is the algebra of continuous functions on the unit circle. This
algebra is often called the Sarason algebra, in honor of D. Sarason who
proved that it is a closed subalgebra of L∞.
By considering operators defined on dense subsets, it is possible to con-
sider Toeplitz, Hankel and multiplication operators defined on L2. In this
spirit, for ϕ ∈ L2 we may define the truncated Toeplitz operator Au
ϕ on the
dense subset Ku ∩ L∞ of Ku by
Au
ϕ(f ) = PKu(ϕf ).
Recent surveys on truncated Toeplitz operators can be found in [7, 19] and
related results appear in [4, 6].
4
3 The norm computation
3.1 Links with Hankel operators and interpolation
Theorem 3.1 below can be used to compute the numerical radius w(T ) of a
bounded operator T . If we maximize Re W (eiθT ) over all θ with 0 ≤ θ < 2π,
we will obtain w(T ) (see also [33]).
Theorem 3.1. [5, Thm. 5, p. 17] Let T be an operator. Then
max{Re λ : λ ∈ W (T )} = lim
a→0+
1
a {kI + aTk − 1} .
We are now able to rephrase the numerical radius computation in terms
of a computation of the norm of an operator, which can be handled by the
algorithm in [14] that we describe in Section 3.2.
We first consider a Blaschke product with distinct zeros. Let a > 0.
We want kI + aSBk = dist((1 + az)/B(z), H ∞), the norm of a truncated
Toeplitz operator, which is also well known to be the norm of the Hankel
operator with symbol (1 + az)/B(z) as in [37, Prop. 2.1].
Similarly, for a general complex t we may consider the norm of the Hankel
operator with symbol (1 + tz)/B(z). Since a finite Blaschke product is
continuous, HB is compact when B is a finite Blaschke product. When the
norm is attained and takes the value γ we have
1 + tz = B(z)g(z) + γh(z),
where h ∈ H ∞ and khk∞ ≤ 1 (in fact it is a Blaschke product). Equivalently,
we can solve the interpolation h(zk) = (1 + tzk)/γ for each k where h ∈ H ∞
with khk∞ ≤ 1 and the zk are the zeros of B.
By Nevanlinna -- Pick theory (see, for example [36]) this is possible if and
only if the matrix M with (j, k) entry
1 − (1 + tzj)(1 + tzk)/γ2
1 − zjzk
(2)
is positive semi-definite. In the case that the zk are real, writing t = teiθ,
this becomes
1 − (1 + t2zjzk + t(zj + zk) cos θ + it(zj − zk) sin θ)/γ2
1 − zjzk
.
5
3.2 The algorithm of Foias and Tannenbaum
We now recall the algorithm from [14] and [15] as applied to the situation de-
scribed above. By restricting to the special case that gives us the numerical
radius, we may make some simplifications, as follows. For the convenience
of the reader we shall mostly use the same notation as in [14].
For ρ > 0, let
and let
Pρ = I −
1
4ρ2 (I + aSB)(I + aS∗
B)
µ∗(z) = 1 − B(z)B(0).
The basic idea of [14] is that Pρ will become singular for various values
of ρ and the largest ρ is the norm of (I + aSB)/2. Combining this with
Theorem 3.1 we will be able to provide the estimates necessary to compute
the numerical radius of SB where B is a finite Blaschke product. We intro-
duce the relevant notation and then recall the main theorem of [14] here for
convenience.
Recalling the notation (x ⊗ y)(w) := hw, yix for x, y, and w in a Hilbert
space H, we note that I − SBS∗
B = µ∗ ⊗ µ∗ on KB, see [39].
Write ν∗(ρ) = P −1
ρ µ∗ and
ν(ρ) := trace(I − SBS⋆
B)P −1
ρ = hν∗(ρ), µ∗i.
The following is the main result of [14].
Theorem 3.2. Let I be an inner function, R a rational function, and let
ρs = max{R(z) : z ∈ σ(SI ) with z = 1} (ρs is the essential spectral radius
of R(SI )). Consider ν(ρ) as a function of ρ in the interval J := (ρs, 1]. If
ν(ρ) is defined on all of J, then we have kR(SI )k = ρs. Otherwise, there
exists an interval (ρ, 1] with ρs < ρ < 1 such that ν(ρ) → ∞ as ρ → ρ+ and
kR(SI )k = ρ.
In the case at hand, I is a finite Blaschke product and the operator we
consider is finite rank so, consequently, ρs = 0. Thus we focus on solving
for ρ. Foias and Tannenbaum give an explicit algorithm to find ρ, which we
recall here.
A computation shows that
1
(cid:18)I −
4ρ2 (cid:0)I + aSB + aS∗
B + a2(SBS∗
B − I) + a2I(cid:1)(cid:19) ν∗(ρ) = µ∗.
6
But
so we have
(I − SBS∗
B)ν∗(ρ) = ν(ρ)µ∗,
1
4ρ2 − a2
B(cid:19) ν∗(ρ) =(cid:18)1 − a2
(cid:18)(cid:18)1 −
Since Bν∗(ρ) ⊥ H 2 we may write Bν∗(ρ) = ν−1z + o(z2). Then
4ρ2(cid:19) I −
a
4ρ2 SB −
a
4ρ2 S∗
4ρ2 ν(ρ)(cid:19) µ∗. (3)
S∗
Bν∗(ρ) = z(ν∗(ρ) − ν∗(ρ)(0)),
SBν∗(ρ) = zν∗(ρ) − Bν−1,
and note that ν∗(ρ)(0) = ν(ρ).
Substituting these into (3), we get
(cid:18)1 −
1
4ρ2 − a2
4ρ2 −
a
4ρ2 z −
a
4ρ2 z(cid:19) ν∗(ρ) +
aB
4ρ2 ν−1 +
= (cid:18)1 − a2
a
4ρ2 zν(ρ)
4ρ2 ν(ρ)(cid:19) µ∗.
Now we see that the coefficient of ν∗(ρ) is 0 (on the unit circle) if
az2
4ρ2 − z(cid:18)1 −
1
4ρ2(cid:19) +
4ρ2 − a2
a and
a
4ρ2 = 0.
Let the roots be z1, z2, so that z1z2 = a
z1 + z2 =
4ρ2 − 1 − a2
a
.
Upon substituting these zeros into (4), we then have for k = 1, 2
(4)
(5)
azkB(zk)
4ρ2
ν−1 +(cid:18) a
4ρ2 + a2
4ρ2 zkµ∗(zk)(cid:19) ν(ρ) = zkµ∗(zk).
We have two unknowns, ν−1 and ν(ρ), and solving for the latter from
these two simultaneous equations gives
ν(ρ)(cid:0)(a + a2z1µ∗(z1))z2B(z2) − (a + a2z2µ∗(z2))z1B(z1)(cid:1)
= 4ρ2(z1µ∗(z1)z2B(z2) − z2µ∗(z2)z1B(z1)).
We may now consider the case that Pρ is singular (and ρ = ρ > ρs), which
happens when
a(z2B(z2) − z1B(z1)) + a2z1z2(µ∗(z1)B(z2) − µ∗(z2)B(z1)) = 0.
7
On using the facts that z1z2 = a/a and µ∗(z) = 1− B(z)B(0), this simplifies
to
(6)
z2B(z2) − z1B(z1) + a(B(z2) − B(z1)) = 0.
This is the key equation for determining z1 and z2, but in the special
case that a is real and the zeros of B are also real, we can go further, since
we have B(z) = B(z) and (6) is simply
Im(z1B(z1)) + a Im B(z1) = 0.
As a → 0, the norm of (I + aSB)/2 → 1/2, so we are interested in values
of ρ near 1/2. Note that for ρ = (1 + δ)/2 where δ is small, and a real, we
see from (5) that
Re z1 = Re z2 =
(1 + δ)2 − 1 − a2
2a
=
2δ + δ2 − a2
2a
.
The equation for z1 = z2 reduces to
Im(zB(z)) = −a Im(B(z)),
and as a → 0, we have Im zB(z) → 0 so zB(z) → ±1.
These formulae will be used extensively in the rest of this section.
Remark 3.3. The Foias -- Tannenbaum approach can be used in principle to
calculate the norm of more general functions of SB, namely, R(SB) for R a
rational function, rather than simply R(z) = 1 + az. The details are more
complicated, but these ideas may shed some light on the Crouzeix conjecture.
3.3 Blaschke products with real zeros
We now apply this to Blaschke products with distinct real zeros and end
this section with several detailed examples. Theorem 3.4 allows us to obtain
the numerical radius of I + aSB from the limit appearing in Theorem 3.1.
Theorem 3.4. Let B be a Blaschke product with distinct zeros z1, . . . , zn ∈
R. Then w(SB) is attained on the real line.
Proof. Suppose that the zeros, z1, . . . , zn with n ≥ 2 are real and lie in the
line segment [a, b]. Rotating the line segment, if necessary, we may assume
that a ≤ b. From the discussion above we know that for t > 0
kI + teiθSBk = k(1 + teiθz) + BH ∞k.
8
We claim that it is enough to show that if we can solve the interpolation
problem f (zk) = 1 + tzk, for small t > 0, and kfk∞ ≤ γ, where
γ = γ(t) = 1 + ct + o(t),
then we can solve h(zk) = 1 + teiθzk with khk∞ ≤ γ + o(t). If this is the
case then we know that for eiθ with 0 < θ < 2π,
k(1 + teiθz) + BH ∞k ≤ kh + (1 + teiθz − h) + BH ∞k ≤ khk ≤ γ + o(t).
Since we take t → 0, this together with Theorem 3.1 yields the result.
We first note that SB can have no eigenvalues on the boundary of W (SB).
Though this is well known, see [11], in this special case, we can give a short
independent proof: If zj is an arbitrary zero of B, we can reorder the zeros
so that zj and zk, with j 6= k, are the first two zeros. The upper-left 2 × 2
corner of the matrix representing SB (see (1)) is then upper triangular and
the (1, 2) entry is m := p1 − zj2p1 − zk2 > 0. The numerical range of
this 2 × 2 block, which is an elliptical disk with foci at zj and zk and minor
axis m, is contained in W (SB). Since the minor axis has nonzero length,
the elliptical disk is nondegenerate and the foci cannot lie on the boundary.
Therefore c > b.
Choose an automorphism ψ of γD with ψ(1) = 1 that takes the point
1 + λt to 1 + eiθλt + O(t2) for λ ∈ [a, b] and all small t. (In fact, the image
of this interval lies on a circular arc that meets γT twice at right angles.)
By the chain rule, the derivative of ψ at the fixed point is the unimodular
constant eiθ. Therefore,
ψ(z) = ψ(1) + ψ′(1)(z − 1) + O(t2) = 1 + eiθ(z − 1) + O(t2).
Thus, for λ ∈ [a, b],
ψ(1 + λt) = 1 + eiθλt + O(t2).
Now consider h := ψ ◦ f . Then khk ≤ γ + o(t) and h(1 + tzk) = 1 +
eiθtzk + O(t2). By adding on another analytic function of norm O(t2), where
the bound depends on n but not on t, we can solve the exact interpolation
problem.
If the zeros of B are not distinct, we may approximate B uniformly by
Blaschke products Bn of the same degree with distinct zeros. In this case,
the numerical radii of SBn converge to the numerical radius of SB and we
obtain the result for general finite Blaschke products.
9
3.4 Degree-2 Blaschke products
We may now use Theorem 3.4 and the algorithm presented in Section 3
to compute the numerical radius of compressions of the shift operator with
Blaschke products with real zeros. We begin with a simple example.
Example 3.5. If B is a degree-2 Blaschke product with real zeros, a1 and
a2, then the numerical radius of SB is
a1 + a2 − a1a2 + 1
2
a1 + a2 + a1a2 − 1
2
(cid:12)(cid:12)(cid:12)(cid:12)
,(cid:12)(cid:12)(cid:12)(cid:12)
(7)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:27) .
max(cid:26)(cid:12)(cid:12)(cid:12)(cid:12)
We are looking for the points at which zB(z) = ±1. In general,
z3 − (a1 + a2)z2 + a1a2z = ±(1 − (a1 + a2)z + a1a2z2).
If the zeros are real and we let z1, z2, z3 denote the three solutions to this
equation, in one case we have
z1 = −1 and Re z2 = Re z3 =
a1 + a2 + 1 − a1a2
2
and in the other case we have
z1 = 1 and Re z2 = Re z3 =
a1 + a2 − 1 + a1a2
2
,
which yields (7).
We can check this result: In the event that the zeros of B are both real,
it follows from the elliptical range theorem that the numerical range is an
elliptical disk with foci at a1 and a2 and major axis of length 1 − a1a2 =
1 − a1a2, in this case. Thus, we obtain the correct result.
In particular, consider
2! .
B(z) = z z − 1
1 − z
2
We find the three solutions to zB(z) = 1 are −1, 3/4 − i√7/4 and 3/4 +
i√7/4, and the solutions to zB(z) = −1 are 1,−1/4 + i√15/4,−1/4 −
i√15/4. Thus we see that w(SB) = 3/4, and indeed one may verify that
k1 + tSBk = 1 +
3
4
t + o(t),
10
as t → 0+. In fact, we know that in this case the numerical radius of SB is
the maximum real value of the ellipse with foci 0 and 1/2 and major axis of
length 1: this is the center of the ellipse plus the length of the semi-major
axis, which is 3/4.
Now, there is an alternative way of calculating k1 + tSBk, namely, to
find the largest γ that makes the Pick matrix singular, see [36, Section 2.1].
Again we take z1 = 0, z2 = 1/2, for which we have seen that the numerical
radius is 3/4. Take t > 0.
The Pick matrix, as given in (2), is
and is singular if
1 − 1/γ2
1 − (1 + t/2)/γ2
1 − (1 + t/2)/γ2
1−(1+t/2)2/γ2
3/4
!
4
3
(γ2 − 1)(γ2 − x2) − (γ2 − x)2 = 0,
where x = 1 + t/2. This gives
γ4 + γ2(−4 + 6x − 4x2) + x2 = 0.
So
γ2 =
2
2 + t + t2 ±p(2 + t + t2)2 − 4(1 + t + t2/4)
and k1 + tSBk = 1 + 3t/4 + O(t2), as expected.
Example 3.6. A second simple example is the Blaschke product B(z) = zn.
In this case, SB is represented by a nilpotent Jordan block of size n with zeros
on the diagonal. The numerical radius is cos(π/(n + 1)), see for example,
[29].
= (1 + t/2) ± t + O(t2),
To show this without referring to known results, we solve zB(z) =
zn+1 = ±1 and take the two values with the largest and smallest real parts.
The cases of n even and n odd need to be considered separately, but we ob-
tain the two sets of points with real part cos(π/(n+1)) and − cos(π/(n+1)),
establishing the result. Since all zeros are real no matter how we rotate, we
see that the numerical range is the closed disk of radius cos(π/(n + 1)).
11
3.5 Degree-3 Blaschke products
Very little is known about the numerical radius of SB for high degree
Blaschke products, but for low degrees (namely degree 3 and degree 4)
we can often compute w(SB) explicitly. Gaaya [17] analyzed the numeri-
cal radius of compressions of the shift operator in the particular case that
. Our techniques also apply to such Blaschke products,
since we may assume that the zero a is real. In this section, we analyze the
numerical radius of all SB with B degree 3 having real zeros.
1−az(cid:17)n
B(z) = (cid:16) z−a
Example 3.7. We now compute the numerical radius of SB when B is a
Blaschke product of degree 3 with real zeros a, b, and c.
In fact, letting
α = a + b + c + abc and β = ab + ac + bc, we see that w(SB) is
max((cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
) .
α +pα2 − 8(β − 1)
4
α −pα2 − 8(β − 1)
4
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
,(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Proof. This can be handled in the same way as the degree-2 case. First
we solve zB(z) = ±1. The equation we need is zB(z) = −1, because the
other has 1 and −1 as a root and therefore the tangent lines do not give the
maximum real part. So we solve
z(z − a)(z − b)(z − c) + (1 − az)(1 − bz)(1 − cz) = 0.
(8)
The left-hand side is
1 − (a + b + c + abc)z + 2(ab + ac + bc)z2
− (a + b + c + abc)z3 + z4 = 1 − αz + 2βz2 − αz3 + z4.
Since all coefficients are real the roots must occur in conjugate pairs, the
roots must be of the form x± iy and u± iv with x, y, u, v real and x2 + y2 =
u2 + v2 = 1. Thus, (8) becomes
(z2 − 2xz + (x2 + y2))(z2 − 2uz + (u2 + v2)) = 0,
and therefore
2(u + x) = a + b + c + abc
and
ab + ac + bc = 1 + 2ux.
So,
2(ux + x2) = (a + b + c + abc)x
12
and thus
2x2 − (a + b + c + abc)x + (ab + ac + bc − 1) = 0.
(9)
Now we can solve the general problem for a degree-3 Blaschke product with
all zeros real:
x =
(a + b + c + abc) ±p(a + b + c + abc)2 − 8(ab + ac + bc − 1)
4
.
(10)
The maximum modulus of the two values of x is the numerical radius of
SB.
We work one particular case in detail. Let B have real zeros at 0, a and
b. Then our computations require us to solve zB(z) = ±1. From (10) we
see that the real parts of the roots satisfy
or
r =
(a + b)r
ab − 1
2
2
+
= 0,
r2 −
a + b ±p(a + b)2 − 8(ab − 1)
4
.
By Theorem 3.4 the numerical range is attained on the real axis. Using the
algorithm from Section 3.2, we know the maximum will occur at
max((cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
a + b +p(a + b)2 − 8(ab − 1)
4
a + b −p(a + b)2 − 8(ab − 1)
4
(11)
We can check that this is correct in particular cases, one of which we do
below.
If a = b, we would like to check that the numerical radius of SB
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
) .
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
,(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
√2 − a2
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
) ,
(12)
w(SB) = max((cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where
In this case,
a
2 −
2
+
a
2
√2 − a2
,(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1 − az(cid:19)2
B(z) = z(cid:18) z − a
1 − az(cid:19)2
zB(z) = z2(cid:18) z − a
.
13
is a composition of two degree-2 Blaschke products. By [26, Theorem 3.6]
(see also [16]) W (SB) is an elliptical disk and the foci are chosen from among
the zeros of B. Writing zB(z) = C(D(z)) with C(0) = D(0) = 0, the zero
of D(z)/z is the one that is not a focus. Therefore, we see that of the
three zeros 0, a, and a of B, the foci must be 0 and a. From (1) the matrix
representation for SB as
a p1 − a2p1 − b2 −bp1 − a2p1 − c2
p1 − b2p1 − c2
0
0
b
0
c
with a = b and c = 0. It follows from the main theorem of [32] that the
length of the minor axis is
(cid:0)tr(A⋆A) − a2 − b2 − c2(cid:1)1/2
which is (cid:0)2(1 − a2)(cid:1)1/2 in this case. So the length of the minor axis is
p2(1 − a2). The center of the ellipse is a/2 and the foci are 0 and a, so the
major axis has length √2 − a2. This agrees with (12).
,
3.6 Degree-4 Blaschke products
There is some work in [22] on numerical ranges of 4 × 4 matrices, although
only in the case of an elliptical disk. The computations are extremely com-
plicated, but here we may take a more transparent approach.
We consider a Blaschke product with real zeros a, b, c, d, where things
are necessarily more complicated. Once again, we must solve
z(−a + z)(−b + z)(−c + z)(−d + z) ± (1 − az)(1 − bz)(1 − cz)(1 − dz) = 0.
For the case
z(−a + z)(−b + z)(−c + z)(−d + z) − (1 − az)(1 − bz)(1 − cz)(1 − dz) = 0,
a computation shows that this is equivalent to
(−1+z)(1−(−1+a+b+c+d+abcd)z+((−1+c)(−1+d)+b(−1+c+d+cd)+
a(−1+c+d+cd+b(1+c+d−cd)))z2−(−1+a+b+c+d+abcd)z3 +z4) = 0.
For
z(−a + z)(−b + z)(−c + z)(−d + z) + (1 − az)(1 − bz)(1 − cz)(1 − dz) = 0,
14
a computation shows that this is equivalent to
(1 + z)(1− (1 + b + c + d− a(−1 + bcd))z + ((1 + c)(1 + d) + b(1 + c + d− cd)+
a(1+c+d−cd−b(−1+c+d+cd)))z2−(1+b+c+d−a(−1+bcd))z3+z4) = 0.
Noting that the techniques used in the solution in Section 3.5 can be
applied here, we see that if we let α denote the negative of the coefficient of
z and β half the coefficient of z2 we obtain:
(1 + z)(1 − αz + 2βz2 − αz3 + z4) = 0.
By the techniques in the previous section
x =
α ±pα2 − 8(β − 1)
4
.
(13)
There is also a standard procedure to solve a quartic (see http://www.sosmath.com/algebra/factor/fac12/fac12.html
for a complete description) and in this case we obtain exact solutions using
this process (or Mathematica).
4 Norms of Truncated Toeplitz Operators
In this section, we look at more general truncated Toeplitz operators.
Recall that for u an inner function Pu denotes the orthogonal projection
of L2 onto Ku. For ϕ ∈ L2 the truncated Toeplitz operator
Au
ϕf := Pu(ϕf ),
f ∈ Ku
is densely defined on K∞
u
compression of the shift operator, Su.
:= H ∞ ∩ Ku. When ϕ(z) = z, we have the
In [18, Corollary 2], the authors provide a general lower bound for kAu
ϕk
for ϕ ∈ L2 and obtain the following as a corollary: If u is an inner function
with zeros accumulating at every point of T and ϕ is a continuous function
on T, then kAu
ϕk = kϕk∞. The authors note that the hypothesis can be
weakened to the following:
Proposition 4.1. Let u be an inner function that is not a finite Blaschke
product. Let ξ be a limit point of the zeros of u. If ϕ ∈ L∞ is continuous
on an open arc containing ξ with ϕ(ξ) = kϕk∞, then kAu
ϕk = kϕk.
15
We present a short proof of their general result in Proposition 4.2 and
show how these same techniques allow the result to be generalized.
Recall that C(T) denotes the algebra of continuous functions on the unit
circle. We let M (H ∞) denote the maximal ideal space of H ∞, or set of
nonzero multiplicative linear functionals with the weak-⋆ topology. Then,
identifying a point z ∈ D with the linear functional that is evaluation at
that point, we may think of D as contained in M (H ∞). Carleson's corona
theorem says that D is dense in M (H ∞). Let
M (H ∞ + C(T)) = M (H ∞) \ D
be the maximal ideal space of the (closed) algebra H ∞ + C(T).
Let Z(u) denote the zeros of an inner function u in M (H ∞) and ZD(u)
the zeros of u in D. If the function f /∈ H ∞, we write f (z) for the Poisson
extension of f evaluated at a point z. Note that the proposition below does
not require u to have zeros in D and can thus be applied readily to inner
functions with a nontrivial singular inner factor.
If u is a bounded harmonic function and v the harmonic conjugate of u
and we let h = eu+iv, then u extends to a continuous function on M (H ∞)
defined by u = log h, see [30, Lemma 4.4]. Thus, for f ∈ L∞ we see that f
is continuous on M (H ∞). There is another way to look at this: Recall that
the maximal ideal space of L∞ is the Shilov boundary of M (H ∞). Then for
f ∈ L∞ and x ∈ M (H ∞) we have
x(f ) = f (x) =Zsupp µx
f dµx,
where f denotes the Gelfand transform of f and supp µx denotes the subset
of M (L∞) that is the support set for the representing measure µx of x. It is
common to write f in place of f . In this way, we may think of the Gelfand
transform of f as a continuous function on M (H ∞), see [31, p. 184]. The
support set of x ∈ M (H ∞) is known to be a weak peak set for H ∞ (see [31,
p. 207]).
Proposition 4.2. Let u be an inner function not invertible in H ∞ + C(T).
For f ∈ L∞, if f (x) = kfk∞ for some x ∈ Z(u), then
dist(f, uH ∞) = kfk∞.
Proof. If x ∈ D, then x is a zero of u and it is clear that
dist(f, uH ∞) ≥ f (x) = kfk∞.
16
Since dist(f, uH ∞) ≤ kfk∞ the result holds. So we may suppose that
x ∈ M (H ∞) \ D and x(f ) = kfk∞. Since dµx is a probability measure, f
must be constant on the support set of x and that constant must be kfk∞.
Therefore,
which completes the proof.
dist(f, uH ∞) ≥ f (x) = kfk∞,
Note that the assumption in Proposition 4.1 implies the assumption in
Proposition 4.2: In Proposition 4.1 we have ϕ ∈ L∞ continuous on an open
arc about ξ for which ϕ(ξ) = kϕk∞ and ξ is a limit point of the zeros of
u, so we may choose x ∈ M (H ∞ + C) in the closure of the zeros of u with
ϕ(x) = kϕk∞.
To see that Proposition 4.2 extends Proposition 4.1, we use Corollary 1
of Garcia and Ross's paper [18] to note that if ϕ ∈ L2, then
sup
{λ∈D:u(λ)=0} ϕ(λ) ≤ kAu
ϕk.
Therefore, if ϕ ∈ L∞ and x ∈ M (H ∞ + C) with x in the closure of the zeros
of u, and ϕ(x) = kϕk∞, then
kϕk∞ = ϕ(x) =
sup
{λ∈D:u(λ)=0} ϕ(λ) ≤ kAu
ϕk ≤ kϕk∞.
We now consider the so-called thin interpolating Blaschke products (de-
fined below). While one can follow the procedure below to obtain estimates
for other interpolating Blaschke products, the estimates will not be as good.
In any event, this gives us information about the essential norm of a Hankel
operator.
Recall that a Blaschke product B is interpolating if the zero sequence
(zn) of B is an interpolating sequence for H ∞; that is, given a bounded
sequence of complex numbers, (wn), there exists f ∈ H ∞ with f (zn) = wn
for all n. Carleson showed that this is equivalent to the existence of δ > 0
such that
.
17
Let
inf
zm − zn
≥ δ.
n Ym6=n(cid:12)(cid:12)(cid:12)(cid:12)
δn := Ym6=n(cid:12)(cid:12)(cid:12)(cid:12)
1 − zmzn(cid:12)(cid:12)(cid:12)(cid:12)
1 − zmzn(cid:12)(cid:12)(cid:12)(cid:12)
zm − zn
If δn → 1 as n → ∞, the interpolating sequence is said to be a thin interpo-
lating sequence. For example, a radial sequence (zn) for which
(1 − zn+1)/(1 − zn) → 0
as n → ∞ is such a sequence. We will apply Earl's theorem (Theorem 4.3)
below to Blaschke products for which the zero sequence is a thin interpo-
lating sequence, [12, Theorem 2]. We isolate Earl's theorem here for easy
reference.
Theorem 4.3 (Earl's Theorem). Suppose that (zn) is an interpolating se-
quence with
inf
n Ym6=n(cid:12)(cid:12)(cid:12)(cid:12)
≥ δ > 0.
zm − zn
1 − zmzn(cid:12)(cid:12)(cid:12)(cid:12)
If (wn) is any bounded sequence of complex numbers, and M is an arbitrary
number greater than
2 − δ2 + 2(1 − δ2)1/2
δ2
sup
n wn,
then there exists an α ∈ R and a Blaschke product B such that
M eiαB(zj) = wj for j = 1, 2, . . . .
From Earl's theorem, we see that if we are given (wn), a bounded se-
quence of complex numbers, we can find g ∈ H ∞ such that g(zn) = wn for
every n and
kgk ≤
2 − δ2 + 2(1 − δ2)1/2
δ2
n wn.
sup
A second result will be useful here as well.
Theorem 4.4. [3, 27] Let f ∈ H ∞ + C(T) and let B be an interpolating
Blaschke product with zeros (zn). Then Bf ∈ H ∞ + C(T) if and only if
f (zn) → 0.
Via the Chang-Marshall theorem, this theorem implies a more general
result for closed subalgebras of L∞ containing H ∞ with u an arbitrary
function in L∞ of norm at most one. We state the version that we will need
below, with a reference to the more general statement.
Theorem 4.5. ([3], Theorem 4.) Let h ∈ H ∞ + C and let u be a function
in H ∞ + C with kuk ≤ 1. If h(1 − u) = 0 on M (H ∞) \ D, then hH ∞[u] ⊆
H ∞ + C.
18
We remind the reader that we identify a function in L∞ with its Gelfand
transform. For f ∈ H ∞ + C, when we evaluate f at a point z ∈ D, we
are evaluating the Poisson extension of the function. In order to state our
results on the disk, we need the following technical lemma.
Lemma 4.6. Let B be an interpolating Blaschke product with zero sequence
(zn) and f ∈ H ∞ + C. Then
max{f (x) : x ∈ Z(B) ∩ M (H ∞ + C)} = lim sup f (zn).
Proof. Using continuity, choose x0 ∈ M (H ∞ + C) to be a point at which
B(x0) = 0 and f (x0) = max{f (x) : x ∈ Z(B) ∩ M (H ∞ + C)}. Let
O = {y ∈ M (H ∞) : f (y)− f (x0) < 1/n}\(cid:0)M (H ∞)\{z : z ≤ 1− 1/n}(cid:1).
Since the Gelfand transform is continuous on M (H ∞), this is an open set.
Since we assume B is interpolating and B(x0) = 0, the point x0 is in the
closure of the zeros of B, [30, p. 83]. Thus there exists zn ∈ D∩O. Therefore,
f (x0) − 1/n ≤ f (zn) ≤ f (x0) + 1/n and zn ≥ 1 − 1/n.
Therefore lim supn f (zn) ≥ f (x0).
Suppose that lim supn f (zn) > f (x0) + α for some α > 0. Then for
all ε < α/2 there exists N (ε) such that supn≥N (ε) f (zn) ≥ f (x0) + α −
ε > f (x0) + α/2. Thus, we find a sequence of points (znN(ε) for which
B(znN(ε) = 0 and
f (znN(ε) > f (x0) + α/2.
Choosing a point x in the closure, we obtain x ∈ M (H ∞ + C) with f (x) >
f (x0), a contradiction.
If B is an interpolating Blaschke product with zeros (zn) we define
δn := inf
.
k>n Ym>n,m6=k(cid:12)(cid:12)(cid:12)(cid:12)
zm − zk
1 − zmzk(cid:12)(cid:12)(cid:12)(cid:12)
Note that δn is a bounded increasing sequence of real numbers and therefore
this sequence converges. We let δ denote this limit.
Theorem 4.7. Let B be an interpolating Blaschke product with zeros (zn)
and let f ∈ H ∞ + C(T). Then
lim sup f (zn) ≤ kf +B(H ∞+C(T))k ≤
2 − δ2 + 2(1 − δ2)1/2
lim sup f (zn).
δ2
19
Proof. For every h ∈ H ∞ + C(T) and x ∈ M (H ∞ + C(T)) with x(B) = 0,
we have
kf + Bhk ≥ x(f ) + x(Bh) = x(f ).
Thus,
kf + B(H ∞ + C(T))k ≥ max{f (x) : x ∈ M (H ∞ + C(T)), x(B) = 0}.
By Proposition 4.6, we obtain the lower inequality.
j=1
For the upper inequality, we note that the conjugate of the function
z−zj
bN (z) := QN
1−zjz lies in H ∞ + C(T) and therefore, writing BN = BbN
we see that BN is the Blaschke product with zero sequence (zn)n>N .
In
particular, δ(BN ) ≥ δN . Now, H ∞ + C(T) is an algebra and 1 = bN bN , so
kf + B(H ∞ + C(T))k = kf + (BbN )(H ∞ + C(T))k = kf + BN (H ∞ + C(T))k.
Choose gN ∈ H ∞, using Earl's theorem, so that gN = f on (zj)j>N and gN
satisfies the norm estimates
kgNk ≤
2 − δ2
N + 2(1 − δ2
N )1/2
δ2
N
n>N wn.
sup
Then f − gN ∈ BN (H ∞ + C(T)) = B(H ∞ + C(T)), by Theorem 4.4. There-
fore,
kf + B(H ∞ + C(T))k = kgN + B(H ∞ + C(T))k
N + 2(1 − δ2
2 − δ2
N )1/2
≤ kgNk ≤
δ2
N
n>N f (zn).
sup
(14)
Now (δn) is an increasing sequence and therefore the constants appearing on
the right-hand side of (14) form a decreasing sequence. Also (supn>N f (zn))
is a decreasing sequence. Taking the limit yields the desired upper bound.
For thin interpolating sequences the result is especially nice. In this case,
δn → 1 implies that δn → 1. Thus we have the following corollary.
Corollary 4.8. Let B be a thin interpolating Blaschke product with zero
sequence (zn) and let f ∈ H ∞ + C(T). Then
kf + B(H ∞ + C(T))k = lim supf (zn).
20
Remark 4.9. Note that in (14) if f (zn) → 0, then f ∈ B(H ∞ + C(T)),
which is the result of Theorem 4.4.
We recall a result of Bessonov [4, Prop. 2.1].
Proposition 4.10. Let u be an inner function and ϕ ∈ H ∞ + C(T). Then
the truncated Toeplitz operator Au
ϕ : Ku → Ku is compact if and only if
ϕ ∈ u(H ∞ + C(T)).
It follows from Theorem 4.4 that if B is an interpolating Blaschke prod-
f is compact if and
uct with zero sequence (zn) and f ∈ H ∞ + C(T), then Au
only if f (zn) → 0.
Ahern and Clark proved the following theorem (see also [20]).
Theorem 4.11. [1, Section 5] Let f be continuous and let u be an inner
function. The operator Au
f is compact if and only if f (eiθ) = 0 for all
eiθ ∈ supp u ∩ T.
We provide a more general result below.
Theorem 4.12. Let f ∈ H ∞ + C(T) and let u be an inner function. The
operator A(un)
is compact for every n ∈ N if and only if
f
lim
z→1− f (z)(1 − u(z)) = 0.
For the other direction, suppose that A(un)
Proof. By our assumption and the corona theorem (which states that D is
dense in M (H ∞)), f (x)(1 − u(x)) = 0 on M (H ∞ + C(T)). By Theo-
rem 4.5, f un ∈ H ∞ + C(T) for every n. Thus, f ∈ un(H ∞ + C(T)) and the
result follows from Proposition 4.10.
is compact for every n. Then,
by Proposition 4.10, f un ∈ H ∞ + C(T) for every n. If x ∈ M (H ∞ + C(T))
with u(x) ≤ r < 1, we know that f = unhn and khnk∞ = kfk. Therefore,
since x ∈ M (H ∞ + C(T)), we have x(f ) = x(u)nx(hn) ≤ x(u)nkfk∞ ≤
rn for every n. Thus x(f ) = 0 if u(x) < 1. Therefore f (x)(1−u(x)) = 0
on M (H ∞ +C(T)). Since D is dense in H ∞ and M (H ∞ +C(T)) = M (H ∞)\
D we have
f
lim
z→1− f (z)(1 − u(z)) = 0.
21
To get the Ahern and Clark result (Theorem 4.11), let f be continuous.
Then f u ∈ H ∞ + C(T) if and only if f vanishes at each discontinuity of u.
This happens if and only if f vanishes at each discontinuity of un. And this,
in turn, happens if and only if f ¯un ∈ H ∞ + C(T) for all n. Thus, Ahern
and Clark's result follows.
In a second survey on truncated Toeplitz operators [7, p. 12], Chalendar,
Ross, and Timotin say that it would be interesting to give an example of
a compact truncated Toeplitz operator with a symbol ψ ∈ u(H ∞ + C(T))
that has no continuous symbol. We present such an example below.
Example 4.13. Let b be an interpolating Blaschke product with zero se-
quence (zn) clustering at every point of the unit circle. Then there exists
f ∈ H ∞ + C(T) with f (zn) → 0 and f (zm) 6= 0 for some m such that Ab
f is
compact and Ab
f has no continuous symbol.
Proof. That such interpolating Blaschke products exist is well known (and,
while they are easy to construct by wrapping around the circle while in-
creasing the modulus at a fast enough rate, it is also clear from the Chang --
Marshall theorem that many of these exist). Choose wn 6= zk for all k such
that ρ(zn, wn) → 0 as n → ∞. If f is the corresponding Blaschke product
with zeros at (wn), then f (zn) → 0 and f (zm) 6= 0 by construction. By
Theorem 4.4 and Proposition 4.10, f ∈ b(H ∞ + C(T)) and Ab
f is compact.
f had a representation with continuous symbol g, we would have
[38, Theorem 3.1] f − g ∈ bH 2 + bH 2. But f (zn) → 0, b(zn) = b(zn) = 0
and therefore g(zn) → 0. But {zn} contains T in its closure and, since g is
continuous, we must have g = 0. Since f (zm) 6= 0, this is impossible.
If Ab
Remark: The proof shows that under these assumptions, if there is a
continuous symbol g, then we must have gb continuous.
Truncated Hankel operators
For an inner function u and a suitable symbol φ the truncated Hankel op-
erator Bu
φ : Ku → zKu = uKu is defined by
φ(f ) = PzKu
Bu
(φf ),
and, as has been observed in [4, Lem. 3.3], we have Bu
uφf for
f ∈ Ku. Since multiplication by u is an isometry from Ku onto uKu,
the study of compactness, boundedness and Schatten-class membership of
truncated Hankel operators generally reduces to that of truncated Toeplitz
operators.
φ(f ) = uAu
22
Thus Example 4.13 provides a function φ ∈ (H ∞ +C(T))∩b(H ∞ +C(T))
φ compact, but with no symbol ψ such that bψ is continuous.
with Bb
Acknowledgements
This work was partially supported by grants from the Simons Foundation
(♯243653 to Pamela Gorkin) and the London Mathematical Society (♯41621
to Jonathan Partington). Since August 2018, PG has been serving as a
Program Director in the Division of Mathematical Sciences at the National
Science Foundation (NSF), USA, and as a component of this position, she
received support from NSF for research, which included work on this pa-
per. Any opinions, findings, and conclusions or recommendations expressed
in this material are those of the authors and do not necessarily reflect the
views of the National Science Foundation.
The authors also wish to thank the referees for their helpful comments.
References
[1] Ahern, P. R. and Clark, D. N., On functions orthogonal to invariant
subspaces. Acta Math. 124 (1970) 191 -- 204.
[2] Allwright, D. J. and Mees, A. I., Stability, extended spaces and numer-
ical ranges. SIAM J. Control Optim. 20 (1982), no. 3, 328 -- 337.
[3] Axler, Sheldon and Gorkin, Pamela, Divisibility in Douglas algebras.
Michigan Math. J. 31 (1984), no. 1, 89 -- 94.
[4] Bessonov, R. V., Fredholmness and compactness of truncated Toeplitz
and Hankel operators. Integral Equations Operator Theory 82 (2015),
no. 4, 451 -- 467.
[5] Bonsall, F.F. and Duncan, J., Numerical ranges of operators on normed
spaces and of elements of normed algebras. London Mathematical So-
ciety Lecture Note Series, 2 Cambridge University Press, London-New
York, 1971.
[6] Camara, M. Cristina and Partington, Jonathan R., Spectral properties
of truncated Toeplitz operators by equivalence after extension. J. Math.
Anal. Appl. 433 (2016), no. 2, 762 -- 784.
23
[7] Chalendar, Isabelle, Fricain, Emmanuel and Timotin, Dan, A survey
of some recent results on truncated Toeplitz operators. Recent progress
on operator theory and approximation in spaces of analytic functions,
59 -- 77, Contemp. Math., 679, Amer. Math. Soc., Providence, RI, 2016.
[8] Crouzeix, Michel, Bounds for analytical functions of matrices. Integral
Equations Operator Theory 48 (2004), no. 4, 461 -- 477.
[9] Crouzeix, M.; Palencia, C., The numerical range is a (1 + √2)-spectral
set. SIAM J. Matrix Anal. Appl. 38 (2017), no. 2, 649 -- 655.
[10] Daepp, Ulrich; Gorkin, Pamela; Shaffer, Andrew and Voss, Karl,
Mobius transformations and Blaschke products: the geometric connec-
tion. Linear Algebra Appl. 516 (2017), 186 -- 211.
[11] Donoghue, William F., Jr., On the numerical range of a bounded oper-
ator. Michigan Math. J., 4 (1957) 261 -- 263.
[12] Earl, J. P., On the interpolation of bounded sequences by bounded
functions. J. London Math. Soc. (2) 2 (1970) 544 -- 548.
[13] Fan, Michael K. H. and Tits, Andr´e L., m-form numerical range and the
computation of the structured singular value. IEEE Trans. Automat.
Control 33 (1988), no. 3, 284 -- 289.
[14] Foias, C. and Tannenbaum, A., On the Nehari problem for a certain
class of L∞-functions appearing in control theory. J. Funct. Anal. 74
(1987), no. 1, 146 -- 159.
[15] Foias C., Tannenbaum A., and Zames, G., Some explicit formulae for
the singular values of certain Hankel operators with factorizable symbol.
SIAM J. Math. Anal. 19 (1988), no. 5, 1081 -- 1089.
[16] Fujimura, Masayo, Inscribed ellipses and Blaschke products. Comput.
Methods Funct. Theory 13 (2013), no. 4, 557 -- 573.
[17] Gaaya, Haykel On the numerical radius of the truncated adjoint shift.
Extracta Math. 25 (2010), no. 2, 165 -- 182.
[18] Garcia, Stephan Ramon, and Ross, William T., The norm of a trun-
cated Toeplitz operator. Hilbert spaces of analytic functions, 59 -- 64,
CRM Proc. Lecture Notes, 51, Amer. Math. Soc., Providence, RI, 2010.
24
[19] Garcia, Stephan Ramon and Ross, William T., Recent progress on trun-
cated Toeplitz operators. Blaschke products and their applications, 275
-- 319, Fields Inst. Commun., 65, Springer, New York, 2013.
[20] Garcia, Stephan Ramon, Ross, William T., and Wogen, Warren R., C*-
algebras generated by truncated Toeplitz operators. Concrete operators,
spectral theory, operators in harmonic analysis and approximation, 181 --
192, Oper. Theory Adv. Appl., 236, Birkhauser/Springer, Basel, 2014.
[21] Garnett, John B., Bounded analytic functions. Pure and Applied Math-
ematics, 96. Academic Press, Inc. [Harcourt Brace Jovanovich, Publish-
ers], New York -- London, 1981.
[22] Gau, Hwa-Long, Elliptic numerical ranges of 4× 4 matrices. Taiwanese
J. Math. 10 (2006), no. 1, 117 -- 128.
[23] Gau, Hwa-Long and Wu, Pei Yuan, Numerical range of S(φ). Linear
and Multilinear Algebra 45 (1998), no. 1, 49 -- 73.
[24] Gau, Hwa-Long and Wu, Pei Yuan, Numerical range and Poncelet prop-
erty. Taiwanese J. Math. 7 (2003), no. 2, 173 -- 193.
[25] Gau, Hwa-Long and Wu, Pei Yuan, Numerical range circumscribed by
two polygons. Linear Algebra Appl. 382 (2004), 155 -- 170.
[26] Gorkin, Pamela and Wagner, Nathan, Ellipses and compositions of fi-
nite Blaschke products. J. Math. Anal. Appl. 445 (2017), no. 2, 1354 --
1366.
[27] Guillory, Carroll, Izuchi, Keiji, and Sarason, Donald, Interpolating
Blaschke products and division in Douglas algebras. Proc. Roy. Irish
Acad. Sect. A 84 (1984), no. 1, 1 -- 7.
[28] Gustafson, Karl E. and Rao, Duggirala K. M., Numerical range. The
field of values of linear operators and matrices. Universitext. Springer-
Verlag, New York, 1997.
[29] Haagerup U. and de la Harpe, P., The numerical radius of a nilpotent
operator on a Hilbert space. Proc. Amer. Math. Soc. 115 (1992), no. 2,
371 -- 379.
[30] Hoffman, Kenneth, Bounded analytic functions and Gleason parts.
Ann. of Math. (2) 86 (1967), 74 -- 111.
25
[31] Hoffman, Kenneth, Banach spaces of analytic functions. Reprint of the
1962 original. Dover Publications, Inc., New York, 1988.
[32] Keeler, Dennis S., Rodman, Leiba, and Spitkovsky, Ilya M., The numer-
ical range of 3 × 3 matrices. Linear Algebra Appl. 252 (1997), 115 -- 139.
[33] Kippenhahn, Rudolf, On the numerical range of a matrix. Translated
from the German by Paul F. Zachlin and Michiel E. Hochstenbach.
Linear Multilinear Algebra 56 (2008), no. 1 -- 2, 185 -- 225.
[34] Li, Chi-Kwong, A simple proof of the elliptical range theorem. Proc.
Amer. Math. Soc. 124 (1996), no. 7, 1985 -- 1986.
[35] Mirman, Boris, UB-matrices and conditions for Poncelet polygon to be
closed. Linear Algebra Appl. 360 (2003), 123 -- 150.
[36] Partington, Jonathan R., Interpolation, identification, and sampling.
London Mathematical Society Monographs. New Series, 17. The
Clarendon Press, Oxford University Press, New York, 1997.
[37] Sarason, D., Generalized interpolation in H ∞. Trans. Amer. Math. Soc.
127 (1967), 179 -- 203.
[38] Sarason, D., Algebraic properties of truncated Toeplitz operators, Oper.
Matrices 1 (2007), no. 4, 491 -- 526.
[39] Sz.-Nagy, B´ela, Foias, Ciprian, Bercovici, Hari and K´erchy, L´aszl´o, Har-
monic analysis of operators on Hilbert space. Second edition. Revised
and enlarged edition. Universitext. Springer, New York, 2010.
Pamela Gorkin
Department of Mathematics, Bucknell University, Lewisburg, PA 17837, U.S.A.
E-mail: [email protected] and
Jonathan R. Partington
School of Mathematics, University of Leeds, Leeds LS2 9JT, U.K.
E-mail: [email protected]
26
|
1906.02503 | 1 | 1906 | 2019-06-06T10:16:09 | Linear perturbations of the Wigner transform and the Weyl quantization | [
"math.FA"
] | We study a class of quadratic time-frequency representations that, roughly speaking, are obtained by linear perturbations of the Wigner transform. They satisfy Moyal's formula by default and share many other properties with the Wigner transform, but in general they do not belong to Cohen's class. We provide a characterization of the intersection of the two classes. To any such time-frequency representation, we associate a pseudodifferential calculus. We investigate the related quantization procedure, study the properties of the pseudodifferential operators, and compare the formalism with that of the Weyl calculus. | math.FA | math |
Linear perturbations of the Wigner transform
and the Weyl quantization
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
Abstract We study a class of quadratic time-frequency representations that, roughly
speaking, are obtained by linear perturbations of the Wigner transform. They satisfy
Moyal's formula by default and share many other properties with the Wigner trans-
form, but in general they do not belong to Cohen's class. We provide a characteriza-
tion of the intersection of the two classes. To any such time-frequency representation,
we associate a pseudodifferential calculus. We investigate the related quantization
procedure, study the properties of the pseudodifferential operators, and compare the
formalism with that of the Weyl calculus.
Key words: Time-frequency analysis, Wigner distribution, Cohen's class, modula-
tion space, pseudodifferential operator, quantization
2010 Mathematics Subject Classification: 42A38,42B35,46F10,46F12,81S30
Dominik Bayer
Universität der Bundeswehr, München
Werner Heisenberg Weg 39
D-85577 Neubiberg, Germany
e-mail: [email protected]
Elena Cordero
Università di Torino, Dipartimento di Matematica,
via Carlo Alberto 10, 10123 Torino, Italy
e-mail: [email protected]
Karlheinz Gröchenig
Faculty of Mathematics, University of Vienna,
Oskar-Morgenstern-Platz 1, A-1090 Vienna, Austria
e-mail: [email protected]
Salvatore Ivan Trapasso
Dipartimento di Scienze Matematiche "G. L. Lagrange", Politecnico di Torino
Corso Duca degli Abruzzi 24, 10129 Torino, Italy
e-mail: [email protected]
1
2
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
1 Introduction
The Wigner transform is a key concept lying at the heart of pseudodifferential op-
erator theory and time-frequency analysis. It was introduced by Wigner [44] as a
quasi-probability distribution in order to extend the phase-space formalism of classi-
cal statistical mechanics to the domain of quantum physics. Subsequently, this line of
thought led to the phase-space formulation of quantum mechanics. In engineering,
the Wigner transform was considered the ideal tool for the simultaneous investigation
of temporal and spectral features of signals, because it enjoys all properties desired
from a good time-frequency representation (except for positivity) [40].
To be precise, the (cross-)Wigner distribution of two signals f , g ∈ L2(Rd) is
defined as
W( f , g)(x, ω) =ÞRd
e−2πiy ·ω f (cid:16)x +
y
2(cid:17) g(cid:16)x −
y
2(cid:17) dy .
(1)
If f = g we write W f (x, ω). This is an example of a quadratic time-frequency
representation, and heuristically W f (x, ω) is interpreted as a measure of the energy
content of the signal f in a "tight" spectral band around ω during a "short" time
interval near x.
The Wigner transform plays a key role in the Weyl quantization and the corre-
sponding pseudodifferential calculus. Quantization is a formalism that associates
a function on phase space (an observable) with an operator on a Hilbert space.
The standard quantization rule in physics is the Weyl correspondence: given the
phase-space observable σ ∈ S ′(R2d) (called symbol in mathematical language) the
corresponding Weyl transform opW(σ) : S(Rd) → S ′(Rd) is (formally) defined by
opW(σ) f (x) =ÞR2d
(2)
e2πi(x−y)·ω σ(cid:16) x + y
2
, ω(cid:17) f (y)dydω.
A formal computation reveals the role of the Wigner transform in this definition,
because
hopW(σ) f , gi = hσ, W(g, f )i,
f , g ∈ S(Rd).
(3)
Whereas the rigorous interpretation of (2) is subtle and requires oscillatory integrals,
the weak formulation (3) is easy to handle and works without problems for distribu-
tional symbols σ. The bracket h f , gi denote the extension to S ′(Rd) × S(Rd) of the
inner product on L2(Rd).
Unfortunately, not all properties which are desired from a time-frequency rep-
resentation are compatible. The Wigner transform is real-valued, but it may take
negative values. This is a serious obstruction to the interpretation of the Wigner trans-
form as a probability distribution or as an energy density of a signal. By Hudson's
theorem [34] only generalized Gaussian functions have positive Wigner transforms.
To obtain time-frequency representations that are positive for all functions, one
takes local averages of the Wigner transform in order to tame the sign oscillations.
This is usually done by convolving W f with a suitable kernel θ. This idea yields a
Linear perturbations of the Wigner transform and the Weyl quantization
3
class of quadratic time-frequency representations, which is called Cohen's class [8].
The time-frequency representations in Cohen's class are parametrized by a kernel
θ ∈ S ′(R2d), and the associated representation is defined as
Qθ ( f , g) ≔ W( f , g) ∗ θ,
f , g ∈ S(Rd).
(4)
Most time-frequency representations proposed so far belong to Cohen's class [7, 33],
and the correspondence between properties of θ and Qθ is well understood [33, 37]. In
many examples Qθ can be interpreted as a perturbation of the Wigner distribution,
while retaining some of its important properties. Thus Cohen's class provides a
unifying framework for the study of time-frequency representations appearing in
signal processing. See for instance [7, 9, 32, 33].
Next, for every time-frequency representation in Cohen's class one can introduce
a quantization rule in analogy to the Weyl quantization (3), namely,
hopθ (σ) f , gi = hσ, Qθ (g, f )i = hσ ∗ θ∗, W(g, f )i,
f , g ∈ S(Rd) ,
(5)
whenever the expressions make sense. Although the new operator opθ (σ) is just a
Weyl operator with the modified symbol σ ∗ θ∗ (whenever defined in S ′(Rd)), the
variety of pseudodifferential calculi given by definition (5) adds flexibility and a new
flavor to the description and analysis of operators.
For example, a first variation of the Wigner transform are the τ-Wigner transforms
Wτ( f , g)(x, ω) =ÞRd
e−2πiy ·ω f (x + τ y)g(x − (1 − τ)y) dy,
f , g ∈ S(Rd) .
(6)
These belong to Cohen's class and possess the kernel θτ ∈ S(R2d) with Fourier
transform
2 )ξ ·η,
(ξ, η) ∈ R2d .
(7)
bθτ (ξ, η) = e−2πi(τ− 1
The corresponding pseudodifferential calculi are the Shubin τ-pseudodifferential
operators opθτ (σ) in formula (5). For the parameter τ = 1/2 this is the Weyl
calculus, for τ = 0 this is the Kohn-Nirenberg calculus. The important Born-Jordan
quantization rule is obtained as an average over τ ∈ [0, 1], see [5] or the textbook
[18].
From (7) we see that the deviation from the Wigner transform is measured by
µ = τ − 1/2. Hence a natural generalization of the τ-Wigner transform is the
replacement of the scalar parameter µ with a matrix expression M = T − (1/2)I,
with T ∈ Rd×d. This gives the family of matrix-Wigner distributions
WM ( f , g)(x, ω) =ÞRd
e−2πiy ·ω f (cid:18)x +(cid:18)M +
1
2
I(cid:19) y(cid:19) g(cid:18)x +(cid:18)M −
1
2
I(cid:19) y(cid:19) dy .
(8)
Again these are members of the Cohen class in (4) with a kernel θ M given by its
Fourier transform
(9)
cθ M (ξ, η) = e−2πiξ ·M η .
4
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
An even more general definition in the spirit of (8) uses an arbitrary linear
mapping of the pair (x, y) ∈ R2d. Let A = (cid:18) A11 A12
A21 A22(cid:19) be an invertible, real-valued
2d × 2d-matrix. We define the matrix-Wigner transform BA of two functions f , g by
BA ( f , g) (x, ω) =ÞRd
e−2πiy ·ω f (A11 x + A12 y) g (A21 x + A22 y)dy .
(10)
Clearly, WM in (8) is a special case by choosing
A = AM =(cid:18) I M + (1/2)I
I M − (1/2)I(cid:19) .
(11)
Once again, every matrix Wigner transform BA is associated with a pseudodif-
ferential calculus or a quantization rule. Given an invertible 2d × 2d -matrix A and
a symbol σ ∈ S ′(R2d), we define the operator σ A by
(cid:10)σ A f , g(cid:11) ≡ hσ, BA (g, f )i ,
f , g ∈ S(Rd) .
(12)
This is then a continuous operator from S(Rd) to S ′(Rd) and the mapping σ → σ A
is a form of quantization similar to the Weyl quantization.
The class of matrix Wigner transforms has already a sizeable history. To our
knowledge they were first introduced in [23] in dimension d = 1 for a different pur-
pose, but the original contribution to the subject went by unnoticed. The first thorough
investigation of the matrix Wigner transforms BA is contained in the unpublished
Ph. D. thesis [1] of the first-named author who studied the general properties of
this class of time-frequency representations and the associated pseudodifferential
operators. Independently, [3] introduced and studied these "Wigner representations
associated with linear transformations of the time-frequency plane" in dimension
d = 1. In [24] matrix Wigner transforms were used for a signal estimation problem.
Recently, in [43] Toft discusses "matrix parametrized pseudo-differential calculi on
modulation spaces", which correspond to the time-frequency representations in (8).
Finally, two of us [11] took up and reworked and streamlined several results of [1] and
determined the precise intersection between matrix Wigner transforms and Cohen's
class.
The goal of this chapter is a systematic survey of the accumulated knowledge
about the matrix Wigner transforms and their pseudodifferential calculi.
In the first part (Section 3) we discuss the general properties of the matrix Wigner
transforms.
(i) We state the main formulas for covariance, the behavior with respect to the
Fourier transform, the analog of Moyal's formulas, and the inversion formula.
(ii) It is well-known that, up to normalization, the ambiguity function and the
short-time Fourier transform are just different versions and names for the Wigner
transform. This is no longer true for the matrix Wigner transforms, so we give precise
conditions on the parametrizing matrix A so that BA can be expressed as a short-time
Fourier transform, up to a phase factor and a change of coordinates.
Linear perturbations of the Wigner transform and the Weyl quantization
5
(iii) Of special importance is the intersection of the class of matrix Wigner
transforms with Cohen's class. After a partical result in [1], it was proved in [11] that
BA belongs to Cohen's class, if and only if A = (cid:16) I M +I /2
I M −I /2(cid:17) for some d × d-matrix
M. Thus in general a matrix Wigner transform does not belong to Cohen's class.
This fact explains why for certain results we have to impose assumptions on the
parametrizing matrix A.
(iv) A further item is the boundedness of the bilinear mapping ( f , g) → BA( f , g)
on various function spaces. These results are quite useful in the analysis of the
mapping properties of the pseudodifferential operators σ A.
In the second part (Section 4) we study the pseudodifferential calculi defined by
the rule (12).
(i) Based on Feichtinger's kernel theorem (see Theorem 1), we first show that every
"reasonable" operator can be represented as a pseudodifferential operator σ A. We
remark that the map (σ, A) 7→ σ A is highly non-injective and, given two matrices
A and B and two symbols σ, ρ, we obtain formulas characterizing the condition
σ A = ρB.
(ii) A large section is devoted to the mapping properties of the pseudodiffer-
ential operator σ A on various function spaces, in particular on L p-spaces and on
modulation spaces.
(iii) Finally, we extend the boundedness results for symbols in the Sjöstrand class
to those pseudodifferential operators for which BA is in Cohen's class.
For most results we will include proofs, but we will omit those proofs that only
require a formal computation. We hope that the self-contained and comprehensive
presentation of matrix Wigner distributions and their pseudodifferential calculi will
offer some added value when compared with the focussed, individual publications.
2 Preliminaries
Notation. We define t2 = t · t, for t ∈ Rd, and x · y is the scalar product on Rd. The
Schwartz class is denoted by S(Rd), the space of temperate distributions by S ′(Rd).
The brackets h f , gi denote the extension to S ′(Rd) × S(Rd) of the inner product
h f , gi =∫ f (t)g(t)dt on L2(Rd). The conjugate exponent p′ of p ∈ [1, ∞] is defined
by 1/p + 1/p′ = 1. The symbol .λ means that the underlying inequality holds up to
a positive constant factor C = C(λ) > 0 that depends on the parameter λ:
f . g ⇒ ∃C > 0 :
f ≤ Cg.
If f . g and g . f we write f ≍ g.
The Fourier transform of a function f ∈ S(Rd) is normalized as
F f (ω) =ÞRd
e−2πix ·ω f (x) dx,
ω ∈ Rd.
6
as
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
For any x, ω ∈ Rd, the modulation Mω and translation Tx operators are defined
Mω f (t) = e2πit ·ω f (t) ,
Tx f (t) = f (t − x) .
Their composition π(x, ω) = MωTx is called a time-frequency shift.
Given a complex-valued function f on Rd, the involution f ∗ is defined as
f ∗(t) ≔ f (−t),
t ∈ Rd .
The short-time Fourier transform (STFT) of a signal f ∈ S ′(Rd) with respect to
the window function g ∈ S(Rd) is defined as
Vg f (x, ω) = h f , π(x, ω)gi = F ( f · Tx g)(ω) =ÞRd
f (y) g(y − x) e−2πiy ·ω dy. (13)
The group of invertible, real-valued 2d × 2d matrices is denoted by
GL (2d, R) =(cid:8)M ∈ R2d×2d det M , 0(cid:9) ,
and we denote the transpose of an inverse matrix by
M# ≡ (M −1)⊤
= (M ⊤)−1,
M ∈ GL (2d, R) .
Let J denote the canonical symplectic matrix in R2d, namely
−Id 0d(cid:19) .
J =(cid:18) 0d Id
Observe that, for z = (z1, z2) ∈ R2d, we have J z = J (z1, z2) = (z2, −z1) , J−1 z =
J−1 (z1, z2) = (−z2, z1) = −J z, and J2 = −I2d×2d.
2.1 Function spaces
Recall that C0(Rd) denotes the class of continuous functions on Rd vanishing at
infinity.
Modulation spaces. Fix a non-zero window g ∈ S(Rd) and 1 ≤ p, q ≤ ∞.
(i) The modulation space M p,q (Rd) consists of all temperate distributions f ∈
S ′(Rd) such that Vg f ∈ L p,q(R2d) (mixed-norm Lebesgue space). The norm on
M p,q is
k f kM p, q = kVg f kL p, q = ÞRd (cid:18)ÞRd
Vg f (x, ω) p dx(cid:19) q/p
dω!1/q
,
(with obvious modifications for p = ∞ or q = ∞). If p = q, we write M p instead of
M p, p.
Linear perturbations of the Wigner transform and the Weyl quantization
7
(ii) The modulation space W(F L p, Lq)(Rd) can be defined as the space of distribu-
tions f ∈ S ′(Rd) such that
k f kW (F L p, L q )(Rd ) := ÞRd (cid:18)ÞRd
Vg f (x, ω) p dω(cid:19) q/p
dx! 1/q
< ∞
(with obvious modifications for p = ∞ or q = ∞). It can also be viewed as a
special case of Wiener amalgam spaces [21]. Though their inventor H.G. Feichtinger
nowadays suggests to call it modulation rather than Wiener amalgam space, since its
definition involves the mixed-Lebesgue norm of the short-time Fourier transform of
f .
The so-called fundamental identity of time-frequency analysis
Vg f (x, ω) = e−2πix ·ωVg f (ω, −x)
(14)
implies that
k f kM p, q =(cid:18)ÞRd
k f Tω gkq
F L p (ω) dω(cid:19)1/q
= k f kW (F L p, L q ).
Hence the W(F L p, Lq) spaces under our consideration are simply the image via the
Fourier transform of the M p,q spaces:
F (M p,q) = W(F L p, Lq).
(15)
The spaces M p,q(Rd) and W(F L p, Lq)(Rd) are Banach spaces. Further, their
definition is independent of the choice of the window g. The reader may find a
comprehensive discussion of these function spaces in [21, 20, 26]. In particular, for
p = q, we obtain
M p(Rd) = W(F L p, L p)(Rd),
1 ≤ p ≤ ∞.
Recall that the class of admissible windows for computing the modulation space
norm can be extended to M1 \ {0} (cf. [26, Thm. 11.3.7] and [15, Thm. 2.2]).
For p = q = 2, we have
M2(Rd) = L2(Rd),
the Hilbert space of square-integrable functions. Among the properties of modulation
spaces we list the following results.
Lemma 1 For 1 ≤ p, q, p1, q1, p2, q2 ≤ ∞, we have
(i) M p1,q1 (Rd) ֒→ M p2,q2 (Rd), if p1 ≤ p2 and q1 ≤ q2.
(ii) If 1 ≤ p, q < ∞, then (M p,q(Rd))′ = M p′,q′
(Rd).
8
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
2.2 Basic Properties of M 1
Here we collect the main properties of the modulation space M1(Rd), also known
as the Feichtinger algebra [19]. By Lemma 1 (ii) we infer its dual space
(M1(Rd))′
= M ∞(Rd)
and the inclusions M1(Rd) ←֓ M p,q (Rd), for every 1 ≤ p, q ≤ ∞. In particular,
M1(Rd) ֒→ L2(Rd) ֒→ M ∞(Rd).
There are plenty of equivalent characterizations for the Feichtinger algebra, we refer
the interested reader to cf. [35].
dow g ∈ M1(Rd), and different windows yield equivalent norms.
Proposition 1 (i) (cid:0)M1(Rd), k·k M 1(cid:1) is a Banach space for any fixed non-zero win-
(ii) (cid:0)M1(Rd), k·k M 1(cid:1) is a time-frequency homogeneous Banach space: for any
z ∈ R2d, f ∈ M1(Rd), one has π (z) f ∈ M1(Rd) and kπ (z) f k M 1 = k f k M 1.
In particular, it is the smallest time-frequency homogeneous Banach space con-
taining the Gaussian function.
(iii) The Schwartz class S(Rd) is a subset of M1(Rd) and L2(Rd) is the completion
of M1(Rd) with respect to k·k L2 norm.
(iv) M1(Rd) is invariant under the Fourier transform, i.e. for any f ∈ M1(Rd) one
has F f ∈ M1(Rd) and kF f k M 1 = k f k M 1.
Recall that the tensor product of two functions f , g : Rd → C it is defined as
f ⊗ g : R2d → C : (x, y) 7→ f ⊗ g (x, y) = f (x) g (y) .
Clearly the tensor product ⊗ is a bilinear bounded mapping from L2(Rd) × L2(Rd)
into L2(cid:0)R2d(cid:1). Feichtinger algebra is well behaved under tensor products and we list
a few results.
Proposition 2 (i) The tensor product
⊗ : M1(Rd) × M1(Rd) → M1(cid:16)R2d(cid:17)
is a bilinear bounded operator.
(ii) M1 enjoys the tensor factorization property1: the space M1(R2d) consists of all
functions of the form
f =n∈N
gn ⊗ hn,
where {gn }, {hn} are (sequences of) functions in M1(Rd) such that
1 That is,
where the symbolb⊗ denotes the projective tensor product (see [35] for the details).
M 1(R2d ) ≃ M 1(Rd )b⊗ M 1(Rd ),
Linear perturbations of the Wigner transform and the Weyl quantization
9
n∈N
kgn k M 1 khn k M 1 < ∞.
(iii) The tensor product is well defined on M ∞: for any f , g ∈ M ∞(Rd), f ⊗ g is the
unique element of M ∞(R2d) such that
h f ⊗ g, φ1 ⊗ φ2i ≡ h f , φ1i hg, φ2i ,
∀φ1, φ2 ∈ M1(Rd).
Just as the (temperate) distributions are related to the Schwartz kernel theorem,
there is an important kernel theorem in the context of time-frequency analysis, the
Feichtinger kernel theorem [22].
Theorem 1 (i) Every distribution k ∈ M ∞(R2d) defines a bounded linear operator
Tk : M1(Rd) → M ∞(Rd) according to
hTk f , gi = hk, g ⊗ f i,
∀ f , g ∈ M1(Rd),
(ii) For any bounded operator T : M1(Rd) → M ∞(Rd) there exists a unique kernel
with kTk k M 1→M ∞ ≤ kk k M ∞.
kT ∈ M ∞(R2d) such that
hT f , gi = hkT , g ⊗ f i,
∀ f , g ∈ M1(Rd).
The proofs of the aforementioned results can be found in the cited references.
We wish to highlight the comprehensive survey [35], where the properties of the
Feichtinger algebra are explored in full generality.
2.3 Bilinear coordinate transformations
Let us summarize the properties of bilinear coordinate transformations in the time-
frequency plane. Given a matrix A = (cid:18) A11 A12
A21 A22(cid:19) ∈ R2d×2d with d × d blocks
Ai j ∈ Rd×d, i, j = 1, 2, we use the symbol T A to denote the transformation acting on
a function F : R2d → C as
T AF (x, y) = F(cid:18)A(cid:18) x
y(cid:19)(cid:19)
= F (A11 x + A12 y, A21 x + A22 y) .
The following lemma collects elementary facts on such transformations.
Lemma 2 (i) For any A, B ∈ R2d×2d we have T ATB = TB A.
(ii) If A ∈ GL (2d, R), the transformation T A is a topological isomorphism on
L2(R2d) with inverse T−1
A
= T
A−1 and adjoint T∗
A
(iii) If A ∈ GL (2d, R), the transformation T A is an isomorphism on M1(R2d).
= det A−1 T
A−1 .
10
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
(iv) For any A ∈ GL (2d, R), f ∈ L2(Rd), x, ω ∈ Rd:
T ATx f = TA−1 x
T A f ,
T AMω f = MA⊤ω T A f .
Proof The only non-trivial issue is the continuity on M1. By [10, Prop. 3.1] we have
kT AF k M 1 ≤ C det A−1(cid:0)det(cid:0)I + A⊤ A(cid:1)(cid:1)1/2
for some constant C > 0.
kF k M 1 ,
(cid:3)
Two special transformations deserve a separate notation. The first is the flip operator
F (x, y) ≡ T I F (x, y) = F (y, x) ,
while the other one is the reflection operator:
I =(cid:18) 0d Id
Id 0d(cid:19) ∈ GL (2d, R) ,
IF (x, y) ≡ T−I F (x, y) = F (−x, −y) .
Sometimes we will also write I = −I ∈ GL (2d, R).
2.4 Partial Fourier transforms
Given F ∈ L1(R2d), we use the symbols F1 and F2 to denote the partial Fourier
transforms
F1F (ξ, y) =cFy (ξ) =ÞRd
F2F (x, ω) =cFx (ω) =ÞRd
e−2πiξ ·t F (t, y) dt,
ξ, y ∈ Rd
e−2πiω·t F (x, t) dt,
x, ω ∈ Rd,
Fx (y) = F (x, y) ,
Fy (x) = F (x, y)
x, y ∈ Rd
are the sections of F at fixed x and y, respectively. The definition is well-posed
whereb· denotes the Fourier transform on L1(Rd) and
thanks to Fubini's theorem, which implies that Fx ∈ L1(cid:16)Rd
Fy ∈ L1(cid:0)Rd
y(cid:17) for a.e. x ∈ Rd and
x(cid:1) for a.e. y ∈ Rd, thus F1F and F2F are indeed well defined. The Fourier
transform F is therefore related to the partial Fourier transforms as
F = F1F2 = F2F1.
Using Plancherel's theorem and properties of modulation spaces (Proposition 1, item
(iv)), the following extension of the partial Fourier transform is routine.
Linear perturbations of the Wigner transform and the Weyl quantization
11
Lemma 3 (i) The partial Fourier transform F2 is a unitary operator on L2(R2d).
In particular,
2 F (x, y) = F −1
F ∗
2 F (x, y) = F2F (x, −y) = TI2 F2F (x, y) ,
where I2 =(cid:18) I 0
0 −I(cid:19).
M ∞(R2d).
(ii) The partial Fourier transform F2 is an isomorphism on M1(R2d) and on
3 Matrix-Wigner distributions
Let us define the main characters of this survey.
Definition 1 Let A = (cid:18) A11 A12
A21 A22(cid:19) ∈ GL (2d, R) . The time-frequency distribution of
Wigner type for f and g associated with A (in short: matrix-Wigner distribution,
MWD) is defined for suitable functions f , g as
BA ( f , g) (x, ω) = F2T A ( f ⊗ g) (x, ω) .
(16)
When g = f , we write BA f for BA ( f , f ).
Explicitly, BA is given by
BA ( f , g) (x, ω) =ÞRd
e−2πiω·y f (A11 x + A12 y) g (A21 x + A22 y)dy.
This definition is meaningful on many function spaces. A first result is for the
triple (M1, L2, M ∞).
Proposition 3 Assume A ∈ GL (2d, R).
(i) If f , g ∈ L2(Rd), then BA( f , g) ∈ L2(R2d) and the mapping BA : L2(Rd) ×
L2(Rd) → L2(R2d) is continuous. Furthermore, span(cid:8)BA( f , g) f , g ∈ L2(Rd)(cid:9)
(ii) If f , g ∈ M1(Rd), then BA( f , g) ∈ M1(R2d) and the mapping BA : M1(Rd) ×
is a dense subset of L2(R2d).
M1(Rd) → M1(R2d) is continuous.
(iii) If f , g ∈ M ∞(Rd), then BA( f , g) ∈ M ∞(R2d) and the mapping BA : M ∞(Rd) ×
M ∞(Rd) → M ∞(R2d) is continuous.
The standard time-frequency representations covered within this framework in-
clude for instance:
• the short-time Fourier transform:
Vg f (x, ω) =ÞRd
e−2πiω·y f (y) g (y − x)dy = BAST ( f , g) (x, ω) ,
(17)
12
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
where
x
AST =(cid:18) 0 I
−I I(cid:19) ;
2(cid:17) g(cid:16) y −
e−2πiω·y f (cid:16)y +
AAmb =(cid:18) 1
2 I I(cid:19) ;
e−2πiω·y f (cid:16)x +
2 I
− 1
I
x
2(cid:17) dy = BA Amb ( f , g) (x, ω) ,
(18)
y
2(cid:17) g(cid:16)x −
y
2(cid:17) dy;
(19)
• the cross-ambiguity function:
Amb ( f , g) (x, ω) =ÞRd
where
• the Wigner distribution:
W ( f , g) (x, ω) =ÞRd
• the Rihaczek distribution:
R ( f , g) (x, ω) =ÞRd
e−2πiω·y f (x) g (x − y)dy = e−2πix ·ω f (x) g (ω).
(20)
The latter two distributions are special cases of the τ-Wigner distribution defined in
(6). For any τ ∈ [0, 1], we have
Wτ ( f , g) (x, ω) = BAτ ( f , g) (x, ω) ,
where
Aτ =(cid:18) I
I − (1 − τ) I(cid:19) .
τI
(21)
The list of elementary properties is in line of those for the short-time Fourier
transform or the Wigner distribution. Mostly the proof is a straightforward com-
putation, and we refer to [1, 11] for the details. The interesting aspect is how the
parametrizing matrix A intervenes in the formulas for BA.
Proposition 4 Let A ∈ GL (2d, R) and f , g ∈ M1(Rd). The following properties
hold:
(i) Interchanging f and g:
BA (g, f ) (x, ω) = BC1 ( f , g) (x, ω),
(x, ω) ∈ R2d,
where
C1 = I AI2 =(cid:18) 0 I
I 0(cid:19)(cid:18) A11 A12
A21 A22(cid:19)(cid:18) I 0
A11 −A12(cid:19) .
0 −I(cid:19) =(cid:18) A21 −A22
In particular, BA f is a real-valued function if and only if A = C, namely
A11 = A21,
A12 = −A22.
Linear perturbations of the Wigner transform and the Weyl quantization
13
(ii) Behaviour of Fourier transforms:
where
BA(cid:16) f , g(cid:17) (x, ω) = det A−1 BC2 ( f , g) (−ω, x) ,
0 −I(cid:19)(cid:16) A−1(cid:17) ⊤(cid:18) 0 I
I 0(cid:19) .
C2 = I2 A# I =(cid:18) I 0
(x, ω) ∈ R2d,
(iii) Fourier transform of a MWD:
F BA ( f , g) (ξ, η) = BAJ ( f , g) (η, ξ) ,
(22)
where
AJ =(cid:18) A11 A12
A21 A22(cid:19)(cid:18) 0 I
−I 0(cid:19) =(cid:18) −A12 A11
−A22 A21(cid:19) .
3.1 Connection to the short-time Fourier transform
We first investigate the relation of the time-frequency representations BA to the ordi-
nary STFT. Whereas the Wigner distribution and the ambiguity transform coincide
with the STFT up to normalization, the time-frequency representation BA can be
written as a STFT only under an extra condition.
Definition 2 A block matrix A = (cid:18) A11 A12
A21 A22(cid:19) ∈ R2d×2d is called left-regular (resp.
right-regular), if the submatrices A11, A21 ∈ Rd×d (resp. A12, A22 ∈ Rd×d) are
invertible.
It is not difficult to prove that A = (cid:18) A11 A12
right-regular) if and only if the matrix2 A# = (cid:0) A−1(cid:1) ⊤
A21 A22(cid:19) ∈ GL (2d, R) is left-regular (resp.
(A#)21 (A#)22(cid:19) is right-
= (cid:18) (A#)11 (A#)12
regular (resp. left-regular).
As a matter of fact, the right-regularity of the matrix AST in (17) stands out
at a first glance and one might guess that this is an essential condition to express
BA( f , g) as a short-time Fourier transform. In fact, this characterization is very
strong, as stated in the subsequent results.
Theorem 2 ([1, Thm. 1.2.5]) Assume that A ∈ GL (2d, R) is right-regular. For every
f , g ∈ M1(Rd) the following formula holds:
BA ( f , g) (x, ω) = det A12−1 e2πi A#
12ω· A11 xVg f (c (x) , d (ω)) ,
x, ω ∈ Rd,
(23)
where
2 Beware that (A#)i j , A#
i j
= (A⊤
i j )−1, i, j = 1, 2.
14
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
c (x) =(cid:16) A11 − A12 A−1
22 A21(cid:17) x,
d (ω) = A#
12ω,
g (t) = g(cid:16)A22 A−1
12 t(cid:17) .
For the sake of clarity, one might use the following formulation:
Theorem 3 Given matrices M, N, P ∈ Rd×d and Q, R ∈ GL(R, d), set
A =(cid:18)
Then A is right-regular and
Q#
Q# N ⊤ M
R(cid:0)Q# N ⊤ M − P(cid:1) RQ#(cid:19) .
BA ( f , g) (x, ω) = det Q e2πiM x ·N ωVg◦R (Px, Qω) ,
x, ω ∈ Rd,
for any f , g ∈ M1(Rd).
Proof The proof is by computation:
det Q e2πiM x ·N ωVg◦R (Px, Qω)
e−2πiQω·y f (y) g (Ry − RPx)dy
= det Q e2πiM x ·N ωÞRd
= det QÞRd
= det QÞRd
=ÞRd
e−2πiω·y f (cid:16)Q#
=BA ( f , g) (x, ω) ,
e−2πiQω·(y−Q# N ⊤ M x) f (y) g (Ry − RPx)dy
e−2πiQω·y f (cid:16) y + Q# N ⊤ M x(cid:17) g(cid:0)Ry + R(cid:0)Q# N ⊤ M − P(cid:1) x(cid:1) dy
y + R(cid:0)Q# N ⊤ M − P(cid:1) x(cid:1) dy
y + Q# N ⊤ M x(cid:17) g(cid:0)RQ#
where A = AM, N, P,Q,R is as claimed.
Remark 1 The peculiar way the blocks of A are combined in
c(x) =(cid:16)A11 − A12 A−1
22 A21(cid:17) x
is a well-known construction in linear algebra and is usually called Schur comple-
ment. The Schur complement comes up many times in our results, ultimately because
of its distinctive role in the inversion of block matrices (cf. for instance [39, Thm.
2.1]).
For distributions associated with right-regular matrices, most results about the
short-time Fourier transform can be formulated for BA: for instance, one can easily
produce orthogonality formulae or a reconstruction formula for BA f . We will study
these issues in more generality in the subsequent sections.
Linear perturbations of the Wigner transform and the Weyl quantization
15
3.2 Main properties of the transformation BA
Having in mind that the entire knowledge on the uncertainty principles could be easily
transposed here, we give a qualitative result in the spirit of Benedick's theorem for
the Fourier transform - which is based on the corresponding uncertainty principle
for the STFT [27, Thm. 2.4.2].
Theorem 4 ([1, Thm. 1.4.3]) Let A ∈ GL (2d, R) be a right-regular matrix. If the
support of BA( f , g) has finite Lebesgue measure, then necessarily f ≡ 0 or g ≡ 0.
Next we characterize the boundedness of BA( f , g) on Lebesgue spaces - which
is a completely established issue for the STFT, cf. [4].
Proposition 5 Assume that A ∈ GL (2d, R) is right-regular. For any 1 ≤ p ≤ ∞ and
q ≥ 2 such that q′ ≤ p ≤ q, f ∈ L p(Rd) and g ∈ L p′
(i) BA( f , g) ∈ Lq(R2d), with
(Rd), we have
kBA( f , g)k L q ≤
k f k L p kgk L p′
det A
1
q det A12
p − 1
1
q det A22
.
p′ − 1
1
q
(24)
(ii) If 1 < p < ∞ then BA( f , g) ∈ C0(R2d). In particular, BA( f , g) ∈ L∞(R2d).
Furthermore, if 1 ≤ p, q ≤ ∞ such that p < q′ or p > q, the map BA( f , g) :
L p(Rd) × L p′
(Rd) → Lq(R2d) is not continuous.
Proof We refer to the proof of [11, Prop. 3.9] for the details concerning the first part.
Item (ii) is a direct application of [4, Prop. 3.2].
(cid:3)
The right-regularity of A is not only a technical condition required for (23) to
hold, but also has unexpected effects on the continuity of BA.
Theorem 5 ([1, Theorem 1.2.9]) Assume A ∈ GL (2d, R) such that det A22 , 0 but
det A12 = 0. Then there exist f , g ∈ L2(Rd) such that BA ( f , g) is not a continuous
function on R2d.
Let us exhibit the orthogonality relations, which extend the Parseval identity to
time-frequency distributions. The generalization of the orthogonality relations was
one of the main motivations for introducing BA in [1].
Theorem 6 ([1, Thm. 1.3.1]) Let A ∈ GL (2d, R) and f1, f2, g1, g2 ∈ L2(Rd). Then
hBA ( f1, g1) , BA ( f2, g2)i L2(R2d ) =
1
det A
h f1, f2i L2(Rd ) hg1, g2i L2(Rd ).
(25)
In particular,
kBA ( f , g)k L2(R2d ) =
1
det A1/2
k f k L2(Rd ) kgk L2(Rd ) .
16
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
Thus, the representation BA,g : L2(Rd) ∋ f 7→ BA ( f , g) ∈ L2(R2d) is a non-trivial
constant multiple of an isometry whenever g . 0.
The proof follows directly from the definition in (16), since F2 is unitary and T A is
a multiple of a unitary operator.
Corollary 1 If {en}n∈N is an orthonormal basis for L2(Rd), then
ndet A1/2 BA (em, en) m, n ∈ No
is an orthonormal basis for L2(R2d).
While the relevance of the orthogonality relations for signal processing or physics
purposes has been sometimes debated [6], they are in fact a useful tool for proving
several properties of the time-frequency distributions that satisfy them. In particular,
orthogonality relations are the main ingredients of a general procedure for recon-
structing a signal from the knowledge of its (cross-)time-frequency distribution with
a given window.
Theorem 7 ([11, Cor. 3.1.7]) Assume A ∈ GL (2d, R) and fix g, γ ∈ L2(Rd) such
that hg, γi , 0. Then, for any f ∈ L2(Rd), the following inversion formula holds:
f =
det A
hg, γi
B∗
A,γ BA,g f ,
A,γ : L2(R2d) → L2(Rd) is the adjoint operator of BA,γ ≡ BA(·, γ), defined
where B∗
as
with
B∗
A,γ H (x) =
1
det AÞRd
T A⋆ F2H (x, y) γ (y) dy,
A⋆
= I2 A−1 ∈ GL (2d, R) .
For right-regular matrices the reconstruction can be made more explicit.
Proposition 6 ([1, Thm. 1.3.3]) Let A ∈ GL (2d, R) be a right-regular matrix, and
g, γ ∈ L2(Rd) such that hg, γi , 0. The following inversion formula (to be interpreted
as vector-valued integral in L2(Rd)) holds for any f ∈ L2(Rd):
1
BA,γ f (x, ω)
e−2πi A#
12ω· A11 x
det A12
Md(ω)Tc(x) gdxdω,
where
f =
hg, γiÞR2d
c (x) =(cid:16) A11 − A12 A−1
22 A21(cid:17) x,
d (ω) = A#
12ω,
g (t) = g(cid:16)A22 A−1
12 t(cid:17) .
Another property that is expected to hold for a time-frequency representation is
the covariance under phase-space shifts.
Linear perturbations of the Wigner transform and the Weyl quantization
17
Theorem 8 ([1, Thm. 1.5.1]) Let A ∈ GL (2d, R). For any f , g ∈ M1(Rd) and
a, b, α, β ∈ Rd, the following formula holds:
BA(cid:0)MαTa f , MβTb g(cid:1) (x, ω) = e2πiσ ·s M(ρ,−s)T(r,σ)BA ( f , g) (x, ω)
= e2πiσ ·se2πi(x ·ρ−ω·s)BA ( f , g) (x − r, ω − σ) ,
(26)
(27)
where
s(cid:19) = A−1(cid:18) a
b(cid:19) ,
(cid:18) r
σ(cid:19) = A⊤(cid:18) α
−β(cid:19) .
(cid:18) ρ
Of course, this result encompasses the covariance formula for the τ-Wigner
distribution with A = Aτ as in (21), cf. [15, Prop. 3.3] and also for the STFT with
A = AST , cf [26, Lem. 3.1.3].
We now cite an amazing representation result for the STFT of a MWD, sometimes
called the magic formula for other distributions (cf. [28]). This relation allows to
painlessly extend our results to general function spaces tailored for the purposes of
time-frequency analysis, namely modulation spaces.
Theorem 9 ([1, Thm. 1.7.1]) Assume A ∈ GL (2d, R) and f , g, ψ, φ ∈ M1(Rd), and
set z = (z1, z2), ζ = (ζ1, ζ2) ∈ R2d. Then,
VB A(φ,ψ)BA ( f , g) (z, ζ ) = e−2πiz2 ·ζ2Vφ f (a, α) Vψ g (b, β),
(28)
where
b(cid:19) = AI2(cid:18) z1
(cid:18) a
β(cid:19) = I2 A#(cid:18) ζ1
(cid:18) α
ζ2(cid:19) =(cid:18) A11z1 − A12ζ2
A21z1 − A22ζ2(cid:19) ,
−(A#)21ζ1 − (A#)22z2(cid:19) .
z2(cid:19) =(cid:18) (A#)11ζ1 + (A#)12 z2
As a concluding remark, we want to underline that the benefits of linear algebra
should be appreciated in view of the very short and simple proofs. This aspect should
not be underestimated: the proof of similar results for certain special members has
lead to quite cumbersome computations (cf. the proofs for the τ-Wigner distributions
in [15]).
3.3 Cohen class members as perturbations of the Wigner transform
We already described the heuristics behind the Cohen class of distributions in the
introduction.
Definition 3 ([26]) A time-frequency distribution Q belongs to the Cohen's class if
there exists a tempered distribution θ ∈ S ′(R2d) such that
Q ( f , g) = W ( f , g) ∗ θ,
∀ f , g ∈ S(Rd).
18
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
Although the Wigner distribution was the main inspiration for the MWDs studied
so far, the connection to the Cohen class is by no means clear. This question is the
point of departure of the paper [11] and the following result completely character-
izes the intersection between these families. We rephrase [11, Thm. 1.1] using the
Feichtinger algebra as follows.
Theorem 10 Let A ∈ GL (2d, R). The distribution BA belongs to the Cohen class if
and only if
A = AM =(cid:18) I M + (1/2)I
I M − (1/2)I(cid:19)
as in (11), for some M ∈ Rd×d. Furthermore, in this case we have
WM ( f , g) ≡ BAM ( f , g) = W ( f , g) ∗ θ M,
f , g ∈ M1,
(29)
where the Cohen's kernel θ M ∈ S′
0(Rd) is
θ M = F ΘM, with ΘM (ξ, η) = e−2πiξ ·M η,
(ξ, η) ∈ R2d .
(30)
If M is invertible, the kernel θ M is explicitly
θ M (x, ω) =
1
det M
e2πix ·M −1ω,
(x, ω) ∈ R2d.
(31)
We say that A = AM is a Cohen-type matrix associated with M ∈ Rd×d.
Remark 2 We mention that, according to the proof of the necessity part in the previous
result, a Cohen-type matrix A should be defined by the following conditions on the
blocks:
A11 = A21 = I,
A12 − A22 = I .
(32)
The choice A22 = M − (1/2)I with M ∈ Rd×d is thus a suitable parametrization,
but by no means the only possible one - and in fact neither the most natural one.
The reason underlying our choice appears if one writes down the explicit formula
for BAM as
WM ( f , g)(x, ω) =ÞRd
e−2πiω·y f (cid:18)x +(cid:18)M +
1
2
I(cid:19) y(cid:19) g(cid:18)x +(cid:18)M −
1
2
I(cid:19) y(cid:19) dy,
(33)
which reveals the similarity with the Wigner distribution. A sort of symmetry with
respect to the Wigner distribution (corresponding to M = 0) immediately stands
out. We interpret these representations as a family of "linear perturbations" of the
Wigner distribution and M as the control parameter, exactly as τ controls the degree
of deviation of τ-Wigner distributions. For this reason, we will refer to A = AM
as the perturbative form of a Cohen-type matrix. The analogy with the τ-Wigner
distributions naturally leads to another representation, hence another choice of A22
in (32). A closer inspection of the kernel (7) and also of (6) reveals that the role of
perturbation parameter is not played by τ, rather by the deviation µ = τ − 1/2. In
Linear perturbations of the Wigner transform and the Weyl quantization
19
this analogy one chooses A21 = T ∈ Rd×d and A22 = −(I − T ) and obtains
WT ( f , g)(x, ω) ≡ BAT ( f , g)(x, ω) =ÞRd
(34)
which should be compared to (6) (see also (21)). Occasionally, we refer to AT as the
affine form of the Cohen-type matrix A. It is clear that the two forms of a Cohen-type
matrix are perfectly equivalent, the connection being
e−2πiω·y f (x + T y) g (x − (I − T )y)dy,
M = T − (1/2)I .
(35)
Therefore, the choice of a form is just a matter of convenience: when studying
the properties of BA as a time-frequency representation, it seems better to explicitly
see the effect of the perturbation M (which could be easily turned off setting M = 0)
and use the perturbative form accordingly. As an example of this, the perturbed
representation of a Gaussian signal is provided.
Lemma 4 ([11, Lem. 4.1]) Consider A = AM ∈ GL (2d, R) as in (11) and ϕλ (t) =
e−πt 2/λ, λ > 0. Then,
WM ϕλ (x, ω) = (2λ)d/2 det (S)−1/2 e−2π x2/λ
· e8π(M ⊤ x ·S−1 M ⊤ x)/λe8πiS−1 ω·M ⊤ x e−2πλω·S−1ω,
(36)
where S = I + 4M ⊤ M ∈ Rd×d.
3.3.1 Main properties of the Cohen class
The properties of a time-frequency distribution belonging to the Cohen class are
intimately related to the structure of the Cohen kernel. There is an established list of
correspondences between the kernel and the properties, which can be used to deduce
the following results. See [7, 11, 36, 38].
Proposition 7 Assume that BA belongs to the Cohen's class. For any f , g ∈ M1(Rd),
the following properties are satisfied:
(i) Correct marginal densities:
ÞRd
BA f (x, ω) dω = f (x)2 ,
ÞRd
In particular, the energy is preserved:
BA f (x, ω) dx =(cid:12)(cid:12) f (ω)(cid:12)(cid:12)2
,
x, ω ∈ Rd .
(ii) Moyal's identity:
ÞR2d
BA f (x, ω) dxdω = k f k2
L2 .
hBA f , BAgi L2(R2d ) = h f , gi2 .
20
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
(iii) Symmetry: for all x, ω ∈ Rd,
BA (I f ) (x, ω) = IBA f (x, ω) = BA f (−x, −ω) ,
BA(cid:16) f(cid:17) (x, ω) = I2BA f (x, ω) = BA (x, −ω).
(iv) Convolution properties: for all x, ω ∈ Rd,
BA ( f ∗ g) (x, ω) = BA f ∗1 BAg,
BA ( f · g) (x, ω) = BA f ∗2 BAg.
Here ∗1 (and ∗2) denotes the convolution with respect to the first (second)
variable.
(v) Scaling invariance: setting Uλ f (t) ≔ λd/2 f (λt), λ ∈ R \ {0}, t ∈ Rd,
BA (Uλ f ) (x, ω) = BA f (cid:16)λx, λ−1ω(cid:17) .
(vi) Strong support property3: the only MWDs in the Cohen class satisfying the
strong correct support properties are Rihaczek and conjugate-Rihaczek distri-
butions.
(vii) Weak support property4: the only MWDs in Cohen's class satisfying the weak
correct support properties are the τ-Wigner distributions with τ ∈ [0, 1].
We now give a few hints on several aspects of interests for both theoretical prob-
lems and applications; extensive discussions on these issues may be found in [1, 11].
Real-valuedness. In view of Proposition 4 (i), BA0
is the only real-valued member of the family BAM .
= W (the Wigner distribution)
More on marginal densities. The marginal densities for a general distribution
BA can be easily computed. For f , g ∈ M1(Rd),
3 Let Q f : R2d
a suitable function space. Recall that Q is said to satisfy the strong support property if
→ C be a time-frequency distribution associated with the signal f : Rd
(x, ω)
t → C in
f (x) = 0 ⇔ Q f (x, ω) = 0,
∀ω ∈ Rd,
f (ω) = 0 ⇔ Q f (x, ω) = 0,
∀x ∈ Rd .
4 With the notation of the previous footnote, we say that Q satisfies the weak support property if,
for any signal f :
πx (suppQ f ) ⊂ C (supp f ) ,
πω (suppQ f ) ⊂ C(cid:16)supp f(cid:17) ,
where πx : R2d
factors (R2d
(x, ω) → Rd
≃ Rd
x × Rd
( x, ω)
x and πω : R2d
ω ) and C (E) is the closed convex hull of E ⊂ Rd.
(x, ω) → Rd
ω are the projections onto the first and second
Linear perturbations of the Wigner transform and the Weyl quantization
21
ÞRd
ÞRd
BA f (x, ω) dω = f (A11 x) f (A21 x),
BA f (x, ω) dx = det A−1 f (cid:16)(A#)12ω(cid:17) f (cid:0)−(A#)22ω(cid:1).
The correct marginal densities are thus recovered if and only if A11 = A21 = I and
(A#)12 = −(A#)22 = I. These conditions force both det A = 1 and the block struc-
ture of A as that of Cohen's type. This fact provides an equivalent characterization of
the distributions BA belonging to the Cohen class: these are exactly those satisfying
the correct marginal densities.
Relation between two distributions. Let A1 = AM1 and A2 = AM2 be two
Cohen-type matrices as in (11). The two distributions WM1 and WM2 are connected
by a Fourier multiplier as follows ([11, Lem. 4.2]). For f , g ∈ M1(Rd),
F WM2 ( f , g) (ξ, η) = e−2πiξ ·(M2−M1)η F WM1 ( f , g) (ξ, η) .
(37)
Furthermore, if M2 − M1 ∈ GL (d, R),
BA2 ( f , g) (x, ω) =
1
det (M2 − M1)
e2πix ·(M2−M1)−1
ω ∗ BA1 ( f , g) (x, ω) .
The proofs follow at once from Theorem 10.
Regularity of the Cohen kernel. Let A = AM ∈ GL (2d, R) be a Cohen-type
matrix and assume in addition that M ∈ GL (d, R). The Cohen kernel associated
with BAM is therefore given by (31), and we can study its regularity with respect to
the scale of modulation spaces: we have ([11, Prop. 4.8])
θ M, F θ M ∈ M1,∞(R2d) ∩ W(cid:16)F L1, L∞(cid:17) (R2d).
This result has an interesting counterpart on the regularity of WM on all modu-
lation spaces, in view of the boundedness of Fourier multipliers with symbols in
W(cid:0)F L1, L∞(cid:1) (cf. [2, Lem. 8]):
Theorem 11 ([11, Thm. 4.10]) Let A = AM ∈ GL (2d, R) be a Cohen-type matrix
with M ∈ GL (d, R) and f ∈ M ∞(Rd) be a signal. Then, for any 1 ≤ p, q ≤ ∞, we
have
W f ∈ M p,q(R2d) ⇐⇒ WM f ∈ M p,q(R2d).
Unfortunately, one cannot go too far if the non-singularity of M is dropped; as a
trivial instance, notice that for M = 0 one has θ0 = δ, and it is easy to verify that
δ ∈ M1,∞(R2d)\W(cid:0)F L1, L∞(cid:1) (R2d), cf. [13].
Perturbation and interferences. The emergence of unwanted artefacts is a well-
known drawback of any quadratic representation. The signal processing literature
is full of strategies to mitigate these effects (see for instance [7, 32, 33]). For what
22
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
concerns the Cohen class, it is folklore that the severity of interferences is somewhat
related to the decay of the Cohen kernel. In fact, a precise formulation of this princi-
ple is rather elusive and recent contributions unravelled further non-trivial fine points
([14, Prop. 4.4 and Thm. 4.6]). We remark that the chirp-like kernel ΘM = F θ M
does not decay at all, and thus no smoothing effect should be expected for the per-
turbed representations. This is confirmed by the experiments in dimension d = 1 in
[3, 11]. The only effect of the perturbation consists of a distortion and relocation
of interferences, but there is no damping. Following the engineering literature, we
suggest that convolution with suitable decaying distributions may provide some im-
provement, probably at the price of loosing other nice properties.
Covariance formula. For any z = (z1, z2) , w = (w1, w2) ∈ R2d, the covariance
formula (26) now reads
WM (π (z) f , π (w) g) (x, ω) = e2πi[ 1
2 (z2+w2)+M (z2−w2)]·(z1−w1)
× MJ(z−w)TTM (z,w)WM ( f , g) (x, ω) ,
(38)
where
TM (z, w) =(cid:18) (1/2) (z1 + w1) + M (w1 − z1)
(1/2) (z2 + w2) + M (z2 − w2)(cid:19)
(z + w) +(cid:18) −M 0
0 M(cid:19) (z − w) .
I + PT =(cid:18) I − T 0
0 T(cid:19) ,
0 −(I − T )(cid:19) ,
PT =(cid:18) −T
=
1
2
0
Alternatively, adopting the affine representation of A (35):
(39)
(40)
we can also write
TT (z, w) =(cid:18) (I − T )z1 + T w1
T z2 + (I − T )w2(cid:19) = (I + PT ) z − PT w.
Boundedness on modulation spaces. We cite here some results on the continuity
of the distributions BAM on the aforementioned spaces. For the sake of clarity, we
report a simplified, unweighted form of [11, Thm. 4.12].
Theorem 12 Let A = AT ∈ GL (2d, R) be a Cohen-type matrix. Let 1 ≤ pi, qi, p, q ≤
∞, i = 1, 2, such that
and
pi, qi ≤ q,
i = 1, 2,
1
p1
+
1
p2
≥
1
p
+
1
q
,
1
q1
+
1
q2
≥
1
p
+
1
q
.
(41)
(42)
Linear perturbations of the Wigner transform and the Weyl quantization
23
(i) If f1 ∈ M p1,q1(Rd) and f2 ∈ M p2,q2 (Rd), then WT ( f1, f2) ∈ M p,q(cid:0)R2d(cid:1), and
the following estimate holds:
kWT ( f1, f2)k M p, q .T k f1 k M p1, q1 k f2 k M p2, q2 .
(ii) Assume further that both T and I − T are invertible (equivalently: AT is right-
regular, or PT is invertible, cf. (39)). If f1 ∈ M p1,q1 (Rd) and f2 ∈ M p2,q2(Rd),
then WT ( f1, f2) ∈ W (F L p, Lq)(cid:0)R2d(cid:1), and the following estimate holds:
kWT ( f1, f2)kW (F L p, L q ) .T (CT )1/q−1/p k f1 k M p1, q1 k f2 k M p2, q2 ,
where
CT = det T det (I − T ) > 0.
(43)
Sharp estimates and continuity results of this type have been given by some of
the authors for the case of τ-Wigner distributions in [16, Lem. 3.1] and [15].
To conclude this section we remark that one can specialize Proposition 5 in
order to characterize the boundedness on Lebesgue spaces at the price of assuming
right-regularity of AM, see [11, Thm. 4.14].
4 Pseudodifferential operators
In this section we discuss the formalism of pseudodifferential operators that is as-
sociated with every time-frequency representation BA. Imitating the time-frequency
analysis of Weyl pseudodifferential operators, we introduce the following general
calculus for pseudodifferential operators.
Theorem 13 ([1, Prop. 2.2.1]) Let A ∈ GL (2d, R) and σ ∈ M ∞(cid:0)R2d(cid:1). The mapping
opA(σ) ≡ σ A defined by duality as
(cid:10)σ A f , g(cid:11) ≡ hσ, BA (g, f )i ,
f , g ∈ M1(Rd)
is a well-defined linear continuous map from M1(Rd) to M ∞(Rd).
The proof easily follows from the continuity of the distribution BA : M1(Rd) ×
M1(Rd) → M1(R2d), from Proposition 3.
Theorem 13, namely
Definition 4 Let A ∈ GL (2d, R) and σ ∈ M ∞(cid:0)R2d(cid:1). The mapping defined in
σ A : M1(Rd) ∋ f 7→ σ A f ∈ M ∞(Rd) :(cid:10)σ A f , g(cid:11) = hσ, BA (g, f )i , ∀g ∈ M1(Rd),
is called quantization rule with symbol σ associated with the matrix-Wigner distri-
bution BAor pseudodifferential operator with symbol σ associated with the matrix-
Wigner distribution BA.
24
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
Using Feichtinger's kernel theorem (Theorem 1), we now provide a number of
equivalent representations for σ A f .
following representations:
Theorem 14 Let A ∈ GL (2d, R). Let T : M1(Rd) → M ∞(Rd) be a continuous
linear operator. There exist distributions k, σ, F ∈ M ∞(cid:0)R2d(cid:1) such that T admits the
1. as an integral operator with kernel k: hT f , gi = Dk, g ⊗ fE for any f , g ∈
M1(Rd);
2. as pseudodifferential operator with symbol σ associated with BA: T = σ A;
3. as a superposition (in weak sense) of time-frequency shifts (also called spreading
representation):
T =ÞR2d
F (x, ω) Tx Mω dxdω.
The relations among k, σ, F and A are the following:
σ = det A F2T Ak,
F = F2T AST k.
(44)
Proof The first representation is exactly the claim of kernel theorem. Now set σ =
det A F2T Ak ∈ M ∞(cid:0)R2d(cid:1): this is a well-defined distribution, since F2 and T A are
isomorphisms on M ∞(cid:0)R2d(cid:1). In particular, for any f , g ∈ M1(Rd) we have
2 σ, g ⊗ fE
A F −1
hT f , gi =Dk, g ⊗ fE
=Ddet A−1 T−1
=Dσ, F2T A(cid:16)g ⊗ f(cid:17)E
=(cid:10)σ A f , g(cid:11) .
= hσ, BA (g, f )i
This proves that T f = σ A f in M ∞(Rd). The relation between the kernel represen-
tation in 1 and the spreading representation in 3 is well-known, e.g. [26]. It can also
be deduced from item 2 from the special matrix AST =(cid:18) 0 I
−I I(cid:19) and
hT f , gi =(cid:10)F, Vf g(cid:11) =(cid:10)F, BAST (g, f )(cid:11) ,
f , g ∈ M1(Rd).
A−1 F −1
2 σ = det A−1 T
Remark 3 Since k = det A−1 T
obtain another representation of the third type with a special spreading function:
1 bσ, one can formally
det AÞR2dbσ(ξ, −(A−1)21 x − (A−1)22 y)e2πiξ ·[(A−1 )11 x+(A−1)22y] f (y)dξdy.
Notice that the inverse of a Cohen-type matrix A = AT has the form
I2 A−1 F −1
σ A f (x) =
(45)
1
Linear perturbations of the Wigner transform and the Weyl quantization
25
A−1
T
=(cid:18) −(I − T ) T
−I(cid:19) ,
I
thus the previous formula becomes
σ A f (x) =ÞR2dbσ(ξ, u)e−2πi(I −T )u ·ξ T−u Mξ f (x)dξdu.
This should be compared with [26, Eq. 14.14] and [15, Eq. 20].
(46)
We now study the relations among pseudodifferential operators associated with
MWDs and the corresponding symbols.
Proposition 8 Let A, B ∈ GL (2d, R) and σ, ρ ∈ M ∞(cid:0)R2d(cid:1). Then,
σ A
= ρB ⇐⇒ σ =
det A
det B
F2T
B−1 AF −1
2 ρ
Proof Assume that T = σ A = ρB. According to Theorem 14, T has a distributional
kernel k such that
σ = det A F2T Ak,
ρ = det B F2TB k.
Therefore,
σ = det A F2T Ak
=
=
det A
det B
det A
det B
F2T AT−1
B F −1
2 ρ
F2T
B−1 AF −1
2 ρ.
On the other side, if σ = det A det B−1 F2T
B−1 AF −1
2 ρ, then for any f , g ∈ M1(Rd)
(cid:10)σ A f , g(cid:11) = hσ, F2T A ( f ⊗ g)i
= hρ, F2TB ( f ⊗ g)i
=(cid:10)det A det B−1 F2T AT−1
=(cid:10)ρB f , g(cid:11) .
B F −1
2 ρ, F2T A ( f ⊗ g)(cid:11)
When the operators are associated with Cohen-type matrices, we have a more explicit
relation that covers the usual rule for τ-Shubin operators (cf. [42, Rem. 1.5]). The
proof is a straightforward application of (37).
Proposition 9 Let A1 = AT1, A2 = AT2 be Cohen-type invertible matrices, and
σ, ρ ∈ M ∞(cid:0)R2d(cid:1). Then,
= σT2
σT1
1
2 ⇐⇒ cσ2 (ξ, η) = e−2πiξ ·(T2−T1)ηcσ1 (ξ, η) .
26
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
It is also interesting to characterize the matrices yielding self-adjoint operators.
Proposition 10 ([1, Prop. 2.2.3])
Let A ∈ GL (2d, R) and σ ∈ M ∞(cid:0)R2d(cid:1). Then
where
ρ = σ,
= ρB,
(cid:16)σ A(cid:17)∗
B = I AI2 =(cid:18) A21 −A22
A11 −A12(cid:19) .
In particular, σ A is self-adjoint if and only if σ = σ (real symbol) and B = A, hence
A21 = A11,
A12 = −A22.
Remark 4 Thus, only matrices of the form(cid:18) P Q
P −Q(cid:19), with P, Q ∈ GL (d, R), give rise
to pseudodifferential operators which are self-adjoint for real symbols. This occurs
for Weyl calculus but not for Kohn-Nirenberg operators (T = 0).
4.1 Boundedness results
4.1.1 Operators on Lebesgue spaces
The boundedness of a pseudodifferential operator σ A associated with BA is inti-
mately related to the boundedness of the distribution BA on certain function spaces,
in view of the duality in the definition of σ A. Let us start this section with two easy
results.
Proposition 11 ([1, Theorems 2.2.6, 2.2.7 and 2.2.9]) Let A ∈ GL (2d, R) and
σ ∈ M ∞(cid:0)R2d(cid:1).
1. If A is right-regular and σ ∈ L1(cid:0)R2d(cid:1), then σ A is a bounded operator on
L2(Rd) such that
(cid:13)(cid:13)σ A(cid:13)(cid:13)L2→L2 ≤
kσk L1
det A121/2 det A221/2
.
2. If σ ∈ L2(cid:0)R2d(cid:1), then σ A is a bounded operator on L2(Rd) such that
(cid:13)(cid:13)σ A(cid:13)(cid:13)L2→L2 ≤
kσk L2
det A1/2
.
If σ ∈ L2(cid:0)R2d(cid:1), then σ A is actually a Hilbert-Schmidt operator, and every Hilbert-
Schmidt operator T possesses a symbol σ ∈ L2(R2d) such that T = σ A.
Linear perturbations of the Wigner transform and the Weyl quantization
27
The study on Lebesgue spaces can be largely expanded thanks to the following
result.
Theorem 15 Let A ∈ GL (2d, R) be right-regular and σ ∈ Lq(cid:0)R2d(cid:1). The quantiza-
tion mapping
σ ∈ Lq(R2d) 7→ σ A ∈ L(cid:16)L p(Rd)(cid:17)
is continuous if and only if q ≤ 2 and q ≤ p ≤ q′, with norm estimate
(cid:13)(cid:13)σ A(cid:13)(cid:13)L p →L p ≤
det A
1
q′ det A12
q′ det A22
kσk L q
1
p − 1
.
1
p′ − 1
q′
Proof Assume f ∈ L p(Rd) and g ∈ L p′
(24) (switch q and q′) and Hölder inequality:
(Rd), with p , 1 nor p , ∞. Therefore, by
(cid:12)(cid:12)(cid:10)σ A f , g(cid:11)(cid:12)(cid:12) = hσ, BA (g, f )i
≤ kσk L q kBA (g, f )k L q′
kσk L q
1
p − 1
1
q′ det A12
det A
≤
q′ det A22
k f k L p kgk L p′ .
p′ − 1
1
q′
The non-continuity is a consequence of [4, Prop. 3.2].
(cid:3)
Remark 5 Note that the closed graph theorem implies non-continuity of the quanti-
zation map. This means that there exists a symbol σ ∈ Lq for which the operator
σ A is not bounded on L p(R2d), cf. [4, Prop. 3.4].
One could also study compactness and Schatten class properties for these opera-
tors. We confine ourselves to prove a result for symbols in the Feichtinger algebra.
Theorem 16 ([1, Thm. 2.2.8]) Let A ∈ GL (2d, R) and σ ∈ M1(cid:0)R2d(cid:1). The operator
σ A ∈ L(cid:0)L2(Rd)(cid:1) belongs to the trace class S1(cid:0)L2(Rd)(cid:1), and the following estimate
holds for the trace class norm:
(cid:13)(cid:13)σ A(cid:13)(cid:13)S1 . det A1/2 kσk M 1 .
Proof The proof follows the outline of [25]. Proposition 11 immediately yields
σ A ∈ L(cid:0)L2(Rd)(cid:1), since M1(cid:0)R2d(cid:1) ⊆ L1(cid:0)R2d(cid:1) ∩ L2(cid:0)R2d(cid:1). In line with the paradigm
of time-frequency analysis of operators (cf. [26, Sec. 14.5]), let us decompose the
action of σ A into elementary pseudodifferential operators with time-frequency shifts
of a suitable function as symbols. The inversion formula for the STFT allows to write
σ =ÞR2dÞR2d
VΦσ (z, ζ ) MζTz Φdzdζ,
for any window function Φ ∈ M1(cid:0)R2d(cid:1) with kΦk L2 = 1. Therefore, for any f , g ∈
M1(Rd):
28
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
(cid:10)σ A f , g(cid:11) = hσ, BA (g, f )i
=ÞR2dÞR2d
=ÞR2dÞR2d
VΦσ (z, ζ )(cid:10)MζTz Φ, BA (g, f )(cid:11) dzdζ
VΦσ (z, ζ )D(cid:0)MζTz Φ(cid:1) A
f , gE dzdζ .
This shows that σ A acts (in an operator-valued sense on L2) as a continuous weighted
superposition of elementary operators:
σ A
=ÞR2dÞR2d
VΦσ (z, ζ )(cid:0)Mζ Tz Φ(cid:1) A
dzdζ .
The action of the building blocks (cid:0)Mζ Tz Φ(cid:1) A can be unwrapped by means of the
magic formula (28) provided one takes Φ = BAϕ for some ϕ ∈ M1(Rd) with
kϕk L2 = det A1/4 (see the orthogonality relations (25)):
D(cid:0)Mζ Tz Φ(cid:1) A
f , gE =(cid:10)MζTz Φ, BA (g, f )(cid:11)
= VB Aϕ BA (g, f ) (z, ζ )
= e2πiz2 ·ζ2Vϕ f (b, β) Vϕ g (a, α),
where a, α, b, β are continuous functions of z and ζ. In particular, we have
f 7→ e2πiz2 ·ζ2(cid:10) f , MβTb ϕ(cid:11) MαTa,
(cid:0)Mζ Tz Φ(cid:1) A : L2(Rd) → L2(Rd)
hence(cid:0)MζTz Φ(cid:1) A is a rank-one operator with trace class norm given by(cid:13)(cid:13)(cid:13)(cid:0)Mζ TzΦ(cid:1) A(cid:13)(cid:13)(cid:13)S1
kϕk2
σ A and compute its norm by means of the estimates for the pieces:
L2 = det A1/2, independent of z, ζ. To conclude, we reconstruct the operator
:
=
(cid:13)(cid:13)σ A(cid:13)(cid:13)S1 ≤ÞR2dÞR2d
VΦσ (z, ζ )(cid:13)(cid:13)(cid:13)(cid:0)MζTz Φ(cid:1) A(cid:13)(cid:13)(cid:13)S1
dzdζ ≤ CA kσk M 1 .
4.1.2 Operators on modulation spaces
We now study the boundedness on modulation spaces of pseudodifferential operators
associated with Cohen-type representations.
Theorem 17 (Symbols in M p,q) Let A = AT ∈ GL (2d, R) be a Cohen-type matrix
and consider indices 1 ≤ p, p1, p2, q, q1, q2 ≤ ∞, satisfying the following relations:
p1, p′
2, q1, q′
2 ≤ q′,
and
Linear perturbations of the Wigner transform and the Weyl quantization
29
1
p1
+
1
p′
2
≥
1
p′
+
1
q′,
1
q1
+
1
q′
2
≥
1
p′
+
1
q′ .
For any σ ∈ M p,q(cid:0)R2d(cid:1), the pseudodifferential operator σ A is bounded from
M p1,q1 (Rd) to M p2,q2 (Rd). In particular,
(cid:13)(cid:13)σ A(cid:13)(cid:13)M p1, q1 →M p2, q2 . A kσk M p, q .
Proof Under the given assumptions on the indices, Theorem 12 implies that
BAT (g, f ) ∈ M p′,q′
2(Rd). Therefore, by
the duality of modulation spaces we obtain
(Rd) for any f ∈ M p1,q1 (Rd) and g ∈ M p′
2,q′
(cid:12)(cid:12)(cid:10)σ A f , g(cid:11)(cid:12)(cid:12) = hσ, BA (g, f )i
≤ kσk M p, q(cid:13)(cid:13)BAM (g, f )(cid:13)(cid:13)M p′, q′
. A kσk M p, q k f k M p1, q1 kgk M p2, q2 .
For symbols in W (F L p, Lq) spaces, we can extend [16, Thm. 1.1] to the matrix
setting.
Theorem 18 (Symbols in W (F L p, Lq)) Let A = AT ∈ GL (2d, R) a right-regular
Cohen-type matrix and consider indices 1 ≤ p, q, r1, r2 ≤ ∞, satisfying the following
relations:
q ≤ p′,
r1, r ′
1, r2, r ′
2 ≤ p.
For any σ ∈ W (F L p, Lq)(cid:0)R2d(cid:1), the pseudodifferential operator σ A is bounded on
Mr1,r2 (Rd); in particular,
(cid:13)(cid:13)σ A(cid:13)(cid:13)M r1, r2 →M r1, r2 . A kσkW ( F L p, L q ) .
Proof We isolate two special cases of Theorem 12. First, for any f ∈ M p1, p2 (Rd)
and g ∈ M p′
2 (Rd) we have
1, p′
kWT (g, f )kW (F L1, L∞) .T k f k M p1, p2 kgk
p′
1, p′
2
M
,
1 ≤ p1, p2 ≤ ∞.
This yields that σ A is bounded on M p1, p2 (Rd), for any 1 ≤ p1, p2 ≤ ∞ and that the
symbol σ is in W(cid:0)F L∞, L1(cid:1)(cid:0)R2d(cid:1), because
(cid:12)(cid:12)(cid:10)σ A f , g(cid:11)(cid:12)(cid:12) = hσ, WT (g, f )i
≤ kσkW (F L∞, L1) kWT (g, f )kW (F L1, L∞)
.T kσkW(F L∞, L1) k f k M p1, p2 kgk
p′
1, p′
2
M
.
The second case requires f , g ∈ M2(Rd), then
kWT (g, f )kW(F L2, L2) .T k f k M 2 kgk M 2 .
30
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
If σ ∈ W(cid:0)F L2, L2(cid:1)(cid:0)R2d(cid:1) = L2(R2d), then σ A is bounded on M2(Rd) = L2(Rd) by
similar arguments:
(cid:13)(cid:13)σ A(cid:13)(cid:13)M 2→M 2 .T kσkW(F L2, L2) .
We proceed now by complex interpolation of the continuous mapping opA on
modulation spaces; in particular, we are dealing with
opA : W(cid:16)F L1, L∞(cid:17)(cid:16)R2d(cid:17) × M p1, p2 (Rd) → M p1, p2 (Rd),
opA : W(cid:16)F L2, L2(cid:17)(cid:16)R2d(cid:17) × M2(Rd) → M2(Rd).
For θ ∈ [0, 1], we have
hW(cid:16)F L1, L∞(cid:17) , W(cid:16)F L2, L2(cid:17)i θ
= W(cid:16)F L p, L p′(cid:17) ,
= Mr1,r2,
2 ≤ p ≤ ∞,
with
1
ri
=
1 − θ
pi
+
1
p
,
i = 1, 2.
pi
(cid:2)M p1, p2, M2,2(cid:3) θ
1 − θ
+
=
θ
2
From these estimates we immediately derive the condition r1, r ′
sion relations for modulation spaces allow to extend the result to W (F L p, Lq)(cid:0)R2d(cid:1)
for any q ≤ p′. Finally, exchanging the role of p and p′ allows us to cover any
p ∈ [1, ∞].
(cid:3)
1, r2, r ′
2 ≤ p. The inclu-
4.2 Symbols in the Sjöstrand class
Another important space of symbols is the Sjöstrand class. It has been introduced
by Sjöstrand [41] to extend the well-behaved Hörmander class S0
0,0 and later rec-
ognized to coincide with the modulation space M ∞,1(R2d). Accordingly, it consists
of bounded symbols with low regularity in general, namely temperate distributions
σ ∈ S ′(R2d) such that
ÞR2d
sup
z ∈R2d
hσ, π(z, ζ )gidζ < ∞.
Nevertheless, they lead to L2-bounded pseudodifferential operators. In fact, much
more is true: the family of Weyl operators with symbols in the Sjöstrand class is an
inverse-closed Banach *-subalgebra of L(L2)(Rd), in the following sense.
Theorem 19
(i) (Boundedness) If σ ∈ M ∞,1(cid:0)R2d(cid:1), then opW(σ) is a bounded operator on
L2(Rd).
Linear perturbations of the Wigner transform and the Weyl quantization
31
(ii) (Algebra property) If σ1, σ2 ∈ M ∞,1(cid:0)R2d(cid:1) and opW(ρ) = opW(σ1)opW(σ2),
then ρ ∈ M ∞,1(cid:0)R2d(cid:1).
(iii) (Wiener property) If σ ∈ M ∞,1(cid:0)R2d(cid:1) and opW(σ) is invertible on L2(Rd), then
= opW(ρ) for some ρ ∈ M ∞,1(cid:0)R2d(cid:1).
These results have been put into the context of time-frequency analysis in [29]
and have been generalized, see [15, 30, 31]. In this section we extend Theorem 19
with respect to MWDs.
[opW(σ)]−1
Let us first provide some conditions on the matrices for which the associated
pseudodifferential operators with symbols in M ∞,1(R2d) are bounded on modulation
spaces.
Theorem 20 ([1, Thm. 2.3.1]) Let σ ∈ M ∞,1(R2d) and assume A ∈ GL (2d, R) is a
left-regular matrix. The pseudodifferential operator σ A is bounded on all modulation
spaces M p,q(Rd), 1 ≤ p, q ≤ ∞, with
1
det A111/p′
det A211/p
(cid:13)(cid:13)σ A(cid:13)(cid:13)M p, q →M p, q . A
Proof Let f , g ∈ M1(Rd) and Φ ∈ M1(cid:0)R2d(cid:1) \ {0}. Then,
·
1
(cid:12)(cid:12)det(A12)#(cid:12)(cid:12)1/q′(cid:12)(cid:12)det(A22)#(cid:12)(cid:12)1/q
kσk M ∞,1 .
(47)
(cid:12)(cid:12)(cid:10)σ A f , g(cid:11)(cid:12)(cid:12) = hσ, BA (g, f )i
= hVΦσ, VΦBA (g, f )i
≤ kVΦσk L∞,1 kVΦBA (g, f )k L1, ∞ ,
where in the last line we used Hölder inequality for mixed-norm Lebesgue spaces.
Let us choose for instance Φ = BAϕ where ϕ ∈ M1(Rd) is the Gaussian function.
We introduce the affine transformations
Pζ (z1, z2) = P1z + P2ζ =(cid:18) A11
Qζ (z1, z2) = Q1 z + Q2ζ =(cid:18) A21
0 (A12)#(cid:19)(cid:18) z1
0 − (A22)#(cid:19)(cid:18) z1
z2(cid:19) +(cid:18)
z2(cid:19) +(cid:18)
0
0
0
−A12
(A11)#
0
− (A21)#
ζ2(cid:19) ,
0 (cid:19)(cid:18) ζ1
ζ2(cid:19) ,
0 (cid:19)(cid:18) ζ1
−A22
in according with the magic formula (28), and using again Hölder's inequality we
get
32
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
kVΦBA (g, f )k L1, ∞ =
≤
=
p, q
z
sup
sup
(ζ1,ζ2)∈R2dÞR2d(cid:12)(cid:12)Vϕ g(cid:0)Pζ (z1, z2)(cid:1)(cid:12)(cid:12)(cid:12)(cid:12)Vϕ f (cid:0)Qζ (z1, z2)(cid:1)(cid:12)(cid:12) dz1dz2
(ζ1,ζ2)∈R2d(cid:13)(cid:13)(cid:0)Vϕ f(cid:1) ◦ Qζ(cid:13)(cid:13)L
det A211/p(cid:12)(cid:12)det (A22)#(cid:12)(cid:12)1/q
det A211/p(cid:12)(cid:12)det (A22)#(cid:12)(cid:12)1/q
(cid:13)(cid:13)(cid:0)Vϕ g(cid:1) ◦ Pζ(cid:13)(cid:13)L
(cid:13)(cid:13)Vϕ g(cid:13)(cid:13)L p′, q′
det A111/p′(cid:12)(cid:12)det (A12)#(cid:12)(cid:12)1/q′
det A11 1/p′(cid:12)(cid:12)det (A12)#(cid:12)(cid:12)1/q′ ,
(cid:13)(cid:13)Vϕ f(cid:13)(cid:13)L p, q
kgk M p′, q′
k f k M p, q
p′, q′
z
≤ C
where the constant C does not depend on f , g or A.
On the other hand,
kVΦσk L∞,1 ≤ CΦ kσk M ∞,1 ,
(48)
where the constant CΦ depends on Φ = BAϕ, hence on A. We conclude by duality
and get the claimed result.
Remark 6 This result broadly generalizes [26, Thm. 14.5.2] and confirms again that
the Sjöstrand's class is a well-suited symbol class leading to bounded operators on
modulation spaces. It is worth to mention that the left-regularity assumption for A
covers any Cohen-type matrix A = AM.
Remark 7 Unfortunately, it is not easy to sharpen the estimate (47) neither in the case
of Cohen-type matrices, the main obstruction being the estimate in (48). The usual
strategy consists of finding a suitable alternative window function Ψ ∈ M1(R2d)
and then estimate kVΨBAφk L1, in order to apply [26, Lem. 11.3.3] and [26, Prop.
11.1.3(a)]. This require cumbersome computations even in the case of τ-distributions
with Ψ Gaussian function, but the result is a uniform estimate for any τ ∈ [0, 1], cf.
[16, Lem. 2.3].
We now prove a similar boundedness result on W(F L p, Lq) spaces. We require
a very special case of the symplectic covariance for the Weyl quantization which is
stable under matrix perturbations - cf. [15, Lem. 5.1] for the τ-Wigner case.
Lemma 5 For any symbol σ ∈ M ∞(cid:0)R2d(cid:1) and any Cohen-type matrix A = AT ∈
GL (2d, R):
F opAT
(σ) F −1
= opAI −T (cid:16)σ ◦ J−1(cid:17) .
Proof We use a formal argument and leave to the reader the discussion on the
function spaces on which it is well defined. Recall the spreading representation of
the operator opAT
from (46)
opAT
f (x) =ÞR2dbσ (ξ, u) e−2πi(I −T )u ·ξT−u Mξ f (x) dudξ.
Since F T−u Mξ F −1 = e2πiu ·ξ Tξ Mu, we obtain
Linear perturbations of the Wigner transform and the Weyl quantization
33
F opAT
F −1 =ÞR2dbσ (ξ, u) e−πi(2T −I )u ·ξTξ Mu dudξ
= opAI −T (cid:16)σ ◦ J−1(cid:17) .
Remark 8 A comprehensive account of the symplectic covariance for perturbed rep-
resentation is out of the scope of this paper. This would require to investigate how the
metaplectic group should be modified in order to accommodate the perturbations.
As a non-trivial example of this issue, we highlight the contribution of de Gosson in
[18] for τ-pseudodifferential operators.
Theorem 21 For any Cohen-type matrix A = AT ∈ GL (2d, R) and any symbol
σ ∈ M ∞,1(cid:0)R2d(cid:1), the operator opA(σ) is bounded on W (F L p, Lq) (Rd) with
kopA(σ)kW ( F L p, L q )→W (F L p, L q ) ≤ CAkσk∞,1
M ,
for a suitable CA > 0.
Proof The proof follows the pattern of [15, Thm. 5.6]. Set σJ = σ ◦ J and consider
the following commutative diagram:
M p,q (Rd)
opI −T (σJ )
/ M p,q (Rd)
F−1
F
W (F L p, Lq) (Rd)
opτ (σ)
W (F L p, Lq) (Rd)
From the formula VG(σJ )(z, ζ ) = VG◦J −1 σ(J z, Jζ ), for any suitable window G,
it follows easily that σ ∈ M ∞,1(cid:0)R2d(cid:1) implies σJ ∈ M ∞,1(cid:0)R2d(cid:1). The operator
opI −T (σJ ) is bounded on M p,q (Rd) as a consequence of Theorem 20 and the claim
follows at once thanks to the previous lemma.
(cid:3)
The significance of the Sjöstrand class as space of symbols comes in many shapes.
In [29] an alternative characterization of the Sjöstrand class was given in terms of a
quasi-diagonalization property satisfied by the Weyl operators with symbol in M ∞,1.
Let us briefly recall the main ingredients of this result. First, fix a non-zero window
ϕ ∈ M1(Rd) and a lattice Λ = AZ2d ⊆ R2d, where A ∈ GL(2d, R), such that
G (ϕ, Λ) is a Gabor frame for L2(Rd). The action of pseudodifferential operators on
time-frequency shifts is described by the entries of the so-called channel matrix, that
is
hopW(σ)π(z)ϕ, π(w)ϕi,
z, w ∈ R2d,
or
M(σ)λ,µ ≔ hopW(σ)π(λ)ϕ, π(µ)ϕi,
λ, µ ∈ Λ,
if we restrict to the discrete lattice Λ. We say that opW is almost diagonalized by
the Gabor frame G(ϕ, Λ) if the associated channel matrix exhibits some sort of off-
diagonal decay - equivalently, if the time-frequency shifts are almost eigenvectors
for opW.
/
/
/
O
O
34
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
Theorem 22 ([29]) Let ϕ ∈ M1(Rd)(Rd) be a non-zero window function such that
G (ϕ, Λ) be a Gabor frame for L2(Rd). The following properties are equivalent:
(i) σ ∈ M ∞,1(R2d).
(ii) σ ∈ M ∞(R2d) and there exists a function H ∈ L1(R2d) such that
hopW (σ) π (z) ϕ, π (w) ϕi ≤ H (w − z) ,
∀w, z ∈ R2d .
(iii) σ ∈ M ∞(cid:0)R2d(cid:1) and there exists a sequence h ∈ ℓ1 (Λ) such that
hopW (σ) π (µ) ϕ, π (λ) ϕi ≤ h (λ − µ) ,
∀λ, µ ∈ Λ.
Corollary 2 Under the hypotheses of the previous theorem, assume that T :
M1(Rd) → M ∞(Rd) is continuous and satisfies one of the following conditions:
(i) hT π (z) ϕ, π (w) ϕi ≤ H (w − z) , ∀w, z ∈ R2d for some H ∈ L1.
(ii) hT π (µ) ϕ, π (λ) ϕi ≤ h (λ − µ) , ∀λ, µ ∈ Λ for some h ∈ ℓ1.
Then T = opW (σ) for some symbol σ ∈ M ∞,1(cid:0)R2d(cid:1) .
The backbone of this result is the interplay between the entries of the channel
matrix of opW and the short-time Fourier transform of the symbol. Theorem 22
can be extended without difficulty to τ-pseudodifferential operators [15]. We now
indicate a further generalization to operators associated with Cohen-type matrices.
Assume that A = AM ∈ GL (2d, R) is a matrix of Cohen's type. The following
result easily follows from the covariance formula (38).
Lemma 6 Let A = AT ∈ GL (2d, R) be a matrix of Cohen's type and fix a non-zero
window ϕ ∈ M1(Rd), then set ΦA = WT ϕ. Then, for any σ ∈ M ∞(cid:0)R2d(cid:1)
(cid:12)(cid:12)(cid:10)σ Aπ (z) ϕ, π (w) ϕ(cid:11)(cid:12)(cid:12) =(cid:12)(cid:12)VΦ A σ (TT (w, z) , J (w − z))(cid:12)(cid:12) =(cid:12)(cid:12)VΦ A σ (x, y)(cid:12)(cid:12) ,
(cid:12)(cid:12)VΦ A σ (x, y)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:10)σ Aπ (z (x, y)) ϕ, π (w (x, y)) ϕ(cid:11)(cid:12)(cid:12) ,
for any x, y, z, w ∈ R2d, where TT is defined in (40) and
and
z (x, y) = x + (I + PT ) J y,
w (x, y) = x + PT J y.
(49)
Proof We have
(cid:12)(cid:12)(cid:10)σ Aπ (z) ϕ, π (w) ϕ(cid:11)(cid:12)(cid:12) = hσ, BA (π (w) ϕ, π (z) ϕ)i
=(cid:12)(cid:12)(cid:10)σ, MJ(w−z)TTT (w,z)BAϕ (x, ω)(cid:11)(cid:12)(cid:12)
=(cid:12)(cid:12)VΦ A σ (TT (w, z) , J (w − z))(cid:12)(cid:12) .
Now, setting x = TT (w, z) and y = J (w − z) we immediately get Eq. (49).
(cid:3)
Theorem 23 Let ϕ ∈ M1(Rd) be a non-zero window function and assume that Λ is
a lattice such that G (ϕ, Λ) is a Gabor frame for L2(Rd). For any Cohen-type matrix
A = AT ∈ GL (2d, R), the following properties are equivalent:
Linear perturbations of the Wigner transform and the Weyl quantization
35
(i) σ ∈ M ∞,1(cid:0)R2d(cid:1).
(ii) σ ∈ M ∞(cid:0)R2d(cid:1) and there exists a function H = HT ∈ L1(Rd) such that
(iii) σ ∈ M ∞(cid:0)R2d(cid:1) and there exists a sequence h = hT ∈ ℓ1 (Λ) such that
(cid:12)(cid:12)(cid:10)σ Aπ (z) ϕ, π (w) ϕ(cid:11)(cid:12)(cid:12) ≤ HT (w − z) ,
(cid:12)(cid:12)(cid:10)σ Aπ (µ) ϕ, π (λ) ϕ(cid:11)(cid:12)(cid:12) ≤ hT (λ − µ) ,
∀w, z ∈ R2d .
∀λ, µ ∈ R2d.
Proof The proof faithfully mirrors the one for Weyl operators [29, Theorem 3.2]. We
detail here only the case (i) ⇒ (ii), the discrete case for hT is similar. For ϕ ∈ M1(Rd)
the T-Wigner distribution ΦT = WT ϕ is in M1(R2d) by Theorem 12. This implies
(cf. [26, Theorem 11.3.7]). The main insight here is that the controlling function
HT ∈ L1(Rd) can be provided by the so-called grand symbol associated with σ,
that the short-time Fourier transform VΦT σ is well-defined for σ ∈ M ∞,1(cid:0)R2d(cid:1)
which is HT (v) = supu ∈R2d VΦT σ(u, v). By definition of M ∞,1(cid:0)R2d(cid:1), we have
HT ∈ L1(R2d), so that Lemma 6 implies
(cid:12)(cid:12)(cid:10)σ Aπ (z) ϕ, π (w) ϕ(cid:11)(cid:12)(cid:12) =(cid:12)(cid:12)VΦT σ (TT (z, w) , J (w − z))(cid:12)(cid:12)
u ∈R2d(cid:12)(cid:12)VΦT σ (u, J (w − z))(cid:12)(cid:12)
= HT (J (w − z)) .
≤ sup
Setting HT = HT ◦ J yields the claim.
(cid:3)
Let us further discuss the main trick of the proof. While the choice of the grand
symbol as controlling function is natural in view of the M ∞,1 norm, the effect of the
perturbation matrix T is confined to the window function ΦT and to the time variable
of the short-time Fourier transform of the symbol. A natural question is then the
following: what happens if we try to control the time dependence of VΦτ σ? Following
the pattern of [15, Thm. 4.3], this remark provides a similar characterization for
symbols belonging to the modulation space W(F L∞, L1) = F M ∞,1.
Theorem 24 Let ϕ ∈ M1(Rd) be a non-zero window function. For any right-regular
Cohen-type matrix A = AT ∈ GL (2d, R), the following properties are equivalent:
(i) σ ∈ W(F L∞, L1)(cid:0)R2d(cid:1).
(ii) σ ∈ M ∞(cid:0)R2d(cid:1) and there exists a function H = HT ∈ L1(Rd) such that
∀w, z ∈ R2d,
(cid:12)(cid:12)(cid:10)σ Aπ (z) ϕ, π (w) ϕ(cid:11)(cid:12)(cid:12) ≤ HT (w − UT z),
where
UT = −(cid:18) (I − T )−1 T
0
0
T −1 (I − T )(cid:19) .
Proof The proof is a straightforward adjustment of the one provided for [15, Thm.
4.3]. Again, we detail here only the case (i) ⇒ (ii) for the purpose of tracking the
36
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
origin of UT . If φ ∈ M1(Rd), φ , 0, then ΦT = WT φ ∈ M1 = W(F L1, L1) by
Theorem 12. For σ ∈ W(cid:0)F L∞, L1(cid:1), we have that VΦT σ is well defined and
HT (x) = sup
y ∈R2d(cid:12)(cid:12)VΦT σ (x, y)(cid:12)(cid:12) ∈ L1(cid:16)R2d(cid:17) .
From Lemma 6 we infer
(cid:12)(cid:12)(cid:10)opT (σ) π (z) ϕ, π (w) ϕ(cid:11)(cid:12)(cid:12) =(cid:12)(cid:12)VΦT σ (Tt (w, z) , J (w − z))(cid:12)(cid:12)
y ∈R2d(cid:12)(cid:12)VΦT σ (TT (w, z) , y)(cid:12)(cid:12)
= HT (TT (w, z)) .
≤ sup
Notice that if AT is right-regular, then I + PT from (39) is invertible (and the converse
holds, too). In particular, we have
(I + PT )−1 (TT (w, z)) = w − (I + PT )−1PT z = w − Uτ z,
and thus HT (TT (w, z)) = HT (cid:0)B−1
T (w − UT z)(cid:1). Define HT = HT ◦ (I + PT )−1; then
HT ∈ L1(R2d) since kHT k L1 = k HT ◦ (I + PT )−1kL1 ≍ k HT kL1 < ∞.
(cid:3)
Remark 9 (i) The almost diagonalization of the (continuous) channel matrix does
not survive the perturbation, but the new result can be interpreted as a measure
of the concentration of the time-frequency representation of opT (σ) along the
graph of the map UT .
(ii) We recover [15, Thm. 4.3], where the same problem was first studied for τ-
operators with τ ∈ (0, 1). One may push further the analogy and generalize the
results for τ = 0 or τ = 1.
(iii) As already remarked for τ-operators in [15], the discrete characterization via
Gabor frames is lost: for a given lattice Λ, the inclusion Uτ Λ ⊆ Λ exclusively
holds for τ = 1/2 (the Weyl transform).
(iv) The boundedness and algebraic properties of AT -operators with symbols in
W(F L∞, L1) proved in [15] rely on the characterization as generalized meta-
plectic operators according to [12, Def. 1.1]. In order to benefit from this frame-
work, it is necessary for UT to be a symplectic matrix, the latter condition being
realized if and only if T is a symmetric matrix (cf. [17, (2.4) and (2.5)] and
notice that T −1(I − T ) = (I − T )T −1).
Acknowledgments
E. Cordero and S. I. Trapasso are members of the Gruppo Nazionale per l'Analisi
Matematica, la Probabilità e le loro Applicazioni (GNAMPA) of the Istituto
Nazionale di Alta Matematica (INdAM). K. Gröchenig acknowledges support from
the Austrian Science Fund FWF, project P31887-N32.
Linear perturbations of the Wigner transform and the Weyl quantization
37
References
1. Bayer, D.: Bilinear Time-Frequency Distributions and Pseudodifferential Operators. PhD The-
sis, University of Vienna (2010)
2. Bényi, A., Gröchenig, K., Okoudjou, K., and Rogers, L. G.: Unimodular Fourier multipliers
for modulation spaces. J. Funct. Anal. 246 (2007), no. 2, 366 -- 384
3. Boggiatto, P., Carypis, E., and Oliaro, A.: Wigner representations associated with linear trans-
formations of the time-frequency plane. In Pseudo-Differential Operators: Analysis, Applica-
tions and Computations (275-288), Springer (2011)
4. Boggiatto, P., De Donno, G., and Oliaro, A.: Weyl quantization of Lebesgue spaces. Math.
Nachr. 282 (2009), no. 12, 1656 -- 1663
5. Boggiatto, P., De Donno, G., and Oliaro, A.: Time-frequency representations of Wigner type
and pseudo-differential operators. Trans. Amer. Math. Soc. 362 (2010), no. 9, 4955 -- 4981
6. Cohen, L.: Time-frequency distributions -- A review. Proc. IEEE 77 (1989), no. 7, 941 -- 981
7. Cohen, L.: Time-frequency Analysis. Prentice Hall (1995)
8. Cohen, L.: Generalized phase-space distribution functions. J. Math. Phys. 7 (1966), no. 5,
781 -- 786
9. Cohen, L.: The Weyl Operator and its Generalization. Springer (2012)
10. Cordero, E., and Nicola, F.: Metaplectic representation on Wiener amalgam spaces and appli-
cations to the Schrödinger equation. J. Funct. Anal. 254 (2008), no. 2, 506 -- 534
11. Cordero, E., and Trapasso, S. I.: Linear Perturbations of the Wigner Distribution and the Cohen
Class. Anal. Appl. - DOI: 10.1142/S0219530519500052 (2018)
12. Cordero, E., Gröchenig, K., Nicola, F., and Rodino, L.: Generalized metaplectic operators and
the Schrödinger equation with a potential in the Sjöstrand class. J. Math. Phys. 55 081506
(2014)
13. Cordero, E., de Gosson, M., and Nicola, F.: Time-frequency analysis of Born-Jordan pseudod-
ifferential operators. J. Funct. Anal. 272 (2017), no. 2, 577 -- 598
14. Cordero, E., de Gosson, M., Dörfler, M., and Nicola, F.: On the symplectic covariance and
interferences of time-frequency distributions. SIAM J. Math. Anal. 50 (2018), no. 2, 2178 -- 2193
15. Cordero, E., Nicola, F., and Trapasso, S. I.: Almost diagonalization of τ-pseudodifferential
operators with symbols in Wiener amalgam and modulation spaces. J. Fourier Anal. Appl. -
DOI: 10.1007/s00041-018-09651-z (2018)
16. Cordero, E., D'Elia, L., and Trapasso, S. I.: Norm estimates for τ-pseudodifferential operators
in Wiener amalgam and modulation spaces. J. Math. Anal. Appl. 471 (2019), no. 1-2, 541 -- 563
17. de Gosson, M.: Symplectic Methods in Harmonic Analysis and in Mathematical Physics.
Springer (2011)
18. de Gosson, M.: Born-Jordan quantization. Fundamental Theories of Physics, Vol. 182, Springer
[Cham], (2016)
19. Feichtinger, H. G.: On a new Segal algebra. Monatsh. Math. 92 (1981), no. 4, 269âĂŞ-289
20. Feichtinger, H. G.: Modulation spaces on locally compact abelian groups, Technical Report,
University Vienna, (1983) and also in Wavelets and Their Applications, M. Krishna, R. Radha,
S. Thangavelu, editors, Allied Publishers (2003), 99 -- 140.
21. Feichtinger, H. G.: Generalized amalgams, with applications to Fourier transform, Canad. J.
Math., 42 (1990), 395 -- 40
22. Feichtinger, H. G., and Gröchenig, K.: Gabor frames and time-frequency analysis of distribu-
tions. J. Funct. Anal. 146 (1997), no. 2, 464 -- 495.
23. Feig, E., and Micchelli, C. A.: L2-synthesis by ambiguity functions. In Multivariate Approx-
imation Theory IV, 143 -- 156, International Series of Numerical Mathematics. Birkhäuser,
Basel, 1989.
24. Goh, S. S., and Goodman, T. N.: Estimating maxima of generalized cross ambiguity functions,
and uncertainty principles. Appl. Comput. Harmon. Anal. 34 (2013), no. 2, 234 -- 251.
25. Gröchenig, K.: An uncertainty principle related to the Poisson summation formula. Studia
Math. 121 (1996), no. 1, 87 -- 104.
38
Dominik Bayer, Elena Cordero, Karlheinz Gröchenig, and S. Ivan Trapasso
26. Gröchenig, K.: Foundations of Time-frequency Analysis. Appl. Numer. Harmon. Anal.,
Birkhäuser (2001)
27. Gröchenig, K.: Uncertainty principles for time-frequency representations. In Advances in
Gabor analysis, 11 -- 30, Appl. Numer. Harmon. Anal., Birkhäuser Boston, Boston, MA, 2003
28. Gröchenig, K.: A pedestrian's approach to pseudodifferential operators. In Harmonic analysis
and applications, 139 -- 169, Appl. Numer. Harmon. Anal., Birkhäuser Boston, Boston, MA,
2006
29. Gröchenig, K.: Time-frequency analysis of Sjöstrand's class. Rev. Mat. Iberoam. 22 (2006),
no. 2, 703 -- 724
30. Gröchenig, K., and Rzeszotnik, Z.: Banach algebras of pseudodifferential operators and their
almost diagonalization. Ann. Inst. Fourier (Grenoble) 58 (2008), no. 7, 2279 -- 2314
31. Gröchenig, K., and Strohmer, T.: Pseudodifferential operators on locally compact abelian
groups and Sjöstrand's symbol class. J. Reine Angew. Math. 613 (2007), 121 -- 146
32. Hlawatsch, F., and Auger, F. (Eds.).: Time-frequency Analysis. John Wiley & Sons (2013)
33. Hlawatsch, F., and Boudreaux-Bartels, G. F.: Linear and quadratic time-frequency signal
representations. IEEE Signal Proc. Mag. 9 (1992), no. 2, 21 -- 67
34. Hudson, R. L.: When is the Wigner quasi-probability density non-negative? Rep. Mathematical
Phys. 6 (1974), no. 2, 249 -- 252
35. Jakobsen, M. S.: On a (no longer) new Segal algebra: a review of the Feichtinger algebra. J.
Fourier Anal. Appl. 24 (2018), no. 6, 1579 -- 1660
36. Janssen, A. J. E. M.: A note on Hudson's theorem about functions with nonnegative Wigner
distributions. SIAM J. Math. Anal. 15 (1984), no. 1, 170 -- 176
37. Janssen, A. J. E. M.: Bilinear time-frequency distributions. In Wavelets and their applications
(Il Ciocco, 1992), 297 -- 311, Kluwer Acad. Publ., Dordrecht, 1994
38. Janssen, A. J. E. M.: Positivity and spread of bilinear time-frequency distributions. In The
Wigner distribution, 1 -- 58, Elsevier Sci. B. V., Amsterdam, 1997
39. Lu, T., and Shiou, S.: Inverses of 2 × 2 block matrices. Comput. Math. Appl. 43 (2002), no.
1-2, 119 -- 129
40. W. Mecklenbräuker and F. Hlawatsch, editors. The Wigner distribution. Theory and applica-
tions in signal processing, Elsevier Science B.V., Amsterdam, 1997
41. Sjöstrand, J.: An algebra of pseudodifferential operators. Math. Res. Lett. 1 (1994), no. 2,
185 -- 192
42. Toft, J.: Continuity properties for modulation spaces, with applications to pseudo-differential
calculus. I. J. Funct. Anal., 207 (2004), no. 2, 399 -- 429
43. Toft, J.: Matrix parameterized pseudo-differential calculi on modulation spaces. In Generalized
Functions and Fourier Analysis, 215 -- 235, Birkhäuser, 2017
44. Wigner, E.: On the Quantum Correction For Thermodynamic Equilibrium. Phys. Rev., 40
(1932), no. 5, 749 -- 759
|
1803.04727 | 1 | 1803 | 2018-03-13T11:01:18 | Characterization of Banach spaces $Y$ satisfying that the pair $ (\ell_\infty^4,Y )$ has the Bishop-Phelps-Bollob\'as property for operators | [
"math.FA"
] | We study the Bishop-Phelps-Bollob\'as property for operators from $\ell_\infty ^4 $ to a Banach space. For this reason we introduce an appropiate geometric property, namely the AHSp-$\ell_\infty ^4$. We prove that spaces $Y$satisfying AHSp-$\ell_\infty ^4$ are precisely those spaces $Y$ such that $(\ell_\infty^4,Y)$ has the Bishop-Phelps-Bollob\'as property. We also provide classes of Banach spaces satisfying this condition. For instance, finite-dimensional spaces, uniformly convex spaces, $C_0(L)$ and $L_1 (\mu)$ satisfy AHSp-$\ell_\infty ^4 $. | math.FA | math |
Characterization of Banach spaces Y satisfying that
the pair (ℓ4
∞, Y ) has the Bishop-Phelps-Bollob´as
property for operators
Mar´ıa D. Acosta, Jos´e L. D´avila, and Maryam Soleimani-Mourchehkhorti
Abstract. We study the Bishop-Phelps-Bollob´as property for operators from ℓ4
∞ to a Banach space. For this reason
∞. We prove that spaces Y satisfying AHSp-ℓ4
we introduce an appropiate geometric property, namely the AHSp-ℓ4
are precisely those spaces Y such that (ℓ4
∞, Y ) has the Bishop-Phelps-Bollob´as property. We also provide classes of
Banach spaces satisfying this condition. For instance, finite-dimensional spaces, uniformly convex spaces, C0(L) and
L1(µ) satisfy AHSp-ℓ4
∞
∞.
1. Introduction
Bishop-Phelps theorem [10] states that every continuous linear functional on a Banach space can be
approximated (in norm) by norm attaining functionals. Bollob´as proved a "quantitative version" of that
result [11]. In order to state such result, we denote by BX, SX and X ∗ the closed unit ball, the unit sphere
and the topological dual of a Banach space X, respectively.
If X and Y are both real or both complex
Banach spaces, L(X, Y ) denotes the space of (bounded linear) operators from X to Y , endowed with its
usual operator norm.
Bishop-Phelps-Bollob´as Theorem (see [12, Theorem 16.1], or [15, Corollary 2.4]). Let X be a Banach
space and 0 < ε < 1. Given x ∈ BX and x∗ ∈ SX ∗ with 1 − x∗(x) < ε2
y∗ ∈ SX ∗ such that y∗(y) = 1, ky − xk < ε and ky∗ − x∗k < ε.
4 , there are elements y ∈ SX and
A lot of attention has been devoted to extending Bishop-Phelps theorem to operators and
interesting results have been obtained about that topic. For instance, we mention the remarkable results by
Lindenstrauss
[23], Bourgain [13] and Gowers
[19].
In 2008 the
study of
extensions of
Bishop-Phelps-Bollob´as theorem to operators was initiated by Acosta, Aron, Garc´ıa and Maestre [2]. In
order to state some of these extensions it will be convenient to recall the following notion.
Definition 1.1. ([2, Definition 1.1]). Let X and Y be either real or complex Banach spaces. The pair
(X, Y ) is said to have the Bishop-Phelps-Bollob´as property for operators (BPBp) if for every 0 < ε < 1
The first author was supported by Junta de Andaluc´ıa grant FQM -- 185 and also by Spanish MINECO/FEDER grant MTM2015-
65020-P. The second author was also partially supported by Junta de Andaluc´ıa grant FQM -- 185. The third author was supported by a
grant from IPM.
1
2
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
there exists 0 < η(ε) < ε such that for every T ∈ SL(X,Y ), if x0 ∈ SX satisfies kT (x0)k > 1 − η(ε), then there
exist an element u0 ∈ SX and an operator S ∈ SL(X,Y ) satisfying the following conditions
kS(u0)k = 1,
ku0 − x0k < ε
and
kS − T k < ε.
Acosta, Aron, Garc´ıa and Maestre showed that the pair (X, Y ) has the BPBp whenever X and Y are
finite-dimensional spaces [2, Proposition 2.4]. They also proved that the pair (X, Y ) has the BPBp in case
that Y has a certain isometric property (called property β of Lindenstrauss), for every Banach space X
[2, Theorem 2.2]. For instance, the spaces c0 and ℓ∞ have such property. In case that the domain is ℓ1 they
obtained a characterization of the Banach spaces Y such that (ℓ1, Y ) has the BPBp [2, Theorem 4.1]. There
was also proved that for several classical spaces Y , and for any positive measure µ, the pair (L1(µ), Y ) have
the BPBp, for instance when Y = L∞(ν) or Y = L1(ν) (see [16], [8] and [17]). For those results the proofs
are involved and interesting. Even so, as far as we know, there are no characterization of the spaces Y such
that the pair (L1([0, 1]), Y ) has the BPBp.
However, the case of X = c0 is quite different from the case X = ℓ1 and seems to be much more
difficult. Now we will list results about this topic where the domain is a space C0(L) (L is a locally compact
Hausdorff space).
∞, Y ) has the BPBp for every positive integer n whenever Y is
uniformly convex [2, Theorem 5.2]. In fact Kim proved that (c0, Y ) also has this property under the same
It was shown that (ℓn
assumption [20, Corollary 2.6]. Aron, Cascales and Kozhushkina showed that (X, C0(L)) has the BPBp if
X is Asplund [7, Corollary 2.6] (see also [14] for an extension of this result). As a consequence, the pair
(c0, C0(L)) has the Bishop-Phelps-Bollob´as property for operators. In the real case, it is also known that
the pair (C(K), C(S)) also satisfies the BPBp, for any compact Hausdorff spaces K and S [3, Theorem 2.5].
Kim, Lee and Lin proved that the pair (L∞(µ), Y ) has BPBp for every positive measure µ, whenever Y is
uniformly convex [22, Theorem 5]. Kim and Lee extended such result for the pair (C(K), Y ) (K is compact
Hausdorff spaces) [21, Theorem 2.2]. In the complex case, Acosta proved that the pair (C0(L), Y ) has the
BPBp for every locally compact Hausdorff space L and any C-uniformly convex space Y [1, Theorem 2.4].
In the complex case L1(µ) is C-uniformly convex space, so the previous result can be applied. There are
some other sufficient conditions on a Banach space Y in order that the pair (c0, Y ) satisfies the BPBp (see
for instance [6, Theorem 2.4]).
However until now there is no characterization of the spaces Y such that (c0, Y ) has the BPBp. Indeed
in the real case it is not known whether or not the pair (c0, ℓ1) has the BPBp. As a consequence of
[9, Theorem 2.1], in case that the pair (c0, ℓ1) has the BPBp, then the pairs (ℓn
∞, ℓ1) satisfy the BPBp
"uniformly" for every n. For this reason in this paper we approach this problem by using in the domain
appropriate finite-dimensional spaces.
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
3
[2, Theorem 4.1],
it is known that (ℓ2
Notice that for dimension 2, since ℓ2
∞ is isometrically isomorphic to ℓ2
1, as a consequence of
∞, Y ) has the BPBp if and only if Y has the approximate
hyperplane series property for convex combinations of two elements. A characterization of the spaces Y
such that (ℓ3
∞, Y ) has the BPBp was shown in [5, Theorem 2.9]. As a consequence of that result, classical
Banach spaces satisfying the previous property were provided. The goal of this paper is to extend the above
mentioned results for the pair (ℓ4
∞, Y ).
Now we briefly describe the content of the paper. In section 2 we introduce a geometric property on a
Banach space Y , namely the AHSp-ℓ4
∞ (see Definition 2.3). We also provide several reformulations of such
property in Proposition 2.12. That result is essential to prove in section 3 that the pair (ℓ4
∞, Y ) has the
BPBp for operators if and only if Y has the AHSp-ℓ4
∞ (see Theorem 3.3). In section 4 we provide examples of
classical spaces satisfying the AHSp-ℓ4
∞. For instance, we check that finite-dimensional spaces and uniformly
convex spaces have this property. It is also satisfied that C0(L, Y ) has the AHSp-ℓ4
∞ whenever Y has the
same property, for any locally compact Hausdorff space L (Proposition 4.4). Lastly we prove that ℓ1 has the
AHSp-ℓ4
for ℓ1, we obtain that (ℓ4
∞ (Proposition 4.6). The proof of that result requires some effort. As a consequence of the result
∞, L1(µ)) has the Bishop-Phelps-Bollob´as property for any positive measure µ.
Throughout this paper we follow the spirit of the results obtained in [5] for ℓ3
much more complicated. We provide a few arguments for that. One reason is that for ℓ3
∞, but the proofs are
∞ any subset of
three extreme points contained in the same face of the unit ball can be applied to any other subset of the
same kind by using an appropriate linear isometry. For dimension 4 the previous statement is not satisfied.
Because of that the image under an operator of any three extreme points in the same face of Bℓ3
choice in order to identify an operator whose domain is ℓ3
∞. In case that the domain of the operator is
ℓ4
∞ we had to find an appropriate choice of the basis in order to identify such operator and to compute its
norm (Proposition 2.11). Another reason is that due to the bigger amount of extreme points in the unit ball
of ℓ4
the property of Y equivalent to the fact that (ℓ4
∞ is more involved. As a consequence,
∞, Y ) has the BPBp is more complicated. That reason also
∞ the description used to identify operators whose domain is ℓ4
is a good
∞
makes to provide examples more difficult.
Throughout the paper we consider real normed spaces. By ℓ4
∞ we denote the space R4, endowed with
the norm given by kxk = max{xi : i ≤ 4}.
2. The approximate hyperplane sum property for ℓ4
∞
In this section we identify the unit ball of L(ℓ4
(Proposition 2.11). We also introduce an intrinsic property on a Banach space Y , namely the AHSp-ℓ4
∞, Y ) with a certain family of elements in Y 4 called M 4
Y
∞ and
show a characterization of that property (see Proposition 2.12).
4
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
Notation 2.1. If Y is a Banach space, we will denote by
M 4
Y = {(yi)i≤4 ∈ (BY )4 : yi1 − yi2 + yi3 ∈ BY , ∀1 ≤ i1 < i2 < i3 ≤ 4}.
Remark 2.2. It is clear that (yi)i≤4 ∈ M 4
Y if (−yi)i≤4 ∈ M 4
Y .
The following notion is analogous to the AHSp-ℓ3
∞ that was used to characterize those spaces Y such
that the pair (ℓ3
∞, Y ) has the BPBp for operators (see [5, Definition 2.1]).
Definition 2.3. A Banach space Y has
the approximate hyperplane sum property for ℓ4
∞
(AHSp-ℓ4
∞) if for every 0 < ε < 1 there is 0 < γ(ε) < ε satisfying the following condition
For every (yi)i≤4 ∈ M 4
y∗(yi) > 1 − γ(ε) for each i ∈ A, then there exists an element (zi)i≤4 ∈ M 4
Y , if there exist a nonempty subset A of {1, 2, 3, 4} and y∗ ∈ SY ∗ such that
Y satisfying kzi − yik < ε for every
i ≤ 4 and kPi∈A zik = A.
Remark 2.4. We recall that Y has the AHSp-ℓ3
∞ if for every ε > 0 there is γ > 0 satisfying the following
condition:
For a subset {yi : i ≤ 3} ⊂ BY with ky1 +y2 −y3k ≤ 1 , if there exist a nonempty subset A of {1, 2, 3} and
y∗ ∈ SY ∗ such that y∗(yi) > 1 − γ for every i ∈ A, then there exists {zi : i ≤ 3} ⊂ BY with kz1 + z2 − z3k ≤ 1
satisfying kzi − yik < ε for every i ≤ 3 and kPi∈A zik = A.
If
(yi)i≤3
(y1, y3, y2, y2) ∈ M 4
satisfies
the assumption in the definition of
Y . This is the key idea to check that AHSp-ℓ4
the AHSp-ℓ3
∞,
∞ implies AHSp-ℓ3
∞.
it
is immediate that
Now we state some basic but useful results.
Lemma 2.5. If {yi : 1 ≤ i ≤ 4} ⊂ BY , the following conditions are equivalent
(1) (y1, y2, y3, y4) ∈ M 4
Y
(2) (y2, y3, y4, −y1) ∈ M 4
Y
(3) (y3, y4, −y1, −y2) ∈ M 4
Y
(4) (y4, −y1, −y2, −y3) ∈ M 4
Y
Proof. Statement (1) is satisfied exactly when the following elements belong to BY
(2.1)
y1 − y2 + y3, y1 − y2 + y4, y1 − y3 + y4, y2 − y3 + y4.
On the other hand, condition (2) means that each of the following elements belongs to BY
(2.2)
y2 − y3 + y4, y2 − y3 − y1, y2 − y4 − y1, y3 − y4 − y1.
As a consequence conditions (1) and (2) are equivalent. By applying this fact we obtain that (2) and (3)
are equivalent. Again by the same argument (3) and (4) are equivalent.
(cid:3)
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
5
Next result shows that the condition stated in Definition 2.3 is trivially satisfied in case that the set A
contains a unique element.
Proposition 2.6. Let Y be a Banach space. Let 0 < ε < 1 and (yi)i≤4 ∈ M 4
Y . Assume that
A ⊂ {1, 2, 3, 4} contains only one element and it is satisfied that
Then there is an element (zi)i≤4 ∈ M 4
Y such that
kyjk > 1 −
ε
6
,
j ∈ A.
kzjk = 1 for j ∈ A and
kzi − yik < ε, ∀1 ≤ i ≤ 4.
Proof. Let 0 < ε < 1 and (yi)i≤4 ∈ M 4
Y . In view of Lemma 2.5 it suffices to show the statement in
case that A = {1}.
Assume that the element y1 satisfies that
ky1k > 1 −
ε
6
> 0.
We define the following real numbers
and the elements in Y
a =
ε
3
,
a2 =
ε
3
,
a3 =
ε
6
,
a4 = −
ε
6
,
z1 =
y1
ky1k
and
zi = (1 − a)yi + aiz1,
i ∈ {2, 3, 4}.
It is trivially satisfied that z1 ∈ SY . We also have that
(2.3)
kz1 − y1k = 1 − ky1k <
In case that 2 ≤ i ≤ 4 we obtain that
kzik ≤ (1 − a) + ai ≤ 1 −
ε
6
< ε.
ε
3
+
ε
3
= 1
and
2
3
It remains to show that for every 1 ≤ i1 < i2 < i3 ≤ 4 it is satisfied that
kzi − yik = k − ayi + aiz1k ≤ a + ai ≤
ε < ε.
In case that {i1, i2, i3} = {2, 3, 4} we obtain
kzi1 − zi2 + zi3k ≤ 1.
kz2 − z3 + z4k = k(1 − a)(y2 − y3 + y4) + (a2 − a3 + a4)z1k = k(1 − a)(y2 − y3 + y4)k ≤ 1 − a < 1.
Otherwise i1 = 1. Notice that for every 1 < i2 < i3 ≤ 4 we have that
a − ai2 + ai3 ≤
ε
6
.
6
Hence
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
kz1 − zi2 + zi3k = k(1 − a)(z1 − yi2 + yi3) + (a − ai2 + ai3)z1k
≤ (1 − a)(1 + kz1 − y1k) + a − ai2 + ai3
(by (2.3))
ε
3(cid:17)(cid:16)1 +
ε
6(cid:17) +
+
−
+
ε
6
ε2
18
ε
6
ε
6
<(cid:16)1 −
= 1 −
= 1 −
< 1.
ε
3
ε2
18
We checked that (zi)i≤4 ∈ M 4
Y and it satisfies all the required conditions.
(cid:3)
Notation 2.7. In what follows we will denote by E1 the subset of Bℓ4
∞
given by
E1 = {x ∈ ℓ4
∞ : 1 = x(1) = kxk}.
In the sequel it will be convenient to use the following notation for the following elements in E1
v1 = (1, 1, 1, 1), v2 = (1, −1, 1, 1), v3 = (1, −1, −1, 1), v4 = (1, −1, −1, −1),
v5 = (1, 1, −1, 1), v6 = (1, 1, −1, −1), v7 = (1, 1, 1, −1), v8 = (1, −1, 1, −1),
By B we will denote the set given by B = {vi : 1 ≤ i ≤ 4}.
The proofs of next assertions are straightforward. For the first one it suffices to check that every
coordinate of an extreme point of E1 belongs to {1, −1}.
Lemma 2.8. It is satisfied that
Lemma 2.9. The set B is a basis of R4 contained in Sℓ4
x(i + 1) − x(i)
x(1) + x(2)
∞
v∗
1(x) =
2
, v∗
i (x) =
2
Ext(cid:0)E1(cid:1) = {vi : i ≤ 8}.
. Moreover the functionals given by
, i = 2, 3 and v∗
4(x) =
x(1) − x(4)
2
,
are elements in S(ℓ4
expressed as
∞)∗ that are the biorthogonal functionals of the basis B. Hence each x in R4 can be
x =
x(1) + x(2)
2
v1 +
3
Xi=2
x(i + 1) − x(i)
2
vi +
x(1) − x(4)
2
v4.
As a consequence of the last assertion in the previous result we obtain the following.
Remark 2.10. The following equalities are satisfied
v5 = v1 − v2 + v3,
v6 = v1 − v2 + v4,
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
7
and
v7 = v1 − v3 + v4,
v8 = v2 − v3 + v4.
Hence (vi)i≤4 ∈ M 4
Y for Y = ℓ4
∞.
The next result shows the connection between operators from ℓ4
∞ to Y and the set M 4
Y .
Proposition 2.11. Every element T ∈ L(ℓ4
∞, Y ) satisfies
So the mapping given by Φ(T ) = (T (vi))i≤4 identifies BL(ℓ4
kT k = max{(cid:13)(cid:13)T (vi1 − vi2 + vi3)(cid:13)(cid:13) : 1 ≤ i1 ≤ i2 ≤ i3 ≤ 4}.
∞,Y ) with M 4
Y .
Proof. In view of Lemma 2.9 the set B = {vi : i ≤ 4} is a basis of R4, so every operator T from ℓ4
∞ to
Y is determined by the element (T (vi))i≤4.
It is clear Ext(cid:0)Bℓ4
∞(cid:1) = Ext(cid:0)E1(cid:1) ∪ Ext(cid:0)−E1(cid:1). Hence
kT k = max{kT (e)k : e ∈ Ext(cid:0)Bℓ4
∞(cid:1)}
= max{kT (e)k : e ∈ Ext(E1)}
(2.4)
= max{kT (vi)k : i ≤ 8},
where we used Lemma 2.8.
In view of Remark 2.10 we have that
{vi : 5 ≤ i ≤ 8} = {vi1 − vi2 + vi3 : 1 ≤ i1 < i2 < i3 ≤ 4}.
From (2.4) and the previous equality we obtain that
It follows that T ∈ BL(ℓ4
kT k = max{(cid:13)(cid:13)T (vi1 − vi2 + vi3)(cid:13)(cid:13) : 1 ≤ i1 ≤ i2 ≤ i3 ≤ 4}.
∞,Y ) if and only if Φ(T ) is an element in M 4
Y .
(cid:3)
We recall that a subset B ⊂ BY ∗ is 1-norming if kyk = sup{y∗(y) : y∗ ∈ B} for each y ∈ Y . Now we
provide a characterization of AHSp-ℓ4
∞ which will be used in the following sections.
Proposition 2.12. Let Y be a Banach space. The following conditions are equivalent:
1) Y has the approximate hyperplane sum property for ℓ4
∞.
2) There is a 1-norming subset B ⊂ SY ∗ such that the condition stated in Definition 2.3 is satisfied
for every y∗ ∈ B.
3) For every 0 < ε < 1 there exists 0 < ν(ε) < ε such that for every element (yi)i≤4 ∈ M 4
Y and each
convex combination P4
i=1 αiyi satisfying
> 1 − ν(ε),
4
Xi=1
(cid:13)(cid:13)(cid:13)
αiyi(cid:13)(cid:13)(cid:13)
8
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
there exist a set A ⊂ {1, 2, 3, 4} and an element (zi)i≤4 ∈ M 4
Y such that
ii) kzi − yik < ε
i) Pi∈A αi > 1 − ε,
iii) kPi∈A zik = A.
for each i ≤ 4,
Moreover, if ρ is the function satisfying condition 2), then condition 3) is also satisfied with the function
ν = ρ2. In case that 3) is satisfied with a function ν, Y has the AHSp-ℓ4
∞ with the function γ(ε) = ν( ε
4 ).
Proof. Clearly 1) implies 2).
2) ⇒ 3)
Assume that Y satisfies condition 2). For each 0 < ε < 1 let be ρ(ε) < ε the positive real number satisfying
Definition 2.3 for every element y∗ ∈ B. We take ν(ε) = (ρ(ε))2.
Let (yi)i≤4 ∈ M 4
Since (−yi)i≤4 ∈ M 4
such that
Y and assume that the convex combination P4
Y , by using (−yi) instead of (yi), if needed, since B is a 1-norming set, there is y∗ ∈ B
i=1 αiyi satisfies (cid:13)(cid:13)P4
i=1 αiyi(cid:13)(cid:13) > 1 − ν(ε).
y∗(cid:18) 4
Xi=1
αiyi(cid:19) =
4
Xi=1
αiy∗(yi) > 1 − ν(ε) = 1 − ρ(ε)2.
By [2, Lemma 3.3] the set A given by A := {i ≤ 4 : y∗(yi) > 1 − ρ(ε)} satisfies
αi ≥ 1 −
ν(ε)
ρ(ε)
> 1 − ε.
Xi∈A
By assumption there is an element (zi)i≤4 ∈ M 4
3) ⇒ 1)
Y such that kzi − yik < ε for each i ≤ 4 and kPi∈A zik = A.
Now we assume that Y satisfies condition 3). Let be 0 < ε < 1 and ν(ε) the positive real number satisfying
the assumption. We will show that γ(ε) = ν( ε
4 ) satisfies Definition 2.3.
Let (yi)i≤4 ∈ M 4
Y and assume that for some nonempty set A ⊂ {1, 2, 3, 4} and y∗ ∈ SY ∗ it is satisfied
that y∗(yi) > 1 − γ(ε) for each i ∈ A. We define the following nonnegative real numbers
αi =( 1
A
0
if i ∈ A
if i ∈ {1, 2, 3, 4}\A.
Clearly P4
i=1 αi = 1 and we also have that
4
Xi=1
(cid:13)(cid:13)(cid:13)
αiyi(cid:13)(cid:13)(cid:13)
= (cid:13)(cid:13)(cid:13)Pi∈A yi(cid:13)(cid:13)(cid:13)
A
y∗(cid:16)Pi∈A yi(cid:17)
A
≥
> 1 − ν(cid:16) ε
4(cid:17).
By assumption there is a set C ⊂ {1, 2, 3, 4} and (zi)i≤4 ∈ M 4
Y such that
4 > 3
4 ,
for each i ≤ 4,
ii) kzi − yik < ε
4
i) Pi∈C αi > 1 − ε
iii) (cid:13)(cid:13)Pi∈C zi(cid:13)(cid:13) = C.
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
9
If A ⊂ C then the proof will be finished since condition iii) implies that (cid:13)(cid:13)Pi∈A zi(cid:13)(cid:13) = A. In case that
there is some i0 ∈ A \ C, we put B = {i ≤ 4 : i 6= i0} and so
1 −
1
A
=
A − 1
A
=Xi∈B
αi ≥Xi∈C
αi >
3
4
.
Then A > 4, which is a contradiction. Hence A ⊂ C and we proved that Y has the AHSp-ℓ4
∞.
Of course, if γ satisfies Definition 2.3, then ρ = γ also satisfies condition 2). In case that 2) is satisfied
with the function ρ, we showed that condition 3) is also satisfied with the function ν = ρ2. Lastly if we
assume that 3) is true with a function ν we know that Y has the AHSp-ℓ4
∞ with the function ε 7→ ν(cid:0) ε
4(cid:1)
(cid:3)
because of the proof of 3) ⇒ 1).
3. A characterization of the spaces Y such that the pair (ℓ4
∞, Y ) has the
Bishop-Phelps-Bollob´as property
In this section we show that the pair (ℓ4
and only if Y has the AHSp-ℓ4
∞, Y ) has the Bishop-Phelps-Bollob´as property for operators if
∞. Some technical results will make the proof easier. As usual, we denote by
co(A) the convex hull of a subset A of a linear space.
Lemma 3.1. If x ∈ E1 and x(i) ≤ x(i + 1) for 2 ≤ i ≤ 3, then x ∈ co (B).
It is satisfied that
E1 = ∪6
k=1co({vi : i ∈ Ak}), where
A1 = {1, 2, 3, 4}, A2 = {1, 3, 4, 5}, A3 = {1, 4, 6, 7},
A4 = {1, 2, 4, 8}, A5 = {1, 4, 5, 6} and A6 = {1, 4, 7, 8}.
Indeed for every 1 ≤ k ≤ 6, {vi : i ∈ Ak} is the image under an appropriate linear isometry on ℓ4
∞ of B.
Proof. If x ∈ E1, by Lemma 2.9 we know that
(3.1)
x =
1 + x(2)
2
v1 +
x(i + 1) − x(i)
2
vi +
1 − x(4)
2
v4.
3
Xi=2
In case that x(2) ≤ x(3) ≤ x(4) notice that x is expressed in (3.1) as a convex combination of {vi : 1 ≤ i ≤ 4}.
For each permutation σ of {2, 3, 4} we define the linear isometry Tσ on ℓ4
∞ given by
Notice that Tσ preserves E1.
Tσ(x) =(cid:0)x(1), x(σ(2)), x(σ(3)), x(σ(4))(cid:1)
(x ∈ R4).
If x ∈ E1 and σ is a permutation of {2, 3, 4} is such that x(σ(2)) ≤ x(σ(3)) ≤ x(σ(4)) we know that the
element Tσ(x) can be expressed as a convex combination of {vi : 1 ≤ i ≤ 4}. Hence x = Tσ−1(Tσ(x)) can be
expressed as a convex combination of {Tσ−1(vi) : 1 ≤ i ≤ 4}. So it suffices to compute the images by Tσ of
{vi : 1 ≤ i ≤ 4}, where σ is any permutation of {2, 3, 4}. Notice that the elements v1 and v4 are invariant
by all these isometries. So it suffices to evaluate the image of the elements {v2, v3}.
10
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
We include the results in the following table, where we denote by I the identity and τi,j the transposition
of the elements i and j on {2, 3, 4}
σ Tσ({v2, v3})
I {v2, v3}
τ2,3 {v3, v5}
τ2,4 {v6, v7}
τ3,4 {v2, v8}
τ2,3 ◦ τ3,4 {v5, v6}
τ2,3 ◦ τ2,4 {v7, v8}
Then we obtained that E1 = ∪6
i=1co{vj : j ∈ Ai}, where the sets Ai are given by
A1 = {1, 2, 3, 4}, A2 = {1, 3, 4, 5}, A3 = {1, 4, 6, 7},
A4 = {1, 2, 4, 8},
A5 = {1, 4, 5, 6} and A6 = {1, 4, 7, 8}.
The next result gives a procedure to change an element u0 close to co{vi : 1 ≤ i ≤ 4} by a new one
satisfying more requirements.
Lemma 3.2. Assume that 0 < ε < 1
2 , x0 ∈ co{vi : 1 ≤ i ≤ 4} and u0 ∈ E1 satisfies that ku0 − x0k < ε.
Then there is v0 ∈ E1 such that kv0 − u0k < 3ε and such that there is a set A ⊂ {i ∈ N : i ≤ 8} satisfying
that u0 and v0 can be written as convex combinations as follows
(cid:3)
and also
u0 =Xi∈A
βivi, v0 = Xi∈A,i≤4
γivi
γi > 0 for some i ⇒ βi > 0.
Proof. By assumption we know that x0 can be written as x0 =P4
i=1 αi = 1.
By Lemma 3.1 u0 ∈ ∪6
P4
v0 = u0 satisfies the statement. Otherwise we can express u0 =Pi∈A βivi, where βi ≥ 0 for each i ∈ A and
Pi∈A βi = 1 and it suffices to prove the claim in the following cases:
i=1co({vi : i ∈ Ak}). If u0 ∈ co{vi : i ∈ A1} = {vi : i ≤ 4} then the element
Case a) A = A2 ∪ A5 = {1, 3, 4, 5, 6}. Since the functional v∗
ball of (ℓ4
∞)∗, in view of Lemma 2.9 and Remark 2.10 we obtain that
2 given by v∗
i=1 αivi, where αi ≥ 0 for i ≤ 4 and
2(x) = x(3)−x(2)
2
belongs to the unit
(3.2)
α2 + β5 + β6 = v∗
2(x0 − u0) ≤ kx0 − u0k < ε.
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
11
Case b) Assume that A = A3 = {1, 4, 6, 7}. The functional v∗ given by v∗(x) = x(4)−x(2)
ball of (ℓ4
∞)∗ and satisfies
2
belongs to the unit
v∗(vi) = 0,
i = 1, 4
v∗(vi) = 1,
i = 2, 3,
and
v∗(vi) = −1,
i = 6, 7.
As a consequence we have that
(3.3)
α2 + α3 + β6 + β7 = v∗(x0 − u0) ≤ kx0 − u0k < ε.
Case c) Assume that A = A4 ∪ A6 = {1, 2, 4, 7, 8}. By Lemma 2.9 and Remark 2.10 we obtain that
(3.4)
α3 + β7 + β8 = v∗
3(x0 − u0) ≤ kx0 − u0k < ε.
In each of the above cases, we take B = A ∩ {1, 2, 3, 4}. Notice that in view of (3.2), (3.3) and (3.4), it
is satisfied that
(3.5)
Xi∈A\B
We will check now that the element v0 =
B ⊂ {1, 2, 3, 4}, clearly v0 ∈ co{v1, v2, v3, v4} ⊂ E1.
Since B ⊂ A the element v0 also satisfies
1
2
.
βi > 1 − ε >
βi < ε ⇒ Xi∈B
Pi∈B βi Pi∈B βivi satisfies the requirements of the claim. Since
1
+(cid:13)(cid:13)(cid:13) Xi∈A\B
βivi(cid:13)(cid:13)(cid:13)
kv0 − u0k ≤ (cid:13)(cid:13)(cid:13)(cid:13)
≤
1
1
1
=
βi
βivi −Xi∈B
− 1 + Xi∈A\B
βivi(cid:13)(cid:13)(cid:13)(cid:13)
Pi∈B βi Xi∈B
Pi∈B βi
Pi∈B βi Xi∈A\B
βi + Xi∈A\B
Pi∈B βi(cid:19)
βi(cid:18)1 +
= Xi∈A\B
< ε(cid:18)1 +
1 − ε(cid:19) < 3ε
βi
1
1
(by (3.5)).
Let us notice that B ⊂ {1, 2, 3, 4} ∩ A. If we take γi = βi
consequence, in case that γi > 0 for some i ∈ B we obtain that βi > 0. So the element v0 satisfies all the
for every i ∈ B then v0 = Pi∈B γivi. As a
Pi∈B βi
required conditions.
(cid:3)
Next result is the version of [5, Theorem 2.9] for ℓ4
∞, where the analogous result was obtained for ℓ3
In our case, the fact that the domain has dimension 4, and so the norm of an element T ∈ L(ℓ4
∞.
∞, Y ) is
makes the proof more
the maximum of the norm of the evaluation of T at eight extreme points of Bℓ4
∞
complicated comparing to the case that the domain has dimension 3.
12
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
Theorem 3.3. For every Banach space Y, the pair (ℓ4
∞, Y ) has the Bishop-Phelps-Bollob´as property if
and only if Y has the approximate hyperplane sum property for ℓ4
∞.
Moreover,
if (ℓ4
∞, Y ) satisfies Definition 1.1 with the function η, then Y has the AHSp-ℓ4
∞ for the function γ (see Definition 2.3), the pair (ℓ4
∞ with
∞, Y )
48(cid:1). In case that Y has the AHSp-ℓ4
γ(ε) = η(cid:0) ε
satisfies BPBp with the function η(ε) = γ2(cid:0) ε
3(cid:1).
satisfies condition 3) in Proposition 2.12 with the function ν(ε) = η(cid:0) ε
Proof. Assume that the pair (ℓ4
∞ with the function γ(ε) = η(cid:0) ε
48(cid:1).
Let us fix 0 < ε < 1. Assume that (yi)i≤4 ∈ M 4
AHSp-ℓ4
∞, Y ) satisfies the BPBp with the function η. We will prove that Y
12(cid:1). As a consequence Y has the
i=1 αiyi is a convex combination satisfying
that
x0 =P4
Y and that P4
αiyi(cid:13)(cid:13)(cid:13)
> 1 − ν(ε).
4
Xi=1
(cid:13)(cid:13)(cid:13)
Let T be the element in BL(ℓ4
∞,Y ) that (yi)i≤4 represents in view of Proposition 2.11. The element
i=1 αivi satisfies that x0 ∈ Sℓ4
∞
and by assumption we know that
4
(cid:13)(cid:13)T (x0)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)
Xi=1
αiyi(cid:13)(cid:13)(cid:13)
> 1 − ν(ε) = 1 − η(cid:16) ε
12(cid:17) > 1 −
ε
12
> 0.
and S ∈ SL(ℓ4
∞,Y ) satisfying the following conditions
∞
Since the pair (ℓ4
(3.6)
∞, Y ) has the BPBp, there are u0 ∈ Sℓ4
ε
12
ku0 − x0k <
kSu0k = 1,
and
S −
(cid:13)(cid:13)(cid:13)
<
ε
12
.
T
kT k(cid:13)(cid:13)(cid:13)
Notice that u0(1) − 1 = u0(1) − x0(1) ≤ ku0 − x0k < 1 and ku0k = 1, so 0 < u0(1) ≤ 1. Since we
clearly have that the element u0 can be written as a convex combination as follows
u0 =
1 + u0(1)
2
(1, u0(2), u0(3), u0(4)) +
1 − u0(1)
2
(−1, u0(2), u0(3), u0(4)),
1 + u0(1) > 0 and S attains its norm at u0, then S also attains its norm at (1, u0(2), u0(3), u0(4)). The
previous element belongs to Sℓ4
∞
and satisfies that
k(1, u0(2), u0(3), u0(4)) − x0k ≤ ku0 − x0k <
ε
12
.
As a consequence, by changing u0 by (1, u0(2), u0(3), u0(4)), if needed, we can assume that u0 ∈ E1.
By Lemma 3.2 there is v0 ∈ E1 such that kv0 − u0k < ε
4 and such that there is a set
A ⊂ {i ∈ N : i ≤ 8} satisfying that u0 and v0 can be written as convex combinations as follows
and also
(3.7)
u0 =Xi∈A
βivi, v0 = Xi∈A,i≤4
γivi
γi > 0 for some i ⇒ βi > 0.
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
13
By (3.6) S attains its norm at u0, hence by Hahn-Banach Theorem there is an element y∗ ∈ SY ∗ such that
y∗(S(u0)) = 1. Since
βi ≥ 0 for each i ∈ A and Pi∈A βi = 1, we have that
As a consequence, by (3.7) we obtain that
(3.8)
The element v0 satisfies that
(3.9)
1 = y∗(S(u0)) =Xi∈A
βiy∗S(vi),
βi ∈ A, βi > 0 ⇒ y∗(cid:0)S(vi)(cid:1) = 1.
i ∈ A, i ≤ 4, γi 6= 0 ⇒ y∗(cid:0)S(vi)(cid:1) = 1.
kv0 − x0k ≤ kv0 − u0k + ku0 − x0k
(by (3.6))
+
ε
12
.
<
=
ε
4
ε
3
By Lemma 2.9 the functionals {v∗
i : 1 ≤ i ≤ 4} belong to B(ℓ4
∞)∗ and are the biorthogonal functionals of the
basis B. In view of (3.9) we get that
(3.10)
αi − γi = v∗
i (x0 − v0) ≤ kx0 − v0k <
ε
3
,
∀i ≤ 4.
We write C := {i ≤ 4 : i ∈ A, γi 6= 0}. If C = {1, 2, 3, 4} then Pi∈C αi = 1. Otherwise, from (3.10) we
obtain that
Xi∈C
αi = 1 − Xi≤4,i6∈C
αi > 1 −
(4 − C) ≥ 1 − ε.
ε
3
Finally we check that (zi)i≤4 = (S(vi))i≤4 is the desired element in M 4
S ∈ SL(ℓ4
∞,Y ) (see Proposition 2.11). By (3.8) we have that
Y . Clearly (zi)i≤4 ∈ M 4
Y since
It remains to show only that kzi − yik < ε for each i ≤ 4. Indeed for each 1 ≤ i ≤ 4 we have that
(cid:13)(cid:13)(cid:13)Xi∈C
zi(cid:13)(cid:13)(cid:13)
≥ y∗(cid:18)Xi∈C
S(vi)(cid:19) = C.
kzi − yik = kS(vi) − T (vi)k
T (vi)
S(vi) −
T (vi)
kT k
+(cid:13)(cid:13)(cid:13)
− T (vi)(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:13)
<
ε
6
T
kT k (cid:13)(cid:13)(cid:13)
kT k(cid:13)(cid:13)(cid:13)
< ε
(by (3.6)).
S −
+ 1 − kT k
We proved that Y satisfies condition 3) of Proposition 2.12 with ν(ε) = η(cid:0) ε
12(cid:1).
14
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
Assume that Y satisfies the AHSp-ℓ4
3) in that result with the function ν(ε) = γ2(ε). We will show that that the pair (ℓ4
with the function η(ε) = ν( ε
3 ) = γ2( ε
3 ).
∞ with the function γ. By Proposition 2.12, Y also satisfies condition
∞, Y ) has the BPBp
Let be 0 < ε < 1 and assume that T ∈ SL(ℓ4
∞,Y ) and x0 ∈ Sℓ4
∞
are such that
By using an appropriate isometry,
if needed,
in view of Lemma 3.1, we can assume that
kT x0k > 1 − η(ε) = 1 − ν(cid:16) ε
3(cid:17).
x0 ∈ co{vi
Proposition 2.11 the set (yi)i≤4 = (T (vi))i≤4 ∈ M 4
: 1 ≤ i ≤ 4}. Let us write x0 as a convex combination x0 = P4
Y . So we have that
i=1 αivi.
In view of
By assumption Y satisfies condition 3) in Proposition 2.12, so there is a (nonempty) set A ⊂ {1, 2, 3, 4},
and (zi)i≤4 ∈ M 4
Y such that
4
Xi=1
(cid:13)(cid:13)(cid:13)(cid:13)
αiyi(cid:13)(cid:13)(cid:13)(cid:13)
= kT (x0)k > 1 − ν(cid:16) ε
3(cid:17).
αi > 1 −
ε
3
Xi∈A
> 0,
kzi − yik <
ε
3
for all i ≤ 4
(3.11)
and
(3.12)
= A.
(cid:13)(cid:13)(cid:13)Xi∈A
zi(cid:13)(cid:13)(cid:13)
Let S be the unique linear operator from ℓ4
(zi)i≤4 ∈ M 4
Sℓ4
. By equation (3.12) the operator S attains its norm at u0 since
∞,Y ) by Proposition 2.11. The element u0 given by u0 =Pi∈A
Y , S ∈ BL(ℓ4
∞
∞ to Y such that S(vi) = zi for every 1 ≤ i ≤ 4. Since
vi belongs to
αi
Pi∈A αi
So S ∈ SL(ℓ4
C = {i ∈ N : i ≤ 4, i /∈ A}. Finally we obtain that
∞,Y ).
From Proposition 2.11 and (3.11) we obtain that kS − T k < ε. We write
1 =
= kS(u0)k ≤ kSk ≤ 1.
kPi∈A αizik
Pi∈A αi
4
αi
1
vi(cid:13)(cid:13)(cid:13)
αivi +Xi∈C
αikvik +Xi∈C
kx0 − u0k =(cid:13)(cid:13)(cid:13)
αivi −Xi∈A
Xi=1
Pi∈A αi
=(cid:13)(cid:13)(cid:13)(cid:16)1 −
Pi∈A αi(cid:17)Xi∈A
Pi∈A αi(cid:12)(cid:12)(cid:12)Xi∈A
≤(cid:12)(cid:12)(cid:12)
≤ Pi∈C αi
Pi∈A αi Xi∈A
αi +Xi∈C
= 2Xi∈C
2ε
3
αi <
< ε
1 −
αi
1
(by (3.11)).
αivi(cid:13)(cid:13)(cid:13)
αikvik
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
We proved that the pair (ℓ4
∞, Y ) has the BPBp with η(ε) = ν(cid:0) ε
3(cid:1) = γ2(cid:0) ε
3(cid:1).
4. Examples of spaces with the approximate hyperplane sum property for ℓ4
∞
15
(cid:3)
The goal of this section is to provide classes of Banach spaces with the approximate hyperplane sum
property for ℓ4
∞.
As we already mentioned in the introduction the pair (X, Y ) has BPBp whenever X and Y are finite-
dimensional normed spaces [2, Proposition 2.4]. By applying this result to X = ℓ4
3.3 we obtain that finite-dimensional spaces have the AHSp-ℓ4
∞, and in view of Theorem
∞. We will also provide a simple direct proof
of this fact.
Proposition 4.1. Every finite-dimensional normed space has the AHSp-ℓ4
∞.
Proof. We will argue by contradiction. Let Y be a finite-dimensional space and assume that Y does
not have the AHSp-ℓ4
∞. So there is ε0 > 0 for which Definition 2.3 is not satisfied. Hence there is a sequence
{γn} of positive real numbers satisfying {γn} → 0 and also for each natural number n, there are an element
(yn
Y and a nonempty set An ⊂ {1, 2, 3, 4} satisfying
i )i≤4 ∈ M 4
(4.1)
and also
(4.2)
(zi)i≤4 ∈ M 4
Y , (cid:13)(cid:13)(cid:13)Xi∈An
(cid:13)(cid:13)(cid:13)Xi∈An
zi(cid:13)(cid:13)(cid:13)
yn
i(cid:13)(cid:13)(cid:13)
> An(cid:0)1 − γn(cid:1)
= An ⇒ max{kyn
i − zik : i ≤ 4} ≥ ε0,
Since Y is finite dimensional, M 4
Y is a compact set of Y 4. By taking into account that {i ∈ N : i ≤ 4}
is finite, and passing to a subsequence, we can assume that there is a set A ⊂ {i ∈ N : i ≤ 4} such that
i }n converges to yi, so (yi)i≤4 ∈ M 4
An = A, for every n, and also that for each i ≤ 4 the sequence {yn
Y .
From condition (4.1) it follows that (cid:13)(cid:13)Pi∈A yi(cid:13)(cid:13) = A. In view of condition (4.2) this is a contradiction. (cid:3)
Recall that a Banach space Y is uniformly convex if for every ε > 0 there is 0 < δ < 1 such that
u, v ∈ BY ,
ku + vk
2
> 1 − δ ⇒ ku − vk < ε .
In such a case, the modulus of convexity of Y is the function defined by
δY (ε) := infn1 −
ku + vk
2
: u, v ∈ BY , ku − vk ≥ εo .
As a consequence of [20, Theorem 2.5] the pair (ℓ4
operators in case that Y is uniformly convex. By Theorem 3.3 uniformly convex spaces have AHSp-ℓ4
∞, Y ) has the Bishop-Phelps-Bollob´as property for
∞.
However, we provide a direct proof of that fact.
16
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
Lemma 4.2. Assume that (yi)i≤4 ∈ M 4
Y , y∗ ∈ SY ∗ and δ > 0 satisfies that
i < k ≤ 4, y∗(yi), y∗(yk) > 1 − δ ⇒ y∗(yj) > 1 − 2δ for each i < j < k.
Proof. Since (yi)i≤4 ∈ M 4
Y , then for every i < j < k we have that kyi − yj + ykk ≤ 1. If we assume
that y∗(yi), y∗(yk) > 1 − δ then we have that
2 − 2δ − y∗(yj) < y∗(yi − yj + yk) ≤ kyi − yj + ykk ≤ 1.
As a consequence,
y∗(yj) > 1 − 2δ.
(cid:3)
Proposition 4.3. Every uniformly convex Banach space has the approximate hyperplane sum property
∞. Moreover, if δY is the modulus of convexity of Y , then Y has the AHSp-ℓ4
∞ with the function
for ℓ4
2
, ε
6(cid:9).
γ(ε) = min(cid:8) δY (ε)
0 < ε < 1, we define γ(ε) = min(cid:8) δY (ε)
a set such that
2
Proof. Assume that Y is a uniformly convex Banach space with modulus of convexity δY . Given
, ε
6(cid:9). Assume that y∗ ∈ SY ∗ , (yi)i≤4 ∈ M 4
Y and ∅ 6= A ⊂ {1, 2, 3, 4} is
By Lemma 2.5 we
can assume
that min A = 1.
In view of Lemma 4.2,
the
set
y∗(yi) > 1 − γ(ε) > 0, ∀i ∈ A.
C = {i ∈ N : 1 ≤ i ≤ max A} satisfies that
y∗(yi) > 1 − 2γ(ε) > 0, ∀i ∈ C.
Hence for every i, j ∈ C we have that
1 − δY (ε) ≤ 1 − 2γ(ε) < y∗ yi + yj
2 ! ≤ y∗ yi
2
kyik + yj
By the definition of the modulus of convexity it follows that
(4.3)
yi
kyik
We will show that there exists (zi)i≤4 ∈ M 4
(cid:13)(cid:13)(cid:13)
− yj(cid:13)(cid:13)(cid:13)
< ε,
∀i, j ∈ C.
Y satisfying that
! ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
yi
kyik + yj
2
.
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(4.4)
kzi − yik < ε, ∀i ≤ 4
and
= C.
By Proposition 2.6 it suffices to show (4.4) in case that there is 2 ≤ k ≤ 4 such that C coincides with
the set Ck = {i ∈ N : i ≤ k}.
So let us fix 2 ≤ k ≤ 4 and define zi = y1
ky1k for every i ∈ Ck. From (4.3) it follows that
(4.5)
kzi − yik < ε,
∀i ∈ Ck
(cid:13)(cid:13)(cid:13)Xi∈C
zi(cid:13)(cid:13)(cid:13)
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
17
and it is trivially satisfied that
(4.6)
In case that k = 2, we put a = γ(ε)
1+γ(ε) , b = γ(ε)
2(1+γ(ε)) and
= Ck.
(cid:13)(cid:13)(cid:13) Xi∈Ck
zi(cid:13)(cid:13)(cid:13)
z3 = (1 − a)y3 + bz1,
z4 = (1 − a)y4 − bz1.
For each i ∈ {3, 4}, it follows that
kzik ≤ 1 − a + b = 1 −
γ(ε)
1 + γ(ε)
+
γ(ε)
2(1 + γ(ε))
≤ 1
and
kzi − yik ≤ a + b =
γ(ε)
1 + γ(ε)
+
γ(ε)
2(1 + γ(ε))
=
3γ(ε)
2(1 + γ(ε))
<
3γ(ε)
2
< ε.
By the last chain of inequalities and (4.5) it is satisfied that
kzi − yik < ε,
∀i ≤ 4.
Since for each i ∈ {3, 4} we have
(4.7)
kz1 − z2 + zik = kzik ≤ 1
and for each j ∈ {1, 2} it is clear that
kzj − z3 + z4k = kz1 − z3 + z4k
= k(1 − 2b)z1 + (1 − a)(−y3 + y4)k
= (1 − a)kz1 − y3 + y4k
(4.8)
≤ (1 − a)(ky1 − y3 + y4k + kz1 − y1k)
≤ (1 − a)(cid:16)1 +(cid:13)(cid:13)(cid:13)
y1
ky1k
− y1(cid:13)(cid:13)(cid:13)(cid:17)
= (1 − a)(2 − ky1k) < (1 − a)(1 + γ(ε)) = 1.
In view of (4.7) and (4.8) we obtain that (zi)i≤4 ∈ M 4
Y .
If k = 3, we take z4 = y4, in this case it is immediate to check that (zi)i≤4 ∈ M 4
Y , also by the definition
of z4 and (4.5) it follows that
kzi − yik < ε,
∀i ≤ 4.
Finally, if k = 4, we have by definition that zi = y1
(zi)i≤4 ∈ M 4
Y and in view of (4.5) we also have that kzi − yik < ε, for each i ∈ {1, 2, 3, 4}.
ky1k for every i ∈ {1, 2, 3, 4}, so it is trivially satisfied that
We showed that for every 2 ≤ k ≤ 4 there exists (zi)i≤4 ∈ M 4
is complete since A ⊂ C.
Y satisfying (4.4) for C = Ck, so the proof
(cid:3)
18
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
As a consequence of [6, Theorem 2.4], [9, Theorem 2.1] and Theorem 3.3, there is also a nontrivial class
of spaces containing uniformly convex spaces and satisfying AHSp-ℓ4
∞.
The next statement can be shown by using the same argument of the analogous result for AHSp-ℓ3
∞ (see
[5, Proposition 2.4]). For this reason we do not include the proof.
Proposition 4.4. Let L be a nonempty locally compact Hausdorff topological space and Y a Banach
space. Then C0(L, Y ) has the AHSp-ℓ4
∞ if, and only if, Y has the AHSp-ℓ4
∞.
It is clear that R has the AHSp-ℓ4
∞, so by the previous result C0(L) has the same property for any locally
compact Hausdorff space L.
Our aim now is to prove that the space ℓ1 has the AHSp-ℓ4
∞. In what follows we will denote by u∗ the
element in ℓ∗
1 given by
∞
The next simple result will be useful for this purpose.
Xn=1
u∗(x) =
x(n)
(x ∈ ℓ1).
Lemma 4.5. Let be s, t ∈ R+, x, y, z ∈ ℓ1. Assume that
1 − s ≤ u∗(x − y + z)
and
kx − y + zk ≤ 1 + t.
Then there is w ∈ ℓ1 such that
w ≥ z,
kw − zk ≤ s + t
and
x − y + w ≥ 0.
Proof. Define the sets given by
P = {k ∈ N : y(k) ≤ (x + z)(k)}
and
N = N\P.
Since u∗ ≥ 0, we have that
1 − s ≤ u∗(x − y + z) ≤ u∗(cid:0)(x − y + z)χP(cid:1)
≤ k(x − y + z)χP k ≤ kx − y + zk
≤ 1 + t.
Hence k(x − y + z)χP k ≥ 1 − s, so
(4.9)
k(x − y + z)χN k = kx − y + zk − k(x − y + z)χP k ≤ 1 + t − (1 − s) = s + t.
Let w ∈ ℓ1 be the element given by
w(k) =(z(k)
(y − x)(k)
if k ∈ P
if k ∈ N.
It is clear that wχP = zχP and (y − x)χN ≥ zχN , so w ≥ z. Since
(x − y + w)χP = (x − y + z)χP ≥ 0
and
(x − y + w)χN = 0,
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
it is satisfied that x − y + w ≥ 0.
In view of (4.9) we also have that
kw − zk = k(w − z)χN k = k(y − x − z)χN k
= k(x − y + z)χN k ≤ s + t.
19
(cid:3)
Notice that in Lemma 4.5 the element x satisfies the same assumptions that z. So under the same
conditions we also obtain an element v ≥ x such that kv − xk ≤ s + t and v − y + z ≥ 0.
Proposition 4.6. There is a function ρ :]0, 1[−→R+ such that the space ℓn
1 satisfies condition 2) of
Proposition 2.12 for such function, for each natural number n. Also the space ℓ1 satisfies the previous
statement. As a consequence, there is a function γ :]0, 1[−→R+ such that the spaces ℓn
approximate hyperplane sum property for ℓ4
∞ with such function (see Definition 2.3).
1 and ℓ1 have the
Proof. We prove the statement for ℓ1. To this purpose we denote by {en} the usual Schauder basis of
226 and E = Ext(B(ℓ1)∗),
∞ with the
ℓ1. It suffices to show that ℓ1 satisfies condition 2) in Proposition 2.12 for ρ(ε) = ε
which is clearly a 1-norming set for ℓ1.
function γ given by γ(ε) = ρ2( ε
In such case we deduce that ℓ1 satisfies the AHSp-ℓ4
4 ) by Proposition 2.12.
From now on 0 < ε < 1 will be fixed and we simply write ρ instead of ρ(ε). Assume that
y∗ ∈ E = Ext(B(ℓ1)∗ ), (ai)i≤4 ∈ M 4
ℓ1 and ∅ 6= C ⊂ {1, 2, 3, 4} is a nonempty set such that
y∗(ai) > 1 − ρ, ∀i ∈ C.
Firstly by Lemma 2.5 we can clearly assume that min C = 1.
In the case that C contains only one element, by Proposition 2.6 condition 2) in Proposition 2.12 is
satisfied for such set C. Otherwise C contains at least two elements.
Since Ext(B(ℓ1)∗) = {z∗ ∈ ℓ∗
onto ℓ1, we can clearly assume that y∗ = u∗. It suffices to show that there exists (zi)i≤4 ∈ M 4
that
1 : z∗(en) = 1, ∀n ∈ N}, by using an appropriate surjective linear isometry
ℓ1 satisfying
kzi − aik < ε, ∀i ≤ 4
and also
u∗(zi) = 1
and
zi ≥ 0, ∀i ∈ C.
We define the set C ′ given by
C ′ := {i ∈ N : 1 ≤ i ≤ max C, u∗(ai) > 1 − 2ρ}.
In case that i, k ∈ C and j ∈ N satisfies i < j < k, by Lemma 4.2 we know that j ∈ C ′. Since C ⊂ C ′, the
previous remark shows that C ′ has consecutive elements. In fact
C ′ = {i ∈ N : 1 = min C ≤ i ≤ max C}.
20
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
We define
Pi = {k ∈ N : ai(k) ≥ 0},
Ni = N\Pi,
bi = aiχPi,
(i ∈ C ′)
Since u∗ ≥ 0, by assumption we have that
(4.10)
Hence
(4.11)
1 − 2ρ < u∗(ai) ≤ u∗(bi) ≤ kbik ≤ kaik ≤ 1 and bi ≥ 0,
∀i ∈ C ′.
kbi − aik = kaiχNik ≤ 1 − kbik < 2ρ,
∀i ∈ C ′.
Now we will distinguish several cases, depending on the number of elements of C ′. Since we assume that C
contains more than one element and C ⊂ C ′, then C ′ also contains at least two elements. So C ′ = {1, 2},
C ′ = {1, 2, 3} or C ′ = {1, 2, 3, 4}.
• Case 1: Assume that C ′ = {1, 2}. Define the elements in ℓ1 given by
So we have that
xi =(bi + (1 − kbik)e1
ai
if i ∈ {1, 2}
if i ∈ {3, 4}.
xi ≥ 0,
and
kxik = 1,
i = 1, 2.
From (4.10) and (4.11) it follows that
(4.12)
kxik ≤ 1
and
kxi − aik < 4ρ,
∀i ≤ 4.
In this case we will change twice the previous vectors in order to get the desired properties. Firstly we define
the set {yi : 1 ≤ i ≤ 4} by
It is clearly satisfied that
yi =( xi
1+8ρ
1+8ρ + (1 − 1
xi
1+8ρ )e1
if i ∈ {1, 2}
if i ∈ {3, 4}.
(4.13)
yi ≥ 0, kyik = 1, for i = 1, 2
and
yi ∈ Bℓ1
for
1 ≤ i ≤ 4.
For each 1 ≤ i ≤ 4 we also have that
(4.14)
yi −
kyi − aik ≤(cid:13)(cid:13)(cid:13)
<(cid:18)1 −
1 + 8ρ
xi
xi
1 + 8ρ(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)
1 + 8ρ(cid:19) +(cid:18)1 −
1
− xi(cid:13)(cid:13)(cid:13)
1 + 8ρ(cid:19) + 4ρ
1
=
16ρ
1 + 8ρ
< 20ρ.
+ 4ρ
+ kxi − aik
(by (4.12))
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
21
Now we fix 1 ≤ i1 < i2 < i3 ≤ 4 and estimate the norm of kyi1 − yi2 + yi3k as follows.
In case that
{1, 2} ⊂ {i1, i2, i3} we have
(4.15)
kyi1 − yi2 + yi3k =(cid:13)(cid:13)(cid:13)
≤
xi1 − xi2 + xi3
1 + 8ρ
(cid:13)(cid:13)(cid:13)
1 + 8ρ
kai1 − ai2 + ai3k + kx1 − a1k + kx2 − a2k
<
kai1 − ai2 + ai3k + 8ρ
1 + 8ρ
(by (4.12))
≤ 1.
Otherwise {1, 2} 6⊂ {i1, i2, i3} and we obtain that
(4.16)
Now we define
kyi1 − yi2 + yi3k ≤(cid:13)(cid:13)(cid:13)
<
xi1 − xi2 + xi3
1 + 8ρ
+ 1 −
kai1 − ai2 + ai3k + 4ρ
(cid:13)(cid:13)(cid:13)
1 + 8ρ
1
1 + 8ρ
8ρ
1 + 8ρ
(by (4.12))
+
< 1 + 4ρ.
1+4ρ + (1 − 1
1+4ρ )e1
zi =( yi
y4
1+4ρ
if i ∈ {1, 2, 3}
if i = 4.
In view of (4.13) we have that
(4.17)
zi ≥ 0, u∗(zi) = 1 for
i = 1, 2
and
zi ∈ Bℓ1
for 1 ≤ i ≤ 4.
For each 1 ≤ i ≤ 4 we obtain that
(4.18)
zi −
kzi − aik ≤(cid:13)(cid:13)(cid:13)
<(cid:18)1 −
1 + 4ρ
yi
yi
+(cid:13)(cid:13)(cid:13)
1 + 4ρ(cid:13)(cid:13)(cid:13)
1 + 4ρ(cid:19) +(cid:18)1 −
1
− yi(cid:13)(cid:13)(cid:13)
1 + 4ρ(cid:19) + 20ρ
1
+ kyi − aik
=
8ρ
1 + 4ρ
+ 20ρ < 28ρ < ε.
(by (4.13) and (4.14))
Now we check that (zi)i≤4 ∈ M 4
{1, 2} ⊂ {i1, i2, i3} it is clear that
ℓ1. For each 1 ≤ i1 < i2 < i3 ≤ 4, we consider the following two cases. If
kzi1 − zi2 + zi3k ≤
kyi1 − yi2 + yi3k
1 + 4ρ
+ 1 −
1
1 + 4ρ
<
1
1 + 4ρ
= 1.
+ 1 −
1
1 + 4ρ
(by (4.15))
Otherwise {1, 2} 6⊂ {i1, i2, i3} and we obtain that
kzi1 − zi2 + zi3k =
kyi1 − yi2 + yi3k
1 + 4ρ
(by (4.16)).
< 1
22
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
In view of (4.17) and (4.18), since we checked that (zi)i≤4 ∈ M 4
• Case 2: Assume that C ′ = {1, 2, 3}. We know that
ℓ1, the proof is finished in case 1.
1 − 4ρ = 2(1 − 2ρ) − 1 < u∗(b1 + b3) − u∗(b2)
(by (4.10))
= u∗(b1 − b2 + b3) ≤ kb1 − b2 + b3k
3
≤ ka1 − a2 + a3k +
kbi − aik
Xi=1
(by (4.11)).
< 1 + 6ρ
We obtained that
(4.19)
1 − 4ρ < u∗(b1 − b2 + b3) ≤ kb1 − b2 + b3k < 1 + 6ρ.
We can apply Lemma 4.5 with b1 playing the role of z, so there is x1 ∈ ℓ1 such that
(4.20)
We define
x1 ≥ b1 ≥ 0,
kx1 − b1k ≤ 10ρ and x1 − b2 + b3 ≥ 0.
xi = bi
for i ∈ {2, 3}
and
x4 = a4.
Notice that in view of (4.11) and (4.20) we have
(4.21)
kxi − aik < 12ρ,
∀i ≤ 4.
As a consequence, for each i ∈ {1, 2, 3}, from (4.10) and (4.20) it follows that
(4.22)
1 − 2ρ < kbik
≤ kxik
If 1 ≤ i1 < i2 < i3 ≤ 4 we obtain that
≤ kxi − aik + kaik < 1 + 12ρ.
3
(4.23)
kxi1 − xi2 + xi3k ≤ kai1 − ai2 + ai3k +
kxij − aij k
Xj=1
< 1 + 36ρ
(by (4.21)).
Now we define a new element (yi)i≤4 in order to have that {yi : i ≤ 3} ⊂ {x ∈ Bℓ1 : u∗(x) = 1}. We put
1+36ρ + (1 − kxik
1+36ρ )e1
yi =( xi
x4
1+36ρ
if
if
i ∈ {1, 2, 3}
i = 4.
For each i ≤ 3, since bi ≥ 0, from (4.20) we also have that xi ≥ 0. In view of (4.22) we deduce that
(4.24)
yi ≥ 0,
u∗(yi) = 1 for
i ∈ {1, 2, 3} and ky4k ≤ 1.
For each 1 ≤ i ≤ 4, in view of (4.22) and (4.21) we obtain that
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
23
(4.25)
xi
yi −
kyi − aik ≤(cid:13)(cid:13)(cid:13)
<(cid:16)1 −
1 + 36ρ
1 + 36ρ(cid:13)(cid:13)(cid:13)
1 + 36ρ(cid:17) + kxik(cid:16)1 −
+(cid:13)(cid:13)(cid:13)
1 − 2ρ
+ kxi − aik
− xi(cid:13)(cid:13)(cid:13)
1 + 36ρ(cid:17) + 12ρ
1
xi
<
38ρ
1 + 36ρ
+ (1 + 12ρ)
+ 12ρ
36ρ
1 + 36ρ
< 86ρ.
On one hand, since xi ≥ 0 for each i ≤ 3 we also have that
(4.26)
ky1 − y2 + y3k =(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)
= 1.
x1 − x2 + x3
1 + 36ρ
x1 − x2 + x3
1 + 36ρ
x1 − x2 + x3
1 + 36ρ
+(cid:16)1 −
+(cid:16)1 −
(cid:13)(cid:13)(cid:13)
+(cid:16)1 −
On the other hand, if 1 ≤ i1 < i2 < 4 then
kx1k − kx2k + kx3k
(by (4.20))
(by (4.23))
1 + 36ρ
kx1 − x2 + x3k
1 + 36ρ
kx1 − x2 + x3k
1 + 36ρ
(cid:17)e1(cid:13)(cid:13)(cid:13)
(cid:17)e1(cid:13)(cid:13)(cid:13)
(cid:17)
(4.27)
xi1 − xi2 + x4
kyi1 − yi2 + y4k ≤(cid:13)(cid:13)(cid:13)
< 1 +
1 + 36ρ
14ρ
1 + 36ρ
1
+
1 + 36ρ(cid:12)(cid:12)(cid:12)
kxi1k − kxi2k(cid:12)(cid:12)(cid:12)
(by (4.22) and (4.23))
(cid:13)(cid:13)(cid:13)
< 1 + 14ρ.
Now we define the element in M 4
ℓ1 satisfying the required conditions in this case. To this purpose we
take
1+14ρ + (1 − 1
1+14ρ )e1
zi =( yi
y4
1+14ρ
if i ∈ {1, 2, 3}
if i = 4.
Let us fix 1 ≤ i ≤ 4. In view of (4.24) it is satisfied that
(4.28)
zi ≥ 0,
u∗(zi) = 1 for i ∈ {1, 2, 3} and kz4k ≤ 1.
By using (4.24) and (4.25) we obtain the following upper estimate
(4.29)
yi
yi
zi −
kzi − aik ≤(cid:13)(cid:13)(cid:13)
<(cid:16)1 −
1 + 14ρ
1 + 14ρ(cid:13)(cid:13)(cid:13)
1 + 14ρ(cid:17) + kyik(cid:16)1 −
+(cid:13)(cid:13)(cid:13)
1
< 28ρ + 86ρ = 114ρ < ε.
+ kyi − aik
− yi(cid:13)(cid:13)(cid:13)
1 + 14ρ(cid:17) + 86ρ
1
24
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
We also have that
(4.30)
kz1 − z2 + z3k =(cid:13)(cid:13)(cid:13)
≤
y1 − y2 + y3
1 + 14ρ
ky1 − y2 + y3k
1 + 14ρ
1
+(cid:16)1 −
+(cid:16)1 −
1 + 14ρ(cid:17)e1(cid:13)(cid:13)(cid:13)
1 + 14ρ(cid:17)
1
In case that 1 ≤ i1 < i2 < 4 we obtain that
= 1
(by (4.26)).
(4.31)
kzi1 − zi2 + z4k =(cid:13)(cid:13)(cid:13)
< 1
yi1 − yi2 + y4
1 + 14ρ
(by (4.27)).
(cid:13)(cid:13)(cid:13)
From equations (4.28), (4.29), (4.30) and (4.31) the element (zi)i≤4 ∈ M 4
ℓ1 satisfies all the required
conditions.
• Case 3: Assume that C ′ = {1, 2, 3, 4}. For each 1 ≤ i1 < i2 < i3 ≤ 4, by the same argument used to obtain
(4.19) in case 2, with (i1, i2, i3) playing the role of (1, 2, 3) there we get that
(4.32)
1 − 4ρ < u∗(bi1 − bi2 + bi3) ≤ kbi1 − bi2 + bi3k < 1 + 6ρ.
In view of Lemma 4.5, there are x1, x4 ∈ ℓ1 such that
(4.33)
xi ≥ bi,
kxi − bik ≤ 10ρ for i = 1, 4,
x1 − b2 + b3 ≥ 0
and
b1 − b3 + x4 ≥ 0.
As a consequence
1 − 4ρ < u∗(b1 − b2 + b4) ≤ u∗(x1 − b2 + b4)
(by (4.32) and (4.33))
≤ kx1 − b2 + b4k
≤ kb1 − b2 + b4k + kx1 − b1k
< ka1 − a2 + a4k + 6ρ + kx1 − b1k
(by (4.11))
≤ 1 + 16ρ
(by (4.33))
and
1 − 4ρ < u∗(b2 − b3 + b4) ≤ u∗(b2 − b3 + x4)
(by (4.32) and (4.33))
≤ kb2 − b3 + x4k
≤ kb2 − b3 + b4k + kx4 − b4k < 1 + 16ρ
(by (4.11) and (4.33)).
From the last two chains of inequalities we have that
1 − 4ρ < u∗(x1 − b2 + b4) ≤ kx1 − b2 + b4k < 1 + 16ρ
1 − 4ρ < u∗(b2 − b3 + x4) ≤ kb2 − b3 + x4k < 1 + 16ρ.
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
25
By applying again Lemma 4.5, there are y1, y4 ∈ ℓ1 satisfying that
(4.34)
yi ≥ xi,
kyi − xik ≤ 20ρ for i = 1, 4,
y1 − b2 + b4 ≥ 0
and
b2 − b3 + y4 ≥ 0.
Now we define y2 = b2 and y3 = b3.
By using (4.34), (4.33) and (4.11), for each i ≤ 4 it is clear that
(4.35)
kyi − aik ≤ kyi − bik + kbi − aik
< 20ρ + 10ρ + 2ρ = 32ρ.
As a consequence, from (4.33) and (4.34) for each i ≤ 4 we also have that
(4.36)
1 − 2ρ < u∗(ai) ≤ u∗(bi) ≤ u∗(yi) ≤ kyik ≤ kyi − aik + kaik < 1 + 32ρ.
In view of (4.33) and (4.34) it is immediate to check that
(4.37)
yi1 − yi3 + yi3 ≥ 0
for any 1 ≤ i1 < i2 < i3 ≤ 4.
For every 1 ≤ i1 < i2 < i3 ≤ 4 we also get that
3
(4.38)
kyi1 − yi2 + yi3k ≤ kai1 − ai2 + ai3k +
kyij − aij k
Xj=1
< 1 + 96ρ
(by (4.35)).
Now we define the elements satisfying the required conditions by
By (4.34) and (4.33) it is clear that yi ≥ 0 for each i ≤ 4. From (4.36) we know that kyik ≤ 1 + 32ρ and so
zi =
yi
1 + 96ρ
+(cid:16)1 −
kyik
1 + 96ρ(cid:17)e1
(1 ≤ i ≤ 4).
we deduce that
(4.39)
We also obtain that
zi ≥ 0
and
u∗(zi) = 1,
∀1 ≤ i ≤ 4.
(4.40)
yi
yi
zi −
kzi − aik ≤(cid:13)(cid:13)(cid:13)
<(cid:16)1 −
<(cid:16)1 −
1 + 96ρ
kyik
1 + 96ρ(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)
1 + 96ρ(cid:17) + kyik(cid:16)1 −
1 + 96ρ(cid:17) +
1 + 32ρ
1 + 96ρ
1 − 2ρ
+ 96ρ + 32ρ
<
98ρ
1 + 96ρ
< 226ρ = ε.
+ kyi − aik
− yi(cid:13)(cid:13)(cid:13)
1 + 96ρ(cid:17) + 32ρ
1
96ρ + 32ρ
(by (4.35))
(by (4.36))
26
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
Finally, notice that for every 1 ≤ i1 < i2 < i3 ≤ 4, since yi ≥ 0 for every i ≤ 4 we have that
(4.41)
kzi1 − zi2 + zi3k =(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)
≤
yi1 − yi2 + yi3
1 + 96ρ
yi1 − yi2 + yi3
1 + 96ρ
kyi1 − yi2 + yi3k
1 + 96ρ
+(cid:16)1 −
+(cid:16)1 −
+ 1 −
1 + 96ρ
kyi1 − yi2 + yi3k
1 + 96ρ
kyi1 − yi2 + yi3k
1 + 96ρ
(cid:17)e1(cid:13)(cid:13)(cid:13)
(cid:17)e1(cid:13)(cid:13)(cid:13)
kyi1k − kyi2k + kyi3k
(by (4.37))
(by (4.38))
= 1.
In view of equations (4.39), (4.40) and (4.41) the proof is also finished in case 3. So we proved the statement
for ℓ1.
The proof for ℓn
(instead of elements in M 4
1 follows from the same argument that we used for ℓ1 by considering elements in M 4
ℓn
(cid:3)
1 )∗) = {z∗ ∈ (ℓn
1 )∗ : z∗(ek) = 1, ∀k ≤ n}.
ℓ1) and the description Ext(B(ℓn
1
Next result follows from the same argument of [5, Theorem 2.7]. It will be useful, for instance, to extend
the previous result to L1(µ) for any positive measure µ.
Theorem 4.7. Assume that Y is a Banach space and γ :]0, 1[−→R+ is a function such that
Y = ∪{Yα : α ∈ Λ}, where {Yα : α ∈ Λ} is a nested family of subspaces of Y satisfying the AHSp-ℓ4
∞
with the function γ. Then Y has the AHSp-ℓ4
∞ with the function ζ(ε) = γ(cid:0) ε
2(cid:1).
Proof. Given 0 < ε < 1, let γ(ε) be the positive real number satisfying Definition 2.3 for each space
Yα.
Assume that (ai)i≤4 ∈ M 4
Y and that for some nonempty set A ⊂ {1, 2, 3, 4} and y∗ ∈ SY ∗, it is satisfied
that y∗(ai) > 1 − γ(cid:0) ε
2(cid:1) for each i ∈ A. Let us choose a real number t such that
2(cid:17) : i ∈ Aoo.
, minny∗(ai) − 1 + γ(cid:16) ε
minn ε
0 < t <
1
4
2
By assumption there exist α0 ∈ Λ and {bi : i ≤ 4} ⊂ BYα0 satisfying
kbi − aik < t
for all i ≤ 4.
Now we define yi = bi
(yi)i≤4 ∈ MY 4
α0
. We clearly have
1+3t for any i ≤ 4. By using that (ai)i≤4 ∈ M 4
Y it is immediate to check that
For each i ∈ A we obtain that
bi
1 + 3t
kyi − aik ≤(cid:13)(cid:13)(cid:13)
(4.42)
(4.43)
ε
2
,
for all i ≤ 4.
+ kbi − aik < 3t + t = 4t <
− bi(cid:13)(cid:13)(cid:13)
y∗(yi) > y∗(ai) − 4t > 1 − γ(cid:16) ε
2(cid:17) > 0.
THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY
27
We define the element z∗ ∈ Y ∗
α0 by
z∗(z) = y∗(z)
(z ∈ Yα0).
Since kz∗k ≤ ky∗k, we get that z∗ ∈ BY ∗
α0
. In view of (4.43) we know that z∗ 6= 0 and we also have that
z∗
kz∗k
(yi) =
y∗
kz∗k
(yi) ≥ y∗(yi) > 1 − γ(cid:16) ε
2(cid:17)
such that
for all i ∈ A.
By assumption there is (zi)i≤4 ∈ M 4
Yα0
i) kzi − yik < ε
2
for each i ≤ 4,
ii) (cid:13)(cid:13)Pi∈A zi(cid:13)(cid:13) = A.
In view of (4.42) we obtain that
kzi − aik ≤ kzi − yik + kyi − aik <
ε
2
+
ε
2
= ε,
∀i ≤ 4.
We proved that Y satisfies Definition 2.3 with the function ζ(ε) = γ(cid:0) ε
2(cid:1).
linearly isometric to some space ℓn
Let us remark that the subspace of L1(µ) generated by a finite number of characteristic functions is
1 . Since the space of simple functions is dense in L1(µ), in view of
Proposition 4.6, we conclude that L1(µ) satisfies the assumption of the previous result and the function γ
does not depend on µ. In view of Theorem 3.3 we obtain that the pairs (ℓ4
∞, L1(µ)) have uniformly BPBp
(cid:3)
for operators for every measure µ. More concretely we deduce the following assertions.
Corollary 4.8. There is a function γ :]0, 1[−→R+ such that L1(µ) satisfies Definition 2.3 with such
∞, L1(µ)) has the BPBp for operators. Moreover
∞, L1(µ)) satisfies Definition 1.1 for such function, for any positive
function, for any positive measure µ. Hence the pair (ℓ4
there is a function η such that the pair (ℓ4
measure µ.
References
[1] M.D. Acosta, The Bishop-Phelps-Bollob´as property for operators on C(K), Banach J. Math. Anal. 10 (2016), no. 2, 307 -- 319.
[2] M.D. Acosta, R.M. Aron, D. Garc´ıa and M. Maestre, The Bishop-Phelps-Bollob´as theorem for operators, J. Funct. Anal. 254 (2008),
no. 11, 2780 -- 2799.
[3] M.D. Acosta, J. Becerra-Guerrero, Y.S. Choi, M. Ciesielski, S.K. Kim, H.J. Lee, M.L. Louren¸co and M. Mart´ın, The Bishop-Phelps-
Bollob´as property for operators between spaces of continuous functions, Nonlinear Anal. 95 (2014), 323 -- 332.
[4] M.D. Acosta, J. Becerra-Guerrero, D. Garc´ıa, S.K. Kim and M. Maestre, Bishop-Phelps-Bollob´as property for certain spaces of
operators, J. Math. Anal. Appl. 414 (2014), 532 -- 545.
[5] M.D. Acosta, J. Becerra-Guerrero, D. Garc´ıa, S.K. Kim and M. Maestre, The Bishop-Phelps-Bollob´as property: a finite-dimensional
approach, Publ. Res. Inst. Math. Sci. 51 (2015), no. 1, 173 -- 190.
[6] M.D. Acosta, D. Garc´ıa, S.K. Kim and M. Maestre, The Bishop-Phelps-Bollob´as property for operators from c0 into some Banach
spaces, J. Math. Anal. Appl. 445 (2017), no. 2, 1188 -- 1199.
[7] R.M. Aron, B. Cascales and O. Kozhushkina, The Bishop-Phelps-Bollob´as theorem and Asplund operators, Proc. Amer. Math. Soc.
139 (2011), no. 10, 3553 -- 3560.
[8] R.M. Aron, Y.S. Choi, D. Garc´ıa and M. Maestre, The Bishop-Phelps-Bollob´as theorem for L(L1(µ), L∞[0, 1]), Adv. Math. 228
(2011), no. 1, 617 -- 628.
[9] R.M. Aron, Y.S. Choi, S.K. Kim, H.J. Lee and M. Mart´ın, The Bishop-Phelps-Bollob´as version of Lindenstrauss properties A and B,
Trans. Amer. Math. Soc. 367 (2015), no. 9, 6085 -- 6101.
[10] E. Bishop and R.R. Phelps, A proof that every Banach space is subreflexive, Bull. Amer. Math. Soc. 67 (1961), 97 -- 98.
[11] B. Bollob´as, An extension to the theorem of Bishop and Phelps, Bull. Lond. Math. Soc. 2 (1970), 181 -- 182.
28
M.D. ACOSTA, J.L. D ´AVILA, AND M. SOLEIMANI
[12] F.F. Bonsall and J. Duncan, Numerical Ranges II, London Math. Soc. Lecture Note Ser., vol. 10, Cambridge Univ. Press, Cambridge,
1973.
[13] J. Bourgain, On dentability and the Bishop-Phelps property, Israel J. Math. 28 (4) (1977), 265 -- 271.
[14] B. Cascales, A.J. Guirao and V. Kadets, A Bishop-Phelps-Bollob´as type theorem for uniform algebras, Adv. Math. 240 (2013),
370 -- 382.
[15] M. Chica, V. Kadets, M. Mart´ın, S. Moreno-Pulido and F. Rambla-Barreno, Bishop-Phelps-Bollob´as moduli of a Banach space, J.
Math. Anal. Appl. 412 (2014), no. 2, 697 -- 719.
[16] Y.S. Choi and S.K. Kim, The Bishop-Phelps-Bollob´as theorem for operators from L1(µ) to Banach spaces with the Radon-Nikod´ym
property, J. Funct. Anal. 261 (2011), no. 6, 1446 -- 1456.
[17] Y.S. Choi, S.K. Kim, H.J. Lee and M. Mart´ın, The Bishop-Phelps-Bollob´as theorem for operators on L1(µ), J. Funct. Anal. 267
(2014), no. 1, 214 -- 242.
[18] J. Diestel and J.J. Uhl, Vector Measures, Am. Math. Soc., Math. Surveys 15, Providence, RI, 1977.
[19] W.T. Gowers, Symmetric block bases of sequences with large average growth, Israel J. Math. 69 (1990), 129 -- 151.
[20] S.K. Kim, The Bishop-Phelps-Bollob´as theorem for operators from c0 to uniformly convex spaces, Israel J. Math. 197 (2013), 425 -- 435.
[21] S.K. Kim and H.J. Lee, The Bishop-Phelps-Bollob´as property for operators from C(K) to uniformly convex spaces, J. Math. Anal.
Appl. 421 (2015), no. 1, 51 -- 58.
[22] S.K. Kim, H.J. Lee and P.K. Lin, The Bishop-Phelps-Bollob´as property for operators from L∞(µ) to uniformly convex Banach spaces,
J. Nonlinear Convex Anal. 17 (2016), no. 2, 243 -- 249.
[23] J. Lindenstrauss, On operators which attain their norm, Israel J. Math. 1 (1963), 139 -- 148.
Universidad de Granada, Facultad de Ciencias, Departamento de An´alisis Matem´atico, 18071 Granada, Spain
E-mail address: [email protected]
Departamento de Matem´atica, Facultad de Ciencias, Universidad de Los Andes, M´erida, 5101 Venezuela
E-mail address: [email protected]
School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box: 19395-5746, Tehran, Iran
E-mail address: [email protected]
|
1606.01071 | 1 | 1606 | 2016-06-03T12:54:24 | On modulus of noncompact convexity for a strictly minimalizable measure of noncompactness | [
"math.FA"
] | In this paper we consider modulus of noncompact convexity $\Delta_{X,\varPhi}$ associated with the strictly minimalizable measure of noncompactness $\varPhi$. We also give some its properties and show its continuity on the interval $[0,\varPhi(\bar{B}_X))$. | math.FA | math |
ON MODULUS OF NONCOMPACT CONVEXITY FOR A
STRICTLY MINIMALIZABLE MEASURE OF
NONCOMPACTNESS
AMRA REKI ´C-VUKOVI ´C1 NERMIN OKI CI ´C2 1 IVAN ARAND- ELOVI ´C2
Abstract. In this paper we consider modulus of noncompact convexity ∆X,φ
associated with the strictly minimalizable measure of noncompactness φ. We
also give some its properties and show its continuity on the interval [0, φ(BX )).
1. Introduction and preliminaries
The theory of measures of noncompactness has many applications in Topol-
ogy, Functional analysis and Operator theory. There are many nonequivalent
definitions of this notion on metric and topological spaces [1],[4]. First of them
was introduced by Kuratowski in 1930. The most important examples of such
functions are: Kuratowski's measure (α), Hausdorff's measure (χ) and measure
of Istratescu (β).
One of the tools that provides classification of Banach spaces considering their
geometrical properties is the modulus of convexity [5]. Its natural generalization
is the notion of noncompact convexity, introduced by K. Goebel and T. Sekowski
[9]. Their modulus was generated with Kuratowski's measure of noncompact-
ness (α). Modulus associated with Hausdorff's measure of noncompactness (χ),
was introduced by Banas [3] and modulus associated with Istratescu's measure of
noncompactness (β) by Dominguez-Benavides and Lopez [8]. In [11], [2] was pre-
sented an abstract unified to this notions, which consider modulus of noncompact
convexity ∆X,φ associated with arbitrary abstract measure of noncompactness φ.
Banas [3] proved that modulus ∆X,χ(ε) is subhomogeneous function in the case
of reflexive space X. Moreover, Prus [11] gave the result connecting continuity of
the modulus ∆X,φ(ε) and uniform Opial condition which imply a normal structure
of the space.
In this paper, we shall prove that modulus ∆X,φ(ε) is subhomogeneous and
a continuous function on the interval [0, φ(BX)), for an arbitrary strictly min-
imalizable measure of noncompactness φ, where X is Banach space having the
Radon-Nikodym property.
Date: Received: xxxxxx; Revised: yyyyyy; Accepted: zzzzzz.
2010 Mathematics Subject Classification. 46B20; 46B22.
Key words and phrases. modulus of noncompact convexity, strictly minimalizable measure
of noncompactness, continuity.
1
2
A. REKI ´C-VUKOVI ´C, N. OKI CI ´C, I. ARAND- ELOVI ´C
2. Preliminaries
Let X be Banach space, B(x, r) denotes the open ball centered at x with radius
r, and BX and SX denote the unit ball and sphere in the given space, respectively.
If A ⊂ X with A and coA we denote closure of the set A, that is convex hull of
A.
Let B be the collection of bounded subsets of the space X. Then function
φ : B → [0, +∞) with properties:
(1) φ(B) = 0 ⇔ B is precompact set,
(2) φ(B) = φ(B), ∀B ∈B,
(3) φ(B1 ∪ B2) = max{φ(B1), φ(B2)}, ∀B1, B2 ∈B,
is the measure of noncompactness defined on X. For more about properties of
the measure of noncompactness see in [1] and [2].
Let φ be a measure of noncompactness. Infinite set A ∈ B is φ-minimal if and
only if φ(A) = φ(B) for any infinite set B ⊆ A.
We say that the measure of noncompactness φ is minimalizable if for every
infinite, bounded set A and for all ε > 0, there exists φ-minimal set B ⊂ A, such
that φ(B) ≥ φ(A) − ε. Measure φ is strictly minimalizable if for every infinite,
bounded set A, there exists φ-minimal set B ⊂ A such that φ(B) = φ(A).
The modulus of noncompact convexity associated to arbitrary measure of non-
compactness φ is the function ∆X,φ : [0, φ(BX)] → [0, 1], defined with
∆X,φ(ε) = inf{1 − d(0, A) : A ⊆ BX , A = coA = A, φ(A) ≥ ε} .
Banas [3] considered modulus ∆X,φ(ε) for φ = χ, where χ is Hausdorff measure
of noncompactness. For φ = α, α is Kuratowski measure of noncompactness,
∆X,α(ε) presents Gobel-Sekowski modulus of noncompact convexity [9], while for
φ = β, where β is a separation measure of noncompactness, ∆X,β(ε) is Dominguez
Benavides -Lopez modulus of noncompact convexity [8].
The characteristic of noncompact convexity of X associated to the measure of
noncompactness φ is defined to be
εφ(X) = sup{ε ≥ 0 : ∆X,φ(ε) = 0} .
Inequalities
hold for moduli ∆X,φ(ε) concerning φ = α, χ, β, and consequently
∆X,α(ε) ≤ ∆X,β(ε) ≤ ∆X,χ(ε) ,
εα(X) ≥ εβ(X) ≥ εχ(X) .
Banach space X has Radon-Nikodym property if and only if every nonempty,
bounded set A ⊂ X is dentable, that is if and only if for all ε > 0 there exists
x ∈ A, such that x /∈ co(A\B(x, ε)), [6].
3. Main results
We started with our main result which is partial generalizations of earlier result
obtained by Banas [3].
ON MODULUS OF NONCOMPACT CONVEXITY ...
3
Theorem 3.1. Let X be a Banach space with Radon-Nikodym property and φ
strictly minimalizable measure of noncompactness. The modulus ∆X,φ(ε) is sub-
homogeneous function, that is
∆X,φ(kε) ≤ k∆X,φ(ε),
(3.1)
for all k ∈ [0, 1], ε ∈ [0, φ(BX)].
Proof. Let η > 0 be arbitrary and ε ∈ [0, φ(BX )]. From the definition of the
modulus ∆X,φ(ε) there exists convex, closed set A ⊆ BX , φ(A) ≥ ε such that
1 − d(0, A) < ∆X,φ(ε) + η .
(3.2)
The set kA ⊆ BX is convex and closed for arbitrary k ∈ [0, 1] and φ(kA) =
kφ(A) ≥ kε. Since φ is strictly minimalizable measure of noncompactness there
is infinite φ-minimal set B ⊂ kA such that
φ(B) = φ(kA) ≥ kε .
Let n0 ∈ N be such that
<
. Since the set B is bounded subset
of Banach space X which has Radon-Nikodym property, we conclude that for
kε
n0
diamB
2
r =
there exists x0 ∈ B such that
kε
n0
that is
x0 /∈ co(cid:2)B\B (x0, r)(cid:3) ,
co(cid:2)B\B (x0, r)(cid:3) ⊂ B ⊂ kA .
We consider the set C = co(cid:2)B\B (x0, r)(cid:3). C is closed and convex set and because
of the properties of the strictly minimalizable measure of noncompactness we have
Moreover,
φ(C) = φ(cid:2)B\B (x0, r)(cid:3) = φ(B) ≥ kε .
1 − d(0, C) ≤ 1 − d(0, kA) = 1 − kd(0, A).
(3.3)
We define the set B ∗ = C +
1 − k
kx0k
arbitrary z ∈ B ∗ is of the form z = y +
x0. B ∗ is a convex and closed set and
1 − k
kx0k
x0, where y ∈ C ⊂ kA and kzk < 1.
So, B ∗ ⊂ BX. From the properties of the measure of noncompactness φ we have
that
φ(B ∗) = φ(cid:18)C +
1 − k
kxk
x(cid:19) = φ(C) ≥ kε .
Since d(0, B ∗) = d(0, C) + 1 − k, than 1 − d(0, B ∗) = k − d(0, C) ≤ k − kd(0, A)
holds, so using inequality (3.2) we have
If we take infimum by all sets B, such that φ(B) ≥ kε, we have
1 − d(0, B ∗) < k(∆X,φ(ε) + η) .
∆X,φ(kε) ≤ k∆X,φ(ε) + kη .
4
A. REKI ´C-VUKOVI ´C, N. OKI CI ´C, I. ARAND- ELOVI ´C
Since η > 0 can be choosen arbitrarily small we obtain
∆X,φ(kε) ≤ k∆X,φ(ε).
(cid:3)
As applications of Theorem we shall state the following corollaries.
Corollary 3.2. Let φ be a strictly minimalizable measure of noncompactness de-
fined on space X with Radon-Nikodym property. The function ∆X,φ(ε) is strictly
increasing on the interval [εφ(X), φ(BX )].
Proof. Let ε1, ε2 ∈ [εφ(X), φ(BX)] and ε1 < ε2. If we put k =
the Theorem 3 we obtain
∆X,φ(ε1) = ∆X,φ(kε2) ≤ k∆X,φ(ε2) < ∆X,φ(ε2) .
ε1
ε2
< 1, then by
(cid:3)
Corollary 3.3. Let φ be a strictly minimalizable measure of noncompactness
defined on space X with Radon-Nikodym property. Inequality
holds for all ε ∈ [0, φ(BX)].
∆X,φ(ε) ≤ ε
Proof. If ε ∈ [0, 1] and if k and ε change roles, and ε take the value ε = 1, than
by using Theorem 3 we have
∆X,φ(ε) ≤ ε∆X,φ(1) ≤ ε .
If ε ∈ (1, φ(BX )], than the monotonicity of the modulus ∆X,φ(ε) provides that
∆X,φ(ε) ≤ ∆X,φ(φ(BX)) = 1 < ε .
(cid:3)
Corollary 3.4. Let φ be a strictly minimalizable measure of noncompactness de-
∆X,φ(ε)
fined on the space X with Radon-Nikodym property. The function f (ε) =
ε
is nondecreasing on the interval [0, φ(BX)] and for ε1 + ε2 ≤ φ(BX) it holds
∆X,φ(ε1 + ε2) ≥ ∆X,φ(ε1) + ∆X,φ(ε2).
(3.4)
Proof. Let ε1, ε2 ∈ [0, φ(BX)] such that ε1 ≤ ε2. We put k =
ε1
ε2
. Then
f (ε1) =
∆X,φ(ε1)
ε1
=
∆X,φ(kε2)
ε1
.
If we use a property of subhomegeneity of the function ∆X,φ(ε) we have
f (ε1) ≤
∆X,φ(ε2)
ε2
= f (ε2) ,
which proves that f (ε) is a nondecreasing function on the interval [0, φ(BX)].
Furthermore
ON MODULUS OF NONCOMPACT CONVEXITY ...
5
∆X,φ(ε1) + ∆X,φ(ε2) ≤ k∆X,φ(ε2) + ∆X,φ(ε2)
=
ε1 + ε2
ε2
∆X,φ(ε2)
≤ (ε1 + ε2)
∆X,φ(ε1 + ε2)
ε1 + ε2
= ∆X,φ(ε1 + ε2).
So, inequality (3.4) is proved.
(cid:3)
Corollary 3.5. Let φ be a strictly minimalizable measure of noncompactness
defined on the space X with Radon-Nikodym property.
∆X,φ(ε2) − ∆X,φ(ε1)
ε2 − ε1
≥
∆X,φ(ε2)
ε2
.
(3.5)
holds for all ε1, ε2 ∈ (ε1(X), φ(BX)], ε1 ≤ ε2.
Proof. Let k =
≤ 1. From the Theorem 3 we have
ε1
ε2
∆X,φ(ε2) − ∆X,φ(ε1) = ∆X,φ(ε2) − ∆X,φ(kε2)
≥ ∆X,φ(ε2) − k∆X,φ(ε2)
=
ε2 − ε1
ε2
∆X,φ(ε2).
Thus the inequality (3.5) is proved.
(cid:3)
4. Continuity of the modulus of noncompact convexity
In this section we shall prove continuity of the modulus of noncompact convex-
ity associated with arbitrary strictly minimalizable measure of noncompactness,
defined on Banach space with Radon-Nikodym property. Now we need the fol-
lowing Lemmas.
Lemma 4.1. Let φ be an arbitrary measure of noncompactness, B ⊂ BX φ-
minimal set and k > 0 arbitrary. Then the set kB is φ-minimal.
Proof. Let B be φ-minimal subset of closed unit ball and D ⊂ kB arbitrary
infinite set for k > 0. If y ∈
D, then y =
y ′ for some y ′ ∈ D. Since y ′ ∈ kB,
then y ′ = kx for some x ∈ B. So y =
is φ-minimal set we have
1
k
1
k
1
k
kx = x ∈ B, therefore
1
k
D ⊂ B. Since B
φ(B) = φ(cid:18) 1
k
D(cid:19) =
1
k
φ(D),
that is φ(D) = kφ(B) = φ(kB). Thus kB is φ-minimal set.
(cid:3)
Lemma 4.2. Let X be a Banach space with Radon-Nikodym property and φ a
strictly minimalizable measure of noncompactness. The modulus of noncompact
convexity ∆X,φ(ε) is left continuous function on the interval [0, φ(BX)).
6
A. REKI ´C-VUKOVI ´C, N. OKI CI ´C, I. ARAND- ELOVI ´C
Proof. Let ε0 ∈ [0, φ(BX )) be arbitrary and let ε < ε0. From the definition of
∆X,φ(ε), for arbitrary η > 0 there exists convex and closed set A ⊂ BX , φ(A) ≥ ε
such that
1 − d(0, A) < ∆X,φ(ε) + η .
Since φ is strictly minimalizable measure of noncompactness there exists infinite
φ-minimal set B ⊂ A such that φ(A) = φ(B) ≥ ε. Moreover, inequality
1 − d(0, B) ≤ 1 − d(0, A) < ∆X,φ(ε) + η
holds for the set B. Let n0 ∈ N be such that
<
. Since B is a bounded
ε
n0
diamB
2
subset of the Banach space X with Radon-Nikodym property ([6]) we conclude
that for r =
there exists x0 ∈ B such that
ε
n0
This means that there is a convex and closed set C = co(cid:2)B\B (x0, r)(cid:3) ⊂ A, where
1 − d(0, C) ≤ 1 − d(0, A) < ∆X,φ(ε) + η .
x0 /∈ co(cid:2)B\B (x0, r)(cid:3) .
Since φ(B) ≥ ε we have that B\B (x0, r) is an infinite set and using properties
of the strictly minimalizable measure of noncompactness φ we obtain
φ(C) = φ(cid:2)B\B (x0, r)(cid:3) = φ(B) ≥ ε .
1 − d(0, C)
Let k = 1 +
convex set such that A∗ ⊆ kC ⊂ kB. From the Lemma 4.1 we have
. We will consider set A∗ = kC ∩ BX. A∗ is a closed and
2
Moreover, inequality
φ(A∗) = φ(kB) = kφ(B) ≥ kε .
1 − d(0, A∗) ≤ 1 − d(0, kC) < 1 − d(0, C) ,
holds, i.e.
1 − d(0, A∗) < ∆X,φ(ε) + η .
(4.1)
Let δ = ε0(cid:18)1 −
1
k(cid:19). For ε ∈ (ε0 − δ, ε0) we have
φ(A∗) ≥ kε > k(ε0 − δ) = k
ε0
k
= ε0 .
If we take infimum in (4.1) by all the sets A∗ such that φ(A∗) > ε0, we obtain
inf{1 − d(0, A∗) : A∗ ⊂ BX, A∗ = co(A∗), φ(A∗) > ε0} ≤ ∆X,φ(ε) + η ,
that is
∆X,φ(ε0) ≤ ∆X,φ(ε) + η .
Since η > 0 was arbitrarily small we conclude lim
ε→ε0
−
∆X,φ(ε) = ∆X,φ(ε0) .
(cid:3)
Lemma 4.3. Let X be a Banach space with Radon-Nikodym property and φ a
strictly minimalizable measure of noncompactness. The function ∆X,φ(ε) is right
continuous on the interval [0, φ(BX)).
ON MODULUS OF NONCOMPACT CONVEXITY ...
7
Proof. Let η > 0 and ε0 ∈ [0, φ(BX )). From the definition of the modulus
of noncompact convexity ∆X,φ(ε), there exists convex and closed set A ⊂ BX,
φ(A) ≥ ε0, such that
1 − d(0, A) < ∆X,φ(ε0) + η .
Since φ is strictly minimalizable measure we conclude that there exists infinite,
φ-minimal set B ⊂ A such that φ(A) = φ(B) ≥ ε0 and
1 − d(0, B) ≤ 1 − d(0, A) < ∆X,φ(ε0) + η .
The set B is bounded subset of the Banach space that has Radon-Nikodym
diamB
property, and because of that for r =
, where n0 ∈ N is such that
ε0
n0
ε0
n0
<
2
we can find x0 ∈ B such that
x0 /∈ co(cid:2)B\B (x0, r)(cid:3) .
This implies that there is convex and closed set C = co(cid:2)B\B (x0, r)(cid:3) ⊂ A, where
1 − d(0, C) ≤ 1 − d(0, A) < ∆X,φ(ε0) + η .
Hence, by the properties of the strictly minimalizable measure of noncompactness
φ we obtain
(1) If φ(C) > ε0, let δ = φ(C) − ε0 and consider arbitrary ε ∈ (ε0, ε0 + δ).
φ(C) = φ(cid:2)B\B (x0, r)(cid:3) = φ(B) ≥ ε0 .
Then φ(C) > ε, hence
inf{1 − d(0, C) : C ⊂ BX , C = coC, φ(C) > ε} ≤ ∆X,φ(ε0) + η .
Moreover,
∆X,φ(ε) ≤ ∆X,φ(ε0) + η .
This completes the proof of the Theorem.
(2) Let φ(C) = ε0. Consider the set B ∗ = kC ∩ BX for k = 1 +
.
B ∗ is closed, convex set and it holds that B ∗ ⊆ kC ⊂ kB. Hence by the
Lemma 4.1 we have
2
1 − d(0, C)
φ(B ∗) = φ(kB) = kφ(B) = kε0 .
Moreover, next inequality holds
1 − d(0, B ∗) ≤ 1 − d(0, kC) < 1 − d(0, C) .
Let δ = ε0(1 − k). Then for ε ∈ (ε0, ε0 + δ) we have φ(B ∗) = kε0 > ε, thus
inf{1 − d(0, B ∗) : B ∗ ⊂ BX, B ∗ = coB ∗, φ(B ∗) > ε} ≤ ∆X,φ(ε0) + η ,
that is
∆X,φ(ε) ≤ ∆X,φ(ε0) + η .
Therefore, we obtain lim
ε→ε0+
∆X,φ(ε) = ∆X,φ(ε0). This completes the proof.
(cid:3)
From Lemma 4.2 and Lemma 4.3 follows that:
8
A. REKI ´C-VUKOVI ´C, N. OKI CI ´C, I. ARAND- ELOVI ´C
Theorem 4.4. Let X be a Banach space with Radon-Nikodym property and φ
a strictly minimalizable measure of noncompactness. The modulus ∆X,φ(ε) is a
continuous function on the interval [0, φ(BX )).
It is known that the Hausdorff measure of noncompactness χ is a strictly mini-
malizable measure of noncompactness in the weakly compactly generated Banach
spaces ([2], Theorem III. 2.7.). Since reflexive spaces are weakly compactly gener-
ated and also have Radon-Nikodym property [10], we conclude that the modulus
of noncompact convexity ∆X,χ, associated to measure of noncompactness χ, is a
continuous function on the interval [0, φ(BX)) in an arbitrary weakly compactly
generated space X.
Acknowledgement. The third author was partially supported by Ministry of
Education, Science and Technological Development of Serbia, Grant No. 174002,
Serbia.
References
[1] R. R. Akhmerov, M. I. Kamenskii, A. S. Potapov, A. E. Rodkina and B. N. Sadovskii,
Measures of Noncompactness and Condensing Operators, Birkhauser Verlag, 1992.
[2] J. M. Ayerbe Toledano, T. Dominguez Benavides and G. Lopez Acedo, Measures of Non-
compactness in Metric Fixed point Theory, Birkhauser Verlag, 1997.
[3] J. Banas, On modulus of noncompact convexity, its properties and applications, Canad.
Math. Bull. 30 (2) (1987), 186 -- 192.
[4] J. Banas and K. Goebel, Measures of Noncompactness in Banach Spaces, Marcel Dekker,
Inc., 1980.
[5] J.A. Clarkson, Uniformly convex spaces, Trans. Amer. Math. Soc. 40 (1936) 394 -- 414.
[6] W.J.Davis and R.P.Phelps, The Radon-Nikodym property and dentable sets in Banach
spaces, Proc. Amer. Math. Soc. 45 (1) (1974), 199 -- 122.
[7] T. Domingez Benavides, Some properties of the set and ball measures of noncompactness
and applications, J. Lond. Math. Soc. (2) 34 (1986), 120 -- 128.
[8] T. Domingez Benavides and G.L.Acedo, Lower bounds for normal structure coefficients,
Proc. Roy. Soc. Edinburgh Sect. A 121 (3-4) (1992), 245 -- 252.
[9] K.Goebel, T.Sekowski, The modulus of noncompact convexity, Ann.Univ. Mariae Curie-
Sklodowska Sect. A 38 (1984) 41 -- 48.
[10] M. Gonzalez, A. Martinez-Abejon, Tauberian operators, Birkhauser Verlag AG, Basel-
Boston-Berlin, 2010.
[11] S. Prus, Banach spaces with the uniform Opial property, Nonlinear Anal. 18(8), 697 -- 704
(1992)
1 Department of Mathematics, University of Tuzla, Univerzitetska 4, Tuzla,
Bosnia and Hercegovina.
E-mail address: [email protected]
E-mail address: [email protected]
2 University of Belgrade - Faculty of Mechanical Engineering,, Kraljice Mar-
ije 16, 11000 Belgrade, Serbia
E-mail address: [email protected]
|
1901.06634 | 1 | 1901 | 2019-01-20T07:32:45 | Some Hermite-Hadamard type inequalities in the class of hyperbolic p-convex functions | [
"math.FA"
] | In this paper, obtained some new class of Hermite-Hadamard and Hermite-Hadamard-Fejer type inequalities via fractional integrals for the p-hyperbolic convex functions. It is shown that such inequalities are simple consequences of Hermite-Hadamard-Fejer inequality for the p-hyperbolic convex function. | math.FA | math |
SOME HERMITE-HADAMARD TYPE INEQUALITIES IN THE
CLASS OF HYPERBOLIC P-CONVEX FUNCTIONS
SILVESTRU SEVER DRAGOMIR AND BERIKBOL T. TOREBEK*
Abstract. In this paper, obtained some new class of Hermite-Hadamard and
Hermite-Hadamard-Fej´er type inequalities via fractional
integrals for the p-
hyperbolic convex functions.
It is shown that such inequalities are simple con-
sequences of Hermite-Hadamard-Fej´er inequality for the p-hyperbolic convex func-
tion.
Contents
Introduction
1.
1.1. Definitions of fractional integrals
1.2. Some generalizations of Hermite-Hadamard and Hermite-Hadamard-
Fej´er inequalities
1.3. Hyperbolic p-convex functions
1.4. Hermite-Hadamard and Hermite-Hadamard-Fej´er inequalities for
hyperbolic p-convex functions
2. Main results
2.1. Fractional analogues of Hermite-Hadamard inequality for hyperbolic
p-convex functions
2.2. Fractional analogues of Hermite-Hadamard-Fej´er inequality for
hyperbolic p-convex functions
Acknowledgements
References
1
2
3
4
5
6
6
8
10
10
1. Introduction
The inequalities for convex functions due to Hermite and Hadamard are found to be
of great importance, for example, see [DP00, PPT92]. According to the inequalities
[H93, H83],
• if u : I → R is a convex function on the interval I ⊂ R and a, b ∈ I with
b > a, then
(1.1)
b
u(cid:18)a + b
2 (cid:19) ≤
1
b − a
Za
u(y)dy ≤
u(a) + u(b)
2
.
2010 Mathematics Subject Classification. Primary 26D10; Secondary 26A33, 35A23.
Key words and phrases. Hermite-Hadamard inequality, Hermite-Hadamard-Fej´er inequality, hy-
perbolic p-convex function, fractional integral.
* Corresponding author. E-mail: [email protected].
1
2
DRAGOMIR AND TOREBEK
For a concave function u, the inequalities in (1.1) hold in the reversed direction.
Definition 1.1. A function u : [a, b] ⊂ R → R is said to be convex if
u(µx + (1 − µ)y) ≤ µu(x) + (1 − µ)u(y)
for all x, y ∈ [a, b] and µ ∈ [0, 1]. We call u a concave function if (−u) is convex.
We note that Hadamard's inequality refines the concept of convexity, and it follows
from Jensen's inequality. The classical Hermite-Hadamard inequality yields estimates
for the mean value of a continuous convex function u : [a, b] → R. The well-known
inequalities dealing with the integral mean of a convex function u are the Hermite-
Hadamard inequalities or its weighted versions. They are also known as Hermite-
Hadamard-Fej´er inequalities.
In [F06], Fej´er obtained the weighted generalization of Hermite-Hadamard inequal-
ity (1.1) as follows.
• Let u : [a, b] → R be a convex function. Then the inequality
Za
u(cid:18)a + b
2 (cid:19)
u(y)v(y)dy ≤
v(y)dy ≤
u(a) + u(b)
Za
Za
2
b
b
b
v(y)dy
(1.2)
holds for a nonnegative, integrable function v : [a, b] → R, which is symmetric
to a+b
2 .
Clearly, for v(x) ≡ 1 on [a, b] we get (1.1).
1.1. Definitions of fractional integrals. Let us give basic definitions of fractional
integrations of the different types.
Definition 1.2. [KST06] The left and right Riemann -- Liouville fractional integrals
a+ and I α
I α
b− of order α ∈ R (α > 0) are given by
and
I α
a+ [f ] (t) =
I α
b− [f ] (t) =
1
Γ (α)
1
Γ (α)
t
Za
b
Zt
(t − s)α−1 f (s)ds, t ∈ (a, b],
(s − t)α−1 f (s)ds, t ∈ [a, b),
respectively. Here Γ denotes the Euler gamma function.
Definition 1.3. [KT16] Let f ∈ L1(a, b). The fractional integrals I α
order α ∈ (0, 1) are defined by
(x − s)(cid:19) u(s)ds, x > a
1
Za
α
exp(cid:18)−
I α
a+u(x) =
1 − α
α
(1.3)
x
a+ and I α
b− of
and
(1.4)
I α
b−u(x) =
b
1
α
Zx
exp(cid:18)−
1 − α
α
(s − x)(cid:19) u(s)ds, x < b
SOME HERMITE-HADAMARD TYPE INEQUALITIES
3
respectively.
1.2. Some generalizations of Hermite-Hadamard and Hermite-Hadamard-
Fej´er inequalities. Here we present some results on the generalization of the above
inequalities.
In [SSYB13], Sarikaya et.
al.
represented Hermite-Hadamard inequality in
Riemann-Liouville fractional integral forms as follows.
• Let u : [a, b] → R be a positive function and u ∈ L1([a, b]). If u is a convex
function on [a, b], then the following inequalities for fractional integrals hold
(1.5)
u(cid:18)a + b
2 (cid:19) ≤
with α > 0.
Γ(α + 1)
2(b − a)α [I α
a u(b) + I α
b u(a)] ≤
u(a) + u(b)
2
In [I16], I¸scan gave the following Hermite-Hadamard-Fej´er integral inequalities via
fractional integrals:
• Let u :
[a, b] → R be convex function with a < b and u ∈ L1([a, b]). If
v : [a, b] → R is nonnegative, integrable and symmetric to (a + b)/2, then the
following inequalities for fractional integrals hold
u(cid:18)a + b
2 (cid:19) [I α
(1.6)
a v(b) + I α
b v(a)] ≤ [I α
a (uv)(b) + I α
b (uv)(a)]
u(a) + u(b)
2
≤
[I α
a v(b) + I α
b v(a)]
with α > 0.
In [KT16] the authors obtained the following generalizations of inequality (1.1) and
(1.2)
1 − α
• Let u : [a, b] → R and u ∈ L1(a, b). If u is a convex function on [a, b], then the
2 (cid:19) ≤
[a, b] → R be convex and integrable function with a < b. If w :
2 , that
following inequalities for fractional integrals I α
u(cid:18)a + b
a u(b) + I α
[I α
• Let u :
[a, b] → R is nonnegative, integrable and symmetric with respect to a+b
is, w(a + b − x) = w(x), then the following inequalities hold
u(cid:18)a + b
b (uw) (a)]
a+ and I α
b u(a)] ≤
a+ hold:
u(a) + u(b)
a w(b) + I α
b w(a)] ≤ [I α
2 (1 − exp (−ρ))
2 (cid:19) [I α
2
;
a (uw) (b) + I α
u(a) + u(b)
[I α
a w(b) + I α
2
≤
b w(a)] .
(1.7)
(1.8)
Many generalizations and extensions of the Hermite-Hadamard and Hermite-
Hadamard-Fej´er type inequalities were obtained for various classes of functions using
fractional integrals; see [C16, CK17, HYT14, I16, JS16, SSYB13, WLFZ12, ZW13]
and references therein.
In [KS18] Kirane and Samet show that most of those re-
sults are particular cases of (or equivalent to) existing inequalities from the litera-
ture. These studies motivated us to consider a new class of functional inequalities
for hyperbolic p-convex functions generalizing the classical Hermite-Hadamard and
Hermite-Hadamard-Fej´er inequalities.
4
DRAGOMIR AND TOREBEK
1.3. Hyperbolic p-convex functions. We consider the hyperbolic functions of a
real argument x ∈ R defined by
sinh x :=
cosh x :=
tanh x :=
coth x :=
ex − e−x
ex + e−x
2
,
,
2
sinh x
cosh x
cosh x
sinh x
,
.
Definition 1.4. [D18a, D18b] We say that a function f : I → R is hyperbolic p-
convex (or sub H-function, according with [A16]) on I, if for any closed subinterval
[a, b] of I we have
(1.9)
for all x ∈ [a, b].
f (x) ≤
sinh[p(b − x)]
sinh[p(b − a)]
f (a) +
sinh[p(x − a)]
sinh[p(b − a)]
f (b)
p-concave on I.
If the inequality (1.9) holds with "≥", then the function will be called hyperbolic
Geometrically speaking, this means that the graph of f on [a, b] lies nowhere above
the p-hyperbolic function determined by the equation
H(x) = H(x; a, b, f ) := A cosh(px) + B sinh(px),
where A and B are chosen such that H(a) = f (a) and H(b) = f (b).
If we take x = (1 − t)a + tb ∈ [a, b], t ∈ [0, 1], then the condition (1.9) becomes
(1.10)
f ((1 − t)a + tb) ≤
sinh[p(1 − t)(b − a)]
sinh[p(b − a)]
f (a) +
sinh[pt(b − a)]
sinh[p(b − a)]
f (b)
for any t ∈ [0, 1]. We have the following properties of hyperbolic p-convex function
on I.
(i): A hyperbolic p-convex function f : I → R has finite right and left deriva-
+(x). The function
tives f ′
f is differentiable on I with the exception of an at most countable set.
−(x) at every point x ∈ I and f ′
−(x) ≤ f ′
+(x) and f ′
(ii): A necessary and sufficient condition for the function f : I → R to be
hyperbolic p-convex function on I is that it satisfies the gradient inequality
f (y) ≥ f (x) cosh[p(y − x)] + Kx,f sin[p(y − x)]
for any x, y ∈ I, where Kx,f ∈ (cid:2)f ′
point x then Kx,f = f ′(x).
−(x), f ′
+(x)(cid:3) . If f is differentiable at the
(iii): A necessary and sufficient condition for the function f to be a hyperbolic
p-convex in I, is that the function
ϕ(x) = f ′(x) − p2
x
Za
f (t)dt
is nondecreasing on I, where a ∈ I.
SOME HERMITE-HADAMARD TYPE INEQUALITIES
5
(iv): Let f : I → R be a two times continuously differentiable function on I.
Then f is hyperbolic p-convex on I if and only if for all x ∈ I we have
For other properties of hyperbolic p-convex functions, see [A16].
f ′′(x) − p2f (x) ≥ 0.
p2
f ′′
f ′′
We observe that
Consider the function fr : (0,∞) → (0,∞), fr(x) = xr with p ∈ R\{0}. If r ∈
(−∞, 0) ∪ [1,∞) the function is convex and if r ∈ (0, 1) it is concave. We have for
r ∈ (−∞, 0) ∪ [1,∞)
r (x) − p2fr(x) = r(r − 1)xr−2 − p2xr = p2xr−2(cid:18)r(r − 1)
− x2(cid:19) , x > 0.
!
r (x) − p2fr(x) > 0 for x ∈ 0,pr(r − 1)
p
,∞! ,
r (x) − p2fr(x) < 0 for x ∈ pr(r − 1)
p
,∞(cid:19) .
which shows that the power function fr for r ∈ (−∞, 0)∪[1,∞) is hyperbolic p-convex
on (cid:18)0,
If r ∈ (0, 1), then
p (cid:19) and hyperbolic p-concave on (cid:18)√r(r−1)
r (x) − p2fr(x) < 0
f ′′
√r(r−1)
and
for any x > 0, which shows that fr is hyperbolic p-concave on (0,∞).
f ′′
p
Consider the exponential function
fµ(x) = exp(µx) = eµx
for µ 6= 0 and x ∈ R. Then
f ′′
µ (x) − p2fµ(x) = µ2eµx − p2eµx = (µ2 − p2)eµx, x > 0.
If µ > p, then fµ is hyperbolic p-convex on R and if µ < p, then fµ is hyperbolic
p-concave on R.
1.4. Hermite-Hadamard and Hermite-Hadamard-Fej´er inequalities for hy-
perbolic p-convex functions. In [D18a], Dragomir obtained Hermite-Hadamard
type inequality in the class of hyperbolic p-convex functions as follows
Theorem 1.1. Assume that the function u : I → R is hyperbolic p-convex function
on I. Then for any a, b ∈ I we have
(cid:19) ≤
2 (cid:19) sinh(cid:18)p(b − a)
(cid:19) .
Remark 1.1. Note that, if p → 0 in (1.11) we get the classical Hermite-Hadamard
inequality (1.1).
tanh(cid:18)p(b − a)
u(cid:18)a + b
f (x)dx ≤
u(a) + u(b)
(1.11)
Za
2
p
p
2
2
b
Hermite-Hadamard-Fej´er type inequalities in the class of hyperbolic p-convex func-
tions was proven in [D18b]:
6
DRAGOMIR AND TOREBEK
Theorem 1.2. Assume that the function u : I → R is hyperbolic p-convex on I and
a, b ∈ I. Assume also that w : I → R is a positive, symmetric and integrable function
on [a, b], then we have
b
u(cid:18)a + b
2 (cid:19)
Za
(1.12)
≤
a + b
2 (cid:19)(cid:21) w(x)dx
cosh(cid:20)p(cid:18)x −
Za
u(x)w(x)dx
b
≤
u(a) + u(b)
2 (cid:19)(cid:21) w(x)dx.
Remark 1.2. Note that, if p → 0 in (1.12) we get the classical Hermite-Hadamard-
Fej´er inequality (1.2).
cosh(cid:20)p(cid:18)x −
cosh−1 p(b − a)
a + b
Za
2
2
b
Theorem 1.3. Assume that the function u : I → R is hyperbolic p-convex on I and
a, b ∈ I. Assume also that w : I → R is a positive, symmetric and integrable function
on [a, b], then
(1.13)
b
Za
≤
+
u(x)w(x)dx
u(a) + u(b)
2
u(a) − u(b)
2
b
2
cosh−1(cid:18)p(b − a)
(cid:19)
Za
(cid:19)
sinh−1(cid:18) p(b − a)
Za
2
b
cosh(cid:20)p(cid:18)x −
sinh(cid:20)p(cid:18)x −
a + b
a + b
2 (cid:19)(cid:21) w(x)dx
2 (cid:19)(cid:21) w(x)dx.
In this section, we formulate the main results of the paper.
2. Main results
2.1. Fractional analogues of Hermite-Hadamard inequality for hyperbolic
p-convex functions. Our first observation is formulated by the following theorem.
Theorem 2.1. Assume that the function u : I → R is hyperbolic p-convex function
on I. Then for any a, b ∈ I we have
(2.1) u(cid:18)a + b
2 (cid:19)Cα(1) ≤ I α
Ra
where Cα(1) :=
2 (cid:1)(cid:3) (b−x)α−1+(x−a)α−1
cosh(cid:2)p(cid:0)x − a+b
cosh−1 (p(b − a))Cα(1),
b−u(a) ≤
a+u(b) + I α
u(a) + u(b)
dx.
Γ(α)
2
b
SOME HERMITE-HADAMARD TYPE INEQUALITIES
7
Proof. Let us suppose that all assumptions of Theorem are satisfied. Let us define
the function w of Theorem 1.2 by
w(x) =
1
Γ(α)(cid:2)(b − x)α−1 + (x − a)α−1(cid:3) , a < x < b.
Clearly, w is a positive, symmetric and integrable function on [a, b]. Moreover, for all
x ∈ (a, b), we get
w(a + b − x) = (b − (a + b − x))α−1 + ((a + b − x) − a)α−1 = w(x).
Moreover, we have
b
Za
u(x)w(x)dx =
1
Γ(α)
b
Za
(b − x)α−1u(x)dx +
1
Γ(α)
b
Za
(x − a)α−1u(x)dx
= I α
b−u(a) + I α
a+u(b).
Therefore, from (1.12) we obtain inequality (2.1).
(cid:3)
Remark 2.1. Inequalities (1.1) and (1.5) are special cases of inequality (2.1).
• If p → 0 we have lim
p→0C1
with the inequality (1.5);
α(1) = 2(b−a)α
Γ(α+1) . In this case, inequality (2.1) coincide
• If p → 0 and α → 1 in (2.1), then we have classical Hermite-Hadamard
inequality (1.1).
Theorem 2.2. Assume that the function u : I → R is hyperbolic p-convex function
on I. Then for any a, b ∈ I we have
(2.2) u(cid:18)a + b
where C2
cosh−1 (p(b − a)) C2
α(1) ≤ I α
α (b−x))+exp(− 1−α
b−u(a) ≤
u(a) + u(b)
α(1) :=
α (x−a))
a+u(b) + I α
2 (cid:1)(cid:3) exp(− 1−α
2 (cid:19)C2
Ra
b
cosh(cid:2)p(cid:0)x − a+b
Proof. Suppose that all assumptions of Theorem are satisfied. Let us define the
function w of Theorem 1.2 by
α(1),
2
α
dx.
w(x) =
1
α(cid:20)exp(cid:18)−
1 − α
α
(b − x)(cid:19) + exp(cid:18)−
1 − α
α
(x − a)(cid:19)(cid:21) , a < x < b.
Clearly, w is a positive, symmetric and integrable function on [a, b]. Moreover, for all
x ∈ (a, b), we get
αw(a + b − x) =(cid:20)exp(cid:18)−
=(cid:20)exp(cid:18)−
(b − (a + b − x))(cid:19) + exp(cid:18)−
(x − a)(cid:19) + exp(cid:18)−
(b − x)(cid:19)(cid:21) = αw(x).
((a + b − x) − a)(cid:19)(cid:21)
1 − α
α
1 − α
α
1 − α
α
1 − α
α
8
DRAGOMIR AND TOREBEK
Moreover, we have
b
Za
u(x)w(x)dx
b
=
exp(cid:18)−
1
Za
α
= I α
b−u(a) + I α
a+u(b).
1 − α
α
b
(b − x)(cid:19) u(x)dx +
1
α
Za
exp(cid:18)−
1 − α
α
(x − a)(cid:19) u(x)dx
Therefore, from (1.12) we obtain inequality (2.2).
(cid:3)
Remark 2.2. Inequalities (1.1) and (1.5) are special cases of inequality (2.2).
• If p → 0 we have lim
p→0C2
α(1) =
coincide with the inequality (1.7);
2 exp(− 1−α
1−α
α (b−a))
. In this case, inequality (2.2)
• If p → 0 and α → 1 in (2.2), then we have classical Hermite-Hadamard
inequality (1.1).
2.2. Fractional analogues of Hermite-Hadamard-Fej´er inequality for hyper-
bolic p-convex functions.
Theorem 2.3. Assume that the function u : I → R is hyperbolic p-convex on I and
a, b ∈ I. Assume also that v : I → R is a positive, symmetric and integrable function
on [a, b], then we have
(2.3)
a+ [u(b)v(b)] + I α
b− [u(a)v(a)]
cosh−1 p(b − a)
C1
α(v),
2
u(cid:18)a + b
2 (cid:19)C1
α(v) ≤ I α
u(a) + u(b)
b
2
≤
2 (cid:1)(cid:3) (b−x)α−1+(x−a)α−1
cosh(cid:2)p(cid:0)x − a+b
Γ(α)
α(v) :=
where C1
Proof. Let us suppose that all assumptions of Theorem are satisfied. Let us define
the function w of Theorem 1.2 by
v(x)dx.
Ra
v(x)
w(x) =
Γ(α)(cid:2)(b − x)α−1 + (x − a)α−1(cid:3) , a < x < b.
Clearly, w is a positive, symmetric and integrable function on [a, b]. Moreover, for all
x ∈ (a, b), we get w(a + b − x) = w(x). Moreover, we have
b
1
Γ(α)
Za
u(x)v(x)(cid:2)(b − x)α−1 + (x − a)α−1(cid:3) dx = I α
Therefore, from (1.12) we obtain inequality (2.3).
a+ [u(b)v(b)] + I α
b− [u(a)v(a)] .
(cid:3)
Remark 2.3. Inequalities (1.2) and (1.6) are special cases of inequality (2.3).
• If p → 0 we have lim
p→0C1
coincide with the inequality (1.6);
α(v) = I α
a+v(b) + I α
b−v(a). In this case, inequality (2.3)
SOME HERMITE-HADAMARD TYPE INEQUALITIES
9
• If p → 0 and α → 1 in (2.3), then we have classical Hermite-Hadamard-Fej´er
inequality (1.2).
The following theorems are proved similarly.
Theorem 2.4. Assume that the function u : I → R is hyperbolic p-convex on I and
a, b ∈ I. Assume also that v : I → R is a positive, symmetric and integrable function
on [a, b], then we have
(2.4)
u(cid:18)a + b
2 (cid:19)C2
α(v) ≤ I α
a+ [u(b)v(b)] + I α
u(a) + u(b)
b− [u(a)v(a)]
cosh−1 p(b − a)
2
C2
α(v),
α (x−a))
α(v) :=
where C2
Remark 2.4. Inequalities (1.2) and (1.6) are special cases of inequality (2.4).
v(x)dx.
α
α (b−x))+exp(− 1−α
b
2
≤
cosh(cid:2)p(cid:0)x − a+b
p→0C2
2 (cid:1)(cid:3) exp(− 1−α
α(v) = I α
coincide with the inequality (1.8);
• If p → 0 we have lim
Ra
a+v(b) + I α
b−v(a). In this case, inequality (2.4)
• If p → 0 and α → 1 in (2.4), then we have classical Hermite-Hadamard-Fej´er
inequality (1.2).
Theorem 2.5. Assume that the function u : I → R is hyperbolic p-convex on I and
a, b ∈ I. Assume also that v : I → R is a positive, symmetric and integrable function
on [a, b], then
a+ [u(b)v(b)] + I α
I α
b− [u(a)v(a)] ≤
u(a) + u(b)
2
+
u(a) − u(b)
2
(2.5)
where
2
cosh−1(cid:18)p(b − a)
sinh−1(cid:18) p(b − a)
(cid:19)C1
(cid:19)S 1
2
α(v)
α(v),
b
S 1
α(v) :=
sinh(cid:20)p(cid:18)x −
Remark 2.5. If α → 1, we have
Za
a + b
2 (cid:19)(cid:21) (b − x)α−1 + (x − a)α−1
Γ(α)
v(x)dx.
and
α→1C1
lim
α(v) =
α→1S 1
lim
α(v) =
b
Za
b
Za
cosh(cid:20)p(cid:18)x −
a + b
2 (cid:19)(cid:21) v(x)dx
sinh(cid:20)p(cid:18)x −
a + b
2 (cid:19)(cid:21) v(x)dx.
In this case, inequality (2.5) coincide with the inequality (1.13).
10
DRAGOMIR AND TOREBEK
Theorem 2.6. Assume that the function u : I → R is hyperbolic p-convex on I and
a, b ∈ I. Assume also that v : I → R is a positive, symmetric and integrable function
on [a, b], then
(2.6)
I α
a+ [u(b)v(b)] + I α
b− [u(a)v(a)] ≤
u(a) + u(b)
2
+
u(a) − u(b)
2
2
cosh−1(cid:18) p(b − a)
sinh−1(cid:18) p(b − a)
(cid:19)C2
(cid:19)S 2
2
α(v)
α(v),
α (b−x))+exp(− 1−α
α (x−a))
α
v(x)dx.
b
α(v) :=
sinh(cid:2)p(cid:0)x − a+b
where S 2
Remark 2.6. If α → 1, we have
Ra
2 (cid:1)(cid:3) exp(− 1−α
Za
cosh(cid:20)p(cid:18)x −
b
α→1C2
lim
α(v) =
a + b
2 (cid:19)(cid:21) v(x)dx
and
α→1S 2
lim
α(v) =
b
Za
sinh(cid:20)p(cid:18)x −
a + b
2 (cid:19)(cid:21) v(x)dx.
In this case, inequality (2.6) coincide with the inequality (1.13).
Acknowledgements
The research of Torebek is financially supported by a grant No.AP05131756 from
the Ministry of Science and Education of the Republic of Kazakhstan. No new data
was collected or generated during the course of research
References
[A16]
[C16]
[CK17]
[DA98]
[DP00]
[D18a]
[D18b]
[F06]
[H93]
[H83]
M. S. S. Ali, On certain properties for two classes of generalized convex functions,
Abstract and Applied Analysis, (2016), Article ID 4652038, 1 -- 7.
F. Chen, Extensions of the Hermite-Hadamard inequality for convex functions via frac-
tional integrals, J. Math. Inequal. 10, 1(2016), 75-81.
H. Chen, U.N. Katugampola, Hermite-Hadamard and Hermite-Hadamard-Fejer type
inequalities for generalized fractional integrals, J. Math. Anal. Appl. 446, 2(2017), 1274-
1291.
S.S. Dragomir, R.P. Agarwal, Two inequalities for differentiable mappings and applica-
tions to special means of real numbers and to trapezoidal formula, Appl. Math. lett. 11,
5 (1998), 91-95.
S.S. Dragomir, C.E.M. Pearce, Selected topics on Hermite-Hadamard inequalities and
applications, RGMIA Monographs, Victoria University, 2000.
S.S. Dragomir, Some inequalities of Hermite-Hadamard type for hyperbolic p-convex
functions. RGMIA Res. Rep. Coll. 21, Art. 13, (2018), 1 -- 11.
S.S. Dragomir, Some inequalities of Fej´er type for hyperbolic p-convex functions.
RGMIA Res. Rep. Coll. 21, Art. 14, (2018), 1 -- 10.
L. Fej´er, Uberdie Fourierreihen, II, Math., Naturwise. Anz Ungar. Akad.Wiss, 24,
(1906), 369-390 (in Hungarian).
J. Hadamard, Etude sur les proprietes des fonctions entieres et en particulier d'une
fonction considree par Riemann, J. Math. Pures et Appl. 58, (1893), 171-215.
Ch. Hermite, Sur deux limites d'une integrale definie, Mathesis 3, (1883), 82.
SOME HERMITE-HADAMARD TYPE INEQUALITIES
11
[HYT14]
[I16]
[JS16]
[KST06]
[KS18]
[KT16]
[P03]
[PPT92]
S.R. Hwang, S.Y. Yeh, K.L. Tseng, Refinements and similar extensions of Hermite-
Hadamard inequality for fractional integrals and their applications, Appl. Math. Com-
putat. 249, (2014), 103-113.
I. Iscan, On generalization of different type inequalities for harmonically quasi-convex
functions via fractional integrals, Appl. Math. Computat. 275, (2016), 287-298.
M. Jleli, B. Samet, On Hermite-Hadamard type inequalities via fractional integrals of a
function with respect to another function, J. Nonlinear Sci. Appl. 9, 3(2016), 1252-1260.
A. A. Kilbas, H. M. Srivastava and J. J. Trujillo. Theory and Applications of Fractional
Differential Equations. North-Holland Mathematics Studies. (2006).
M. Kirane, B. Samet, Discussion of some inequalities via fractional integrals, Journal of
Inequalities and Applications. 2018, (2018), Art. 19, 1 -- 10.
M. Kirane, B.T. Torebek, Hermite-Hadamard, Hermite-Hadamard-Fejer, Dragomir-
Agarwal and Pachpatte type inequalities for convex functions via fractional integrals,
(2016). arXiv:1701.00092v1
B.G. Pachpatte, On some inequalities for convex functions, RGMIA Res. Rep. Coll. 6
(E), 2003.
J.E. Pecari´c, F. Proschan, Y.L. Tong, Convex Functions, Partial Orderings and Statis-
tical Applications, Academic Press, Boston, 1992.
[SSYB13] M.Z. Sarikaya, E. Set, H. Yaldiz, N. Ba¸sak, Hermite-Hadamard's inequalities for frac-
tional integrals and related fractional inequalities, Math. Comput. Modelling 57, 9-10
(2013), 2403-2407.
[WLFZ12] J. Wang, X. Li, M. Feckan, Y. Zhou, Hermite-Hadamard-type inequalities for Riemann-
Liouville fractional integrals via two kinds of convexity, Appl. Anal. 92, 11(2012), 2241-
2253.
Y. Zhang, J. Wang, On some new Hermite-Hadamard inequalities involving Riemann-
Liouville fractional integrals, J. Inequal. Appl. 2013, 220(2013), 27 pp.
[ZW13]
Silvestru Sever Dragomir
College of Engineering and Science, Victoria University,
PO Box 14428, Melbourne City, MC 8001, Australia.
E-mail address: [email protected]
Berikbol T. Torebek
Institute of Mathematics and Mathematical Modeling
125 Pushkin str., 050010 Almaty, Kazakhstan
Al-Farabi Kazakh National University.
71 Al-Farabi ave., 050040 Almaty, Kazakhstan
E-mail address: [email protected]
|
1806.10849 | 2 | 1806 | 2019-02-03T07:06:04 | Linear functions and duality on the infinite polytorus | [
"math.FA",
"math.CV"
] | We consider the following question: Are there exponents $2<p<q$ such that the Riesz projection is bounded from $L^q$ to $L^p$ on the infinite polytorus? We are unable to answer the question, but our counter-example improves a result of Marzo and Seip by demonstrating that the Riesz projection is unbounded from $L^\infty$ to $L^p$ if $p\geq 3.31138$. A similar result can be extracted for any $q>2$. Our approach is based on duality arguments and a detailed study of linear functions. Some related results are also presented. | math.FA | math | LINEAR FUNCTIONS AND DUALITY
ON THE INFINITE POLYTORUS
OLE FREDRIK BREVIG
Abstract. We consider the following question: Are there exponents 2 < p <
q such that the Riesz projection is bounded from Lq to Lp on the infinite
polytorus? We are unable to answer the question, but our counter-example
improves a result of Marzo and Seip by demonstrating that the Riesz projection
is unbounded from L∞ to Lp if p ≥ 3.31138. A similar result can be extracted
for any q > 2. Our approach is based on duality arguments and a detailed
study of linear functions. Some related results are also presented.
.
A
F
h
t
a
m
[
2
v
9
4
8
0
1
.
6
0
8
1
:
v
i
X
r
a
1. Introduction
Let T∞ = T × T × ··· denote the countably infinite cartesian product of the
torus T = {z ∈ C : z = 1}. We equip the T∞ with its Haar measure µ∞, which is
equal to the infinite product of the normalized Lebesgue arc measure on T in each
variable. Let 1 ≤ p ≤ ∞. Every f in Lp(T∞) has a Fourier series expansion
f (z) = Xα∈Z∞
0
cαzα
where the Fourier coefficients are defined in the standard way and α ∈ Z∞
0 means
that the multi-index α contains only a finite number of non-zero components. The
Riesz projection on T∞ is defined by
(1)
P f (z) = Xα∈N∞
0
cαzα.
The initial motivation for the present paper is the following.
Question. What is the largest p = p∞ such that the Riesz projection (1) is bounded
from L∞(T∞) to Lp(T∞)?
The Riesz projection is certainly a contraction on the Hilbert space L2(T∞) and
since kfkL2(T∞) ≤ kfkL∞(T∞), we get that p∞ ≥ 2. This question has previously
been investigated by Marzo and Seip [8] who demonstrated that p∞ ≤ 3.67632. We
will obtain the following improvement.
Theorem 1. p∞ ≤ p = 3.31138 . . ., where p denotes the unique positive solution
of the equation
Γ(cid:16)1 +
1
p
p
2(cid:17)
=
2
√π
.
Date: February 5, 2019.
2010 Mathematics Subject Classification. Primary 42B05. Secondary 42B30, 46E30.
1
2
OLE FREDRIK BREVIG
For 2 ≤ p ≤ q ≤ ∞, let kPkq,p denote the norm of the Riesz projection from
Lq(T∞) to Lp(T∞). In the case that the Riesz projection is unbounded, we use the
convention kPkq,p = ∞. As explained in [8], for each fixed 2 ≤ q ≤ ∞ there is a
number 2 ≤ pq ≤ q, called the critical exponent, with the property that
(2)
kPkp,q =(1
if p ≤ pq,
∞ if p > pq.
The dichotomy (2) is a direct consequence of the fact that we are on the infinite
polytorus. Let f be a function in the unit ball of Lq(T∞) such that kP fkLp(T∞) > 1.
Consider the function
f2(z) = f (z1, z3, z5, . . .) · f (z2, z4, z6, . . .)
which is also in the unit ball of Lq(T∞). The Riesz projection (1) acts independently
on the variables, so we find that
P f2(z) = P f (z1, z3, z5, . . .) · P f (z2, z4, z6, . . .)
which implies that kP f2kLp(T∞) = kP fk2
Lp(T∞) > kP fkLp(T∞). This procedure can
be repeated and so we obtain (2). The example from [8] producing p∞ ≤ 3.67632
is a function of only two variables.
The present paper is inspired by [3], where linear functions are used as building
blocks in an similar way to what was just described to construct a counter-example
related to Nehari's theorem for Hankel forms on T∞. The example from [3] improves
on an earlier example from [9] by replacing a function of two variables by a linear
function in an infinite number of variables.
Our approach differs from that of [8] (and [2]) in that we do not attempt to
directly construct a counter-example, but instead use duality arguments to infer
its existence. This approach leads us to consider the Hardy spaces H p(T∞), which
are the subspaces of Lp(T∞) consisting of elements such that P f = f . A standard
argument involving the Hahn -- Banach theorem (see e.g. [5, Sec. 7.2]) yields that
(3)
inf
P ψ=ϕ kψkLq(T∞) = kϕk(Hr (T∞))∗ = sup
f ∈Hr (T∞)
hf, ϕiL2(T∞)
kfkHr(T∞)
for 1 ≤ r < ∞ and q−1 + r−1 = 1. We will choose ϕ and try to find the optimal f
in H r(T∞) attaining the supremum. This will ensure the existence of ψ in Lq(T∞)
attaining the infimum, which be our counter-example through (2).
We shall see in Section 3 that if we know the optimal f in the supremum on the
right hand side of (3), we can use Hölder's inequality to construct the element ψ
in Lq(T∞) of minimal norm such that P ψ = ϕ, thereby attaining the infimum on
the left hand side of (3).
As in [3] we will primarily be working with linear functions, which are of the
form
(4)
f (z) =
cjzj.
∞
Xj=1
H 2(T∞) =Pj≥1 cj2 and we easily check that kfkH∞(T∞) =Pj≥1 cj.
Clearly, kfk2
For 1 ≤ p < ∞, optimal norm estimates are given by Khintchine's inequality.
LINEAR FUNCTIONS AND DUALITY ON THE INFINITE POLYTORUS
3
Define
ap = min(cid:18)1, Γ(cid:16)1 +
p
2(cid:17)
1
p(cid:19)
and
bp = max(cid:18)1, Γ(cid:16)1 +
p
2(cid:17)
1
p(cid:19) .
If f is a linear function (4) and 1 ≤ p < ∞, then we restate a result from [7] as
(5)
apkfkH 2(T∞) ≤ kfkHp(T∞) ≤ bpkfkH 2(T∞)
and the constants in (5) are optimal. We shall obtain the following companion
inequality for dual norms, which might be of independent interest.
Theorem 2. Let 1 ≤ p < ∞. If f is a linear function (4), then
(6)
p kfkH 2(T∞).
b−1
p kfkH 2(T∞) ≤ kfk(Hp(T∞))∗ ≤ a−1
The constants are optimal.
Remark. In the case p = ∞, it is easy to deduce by similar considerations (Lemma 4)
that kfk(H∞(T∞))∗ = supj≥1 cj if f is a linear function (4).
Optimality of the constants containing the Gamma function in (5) and (6) both
arise from the function
z1 + z2 + ··· + zd
f (z) =
√d
as d → ∞ through the central limit theorem.
In view of (2) and (3), we can
therefore obtain the following general result. Note that Theorem 1 corresponds to
the particular case q = ∞, since Γ(3/2) = √π/2.
Theorem 3. Let 2 ≤ p ≤ q ≤ ∞ and set q−1 + r−1 = 1. If
then the Riesz projection is unbounded from Lq(T∞) to Lp(T∞).
Γ(cid:16)1 +
1
p
p
2(cid:17)
Γ(cid:16)1 +
r
2(cid:17)
1
r
> 1,
Remark. Theorem 3 is an improvement on the same statement with requirement
p/2·r/2 > 1, which can be deduced from a one-variable example found in [2, Sec. 4].
Here is an alternative example to that of [2] obtained by our approach using the
Hahn -- Banach theorem. For w ∈ D, the functional of point evaluation f 7→ f (w)
has norm (1 − w2)−1/r on H r(T) and the analytic symbol is ϕw(z) = (1 − wz)−1.
Hence, if w = ε > 0 then kϕεk(Hr (T))∗ = 1 + r−1ε2 + O(ε4) as ε → 0. Furthermore,
kϕεkHp(T) =(cid:13)(cid:13)(1 − εz)−p/2(cid:13)(cid:13)
2/p
H 2(T) = 1 +
ε2 + O(ε4),
p
4
so we obtain the desired counter-example as soon as r−1 > p/4 in view of (2). The
optimal ψw in Lq(T) for this functional can be found in [4, Thm. 6.1], and we note
that it is similar (but not equal to) the counter-example constructed in [2].
The present paper is organised into two additional sections.
In Section 2 we
prove Theorem 2 and Theorem 3. Section 3 is devoted to constructing the element
ψ in Lq(T∞) for 1 < q ≤ ∞ of minimal norm such that P ψ(z) = z1 + z2 + ··· + zd,
thereby realising the infimum (3) in this special case, which is of particular interest
due to the crucial role it plays in the proof of Theorem 2 and Theorem 3.
4
OLE FREDRIK BREVIG
2. Linear functions on T∞
In preparation for the proof of Theorem 2 and Theorem 3, let us recall some basic
facts about linear functions and projections on T∞. The projection Ad obtained
by formally setting zj = 0 for j > d has the representation
Adf (z1, z2, . . .) =ZT∞
f (z1, z2, . . . , zd, zd+1, zd+2, . . .) dµ∞(zd+1, zd+2, . . .).
Since Adf is a function the first d variables, we take Lp norm with respect to these
variables and use the triangle inequality to obtain
(7)
Let k ∈ Z. We say that f is k-homogeneous if
kAdfkLp(T∞) ≤ kfkLp(T∞).
f (eiθz1, eiθz2, eiθz3, . . .) = ekiθf (z1, z2, z3, . . .).
Clearly every f in Lp(T∞) can be decomposed in k-homogeneous parts, say
(8)
f (z) =Xk∈Z
fk(z),
where fk is k-homogeneous. The following simple lemma is well-known, but we
include a short proof for the readers convenience.
Lemma 4. Let 1 ≤ p ≤ ∞ and suppose that f in Lp(T∞) is decomposed as in (8).
Then kfkkLp(T∞) ≤ kfkLp(T∞) for every k ∈ Z.
Proof. By the decomposition (8), we find that
By the triangle inequality and interchanging the order of integration, we obtain
f (eiθz1, eiθz2, eiθz3, . . .) e−kiθ dθ
2π
.
−π
fk(z) =Z π
−πZT∞(cid:12)(cid:12)f (eiθz1, eiθz2, eiθz3, . . .)(cid:12)(cid:12)
p
dµ∞(z)
dθ
2π
= kfkp
Lp(T∞),
kfkkp
Lp(T∞) ≤Z π
since for each θ the rotation zj 7→ eiθzj does not change the Lp(T∞) norm of f . (cid:3)
Let Lin(T∞) denote the space of linear functions (4). Lemma 4 states that
the projection from H p(T∞) to Lin(T∞) ∩ H p(T∞) is contractive. This fact is
crucial to the proof of Theorem 2 and Theorem 3 since it allows us to compute the
(H p(T∞))∗ norm of a linear function ϕ by testing only against functions f from
Lin(T∞) ∩ H p(T∞).
In view of Khintchine's inequality (5), the space Lin(T∞) ∩ H p(T∞) consists of
linear functions (4) with square summable coefficients for each 1 ≤ p < ∞, although
the norms are generally different.
Armed with these preliminaries, we will now obtain the key new ingredient
needed in the proofs of Theorem 2 and Theorem 3.
Lemma 5. Let 1 ≤ p < ∞ and set ϕd(z) = (z1 + ··· + zd)/√d. Then
kϕdk(Hp(T∞))∗ = kϕdk−1
Hp(T∞).
Proof. For the lower bound, we simply note that since ϕd is in H p(T∞) we obtain
(9)
kϕdk(Hp(T∞))∗ = sup
f ∈Hp(T∞)
hf, ϕdiH 2(T∞)
kfkHp(T∞) ≥
H 2(T∞)
kϕdk2
kϕdkHp(T∞)
= kϕdk−1
Hp(T∞).
1
d
d
Xk=1
fk(z) =
c1 + ··· + cd
d
zj = λϕd(z).
d
Xj=1
The triangle inequality therefore allows us to conclude that
(12)
λkϕdkHp(T∞) ≤
1
d
kfkkHp(T∞) = kfkHp(T∞).
d
Xk=1
LINEAR FUNCTIONS AND DUALITY ON THE INFINITE POLYTORUS
5
For the upper bound, we first use (7) and Lemma 4 to the effect that
(10)
kϕdk(Hp(T∞))∗ = sup
f ∈Hp(T∞)
hf, ϕdiH 2(T∞)
kfkHp(T∞)
= sup
f ∈Lin(Td)
hf, ϕdiH 2(T∞)
kfkHp(T∞)
.
Any non-trivial element f in Lin(Td) is of the form
d
f (z) =
Xj=1
with at least one non-zero coefficient. Define
λ = hf, ϕdiH 2(T∞) =
(11)
cjzj
c1 + ··· + cd
√d
.
After rotating each of the variables if necessary, we may assume that cj ≥ 0 for
1 ≤ j ≤ d so that λ > 0 whenever f is a non-trivial element in Lin(Td).
For 1 ≤ k ≤ d, let fk denote the polynomial obtained by replacing the coefficient
sequence (c1, . . . , cd) of f with the shifted sequence
(ck, ck+1, . . . , cd, c1, . . . , ck−1).
By symmetry, we find that kfkkHp(T∞) = kfkHp(T∞). Note also that
Using (10) with (11) and (12), we obtain the upper bound
kϕdk(Hp(T∞))∗ = sup
f ∈Hp(T∞)
hf, ϕdiH 2(T∞)
kfkHp(T∞) ≤
λ
λkϕdkHp(T∞)
= kϕdk−1
Hp(T∞)
which, when combined with the lower bound (9), completes the proof.
(cid:3)
Another viewpoint is to consider (zj)j≥1 a sequence of independently distributed
random variables on the torus and f (z) =Pj≥1 cjzj as a weighted random walk in
the plane. The norms kfkHp(T∞) can now be interpreted as moments of this random
walk. A simple computation (see Section 3) gives that kz1 + z2kH 1(T∞) = 4/π and
it is demonstrated in [1] that
kz1 + z2 + z3kH 1(T∞) =
3
16
21/3
π4 Γ6(cid:18) 1
3(cid:19) +
27
4
22/3
π4 Γ6(cid:18) 2
3(cid:19) = 1.57459 . . .
In general it is difficult to compute kfkHp(T∞) even for simple linear polynomials
f (when p is not an even integer). However, the central limit theorem gives that
p
lim
(13)
√d
z1 + z2 + ··· + zd
2(cid:17) ,
= Γ(cid:16)1 +
since (z1 + z2 + ··· + zd)/√d has a limiting complex normal distribution.
=ZC Zpe−Z2 dZ
d→∞(cid:13)(cid:13)(cid:13)(cid:13)
We are now ready to prove Theorem 2. To conform with the notations of the
present section and to make the proof clearer, we consider now ϕ in (H p(T∞))∗
and f in H p(T∞), so ϕ plays the role of f in the statement of the theorem.
(cid:13)(cid:13)(cid:13)(cid:13)
Hp(T∞)
π
p
6
OLE FREDRIK BREVIG
Proof of Theorem 2. Let ϕ be a linear function in (H p(T∞))∗. By Lemma 4, the
Cauchy -- Schwarz inequality and Khintchine's inequality (5), we find that
kϕk(Hp(T∞))∗ = sup
f ∈Lin(T∞)
hf, ϕiH 2(T∞)
kfkHp(T∞)
≤ sup
f ∈Lin(T∞)
kfkH 2(T∞)kϕkH 2(T∞)
kfkHp(T∞)
≤ kϕkH 2(T∞)
ap
.
Conversely, Khintchine's inequality (5) also gives that
kϕk2
kϕkHp(T∞) ≥ kϕkH 2(T∞)
hf, ϕiH 2(T∞)
kfkHp(T∞) ≥
kϕk(Hp(T∞))∗ = sup
f ∈Hp(T∞)
H 2(T∞)
bp
,
since ϕ is in H p(T∞). To prove optimality of the constants, we appeal to Lemma 5
(cid:3)
and consider ϕd(z) = (z1 + ··· + zd)/√d for d = 1 and as d → ∞.
Theorem 3 also follows easily from Lemma 5 and (13).
Proof of Theorem 3. Let 2 ≤ p ≤ q ≤ ∞ and set q−1 + r−1 = 1. Suppose that
(14)
> 1.
p
r
1
p
1
r
Γ(cid:16)1 +
2(cid:17)
Γ(cid:16)1 +
2(cid:17)
We want to to prove that the Riesz projection is unbounded from Lq(T∞) to
Lp(T∞). In view of (2), it is sufficient to find ψ in Lq(T∞) such that
kP ψkLp(T∞)
kψkLq(T∞)
> 1.
We pick ψd in Lq(T∞) of minimal norm such that P ψd = ϕd, where ϕd denotes the
function from Lemma 5. By (3) and Lemma 5, we obtain
kP ψdkLp(T∞)
kψdkLq(T∞)
= kϕdkLp(T∞)kϕdkLr(T∞).
By (13) and our assumption (14), the right hand side is strictly larger than 1 for
some sufficiently large d.
(cid:3)
3. Minimal Lq(T∞) norm
We will now solve the following problem: For 1 < q ≤ ∞, find the element ψ in
Lq(T∞) of minimal norm such that
P ψ(z) = z1 + z2 + ··· + zd = ϕ(z).
The strict convexity of Lq(T∞) when 1 < q < ∞ means that the minimizer is
unique. Uniqueness of the minimizer holds also for q = ∞, but in this case it is a
consequence of the continuity of ϕ on the polytorus (see e.g. [5, Sec. 8.2]).
In view of (3) and (the proof of) Lemma 5, we know that ψ satisfies
(15)
kψkLq(T∞) = hϕ, ψiL2(T∞)
kϕkLp(T∞)
=
d
kϕkLp(T∞)
with p−1 + q−1 = 1. On the left hand side of (15) we have attained equality in
Hölder's inequality, which implies that ψ = Cϕp−1 almost everywhere. Inserting
LINEAR FUNCTIONS AND DUALITY ON THE INFINITE POLYTORUS
7
this into the norm expression kψkLq(T∞) in (15) and using that (p − 1)q = p, we
find that C = dkϕk−p
Lp(T∞). From Hölder's inequality and (15) we also see that
hϕ,ψiL2(T∞) ≤ kϕkLp(T∞)kψkLq(T∞) = hϕ, ψiL2(T∞),
which is only possible if ϕψ ≥ 0 almost everywhere. Combining these observations
yields that
ψ(z) =
d
kϕkp
Lp(T∞)ϕ(z)p−2ϕ(z)
is the element in Lq(T∞) of minimal norm such that P ψ(z) = z1+z2+···+zd = ϕ(z)
for 1 < p ≤ ∞ and p−1 + q−1 = 1. Note that ψ is 1-homogeneous, which we knew
in advance by Lemma 4. We can also directly verify that
ZT∞
ψ(z) zj dm∞(z) =ZT∞
since ψ inherits the symmetry of ϕ.
ψ(z)
z1 + z2 + ··· + zd
d
dµ∞(z) = 1,
When d = 2, we can actually compute the Fourier series explicitly. We begin by
using the trick z1 + z2 = z2(1 + z1z2) to write ψ(z) = z2Ψ(z1z2), where
Ψ(z) =
2
k1 + zkp
Lp(T)1 + zp−2(1 + z).
Then we get that
k1 + zkp
2
Lp(T)
=
1
2Z π
−π 1 + eiθp dθ
2π
= 2p−1Z π
−π
cosp(cid:18) θ
2(cid:19) dθ
2π
=
2p
π Z π/2
0
cosp(ϑ) dϑ.
Similarly, we compute:
Z π
−π 1 + eiθp−2(1 + eiθ) e−ikθ dθ
2π
−π
= 2p−1Z π
= 2p−1Z π/2
π Z π/2
2p
=
0
−π/2
2(cid:19) e−i(k−1/2)θ dθ
cosp−1(cid:18) θ
2π
cosp−1(ϑ) e−i(2k−1)ϑ dϑ
π
cosp−1(ϑ) cos((1 − 2k)ϑ) dϑ
The latter integral, which contains the former as the special case k = 0, 1 is known
(see e.g. [6, p. 399]) and we obtain that
Z π
−π 1 + eiθp−2(1 + eiθ) e−ikθ dθ
2π
=
1
p Beta(cid:16) p+1−2k+1
2
, p−1+2k+1
2
(cid:17)
for Beta(x, y) = Γ(x)Γ(y)/Γ(x + y). Combining everything, we find that
ψ(eiθ1 , eiθ2) =Xk∈Z
Γ(1 + p/2)Γ(p/2)
Γ(1 + p/2 − k)Γ(p/2 + k)
eikθ1 ei(1−k)θ2 .
Acknowledgements
The author would like to extend his gratitude to A. Bondarenko, H. Hedenmalm,
E. Saksman and K. Seip for an interesting discussion which culminated in the
material presented in Section 3 and to the referee for a helpful suggestion.
8
OLE FREDRIK BREVIG
References
1. J. M. Borwein, D. Nuyens, A. Straub, and J. Wan, Some arithmetic properties of short random
walk integrals, Ramanujan J. 26 (2011), no. 1, 109 -- 132.
2. O. F. Brevig, J. Ortega-Cerdà, K. Seip, and J. Zhao, Contractive inequalities for Hardy spaces,
Funct. Approx. Comment. Math. 59 (2018), no. 1, 41 -- 56.
3. O. F. Brevig and K.-M. Perfekt, Failure of Nehari's theorem for multiplicative Hankel forms
in Schatten classes, Studia Math. 228 (2015), no. 2, 101 -- 108.
4. B. J. Cole and T. W. Gamelin, Representing measures and Hardy spaces for the infinite polydisk
algebra, Proc. London Math. Soc. (3) 53 (1986), no. 1, 112 -- 142.
5. P. L. Duren, Theory of H p spaces, Pure and Applied Mathematics, Vol. 38, Academic Press,
New York-London, 1970.
6. I. S. Gradshteyn and I. M. Ryzhik, Table of integrals, series, and products, eighth ed., Else-
vier/Academic Press, Amsterdam, 2015.
7. H. König and S. Kwapień, Best Khintchine type inequalities for sums of independent, rotation-
ally invariant random vectors, Positivity 5 (2001), no. 2, 115 -- 152.
8. J. Marzo and K. Seip, L∞ to Lp constants for Riesz projections, Bull. Sci. Math. 135 (2011),
no. 3, 324 -- 331.
9. J. Ortega-Cerdà and K. Seip, A lower bound in Nehari's theorem on the polydisc, J. Anal.
Math. 118 (2012), no. 1, 339 -- 342.
Department of Mathematical Sciences, Norwegian University of Science and Tech-
nology (NTNU), NO-7491 Trondheim, Norway
E-mail address: [email protected]
|
1502.01561 | 1 | 1502 | 2015-02-05T14:14:55 | An implicit function theorem for non-smooth maps between Fr\'echet spaces | [
"math.FA"
] | We prove an inverse function theorem of Nash-Moser type for maps between Fr\'echet spaces satisfying tame estimates. In contrast to earlier proofs, we do not use the Newton method, that is, we do not use quadratic convergence to overcome the lack of derivatives. In fact, our theorem holds when the map to be inverted is not C^2 | math.FA | math |
An implicit function theorem for non-smooth
maps between Fr´echet spaces.
CEREMADE, Universit´e Paris-Dauphine, 75016 Paris, France
Ivar Ekeland, Eric S´er´e
November 2013
1 Introduction
In this paper, we prove a "hard" inverse function theorem, that is, an inverse
function theorem for maps F which lose derivatives: F (u) is less regular than u.
Such theorems have a long history, starting with Kolmogorov in the Soviet Union
([2], [3], [4]) and Nash in the United States ([15]), and it would be impossible,
in such a short paper, to give a full account of the developments which have
occured since. Important contributions have been made since by Hormander,
Zehnder, Mather, Sergeraert, Tougeron, Hamilton, Hermann, Craig, Dacorogna,
Bourgain, Berti and Bolle, and lately by Villani and Mouhot. However, all the
results which we are aware of require the function F to be inverted to be at
least C2; in the Kolmogorov-Arnol'd-Moser tradition, for instance, one uses the
fast convergence of Newton's method to overcome the loss of derivatives.
In
contrast, we make no smoothness assumption on F , only that is continuous and
Gateaux-differentiable.
We will overcome the loss of derivatives by using a new version of the "soft"
inverse function theorem (between Banach spaces), the proof of which is given
in [12], namely:
Theorem 1 Let X and Y be Banach spaces, with respective norms kxk and
kyk′. Let f : X → Y be continuous and Gateaux-differentiable, with f (0) = 0.
Assume that the derivative Df (x) has a right-inverse [Df (x)]−1
r , uniformly
bounded in a neighbourhood of 0:
Df (x) [Df (x)]−1
r h = h
sup(cid:8)(cid:13)(cid:13)[Df (x)]−1
r (cid:13)(cid:13) kxk ≤ R(cid:9) < m
Then, for every y ∈ Y with kyk′ ≤ Rm−1
k¯xk ≤ R, such that f (¯x) = ¯y and k¯xk ≤ mk¯yk′.
there is some ¯x ∈ X with
As we just said, we will make no attempt to review the literature on hard
inverse function theorems; see the survey by Hamilton [10] for an account up to
1
1982. We have drawn inspiration from the version in [1], which itself is inspired
from Hormander's result [11]. We have also learned much from the work on the
nonlinear wave equation by Berti and Bolle, [5], [6], [7], [8] whom we thank for
extensive discussions.
In the section 2, we state our main result, Theorem 1, and we derive it from
an approximation procedure, which is described in Theorem 2. We also given
some variants of Theorem 1, for instance an implicit function theorem and we
describe a particular case when we can gain some regularity. Theorem 2 is
proved in section 3, and the theoretical part is thus complete. The next two
sections are devoted to applications. Section 5 revisits the classical isometric
imbedding problem, which was the purpose for Nash's original work. This is
somewhat academic, since it known now that it can be treated without resorting
to a hard inverse function theorem (see [9]), but it gives us the opportunity to
show on a simple example how our result improves, for instance, on those of
Moser [14].
2 The setting
2.1 The spaces
Let (Xs, k · ks)s≥0 be a scale of Banach spaces:
0 ≤ s1 ≤ s2 =⇒ (Xs2 ⊂ Xs1 and k · ks1 ≤ k · ks2 )
We shall assume that there exists a sequence of projectors ΠN : X0 → EN
where EN ⊂ Ts≥0 Xs is the range of ΠN , with Π0 = 0, EN ⊂ EN +1 and
SN≥1 EN is dense in each space Xs for the norm k · ks. We assume that for
any finite constant A there is a constant CA
numbers s, d satisfying s + d ≤ A:
1 > 0 such that, for all nonnegative
kΠN uks+d ≤ CA
1 N dkuks
(1)
(2)
Note that these properties imply some interpolation inequalities, for 0 ≤ t ≤
k(1 − ΠN )uks ≤ CA
1 N−dkuks+d
1 and 0 ≤ s1 , s2 ≤ A , and for a new constant C(2)
A (see e.g. [8]):
kxkts1+(1−t)s2 ≤ CA
2 kxkt
s1kxk1−t
s2
.
(3)
If all these properties are satisfied, we shall say that the scale (Xs) , s ≥ 0, is
regular, and we shall refer to the ΠN as smoothing operators. Let (Ys, k ·k′s)s≥0
be another regular scale of Banach spaces. We shall denote by Π′N : Y0 →
E′N ⊂ Ts≥0 Ys the smoothing operators. In the sequel, Bs(R) (resp. B′s(R))
will denote the open ball of center 0 and radius R in Xs (resp. Ys)
2
2.2 The map
Recall that a map F : X → Y , where X and Y are Banach spaces, is Gateaux-
differentiable at x if there is a linear map DF (x) : X → Y such that:
∀h ∈ X,
1
t
lim
t→0
[F (x + th) − F (x)] = DF (x) h
In the following, R > 0 and S > 0 are prescribed, with possibly S = ∞
Definition 2 We shall say that F : B0(R) → Y0 is roughly tame with loss of
regularity µ if:
(a) F is continuous and Gateaux-differentiable from B0(R) ∩ Xs to Ys for any
s ∈ [0, S).
(b) For any A ∈ [0, S) there is a finite constant KA such that, for all s < A
and x ∈ B0 (R):
∀h ∈ X, kDF (x)hk′s ≤ K A(khks + kxkskhk0)
(4)
(c) For x ∈ B0(R) ∩ EN , the linear maps LN (x) : EN → E′N defined by
LN = Π′N DF (x) EN have a right-inverse, denoted by [LN (x)]−1
r . There
is a constant µ > 0 and, for any A ∈ [0, S), a positive constant γA, such
that, for all s < A and x ∈ B0 (R) we have:
∀k ∈ E′N , k[LN (x)]−1
r kks ≤
1
γA N µ(kkk′s + kxkskkk′0)
(5)
In our assumptions, there is no regularity loss between x and F (x), or more
exactly the regularity loss, if there is one, has been absorbed by translating the
indexation of the spaces Ys. The number S represents the maximum regular-
ity available, and the constant µ may be interpreted as the loss of derivatives
incurred when solving the linearized equation LN (x)h = k. Note that we need
µ < S to start the process.
When trying to solve F (x) = y, it is thus natural to assume that y − F (0)
is small in Yµand look for x in X0. This was done in [12] by assuming that
DF (x) has a right inverse which satisfied estimates similar to (4) and (5), but
which were independent of the base point x. In the present work, since the tame
estimates depend on x with loss of regularity, we will have to assume that y is
small in a more regular space Yδ, with δ > µ.
2.3 The main result
We will need an assumption relating µ, δ and S with µ < δ and S > µ. Here it
is:
Condition 3 There is some κ such that:
3
1 < κ < 2 and min{κ2, κ + 1}µ < δ
κ2
κ − 1
µ < S
(6)
(7)
Inequality (7) imposes S > 4µ, since 4 is the minimum value of κ2
κ−1 , attained
when κ = 2. On the other hand, min{κ2, κ + 1} is an increasing function of κ,
which coincides with κ2 when κ ≤ (cid:0)1 + √5(cid:1) /2 and coincides with κ + 1 when
κ >(cid:0)1 + √5(cid:1) /2. So inequality (6) imposes δ > 1, attained when κ = 1.
Let us represent condition (3) geometrically. Define a real function ϕ on
(0, 3] by:
2 x(cid:16)1 −q1 − 4
x(cid:17) + 1 if
4 (cid:16)1 −q1 − 4
x(cid:17)
x2
if
2
4 < x ≤ 3+√5√5−1
x ≥ 3+√5√5−1
(8)
1
ϕ (x) =
µ(cid:17)
µ ≥ ϕ(cid:16) S
δ
Proposition 4 (µ, δ, S) ∈ R2
µ ≤ 3 and δ
Proof. Follows immediately from inverting formulas (6) and (7).
+ satisfies condition (3) if and only if δ
µ ≥ 3 or
We have represented the admissible region for (cid:16) S
region Ω above the curve.
µ , δ
µ(cid:17) on Figure 1: it is the
Figure 1: The function ϕ (x)
4
The parameter κ decreases from κ = 2 (corresponding to S/µ = 3) to κ = 1
,
(corresponding to S/µ → ∞) along the curve. Note the kink at S/µ = 3+√5√5−1
2(cid:0)3 + √5(cid:1) corresponding to κ = 1
2(cid:0)1 + √5(cid:1) (the golden ratio). In the
sequel, we will separate the case κ ≤ 1+√5
(to the right) from the case κ ≥ 1+√5
δ/µ = 1
2
2
(to the left):
1 < κ ≤ 1+√5
1+√5
2 ≤ κ < 2
2
δ
µ > κ2 ≥ 1
µ > κ + 1 ≥ 3+√5
2
δ
S
µ > κ2
κ−1 ≥ 3+√5√5−1
µ > κ2
κ−1 ≥ 4
S
Theorem 5 Assume F : B0 (R) ∩ Xs → Ys, 0 ≤ s < S, is roughly tame with
loss of regularity µ. Suppose F (0) = 0. Let δ > 0 and α > 0 be such that
δ
µ
α
µ
µ(cid:19)
> ϕ(cid:18) S
µ − ϕ(cid:18) S
< min(cid:26) δ
µ(cid:19) ,
S
µ − ϕ−1(cid:18) δ
µ(cid:19)(cid:27)
(9)
(10)
Then one can find ρ > 0 and C > 0 such that, for any y ∈ Yδ with kyk′δ ≤ ρ
, there some x ∈ Xα such that:
F (x) = y
kxk0 ≤ 1
kxkα ≤ C kykδ
It follows that the map F sends X0 into Y0, while F −1 sends Yδ into X0.
This parallels the situation with the linearized operator DF (0), which sends X0
into Y0 while DF (0)−1 sends Yµ into X0, with µ < δ. More precisely, we have:
• if F (¯x) = ¯y, then F (X0) contains some δ-neighbourhood of ¯y
• if F (¯x) = ¯y, then F −1 (¯y + Yδ) contains some α-neighbourhood of ¯x
S is the maximal regularity on x and δ is the minimal regularity on y that we
will need in the approximation procedure, bearing in mind that F (0) = 0 (see
Corollary 6 below for the case when F (0) = ¯y 6= 0 has fininte regularity). Note
that we may have δ > S: this simply means that the right-hand side y is more
regular than the sequence of approximate solutions xn that we will construct.
α > 0 is the regularity of the solution x. Note the significance of 9) and (10)
µ > ϕ(cid:16) S
take together. The inequality δ
more regular than needed (or, alternatively, that the full range of S has not been
used), and this "excess regularity", measured by the difference δ
µ(cid:17) tells us that the right-hand side y is
µ(cid:17) (or,
µ(cid:17) ) can be diverted to x. For instance, if S/µ → ∞,
alternatively, S
the total loss of regularity δ − α between y and x satisfies
µ − ϕ−1(cid:16) δ
µ − ϕ(cid:16) S
µ(cid:19)
δ − α > µϕ(cid:18) S
5
and can be made as close to µ as one wishes: we can start the iterative procedure
xn from a very regular initial point. However, we lose control on ρ (which goes
to zero) and C (which goes to infinity). On the other hand, when δ/µ > 3,
that is, when we are not worried about the loss of regularity, then S can be any
number larger than µ, that is, we need very little regularity to start with.
We now go from the case F (0) = 0 to the case F (¯x) = ¯y. There is some sub-
telty there because 0 belong to all the Xs, while ¯y does not, and puts adittional
limits to the regularity. We shall say that F is roughly tame at ¯x if F (x − ¯x)
is roughly tame at 0.
Corollary 6 Suppose ¯x ∈ XS1, ¯y ∈ YS2 and F (¯x) = ¯y. Assume F (x) sends
Xs ∩ B0 (¯x, R) into Ys for every s ≥ 0, and is roughly tame at ¯x with loss of
regularity µ. Set S = min{S1, S2}. Let δ and α satisfy (9) and (10). Then
one can find ρ > 0 and C > 0 such that, for any y with k¯y − yk′δ ≤ ρ , there
is a solution x of the equation F (x) = y, with kx − ¯xk0 ≤ 1 and kx − ¯xkα ≤
Cky − ¯yk′δ .
Proof. Consider the map Φ (x) := F (x + ¯x) − ¯y.
It is roughly tame, with
F (0) = 0, and we can apply the preceding Theorem with S = min {S1, S2}.
The result follows
We now deduce an implicit function theorem. Let V be a Banach space and
let F : B (R, X0 × V ) ∩ (Xs × V ) → Ys, 0 ≤ s < S satisfy the following:
Definition 7 (a') F is continuous and Gateaux-differentiable for any s ∈ [0, S).
We write:
DF (x, v) = (DxF (x, v) , DvF (x, v))
(b') For any A ∈ [0, S) there is a finite constant KA such that, for all s < A
and (x, v) ∈ B (R, X0 × V ):
∀h ∈ X, kDF (x, v)hk′s ≤ K A(khks + kxkskhk0)
(c') For x ∈ (x, v) ∈ B (R, X0 × V ) ∩ (EN × V ), the linear maps LN (x, v) :
EN → E′N defined by LN = Π′N DxF (x, v)EN have a right-inverse, de-
noted by [LN (x, v)]−1
r . There is a constant µ > 0 and, for any A ∈ [0, S),
a positive constant γA, such that, for all s < A and (x, v) ∈ B (R, X0 × V )
we have:
∀k ∈ E′N , k[LN (x, v)]−1
r kks ≤
1
γA N µ(kkk′s + kxkskkk′0)
Corollary 8 Assume (a'), (b'), (c' ) are satisfied and F (0, 0) = 0 . Take any
α with 0 < α < S − 4µ. Then one can find ρ > 0 and C > 0 such that, for any
v with kvk ≤ ρ , there is a some x such that:
F (x, v) = 0
kxk0 ≤ 1
kxkα ≤ C kvk
6
Proof. Consider the Banach scale Xs × V and Ys × V with the natural norms.
Consider the map Φ (x, v) = (F (x, v) , v) from Xs × V , 0 ≤ s < S, into Ys × V .
It is roughly tame with Φ (0, 0) = (0, 0) and we can apply the preceding Theorem
with δ = ∞. Condition (10) becomes
α
<
µ
S
µ − 4
2.4 A particular case
In the case when F (x) = Ax + G (x) where A is linear, we can improve the
regularity.
Proposition 9 Suppose F (x) = Ax + G (x), where A : Xs+ν → Ys is a contin-
uous linear operator, independent of the base point x, and G satisfies G (0) = 0.
Suppose moreover that:
(a) G is continuous and Gateaux-differentiable from B0(R) ∩ Xs to Ys for any
s ∈ [0, S).
(b) For any A ∈ [0, S) there is a finite constant KA such that, for all s ≤ A
and x ∈ B0 (R):
∀h ∈ X, kDG(x)hk′s ≤ K A(khks + kxkskhk0)
(c) For x ∈ B0(R) ∩ EN , the linear maps LN (x) : EN → E′N have a right-
inverse, denoted by [LN (x)]−1
r . There is a constant µ > 0 and, for any
A ∈ [0, S), a positive constant γA, such that, for all s ∈ [0, S) and
x ∈ B0 (R) we have:
∀k ∈ E′N , k[LN (x)]−1
r kks ≤
1
γA N µ(kkk′s + kxkskkk′0)
(d) The EN are A-invariant.
Let δ > 0 and α > 0 be such that
δ
µ
α
µ
µ(cid:19)
> ϕ(cid:18) S
µ(cid:19) ,
µ − ϕ(cid:18) S
< min(cid:26) δ
S
µ − ϕ−1(cid:18) δ
µ(cid:19)(cid:27)
. Then one can find ρ > 0 and C > 0 such that, for any y ∈ Yδ with kyk′δ ≤ ρ
, there some x ∈ Xα such that:
F (x) = y
kxk0 ≤ 1
kxkα ≤ C kykδ
7
In this situation, a direct application of Theorem 5 would give a loss or
regularity of µ + ν. Proposition 9 tells us that the µ is enough: the loss of
regularity due to the linear part can be circumventend.
2.5 The approximating sequence
Theorem 5 is proved by an approximation procedure: we construct by induc-
tion a sequence xn having certain properties, and we show that it converges
to the desired solution. We now describe that sequence, and give the proof of
convergence. The actual construction of the sequence is postponed to the next
section.
Given an integer N and a real number α > 1, we shall denote by E [N α] the
integer part of N α:
E [N α] ≤ N α < E [N α] + 1
Theorem 10 Assume that µ, δ, S and κ satisfy condition ??. Choose σ and β
such that:
κ2
κ − 1
µ < κβ < σ < S
κβ > κµ + σ −
δ
κ
β > µ + σ − δ
Impose, moreover:
• For 1 < κ ≤ 1+√5
2
• For 1+√5
2 ≤ κ < 2
(11)
(12)
(13)
Then one can find N0 ≥ 2, ρ > 0 and c > 0 such that, for any y ∈ Y with
kyk′δ ≤ ρ, there are sequences (xn)n≥1 in B0(1) and Nn := N (κn)
satisfying:
0
• For 1 < κ ≤ 1+√5
2
,
Π′NnF (xn) = Π′Nn−1y and xn ∈ ENn
Π′Nn F (xn) = Π′Nn y and xn ∈ ENn
• For 1+√5
2 ≤ κ < 2
And in both cases:
kx1k0 ≤ cN µ
kx1kσ ≤ cN β
1 kyk′δ and kxn+1 − xnk0 ≤ c N κβ−σ
1 kyk′δ and kxn+1 − xnkσ ≤ c N κβ
n kyk′δ
n
kyk′δ
(14)
(15)
(16)
(17)
8
κ2
The set of admissible σ and β is non-empty. Indeed, because of (7), we have
κ−1 µ < S, so we can find κβ and σ satisfying condition (11). For 1 < κ ≤
(cid:0)1 + √5(cid:1) /2, we have δ > µκ2, so κµ − δ/κ < 0 and κβ can satisfy both (11)
and (12). For κ ≥(cid:0)1 + √5(cid:1) /2, we have δ − µ > κµ, so µ + σ − δ < σ − κµ and
condition (13) is satisfied provided β > σ − κµ, or κβ > κσ − κ2µ. If σ satisfies
(11), we have κσ − κ2µ > (κ − 1) σ > 0, so we can find κβ satisfying (11) and
(13).
Note that the estimate on kxn+1 − xnkσ blows up very fast when n → ∞,
while the estimate on kxn+1 − xnk0 goes to zero very fast, since κβ − σ < 0.
Using the interpolation inequality (3), this will enable us to maintain control of
some intermediate norms.
The proof of Theorem 10 is postponed to the next section. We now show
that it implies Theorem 5. Let u begin with an estimate:
3 such that, for all s ∈
Lemma 11 Given 0 < A < S there is a constant CA
[0, A] , all integers N, P ≥ 0 and any x ∈ B0(1) ∩ Xs :
3 kxks
3 N−skxks
3 N−skxks
Proof. The function ϕ : t ∈ [0, 1] → F (tx) has derivative d
dt ϕ = DF (tx)x and
by the tame estimates (4) on DF (x), we have k d
dt ϕ(t)k′s ≤ 2KAkxks. Since
ϕ(0) = 0, this gives our first estimate. Combining it with (2), we get the second
one, and applying (1) with d = 0, we get the third one.
kF (x)k′s ≤ CA
k(1 − Π′N )F (x)k′0 ≤ CA
kΠ′N +P (1 − Π′N )F (x)k′0 ≤ CA
Let us now prove Theorem 5. Since κβ − σ < 0, the inequalities (16) imply
that the sequence (xn) is Cauchy in X0, and has a limit ¯x with k¯xk0 ≤ Ck¯yk′δ,
where
C = c(N µ
N κβ−σ
n
)
1 +Xn≥1
n k¯yk′δ, with:
n N β
1 +
Then F (xn) converges to F (¯x) in Y0 , by the continuity of F : Xs → Y .
(17) implies that kxnkσ ≤ C′nN κβ
s .Similarly,
C′n := cN−κβ
Xi=1
and C′ := supn C′n < ∞, so that kxnkσ ≤ C′N κβ
n k¯yk′δ for all n.
The case 1 < κ ≤ 1+√5
. We have, by (14):
N κβ
2
n−1
i !
F (xn) = (1 − Π′Nn )F (xn) + Π′Nn−1 ¯y .
(18)
Then Lemma 11 gives the estimate
k(1 − Π′Nn)F (xn)k′0 ≤ C(3)
δ N−s
n kxnks
for all s < δ
9
Substituting kxnkσ ≤ C′N κβ
n kyk′δ, we get:
k(1 − Π′Nn )F (xn)k′0 ≤ C′ C(3)
δ N κβ−σ
n
kyk′δ
for all s ≤ δ
By (11), the exponent (κβ − σ) is negative. So (1 − Π′Nn )F (xn) converges
to zero in Y0. Now, using the inequality (2) we get
k(1 − Π′Nn−1)¯yk′0 ≤ C(1)
δ N−δ
n−1k¯yk′δ
for all s ≤ δ
so Π′Nn−1 ¯y converges to ¯y in Y0. So both terms on the right-hand side of (18)
converge to zero, and we get F (¯x) = ¯y, as announced.
Together with the interpolation inequality (3), conditions (16) and (17) im-
ply:
kxn+1 − xnk(1−t)σ ≤ ctN κβ−tσ
n
kyk′δ
The exponent on the right-hand side is negative for t > κβ/σ, so that
(1 − t) σ < σ − κβ. Arguing as above, if follows that k¯xkα ≤ C kyk′δ, provided:
where:
A1 =((κ, β, σ)
α < sup
A1 {σ − κβ }
κ−1 µ < κβ < σ < S )
σ − κβ < δ
κ − κµ
κ2
Set α′ = α/µ, σ′ = σ/µ, β′ = β/µ, δ. = δ/µ, S′ = S/µ. The problem
becomes:
1
α′ < sup
A′
{σ′ − κβ′ }
A′1 =((κ, β′, σ′)
κ−1 < κβ′ < σ′ < S′ )
σ′ − κβ′ < δ′
κ − κ
κ2
For given κ, Figure 3 gives the admissible (β, σ) region in the case S′ > δ′/κ− κ
(upper horizontal line) and in the case S′ < δ′/κ − κ (lower horizontal line).
The admissible region is to the right of the vertical β′ = κ/ (κ − 1), both in the
case S′ > δ′/κ − κ (right line) and in the case S′ < δ′/κ − κ (left line)
10
Figure 3: The admissible (β, σ) region in the first case
The maximum is attained at the upper left corner of the admissible region,
which is the point (β′, min{S′, κβ′ − κ + δ′/κ}), with β′ = κ (κ − 1)−1. Hence:
sup
A1 {σ′ − κβ′} = min(cid:26)S′ −
κ2
κ − 1
,
δ′
κ − κ(cid:27)
(19)
The case 1+√5
replace Π′Nn−1 ¯y by Π′Nn ¯y in (18).
We now have:
2 ≤ κ < 2 The argument is the same, except that we have to
2
α′ < sup
A′
{σ′ − κβ′ }
A′2 =(cid:26)(κ, β′, σ′)
σ′ < β′ + δ − 1
κ−1 < κβ′ < σ′ < S′ (cid:27)
κ2
For given κ, the admissible (β, σ) region is given in Figure 4, in the case
S > δ−1+κ/ (κ − 1) (upper horizontal line) and in the case S < δ−1+κ/ (κ − 1)
(lower horizontal line). The vertical is β = κ/ (κ − 1).
11
Figure 4: The admissible (β, σ) region in the second case
Again the maximum is attained in the upper left corner, which is the point
(β′, min{S′, δ′ − 1 + β′}) with β′ = κ (κ − 1). This gives:
sup
A2 {σ′ − κβ′ } = min(cid:26)S′ −
κ2
κ − 1
, δ′ − 1 − κ(cid:27)
(20)
Putting (19) and (20) together gives formula (10)
3 Proof of Theorem 2
We work under the assumptions of Theorem 10. So µ, δ, S, κ, σ, β, ¯y are given.
Note that we may have σ < δ.
1 , CA
We assume ¯y 6= 0 (the case ¯y = 0 is obvious). We fix A = σ , and the
3 , K A, γA of (1, 2, 3, 4, 5) and Lemma 11 are simply
constants CA
denoted C1, CA
3 , K , γ. The proof will make use of a certain number
of constants, which we list here to make sure that they do not depend on the
iteration step and can be fixed at the beginning.
2 , CA
2 , CA
Recall that 2(α) is the integer part of 2α, and set Pn = E(cid:2)2κn(cid:3). There is a
constant g > 1 such that for all N0 ≥ 2 and n ≥ 0,
n ≤ Pn+1 ≤ g P κ
g−1P κ
(21)
n
12
We define constants B0, B1 and B2 by:
)−1
n
P κβ−σ
B1 := sup
B0 = (P1 +Xn≥1
n {P −β
n nB1P −(κ−1)β
B2 := sup
n (P β
n
1 + X1≤i≤n−1
P κβ
i
) n ≥ 1}
+ 1 n ≥ 1o
(22)
(23)
(24)
We shall use Theorem 1 to construct inductively the sequence xn, thanks to
a sequence of carefully chosen norms. For this purpose, we will have to take c
large and ρ small.
4 Choice of N0
For 1 < κ ≤ 1+√5
ϕ2 (n) := 2 B1C3 c P β−κ(β−µ)
2
n
, we consider the following function of n:
+C1(cid:16)P σ−δ/κ−κ(β−µ)
n
gδ/κ + P (σ−δ)+−(δ−σ)+/κ−κ(β−µ)
n
(25)
By condition (11) and (12), all the exponents are negative. So we may pick
n0 so large that:
ϕ2 (n0) ≤ γc (B2 + 2)−1 g−µ for all n ≥ n0
(26)
For 1+√5
2 ≤ κ < 2, we consider the following function of N0 and n:
(cid:17)
(cid:17)
ϕ1 (n0) := 2 B1C3 c P β−κ(β−µ)
n
+ g(σ−δ)+ P κ(σ−δ)+−(δ−σ)+−κ(β−µ)
n
+C1(cid:16)P σ−δ−κ(β−µ)
n
(27)
By condition (11) and (13), all the exponents are negative. So we may pick
N0 so large that:
ϕ1 (n0) ≤ γc (B2 + 2)−1 g−µ for all n ≥ n0
(28)
In both cases we set N0 = Pn0 , and Nn = E(cid:2)N nκ
expressions (25) and (26) are less than γc (B2 + 2)−1 g−µ when one substitutes
Nn for Pn.
0 (cid:3) = E(cid:2)2(non)κ(cid:3). So the
4.1 Construction of the initial point
The case 1 < κ ≤ 1+√5
kΠ′N0 ¯ykσ ≤ C1N (σ−δ)+
0
2
number t. We choose the norm
. Thanks to inequality (2), kΠ′N0 ¯yk0 ≤ C1k¯yk′δ and
k¯yk′δ, where t+ denotes the positive part of the real
N0(x) = kxk0 + N−(σ−δ)+
0
kxkσ
13
on EN1 and the norm:
N ′0(y) = kyk′0 + N−(σ−δ)+
0
kyk′σ
on E′N1. For these norms, EN1 and E′N1 are Banach spaces. Note that
N ′0(Π′N0y) < 2C1kykδ
for y ∈ E′N1
(29)
For N0 (x) ≤ 1, we define
f (x) := Π′N1 F (x) ∈ E′N1
The function f is continuous and Gateaux-differentiable for the norms N0
and N ′0, with f (0) = 0. Moreover, using the tame estimate (5) and applying
assumption (1) to kxkσ, we find that:
sup(cid:8)(cid:13)(cid:13) [Df (x)]−1k(cid:13)(cid:13)0 N (x) ≤ 1(cid:9) <
sup(cid:8)(cid:13)(cid:13) [Df (x)]−1k(cid:13)(cid:13)σ N (x) ≤ 1(cid:9) <
hence:
2N µ
1
γ kkk′0
N µ
(kkk′σ + N (σ−δ)+
1
γ
0
kkk′0)
(30)
(31)
(32)
(33)
(34)
sup(cid:8)N0([Df (x)]−1k) N (x) ≤ 1(cid:9) <
3N µ
1
γ N ′0(k)
By Theorem 1, we can solve f (¯u) = ¯v with N0(¯u) ≤ 1 if N ′0(Π′N0 ¯y) ≤
γ (3N µ
1 )−1. By (29), this is fulfilled provided:
k¯yk′δ ≤
γ
6C1N µ
1
=: ρ .
In addition, Theorem 1 tells us that we have the estimate:
1 γ−1 k¯yk′δ
N0(¯u) ≤ 3N µ
1 γ−1N ′0(cid:0)Π′N0 ¯y(cid:1) ≤ 6C(1)N µ
If (33 is satisfied, x1 := ¯u is the desired solution in EN1 of the projected
equation Π′N1F (x1) = Π′N0 ¯y, with N0(¯u) ≤ 1. Let us check conditions (17) and
(16). We have, by (34):
N0(x1) = kx1k0 + N−(σ−δ)+
0
kx1kσ ≤ R = 6C(1)N µ
1 γ−1k¯yk′δ
Since N0 ≤ g1/κN 1/κ
1
, we find:
kx1k0 + (gN1)−κ−1(σ−δ)+ kx1kσ ≤ 6C(1)N µ
1 γ−1k¯yk′δ
Since µ + κ−1(σ − δ)+ < β, this yields kx1k0 ≤ cN µ
1 k¯yk′δ as required, with
cN β
1 k¯yk′δ and kx1kσ ≤
c := 6C(1)g(σ−δ)+/κγ−1
(35)
14
The case 1+√5
arguments. Replace N0 by N1, so that the norms become:
2 ≤ κ < 2 Very few modifications are needed in the above
N0(x) = kxk0 + N−(σ−δ)+
N ′0(y) = kyk′0 + N−(σ−δ)+
1
1
kxkσ
kyk′σ
and define as above f (x) := Π′N1F (x) ∈ E′N1 . Because (14) is replaced by
(15), we now consider Π′N1 ¯y ∈ E′N1 . Estimates (29) and (32) still hold. Using
Theorem 1 as before, we will be able do find some ¯u ∈ EN1 with N0(¯u) ≤ 1
solving f (¯u) = Π′N1 ¯y provided ¯y satisfies (33). The estimate (34) still holds:
N0(x1) = kx1k0 + N−(σ−δ)+
1
kx1kσ ≤ 6C1N µ
1 γ−1k¯yk′δ
Since µ + (σ − δ)+ < β, this yields kx1k0 ≤ cN µ
1 k¯yk′δ and kx1kσ ≤ cN β
as above, with
c := 6C(1)γ−1
1 k¯yk′δ
(36)
Diminishing ρ, if necessary, we can always assume that, in both cases, ρ and
c satisfy the constraint, where B0 is defined by (22):
cρ < B(0)
(37)
4.2
Induction
The case 1 < κ ≤ 1+√5
. Assume that the result has been proved up to n.
In other words, define c by (35), and assume we have found ρ with cρ < B(0)
such that, for ¯y ∈ Y with k¯yk′δ ≤ ρ, there exists a sequence x1,··· , xn satisfying
(14), (17), and (16). To be precise:
2
kxi+1 − xik0 ≤ c N κβ−σ
kxi+1 − xikσ ≤ c N κβ
i k¯yk′δ
i
k¯yk′δ
for i ≤ n − 1
for i ≤ n − 1
(38)
(39)
Since x1,··· , xn satisfy (16), and k¯yk′δ ≤ ρ, this will imply that kxnk0 ≤
1 − ηn with:
i
N µ
ηn := Pi≥n N κβ−σ
1 +Pn≥1 N κβ−σ
Since x1,··· , xn satisfy (17), we also have:
1 + X1≤i≤n−1
kxnkσ ≤ c (N β
n
≤ cρXi≥n
N κβ
i
)k¯yk′δ
N κβ−σ
i
.
(40)
Using the constant B1 defined in (22), this becomes:
kxnkσ ≤ B1 c N β
nk¯ykδ
(41)
15
We are going to construct xn+1 so that (14), (17), and (16) hold for i ≤ n.
Write:
xn+1 = xn + ∆xn
By the induction hypothesis, xn ∈ ENn and Π′NnF (xn) = Π′Nn−1 ¯y. The
equation to be solved by ∆xn may be written in the following form:
fn (∆xn) = en + ∆yn−1
fn(u) := ΠNn+1 (F (xn + u) − F (xn))
en := ΠNn+1(ΠNn − 1)F (xn)
∆yn−1 := ΠNn (1 − ΠNn−1)¯y
(42)
(43)
(44)
(45)
The function fn is continuous and Gateaux-differentiable with f (0) = 0. We
will solve equation (42) by applying Theorem 1. We choose the norms:
Nn(u) = kuk0 + N−σ
N ′n(v) = kvk′0 + N−σ
n kukσ on ENn+1
n kvk′σ on E′Nn+1
(46)
(47)
Let Rn := cN κβ−σ
n
k¯ykδ . If Nn(u) ≤ Rn, we have, by (40)
kuk0 ≤ Nn(u) < Rn < cN κβ−σ
ρ < ηn
n
so that kxn + uk0 ≤ 1 and the function fn is well-defined by (43). Using (41)
and (46), we find that, for Nn(u) ≤ Rn, we have kukσ ≤ N σ
n k¯ykδ,
and hence, with the constants B1 and B2defined by (23) and (24):
n Rn = cN κβ
kxn + ukσ ≤ c (B1 N β
n + N κβ
n )k¯ykδ ≤ B2 c N κβ
n k¯yk′δ
(48)
Plugging this into the tame inequality (5), and taking into account that
ck¯ykδ = RnN σ−κβ
n
then gives:
sup(cid:8)(cid:13)(cid:13) [Dfn(u)]−1k(cid:13)(cid:13)0 Nn(u) ≤ Rn(cid:9) < 2N µ
sup(cid:8)(cid:13)(cid:13) [Dfn(u)]−1k(cid:13)(cid:13)σ Nn(u) ≤ Rn(cid:9) < N µ
n+1γ−1kkk′0
n+1γ−1(kkk′σ + B2cN κβ
n+1γ−1(kkk′σ + B2 Rn N σ
= N µ
(49)
n kkk′0k¯yk′δ)
nkkk′0)
(50)
Hence:
sup(cid:8)Nn([Dfn(u)]−1k) Nn(u) ≤ Rn(cid:9) < γ−1(cid:16)B(2) + 2(cid:17) N µ
By Theorem 1, we will be able to solve (42) with Nn(∆xn) ≤ Rn provided:
N ′n (en + ∆yn−1) ≤ γ(cid:16)B(2) + 2(cid:17)−1
n+1Rn = γc(cid:16)B(2) + 2(cid:17)−1
n+1N ′n(k)
n+1N κβ−σ
N−µ
N−µ
k¯ykδ
(51)
n
16
We now compute N ′n (en + ∆yn−1). Applying inequality (2) to (45), we get:
(52)
k∆yn−1k′0 ≤ C1N−δ
n−1k¯yk′δ
k∆yn−1k′σ ≤ C1N−(δ−σ)+
n−1
N (σ−δ)+
n
k¯yk′δ
Because of the induction hypothesis, we get
kxnkσ ≤ c (N β
1 + X1≤i≤n−1
N κβ
i
)k¯yk′δ ≤ B1 c N β
nk¯yk′δ
So, combining Lemma 11 with conditions (17) and (16), we find that
kenk′0 ≤ C(3)N−σkxnkσ ≤ B1C3 c N β−σ
kenk′σ ≤ C(3)kxnkσ ≤ B1C3 c N β
nk¯yk′δ
n
k¯yk′δ
(53)
(54)
(55)
(56)
We now check (51). Using the estimates (52), (53), ( 55), (56), and remem-
bering (21), we have :
N ′n(en) ≤ 2B1C3 c N β−σ
N ′n(∆yn−1) ≤ C1k¯yk′δ(cid:16)N−δ
N ′n(en) + N ′n(∆yn−1) ≤h2 B1C3 c N β−σ
n−1 + N−σ
n + C1(cid:16)N−δ
k¯yk′δ
n
n−1N−(δ−σ)+
N (σ−δ)+
(cid:17)
n−1
n−1 + N−σ−(δ−σ)+
n−1
n
Condition (51) is satisfied provided:
(57)
(58)
N (σ−δ)+
n
(cid:17)ik¯yk′δ
(59)
2 B1C3 c N β
n +C1(cid:16)N σ−δ/κ
n
gδ/κ + N (σ−δ)+−(δ−σ)+/κ
(cid:17) ≤ γc (B2 + 2)−1 g−µN κ(β−µ)
n
n
Dividing by N κ(β−µ)
n
(60)
on both sides, we find that the sequence on the left-
hand side is just a subsequence of ϕ2 (n), where ϕ2 is defined by (25), and so
(60) follows directly from (26), that is, from the construction of N0. So we
may apply Theorem 1, and we find u = ∆xn with Nn(∆xn) ≤ Rn solving (42).
Since Nn(∆xn) ≤ Rn, we have k∆xnkσ ≤ N σ
n k¯yk′δ and k∆xnk0 ≤
Rn = cN κβ−σ
k¯yk′δ , so inequalities (16) and (17) are satisfied. The induction is
proved.
n Rn = cN κβ
n
The case 1+√5
Π′Nn ¯y. The system (42) (43), (44), 45) becomes:
2 ≤ κ < 2. The induction hypothesis becomes Π′Nn F (xn) =
fn (∆xn) = en + ∆yn
fn(u) := ΠNn+1 (F (xn + u) − F (xn))
en := ΠNn+1(ΠNn − 1)F (xn)
∆yn := ΠNn (1 − ΠNn )¯y
17
Using Theorem 1 in the same way, we find that we can find ∆xn satisfying
these equations and the estimates (16) and (17) provided:
2 B(3) c N β
n + g(σ−δ)+ N κ(σ−δ)+−(δ−σ)+
n
n +C1(cid:16)N σ−δ
But the left-hand side is just ϕ1 (n, N0) N κ(β−µ)
(61)
, where ϕ1 is defined by (27),
and so (61) follows directly from (28), that is, from the construction of N0. The
induction is proved in this case as well.
n
(cid:17) ≤ γc(cid:16)B(2) + 2(cid:17)−1
g−µN κ(β−µ)
n
4.3 Proof of Proposition 9
We will take advantage of the special form of en := ΠNn+1(ΠNn − 1)F (xn).The
proof is the same, with the estimates (55) and (56) derived as follows. Since
xn ∈ ENn , and ENn is A-invariant, we have ΠNnAxn = Axn and:
en := ΠNn+1(ΠNn − 1)F (xn)
= ΠNn+1(ΠNn − 1) (Axn + G(xn))
= ΠNn+1(ΠNn − 1)G(xn)
Lemma 11 holds for G (though no longer for F ), so that estimates (55) and
(56) follow readily.
5 An isometric imbedding
We will use the same example as Moser in his seminal paper [14], who himself
follows Nash [15]. Suppose we are given a Riemannian structure g0 on the two-
dimensional torus T2 = (R/Z)2, and an isometric imbedding into Euclidian R5.
In other words, we know x0 =(cid:0)x0
(cid:18) ∂x0
1, ..., x0
5(cid:1) with:
∂θj(cid:19) = g0
∂x0
∂θi
,
i,j (θ1, θ2)
i,j = g0
where g0
j,i, so there are three equations for five unknown functions. If we
slightly perturb the Riemannian structure, does the imbedding still exist ? If
we replace g0 on the right-hand side by some g sufficiently close to g0, can we
still find some x : T2 → R5 which solves the system ?
We consider the Sobolev spaces H s(cid:0)T2; R5(cid:1) and we assume that x0 ∈ H µ,
with µ > 3, and g0 ∈ H σ. Define F =(cid:0)F i,j(cid:1) by:
∂θi (cid:0)x + x0(cid:1) ,
Fi,j(cid:0)x + x0(cid:1) =(cid:18) ∂
∂θj (cid:0)x + x0(cid:1)(cid:19) − g0
Clearly F (0) = 0. For s ≥ 3/2, we know H s is an algebra, so if s ≥ µ
and x ∈ H s, the first term on the right is in H s−1. On the other hand, the
right-hand side cannot be more regular than g0, which is in H σ. So F sends H s
(62)
∂x
18
into H s−1 for µ ≤ s ≤ σ + 1. The function F is quadratic, hence smooth, and
we have:
[DF (x) u]i,j =(cid:18) ∂x
∂θi
,
∂x
∂θj(cid:19)
∂u
∂θj(cid:19) +(cid:18) ∂u
∂θi
,
(63)
(64)
We need to invert the derivative DF (x), that is, to solve the system
DF (x) u = vi,j
It is an undetermined system, since there are three equations for five un-
knowns. Following Nash, and Moser, we impose two additional conditions:
(cid:18) ∂x
∂θ1
, u(cid:19) =(cid:18) ∂x
∂θ2
, u(cid:19) = 0
Differentiating, and substituting into (63), we find:
− 2(cid:18) ∂2x
∂θi∂θj
, u(cid:19) = vi,j
(65)
(66)
So any solution ϕ of (65), (66) is also a solution of (64). The five equations
(65), (66) can be written as:
M (x (θ)) u (θ) =(cid:18) 0
v (θ) (cid:19)
with u = (cid:0)ui(cid:1) , 1 ≤ i ≤ 5, v = (cid:0)v11, v12, v22(cid:1) and M (x (θ)) a 5 × 5 matrix.
These are no longer partial differential equations. If
det M (x (θ)) = det(cid:18) ∂x
∂θ1
,
∂x
∂θ2
,
∂2x
∂θ2
1
,
∂2x
∂θ1∂θ2
,
∂2x
∂θ2
2(cid:19) 6= 0
(67)
they can be solved pointwise. Set L (x (θ)) := M (x (θ))−1, and denote by M (x)
and L (x) the operators u (θ) → M (x (θ)) u (θ) and v (θ) → L (x (θ)) v (θ).
Since x0 ∈ H µ, with µ > 3, we have x0 ∈ C2, so M(cid:0)x0 (θ)(cid:1) is well-defined
and continuous with respect to θ. If det M(cid:0)x0 (θ)(cid:1) 6= 0 for all θ ∈ T2, then
there will be some R > 0 and some ε > 0 such that det M (x (θ)) ≥ ε for all
x with (cid:13)(cid:13)x − x0(cid:13)(cid:13)µ ≤ R. So the operator L (x) is a right-inverse of DF (x) on
(cid:13)(cid:13)x − x0(cid:13)(cid:13)µ ≤ R, and we have the uniform estimates, valid on x0 + Bµ (R) and
s ≥ 0
kDF (x) uk0 ≤ C0 kxk1 kuk1
kDF (x) uks ≤ Cs(cid:0)kxks+1 kuk1 + kxk1 kuks+1(cid:1)
kL (x) vk0 ≤ C0 kxkµ kvk0
kL (x) vks ≤ C′s(cid:16)kxkµ+s kvk0 + kxkµ kvks(cid:17)
19
This means that the tame estimate (4) is satisfied. However, (5) requires
a proof. For this, we have to build a sequence of projectors ΠN satisfying the
estimates (1) and (2). For this, we use a multiresolution analysis of L2 (R) (see
[13]). Recall that it is an increasing sequence VN , N ∈ Z, of closed subspaces of
L2 (R) with the following properties:
• ∩N =+∞N =−∞
VN = {0} and ∪N =+∞N =−∞
• u (t) ∈ VN ⇐⇒ u (2t) ∈ VN +1
• ∀k ∈ Z, u (t) ∈ V0 ⇐⇒ u (t − k) ∈ V0
• there is a function ϕ (t) ∈ V0 such that the ϕ (t − k) , k ∈ Z, constitute a
VN is dense in L2
Riesz basis of L2.
It is known ([13], Theorem III.8.3) that for every r ≥ 0 there is a multireso-
lution analysis of L2 (R) such that:
• the ϕ (t − k) , k ∈ Z, constitute an orthogonal basis of V0
• ϕ (t) is Cr and has compact support: there is some a (depending on r)
such that t ≥ a =⇒ ϕ (t) = 0.
We choose r so large that Cr ⊂ H S. Set ϕN (t) := 2N/2ϕ(cid:0)2N t(cid:1). For N
large enough, say N ≥ N0, the function ϕN has its support in ] − 1/2, 1/2[ so
we can consider it as a function on R/Z, and the ϕN,k (θ) := 2N/2ϕ(cid:0)2N θ − k(cid:1),
for 0 ≤ k ≤ 2N − 1, constitute an orthonormal basis of VN . In this way, we get
a multiresolution analysis on L2 (R/Z). Setting
ΦN,k (θ) = 2N ϕ(cid:0)2N θ1 − k1(cid:1) ϕ(cid:0)2N θ2 − k2(cid:1)
EN = Span(cid:8)ΦN,k (θ) k = (k1, k2) , 0 ≤ k1, k2 ≤ 2N − 1(cid:9)
we get a multiresolution analysis of L2 (T2), and the ΦN,k (θ) are an orthonormal
N constitute a multiresolution analysis of L2 (T2)5 =
basis of EN . Finally, the E5
L2(cid:0)T2, R5(cid:1). Denote by ΠN the orthogonal projection:
< ΦN,k, ui > ΦN,k
(ΠN u)i =Xi,k
Introduce an orthonormal basis of wavelets associated with this multireso-
N(cid:1)N≥0 of L2(cid:0)T2, R5(cid:1). More precisely, the ΦN0,k, 0 ≤ k1, k2 ≤
lution analysis(cid:0)E5
2N − 1, span EN0 , and one can find a Cr function Ψ with compact support such
that the ΨN,k = 2N Ψ(cid:0)2N θ1 − k1, 2N θ2 − k2(cid:1) span the orthogonal complement
of EN−1 in EN . The ΦN0,k and the ΨN,k for N ≥ N0 constitute an orthonormal
basis for L2 (T2). We have, for u ∈ L2 (T2)
ui =Xk
Xi=1
L2 =
5
kuk2
hui, ΦN0,kiΦN0,k + XN≥N0,Xk
Xk
hui, ΦN0,ki2 + Xn≥N0,Xk
hui, ΨN,kiΨN,k
hui, Ψn,ki2
20
5
kΠN uk2 =
It follows from the definition that, for N ≥ N0, we have:
hui, ΦN0,ki2 + XN0≤n≤N,Xk
hui, Ψn,ki2
Xk
Xi=1
Xi=1 Xn≥N,Xk
kΠN u − uk2
L2 =
5
hui, Ψn,ki2
It is known (see [13] Theorem III.10.4) that:
C1 kuk2
H s ≤ Xk∈(KN )5
If v = ΠN u, we must have hv, Ψn,ki = 0 for all n ≥ N + 1, so that:
22N0shu, ΦN0,ki2+ Xn≥N0, Xk∈(Kn)5
22nshu, Ψn,ki2 ≤ C2 kuk2
H s
C1 kΠN uk2
kΠN u − uk2
H s ≤ 22N s kΠN uk2
L2 ≤ 2−2N s Xn≥N, Xk∈(Kn)5
≤ 2−2N sC2 kuk2
So estimates (1) and (2) have been proved.
Finally, we prove (5). We have:
H s
L2 ≤ 22N s kuk2
L2
22nshu, Ψn,ki2
(ΠN u)i (θ) =Xk
(M (x (θ)) ΠN u)j (θ) =Xk Xi
ui
kϕN,k (θ) , 1 ≤ i < 5
ϕN,k (θ)(cid:16)M j
i (x (θ)) ui
k(cid:17)
So ΠN M (x) ΠN is a (cid:0)2N − 1(cid:1) ×(cid:0)2N − 1(cid:1) matrix of 5 × 5 matrices mk,k′ ,
with:
ϕN,k (θ) ϕN,k′ (θ) M (x (θ))
mk,k′ =ZT2
Looking at the supports of ϕN,k and ϕN,k′, we see that there is a band along
the diagonal outsides which mkk′ vanishes.
mk,k′ = 0 if max
i=1,2 ki − k′i > 21−N a
(68)
Choose some ε > 0 and N so large that maxi=1,2(cid:12)(cid:12)θ − 2−N ki(cid:12)(cid:12) ≤ 21−N a
implies that (cid:13)(cid:13)M (x (θ)) − M(cid:0)x(cid:0)2−N k(cid:1)(cid:1)(cid:13)(cid:13) ≤ ε. Then:
(cid:13)(cid:13)(cid:13)(cid:13)
ϕN,k (θ) ϕN,k′ (θ) M(cid:0)x(cid:0)2−N k(cid:1)(cid:1) dθ(cid:13)(cid:13)(cid:13)(cid:13)
mk,k′ −ZT2
≤ εZT2 ϕN,k (θ)ϕN,k′ (θ) dθ
21
Using the fact that the system ϕN,k is orthonormal, we get:
(cid:13)(cid:13)mk,k′ − δk,k′ M(cid:0)x(cid:0)2−N k(cid:1)(cid:1)(cid:13)(cid:13) ≤ ε (max ϕ)2
In addition, for every k, (68) gives us at most 4a2 non-zero values for k′.
is a small perturbation of the diagonal matrix ∆N :=
So the matrix mk,k′
∆−1
N
So (ΠN M (x) ΠN ) is invertible for ε small enough, for instance:
(ΠN M (x) ΠN )−1 =(cid:2)I + ∆−1
N (ΠN M (x) ΠN − ∆N )(cid:3)−1
N (cid:13)(cid:13) ≤ maxθ(cid:13)(cid:13)(cid:13)
δk,k′ M(cid:0)2−N k(cid:1) , k ∈ KN , which is invertible by (67). More precisely, we have:
M (x (θ))−1(cid:13)(cid:13)(cid:13)
with kΠN M (x) ΠN − ∆Nk ≤ ε4a2 (max ϕ)2 and(cid:13)(cid:13)∆−1
θ (cid:13)(cid:13)(cid:13)
M (x (θ))−1(cid:13)(cid:13)(cid:13)
N (cid:13)(cid:13) = 2(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)
(ΠN M (x) ΠN )−1(cid:13)(cid:13)(cid:13) ≤ 2(cid:13)(cid:13)∆−1
Estimate (5) then follows immediately
We now introduce the new spaces X s := H s+µ and Y s := H s+µ−1. We
denote their norms by kxk∗s := kxks+µ and kyk∗s := kyks+µ−1 respectively, so
the above estimates become:
M (x (θ))−1(cid:13)(cid:13)(cid:13)(cid:13)
ε4a2 (max ϕ)2 max
and we have:
min
.
<
1
2
θ
kDF (x) uk∗s ≤ Cs(cid:0)kxk∗s kuk∗0 + kxk∗0 kuk∗s(cid:1)
u kL (x) vk∗s ≤ C′s(cid:16)kxk∗s+µ kvk∗1 + kxk∗0 kvk∗s+1(cid:17) for s ≥ 0
for s ≥ 0
and the range for s becomes 0 ≤ s ≤ S with S = σ + µ− 1.We have proved that
all the conditions of Corollary 6 are satisfied, with x0 ∈ H µ = X 0, g0 ∈ H σ =
Y σ−µ+1 and S = σ − µ + 1. Hence:
Theorem 12 Take any µ > 3. Suppose x0 ∈ H µ and g0 ∈ H σ with σ ≥
5µ− 1 > 14. Suppose the determinant (67) does not vanish. Set S := σ − µ + 1.
Then, for any δ and α such that:
δ
µ(cid:19)
µ ≥ ϕ(cid:18) S
µ(cid:19) ,
µ − ϕ(cid:18) S
< min(cid:26) δ
α
µ
S
µ − ϕ−1(cid:18) δ
µ(cid:19)(cid:27)
there is some ρ > 0 and some C > 0 such that, for any g with (cid:13)(cid:13)g − g0(cid:13)(cid:13)δ+µ−1 ≤
ρ , there is x ∈ H µ+α such that:
∂x
,
∂θi
∂θj(cid:19) = gi,j (θ1, θ2)
(cid:18) ∂x
(cid:13)(cid:13)x − x0(cid:13)(cid:13)µ ≤ 1
kx − x0kµ+α ≤ C(cid:13)(cid:13)g − g0(cid:13)(cid:13)µ+δ−1
22
∂x
,
∂θi
∂θj(cid:19) = gi,j (θ1, θ2)
(cid:18) ∂x
(cid:13)(cid:13)x − x0(cid:13)(cid:13)µ ≤ 1
kx − x0kµ+α ≤ C(cid:13)(cid:13)g − g0(cid:13)(cid:13)µ+δ−1
Moser [14] finds that if g, g0 ∈ Cr+40 and x0 ∈ Cr for some r ≥ 2, and if
(cid:12)(cid:12)g − g0(cid:12)(cid:12)r is sufficiently small, then we can solve the problem. Although he made
no effort to get optimal differentiability assumptions, we note that our loss of
regularity is substantially smaller (σ − µ ≥ 4µ − 1 > 11 instead of 40).
Note that when µ → 3, we have σ → 14 and δ → ∞. In another direction:
Corollary 13 Suppose x0 ∈ C∞, g0 ∈ C∞, and the determinant (67) does not
vanish. Then, for any δ > µ > 3 and any α < δ − µ, there is some ρ > 0
and some C > 0 such that, for any g with (cid:13)(cid:13)g − g0(cid:13)(cid:13)δ+µ−1 ≤ ρ , there is some
x ∈ H α+µ such that:
References
[1] Alinhac, Serge and G´erard, Patrick, "Op´erateurs Pseudo-diff´erentiels et
Th´eor`eme de Nash-Moser", Inter´editions et ´Editions du CNRS, Paris
(1991). English translation: "Pseudo-differential Operators and the Nash-
Moser Theorem", Graduate Studies in Mathematics vol. 82, AMS, Rhode
Island, (2000)
[2] Arnol'd, Vladimir I. "Small divisors", Dokl. Akad. Nauk CCCP 137 (1961),
255-257 and 138 (1961), 13-15 (Russian)
[3] Arnol'd, Vladimir I. "Small divisors I ", Izvestia Akad. Nauk CCCP 25
(1961), 21-86 (Russian)
[4] Arnol'd, Vladimir I. "Small divisors II ", Ouspekhi Mathematitcheskikh
Nauk 18 (1963), 81-192 (Russian)
[5] Berti, Massimiliano, and Bolle, Philippe. "Cantor families of periodic so-
lutions for completely resonant nonlinear wave equations ", Duke Mathe-
matical Journal, 134 (2006), 359-419
[6] Berti, Massimiliano, and Bolle, Philippe. "Cantor families of periodic so-
lutions for completely resonant nonlinear wave equations ", NoDEA 15
(2008), 247-276
[7] Berti, Massimiliano, and Bolle, Philippe. "Sobolev periodic solutions of non-
linear wave equations in higher spatial dimensions", Archive for Rational
Mechanics and Analysis, 195 (2010), 609 -642
[8] Berti, Massimiliano, Bolle, Philippe, Procesi, Michela. "An abstract Nash-
Moser theorem with parameters and applications to PDEs", Ann. IHP C
27, no. 1 (2010), 377-399
23
[9] Gunther, Mathias, "Isometric embeddings of Riemannian manifolds",
Proc. ICM Kyoto (1990), 1137-1143
[10] Hamilton, Richard, "The Inverse Function Theorem of Nash and Moser "
Bull. AMS 7 (1982), 165-222
[11] Hormander, Lars, "The Boundary Problems of Physical Geodesy". Arch.
Rat. Mech. An. 62 (1976), 1-52
[12] Ekeland, Ivar, "An inverse function theorem in Fr´echet spaces", Ann. IHP
C 28, no. 1 (2011), 91-105
[13] Meyer, Yves, "Ondelettes ", Hermann, Paris, 1990
[14] Moser, Jurgen, "A new technique for the construction of solutions to non-
linear differential equations", Proc. Nat. Acad. Sci. USA, 7 (1961), 1824-31
[15] Nash, John, "C1 isometric imbeddings" Ann. of Math. (2) 60 (1954). 383–
396.
24
|
1103.0047 | 1 | 1103 | 2011-02-28T22:48:51 | Stabilizing isomorphisms from $\ell_p(\ell_2)$ into $L_p[0,1]$ | [
"math.FA"
] | Let $1<p\not=2<\infty$, $\epsilon>0$ and let $T:\ell_p(\ell_2)\overset{into}{\rightarrow}L_p[0,1]$ be an isomorphism. Then there is a subspace $Y\subset \ell_p(\ell_2)$ $(1+\epsilon)$-isomorphic to $\ell_p(\ell_2)$ such that: $T_{|Y}$ is an $(1+\epsilon)$-isomorphism and $T(Y)$ is $K_p$-complemented in $L_p[0,1]$, with $K_p$ depending only on $p$. Moreover, $K_p\le (1+\epsilon)\gamma_p$ if $p>2$ and $K_p\le (1+\epsilon)\gamma_{p/(p-1)}$ if $1<p<2$, where $\gamma_r$ is the $L_r$ norm of a standard Gaussian variable. | math.FA | math |
Stabilizing isomorphisms from ℓp (ℓ2) into Lp [0, 1]
Ran Levy∗ and Gideon Schechtman†
November 10, 2018
Abstract
Let 1 < p 6= 2 < ∞, ε > 0 and let T : ℓp(ℓ2)
into
→ Lp[0, 1] be an isomorphism Then
there is a subspace Y ⊂ ℓp(ℓ2) (1 + ε)-isomorphic to ℓp(ℓ2) such that: TY is an (1 + ε)-
isomorphism and T (Y ) is Kp-complemented in Lp [0, 1], with Kp depending only on p.
Moreover, Kp ≤ (1 + ε)γp if p > 2 and Kp ≤ (1 + ε)γp/(p−1) if 1 < p < 2, where γr is
the Lr norm of a standard Gaussian variable.
1
Introduction
Let B be one of the 5 "classical" subspaces of Lp = Lp[0, 1]; by these we mean ℓp,
ℓ2, ℓp ⊕p ℓ2, ℓp(ℓ2) and Lp itself. Here ℓp(ℓ2) is the space of (say, real) matrices A =
(ai,j)∞
i,j)p/2)1/p. It is well known that these
five spaces isometrically embed in Lp, 1 ≤ p < ∞, and, for 1 < p < ∞, have embeddings
which are complemented. (For p = 1 this last statement holds only for L1 and ℓ1.)
i,j=1 with norm kAkℓp(ℓ2) = (P∞
j=1(P∞
i=1 a2
It was known for some time that if X is any subspace of Lp, 1 < p < ∞, isomorphic
to one of these B then there is a subspace of X isomorphic to B and complemented in
Lp. For ℓp see [KP], for ℓ2, [PR]. The case ℓp ⊕p ℓ2 follows from these two results. The
quite complicated case of Lp was proved in [JMST] (and for p = 1 perviously in [ES]).
The case of ℓp(ℓ2) can be proved using a variation of the method of [JMST] (and there
is also a much simpler proof for p > 2) and was known to the second named author for
a long time but not published (the simpler proof for p > 2 is included in [HOS]).
Recently, there were three paper which address this property for some of the spaces
B above and related questions again. This was done mostly because some strength-
ening of this property was needed for other purposes. Firstly, Haydon, Odell and
∗This is part of the first author MSc thesis written at the Weizmann Institute under the supervision of
the second author
†supported in part by the Israel Science Foundation and by the Israel-US binational Science Foundation
1
Schlumprecht proved in [HOS, Theorem 6.8] that, for p > 2, any subspace of Lp iso-
morphic to ℓp(ℓ2) contains a subspace (1 + ε)-isomorphic to ℓp(ℓ2) and complemented
in Lp by means of a projection of norm at most (1 + ε)γp, γp being the Lp norm of a
standard Gaussian variable. Their proof uses the fact that a similar theorem holds for
the space ℓ2 in Lp; i.e., any subspace of Lp, p > 2, isomorphic to ℓ2 contains a subspace
(1 + ε)-isomorphic to ℓ2 and complemented in Lp by means of a projection of norm
at most (1 + ε)γp. This later deep fact, hidden already in Aldous' [Ald], was recently
given a simpler proof by Alspach [Als]. The third paper is by Dosev, Johnson and the
second named author [DJS, Theorem 3.4] from which the following follows:
For each 1 < p < ∞ there is a constant Kp, depending only on p such that if
T : Lp → Lp is an isomorphism (into) then there is a subspace X of Lp Kp-isomorphic
to Lp such that TX is a Kp-isomorphism and T (X) is Kp-complemented in Lp.
We remark that a similar theorem for p = 1 (with the constant K1 arbitrarily close to
1) is due to Enflo and Starbird [ES]; see also [R] for a somewhat simpler exposition.
The main purpose of the current paper is to prove a similar theorem for ℓp(ℓ2).
Theorem 1. Let 1 < p 6= 2 < ∞, ε > 0 and let T : ℓp(ℓ2) onto→ X be an isomorphism
where X ⊂ Lp [0, 1]. Then there is a subspace Y ⊂ ℓp(ℓ2) (1 + ε)-isomorphic to ℓp(ℓ2)
such that TY is an (1+ε)-isomorphism and T (Y ) is complemented in Lp [0, 1] by means
of a projection of norm at most (1 + ε)γp if p > 2 and (1 + ε)γp/(p−1) if 1 < p < 2.
For p > 2 this theorem follows easily from the result of [HOS] mentioned above so
the main innovation here is the case 1 < p < 2. However, since the addition needed to
present a uniform proof for 1 < p < 2 and p > 2 is minimal, we shall prove both cases.
The proof is very much in the spirit of [DJS] but we wrote it in such a manner that
one does not need to refer to that paper.
2 Preliminaries
The proofs below will assume familiarity with basic techniques of Banach space theory.
In particular techniques related to bases. We shall use freely notions like unconditional
bases, block bases, small perturbations of bases, gliding hump arguments and similar
notions. They can all be found in the first chapter of [LT-I].
Recall that the Haar system is the following sequence of functions on [0, 1]:
h0,0(t) ≡ 1 and, for n = 0, 1, . . . and i = 1, 2, . . . , 2n,
if t ∈ ((2i − 2)2−(n+1), (2i − 1)2−(n+1))
1
−1 if t ∈ ((2i − 1)2−(n+1), 2i2−(n+1))
0
otherwise
hn,i(t) =
This system forms an unconditional basis for Lp = Lp[0, 1] for each 1 < p < ∞ (but
not in L1 in which it is, in its natural order, a non-unconditional Schauder basis) see
e.g. [LT-II]. We denote by Hp its unconditional constant; i.e.,
kX an,ihn,ikp ≤ HpkX εn,ian,ihn,ikp
for any sequence of coefficients {an,i} and any sequence of signs {εn,i}. (We shall use
the real field although all the arguments easily carry over to the complex field with
2
minimal changes, one of which should be in this definition). The best constant Hp is
known and is of order max{p, p/(p − 1)}.
Recall Khinchine's inequality; For 1 ≤ p < ∞,
n
Ap(
Xi=1
a2
i )1/2 ≤ (Ave±
n
Xi=1
n
±aip)1/p ≤ Bp(
Xi=1
a2
i )1/2
for all n and all coefficients {ai}n
i=1. The best constants Ap and Bp are known and in
particular Ap is between 2−1/2 and 1 for 1 ≤ p ≤ 2 and 1 for p > 1 and Bp is 1 for
1 ≤ p ≤ 2 and of order p1/2 for p > 2.
We shall make an intensive use of the square function with respect to the Haar
system. For f ∈ Lp, 1 < p < ∞, with expansion f = P an,ihn,i we denote its square
function with respect to the Haar system by
S(f ) = (X a2
n,ih2
n,i)1/2.
The unconditionality of the Haar system and Khinchine's inequality easily imply that
H −1
p ApkS(f )kp ≤ kf kp ≤ HpBpkS(f )kp, 1 < p < ∞.
Equi-integrability of some sets of functions will play an important role in the sequel.
Recall that a set F of Lebesgue integrable functions on [0, 1] is said to be equi-integrable
if for all ε > 0 there is a δ > 0 such that for every subset A of [0, 1] of measure at most
δ, RA f dt < ε for all f ∈ F . Equivalently, if For for all ε > 0 there is a positive R such
that R f 1f >Rdt < ε for all f ∈ F .
For 0 < r < ∞ the set F is r-equi-integrable if {f r ; f ∈ F } is equi-integrable.
Finally, by a K-isomorphism we mean a linear map T from one normed space X
into another Y such that A−1kxk ≤ kT xk ≤ Bkxk for all x ∈ X with AB ≤ K. In
particular, for small ε a (1 + ε)-isomorphism T does not necessarily almost preserve the
norm of each x (as is sometimes assumed in other places) but of course some multiple
of T does.
3 Stabilazing embeddings of ℓ2 into Lp
In this section we consider the analogue of Theorem 1 for the space ℓ2 instead of
ℓp(ℓ2). As we indicated above this theorem is known although not simple, especially if
one wants to achieve the best constants. (For a somewhat weaker Theorem, in terms
of the constants achieved, see Theorem 3.1 in [PR].) The purpose of this section is to
survey its proof and point the reader to the relevant references.
Theorem. Let ε > 0 and let T : ℓ2 → Lp [0, 1], 1 ≤ p < ∞, be an isomorphism.
Then there is an infinite dimensional subspace X ⊆ ℓ2 such that TX is an (1 + ε)-
isomorphism and, for 1 < p < ∞, T X is (1 + ε) γp-complemented in Lp. Here, for
p > 2, γp is the Lp norm of a standard Gaussian variable and, for 1 < p < 2, the L p
p−1
norm of such a variable.
Note first that in order to prove the first part of the theorem, the existence of a
(1 + ε)-isomorphism on X ⊆ ℓ2, it is enough to prove that T X contains a Y which is
(1 + ε)-isomorphic to ℓ2. Indeed if this is the case let {ei}∞
i=1 ⊆ X be an orthonormal
3
basis of X and {fi}∞
Since {T ei}∞
as small perturbation as we wish of a block basis of {fi}∞
i=1 is a weakly null sequence in T X for which {kT eik}∞
i=1 ⊆ T X a basis (1 + ε)-equivalent to an orthonormal basis of ℓ2.
i=1 is bounded and
j=1 is
j=1 is as close
j=1 is then the subspace we are
after. That T X contains a subspace (1 + ε)-isomorphic to ℓ2 is by now well known and
follows from the stability of Lp (see [KM]).
bounded away from zero, we may find a subsequence (cid:8)eij(cid:9)∞
as we want to a constant sequence. X0 = span(cid:8)eij(cid:9)∞
j=1 such that (cid:8)T eij(cid:9)∞
i=1 and(cid:8)(cid:13)(cid:13)T eij(cid:13)(cid:13)(cid:9)∞
We are left with the problem of complementation, and especially the norm of the
best projection. For p > 2 the theorem is specifically stated and proved in [HOS].
There is a simpler proof in [Als]. For 1 ≤ p < 2 this follows, as we shall indicate
momentarily from [Ald]. This is not at all an easy paper to follow and it would be
nice if somebody finds an easier proof maybe a-la-[Als]. We shall only sketch how to
get the result from [Ald] and then give a much simpler argument which however gives
a somewhat weaker estimate.
Let Y ⊆ Lp [0, 1], 1 < p < 2, be isomorphic to ℓ2. We would like to find a subspace
Y0 ⊆ Y (1 + ε)-isomorphic to ℓ2 and (1 + ε) γ p
p−1
-complemented in Lp [0, 1].
By [KP] the unit ball of Y is p-equi-integrable and the L1 and Lp norms are equiv-
alent on Y . From here on we shall use the notations of [Ald]. By the combination of
Proposition 3.9 and Theorem 3.10 there, there is a uniformly integrable sequence Xn
in Y such that i (Xn) wm→ σ (q, α) for some 1 < q ≤ 2. Recall that i (Xn) is the random
wm→ ξ denotes hf, ξni w→ hf, ξi
measure δXn and that for random measures ξn, ξ, ξn
w
→ denotes weak convergence in L1). Finally, for a function
for all f ∈ C (R) (and
α ≥ 0, σ (q, α) is the random measure whose characteristic function is e−αqtq
; i.e., it
is a mixture of symmetric q-stable random variables.
In our case, by Proposition 3.11 of [Ald] the only possible value for q is q = 2. So
we get a sequence Xn which tends in some sense to a mixture of Gaussian variables.
Proposition 3.11 and its proof then say that some subsequence of Xn is (1 + ε) equiv-
alent to the unit vector basis of ℓ2. The proof really gives more: some subsequence
of Xn is arbitrarily close, in Lp norm to a sequence of the form αZn where, given α,
Zn are i.i.d N (0, 1). This means that after a change of density we may assume that
{Xn} is a small perturbation in the Lp norm of a sequence of i.i.d N (0, 1) variables
Zn. By a change of density we mean an operator of the form Tϕ : Lp → Lp([0, 1], ϕdt),
Tϕf = f
ϕ1/p where ϕ is a density which is strictly positive on the union of the sup-
It is thus enough to show that the span of such a sequence is
ports of the Xn-s.
it is clearly enough to show that
γ p
p−1
complemented in Lp. Since {Zn}∞
n=1 ⊆ L p
p−1
the orthogonal projection P : L2
erator on L p
p−1
= γ p
kP f k p
p−1
p−1
kP f k2 ≤ γ p
. Since for all f ∈ L p
p−1
kf k2 ≤ γ p
p−1
onto
→ [Zn] has norm γ p
p−1
when considered as an op-
(so also f ∈ L2) P f is a Gaussian variable
kf k p
p−1
.
p−1
This concludes the (admitadly very rough) sketch of the proof of Theorem 3
which does not use [Ald], but gives a somewhat worse constant than (1 + ε) γ p
p−1
{fn}∞
basis of ℓ2. Let gn = fnp−1 signfn so that kgnk p
We now present a sketch of a proof of the complementation part for 1 < p < 2
. Let
n=1 be a normalized sequence in T X which is (1 + ε) equivalent to the unit vector
= 1 and hgn, fni = 1. Passing to
w→ 0, passing to another subsequence
n=1 are arbitrarily close to a biorthogonal
a subsequence we may assume gn
we may assume that {fn}∞
n=1 and {gn − g}∞
w→ g. Since fn
p−1
4
sequence; i.e.
hfn, gm − gi − δnm < ε
Xn,m
Note that 1 − ε ≤ kgn − gk ≤ 2 for all n (the lower bound follows from hfn, gn − gi ≥
1 − ε). We may also assume, by passing to a further subsequence, that gn − g is a
martingale difference sequence (or a block basis of the Haar system). Consequently,
for all coefficients {an}∞
n=1
p−1
kP an (gn − g)k p
where Hp is the unconditionality constant of the Haar basis and Bp = γ p
p−1
2 constant of L p
p−1
easily seen that P is a projection of norm ≤ 2KpBp (1 + ε).
n(cid:1)1/2
. Define now P : Lp → [fn] by P f = P∞
≤ HpE kP ±an (gn − g)k p
≤ 2HpBp(cid:0)P a2
p−1
is the type
n=1 hf, gn − gi fn then it is
4 Proof of the main result
Here we shall prove Theorem 1. We shall denote by {ei,j}∞
ℓp(ℓ2) i.e.,
i,j=1 the canonical basis of
1/p
a2
i,j!p/2
∞
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi,j=1
ai,jei,j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
∞
=
Xj=1 ∞
Xi=1
i,j=1 ⊆ R. By passing to a subsequence in each column of {ei,j}∞
for all {ai,j}∞
i,j=1, a
gliding hump argument (applied in the order (11) , (12) , (21) , (13) , (22) , (31) , ... of the
indices) and a simple perturbation argument, we can assume that for some infinite
subsequences Nj ⊆ N , {T ei,j}∞
is a block basis of the Haar system in Lp. By
that we mean that if the perturbed operator satisfies the conclusion of the theorem so
does the original one. Also, since {ei,j}∞
spans an isometric ℓp(ℓ2), and we are
anyhow interested only in a subspace of ℓp(ℓ2) isometric to ℓp(ℓ2), we may also assume
that Nj = N for all j; i.e., {T ei,j}∞
i,j=1 is a block basis of the Haar system. We shall
assume that from now on.
j=1,i∈Nj
j=1,i∈Nj
is equi-integrable.
are basically taken from [DJS]. We repeat the proofs for completeness.
Given a finite E ⊆ N and i ∈ N set vi (E) = S(cid:16)Pj∈E T ei,j(cid:17). The next two lemmas
Lemma 1. For all finite E ⊆ N and 1 < p < 2 the convex hull of (cid:8)v2
i (E)(cid:9)i∈N is p/2-
equi-integrable; i.e., the set of p/2 powers of functions in the convex hull of(cid:8)v2
i (E)(cid:9)i∈N
(cid:8)v2
i(cid:9)i∈N is not p/2-equi-integrable. Then, there exists ε0 > 0, a sequence {u2
disjoint convex blocks of (cid:8)v2
of N and Pi∈σj
Proof: Fix a finite E ⊆ N and write vi = vi (E). Assume that the convex hull of
j }j∈N of
i where σj are disjoint subsets
ujp > ε0 for all j ∈ N.
j } to be disjointly supported with respect
i=1 is
i,j ≤ 1) and disjoint subsets Bj such that RBj
i } follows from the easy fact that the convex hull of a finite subset of {v2
The fact that we can choose the sequence {u2
to {v2
p/2-equi-integrable.
i(cid:9)i∈N (i.e. u2
j =Pi∈σj
α2
i,jv2
i }∞
α2
5
Now, For all {am}∞
m=1 ∈ ℓ2, we obtain the following inequality, contradicting p < 2.
m=1 a2
(cid:0)P∞
m=1 a2
≥ E−1/p kT k−1 H −1
= E−1/p kT k−1 H −1
≥ E−1/p(cid:13)(cid:13)(cid:13)P∞
m=1 amPi∈σm αi,mPj∈E ei,j(cid:13)(cid:13)(cid:13)ℓp(ℓ2)
m(cid:1)p/2(cid:17)1/p
m(cid:1)1/2 = E−1/p(cid:16)Pj∈E(cid:0)P∞
p Ap(cid:13)(cid:13)(cid:13)
m=1 amPi∈σm αi,mPj∈E T ei,j(cid:17)(cid:13)(cid:13)(cid:13)p
S(cid:16)P∞
m (E)(cid:1)p/2 dµ(cid:17)1/p
p Ap(cid:16)R (cid:0)P∞
m (E) χBm(cid:1)p/2 dµ(cid:17)1/p
p Ap(cid:16)R (cid:0)P∞
m (E) dµ(cid:17)1/p
p Ap(cid:16)P∞
p Apε0 (P∞
= E−1/p kT k−1 H −1
≥ E−1/p kT k−1 H −1
up
m=1 amp)1/p .
m=1 ampRBm
≥ E−1/p kT k−1 H −1
m=1 a2
m=1 a2
mu2
mu2
Lemma 2. There are successive convex combinations νk (·) of (cid:8)v2
Λ (E) in Lp/2. Λ (E) is a L+
all finite E ⊆ N νk (E) →
k→∞
Λ, on the finite subsets on N.
p/2 additive valued measure,
i (·)(cid:9) such that for
nv2
n=1 α2
n(cid:1) :P α2
bounded in Lp/2, by a result of Nikishin [NI] for each ε > 0 there is a set D = Dε ⊂ [0, 1]
Proof: Case 1 (1 < p < 2): Let V = (cid:8)(cid:0)P∞
n ≤ 1(cid:9). Since V is
of measure larger than 1 − ε such that supv∈V RD vdµ < ∞. As in the proof of [JMST]
Lemma 6.4 (or see Proposition 5.2 in [DJS] for more details), we can find successive
convex combinations νk(·) of the v2
n(·) such that νk(E)1D converges pointwise and in L1
to Λ(E)1D for every finite E ⊂ N, where Λ (E) is L+
1 -valued).
By passing to a subsequence of the νk and a simple diagonal argument we can find,
for every εn → 0 a sequence of sets Dn with µ(Dn) > 1 − εn and such that νk(E)1Dn
converges, as k → ∞, pointwise and in L1 to Λ(E)1Dn for every finite E ⊂ N and every
n. In particular, νk(E) converges pointwise to Λ(E) for every finite E ⊂ N. It remains
to show that the convergence is also in Lp/2 (on the whole interval). Since for each E,
{νk (E)}k∈N is p/2-equi-integrable, it follows that, given any δ > 0, if n is large enough
0 -valued (and Λ1D is L+
n
νk(E)p/2dµ < δ for all k. Consequently, also RDc
RDc
lim supk→∞R νk(E) − Λ(E)p/2dµ ≤ lim supk→∞RDn
n
≤ lim supk→∞ k(νk(E) − Λ(E))1Dn kp/2
= 2δ.
νk(E) − Λ(E)p/2dµ + 2δ
1 + 2δ
Λ(E)p/2dµ ≤ δ and
Case 2 (2 < p < ∞):
1/2
T ei,j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p
p/2 = kvi (E)kp ≤ Hp(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xj∈E
(cid:13)(cid:13)v2
i (E)(cid:13)(cid:13)
i(cid:9)∞
Thus, there exists a subsequence of(cid:8)v2
i=1 such that Pi∈σk
ei,j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)ℓp(ℓ2)
≤ Hp kT k(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
i = 1 and Pi∈σk
i=1 that converges weakly in Lp/2 [0, 1] for every
E ∈ 2N. Denote the limit by Λ (E). By the reflexivity of Lp/2 [0, 1] and another diagonal
argument, there exists a sequence {σk}∞
k=1 of disjoint finite subsets of the integers and
non-negative numbers {αi}∞
i (E) → Λ (E)
= Hp kT k E1/p .
Xj∈E
α2
i v2
α2
6
as k → ∞ for all finite E where the convergence is in the Lp/2 [0, 1] norm. Set νk (·) =
i (·). νk is clearly additive on the finite subsets of N.
α2
i v2
It is clear that the sequence {Λ ({j})}∞
Pi∈σk
tor basis of ℓp/2; i.e., putting Λj = Λ ({j}), (cid:13)(cid:13)(cid:13)P a2
sequences of coefficients {aj}∞
stant to be arbitrarily close to 1 by blocking the Λj-s.
j Λj(cid:13)(cid:13)(cid:13)p/2
j=1. Next we would like to improve the equivalence con-
≈ (P ajp)2/p for all the
j=1 is positively equivalent to the unit vec-
Lemma 3. Let 1 < p < ∞, ε > 0 and εk ց 0. There are successive disjoint σk ⊆
N k = 1, 2, ... and coefficients {αj}∞
j Λj, for all
{ak}∞
α2
k=1,
akp!2/p
∞
Xk=1
∞
≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xk=1
In addition, there are disjoint sets {Bk}∞
a2
j=1 such that putting φk = Pj∈σk
akp!2/p
≤ (1 + ε) ∞
kφk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p/2
Xk=1
k(cid:13)(cid:13) ≤ εk.
k=1 such that (cid:13)(cid:13)φk1Bc
.
[AK], Lemma 5.2.8) we may
j=1 such that
→ 0.
Proof: By the subsequence splitting lemma (see e.g.
assume, passing to a subsequence, that there are disjoint sets {Aj}∞
jo∞
j=1
j 1Ac
is equi-integrable. For p > 2 this already implies that R Λp/2
nΛp/2
Otherwise, there is an R > 0 and δ > 0 such that R Λp/2
infinite J ⊆ N and it would follow that for j ∈ J R Λj ≥ δR1−p/2 and for all coefficients
{ai}, R (cid:16)P a2
, contradicting the
positive equivalence to the unit vector basis of ℓp/2. We can now take the σk-s to be
singletons and Bk-s to be some subsequence of the Aj-s.
j 1Λj ≤R ≥ δ for j ∈ J for some
≥ (cid:16)R P a2
2 )(cid:16)P a2
j Λj(cid:17)p/2
j Λj(cid:17)p/2
j(cid:17)p/2
≥ δp/2R
j 1Ac
2 (1− p
p
j
For 1 < p < 2, the proof is a bit more complicated. Set hj = Λj1Ac
j
and let δk ց 0
be a sequence to be determined later. By the equi-integrability, there exists an R > 0
such that
We also have,
n2/p
1
Z (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=1
1
n
n2/p
Z (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
j 1{hj≥R}dµ ≤
δp/2
1
2
.
dµ = Rp/2 1
n
np/2 ≤
δp/2
1
2
n
Xi=1Z hp/2
R(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=1
p/2
n
p/2
n
1
p/2
1
n
n2/p
dµ ≤
hj 1{hj≥R}(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=1
hj 1{hj<R}(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dµ <Z (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
n2/p Pn
p/2 dµ =R (cid:12)(cid:12)(cid:12)
n2/p Pn
1(cid:12)(cid:12)
R (cid:12)(cid:12)m11Bc
≤R (cid:12)(cid:12)(cid:12)
n2/p Pn
(n2−n1)2/p Pn2
Z (cid:12)(cid:12)m21Bc
2(cid:12)(cid:12)
1
1
1
7
p/2
dµ
p/2
j=1 Λj1Bc
j=1 Λj1Ac
dµ ≤ δp/2
1
.
1(cid:12)(cid:12)(cid:12)
j(cid:12)(cid:12)(cid:12)
p/2 dµ ≤ δp/2
.
2
for n sufficiently large. Set m1 = 1
j=1 Λj and B1 = ∪n
j=1Aj. Then
Similarly, we find m2 =
j=n1+1 Λj and B2 = ∪n2
j=n1+1Aj that satisfy
Continuing in this manner we find mi =
p/2 dµ ≤ δp/2
i(cid:12)(cid:12)
satisfying R (cid:12)(cid:12)mi1Bc
away from zero. Consequently, putting Φk = mk
choice of the δk-s, {Φk, Bk} satisfy the conclusion of the lemma.
kmkkp/2
i
1
(ni−ni−1)2/p Pni
j=ni−1+1 Λj and disjoint {Bj}
. The sequence {kmkkp/2} is bounded and bounded
we get that for an appropriate
Proof of Theorem 1 In the following let σk and αj be as in the statement of
l=1 be successive finite subsets of N and
<
Lemma 3. Let εl,k > 0 and for each k let {σl,k}∞
{βi,k}i∈σl,k
εl,k. Then
i,kν2
β2
β2
i (σk) − Λ (σk)(cid:13)(cid:13)(cid:13)p/2
α2
β2
i,kν2
coefficients such thatPi∈σl,k
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
j Xi∈σl,k
Xj∈σk
αjPi∈σl,k
Put fl,k =Pj∈σk
(cid:13)(cid:13)(cid:13)Pl,k al,kfl,k(cid:13)(cid:13)(cid:13)ℓp(ℓ2)
α2
j∈σk
i (j) − Xj∈σk
i,k = 1 and(cid:13)(cid:13)(cid:13)Pi∈σl,k
j Λj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p/2
≤(cid:18)max
l,kS2 (T (fl,k))(cid:13)(cid:13)(cid:13)
≈ (cid:13)(cid:13)(cid:13)Pl,k a2
= (cid:13)(cid:13)(cid:13)Pl,k a2
jPi∈σl,k
l,kPj∈σk
j Λj(cid:13)(cid:13)(cid:13)
≈ (cid:13)(cid:13)(cid:13)Pl,k a2
l,kPj∈σk
l,k(cid:17)p/2(cid:19)1/p
≈ (cid:18)P∞
k=1(cid:16)P∞
l=1 a2
α2
α2
1+ε
1+ε
p/2
p/2
1/2
1/2
βi,kei,j. Then, if the εl,k-s are small enough,
αjp(cid:19)2/p
εl,k.
i,kν2
β2
1/2
p/2
i (j)(cid:13)(cid:13)(cid:13)
.
This shows that {fl,k}∞,∞
l=1,k=1 is equivalent to the unit vector basis of ℓp(ℓ2). The
constant of the equivalence depends however on kT k ,(cid:13)(cid:13)T −1(cid:13)(cid:13) and p. We next correct
this: we can find, for each k, a block basis gu,k =Pl∈τu,k
kP∞
Since for each k gu,k is supported only on columns in σk, we get from that that {gu,k}u,k
is (1 + ε)-equivalent to the unit vector basis of ℓp(ℓ2).
γl,kfl,k, u = 1, 2, ..., such that
≈ (cid:0)P a2
u=1 augu,kkℓp(ℓ2)
u(cid:1)1/2.
1+ε
Next we would like to apply a similar stabilization procedure to {T (gu,k)}u,k. This
of course is more complicated since this sequence does not lie in ℓp(ℓ2) any more.
Note that, by passing to further subsequences we may assume that for some positive
constant b (depending on T ), and for all k and {an} with P∞
n=1 a2
n = 1
where εk is the preassigned sequence appearing in Lemma 3. In particular,
< εk
S2 ∞
Xn=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(S2 (P∞
≤ εp/2
n=1 anT gn,k) 1Bc
anT gn,k! − bφk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p/2
k(cid:13)(cid:13)
p/2
p/2
k + bp/2εp/2
n=1 anT gn,k) − bφk)1Bc
k(cid:13)(cid:13)
This shows that any sequence of the form {S(P∞
P∞
n=1 a2
k
.
(cid:13)(cid:13)S2 (P∞
8
p/2
p/2
p/2
p/2
+ bp/2(cid:13)(cid:13)φk1Bc
k(cid:13)(cid:13)
n=1 an,kT gn,k)}∞
k=1 with, say,
n,k = 1 for all k is essentially (with respect to the Lp norm) disjointly supported.
We would like to have a similar statement with the S removed. This is where the Haar
functions (rather than some other unconditional basis of Lp) play a role. First we may
assume that the sets Bk are each a finite union of dyadic intervals. Next notice that
for each k if l is large enough then T fl,k is a dyadic simple function such that each
Haar function appearing in its expansion (with non zero coefficient) has support which
is either contained or disjoint of Bk. Consequently,
S (T fl,k1Bk ) = S (T fl,k) 1Bk and S(cid:0)T fl,k1Bc
k(cid:1) = S (T fl,k) 1Bc
k
In particular, this implies that for some constant Kp (depending on p only), for all k
and all {an}∞
n=1,
.
(1)
n=1 anT gn,k) 1Bc
(cid:13)(cid:13)(P∞
k(cid:13)(cid:13)p
n=1 anT gn,k) 1Bc
≤ Kp(cid:13)(cid:13)S2 (P∞
≤ Kp(cid:0)1 + bp/2(cid:1)1/p
1/2
p/2
n=1 a2
k(cid:13)(cid:13)
n(cid:1)1/2 .
ε1/2
k
(cid:0)P∞
We thus get that if εk are chosen small enough, the two sequences {T gn,k1Bk }∞
and {T gn,k}∞
[(T gn,k) 1Bk ]∞
show that for some subspace X ⊂ ℓp(ℓ2), spanned by blocks {un,k}∞
n,k=1 are small perturbations one of the other. Define T : ℓp(ℓ2) →
n,k=1 by T en,k = (T gn,k) 1Bk . The perturbation above is such that if we
n,k=1 of {en,k}∞
1-equivalent to the unit vector basis of ℓp(ℓ2), and such that n T un,ko∞
(1 + ε)-equivalent to a multiple of the unit vector basis of ℓp(ℓ2) and h T un,ki∞
(1 + ε) γp/(p−1)-complemented in Lp, then the same (with 1 + 2ε replacing 1 + ε) holds
for {T un,k}∞
n,k=1
n,k=1
n,k=1
is
is
n,k=1
n,k=1.
The way we choose un,k is similar to the way we have chosen gn,k: since for each
k {T gn,k}∞
k=1 is equivalent to the unit vector basis of ℓ2 we can find blocks {un,k}
of {gn,k} such that {(T un,k) 1Bk } (which are blocks of {(T gn,k) 1Bk }∞
n,k=1) are (1 + ε)
equivalent to the unit vector basis of ℓ2 and (1 + ε) γp complemented in Lp. Call the
projection Pk. Note also that we may assume that the ℓ2 norm of the coefficients of
{T un,k} relative to {T fn,k} all differ by a multiplicative constant of at most 1 + ε.
(This is important since we want TX to be a (1 + ε)-isomorphism). Finally, define
P : Lp → [(T un,k) 1Bk ] by P f =P Pk (f 1Bk ).
5 Concluding remarks
1. We first remark that the constants γp and γp/(p−1) that appear in the statement of
Theorem 1 (and also in the theorem in section 3) are best possible. Actually, these are
lower bounds on the norm of the best projection onto an (isometric) ℓ2 subspace of Lp.
This was proved in [GLR].
2. If one wants to avoid the use of the material in section 3 at the price of getting
worse constants, still depending only on p, one can use Theorem 3.1 in [PR] instead.
3. As we already remarked, it would be nice if somebody comes up with a simpler
proof of the theorem of section 3. Maybe along the lines of [Als].
4. Since ℓ2 is not isomorphic to a complemented subspace of L1 there is of course
no complete analogue of Theorem 1 for L1. However, the weaker statement that any
isomorphism T : ℓ1(ℓ2) → L1 stabilizes; i.e., is a (1 + ε)-isomorphism when restricted
9
to some subspace of ℓ1(ℓ2) (1 + ε)-isomorphic to ℓ1(ℓ2) , is still possible. Of course
the proof as is written above makes an heavy use of the unconditionality of the Haar
system and thus cannot be used. However, note that for most of the proof we could
replace the Haar system with any unconditional basic sequence containing T (ℓ1(ℓ2)),
even with constant depending on T . So we could use T ei,j as such a sequence. The
problem is in (1) where we use the "eventual commutativity" of the square function
operation and the restriction to a given dyadic set. This seems like a not very essential
use of the Haar system but we couldn't overcome it easily and we leave it for future
research.
References
[Ald] D. J. Aldous, Subspaces of L1, via Random Measures, Trans. of Amer. Math.
Soc., 267, No. 2 (Oct., 1981), pp. 445-463.
[Als] D. E. Alspach, Good ℓ2-subspaces of Lp, p > 2, Banach J. Math. Anal. 3 (2009),
no. 2, 49 -- 54
[AK] F. Albiac and N.J. Kalton, Topics in Banach space theory, Graduate Texts in
Mathematics, 233, Springer, New York, 2006.
[DJS] D. Dosev, W.B. Johnson and G. Schechtman, Comutators on Lp, 1 ≤ p < ∞,
http://arXiv.org/abs/1102.0137
[ES] P. Enflo and T. Starbird, subspaces of L1 containing L1, Studia Math. 65(1979),
203-225.
[GLR] Y. Gordon, D. R. Lewis and J. R. Retherford, Banach ideals of operators with
applications, J. Functional Analysis 14 (1973), 85129.
[HOS] R. Haydon, E. Odell and T. Schlumprecht, Small Subspaces of Lp, Annals of
Math., to appear.
[JMST] W.B. Johnson, B. Maurey, G. Shechtman and L. Tzafriri, Symmetric struc-
tures in Banach spaces, Memoirs Amer. Math. Soc., 217 (1979).
[KM] J.L. Krivine and B. Maurey, Espaces de Banach stables, (French), Israel J. Math.
39 (1981), no. 4, 273-295.
[KP] M. I. Kadec and A. Pelczynski, Bases, lacunary sequences and complemented
subspaces in the spaces Lp, Studia Math. 21 (1962), pp. 161-176.
[LT-I] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces I, Sequence spaces,
Springer-Verlag, Berlin-New York, 1977.
[LT-II] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces II, Function spaces,
Springer-Verlag, New York, 1979.
[NI] E.M. Nikishin, Resonance theorems and superlinear operators, Russ. Math. Surv.
25, no. 6, (1970), pp. 125-187.
[PR] A. Pelczynski and H. P. Rosenthal, Localization techniques in Lp spaces, Studia
Math. 52, (1975), pp. 263 -- 289.
[R] H. P. Rosenthal, Embeddings of L1 in L1, Contemporary Math. 29(1984), 335-349.
10
R. Levy
Department of Mathematics
Weizmann Institute of Science Weizmann Institute of Science
Rehovot, Israel
[email protected]
G. Schechtman
Department of Mathematics
Rehovot, Israel
[email protected]
R. Levy current address:
IBM Research Lab
Haifa, Israel
[email protected]
11
|
1210.3696 | 1 | 1210 | 2012-10-13T10:55:15 | Szlenk and $w^\ast$-dentability indices of the Banach spaces $C([0,\alpha])$ | [
"math.FA"
] | Let $\alpha$ be an infinite ordinal and $\gamma$ the unique ordinal satisfying $\omega^{\omega^\gamma}\leq \alpha < \omega^{\omega^{\gamma+1}}$. We show that the Banach space $C([0,\,\alpha])$ of all continuous scalar-valued functions on the compact ordinal interval $[0,\,\alpha]$ has Szlenk index equal to $\omega^{\gamma+1}$ and $w^\ast$-dentability index equal to $\omega^{1+\gamma+1}$. | math.FA | math |
Szlenk and w∗-dentability indices of the Banach
spaces C([0, α])
Philip A.H. Brooker∗†
February 24, 2012
Abstract
Let α be an infinite ordinal and γ the unique ordinal satisfying ωωγ ≤
α < ωωγ+1
. We show that the Banach space C([0, α]) of all continuous
scalar-valued functions on the compact ordinal interval [0, α] has Szlenk
index equal to ωγ+1 and w∗-dentability index equal to ω1+γ+1.
1
Introduction
The Szlenk index is an ordinal-valued isomorphic invariant of a Banach space that
was introduced in [21]. There it is used to show that the class of separable, reflexive
Banach spaces contains no universal element, thereby solving a problem posed by
Banach and Mazur in the Scottish Book. Since then the Szlenk index has found
a variety of uses in the study of Banach space geometry, a survey of which can be
found in [14]. One of the main applications of the Szlenk index is in the study of
C(K) spaces and their operators, as witnessed in particular by the work of Alspach
[2], Alspach and Benyamini [1], Benyamini [3], Bourgain [5] and Gasparis [8]; we
refer to the survey article [17] for a detailed discussion of this topic. The purpose
of the current paper is to enlarge the class of C(K) spaces for which the Szlenk
index of C(K) is known. We shall also discuss the related w∗-dentability index for
the same class of C(K) spaces.
It is a classical result of Mazurkiewicz and Sierpinski [15] that every countable,
compact Hausdorff space is homeomorphic to an ordinal interval [0, α] equipped
∗The author gratefully acknowledges the financial support of a Lift-Off Fellowship from the
Australian Mathematical Society. This research was undertaken whilst the author was a Visiting
Fellow at the Mathematical Sciences Institute of the Australian National University.
†E-mail address: [email protected]
1
with its order topology, for some α < ω1. The linear isomorphic classification of
C(K) spaces with K countable is due to Bessaga and Pe lczy´nski [4], who showed
that for ordinals ω ≤ α < β < ω1, C([0, α]) is isomorphic to C([0, β]) if and
only if β < αω. In particular, it follows that each C(K) space with K countable
is isomorphic to the space C([0, ωωγ ]) for a unique countable ordinal γ. The
computation of the Szlenk indices of the Banach spaces C(K) with K countable
is due to Samuel [19]; drawing upon deep results of Alspach and Benyamini [1],
Samuel showed that the Szlenk index of C([0, ωωγ ]) is ωγ+1 for each countable
ordinal γ.
The first extension of Samuel's result was achieved by Lancien [13], who used
Samuel's result and a separable-reduction argument to show that if K is a (scat-
tered) compact Hausdorff space of countable height, then the Szlenk index of C(K)
is ωγ+1, where γ is the unique ordinal such that the height of K belongs to the
ordinal interval [ωγ, ωγ+1). H´ajek and Lancien later gave in [9] a 'direct' proof
of Samuel's result, the existence of which was conjectured by Rosenthal in [17];
in particular, they computed the Szlenk indices of the Banach spaces C([0, α])
for ordinals α < ω1ω without appeal to Alspach and Benyamini's results from [1].
Their result says that for ω ≤ α < ω1ω the Szlenk index of C([0, α]) is ωγ+1, where
γ is the unique ordinal satisfying ωωγ ≤ α < ωωγ+1. As the Szlenk index of c0(X)
coincides with the Szlenk index of X for every infinite dimensional Banach space
X, it follows easily that the statement of H´ajek and Lancien's result holds in fact
for all ordinals α < ω1ωω (see [6, p.2232] for details.)
The main purpose of the current paper is to determine the Szlenk index of
C([0, α]) for α an arbitrary ordinal. In particular, we shall extend the previous
results of Samuel and H´ajek-Lancien, showing that for α ≥ ω the Szlenk index
of C([0, α]) is ωγ+1, where γ is the unique ordinal satisfying ωωγ ≤ α < ωωγ+1
(Theorem 2.6). The computation by H´ajek and Lancien of Szlenk indices of the
spaces C([0, α]), α < ω1, makes use of the aforementioned isomorphic classification
of the spaces C([0, α]), α < ω1, by Bessaga and Pe lczy´nski. That the statement
of the Bessaga-Pe lczy´nski classification theorem does not hold in general for the
spaces C([0, α]) when α ≥ ω1 is the reason that the argument of H´ajek and Lan-
cien does not yield the Szlenk indices of the spaces C([0, α]) for arbitrary α. In
the current paper we avoid an appeal to the Bessaga-Pe lczy´nski theorem, work-
ing instead through decompositions of the spaces C([0, α]) into c0-direct sums of
smaller spaces of continuous functions on compact ordinals (cf. Lemma 2.2) and
isomorphisms of such c0-direct sums (cf. Lemma 2.3).
In Section 3 we shall outline how the techniques developed in [10] can be com-
bined with the arguments used in the proof of Theorem 2.6 of the current paper
to show that for α and γ as in the preceding paragraph, the w∗-dentability index
of C([0, α]) is ω1+γ+1 (Theorem 3.2).
2
We now detail most of the necessary terminology and background results for the
current paper. As usual, ω denotes the first infinite ordinal and ω1 denotes the first
uncountable ordinal. For K a compact Hausdorff space, C(K) is the Banach space
of continuous scalar-valued functions on K, equipped with the supremum norm.
For α an ordinal, the ordinal interval [0, α] is a compact Hausdorff space when
equipped with its order topology. We denote by C0([0, α]) the closed subspace of
C([0, α]) consisting of all elements of C([0, α]) that vanish at α. It is well-known
and easy to show that C0([0, α]) is isomorphic to C([0, α]) whenever α ≥ ω. For
ordinals ξ ≤ α and f ∈ C0([0, ξ]), we define fξ, α ∈ C0([0, α]) by setting
fξ, α(ζ) =(f (ζ)
0
if ζ ≤ ξ,
if ζ > ξ,
0 ≤ ζ ≤ α .
It is clear that the operator Jξ, α : C0([0, ξ]) −→ C0([0, α]) : f 7→ fξ, α is an
isometric linear embedding of C0([0, ξ]) into C0([0, α]).
For a Banach space X we write BX for the set {x ∈ X kxk ≤ 1}. If Y is a
Banach space that is isomorphic to X, we write X ≈ Y . If S is a nonempty set,
c0(S, X) is defined to be the linear space
{f : S −→ X {s ∈ S kf (s)k > ǫ} is finite for every ǫ > 0}
which we equip with the complete norm kf k := sup{kf (s)k s ∈ S}. For a
nonempty subset R ⊆ S, we denote by UR the canonical isometric linear embedding
of c0(R, X) into c0(S, X). The dual space c0(S, X)∗ is naturally identified via
isometric linear isomorphism with the Banach space ℓ1(S, X ∗).
The Szlenk index is defined as follows. Let X be an Asplund space and B ⊆ X ∗.
Define
sǫ(B) = {x ∈ B diam(B ∩ V ) > ǫ for every w∗-open V ∋ x} .
ǫ (B) = B, sβ+1
ǫ
(B) = sǫ(sβ
ǫ (B)) for each
ǫ (B) =Tσ<β sσ
We iterate sǫ transfinitely as follows: s0
ordinal β and sβ
ǫ (B) whenever β is a limit ordinal.
The ǫ-Szlenk index of X, denoted Sz(X, ǫ), is the first ordinal β such that
sβ
ǫ (BX ∗) = ∅. The Szlenk index of X is the ordinal Sz(X) := supǫ>0 Sz(X, ǫ).
Note that Sz(X, ǫ) exists for every Asplund space X and ǫ > 0 by following
well-known characterisation of Asplund spaces: a Banach space is Asplund if and
only if every bounded subset of its dual admits w∗-open slices of arbitrarily small
norm diameter [7, Theorem 5.2]. The ordinal index Sz(X) is thus defined for
every Asplund space X, and the definition cannot be extended beyond the class
of Asplund spaces. It is worth noting that the definition of the Szlenk index used
in the current paper (and many others) differs from that introduced by Szlenk in
3
[21], however the two definitions give the same index on separable Banach spaces
containing no copy of ℓ1.
The following proposition collects some basic facts regarding the Szlenk index.
Proposition 1.1. Let X and Y be Asplund spaces.
(i) If X is isomorphic to a subspace of Y , then Sz(X) ≤ Sz(Y ). In particular,
the Szlenk index is an isomorphic invariant of an Asplund space.
(ii) If γ is an ordinal and ǫ > 0 is such that Sz(X, ǫ) > ωγ, then Sz(X) ≥ ωγ+1.
It follows that Sz(X) = ωα for some ordinal α.
(iii) Sz(X) = 1 if and only dim(X) < ∞.
Details of the proofs of assertions (i) and (ii) of Proposition 1.1 can be found
in [11, §2.4]. Verification of (iii) is elementary.
The characterisation of those compact Hausdorff spaces K for which C(K)
is an Asplund space is due to Namioka and Phelps; they showed in [16] that
a Banach space C(K) is Asplund if and only if K is scattered. As every ordinal
interval [0, α] is scattered and compact when equipped with its order topology, the
spaces C([0, α]) are Asplund spaces and their Szlenk index is defined. Information
regarding topological properties of ordinals can be found in, e.g., [20, §8.6].
Important to our analysis of the spaces C([0, α]) and their duals is the fact that
for a scattered, compact Hausdorff space K, the dual space C(K)∗ is naturally
identified with ℓ1(K); this is due to Rudin [18, Theorem 6]. The dual of C0([0, α])
is naturally identified with ℓ1([0, α)).
2 The Szlenk index of C([0, α])
We begin our computations of Szlenk indices by gathering some preliminary results
that we shall need. The first such result is the following proposition that provides
a way to obtain an upper estimate of the Szlenk index of a Banach space.
Proposition 2.1 ([9]). Let X be a Banach space and η an ordinal. Assume that
∀ǫ > 0 ∃δ(ǫ) ∈ (0, 1)
sη
ǫ (BX ∗) ⊆ (1 − δ(ǫ))BX ∗ .
Then
Sz(X) ≤ ηω .
Lemma 2.2. Let ξ and ζ be ordinals satisfying 0 < ζ ≤ ξ and ω ≤ ξ. Then
C0([0, ξζ]) ≈ C0([0, ζ]) ⊕ c0(ζ, C0([0, ξ])).
4
Lemma 2.2 is essentially noted by Bessaga and Pe lczy´nski in their proof of [4,
2.4]; for the sake of completeness, we give here the details of their sketch proof.
Proof. We may write C0([0, ξζ]) = Y ⊕ Z, where Y consists of all elements of
C0([0, ξζ]) that are constant on the intervals (ξσ, ξ(σ + 1)], 0 ≤ σ < ζ, and Z
consists of all elements of C0([0, ξζ]) vanishing at the points ξσ, 1 ≤ σ ≤ ζ. The
lemma then follows from the routine observation that Y and Z are isometrically
isomorphic to C0([0, ζ]) and c0(ζ, C0([0, ξ])) respectively.
We have the following consequence of Lemma 2.2.
Lemma 2.3. Let γ be an ordinal and 1 < n < ω. Then
C0([0, ωωγn]) ≈ c0(ωωγ
, C0([0, ωωγ
])) .
Proof. We proceed via induction on n. For the case n = 2, note that an application
of Lemma 2.2 with ξ = ζ = ωωγ yields
C0([0, ωωγ 2]) ≈ C0([0, ωωγ
]) ⊕ c0(ωωγ
, C0([0, ωωγ
])) ≈ c0(ωωγ
, C0([0, ωωγ
])) ,
as desired. Similarly, if C0([0, ωωγ n]) ≈ c0(ωωγ , C0([0, ωωγ ])), then applying Lemma 2.2
with ζ = ωωγ and ξ = ωωγn yields
C0([0, ωωγ(n+1)]) ≈ C0([0, ωωγ
≈ C0([0, ωωγ
≈ C0([0, ωωγ
≈ c0(ωωγ
]) ⊕ c0(ωωγ
]) ⊕ c0(ωωγ
]) ⊕ c0(ωωγ
])) ,
, C0([0, ωωγ
, C0([0, ωωγ n]))
, c0(ωωγ
, C0([0, ωωγ
]))
, C0([0, ωωγ
])))
which completes the proof.
The last of the preliminary results that we shall require is the following gener-
alisation of [9, Lemma 3.3].
Lemma 2.4. Let α, β and ξ be ordinals such that ξ < α, let S be a set, ∅ ( R ⊆ S
ξ, αzrk > 1 − ǫ, then
and ǫ > 0.
(J ∗
3ǫ(Bc0(S, C0([0, α]))∗) and Pr∈R kJ ∗
If (zs)s∈S ∈ sβ
ǫ (Bc0(R, C0([0, ξ]))∗).
ξ, αzr)r∈R ∈ sβ
Proof. We proceed via transfinite induction on β. The assertion of the lemma
is clearly true for β = 0. Suppose that σ is an ordinal such that the assertion
of the lemma holds for all β ≤ σ; we will show that the lemma holds also for
ξ, αzrk > 1 − ǫ
(Bc0(R, C0([0, ξ]))∗). Since we intend to show that (zs)s∈S /∈
3ǫ(Bc0(S, C0([0, α]))∗), hence
β = σ + 1. Let (zs)s∈S ∈ Bc0(S, C0([0, α]))∗ be such that Pr∈R kJ ∗
and (J ∗
sσ+1
3ǫ (Bc0(S, C0([0, α]))∗), we may assume that (zs)s∈S ∈ sσ
ξ, αzr)r∈R /∈ sσ+1
ǫ
5
ξ, αzr)r∈R ∈ sσ
(J ∗
open subset V of c0(R, C0([0, ξ]))∗ containing (J ∗
sσ
ǫ (Bc0(R, C0([0, ξ]))∗) by the induction hypothesis. So there is a w∗-
ξ, αzr)r∈R and such that diam(V ∩
ξ, αzrk > 1 − ǫ, we may assume that
ǫ (Bc0(R, C0([0, ξ]))∗)) ≤ ǫ. Since Pr∈R kJ ∗
V ∩ (1 − ǫ)Bc0(R, C0([0, ξ]))∗ = ∅ .
(2.1)
Define
J : c0(R, C0([0, ξ])) −→ c0(R, C0([0, α])) : (xr)r∈R 7→ (Jξ, αxr)r∈R ,
R)−1(V ), so that W is a w∗-open set containing (zs)s∈S, and let
ξ,αurk > 1 − ǫ by (2.1) and
so that URJ is an isometric linear embedding of c0(R, C0([0, ξ])) into c0(S, C0([0, α])).
Let W = (J ∗U ∗
(us)s∈S ∈ W ∩ sσ
(us)s∈S ∈ sσ
by the induction hypothesis. Suppose (us)s∈S, (vs)s∈S ∈ W ∩ sσ
Then
3ǫ(Bc0(S, C0([0, α]))∗). Then Pr∈R kJ ∗
3ǫ(Bc0(S, C0([0, α]))∗) by assumption, hence J ∗U ∗
3ǫ(Bc0(S, C0([0, α]))∗).
ǫ (Bc0(R, C0([0, ξ]))∗)
R(us)s∈S ∈ sσ
kJ ∗U ∗
R(us)s∈S − J ∗U ∗
R(vs)s∈Sk ≤ diam(V ∩ sσ
ǫ (Bc0(R, C0([0, ξ]))∗)) ≤ ǫ .
Moreover, since kJ ∗U ∗
R(us)s∈Sk > 1 − ǫ, we have
Xs∈S\R
Xs∈S\R
kusk +Xr∈R
kvsk +Xr∈R
kur[ξ, α)k < ǫ ,
kvr[ξ, α)k < ǫ .
and similarly,
It follows that
k(us)s∈S − (vs)s∈Sk ≤ kJ ∗U ∗
R(us)s∈S − J ∗U ∗
R(vs)s∈Sk + Xs∈S\R
kus − vsk +Xr∈R
k(ur − vr)[ξ, α)k
≤ ǫ + ǫ + ǫ = 3ǫ .
In particular, diam(W ∩sσ
We have now shown that the assertion of the lemma passes to successor ordinals.
As the assertion of the lemma passes readily to limit ordinals, the proof is
3ǫ(Bc0(S, C0([0, α]))∗)) ≤ 3ǫ, hence (zs)s∈S /∈ sσ+1
3ǫ (Bc0(S, C0([0, α]))∗).
complete.
We are now ready to determine upper estimates for the Szlenk indices of the
Banach spaces C0([0, ωωγ ]), where γ is an arbitrary ordinal.
Proposition 2.5. Let γ be an ordinal and 0 < n < ω. Then
Sz(c0(ωωγ
, C0([0, ωωγn]))) ≤ ωγ+1 .
6
Proof. We proceed by induction on γ, first establishing the proposition in the case
γ = 0 and n = 1. By Proposition 2.1, it suffices to show that
∀ǫ > 0 sǫ(Bc0(ω, C0([0, ω]))∗) ⊆(cid:16)1 −
3(cid:17)Bc0(ω, C0([0, ω]))∗ .
ǫ
Suppose by way of contraposition that there is ǫ > 0 and (zl)l<ω ∈ sǫ(Bc0(ω, C0([0, ω]))∗)
such that k(zl)l<ωk > 1 − ǫ/3. Since
k(zl)l<ωk = supnXr∈R
0 < m < ω, R ⊆ ω, 0 < R < ∞o ,
there exists m < ω and a nonempty finite set R ⊆ ω such that
kJ ∗
m, ωzrk (cid:12)(cid:12)(cid:12)
Xr∈R
kJ ∗
m, ωzrk > 1 −
ǫ
3
.
By Lemma 2.4, this implies that (J ∗
1. By Proposition 1.1(iii), this in turn implies that c0(R, C0([0, m])) is infinite di-
mensional; however, this is impossible since dim(c0(R, C0([0, m]))) = mR < ∞.
With this contradiction we have now established the assertion of the proposition
for γ = 0 and n = 1.
m, ωzr)r∈R ∈ sǫ/3(Bc0(R, C0([0, m]))∗), hence Sz(c0(R, C0([0, m]))) >
Next we show that if β is an ordinal such that the assertion of the proposition
holds for γ = β and n = 1, then the proposition is true for γ = β and all 0 < n < ω.
Let 1 < m < ω and note that, by Lemma 2.3, for any ordinal β we have
c0(ωωβ
, C0([0, ωωβm])) ≈ c0(ωωβ
, c0(ωωβ
, C0([0, ωωβ
]))) ≈ c0(ωωβ
, C0([0, ωωβ
])) .
Assuming the proposition is true for n = 1 and γ = β, we deduce that
Sz(c0(ωωβ
, C0([0, ωωβm]))) = Sz(c0(ωωβ
, C0([0, ωωβ
]))) ≤ ωβ+1 ,
as desired.
It now remains to show that if β is an ordinal such that the assertion of the
proposition holds for all γ < β and 0 < n < ω, then the assertion of the proposition
holds for n = 1 and γ = β. Take such β and note that, by Proposition 2.1, it suffices
to show that
∀ǫ > 0 sωβ
ǫ (Bc0(ωωβ , C0([0, ωωβ ]))∗) ⊆(cid:16)1 −
3(cid:17)Bc0(ωωβ , C0([0, ωωβ ]))∗ .
ǫ
(2.2)
Suppose by way of contraposition that there is ǫ > 0 and (zη)η<ωωβ ∈ sωβ
with k(zη)η<ωωβ k > 1 − ǫ/3. Since
ǫ (Bc0(ωωβ , C0([0, ωωβ ]))∗)
k(zη)η<ωωβ k = supnXr∈R
kJ ∗
ωωζ m, ωωβ zrk (cid:12)(cid:12)(cid:12)
ζ < β, 0 < m < ω, R ⊆ ωωβ
, 0 < R < ∞o ,
7
there exists ζ < β, 0 < m < ω and a nonempty finite set R ⊆ ωωβ such that
kJ ∗
ωωζ m, ωωβ zrk > 1 − ǫ/3 .
Xr∈R
By Lemma 2.4, this implies that (J ∗
Sz(c0(R, C0([0, ωωζm]))) > ωβ. By the induction hypothesis, it follows that
ωωζ m, ωωβ zr)r∈R ∈ sωβ
ǫ/3(Bc0(R, C0([0, ωωζ m]))∗), hence
ωβ < Sz(c0(R, C0([0, ωωζm]))) ≤ ωζ+1 ≤ ωβ ,
which is absurd. Thus (2.2) holds, and the assertion of the proposition holds for
n = 1 and γ = β. The inductive proof is now complete.
Theorem 2.6. Let α ≥ ω and let γ be the unique ordinal satisfying ωωγ ≤ α <
ωωγ+1. Then
Sz(C([0, α])) = ωγ+1 .
Proof. Let n < ω be such that ωωγn > α, so that C([0, α]) is isomorphic to a
subspace of C0([0, ωωγ n]), hence isomorphic to a subspace of c0(ωωγ , C0([0, ωωγ n])).
Then, by Proposition 1.1(i) and Proposition 2.5,
Sz(C([0, α])) ≤ Sz(c0(ωωγ
, C0([0, ωωγn]))) ≤ ωγ+1.
(2.3)
To obtain the reverse inequality, we consider the functionals δξ ∈ BC([0, α])∗, ξ ≤ α,
where hδξ, f i = f (ξ) for each f ∈ C([0, α]). As the map [0, α] −→ C([0, α])∗ is an
order-w∗ homeomorphic embedding, a straightforward induction shows that δωζ ∈
sζ
1(BC([0, α])∗) whenever ωζ ≤ α. In particular, sωγ
1 (BC([0, α])∗) ∋ δωωγ is nonempty,
hence Sz(C([0, α])) > ωγ. By Proposition 1.1(ii), Sz(C([0, α])) ≥ ωγ+1, and we
are done.
3 The w∗-dentability index of C([0, α])
In this section we discuss the w∗-dentability indices of the spaces C([0, α]), where
α is an arbitrary ordinal. For a (real) Asplund space X, the definitions of the ǫ-
w∗-dentability index Dz(X, ǫ) of X and the w∗-dentability index Dz(X) of X are
essentially the same as for the Szlenk indices Sz(X, ǫ) and Sz(X), the difference
being that in the derivation on w∗-compact sets we remove only w∗-slices of small
norm diameter (for x ∈ X and t ∈ R, let H(x, t) = {x∗ ∈ X ∗ x∗(x) > t}; for
B ⊆ X ∗, a w∗-slice of B is any set of the form H(x, t) ∩ B, where x ∈ X and
t ∈ R.) To be precise, let X be an Asplund space and B ⊆ X ∗. Define
dǫ(B) = {x∗ ∈ B diam(B ∩ H(x, t)) > ǫ for every w∗-slice H(x, t) ∋ x∗} .
8
ǫ (B) = B, dβ+1
ǫ
(B) = dǫ(dβ
ǫ (B)) for each
ǫ (B) whenever β is a limit ordinal.
We iterate dǫ transfinitely, setting d0
ordinal β and dβ
ǫ (B) =Tσ<β dσ
Define Dz(X, ǫ) to be the first ordinal β such that dβ
ǫ (BX ∗) = ∅, and Dz(X) :=
supǫ>0 Dz(X, ǫ). Similarly to the Szlenk index, the w∗-dentability index Dz(X) is
defined for every Asplund space X.
The natural analogues of parts (i) and (ii) of Proposition 1.1 hold for the w∗-
dentability index, with similar proofs. In particular, Dz(X) ≤ Dz(Y ) whenever
X is a subspace of Y , and Dz(X) > ωγ implies Dz(X) ≥ ωγ+1. For part (iii), the
analogous result for the w∗-dentability index is that Dz(X) ≤ ω if and only if X
is superreflexive; this is due to Lancien [12]. Moreover, it is clear that Sz(X) ≤
Dz(X) for all Asplund spaces X; conversely, we have the following:
Proposition 3.1 ([14]). Let X be an Asplund space and L2(X) the Banach space of
all (equivalence classes of ) Bochner integrable functions f : [0, 1] −→ X, equipped
with its usual norm. Then
Dz(X) ≤ Sz(L2(X)) .
Proposition 3.1 was used in [10] to show that for ordinals ωωγ ≤ α < ωωγ+1 <
ω1, the w∗-dentability index of C([0, α]) is ω1+γ+1. The authors of [10] then ex-
tend their result to a certain nonseparable setting by showing that for a scattered
compact Hausdorff space K of countable height, the w∗-dentability index of C(K)
is ω1+γ+1, where γ is the unique (countable) ordinal such that the height of K
belongs to the ordinal interval [ωγ, ωγ+1). The following result extends the main
result of [10] in a different direction.
Theorem 3.2. Let α ≥ ω and let γ be the unique ordinal satisfying ωωγ ≤ α <
ωωγ+1. Then
Dz(C([0, α])) = ω1+γ+1 .
We shall only sketch the proof of Theorem 3.2, as the differences between the
proofs of Theorem 2.6 and Theorem 3.2 are completely analogous to the differences
between the proofs of the separable cases established in [9] and [10] (we note that
although it is essentially possible to prove Theorem 2.6 and Theorem 3.2 simul-
taneously by estimating the Szlenk index of L2(µ, C([0, α])), where µ is assumed
to be either counting measure on a singleton or Lebesgue measure on [0, 1], re-
spectively, we feel it would obscure the main ideas of the proof of Theorem 2.6 to
do so). Theorem 3.2 follows readily from the Szlenk index estimate given by the
following result.
Proposition 3.3. Let γ be an ordinal and 0 < n < ω. Then
Sz(L2(c0(ωωγ
, C([0, ωωγn])))) ≤ ω1+γ+1 .
9
The main difficulty in establishing Proposition 3.3 is to prove the following vari-
ant of Proposition 2.4; the proof combines ideas from the proofs of Proposition 3.3
and [10, Lemma 6]
Lemma 3.4. Let 0 < n < ω and let ζ and γ be ordinals satisfying either ζ = γ = 0
or ωωζn < ωωγ . Let ∅ ( R ⊆ ωωγ , ǫ > 0 and let J denote the canonical embedding of
L2(c0(R, C([0, ωωζn]))) into L2(c0(ωωγ , C([0, ωωγ ]))). Let β be an ordinal. If z ∈
sβ
3ǫ(BL2(c0(ωωγ , C([0, ωωγ ])))∗ ) and kJ ∗zk2 > 1−ǫ2, then J ∗z ∈ sβ
ǫ (BL2(c0(R, C([0, ωωζ n])))∗).
The estimate Dz(C([0, α])) ≤ ω1+γ+1 follows readily from Proposition 3.1 and
Proposition 3.3. For the reverse inequality, note that for the case γ ≥ ω we have
Dz(C([0, α])) ≥ Sz(C([0, α])) > Sz(C([0, α]), 1) ≥ ωγ = ω1+γ ,
so that the required estimate follows by the aforementioned w∗-dentability index
version of Proposition 1.1(ii). The case γ < ω follows from the fact established in
[10, Proposition 11] that Dz(C([0, ωωγ ]), 1/2) > ω1+γ for every γ < ω.
References
[1] D.E. Alspach and Y. Benyamini, C(K) quotients of separable L∞ spaces.
Israel J. Math. 32 (1979) 145 -- 160.
[2] D.E. Alspach, Quotients of C[0, 1] with separable dual. Israel J. Math. 29
(1978) 361 -- 384.
[3] Y. Benyamini, An extension theorem for separable Banach spaces. Israel J.
Math. 29 (1978) 24 -- 30.
[4] C. Bessaga and A. Pe lczy´nski, Spaces of continuous functions. IV. On isomor-
phical classification of spaces of continuous functions. Studia Math. 19 (1960)
53 -- 62.
[5] J. Bourgain, The Szlenk index and operators on C(K)-spaces. Bull. Soc.
Math. Belg. S´er. B 31 (1979) 87 -- 117.
[6] P. Brooker, Direct sums and the Szlenk index. J. Funct. Anal. 260 (2011)
2222 -- 2246.
[7] R. Deville, G. Godefroy, and V. Zizler, Smoothness and renormings in Banach
spaces, volume 64 of Pitman Monographs and Surveys in Pure and Applied
Mathematics. Longman Scientific & Technical, Harlow, 1993.
10
[8] I. Gasparis, Operators on C(K) spaces preserving copies of Schreier spaces.
Trans. Amer. Math. Soc. 357 (2005) 1 -- 30.
[9] P. H´ajek and G. Lancien, Various slicing indices on Banach spaces. Mediterr.
J. Math. 4 (2007) 179 -- 190.
[10] P. H´ajek, G. Lancien, and A. Proch´azka, Weak∗ dentability index of spaces
C([0, α]). J. Math. Anal. Appl. 353 (2009) 239 -- 243.
[11] P. H´ajek, V. Montesinos Santaluc´ıa, J. Vanderwerff, and V. Zizler, Biorthog-
onal systems in Banach spaces. CMS Books in Mathematics/Ouvrages de
Math´ematiques de la SMC, 26. Springer, New York, 2008.
[12] G. Lancien, On uniformly convex and uniformly Kadec-Klee renormings.
Serdica Math. J. 21 (1995) 1 -- 18.
[13] G. Lancien, On the Szlenk index and the weak∗-dentability index. Quart. J.
Math. Oxford Ser. (2) 47 (1996) 59 -- 71.
[14] G. Lancien, A survey on the Szlenk index and some of its applications. RAC-
SAM Rev. R. Acad. Cienc. Exactas F´ıs. Nat. Ser. A Mat. 100 (2006) 209 -- 235.
[15] S. Mazurkiewicz and W. Sierpinski, Contribution `a la topologie des ensembles
d´enombrables. Fund. Math. 1 (1920) 17 -- 27.
[16] I. Namioka and R.R. Phelps, Banach spaces which are Asplund spaces. Duke
Math. J. 42 (1975) 735 -- 750.
[17] H.P. Rosenthal, The Banach spaces C(K),
in: W.B. Johnson, J. Linden-
strauss (Eds.), Handbook of the Geometry of Banach spaces, vol. 2, Elsevier,
Amsterdam, 2003, pp. 1547 -- 1602.
[18] W. Rudin, Continuous functions on compact spaces without perfect subsets.
Proc. Amer. Math. Soc. 8 (1957) 39 -- 42.
[19] C. Samuel,
Indice de Szlenk des C(K),
in: S´eminaire de G´eom´etrie des
Espaces de Banach, vols. I-II, Publications Math´ematiques de l'Universit´e
Paris VII, Paris, 1983, pp. 81 -- 91.
[20] Z. Semadeni, Banach spaces of continuous functions. Vol. I. PWN -- Polish
Scientific Publishers, Warsaw, 1971. Monografie Matematyczne, Tom 55.
[21] W. Szlenk, The non-existence of a separable reflexive Banach space universal
for all separable reflexive Banach spaces. Studia Math. 30 (1968) 53 -- 61.
11
|
1610.03462 | 1 | 1610 | 2016-10-11T18:47:40 | $(\alpha,\beta)$-A-Normal Operators in Semi-Hilbertian Spaces | [
"math.FA"
] | In this paper we introduce and prove some properties of $(\alpha;\beta)$-normal operators according to semi-Hilbertian space structures. Furthermore we s,ate various inequalities between the A-operator norm and A-numerical radius of $(\alpha,\beta)$-normal operators in semi Hilbertian spaces. | math.FA | math |
( α, β)-A-NORMAL OPERATORS IN
SEMI-HILBERTIAN SPACES
Abdelkader Benali [1] and Ould Ahmed Mahmoud Sid Ahmed [2]
[1] Faculty of science, Mathematics Department,University of Hassiba
Benbouali, Chlef Algeria. B.P. 151 Hay Essalem, chlef 02000, Algeria.
[2] Mathematics Department, College of Science. Aljouf University
[email protected]
Aljouf 2014. Saudi Arabia
[email protected]
October 12, 2016
Abstract
Let H be a Hilbert space and let A be a positive bounded operator on H. The
semi-inner product hu viA := hAu vi, u, v ∈ H induces a semi-norm k. kA
on H. This makes H into a semi-Hilbertian space. In this paper we introduce and
prove some proprieties of (α, β)-normal operators according to semi-Hilbertian space
structures. Furthermore we state various inequalities between the A-operator norm
and A-numerical radius of (α, β)-normal operators in semi Hilbertian spaces.
Keywords. Semi-Hilbertian space, A-selfadjoint operators,A-normal operators, A-
positive operators, (α, β)-normal operators.
Mathematics Subject Classification: Primary 46C05, Secondary 47A05.
1
INTRODUCTION AND PRELIMINARIES
RESULTS
One of the most important subclasses of the algebra of all bounded linear operators acting
on Hilbert space, the class of normal operators (T T ∗ = T ∗T ).They have been the object
of some intensive studies. The theory of these operators was investigated in [5] and [20].
This class has been generalized, in some sense, to the larger sets of so-called quasinormal,
hyponormal, isometry, partial isometry, m-isometries operators on Hilbert spaces.
Recently, these classes of operators have been generalized by many authors when an
additional semi-inner product is considered (see [2, 3, 4, 17, 18, 21] ) and other papers.
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
2
In this framework, we show that many results from [7, 8, 12] remain true if we consider
an additional semi-inner product defined by a positive semi-definite operator A. We are
interested to introducing a new concept of normality in semi-Hilbertian spaces.
The contents of the paper are the following. In Section 1, we give notation and results
about the concept of A-adjoint operators that will be useful in the sequel. In Section 2 we
introduce the new concept of normality of operators in semi-Hilbertian space (H, h. .iA),
called (α, β)-A-normality and we investigate various structural properties of this class of
operators. In Section 3,we state various inequalities between the A-operator norm and
A-numerical radius of (α, β)-A-normal operators.
We start by introducing some notations. The symbol H stands for a complex Hilbert
space with inner product h. . i and norm k.k . We denote by B(H) the Banach algebra
of all bounded linear operators on H, I = IH being the identity operator. B(H)+ is the
cone of positive (semi-definite) operators, i.e.,
B(H)+ = {A ∈ B(H) : hAu,
ui ≥ 0, ∀ u ∈ H }. For every T ∈ L(H) its range is
denoted by R(T ), its null space by N (T ) and its adjoint by T ∗. If M ⊂ H is a closed
subspace, PM is the orthogonal projection onto M. The subspace M is invariant for T
if TM ⊂ M. We shall denote the set of all complex numbers and the complex conjugate
of a complex number λ by C and λ, respectively. The closure of R(T ) will be denoted by
R(T ), and we shall henceforth shorten T − λI by T − λ. In addition, if T, S ∈ B(H) then
T ≥ S means that T − S ≥ 0. .
Any A ∈ B(H)+ defines a positive semi-definite sesquilinear form, denoted by
h. .iA : B(H) × B(H) → C, hu viA = hAu vi.
1
1
2 u A
We remark that hu viA = hA
2 vi. The semi-norm induced by h..iA, which is
2 uk. This makes H into a semi-
denoted by k.kA, is given by kukA = hu ui
Hilbertian space. Observe that kukA = 0 if and only if u ∈ N (A). Then k.kA is a norm if
and only if A is an injective operator, and the semi-normed space (B(H),k.kA) is complete
if and only if R(A) is closed. Moreover h iA induced a seminorm on a certain subspace
of B(H), namely, on the subset of all T ∈ B(H) for witch there exists a constant c > 0
such that kT ukA ≤ ckukA for every u ∈ H (T is called A-bounded). For this operators it
holds
A = kA
1
2
1
kTkA =
sup
u∈R(A),u6=0
It is straightforward that
kT ukA
kukA
< ∞.
kTkA = sup{hT u viA : u, v ∈ H and kukA ≤ 1,kvkA ≤ 1 }.
Definition 1.1. ([2]) For T ∈ B(H), an operator S ∈ B(H) is called an A-adjoint of T
if for every u, v ∈ H
hT u viA = hu SviA,
i.e., AS = T ∗A.
If T is an A-adjoint of itself, then T is called an A-selfadjoint operator (cid:0)AT = T ∗A(cid:1).
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
3
It is possible that an operator T does not have an A-adjoint, and if S is an A-adjoint of
T we may find many A-adjoints; In fact, in AR = 0 for some R ∈ B(H),then S + R is an
A-adjoint of T . The set of all A-bounded operators which admit an A-adjoint is denoted
by BA(H). By Douglas Theorem (see [6, 10] ) we have that
BA(H) = (cid:8) T ∈ B(H) / R(T ∗A) ⊂ R(A) (cid:9).
If T ∈ BA(H), then there exists a distinguished A-adjoint operator of T , namely,the
reduced solution of equation AX = T ∗A, i.e., A†T ∗A. This operator is denoted by T ♯.
Therefore, T ♯ = A†T ∗A and
AT ♯ = T ∗A, R(T ♯) ⊂ R(A) and N (T ♯) = N (T ∗A).
Note that in which A† is the Moore-Penrose inverse of A. For more details see [2, 3, 4].
In the next proposition we collect some properties of T ♯ and its relationship with the
seminorm k . kA. For the proof see [2, 3, 4] .
Proposition 1.1. Let T ∈ BA(H). Then the following statements hold.
(1) T ♯ ∈ BA(H), (T ♯)♯ = PR(A)T PR(A) and (T ♯)♯)♯ = T ♯.
(2) If S ∈ BA(H) then T S ∈ BA(H) and (T S)♯ = S♯T ♯.
(3) T ♯T and T T ♯ are A-selfadjoint.
1
2
1
2
A.
(4) kTkA = kT ♯kA = kT ♯Tk
(5) kSkA = kT ♯kA for every S ∈ B(H) which is an A-adjoint of T.
(6) If S ∈ BA(H) then kT SkA = kSTkA.
A = kT T ♯k
We recapitulate very briefly the following definitions. For more details, the interested
reader is referred to [2, 4, 21] and the references therein.
Definition 1.2. Any operator T ∈ BA(H) is called
(1) A-normal if T T ♯ = T ♯T.
(2) A-isometry if T ♯T = PR(A).
(3) A-unitary if T ♯T = T T ♯ = PR(A).
In [21], the A-spectral radius of an operator T ∈ B(H), denoted rA(T ) is defined as
rA(T ) = lim sup
n−→∞ kT nk
1
n
A
and the A-numerical radius of an operator T ∈ B(H), denoted by ωA(T ) is defined as
wA(T ) = sup(cid:8) hT u uiA , u ∈ H ,kukA = 1 (cid:9).
It is a generalization of the concept of numerical radius of an operator. Clearly, ωA
defines a seminorm on B(H). Furthermore, for every u ∈ H,
hT u uiA ≤ ωA(T )kuk2
A .
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
4
Remark 1.1. If T ∈ BA(H) is A-selfadjoint,then kTkA = wA(T ) (see [21]).
Theorem 1.1. ([21], Theorem 3.1 )
A necessary and sufficient condition for an operator T ∈ BA(H) to be A-normal is that
(1) R(T T ♯) ⊂ R(A) and
(2) kT ♯T ukA = kT T ♯ukA for all u ∈ H.
2
PROPERTIES OF (α, β)-A-NORMAL OPERA-
TORS
In this section we define the class of (α, β)-A-normal operators according to semi-Hilbertian
space structures and we give some their proprieties.
Let (α, β) ∈ R2 such that 0 ≤ α ≤ 1 ≤ β, an operator T ∈ B(H) is said to be
(α, β)-normal [7, 19] if
which is equivalent to the condition
α2T ∗T ≤ T T ∗ ≤ β2T ∗T,
αkT uk ≤ kT ∗uk ≤ βkT uk
for all u ∈ H. For α = 1 = β is a normal operator. For α = 1, we observe from the left
inequality that T ∗ is hyponormal and for β = 1, from the right inequality we obtain that
T is hyponormal. In recent work, Senthilkumar [22] introduced p-(α, β)-normal operators
as a generalization of (α, β)- normal operators . An operator T ∈ B(H) is said to be
p-(α, β)- normal operators for 0 < p ≤ 1 if
α2(T ∗T )p ≤ (T T ∗)p ≤ β2(T ∗T )p, 0 ≤ α ≤ 1 ≤ β.
When p = 1, this coincide with (α, β)-normal operators.
Now we are going to consider an extension of the notion of (α, β) -normal operators,
similar to those extensions of the notion of normality to A-normality and hyponormality
to A-hyponormality (see [18, 21]).
Definition 2.1. ([18]) Let A ∈ B(H)+ and T ∈ B(H). We say that T is an A-positive if
AT ∈ B(H)+ which is equivalent to the condition
hT u uiA ≥ 0 ∀ u ∈ H.
We note T ≥A 0.
Definition 2.2. ([18]) An operator T ∈ BA(H) is said to be A-hyponormal if T ♯T − T T ♯
is A-positive i.e., T ♯T − T T ♯ ≥A 0.
Proposition 2.1. ([18]) Let T ∈ BA(H). Then T is A-hyponormal if and only if
kT ukA ≥ kT ♯ukA for all u ∈ H.
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
5
As a generalization of A-normal and A-hyponormal operators, we introduce (α, β)-A-
normal operators.
Definition 2.3. An operator T ∈ BA(H) is said to be (α, β)-A-normal for 0 ≤ α ≤ 1 ≤ β,
if
which is equivalent to the condition
β2T ♯T ≥A T T ♯ ≥A α2T ♯T,
βkT ukA ≥ kT ♯ukA ≥ αkT ukA, for all u ∈ H.
When A = I (the identity operator), this coincide with (α, β)-normal operator.
• For α = 1 = β is a A normal operator.
• For β = 1, we observe from the right inequality that T is A-hyponormal.
• For α = 1, and N (A) is invariant subspace for T from the right inequality we obtain
that T ♯ is A-hyponormal.
Remark 2.1. (1) Every A-normal operator is (α, β)-A-normal operator.
(2) If A is injective,then (1, 1)-A-normal is A-normal operator.
(3) If R(T T ♯) ⊂ R(A),then (1, 1)-A-normal is A-normal operator.
We give an example of (α, β)-A-normal operator which is neither A-normal nor A-
hyponormal.
Example 2.1. Let A = (cid:18) 1 0
0 2 (cid:19) and T = (cid:18) 1 2
A ≥ 0,R(T ∗A) ⊂ R(A)and T ♯ = (cid:18) 1 0
0 1 (cid:19) ∈ B(C2). It easy to check that
1 1 (cid:19) , T ♯T 6= T T ♯and kT ukA 6≥ kT ♯ukA.
T is neither A-normal nor A-hyponormal. Moreover
10T ♯T ≥A T T ♯ ≥A
1
6
T ♯T.
So T is (
1
√6
,√10)-A-normal operator.
The following theorem gives a necessary and sufficient conditions that an operator to
be (α, β)-A-normal. It is similar to [14, Theorem 2.3].
Theorem 2.1. Let T ∈ BA(H) and (α, β) ∈ R2 such that 0 ≤ α ≤ 1 ≤ β. Then T is
(α, β)-A-normal if and only if the following conditions are satisfied
λ2T T ♯ + 2α2λT ♯T + T T ♯ ≥A 0,
and
for all λ ∈ R
λ2T ♯T + 2λT T ♯ + β4T ♯T ≥A 0,
for all λ ∈ R
(1)
(2).
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
6
Proof. Assume that the conditions (1) and (2) are satisfied and prove that T is (α, β)-A-
normal.
In fact we have by using elementary properties of real quadratic forms
Similarly
2
2
A + 2α2λkT uk2
λ2T T ♯ + 2α2λT ♯T + T T ♯ ≥A 0
⇔ (cid:10)(λ2T T ♯ + 2α2λT ♯T + T T ♯)u u(cid:11) ≥A 0, ∀ u ∈ H, ∀ λ ∈ R
A +(cid:13)(cid:13)T ♯u(cid:13)(cid:13)
⇔ λ2(cid:13)(cid:13)T ♯u(cid:13)(cid:13)
A ≥A 0, ∀ u ∈ H, ∀ λ ∈ R
⇔ αkT ukA ≤ (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A , ∀ u ∈ H.
λ2T ♯T + 2λT T ♯ + β4T ♯T ≥A 0
⇔ (cid:10)(λ2T ♯T + 2λT T ♯ + β4T ♯T )u u(cid:11) ≥A 0, ∀ u ∈ H, ∀ λ ∈ R
A + β4 kT uk2
⇔ λ2 kT uk2
A ≥A 0, ∀ u ∈ H, ∀ λ ∈ R
⇔ (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ β kT ukA , ∀ u ∈ H.
A + 2λ(cid:13)(cid:13)T ♯u(cid:13)(cid:13)
αkT ukA ≤ (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ β kT ukA , ∀ u ∈ H.
2
Consequently
So T is (α, β)-A-normal as desired.
The proof of the converse seems obvious.
Proposition 2.2. Let T ∈ BA(H) such that N (A) is invariant subspace for T and let
(α, β) ∈ R2 such that 0 < α ≤ 1 ≤ β . Then T is an (α, β)-A-normal if and only if T ♯ is
1
(
β
Proof. First assume that T is (α, β)-A-normal operator. We have for all u ∈ H
)-A-normal operator.
1
α
,
On the other hand,since N (A) is invariant subspace for T we observe that T PR(A) =
PR(A)T and APR(A) = PR(A)A = A and it follows that
It follows that
Consequently
αkT ukA ≤ (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ β kT ukA .
1
β (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ kT ukA and kT ukA ≤
1
α (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A .
= (cid:13)(cid:13)(cid:13)PR(A)T PR(A)u(cid:13)(cid:13)(cid:13)A
(cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)♯u(cid:13)(cid:13)(cid:13)A
β (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ (cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)♯u(cid:13)(cid:13)(cid:13)A ≤
1
= kT ukA .
α (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A
1
)-A-normal operator.
for all u ∈ H. Therefore T ♯ is (
1
β
,
1
α
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
7
Conversely assume that T ♯ is (
1
β
,
1
α
)-A-normal operator. We have
for all u ∈ H, and from which it follows that
1
β (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ (cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)♯u(cid:13)(cid:13)(cid:13)A ≤
β (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ kT ukA ≤
1
1
α (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A
1
α (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A
for all u ∈ H. Hence
This completes the proof.
αkT ukA ≤ (cid:13)(cid:13)T ♯u(cid:13)(cid:13)A ≤ β kT ukA , ∀ u ∈ H.
The following corollary is a immediate consequence of Proposition 2.2.
Corollary 2.1. Let T ∈ BA(H) such that N (A) is invariant subspace for T and let
(α, β) ∈ R2 such that 0 < α ≤ 1 ≤ β and αβ = 1. Then T is an (α, β)-A-normal if and
only if T ♯ is (α, β)-A-normal operator.
Remark 2.2. (α, β)-A-normality is not translation invariant, more precisely, there exists
an operator T ∈ BA(H) that T is (α, β)-A-normal,but T + λ is not (α, β)-A-normal for
some λ ∈ C. The following example shows that such operators exist:
0 1 (cid:19) and S = T + I =
0 2 (cid:19) ,T = (cid:18) 1 2
Example 2.2. Consider the operators A = (cid:18) 1 0
0 2 (cid:19) ∈ B(R2). It is easily to check that T is (
(cid:18) 2 2
,√10)-A-normal, but S is not
,√10)-A-normal. So (α, β)-A-normality is not translation-invariant.
1
√6
1
√6
(
Similarly to [12], we define the following quantities
and
µ1
A(T ) = inf(cid:26) RehT u uiA
kT ukA
A(T ) = sup(cid:26) RehT u uiA
kT ukA
µ2
, kukA = 1, T u /∈ N (A
, kukA = 1, T u /∈ N (A
1
2 ) (cid:27)
1
2 ) (cid:27).
A.Saddi [21, Corollary 3.2 ] have shown that if T is A-normal operator such that N (A)
is invariant subspace for T , then T − λ is A-normal. In [18, Theorem 2.7 ] the authors
proved this property for A-hyponormal operators. In the following theorem we extend
these results to (α, β)-A-normal operators. This is a generalization of [12, Theorem 2.1].
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
8
Theorem 2.2. Let T ∈ BA(H) such that N (A) is invariant subspace for T and 0 ≤ α ≤
1 ≤ β. The following statements hold.
(1) If T is (α, β)-A-normal, then λT is (α, β)-A-normal for λ ∈ C .
(2) If T is (α, β)-A-normal, then T + λ for λ ∈ C is (α, β)-A-normal, if one of the
following conditions holds:
(i) µ1
A(λT ) ≥ 0
A(λT ) < 0, λ2 + 2λkTkAµ1
(ii) µ1
Proof. (1) Since N (A) is invariant subspace for T we observe that T PR(A) = PR(A)T and
APR(A) = PR(A)A = A. Let T be (α, β)-A-normal then
A(λT ) > 0.
β2T ♯T ≥A T T ♯ ≥A α2T ♯T ⇔ β2λ2T ♯T ≥A λ2T T ♯ ≥A λ2α2T ♯T
⇔ AλT ♯λT ≥ AλT λT ♯ ≥ α2AλT ♯λT
⇔ APR(A)λT ♯λT ≥ APR(A)λT λT ♯ ≥ α2APR(A)λT ♯λT
⇔ β2A(λT )♯(λT ) ≥ A(λT )(λT )♯ ≥ α2A(λT )♯(λT )
⇔ β2(λT )♯(λT ) ≥A (λT )(λT )♯ ≥A α2(λT )♯(λT ).
Therefore λT is (α, β)-A-normal operator.
(2) Assume that T is (α, β)-A-normal and the condition (i) holds. We need to prove that
α2D(cid:0)T + λ(cid:1)♯(cid:0)T + λ(cid:1)u uEA ≤ D(cid:0)T + λ(cid:1)(cid:0)T + λ(cid:1)♯u uEA
D(cid:0)T + λ(cid:1)(cid:0)T + λ(cid:1)♯u uEA ≤ β2D(cid:0)T + λ(cid:1)♯(cid:0)T + λ(cid:1)u uEA
.
In order To verify (2.1) we have
(2.1)
α2D(cid:0)T + λ(cid:1)♯(cid:0)T + λ(cid:1)u uEA
= α2(cid:26)(cid:10)T ♯T u u(cid:11)A +(cid:10)λT ♯u u(cid:11)A +DλPR(A)T u uEA
+λ2DPR(A)u uEA(cid:27)
The condition (i) implies that 2Re(cid:10)λT u u(cid:11)A ≥ 0 and it follows that
= α2(cid:26)(cid:10)T ♯T u u(cid:11)A + 2Re(cid:10)λT u u(cid:11)A + λ2 kuk2
A(cid:27)
A(cid:27).
≤ α2(cid:10)T ♯T u u(cid:11)A + α2(cid:26)2Re(cid:10)λT u u(cid:11)A + λ2 kuk2
A(cid:27)
α2D(cid:0)T + λ(cid:1)♯(cid:0)T + λ(cid:1)u uEA ≤ (cid:26)(cid:10)T T ♯u u(cid:11)A + 2Re(cid:10)λT u u(cid:11)A + λ2 kuk2
= D(cid:0)T + λ(cid:1)(cid:0)T + λ(cid:1)♯u uEA
A(cid:27)
= (cid:26)(cid:10)T T ♯u u(cid:11)A + 2Re(cid:10)λT u u(cid:11)A + λ2 kuk2
≤ β2D(cid:0)T + λ(cid:1)♯(cid:0)T + λ(cid:1)u uEA
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
9
and hence T + λ is (α, β)-A-normal. On the other hand if the condition (ii) is satisfied
then we have for λ 6= 0
A(λT )
λ2 + 2λ kTkA µ1
= λ2 + 2λ(cid:18) sup
≤ λ2 + 2 inf
≤ λ2 + 2Re(cid:10)λT u u(cid:11)A .
kukA=1
kukA=1kT ukA(cid:19)(cid:18) inf(cid:26) Re(cid:10)λT u u(cid:11)A
λkT ukA
Re(cid:10)λT u u(cid:11)A
A similar argument used as above shows that T + λ is (α, β)-A-normal.
, kukA = 1, T u /∈ N (A
1
2 ) (cid:27)(cid:19)
Corollary 2.2. Let T ∈ BA(H) be an (α, β)-A-normal operator. The following statement
hold
(1) If µ1
(1) If µ2
A(T ) ≥ 0 then T + λ is (α, β)-A-normal for every λ > 0.
A(T ) ≤ 0 then T + λ is (α, β)-A-normal for every λ < 0.
A(λT ) = µ1
Proof. (1) For every λ > 0 we have µ1
2.2 (i) we have that T + λ is an (α, β)-A-normal.
A(λT ) = µ1
A(T ) ≥ 0.By using Theorem
(2) For every λ < 0 we have µ1
that T + λ is an (α, β)-A-normal.
A(λT ) = −µ2
A(T ) ≥ 0.By using Theorem 2.2 (ii) we have
Lemma 2.1. ([18], Lemma 2.1 ) Let T, S ∈ B(H) such that T ≥A S and let R ∈ BA(H).
Then the following properties hold
(1) R♯T R ≥A R♯SR.
(2) RT R♯ ≥A RSR♯.
(3) If R is A-selfadjoint then RT R ≥A RSR.
Proposition 2.3. Let T, V ∈ BA(H) such that N (A) is invariant subspace for both T
and V .If T is an (α, β)-A-normal (0 ≤ α ≤ 1 ≤ β) and V is an A-isometry , then V T V ♯
is an (α, β)-A-normal operator.
Proof. Assume that β2T ♯T ≥A T T ♯ ≥A α2T ♯V and V ♯V = PR(A). This implies
β2(cid:0)V T V ♯(cid:1)♯(cid:0)V T V ♯(cid:1) = β2(cid:18)(cid:0)V ♯)♯T ♯V ♯V T V ♯(cid:19)
= β2(cid:18)PR(A)V PR(A)T ♯PR(A)T V ♯(cid:19)
= β2(cid:18)V PR(A)T ♯T(cid:0)V PR(A)(cid:1)♯(cid:19)
≥A V PR(A)T T ♯(cid:0)V PR(A)(cid:1)♯
≥A (cid:0)V T V ♯(cid:1)(cid:0)V T V ♯(cid:1)♯.
(by Lemma 2.1)
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
10
Similarly, we have
(cid:0)V T V ♯(cid:1)(cid:0)V T V ♯(cid:1)♯ = V PR(A)T T ♯(cid:0)V PR(A)(cid:1)♯
≥A α2V PR(A)T ♯T(cid:0)V PR(A)(cid:1)♯
≥A α2(cid:0)V T V ♯(cid:1)♯(cid:0)V T V ♯(cid:1).
(by Lemma 2.1)
The conclusion holds.
Proposition 2.4. Let T, S ∈ BA(H) such that T is (α, β)-A-normal and S is A-selfadjoint.
If T ♯S = ST ♯ then T S is (α, β)-A-normal.
Proof. Since T is (α, β)-A-normal we have for u ∈ H
α kT SukA ≤ (cid:13)(cid:13)T ♯Su(cid:13)(cid:13)A ≤ β kT SukA .
On the other hand
(cid:13)(cid:13)T ♯Su(cid:13)(cid:13)
This implies
2
A = (cid:10)T ♯Su T ♯Su(cid:11)A = (cid:10)AST ♯u ST ♯u(cid:11) = (cid:10)(T S)♯u (T S)♯u(cid:11)A = (cid:13)(cid:13)(T S)♯u(cid:13)(cid:13)
2
A .
αkT SukA ≤ (cid:13)(cid:13)(T S)♯u(cid:13)(cid:13)A ≤ β kT SukA .
Proposition 2.5. Let T, S ∈ BA(H) such that T is (α, β)-A-normal and S is A-unitary.
If T S = ST and N (A) is invariant subspace for T then T S is (α, β)-A-normal.
Proof. Since N (A) is invariant subspace for T we observe that T PR(A) = PR(A)T and
T ♯PR(A) = PR(A)T ♯. Let S be A-unitary then S♯S = SS♯ = PR(A).
Now it is easy to see that
β2(cid:18)(T S)♯(T S)(cid:19) = β2(cid:18)T ♯S♯ST(cid:19) = β2(cid:18)T ♯PR(A)T(cid:19) = β2(cid:18)PR(A)T ♯T PR(A)(cid:19).
By using the fact that T is (α, β)-A-normal,it follows immediately from Lemma 2.1 that
β2(cid:18)(T S)♯(T S)(cid:19) ≥A (cid:18)PR(A)T T ♯PR(A)(cid:19)
}
{z
(1)
Notice that (1) gives
≥A α2 (PR(A)T ♯T PR(A)(cid:19)
}
{z
(2)
.
(PR(A)T T ♯PR(A) = T PR(A)T ♯ = T SS♯T ♯ = T S(T S)♯
and similarly (2) gives
(PR(A)T ♯T PR(A) = T ♯PR(A)T = T ♯S♯ST = (T S)♯(T S).
So
Hence T S is (α, β)-A-normal operator.
β2(T S)♯(T S) ≥A T S(T S)♯ ≥A α2(T S)♯(T S).
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
11
The following example proves that even if T and S are (α, β)-A–normal operators,
their product T S is not in general (α, β)-A-normal operator.
Example 2.3. (1) Consider T =
0 0 1
0 1 0
1 0 0
−1
0
0 −1
0
0
0
0 −1
and S =
which are
0 −1 (cid:19) which are (α, β)-I2-normal whereas
0
(α, β)-I3-normal and their product is (α, β)-I3-normal.
1 1 (cid:19) and S = (cid:18) −1
−1 −1 (cid:19) is not (α, β)-I2-normal.
(2) Consider T = (cid:18) 1 0
their product T S = (cid:18) −1
Theorem 2.3. Let T, S ∈ BA(H) such that T is (α, β)-A-normal (0 ≤ α ≤ 1 ≤ β) and
S is (α′, β′) -A-normal (0 ≤ α′ ≤ 1 ≤ β′). Then the following statements hold:
(1) If T ♯S = ST ♯,then T S is (αα′, ββ′)-A-normal operator.
0
(2) If S♯T = T S♯,then ST is (αα′, ββ′)-A-normal operator.
Proof. (1) Since T is (α, β)-A-normal and S is (α′, β′)- A-normal, it follows that for all
u ∈ H
αα′ kT SukA ≤ α′(cid:13)(cid:13)T ♯Su(cid:13)(cid:13)A = α′(cid:13)(cid:13)ST ♯u(cid:13)(cid:13)A
≤ (cid:13)(cid:13)S♯T ♯u(cid:13)(cid:13)A
= (cid:13)(cid:13)(T S)♯u(cid:13)(cid:13)A
≤ β′(cid:13)(cid:13)ST ♯u(cid:13)(cid:13)A
= β′(cid:13)(cid:13)T ♯Su(cid:13)(cid:13)A
≤ ββ′ kT SukA .
αα′ kST ukA ≤ (cid:13)(cid:13)(T S)♯u(cid:13)(cid:13)A ≤ ββ′ kT SukA .
It follows that
The proof of the second assertion is completed in much the same way as the first assertion.
The following example shows that the power of (α, β)-A-normal operator not neces-
sarily an (α, β)-A-normal.
is (
1
√6
ther (
0 2 (cid:19) and T = (cid:18) 1 2
Example 2.4. Let A = (cid:18) 1 0
0 1 (cid:19) ∈ B(C2). By Example 2.1,T
,√10)-A-normal.However by direct computation one can show that T 2 is is nei-
,√10)-A-normal nor (
1
√6
, (√10)22(cid:1)-A-normal.
(cid:0)(cid:0) 1
√6(cid:1)22
Question. If T ∈ BA(H) which is (α, β)-A-normal operator, is that true T n is (αn2, βn2)-
A-normal operator?
, 100)-A-normal i.e., T 2 is
, 10)-A-normal. But it is (
1
36
1
6
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
12
Remark 2.3. Let T ∈ BA(H) ,then
(1) If T is A-normal, then rA(T ) = kTkA (see,[21, Corollary 3.2]).
(2) If T is A-hyponormal, then rA(T ) = kTkA (see,[18, Theorem 2.6]).
The following theorem presents a generalization of these results to (α, β)-A-normal. Our
inspiration cames from [12, Theorem 2.5].
Theorem 2.4. Let T ∈ BA(H) be an (α, β)-A-normal such that T 2n
for every n ∈ N,too. Then, we have
is (α, β)-A-normal
1
β kTkA ≤ rA(T ) ≤ kTkA .
Proof. It is we know that if T ∈ BA(H) then
and if T is A-selfadjoint then
(cid:13)(cid:13)T ♯T(cid:13)(cid:13)A = (cid:13)(cid:13)T T ♯(cid:13)(cid:13)A = kTk2
A
(cid:13)(cid:13)T 2(cid:13)(cid:13)A = kTk2
A .
From the definition of (α, β)-A-normal operator and Lemma 2.1.1 we deduce that
and so
Hence
We have
β2(cid:0)T ♯(cid:1)2T 2 ≥A (cid:0)T ♯T(cid:1)2 ≥A α2(cid:0)T ♯(cid:1)2T 2
sup
kukA=1D(cid:0)T ♯(cid:1)2T 2u uEA ≥
1
β2
sup
kukA=1D(cid:0)T ♯T(cid:1)2u uEA
.
2
A
=
1
β2 kTk4
A .
A ≥
2
(cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)2
T 2(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)2n
1
β2 (cid:13)(cid:13)(cid:13)(cid:0)T ♯T(cid:1)2(cid:13)(cid:13)(cid:13)
T 2n(cid:13)(cid:13)(cid:13)A ≥
1
.
A
1
2n
lim sup
β2n+1−2 kTk2n+1
n−→∞(cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)2n(cid:13)(cid:13)(cid:13)
n−→∞(cid:18)(cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)2n(cid:13)(cid:13)(cid:13)A(cid:13)(cid:13)T 2n(cid:13)(cid:13)A(cid:19) 1
T 2n(cid:13)(cid:13)(cid:13)A(cid:19) 1
n−→∞(cid:18)(cid:13)(cid:13)(cid:13)(cid:0)T ♯(cid:1)2n
2n
A
≥ lim
≥ lim
1
β2 kTk2
A .
≥
n−→∞(cid:13)(cid:13)T 2n(cid:13)(cid:13)
2n
1
2n
A
Now using a mathematical induction, we observe that for every positive integer number
n,
rA(T )2 = rA(T ♯)rA(T ) = lim sup
Therefore, we get
This completes the proof.
1
β kTkA ≤ rA(T ) ≤ kTkA .
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
13
Let H⊗H denote the completion, endowed with a reasonable uniform crose-norm,
of the algebraic tensor product H⊗H of H with H. Given non-zero T, S ∈ B(H), let
T ⊗ S ∈ B(H⊗H) denote the tensor product on the Hilbert space H⊗H, when T ⊗ S is
defined as follows
hT ⊗ S(ξ1 ⊗ η1) (ξ2 ⊗ η2)i = hT ξ1 ξ2ihSη1 η2i.
The operation of taking tensor products T ⊗ S preserves many properties of T, S ∈ B(H),
but by no means all of them. Thus, whereas T ⊗ S is normal if and only if T and S are
normal [15], there exist paranormal operators T and S such that T ⊗ S is not paranormal
[1]. In [9], Duggal showed that if for non-zero T, S ∈ B(H), T ⊗ S is p-hyponormal if
and only if T and S are p-hyponormal. Thus result was extended to p-quasi-hyponormal
operators in [16] .
Recall that for T ∈ BA(H) and S ∈ BB(H), T ⊗ S is (α, β)-(A ⊗ B)-normal operator
with 0 ≤ α ≤ 1 ≤ β, if
or equivalently
β2(cid:0)T ⊗ S(cid:1)♯(cid:0)T ⊗ S(cid:1) ≥A⊗B (cid:0)T ⊗ S(cid:1)(cid:0)T ⊗ S(cid:1)♯ ≥A⊗B α2(cid:0)T ⊗ S(cid:1)♯(cid:0)T ⊗ S(cid:1)
α(cid:13)(cid:13)(cid:0)T ⊗ S(cid:1)(cid:0)u ⊗ v(cid:1)(cid:13)(cid:13)A⊗B ≤ (cid:13)(cid:13)(cid:13)(cid:0)T ⊗ S(cid:1)♯(cid:0)u ⊗ v(cid:1)(cid:13)(cid:13)(cid:13)A⊗B ≤ β(cid:13)(cid:13)(cid:0)T ⊗ S(cid:1)(cid:0)u ⊗ v(cid:1)(cid:13)(cid:13)A⊗B ,
for all u, v ∈ H.
Lemma 2.2. ( [18], Lemma 3.1 )
Let Tk, Sk ∈ B(H), k = 1, 2 and Let A, B ∈ B(H)+, such that T1 ≥A T2 ≥A 0 and
S1 ≥B S2 ≥B 0, then
(cid:0)T1 ⊗ S1(cid:1) ≥A⊗B (cid:0)T2 ⊗ S2(cid:1) ≥A⊗B 0.
Proposition 2.6. ( [18], Proposition 3.2 ) Let T1, T2, S1, S2 ∈ B(H) and let A, B ∈
B(H)+ such that Tk is A- positive and Sk is B-positive for k = 1, 2. If T1 6= 0 and
S1 6= 0,then the following conditions are equivalents
(1) T2 ⊗ S2 ≥A⊗B T1 ⊗ S1
(2) there exists d > 0 such that dT2 ≥A T1 and d−1S2 ≥B S1.
The following theorem gives a necessary and sufficient condition for T ⊗S to be (α, β)-
-A ⊗ B-normal operator when T and S are both nonzero operators.
Proposition 2.7. Let T ∈ BA(H) and let S ∈ BB(H) with T 6= 0 and S 6= 0. Let
(α, β) ∈ R2 and (α′, β′) ∈ R2 such that 0 ≤ α, α′ ≤ 1 and 1 ≤ β, 1 ≤ β′. The following
properties hold:
(1) If T is an (α, β)-A normal and S is an (α′, β′)-B-normal ,then T ⊗ S is a (αα′, ββ′)-
A ⊗ B-normal operator.
(2) If T ⊗ S is an (α, β)-A ⊗ B-normal, then there exist two constants d > 0 and d0 > 0
such that T is (cid:0)qd−1
0 α,√dβ(cid:1)-A-normal and S is (cid:0)√d0,
1
√d(cid:1)-B-normal operator.
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
14
Proof. Assume that T is an (α, β)-A normal and S is an (α′, β′)-B-normal. By assumptions
we have
and
β2T ♯T ≥A T T ♯ ≥A α2T ♯T
β′2S♯S ≥B SS♯ ≥B α′2S♯S.
It follows from the inequalities above and Lemma 2.2 that
β2β′2T ♯T ⊗ S♯S ≥A⊗B T T ♯ ⊗ SS♯ ≥A⊗B α2α′2T ♯T ⊗ S♯S
and so
(ββ′)2(cid:0)T ⊗ S(cid:1)♯(cid:0)T ⊗ S(cid:1) ≥A⊗B (cid:0)T ⊗ S(cid:1)(cid:0)T ⊗ S(cid:1)♯ ≥A⊗B (αα′)2(cid:0)T ⊗ S(cid:1)♯(cid:0)T ⊗ S(cid:1).
Hence,T ⊗ S is a (αα′, ββ′)-A ⊗ B-normal operator.
Conversely assume that T ⊗ S is a (α, β)-A ⊗ B-normal operator.
We have
β2T ♯T ⊗ S♯S ≥A⊗B T T ♯ ⊗ SS♯ ≥A⊗B α2T ♯T ⊗ S♯S.
β2T ♯T ⊗ S♯S ≥A⊗B T T ♯ ⊗ SS♯
So
and
T T ♯ ⊗ SS♯ ≥A⊗B α2T ♯T ⊗ S♯S
(2.2)
(2.3)
We deduce from inequality (2.2) and Proposition 2.6 that there exists a constant d > 0
such that
It easily to see that
dβ2 sup
kukA=1(cid:10)T T ♯u u(cid:11)A
and so
Thus, dβ2 ≥ 1. Similarly, we obtain d−1 ≥ 1.
On the other had by inequality (2.3) and we can find a constant d0 > 0 satisfies
dβ2T ♯T ≥A T T ♯
and
d−1S♯S ≥B SS♯
kukA=1(cid:10)T ♯T u u(cid:11)A ≥ sup
dβ2(cid:13)(cid:13)T ♯T(cid:13)(cid:13)A ≥ (cid:13)(cid:13)T T ♯(cid:13)(cid:13)A .
qd−1
d0T T ♯ ≥A α2T ♯T
and
d−1
0 SS♯ ≥B S♯S
0 α ≤ 1 and d0 ≤ 1.
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
15
Consequently we have
and
(cid:0)√dβ(cid:1)2T ♯T ≥A T T ♯ ≥A (cid:0)qd−1
(cid:0) 1
S♯S ≥B SS♯ ≥B (cid:0)pd0(cid:1)2
√d(cid:1)2
0 α(cid:1)2T ♯T
S♯S.
This proof is completes.
Theorem 2.5. Let T, S ∈ BA(H) such that T is (α, β)-A-normal and S is (α′, β′)-A -
normal operators with 0 ≤ α ≤ 1 ≤ β and 0 ≤ α′ ≤ 1 ≤ β′. The following statements
hold:
(1) If T ♯S = ST ♯, then T S ⊗ T is (α2α′, β2β′)-(A ⊗ A)-normal operator and T S ⊗ S is
(αα′2, ββ′2)-A ⊗ A-normal operator
(2) If S♯T = T S♯,then ST ⊗ T is (α′α2, β′β2)-(A ⊗ A)-normal operator and ST ⊗ S is
(α′2α, β′2β)-A ⊗ A-normal operator.
Proof. The proof is an immediate consequence of Theorem 2.3 and Proposition 2.7.
3
INEQUALITIES INVOLVING A-OPERATOR
NORMS AND A-NUMERICAL RADIUS OF
(α, β)-A-NORMAL OPERATORS
Drogomir and Moslehian [7] have given various inequalities between the operator norm
and the numerical radius of (α, β)-normal operators in Hilbert spaces.
Motivated by this work, we will extended some of these inequalities to A-operator norm
and A-numerical radius ωA of (α, β)-A-normal in semi-Hilbertian spaces by employing
some known results for vectors in inner product spaces. We start with the following
lemma reproduced from [13].
Lemma 3.1. Let r ∈ R and u, v ∈ H such that kukA ≥ kvkA and u, v /∈ N (A) ,then the
following inequalities hold
ku − vk2
A if r ≥ 1
A
r2 kuk2r−2
and
(3.1)
kuk2r
A + kvk2r
A − 2kukr
A kvkr
A
Rehu viA
kukA kvkA ≤
ku − vk2
Theorem 3.1. T ∈ BA(H) be an (α, β)-A-normal operator. Then
2βrωA(T 2) + β2r−2(cid:13)(cid:13)βT − T ♯(cid:13)(cid:13)
kvk2r−2
A
2
(cid:18)α2r + β2r(cid:19)kTk2
A ≤
and
A if r < 1.
A if r ≥ 1
(3.2)
2βrωA(T 2) +(cid:13)(cid:13)βT − T ♯(cid:13)(cid:13)
2
A if r < 1.
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
16
Proof. Firstly , assume that r ≥ 1 and let u ∈ H with kukA = 1. Since T is (α, β)-A-
normal
we have
α2 kT uk2
A ≤ β2 kT uk2
A
2
A ≤ (cid:13)(cid:13)T ♯u(cid:13)(cid:13)
A ≤ β2r kT uk2r
(cid:18)α2r + β2r(cid:19)kT uk2r
A +(cid:13)(cid:13)T ♯u(cid:13)(cid:13)
2r
A .
Applying Lemma 3.1 with the choices u0 = β T u and v0 = T ♯u we get
2r
kβT uk2r
From which , it follows that
A +(cid:13)(cid:13)T ♯u(cid:13)(cid:13)
A −2kβT ukr−1
A (cid:13)(cid:13)T ♯u(cid:13)(cid:13)
r−1
A Re(cid:10)βT u T ♯u(cid:11)A ≤ r2 kβT uk2r−2
A
(cid:13)(cid:13)βT u − T ♯u(cid:13)(cid:13)
(3.3)
2
A .
(cid:18)α2r + β2r(cid:19)kT uk2r
A (cid:12)(cid:12)(cid:10)βT 2u u(cid:11)A(cid:12)(cid:12)
(cid:13)(cid:13)βT u − T ♯u(cid:13)(cid:13)
Taking the supremum in (3.4) over u ∈ H,kukA = 1 and using the fact that
A (cid:13)(cid:13)T ♯u(cid:13)(cid:13)
+r2 kβT uk2r−2
A ≤ 2kβT ukr−1
2
A .
r−1
A
(3.4)
sup
kukA=1(cid:12)(cid:12)(cid:10)T 2u u(cid:11)iA(cid:12)(cid:12) = ωA(T 2)
(cid:18)α2r + β2r(cid:19)kTk2r
A ≤ 2βr kTk2r−2
A
ωA(T 2) + r2β2r−2 kTk2r−2
A
(cid:13)(cid:13)βT − T ♯(cid:13)(cid:13)
2
A .
we get
So
(cid:18)α2r + β2r(cid:19)kTk2
A ≤ 2β2rωA(T 2) + r2β2r−2(cid:13)(cid:13)βT − T ♯(cid:13)(cid:13)
2
A
which is the first inequality in (3.2).
By employing a similar argument to that used in the first inequality in (3.1) , gives the
second inequality of (3.2).
Theorem 3.2. Let T ∈ BA(H) be an A(α.β)-normal operator. Then
ωA(T )2 ≤
Proof. Since for all u, v and e ∈ H
1
2(cid:18)β kTk2
A + ωA(T 2)(cid:19).
(3.5)
hu viA − hu eiA he viA ≥ hu eiA he viA − hu viA
we have by applying the inequalities reproduced from [11]
kukA kvkA ≥ hu viA − hu eiA he viA + hu eiA he viA ≥ hu viA
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
that
hu eiAhe viA ≤
1
2(cid:18)kukA kvkA + hu viA(cid:19)
for all u, v, e ∈ H with kekA = 1.
Let x ∈ H with kxkA = 1 and choosing in (3.6) u = T x , v = T ♯x and e = x we get
hT x xiA(cid:12)(cid:12)(cid:10)x T ♯x(cid:11)A(cid:12)(cid:12) ≤
Since T is (α, β)-A-normal, it follows that
1
2(cid:18)kT xkA(cid:13)(cid:13)T ♯x(cid:13)(cid:13)A +(cid:12)(cid:12)(cid:10)T x T ♯x(cid:11)A(cid:12)(cid:12)(cid:19).
2(cid:18)β kT xk2
A +(cid:12)(cid:12)(cid:10)T 2x x(cid:11)A(cid:12)(cid:12)(cid:19).
1
hT x xiA2 ≤
Tanking the supremum over x ∈ H kxkA = 1, we get the desired inequality in (3.5).
Theorem 3.3. Let T ∈ BA(H) be an (α, β)-A-normal operator and λ ∈ C. Then
17
(3.6)
(3.7)
(3.8)
(3.9)
αkTk2
A ≤ ωA(T 2) +
2
A
.
2β(cid:13)(cid:13)T − λT ♯(cid:13)(cid:13)
(cid:0)1 + λα(cid:1)2
Proof. For λ = 0, the inequality (3.9) is obvious. Assume that λ 6= 0.From the following
inequality [13]
1
2(cid:18)kuk + kvk(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
u
kuk −
v
kvk(cid:13)(cid:13)(cid:13)(cid:13) ≤ ku − vk , ; u, v ∈ H − {0}
which is well known in the literature as the Dunkl-Williams inequality, it follows that
A simple computation shows that
1
2(cid:0)kukA + kvkA(cid:1)(cid:13)(cid:13)(cid:13)(cid:13)
kukA −
u
(cid:13)(cid:13)(cid:13)(cid:13)
which shows that
u
kukA −
v
kvkA(cid:13)(cid:13)(cid:13)(cid:13)A ≤ ku − vkA for all u, v ∈ H / u, v /∈ N (A).
2
A
v
kvkA(cid:13)(cid:13)(cid:13)(cid:13)
= 2 − 2
Rehu viA
kukA kvkA ≤
A
4ku − vk2
(cid:0) kukA + kvkA(cid:1)2
kukA kvkA − hu viA
kukA kvkA
and so
≤
A
2ku − vk2
(cid:0)kukA + kvkA(cid:1)2 ,
for all u, v ∈ H / u, v /∈ N (A)
kukA kvkA ≤ hu viA +
2kukA kvkA
(cid:0)kukA + kvkA(cid:1)2 ku − vk2
A .
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
Let x ∈ H with kxkA = 1 and consider u = T x and v = λT ♯x with x /∈ N (A
N (A
2 T ♯) we obtain
1
kT xkA(cid:13)(cid:13)λT ♯x(cid:13)(cid:13)A ≤ (cid:12)(cid:12)(cid:10)T x λT ♯x(cid:11)A(cid:12)(cid:12) +
Since T being (α, β)-A-normal operator, we get
αkT xk2
A ≤ (cid:12)(cid:12)(cid:10)T 2x x(cid:11)A(cid:12)(cid:12) +
2
A .
2kT xkA(cid:13)(cid:13)λT ♯x(cid:13)(cid:13)A
(cid:0) kT xkA + kλT ♯xkA(cid:1)2 (cid:13)(cid:13)T x − λT ♯x(cid:13)(cid:13)
2β kT xk2
A
(cid:0)kT xkA + αλ kT xkA(cid:1)2 (cid:13)(cid:13)T x − λT ♯x(cid:13)(cid:13)
2
A .
18
1
2 T ) =
(3.10)
(3.11)
(3.12)
Tanking the supremum over x ∈ H; kxkA = 1, we get the desired inequality in (3.9).
Theorem 3.4. Let T ∈ BA(H) be an (α, β)-A-normal operator and λ ∈ C. Then
(cid:20)α2 −(cid:0) 1
λ
+ β(cid:1)2(cid:21)kTk4
A ≤ ωA(T 2).
Proof. We apply the following inequality inspired from [8]
1
λ2 kuk2
A ku − vk2
A
0 ≤ kuk2
A − hu viA ≤
A kvk2
for all u, v ∈ H and λ ∈ C , λ 6= 0.
Let x ∈ H and set u = T x and v = T ♯x in (3.11) we get
λ2 kT xk2
A ≤ (cid:12)(cid:12)(cid:10)T 2x x(cid:11)A(cid:12)(cid:12)
α2 kT xk4
2 +
1
A(cid:0)1 + λβ(cid:1)2 kT xk2
A .
Tanking the supremum over x ∈ H kxkA = 1, we get the desired inequality in (3.10).
References
[1] T. Ando, Operators with a norm condition, Acta Sci. Math.(Szeged) 33 (1972),169 -
178.
[2] M.L. Arias, G. Corach, M.C. Gonzalez, Partial isometries in semi-Hilbertian spaces,
Linear Algebra Appl. 428 (7) (2008) 1460-1475.
[3] M.L. Arias, G. Corach, M.C. Gonzalez,Metric properties of projections in semi- Hilber-
tian spaces,Integral Equations Operator Theory 62 (1) (2008) 11–28.
[4] M. L. Arias, G. Corach, M. C. Gonzalez, Lifting properties in operator ranges,Acta
Sci. Math. (Szeged) 75:3-4(2009), 635–653.
[5] J.B. Conway, A Course in Functional Analysis, Second Edition, Springer Verlag, Berlin
- Heildelberg - New York 1990.
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
19
[6] R.G. Douglas, On majorization, factorization and range inclusion of operators in
Hilbert space, Proc. Amer. Math. Soc. 17 (1966) 413-416.
[7] S. S. Dragomir and M. S. Moslehian, Some inequalities for (α, β)-normal operators in
Hilbert spaces, Facta Universitatis, vol. 23,(2008) pp. 3947.
[8] S. S. Dragomir,Inequalities for normal operators in Hilbert spaces, Applicable Analysis
and Discrete Mathematics, 1 (2007), 92-110.
[9] B.P.Duggal,Tensor products of operators-strong stability and p-hyponormality.Glasgow
Math Journal. 42( 2000),371–381.
[10] P.A. Fillmore, J.P.Williams, On operator ranges, Adv. Math. 7 (1971). 254–281.
[11] C.F. Dunkl and K.S. Williams, A simple norm inequality, Amer. Math. Monthly,
71(1) (1964), 4344.
[12] R. Eskandari, F. Mirzapour, and A. Morassaei, More on (α, β)-Normal Operators
in Hilbert Spaces,Abstract and Applied Analysis Volume 2012, Article ID 204031, 11
pages.
[13] A. Goldstein, J.V. Ryff and L.E. Clarke, Problem 5473, Amer. Math. Monthly, 75(3)
(1968), 309.
[14] A. Gupta and P.Sharma,(α, β)-Normal Composition Operators, Thai Journal of
Mathematics Volume 14 (2016) Number 1 : 83-92.
[15] J.C. Hou, On the tensor products of operators, Acta Math. Sinica (N.S.) 9 (1993),
no. 2, 195 –202.
[16] I.H. Kim,On (p; k)-quasihyponormal operators. Mathematical inequalities and Ap-
plications. Vol. 7,Number 4 (2004),629–638.
[17] O.A.Mahmoud Sid Ahmed and A.Saddi, A-m-Isomertic operators in semi-Hilbertian
spaces,Linear Algebra and its Applications 436 (2012) 3930-3942.
[18] O.A.Mahmoud Sid Ahmed and A.Benali, Hyponormal and k-quasi-hyponormal op-
erators on Semi-Hilbertian spaces The Australian Journal of Mathematical Analysis
and Applications.Volume 13, Issue 1, Article 7, (2016), pp. 1–22.
[19] M.S. Moslehian, On (α, β)-normal operators in Hilbert spaces, IMAGE, 39 (2007)
Problem 39–4.
[20] C.R. Putnam, On normal operators in Hilbert space, Amer. J. Math. 73(1951), 357-
362.
[21] A.Saddi, A-Normal operators in Semi-Hilbertian spaces.The Australian Journal of
Mathematical Analysis and Applications.Volume 9, Issue 1, Article 5, (2012), pp. 1–
12.
Abdelkader Benali and Ould Ahmed Mahmoud Sid Ahmed
20
[22] D. Senthilkumar,On p-(α, β)-Normal Operators, Applied Mathematical Sciences, Vol.
8, 2014, no. 41, 2041 - 2052.
|
1103.3610 | 1 | 1103 | 2011-03-18T12:14:55 | Harmonic analysis of weighted $L^p$-algebras | [
"math.FA"
] | Let $G$ be a locally compact, compactly generated group of polynomial growth and let $\omega$ be a weight on $G$. Under proper assumptions on the weight $\omega$, the Banach space $L^p(G,\omega)$ is a Banach \ast-algebra. In this paper we give examples of such weighted $L^p$-algebras and we study some of their harmonic analysis properties, such as symmetry, existence of functional calculus, regularity, weak Wiener property, Wiener property, existence of minimal ideals of a given hull. | math.FA | math |
Harmonic analysis of weighted Lp-algebras
Yu. N. Kuznetsova
C. Molitor-Braun ∗
Abstract
Let G be a locally compact, compactly generated group of polynomial growth and let ω
be a weight on G. Under proper assumptions on the weight ω, the Banach space Lp(G, ω)
is a Banach ∗-algebra. In this paper we give examples of such weighted Lp-algebras and we
study some of their harmonic analysis properties, such as symmetry, existence of functional
calculus, regularity, weak Wiener property, Wiener property, existence of minimal ideals of a
given hull. 1
1
Introduction
Weights and weighted function spaces play an important role in mathematics.
In essence, a
weight makes it possible to study the behaviour of functions around a certain point, ignoring
their oscillations at infinity, or on the contrary, to amplify the asymptotic behaviour of a function.
More precisely, introducing a weight means modelling in a quantitative manner the decay of the
functions to be studied. This has numerous applications in numerical mathematics and is quite
often used for concrete applications (signal theory, Gabor analysis, sampling theory,...), see for
instance ([Gro-Lei], [Da-Fo-Gro], [Gro-Lei1], [Fe-Gro-Lei]).
On the other hand, weights appear naturally in analysis: in inequalities relating the norm of a func-
tion to the norm of its derivatives, in extension theorems, etc.; see, e.g., a survey of L. D. Kudryavt-
sev and S. M. Nikol'sky [Kud-S.Ni]. One of the areas where weighed spaces are applied most
intensively is the theory of boundary value problems for partial differential equations (see the
surveys [Kud-S.Ni], [Glu-Sav]). By the way, using the Laplace transform also means working in a
weighted function space.
In representation theory, which interests us most, weights occur for instance in the following way.
If G denotes a locally compact group and (T, V ) is a continuous representation of G, then the
maps ω : x 7→ kT (x)kop and ω : x 7→ max(kT (x)kop,kT (x−1kop) are weights, the last one being
symmetric (ω(x−1) = ω(x), ∀x). For any one of these weights, the map
f 7→ T (f ) :=ZG
f (x)T (x)dx
is a representation of the weighted function algebra
L1(G, ω) := {f : G → C f measurable and ZG f (x)ω(x)dx < +∞}.
Subexponential weights like the ones introduced in 1.2.3 below, appear in the context of nilpotent
Lie groups. In fact, let G be a connected, nilpotent Lie group. Let G1 be the derived group of
G, i. e. the closed subgroup generated by the elements of the form [x, y] = x−1y−1xy, x, y ∈ G.
Let U be a generating neighbourhood of the identity e in G and V = U ∩ G1 the corresponding
neighbourhood of e in G1. Let
xU := inf{n ∈ N x ∈ U n} for x 6= e, eU = 1,
∗Supported by the research grant F1R-MTH-PUL-10NCHA of the University of Luxembourg
1keywords: Weighted Lp-algebras, symmetry, functional calculus, regularity, weak Wiener property, Wiener
property, minimal ideals of a given hull -- 2010 Mathematics Subject Classification: 43A15, 22D15, 22D20
1
similarly for xV . Then it is shown in [Al]), that for any weight ω on G which is submultiplicative,
i. e. such that ω(xy) ≤ ω(x)ω(y) for all x, y,
ωG1(x) ≤ eCx
1
2
V , ∀x ∈ G1,
for some constant C. By the way, on any compactly generated locally compact group, with
generating neighbourhood U , every submultiplicative weight ω is exponentially bounded, i. e.
satisfies a relation of the form ω(x) ≤ eKxU for K = ln supx∈U ω(x).
If the weight ω is submultiplicative, then the weighted function space L1(G, ω) is a Banach algebra
for convolution, and even a Banach ∗-algebra if the weight is symmetric. The advantage of
Banach ∗-algebras over just Banach spaces is clear. They have a much richer structure which may
be studied via representation theory and harmonic analysis techniques. In this way, interesting
problems arise. Let us just mention the question of their ideal theory, problems of generalized
spectral synthesis, of symmetry of the algebra, of invertibility, factorization problems. All these
questions make sense for the weighted algebra L1(G, ω). Moreover, it is the harmonic analysis
properties for algebras, that make the weighted algebras L1(G, ω) interesting for some of the
concrete applications mentioned in the beginning (see for instance [Gro-Lei1]).
On the other hand, the importance of Lp-spaces of the form Lp(G) or Lp(G, ω), 1 < p < +∞,
is well known in functional analysis. It would be attractive to extend the theory of convolution
algebras to the Lp-case, because Lp spaces are reflexive -- not a common property among Banach
algebras. Unfortunately, if G is not compact and if p 6= 1, Lp(G) is not an algebra for convolution.
Nevertheless, for appropriate groups G and weights ω, the weighted Lp-spaces
Lp(G, ω) := {f : G → C f measurable and kfkp,ω :=(cid:0)ZG f (x)pω(x)pdx(cid:1)
1
p < +∞}
may be algebras. A sufficient condition on the group G for the existence of weighted Lp-algebras,
is that the group is σ-compact. In that case, there are even a lot of such weighted Lp-algebras.
In 1.2.2 we show that any positive symmetric submultiplicative function multiplied by any Lp-
algebra weight produces again an Lp-algebra weight. This makes it possible to construct Lp-algebra
weights with all kinds of different growth behaviors.
In the context of weighted Lp-algebras, let us mention the works of Wermer [We], Nikol'ski [N.Ni],
Feichtinger [Fei], and recently ([Ku1]-[Ku4]). Most of these papers concentrate mainly on the ques-
tion whether the corresponding Lp-spaces are algebras. The only well-studied case is Lp(Z, ω), see,
e.g., a long paper of El-Fallah, Nikol'ski and Zarrabi [Fa-N.Ni-Za]. This is mainly for the reason
that in the problems of weighted approximation by polynomials, as initiated by S. N. Bernstein
[Be], Lp(Z, ω) algebras play a distinguished role [N.Ni]. But this is not the only possible applica-
tion of weighted Lp-algebras. Similarly as for L1-algebras, the weights can be used in numerical
mathematics to model the decay of the functions to be used and allow numerical computations. On
the other hand, weighted Lp-algebras may turn out to be important examples for people working
in Banach algebra theory or operator theory. Such algebras have already been used successfully
in the interpolation theory and in questions of factorization [Bl-Ka-Ra].
As for possible harmonic analysis properties of an arbitrary Banach ∗-algebra A, several questions
have particularly caught the interest of mathematicians:
Is the algebra symmetric, i. e. does every self-adjoint element have a real spectrum?
Is the algebra A regular, i. e. do the elements of the algebra separate points from closed sets in
Prim∗A, the space of kernels of topologically irreducible unitary representations?
Does the algebra have the weak Wiener property, i. e.
is every proper, closed, two-sided ideal
annihilated by an algebraically irreducible representation?
Does the algebra have the Wiener property, i. e. does the previous property hold for topologically
irreducible unitary representations?
Do there exist minimal ideals of a given hull?
The list of authors who studied group algebras L1(G) and their properties is long and is outside
the scope of this paper. For weighted group algebras L1(G, ω), let us mention among others the
2
following: In the abelian case, the systematic study of such properties for weighted group algebras
L1(G, ω) goes back to Beurling ([Beu1], [Beu2]), Domar [Do] and Vretblad [Vr] among others. In
the non-abelian case, one may refer to Hulanicki ([Hu1], [Hu2]), Pytlik ([Py1], [Py2]), as well as
to more recent studies ([Dz-Lu-Mo]), [Fe-Gro-Lei-Lu-Mo] and [Fe-Gro-Lei]). In [Fe-Gro-Lei] for
instance, the question of the symmetry for weighted group algebras L1(G, ω) is completely solved
for compactly generated groups with polynomial growth. Let us mention in this respect, that
these abstract problems may be quite important for concrete applications. Hence Grochenig and
Leinert [Gro-Lei] point out that the theory of symmetric group algebras is an important tool to
solve problems about Gabor frames, motivated by signal theory.
A systematic study of harmonic analysis properties of weighted Lp-algebras Lp(G, ω) should also
be of importance, as well for applications and for more abstract mathematical problems. In [Ku3]
some harmonic analysis properties like the regularity are studied in the case of abelian groups. But
not much seems to have been done up to now in the non-abelian case. Hence the main purpose
of this paper will be to study harmonic analysis properties in the context of non-abelian weighted
Lp-algebras.
In the present paper, we work on general compactly generated groups with polynomial growth. The
weight ω is supposed to satisfy ω−q ∗ ω−q ≤ Cω−q for some constant C > 0, where 1
q = 1. This
ensures Lp(G, ω) to be an algebra ([We], [Ku1]). We also assume the weight to be submultiplicative.
We start by giving examples of Lp(G, ω) ∗-algebras, as well for polynomially growing weights as for
sub-exponentially growing weights. We then address the questions raised previously: We prove the
symmetry of the Lp-algebra Lp(G, ω), if either G is abelian and ω satisfies the same condition as for
the case of L1(G, ω) (condition (S), [Fe-Gro-Lei-Lu-Mo], [Fe-Gro-Lei]) or if ω is polynomial in the
sense of Pytlik [Py2] (and G not necessarily abelian). The same hypothesis as in [Dz-Lu-Mo], i. e.
the non-abelian Beurling-Domar condition (BDna) allows us to construct a functional calculus on
a total subset of the algebra Lp(G, ω) and to show regularity, as well as the weak Wiener property.
If the Lp-algebra is moreover symmetric, we also get the Wiener property and the existence of
minimal ideals of a given hull. Let us recall that the (BDna) condition is defined as follows in
p + 1
[Dz-Lu-Mo]: Let G = Sn U n, where U is a relatively compact, generating neighbourhood of e in
G. We define s(n) := supx∈U n ω(x). Then the weight ω satisfies (BDna) if and only if
Xn∈N,n≥ee (cid:0) ln(ln n)(cid:1) ln(cid:0)s(n)(cid:1)
1 + n2
< +∞.
This condition is independent of the choice of the generating neighbourhood U and is only slightly
stronger than the conditions used by Domar and Beurling in the abelian case (see [Dz-Lu-Mo] for
additional comments). These results relay on the corresponding results for L1-algebras L1(G, ω).
The question of whether, more generally, condition (S) or the GRS-condition as defined in [Fe-Gro-Lei-Lu-Mo]
and [Fe-Gro-Lei] imply the symmetry of the algebra Lp(G, ω), as they do for L1(G, ω), is still an
open problem. Finally, let us point out that although these algebras Lp(G, ω) have certain nice
harmonic analysis properties, they are not amenable if p > 1 and G non-discrete, as they don't
have bounded approximate identities [Ku2].
Let us also mention, that we assume our weights to be submultiplicative, in order for L1(G, ω)
to be a ∗-algebra. Therefore we can rely on the known L1(G, ω)-results. But there are examples
of weights which are not submultiplicative and which produce nevertheless Banach ∗-algebras
Lp(G, ω) (see for instance [Ku2]). Studying these weights and the properties of the corresponding
Lp-algebras is still a challenge.
1.1 Assumptions on groups and weights
We suppose in this paper that G is a compactly generated locally compact group. This group G
is a group of polynomial growth if there is a relatively compact generating neighbourhood of the
identity U such that U n ≤ CnQ for some constants C, Q. It is known that Q does not depend on
the choice of U . The class of such groups will be denoted by [PG], and for the rest of the paper
we assume that G is [PG].
3
If U is a relatively compact generating neighbourhood of the identity, we define
xU := inf{n x ∈ U n}.
When the choice of U is not important, we write simply x. In the case G = R, x may also
denote the absolute value of x, and the results of this paper remain correct.
Let
be a measurable function (weight ) such that
ω : G → [1, +∞[
ω(xy) ≤ ω(x)ω(y),
ω(x) = ω(x−1),
ω−q ∗ ω−q ≤ ω−q,
∀x, y ∈ G
∀x ∈ G
p + 1
where 1
q = 1, p > 1. These conditions are sufficient for Lp(G, ω) to be a ∗-Banach algebra ([We],
[Ku1]). We will say that (G, ω) satisfies (LPAlg) (Lp-algebra) if these conditions are satisfied.
It is often easier to check that ω satisfies conditions (LPAlg) with some constants C1, C2: ω(xy) ≤
C1ω(x)ω(y) and ω−q ∗ ω−q ≤ C2ω−q. But a renormalizing ω1 = Cω with C = max(C1, C1/q
)
gives an equivalent weight satisfying (LPAlg).
Since every group in [PG] is amenable, it follows from (LPAlg), by [Ku4, Theorem 3.2], that
ω−q ∈ L1(G). We may assume, without loss of generality, that the weight ω is continuous ([Fei]).
This will be assumed for the rest of the paper, except for some examples which depend on the
discontinuous function x = xU .
2
1.2 Examples of weights
1.2.1 Polynomial weights
On every group of polynomial growth, the weight ω(x) = (1 + x)D satisfies (LPAlg) for D
sufficiently large ([Fei]), so Lp(G, ω) is an algebra.
1.2.2 Products of weights
Let u be a positive submultiplicative function on G, u(x) = u(x−1), and let w1 be a weight
satisfying (LpAlg). Then w(x) = u(x)w1(x) also satisfies (LpAlg). In particular, any such sub-
multiplicative function u, multiplied by (1 +x)D for D sufficiently large, is a (LpAlg)-weight. To
prove this, we need to check only the last condition:
(w−q ∗ w−q)(x) =ZG
w−q(y)w−q(y−1x)dy =ZG
u−q(y)u−q(y−1x)w−q
1 (y)w−q
1 (y−1x)dy.
From u(x) ≤ u(y)u(y−1x) we have u(x)−q ≥ u(y)−qu(y−1x)−q, so the integral above is bounded
by
(w−q ∗ w−q)(x) ≤ u−q(x)ZG
1 (y)w−q
w−q
1 (y−1x)dy ≤ u−q(x)w−q
1 (x) = w−q(x).
1.2.3 Non-polynomial weight
By the reasoning of section 1.2.2, ω(x) = exγ
(1 + x)D with 0 < γ ≤ 1 is an Lp-algebra weight
on any group in [PG] for all D sufficiently large.
Moreover, it can be shown that the weight ω(x) = exγ
itself with 0 < γ < 1 satisfies (LPAlg)
for all p > 1. For G = R this example is contained in the very first paper of Wermer [We] on
Lp-algebras.
4
Let G = ∪U n, where U = U −1 and U n ≤ CnQ. Denote Un = U n \ U n−1 and assume that G is
non-compact. We define
ωn ≡ ωUn = enγ
,
0 < γ < 1, and show that Lp(G, ω) is an algebra for every p > 1 (the case p = 1 is known). For
this, we check the sufficient condition ω−q ∗ ω−q ≤ C′ω−q.
Denote u = ω−q, then
un ≡ uUn = e−qnγ
.
Take x ∈ Um and estimate
(x).
u ∗ u
u
um ZG
1
u ∗ u
u
(x) =
u(y)u(y−1x)dy =Xn
un
um ZUn
u(y−1x)dy.
If y ∈ Un = U −1
Then Unx = ∪kU x
as
nk, Uk = ∪nU x
n , and y−1x ∈ Uk, then max(n− m, m− n) ≤ k ≤ n+ m. Denote U x
nk = Uk∩(Unx).
nk. In particular, U x
nk ≤ min(Un,Uk). We can rewrite the sum
un
(x) =Xn,k
um
u ∗ u
u
ukU x
nk.
Note that un decreases and C0 =RG u < ∞. Split now the sum into four parts:
um
um
1) n ≥ m. Then un ≤ um, and
(sum1) ≤Xn,k
nk ≤Xk
2) Similarly, if n < m but k ≥ m then uk ≤ um and
nk ≤Xn
(sum2) ≤Xn,k
unU x
ukU x
um
um
ukUk =ZG
unUn =ZG
u = C0.
u = C0.
3) m/2 < n < m, m/2 < k < m. Then max(un, uk) ≤ u[m/2]+1 ≤ exp(−qmγ/2γ);
m
m
Xk=[m/2]+1
(sum3) ≤
≤
m
Xn=[m/2]+1
Xn=1
eq(mγ −2mγ 2−γ )U x
n,k
eqmγ (1−21−γ )Un ≤ eqmγ (1−21−γ )CmQ · m.
Since 21−γ > 1, the coefficient in the exponent is negative, so this expression tends to zero as
m → ∞. Thus, this is bounded by a constant C1.
m − k ≤ n < m which reduces to the first one by exchanging k and n. Here uk ≤ um−n, so
4) The only complicated case is n ≤ m/2, m − n ≤ k < m, and the symmetric case k ≤ m/2,
m
Xk=m−n
U x
nk ≤
[m/2]
Xn=0
unum−n
um Un
eq(mγ −nγ −(m−n)γ )nQ.
(sum4) ≤
[m/2]
unum−n
um
[m/2]
Xn=0
Xn=0
≤ C
f (x) =Z x/2
0
Denote
clearly our sum is bounded for all m if and only if f is bounded on R+.
tQeq(xγ −tγ −(x−t)γ )dt;
5
tQeqxγ (1−tγ −(1−t)γ )dt.
By changing the variable to s = t/x we have:
f (x) =Z 1/2
xQsQeqxγ (1−sγ −(1−s)γ )xds = xQ+1Z 1/2
There is a classical theorem [Erd, §2.4] for integrals of the type
0
0
F (x) =Z β
α
g(t)exh(t)dt.
Suppose that:
h is real-valued and continuous at t = α;
h′ exists and is continuous for α < t ≤ β;
h′ < 0 for α < t < α + η with some η > 0;
h(t) ≤ h(α) − ε for some ε > 0 and all t ∈ [α + η, β];
h′(t) ∼ −a(t − α)ν−1 as t → α, where ν > 0;
g(t) ∼ b(t − α)λ−1 as t → α, where λ > 0.
Γ(cid:16) λ
ν(cid:17)(cid:16) ν
ax(cid:17)λ/ν
f (x) ∼
b
ν
exh(α)
Then
as x → ∞.
We can apply this theorem to
F (x) =Z 1/2
0
tQex(1−tγ −(1−t)γ )dt.
In this case g(t) = tQ, h(t) = 1 − tγ − (1 − t)γ, α = 0, β = 1/2. The derivative h′(t) =
−γ(cid:0)tγ−1 − (1− t)γ−1(cid:1) is negative on (0, 1/2) since γ − 1 < 0 (so that tγ−1 decreases) and t < 1− t.
It follows that h decreases, so all the conditions hold. We have b = 1, λ = Q + 1, a = ν = γ. Thus,
F (x) ∼
1
γ
If we return to f , then we get
γ (cid:17)Å γ
Γ(cid:16) Q + 1
γxã(Q+1)/γ
ex·0 ≡ C2x−(Q+1)/γ.
f (x) = xQ+1F (qxγ ) ∼ C2xQ+1(qxγ )−(Q+1)/γ = C2q−(Q+1)/γ ≡ C3.
It follows that there is a constant C4 such that f (x) ≤ C4 for all x > 0.
Now, collecting all together, we have
u ∗ u
u
(x) ≤ 2C0 + C1 + C4 ≡ C′,
what completes the proof.
1.2.4 Fast-growing weights
If the weight is submultiplicative, as we always assume, then it can grow at most exponentially.
But Lp(G, ω) is in general not an algebra with the weight ω(x) = ex. We will show this for
G = R. Take nonnegative f, g ∈ Lp(R, ex), then F = exf, G = exg are in Lp(R);
(f ∗ g)(s) =Z ∞
−∞
f (t)g(s − t)dt ≥Z s
0
F (t)e−tG(s − t)e−s+tdt = e−sZ s
0
F (t)G(s − t)dt.
6
Let F+ = F · I[0,+∞), G+ = G · I[0,+∞), where I[0,+∞) is the characteristic function of the interval
[0, +∞). If we assume that f ∗ g ∈ Lp(R, ω), then from the formula above F+ ∗ G+ ∈ Lp(R). Since
for every F+, G+ ∈ Lp([0, +∞)) we have e−tF+,e−tG+ ∈ Lp(R, ω), it follows that Lp([0, +∞))
is a convolution algebra if Lp(R, ω) is so. But it is well-known that this is not true, if p 6= 1.
Nevertheless, by 1.2.1 and 1.2.2, Lp(R, ω1) is an algebra for ω1(x) = (1 + x)Dex.
There are, moreover, super-exponential weights with which Lp(G, ω) is an algebra. One example
is ω(x) = ex2
on the real line, found by El Kinani [Ki]. But this weight is not submultiplicative.
2 First properties of weighted algebras
2.1 Known inequalities
The conditions (LPAlg) guarantee that Lp(G, ω) is an algebra, that Lp(G, ω) is translation-
invariant and kxfkp,ω ≤ ω(x)kfkp,ω, where xf (t) = f (x−1t), as
kxfkp
p,ω =Z f (x−1t)pω(t)pdt =Z f (y)pω(xy)pdy ≤ ω(x)pkfkp
p,ω.
Also under (LPAlg) the following is known:
• L1(G, ω) ⊂ L1(G), and kfk1 ≤ kfk1,ω for all f ∈ L1(G, ω)
• Lp(G, ω) ⊂ L1(G), and kfk1 ≤ Ckfkp,ω for all f ∈ Lp(G, ω), with C =(cid:0)R ω−q(x)dx(cid:1)1/q
• L1(G, ω) ∗ Lp(G, ω) ⊂ Lp(G, ω), and
kf ∗ gkp,ω ≤ kfk1,ωkgkp,ω
for all f ∈ L1(G, ω), g ∈ Lp(G, ω)
(2.1)
If ω(x) = ω(x−1), the usual involution f ∗(x) = f (x−1) is an isometry on Lp(G, ω): kf ∗kp,ω =
kfkp,ω (recall that every group of polynomial growth is unimodular).
2.2 Approximate units
Under the assumption that ω is continuous, the proof of ([Ku2]) of the fact that the measurable,
bounded functions of compact support are dense in Lp(G, ω), shows that the same is also true
for the set Cc(G) of continuous functions with compact support. By ([Ku2]), there exists a net
(fs)s of measurable, bounded functions with compact support which form a bounded approximate
identity in L1(G, ω) and an (unbounded) approximate identity in Lp(G, ω). In fact, in the same
way it can be proved that if Vs runs through a basis of compact, symmetric neighbourhoods of
the identity e in G, then every family fs such that 0 ≤ fs ≤ 1, kfsk1 = 1, suppfs ⊂ Vs, is an
approximate identity with properties as above. It is easy to see that these functions fs may be
chosen to be continuous and self-adjoint, fs = f ∗
s . Moreover, it will be convenient to have Vs ⊂ K,
where K is a fixed compact set.
2.3 Representations of the weighted algebras
For every non-degenerate ∗-representation T of Lp(G, ω) in a Hilbert space H, there is a unitary
continuous representation V of G such that
T (f ) =ZG
V (x)f (x)dx
(2.2)
for all f . This is proved exactly like in [H-R, Theorem 22.7], though the assumption (ii) of this
theorem does not hold. Let us make the following remarks on the proof.
(1) If T is cyclic with the cyclic vector ξ, one defines V as follows: on the dense subspace of vectors
of the type T (f )ξ, f ∈ Lp(G, ω), put V (x)(cid:0)T (f )ξ(cid:1) = T (xf )ξ; it may be easily shown that this is
7
an isometry, so V (x) extends to a unitary operator on H. It is also straightforward that V is a
representation.
(2) To get the equality (2.2), we need to prove that every coefficient x 7→ hV (x)T (f )ξ, ηi, where
f ∈ Lp(G, ω), and ξ, η ∈ H, is measurable (this is [H-R, 22.3i]). But we have even more: coefficients
of V are continuous since by [Ku2] the mapping x 7→ xf from G to Lp(G, ω) is continuous for every
f . From this, by [H-R, 22.3] we get some representation T of Lp(G, ω) defined by
T (f ) =ZG
V (x)f (x)dx
(2.3)
(3) To prove that T = T , take a vector η ∈ H and the cyclic vector ξ ∈ H. For the linear functional
H(f ) = hT (f )ξ, ηi on Lp(G, ω) there is a function h ∈ Lq(G, ω) such that H(f ) = R f (x)h(x)dx.
Then for any f, g ∈ Lp(G, ω) we have, with all integrals absolutely converging,
h T (f )T (g)ξ, ηi =ZGhV (x)f (x)T (g)ξ, ηidx =ZGhV (x)T (g)ξ, ηif (x)dx
xg(y)h(y)dyã f (x)dx
h(y)(f ∗ g)(y)dy
=ZGhT (xg)ξ, ηif (x)dx =ZG
=ZG
= hT (f ∗ g)ξ, ηi = hT (f )T (g)ξ, ηi.
g(x−1y)h(y)dyã f (x)dx =ZG
ÅZG
ÅZG
It follows that T = T on the dense subspace of vectors of the type T (f )ξ, f ∈ Lp(G, ω), and as a
consequence on the whole H.
(4) In general, T can be expanded into a direct sum of cyclic representations Tα [H-R, 21.13].
Every Tα is given by the formula (2.3) with some Vα; then T is equal to the same integral (2.3)
with V = ⊕Vα.
Further, by [H-R, 22.6], T and V are irreducible or not simultaneously.
In particular, this gives us the identification ÿ Lp(G, ω) = "G.
3 Symmetry
3.1
The notion of symmetry plays an important role in the theory of Banach ∗-algebras. It may be
defined as follows:
Let A be a Banach ∗-algebra and let a ∈ A. We will denote the spectrum of a in A by σA(a)
and the spectral radius of a in A by rA(a). Then the algebra A is said to be symmetric if
σA(a∗a) ⊂ [0, +∞[ for all a ∈ A, or, equivalently, if σA(a) ⊂ R for all a = a∗ ∈ A.
For abelian Banach ∗-algebras the symmetry is equivalent to the fact that all the characters of
the algebra are unitary.
Let (G, ω) satisfy (LPAlg). Let us recall the following definitions for the weight ω (see for instance
[Fe-Gro-Lei-Lu-Mo] and [Fe-Gro-Lei]).
Definition 3.1. a) The weight ω on G is said to satisfy the GRS-condition (or GNR-condition),
if
lim
n→+∞
ω(xn)
1
n = 1,
∀x ∈ G.
(GRS)
b) The weight ω is said to satisfy condition (S), if, for every generating, relatively compact neigh-
bourhood U of G,
1
n = 1.
(S)
lim
n→+∞
sup
x∈U n
ω(x)
8
These conditions are linked to the symmetry of weighted group algebras. Among others, the
following results are known:
For G = Z, l1(Z, ω) is symmetric if and only if limn→+∞ ω(n)
If G ∈ [P G], then L1(G) is symmetric ([Lo]).
If G ∈ [P G] and ω satisfies condition (S), then L1(G, ω) is symmetric ([Fe-Gro-Lei-Lu-Mo]).
The final version of results of this type is due to Fendler, Grochenig and Leinert ([Fe-Gro-Lei]).
They prove:
n = 1 ([Nai]).
1
Theorem 3.2. Let G ∈ [P G] and let ω be a weight on G. Then the following are equivalent:
(i) ω satisfies the GRS-condition.
(ii) ω satisfies condition (S).
(iii) L1(G, ω) is symmetric.
(iv) σL1(G,ω)(f ) = σL1(G)(f ),
∀f ∈ L1(G, ω).
The last three results are based on a method developped by Ludwig ([Lu]). Previously, using a
result of Hulanicki ([Hu1]), Pytlik ([Py2]) had already proved the following:
Theorem 3.3. If the weight ω satisfies
ω(xy) ≤ C(ω(x) + ω(y)),
∀x, y ∈ G
(3.1)
for some positive constant C and if ω−1 ∈ Lp(G) for some 0 < p < +∞, then L1(G, ω) is
symmetric.
Pytlik calls a weight satisfying (3.1) a polynomial weight. In particular, weights of the form
for some positive D, where
ω(x) = (1 + x)D,
x := inf{n x ∈ U n},
satisfy ω(xy) ≤ C(ω(x)+ ω(y)) for all x, y ∈ G, and hence give symmetric weighted group algebras
L1(G, ω), as Pytlik already noticed in ([Py1]).
By ([Py2]) every weight satisfying ω(xy) ≤ C(cid:0)ω(x) + ω(y)(cid:1) is dominated by a weight of the form
K(1 + x)D, K ≥ 1, D > 0. It is then easy to check that all these weights satisfy condition (S).
So the result of Pytlik is a particular case of the result of Fendler, Grochenig and Leinert.
3.2
Our aim is to study symmetry for weighted Lp-algebras. For the rest of this section we hence
assume that (G, ω) satisfies (LPAlg), in order to be sure that Lp(G, ω) is an algebra. The question
is whether condition (S) will also imply the symmetry of Lp(G, ω). We need some preliminary
result.
Lemma 3.4. The weight ω satisfies condition (S) if and only if ω(x) = O(eεx) for all ε > 0,
where x = inf{n x ∈ U n}.
Proof. Let us assume that ω(x) ≤ C(ε)eεx for some constant C(ε), ε > 0. Then
and
x ∈ U k ⇒ x ≤ k
⇒ ω(x) ≤ C(ε)eεk
ω(x)
sup
x∈U k
1
k ≤ C(ε)
1
k eε.
9
So
1 ≤ lim
k→+∞
sup
x∈U k
ω(x)
1
k ≤ eε.
As this has to be true for all ε > 0, limk→+∞ supx∈U k ω(x)
Conversely, let us assume that there exists ε > 0 such that ω(x) is not O(eεx). So, for every
k ∈ N, there exists xk ∈ G such that ω(xk) > keεxk > k. As ω is bounded on each U n, this
implies that "xk → ∞", which means the following: Let n(k) := xk. Then the sequence (n(k))k
admits a subsequence (n(r)) such that
k = 1 and ω satisfies (S).
1
if r → +∞
n(r) → +∞,
xr ∈ U n(r) \ U n(r)−1
ω(xr) > reεn(r) > r
sup
ω(x)
x∈U n(r)
1
n(r) ≥ ω(xr)
1
n(r) > r
1
n(r) eε ≥ eε.
limr→+∞ sup
ω(x)
x∈U n(r)
1
n(r) ≥ eε > 1.
Hence
Thus ω does not satisfy condition (S).
We may now prove the symmetry result for abelian groups:
Theorem 3.5. Let G be an abelian group such that (G, ω) satisfies (LPAlg). Let us assume that
ω satisfies condition (S). Then Lp(G, ω) is a symmetric Banach ∗-algebra.
Proof. According to ([Ku1]), every character χ of Lp(G, ω) is of the form
χ(f ) =ZG
f (x)σ(x)dx =ZG(cid:16)f (x)ω(x)(cid:17)(cid:16)σ(x)ω−1(x)(cid:17)dx,
∀f ∈ Lp(G, ω),
where σ is a (possibly unbounded) character of the group G. As f ω ∈ Lp(G) is arbitrary,
σ(x)ω(x)−1 ∈ Lq(G), with 1
q = 1, and, as ω(x) ≤ C(ε)eεx for all ε > 0,
p + 1
σ(x)e−εx ≤ σ(x)C(ε)ω(x)−1 ∈ Lq(G).
Let us assume that σ is not unitary. Then there exists x0 ∈ G and δ > 0 such that σ(x0) > 1 + δ.
By continuity, there is a non-empty open subset V of G such that σ(x) > 1 + δ
2 , for all x ∈ V .
Hence, for x ∈ V n, x = x1 · x2 ··· xn with xj ∈ V for all j, and σ(x) =Qn
2 )n.
On the other hand, if V ⊂ U k where U is a (relatively compact) generating neighbourhood of the
identity then V n ⊂ U kn, and so x ≤ kn for x ∈ V n. This implies that, for all n (and for all
ε > 0),
j=1 σ(xj ) ≥ (1 + δ
dx
(3.2)
+ ∞ > ZG(cid:16)σ(x)e−εx(cid:17)q
≥ ZV n σ(x)qe−εqxdx
≥ (cid:0)1 +
≥ (cid:16)(cid:0)1 +
2(cid:1) · e−εk > 1, then (3.5) tends to +∞ with n. This is a
2(cid:1)qnZV n
2(cid:1) · e−εk(cid:17)qn
e−εqkndx
V .
(3.3)
δ
δ
(3.4)
(3.5)
contradiction which shows that σ, and hence χ are unitary. So Lp(G, ω) is symmetric.
But we can choose ε so that (cid:0)1 + δ
Example: L2(R, e√x) is a symmetric Banach ∗-algebra (section 1.2.3).
10
3.3
Before studying the non-abelian case, let us first recall the generalized Minkowski relation: Let
X, Y be measure spaces and let F be a measurable function on X × Y . Then, for all p ≥ 1,
ZX(cid:16)ZY F (x, y)dy(cid:17)p
dx!
1
p
≤ZY (cid:16)ZX F (x, y)pdx(cid:17)
1
p
dy.
We need the following relation:
Lemma 3.6. Let us assume that the weight ω satisfies (LPAlg) and is polynomial in the sense of
Pytlik, i. e. that it satisfies
for some constant C > 0. Then
ω(xy) ≤ C(cid:16)ω(x) + ω(y)(cid:17),
∀x, y ∈ G,
kf ∗ gkp,ω ≤ C(cid:16)kfkp,ω kgk1 + kgkp,ω kfk1(cid:17),
∀f, g ∈ Lp(G, ω) ⊂ L1(G).
Proof.
kf ∗ gkp,ω = (cid:16)ZG ZG
≤ C(cid:16)ZG ZG
≤ C(cid:16)ZG ZG
f (y)g(y−1x)dypω(x)pdx(cid:17)
f (y)g(y−1x)(cid:0)ω(y) + ω(y−1x)(cid:1)dypdx(cid:17)
+ C(cid:16)ZG ZG
f (y)g(y−1x)ω(y)dypdx(cid:17)
1
p
= I + II
1
p
1
p
f (y)g(y−1x)ω(y−1x)dypdx(cid:17)
1
p
by the triangle inequality for k · kp. Then the generalized Minkowski inequality implies that
1
p
I ≤ C(cid:16)ZG(cid:0)ZG f (y)g(y−1x)ω(y)dy(cid:1)p
dx(cid:17)
≤ C(cid:16)ZG(cid:0)ZG f (xu−1)g(u)ω(xu−1)du(cid:1)p
dx(cid:17)
≤ CZG(cid:16)ZG f (xu−1)pg(u)pω(xu−1)pdx(cid:17)
= Ckfkp,ω kgk1
1
p
(y = xu−1)
1
p
du
and
II ≤ C(cid:16)ZG(cid:0)ZG f (y)g(y−1x)ω(y−1x)dy(cid:1)p
dx(cid:17)
≤ CZG(cid:16)ZG f (y)pg(y−1x)pω(y−1x)pdx(cid:17)
= Ckgkp,ω kfk1.
1
p
1
p
dy
We may now use the methods of Pytlik ([Py1], [Py2]) to show the symmetry of the algebra Lp(G, ω)
for polynomial weights. This is done via the following lemmas.
Lemma 3.7. Let (G, ω) satisfy (LPAlg). Then, for any f ∈ Lp(G, ω) ⊂ L1(G), r1(f ) ≤ rp,ω(f ),
where r1(f ) denotes the spectral radius of f in L1(G) and rp,ω(f ) denotes the spectral radius of f
in Lp(G, ω).
11
Proof. From
we deduce
kfk1 =ZG f (x)ω(x)ω−1(x)dx ≤(cid:0)ZG
n kf ∗nk
kf ∗nk
1
n
1
1
n
p,ω ,
where f ∗n = f ∗ f ∗ ··· ∗ f (n factors). Hence, for n → +∞,
1 ≤ C
ω−q(x)dx(cid:1)
1
q kfkp,ω = Ckfkp,ω
r1(f ) ≤ rp,ω(f ).
Lemma 3.8. Let (G, ω) satisfy (LPAlg). Let us assume that ω is polynomial in the sense of
Pytlik, i. e. that
Then
ω(xy) ≤ C(cid:0)ω(x) + ω(y)(cid:1),
r1(f ) = rp,ω(f ),
∀x, y ∈ G.
∀f ∈ Lp(G, ω).
Proof. By the methods of Pytlik ([Py2]), (3.6) gives
and, by induction,
So,
kf ∗ fkp,ω ≤ 2Ckfkp,ω kfk1
kf ∗2n
kp,ω ≤ (2C)nkfkp,ω kfk2n−1
1
.
rp,ω(f ) =
lim
n→+∞kf ∗2n
p,ω
k2−n
(2C)n·2−n
n→+∞
lim
≤
= kfk1.
kfk2−n
p,ω kfk1−2−n
1
Finally,
and
We finally get:
rp,ω(f ) = rp,ω(f ∗n)
1
n ≤ kf ∗nk
1
n
1 ,
∀n,
rp,ω(f ) ≤ lim
n→+∞kf ∗nk
1
n
1 = r1(f ).
Theorem 3.9. Let (G, ω) satisfy (LPAlg). Let us assume that ω is polynomial in the sense of
Pytlik, i. e. that
ω(xy) ≤ C(cid:0)ω(x) + ω(y)(cid:1),
∀x, y ∈ G.
Then Lp(G, ω) is a symmetric Banach ∗-algebra.
Proof. This follows from Lemma 3.1 of ([Fe-Gro-Lei]) applied to A := Lp(G, ω) and B := L1(G),
and from the result of Losert ([Lo]) about the symmetry of L1(G). As a matter of fact, these
results imply that
σL1(G)(f ) = σLp(G,ω)(f ),
∀f = f ∗ ∈ Lp(G, ω),
and, as L1(G) is symmetric, σL1(G)(f ) ⊂ R.
Example: We know that for D > 0 sufficiently large, ω(x) = (1 + x)D, where x = inf{n x ∈
U n}, gives rise to an Lp-algebra (1.2.1). Hence this algebra is symmetric.
One may conjecture that, more generally, Lp(G, ω) is a symmetric Banach ∗-algebra, if (G, ω)
satisfies (LPAlg) and ω satisfies condition (S), or, equivalently, the GRS-condition. But this
remains an open question.
12
4 Functional calculus
4.1
Let (G, ω) satisfy (LPAlg). The aim of the following section is the construction of functional
calculus for all continuous functions f with compact support such that f = f ∗, where f ∗(x) =
f (x−1), for all x ∈ G. We will follow the method of ([Hu2], [Dz-Lu-Mo]) and use their results. To
use this method, we have to bound
u(nf ) :=
nkf ∗k
ik
k!
∞
Xk=1
in Lp(G, ω) and show that there are "enough" functions ϕ : R → R, periodic with period 2π, with
ϕ(0) = 0, such that
ϕ{f} := Xn∈Z
u(nf ) ϕ(n)
converges in Lp(G, ω). For more details on functional calculus, see among others ([Di], [Dz-Lu-Mo]).
Let us first recall that, for all continuous functions g with compact support, kgk1 ≤ kgk1,ω and
kgk1 ≤ Ckgkp,ω for some positive constant C. Moreover,
ku(nf )k ≤
1
k!
+∞
Xk=1
nkkfkk ≤ enkf k
in any convenient norm k · k. So for any continuous function f with compact support such that
f = f ∗, the series defining u(nf ) converges in L1(G), L1(G, ω) and Lp(G, ω) to the same element,
i. e. the notation u(nf ) represents a function belonging to L1(G, ω) ∩ Lp(G, ω) ⊂ L1(G). We
will now deduce a bound for ku(nf )kp,ω from the bound for ku(nf )k1,ω which was established in
([Dz-Lu-Mo]). Let us recall the following notations and facts from ([Dz-Lu-Mo]): There exists a
constant C > 1 such that
where x = inf{n x ∈ U n} for an arbitrary (relatively compact) generating neighbourhood U .
We denote
ω(x) ≤ eCx, ∀x ∈ G,
s(n)
:= sup
x∈U n
ω(x),
∀n ∈ N∗
s(0)
ω1(x)
ω2(x)
s2(n)
:= 1
:= s(x)
:= eCx
:= sup
x∈U n
ω2(x) = eCn.
We also consider an arbitrary increasing function r : N → N, which will be specified later. It is
shown in ([Dz-Lu-Mo]) that there exist positive constants C1, C2 such that
ku(nf )k1,ω ≤ C1(1 + n)(1 + nr(n))
∀n ∈ Z,
(4.1)
where Q denotes the power appearing in the polynomial growth condition of the group G, i. e.
U n ≤ KnQ, for all n ∈ N∗, for some positive constant K.
From the formal representation u(f ) = eif − 1 we get the following identity valid also in the
non-unital case:
Q
2 s(nr(n))eC2(cid:0)
n
s2(r(n))(cid:1),
u(nf ) = einf − 1 = ei(n−1)f ∗ eif − 1 = (u((n − 1)f ) + 1) ∗ (u(f ) + 1) − 1 =
= u((n − 1)f ) ∗ u(f ) + u((n − 1)f ) + u(f ).
13
By induction it follows that
u(nf ) = nu(f ) +
n−1
Xk=1
u(kf ) ∗ u(f ).
From (2.1), we get an estimate
ku(nf )kp,ω ≤ nku(f )kp,ω +
n−1
Xk=1
ku(kf )k1,ωku(f )kp,ω.
Using the bound for ku(kf )k1,ω obtained in ([Dz-Lu-Mo]) and recalled in (4.1), we get
ku(nf )kp,ω ≤ Kn + K1
Xk=1
n−1
(1 + k)(1 + kr(k))
≤ Kn + K1(n + 1)(1 + nr(n))
Q
2 s(cid:0)kr(k)(cid:1)eC2(cid:0)
eC2(cid:0)
Xk=1
2 s(cid:0)nr(n)(cid:1)
n−1
k
s2(r(k))(cid:1)
s2(r(k))(cid:1),
k
Q
for some constants K, K1, as the functions s and r are increasing. (Here 1 + k could be bounded
by n as well, but we choose n+ 1 to comply with assumptions of [Dz-Lu-Mo]). As in ([Dz-Lu-Mo]),
we put r(n) := ln(ln n) + 1, for n ≥ ee. Hence, for k ≥ max(ee, eC),
s2(r(k)) = eCr(k) ≥ eC ln(ln k) = (ln k)C
eC2(cid:0)
s2 (r(k))(cid:1) ≤ e
(ln x)C is increasing for x ≥ eC . This allows to estimate the sum over k by
as the function f (x) = x
(ln k)C ≤ e
(ln n)C ,
and
C2
C2
n
k
k
C2 n
ne
(ln n)C . Moreover, as in ([Dz-Lu-Mo]), for n ≥ ee,
ln(ln n) ≤ r(n) ≤ 2 ln(ln n) ≤ 2n
s(cid:0)nr(n)(cid:1) ≤ s(n)r(n) ≤ s(n)2 ln(ln n).
Finally, noticing that nf = (−n)(−f ), we may compute ku(nf )kp,ω even for negative n (by
replacing the constants depending on f by the sup of the corresponding constants for f and −f ).
So, there exist positive constants A1, A2 (depending on f and ω) such that
ku(nf )kp,ω ≤ A1(1 + n)2(1 + n2)
Q
2 s(n)2 ln(ln n)e
A2(cid:0)
n
(ln n)C(cid:1)
for all n ≥ max(ee, eC). We thus obtain a similar bound as for ku(nf )k1,ω, except that the factor
(1 + n) has been replaced by (1 + n)2. Of course the constants are slightly different too. They
depend on f and ω. We may conclude exactly as in ([Dz-Lu-Mo]).
4.2
Let us recall the non-abelian Beurling-Domar condition (BDna) given by
Xn∈N,n≥ee (cid:0) ln(ln n)(cid:1) ln(cid:0)s(n)(cid:1)
1 + n2
< +∞.
It is independent of the choice of the generating neighbourhood U used to compute s(n). See
([Dz-Lu-Mo]) for more details on that condition. We then have the following result:
14
Theorem 4.1. Let (G, ω) satisfy (LPAlg). Let us assume that moreover the weight ω satisfies the
(BDna) condition. Let f = f ∗ be a continuous function with compact support. Then, given a, b, ε
such that 0 < a < a + ε < b − ε < b < 2π, there exists a function ψ : R → R, continuous, periodic
of period 2π such that suppψ ∩ [0, 2π] ⊂ [a, b], ψ ≡ 1 on [a + ε, b − ε] and
Xn∈Zku(nf )kp,ω ψ(n) < +∞.
Hence this defines a function
ψ{f} := Xn∈Z
ψ(n)u(nf ) ∈ Lp(G, ω) ∩ L1(G, ω)
and the properties of functional calculus are satisfied, i. e.
for every character χ of the abelian Banach ∗-subalgebra of Lp(G, ω) generated by f ,
χ(ψ{f}) = ψ(χ(f ))
π(ψ{f}) = ψ(π(f )),
(ϕψ){f} = ϕ{f} ∗ ψ{f},
∀π ∈ ÿ Lp(G, ω) ≡ G,
if the functions ϕ and ϕψ still have the correct properties to allow functional calculus.
Proof. See ([Dz-Lu-Mo]), pages 337 to 345. Here we use again the argument that if a series
converges in L1(G, ω) and Lp(G, ω), then it also converges in L1(G) and the limit is the same in
the three spaces.
4.3
Examples: a) If G ∈ [P G] and ω(x) = K(1 + x)D for K ≥ 1 and D > 0 large enough, then
(G, ω) satisfies (LPAlg) (1.2.1). It is easy to check, that ω also verifies (BDna) and so functional
calculus exists.
b) Let (G, ω) satify (LPAlg) and let us assume that ω(xy) ≤ C(ω(x) + ω(y)) for all x, y ∈ G.
By ([Py2]), such a weight is bounded by a weight of the form K(1 + x)D and hence (BDna) is
verified. Functional calculus exists.
c) If G ∈ [P G], then
0 < γ < 1,
ω(x) := eCxγ
,
is such that (G, ω) satisfies (LPAlg) (1.2.3) and the weight ω verifies (BDna) ([Dz-Lu-Mo]). Func-
tional calculus exists.
Remarks: a) Condition (BDna) is independent of the choice of the generating neighbourhood U .
b) Condition (BDna) is only slightly more restrictive than the well known Beurling-Domar condi-
tion in the abelian case.
c) If ω satisfies (BDna), it also verifies condition (S).
See ([Dz-Lu-Mo]) for more details.
Functional calculus is a very useful tool to prove different harmonic analysis properties, as will be
shown in the rest of this paper.
5 Regularity
5.1
For abelian Banach algebras, regularity is defined as follows by Silov [Si]: Let A be an abelian
Banach algebra and let ∆(A) denote the space of characters of A. Then A is said to be regular
15
if, given any ϕ ∈ ∆(A) and any closed set F ⊂ ∆(A) not containing ϕ, there exists x ∈ A such
that x(ϕ) = ϕ(x) = 1 and xF ≡ 0, where x denotes the Gelfand transform of x.
In the non-abelian case ∆(A) should be replaced by the space P rim∗A defined as the set of all
kernels of topologically irreducible ∗-representations of A. The set P rim∗A is equipped with the
hull-kernel topology.
5.2
As previously, we assume that (G, ω) satisfies (LPAlg). Let Cc(G) denote the set of continuous
functions with compact support on G. It is obvious that Cc(G) ⊂ Lp(G, ω) ⊂ L1(G), that Cc(G)
is dense in Lp(G, ω) and in L1(G), that for all π ∈ G ≡ ÿ Lp(G, ω),
kπ(f )kop ≤ kfk1 ≤ Ckfkp,ω,
∀f ∈ Lp(G, ω),
resp. kπ(f )kop ≤ kfk1 for all f ∈ L1(G). Moreover, we have seen in the previous section
that functional calculus is possible on the self-adjoint elements of Cc(G), provided the weight ω
satisfies (BDna). This imlies that the arguments of ([Dz-Lu-Mo], pages 350 and 351) remain valid.
In particular, we have the following results:
Theorem 5.1. Let (G, ω) satisfy (LPAlg). Let us assume that the weight ω verifies (BDna). We
then have:
(i) The map
Ψ : Prim∗L1(G) → Prim∗Lp(G, ω)
kerπ 7→ kerπ ∩ Lp(G, ω)
is a homeomorphism.
(ii) In particular, Prim∗Lp(G, ω), Prim∗L1(G, ω) and Prim∗L1(G) are homeomorphic.
(iii) Given any ρ ∈ G and any open neighbourhood N of ρ in G, resp. any open neighbourhood N1
of kerρ ∩ Lp(G, ω) in Prim∗Lp(G, ω), there exists f ∈ Lp(G, ω) such that ρ(f ) 6= 0 and π(f ) = 0
for all π ∈ G \ N , resp. for all π such that kerπ ∩ Lp(G, ω) ∈ Prim∗Lp(G, ω) \ N1.
Proof. See ([Dz-Lu-Mo]). Part of the argument relies heavily on functional calculus and on the
∗-regularity of groups with polynomial growth.
Point (iii) of the previous theorem, which is often called Domar's property, corresponds to the
regularity of abelian Banach algebras.
6 Weak Wiener property
Let us recall the following definitions:
Definition 6.1. Let A be a Banach algebra.
(i) A representation (T, V ) of A on a vector space V is said to be algebraically irreducible, if there
are no non-trivial T -invariant subspaces in V .
(ii) The algebra A is said to have the weak Wiener property, if every proper closed two-sided ideal
of A is contained in the kernel of an algebraically irreducible representation.
Let (G, ω) satisfy (LPAlg) and let (fs)s be an approximate unit of Lp(G, ω) with the properties
discussed in 2.2.
In ([Dz-Lu-Mo]) it is shown that, provided ω satisfies (BDna), there exists a periodic function ϕ
of period 2π with ϕ(1) = 1, ϕ ≡ 0 in a neighbourhood of 0, such that ϕ{fs} is defined in L1(G, ω).
By our section on functional calculus in Lp(G, ω), the same ϕ{fs} also converges in Lp(G, ω), i.
e. ϕ{fs} ∈ L1(G, ω) ∩ Lp(G, ω) for all s. Moreover, in ([Dz-Lu-Mo]) it is shown that
kϕ{fs} ∗ f − fk1,ω → 0
16
for all continuous functions f with compact support in G. Hence, for any f, g ∈ Cc(G), f ∗ g ∈
Cc(G) ⊂ L1(G, ω) ∩ Lp(G, ω) and
kϕ{fs} ∗ f ∗ g − f ∗ gkp,ω ≤ kϕ{fs} ∗ f − fk1,ωkgkp,ω.
So
This gives the following result:
kϕ{fs} ∗ f ∗ g − f ∗ gkp,ω → 0.
(6.1)
Lemma 6.2. Under the assumptions above, let I be a proper closed two-sided ideal of Lp(G, ω).
Then there exists s such that ϕ{fs} /∈ I.
Proof. Let us assume that ϕ{fs} ∈ I, for all s. Then ϕ{fs}∗ f ∗ g ∈ I for all s and all f, g ∈ Cc(G).
As I is closed, the relation (6.1) shows that f ∗ g ∈ I for all f, g ∈ Cc(G). But this implies that
I = Lp(G, ω), by density, which is a contradiction.
We are now able to prove the weak Wiener property:
Theorem 6.3. Let (G, ω) be (LPAlg). Let us also assume that the weight ω satisfies (BDna).
Then the algebra Lp(G, ω) has the weak Wiener property.
Proof. The proof is standard, but we repeat it for the sake of completeness. Let I be a proper,
closed, two-sided ideal of Lp(G, ω). Let s and ϕ be such that ϕ{fs} /∈ I. Let ψ be another function
such that functional calculus ψ{fs} is possible and such that ψ ≡ 1 on the support of ϕ. Such a
ψ exists by theorem 4.1. Then
Let us consider the algebra A := Lp(G, ω)/I. We have
ψ{fs} ∗ ϕ{fs} = (ψϕ){fs} = ϕ{fs}.
ϕ{fs} ∈ A
0 6=
ψ{fs} − 1) ∗
(
ϕ{fs} = 0 in A ⊕ C,
where the dot denotes the equivalence class in the quotient space Lp(G, ω)/I. Hence, by ([Bo-Du])
ψ{fs} /∈ rad(A), where rad(A)
ψ{fs} − 1 is not invertible in A ⊕ C, i. e. 1 ∈ σA(
denotes the radical of A. This implies that there exists an algebraically irreducible representation
ψ{fs}) 6= 0. We then define the non-trivial algebraically irreducible
( T , V ) of A such that T (
representation (T, V ) of Lp(G, ω) by T (f ) := T ( f ). By construction, I ⊂ kerT . Hence Lp(G, ω)
is weakly Wiener.
ψ{fs}). So
7 Wiener property
We start with the following definition:
Definition 7.1. Let A be a Banach ∗-algebra.
(i) A representation (T, V ) of A on a Banach space V is said to be topologically irreducible, if
there are no non-trivial closed T -invariant subspaces in V . In particular, this definition is applied
to unitary representations (T,H) on Hilbert spaces H.
(ii) The algebra A is said to have the Wiener property, if every proper closed two-sided ideal in A
is contained in the kernel of a topologically irreducible unitary representation of A.
It is well known that every symmetric Banach ∗-algebra which has the weak Wiener property, also
has the Wiener property (see [Le1] and [Le2]). This leads us to the following result:
17
Theorem 7.2. Let (G, ω) satisfy (LPAlg). Then the algebra Lp(G, ω) has the Wiener property in
the following cases:
a) G is abelian and ω satisfies (BDna).
b) G is non abelian and ω(xy) ≤ C(ω(x) + ω(y)), for all x, y ∈ G, for some constant C > 0.
c) G is non abelian and ω(x) = K(1 +x)D for D large enough, with x = inf{n x ∈ U n}, where
U is an arbitrary, relatively compact, generating neighbourhood of the identity.
If one could prove that property (S) implies the symmetry of the algebra Lp(G, ω) (conjecture),
then the property (BDna) would imply the Wiener property.
8 Minimal ideals of a given hull
As always, we assume that (G, ω) satisfies (LPAlg). We also suppose that ω satisfies (BDna). For
details of the following we refer to ([Dz-Lu-Mo]). We will use the following notations:
Let us denote by Φ the set of functions ϕ from R to R, periodic of period 2π, with ϕ(0) = 0, with
suppϕ ∩ [0, 2π] compact contained in ]0, 2π[, which operate on the set of continous, self-adjoint
functions with compact support in the algebras L1(G, ω) and Lp(G, ω). The construction of such
functions is described in more details in ([Dz-Lu-Mo]). See also theorem 4.1.
For any closed ideal I of Lp(G, ω), we define the hull of I by
h(I) := {kerπ ∈ Prim∗Lp(G, ω) I ⊂ kerπ}.
For any compact subset C of Prim∗Lp(G, ω) (endowed with the hull-kernel topology), let us define
C := {π ∈ G kerπ ∈ C}
and
kfkC := sup
π∈ C kπ(f )kop
m(C)
:= {ϕ{f} f = f ∗, f ∈ Cc(G),kfk1 ≤ 1, ϕ ∈ Φ, ϕ ≡ 0 on a neighbourhood of [−kfkC,kfkC]}
Let j(C) be the closed two-sided ideal of Lp(G, ω) generated by m(C). As in ([Dz-Lu-Mo]), one
may prove:
Lemma 8.1. The hull of j(C) is C.
Proof. See ([Dz-Lu-Mo]).
Theorem 8.2. Let (G, ω) satisfy (LPAlg). Let the weight ω satisfy (BDna). We also assume that
the algebra Lp(G, ω) is symmetric. Let C be a closed subset of Prim∗Lp(G, ω) (≡ Prim∗L1(G, ω) ≡
Prim∗L1(G)). There exists a closed two-sided ideal j(C) of Lp(G, ω) with h(j(C)) = C, which is
contained in every two-sided closed ideal I with h(I) = C.
This is in particular the case if G is either abelian and ω satisfies (BDna) or if G is non-abelian
and ω satisfies the relationship ω(xy) ≤ C(cid:0)ω(x)+ω(y)(cid:1) for all x, y ∈ G, for some positive constant
C.
Proof. See ([Dz-Lu-Mo]).
9 A symmetric algebra having infinite-dimensional irreducible
representations
Using C. Read's example of a quasi-nilpotent operator on ℓ1 with no nontrivial invariant subspaces,
one can construct an example of a weighted algebra Lp(R, ω) for any p > 1 which is symmetric
and has infinite-dimensional topologically irreducible representations. Up to our knowledge, the
first application of this type is due A. Atzmon ([Atz]).
Let T : ℓ1 → ℓ1 be the operator constructed in [Re]. It is quasi-nilpotent, so that
kTk ≤ 1,
kT nk1/n → 0, n → ∞.
(9.1)
18
9.1 Definition of the weights and symmetry
We will introduce a family of weights on R which will depend on p ≥ 1. For p = 1, we take as a
weight
(9.2)
ω(x) = max{ kexTk, ke−xTk}.
Obviously, ω is submultiplicative. Thus, L1(R, ω) is an algebra. For p > 1, we put ω1(x) =
ω(x)(1+x)2; by sections 1.2.1 and 1.2.2, this is an Lp-algebra weight (possibly after multiplication
by a constant).
Next we prove the following estimate:
Lemma 9.1. Let ω be defined by (9.2). Then ω(x) = O(exp(εx)), x → ∞, for any ε > 0.
Proof. Let ε > 0 be given. It is enough to estimate ω(x) for x > 0. By (9.1), there is N (ε) such
that kT nk < εn for all n ≥ N (ε). Separate the series into two parts: ω(x) = Ω1(x) + Ω2(x), where
Ω1(x) = Xn<N (ε)
Ω2(x) = Xn≥N (ε)
kT nkxn
n!
kT nkxn
n!
,
.
For Ω2, we have:
For Ω1:
Now,
εnxn
n! ≤ exp(εx).
Ω2(x) ≤ Xn≥N (ε)
n!εn ≤ ε−N (ε) Xn<N (ε)
xnεn
Ω1(x) ≤ Xn<N (ε)
xnεn
n! ≤ ε−N (ε) exp(εx).
what proves the lemma.
ω(x) ≤ exp(εx)(cid:0)1 + ε−N (ε)(cid:1) = C(ε) exp(εx),
Corollary 9.2. The algebra L1(R, ω) and every algebra Lp(R, ω1) with p > 1 are symmetric.
Proof. It is easy to see that ω1(x) = O(exp(εx)) for any ε > 0 as well. By lemma 3.4, ω and
ω1 satisfy condition (S), so by theorems 3.2 and 3.5 the algebras L1(R, ω) and Lp(R, ω1) are
symmetric.
9.2 An infinite-dimensional irreducible representation
Let A stand for L1(R, ω) if p = 1 and for Lp(R, ω1) if p > 1. Now we can put
U (f ) =ZR
exp(xT )f (x)dx
for any f ∈ A. First of all, we will show that this integral converges absolutely. We can estimate
kU (f )k ≤ZR k exp(T x)k f (x)dx ≤ZR
ω(x)f (x)dx.
If p = 1, this equals kfk1,ω. If p > 1,
ZR
ω(x)f (x)dx =ZR
ω1(x)f (x)(1 + x)−2dx ≤ kω1fkpk(1 + x)−2kq = Cqkfkp,ω1.
In both cases, we see that kU (f )k ≤ CkfkA, so U is continuous. Clearly U is a homomorphism.
19
It remains now to show that U is topologically irreducible. This will follow from the fact that
closed invariant subspaces of U are invariant under T .
Suppose that Z ⊂ ℓ1 is a nonzero invariant subspace of U . Take z ∈ Z, z 6= 0. Let Iε be the
indicator function of [0, ε] and let ξε = ε−1Iε. Then
U (ξε) − I =
=
0
1
ε Z ε
ε Z ε
1
0
∞
exT dx −
(xT )n
Xn=1
n!
1
ε Z ε
0
dx =
Idx =
∞
0
(
1
ε Z ε
Xn=0
ε Z ε
0 (cid:16)xT +
Xn=2
∞
1
(xT )n
n! − I)dx
(xT )n
n! (cid:17)dx.
Now, assuming that 0 < ε < 1, we have
kU (ξε) − I −
ε
2
Tk = k
≤
Thus, ε−1(cid:0)U (ξε) − I(cid:1) − 1
∞
0
0
1
1
1
∞
ε
2
T +
x dx T −
Xn=0
ε Z ε
Xn=2
ε Z ε
ε Z ε
2 T → 0, as ε → 0, or T = lim
ε Z ε
dx ≤
(n + 2)!
xn
x2
ε→0
1
0
0
∞
(xT )n
1
ε Z ε
k(xT )nk
0
1
n!
dx
Xn=2
n! (cid:17)dxk ≤
ε Z ε
x2exdx ≤
2ε−1(cid:0)U (ξε) − I(cid:1). If now Z is an invariant
e x2dx =
< ε2.
e
ε ·
ε3
3
0
subspace for U , then its closure ¯Z is invariant for T , so ¯Z is trivial, and we are done.
Remarks. It is clear that this example can be extended to Rn: replace exT by eη(x)T , where η is
a nonzero linear form on Rn. One can show also that ω(x) > C exp(x/ ln x), and so the algebras
which we construct are not regular.
References
[Atz]
[Al]
[Be]
[Beu1]
[Beu2]
[Bl-Ka-Ra]
[Bo-Du]
[Da-Fo-Gro]
A. Atzmon, Irreducible representations of abelian groups, Lect. Notes Math.
1359 (1987), 83-92.
D. Alexandre, Id´eaux minimaux d'alg`ebres de groupes, PhD thesis, Metz
(2000).
S. N. Bernstein, Le probl`eme de l'approximation des fonctions continues sur
tout l'axe r´eel et l'une de ses applications, Bull. Math. Soc. France, 52 (1924),
399-410.
A. Beurling, Sur les int´egrales de Fourier absolument convergentes et
leur application `a une transformation fonctionnelle, 9`eme Congr`es des
Math´ematiciens scandinaves, aout 1938, Helsinki (1939), 345-366.
A. Beurling, Sur le spectre des fonctions, Analyse Harmonique, Colloques In-
ternationaux du Centre National de la Recherche Scientifique, no. 15, CNRS
Paris (1949), 9-29.
A. Blanco, S: Kaijser, T. J. Ransford, Real interpolation of Banach algebras
and factorization of weakly compact homomorphisms, J. Funct. Anal. 217
(2004), 126-141.
F. F. Bonsall, J. Duncan, Complete normed algebras. Ergebnisse der Mathe-
matik und ihrer Grenzgebiete, Band 80, Springer-Verlag, New York, Heidel-
berg, 1973.
S. Dahlke, M. Fornasier, K. Grochenig, Optimal Adaptive Computations in
the Jaffard Algebra and Localized Frames, J. Approx. Theory 162, no. 1
(2010), 153-185.
20
[Di]
[Do]
J. Dixmier, Op´erateurs de rang fini dans les repr´esentations unitaires, Inst.
Hautes Etudes Sci. Publ. Math. (1960), 13-25.
Y. Domar, Harmonic analysis based on certain commutative Banach algebras,
Acta Math. 96 (1956), 1-66.
[Dz-Lu-Mo]
J. Dziubanski, J. Ludwig and C. Molitor-Braun, Functional Calculus in
Weighted Group Algebras, Rev. Mat. Complut. 17, N´um. 2, (2004), 321-357.
[Erd]
A. Erdelyi, Asymptotic expansions. Dover Publications, 1956.
[Fa-N.Ni-Za]
O. El-Falla, N. K. Nikol'skiı, M. Zarrabi, Estimates for resolvents in Beurling-
Sobolev algebras. St. Petersburg Math. J. 10 (1999), no. 6, 901 -- 964.
[Fei]
[Fe-Gro-Lei]
H. G. Feichtinger, Gewichtsfunktionen auf lokalkompakten Gruppen, Sitzber.
Osterr. Akad. Wiss. Abt. II, 188, no. 8 -- 10 (1979), 451 -- 471.
G. Fendler, K. Grochenig and M. Leinert, Symmetry of Weighted L1-Algebras
and the GRS-Condition, Bull. London Math. Soc. 38, no. 4 (2006), 625-635.
[Fe-Gro-Lei-Lu-Mo] G. Fendler, K. Grochenig, M. Leinert, J. Ludwig and C. Molitor-Braun,
Weighted group algebras on groups of polynomial growth, Math. Z. 245
(2003), 791-821.
[Glu-Sav]
[Gro-Lei]
[Gro-Lei1]
[H-R]
[Hu1]
[Hu2]
[Ki]
V. P. Glushko, Yu. B. Savchenko, Higher-order degenerate elliptic equations:
Spaces, operators, boundary-value problems, J. Math. Sci. 39 no. 6 (1987),
3088-3148.
K. Grochenig, M. Leinert, Wiener's lemma for twisted convolution and Gabor
frames, J. Amer. Math. Soc., 17, no. 1 (2004), 1-18.
K. Grochenig, M. Leinert, Symmetry and inverse closedness of matrix al-
gebras and functional calculus for infinite matrices, Trans. Amer. Math.
Soc. 358, no. 6 (2006), 2695-2711.
E. Hewitt, K. A. Ross, Abstract harmonic analysis I, II. Springer, 3rd printing
(1997).
A. Hulanicki, On the Spectrum of Convolution Operators on Groups with
polynomial Growth, Inventiones math. 17 (1972), 135-142.
A. Hulanicki, A functional calculus for Rockland operators on nilpotent Lie
groups, Studia math. 78 (1984), 253-266.
A. El Kinani, On generalized Beurling algebras. Rend. Circ. Mat. Palermo
(2) 56 (2007), no. 3, 369-380.
[Kud-S.Ni]
L. D. Kudryavtsev, S. M. Nikol'skii, Spaces of differentiable functions of
several variables and imbedding theorems. In: S.M. Nikol'skii (ed.), Analysis
III, Encycl. Math. Sci., 26, Springer (1990), 1140 (Transl. from Russian).
[Ku1]
[Ku2]
[Ku3]
Yu. N. Kuznetsova, Weighted Lp-algebras on groups, Funktsional. Anal. i
Prilozhen. 40 (3) (2006), 82-85; translation in Funct. Anal. Appl. 40 (3)
(2006), 234-236.
Yu. N. Kuznetsova, Invariant weighted algebras Lw
(2008), 567-576.
p (G), Mat. Zametki 84 (4)
Yu. N. Kuznetsova, Constructions of
Sbornik 200 (2) (2009), 75 -- 88.
regular algebras Lw
p (G). Mat.
21
[Ku4]
[Le1]
[Le2]
[Lo]
[Lu]
[Nai]
[N.Ni]
[Py1]
[Py2]
[Re]
[Si]
[Vr]
[We]
Yu. N. Kuznetsova, Example of a weighted algebra Lw
discrete group, J. Math. Anal. Appl. 353 (2009), 660-665.
p (G) on an uncoutable
H. Leptin, On group algebras of nilpotent Lie groups, Studia Math. 47 (1973),
37 -- 49.
H. Leptin, Ideal theory in group algebras of locally compact groups, Inven-
tiones Math. 31 (1976), 259 -- 278.
V. Losert, On the structure of groups with polynomial growth II, J. London
Math. Soc. 63 (2001), 640-654.
J. Ludwig, A class of symmetric and a class of Wiener group algebras, J.
Funct. Anal. 31 (1979), 187-194.
M. A. Naimark, Normed algebras, Wolters-Noordhoff Publishing (1972).
N. K. Nikol'skiı, Selected problems of weighted approximation and spectral
analysis. Trudy Math. Inst. Steklov 120 (1974).
T. Pytlik, On the spectral radius of elements in group algebras, Bull. Acad.
Polon. Sci. S´er. Sci. Math. Astronom. Phys. 21 (1973), 899-902.
T. Pytlik, Symbolic calculus on weighted group algebras, Studia Math. 73
(1982), 169-176.
C. J. Read, Quasinilpotent operators and the invariant subspace problem. J.
London Math. Soc. (2) 56 (1997), 595-606.
G. E. Silov, On regular normed rings. Trav. Inst. Math. Stekloff 21 (1947).
A. Vretblad, Spectral analysis in weighted L1 spaces on R, Ark. Mat. 11
(1973), 109-138.
J. Wermer, On a class of normed rings, Ark. Mat. 2 (1954), Hf. 6, 537 -- 551.
Julia Kuznetsova, Unit´e de Recherche en Math´ematiques, Universit´e du Luxembourg, 6, rue
Coudenhove-Kalergi, L-1359 Luxembourg, Luxembourg, [email protected]
Carine Molitor-Braun, Unit´e de Recherche en Math´ematiques, Universit´e du Luxembourg, 6, rue
Coudenhove-Kalergi, L-1359 Luxembourg, Luxembourg, [email protected]
22
|
1807.02618 | 1 | 1807 | 2018-07-07T06:25:40 | Results on the Spectral Schwartz Distribution | [
"math.FA",
"math.OA"
] | The resolvent of an operator in a Banach space is defined on an open subset of the complex plane and is holomorphic. It obeys the resolvent equation. A generalization of this equation to Schwartz distributions is defined and a Schwartz distribution, which satisfies that equation is called a resolvent distribution. Its restriction to the subset, where it is continuous, is the usual resolvent function. Its complex conjugate derivative is,but a factor, the spectral Schwartz distribution, which is carried by a subset of the spectral set of the operator. The spectral distribution yields a spectral decomposition. The spectral distribution of a matrix and a unitary operator are given. If the the operator is a self-adjoint operator on a Hilbert space, the spectral distribution is the derivative of the spectral family. We calculate the spectral distribution of the multiplication operator and some rank one perturbations. These operators are not necessarily self adjoint and may have discrete real or imaginary eigenvalues or a nontrivial Jordan decomposition. | math.FA | math |
RESULTS ON THE SPECTRAL SCHWARTZ DISTRIBUTION
WILHELM VON WALDENFELS
UNIVERSIT AT HEIDELBERG
Abstract. The resolvent of an operator in a Banach space is defined on an open subset of the
complex plane and is holomorphic.
It obeys the resolvent equation. A generalization of this
equation to Schwartz distributions is defined and a Schwartz distribution, which satisfies that
equation is called a resolvent distribution. Its restriction to the subset, where it is continuous,
is the usual resolvent function.
Its complex conjugate derivative is ,but a factor, the spectral
Schwartz distribution , which is carried by a subset of the spectral set of the operator. The spectral
distribution yields a spectral decomposition. The spectral distribution of a matrix and a unitary
operator are given. If the the operator is a self-adjoint operator on a Hilbert space, the spectral
distribution is the derivative of the spectral family. We calculate the spectral distribution of the
multiplication operator and some rank one perturbations. These operators are not necessarily self
adjoint and may have discrete real or imaginary eigenvalues or a nontrivial Jordan decomposition.
Keywords: Schwartz distributions, Distribution Kernels, Resolvent, Spectral Decomposition,
Spectral Distribution, Multiplication Operator.
AMS classification: 47A10 Spectrum,Resolvent; 46F2046A20 Distributions as Boundary Values
of Analytic Functions, 46F2546A24 Distributions on infinite dimensional spaces.
1. Definition and Basic Properties
1.1. Introduction. In a study of radiation transfer Garyi V. Efimov, Rainer Wehrse and the
author [1] developed a method, how to calculate the spectral decomposition of a multiplication
operator in Rn perturbed by a rank one operator . By Krein's formula it is possible to calculate the
resolvent R(z) and then one investigates the singularities of the function z 7→ hf1R(z)f2i, where
In [9, chapter 3] the complex derivative derivative 1/π∂hf1R(z)f2i is
f1, f2 are test-functions.
called the bracket hf1M (z)f2i of the spectral Schwartz distribution M (z), which therefor was only
scalarly defined. There was developed a rudimentary theory and calculated four explicit examples
[9, chapter 4], originated from the theory of quantum stochastic processes. Inspecting the examples
it turned out, that the resolvent it self, not only the brackets between test functions can be extended
to an operator valued Schwartz distribution and under this much stronger regularity condition a
deeper theory of the spectral Schwartz distribution can be established. We calculate the spectrum of
the multiplication operator, and two of its perturbations by a rank one operator. The examples are
not necessarily self adjoint. In the second example we have imaginary eigenvalues and a nontrivial
Jordan decomposition.
1.2. Schwartz distributions. We recall the definition of a distribution given byf L.Schwartz [6].
A distribution on on an open set G ⊂ Rn is a linear functional T on the space of test functions
D(G), that is the space of infinitely differentiable functions of compact support ( or C∞
c -functions)
with support in G. The functional T is continuous in the following way: if the functions ϕn ∈ D(G)
have their support in a common compact set and if they and all their derivatives converge to 0
1
T (ϕ) =Z T (x)ϕ(x)dx.
The integral notation is the usual notation of physicists and has been adopted by Schwartz with
minor modifications in his articles on vector-valued distributions [7][8]. It emphasizes the fact, that
a distribution can be considered as a generalized function. This is the notation in the Russian
literature [2]. Similar is de Rham's formulation [5] : T can be considered as a current of degree
0 applied to a form of degree n namely ϕ(x1,··· , xn)dx1 ··· dxn. We use variable transforms of
distributions accordingly.
In order to include Dirac's original ideas, L.Schwartz introduced the distribution kernels [7]. A
distribution kernel K(x, y) on Rm × Rn is a distribution on that space. If T is a distribution on Rn
one may associate to it a kernel T (x − y) on Rn × Rn by
ZZ T (x − y)ϕ(x)ψ(y)dxdy =Z dxT (x)ϕ(x)Z dyψ(y − x) =Z dx(T ∗ ψ)(x)ϕ(x),
where T ∗ ψ denotes the convolution. We have often to deal with the so called δ-function given
by δ(ϕ) = ϕ(0) and the generalized function P/x = (d/dx) ln x, where P stands for principal
value.We have
ZZ δ(x − y)ϕ(x, y)dxdy =Z dxϕ(x, x).
uniformly, then the T (ϕn) → 0. The space of distributions on Rn is denoted by D′(Rn). We write
often
So δ(x − y) = δ(y − x).
If S(x, y) and T (y, z) are two kernels one may form a distribution of three variables S(x, y)T (y, z),
if this exists. If the kernels are S(x − y) and T (y − z) then S(x − y)T (y − z) exists and is given by
ZZZ dxdydzS(x − y)T (y − z)ϕ(x, y, z) =ZZZ dudvdyS(u)T (v)ϕ(u + y, y, y − v).
We obtain for any distribution T the relation
(1)
δ(x − y)T (y − z) = δ(x − y)T (x − z).
We have the formulae ( for the second equation see [9, p.74])
(2)
(3)
ZZZ δ(x − y)δ(y − z)ϕ(x, y, z)dxdydz =Z ϕ(x, x, x)dx
P
x − ω
y − ω(cid:19) + π2δ(x − ω)δ(y − ω)
P
y − ω
=
1
y − x(cid:18) P
x − ω − P
1.3. Notion of the Spectral Distribution. Assume we have a Banach space V and denote by
L(V ) the space of all bounded linear operators from V to V provided with the usual operator norm,
Assume a subspace D ⊂ V and an operator A : D → V . Then the complex plane splits into two
sets, the resolvent set , where the resolvent R(z) = (z − A)−1 exists, and the spectral set, where it
does not. The resolvent set is open, the spectral set is closed. In the resolvent set the resolvent
obeys one of the two equivalent resolvent equations
(4)
(5)
R(z1) − R(z2) = (z2 − z1)R(z1)R(z2)
R(z1)R(z2) = 1/(z2 − z1)(R(z1) − R(z2).
Go the other way round and assume an open set G ⊂ C and a function R(z) : G → L(V ) satisfying
the resolvent equation. The function R(z) is holomorphic in G. The subspace D = R(z)V is a
2
subset independent of z ∈ G. If R(z0) is injective for one z0 ∈ G, then R(z) is injective for all z ∈ G
and there exists a mapping A : D → V such that
(z − A)R(z)f = f for f ∈ V
R(z)(z − A)f = f for f ∈ D
The operator A is closed and R(z) is the resolvent of A.[4] There exists not always such an operator
corresponding to a resolvent , e.g. the function R(z) = 0 for all z fulfills the resolvent equation and
R(z) is surely not injective.
We use the second equation (5) and formulate:
Definition 1. Assume a Schwartz distribution ϕ 7→ R(ϕ) on an open set G ⊂ C , which satisfies
the equation
R(ϕ1)R(ϕ2) =Z d2z1R(z1)ϕ1(z1)Z d2z2ϕ2(z2)/(z2 − z1)
−Z d2z2R(z2)ϕ2(z2)Z d2z1ϕ1(z1)/(z2 − z1)
then we say, that R(ϕ) satisfies the distribution resolvent equation and call R(ϕ) a resolvent dis-
tribution. Here d2z = dxdy is the surface elament on the complex plane for z = x + iy.
If R(z) obeys the resolvent equation in G it might be, that there is Schwartz distribution on
an open set G′ ⊃ G extending R(z) and obeying the distribution resolvent equation. Assume a
distribution R(z) satisfying the distribution resolvent equation , whose restriction on an open subset
of the complex plane equals a continuous function, then the restriction satisfies the usual resolvent
equation and is holomorphic. We call the set, where the distribution R(z) is not continuous, the
spectral set. It is closed in G′.
If G ⊂ C is open and f : G → C, f (z) = f (x + iy) has a continuous derivative, set
∂y(cid:19) .
2(cid:18) ∂f
2(cid:18) ∂f
∂f
∂y(cid:19)
∂x − i
1
=
1
=
∂f
+ i
∂f =
df
dz
(6)
∂f =
df
dz
∂x
The operator ∂ is the usual derivative, ∂ is called the complex conjugate derivative. The function
f is holomorphic if and only if ∂f = 0.
In an analogous way one defines these derivatives for Schwartz distributions. If T is a distribution
on C, then ∂T (ϕ) = −T (∂ϕ). The function z 7→ 1/z is locally integrable and one obtains
(7)
Assume f to be defined and holomorphic for the elements x + iy ∈ G, y 6= 0, and that f (x ± i0)
exists and is continuous, then
∂(1/z) = πδ(z).
(8)
∂f (x + iy) = (i/2)(f (x + i0) − f (x − i0))δ(y)
The last equation holds for operator valued functions as well. To prove these assertions , we need
the following variant of Gauss' theorem
Lemma 1. Assume G ⊂ C an open subset , f : G → C continuously differentiable and G0 ⊂ G an
open subset with compact closure and smooth border G′
0, then
ZG0
∂f (z)d2z = −
3
i
2ZG′
0
f (z)dz
Here dz denotes the line element directed in the sense, that the interior of G0 is on the left hand
side
Definition 2. Assume R(z) satisfying the distribution resolvent equation, then we call
the spectral Schwartz distribution of R(z).
M (z) = (1/π)∂R(z)
1.4. Basic properties. If the resolvent distribution R(z) is in an open set the extension of the
resolvent function of an operator A, we call M (z) a spectral distribution of A. Remark, that there
may be many spectral distributions of A. As R(z) is holomorphic on an open subset, where it is
continuous, M (z) is supported by the spectral set of R(z).
Using the abbreviation r(z) = 1/z we may the the distribution resolvent equation write in the
form
(9)
R(ϕ1)R(ϕ2) = −R((r ∗ ϕ1)ϕ2 + ϕ1(r ∗ ϕ2))
The following theorem shows, that the spectral distribution can be considered as the generalization
of the family of eigenprojectors in finite dimensional case. The operator M (z) corresponds to
eigenprojecto of the eigenvalue z. We have a generalized orthogonality relation and the fact, that
the product with the resolvent means inserting the corresponding eigenvalue.
Theorem 1. The spectral distribution is multiplicative. So assume two C∞
c
functions ϕ1, ϕ2, then
M (ϕ1)M (ϕ2) = M (ϕ1ϕ2)
or M (z1)M (z2) = δ(z1 − z2)M (z1).
Conversely assume a distribution M (z) in G ⊂ C obeying the equation M (z1)M (z2) = δ(z1 −
z2)M (z1). Assume that the integral
R(z) =Z d2ζM (ζ)/(z − ζ),
Z R(z)ϕ(z)d2z =Z M (ζ)d2ζZ ϕ(z)/(z − ζ)d2z
exists for ϕ ∈ C∞
c , then R(z) fulfills the resolvent equation for distributions. Furthermore
M (z1)R(z2) = M (z1)/(z2 − z1),
Z M (z1)ϕ1(z1)d2z1Z R(z2)ϕ2(z2)d2z2 =Z M (z1)ϕ(z1)Z ϕ2(z2)/(z2 − z1)d2z2d2z1.
Proof. Use the equation ∂ r = πδ and hence r ∗ ∂ϕ = r ∗ ∂δ ∗ ϕ = ∂δ ∗ r ∗ ϕ = ∂ r ∗ ϕ = πϕ and
calculate
M (ϕ1)M (ϕ2) = π−2R(∂ϕ1)R(∂ϕ2) = −π−2R((r ∗ ∂ϕ1)∂ϕ2 + ∂ϕ1(r ∗ ∂ϕ2))
In order to prove the converse assertion, observe that it means that r∗ M is a resolvent distribution.
Now M (ϕ) = −1/πR(∂ϕ) and (r ∗ M )(ϕ) = −M (r ∗ ϕ). Hence
= −πR(ϕ1∂ϕ2 + (∂ϕ1)ϕ2) = −πR(∂(ϕ1ϕ2)) = M (ϕ1ϕ2).
R(ϕ1)R(ϕ2) = M (r ∗ ϕ1)M (r ∗ ϕ2) = M ((r ∗ ϕ1)(r ∗ ϕ2))
= −1/πR(∂((r ∗ ϕ1)(r ∗ ϕ2)) = −R((r ∗ ϕ1)ϕ2 + ϕ1(r ∗ ϕ2))
4
The last equation says M (ϕ1)R(ϕ2) = −M (ϕ1(r ∗ ϕ2)). In fact
M (ϕ1)R(ϕ2) = −
1
π
R(∂ϕ1)R(ϕ2) =
1
π
R((r ∗ ∂ϕ1)ϕ2 + ∂ϕ1(r ∗ ϕ2))
=
1
π
R(πϕ1ϕ2 + ∂ϕ1(r ∗ ϕ2)) =
1
π
R(∂(ϕ1(r ∗ ϕ2)) = −M (ϕ1(r ∗ ϕ2)).
The operator M (z) is a generalized eigen-projector. In fact if R(z) = 1/(z − A) then
M (z1)(1/(z2 − A) = M (z1)/(z2 − z1). The relation M (z1)M (z2) = δ(z1 − z2)M (z1) is a generalized
orthonormality relation.
Remark 1. If A ∈ L(V ) is an operator and if the distribution R(z) fulfills the equation AR(z) =
−1 + zR(z) in the sense of distributions, then it fulfills the resolvent equation (4) R(z1) − R(z2) =
(z2 − z1)R(z1)R(z2). and it is not necessarily the resolvent distribution of A . If it is a resolvent
distribution , by differentiation one obtains in this case AM (z) = zM (z), as ∂z = 0.So M (z) is an
eigenprojector for the eigenvalue z.
One proves by a partition of unity
Proposition 1. Assume a bounded set G ⊂ C and a family Gi, i = 1,··· , l of open, pairwise
disjoint subsets of G, furthermore an open subset G0 ⊃ (G \Sl
1 Gi) and assume a distribution M
on G, whose restriction to G0 vanishes and whose restriction to Gi is multiplicative for i = 1,··· , l.
Then M is multiplicative on G if and only if M (ϕ)M (ψ) = 0 , under the condition that ϕ has its
support in Gi and ψ has its support in Gj with i, j ∈ [1, l], i 6= j.
Proposition 2. Assume an open bounded set G ⊂ C and a family Gi, i = 1,··· , l of open, pairwise
disjoint subsets of G. Assume in each subset Gi a subset Gi,0 of Lebesgue measure 0 and a resolvent
function R(z) on G \S Gi,0 and for each i a sequence of subsets Gi ⊃ Gi,n ↓ Gi,0, such that
R(z)ϕ(z)d2z
Ri(ϕ) = lim
n→∞ZGi\Gi,n
for ϕ ∈ D(Gi) exists and defines a resolvent distribution on Gi. Then the distribution R defined
by the the function R(z) on G \ S Gi,0 and by the distributions Ri on Gi fulfills the resolvent
distribution equation.
Proof. Call Mi = 1/π∂Ri. We have to show, that Mi(ϕ1)Mj(ϕ2) = 0 , if the support of ϕ1 is in
Gi and the support of ϕ2 is in Gj for i 6= j. We have using (5)
d2z2ϕ1(z1)ϕ2(z2)R(z1)R(z2) = I + II
d2z1ZGj \Gj,n
ZGi\Gi,m
d2z1ZGj\Gj,n
z1 − z2
d2z2ϕ2(z2)/(z2 − z1) = ψn(z1) → ϕ1(z1)ZGj
d2z2ϕ1(z1)ϕ2(z2)R(z1)
1
= −ZGi\Gi,m
d2zR(z)ψn(z)
d2z2ϕ2(z2)/(z2 − z1) = ψ(z1)
with
The sequence
I = −ZGi\Gi,m
ϕ1(z1)ZGj \Gj,n
in the sense of D(Gi). We have I → −Ri(ψ) = −R(ψ) = −R(ϕ1(r ∗ ϕ2)). Similarly II →
−R(ϕ2(r ∗ ϕ1)). By the proof of theorem 1, one sees that Mi(ϕ1)Mj(ϕ2) = M (ϕ1ϕ2) = 0.
5
Proposition 3. If the resolvent R(z) has a pole of order n in a point z0, then R(z) is holomorphic
for z in a neighborhood of z0, z 6= z0 and behaves like b0(z − z0)−1 +··· bn−1(z − z0)−n near z0. We
have b0 = p and bj = aj, where p2 = p and an = 0 and ap = pa. Define
where the distribution [6]
Z d2z P
zk ϕ(z) = lim
n−1
P
bk
R(z) =
(z − z0)k+1 ,
Xk=0
zk ϕ(z)d2z = (−1)kZ 1
ε↓0Zz>ε
1
∂kϕ(z)d2z.
z
The distribution R(z) fulfills the resolvent distribution equation. The spectral distribution is
M (z) =
(−1)k
k!
pak∂kδ(z − z0)
n−1
Xk=0
The proof follows directly from Leibniz's formula.
Proposition 4. Assume a compact set K in the real line and an open neighborhood G = G0 × I of
K, where G0 is an open set containing K and I is an open interval containing 0 , and a resolvent
function R(z) holomorphic in G \ K. Define the function Ru on I by Ru(x) = R(x + iu) and
assume, that for any C∞-function ϕ with support in G0 the limits of Ru(ϕ) = R dxRu(x)ϕ(x) for
u → 0+ and for u → 0− exist. If ϕ is a C∞
c - function on G, define ϕu(x) = ϕ(x + iu). Then
u 7→ Ru(ϕu) is integrable and defines the distribution R(ϕ) on G by
R(ϕ) =Z duRu(ϕu) =Z duZ dxR(x + iu)ϕ(x + iu).
The integral defines a resolvent distribution on G.
Proof. Define for m = 0, 1,··· and for any test function ϕ on G0 the norm
kϕkm = max{∂k
xϕ(x) : x ∈ G0, k = 1,··· , m}.
For a distribution T on G0 define accordingly Nm(T ) = sup{kT (ϕ)k : kϕkm ≤ 1} Denote by R0±
the limites of Ru. As for distributions weak convergence implies strong convergence , there exists an
m such that Nm(Ru−R0+) → 0 and Nm(Ru−R0−) → 0 for u → 0+ or u → 0− resp.. The function
u 7→ Ru(ϕu) is continuous for u 6= 0 and has right and left limits at u = 0. Hence it is integrable.
Define a C∞ function α on the real line with α ≥ 0, α(x) = 0 for x ≥ 1, α(x) = 1 forx ≤ 1/2.
Put αε(u) = α(u/ε). Then for zj = xj + iuj, j = 1, 2
R(ϕ1)R(ϕ2) =ZZ du1du2Ru1 (ϕ1,u1 )Ru2 (ϕ2,u2 ) = lim
ε→0ZZ du1du2Ru1 (ϕ1,u1 )Ru2 (ϕ2,u2 )(1−αε(u1−u2))
On the other hand
ZZ du1du2Ru1 (ϕ1,u1 )Ru2 (ϕ2,u2 )(1−αε(u1−u2)) =ZZ d2z1d2z2R(z1)R(z2)ϕ1(z1)ϕ2(z2)(1−αε(u1−u2)).
In the second term the quantities R(z1), R(z2) are usual resolvent functions. We apply the resolvent
equation (5) and split the integral into two terms Iε and IIε.
Iε = −ZZ d2z1d2z2R(z1)1/(z1 − z2)ϕ1(z1)ϕ2(z2)αε(u1 − u2) = −Z duR(ψu,ε)
6
with
Put
ψu,ε(x) = ϕ1(x + iu)Z dx2du2ϕ2(x2 + iu2)
We have to estimate
∂k
x (ψu(x)− ψu,ε(x)) =
ψu(x) = ϕ1(x + iu)Z dx2du2ϕ2(x2 + iu2)
Xj=1(cid:18)k
x ϕ1(x + iu)Z dx2du2∂j
j(cid:19)∂k−j
k
The integral can be estimated with some constants C, K
xϕ2(x+iu−x2−iu2)
Z dx2du2∂j
Hence kψu − ψu,εkm = O(ε ln ε) uniformly in u and
αε(u2) ≤ CZZx′≤K,u′≤ε
x2 + iu2
1
1
x + iu − x2 − iu2
1
(1 − αε(u − u2)).
x + iu − x2 − iu2
.
1
xϕ2(x2 + iu2)
x + iu − x2 − iu2
αε(u− u2)
dx′du′
1
√x′2 + u′2
= O(ε ln ε)
Iε → I = −ZZ d2z1d2z2R(z−1)1/(z1 − z2)ϕ1(z1)ϕ2(z2)
By interchanging the roles of 1 and 2, one obtains the corresponding result for IIε.
Theorem 2. If R is the resolvent of a bounded operator and M is a resolvent distribution extending
R defined on the whole complex plane , then the support of M is compact and M (1) is defined ,
where here 1 is the constant function 1, and
In this case we call M complete.
M (1) =Z d2zM (z) = 1L(V )
Proof. Assume, that the support of M is contained in the circle of radius r. Assume a test function
ϕ constant 1 on this circle, then
M (1) = M (ϕ) = − 1
π R d2zR(z)∂ϕ(z) = − 1
π Rz>r d2zR(z)∂ϕ(z) =
π Zz>r
−
d2z∂(R(z)ϕ(z)) =
1
1
2πiZΓ
dzR(z) = 1L(V ),
where Γ is the circle of radius r run in the anti-clockwise sense. Now the residuum in infinity of
R(z) = 1/(z − A) equals 1L(V ).
1.5. Examples.
(1) Finite dimensional matrix. By Jordan's normal form one obtains that
M (z) =Xi
piXk
(1/k!)(−1)kak
i ∂kδ(z − λi),
where the λi are the eigenvalues,the pi are the eigenprojectors, pipj = δij and the ai are
nilpotent and aipj = δij ai. Instead of ∂ we could have chosen any other linear combination
D of ∂x and ∂y, such that Dz = 1. This an example, that there are many resolvent
distributions extending a resolvent function.
7
We cite another example affirming remark 1. Assume A = 0, then zR(z) = 1, so
R(z) = 1/z − πCδ(z), where C is an arbitrary matrix. Then M (z) = δ(z) − C∂δ(z) and
M (ϕ)M (ψ) = M (ϕψ) if and only if C2 = 0, this might not be the case.
(2) Assume that V is a Hilbert space and that U is a unitary operator.
Z M (z)ϕ(z)d2z =
∞
Xl=−∞
U l 1
2π Z e−iϑlϕ(eiϑ)dϑ.
(3) Assume V to be a Hilbert space and A to be a self adjoint operator. Let E(x), x ∈ R be
the spectral family of A,then
where the derivative is in the sense of distributions.
M (x + iy) = E′(x)δ(y),
2. Discussion of the Multiplication Operator and some Rank One Perturbations
2.1. C∞ Multiplication Operator. Assume an open bounded set G ⊂ Rn and a C∞ bounded
function P : G → R , which is bonded with all its derivatives and with ∇P (y) 6= 0 for all y ∈ G.
Consider V = L2(G) and the operator Ω : V → V with Ωf (y) = P (y)f (y) . The operator Ω is
bounded and its resolvent is R(z) = 1/(z − Ω) for Imz 6= 0. For any test function on the complex
plane the integral
R(ϕ) =Z d2zϕ(z)/(z − Ω),
(R(ϕ)f )(y) =Z d2zϕ(z)/(z − P (y))f (y)
exists and we define the resolvent distribution in that way. A short calculation shows, that the
resolvent distribution equation is fulfilled. We find for the spectral distribution
M (ϕ)f (y) = ϕ(P (y))f (y).
The open mapping theorem ensures that the set P (G) = {x ∈ R : ∃y ∈ G, P (y) = x} is open. The
set
is a (n− 1)-dimensional C∞-submanifold of Rn. With the help of the theorem of implicit functions
one establishes the lemma
S(x) = {y ∈ G : P (y) = x}
Lemma 2. Assume a point y0 ∈ G and set x0 = P (y0). The exists an open neighborhood U1 ⊂ R
of x0 and an open subset U2 ⊂ Rn−1 and an injective C∞ mapping Ψ : U1 × U2 → G, such that
P (Ψ(x, u)) = x
and Ψ : U1 × U2 7→ Ψ(U1 × U2) is a diffeomorphism. Then Ψ(U1 × U2) is an open neighborhood of
y0 and for fixed x the mapping Ψx : u ∈ U2 7→ Ψ(x, u) is a C∞-chart of S(x). If
is the absolute value of the Jacobi's determinant and
J(Ψ) = det(∂Ψ/∂x, (∂Ψ/∂ui)i=1,··· ,n−1))
Γ(Ψx) = (det(∂Ψx/∂ui.∂Ψx/∂uj)i,j=1,··· ,n−1)1/2
is the square root of Gram's determinant and dσx(y) is the euclidean surface element on S(x) for
y ∈ S(x), where the point denotes the scalar product. Then J(Ψ)(x, u) = Γ(Ψx)(x, u)/∇P (x, u)
and dy = dx σx(y)/∇P (y)
8
We introduce dτx(y) = dσx(y)/∇P (y). So
(10)
dy = dxdσx(y)/∇P (y) = dxdτx(y)
As any function of compact support on G can be represented as a finite sum of functions which
have their support in a chart like in the preceding lemma, we obtain the proposition
Corollary 1. If f is a continuous function of compact support on G, then
Z f (y)dy =Z dxZy∈S(x)
dτx(y)
Lemma 3. Assume a function f ∈ D(G), then
ω ∈ P (G) 7→Zy∈S(ω)
f (y)dτx(y)
is in D(P (G))where P (G) is the image of G in R
Proof. Assume at first, that the support of f is contained in an open set Ψ(U1 × U2) like in lemma
2 . Using again this lemma we have
and this is surely C∞. A partition of unity finishes the proof.
Zy∈S(ω)
f (y)dτ (y) =Z duf (Ψ(ω, u)J(ω, u)
If T is a distribution on the open set P (G), then consider T (P (y)).
If f has its support in
U1 × U2, then as distributions transform like functions
Z T (P (y))f (y)dy =ZZ dxduT (P (Ψ(x, u))f (Ψ(x, u))J(Ψ)(x, u) =Z dxT (x)Z duf (Ψ(x, u))Γ(Ψx)(u)
Hence we define for test function f ∈ D(G)
(11)
T (Ω)(f ) =Z T (P (y))f (y)dy =Z dxT (x)Zy∈S(x)
f (y)dτx(y).
Especially
(12)
δ(x − Ω)(f ) =Zy∈S(x)
f (y)dτx(y)
and M (z) = δ(x − Ω)δ(y). This measure on Rn has been treated by Gelfand-Schilow under the
name δ(x − P )[2]. We cite the definition [3]
Definition 3. Assume a vector space V and and a linear mapping A : V → V . A linear functional
F : V → C is called a generalized eigenvector for the eigenvalue x if F (Af ) = xF (f ) for all f ∈ V..
Obviously the operator Ω leaves D(G) invariant. We define the left generalized ket-vector hδy
as the kernel hδy(w) = δ(y − w) applied to f in the following way
hδyfi =Z δ(y − w)f (w)dw = f (y)
hδyΩfi = P (y)f (y) = P (y)hδyfi
hence
and hδy is a generalized left eigenvector for the eigenvalue P (y). . Similar we define the generalized
right eigenvector δyi by hfδyi = f (y). For f ∈ L2(G) the bra-vector hf is given by the functional
9
g 7→ hfg = hfgi =R f (y)g(y)dy. Similar relations hold for the ket-vector fi. Following this idea
we define hδyδy ′i =R dwδ(y − w)δ(y′ − w) and finally
hδyδy ′i = δ(y − y′).
This is the orthogonality relation for the generalized eigenvectors. The relation
states the completeness of the eigenvectors. Finally
hfgi =ZZ dydy′f (y)g(y′)hδyδy ′i.
δ(x − Ω) =Zy∈S(x) δyihδydτx(y).
2.2. Perturbations of the Multiplication Operator 1. We perturb the multiplication operator
Ω and define the operator
H = Ω + gihh
(13)
with g, h ∈ D(G). The resolvent can be easily calculated
(14)
where RΩ(z) = 1/(z − Ω) is the resolvent of Ω . The ket-vector RΩ(z)gi is the functional f ∈
D(G) 7→ hfRΩ(z)gi and similarly defined is the bra-vector hhRΩ(z).
R(z) = RΩ(z) + RΩ(z)gihhRΩ(z)/C(z),
C(z) = 1 − hhRΩ(z)gi,
Recall the formula
C(z) → 1 − hh P
x − Ωgi ± iπhhδ(x − Ω)gi = C1(x) ± iπC2(x)
So there are jumps of C(z) on the real axis contained in P (K), if the supports of g.h are contained
in the compact set K ⊂ G. If z /∈ P (K) and C(z) 6= 0, then R(z) exists as function and z belongs
to the resolvent set. We assume, that the jumps on P (K) are are isolated, more precisely that there
exists an open neighborhood G0 ⊂ C of P (K), such that C(z) is 6= 0 and holomorphic in G0 \ P (K)
and C(x ± i0) 6= 0 for x ∈ G0 ∩ R. We may assume, that G0 = G1 × I, where G1 is an open subset
of the real line containing P (K) and I is an open interval containing 0.
Lemma 5. Assume a C∞ test function ϕ with support in G1 and put Ru(x) = R(x + iu) then for
ϕ ∈ D(G1) the operator R dxR(x + iu)ϕ(x) = Ru(ϕ) on L2(G) converges for u → 0+ or u → 0−
in operator norm to an operator called R±0(ϕ).
10
1
x ± iε →
1
x ± i0
= P
x ∓ iπδ(x) for ε ↓ 0,
1
1
f (y)g(y)
z − P (y)
where P denotes the principal value and obtain the lemma
Lemma 4. Assume f1, f2 ∈ D(G), then for z → x ± i0 uniformly in x
x ± i0 − Ωf2i = hf1 P
Z dy
x − Ωf2i =Z dω P
hf1 P
We use Dirac's notation hf1Af2i = hf1Af2i, which is in many cases convenient. We calculate
z − Ωf2i → hf1
x − ω Zy∈S(ω)
hf1δ(x − Ω)f2i =Zy∈S(x)
x − Ωf2i ∓ iπhf1δ(x − Ω)f2i
f 1(y)f2(y)dτ (y),
f 1(y)f2(y)dτ (y)
= hf1
with
for z → x ± i0
Proof. Rewrite equation (14)
R(z) =
1
1
1
z − Ω
+ZZ dw1dw2
+ZZ dw1dw2
w2 − w1
1
(
z − w1
z − w1 −
1
1
=
δ(w1 − Ω)gihhδ(w2 − Ω)/C(z)
z − w2
1
z − Ω
Then for u 6= 0
Z dxR(x + iu)ϕ(x) = Ru(ϕ) = ψu(Ω) +ZZ dw1dw2χu(w1, w2)δ(w1 − Ω)gihhδ(w2 − Ω)
z − w2
)δ(w1 − Ω)gihhδ(w2 − Ω)/C(z)
The function ψu(w) = R dxϕ(x)/(x + iu − w) is in D(G1) and converges in this sense to ψ+0 or
ψ−0.
χu(w1, w2) =Z dxϕ(x)
= −Z dx
)Z 1
0
dt
=Z dx
d
dx
(
ϕ(x)
C(x + iu)
1
1
(
x + iu − w1 −
dt
ϕ(x)
d
dx
w2 − w1
C(x + iu)Z 1
x + iu − w1 − t(w2 − w1)
ϕ(x)
1
0
ωu(x) =
d
dx
(
C(x + iu)
)
)/C(x + iu)
1
x + iu − w2
1
x + iu − w1 − t(w2 − w1)
=ZZ dxdt
ωu(x + (1 − t)w1 + tw2))
x + iu
The function x 7→ ωu(x) is in D(G1) and converges for u → ±0 in this sense , as the function
x 7→ C(x + iu) is in C∞ and converges to x 7→ C(x ± 0) for x → +0 or x → −0 uniformly , the
analogue holds for all derivatives. Hence we obtain
χ±0(w1, w2) =ZZ dxdt
ω±0(x + (1 − t)w1 + tw2))
.
x ± i0
Assume f1, f2 ∈ L2(G), then for ϑ = ±0
ZZ dw1dw2(χu(w1, w2) − χϑ(w1, w2))hf1δ(w1 − Ω)gihhδ(w2 − Ω)f2i
≤ max{χu(w1, w2) − χϑ(w1, w2)) : w1, w2 ∈ P (K)}ZZ dw1dw2hf1δ(w1−Ω)gihhδ(w2−Ω)f2i.
Observe, using corollary 1 , that e.g. R dwhfδ(w− Ω)gi ≤ N (f )N (g), where N (.) is the L2-norm,
and conclude from there that the last expression converges to 0 in operator norm,
A consequence of proposition 4 and lemma 5
Proposition 5. Define for a C∞ test function ϕ with support in G0 the operator valued distribution
R(ϕ) = lim
ε↓0ZZu>ε
dxduR(x, u)ϕ(x, u) = lim
ε↓0Zu>ε
duRu(ϕu)
with ϕu(x) = ϕ(x + iu). Then R(ϕ) extends the function R(z) on G0 \ P (K) to a distribution on
G0, which fulfills the resolvent distribution equation. We have
M (x + iy) =µ(x)δ(y),
µ(x) =
1
2πi
11
(R(x − i0) − R(x + i0).
We consider for u 6= 0 the sesquilinear form
f1, f2 ∈ D(G) → B(Ru(x))(f1, f2) = hf1Ru(x)f2i.
By lemma 4 the bracket converges uniformly in x for x → ±0 to a sesquilinear form B(R±0)(x).
We obtain
hf1R±0(ϕ)f2i =Z dxϕ(x)B(R±0)(x)(f1, f2),
and define
Hence
hf1κ(x)f2i = B(µ(x))(f1, f2) =
1
2πi
(B(R−0)(x)(f1, f2) − B(R+0(x)(f1, f2)).
hf1µ(ϕ)f2i =Z dxϕ(x)hf1κ(x)f2i
From the formula
and κ(x) appears as the restriction of the sesquilinear form f1, f2 ∈ L2(G) 7→ hf1µ(x)f2i, which
is a distribution, to f1, f2 ∈ D(G), where x 7→ hf1κ(x)f2i is a continuous function.
B(R±0(x)) = P
( P
x − Ω ∓ iπδ(x − Ω))gihh( P
x − Ω − ∓iπδ(x − Ω)).
C1 ± iπC2
one establishes by straight forward calculations
x − Ω ∓ iπδ(x − Ω) +
1
Proposition 6. Recall
C1(x) =1 − hh P
x − Ωgi
Assume, that g, h are real and that
C2(x) =hhδ(x − Ω)gi
C(x ± i0)2 = C1(x) ± iπC2(x)2 = C1(x)2 + π2C2(x)2 6= 0
for x ∈ G0 ∩ R. We define the following bra- and ket-vectors as functionals over D(G).
A =A(x) = P
x − Ωgi
B =B(x) = δ(x − Ω)gi
A′ =A′(x) = hh P
x − Ω
B′ =B′(x) = hhδ(x − Ω)
For x ∈ P (K) one obtains
κ(x) = δ(x − Ω) +
1
C2
1 + π2C2
2
(AC2A′ + AC1B′ + BC1A′ − π2BC2B′)
If C2(x) 6= 0, we may write
κ(x) = δ(x − Ω) −
BB′
C2
= δ(x − Ω) −
+
BB′
C2
+
1
1 + π2C2
(C2
1
(AC2
2 A′ + AC1C2B′ + BC1C2A′ + C2
1 BB′)
2 )C2
(C1B + AC2)(C1B′ + A′C2) = p(x) + α(x)ihα′(x)
(C2
1 + π2C2
2 )C2
with
p(x) = δ(x − Ω) −
δ(x − Ω)gihhδ(x − Ω)
α(x)i = N (x)(C1(x)δ(x − Ω)gi + C2(x) P
x − Ωgi)
hα′(x) = N (x)(C1(x)hhδ(x − Ω) + C2(x)hh P
x − Ω
hhδ(x − Ω)gi
12
and N (x)2 = 1/((C2
1 + π2C2
2 )C2).
We discuss the case C2(x) = 0. If h = g, then C2(x) = RS(x) dτ (y)g(y)2 = 0 implies , that g on
S(x) vanishes and hence B = 0, B′ = 0. In the general case, remark that B = 0 or B′ = 0 imply
that C2 = 0. Let us assume, that g, h are in such form ,that C2(x) = 0 implies B(x) = 0, B′(x) = 0.
Then
(15)
κ(x) =(p(x) + α(x)ihα′(x)
δ(x − Ω)
for C2(x) 6= 0
for C2(x) = 0
If g = h, then µ(x) is positive definite.
Proposition 7. The operator H maps D(G) → D(G). So we may formulate
Hκ(x) = κ(x)H = xκ(x)
Hα(x)i = xα(x)i
Hp(x) = p(x)H = xp(x)
hα′(x)H = xhα′(x)
These equations have to be understood as equations for sesquilinear forms bracketed between func-
tions in D(G) or as functionals on D(G) . So κ(x) and p(x) are generalized eigenprojectors and
α(x)i is a generalized right eigenvector and hα′(x) is a generalized left eigenvector, all for the
eigenvalue x.
The proof is done by straight forward computation.
Proposition 8. We have the orthogonality relations
κ(x)κ(x′) = δ(x − x′)κ(x)
p(x)α(x′)i = 0, hα′(x)p(x′) = 0
p(x)p(x′) = δ(x − x′)p(x)
hα′(x)α(x′)i = δ(x − x′)
These relations have to be understood in the sense of distributions, e.g. the first one signifies
(Z dx1ϕ1(x1)hf1κ(x1))(Z dx2ϕ2(x2)κ(x2)f2i =Z dxϕ1(x)ϕ2(x)hf1κ(x)f2i
for f1, f2 ∈ D(G) and ϕ1, ϕ2 ∈ D(P (G)) and we show, that the expressions make sense.
Proof. Assume f1 ∈ D(G), by the mapping f2 ∈ D(G) → hf1κ(x)f2i we define a distribution
called hf1κ(x). For a test function ϕ ∈ D(P (G)) the integral R dxϕ(x)hf1κ(x) exists and we have
(Z dxϕ(x)hf1κ(x))(f2) =Z dxϕ(x)hf1κ(x)f2i = hf1µ(ϕ)f2i.
Hence
Similarly define κ(x)fi and obtain
So we can form the scalar product
Z dxϕ(x)hf1κ(x) = hf1µ(ϕ) ∈ L2(G).
Z dxϕ(x)κ(x)fi = µ(ϕ)f2i ∈ L2(G).
(Z dx1ϕ1(x1)hf1κ(x1))(Z dx2ϕ2(x2)κ(x2)f2i = hf1µ(ϕ1)µ(ϕ2)f2i = hf1µ(ϕ1ϕ2)f2i
13
This proves the first equation. We check the last equation
hα′(x)α(x′)i = N (x)N (x′)(C1(x)hhδ(x−Ω)+C2(x)hh P
x − Ω
)(C1(x′)δ(x′−Ω)gi+C2(x′) P
x′ − Ωgi)
= N (x)N (x′)ZZ dw1dw2hhδ(w1 − Ω)δ(w2 − Ω)gi
Use δ(w1 − Ω)δ(w2 − Ω) = δ(w1 − w2)δ(w1 − Ω) and continue
(C1(x)δ(x − w1) + C2(x) P
x − w1
= N (x)N (x′)Z dwC2(w)(C1(x)δ(x − w) + C2(x) P
x − w
Use the properties of the δ- function and equation (3) and continue
)(C1(x′)δ(x′ − w2) + C2(x) P
)
x′ − w2
)(C1(x′)δ(x′ − w) + C2(x′) P
x′ − w
).
= N (x)N (x′)(C2(x)C1(x)2δ(x−x′)+C2(x)C2(x′)(C1(x′) P
x − x′ +C1(x) P
x′ − x
1
x′ − x
)+π2C2(x)2δ(x−x′))
( P
x − w − P
x′ − w
))
+ C2(x)C2(x′)Z dwC2(w)
x − x′ (− P
x′ − Ω
=Z dwC2(w)
= hh
1
+ P
)gi
x − Ω
x − x′ (− P
1
x′ − w
+ P
x − w
))
Now
C1(x′) P
x − x′ + C1(x) P
x′ − x
=
C1(x′) − C1(x)
x − x′
The verification of the two other relations are left to the reader.
2.3. Perturbation of the Multiplication Operator 2. This example is a caricature of the
eigenvalue problem arising in the theory of radiation transfer in a gray atmosphere in plan parallel
geometry [1]. We consider for some c > 1 the set G =]−c,−1[∪]1, c[⊂ R, the multiplication operator
Ωf (y) = yf (y) and two real C∞ functions f, g on R with g(x) > 0 for 1 < x < c and −c < x < −1
and 0 outside these two open interval. We assume g(y) = g(−y) and h(y) = −g(y) for y > 1 and
h(y) = g(y) for y < −1. We study H = Ω + gihh and obtain
= 1 +Z c
g(y)h(y)
z − y
z2 − y2 .
2yg(y)2
dy
1
We have for z = x + iu
dy
C(z) = 1 −ZG
ImC(z) =Z c
1
dy
C(0) = 1 −Z c
1
dy
2g(y)2
y
,
4y2xug(y)2
(x2 − u2 − y2)2 + 4x2u2
C(iu) = 1 −Z c
1
dy
2yg(y)2
u2 + y2
Hence C(z) = 0 implies xu = 0 , so either x or u or both vanish. We have
So C(iu) is monotonic increasing for increasing u2 and goes to 1 for u2 → ∞. If C(0) < 0, there
exists exactly one u0 > 0 such that C(iu0) = 0, if C(0) > 0, then C(iu) > 0 for all u.
For x ≤ 1 we have
C(x) = 1 −Z c
1
14
dy
2yg(y)2
y2 − x2
and is monotonic decreasing for increasing x. If C(0) > 0 and C(1) < 0 there exists exactly one
x0 with 0 < x0 < 1, such that C(x0) = 0. If C(0) < 0 there does not exist such an x. We do not
discuss the case C(0) > 0 and C(1) ≥ 0. In case C(0) = 0 we have a double zero. For x ≥ c we
have C(x) ≥ 1. Hence C(x) does not vanish for x ≥ c.
The singularities of the resolvent are the slits [−c,−1] and [1, c] and the zeros of C(z). In the
neighborhood of a zero of C(z) we may define a resolvent distribution with the help of proposition
3. We discuss the behavior of the resolvent in the neighborhood of the slits. We have
C(x ± i0) = C1(x) ± iπC2(x) = 1 −Z dyg(y)h(y) P
x − y ± iπg(x)h(x).
There exists a neighborhood G1 ⊂ R of [−1,−c] ∪ [1, c] and an open interval I containing 0, such
that R(z) is holomorhic in (G1 × I) \ ([−1,−c] ∪ [1, c]) and C(x ± i0) 6= 0 for x ∈ G1. So we can
apply prop. 4 and define a distribution R extending the resolvent function R(z) to G1 × I and
fulfilling the distribution resolvent equation . From the local definition in the neighborhood of the
singularities we can define a resolvent distribution extending R(z) to C with the help of proposition
2.
Proposition 9. We calculate the spectral distribution. If there are two zeros 6= 0 we obtain
M (z) = M (x + iu) = r+δ(z − z0) + r−δ(z + z0) + δ(u)µ(x)
where r± are the residues of R(z) at the points ±z0 and
µ(x) =
1
2πi
(R(x − i0) − R(x + i0).
We identify µ(x) with its restriction κ(x) as bilinear form on D(C) and obtain
−δ(z + z0) + δ(u)αxihα′
x
+δ(z − z0) + α−ihα′
M (z) = α+ihα′
with the right resp. left usual eigenvectors
1
1
α±i =
± =
and for x ∈ G the right, resp. left generalized eigenvectors
x =
phh(±z0 − Ω)−2gi
(C1(x)δxi + h(x)A(x))
±z0 − Ωgi
αxi =
hα′
1
1 + π2C2
2
hα′
pC2
with
1
±z0 − Ω
1
phh(±z0 − Ω)−2gihh
pC2
1 + π2C2
2
1
(C1(x)hδx + g(x)A′(x))
C2(x) = g(x)h(x)
A′(x) = hh P
x − Ω
B′(x) = h(x)hδx
+ O(1) = z−2a + z−1p0 + O(1)
C1(x) == 1 −Z dyg(y)h(y) P
x − y
A(x) = P
x − Ωgi
B(x) = g(x)δxi
In the case C(0) = 0 we obtain
R(z) = z−2 Ω−1gihhΩ−1
hhΩ−3gi
and
+ z−1 Ω−2gihhΩ−1 + Ω−1gihhΩ−2
hhΩ−3gi
M (z) = M (x + iu) = p0δ2(z) − a∂δ2(z) + µ(x)δ(u),
15
where µ(x) = αxihα′
We obtain for ϑ, ϑ′ = ± and x, x′ ∈ R the orthogonality relations
x is given by the formula above. Here p2 = p and a2 = 0 and ap = pa = a.
hα′
hαxαx′i = δ(x − x′)
Analogous relations hold for the case C(0) = 0. The spectral distribution is complete, i.e.
xαϑi = hαϑαxi = 0
ϑαϑ′i = δϑ,ϑ′
hα′
M (1) =Z ]d2zM (z) = Xϑ=±
ϑihαϑ +Zg
α′
dxα′
xihαx = 1
Proof. Assume at first, that C(z) has two zeros ±z0 = ±iu0 or ±z0 = ±x0. The residuum of the
complex function R(z) at the points ±z0 is given by
±z0 − Ω(cid:1)/C′(±z0) =(cid:0)
±z0 − Ωgihh
±z0 − Ωgihh
hh(±z0 − Ω)−2gi
r± =(cid:0)
1
±z0 − Ω(cid:1)
1
1
1
1
Using the results of proposition 3 we obtain
M (z) = M (x + iu) = r+δ2(z − z0) + r−δ2(z + z0) + δ(u)µ(x)
where
We consider the case C(0) = 0 .We expand at the origin
µ(x) =
1
2πi
(R(x − i0) − R(x + i0)
C(z) = 1 +
∞
Xn=0
znhhΩ−(n+1)gi = z2hhΩ−3gi + O(z3) = −z2Z ∞
1
2g(y)/y3dy + O(z3)
R(z) = z−2 Ω−1gihhΩ−1
hhΩ−3gi
+ z−1 Ω−2gihhΩ−1 + Ω−1gihhΩ−2
hhΩ−3gi
+ O(1) = z−2a + z−1p0 + O(1)
and
Here p2 = p and a2 = 0 and ap = pa = a.
M (z) = M (x + iu) = p0δ2(z) − a∂δ2(z) + µ(x)δ(u).
In both cases we can use for f1, f2 ∈ D(R) the relation hf1µ(x)f2i = hf1κ(x)f2i and κ(x) is
given by the formula
κ(x) = δ(x − Ω) +
1
C2
1 + π2C2
2
(AC2A′ + AC1B′ + BC1A′ − π2BC2B′)
=
1
C2
1 + π2C2
2
(Ah(x)g(x)A′ + AC1h(x)hδx + g(x)δxiC1A′ + C2
1 h(x)2g(x)2δxihδx) = αxihα′
x
as δ(x − Ω) = δxihδx. The proof of the orthogonality relations can be deduced from the relation
κ(x)κ(x′) = δ(x − x′)κ(x), like in the proof of prop 8 or can be done by hand. That is left to the
reader. For the completeness refer to theorem 2.
16
References
[1] G.V.Efimov,W. von Waldenfels, R.Wehrse, Analytical solution of the non-discretized radiative transfer equation
for a slab of finite optical depth. J.Spectrosc.Radiative Transfer 53,59-74 (1995)
[2] I. Gelfand,N.Vilenkin. Generalized functions. vol1(Academic Press, New York 1964)
[3] I. Gelfand,N.Vilenkin. Generalized functions. vol4(Academic Press, New York 1964)
[4] E.Hille, R.S.PHillips. Functional Analysis and Semigroups.(Amer.Math.Soc.Providence 1968)
[5] G.de Rham Vari´et´es differentiables.( Hermann,Paris,1955).
[6] L.Schwartz.Th´eorie des distributions I (Hermann,Paris ,1951)
[7] L.Schwartz. Th´eorie des noyaux. Proceedings of the international congress of mathematicians , 1950, vol !,
p.220-230.
[8] L.Schwartz. Th´eorie des distributions `a valeurs vectorielles . Ann. Inst. Fourier 7(1957),p.1-142
[9] W.von Waldenfels. A measure theoretical approach to quantum stochastic processes. Springerverlag 2014. Lec-
ture Notes in Physics 878.(Springer Verlag Berlin Heidelberg 20014)
17
|
1610.01997 | 1 | 1610 | 2016-10-06T19:10:18 | Complete Nevanlinna-Pick kernels | [
"math.FA"
] | We give a new treatment of Quiggin's and McCullough's characterization of complete Nevanlinna-Pick kernels. We show that a kernel has the matrix-valued Nevanlinna-Pick property if and only if it has the vector-valued Nevanlinna-Pick property. We give a representation of all complete Nevanlinna-Pick kernels, and show that they are all restrictions of a universal complete Nevanlinna-Pick kernel. | math.FA | math |
Complete Nevanlinna-Pick Kernels
Jim Agler ∗
John E. McCarthy †
University of California at San Diego, La Jolla California 92093
Washington University, St. Louis, Missouri 63130
Abstract
We give a new treatment of Quiggin's and McCullough's characterization of com-
plete Nevanlinna-Pick kernels. We show that a kernel has the matrix-valued Nevanlinna-
Pick property if and only if it has the vector-valued Nevanlinna-Pick property. We
give a representation of all complete Nevanlinna-Pick kernels, and show that they are
all restrictions of a universal complete Nevanlinna-Pick kernel.
0
Introduction
Let X be an infinite set, and k a positive definite kernel function on X, i.e. for any finite
collection x1, . . . , xn of distinct points in X, and any complex numbers {ai}n
i=1, the sum
n
Xi,j=1
ai¯ajk(xi, xj) ≥ 0,
(0.1)
with strict inequality unless all the ai's are 0. For each element x of X, define the function
kx on X by kx(y) := k(x, y). Define an inner product on the span of these functions by
hX aikxi , X bjkyj i = X ai¯bjk(xi, yj),
and let H = Hk be the Hilbert space obtained by completing the space of finite linear
combinations of kxi's with respect to this inner product. The elements of H can be thought
of as functions on X, with the value of f at x given by hf, kxi.
A multiplier of H is a function φ on X with the property that if f is in H, so is φf . The
Nevanlinna-Pick problem is to determine, given a finite set x1, . . . , xn in X, and numbers
∗Partially supported by the National Science Foundation
†Partially supported by National Science Foundation grant DMS 9531967.
1
λ1, . . . , λn, whether there exists a multiplier φ of norm at most one that interpolates the
data, i.e. satisfies φ(xi) = λi for i = 1, . . . , n.
If φ is a multiplier of H, we shall let Tφ denote the operator of multiplication by φ. Note
φ kx = φ(x)kx. So if φ interpolates the data (xi, λi), then
φ , and on this
that the adjoint of Tφ satisfies T ∗
the n-dimensional space spanned by {kxi : 1 ≤ i ≤ n} is left invariant by T ∗
subspace the operator T ∗
φ is the diagonal
λ1
. . .
λn
(0.2)
with respect to the (not necessarily orthonormal) basis {kxi}. For a given set of n data
points (x1, λ1), . . . , (xn, λn), let Rx,λ be the operator in (0.2), i.e. the operator that sends
kxi to λikxi. A necessary condition to solve the Nevanlinna-Pick problem is that the norm of
Rx,λ be at most 1; the kernel k is called a Nevanlinna-Pick kernel if this necessary condition
is also always sufficient.
Notice that Rx,λ is a contraction on sp{kxi : 1 ≤ i ≤ n} if and only if (1 − R∗
x,λRx,λ) is
positive on that space. As
h(1 − R∗
x,λRx,λ)
n
Xi=1
aikxi,
n
Xj=1
ajkxj i =
n
Xi,j=1
aiaj(1 − λjλi)hkxi, kxj i,
it follows that the contractivity of Rx,λ on sp{kxi : 1 ≤ i ≤ n} is equivalent to the positivity
of the n-by-n matrix
The classical Nevanlinna-Pick theorem asserts that the Szego kernel
(cid:16)(1 − λjλi)hkxi, kxj i(cid:17)n
i,j=1
.
(0.3)
k(x, y) =
1
1 − ¯xy
on the unit disk is a Nevanlinna-Pick kernel. The condition is normally stated in terms of
the positivity of (0.3), but as we see that is equivalent to the contractivity of (0.2).
The matrix-valued Nevanlinna-Pick problem is as follows. Fix some auxiliary Hilbert
space, which for notational convenience we shall assume to be the finite-dimensional space
Cν. The tensor product H ⊗ Cν can be thought of as a space of vector valued functions on
X. A multiplier of H ⊗ Cν is now a ν-by-ν matrix valued function Φ on X with the property
that whenever
f1
...
fν
∈ H ⊗ Cν
2
f1
...
fν
Φ
∈ H ⊗ Cν.
then
The matrix Nevanlinna-Pick problem is to determine, given points x1, . . . , xn and matrices
Λ1, . . . , Λn, whether there is a multiplier Φ of norm at most one that interpolates: Φ(xi) =
Λi.
Fix a (not necessarily orthonormal) basis {eα}ν
α=1 for Cν. As before,
T ∗
Φkx ⊗ v = kx ⊗ Φ(x)∗v,
so if M is the span of {kxi ⊗ eα : 1 ≤ i ≤ n, 1 ≤ α ≤ ν}, a necessary condition for the
Nevanlinna-Pick problem to have a solution is that the nν-by-nν matrix
Rx,Λ : kxi ⊗ eα 7→ kxi ⊗ Λ∗
i eα
(0.4)
be a contraction. We shall call the kernel k a complete Nevanlinna-Pick kernel if, for all
finite ν and all positive n, the contractivity of Rx,Λ is also a sufficient condition to extend
Φ to a multiplier of all of H × Cν of norm at most one.
In Section 1 we give a classification of all complete Nevanlinna-Pick kernels. This was
originally done by S. McCullough in [7] in the context of the Carath´eodory interpolation
problem. The Nevanlinna-Pick problem was studied by P. Quiggin, who in [8] established
the sufficiency of the condition in Theorem 1.2, and in [?] established the necessity.
In Section 2 we show that if a kernel has the Nevanlinna-Pick property for row vectors
In
of length ν, then it has the Nevanlinna-Pick property for µ-by-ν matrices for all µ.
particular, having the vector-valued Nevanlinna-Pick property is equivalent to having the
complete Nevanlinna-Pick property.
In Section 3, we show that all complete Nevanlinna-Pick kernels have the form
k(x, y) =
δ(x)δ(y)
1 − F (x, y)
where δ is a nowhere vanishing function and F : X × X → D is a positive semi-definite
function.
In Section 4 we introduce the universal complete Nevanlinna-Pick kernels am defined on
the unit ball Bm of an m-dimensional Hilbert space (m may be infinite) by
am(x, y) =
1
1 − hx, yi
.
These kernels are universal in the sense that, up to renormalization, every complete Nevanlinna-
Pick kernel is just the restriction of an am to a subset of Bm.
3
1 Characterization of Complete Nevanlinna-Pick ker-
nels
To simplify notation, we shall let ki denote kxi, and kij denote hki, kji = k(xi, xj). First
we want a lemma that says that we can break H up into summands on each of which k is
irreducible, i.e. kij is never 0. For convenience, we shall defer the proof of the lemma until
after the proof of the theorem.
Lemma 1.1 Suppose k is a Nevanlinna-Pick kernel on the set X. Then X can be parti-
tioned into disjoint subsets Xi such that if two points x and y are in the same set Xi, then
k(x, y) 6= 0; and if x and y are in different sets of the partition, then k(x, y) = 0.
A reducible kernel will have the (complete) Nevanlinna-Pick property if and only if each
irreducible piece does, so we shall assume k is irreducible.
Theorem 1.2 A necessary and sufficient condition for an irreducible kernel k to be a com-
plete Nevanlinna-Pick kernel is that, for any finite set {x1, . . . , xn} of n distinct elements
of X, the (n − 1)-by-(n − 1) matrix
Fn = 1 −
is positive semi-definite.
kinknj
kijknn!n−1
i,j=1
(1.3)
Proof: Let x1, . . . , xn−1 and Λ1, . . . , Λn−1 be chosen, let M be the span of {ki ⊗ eα : 1 ≤
i ≤ n − 1, 1 ≤ α ≤ ν}, and define Rx,Λ on M by (0.4). The operator Rx,Λ is a contraction
if and only if I − R∗
x,ΛRx,Λ ≥ 0. Calculate
h(I − R∗
x,ΛRx,Λ)Xi,α
aα
i ki ⊗ eα , Xj,β
aβ
j kj ⊗ eβi = Xi,α,j,β
i ¯aβ
aα
j kij(heα, eβi − hΛjΛ∗
i eα, eβi) (1.4)
A necessary and sufficient condition to be able to find a matrix Λn so that the extension
neα for each α has the same norm as Rx,Λ is:
x,Λ of Rx,Λ that sends kxn ⊗ eα to kxn ⊗ Λ∗
R∼
whenever Λ1, . . . , Λn−1 are chosen so that
on ∨{ki ⊗ eα : 1 ≤ i ≤ n − 1, 1 ≤ α ≤ ν}, then
I − R∗
x,ΛRx,Λ ≥ 0
P − (P R∼
x,ΛP )∗(P R∼
x,ΛP ) ≥ 0,
4
(1.5)
(1.6)
where P is the orthogonal projection from ∨{ki ⊗ eα : 1 ≤ i ≤ n, 1 ≤ α ≤ ν} onto the
orthogonal complement of ∨{kn ⊗ eα : 1 ≤ α ≤ ν}. (This was first proved in [1] in the scalar
case, and a proof of the matrix case is given in [3]. Notice that (1.6) does not depend on
the choice of Λn. We use ∨ to denote the closed linear span of a set of vectors.)
That (1.5) always implies (1.6) for any choice of x and Λ is not only necessary, but
also sufficient for k to be a complete Nevanlinna-Pick kernel. Sufficiency is proved by an
inductive argument that if one can always extend a multiplier defined on a finite set to any
other point without increasing the norm, then one can extend the multiplier to all of X.
In the absence of any a priori simplifying assumptions about the multiplier algebra of H
being large, the proof of this inductive argument is subtle, and is originally due to Quiggin
[8, Lemma 4.3].
Using the fact that
we can calculate that
P (ki ⊗ eα) = (ki −
kin
knn
kn) ⊗ eα,
h(P − (P R∼
x,ΛP )∗(P R∼
x,ΛP )Xi,α
aα
i ki ⊗ eα , Xj,β
aβ
j kj ⊗ eβi
equals
j kij 1 −
i ¯aβ
aα
Xi,α,j,β
kinknj
kijknn!hheα, eβi − hΛjΛ∗
i eα, eβii
(1.7)
Comparing (1.4) and (1.7), we see that we want that whenever the matrix whose (i, α)th
column and (j, β)th row is given by
kij(heα, eβi − hΛjΛ∗
i eα, eβi)
(1.8)
is positive, then the Schur product of this matrix with Fn ⊗ J is positive, where J is the
ν-by-ν matrix all of whose entries are 1. As the Schur product of two positive matrices is
positive, the positivity of (1.3) is immediately seen to be a sufficient condition for k to be
a complete Nevanlinna-Pick kernel.
We shall prove necessity by induction on n. The case n = 2 holds by the Cauchy-Schwarz
inequality. So assume that Fn−1 is positive, and we shall prove that Fn is positive.
Note first the sort of matrices that can occur in (1.8). For each i and α, one can choose
the vector Λ∗
i eα arbitrarily. In particular, let G be any positive (n−1)-by-(n−1) matrix, let
ε > 0, and choose {eα} so that heα, eβi = εδα,β + 1. Choose vectors vi so that hvi, vji = Gij.
5
Let ν = n − 1, and choose Λ∗
(1.8) becomes
i to be the rank one matrix that sends each eα to vi. Then
kij(εδα,β + 1 − Gij).
(1.9)
We know that Fn has the property that if G is a positive matrix and the (n−1)ν-by-(n−1)ν
matrix (1.9) is positive, then the Schur product of Fn ⊗ J with (1.9) is also positive. Denote
by K the (n − 1)-by-(n − 1) matrix whose (i, j) entry is kij, and let · denote Schur product.
By letting ε tend to zero, we get that whenever G ≥ 0 and
[K · (J − G)] ⊗ J ≥ 0,
[Fn ⊗ J] · ([K · (J − G)] ⊗ J) ≥ 0,
then
which is the same as saying
K · (J − G) ≥ 0 ⇒ Fn · K · (J − G) ≥ 0.
(1.10)
Let L be the rank one positive (n − 1)-by-(n − 1) matrix given by
and let G be the matrix given by
Lij =
ki(n−1)k(n−1)j
k(n−1)(n−1)
,
Gij = 1 −
Lij
kij
.
Then G is the matrix that agrees with Fn−1 in the first (n − 2) rows and columns, and all
the entries in the (n−1)st row and column are zero. Therefore G is positive by the inductive
hypothesis. Moreover, K · (J − G) = L and so is positive. Therefore Fn · L is positive. But
L is rank one, so 1/L (the matrix of reciprocals) is also positive, and therefore
Fn · L · 1/L = Fn ≥ 0,
as desired.
✷
Proof of Lemma 1.1: Let Xx = {y : k(x, y) 6= 0}. We need to show that for any
two points x and y, the sets Xx and Xy are either equal or disjoint. This is equivalent to
proving that if k(x, z) 6= 0 and k(y, z) 6= 0, then k(x, y) 6= 0.
6
Assume this fails. Consider the 2-by-2 matrix T ∗ defined on the linear span of kx and
ky by
T ∗kx = kx
T ∗ky = −ky
This has norm one, because k(x, y) = 0. By the hypothesis that k is a Nevanlinna-Pick
kernel, T ∗ can be extended to the space spanned by kx, ky and kz so that the new operator
has the same norm and has kz as an eigenvector. But for this to hold, from equation (1.7)
we would need
2 0(cid:19) · kxx − kxz2
(cid:18) 0 2
kzz
kyx − kyzkzx
kzz
! ≥ 0.
kxy − kxzkzy
kzz
kyy − kyz2
kzz
(1.11)
But the Schur product of the two matrices in (1.11) is zero on the diagonal, non-zero off
the diagonal, and therefore cannot be positive.
✷
By the same argument as in the theorem, an irreducible kernel will have the (scalar)
Nevanlinna-Pick property if and only if whenever G is positive and rank one, (1.10) holds.
We do not know how to classify such kernels in the sense of Theorem 1.2.
The positivity of Fn can be expressed in other ways. The proof that Fn being positive
is equivalent to 1/K having only one positive eigenvalue below is due to Quiggin [8].
Corollary 1.12 A necessary and sufficient condition for the irreducible kernel k to have
the complete Nevanlinna-Pick property is that for any finite set x1, . . . , xn, the matrix
Hn := 1
kij!n
i,j=1
has exactly one positive eigenvalue (counting multiplicity).
Proof: As all the diagonal entries of Hn are positive, Hn must have at least one positive
eigenvalue.
The condition that Fn+1 be positive is equivalent to saying
Mn := kn+1,n+1
ki,n+1kn+1,j
−
1
kij!n
i,j=1
≥ 0,
(1.13)
because ki,n+1kn+1,j is rank one so its reciprocal is positive. But (1.13) says that Hn is less
than or equal to a rank one positive operator, so has at most one positive eigenvalue.
Conversely, any symmetric matrix
(cid:18) A B
B∗ C (cid:19)
7
with C invertible is congruent to
(cid:18) A − BC −1B∗
0
0
C(cid:19) .
(The top left entry is called the Schur complement of C.) Applying this to Hn with C the
(n, n) entry, we get that Hn is congruent to
(cid:18) −Mn−1
0
0
1
knn (cid:19) .
So if Hn has only one positive eigenvalue, −Mn−1 must be negative semi-definite, and
therefore Fn must be positive semi-definite.
✷
As an application of the Corollary, consider the Dirichlet space of holomorphic functions
on the unit disk with reproducing kernel k(w, z) = 1
1− ¯wz . It is shown in [1] that this
is a Nevanlinna-Pick kernel, and in the course of the proof it is established that 1 − 1/k
is positive semi-definite (because all the coefficients in the power series are positive).
It
then follows at once from Corollary 1.12 that the Dirichlet kernel is actually a complete
¯wz log
1
Nevanlinna-Pick kernel.
2 Vector-valued Nevanlinna-Pick kernels
Let Mµ,ν denote the µ-by-ν matrices. Let us say that a kernel k has the n-point Mµ,ν
Nevanlinna-Pick property if, for any points x1, . . . , xn, and any matrices Λ1, . . . , Λn in Mµ,ν,
there exists a multiplier Ψ,
Ψ : H ⊗ Cν → H ⊗ Cµ,
such that Ψ(xi) = Λi, 1 ≤ i ≤ n, and
kTΨk = kT ∗
Ψk = kT ∗
Ψsp{kxi ⊗Cµ: 1≤i≤n}k.
We shall say that k is a vector-valued Nevanlinna-Pick kernel if k has the n point M1,ν
Nevanlinna-Pick property for all n and ν.
Theorem 2.1 Let ν ≥ n − 1. Then k has the n-point Mµ,ν Nevanlinna-Pick property for
some positive integer µ if and only if it has the property for all positive integers µ.
Proof: It is clear that the n-point Mµ,ν Nevanlinna-Pick property implies the n-point
Mπ,ν Nevanlinna-Pick property for all π smaller than µ. So it is sufficient to prove that the
8
n-point M1,ν Nevanlinna-Pick property implies the n-point Mµ,ν Nevanlinna-Pick property
for all µ.
As in the proof of Theorem 1.2, the kernel k has the n-point Mµ,ν Nevanlinna-Pick
property if and only if the positivity of the matrix
hkij(heα, eβiCµ − hΛjΛ∗
i eα, eβiCν )ii,j=n;α,β=µ
i,j=1;α,β=1
(2.2)
implies the positivity of the Schur product of (2.2) with Fn+1 ⊗ Jµ. Again, as in the proof of
Theorem 1.2, this implies that whenever K ·(Jn −G) is positive, then so is Fn+1 ·K ·(Jn −G),
for G any positive n-by-n matrix of rank less than or equal to max(ν, n).
So, if k has the n-point M1,ν Nevanlinna-Pick property, then we can choose G to be
the rank (n − 1) matrix used in the proof of Theorem 1.2, and conclude that Fn+1 has
to be positive. But the positivity of Fn+1 clearly implies that k has the n-point Mµ,ν
Nevanlinna-Pick property for all values of µ and ν.
✷
Corollary 2.3 The kernel k is a complete Nevanlinna-Pick kernel if and only if it is a
vector-valued Nevanlinna-Pick kernel.
See [3] for another approach to describing Mν,ν Nevanlinna-Pick kernels when there
is a distinguished operator (or tuple of operators) acting on H for which all the kx's are
eigenvectors.
3 Representation of Complete Nevanlinna-Pick ker-
nels
It is a consequence of Theorem 1.2 that all complete Nevanlinna-Pick kernels have a very
specific form.
Theorem 3.1 The irreducible kernel k on X is a complete Nevanlinna-Pick kernel if and
only if there is a positive semi-definite function F : X × X → D and a nowhere vanishing
function δ on X so that
k(x, y) =
δ(x)δ(y)
1 − F (x, y)
.
(3.2)
Proof: (Sufficiency): If k has the form of (3.2), then 1/k is a rank-one operator minus
a positive operator, so has exactly one positive eigenvalue, and the result follows from
Corollary 1.12.
9
(Necessity): Suppose k is a complete Nevanlinna-Pick kernel. Fix any point x0 in X.
Then the kernel
F (x, y) = 1 −
k(x, x0)k(x0, y)
k(x, y)k(x0, x0)
(3.3)
is positive semi-definite by Theorem 1.2. Let
δ(x) =
k(x0, x)
qk(x0, x0)
.
It is immediate that equation (3.2) is satisfied. As k(x, x) is positive and finite for all x,
F (x, x) must always lie in [0, 1); as F (x, y) is a positive semi-definite kernel, it follows that
F (x, y) < 1 for all x, y.
✷
Any positive definite kernel k(x, y) can be rescaled by multiplying by a nowhere-vanishing
rank-one kernel δ(x)δ(y). Let j(x, y) = δ(x)δ(y)k(x, y). Then the Hilbert space Hj is just
a rescaled copy of Hk: a function f is in Hk if and only if δf is in Hj, so Hj = δHk. The
multipliers of Hk and Hj are the same, and one space has the complete Nevanlinna-Pick
property if and only if the other one does (the matrices Fn are identical, as the scaling
factors cancel). We shall say that the kernel k is normalized at x0 if k(x0, x) = 1 for all
x; this is equivalent to scaling the kernel by qk(x0, x0)
, and means that in (4.1) δ can be
chosen to be one, and F (x, y) becomes 1 −
k(x, y)
k(x0, x)
1
.
4 The Universal Complete Nevanlinna-Pick Kernels
It follows from Theorem 3.1 that there is a universal complete Nevanlinna-Pick kernel
(actually a family of them, indexed by the cardinal numbers). Let l2
m be m-dimensional
Hilbert space, where m is any cardinal bigger than or equal to 1. Let Bm be the unit ball
in l2
m, and define a kernel am on Bm by
am(x, y) =
1
1 − hx, yi
(4.1)
m be the completion of the linear span of the functions {am(·, y) : y ∈ Bm}, with inner
m are
Let H 2
product defined by ham(·, y), am(·, x)i = am(x, y). We shall show that the spaces H 2
universal complete Nevanlinna-Pick spaces.
10
Theorem 4.2 Let k be an irreducible kernel on X. Let m be the rank of the Hermitian
form F defined by (3.3). Then k is a complete Nevanlinna-Pick kernel if and only if there
is an injective function f : X → Bm and a nowhere vanishing function δ on X such that
k(x, y) = δ(x)δ(y) am(f (x), f (y)).
(4.3)
Moreover if this happens, then the map kx 7→ δ(x)(am)f (x) extends to an isometric linear
embedding of Hk into δH 2
m.
If in addition there is a topology on X so that k is continuous on X × X, then the map
f will be a continuous embedding of X into Bm.
Proof: (Sufficiency): Any kernel of the form (4.3) is of the form (3.2).
(Necessity): Suppose k is a complete Nevanlinna-Pick kernel. As F is positive semi-
m) and a
m so that F (x, y) = hf (x), f (y)i. Moreover, as F takes value in D, f actuallly
definite, there exists a Hilbert space of dimension m (which we shall take to be l2
map f : X → l2
maps into Bm. It now follows from Theorem 3.1 that k has the form (4.3).
The linear map that sends kx to the function
X} by (4.3) and gives the desired embedding.
δ(x)
1 − hf (x), ·i
is an isometry on ∨{kx : x ∈
If f (x) = f (y) then kx = ky; as k is positive definite, this implies x = y.
Finally, f can be realised as the composition of the four maps
x 7→ kx
kx
7→ δ(x)am(f (x), ·)
δ(y)am(y, ·)
7→ am(y, ·)
am(y, ·)
7→ y
The fourth map is continuous by direct calculation, the second is an isometry by the theorem,
and the first and third maps are continuous if k is continuous.
✷
Note that if one first normalizes k at some point, δ can be taken to be 1 in Theorem 4.2.
For m = 1, the space H 2
m is the regular Hardy space on the unit disk. For larger m, it is
a Hilbert space of analytic functions on the ball Bm. Thus every reproducing kernel Hilbert
space with the complete Nevanlinna-Pick property is a restriction of a space of analytic
functions.
It was shown in [2] that the Sobolov space W 1,2[0, 1], the functions g on the unit interval
0 g2 + g′2dx is finite, has the Nevanlinna-Pick property. It follows from [8,
for which R 1
11
Cor. 6.5] that the condition of Corollary 1.12 is satisfied, so the Sobolov space has the
complete Nevanlinna-Pick property. We can normalize W 1,2[0, 1] at 1 say, by calculating
that k1(t) = cosinh(1) cosh(t), and hence δ(t) = qsinh(1) cosh(1)cosech(t). Therefore there
is a continuous embedding f : [0, 1] → Bℵ0 so that if g is any function in W 1,2[0, 1], then
(δ.g) ◦ f −1 extends off the curve f ([0, 1]) to be analytic on all of Bℵ0 - even though δ.g need
not be analytic in any neighborhood of the unit interval on which it is originally defined.
After normalization, every separable reproducing kernel Hilbert space with the complete
Nevanlinna-Pick property is the restriction of the space H 2
ℵ0 to a subspace spanned by a
set of kernel functions, which is why we call this space universal. The kernel k is just the
restriction of aℵ0 to a subset of Bℵ0.
Let A be a normed algebra of functions on a set X with the complete Nevanlinna-Pick
property, i.e. there exists a positive definite function k on X × X such that there is a
function f in A ⊗ Mk of norm at most one and with f (xi) = Λi if and only if the nk-by-nk
matrix
k(xi, xj) ⊗ [Ik − Λ∗
i Λj]
is positive. It is then immediate that A is the multiplier algebra of Hk, and k is a complete
Nevanlinna-Pick kernel.
If Hk is separable, k is therefore the restriction of aℵ0 to some
subset of Bℵ0. By the Pick property, every function in A extends to an element of the
multiplier algebra of H 2
ℵ0 without increasing the norm. So every separably acting algebra
with the complete Nevanlinna-Pick property embeds isometrically in the multiplier algebra
of H 2
ℵ0.
It is probably the universality of the kernel am which is responsible for the recent surge
of interest in it - see e.g. [3, 4, 5, 6].
References
[1] J. Agler. Some interpolation theorems of Nevanlinna-Pick type. To appear.
[2] J. Agler. Nevanlinna-Pick interpolation on Sobolov space. Proc. Amer. Math. Soc.,
108:341 -- 351, 1990.
[3] J. Agler and J.E. McCarthy. Nevanlinna-pick kernels and localization. To appear.
[4] A. Arias and G. Popescu. Noncommutative interpolation and Poisson transforms. To
appear.
12
[5] W.B. Arveson. Subalgebras of C*-algebras III: Multivariable operator theory. To appear.
[6] K.R. Davidson and D.R. Pitts. Nevanlinna-Pick interpolation for non-commutative
analytic Toeplitz algebras. To appear.
[7] S. McCullough. Carath´eodory interpolation kernels. Integral Equations and Operator
Theory, 15(1):43 -- 71, 1992.
[8] P. Quiggin. For which reproducing kernel Hilbert spaces is Pick's theorem true? Integral
Equations and Operator Theory, 16(2):244 -- 266, 1993.
13
|
1512.00610 | 1 | 1512 | 2015-12-02T08:37:44 | Some results on almost square Banach spaces | [
"math.FA"
] | We study almost square Banach spaces under a topological point of view. Indeed, we prove that the class of Banach spaces which admits an equivalent norm to be ASQ is that of those Banach spaces which contain an isomorphic copy of $c_0$. We also prove that the symmetric projective tensor products of an almost square Banach space have the strong diameter two property | math.FA | math |
SOME RESULTS ON ALMOST SQUARE BANACH SPACES
JULIO BECERRA GUERRERO, GIN´ES L ´OPEZ-P´EREZ AND ABRAHAM RUEDA
ZOCA
Abstract. We study almost square Banach spaces under a topological
point of view. Indeed, we prove that the class of Banach spaces which
admits an equivalent norm to be ASQ is that of those Banach spaces
which contain an isomorphic copy of c0. We also prove that the symmet-
ric projective tensor products of an almost square Banach space have
the strong diameter two property.
1. Introduction
The study of the size of slices, non-empty relatively weakly open and
convex combination of slices of the unit ball of a Banach space has emerged
in the last few years. A Banach space X is said to have the slice diameter two
property (respectively diameter two property, strong diameter two property)
if every slice (respectively non-empty relatively weakly open subset, convex
combination of slices of the unit ball) has diameter two. Such properties,
which have been proved to be different in an extreme way [7], have show to
have strong links with other properties of the geometry of a Banach space
such as having octahedral norms [9]. Under this frame, a new class of Banach
spaces have recently appeared: the so-called almost square Banach spaces.
According to [3], a Banach space X is said to be
(1) locally almost square (LASQ) if for every x ∈ SX there exists a
sequence {yn} in BX such that kx ± ynk → 1 and kynk → 1.
(2) weakly almost square (WASQ) if for every x ∈ SX there exists a
sequence {yn} in BX such that kx±ynk → 1, kynk → 1 and {yn} → 0
weakly.
(3) almost square (ASQ) if for every x1, . . . , xn elements of SX there
exists a sequence {yn} in SX such that kynk → 1 and kxi ± ynk → 1
for every i ∈ {1, . . . , n}.
On the one hand it is obvious that WASQ Banach spaces are LASQ Ba-
nach spaces, and it is also known that ASQ Banach spaces are WASQ Ba-
nach spaces [3, Theorem 2.8]. On the other hand there are several examples
of Banach spaces which are ASQ as c0 [3, Example 3.1], the Hagler space
JH [6, Lemma 3.1], non-reflexive M -embedded Banach spaces [3, Corollary
4.3]...
1
2
J. Becerra, G. L´opez and A. Rueda
In [3] it is pointed out the nice relation between almost square Banach
spaces and diameter two properties.
Indeed, LASQ (respectively WASQ,
ASQ) Banach spaces enjoy to have the slice diameter two property (respec-
tively the diameter two property, strong diameter two property). However,
in such paper not only do the autors study almost square Banach spaces
from a geometrical point of view but also from an isomorphic one.
Indeed, every ASQ Banach space contains an isomorphic copy of c0 [3,
Lemma 2.6]. In addition, as the property of being an ASQ Banach space
is preserved under taking ℓ∞ sums with any other Banach space, Banach
spaces containing a complemented copy of c0 can be equivalently renormed
to be ASQ. Consequently, a separable Banach space X can be equivalently
renormed to be ASQ if, and only if, X contains an isomorphic copy of c0
[3, Corollary 2.10]. Moreover, the authors ask whether the hypothesis of
separability can be eliminated.
The main aim of this note is to provide a positive answer to the above
question, proving that every Banach space containing an isomorphic copy
of c0 can be equivalently renormed to be ASQ. To this aim, in section 2, we
will firstly renorm the Banach space ℓ∞ to be ASQ. Then we will prove that
every Banach space containig an isomorphic copy of c0 can be equivalently
renormed to be ASQ by giving a suitable renorming of the bidual space.
Section 3 will be devoted to analyse the the relations between ASQ Banach
spaces and the strong diameter two property in symmetric projective ten-
sor products. In the last few years several papers have appeared related to
diameter two properties in tensor products spaces (see [1, 2, 8]) and, even
though it has been proved stability results of the slice diameter two property
and strong diameter two property in projective tensor products of Banach
spaces (see [8]), such results seem to be unknown for symmetric tensor prod-
ucts. So, in Section 3, we shall prove that the symmetric projective tensor
products of ASQ Banach spaces have the strong diameter two property. Fi-
nally, Section 4 we will be devoted to exhibit open problems related to ASQ
spaces.
We shall now introduce some notation. We will consider real Banach
spaces. BX, respectively SX, stands for the closed unit ball, respectively the
unit sphere, of the Banach space X. We denote by X ∗ the topological dual
space of X. Given I a non-empty set and U an ultrafilter on I, recall that U
is said to be principal if there exists i ∈ I such that U := {Y ⊆ I / i ∈ Y }.
In other case, U is said to be non-principal (see [17] for background). Given
f : I −→ R a bounded function we will denote by
f
lim
U
the limit of f under the ultrafilter U . It is well known that in the particular
case I = N, there are non-principal ultrafilters. Moreover, given U a non-
principal ultrafilter it follows that
lim
U
x = lim
n→∞
x(n)
Some results on almost square Banach spaces
3
for every convergent sequence x.
Finally, according to [1], for a Banach space X and N ∈ N, we will denote
under the norm given by
by b⊗π,s,N X the symmetric projective N-tensor product of X. This space is
the completion of the linear space generated by(cid:26)xN :=
x ⊗ . . . ⊗ x : x ∈ X(cid:27)
i , λi ∈ R, xi ∈ SX ∀i ∈ {1, . . . , k}) .
kzk := inf( kXi=1
kXi=1
λi : z =
It is well known that its topological dual space is identified with the space
of all N -homogeneous and bounded polynomials on X by the action
λixN
N
P(cid:0)xN(cid:1) := P (x) ∀x ∈ X
for each N -homogeneous and bounded polynomial P (see [13] for back-
ground). We will denote by P(N X, Y ) the space of N -homogeneous poly-
nomials from X to Y .
2. A renorming theorem for ASQ Banach spaces
It is clear that each Banach space containing a complemented copy of c0
can be equivalently renormed to be ASQ. Since c0 is not complemented in
ℓ∞ [11, Theorem 5.15] it seems natural to wonder whether ℓ∞, which is not
an ASQ space, can be equivalently renormed to be ASQ.
In order to exhibit such a norm on ℓ∞, consider U a non-principal ultra-
filter on N and define, given x ∈ ℓ∞, the following
lim(x) := lim
U
x.
Then lim : ℓ∞ −→ R is linear and continuous. In fact, it is easy to prove
that k lim k = 1.
Now we are ready to prove the following
Theorem 2.1. There exists an equivalent norm on ℓ∞, say · , such that
the Banach space (ℓ∞, · ) is an ASQ Banach space.
Proof. Consider on ℓ∞ the norm given by
x := max(cid:26) lim(x), sup
n
x(n) − lim(x)(cid:27) ,
Let us prove that the norm defined above is equivalent to the classical one
on ℓ∞. To this aim consider x ∈ ℓ∞. Now, on the one hand
x ≤ max(cid:26) lim(x), sup
n
x(n) + lim(x)(cid:27) ≤ kxk∞ + kxk∞ = 2kxk∞,
as k lim k = 1. On the other hand
x ≥ sup
x(n) − lim(x) ≥ sup
x(n) − lim(x).
n
n
4
Now
J. Becerra, G. L´opez and A. Rueda
kxk∞ ≤ x + lim(x) ≤ lim(x) + sup
n
x(n) − lim(x) + lim(x)
≤ 2( lim(x)+sup
n
x(n)−lim(x)) ≤ 4 max(cid:26) lim(x), sup
n
x(n) − lim(x)(cid:27) = 4x.
So · and k · k∞ are equivalent norms. Let us now prove that (ℓ∞, · )
is an ASQ Banach space. To this aim pick x1, . . . , xn ∈ Sℓ∞ and ε > 0.
Given i ∈ {1, . . . , n} consider the sets
Ai := {n ∈ N / xi(n) − lim xi < ε}.
Then A1, . . . , An ∈ U by the definition of the limit by ultrafilter, so A :=
Ai ∈ U as a finite intersection of elements of U . Since U is an ultrafilter
nTi=1
we have that A 6= ∅, so pick n ∈ A. Now let us estimate xi ± en ∀i ∈
{1, . . . , n}. To this aim pick i ∈ {1, . . . , n}. Then, on the one hand,
since lim en = 0. On the other hand
lim(xi ± en) = lim(xi)
sup
xi(k) ± en(k) − lim(xi ± en)
k
= max(sup
≤ max(sup
k6=n
k6=n
xi(k) − lim(xi), xi(n) − lim(xi) ± 1)
xi(k) − lim(xi), xi(n) − lim(xi) + 1)
≤ max(cid:26)sup
n
xi(n) − lim(xi), 1 + ε(cid:27) .
Consequently, by definition of the norm · , one has
xi ± en ≤ max{xi, 1 + ε} = 1 + ε.
Moreover, check that
and
lim(en) = 0
sup
en(k) − lim(ek) = 1.
k
Thus en ∈ Sℓ∞. From [3, Proposition 2.1] we get that (ℓ∞, · ) is ASQ
Banach space, as desired.
Remark 2.2. From the above proof it follows that, given x1, . . . , xn ∈ Sℓ∞
and ε > 0 we can find y ∈ Sc0 such that xi ±y ≤ 1+ε. Roughly speaking
we can say that the fact that ℓ∞ under the norm of above Theorem is ASQ
relies on the subspace c0. This simple observation will be the key of the
general renorming result.
Some results on almost square Banach spaces
5
As we have pointed out above, the Banach space ℓ∞ plays an important
role as example of Banach space containing an isomorphic copy of c0 which
can be equivalently renormed to be ASQ. However, as dual Banach spaces
containing an isomorphic copy of c0 actually contain a complemented copy
of ℓ∞ [16, Proposition 2.e.8], we can deduce our general result from this
particular example by giving a suitable renorming in the bidual space.
Theorem 2.3. Let X be a Banach space containing an isomorphic copy of
c0. Then there exists an equivalent norm on X such that X is an ASQ space
under the new norm.
Proof. Assume that X contains a subspace Y which is isometric to c0.
As Y ∗∗ ⊆ X ∗∗ is linearly isometric to ℓ∞, then Y ∗∗ is complemented in
X ∗∗ [11, Proposition 5.13].
Then we can consider on X ∗∗ an equivalent norm so that
and such norm agrees with the original one of Y ∗∗.
X ∗∗ = Y ∗∗ ⊕∞ Z,
Now we can consider on Y ∗∗ the norm defined in Theorem 2.1, so Y ∗∗
becomes into an ASQ space and agrees with the original norm on Y ⊆ X.
This defines an equivalent norm on X ∗∗ which we will denote by k·k. Clearly
X ∗∗ is an ASQ space [3, Proposition 5.7]. Our aim is to prove that X is an
ASQ space following similar ideas to [3, Proposition 5.7] and Remark 2.2.
To this aim pick x1, . . . , xn ∈ SX and ε > 0. Now xi ∈ X ∗∗ = Z ⊕∞ Y ∗∗
for each i ∈ {1, . . . , n}, so we can find zi ∈ Z and yi ∈ Y ∗∗ such that xi =
(zi, yi) ∀i ∈ {1, . . . , n}. We can assume, making a perturbation argument if
necessary, that yi 6= 0 ∀i ∈ {1, . . . , n}. From Remark 2.2 we can find y ∈ Sc0
such that
(2.1)
Define z := (0, y) ∈ Sc0 ⊆ X. Then
kxi ± zk = max{kzik, kyi ± yk} ≤ max{1, kyi ± yk}
yi
kyik
(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:26)1,(cid:13)(cid:13)(cid:13)(cid:13)kyik(cid:18) yi
± y(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1 + ε.
± y(cid:19) ± (1 − kyik)y(cid:13)(cid:13)(cid:13)(cid:13)(cid:27)
kyik
(2.1)
≤ max {1, kyik(1 + ε) + (1 − kyik)kyk} ≤ 1 + ε.
To sum up we have proved that given x1, . . . , xn ∈ SX and ε > 0 we can
find z ∈ SX such that
kxi ± zk ≤ 1 + ε.
Thus X is an ASQ space under the new equivalent norm, so we are done.
6
J. Becerra, G. L´opez and A. Rueda
Above Theorem allows us to strengthen [4, Proposition 4.7], where it is
proved that every Banach space containing an isomorphic copy of c0 can be
equivalently renormed to have the strong diameter two property.
Moreover, from Theorem 2.3 and [3, Theorem 2.4] we get the following
Corollary 2.4. Let X be a Banach space. Then there exists an equivalent
norm on X such that X is an ASQ Banach space under the new norm if,
and only if, X contains an isomorphic copy of c0.
In [3] the relation between ASQ Banach spaces and the intersection prop-
erty is pointed out. Recall that a Banach space X has the intersection
property if for every ε > 0 there exist x1, . . . , xn ∈ X such that kxik < 1
and such that if y ∈ X verifies that kxi − yk ≤ 1 for every i ∈ {1, . . . , n}
then kyk ≤ ε. Given 0 < ε < 1, X is said to ε-fail the intersection property
if γ(ε) = 1, where
γ(ε) :=
sup
x1,...,xn∈B[0,1)
inf
y∈B(ε,1]
max
1≤i≤n
kxi − yk ∀ 0 < ε < 1,
and BI := {x ∈ X / kxk ∈ I} for each I ⊆ R+. Finally, a Banach space
is said to fail the intersection property if X ε-fails the intersection property
for some 0 < ε < 1.
On the one hand, it is known that a Banach space X is ASQ if, and only
if, X ε-fails the intersection property for every 0 < ε < 1 [3, Proposition
6.1]. On the other hand, it is known that a Banach space admits an equiv-
alent norm which fails the intersection property if, and only if, X contains
an isomorphic copy of c0 [14, Theorem 1.7]. Now we can improve above
Theorem as an straightforward application of Corollary 2.4.
Theorem 2.5. Let X be a Banach space. Then X admits an equivalent
norm which ε-fails the intersection property for each 0 < ε < 1 if, and only
if, X contains an isomorphic copy of c0.
3. ASQ Banach spaces and symmetric tensor products
One of the most important fact when one tries to prove that ASQ Banach
spaces have the strong diameter two property is that in such spaces we have
a lot of weakly-null sequences which are equivalent to the c0 basis (see [3,
Lemma 2.6]). So, in order to prove that the symmetric projective tensor
products of an ASQ Banach space have the strong diameter two property,
we shall begin by proving the following Lemma, which asserts that such
sequences are still weakly-null when they are considered in the tensor space.
Lemma 3.1. Let X be a Banach space and consider {yn} ⊆ SX a sequence
equivalent to the usual basis of c0. Pick N ∈ N. Then {eN
n } → 0 in the weak
topology of Y := b⊗π,s,N X.
Proof. Pick P a N -homogenous polynomial on X. Define Y := span{en :
n ∈ N}, which is a subspace of X which is isomorphic to c0. Consider
Some results on almost square Banach spaces
7
Q = PY , which is a polynomial in Y . In fact, if P (x) = M (x, . . . , x) for
suitable N -lineal form M , then
Q(y) = MY N (y, . . . , y) ∀y ∈ Y.
For each n ∈ N it follows that
P (en) = Q(en).
Moreover, as each polynomial on Y is weakly sequentially continuous (be-
cause Y is linearly isomorphic to c0 and such space has the polynominal
Dunford-Pettis property [15]) we conclude that {Q(en)} → 0, so we are
done.
Note that given a Banach space X with the strong diameter two property
we can find, in every convex combination of slices of its unit ball, elements
whose norm is as close to 1 as desired (see [8, Lemma 2.1]). In order to prove
the announced result we shall before verify that such property is satisfied
by every symmetric projective tensor product of an ASQ Banach space.
Lemma 3.2. Let X be an ASQ space and N ∈ N. Consider Y := b⊗π,s,N X
and C :=Pk
i=1 λiS(BY , Pi, αi) a convex combination of slices of BY . Then,
for each ε > 0, we can find f ∈ SX ∗ and yN
i ∈ Si such that
f (yi) > 1 − ε ∀i ∈ {1, . . . , k}.
Proof. For each i ∈ {1, . . . , k} consider xi ∈ SX such that xN
i ∈ Si, in other
words, Qi(xi) > 1−αi for each i ∈ {1, . . . , k}. Since Qi is an N -homogeneous
polynomial for each i ∈ {1, . . . , k} we can find ε0 > 0 such that
0 < ε < ε0 ⇒ Qi(cid:18) xi
1 + ε(cid:19) > 1 − αi ∀i ∈ {1, . . . , k}.
Following [3, Lemma 2.6] we can find {yn} a weakly-null sequence in SX
which is 1 + ε-isometric to the usual basis of c0 such that
kxi ± ynk → 1 ∀i ∈ {1, . . . , k}.
Now, as {yN
n } → 0 weakly in Y because of Lemma 3.1 we conclude that
{(xi ± yn)N } → xi in the weak topology of Y for each i ∈ {1, . . . , k}. In
fact, given i ∈ {1, . . . , k}, one has that {(xi ± yn)N } → xi if, and only if,
{(xi ± yn − xi)N } → 0 [12, Lemma 1.1].
From facts above we can find n large enough to ensure
and
kxi ± ynk
1 + ε
≤ 1 ∀i ∈ {1, . . . , k},
Qi(cid:18) xi ± yn
1 + ε (cid:19) > 1 − αi ∀i ∈ {1, . . . , k}.
8
J. Becerra, G. L´opez and A. Rueda
Now, on the one hand,Pk
∈ C. On the other hand, consider
f ∈ SX ∗ such that f (yn) = 1. As f (xi ± yn) ≤ 1 + ε for each i ∈ {1, . . . , k}
we conclude that
i=1 λi
(xi±yn)N
1+ε
f (xi) ≤ ε ∀i ∈ {1, . . . , k}.
Now, defining yi := xi+yn
1+ε , one has
f (xi) + f (yn)
1 + ε
f (yi) =
>
1 − ε
1 + ε
∀i ∈ {1, . . . , k}.
As 0 < ε < ε0 was arbitrary we get the desired result.
Now we are ready to prove the main result of the section.
has the strong diameter two property.
Theorem 3.3. Let X be an ASQ space and N ∈ N. Then Y := b⊗π,s,N X
such that Pk
Proof. Pick k ∈ N and Si := S(BY , Qi, αi) slices of BY , where Q1, . . . , Qk
are norm-one N -homogeneous polynomials on X and pick λ1, . . . , λk ∈ [0, 1]
i=1 λiSi a convex combination of
slices. Our aim is to prove that diam(C) = 2. To this aim pick ε > 0.
In view of above lemma we can find xN
i ∈ Si for each i ∈ {1, . . . , k} and
g ∈ SX ∗ such that
i=1 λi = 1. Define C := Pk
g(xi) > 1 − ε ∀i ∈ {1, . . . , k}.
Now, again from computations of above Lemma, we can find y ∈ SX such
that
(xi ± y)N
1 + ε
∈ Si ∀i ∈ {1, . . . , k}.
(xi±y)N
1+ε
∈ C. On the other hand, let us
i=1 λi
estimate
Now, on the one hand, Pk
diam(C) ≥(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
kXi=1
λi
(xi + y)N − (xi − y)N
1 + ε
.
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
To this aim, pick f ∈ SX ∗ such that f (y) = 1. As f (xi ± y) ≤ 1 + ε ∀i ∈
{1, . . . , k} we conclude that
Now let us argue by cases:
f (xi) ≤ ε ∀i ∈ {1, . . . , k}.
• If N is odd, define P (z) := f (z)N for each z ∈ X. Then
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
kXi=1
kXi=1
λi
=
λi
(xi + y)N − (xi − y)N
1 + ε
λi
P (x + y) − P (x − y)
1 + ε
f (xi + y)N − f (xi − y)N
f (xi + y)N + f (y − xi)N
1 + ε
≥
(1 − ε)N + (1 − ε)N
1 + ε
λi = 2
1 + ε
(1 − ε)N
1 + ε
.
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=
λi
≥
kXi=1
kXi=1
kXi=1
Some results on almost square Banach spaces
9
• If N is even, define P (z) := f (z)N −1g(z) for each z ∈ X. Again,
kP k ≤ 1. Now we conclude
λi
(xi + y)N − (xi − y)N
1 + ε
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
kXi=1
=
λi
kXi=1
≥
λi
kXi=1
P (xi + y) − P (xi − y)
1 + ε
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
f (xi + y)N −1g(xi + y) − f (xi − y)N −1g(xi − y)
g(xi)(f (xi + y)N −1 − f (xi − y)N −1) + g(y)(f (xi + y)N −1 + f (xi − y)N −1)
.
=
kXi=1
λi
1 + ε
1 + ε
Now, as N − 1 is odd, we can estimate f (xi + y)N −1 − f (xi − y)N −1
by 2(1 − ε)N −1 as in the above case for each i ∈ {1, . . . , k}. On
the other hand, check that g(y) ≤ 2ε as g(xi ± y) ≤ 1 + ε and
g(xi) > 1 − ε for each i ∈ {1, . . . , k}. Bearing in mind that g(xi) >
1 − ε ∀i ∈ {1, . . . , k} we conclude that
diam(W ) ≥
2(1 − ε)N − 4ε(1 + ε)N −1
1 + ε
.
In any case, as 0 < ε < ε0 was arbitrary we conclude that diam(C) = 2
as desired.
Remark 3.4. Check that last above estimates are similar to the ones of [1].
Consequently, in [1, Proposition 2.4] can be obtained the strong diameter
two property under the same assumptions.
Now, as an easy consequence of Theorems 2.3 and 3.3, we get the following
Corollary 3.5. Let X be a Banach space which contains an isomorphic copy
of c0. Then there exists an equivalent norm on X such that for each N ∈ N
two property.
the projective symmetric tensor product b⊗π,s,N X has the strong diameter
According to [9], the norm on a Banach space X is said to be octahedral
if for every ε > 0 and every Y ⊆ X finite-dimensional subspace there exists
x ∈ SX such that
ky + λxk ≥ (1 − ε)(kyk + λ) ∀y ∈ Y, ∀λ ∈ R.
It is known that a Banach space X an octahedral norm if, and only if, X ∗ has
the weak-star strong diameter two property [9, Theorem 2.1]. Consequently,
from Theorem 3.3 we conclude that given an ASQ Banach space X then
P(N X) has an octahedral norm for each N ∈ N. However, we can go
further bearing in mind the results of [8].
Corollary 3.6. Let X and Y be Banach spaces. If X is an ASQ space and
Y has an octahedral norm, then P(N X, Y ) has an octahedral norm for each
N ∈ N.
10
J. Becerra, G. L´opez and A. Rueda
Proof. As P(N X, Y ) and L(b⊗π,s,N X, Y ) are linearly isometric [13] for each
N ∈ N, the Corollary follows from Theorem 3.3, [9, Theorem 2.1] and [8,
Theorem 2.5].
4. Some remarks and open questions
In [3] it is posed as an open question whether there exists a dual Banach
space which is ASQ. In view of Theorem 2.1 it seems natural take advantage
of Theorem 2.3 for dual Banach spaces. Unfortunately, the technique ex-
posed in such result does not respect the duality of Banach spaces. Indeed,
we have the following
Proposition 4.1. Let · be the ASQ norm on ℓ∞. Then Ext(Bℓ∞) = ∅.
As a consequence, (ℓ∞, · ) is not isometric to any dual Banach space.
Proof. Consider x ∈ Sℓ∞. Then we have the following considerations:
(1) If lim(x) < 1 then there exists ε > 0 such that lim(x) ± ε < 1.
Now consider
y := x + ε1 z := x − ε1.
Then clearly lim(y) ≤ 1 and lim(z) ≤ 1. On the other hand,
given n ∈ N one has
y(n) − lim(y) = yn + ε − lim(x) − ε = x(n) − lim(x) ≤ x ≤ 1.
So y ∈ Bℓ∞. By a similar argument we have that z ∈ Bℓ∞. As
x = y+z
2 we get that x /∈ Ext(Bℓ∞) in this case.
(2) If lim(x) = 1, we shall assume with no loss of generality that
lim(x) = 1. Then supn x(n) − 1 ≤ 1 from where we conclude that
x(n) ≥ 0 ∀n ∈ N. Moreover as lim(x) = 1 then we can find ε > 0
and n ∈ N such that x(n) ≥ ε. We claim that
Indeed,
x ± εen ∈ Bℓ∞.
lim(x ± εen) = lim(x).
Moreover, given k ∈ N,
x(k) ± εen(k) − lim(x ± εen) = x(k) ± εδkn − lim(x)
If k 6= n clearly last quantity is less than or equal to x ≤ 1.
Moreover, if k = n then we have
x(n) ± ε − lim(x) = x(n) ± ε − 1 ≤ x(n) − 1 + ε = 1 − x(n) + ε ≤ 1.
Thus x ± εen ∈ Bℓ∞, so x /∈ Ext(Bℓ∞).
This proves that Ext(Bℓ∞) = ∅. Now ℓ∞ is not a dual Banach space as
an easy consequence of Krein-Milman theorem.
Some results on almost square Banach spaces
11
From above Proposition we get that we can not give a dual renorming
of the bidual of a Banach space containing an isomorphic copy of c0 using
the ideas of Theorem 2.3. Indeed, given X a Banach space containing an
isomorphic copy f c0 we have considered a renorming of X ∗∗ such that
X ∗∗ = ℓ∞ ⊕∞ Z,
where we consider on ℓ∞ the renorming of Theorem 2.1. By above Propo-
sition the unit ball of ℓ∞ does not have any extreme point. Consequently,
so does the unit ball of X ∗∗ and, again by Krein-Milman theorem, X ∗∗ can
not be isometric to any dual Banach space. So it remains open wheter there
exists a dual Banach space which is ASQ. It is even open wether there exists
a dual Banach space failing the intersection property [10, Section 4]
It is also posed as an open question in [3] if there exists any LASQ Banach
space which fails to be WASQ. In this direction recall that it is proved that
every Banach space which contains an isomorphic copy of c0 can be equiva-
lently renormed to have the slice diameter two property and whose new unit
ball contains non-empty relatively weakly open subset whose diameter is as
small as desired [5, Theorem 2.4]. Similarly, it is proved that every Banach
space which contains an isomorphic copy of c0 can be equivalently renormed
to have the diameter two property and whose new unit ball contains convex
combinations of slices whose diameter is as small as desired [7, Theorem
2.5]. Thus we can go further and pose the following:
Question 4.2. Let X be a Banach space containing an isomorphic copy of
c0.
(1) Is there an equivalent norm on X which is LASQ and fails to be
WASQ (or even its new unit ball contains non-empty relatively
weakly open subsets whose diameter is as small as desired)?
(2) Is there an equivalent norm on X which is WASQ and fails to be
ASQ (or even its new unit ball contains convex combinations of slices
whose diameter is as small as desired)?
Finally, bearing in mind the results exposed in Section 3, we will pose the
following
an ASQ Banach space?
Question 4.3. Let X be an ASQ Banach space and pick N ∈ N. Isb⊗π,s,N X
References
[1] M.D. Acosta and J. Becerra Guerrero, Weakly open sets in the unit ball of some
Banach spaces and the centralizer, Jour. Funct. Anal. 259 (2010) 842-856.
[2] M.D. Acosta, J. Becerra Guerrero and A. Rodr´ıguez-Palacios, Weakly open sets
in the unit ball of the projective tensor product of Banach spaces, J. Math. Anal.
Appl. 383 (2011) 461-473.
[3] T. Abrahamsen, J. Langemets and V. Lima, Almost square Banach spaces, Jour.
Math. Anal. App. 434, 2 (2016), 1549-1565.
[4] T.A. Abrahamsen, V. Lima and O. Nygaard, Remarks on diameter two proper-
ties, J. Convex Anal. (2013).
12
J. Becerra, G. L´opez and A. Rueda
[5] J. Becerra Guerrero, G. L´opez P´erez and A. Rueda Zoca, Big slices versus rel-
atively weakly open subsets in Banach spaces, J. Math. Anal. App. 428 (2015)
855-865.
[6] J. Becerra Guerrero, G. L´opez P´erez and A. Rueda Zoca, Diameter two properties
in James spaces, Ban. Jour. Math. Anal, vol. 9, n.4 (2015).
[7] J. Becerra Guerrero, G. L´opez P´erez and A. Rueda Zoca, Extreme differences
between weakly open subsets and convex combination of slices in Banach spaces,
Adv. in Math. 269 (2015), 56-70.
[8] J. Becerra Guerrero, G. L´opez P´erez and A. Rueda Zoca, Octahedral norms in
spaces of operators, J. Math.Anal.Appl. 427 (2015), 171-184.
[9] J. Becerra Guerrero, G. L´opez P´erez and A. Rueda Zoca, Octahedral norms and
convex combination of slices in Banach spaces, Jour. Funct. Anal. 266 (2014)
2424-2435
[10] E. Behrends and P. Harmand, Banach spaces which are proper M-ideals, Studia
Math. 81 (1985), no. 2, 159-169.
[11] M. Fabian, P. Habala, P. H´ajek, V. Montesinos, J. Pelant, V. Zizler, Func-
tional Analysis and Infinite-Dimensional Geometry, CMS Books in Mathematics.
Springer-Velag New York 2001.
[12] Farmer, J.D., and W.B. Johnson, Polynomial Schur and polynomial Dunford-
Pettis properties, Contemp. Math. 144 (1993), 95-105.
[13] K. Floret, Natural norms on symmetric tensor products of Banach spaces, Note
Mat. 171 (1997) 153-188.
[14] P. Harmand and T. S. S. R. K. Rao, An intersection property of balls and relations
with M-ideals, Math. Z. 197 (1988), no. 2, 277-290.
[15] J.A. Jaramillo, A. Prieto and I. Zalduendo, Sequential convergences and Dunford-
Pettis properties, Mathematica 25 (2000), 467-475.
[16] J. Lindestrauss, L. Tzafriri, Classical Banach spaces, Springer Verlag Berlin
(1977).
[17] A.Wilansky, Topology for Analysis, John Wiley (1970).
Universidad de Granada, Facultad de Ciencias. Departamento de An´alisis
Matem´atico, 18071-Granada (Spain)
E-mail address: [email protected], [email protected], [email protected]
|
1812.11299 | 1 | 1812 | 2018-12-29T06:47:07 | New counterexamples on Ritt operators, sectorial operators and R-boundedness | [
"math.FA"
] | Let $\mathcal D$ be a Schauder decomposition on some Banach space $X$. We prove that if $\mathcal D$ is not $R$-Schauder, then there exists a Ritt operator $T\in B(X)$ which is a multiplier with respect to $\mathcal D$, such that the set $\{T^n\, :\, n\geq 0\}$ is not $R$-bounded. Likewise we prove that there exists a bounded sectorial operator $A$ of type $0$ on $X$ which is a multiplier with respect to $\mathcal D$, such that the set $\{e^{-tA}\, : \, t\geq 0\}$ is not $R$-bounded. | math.FA | math |
NEW COUNTEREXAMPLES ON RITT OPERATORS, SECTORIAL
OPERATORS AND R-BOUNDEDNESS
LORIS ARNOLD AND CHRISTIAN LE MERDY
Abstract. Let D be a Schauder decomposition on some Banach space X. We prove that
if D is not R-Schauder, then there exists a Ritt operator T ∈ B(X) which is a multiplier
with respect to D, such that the set {T n : n ≥ 0} is not R-bounded. Likewise we prove
that there exists a bounded sectorial operator A of type 0 on X which is a multiplier with
respect to D, such that the set {e−tA : t ≥ 0} is not R-bounded.
2000 Mathematics Subject Classification: 47A99, 46B15.
R-boundedness plays a prominent role in the study of sectorial operators and Ritt opera-
tors. Namely the notions of R-sectorial operators and R-Ritt operators have been instrumen-
tal in the development of H ∞-functional calculus, square function estimates and applications
to maximal regularity and to many other aspects of the harmonic analysis of semigroups (in
either the continuous or the discrete case).
The existence of sectorial operators which are not R-sectorial was discovered by Kalton
and Lancien in their paper solving the Lp-maximal regularity problem [6]. The existence
of Ritt operators which are not R-Ritt was established a bit later by Portal [14]. More
recently, Fackler [4] extended the work of Kalton-Lancien in various directions. In contrast
with [6], which focused on existence results, [4] supplied explicit constructions of sectorial
operators which are not R-sectorial. Further it is easy to derive from the latter paper explicit
constructions of Ritt operators which are not R-Ritt. In [4, 6, 14], sectorial operators which
are not R-sectorial (resp. Ritt operators which are not R-Ritt) are defined as multipliers
with respect to Schauder decompositions having various "bad" properties.
In particular,
these Schauder decompositions cannot be R-Schauder (see Lemma 0.2).
The aim of this note is two-fold. First we show that given any Schauder decomposition
D which is not R-Schauder, one can define a sectorial operator A which is a multiplier with
respect to D and which is not R-sectorial (resp. a Ritt operator T which is a multiplier
with respect to D and which is not R-Ritt). Second we strengthen these negative results in
both cases by showing that A can be chosen bounded and such that {e−tA : t ≥ 0} is not
R-bounded, whereas T is taken such that {T n : n ≥ 0} is not R-bounded. (See Remark 0.6
for more comments.)
Date: January 1, 2019.
1
2
L. ARNOLD AND C. LE MERDY
In addition to the above mentioned papers, we refer the reader to [3, 8, 12, 16] for rele-
vant information on R-sectorial and R-Ritt operators. We also mention [9] which contains
examples of Ritt operators which are not R-Ritt. They are of a different nature to those in
[14].
We now introduce the relevant definitions and constructions to be used in this paper.
Throughout we let X be a complex Banach space and we let B(X) denote the Banach
algebra of all bounded operators on X. We let IX denote the identity operator on X.
Let (εj)j≥1 be an independent sequence of Rademacher variables on some probability space
(Ω, dP). Given any x1, . . . , xk in X, we set
k
k
(cid:13)(cid:13)(cid:13)
Xj=1
εj ⊗ xj(cid:13)(cid:13)(cid:13)R,X
k
Xj=1
= ZΩ(cid:13)(cid:13)(cid:13)
εj(u)xj(cid:13)(cid:13)(cid:13)X
dP(u).
(cid:13)(cid:13)(cid:13)
Xj=1
εj ⊗ Tj(xj)(cid:13)(cid:13)(cid:13)R,X
≤ K (cid:13)(cid:13)(cid:13)
Xj=1
εj ⊗ xj(cid:13)(cid:13)(cid:13)R,X
.
k
Then we say that a subset F ⊂ B(X) is R-bounded provided that there exists a constant
K ≥ 0 such that for any k ≥ 1, for any T1, . . . , Tk in F and for any x1, . . . , xk in X,
We refer the reader to e.g. [5, Chap. 8] for basic information on R-boundedness.
For any ω ∈ (0, π), we let Σω = {λ ∈ C∗ : Arg(λ) < ω}. Let A be a densely defined
closed operator A : D(A) → X, with domain D(A) ⊂ X. Let σ(A) denote the spectrum of
A and let R(λ, A) = (λIX − A)−1 denote the resolvent operator for λ /∈ σ(A). We say that
A is sectorial of type ω ∈ (0, π) if σ(A) ⊂ Σω and for any θ ∈ (ω, π), the set
(0.1)
(cid:8)λR(λ, A) : λ ∈ C \ Σθ(cid:9)
is bounded. We further say that A is sectorial of type 0 if it is sectorial of type ω for any
ω ∈ (0, π).
Note that if A is sectorial of type ω and A is invertible, then A−1 ∈ B(X) is sectorial of
type ω as well. This readily follows from the fact that for any λ /∈ Σω, we have λ−1 /∈ Σω
and
(0.2)
λR(λ, A−1) = IX − λ−1R(λ−1, A).
We recall that A is sectorial of type < π
2 if and only if −A generates a bounded analytic
semigroup. In this case, the latter is denoted by (e−tA)t≥0.
Next we say that A is R-sectorial of R-type ω ∈ (0, π) if A is sectorial of type ω and for
any θ ∈ (ω, π), the set (0.1) is R-bounded.
The following lemma is a straightforward consequence of (0.2).
Lemma 0.1. Let A be R-sectorial of R-type ω and assume that A is invertible. Then A−1
is also R-sectorial of R-type ω.
Let A be a sectorial operator of type < π
2 . We recall that by [16], A is R-sectorial of
R-type < π
(0.3)
2 if and only if the two sets
(cid:8)e−tA : t ≥ 0(cid:9)
are R-bounded.
and
(cid:8)tAe−tA : t ≥ 0(cid:9)
COUNTEREXAMPLES R-BOUNDEDNESS
3
Let T ∈ B(X). We say that T is a Ritt operator if the two sets
(0.4)
(cid:8)T n : n ≥ 0(cid:9)
and
(cid:8)nT n(IX − T ) : n ≥ 1(cid:9)
are bounded. We further say that T is R-Ritt if these two sets are R-bounded.
These notions are closely related to sectoriality. Indeed let D = {λ ∈ C :
λ < 1} be
the open unit disc. Then T is a Ritt operator if and only if σ(T ) ⊂ D ∪ {1} and IX − T is
sectorial of type < π
2 . Further in this case, T is R-Ritt if and only if IX − T is R-sectorial of
R-type < π
2 . We refer the reader to [3, 12] and the references therein for these results and
various informations on Ritt operators and their applications.
We recall from [13, Section 1.g] that a Schauder decomposition on X is a sequence D =
{Xn : n ≥ 1} of closed subspaces of X such that for any x ∈ X, there exists a unique
n=1 xn. For any n ≥ 1,
we let pn ∈ B(X) be the projection defined for x as above by pn(x) = xn. For any integer
sequence (xn)n≥1 of X such that xn ∈ Xn for any n ≥ 1 and x = P∞
N ≥ 1, consider their sum PN = PN
n=1 pn . This is a projection and the set
{PN : N ≥ 1}
(0.5)
is bounded.
We say that D is an R-Schauder decomposition if this set is actually R-bounded. Then a
Schauder basis is called R-Schauder if its associated Schauder decomposition is R-Schauder.
n=1 cn −cn+1
is finite (in which case we say that the sequence has a bounded variation). Then c has a
limit. Let ℓc denote this limit and set
Let c = (cn)n≥1 be a sequence of complex numbers. Assume that the sum P∞
var(c) = ℓc +
∞
Xn=1
cn − cn+1 .
argument, using the boundedness of {PN : N ≥ 1}. Let Mc : X → X be defined by
For any x ∈ X, the series Pn cnpn(x) converges. This follows from an Abel transformation
Mc(x) = P∞
n=1 cnpn(x), then we actually have
∞
(0.6)
Mc(x) = ℓcx +
(cN − cN +1)PN (x) ,
x ∈ X.
XN =1
This implies that
(0.7)
kMck ≤ var(c) sup
N ≥1
kPN k.
Let (an)n≥1 be a nondecreasing sequence of (0, ∞). Then we may define an operator
A : D(A) → X as follows. We let D(A) be the space of all x ∈ X such that the series
Pn anpn(x) converges and for any x ∈ D(A), we set
∞
(0.8)
A(x) =
anpn(x).
Xn=1
Such operators were first introduced in [2, 15]. It is well-known that
(0.9)
σ(A) = (cid:8)an : n ≥ 1(cid:9)
4
L. ARNOLD AND C. LE MERDY
and that A is a sectorial operator of type 0 (see [6, 10]). More precisely, for any λ ∈ C \ R+,
R(λ, A) is the operator Mc(λ) associated with the sequence c(λ) = (cid:0)(λ − an)−1(cid:1)n≥1 and for
any θ ∈ (0, π), we have
(0.10)
Kθ = sup(cid:8)λvar(c(λ)) : λ ∈ C \ Σθ(cid:9) < ∞,
see e.g. [10, Section 2]. This estimate and (0.7) show that A is sectorial of type 0.
We note for further use that by (0.9), the above operator A is invertible.
In the sequel, any sectorial operator A of this form will be called a D-multiplier.
Likewise let c = (cn)n≥1 be a nondecreasing sequence of (0, 1). Then c has a bounded
variation, which allows the definition of T = Mc ∈ B(X) given by
(0.11)
T (x) =
∞
Xn=1
cnpn(x),
x ∈ X.
It turns out that T is a Ritt operator on X. Indeed, let A be the sectorial operator (0.8)
associated with the sequence (an)n≥1 defined by an = (1 − cn)−1. Then IX − T = A−1 is
sectorial of type 0 and σ(T ) ⊂ [0, 1], which ensures that T is a Ritt operator.
In the sequel, any Ritt operator T of this form will be called a D-multiplier.
The following is well-known to specialists.
Lemma 0.2. Let D = {Xn : n ≥ 1} be an R-Schauder decomposition on X.
(a) Any sectorial operator A on X which is a D-multiplier is R-sectorial or R-type 0.
(b) Any Ritt operator T ∈ B(X) which is a D-multiplier is R-Ritt.
Proof. Let F = {PN : N ≥ 1}. Let A be given by (0.8) and let θ ∈ (0, π). We may assume
that limn an = ∞. It follows from the above discussion that for any λ ∈ C \ R+,
(cid:2)λR(λ, A)(cid:3)(x) =
∞
XN =1
λ(cid:0)c(λ)N − c(λ)N +1(cid:1) PN (x),
x ∈ X,
with c(λ) = (cid:0)(λ − an)−1(cid:1)n≥1. This implies that
(cid:8)λR(λ, A) : λ ∈ C \ Σθ(cid:9) ⊂ Kθ · acoso(F ),
where Kθ is given by (0.10) and acoso(F ) stands for the the closure of the absolute convex
hull of F in the strong operator topology of B(X). Since F is R-bounded, acoso(F ) is R-
bounded as well, see e.g. [5, Subsection 8.1.e]. Then the set (0.1) is R-bounded, which shows
(a).
Let T be given by (0.11). It follows from the above discussion that T = IX − A−1 for some
sectorial operator A on X which is a D-multiplier. By (a) and Lemma 0.1, A−1 is R-sectorial
of type < π
(cid:3)
2 . This entails that T is R-Ritt.
Our main result is the following.
Theorem 0.3. Let D be a Schauder decomposition on X and assume that D is not R-
Schauder.
COUNTEREXAMPLES R-BOUNDEDNESS
5
(a) There exists a sectorial operator A on X which is a D-multiplier, such that the set
is not R-bounded.
(cid:8)e−tA−1
: t ≥ 0(cid:9)
(b) There exists a Ritt operator T ∈ B(X) which is a D-multiplier, such that the set
is not R-bounded.
(cid:8)T n : n ≥ 0(cid:9)
Proof. We introduce QN = IX − PN for any N ≥ 1. The idea of the proof is to construct A
(resp. T ) such that each QN is close to e−tA−1 for some t ≥ 0 (resp. to T n for some n ≥ 0).
Let c = (cn)n≥1 be a complex sequence with a bounded variation and let N ≥ 1 be a fixed
integer. For any x ∈ X, QN (x) = P∞
Xn=1
(cid:0)Mc − QN(cid:1)(x) =
N
n=N +1 pn(x) hence
cnpn(x) +
∞
(cn − 1)pn(x).
Xn=N +1
On the one hand, we have
N
Xn=1
cnpn(x) = cN PN (x) +
N −1
(cn − cn+1)Pn(x).
Xn=1
On the other hand,
∞
∞
(cn − 1)pn(x) =
Xn=N +1
Xn=N +1
(cn − 1)(cid:0)Qn−1(x) − Qn(x)(cid:1)
∞
= (cN +1 − 1)QN (x) +
(cn+1 − cn)Qn(x).
Xn=N +1
Let K = supN ≥1 kPN k. If follows from these identities that
(cid:13)(cid:13)Mc − QN(cid:13)(cid:13) ≤ (1 + K)(cid:18)N −1
Xn=1
cn+1 − cn + cN + 1 − cN +1 +
∞
Xn=N +1
cn+1 − cn(cid:19).
Let (an)n≥1 be a nondecreasing sequence of (0, ∞), with limn an = ∞, and let A be the
associated sectorial operator defined by (0.8). Let t > 0 and apply the above with
cn = e−ta−1
n ,
n ≥ 1.
Then cn ∈ (0, 1) for any n ≥ 1, the sequence (cn)n≥1 is nondecreasing (hence has a bounded
variation) and Mc = e−tA−1. Further limn cn = 1. Consequently we have
cN = cN = e−ta−1
N ,
1 − cN +1 = 1 − cN +1 = 1 − e−ta−1
N+1,
N −1
Xn=1
cn+1 − cn = cN − c1 ≤ cN = e−ta−1
N ,
6
L. ARNOLD AND C. LE MERDY
∞
Xn=N +1
cn+1 − cn = 1 − cN +1 = 1 − e−ta−1
N+1,
and hence
(0.12)
We apply the above with
(cid:13)(cid:13)e−tA−1
− QN(cid:13)(cid:13) ≤ 2(1 + K)(cid:0)e−ta−1
N + (1 − e−ta−1
N+1)(cid:1).
an = (n!)3,
n ≥ 1.
Next we consider the sequence (tN )N ≥1 of positive integers given by we set
tN = N(N!)3,
N ≥ 1.
Then by (0.12), we have
(cid:13)(cid:13)e−tN A−1
− QN(cid:13)(cid:13) ≤ 2(1 + K)(cid:0)e−N + (1 − e− N
for any N ≥ 1. This estimate implies that
(N+1)3 )(cid:1) ≤ 2(1 + K)(cid:0)e−N + N −2(cid:1)
∞
Let C be the above sum. Then for any x1, . . . , xk in X, we have
XN =1(cid:13)(cid:13)e−tN A−1
− QN(cid:13)(cid:13) < ∞.
k
XN =1
εN ⊗(cid:0)e−tN A−1
(cid:13)(cid:13)(cid:13)
− QN(cid:1)(xN )(cid:13)(cid:13)(cid:13)R,X
k
k
≤
XN =1(cid:13)(cid:13)e−tN A−1
− QN(cid:13)(cid:13)kxN k
≤ C sup(cid:8)kxN k : 1 ≤ N ≤ k(cid:9)
≤ C (cid:13)(cid:13)(cid:13)
XN =1
(xN )(cid:13)(cid:13)(cid:13)R,X
.
k
k
Hence
k
(cid:13)(cid:13)(cid:13)
XN =1
εN ⊗ xN(cid:13)(cid:13)(cid:13)R,X
+ C (cid:13)(cid:13)(cid:13)
XN =1
R-bounded. The above estimate therefore shows that the set (cid:8)e−tA−1
the sequence cn = e−a−1
above argument shows that {T n : n ≥ 1} is not R-bounded, which proves (b).
εN ⊗ xN(cid:13)(cid:13)(cid:13)R,X
: t ≥ 0(cid:9) cannot be
To prove (b), we consider T = e−A−1. Then T is the Ritt operator defined by (0.11) for
n . For any N ≥ 1, tN is an integer and T tN = e−tN A−1. Hence the
(cid:3)
By assumption, the set {PN : N ≥ 1} is not R-bounded, hence {QN : N ≥ 1} is nor
εN ⊗ QN (xN )(cid:13)(cid:13)(cid:13)R,X
R-bounded. This proves (a).
≤ (cid:13)(cid:13)(cid:13)
XN =1
εN ⊗ e−tN A−1
.
Theorem 0.3 provides a converse to Lemma 0.2, as follows.
Corollary 0.4. Let D be a Schauder decomposition on X. Then D is R-Schauder if and
only if any sectorial operator on X which is a D-multiplier is R-sectorial, if and only if any
Ritt operator on X which is a D-multiplier is R-Ritt.
COUNTEREXAMPLES R-BOUNDEDNESS
7
Proof. Let D be a Schauder decomposition on X which is not R-Schauder. Let A be verifying
(a) in Theorem 0.3, and let B = A2. Then B is a sectorial operator on X which is a D-
multiplier. Assume that B is R-sectorial, with some R-type ω ∈ (0, π). By Lemma 0.1, its
inverse B−1 is R-sectorial R-type ω as well. Hence by [7, Proposition 3.4], A−1 = (B−1)
is R-sectorial of R-type ω
: t ≥ 0} is
R-bounded, a contradiction. Hence B is not R-sectorial.
2 . This implies (see (0.3)) that the set {e−tA−1
2 < π
1
2
Combining the above fact with Theorem 0.3 (b) and Lemma 0.2, we deduce both 'if and
(cid:3)
only if' results.
It follows from [4, 6] that if X has an unconditional basis and X is not isomorphic to a
Hilbert space, then X has a Schauder basis which is not R-Schauder. The above theorem
therefore applies to all these spaces.
Further the arguments in [6, Theorem 3.7 & Corollary 3.8] show that we actually have the
following.
Corollary 0.5. Let X be isomorphic to a separable Banach lattice and assume that X is
not isomorphic to a Hilbert space.
(a) There exists a bounded sectorial operator A of type 0 on X such that {e−tA : t ≥ 0}
is not R-bounded.
(b) There exists a Ritt operator T ∈ B(X) such that the set {T n : n ≥ 0} is not
R-bounded.
Remark 0.6. This final remark compares the above corollary with existing results. Let X
be isomorphic to a separable Banach lattice without being isomorphic to a Hilbert space.
(1) It follows from [14] that there exists a Ritt operator T ∈ B(X) such that T is not
R-Ritt. Recall that by definition, T is not R-Ritt if and only if one of the two sets in (0.4)
is not R-bounded. Part (b) of Corollary 0.5 strengthens [14] by providing a Ritt operator
T on X for which we know that the first of the two sets in (0.4) is not R-bounded. This
is an important step in the understanding of the class of power bounded operators T such
that {T n : n ≥ 0} is R-bounded. This class will be investigated in a future paper (in
preparation). We refer to [11] for the study of invertible operators T ∈ B(X) such that
{T n : n ∈ Z} is R-bounded.
(2) The existence of a sectorial operators A of type 0 on X such that {e−tA : t ≥ 0} is
not R-bounded follows from [14]. Part (a) of Corollary 0.5 shows that this can be achieved
with a bounded A. We refer to [1] for various results on bounded C0-semigroups (Tt)t≥0 on
Banach space such thatthe set {Tt : t ≥ 0} is/is not R-bounded.
Acknowledgements. The authors were supported by the French "Investissements d'Avenir"
program, project ISITE-BFC (contract ANR-15-IDEX-03).
[1] L. Arnold, γ-boundedness of C0-semigroups and their H ∞-functional calculi, Preprint 2018.
[2] J.-B. Baillon, P. Cl´ement, Examples of unbounded imaginary powers of operators, J. Funct. Anal. 100
(1991), no. 2, 419-434.
References
8
L. ARNOLD AND C. LE MERDY
[3] S. Blunck, Maximal regularity of discrete and continuous time evolution equations, Studia Math. 146
(2001), no. 2, 157-176.
[4] S. Fackler, The Kalton-Lancien theorem revisited: maximal regularity does not extrapolate, J. Funct.
Anal. 266 (2014), 121-138.
[5] T. Hytonen, J. van Neerven, M. Veraar, L. Weis, Analysis in Banach spaces. Vol. II., Results in Mathe-
matics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics 67, Springer, Cham,
2017. xxi+616 pp.
[6] N. Kalton and G. Lancien, A solution to the problem of Lp-maximal regularity, Math. Z. 235 (2000),
no. 3, 559-568.
[7] N. Kalton, P. Kunstmann and L. Weis, Perturbations and interpolation theorems for H ∞-calculus with
applications to differential operators, Math. Ann. 336 (2006), 747-801.
[8] P. Kunstmann and L. Weis, Maximal Lp-regularity for parabolic equations, Fourier multiplier theorems
and H ∞-functional calculus, in "Functional analytic methods for evolution equations", pp. 65-311,
Lecture Notes in Math., 1855, Springer, Berlin, 2004.
[9] F. Lancien and C. Le Merdy, On functional calculus properties of Ritt operators, Proc. Roy. Soc. Edin-
burgh Sect. A 145 (2015), no. 6, 1239-1250.
[10] G. Lancien, Counterexamples concerning sectorial operators, Arch. Math. (Basel) 71 (1998), no. 5,
388-398.
[11] C. Le Merdy, γ-bounded representations of amenable groups, Adv. Math. 224 (2010), no. 4, 1641-1671.
[12] C. Le Merdy, H ∞-functional calculus and square function estimates for Ritt operators, Rev. Mat.
Iberoam. 30 (2014), no. 4, 1149-1190.
[13] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces I, Ergebnisse der Mathematik und ihrer
Grenzgebiete, Vol. 92. Springer-Verlag, Berlin-New York, 1977. xiii+188 pp.
[14] P. Portal, Discrete time analytic semigroups and the geometry of Banach spaces, Semigroup Forum 67
(2003), no. 1, 125-144.
[15] A. Venni, A counterexample concerning imaginary powers of linear operators, in "Functional analysis
and related topics, 1991 (Kyoto)", pp. 381-387, Lecture Notes in Math., 1540, Springer, Berlin, 1993.
[16] L. Weis, Operator-valued Fourier multiplier theorems and maximal Lp-regularity, Math. Ann. 319 (2001),
no. 4, 735-758.
E-mail address: [email protected]
Laboratoire de Math´ematiques de Besanc¸on, UMR 6623, CNRS, Universit´e Bourgogne
Franche-Comt´e, 25030 Besanc¸on Cedex, FRANCE
E-mail address: [email protected]
Laboratoire de Math´ematiques de Besanc¸on, UMR 6623, CNRS, Universit´e Bourgogne
Franche-Comt´e, 25030 Besanc¸on Cedex, FRANCE
|
1802.03305 | 3 | 1802 | 2018-03-19T22:43:01 | Maps on probability measures preserving certain distances --- a survey and some new results | [
"math.FA",
"math-ph",
"math.MG",
"math-ph"
] | Borel probability measures living on metric spaces are fundamental mathematical objects. There are several meaningful distance functions that make the collection of the probability measures living on a certain space a metric space. We are interested in the description of the structure of the isometries of such metric spaces. We overview some of the recent results of the topic and we also provide some new ones concerning the Wasserstein distance. More specifically, we consider the space of all Borel probability measures on the unit sphere of a Euclidean space endowed with the Wasserstein metric $W_p$ for arbitrary $p \geq 1,$ and we show that the action of a Wasserstein isometry on the set of the Dirac measures is induced by an isometry of the underlying unit sphere. | math.FA | math | MAPS ON PROBABILITY MEASURES PRESERVING CERTAIN DISTANCES - A
SURVEY AND SOME NEW RESULTS
DÁNIEL VIROSZTEK
Dedicated to the memory of Professor Dénes Petz
ABSTRACT. Borel probability measures living on metric spaces are fundamental math-
ematical objects. There are several meaningful distance functions that make the col-
lection of the probability measures living on a certain space a metric space. We are in-
terested in the description of the structure of the isometries of such metric spaces. We
overview some of the recent results of the topic and we also provide some new ones
concerning the Wasserstein distance. More specifically, we consider the space of all
Borel probability measures on the unit sphere of a Euclidean space endowed with the
Wasserstein metric Wp for arbitrary p ≥ 1, and we show that the action of a Wasserstein
isometry on the set of the Dirac measures is induced by an isometry of the underlying
unit sphere.
8
1
0
2
r
a
M
9
1
]
.
A
F
h
t
a
m
[
3
v
5
0
3
3
0
.
2
0
8
1
:
v
i
X
r
a
1. INTRODUCTION
The study of isometries of various metric spaces has a huge literature. Some results
that describe the structure of the isometries of some highly important spaces are very
well-known. From our viewpoint, the most interesting classical result is the Banach-
Stone theorem that describes the surjective linear isometries between the function spaces
C (X ) and C (Y ), where X and Y are compact Hausdorff spaces. The Banach-Stone the-
orem says that every such isometry is the composition of an isometry induced by a
homeomorphism between the underlying spaces X and Y and a trivial isometry. (We
will make this statement precise later.) Another example of the well-know classical
results on isometries is the Mazur-Ulam theorem which states that a surjective isom-
etry between real normed spaces is necessarily affine. For a comprehensive study of
isometries, moreover, other types of preserver problems, we refer to the monographs
[4, 5, 11].
Isometries of spaces of measures (or distribution functions) have also been stud-
ied extensively. In a series of papers, Lajos Molnár (partially with Gregor Dolinar) de-
scribed the isometries of the distribution functions with respect to the Kolmogorov-
Smirnov metric and the Lévy metric (see [3, 9, 10]). As a substantial generalization of
Molnár's result on the Lévy isometries, György Pál Gehér and Tamás Titkos managed
to describe the surjective isometries of the space of all Borel probability measures on a
separable real Banach space with respect to the Lévy-Prokhorov distance [7]. Gehér also
described the surjective isometries of the probability measures on the real line with re-
spect to the Kuiper metric [6]. The Wasserstein isometries have been investigated by
2010 Mathematics Subject Classification. Primary: 46E27, 54E40.
Key words and phrases. Wasserstein isometies, unit sphere.
The author was supported by the ISTFELLOW program of the Institute of Science and Technology
Austria (project code IC1027FELL01) and partially supported by the Hungarian National Research, De-
velopment and Innovation Office – NKFIH (grant no. K124152).
1
2
DÁNIEL VIROSZTEK
Jérome Bertrand and Benoit R. Kloeckner on various spaces with the special choice of
the parameter p = 2 [8, 1].
Our goal is to study the Wasserstein isometries on probability measures defined on
unit spheres for an arbitrary parameter p ≥ 1. We make some progress in the direction
of a Banach-Stone-type result, that is, we show that the action of a Wasserstein isom-
etry on the set of the Dirac measures is induced by an isometry of the underlying unit
sphere.
2. OPTIMAL TRANSPORT
2.1. Motivation. Let us consider the following problem. There are m producers of
a certain product, say, x1,..., xm, and there are n customers which are denoted by
y1,..., yn. The producer xi offers pi unit of the product and the customer y j needs
q j unit of it. Assume that we are in the fortunate situation when the total demand co-
j =1 q j . For the sake of simplicity, we
assume that the aforementioned quantities are equal to 1. Let us denote the cost of
transferring a unit of product from xi to y j by c(i, j ). A transference plan (or transport
plan) is a declaration of the amounts of the product that are to be transferred from the
sources to the targets. Let t(i, j ) denote the amount that is to be transferred from xi
incides with the total supply, that is, Pm
i =1 pi =Pn
to y j . Then a transference plan is an array of nonnegative real numbers©t(i, j )ªm n
such thatPn
We are interested in finding the minimal cost of transferring the product from the
j =1 t(i, j ) = pi andPm
i =1 t(i, j ) = q j for all i and j.
producers to the customers. Clearly, the minimal cost is
i =1, j =1
(1)
c(i, j )t(i, j )
infXi , j
where the infimum runs over all transport plans. The quantity (1) is called the optimal
transport cost between the probability measures©piªm
i =1 and©q jªn
j =1 .
2.2. The mathematical treatment of more general optimal transport problems. The
optimal transport cost may be defined between any Borel probability measures on suf-
ficiently nice spaces. The key notion which is needed to define optimal transport cost
between general probability measures is the coupling, which is a basic concept in prob-
ability theory with a lot of applications that are different from optimal transport (see,
e.g., [14, Chapter 1]).
Definition 1 (Coupling). Let X and Y be Polish (that is, separable and complete) metric
spaces and let µ and ν be Borel probability measures on X and Y , respectively. A Borel
probability measure π on X × Y is said to be a coupling of µ and ν if the marginals of π
are µ and ν, that is, π(A × Y ) = µ(A) and π(X × B) = ν(B) for all Borel sets A ⊂ X and
B ⊂ Y .
Let us denote the set of all couplings of the probability measures µ and ν by Π¡µ, ν¢ .
Now, let c(x, y) stand for the cost of transporting one unit of mass from x ∈ X to y ∈ Y .
(In this contex, the word "mass" refers to something that is to be transferred.) The
optimal transport cost between the measures µ and ν is defined as
(2)
C¡µ, ν¢ := inf
π∈Π(µ,ν)ZX ×Y
c(x, y)dπ(x, y).
ISOMETRIES ON PROBABILITY MEASURES
3
2.3. Metric properties of the optimal transport cost. One may expect that the quan-
tity (2) serves as a distance between probability measures. In general,¡µ, ν¢ 7→ C¡µ, ν¢
is not a metric, but there are some important special cases when it is indeed a metric.
When the cost function is defined in terms of a metric appropriately, then the optimal
transport cost is (in a very simple correspondence with) a metric on measures. The
Wasserstein distances are metrics on measures that are defined as very simple func-
tions of optimal transport costs induced by special cost functions. In order to define
Wasserstein distances, first we need to define Wasserstein spaces. Here and throughout,
let P(X ) denote the set of all Borel probability measures on a metric space X .
Definition 2 (Wasserstein spaces). Let (X , d) be Polish metric space and let 1 ≤ p < ∞.
The Wasserstein space of order p is defined as
Pp (X ) :=½µ ∈ P(X )¯¯¯¯
ZX
d(x0, x)p dµ(x) < ∞ for some (hence all) x0 ∈ X¾.
In words, the Wasserstein space of order p consists of the probability distributions
that have finite moment of order p. Clearly, if the metric d is bounded on X , then we
have Pp(X ) = P(X ) for all p ∈ [1, ∞).
Now we are in the position to define the Wasserstein distances.
Definition 3 (Wasserstein distances). With the same conventions as in Definition 2,
the Wasserstein distance of order p between µ ∈ Pp(X ) and ν ∈ Pp(X ) is defined by the
formula
(3)
Wp¡µ, ν¢ :=µ
inf
π∈Π(µ,ν)ZX ×X
d(x, y)p dπ(x, y)¶
1
p
.
It can be shown that the Wasserstein distance of order p (or p-Wasserstein distance)
is a true metric on Pp(X ) (see, e.g., [14, Chapter 6], or [2] for the special case p = 1).
The Wasserstein distances encode valuable geometric information as they are defined
in terms of the underlying geometry. In particular we have Wp¡δx, δy¢ = d(x, y) for any
Polish space (X , d) and any x, y ∈ X and 1 ≤ p < ∞. (Here and throughout, δx denotes
the Dirac measure concentrated on the point x ∈ X .) Consequently, the Polish space
X can be embedded isometrically into the measure space Pp (X ) by the map x 7→ δx
for any 1 ≤ p < ∞. Moreover, any isometry of X induces a p-Wasserstein isometry on
Pp(X ) (for any p) by the push-forward of measures. This latter concept is of a particular
importance in measure theory and it will play a crucial role throughout this paper,
hence we define it in a quite general context.
Definition 4 (Push-forward). Let (X , A ) and (Y , B) be measurable spaces and let µ be
complex measure on X . Let ψ : (X , A ) → (Y , B) be a measurable map. Then the push-
forward of the measure µ by the map ψ is denoted by ψ#µ and it is defined by
ψ#µ(B) := µ¡ψ−1(B)¢
(B ∈ B).
Indeed, it is easy to see that for an isometry ψ : X → X the induced push-forward
of measures ψ# : Pp (X ) → Pp (X ) is a p-Wasserstein isometry for any p. So, we have a
natural group homomorphism from the isometry group of X into the isometry group
of Pp (X ) which looks as follows:
(4)
# : Isom X → Isom Pp(X );
ψ 7→ ψ#.
4
DÁNIEL VIROSZTEK
2.4. Some remarkable properties of the Wasserstein distance of order 1. In the se-
quel we recall two interesting properties of the distance W1 (which is also commonly
called the Kantorovich–Rubinstein distance) on particular Polish metric spaces.
Example 1. The total variation distance is a well known metric on P(X ) defined by the
formula
dT V¡µ, ν¢ = sup
B ∈BX¯¯µ(B) − ν(B)¯¯
,
where BX denotes the collection of all Borel sets of X .
Let (X , d) be a discrete metric space, that is, X is a nonempty set and d : X × X →
[0, ∞) is defined by
d(x, y) =(0, if x = y,
1, if x 6= y.
Then the total variation distance coincides with the Wasserstein distance of order 1 on
P(X ) = P1(X ) (see [2] and [13]).
Example 2. Let X = R equipped with the usual (Euclidean) metric. In this special case,
the Wasserstein distance of order 1 can be expressed explicitly in terms of the cumula-
tive distribution functions by the formula
(5)
(6)
W1(µ, ν) =ZR
F (x) −G(x) ,
where F (x) = µ((−∞, x]) and G(x) = ν((−∞, x]). This result is due to Vallender [13].
3. ISOMETRIES OF MEASURE SPACES: AN OVERVIEW OF THE LITERATURE
The study of isometries of measure spaces is an extensive topic in the area of pre-
server problems. Throughout this paper, by measure spaces we mean collections of
Borel probability measures on Polish metric spaces. Different notions of distance lead
to different geometry on measure spaces. In order to understand a geometric struc-
ture one has to face several challenges. One of the most fundamental characteristics of
a geometric structure is its isometry group, so the description of the isometries belongs
certainly to the important challenges.
In the sequel we recall some results on isometries of measure spaces. Certainly, this
enumeration of the relevant works is far from being complete. As the main result of
this note is a step in the direction of a Banach-Stone-type result, first we shall recall the
famous Banach-Stone theorem.
Theorem 5 (Banach-Stone). Let X and Y be compact, Hausdorff topological spaces and
let C (X ) and C (Y ) denote the spaces of all continuous complex-valued functions on X
and Y , respectively (equipped with the supremum norm). Let T : C (X ) → C (Y ) be a sur-
jective, linear isometry. Then there exists a homeomorphism ϕ : Y → X and a function
u ∈ C (Y ) with¯¯u(y)¯¯ = 1 for all y ∈ Y such that
(T f )(y) = u(y) f ¡ϕ(y)¢
for all y ∈ Y and f ∈ C (X ).
Seemingly, this classical result does not have any connection with measures.
Let us remark that the Banach-Stone theorem describes the structure of the sur-
jective linear isometries between unital commutative C ∗-algebras. (By the Gelfand-
Naimark theorem, any such algebra is isometrically ∗-isomorphic to C (K ) for some
compact Hausdorff space K ; the ∗ operation on C (K ) is the pointwise conjugation, that
ISOMETRIES ON PROBABILITY MEASURES
5
is, f ∗(k) = f (k) for all k ∈ K .) It states that any surjective linear isometry is necessarily
an algebra ∗-isomorphism - up to multiplication by a fixed function of modulus 1.
The Kadison theorem is a generalization of the Banach-Stone theorem for not necessar-
ily commutative unital C ∗-algebras. It says that a surjective linear isometry between
unital C ∗-algebras can be obtained as a Jordan ∗-isomorphism multiplied by a fixed
unitary element. (A Jordan ∗-isomorphism is a bijective linear map J that respects the
∗ operation and preserves the square, that is, J¡a2¢ = J(a)2 for all a.)
There are several results in the large area of preserver problems which state that "any
isometry between certain extra structures built on sets (say, function spaces, measure
spaces, etc.) is necessarily driven by some sufficiently nice transformation between
the underlying sets". Such results are called Banach-Stone-type theorems for obvious
reasons.
3.1. Banach-Stone-type results on isometries of measure spaces. In this subsection
we recall some recent Banach-Stone-type results concerning measure spaces.
3.1.1. Kolmogorov-Smirnov isometries. The first non-classical result that we recall here
is the theorem of Dolinar and Molnár on the isometries of the space of probability
distributions on the real line with respect to the Kolmogorov-Smirnov metric. The
Kolmogorov-Smirnov distance of the Borel probability measures µ, ν ∈ P(R) is defined
by the formula
(7)
dK S¡µ, ν¢ :=¯¯¯¯Fµ − Fν¯¯¯¯∞
= sup
x∈R¯¯Fµ(x) − Fν(x)¯¯
,
where Fη stands for the cumulative distribution function of the measure η for any η ∈
P(R), that is, Fη(x) = η((−∞, x]) .
Note that the Kolmogorov-Smirnov distance is closely related to the total variation
distance (introduced in Example 1) and the 1-Wasserstein distance on P1(R) (see Ex-
ample 2). It is clear by the comparison of the formulas (5) and (7) that dK S¡µ, ν¢ ≤
dT V¡µ, ν¢ always holds, and by the comparison of the formulas (6) and (7) one may
observe that the 1-Wasserstein distance is just the L1 distance of the distribution func-
tions while the Kolmogorov-Smirnov distance is the L∞ distance of them. The theorem
of Dolinar and Molnár reads as follows.
Theorem 6 ([3]). Let φ : P(R) → P(R) be a surjective Kolmogorov-Smirnov isometry, that
is, a bijection on P(R) with the property that
Then either there exists a strictly increasing bijection ψ : R → R such that
where Fη(x−) denotes the left limit of the distribution function Fη at the point x - note
that Fη(x−) = η((−∞, x)).
Moreover, any transformation of the form (8) or (9) is a surjective Kolmogorov-Smirnov
isometry.
or there exits a strictly decreasing bijection ψ : R → R such that
dK S¡φ(µ), φ(ν)¢ = dK S¡µ, ν¢
Fφ(µ)(t) = Fµ¡ψ(t)¢
Fφ(µ)(t) = 1 − Fµ¡ ψ(t)−¢
¡µ, ν ∈ P(R)¢.
¡t ∈ R, µ ∈ P(R)¢,
¡t ∈ R, µ ∈ P(R)¢,
(8)
(9)
6
DÁNIEL VIROSZTEK
Although Theorem 6 is formulated in terms of distribution functions, it can be easily
reformulated in terms of measures as follows: for any surjective Kolmogorov-Smirnov
isometry φ : P(R) → P(R) there is a homeomorphism ϕ : R → R such that
φ(µ) = ϕ#µ
¡µ ∈ P(R)¢,
where ϕ# is the push-forward induced by ϕ (see Definition 4). Indeed, if φ acts on P(R)
such that (8) holds, then ϕ = ψ−1, that is, φ =¡ψ−1¢# and if φ acts on P(R) such that (9)
holds, then then ϕ =¡ ψ¢−1 , that is, φ =³¡ ψ¢−1´#
The key idea of the result of Dolinar and Molnár is the observation that the Dirac
distributions can be characterized in terms of the Kolmogorov-Smirnov metric. To pre-
cisely state the characterization we shall introduce the following notation: for a metric
space (Y , ρ) and a set S ⊂ Y let U (S) be defined by
.
Now let us consider the special metric space (P(R), dK S). The metric characterization
of the trivial distributions reads as follows.
U (S) :=©y ∈ Y¯¯ρ¡y, s¢ = 1 for all s ∈ Sª.
(10)
A measure µ ∈ P(R) is a Dirac mass ⇐⇒ U¡U¡©µª¢¢ =©µª.
Such characterizations of Dirac measures will play a crucial role in several following
results.
3.1.2. Lévy isometries. The Lévy distance of the measures µ, ν ∈ P(R) is defined as fol-
lows:
dLE¡µ, ν¢
dLE¡µ, ν¢
Let us remark that the equivalent definition
= inf©ε > 0¯¯ µ((−∞, t − ε]) − ε ≤ ν((−∞, t]) ≤ µ((−∞, t + ε]) + ε for all t ∈ Rª.
= sup©ε > 0¯¯ ν((−∞, t]) + ε < µ((−∞, t − ε])
or ν((−∞, t]) − ε > µ((−∞, t + ε]) for some t ∈ Rª
offers another viewpoint to understand the Lévy metric. The importance of the Lévy
distance comes from the fact that (just like some other metrics) it metrizes the topology
of weak convergence in P(R). This type of convergence is of a particular importance in
probability theory.
Molnár's theorem reads as follows.
Theorem 7 ([10]). Let φ : P(R) → P(R) be a surjective Lévy isometry, that is, a bijection
on P(R) with the property that
Then there is a constant c ∈ R such that either
dLE¡φ(µ), φ(ν)¢ = dLE¡µ, ν¢
Fφ(µ)(t) = Fµ (t + c)
¡µ, ν ∈ P(R)¢ .
¡t ∈ R, µ ∈ P(R)¢
Fφ(µ)(t) = 1 − Fµ ((−t + c)−)
¡t ∈ R, µ ∈ P(R)¢
Moreover, any transformation of any of the forms (11), (12) is a surjective Lévy isome-
try on P(R).
(11)
or
(12)
holds.
ISOMETRIES ON PROBABILITY MEASURES
7
The easy part of Theorem 7 says that, similarly to the Wasserstein distances, the Lévy
metric has the property that the map ψ 7→ ψ# is a group homomorphism from Isom R
into Isom P(R) (where the latter group consists of all surjective isometries of P(R) with
respect to the Lévy metric). The difficult part of Theorem 7 says that this group homo-
morphism is in fact onto, hence a group isomorphism.
The key idea is a metric characterization of the Dirac distributions (see equation
(10)), similarly to the proof of the result in [3].
3.1.3. Kuiper isometries. The Kuiper distance of the probability measures µ, ν ∈ P(R) is
given by the formula
(13)
where I = {I ⊂ R #I > 1 and I is connected}, that is, I denotes the set of all non-degenerate
intervals of R. It is clear from the definitions that the inequality
,
dK U¡µ, ν¢ := sup
I ∈I¯¯µ(I ) − ν(I )¯¯
0 ≤ dK S¡µ, ν¢ ≤ dK U¡µ, ν¢ ≤ dT V¡µ, ν¢ ≤ 1
¡µ, ν ∈ P(R)¢
holds (compare the formula (13) to the formulas (5) and (7)).
The theorem of Gehér on the isometries of P(R) with respect to the Kuiper metric
reads as follows.
Theorem 8 ([6]). Let φ : P(R) → P(R) be a surjective Kuiper isometry, that is, a bijection
on P(R) with the property that
Then there exists a homeomorphism g : R → R such that
dK U¡φ(µ), φ(µ)¢ = dK U¡µ, ν¢
¡µ, ν ∈ P(R)¢.
Moreover, every transformation of this form is a surjective Kuiper isometry on P(R).
φ(µ) = g#µ
¡µ ∈ P(R)¢.
3.1.4. Lévy-Prokhorov isometries. As mentioned before, the Lévy distance is an impor-
tant metric on P(R) as it metrizes the weak convergence in P(R). In 1956 Prokhorov
introduced a metric which metrizes the weak convergence in P(X ) for a general Polish
metric space (X , d) [12]. Now we call this metric Lévy-Prokhorov distance although it
does not coincide with the Lévy metric in the special case X = R. The Lévy-Prokhorov
distance is defined as follows:
where
dLP¡µ, ν¢ = inf©ε > 0¯¯µ(A) ≤ ν¡Aε¢ + ε for all A ∈ BXª,
Aε = [x∈A
Bε(x) and Bε(x) =©y ∈ X¯¯ d(x, y) < εª.
Gehér and Titkos considered the problem of determining the Lévy-Prokhorov isome-
tries of P(X ) in the case when X is a separable real Banach space which is a bit less
general setting than the setting of Polish metric spaces (which is the most general pos-
sible setting) [7]. Their result reads as follows.
Theorem 9 ([7]). Let (X , ·) be a separable real Banach space and let φ : P(X ) → P(X )
be a surjective Lévy-Prokhorov isometry, that is, assume that
holds. Then there exists a surjective affine isometry ψ : X → X which induces φ, that is,
we have
dLP¡φ(µ), φ(µ)¢ = dLP¡µ, ν¢
¡µ, ν ∈ P(X )¢
(14)
φ(µ) = ψ#µ
¡µ ∈ P(X )¢ .
8
DÁNIEL VIROSZTEK
Moreover, any transformation of the form (14) is a surjective Lévy-Prokhorov isometry.
Similarly to the case of the Lévy distance, we learned that the map ψ 7→ ψ# is a group
homomorphism from Isom X into Isom P(X ) (easy), and that this homomorphism is
actually onto (difficult).
The observation (10) plays an important role in the proof of the result of Gehér and
Titkos, as well. However, the general setting of separable real Banach spaces required
the development of other involved techniques.
3.1.5. 2-Wasserstein isometries on negatively curved spaces. We have noted before that
if we consider Wasserstein distances on Pp (X ) for a Polish space X , then the push-
forward of measures by an isometry of X is always a p-Wasserstein isometry on Pp (X ),
no matter what the value of the parameter p is. (See equation (4)).
The question naturally appears: are there isometries of Pp(X ) that can not be ob-
tained this way? In other words: are there non-trivial isometries of Pp (X )? (Following
the terminology of [1], we call a p-Wasserstein isometry φ of Pp(X ) trivial if φ = ψ# for
some isometry ψ : X → X . Moreover, an isometry φ is called shape-preserving if for any
µ ∈ Pp (X ) there exists an isometry ψµ : X → X such that φ(µ) =¡ψµ¢# µ. The isometries
that are not even shape-preserving are called exotic isometries.)
The result of Bertrand and Kloeckner states that if X is a negatively curved space,
then all the 2-Wasserstein isometries of P2(X ) are trivial [1]. In other words, the mea-
sure space P2(X ) is isometrically rigid.
The precise statement reads as follows.
Theorem 10 ([1]). Let X be a negatively curved geodesically complete Hadamard space.
Let φ : P2(X ) → P2(X ) be a 2-Wasserstein isometry, that is, assume that
Then there is an isometry ψ : X → X such that
W2¡φ(µ), φ(µ)¢ = W2¡µ, ν¢
¡µ, ν ∈ P2(X )¢.
φ(µ) = ψ#µ
¡µ ∈ P2(X )¢ .
Note that with the terminology borrowed from [1] the results of [10] and [7] can be
rephrased as follows: the measure space P(R) equipped with the Lévy metric is isomet-
rically rigid, and the measure space P(X ) for a real separable Banach space X equipped
with the Lévy-Prokhorov metric is also isometrically rigid.
3.2. Non-Banach-Stone-type results on isometries of measure spaces. Quite surpris-
ingly, the probability measures on Euclidean spaces have non-trivial 2-Wasserstein
isometries, as well. Furthermore, in the special case of the real line we have also ex-
otic isometries (recall that exotic means that it does not preserve the shape of the mea-
sures).
The precise statements of Kloeckner about the isometries of P2 spaces over Euclidean
spaces read as follows.
3.2.1. The case of the real line.
Theorem 11 ([8]). The isometry group of the space P2(R) with respect to the 2-Wasserstein
metric is a semidirect product
(15)
Isom R ⋉ Isom R.
In (15) the left factor is the image of # (recall that # was introduced in (4)) and the right
factor consists of all isometries that fix pointwise the set of Dirac measures. Moreover,
ISOMETRIES ON PROBABILITY MEASURES
9
the right factor decomposes as Isom R = C2 ⋉ R, where the C2 factor (the group of order
2) is generated by a non-trivial involution that preserve shapes and the R factor is a flow
of exotic isometries.
The question naturally appears: how do the elements of the right factor of (15) look
like? That is, how does a 2-Wasserstein isometry that fixes all Dirac measures look like?
The description of the exotic isometries is beyond the scope of this paper, we refer
to the original work of Kloeckner [8]. However, the description of the non-trivial but
still shape-preserving isometries is easy; the reader will find it in the explanation of
Theorem 12, because the behavior of the shape-preserving isometries is independent
of the dimension of the underlying Euclidean space. Keep in mind that C2 = O(1).
3.2.2. The case of Rn for n ≥ 2.
Theorem 12 ([8]). For n ≥ 2, the 2-Wasserstein isometry group of P2 (Rn) is a semidirect
product
(16)
where the action of an element T ∈ Isom Rn on O(n) is the conjugacy by its linear part T .
The left factor in (16) is the image of # (see (4)) and each element of the right factor
Isom Rn ⋉ O(n)
fixes all Dirac measures and preserves shapes.
So in "higher" dimensions we do not have exotic isometries but we still have non-
trivial isometries. We need some notation to explain how the non-trivial isometries
look like.
Given a µ ∈ P2 (Rn), the center of mass of µ is denoted by cµ, that is, cµ =RRn xdµ(x).
Furtheremore, for any y ∈ Rn, the associated translation is denoted by η y, that is, η y (x) =
x + y (x ∈ Rn). Now we can describe the non-trivial isometries: for any ϕ ∈ O(n), the
map
is a 2-Wasserstein isometry which leaves the Dirac measures invariant.
µ 7→³ηcµ´#
◦ ϕ# ◦³η−1
cµ´#
(µ)
4. WASSERSTEIN ISOMETRIES ON P¡Sn−1¢
After having reviewed some recent results in the topic, now we turn to the main
problem of the current paper which is the description of the p-Wasserstein isometries
on measures defined on unit balls of Euclidean spaces.
Let n ≥ 2 be arbitrary and let us consider the separable metric space
Sn−1 :=½x ∈ Rn¯¯¯¯
x =
1
2¾,
we present soon works for all p ≥ 1, so from now on, let p ∈ [1, ∞) be arbitrary.
where . denotes the Euclidean norm. We consider the Euclidean distance d(x, y) =
on the unit sphere Sn−1. Clearly, the Euclidean distance is bounded on any
¯¯¯¯x − y¯¯¯¯
unit ball, so for all n ≥ 2 we have Pp¡Sn−1¢ = P¡Sn−1¢ for all p ≥ 1. Our arguments that
Claim 13. Set µ ∈ P¡Sn−1¢ . The followings are equivalent.
(2) There exists ν ∈ P¡Sn−1¢ such that Wp(µ, ν) = 1.
(1) µ is a Dirac measure, that is, there exists x ∈ Sn−1 such that µ = δx .
10
DÁNIEL VIROSZTEK
Proof. The implication (1) =⇒ (2) is clear by the following short argument. For any
. Therefore, if
µ = δx, then by the choice ν = δ−x we have
x, y ∈ Sn−1 and for any p ≥ 1, we have Wp¡δx, δy¢ = d(x, y) =¯¯¯¯x − y¯¯¯¯
Wp¡µ, ν¢ = Wp (δx, δ−x ) = d(x, −x) = x − (−x) = 1.
The proof of the direction (2) =⇒ (1) is a bit more complicated. We have to show that
if µ ∈ P¡Sn−1¢ is not a Dirac measure, then Wp(µ, ν) < 1 holds for all ν ∈ P¡Sn−1¢ . So,
assume that µ ∈ P¡Sn−1¢ is not a Dirac measure and let y ∈ Sn−1 be arbitrary. Then
there exists some ε > 0 such that
(17)
µ¡©x ∈ Sn−1¯¯ d(x, y) ≤ 1 −  > 0.
(Otherwise, µ would be equal to δ−y .) Let us denote by η this positive number appear-
ing on the left hand side of (17) in the sequel. The estimation
d(x, y)p dµ(x)
ZS n−1
=Z{x∈S n−1 d(x,y)≤1−ε}
d(x, y)p dµ(x) +Z{x∈S n−1 d(x,y)>1−ε}
d(x, y)p dµ(x)
≤ η(1 − ε)p + (1 − η) · 1 < 1
shows that the map
Sn−1 → [0,1];
y 7→ZS n−1
d(x, y)p dµ(x)
is strictly less than 1 everywhere. Therefore, we have
(18)
ZS n−1ZS n−1
d(x, y)p dµ(x)dν(y) < 1
for any Borel probability measure ν. The map (x, y) 7→ d(x, y)p is bounded and both
µ and ν are probability measures, hence Fubini's theorem can be applied to show that
the integral on the left hand side of (18) is equal to
ZS n−1×S n−1
d(x, y)p d¡µ × ν¢ (x, y).
The measure µ×ν is clearly a coupling of µ and ν, hence by the definition of the Wasser-
stein distance (see eq. (3)) we have
p
W
d p (x, y)dπ(x, y)
π∈Π(µ,ν)ZS n−1×S n−1
p ¡µ, ν¢ = inf
≤ZS n−1×S n−1
p ¡µ, ν¢ < 1 which means that Wp¡µ, ν¢ < 1. The measure ν ∈
d(x, y)p d¡µ × ν¢ (x, y).
(cid:3)
So, we deduced that W
p
P¡Sn−1¢ was arbitrary, hence the proof is done.
The following result may be considered as a first step in the direction of a Banach-
Stone-type result on the structure of the Wasserstein isometries of probability mea-
sures on unit spheres.
Then there exists an isometry T : Sn−1 → Sn−1 such that
Wp¡φ(µ), φ(ν)¢ = Wp¡µ, ν¢
φ(δx) = T#δx, that is, φ(δx) = δT (x)
¡µ, ν ∈ P¡Sn−1¢¢.
¡x ∈ Sn−1¢.
ISOMETRIES ON PROBABILITY MEASURES
11
Theorem 14. Let φ : P¡Sn−1¢ → P¡Sn−1¢ be a (not necessarily surjective) Wasserstein
isometry, that is, a map satisfying
is a Dirac measure. So, φ sends Dirac measures to Dirac measures. That is, there exists
a map T : Sn−1 → Sn−1 such that φ(δx) = δT (x) holds for all x ∈ Sn−1. We have to show
Proof. Let x ∈ Sn−1 be arbitrary. By Claim 13, there exists a ν ∈ P¡Sn−1¢ such that
Wp (δx, ν) = 1. By assumption, Wp¡φ(δx), φ(ν)¢ = 1. By Claim 13, this means that φ(δx)
that T is an isometry. But this is clear, because Wp¡δx, δy¢ = d(x, y). Indeed, by this
elementary fact we have
for every x, y ∈ Sn−1. The proof is done.
d(x, y) = Wp¡δx, δy¢ = Wp¡φ(δx), φ¡δy¢¢
= Wp¡δT (x), δT (y)¢ = d¡T (x), T (y)¢
(cid:3)
Final remarks. Let us emphasize that we did not assume the surjectiviy of the isome-
tries in our previous arguments.
Naturally our most concrete future plan is to discover wheter the measure spaces on
the unit balls are isometrically rigid, or we also have some non-trivial isometries (let
alone exotic isometries). We believe that the answer depends on the dimension n and
on the value of the parameter p, as well.
Acknowledgement. The author is grateful to György Pál Gehér and Tamás Titkos for
drawing his attention to some of the works listed in the Bibliography, and for useful
discussions. The author is also grateful to the anonymous referee for many valuable
comments and suggestions that helped to improve the presentation of the paper sub-
stantially.
REFERENCES
[1] J. Bertrand, and B. Kloeckner, A geometric study of Wasserstein spaces: isometric rigidity in negative
curvature, Int. Math. Res. Notices 2016(5) (2016), 1368-1386.
[2] R. L. Dobrushin, Prescribing a system of random variables by conditional distributions, Theory
Probab. Appl. 15 (1970), 458–486.
[3] G. Dolinar, and L. Molnár, Isometries of the space of distribution functions with respect to the Kol-
mogorov–Smirnov metric, J. Math. Anal. Appl. 348 (2008), 494–498.
[4] R. J. Fleming, and J. E. Jamison, Isometries on Banach Spaces: Function Spaces, Chapman &
Hall/CRC Monographs and Surveys in Pure and Applied Mathematics, 129. Chapman & Hall/CRC,
Boca Raton, FL, 2003.
[5] R. J. Fleming, and J. E. Jamison, Isometries on Banach spaces: Vector-valued Function Spaces, Chap-
man & Hall/CRC Monographs and Surveys in Pure and Applied Mathematics, 138. Chapman &
Hall/CRC, Boca Raton, FL, 2008.
[6] Gy. P. Gehér, Surjective Kuiper isometries, Houston Journal of Mathematics (2018), in press.
[7] Gy. P. Gehér, and T. Titkos, A characterisation of isometries with respect to the Lévy-Prokhorov metric,
Annali della Scuola Normale Superiore di Pisa - Classe di Scienze (2018), in press.
[8] B. Kloeckner, A geometric study of Wasserstein spaces: Euclidean spaces, Annali della Scuola Nor-
male Superiore di Pisa - Classe di Scienze IX, 2 (2010), 297-323.
[9] L. Molnár, Kolmogorov-Smirnov isometries and affine automorphisms of spaces of distribution func-
tions, Cent. Eur. J. Math. 9 (2011), 789-796.
12
DÁNIEL VIROSZTEK
[10] L. Molnár, Lévy isometries of the space of probability distribution functions, J. Math. Anal. Appl. 380
(2011), 847-852.
[11] L. Molnár, Selected Preserver Problems on Algebraic Structures of Linear Operators and on Function
Spaces, Lecture Notes in Mathematics, Vol. 1895, Springer, 2007
[12] Yu. V. Prokhorov, Convergence of random processes and limit theorems in probability theory, Theory
Probab. Appl. 1 (1956), 157–214.
[13] S. S. Vallender, Calculation of the Wasserstein distance between probability distributions on the line,
Theory Probab. Appl. 18 (1973), 784–786.
[14] C. Villani, Optimal Transport, Old and New, Springer, 2009.
INSTITUTE OF SCIENCE AND TECHNOLOGY AUSTRIA, AM CAMPUS 1, 3400 KLOSTERNEUBURG, AUS-
TRIA
E-mail address: [email protected]
URL: http://pub.ist.ac.at/dviroszt
|
1104.0770 | 1 | 1104 | 2011-04-05T08:36:51 | A weak local irregularity property in S^\nu spaces | [
"math.FA"
] | Although it has been shown that, from the prevalence point of view, the elements of the S^ \nu spaces are almost surely multifractal, we show here that they also almost surely satisfy a weak uniform irregularity property. | math.FA | math |
A weak local irregularity property in S ν spaces
Marianne Clausel∗and Samuel Nicolay†‡
November 16, 2018
Abstract
Although it has been shown that, from the prevalence point of view,
the elements of the S ν spaces are almost surely multifractal, we show here
that they also almost surely satisfy a weak uniform irregularity property.
1
Introduction
The uniform regularity defined from the Holder spaces Cα(Rd) is one of the
most popular concepts for the uniform regularity.
It has been introduced to
study smoothness properties of functions such as the Weierstrass function (see
[15]). Indeed, many "historical" functions share the same property (see [20]):
there exist H ∈ (0, 1) and a constant C > 0 such that the function f satisfies
on some interval I,
∀x, y ∈ I, f (x) − f (y) ≤ Cx − yH ,
and
∀x, y ∈ I,
sup
f (x) − f (y) ≥
(u,v)∈[x,y]2
1
C
x − yH .
(1)
(2)
It has been shown in [8] that this behavior is the typical behavior of the functions
belonging to CH (Rd), in the sense of prevalence (the notion of prevalence will
be defined in the sequel, see section 2.3). In other words, almost every function
of CH (Rd) satisfies Relations (1) and (2).
However, in many cases, the functions do not satisfy Relations (1) and (2). It
is in particular the case of the so-called multifractal functions (see Definition 5),
originally introduced in the context of turbulence and now used in many fields
of science (see e.g. [1, 2, 14, 17, 26]). It can be shown that in several functional
spaces, almost every function (in the sense of prevalence) is multifractal (see
[13, 4]).
In this paper, we aim at investigating the typical irregularity properties of
the elements of the Sν spaces. These functional spaces, defined in [18, 3], give
rise to an efficient multifractal formalism (see section 2), i.e. an heuristic method
∗Universit´e de Lyon, CNRS, INSA de Lyon, Institut Camille Jordan UMR 5208, Batiment
L. de Vinci, 20 av. Albert Einstein, F-69621 Villeurbanne Cedex, France.
†Universit´e de Li`ege, Institut de Math´ematique, Grande Traverse, 12, Batiment B37, B-
4000 Li`ege (Sart-Tilman), Belgium.
‡Corresponding author. Email: [email protected]. Phone: +32(0)43669433. Fax:
+32(0)43669547.
1
to study the pointwise regularity of a function from a global point of view. It
has been shown in [4] that almost every element of Sν is multifractal. We
show here that, although they are multifractal, the elements of Sν also satisfy
almost surely a weak uniform irregularity property. To this end, we introduce a
concept of weak uniform irregularity, define the irregularity exponent H(x0) of
a function at a given point x0 and prove that, in the sense of prevalence, there
exists α (depending on Sν) such that almost every function of Sν satisfy
H(x) = α,
for any x.
The paper is organized as follows. In the next section, we recall some defi-
nitions about the multifractal analysis and the Sν spaces. Next, we define the
local irregularity exponent. To prove our main result, we will need wavelet cri-
teria for the irregularity; these are stated in Section 3. The prevalent result is
obtained in the last section.
2 Multifractal models
The multifractal analysis aims to study the smoothness of very irregular signals.
For such a function, it is meaningless to try to characterize its regularity at a
given point, since the pointwise regularity can abruptly change. One rather
tries to determine the so-called spectrum of singularities, which gives the size
of the set of points that share the same pointwise regularity. We first have
to precise what is meant by "pointwise regularity" and "size of set". The Sν
spaces have been introduced in order to provide an efficient multifractal formal-
ism, i.e. a method that allows the computation of the multifractal spectrum in
many practical cases. Although this technique does not always lead to the right
spectrum, it has been shown that the Sν-based multifractal formalism gives the
right answer for almost every element of Sν. Here the term "almost every" has
to be clearly stated, since one can not use the usual Lebesgue measure for the
infinite dimensional settings. Finally, we introduce another index of regularity,
called the local irregularity exponent.
2.1 Pointwise regularity, Hausdorff dimension and multi-
fractal spectrum
The notion of the pointwise regularity which we will use here is based on the
characterization of the Holder spaces Cα(x0).
Definition 1 Let x ∈ Rd and α > 0. A locally bounded function f : Rd → C
belongs to Cα(x0) if there exist C > 0 and a polynomial P of degree strictly
lower than α such that, in a neighborhood of x0,
f (x) − P (x − x0) ≤ Cx − x0α.
The Holder exponent of f at x0 is the quantity
H(x0) = sup{α > 0 : f ∈ Cα(x0)}.
2
The multifractal analysis provides a description of the collection of the
Holder exponents of a function by way of the multifractal spectrum (also called
the singularity spectrum), which associates to a value h the size of the set of
points for which the Holder exponent is h. By size, one usually means Hausdorff
dimension; this notion of dimension is usually preferred because it relies on a
measure. We will only give the necessary definitions; the interested reader is
referred to e.g. [12, 23].
Definition 2 Let S be a Borelian subset of Rd and Γǫ(S) be the collection of all
the countable ǫ-coverings of S, i.e. the collection of all the countable coverings of
S by sets whose diameter is lower than ǫ. The δ-dimensional Hausdorff measure
of S is
mδ(S) = lim
ǫ→0
inf
(Sj )j ∈Γǫ(S)Xj
Sjδ,
where Sj denotes the diameter of the set Sj.
It is easy to check that if mδ0 (S) is finite, then mδ(S) = ∞ if δ < δ0 and
mδ(S) = 0 if δ > δ0. We are then naturally led to the following definition.
Definition 3 The Hausdorff dimension dim S of a non-empty Borelian subset
S of Rd is given by
dim S = inf{δ : mδ(S) = 0}.
One sets dim ∅ = −∞.
If S is empty or uncountable, one has dim S = sup{δ : mδ(S) = ∞}. We are
now able to introduce the multifractal spectrum of a function.
Definition 4 The multifractal spectrum of a locally bounded function f : Rd →
C is the application
d : (0, ∞] → [−∞, d] h 7→ dim{x ∈ Rd : H(x) = h}.
Finally, let us introduce some usual denominations.
Definition 5 A function is called multifractal if the associated multifractal
spectrum takes more than one real value on a support with non-zero Hausdorff
dimension. A function that is not multifractal is a monofractal function. If the
spectrum only takes one value, the function is said to be monoHolder.
2.2 Bases of wavelets and S ν spaces
One of the remarkable aspects of the wavelets is that they provide bases of
the space L2(Rd). Historically, the first orthonormal wavelet basis is the Haar
basis, constructed long before the introduction of the term "wavelet". In [19],
it is shown that, under some suitable conditions, the multifractal spectrum of
a function can be estimated from its wavelet coefficients. The Sν spaces were
introduced in [18] to improve the classical wavelet-based multifractal formalism.
The Sν spaces are closely related to the Besov spaces.
Let us briefly recall some definitions and notations (for more precisions, see
e.g. [24, 10, 21]). Under some general assumptions, there exist a function φ and
2d − 1 functions (ψ(i))1≤i<2d called wavelets such that
{φ(x − k) : k ∈ Zd} ∪ {ψ(i)(2j − k) : 1 ≤ i < 2d, k ∈ Zd, j ∈ N}
3
form a basis of L2(Rd). A function f ∈ L2(Rd) can be decomposed as follows,
f = Xk∈Zd
Ckφ(· − k) +
∞
2d−1
Xj=1 Xk∈Zd
Xi=1
c(i)
j,kψ(i)(2j · −k),
where
and
Ck = ZRd
f (x)φ(x − k) dx
j,k = 2dj ZRd
c(i)
f (x)ψ(i)(2jx − k) dx.
(3)
Let us remark that we do not choose the (usual) L2 normalization for the
wavelets, but rather an L∞ normalization, which is better fitted to the study
of the Holderian regularity. Expressions such as (3) can make sense in more
general settings (e.g. if f is a distribution). Hereafter, we will assume that the
wavelets belong to Cγ(Rd) with γ ≥ α + 1, and that the functions {∂sφ}s≤γ,
{∂sψ}s≤γ have fast decay. Moreover, for the sake of simplicity, when dealing
with the Sν spaces, we will suppose that the application f is defined on the
torus Td = Rd/Zd. Let
Λ = (cid:8)(i, j, k) : 1 ≤ i < 2d, j ∈ N, k ∈ {0, . . . , 2j − 1}d(cid:9) .
If (i, j, k) ∈ Λ, the periodized
ψ(i)
j,k = Xl∈Zd
ψ(cid:0)2j(· − l) − k(cid:1)
form a basis of the one-periodic functions in L2([0, 1]d). We will denote (c(i)
or (cλ)λ∈Λ the wavelet coefficients of a function in L2([0, 1]d).
j,k)(i,j,k)∈Λ
Let us now introduce the Sν spaces.
Definition 6 For a sequence c = (cλ)λ∈Λ, C > 0 and α ∈ R, we define
Ej(C, α)[c] = {(i, k) : c(i)
j,k ≥ C2−jα}.
The wavelet profile νc of c is defined as
νc(α) = lim
ǫ→0+
lim sup
j→∞
log2 #Ej (1, α + ǫ)[c]
j
,
with α ∈ R. If c represents the wavelet coefficients of a function f , one sets
Ej(C, α)[f ] = Ej (C, α)[c] and νf = νc.
Clearly, νc is non-decreasing, right-continuous and non-negative (lower than d)
when not equal to −∞. It gives, in some way, the asymptotic behavior of the
number of coefficients of c that have a given order of magnitude.
Definition 7 Let ν : R → {−∞} ∪ [0, d] be a non-decreasing, right-continuous
function such that there exists α0 for which ν(α) = −∞ if α < α0 and ν(α) ∈
[0, d] otherwise. A sequence distribution f belongs to Sν if its wavelet coefficients
satisfy νf (α) ≤ ν(α) for any α ∈ R.
4
These spaces are robust, in the sense that their definition does not depend on
the choice of the wavelet basis (see [18]). Roughly speaking, a function belongs
to Sν if for each scale j and every α ∈ R, there are about 2ν(α)j coefficients c(i)
j,k
that are larger than 2−αj. The following definition is equivalent to the preceding
one (see [3]).
Definition 8 If ν is a function as given in Definition 7, f belongs to Sν if for
any α ∈ R, ǫ > 0 and C > 0, there exists J such that j ≥ J implies
#Ej (C, α)[f ] ≤ 2(ν(α)+ǫ)j.
A topological framework for the spaces Sν will also be needed (further details
can be found in [3, 4]). The ancillary spaces are useful to obtain a structure of
complete metric space on Sν.
Definition 9 Let m, n ∈ N; f belongs to Em,n if there exist C > 0 such that
#Ej (C, αm)[f ] ≤ C2ν(αm)+ǫn)j,
(4)
for any j, where (αm)m is any dense sequence in R and (ǫn)n is decreasing to
zero.
If for f ∈ Em,n, the infimum of the constants C satisfying the inequality (4) is
noted dm,n(f, 0), the distance dm,n(f, g) = dm,n(f − g, 0) makes (Em,n, dm,n) a
metric space. It can be shown that Sν = ∩m,nEm,n and that with the distance
d(f, g) = Xm,n≥0
2−m−n dm,n(f, g)
1 + dm,n(f, g)
,
the space (Sν, d) is a complete separable metric space. The distance d may
depend on the sequences chosen in (4), but the induced topology does not. The
Borel σ-algebra relative to this topology will be denoted B(Sν).
2.3 Prevalence of multifractal functions in S ν
In a finite dimensional space, a property "holds almost everywhere" if the set
of points for which this property is not satisfied vanishes for the Lebesgue mea-
sure. The Lebesgue measure has a preponderant role, as there is no other
σ-finite translation invariant measure. Unfortunately, there is such measure in
the infinite dimensional Banach spaces. The notion of prevalence provides the
analogue of "Lebesgue measure zero" in complete metric vector spaces.
In [5], to recover the notion of "almost every" in infinite vector spaces, a well-
known characterization of Lebesgue measure zero subsets of Rd is generalized:
In Rd, a Borel set B has Lebesgue measure zero if and only if there exists a
compactly supported probability measure µ such that µ(B + x) = 0 for any
x ∈ Rd. This characterization can be turned into a definition in more general
settings and leads to the concept of Haar-null set, which provides the analogue
of "Lebesgue measure zero" set for infinite dimensional spaces.
Definition 10 Let E be a complete metric vector space. A Borel set B ⊂ E is
Haar-null if there exists a Borel probability measure, strictly positive on some
compact set K ⊂ E such that µ(B + x) = 0 for any x ∈ E. A subset of E is
Haar-null if it is included in a Haar-null Borel subset of E. The complement of
a Haar-null set is called a prevalent set.
5
In [4], the prevalent behavior of almost every function of Sν has been studied
and it has been proved that almost every function of Sν is multifractal. More
precisely, an upper bound of the spectrum d of any f ∈ Sν is first given:
if
αmax = inf h≥α0 {h/ν(h)} and
dν (h) = (cid:26) h suph′∈(0,h]{ν(h′)/h′}
1
if h ≤ αmax
otherwise
,
one has, for all f ∈ Sν and h ∈ R, d(h) ≤ dν(h). Then it is shown that the sets
{f ∈ Sν : d(h) = dν (h) if h ≤ αmax and d(h) = −∞ otherwise}
and
{f ∈ Sν : the pointwise regularity is almost everywhere αmax}
are prevalent. The "typical elements" (in the sense of the prevalence) of Sν are
thus multifractal and do not satisfy the classical law of the iterated logarithm
almost everywhere.
2.4 Local regularity exponents
We introduce here another notion of local regularity, which will be used to
state that a weak irregularity condition is satisfied for almost every multifractal
function of Sν. To this end, we first need to recall different concepts of global
regularity, based on the global Holder regularity.
If Ω is an open subset of Rd, for any h ∈ Rd, we will denote by Ωh the set
Ωh = {x ∈ Rd : [x, ([α] + 1)h] ⊂ Ω},
where the value α > 0 will be implied by the context and [α] denotes the greatest
integer lower than α. We will need the classical notion of finite difference.
Definition 11 Let x, h ∈ Rd and f : Rd → C; the first order difference of f is
∆1
hf (x) = f (x + h) − f (x). For n ≥ 2, the difference of order n is defined by
∆n
hf (x) = ∆n−1
h ∆1
hf (x).
We can now introduce the Holder spaces Cα(Ω).
Definition 12 Let Ω be an open subset of Rd and α > 0; a bounded function
f defined on a subset of Rd belongs to Cα(Ω) if there exist C, r0 > 0 such that
for any r ≤ r0,
k∆[α]+1
h
f (x)kL∞(Ωh) ≤ Crα.
sup
h≤r
The function f is said to be uniformly Holderian on Ω if for some α > 0, f
belongs to Cα(Ω).
We will also need a notion of uniform irregularity.
Definition 13 Let Ω be an open subset of Rd, α > 0 and β ∈ R; a bounded
function f defined on a subset of Rd belongs to I α(Ω) if there exist C, r0 > 0
such that for any r ≤ r0,
k∆[α]+1
h
f (x)kL∞(Ωh) ≥ Crα.
sup
h≤r
A function that belongs to I α(Ω) is said to be uniformly irregular with exponent
α.
6
Let us remark that the statement f ∈ I α(Ω) is not equivalent to f /∈ Cα(Ω). In
the latter case, for any C > 0, there exists a sequence (rn)n (depending on C)
decreasing to zero for which
k∆[α]+1
h
f kL∞(Ωh) ≥ Crα
n .
sup
h≤rn
We are thus naturally led to the following definition.
Definition 14 Let Ω be an open subset of Rd, α > 0 and β ∈ R; a bounded
function belong to Cα
w(Ω) if f is defined on Ω and if for any C > 0, there exists
a decreasing sequence (rn)n decreasing to zero such that
k∆[α]+1
h
f kL∞(Ω) ≤ Crα
n ,
sup
h≤rn
for any n ∈ N. The set Cα
belong to Cα
β,w(Ω) is simply denoted Cα
w(Ω). A function that
w(Ω) is said to be weakly uniformly Holder with exponent α on Ω.
Definition 15 The upper uniform Holder exponent of a weakly uniformly Holder
function f on an open set Ω is defined as
H(Ω) = sup{α > 0 : f ∈ Cα
w(Ω)}.
We now define the local irregular exponent, using the same approach as in
[25].
Definition 16 A sequence (Ωn)n of open subsets of Rd is decreasing to x0 ∈ Rd
if
• m < n implies Ωn ⊂ Ωm,
• Ωn → 0 as n → ∞,
• ∩nΩn = {x0}.
The following lemma is needed.
Lemma 1 If (Ωn)n and (Ω′
x0, then
n)n are two sequences of open sets that decrease to
sup
n∈N
{H(Ωn)} = sup
n∈N
{H(Ω′
n)}.
Proof. Let us suppose that supn∈N{H(Ωn)} > supn∈N{H(Ω′
an index n1 such that H(Ωn1 ) > supn∈N{H(Ω′
B(x0, r) ⊂ Ωn1; since (Ω′
that Ω′
n)}. There exists
n)}. Now let r > 0 be such that
n)n is decreasing to x0, there exists an index n2 such
n2 ⊂ B(x0, r). One thus have
H(Ω′
n2 ) ≥ H(Ωn1 ) > sup
n∈N
{H(Ω′
n)},
which leads to a contradiction.
Definition 17 If f is a bounded function, the local irregularity exponent of f
at x0 is
where (Ωn)n is a sequence of open sets decreasing to x0.
H(x0) = sup
n
{H(Ωn)},
7
3 Wavelet criteria for the uniform irregularity
and characterization of the local irregularity
exponent
In this section, we first give necessary and sufficient conditions for a function to
belong to I α(Ω). Next we characterize the local irregularity exponent in terms
of wavelet coefficients.
3.1 Wavelet criteria for the uniform irregularity
In what follows, we will assume that the multiresolution analysis is compactly
supported (see [9]). The following result is shown in [19]: in R, if the wavelet
basis belongs to CM (R), there exist a fast decaying function ΨM such that ψ =
∆M
1/2ΨM . Furthermore, the function ΨM can be picked compactly supported
with support included in this of ψ. In Rd, we will use the tensor product wavelet
basis (see [24, 10]),
ψ(i)(x) = Ψ(1)(x1) · · · Ψ(d)(xd),
where Ψ(i) (i ∈ {1, . . . , d}) are either ψ or φ, but at least one of them must equal
ψ. We will also use the following notations: given j ∈ N, Ω an open subset of
Rd and a family of wavelets ψ(i)
j,k, we set
and
Ij = {(i, k) : supp(ψ(i)
j,k) ⊂ Ω}
kc(·)
j,·kℓ∞(Ω) = sup
Ij
c(i)
j,k.
The uniform regularity of a function is related to the decay rate of its wavelet
coefficients (see [24]). Let f be a bounded function and α ∈ (0, 1); f belongs
to Cα(Ω) (where Cα(Ω) denotes the homogeneous version of the Holder space
Cα(Ω)) if and only if there exists C > 0 such that for any (j, k) such that for
any j ≥ 0,
kc(·)
j,·kℓ∞(Ω) ≤ C2−αj.
(5)
The following result gives a sufficient condition to belong to I α(Ω).
Theorem 1 Let α > 0, f ∈ Cα(Ω) and set M = [α] + 1. If there exists C > 0
and γ > 1 such that, for any j ≥ 0,
max{
sup
j≤l≤j+log2 j
then f ∈ I α(Ω).
kc(·)
l,· kℓ∞(Ω), 2−jM sup
j−log2 j≤l≤j
(2lM kc(·)
l,· kℓ∞(Ω))} ≥ C2−jαjγ,
(6)
Now, if f belongs to I α(Ω), there exist C > 0 and β ∈ (0, 1) such that for
any integer j ≥ 0,
max{
sup
j≤l≤j+log2 j
kc(·)
l,· kℓ∞(Ω), 2−jM jβ
sup
j−log2 j≤l≤j
(2lM kc(·)
l,· kℓ∞(Ω))} ≥ C2−jα.
(7)
8
Proof. To prove the first part of the theorem, let us suppose that f ∈ Cα
w(Rd)
and let C > 0. As shown in [8], there exists some increasing sequence of integers
(jn)n such that for any n ∈ N and any j ≥ jn,
sup
h≤2−j
k∆M
h f kL∞(Ωh) ≤ C2−jnα.
(8)
Let us show that this inequality leads to a contradiction.
For the sake of simplicity, let us suppose that Ψ(1) = ψ (let us recall that
one of the Ψ(i) is ψ). We then have, by definition of the wavelet coefficients, for
any j ≥ 0 and any (i, k) ∈ Ij ,
j,k = 2jdZRd
c(i)
= 2jdZRd
= 2jdZRd
f (x)Ψ(1)(2jx1 − k1) · · · Ψ(d)(2jxd − kd) dx
f (x)∆M
1/2ΨM (2jx1 − k1) · · · Ψ(d)(2jxd − kd) dx
∆M
1/2j+1e1 f (x)ΨM (2jx1 − k1) · · · Ψ(d)(2jxd − kd) dx,
where e1 = (1, 0, . . . , 0). Using the assumptions on the support of ΨM and the
definition of Ij, one has, for any n ∈ N, any j ≥ jn and any (i, k) ∈ Ij,
j,k ≤ 2jdZΩ
c(i)
∆M
1/2j+1e1 f (x)ΨM (2jx1 − k1) · · · Ψ(d)(2jxd − kd) dx
≤ C2jd2−jnαZRd
= C2−jnαkΨM ⊗ · · · ⊗ Ψ(d)kL1(Rd),
ΨM (2jx1 − k1) · · · Ψ(d)(2jxd − kd) dx
by using Relation (8).
For n ∈ N, let ln = jn + γ[log2 jn]; for n sufficiently large, one has ln −
log2 ln ≥ jn. Therefore, the following relations hold for n sufficiently large (and
any (i, k) ∈ Ij ),
sup
ln≤j≤ln+log2 ln
c(i)
j,k ≤ C2−jnα ≤ C2−lnαlγ ′
n
and
sup
ln−log2 ln≤j≤ln
2jM c(i)
j,k ≤ C2lnM 2−jnα ≤ C2lnM 2−lnαlγ ′
n ,
which is in contradiction with the relation (6).
To prove the second part of the theorem we will use the following result (see
∞,∞(Ω) ∩ Cγ(Ω), the wavelet characteri-
[24]): Let f ∈ Cγ(Ω); since Cγ(Ω) ⊂ B0
zations of these two functional spaces lead to the existence of a constant C > 0
that does not depend on the function such that for any h ∈ Rd and any x ∈ Ωh,
one has
∆[γ]+1
h
f (x) ≤ C sup
j∈N
kc(·)
j,· kl∞(Ω) ,
and
∆[γ]+1
h
f (x) ≤ Chγ sup
j∈N
{2jγkc(·)
j,·kl∞(Ω)}.
9
(9)
(10)
Let us assume f ∈ I α(Ω) and that Property (7) is not satisfied. In this case,
for any C > 0 and any β ∈ (0, 1), there exists an increasing sequence of integers
(jn)n such that, for any n ∈ N,
max{
sup
jn≤l≤jn+log2 jn
2−jnM jβ
kc(·)
l,· kℓ∞(Ω),
sup
jn−log2 jn≤l≤jn
(2lM kc(·)
l,· kℓ∞(Ω))}
≤ C2−jnα.
(11)
Let us fix C > 0, x ∈ Ωh and let n0 ∈ N, h ∈ Rd be such that h ≤ 2−jn0 . By
definition of I α(Ω), we have to show that f ∈ Cα
w(Ω). We will use the following
notations:
f−1 = Xk∈Zd
Ckφ(· − k),
fj =
2d−1
Xi=1 Xk∈Zd
c(i)
j,kψ(2j · −k),
with j ≥ 0. Since f is uniformly Holder, fj and Pj≥−1 fj converge uniformly
on any compact set and
∆M
h f = Xj≥−1
∆M
h fj.
We first consider the function g1 = Pjn0
j=−1 fj. Let us fix γ ∈ ([α] + 1 −
β, [α] + 1). The regularity of the wavelets and property (10) imply the existence
of C > 0 not depending on n0, x and h such that
∆M
h g1(x) ≤ Chγ sup
l≤jn0 (cid:16)2lγkc(·)
l,· kℓ∞(Ω)(cid:17) .
(12)
Since f ∈ Cα(Ω), there exists a constant such that, for any n ∈ N,
sup
l≤jn0 −log2 jn0
(2lγkc(·)
l,· kℓ∞(Ω)) ≤ C2(jn0 −log2 jn0 )(γ−α) = C′ 2jn0 (γ−α)
jγ−α
n0
.
On the other hand, we have, using relation (11),
sup
jn0 −log2 jn0 ≤l≤jn0
(2lγkc(·)
l,· kℓ∞(Ω))
=
sup
jn0 −log2 jn0 ≤l≤jn0
(2l(γ−M)2lM kc(·)
l,· kℓ∞(Ω))
≤ C2(jn0 −log2 jn0 )(γ−M)2jn0 (M−α)/jβ
n0
= C′jM−γ
≤ C′2jn0 (γ−α).
2jn0 (γ−α)/jβ
n0
n0
This implies that for n0 sufficiently large (since 0 < M − γ ≤ β),
(2lγkc(·)
l,· kℓ∞(Ω)) ≤ C2jn0 (γ−α),
sup
l≤jn0
and hence, using inequality (12),
∆M
h g1(x) ≤ C2−jn0 α,
(13)
10
since h has been chosen adequately.
Let us now consider g2 = Pj>jn0
gives the following relation,
fj. Property (9) applied to g2 directly
∆M
h g2 ≤ C sup
l>jn0
(kc(·)
l,· kl∞(Ω)).
(14)
Once again, since f ∈ Cα(Ω), we have
sup
l≥jn0 +log2 jn0
l,· kℓ∞(Ω) ≤ C2−(jn0 +log2 jn0 )α = C′ 2−jn0 α
kc(·)
jα
n0
.
Now, Relation (11) implies
sup
jn0 ≤l≤jn0 +log2 jn0
kc(·)
l,· kℓ∞(Ω) ≤ C2−jn0 α.
These last inequalities lead to the following relation for n0 sufficiently large,
∆M
h g2(x) ≤ C2−jn0 α,
(15)
thanks to relation (14).
Putting relations (13) and (15) together, we obtain
∆M
h f (x) = ∆M
h (g1 + g2)(x) ≤ C2−jn0 α
for n0 sufficiently large, that is f ∈ Cα
w(Ω), which is impossible, since f ∈ I α(Ω).
3.2 Characterization of the local irregularity exponent
The preceding result leads to the following characterization of the irregularity
exponent.
Corollary 1 Let α > 0; if f ∈ Cα(Ω), then the irregularity exponent of f on
Ω equals α if and only if
log2 max{
sup
j≤l≤j+log2 j
kc(·)
l,· kℓ∞(Ω), 2−jM
sup
j−log2 j≤l≤j
(2lM kc(·)
l,· kℓ∞(Ω))}
−j
lim
j→∞
= α,
where M = [α] + 1.
Proof. Theorem 1 directly yields that if f ∈ Cα(Ω), the irregularity exponent
of f on Ω equals α if and only if
log2 max{
sup
j≤l≤j+log2 j
lim sup
j→∞
kc(·)
l,· kℓ∞(Ω), 2−jM sup
j−log2 j≤l≤j
−j
(2lM kc(·)
l,· kℓ∞(Ω))}
= α.
We have to prove that the lim sup can be replaced by a limit. Since f ∈ Cα(Ω),
the relation (5) implies that
log2 max{
sup
j≤l≤j+log2 j
lim inf
j→∞
kc(·)
l,· kℓ∞(Ω), 2−jM sup
j−log2 j≤l≤j
−j
11
(2lM kc(·)
l,· kℓ∞(Ω))}
= α.
is always larger than
lim inf
j→∞
log2 max{supj≤l≤j+log2 j 2−lα, 2−jM supj−log2 j≤l≤j 2l(M−α)}
−j
= α,
which is sufficient to conclude.
We can now state the local version of the previous result.
Theorem 2 Let α > 0; if f ∈ Cα(Ω) and x0 ∈ Rd, then
H(x0) = α
if and only if
log2 max{
sup
j≤l≤j+log2 j
kc(·)
l,· kℓ∞(B(x0,r)),
lim
r→0
lim
j→∞
where M = [α] + 1.
2−jM sup
j−log2 j≤l≤j
−j
(2lM kc(·)
l,· kℓ∞(B(x0,r)))}
= α,
4 A prevalent result on the Sν spaces
Prevalence supplies a natural definition of "almost every" which is translation
invariant and where no specific measure plays a particular role. To prove our
prevalent result, we first need to introduce the method we will use. We will also
need some properties obtained in [4].
4.1 The stochastic process technique
Our result concerning the prevalence relies on the stochastic process technique.
Let us recall that a random element X on a complete metric space E is a
measurable mapping X defined on a probability space (Ω, A, P ) with values in
E. For a random element on E, one can define a probability on E by the formula
PX (A) = P {X ∈ A}.
Replacing the measure µ in Definition 10 of a Haar-null set with µ = PX , we
see that in order to prove that a set is Haar-null, it is sufficient to check that
for any f ∈ E,
PX (A + f ) = 0.
The stochastic process that we will use here is a random wavelet series
associated in a proper to ν. To this end, for each j ≥ 0, let us define as in [4],
Fj (α) = (cid:26) 0
2−jd sup{j2, 2jν(α)}
if α < αmin ,
if α ≥ αmin .
(16)
Since Fj is non-decreasing and piecewise continuous, it is the repartition function
of some probability distribution associated to a probability law ρj supported on
[αmin, ∞], where
αmin = inf{α : ν(α) ≥ 0}.
12
The following remark is made in [4]. If ρj is the probability distribution whose
repartition function Fj is defined by (16), then there exists some sequence of
random numbers (cλ)λ∈Λ with independent phase and moduli such that for
any j, ρj is the common law of − log2 c(i)
j,k/j and satisfies the two following
conditions:
2jdρj((−∞, α])
= ν(α),
∀α ∈ R,
lim
j→∞
j
and
∀α ≥ αmin,
2jdρj((−∞, α]) ≥ j2.
(17)
(18)
Starting from these results, we will use the following random wavelet series
associated to ν,
Xν = Xλ∈Λ
cλψλ.
(19)
It is shown in [3] that the metric topology d(·, ·) on Sν makes it a Polish space,
which is a very good framework for prevalence. Moreover, as proved in [4], the
measure PX is a Borel measure (relatively to this topology). In other words, we
can use the stochastic process technique with Xν.
4.2 Prevalent irregularity properties in S ν
We are now ready to prove the following result.
Theorem 3 The following set is prevalent in Sν,
{f : ∀x ∈ Td, the local irregularity exponent of f at x is H(x) = αmin},
where
αmin = inf{α : ν(α) ≥ 0}.
From this point of view, the "typical elements" of Sν are multifractal and satisfy
a weak uniform irregularity property.
From what precede, it is sufficient to show the following result.
Proposition 1 Let Xν be the random wavelet series defined by (19). Then, for
any f ∈ Sν, the local irregularity exponent H(x) of f + Xν at x is equal to αmin
almost surely.
Proof. Since Sν ⊂ Cαmin(Td) then for any f ∈ Sν, f + Xν ∈ Cαmin (Td) almost
surely. Let us now fix m ∈ N and define, for any r ∈ Zd,
Tr,m =
d
(
Yn=1
rn
2m ,
rn + 1
2m )
so that Td = ∪rTr,m. We aim at showing that the equality
H(Tr,m) = αmin
holds almost surely for any r, m; in this case, Theorem 2 directly yields the
required result.
For λ ∈ Λ, we will denote as usual cλ the wavelet coefficients associated to
Xν and dλ the wavelet coefficients associated to f ∈ Sν. Let us first remark
13
that if for some fixed λ ∈ Λ, ℜ(cλdλ) ≤ 0 i.e. cλ is in the complex half-plane
opposite to dλ, then
cλ − dλ ≥ cλ
Therefore, for any λ ∈ Λ,
P (f : cλ − dλ ≥ cλ) ≥ P (f : ℜ(cλdλ) ≤ 0) ≥ 1/2.
(20)
If we define for any N ∈ N and λ = (i, j, k) ∈ Λ,
AN,λ = {f : (∃k′ ∈ k + [0, N ]d : c(i)
j,k′ − d(i)
j,k′ ≥ c(i)
j,k′ )},
then, thanks to the independence of the wavelet coefficients of Xν and inequal-
ity (20), we have
P (AN,λ) ≥ 1 − 2−N d
for any N and λ.
Now, for N ∈ N, let us set
Br,m,n = {f : kc(·)
j,·kℓ∞(Tr,m) ≥ 2−j(αmin+1/n)}.
If for any i, we have supp(ψ(i)) ⊂ [0, M ]d, then supp(ψλ) ⊂ Tr,m if and only if,
for any l ∈ {1, . . . , d}, we have rl2j−m ≤ kl ≤ (rl + 1)2j−m − M . Therefore, for
any fixed j ≥ m,
P (Br,m,n) = 1 − (1 − 2j(ν(αmin+1/n)−d))2d(j−m)
≥ 1 − exp(−2−md2jν(αmin+1/n)).
Let us choose N = 2j−m; for any m, n and any j ≥ m, we have
kc(·)
j,· − d(·)
P (f : \r
≥ (cid:0)P (f : ( [(i,k)∈Ij
≥ (cid:0)1 − 2−2d(j−m)
Moreover, for any n,
j,· kℓ∞(Tr,m) ≥ 2−j(αmin+1/n))
A2j−m ,λ) ∩ B2r,m+1,n)(cid:1)2m
− exp(−2−md2jν(αmin+1/n))(cid:1)2md
.
Xm Xj≥m
(1 −(cid:0)1 − 2−2d(j−m)
− exp(−2−md2jν(αmin+1/n))(cid:1)2md
) < ∞.
Therefore, the Borel-Cantelli lemma implies that for any n, there exists m0 ∈ N
such that, for any m ≥ m0, any r and any j ≥ m, the inequality
kc(·)
j,· − d(·)
j,· kℓ∞(Tr,m) ≥ 2−j(αmin+1/n)
(21)
holds almost surely. Using inequality (21), we see that for any n ∈ N,
max{
sup
j≤l≤j+log2 j
kc(·)
l,· − d(·)
l,· kℓ∞(Tr,m),
2−j([αmin]+1)
sup
j−log2 j≤l≤j
14
(2l([αmin]+1)kc(·)
l,· − d(·)
l,· kℓ∞(Tr,m)}
is larger than 2−j(αmin+1/n) for any r, m almost surely. Moreover, since Xν +f ∈
Cαmin(Td) almost surely, we also almost surely have that, for any j and any r, m,
max{
sup
j≤l≤j+log2 j
l,· kℓ∞(Tr,m),
kc(·)
l,· − d(·)
2−j([αmin]+1)
sup
j−log2 j≤l≤j
(2l([αmin]+1)kc(·)
l,· − d(·)
l,· kℓ∞(Tr,m)}
is lower than 2−jαmin. These two bounds allows us to say that for any n ∈ N,
the quantity
log2 max{
sup
j≤l≤j+log2 j
2−j([αmin]+1)
kc(·)
l,· − d(·)
l,· kℓ∞(Tr,m),
sup
(2l([αmin]+1)kc(·)
l,· − d(·)
l,· kℓ∞(Tr,m)}
lim
j→∞
j−log2 j≤l≤j
−j
almost surely belongs to [αmin, αmin + 1/n], for any r, m. Since this relation is
valid for any n, Corollary 1 allows to conclude.
References
[1] P. Abry, P. Goncalves and J. L´evy-V´ehel, Lois d'´echelle, Fractales et On-
delettes, Hermes (2002).
[2] A. Arneodo, B. Audit, N. Decoster, J.-F. Muzy and C. Vaillant,
Wavelet-based multifractal formalism: Applications to DNA sequences, satellite
images of the cloud structures and stock market data, in The Science of Disaster,
A. Bunde, J. Kropp and H.J. Schellnhuber eds., Springer (2002).
[3] J.M. Aubry, F. Bastin, S. Dispa and S. Jaffard, Topological properties of
the sequences spaces S ν , J. Math. Anal. Appl., Vol. 321, pp. 364 -- 387 (2006)
[4] J.M. Aubry, F. Bastin and S. Dispa, Prevalence of multifractal functions in
S ν spaces, J. Fourier Anal. Appl., Vol. 13, pp. 175 -- 185 (2007)
[5] J. Christensen, On sets of Haar measure zero in Abelian Polish groups, Israel
J. Math., Vol 13, pp. 255-260 (1972).
[6] M. Clausel, Quelques notions d'irr´egularit´e uniforme et ponctuelle : le point de
vue ondelettes, Ph. D. Thesis, Universit´e Paris XII (2008).
[7] M. Clausel and S. Nicolay, A multifractal formalism for pointwise anti-
Holderian irregularity, submitted.
[8] M. Clausel ans S. Nicolay, Some prevalent results about strongly monoHolder
functions, Nonlinearity, Vol. 23, pp. 2101 -- 2116 (2010).
[9] I. Daubechies, Orthonormal bases of compactly supported wavelets, Comm. Pure
Appl. Math., Vol. 41, pp. 909 -- 996 (1988).
[10] I. Daubechies, Ten Lectures on Wavelets, SIAM (1992).
[11] S. Dispa, Beyond Besov spaces, S ν spaces: Topology and prevalent properties,
Ph. D. Thesis, University of Li`ege (2006).
15
[12] K. Falconer, Fractal Geometry: Mathematical Foundations and Applications,
Wiley and sons (1990).
[13] . A. Fraysse and S. Jaffard, How smooth is almost every function in a Sobolev
space?, Rev. Mat. Iberoamericana, Vol. 22, pp. 663-682 (2006).
[14] T. Halsey, M. Jensen, L. Kadanoff, I. Procaccia and B. Shraiman, Frac-
tal measures and their singularities: The characterization of strange sets, Phys.
Rev. A, Vol. 33, pp. 1141 -- 1151 (1986).
[15] G.H. Hardy, Weierstrass's non differentiable function, Trans. Amer. Math. Soc.,
Vol. 17, pp.301 -- 325 (1916).
[16] B. Hunt, T. Sauer, J. Yorke, Prevalence: a translation invariance "almost
every" on infinite dimensional spaces, Bull. Amer. (1992).
[17] S. Jaffard, Y. Meryer and R. Ryan, Wavelets: Tools for Science and Tech-
nology, SIAM (2001).
[18] S. Jaffard, Beyond Besov spaces, part I: Distribution of wavelet coeficients, J.
Fourier Anal. Appl., Vol. 10, pp. 221-246 (2004).
[19] S. Jaffard, Wavelet techniques in multifractal analysis, fractal geometry and
applications, Proc. Symp. Pure Math., Vol. 72, pp. 91 -- 151 (2004).
[20] S. Jaffard, S. Nicolay, Pointwise smoothness of space-filling functions, Appl.
Comput. Harmon. Anal., Vol. 26, pp. 181 -- 199 (2009).
[21] S. Mallat, A wavelet tour of signal processing, Academic Press (1998).
[22] B.B. Mandelbrot, Fractals: Form, chance and dimension, Freeman (1977).
[23] P. Mattila, Geometry of Sets and Measures in Euclidian Spaces, Cambridge
University Press (1995).
[24] Y. Meyer, Ondelettes et op´erateurs, Hermann (1990).
[25] J. L´evy-V´ehel and S. Seuret, The local Holder function of a continuous func-
tion, Appl. Comput. Harmon. Anal., Vol 13, pp. 263 -- 276 (2002).
[26] K.R. Sreenivasan, Fractals and multifractals in turbulence, Ann. Rev. Fluid
Mech., Vol. 23, pp. 539 -- 600 (1991).
16
|
1104.4660 | 1 | 1104 | 2011-04-24T22:12:03 | Summable families in tempered distribution spaces | [
"math.FA",
"quant-ph"
] | In this note we define summable families in tempered distribution spaces and we state some their properties and characterizations. Summable families are the analogous of summable sequences in separable Hilbert spaces, but in tempered distribution spaces, having elements (functional) realizable as generalized vectors indexed by real Euclidean spaces (not pointwise defined ordered families of scalars indexed by real Euclidean spaces in the sense of distributions). Any family we introduce here is summable with respect to every tempered system of coefficients belonging to a certain normal space of distributions, in the sense of superpositions. The summable families we present in this note are one possible rigorous and simply manageable mathematical model for the infinite families of vector-states appearing in the formulation of the continuous version of the celebrated Principle of Superpositions in Quantum Mechanics. | math.FA | math |
Summable families in tempered distribution
spaces
David Carf`ı
Abstract
In this note we define summable families in tempered distribu-
tion spaces and state some their properties and characterizations.
Summable families are the analogous of summable sequences in sep-
arable Hilbert spaces, but in tempered distribution spaces having el-
ements (functional) realizable as generalized vectors indexed by real
Euclidean spaces (not pointwise defined ordered families of scalars in-
dexed by real Euclidean spaces in the sense of distributions). Any
family we introduce here is summable with respect to every tempered
system of coefficients belonging to a certain normal spaces of distri-
butions, in the sense of superpositions. The summable families we
present in this note are one possible rigorous and simply manageable
mathematical model for the infinite families of vector-states appearing
in the formulation of the continuous version of the celebrated Principle
of Superpositions in Quantum Mechanics.
1 Characterizations of Dfamilies (*)
As we proved the summability theorem of Schwartz families and character-
ization of summability, in a perfectly analogous way, it can be proved the
following theorem.
The Dfamilies in D′
n can be defined analogously to the Sfamilies in S ′
n.
Definition (family of tempered distributions of class D). Let v be
n indexed by the Euclidean space Rm.
a family of distributions in the space D′
1
The family v is called a Schwartz family of class S or even Dfamily if,
for each test function φ ∈ Dn, the image of the test function φ by the family
v - that is the function v(φ) : Rm → K defined by
v(φ)(p) := vp(φ),
for each index p ∈ Rm - belongs to the space of test functions Dm. We shall
denote the set of all Dfamilies by D(Rm, D′
n).
The following theorem holds since the Corollary of page 91 of Dieudonn´e-
Schwartz seminal paper holds true because D′
n is an LF-space.
Theorem (basic properties on Dfamilies). Let v ∈ D(Rm, D′
n) be a
family of distributions. Then the following assertions hold and are equivalent:
1)
2)
3)
4)
for every a ∈ D′
is a distribution;
m the composition u = a ◦bv, i.e., the functional
u : Dn → K : φ 7→ a (bv(φ)) ,
the operator bv is transposable;
the operator bv is (σ(Dn), σ(Dm))-continuous from Dn to Dm;
the operator bv is a strongly continuous from (Dn) to (Dm).
2 Algebraic EFamilies and Esummable fami-
lies
Let us begin with a family of tempered distribution which is not of class S.
Example (a family that is not of class S). Let u be a distribution
in S ′
n and let v be the constant family in S ′
n, indexed by the Euclidean space
Rm, defined by vy = u, for each point y ∈ Rm. Then, if the distribution u
2
is different from zero, v is not of class S. In fact, let φ ∈ S(Rn, K) be such
that u(φ) 6= 0; for every point-index y ∈ Rm, we have
v(φ)(y) = vy(φ) =
= u(φ)1Rm(y),
where, 1Rm is the constant K-functional on Rm with value 1. Thus, the
function v(φ) is a constant K-functional on Rm different from zero, and so it
cannot live in the space Sm.
The preceding example induces us to consider other classes of families in
addition to the Sfamilies, for this reason, we will give the following definitions.
We shall denote by Cm the space C 0(Rm, K) of continuous functions de-
fined on the Euclidean space Rm and with values in the scalar field K.
Definition (Efamilies and algebraically Esummable families). Let
E be a subspace of the function space F (Rm, K) (without any topology) con-
taining the space Sm. If v is a family in the distribution space S ′
n indexed by
Rm, we say that the family v is an Efamily if, for every test function φ in
Sn, the image v(φ) of the test function by the family v lies in the subspace
E. An Efamily v is said to be algebraically Esummable or E ∗summable
if, for every functional a in the algebraic dual E ∗, the functional
ZRm
av : Sn → K : φ 7→ a(v(φ))
is a tempered distribution living in S ′
n. More generally, if F is a part of
the algebraic dual E ∗ we say that the family v is F summable if the above
functional φ 7→ a(v(φ)) is a tempered distribution living in S ′
n, for every
functional a in F .
Example. With the preceding new definition, the family of the above
n, where by Em we (in standard way) denote the
example is a Emfamily in S ′
space C ∞(Rm, K) of smooth function from Rm into K.
Moreover, for every tempered distribution a in E ′
m, we have
a(v(φ)) = a(u(φ)1Rm) =
= u(φ)a(1Rm) =
= u(φ)ZRm
a,
3
where byRRm a we denote the integral of the distribution a (we recall that the
compact support distributions are integrable and their integrals is defined as
their value on the constant unit functional 1 Rm), so that
ZRm
av =(cid:18)ZRm
a(cid:19) u,
and the family v is E ′
msummable.
3
EFamilies and Esummable families
Remark. Let E be a subspace of the function space F (Rm, K) (without any
topology) containing the space S m and let w be a Hausdorff locally convex
topology on the subspace E.
• If the topological vector space (Sm) is continuously imbedded in the
w is continuously imbedded in
m. In this case, in Distribution Theory, we say that the dual
space Ew, then, the topological dual E ′
the space S ′
E ′
w is a space of tempered distribution on Rm.
m is contained in the dual E ′
• Moreover, if the topological vector space Ew is continuously imbedded
in the space (Cm), then the dual C ′
w. In other
terms, every Radon measure on Rm with compact support belongs to
the dual E ′
w and, in particular, the Dirac family is contained in the
topological dual E ′
w. Since the Dirac basis in sequentially total in the
space of tempered distributions S ′
m (the linear hull of the Dirac basis
is dense in S ′
w is continuously imbedded in the
space S ′
m itself, the Dirac family shall be sequentially total also in the
topological vector space (E ′
w)σ, that is sequentially dense with respect
to the weak* topology σ(E ′, E).
m) and since the space E ′
Now we can give two new definitions.
Definition (Efamilies and Esummable families). Let E be a subspace
of the space F (Rm, K) containing the space Sm and endowed with a locally
convex linear topology w. If v is a family in the distribution space S ′
n indexed
4
by Rm. We say that the family v is an Efamily if, for every test function φ
in Sn, the image v(φ) of the test function by the family v lies in the subspace
E. An Efamily is said to be Esummable if for every tempered distribution
a in the topological dual E ′
w the functional
Sn → K : φ 7→ a(v(φ))
is a tempered distribution in S ′
n.
4 Normal spaces of distributions and summa-
bility
Definition (of normal space of test function for S ′
m). We will call a
locally convex topological vector space E a normal space of test functions
for the distribution space S ′
m if it verifies the following properties
• the space E is an algebraic subspace of the space Cm;
• the space E contains the space Sm;
• the topological vector space (Sm) is continuously imbedded and dense
in the topological vector space E;
• the topological vector space E is continuously imbedded in the space
(Cm).
In these conditions the dual E ′ is called a normal space of tempered
distributions on Rm.
Theorem (on the Efamily generated by a linear and continuous
operator). Let E be a normal space of test functions for the space S ′
m, let
A : Sn → E be a linear and continuous operator of the space (Sn) into the
space E and let δ be the Dirac family in C ′
m. Then, the family of functionals
A∨ := (δp ◦ A)p∈Rm
is a family of distribution in S ′
n and it is an Efamily.
5
We can prove that:
Theorem. Let E be a normal space of test functions for the distribu-
n (obviously indexed by the m-
m. Then, every Efamily in S ′
tion space S ′
dimensional Euclidean space) is Esummable.
References
[1] J. Barros-Neto, An Introduction to the theory of distributions, Marcel
Dekker, Inc. NewYork, 1981
[2] N. Boccara, Functional analysis, an introduction for physicists, Academic
press, Inc. 1990
[3] J. Horvath, Topological Vector Spaces and Distributions (Vol.I), Addison-
Wesley Publishing Company, 1966
[4] L. Schwartz, Th´eorie des distributions, Hermann, Paris 1966
[5] J. Dieudonn´e, "La dualit´e dans les espaces vectoriels topologiques",
Annales scietifiques de l'E.N.S. 3 es´erie, tome 59 p. 107 - 139, 1942.
http://www.numdam.org
[6] J. Dieudonn´e L. Schwartz, "La dualit´e dans les espaces (F ) and
tome 1 p. 61 - 101, 1949.
(LF)", Annales de l'institut Fourier,
http://www.numdam.org
[7] P. A. M. Dirac, The principles of Quantum Mechanics, Oxford Claredon
press, 1930.
[8] D. Carf`ı, "S-linear operators in quantum mechanics and in eco-
nomics", (APPS), volume 6 pp.7-20, 2004. (no.1 electronic edition, )
http://vectron.mathem.pub.ro/apps/v6/a6.htm
[9] D. Carf`ı, "Dirac-orthogonality in the space of tempered distributions",
Journal of computational and applied mathematics, vol. 153, pp.99-107,
numbers 1-2, 1 april 2003.
6
[10] D. Carf`ı, "S-diagonalizable operators in quantum mechanics", Glasnik
Mathematicki, vol. 40 n.2, pp. 267-307, 2005.
[11] D. Carf`ı, "On the Schodinger's equation associated with an operator...",
Rendiconti del Seminario Matematico di Messina, n. 8 serie II, 2001.
[12] D. Carf`ı, "Quantum statistical systems with a continuous range of
states", series of Advances in Mathematics for Applied Sciences, vol. 69,
pp. 189 - 200, edited by World Scientific. 2005.
[13] D. Carf`ı, "Dyson formulas for Financial and Physical evolutions in S ′
n",
Proceedings of the "VIII Congresso SIMAI", Ragusa, Baia Samuele, 22 -
26 Maj 2006.
[14] D. Carf`ı, "Prigogine approach to irreversibility for Financial and Physi-
cal applications", Supplemento Atti dell'Accademia Peloritana dei Peri-
colanti di Messina, Proceedings Thermocon'05, 2006. ISSN: 0365-0359.
http://antonello.unime.it/atti/
[15] D. Carf`ı, "S-convexity in the space of Schwartz distributions and appli-
cations", Supplemento ai Rendiconti del Circolo Matematico di Palermo,
Serie II - Numero 77 - pp. 107-122, Anno 2006,. ISSN: 0009-725X.
[16] D. Carf`ı, "S-Linear Algebra in Economics and Physics", Applied Sci-
ences (APPS), vol. 9, 2007, ISSN 1454-5101, forthcoming volume.
David Carf`ı
Faculty of Economics
University of Messina
[email protected]
7
|
1904.11457 | 3 | 1904 | 2019-10-09T19:39:58 | Non-geodesic Spherical Funk Transforms with One and Two Centers | [
"math.FA"
] | We study non-geodesic Funk-type transforms associated with cross-sections of the n-sphere by k-dimensional planes passing through an arbitrary fixed point inside the sphere. The main results include injectivity conditions for these transforms, inversion formulas, and connection with geodesic Funk transforms. We also show that, unlike the case of planes through a single common center, the integrals over spherical sections by planes through two distinct centers provide the corresponding reconstruction problem a unique solution. | math.FA | math |
NON-GEODESIC SPHERICAL FUNK TRANSFORMS
WITH ONE AND TWO CENTERS
M. AGRANOVSKY AND B. RUBIN
Abstract. We study non-geodesic Funk-type transforms on the
unit sphere Sn in Rn+1 associated with cross-sections of Sn by k-
dimensional planes passing through an arbitrary fixed point inside
the sphere. The main results include injectivity conditions for these
transforms, inversion formulas, and connection with geodesic Funk
transforms. We also show that, unlike the case of planes through
a single common center, the integrals over spherical sections by
planes through two distinct centers provide the corresponding re-
construction problem a unique solution.
1. Introduction
Let Sn be the unit sphere in Rn+1. Given a point a inside Sn, we
denote by Gra(n + 1, k), 1 ≤ k ≤ n, the Grassmann manifold of k-
dimensional affine planes in Rn+1 passing through a. The aim of the
paper is to study injectivity of the generalized Funk transform
(Faf )(τ ) = ZSn∩τ
f (x) dσ(x),
τ ∈ Gra(n + 1, k),
(1.1)
and obtain inversion formulas for Fa in suitable classes of functions.
The classical case Fa = Fo, when a = o is the origin, goes back
to the pioneering works by Funk [2, 3] (n = 2), which were inspired
by Minkowski [10]. A generalization of the Funk transform Fo to ar-
bitrary 1 ≤ k ≤ n is due to Helgason [8]; see also [9, 18, 20] and
references therein. Operators of this kind play an important role in
convex geometry, spherical tomography, and various branches of Anal-
ysis [4, 6, 7, 20, 13, 14].
The case when a differs from the origin is relatively new in mod-
ern literature, though Funk-type transforms on S2 for noncentral plane
sections were considered by Gindikin, Reeds, and Shepp [7] in the
framework of the kappa-operator theory. One should also mention
non-geodesic Funk-type transforms studied by Palamodov [14, Section
5.2]. Inversion formulas for these transforms were obtained in terms of
2010 Mathematics Subject Classification. Primary 44A12; Secondary 37E30.
1
2
M. AGRANOVSKY AND B. RUBIN
delta functions and differential forms. Operators (1.1) with a 6= o are
non-geodesic too, however, they differ from those in [14]. In particular,
they are non-injective. Non-geodesic Funk-type transforms over sub-
spheres of fixed radius were studied by the second co-author in [17],
where the results fall into the scope of number theory.
The case a 6= o in (1.1) with k = n was considered by Salman;
see [23] for n = 2 and [24] for any n ≥ 2. To avoid non-uniqueness,
he imposed restriction on the support of the functions that makes his
operator different from ours. The stereographic projection method of
[23, 24] makes it possible to reduce inversion of Salman's operator to
the similar problem for a certain Radon-like transform over spheres in
Rn.
The next step was due to Quellmalz [15] for n = 2, who expressed
Fa through the totally geodesic Funk transform Fo and thus explicitly
inverted this operator on a certain subclass of continuous functions. If
a = o this subclass consists of even functions on Sn. The results from
[15] were generalized by Quellmalz [16] and Rubin [21] to any n ≥ 2
with k = n. The paper [21] also contains an alternative inversion
method for Salman's operators.
Our aim in the present article is two-fold. First, we characterize
the kernel (the null subspace) of Fa and the subclass of all continuous
functions on which Fa is injective. We also obtain inversion formulas
for Fa on that subclass for any 1 ≤ k ≤ n and thus generalize the
corresponding results from [21].
Second, to achieve uniqueness in the reconstruction problem, we con-
sider sections by planes through two distinct centers. To the best of
our knowledge, this approach is new. We shall prove that for any pair
of distinct points a and b inside the sphere, the kernels of the corre-
sponding transforms Fa and Fb have trivial intersection. The latter
means that, unlike the case of a single center, the collection of data
from two distinct centers provides the reconstruction problem a unique
solution. We also develop an analytic procedure of the reconstruction,
that reduces to a certain dynamical system on Sn.
The results of this paper extend to the case when the point a lies
outside Sn, and to arbitrary pairs of distinct centers a, b in Rn+1. We
plan to address these cases elsewhere.
Plan of the Paper. Section 2 contains notation and necessary prelim-
inaries related to Mobius-type automorphisms of the sphere. In Section
3 we describe the kernel of the operator Fa on continuous functions and
characterize the subclass of functions on which Fa is injective. We also
obtain an explicit inversion formula for Fa on that subclass. Section
NON-GEODESIC FUNK TRANSFORMS
3
4 deals with the system of two equations, Faf = g, Fbf = h, corre-
sponding to distinct centers a and b inside the sphere. Unlike the case
of a single common center, such a system determines f uniquely and
the function f can be reconstructed by a certain pointwise convergent
series. Norm convergence of this series is studied in Section 5. It turns
out that the series does not converge uniformly on the entire sphere
Sn (only on some compact subsets of Sn), however, it converges in the
Lp(Sn)-norm for all 1 ≤ p ≤ p0, p0 = n/(k − 1), and this bound is
sharp. In Section 6 we prove Theorem 3.1, which was formulated with-
out proof in Section 3. This theorem plays a key role in the paper. It
states that the shifted transform Fa is represented as Fa = NaFoMa,
where Na and Ma are the suitable bijections and Fo is the classical
Funk transform corresponding to a = o.
The main results are contained in Theorems 3.4, 4.2, 5.2, and 5.4.
2. Preliminaries
2.1. Notation. In the following, Bn+1 = {x ∈ Rn+1 : x < 1} is the
open unit ball in Rn+1, Sn is its boundary, x · y is the usual dot prod-
uct. The notation C(Sn) and Lp(Sn) for the corresponding spaces of
continuous and Lp functions on Sn is standard. If x is the variable of
integration over Sn, then dx stands for the O(n + 1)-invariant surface
area measure on Sn, so that RSn dx = 2π(n+1)/2/Γ((n + 1)/2). We write
dσ(x) for the induced surface area measure on lower dimensional spher-
ical sections. The letter x can be replaced by another one, depending
on the context.
We denote by Mn,m the space of real matrices having n rows and
m columns; M′ is the transpose of the matrix M, Im is the identity
m × m matrix. For n ≥ m, St(n, m) = {M ∈ Mn,m : M′M = Im}
denotes the Stiefel manifold of orthonormal m-frames in Rn; Gra(n, m)
is the Grassmann manifold of m-dimensional affine planes in Rn passing
through a fixed point a. We will be mainly dealing with the manifolds
St(n+1, n+1−k), Gra(n + 1, k), and Gro(n + 1, k) (i.e. a = o), 1 ≤
k ≤ n. Given a frame ξ ∈ St(n+1, n+1−k), the notation ξ⊥ stands
for the k-dimensional linear subspace orthogonal to ξ; {ξ} denotes an
(n + 1 − k)-dimensional linear subspace spanned by ξ. All points in
Rn+1 are identified with the corresponding column vectors.
2.2. Spherical Automorphisms. We recall some basic facts; see,
e.g., Rudin [22, Section 2.2.1)], Stoll [25, Section 2.1]. Given a point
a ∈ Bn+1 \ {o}, we denote by Pa and Qa = In+1 − Pa the orthogo-
nal projections of Rn+1 onto the direction of a and the subspace a⊥,
4
M. AGRANOVSKY AND B. RUBIN
respectively. If x ∈ Rn+1, then
Pax =
a · x
a2 a.
Let
ϕax =
which is a one-to-one Mobius transformation satisfying
a − Pax − saQax
,
1 − x · a
sa =p1 − a2,
ϕa(o) = a,
ϕa(a) = o,
ϕa(ϕax) = x,
1 − ϕax2 =
(1 − a2)(1 − x2)
(1 − x · a)2
,
x · a 6= 1.
(2.1)
(2.2)
(2.3)
(2.4)
If x ∈ Sn, then
1 − a · ϕax
1 + a · ϕax
=
1 − a2
a − x2 .
Properties (2.2)-(2.3) can be checked by straightforward computation.
By (2.3), ϕa maps the ball Bn+1 onto itself and preserves Sn.
Remark 2.1. It is known that the ball Bn+1 with the relevant metric
can be considered as the Poincar´e model of the real (n+ 1)-dimensional
hyperbolic space Hn+1. There is an intimate connection between the
Mobius transformations of Bn+1 and the group O(1, n + 1) in the hy-
perboloid model of Hn+1. In the present article we do not exploit this
connection. An interested reader may be referred, e.g., to Beardon
[1, Section 3.7], Gehring, Martin, Palka [5, Section 3.7], Mostow [12,
Theorem 1.1].
Lemma 2.2. For any f ∈ L1(Sn),
(f ◦ ϕa)(y)
(1 − a · y)n dy,
sa =p1 − a2.
f (x) dx = sn
aZSn
ZSn
(2.5)
Proof. We write x in spherical coordinates
x = √1 − u2 θ + ua,
to obtain
a =
,
a
a
u ≤ 1,
θ ∈ Sn ∩ a⊥,
f (x) dx =
1
Z−1
ZSn
By (2.1),
(1−u2)(n−2)/2 du ZSn∩ a⊥
√1 − v2 θ + va,
ϕax = −
f(cid:16)√1−u2 θ+u a(cid:17) dθ.
v = a − u
1 − au
.
(2.6)
(2.7)
NON-GEODESIC FUNK TRANSFORMS
5
Note that the map u → v is an involution. Changing variable
and taking into account that
u = a − v
1 − av
du
dv
= a2 − 1
(1 − av)2 ,
1 − u2 =
(1 − a2)(1 − v2)
(1 − av)2
,
we have
ZSn
f (x) dx = (1 − a2)n/2
1
Z−1
(1−v2)(n−2)/2
(1 − av)n dv
× ZSn∩ a⊥
aZSn
= sn
f p1 − a2 √1 − v2
1 − av
(f ◦ ϕa)(y)
(1 − a · y)n dy,
as desired.
θ + a − v
1 − av
a! dθ
(cid:3)
We also define the reflection τa : Sn → Sn about the point a ∈ Bn+1:
(2.8)
(a2 − 1) x + 2(1 − x · a) a
τax =
.
x − a2
It assigns to x ∈ Sn the antipodal point τax ∈ Sn that lies on the line
passing through x and a. A similar reflection map about the origin o
is denoted by τo, so that τox = −x.
The map ϕa intertwines reflections τa and τo, that is,
ϕaτa = τoϕa.
(2.9)
Indeed, ϕa maps chords of the ball onto chords. Hence, for any x ∈ Sn,
the segment [x, τax] is mapped onto the segment [ϕax, ϕaτax]. Since
the first segment contains a, the second one contains ϕa(a) = o. The
latter means that the points ϕax and ϕaτax are symmetric with respect
to the origin, that is, ϕaτax = τoϕax.
Lemma 2.3. If f ∈ L1(Sn) and a ∈ Bn+1, then
a − x2(cid:19)n
f (x) (cid:18) 1 − a2
a − x2(cid:19)n
f (τax) (cid:18) 1 − a2
f (τax) dx =ZSn
f (x) dx =ZSn
ZSn
ZSn
(2.11)
(2.10)
dx,
dx.
6
M. AGRANOVSKY AND B. RUBIN
Proof. By (2.9) and (2.5),
ZSn
f (τax) dx = ZSn
f (ϕaτoϕax) dx
(set x = ϕaτoy)
= (1 − a2)n/2ZSn
= ZSn
1 + a · y(cid:19)n
(1 − a · y)n (cid:18)1 − a · y
(f ◦ ϕa)(y)
1 + a · ϕax(cid:19)n
f (x) (cid:18) 1 − a · ϕax
dx.
dy
It remains to apply (2.4). The second equality follows from the first
one: just replace f (x) by f (τax) and use τaτax = x.
(cid:3)
3. The Shifted Funk Transform
3.1. Inversion Procedure. The following theorem establishes con-
nection between the shifted Funk transform
(Faf )(τ ) = ZSn∩ τ
f (x) dσ(x),
τ ∈ Gra(n + 1, k),
(3.1)
and the classical Funk transform Fo = Faa=o that takes functions on Sn
to functions on Gro(n + 1, k). Given a function f on Sn and a function
Φ on Gro(n + 1, k), we denote
(f ◦ ϕa)(y),
(NaΦ)(τ ) = Φ(ϕaτ ),
1−a · y(cid:19)k−1
(Maf )(y) =(cid:18) sa
where sa =p1 − a2 and ϕa is an automorphism (2.1).
Theorem 3.1. Let 1 ≤ k ≤ n, a ∈ Bn+1. If f ∈ C(Sn), then
Faf = NaFoMaf.
(3.2)
(3.3)
The proof of this theorem is given in Section 6.
The Funk transform Fo is injective on the subspace C +(Sn) of even
functions, whilst the subspace C −(Sn) of odd functions is the kernel of
Fo in C(Sn); see, e.g., [9, 18, 19, 20]. We denote by Fo the restriction
of Fo onto C +(Sn).
There exist several different approaches to inversion of Fo. We recall
one of them. Given ϕ = Fof , f ∈ C +(Sn), consider the mean value
operator
(F ∗
x ϕ)(r) =
Z
{ζ∈Gro(n+1,k): d(x,ζ)=r}
ϕ(ζ) dm(ζ),
0 < r < 1,
(3.4)
NON-GEODESIC FUNK TRANSFORMS
7
where integration is performed with respect to the relevant probability
measure over the set of all planes ζ ∈ Gro(n + 1, k) at geodesic distance
d(x, ζ) = cos−1 r from x.
Theorem 3.2. (cf. [19, Theorem 5.3]) A function f ∈ C +(Sn) can be
reconstructed from ϕ = Fof by
o ϕ)(x)
f (x) ≡ ( F −1
s→1(cid:18) 1
= lim
2s
∂
∂s(cid:19)k" π−k/2
Γ(k/2)
(3.5)
s
Z0
(s2−r2)k/2−1 (F ∗
x ϕ)(r) rk dr#.
In particular, for k even,
( F −1
o ϕ)(x) = lim
s→1
1
2πk/2(cid:18) 1
2s
∂
∂s(cid:19)k/2
[sk−1(F ∗
x ϕ)(s)].
(3.6)
The limit in these formulas is understood in the sup-norm.
Now we proceed to inversion of Fa, which, by Theorem 3.1, is fac-
torized as Fa = NaFoMa. Here the operators Ma and Na are injective,
so that
(M −1
a f )(x) = (1 − a · ϕax)k−1 (f ◦ ϕa)(x),
N −1
a Φ = Φ ◦ ϕa. (3.7)
The following definition is motivated by the factorization Fa = NaFoMa
and nicely agrees with the case a = o.
Definition 3.3. A function f ∈ C(Sn) is called a-even (or a-odd) if
Maf is even (or odd, resp.) in the usual sense. The subspaces of all
a-even and a-odd continuous functions on Sn will be denoted by C +
a (Sn)
and C −
a (Sn) will be
denoted by Fa.
Theorem 3.4. Let 1 < k ≤ n. Then ker (Fa) = C −
stricted operator Fa is injective. A function f ∈ C +
reconstructed from g = Faf by
f ≡ F −1
a (Sn), respectively. The restriction of Fa onto C +
a (Sn) and the re-
a (Sn) can be uniquely
a g = M −1
a
F −1
o N −1
a g,
(3.8)
where M −1
a , F −1
o
, and N −1
a
are defined by (3.7) and Theorem 3.2.
This statement is an immediate consequence of (3.3) and the corre-
sponding results for Fo.
Remark 3.5. In the case k = 1, which is not included in Theorem 3.4,
the plane τ is a line and the integral (1.1) is a sum of the values of f
8
M. AGRANOVSKY AND B. RUBIN
at the points where this line intersects the sphere. If x is one of such
points and La,x is the line through a and x, then
(Faf )(La,x) = f (x) + f (τax).
(3.9)
The a-odd functions, for which f (x) = −f (τax), form the kernel of the
operator (3.9). An a-even function f , satisfying f (x) = f (τax), can be
reconstructed from (Faf )(La,x) by the formula
f (x) =
1
2
(Faf )(La,x).
(3.10)
3.2. Alternative description of the subspaces C ±
a (Sn). We set
ρa(x) =(cid:18) 1 − a2
a − x2(cid:19)k−1
where τa is the reflection (2.8).
,
(Waf )(x) = ρa(x)f (τax),
(3.11)
Lemma 3.6. The operator Wa is an involution, i.e., WaWaf = f .
Proof. The statement is obvious for a = o, when (W0f )(x) = f (−x).
It is also obvious for any a ∈ Bn if k = 1. In the general case, taking
into account that τaτax = x, we have
(WaWaf )(x) =(cid:20) 1 − a2
a − x2
(1 − a · ϕax) (1 − a · ϕaτax)
(1 + a · ϕax) (1 + a · ϕaτax)
=
f (x).
a − τax2(cid:21)k−1
1 − a2
(1 − a · ϕax) (1 + a · ϕax)
(1 + a · ϕax) (1 − a · ϕax)
= 1.
By (2.4) and (2.9), the expression in square brackets can be written as
This gives the result.
Theorem 3.7. A function f ∈ C(Sn) is a-even (or a-odd) if and only
if f = Waf (or f = −Waf , respectively).
Proof. By Definition 3.3, f ∈ C(Sn) is a-even if and only if (Maf )(y) =
(Maf )(−y) for all y ∈ Sn. The latter is equivalent to
(cid:3)
(f ◦ ϕa)(y) =(cid:18)1 − a · y
1 + a · y(cid:19)k−1
(f ◦ ϕa)(−y),
or (set y = ϕax and use (2.4) and (2.9))
f (x) =(cid:18)1 − a · ϕax
1 + a · ϕax(cid:19)k−1
f (τax) = ρa(x)f (τax) = (Waf )(x).
The proof for the a-odd functions is similar.
(cid:3)
NON-GEODESIC FUNK TRANSFORMS
9
Corollary 3.8. Every function f ∈ C(Sn) can be represented as a sum
of its a-even and a-odd parts. Specifically,
f = f +
a + f −
a ,
f ±
a =
f ± Waf
2
.
(3.12)
Proof. The first equality follows from the second one. Further, by
Lemma 3.6,
Waf ±
a =
Waf ± WaWaf
Waf ± f
=
2
2
a is a-even and f −
a is a-odd.
= ±f ±
a .
Hence, by Theorem 3.7, f +
(cid:3)
4. Reconstruction from Two Centers
As we have seen in Section 3, a generic function f ∈ C(Sn) cannot
be reconstructed from Faf . Because Faf = Faf +
a , one can reconstruct
only the a-even part f +
a of f , whilst the a-odd part is lost. Our aim is
to show that complete reconstruction becomes possible if we consider
two distinct centers instead of one. Specifically, let a, b ∈ Bn+1, a 6= b.
Consider the system of two equations
(4.1)
and suppose that a function f ∈ C(Sn) satisfies this system. Then
Faf +
a = g, Fbf +
Faf = g,
Fbf = h,
b = h, and therefore, by (3.12),
f + Waf
f + Wbf
= F −1
a g,
2
f +
b ≡
2
= F −1
b h,
f +
a ≡
where
(Waf )(x) = ρa(x)f (τax),
(Wbf )(x) = ρb(x)f (τbx).
(4.2)
Setting
g1 = 2 F −1
a g,
h1 = 2 F −1
b h,
we obtain a pair of functional equations
f = g1 − Waf,
f = h1 − Wbf.
(4.3)
Then we substitute f from the second equation into the right-hand side
of the first one to get
f = W f + q,
W = WaWb,
Iterating (4.4), we obtain
m−1
q = g1 − Wah1.
(4.4)
f = W mf +
W jq;
m = 1, 2, . . . .
(4.5)
Xj=0
This equation generates a dynamical system on Sn.
10
M. AGRANOVSKY AND B. RUBIN
Lemma 4.1. Let a∗ and b∗ be the points on Sn that lie on the straight
line through a and b. Suppose that a is closer to a∗ than b. If W =
(W mf )(x) = 0 for all x ∈ Sn \ {a∗} and 1 < k ≤ n.
WaWb, then lim
m→∞
If k = 1 and x ∈ Sn \ {a∗}, then lim
Proof. We observe that
(W mf )(x) = f (b∗).
m→∞
(W f )(x) = (WaWbf )(x) = ρa(x)ρb(τax)f (τbτax).
(4.6)
Denote
ρ(x) = ρa(x)ρb(τax) =(cid:20) (1 − a2) (1 − b2)
a − x2 b − τax2 (cid:21)k−1
Then (W f )(x) = ρ(x)f (Tx) and, by iteration,
(W mf )(x) = ωm(x) f (Tm+1x),
ωm(x) =
, T = τbτa.
(4.7)
m
ρ(Tjx).
(4.8)
Yj=0
For any x 6= a∗, the mapping T preserves the circle Cx,a,b in the 2-plane
spanned by x, a and b, and leaves the points a∗ and b∗ fixed. A simple
geometric consideration in the 2-plane shows that the distance from
the points Tjx ∈ Cx,a,b to b∗ monotonically decreases, and therefore,
the sequence Tjx has a limit. This limit must be a fixed point of the
mapping T, and hence Tjx → b∗ as j → ∞. Because ρ is continuous,
it follows that
(4.9)
ρ(Tjx) = ρ(b∗).
lim
j→∞
Using this fact, let us show that if k > 1, then
lim
m→∞
ωm(x) = 0.
(4.10)
Once (4.10) has been proved, the statement of the lemma for k > 1
will follow because the factor f (Tm+1x) has finite limit f (b∗).
To prove (4.10), it suffices to show that
ρ(b∗) < 1,
where, by (4.7),
Let
ρ(b∗) =(cid:20)(1 − a2) (1 − b2)
a − b∗2 a∗ − b2 (cid:21)k−1
.
(4.11)
(4.12)
a = a∗+ t(b∗−a∗),
b = a∗+ s(b∗−a∗),
0 < t < s < 1.
(4.13)
Taking into account that a∗ = b∗ = 1 and using (4.13), we obtain
1−a2 = 2t(1−t)(1−a∗· b∗),
1−b2 = 2s(1−s)(1−a∗· b∗), (4.14)
NON-GEODESIC FUNK TRANSFORMS
11
a−b∗2 = 2(1−t)2(1−a∗· b∗),
a∗−b2 = 2s2(1−a∗· b∗).
Hence
ρ(b∗) =(cid:20)t(1 − s)
s(1 − t)(cid:21)k−1
< 1.
(4.15)
The last inequality is an immediate consequence of the assumption
0 < t < s < 1.
The case k = 1 is simpler.
In this case ρ(x) = 1, and therefore,
(cid:3)
(W mf )(x) = f (Tm+1x) → f (b∗) as m → ∞, x ∈ Sn \ {a∗}.
a (Sn) ∩ C −
The above reasoning yields the following preliminary conclusion. If
a 6= b, then, by Theorem 3.4, the kernel of the map f → (Faf, Fbf ),
b (Sn). But if f is odd with respect to both
f ∈ C(Sn), is C −
a and b, then, by Theorem 3.7, W f = f . By Lemma 4.1 it follows
that f (x) = 0 for all x ∈ Sn \ {a∗}. However, since f is continuous, we
must have f = 0 everywhere on Sn. In particular, it follows that f can
be reconstructed from the knowledge of Faf and Fbf , or, what is the
same, from f +
a and f +
b .
More precisely, we have the following result.
Theorem 4.2. Let Wa and Wb be involutions (4.2), 1 < k ≤ n. If the
system of equations Faf = g and Fbf = h has a solution f ∈ C(Sn),
then this solution is unique and can defined by the pointwise convergent
series
∞
f (x) =
W jq(x),
Xj=0
a g−Wa F −1
x 6= a∗,
a and F −1
b
(4.16)
being defined
b h], F −1
where W = WaWb, q = 2 [ F −1
as in (3.8). Alternatively,
f (x) =
W jr(x),
x 6= b∗,
(4.17)
∞
Xj=0
a g].
b h − Wb F −1
where W = WbWa and r = 2 [ F −1
Proof. To prove (4.16), it suffices to pass to the limit in (4.5), taking
into account that, by Lemma 4.1, the remainder (W mf )(x) of the series
(4.16) converges to zero for every x 6= a∗. An alternative formula (4.17)
then follows if we interchange a and b, g and h.
Remark 4.3. In the case k = 1, a function f ∈ C(Sn) can be recon-
structed from the system Faf = g, Fbf = h as follows. By Lemma 4.1,
(W mf )(x) → f (b∗) as m → ∞. Hence
(cid:3)
∞
f (x) =
Xj=0
q(Tj+1x) + f (b∗),
x 6= a∗, T = τbτa,
(4.18)
12
M. AGRANOVSKY AND B. RUBIN
where q(x) = 2 [( F −1
a g)(x) − ( F −1
( F −1
a g)(x) =
1
2
g(La,x),
b h)(τax)]. By (3.10),
1
2
b h)(τax) =
( F −1
h(Lb,τax),
where the line La,x passes through a and x and Lb,τax passes through b
and τax. It follows that
q(x) = g(La,x) − h(Lb,τax).
(4.19)
Similarly, ( W mf )(x) → f (a∗), and we have
∞
f (x) =
Xj=0
r( Tj+1x) + f (a∗),
x 6= b∗,
T = τaτb,
(4.20)
r(x) = h(Lb,x) − g(La,τbx).
(4.21)
The series (4.18) and (4.20) reconstruct f up to unknown additive
constants f (a∗) or f (b∗), where a∗ and b∗ are the endpoints of the chord
through a and b. However, complete reconstruction is still possible, if
we apply symmetrization, by summing (4.18) and (4.20). This gives
the following result.
Theorem 4.4. Let k = 1. Then
2f (x) =
∞
Xj=0
q(Tj+1x)+
∞
Xj=0
r( Tj+1x)+Fa(La,b),
x 6= a∗, b∗, (4.22)
where q and r are defined by (4.19) and (4.21), respectively, La,b is
the line through a and b, and Fa(La,b) = f (a∗) + f (b∗) (= Fb(La,b)) is
known. The values of f at the points a∗ and b∗ can be reconstructed by
continuity.
5. Norm Convergence of the Reconstructing Series
Reconstruction of f by the pointwise convergent series (4.16) and
(4.17) gives a little possibility to control the accuracy of the result
because the rate of the pointwise convergence depends on the point.
Therefore, it is natural to look at the convergence in certain normed
spaces. Below we explore such convergence in the spaces C(Sn) and
Lp(Sn). As above, we keep the notation a∗ and b∗ for the endpoints of
the chord through a and b.
Consider the most interesting case k > 1. By (4.5), the convergence
of the series (4.16) to f is equivalent to convergence of its remainder
W mf to 0 as m → ∞. Thus, it suffices to confine to W mf .
point of the mapping T, we have
(W mf )(a∗) = ρ(a∗)m+1f (a∗),
Qj=0
ρ(a∗) =(cid:20)(1 − a2)(1 − b2)
a − a∗2b − b∗2 (cid:21)k−1
.
NON-GEODESIC FUNK TRANSFORMS
13
We first note that the series (4.16) may diverge at the point a∗.
ρ(Tja∗) and a∗ is a fixed
Indeed, because (W mf )(a∗) = f (Tm+1a∗)
m
Suppose that a and b are symmetric with respect to the origin and
a = b = 1/2. Then
ρ(a∗) =(cid:20) (1 + a)(1 + b)
(1 − a)(1 − b)(cid:21)k−1
= 9k−1,
and therefore (W mf )(a∗) = 9(k−1)(m+1)f (a∗) → ∞ as m → ∞ when-
ever f (a∗) 6= 0. The latter means that if f (a∗) 6= 0, then the series
(4.16) diverges at a∗ and its uniform convergence on the entire sphere
fails. Below it will be shown that the uniform convergence of this series
fails for any a, b ∈ Bn+1.
the dynamics of involved reflections.
To understand the type of convergence, we need a deeper insight in
5.1. Dynamics of the Double Reflection Mapping T = τbτa. We
know that the trajectory {Tmx : m = 0, 1, 2, . . .} of any point x ∈
Sn \ {a∗} converges to the point b∗, which is the endpoint of the chord
containing a and b. Let us specify the character of this convergence.
a → Sn
a of b∗ and any compact set K ⊂ Sn
Lemma 5.1. The mapping T = τbτa maps the punctured sphere Sn
a =
Sn\{a∗} onto itself. The point b∗ is the attracting point of the dynamical
system T m : Sn
a uniformly on compact subsets, that is, for any
open neighborhood U ⊂ Sn
a there
exists m such that T mK ⊂ U for all m ≥ m.
Proof. The first statement is obvious, because Ta∗ = a∗ and T−1a∗ =
τaτba∗ = a∗. The second statement follows by a standard argument for
monotone pointwise convergence on compacts.
In fact, it suffices to
prove this statement for the sets U and K having the form
K = Kδ = Sn \ B(a∗, δ),
U = Uε = B(b∗, ε),
where B(a∗, ε) and B(b∗, δ) are geodesic balls in Sn of sufficiently small
radii.
The pointwise convergence yields that for any fixed x0 ∈ Kδ there
exists a number m0 such that Tm0x0 ∈ Uε. By continuity, the same is
true for every x in some neighborhood Vx0 of x0. Thus, the compact
Kδ is covered by open sets Vx, x ∈ Kδ, and therefore we can cover Kδ
14
M. AGRANOVSKY AND B. RUBIN
by a finite family {Vx1, . . . VxM}. For each xi, there is a number mi such
that Tmixi ∈ Uε. Setting m = max{m1, ..., mM}, we have
TmKδ ⊂ Uε.
A simple geometric consideration shows that the sequence Tm+1Kδ
monotonically decreases, i.e., Tm+1Kδ ⊂ TmKδ. Hence TmKδ ⊂ Uε for
all m ≥ m.
5.2. Uniform Convergence on Compact Subsets of the Punc-
tured Sphere.
(cid:3)
Theorem 5.2. If f ∈ C(Sn), then the series (4.16) converges to f
uniformly on compact subsets of the punctured sphere Sn \ {a∗}.
Proof. Consider the remainder (W mf )(x) of the series (4.16). By (4.8),
m
(W mf )(x) = ωm(x) f (Tm+1x),
ωm(x) =
ρ(Tjx).
Yj=0
Because ρ(b∗) < 1 (see (4.15)), for a fixed γ satisfying ρ(b∗) < γ < 1,
there is an open neighborhood U ⊂ Sn \ {a∗} of the point b∗ such
that 0 < ρ(y) < γ for all y ∈ U. On the other hand, Lemma 5.1
says that there exists m such that TmKδ ⊂ U for m ≥ m and hence
0 < ρ(Tmx) < γ for all x ∈ Kδ and m ≥ m. Thus
ρ(Tjx)
m
ωm(x) ≤ γm−m max
x∈Kδ
Yj=0
for all m ≥ m and all x ∈ Kδ. It follows that ωm(x) → 0 as m →
∞ uniformly on Kδ. Since f (T mx) ≤ kfkC(S n), we conclude that
W mf → 0 uniformly on Kδ. This gives the result.
5.3. Lp-Convergence.
Lemma 5.3. The operators Wa, Wb, W = WaWb, and W = WbWa
are isometries of the space Lp0(Sn) with p0 = n/(k − 1).
Proof. The statement about Wa follows from (2.11), which reads
(cid:3)
ZSn (cid:0)ρa(x)(cid:1)p0f (τax)dx =ZSn
f (x) dx.
The equality holds for any f ∈ L1(Sn) and therefore, if f ∈ Lp0(Sn),
then, using f (x)p0 instead of f , we obtain kWafkp0 = kfkp0. The
statement for Wb follows analogously. The operators W and W are
also isometries, as the products of two isometries.
(cid:3)
NON-GEODESIC FUNK TRANSFORMS
15
Theorem 5.4. Let f ∈ C(Sn), p0 = n/(k − 1). The series (4.16) and
(4.17) converge to f in the norm of Lp(Sn) for any 1 ≤ p < p0. The
convergence to f fails in any space Lp(Sn) with p0 ≤ p ≤ ∞.
Proof. It is clear that f belongs to Lp(Sn) for any 1 ≤ p ≤ ∞. Fix
δ > 0 and consider the function W mf = (WaWb)mf . Suppose that
p < p0 and set r = p0/p > 1. We write
kW mfkp
p =
ZB(a∗,δ)
(W mf )(x)p dx +ZKδ
(W mf )(x)pdx
= I1(m, δ) + I2(m, δ),
where, as above, Kδ = Sn \ B(a∗, δ). By Holder's inequality,
I1(m, δ) ≤ ZB(a∗,δ) (cid:16)(W mf )(x)p(cid:17)r
dx!1/r ZB(a∗,δ)
dx!r/(r−1)
(5.1)
.
Owing to Lemma 5.3, the operator W m preserves the Lp0-norm, and
therefore
ZB(a∗,δ) (cid:16)(W mf )(x)p(cid:17)r
p0 = kfkp
≤ kW mfkp0/r
p0.
Hence
dx!1/r
= ZB(a∗,δ)
(W mf )(x)p0dx!1/r
(5.2)
where A(δ) is the n-dimensional surface area of the geodesic ball B(a∗, δ).
For the second integral in (5.1) we have
I1(m, δ) ≤ A(δ)r/(r−1)kfkp
p0,
I2(m, δ) < σn sup
x∈Kδ (W mf )(x)p,
(5.3)
where σn is the area of the unit sphere Sn.
Now we fix sufficiently small ε > 0. Using (5.2), let us choose δ > 0
so that I1(m, δ) < ε/2 for all m ≥ 0. By Theorem 5.2, the inequality
(5.3) implies that there exists m = m(δ) such that I2(m, δ) < ε/2 for
all m ≥ m. Hence, by (5.1), kW mfkp
p < ε for m ≥ m, and therefore
W mf tends to 0 as m → ∞ in the Lp-norm. The latter gives the
desired convergence of the series (4.16).
On the other hand, if p > p0, then, by Holder's inequality, kfkp0 =
kW mfkp0 ≤ c kW mfkp, c = const > 0. It follows that the Lp-norm of
the remainder W mf of the series does not tend to 0 as m → ∞, unless
f = 0.
The proof for W f is similar.
(cid:3)
16
M. AGRANOVSKY AND B. RUBIN
Remark 5.5. As we can see, the iterative method in terms of the series
(4.16) and (4.17) does not provide uniformly convergent reconstruction
of continuous functions. The reconstruction is guaranteed only in the
Lp-norm with 1 ≤ p < p0 = n/(k − 1). For instance, in the case of
the hyperplane sections, when k = n and p0 = 1 + 1/(n − 1), the L2-
convergence fails because p0 does not exceed 2. The less the dimension
k is, the greater exponent p can be chosen. The case p = 1 works for
all 1 < k ≤ n.
6. Proof of Theorem 3.1
We recall that a ∈ Bn+1, sa =p1 − a2, and ϕa is an automorphism
(2.1). The following lemma allows us to exploit the language of Stiefel
manifolds when dealing with affine planes.
Lemma 6.1. Let 1 ≤ k ≤ n. The map ϕa extends as a bijection from
Gra(n + 1, k) onto Gro(n + 1, k). Specifically, if τ ∈ Gra(n + 1, k) is
defined by
τ = {x ∈ Rn+1 : ξ′x = ξ′a},
ξ ∈ St(n+1, n+1−k),
then ζ ≡ ϕaτ ∈ Gro(n + 1, k) has the form
ζ = {y ∈ Rn+1 : η′y = 0},
η ∈ St(n+1, n+1−k),
where
(6.1)
(6.2)
η = −(Aξ) α−1/2,
(6.3)
Conversely, if ζ ∈ Gro(n + 1, k) is defined by (6.2), then τ ≡ ϕaζ
α = (Aξ)′(Aξ).
A = saPa + Qa,
has the form (6.1) with
ξ = (A1η) β−1/2,
A1 = Pa + saQa,
β = (A1η)′(A1η).
(6.4)
Proof. Let τ ∈ Gra(n + 1, k) be defined by (6.1). Then
ζ ≡ ϕaτ = {y ∈ Rn+1 : ξ′(ϕay − a) = 0}.
By (2.1),
ϕay − a =
saAy
1 − y · a
.
(6.5)
Hence ζ = {y ∈ Rn+1 : (Aξ)′y = 0}. Now (6.2) follows if we represent
the (n + 1) × (n + 1 − k) matrix Aξ in the polar form
Aξ = η α1/2,
α = (Aξ)′(Aξ),
η = (Aξ) α−1/2;
(6.6)
see, e.g., [11, pp. 66, 591].
Conversely, let ζ ∈ Gro(n + 1, k) be defined by (6.2). Then
τ ≡ ϕaζ = {x ∈ Rn+1 : η′ϕax = 0}.
NON-GEODESIC FUNK TRANSFORMS
17
By (2.1), the equality η′ϕax = 0 is equivalent to
(Paη + saQaη)′x = η′a or
(A1η)′x = η′a.
(6.7)
We write A1η in the form A1η = ξ β1/2 with β = (A1η)′(A1η) and
ξ = (A1η) β−1/2 ∈ St(n+1, n+1−k). Then (6.7) yields ξ′x = β−1/2η′a.
To complete the proof, it remains to note that β−1/2η′a = ξ′a. Indeed,
ξ′a = β−1/2(A1η)′a = β−1/2(Paη + saQaη)′a
= β−1/2(η′(Paa + saη′Qaa) = β−1/2η′a.
(cid:3)
Proof of the Theorem. The case k = 1 is almost obvious; cf. Remark
4.2. Assuming 1 < k ≤ n, let τ ∈ Gra(n + 1, k) have the form (6.1)
and write
(Faf )(τ ) ≡ (Faf )(ξ) =
f (x) dσ(x),
ξ ∈ St(n+1, n+1−k).
Z
{x∈Sn:ξ′(x−a)=0}
We make use of the standard approximation machinery. Given a suffi-
ciently small ε > 0, let
(Fa,εf )(ξ) =ZSn
f (x) ωε(ξ′(x − a)) dx,
(6.8)
where ωε is a smooth bump function supported on the ball in Rn+1−k of
radius ε with center at the origin, so that lim
for any function g which is continuous in a neighborhood of the origin.
ε→0Rt<ε ωε(t) g(t) dx = g(o)
STEP I. Let us show that
lim
ε→0
(Fa,εf )(ξ) = (1 − ξ′a2)−1/2(Faf )(ξ).
(6.9)
We pass to bispherical coordinates (see, e.g., [20, p. 31])
x =(cid:20) ϕ sin θ
ψ cos θ (cid:21) , ϕ ∈ Sn ∩ ξ⊥, ψ ∈ Sn ∩ {ξ},
0≤ θ≤ π/2, (6.10)
dx = sink−1 θ cosn−k θ dθdϕdψ,
18
M. AGRANOVSKY AND B. RUBIN
and set s = cos θ. This gives
1
Z0
dϕ
sn−k(1 − s2)(k−2)/2ds ZSn∩ξ⊥
(cid:21)(cid:19) ωε(sψ − ξ′a) dψ
(Fa,εf )(ξ) =
sψ
f(cid:18)(cid:20) ϕ √1 − s2
H(y) ωε(y − ξ′a) dy,
× ZSn∩{ξ}
= ZRn−k+1
(6.11)
where
H(y) = (1−y2)(k−2)/2 ZSn∩ξ⊥
f(cid:18)(cid:20) ϕp1−y2
y
(cid:21)(cid:19) dϕ
if y ≤ 1 and H(y) = 0, otherwise. Passing to the limit, we obtain
(Fa,εf )(ξ) = H(ξ′a),
lim
ε→0
where
f(cid:18)(cid:20) ϕp1−ξ′a2
H(ξ′a) = (1−ξ′a2)(k−2)/2 ZSn∩ξ⊥
If the argument of f is denoted by x, then x − a lies in the subspace
perpendicular to ξ. Further, the integration in (6.12) is performed over
the (k − 1)-dimensional sphere of radius p1−ξ′a2. Switching to the
surface area measure, we can write (6.12) as
(cid:21)(cid:19) dϕ.
(6.12)
ξ′a
H(ξ′a) = (1−ξ′a2)−1/2
Z
{x∈Sn: ξ′(x−a)=0}
f (x) dσ(x),
as desired.
STEP II. Let us obtain an alternative expression for the limit (6.9),
now in terms of the automorphism ϕa. By Lemma 2.2,
(Fa,εf )(ξ) = sn
aZSn
(f ◦ ϕa)(y)
(1 − a · y)n ωε(ξ′[ϕay − a]) dy,
where
ξ′[ϕay − a] = −
saξ′Ay
1 − a · y
= −
sa(Aξ)′y
1 − a · y
,
A = saPa + Qa;
NON-GEODESIC FUNK TRANSFORMS
19
see (6.5). Denote
Then
f (y) = sn
a
(f ◦ ϕa)(y)
(1 − a · y)n .
f (y) ωε(cid:18)sa(Aξ)′y
1 − a · y(cid:19) dy.
(Fa,εf )(ξ) =ZSn
As in (6.6), the polar decomposition yields
Aξ = η α1/2,
α = (Aξ)′(Aξ),
η = (Aξ) α−1/2.
(6.13)
Then we pass to bispherical coordinates (cf. (6.10))
y =(cid:20) ϕ sin θ
ψ cos θ (cid:21) , ϕ ∈ Sn ∩ η⊥, ψ ∈ Sn ∩ {η},
0≤ θ≤ π/2, ,
dy = sink−1 θ cosn−k θ dθdϕdψ,
and set s = cos θ. This gives
sa α1/2sψ
1 − a · (√1 − s2 ϕ + sψ)(cid:19) dψ,
(Fa,εf )(ξ) =
1
Z0
sn−k(1 − s2)(k−2)/2ds ZSn∩η⊥
dϕ
sψ
(cid:21)(cid:19) ωε(cid:18)
f(cid:18)(cid:20) √1 − s2ϕ
× ZSn∩{η}
or (set z = sψ ∈ {η} ∼ Rn+1−k, z < 1)
(Fa,εf )(ξ) = Zz<1
× ZSn∩η⊥
(1 − z2)(k−2)/2dz
(cid:21)(cid:19) ωε
f(cid:18)(cid:20) p1 − z2ϕ
z
sa α1/2z
1 − a · (p1 − z2 ϕ + z)! dϕ,
sa α1/2z
=
Λz
,
(6.14)
We set
t ≡ t(z) =
Λ = sa α1/2,
1 − a · (p1 − z2 ϕ + z)
1 − h(z)
h(z) = a · (p1 − z2 ϕ + z),
so that t = o if and only if z = o, where o is the origin in the corre-
sponding space. Further, we write (6.14) as
Φ(t, z) ≡ Λz − t + th(z) = 0
20
M. AGRANOVSKY AND B. RUBIN
and denote m = n+1−k. Because the m×m matrix (∂Φi/∂zj)(o, o) =
Λ is invertible, there exists an inverse function z = z(t), which is well-
defined and differentiable in a small neighborhood of t = 0. Hence, for
sufficiently small ε > 0,
(1 − z(t)2)(k−2)/2ωε(t) det(z′(t)) dt
(Fa,εf )(ξ) = Zt<ε
× ZSn∩η⊥
z′(t) = −(cid:20)∂Φ(t, z)
∂z
z(t)
f(cid:18)(cid:20) p1 − z(t)2ϕ
(cid:21)(cid:19) dϕ,
(cid:21)−1 ∂Φ(t, z)
(Fa,εf )(ξ) =det(z′(o)) ZSn∩η⊥
∂t
lim
ε→0
where
Passing to the limit, we obtain
,
z = z(t).
f (ϕ) dϕ,
where
z′(o) = (1 − a · ϕ)Λ−1 = s−1
This gives
lim
ε→0
(Fa,εf )(ξ) =
Note that
α = (Aξ)′(Aξ).
a (1 − a · ϕ) α−1/2,
det(α)1/2 ZSn∩η⊥
sk−1
a
(f ◦ ϕa)(ϕ)
(1 − a · ϕ)k−1 dϕ.
α = (Aξ)′(Aξ) = (saPaξ + Qaξ)′(saPaξ + Qaξ)
= In+1−k − ξ′aa′ξ,
and therefore det(α) = det(In+1−k − ξ′aa′ξ). The last expression can
be transformed by making use of the known fact from Algebra (see,
e.g., [11, Theorem A3.5]). Specifically, if U and V are m× n and n× m
matrices, respectively, then
(6.15)
By this formula, det(α) = 1 − (a′ξ)(ξ′a) = 1 − ξ′a2. Thus, changing
notation, as in (3.2), we have
det(Im + UV) = det(In + VU).
lim
ε→0
(Fa,εf )(ξ) = (1 − ξ′a2)−1/2 ZSn∩η⊥
where η = (Aξ) α−1/2; cf. (6.13).
(Maf )(y) dσ(y),
(6.16)
NON-GEODESIC FUNK TRANSFORMS
21
STEP III. Comparing (6.9) with (6.16) and switching backward to
the Grassmannian language (use Lemma 6.1), we obtain the statement
of the theorem.
References
1. A. Beardon The Geometry of Discrete Groups, Graduate Texts in Mathemat-
ics, Springer-Verlag, New York Inc., 1983.
2. P.G. Funk, Uber Flachen mit lauter geschlossenen geodatischen Linien", Thesis,
Georg-August-Universitat Gottingen, 1911.
3. P.G. Funk, Uber Flachen mit lauter geschlossenen geodatschen Linen. Math.
Ann., 74 (1913), 278 -- 300.
4. R.J. Gardner, Geometric Tomography (second edition). Cambridge University
Press, New York, 2006.
5. F.W. Gehring, G.J. Martin, B.P. Palka, An Introduction to the Theory of
Higher-Dimensional Quasiconformal Mappings, AMS, Providence, Rhode Is-
land, 2017.
6. I.M. Gelfand, S.G. Gindikin, M.I. Graev. Selected Topics in Integral Geometry,
Translations of Mathematical Monographs, AMS, Providence, Rhode Island,
2003.
7. S. Gindikin, J. Reeds, L. Shepp, Spherical tomography and spherical integral
geometry. In Tomography, impedance imaging, and integral geometry (South
Hadley, MA, 1993), 83 -- 92, Lectures in Appl. Math., 30, Amer. Math. Soc.,
Providence, RI (1994).
8. S. Helgason, The totally geodesic Radon transform on constant curvature
spaces. Contemp. Math., 113 (1990), 141 -- 149.
9. S. Helgason,
Integral geometry and Radon transform. Springer, New York-
Dordrecht-Heidelberg-London, 2011.
10. H. Minkowski, Uber die Korper konstanter Breite [in Russian]. Mat. Sbornik.
25 (1904), 505 -- 508; German translation in Gesammelte Abhandlungen 2, Bd.
(Teubner, Leipzig, (1911), 277 -- 279.
11. R.J. Muirhead, Aspects of multivariate statistical theory, John Wiley & Sons.
Inc., New York, 1982.
12. G.D. Mostow, Quasi-conformal mappings in n-space and the rigidity of hyper-
bolic space forms. Inst. Hautes ´Etudes Sci. Publ. Math. 34 (1968), 53-104.
13. V. Palamodov. Reconstructive Integral Geometry. Monographs in Mathematics,
98. Birkhauser Verlag, Basel, 2004.
14. V.P. Palamodov, Reconstruction from Integral Data. Monographs and Research
Notes in Mathematics. CRC Press, Boca Raton, FL, 2016.
15. M. Quellmalz, A generalization of the Funk-Radon transform. Inverse Problems
33, no. 3, 035016, 26 pp. (2017).
16. M. Quellmalz, The Funk-Radon transform for hyperplane sections through a
common point, Preprint, arXiv:1810.08105 (2018).
17. B. Rubin, Generalized Minkowski-Funk transforms and small denominators on
the sphere. Fractional Calculus and Applied Analysis (2) 3 (2000), 177 -- 203.
18. B. Rubin, Inversion formulas for the spherical Radon transform and the gener-
alized cosine transform. Advances in Appl. Math., 29 (2002), 471 -- 497.
22
M. AGRANOVSKY AND B. RUBIN
19. B. Rubin, On the Funk-Radon-Helgason inversion method in integral geometry.
Contemp. Math., 599 (2013), 175 -- 198.
20. B. Rubin, Introduction to Radon transforms: With elements of fractional calcu-
lus and harmonic analysis (Encyclopedia of Mathematics and its Applications),
Cambridge University Press, 2015.
21. B. Rubin, Reconstruction of
in-
tegrals over hyperplane sections, Analysis and Mathematical Physics,
https://doi.org/10.1007/s13324-019-00290-1, 2019.
functions on the sphere from their
22. W. Rudin, Function Theory in the Unit Ball of C n, Springer, 1980.
23. Y. Salman, An inversion formula for the spherical transform in S 2 for a special
family of circles of integration. Anal. Math. Phys., 6, no. 1 (2016), 43 -- 58.
24. Y. Salman, Recovering functions defined on the unit sphere by integration on
a special family of sub-spheres. Anal. Math. Phys. 7, no. 2 (2017), 165 -- 185.
25. M. Stoll, Harmonic and subharmonic function theory on the hyperbolic ball.
LMS Lecture Notes in Mathematics, vol. 155, Cambridge University Press,
2016.
Department of Mathematics, Bar Ilan University, Ramat-Gan, 5290002,
and Holon Institute of Technology, Holon, 5810201, Israel
E-mail address: [email protected]
Department of Mathematics, Louisiana State University, Baton Rouge,
Louisiana 70803, USA
E-mail address: [email protected]
|
1809.06525 | 4 | 1809 | 2019-02-12T05:39:29 | Convergence analysis of a variable metric forward-backward splitting algorithm with applications | [
"math.FA"
] | The forward-backward splitting algorithm is a popular operator-splitting method for solving monotone inclusion of the sum of a maximal monotone operator and a cocoercive operator. In this paper, we present a new convergence analysis of a variable metric forward-backward splitting algorithm with extended relaxation parameters in real Hilbert spaces. We prove that this algorithm is weakly convergent when certain weak conditions are imposed upon the relaxation parameters. Consequently, we recover the forward-backward splitting algorithm with variable step sizes. As an application, we obtain a variable metric forward-backward splitting algorithm for solving the minimization problem of the sum of two convex functions, where one of them is differentiable with a Lipschitz continuous gradient. Furthermore, we discuss the applications of this algorithm to the fundamental of the variational inequalities problem, constrained convex minimization problem, and split feasibility problem. Numerical experimental results on LASSO problem in statistical learning demonstrate the effectiveness of the proposed iterative algorithm. | math.FA | math |
Convergence analysis of a variable metric forward-backward splitting
algorithm with applications
Fuying Cuia, Yuchao Tanga, Chuanxi Zhua
a Department of Mathematics, Nanchang University, Nanchang 330031, P.R. China
Abstract. The forward-backward splitting algorithm is a popular operator-splitting method for
solving monotone inclusion of the sum of a maximal monotone operator and a cocoercive operator.
In this paper, we present a new convergence analysis of a variable metric forward-backward splitting
algorithm with extended relaxation parameters in real Hilbert spaces. We prove that this algorithm
is weakly convergent when certain weak conditions are imposed upon the relaxation parameters.
Consequently, we recover the forward-backward splitting algorithm with variable step sizes. As
an application, we obtain a variable metric forward-backward splitting algorithm for solving the
minimization problem of the sum of two convex functions, where one of them is differentiable with
a Lipschitz continuous gradient. Furthermore, we discuss the applications of this algorithm to the
fundamental of the variational inequalities problem, constrained convex minimization problem, and
split feasibility problem. Numerical experimental results on LASSO problem in statistical learning
demonstrate the effectiveness of the proposed iterative algorithm.
Key words: Forward-backward splitting algorithm; Monotone inclusion; Variable metric; Split
feasibility problem.
AMS Subject Classification: 90C25; 47H05.
1
Introduction
Let H be a real Hilbert space. The forward-backward splitting algorithm is a classical operator-
splitting algorithm, which solves the monotone inclusion problem,
find x ∈ H,
such that 0 ∈ Ax + Bx,
(1.1)
where A : H → 2H is a maximal monotone operator and B : H → H is a β-cocoercive operator,
for some β > 0. The forward-backward splitting algorithm, which dates back to the original work of
Lions and Mercier [1], has been studied and reported extensively in the literature, for example [2 -- 6].
The emergence of compressive sensing theory and large-scale optimization problems associated with
signal and image processing has resulted in the forward-backward splitting algorithm receiving much
attention in recent years. A forward-backward splitting algorithm with relaxation and errors in
Hilbert spaces was proposed by Combettes [4]. More precisely, let x0 ∈ H, set
xk+1 = xk + λk(JγkA(xk − γk(Bxk + bk)) + ak − xk),
k ≥ 0,
(1.2)
1
where {γk} ⊂ (0, 2β), {λk} ⊂ (0, 1], {ak} and {bk} are absolutely summable sequences in H.
In
addition, JγkA := (I + γkA)−1 denotes the resolvent of operator A with index γk > 0. Combettes [4]
proved the convergence of the iterative scheme (1.2) when certain conditions are imposed upon
the parameters. Jiao and Wang [7] generalized the iterative scheme (1.2) by extending the work of
Combettes [4]. They proved the convergence of (1.2) by requiring the parameters {λk} to comply with
the requirement {λk} ⊂(cid:16)0,
is strictly larger than
one when {γk} ⊂ (0, 2β). Further, Combettes and Yamada [8] improved the range of the relaxation
parameters {λk} in (1.2) to (0, 4β−γk
.
Therefore, the range of {λk} in the work of Combettes and Yamada [8] is larger than that of Jiao
and Wang [7].
2β+γk(cid:17) when bk = 0. It is easy to see that
2β+γk
2β ). After a simple calculation, we know that 4β−γk
2β > 4β
2β+γk
4β
4β
In the case when γk = γ and ak = bk = 0, the iterative scheme (1.2) is reduced to the forward-
backward splitting algorithm with a constant step size [9],
xk+1 = xk + λk(JγA(xk − γBxk) − xk),
k ≥ 0,
(1.3)
where γ ∈ (0, 2β) and {λk} ⊂ (0, 4β−γ
2β ). Bauschke and Combettes [9] obtained the convergence of
the iterative algorithm (1.3) by adopting the Krasnosekii-Mann iteration for computing the fixed
points of nonexpansive operators. The forward-backward splitting algorithm with constant step size
(1.3) is usually considered to be stationary, whereas the forward-backward splitting algorithm with
variable step sizes (1.2) is referred to as non-stationary.
It is worth mentioning that by letting λk = 1, then (1.3) reduces to the classical forward-backward
splitting algorithm. More precisely, the iterative sequence {xk} is defined by
xk+1 = JγA(xk − γBxk),
k ≥ 0.
(1.4)
In the context of convex optimization, the forward-backward splitting algorithm is equivalent to
the so-called proximal gradient algorithm (PGA) applied to solve the following convex minimization
problem,
min
x∈H
f (x) + g(x),
(1.5)
where f : H → R is convex, differentiable with an L-Lipschitz continuous gradient for some L > 0
and g : H → (−∞, +∞] is a proper, lower semicontinuous, convex function. The convex optimization
problem (1.5) has found widespread application in signal and image processing, for example [10 -- 12].
As a consequence of [4], Combettes and Wajs [13] employed the forward-backward splitting algorithm
(1.2) to solve the minimization problem (1.5). The obtained iterative algorithm is defined as
xk+1 = xk + λk(proxγkg(xk − γk(∇f (xk) + bk)) + ak − xk),
k ≥ 0,
(1.6)
where {γk} ⊂ (0, 2/L), {λk} ⊂ (0, 1], and {ak}, {bk} are absolutely summable sequences in H.
proxγg denotes the proximity operator of g with index γ > 0.
In addition, Combettes and Wajs
[13] presented applications of this algorithm to many concrete convex optimization problems. This
iterative algorithm (1.6) was subsequently improved by Combettes and Yamada [8] who extended
the range of the relaxation parameters {λk}.
2
Inspired by solving large-scale convex optimization problems arising in image processing, machine
learning, and economic management, many efficient primal-dual splitting algorithms have been pro-
posed for structured monotone inclusions involving maximal monotone operators and single-valued
Lipschitz or cocoercive monotone operators, for example [14, 15]. Although these monotone inclu-
sions are more complicated than the monotone inclusion problem (1.1), they can be transformed into
the form of this problem in a suitable product space. Therefore, it is natural to consider using the
forward-backward splitting algorithm (e.g., (1.2) or (1.3)) to solve the equivalent monotone inclusion
problem. Because the backward steps cannot be decomposed, direct use of the forward-backward
splitting algorithm often fails to obtain a completely splitting algorithm. Many researchers attempted
to overcome this difficulty by investigating variable metric operator splitting algorithms. The use of
a suitable variable metric enables the implicit step of backward splitting to be easily decomposed.
For example, the primal-dual hybrid gradient algorithm [16] (also known as the primal-dual of the
Chambolle-Pock algorithm [17]) is equivalent to the variable metric proximal point algorithm [18,19].
We refer the readers to a subsequent paper [20] for more details. Vu [21] proposed a variable metric
extension of the forward-backward-forward splitting algorithm [3] for solving monotone inclusion of
the sum of a maximal monotone operator and a monotone Lipschitzian operator in Hilbert spaces.
Liang [22] proposed a variable metric multi-step inertial operator-splitting algorithm for solving the
monotone inclusion problem (1.1). Bonettini et al. [23] developed a scaled inertial forward-backward
splitting algorithm for solving (1.1) in the context of convex minimization. Neither of the respective
algorithms in the work by Liang [22] and Bonettini et al. [23] was compatible with the relaxation
strategy. The variable metric forward-backward splitting algorithm was originally studied in finite-
dimensional Hilbert spaces [2, 24]; however, the methods in these studies either had to be strongly
monotone to study the convergence rate or they did not make use of the cocoercive property of B
in (1.1). For infinite-dimensional Hilbert spaces, Combettes and Vu [25] proposed a variable metric
forward-backward splitting algorithm to solve (1.1) and analyzed its weak and strong convergence.
This algorithm is defined as follows. Let x0 ∈ H, and set
(1.7)
( yk = xk − γkUk(Bxk + bk),
xk+1 = xk + λk(JγkUkA(yk) + ak − xk).
where {Uk} ⊂ Pα(H), {λk} ⊂ (0, 1], {γk} ⊂ (0, 2β), {ak} and {bk} are absolutely summable sequences
in H. This algorithm (1.7) includes a variable metric, variable step sizes, relaxation parameter, and
errors. It includes nearly all of the forward-backward type of splitting algorithms mentioned above.
For example, by letting Uk = I in (1.7), it is reduced to (1.2). The relaxation parameters {λk} in
(1.2) are observed to be strictly larger than that based on the work of Combettes and Yamada [8].
While preparing this manuscript, we discovered that in Chapter 5 of the dissertation [26], Simoes
generalized the variable metric forward-backward splitting algorithm by replacing the relaxation
parameters {λk} in (1.7) with self-adjoint, strong positive linear operators. However, this approach
still requires the maximum eigenvalue of the operators to be smaller than one.
The purpose of this paper is to introduce a new convergence analysis for the variable metric
forward-backward splitting algorithm (1.7) with an extended range of relaxation parameters. We
prove the weak convergence of the variable metric forward-backward splitting algorithm by setting
3
the relaxation parameter {λk} larger than one in real Hilbert spaces. To achieve this goal, we make
full use of the averaged and firmly nonexpansive property of operators JγkUkA(I−γkUkB) and JγkUkA,
where λk > 0 and Uk ∈ Pα(H). In contrast, existing solutions mainly rely on JγkUkA being firmly
nonexpansive. Consequently, we obtain the convergence of the forward-backward splitting algorithm
with variable step sizes. Moreover, we impose a slightly weak condition on the relaxation parameters
to ensure the convergence of this algorithm. The results we obtained complement and extend those
of Combettes and Yamada [8]. As an application, we obtain the variable metric forward-backward
splitting algorithm for solving the minimization problem (1.5). We also present the application of
this algorithm to the variational inequalities problem, constrained convex minimization problem, and
split feasibility problem. To the best of our knowledge, the iterative algorithms we obtained are the
most general ones for solving these problems. Finally, we conduct numerical experiments on LASSO
problem to validate the effectiveness of the proposed iterative algorithm.
The remainder of this paper is organized as follows. Section 2 reviews selected notations and
lemmas on monotone operator theory and presents some technical lemmas. In Section 3, we prove the
main convergence results of the variable metric forward-backward splitting algorithm with relaxation
in real Hilbert spaces. Consequently, we obtain several corollaries of some special cases. Section
4 presents our use of the proposed iterative algorithm to solve three typical optimization problems
include the variational inequalities problem, constrained convex minimization problem, and split
feasibility problem. In Section 5, we present preliminary numerical results on LASSO problem to
illustrate the performance of the proposed iterative algorithm. Finally, we provide our conclusions.
2 Preliminaries
In this section, we recall selected concepts and lemmas that are commonly used in the context of
convex analysis and monotone operator theory. Most of them can be found in [9, 27]. Throughout
this paper, let H be a real Hilbert space. The inner product and the associated norm of Hilbert
space H are denoted by h·,·i and k · k, respectively. I denotes the identity operator and the symbols
⇀ and → denote weak and strong convergence.
We first recall selected basic notations and definitions. Let A : H → 2H be a set-valued operator.
We denote its domain, range, graph, and zeros by dom A = {x ∈ HAx 6= ∅}, ran A = {u ∈ H(∃x ∈
H)u ∈ Ax}, gra A = {(x, u) ∈ H × Hu ∈ Ax}, and zer A = {x ∈ H0 ∈ Ax}, respectively.
Definition 2.1. Let A : H → 2H be a set-valued operator. A is said to be monotone, if
hx − y, u − vi ≥ 0,
∀(x, u), (y, v) ∈ gra A.
Moreover, A is said to be maximal monotone, if its graph is not strictly contained in the graph of
any other monotone operator on H.
A well-known example of a maximal monotone operator is the subgradient mapping of a proper,
lower semicontinuous convex function f : H → (−∞, +∞] defined by
∂f : H → 2H : x 7→ {u ∈ Hf (y) ≥ f (x) + hu, y − xi,∀y ∈ H}.
4
Definition 2.2. Let A : H → 2H be a maximal monotone operator. The resolvent operator of A
with index λ > 0 is defined as
JλA = (I + λA)−1.
According to the Minty theorem, the resolvent operator JλA is defined everywhere on Hilbert
space H, and JλA is firmly nonexpansive.
Let us recall the definition of the proximity operator, which was first introduced by Moreau [28].
Let f ∈ Γ0(H), where Γ0(H) denotes the set of all proper lower semicontinuous convex functions
f : H → (−∞, +∞]. The proximity operator of f with index λ > 0 is defined by
proxλf : H → H : x 7→ arg min
y∈H
1
2ky − xk2 + λf (y).
In fact, the resolvent operator of the subdifferential operator of any f ∈ Γ0(H) with index λ > 0 is
the proximal operator of f with index λ > 0, that is
proxλf = (I + λ∂f )−1.
Therefore, the proximity operators have the same property as the resolvent operators.
Definition 2.3. Let B : H → H be a single-valued operator. Let β > 0, then B is said to be
β-cocoercive, if
hx − y, Bx − Byi ≥ βkBx − Byk2,
∀x, y ∈ H.
The β-cocoercive operator is also known as a β-inverse strongly monotone operator (β-ism), for
β -Lipschitz
example [29]. It is easy to see from the above definition that a β-cocoercive operator is 1
continuous, i.e., kBx − Byk ≤ 1
βkx − yk.
Next, we recall the definitions of nonexpansive and related mappings. These mappings often
appear in the convergence analysis of optimization algorithms.
Definition 2.4. Let C be a nonempty subset of H. Let T : C → H, then
(i) T is considered to be nonexpansive, if
(ii) T is considered to be firmly nonexpansive, if
kT x − T yk ≤ kx − yk,
∀x, y ∈ C.
kT x − T yk2 ≤ kx − yk2 − k(I − T )x − (I − T )yk2,
∀x, y ∈ C.
(iii) T is referred to as α-averaged, α ∈ (0, 1), if there exists a nonexpansive mapping S such that
T = (1 − α)I + αS.
It follows immediately that a firmly nonexpansive mapping is a nonexpansive mapping and an
α-averaged mapping is also nonexpansive.
We denote by F ix(T ) the set of fixed pints of a mapping T , that is F ix(T ) = {x ∈ Hx = T x}.
Lemma 2.1. (Demiclosedness Principle) Let C be a nonempty subset of H. Let T : C → H be a
nonexpansive mapping with F ix(T ) 6= ∅. If {xk} is a sequence in C that converges weakly to x and
if {(I − T )xk} converges strongly to y, then (I − T )x = y; in particular, if y = 0, then x ∈ F ix(T ).
5
The following proposition provides some equivalent definitions of the firmly nonexpansive map-
pings. This proposition can be found in Proposition 4.2 of [27].
Proposition 2.1. Let C be a nonempty subset of H. Let T : C → H, then the following are
equivalent
(i) T is firmly nonexpansive;
(ii) I − T is firmly nonexpansive;
(iii) 2T − I is nonexpansive;
(iv) hx − y, T x − T yi ≥ kT x − T yk2,
∀x, y ∈ C;
From Proposition 2.1 (iii) and (iv), we know that if T is firmly nonexpansive, then T is 1
2 -averaged,
and a 1-cocoercive operator is firmly nonexpansive.
The following proposition is taken from Proposition 4.25 of [27].
Proposition 2.2. Let C be a nonempty subset of H. Let T : C → H, then T is α-averaged if and
only if
kT x − T yk2 ≤ kx − yk2 −
.
1 − α
α k(I − T )x − (I − T )yk,
∀x, y ∈ C.
The following lemma provides a relation between an operator T with its complement I − T .
Lemma 2.2. Let C be a nonempty subset of H. Let T : C → H, then
(i) T is nonexpansive if and only if the complement I − T is 1
(ii) T is α-averaged if and only if the complement I − T is 1
2 -cocoercive;
2α -cocoercive.
We refer interested readers to [27] for further properties of nonexpansive, firmly nonexpansive,
and α-averaged nonlinear mappings.
We recall the results of the composition of two averaged operators. The following lemma first
appeared in [30] after which it was extended to a finite family of composition-averaged operators [8].
Lemma 2.3. Let C be a nonempty subset of H. Let T1 : C → H is α1-averaged and T2 : C → H is
α2-averaged. Then
T := T1T2 is
α1 + α2 − 2α1α2
1 − α1α2
− averaged.
Remark 2.1. (i) It is worth mentioning that two other results of the combination of averaged
-averaged.
operators were reported. From Proposition 4.32 of [27], T := T1T2 is α =
2
1
1+
max(α1,α2)
1−α1α2
α1+α2−2α1α2
From Byrne [29], T := T1T2 is bα = α1 + α2 − α1α2-averaged.
is smaller than the other two constants α and bα.
(ii) The constant bα is used in [7] to show the upper bound of the relaxation parameter λk such
It is not difficult to verify that
that λk < 1
bα .
We employ the following previously used notation [25]. Let B(H, G) be the spaces of bounded
linear operators from Hilbert space H to Hilbert space G. The norm of L ∈ B(H, G) is defined as
6
kLxk
kLk = supx∈H
the adjoint of L. The Loewner partial ordering on S(H) is defined by, for any U, V ∈ S(H),
kxk . We set B(H) = B(H, H) and S(H) = {L ∈ B(H)L = L∗}, where L∗ denotes
U (cid:23) V ⇔ hU x, xi ≥ hV x, xi,∀x ∈ H.
Let α ∈ [0, +∞), set
We denote √U as the square root of U ∈ Pα(H). Moreover, for every U ∈ Pα(H), we define a
Pα(H) = {U ∈ S(H)U (cid:23) αI}.
semi-scalar product and a semi-norm (a scalar product and a norm , if α > 0 by
(∀x ∈ H)(∀y ∈ H)
hx, yiU = hU x, yi and
kxkU =phU x, xi.
We borrow the following results on monotone operators in a variable metric setting from Com-
bettess work [25].
Lemma 2.4. Let A : H → 2H be maximal monotone, let α ∈ (0, +∞), let U ∈ Pα(H) and let HU −1
be the real Hilbert space with the scalar product hx, yiU −1 = hU −1x, yi,∀x, y ∈ H. Then the following
hold:
(i) U A : H → 2H is maximal monotone;
(ii) JU A : H → 2H is 1-cocoercive, i.e., firmly nonexpansive. More precisely,
U −1,
U −1 − k(I − JU A)x − (I − JU A)yk2
∀x, y ∈ H.
kJU Ax − JU Ayk2
U −1 ≤ kx − yk2
(2.1)
(iii) JU A = (U −1 + A)−1 ◦ U −1
Let U ∈ Pα(H) for some α > 0. The proximity operator of f ∈ Γ0(H) relative to the metric
induced by U is defined by
proxU
f : H → H : x 7→ arg min
2kx − yk2
y∈H(cid:18) 1
U + f (y)(cid:19) .
We have proxU
f = JU −f ∂f and we can write proxI
f = proxf .
We make full use of the following lemmas to obtain the weak convergence of the considered
In the following, we
+(N) the set of summable sequences in [0, +∞), where N is a set of nonnegative integer
iterative sequence. Both of the two lemmas were previously reported [31].
denote by ℓ1
numbers.
Lemma 2.5. Let α ∈ (0, +∞), and let {Wk} be in Pα(H), let C be a nonempty subset of H, and
let {xk} be a sequence in H such that
kxk+1 − zkWk+1 ≤ (1 + ηk)kxk − zkWk
+ ǫk,∀z ∈ C,
(2.2)
where {ηn} ⊂ ℓ1
converges.
+(N) and {ǫk} ⊂ ℓ1
+(N). Then {xk} is bounded and, for every z ∈ C, (kxk − zkWk
)
7
Lemma 2.6. Let α ∈ (0, +∞), and let {Wk} and W be in Pα(H) such that Wk → W pointwise as
k → +∞, as is the case when
supk∈N kWkk < +∞ and (∃{ηk} ⊂ ℓ1
+(N))(1 + ηk)Wk (cid:23) Wk+1.
Let C be a nonempty subset of H, and let {xk} be a sequence in H such that (2.2) is satisfied. Then
{xk} converges weakly to a point in C if and only if every weak sequential cluster point of {xk} is in
C.
The following lemma can be found in Corollary 2.14 in the book by Bauschke and Combettes [27].
Lemma 2.7. Let x ∈ H, y ∈ H, and α ∈ R. Then
kαx + (1 − α)yk2 = α kxk2 + (1 − α)kxk2 − α(1 − α)kx − yk2
(2.3)
3 Variable metric forward-backward splitting algorithm
In this section, we study the convergence of the variable metric forward-backward splitting algorithm.
First, we prove the following useful lemmas.
γkU k
Lemma 3.1. Let B : H → H be a β-cocoercive operator. Let α > 0, and let U ∈ Pα(H). Let HU −1
be a real Hilbert space with the scalar product hx, yiU −1 = hU −1x, yi,∀x, y ∈ H. Then I − γU B is a
2β -averaged operator on HU −1, for any γ ∈ (0, 2β
Proof. Let x, y ∈ H. Because B is β-cocoercive, we have
kU k ).
hU Bx − U By, x − yiU −1 = hBx − By, x − yi
≥ βkBx − Byk2.
On the other hand, we obtain
kU Bx − U Byk2
U −1 ≤ kUk · kBx − Byk2.
From (3.1) and (3.2), we obtain
hU Bx − U By, x − yiU −1 ≥
β
kUk · kU Bx − U Byk2
U −1,
(3.1)
(3.2)
(3.3)
which means that U B is
β
kU k -cocoercive on HU −1. Then γU Bx is
β
γkU k -cocoercive. By Lemma 2.2
2β -averaged operator on HU −1.
(ii), I − γU B is γkU k
Lemma 3.2. Let A : H → 2H be maximal monotone. Let α ∈ (0, +∞), and let U ∈ Pα(H). Let
HU −1 be a real Hilbert space with the scalar product hx, yiU −1 = hU −1x, yi,∀x, y ∈ H. Let B : H → H
be a β-cocoercive operator. Then, for any γ ∈ (0, 2β
4β−γkU k -averaged on HU −1.
kU k ), JγU A(I − γU B) is
2β
8
Proof. Because A is maximal monotone, then for any γ > 0, γU A is maximal monotone. According to
Lemma 2.4 (ii), JγU A is 1-cocoercive on HU −1. Then JγU A is 1
2 -averaged. Lemma 3.1 determines that
I−γU B is γkU k
2β -averaged. Therefore, we apply Lemma 2.3, from which we know that JγU A(I−γU B)
1
2 + γkU k
1 − 1
2β − γkU k
2 · γkU k
2β
2β
=
2β
4β − γkUk
,
(3.4)
is
α1 + α2 − 2α1α2
1 − α1α2
=
which is the averaged operator.
Lemma 3.3. Let H be a real Hilbert space. Let A : H → 2H be a maximal monotone operator. Let
B : H → H be a β-cocoercive operator, for some β > 0. Suppose that Ω :=zer(A + B) 6= ∅. Let
γk > 0, α > 0, and {Uk} ⊂ Pα(H). Then the following are equivalent:
(i) x∗ ∈ zer(A + B).
(ii) x∗ = JγkUkA(I − γkUkB)(x∗), for any γk > 0.
(iii) x∗ = ( U −1
k −γkB
k +γkA
)x∗.
α
)−1 ◦ ( U −1
α
Proof. (i)⇔(ii) Let x∗ ∈ zer(A + B), then we have
0 ∈ γkAx∗ + γkBx∗
⇔ 0 ∈ γkUkAx∗ + γkUkBx∗
⇔ x∗ − γkUkBx∗ ∈ x∗ + γkUkAx∗
⇔ x∗ = (I + γkUkA)−1(x∗ − γkUkBx∗)
⇔ x∗ = JγkUkA(I − γkUkB)(x∗)
(ii)⇔ (iii) Let x∗ = JγkUkA(I − γkUkB)x∗, then
x∗ − γkUkBx∗ ∈ x∗ + γkUkAx∗
k x∗ − γkBx∗ ∈ U −1
⇔ U −1
k x∗ + γkAx∗
U −1
U −1
k − γkB
k + γkA
⇔ (
α
U −1
⇔ x∗ = (
)x∗ ∈ (
k + γkA
α
U −1
k − γkB
)−1 ◦ (
)x∗
α
α
)x∗.
Lemma 3.4. Let H be a real Hilbert space. Let A : H → 2H be a maximal monotone operator. Let
B : H → H be a β-cocoercive operator, for some β > 0. Let r > 0 and s > 0, and let U, V ∈ Pα(H).
Define a variable metric forward-backward operator TrU := JrU A(I − rU B). Then, for any x ∈ H,
we have
r
s
where λmin(U −1) represents the minimum eigenvalue of U −1.
λmin(U −1)(cid:13)(cid:13)(cid:13)(cid:16)U −1 −
kTrU x − TsV xk ≤
V −1(cid:17) (x − TsV x)(cid:13)(cid:13)(cid:13) ,
1
Proof. Let x ∈ H, in which case we have
U −1x − U −1TrU x
r
− Bx ∈ ATrU x,
9
V −1x − V −1TsV x
s
− Bx ∈ ATsV x.
It follows from the monotonicity of operator A that
r
s
−
Then
U −1x − U −1TrU x
V −1x − V −1TsV x
(cid:28)TrU x − TsV x,
kTrU x − TsV xk2
(cid:29) ≥ 0.
s (cid:19) (x − TsV x)(cid:29) .
Because of the Cauchy-Schwarz inequality and the fact that λmin(U −1)kxk2 ≤ kxk2
we obtain
V −1(cid:17) (x − TsV x)(cid:13)(cid:13)(cid:13) .
U −1 ≤ r(cid:28)TrU x − TsV x,(cid:18) U −1
λmin(U −1)(cid:13)(cid:13)(cid:13)(cid:16)U −1 −
kTrU x − TsV xk ≤
r −
V −1
r
s
1
U −1 , for any x ∈ H,
We are ready to state our main theorems and present their convergence analysis.
Theorem 3.1. Let H be a real Hilbert space. Let A : H → 2H be maximal monotone. Let B : H → H
be β-cocoercive, for some β > 0. Suppose that Ω := zer(A + B) 6= ∅. Let α > 0, {ηk} ∈ ℓ1
+(N), and
{Uk} ⊂ Pα(H) such that
µ = sup
k∈N kUkk < +∞ and
(1 + ηk)Uk+1 (cid:23) Uk,
∀k ∈ N.
(3.5)
Let {γk} ⊂ (0, 2β
in H such that P+∞
kUkk ), and {λk} ⊂ (0, 1
k=0 λkkakk < +∞ andP+∞
αk
), where αk =
2β
4β−γkkUkk . Let {ak} and {bk} be two sequences
k=0 λkkbkk < +∞. Let x0 ∈ H, and set
(3.6)
(3.7)
(3.8)
( yk = xk − γkUk(Bxk + bk),
xk+1 = xk + λk(JγkUkA(yk) + ak − xk),
Then, we have
(i) For any x∗ ∈ Ω, limk→+∞ kxk − x∗kU −1
k
exists;
Suppose that 0 < λ ≤ λk ≤ 1
αk − τ , where τ ∈ (0, 1
αk − λ), then
(ii) limk→+∞ kxk − JγkUkA(xk − γkUkBxk)k = 0;
Suppose that 0 < γ ≤ γk, then
(iii) {xk} converges weakly to a point in Ω.
Further, suppose that γk ≤ 2β−ǫ
(iv) Bxk → Bx∗ as k → +∞, where x∗ ∈ Ω.
Proof. According to condition (3.5), we have
µ , where ǫ ∈ (0, 2β − µγ). Then
kU −1
k k ≤
1
α
Hence,
, U −1
k ∈ P 1
µ
(H),
and (1 + ηk)U −1
k (cid:23) U −1
k+1.
(1 + ηk)kxk2
k ≥ kxk2
U −1
U −1
k+1
,
∀x ∈ H.
10
For the sake of convenience, let
xk+1 = xk + λk(Jγk UkA(xk − γkUkBxk) − xk).
Then, iterative scheme (3.6) can be rewritten as
xk+1 = xk+1 + λkek,
(3.9)
(3.10)
where ek = JγkUkA(yk)− JγkUkA(xk − γkUkBxk) + ak such thatP+∞
JγkUkA is nonexpansive on HU −1
, we have
k
k=0 λkkekk < +∞. In fact, because
k
λkkekk ≤ √µλkkekkU −1
≤ √µλkkyk − (xk − γkUkBxk)kU −1
≤ µγkλkkbkk +r 1
λkkbkk +r 1
λkkakk.
λkkakk
≤ µ
2β
α
α
α
k
+ √µλkkakkU −1
k
(3.11)
Notice thatP+∞
k=0 λkkakk < +∞ andP+∞
From Lemma 3.2, we know that JγkUkA(I − γkUkB) is
k=0 λkkekk < +∞.
4β−γkkUkk ,
then there exist nonexpansive mappings Rk such that JγkUkA(I − γkUkB) = (1 − αk)I + αkRk.
Consequently, the iterative sequence {xk+1} in (3.9) is equivalent to
k=0 λkkbkk < +∞, (3.11) implies thatP+∞
4β−γkkUkk -averaged. Let αk =
2β
2β
xk+1 = (1 − λk)xk + λk((1 − αk)xk + αkRkxk)
= (1 − λkαk)xk + λkαkRkxk.
(3.12)
(i) Let x∗ ∈ zer(A + B), according to Lemma 3.3, x∗ = JγkUkA(I − γkUkB)(x∗). Then x∗ = Rkx∗.
From (3.8), (3.10), and (3.12), we obtain
kxk+1 − x∗kU −1
k
k
k
)
+ λkkekkU −1
k+1 ≤p(1 + ηk)kxk+1 − x∗kU −1
≤p(1 + ηk)(kxk+1 − x∗kU −1
≤p(1 + ηk)k(1 − λkαk)(xk − x∗) + λkαk(Rkxk − x∗)kU −1
+p(1 + ηk)r 1
≤ (1 + ηk)kxk − x∗kU −1
α λkkekk. Because P+∞
λkkekk
+ ǫk,
α
k
k
where ǫk = p(1 + ηk)q 1
k=0 λkkekk < +∞ and P+∞
P∞
k=0 kǫkk < +∞. On the basis of Lemma 2.5, we conclude that limk→+∞ kxk − x∗kU −1
Moreover, {kxk − x∗k} is bounded. Let M > 0 such that supk≥0 kxk − x∗k ≤ M .
(ii) With the help of the inequality kx + yk2 ≤ kxk2 + 2hy, x + yi, ∀x, y ∈ H. We obtain
k=0 kηkk < +∞, then
exists.
k
(3.13)
kxk+1 − x∗k2
k+1 ≤ (1 + ηk)kxk+1 − x∗k2
U −1
U −1
k
= (1 + ηk)kxk+1 − x∗ + λkekk2
U −1
k
11
= (1 + ηk)(kxk+1 − x∗k2
≤ (1 + ηk)kxk+1 − x∗k2
U −1
k
U −1
k
+ 2λkhek, xk+1 − x∗iU −1
+ 2(1 + ηk)MkU −1
k kλkkekk.
)
k
From Lemma 2.7 and (3.9) we derive that
kxk+1 − x∗k2
U −1
k
= k(1 − λk)(xk − x∗) + λk(JγkUkA(xk − γkUkBxk) − x∗)k2
= (1 − λk)kxk − x∗k2
− λk(1 − λk)kxk − JγkUkA(xk − γkUkBxk)k2
+ λk k(JγkUkA(xk − γkUkBxk) − x∗)k2
U −1
U −1
U −1
.
k
k
k
U −1
k
Because JγkUkA(I − γkUkB) is αk-averaged, it follows from Proposition 2.2 that
(3.14)
(3.15)
kJγkUkA(xk − γkUkBxk) − x∗k2
k ≤ kxk − x∗k2
U −1
k −
U −1
1 − αk
αk
Substituting (3.16) into (3.15) yields,
kxk − JγkUkA(xk − γkUkBxk)k2
.
U −1
k
(3.16)
k¯xk+1 − x∗k2
k ≤ kxk − x∗k2
U −1
k − λk(
U −1
Combining (3.17) with (3.14), we obtain
1
αk − λk)kxk − JγkUkA(xk − γkUkBxk)k2
U −1
k
.
(3.17)
kxk+1 − x∗k2
k+1 ≤ (1 + ηk)kxk − x∗k2
U −1
U −1
k
+ 2(1 + ηk)MkU −1
k kλkkekk
− (1 + ηk)λk(
which implies that
1
αk − λk)kxk − JγkUkA(xk − γkUkBxk)k2
U −1
k
,
(3.18)
1
αk − λk)kxk − JγkUkA(xk − γkUkBxk)k2
λk(
≤ (1 + ηk)kxk − x∗k2
Observe that limk→+∞ kxk − x∗kU −1
the above inequality and considering the condition on {λk}, we obtain
k − kxk+1 − x∗k2
exists and P+∞
U −1
k+1
U −1
U −1
k
k
+ 2(1 + ηk)MkU −1
k kλkkekk.
(3.19)
k=0 λkkekk < +∞. Then by letting k → +∞ in
k→+∞kxk − JγkUkA(xk − γkUkBxk)kU −1
lim
k
= 0.
(3.20)
Because the two norms k · kU −1
from (3.20) that
k
and k · k defined on the Hilbert spaces H are equivalent, it follows
k→+∞kxk − JγkUkA(xk − γkUkBxk)k = 0.
lim
(3.21)
(iii) In this part, we prove that the sequence {xk} converges weakly to a point in Ω. In fact, let
¯x be a weak sequential cluster point of {xk}, then there exists a subsequence {xkn} ⊂ {xk} such
that xkn ⇀ ¯x. Because {γk} ⊂ (γ, 2β
α ) is bounded, there exists a subsequence of {γk}
converges to γ ∈ (γ, 2β
α ). Without loss of generality, we may assume that γkn → γ. According to
condition (3.5), it follows from Lemma 2.6 that there exists U −1 ∈ P 1
k → U −1
pointwise.
kUkk ) ⊂ (γ, 2β
(H) such that U −1
µ
12
+
+
+
+
+
≤ kxkn − Jγkn Ukn A(xkn − γknUknBxkn)k
kn
µ
γ k(U −1
µ
γ k(U −1
kn
kn
γ − U −1
γkn)(xkn − JγU A(xkn − γU Bxkn))k
γkn − U −1γkn)(xkn − JγU A(xkn − γU Bxkn))k
≤ kxkn − Jγkn Ukn A(xkn − γknUknBxkn)k
µ
γαγ − γknkxkn − JγU A(xkn − γU Bxkn)k
2β
µ
α k(U −1
γ
kn − U −1)(xkn − JγU A(xkn − γU Bxkn))k.
With the help of Lemma 3.4, we make the following estimation,
kxkn − JγU A(xkn − γU Bxkn)k
≤ kxkn − Jγkn Ukn A(xkn − γknUknBxkn)k
+ kJγkn Ukn A(xkn − γknUknBxkn) − JγU A(xkn − γU Bxkn)k
≤ kxkn − Jγkn Ukn A(xkn − γknUknBxkn)k
1
k )k(cid:18)U −1
kn −
γkn
γ
U −1(cid:19) (xkn − JγU A(xkn − γU Bxkn))k
λmin(U −1
(3.22)
Because {kxkn − JγU A(xkn − γU Bxkn)k} is bounded, it follows from the conditions above, and we
can conclude from (3.22) that
kxkn − JγU A(xkn − γU Bxkn)k → 0 as kn → +∞.
(3.23)
As JγU A(I − γU B) is nonexpansive, based on the demiclosedness property of nonexpansive mapping,
we deduce that ¯x = JγU A(¯x − γU B ¯x), which means that ¯x ∈ zer (A + B). Because ¯x is arbitrary,
together with conclusion (i), we can conclude from Lemma 2.6 that {xk} converges weakly to a point
in zer(A + B).
(iv) On the other hand, as JγkUkA is firmly nonexpansive, it follows that we have
k
U −1
k
U −1
kJγkUkA(xk − γkUkBxk) − x∗k2
≤ kxk − γkUkBxk − (x∗ − γkUkBx∗)k2
− k(I − JγkUkA)(xk − γkUkBxk) − (I − JγkUkA)(x∗ − γkUkBx∗)k2
= kxk − x∗ − (γkUkBxk − γkUkBx∗)k2
− kxk − JγkUkA(xk − γkUkBxk) − (γkUkBxk − γkUkBx∗)k2
= kxk − x∗k2
+ kγkUkBxk − γkUkBx∗k2
− kxk − JγkUkA(xk − γkUkBxk) − (γkUkBxk − γkUkBx∗)k2
k − 2hxk − x∗, γkUkBxk − γkUkBx∗iU −1
U −1
U −1
U −1
U −1
U −1
.
k
k
k
k
k
U −1
k
Because B is β-cocoercive, we have that
hxk − x∗, γkUkBxk − γkUkBx∗iU −1
k ≥ γkβ kBxk − Bx∗k2 .
13
(3.24)
(3.25)
In addition, we have
kγkUkBxk − γkUkBx∗k2
U −1
k ≤ γ2
≤ µγ2
k kUkkkBxk − Bx∗k2
k kBxk − Bx∗k2 .
Substituting (3.25) and (3.26) into (3.24), we obtain
kJγkUkA(xk − γkUkBxk) − x∗k2
≤ kxk − x∗k2
− k(xk − JγkUkA(xk − γkUkBxk)) − (γkUkBxk − γkUkBx∗)k2
k − γk(2β − γkµ)kBxk − Bx∗k2
U −1
U −1
k
U −1
k
(3.26)
.
(3.27)
The combination of (3.27) with (3.15) yields
kxk+1 − x∗k2
k ≤ kxk − x∗k2
k − λkγk(2β − γkµ)kBxk − Bx∗k2
U −1
U −1
− λk kxk − JγkUkA(xk − γkUkBxk) − (γkUkBxk − γkUkBx∗)k2
− λk(1 − λk)kxk − JγkUkA(xk − γkUkBxk)k2
U −1
.
k
U −1
k
(3.28)
Further, on the basis of (3.28) and (3.14), we obtain
kxk+1 − x∗k2
k+1 ≤ (1 + ηk)kxk − x∗k2
U −1
k − (1 + ηk)λkγk(2β − γkµ)kBxk − Bx∗k2
U −1
− (1 + ηk)λk kxk − JγkUkA(xk − γkUkBxk) − (γkUkBxk − γkUkBx∗)k2
− (1 + ηk)λk(1 − λk)kxk − JγkUkA(xk − γkUkBxk)k2
+ 2λk(1 + ηk)MkU −1
k kkekk,
U −1
k
U −1
k
(3.29)
(3.30)
which implies that
λkγk(2β − γkµ)kBxk − Bx∗k2 ≤ (1 + ηk)kxk − x∗k2
k − kxk+1 − x∗k2
U −1
U −1
k+1
− (1 + ηk)λk(1 − λk)kxk − JγkUkA(xk − γkUkBxk)k2
+ 2λk(1 + ηk)MkU −1
k kkekk.
U −1
k
By the conditions on {γk} and {λk}, and together with conclusions (i), (ii) and the fact that
k=0 λkkekk < +∞, letting k → +∞ in the above inequality, we obtain
P+∞
Bxk → Bx∗ as k → +∞.
(3.31)
This completes the proof.
Remark 3.1. Because the upper bound of the relaxation parameter {λk} in Theorem 3.1 is governed
by the averaged constant of the variable metric forward-backward operator, Theorem 3.1 provides a
larger selection of the relaxation parameter and errors than Theorem 4.1 of Combettes [25].
14
Remark 3.2. If we assume that λk ∈ (λ, 1], then we reaffirm the conclusion that P+∞
k=0 kBxk −
Bx∗k2 < +∞ as in Theorem 4.1 of the paper by Combettes [25]. In fact, from inequality (3.30), we
have
λγǫkBxk − Bx∗k2 ≤ λkγk(2β − γkµ)kBxk − Bx∗k2
≤ (1 + ηk)kxk − x∗k2
+ 2λk(1 + ηk)MkU −1
By summing the above inequality from zero to infinity, we have
k − kxk+1 − x∗k2
U −1
k kkekk.
U −1
k+1
λγǫ
+∞Xk=0
≤ kx0 − x∗k2
kBxk − Bx∗k2
+∞Xk=0
U −1
+
0
+
+∞Xk=0
2λk(1 + ηk)MkU −1
k kkekk,
ηk sup
k≥0 kxk − x∗k2
U −1
k
k=0 kBxk − Bx∗k2 < +∞.
which implies thatP+∞
Remark 3.3. In view of Theorem 3.1 (iii), the iterative sequence generated by (3.6) converges weakly
to a point in Ω. The strong convergence of {xk} requires xk → x∗, x∗ ∈ Ω. Similar to Theorem 4.1
of Combettes [25], we need to assume that one of the following conditions holds.
(i) lim inf k→+∞ dΩ(xk) = 0;
(ii) A or B is demiregular at every point in Ω;
(iii) intΩ 6= ∅ and there exists {vk} ∈ ℓ1
+(N) such that (1 + vk)Uk (cid:23) Uk+1.
Because the proof is the same as that of Combettes [25], we omit it here.
Next, we impose a slightly weaker condition on the iterative parameter λk than in Theorem 3.1
to ensure the weak convergence of the iterative sequence {xk}.
Theorem 3.2. Let H be a real Hilbert space. Let A : H → 2H be maximal monotone. Let B : H → H
be β-cocoercive, for some β > 0. Suppose that Ω := zer(A + B) 6= ∅. Let α > 0, {ηk} ∈ ℓ1
+(N) and
{Uk} ∈ Pα(H) such that
µ = sup
k∈N kUkk < +∞ and
(1 + ηk)Uk+1 (cid:23) Uk,
∀k ∈ N.
(3.32)
exists;
Suppose that
k=0 λk( 1
Let the iterative sequence {xk} be defined by (3.6). Then, we have
(i) For any x∗ ∈ Ω, limk→+∞ kxk − x∗kU −1
(a) P+∞
(b) 0 < γ ≤ γk ≤ 2β−ǫ
(c)P+∞
k=0 γk+1−γk < +∞,P+∞
k=0 γk+1kUk+1k−γkkUkk < +∞, andP+∞
αk − λk) = +∞, where αk =
µ , where ǫ ∈ (0, 2β − µγ);
4β−γkkUkk ;
for any x ∈ H.
2β
k
k=0 kU −1
k x−U −1
k+1xk < +∞,
15
Then,
(ii) limk→+∞ kxk − JγkUkA(xk − γkUkBxk)k = 0;
(iii) {xk} converges weakly to a point in Ω;
Further, suppose that λk ≥ λ > 0. Then
(iv) Bxk → Bx∗ as k → +∞, where x∗ ∈ Ω.
Proof. (i) Let x∗ ∈ Ω,
limk→+∞ kxk − x∗kU −1
k
(ii) From (3.19), we obtain
it follows from the same proof of Theorem 3.1 (i) and we know that
exists. Then, {kxk − x∗k} is bounded. Let M := supk≥0 kxk − x∗k.
αk − λk)kxk − JγkUkA(xk − γkUkBxk)k2
U −1
k
U −1
0
+
1
α
M 2
+∞Xk=0
ηk + 2
1
α
+∞Xk=0
(1 + ηk)M λkkekk.
(3.33)
1
λk(
+∞Xk=0
≤ kx0 − x∗k2
k=0 ηk < +∞ andP+∞
1
+∞Xk=0
BecauseP+∞
k=0 λkkekk < +∞, then
λk(
αk − λk)kxk − JγkUkA(xk − γkUkBxk)k2
U −1
k
< +∞.
(3.34)
Let Tk = JγkUkA(I−γkUkB). By condition (a), (3.34) implies that lim inf k→+∞ kxk − TkxkkU −1
0. Consequently, lim inf k→+∞ kxk − Tkxkk = 0. Because Tk is αk-averaged, where αk =
there exists nonexpansive mappings Rk on HU −1
RkxkkU −1
= 0. Next, we prove that limk→+∞ kxk − Rkxkk = 0.
such that Tk = (1−αk)I+αkRk. Then, lim inf k→+∞ kxk−
k
k
Using formulation (3.10) and the fact that Rk+1 is nonexpansive on HU −1
, we have
=
4β−γkkUkk ,
k
2β
k+1
kxk+1 − Rk+1xk+1kU −1
k+1
k+1
k+1
k+1
+ λkkekkU −1
(3.10)
= kxk+1 − Rk+1xk+1 + λkekkU −1
≤ kxk+1 − Rk+1xk+1kU −1
= k(1 − λkαk)xk + λkαkRkxk − Rk+1xk+1kU −1
= k(1 − λkαk)(xk − Rkxk) + Rkxk − Rk+1xk+1kU −1
≤ (1 − λkαk)kxk − RkxkkU −1
+ kRk+1xk − Rk+1xk+1kU −1
≤ (1 − λkαk)kxk − RkxkkU −1
+ kRkxk − Rk+1xkkU −1
+ λkkekkU −1
+ kRkxk − Rk+1xkkU −1
k+1
k+1
k+1
k+1
k+1
k+1
k+1
+ λkkekkU −1
k+1
+ λkkekkU −1
k+1
+ kxk − xk+1kU −1
k+1
k+1
+ λkkekkU −1
k+1
≤ kxk − RkxkkU −1
k+1
+ kRkxk − Rk+1xkkU −1
k+1
+ 2r 1
α
λkkekk.
(3.35)
On the other hand, using the relation Rk = (1 − 1
αk
)I + 1
αk
Tk and Lemma 3.4, we have
kRkxk − Rk+1xkkU −1
k+1
16
1
1
1
k+1
k+1
k+1
k+1
αk+1
αk+1
αk+1
αk+1
)xk −
1
Tk+1xk(cid:13)(cid:13)(cid:13)(cid:13)U −1
Tk+1xk(cid:13)(cid:13)(cid:13)(cid:13)U −1
Tkxk(cid:13)(cid:13)(cid:13)(cid:13)U −1
k+1(cid:17) +
k+1(cid:17) + 2r 1
k+1(cid:17)
k+1)(xk − Tk+1xk)(cid:13)(cid:13)
k+1(cid:17)
+ kTkxkkU −1
+ kTkxkkU −1
+ kTkxkkU −1
+ kTkxkkU −1
k+1
k+1
k+1
k+1
1
1
1
1
1
k+1
k+1
≤
αk+1
αk+1
1
αk
1
αk
1
αk
1
αk
)xk +
Tkxk −
Tkxk −
Tkxk −
Tkxk − (1 −
1
αk+1 −
1
αk+1 −
1
αk(cid:12)(cid:12)(cid:12)(cid:12)kxkkU −1
αk(cid:12)(cid:12)(cid:12)(cid:12)kxkkU −1
=(cid:13)(cid:13)(cid:13)(cid:13)(1 −
≤(cid:12)(cid:12)(cid:12)(cid:12)
+(cid:13)(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)(cid:13)
≤(cid:12)(cid:12)(cid:12)(cid:12)
+(cid:13)(cid:13)(cid:13)(cid:13)
Tk+1xk(cid:13)(cid:13)(cid:13)(cid:13)U −1
2β γkkUkk − γk+1kUk+1k(cid:16)kxkkU −1
2β γkkUkk − γk+1kUk+1k(cid:16)kxkkU −1
2β γkkUkk − γk+1kUk+1k(cid:16)kxkkU −1
+ 2r 1
k − γkU −1
2β γkkUkk − γk+1kUk+1k(cid:16)kxkkU −1
+ 2r 1
+ 2r 1
γk+1(cid:13)(cid:13)(γk+1U −1
γ γk+1 − γk(cid:13)(cid:13)U −1
α (cid:13)(cid:13)(U −1
k − U −1
≤
≤
µ
γ
≤
2β
α
α
α
µ
µ
1
1
k (xk − Tk+1xk)(cid:13)(cid:13)
k+1)(xk − Tk+1xk)(cid:13)(cid:13) .
1
αk+1kTkxk − Tk+1xkkU −1
k+1
αkTkxk − Tk+1xkk
The combination of (3.36) with (3.35) yields
kxk+1 − Rk+1xk+1kU −1
k+1
1
≤ kxk − RkxkkU −1
+
+ kTkxkkU −1
k+1(cid:17)
α
+ 2r 1
+ 2r 1
α
α
+ 2r 1
+ 2r 1
α
µ
k+1
k+1
2β
µ
γ
k − U −1
2β γkkUkk − γk+1kUk+1k(cid:16)kxkkU −1
k (xk − Tk+1xk)(cid:13)(cid:13)
k+1)(xk − Tk+1xk)(cid:13)(cid:13) + 2r 1
2β γkkUkk − γk+1kUk+1k(cid:16)kxkkU −1
k (xk − Tk+1xk)(cid:13)(cid:13)
k+1)(xk − Tk+1xk)(cid:13)(cid:13) + 2r 1
γ γk+1 − γk(cid:13)(cid:13)U −1
α (cid:13)(cid:13)(U −1
γ γk+1 − γk(cid:13)(cid:13)U −1
α (cid:13)(cid:13)(U −1
k − U −1
λkkekk.
λkkekk
µ
γ
2β
+
α
α
µ
1
k+1
≤ (1 + ηk)kxk − RkxkkU −1
k
With the help of Lemma 2.5, we can conclude from (3.37) that limk→+∞ kxk− RkxkkU −1
limk→+∞ kxk − Rkxkk = 0. As a consequence, limk→+∞ kxk − Tkxkk = 0.
k
(iii) and (iv) can be proven using the same proof as Theorem 3.1.
(3.36)
+ kTkxkkU −1
k+1(cid:17)
(3.37)
= 0. Hence,
17
Remark 3.4. In Theorem 3.2, we obtain the weak convergence of the iterative sequence generated
by (3.6) with a weaker condition on {λk} than Theorem 3.1.
In Theorems 3.1 and 3.2, let Uk = I, in which case we obtain the following corollary, which shows
the convergence of the forward-backward splitting algorithm with variable step sizes.
Corollary 3.3. Let H be a real Hilbert space. Let A : H → 2H be maximal monotone. Let B : H →
H be β-cocoercive, for some β > 0. Suppose that Ω = zer(A+B) 6= ∅. Let {γk} ⊂ (0, 2β), and {λk} ⊂
(0, 1
k=0 λkkakk < +∞
αk
. Let {ak} and {bk} be two sequences in H such that P+∞
), where αk = 2β
k=0 λkkbkk < +∞. Let x0 ∈ H, and set
4β−γk
andP+∞
( yk = xk − γk(Bxk + bk),
xk+1 = xk + λk(JγkA(yk) + ak − xk).
(3.38)
Then, we have
(i) for any x∗ ∈ Ω, limk→+∞ kxk − x∗k exists;
Suppose that
(a1) 0 < λ ≤ λk;
(a2) λk ≤ 1
(a3) 0 < γ ≤ γk;
(a4) γk ≤ 2β − ǫ, where ǫ ∈ (0, 2β − γ).
αk − τ , where τ ∈ (0, 1
αk − λ);
αk − λk) = +∞ andP+∞
k=0 λk( 1
(a5) P+∞
If the conditions of (a1)-(a2) or (a3)-(a5) hold, then we have
(ii) limk→+∞ kxk − JγkA(xk − γkBxk)k = 0;
If the conditions of (a1)-(a3) or (a3)-(a5) hold, then we have
k=0 γk+1 − γk < +∞.
(iii) {xk} converges weakly to a point in Ω;
If the conditions of (a1)-(a3) or (a1), (a3)-(a5) hold, then we have
(iv) Bxk → Bx∗ as k → +∞, where x∗ ∈ Ω.
Remark 3.5. Under the condition (a1)-(a3), Corollary 3.3 reaffirms Proposition 4.4 of Combettes
and Yamada [8]. In addition, we obtain the convergence of the iterative scheme (3.38) under the
condition (a3)-(a5), which provides a weaker assumption on the relaxation parameters λk than the
condition (a1) and (a2). Consequently, the obtained results improve and generalize Proposition 4.4
of Combettes and Yamada [8].
As an application of Theorems 3.1 and 3.2, we can obtain the following convergence results for
solving convex minimization problem (1.5).
Corollary 3.4. Let H be a real Hilbert space. Let g : H → (−∞, +∞] be a proper, lower semi-
continuous, convex function. Let f : H → R be convex and differentiable with a 1/β-Lipschitz
continuous gradient. Assume that Ω is the set of solutions of problem (1.5) and Ω 6= ∅. Let x0 ∈ H,
and set
yk = xk − γkUk(∇f (xk) + bk),
k
U −1
γkg (yk) + ak − xk),
xk+1 = xk + λk(prox
18
(3.39)
where {Uk}, {γk}, {λk}, {ak}, and {bk} satisfy the same conditions as in Theorem 3.1 or Theorem
3.2.
Then the following hold:
exists;
k
k
U −1
γkg (xk − γkUk∇f (xk))k = 0;
(i) For any x∗ ∈ Ω, limk→+∞ kxk − x∗kU −1
(ii) limk→+∞ kxk − prox
(iii) {xk} converges weakly to a point in Ω;
(iv) ∇f (xk) → ∇f (x∗) as k → +∞, where x∗ ∈ Ω.
Proof. Because f is convex differentiable, according to the Baillon-Haddad theorem, ∇f is β-cocoercive.
From the definition of the proximity operator on the Hilbert space HU −1, we know that
prox
U −1
γkg (u) = JγkUk∂g(u).
k
(3.40)
Set A = ∂g and B = ∇f in Theorem 3.1 or Theorem 3.2 and this enables us to confirm the conclusions
of Corollary 3.4.
4 Applications
In this section, we present our study of several applications of the variable metric forward-backward
splitting algorithm.
4.1 Application to variational inequality problem
Consider the following variational inequality problem (VIP):
find x∗ ∈ C,
such that hBx∗, y − x∗i ≥ 0,
∀y ∈ C,
where C is a nonempty closed convex subset of H, and B : H → H is a nonlinear operator.
Recall the indicator function δC, which is defined as
δC (x) =( 0,
x ∈ C
+ ∞, otherwise.
(4.1)
(4.2)
The proximal operator of δC is well known to be the metric projection on C, which is defined by
PC(x) = proxδC (x) = arg min
y∈C kx − yk.
The normal cone operator of C is NC, which is defined by
NC(x) =( {whw, y − xi ≤ 0,∀y ∈ C},
∅,
x ∈ C
otherwise.
Then, VIP (4.1) is equivalent to the following monotone inclusion problem:
0 ∈ Bx + NC(x).
19
(4.3)
(4.4)
Assuming that B is β-cocoercive, then (4.4) is a special case of the monotone inclusion problem (1.1).
Let A = NC, then we know that JγU A = P U −1
, for any γ > 0 and U ∈ Pα(H). The operator P U −1
denotes the projector onto a nonempty closed convex subset C of H relative to the norm k · kU −1.
More precisely,
C
C
P U −1
C
(x) = arg min
y∈C kx − ykU −1.
On the basis of Theorems 3.1 and 3.2, we obtain the following convergence theorem to solve the VIP
(4.1).
Theorem 4.1. Let H be a real Hilbert space. Let B : H → H be a β-cocoercive operator. We denote
by Ω the solution set of VIP (4.1) and assume that Ω 6= ∅. Let x0 ∈ H, set
yk = xk − γkUk(Bxk + bk),
xk+1 = xk + λk(P
U −1
C
k
(yk) + ak − xk),
(4.5)
where {Uk}, {γk}, {λk}, {ak}, and {bk} satisfy the same conditions as in Theorem 3.1 or Theorem
3.2.
Then the following hold:
(i) For any x∗ ∈ Ω, limk→+∞ kxk − x∗kU −1
(ii) limk→+∞ kxk − P
(iii) {xk} converges weakly to a point in Ω;
(iv) Bxk → Bx∗ as k → +∞, where x∗ ∈ Ω.
(xk − γkUkAxk)k = 0;
U −1
C
k
k
exists;
4.2 Application to constrained convex minimization problem
Consider the following constrained convex minimization problem:
min f (x)
s.t. x ∈ C,
(4.6)
where C is a nonempty closed convex subset of H, and f : H → R is a proper closed convex
differentiable function with a Lipschitz continuous gradient.
It follows from the definition of the indicator function that constrained convex minimization
problem (4.6) is equivalent to the following unconstrained minimization problem:
It is obvious that problem (4.7) is a special case of (1.5). Therefore, by taking g(x) = δC (x), we
obtain the following convergence theorem for solving constrained convex minimization problem (4.6).
min
x∈H
f (x) + δC (x).
(4.7)
Theorem 4.2. Let H be a real Hilbert space. Let f : H → R be a proper, closed convex function
such that f is differentiable with an L-Lipschitz continuous gradient. We denote by Ω the solution
set of the constrained convex minimization problem (4.1) and assume that Ω 6= ∅. Let x0 ∈ H, and
set
yk = xk − γkUk(∇f (xk) + bk),
xk+1 = xk + λk(P
U −1
C
k
(yk) + ak − xk),
20
(4.8)
where {Uk}, {γk}, {λk}, {ak}, and {bk} satisfy the same conditions as in Theorem 3.1 or Theorem
3.2.
Then the following hold:
(i) For any x∗ ∈ Ω, limk→+∞ kxk − x∗kU −1
(ii) limk→+∞ kxk − P
(iii) {xk} converges weakly to a point in Ω;
(iv) ∇f (xk) → ∇f (x∗) as k → +∞, where x∗ ∈ Ω.
(xk − γkUk∇f (xk))k = 0;
exists;
U −1
C
k
k
4.3 Application to split feasibility problem
Consider the split feasibility problem (SFP) as follows:
find x ∈ C,
such that Lx ∈ Q,
(4.9)
where C and Q are nonempty, closed convex subsets of Hilbert spaces H and G, respectively. L :
H → G is a bounded linear operator. SFP (4.9) was first introduced by Censor and Elfving [32] in
a finite dimensional Hilbert space and has since been extensively studied by many authors, see, for
example [33, 34] and references therein.
SFP (4.9) is closely related to constrained convex minimization problem (4.6). More precisely,
the corresponding constrained convex minimization problem of SFP (4.9) is,
min
x
1
2kx − PQ(Lx)k2
s.t. x ∈ C.
(4.10)
Let x∗ be a solution of SFP (4.9), then x∗ is a solution of (4.10). Conversely, let x∗ be a solution of
(4.10) and f (x) := 1
2kx − PQ(Lx)k2 = 0, then x∗ is a solution of SFP (4.9). Under the assumption
that the solution set of SFP (4.9) is nonempty, SFP (4.9) and constrained convex minimization
problem (4.10) are equivalent.
The function f (x) = 1
L∗(Lx − PQ(Lx)) is
(4.9).
2kx− PQ(Lx)k2 is convex differentiable and the gradient operator ∇f (x) =
1
kLk2 -cocoercive. Therefore, we obtain the following theorem for solving SFP
Theorem 4.3. Let H and G be real Hilbert spaces. Let L : H → G be a bounded linear operator.
Let C and Q be nonempty closed and convex subsets of H and G, respectively. We denote by Ω the
solution set of SFP (4.9) and assume that Ω 6= ∅. Let x0 ∈ H, and set
yk = xk − γkUk(L∗(Lxk − PQ(Lxk)) + bk),
xk+1 = xk + λk(P
U −1
C
k
(yk) + ak − xk),
(4.11)
where {Uk}, {γk}, {λk}, {ak}, and {bk} satisfy the same conditions as in Theorem 3.1 or Theorem
3.2.
Then the following hold:
(i) For any x∗ ∈ Ω, limk→+∞ kxk − x∗kU −1
k
exists;
21
U −1
C
k
(xk − γkUkL∗(Lxk − PQ(Lxk)))k = 0;
(ii) limk→+∞ kxk − P
(iii) {xk} converges weakly to a point in Ω;
(iv) L∗(Lxk − PQ(Lxk)) → L∗(Lx∗ − PQ(Lx∗)) as k → +∞, where x∗ ∈ Ω.
Remark 4.1. To the best of our knowledge, the proposed iterative algorithms (4.5), (4.8), and
(4.11) are the most general ones for solving variational inequality problem (4.1), constrained convex
minimization problem (4.6), and split feasibility problem (4.9), respectively. Most of the existing
algorithms [7, 29, 34 -- 36] are special cases of ours.
5 Numerical experiments
In this section, we apply the proposed iterative algorithm (3.39) to solve the famous LASSO problem
[37]. All the experiments are performed on a standard Lenovo Laptop with Intel (R) Core (TM)
i7-4712MQ 2.3 GHZ CPU and 4 GB RAM. We run the program with MATLAB 2014a.
Let's recall the LASSO problem:
min
x∈Rn
1
2kAx − bk2
2
s.t. kxk1 ≤ t,
(5.1)
where A ∈ Rm×n, b ∈ Rm and t > 0. Define C := {xkxk1 ≤ t}, by using the indicator function, we
see that (5.1) is equivalent to the following unconstrained optimization problem
min
x
1
2kAx − bk2
2 + δC (x),
(5.2)
which is a special case of the general optimization problem (1.5). Let f (x) = 1
2 and
g(x) = δC (x), then we can apply iterative algorithm (3.39) to solve (5.2). Notice that the gradient
of f (x) is ∇f (x) = AT (Ax − b) and the Lipschitz constant of ∇f is L := kAk2. Besides, the
proximity operator of indicator function δC(x) is the orthogonal projection onto the closed convex
set C. Although it has no closed-form solution, it can be calculated in a polynomial time.
2kAx − bk2
In the tests, the true signal x ∈ Rn has k non-zero elements, which is generated from uniform
distribution in the interval [−2, 2]. The system matrix A ∈ Rm×n is generated from standard
Gaussian distribution. The observed signal b is given by b = Ax. In the experiment, we set m = 240,
n = 1024 and k = 40. The stopping criterion is defined as,
kxk+1 − xkk2
kxkk2
≤ ε,
(5.3)
where ε > 0 is a small constant. We test the performance of the proposed iterative algorithm with
different choices of the step size γk and the relaxation parameter λk. For simplicity, we set them
as constant during the iteration process. According to Corollary 3.4, we know that γk ∈ (0, 2
L ) and
λk ∈ (0, 4−γkL
). The obtained numerical results are listed in Table 1, in which we report the number
of iterations ("Iter"), the objective function value ("Obj") and the error between the recovered signal
and the true signal ("Err"). We can see from Table 1 that when the step size γk is fixed, a large
relaxation parameter λk leads to a faster convergence. At the same time, the larger the step size,
the faster the algorithm converges.
2
22
Table 1: Numerical results for different choices of γk and λk for solving the LASSO problem (5.1)
ε = 10−6
Err
γk
λk
0.2
0.4
0.6
0.8
1
1.2
1.5
1.75
0.2
0.4
0.6
0.8
1
1.2
1.5
0.2
0.4
0.6
0.8
1
1.05
1
2L
1
L
1.9
L
Iter
16553
9457
6765
5319
4408
3777
3123
2736
9455
5319
3776
2955
2440
2085
1718
5550
3086
2178
1698
1398
1340
0.0246
0.0123
0.0082
0.0062
0.0049
0.0041
0.0033
0.0028
0.0123
0.0062
0.0041
0.0031
0.0025
0.0020
0.0016
0.0065
0.0032
0.0022
0.0016
0.0013
0.0012
Obj
0.0020
4.9063e − 4
2.1851e − 4
1.2296e − 4
7.8535e − 5
5.4500e − 5
3.4835e − 5
2.5671e − 5
4.9106e − 4
1.2280e − 4
5.4633e − 5
3.0621e − 5
1.9556e − 5
1.3549e − 5
8.6760e − 6
1.3624e − 4
3.4008e − 5
1.5100e − 5
8.4390e − 6
5.3842e − 6
4.8598e − 6
ε = 10−8
Err
2.4764e − 4
1.2383e − 4
8.2498e − 5
6.1880e − 5
4.9492e − 5
4.1228e − 5
3.2952e − 5
2.8267e − 5
1.2381e − 4
6.1913e − 5
4.1206e − 5
3.0909e − 5
2.4670e − 5
2.0554e − 5
1.6469e − 5
6.5144e − 5
3.2508e − 5
2.1657e − 5
1.6253e − 5
1.2973e − 5
1.2350e − 5
Obj
1.9778e − 7
4.9432e − 8
2.1935e − 8
1.2339e − 8
7.8918e − 9
5.4756e − 9
3.4975e − 9
2.5735e − 9
4.9417e − 8
1.2352e − 8
5.4698e − 9
3.0772e − 9
1.9601e − 9
1.3605e − 9
8.7329e − 10
1.3675e − 8
3.4038e − 9
1.5104e − 9
8.5059e − 10
5.4181e − 10
4.9099e − 10
Iter
32336
17357
12035
9272
7570
6412
5231
4543
17356
9271
6412
4931
4021
3402
2771
9711
5167
3565
2737
2229
2131
In order to more visualize the effect of iterative parameters on the value of the function, Figure 1
shows the objective function value against the number of iterations. Further, we plot the true signal
and the recovered signal in Figure 2 for the parameters of γk = 1.9
L , λk = 1.05 and the stopping
criterion ε = 10−8. We can see from Figure 2 that the true signal is successfully reconstructed.
l
e
u
a
v
n
o
i
t
c
n
u
f
e
v
i
t
c
e
b
O
j
6000
5000
4000
3000
2000
1000
0
100
(a)
λ
k = 0.2
λ
k = 0.4
λ
k = 0.6
λ
k = 0.8
λ
k = 1
λ
k = 1.2
λ
k = 1.5
λ
k = 1.75
5000
(b)
l
e
u
a
v
n
o
i
t
c
n
u
f
e
v
i
t
c
e
b
O
j
4000
3000
2000
1000
λ
k
λ
k
λ
k
λ
k
λ
k
λ
k
λ
k
= 0.2
= 0.4
= 0.6
= 0.8
= 1
= 1.2
= 1.5
5000
(c)
l
e
u
a
v
n
o
i
t
c
n
u
f
e
v
i
t
c
e
b
O
j
4000
3000
2000
1000
λ
k
λ
k
λ
k
λ
k
λ
k
λ
k
= 0.2
= 0.4
= 0.6
= 0.8
= 1
= 1.05
105
0
100
105
0
100
iteration
iteration
102
iteration
104
Figure 1: The objective function value against the number of iterations for the LASSO problem. (a)
γk = 1
L and (c) γk = 1.9
L .
2L , (b) γk = 1
6 Conclusions
In this paper, we proposed a new convergence analysis of the variable metric forward-backward
splitting algorithm (1.7) with extended relaxation parameters. Based on the averaged operator
JγkUkA(I − γkUkB) and the firmly nonexpansive JγkUkA on the Hilbert spaces HU −1
, we proved
k
23
l
e
u
a
V
l
a
n
g
S
i
2
1.5
1
0.5
0
−0.5
−1
−1.5
−2
0
true signal
recovered signal
200
400
600
Index
800
1000
1200
Figure 2: The recovered sparse signal versus the true k-sparse signal.
the weak convergence of this algorithm. Compared to existing work, we imposed a slightly weak
condition on the relaxation parameters to ensure the convergence of the forward-backward splitting
algorithm when using the variable metric and variable step sizes. Our results complemented and
extended the corresponding results of Combettes and Yamada [8]. Furthermore, we obtained several
general iterative algorithms for solving the variational inequality problem, the constrained convex
minimization problem, and the split feasibility problem, respectively. These results generalized and
improved the known results in the literature. Numerical experimental results on LASSO problem
showed that the step size γk and relaxation parameter λk had much impact on the convergence speed
of the proposed iterative algorithm. The larger the step size, the faster the algorithm converged.
The over-relaxation parameter λk (λk > 1) performed better than the under-relaxation parameter
λk (λk ≤ 1).
Acknowledgement
This work was supported by the National Natural Science Foundations of China (11661056, 11771198,
11401293), the Postdoctoral Research Foundation of China (2015M571989) and the Postdoctoral
Science Foundation of Jiangxi Province (2015KY51).
Conflict of interest
The authors declare no conflict of interest.
References
[1] P.L. Lions and B. Mercier. Splitting algorithms for the sum of two nonlinear operators. SIAM
J. Numer. Anal., 16(6):964 -- 979, 1979.
[2] George H.G. Chen and R.T. Rockafellar. Convergence rates in forward-backward splitting. SIAM
J. Optim., 7(2):421 -- 444, 1997.
24
[3] P. Tseng. A modified forward-backward splitting method for maximal monotone mappings.
SIAM J. Control Optim., 38(2):431 -- 446, 2000.
[4] P.L. Combettes. Solving monotone inclusions via compositions of nonexpansive averaged oper-
ators. Optimization, 53:475 -- 504, 2004.
[5] G. Lopez, V. Martin-Marquez, F. Wang, and H.K. Xu. Forward-backward splitting method
for accretive operators in banach spaces. Abstr. Appl. Anal., 2012(Article ID 109236):25 pages,
2012.
[6] C. Zong, Y.T. Tang, Y.J. Cho. Convergence analysis of an inexact three-operator splitting
algorithm. Symmetry, 10:563, 2018.
[7] H.W. Jiao and F.H. Wang. On an iterative method for finding a zero to the sum of two maximal
monotone operators. J. Appl. Math., 2014(414031):1 -- 5, 2014.
[8] P.L. Combettes and I. Yamada. Compositions and convex combinations of averaged nonexpan-
sive operators. J. Math. Anal. Appl., 425:55 -- 70, 2015.
[9] H.H. Bauschke and P.L. Combettes. Convex Analysis and Motonone Operator Theory in Hilbert
Spaces. Springer, London, second edition, 2017.
[10] A. Beck and M. Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse
problems. SIAM J. Imaging Sci., 2:183 -- 202, 2009.
[11] A. Beck and M. Teboulle. Fast gradient-based algorithms for constrained total variation image
denoising and deblurring problems. IEEE Trans. Image Process., 18(11):2419 -- 2434, 2009.
[12] J.F. Cai, E.J. Candes, and Z. Shen. A singular value thresholding algorithm for matrix comple-
tion. SIAM J. Optim., 20:1956 -- 1982, 2010.
[13] P.L. Combettes and V.R. Wajs. Signal recovery by proximal forward-backward splitting. Mul-
tiscale Model. Simul., 4:1168 -- 1200, 2005.
[14] P. L. Combettes and J.C. Pesquet. Primal-dual splitting algorithm for solving inclusions with
mixtures of composite, lipschitzian, and parallel-sum type monotone operators. Set-Valued Var.
Anal., 20(2):307 -- 330, 2012.
[15] B.C. Vu. A splitting algorithm for dual monotone inclusions involving cocoercive operators.
Adv. Comput. Math., 38:667 -- 681, 2013.
[16] E. Esser, X. Zhang, and T. Chan. A general framework for a class of first order primal-dual
algorithms for convex optimization in imaging science. SIAM J. Imaging Sci., 3(4):1015 -- 1046,
2010.
[17] A. Chambolle and T. Pock. A first-order primal-dual algorithm for convex problems with
applications to imaging. J. Math. Imaging Vision, 40(1):120 -- 145, 2011.
25
[18] J.V. Burke and M.J. Qian. A variable metric proximal point algorithm for monotone operators.
SIAM J. Control Optim., 37(2):353 -- 375, 1998.
[19] L.A. Parente, P.A. Lotito, and M.V. Solodov. A class of inexact variable metric proximal point
algorithms. SIAM J. Optim., 19(1):240 -- 260, 2008.
[20] B.S. He and X.M. Yuan. Convergence analysis of primal-dual algorithms for a saddle-point
problem: from contraction perspective. SIAM J. Imaging Sci., 5(1):119 -- 149, 2012.
[21] B.C. Vu. A variable metric extension of the forward-backward-forward algorithm for monotone
operators. Numer. Funct. Anal. Optim., 34(9):1050 -- 1065, 2013.
[22] J. Liang. Convergence rates of first-order operator splitting methods. PhD thesis, 2016.
[23] S. Bonettini, F. Porta, and V. Ruggiero. A variable metric forward-backward method with
extrapolation. SIAM J. Sci. Comput., 38(4):A2558 -- A2584, 2016.
[24] P.A. Lotito, L.A. Parente, and M.V. Solodov. A class of variable metric decomposition methods
for monotone variational inclusions. J. Convex Anal., 16:857 -- 880, 2009.
[25] P.L. Combettes and B.C. Vu. Variable metric forward-backward splitting with applications to
monotone inclusions in duality. Optimization, 63(9):1289 -- 1318, 2014.
[26] M. Simoes. On some aspects of inverse problems in image processing. PhD thesis, 2017.
[27] H.H. Bauschke and P.L. Combettes. Convex Analysis and Motonone Operator Theory in Hilbert
Spaces. Springer, London, 2011.
[28] J.J. Moreau. Fonctions convexes duales et points proximaux dans un espace hilbertien. C. R.
Acad. Sci., Paris Ser. A Math, 255:2897 -- 2899, 1962.
[29] C. Byrne. A unified treatment of some iterative algorithms in signal processing and image
reconstruction. Inverse Probl., 20(1):103 -- 120, 2004.
[30] N. Ogura and I. Yamada. Non-strictly convex minimization over the fixed point set of the
asymptotically shrinking nonexpansive mapping. Numer. Funct. Anal. Optim., 23:113 -- 137,
2002.
[31] P.L. Combettes and V.R. Wajs. Variable metric quasi-fejer monotonicity. Nonlinear Anal.,
78:17 -- 31, 2013.
[32] Y. Censor and T. Elfving. A multiprojection algorithm using bregman projections in a product
space. Numer. Algorithms, 8:221 -- 239, 1994.
[33] H.K. Xu. A variable krasnoselskii-mann algorithm and the multiple-set split feasibility problem.
Inverse Probl., 22:2021 -- 2034, 2006.
[34] H.K. Xu. Iterative methods for the split feasibility problem in infinite dimensional hilbert spaces.
Inverse Probl., 26:105018(17pp), 2010.
26
[35] Q. Yang and J. Zhao. Generalized km theorems and their applications. Inverse Probl., 22:833 --
844, 2006.
[36] H.K. Xu. Averaged mappings and the gradient-projection algorithm. J. Optim. Theory Appl.,
150:360 -- 378, 2011.
[37] R. Tibshirani. Regression shrinkage and selection via the LASSO. J. R. Stat. Soc. Ser. B Stat.
Methodol., 58:267 -- 288, 1996.
27
|
1402.2123 | 4 | 1402 | 2018-01-11T09:13:06 | On Qian's problem for $\mathcal{L}_{\infty}$-spaces | [
"math.FA"
] | In this paper we devote to study Qian's problem for $\mathcal{L}_{\infty}$-spaces. Firstly, a positive answer to Qian's problem for $C(K)$-spaces is given by the assumption that $K$ has the C$\check{e}$ch-Stone property. Secondly, we obtain quantitative characterizations of separably injective spaces that turn out to give a positive answer to Qian's problem of 1995 in the setting of separable universality. Thirdly, we prove a sharpen quantitative and generalized Sobczyk theorem, which gives sharpen constants ($\alpha,\gamma$) for Qian's Problem. Finally, we give a more generalized Figiel theorem for $\mathcal{L}_{\infty}$-spaces. | math.FA | math | On Qian's problem for L∞-spaces
Duanxu Dai
College of Mathematics and Computer Science
Quanzhou Normal University
Quanzhou 362000, China
E-mail: [email protected]
8
1
0
2
n
a
J
1
1
]
.
A
F
h
t
a
m
[
4
v
3
2
1
2
.
2
0
4
1
:
v
i
X
r
a
Abstract
In this paper we devote to study Qian's problem for L∞-spaces.
Firstly, a positive answer to Qian's problem for C(K)-spaces is given
by the assumption that K has the Cech-Stone property. Secondly,
we obtain quantitative characterizations of separably injective spaces
that turn out to give a positive answer to Qian's problem of 1995 in
the setting of separable universality. Thirdly, we prove a sharpen
quantitative and generalized Sobczyk theorem, which gives sharpen
constants (α, γ) for Qian's Problem. Finally, we give a more gener-
alized Figiel theorem for L∞-spaces.
1
Introduction
Mazur and Ulam [24] in 1932 proved that every surjective isometry between
two Banach spaces X and Y is necessarily affine. Since then, properties of
isometries and generalizations there of between Banach spaces has continued
for 86 years. On this period, many significant problems about perturbation
2010 Mathematics Subject Classification: Primary 46B04, 46B20; Secondary 46A22,
54C60.
Key words and phrases: ε-Isometry, Stability, Figiel theorem, Banach space.
Supported by the Natural Science Foundation of China (Grant No. 11601264)
and the Outstanding Youth Scientific Research Personnel Training Program of Fu-
jian Province and the High level Talents Innovation and Entrepreneurship Project of
Quanzhou City and the Research Foundation of Quanzhou Normal University(Grant
No. 2016YYKJ12).
1
2
D. Dai
properties of surjective ε-isometries were proposed and solved by numer-
ous mathematicians. In particular, we mention the Hyers-Ulam problem
[20] (see, for instance, [18], [19], and [25]). In 1968, Figiel [17] showed the
following remarkable result(Figiel theorem): For every standard isometry
f : X → Y there is a linear operator T : L(f ) → X with kT k = 1 so that
T f = Id on X, where L(f ) is the closure of span f (X) in Y (see also [7]
and [15]).
Definition 1.1. Let X, Y be two Banach spaces, ε ≥ 0, and let f : X → Y
be a mapping.
(1) f is said to be an ε-isometry if
(1.1)
kf (x) − f (y)k − kx − yk ≤ ε for all x, y ∈ X.
In particular, a 0-isometry f is simply called an isometry.
(2) We say an ε-isometry f is standard if f (0) = 0.
(3) A standard ε-isometry f is (α, γ)-stable if there exist α, γ > 0 and a
bounded linear operator T : L(f ) → X with kT k ≤ α such that
(1.2)
kT f (x) − xk ≤ γε,
for all x ∈ X.
In this case, we also simply say f is stable, if no confusion arises.
(4) A pair (X, Y ) of Banach spaces X and Y is said to be stable if every
standard ε-isometry f : X → Y is (α, γ)-stable for some α, γ > 0.
(5) A pair (X, Y ) of Banach spaces X and Y is called (α, γ)-stable for
some α, γ > 0 if every standard ε-isometry f : X → Y is (α, γ)-stable.
The study of non-surjective ε-isometries has also been considered (see, for
instance, [5], [10], [11], [13], [14], [25], [28], [30] and [32]). Qian[28] proposed
the following problem in 1995.
Problem 1.2. Is it true that for every pair (X, Y ) of Banach spaces X and
Y there exists γ > 0 such that every standard ε-isometry f : X → Y is
(α, γ)-stable for some α > 0?
However, Qian [28] presented a counterexample showing that if a sepa-
rable Banach space Y contains an uncomplemented closed subspace X then
for every ε > 0 there is a standard ε-isometry f : X → Y which is not
stable. Recently, Cheng et al [12] showed the following sharp weak stability
version.
On Qian's problem for L∞-spaces
3
Theorem 1.3 (Cheng et al). Let X and Y be Banach spaces, and let f :
X → Y be a standard ε-isometry for some ε ≥ 0. Then for every x∗ ∈ X ∗,
there exists φ ∈ Y ∗ with kφk = kx∗k ≡ r such that
hφ, f (x)i − hx∗, xi ≤ 2εr, f or all x ∈ X.
For study of the stability of ε-isometries of Banach spaces, the following
question was proposed in [11].
Problem 1.4. Is there a characterization for the class of Banach spaces X
satisfying given any X ∈ X and Banach space Y , the pair (X, Y ) is ((α, γ)-,
resp.) stable?
Every space X of this class is said to be a universally ((α, γ)-, resp.) left-
stable space. On one hand, Cheng, Dai, Dong et.al. [11] proved that every
injective Banach space is a universally left-stable space. On the other hand,
the first two authors Cheng and Dai, together with others [5] showed that
every universally left-stable space is just a cardinality injective Banach space
(i.e., a Banach space which is complemented in every superspace with the
same cardinality) and they also showed that a dual space is injective (i.e.,
X is complemented in ℓ∞(BX ∗))if and only if it is a universally left-stable
space.
This paper devotes to study Qian's problem for L∞-spaces as follows. In
Section 3, we obtain a weak positive answer to Qian's problem for C(K)-
spaces (see Corollary 3.5). Then by assuming that K has the Cech-Stone
property, a positive answer for such a C(K)-space is given. The following
Problem 1.5 is also very natural.
Problem 1.5. Is there a characterization for the class of Banach spaces S
satisfying given any X ∈ S and separable Banach space Y , the pair (X, Y )
is ((α, γ)-, resp.) stable?
Every space X of this class is said to be a separably universally (resp.
(α, γ)) left-stable space. In Section 4, we will show that all of these spaces
of the class S coincide with separably injective Banach spaces. We here
refer the reader to a very excellent paper [4] by Avil´es-S´anchez-Castillo-
Gonz´alez-Moreno for further information about injective Banach spaces and
separably injective Banach spaces where they resolved (under an additional
assumption) a long standing problem proposed by Lindenstrauss in the
middle sixties.
4
D. Dai
The following theorem was proved by Sobczyk [31] which says that c0
is separable separably injective space( A Banach space X is said to be λ-
separably injective if it has the following extension property: Every bounded
linear operator T from a closed subspace of a separable Banach space into
X can be extended to be a bounded operator on the whole space with its
norm at most λkT k. In this case, X is said to be separably injective if it
is λ-separably injective for some λ ≥ 1 [34]) while Zippin [34] showed that
c0 is the unique separable separably injective space, up to an isomorphism.
In 2014, Cheng, Dai et al [11] proved that c0, up to an isomorphism, is
the unique separable space such that the couple (c0, Y ) is stable for every
separable space Y .
Theorem 1.6 (Sobczyk theorem [31]). Let X be a separable Banach space.
If E is a closed subspace of X and T : E → c0 is a bounded operator then
there exists an operator eT : X → c0 such that eT E = T and keT k ≤ 2kT k.
In section 5, we also prove a sharpen quantitative and generalized Sobczyk
theorem (see Theorem 1.6), that is, Theorem 5.1, which gives examples of
nonseparable separably injective spaces X (but not injective, i.e., X is not
complemented in ℓ∞(BX ∗)) such that for some sharpen constants α, γ > 0,
the couple (X, Y ) is (α, γ)-stable for every separable space Y . In Section
6, we prove a more generalized Figiel theorem for L∞,λ-spaces ( see [3], [4],
[8]).
All symbols and notations in this paper are standard. We use X to denote
a real Banach space and X ∗ its dual. BX , ext (BX ∗) and SX denote the closed
unit ball of X, the set of all extremal points of BX ∗ and the unit sphere
of X, respectively. For a subset A ⊂ X, A and card (A) stand respectively
for the closure of A, the cardinality of A. Given a bounded linear operator
T : X → Y , T ∗ : Y ∗ → X ∗ stands for its conjugate operator. We denote by
d(X, Y ) = inf{kT k · kT −1k : T is an isomorphism between X and Y } the
Banach-Mazur distance between X and Y .
2 Preliminaries
Recall that a Banach space X is said to be λ-(resp. separably injective)
injective if it has the following extension property: Every bounded linear
operator T from a closed subspace of a (resp. separable) Banach space into
X can be extended to be a bounded operator on the whole space with its
norm at most λkT k (see, for instance, [1], [4], [16], [33], [34]). In this case,
On Qian's problem for L∞-spaces
5
X is said to be injective (resp. separably injective) if it is λ-(resp. separably
injective) injective for some λ ≥ 1.
The following Proposition 2.1 follows easily from Remark 2.3.
Proposition 2.1. A (resp. separable) Banach space X is λ-(resp. separa-
bly injective) injective if and only if it is λ-complemented in every (resp.
separable) superspace (i.e., a normed linear space which contains X).
The following Proposition 2.2 was proved by Avil´es, S´anchez, Castillo,
Gonz´alez and Moreno (see [4, Prop. 3.2]).
Proposition 2.2. (1) If a Banach space X is λ-separably injective, then it
is 3λ-complemented in every superspace Y such that Y /X is separable.
(2) If a Banach space X is λ-complemented in every superspace Y such
that Y /X is separable, then X is λ-separably injective.
Remark 2.3. For any set Γ, that ℓ∞(Γ) is 1-injective follows from the
Hahn-Banach theorem.
Recall that S is the class of Banach spaces satisfying given any X ∈ S
and separable Banach space Y , the pair (X, Y ) is ((α, γ)-, resp.) stable.
Every space X of this class is said to be a separably universally ((α, γ)-,
resp.) left-stable space. In section 3, we completely solve Problem 1.5. That
is, we prove that all of these spaces of the class S coincide with separably
injective Banach spaces.
Lemma 2.4. Suppose that X, Y are Banach spaces. Let ε ≥ 0. Assume that
f is a ε− isometry from X into Y with f (0) = 0. Then for every w∗-dense
subset Ω ⊂ ext (BX ∗) there is a bounded linear operator T : Y → ℓ∞(Ω)
such that
kT f (x) − xk ≤ 2ε,
for all x ∈ X.
Proof. By Theorem 1.3, for every x∗ ∈ Ω, there exists a functional Q(x∗) ∈
SY ∗ such that
hQ(x∗), f (x)i − hx∗, xi ≤ 2ε, f or all x ∈ X.
We now define a mapping T : Y → ℓ∞(Ω) by
T (y) = {Q(x∗)(y)}x∗∈Ω.
It is clear that T is a bounded linear operator with norm one and
kT f (x) − xk = sup
x∗∈Ω
Q(x∗)f (x) − x∗(x) ≤ 2ε,
for all x ∈ X.
6
D. Dai
The following Lemma 2.5 follows from Qian's counterexample in [28] (see
also [11]).
Lemma 2.5. Let X be a closed subspace of Banach space Y . If card (X) =
card (Y ), then for every ε > 0 and every bijective mapping g : X → BY
with g(0) = 0, there is a standard ε-isometry f : X → Y defined for all
x ∈ X by f (x) = x + ε
2g(x) such that
(1) L(f ) ≡ span f (X) = Y ;
(2) X is λ complemented in Y whenever f is (λ, γ) stable for some λ,
γ > 0.
3 On Qian's problem for C(K)-spaces
Recall that a dual Banach space Y ∗ is said to have the point of weak star to
norm continuity property (in short, w∗-PCP) if every nonempty bounded
subset of Y ∗ admits relative weak star neighborhoods of arbitrarily small
norm diameter. For example, if Y is an Asplund space, then Y ∗ has the
w∗-PCP (see, for instance, [26]).
Recall that a set valued mapping F : X → 2Y is said to be usco pro-
vided it is nonempty compact valued and upper semicontinuous, i.e., F (x)
is nonempty compact for each x ∈ X and {x ∈ X : F (x) ⊂ U} is open in X
whenever U is open in Y . We say that F is usco at x ∈ X if F is nonempty
compact valued and upper semicontinuous at x, i.e., for every open set V
of Y containing F (x) there exists a open neighborhood U of X such that
F (U) ⊂ V . Therefore, F is usco if and only if F is usco at each x ∈ X.
Recall that a mapping ϕ : X → Y is called a selection of F if ϕ(x) ∈ F (x)
for each x ∈ X, moreover, we say ϕ is a continuous (linear) selection of
F if ϕ is a continuous (linear) mapping. We denote the graph of F by
G(F ) ≡ {(x, y) ∈ X × Y : y ∈ F (x)}, we write F1 ⊂ F2 if G(F1) ⊂ G(F2).
A usco mapping F is said to be minimal if E = F whenever E is a usco
mapping and E ⊂ F (see, for instance, [13], [26, page 19, 102-109]).
The following Problem 3.1 is equivalent to Problem 1.2.
Problem 3.1. Does there exist a constant γ > 0 depending only on X and
Y with the following property: For each ε-isometry f : X → Y with f (0) = 0
there is a w∗ − w∗ continuous linear selection Q of the set-valued mapping
Φ from X ∗ into 2L(f )∗ defined by
Φ(x∗) = {φ ∈ L(f )∗ : hφ, f (x)i − hx∗, xi ≤ γkx∗kε,
for all x ∈ X},
where L(f ) = span f (X)?
On Qian's problem for L∞-spaces
7
The following Lemma 3.2 was motivated by Dai et.al. in [13, Lemma
4.2]. By an analogous argument we conclude the result on w∗ − w∗ usco
mappings, which will be used to prove Corollary 3.5.
Lemma 3.2. Suppose that X, Y are Banach spaces. Let ε ≥ 0. Assume
that f is a ε− isometry from X into Y with f (0) = 0, H itself is a Baire
subspace contained in SX ∗ with respect to w∗-topology. If we define a set-
valued mapping Φ1 : SX ∗ → 2SL(f )∗ by
Φ1(x∗) = {φ ∈ SL(f )∗ : hφ, f (x)i − hx∗, xi ≤ 4ε,
for all x ∈ X},
where L(f ) = span f (X), then
(i) Φ1 is convex w∗-usco at each point of SX ∗.
(ii) There exists a minimal convex w∗ − w∗ usco mapping contained in
Φ1.
(iii) If, in addition, Y ∗ has the w∗-PCP (especially, if Y is an Asplund
space) or Y is separable, then there exists a selection Q of Φ1 such that Q
is w∗ − w∗ continuous on a w∗-dense Gδ subset of H.
Proof. (i) It follows easily from [13, Lemma 4.2 (i)].
(ii)By Zorn Lemma (see [13, Lemma 4.2 (ii)] or [26, Prop.7.3, p.103])
there exists a minimal convex w∗ − w∗ usco mapping contained in Φ1.
(iii) By (ii) there is a minimal convex w∗ −w∗ usco mapping F ⊂ Φ1, and
H itself is a Baire space with respect to w∗-topology, and Y ∗ has the w∗-
PCP (especially, if Y is an Asplund space) or Y is separable, which follows
easily from [26, Lemma 7.14, p.106-107] and [13, Lemma 4.2 (iii)].
Remark 3.3. The above Lemma 3.2 also holds if we substitute Y ∗ and SY ∗
for L(f )∗ and SL(f )∗, respectively.
Lemma 3.4. Suppose that X, Y are Banach spaces. Let ε ≥ 0. Assume
that f is a ε− isometry from X into Y with f (0) = 0. Then
(1) for every w∗-dense subset Ω ⊂ ext (BX ∗) there is a bounded linear
operator T : Y → ℓ∞(Ω) such that
kT f (x) − xk ≤ 2ε,
for all x ∈ X.
(2) If Y ∗ has the w∗-PCP or Y is separable, then there exists a w∗-
dense Gδ subset Ω ⊂ ext BX ∗ such that there is a bounded linear operator
T : Y → C(Ω) such that
kT f (x) − xk ≤ 2ε,
for all x ∈ X.
8
D. Dai
Proof. (1) By Theorem 1.3, for every x∗ ∈ Ω, there exists a functional
Q(x∗) ∈ SY ∗ such that
hQ(x∗), f (x)i − hx∗, xi ≤ 2ε, f or all x ∈ X.
We now define a mapping T : Y → ℓ∞(Ω) by
T (y) = {Q(x∗)(y)}x∗∈Ω.
It is clear that T is a bounded linear operator with norm one and
kT f (x) − xk = sup
x∗∈Ω
Q(x∗)f (x) − x∗(x) ≤ 2ε,
for all x ∈ X.
(2) Since ext (BX ∗) itself is a Baire space in its relative w∗-topology (see
[?, p.217, line 17-19 ]), it follows from Lemma 3.2 that there is a w∗− dense
Gδ subset Ω in ext (BX ∗) such that there is a w∗ − w∗ continuous selection
Q of Φ1 on Ω satisfying that for every x ∈ X and x∗ ∈ Ω, the following
inequality holds :
hQ(x∗), f (x)i − hx∗, xi ≤ 2ε.
Let T : Y → ℓ∞(Ω) be defined as in (i). Therefore, T (y) ∈ C(Ω) and
kT f (x) − xk ≤ 2ε,
for all x ∈ X.
Corollary 3.5. Suppose that X = C(K) for a compact Hausdorff space
K and Y ∗ has the w∗-PCP (especially, if Y is an Asplund space) or Y is
separable. Let ε ≥ 0. Assume that f is a standard ε-isometry from X into
Y . Then there exists a dense Gδ subset Ω of K such that there is a bounded
linear operator T : Y → C(Ω) such that T f − Id is uniformly bounded by
2ε on X.
Proof. It suffices to note that ext (BX ∗) = {±δt : t ∈ K} and {δt : t ∈ K}
is a compact Baire space norming for X, and then apply Lemma 3.2 and
Lemma 3.4 to conclude the results we desired by substituting {δt : t ∈ K}
respectively for H and ext (BX ∗) everywhere.
Now we introduce a new notion the so called the Cech-Stone property
that a topological space K is said to have the Cech-Stone property provided
that for every Gδ dense subset S ⊂ K there is a dense subset Ω ⊂ S such
On Qian's problem for L∞-spaces
9
that K is the Cech-Stone compactification of Ω. For example, βN, the Cech-
Stone compactification of N has the Cech-Stone property since every Gδ
dense subset of it must contain the discrete space N.
As a consequence of Corollary 3.5 we have
Theorem 3.6. Suppose that X = C(K) where K is a compact Hausdorff
space and admitting the Cech-Stone property. If either Y ∗ has the w∗-PCP
or Y is separable, then the pair (X, Y ) is (1, 2)-stable.
4 A quantitative characterization of separa-
bly injective Banach spaces
In this section, we combine Lemma 3.4 with some results from [22] by
Johnson-Oikhberg (Lindenstrass[23], Rosenthal [27], S´anchez [29] and Castillo-
Moreno [9]) and from [4] by Avil´es-S´anchez-Castillo-Gonz´alez-Moreno to
conclude a quantitative characterization of separably injective Banach spaces
which completely solves Problem 1.5.
Theorem 4.1. (i) If X is a λ-separably injective Banach space, then the
pair (X, Y ) is (3λ, 6λ) stable for every separable Banach space Y .
(ii) If the pair (X, Y ) is (λ, γ) stable for every separable Banach space
Y , then X is a λ-separably injective Banach space.
Proof. (i) Since Y is separable, it follows from Lemma 3.4 that for every
w∗-dense subset Ω ⊂ ext (BX ∗), there is a bounded linear operator T : Y →
ℓ∞(Ω) such that
kT f (x) − xk ≤ 2ε,
for all x ∈ X.
Let Z = span {Tf(X) ∪ X}. It follows from the continuity of T that Z/X
is separable quotient space since Y is separable. Since X is λ- separable
injective, it follows from Proposition 2.2 that X is 3λ-complemented in Z.
Therefore, there is a bounded linear operator P : Z → X with kP k ≤ 3λ
such that
kP T f (x) − x = kP T f (x) − P xk ≤ 6λε,
for all x ∈ X,
where P T : L(f ) → X satisfies that kP T k ≤ 3λ.
(ii) By Proposition 2.2, it suffices to show that X is λ-complemented
in every superspace Y such that Y /X is separable. Let Y = X + Y /X be
the algebraic direct sum. Since Y /X is separable, card (X) = card (Y ).
10
D. Dai
It follows from Qian's counterexample (i.e., Lemma 2.5) that there is an
ε-isometry f : X → Y such that Y = L(f ) and f (0) = 0. Hence by the
assumption, there is a projection P : Y → X with kP k ≤ λ and we complete
the proof.
Recall that a compact Hausdorff space K is said to be an F -space if
disjoint open Fσ sets have disjoint closures. For example, βN, the Cech-Stone
compactification of N and βN\N are F -spaces. Since C(K) is 1-separably
injective for every F -space K (see, for instance, [4, p.202-203], [23]), we have
Corollary 4.2. For every compact F -space K (for example, K = βN\N),
the pair (C(K), Y ) (resp. (ℓ∞/c0, Y ) ) is (3, 6) stable for every separable
Banach space Y .
Proof. It is sufficient to note that ℓ∞/c0 is linearly isometric to C(βN\N).
Recall that a compact space K has height n if K (n) = ∅, where we write
K ′ for the derived set of K and K (n+1) = (K (n))′. Since C(K) is (2n − 1)-
separably injective for every K of height n (see, for instance, [4, p.203]), we
have
Corollary 4.3. For every compact space K of height n, the pair (C(K), Y )
is (6n − 3, 12n − 6) stable for every separable Banach space Y .
To combine Theorem 4.1 with the results of Johnson-Oikhberg [22] that
for every family of λ-separably injective spaces {Ei}i∈Λ, (Pi∈Λ Ei)ℓ∞ and
(Pi∈Λ Ei)c0) are respectively λ-separably injective and 2λ2-separably injec-
tive, which was also proved by Rosenthal [27], S´anchez [29] and Castillo-
Moreno [9] with the estimates for the constant, respectively λ(1+λ)+, (3λ2)+
and 6(1 + λ), we have the following corollaries.
Corollary 4.4. The pair ((Pi∈Λ Ei)ℓ∞, Y ) is (3λ, 6λ) stable for every sep-
arable Banach space Y , where {Ei}i∈Λ is a family of λ-separably injective
spaces.
Corollary 4.5. The pair ((Pi∈Λ Ei)c0), Y ) is (6λ2, 12λ2) (resp. (3λ(1 +
λ)+, 6λ(1 + λ)+), ((9λ2)+, (18λ2)+) and (18(1 + λ), 36(1 + λ)) stable for
every separable Banach space Y , where {Ei}i∈Λ is a family of λ-separably
injective spaces.
On Qian's problem for L∞-spaces
11
Remark 4.6. There are many other examples for separably injective Ba-
nach spaces, such as the Johnson-Lindenstrauss spaces [21], Benyamini-
space which is an M-space nonisomorphic to a C(K)-space [6] and the WCG
nontrivial twisted sums of c0(Γ) constructed by Argyros, Castillo, Granero,
Jimenez and Moreno [2] (see, for instance, [4]).
Qian [28] proved that the pair (Lp, Lp) is stable for 1 < p < ∞. Semrl
and Vaisala [30] gave a sharp estimate for the constant pair (α, γ) with
γ = 2. Therefore, it is very natural to ask:
Problem 4.7. Is it true that the following pairs are stable for 1 ≤ p ≤ ∞
and p 6= q < ∞?
p )ℓ∞, (P∞
p )c0); (2)((P∞
p )c0, (P∞
n=1 ln
n=1 ln
n=1 ln
n=1 ln
n=1 lp)ℓ∞, (P∞
n=1 ℓ∞)lp); (4) ((P∞
n=1 ℓ∞)lp, (P∞
n=1 c0)lp, (P∞
n=1 Lp)ℓ∞); (6) ((P∞
n=1 Lp)ℓ∞, (P∞
n=1 ℓp)c0, (P∞
n=1 Lp)c0); (8) ((P∞
n=1 Lp)c0, (P∞
n=1 Lp)ℓq, (P∞
n=1 lp)ℓq); (10) ((P∞
n=1 lp)ℓq , (P∞
(1) ((P∞
(3) ((P∞
(5) ((P∞
(7) ((P∞
(9) ((P∞
p )ℓ∞);
n=1 lp)ℓ∞);
n=1 c0)lp);
n=1 lp)c0).
n=1 Lp)ℓq ).
case, it is not true for (6), (7) and (8) since (P∞
(P∞
n=1 ℓ∞)lp, (P∞
true for (3), (4) and (5) since (P∞
It is true for (1), (2), (3), (4) and (5) if p = ∞ as we have proved. In this
n=1 L∞)c0 and
n=1 ℓ∞)c0 are not complemented in ℓ∞. If 1 ≤ p < ∞, then it is also not
n=1 Lp)ℓ∞
are not complemented in ℓ∞. However, we do not know if it is true or not
for the above problem 4.7 in general case.
n=1 c0)ℓ∞, (P∞
n=1 lp)ℓ∞ and (P∞
5 A quantitative and generalized Sobczyk the-
orem
If Ei is a λ-injective Banach spaces for each i ∈ Λ (A Banach space X
is said to be λ-injective if it is λ-complemented in every superspace i.e., a
Banach space which contains X), then by Theorem 1.3 we have the following
Theorem 5.1 which gives sharpen constants (α, γ) for Qian's Problem. In
some sense, it could be seen as a quantitative and generalized Sobczyk
theorem (See Theorem 1.6).
Theorem 5.1. Let Λ and Γi for each i ∈ Λ are index sets. Suppose that
one of the following three statements holds
i) X is isomorphic to Z = (Pi∈Λ c0(Γi))ℓ∞ and λ > d(X, Z);
ii) X is isomorphic to Z = (Pi∈Λ ℓ∞(Γi))c0 and λ > d(X, Z);
12
D. Dai
iii) X = (Pi∈Λ Ei)c0 and {Ei}i∈Λ is a family of λ-injective Banach
spaces.
Then (X, Y ) is (2λ, 4λ)-stable for every separable Banach space Y .
Proof. i) Let X be a Banach space isomorphic to (Pi∈Λ c0(Γi))ℓ∞ and T :
X → (Pi∈Λ c0(Γi))ℓ∞ be an isomorphism such that kT k · kT −1k < λ. For
each n ∈ Λ and m ∈ Γn, let enm ∈ (Pi∈Λ c0(Γi))ℓ∞ with the standard
biorthogonal functionals e∗
ij(enm) = δinδjm.
For all n ∈ Λ and m ∈ Γn, let xnm ∈ X be such that T (xnm) = enm. Let
T ∗ : Z ∗ → X ∗ be the conjugate operator of T . Then
nm ∈ (Pi∈Λ c0(Γi))∗
ℓ∞ such that e∗
T (x) = { X
m∈Γn
(T ∗e∗
nm)(x)enm}n∈Λ
and
x = T −1{ X
m∈Γn
(T ∗e∗
nm)(x)enm}n∈Λ,
for all x ∈ X.
For all n ∈ Λ and m ∈ Γn, let x∗
nm ∈ kT kBX ∗. It follows from
Theorem 1.3 that for every n ∈ Λ and m ∈ Γn, there exists a functional
φnm ∈ kT kBY ∗ with kφnmk = kx∗
nm = T ∗e∗
nmk such that
(5.1)
hφnm, f (x)i − hx∗
nm, xi ≤ 2εkT k, for all x ∈ X.
It follows from the w∗ − w∗ continuity of T ∗ that for each n ∈ Λ, x∗
in the w∗-topology of X ∗ Since e∗
nm → 0 in the w∗-topology of Z ∗. Let
nm → 0
K = {ψ ∈ kT kB(Y ∗) : hψ, f (x)i ≤ 2εkT k, for all x ∈ X}.
Then K is a nonempty w∗-compact subset of Y ∗. Since Y is separable,
(kT kBY ∗, w∗) is metrizable. Let d be a metric such that (kT kBY ∗, d) is
homeomorphic to (kT kBY ∗, w∗). Since for each n ∈ Λ, (x∗
nm) is a w∗-null
net in X ∗, inequality (5.1) implies that for each n ∈ Λ, every w∗-cluster point
φ of (φnm) is in K such that kφk ≤ kT k, which yields that d(φnm, K) → 0
for each n ∈ Λ. Hence, for each n ∈ Λ, there is a net (ψnm) ⊂ K such that
d(φnm, ψnm) → 0, or equivalently, φnm − ψnm → 0 in the w∗-topology of Y ∗.
Let S : Y → X be defined for every y ∈ Y by
S(y) = T −1{ X
m∈Γn
hφnm − ψnm, yienm}n∈Λ ∈ X.
Hence
and
kSk ≤ 2kT k · kT −1k < 2λ
On Qian's problem for L∞-spaces
13
kSf (x) − xk = kT −1{ X
hφnm − ψnm, f (x)ienm}n∈Λ − T −1{ X
hx∗
nm, xienm}n∈Λk
≤ kT −1k sup
n∈Λ
≤ kT −1k · sup
n∈Λ
≤ kT −1k(sup
n∈Λ
(k X
m∈Γn
sup
m∈Γn
sup
m∈Γn
m∈Γn
hφnm − ψnm, f (x)ienm − X
m∈Γn
hx∗
nm, xienmk)
m∈Γn
hφnm, f (x)i − hx∗
nm, xi − hψnm, f (x)i
hφnm, f (x)i − hx∗
nm, xi + sup
n∈Λ
hψnm, f (x)i)
sup
m∈Γn
≤ 4εkT k · kT −1k < 4ελ.
ii-iii) For each i ∈ Λ, Γi denotes by BE ∗
i . It suffices to show this case
that X = (Pi∈Λ Ei)c0. Let J : X = (Pi∈Λ Ei)c0 → (Pi∈Λ ℓ∞(BE ∗
i ))c0 =
(Pi∈Λ ℓ∞(Γi))c0 be the canonical embedding. For each n ∈ Λ, let Qn :
(Pi∈Λ ℓ∞(Γi))c0 → ℓ∞(Γn) be the canonical projection. Let Pn : ℓ∞(Γn) →
let enm ∈ (Pi∈Λ ℓ∞(Γi))c0 with the standard biorthogonal functionals e∗
((Pi∈Λ ℓ∞(Γi))c0)∗ such that e∗
En be a family of projections with kPnk ≤ λ. For each n ∈ Λ and m ∈ Γn,
nm ∈
ij(enm) = δinδjm. Then
x = X
{(e∗
nm)(x)}m∈Γn for all x ∈ X.
n∈Λ
By Theorem 1.3, for each n ∈ Λ and m ∈ Γn, there exists φnm ∈ BY ∗ with
kφnmk = ke∗
nmk such that
hφnm, f (x)i − he∗
nm, xi ≤ 2ε, for all x ∈ X.
Clearly, e∗
nm → 0 uniformly for each m ∈ Γn in the w∗-topology of Z ∗. Let
K = {ψ ∈ B(Y ∗) : hψ, f (x)i ≤ 2ε, for all x ∈ X}.
Since Γn can be well ordered for every n ∈ Λ, we write
Γn = {0, 1, 2, · · · , w0, w0 + 1, · · · , w1, · · · ≺ Γn},
where Γn also denotes by its ordinal number. It follows from i) that for
each n ∈ Λ, there is a net (ψn0) ⊂ K such that d(φn0, ψn0) → 0. We can
choose (ψnm) ⊂ K such that for every n ∈ Λ and m ∈ Γn, d(φnm, ψnm) ≤
d(φn0, ψn0) or equivalently, (φnm − ψnm) → 0 uniformly for each m ∈ Γn
in the w∗-topology of Y ∗. Let Q : Y → (Pi∈Λ ℓ∞(Γi))c0 be defined for all
y ∈ Y by
Q(y) = X
{hφnm − ψnm, yi}m∈Γn ∈ (X
ℓ∞(Γi))c0,
n∈Λ
i∈Λ
14
which yields that
D. Dai
kQ(y)k ≤ (
sup
kφnm − ψnmk)kyk ≤ 2kyk.
n∈Λ,m∈Γn
Thus
Let S : Y → X be defined for all y ∈ Y by
kQk ≤ 2.
S(y) = X
PnQnQ(y) = X
n∈Λ
n∈Λ
Pn{hφnm − ψnm, yi}m∈Γn.
Hence
and
kSk = sup
n∈Λ
kPnQnQk ≤ 2λ
kSf (x) − xk = kX
Pn{hφnm − ψnm, f (x)i}m∈Γn − X
Pn{he∗
nm, xi}m∈Γnk
n∈Λ
n∈Λ
hφnm, f (x)i − he∗
nm, xi − hψnm, f (x)i
hφnm, f (x)i − he∗
nm, xi + sup
n∈Λ
hψnm, f (x)i)
sup
m∈Γn
≤ λ sup
n∈Λ
≤ λ(sup
n∈Λ
sup
m∈Γn
sup
m∈Γn
≤ 4ελ.
Thus, our proof is completed.
6 A generalized Figiel theorem for L∞,λ-spaces
Recall that a Banach space X is said to be a L∞,λ-space if every finite
dimensional subspace F of X is contained in another finite dimensional
subspace E of X such that d(E, ℓdim E
∞ ) ≤ λ. In this case, a Banach space
X is said to be a L∞-space if for some λ > 0, X is a L∞,λ-space(see, for
instance, [3], [4], [8]). For example, a λ-separably injective Banach space is
a L∞,9λ+-space (see [4, p.199, Prop.3.5 (a)])and every C(K)-space is also a
L∞-space.
Theorem 6.1. Suppose that X is a L∞,λ-space and Y is a Banach space.
Then for every standard ε-isometry f : X → Y , there is a bounded linear
operator T : Y → X ∗∗ such that T f − Id is uniformly bounded by 2λε on
X.
On Qian's problem for L∞-spaces
15
Proof. By Lemma 3.4, for every w∗-dense subset Ω ⊂ ext (BX ∗) there is a
bounded linear operator S : Y → ℓ∞(Ω) with norm one such that
kSf (x) − xk ≤ 2ε,
for all x ∈ X.
Let X = ∪i∈IEi such that for every i, j ∈ (I, (cid:23)), i (cid:23) j if and only if
Ei ⊇ Ej satisfying that for each i ∈ I, dim Ei < ∞ and d(Ei, ℓdim Ei
∞ ) ≤ λ.
Hence for each i ∈ I, there exists a projection Pi : ℓ∞(Ω) → Ei such that
kPik < λ + 1
. Since {Pi}i∈I is uniformly bounded on Bℓ∞(Ω), it follows
from the Arzel`a-Ascoli theorem that there is a subnet {δi}i∈Λ of I for an
partial order set Λ such that P : ℓ∞(Ω) → X ∗∗ is well defined by
1+dim Ei
P (y) = w∗ − lim
i∈Λ
Pδi(y), for all y ∈ ℓ∞(Ω),
which yields that
Hence
kP k ≤ λ and P X = Id.
kT f (x) − xk ≤ 2ελ,
for all x ∈ X,
where T = P S : Y → X ∗∗ with kT k ≤ λ.
7 Acknowledgements
This work was partially done based on many discussions with Professor
W.B. Johnson while the author was visiting Texas A&M University and in
Analysis and Probability Workshop at Texas A&M University which was
funded by NSF Grant. The author would like to thank Professor W.B.
Johnson and Professor Th. Schlumprecht for the invitation. This work is
also a part of the author's Ph.D. thesis under the supervision of Professor
Lixin Cheng.
The author was Supported by the Natural Science Foundation of China
(Grant No. 11601264) and the Outstanding Youth Scientific Research Per-
sonnel Training Program of Fujian Province and the High level Talents In-
novation and Entrepreneurship Project of Quanzhou City and the Research
Foundation of Quanzhou Normal University(Grant No. 2016YYKJ12).
References
[1] F. Albiac, N.J. Kalton, Topics in Banach Space Theory, Graduate
Texts in Mathematics 233, Springer, New York, 2006.
16
D. Dai
[2] S.A. Argyros, J.M.F. Castillo, A.S. Granero, M. Jim´enez, J.P. Moreno,
Complementation and embeddings of c0(I) in Banach spaces, Proc.
Lond. Math. Soc. 85 (2002) 742-772.
[3] S.A. Argyros, R.G. Haydon, A hereditarily indecomposable L∞-space
that solves the scalar-plus-compact problem, Acta Math. 206 (1) (2011)
1-54.
[4] A. Avil´es, F.C. S´anchez, Jes´us M. F. Castillo, M. Gonz´alez and Y.
Moreno, On separably injective Banach Spaces, Advances in Mathe-
matics 234 (2013) 192-216.
[5] L. Bao, L. Cheng, Q. Cheng, D. Dai, On universally left-stability of
ε-isometry, Acta Math. Sin., Engl. Ser. 29 (11) (2013) 2037 -- 2046.
[6] Y. Benyamini, An M-space which is not isomorphic to a C(K)-space,
Israel J. Math. 28 (1977) 98-104.
[7] Y. Benyamini, J. Lindenstrauss, Geometric Nonlinear Functional
Analysis I, Amer. Math. Soc. Colloquium Publications, Vol.48, Amer.
Math. Soc., Providence, RI, 2000.
[8] J. Bourgain, F. Delbaen, A class of special L∞-space, Acta Math.
145 (1980) 155-176.
[9] J.M.F. Castillo, Y. Moreno, Sobczyk's theorem and the bounded ap-
proximation property, Studia Math. 201 (2010) 1-19.
[10] L. Cheng, Y. Dong, W. Zhang, On stability of nonlinear non-surjective
ε-isometries of Banach spaces, J. Funct. Anal. 264 (2013) 713 -- 734.
[11] L. Cheng, D. Dai, Y. Dong et.al., Universal stability of Banach spaces
for ε-isometries, Studia Math. 221 (2014),141-149. arXiv:1301.3374v4.
[12] L. Cheng, Q. Cheng, K. Tu, J. Zhang A universal theorem for stability
of ε-isometries of Banach spaces, J. Funct. Anal. 269 (2015) 199 -- 214.
[13] D. Dai, Y. Dong, On stability of Banach spaces via nonlinear ε-
isometries, J. Math. Anal. Appl. 414 (2014) 996 -- 1005.
[14] S.J. Dilworth, Approximate isometries on finite-dimensional normed
spaces, Bull. London Math. Soc. 31 (1999) 471 -- 476.
On Qian's problem for L∞-spaces
17
[15] Y. Dutrieux, G. Lancien, Isometric embeddings of compact spaces into
Banach spaces, J. Funct. Anal. 255 (2008) 494 -- 501.
[16] M. Fabian, P. Habala, P. H´ajek, V. Montesinos, V. Zizler, Banach
Space Theory. The Basis for Linear and Nonlinear Analysis, Springer,
New York, 2011.
[17] T. Figiel, On non linear isometric embeddings of normed linear spaces,
Bull. Acad. Polon. Sci. Math. Astro. Phys. 16 (1968) 185 -- 188.
[18] J. Gevirtz, Stability of isometries on Banach spaces, Proc. Amer.
Math. Soc. 89 (1983) 633 -- 636.
[19] P.M. Gruber, Stability of isometries, Trans. Amer. Math. Soc. 245
(1978) 263 -- 277.
[20] D.H. Hyers, S.M. Ulam, On approximate isometries, Bull. Amer.
Math. Soc. 51 (1945) 288 -- 292.
[21] W.B. Johnson, J. Lindenstrauss, Some remarks on weakly compactly
generated Banach spaces, Israel J. Math. 17 (1974) 219-230.
[22] W.B. Johnson, T. Oikhberg, Separable lifting property and extensions
of local reflexivity, Illinois J. Math. 45 (2001) 123-137.
[23] J. Lindenstrauss, On the extension of operators with range in a C(K)
space, Proc. Amer. Math. Soc. 15 (1964) 218-225.
[24] S. Mazur, S. Ulam, Sur les transformations isom´etriques d'espaces
vectoriels norm´es, C.R. Acad. Sci. Paris 194 (1932) 946 -- 948.
[25] M. Omladic, P. Semrl, On non linear perturbations of isometries,
Math. Ann. 303 (1995) 617 -- 628.
[26] R.R. Phelps, Convex functions, monotone operators and differentia-
bility, Lecture Notes in Mathematics 1364, Springer, 1989.
[27] H.P. Rosenthal, The complete separable extenson property, J. Oper-
ator Theory 43 (2000) 329-374.
[28] S. Qian, ε-Isometric embeddings, Proc. Amer. Math. Soc. 123 (1995)
1797 -- 1803.
18
D. Dai
[29] F. Cabello S´anchez, Yet another proof of Sobczyk's theorem, in: Meth-
ods in Banach Space Theory, in: London Math. Soc. Lecture Notes,
Vol.337, Cambridge Univ. Press, 2006, pp. 133-138.
[30] P. Semrl, J. Vaisala, Nonsurjective nearisometries of Banach spaces,
J. Funct. Anal. 198 (2003) 268 -- 278.
[31] A. Sobczyk, Projection of the space (m) on its subspace c0, Bull. Amer.
Math. Soc. 47 (1941) 938 -- 947.
[32] J. Tabor, Stability of surjectivity, J. Approx. Theory 105 (2000) 166 --
175.
[33] J. Wolfe, Injective Banach spaces of type C(T ), Israel J. Math. 18
(1974) 133 -- 140.
[34] M. Zippin, The separable extension problem, Israel J. Math. 26 (1977)
372 -- 387.
|
1212.4134 | 1 | 1212 | 2012-12-17T20:45:28 | Orthonormal bases generated by Cuntz algebras | [
"math.FA"
] | We show how some orthonormal bases can be generated by representations of the Cuntz algebra; these include Fourier bases on fractal measures, generalized Walsh bases on the unit interval and piecewise exponential bases on the middle third Cantor set. | math.FA | math |
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
Abstract. We show how some orthonormal bases can be generated by representations of the
Cuntz algebra; these include Fourier bases on fractal measures, generalized Walsh bases on the unit
interval and piecewise exponential bases on the middle third Cantor set.
Contents
Introduction
1.
2. QMF bases and representations of the Cuntz algebra
3. Orthonormal bases generated by Cuntz algebras
3.1. Piecewise exponential bases on fractals
3.2. Walsh bases
References
1
3
8
9
11
15
The Cuntz algebra ON , [Cun77] is the C∗-algebra generated by N isometries Si, i = 0, . . . , N − 1
with the properties:
1. Introduction
(1.1)
S∗i Sj = δij, i, j = 0, . . . , N − 1,
SiS∗i = I.
N−1
Xi=0
The Cuntz algebras are ubiquitous in analysis, but we draw our inspiration from wavelet theory.
The role played by the Cuntz algebras in wavelet theory was described in the work of Bratteli
and Jorgensen [BJ02a, BJ02b, BEJ00, BJ97]. Orthonormal wavelet bases are constructed from
various choices of quadrature mirror filters (QMF) (see [Dau92]). These filters are in one-to-one
correspondence with certain representations of the Cuntz algebra. In section 2, we will show how
the ideas of Bratteli and Jorgensen carry over without too much difficulty in a more general setting
associated to some non-linear dynamics. We describe here this setting and give some examples.
Definition 1.1. Let X be a compact metric space and µ a Borel probability measure on X. Let
r−1(z) = N for µ.a.e. z ∈ X, where ·
r : X → X an N -to-1 onto Borel measurable map, i.e.
indicates cardinality. We say that µ is strongly invariant (for r) if for every continuous function f
on X the following invariance equation is satisfied:
2000 Mathematics Subject Classification. 42C10,28A80,42C40.
Key words and phrases. Cuntz algebras, fractal, Fourier basis, Hadamard matrix, quadrature mirror filter.
1
2
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
(1.2)
Z f dµ =
1
N Z Xr(w)=z
f (w)dµ(z)
Assumption. In this paper µ will be a strongly invariant measure for the N -to-1 map r : X → X
as in Definition 1.1
Example 1.2. Let T = {z ∈ C : z = 1} be the unit circle. Let r(z) = zN , z ∈ T. Let µ be the
Haar measure on T. Then µ is strongly invariant. An equivalent system can be realized on[0, 1]
with r(x) = N x mod 1, x ∈ [0, 1] with the Lebesgue measure dx on [0, 1]. We can identify the unit
circle T with the unit interval [0, 1] by z = e2πix.
Example 1.3. Let Γ be a countable discrete abelian group. Let α : Γ → Γ be an endomorphism
of Γ such that α(Γ) has finite index N in Γ and
∩n≥0 αn(Γ) = {0}
(1.3)
Let Γ be the compact dual group and let µ be the Haar measure on Γ, µ(Γ) = 1. Denote by α∗
the dual endomorphism on Γ, w 7→ w ◦ α (w ∈ Γ). Observe that α∗ is surjective, Ker α∗ = N so
α∗−1(z) = N for all z ∈ Γ, and condition (1.3) implies that ∪n≥0Ker α∗n is dense in Γ.
Proposition 1.4. The Haar measure on Γ is strongly invariant for α∗.
Proof. To prove the strong invariance relation (1.2) it is enough to check it on characters on Γ,
which by Pontryagin duality are given by the elements of Γ and we denote them by eγ(w) = w(γ),
γ ∈ Γ, w ∈ Γ. Fix γ ∈ Γ, γ 6= 0. Pick an element g0 ∈ Γ such that eγ(g0) 6= 1. We have
ZΓ
1
N Xα∗(w)=z
eγ(w)dµ(z) =ZΓ
1
eγ(w)dµ(z) =ZΓ
1
N Xα∗(u)=z
eγ(u − g0)dµ(z)
N Xα∗(w)=z−α∗(g0)
= eγ(g0)ZΓ
N Xα∗(u)=z
1
eγ(u)dµ(z).
1
strong invariance of µ is obtained.
Since eγ(g0) 6= 1 it follows that RΓ
N Pα∗(w)=z eγ(w)dµ(z) = 0. Since RΓ eγ(z)dµ(z) = 0 the
Example 1.5. We consider affine iterated function systems with no overlap. Let R be a d × d
expansive real matrix, i.e., all the eigenvalues of R have absolute value strictly greater than 1.Let
B ⊂ Rd a finite set such that N = B. Define the affine iterated function system
(cid:3)
(1.4)
By [Hut81] there exists a unique compact subset XB of Rd which satisfies the invariance equation
(x ∈ Rd, b ∈ B)
τb(x) = R−1(x + b)
(1.5)
XB = ∪b∈Bτb(XB)
XB is called the attractor of the iterated function system (τb)b∈B. Moreover XB is given by
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
3
(1.6)
XB =( ∞
Xk=1
R−kbk : bk ∈ B for all k ≥ 1)
Also, from [Hut81], there is a unique probability measure µB on Rd satisfying the invariance
equation
(1.7)
Z f dµB =
1
N Xb∈BZ f ◦ τbdµB
for all continuous compactly supported functions f on R. We call µB the invariant measure for the
IFS (τb)b∈B. By [Hut81], µB is supported on the attractor XB. We say that the IFS has no overlap
if µB(τb(XB) ∩ τ′b(XB)) = ∅ for all b 6= b′ in B.
Assume that the IFS (τb)b∈B has no overlap. Define the map r : XB → XB
r(x) = τ−1
b
(x), if x ∈ τb(XB)
(1.8)
Then r is an N -to-1 onto map and µB is strongly invariant for r. Note that r−1(x) = {τb(x) : b ∈ B}
for µB.a.e. x ∈ XB.
Example 1.6. Let r be a rational map on the complex sphere C∞. Let J be its Julia set. Then
by [Bro65], [OP72] there exists a stongly invariant measure µ supported on J, which is non-atomic.
The Julia set is invariant for r and the restriction r : J → J is a N -to-1 onto map where N = deg(r).
We will show in Section 2 Proposition 2.7 how representations of the Cuntz algebra are obtained
from a choice of a quadrature mirror filter (QMF) basis (Definition 2.4. Then we show how QMF
bases can be constructed using some unitary matrix valued functions (Theorem 2.12). This gives
us a large variety of representations of the Cuntz algebras, which we use in Section 3 to construct
various orthonormal bases.
The central result of the paper is Theorem 3.1, where we present a general criterion for a Cuntz
algebra representation to generate an orthonormal basis. As a corollary (Theorem 3.5), when
applied to some affine iterated function systems, we obtain a construction of piecewise exponential
bases on some Cantor fractal measures which extends a result of Dutkay and Jorgensen [DJ06b].
In particular, we obtain piecewise exponential orthonormal bases on the middle third Cantor set
(Example 3.8) which is known [JP98] not to have any orthonormal bases of exponential functions.
Another corollary to our Theorem 3.1 gives us a construction of generalized Walsh bases on the
unit interval starting from any unitary N × N matrix with constant first row.
2. QMF bases and representations of the Cuntz algebra
Definition 2.1. A quadrature mirror filter (QMF) for r is a function m0 in L∞(X, µ) with the
property that
(2.1)
1
N Xr(w)=z
m0(w)2 = 1,
(z ∈ X)
4
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
As shown by Dutkay and Jorgensen [DJ05, DJ07], every QMF gives rise to a wavelet theory.
Various extra conditions on the filter m0 will produce wavelets in L2(R) [Dau92], on Cantor sets
[DJ06a, MP11], on Sierpinski gaskets [DMP08] and many others.
Theorem 2.2. [DJ05, DJ07] Let m0 be a QMF for r. Then there exists a Hilbert space H, a
representation π of L∞(X) on H, a unitary operator U on H and a vector ϕ in H such that
(i) (Covariance)
(2.2)
U π(f )U∗ = π(f ◦ r),
(f ∈ L∞(X))
(ii) (Scaling equation)
(2.3)
U ϕ = π(m0)ϕ
(iii) (Orthogonality)
(2.4)
(iv) (Density)
(2.5)
hπ(f )ϕ , ϕi =Z f dµ,
(f ∈ L∞(X))
Definition 2.3. The system (H, U, π, ϕ) in Theorem 2.2 is called the wavelet representation asso-
ciated to the QMF m0.
span(cid:8)U−nπ(f )ϕ : f ∈ L∞(X), n ≥ 0(cid:9) = H
To construct a multiresolution, as in [Dau92], for a wavelet representation, one needs a QMF
basis.
Definition 2.4. A QMF basis is a set of N QMF's m0, m1, . . . , mN−1 such that
(2.6)
mi(w)mj (w) = δij,
1
(i, j ∈ {0, . . . , N − 1}, z ∈ X)
N Xr(w)=z
We can interpret these conditions in terms of a conditional expectation:
Definition 2.5. Let B be the Borel sigma-algebra on X and r−1(B) be the sigma-algebra
r−1(B) = {r−1(B) : B ∈ B}. Note that the r−1(B)-measurable functions are of the form f ◦ r,
where f is Borel measurable.
The conditional expectation from to B to r−1(B) is defined by
(z ∈ X)
E(f )(z) =
f (w),
1
(2.7)
N Xr(w)=z
Alternatively E(f ) can be defined, up to µ-measure zero as a r−1(B)-measurable function such that
(2.8)
for all g ∈ L∞(X, µ).
Z f g ◦ rdµ =Z E(f )g ◦ rdµ,
Proposition 2.6. A set of functions (mi)N−1
i=0
(2.9)
E(mimj) = δij,
in L∞(X, µ) is a QMF basis if and only if
(i, j ∈ {0, . . . N − 1})
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
5
In this case any function f ∈ L2(X, µ) can be written in the QMF basis as
(2.10)
f =
E(f mi)mi
N−1
Xi=0
Proof. The first statement is clear. For the second, define for f ∈ L2(X, µ) the vector-valued
function F (f )(z) = (f (w))r(w)=r(z) ∈ CN . Note that the QMF basis property implies that
(F ( 1√N
mi)(z))N−1
i=0
F (f )(z) =
mi)(z) =
E(f mi)(z)F (mi)(z)
N−1
is an orthonormal basis in Cn. Then for z ∈ X
Xi=0 (cid:28)F (f )(z) , F (
mi)(z)(cid:29)CN
1
√N
1
√N
F (
N−1
Xi=0
Then looking at the first component (since r(z) = r(z) one can take w = z) we get (2.10).
(cid:3)
Next, we show how a QMF basis induces a representation of the Cuntz algebra.
Proposition 2.7. Let (mi)N−1
i=0 be a QMF basis. Define the operators on L2(X, µ)
(2.11)
Then the operators Si are isometries and they form a representation of the Cuntz algebra ON , i.e.
Si(f ) = mif ◦ r,
i = 0, . . . , N − 1
(2.12)
S∗i Sj = δij,
i, j = 0, . . . , N − 1,
The adjoint of Si is given by the formula
SiS∗i = I
N−1
Xi=0
(2.13)
S∗i (f )(z) =
1
N Xr(w)=z
mi(w)f (w)
Proof. We compute the adjoint: take f , g in L2(X, µ). We use the strong invariance of µ.
hS∗i f , gi =Z f mig ◦ r dµ =Z 1
N Xr(w)=z
mi(w)f (w)g(z)dµ(z)
Then (2.13) follows. The Cuntz relations in (2.12) are then easily checked with Proposition 2.6. (cid:3)
Every QMF basis generates a multiresolution for the wavelet representation associated to m0.
Since the ideas are simple and are the same as in the classical wavelet theory presented in [Dau92],
we omit the proof. Note though, that the intersection of the resolution spaces might be non-trivial
(for example, if m0 = 1 then 1 is contained in this intersection).
Proposition 2.8. Let (mi)N−1
i=0
associated to m0. Define
be a QMF basis. Let (H, U, π, ϕ) be the wavelet representation
(2.14)
(2.15)
V0 := span{π(f )ϕ : f ∈ L∞(X)} ,
Vn = U−nV0, n ∈ Z
ψi = U−1π(mi)ϕ,
i = 1, . . . , N − 1
6
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
(2.16)
Then
Wi := span{π(f )ψi : f ∈ L∞(X)}
(i) ∪n∈ZVn = H
(ii) V1 = V0 ⊕ W1 ⊕ ··· ⊕ WN−1
(iii) If ∩n∈ZVn = {0} then
Mn∈Z
U n (W1 ⊕ ··· ⊕ WN−1) = H
A particular case which we will use in Section 3, is that of QMF bases generated by Hadamard
matrices which are defined from a finite set B and its spectrum Λ.
Definition 2.9. Denote by eλ(x) := e2πiλ·x for λ, x ∈ Rd. Let B be a finite subset of Rd, B =: N .
We say that a finite set Λ in Rd is a spectrum for B if Λ = N and the matrix
1
√N
[e2πib·λ]λ∈Λ
b∈B
is unitary. Let B and L be finite subsets of Zd, B =: N = L and let R be an expansive d × d
integer matrix. We say that (B, L) is a Hadamard pair with scaling factor R if L is a spectrum for
R−1B; equivalently, the matrix
is unitary.
1
√N
[e2πiR−1b·l]l∈L
b∈B
Example 2.10. Consider the setting in Example 1.5. We have the following equivalence:
Proposition 2.11. A finite set Λ in Rd is a spectrum for R−1B if and only if (eλ)λ∈Λ is a QMF
basis. Let L be a finite subset of Zd. Then (B, L) is a Hadamard pair with scaling factor R if and
only if (el)l∈L is a QMF basis.
Proof. We have
1
N Xr(w)=z
1
N Xb∈B
eλ(w)eλ′ (w) =
e2πiτb(z)·(λ−λ′) =
e2πiR−1(z+b)·(λ−λ′)
1
N Xb∈B
= e2πiR−1(z)·(λ−λ′) 1
e2πiR−1b·(λ−λ′)
N Xb∈B
Thus, the QMF basis condition is equivalent to
which is exactly the orthogonality of the columns of the matrix
e2πiR−1b·(λ−λ′) = δλλ′
1
N Xb∈B
The equivalence for Hadamard pairs follows as a particular case.
(cid:3)
1
√N
[e2πiR−1b·λ]λ∈Λ
b∈B
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
7
If B is a finite set and R−1B has spectrum Λ, then the set {eλ : λ ∈ Λ} is a QMF basis, by
Proposition 2.11. Then, with Proposition 2.7, the operators Sλf = eλf ◦ r form a representation
of the Cuntz algebra. Such representations were studied in [DJ12].
The next theorem shows how QMF bases can be constructed from unitary matrix valued functions
as in the work of Bratteli and Jorgensen [BJ02a, BJ02b, BEJ00, BJ97], now in a more general
context.
Theorem 2.12. Fix (mi)N−1
following two sets:
i=0 a QMF basis. There is a one-to-one correspondence between the
(i) QMF bases (m′i)N−1
i=0
(ii) Unitary valued maps A : X → UN (C)
Given a QMF basis (m′i)N−1
i=0
the matrix A with entries
(1) Aij(z) =
m′i(w)mj(w),
(z ∈ X, i, j = 0, . . . , N − 1)
is unitary.
Given a unitary-valued map A : X → UN (C), the functions form a QMF basis
(2) m′i(z) =
Aij(r(z))mj(z),
(z ∈ X, i = 0, . . . N − 1)
1
N Xr(w)=z
N−1
Xj=0
These correspondences are inverse to each other.
Proof. The result requires some simple computations
N−1
Xj=0
N 2 Xw,w′
1
Aij(z)Ai′j(z) =
m′i(w)m′i′ (w′) ·Xj
Xj
1
N 2 Xj Xr(w)=z
m′i(w)mj (z) · Xr(w′)=z
m′i′(w′)mj(w′) =
mj(w)mj(w′) =
m′i(w)m′i′(w′)δw,w′ = δii′
1
N Xw,w′
mj(w)mj(w′) = δww′
Note that we used the equality
which follows from the fact that the matrix
1
√N
[mi(w)]i=0,...N−1
w∈r−1(z)
is unitary, which, in turn, is a consequence of the QMF property. Hence A is unitary.
If A is unitary, we check the QMF relations:
1
N Xr(w)=z
N Xk,l
1
m′i(w)m′j(w) =
1
Aik(r(w))mk(w)Xl
N Xw Xk
mk(w)ml(w) =Xk,l
Aik(z)Ajl(z)Xw
Ajk(z)Ajl(z)δkl = δij
Ajl(r(w))ml(w) =
8
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
Hence (m′i)N−1
i=0
is a QMF basis.
The fact that the two correspondences are inverse to each other follows from the next computation:
Xj
Aij(r(z))mj (z) =Xj
1
N Xr(w)=r(z)
m′i(w)mj (w)
mj(z) = Xr(w)=r(z)
m′i(w)δwz = m′i(z)
= Xr(w)=r(z)
m′i(w) ·
1
N Xj
mj(w)mj(z)
(cid:3)
Remark 2.13. Note that the equation (1) can be reformulated as Aij(r(z)) = E(m′imj). The con-
ditional expectation E can be regarded as a L∞(X, µ)-valued inner product hf , giL∞(X,µ) = E(f g)
for f , g ∈ L∞(X, µ). The QMF basis condition is equivalent to the orthogonality of (mi)N−1
i=0 with
respect to this inner product. Since the dimension of L∞(X, µ) as a module over E(L∞(X, µ)) =
L∞(X, r−1(B), µ) is N , the completeness is automatic, so (mi)N−1
is an orthonormal basis for this
inner product. Thus A ◦ r is the change of base matrix from (mi) to (m′i). Equation (2) can be
understood in the sense that a unitary matrix maps orthonormal bases into orthonormal bases.
i=0
3. Orthonormal bases generated by Cuntz algebras
Next, we present the central result of our paper. It gives a general criterion for a family generated
by the Cuntz isometries to be an orthonormal basis.
Theorem 3.1. Let H be a Hilbert space and (Si)N−1
i=0 be a representation of the Cuntz algebra ON .
Let E be an orthonormal set in H and f : X → H a norm continuous function on a topological
space X with the following properties:
i=0 SiE.
(i) E = ∪N−1
(ii) span{f (t) : t ∈ X} = H and f (t)= 1, for all t ∈ X.
(iii) There exist functions mi : X → C, gi : X → X, i = 0, . . . , N − 1 such that
(3.1)
S∗i f (t) = mi(t)f (gi(t)),
t ∈ X.
(iv) There exist c0 ∈ X such that f (c0) ∈ spanE.
(v) The only function h ∈ C(X) with h ≥ 0, h(c) = 1, ∀ c ∈ {x ∈ X : f (x) ∈ spanE}, and
(3.2)
h(t) =
are the constant functions.
N−1
Xi=0
mi(t)2h(gi(t)),
t ∈ X
Then E is an orthonormal basis for H.
Proof. Define
h(t) :=Xe∈E
t ∈ X
where P is the orthogonal projection onto the closed linear span of E.
Since t 7→ f (t) is norm continuous we get that h is continuous. Clearly h ≥ 0. Also, if f (c) ∈
spanE, then P f (c)= f (c)= 1 so h(c) = 1. In particular, from (ii) and (iv), h(c0) = 1. We
hf (t) , ei2 = P f (t)2,
By (v), h is constant and, since h(c0) = 1, h(t) = 1 for all t ∈ X. Then P f (t)= 1 for all
t ∈ X. Since f (t)= 1 it follows that f (t) ∈ spanE for all t ∈ X. But the vectors f (t) span H so
spanE = H and E is an orthonormal basis.
(cid:3)
Remark 3.2. The operators of the form
Rh(t) =
mi(t)2h(gi(t)),
t ∈ X, h ∈ C(X),
N−1
Xi=0
that appear in (3.2), are sometimes called Ruelle operators or transfer operators, see e.g. [Bal00].
3.1. Piecewise exponential bases on fractals. We apply Theorem 3.1 to the setting of Example
2.10, in dimension d = 1 for affine iterated function systems, when the set 1
R B has a spectrum L.
Definition 3.3. Let L in R, L = N , R > 1 such that L is a spectrum for the set 1
R B. We say
that c ∈ R is an extreme cycle point for (B, L) if there exists l0, l1, . . . , lp−1 in L such that, if c0 = c,
c1 = c0+l0
= c0, and mB(ci) = 1 for i = 0, . . . , p − 1
where
R . . . cp−1 = cp−2+lp−2
R , c2 = c1+l1
then cp−1+lp−1
R
R
mB(x) =
e2πibx x ∈ R.
1
N Xb∈B
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
9
check (3.2). Since the sets SiE, i = 0, . . . N − 1 are mutually orthogonal, the union in (i) is disjoint.
Therefore for all t ∈ X :
h(t) =
N−1
Xi=0 Xe∈E
hf (t) , Siei2 =
hS∗i f (t) , ei2 =
hf (gi(t)) , ei2 =
N−1
Xi=0
mi(t)2Xe∈E
N−1
Xi=0 Xe∈E
Xi=0
N−1
=
mi(t)2h(gi(t))
Definition 3.4. We denote by L∗ the set of all finite words with digits in L, including the empty
word. For l ∈ L let Sl be given as in (2.11) where ml is replaced by the exponential el.
If
w = l1l2 . . . ln ∈ L∗ then by Sw we denote the composition Sl1Sl2 . . . Sln.
Theorem 3.5. Let B ⊂ R, 0 ∈ B, B = N , R > 1 and let µB be the invariant measure associated
to the IFS τb(x) = R−1(x + b), b ∈ B. Assume that the IFS has no overlap and that the set 1
R B
has a spectrum L ⊂ R, 0 ∈ L. Then the set
E(L) = {Swe−c : c is an extreme cycle point for (B, L), w ∈ L∗}
is an orthonormal basis in L2(µB). Some of the vectors in E(L) are repeated but we count them
only once.
Proof. Let c be an extreme cycle point. Then mB(c) = 1. Using the fact that we have equality in
the triangle inequality (1 = mB(c) ≤ 1
N Pb∈Be2πibc = 1) , and since 0 ∈ B, we get that e2πibc = 1
so bc ∈ Z for all b ∈ B. Also there exists another extreme cycle point d and l ∈ L such that d+l
R = c.
Then we have: Sle−c(x) = e2πilxe2πi(Rx−b)(−c), if x ∈ τb(XB). Since bc ∈ Z and R(−c) + l = −d,
we obtain
(3.3)
Sle−c = e−d
10
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
We use this property to show that the vectors Swe−c, Sw′e−c′ are either equal or orthogonal for
w, w′ in L∗ and c, c′ extreme cycle points for (B, L). Using (3.3), we can append some letters at
the end of w and w′ suh that the new words have the same length:
Swe−c = Swαe−d, Sw′e−c′ = Sw′βe−d′,
wα = w′β where d, d′ are cycle points.
Moreover, repeating the letters for the cycle points d and d′ as many times as we want, we can
assume that α ends in a repetition of the letters associated to d and similarly for β and d′. But,
since wα = w′β, the Cuntz relations imply that Swαe−d ⊥ Sw′βe−d′ or wα = w′β. Assume
w ≤ w′. Then α = w′′β for some word w′′. Then Swαe−d ⊥ Sw′βe−d iff Sαe−d ⊥ Sw′′βe−d′.
Also, α consists of repetitions of the digits of the cycle associated to d and similarly for d′. So
Sαe−d = e−f , Sw′′βe−d′ = e−f ′, and all points d, d′, f, f′, c, c′ all belong to the same cycle. So the
only case when Swe−c is not orthogonal to Sw′e−c′ is when they are equal.
Next we check that the hypotheses of Theorem 3.1 are satisfied. We let f (t) = e−t ∈ L2(µB). To
check (i) we just to have to see that e−c ∈ ∪l∈LSlE(L). But this follows from (3.3). Requirement
(ii) is clear. For (iii) we compute
S∗l e−t(x) =
1
N Xb∈B
e−2πil· 1
R (x+b)e−2πit· 1
R (x+b) = e−2πx· 1
R (t+l) 1
N Xb∈B
e−2πib( t+l
R ) =
So (iii) is satisfied with ml(t) = mB( t+l
For (iv) take c0 = −c for any extreme cycle point ( 0 is always one). For (v), take h continuous
= mB(cid:18) t + l
R (cid:19) e
R ), gl(t) = t+l
R .
(x)
− t+l
R
on R , 0 ≤ h ≤ 1, h(c) = 1 for all c with e−c ∈ spanE(L), and
h(cid:18) t + L
2
h(t) =Xl∈L(cid:12)(cid:12)(cid:12)(cid:12)
R (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
mB(cid:18) t + l
R (cid:19) := Rh(t)
R−1 , b ≥ maxL
In particular, we have h(c) = 1 for every extreme cycle point c. Assume h 6≡ 1. First we will restrict
our attention to t ∈ I := [a, b] with a ≤ minL
R−1 , and note that gl(I) ⊂ I for all l ∈ L. Let
m = mint∈I h(t). Then let h′ = h − m, assume m < 1. Then Rh′(t) = h′(t) for all t ∈ R, h′ has a
zero in I and h ≥ 0 on I, h′(z0) = 0. But this implies that mB(gl(z0))2h′(gl(z0)) = 0 for all l ∈ L.
Since Pl∈LmB(gl(z0))2 = 1, it follows that for one of the l0 ∈ L we have h′(gl0(z0)) = 0. By
induction, we can find zn = gln−1 ··· gl0z0 such that h′(zn) = 0. We prove that z0 is a cycle point.
Suppose not. Since mB has finitely many zeros, for n large enough gαk ··· gα1zn is not a zero for
mB, for any choice of digits α1, . . . , αk in L. But then, by using the same argument as above we get
that h′(gαk ··· gα1zn) = 0 for any α1, . . . , αk ∈ L. The points {gαk ··· gα1zn : α1, ...αk ∈ L, k ∈ N}
are dense in the attractor XL of the IFS {gl}l∈L, thus h′ is constant 0 on XL. But the extreme
cycle points c are in XL and since h(c) = 1 we have 0 = h′(c) = 1 − m, so m = 1. Thus h = 1 on
I. Since we can let a → −∞ and b → ∞ we obtain that h ≡ 1.
(cid:3)
Remark 3.6. The functions in E(L) are piecewise exponential. The formula for Sl1...lne−c is
(3.4)
Sl1...lne−c(x) = eα(b,l,c) · el1+Rl2+...+Rn−1ln−1+Rn(−c)(x)
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
11
where α(b, l, c) = −[b1l2 + (Rb1 + b2)l3 + ... + (Rn−2b1 + ... + bn−1)ln] + (Rn−1b1 + ... + bn) · c if
x ∈ τb1...τbnXB . We have
If x ∈ τb1...τbnXB then rx ∈ τb2...τbnXB, rn−1x ∈ τbnXB. So
Sl1...Slne−c(x) = el1(x)el2 (rx)...eln (rn−1x)ec(rnx)
rx = Rx − b1
r2x = Rrx − b2 = R2x − Rb1 − b2
...
rn−1x = Rn−1x − Rn−2b1 − ... − Rbn−2 − bn−1
rnx = Rnx − Rn−1b1 − Rn−2b2 − ... − Rbn−1 − bn.
The rest follows from a direct computation.
Corollary 3.7. In the hypothesis of Theorem 3.1, if in addition B, L ⊂ Z and R ∈ Z, then there
exists a set Λ such that {eλ : λ ∈ Λ} is an orthonormal basis for L2(µB).
Proof. If everything is an integer then, it follows from Remark 3.6 that Swe−c is an exponential
function for all w and extreme cycle points c. Note that, as in the proof of Theorem 3.1, bc ∈ Z for
all b ∈ B.
(cid:3)
2
2
3
so 0 ≤ c ≤ 3/4
= 1, therefore c ∈ 1
3{0, 2} has spectrum L = {0, 3/4}. We look for the extreme cycle points for (B, L).
Example 3.8. We consider the IFS that generates the middle third Cantor set: R = 3, B = {0, 2}.
The set 1
We need mB(−c) = 1 so 1+e2πi2c
Z. Also c has to be a cycle for the IFS
g0(x) = x/3, g3/4(x) = x+3/4
3−1 = 3/8. Thus, the only extreme cycle is {0}. By
Theorem 3.1 E = {Sw1 : w ∈ {0, 3/4}∗} is an orthonormal basis for L2(µB). Note also that the
numbers e2πiα(b,l,c) in formula (3.4) are ±1 because 2πiB · L ⊂ πiZ.
3.2. Walsh bases. In the following, we will focus on the unit interval, which can be regarded as
the attractor of a simple IFS and we use step functions for the QMF basis to generate Walsh-type
bases for L2[0, 1].
Example 3.9. The interval [0, 1] is the attractor of the IFS τ0x = x
2 , and the invariant
measure is the Lebesgue measure on [0, 1]. The map r defined in Example 1.5 is rx = 2xmod1. Let
m0 = 1, m1 = χ[0,1/2) − χ[1/2,1). It is easy to see that {m0, m1} is a QMF basis. Therefore S0, S1
defined as in Proposition 2.7 form a representation of the Cuntz algebra O2.
Proposition 3.10. The set E := {Sw1 : w ∈ {0, 1}∗} is an orthonormal basis for L2[0, 1], the
Walsh basis.
2 , τ1x = x+1
Proof. We check the conditions in Theorem 3.1. To see that (i) holds note that S01 = 1. Define
f (t) = et, t ∈ R. (ii) is clear. For (iii) we compute
S∗1 et(x) =
S∗1 et(x) =
1
2
1
2
(e2πit·x/2 + e2πit·(x+1)/2) = e2πit·x/2 1
2
(e2πit·x/2 − e2πit·(x+1)/2) = e2πit·x/2 1
2
(1 + e2πit/2)
(1 − e2πit/2)
12
Thus (iii) holds with m0(t) = 1
it follows that (iv) holds.
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
2 (1 + e2πit/2), m1(t) = 1
2 (1 − e2πit/2), g0(t) = g1(t) = t
2 . Since e0 = 1
For (v) take h continuous on R, 0 ≤ h ≤ 1, h(c) = 1 for all c ∈ R with et ∈ spanE, in particular
h(0) = 1 and
2
h(t) =(cid:12)(cid:12)(cid:12)(cid:12)
1
2
(1 + e2πit/2)(cid:12)(cid:12)(cid:12)(cid:12)
1
2
h(t/2) +(cid:12)(cid:12)(cid:12)(cid:12)
(1 − e2πit/2)(cid:12)(cid:12)(cid:12)(cid:12)
2
h(t/2) = h(t/2)
Then h(t) = h(t/2n) for all t ∈ R, n ∈ N. Letting n → ∞ and using the continuity of h, we get
h(t) = h(0) = 1 for all t ∈ R. Since all conditions hold, we get that E is an orthonormal basis.
That E is actually the Walsh basis follows from the following calculations: for w = n in {0, 1}∗
let n =Pi xi2i be the base 2 expansion of n. Because S0f = f ◦ r, S1f = m1f ◦ r and m0 ≡ 1 we
obtain the following decomposition:
Sw1(x) = m1(ri1x) · m1(ri2x)··· m1(rik x), where i1, i2, . . . , ik correspond to those i with xi = 1.
Also m1(rix) = m1(2ixmodi) are the Rademacher functions and thus we obtain the Walsh basis
(see e.g. [SWS90]).
The Walsh bases can be easily generalized by replacing the matrix
(cid:3)
1
1
√2(cid:18)1
1 −1(cid:19)
which appears in the definition of the filters m0, m1, with an arbitrary unitary matrix A with
constant first row and by changing the scale from 2 to N .
Theorem 3.11. Let N ∈ N, N ≥ 2. Let A = [aij] be an N × N unitary matrix whose first row is
N , x ∈ R, j = 0, . . . , N − 1 with the attractor [0, 1] and
constant
invariant measure the Lebesgue measure on [0, 1]. Define
. Consider the IFS τjx = x+j
1√N
mi(x) = √N
N−1
Xj=0
aijχ[j/N,(j+1)/N ](x)
i=0
is a QMF basis. Consider the associated representation of the Cuntz algebra ON .
Then {mi}N−1
Then the set E := {Sw1 : w ∈ {0, ...N − 1}∗} is an orthonormal basis for L2[0, 1].
Proof. We check the conditions in Theorem 3.1. Let f (t) = et, t ∈ R.
To check (i) note that S01 ≡ 1. (ii) is clear. For (iii) we compute:
1
N
S∗ket =
Xj=0
mk(τjx)et(τjx) =
1
√N
So (iii) is true with mk(t) = 1√N PN−1
(iv) is true with c0 = 0. For (v) take h ∈ C(R), 0 ≤ h ≤ 1, h(c) = 1 for all c ∈ R with ec ∈ spanE
akje2πit·(x+j)/N = e2πit·x/N 1
√N
j=0 akje2πit·j/N and gk(t) = t
N .
akje2πit·j/N
Xj=0
Xj=0
( in particular h(0) = 1), and
N−1
N−1
N−1
h(t) =
N−1
Xk=0
mk(t)2h(t/N ) = h(t/N )
N−1
Xk=0
1
N
N−1
Xj=0
akje−2πit·j/N2 = h(t/N ) ·
1
N Av2
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
13
where v = (e−2πit·j/N )N−1
j=0 . Since A is unitary, Av2= v2= N . Then h(t) = h(t/N n). Letting
n → ∞ and using the continuity of h we obtain that h(t) = 1 for all t ∈ R. Thus, Theorem 3.1
implies that E is an orthonormal basis.
(cid:3)
Remark 3.12. We can read the constants that appear in the step function Sw1 from the tensor
of A with itself n times, where n is the length of the word w.
Let A be an N × N matrix, B an M × M matrix. Then A ⊗ B has entries :
(A ⊗ B)i1+M i2,j1+M j2 = ai1j1bi2j2,
Ab0,0
Ab1,0
...
A ⊗ B =
...
AbM−1,0 AbM−1,1
i1, j1 = 0, . . . , N − 1, i2, j2 = 0, . . . , M − 1
Ab0,1
Ab1,1
Ab0,M−1
Ab1,M−1
···
···
. . .
··· AbM−1,M−1
...
The matrix A⊗n is obtained by induction, tensoring to the left: A⊗n = A ⊗ A⊗(n−1).
Thus A ⊗ A ⊗ A ⊗ ··· ⊗ A, n times, has entries
A⊗n
i0+N i1+N 2i2+···+N n−1in−1,j0+N j1+···+N n−1jn−1
= ai0j0ai1j1 . . . ain−1jn−1
Now compute for i0, . . . in−1 ∈ {0, . . . , N − 1}:
Si0...in−11(x) = mi0(x)mi1(rx) . . . min−1(rn−1x)
N n , k+1
Suppose x ∈ [ k
0 ≤ j0, . . . , jn−1 < N .
Then x ∈ [ j0
N , j0+1
N ), rx = (N x)mod1 ∈ [ j1
mi0(x) = √N ai0j0, mi1(rx) = √N ai1j1, . . . ,min−1(rn−1x) = √N ain−1jn−1 hence
N n ), 0 ≤ k < N n and k = N n−1j0 + N n−2j1 + ··· + N jn−2 + jn−1, where
N ), so
N ), . . . , rn−1x = (N n−1x)mod1 ∈ [ jn−1
N , jn−1+1
N , j1+1
Si0...in−11(x) = √N nai0j0 . . . ain−1jn−1 = √N nA⊗n
i0+N i1+N 2i2+···+N n−1in−1,j0+N j1+···+N n−1jn−1
14
DORIN ERVIN DUTKAY, GABRIEL PICIOROAGA, AND MYUNG-SIN SONG
Figure 1. Walsh functions Sw1 for words w of length 2.
Example 3.13. The pictures in Figure 1 show the Walsh functions that correspond to the scale
N = 4 and the matrix
1
1
2
2
√2
√2
2 −
2
0
0
1
1
2
2
A =
for the words of length 2, indicated at the top.
1
1
2
2
0
0
√2
√2
2 −
2
− 1
2 − 1
2
ORTHONORMAL BASES GENERATED BY CUNTZ ALGEBRAS
15
References
[Bal00] Viviane Baladi. Positive transfer operators and decay of correlations, volume 16 of Advanced Series in
Nonlinear Dynamics. World Scientific Publishing Co. Inc., River Edge, NJ, 2000.
[BEJ00] Ola Bratteli, David E. Evans, and Palle E. T. Jorgensen. Compactly supported wavelets and representations
[BJ97]
of the Cuntz relations. Appl. Comput. Harmon. Anal., 8(2):166 -- 196, 2000.
Ola Bratteli and Palle E. T. Jorgensen. Isometries, shifts, Cuntz algebras and multiresolution wavelet
analysis of scale N . Integral Equations Operator Theory, 28(4):382 -- 443, 1997.
[BJ02a] Ola Bratteli and Palle Jorgensen. Wavelets through a looking glass. Applied and Numerical Harmonic
Analysis. Birkhauser Boston Inc., Boston, MA, 2002. The world of the spectrum.
[BJ02b] Ola Bratteli and Palle E. T. Jorgensen. Wavelet filters and infinite-dimensional unitary groups. In Wavelet
analysis and applications (Guangzhou, 1999), volume 25 of AMS/IP Stud. Adv. Math., pages 35 -- 65. Amer.
Math. Soc., Providence, RI, 2002.
[Bro65] Hans Brolin. Invariant sets under iteration of rational functions. Ark. Mat., 6:103 -- 144 (1965), 1965.
[Cun77]
[Dau92]
Joachim Cuntz. Simple C ∗-algebras generated by isometries. Comm. Math. Phys., 57(2):173 -- 185, 1977.
Ingrid Daubechies. Ten lectures on wavelets, volume 61 of CBMS-NSF Regional Conference Series in Ap-
plied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1992.
Dorin Ervin Dutkay and Palle E. T. Jorgensen. Hilbert spaces of martingales supporting certain
substitution-dynamical systems. Conform. Geom. Dyn., 9:24 -- 45 (electronic), 2005.
[DJ05]
[DJ06a] Dorin E. Dutkay and Palle E. T. Jorgensen. Wavelets on fractals. Rev. Mat. Iberoam., 22(1):131 -- 180, 2006.
[DJ06b] Dorin Ervin Dutkay and Palle E. T. Jorgensen. Iterated function systems, Ruelle operators, and invariant
[DJ07]
[DJ12]
projective measures. Math. Comp., 75(256):1931 -- 1970 (electronic), 2006.
Dorin Ervin Dutkay and Palle E. T. Jorgensen. Martingales, endomorphisms, and covariant systems of
operators in Hilbert space. J. Operator Theory, 58(2):269 -- 310, 2007.
Dorin Ervin Dutkay and Palle E. T. Jorgensen. Spectral measures and Cuntz algebras. Math. Comp.,
81(280):2275 -- 2301, 2012.
[DMP08] Jonas D'Andrea, Kathy D. Merrill, and Judith Packer. Fractal wavelets of Dutkay-Jorgensen type for the
Sierpinski gasket space. In Frames and operator theory in analysis and signal processing, volume 451 of
Contemp. Math., pages 69 -- 88. Amer. Math. Soc., Providence, RI, 2008.
John E. Hutchinson. Fractals and self-similarity. Indiana Univ. Math. J., 30(5):713 -- 747, 1981.
Palle E. T. Jorgensen and Steen Pedersen. Dense analytic subspaces in fractal L2-spaces. J. Anal. Math.,
75:185 -- 228, 1998.
[Hut81]
[JP98]
[MP11] Matilde Marcolli and Anna Maria Paolucci. Cuntz-Krieger algebras and wavelets on fractals. Complex Anal.
Oper. Theory, 5(1):41 -- 81, 2011.
[OP72] Marilyn K. Oba and Tom S. Pitcher. A new characterization of the F set of a rational function. Trans.
Amer. Math. Soc., 166:297 -- 308, 1972.
[SWS90] F. Schipp, W. R. Wade, and P. Simon. Walsh series. Adam Hilger Ltd., Bristol, 1990. An introduction to
dyadic harmonic analysis, With the collaboration of J. P´al.
[Dorin Ervin Dutkay] University of Central Florida, Department of Mathematics, 4000 Central
Florida Blvd., P.O. Box 161364, Orlando, FL 32816-1364, U.S.A.,
E-mail address: [email protected]
[Gabriel Picioroaga] University of South Dakota, Department of Mathematical Sciences, 414 E.
Clark St., Vermillion, SD 57069, U.S.A.,
E-mail address: [email protected]
[Myung-Sin Song] Department of Mathematics and Statistics, Southern Illinois University Ed-
wardsville, Campus Box 1653, Science Building, Edwardsville, IL 62026, U.S.A.,
E-mail address: [email protected]
|
1302.6019 | 1 | 1302 | 2013-02-25T08:58:45 | On dentability in locally convex vector spaces | [
"math.FA"
] | For a locally convex vector space (l.c.v.s.) $E$ and an absolutely convex neighborhood $V$ of zero, a bounded subset $A$ of $E$ is said to be $V$-dentable (respectively, $V$-f-dentable) if for any $\epsilon>0$ there exists an $x\in A$ so that $$x\notin \bar{co} (A\setminus (x+\epsilon V)) $$ (respectively, so that $$ x\notin {co} (A\setminus (x+\epsilon V))). $$ Here, "$\bar{co}$" denotes the closure in $E$ of the convex hull of a set. We present a theorem which says that for a wide class of bounded subsets $B$ of locally convex vector spaces the following is true: $(V)$ every subset of $B$ is $V$-dentable if and only if every subset of $B$ is $V$-f-dentable. The proof is purely geometrical and independent of any related facts. As a consequence (in the particular case where $B$ is complete convex bounded metrizable subset of a l.c.v.s.), we obtain a positive solution to a 1978-hypothesis of Elias Saab (see p. 290 in "On the Radon-Nikodym property in a class of locally convex spaces", Pacific J. Math. 75, No. 1, 1978, 281-291). | math.FA | math |
ON DENTABILITY IN LOCALLY CONVEX VECTOR SPACES
Oleg Reinov and Asfand Fahad
§0. Preliminaries
For a locally convex vector space (l.c.v.s.) E and an (absolutely convex) neighborhood
V of zero, a bounded subset A of E is said to be V -dentable (respectively, V -f-dentable )
if for any ǫ > 0 there exists an x ∈ A so that
(respectively, so that
x /∈ co (A \ (x + ǫV ))
x /∈ co (A \ (x + ǫV )) ).
Here, "co " denotes the closure in E of the convex hull of a set.
We present a theorem which says that for a wide class of bounded subsets B of locally
convex vector spaces the following is true:
(V ) every subset of B is V -dentable if and only if every subset of B is V -f-dentable.
The proof is purely geometrical and independent of any related facts.
As a consequence (in the particular case where B is complete convex bounded metriz-
able subset of a l.c.v.s.), we obtain a positive solution to a 1978-hypothesis of an Amer-
ican mathematician Elias Saab (see p. 290 in "On the Radon-Nikodym property in a
class of locally convex spaces", Pacific J. Math. 75, No. 1, 1978, 281-291).
§1. A proposition
Proposition.. Let B be a bounded sequentially complete convex metrizable subset of
a locally convex vector space E, V is a neighborhood of zero in E. The following are
equivalent:
1). B is subset V -dentable 1;
2). B is subset V -f-dentable.
The research is supported by the Higher Education Commission of Pakistan and by grant 12-01-00216
of RFBR.
z AMS Subject Classification 2010: 46B22 Radon-Nikodym, Krein-Milman and related properties;
46A55 Convex sets in topological linear spaces; Choquet theory.
Key words: dentability, dentable sets, locally convex spaces.
1"subset Π" means that every subset of a set is Π.
1
Typeset by AMS-TEX
Proof. Clearly, 1) =⇒ 2). Let d be a metric in B which gives the topology on B,
induced from E. For x ∈ B, denote by Dε(x) the e-ball in B with a center at x. We
may and do assume that V is absolutely convex.
Given 2), suppose that there is a subset K ⊂ B, which is not V -dentable, i.e. there
exists an ε > 0 such that x ∈ co (K \ (2εV + x)) for all x ∈ K. Put (for a later use)
k(i, 0) = 0 if i = 1, 2, . . . . Fix x01 ∈ K and take x1j ∈ K (j = 1, 2, . . . , k(0, 1)) such that
x1j /∈ x01 + 2εV,
i.e. x1j − x01 /∈ 2εV ;
d(x01,
k(0,1)
X
j=1
α1jx1j) < ε0/8,
Dε0(x01) ⊂ ε/8 V + x01,
and
where α1j ≥ 0, Pj α1j = 1 and ε0 is such that ε0 < ε/8.
Let ε1 > 0 be such a number that 0 < ε1 < ε0/2, x1j − x01 /∈ (ε + ε1)V, Dε1(x1j ) ⊂
ε/8 V + x1j and Dε1(x1j) ∩ [(ε + ε1)V + x01] = ∅ for all j = 1, 2, . . . , k(0, 1) =: n(0).
For every x1m take (x2j) ⊂ K (j = k(1, m − 1) + 1, . . . , k(1, m)) so that
x2j − x1m /∈ 2εV,
d(x1m,
k(1,m)
X
j=k(1,m−1)+1
α2jx2j) < ε1/4
and
d(
k(0,1)
X
j1=1
α1j1
k(1,m)
X
j2=k(1,m−1)+1
α2j2x2j2 ,
k(0,1)
X
j1=1
α1j1x1j1 ) < ε1/4,
where α2j ≥ 0, Pk(1,m)
next step).
j2=k(1,m−1)+1 α2j2 = 1 (we use the continuity of the metric d; see the
Let n(1) = k(1, n(0)) and ε2 be such that 0 < ε2 < ε1/2, x1m − x2j /∈ (ε + ε2)V,
Dε2(x2j ) ⊂ ε/8 V + x2j and Dε2 (x2j) ∩ [(ε + ε2)V + x1m] = ∅ for every pair of indices
m = 1, 2, . . . , n(0) and j = k(1, m − 1) + 1, . . . , k(1, m).
Now, find a δ2 > 0, δ2 < ε2/42, with a property that it follows from (x2j) ⊂ B and
maxj d(x2j, x2j) < δ2 that for all m = 1, 2, . . . , k(0, 1)
k(1,m)
X
d(
α2j2x2j2,
k(1,m)
X
j2=k(1,m−1)+1
j2=k(1,m−1)+1
α2j2 x2j2) < ε2/42
and
d(
k(0,1)
X
j1=1
α1j1
k(1,m)
X
α2j2x2j2,
j2=k(1,m−1)+1
α1j1
k(0,1)
X
j1=1
2
k(1,m)
X
j2=k(1,m−1)+1
α2j2 x2j2) < ε2/42
For every x2m take (x3j) ⊂ K (j = k(2, m − 1) + 1, . . . , k(2, m)) so that
x2m − x3j /∈ 2εV,
d(x2m,
k(2,m)
X
j=k(2,m−1)+1
α3jx3j) < δ2 (< ε2/42),
where α3j ≥ 0, Pk(2,m)
the corresponding three inequalities combining the last ones.
j=k(2,m−1)+1 α3j = 1. Putting x2j := Pk(2,m)
j=k(2,m−1)+1 α3jx3j, we get
And one more step (of induction).
Let n(2) = k(1, n(1)) and let ε3 be such that 0 < ε3 < ε2/2, x2m − x3j /∈ (ε + ε3)V,
Dε3(x3j ) ⊂ ε/8 V + x3j and Dε3 (x3j) ∩ [(ε + ε3)V + x2m] = ∅ for every pair of indices
m = 1, 2, . . . , n(1) and j = k(2, m − 1) + 1, . . . , k(2, m).
Find a δ3 > 0, δ3 < ε3/43 such that it follows from (x3j) ⊂ B and maxj d(x3j, x3j) <
δ3 that for all m2 = 1, 2, . . . , n(1) and m1 = 1, 2, . . . , n(0) :
k(2,m2)
X
d(
α3j3x3j3 ,
k(2,m2)
X
j3=k(2,m2−1)+1
j3=k(2,m2−1)+1
α3j3 x3j3) < ε3/43,
k(1,m1)
X
d(
α2j2
k(2,m2)
X
j2=k(1,m1−1)+1
j3=k(2,m2−1)+1
α3j3x3j3,
k(1,m1)
X
α2j2
k(2,m2)
X
j2=k(1,m1−1)+1
j3=k(2,m2−1)+1
α3j3 x3j3 ) < ε3/43
and
d(
k(0,1)
X
j1=1
α1j1
k(1,m1)
X
α2j2
k(2,m2)
X
α3j3 x3j3,
j2=k(1,m1−1)+1
j3=k(2,m2−1)+1
k(0,1)
X
j1=1
α1j1
k(1,m1)
X
α2j2
k(2,m2)
X
j2=k(1,m1−1)+1
j3=k(2,m2−1)+1
α3j3 x3j3) < ε3/43.
For every x3m2 (m2 = 1, . . . , n(1)) take (x4j) ⊂ K (j = k(3, m2 − 1) + 1, . . . , k(3, m2))
so that
x3m2 − x4j /∈ 2εV,
d(x3m2,
k(3,m2)
X
j=k(3,m2−1)+1
α4jx4j) < δ3 (< ε3/43),
where α4j ≥ 0, Pk(3,m2)
four corresponding inequalities combining the last ones.
j=k(3,m2−1)+1 α4j = 1. Putting x3j3 := Pk(3,m2)
j=k(3,m2−1)+1 α4jx4j, we get
3
Then, let n(3) = k(2, n(2)) and let ε4 be such that 0 < ε4 < ε3/2, x3m − x4j /∈
(ε + ε4)V, Dε4(x4j ) ⊂ ε/8 V + x4j and Dε4 (x4j) ∩ [(ε + ε4)V + x3m] = ∅ for every
corresponding pair x3m, x4j etc.
We think that induction is clear.
By this induction, we get a subset (xim)i,m ⊂ K with properties like ones just indi-
i=0,m=1 (with n(−1) = 1)
i=0 in mind, we, for any q = 1, 2, . . . ; i = 0, 1, 2, . . . and m = 1, 2, . . . , n(i − 1),
cated above for a part of the set. Having this sequences (xim)∞,n(i−1)
and (εi)∞
put
ϕ(i, m; q) :=
k(i,m)
X
αi+1jm+1
k(i+1,jm+1)
X
αi+2jm+2
jm+1=k(i,m−1)+1
jm+2=k(i+1,jm+1−1)+1
k(i+2,jm+2)
X
αi+3jm+3 · · ·
k(i+q−1,jm+q−1)
X
jm+3=k(i+2,jm+2−1)+1
jm+q =k(i+q−1,jm+q−1−1)+1
αi+qjm+q xi+qjm+q ,
Then, by triangle inequality, d(ϕ(i, m; q), xim) ≤ εi and d(ϕ(i, m; q1), ϕ(i, m; q2)) →
q1,q2
zim in B
0. Therefore, for any i and m, there exists zim ∈ co K for which ϕ(i, m; q) →
q
and
(i) d(zim, xim) ≤ εi,
(ii) zim =
k(i,m)
X
j=k(i,m−1)+1
αi+1jzi+1j .
To obtain the second equality, it is enough to compare the definition of ϕ(i, m; q) with
corresponding formulas for ϕ(i, m; m + 1) and ϕ(i + 1, jm+1; q) to note that
ϕ(i, m; q + 1) =
k(i,m)
X
jm+1=k(i,m−1)+1
αi+1jm+1 [ϕ(i + 1, jm+1; q)].
Taking a limit, as q → +∞, we get (ii).
Fix i = 1, 2, . . . ; m = 1, 2, . . . , n(i − 1) and j = k(i, m − 1) + 1, . . . , k(i, m). Since
xim − xi+1j /∈ (ε + εi+1)V, we get from (i) and the construction that, e.g., zim − zi+1j /∈
(ε/10)V. Together with (ii), this gives us a contradiction.
§2. Main Theorem
The proof of the following simple assertion can be given by a reader:
Lemma.. Let E be a locally convex vector space and let B ⊂ E be closed bounded
convex sequentially complete and having the property that for every M ⊂ B and for
x ∈ M there exists a sequence xn ∈ M such that lim
xn = x. If B is not V -dentable
n
then there exists a countable set A ⊆ B which is not V -dentable
Proposition and Lemma yields the following theorem.
4
Theorem.. Let E be a locally convex vector space, V is a neighborhood of zero in E and
let B ⊂ E have the following properties: (i) it is closed bounded convex and sequentially
complete; (ii) for every M ⊂ B and for x ∈ M there exists a sequence xn ∈ M such
that limn xn = x; (iii) each separable subset of B is metrizable.
Then the following are equivalent:
(i) B is subset V -dentable;
(ii) B is subset V -f-dentable.
Oleg Reinov:
St. Petersburg State University, Dept. Math. and Mech.,
198904, Saint Petersburg, Russia and
Abdus Salam School of Mathematical Sciences,
68-B, New Muslim Town, Lahore 54600, PAKISTAN
E-mail address: [email protected]
Asfand Fahad:
Abdus Salam School of Mathematical Sciences,
68-B, New Muslim Town, Lahore 54600, PAKISTAN
E-mail address: [email protected]
5
|
1001.4552 | 1 | 1001 | 2010-01-25T22:00:49 | Existence of fixed points for a particular multifunction | [
"math.FA"
] | We prove a fixed point theorem for a particular multifunction from the unit sphere of a reflexive Banach space with the Kadec-Klee property into itself. | math.FA | math |
Existence of fixed points for a particular multifunction
BIAGIO RICCERI
Department of Mathematics
University of Catania
Viale A. Doria 6
95125 Catania, Italy
E-mail: [email protected]
Abstract. In this paper, we prove that if E is an infinite-dimensional reflexive real Banach
space possessing the Kadec-Klee property, then, for every compact function from the unit
sphere S of E to the dual E∗ satisfying the condition inf x∈S kf (x)kE∗ > 0, there exists
x ∈ S such that
f (x)(x) = kf (x)kE∗ .
Key words and phrases: Kadec-Klee property, reflexivity, upper semicontinuous mul-
tifunction, fixed point, Fan-Kakutani theorem.
2010 Mathematics Subject Classification: 46B10, 46B20, 47H04, 47H10.
Here and in the sequel, E is a reflexive real Banach space, with dual E∗. Set
S = {x ∈ E : kxk = 1} .
Let f : S → E∗ be a continuous function.
In this very short paper, we are interested in the existence of some x ∈ S such that
Clearly, if, for each x ∈ S, we set
f (x)(x) = kf (x)kE∗ .
Φf (x) := {y ∈ S : f (x)(y) = kf (x)kE∗} ,
our problem is equivalent to finding a fixed point of the multifunction Φf .
Note that, by reflexivity, we have Φf (x) 6= ∅ for all x ∈ S. However, the mere
continuity of f is not enough to guarantee the existence of solutions to our problem.
1
Let us recall that, when dim(E) = ∞, E is said to have the Kadec-Klee property if
for every sequence {xn} in S weakly converging to x ∈ S, one has limn→∞ kxn − xk = 0.
Also, we say that a function ψ : S → E∗ is compact if it is continuous and ψ(S) is
relatively compact.
Here is our contribution about the above problem.
THEOREM 1. - Let E be infinite-dimensional and have the Kadec-Klee property, and
let f : S → E∗ be a compact function such that
inf
x∈S
kf (x)kE∗ > 0 .
(1)
Then, there exists x ∈ S such that
f (x)(x) = kf (x)kE∗ .
Let us recall that a multifunction F : X → 2Y between two topological spaces is said
to be upper semicontinuous if, for each closed set C ⊆ Y , the set
F −(C) := {x ∈ X : F (x) ∩ C 6= ∅}
is closed in X.
Proof of Theorem 1. Consider the multifunction Ψ : E∗ → 2S defined by putting
Ψ(ϕ) = {x ∈ S : ϕ(x) = kϕkE∗ }
for all ϕ ∈ E∗. Let us show that the restriction of Ψ to E∗\{0} is upper semicontinuous. To
this end, let C ⊆ S be a (non-empty) closed set and let {ϕn} be a sequence in Ψ−(C) \ {0}
converging in E∗ to ϕ 6= 0. We have to show that ϕ ∈ Ψ−(C). For each n ∈ N, choose
xn ∈ C so that
ϕn(xn) = kϕnkE∗ .
(2)
By reflexivity, there is a subsequence {xnk } weakly converging to some x ∈ E. Reading
(2) with nk instead of n and passing to the limit for k → ∞, we get
ϕ(x) = kϕkE∗ .
Since ϕ 6= 0, we have x ∈ S. Consequently, since E has the Kadec-Klee property, {xnk }
converges strongly to x. Therefore, since C is closed, we have x ∈ C. Hence, x ∈ Ψ(ϕ) ∩ C
and so ϕ ∈ Ψ−(C). Next, observe that, for each ϕ ∈ E∗ \ {0}, the set Ψ(ϕ) is bounded,
closed and convex, and so it is weakly compact, by reflexivity. But, since E has the
Kadec-Klee property, each weakly compact subset of S is compact, in view of the Eberlein-
Smulyan theorem. So, each set Ψ(ϕ), with ϕ 6= 0, is compact. Next, note that, by (1),
K := f (S) is a compact set in E∗ which does not contain 0. Consequently, by the
upper semicontinuity of Ψ(E∗\{0}), the set Ψ(K) is compact ([2], Theorem 7.4.2). Since
2
dim(E) = ∞, there is a continuous function ω : B → S such that ω(x) = x for all x ∈ S,
where B is the closed unit ball of E. Finally, denote by Y the closed convex hull of Ψ(K)
and set
G(x) = Ψ(f (ω(x)))
for all x ∈ Y . So, G is an upper semicontinuous multifunction (as the composition of the
upper semicontinuous multifunction Ψ(E∗\{0}) and the continuous function f ◦ ω) with
non-empty, closed and convex values, from the compact convex set Y into itself. Then,
by the Fan-Kakutani theorem ([1]), there exists x ∈ Y such that x ∈ G(x). Then, since
x ∈ S, we have ω(x) = x, and so x ∈ Φf (x), as desired.
△
Some remarks about the assumptions of Theorem 1 are now in order.
Assume that (E, h·, ·, i) is a Hilbert space. Consider the continuous function f : S →
E∗ defined by putting
f (x)(y) = −hx, yi
for all x ∈ S, y ∈ E. In this case, we have
and
f (x)(x) = −1
kf (x)kE∗ = 1
for all x ∈ S. This example shows, at the same time, that Theorem 1 is no longer true if
either the infinite dimensionality of E or the compactness of f is removed.
Also, note that, concerning the compactness of f , a more sophisticated example can
be provided in any infinite-dimensional Banach space. Actually, in this case, E. Michael
([3]) proved that there exists a continuous function f : S → E∗ such that
and
for all x ∈ S.
f (x)(x) = 0
kf (x)kE∗ = 1
Concerning the necessity of (1), consider the following example.
Let E be the space of all absolutely continuous functions u : [0, 1] → R with u′ ∈
L2([0, 1]). In other words, let E be the usual Sobolev space H 1(0, 1), with the usual norm
kuk = (cid:18)Z 1
0
u′(t)2dt +Z 1
0
u(t)2dt(cid:19)
1
2
.
Consider the continuous function f : S → E∗ defined by putting
f (u)(v) = −Z 1
0
u(t)v(t)dt
3
for all u ∈ S, v ∈ E. Note that E has the Kadec-Klee property since it is a Hilbert
space. Moreover, since E is compactly embedded into C0([0, 1]), the function f is compact.
Finally, we have
f (u)(u) = −Z 1
0
u(t)2dt < 0 < Z 1
0
u(t)2dt ≤ kf (u)kE∗
for all u ∈ S.
We conclude by proposing the following problem.
PROBLEM 1.
- Let E be an infinite-dimensional reflexive real Banach space such
that, for each compact function f : S → E∗ satisfying (1), there exists x ∈ S for which
Then, does E possess the Kadec-Klee property ?
f (x)(x) = kf (x)kE∗ .
References
[1] K. FAN, Fixed-point and minimax theorems in locally convex topological linear spaces,
Proc. Nat. Acad. Sci. U.S.A., 38 (1952), 121-126.
[2] E. KLEIN and A. C. THOMPSON, Theory of correspondences, John Wiley & Sons,
1984.
[3] E. MICHAEL, Continuous selections avoiding a set, Topology Appl., 28 (1988), 195-
213.
4
|
1405.6570 | 2 | 1405 | 2015-02-11T09:15:46 | Self-Adjointness criterion for operators in Fock spaces | [
"math.FA",
"math-ph",
"math-ph"
] | In this paper we provide a criterion of essential self-adjointness for operators in the tensor product of a separable Hilbert space and a Fock space. The class of operators we consider may contain a self-adjoint part, a part that preserves the number of Fock space particles and a non-diagonal part that is at most quadratic with respect to the creation and annihilation operators. The hypotheses of the criterion are satisfied in several interesting applications. | math.FA | math | Self-Adjointness criterion for operators in Fock spaces
Marco Falconi∗
IRMAR and Centre Henri Lebesgue; Université de Rennes I
Campus de Beaulieu, 263 avenue du Général Leclerc
CS 74205, 35042 RENNES Cedex
(Dated: July 31, 2017)
In this paper we provide a criterion of essential self-adjointness for operators in the
tensor product of a separable Hilbert space and a Fock space. The class of operators
we consider may contain a self-adjoint part, a part that preserves the number of Fock
space particles and a non-diagonal part that is at most quadratic with respect to the
creation and annihilation operators. The hypotheses of the criterion are satisfied in
several interesting applications.
Keywords: Essential Self-Adjointness, Fock Spaces, Interacting Quantum Field Theories, Nelson
Hamiltonian, Pauli-Fierz Hamiltonian.
5
1
0
2
b
e
F
1
1
]
.
A
F
h
t
a
m
[
2
v
0
7
5
6
.
5
0
4
1
:
v
i
X
r
a
1.
INTRODUCTION
Let H1, H2 be separable Hilbert spaces. We consider the following space:
(1)
H = H1 ⊗ Γs(H2) ;
where Γs(K ) is the symmetric Fock space based on K [see 7, 9, 24, for mathematical presen-
tations of Fock spaces and second quantization]. The symmetric structure of the Fock space
does not play a role in the argument: in principle it is possible to formulate the same criterion
for anti-symmetric Fock spaces H1⊗Γa(H2). We focus on symmetric spaces, the corresponding
antisymmetric results should be deduced without effort.
We are interested in proving a criterion of essential self-adjointness for densely defined
operators of the form:
(2)
H = H01 ⊗ 1 + 1 ⊗ H02 + HI ;
with suitable assumptions on H01, H02 and HI. Operators based on these spaces and with such
structure are crucial in physics, to describe the quantum dynamics of interacting particles and
fields.
∗ [email protected]
2
Self-adjointness of operators in Fock spaces has been widely studied, in particular in the
context of Constructive Quantum Field Theory [e.g. 13–15, 21, 25, 26] and Quantum Electro-
Dynamics [e.g. 1, 4–6, 16, 18, 20, 22, 27]. A variety of advanced tools has been utilized, for
even "simple" systems present technical difficulties to overcome: many questions still remain
unsolved.
In some favourable situations, however, it is possible to take advantage of the peculiar
structure of the Fock space and prove essential self-adjointness with almost no effort. The
idea first appeared in a paper by Ginibre and Velo [13]; and the author utilized it in [2, 11] for
the Nelson model with cut off: essential self-adjointness can be proved with less assumptions
than using the Kato-Rellich Theorem (and that becomes particularly significative in dimen-
sion two), see Section 4.2. Another remarkable application is the Pauli-Fierz Hamiltonian
describing particles coupled with a radiation field. For general coupling constants, essential
self-adjointness has been first proved in a probabilistic setting, using stochastic integration
[17, 18]. In this paper we prove the same result directly in Section 4.3, applying the criterion
formulated in Assumptions A0, AI and Theorem 3.1.
In the literature, self-adjointness of operators in Fock spaces has been studied using various
tools of functional analysis: the Kato-Rellich and functional integration arguments mentioned
above are two examples, as well as the Nelson commutator theorem [10]. For each particular
system, a strategy is utilized ad hoc: the more complicated is the correlation between H1 and
Γs(H2), the more difficult is the strategy. We realized that, if we take suitable advantage of
the fibered structure of the Fock space, the type of interaction between the spaces is not so
relevant. This was a strong motivation to study the problem from a general perspective. Due
to the variety of possible applications, an effort has been made to formulate the necessary
assumptions in a general form. Roughly speaking, the essential requirement is that the part
of HI that does not commute with the number operator of Γs(H2) is at most quadratic with
respect to the creation and annihilation operators. As anticipated, the space H1 does not play
a particular role, as long as HI behaves sufficiently well with respect to H01.
1. Paper organization.
In Section 1.2 we introduce the notation, and recall some basic definitions of operators
in Fock spaces. In Section 2 we formulate the necessary assumptions on the operator H.
In Section 3 we prove the criterion. In Section 4 we outline some of the most interesting
3
applications. Finally in Section 5 we give some conclusive remarks, and an extension of the
criterion to semi-bounded quartic operators.
2. Definitions and notations.
• Let K be a separable Hilbert space. Then the symmetric Fock space Γs(K ) is defined
as the direct sum:
Γs(K ) =
∞Mn=0
K ⊗sn ,
where K ⊗sn is the n-fold symmetric tensor product of K , and K ⊗s0 := C.
• Let h : K ⊇ D(h) → K be a densely defined self-adjoint operator on a separable Hilbert
space K . Its second quantization dΓ(h) is the self-adjoint operator on Γs(K ) defined by
dΓ(h)D(h)⊗s n =
nXk=1
1 ⊗ ··· ⊗ h{z}k
⊗··· ⊗ 1 .
Let u be a unitary operator on K . We define Γ(u) to be the unitary operator on Γs(K )
given by
Γ(u)K ⊗s n =
nOk=1
u .
If eith is a group of unitary operators on K , Γ(eith) = eitdΓ(h).
• N := dΓ(1) the number operator of Γs(H2).
• H0 := H01 ⊗ 1 + 1 ⊗ H02; the free Hamiltonian.
• If X is a self-adjoint operator on a Hilbert space, we denote by D(X) its domain, by
qX(·,·) the form associated with X and by Q(X) the form domain.
• Let K be a Hilbert space; {K (j)}j∈N a collection of disjoint subspaces of K ; X an oper-
ator densely defined on K . We say that {K (j)}j∈N is invariant for X if ∀j ∈ N, X maps
D(X) ∩ K (j) → K (j), and D(X) ∩ K (j) is dense in K (j).
• Let K be a Hilbert space; {K (j)}j∈N a collection of disjoint closed subspaces of K such
that Lj∈N
K (j) = K . Then we call the collection complete, and we define the dense
4
subset f0(K (·)) of K as:
(3)
f0(K (·)) =nφ ∈ K ,∃n ∈ N s.t. φ ∈
nMj=0
K (j)o .
Also, we denote by 1j(K (·)) the orthogonal projection on K (j), by 1≤n(K (·)) the orthog-
j=0
K (j).
onal projection onLn
• Let K ∋ f, g be two elements of a separable Hilbert space. We define the creation a∗(f )
and annihilation a(f ) operators on Γs(K ) by their action on n-fold tensor products (with
a(f )φ0 = 0 for any φ0 ∈ K ⊗s0 = C):
a(f )g⊗n = √n hf, giK g⊗(n−1)
a∗(f )g⊗n = √n + 1 f ⊗s g⊗n .
They extend to densely defined closed operators and are adjoint of each other: we denote
again by a#(f ) their closures. For any f ∈ K , D(a∗(f )) = D(a(f )) with
D(a(f )) =nφ ∈ Γs(K ) :
∞Xn=0
(n + 1)khf (x), φn+1(x, Xn)iK (x)k2
K ⊗sn(Xn) < +∞o ,
where φn+1 = φ(cid:12)(cid:12)K ⊗sn+1; also D(a(f )) ⊃ D(dΓ(1)1/2), D(a(f )) ⊃ f0(K (·)). They satisfy
the Canonical Commutation Relations [a(f1), a∗(f2)] = hf1, f2iK on suitable domains
(e.g. f0(K (·))).
• We decompose Γs(H2) in its subspaces with fixed number of particles as usual: ∀n ∈ N,
}n∈N is a complete
2 = C. Then {H (n)
, with the convention H (0)
define H (n)
:= H ⊗sn
2
2
2
collection of closed disjoint subspaces of Γs(H2) invariant for N .
• Let X be an operator on H . We say that X is diagonal if {H1 ⊗ H (n)
for X; X is non-diagonal if for all n ∈ N and φ ∈ D(X) ∩ H1 ⊗ H (n)
}n∈N is invariant
, Xφ /∈ H1 ⊗ H (n)
.
2
2
2
2. ASSUMPTIONS ON H
In this section we discuss Assumptions A0 and AI(A′
I ). In Section 4 below they are checked
in concrete examples.
We recall that our Hilbert space H has the form
H = H1 ⊗ Γs(H2) ;
5
while the operator is
H = H01 ⊗ 1 + 1 ⊗ H02 + HI .
We separate the assumptions on H0 from the ones on HI, to improve readability. On HI
we require either Assumption AI or Assumption A′
I . In AI the non-diagonal part of HI can
be more singular: that restricts the diagonal part to be at most quadratic in the creation
and annihilation operators. In A′
I on the other hand is assumed more regularity on the non-
diagonal part of HI , allowing for a more singular diagonal part.
Assumption A0. H01 and H02 are semi-bounded self-adjoint operators. We denote respectively
by −M1 and −M2 their lower bounds. Furthermore, ∀t ∈ R, {H (n)
}n∈N is invariant for eitH02 .
2
This is quite natural. In physical systems the Hamiltonian is often split in a part describing
the free dynamics (usually a self-adjoint and positive unbounded operator), and an interaction
part. The invariance of the n-particles subspaces is also a usual feature of free quantum
theories: let h02 be a semi-bounded self-adjoint operator on the one-particle space H2; then
the second quantization dΓ(h02) is self-adjoint, and the group Γ(eith02 ) generated by it satisfies
the assumption.
Assumption AI . HI is a symmetric operator on H , with a domain of definition D(HI) such
that D(H0) ∩ D(HI ) is dense in H . Furthermore ∀φ ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H1 ⊗ H (n)
2
,
(4)
HI φ ∈
2Mi=−2
H1 ⊗ H (n+i)
2
.
Also, HI satisfies the following bound: ∀n ∈ N ∃C > 0 such that ∀ψ ∈ H , ∀φ ∈ Q(H01 ⊗ 1) ∩
Q(1 ⊗ H02) ∩ H1 ⊗ H (n)
2
:
hψ, HI φiH 2 ≤ C 2
(5)
2Xi=−2
kψn+ik2
H1⊗H (n+i)
2
h(n + 1)2kφk2
H1⊗H (n)
2
+q1⊗H02(φ, φ) + (M1 + M2 + 1)kφk2
+ (n + 1)(cid:16)qH01⊗1(φ, φ)
2 (cid:17)i ;
H1⊗H (n)
where we define ψn := 1 ⊗ 1n(H (·)
2
)ψ.
Consider Assumption AI . First of all, HI has to be sufficiently regular, i.e. relatively
bounded by H0 (in some sense) when restricted to the subspaces H1 ⊗ H (n)
require that HI is at most quadratic in the annihilation and creation operators, as reflected
. Essentially, we
2
by the n-dependence in (5).
6
Assumption A′
I . HI is a symmetric operator on H , with a domain of definition D(HI) such
that D(H0) ∩ D(HI ) is dense in H . Furthermore ∀φ ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H1 ⊗ H (n)
2
,
(6)
HI φ ∈
2Mi=−2
H1 ⊗ H (n+i)
2
.
Also, HI = Hdiag + H2 with the following properties:
i) Hdiag is diagonal; H2 is non-diagonal.
ii) Hdiag satisfies the following bound. ∀n ∈ N ∃C(n) > 0 such that ∀ψ ∈ H , ∀φ ∈ Q(H01 ⊗
1) ∩ Q(1 ⊗ H02) ∩ H1 ⊗ H (n)
2
:
(7)
hψ, HdiagφiH 2 ≤ C 2(n)kψnk2
2 (cid:16)qH01⊗1(φ, φ) + q1⊗H02(φ, φ) + (M1 + M2
2 (cid:17) .
iii) H2 satisfies the following bound. ∀n ∈ N ∃C > 0 such that ∀ψ ∈ H , ∀φ ∈ H1 ⊗ H (n)
+1)kφk2
H1⊗H (n)
H1⊗H
(n)
2
:
(8)
hψ, H2φiH ≤ C(n + 1)kφkH1⊗H (n)
2
2Xi=−2
i6=0
kψn+ikH1⊗H (n+i)
2
.
Assumption A′
I is similar to Assumption AI . However since the non-diagonal quadratic
part H2 is more regular than before, we can be less demanding on the diagonal part Hdiag: it
has still to be bounded in a suitable sense by H0, but it can be non-quadratic with respect to
the creation and annihilation operators.
Remark 2.1. In some applications, there is a decomposition of H1 invariant for H. For example,
it may happen that H1 is also a Fock space but H leaves invariant each sector with fixed
number of particles. In this situation, we can prove essential self-adjointness with little less
regularity on the assumptions. In particular, Assumption AI would be changed in:
HI is a symmetric operator on H , with a domain of definition D(HI ) such that D(H0) ∩
D(HI ) is dense in H . Furthermore there exists a complete collection {H (j)
1 ⊗ Γs(H2)}j∈N
invariant for H0 and HI such that: ∀φ ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H (j)
1 ⊗ H (n)
2
,
HI φ ∈
2Mi=−2
H (j)
1 ⊗ H (n+i)
2
.
Also, HI satisfies the following bound: ∀j, n ∈ N ∃C(j) > 0 such that ∀ψ ∈ H , ∀φ ∈
Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H (j)
1 ⊗ H (n)
2
:
7
hψ, HI φiH 2 ≤ C 2(j)
2Xi=−2
kψj,n+ik2
H (j)
1 ⊗H (n+i)
2
h(n + 1)2kφk2
H (j)
1 ⊗H (n)
2
+q1⊗H02(φ, φ) + (M1 + M2 + 1)kφk2
+ (n + 1)(cid:16)qH01⊗1(φ, φ)
2 (cid:17)i ;
1 ⊗H (n)
H (j)
where we define ψj,n := 1j(H (·)
1
) ⊗ 1n(H (·)
2
)ψ.
Theorem 3.1 would then read:
Assume A0 and AI(A′
f0(H (·)
1 ⊗ H (·)
).
2
I ). Then H is essentially self adjoint on D(H01 ⊗ 1) ∩ D(H02 ⊗ 1) ∩
3. DIRECT PROOF OF SELF-ADJOINTNESS
In this section we present the criterion of essential self-adjointness . The strategy is to
prove that Ran(H ± i) is dense in H , by an argument of reductio ad absurdum. As already
discussed, the non-diagonal part of HI is at most quadratic with respect to the annihilation and
creation operators of Γs(H2), and that plays a crucial role in the proof. We prove Theorem 3.1
assuming AI ; the other case being analogous.
Theorem 3.1. Assume A0 and AI (A′
1) ∩ H1 ⊗ f0(H (·)
).
2
I ). Then H is essentially self adjoint on D(H01⊗1)∩D(H02⊗
Proof. Let ψ ∈ H , z ∈ C with Imz 6= 0. Suppose that ∀φ ∈ D(H01 ⊗ 1) ∩ D(1 ⊗ H02) ∩ H1 ⊗
f0(H (·)
):
2
(9)
hψ, (H − z)φiH = 0 .
Then it suffices to show that ψ = 0. This is done in few steps. Let n ∈ N and φn ∈ D(H01 ⊗
. For all n ∈ N, the space Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H1 ⊗ H (n)
1) ∩ D(1 ⊗ H02) ∩ H1 ⊗ H (n)
2
2
with the scalar product:
(10)
h· , ·iXn = qH01⊗1(· , · ) + q1⊗H02(· , · ) + (M1 + M2 + 1)h· , ·iH1⊗H (n)
2
is complete, and therefore a Hilbert space. We denote it by Xn. Then (9) together with As-
sumption A0 imply, since φn ∈ D(H01 ⊗ 1) ∩ D(1 ⊗ H02):
(11)
hψn, φniXn = (z + M1 + M2 + 1)hψn, φniH1⊗H (n)
2 − hψ, HI φniH .
8
Use bound (7) and then Riesz's Lemma on Xn: it follows that ψn ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩
H1 ⊗ H (n)
for any n ∈ N.
2
). Then ∃{φ(α)}α∈N such that ∀α ∈ N,
n → φn in the topology induced
); and ∀n ∈ N, φ(α)
Let φ ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H1 ⊗ f0(H (·)
2
φ(α) ∈ D(H01 ⊗ 1) ∩ D(1⊗ H02) ∩ H1 ⊗ f0(H (·)
by k·kXn
. Furthermore ∀α ∈ N:
2
(12)
hψ, (H − z)φ(α)iH = 0 .
Since ψn ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H1 ⊗ H (n)
∀φ ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H1 ⊗ f0(H (·)
):
2
2
, we can take the limit of (12) and obtain,
(13)
qH01⊗1(ψ, φ) + q1⊗H02(ψ, φ) + hψ, HI φiH = zhψ, φiH .
Hence we can choose φ = ψ≤n := 1 ⊗ 1≤n(H (·)
2
taking the imaginary part we obtain:
)ψ in (13). Then, using Assumption A0 and
(14)
Im(z)hψ≤n, ψ≤ni = Im(hψ − ψ≤n, HIψ≤ni) .
Now, by Assumption AI (the equality holds on the suitable domain):
HI(cid:0)1 ⊗ 1≤n(H (·)
2
)(cid:1) =(cid:0)1 ⊗ 1≤n+2(H (·)
2
)(cid:1)HI(cid:0)1 ⊗ 1≤n(H (·)
2
)(cid:1) .
Furthermore 1 ⊗ 1≤n+2(H (·)
2
)(ψ − ψ≤n) = ψn+1 ⊕ ψn+2. Then Equation (14) becomes:
Im(z)hψ≤n, ψ≤ni =
2Xi=1
Im(hψn+i, HI ψ≤ni) .
Using the symmetry of HI , and (4) we obtain:
(15)
Im(z)hψ≤n, ψ≤ni = Im(hψn+2, HIψni + hψn+1, HI ψni + hψn+1, HI ψn−1i) .
Now bound (15) using (5); then we obtain ∀n ∈ N:
Imz
nXi=0
(16)
kψik2 ≤ Chkψn+1k(cid:16)(n + 1)(cid:0)kψnk + kψn−1k(cid:1) + √n + 1(cid:0)kψnkXn + kψn−1kXn−1(cid:1)(cid:17)
+kψn+2k(cid:16)(n + 1)kψnk + √n + 1kψnkXn(cid:17)i
Xn+i2i .
kψn+i1k2 +
(n + 1)−1kψn+i2k2
≤ 2C(n + 1)h 2Xi1=0
0Xi2=−1
For all α > 0 define:
S :=
∞Xn=0
kψnk2 ; Sα :=
∞Xn=0
(n + α)−1kψnk2
Xn .
9
ψ ∈ H , hence S is finite. We prove that also Sα is finite. Using equation (13) with φ = ψn we
obtain, for all n ∈ N:
(17)
(n + α)−1kψnk2
Xn = (n + α)−1(z + M1 + M2 + 1)kψnk2 − (n + α)−1hψ, HI ψni .
Now, we can use bound (5) on (n + α)−1hψ, HI ψni, obtaining
(18)
2Xi=−2
≤ C 2(α)
(n + α)−2hψ, HI ψniH 2 ≤ C 2(n + α)−2
kψn+ik2h(n + 1)2kψnk2 + (n + 1)kψnk2
Xni
kψn+ik2hkψnk2 + (n + α)−1kψnk2
Xni ,
2Xi=−2
for some C(α) > 0. The only terms we need to deal with are (n + α)−1kψn+ik2kψnk2
the fact that for any ε, a, b > 0, ab ≤ 1
(n + α)−1kψn+ik2kψnk2
εkψn+ik4(cid:17) .
2 (εa2 + 1
ε b2), obtaining
1
2(cid:16)ε(n + α)−2kψnk4
Xn ≤
Xn +
(19)
1
. We use
Xn
Combining (19) with (18), and applying to Equation (17), we obtain the following bound: for
all ε, α > 0, ∃C(α, ε) > 0 such that
(20)
(n + α)−1kψnk2
Xn ≤ C(α, ε)
2Xi=−2
kψn+ik2 + ε(n + α)−1kψnk2
Xn .
Fix ε < 1, then for all α > 0, ∃C(α) > 0 such that ∀¯n ∈ N:
(21)
¯nXn=0
(n + α)−1kψnk2
Xn ≤ C(α)S ;
uniformly in ¯n. Then we can take the limit ¯n → ∞ and obtain Sα < ∞.
Remark. The bound of Equation (20) could seem to follow from an implicit smallness condition
on the interaction HI. As it will become clearer with the examples of Section 4, it is not the case.
Roughly speaking, Assumption AI allows for interaction parts that are at most as singular as
(H0 + M1 + M2)1/2(N + 1)1/2.
Now return to Equation (16). There exists n∗ ∈ N such that ∀n ≥ n∗:
1
2
S ≤
nXi=0
kψik2 ≤ S .
Hence summing in n∗ ≤ n ≤ ¯n on both sides of (16) we obtain for all ¯n > n∗:
1
2
S
¯nXn=n∗
(n + 1)−1 ≤
¯nXn=n∗
(n + 1)−1
nXi=0
kψik2 ≤ 2
C
Imz
(3S + S1 + S2) .
The bound on the right hand side is uniform in ¯n: that is absurd, unless S = S1 = S2 = 0 ⇔
ψ = 0.
10
Once essential self-adjointness is established, it is possible to give the following character-
ization of the domain of self-adjointness D(H).
Proposition 3.2. Assume A0 and AI (A′
I ). If exists K self-adjoint operator with domain D(K)
such that:
i) D(H0) ∩ D(K) is dense in H ; H1 ⊗ f0(H (·)
2
) is dense in D(K).
ii) There exists 0 < ε < 1 such that ∃C(ε) > 0, ∀φ ∈ D(H0) ∩ D(K):
kHI φk ≤ εkH0φk + C(ε)(kKφk + kφk) .
(22)
Then D(H) ∩ D(K) = D(H0) ∩ D(K).
Proof. Using bound (22), we have ∀φ ∈ D(H0) ∩ D(K):
(23)
kHφk ≤ (ε + 1)kH0φk + C(ε)(kKφk + kφk) .
Then D(H) ⊇ D(H0) ∩ D(K). Now let φ ∈ D(H) ∩ D(K): using (22)
(24)
kH0φk ≤ εkH0φk + kHφk + C(ε)(kKφk + kφk) ;
since ε < 1, D(H0) ⊇ D(H) ∩ D(K).
4. APPLICATIONS
It is possible to apply Theorem 3.1 in several situations of mathematical and physical in-
terest. We present and discuss some of them in this section; not before a brief discussion of the
"boundaries" of Theorem 3.1: it may be interesting to see how its proof fails when we consider
operators that are more than quadratic in the annihilation/creation operators; and to define a
quadratic operator that is not sufficiently regular for Assumption AI (A′
I ) to hold. According to
this purpose, we will consider simple toy models on Γs(C). We denote by a# the corresponding
annihilation/creation operators.
Let's consider a simple trilinear Hamiltonian on Γs(C):
H3 = a∗a + a∗a∗a∗ + aaa .
The free part is H0 = a∗a, and the interaction part is HI = a∗a∗a∗ + aaa. Assumption A0 is
satisfied, and Assumption A′
I is slightly modified: i now ranges from −3 to 3, and bounds (7)
and (8) are replaced by the simple bound:
hψ, HI φiH ≤ C(n + 1)3/2kφkH (n)(cid:0)kψn+3kH (n+3) + kψn−3kH (n−3)(cid:1) .
The proof of Theorem 3.1 carries on, almost unchanged, up to Equation (16) that would now
read
Imz
nXi=0
kψik2 ≤ C(n + 1)3/2kψn+3k2 .
11
However if we now take the sum in n from n∗ to ¯n (where n∗ is such that 1
i=0kψik2 ≤
n=0(n +
1)−3/2 converges. Hence the proof fails, and analogously would fail for any higher order poly-
kψk2 for all n ≥ n∗) we cannot conclude that kψk must be zero, because the series P∞
2kψk2 ≤Pn
nomial of the annihilation/creation operators.
On the other hand, we introduce now a quadratic model for which Assumption AI(A′
I ) fails
to hold, and thus Theorem 3.1 cannot be applied. For the following operator on L2(R) ⊗ Γs(C)
Assumption AI is satisfied:
H∂a = −∂2
x + a∗a − i∂x(a∗ + a) + a∗a∗ + aa ,
where H0 = −∂2
is coupled with the quadratic term
x + a∗a and HI = −i∂x(a∗ + a) + a∗a∗ + aa. If, however, the derivative operator
H∂aa = −∂2
x + a∗a − i∂x(a∗a∗ + aa) ,
AI (A′
I) is no longer satisfied. The interaction in this case would be of type H 1/2
0 N , and therefore
too singular: Theorem 3.1 does not hold for H∂aa.
Throughout the section we will adopt the following notations, in addition to the ones of
Section 1.2. Let K a Hilbert space; we denote by L(K ) the set of bounded operators on K
and by ·L(K ) the operator norm. It is also useful to define the annihilation/creation operator
valued distributions a#(x), x ∈ Rd. Let f ∈ L2(Rd), a#(f ) the annihilation/creation operators
on Γs(L2(Rd)). Then the operator valued distributions a#(x) acting on L2(Rd), with values on
Γs(L2(Rd)), are defined by:
(a∗, f ) ≡ZRd
a∗(x)f (x)dx := a∗(f ) ; (a, f ) ≡ZRd
a(x) ¯f (x)dx := a(f ) .
They satisfy the commutation relations (inherited by the CCR) [a(x), a∗(y)] = δ(x − y).
1. Hamiltonians of identical bosons.
The criterion applies to operators in the Fock space Γs(K ), for any separable Hilbert space
K . Simply choose H1 ≡ C and H2 ≡ K ; then C ⊗ Γs(K ) ≈ Γs(K ) up to an unitary isomor-
phism.
12
(25)
An example is given by the following class of operators. Let K = L2(Rd); h0 a positive
self adjoint operator on L2(Rd) (the one-particle free Hamiltonian). Furthermore, let V1 ∈
L2(Rd), V2, V3 ∈ L2(R2d), with V2 = V 2, and V4(·) : Rd → R, such that V4(x) = V4(−x) and
V4 (h0 + 1)−1/2 ∈ L(L2(Rd)). Consider
H = dΓ(h0) +ZRd(cid:16)V1(x)a∗(x) + V 1(x)a(x)(cid:17)dx +ZR2d(cid:16)V2(x, y)a∗(x)a(y) + V3(x, y)a∗(x)
V4(x − y)a∗(x)a∗(y)a(x)a(y)dxdy .
We make the following identifications: H01 ≡ 0, H02 ≡ dΓ(h0), Hdiag ≡ R (V4a∗a∗aa + V2a∗a),
H2 ≡R (V1a∗ + V 1a) +R (V3a∗a∗ + V 3aa). Assumption A0 is trivial to verify; and Assumption A′
s(Rnd) ∩ Q(dΓ(h0)),
follows from standard estimates on Fock space: let ψ ∈ Γs(L2(Rd)), φn ∈ L2
n ∈ N, then
a∗(y) + V 3(x, y)a(x)a(y)(cid:17)dxdy +
1
2ZR2d
I
(26)
(27)
hψ, Hdiagφni ≤(cid:16)nkV2k2kφnk + V4 (h0 + 1)−1/2L(L2(Rd))(cid:0)n3/2k(dΓ(h0))1/2φnk
+n2kφnk(cid:1)(cid:17)kψnk ;
2Xi=−2
kψn+ik .
hψ, H2φni ≤ 2(cid:16)√n + 1kV1k2 + (n + 1)kV3k2(cid:17)kφnk
i6=0
Hence we can apply Theorem 3.1; and prove essential self-adjointness of H in D(dΓ(h0)) ∩
f0(L2(Rd)(·)). We can also apply Proposition 3.2 with K ≡ N 3, i.e. D(H)∩D(N 3) = D(dΓ(h0))∩
D(N 3). Observe that if d = 3, the well-known many body Hamiltonian with Coulomb pair
interaction
HC = dΓ(−∆) ±
1
2ZR6
1
x − y
is just the special case h0 = −∆, V1 = V2 = V3 = 0 and V4 = ±x−1.
a∗(x)a∗(y)a(x)a(y)dxdy ,
2. Nelson-type Hamiltonians.
We consider now the dynamics of different species of particles (or fields) interacting. A
typical example is the Nelson Hamiltonian. It was introduced in a rigorous way by Nelson
[20] to describe nucleons in a meson field, and studied by several authors [e.g. 1, 8, 10, 12].
Let H = L2(Rpd) ⊗ Γs(L2(Rd)): the first space corresponds to n non-relativistic particles;
the second to a scalar relativistic field. Let ω be a positive self-adjoint operator on L2(Rd) (the
loc(Rd, R+) an external potential acting on
dispersion relation of the relativistic field), V ∈ L2
the particles. The interaction between the particles and the field is linear in the creation and
annihilation operators a# corresponding to the field. Let v : R2d → C such that
13
• (1 − ∆x)−1/2kv(x,·)k2
L2
(k)(Rd)(1 − ∆x)−1/2 ∈ L(L2
(x)(Rd));
• for all k ∈ Rd, v(x, k)(1 − ∆x)−1/2 ∈ L(L2
(x)(Rd)), with v(x,·)(1 − ∆x)−1/2L(L2
(x)(Rd)) ∈
L2
(k)(Rd).
Then we define the Nelson Hamiltonian:
(28)
HN =(cid:16) pXi=1
−∆xi + V (xi)(cid:17) ⊗ 1 + 1 ⊗ dΓ(ω) +
pXi=1
a∗(v(xi,·)) + a(v(xi,·)) .
The function v describes the coupling between the particles and the relativistic field. The
assumptions above imply that it has a good behaviour both for high and small momenta; in
particular in three-dimensions it acts as an UV cutoff function.
Remark. The model of Nelson [20] was much more specific: d = 3, ω(k) =pk2 + µ2 with µ > 0,
V = 0 and v(x, k) = λ(2π)−3/2(2ω(k))−1/2e−ik·x1 · ≤σ(k) with λ, σ > 0. With these assumptions,
v ∈ L∞(R3, L2(R3)), ω−1/2v ∈ L∞(R3, L2(R3)); then HN (the Nelson model with UV cut off) is
self-adjoint by the Kato-Rellich Theorem. However, if we consider d = 2 and µ = 0 (massless
relativistic field), the Kato-Rellich Theorem is not applicable because ω−1/2v /∈ L∞(R2, L2(R2))
due to an infrared divergence. Instead assumptions A0 and A′
I are still satisfied, thus Theo-
rem 3.1 can be used.
In order to check Assumptions A0 and AI on (28), we make the (straightforward) identifica-
tions: H1 ≡ L2(Rpd), H2 ≡ L2(Rd), H01 ≡Pi −∆xi + V (xi), H02 ≡ dΓ(ω), HI ≡Pi a∗(v(xi,·)) +
a(v(xi,·)). We do not need to introduce a decomposition of H1. Assumption A0 is satisfied: for
loc(Rd, R+), −∆ + V (·) is a positive self-adjoint operator, and the vectors with fixed
all V ∈ L2
number of particles are invariant for the evolution associated with the positive self-adjoint op-
erator dΓ(ω). Furthermore, since H01⊗ 1 and 1⊗ H02 are positive self-adjoint commuting oper-
ators, H0 is a positive self-adjoint operator with domain D(H0) = D(H01 ⊗ 1)∩ D(1⊗ H02). As-
sumption AI is also satisfied by usual estimates: ∀ψ ∈ H , ∀φn ∈ L2(Rpd)⊗L2
s(Rnd)∩Q(H01⊗1),
n ∈ N,
(29)
hψ, HI φni ≤p2p(cid:0)2√nkv(x,·)(1 − ∆x)−1/2L(L2
(x))(cid:1)(cid:16)(cid:13)(cid:13)(cid:13)(cid:0) pXi=1
(1 − ∆x)−1/21/2
L(L2
(x))kL2
(k)
+ (1 − ∆x)−1/2kv(x,·)k2
L2
(k)
−∆xi(cid:1)1/2φn(cid:13)(cid:13)(cid:13) + √pkφnk(cid:17) 1Xi=−1
i6=0
kψn+ik .
Then HN is essentially self-adjoint on D(H0) ∩ f0(L2(Rpd) ⊗ L2(Rd)(·)).
14
(30)
Γs(L2(Rd)) ⊗ Γs(L2(Rd)) in the following way. Define
eHN = dΓ(−∆ + V ) ⊗ 1 + 1 ⊗ dΓ(ω) +ZRd
ψ∗(x)(cid:0)a∗(v(x,·)) + a(v(x,·))(cid:1)ψ(x)dx ,
Let HNs be the restriction of HN to L2
s(Rpd) ⊗ Γs(L2(Rd)). It is possible to extend HNs to
where ψ# are the creation and annihilation operators corresponding to the first Fock space.
Then HNs and eHN agree on the p-particle sector L2
s(Rpd) ⊗ Γs(L2(Rd)) of Γs(L2(Rd)) ⊗
Γs(L2(Rd)). The self-adjointness of eHN still follows from Theorem 3.1 using the bound (29):
it is sufficient to choose for H1 ≡ Γs(L2(Rd)) the decomposition in finite particle vectors
{H (j)
s(Rjd) ⊗ Γs(H2)}j∈N. Let H0 ≡ dΓ(−∆ + V ) ⊗ 1 + 1⊗ dΓ(ω), then the
1 ⊗ Γs(H2)}j∈N ≡ {L2
domain of essential self-adjointness for eHN is D(H0)∩ f0(L2(Rd)(·) ⊗ L2(Rd)(·)). Let N1 and N2
applying Proposition 3.2 we also obtain D(eHN ) ∩ D(N 2
be the number operators corresponding to the first and second Fock space respectively. Then
2 ) = D(H0) ∩ D(N 2
1 + N 2
1 + N 2
2 ).
3. Pauli-Fierz Hamiltonian.
The last example considered is an operator describing the dynamics of rigid charges and
their radiation field interacting. The model was introduced by Pauli and Fierz [23], and has
been extensively studied by a mathematical standpoint. See Spohn [27, and references thereof
contained] for a detailed presentation.
Let H (spin) = (⊗pC2[ d
2 ])⊗L2(Rpd)⊗Γs(Cd−1⊗L2(Rd)), H = L2(Rpd)⊗Γs(Cd−1⊗L2(Rd)): the
first space corresponds to p spin- 1
2 particles, the second to spinless particles. Let χ ∈ L2(Rd),
loc(Rpd, R+), ω = k, mj > 0, qj ∈ R for all j = 1, . . . , p. Furthermore, let eλ : Rd → Rd
V ∈ L2
such that for almost all k ∈ Rd, k · eλ(k) = 0 and eλ(k) · eλ′(k) = δλλ′ for all λ, λ′ = 1, . . . , d − 1.
Then we define the electromagnetic vector potential in the Coulomb gauge as
(31)
A(x) =
ZRd
d−1Xλ=1
eλ(k)(cid:16)a∗
λ(k)χ(k)eik·x + aλ(k) ¯χ(k)eik·x(cid:17)dk ;
where a#
λ are the creation and annihilation operators of Γs(Cd−1⊗L2(Rd)) satisfying the canon-
λ′ (k′)] = δλλ′δ(k − k′); the (spinless) Pauli-Fierz Hamilto-
ical commutation relations [aλ(k), a∗
nian on H is then
(32) HP F =
pXj=1
1
2mj(cid:0)−i∇j ⊗ 1 + qjA(xj)(cid:1)2 + V (x1, . . . , xp) ⊗ 1 + 1 ⊗
ZRd
d−1Xλ=1
ω(k)a∗
λ(k)aλ(k)dk .
The function χ plays the role of an ultraviolet cut off in the interaction, and is usually inter-
preted as the Fourier transform of the particles' charge distribution. Let {σ(µ)}d
µ=1 the 2[ d
2 ]×2[ d
2 ]
15
, j = 1, . . . , p the operator
2 ]) acting as σ(µ) on the j-th space of the tensor product. Then the spin- 1
2 Pauli-Fierz
matrices satisfying σ(µ)σ(ν) + σ(ν)σ(µ) = 2δµν Id. Also, denote by σ(µ)
on (⊗pC2[ d
Hamiltonian on H (spin) = (⊗pC2[ d
2 ]) ⊗ H can be written as:
pXj=1
j ⊗(cid:16)∂(µ)
qj X1≤µ<ν≤d
H (spin)
P F = 1 ⊗ HP F +
σ(µ)
j σ(ν)
(33)
i
2
j
j A(ν)(xj) − ∂(ν)
j A(µ)(xj)(cid:17) ;
where A(µ)(x) is the µ-th component of the vector A(x).
The quadratic form corresponding to the Pauli-Fierz Hamiltonian is bounded from below,
so it is possible to define at least one self-adjoint extension by means of the Friedrichs Exten-
sion Theorem. This type of information is not completely satisfactory, since infinitely many
extensions may exist, each one dictating a different dynamics for the system. For small val-
ues of the ratios q2
j /mj between charge and mass of the particles, and if χ, χ/√ω ∈ L2(Rd), a
unique self-adjoint extension is given by KLMN Theorem. For arbitrary values of the ratios
j /mj, it is possible to prove essential self-adjointness of both HP F and H (spin)
q2
(for the spin op-
erator we need in addition ωχ ∈ L2(Rd)) by means of Theorem 3.1, under the sole assumption
χ ∈ L2(Rd). As discussed in Section 1, an analogous result (on a slightly different domain)
has been obtained with an argument of functional integration by Hiroshima [18]. If the de-
P F
pendence on x of A(x) is more general, functional integration methods may not be applicable;
however Theorem 3.1 still holds.
In the following discussion we will focus on a simplified model, for the sake of clarity.
Assumptions A0 and AI are checked on HP F with p = 1, m = 1/2 and q = −1, i.e.: H ≡
L2(Rd) ⊗ Γs(Cd−1 ⊗ L2(Rd)) and
(34)
H ≡(cid:0)i∇x ⊗ 1 + A(x)(cid:1)2 + V (x) ⊗ 1 + 1 ⊗
ZRd
d−1Xλ=1
ω(k)a∗
λ(k)aλ(k)dk .
Observe that, since we are in the Coulomb gauge, ∇x·A(x) = 0 hence [−i∇x⊗1, A(x)] = 0 on
a suitable dense domain. Rewrite H in the following form, to identify the free and interaction
parts:
(35)
H =(cid:0)−∆x + V (x)(cid:1) ⊗ 1 + 1 ⊗
ZRd
We identify H01 ≡ −∆ + V , H02 ≡ PλRRd ωa∗
d−1Xλ=1
λaλ and HI ≡ 2iA · (∇ ⊗ 1) + A2. Assumption A0
is satisfied, as in the Nelson model (28) above. For the interaction part, we have the following
ω(k)a∗
λ(k)aλ(k)dk + 2iA(x) · (∇x ⊗ 1) + A2(x) .
16
bounds: ∀ψ ∈ H , ∀φn ∈ L2(Rd) ⊗ (Cd−1 ⊗ L2(Rd))⊗sn ∩ Q(H01 ⊗ 1), n ∈ N,
1Xi=−1
hψ,A(x) · (∇x ⊗ 1)φni ≤p2(d − 1)kχk2√n + 1k(∇x ⊗ 1)φnk
(36)
hψ,A2(x)φni ≤ 2(d − 1)kχk2(n + 1)kφnk
2Xi=−2
kψn+ik .
kψn+ik ;
i6=0
Hence Assumption AI is satisfied. Then H is essentially self-adjoint on D(H0) ∩ f0(L2(Rd) ⊗
(Cd−1 ⊗ L2(Rd))(·)).
Remark. Neither non-negativity of the Pauli-Fierz operator nor smallness of the coupling con-
stant are necessary to prove essential self-adjointness by means of Theorem 3.1. Using oper-
ator methods (commutator estimates), self-adjointness of HP F with V = 0 has been proved for
general coupling constants in [16], but the non-negativity was needed to associate a unique
self-adjoint operator to the quadratic form. Theorem 3.1 relies on different assumptions, and
takes advantage of the fibered structure of the Fock space: boundedness from below of the oper-
ator is, in general, not necessary. In fact, the Hamiltonians considered in Sections 4.1 and 4.2
are possibly unbounded from below, as well as the following extension (37) of the Pauli-Fierz
Hamiltonian to infinite degrees of freedom (for the particles). As outlined in Section 5, if we
assume boundedness from below, Theorem 3.1 can be extended to operators quartic in the
creation/annihilation operators (see Assumptions BH , BI and Theorem 5.1).
Let mj = 1/2, qj = −1 and V = Pp
i=1 Vext(xi) + Pi<j Vpair(xi − xj) such that Vext ∈
loc(Rd, R+), Vpair(x) = Vpair(−x) and Vpair(1 − ∆)−1/2 ∈ L(L2(Rd)). Under these assump-
L2
s(Rpd) ⊗ Γs(Cd−1 ⊗ L2(Rd)). The physical
tions define HP Fs as the restriction of (32) to L2
interpretation is a system of p identical bosonic charges subjected to an external potential,
interacting via pair interaction and with their radiation field. As we did for the Nelson model
in (30), we can extend HP Fs to Γs(L2(Rd)) ⊗ Γs(Cd−1 ⊗ L2(Rd)):
ψ∗(x)n(cid:0)i∇x ⊗ 1 + A(x)(cid:1)2 + Vext(x)oψ(x)dx +
1
ψ(x)ψ(y)dxdy + 1 ⊗
Vpair(x − y)ψ∗(x)ψ∗(y)
ZRd
eHP F =ZRd
ω(k)a∗
λ(k)aλ(k)dk .
(37)
2ZR2d
d−1Xλ=1
We would like to prove essential self-adjointness by means of Theorem 3.1. Identify H01 ≡
R ψ∗(−∆ + Vext)ψ; H02 ≡PλRRd ωa∗
2R Vpairψ∗ψ∗ψψ; and
{H (j)
s(Rjd) ⊗ Γs(Cd−1 ⊗ L2(Rd))}j∈N. Then Assumptions A0 and AI are
1 ⊗ Γs(H2)}j∈N ≡ {L2
satisfied using bounds analogous to (36) and (26) (for Vpair), for each fixed j ∈ N. Hence eHP F
is essentially self-adjoint on D(H01 ⊗ 1) ∩ D(1 ⊗ H02) ∩ f0(L2(Rd)(·) ⊗ (Cd−1 ⊗ L2(Rd))(·)).
λaλ; HI ≡R ψ∗(2iA · (∇ ⊗ 1) + A2)ψ + 1
5. CONCLUSIVE REMARKS
17
The examples of the preceding section are not exhaustive: we focused on them because
of their relevance in physical and mathematical literature. The application to operators on
curved space-time, or to anti-symmetric systems may also lead to results of interest.
The Assumptions A0, AI and A′
I are easy to check: in the examples above follow from ba-
sic estimates of creation and annihilation operators. The proof of Theorem 3.1 itself is not
complicated, and relies on the direct sum decomposition of Γs(H2) and the structure of the
interaction with respect to the latter. Hence this criterion gives, in our opinion, a simple yet
powerful tool to prove essential self-adjointness in Fock spaces, tailored to take maximum
advantage of their structure.
If we assume that H is bounded from below, we can take inspiration from Masson and
McClary [19] and extend our criterion to accommodate quartic operators. The modified as-
sumptions and theorem would then read:
Assumption BH . H is a densely defined symmetric operator on H = H1 ⊗ Γs(H2) bounded
from below. H01 and H02 are self-adjoint operators bounded from below such that ∀t ∈ R,
{H (n)
}n∈N is invariant for eitH02 .
2
Assumption BI. HI is a symmetric operator on H , with a domain of definition D(HI )
such that D(H0) ∩ D(HI ) is dense in H . Furthermore exists a complete collection {H (j)
1 ⊗
1 ⊗ H (n)
Γs(H2)}j∈N invariant for H0 and HI such that: ∀φ ∈ Q(H01 ⊗ 1) ∩ Q(1 ⊗ H02) ∩ H (j)
,
2
(38)
HI φ ∈
4Mi=−4
H (j)
1 ⊗ H (n+i)
2
.
Also, HI satisfies the following bound: ∀j, n ∈ N ∃C(j) > 0 such that ∀ψ ∈ H , ∀φ ∈ Q(H01 ⊗
1) ∩ Q(1 ⊗ H02) ∩ H (j)
1 ⊗ H (n)
2
:
(39)
H (j)
1 ⊗H (n+i)
4Xi=−4
kψj,n+ik2
h(n + 1)4kφk2
hψ, HI φiH 2 ≤ C 2(j)
+ (n + 1)2(cid:16)qH01⊗1(φ, φ)
2 (cid:17)i .
Theorem 5.1. Assume BH and BI . Then H is essentially self adjoint on D(H01 ⊗ 1) ∩ D(H02 ⊗
1) ∩ f0(H (·)
+q1⊗H02(φ, φ) + (M1 + M2 + 1)kφk2
1 ⊗ H (·)
H (j)
1 ⊗H (n)
2
).
2
2
H
(j)
1 ⊗H
(n)
18
Remark. An attempt to extend the results of [19] can be found in [3]. Theorem 5.1 is a gen-
eralization of both: it can be applied to more singular situations and a more general class of
spaces.
The proof of Theorem 3.1 can be adapted to Theorem 5.1, making use of the inferior bound
for H. We remark that Assumption BH, by itself, implies that H has at least one self-adjoint
extension: it may be tricky to prove for general operators. Theorem 5.1 essentially states
that for regular enough quartic interactions, existence of a particular self-adjoint extension
(the Friedrichs one) is equivalent to its uniqueness. It may have interesting applications in
CQFT: e.g. the d-dimensional (bounded from below) Yd and (λϕ(x)4)d models with cut offs
have interactions that are at most quartic and regular. It is our hope that the ideas utilized
in this paper could contribute to improve the mathematical insight on interacting quantum
field theories, and could be developed to study self-adjointness of more singular systems.
ACKNOWLEDGMENTS
This work has been supported by the Centre Henri Lebesgue (programme "Investissements
d'avenir" - ANR-11-LABX-0020-01). The author would like to thank Giorgio Velo, that has
suggested to him the idea of a direct proof of self-adjointness on Fock spaces. Also, he would
like to thank Zied Ammari and Francis Nier for precious advices and stimulating discussions
during the redaction of the paper.
[1] Z. Ammari. Asymptotic completeness for a renormalized nonrelativistic Hamiltonian in quantum
field theory: the Nelson model. Math. Phys. Anal. Geom., 3(3):217–285, 2000. ISSN 1385-0172.
doi:10.1023/A:1011408618527. URL http://dx.doi.org/10.1023/A:1011408618527.
[2] Z. Ammari and M. Falconi. Wigner measures approach to the classical limit of the Nelson model:
Convergence of dynamics and ground state energy. J. Stat. Phys., 157(2):330–362, 10 2014. doi:
10.1007/s10955-014-1079-7. URL http://dx.doi.org/10.1007/s10955-014-1079-7.
[3] A. Arai.
A theorem on essential self-adjointness with application to hamiltonians in
nonrelativistic quantum field theory.
J. Math. Phys., 32(8):2082–2088, 1991.
URL
http://dx.doi.org/10.1063/1.529178.
[4] V. Bach, J. Fröhlich, and I. M. Sigal. Quantum electrodynamics of confined nonrelativistic par-
ticles. Adv. Math., 137(2):299–395, 1998. ISSN 0001-8708. doi:10.1006/aima.1998.1734. URL
http://dx.doi.org/10.1006/aima.1998.1734.
19
[5] V. Bach, J. Fröhlich, and I. M. Sigal. Spectral analysis for systems of atoms and molecules coupled
to the quantized radiation field. Comm. Math. Phys., 207(2):249–290, 1999. ISSN 0010-3616. doi:
10.1007/s002200050726. URL http://dx.doi.org/10.1007/s002200050726.
[6] V. Bach, J. Fröhlich, I. M. Sigal, and A. Soffer. Positive commutators and the spectrum of Pauli-
Fierz Hamiltonian of atoms and molecules. Comm. Math. Phys., 207(3):557–587, 1999. ISSN
0010-3616. doi:10.1007/s002200050737. URL http://dx.doi.org/10.1007/s002200050737.
[7] J. M. Cook. The mathematics of second quantization. Proc. Nat. Acad. Sci. U. S. A., 37:417–420,
1951. ISSN 0027-8424.
[8] J. Dereziński and C. Gérard. Asymptotic completeness in quantum field theory. Massive
Pauli-Fierz Hamiltonians. Rev. Math. Phys., 11(4):383–450, 1999.
ISSN 0129-055X.
doi:
10.1142/S0129055X99000155. URL http://dx.doi.org/10.1142/S0129055X99000155.
[9] J. Dereziński and C. Gérard. Mathematics of quantization and quantum fields.
Cam-
bridge Monographs
on Mathematical Physics. Cambridge University Press, Cam-
bridge, 2013.
ISBN 978-1-107-01111-3.
doi:10.1017/CBO9780511894541.
URL
http://dx.doi.org/10.1017/CBO9780511894541.
[10] J. Dereziński and V. Jakšić.
Spectral
theory of Pauli-Fierz operators.
J. Funct.
Anal., 180(2):243–327, 2001.
ISSN 0022-1236.
doi:10.1006/jfan.2000.3681.
URL
http://dx.doi.org/10.1006/jfan.2000.3681.
[11] M. Falconi. Classical limit of the Nelson model. PhD thesis, Dottorato di Ricerca in Matematica
XXIV ciclo, Università di Bologna, 2012.
[12] C. Gérard, F. Hiroshima, A. Panati, and A. Suzuki. Infrared problem for the Nelson model on static
space-times. Comm. Math. Phys., 308(2):543–566, 2011. ISSN 0010-3616. doi:10.1007/s00220-
011-1289-7. URL http://dx.doi.org/10.1007/s00220-011-1289-7.
[13] J. Ginibre and G. Velo. Renormalization of a quadratic interaction in the Hamiltonian formalism.
Comm. Math. Phys., 18:65–81, 1970. ISSN 0010-3616.
[14] J. Glimm. Yukawa coupling of quantum fields in two dimensions. I. Comm. Math. Phys., 5:343–
386, 1967. ISSN 0010-3616.
[15] J. Glimm and A. Jaffe. Collected papers. Vol. 2. Birkhäuser Boston, Inc., Boston, MA, 1985. ISBN
0-8176-3272-7. Constructive quantum field theory. Selected papers, Reprint of articles published
1968–1980.
[16] D. Hasler and I. Herbst. On the self-adjointness and domain of Pauli-Fierz type Hamiltonians.
Rev. Math. Phys., 20(7):787–800, 2008. ISSN 0129-055X. doi:10.1142/S0129055X08003389. URL
http://dx.doi.org/10.1142/S0129055X08003389.
[17] F. Hiroshima. Essential self-adjointness of translation-invariant quantum field models for ar-
bitrary coupling constants. Comm. Math. Phys., 211(3):585–613, 2000. ISSN 0010-3616. doi:
10.1007/s002200050827. URL http://dx.doi.org/10.1007/s002200050827.
[18] F. Hiroshima. Self-adjointness of the Pauli-Fierz Hamiltonian for arbitrary values of coupling
20
constants. Ann. Henri Poincaré, 3(1):171–201, 2002. ISSN 1424-0637. doi:10.1007/s00023-002-
8615-8. URL http://dx.doi.org/10.1007/s00023-002-8615-8.
[19] D. Masson and W. K. McClary. On the self-adjointness of the (g(x)φ4)2 hamiltonian. Comm. Math.
Phys., 21(1):71–74, 1971. URL http://projecteuclid.org/euclid.cmp/1103857260.
[20] E. Nelson. Interaction of nonrelativistic particles with a quantized scalar field. J. Math. Phys., 5
(9):1190–1197, 1964. doi:10.1063/1.1704225. URL http://link.aip.org/link/?JMP/5/1190/1.
[21] E. Nelson. A quartic interaction in two dimensions. In Mathematical Theory of Elementary Par-
ticles (Proc. Conf., Dedham, Mass., 1965), pages 69–73. M.I.T. Press, Cambridge, Mass., 1966.
[22] T. Okamoto and K. Yajima.
Complex scaling technique in nonrelativistic massive QED.
Ann. Inst. H. Poincaré Phys. Théor., 42(3):311–327, 1985.
ISSN 0246-0211.
URL
http://www.numdam.org/item?id=AIHPB_1985__42_3_311_0.
[23] W. Pauli and M. Fierz. Zur theorie der emission langwelliger lichtquanten. Il Nuovo Cimento, 15
(3):167–188, 1938.
[24] M. Reed and B. Simon. Methods of modern mathematical physics. II. Fourier analysis, self-
adjointness. Academic Press, New York, 1975.
[25] L. Rosen. A λφ2n field theory without cutoffs. Comm. Math. Phys., 16:157–183, 1970. ISSN
0010-3616.
[26] I. Segal. Construction of non-linear local quantum processes. I. Ann. of Math. (2), 92:462–481,
1970. ISSN 0003-486X.
[27] H. Spohn. Dynamics of charged particles and their radiation field. Cambridge Univer-
sity Press, Cambridge, 2004.
ISBN 0-521-83697-2. doi:10.1017/CBO9780511535178. URL
http://dx.doi.org/10.1017/CBO9780511535178.
|
1809.02696 | 1 | 1809 | 2018-09-07T22:07:35 | B*-algebras over ultranormed fields | [
"math.FA",
"math.RA"
] | This article is devoted to the investigation of $B^*$-algebras, dual and annihilator ultranormed algebras. Their structure is studied in the paper. Extensions of algebras and fields are considered and using them core radicals and radicals are investigated. Moreover, for this purpose also $*$-algebras and finely regular algebras are studied. Relations with operator theory and realizations of these algebras by operator algebras are outlined. | math.FA | math |
B∗-algebras over ultranormed fields.
S.V. Ludkowski.
17 May 2018
Abstract
This article is devoted to the investigation of B∗-algebras, dual
and annihilator ultranormed algebras. Their structure is studied in the
paper. Extensions of algebras and fields are considered and using them
core radicals and radicals are investigated. Moreover, for this purpose
also ∗-algebras and finely regular algebras are studied. Relations with
operator theory and realizations of these algebras by operator algebras
are outlined.
1
1
Introduction.
Algebras and operator algebras over the real field R and the complex field
C were intensively studied. They have found many-sided applications. For
them a lot of results already was obtained (see, for example, [6, 14, 19, 25]
and references therein). Among them dual algebras and annihilator algebras
play very important role. But for such algebras over ultranormed fields com-
paratively little is known because of their specific features and additional
difficulties arising from structure of fields [1, 12, 16, 17, 20, 22, 31, 38].
Many results in the classical case use the fact that the real field R has
the linear ordering compatible with its additive and multiplicative structure
1key words and phrases: operator; algebra; ideal; infinite dimension; field; ultranorm
Mathematics Subject Classification 2010: 12J05; 14F30; 16D60; 16E40; 46B28
address: Dep. Appl. Mathematics, Moscow State Techn. Univ. MIREA, av. Vernadksy
78, Moscow, 119454, Russia
e-mail: [email protected]
1
and that the complex field C is algebraically closed and norm complete and
locally compact and is the quadratic extension of R, also that there are not
any other commutative fields with archimedean multiplicative norms and
complete relative to their norms besides these two fields.
For comparison, in the non-archimedean case the algebraic closure of the
field Qp of p-adic numbers is not locally compact. Each ultranormed field
can be embedded into a larger ultranormed field. There is not any ordering
of an infinite ultranormed field such as Qp, Cp or Fp(t) compatible with its
algebraic structure.
In their turn, non-archimedean analysis, functional analysis and represen-
tations theory of groups over non-archimedean fields develop fast in recent
years [30, 31, 32, 33, 11, 23, 24]. This is motivated not only by needs of
mathematics, but also their applications in other sciences such as physics,
quantum mechanics, quantum field theory, informatics, etc. (see, for exam-
ple, [2, 3, 4, 13, 18, 29, 36, 37] and references therein).
This article is devoted to ultranormed B∗-algebras, dual algebras and an-
nihilator algebras over non-archimedean fields. Their structure is studied in
the paper. Extensions of algebras and fields are considered and using them
core radicals and radicals are investigated. Moreover, for this purpose also
∗-algebras and finely regular algebras are studied. Theorems about idempo-
tents of algebras and their orthogonality are proven. Division subalgebras
related with idempotents are investigated. Relations with operator theory
and realizations of these algebras by operator algebras are outlined. Then
B∗-algebras are defined and their properties studied. Theorems about their
embeddings into operator algebras are proved.
All main results of this paper are obtained for the first time. They can
be used for further studies of ultranormed algebras and operator algebras
on non-archimedean Banach spaces, their cohomologies, spectral theory of
operators, the representation theory of groups, algebraic geometry, PDE,
applications in the sciences, etc.
2 Ultranormed algebras and ∗-algebras.
To avoid misunderstandings we first give our definitions and notations.
1. Notation. Let F be an infinite field supplied with a multiplicative
non-trivial ultranorm · F relative to which it is complete, so that F is non-
2
discrete and ΓF ⊂ (0, ∞) = {r ∈ R : 0 < r < ∞}, where ΓF := {xF :
x ∈ F \ {0}}, whilst as usually xF = 0 if and only of x = 0 in F , also
x + yF ≤ max(xF , yF ) and xyF = xF yF for each x and y in F . We
consider fields with multiplicative ultranorms if something other will not be
specified.
If F is such a field, we denote by En(F ) the class containing F and
all ultranormed field extensions G of F so that these G are norm complete
and · GF = · F . By En we denote the class of all infinite non-trivially
ultranormed fields F which are norm complete.
Henceforward, the terminology is adopted that a commutative field is
called shortly a field, while a noncommutative field is called a skew field or
a division algebra.
2. Definitions. By c0(α, F ) is denoted a Banach space consisting of all
vectors x = (xj : ∀j ∈ α xj ∈ F ) satisfying the condition
card{j ∈ α :
xj > ǫ} < ℵ0 for each ǫ > 0
and furnished with the norm
(1) x = supj∈α xj,
where α is a set. For locally convex spaces X and Y over F the family of all
linear continuous operators A : X → Y we denote by L(X, Y ). For normed
spaces X and Y the linear space L(X, Y ) is supplied with the operator norm
(2) A := supx∈X\{0} Ax/x.
For locally convex spaces X and Y over F the space L(X, Y ) is furnished
with a topology induced by a family of semi-norms
(3) Ap,q := supx∈X,p(x)>0 q(Ax)/p(x)
for all continuous semi-norms p on X and q on Y .
Speaking about Banach spaces and Banach algebras we undermine that
a field over which it is defined is ultranorm complete.
If X = c0(α, F ), then to each A ∈ L(X, X) an infinite matrix (Ai,j : i ∈
α, j ∈ α) corresponds in the standard basis {ej : j ∈ α} of X, where
for each x ∈ X = c0(α, F ).
(4) x = Pj xjej
For a subalgebra V of L(X, X) an operation B 7→ Bt from V into L(X, X)
will be called a transposition operation if it is induced by that of its infinite
matrix such that (aA + bB)t = aAt + bBt and (AB)t = BtAt and (At)t = A
for every A and B in V and a and b in F , that is (At)i,j = Aj,i for each i and
j in α. Then V t := {A : A = Bt, B ∈ V }.
An operator A in L(X, X) is called symmetric if At = A.
3
By L0(X, X) is denoted the family of all continuous linear operators U :
X → X matrices (Ui,j : i ∈ α, j ∈ α) of which fulfill the conditions
(5) ∀i ∃ limj Uj,i = 0 and ∀j ∃ limi Uj,i = 0.
For an algebra A over F , F ∈ En, it is supposed that an ultranorm · A
on A satisfies the conditions:
aA ∈ (ΓF ∪ {0}) for each a ∈ A, also
aA = 0 if and only if a = 0 in A,
taA = tF aA for each a ∈ A and t ∈ F ,
a + bA ≤ max(aA, bA) and
abA ≤ aAbA for each a and b in A.
For short it also will be written · instead of · F or · A.
3. Theorem. Let V be a subalgebra in L(X, X) such that V t = V . Then
J is a left or right ideal in V if and only if J t is a right or left respectively
ideal in V .
Proof. For each A and B in V we get (ABt)t = BAt and Bt ∈ V and
At ∈ V , since V t = V . Therefore, for a right ideal J we deduce that ∀A ∈
J ∀B ∈ V (ABt ∈ J) ⇔ (BAt ∈ J t). Moreover, ∀B ∈ V ∃U ∈ V U t = B.
The similar proof is for a left ideal J.
4. Theorem. Let X = c0(α, F ), where F ∈ En. Then the class Lc(X, X)
of all compact operators T : X → X is a closed ideal in L(X, X), also
Lt,c(X, X) := {A : A ∈ Lc(X, X) & At ∈ Lc(X, X)} is a closed ideal in
L0(X, X).
Proof. By the definition of a compact operator T ∈ Lc(X, X) if and
only if for the closed unit ball B (of radius 1 and with 0 ∈ B) in X its image
T B is a compactoid in X (see Ch. 4 in [31]). Therefore, if A ∈ L(X, X),
then AB is bounded and convex in X, consequently, T A ∈ Lc(X, X). On
the other hand, if C is a compactoid in X, then AC is a compactoid in X,
hence AT ∈ Lc(X, X). Thus Lc(X, X) is the ideal in L(X, X).
Suppose that Tn is a fundamental sequence in Lc(X, X) relative to the
operator norm topology. Then its limit T = limn Tn exists in L(X, X), since
L(X, X) is complete relative to the operator norm topology. Let ǫ > 0.
There exists m ∈ N such that T − Tn < ǫ for each n > m. Since Tn
is the compact operator, there exists a finite set a1, ..., al in X such that
(TnB) ⊆ B(X, 0, ǫ) + Co(a1, ..., al), where
Co(a1, ..., al) = {x ∈ X : x = t1a1 + ... + tlal, t1 ∈ B(F, 0, 1, ),
..., tl ∈ B(F, 0, 1)} and B(X, y, r) := {z ∈ X :
z − y ≤ r},
0 < r, ¯U denotes the closure of a set U in a topological space. Therefore,
4
if x ∈ T B, then there exists y ∈ TnB such that x − y < ǫ, consequently,
x ∈ B(X, 0, ǫ) + Co(a1, ..., al) due to the ultrametric inequality, hence
T B ⊆ B(X, 0, ǫ) + Co(a1, ..., al).
This means that the operator T is compact. Thus Lc(X, X) is closed in
L(X, X).
The mapping U 7→ U t is continuous from L(X, X) into L(X, X), since
U = U t = supi∈α, j∈α Ui,j for each U ∈ L(X, X). In view of Theorem 4.39
in [31] for each A ∈ Lt,c(X, X) and ǫ > 0 operators S and R in L(X, X) exist
such that SX and RX are finite dimensional spaces over F and A − S < ǫ
and At − R < ǫ. Therefore, Lt,c(X, X) ⊂ L0(X, X) and Lt,c(X, X) is
the ideal in L0(X, X). On the other hand, L0(X, X) is closed in L(X, X),
consequently, Lt,c(X, X) is closed in L0(X, X).
5. Definition. Suppose that F is an infinite field with a nontrivial
non-archimedean norm such that F is norm complete, F ∈ En and of the
characteristic char(F ) 6= 2 and B2 = B2(F ) is the commutative associative
algebra with one generator i1 such that i2
1 = −1 and with the involution
(vi1)∗ = −vi1 for each v ∈ F . Let A be a subalgebra in L(X, X) such that A
is also a two-sided B2-module, where X = c0(α, F ) is the Banach space over
F , α is a set. We say that A is a ∗-algebra if there is
(1) a continuous bijective surjective F -linear operator I : A → A such
that
(2) I(ab) = (Ib)(Ia) and
(3) I(ga) = (Ia)g∗ and I(ag) = g∗(Ia)
(4) IIa = a
(5) (θ(y))(ax) = (θ((Ia)y))(x)
for every a and b in A and g ∈ B2 and x and y in X, where θ : X ֒→ X ′
is the canonical embedding of X into the topological dual space X ′ so that
θ(y)x = Pj∈α yjxj. For short we can write a∗ instead of Ia. The mapping I
we call the involution. An element a ∈ A we call self-adjoint if a = a∗.
6. Lemma. Let A be a subalgebra of L(X, X) with transposition and
At = A, where X = c0(α, F ), F ∈ En, char(F ) 6= 2. Then the minimal
∗-algebra K generated by A and B2 has an embedding ψ into L(U, U) such
that ψ(B2) is contained in the center Z(K) of K, where U = X ⊕ X.
Proof. We put ψ(a) := (cid:16)a 0
−a 0(cid:17) and (ψ(a))∗ := (cid:16)at 0
0 at(cid:17)
and (ψ(ai1))∗ := (cid:16)0 −at
at 0(cid:17) for each a ∈ A, since at ∈ A. Therefore, the minimal
algebra containing ψ(A) and ψ(Ai1) is the ∗-subalgebra in L(U, U). Then
0 a(cid:17) and ψ(ai1) := (cid:16) 0 a
5
−IX 0(cid:17), where IX is the unit operator on X.
(ψ(i1))2 = −IU and ψ(i1) = (cid:16) 0 IX
Thus ψ(ai1) = ψ(a)ψ(i1) = ψ(i1)ψ(a) = ψ(i1a) for each a ∈ A and hence
ψ(B2) ⊂ Z(K), where Z(K) denotes the center of the algebra K.
7. Lemma. Let A be a ∗-algebra over F (see Definition 5), then each
1 = a1
element a ∈ A has the decomposition a = a0 + a1i1 with a∗
in A.
0 = a0 and a∗
Proof. Put a0 = (a + a∗)/2, a1 = (ai∗
1 + i1a∗)/2, since char(F ) 6= 2.
Then a0 and a1 are in A, since A is the two-sided B2-module and a∗ ∈ A and
1 ∈ B2 and i1 ∈ B2 and i∗
1 = −i1. The algebra A is associative. Therefore,
a∗
j = aj and (i1aj)∗i1 = a∗
j i∗
1i1 = aj = i1(aji1)∗ for j = 0 and j = 1.
Consider the particular case:
if a = a∗, then a0 = a and (a1i1)∗ = (a + i1ai1)∗/2 = a1i1.
The latter together with a∗
1 = a1 implies that −i1a1 = a1i1 if a = a∗. On
the other hand, a = 2a1i1 − i1ai1 and a∗ = −2a1i1 − i1ai1 if a = a∗. Thus
4a1i1 = 0 and hence a1 = 0, that is, ai1 = i1a if a = a∗, since a1 = a1i1i∗
1 and
char(F ) 6= 2. This implies that a1i1 = i1a1 for each a ∈ A, consequently, the
decomposition is valid a = a0 + a1i1 with the self-adjoint elements a∗
0 = a0
and a∗
1 = a1 in A.
3 Dual and annihilator ultranormed algebras.
1. Definition. Let A be a topological algebra over a field F and let S be a
subset of A. The left annihilator is defined by L(A, S) := {x ∈ A : xS = 0}
and the right annihilator is R(A, S) := {x ∈ A : Sx = 0}, shortly they also
will be denoted by Al(S) := L(A, S) and Ar(S) := R(A, S).
2. Definition. An algebra A is called an annihilator algebra if conditions
(1 − 3) are fulfilled:
(1) Al(A) = Ar(A) = 0 and
(2) Al(Jr) 6= 0 and
(3) Ar(Jl) 6= 0
for all closed right Jr and left Jl ideals in A.
If for all closed (proper or improper) left Jl and right Jr ideals in A
(4) Al(Ar(Jl)) = Jl and
(5) Ar(Al(Jr)) = Jr
then A is called a dual algebra.
6
If A is a ∗-algebra (see Definitions 2.5) and for each x ∈ A elements a ∈ A
and a1 ∈ A exist such that an ultranorm on A for these elements satisfies
the following conditions
(6) axx∗a∗
1 = x2 and aa∗
1 ≤ 1,
then the algebra A is called finely regular.
3. Theorem. If A is an ultranormed annihilator finely regular Banach
algebra, then A is dual.
Proof. Consider arbitrary x ∈ A and take elements a ∈ A and a1 ∈ A
1 ≤ xx∗, hence
fulfilling conditions 2(6), then x2 = axx∗a∗
x ≤ x∗. Substituting x by x∗ we deduce analogously that x∗ ≤ x,
consequently, x = x∗.
1 ≤ axx∗a∗
For a closed left ideal Jl in A if x ∈ Jl ∩ (Ar(Jl))∗, then xx∗ = 0, con-
sequently, x = 0 by Formula 2(6) and hence Jl ∩ (Ar(Jl))∗ = 0. Then
Vl := Jl ⊕ (Ar(Jl))∗ is a left ideal in A, since Ar(Jl) is the closed right ideal
in A and (Ar(Jl))∗ is the closed left ideal in A.
For an arbitrary x ∈ Vl there exist elements y ∈ Jl and z ∈ (Ar(Jl))∗
such that x = y + z. Therefore, xz∗ = zz∗ and xy∗ = yy∗. Using con-
ditions 2(6) we choose elements a ∈ A, a1 ∈ A, b ∈ A and b1 ∈ A with
aa∗
1 = y2
1 ≥ axz∗a∗
and hence xz∗ ≥ axz∗a∗
1 = z2 and xy∗ ≥
bxy∗b∗
1 = byy∗b∗
1 = y2. Therefore, x ≥ z and x ≥ y.
Thus Vl is the closed left ideal in A.
1 ≤ 1 such that azz∗a∗
1 = z2 and byy∗b∗
1 = azz∗a∗
1 ≤ 1 and bb∗
1 ≥ bxy∗b∗
From Condition 2(3) it follows that a nonzero element a ∈ A exists such
that Vla = (0), consequently, Jla = (0) and (Ar(Jl))∗a = (0). Then from the
inclusion a ∈ Ar(Jl) and hence a∗ ∈ (Ar(Jl))∗ it follows that a∗a = 0. The
latter contradicts the supposition that the algebra A is completely regular.
Thus Vl = A and analogously for each closed right ideal Jr in A the equality
A = Vr is valid, where Vr = Jr ⊕ (Al(Jr))∗.
Particularly, for Jr = Ar(Jl) it implies that A = Ar(Jl) ⊕ (Al(Ar(Jl)))∗.
The involution of both sides of the latter equality gives A = (Ar(Jl))∗ ⊕
Al(Ar(Jl)), since Jl ⊆ Al(Ar(Jl)). Thus Jl = Al(Ar(Jl)) for each closed left
ideal Jl in A and the involution leads to the equality Jr = Ar(Al(Jr)) for
each closed right ideal Jr in A. Thus conditions 2(4, 5) are fulfilled.
4. Definition. If idempotents w1 and w2 of an algebra A satisfy the
conditions w1w2 = 0 and w2w1 = 0, then it is said that they are orthogonal.
A family {wj : j} of idempotents is said to be orthogonal, if and only if every
two distinct of them are orthogonal. An idempotent p is called irreducible,
7
if it can not be written as the sum of two mutually orthogonal idempotents.
5. Definition. For two Banach algebras A and B over an ultranormed
field F , F ∈ En, we consider the completion A ⊗F B relative to the projective
tensor product topology (see [28, 31]) of the tensor product A ⊗F B over the
field F such that A ⊗F B is also a Banach algebra into which A and B have
natural F -linear embeddings π1 and π2.
For a Banach algebra B over an ultranormed field F , F ∈ En, and an
element x ∈ B we say that x has a left core quasi-inverse y if for each H ∈
En(F ) an element y ∈ BH exists satisfying the equality x + y + yx = 0, where
BH = B ⊗F H. Similarly is defined a right core quasi-inverse. Particularly, if
only H = F is considered they are shortly called a left quasi-inverse and a
right quasi-inverse correspondingly.
For a unital Banach algebra A over F , where F ∈ En, if an element
x ∈ A has the property: for each field extension G ∈ En(F ) the left inverse
(1 + yx)−1
exists in AG for each y ∈ AG, then we call x a generalized core
nil-degree element. The family of all generalized core nil-degree elements of
A we call a core radical and denote it by Rc(A).
l
6. Proposition. Let A be a unital Banach algebra over F , where F ∈
En. Then
l
Rc(A) = T{A∩Jl : G ∈ En(F ) & Jl is a proper maximal left ideal in AG}.
Proof. Consider an element x ∈ A such that for each G ∈ En(F ) (see
Subsection 2.1) and each maximal left ideal Jl in AG the inclusion x ∈ Jl
is valid. If an element y ∈ AG is such that (1 + yx)−1
does not exist, then
an element z = 1 + yx belongs to some left ideal J in AG. Since AG is the
unital algebra, then z belongs to some proper maximal left ideal M such
that J ⊂ M. But yx also belongs to M, since x belongs to each maximal
left ideal, consequently, 1 = z − yx ∈ M. The latter is impossible, since M
is the proper left ideal in AG. This means that the left inverse (1 + yx)−1
exists for every G ∈ En(F ) and y ∈ AG. Thus x belongs to the core radical.
Vice versa. Let now x ∈ Rc(A). Suppose the contrary that a field ex-
tension G ∈ En(F ) and a proper maximal left ideal Jl in AG exist such that
x /∈ Jl. Consider the set V of all elements z = b − yx with b ∈ Jl and
y ∈ AG. Evidently V is the left ideal in AG containing Jl, but Jl is maximal,
consequently, V = AG. This implies that 1 = b − yx for some b ∈ Jl and
y ∈ AG. Therefore the element b = 1 + yx has not a left inverse. But this
contradicts the supposition made above.
l
7. Proposition. Suppose that A is a unital Banach algebra over F ,
8
where F ∈ En. Then
(x ∈ Rc(A)) ⇔ (∀G ∈ En(F ) ∀y ∈ AG ∃(1 + yx)−1 ∈ AG).
Proof. If ∀G ∈ En(F ) ∀y ∈ AG ∃(1+yx)−1 ∈ AG, then ∀G ∈ En(F ) ∀y ∈
l ∈ AG, consequently, x ∈ Rc(A), where as usually (1 + yx)−1
AG ∃(1 + yx)−1
notates the inverse of 1 + yx.
Vice versa. Let x ∈ Rc(A). Then by the definition of the core radical
∀G ∈ En(F ) ∀y ∈ AG ∃(1 + yx)−1
l ∈ AG. For G ∈ En(F ) denote by 1 + b a
left inverse of 1 + yx in AG, that is (1 + b)(1 + yx) = 1. This implies that
1 + yx is the right inverse of 1 + b in AG and b = −byx − yx. From x ∈ Rc(A)
it follows that b ∈ Rc(AG), since x ∈ Jl and hence y ∈ Jl for each proper
maximal left ideal Jl in AH and each H ∈ En(G). This means that for every
H ∈ En(G) and z ∈ AH a left inverse (1 + zb)−1
exists in AH , particularly,
for z = 1 also. On the other hand, the right inverse is (1 + zb)−1
r = 1 + yx
as it was already proved above. Therefore the inverse (i.e.
left and right
simultaneously) (1 + b)−1 = 1 + yx exists. Thus 1 + b is the inverse of 1 + yx
in AG.
l
8. Proposition. Let A be a unital Banach algebra over F , where F ∈
En. Then
Rc(A) = T{A∩Jr : G ∈ En(F ) & Jr is a proper maximal right ideal in AG}.
Moreover, Rc(A) is the two-sided ideal in A.
Proof. Consider the class Qe(A) of all elements x ∈ A such that for each
exists in AG for each
field extension G ∈ En(F ) the right inverse (1 + xy)−1
r
y ∈ AG. Analogously to the proof of Proposition 6 we infer that
Similarly to the proof of Proposition 7 we deduce that
Qe(A) = T{Jr : G ∈ En(F ) & Jr is a proper maximal right ideal in AG}.
(x ∈ Qe(A)) ⇔ (∀G ∈ En(F ) ∀y ∈ AG ∃(1 + xy)−1 ∈ AG).
Suppose that G ∈ En(F ), x ∈ A, y ∈ AG and the inverse element exists
(1 + yx)−1 = 1 + b in AG. Then (1 + xy)(1 − xy − xby) − 1 = −x((1 + yx)(1 +
b) − 1)y = 0 and (1 − xy − xby)(1 + xy) − 1 = −x((1 + b)(1 + yx) − 1)y = 0,
consequently, 1 − xy − xby = (1 + xy)−1. Analogously if the inverse element
(1+yx)−1 exists, then (1+xy)−1 also exists. This implies that Qe(A) = Rc(A)
and hence the core radical is the two-sided ideal in A.
9. Proposition. Suppose that A is a unital Banach algebra over F ,
where F ∈ En. Then an extension field H = HF ∈ En(F ) exists such that
Rc(A) = A∩R(AH ), where R(AH) denotes the radical of the algebra AH over
H. Moreover, H can be chosen algebraically closed and spherically complete.
Proof. Consider an arbitrary element x ∈ A \ Rc(A). This means that
9
a field extension G = Gx ∈ En(F ) and an element y ∈ AG exist such that
the element (1 + yx) has not the left inverse in AG. For the family G :=
{Gx : x ∈ A \ Rc(A), Gx ∈ En(F )} a field H = HF ∈ En(F ) exists such that
Gx ⊆ H for each x ∈ A \ Rc(A) due to Proposition V.3.2.2 [8] and since the
multiplicative ultranorm · F can be extended to a multiplicative ultranorm
· H on H (see Proposition 5 in Section VI.3.3 [9], Krull's existence theorem
14.1 and Theorem 14.2 in [33] or 3.19 in [31], Lemma 1 and Proposition 1 in
[10])).
If HF is not either algebraically closed or spherically complete, one can
take the spherical completion of its algebraic closure ¯HF (see Corollary 3.25,
Theorem 4.48 and Corollary 4.51 in [31]). Then also ¯HF ∈ En(F ). Denote
shortly ¯HF by H.
Therefore, if G ∈ G, then from y ∈ AG it follows that y ∈ AH. For each
x ∈ A \ Rc(A) an element y ∈ AG exists such that AG(1 + yx) is a left proper
ideal in AG, consequently, AG(1 + yx) ⊗GH = AH (1 + yx) is a left proper
ideal in AH , since H ⊂ Z(AH). Therefore, (1 + yx) has not a left inverse in
AH.
Thus for each x ∈ A \ Rc(A) and G = Gx ∈ G and element y ∈ AH exists
such that (1+yx) has not a left inverse in AH . Therefore, A∩R(AH ) ⊂ Rc(A).
On the other hand, if x ∈ Rc(A), then x ∈ R(AH) according to the definition
of Rc(A) in §5. Thus Rc(A) = A ∩ R(AH ) for the fields H constructed above.
10. Theorem. Let A be a unital Banach algebra over F , where F ∈ En.
Then an extension field K = KF ∈ En(F ) exists such that
(1) Rc(AK) = R(AK).
Moreover, K can be chosen algebraically closed and spherically complete.
Proof. Put K1 = H, where H = HF is given by Proposition 9. Then by
induction take Kn+1 = HKn for each natural number n = 1, 2, 3, .... There are
isometric embeddings Kn ֒→ Kn+1 for each n. Let K be the norm completion
of K∞ := S∞
n=1 Kn, hence K ∈ En(F ). In addition each field Kl can be chosen
algebraically closed and spherically complete due to Proposition 9. Moreover,
it is possible to take as K the spherical completion of the algebraic closure
of K∞ (see Corollary 3.25, Theorem 4.48 and Corollary 4.51 in [31]).
In view of Proposition 9 Rc(AKl) = AKl ∩ R(AKl+1) for each natural
number l. Let x ∈ Rc(AK), that is for each G ∈ En(K) and y ∈ AK a left
inverse (1 + yx)−1
exists in AK. The algebra A ⊗F K∞ over the field K∞ is
everywhere dense in AK = A ⊗F K. Therefore, there exist sequences xn and
yn in AK such that xn ∈ AKn and yn ∈ AKn for each n and limn xn = x and
l
10
limn yn = y. Since (1 + z) is invertible in AK for each z ∈ AK with z < 1,
then a natural number m exits such that a left inverse (1 + ynxn)−1
exists
for each n > m.
l
From G ∈ En(K) and Kl ∈ En(F ), Kl ⊆ K it follows that G ∈ En(Kl)
for each l = 1, 2, 3, .... On the other hand, an element y ∈ AK can be any
marked element in particularly belonging to AKl. Thus Sl Rc(AKl) is dense
in Rc(AK). Similarly considering G = K one gets that Sl R(AKl) is dense in
R(AK). Mentioning that Sl AKl is dense in A ⊗F K∞ one gets that Sl AKl is
dense in AK. Therefore, we infer that
Rc(AK) = clAK ([
l
Rc(AKl)) = clAK ([
(AKl ∩ R(AKl+1)))
l
= clAK ([
l
R(AKl+1)) = R(AK),
where clAK B denotes the closure of a subset B, B ⊂ AK, in AK.
11. Proposition. Let A be a Banach algebra over F , F ∈ En, also
let a field K fulfill Condition 10(1) for A1, where A1 = A if 1 ∈ A, while
A1 = A ⊕ 1F if 1 /∈ A. Then an element x ∈ AK is not core left quasi-
invertible if and only if Jl,G := {z + zx : z ∈ AG} is a proper left ideal in
AG for each G ∈ En(K). If so Jl,G is a proper regular left ideal in AG such
that x /∈ Jl,G.
Proof. By virtue of Theorem 10 Rc(A1,K) = R(A1,K). Hence for each
G ∈ En(K) an element x ∈ AK is not core left quasi-invertible in AG if
and only if it does not belong to R(A1,K). If u = y + yx and v = z + zx
belong to Jl,G, b and c are in G, where y and z belong to AG, then bu + cv =
(by + cz) + (by + cz)x, consequently, cu + bv ∈ Jl,G. That is AGJl,G ⊆ Jl,G. If
Jl,G is not a proper left ideal, then Jl,G = AG. This implies that an element
z ∈ AG exists such that x + zx = −x. The latter is equivalent to the equality
x + zx + x = 0. Thus z is a left quasi-inverse of x.
Vise versa if x has a left quasi-inverse in AG, then x ∈ Jl,G, hence −zx ∈
Jl,G. Therefore, z = (z + zx) − zx ∈ Jl,G for each z ∈ AG, consequently,
AG = Jl,G. Thus if Jl,G is a proper left ideal, then x /∈ Jl,G. Mention that
the element w = −x is unital modulo the proper left ideal Jl,K, consequently,
this ideal is regular.
12. Proposition. Suppose that A is a Banach algebra over F , F ∈ En,
also a field K satisfies Condition 10(1) for A1. Then the following conditions
are equivalent:
11
(1) an element x ∈ AK possesses a left quasi-inverse in AG for each
G ∈ En(K);
(2) for every G ∈ En(K) and a maximal regular proper left ideal Ml,G in
AG an element y ∈ AG exists such that x + y + yx ∈ Ml,G.
Proof. If an element x ∈ AK possesses a left quasi-inverse yG in AG for
each G ∈ En(K), then x + yG + yGx ∈ Ml,G for each maximal regular proper
left ideal Ml,G in AG due to Theorem 10.
Vise versa suppose that Condition (2) is fulfilled, but x is not left quasi-
invertible in AG for some G ∈ En(K). Then Jl,G is a regular proper left ideal
in AG according to Proposition 11. Therefore, a maximal regular proper left
ideal Ml,G in AG exists containing Jl,G. Thus an element y ∈ AG exists such
that x + y + yx ∈ Ml,G. On the other hand, the inclusion y + yx ∈ Jl,G is
accomplished, consequently, y ∈ Ml,G and hence −zx ∈ Ml,G for each z ∈ AG.
This implies that z ∈ Ml,G for each z ∈ AG, since z = −zx + (z + zx). But
this leads to the contradiction AG = Ml,G. Thus (2) ⇒ (1).
13. Proposition. Suppose that A is a Banach annihilator algebra over
an ultranormed field F , F ∈ En. Then a field extension K, K ∈ En(F ),
exists such that if an element −p ∈ AK is not core left quasi-invertible, then
a nonzero element x ∈ AK \ {0} exist satisfying the equation px = x.
Proof. We take a field K, K ∈ En(F ), given by Theorem 10 for a unital
algebra E = A1, where E = A ⊕ 1F if 1 /∈ A, while E = A if 1 ∈ A.
Therefore, Rc(EK) = R(EK).
By virtue of Proposition 11 Jl,K := {yp − p : y ∈ AK} is a regular proper
left ideal in AK. Since EK is the unital Banach algebra over K, then it is with
continuous inverse. Hence if A is not unital, then AK is with the continuous
quasi-inverse. Mention that an element v is a left quasi-inverse of q in AK if
and only if 1 + v is a left inverse of 1 + q in EK.
Therefore, if 1 /∈ A, then a bijective correspondence exists: Q is a left
(maximal) ideal of EK which is not contained entirely in A if and only if
Q ∩ AK is a regular (maximal respectively) left ideal of AK. If 1 ∈ A, then
each left ideal in AK is regular.
Recall that a ring B satisfying the identities
(1) L(B, B) = (0) and R(B, B) = (0) is called annihilator, where
(2) L(B, S) = {x ∈ B : xS = (0)} and R(B, S) = {x ∈ B : Sx = (0)}
denote a left annihilator and a right annihilator correspondingly of a subset
S in B. Thus
(3) L(AK, AK) = (0) and R(AK, AK) = (0),
12
since AK = A ⊗F K, since by the conditions of this proposition L(A, A) = (0)
and R(A, A) = (0), also A and AK are Banach algebras. Next we take the
closure clAK (Jl,K) of Jl,K in AK. Therefore, R(AK, clAK (Jl,K)) is not nil,
R(AK, clAK (Jl,K)) 6= (0).
Suppose that x is a nonzero element in R(AK, clAK (Jl,K)), consequently,
x ∈ R(AK, Jl,K).
If z ∈ R(AK, Jl,K), then y(pz − z) = (yp − y)z = 0 for each y ∈ AK. From
L(AK, AK) = (0) and R(AK, AK) = (0) it follows that pz − z = 0. Vise versa,
if pz − z = 0 for some z ∈ AK, then (yp − y)z = y(pz − z) = 0 and hence
z ∈ R(AK, Jl,K). Therefore,
(4) R(AK, Jl,K) = {z ∈ AK : pz = z}.
Thus px = x.
14. Theorem. Suppose that A is a Banach annihilator algebra over a
field F ∈ En such that Rc(A) = R(A) and Mr is a proper maximal closed right
ideal in A satisfying the condition L(A, Mr) ∩ R(A) = (0). Then L(A, Mr)
contains an idempotent p and
(1) L(A, Mr) = Ap and
(2) Mr = {z − pz : z ∈ A}.
Proof. A nonzero element b in L(A, Mr) exists, since L(A, Mr) 6= (0),
since Mr is a proper right ideal in A. Therefore, Mr ⊂ R(A, {b}) 6= A and
consequently,
(3) R(A, {b}) = Mr,
since the right ideal Mr is maximal. The element b does not belong to R(A),
since L(A, Mr) ∩ R(A) = (0) by the conditions of this theorem.
In view of Theorem 10 and Propositions 11 and 12 a scalar t ∈ F and
an element y ∈ A exist such that the element −p = tb + yb has not a left
quasi-inverse in AG for each G ∈ En(F ). Thus p 6= 0 and p ∈ L(A, Mr).
By virtue of Proposition 13 a nonzero element x ∈ A \ (0) exists such that
px = x, consequently, (p2 − p)x = 0.
Suppose that p2 − p is not nil, p2 − p 6= 0. We have p2 − p ∈ L(A, Mr).
Taking b = p2 −p in (3) one gets R(A, p2 −p) = Mr, consequently, (p2 −p)x ∈
Mr and inevitably x = px = 0. This leads to the contradiction. Thus p2 = p.
On the other hand, p ∈ L(A, Mr) and p is not nil. Taking b = p in
(3) provides Mr = R(A, {p}) and R(A, {p}) = {z − pz : z ∈ A}, since
p(y − py) = py − p2y = 0, also if pz = 0, then z = z − pz. Therefore,
L(A, Mr) = Ap due to 13(4) and since p is the idempotent.
15. Corollary. If conditions of Theorem 14 are fulfilled, then Mr is a
13
maximal right ideal and L(A, Mr) is a minimal left ideal, also pA is a minimal
right ideal and L(A, pA) is a maximal left ideal.
16. Theorem. Let A be a Banach annihilator algebra over a field F ∈
En such that Rc(A) = R(A), let also Jl be a minimal left (may be closed)
ideal which is not contained in R(A), Jl \ R(A) 6= ∅. Then Jl contains an
idempotent p for which Jl = Ap and R(A, Ap) = {x − px : x ∈ A}.
Proof. Take x ∈ Jl \ R(A). From Propositions 11 and 12 it follows that
b ∈ F and y ∈ A exist such that the element −p = bx + yx has not a left
quasi-inverse, consequently, p 6= 0.
In view of Proposition 13 an element v ∈ A exists having the property
pv = v. Therefore, Yl := {z ∈ Jl : xv = 0} is a left ideal such that it is
contained in Jl and Jl 6= Yl, since p ∈ Jl \ Yl. This ideal Yl is closed, if Jl
is closed. The ideal Jl is minimal, hence Yl = (0). This implies that zv 6= 0
if z ∈ Jl \ {0}.On the other hand, p2 − p ∈ Jl and (p2 − p)v = 0, hence
p2 − p = 0. Thus p is the idempotent.
For each z ∈ Ap the condition z = zp is valid, consequently, Ap is a
closed left ideal contained in Jl and hence Ap = Jl, since the left ideal Jl is
minimal. Therefore, R(A, Jl) = {x − px : x ∈ A}.
17. Lemma. If A is a Banach annihilator semi-simple algebra over a
field F ∈ En with Rc(A) = R(A) and J is a left (or right, or two-sided) ideal
in A such that J 2 = (0), then J = (0).
Proof. Suppose that J is a left ideal in A with J 2 = (0). Therefore,
(tx + yx)2 = 0 for every t ∈ F , x ∈ J and y ∈ A, since tx + yx ∈ J. In
this case the element z = tx + yx has the left quasi-inverse −z. By virtue of
Propositions 11 and 12 x ∈ R(A), since Rc(A) = R(A) by the conditions of
this lemma. The algebra A is semi-simple, consequently, J = (0).
For a right ideal or a two-sided ideal the proof is analogous.
18. Lemma. If A is a Banach annihilator semi-simple algebra over a
field F ∈ En with Rc(A) = R(A) and Jr is a right minimal ideal in A, then
a closed two-sided ideal Y = Y (Jr) generated by Jr is minimal and closed in
A.
Proof. If X is a closed two-sided ideal contained in Y , then Jr ∩ X is a
right ideal contained in Jr, consequently, either Jr ∩ X = Jr or Jr ∩ X = (0),
since Jr is minimal. If Jr ∩ X = Jr, then Y ⊂ X, hence Y = X.
If Jr ∩ X = (0), then JrX ⊂ Jr ∩ X = (0), consequently, Jr ⊂ L(A, X).
Then L(A, X) is the closed two-sided ideal, consequently, Y ⊂ L(A, X).
Therefore, X ⊂ L(A, X) and consequently, X 2 = (0). Applying Lemma
14
17 we get that X = (0).
Thus Y is minimal.
19. Theorem. Let A be a Banach annihilator semi-simple algebra over
a field F ∈ En with Rc(A) = R(A). Then the sum of all left (or right) ideals
of A is dense in A.
Proof. Suppose that U is a sum of all minimal right ideals and ¯U is its
closure in A. If ¯U 6= A, then ¯U is the closed right ideal in A, consequently, a
nonzero element y in A exists such that y ¯U = (0). This implies that y belongs
to all left annihilators of all minimal right ideals and hence it belongs to the
intersection V of all maximal left regular ideals. In view of Proposition 3.8
one gets that this intersection is Rc(A). By the conditions of this theorem
Rc(A) = R(A), hence V is zero, since A is semi-simple. Thus y = 0 providing
the contradiction. Thus ¯U = A.
20. Proposition. Let conditions of Theorem 19 be fulfilled and let J be
a right ideal in A. Then J contains a minimal right ideal and an irreducible
idempotent s.
Proof. Suppose that J does not contain a minimal right ideal and sA is
a minimal right ideal for some irreducible idempotent s in A. This implies
that J ∩ (sA) = (0). Hence for each a ∈ A either asA = (0) or asA is also a
minimal right ideal, consequently, (asA) ∩ J = (0) for all a ∈ A and hence
(as) ∩ J = (ass) ∩ J ⊂ (asA) ∩ J = (0) for all a ∈ A. Thus (aS) ∩ J = (0).
Therefore JAs = (0), since JAs ⊂ (As) ∩ J. This means that JAs = (0) for
all minimal left ideals As. In view of Theorem 19 JA = (0), consequently,
J = (0).
21. Proposition. If conditions of Theorem 19 are satisfied and s is an
irreducible idempotent in A, then sA and As are minimal right and left ideals
correspondingly.
Proof. Suppose that sA is not minimal. By virtue of Proposition 20 it
contains a minimal right ideal rA such that rA 6= sA, rA ⊂ sA. Then an
element a ∈ A exists such that r = sa, consequently, rs = sas ∈ rA. This
implies that t is a nonzero idempotent contained in rA such that the element
t = rs satisfies the equalities st = ts = t and s−t is also a nonzero idempotent
providing the contradiction, since s = t + (s − t) and t(s − t) = (s − t)t = 0,
but s is irreducible by the conditions of this proposition. Thus sA is minimal.
22. Proposition. If conditions of Theorem 19 are satisfied and J is a
closed two-sided ideal in A, then L(A, J) = R(A, J) and J + R(A, J) is dense
in A.
15
Proof. In view of Lemma 17 J ∩ R(A, J) = (0), since J ∩ R(A, J) =: V
is the right ideal possessing the property V 2 = V . Therefore, R(A, J)J = (0)
and hence R(A, J) ⊂ L(A, J). Similarly L(A, J) ⊂ R(A, J), consequently,
L(A, J) = R(A, J).
If J + R(A, J) would be not dense in A, then its closure should be a
proper ideal in A, consequently, a nonzero element x in A exists such that
(J +R(A, J))x = (0). Therefore J(αx+xy) = (0) and R(A, J)(αx+xy) = (0)
for each y ∈ A and α ∈ F , hence (αx + xy) ∈ R(A, J) and consequently,
(αx + xy)2 = 0 for each y ∈ A and α ∈ F . But in the semi-simple algebra A
with Rc(A) = R(A) this is impossible for x 6= 0.
23. Proposition.
If conditions of Theorem 19 are met and J is a
minimal closed two-sided ideal in A, then J is an annihilator algebra with
Rc(J) = R(J). If in addition A is dual, then J is also dual.
Proof. If x ∈ J and Jx = (0), then x = 0, since J ∩ R(A, J) = (0) due
to Proposition 22. Analogously if xJ = (0) and x ∈ J, then x = 0. Thus
L(A, J) = R(A, J) = (0).
If Vl is a closed left ideal in J, then (J + L(A, J))Vl = JVl ⊂ Vl, hence
AVl ⊂ Vl, since J + L(A, J) is dense in A by Proposition 22. Thus Vl is the
closed left ideal in A.
Put Hl = Vl + R(A, J). Then either Hl is dense in A or R(A, Hl) 6= (0).
From Lemmas 17, 18 and Proposition 20 one gets J ∩ R(A, Hl) 6= (0) and
hence J ∩ R(A, Vl) 6= (0). Analogously J ∩ L(A, Vr) 6= (0) for a closed right
ideal Vr in J.
Suppose now that the algebra A is dual.
In view of Lemma 17 and
Proposition 22 if x ∈ J and [L(A, Vr) ∩ J]x = (0), then x ∈ R(A, L(A, Vr) ∩
J) = clA(R(L(A, Vr)) + R(A, J)) = clA(Vr + R(A, J)) = clA(Vr + L(A, J)).
Then (Vr + L(A, J))J = VrJ ⊂ Vr, since Vr is a right ideal in J, consequently,
clA(Vr + L(A, J))J ⊂ Vr, hence xJ ⊂ Vr and consequently, L(A, Vr)xJ ⊂
L(A, Vr)Vr = (0). On the other hand, L(A, Vr)xR(A, J) = (0), since x ∈ J,
consequently, L(A, Vr)x(J + R(A, J)) = (0). We have that J + R(A, J) is
dense in A due to Proposition 22, hence L(A, Vr)xA = (0) and consequently,
L(A, Vr)x ⊂ L(A, A) = (0). From the duality of A it follows that x ∈ Vr.
Therefore, R(J, L(J, Vr)) = Vr and similarly L(J, R(J, Vl)) = Vl. Thus J is
also dual.
24. Theorem. Let A be a Banach semi-simple annihilator algebra over
F ∈ En with Rc(A) = R(A). Then A is the completion of the direct sum of all
its minimal closed two-sided ideals Hk each of which is a simple annihilator
16
algebra over F . Moreover, if A is dual, then each Hk is simple and dual.
Proof. By virtue of Proposition 20 each closed minimal two-sided ideal J
in A contains a minimal right ideal Vr, hence J = Vr according to Lemma 18.
Then the closure clAVr is a closed minimal two-sided ideal for each minimal
right ideal Vr due to the same lemma. According to Proposition 23 clAVr
is the annihilator algebra, which is also dual if A is dual. If H is a closed
two-sided ideal in clAVr, then it is such in A also. But clAVr is minimal,
hence the algebra clAVr is simple.
By virtue of Theorem 19 the sum of all minimal right ideals Vr is dense in
A. Let K and M be two minimal closed two-sided ideals which are different,
K 6= M. Therefore KM ⊂ K ∩ M = (0), since K ∩ M is the closed two-sided
ideal contained in minimal closed two-sided ideals K and in M and different
from them. If x + y = 0 for some x ∈ K and y ∈ M, then Kx = (0) and
My = (0), consequently, (xA)2 ⊂ K(xA) = (0) and analogously (yA)2 = (0).
Therefore xA = (0) and yA = (0), since A is semi-simple, consequently, x = 0
and y = 0. Thus the considered sum is direct.
25. Theorem. If A is a Banach simple annihilator algebra over a field
F ∈ En with Rc(A) = R(A), if also p is an irreducible idempotent, then
pAp =: H is an ultranormed division algebra over F . Moreover, if A and F
are ultranormed and A is commutative, then a multiplicative ultranorm · H
on H exists extending that of F such that it induces a topology on H not
stronger than the topology inherited from A.
Proof. From the conditions of this proposition it follows that pH =
Hp = H, since p2 = p and the algebra A is associative. Evidently, H is the
algebra over F , since A is the algebra over F . The restriction of p to H is the
identity on H, since ps = p2s = p(ps) for each s ∈ A and hence pr = r for
each r ∈ H, similarly rp = r for each r ∈ H and hence pr = rp = r = prp.
For each nonzero element r in H the set Ar is a left ideal in A and Ar 6= (0)
due to Condition 1(1). In view of Propositions 20 and 21 Ar ⊂ Ap and Ap is
a minimal left ideal, since p is the irreducible idempotent. Thus Ar = Ap and
hence an element y ∈ A exists such that yr = p2 = p, consequently, pyr =
py(pr) = (pyp)r. Therefore, (pyp)r = (pyp)(prp) = pyprp = pyr = pp = p,
consequently, pyp is a left inverse of r in H. Similarly r has a right inverse
in H. Thus H is the division algebra such that F is isomorphic with F p and
F p ⊂ H. From the continuity of the algebraic operations on A it follows
that they are continuous on H. The norm on A induces a norm on H, since
H ⊂ A. Since H is the topological ring with the continuous quasi-inverse
17
and H possesses the unit, then H is with the continuous inverse.
If A and F are ultranormed and A is commutative, then the ultranorm
·A on A induces the ultranorm on H and H is also commutative. Therefore,
pA = p2A ≤ p2
A and hence 1 ≤ pA. On the other hand, on H as the field
extension of F there exists a multiplicative ultranorm · H extending · F
that of the field F (see Proposition 5 in Section VI.3.3 [9], Krull's existence
theorem 14.1 and Theorem 14.2 in [33] or 3.19 in [31]). We have that 1F = 1,
1 ≤ pA, also p plays the role of the unit in H, while bxA = bF xA for
each b ∈ F and x ∈ A.
If A is not unital, we consider the algebra A1 obtained from it by adjoining
the unit. The norms on A and F induce the norm on A1 = A⊕F . Therefore,
it is sufficient to consider the case of the unital algebra A. Mention that
(1−p)2 = 1−p and A(1−p) is the ideal in A such that A = Ap+A(1−p) with
Ap ∩ A(1 − p) = (0). Moreover, Ap = pAp = Ap2, since A is commutative.
This implies that H is isomorphic with the quotient algebra J = A/(A(1−p)).
Then the ultranorm on A induces the quotient ultranorm on J such that
xyJ ≤ xJyJ and xypJ ≤ xpJypJ for each x and y in J, since pxp = xp
and xpyp = xyp for each elements x and y in the commutative algebra A.
At the same time, xypH = pxppypH = xpHypH for each x and y in A.
The ultranorm · A on F p induced from A is equivalent with the mul-
tiplicative ultranorm · F on F , since F p is isomorphic with F and conse-
quently, xpypzA = xpAypAzA for every xp ∈ F p, yp ∈ F p and z in A,
since xpF = xpA. Then xypJ = xpJ ypJ if xp ∈ F p and yp ∈ F p,
where x and y are in A. The inequality p−1
J xpJ ≤ xJ is also fulfilled for
each x ∈ A. Therefore, H can be supplied with a multiplicative norm · H
extending that of F and satisfying the inequality xH ≤ xJ for each x ∈ H
according to Theorems 1.15 and 1.16 [16].
26. Proposition. Suppose that A is a Banach simple annihilator algebra
over a field F , F ∈ En, also Rc(A) = R(A). Then a maximal family of
orthogonal irreducible idempotents {wj : j ∈ J} exists such that Pj Awj and
Pj wjA are dense in A.
Proof. In view of Proposition 20 there are irreducible idempotents wj
in A. Each right ideal B in A contains a minimal right ideal, consequently,
it contains an irreducible idempotent. By virtue of Zorn's lemma (see [15]or
[21]) a maximal orthogonal system {wj : j ∈ J} of irreducible idempotents
wj exists. Let C = Pj Awj be the sum of all such left ideals. Suppose that
clAC 6= A. Then clAC is a closed left ideal. Therefore, Ar(clAC) is the right
18
ideal different from zero. This implies that Ar(clAC) contains an irreducible
idempotent p orthogonal to each wj. But this is impossible, since the family
{wj : j ∈ J} is maximal. It remains that C is dense in A. Similarly Pj wjA
is dense in A.
27. Proposition. Let F be a field and let {Kj : j ∈ P } be a family of
division algebras such that F is contained in the center Z(Kj) of Kj for each
j ∈ P , where P is a set. Then a minimal division algebra K exists such that
Kj ⊆ K for each j ∈ P .
Proof. Since Kj is a division algebra, then its center Z(Kj) is a field.
Take the tensor product T = Nj∈P Kj of Kj as algebras over the field F .
Therefore, T is an algebra over F so that T may be noncommutative if
at least one of Kj is noncommutative. For each Kj a natural embedding
hj : Kj ֒→ K exists. Moreover, T contains the unit element which can be
identified with the unit of the field F .
For each proper left ideal B in T the intersection Bj = B ∩ hj(Kj) is a
left ideal of Kj. In view of Theorem I.9.1 in [7] Bj = (0), since F ⊂ T /B
and the unit is unique in the associative algebra T /B. Particularly, for a
maximal proper left ideal B in T this induces the embedding t ◦ hj of Kj into
the quotient algebra T /B over the field F for each j ∈ P , where t : T → T /B
denotes the quotient F -linear mapping.
Then equations ajxj = bj and yjaj = bj with aj 6= 0 and bj in hj(Kj)
have unique solutions xj and yj in hj(Kj) for each j ∈ P . For an arbitrary
a ∈ T /B take an element c ∈ t−1(a). Then c + B = t−1(a) and consequently,
hj(Kj) ∩ t−1(a) = hj(Kj) ∩ {c}, where {c} denotes the singleton in T . At
the same time, qu = 0 for some q and u in T implies t((q + B)(u + B)) = 0
in T /B. Therefore, equations ax = b and ya = b with a 6= 0 and b in T /B
have unique solutions x and y in T /B, since B is the proper maximal left
ideal in T and hj(Kj) ∩ hi(Ki) = F for each j 6= i in P . From Theorem
9.2 and Corollary 9.3 in [27] it follows that an embedding of T /B into a
unique-division algebra L over F exists. Taking the intersection of all such
algebras L one gets a minimal unique-division algebra K over F containing
T /B. Thus the embedding of Kj into the division algebra K exists for each
j ∈ P .
28. Theorem. If conditions of Proposition 27 are fulfilled and each Kj
is a Hausdorff topological division algebra with a topology τj such that
(i) τjKi∩Kj = τiKi∩Kj
for each i and j in P , then a Hausdorff topology τ on K exists such that an
19
embedding hj : Kj → K is a homeomorphism of (Kj, τj) onto (hj(Kj), τ ∩
hj(Kj)) for each j ∈ P . Moreover, if each Kj is ultranormed and
(ii) · Kj Ki∩Kj = · Kj Ki∩Kj
for each i and j in P , where · Kj denotes an ultranorm on Kj, then K is
ultranormed.
Proof. Consider on the weak product S = Q′
j∈P Kj the box product
topology, where each s ∈ S has the form s = (sj : ∀j sj ∈ Kj, card{j :
sj 6= ej} < ℵ0), where ej = 1 denotes the unit element in Kj. It induces the
corresponding topology ap on the tensor product T , where T is the quotient
algebra S/M of S by the submodule M having elements of the form
(1) (x) + (y) − (z) with xi + yi = zi for one index i ∈ P and with
xj = yj = zj for each j 6= i in P ;
(2) (x) − (y) with xi = byi for one index i in P and xj = yj for each
j 6= i in P for every b ∈ F , (x), (y) and (z) in S (see also Chapter 3 in
[7]). The algebra T is supplied with the multiplication prescribed by the rule
(x)(y) = Nj∈P xjyj for each (x) and (y) in T . Due to condition (i) for each
i there exists an algebraic topological embedding of Ki into T .
The algebra K over F is obtained as the unique-division algebra K over
F containing T /B (see the proof of Proposition 27).
The algebra T is unital, since Kj is unital for each j. There exists a
neighborhood Wj of 1 in Kj such that the inversion is continuous on Wj for
each j. Take W = Q′
j∈P Wj, hence W is a neighborhood of 1 in T such
that the inversion is continuous on W , since S is supplied with the box
topology. Therefore, if B is a left maximal ideal in T , then B is closed in T ,
since algebraic operations on T are continuous and T is with the continuous
inverse on W .
Therefore, the box topology ap on T induces the quotient T1-topology bp
on T /B, since ap is the Hausdorff topology and B is closed in T . Consider a
base U of a topology τ on K satisfying the conditions:
(3) Ux = U0 + x for each x ∈ K,
(4) Ux = xU1 = U1x for each nonzero x in K,
(5) U0 ∩ (T /B) is the base of neighborhoods of zero in the bp topology on
T /B, where Ux denotes a base of neighborhoods of an element x in K such
that Ux ⊂ U;
(6) for each E and D in U0 there exists C ∈ U0 such that C ⊂ E ∩ D;
(7) TV ∈U0 V = {0};
20
(8) for each E ∈ U0 there exists D ∈ U0 such that (D + D) ⊂ E and
(D + 1)2 ⊂ (E + 1);
(9) for each E ∈ U0 there exists D ∈ U0 such that −D ⊂ E and (D +
1)−1 ⊂ (E + 1),
(10) UF provides the base of the τjF topology on F .
This is possible, since F ⊂ Ki and condition (i) is fulfilled for each i
and j in P and since the bp topology on T /B satisfies analogous to (3 − 10)
conditions due to Theorem 1.3.12 in [5].
Each element of K is obtained from elements of T /B by a finite number of
algebraic operations. Therefore, the intersection of all such bases U satisfying
conditions (3) − (9) provides a minimal base possessing these properties. In
view of Theorem 1.3.12 in [5] this induces a Hausdorff topology τ on K.
From the construction above it follows that τ ∩(T /B) = bp, consequently,
τ ∩hj(Kj) is equivalent with the topology τ ∩hj(Kj) on hj(Kj) inherited from
(T /B, bp) for each j ∈ P , where hj is the algebraic embedding as in subsection
27. Therefore hj is the homeomorphism of (Kj, τj) onto (hj(Kj), τ ∩ hj(Kj))
for each j ∈ P .
In particular, if Kj is ultranormed for each j, then T is ultranormed by
x = supj xjKj , where hx = hx for each h ∈ F and x ∈ T . Such
ultranormed topology is not stronger than the ap topology. By condition (ii)
of this theorem ultranorms · Kj and · Ki on Ki ∩ Kj are equivalent for
each i and j in P , hence there exists an algebraic isometric embedding of
Ki into T for each i. On the other hand, F ⊂ Ki for each i. This induces
the quotient ultranorm on T /B relative to which hj is continuous for each j.
Therefore, U0 and U1 on K can be chosen countable and such that V +V ⊆ V
for each V ∈ U0, W W ⊆ W for each W ∈ U1. Thus K is ultranormable:
x + yK ≤ max(xK, yK) and xyK ≤ xKyK for each x and y in K.
29. Corollary. If conditions of Theorem 28 are satisfied, then a com-
pletion K of K relative to a left uniformity lτ induced by τ exists such that
K is a division algebra. Moreover, if Condition 28(ii) is fulfilled, then K is
the Banach division algebra.
Proof. Consider on the multiplicative group K ∗ of nonzero elements of
K the left uniformity lτ induced by τ . In view of §8.1.17 and Theorem 8.3.10
in [15] and conditions 28(3, 4) the completion K of K relative to lτ is the
unique-division algebra over F .
If in addition Condition 28(ii) is fulfilled, then we take K as the comple-
tion of K relative to its ultranorm.
21
30. Theorem. Let A be a simple annihilator Banach algebra over an
ultranormed field F , F ∈ En, with Rc(A) = R(A), and let {wj} be a maximal
orthogonal system of irreducible idempotents in it. Then an ultranormed
Banach division algebra G exists such that wjAwj ⊂ G for each irreducible
idempotent wj in A, also Pi,j wiAwj is dense in A.
Proof. Suppose that {wj} is a maximal orthogonal system of irreducible
idempotents in the algebra A. For a chosen idempotent wi one gets the two-
sided non nil ideal AwiA. Since A is simple, then clA(AwiA) = A, where
clAS denotes the closure of a subset S in A. Therefore, wjAwiAwj 6= (0)
for each j. Moreover, wjAwiAwj ⊆ wjAwj = Gjwj, where Gj is a division
algebra over F according to Theorem 25.
Consider the algebra Aj := AGj = A ⊗F Gj obtained from A by extension.
For each j elements xj and yj in Aj exist such that wjxjwiyjwj = wj. Put
wj,i = wjxjwi and wi,j = wiyjwj and wj,k = wj,iwi,k. Therefore, wj,i and wi,j
belong to Aj and wj,j = wj, since w2
i = wi. Then one infers that
wj1,k1wk1,k2 = wj1xj1wiwiyk1wk1wk1xk1wiwiyk2wk2
= wj1xj1wi(yk1wk1xk1wi)yk2wk2 and
wj1,k2 = wj1,iwi,k2 = wj1xj1wiwiyk2wk2
= wj1xj1wiyk2wk2.
Mention that this construction implies wi,jwj,i ∈ wiAwi = Giwi and
consequently, wi,jwj,i = bwi for a scalar b = bi,j ∈ Gi. The multiplication of
both sides of the latter equality on the left by wj,i and on the right by wi,j
leads to
wj,iwi,jwj,iwi,j = w2
j = bwj,iwi,j = bw2
j ,
consequently, wj = bwj and hence b = 1. Thus
bwj1,k1wk1,k2 = wj1,iwi,k1wk1,iwi,k2 = wj1,iwiwi,k2 = wj1,k2,
since wj,iwi = wj,i, wiwi,j = wi,j. Then
wj1,k1wj2,k2 = wj1,iwi,k1wj2,iwi,k2 and
wi,k1wj2,i = wiyk1wk1wj2xk2wi, consequently,
(1) wj1,k1wj2,k2 = 0 and
(2) wk1wj2 = 0 for each k1 6= j2, also
(3) wj1,k1wk1,k2 = wj1,k2 for every j1, k1, k2.
Thus the set wjAwk is composed of elements which are multiples of the
element wj,k, consequently, wjxwk,j ∈ wjAwj = Gjwj, where the division
algebra Gj is over the field F according to Theorem 25. Therefore, a scalar
b ∈ Gj exists such that wjxwk,j = bwj. Multiplying on the right by wj,k
22
and using (3) we infer that wjxwk = bwj,k, where b = b(j, k, x) ∈ Gj. This
implies that Pj,k wjAwk =: B ⊂ A, where B is an algebra over F .
By virtue of Theorem 28 and Corollary 29 an ultranormed Banach divi-
sion algebra G exists such that Gj ⊂ G for each j, since Gj is the algebra
over F for each j, also since A is the ultranormed Banach algebra.
We put AG = A ⊗F G, that is AG is the right G-module and the algebra
over F . Thus Pj wjAwj =: E ⊂ AG.
On the other hand, wjAwj ⊂ A as the algebra over F for each j, since
wj ∈ A for each j. Mention that the sum of all wjAwk contains the F -linear
span Y of the set (Pj wjA)(Pk Awk). The multiplication and addition are
continuous on A, hence Y is dense in the F -linear span X of (Pj wjA)A,
since Pk Awk is dense in A. In its turn X is dense in the F -linear span V of
A2, since Pj wjA is dense in A. Therefore, E is dense in A, since V is the
two-sided ideal in A which is necessarily dense in A.
31. Definition. Let X be a Banach space over an ultranormed field F ,
F ∈ En, such that X also has the structure of a right G-module, where G
is a division algebra over F . An operator s ∈ L(X, X) will be called (right)
quasi finite dimensional if its range s(X) is contained in a finite direct sum
x1G ⊕ ... ⊕ xnG embedded into X and such that s is right G-linear, that is
s(xb) = (sx)b for each x ∈ X and b ∈ G, where x1,...,xn are nonzero vectors
belonging to X.
32. Theorem. Let A be a simple annihilator Banach algebra with
Rc(A) = R(A) over an ultranormed field F , F ∈ En. Then an ultranormed
Banach division algebra G exists such that AG := A ⊗F G has an embedding
T into the algebra L(X, X), where X is a Banach space over F and a right
G-module, such that
(1) T (AG) contains all (right) quasi finite dimensional operators so that
T (AG) is a Banach subalgebra in L(X, X) and
(2) a dense subalgebra B in AG exists whose image T (B) consists of quasi
finite dimensional operators.
Proof. Let an ultranormed Banach division algebra G be provided by
Theorem 30. Then wiAwi = Gi ⊆ G for each i and hence wiAGwi = G, since
G is the division algebra over F .
Denote for short AG by A. Then clA(AwiAx) 6= (0) for each x 6= 0,
consequently, AwiAx 6= (0) and hence
(1) wiAx 6= (0) for each x 6= 0 in A.
Next we consider a left regular representation of the algebra A by op-
23
erators Lx for each x ∈ A, where Lxy := xy for each y ∈ A. From prop-
erty (1) it follows that the left regular representation A ∋ x 7→ Lx is the
F -linear isomorphism. On the other hand, Gj ⊂ A and Gj ⊂ G and
GjG = GGj = G for each j. The operator Lx is right G-linear for each
x ∈ A, that is Lx(yb) = (Lxy)b for each y ∈ A and b ∈ G, since A and G
are associative algebras over the field F , also A has the structure of the right
G-module.
In view of Formulas 30(2, 3) the operator Lwk,i maps the one dimensional
over G right module wiAwj into wkAwj, also Lwj,k wlAwi = (0) for k 6=
l. Since the sum Pj wiAwj is dense in wiA, then Lwk,iwiAwj is the one-
dimensional over G right module. Therefore, the operator Lx is quasi finite
dimensional for each x in B := Pj,k wjAwk.
Suppose now that V is a one dimensional operator in wiA over G and
b ∈ wiA is an element such that bA 6= 0. Therefore, wiA = Gb ⊕ N(V ),
where N(V ) := {x ∈ wiA : V x = 0}.
Suppose that L(A, N(V )) = (0). This implies that L(A, N(V )A) =
L(A, N(V )) = (0), since the closed right G-module MN (V )A generated by
N(V )A has the natural embedding ψ into A and ψMN (V )A is a right ideal
in A. Therefore A = ψMN (V )A. On the other hand, N(V ) = wiN(V ),
consequently, MwiN (V )wiA = wiA. Then the identity wiAwi = wiG would
imply that N(V ) = wiA providing the contradiction. This implies that
L(A, N(V )) 6= (0).
Take now x 6= 0 in L(A, N(V )). Let xb = 0, hence wiAx = MGb⊕N (V )x =
(0) contradicting Property (1), consequently, xb 6= 0 and hence xbA is a
non null right ideal in A. Then xbA = wiA, since wiA is the minimal right
ideal in A and xbA ⊆ wiA. Thus an element y ∈ A exists fulfilling the
condition yxb = V b. This implies that the operators V and Lyx coincide
on Gb. Mention that LyxN(V ) = (0) = V N(V ), since x ∈ L(A, N(V )),
hence Lyx = V . Thus all right G-linear one dimensional over G operators
are among Lx, where x ∈ A.
Assume that A is a Banach algebra, then G provided by Theorem 30 is
also a Banach division algebra over F . By the continuity of the multiplication
in A, one gets that wiA is a closed F -linear subspace in A, consequently,
Lx ≤ x for each x ∈ A, since Lxy = xy ≤ xy for each y ∈ A.
Therefore A ∋ x 7→ Lx is the continuous isomorphism into L(X, X) and
each Lx is the limit relative to the operator norm topology of quasi finite
dimensional operators Lxn with xn ∈ B for each n ∈ N.
24
4 B∗-algebras.
1. Definition. Let A be an ultranormed algebra over F ∈ En satisfying the
following conditions:
(1) A is a Banach ∗-algebra and
(2) there exists a bilinear form (·, ·) : X 2 → F such that (x, y) ≤ qxy
for all x and y in A, where 0 < q < ∞ is a constant independent of x and y,
(3) (x, y) = (y, x) and (x, y) = (x∗, y∗) for each x and y in A,
(4) if (x, y) = 0 for each y ∈ A, then x = 0;
(5) (xy, z) = (x, zy∗) for every x, y and z in A,
(6) xx∗ 6= 0 for each nonzero element x ∈ A \ (0).
Then we call A a B∗-algebra.
2. Lemma. For a ∗-subalgebra A of L(X, X) with X = c0(N, F ),
F ∈ En, a bilinear form (·, ·) satisfying conditions 1(2, 3, 5) exists.
Proof. We put (x, y) = T r(x∗Sy), where S is a marked compact operator
such that S∗ = S, S ∈ Lc(X, X), X = c0(N, F ), the trace T r(C) = Pj Cj,j
is defined for each compact operator C ∈ Lc(X, X).
In view of Theorem
2.4 T r(x∗Sy) exists for each x and y in L(X, X). Since T r(C) ≤ C and
x∗Sy ≤ xSy, then condition 1(2) is valid. From T r(C ∗) = (T r(C))∗ =
T r(C) for each C ∈ Lc(X, X) and (x∗Sy)∗ = y∗Sx property 1(3) follows,
since t∗ = t for each t ∈ F . Then using the identity T r(CD) = T r(DC) =
Pk,j Ck,jDj,k for each C ∈ Lc(X, X) and D ∈ L(X, X) we deduce that
(xy, z) = T r(y∗x∗Sz) = T r(x∗Szy∗) = (x, zy∗) for every x, y and z in
A, since (xy)∗ = y∗x∗.
3. Lemma. If conditions of Lemma 2 are satisfied and Lc(X, X) ⊆ A,
then conditions 1(4, 6) are also valid.
Proof. Choose S ∈ Lc(X, X) for which the decomposition S = T −1Y T
is such that T : X → X is an automorphism of the Banach space X and
S∗ = S, also Y ej = Yj,jej with Yj,j 6= 0 for each j, while Yi,j = 0 for each
i 6= j, where {ek : k ∈ N} is the standard basis of X. Then we get property
1(4), since T r(x∗Sy) ∈ F .
On the other hand, (ax)(ax)∗ = a(xx∗)a∗ and (xx∗)∗ = xx∗. Therefore,
considering a ∈ A of the form a = Pk,j aj,kEj,k with aj,k ∈ F one finds
coefficients aj,k such that (ax)(ax)∗ 6= 0, since Ej,k ∈ Lc(X, X) for each j, k
and Lc(X, X) ⊆ A, where Ej,k = e′
j = θ(ej) for each j (see also
Definition 2.5). Mention that (ax)(ax)∗ 6= 0 implies that xx∗ 6= 0, since the
algebra A is associative. Thus property 1(6) also is fulfilled.
j ⊗ ek, e′
25
4. Lemma. If Jr and Jl are proper or improper right and left ideals in a
B∗-algebra A, then L(A, Jr) and R(A, Jl) are orthogonal relative to the family
r and J ∗
of bilinear functionals {(·, ·)a : a ∈ A} complements of the sets J ∗
in
l
the Banach space A, where (x, y)a = (ax, ay) for every a, x and y in A.
r ) = 0. This means that x ∈ A ⊖ J ∗
Proof. If x ∈ L(A, Jr), then xJr = (0), hence (axJr, aA) = 0 for each
r ) = 0 by identity 1(5) and inevitably
r relative to {(·, ·)a : a ∈ A},
r . Similarly R(A, Jl) is the
in A as the Banach space relative to the family
a ∈ A and consequently, (ax, aAJ ∗
(ax, aJ ∗
that is L(A, Jr) is the orthogonal complement of J ∗
orthogonal complement of J ∗
l
{(·, ·)a : a ∈ A} of bilinear functionals.
5. Proposition. Any B∗-algebra A is dual.
Proof.
If Jr and Jl are right and left ideals in A, then by Lemma
r ) = A ⊖ (A ⊖ Jr) = Jr and analogously
4 R(A, L(A, Jr)) = R(A, A ⊖ J ∗
L(A, R(A, Jl)) = Jl, since A∗ = A and (J ∗
r )∗ = Jr.
6. Theorem. Any B∗-algebra A over a field F ∈ En with Rc(A) =
R(A) is representable as the direct sum of its two-sided minimal closed ideals
which are simple B∗-algebras and pairwise orthogonal relative to the family
of bilinear functionals {(·, ·)a : a ∈ A}.
Proof. By virtute of Theorem 3.24 and Proposition 5 the algebra A is the
completion (relative to the ultranorm) of the direct sum of its minimal closed
two-sided ideals which are simple dual subalgebras. Consider a two-sided
minimal closed non null ideal J in A. The involution mapping x 7→ Ix = x∗
provides from it the minimal closed two-sided ideal J ∗ due to Condition
2.5(1).
Suppose that J ∗ 6= J, then JJ ∗ = (0), since the ideal J is minimal. From
aJ ⊂ J and Ja ⊂ J for each a ∈ A we deduce that AJJ ∗A = (0). Together
with condition 1(6) imposed on the B∗-algebra this would imply that x = 0
for each x ∈ J contradicting J 6= (0). Thus J ∗ = J.
Mention that properties 1(1 − 3) and 1(5) for J are inherited from that
of A. Then condition 1(6) on A implies that J 2 6= (0), since J ∗ = J and
AJ ⊆ J, also JA ⊆ J. But J is minimal, hence J 2 = J. Therefore, property
1(4) on J follows from that of on A and 1(5) and J 2 = J, since for each u ∈ J
there exists x and y in J with u = xy and (u, z) = (xy, z) = (x, zy∗) for all
z ∈ A, also since zy∗ ∈ J. Then for each y ∈ J \ (0) an element x ∈ J \ (0)
exists such that xy 6= 0, hence u = xy ∈ J \ (0). Then we have that uu∗ 6= 0
by 1(6) on A. Hence (xy)(xy)∗ 6= 0, consequently, yy∗ 6= 0, since the algebra
A is associative and x(yy∗)x∗ 6= 0. Therefore property 1(6) on J is valid.
26
Thus J is the B∗-algebra.
If J and S are two distinct minimal closed two-sided ideals in A, then
JS = (0). From Lemma 4 it follows that S ⊂ R(A, J) = A⊖J ∗ = A⊖J. Thus
these ideals J and S are orthogonal relative to the family {(·, ·)a : a ∈ A} of
bilinear functionals.
Using condition 1(4) and Lemma 4 we infer that A is the direct sum of
its two-sided minimal closed ideals.
7. Theorem. Let A be a simple B∗-algebra over a field F ∈ En with
Rc(A) = R(A) and let a division algebra G be provided by Theorem 3.30.
Then the following conditions are equivalent:
(1) AG is finite dimensional over G;
(2) AG is unital;
(3) the center Z(AG) of AG is non null.
Proof. Let {wj : j ∈ Λ} be a maximal system of irreducible idempotents
provided by Theorem 3.30.
(1) ⇒ (2). If AG is finite dimensional over G, then according to Theorem
6 a maximal system {wj : j ∈ Λ} of irreducible idempotents is finite, that is
card(Λ) < ℵ0. Then their sum w = Pj∈Λ wj is the idempotent fulfilling the
condition x = Pj∈Λ xwj = xw and x = Pj∈Λ wjx = wx. Thus w is the unit
in AG.
(2) ⇒ (3). If AG contains a unit w, then Z(AG) contains w, consequently,
Z(AG) is non null.
(3) ⇒ (1). Let Z(AG) 6= (0) and x be a non zero element of Z(AG), x 6= 0.
In view of Theorem 3.30 xwj = (xwj)wj = wjxwj = w2
j xwj, hence xwj =
bjwj = wjbjwj, where bj ∈ G. Thus (bjwj)wj = wj(bjwj). Therefore x =
Pj xwj = Pj bjwj and hence bjwj,k = bjwjwj,k = xwjwj,k = xwj,k = wj,kx =
wj,kwkx = wj,kxwk = wj,kbkwk. Similarly bkwk,j = wk,jbjwj, consequently,
bjwj,kwk,j = bjwj = wj,kbkwkwk,j = wj,kbkwk,j and hence Pj bjwj = bkwk +
Pj,j6=k wj,kbkwk,j = Pj wj,kbkwk,j.
Mention that wjAGwj = Gwj for each j, where wj plays the role of the
unit in Gwj. Then
Gwj ⊇ wj(wj,kAGwk,j)wj = wj,kAGwk,j
= wj,k(wkAGwk)wk,j ⊇ wj,k(wk,jAGwj,k)wk,j = wjAGwj = Gwj
for each j and k, hence Gwk ∋ b 7→ wj,kbwk,j ∈ Gwj is the isomorphism of
ultranormed algebras Gwj with Gwk for each j and k.
Therefore the sum Pj wj,kbkwk,j = Pj wj,kwkbkwkwk,j may converge only
if it is finite. Thus the algebra AG is finite dimensional over G.
27
8. Notation. For the Banach space X = c0(α, F ) over a field F ∈ En by
Ld(X, X) is denoted the space of all bounded F -linear operators U : X → X
satisfying the condition limj,k Uj,k = 0, that is for each t > 0 a finite subset
γ in a set α exists such that Uj,k < t for each j and k with either j ∈ α \ γ
or k ∈ α \ γ. Then for a division algebra H over F and a Banach two-sided
H-module XH = c0(α, H) by Lr,d(XH, XH) we denote the Banach right H-
module of all bounded F -linear right H-linear operators C from XH into XH
of the class Ld, that is C(xb) = (Cx)b for each x ∈ XH and b ∈ H.
9. Theorem. Let A be a spherically complete simple B∗-algebra over
a spherically complete field F ∈ En with Rc(A) = R(A). Let also G be a
division algebra provided by Theorem 3.30 such that s1/2 ∈ G for each s ∈ G,
also G ⊂ A and G∗ = G. Then a Banach two-sided G-module XG exist such
that A and Lr,d(XG, XG) are isomorphic as the Banach right G-modules and
as F -algebras.
Proof. By the conditions of this theorem a division algebra G is such
that wAw ⊂ Gw for each irreducible idempotent w in A. Put H = G ∩ G∗.
From G = G∗ it follows that H = G. If b ∈ H, then b1/2 ∈ G and (b1/2)∗ =
(b∗)1/2 ∈ G, since H ∗ = H, consequently, b1/2 ∈ H.
For each irreducible idempotent w described in the proof of Theorem 3.32
ww∗ 6= 0, since A is the B∗-algebra over F . Then (ww∗)(ww∗)∗ 6= 0, hence
ww∗ww∗ 6= 0 and consequently, w∗ww∗ 6= 0 implying that ww∗w 6= 0, since
(w∗ww∗)∗ = ww∗w and c∗∗ = c for each c ∈ A. Therefore w∗w 6= 0 also.
Since w is the irreducible idempotent and A∗ = A, then w∗ is the irre-
ducible idempotent in the B∗-algebra A. Then we deduce that w∗ww∗ ∈
(w∗AGw∗)w∗ ⊆ G∗w∗ = (wG)∗, since A∗ = A, consequently, an element
s ∈ G∗ \ (0) exists such that w∗ww∗ = sw∗, since w∗ww∗ 6= 0. The latter im-
plies w∗ww∗w = sw∗w. But the elements w∗ww∗w and w∗w are self-adjoint,
hence sw∗w = w∗ws∗ and consequently,
w∗w(s∗)−1 = s−1w∗w.
We put v = s−1w∗w, hence
v∗ = w∗w(s∗)−1 = s−1w∗w = v and
v2 = s−1w∗ws−1w∗w = s−1w∗ww∗w(s∗)−1
= (s−1(sw∗w))(s∗)−1 = w∗w(s∗)−1
= s−1w∗w = v.
Thus v is the self-adjoint idempotent. On the other hand, AGv = AGs−1w∗w ⊆
AGw and AGv 6= 0 and the idempotent w is irreducible, hence the idempotent
v is also irreducible, since AGw is the non null minimal left ideal in AG.
28
Then from the proof of Theorem 3.30 it follows that (vAGv)∗ = v∗A∗
Gv∗ =
vAGv is the self-adjoint division algebra for each such irreducible self-adjoint
idempotent v, consequently, vAGv ⊆ Hv. By the conditions of this theorem
we have A = AG.
The algebra A is simple, that is by the definition each its two-sided ideal
coincides with either (0) or A.
Next we take a maximal orthogonal system {wj : j ∈ Λ} of self-adjoint
idempotents in A and for them elements wj,k as in Theorem 3.30, where Λ is a
set. Hence wj,kw∗
j,k = bwj.
Then bwj = wjb∗, since w∗
j,k. Moreover, b 6= 0,
since wj,k is non null and hence wj,kw∗
j,k is non null. For vj,k = (bj,k)−1/2 wj,k
we deduce that vj,kv∗
j,k ∈ wjAwj and b = bj,k ∈ H exists such that wj,kw∗
j = wj and (wj,kw∗
j,k)∗ = wj,kw∗
j,k = wj, since
b−1/2wj,kw∗
= wj(b1/2)∗(b−1/2)∗ = wj(b−1/2b1/2)∗ = wj,
j,k(b−1/2)∗ = b−1/2bwj(b−1/2)∗
Thus it is possible to choose an element wj,k such that wj,kw∗
since A is associative and b−1/2 ∈ H for each non null b in H, where b = bj,k.
j,k = wj
for each k. Taking a marked element j = j0 and setting wk,j = w∗
j,k and
wl,k = wl,jwj,k for each l and k one gets w∗
l,j = wk,jwj,l = wk,l and
wk,k = wk, also wk,lwi,h = δl,iwk,h for every h, i, k, l. Thus elements wl,k can
be chosen such that w∗
l,k = wk,l for each l and k.
l,k = w∗
j,kw∗
If prove the statement of this theorem for the spherical completion H
of H, then it will imply the statement of this theorem for H. So the case
of the spherically complete division algebra H is sufficient. Then A and H
considered as the Banach spaces over the spherically complete field F are
isomorphic with c0(α, F ) and H with c0(β, F ) due to Theorems 5.13 and
5.16 in [31], where β ⊂ α.
From the proof of Theorem 3.32 it follows that the sum B := Pj,k wjAwk
is dense in A. Conditions 1(2, 3, 5) imply that (xy, z) = (y, x∗z), since t∗ = t
for each t ∈ F . Therefore, from properties 1(2, 3, 5) it follows that if j 6= h
or k 6= l, then (wjxwk, whzwl) = 0 for each x and z in A, since
(wjxwk, whzwl) = (wjx, whzwlw∗
k) = (wjx, whz(wlwk)) = (wjx, 0) = 0 for
each k 6= l, also
(whzwl, wjxwk) = (zwl, w∗
hwjxwk) = (zwl, (whwj)xwk) = (zwl, 0) = 0 for
each j 6= h.
Thus the set {wj,k : j, k} is complete and (wj,kH, wh,lH) = (0) for each
j 6= h or k 6= l, where the latter property is interpreted as the orthogonality.
29
Together with property 1(4) this implies that each element x ∈ A has the
form x = Pj,k∈Λ wj,kxj,k with limj,k wj,kxj,k = 0, since AH is the right H-
module, also A is isomorphic with AG as the F -algebra and the right G-
module, where the series may be infinite, xj,k ∈ H for each j, k ∈ Λ, where
Λ denotes the corresponding set.
Take the two-sided Banach H-module XH = c0(Λ, H) and to each element
x ∈ B one can pose the operator Tx such that e∗
j Txek = xj,kξj,k, where ξj,k ∈ F
and ξj,k = wj,k for each j and k in Λ, since a ∈ (ΓF ∪ {0}) for each
a ∈ A, where B := Pj,k wjAwk (see above). Then Tx ∈ Lr,d(XH, XH) and
the mapping T : B → Lr,d(XH, XH) is the isometry having the isometrical
extension T : A → Lr,d(XH, XH). The property wj,kw∗
j,k = wj 6= 0 given
above provides wj,k 6= 0 for each j and k ∈ Λ, consequently, T is bijective
from A onto Lr,d(XH, XH), since A is simple.
For each S and V in Lr,d(XH, XH) one has SV (xb) = S(V x)b = (SV x)b
for each b ∈ H and x ∈ XH. Moreover, (SV )j,k ≤ supm Sj,mVm,k, conse-
quently, limj,k(SV )j,k = 0, that is SV ∈ Lr,d(XH, XH). Hence verifying other
properties one gets that Lr,d(XH, XH) also has the F -algebra structure. From
the construction of AH it follows that AH is the F -algebra, since H and A
are F -algebras. Mention that moreover, AH as the F -algebra is isomorphic
with the Banach F -algebra Lr,d(XH, XH). By the conditions of this theorem
AH is isomorphic with A as the F -algebra and the right H-module.
10. Theorem. Let A be a spherically complete simple B∗-algebra over
a spherically complete field F ∈ En with Rc(A) = R(A) and Z(A) = F . Let
also G be a division algebra provided by Theorem 3.30 such that s1/2 ∈ G
for each s ∈ G. Then a division subalgebra H of G and a Banach two-sided
H-module XH exist such that AH and Lr,d(XH, XH) are isomorphic as the
Banach right H-modules and as F -algebras.
Proof. In this case H = G ∩ G∗ and instead of A we consider AH =
A ⊗F H.
The B∗-algebra A is simple and central, Z(A) = F , hence the right H-
module AH is simple due to Satz 5.9 in [20] and Theorem 7 above. We denote
AH shortly by A and the rest of the proof is similar to that of Theorem 9.
From Theorems 6, 9 and 10 the corollary follows.
11. Corollary. Suppose that A is a spherically complete B∗-algebra over
a spherically complete field F ∈ En with Rc(A) = R(A) and G is a division
algebra given by Theorem 3.30 so that s1/2 ∈ G for each s ∈ G such that
either (i) G ⊂ A and G∗ = G or (ii) Z(A) = F . Then a division subalgebra
30
H in G with H ∗ = H and two-sided H-modules Xk,H exist such that AH is
the direct sum of Lr,d(Xk,H, Xk,H).
References
[1] Y. Amice. "Interpolation p-Adique"// Bull. Soc. Math. France 92(1964),
117-180.
[2] V. Anashin. "Automata finitness criterion in terms of van der Put series
of automata functions" // p-Adic Numbers Ultrametric Anal. Appl. 4:
2 (2012), 151-160.
[3] I.Ya. Aref'eva. "Holographic relation between p-adic effective action and
string field theory" // Proc. Steklov Inst. Math. 285 (2014), 26-29.
[4] I.Ya. Aref'eva, B. Dragovich, P.H. Frampton, I.V. Volovich. "Wave func-
tions of the universe and p-adic gravity"// Int. J. Modern Phys. 6 (1991),
4341-4358.
[5] A. Arhangel'skii, M. Tkachenko. "Topological groups and related struc-
tures" (Amsterdam: Atlantis Press, 2008).
[6] E. Beckenstein, L. Narici, C. Suffel. "Topological algebras" (Amsterdam:
North-Holland Publishing Company, 1977).
[7] N. Bourbaki. "Alg`ebre" Ch. 1-3 (Berlin: Springer-Verlag, 2007).
[8] N. Bourbaki. "Premi`ere partie. Les
fondamentales de
l'analyse. XI. Livre II. Alg`ebre. Ch. IV. Polynomes et fractions ra-
tionnelles. Ch. V. Corps commutatifs" (Paris: Hermann, 1950).
structures
[9] N. Bourbaki. "Alg`ebre commutative" Ch. 1-7 (Paris; Hermann, 1961-
1965).
[10] B. Diarra. "Ultraproduits ultrametriques de corps values"// Ann. Sci.
Univ. Clermont II, S´er. Math. 22 (1984), 1-37.
[11] B. Diarra. "On reducibility of ultrametric almost periodic linear repre-
sentations"// Glasgow Math. J. 37 (1995), 83-98.
31
[12] B. Diarra, S.V. Ludkovsky "Spectral integration and spectral theory for
non-Archimedean Banach spaces"// Int. J. Math. and Math. Sci. 31: 7
(2002), 421-442.
[13] B. Dragovich. "On measurements, numbers and p-adic mathematical
physics" // p-Adic Numbers Ultrametric Anal. Appl. 4: 2 (2012), 102-
108.
[14] N. Dunford, J.C. Schwartz. "Linear operators" (New York; J. Wiley and
Sons, Inc., 1966).
[15] R. Engelking. "General topology" (Moscow: Mir, 1986).
[16] A. Escassut. "Analytic elements in p-adic analysis" (Singapore: World
Scientific, 1995).
[17] A. Escassut. "Ultrametric Banach algebras" (New Jersey: World Scien-
tific, 2003).
[18] C.J. Isham. "Topological and global aspects of quantum theory". In:
"Relativity, groups and topology.II" 1059-1290, (Les Hauches, 1983).
Editors: R. Stora, B.S. De Witt (Amsterdam: Elsevier Sci. Publ., 1984).
[19] R.V. Kadison, J.R. Ringrose, "Fundamentals of the theory of operator
algebras" (New York: Acad. Press, 1983).
[20] I. Kersten. "Brauergruppen von Korpern" (Braunschweig: Friedr.
Vieweg and Sons, 1990).
[21] K. Kunen. "Set theory" (Amsterdam: North-Holland Publishing Co.,
1980).
[22] S.V. Ludkowski. "Non-archimedean antiderivations and calculus of op-
erators with local spectra"// Far East J. of Mathem. Sciences, 99: 4
(2016), 455-489.
[23] S.V. Ludkovsky. "Quasi-invariant and pseudo-differentiable measures in
Banach spaces" (New York: Nova Science Publishers, Inc., 2008).
32
[24] S.V. Ludkovsky. "Stochastic processes in non-archimedean Banach
spaces, manifolds and topological groups" (New York: Nova Science
Publishers, Inc., 2010).
[25] M.A. Naimark. "Normed rings" (Moscow: Nauka, 1968).
[26] L. Narici, E. Beckenstein. "Topological vector spaces" (New York:
Marcel-Dekker Inc., 1985).
[27] B. H. Neumann. "Embedding non-associative rings in division rings"//
Proc. London Math. Soc. 1 (1951), 241-256.
[28] M. van der Put, J. van Tiel. "Espaces nucl´eaires non-archim´ediens"//
Indag. Mathem. 29 (1967), 556-561.
[29] A.F. Revuzhenko. "Mathematical analysis of non-archimedean vari-
able functions. Specialized mathematical apparatus for stuctural geo-
environment level description" (Novosibirsk: Nauka, 2012).
[30] A. Robert. "Representations p-adiques irr´eductibles de sous-groupes ou-
verts de SL2(Zp)"// C.R. Acad. Sci. Paris S´er. I Math. 298: 11 (1984),
237-240.
[31] A.C.M. van Rooij. "Non-Archimedean functional analysis" (New York:
Marcel Dekker Inc., 1978).
[32] A.C.M. van Rooij, W.H. Schikhof. "Groups representations in non-
Archimedean Banach spaces"// Bull. Soc. Math. France. Memoire. 39-
40 (1974), 329-340.
[33] W.H. Schikhof. "Ultrametric calculus" (Cambridge: Cambr. Univ.
Press, 1984).
[34] W.H. Schikhof. "Non-Archimedean calculus". Nijmegen: Math. Inst.,
Cath. Univ., Report 7812, 130 pages, 1978.
[35] J. van Tiel. "Espaces localement K-convexes, I-III"// Indag. Mathe-
maticae 27 (1965), 249-289.
33
[36] K.S. Viswanathan. "Colliding gravitational plane waves and black hole
creation" // p-Adic Numbers Ultrametric Anal. Appl. 4: 2 (2012), 143-
150.
[37] V.S. Vladimirov, I.V. Volovich, E.I. Zelenov. "p-adic analysis and math-
ematical physics" (Moscow: Nauka, 1994).
[38] A. Weil. "Basic number theory" (Berlin: Springer, 1973).
34
|
1112.0158 | 1 | 1112 | 2011-12-01T12:30:40 | Fusion Frames and the Restricted Isometry Property | [
"math.FA"
] | We will show that tight frames satisfying the restricted isometry property give rise to nearly tight fusion frames which are nearly orthogonal and hence are nearly equi-isoclinic. We will also show how to replace parts of the RIP frame with orthonormal sets while maintaining the RIP property. | math.FA | math |
FUSION FRAMES AND THE RESTRICTED ISOMETRY
PROPERTY
BERNHARD G. BODMANN, JAMESON CAHILL, AND PETER G. CASAZZA
Abstract. We will show that tight frames satisfying the restricted
isometry property give rise to nearly tight fusion frames which are nearly
orthogonal and hence are nearly equi-isoclinic. We will also show how to
replace parts of the RIP frame with orthonormal sets while maintaining
the RIP property.
1. Introduction
Fusion frames are a generalization of frames. Fusion frames were intro-
duced in [8] under the name frames of subspaces and quickly found ap-
plication to problems in sensor networks, distributed processing and more
[5, 9, 10, 17]. For a comprehensive view of the papers on fusion frames
we refer the reader to www.fusionframes.org. While frames decompose a
vector into scalar coefficients, fusion frames decompose a vector into vec-
tor coefficients which can be locally processed and later combined. Fusion
frames are designed to handle modern techniques for information processing
which today emphasizes distributed processing. They allow data processing
to become a two step process where we first perform local processing at
individual nodes in the system and this is followed by integration of these
results at a central processor. This has application to packet-based network
communications, sensor networks, radar imaging and more [4]. This hier-
archical processing helps to design systems which are robust against noise,
data loss, and erasures [1, 9, 10, 17, 16]. Much of the work on fusion frames
has surrounded the construction of fusion frames with specialized properties
[2, 6, 7, 20].
Our goal here is to use tools from compressed sensing, namely matrices
with the restricted isometry property, to construct fusion frames with very
strong properties. Compressed sensing is a very hot topic today because of
its broad application to problems in sparse signal recovery. There is so much
literature in this area it is not possible to adequately represent it here so we
refer the reader to two recent tutorials on the subject and their references
[13, 19]. A fundamental tool in this area is the restricted isometry property
(RIP) (See section 4 for definitions). This is a very powerful property for a
B. G. Bodmann was supported by NSF DMS 1109545 and by AFOSR FA9550-11-1-
0245. J. Cahill and P. G. Casazza were supported by NSF DMS 1008183, DTRA/NSF:
ATD 1042701, AFOSR FA9550-11-1-0245.
1
2
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
family of vectors {ϕi}M
i=1 in HN which yields that subsets of a fixed size are
nearly orthonormal. As such, it is quite difficult to produce such families
of vectors of the needed sizes and they are constructed by probabilistic
methods. It is a fundamental open problem in the area to give a concrete
construction of RIP vectors of the appropriate sizes.
In this paper, we will use tight frames of RIP matrices to construct fusion
frames with some very strong properties. First we will show that we can
construct nearly tight fusion frames which still have the RIP property. Next,
we will construct fusion frames with additional strong properties such as
being nearly equi-isoclinic. Finally, we will see how to replace subsets of our
RIP family with orthonormal sequences while tracking the change in the
RIP constants.
2. Frames and Fusion Frames
Fusion frames are a generalization of frames.
Definition 2.1. A family of vectors {ϕi}i∈I in a Hilbert space H is a frame
for H if there are constants 0 < A ≤ B < ∞ so that for all ϕ ∈ H we have
Akϕk2 ≤Xi∈I
hϕ, ϕii2 ≤ Bkϕk2.
The numbers A, B are lower (respectively, upper) frame bounds for the
frame. If A = B it is an A-tight frame and if A = B = 1, it is a Parseval
frame. If kϕik = c for all i ∈ I this is an equal norm frame and if c = 1 it is
a unit norm frame. The analysis operator of the frame is T : HN → ℓ2(M )
given by
M
T (ϕ) =
Xi=1
hϕ, ϕiiei,
where {ei}M
operator is T ∗ and is given by
i=1 is the coordinate orthonormal basis of ℓ2(M ). The synthesis
T ∗ M
Xi=1
aiei! =
M
Xi=1
aiϕi.
The frame operator is the positive self-adjoint invertible operator S = T ∗T
and satisfies
S(ϕ) =
M
Xi=1
hϕ, ϕiiϕi.
Reconstruction is given by
M
M
ϕ =
Xi=1
hϕ, ϕiiS−1ϕi =
Xi=1
hϕ, S−1/2ϕiiS−1/2ϕi.
In particular, {S−1/2ϕi}M
i=1 is a Parseval frame for HN .
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
3
Frame theory has application to a wide variety of problems in signal pro-
cessing and much more (see the monographs [14, 11] for a comprehensive
view). Fusion frames are a generalization of frames and were introduced in
[8]. While frames decompose a signal into scalar coefficients, fusion frames
decompose signals into vector coefficients which can then be locally pro-
cessed and later combined to draw global conclusions.
Definition 2.2. Given a Hilbert space H and a family of closed subspaces
{Wi}i∈I with associated positive weights vi, i ∈ I, a collection of weighted
subspaces {(Wi, vi)}i∈I is a fusion frame for H if there exist constants 0 <
A ≤ B < ∞ satisfying
Akϕk2 ≤Xi∈I
v2
i kPiϕk2 ≤ Bkϕk2
for any ϕ ∈ H,
where Pi is the orthogonal projection onto Wi.
The constants A and B are called fusion frame bounds. A fusion frame is
called tight if A and B can be chosen to be equal, Parseval if A = B = 1,
and orthonormal if
H = ⊕i∈I Wi.
For 0 < ǫ < 1, the fusion frame is ǫ-nearly tight if there is a constant C so
that A = 1
1+ǫ C, B = (1 + ǫ)C. The fusion frame is equi-dimensional if all
its subspaces Wi have the same dimension.
Notation 2.3. If {Wi}i∈I are subspaces of HN , we define the space
Xi∈I
⊕Wi!ℓ2
with inner product given by
kψik2 < ∞},
The fusion frame operator is the positive, self-adjoint and invertible op-
erator SW : H → H given by
SW ϕ =Xi∈I
v2
i Piϕ, for all ϕ ∈ H.
hψi, ψii.
= {{ψi}i∈I ψi ∈ Wi and Xi∈I
D{ψi}i∈I ,{ ψi}i∈IE =Xi∈I
⊕Wi!ℓ2
T : HN → Xi∈I
T (ϕ) = {viPiϕ}i∈I .
T ∗ ({ψi}i∈I ) =Xi∈I
viψi.
,
The analysis operator of the fusion frame is the operator
given by
The synthesis operator of the fusion frame is T ∗ and is given by
4
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
It is known [10] that {Wi, vi}i∈I is a fusion frame with fusion frame bounds
A, B if and only if AI ≤ SW ≤ BI. Any signal ϕ ∈ H can be reconstructed
[10] from its fusion frame measurements {viPiϕ}i∈I by performing
viS−1(viPiϕ).
ϕ =Xi∈I
A frame {ϕi}i∈I can be thought of as a fusion frame of one dimensional
subspaces where Wi = span {ϕi} for all i ∈ I. The fusion frame is then
{Wi,kϕik}. A difference between frames and fusion frames is that for frames,
an input signal ϕ ∈ H is represented by a collection of scalar coefficients
{hϕ, ϕii}i∈I that measure the projection of the signal onto each frame vector
ϕi, while for fusion frames, an input signal ϕ ∈ H is represented by a
collection of vector coefficients {ΠWi(ϕ)}i∈I corresponding to projections
onto each subspace Wi.
Much work has been put into the construction of fusion frames with spec-
ified properties [2, 6, 7]. We also have a generalization of fusion frames using
non-orthogonal projections [3].
There is an important connection between fusion frame bounds and bounds
from frames taken from each of the fusion frame's subspaces [?].
Theorem 2.4. For each i ∈ I, let vi > 0 and Wi be a closed subspace of H,
and let {ϕij}j∈Ji be a frame for Wi with frame bounds Ai, Bi. Assume that
0 < A = infi∈I Ai ≤ supi∈I Bi = B < ∞. Then the following conditions
hold:
(1) {Wi, vi}i∈I is a fusion frame for H.
(2) {viϕij}i∈I,j∈Ji is a frame for H.
In particular, if {Wi, vi}j∈Ji}i∈I is a fusion frame for H with fusion frame
bounds C, D, then {viϕij}i∈I,j∈Ji is a frame for H with frame bounds AC, BD.
Also, if {viϕij}i∈I,j∈Ji is a frame for H with frame bounds C, D, then {Wi, vi,}j∈Ji}i∈I
is a fusion frame for H with fusion frame bounds C
Corollary 2.5. For each i ∈ I, let vi > 0 and Wi be a closed subspace of
H. The following are equivalent:
B , D
A .
(1) {Wi, vi}i∈I is a fusion frame for H with fusion frame bounds A, B.
(2) For every orthonormal basis {eij}j∈Ki for Wi, the family {vieij}i∈I,j∈Ki
(3) For every Parseval frame {ϕij}i∈I,j∈Ji for Wi, the family {viϕij}i∈I,j∈Ji
is a frame for H with frame bounds A, B.
is a frame for H with frame bounds A, B.
Corollary 2.6. For each i ∈ I, let vi > 0 and Wi be a closed subspace of
H. The following are equivalent:
(1) {Wi, vi}i∈I is a Parseval fusion frame for H.
(2) For every orthonormal basis {eij}j∈Ki for Wi, the family {vieij}i∈I,j∈Ki
(3) For every Parseval frame {ϕij}i∈I,j∈Ji for Wi, the family {viϕij}i∈I,j∈Ji
is a Parseval frame for H.
is a Parseval frame for H.
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
5
3. ǫ-Riesz Sequences
For our work we will need some information concerning ǫ-Riesz sequences.
Definition 3.1. A family of vectors {ϕi}N
i=1 in HN is a Riesz basis with
lower (resp. upper) Riesz bounds 0 < A ≤ B < ∞ if for all scalars {ai}N
we have
i=1
A
N
Xi=1
ai2 ≤ k
N
Xi=1
aiϕik2 ≤ B
N
Xi=1
ai2.
This family of vectors is an ǫ-Riesz basis for HN if for all scalars {ai}N
we have
i=1
1
1 + ǫ
N
Xi=1
ai2 ≤ k
N
Xi=1
aiϕik2 ≤ (1 + ǫ)
N
Xi=1
ai2.
The vectors are an ǫ-Riesz sequence if they are an ǫ-Riesz basis for their
span.
As one can see, ǫ-Riesz sequences are nearly orthonormal. The next few
lemmas will formalize this statement. First we recall that for a linearly
independent set of vectors {ϕi}N
i=1 in HN , the frame bounds of this family
equal the Riesz bounds. It follows that if S is the frame operator for {ϕi}N
then {S−1/2ϕi}N
Proposition 3.2. Let {ϕi}M
ǫ-Riesz sequence. Then for every partition {Ij}r
for all scalars {ai}M
i=1 be a family of unit norm vectors which is a
j=1 of {1, 2, . . . , M} we have
i=1
i=1 is an orthonormal basis for HN .
i=1
1
(1 + ǫ)
r
Xj=1
kXi∈Ij
aiϕik2 ≤
M
Xi=1
ai2 ≤ (1 + ǫ)
r
Xj=1
kXi∈Ij
aiϕik2.
Hence,
1
(1 + ǫ)2
r
Xj=1
kXi∈Ij
aiϕik2 ≤ k
M
Xi=1
aiϕik2 ≤ (1 + ǫ)2
r
Xj=1
kXi∈Ij
aiϕik2.
6
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
Proof. We compute
r
1
(1 + ǫ)
Xj=1
kXi∈Ij
aiϕik2 ≤
1
(1 + ǫ)
ai2
r
Xj=1
(1 + ǫ)Xi∈Ij
ai2
r
=
j=1Ij
ai2
= Xi∈∪r
Xj=1Xi∈Ij
Xj=1
(1 + ǫ)kXi∈Ij
kXi∈Ij
= (1 + ǫ)
Xj=1
≤
r
r
aiϕik2
aiϕik2.
For the hence, we combine the first part of the proposition with the fact
that
M
1
M
M
1 + ǫ
ai2 ≤ k
Xi=1
Lemma 3.3. If {ϕi}N
i=1 is an ǫ-Riesz basis for HN and let S be the frame
operator. Then
aiϕik2 ≤ (1 + ǫ)
Xi=1
Xi=1
ai2.
(cid:3)
Hence,
In general, if 0 < a then
1
1 + ǫ
I ≤ S ≤ (1 + ǫ)I.
1
1 + ǫ
I ≤ S−1 ≤ (1 + ǫ)I.
Hence, if a > 0 then
1
(1 + ǫ)a I ≤ Sa ≤ (1 + ǫ)aI.
1
(1 + ǫ)a I ≤ S−a ≤ (1 + ǫ)aI.
Proof. Let T be the analysis operator for the Riesz basis. By the definition,
for any scalars {ai}N
kT ∗({ai}N
aiϕik2 ≤ (1 + ǫ)k{ai}N
i=1)k2 = k
i=1 we have
i=1k2.
N
And similarly,
Xi=1
i=1)k2 ≥
kT ∗({ai}N
1
1 + ǫk{ai}N
i=1k2.
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
7
It follows that T satisfies the same inequalities. For any ϕ ∈ HN and any
0 < a we have
hSaϕ, ϕi = h(T ∗T )aϕ, ϕi
= h(T ∗T )a/2ϕ, (T ∗T )a/2ϕi
= k(T ∗T )a/2ϕk2
≤ k(T ∗T )a/2k2kϕk2
= kT ∗Tkakϕk2
≤ (1 + ǫ)akϕk2.
This shows that Sa ≤ (1 + ǫ)aI. The lower bound is derived similarly. (cid:3)
Finally, we need to measure the angle between spaces spanned by disjoint
subsets of a ǫ-Riesz sequence.
Proposition 3.4. Let {ϕi}M
i=1 be an ǫ-Riesz sequence and choose any par-
tition {I1, I2} of {1, 2, . . . , M}. If ϕ ∈ span {ϕi}i∈I1 and ψ ∈ span {ϕi}i∈I2
are unit vectors, then
hϕ, ψi < 2ǫ(cid:16)1 +
ǫ
2(cid:17).
Proof. Let ϕ =Pi∈I1 aiϕi and ψ =Pi∈I2 aiϕi and we compute
M
1
1 + ǫ
Xi=1
ai2 ≤ kϕ + ψk2
= kϕk2 + kψk2 + 2Rehϕ, ψi
≤ (1 + ǫ)
ai2.
M
Xi=1
Hence,
M
2Rehϕ, ψi ≤ (1 + ǫ)
ai2 − (kϕk2 + kψk2)
M
Xi=1
Xi=1
= (1 + ǫ −
≤ (1 + ǫ)
= ǫ
2 + ǫ
1 + ǫ
M
Xi=1
ai2 − (
1 + ǫ
M
1
Xi=1
ai2.
ai2 +
1
1 + ǫ Xi∈I2
ai2)
1
1 + ǫ Xi∈I1
ai2
8
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
Next, we observe that hϕ, ψi = maxλ=1 Rehϕ, λψi. Thus, we obtain to-
gether with Proposition 3.2,
hϕ, ψi ≤ ǫ(1 +
≤ ǫ(1 +
ǫ
2
ǫ
2
)
1
1 + ǫ
ai2 +Xi∈I2
Xi∈I1
)(kϕk2 + kψk2) = 2ǫ(1 +
ai2
ǫ
2
).
(cid:3)
4. Fusion Frames and the Restricted Isometry Property
In this section we will show how to use tight frames of vectors which have
the ǫ-restricted isometry property to construct ǫ-nearly tight fusion frames.
Definition 4.1. A family of vectors {ϕi}M
i=1 in HN has the restricted isom-
etry property with constant 0 < ǫ < 1 for sets of size s ≤ N if for every
I ⊂ {1, 2, . . . , M} with I ≤ s, the family {ϕi}i∈I is an ǫ-Riesz basis for its
span.
The restricted isometry property is one of the cornerstones of compressed
sensing. Compressed sensing is one of the most active area of research today
and so we refer the reader to the tutorials [13, 19] and their references for
a background in the area. It is known that the optimal ǫ above is on the
order of
s
N
ǫ ∼
log
M
s
.
Now we will see how tight frames of restricted isometry vectors with
constant ǫ will produce nearly tight fusion frames.
Theorem 4.2. Let {ϕi}M
RIP with constant ǫ for sets of size s. Then for any partition {Ij}K
{1, 2, . . . , M} with Ij ≤ s if we let
i=1 be a unit norm tight frame for HN which has
j=1 of
Wj = spani∈Ij ϕi,
then {Wj, 1}K
j=1 is a fusion frame with fusion frame bounds
M
(1 + ǫ)N
,
M (1 + ǫ)
N
.
Moreover, if L ⊂ {1, 2, . . . , K} and for j ∈ L we have Jj ⊂ Ij with
PK
j=1 Jj ≤ s then for all scalars we have
K
L
L
1
1 + ǫ
Xj=1
kXi∈Jj
aiϕik2 ≤ k
Xj=1Xi∈Jj
aiϕik2 ≤ (1 + ǫ)
Xj=1
kXi∈Jj
aiϕik2.
To prove the theorem we need a lemma.
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
9
Lemma 4.3. Under the assumptions of the theorem, if Pj is the orthogonal
projection of HN onto Wj, then for any ϕ ∈ HN we have:
1
1 + ǫXi∈I
hϕ, ϕii2 ≤ kPjϕk2 ≤ (1 + ǫ)Xi∈I
hϕ, ϕii2.
Hence,
M
(1 + ǫ)N kϕk2 ≤
K
Xj=1
kPjϕk2 ≤ (1 + ǫ)
M
N kϕk2.
Proof. Let S be the frame operator:
Sϕ =Xi∈I
hϕ, ϕiiϕi, for all ϕ ∈ HN .
Let {ej}M
j=1 be the eigenbasis for S with eigenvalues
(1 + ǫ) ≥ λ1 ≥ ··· ≥ λI ≥
1
1 + ǫ ≥ 0 ≥ 0 ≥ ··· ≥ 0.
Then
So
On the other hand,
and so
That is,
Pjϕ =
I
Xj=1
hϕ, ejiej.
kPjϕk2 =
I
Xi=1
hϕ, eii2.
Sϕ =
I
Xj=1
λjhϕ, ejiej,
hSϕ, ϕi =
I
Xj=1
λjhϕ, eji2.
I
kPjϕk2 =
hϕ, eji2
Xj=1
≤ (1 + ǫ)
I
Xj=1
λjhϕ, eji2
= (1 + ǫ)hSϕ, ϕi
I
= (1 + ǫ)
hϕ, ϕji2
Xj=1
10
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
The other inequality is similar.
For the hence, we just observe that
hϕ, ϕii2 =
M
N kϕk2.
M
Xi=1
(cid:3)
Proof of Theorem 4.2:
For each j = 1, 2, . . . , K let Pj be the othogonal projection of HN onto
K
Wj. Then by the Lemma 4.3, for any ϕ ∈ HN we have:
Xj=1
hϕ, ϕji2 = (1+ǫ)
kPjϕk2 ≤ (1+ǫ)
M
Xi=1
K
Xj=1Xi∈Ij
hϕ, ϕii2 = (1+ǫ)
M
N kϕk2.
Similarly,
kPjϕk2 ≥
1
(1 + ǫ)
K
Xj=1
K
Xj=1Xi∈Ij
This completes the proof.
hϕ, ϕji2 =
1
(1 + ǫ)
M
Xi=1
hϕ, ϕii2 =
1
(1 + ǫ)
M
N kϕk2.
5. Nearly Equi-Isoclinic Fusion Frames and the Restricted
Isometry Property
In this section, we will see how to use tight frames of vectors with the
restricted isometry property to construct nearly equi-isoclinic fusion frames.
Definition 5.1. Given two subspaces W1, W2 of a Hilbert space H with
dim W1 = k ≤ dim W2 = ℓ, the principal angles (θ1, θ2, . . . θk) between the
subspaces are defined as follows: The first principal angle is
θ1 = min{arccos hϕ, ψi : ϕ ∈ SW1, ψ ∈ SW2}
where SWi = {ϕ ∈ Wi : kϕk = 1}. Two vectors ϕ1, ψ1 are called principal
vectors if they give the minimum above.
The other principal angles and vectors are then defined recursively via
θi = min{arccos hϕ, ψi : ϕ ∈ SW1, ψ ∈ SW2, and ϕ ⊥ ϕj, ψ ⊥ ψj, 1 ≤ j ≤ i−1}.
Definition 5.2. Two k-dimensional subspaces W1, W2 of a Hilbert space
are isoclinic with parameter λ, if the angle θ between any ϕ ∈ W1 and its
orthogonal projection P ϕ in W2 is unique with cos2 θ = λ.
Multiple subspaces are equi-isoclinic if they are pairwise isoclinic with the
same parameter λ.
An alternative definition is given in [12] where two subspaces are called
isoclinic if the stationary values of the angles of two lines, one in each sub-
space, are equal. The geometric characterization given by Lemmens and
Seidel [18] is that when a sphere in one subspace is projected onto the other
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
11
subspace, then it remains a sphere, although the radius may change. This is
all equivalent to the principal angles between the subspaces being identical.
Much work has been done on finding the maximum number of equi-
isoclinic subspaces given the dimensions of the overall space and the sub-
spaces (and often the parameter λ). Specifically, Seidel and Lemmens [18]
give an upper bound on the number of real equi-isoclinic subspaces and
Hoggar [15] generalizes this to vector spaces over R and C.
Definition 5.3. Two K-dimensional subspaces W1, W2 with associated or-
thogonal projections P1 and P2 are isoclinic with parameter λ ≥ 0 if
P1P2P1 = λP1 and P2P1P2 = λP2 .
A family of subspaces {Wj} is ǫ-nearly equi-isoclinic if there exists λ ≥ 0
such that for every two subspaces Pi and Pj, i 6= j,
(λ − ǫ2)P1 ≤ P1P2P1 ≤ (λ + ǫ2)P1 and (λ − ǫ2)P2 ≤ P2P1P2 ≤ (λ + ǫ2)P2 .
We will call a equi-dimensional fusion frame {Wi}K
i=1 ǫ-nearly equi-isoclinic
if its subspaces {Wi}K
i=1 are ǫ-nearly equi-isoclinic.
It can be checked that a fusion frame {Wi, 1}K
i=1 is ǫ-nearly equi-isoclinic
if and only if the squared cosines of the principal angles between any two of
its subspaces are within ǫ2 of a fixed λ.
A related property is:
Definition 5.4. A fusion frame {Wi, vi}K
i=1 is ǫ-nearly orthogonal if when-
ever we take unit vectors ϕ ∈ Wi and ψ ∈ Wj for 1 ≤ i 6= j ≤ K we have
hϕ, ψi < ǫ.
An ǫ-nearly orthogonal fusion frame is ǫ-nearly equi-isoclinic by default
in the sense that it satisfies the definition with λ = 0.
Theorem 5.5. Let {ϕi}M
i=1 be a unit norm tight frame for HN which has
the restricted isometry property with constant ǫ for sets of size s. Then for
any partition {Ij}K
j=1 of {1, 2, . . . , M} with Ij ≤ s
2 if we let
Wj = spani∈Ij ϕi,
then {Wj, 1}K
j=1 is a ǫ-tight fusion frame with fusion frame bounds
M
(1 + ǫ)N
,
M (1 + ǫ)
N
.
Moreover, this is a 2ǫ(1 + ǫ)2-nearly orthogonal fusion frame and hence it is
a 2ǫ(1 + ǫ)2-nearly equi-isoclinic fusion frame.
Proof. The first part of the theorem is immediate by Theorem 4.2 and the
moreover part is immediate by Proposition 3.4.
(cid:3)
12
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
6. The Restricted Isometry Property with Orthonormal
Subsets
A natural problem is the following:
Problem 6.1. Can we construct a family of vectors {ϕi}M
i=1 in HN with the
restricted isometry property with constant 0 < ǫ < 1 for sets of size s our of
orthonormal bases for HN ? Or, can they be constructed from orthonormal
sequences each having s elements?
We will now look at how we might try to alter a family of vectors with the
RIP property to a set which contains orthonormal sequences with s vectors
each. We will need a lemma for this proof.
Lemma 6.2. Let W1, W2 be subspaces of HN and let T : W1 → W2 be a
surjection which satisfies
kϕ − T ϕk2 ≤ ǫkϕk2, for all ϕ ∈ W1.
Let P1 be the orthogonal projection of HN onto W1. Then
(1 − ǫ)2kψk2, for all ψ ∈ W2.
kψ − P1ψk2 ≤ 4
ǫ
Hence,
kP1ψk2 ≥ (1 −
Proof. First note that for any ϕ ∈ W1
4ǫ
(1 − ǫ)2 )kψk2.
(1 − ǫ)2kϕk2 ≤ kT ϕk2 ≤ (1 + ǫ)2kϕk2.
Next we have for any ϕ ∈ W1
kϕ− T ϕk2 = kϕ− P1T ϕk2 = kP1(I − T )ϕk2 +k(I − P1)(I − T )ϕk2 ≤ ǫkϕk2.
Let ψ ∈ W2. Choose ϕ ∈ W1 so that T ϕ = ψ. Now we compute
kψ − P1ψk = kψ − P1T ϕk
≤ kψ − ϕk + kϕ − P1T ϕk
≤ kT ϕ − ϕk + kϕ − P1T ϕk
≤ √ǫkϕk + √ǫkϕk
≤ 2√ǫkT −1ψk
≤ 2√ǫkT −1kkψk
≤ 2
√ǫ
1 − ǫkψk.
For the hence, we note that by Pythagoras
kP1ψk2 = kψk2 − k(I − P1)ψk2
(1 − ǫ)2 )kψk2 .
≥ (1 −
4ǫ
(cid:3)
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
13
Now we are ready for the construction of RIP families which contain
orthonormal sets.
Theorem 6.3. Let {ϕi}M
i=1 be a family of vectors in HN having the restricted
isometry property with constant 0 < ǫ < 1 for sets of size s. Partition
{1, 2, . . . , M} into sets {Ij}K
j=1 with Ij ≤ s for all j = 1, 2, . . . , K. For each
j let Sj be the frame operator for {ϕi}i∈Ij . For K1 ≤ K, replace for ach
j ≤ K1 the family {ϕi}i∈Ij by {S−1/2
ϕi}i∈Ij , which is an orthonormal basis
for its span. Then {S−1/2
ϕi}i∈Ij ;j=1,2,...,K1 ∪ {ϕi}i∈Ij :K1+1≤j≤K =: {ψi}M
has the restricted isometry property and for sets J ⊂ {1, 2, . . . , M} with
J ≤ s we have for all families of scalars {ai}i∈J ,
i=1
j
j
(1 + ǫ)2
− 4ǫ(1 + ǫ)pK1(cid:21) Xi∈J
ai2!1/2
(cid:20) 1 − 4ǫ/(1 − ǫ)2
aiψik ≤h((1 + ǫ)3/2 + 4ǫ(1 + ǫ)pK1i Xi∈J
ai2!1/2
.
≤ kXi∈J
Proof. Choose a subset J ⊂ {1, 2, . . . , M with J ≤ s and let Jj = J ∩ Ij for
all j = 1, 2, . . . , K. For each 1 ≤ j ≤ K1 let Pj be the orthogonal projection
of HN onto span {S−1/2
ϕi}i∈Jj . Choose any scalars {ai}i∈Jj :j=1,2...,K. Then
aiS−1/2
aiS−1/2
ϕi +
K1
K1
K
j
j
j
ϕik
(1) k
Xj=1
Pj Xi∈Jj
aiS−1/2
j
ϕik
We will consider the above two sums separately. By Lemma 3.3 we have
(I − S−1/2
j
)2 ≤(cid:18)1 −
1
√1 + ǫ(cid:19)2
ǫ
1 + ǫ
I.
I ≤
Applying Lemma 3.3 and using T = S−1/2 in Lemma 6.2 we have for all
j = 1, 2, . . . , K1
k(I − Pj)Xi∈Jj
aiS−1/2
j
ϕik ≤
1/2
4 ǫ
1+ǫ
(1 − ǫ
ai2
1+ǫ )2
Xi∈Jj
ai2
= 4ǫ(1 + ǫ)
Xi∈Jj
1/2
.
≤ k
K1
Xj=1Xi∈Jj
K1
aiS−1/2
j
ϕi +
≤ k
Xj=1
Pj Xi∈Jj
K
Xj=K1+1Xi∈Jj
Xj=K1+1Xi∈Jj
aiS−1/2ϕi +
K
aiϕik−k
(I−Pj)Xi∈Jj
aiϕik
Xj=1
Xj=K1+1Xi∈Jj
Xj=1
(I − Pj)Xi∈Jj
aiϕik + k
K1
14
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
Hence,
K1
(2)
k
Xj=1
(I − Pj)Xi∈Jj
aiS−1/2
j
K1
ϕik ≤
j
K1
1/2
aiS−1/2
Xj=1
≤ 4ǫ(1 + ǫ)
k(I − Pj)Xi∈Jj
ϕik
ai2
Xj=1
Xi∈Jj
≤ 4ǫ(1 + ǫ)pK1
ai2
Xj=1Xi∈Jj
K1
1/2
For the second term, since the vector
K1
Xj=1
Pj Xi∈Jj
aiS−1/2
j
ϕi +
K
Xj=K1+1Xi∈Jj
aiϕi,
is contained in the span of the vectors {ϕi}i∈Jj :j=1,2,...,K and
K
which is an ǫ-Riesz sequence, we have by Proposition 3.2
Jj = J ≤ s,
Xj=1
Since {S−1/2
j
(3)
aiϕik2
kXi∈Jj
aiϕik2
kXi∈Jj
aiϕik2
aiϕik2
ai2
j
K
K
1
K1
K1
K1
j
j
ϕi +
aiS−1/2
aiS−1/2
aiS−1/2
ϕik2 +
(1 + ǫ)2
kPj Xi∈Jj
Xj=1
Pj Xi∈Jj
Xj=1
≤ k
≤ (1 + ǫ)2
kPj Xi∈Jj
Xj=1
Xj=K1+1
Xj=K1+1Xi∈Jj
Xj=K1+1
ϕik2 +
ϕi}i∈Jj is an orthonormal set, we have
kXi∈Jj
Xj=K1+1
kPj Xi∈Jj
Xj=1
Xj=K1+1Xi∈Jj
Xj=1
kXi∈Jj
ϕik2 + (1 + ǫ)
Xj=1Xi∈Jj
ai2
ϕik2 +
aiS−1/2
aiS−1/2
≤
K1
=
K1
j
j
K1
K
K
K
ai2 + (1 + ǫ)
≤ (1 + ǫ)Xi∈J
K
Xj=K1+1Xi∈Jj
ai2.
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
15
Similarly, applying the hence from Lemma 6.2 we have
(4)
K1
Xj=1
kPj Xi∈Jj
Xj=1
≥ (1 − 4ǫ/(1 − ǫ)2)
= (1 − 4ǫ/(1 − ǫ)2)
K1
aiS−1/2
j
ϕik2 +
K
Xj=K1+1
ϕik2 +
kXi∈Jj
1
aiϕik2
K
(1 + ǫ)
aiS−1/2
j
kXi∈Jj
Xj=1Xi∈Jj
K1
ai2 +
1
(1 + ǫ)
ai2
Xj=K1+1Xi∈Jj
ai2
K
Xj=K1+1Xi∈Jj
≥ (1 − 4ǫ/(1 − ǫ)2)Xi∈J
Putting this second part together we have
ai2.
K1
Pj Xi∈Jj
Xj=1
k
≤ (1 + ǫ)2
kPj Xi∈Jj
Xj=1
K1
aiS−1/2
j
ϕi +
aiϕik2
K
Xj=K1+1Xi∈Jj
Xj=K1+1
ϕik2 +
K
aiS−1/2
j
≤ (1 + ǫ)3Xi∈J
ai2.
K
kXi∈Jj
aiϕik2
K1
j
1
K1
ϕi +
aiS−1/2
Pj Xi∈Jj
Xj=1
k
(1 + ǫ)2
kPj Xi∈Jj
Xj=1
1 − 4ǫ/(1 − ǫ)2
aiS−1/2
j
aiϕik2
Xj=K1+1Xi∈Jj
Xj=K1+1
ϕik2 +
K
kXi∈Jj
aiϕik2
≥
(1 + ǫ)2 Xi∈J
ai2.
Similarly,
(5)
≥
And by equation 4 we can continue this inequality to
Finally, combining equations 1, 2, and 3 we have:
K1
k
Xj=1Xi∈Jj
aiS−1/2
j
ϕi +
≤ k
K1
Xj=1
Pj Xi∈Jj
aiS−1/2ϕi +
K
Xj=K1+1Xi∈Jj
K
aiϕik
Xj=K1+1Xi∈Jj
(I − Pj)Xi∈Jj
Xj=1
aiϕik + k
K1
aiS−1/2
j
ϕik
16
B.G. BODMANN, J. CAHILL, AND P.G. CASAZZA
K1
≤ (1 + ǫ)3/2 Xi∈J
ai2!1/2
+ 4ǫ(1 + ǫ)pK1
Xj=1Xi∈Jj
ai2!1/2
≤h((1 + ǫ)3/2 + 4ǫ(1 + ǫ)pK1i Xi∈J
.
ai2
Similarly, combining equations 1, 4 and 5 we have
1/2
K1
k
Xj=1Xi∈Jj
K1
ϕi +
≥ k
j
aiS−1/2
Pj Xi∈Jj
Xj=1
1 − 4ǫ/(1 − ǫ)2
≥
(1 + ǫ)2
Xi∈J
≥(cid:20) 1 − 4ǫ/(1 − ǫ)2
(1 + ǫ)2
K
K
K1
aiϕik
aiS−1/2ϕi +
Xj=K1+1Xi∈Jj
Xj=1
(I − Pj)Xi∈Jj
Xj=K1+1Xi∈Jj
aiϕik − k
− 4ǫ(1 + ǫ)pK1
ai2!1/2
Xj=1Xi∈Jj
ai2!1/2
− 4ǫ(1 + ǫ)pK1(cid:21) Xi∈J
K1
.
aiS−1/2
j
ϕik
1/2
ai2
(cid:3)
So we can maintain the restricted isometry property after replacement of
some K1 groups of s vectors in the RIP family by orthonormal sets as long
as
0 <(cid:20) 1 − 4ǫ/(1 − ǫ)2
(1 + ǫ)2
− 4ǫ(1 + ǫ)pK1(cid:21)
Solving for K1 we have
K1 <
1
16ǫ2
(1 − 4ǫ/(1 − ǫ)2)2
(1 + ǫ)6
.
So for sufficiently small ǫ, the fraction on the right hand side is close to one
and we can let K1 grow like 1/ǫ2.
References
[1] B.G. Bodmann, Optimal linear transmission by loss-insensitive packet encoding,
Applied and Computational Harmonic Analysis 22 (2007) 274-285.
[2] B.G. Bodmann, P.G. Casazza, J. Peterson, I. Smalyanu and J.C. Tremain, Equi-
isoclinic fusion frames and mutually unbiased basic sequences, Preprint.
[3] J. Cahill, P.G. Casazza and S. Li, Non-orthogonal fusion frames and the sparsity
of fusion frame operators, Preprint.
[4] R. Calderbank, P.G. Casazza, A Heinecke, G. Kutyniok and A. Pezeshki, Sparse fu-
sion frames: existence and construction, Advances in Computational Mathematics,
35 No. 1 (2011) pp. 1-31.
[5] P.G. Casazza and M. Fickus, Minimizing fusion frame potential, Acta. Appl. Math
107 No. 103 (2009) 7 -- 24.
FUSION FRAMES AND THE RESTRICTED ISOMETRY PROPERTY
17
[6] P.G. Casazza, M. Fickus, A. Heinecke, Y. Wang and Z. Zhou, Spectral Tetris fusion
frame constructions, Preprint.
[7] P.G. Casazza, M. Fickus, D. Mixon, Y. Wang and Z. Zhou, Constructing tight
fusion frames, Applied and Computational Harmonic Analysis 30 (2011) 175-187.
[8] P. G. Casazza and G. Kutyniok, Frames of subspaces, Wavelets, frames and oper-
ator theory, Con- temp. Math., vol. 345, Amer. Math. Soc., Providence, RI, 2004,
pp. 87-113.
[9] P. G. Casazza and G. Kutyniok, Robustness of Fusion Frames under Erasures of
Subspaces and of Local Frame Vectors, Radon transforms, geometry, and wavelets
(New Orleans, LA, 2006), 149 -- 160, Contemp. Math. 464, Amer. Math. Soc., Prov-
idence, RI, 2008.
[10] P. G. Casazza, G. Kutyniok, and S. Li, Fusion frames and distributed processing,
Appl. Comput. Harmon. Anal. 25 (2008), no. 1, 114-132.
[11] O. Christensen, An Introductin to Frames and Riesz Bases, Birkhauser, Boston
(2003).
[12] B. Et-Taoui, Equi-isoclinic planes in Euclidean even dimensional spaces. Adv.
Geom. 7 (2007), no. 3, 379 -- 384.
[13] M. Fornasier and H. Rauhut, Compressive Sensing,
In O. Scherzer, Ed.
Handbook of Mathematical Methods in Imaging, Springer (2011) 187-228.
(hppt://rauhut.ins.uni-bonn.de/CSFornasierRauhut.pdf)
[14] K. Grochenig, Foundations of time-frequency analysis, Applied and Numerical Har-
monic Analysis, Birkhauser, Boston (2001).
[15] S. G. Hoggar, New sets of equi-isoclinic n-planes from old, Proc. Edinburgh Math.
Soc. (2) 20 (1976/77), 287 -- 291.
[16] G. Kutyniok, A. Peszeshki, A.R. Calderbank, Fusion frames and robust dimension
reduction, In: Proc. 42nd Annual Conference on Information Sciences and Systems
(CISS), Princeton University, Princeton, N.J. (2008) 264-268.
[17] G. Kutyniok, A. Pezeshki, A.R. Calderbank, T. Liu, Robust Dimension Reduction,
Fusion Frames, and Grassmannian Packings, Appl. Comput. Harmon. Anal. 26,
64 -- 76 (2009).
[18] P. W. H. Lemmens, J. J. Seidel, Equi-isoclinic subspaces of Euclidean spaces, Ned-
erl. Akad. Wetensch. Proc. Ser. A 76 Indag. Math. 35 (1973), 98 -- 107.
[19] H. Rauhut, Compressive sensing and structured random matrices,
In M.
Fornasier, Ed., Theoretical Foundations and Numerical Methods for Sparse
Recovery, 9 of Radon Series Comp. Appl. Math. deGruyter (2010) 1-92.
(http://rauhut.ins.uni-bonn.de/LinzRauhut.pdf)
[20] P.G. Massey, M.A. Ruiz and D. Stojanoff, The structure of minimizers of the frame
potential on fusion frames, Journal of Fourier Analysis and Applications 16 No. 4
(2010) 514-543.
Department of Mathematics, University of Houston, Houston, TX 77204-
3008
E-mail address: [email protected]
Department of Mathematics, University of Missouri, Columbia, MO 65211-
4100
E-mail address: [email protected]
Department of Mathematics, University of Missouri, Columbia, MO 65211-
4100
E-mail address: [email protected]
|
1409.3117 | 2 | 1409 | 2015-09-07T13:37:40 | Failure of Nehari's Theorem for Multiplicative Hankel Forms in Schatten Classes | [
"math.FA"
] | Ortega-Cerd\`a -- Seip demonstrated that there are bounded multiplicative Hankel forms which do not arise from bounded symbols. On the other hand, when such a form is in the Hilbert-Schmidt class $\mathcal{S}_2$, Helson showed that it has a bounded symbol. The present work investigates forms belonging to the Schatten classes between these two cases. It is shown that for every $p>(1-\log{\pi}/\log{4})^{-1}$ there exist multiplicative Hankel forms in the Schatten class $\mathcal{S}_p$ which lack bounded symbols. The lower bound on $p$ is in a certain sense optimal when the symbol of the multiplicative Hankel form is a product of homogeneous linear polynomials. | math.FA | math |
form on ℓ2 × ℓ2 is given by
(1)
(a, b) =
mnambn,
∞Xm=1
∞Xn=1
FAILURE OF NEHARI'S THEOREM FOR MULTIPLICATIVE
HANKEL FORMS IN SCHATTEN CLASSES
OLE FREDRIK BREVIG AND KARL-MIKAEL PERFEKT
Abstract. Ortega-Cerd`a -- Seip demonstrated that there are bounded multi-
plicative Hankel forms which do not arise from bounded symbols. On the
other hand, when such a form is in the Hilbert -- Schmidt class S2, Helson
showed that it has a bounded symbol. The present work investigates forms
belonging to the Schatten classes between these two cases. It is shown that for
every p > (1 − log π/ log 4)−1 there exist multiplicative Hankel forms in the
Schatten class S
p which lack bounded symbols. The lower bound on p is in a
certain sense optimal when the symbol of the multiplicative Hankel form is a
product of homogeneous linear polynomials.
1. Introduction
For a sequence = (1, 2, 3, . . . ) ∈ ℓ2 its corresponding multiplicative Hankel
which initially is defined at least for finitely supported a, b ∈ ℓ2. Such forms are
naturally understood as small Hankel operators on the Hardy space of the infinite
polydisc, H 2(D∞). Therefore, one is led to investigate the relationship between the
symbol -- a function on the polytorus T∞ generating the Hankel form -- and the
properties of the corresponding Hankel operator.
In the classical setting, (additive) Hankel forms are realized as Hankel operators
on the Hardy space in the unit disc, H 2(D). Nehari's theorem [8] states that every
bounded Hankel form is generated by a bounded symbol on the torus T.
On the infinite polydisc, the study of the corresponding statement was initiated
by H. Helson [4, pp. 52 -- 54], who raised the following questions.
Question 1. Does every bounded multiplicative Hankel form have a bounded symbol
ψ on the polytorus T∞?
Question 2. Does every multiplicative Hankel form in the Hilbert -- Schmidt class S2
have a bounded symbol?
Helson himself [5] gave a positive answer to Question 2. Ortega-Cerd`a and Seip
[9] proved that there are bounded multiplicative Hankel forms that do not have
bounded symbols, using an idea of Helson [6], and hence gave a negative answer
Date: August 14, 2018.
2010 Mathematics Subject Classification. Primary 47B35. Secondary 30B50.
Key words and phrases. Hankel forms, infinite-dimensional torus, Schatten class, Nehari's
theorem, Dirichlet series.
The first author is supported by Grant 227768 of the Research Council of Norway.
1
2
OLE FREDRIK BREVIG AND KARL-MIKAEL PERFEKT
to Question 1. Furthermore, their argument also quickly produces that there are
compact Hankel forms without bounded symbols (see Lemma 1). In light of these
results, a next natural question to ask is:
Question 3. Does there exist a Hankel form belonging to a Schatten class Sp,
2 < p < ∞, without a bounded symbol? If so, for which values of p does such a
form exist?
We will answer the first part of this question, by showing that for every
p > p0 =(cid:18)1 −
log π
log 4(cid:19)−1
≈ 5.738817179,
there are multiplicative Hankel forms in Sp which do not have bounded symbols.
Our construction relies on independent products of homogeneous linear symbols
and is optimal when testing against products of linear homogeneous polynomials,
see Theorem 4.
It is quite tempting to further conjecture that forms without
bounded symbols can be found in Sp for every p > 2, but our method does not
substantiate this claim.
The paper is organized into two further sections. Section 2 reviews the connection
between multiplicative Hankel forms, the Hardy space of Dirichlet series, and the
Hardy space of the infinite polydisc. In Section 3 the main results are proven.
We let H 2 denote the Hilbert space of Dirichlet series
2. Preliminaries
(2)
f (s) =
ann−s
with square summable coefficients.
coefficients bn and n, respectively, a computation shows that
If g and ϕ are Dirichlet series in H 2 with
which associates the finite non-negative multi-index κ(n) = (κ1, κ2, κ3, . . . ) to n.
The Bohr lift of the Dirichlet series (2) is the power series
(3)
Bf (z) =
anzκ(n),
where z = (z1, z2, z3, . . . ). Hence (3) is a power series in countably infinite number
of variables, but each term contains only a finite number of variables.
Under the Bohr lift, H 2 corresponds to the infinite dimensional Hardy space
H 2(D∞), which we view as a subspace of L2(T∞). We refer to [3] for the details,
mentioning only that the Haar measure of the compact abelian group T∞ is simply
the product of the normalized Lebesgue measures of each variable. In particular,
H 2(Dd) is a natural subspace of H 2(D∞).
A key tool in the study of Hardy spaces of Dirichlet series is the Bohr lift [1]. For
any n ∈ N, the fundamental theorem of arithmetic yields the prime factorization
hf g, ϕiH 2 = (a, b).
n =
pκj
j ,
∞Xn=1
∞Yj=1
∞Xn=1
NEHARI'S THEOREM FOR MULTIPLICATIVE HANKEL FORMS IN Sp
3
A formal computation shows that
hBf Bg, BϕiL2 (T∞) = hf g, ϕiH 2,
allowing us to compute the multiplicative Hankel form (1) on T∞. In the remainder
of this paper we work exclusively in the polydisc, with no reference to Dirichlet
series. Therefore, we drop the notation B and study Hankel forms
f, g ∈ H 2(D∞).
Hϕ(f g) = hf g, ϕiL2(T∞),
(4)
In the previous considerations we had that ϕ ∈ H 2(D∞), but there is nothing to
prevent us from considering arbitrary symbols from L2(T∞). Hence, each ϕ ∈
L2(T∞) induces by (4) a (possibly unbounded) Hankel form Hϕ on H 2(D∞) ×
H 2(D∞). Of course, this is not a real generalization. Each form Hϕ is also induced
by a symbol ψ ∈ H 2(D∞); letting ψ = P ϕ we have Hϕ = Hψ, where P denotes the
orthogonal projection of L2(T∞) onto H 2(D∞).
Note that if ψ ∈ L∞(T∞), then the corresponding multiplicative Hankel form is
bounded, since
Hψ(f g) = hf g, ψi ≤ kfk2 kgk2k ψk∞.
We say that Hϕ has a bounded symbol if there exists a ψ ∈ L∞(T∞) such that
Hϕ = Hψ. As mentioned in the introduction, it was shown in [9] that not every
bounded multiplicative Hankel form has a bounded symbol.
On the polydisc the Hankel form Hϕ is naturally realized as a (small) Hankel
operator Hϕ, which when bounded acts as an operator from H 2(D∞) to the anti-
analytic space H 2(D∞). Letting P denote the orthogonal projection of L2(T∞)
onto H 2(D∞), we have at least for polynomials f ∈ H 2(D∞) that
(5)
Hϕf = P (ϕf ).
It is clear that when written in standard bases, the form Hϕ and the operator Hϕ
both correspond to the same infinite matrix
M =
1
2
3
...
2
4
6
...
3
6
9
...
···
···
···
. . .
.
Finally, we briefly recall the definition of the Schatten classes Sp, 0 < p < ∞.
Assume that the Hankel form Hϕ is compact. Let Λ = {λk}∞
k=1 denote the singular
value sequence of M, which of course is the same as the singular value sequence of
the operator Hϕ. The form Hϕ, or equivalently the operator Hϕ, is in the Schatten
class Sp if Λ ∈ ℓp, and
kHϕkSp = kHϕkSp = kΛkℓp.
3. Results
To prove that there for each p > p0 exist multiplicative Hankel forms in Sp
without bounded symbols, we will assume that every Hϕ ∈ Sp has a bounded
symbol and derive a contradiction. We begin with the following routine lemma.
Lemma 1. Let p ≥ 1. Assume that every Hϕ ∈ Sp has a bounded symbol on T∞.
Then there is a constant Cp ≥ 1 with the property that every Hϕ ∈ Sp has a symbol
ψ ∈ L∞(T∞) with Hϕ = Hψ and such that kψk∞ ≤ CpkHϕkSp.
4
OLE FREDRIK BREVIG AND KARL-MIKAEL PERFEKT
Proof. We will define a lifting operator and show that it has to be continuous by
appealing to the closed graph theorem.
Let BH denote the space of bounded multiplicative Hankel forms. By a standard
argument it is isomorphic to the dual space of the weak product H 2 ⊙ H 2 [6]. In
particular BH is a Banach space under the operator norm. It follows that SpH is
also a Banach space, where SpH denotes the space of multiplicative Hankel forms
in Sp equipped with the norm of Sp.
Now we define
X = L∞(T∞) ∩(cid:0)L2(T∞) ⊖ H 2(T∞)(cid:1) ,
Y = L∞(T∞)/X.
Y is a Banach space under the norm kϕkY = inf {kψk∞ : ψ − ϕ ∈ X}, seeing as
X is a closed subspace of L∞(T∞). Since by assumption every Hϕ ∈ SpH has a
symbol ψ ∈ L∞(T∞), we can define a map T : SpH → Y by T (Hϕ) = ψ. This
is a well-defined linear map since Hϕ = 0 for a symbol ϕ ∈ L∞(T∞) if and only
if ϕ ∈ X. An obvious computation verifies that T is a closed operator, hence
continuous. Therefore, there is a Cp ≥ 1 such that
kT (Hϕ)kY ≤ CpkHϕkSp.
The statement of the lemma follows immediately.
(cid:3)
Given the assumption of the lemma, we hence have for each polynomial f and
form Hϕ ∈ Sp that
hf, ϕi = Hϕ(f · 1) = Hψ(f · 1) = hf, ψi ≤ kψk∞kfk1 ≤ CpkHϕkSpkfk1,
where k · k1 denotes the norm of L1(T∞). We thus obtain
(6)
hf, ϕi
kHϕkSp kfk1 ≤ Cp
for every polynomial f and every Hϕ ∈ Sp. To prove our main result we will
construct a sequence of polynomials and finite rank forms to show that no finite
constant Cp satisfying (6) exists for p > p0, thus obtaining a contradiction to the
assumption of Lemma 1. We will require the following lemma.
Lemma 2. Suppose that ϕ1, ϕ2, . . . , ϕm are symbols that depend on mutually
separate variables and which generate the multiplicative Hankel forms Hϕj ∈ Sp,
1 ≤ j ≤ m. Then
(7)
where ϕ = ϕ1ϕ2 ··· ϕm.
Proof. For 1 ≤ j ≤ m, we let Xj denote the Hardy space of precisely the variables
that the symbol ϕj depends on, and if necessary let X0 denote the Hardy space of
the remaining variables, so that -- as tensor products of Hilbert spaces -- we have
kHϕkSp = kHϕ1kSp kHϕ2kSp ··· kHϕmkSp,
H 2(D∞) = X0 ⊗ X1 ⊗ X2 ⊗ ··· Xm.
similarly to (5) for 0 ≤ j ≤ m. Now, if fj ∈ Xj, 0 ≤ j ≤ m, we observe that
We set ϕ0 = 1 and consider the small Hankel operators eHϕj : Xj → Xj, defined
and hence Hϕ = eHϕ0 ⊗ eHϕ1 ⊗ ··· ⊗ eHϕm.
Hϕ(f0f1 ··· fm) = eHϕ0(f0)eHϕ1(f1) ··· eHϕm(fm),
NEHARI'S THEOREM FOR MULTIPLICATIVE HANKEL FORMS IN Sp
5
way. From this, a short computation shows that
singular values λ of Hϕ are obtained as products λ = λ1λ2 ··· λm, where λj is a
Note that eHϕ0 has the sole singular value 1, of multiplicity 1. It follows that all
singular value of eHϕj , see [2]. The multiplicity of λ is also obtained in the expected
Finally, we have Hϕj = eHϕ0 ⊗ eHϕj , where we now regard eHϕ0 as an operator on
follows that kHϕjkSp = keHϕjkSp , completing the proof.
the Hardy space of the variables of which ϕj is independent. Arguing as above, it
(cid:3)
kHϕkSp = keHϕ1kSp keHϕ2kSp ··· keHϕmkSp .
If f1, f2, . . . , fm are polynomials depending on the same separate variables as
ϕ1, ϕ2, . . . , ϕm, respectively, and we set f = f1f2 ··· fm, then
hf, ϕi = hf1, ϕ1ihf2, ϕ2i···hfm, ϕmi,
kfk1 = kf1k1 kf2k1 ···kfmk1.
Let S be the shift operator Sf (z1, z2, . . .) = f (z2, z3, . . .). Suppose that we can
find polynomials f and ϕ, both depending on the first d variables z1, z2, . . . , zd,
satisfying
(8)
kHϕkSp kfk1
Then, for 1 ≤ j ≤ m, consider the functions
and
ϕj(z) = Sd(j−1)ϕ(z)
hf, ϕi
> 1.
fj(z) = Sd(j−1)f (z).
With Φ = ϕ1ϕ2 ··· ϕm and F = f1f2 ··· fm, Lemma 2 yields
hF, Φi
kHΦkSp kFk1
=(cid:18) hf, ϕi
kHϕkSp kfk1(cid:19)m
→ ∞,
m → ∞,
giving us the sought contradiction to (6). We realize this scheme in the next
theorem.
Theorem 3. For every p > p0 there is a multiplicative Hankel form Hϕ ∈ Sp which
does not have a bounded symbol.
Proof. Let d be a large positive integer to be chosen later. Consider the symbol
ϕ(z) =
z1 + z2 + z3 + ··· + zd
√d
.
It is clear that the sequence = (n)∞
n=1 for the matrix of Hϕ is given by
n =(1/√d if n = pj and 1 ≤ j ≤ d
otherwise
0
,
where pj denotes the jth prime. In other terms, the matrix M of Hϕ, with all zero
rows and columns omitted, is the (d + 1) × (d + 1) matrix
1
√d
0
1
1
...
1
1 1
0 0
0 0
...
...
0 0
···
···
···
. . .
···
1
0
0
...
0
.
6
OLE FREDRIK BREVIG AND KARL-MIKAEL PERFEKT
This matrix is easily seen to have the singular values 1 (with multiplicity 2) and 0
(with multiplicity d − 1), and thus
kHϕkSp = 2
1
p .
We choose f (z) = ϕ(z). Then hf, ϕi = 1, and, moreover, the central limit theorem
for Steinhaus variables gives us that
lim
d→∞kfk1 = lim
d→∞
E(cid:18)z1 + z2 + z3 + ··· + zd
√d
(cid:19) =
√π
2
.
In particular, for each δ > 0 we have for sufficiently large d that
√π
2
+ δ.
kfk1 ≤
We now observe that p = p0 is the solution of the equation 21/p · √π/2 = 1, and
hence if p > p0 we may find δ > 0 small enough that
kHϕkSp · kfk1 ≤ 21/p ·(cid:18)√π
2
+ δ(cid:19) < 1.
This implies that if d is large enough, f and ϕ satisfy (8). This completes the proof
by appealing to the discussion preceding the statement of the theorem.
(cid:3)
Our result is optimal for symbols which are independent products of linear homo-
geneous polynomials and test functions of the same form, as shown by the following
result.
Theorem 4. Suppose p ≤ p0 and consider
and
ϕ(z) = a1z1 + a2z2 + ··· + adzd
f (z) = b1z1 + b2z2 + ··· + bdzd,
for aj, bj ∈ C. Then hf, ϕi ≤ kHϕkSpkfk1.
Proof. By the Cauchy -- Schwarz inequality and Parseval's formula, it is clear that
hf, ϕi ≤ kakℓ2kbkℓ2.
Straightforward computations with the matrix M of Hϕ show that
MM ∗
=
ℓ2
kak2
0
0
...
0
0
0
a1a1
a1a2
a2a2
a2a1
...
...
ada1 ada2
···
···
···
. . .
···
0
a1ad
a2ad
...
adad
.
Here we have again omitted zero rows and columns. Note that the lower right
block has rank 1. By considering the vector (0, a1, a2, . . . , ad) it is clear that it
has the sole eigenvalue kak2
ℓ2. Thus, the singular value sequence of M is Λ =
{kakℓ2, kakℓ2, 0, . . . , 0}, and hence
kHϕkSp = 21/pkakℓ2.
We use the optimal Khintchine inequality for Steinhaus variables [7, 10], p = 1, and
obtain
√π
2 kbkℓ2.
kfk1 ≥
NEHARI'S THEOREM FOR MULTIPLICATIVE HANKEL FORMS IN Sp
7
The hypothesis that p ≤ p0 implies that 21/p√π/2 ≥ 1, and the proof is finished
by the following chain of inequalities.
kHϕkSp · kfk1 ≥ 21/p · kakℓ2 ·
√π
2 · kbkℓ2 ≥ kakℓ2 · kbkℓ2 ≥ hf, ϕi.
(cid:3)
References
1. H. Bohr, Ueber die Bedeutung der Potenzreihen unendlich vieler Variablen in der Theorie
der Dirichletschen Reihe, Nachr. Akad. Wiss. Gottingen Math. -- Phys. 1913 (1913), 441 -- 488.
2. A. Brown and C. Pearcy, Spectra of tensor products of operators, Proc. Amer. Math. Soc. 17
(1966), no. 1, 162 -- 166.
3. H. Hedenmalm, P. Lindqvist, and K. Seip, A Hilbert space of Dirichlet series and systems of
dilated functions in L2(0, 1), Duke Math. J. 86 (1997), no. 1, 1 -- 37.
4. H. Helson, Dirichlet series, Henry Helson, 2005.
5.
6.
7. H. Konig, On the best constants in the Khintchine inequality for Steinhaus variables, Israel
, Hankel forms and sums of random variables, Studia Math. 176 (2006), 85 -- 92.
, Hankel forms, Studia Math. 198 (2010), 79 -- 84.
J. Math. (2013), 1 -- 35.
8. Z. Nehari, On bounded bilinear forms, Ann. of Math. (1957), 153 -- 162.
9. J. Ortega-Cerd`a and K. Seip, A lower bound in Nehari's theorem on the polydisc, J. Anal.
Math. 118 (2012), no. 1, 339 -- 342.
10. J. Sawa, The best constant in the Khintchine inequality for complex Steinhaus variables, the
case p = 1, Studia Math. 81 (1985), no. 1, 107 -- 126.
Department of Mathematical Sciences, Norwegian University of Science and Tech-
nology (NTNU), NO-7491 Trondheim, Norway
E-mail address: [email protected]
Department of Mathematical Sciences, Norwegian University of Science and Tech-
nology (NTNU), NO-7491 Trondheim, Norway
E-mail address: [email protected]
|
1912.13266 | 1 | 1912 | 2019-12-31T11:06:22 | Invertibility, Fredholmness and kernels of dual truncated Toeplitz operators | [
"math.FA"
] | Asymmetric dual truncated Toeplitz operators acting between the orthogonal complements of two (eventually different) model spaces are introduced and studied. They are shown to be equivalent after extension to paired operators on $L^2(\mathbb T) \oplus L^2(\mathbb T)$ and, if their symbols are invertible in $L^\infty(\mathbb T)$, to asymmetric truncated Toeplitz operators with the inverse symbol. Relations with Carleson's corona theorem are also established. These results are used to study the Fredholmness, the invertibility and the spectra of various classes of dual truncated Toeplitz operators. | math.FA | math |
INVERTIBILITY, FREDHOLMNESS AND KERNELS
OF DUAL TRUNCATED TOEPLITZ OPERATORS
M. CRISTINA C AMARA, KAMILA KLI´S -- GARLICKA, BARTOSZ LANUCHA,
AND MAREK PTAK
Abstract. Asymmetric dual truncated Toeplitz operators acting
between the orthogonal complements of two (eventually different)
model spaces are introduced and studied. They are shown to be
equivalent after extension to paired operators on L2(T) ⊕ L2(T)
and, if their symbols are invertible in L∞(T), to asymmetric trun-
cated Toeplitz operators with the inverse symbol. Relations with
Carleson's corona theorem are also established. These results are
used to study the Fredholmness, the invertibility and the spectra
of various classes of dual truncated Toeplitz operators.
0. Introduction
Toeplitz operators have been for a long time one of the most studied
classes of nonselfadjoint operators ([3]). They are defined as compres-
sions of multiplication operators on L2(T), to the Hardy space of the
unit disk H 2(D). Dual Toeplitz operators are analogously defined on
the orthogonal complement of H 2(D), identified as usual with a sub-
space of L2(T), as multiplication operators followed by projection onto
L2(T) ⊖ H 2(D). Although they differ in various ways from Toeplitz
operators, they also share many properties, which is not surprising
given that they are anti-unitarily equivalent. The algebraic and spec-
tral properties of dual Toeplitz operators, and the extent to which their
properties are parallel to those of Toeplitz operators on H 2(D), were
studied in [20].
2010 Mathematics Subject Classification. Primary 47B35, Secondary 47B32,
30D20.
Key words and phrases. truncated Toeplitz operator, dual truncated Toeplitz
operator, paired operator, equivalence after extension, Hardy space, model space,
invariant subspaces for unilateral shift.
The work of the first author was partially supported by FCT/Portugal through
UID/MAT/04459/2019. The research of the second and the fourth authors was
financed by the Ministry of Science and Higher Education of the Republic of Poland.
1
2
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
Truncated Toelitz operators, defined as compressions of multiplica-
tion operators to closed subspaces of H 2(D) which are invariant for the
backward shift S∗, called model spaces, have also generated great inter-
est, partly motivated by Sarason's paper [16]. Their study, as well as
that of asymmetric truncated Toeplitz operators later introduced in [7],
raised many interesting questions and has led to new and sometimes
surprising results, see for example [2, 5, 7, 9]. It is natural to consider
dual truncated Toeplitz operators, defined analogously as compressions
of multiplication operators to the orthogonal complement of a model
space in L2(T). These operators were very recently introduced and
studied in [14, 15, 12]. It turns out that, in this case, they behave very
differently from truncated Toeplitz operators. For instance, the symbol
of a dual truncated Toeplitz operator is unique and the only compact
operator of that kind is the zero operator, in sharp contrast with what
happens with truncated Toeplitz operators on model spaces.
In this paper we study the kernels and various spectral properties,
such as Fredholmness and invertibility, of dual truncated Toeplitz oper-
ators. The results are applied to describe the spectra of dual truncated
Toeplitz operators in several classes including, as particular cases, the
dual truncated shift and its adjoint. We do this by using a novel ap-
proach to dual truncated Toeplitz operators and their asymmetric ana-
logues, defined similarly between the orthogonal complements of two
possibly different model spaces. This involves proving their equivalence
after extension to paired operators in L2(T) ⊕ L2(T), defined in Section
2, and establishing connections with the corona theorem. This allows
moreover to show that, whenever their symbol is invertible in L∞(T),
dual truncated Toeplitz operators are in fact equivalent after extension
to truncated Toeplitz operators with the inverse symbol.
The paper is organized as follows. In Section 1 we introduce asym-
metric dual truncated Toeplitz operators and present some basic prop-
erties, while in Section 2 we recall the concepts of paired operator and
equivalence after extension between two Banach spaces.
In Section
3 we study the solvability of certain equations involving asymmetric
dual truncated Toeplitz operators in connection with equations involv-
ing paired operators. In Section 5 we show that dual truncated Toeplitz
operators are equivalent after extension to truncated Toeplitz opera-
tors with the inverse symbol, if the latter is invertible in L∞(T). In
Section 6 we study the kernels of asymmetric dual truncated Toeplitz
operators in terms of explicitly defined isomorphisms with kernels of
other operators. We show in particular that the kernels of a dual trun-
cated Toeplitz operator and its adjoint are isomorphic and related by
DUAL TRUNCATED TOEPLITZ OPERATORS
3
the usual conjugation on a model space. In Section 7 we present suffi-
cient conditions for a dual truncated Toeplitz operator to be injective
or invertible in terms of certain corona pairs, i.e., pairs of functions
satisfying the hypotheses of Carleson's corona theorem ([17, 11]). We
use the previous results to study the Fredholmness, invertibility and
spectra of several classes of dual truncated Toeplitz operators.
1. Elementary properties
H 2
Let P +, P − be the orthogonal projections from L2 onto H 2 and
− = ¯zH 2, respectively. We have P − = I − P +.
Recall that for ϕ ∈ L∞ = L∞(T) a Toeplitz operator Tϕ : H 2 → H 2
is defined by Tϕf = P +(ϕf ) for f ∈ H 2.
For an inner function θ define the model space Kθ = H 2 ⊖ θH 2 and
let Pθ and Qθ be the orthogonal projections from L2 = L2(T) onto the
model space Kθ and its orthogonal complement (Kθ)⊥ = L2 ⊖ Kθ =
H 2
− ⊕ θH 2, respectively.
Let α, θ be inner functions and let ϕ ∈ L2. An asymmetric truncated
Toeplitz operator Aθ,α
ϕ is defined by
Aθ,α
ϕ f = PαϕPθf,
for f ∈ Kθ ∩ L∞ (see [7, 5]), whereas the asymmetric dual truncated
Toeplitz operator Dθ,α
is defined by
ϕ
Dθ,α
ϕ f = QαϕQθf = P −ϕf + αP + ¯αϕf
for f ∈ (Kθ)⊥ ∩ L∞, which is a dense subset of (Kθ)⊥. If Aθ,α
ϕ or Dθ,α
ϕ
have a bounded extension to Kθ or (Kθ)⊥, respectively, we denote them
also by Aθ,α
ϕ and
Dθ,θ
ϕ we write Aθ
We start with some elementary properties of asymmetric dual trun-
cated Toeplitz operators. These properties were proved in [14] for
α = θ.
ϕ , respectively. When α = θ instead of Aθ,θ
ϕ, respectively.
ϕ and Dθ,α
ϕ and Dθ
Proposition 1.1. Let ϕ ∈ L2. Then Dθ,α
ϕ
ϕ ∈ L∞ and, in that case, kDθ,α
ϕ k = kϕk∞.
is bounded if and only if
Proof. Let ϕ ∈ L2 and take f ∈ H ∞. Then θf ∈ θH ∞ ⊂ θH 2,
Dθ,α
ϕ (θf ) = (P − + αP + ¯α)(ϕθf ) ∈ H 2
− ⊕ αH 2
and
kDθ,α
ϕ (θf )k2 = kP −ϕθf k2 + kαP + ¯αϕθf k2
= kP −ϕθf k2 + kP + ¯αϕθf k2 > kT ¯αθϕ(f )k2.
4
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
ϕ
If Dθ,α
is bounded, then, for some C > 0, kDθ,α
ϕ (θf )k2 6 Ckf k2, so
by the above kT ¯αθϕ(f )k2 6 Ckf k2. Since this holds for any f ∈ H ∞,
it follows that T ¯αθϕ is bounded in H 2 and therefore ¯αθϕ ∈ L∞, which
implies that ϕ ∈ L∞. Moreover, kϕk∞ = kT ¯αθϕk 6 kDθ,α
ϕ k.
On the other hand, if ϕ ∈ L∞, then Dθ,α
is clearly a bounded op-
Indeed, for any f ∈ (Kθ)⊥ we then
ϕ
erator from (Kθ)⊥ into (Kα)⊥.
have
Moreover, kDθ,α
kDθ,α
ϕ k 6 kϕk∞.
ϕ f k = kQαϕf k 6 kϕf k 6 kϕk∞kf k.
(cid:3)
Taking the result of Proposition 1.1 into account, we assume from
now on that ϕ ∈ L∞. It is also easy to see that (Dθ,α
ϕ )∗ = Dα,θ
ϕ .
Proposition 1.2. For ϕ ∈ L∞, we have that Dθ,α
only if ϕ = 0.
ϕ
is compact if and
ϕ is compact and let fn ∈ (Kθ)⊥, with fn weakly
Proof. Assume that Dθ,α
convergent to 0 (fn ⇀ 0). Then kDθ,α
ϕ fnk → 0. Note that for fn =
θ fn with fn ∈ H 2, we have fn ⇀ 0 (in θH 2) if and only if fn ⇀ 0
ϕ (θ fn)k → 0. Since
(in H 2).
ϕ (θ fn)k (see the proof of Proposition 1.1), we have
kT ¯αθϕ fnk 6 kDθ,α
that kT ¯αθϕ fnk → 0 whenever fn ⇀ 0. So, if Dθ,α
ϕ is compact, then T ¯αθϕ
is also compact and therefore ϕ = 0.
(cid:3)
It follows that if fn ⇀ 0, then kDθ,α
Remark 1.3. Proposition 1.2 implies, in particular, that the only sym-
bol for the zero asymmetric dual truncated Toeplitz operator is ϕ = 0.
Since the question of Dθ,α
ϕ being the zero operator is equivalent to ϕ
being a multiplier from (Kθ)⊥ into Kα, we conclude also that there are
no non-trivial L∞-multipliers from (Kθ)⊥ into Kα.
In contrast with
this, the question of whether there are non-trivial multipliers from Kα
into (Kθ)⊥, which is equivalent to Aα,θ
ϕ being the zero operator, has a
positive answer ([7, 5]).
2. Paired operators and equivalence after extension
For a Banach space X denote by L(X) the space of all bounded
linear operators A : X → X. Let P ∈ L(X) be a projection and let
Q = I − P be its complementary projection. An operator of the form
AP + BQ or P A + QB, where A, B ∈ L(X), is called a paired operator
([13]).
Paired operators are closely connected with operators of the form
P CP Im P and QCQIm Q, where C ∈ L(X), which are called general
Wiener-Hopf operators or operators of Wiener-Hopf type ([13, 18]). To
DUAL TRUNCATED TOEPLITZ OPERATORS
5
understand this relation, it will be useful to introduce here the concept
of equivalence after extension for operators.
Definition 2.1 (Equivalence after extension, [1]). Let X, X, Y , Y be
Banach spaces and let us use the term operator to mean a bounded
linear operator.
The operators T : X → X and S : Y → Y are said to be (alge-
braically and topologically) equivalent if and only if T = ESF where
E and F are invertible operators; in that case we use the notation
T ∼ S.
The operators T and S are equivalent after extension (T ⋆∼ S) if
and only if there exist (possibly trivial) Banach spaces X0, Y0, called
extension spaces, and invertible operators E : Y ⊕Y0 → X ⊕X0, F : X ⊕
X0 → Y ⊕ Y0, such that
0
0
(cid:20)T
0 IX0(cid:21) = E(cid:20)S
0 IY0(cid:21) F.
Clearly, if T ∼ S, then T ⋆∼ S. Operators that are equivalent after
extension share many properties. In particular we have the following.
Theorem 2.2 ([1]). Let T and S be two operators, T : X → X,
S : Y → Y , and assume that T ⋆∼ S. Then
(1) ker T is isomorphic to ker S, i.e., ker T ≃ ker S;
(2) Im T is closed if and only if Im S is closed and, in that case,
X/ Im T ≃ Y / Im S;
(3) if one of the operators T , S is generalized (left, right) invertible
([13, 18]), then the other is generalized (left, right) invertible
too;
(4) T is Fredholm if and only if S is Fredholm and, in that case,
dim ker T = dim ker S, codim Im T = codim Im S.
It is not difficult to see that P CP Im P (respectively, QCQIm Q) is
equivalent after extension to CP + Q (respectively, P + CQ) and
(2.1)
because
and
where
CP + Q ∼ P C + Q ∼ P CP + Q,
CP + Q = (P CP + Q)(I + QCP )
P C + Q = (I + P CQ)(P CP + Q),
(I + P CQ)−1 = I − P CQ,
(I + QCP )−1 = I − QCP,
(and analogously P + CQ ∼ P + QC ∼ P + QCQ).
6
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
As an example of two operators which are equivalent after extension
we have the following.
Theorem 2.3 ([7]). Let ϕ ∈ L∞ and let α, θ be inner functions. Then
Aθ,α
ϕ = PαϕPθKθ and TΦ is the block Toeplitz operator
ϕ
⋆∼ TΦ where Aθ,α
on H 2 ⊕ H 2 with symbol Φ = (cid:20) ¯θ
ϕ α(cid:21).
0
Equivalence after extension for two operators T and S implies that
there is a strong connection between the solvability of the equations
T ϕ = ψ and Sx = y, in particular as regards the existence and unique-
ness of solutions.
In the next section we study the relations between the solutions of
the equations
and
Dθ,α
ϕ f = g,
f ∈ (Kθ)⊥, g ∈ (Kα)⊥
(AP + + BP −)Φ = Ψ, Φ, Ψ ∈ L2 ⊕ L2,
where AP + + BP − is a certain paired operator on L2 ⊕ L2, as a first
step to establishing the equivalence after extension of Dθ,α
to a paired
operator AP + + BP −. Here P ±(L2 ⊕ L2) = P ±L2 ⊕ P ±L2.
ϕ
3. Solvability relations
Let ϕ ∈ L∞ and let α, θ be inner functions. We define
A = (cid:20) ϕθ −1
ϕθ ¯α
0(cid:21) , B = (cid:20) ϕ
¯αϕ −1(cid:21) .
0
Theorem 3.1.
(1) Let f−, g− ∈ H 2
−, f+, g+ ∈ H 2. Then
Dθ,α
ϕ (f− + θ f+) = g− + αg+
implies that
(AP + + BP −)Φ = Ψ,
where Φ, Ψ ∈ L2 ⊕ L2 are given by
f− + f+
Φ =(cid:20)φ1
Ψ =(cid:20)ψ1
(1 + ¯α)Pαϕ(f− + θ f+)(cid:21) ,
φ2(cid:21) = (cid:20)
ψ2(cid:21) = (cid:20)g− + αg+
¯αg− + g+(cid:21) .
(2) Let Φ, Ψ ∈ L2 ⊕ L2, Φ = (cid:20)φ1
φ2(cid:21) , Ψ = (cid:20)ψ1
ψ2(cid:21) . Then
(AP + + BP −)Φ = Ψ
DUAL TRUNCATED TOEPLITZ OPERATORS
7
implies that
Dθ,α
ϕ (f− + θ f+) = g− + αg+,
where f−, g− ∈ H 2
− and f+, g+ ∈ H 2 are given by
f− = P −φ1,
f+ = P +φ1,
g− = P −ψ1,
g+ = P +ψ2.
Proof. To prove (1) assume that f−, g− ∈ H 2
by definition Dθ,α
θ f+) = g− + αg+ means that there exists ψ+ ∈ Kα such that
ϕ (f− + θ f+) = Qαϕ(f− + θ f+), the equality Dθ,α
− and f+, g+ ∈ H 2. Since
ϕ (f− +
ϕ(f− + θ f+) = g− + αg+ + ψ+.
Equivalently, there exist ψ+ ∈ H 2 and ψ− ∈ H 2
− such that
i.e.,
(cid:26) ϕf− + ϕθ f+ = g− + αg+ + ψ+,
¯αψ+ = ψ−,
(cid:26)
ϕθ f+ − ψ+ + ϕf− = g− + αg+,
θ ¯αϕ f+ + ¯αϕf− − ψ− = ¯αg− + g+.
This last system of equations can be written in matrix form as
(cid:18)(cid:20) ϕθ −1
ϕθ ¯α
0(cid:21) P + +(cid:20) ϕ
¯αϕ −1(cid:21) P −(cid:19)(cid:20)φ1
φ2(cid:21) = (cid:20)ψ1
ψ2(cid:21)
0
with
and
Ψ = (cid:20)ψ1
ψ2(cid:21) = (cid:20)g− + αg+
¯αg− + g+(cid:21)
Φ = (cid:20)φ1
φ2(cid:21) = (cid:20) f− + f+
ψ− + ψ+(cid:21) = (cid:20)
(1 + ¯α)(ϕf− + ϕθ f+ − g− − αg+)(cid:21) .
f− + f+
Now the result for Φ follows from the fact that
g− = P −(ϕf− + ϕθ f+)
and
To prove (2) let Φ = (cid:20)φ1
g+ = P + ¯α(ϕf− + ϕθ f+).
φ2(cid:21) ∈ L2 ⊕ L2, Ψ = (cid:20)ψ1
ψ2(cid:21) ∈ L2 ⊕ L2 and put
φi± = P ±φi, i = 1, 2. Then (AP + + BP −)Φ = Ψ can be written as a
system of equations
(cid:26)
ϕθφ1+ − φ2+ + ϕφ1− = ψ1,
θ ¯αϕφ1+ + ¯αϕφ1− − φ2− = ψ2,
8
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
which is equivalent to
(cid:26)
ϕφ1− + ϕθφ1+ = ψ1 + φ2+,
¯α(ϕφ1− + ϕθφ1+) = ψ2 + φ2−.
Moreover, the above is equivalent to
(cid:26) ϕ(φ1− + θφ1+) = ψ1 + φ2+,
¯α(ψ1 + φ2+) = ψ2 + φ2−.
The first equation in the system above implies that
Dθ,α
ϕ (φ1− + θφ1+) = Qαϕ(φ1− + θφ1+)
= Qα(ψ1 + φ2+) = P −ψ1 + αP + ¯α(ψ1 + φ2+)
and the second equation gives
αP + ¯α(ψ1 + φ2+) = αP +ψ2,
that is,
Dθ,α
ϕ (P −φ1 + θP +φ1) = P −ψ1 + αP +ψ2.
(cid:3)
The relations in Theorem 3.1 imply that Dθ,α
ϕ and AP + + BP − share
many properties. Indeed, in the next section we show that the former
is equivalent after extension to the latter.
4. Equivalence after extension of Dθ,α
ϕ to a paired
operator
Let us introduce some notations (see [6]). If H, K1, K2 are Hilbert
spaces and A1 : H → K1, A2 : H → K2, B1 : K1 → H, B2 : K2 → H are
bounded linear operators, we define
A1 ⋄ A2 : H → K1 ⊕ K2,
(A1 ⋄ A2)h = A1h ⊕ A2h,
and
B1 ⊞ B2 : K1 ⊕ K2 → H,
(B1 ⊞ B2)(f + g) = B1f + B2g.
In what follows, α and θ are inner functions, ϕ ∈ L∞ and
(4.1)
A = (cid:20) ϕθ −1
ϕθ ¯α
0(cid:21) , B = (cid:20) ϕ
¯αϕ −1(cid:21) ,
0
as at the beginning of Section 3.
Proposition 4.1. Let ϕ ∈ L∞ and let α, θ be inner functions. Then
Dθ,α
ϕ
⋆∼ QαϕQθ ⊞ Pα,
where QαϕQθ ⊞ Pα : (Kθ)⊥ ⊕ Kα → L2.
DUAL TRUNCATED TOEPLITZ OPERATORS
9
Proof. An easy computation shows that
(cid:20)Dθ,α
ϕ
0
0
IKα(cid:21) = E1(cid:20)QαϕQθ ⊞ Pα
0
0
I{0}(cid:21) F1,
where F1 : (Kθ)⊥ ⊕ Kα → ((Kθ)⊥ ⊕ Kα) ⊕ {0} is defined for f− ∈ H 2
−,
f+ ∈ H 2 and gα ∈ Kα by
F1((f− + θ f+) ⊕ gα) = [(f− + θ f+) ⊕ gα] ⊕ 0,
and E1 : L2 ⊕ {0} → (Kα)⊥ ⊕ Kα is defined for f ∈ L2 by
E1(f ⊕ 0) = Qαf ⊕ Pαf.
Clearly, F1 and E1 are invertible.
(cid:3)
Theorem 4.2. Let ϕ ∈ L∞ and let α, θ be inner functions. Then
Dθ,α
ϕ
⋆∼ AP + + BP −,
where AP + + BP − : L2 ⊕ L2 → L2 ⊕ L2 is a paired operator with A, B
given by (4.1).
Proof. Given the result of Proposition 4.1 and the fact that ⋆∼ is an
equivalence relation and thus transitive, we only have to prove that
QαϕQθ ⊞ Pα
⋆∼ AP + + BP −.
To this end, we note that QαϕQθ ⊞ Pα is obviously equivalent after
extension to
(cid:20)QαϕQθ ⊞ Pα
0
0
IL2(cid:21) : ((Kθ)⊥ ⊕ Kα) ⊕ L2 → L2 ⊕ L2.
Using the relations from Section 3 and rewriting them appropriately,
we get (as can be verified independently)
(cid:20)QαϕQθ ⊞ Pα
0
0
IL2(cid:21) = E(cid:20) ϕP − + ϕθP + −P +
¯αϕP − + ϕ ¯αθP + −P −(cid:21) F,
where F : ((Kθ)⊥ ⊕ Kα) ⊕ L2 → L2 ⊕ L2,
F = (cid:20)
(P − + P + ¯θ) ⊞ 0
(P −ϕ ¯α + P +ϕ)Qθ ⊞ (−Pα) −(P − + αP +)(cid:21) ,
0
and E : L2 ⊕ L2 → L2 ⊕ L2,
E = (cid:20)αP − ¯α αP +
P − (cid:21) .
P + ¯α
The operators F and E are invertible by Lemmas 4.3 and 4.4 below. (cid:3)
The proofs of the following two lemmas are straightforward.
10
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
Lemma 4.3. The operator F : ((Kθ)⊥ ⊕ Kα) ⊕ L2 → L2 ⊕ L2,
F = (cid:20)
(P − + P + ¯θ) ⊞ 0
(P +ϕ + P −ϕ ¯α)Qθ ⊞ (−Pα) −(P − + αP +)(cid:21) ,
0
is invertible and F −1 : L2 ⊕ L2 → ((Kθ)⊥ ⊕ Kα) ⊕ L2,
F −1 = (cid:20)(P − + θP +) ⋄ Pαϕ(P − + θP +)
ϕ ¯α(P − + θP +)
0 ⋄ (−Pα)
−(P − + P + ¯α)(cid:21) .
Lemma 4.4. The operator E : L2 ⊕ L2 → L2 ⊕ L2,
E = (cid:20)αP − ¯α αP +
P − (cid:21) ,
P + ¯α
is invertible and E−1 = E.
Corollary 4.5. Let θ be an inner function. Then
Dθ
ϕ
⋆∼ A0P + + B0P −,
where
(4.2)
A0 = (cid:20)ϕθ −1
ϕ
0(cid:21) , B0 = (cid:20) ϕ
¯θϕ −1(cid:21) .
0
In what follows GL∞ will denote the set of all invertible elements of
the algebra L∞.
Corollary 4.6. The operator Dθ,α
ϕ is semi-Fredholm (respectively Fred-
holm) if and only if AP + + BP − is semi-Fredholm (respectively Fred-
holm) on L2 ⊕ L2 and, in that case, we have ϕ ∈ GL∞.
Proof. The equivalence is a consequence of Theorem 2.2 and Theorem
4.2. To prove that ϕ is invertible note that det A = det B = ϕ.
Since a necessary condition for the operator AP + + BP − to be semi-
Fedholm is that
ess inf t∈T det A(t) > 0,
ess inf t∈T det B(t) > 0
([13, Chapter V, Theorem 5.1]), we conclude that if Dθ,α
Fredholm, then ϕ ∈ GL∞.
ϕ
is semi-
(cid:3)
In particular, we conclude that if Dθ
ϕ is invertible, then ϕ ∈ GL∞
([14, Proposition 2.4]) and then, denoting by ϕ(T) the essential range
of ϕ ∈ L∞, we have
(4.3)
ϕ(T) ⊂ σe(Dθ
ϕ) ⊂ σ(Dθ
ϕ)
(see also [15, Theorem 4.1]).
DUAL TRUNCATED TOEPLITZ OPERATORS
11
5. Equivalence relations between Dθ,α
Toeplitz operators
ϕ and truncated
In view of Corollary 4.6, the case where ϕ is an invertible element of
L∞ becomes particularly interesting. Assume then that
ϕ ∈ GL∞.
This implies that, for A, B defined by (4.1), we have A, B ∈ G(L∞)2×2
with
A−1 = (cid:20) 0 ¯θαϕ−1
α (cid:21) , B−1 = (cid:20)ϕ−1
¯α −1(cid:21) .
−1
0
In that case
AP + + BP − = B(P +B−1AP + + P −)(I + P −B−1AP +),
where B (identified with multiplication by B on L2 ⊕ L2) is invertible,
as well as I + P −B−1AP + because
(I + P −B−1AP +)−1 = I − P −B−1AP +.
Therefore AP + + BP − is equivalent to P +CP + + P −, where
C = B−1A = (cid:20)θ −ϕ−1
0 −¯α (cid:21) .
It follows from Theorem 4.2 that Dθ,α
is equivalent after extension to
P +CP + + P −. It is easy to see that P +CP + + P − ∼ P +GP + + P −,
where
ϕ
G = (cid:20)0 1
1 0(cid:21) C(cid:20) 0 1
−1 0(cid:21) = (cid:20) ¯α
ϕ−1 θ(cid:21)
0
(since G and C differ by constant factors). Therefore we have that
Dθ,α
ϕ
⋆∼ P +GP + + P − ⋆∼ P +GP +H 2⊕H 2 = TG
⋆∼ Aα,θ
ϕ−1,
where we took Theorem 2.3 into account. We have thus proved the
following:
Theorem 5.1. Let α, θ be inner functions. If ϕ ∈ GL∞, then Dθ,α
Aα,θ
⋆∼
ϕ
ϕ−1.
Corollary 5.2. If ϕ ∈ GL∞, then Dθ
ϕ
⋆∼ Aθ
ϕ−1.
Corollary 5.3. The operator Dθ
if and only if ϕ ∈ GL∞ and Aθ
ϕ is Fredholm (respectively, invertible)
ϕ−1 is Fredholm (respectively, invertible).
ϕ is Fredholm, then by Corollary 4.6 we have ϕ ∈ GL∞
ϕ−1 is Fredholm. Conversely, if
ϕ is
(cid:3)
Proof. If Dθ
and since Dθ
ϕ
ϕ ∈ GL∞ and Aθ
also Fredholm. The proof for invertibility is analogous.
ϕ−1 is Fredholm, then Corollary 5.2 implies that Dθ
ϕ−1, it follows that Aθ
⋆∼ Aθ
12
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
6. Kernel Isomorphisms
By Theorem 2.2, the kernels of two operators that are equivalent
after extension, are isomorphic. Using the relations from Section 2, we
describe here several of those isomorphisms. We use the same notation
as in Section 4.
Theorem 6.1. The map
N : ker Dθ,α
ϕ → ker(AP + + BP −),
N (f− + θ f+) = (cid:20)
f− + f+
ϕ(1 + ¯α)(f− + θ f+)(cid:21) , where f− ∈ H 2
−, f+ ∈ H 2,
is an isomorphism and
ker Dθ,α
ϕ = (P − + θP +) P1(ker(AP + + BP −)),
where P1(cid:18)(cid:20)x
y(cid:21)(cid:19) = x.
Proof. Since
(1 + ¯α)Pαϕ(f− + θ f+) = ϕ(1 + ¯α)(f− + θ f+) − (1 + ¯α)Dθ,α
it follows from Theorem 3.1 that Dθ,α
ϕ (f− + θ f+),
(AP + + BP −)(cid:20)
ϕ (f− + θ f+) = 0 if and only if
f− + f+
ϕ(1 + ¯α)(f− + θ f+)(cid:21) = (cid:20)0
0(cid:21) .
Thus N is well defined and injective. To see that N is also surjective
note that if
(AP + + BP −)(cid:20)φ1
φ2(cid:21) = (cid:20)0
0(cid:21) ,
then
φ2 = ϕθP +φ1 + ϕP −φ1 + ϕθ ¯αP +φ1 + ¯αϕP −φ1
= ϕ(P −φ1 + θP +φ1) + ϕ ¯α(P −φ1 + θP +φ1)
= ϕ(1 + ¯α)(P −φ1 + θP +φ1),
which means that φ2 is determined by φ1. Thus by Theorem 3.1(2), if
Φ = (cid:20)φ1
φ2(cid:21) is in ker(AP + + BP −), then Φ = N (P −φ1 + θP +φ1).
(cid:3)
Theorem 6.2. The map
N∗ : ker(Dθ,α
N∗(g− + αg+) = (cid:20)g−
ϕ )∗ → ker(AP + + BP −)∗,
g+(cid:21) , where g− ∈ H 2
−, g+ ∈ H 2,
is an isomorphism.
DUAL TRUNCATED TOEPLITZ OPERATORS
13
ϕ )∗ = Dα,θ
¯ϕ and (AP + + BP −)∗ = P +A∗ + P −B∗,
Proof. We have (Dθ,α
where A∗ = ¯AT , B∗ = ¯BT and we identify A∗ and B∗ with the corre-
sponding multiplication operators on L2 ⊕ L2. Now, let Φ = (cid:20)φ1
φ2(cid:21) ∈
L2 ⊕ L2. Then (AP + + BP −)∗Φ = 0 if and only if P +A∗Φ = 0 and
P −B∗Φ = 0. The last two conditions are equivalent to
A∗Φ ∈ (H 2
− ⊕ H 2
−)
and B∗Φ ∈ (H 2 ⊕ H 2),
that is,
(cid:20) ¯ϕ¯θ
−1
¯ϕ¯θα
0 (cid:21)(cid:20)φ1
φ2(cid:21) ∈ (H 2
− ⊕ H 2
−), and (cid:20) ¯ϕ α ¯ϕ
0 −1(cid:21)(cid:20)φ1
φ2(cid:21) ∈ (H 2 ⊕ H 2).
In other words, φ1 ∈ H 2
−, φ2 ∈ H 2 and
(cid:26) ¯θ( ¯ϕφ1 + α ¯ϕφ2) ∈ H 2
¯ϕφ1 + α ¯ϕφ2 ∈ H 2.
−,
The two conditions above can be written as ¯ϕ(φ1 + αφ2) ∈ Kθ. There-
fore, (AP + + BP −)∗Φ = 0 if and only if φ1 ∈ H 2
−, φ2 ∈ H 2 and
Dα,θ
¯ϕ (φ1 + αφ2) = 0. This implies that N∗ is well defined and surjec-
tive. It is now easy to see that N∗ is injective.
(cid:3)
A conjugation on a Hilbert space H is an antilinear isometric invo-
lution (see for instance [10]). In what follows let Cθ : L2 → L2 denote
the conjugation defined as
Cθ(f ) = θ¯z ¯f
for f ∈ L2.
The conjugation Cθ preserves both the model space Kθ and its orthog-
onal complement (Kθ)⊥, (i.e., CθPθ = PθCθ), and therefore induces a
conjugation in Kθ and in (Kθ)⊥, which we also denote by Cθ. This
conjugation plays an important role in the study of truncated Toeplitz
operators.
Theorem 6.3. The map
ND : ker Dθ
ϕ → ker(Dθ
ϕ)∗,
ND(f− + θ f+) = z f+ + θzf−, where f− ∈ H 2
−, f+ ∈ H 2,
is an isomorphism and
ker(Dθ
ϕ)∗ = Cθ(ker Dθ
− and f+ ∈ H 2. Then Dθ
Proof. Let f− ∈ H 2
if ϕ(f− + θ f+) ∈ Kθ, that is, if and only if
ϕ).
ϕ(f− + θ f+) = 0 if and only
θ¯zϕ(f− + θ f+) = ¯ϕ (z f+ + θ zf−) ∈ Kθ.
14
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
The above means that
Dθ
¯ϕ(g− + θg+) = 0
with g− = z f+ ∈ H 2
Thus ND is an isomorphism and we have
−, g+ = zf− ∈ H 2 (i.e., g− + θg+ = ND(f− + θ f+)).
ker(Dθ
ϕ)∗ = θ¯zker Dθ
ϕ = Cθ(ker Dθ
ϕ).
Corollary 6.4. If Dθ
ϕ is Fredholm, then it has Fredholm index 0.
(cid:3)
Corollary 6.5. If Dθ
ker Dθ
ϕ = {0}.
ϕ is Fredholm, then it is invertible if and only if
Theorem 6.6. If ϕ ∈ GL∞, then the map
NDA : ker Dθ,α
ϕ → ker Aα,θ
ϕ−1,
NDA(f− + θ f+) = ϕ(f− + θ f+), where f− ∈ H 2
−, f+ ∈ H 2,
is an isomorphism and
ker Dθ,α
ϕ = ϕ−1 ker Aα,θ
ϕ−1.
if and only if f ∈ (Kθ)⊥ and
Proof. Let f ∈ L2. Then f ∈ ker Dθ,α
ϕf ∈ Kα. In other words, g = ϕf ∈ Kα and ϕ−1g = f ∈ (Kθ)⊥, that
is, g ∈ ker Aα,θ
(cid:3)
ϕ
ϕ−1.
7. Dual truncated Toeplitz operators and the corona
theorem
Corona problems, seen as left invertibility problems, have a strong
connection with the invertibility and Fredholmness of block Toeplitz
operators (see for instance [4] and references in it). In this section we
extend some of those connections to paired operators and apply them
to the study of injectivity and invertibility of dual truncated Toeplitz
operators.
Let CP ± denote the sets of corona pairs, i.e., pairs of functions
satisfying the so called corona conditions in D and C \ (D ∪ T), denoted
here by D±, respectively:
CP + = (cid:26)h+ = (cid:20)h1+
CP − = (cid:26)h− = (cid:20)h1−
h2+(cid:21) ∈ (H ∞ ⊕ H ∞) :
h2−(cid:21) ∈ (H ∞ ⊕ H ∞) :
inf
z∈D+
inf
z∈D−
(h1+(z) + h2+(z)) > 0(cid:27) .
(h1−(z) + h2−(z)) > 0(cid:27) .
Obviously, h− ∈ CP − if and only if h− ∈ CP +.
DUAL TRUNCATED TOEPLITZ OPERATORS
15
By the corona theorem h± ∈ CP ± if and only if there exist h± with
h+ ∈ (H ∞ ⊕ H ∞) and h− ∈ H ∞ ⊕ H ∞ such that
(7.1)
(h±)T h± = 1
in D±.
Theorem 7.1. Let A, B ∈ (L∞)2×2 be such that det A = ϕf+ and
det B = ϕf−, where ϕ ∈ GL∞ and f+, f− ∈ H ∞ \ {0}. If there exist
h± ∈ CP ± satisfying
(7.2)
with Ah+(t) 6= 0 a. e. on T, then the operator AP + + BP − is injective
in L2 ⊕ L2.
Ah+ + Bh− = 0
Proof. Let Φ = (cid:20)φ1
φ2(cid:21) ∈ ker(AP ++BP −) ⊂ L2⊕L2. Taking Φ± = P ±Φ
we have
(7.3)
AΦ+ = −BΦ−.
So, using (7.2), we can write
Taking determinants on both sides, we get
A(cid:2)h+ Φ+(cid:3) = −B(cid:2)h− Φ−(cid:3) .
which, since ϕ ∈ GL∞, is equivalent to
ϕf+ det(cid:2)h+ Φ+(cid:3) = −ϕf− det(cid:2)h− Φ−(cid:3) ,
f+ det(cid:2)h+ Φ+(cid:3) = −f− det(cid:2)h− Φ−(cid:3) .
Moreover, since the left hand side of the above equality represents a
function in H 2 while the right hand side represents a function in H 2
−,
both sides must be zero. It follows that
so there exist non-zero linear combinations, with δ1± and δ2± defined
a. e. on T,
det(cid:2)h± Φ±(cid:3) = 0,
Multiplying on the left by (h±)T , by (7.1) we have
δ1±h± + δ2±Φ± = 0.
δ1± = −δ2±(h±)T Φ±.
Therefore there exist δ± 6= 0 a. e. on T such that
(7.4)
Φ± = δ±h± on T.
From (7.2) and (7.4), AΦ+ = −BΦ− if and only if δ+Ah+ = −δ−Bh−,
or (δ+ − δ−)Ah+ = 0. This however happens if and only if δ+ = δ− (as
Ah+(t) 6= 0 a. e. on T). Since δ+ = (h+)T Φ+ ∈ H 2, δ− = (h−)T Φ− ∈
H 2
(cid:3)
− and δ+ = δ−, we have δ± = 0 and therefore Φ± = 0.
16
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
Theorem 7.2. Let A, B ∈ (L∞)2×2 be invertible matrices such that
det(B−1A) = f−zkf+ with f+, f− ∈ GH ∞, k ∈ Z. If there exist h± ∈
CP ± satisfying
Ah+ + Bh− = 0,
(7.5)
then AP + + BP − is invertible, injective (and not surjective), or sur-
jective (and not injective) if and only if k = 0, k > 0 or k < 0,
respectively.
Proof. If h± satisfy (7.5), then B−1Ah+ = −h−, where det(B−1A) =
f−zkf+ with f+, f− ∈ GH ∞, k ∈ Z and h± ∈ CP ±. Therefore, by
[4, Theorem 4.5] and the proof given there, we have that the Toeplitz
operator TB−1A : (H 2 ⊕ H 2) → (H 2 ⊕ H 2) is invertible, only injective,
or only surjective if and only if k = 0, k > 0 or k < 0, respectively. We
conclude that the same holds for AP + + BP − because (by (2.1)),
TB−1A
⋆∼ B−1AP + + P − ∼ AP + + BP −.
(cid:3)
Taking Theorem 4.2 into account, we also have the following.
Theorem 7.3. Let θ, α be inner functions such that α 6 θ (i.e., α
divides θ).
(1) If ϕ ∈ GL∞, and there exist h1+, h2+, h2− ∈ H ∞ satisfying
(7.6)
(7.7)
and
(cid:20)h1+
h2+(cid:21) ∈ CP +,
(cid:20)¯αh1+
h2− (cid:21) ∈ CP −
ϕ(h2− + θh2+) = h1+,
then ker Dθ,α
ϕ = {0}.
(2) If, moreover, θ ¯α is a finite Blaschke product of degree k, then
is injective; it is invertible if and only if α = cθ with
ϕ
Dθ,α
c ∈ C, c = 1; so in particular Dθ
ϕ is invertible.
Proof. To prove (1) note that for A, B given by (4.1), we have
det A = ϕθ ¯α,
det B = −ϕ,
where θ ¯α ∈ H ∞. For h1− = ¯αh1+ ∈ H ∞, we have from (7.7) that
(cid:26) ϕθh2+ + ϕh2− − h1+ = 0,
¯αϕθh2+ + ¯αϕh2− − h1− = 0,
which can be written as
A(cid:20)h2+
h1+(cid:21) + B(cid:20)h2−
h1−(cid:21) = 0.
DUAL TRUNCATED TOEPLITZ OPERATORS
17
By (7.6), (cid:20)h2+
h1+(cid:21) ∈ CP + and (cid:20)h2−
h1−(cid:21) ∈ CP −, so Theorem 7.1 implies
is also injective.
that AP + + BP − is injective. By Theorems 2.2 and 4.2 the operator
Dθ,α
ϕ
Now we prove (2). If θ ¯α is a finite Blaschke product of degree k,
then we can factorize θ ¯α = R−zkR+, where R± are rational functions
such that R+, ¯R− ∈ GH ∞. Since det(B−1A) = −θ ¯α, the result follows
from Theorems 7.2 and 4.2.
(cid:3)
Dual truncated Toeplitz operators can also be related to corona prob-
lems by using Corollary 5.2 and the known relations between truncated
Toeplitz operators and the corona theorem ([7, 8]), as in the following
theorem which will be used in the next section to describe the spectrum
of a class of dual truncated Toeplitz operators with analytic symbols.
Theorem 7.4. Let θ be an inner function. If ϕ ∈ GL∞ and there exist
h+ ∈ H ∞ ⊕ H ∞, h− ∈ CP + such that Gh+ = h− with
G = (cid:20) ¯θ
ϕ−1 θ(cid:21) ,
0
then Dθ
ϕ is invertible if and only if h+ ∈ CP +.
Proof. In this case Dθ
ϕ−1 is invertible, and
this is equivalent to the Toeplitz operator TG being invertible. The
result now follows from Theorem 3.11 in [8].
(cid:3)
ϕ is invertible if and only if Aθ
8. Fredholmness, inwertibility and spectra of dual
truncated Toeplitz operators
In this section we apply the previous results to study several classes
of dual truncated Toeplitz operators. In what follows θ is always an
inner function.
8.1. Analytic symbols. We start by using the results of Theorem
7.4 to describe the spectrum of dual truncated Toeplitz operators with
analytic symbols of a particular type.
Theorem 8.1. If ϕ ∈ H ∞ and ¯θϕ ∈ H ∞, then σ(Dθ
ϕ) = clos ϕ(D).
Proof. If λ ∈ ϕ(T), then ϕ − λ /∈ GL∞, so Dθ
λ /∈ ϕ(T), then ϕ − λ ∈ GL∞ and we have
ϕ−λ is not Fredholm. If
(cid:20) ¯θ
1
ϕ−λ
0
θ(cid:21)(cid:20)ϕ − λ
0 (cid:21) = (cid:20)¯θ(ϕ − λ)
1
(cid:21) .
18
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
Since the right hand side of the above equality belongs to CP −, by
Theorem 7.4 we have that Dθ
ϕ−λ in invertible if and only if (cid:20)ϕ − λ
0 (cid:21) ∈
CP +. The latter is equivalent to (ϕ − λ) ∈ GH ∞, i.e.,
ϕ(z) − λ > 0.
inf
z∈D
Since ϕ(T) ⊂ clos ϕ(D), we conclude that σ(Dθ
ϕ) = clos ϕ(D).
(cid:3)
The assumptions of Theorem 8.1, are satisfied in particular if ϕ ∈
Kzθ ∩ L∞ since, in that case, ϕ ∈ H ∞ and ¯θϕ ∈ H 2 ∩ L∞ = H ∞ (see
[15, Theorem 4.3]).
Several other results regarding analytic symbols will be obtained
from the properties studied below.
8.2. Symbols with analytic or co-analytic inverse. We consider
now symbols ϕ ∈ GL∞ such that ϕ−1 is in H ∞ or in H ∞.
In that
case, by Corollary 5.2, Dθ
ϕ is equivalent after extension to a truncated
Toeplitz operator Aθ
ϕ−1; therefore we start by recalling the following.
Theorem 8.2 ([7, 8]). Let g+ ∈ H ∞ and denote by gi
of g+. Then:
+ the inner factor
(1) Aθ
g+ is Fredholm if and only if γ = GCD(θ, gi
Blaschke product and ¯γ(cid:20) θ
g+ is invertible if and only if (cid:20) θ
g+(cid:21) ∈ CP +,
g+(cid:21) ∈ CP +,
(2) Aθ
+) is a finite
(3) ker Aθ
g+ = θ
γ Kγ.
Using Theorem 2.2 and Corollary 5.2 we now obtain the following.
Theorem 8.3. Let ϕ ∈ GL∞. Assume that ϕ−1 ∈ H ∞ and let ϕ−1 =
βa+ be its standard inner -- outer decomposition (β -- inner, a+ -- outer).
Denote γ = GCD(θ, β). Then
(1) Dθ
ϕ is Fredholm if and only if γ is a finite Blaschke product,
(2) Dθ
ϕ is invertible if and only if (cid:20)θ
β(cid:21) ∈ CP +,
(3) ker Dθ
ϕ = βa+
θ
γ Kγ ⊂ θβ
γ H 2.
Proof. Parts (1) and (2) follow from Theorem 8.2, Theorem 2.2 and
Corollary 5.2. Note that the condition (cid:20)θ
β(cid:21) ∈ CP + in (2) is equivalent
DUAL TRUNCATED TOEPLITZ OPERATORS
19
to (cid:20) θ
ϕ−1(cid:21) ∈ CP + in this case. On the other hand, by Theorem 6.6,
ker Dθ
ϕ = ϕ−1 ker Aθ
ϕ−1 = βa+
θ
γ Kγ.
Corollary 8.4. If ϕ ∈ GH ∞, then Dθ
ϕ is invertible.
Proof. In this case β ∈ C with β = 1, therefore (cid:20)θ
β(cid:21) ∈ CP +.
(cid:3)
(cid:3)
Corollary 8.5. If θ is a singular inner function, then Dθ
if and only if it is invertible.
ϕ is Fredholm
Proof. Using the same notation as in Theorem 8.3 we have that γ =
GCD(θ, β) is either a constant or a singular inner function. So either
ker Dθ
ϕ is equivalent to
its invertibility by Corollary 6.5, or ker Dθ
ϕ is infinite dimensional and
in this case Dθ
(cid:3)
ϕ = {0} and in that case the Fredholmness of Dθ
ϕ is neither Fredholm nor invertible.
¯β = β θ
Corollary 8.6. Let β be an inner function and γ = GCD(θ, β). Then
ker Dθ
¯β is Fredholm if and only if γ is a
θ(cid:21) ∈ CP +.
finite Blaschke product, and it is invertible if and only if (cid:20)β
γ Kγ. The operator Dθ
It follows from Corollary 8.6 that we have
ker Dθ
¯θ = θKθ
and θ ∈ ker Dθ
¯θ if and only if θ(0) = 0 (cf. [14, Example 2.5]).
If ϕ ∈ GL∞ and ϕ−1 ∈ H ∞, then we can reduce the study of Fred-
¯ϕ and use Theo-
ϕ)∗ = Dθ
holmness and invertibility of Dθ
rem 8.3. Regarding the kernel, we have the following:
ϕ to that of (Dθ
Corollary 8.7. Let ϕ ∈ GL∞. Assume that ϕ−1 ∈ H ∞ and let ϕ−1 =
βa+ be its standard inner -- outer decomposition (β -- inner, a+ -- outer).
Denote γ = GCD(θ, β). Then
ker Dθ
ϕ = Cθ(ker(Dθ
ϕ)∗) = (cid:18)β
γ(cid:19)a+ ¯zKγ ⊂ (cid:18) β
γ(cid:19)H 2
−.
8.3. Rational symbols. Let now ϕ be continuous on the unit circle
T, ϕ ∈ C(T). In this case we can describe the essential spectrum of
Dθ
ϕ (see also [15, Theorem 4.3]).
For an inner function θ,
Σ(θ) = {w ∈ clos D :
lim inf
z→w
θ(z) = 0}.
20
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
Theorem 8.8. If ϕ ∈ C(T), then σe(Dθ
ϕ) = ϕ(T).
Proof. By (4.3), we have ϕ(T) ⊂ σe(Dθ
then ϕ − λ ∈ GL∞ and Dθ
By Theorem 5.3 in [8] we have that this happens if and only if 0 /∈
g(Σ(θ) ∩ T) with g = 1
(cid:3)
ϕ). Conversely, if λ ∈ C \ ϕ(T),
is Fredholm.
ϕ is Fredholm if and only if Aθ
ϕ−λ, which is always true.
ϕ−λ
1
Corollary 8.9. If ϕ ∈ C(T), then Dθ
invertible in C(T).
ϕ is Fredholm if and only if ϕ is
We can obtain further results if ϕ is rational and continuous on T,
i.e., ϕ ∈ R. We start by describing the kernels of dual truncated
Toeplitz operators with rational symbols.
Theorem 8.10. Let R ∈ R and R = P/Q, where P and Q are poly-
nomials without common zeros. Then f ∈ ker Dθ
R if and only if there
is a decomposition f = f− + θ f+, f− ∈ H 2
−, f+ ∈ H 2 such that there
are polynomials P1, P2 with deg(P2) 6 max{deg(P ), deg(Q)} − 1 such
that
(8.1)
f− = P1
P ,
f+ = P2
P ,
Proof. Let f = f− + θ f+ with f− ∈ H 2
R if
and only if R(f− + θ f+) ∈ Kθ. The latter happens if and only if there
exist k+ ∈ H 2 and k− ∈ H 2
P1
Q ∈ Kθ.
Q + θ P2
−, f+ ∈ H 2. Then f ∈ ker Dθ
− such that
Qf− + P
Q θ f+ = k+,
¯θk+ = k−,
(cid:26) P
or equivalently,
(8.2)
(cid:26) P f− = k+Q − P θ f+,
P f+ = k−Q − ¯θP f−.
By a generalization of Liouville's theorem, both sides of the first
equation in (8.2) must be equal to a polynomial P1 such that P1
P ∈ H 2
−
and, analogously, both sides of the second equation in (8.2) must be
P ∈ H 2 and P2 = Qk− − P ¯θf−.
equal to a polynomial P2 such that P2
So the degree of P2 is appropriate and
f+ = P1
k+ = P1
Q + θ P2
Q ∈ Kθ.
Q + θ P
Q
Conversely, if P1 and P2 are polynomials satisfying desired condi-
−, f+ ∈ H 2
P we have that f− ∈ H 2
tions, then for f− = P1
and
P , f+ = P2
Q (cid:0) P1
so that f = f− + θ f+ ∈ ker Dθ
R.
R(f− + θ f+) = P
P + θ P2
P (cid:1) = P1
Q + θ P2
Q ∈ Kθ,
(cid:3)
DUAL TRUNCATED TOEPLITZ OPERATORS
21
The previous theorem enables us to characterize the points λ ∈
σ(Dθ
R) for R ∈ R, as follows.
Theorem 8.11. If R ∈ R, then
σ(Dθ
R) = R(T) ∪ σp(Dθ
R) = σe(Dθ
R) ∪ σp(Dθ
R).
Proof. From Theorem 8.8 we have σe(Dθ
λ /∈ R(T), then Dθ
it is injective by Corollary 6.5. Therefore λ ∈ σ(Dθ
ker Dθ
If
R−λ is Fredholm, so it is invertible if and only if
R) if and only if
(cid:3)
ϕ−λ 6= {0}, i.e., λ is an eigenvalue of Dθ
ϕ.
R) = R(T) ⊂ σ(Dθ
R).
As an application we study the spectra of the dual truncated shift
Dθ
z and its adjoint -- the dual truncated backward shift Dθ
¯z. We start by
studying their kernels. The next result is a consequence of Theorems
8.10 and 8.11.
Theorem 8.12. If λ ∈ T ∪ D− or if λ ∈ D and θ(0) 6= 0 then
ker Dθ
¯z−¯λ = {0}. If λ ∈ D, θ(0) = 0, then
z−λ = ker Dθ
and
ker Dθ
ker Dθ
−
z−λ = span(cid:8) 1
¯z−¯λ = span(cid:8) θ
z−λ(cid:9) ⊂ H 2
1−λz(cid:9) ⊂ θH 2.
Proof. By Theorem 8.10, f ∈ ker Dθ
z−λ if and only if there are constants
A, B ∈ C such that f = f− + θ f+ with f− = A
z−λ. Note
that if λ ∈ T∪ D−, then A = 0, and if λ ∈ T∪ D, then B = 0. Therefore
z−λ and f+ = B
(1) if λ ∈ T, then ker Dθ
(2) if λ ∈ D−, then ker Dθ
(3) if λ ∈ D, then ker Dθ
z−λ = {0};
z−λ = {θ B
z−λ = { A
z−λ : B ∈ C, θB ∈ Kθ};
z−λ : A ∈ C ∩ Kθ} = {0}.
So ker Dθ
have ker Dθ
z−λ = {0} unless λ ∈ D and θ(0) = 0. By Theorem 6.3, we
(cid:3)
¯z−¯λ = Cθ(ker Dθ
z−λ).
Corollary 8.13.
(1) σe(Dθ
z) = σe(Dθ
¯z) = T;
(2) σ(Dθ
σp(Dθ
θ(0) = 0.
z) and σ(Dθ
¯z), where σp(Dθ
¯z) are the disjoint union of T with σp(Dθ
z) = ∅ if θ(0) 6= 0 and σp(Dθ
z) =
z) = D if
References
[1] H. Bart, V. `E. Tsekanovskiı, Matricial coupling and equivalence after extension,
Operator theory and complex analysis (Sapporo, 1991), 143 -- 160, Oper. Theory
Adv. Appl. 59, Birkhauser, Basel, 1992.
[2] H. Bercovici, D. Timotin, Truncated Toeplitz operators and complex symme-
tries, Proc. Amer. Math. Soc. 146 (2018) 261 -- 266.
22
M. C. C AMARA, K. KLI´S -- GARLICKA, B. LANUCHA, AND M. PTAK
[3] A. Bottcher, B. Silberman., Analysis of Toeplitz Operators, 2nd edition,
Springer-Verlag, Berlin, 2006.
[4] M. C. Camara, C. Diogo, L. Rodman, Fredholmness of Toeplitz operators and
corona problems, J. Funct. Anal. 259 (2010), no. 5, 1273 -- 1299.
[5] C. Camara, J. Jurasik, K. Kli´s-Garlicka, M. Ptak, Characterizations of asym-
metric truncated Toeplitz operators, Banach J. Math. Anal. 11 (2017), no. 4,
899 -- 922.
[6] C. Camara, K. Kli´s-Garlicka and M. Ptak, Asymmetric truncated
Toeplitz operators and conjugations, Filomat, to appear.
[7] M. C. Camara, J. R. Partington, Asymmetric truncated Toeplitz operators and
Toeplitz operators with matrix symbol, J. Operator Theory 77 (2017), no. 2,
455 -- 479.
[8] M. C. Camara, J. R. Partington, Spectral properties of truncated Toeplitz op-
erators by equivalence after extension, J. Math. Anal. Appl. 433 (2016), no. 2,
762 -- 784.
[9] S. R. Garcia, J. Mashreghi, W.T. Ross, Introduction to model spaces and their
operators, Cambridge University Press, 2016.
[10] S. R. Garcia, M. Putinar, Complex symmetric operators and applications,
Trans. Amer. Math.Soc., 358 (2006), 1285 -- 1315.
[11] J. B. Garnett, Bounded analytic functions, Academic Press, New York, 1981.
[12] Y. Hu, J. Deng, T. Yu, L. Liu, Y. Lu, Reducing Subspaces of the Dual Truncated
Toeplitz Operator, J. Funct. Spaces (2018), Art. ID 7058401, 9 pp.
[13] S. G. Mikhlin, S. Prossdorf, Singular integral operators, Translated from the
German by Albrecht Bottcher and Reinhard Lehmann. Springer-Verlag, Berlin,
1986.
[14] X. Ding, Y. Sang, Dual truncated Toeplitz operators, J. Math. Anal. Appl. 461
(2018), no. 1, 929 -- 946.
[15] Y. Sang, Y. Qin, X. Ding, Dual truncated Toeplitz C ∗-algebras, Banach J.
Math. Anal. 13 (2019), no. 2, 275 -- 292.
[16] D. Sarason Algebraic properties of truncated Toeplitz operators, Oper. Matrices
1 (2007), 491 -- 526.
[17] D. Sarason, Doubly shift-invariant spaces in H 2, J. Operator Theory 16 (1986),
75 -- 97.
[18] F.-O. Speck, General Wiener-Hopf factorization methods. With a foreword by
E. Meister. Research Notes in Mathematics, 119. Pitman (Advanced Publish-
ing Program), Boston, MA, 1985.
[19] F.-O. Speck, Wiener-Hopf factorization through an intermediate space, Integral
Equations Operator Theory 82 (2015), no. 3, 395 -- 415.
[20] K. Stroethoff and D. Zheng, Dual Toeplitz operators, Trans. AMS 354 (2002),
no. 6, 2495 -- 2520.
DUAL TRUNCATED TOEPLITZ OPERATORS
23
M. Cristina Camara Center for Mathematical Analysis, Geometry
and Dynamical Systems, Mathematics Department, Instituto Superior
T´ecnico Universidade de Lisboa, Av. Rovisco Pais, 1049-001 Lisboa,
Portugal
E-mail address: [email protected]
Kamila Kli´s-Garlicka, Department of Applied Mathematics, Univer-
sity of Agriculture, ul. Balicka 253C, 30-198 Krak´ow, Poland
E-mail address: [email protected]
Bartosz Lanucha, Institute of Mathematics, Maria Curie-Sk lodowska
University, pl. M. Curie-Sk lodowskiej 1, 20-031 Lublin, Poland
E-mail address: [email protected]
Marek Ptak, Department of Applied Mathematics, University of
Agriculture, ul. Balicka 253C, 30-198 Krak´ow, Poland
E-mail address: [email protected]
|
1109.3137 | 1 | 1109 | 2011-09-14T17:06:01 | After the Explosion: Dirichlet Forms and Boundary Problems for Infinite Graphs | [
"math.FA",
"math.PR"
] | Formal Laplace operators are analyzed for a large class of resistance networks with vertex weights. The graphs are completed with respect to the minimal resistance path metric. Compactness and a novel connectivity hypothesis for the completed graphs play an essential role. A version of the Dirichlet problem is solved. Self adjoint Laplace operators and the probability semigroups they generate are constructed using reflecting and absorbing conditions on subsets of the graph boundary. | math.FA | math |
After the Explosion:
Dirichlet Forms and
Boundary Problems for
Infinite Graphs
Robert Carlson
Department of Mathematics
University of Colorado at Colorado Springs
[email protected]
June 8, 2018
Abstract
Formal Laplace operators are analyzed for a large class of resis-
tance networks with vertex weights. The graphs are completed with
respect to the minimal resistance path metric. Compactness and a
novel connectivity hypothesis for the completed graphs play an essen-
tial role. A version of the Dirichlet problem is solved. Self adjoint
Laplace operators and the probability semigroups they generate are
constructed using reflecting and absorbing conditions on subsets of
the graph boundary.
Mathematics Subject Classification. Primary 34B45
Keywords. boundary value problems on networks, resistance networks,
Dirichlet forms, Markov chain explosions.
1
1
Introduction
This work has its roots in the challenge of extending differential equation
models for diffusion or wave propagation from domains in Euclidean space
to infinite graphs intended to resemble biological transport systems such as
the arteries of the human circulatory system. Such biological systems can in-
clude enormous numbers of branching segments. Short time transport across
the network is essential, so treelike structures with small numbers of large
edges and vast collections of microscopic edges are typical. Faced with such
complex heterogeneous structures, one hopes that appropriate infinite graph
models will suggest useful structural features and robustly posed problems.
Building on an earlier 'quantum graph' analysis of such problems [1], this
work uses infinite graph and operator theoretic methods to treat a class of
continuous time Markov chains. Recall that continuous time Markov chains
use a system of constant coefficient differential equations
dP
dt
= QP, P (0) = I.
(1.1)
to describe the evolution of probability densities X(t) = X(0)P (t) on a finite
or countably infinite set of states. An associated graph may be constructed
by connecting states (vertices) i and j with an edge if Qij 6= 0.
In the finite state case the solution of (1.1) is simply P (t) = eQt. When
the set of states is infinite the formal description of the operator Q may not be
adequate to determine the semigroup eQt, an issue known in probability as the
problem of explosions. Infinite graph models inspired by biological transport
systems will typically face the explosion problem. By imposing restrictions
on both the form of the Markov chain generator Q and the structure of the
associated graph viewed as a metric space, this work provides a resolution in
terms of 'reflecting' and 'absorbing' behavior at a graph boundary.
It will be advantageous to use the Dirichlet form theory [4, 9]. To that end,
consider a graph G whose edges e = [u, v] are equipped with positive weights
R(u, v) which are interpreted as edge length. With C(u, v) = 1/R(u, v), a
symmetric bilinear form for functions on the vertex set is defined by
B(f, g) =
1
2 X
v∈V
X
u∼v
C(u, v)(f (v) − f (u))(g(v) − g(u)).
(1.2)
Each vertex is also given a positive weight µ(v). Formal semigroup generators
2
∆µ are defined by
∆µf (v) =
1
µ(v) X
u∼v
C(u, v)(f (v) − f (u)).
(1.3)
R(u, v) is often interpreted as electrical resistance. The electrical network
analogy is treated at length in [6, 18]. The recent work [10] treats electrical
currents in a context similar to this paper, while [11] treats related topological
questions. An analysis of function theory on infinite trees motivated by
modeling the human lungs is in [19]. With the domain of functions with finite
support, ∆µ is a symmetric operator on l2(µ). In contrast to this paper, other
recent work [3, 13, 15, 16] has stressed cases when this symmetric operator
is essentially selfadjoint, and so behavior at the graph boundary is not an
issue.
The vertex set V of an edge weighted locally finite graph G can be
equipped with a metric d(u, v) obtained by minimizing the sum of the edge
lengths of paths from u to v. By completing this metric space we obtain a
metric space G in which one can discuss features like the graph boundary and
compactness. If G is a tree, then distinct points of G can be separated by
deleting a suitable edge. Generalizing this idea, our graphs will be required
to have 'weakly connected' completions, with the property that, for any two
distinct points, any path joining them must include an edge from a finite set.
This generalization identifies a rich class of edge weighted graphs with useful
topological and function theoretic properties.
The properties of weakly connected graph completions are developed in
the second section.
In addition to trees, arbitrary graphs with finite vol-
ume have weakly connected completions. This class is also preserved if we
add suitably constrained edge sequences to a graph. Weakly connected com-
pletions are totally disconnected metric spaces. When also compact, these
spaces are topologically stable with respect to decrease of the metric. The
weakly connected class will be characterized using the separation of points
property for an algebra of 'eventually flat' functions.
The third section treats the bilinear forms, vertex weights, and associ-
ated operators. The choice of vertex weights typically used for discrete time
Markov chains are contrasted with weights making ∆µ resemble a discretized
second derivative. The bilinear form is used to construct several 'Sobolev
style' Hilbert spaces on G whose elements extend continuously to G.
Two main problems are treated in the fourth section. The first, a version
of the Dirichlet problem, asks for conditions under which continuous func-
3
tions on ∂G have a unique harmonic extension to G. An example shows that
a lack of compactness can lead to a negative result. Using assumptions of
compactness and weak connectivity, a general positive result is established.
The second problem is the resolution of the explosion problem in terms of
reflecting and absorbing boundary conditions. The semigroups generated by
the operators defined using these boundary conditions are positivity preserv-
ing contractions on l1(µ).
Despite the connections with probability, this work will not explicitly use
probabilistic techniques or interpretations. We simply mention the classic
work [7],and the recent works [9, 17, 21] as pointers to the enormous literature
related to analysis of infinite state Markov chains.
2 Weakly connected graphs
2.1 Topology
G will denote a simple graph with a countable vertex set V and a countable
edge set E. Each vertex will have at least one and at most finitely many
incident edges. Vertices of degree 1 are boundary vertices; the rest are interior
vertices. G is assumed to have edge weights (resistances). That is, there is
a function R : E → (0, ∞), denoted by R(u, v) when [u, v] ∈ E. General
references on graphs are [2, 5]. Edge weights, considered the length of the
edges, are commonly identified with electrical network resistance [6] or [18],
and then edge conductance is the reciprocal C(u, v) = 1/R(u, v) if R(u, v) >
0, and 0 otherwise.
A finite path (sometimes called a walk) in G connecting vertices u and v
is a finite sequence of vertices u = v0, v1, . . . , vK = v such that [vk, vk+1] ∈ E
for k = 0, . . . , K − 1. G is connected there is a finite path from u to v for all
u, v ∈ V. Define a metric on (the vertices of) G by
d(u, v) = inf
γ X
k
R(vk+1, vk),
(2.1)
the infimum taken over all finite paths γ joining u and v. If there is no finite
path from u to v then d(u, v) = ∞. G, with the extended metric d, will
denote the metric space completion [20, p. 147] of G.
Extending the combinatorial notion of path, a path in G will be a sequence
{vk} with vk ∈ V, [vk, vk+1] ∈ E, where the index set may be finite (finite
4
path), the positive integers (a ray), or the integers (a double ray). The role of
continuous paths in G is played by paths going from u ∈ G to v ∈ G, which in
the double ray case requires limk→−∞ d(vk, u) = 0 and limk→∞ d(vk, v) = 0.
The ray case is similar. A path for which all vertices are distinct is a simple
path. If G is connected then there is a path joining any pair of points u, v ∈ G.
Modifying ideas from [1], say that G is weakly connected if for every pair of
distinct points u, v ∈ G there is a finite set W of edges in G such that every
path from u to v contains an edge from W .
One may extend G to a metric graph Gm by identifying the combinatorial
edge [u, v] with an interval of length R(u, v). With the usual metric on Gm,
its vertex set will be isomorphic to G. By this device some of the results
of [1], which should be consulted for more details, carry over to the present
context. The next result is a simple example.
Proposition 2.1. If T is a tree then T is weakly connected.
The volume of a graph is defined as the sum of its edge lengths,
vol(G) = X
[v1,v2]∈E
R(v1, v2).
Finite volume graphs also have weakly connected completions [1].
Proposition 2.2. If vol(G) < ∞ then G is weakly connected.
Proof. The main case considers distinct points x and y in G \ G. Remove a
finite set of edges from E so that the remaining edgeset E1 satisfies
X
[v1,v2]∈E1
R(v1, v2) <
d(x, y)
2
.
Proceeding with a proof by contradiction, suppose (. . . , v−1, v0, v1, . . . ) is
a path from x to y using only edges in E1. Then for n sufficiently large
d(v−n, vn) > d(x,y)
, but there is a simple path from v−n to vn using only edges
from E1, so d(v−n, vn) < d(x,y)
.
2
2
The next result gives conditions allowing edges to be added to a weakly
connected graph without disturbing that property.
5
Theorem 2.3. Suppose the graph G0, with vertex set V, has a weakly con-
nected completion G0. Using the same vertex set V, enlarge G0 to a graph G1
by adding a sequence E1 of edges en whose lengths Rn satisfy limn→∞ Rn = 0.
Assume there is a positive constant C such that
dG1(u, v) ≤ dG0(u, v) ≤ CdG1(u, v),
u, v ∈ V.
Then G1 is weakly connected.
Proof. First note that G1 \ G1 = G0 \ G0, since the set of Cauchy sequences of
vertices has not changed. It will suffice to consider distinct points u and v in
G1 \ G1 which are joined by a path in G1. Let W0 be a finite set of edges in
G0 such that every path in G0 from u to v contains an edge [v1, v2] from W0.
Pick ǫ > 0 such that ǫ < R(v1, v2) for all edges [v1, v2] ∈ W0. Find N so
that the lengths Rn of edges en ∈ E1 satisfy Rn < ǫ/C for n > N. Let W1
be the set of edges W0 ∪ {e1, . . . , eN } in G1. Suppose there is a path γ in G1
joining u to v, but not containing any edge from W1.
Let Γ be the set of edges ej ∈ E1 which are in γ. Γ is not empty since
γ contains at least one edge not in G0. Since γ contains no edge from W1,
the edges ej ∈ Γ have length Rj < ǫ/C.
If an edge ej = [vj, vj+1] ∈ Γ
has dG0(vj, vj+1) < ǫ, then ej can be replaced by a finite path γj in G0 with
length at most CRj, and containing no edge from W0. Thus there is at
least one ej = [vj, vj+1] ∈ Γ with dG0(vj, vj+1) ≥ ǫ. But the inequalities
dG1(vj, vj+1) < ǫ/C while dG0(vj, vj+1) ≥ ǫ contradict the hypotheses, so no
such path from u to v exists, and G 1 is weakly connected.
The conclusion of Theorem 2.3 may be false if the edge lengths Rn have
a positive lower bound. Start with G0 which is be weakly connected and
connected. Take distinct points x, y ∈ G0 \ G0. Suppose γ is a path from x
to y. Take sequences of vertices xn, yn from γ with xn → x and yn → y such
that xn 6= yn, and [xn, yn] is not an edge in G0, Form G1 by adding edges
[xn, yn] to G0, with R(xn, yn) = dG0(xn, yn). Since dG1(u, v) = dG0(u, v) for all
vertices u, v, the vertex sets for G0 and G1 are isometric. The edge lengths
R(xn, yn) have a positive lower bound, and G1 is not weakly connected,
Theorem 2.4. Assume G is weakly connected.
If U and V are disjoint
compact subsets of G, then there is a finite set W of edges in G such that
every path from U to V contains an edge in W .
6
Proof. Since G is weakly connected, if u ∈ U and v ∈ V there is a finite set
W (u, v) of edges in G such that every path from u to v contains an edge in W .
Take ǫ > 0 such that ǫ < R(v1, v2) for all edges [v1, v2] ∈ W . If z1 ∈ Bǫ(u),
the open ǫ ball centered at u, and z2 ∈ Bǫ(v), then every path from z1 to z2
contains an edge in W (u, v). The collection {Bǫ(u) × Bǫ(v), u ∈ U, v ∈ V }
is an open cover of the compact set U × V , so there is a finite subcover
Bǫn(un) × Bǫn(vn) for n = 1, . . . , N. Take W = ∪nW (un, vn).
Suppose W is a nonempty finite set of edges in E. For x ∈ G, let UW (x)
be the set of points y ∈ G which can be connected to x by a path containing
no edge of W .
Lemma 2.5. For all x ∈ G the set UW (x) is both open and closed in G.
Proof. Take ǫ > 0 such that ǫ < R(v1, v2) for all edges [v1, v2] ∈ W . Suppose
y and z are vertices with d(z, y) < ǫ, so there is a path of length smaller
than ǫ from y to z. If there is a path from y to x containing no edge from
W , then by concatenating these paths there is a path from z to x containing
no edge from W . This shows that UW (x) and the complement of UW (x) are
both open.
Theorem 2.6. A weakly connected G is totally disconnected.
Proof. Suppose v1 and v2 are distinct points in G, with W being a finite set of
edges in G such that every path from v1 to v2 contains an edge from W . The
set UW (v1) is both open and closed. Since UW (v1) and UW (v2) are disjoint,
UW (v2) ⊂ U c
W (v1), the complement of UW (v1) in G. Thus v1 and v2 lie in
different connected components.
If G is totally disconnected and compact, it has a rich collection of clopen
sets, that is sets which are both open and closed. In fact [1] or [12, p. 97] for
any x ∈ G and any ǫ > 0 there is a clopen set U such that x ∈ U ⊂ Bǫ(x).
In particular any compact subset of G can then be approximated by a clopen
set.
Changing the metric on G may change the completion G and the functions
on G that extend continuously to graphbar. Given G and two weight functions
R0 and R1, let G0 and G 1 denote the completions of G with respect to the
associated metrics d0 and d1. Say that the weight function R1(u, v) is smaller
than the weight function R0(u, v) if R1(u, v) ≤ R0(u, v) for every edge [u, v] ∈
E. If R1 is smaller than R0, the metric d1 on G extends to a pseudometric
7
[20, p. 140-141] on G 0; that is, there may be distinct points x, y ∈ G0 with
d1(x, y) = 0. The next result establishes a stability property for weakly
connected graph completions
Theorem 2.7. Suppose G0 is weakly connected. If R1 is smaller than R0 then
the pseudometric on G 0 induced by R1 is a metric. If in addition (G0, d0) is
compact, then (G 0, d0) and (G0, d1) are homeomorphic.
Proof. The usual construction [20, p. 147] identifies the completion of a
metric space with equivalence classes of Cauchy sequences, two sequences
{xn} and {yn} being equivalent when d(xn, yn) → 0. Since G is locally finite,
it is only possible to have distinct points x, y ∈ G0 with d1(x, y) = 0 if
x, y ∈ G0 \ G.
Suppose x and y are distinct points of G0 \ G, and that W is a finite set of
edges such that any path from x to y contains an edge from W . Let {xn} and
{yn} be sequences in V with xn → x and yn → y. By Lemma 2.5 there is an
N such that n ≥ N implies any path in G from xn to yn must contain an edge
from W . Let 0 < ǫ < R1(u, v) for all edges [u, v] ∈ W . Then d1(xn, yn) ≥ ǫ
for all n ≥ N, so the sequences {xn} and {yn} are still inequivalent Cauchy
sequences with respect to d1, showing that d1 is a metric on G0.
Let ι denote the identity map from (G0, d0) to (G0, d1), which is distance
reducing, so continuous. If (G0, d0) is compact, then so is (G0, d1). Following
[8, p.123], suppose K ⊂ (G0, d0) is closed. Then (ι−1)−1(K) = ι(K) is
compact, so closed, and ι−1 is continuous.
2.2 Function theory
There is an algebra of functions with pointwise addition and multiplication
well matched to the weakly connected graph completions. Define the 'even-
tually flat' functions A to be the real algebra of functions φ : V → R such
that the set of edges [u, v] in E with φ(u) 6= φ(v) is finite.
Lemma 2.8. All functions φ ∈ A extend continuously to G.
Proof. Let W be the finite set of edges [u, v] with φ(u) 6= φ(v). Suppose
x ∈ G, xn ∈ G, and xn → x. If x ∈ G then xn = x for n sufficiently large.
Suppose x ∈ G \ G. By Lemma 2.5 the set U(W )(x) is open, so there is
an N such that n ≥ N implies xn ∈ UW (x) and φ(xn) = φ(xn+1). Take
φ(x) = limn→∞ φ(xn).
8
Making use of this lemma, functions φ ∈ A are extended continuously to
functions on G.
Lemma 2.9. If G is connected, then any φ ∈ A has finite range. For c ∈ R,
φ−1(c) is a clopen set in G.
Proof. Suppose φ ∈ A is not constant, and suppose u ∈ V. Find a path
(u = v0, v1, . . . , vN ) such that φ(vn) = φ(u) for n ≤ N and φ(vN ) 6= φ(w) for
some w adjacent to vN . Since the set of such vertices vN is finite, φ(u) has
one of a finite set of values.
For c ∈ R, φ−1(c) is a closed set, and its complement in G is the union of
a finite collection of closed sets.
The next result shows that A separates points of G if and only if G is
weakly connected.
Theorem 2.10. Suppose G is weakly connected. If x and y are distinct points
of G, there is a function φ ∈ A whose range is {0, 1} such that φ(z) = 0 for z
in an open neighborhood U of x, and φ(z) = 1 for z in an open neighborhood
V of y.
Conversely, if A separates points of G, then G is weakly connected.
Proof. Let W be a finite set of edges in G such that every path from x to y
contains an edge from W . By Lemma 2.5 the set UW (x) is both open and
closed in G, as is U c
W (x) Define φ(z) = 0 for z ∈ UW (x) and φ(z) = 1 for
z ∈ U c
W (x). For every vertex v and adjacent vertex w we have φ(v) = φ(w)
unless [v, w] ∈ W . Since W is finite, φ ∈ A.
In the other direction, suppose A separates points of G. Let x and y be
distinct points in G, and suppose φ ∈ A with φ(x) < φ(y). Let W be the
finite set of edges [u, v] such that φ(u) 6= φ(v). If γ is any path starting at
x which contains no edge from W , then φ must be constant along γ. That
is, every path from x to y must contain an edge from W , so G is weakly
connected.
Corollary 2.11. Suppose G is weakly connected. If Ω and Ω1 are nonempty
disjoint compact subsets of G, then there is a function f ∈ A such that
0 ≤ f ≤ 1,
f (x) = 1,
x ∈ Ω,
f (y) = 0,
y ∈ Ω1.
9
Proof. First fix y ∈ Ω1. Using Theorem 2.10, find a finite cover U1, . . . , UN of
Ω by open sets with corresponding functions f1, . . . , fN which satisfy fn(z) =
1 for all z in some open neighborhood Vy of y, while fn(z) = 0 for all z in
Un.
Define Fy = f1 · · · fN . Find a finite collection F1, . . . , FN whose corre-
sponding open sets V1, . . . , VN cover Ω1. The function f = (1−F1) · · · (1−FN )
has the desired properties.
The combination of Theorem 2.10 and the Stone-Weierstrass Theorem
yields the next result.
Theorem 2.12. If G is weakly connected and compact, then A is uniformly
dense in the continuous functions on G.
3 Quadratic forms
3.1 Weights and forms
Vertex weights and the corresponding measure are now added to the edge
weighted graph G. A vertex weight function µ : V → (0, ∞) provides the
Borel measure µ(U) = Pv∈U µ(v) on G, which may be extended [20, p. 257]
to G by defining the measure of G \ G to be 0. The real Hilbert space l2(µ)
will consist of functions f : V → R with Pv∈V f (v)2µ(v) < ∞, and inner
product hf, giµ = Pv∈V f (v)g(v)µ(v). The set of functions f : V → R which
are 0 at all but finitely many vertices is denoted by DK. Also introduce
DA,µ = A ∩ l2(µ).
The next proposition collects basic facts about the symmetric bilinear
forms induced by the edge conductances C(u, v). Closely related results
using the smaller domain DK are in [4, p. 20] and [16].
Proposition 3.1. Suppose C : V × V → [0, ∞) satisfies C(u, v) = C(v, u),
with C(u, v) > 0 if and only if [u, v] ∈ E. Given a vertex weight µ, the
symmetric bilinear form
B(f, g) =
1
2 X
v∈V
X
u∼v
C(u, v)(f (v) − f (u))(g(v) − g(u)),
f, g ∈ DA,µ, (3.1)
has a nonnegative quadratic form B(f, f ), and satisfies
B(f, g) = h∆µf, giµ = hf, ∆µgiµ,
(3.2)
10
where
∆µf (v) =
1
µ(v) X
u∼v
C(u, v)(f (v) − f (u)).
(3.3)
Proof. The nonnegativity of the quadratic form is immediate from the defi-
nition. Note that for any f ∈ A there are only finitely many vertices v ∈ V
for which f (v) − f (u) is nonzero if u is adjacent to v.
To identify the operator ∆µ, start with
2B(f, g) = X
v∈V
g(v)X
u∼v
C(u, v)(cid:16)f (v) − f (u)(cid:17)
(3.4)
−X
v∈V
(cid:16)X
u∼v
C(u, v)g(u)(f (v) − f (u))(cid:17)
Suppose a graph edge e has vertices v1(e) and v2(e). The second sum over
v ∈ V in (3.4) can be viewed as a sum over edges, with each edge contributing
the terms C(v1, v2)g(v1)(f (v2) − f (v1)) and C(v1, v2)g(v2)(f (v1) − f (v2)).
Using this observation to change the order of summation gives
X
v∈V
(cid:16)X
u∼v
C(u, v)g(u)(f (v) − f (u))(cid:17)
= X
e∈E
C(v1(e), v2(e))(cid:16)g(v1)(f (v2) − f (v1)) + g(v2)(f (v1) − f (v2))(cid:17)
= X
u∈V
g(u)X
v∼u
C(u, v)(f (v) − f (u))
Employing this identity in (3.4) gives
2B(f, g) = 2X
v
µ(v)g(v)[
1
µ(v) X
u∼v
C(u, v)(f (v) − f (u))].
(3.5)
With respect to the standard basis consisting of functions δw : V → R
with δw(w) = 1 and δw(v) = 0 for v 6= w, the operators ∆µ have the matrix
representation
Q(v, w) = n
v = w
v ∼ w
otherwise
o,
v, w ∈ V.
µ−1(w)Pu∼w C(u, w),
−µ−1(v)C(v, w),
0,
11
If v is fixed, then summing on w gives
Q(v, w) =
X
w
1
µ(v) X
u∼v
C(u, v) −
1
µ(v) X
u∼v
C(u, v) = 0,
so −Q(v, w) is a Q-matrix in the sense of Markov chains [17, p. 58].
In the Q - matrix formulation the matrix entries represent transition rates,
so decreasing the vertex measure µ increases the rates. Consistent with the
boundary value themes of this work an interesting choice is to take the vertex
weight µ0(v) to be half the sum of the lengths of the incident edges,
µ0(v) =
1
2 X
u∼v
R(u, v).
This choice makes the vertex measure consistent with the previously defined
graph volume,
µ0(G) = X
e∈E
R(e) = vol(G).
If vol(G) < ∞, then l2(µ0) will include all functions in A.
With respect to this measure
∆µ0f (v) = µ−1
0 (v)X
u∼v
C(u, v)(f (v) − f (u)) =
2
Pu∼v R(u, v) X
u∼v
f (v) − f (u)
R(u, v)
.
This operator resembles the symmetric second difference operator from nu-
merical analysis. Like the classical Laplace operator in Euclidean space, ∆µ0
exhibits quadratic scaling behavior. Suppose v0, v1 ∈ V have equal num-
bers of incident edges, and there is a bijection of vertex neighborhoods with
R(v1, w1) = ρR(v0, w0) for ρ > 0, and wj ∼ vj.
If f : V → R satisfies
f (v1) = f (v0) and f (w1) = f (w0), then
∆µ0f (v1) = ρ−2∆µ0f (v0).
The vertex weight µ0 is typically distinct from µ(v) = Pu∼v C(u, v), a
choice which appears in the study of discrete time Markov chains [6, p. 40],
[17, p. 73], [18, p. 18] with transition probabilities p(u, v) = µ−1(v)C(u, v)
for u 6= v.
12
3.2 Continuity
The next result considers continuous extension of functions to G when the
quadratic form is finite.
Theorem 3.2. Suppose G is connected. Using the metric of (2.1), functions
f : V → R with B(f, f ) < ∞ are uniformly continuous on G, and so f
extends uniquely to a continuous function on G.
Proof. If v, w ∈ V and γ = (v = v0, v1, . . . , vK = w) is any finite simple path
from v to w, then the Cauchy-Schwarz inequality gives
[f (vk+1) − f (vk)]
X
k
2
C 1/2(vk+1, vk)
C 1/2(vk+1, vk)(cid:12)(cid:12)(cid:12)
f (w) − f (v)2 = (cid:12)(cid:12)(cid:12)
≤ X
k
[C(vk+1, vk)(f (vk+1) − f (vk))2]X
R(vk+1, vk)
k
≤ 2B(f, f )X
k
R(vk+1, vk).
There is a simple path with Pk R(vk+1, vk) ≤ 2d(v, w), so
f (w) − f (v)2 ≤ 4B(f, f )d(v, w),
(3.6)
which shows f is uniformly continuous on G. By [20, p. 149] f extends
continuously to G.
The bilinear form may be used to define a 'Sobolev style' Hilbert space
H 1(µ) with inner product
hf, giµ,1 = X
v
f (v)g(v)µ(v) + B(f, g).
Let H 1
0 (µ) be the closure of DK in H 1(µ).
Lemma 3.3. If G is connected with finite diameter, then there is a constant
C such that
f (v) ≤ Ckf kµ,1,
sup
v∈V
(3.7)
so a Cauchy sequence in H 1(µ) is a uniform Cauchy sequence. The functions
f in the unit ball of H 1 are uniformly equicontinuous [20, p. 29].
13
Proof. Fixing a vertex v0, (3.6) gives
f (v) ≤ f (v0) + f (v) − f (v0) ≤ kf kµ,1/pµ(v0) + 2kf kµ,1diam(G)1/2,
which is (3.7). The uniform equicontinuity follows from (3.6).
If f ∈
0 (µ), then f has a unique continuous extension to G which is zero at all
Theorem 3.4. Suppose G is connected and has finite diameter.
H 1
points x ∈ G \ G.
Proof. Any function f ∈ H 1
0 (µ) is the limit in H 1(µ) of a sequence fn from
DK. The functions f and fn have unique continuous extensions to G by
Theorem 3.2. The extended functions fn satisfying fn(x) = 0 for all x ∈
G \ G. By Lemma 3.3 the sequence fn converges to f uniformly on G, so the
extensions fn converge uniformly to the extension f on G. Thus f (x) = 0
for all x ∈ G \ G.
4 Boundary value problems
4.1 The basic Laplacian
Let SK,µ denote the operator ∆µ on l2(µ) with the domain DK.
Proposition 4.1. The operator SK,µ is symmetric and nonnegative on l2(µ).
The adjoint operator S∗
K,µ on l2(µ) acts by
(S∗
K,µh)(v) = ∆µh(v) =
1
µ(v) X
u∼v
C(u, v)(h(v) − h(u))
on the domain consisting of all h ∈ l2(µ) for which ∆µh ∈ l2(µ).
Proof. The symmetry and nonnegativity of SK,µ are given by (3.2). Since
SK,µ is densely defined, S∗
K,µ is the operator whose graph is the set of pairs
(h, k) ∈ l2(µ) ⊕ l2(µ) such that
hSK,µf, hiµ = hf, kiµ
for all f ∈ DK. Suppose fv = 1
µ(v) δv Then for any h in the domain of S∗
K,µ,
k(v) = (S∗
K,µh)(v) = hSK,µfv, hiµ = X
C(u, w)(fv(w) − fv(u))]h(w)
[X
u∼w
w
14
=
1
µ(v) X
u∼v
C(u, v)(h(v) − h(u)).
Proposition 4.1 provides a basic Laplace operator, the Friedrich's exten-
sion [14, pp. 322-326] of SK,µ, whose domain is a subset of H 1
0 (µ) the closure
of DK in H 1(µ). Let LK,µ denote the Friedrich's extension of SK,µ. Several
features of LK,µ are implied by the condition µ(G) < ∞.
Proposition 4.2. Suppose G is connected, with finite diameter and infinitely
many vertices. If µ(G) < ∞, f ∈ domain(LK), and kf kµ = 1, then LK,µ has
the strictly positive lower bound
hLK,µf, f iµ = B(f, f ) ≥
1
4µ(G)diam(G)
,
(4.1)
Proof. The Friedrich's extension LK,µ of the nonnegative symmetric operator
SK,µ has the same lower bound, so it suffices to consider functions f ∈ DK.
Since kf kµ = 1 there must be some vertex v where f 2(v) ≥ µ−1(G). Since f
has finite support, there is another vertex u with f (u) = 0. An application
of (3.6) gives
µ−1(G) ≤ f 2(v) = [f (v) − f (u)]2 ≤ 4B(f, f )d(u, v).
Proposition 4.3. Suppose G is connected, G is compact, and µ(G) is finite.
Let S1,µ be a symmetric extension of SK,µ in l2(µ) whose associated quadratic
form is
hS1,µf, f iµ = B(f, f ).
Then the Friedrich's extension L1,µ of S1,µ has compact resolvent.
Proof. The resolvent of L1,µ maps a bounded set in l2(µ) into a bounded set in
H 1(µ). Suppose fn is a bounded sequence in l2(µ), with gn = (L1 − λI)−1fn.
By Lemma 3.3 and the Arzela-Ascoli Theorem [20, p. 169] the sequence gn
has a uniformly convergent subsequence, which converges in l2(µ).
15
4.2 Harmonic functions
When considering harmonic functions, it will be convenient to treat vertices
of degree one (boundary vertices) as part of the boundary of G. For instance,
this convention allows finite graphs with degree one vertices to have noncon-
stant harmonic functions. Let Vint denote the set of interior vertices, that is
vertices with degree at least 2. Say that a function f : V → R is harmonic
on G if
f (v) =
1
Pu∼v C(u, v) X
u∼v
C(u, v)f (u)
(4.2)
for all vertices v ∈ Vint. Since the value f (v) of a harmonic function is the
weighted average of the values at the adjacent vertices, f has a minimum or
maximum at some v ∈ Vint of a connected graph if and only if f is constant.
The set of harmonic functions is independent of the vertex weight µ, but
harmonic functions may be described as those for which
∆µf (v) =
1
µ(v) X
u∼v
C(u, v)(f (v) − f (u)) = 0,
v ∈ Vint.
Proposition 4.4. Suppose G is connected and G is compact. If G \ G 6= ∅,
then 0 is not an eigenvalue of an operator LK,µ.
Proof. If 0 were an eigenvalue of LK,µ, then there would be a corresponding
eigenfunction f which is positive somewhere. By Theorem 3.4, f would
extend continuously to G, with f (x) = 0 for all x ∈ G \ G. Since G is
compact, f has a positive maximum at some x ∈ G. Since (4.2) holds at all
vertices, f is constant, but then G \ G 6= ∅ implies f = 0.
Define ∂G to be the complement of Vint in G. The Dirichlet problem asks
whether every continuous function F : ∂G → R has a continuous extension
f : G → R which is harmonic on G. The following example shows that the
Dirichlet problem is not always solvable.
As shown in Figure A, construct a graph G beginning with a sequence
of vertices vn and with edges (vn, vn+1), n = 0,1,2,. . . . Take R(vn, vn+1) =
2−n−1. For each of the vertices vn introduce 2n additional vertices un,k, and
edges (vn, un,k) of length 1. In this example, ∂G consists of the boundary
vertices un,k together with the boundary point at v = limn vn. All of these
points are isolated in ∂G.
16
Suppose the function F satisfies
F (un,k) = 0, F (v) = 1.
A computation with (4.2) shows that any harmonic extension f which is
continuous on G must satisfy the contradictory requirements
lim
n→∞
f (vn) = 1,
lim
n→∞
f (vn) = 3/4.
...
0
1
Figure A
The graph G in this example is connected with finite diameter, G is weakly
connected, but G is not compact. If G is required to be compact, there is a
positive result.
Theorem 4.5. Suppose G is connected, while G is weakly connected and
compact. Every continuous function F : ∂G → R has a unique extension to
a continuous function f : G → R that is harmonic on G.
Proof. The proof breaks into two main parts. For the first part, assume that
G has no boundary vertices, so F is defined on G \ G. The second part of the
proof will extend this partial result to the full theorem.
The uniqueness is a standard consequence of the maximum principle since
the difference of two solutions would be 0 on G \ G. By the Tietze extension
theorem [20, p. 179], F may be extended to a continuous function on G. By
Theorem 2.12, for any n > 0 there is a gn ∈ A with
max
x∈G
F (x) − gn(x) ≤ 1/n.
Pick vertex weights µ with µ(G) < ∞. Since gn ∈ A, the function ∆µgn
has the value 0 except at a finite set of vertices. The compactness of G
17
means that G has finite diameter. By Proposition 4.2 the operator LK,µ has
a strictly positive lower bound, so the equation
−Lµhn = ∆µgn,
hn(cid:12)(cid:12)G\G = 0,
has a solution. The function fn = gn + hn is harmonic on V, continuous on
G, and satisfies
max
x∈G\G
F (x) − fn(x) ≤ 1/n.
Since {fn} converges uniformly on G \ G, the maximum principle implies
it is a uniformly Cauchy sequence on G, so there is a continuous limit f on
G. At each v ∈ V
f (v) = lim
n→∞
fn(v) = lim
n→∞
1
Pu∼v C(u, v) X
u∼v
C(u, v)fn(u)
=
1
Pu∼v C(u, v) X
u∼v
C(u, v)f (u),
so f is harmonic on G. This completes the first main part of the proof.
The second main step involves extending the theorem to include boundary
vertices. For each edge e = [v0, vb] with boundary vertex vb and edge length
re = C(v0, vb)−1, delete the vertex vb and edge e from G. For n = 1, 2, 3, . . . ,
add a sequence of new vertices vn and edges en = [vn, vn+1]. Assume the
edges en have length rn with P rn = re. The resulting graph G1 will have a
completion G 1 containing a point w = limn→∞ vn.
Suppose a continuous function G : G1 \ G1 → R is given, and is extended
to the harmonic function g on G1. With C(vn, vn+1) = r−1
n we have
g(vn) =
1
1/rn−1 + 1/rn
[
1
rn−1
g(vn−1) +
1
rn
g(vn+1)]
=
rn
rn−1 + rn
g(vn−1) +
rn−1
rn−1 + rn
g(vn+1),
(g(vn) − g(vn−1))/rn−1 = (g(vn+1) − g(vn))/rn,
or
and so g is a linear function of the distance along the path from v0 to w.
At v0 the fact that g is harmonic means
C(u, v0)(g(v0) − g(u)) = 0.
X
u∼v0
18
The linearity of g along the path means that
C(v1, v0)(g(v0) − g(v1)) = r−1
0 (g(v0) − g(v1)) = r−1
e (g(v0) − g(w)).
Now define the function f : G → R by f (v) = g(v) for v ∈ G1 ∩G, and f (vb) =
g(w). The function f is harmonic on G, and extends to the continuous
function F : G → R where F (x) = G(x) for x ∈ G \ G, while F (vb) = G(w)
for each boundary vertex vb.
4.3 Boundary conditions and operators
In this section absorbing and reflecting boundary conditions are used to con-
struct distinct nonnegative self adjoint extensions of SK,µ. The constructed
operators extend to semigroup generators which are positivity preserving
contractions on l1(µ).
Given a closed set Ω ⊂ {G \ G}, let AΩ denote the subalgebra of A
vanishing on Ω. Define the domain
DΩ = AΩ ∩ l2(µ).
(4.3)
Let SΩ,µ denote the operator with domain DΩ acting on l2(µ) by SΩf = ∆µf .
By Proposition 3.1 the operator SΩ,µ is nonnegative and symmetric, with
quadratic form hSΩ,µf, f iµ = B(f, f ). Let LΩ,µ denote the Friedrich's exten-
sion of SΩ,µ, and note that the domain of LΩ,µ is a subset of H 1(µ). A slight
modification of the proof of Theorem 3.4 shows that every function fj in the
domain of LΩ,µ extends continuously to G with fj(x) = 0 for x ∈ Ω.
Given a compact set Ω ⊂ G, let NR(Ω) = {z ∈ G d(z, Ω) ≤ R} be the
set of points whose distance from Ω is at most R.
Theorem 4.6. Suppose G is connected, weakly connected, and compact.
Assume that for j = 1, 2 the sets Ω(j) ⊂ {G \ G} are compact, and that
µ(NR(Ω(j))) < ∞ for some R > 0. If Ω(1) 6= Ω(2), then LΩ(1),µ and LΩ(2),µ
have distinct domains.
Proof. Reversing the roles of Ω(1) and Ω(2) if necessary, we may assume
Ω(1) 6⊂ Ω(2). Let Ω(2)c denote the complement of Ω(2). Find x ∈ Ω(1) ∩
Ω(2)c, and an r with 0 < r < R such that Nr(x) ⊂ Ω(2)c.
The sets {x} and V = {y ∈ Gd(x, y) ≥ r} are disjoint and compact.
Corollary 2.11 shows there is a function f ∈ A with f (x) = 1, while f (z) = 0
19
for all z ∈ V . Since µ(Nr(x)) < ∞, the function f is in the domain of LΩ(2),µ,
but f is not in the domain of LΩ(1),µ.
Since the operators L = LΩ,µ are nonnegative self adjoint on l2(µ), the
operators of the semigroup exp(−tLΩ,µ) are l2(µ) contractions for t ≥ 0. For
applications to probability this is not sufficient. Let Quad(L) denote the
domain of L1/2. The method of [4, p. 20] will show that the quadratic forms
Q(f ) = hL1/2f, L1/2f iµ,
f ∈ Quad(L)
associated to LΩ,µ are Dirichlet forms.
There are two conditions to check. The first condition is that f ∈
Quad(L) implies f ∈ Quad(L) and B(f , f ) ≤ B(f, f ). Since the form is
B(f, f ) =
1
2 X
v∈V
X
u∼v
C(u, v)(f (v) − f (u))2,
the first condition holds for f ∈ DΩ. If f ∈ Quad(L) then there is a sequence
fn ∈ DΩ with
hf, f iµ + B(f, f ) = lim
n→∞
hfn, fniµ + B(fn, fn).
It follows that
hf , f iµ + B(f , f ) = lim
n→∞
hfn, fniµ + B(fn, fn),
and B(f , f ) ≤ B(f, f ).
The second condition is that if f ∈ Quad(L) and g ∈ l2(µ) with g(v) ≤
f (v) and g(v) − g(u) ≤ f (v) − f (u) for all u, v ∈ V, then g ∈ Quad(L)
and Q(g) ≤ Q(f ). This is even more transparent than the first condition.
Again quoting [4, p. 12-13], the following result is established.
Theorem 4.7. For t ≥ 0 the semigroups exp(−LΩ,µt) on l2(µ) are positivity
preserving contractions on lp(µ) for 1 ≤ p ≤ ∞
Arguing by analogy, one expects the operator L = LΩ,µ to exhibit 'absorp-
tion' at Ω and 'reflection' at Ωc. This analogy may be tested by estimating
the rate of decay of a probability density function initially supported near
the 'reflecting' boundary.
20
Suppose G is compact, connected and weakly connected. Using Theo-
rem 2.6 and the subsequent remarks, let V be a clopen neighborhood of
Ω, with U = V c. The edge boundary of U will be the set ∂eU of edges
e = [u, v] ∈ G with u ∈ U and v ∈ V .
The edge boundary ∂eU is a finite set since by Theorem 2.4 there is a
finite set of edges W such that any path from U to V contains an edge from
W , and ∂eU ⊂ W . Let 1U be the indicator function of U, with 1U (x) = 1 for
x ∈ U and 1U (x) = 0 for x ∈ V . The function 1U is in A. If µ(V ) < ∞ then
1U is in the domain of L.
Suppose p0 : V → [0, ∞) is a bounded probability density with p0(v) ≥ 0
and PG p0(v)µ(v) = 1. For t ≥ 0 define p(t, v) = exp(−tL)p0(v) and denote
the density integrated over U by PU (t) = Pv∈U p(t, v)µ(v).
Using p0 ∈ l2(µ), we find that for t > 0,
d
dt
PU (t) =
d
dt
hexp(−tL)p0, 1U iµ = −hL exp(−tL)p0, 1U iµ
= −hexp(−tL)p0, L1U iµ = − X
(u,v)∈∂eU
C(u, v)(p(t, u) − p(t, v)).
Since the semigroup is positivity preserving and a contraction on l∞(µ),
d
dt
PU (t) ≥ −kp0k∞ X
(u,v)∈∂eU
C(u, v).
(4.4)
That is, the decay rate for PU (t) is controlled by what happens at ∂eU,
without regard to the rest of the boundary of U.
21
References
[1] R. Carlson. Boundary Value Problems for Infinite Metric Graphs Anal-
ysis on Graphs and Its Applications, PSPM 77 (2008), pp. 355 -- 368.
[2] F. Chung. Spectral Graph Theory. American Mathematical Society,
Providence, 1997.
[3] Y. Colin de Verdiere, N. Torki-Hamza, and F. Truc. Essential self-
adjointness for combinatorial Schr´'odinger operators II - Metrically non-
complete graphs Mathematical Physics, Analysis, and Geometry 14,
(2011) 21-38.
[4] E.B. Davies. Heat Kernels and Spectral Theory. Cambridge University
Press, 1990.
[5] R. Diestel. Graph Theory. Springer, 2005.
[6] P. Doyle and J. Snell. Random Walks and Electrical Networks. Mathe-
matical Association of America, Washington, D.C., 1984.
[7] W. Feller. On Boundaries and Lateral Conditions for the Kolmogorov
Differential Equations The Annals of Mathematics 65 (1957), pp. 527 --
570.
[8] G. Folland. Real Analysis. John Wiley and Sons, New York, 1984.
[9] M. Fukushima and Y. Oshima and M. Takeda. Dirichlet Forms and
Symmetric Markov Processes. de Gruyter, Berlin, 1994.
[10] A. Georgakopoulos. Uniqueness of electrical currents in a network of
finite total resistance J. London Math. Soc. 82 (2010). pp. 256 -- 272
[11] A. Georgakopoulos. Graph topologies induced by edge lengths Discrete
Mathematics 311 (2011). pp. 1523 -- 1542
[12] J. Hocking and G. Young. Topology. Addison-Wesley, 1961.
[13] P.E.T. Jorgensen. Essential selfadjointness of the graph-Laplacian J.
Math. Phys. Vol. 49, No. 7 (2008)
[14] T. Kato. Perturbation Theory for Linear Operators. Springer-Verlag,
New York, 1995.
22
[15] M. Keller and D. Lenz. Dirichlet Forms and Stochastic Completeness of
Graphs and Subgraphs Preprint (2009)
[16] M. Keller and D. Lenz. Unbounded Laplacians on Graphs: Basic Spectral
Properties and the Heat Equation Mmath. Model. Nat. Phenom. Vol.
5, No. 2 (2010)
[17] T. Liggett. Continuous Time Markov Processes. American Mathemati-
cal Society, Providence, 2010.
[18] R. Lyons and Y. Peres. Probability on Trees and Networks.
[19] B. Maury, D. Salort, and C. Vannier. Trace theorem for trees and appli-
cation to the human lungs Networks and Heterogeneous Media 4 (2009),
pp. 469-500.
[20] H. Royden. Real Analysis. Macmillan, New York, 1988.
[21] W. Woess. Denumerable Markov Chains. European Mathematical So-
ciety, 2009.
23
|
1003.2285 | 5 | 1003 | 2010-09-01T12:39:15 | On diagonalizable operators in Minkowski spaces with the Lipschitz property | [
"math.FA"
] | A real semi-inner-product space is a real vector space $\M$ equipped with a function $[.,.] : \M \times \M \to \Re$ which is linear in its first variable, strictly positive and satisfies the Schwartz inequality. It is well-known that the function $||x|| = \sqrt{[x,x]}$ defines a norm on $\M$. and vica versa, for every norm on $X$ there is a semi-inner-product satisfying this equality. A linear operator $A$ on $\M$ is called \emph{adjoint abelian with respect to $[.,.]$}, if it satisfies $[Ax,y]=[x,Ay]$ for every $x,y \in \M$. The aim of this paper is to characterize the diagonalizable adjoint abelian operators in finite dimensional real semi-inner-product spaces satisfying a certain smoothness condition. | math.FA | math |
ON DIAGONALIZABLE OPERATORS IN MINKOWSKI SPACES
WITH THE LIPSCHITZ PROPERTY
ZSOLT L ´ANGI
Abstract. A real semi-inner-product space is a real vector space M equipped
with a function [., .] : M × M → R which is linear in its first variable, strictly
positive and satisfies the Schwartz inequality. It is well-known that the function
x = p[x, x] defines a norm on M. and vica versa, for every norm on X there
is a semi-inner-product satisfying this equality. A linear operator A on M is
called adjoint abelian with respect to [., .], if it satisfies [Ax, y] = [x, Ay] for
every x, y ∈ M. The aim of this paper is to characterize the diagonalizable
adjoint abelian operators in finite dimensional real semi-inner-product spaces
satisfying a certain smoothness condition.
1. Introduction and preliminaries
A real semi-inner-product space is a real linear space M equipped with a function
[., .] : M × M → R, called a semi-inner-product, such that
(1) [., .] is linear in the first variable,
(2) [x, x] ≥ 0 for every x ∈ M, and [x, x] = 0 yields that x = 0,
(3) [x, y]2 ≤ [x, x] · [y, y] for every x, y ∈ M.
These spaces were introduced in 1961 by Lumer [8], and have been extensively
studied since then (cf., for example [1]). It was remarked in [8] that in a real semi-
inner-product space M, the function x = p[x, x] defines a norm. The converse
also holds, i.e. if M is a real linear space, then for every real norm . : M → R, there
is a semi-inner-product [., .] : M × M → R satisfying x = p[x, x]. Furthermore,
the semi-inner-product determined by a norm is unique if, and only if, its unit
ball is smooth; that is, if the unit sphere has a unique supporting hyperplane at
its every point. By [4], in this case the semi-inner-product is homogeneous in the
second variable; i.e., [x, λy] = λ[x, y] for any x, y ∈ M and λ ∈ R.
We say that a real semi-inner-product is continuous, if for every x, y, z ∈ M
[4] or
with [x, x] = [y, y] = [z, z] = 1, λ → 0 yields that [x, y + λz] → [x, y] (cf.
[5]). It is well-known that the semi-inner-product determined by a smooth norm is
continuous; it follows, for example, from E∗ on page 118 of [11] and Theorem 3 of
[4].
A linear operator A is called adjoint abelian with respect to a semi-inner-product
[., .], if it satisfies [Ax, y] = [x, Ay] for every x, y ∈ M (cf., for instance [2] and [3]).
1991 Mathematics Subject Classification. 47A05, 52A21, 46B25.
Key words and phrases. semi-inner-product space, Minkowski space, norm, adjoint abelian.
1
2
Z. L ´ANGI
In the following, M denotes a smooth Minkowski space; that is, a real finite
dimensional smooth normed space, and . and [., .] denote the norm and the
induced semi-inner-product of M, respectively. We denote by S the unit sphere
with respect to the norm, i.e., we set S = {x ∈ M : x = 1}. We say that the
semi-inner-product [., .] has the Lipschitz property, if for every x ∈ S, there is a
real number κ such that for every y, z ∈ S, we have [x, y] − [x, z] ≤ κy − z.
We note that in a similar way, a differentiability property of semi-inner-products
was defined in [5], and that any semi-inner-product satisfying that differentiability
property satisfies also the Lipschitz property.
The aim of this paper is to characterize the diagonalizable adjoint abelian op-
erators in finite dimensional spaces with a semi-inner-product that satisfies the
Lipschitz property.
To formulate our main result, we need the following notions and notations. An
isometry of M is an operator A : M → M satisfying Ax = x for every x ∈ M,
or, equivalently, [Ax, Ay] = [x, y] for every x, y ∈ M (cf. [6]). For the properties of
isometries in Minkowski spaces, the interested reader is referred to [9].
For the following definition, see also [4].
Definition 1. If x, y ∈ M and [x, y] = 0, we say that x is transversal to y, or y is
normal to x. If X, Y ⊂ M such that [x, y] = 0 for every x ∈ X and y ∈ Y , we say
that X is transversal to Y or Y is normal to X.
Definition 2. Let U and V be linear subspaces of M such that M = U ⊕ V . If for
every xu, yu ∈ U and xv, yv ∈ V , we have
[xu + xv, yu + yv] = [xu, yu] + [xv, yv],
then we say that the semi-inner-product [., .] is the direct sum of [., .]U and [., .]V ,
and denote it by [., .] = [., .]U + [., .]V . If there are no such linear subspaces of M,
we say that [., .] is non-decomposable.
We remark that if [., .] = [., .]U + [., .]V for some linear subspaces U and V , then
U and V are both transversal and normal, and that the converse does not hold. We
note also that any two semi-inner-product spaces can be added in this way (cf. [5]).
Definition 2 can be formulated for finitely many subspaces as well in the natural
way. For the simplicity of notation, we mean that every semi-inner-product space
is the direct sum of itself.
Let A : M → M be a linear operator, and let λ1 > λ2 > . . . > λk ≥ 0 be the
absolute values of the eigenvalues of A. If λi is an eigenvalue of A, then Ei denotes
the eigenspace of A belonging to λi, and if λi is not an eigenvalue, we set Ei = {0}.
We define E−i similarly with −λi in place of λi, and set ¯Ei = span(Ei ∪ E−i). Our
main theorem is the following.
Theorem 1. Let M be a smooth Minkowski space such that the induced semi-inner-
product [., .] : M × M → R satisfies the Lipschitz condition, and let A : M → M be
a diagonalizable linear operator. Then A is adjoint abelian with respect to [., .] if,
and only if, the following hold.
(1) [., .] is the direct sum of its restrictions to the subspaces ¯Ei, i = 1, 2, . . . , k;
(2) for every value of i, the subspaces Ei and E−i are both transversal and
normal;
ON ADJOINT ABELIAN OPERATORS
3
(3) for every value of i, the restriction of A to ¯Ei is the product of λi and an
isometry of ¯Ei.
From Theorem 1, we readily obtain the following corollary.
Corollary 1. Let M be a smooth Minkowski space such that the induced semi-inner-
product [., .] satisfies the Lipschitz condition. Then the following are equivalent.
(1) [., .] is non-decomposable;
(2) every diagonalizable adjoint abelian linear operator of M is a scalar multiple
of an isometry of M.
Note that if A is not diagonal, then we may apply Theorem 1 for the span of the
eigenspaces of A.
Corollary 2. Let M be a smooth Minkowski space such that the induced semi-inner-
product [., .] : M × M → R satisfies the Lipschitz condition, and let A : M → M
be an adjoint abelian linear operator with respect to [., .]. Then (1), (2) and (3) in
Theorem 1 hold for A .
If [., .] = [., .]U + [., .]V for some subspaces U and V , and u ∈ U and v ∈ V ,
then, by Theorem 1, S ∩ span{u, v} is an ellipse. This observation is proved, for
example, in Statement 1 of [5]. Thus, we have the following.
Corollary 3. Let M be a smooth Minkowski space such that the induced semi-
inner-product satisfies the Lipschitz condition. If no section of the unit sphere S
with a plane is an ellipse with the origin as its centre, then every diagonalizable
adjoint abelian operator of M is a scalar multiple of an isometry of M.
In the proof of Theorem 1, we need the following lemma.
Lemma 1. Let M be a smooth Minkowski space. Let ., [., .] and S denote the
norm, the associated semi-inner-product and the unit sphere of M. Then the fol-
lowing are equivalent.
(1) [., .] satisfies the Lipschitz condition;
(2) for every x ∈ M, the function fx : M → R, fx(y) = [x, y] is uniformly
continuous on M; that is, for every x ∈ M and ε > 0 there is a δ > 0 such
that y, z ∈ M and y − z < δ imply [x, y] − [x, z] < ε;
(3) for every x ∈ M and any sequences {yn}, {zn} in M, if yn − zn → 0,
then [x, yn] − [x, zn] → 0.
Proof. Note that (2) and (3) are equivalent. We prove that (1) and (3) are equiva-
lent.
First we show that (1) yields (3). Observe that since [x, y] is homogeneous in x,
it suffices to prove (3) for x ∈ S. Let x ∈ S, and assume that there is a number
κ ∈ R such that for every y, z ∈ S, we have [x, y] − [x, z] < κy− z. Consider the
sequences {yn} and {zn} in M, and assume that yn− zn → 0. Since a continuous
function is uniformly continuous on any compact set and since the unit ball of M
is compact, we may assume that yn ≥ 1 and that zn ≥ 1 for every n. Let
wn = zn
yn yn. Observe that, by the definition of semi-inner-product, [u, v] ≤ 1
4
Z. L ´ANGI
yn
wn
obtain that
for any u, v ∈ S. Then, from (cid:12)(cid:12)(cid:12)hx, yn
[x, yn] − [x, zn] ≤ [x, yn] − [x, wn] + [x, wn] − [x, zn] =(cid:12)(cid:12)zn − yn(cid:12)(cid:12)·
·(cid:12)(cid:12)(cid:12)(cid:12)
yn(cid:21)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:20)x,
Note that wn − yn = (cid:12)(cid:12)yn − zn(cid:12)(cid:12) ≤ yn − zn → 0 and that wn − zn ≤
yni(cid:12)(cid:12)(cid:12) ≤ 1 and from the triangle inequality, we
wn(cid:21) −(cid:20)x,
wn − yn + yn − zn → 0, from which it follows that [x, yn] − [x, zn] → 0.
+ zn ·(cid:12)(cid:12)(cid:12)(cid:12)
(cid:20)x,
Assume that (1) does not hold. Then there is a point x ∈ S and sequences
yn, zn ∈ S such that [x, yn] − [x, zn] = κnyn − zn where κn → ∞. We may
assume that κn > 0 for every n, and since S is compact, also that yn → y and
zn → z for some y, z ∈ S. Note that κn → ∞ implies y = z. Let δn = yn − zn,
and assume that δn > 0 for every n. Observe that as yn and zn converge to the
same point, we have δn → 0, and, as [x, y] is continuous in y ∈ S for every x ∈ S,
we have also that κnyn − zn = κnδn → 0. Let un = yn
. Then
un − vn = 1
κn → 0, and [x, un] − [x, vn] = 1, and hence, (3) does not hold. (cid:3)
and vn = zn
κnδn
≤ yn − zn + κwn − zn.
zn
zn(cid:21)(cid:12)(cid:12)(cid:12)(cid:12)
κnδn
2. Proof of Theorem 1
Assume that A is adjoint abelian. Let µ and ν be two different eigenvalues of A
and let x and y be eigenvectors belonging to µ and ν, respectively. Then,
µ[x, y] = [Ax, y] = [x, Ay] = ν[x, y],
which yields that x is transversal to y. Thus, any two eigenspaces, belonging to
distinct eigenvalues, are both transversal and normal, which, in particular, proves
(2) (for isometries, see this observation in [7]). Recall that an Auerbach basis of
a Minkowski space is a basis in which any two distinct vectors are transversal and
normal to each other, and that in every norm there is an Auerbach basis. Note
that the restriction of a norm to a linear subspace is also a norm, and thus, we
may choose Auerbach bases in each eigenspace separately, which, by the previous
observation, form an Auerbach basis in the whole space. Let x ∈ M, and observe
that x has a unique representation of the form x = Pk
i=1 xi, where xi ∈ ¯Ei. To
prove Theorem 1, we need the following lemma.
Lemma 2. If z ∈ ¯Ei for some value of i, then [z, x] = [z, xi].
Proof. Assume that z ∈ ¯Ei for some i ∈ {1, 2, . . . , k}.
Case 1, i = 1.
If λ1 = 0, then A is the zero operator, and the assertion immediately follows. Let
us assume that λ1 > 0. As A is adjoint abelian, we have that [A2z, x] = [Az, Ax] =
[z, A2x]. Observe that A2z = λ2
i xi. Thus,
i=1 λ2
(1)
1z, and that A2x =Pk
xi#
[z, x] ="z,
1(cid:19)n
Xi=1(cid:18) λ2
i
λ2
k
for every positive integer n. Since [., .] is continuous in both variables, we obtain
that the limit of the right-hand side of (1) is [z, x1], and hence, [z, x] = [z, x1].
ON ADJOINT ABELIAN OPERATORS
5
Case 2, i > 1 and λi 6= 0.
We prove by induction on i. Let us assume that [y, x] = [y, xj] for every y ∈ ¯Ej
and j = 1, 2, . . . , i − 1.
First, set w = Pi
¯Ej ). We show that [z, w] = [z, xi].
Note that AFi is invertible, adjoint abelian, and its inverse is also adjoint abelian.
Let Bi denote the inverse of AFi , and observe that the absolute values of the
eigenvalues of Bi are 1
and its eigenspaces are Ej and E−j, where j = 1, 2, . . . , i.
λj
Thus, we have 1
λ2
i
j=1 xj and Fi = span(Si
i z, w] = [Biw, Biw] = [z, B2
[z, w] = [B2
i w] and
j=1
[z, w] =
z,
i
j!n
Xj=1 λ2
i
λ2
xj
for every positive integer n. By the continuity of the semi-inner-product, and since
the limit of the right-hand side is [z, xi], we have the desired equality.
Now we show that [z, x] = [z, xi]. Similarly like before, we obtain that
j
λ2
1
λ2
j
k
z,
= 0, and
n
j
λ2
n
j
λ2
other hand,
(cid:12)(cid:12)(cid:12)(cid:12)
xj(cid:12)(cid:12)(cid:12)(cid:12)
Case 3, i > 1 and λi = 0.
[z, x] =
integer n. Thus, by Lemma 1, we have that [z, x] = [z, xi].
xj
i!n
Xj=1 λ2
(cid:12)(cid:12)(cid:12)(cid:12)Pk
for every positive integer n. Observe that limn→∞(cid:12)(cid:12)(cid:12)(cid:12)
j=i+1(cid:16) λ2
i(cid:17)
xj(cid:21) = [z, xi] for every positive
that, by the previous paragraph, (cid:20)z,Pi
j=1(cid:16) λ2
i(cid:17)
Note that in this case i = k and ¯Ek = Ek. Let F = span(Sk−1
xf = Pk−1
j=1 xj. By Cases 1 and 2, we have [xf , xk] = Pk−1
xj
= 0.
¯Ei) and set
j=1 [xj , xk] = 0. On the
xj
=
A2xk,
Subcase 3.1, dim G = 2.
xf and e2 = xk
Thus, we obtained that F and Ek are both transversal and normal. Now we let
G = span{xf , xk, z}.
Let e1 = xf
xk . Since F and Ek are transversal and normal, the pair
j=1 [., .] ¯Ej , which
{e1, e2} is an Auerbach basis in G. By Cases 1 and 2, [., .]F =Pk−1
yields that [e1, α1e1 + α2e2] = α1 for any α1, α2 ∈ R.
Now we identify G with the Euclidean plane R2 by α1e1 + α2e2 7→ (α1, α2); or
in other words, we assume that e1 and e2 are the standard basis of an underlying
Euclidean plane. We need to show that [e2, α1e1 + α2e2] = α2 for any α1, α2 ∈ R,
or, equivalently, that the unit circle S ∩ G of the subspace G is the Euclidean unit
circle.
[xk, xf ] =
xk, A2
k−1
Xj=1
k−1
Xj=1
1
λ2
j
i=1
Since M is smooth, S∩G is a convex differentiable curve. Consider the Descartes
coordinate system induced by the standard basis e1 and e2, and note that the
lines x = 1, x = −1, y = 1 and y = −1 support conv S. Thus, for every value
6
Z. L ´ANGI
of x ∈ (−1, 1), there is exactly one point of S with x as its x-coordinate and
nonnegative y-coordinate. We represent the points of S ∩ G with nonnegative y-
coordinates as the union of the graph of a function x 7→ f (x) with x ∈ [−1, 1], and
(possibly) two segments on the lines with equations x = 1 and x = −1. We express
the equality [e1, α1e1 + α2e2] = α1 with the function f .
We may assume that v = α1e1 +α2e2 ∈ S∩G. Consider the case that v = x0e1 +
f (x0)e2 for some x0 ∈ (−1, 1). Then we have [e1, v] = [e1, x0e1 +f (x0)e2] = x0. Let
vp denote the projection of e1 onto the line {λv : λ ∈ R} parallel to the supporting
line of conv S at v, and let ep denote the projection of v onto the line {λe1 : λ ∈ R}
parallel to the supporting line of conv S at e1 (cf. Figure 1). Let vp = µv and
observe that ep = x0e1. We note that, by the construction of the semi-inner-product
described, for example in [8], we have that [e1, v] = µ and [v, e1] = x0. Hence the
triangle with vertices o, e1, v is similar to the triangle with vertices o, vp, ep, with
similarity ratio x0. From this, we obtain that vp = x0v = x2
0e1 + x0f (x0)e2. As the
line, passing through e1 and vp, is parallel to the supporting line of conv S at v, we
have
which is an ordinary differential equation for f with the initial condition f (0) = 1.
x0f (x0)
1 − x2
0
f ′(x0) = −
,
Figure 1. An illustration for Subcase 3.1
We omit an elementary computation that shows that the solution of this differ-
ential equation is y = √1 − x2. Thus, we obtain that S ∩ G is the Euclidean unit
circle, which yields, in particular, that [z, xf + xk] = [z, xk].
Subcase 3.2, dim G = 3.
Set e1 = xf
xf and choose an Auerbach basis {e2, e3} in span{xk, z}. Then the set
{e1, e2, e3} is an Auerbach basis in G. Furthermore, since F and Ek are transversal
and normal, {e1, v} is an Auerbach basis in its span for any v ∈ span{xk, z} with
v = 1. Thus, applying the argument in Subcase 3.1 for the subspace span{e1, v},
we obtain that S ∩ span{e1, v} is the ellipse with semiaxes e1 and v. Note that this
property and S ∩ span{e1, e2} determines the norm.
ON ADJOINT ABELIAN OPERATORS
7
Consider the semi-inner-product defined by [β1e1 + β2e2 + β3e3, α1e1 + α2e2 +
α3e3]′ = β1α1 + [β2e2 + β3e3, α2e2 + α3e3]. We show that [., .]′ and [., .] define the
same norm, which, as a smooth norm uniquely determines its semi-inner-product,
yields that [., .]′ = [., .].
Let v = α2e2 + α3e3 with v = 1 be arbitrary. Note that if µe1 + νv = 1,
then
[µe1 + νv, µe1 + νv]′ = µ2 + ν 2[v, v] = µ2 + ν 2 = 1,
which, in span{e1, v}, is the equation of the ellipse with semiaxes e1 and v. As
the restrictions of [., .]′ and [., .] to span{e2, e3} are clearly equal, we obtain that
[., .] = [., .]′, which, in particular, implies that [z, xf + xk] = [z, xk].
(cid:3)
By Lemma 2, we have that (1) of Theorem 1 holds. Thus, it remains to show
that (3) also holds. Without loss of generality, let us assume that k = 1, and that
λ1 = 1. Then every x ∈ M can be decomposed as x = x1 + y1 with x1 ∈ E1 and
y1 ∈ E−1. Hence,
[A(x1 + y1), A(x1 + y1)] = [x1 + y1, A2(x1 + y1)] = [x1 + y1, x1 + y1],
and thus, A is an isometry.
i1
λi
A = id on ¯Ei. Hence,
Finally, we show that if (1), (2) and (3) holds, then A is adjoint abelian. Let
x = Pk
i=1 xi and y = Pk
yi with xi, yi ∈ ¯Ei. Assume, first, that λk 6= 0, which
means that A is invertible. Note that ¯Ei is an invariant subspace of A for every
A(cid:17)−1
value of i, and that (cid:16) 1
Xi=1
λi(cid:20)xi,
λi(cid:20) 1
Ayi(cid:21) =
Axi, yi(cid:21) =
Xi=1
and the assertion follows. If A is not invertible, we may apply a slighly modified
argument.
λi"xi,(cid:18) 1
λi
A(cid:19)−1
[xi, Ayi] = [x, Ay],
yi# =
[Ax, y] =
[Axi, yi] =
k
Xi=1
λi
k
= 1
λi
k
1
λi
k
Xi=1
k
Xi=1
Acknowledgements. The author is indebted to ´A. G. Horv´ath for proposing
this project and for his helpful comments.
References
[1] S. S. Dragomir, Semi-inner-products and Applications Nova Science Publishers, Hauppauge
NY, 2004.
[2] R. J. Fleming and J. E. Jamison, Hermitian and adjoint abelian operators on certain Banach
spaces, Pacific J. Math. 52 (1974), 67-84.
[3] R. J. Fleming and J. E. Jamison, Adjoint abelian operators on Lp and C(K), Trans. Amer.
Math. Soc. 217 (1976), 87-98.
[4] J. R. Giles, Classes of semi-inner product spaces, Trans. Amer. Math. Soc. 129 (1967),
436-446.
[5] ´A. G. Horv´ath, Semi-indefinite product and generalized Minkowski spaces, J. Geom. Phys.
60 (2010), 1190-1208.
[6] D. O. Koehler, A note on some operator theory in certain semi-inner-product spaces, Proc.
Amer. Math. Soc. 30(2) (1971), 363-366.
[7] D. Kohler and P. Rosenthal, On isometries of normed linear spaces, Studia Math. 36 (1970),
213-216.
[8] G. Lumer, Semi-inner-product spaces, Trans. Amer. Math. Soc. 100 (1961), 29-43.
8
Z. L ´ANGI
[9] H. Martini and M. Spirova, Reflections in strictly convex Minkowski planes, Aequationes
Math. 78 (2009), 71-85.
[10] R. E. Megginson, An Introduction to Banach Space Theory, Graduate texts in mathematics
183, Springer-Verlag, New York, 1998.
[11] A. W. Roberts, D. E. Varberg, Convex functions, Pure and Applied Mathematics 57, Aca-
demic Press [A subsidiary of Harcourt Brace Jovanovich, Publishers], New York-London,
1973.
Zsolt L´angi, Dept. of Geometry, Budapest University of Technology, Budapest,
Egry J´ozsef u. 1., Hungary, 1111
E-mail address: [email protected]
|
1008.3441 | 1 | 1008 | 2010-08-20T06:13:58 | On bounds of Tsallis relative entropy and an inequality for generalized skew information | [
"math.FA"
] | Quantum entropy and skew information play important roles in quantum information science. They are defined by the trace of the positive operators so that the trace inequalities often have important roles to develop the mathematical theory in quantum information science. In this paper, we study some properties for information quantities in quantum system through trace inequalities. Especially, we give upper bounds and lower bounds of Tsallis relative entropy, which is a one-parameter extension of the relative entropy in quantum system. In addition, we compare the known bounds and the new bounds, for both upper and lower bounds, respectively. We also give an inequality for generalized skew information by introducing a generalized correlation measure. | math.FA | math |
On bounds of Tsallis relative entropy and an inequality for
generalized skew information
Shigeru Furuichi1∗
1Department of Computer Science and System Analysis,
College of Humanities and Sciences, Nihon University,
3-25-40, Sakurajyousui, Setagaya-ku, Tokyo, 156-8550, Japan
Abstract. Quantum entropy and skew information play important roles in quantum in-
formation science. They are defined by the trace of the positive operators so that the trace
inequalities often have important roles to develop the mathematical theory in quantum infor-
mation science. In this paper, we study some properties for information quantities in quantum
system through trace inequalities. Especially, we give upper bounds and lower bounds of Tsallis
relative entropy, which is a one-parameter extension of the relative entropy in quantum sys-
tem. In addition, we compare the known bounds and the new bounds, for both upper and lower
bounds, respectively. We also give an inequality for generalized skew information by introducing
a generalized correlation measure.
Keywords : Tsallis relative entropy, trace inequality, skew information, correlation mea-
sure and positive operators
2000 Mathematics Subject Classification : 15A45, 47A63 and 94A17
1
Introduction
Two important theorems in quantum information theory have been proved in [42] and [26, 43].
In the paper [42], it was shown the relation between quantum entropy (von Neumann entropy)
[50] and source coding theorem in quantum system (non-commutative system). In the papers
[26, 43], it was also shown the relation between the Holevo bound, which is considered to
be mutual information in quantum information theory, and coding theorem for the classical-
quantum channel. Especially, in these papers [26, 42, 43], the role of von Neumann entropy
and Holevo bound in quantum information were clarified so that quantum information theory
has been progressed for around fifteen years [35, 38]. Before such developments of quantum
information, von Nuemann entropy and related entropies such as relative entropy [49] and mutual
entropy [36] were studied in both direction from physics and mathematics [37, 51]. The study
for a generalization on information entropy in classical system (commutative system) has a long
history and we have many literatures [7, 39, 45]. See also [1, 8] and references therein. As
for one of the generalizations of entropy, we have studied the Tsallis entropy and the Tsallis
relative entropy in quantum system [11, 12, 13, 14, 15, 16, 17]. In the present paper, we study
on the bounds of the Tsallis relative entropy which is a one-parameter extension of the Umegaki
relative entropy.
As one of the mathematical studies on the topics about entropy theory, skew information
[52, 53] and its concavity problem are famous. The concavity problem for skew information
∗E-mail:[email protected]
1
was generalized by F.J.Dyson, and it was proven by E.H.Lieb in [30].
It is also known that
skew information presents the degree of noncommutativity between a certain quantum state
represented by a density operator ρ (which is a positive operator with an unit trace) and an
observable represented by self-adjoint operator H, therefore an uncertainty relation using skew
information has been studied in [11, 18, 28, 31, 34, 32, 54]. See [24, 40, 41] for the original
uncertainty relations which were represented by the trace inequalities and they show the uncer-
tainty principle which is a fundamental concept in quantum mechanical physics. In the present
paper, we give a trace inequality for a generalized skew information by introducing a generalized
correlation measure.
2 Upper bounds of Tsallis relative entropy
1
Firstly we give some notation. We denote the generalized exponential function by expν(x) ≡
ν if 1+νx > 0, otherwise it is undefined and its inverse function (generalized logarithmic
(1+νx)
function) by lnν x ≡ xν
−1
, for ν ∈ (0, 1] and x ≥ 0. The functions expν (x) and lnν x converge
ν
to ex and log x as ν → 0, respectively. Note that the definition of the generalized logarithmic
In this paper, we define the gen-
function lnν(X) for a positive operator X is well-defined.
ν , if
eralized exponential function expν(X) for a positive operator X by expν(X) ≡ (I + νX)
1
T rh(I + νX)
1
νi ∈ R.
The Tsallis relative entropy in quantum system (noncommutative system) is defined for
the positive operators in the following manner. For the study of entropies from mathematical
viewpoint, the condition such as an unit trace is often relaxed. See p.274 of [5] or [11, 12]. See
also [45, 46, 47, 48] for the original Tsallis entropy and its advances in statistical physics.
Definition 2.1 The Tsallis relative entropy is defined by
Dν (XY ) ≡
Tr[X − X 1−νY ν ]
ν
= Tr[X 1−ν (lnν X − lnν Y )]
for positive operators X, Y and ν ∈ (0, 1].
We have the following proposition, which gives an upper bound of the Tsallis relative entropy.
Proposition 2.2 ([11]) For positive operators X, Y and ν ∈ (0, 1], the following inequality
holds.
Dν(XY ) ≤ −T rhX lnν(cid:16)X−1/2Y X−1/2(cid:17)i .
(1)
The further upper bound of the right hand side was given by T.Furuta in [21], with the
generalized Kantorovich constant:
Dν (XY ) ≤ −T rhX lnν(cid:16)X−1/2Y X−1/2(cid:17)i ≤(cid:18) 1 − K(ν, h)
(cid:19) T r[X]1−νT r[Y ]ν + Dν (XY ),
where X 1/2 lnν(cid:0)X−1/2Y X−1/2(cid:1) X 1/2 is often called the Tsallis relative operator entropy for
positive operators X, Y and K(ν, h) is the generalized Kantorovich constant defined for ν ∈ R
and h ∈ (0, ∞) with h 6= 1:
ν
K(ν, h) ≡
(hν − h)
(ν − 1)(h − 1)(cid:18) (ν − 1)
ν
hν − 1
(hν − h)(cid:19)ν
.
2
m ) = K(−1, M
We note that K(2, M
4mM , which is often called the Kantorovich constant.
See [22, 23] for details. Since we have K(0, h) = K(1, h) = 1 and dK(ν,h)
ν=0 = − log S(h),
the above inequalities recover the following inequalities (the first inequality below was originally
given in [25]) :
m ) = (M +m)2
dν
U (XY ) ≤ −T rhX log(cid:16)X−1/2Y X−1/2(cid:17)i ≤ T r[X] log S(h) + U (XY ),
in the limit ν → 0. Where U (XY ) ≡ T r[X (log X − log Y )] is the relative entropy introduced
by Umegaki in [49], X 1/2 log(cid:0)X−1/2Y X−1/2(cid:1) X 1/2 is the relative operator entropy introduced
is the Specht's ratio [44], where h ≡ M1M2
m1m2
in [4, 10] and S(h) ≡ h
> 1 for 0 < m1I ≤
h−1
1
1
h−1
e log h
X ≤ M1I and 0 < m2I ≤ Y ≤ M2I. (See [21] for details.)
We also have the following proposition.
Proposition 2.3 For positive operators X, Y and ν ∈ (0, 1], the following inequality holds.
Dν(XY ) ≤
T r[(X − Y )+]
ν
,
(2)
where A+ ≡ 1
2 (A + A) and A ≡ (A∗A)1/2 for any operator A.
Proof: It immediately follows from the trace inequality proven by K.M.R.Audenaert et.al.
in [2]:
T r[AsB1−s] ≥
1
2
T r[A + B − A − B]
(3)
for positive operators A, B and s ∈ [0, 1].
From the above two propositions, we have two different upper bounds for the Tsallis relative
entropy. It is quite natural to consider that we have the following inequality for positive operators
X and Y and ν ∈ (0, 1]:
which is equivalent to
− T rhX lnν(cid:16)X−1/2Y X−1/2(cid:17)i ≤
T r[(X − Y )+]
ν
,
T r [X♯ν Y ] ≥
1
2
T r[X + Y − X − Y ],
(4)
(5)
where X♯ν Y ≡ X 1/2(X−1/2Y X−1/2)νX 1/2 is ν-power mean. Here we note that we have
T r[X 1−νY ν] ≥ T r [X♯ν Y ]
from the inequality (1). However we have the counter-example for the inequality (5) in the
following. We take ν = 1/2 and
X =(cid:18) 10 7
7
5 (cid:19) , Y =(cid:18) 16 6
3 (cid:19) .
6
Then we have
T r [X♯ν Y ] −
1
2
T r[X + Y − X − Y ] ≃ −0.510619.
Therefore the inequality (4) does not hold in general. That is, we can conclude that neither the
inequality (1) nor the inequality (2) is uniformly better than the other.
3
3 Lower bounds of Tsallis relative entropy
We firstly note that the Tsallis relative entropy defined in Definition 2.1 is not always nonneg-
ative. (If we impose on the condition such as an unit trace for two positive operators X ad Y ,
then the Tsallis relative entropy has a nonnegativity.) Therefore it is natural to have an interest
in the lower bound of the Tsallis relative entropy defined for two positive operators. We have
the following proposition, which gives a lower bound of the Tsallis relative entropy.
Proposition 3.1 ([11]) For positive operators X, Y and ν ∈ (0, 1], the generalized Bogolivbov
inequality holds.
Dν (XY ) ≥
T r[X] − (T r[X])1−ν (T r[Y ])ν
ν
(6)
(7)
If we take the limit ν → 0, Proposition 3.1 recovers the original Peierls-Bogoliubov inequality
[3, 27]:
U (XY ) ≥ T r [X (log T r[X] − log T r[Y ])] .
We also easily find that the right hand side in the inequality (6) is nonnegative, if we have
the relation such that T r[X] ≥ T r[Y ]. We also have the following lower bound for the Tsallis
relative entropy.
Theorem 3.2 For positive operators X, Y and ν ∈ (0, 1], if we have I ≤ Y ≤ X, then we have
the following inequality
Dν(XY ) ≥ T rhX 1−ν lnν(cid:16)Y −1/2XY −1/2(cid:17)i .
It is notable that the condition X ≥ Y assures the nonnegativity of the right hand side in
the inequality (7). To prove Theorem 3.2, we use the following lemmas.
Lemma 3.3 ([19]) For positive operators X, Y and ν ∈ (0, 1], we have
T r[expν (X + Y )] ≤ T r[expν(X) expν(Y )].
(8)
Here we give a slightly different version of a variational expression for the Tsallis relative
entropy. It can be proven by the similar way to Theorem 2.1 in [13]. However we give the proof
for the convenience of readers as to be a self-contained article.
Lemma 3.4 For any ν ∈ (0, 1] and any d ∈ [0, ∞), we have the following relations.
(i) If A and Y are positive operators, then we have
d lnν(cid:18) T r[expν(A + lnν Y )]
d
(cid:19) = max(cid:8)T r[X 1−ν A] − Dν (XY ) : X ≥ 0, T r[X] = d(cid:9) .
(ii) If X and B are positive operators with T r[X] = d, then
Dν (X expν(B)) = max(cid:26)T r[X 1−νA] − d lnν(cid:18) T r[expν(A + B)]
d
(cid:19) : A ≥ 0(cid:27) .
Proof: For the case of ν = 1, it is trivial so that we assume ν ∈ (0, 1). Since we have
a
x = 0 for ν ∈ (0, 1), it is also trivial for the case of X = 0, thus we assume X 6= 0.
limx→0 x lnν
4
(1) We define
Fν (X) ≡ T r[X 1−νA] − Dν(XY )
for a positive operator X with T r[X] = d < ∞. If we take the Schatten decomposition X =
P∞j=1 µjEj, where all Ej, (j = 1, 2, · · · , ∞) are projections of rank one with P∞j=1 Ej = I
and µj ≥ 0, (j = 1, 2, · · · , ∞) with P∞j=1 µj = d, then we rewrite
j T r[EjA] +
µ1−ν
j T r[EjY ν] −
1
ν
1
ν
µjT r[Ej](cid:27) .
Then we have
∞
Fν
Xj=1
∂2
∂µ2
j
Xj=1(cid:26)µ1−ν
∞
µjEj
=
µjEj
∞
Xj=1
Fν
= −ν(1 − ν)µ−ν−1
j
T r(cid:20)Ej(cid:18)A +
1
ν
Y ν(cid:19)(cid:21) ≤ 0,
which means Fν is concave function. Thus we find Fν (X) attains its maximum at a
certain positive operator X0 with T r[X0] = d. Then for any self-adjoint operators S with
T r[S] = 0 (since for any t ∈ R, T r[X0 + tS] = d which is a condition on the positive
operator defined on the domain of the function Fν), there exists a positive operator X0
such that
0 =
Fν(X0 + tS)t=0 = (1 − ν)T r(cid:20)S(X−ν
d
dt
ν X−ν
0 Y ν = cI for c ∈ R. Thus we have
0 A +
1
ν
X−ν
0 Y ν)(cid:21) ,
so that X−ν
0 A + 1
X0 = d
expν(A + lnν Y )
T r[expν (A + lnν Y )]
.
ν and satisfying the condition T r[X0] = d. By the formulae lnν
y
x =
by putting c = 1
lnν y + yν lnν
1
x and lnν
1
x = −x−ν lnν x, we have
Fν (X0) = d1−ν T r[{expν(A + lnν Y )}1−ν A]
−d1−ν T r" {expν(A + lnν Y )}1−ν
= d1−ν T r[{expν(A + lnν Y )}1−ν A]
T r[expν(A + lnν Y )]1−ν
T r[expν (A + lnν Y )]1−ν lnν expν(A + lnν Y )
1
d T r[expν(A + lnν Y )]! − lnν Y!#
T r[expν(A + lnν Y )]1−ν {A + lnν Y
T r[expν(A + lnν Y )]1−ν − d1−ν T r" {expν(A + lnν Y )}1−ν
T r[expν(A + lnν(A + lnν Y ))](cid:19) − lnν Y(cid:27)(cid:21)
+ {expν(A + lnν Y )}ν lnν(cid:18)
d
= −d1−ν T r[expν(A + lnν Y )]ν lnν(cid:18)
d
T r[expν(A + lnν Y )](cid:19)
= d lnν
T r[expν(A + lnν Y )]
.
d
(2) It follows from (1) that the functional
g(A) ≡ d lnν(cid:18) T r[expν(A + B)]
d
(cid:19) ,
d ≡ T r[X] < ∞
5
defined on the set of all positive operator is convex, due to triangle inequality on max.
Now let A0 = lnν X − B, and define
Gν(A) ≡ T r[X 1−ν A] − d lnν(cid:18) T r[expν(A + B)]
d
(cid:19) ,
which is concave on the set of all positive operator. Then for any self-adjoint operators S,
there exists a positive operator A0 such that
d
dt
Gν(A0 + tS)t=0 = T r[X 1−νS] − d(cid:18) T r[X]
d (cid:19)ν−1 T r[S(I + ν lnν X)
d
= T r[X 1−νS] − T r[SX 1−ν ] = 0,
1−ν
ν ]
using the formulae d
attaines the maximum
dx lnν(x) = xν−1 and d
dx expν (x) = (1 + νx)
1
ν −1. Therefore Gν (A)
Gν (A0) = T r[X 1−ν(lnν X − B)] − d lnν(cid:18) T r[expν(lnν X − B + B)]
d
(cid:19)
d (cid:19)
= T r[X 1−ν(lnν X − B)] − d lnν(cid:18) T r[X]
= T r[X 1−ν(lnν X − lnν expν (B))]
= Dν(X expν(B)).
If we take d = 1 and the limit ν → 0, then Lemma 3.4 recovers Lemma 2.1 in [25] for posi-
tive operators A and B. It is notable that the original variational expressions for the Umegaki
relative entropy proved by F.Hiai and D.Petz in [25] holds for Hermitian matrices A and B. We
are now in a position to prove Theorem 3.2.
Proof of Theorem 3.2: Putting B = lnν Y and A = lnν Y −1/2XY −1/2 in (ii) of Lemma 3.4
under the assumption of I ≤ Y ≤ X which assures A ≥ 0 and B ≥ 0, and then using Lemma
3.3, we have
Dν(XY ) = Dν (X expν(lnν Y ))
= Dν (X expν(B))
≥ T r[X 1−ν A] − T r[X] lnν(cid:18) T r[expν(A + B)]
(cid:19)
≥ T r[X 1−ν A] − T r[X] lnν(cid:18) T r[expν(A) expν(B)]
= T r[X 1−ν lnν Y −1/2XY −1/2] − T r[X] lnν T r[Y −1/2XY −1/2Y ]
T r[X]
T r[X]
T r[X]
(cid:19)
!
= T r[X 1−ν lnν Y −1/2XY −1/2].
Remark 3.5
(I) The trace inequality (7) is equivalent to the following trace inequality:
T rhX 1−νnX ν − Y ν + I −(cid:16)Y −1/2XY −1/2(cid:17)νoi ≥ 0.
6
Therefore, if the following matrix inequality:
X ν − Y ν + I −(cid:16)Y −1/2XY −1/2(cid:17)ν
≥ 0
(9)
holds, then the trace inequality (7) immediately holds. However the matrix inequality (9)
does not hold in general, since we have the following counter-examples.
(i) If we take ν = 1 and
X =(cid:18) 2 1
1 4 (cid:19) , Y =(cid:18) 1 0
0 2 (cid:19) ,
satisfying the condition I ≤ Y ≤ X (which is the assumption of Proposition 3.2),
then one of the eigenvalues of the Hermitian matrix X − Y + I − Y −1/2XY −1/2 takes
a negative value.
(ii) If we take ν = 1 and
X =
1
9(cid:18) 2 1
1 5 (cid:19) , Y =
1
3(cid:18) 1 0
0 2 (cid:19) ,
satisfying the condition X ≤ Y ≤ I, then one of the eigenvalues of the Hermitian
matrix X − Y + I − Y −1/2XY −1/2 takes a negative value.
(II) Our next concern moves to the assumption of Proposition 3.2. We easily find that a
counter-example for the trace inequality (7), in the case that our assumption I ≤ Y ≤ X
is not satisfied. For example, if we take ν = 1 and
X =
1
15(cid:18) 10 −3
10 (cid:19) , Y =
−3
1
10(cid:18) 1 1
1 2 (cid:19) ,
which does not satisfy the assumption I ≤ Y ≤ X (but satisfy 0 < Y ≤ X ≤ I), then
T r[X − Y + I − Y −1/2XY −1/2] ≃ −20.9667.
Thus the inequality (7) does not hold in general for arbitrary positive operators X and Y .
From (I) and (II), we may claim that Proposition 3.2 is not a trivial result.
Closing this section, we give a comment on the comparison of two lower bounds for the
Tsallis relative entropy. Under the condition I ≤ Y ≤ X, we may have a conjecture such as
T rhX 1−ν lnν(cid:16)Y −1/2XY −1/2(cid:17)i ≥
T r[X] − (T r[X])1−ν (T r[Y ])ν
ν
(10)
for positive operators X and Y and ν ∈ (0, 1]. The inequality (10) is equivalent to the following
inequality
T rhX 1−ν(cid:16)Y −1/2XY −1/2(cid:17)νi + (T r[X])1−ν (T r[Y ])ν ≥ T r[X 1−ν] + T r[X].
Here we take two positive definite matrices
(11)
X =(cid:18) 10 5
5 (cid:19) , Y =(cid:18) 1 0
0 2 (cid:19) ,
5
satisfying the condition I ≤ Y ≤ X. Then for ν = 0.1, we have
T rhX 1−ν(cid:16)Y −1/2XY −1/2(cid:17)νi + (T r[X])1−ν (T r[Y ])ν −(cid:0)T r[X 1−ν] + T r[X](cid:1) ≃ 0.508133.
7
For ν = 0.9, we also have
T rhX 1−ν(cid:16)Y −1/2XY −1/2(cid:17)νi + (T r[X])1−ν (T r[Y ])ν −(cid:0)T r[X 1−ν] + T r[X](cid:1) ≃ −1.1696.
Therefore the inequality (10) does not hold in general. That is, we can conclude that neither
the inequality (6) nor the inequality (7) is uniformly better than the other. This result supports
that our Theorem 3.2 is meaningful, in the sense of the comparison with Proposition 3.1.
4 An inequality for a generalized skew information
The uncertainty principle is a fundamental concept in quantum mechanical physics. It is repre-
sented by the famous Heisenberg uncertainty relation such as a trace inequality [24]:
Vρ(A)Vρ(B) ≥
T r[ρ[A, B]]2
1
4
(12)
for a quantum state ρ and two observables A and B. Where the variance for a quantum state ρ
and an observable H is defined by Vρ(H) ≡ T r[ρ (H − T r[ρH]I)2] = T r[ρH 2] − T r[ρH]2. The
further strong result was given by Schrodinger [41]:
Vρ(A)Vρ(B) − Re {Covρ(A, B)} 2 ≥
T r[ρ[A, B]]2,
1
4
(13)
where the covariance is defined by Covρ(A, B) ≡ T r[ρ (A − T r[ρA]I) (B − T r[ρB]I)]. Due to its
importance in quantum physics, the uncertainty relation has been studied by many researchers.
Especially, some important results have been studied in the relation to the skew information
representing a quantum uncertainty from the viewpoints of quantum information science. Here
we firstly review about it. As it has been shown in [29, 34, 55], we do not have the uncertainty
relation type inequality for the Wigner-Yanase skew information [52]:
Iρ(H) ≡
1
2
T rh(i[ρ1/2, H0])2i = T r[ρH 2] − T r[ρ1/2Hρ1/2H],
(14)
where H0 ≡ H − T r[ρH]I for a density operator ρ and an observable H. That is, the following
trace inequality did not hold in general [29, 34, 55]:
Iρ(A)Iρ(B) ≥
T r[ρ[A, B]]2
1
4
(15)
for a density operator ρ and observables A and B. Where [X, Y ] ≡ XY − Y X is a commutator.
The counter example was given as follows.
Counter-example 4.1 ([55]) We take
1
ρ =
4(cid:18) 3 0
then we have Iρ(A)Iρ(B) = (cid:16)1 −
0 1 (cid:19) , A =(cid:18) 0
2 (cid:17)2
does not hold in general.
and 1
√3
i
−i 0 (cid:19) , B =(cid:18) 0 1
0 (cid:19) ,
1
4 T r[ρ[A, B]]2 = 1
4 . Therefore the inequality (15)
As a one-parameter generalization, the Wigner-Yanase-Dyson skew information was defined
by
Iρ,α(H) ≡
1
2
T r(cid:2)(i[ρα, H0])(i[ρ1−α, H0])(cid:3)
= T r[ρH 2] − T r[ραHρ1−αH], α ∈ [0, 1]
(16)
8
where H0 ≡ H − T r[ρH]I for a density operator ρ and an observable H.
In [32], S.Luo introduced a new quantity such as
Uρ(H) ≡qVρ(H)2 − (Vρ(H) − Iρ(H))2
(17)
for a density operator ρ and an observable H. Then he succeeded to establish the uncertainty
relation type inequality as follows:
Uρ(A)Uρ(B) ≥
T r[ρ[A, B]]2
1
4
(18)
for a density operator ρ and observables A and B. He also introduced the quantity associated
to Wigner-Yanase skew information,
Jρ(H) ≡
1
2
T r(cid:20)(cid:16)inρ1/2, H0o(cid:17)2(cid:21) ,
where the anti-commutator is defined by {X, Y } ≡ XY + Y X for any operator X and Y . Then
2 (Iρ(H) + Jρ(H)). Therefore, Luo's
we have the relation Uρ(H) =pIρ(H)Jρ(H) and Vρ(H) = 1
inequality (18) refines Heisenberg's one (12), since Uρ(H) ≤ Vρ(H).
K.Yanagi recently gave the generalization of the inequality (18) as follows [54]:
Uρ,α(A)Uρ,α(B) ≥ α(1 − α)T r[ρ[A, B]]2
(19)
for α ∈ [0, 1], a density operator ρ and observables A and B. Where Uρ,α(H) was defined by
for a density operator ρ and an observable H.
Uρ,α(H) ≡qVρ(H)2 − (Vρ(H) − Iρ,α(H))2
(20)
In addition, quite recently, we gave the Schrodinger uncertainty relation for mixed states in
[18]:
Uρ(A)Uρ(B) − Re {Corrρ(A, B)} 2 ≥
T r[ρ[A, B]]2,
1
4
(21)
where the correlation measure is defined for arbitrary operators X and Y by
Corrρ(X, Y ) ≡ T r[ρX∗Y ] − T r[ρ1/2X∗ρ1/2Y ].
On the other hand, S.Luo showed the trace inequality representing the relation between the
original Wigner-Yanase skew information and the correlation measure in [32]:
Iρ(A)Iρ(B) ≥ Re {Corrρ(A, B)}2
(22)
for a density operator ρ and two observables A and B.
It is remarkable that, if a quantum state (density operator) ρ is a pure state (i.e., ρ2 = ρ),
then the inequality (22) recovers
Vρ(A)Vρ(B) ≥
T r[ρ {A, B}]2,
1
4
since Iρ(H) = Vρ(H) and Corrρ(A, B) = Covρ(A, B) if ρ is a pure state. Where the covariance
is defined by Covρ(A, B) ≡ T r[ρAB] − T r[ρA]T r[ρB].
In addition, defining a one-parameter extended correlation measure for α ∈ [0, 1] and arbi-
trary operators X and Y by
Corrρ,α(X, Y ) ≡ T r[ρX∗Y ] − T r[ραX∗ρ1−αY ].
9
we have the following inequality:
Iρ,α(A)Iρ,α(B) ≥ Re {Corrρ,α(A, B)}2 .
(23)
for a density operator ρ, two observables A, B and α ∈ [0, 1], putting ε = 0 in Theorem III.4 of
[55].
If we take α = 1
To give the further generalized trace inequality, we give the following definition.
2 , then the inequality (23) recovers the inequality (22).
Definition 4.2 Let f and g be the operator monotone functions. Let (f, g) be a monotonic pair.
Where (f, g) is called a monotonic pair if (f (a) − f (b))(g(a) − g(b)) ≥ 0 for any a, b ∈ D ⊂ R,
for two functions f and g on the domain D ⊂ R. For a density operator ρ and an observable
H, we define (f, g)-skew information Iρ,(f,g)(H) by
Iρ,(f,g)(H) ≡
1
2
T r [(i [f (ρ), H0]) (i [g(ρ), H0])] ,
(24)
where H0 ≡ H − T r[ρH]I. For a density operator ρ and any operators X, Y , we also define
(f, g)-correlation measure Corrρ,(f,g)(X, Y ) by
Corrρ,(f,g)(X, Y ) ≡ T r [f (ρ)g(ρ)X∗Y ] − T r [f (ρ)X∗g(ρ)Y ] .
(25)
Then we can prove the following theorem.
Theorem 4.3 For (f, g)-skew informations Iρ,(f,g)(A), Iρ,(f,g)(B) and (f, g)-correlation mea-
sure Corrρ,(f,g)(A, B), we have
Iρ,(f,g)(A)Iρ,(f,g)(B) ≥(cid:12)(cid:12)Re(cid:8)Corrρ,(f,g)(A, B)(cid:9)(cid:12)(cid:12)
2 .
(26)
Proof: For Corrρ,(f,g)(X, Y ), we have the following properties, that is, Corrρ,(f,g)(X, Y ) is
a sesquilinear form and Hermitian and Corrρ,(f,g)(A, A) has the nonnegativity for a self-adjoint
operator A, since (f, g) is a monotonic pair [6, 9]. Then for self-adjoint operators A and B, we
have for any t ∈ R
0 ≤ Corrρ,(f,g)(tA + B, tA + B)
= T r[f (ρ)g(ρ)(tA + B)(tA + B)] − T r[f (ρ)(tA + B)g(ρ)(tA + B)]
= (cid:0)T r[f (ρ)g(ρ)A2] − T r[f (ρ)Ag(ρ)A](cid:1) t2 + 2Re (T r[f (ρ)g(ρ)AB] − T r[f (ρ)Ag(ρ)B]) t
+(cid:0)T r[f (ρ)g(ρ)B2] − T r[f (ρ)Bg(ρ)B](cid:1) .
Thus we have
Re {T r[f (ρ)g(ρ)AB] − T r[f (ρ)Ag(ρ)B]}2
≤(cid:0)T r[f (ρ)g(ρ)A2] − T r[f (ρ)Ag(ρ)A](cid:1)(cid:0)T r[f (ρ)g(ρ)B2] − T r[f (ρ)Bg(ρ)B](cid:1) ,
which implies the theorem, since we have
Iρ,(f,g)(H) = T r[f (ρ)g(ρ)H 2] − T r[f (ρ)Hg(ρ)H].
Theorem 4.3 recovers the inequality (23), putting f (x) = xα and g(x) = x1−α for α ∈ [0, 1].
We also have the following corollary.
10
Corollary 4.4 For α ∈ [0, 1], a density operator ρ and two observables A, B, we have
where a one-parameter extended correlation measure is defined by
Kρ,α(A)Kρ,α(B) ≥(cid:12)(cid:12)(cid:12)
(cid:19)2
RenCorr(K)
ρ,α (A, B)o(cid:12)(cid:12)(cid:12)
X∗Y# − T r(cid:20)(cid:18) ρα + ρ1−α
2
Corr(K)
ρ,α (X, Y ) ≡ T r"(cid:18) ρα + ρ1−α
2
2
,
(27)
(cid:19) X∗(cid:18) ρα + ρ1−α
2
(cid:19) Y(cid:21)
for α ∈ [0, 1], a density operator ρ and any operators X and Y , and a one-parameter extended
Wigner-Yanase skew information [20] is defined by Kρ,α(H) ≡ Corr(K)
ρ,α (H, H) for α ∈ [0, 1], a
density operator ρ and a self-adjoint operator H.
Proof: Put f (x) = g(x) = xα+x1−α
2
in Theorem 4.3.
5 Conclusion
As we have seen, we have studied the properties of the fundamental information measure in
quantum system, namely the Tsallis relative entropy and the generalized skew information
through the trace inequality with the mathematical tools in matrix analysis. Closing conclusion,
we give the following quantity which generalizes the Tsallis relative entropy and the Wigner-
Yanase-Dyson skew information. For two positive operators X and Y , and a self-adjoint operator
H, we define
Lt(X, Y ; H) ≡ T r[XH 2] − T r[X tHY 1−tH],
t ∈ [0, 1].
Then using Lt(X, Y ; H), the Wigner-Yanase-Dyson skew information can be rewritten by
Iα(ρ, H) ≡ Lα(ρ, ρ; H) = T r[ρH 2] − T r[ραHρ1−αH].
Also the Tsallis relative entropy can be rewritten by
Dν (XY ) ≡
1
ν
L1−ν(X, Y ; I) =
T r[X − X 1−νY ν ]
ν
,
ν ∈ (0, 1].
It may be important to study the mathematical properties of the quantity Lt(X, Y ; H) in
the future, since it covers both the Tsallis relative entropy (one-parameter extended relative
entropy) and the Wigner-Yanase-Dyson skew information as special cases.
Acknowledgement
The author was supported in part by the Japanese Ministry of Education, Science, Sports and
Culture, Grant-in-Aid for Encouragement of Young Scientists (B), 20740067
References
[1] J.Acz´el and Z.Dar´oczy, On measures of information and their characterizations, Academic
Press, 1975.
[2] K.M.R.Audenaert
et.al., Discriminating
states:
The quantum chernoff bound,
Phys.Rev.Lett.,Vol.98(2007),160501.
11
[3] N.Bebiano, J. da Providencia, Jr. and R.Lemos, Matrix inequalities in statistical mechanics,
Linear Algebr. Appl. Vol.376(2004),pp.265-273.
[4] V.P.Belavkin and P.Staszewski, C∗-algebraic generalization of relative entropy and entropy,
Ann. Inst. Henri Poincar´e, Sec. A, Vol.37(1982), pp.51-58.
[5] R.Bhatia, Matrix Analysis, Springer, 1997.
[6] J.-C.Bourin, Some inequalities for norms on matrices and operators, Linear Alg.Appl.,
Vol.292(1999), pp.139-154.
[7] I. Csisz´ar, Axiomatic characterizations of information measures, Entropy, Vol.10(2008),
pp.261-273.
[8] B.Ebanks, P.Sahoo and W.Sander, Characterizations of information measure, World Scien-
tific, 1998.
[9] J.I.Fujii, A trace inequality arising from quantum information theory, Linear Alg. Appl.,
Vol.400(2005), pp.141-146.
[10] J.I.Fujii and E.Kamei, Relative operator entropy in noncommutative information theory,
Math. Japonica, Vol.34(1989), pp.341-348.
[11] S.Furuichi, K.Yanagi and K.Kuriyama, Fundamental properties of Tsallis relative entropy,
J.Math.Phys., Vol.45(2004), pp.4868-4877.
[12] S.Furuichi, K.Yanagi and K.Kuriyama, A note on operator inequalities of Tsallis relative
opeartor entropy, Linear Alg. Appl.,Vol.407(2005), pp.19-31.
[13] S.Furuichi, Trace inequalities in nonextensive statistical mechanics, Linear Alg.Appl.,
Vol.418(2006), pp.821-827.
[14] S.Furuichi, A note on a parametrically extended entanglement-measure due to Tsallis rel-
ative entropy, INFORMATION, Vol.9,No.6,pp.837-844(2006).
[15] S.Furuichi, Tsallis entropies and their theorems, properties and applications, pp.1-86, in the
book "Aspects of Optical Sciences and Quantum Information", Research Signpost, Edited
by M.Abdel-Aty, 2007.
[16] S.Furuichi, Matrix trace inequalities on Tsallis entropies, Journal of inequalities in pure
and applied mathematics, Vol.9(2008), Issue 1, Article 1, 7pp.
[17] S.Furuichi, A review of the mathematical properties of the Tsallis entropies, AIP Conference
Proceedings: Current Themes in Engineering Science 2007: Selected Presentations at the
World Congress on Engineering-2007,Vol.1045(1),pp.11 - 20,(September 9, 2008).
[18] S.Furuichi, Schrodinger uncertainty relation with Wigner-Yanase skew information, to ap-
pear in Phys.Rev.A.
[19] S.Furuichi and M.Lin, A matrix trace inequality and its application, Liniear Alg. Appl.
Vol.433(2010), pp.1324-1328.
[20] S.Furuichi, K.Yanagi and K.Kuriyama, Trace inequalities on a generalized Wigner-Yanase
skew information, J.Math.Anal.Appl.,Vol.356(2009),pp.179-185.
12
[21] T.Furuta, Reverse inequalities involving two relative operator entropies and two relative
entropies, Linear Alg.Appl., Vol. 403(2005), pp. 24-30.
[22] T.Furuta, Invitation to linear operators, Taylor and Francis, 2001.
[23] T.Furuta, J.Mi´ci´c Hot, J.Recari´c and Y.Seo, Mond-Pecari´c method in operator inequalities,
Element, Zagreb, 2005.
[24] W.Heisenberg, Uber den anschaulichen Inhalt der quantummechanischen Kinematik und
Mechanik, Zeitschrift fur Physik, Vol.43(1927), pp.172-198.
[25] F.Hiai and D.Petz, The Golden-Thompson trace inequality is complemented, Linear Alg.
Appl.,Vol.181(1993), pp.153-185.
[26] A.S.Holevo, The capacity of quantum channel with general signal states, IEEE.Trans.IT,
Vol.44(1998), pp.269-273.
[27] K.Huang, Statistical Mechanics, John Wiley and Sons, New York, 1987.
[28] C.K.Ko, Comments on conjectures of trace inequalities on a generalized Wigner-Yanase
skew information, J.Math.Anal.Appl.,Vol.369(2010),pp.164-167.
[29] H.Kosaki, Matrix trace inequality related to uncertainty principle, International Journal of
Mathematics,Vol.16(2005),pp.629-646.
[30] E.H.Lieb, Convex trace functions and the Wigner-Yanase-Dyson conjecture, Adv.
Math.,Vol.11(1973),pp.267-288.
[31] M.Lin, On a conjecture of S.Furuichi et al., personal communications.
[32] S.Luo, Heisenberg uncertainty relation for mixed states, Phys.Rev.A,Vol.72(2005),042110.
[33] S. Luo, Quantum versus classical uncertainty, Theor. Math. Phys. Vol.143(2005),pp.681-
688.
[34] S.Luo and Q.Zhang, On skew information,
Information Theory,
Vol.50(2004),pp.1778-1782, and Correction to "On skew information", IEEE Trans. In-
formation Theory, Vol.51(2005),p.4432.
IEEE Trans.
[35] M.A.Nielsen and I.L.Chuang, Quantum computation and quantum information, Cambridge
University Press, 2000.
[36] M.Ohya, On compound state and mutual information in quantum information theory,
IEEE.Trans.Information Theory, Vol.29(1983), pp.770-774.
[37] M.Ohya and D.Petz, Quantum entropy and its use, Springer, 1993.
[38] D.Petz, Quantum information theory and quantum statistics, Springer, 2008.
[39] A.R´enyi, On measures of entropy and information, in Proc. 4th Berkeley Symp., Math-
ematical and Statistical Probability, Berkeley, CA: Univ. Calif. Press, Vol. 1(1961), pp.
547-561.
[40] H.P.Robertson, The uncertainty principle, Phys.Rev.,Vol.34(1929),pp.163-164.
[41] E.Schrodinger, About Heisenberg uncertainty relation, Proc.Prussian Acad.Sci.,Phys.Math.
Section,Vol.XIX(1930),p.293.
13
[42] B.Schumacher, Quantum coding, Phys.Rev.A, Vol.51(1995), pp.2738-2747.
[43] B.Schumacher and M.D.Westmoreland, Sending classical information via noisy quantum
channel, Phys.Rev.A, Vol.56(1997), pp.131-138.
[44] W.Specht, Zer Theorie der elementaren Mittel, Math.Z.,Vol.74(1960), pp.91-98.
[45] C. Tsallis, Possible generalization of Bolzmann-Gibbs statistics, J.Stat. Phys., Vol.
52(1988), pp. 479-487.
[46] C. Tsallis et al. In: S. Abe and Y. Okamoto, Editors, Nonextensive Statistical Mechan-
ics and its Applications, Springer, 2001. See also the comprehensive list of references at
http://tsallis.cat.cbpf.br/biblio.htm.
[47] C.Tsallis, Introduction to Nonextensive Statistical Mechanics: Approaching a Complex
World, Springer, 2009.
[48] C. Tsallis, Entropy, in Encyclopedia of Complexity and Systems Science, Springer, Berlin,
2009.
[49] H.Umegaki, Conditional expectation in an operator algebra, IV (entropy and informa-
tion),Kodai Math.Sem.Rep., Vol.14(1962), pp.59-85.
[50] J.von
Neumann,Thermodynamik
quantenmechanischer
Gesamtheiten,Gottinger
Nachrichen,pp.273-291(1927).
[51] A.Wehrl, General properties of entropy, Rev.Mod.Phys.,Vol.50(1978), pp.221-260.
[52] E.P.Wigner and M.M.Yanase, Information content of distribution, Proc.Nat.Acad.Sci.
U.S.A., Vol.49(1963), pp.910-918.
[53] E.P.Wigner and M.M.Yanase, On the positive semidefinite nature of certain matrix expres-
sion, Canad. J. Math.,Vol.16(1964),pp.397-406.
[54] K.Yanagi,
Uncertainty
relation
on Wigner-Yanase-Dyson
skew information,
J.Math.Anal.Appl.,Vol.365(2010),pp.12-18.
[55] K.Yanagi, S.Furuichi and K.Kuriyama, A generalized skew information and uncertainty
relation, IEEE Trans. Information Theory, Vol.51(2005),pp.4401-4404.
[56] K.Yanagi, S.Furuichi and K.Kuriyama, On generalized skew information, Proc. of 2005
Symp. on Appl. Func. Anal., pp.113-123.
14
|
1403.1731 | 3 | 1403 | 2016-04-28T09:55:56 | Hardy-Littlewood inequalities and Fourier multipliers on SU(2) | [
"math.FA"
] | In this paper we prove a noncommutative version of Hardy-Littlewood inequalities relating a function and its Fourier coefficients on the group $SU(2)$. As a consequence, we use it to obtain lower bounds for the $L^p-L^q$ norms of Fourier multipliers on the group $SU(2)$, for $1 < p \leq 2 \leq q < 1$. In addition, we give upper bounds of a similar form, analogous to the known results on the torus, but now in the noncommutative setting of $SU(2)$. | math.FA | math |
Hardy-Littlewood-Paley inequalities
and Fourier multipliers on SU(2)
Rauan Akylzhanov
Department of Mathematics
Imperial College London
180 Queen's Gate, London SW7 2AZ
United Kingdom
E-mail address [email protected]
Erlan Nurlustanov
Department of Mathematics
Moscow State University, Kazakh Branch
Astana, Kazakhstan
E-mail address [email protected]
Michael Ruzhansky
Department of Mathematics
Imperial College London
180 Queen's Gate, London SW7 2AZ
United Kingdom
E-mail address [email protected]
November 8, 2018
Abstract
In this paper we prove noncommutative versions of Hardy -- Little-
wood and Paley inequalities relating a function and its Fourier coef-
ficients on the group SU(2). As a consequence, we use it to obtain
The third author was supported by the EPSRC Grant EP/K039407/1.
2010 Mathematics Subject Classification: Primary 35G10; 35L30; Secondary 46F05;
Key words and phrases: Fourier multipliers, Hardy-Littlewood inequality, Paley in-
equality, noncommutative harmonic analysis.
1
2
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
lower bounds for the Lp -- Lq norms of Fourier multipliers on the group
SU(2), for 1 < p ≤ 2 ≤ q < ∞. In addition, we give upper bounds
of a similar form, analogous to the known results on the torus, but
now in the noncommutative setting of SU(2).
1
Introduction
Let Tn be the n-dimensional torus and let 1 < p ≤ q < ∞. A sequence
λ = {λk}k∈Zn of complex numbers is said to be a multiplier of trigonometric
Fourier series from Lp(Tn) to Lq(Tn) if the operator
Tλf (x) = Xk∈Zn
λkbf (k)eikx
is bounded from Lp(Tn) to Lq(Tn). We denote by mq
tipliers.
p the set of such mul-
Many problems in harmonic analysis and partial differential equations
can be reduced to the boundedness of multiplier transformations. There
arises a natural question of finding sufficient conditions for λ ∈ mp
p. The
topic of mq
p multipliers has been extensively researched. Using methods
such as the Littlewood-Paley decomposition and Calderon-Zygmund theory,
it is possible to prove Hormander-Mihlin type theorems, see e.g. Mihlin
[Mih57, Mih56], Hormander [Hor60], and later works.
Multipliers have been then analysed in a variety of different settings, see
e.g. Gaudry [Gau66], Cowling [Cow74], Vretare [Vre74]. The literature on
the spectral multipliers is too rich to be reviewed here, see e.g. a recent
paper [CKS11] and references therein. The same is true for multipliers on
locally compact abelian groups, see e.g. [Arh12], or for Fourier or spectral
multipliers on symmetric spaces, see e.g. [Ank90] or [CGM93], resp. We refer
to the above and to other papers for further references on the history of mq
p
multipliers on spaces of different types.
In this paper we are interested in questions for Fourier multipliers on
compact Lie groups, in which case the literature is much more sparse: in
the sequel we will make a more detailed review of the existing results. Thus,
in this paper we will be investigating several questions in the model case
of Fourier multipliers on the compact group SU(2). Although we will not
explore it in this paper, we note that there are links between multipliers on
SU(2) and those on the Heisenberg group, see Ricci and Rubin [RR86].
In general, most of the multiplier theorems imply that λ ∈ mp
p for all 1 <
p < ∞ at once. In [Ste70], Stein raised the question of finding more subtle
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
3
sufficient conditions for a multiplier to belong to some mp
p, p 6= 2, without
implying also that it belongs to all mp
p, 1 < p < ∞. In [NT00], Nursultanov
and Tleukhanova provided conditions on λ = {λk}k∈Z to belong to mq
for the range 1 < p ≤ 2 ≤ q < ∞. In particular, they established lower
and upper bounds for the norms of multiplier λ ∈ mq
p which depend on
parameters p and q. Thus, this provided a partial answer to Stein's question.
Let us recall this result in the case n = 1:
p
Theorem 1.1. Let 1 < p ≤ 2 ≤ q < ∞ and let M0 denote the set of all
finite arithmetic sequences in Z. Then the following inequalities hold:
1
q − 1
Q1+ 1
. kTλkLp→Lq . sup
k∈N
1
q − 1
k1+ 1
p
λ∗
m,
kXm=1
sup
Q∈M0
p (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xm∈Q
λm(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where λ∗
of elements in the arithmetic progression Q .
m is a non-increasing rearrangement of λm, and Q is the number
In this paper we study the noncommutative versions of this and other
related results. As a model case, we concentrate on analysing Fourier multi-
pliers between Lebesgue spaces on the group SU(2) of 2×2 unitary matrices
with determinant one. Sufficient conditions for Fourier multipliers on SU(2)
to be bounded on Lp-spaces have been analysed by Coifman-Weiss [CW71b]
and Coifman-de Guzman [CdG71], see also Chapter 5 in Coifman and Weiss'
book [CW71a], and are given in terms of the Clebsch-Gordan coefficients of
representations on the group SU(2). A more general perspective was pro-
vided in [RW13] where conditions on Fourier multipliers to be bounded on
Lp were obtained for general compact Lie groups, and Mihlin-Hormander
theorems on general compact Lie groups have been established in [RW15].
Results about spectral multipliers are more known, for functions of the
Laplacian (N. Weiss [Wei72] or Coifman and Weiss [CW74]), or of the sub-
Laplacian on SU(2), see Cowling and Sikora [CS01]. However, following
[CW71b, CW71a, RW13, RW15], here were are rather interested in Fourier
multipliers.
In this paper we obtain lower and upper estimates for the norms of
Fourier multipliers acting between Lp and Lq spaces on SU(2). These es-
timates explicitly depend on parameters p and q. Thus, this paper can be
regarded as a contribution to Stein's question in the noncommutative set-
ting of SU(2). At the same time we provide a noncommutative analogue of
Theorem 1.1. Briefly, let A be the Fourier multiplier on SU(2) given by
cAf (l) = σA(l)bf (l), for σA(l) ∈ C(2l+1)×(2l+1), l ∈
1
2
N0,
4
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
where we refer to Section 2 for definitions and notation related to the Fourier
analysis on SU(2). For such operators, in Theorem 3.1, for 1 < p ≤ 2 ≤ q <
∞, we give two lower bounds, one of which is of the form
(1.1)
sup
l∈ 1
N0
2
1
(2l + 1)1+ 1
q − 1
p
1
2l + 1 Tr σA(l) . kAkLp(SU(2))→Lq(SU(2)).
A related upper bound
(1.2)
kAkLp(SU(2))→Lq(SU(2)) . sup
s>0
will be given in Theorem 4.1.
s Xl∈ 1
2
N0
kσA(l)kop≥s
(2l + 1)2
1
p − 1
q
.
The proof of the lower bound is based on the new inequalities describ-
ing the relationship between the "size" of a function and the "size" of its
Fourier transform. These inequalities can be viewed as a noncommutative
SU(2)-version of the Hardy-Littlewood inequalities obtained by Hardy and
Littlewood in [HL27]. To explain this briefly, we recall that in [HL27], Hardy
and Littlewood have shown that for 1 < p ≤ 2 and f ∈ Lp(T), the following
inequality holds true:
(1.3)
Xm∈Z
(1 + m)p−2bf (m)p ≤ Kkfkp
Lp(T),
arguing this to be a suitable extension of the Plancherel identity to Lp-
spaces. While we refer to Section 1 and to Theorem 2.1 for more details on
this, our analogue for this is the inequality
(1.4) Xl∈ 1
2
N0
(2l + 1)(2l + 1)
5
2 (p−2)kbf (l)kp
HS ≤ ckfkp
Lp(SU(2)),
1 < p ≤ 2,
which for p = 2 gives the ordinary Plancherel identity on SU(2), see (2.1).
We refer to Theorem 2.2 for this and to Corollary 2.3 for the dual statement.
For p ≥ 2, the necessary conditions for a function to belong to Lp are usually
harder to obtain. In Theorem 2.8 we give such a result for 2 ≤ p < ∞ which
takes the form
(1.5)
Xl∈ 1
2
N0
(2l + 1)p−2 sup
k∈ 1
2
k≥l
N0
1
2k + 1(cid:12)(cid:12)(cid:12)Trbf (k)(cid:12)(cid:12)(cid:12)
p
≤ ckfkp
Lp(SU(2)),
2 ≤ p < ∞.
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
5
In turn, this gives a noncommutative analogue to the known similar result
on the circle (which we recall in Theorem 2.7). Similar to (1.1), the averaged
trace appears also in (1.5) -- it is the usual trace divided by the number of
diagonal elements in the matrix.
In [Hor60] Hormander proved a Paley-type inequality for the Fourier
transform on RN . In this paper we obtain an analogue of this inequality on
the group SU(2).
The results on the group SU(2) are usually quite important since, in
view of the resolved Poincar´e conjecture, they provide information about
corresponding transformations on general closed simply-connected three-
dimensional manifolds (see [RT10] for a more detailed outline of such rela-
tions). In our context, they give explicit versions of known results on the
circle T or on the torus Tn, in the simplest noncommutative setting of SU(2).
At the same time, we note that some results of this paper can be extended
to Fourier multipliers on general compact Lie groups. However, such analysis
requires a more abstract approach, and will appear elsewhere.
The paper is organised as follows. In Section 2 we fix the notation for the
representation theory of SU(2) and formulate estimates relating functions
with its Fourier coefficients: the SU(2)-version of the Hardy -- Littlewood and
Paley inequalities and further extensions. In Section 3 we formulate and
prove the lower bounds for operator norms of Fourier multipliers, and in
Section 4 the upper bounds. Our proofs are based on inequalities from
Section 2. In Section 5 we complete the proofs of the results presented
in previous sections.
We shall use the symbol C to denote various positive constants, and Cp,q
for constants which may depend only on indices p and q. We shall write
x . y for the relation x ≤ Cy, and write x ∼= y if x . y and y . x.
The authors would like to thank V´eronique Fischer for useful remarks.
2 Hardy-Littlewood and Paley inequalities
on SU(2)
The aim of this section is to discuss necessary conditions and sufficient con-
ditions for the Lp(SU(2))-integrability of a function by means of its Fourier
coefficients. The main results of this section are Theorems 2.2, 2.4 and 2.8.
These results will provide a noncommutative version of known results of
this type on the circle T. The proofs of most of the results of this Section
are given in Section 5.
6
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
First, let us fix the notation concerning the representations of the com-
pact Lie group SU(2). There are different types of notation in the literature
for the appearing objects - we will follow the notation of Vilenkin [Vil68],
as well as that in [RT10, RT13]. Let us identify z = (z1, z2) ∈ C1×2, and
let C[z1, z2] be the space of two-variable polynomials f : C2 → C. Consider
mappings
tl : SU(2) → GL(Vl),
(tl(u)f )(z) = f (zu),
2
where l ∈ 1
N0 is called the quantum number, N0 = N∪{0}, and where Vl is
the (2l + 1)-dimensional subspace of C[z1, z2] containing the homogeneous
polynomials of order 2l ∈ N0, i.e.
Vl = {f ∈ C[z1, z2] : f (z1, z2) =
2lXk=0
The unitary dual of SU(2) is
akzk
1 z2l−k
2
,
{ak}2l
k=0 ⊂ C}.
\SU(2) ∼= {tl ∈ Hom(SU(2), U(2l + 1)) : l ∈
1
2
N0},
where U(d) ⊂ Cd×d is the unitary matrix group, and matrix components
mn ∈ C ∞(SU(2)) can be written as products of exponentials and Legendre-
tl
Jacobi functions, see Vilenkin [Vil68]. It is also customary to let the indices
m, n to range from −l to l, equi-spaced with step one. We define the Fourier
transform on SU(2) by
bf (l) := ZSU(2)
f (u) = Xl∈ 1
N0
2
f (u)tl(u)∗ du,
(2l + 1) Trbf (l)tl(u).
with the inverse Fourier transform (Fourier series) given by
The Peter-Weyl theorem on SU(2) implies, in particular, that this pair of
transforms are inverse to each other and that the Plancherel identity
(2.1)
L2(SU(2)) = Xl∈ 1
kfk2
(2l + 1)kbf (l)k2
holds true for all f ∈ L2(SU(2)). Here kbf (l)k2
N0
2
ℓ2(SU(2))
HS =: kbfk2
HS = Trbf (l)bf (l)∗ denotes the
Hilbert-Schmidt norm of matrices. For more details on the Fourier transform
on SU(2) and on arbitrary compact Lie groups, and for subsequent Fourier
and operator analysis we can refer to [RT10].
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
7
There are different ways to compare the "sizes" of f and bf . Apart from
the Plancherel's identity (2.1), there are other important relations, such
as the Hausdorff-Young or the Riesz-Fischer theorems. However, such esti-
mates usually require the change of the exponent p in Lp-measurements of
f and bf . Our first results deal with comparing f and bf in the same scale
of Lp-measurements. Let us remark on the background of this problem. In
[HL27, Theorems 10 and 11], Hardy and Littlewood proved the following
generalisation of the Plancherel's identity.
Theorem 2.1 (Hardy -- Littlewood [HL27]). The following holds.
1. Let 1 < p ≤ 2. If f ∈ Lp(T), then
Xm∈Z
(1 + m)p−2bf (m)p ≤ Kpkfkp
Lp(T),
where Kp is a constant which depends only on p.
2. Let 2 ≤ p < ∞. If {bf (m)}m∈Z is a sequence of complex numbers such
that
Xm∈Z
(1 + m)p−2bf (m)p < ∞,
(2.2)
(2.3)
then there is a function f ∈ Lp(T) with Fourier coefficients given by
bf (m), and
kfkp
pXm∈Z
Lp(T) ≤ K ′
(1 + m)p−2bf (m)p.
Hewitt and Ross [HR74] generalised this theorem to the setting of com-
pact abelian groups. Now, we give an analogue of the Hardy -- Littlewood
Theorem 2.1 in the noncommutative setting of the compact group SU(2).
Theorem 2.2. If 1 < p ≤ 2 and f ∈ Lp(SU(2)), then we have
(2.4)
(2l + 1)
5
2
Lp(SU(2)).
HS ≤ cpkfkp
Xl∈ 1
2
N0
p−4kbf (l)kp
We can write this in the form more resembling the Plancherel identity,
namely, as
(2.5)
Xl∈ 1
2
N0
(2l + 1)(2l + 1)
5
2 (p−2)kbf (l)kp
HS ≤ cpkfkp
Lp(SU(2)),
providing a resemblance to both (2.2) and (2.1). By duality, we obtain
8
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
Corollary 2.3. If 2 ≤ p < ∞ and Pl∈ 1
Lp(SU(2)) ≤ cp Xl∈ 1
kfkp
f ∈ Lp(SU(2)) and we have
(2.6)
N0
2
2
(2l + 1)
N0
HS < ∞, then
5
2 p−4kbf (l)kp
(2l + 1)
HS.
5
2 p−4kbf (l)kp
For p = 2, both of these statements reduce to the Plancherel identity
(2.1).
In [Hor60] Hormander proved a Paley-type inequality for the Fourier
transform on RN . We now give an analogue of this inequality on the group
SU(2).
Theorem 2.4. Let 1 < p ≤ 2. Suppose {σ(l)}l∈ 1
matrices σ(l) ∈ C(2l+1)×(2l+1) such that
(2.7)
2
Kσ := sup
s>0
(2l + 1)2 < ∞.
s Xl∈ 1
2
N0
kσ(l)kop≥s
N0 is a sequence of complex
Then we have
(2.8)
Xl∈ 1
2
N0
(2l + 1)p( 2
p − 1
2 )kbf (l)kp
HSkσ(l)k2−p
op . Kσ
2−pkfkp
Lp(SU(2)).
It will be useful to recall the spaces ℓp(\SU(2)) on the discrete unitary dual
\SU(2). For general compact Lie groups these spaces have been introduced
and studied in [RT10, Section 10.3]. In the particular case of SU(2), for a
sequence of complex matrices σ(l) ∈ C(2l+1)×(2l+1) they can be defined by
the finiteness of the norms
(2.9)
and
(2.10)
kσkℓp(\SU(2)) :=Xl∈ 1
2
N0
(2l + 1)p( 2
p − 1
2 )kσ(l)kp
HS
1
p
,
1 ≤ p < ∞,
kσkℓ∞(\SU(2)) := sup
l∈ 1
2
N0
(2l + 1)− 1
2kσ(l)kHS.
Among other things, it was shown in [RT10, Section 10.3] that these spaces
the Hausdorff-Young inequality
(2.11)
are interpolation spaces, they satisfy the duality property and, with σ = bf ,
Xl∈ 1
HS
2 )kbf (l)kp′
≡ kbfkℓp′ (\SU(2))
. kfkLp(SU(2)), 1 ≤ p ≤ 2.
(2l + 1)p′( 2
p′ − 1
1
p′
N0
2
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
9
Further, we recall a result on the interpolation of weighted spaces from
[BL76]:
Theorem 2.5 (Interpolation of weighted spaces). Let us write dµ0(x) =
ω0(x)dµ(x), dµ1(x) = ω1(x)dµ(x), and write Lp(ω) = Lp(ωdµ) for the
weight ω. Suppose that 0 < p0, p1 < ∞. Then
(Lp0(ω0), Lp1(ω1))θ,p = Lp(ω),
where 0 < θ < 1, 1
p = 1−θ
p0
From this we obtain:
+ θ
p1
, and ω = w
p 1−θ
p0
0 w
p θ
p1
1
.
Corollary 2.6. Let 1 < p ≤ b ≤ p′ < ∞. If {σ(l)}l∈ 1
(2.7) with constant Kσ, then we have
2
N0 satisfies condition
(2.12) Xl∈ 1
2
N0
(2l + 1)b( 2
b − 1
2 )(cid:18)kbf (l)kHSkσ(l)k
1
b
1
b − 1
p′
op (cid:19)b
. (Kσ)
1
b − 1
p′ kfkLp(SU(2)).
This reduces to (2.11) when b = p′ and to (2.8) when b = p.
Proof. We consider a sub-linear operator A which takes a function f to its
Fourier transform bf (l) divided by √2l + 1 i.e.
f 7→ Af =:( bf (l)
√2l + 1)l∈ 1
2
,
N0
where
bf (l) = ZSU(2)
f (u)tl(u)∗ u ∈ C(2l+1)×(2l+1), l ∈
1
2
N0.
The statement follows from Theorem 2.5 if we regard the left-hand sides
of inequalities (2.8) and (2.11) as an kAfkLp-norm in a weighted sequence
space over 1
and
2
w1(l) = (2l + 1)2, l ∈ 1
N0 with the weights given by w0(l) = (2l + 1)2kσ(l)k2−p
N0.
op
2
Coming back to the Hardy -- Littlewood Theorem 2.1, we see that the
convergence of the series (2.3) is a sufficient condition for f to belong to
Lp(T), for p ≥ 2. However, this condition is not necessary. Hence, there
arises the question of finding necessary conditions for f to belong to Lp. In
other words, there is the problem of finding lower estimates for kfkLp in
terms of the series of the form (2.3). Such result on Lp(T) was obtained by
Nursultanov and can be stated as follows.
10
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
Theorem 2.7 ([Nur98a]). If 2 < p < ∞ and f ∈ Lp(T), then we have
(2.13)
∞Xk=1
kp−2sup
e∈M
e≥k
p
1
e(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xm∈ebf (m)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where M is the set of all finite arithmetic progressions in Z.
≤ Ckfkp
Lp(T),
We now present a (noncommutative) version of this result on the group
SU(2).
Theorem 2.8. If 2 < p < ∞ and f ∈ Lp(SU(2)), then we have
p
1
N0
k∈ 1
2
k≥l
(2l + 1)p−2 sup
2k + 1(cid:12)(cid:12)(cid:12)Trbf (k)(cid:12)(cid:12)(cid:12)
f (x)g(x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ZSU(2)
f (x)g(x) dx = Xl∈ 1
(2l + 1) Trbf (l)bg(l)∗.
kfkLp(SU(2)) = sup
g∈Lp′
Lp′ =1
kgk
N0
.
2
ZSU(2)
Using Plancherel's identity (2.1), we get
(2.14) Xl∈ 1
2
N0
≤ ckfkp
Lp(SU(2)).
For completeness, we give a simple argument for Corollary 2.3.
Proof of Corollary 2.3. The application of the duality of Lp spaces yields
It is easy to see that
5
2 − 4
p + 5
2 − 4
p′ ,
(2l + 1) = (2l + 1)
(cid:12)(cid:12)(cid:12)Trbf (l)bg(l)∗(cid:12)(cid:12)(cid:12) ≤ kbf (l)kHSkbg(l)kHS.
Using these inequalities, applying Holder inequality, for any g ∈ Lp′ with
kgkLp′ = 1, we have
5
2 − 4
2
2
N0
N0
(2l + 1)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(2l + 1) Trbf (l)bg(l)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xl∈ 1
≤ Xl∈ 1
≤Xl∈ 1
pXl∈ 1
HS
2 p−4kbf (l)kp
≤Xl∈ 1
(2l + 1)
N0
N0
5
1
2
2
2
N0
5
2 − 4
5
(2l + 1)
1
p′
pkbf (l)kHS(2l + 1)
p′ kbg(l)kHS
HS
2 p′−4kbg(l)kp′
HS
2 p−4kbf (l)kp
kgkLp′ ,
1
p
5
(2l + 1)
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
11
where we used Theorem 2.2 in the last line. Thus, we have just proved that
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ZSU(2)
f (x)g(x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xl∈ 1
2
N0
(2l + 1) Trbf (l)bg(l)∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤Xl∈ 1
N0
2
(2l + 1)
1
p
5
HS
2 p−4kbf (l)kp
kgkLp′ .
Taking supremum over all g ∈ Lp′(SU(2)), we get (2.6). This proves Corol-
lary 2.3.
3 Lower bounds for Fourier multipliers on
SU(2)
Let A : C ∞(SU(2)) → D′(SU(2)) be a continuous linear operator. Here we
are concerned with left-invariant operators which means that A◦ τg = τg ◦ A
for the left-translation τgf (x) = f (g−1x). Using the Schwartz kernel theo-
rem and the Fourier inversion formula one can prove that the left-invariant
continuous operator A can be written as a Fourier multiplier, namely, as
cAf (l) = σA(l)bf (l),
for the symbol σA(l) ∈ C(2l+1)×(2l+1). It follows from the Fourier inversion
formula that we can write this also as
(3.1)
Af (u) = Xl∈ 1
2
N0
(2l + 1) Tr tl(u)σA(l)bf (l),
where the symbol σA(l) is given by
σA(l) = tl(e)∗Atl(e) = Atl(e),
where e is an identity matrix in SU(2), and (Atl)mk = A(tl
mk) is defined
component-wise, for −l ≤ m, n ≤ l. We refer to operators in these equivalent
forms as (noncommutative) Fourier multipliers. The class of these operators
on SU(2) and their Lp-boundedness was investigated in [CW71b, CW71a],
and on general compact Lie groups in [RW13]. In particular, these authors
proved Hormander -- Mikhlin type multiplier theorems in those settings, giv-
ing sufficient condition for the Lp-boundedness in terms of symbols. These
12
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
conditions guarantee that the operator is of weak (1,1)-type which, com-
bined with a simple L2-boundedness statement, implies the boundedness
on Lp for all 1 < p < ∞.
For a general (non-invariant) operator A, its matrix symbol σA(u, l) will
also depend on u. Such quantization (3.1) has been consistently developed
in [RT10] and [RT13]. We note that the Lp-boundedness results in [RW13]
also cover such non-invariant operators.
For a noncommutative Fourier multiplier A we will write A ∈ M q
p (SU(2))
if A extends to a bounded operator from Lp(SU(2)) to Lq(SU(2)). We in-
troduce a norm k · k on M q
p (SU(2)) by setting
kAkM q
p := kAkLp→Lq .
Thus, we are concerned with the question of what assumptions on the sym-
p . The sufficient conditions on σA for A ∈ M p
bol σA guarantee that A ∈ M q
p
were investigated in [RW13]. The aim of this section is to give a necessary
condition on σA for A ∈ M q
p , for 1 < p ≤ 2 ≤ q < ∞.
Suppose that 1 < p ≤ 2 ≤ q < ∞ and that A : Lp(SU(2)) → Lq(SU(2)) is
a Fourier multiplier. The Plancherel identity (2.1) implies that the operator
A is bounded from L2(SU(2)) to L2(SU(2)) if and only if supl kσA(l)kop < ∞.
Different other function spaces on the unitary dual have been discussed in
[RT10]. Following Stein, we search for more subtle conditions on the symbols
of noncommutative Fourier multipliers ensuring their Lp − Lq boundedness,
and we now prove a lower estimate which depends explicitly on p and q.
Theorem 3.1. Let 1 < p ≤ 2 ≤ q < ∞ and let A be a left-invariant
operator on SU(2) such that A ∈ M q
n∈{−l,...,+l}σA(l)nn
p (SU(2)). Then we have
min
(3.2)
sup
l∈ 1
N0
2
. kAkLp(SU(2))→Lq(SU(2)),
(3.3)
sup
l∈ 1
N0
2
. kAkLp(SU(2))→Lq(SU(2)).
q
q
1
(2l + 1)
p′ + 1
Tr σA(l)
p′ + 1
(2l + 1)1+ 1
One can see a similarity between (3.2), (3.3) and (1.1) as
(3.4)
1
(2l + 1)
sup
l∈ 1
N0
2
1
p′ + 1
q
1
2l + 1Tr σA(l) . kAkLp(SU(2))→Lq(SU(2)).
We also note that estimates (3.2) and (3.3) can not be immediately com-
pared because the value of the trace in (3.3) depends on the signs of the
diagonal entries of σA(l).
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
13
Proof of Theorem 3.1. In [GT80] it was proven that for any l ∈ 1
exists a basis for tl ∈ \SU(2) and a diagonal matrix coefficient tl
some n, −l ≤ n ≤ l), such that
N0 there
2
nn (i.e. for
(3.5)
nnkLp(SU(2)) ∼=
ktl
1
(2l + 1)
.
1
p
p (SU(2)). Let us fix an arbitrary l0 ∈ 1
Now, we use this result to establish a lower bound for the norm of A ∈
M q
N0 and the corresponding diag-
onal element tl0
nn. We consider fl0(g) such that its matrix-valued Fourier
coefficient
2
(3.6)
cfl0(l) = diag(0, . . . , 1, 0, . . .)δl
l0
has only one non-zero diagonal coefficient 1 at the nth diagonal entry. Then
by the Fourier inversion formula we get fl0(g) = (2l0 + 1)tl0
nn(g). By defini-
tion, we get
kAkLp→Lq = sup
f 6=0
kPl∈ 1
2
N0
kfkLp(SU(2))
(2l + 1) Tr tl(u)σA(l)bf (l)kLq(SU(2))
kPl∈ 1
(2l + 1) Tr tl(u)σA(l)cfl0(l)kLq(SU(2))
kfl0kLp(SU(2))
N0
2
.
≥
Recalling (3.6), we get
kAkLp→Lq & k(2l0 + 1) Tr tl0(g)σA(l0)cfl0(l)kLq(SU(2))
kfl0kLp(SU(2))
.
Setting h(g) := (2l0 + 1) Tr tl0(g)σA(l0)cfl0(l0), we havebh(l) = 0 for l 6= l0,
andbh(l0) = σA(l0)cfl0(l0). Consequently, we get
l > l0,
sup
k∈ 1
N0
2
k≥l
1
2k + 1(cid:12)(cid:12)(cid:12)Trbh(k)(cid:12)(cid:12)(cid:12) =(0,
1
2l0+1 σA(l0)nn , 1 ≤ l ≤ l0.
Using this, Theorem 2.8 and (3.5), we have
l0Xl=1
(2l + 1)q−2(cid:18) 1
2l0 + 1 σA(l0)nn(cid:19)q! 1
q
(2l0 + 1)1− 1
p
,
kAkLp→Lq &
14
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
where l0 is an arbitrary fixed half-integer. Direct calculation now shows that
l0Xl=1
(2l + 1)q−2(cid:18) 1
2l0 + 1 σA(l0)nn(cid:19)q! 1
q
(2l0 + 1)1− 1
p
=
1
2l0 + 1 σA(l0)nn
=
1
2l0 + 1 σA(l0)nn
(2l0 + 1)1− 1
(2l0 + 1)1− 1
q
q
(2l + 1)q−2! 1
l0Xl=1
(2l0 + 1)1− 1
p ∼= σA(l0)nn
p′ + 1
(2l0 + 1)
1
p
q
.
Taking infimum over all n ∈ {−l0,−l0 +1, . . . , l0−1, l0} and then supremum
over all half-integers, we have
kAkLp→Lq & sup
l∈ 1
2
N0
min
n∈{−l,...,+l}σA(l)nn
(2l + 1)
1
p′ + 1
q
.
This proves estimate (3.2). Now, we will prove estimate (3.3). Let us fix some
l0 ∈ 1
N0 and consider now fl0(u) := (2l0 + 1)χl0(u), where χl0(u) = Tr tl0(u)
is the character of the representation tl0. Then, in particular, we have
2
(3.7)
cfl0(l) =(I2l+1,
0,
l = l0,
l 6= l0,
where I2l+1 ∈ C(2l+1)×(2l+1) is the identity matrix. Using the Weyl character
formula, we can write
χl0(u) =
eikt,
l0Xk=−l0
0
0
characters are central. Further, the application of the Weyl integral formula
yields
e−it(cid:19) v. The value of χl0(u) does not depend on v since
where u = v−1(cid:18)eit
kfl0kLp(SU (2)) = (2l0+1)kχl0kLp(SU (2)) = (2l0+1)
2π
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2πZ0
l0Xk=−l0
It is clear that(cid:12)(cid:12)(cid:12)ei(−l0−1)tPl0
ei(k+l0+1)t(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)P2l0+1
k=1 eikt(cid:12)(cid:12)(cid:12). We call D2l0+1(t) :=
P2l0+1
k=1 eikt the Dirichlet kernel. Then, we apply [Nur98a, Corollary 4] to the
Dirichlet kernel D2l0+1(t), to get
eikt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2 sin2 t
k=−l0
dt
1
p
p
.
(3.8)
kχl0kLp(SU(2)) . kD2l0+1kLp(0,2π) ∼= (2l0 + 1)1− 1
p .
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
15
By definition, we get
kAkLp→Lq = sup
f 6=0
kPl∈ 1
2
N0
kfkLp(SU(2))
(2l + 1) Tr tl(u)σA(l)bf (l)kLq(SU(2))
kPl∈ 1
(2l + 1) Tr tl(u)σA(l)cfl0(l)kLq(SU(2))
kfl0kLp(SU(2))
N0
2
.
≥
Recalling (3.7), we obtain
kAkLp→Lq & k(2l0 + 1) Tr tl0(g)σA(l0)kLq(SU(2))
kfl0kLp(SU(2))
.
Setting h(g) := (2l0 + 1) Tr tl0(g)σA(l0), we have bh(l) = 0 for l 6= l0, and
bh(l0) = σA(l0). Consequently, we get
2k + 1(cid:12)(cid:12)(cid:12)Trbh(k)(cid:12)(cid:12)(cid:12) =(0,
l > l0,
1 ≤ l ≤ l0.
2l0+1 Tr σA(l0) ,
sup
k∈ 1
N0
2
k≥l
1
1
Using this and Theorem 2.8, we have
l0Xl=1
kAkLp→Lq &
(2l + 1)q−2(cid:18) 1
2l0 + 1 Tr σA(l0)(cid:19)q! 1
q
(2l0 + 1)(2l0 + 1)1− 1
p
,
where l0 is an arbitrary fixed half-integer. Direct calculation shows that
l0Xl=1
(2l + 1)q−2(cid:18) 1
2l0 + 1 Tr σA(l0)(cid:19)q! 1
q
(2l0 + 1)(2l0 + 1)1− 1
1
p
=
2l0 + 1 Tr σA(l0)
=
1
2l0 + 1 Tr σA(l0)
(2l0 + 1)1− 1
q
(2l0 + 1)(2l0 + 1)1− 1
q
(2l + 1)q−2! 1
l0Xl=1
(2l0 + 1)(2l0 + 1)1− 1
p ∼= Tr σA(l0)
p′ + 1
(2l0 + 1)1+ 1
q
p
.
Taking supremum over all half-integers, we have
kAkLp→Lq & sup
l∈ 1
2
N0
Tr σA(l)
(2l + 1)1+ 1
p′ + 1
q
.
This proves the estimate (3.3)
16
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
4 Upper bounds for Fourier multipliers on
SU(2)
In this section we give a noncommutative SU(2) analogue of the upper
bound for Fourier multipliers, analogous to the one on the circle T in The-
orem 1.1 (see also [Nur98b, NT11] for the circle case).
Theorem 4.1. If 1 < p ≤ 2 ≤ q < ∞ and A is a left-invariant operator on
SU(2), then we have
Proof. Since A is a left-invariant operator, it acts on f via the multipication
(4.1)
kAkLp(SU(2))→Lq(SU(2)) . sup
s>0
of bf by the symbol σA
(4.2)
where
1
p − 1
q
.
(2l + 1)2
kσA(l)kop>s
2
N0
s Xl∈ 1
cAf (π) = σA(π)bf (π),
σA(π) = π(x)∗Aπ(x)(cid:12)(cid:12)x=e
.
Let us first assume that p ≤ q′. Since q′ ≤ 2, for f ∈ C ∞(SU(2)) the
Hausdorff-Young inequality gives
(4.3)
(2l + 1)2− q′
kAfkLq(SU(2)) ≤ kcAfkℓq′ (\SU(2)) = kσAbfkℓq′ (\SU(2))
HS
2 kσA(l)bf (l)kq′
HS
opkbf (l)kq′
2 kσA(l)kq′
= Xl∈\SU(2)
≤ Xl∈\SU(2)
(2l + 1)2− q′
1
q′
1
q′
.
The case q′ ≤ (p′)′ can be reduced to the case p ≤ q′ as follows. The appli-
cation of Theorem 4.2 with G = SU(2) and µ = {Haar measure on SU(2)}
yields
(4.4)
kAkLp(SU(2))→Lq(SU(2)) = kA∗kLq′ (SU(2))→Lp′ (SU(2)).
A(l)
The symbol σA∗(l) of the adjoint operator A∗ equals to σ∗
(4.5)
σA∗(l) = σ∗
A(l),
1
2
l ∈
N0,
(4.6)
(4.7)
sup
s>0
1
r
kσ(tl)kr
op>s
sup
s>0
q = 1
p′ = 1
p − 1
q′ − 1
s Xtl∈\SU(2)
Xl∈\SU(2)
(2l + 1)2
.
kAfkLq(SU(2)) .
(2l + 1)2
sr Xtl∈\SU(2)
(2l + 1)2
s Xtl∈\SU(2)
=
kσA(tl)kop>s
kσ(tl)kop>s
sup
s>0
sup
s>0
=
1
r
sup
s>0
s Xtl∈\SU(2)
kσ(tl)kr
Further, it can be easily checked that
(2l + 1)2− q′
2 kσA(l)kq′
1
q′
HS
opkbf (l)kq′
kfkLp(SU(2)),
f ∈ Lp(SU(2)),
1
r
op>s
kfkLp(SU(2)).
(2l + 1)2
(2l + 1)2
s Xtl∈\SU(2)
(2l + 1)2
s Xtl∈\SU(2)
= sup
s>0
kσA(tl)kop>s
kσA(tl)kop>s
1
r
1
r
1
r
1
r
.
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
17
and its operator norm kσA∗(l)kop equals to kσA(l)kop. Now, we are in a
position to apply Corollary 2.6. Set 1
q . We observe that with σ(tl) :=
opI2l+1, l ∈ 1
N0 and b = q′, the assumptions of Corollary 2.6 are
kσA(tl)kr
satisfied and we obtain
p − 1
r = 1
2
in view of 1
r . Thus, for 1 < p ≤ 2 ≤ q < ∞, we obtain
This completes the proof.
For the completness, we give a short proof of Theorem 4.2 used in the
proof.
Theorem 4.2. Let (X, µ) be a measure space and 1 < p, q < ∞. Then we
have
(4.8)
where A∗ : Lq′(X, µ) → Lp′(X, µ) is the adjoint of A.
kAkLp(X,µ)→Lq(X,µ) = kA∗kLq′ (X,µ)→Lp′ (X,µ),
18
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
Proof of Theorem 4.2. Let f ∈ Lp∩ L2 and g ∈ Lq′ ∩ L2. By Holder inequal-
ity, we have
(4.9)
(Af, g)L2 = (A∗g, f )L2 ≤ kA∗gkLp′kfkLp ≤ kA∗kLq′ →Lp′kgkLq′kfkLp.
Thus, we get
(4.10)
kAkLp→Lq ≤ kA∗kLq′ →Lp′ .
Analogously, we show that
(4.11)
kA∗kLq′ →Lp′ ≤ kAkLp→Lq .
The combination of (4.10) and (4.11) yields
This completes the proof.
kAkLp→Lq = kA∗kLq′ →Lp′ .
5 Proofs of Theorems from Section 2
Proof of Theorem 2.4. Let µ give measure kσ(tl)k2
N0 to the
set consisting of the single point {tl}, tl ∈ \SU(2), and measure zero to a set
op(2l + 1)2, l ∈ 1
2
which does not contain any of these points, i.e.
µ{tl} := kσ(tl)k2
op(2l + 1)2.
(5.1)
We define the space Lp(\SU(2), µ), 1 ≤ p < ∞, as the space of complex (or
real) sequences a = {al}l∈ 1
N0 such that
2
alpkσ(tl)k2
We will show that the sub-linear operator
kakLp(\SU(2),µ) :=Xl∈ 1
A : Lp(SU(2)) ∋ f 7→ Af =(
is well-defined and bounded from Lp(SU(2)) to Lp(\SU(2), µ) for 1 < p ≤ 2.
op(2l + 1)2
√2l + 1kσ(tl)kop)tl∈\SU(2)
kbf (tl)kHS
∈ Lp(\SU(2), µ)
< ∞.
N0
2
1
p
In other words, we claim that we have the estimate
(5.2)
kAfkLp(\SU(2),µ) = Xtl∈\SU(2)
√2l + 1kσ(tl)kop!p
kbf (tl)kHS
kσ(tl)k2
1
p
op(2l + 1)2
2−p
p
σ
. K
kfkLp(SU(2)),
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
19
which would give (2.8) and where we set Kσ := sups>0 s Ptl∈\SU(2)
kσ(tl)kop≥s
(2l + 1)2.
We will show that A is of weak type (2,2) and of weak-type (1,1). For
definition and discussions we refer to Section 6 where we give definitions of
weak-type, formulate and prove Marcinkiewicz interpolation Theorem 6.1.
More precisely, with the distribution function ν as in Theorem 6.1, we show
that
(5.3)
ν\SU(2)(y; Af ) ≤ (cid:18) M2kfkL2(SU(2))
y
(5.4)
ν\SU(2)(y; Af ) ≤
M1kfkL1(SU(2))
y
(cid:19)2
with norm M2 = 1,
with norm M1 = Kσ.
Then (5.2) would follow by Marcinkiewicz interpolation Theorem 6.1. Now,
to show (5.3), using Plancherel's identity (2.1), we get
y2ν\SU(2)(y; Af ) ≤ kAfk2
= Xtl∈\SU(2)
Lp(\SU(2),µ)
:= Xtl∈\SU(2)
(2l + 1)kbf (tl)k2
√2l + 1kσ(tl)kop!2
kbf (tl)kHS
HS = kbfk2
= kfk2
ℓ2(\SU(2))
kσ(tl)k2
op(2l+1)2
L2(SU(2)).
Thus, A is of type (2,2) with norm M2 ≤ 1. Further, we show that A is of
weak-type (1,1) with norm M1 = C; more precisely, we show that
> y} . Kσ kfkL1(SU(2))
y
.
op(2l + 1)2 taken over
> y. From the definition of the
(5.5)
ν\SU(2){tl ∈ \SU(2) :
The left-hand side here is the weighted sumPkσ(tl)k2
those tl ∈ \SU(2) for which
Fourier transform it follows that
kbf (tl)kHS
√2l + 1kσ(tl)kop
kbf (tl)kHS
√2l + 1kσ(tl)kop
√2l + 1kfkL1(SU(2)).
kbf (tl)kHS ≤
kbf (tl)kHS
√2l + 1kσ(tl)kop ≤ kfkL1(SU(2))
kσ(tl)kop
.
y <
Therefore, we have
Using this, we get
(tl ∈ \SU(2) :
kbf (tl)kHS
√2l + 1kσ(tl)kop
> y) ⊂(cid:26)tl ∈ \SU(2) : kfkL1(SU(2))
kσ(tl)kop
> y(cid:27)
20
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
for any y > 0. Consequently,
µ(tl ∈ \SU(2) :
kbf (tl)kHS
√2l + 1kσ(tl)kop
Setting v :=
kf kL1(SU(2))
y
, we get
> y) ≤ µ(cid:26)tl ∈ \SU(2) : kfkL1(SU(2))
kσ(tl)kop
> y(cid:27) .
(5.6) µ(tl ∈ \SU(2) :
kbf (tl)kHS
√2l + 1kσ(tl)kop
> y) ≤ Xtl∈\SU(2)
kσ(tl)kop≤v
kσ(tl)k2
op(2l+1)2.
We claim that
(5.7)
In fact, we have
Xtl∈\SU(2)
kσ(tl)kop≤v
Xtl∈\SU(2)
kσ(tl)kop≤v
kσ(tl)k2
op(2l + 1)2 . Kσv.
kσ(tl)k2
op(2l + 1)2 = Xtl∈\SU(2)
kσ(tl)kop≤v
(2l + 1)2
We can interchange sum and integration to get
kσ(tl)k2
op
Z0
dτ.
Xtl∈\SU(2)
kσ(tl)kop≤v
(2l + 1)
kσ(tl)k2
op
Z0
dτ =
v2Z0
dτ Xtl∈\SU(2)
1
2 ≤kσ(tl)kop≤v
τ
(2l + 1)2.
Further, we make a substitution τ = s2, yielding
v2Z0
dτ Xtl∈\SU(2)
1
2 ≤kσ(tl)kop≤v
(2l + 1)2 = 2
vZ0
τ
Since
(2l + 1)2
s≤kσ(tl)kop≤v
s ds Xtl∈\SU(2)
vZ0
≤ 2
s ds Xtl∈\SU(2)
s≤kσ(tl)kop
(2l + 1)2.
s Xtl∈\SU(2)
s≤kσ(tl)kop
(2l + 1)2 ≤ sup
s>0
s Xtl∈\SU(2)
s≤kσ(tl)kop
(2l + 1)2 =: Kσ
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
21
is finite by the definition of Kσ, we have
2
vZ0
s ds Xtl∈\SU(2)
s≤kσ(tl)kop
(2l + 1)2 . Kσv.
This proves (5.7). We have just proved inequalities (5.3), (5.4). Then by
using
Marcinkiewicz' interpolation theorem (Theorem 6.1 from Section 6) with
p1 = 1, p2 = 2 and 1
2 we now obtain
p = 1 − θ + θ
Xl∈ 1
2
N0
√2l + 1kσ(π)kop!p
kbf (π)kHS
kσ(π)k2
1
p
op(2l + 1)2
= kAfkLp(\SU(2),µ)
2−p
p
σ
. K
kfkLp(SU(2)).
This completes the proof.
Now we prove the Hardy -- Littlewood type inequality given in Theorem
2.2.
Proof of Theorem 2.2. Let ν give measure
(2l + 1)4 to the set consisting of
the single point l, l = 0, 1
2 , 2, . . ., and measure zero to a set which does
not contain any of these points. We will show that the sub-linear operator
2 , 1, 3
1
T f := {(2l + 1)
5
2kbf (l)kHS}l∈ 1
2
N0
is well-defined and bounded from Lp(SU(2)) to Lp( 1
2
with
N0, ν) for 1 < p ≤ 2,
kT fkLp(\SU(2),ν) =Xl∈ 1
2
N0(cid:16)(2l + 1)
5
2kbf (l)kHS(cid:17)p
1
p
.
· (2l + 1)−4
This will prove Theorem 2.2. We first show that T is of type (2, 2) and weak
type (1, 1). Using Plancherel's identity (2.1), we get
kT fk2
Lp(\SU(2),ν)
= Xl∈ 1
2
N0
(2l + 1)
5p
2 −4kbf (l)k2
2
N0
HS = Xl∈ 1
= kbfk2
HS
(2l + 1)kbf (l)k2
= kfk2
ℓ2(\SU(2))
L2(SU(2)).
22
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
Thus, T is of type (2, 2).
Further, we show that T is of weak type (1,1); more precisely we show
that
(5.8)
ν{l ∈
1
2
N0 : (2l + 1)
The left-hand side here is the sumX 1
5
which (2l + 1)
follows that
5
2kbf (l)kHS > y} ≤
4
3
kfkL1(SU(2))
y
.
(2l + 1)4 taken over those l ∈ 1
2
N0 for
Therefore, we have
y < (2l + 1)
5
5
2 + 1
2kfkL1(SU(2)).
√2l + 1kfkL1(SU(2)).
2kbf (l)kHS > y. From the definition of the Fourier transform it
N0 : (2l + 1) >(cid:18) y
kbf (l)kHS ≤
2kbf (l)kHS ≤ (2l + 1)
2kbf (l)kHS > y(cid:27) ⊂(l ∈
3) .
2kbf (l)kHS > y(cid:27) ≤ ν(l ∈
kfkL1(cid:19) 1
kf kL1(SU(2))(cid:17) 1
3 . Now, we estimate ν(cid:8)l ∈ 1
N0 : (2l + 1) > w(cid:9).
ν(l ∈
3) =
N0 : (2l + 1) >(cid:18) y
∞Xn>w
N0 : (2l + 1) >(cid:18) y
3)
kfkL1(cid:19) 1
kfkL1(cid:19) 1
1
n4
1
2
1
2
1
2
.
2
y
Using this, we get
(cid:26)l ∈
1
2
N0 : (2l + 1)
5
N0 : (2l + 1)
5
1
2
ν(cid:26)l ∈
We set w := (cid:16)
By definition, we have
for any y > 0. Consequently,
In order to estimate this series, we introduce the following lemma.
Lemma 5.1. Suppose β > 1 and w > 0. Then we have
(5.9)
∞Xn>w
1
nβ ≤( β
β−1 ,
1
β−1
w ≤ 1,
wβ−1 , w > 1.
1
The proof is rather straightforward. Now, suppose w ≤ 1. Then applying
this lemma with β = 4, we have
1
n4 ≤
4
3
.
∞Xn>w
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
23
Since 1 ≤ 1
w3 , we obtain
1
n4 ≤
4
3 ≤
4
3
1
w3
.
3 , we finally obtain
∞Xn>w
kf kL1 (SU(2))(cid:17) 1
N0 : (2l + 1) >(cid:18) y
y
Recalling that w =(cid:16)
ν(l ∈
1
2
kfkL1(cid:19) 1
3) =
1
n4 ≤
4
3
kfkL1(SU(2))
y
.
∞Xn>w
Now, if w > 1, then we have
1
n4 ≤
1
3
1
w3 =
4
3
kfkL1
y
.
∞Xn>w
Finally, we get
ν(l ∈
1
2
This proves (5.8).
N0 : (2l + 1) >(cid:18) y
kfkL1(cid:19) 1
3) ≤
4
3
kfkL1(SU(2))
y
.
By Marcinkiewicz interpolation Theorem 6.1 with p1 = 1, p2 = 2, we
obtain
Xl∈ 1
2
N0
(2l + 1)
5
HS
2 p−4kbf (l)kp
1
p
= kT fkLp(\SU(2),ν) ≤ cpkfkLp(SU(2)).
This completes the proof of Theorem 2.2. .
Now we prove Theorem 2.8.
Proof of Theorem 2.8. We first simplify the expression for Trbf (k). By def-
inition, we have
bf (k) = ZSU(2)
f (u)T k(u)∗ du,
k ∈
1
2
N0,
where T k is a finite-dimensional representation of \SU(2) as in Section 2.
Using this, we get
(5.10)
Trbf (k) = ZSU(2)
f (u)χk(u) du,
24
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
where χk(u) = Tr T k(u), k ∈ 1
N0, where we changed the notation from
tk to T k to avoid confusing with the notation that follows. The characters
χk(u) are constant on the conjugacy classes of SU(2) and we follow [Vil68]
to describe these classes explicitly.
2
It is well known from linear algebra that any unitary unimodular matrix
1 , where u1 ∈ SU(2) and δ is a
u can be written in the form u = u1δu−1
diagonal matrix of the form
(5.11)
δ = e
it
2
0
2! ,
0
e− it
it
2 and 1
λ = e− it
2 are the eigenvalues of u. Moreover, among the
where λ = e
matrices equivalent to u there is only one other diagonal matrix, namely,
the matrix δ′ obtained from δ by interchanging the diagonal elements.
Hence, classes of conjugate elements in SU(2) are given by one parameter
t, varying in the limits −2π ≤ t ≤ 2π, where the parameters t and −t give
one and the same class. Therefore, we can regard the characters χk(u) as
functions of one variable t, which ranges from 0 to 2π.
The special unitary group SU(2) is isomorphic to the group of unit
quaternions. Hence, the parameter t has a simple geometrical meaning -
it is equal to angle of rotation which corresponds to the matrix u.
Let us now derive an explicit expression for the χk(u) as function of t. It
was shown e.g. in [RT10] that T k(δ) is a diagonal matrix with the numbers
e−int, −k ≤ n ≤ k on its principal diagonal.
1 . Since characters are constant on conjugacy classes of
Let u = u1δu−1
elements, we get
(5.12)
χk(u) = χk(δ) = Tr T k(δ) =
eint.
kXn=−k
It is natural to express the invariant integral over SU(2) in (5.10) in new
parameters, one of which is t.
Since special unitary group SU(2) is diffeomorphic to the unit sphere S3
in R4 (see, e.g., [RT10]), with
x3 + ix4
−x3 + ix4 x1 − ix2(cid:19) ←→ ϕ(u) = x = (x1, x2, x3, x4) ∈ S3,
SU(2) ∋ u =(cid:18) x1 + ix2
ZSU(2)
f (u)χk(u) du =ZS3
f (x)χk(x) dS,
we have
(5.13)
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
25
where f (x) := f (ϕ−1(x)), and χk(x) := χk(ϕ−1(x)). In order to find an
explicit formula for this integral over S3, we consider the parametrisation
x1 = cos
t
2
,
x2 = v,
x3 =rsin2 t
x4 =rsin2 t
2 − v2 · cos h,
2 − v2 · sin h,
(t, v, h) ∈ D,
where D = {(t, v, h) ∈ R3 : v ≤ sin t
2 , 0 ≤ t, h ≤ 2π}.
The reader will have no difficulty in showing that dS = sin t
2 dtdvdh.
Therefore, we have
f (h, v, t)χk(t) sin
t
2
dhdvdt.
f (h, v, t)χk(t) sin
t
2
dhdvdt.
Combining this and (5.13), we get
ZS3
f (x)χk(t)dS =ZD
Trbf (k) =ZD
Thus, we have expressed the invariant integral over SU(2) in the parameters
t, v, h. The application of Fubini's Theorem yields
ZD
f (h, v, t)χk(t) sin
t
2
dhdvdt =
2πZ0
χk(t) sin
t
2
Combining this and (5.12), we obtain
dt
sin t
2Z− sin t
2
dv
2πZ0
f (h, v, t) dh.
2πZ0
dt
kXn=−k
eint sin
t
2
Trbf (k) =
sin t
2Z− sin t
2
dv
2πZ0
f (h, v, t) dh.
Interchanging summation and integration, we get
2πZ0
kXn=−k
eint sin
t
2
Trbf (k) =
dt
sin t
2Z− sin t
2
dv
2πZ0
f (h, v, t) dh.
By making the change of variables t → 2t, we get
(5.14)
πZ0
kXn=−k
Trbf (k) =
e−i2nt · 2 sin t dt
sin tZ− sin t
dv
2πZ0
f (h, v, 2t) dh.
26
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
Let us now apply Theorem 2.7 in Lp(T). To do this we introduce some
notation. Denote
F (t) := 2 sin t
sin tZ− sin t
2πZ0
f (h, v, 2t) dh dv,
t ∈ (0, π).
We extend F (t) periodically to [0, 2π), that is F (x + π) = F (x). Since
f (t, v, h) is integrable, the integrability of F (t) follows immediately from
Fubini's Theorem. Thus function F (t) has a Fourier series representation
where the Fourier coefficients are computed by
F (t) ∼Xk∈Z bF (k)eikt,
2π Z[0,2π]
bF (k) =
1
F (t)e−ikt dt.
Let Ak be a 2k + 1-element arithmetic sequence with step 2 and initial term
−2k, i.e.,
Ak = {−2k,−2k + 2, . . . , 2k} = {−2k + 2j}2k
j=0.
Using this notation and (5.14), we have
(5.15)
Define
Trbf (k) = Xn∈Ak bF (n).
B = {Ak}∞
k=1.
Using the fact that B is a subset of the set M of all finite arithmetic pro-
gressions, and (5.15), we have
(5.16)
1
e(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi∈e bF (i)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
.
N0, then m runs over N. Using (5.16),
sup
k∈ 1
N0
2
2k+1≥2l+1
1
2k + 1(cid:12)(cid:12)(cid:12)Trbf (k)(cid:12)(cid:12)(cid:12) ≤ sup
e∈B
e≥2l+1
Denote m := 2l + 1. If l runs over 1
2
we get
(5.17) Xl∈ 1
2
N0
(2l + 1)p−2 sup
k∈ 1
2
N0
2k+1≥2l+1
e≥2l+1
1
e∈M
e(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi∈e bF (i)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ sup
2k + 1(cid:12)(cid:12)(cid:12)Trbf (k)(cid:12)(cid:12)(cid:12)
mp−2 sup
≤Xm∈N
e∈M
e≥m
1
p
p
.
1
e(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi∈e bF (i)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
27
Application of inequality (2.13) yields
(5.18)
Xm∈N
e∈M
e≥m
mp−2 sup
πZ0
p
1
e(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi∈e bF (i)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2πZ0
sin tZ− sin t
dv
sin t dt
F (t)p dt .
πZ0
Using Holder inequality, we obtain
≤ ckFkp
Lp(0,2π).
f (h, v, 2t)p dh.
By making the change of variables t → t
get
2 in the right hand side integral, we
πZ0
F (t)p dt .
2πZ0
sin
t
2
dt
sin t
2Z− sin t
2
dv
2πZ0
f (h, v, t)p dh
1
p
.
Thus, we have proved that
(5.19)
kFkLp(0,π) ≤ cpkfkLp(SU(2)),
where cp depending only on p. Combining (5.16), (5.17) and (5.19), we
obtain
Xm∈N
mp−2 sup
k∈ 1
N0
2
2k+1≥m
p
1
2k + 1(cid:12)(cid:12)(cid:12)Trbf (k)(cid:12)(cid:12)(cid:12)
≤ ckfkp
Lp(SU(2)).
This completes the proof.
6 Marcinkiewicz interpolation theorem
In this section we formulate and prove Marcinkiewicz interpolation theorem
for linear mappings between G and the space of matrix-valued sequences Σ
that will be realised via
Σ :=(cid:8)h = {h(π)}π∈ bG, h(π) ∈ Cdπ×dπ(cid:9) .
Thus, a linear mapping A : D′(G) → Σ takes a function to a matrix valued
sequence, i.e.
where
f 7→ Af =: h = {h(π)}π∈ bG,
h(π) ∈ Cdπ×dπ , π ∈ bG.
28
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
We say that a linear operator A is of strong type (p, q), if for every f ∈
Lp(G), we have Af ∈ ℓq(bG, Σ) and
where M is independent of f , and the space ℓq(bG, Σ) defined by the norm
kAfkℓq( bG,Σ) ≤ MkfkLp(G),
1
p
(6.1)
khkℓq( bG,Σ) :=Xπ∈ bG
dp( 2
p − 1
2 )kh(π)kp
HS
(2.9). The least M for which this is satisfied is taken to be the strong (p, q)-
norm of the operator A.
Denote the distribution functions of f and h by µG(t; f ) and ν bG(u; h),
respectively, i.e.
µG(x; f ) := Zu∈G
ν bG(y; h) := Xπ∈ bG
kh(π)kHS
√dπ
f (u)≥x
du, x > 0,
d2
π, y > 0.
≥y
(6.2)
(6.3)
Then
Lp(G) =ZG
kfkp
=Xπ∈ bG
ℓq( bG,Σ)
khkq
f (u)p du = p
+∞Z0
π(cid:18)kh(π)kHS√dπ (cid:19)q
d2
xp−1µG(x; f ) dx,
= q
+∞Z0
uq−1ν bG(y; h) dy.
A linear operator A : D′(SU(2)) → Σ satisfying
(6.4)
ν bG(y; Af ) ≤(cid:18) M
y kfkLp(G)(cid:19)q
is said to be of weak type (p, q); the least value of M in (6.4) is called weak
(p, q) norm of A.
Every operation of strong type (p, q) is also of weak type (p, q), since
y(cid:0)ν bG(y; Af )(cid:1) 1
q ≤ kAfkLq( bG) ≤ MkfkLp(G).
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
29
Theorem 6.1. Let 1 ≤ p1 < p < p2 < ∞. Suppose that a linear operator A
from D′(G) to Σ is simultaneously of weak types (p1, p1) and (p2, p2), with
norms M1 and M2, respectively, i.e.
(6.5)
(6.6)
ν bG(y; Af ) ≤(cid:18) M1
ν bG(y; Af ) ≤(cid:18) M2
y kfkLp1 (G)(cid:19)p1
y kfkLp2 (G)(cid:19)p2
,
.
Then for any p ∈ (p1, p2) the operator A is of strong type (p, p) and we have
(6.7)
0 < θ < 1,
kAfkℓp( bG,Σ) ≤ M 1−θ
1 M θ
2kfkLp(G),
where
1
p
=
1 − θ
p1
+
θ
p2
.
The proof is done in analogy to Zygmund [Zyg56] adapting it to our
setting.
Proof. Let f ∈ Lp(G). We have to prove inequality (6.7). By definition, we
have
(6.8)
kAfkp
ℓp( bG,Σ)
d2
π(cid:18)kAf (π)kHS
√dπ (cid:19)p
=
=Xπ∈ bG
+∞Z0
pxp−1ν bG(x; Af ) dx.
For a fixed arbitrary z > 0 we consider the decomposition
f = f1 + f2,
where f1 = f whenever f < z, and f1 = 0 otherwise; thus f2 > z or else
f2 = 0. Since f ∈ Lp(G) the same holds for f1 and f2; it follows that f1 is
in Lp1(G) and f2 ∈ Lp2(G). Hence Af1 and Af2 exist, by hypothesis, and
so does Af = A(f1 + f2). It follows that
(6.9)
The inequality
f1 = min(f, z),
f = f1 + f2.
kA(f1 + f2)(π)kHS ≤ kAf1(π)kHS + kAf2(π)kHS, π ∈ bG
leads to an inclusion
√dπ
(cid:26)π ∈ bG : kAf (π)kHS
≥ y(cid:27) ⊂
⊂(cid:26)π ∈ bG : kAf1(π)kHS
√dπ
≥
y
2(cid:27)[(cid:26)π ∈ bG : kAf2(π)kHS
√dπ
y
2(cid:27) .
≥
30
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
Then applying assumptions (6.5) and (6.6) to f1 and f2, we obtain
(6.10) ν bG(y; Af ) ≤ ν bG(
y
2
y
2
; Af2)
; Af1) + ν bG(
≤ M p1
1 y−p1kf1kp1
Lp1 (G) + M p2
1 y−p2kf2kp2
Lp2 (G).
The right side here depends on z and the main idea of the proof consists in
defining z as a suitable monotone functions of t, z = z(t), to be determined
later. By (6.9)
µG(t; f1) = µG(t; f ),
µG(t; f1) = 0,
for 0 < t ≤ z,
for t > z,
µG(t; f2) = µG(t + z; f ),
for t > 0.
Here, the last equation is a consequence of the fact that wherever f2 6= 0
we must have f1 = z, and so the second equation (6.9) takes the form
f = z + f2.
It follows from (6.10) that the last integral in (6.8) is less than
(6.11) M p1
1
(6.12) I1 =
=
p2
p2
p1
p1
dt
dy
dy+
= M p1
1 p1
+ M p2
2 p2
+ M p2
2
f1(u)p1 du
yp−p1−1
ZG
+∞Z0
yp−p2−1
f2(u)p1 du
ZG
+∞Z0
xp1−1µG(x; f ) dx
yp−p1−1
zZ0
+∞Z0
(x − z)p2−1µG(x; f ) dx
yp−p2−1
+∞Zz
+∞Z0
tp−p1−1
up1−1µG(u; f ) du
zZ0
+∞Z0
xp1−1µG(x; f )
yp−p1−1 dy
AxZ0
+∞Z0
+∞Z0
Ap−p1
p − p1
dx
dt
=
xp1−1+p−p1µG(x; f ) dx.
Set z(y) = A
change the order of integration in I1
y . Denote by I1 and I2 the two double integrals last written. We
dt.
(6.13) I2 = M p2
2 p2
= M p2
2 p2
dy
dy
yp−p2−1
(x − z)p2−1µG(x; f ) dx
+∞Zz
+∞Z0
yp−p2−1
xp2−1µG(x + z; f ) dx
+∞Z0
+∞Z0
yp−p2−1
xp2−1µG(x; f2) dx
+∞Z0
+∞Z0
xp2−1µG(x; f2)yp−p2−1 dy
+∞Z0
+∞Z0
xp2−1µG(x; f2)yp−p2−1 dy
+∞Z0
+∞Z
xp2−1µG(x; f2)
yp−p2−1 dy
+∞ZAx
+∞Z0
+∞Z0
+∞Z0
xp2−1+p−p2µG(x; f2) dx
Ap−p2
p2 − p
Ap−p2
p2 − p
M p2
2 p2
M p2
2 p2
≤
Ax
1
ξ
dy
dx
dx
dx
= M p2
2 p2
= M p2
2 p2
= M p2
2 p2
= M p2
2 p2
=
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
31
Similarly, making a substitution x− z → x and using (6.9) we see that I2 is
xp2−1+p−p2µG(x; f ) dx.
Collecting estimates (6.11), (6.12), (6.13) we see that integral in (6.8) does
not exceed
(6.14)
M p1
1 p1
Ap−p1
p − p1
+∞Z0
xp−1µG(x; f ) dx + M p2
2 p2
Ap−p2
p2 − p
+∞Z0
xp−1µG(x; f2)dx.
Now, using the identity
+∞Z0
xp−1µG(x; f ) dx =ZG
f (u)p du = kfkp
Lp(G),
and inequalities (6.8) and (6.14) we get
kAfkp
ℓp( bG) ≤(cid:18)M p1
1 p1
Ap−p1
p − p1
+ M p2
2 p2
Ap−p2
p2 − p(cid:19)p
kfkp
ℓp( bG)
.
32
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
Next we set
p1
p2
A = M
p2−p1
p1−p2
1 M
2
.
A simple computation shows that
M p1
1 Ap−p1 = M p2
p1(p2−p)
2 Ap−p2 = M
p2−p1
1
p2(p1−p)
p1−p2
M
2
= M 1−θ
1 M θ
2 ,
1
p
=
1 − θ
p1
+
θ
p2
.
Finally, we have
where
This completes the proof.
References
1 M θ
2kfkLp(G),
kAfkℓp( bG) ≤ Kp,p1,p2M 1−θ
Kp,p1,p2 =(cid:18) p1
p − p1
+
p
p2
p2 − p(cid:19) 1
.
[Ank90]
J.-P. Anker. Lp Fourier multipliers on Riemannian symmetric
spaces of the noncompact type. Ann. of Math. (2), 132(3):597 --
628, 1990.
[Arh12] C. Arhancet. Unconditionality, Fourier multipliers and Schur
multipliers. Colloq. Math., 127(1):17 -- 37, 2012.
[BL76]
J. Bergh and J. Lofstrom. Interpolation spaces. An introduction.
Springer-Verlag, Berlin-New York, 1976. Grundlehren der Math-
ematischen Wissenschaften, No. 223.
[CdG71] R. R. Coifman and M. de Guzm´an. Singular integrals and multi-
pliers on homogeneous spaces. Rev. Un. Mat. Argentina, 25:137 --
143, 1970/71. Collection of articles dedicated to Alberto Gonz´alez
Dom´ınguez on his sixty-fifth birthday.
[CGM93] M. Cowling, S. Giulini, and S. Meda. Lp-Lq estimates for func-
tions of the Laplace-Beltrami operator on noncompact symmetric
spaces. I. Duke Math. J., 72(1):109 -- 150, 1993.
[CKS11] M. G. Cowling, O. Klima, and A. Sikora. Spectral multipliers for
the Kohn sublaplacian on the sphere in Cn. Trans. Amer. Math.
Soc., 363(2):611 -- 631, 2011.
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
33
[Cow74] M. G. Cowling. Spaces Aq
p and Lp-Lq Fourier multipliers. PhD
thesis, The Flinders University of South Australia, 1974.
[CS01] M. Cowling and A. Sikora. A spectral multiplier theorem for a
sublaplacian on SU(2). Math. Z., 238(1):1 -- 36, 2001.
[CW71a] R. R. Coifman and G. Weiss.
Analyse harmonique non-
commutative sur certains espaces homog`enes. Lecture Notes in
Mathematics, Vol. 242. Springer-Verlag, Berlin, 1971. ´Etude de
certaines int´egrales singuli`eres.
[CW71b] R. R. Coifman and G. Weiss. Multiplier transformations of
functions on SU(2) and P2. Rev. Un. Mat. Argentina, 25:145 --
166, 1971. Collection of articles dedicated to Alberto Gonz´alez
Dom´ınguez on his sixty-fifth birthday.
[CW74] R. R. Coifman and G. Weiss. Central multiplier theorems for
compact Lie groups. Bull. Amer. Math. Soc., 80:124 -- 126, 1974.
[Gau66] G. I. Gaudry. Multipliers of type (p, q). Pacific J. Math., 18:477 --
488, 1966.
[GT80]
S. Giulini and G. Travaglini. Lp-estimates for matrix coefficients
of irreducible representations of compact groups. Proc. Amer.
Math. Soc., 80(3):448 -- 450, 1980.
[HL27] G. H. Hardy and J. E. Littlewood. Some new properties of Fourier
constants. Math. Ann., 97(1):159 -- 209, 1927.
[Hor60]
L. Hormander. Estimates for translation invariant operators in
Lp spaces. Acta Math., 104:93 -- 140, 1960.
[HR74]
E. Hewitt and K. A. Ross. Rearrangements of Lr Fourier series on
compact Abelian groups. Proc. Lond. Math. Soc. (3), 29:317 -- 330,
1974.
[Mih56]
S. G. Mihlin. On the theory of multidimensional singular integral
equations. Vestnik Leningrad. Univ., 11(1):3 -- 24, 1956.
[Mih57]
S. G. Mihlin. Singular integrals in Lp spaces. Dokl. Akad. Nauk
SSSR (N.S.), 117:28 -- 31, 1957.
34
[NT00]
R. Akylzhanov, E. Nursultanov and M. Ruzhansky
E. D. Nursultanov and N. T. Tleukhanova. Lower and upper
bounds for the norm of multipliers of multiple trigonometric
Fourier series in Lebesgue spaces. Funct. Anal. Appl., 34(2):86 --
88, 2000.
[NT11]
E. Nursultanov and S. Tikhonov. Net spaces and boundedness of
integral operators. J. Geom. Anal., 21(4):950 -- 981, 2011.
[Nur98a] E. Nursultanov. Net spaces and inequalities of Hardy-Littlewood
type. Sb. Math., 189(3):399 -- 419, 1998.
[Nur98b] E. D. Nursultanov. On multipliers of Fourier series in a trigono-
metric system. Mat. Zametki, 63(2):235 -- 247, 1998.
[RR86]
F. Ricci and R. L. Rubin. Transferring Fourier multipliers from
SU(2) to the Heisenberg group. Amer. J. Math., 108(3):571 -- 588,
1986.
[RT10] M. Ruzhansky and V. Turunen. Pseudo-differential operators
and symmetries. Background analysis and advanced topics, vol-
ume 2 of Pseudo-Differential Operators. Theory and Applications.
Birkhauser Verlag, Basel, 2010.
[RT13] M. Ruzhansky and V. Turunen. Global quantization of pseudo-
differential operators on compact Lie groups, SU(2), 3-sphere,
and homogeneous spaces. Int. Math. Res. Not. IMRN, (11):2439 --
2496, 2013.
[RW13] M. Ruzhansky and J. Wirth. On multipliers on compact Lie
groups. Funct. Anal. Appl., 47(1):87 -- 91, 2013.
[RW15] M. Ruzhansky and J. Wirth. Lp Fourier multipliers on compact
Lie groups. Math. Z., 280: 621 -- 642, 2015.
[Ste70]
[Vil68]
E. M. Stein. Singular integrals and differentiability properties
of functions. Princeton Mathematical Series, No. 30. Princeton
University Press, Princeton, N.J., 1970.
N. J. Vilenkin. Special functions and the theory of group represen-
tations. Translated from the Russian by V. N. Singh. Translations
of Mathematical Monographs, Vol. 22. American Mathematical
Society, Providence, R. I., 1968.
Hardy-Littlewood-Paley inequalities and Fourier multipliers on SU(2)
35
[Vre74]
L. Vretare. On Lp Fourier multipliers on a compact Lie-group.
Math. Scand., 35:49 -- 55, 1974.
[Wei72] N. J. Weiss. Lp estimates for bi-invariant operators on compact
Lie groups. Amer. J. Math., 94:103 -- 118, 1972.
[Zyg56] A. Zygmund. On a theorem of Marcinkiewicz concerning inter-
polation of operations. J. Math. Pures Appl. (9), 35:223 -- 248,
1956.
|
1005.3528 | 1 | 1005 | 2010-05-19T19:12:27 | Thin-very tall compact scattered spaces which are hereditarily separable | [
"math.FA",
"math.GN"
] | We strengthen the property $\Delta$ of a function $f:[\omega_2]^2\rightarrow [\omega_2]^{\leq \omega}$ considered by Baumgartner and Shelah. This allows us to consider new types of amalgamations in the forcing used by Rabus, Juh\'asz and Soukup to construct thin-very tall compact scattered spaces. We consistently obtain spaces $K$ as above where $K^n$ is hereditarily separable for each $n\in\N$. This serves as a counterexample concerning cardinal functions on compact spaces as well as having some applications in Banach spaces: the Banach space $C(K)$ is an Asplund space of density $\aleph_2$ which has no Fr\'echet smooth renorming, nor an uncountable biorthogonal system. | math.FA | math |
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH
ARE HEREDITARILY SEPARABLE
CHRISTINA BRECH AND PIOTR KOSZMIDER
Abstract. We strengthen the property ∆ of a function f : [ω2]2 → [ω2]≤ω
considered by Baumgartner and Shelah. This allows us to consider new types of
amalgamations in the forcing used by Rabus, Juh´asz and Soukup to construct
thin-very tall compact scattered spaces. We consistently obtain spaces K as
above where K n is hereditarily separable for each n ∈ N. This serves as a
counterexample concerning cardinal functions on compact spaces as well as
having some applications in Banach spaces: the Banach space C(K) is an
Asplund space of density ℵ2 which has no Fr´echet smooth renorming, nor an
uncountable biorthogonal system.
1. Introduction
Given a compact scattered space K, we call the derivative of K (denoted by
K ′) the subset of K formed by its accumulation points and we inductively define
K (α) = (K (β))′ if α = β+1 and K (α) = Tβ<α K (β) if α is a limit ordinal. The height
of K, ht(K), is the smallest ordinal α such that K (α) is finite and nonempty, and
the width of K, wd(K), is the supremum of the cardinalities K (α) \K (α+1) for α <
ht(K). We call K = Sα<ht(K) K (α) \ K (α+1) the Cantor-Bendixson decomposition
of K and K (α) \ K (α+1) its αth Cantor-Bendixson level.
The purpose of this work is to show that the existence of compact hereditarily
separable scattered spaces of height ω2 is consistent with the usual axioms of set
theory. For a given ordinal θ let us consider the following notation:
• A cw(θ) space is a compact scattered space of countable width and height
equal to θ.
• A hs(θ) space is a compact scattered space which is hereditarily separable
and of height equal to θ.
cw(ω1) spaces are usually called thin-tall spaces and cw(ω2) spaces are the thin-
very tall spaces. First we remark that any hs(θ) space is a cw(θ) space as the
Cantor-Bendixson levels form discrete subspaces. Whether there is or not in ZFC
a cw(ω1) space was a question posed by Telg´arsky in 1968 (unpublished) and first
(consistently) answered by Ostaszewski [21], using ♦. Rajagopalan constructed the
first ZFC example of a cw(ω1) in [23]. Further, Juh´asz and Weiss generalized these
1991 Mathematics Subject Classification. Primary 54G12; Secondary 03E35, 46B26.
The research has been a part of Thematic Project FAPESP (2006/02378-7). The first au-
thor was supported by scholarships from CAPES (3804/05-4) and CNPq (140426/2004-3 and
202532/2006-2). She would like to thank Stevo Todorcevic and the second author, her Ph.D.
advisors at the University of Sao Paulo and at the University of Paris 7, under whose supervision
the results of this paper were obtained.
The second author was partially supported by Polish Ministry of Science and Higher Education
research grant N N201 386234.
1
2
CHRISTINA BRECH AND PIOTR KOSZMIDER
results (and simplified their proofs) in [12] proving in ZFC that for any ordinal
θ < ω2, there is a cw(θ) space.
For higher θ's the situation changes: in any model of CH there are no cw(ω2)
spaces and Just proved in [13] that neither are there such spaces in the Cohen model
(where ¬ CH holds). On the other hand, Baumgartner and Shelah [2] constructed
by forcing the first consistent example of a cw(ω2) space. An interesting point of
this forcing construction was the use of a new combinatorial device called a function
with the property ∆.
The main purpose of this work is to prove the consistency of the existence of
a hs(ω2) space. In fact, our space has even stronger properties: each of its finite
powers is hereditarily separable. Whether consistently there are hs(ω3) or even
cw(ω3) spaces remains a well-known open question. On the other hand, Mart´ınez
in [17] adopted the method of [2] to obtain the consistency of the existence of cw(θ)
spaces for each θ < ω3.
It follows from an old result of Shapirovskiı [25] that for any compact space K,
hd(K) ≤ hL(K)+. Our construction shows that the dual inequality does not follow
from ZFC, since for our compact space K, we have that hL(K) = ℵ2 6≤ ℵ1 =
hd(K)+. Nevertheless, the dual inequality holds under GCH for regular spaces:
since the weight w(K) of a regular space K is less or equal to 2d(K) (see, for example,
[8]), we trivially conclude that hL(K) ≤ w(K) ≤ 2d(K) = d(K)+ ≤ hd(K)+ under
GCH.
Turning to properties of Banach spaces, let us first recall some definitions and
results: a Banach space X is an Asplund space if every continuous and convex
real-valued function on X is Fr´echet smooth at all points of a Gδ dense subset of
X. For separable Banach spaces, this is equivalent to admitting a Fr´echet smooth
renorming (see [4]). Namioka and Phelps proved in [19] that C(K) is Asplund if
and only if K is scattered. Thus, our C(K) is an Asplund space.
Haydon constructed in [7] the first nonseparable Asplund space C(K) which does
not admit a Fr´echet smooth renorming, concluding that the situation changes for
nonseparable Asplund spaces. Later, Jim´enez Sevilla and Moreno analyzed in [10]
the structural properties of the space C(K), where K is the well-known Kunen line
constructed under CH (see [20]). They showed, for the Kunen line K, that C(K)
is also a nonseparable Asplund space with no Fr´echet smooth renorming.
The weight of our space K is ℵ2, so that C(K) is an Asplund space of density ℵ2.
The fact that K is compact scattered and every finite power of K is hereditarily
separable implies, in the same way as for the Kunen line, that C(K) does not admit
any Fr´echet smooth renorming, but as in the case of the Kunen line we do not know
if it admits a Gateaux smooth renorming, or a Fr´echet smooth bump function.
A biorthogonal system on a Banach space X is a family (xα, ϕα)α<κ ⊆ X × X ∗
such that ϕα(xβ ) = δα,β and a semi-biorthogonal system on a Banach space X is
a sequence (xα, ϕα)α<κ ⊆ X × X ∗ such that ϕα(xβ ) = 1, if α = β, ϕα(xβ) = 0, if
α < β and ϕα(xβ ) ≥ 0 if β < α. Todorcevic showed in [29] (Theorem 9 together
with the results of [3]) the existence of uncountable semi-biorthogonal systems in
Banach spaces C(K) of density strictly greater than ℵ1. On the other hand, the fact
that our space K is compact scattered and every finite power of K is hereditarily
separable implies, in the same way as for the Kunen line, that C(K) does not admit
an uncountable biorthogonal system.
It follows that Todorcevic's result cannot
be improved in ZFC by replacing the existence of uncountable semi-biorthogonal
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
3
systems by the existence of uncountable biorthogonal systems in spaces C(K) of
large density. On the other hand it is proved in [29] that it is consistent that every
nonseparable Banach space has an uncountable biorthogonal system, showing that
the existence of a Banach space like ours or Kunen's cannot be proved in ZFC.
Our construction is based on the Juh´asz and Soukup [11] interpretation of Rabus'
work [22], where he modified the Baumgartner-Shelah forcing from [2] to obtain a
countably tight space which is initially ω1-compact and noncompact, answering a
question of Dow and van Douwen.
This paper is organized as follows: we finish this section by reviewing the method
of Juh´asz and Soukup and some related results and definitions which we will need
afterwards. In Section 2 we prove the key lemma which enables us to prove our
main result in a straightforward way. This lemma introduces a new way of amal-
gamating conditions in forcings which add thin-very tall spaces. One can apply
these amalgamations in the generic construction if one strengthens the property ∆
of a function involved in the forcing. In Section 3, we introduce the strong prop-
erty ∆ and assuming the existence of a function which satisfies it, we prove the
main results and analyze their consequences in topological and functional analytic
terms. Section 4 is devoted to establishing the consistency of the existence of a
function with the strong property ∆. Section 4 is due to the second author and the
remaining sections to the first author.
The notation and terminology used are those of [11]. Given a set X, ℘(X) is the
power set of X and, given a cardinal κ, [X]κ (resp. [X]≤κ and [X]<κ) denotes the
family of subsets of X of cardinality equal to κ (resp.
less or equal to κ and less
than κ).
Let us start by recalling the definition of the property ∆:
Definition 1.1 (Baumgartner, Shelah, [2], p.122). A function f : [ω2]2 → [ω2]≤ω
has the property ∆ if f ({ξ, η}) ⊆ min{ξ, η} for all {ξ, η} ∈ [ω2]2 and for any
uncountable family A of finite subsets of ω2, there are distinct a, b ∈ A such that
for any τ ∈ a ∩ b, any ξ ∈ a \ b and any η ∈ b \ a we have:
1) a ∩ b ∩ min{ξ, η} ⊆ f ({ξ, η});
2) τ < ξ ⇒ f ({τ, η}) ⊆ f ({ξ, η});
3) τ < η ⇒ f ({τ, ξ}) ⊆ f ({ξ, η}).
Now, we fix a function f : [ω2]2 → [ω2]≤ω with the property ∆.
Definition 1.2 (Juh´asz, Soukup [11], Definition 2.1). Let Pf be the forcing formed
by conditions p = (Dp, hp, ip) where:
1. Dp ∈ [ω2]<ω;
2. hp : Dp → ℘(Dp) and for all ξ ∈ Dp, max hp(ξ) = ξ;
3. ip : [Dp]2 → [Dp]<ω and for all ξ, η ∈ Dp, ξ < η, we have that:
(a) if ξ ∈ hp(η), then hp(ξ) \ hp(η) ⊆ Sγ∈ip({ξ,η}) hp(γ),
(b) if ξ /∈ hp(η), then hp(ξ) ∩ hp(η) ⊆ Sγ∈ip({ξ,η}) hp(γ),
(c) ip({ξ, η}) ⊆ f ({ξ, η});
ordered by p ≤ q if Dp ⊇ Dq, for all ξ ∈ Dq, hp(ξ) ∩ Dq = hq(ξ) and ip[Dq]2 = iq.
To simplify notation, it is convenient to define the following:
4
CHRISTINA BRECH AND PIOTR KOSZMIDER
Definition 1.3 (Juh´asz, Soukup [11]). Given finite nonempty sets of ordinals x
and y such that max x < max y, we define
x ∗ y = (cid:26) x \ y
x ∩ y
if max x ∈ y,
if max x /∈ y.
We now rewrite conditions 3.(a) and (b) of the definition of the forcing as
hp(ξ) ∗ hp(η) ⊆ [γ∈ip({ξ,η})
hp(γ).
To define the space Kf , fix the ground model V and a generic filter G.
Definition 1.4 (Juh´asz, Soukup [11], Definition 2.3). For each ξ < η < ω2, working
in V Pf , let
h(ξ) = [p∈G
hp(ξ)
and
i({ξ, η}) = [p∈G
ip({ξ, η}),
and let Lf be the topological space (ω2, τ ), where τ is the topology on ω2 which
has the family of sets
{h(ξ) : ξ < ω2} ∪ {ω2 \ h(ξ) : ξ < ω2}
as a topological subbasis. We call h(ξ) the generic neighborhood of ξ.
From Theorem 1.5 of [11], it follows that for all ξ < ω2, h(ξ) is a compact
subspace of (ω2, τ ) and it easy to check that
{h(ξ) \ [η∈F
h(η) : F ∈ [ξ]<ω} forms a local topological basis at ξ.
(+)
Therefore Lf is a locally compact scattered zero-dimensional space.
We are now ready to define Kf :
Definition 1.5. In V Pf , Kf is the one-point compactification of Lf . The point of
compactification is denoted ∗, thus Kf \ Lf = {∗}.
In particular, we use the following results.
Theorem 1.6 (Rabus [22], Lemma 4.1; Juh´asz, Soukup [11], Lemma 2.8). Pf
satisfies c.c.c.
Proposition 1.7. V Pf satisfies "Kf is a compact scattered zero-dimensional space".
2. Amalgamating conditions
In this section, we present the key lemma needed to prove our main result. Let
us start with some preliminaries and auxiliary lemmas.
Definition 2.1. Let p1 = (D1, h1, i1), p2 = (D2, h2, i2) ∈ Pf be two conditions.
We say that p1 and p2 are isomorphic conditions if there is an order-preserving
bijective function e : D1 → D2 satisfying the following conditions:
(a) if ξ, η ∈ D1, then ξ ∈ h1(η) if and only if e(ξ) ∈ h2(e(η));
(b) if ξ ∈ D1 ∩ D2, then e(ξ) = ξ.
In this case, if the order-preserving bijection e is such that ξ ≤ e(ξ) for every ξ ∈ D1
we say that p1 is lower than p2.
For example we have the following:
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
5
Lemma 2.2. Let p1 = (D1, h1, i1), p2 = (D2, h2, i2) ∈ Pf be two isomorphic
conditions and let e : D1 → D2 be the order-preserving bijection. Then for every
ξ ∈ D1 ∩ D2,
(a) (h1(ξ) ∪ h2(ξ)) ∩ D1 = h1(ξ),
(b) (h1(ξ) ∪ h2(ξ)) ∩ D2 = h2(ξ),
(c) h1(ξ) = e−1[h2(ξ)].
Proof. Directly from Definition 2.1.
(cid:3)
Definition 2.3 (Juh´asz, Soukup [11]). Given p1 = (D1, h1, i1), p2 = (D2, h2, i2) ∈
Pf , define a mapping δ2 : dom(δ2) → D1 ∩ D2, where
dom(δ2) = {η ∈ D2 : there is δ ∈ D1 ∩ D2 such that η ∈ h2(δ)}
and
δ2(η) = min{δ ∈ D1 ∩ D2 : η ∈ h2(δ)}.
Lemma 2.4. Suppose that p1 = (D1, h1, i1) and p2 = (D2, h2, i2) ∈ Pf are two
conditions. Then,
(a) for all η ∈ dom(δ2) \ D1, we have that η < δ2(η) and
(b) for all η ∈ D1 ∩ D2 we have η ∈ dom(δ2) and δ2(η) = η.
Proof. Directly from Definition 2.3.
(cid:3)
We prove the next lemma, for the reader's convenience.
Lemma 2.5 (Juh´asz, Soukup [11]). Let p1 = (D1, h1, i1), p2 = (D2, h2, i2) ∈ Pf
be two isomorphic conditions. If ξ ∈ D1 ∩ D2, then
Proof. Let ξ ∈ D1 ∩ D2.
h2(ξ) = δ−1
2 [h1(ξ)].
Suppose that η ∈ dom(δ2) and δ2(η) ∈ h1(ξ). Since δ2(η), ξ ∈ D1 ∩ D2, it fol-
lows from 2.1.(a) and (b) that δ2(η) ∈ h2(ξ). Suppose that η /∈ h2(ξ). Then,
η ∈ h2(δ2(η)) ∗ h2(ξ) so that there is δ ∈ i2({δ2(η), ξ}) such that η ∈ h2(δ),
which contradicts the minimality of δ2(η) and concludes the proof of the inclu-
sion δ−1
2 [h1(ξ)] ⊆ h2(ξ).
Reciprocally, if η ∈ h2(ξ), then η ∈ dom(δ2). Suppose that δ2(η) /∈ h2(ξ). Then,
η ∈ h2(δ2(η)) ∗ h2(ξ) so that there is δ ∈ i2({δ2(η), ξ}) such that η ∈ h2(δ), which
contradicts the minimality of δ2(η). So, δ2(η) ∈ h2(ξ) and since δ2(η), ξ ∈ D1 ∩ D2,
it follows from 2.1.(a) and (b) that δ2(η) ∈ h1(ξ), concluding the proof of the
lemma.
(cid:3)
In the proof of c.c.c., Rabus, Juh´asz and Soukup considered the minimal amal-
gamation which is constructed in a symmetric way with respect to both of the
conditions being extended. We will consider an asymmetric amalgamation. The
lack of symmetry in our amalgamation is the result of using two functions, δ2 and
e, in the definition of the amalgamation. The final auxiliary lemma below char-
acterizes the sets given by the operation ∗ for elements of the extended condition.
The role of the function g will be played by δ2 or by e.
Lemma 2.6. Let p = (Dp, hp, ip) ∈ Pf and let Dq ∈ [ω2]<ω, hq : Dq → ℘(Dq) and
g : dom(g) → Dp be such that
(i) Dp ⊆ Dq; dom(g) ⊆ Dq,
6
CHRISTINA BRECH AND PIOTR KOSZMIDER
(ii) for all ξ ∈ Dp ∩ dom(g) we have g(ξ) = ξ,
(iii) for all ξ ∈ Dp, hq(ξ) = hp(ξ) ∪ g−1[hp(ξ)].
Then, for all ξ, η ∈ Dp, ξ < η, we have that
(hq(ξ) ∗ hq(η)) ∩ Dp = hp(ξ) ∗ hp(η)
and
(hq(ξ) ∗ hq(η)) ∩ (Dq \ Dp) = g−1[hp(ξ) ∗ hp(η)] ∩ (Dq \ Dp).
Proof. Since ξ, η ∈ Dp, ξ < η, by (ii) we have that ξ ∈ hp(η) if and only if ξ ∈ hq(η),
so (ii) obviously gives (hq(ξ) ∗ hq(η)) ∩ Dp = hp(ξ) ∗ hp(η).
Now suppose ξ ∈ hq(η), so ξ ∈ hp(η) and so by (iii)
(hq(ξ) \ hq(η)) ∩ (Dq \ Dp) = (hq(ξ) ∩ (Dq \ Dp)) \ (hq(η) ∩ (Dq \ Dp))
= (g−1[hp(ξ)]∩(Dq \ Dp))\ (g−1[hp(η)]∩(Dq \ Dp)) = g−1[hp(ξ)\ hp(η)]∩(Dq \ Dp).
On the other hand, if ξ /∈ hq(η), then ξ /∈ hp(η) and so by (iii)
(hq(ξ) ∩ hq(η)) ∩ (Dq \ Dp) = (hq(ξ) ∩ (Dq \ Dp)) ∩ (hq(η) ∩ (Dq \ Dp))
= (g−1[hp(ξ)]∩(Dq \Dp))∩(g−1[hp(η)]∩(Dq \Dp)) = g−1[hp(ξ)∩hp(η)]∩(Dq \Dp),
concluding the proof of the lemma.
(cid:3)
Now we go to our key lemma: a strong hypothesis about the behaviour of the
function f allows us to amalgamate two isomorphic conditions, one lower than the
other, into a common extension q in such a way that h(ξ) ∩ Dq ⊆ h[e(ξ)] ∩ Dq for
ξ in the domain of the lower of the two conditions.
Lemma 2.7. Let p1 = (D1, h1, i1), p2 = (D2, h2, i2) ∈ Pf be two isomorphic
conditions and suppose p1 is lower than p2. Let e : D1 → D2 be the order-preserving
bijective function and assume that
(A) if ξ, η ∈ D1 ∩ D2 and ξ 6= η, then i1({ξ, η}) = i2({ξ, η});
(B) for all ζ ∈ D1 ∩ D2, all ξ ∈ D1 \ D2 and all η ∈ D2 \ D1:
(i) if ζ < ξ, then f ({ζ, η}) ⊆ f ({ξ, η});
(ii) D1 ∩ ξ ∩ η ⊆ f ({ξ, η}).
Then there is q ∈ Pf , q ≤ p1, p2, such that for all ξ ∈ D1 and all η ∈ D2:
ξ ∈ hq(η) if and only if e(ξ) ∈ h2(η).
Proof. We define q = (Dq, hq, iq) by: Dq = D1 ∪ D2;
and
if ξ ∈ D1,
if ξ ∈ D2;
if ξ, η ∈ D1,
if ξ, η ∈ D2,
otherwise.
2 [h1(ξ)]
h2(ξ) ∪ e−1[h2(ξ)]
hq(ξ) = (cid:26) h1(ξ) ∪ δ−1
iq({ξ, η}) =
i1({ξ, η})
i2({ξ, η})
f ({ξ, η}) ∩ Dq
Note that (A) implies that the set iq({ξ, η}) is well-defined for any ξ, η ∈ D1 ∩ D2,
ξ 6= η; clearly iq is well-defined for the other pairs. Also, if ξ ∈ D1 ∩D2, then the set
hq(ξ) is well-defined because both of the conditions reduce to hq(ξ) = h1(ξ) ∪ h2(ξ)
by Lemmas 2.5 and 2.2.(c).
We have to show that q ∈ Pf , i.e., that q satisfies conditions 1, 2 and 3 from
Definition 1.2. The fact that q satisfies conditions 1.2.1 and 1.2.3.(c) follows directly
from the definition of q and from the fact that p1, p2 ∈ Pf . Condition 1.2.2.
is
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
7
satisfied because p1, p2 ∈ Pf and the functions e and δ2 are nondecreasing. In what
follows we will be using Lemma 2.6 for p = p1, p2 and g = δ2, e respectively. The
hypothesis of the lemma is satisfied for these objects by 2.4.(b) and 2.1.(b).
Now we check conditions 1.2.3.(a) and (b). Let ξ, η ∈ Dq, ξ < η, and we consider
the following cases:
Case 1. ξ, η ∈ D1.
It follows from the definition of q and from Lemma 2.6 that
(hq(ξ) ∗ hq(η)) ∩ D1 = h1(ξ) ∗ h1(η)
and
(hq(ξ) ∗ hq(η)) ∩ (D2 \ D1) = δ−1
2 [h1(ξ) ∗ h1(η)] ∩ (D2 \ D1).
Now let ζ ∈ hq(ξ) ∗ hq(η).
Subcase 1.1. ζ ∈ D1.
In this subcase, ζ ∈ h1(ξ) ∗ h1(η) and there is γ ∈ i1({ξ, η}) = iq({ξ, η}) such
that ζ ∈ h1(γ) ⊆ hq(γ), as we wanted.
Subcase 1.2. ζ ∈ D2 \ D1.
In this subcase, δ2(ζ) ∈ h1(ξ) ∗ h1(η) and, since δ2(ζ), ξ, η ∈ D1 and p1 ∈ Pf ,
there is γ ∈ i1({ξ, η}) = iq({ξ, η}) such that δ2(ζ) ∈ h1(γ). Since γ ∈ D1, it follows
by the definition of q that ζ ∈ hq(γ), as we wanted.
Case 2. ξ, η ∈ D2.
It follows from the definition of q and from Lemma 2.6 that
(hq(ξ) ∗ hq(η)) ∩ D1 = h1(ξ) ∗ h1(η)
and
(hq(ξ) ∗ hq(η)) ∩ (D2 \ D1) = e−1[h1(ξ) ∗ h1(η)] ∩ (D2 \ D1).
Now let ζ ∈ hq(ξ) ∗ hq(η).
Subcase 2.1. ζ ∈ D1 \ D2.
In this subcase, e(ζ) ∈ h2(ξ) ∗ h2(η) and, since e(ζ), ξ, η ∈ D2 and p2 ∈ Pf , there
is γ ∈ i2({ξ, η}) = iq({ξ, η}) such that e(ζ) ∈ h2(γ). Since γ ∈ D2, it follows by
the definition of q that ζ ∈ hq(γ), as we wanted.
Subcase 2.2. ζ ∈ D2.
In this subcase, ζ ∈ h2(ξ) ∗ h2(η) and there is γ ∈ i2({ξ, η}) = iq({ξ, η}) such
that ζ ∈ h2(γ) ⊆ hq(γ), as we wanted.
Case 3. ξ ∈ D1 \ D2 and η ∈ D2 \ D1.
Here we fix ζ ∈ hq(ξ) ∗ hq(η) and we consider the following subcases:
Subcase 3.1. ζ ∈ D1.
In this subcase, ζ ∈ D1 ∩ ξ ∩ η and it follows from (B).(ii) that ζ ∈ f ({ξ, η}).
Hence, ζ ∈ D1 ∩ f ({ξ, η}) ⊆ Dq ∩ f ({ξ, η}) = iq({ξ, η}). Taking γ = ζ, we conclude
that ζ ∈ hq(γ) and γ ∈ iq({ξ, η}), as we wanted.
Subcase 3.2. ζ ∈ D2 \ D1.
First note that, regardless of the fact whether hq(ξ) ∗ hq(η) = hq(ξ) ∩ hq(η) or
hq(ξ)∗hq(η) = hq(ξ)\hq(η), the assumption ζ ∈ hq(ξ)∗hq(η) implies that ζ ∈ hq(ξ)
8
CHRISTINA BRECH AND PIOTR KOSZMIDER
In this subcase, it follows from the definition of hq(ξ) that δ2(ζ) ∈ h1(ξ), so that
δ2(ζ) ∈ D1∩ξ∩η; and it follows from (B).(ii) that δ2(ζ) ∈ f ({ξ, η})∩Dq = iq({ξ, η}).
By the definition of δ2(ζ), ζ ∈ h2(δ2(ζ)) ⊆ hq(δ2(ζ)). Taking γ = δ2(ζ), we have
that ζ ∈ hq(γ) and γ ∈ iq({ξ, η}), concluding the proof of this subcase.
Case 4. ξ ∈ D2 \ D1 and η ∈ D1 \ D2.
Again we fix ζ ∈ hq(ξ) ∗ hq(η) and we consider the following subcases:
Subcase 4.1. ζ ∈ D1.
The proof in this subcase follows identically to the proof of Subcase 3.1.
Subcase 4.21. ζ ∈ D2 \ D1.
We start this last subcase by proving the following:
Fact 1. {ζ, ξ} ∩ dom(δ2) is a nonempty set such that min δ2[{ζ, ξ}] < η and if both
ζ and ξ are in dom(δ2), then δ2(ζ) 6= δ2(ξ).
Proof of Fact 1. First remark that, from the definition of ∗, it follows that if ξ /∈
hq(η), then ζ ∈ hq(ξ) ∗ hq(η) = hq(ξ) ∩ hq(η) and therefore ζ ∈ hq(η). Analogously,
if ξ ∈ hq(η), then ζ ∈ hq(ξ) ∗ hq(η) = hq(ξ) \ hq(η) and therefore ζ /∈ hq(η). So,
{ζ, ξ} ∩ hq(η) = 1.
From the definition of q we have that, since ξ, ζ 6∈ D1 and η ∈ D1 in this subcase,
the above means that
{ζ, ξ} ∩ δ−1
2 [h1(η)] = 1,
so that {ζ, ξ} ∩ dom(δ2) is a nonempty set. The above observation also implies that
δ2[{ζ, ξ}] ∩ h1(η) 6= ∅ and so, min δ2[{ζ, ξ}] ≤ η. Now since η /∈ D2 and the range
of δ2 is included in D1 ∩ D2, the inequality must be strict.
Finally, we have seen that ζ ∈ hq(η) if and only if ξ /∈ hq(η) and so, if both ζ and
ξ are in the domain of δ2, it follows that δ2(ζ) ∈ h1(η) if and only if δ2(ξ) /∈ h1(η),
so that δ2(ζ) 6= δ2(ξ), concluding the proof of Fact 1.
Take θ = min{δ2(ξ), δ2(ζ)} and note that θ 6= ξ since ξ ∈ D2 \ D1 and the range
of δ2 is included in D1 ∩ D2. We go now to the following subcases:
Subcase 4.2.1. θ < ξ.
Here, θ 6= δ2(ξ) and therefore δ2(ζ) = θ ∈ D1 ∩ ξ ∩ η. From condition (B).(ii),
it follows that δ2(ζ) ∈ f ({ξ, η}) ∩ Dq = iq({ξ, η}). Since ζ ∈ h2(δ2(ζ)) ⊆ hq(δ2(ζ)),
taking γ = δ2(ζ), we have that ζ ∈ hq(γ) and γ ∈ iq({ξ, η}), as we wanted.
Subcase 4.2.2. θ > ξ.
Note that ζ ∈ hq(ξ) ∗ hq(η) ⊆ hq(ξ). Since ζ and ξ satisfying the hypothesis of
the Case 4.2. are in D2 \ D1, it follows from the definition of hq that ζ ∈ h2(ξ). To
finish, let us show the following:
Fact 2. ζ ∈ h2(ξ) ∗ h2(θ).
Proof of Fact 2. First suppose θ = δ2(ξ). If ζ 6∈ dom(δ2), then ζ /∈ h2(δ2(ξ)). If
ζ ∈ dom(δ2), from Fact 1 and the minimality of δ2(ζ), it follows that ζ /∈ h2(δ2(ξ)).
Since ξ ∈ h2(δ2(ξ)), we have that ζ ∈ h2(ξ) \ h2(δ2(ξ)) = h2(ξ) ∗ h2(θ).
1This case is similar to Subcase 2.2 in the proof of Claim 2.7.2 of [11].
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
9
Now suppose θ = δ2(ζ). Analogously we prove that ξ /∈ h2(δ2(ζ)) and ζ ∈
h2(ξ) ∩ h2(δ2(ζ)) = h2(ξ) ∗ h2(θ), concluding the proof of Fact 2.
Finally, since p2 ∈ Pf , there is γ ∈ i2({ξ, θ}) such that ζ ∈ h2(γ) ⊆ hq(γ). By
condition (B).(i), which can be used by Fact 1, we have that
i2({ξ, θ}) ⊆ f ({ξ, θ}) ∩ D2 ⊆ f ({ξ, η}) ∩ Dq = iq({ξ, η}).
Hence, γ ∈ iq({ξ, η}) and ζ ∈ hq(γ), concluding the proof of Subcase 4.2, Case 4
and thus concluding the proof of Claim 2.
Now we know that q ∈ Pf and let us check the other conclusions: it follows easily
from the definition of q and Lemma 2.5 that q ≤ p1 and analogously it follows from
the definition of q and Lemma 2.4 that q ≤ p2.
Finally, we verify the condition we want q to satisfy, that is, ξ ∈ h2(η)∪e−1[h2(η)]
let ξ ∈ D1 and η ∈ D2 and we consider again the
if and only if e(ξ) ∈ h2(η):
following cases:
Case 1. ξ ∈ D1 ∩ D2.
It follows from the fact that in this case e(ξ) = ξ.
Case 2. ξ ∈ D1 \ D2.
In this case, ξ ∈ h2(η) ∪ e−1[h2(η)] if and only if ξ ∈ e−1[h2(η)] if and only if
(cid:3)
e(ξ) ∈ h2(η), concluding the proof of the lemma.
3. The main results
To apply the key lemma proved in the previous section, the function f on which
the forcing Pf depends must satisfy a stronger version of the property ∆:
[ω2]2 → [ω2]≤ω has the strong property ∆ if
Definition 3.1. A function f :
f ({ξ, η}) ⊆ min{ξ, η} for all {ξ, η} ∈ [ω2]2 and for any uncountable ∆-system A
of finite subsets of ω2, there are distinct a, b ∈ A and an order-preserving bijection
e : a → b which is the identity on a ∩ b and such that ξ ≤ e(ξ) for all ξ ∈ a and for
any τ ∈ a ∩ b, any ξ ∈ a \ b and any η ∈ b \ a we have:
1) a ∩ min{ξ, η} ⊆ f ({ξ, η});
2) τ < ξ ⇒ f ({τ, η}) ⊆ f ({ξ, η});
3) τ < η ⇒ f ({τ, ξ}) ⊆ f ({ξ, η}).
Finally we arrive at the main result of this paper.
Theorem 3.2. If f : [ω2]2 → [ω2]≤ω has the strong property ∆, then V Pf satisfies
"for all n ∈ N, K n
f is hereditarily separable".
Proof. We prove this by induction on n ∈ N: in V Pf , fix n ∈ N and suppose that
for all 0 ≤ i < n, K i
f = {∗}) and let us show
that K n
f is hereditarily separable. We will be using a well-known fact that a regular
space is hereditarily separable if and only if it has no uncountable left-separated
sequence (see Theorem 3.1 of [24]).
f is hereditarily separable (take K 0
In V , suppose ( xα)α<ω1 is a sequence of names such that Pf forces that ( xα)α<ω1
f of cardinality ℵ1 and for each α < ω1, we have
is a left-separated sequence in K n
that xα = ( xα
n), where each xα
i
1 , . . . , xα
is a name for an element of Kf .
10
CHRISTINA BRECH AND PIOTR KOSZMIDER
Notice that if
Pf (cid:13) ∃1 ≤ i ≤ n, ∃X ⊆ ω1, X = ℵ1 such that ∀α, β ∈ X,
i = xβ
xα
i ,
then
Pf (cid:13) ∃1 ≤ i ≤ n, ∃X ⊆ ω1, X = ℵ1 such that (( xα
1 , . . . , xα
i−1, xα
i+1, . . . , xα
n))α∈X
is a left-separated sequence in K n−1
f
,
contradicting the inductive hypothesis. Therefore, we can assume without loss of
β and
generality that Pf forces that for all 1 ≤ i ≤ n and all α < β < ω1, xα
i
xα
i ∈ Lf = Kf \ {∗}.
6= xi
By assertion (+) following Definition 1.4, for each α < ω1, there are names
F α
1 , . . . , F α
n for finite subsets of ω2 such that Pf forces that
∀α < ω1 ∀1 ≤ i ≤ n xα
i ∈ h( xα
i ) \ [ξ∈ F α
i
h(ξ)
and
∀α < β < ω1 ∃1 ≤ i ≤ n xα
i /∈ h( xβ
i ) \ [ξ∈ F β
i
h(ξ).
For each α < ω1, let pα = (Dα, hα, iα) ∈ Pf , xα
1 , . . . , xα
n ∈ ω2 and F α
1 , . . . , F α
n ⊆
ω2 be finite such that
pα (cid:13) ∀1 ≤ i ≤ n xα
i = xα
i and F α
i = F α
i .
By Lemma 2.2 of [11], we can assume without loss of generality that for all
α < ω1 and all 1 ≤ i ≤ n, F α
i ⊆ Dα and xα
i ∈ Dα.
By the ∆-system Lemma, we can assume as well that (Dα)α<ω1 forms a ∆-
system with root D. Since for each pair {ξ, η} ⊆ D and each α < ω1, we have that
iα({ξ, η}) ∈ [f ({ξ, η})]<ω, we may assume that for all α < β < ω1, if ξ, η ∈ D,
ξ 6= η, then iα({ξ, η}) = iβ({ξ, η}).
By thinning out, we can assume without loss of generality that (Dα)α<ω1 forms
a ∆-system with root D such that for every α < β < ω1:
• pα is isomorphic to pβ;
• pα is lower than pβ;
• if eαβ : Dα → Dβ is the order-preserving bijective function, then eαβ(xα
i ) =
xβ
i , for all 1 ≤ i ≤ n.
Finally, we may assume that for all 1 ≤ i ≤ n we have: either xα
for all
i /∈ D for all α < ω1 and actually the second case holds by our
i = xβ
i
α < β < ω1; or xα
initial assumption about the sequence.
Since f has the strong property ∆, there are α < β < ω1 such that for all ζ ∈ D,
all ξ ∈ Dα \ D and all η ∈ Dβ \ D:
(i) Dα ∩ ξ ∩ η ⊆ f ({ξ, η});
(ii) if ζ < ξ, then f ({ζ, η}) ⊆ f ({ξ, η});
(iii) if ζ < η, then f ({ζ, ξ}) ⊆ f ({ξ, η}).
Note that pα and pβ satisfy the hypothesis of Lemma 2.7. Hence, there is
q ≤ pα, pβ in Pf such that for all ξ ∈ Dα and all η ∈ Dβ,
ξ ∈ hq(η) if and only if eαβ(ξ) ∈ hpβ (η).
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
11
Then, for all 1 ≤ i ≤ n and all ξ ∈ Dβ, we have that
i ∈ hq(ξ) if and only if xβ
xα
i ∈ hpβ (ξ).
So we have that
i ∈ hq(xβ
xα
i ) \ [ξ∈F β
i
hq(ξ).
But q ≤ pα, pβ and then
q (cid:13) ∀1 ≤ i ≤ n,
i = xα
xα
i ,
xβ
i = xβ
i and F β
i = F β
i .
Therefore,
q (cid:13) ∀1 ≤ i ≤ n,
contradicting the hypothesis about xα
xα
i = xα
i ∈ h(xβ
h(ξ) = h( xβ
i ) \ [ξ∈ F β
i and F β
i , xβ
i .
i
i ) \ [ξ∈ F β
i
h(ξ),
(cid:3)
Corollary 3.3. It is relatively consistent with ZFC that there is a hereditarily
separable compact scattered space of height ω2.
Proof. Since each level of the Cantor-Bendixson decomposition of Kf is a discrete
subset of Kf , it follows that every level of it is countable. But Kf = ℵ2 and Kf =
f =
(cid:3)
, so that ht(Kf ) ≥ ω2. It is easy to see that Tα<ω2 K (α)
Sα<ht(Kf ) K (α)
{∗} concluding that ht(Kf ) = ω2.
f \ K (α+1)
f
Corollary 3.4. It is relatively consistent with ZFC that there is a hereditarily
separable compact space with hereditary Lindelof degree equal to ℵ2. In particular, it
is relatively consistent with ZFC that there is a compact space K such that hL(K) 6≤
hd(K)+".
Proof. It follows from the fact that hL(Kf ) ≤ Kf = ℵ2 and that {Kf \ K (α)
: α <
ω2} is an open covering of Kf \ {∗} which does not admit a subcovering of strictly
smaller cardinality.
(cid:3)
f
Corollary 3.5. It is relatively consistent with ZFC that there is an Asplund space
C(K) of density ℵ2 which does not admit any Fr´echet smooth renorming and which
does not contain an uncountable biorthogonal system.
Proof. Since every finite power of Kf is hereditarily separable, Lemma 4.37 and
Theorem 4.38 of [6] imply that C(Kf ) is hereditarily Lindelof relative to its point-
wise convergence topology. But for compact scattered spaces K, the pointwise
convergence topology and the weak topology of C(K) coincide (see Theorem 7.4 of
[20]), so that C(Kf ) is hereditarily Lindelof relative to its weak topology.
Now, if C(Kf ) admits a Fr´echet smooth renorming, by Corollaries 8.34 (due to
Mazur [18]) and 8.36 of [6] (due to Jim´enez Sevilla and Moreno [10]) it contains an
uncountable bounded subset A such that for every x0 ∈ A, x0 is not in the (norm-)
closed convex hull of A \ {x0}, that is, x0 /∈ conv(A \ {x0}). Since the weak and
norm convex closures coincide in Banach spaces, A turns out to be an uncountable
discrete family of C(Kf ) relative to its weak topology, which contradicts the fact
that C(Kf ) is hereditarily Lindelof relative to its weak topology.
Now,
if C(Kf ) admits an uncountable biorthogonal system (xα, ϕα)α<ω1 ⊆
C(Kf ) × C(Kf )∗, then {xα : α < ω1} is an uncountable discrete family of C(Kf )
12
CHRISTINA BRECH AND PIOTR KOSZMIDER
relative to its weak topology, contradicting the fact that C(Kf ) is hereditarily Lin-
delof relative to its weak topology.
(cid:3)
One should compare the above corollary to Theorem 4.41 of [6] (due to Os-
taszewski [21]) and to Corollary 8.37 of [6] (due to Jim´enez Sevilla and Moreno
[10]).
4. The existence of the required function f
In this section we prove the consistency of the existence of a function with the
It turns out that we are even able to prove the consistency
strong property ∆.
of the existence of such a function with its range included in the family of finite
(rather than countable) subsets of ω2. The method is quite involved but, as shown
at the end of this section, forcings preserving CH (as in [2]) cannot serve for this
purpose even if we were interested in a function with its range included in countable
subsets of ω2.
4.1. Forcing with side conditions in Velleman's simplified morasses. To
construct a forcing which adds the required auxiliary function on pairs of ω2 we
will need a family of countable subsets of ω2 with some strong properties. The
following proposition establishes a list of the most useful properties:
Proposition 4.1. It is relatively consistent with ZFC+CH that there exists a family
F ⊆ [ω2]ω which satisfies the following properties:
1) (F , ⊆) is well-founded (thus, one can talk about rank(X) for X ∈ F );
2) F is stationary in [ω2]ω (see [1]);
3) If α ∈ X, Y ∈ F and rank(X) ≤ rank(Y ), then X ∩ α ⊆ Y ∩ α;
If M is a countable elementary submodel of H(ω3) containing ω1, ω2, F and X =
M ∩ ω2 ∈ F , then
4) M ∩ ω1 = rank(X);
5) Y ⊂ X, Y ∈ F implies Y ∈ M ;
6) X1, ..., Xn ∈ F for n ∈ N and rank(Xi) < rank(X) for 1 ≤ i ≤ n implies that
there is Z ∈ F such that Z ∈ M and X ∩ (X1 ∪ ... ∪ Xn) ⊆ Z.
Proof. We will prove that a simplified Velleman's (ω1, 1)-morass (see [30]) which
is a stationary coding set (see [31]) satisfies the above properties. The proof relies
heavily on the properties of Velleman's morasses obtained in [15]. We will often
refer to this paper, in particular we adopt definitions of simplified morass and
stationary coding set from this paper (section 2). The consistency of the existence
of such morasses can be immediately obtained from the corresponding proof for
semimorasses in [14], Theorem 3 Section 2.
1) follows from Definition 2.1 of [15] and 2) from the fact that F is assumed to
be a stationary coding set. To prove 3) apply 2.5 of [15]. Now 4) is Fact 2.7 of [15],
5) is Fact 2.6 of [15] To obtain 6) apply Fact 2.8 of [15] to each Xi obtaining Z(Xi)
such that Z(Xi) ∈ M ∩ F and Xi ∩ X ⊆ Z(Xi). Now use the elementarity of M
and the directedness of F (see Definition 2.1. of [15]) to obtain Z as in 6).
(cid:3)
Now we will adopt a few facts from [16] and [15] concerning forcing with side
conditions in F . As explained in these papers, to use elements of F as side con-
ditions means to use forcings P whose conditions are of the form (p, A) where p
is a finite condition of a natural forcing adding the structure in question and A is
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
13
a finite subset of F . This is like using models as side conditions in the method
of forcing with models as side conditions developed by Todorcevic (see [27]). The
order is given by the forcing order on the first coordinate and inverse inclusion on
the second coordinate. In addition we require the existence of some natural projec-
tions of p onto the elements of A as a part of the definition of the forcing notion.
The properties 1) - 6) above allow us to perform many maneuvers with ease; also
the definitions are simpler. This method appears to be equivalent to the variant
of Todorcevic's method where one employs matrices of models (see [28] Section 4,
for an example with detailed definitions). The price we need to pay for this con-
venience is that P is not proper (unlike Todorcevic's forcings,) but only F -proper,
i.e., there is a club C ⊆ [ω2]ω such that for models M ≺ H(ω3) such that M ∈ F ∩ C
and p ∈ P ∩ M , there are (P, M )-generic conditions stronger than p. As F may
be assumed to be stationary, F -properness implies the preservation of ω1 (proof as
for proper forcings, see [1]). The preservation of bigger cardinals follows from the
ω2-chain condition. Note that the fact that the forcing is not proper but preserves
cardinals is no limitation in the applications that one seeks here, i.e., consistent
existence of structures of sizes bigger than ω1. Let us describe basic notions related
to forcing with side conditions in F that we will use.
Definition 4.2. Suppose F ⊆ [ω2]ω. We say that a forcing notion P is F -proper if
there is θ > (2P )+ and a club set C ⊆ [H(θ)]ω such that whenever p ∈ M ∈ C and
M ∩ ω2 ∈ F then there is a (P, M )-generic p0 ≤ p, i.e., D ∩ M is predense below
p0 for every D ∈ M which is dense in P .
Fact 4.3. Suppose F ⊆ [ω2]ω is a stationary set and P is an F -proper forcing
notion, then P preserves ω1.
Proof. The proof is a straightforward version of Shelah's paradigmatic proof of
preservation of ω1 by proper forcings (see [26] or [1]).
(cid:3)
The following definition and lemmas are formulations of well-known techniques
(originated in Shelah's use of elementary submodels in forcing) and will simplify
our further arguments.
Definition 4.4. Let P be a notion of forcing, q ∈ P and let θ > (2P )+. Suppose
M ≺ H(θ) and P, π1, ..., πk ∈ M . We say that a formula φ(x0, x1, ..., xk) well
reflects q in (M ; π1, ..., πk) whenever the following are satisfied:
i) φ(q, π1, ..., πk) holds in H(θ);
ii) whenever s ∈ M is such that φ(s, π1, ..., πk) holds in M , then q and s are
compatible.
Definition 4.5. Suppose F ⊆ [ω2]ω and suppose P is a notion of forcing. We say
that P is simply F -proper if there is θ such that whenever
a) p ∈ P ,
b) M ≺ H(θ), M countable,
c) p, P, F ∈ M ,
d) M ∩ ω2 ∈ F ,
then there is p0 ≤ p such that if q ≥ p0, then there are π1, ..., πk ∈ M and a formula
φ(x0, x1, ..., xk) which well reflects q in (M, π1, ..., πk).
Lemma 4.6. If P is simply F -proper, then P is F -proper.
14
CHRISTINA BRECH AND PIOTR KOSZMIDER
Proof. We will prove that whenever M, p are as in a) - d) of Definition 4.5, then
p0 is a (P, M )-generic condition. Let D ∈ M be dense, we will show that D ∩ M
is predense below p0. Let q ≤ p0, we may w.l.o.g. assume that q ∈ D. Let
π1, ..., πk ∈ M and φ(x0, x1, ..., xk) be such that φ(x0, x1, ..., xk) well reflects q in
(M, π1, ..., πk). By i) of Definition 4.4, we have φ(q, π1, ...πk) in H(θ). By its
elementarity, M satisfies the formula "∃x ∈ P φ(x, π1, ...πk) & x ∈ D". So let
s ∈ M witness this fact. Now by Definition 4.4.ii), s and q are compatible, so
D ∩ M contains a condition compatible with q which proves that D ∩ M is predense
below q which completes the proof.
(cid:3)
4.2. Adding a function with the strong property ∆. Fix a family F ⊆ [ω2]ω
satisfying 1) - 6) of Proposition 4.1. We will assume familiarity of the reader with
elementary submodels of structures H(θ). In particular we will make use of facts
such as that countable elements of such models are their subsets or that such models
contain ω. See [5] for more on this subject. We consider the following forcing P
whose conditions p are of the form: p = (ap, fp, Ap) where
a) ap ∈ [ω2]<ω;
b) fp : [ap]2 → [ω2]<ω;
c) Ap ∈ [F ]<ω;
d) fp(α, β) ⊆ T{X : X ∈ Ap, α, β ∈ X} ∩ min{α, β} for any distinct α, β ∈ ap.
The order is just the inverse inclusion, i.e., p ≤ q if and only if ap ⊇ aq, fp ⊇ fq,
Ap ⊇ Aq.
Fact 4.7. P is simply F -proper.
Proof. Let θ = ω3 and let M and p be as in a) - d) of Definition 4.5. The existence
of such an M follows from the stationarity of F . Let X0 = M ∩ ω2. Let p0 =
(ap, fp, Ap ∪ {X0}). Finally let q ≤ p0. The proof consists of using Lemma 4.6 and
finding the parameters π1, ..., πk ∈ M and a formula φ(x0, x1, ..., xk) which well
reflects q in (M, π1, ..., πk).
Define qM = (aq ∩ M, fqM, Aq ∩ M ).
Introduce notation δ = M ∩ ω1 =
rank(M ), where the second equality follows from 4) of Proposition 4.1. Note that
Aq ∩ M = AqM = {X ∈ Aq : X ⊂ X0}. This follows from 5) of Proposition 4.1.
The fact that [M ]<ω ⊆ M implies that aqM , AqM ∈ M . Also as d) of the definition
of the forcing is satisfied for q and α, β ∈ aq, we have that fq(α, β) ⊆ X0 = M ∩ ω2
for α, β ∈ aq ∩ X0. So, we may conclude that fqM ∈ M , in other words we have
qM ∈ M ∩ P . It is clear that qM ≤ p. By 6) of Proposition 4.1 and the fact that
[M ]<ω ⊆ M , in M there is a Z ∈ F such that S{X ∩ M : rank(X) < δ, X ∈
Aq} ⊆ Z. Let φ(x0, x1, x2, x3, x4) be the formula which says that x0 is a condition
of the partial order x4 which extends in x4 the condition x3 and such that the
difference between the first coordinate of x0 and x2 is disjoint from x1.
Claim. φ(x0, x1, x2, x3, x4) well-reflects q in (M, Z, aqM , qM, P ).
It is clear that φ(q, Z, aqM , qM, P ) holds in H(ω3). Now
Proof of the Claim.
let s ∈ M be a condition satisfying φ(s, Z, aqM , qM, P ) i.e., s extends in P the
condition qM and as \ aqM is disjoint from Z. Define the common extension r of q
and s as follows: ar = as ∪ aq, fr = fs ∪ fq ∪ h, Ar = As ∪ Aq, where h({α, β}) = ∅
for {α, β} ∈ [as ∪ aq]2 − ([as]2 ∪ [aq]2). Such an fr is a function on [ar]2 since
qM ≥ q, s. Clearly all clauses of the definition of the forcing P but d) are trivially
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
15
satisfied by r. So let us prove d). Let α, β ∈ ar and X ∈ Ar, we will consider a few
cases.
Case 1. α, β ∈ as, X ∈ As
It is trivial because s ∈ P .
Case 2. α, β ∈ aq, X ∈ Aq
It is trivial because q ∈ P .
Case 3. α, β ∈ as, X ∈ Aq.
Since φ(s, Z, aqM , qM, P ) holds in M we have that either rank(X) ≥ δ =
rank(M ∩ ω2) = rank(X0) in which case d) is satisfied because fr({α, β}) =
fs({α, β}) ⊆ X0 ∩ min{α, β} ⊆ X ∩ min{α, β} by d) for s and 3) of Proposition 4.1
or rank(X) < δ and then by the definition of φ and Z we get that α, β ∈ as ∩ aq,
so we are again in Case 2.
Case 4. α, β ∈ aq, X ∈ As.
This means that α, β ∈ M , because s ∈ M , i.e., α, β ∈ as ∩ aq so we are again
in Case 1.
Case 5. α ∈ as \ aq and β ∈ aq \ as.
Then h({α, β}) = ∅.
The proof of the claim completes the proof of Fact 4.7.
(cid:3)
Definition 4.8. For p ∈ P , call the set ap ∪ f [[ap]2] ∪S Ap the support of p and
denote it by supp(p).
Definition 4.9. We say that two conditions p, q of P are isomorphic (via π :
supp(p) → supp(q)) if π : supp(p) → supp(q) is an order preserving bijection
constant on supp(p) ∩ supp(q) and
i) π[ap] = aq;
ii) {π[X] : X ∈ Ap} = Aq;
iii) fq({π(α), π(β)}) = π[fp({α, β})] for all α, β ∈ ap.
Lemma 4.10. Suppose p, q ∈ P are isomorphic via π : supp(p) → supp(q). Then
they are compatible.
Proof. Define the common extension r of p and q as follows: ar = ap ∪ aq, fr =
fp ∪fq ∪h, Ar = Ap ∪Aq, where h({α, β}) = ∅ for {α, β} ∈ [ap ∪aq]2 −([ap]2 ∪[aq]2).
The only non-automatic condition of the definition of P which needs to be checked
is d).
Case 1. α, β ∈ ar.
If X ∈ Ar, we are trivially done.
If X ∈ Aq and α, β ∈ X, then α, β ∈
supp(p) ∩ supp(q), hence α, β ∈ ap ∩ aq and hence again use d) for q.
Case 2. α, β ∈ aq.
Similar to the previous case.
Case 3. α ∈ ar \ aq, β ∈ aq \ ar.
In this case h is empty.
(cid:3)
16
CHRISTINA BRECH AND PIOTR KOSZMIDER
Fact 4.11. Assuming CH the forcing P is ω2-c.c. Thus by Fact 4.7, Lemma 4.6
and Fact 4.3, P preserves cardinals.
Proof. By the previous lemma the proof is a standard application of the ∆-system
lemma to the sequence of supports {supp(pξ) : ξ < ω2} of some conditions pξ ∈ P
under our cardinal arithmetic assumption.
(cid:3)
Theorem 4.12. In V P there is a function f : [ω2]2 → [ω2]<ω with the strong
property ∆.
Proof. Clearly, we claim that f = S{fp : p ∈ G} defines such a function, where G
is a P -generic over V . Let f be a name for it.
Fix a set A = { aα : α < ω1} of P -names for elements of an uncountable ∆-system
of n-tuples a = { ai : i < n} of elements of ω2 for which there are bijections e as in
Definition 3.1 (any uncountable ∆-system has an uncountable such a subsystem).
Fix a condition p ∈ P .
Take a model M ≺ H(ω3) such that M ∩ ω2 = X0 ∈ F and p ∈ P ∩ M ; F ∈ M
and { aα : α < ω1} ∈ M . We will show that there are α1 < α2 < ω1 and r ≤ p such
that r forces 1), 2), 3) of Definition 3.1 for aα1 and aα2 .
First take a condition p0 ≤ p as in Fact 4.7, i.e., ap0 = ap, fp0 = fp, Ap0 = Ap ∪
X0. Take q ≤ p0 and α1 ∈ ω1 such that there is b such that b \ M 6= ∅, q (cid:13) aα1 = b
and b ⊆ aq. This can be done as { aα : α < ω1} is a sequence of names for an
uncountable ∆-system of sets and M = ω. Proceed as in the proof of Fact 4.7,
i.e., choose Z and φ as in Fact 4.7. So, we have φ(q, Z, aqM , qM, P ) in H(ω3) and
so by the elementarity of M , we can find an s and α2 such that φ(s, Z, aqM , qM, P )
holds in M and moreover there is a such that a \ (b ∩ M ) ∈ [M \ Z]<ω and such
that s (cid:13) aα2 = a and a ⊆ as. Now we will obtain another amalgamation r of s and
q which will force 1), 2) and 3) of Definition 3.1. Let ar = as ∪ aq, fr = fs ∪ fq ∪ h.
For ξ ∈ as \ aq and η ∈ aq \ as:
∗∗)
where
h({ξ, η}) = [A ∪ B ∪ C] ∩ D
A = a ∩ min{ξ, η}
B = [{fs({τ, ξ}) : τ ∈ a ∩ b, τ < η}
C = [{fq({τ, η}) : τ ∈ a ∩ b, τ < ξ}
D = min{ξ, η} ∩\{X ∈ Aq : ξ, η ∈ X, rank(X) ≥ δ}
First let us check that r is a common extension of q and s. The proof also
follows the cases as in the Claim in the proof of Fact 4.7. All are checked in the
same manner except for Case 5 where one may assume that X ∈ Aq as β 6∈ M .
This time the inclusion in the set D guarantees that d) holds in Case 5.
Now we will check 1), 2) and 3) of Definition 3.1 for a, b as above and fr. This
will be enough since r (cid:13) aα1 = b, aα2 = a and r (cid:13) fr ⊆ f . Suppose ξ ∈ a \ b and
η ∈ b \ a. By the form of the definition of fr({ξ, η}) = h({ξ, η}) it will be enough
to prove that the sets A, B and C are actually included in min{ξ, η} ∩ X for any
X ∈ Aq such that rank(X) ≥ δ and ξ, η ∈ X. So, let X ∈ Aq be any such element
that rank(X) ≥ δ and ξ, η ∈ X.
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
17
For 1) of Definition 3.1, note that since X0 = M ∩ ω2 and rank(X0) = δ we have
that M ∩ min{ξ, η} is included in X by 3) of Proposition 4.1. Hence, as a ⊆ M , we
have a ∩ min{ξ, η} ⊆ min{ξ, η} ∩ X, that is we obtain 1).
To get 2) of Definition 3.1 assume that τ ∈ a ∩ b and τ < ξ, hence min{τ, η} ≤
min{ξ, η}. Note again, by 3) of Proposition 4.1, that M ∩ ξ ⊆ X ∩ ξ which implies
in this case that τ ∈ X. Hence, since τ, η ∈ aq, by d) of the definition of the forcing,
we have that fq({τ, η}) ⊆ X ∩ min{τ, η} ⊆ X ∩ min{ξ, η}, so we obtain 2).
To get 3) of Definition 3.1 assume that τ ∈ a ∩ b and τ < η, hence min{τ, ξ} ≤
min{ξ, η}. We have that τ, ξ ∈ M ∩ ξ, and again M ∩ ξ ⊆ X. Hence, since ξ, τ ∈ as,
by d) of the definition of the forcing, and the fact that s ∈ M , we have that
fs({τ, ξ}) ∩ min{τ, ξ} ⊆ X ∩ min{τ, ξ} ⊆ X ∩ min{ξ, η}, so we obtain 3).
(cid:3)
Remark 4.13. For any k ≤ ω and any uncountable ∆-system one can have k sets
satisfying 1), 2) and 3) of Definition 3.1. This follows from the Dushnik-Miller
theorem see [9] Theorem 9.7.
4.3. CH and the strong property ∆. In this section we prove that CH implies
that there is no function f such as in the previous section, even if we allow f to
take countable sets as values. This also proves that the strong property ∆ cannot
be obtained as in Baumgartner and Shelah [2], that is, by a forcing which preserves
CH.
Proposition 4.14. (CH) There is no f : [ω2]2 → [ω2]ω such that for every ∆-
system A of finite subsets of ω2 of cardinality ℵ1, there exist distinct a, b ∈ A such
that
∗)
∀ξ ∈ a \ b∀η ∈ b \ a
a ∩ ξ ∩ η ⊆ f ({ξ, η}).
Proof. Suppose that f : [ω2]2 → [ω2]ω. For an A ⊆ ω2 and ξ ∈ ω2 \ A define
fA,ξ : A → [A]ω,
fA,ξ(η) = f ({η, ξ}) ∩ A ∀η ∈ A.
Let M ≺ H(ω3) be closed under its countable subsets (here we use CH) M = ω1,
ω1 ⊆ M ; ω1, ω2, f ∈ M and such that sup(M ∩ ω2) = γ has an uncountable
cofinality.
By recursion construct a sequence (αξ, βξ)ξ<ω1 which satisfies:
1) ω1 < αξ, βξ ∈ M ∩ γ;
2) αξ < βξ < αη for all ξ < η;
3) fAη,βη = fAη,γ where Aη = {αξ, βξ : ξ < η};
4) αη 6∈ f ({βη, γ}).
To justify that this construction can be carried out assume that we have Aη
satisfying 1)-4) and let us show how to obtain αη, βη. As Aη ⊆ M and M is closed
under its countable sets we have Aη ∈ M . Also fAη,γ ∈ M as M is closed under
countable sets. Hence, by the elementarity there is βη ∈ M \ sup(Aη) such that
fAη,βη = fAη ,γ and cf (βη) = ω1. Now f ({βη, γ}) ∩ M is in M again, so using
the fact that cf (βη) = ω1 we can find αη ∈ M satisfying Aη < αη < βη and
αη 6∈ f ({βη, γ}) which completes the construction.
Now define γη < ω1 such that for ξ < η < ω1 we have
γξ < γη 6∈ [{f ({βξ, βη}) : ξ < η}.
18
CHRISTINA BRECH AND PIOTR KOSZMIDER
Finally define A = {{γξ, αξ, βξ} : ξ < ω1}. Suppose ξ < η. Note that, as by 4),
αξ 6∈ f ({βξ, γ}) and by 3), f ({βξ, γ}) = f ({βξ, βη}), we have
αξ ∈ (βξ ∩ βη) \ f ({βξ, βη}).
But on the other hand, by the definition of γη, we have
γη ∈ (βξ ∩ βη) \ f ({βξ, βη}),
which shows that the inclusion *) of the proposition holds for no a, b ∈ A.
(cid:3)
References
1. J. E. Baumgartner, Applications of the proper forcing axiom, Handbook of set-theoretic topol-
ogy, North-Holland, Amsterdam, 1984, pp. 913 -- 959.
2. J. E. Baumgartner and S. Shelah, Remarks on superatomic Boolean algebras, Ann. Pure Appl.
Logic 33 (1987), no. 2, 109 -- 129.
3. J. M. Borwein and J. D. Vanderwerff, Banach spaces that admit support sets, Proc. Amer.
Math. Soc. 124 (1996), no. 3, 751 -- 755.
4. R. Deville, G. Godefroy, and V. Zizler, Smoothness and renormings in Banach spaces, Pitman
Monographs and Surveys in Pure and Applied Mathematics, vol. 64, Longman Scientific &
Technical, Harlow, 1993.
5. A. Dow, An introduction to applications of elementary submodels to topology, Topology Proc.
13 (1988), no. 1, 17 -- 72.
6. P. H´ajek, V. Montesinos Santaluc´ıa, J. Vanderwerff, and V. Zizler, Biorthogonal systems in
Banach spaces, CMS Books in Mathematics/Ouvrages de Math´ematiques de la SMC, 26,
Springer, New York, 2008.
7. R. Haydon, A counterexample to several questions about scattered compact spaces, Bull. Lon-
don Math. Soc. 22 (1990), no. 3, 261 -- 268.
8. R. Hodel, Cardinal functions. I, Handbook of set-theoretic topology, North-Holland, Amster-
dam, 1984, pp. 1 -- 61.
9. T. Jech, Set theory, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2003, The
third millennium edition, revised and expanded.
10. M. Jim´enez Sevilla and J.-P. Moreno, Renorming Banach spaces with the Mazur intersection
property, J. Funct. Anal. 144 (1997), no. 2, 486 -- 504.
11. I. Juh´asz and L. Soukup, How to force a countably tight, initially ω1-compact and noncompact
space?, Topology Appl. 69 (1996), no. 3, 227 -- 250.
12. I. Juh´asz and W. Weiss, On thin-tall scattered spaces, Colloq. Math. 40 (1978/79), no. 1,
63 -- 68.
13. W. Just, Two consistency results concerning thin-tall Boolean algebras, Algebra Universalis
20 (1985), no. 2, 135 -- 142.
14. P. Koszmider, Semimorasses and nonreflection at singular cardinals, Ann. Pure Appl. Logic
72 (1995), no. 1, 1 -- 23.
15.
16.
, On strong chains of uncountable functions, Israel J. Math. 118 (2000), 289 -- 315.
, Universal matrices and strongly unbounded functions, Math. Res. Lett. 9 (2002),
no. 4, 549 -- 566.
17. J. C. Mart´ınez, A consistency result on thin-very tall Boolean algebras, Israel J. Math. 123
(2001), 273 -- 284.
18. S. Mazur, Uber schwache Konvergenz in den Raumen Lp, Studia Math. 4 (1933), 129 -- 133.
19. I. Namioka and R. R. Phelps, Banach spaces which are Asplund spaces, Duke Math. J. 42
(1975), no. 4, 735 -- 750.
20. S. Negrepontis, Banach spaces and topology, Handbook of set-theoretic topology, North-
Holland, Amsterdam, 1984, pp. 1045 -- 1142.
21. A. J. Ostaszewski, A countably compact, first-countable, hereditarily separable regular space
which is not completely regular, Bull. Acad. Polon. Sci. S´er. Sci. Math. Astronom. Phys. 23
(1975), no. 4, 431 -- 435.
22. M. Rabus, An ω2-minimal Boolean algebra, Trans. Amer. Math. Soc. 348 (1996), no. 8,
3235 -- 3244.
23. M. Rajagopalan, A chain compact space which is not strongly scattered, Israel J. Math. 23
(1976), no. 2, 117 -- 125.
THIN-VERY TALL COMPACT SCATTERED SPACES WHICH ARE HS
19
24. J. Roitman, Introduction to modern set theory, Pure and Applied Mathematics (New York),
John Wiley & Sons Inc., New York, 1990, A Wiley-Interscience Publication.
25. B. `E. Shapirovskiı, Cardinal invariants in compacta, Seminar on General Topology, Moskov.
Gos. Univ., Moscow, 1981, pp. 162 -- 187.
26. S. Shelah, Proper forcing, Lecture Notes in Mathematics, vol. 940, Springer-Verlag, Berlin,
1982.
27. S. Todorcevic, A note on the proper forcing axiom, Axiomatic set theory (Boulder, Colo.,
1983), Contemp. Math., vol. 31, Amer. Math. Soc., Providence, RI, 1984, pp. 209 -- 218.
28.
29.
, Directed sets and cofinal types, Trans. Amer. Math. Soc. 290 (1985), no. 2, 711 -- 723.
, Biorthogonal systems and quotient spaces via Baire category methods, Math. Ann.
335 (2006), no. 3, 687 -- 715.
30. D. Velleman, Simplified morasses, J. Symbolic Logic 49 (1984), no. 1, 257 -- 271.
31. W. S. Zwicker, Pkλ combinatorics. I. Stationary coding sets rationalize the club filter, Ax-
iomatic set theory (Boulder, Colo., 1983), Contemp. Math., vol. 31, Amer. Math. Soc., Prov-
idence, RI, 1984, pp. 243 -- 259.
IMECC-UNICAMP, Caixa Postal 6065, 13083-970, Campinas, SP, Brazil
E-mail address: [email protected]
Instytut Matematyki; Politechnika L´odzka, ul. W´olcza´nska 215; 90-924 L´od´z, Poland
E-mail address: [email protected]
|
1705.02188 | 1 | 1705 | 2017-05-05T12:21:07 | Estimates for Tsallis relative operator entropy | [
"math.FA"
] | We give the tight bounds of Tsallis relative operator entropy by using Hermite-Hadamard's inequality. Some reverse inequalities related to Young inequalities are also given. In addition, operator inequalities for normalized positive linear map with Tsallis relative operator entropy are given. | math.FA | math |
M athematical
I nequalities
& A pplications
www.ele-math.com
ESTIMATES FOR TSALLIS RELATIVE OPERATOR ENTROPY
HAMID REZA MORADI, SHIGERU FURUICHI AND NICUS¸ OR MINCULETE
Submitted to Math. Inequal. Appl.
Abstract. We give the tight bounds of Tsallis relative operator entropy by using Hermite-Hadamard's
inequality. Some reverse inequalities related to Young inequalities are also given. In addition,
operator inequalities for normalized positive linear map with Tsallis relative operator entropy are
given.
1. Introduction and Preliminaries
The operator theory related to inequalities in Hilbert space is studied in many
papers. Let A, B be two operators in a Hilbert space H . An operator A is said to
be strictly positive (denoted by A > 0 ) if A is positive and invertible. For two strictly
positive operators A, B and p ∈ [0, 1], p -power mean A#pB is defined by
A#pB := A
2 BA− 1
1
2 ,
A
1
2(cid:16)A− 1
2(cid:17)p
and we remark that A#pB = A1−pBp if A commutes with B . We also use the symbol
2 for r ∈ R . The weighted operator arithmetic mean is
A♮rB := A
defined by
2 BA− 1
2(cid:16)A− 1
2(cid:17)r
A
1
1
A∇pB := (1− p)A + pB,
for any p ∈ [0, 1] .
The relative operator entropy S (AB) in [5] is defined by
2(cid:17) A
1
2 ,
2(cid:16)log A− 1
S (AB) := A
2 BA− 1
1
where A and B are two invertible positive operators on a Hilbert space. As a parametric
extension of the relative operator entropy, Yanagi, Kuriyama and Furuichi [19] defined
Tsallis relative operator entropy which is an operator version of Tsallis relative entropy
in quantum system due to Abe [1], also see [9]:
A
Tp (AB) :=
1
2(cid:16)A− 1
2 BA− 1
p
2(cid:17)p
A
1
2 − A
,
p ∈ (0, 1].
Mathematics subject classification (2010): Primary 47A63, Secondary 46L05, 47A60.
Keywords and phrases: Relative operator entropy; Tsallis relative operator entropy; operator inequal-
ity; Hermite-Hadamard's inequality; Young's inequality.
The authors would like to thank editor and referee for providing valuable comments to improve our manuscript. The
author (S.F.) was partially supported by JSPS KAKENHI Grant Number 16K05257.
1
Notice that Tp (AB) can be written by using the notation of A#pB as follows:
Tp (AB) :=
A#pB− A
p
,
p ∈ (0, 1].
The relation between relative operator entropy S (AB) and Tsallis relative opera-
tor entropy Tp (AB) was considered in [9, 20], as follows:
A− AB−1A ≤ T−p (AB) ≤ S (AB) ≤ Tp (AB) ≤ B− A.
(1)
Some deeper properties of Tsallis relative operator entropy were proved in [6, 11,
13, 20].
The main result of the present paper is a set of bounds that are complementary to
(1). Some of our inequalities improve well-known ones. Among other inequalities, it is
shown that if A, B are invertible positive operators and p ∈ (0, 1], then
2 BA− 1
2 + I
A
1
2 A− 1
2
!p−1
2 BA− 1
(cid:16)A− 1
2 − I(cid:17) A
1
2 ≤ Tp (AB)
1
2
≤
(A#pB− A♮p−1B + B− A),
which is a considerable refinement of (1), where I is the identity operator. We also
prove a reverse inequality involving Tsallis relative operator entropy Tp (AB) .
2. Refinements of the inequalities (1) via Hermite-Hadamard inequality
An important ingredient in our approach is the following:
Let f be a convex function on [a, b] ⊆ R ; the well-known Hermite-Hadamard's
inequality can be expressed as
f(cid:18) a + b
2 (cid:19) ≤
1
b− a
b
Za
f (t) dt ≤
f (a) + f (b)
2
.
(2)
THEOREM 1. For any invertible positive operator A and B such that A ≤ B , and
p ∈ (0, 1] we have
2 BA− 1
1
A
2 A− 1
2
2 + I
!p−1
(cid:16)A− 1
2 BA− 1
2 − I(cid:17)A
1
2 ≤ Tp (AB)
(3)
1
2
≤
(A#pB− A♮p−1B + B− A).
2
Proof. Consider the function f (t) = t p−1, p ∈ (0, 1]. It is easy to check that f (t)
is convex on [1, ∞) . Bearing in mind the fact
x
Z1
t p−1dt =
xp − 1
p
,
and utilizing the left-hand side of Hermite-Hadamard inequality, one can see that
2 (cid:19)p−1
(cid:18) x + 1
(x− 1) ≤
xp − 1
p
,
(4)
where x ≥ 1 and p ∈ (0, 1]. On the other hand, it follows from the right-hand side of
Hermite-Hadamard inequality that
p ≤(cid:18) xp−1 + 1
xp − 1
2
(cid:19) (x− 1),
for each x ≥ 1 and p ∈ (0, 1].
2 BA− 1
Replacing x by A− 1
get the desired result (3).
2 in (4) and (5), and multiplying A
(5)
1
2 on both sides, we
REMARK 1. Simple calculation gives for all x ≥ 1 and p ∈ (0, 1] ,
1
x ≤(cid:18) x + 1
2 (cid:19)p−1
(x− 1) ≤
p ≤(cid:18) xp−1 + 1
xp − 1
2
(cid:19) (x− 1) ≤ x− 1,
(6)
0 ≤ 1−
which means
0 ≤ A− AB−1A ≤ A
1
2 A− 1
≤ Tp (AB)
≤
≤ B− A,
1
2
2 BA− 1
2 + I
2
!p−1
2 BA− 1
(cid:16)A− 1
1
2
2 − I(cid:17) A
(A#pB− A♮p−1B + B− A)
for any invertible positive operators A and B such that A ≤ B , and p ∈ (0, 1] . There-
fore, our inequalities (3) improve the inequalities (1) for the case A ≤ B.
PROPOSITION 1. For x ≥ 1 and 1
2 ≤ p ≤ 1 ,
2 (cid:19)p−1
√x ≤(cid:18) x + 1
x− 1
(x− 1).
(7)
Proof. In order to prove (7), we set the function f p (x) ≡(cid:0) x+1
where
d p =(cid:0) x+1
d p ≥ 0 for x ≥ 1 . Thus,
√2x−√x+1
√x(x+1) ≥ 0 for x ≥ 1 . Therefore, we have the inequality
ln(cid:0) x+1
2 (cid:1) ,
x ≥ 1 and 1
2 ≤ p ≤ 1 . Since
we have f p(x) ≥ f1/2(x) =
(7).
2 (cid:1)p−1
− 1√x
2 (cid:1)p−1
d f p(x)
d f p(x)
3
REMARK 2. The first inequality (3) gives tight lower bound for the Tsallis relative
operator entropy Tp(AB) more than the eighth inequality in [8, Theorem 2.8 (i)], due
to Proposition 1.
COROLLARY 1. For any invertible positive operators A and B such that A ≥ B ,
and p ∈ (0, 1] , we have
A#pB− A♮p−1B ≤
1
2
Proof. Put t = 1
≤ A
1
2 + I
2 BA− 1
(A#pB− A♮p−1B + B− A) ≤ Tp (AB)
2 A− 1
(cid:16)A− 1
x ≤ 1 in the inequalities (6).
!p−1
2 BA− 1
2
2 − I(cid:17) A
1
2 ≤ A♮p+1B− A#pB ≤ 0.
REMARK 3. By numerical computations, we have (cid:16) xp−1+1
3 , x = 0.1 . Thus we do not have ordering between (cid:16) xp−1+1
2 (cid:17) (x − 1) − 2(x−1)
x+1 ≃
3 , x = 0.01 , and we also have (cid:16) xp−1+1
x+1 ≃ 0.216868
2 (cid:17) (x− 1) and 2(x−1)
2 ≤ p ≤ 1 so that there is no ordering between the second inequality
−0.83219 for p = 2
for p = 2
for 0 < x ≤ 1 and 1
2 (cid:17) (x− 1)− 2(x−1)
in Corollary 1 and the sixth inequality in [8, Theorem 2.8 (ii)].
x+1
THEOREM 2. For any invertible positive operators A and B such that A ≤ B and
p ∈ (0, 1] , we have the following inequalities,
Lp (A, B) + Kp (A, B) ≤ Tp (AB) ≤ Rp (A, B) + Kp (A, B) ,
Jp (A, B)− 2Rp (A, B) ≤ Tp (AB) ≤ Jp (A, B)− 2Lp (A, B) ,
(8)
(9)
and
where
2
Kp (A, B) ≡ A1/2(cid:16) A−1/2BA−1/2+I
Jp (A, B) ≡ 1
Lp (A, B) ≡ 1
Rp (A, B) ≡ 1
2 (A#pB− A♮p−1B + B− A) ,
24 (p− 1) (p− 2) (A#pB− 3A♮p−1B + 3A♮p−2B− A♮p−3B) ,
24 (p− 1) (p− 2) (A♮3B− 3A♮2B + 3B− A) .
(cid:0)A−1/2BA−1/2 − I(cid:1)A1/2,
(cid:17)p−1
Proof. According to [18, Theorem 1], if f : [a, b] → R is a twice differentiable
function that there exists real constants m and M so that m ≤ f ′′ ≤ M , then
m
(b− a)2
24 ≤
1
b− a
b
Za
f (t) dt − f(cid:18) a + b
2 (cid:19) ≤ M
(b− a)2
24
,
m
(b− a)2
12 ≤
f (a) + f (b)
2
−
1
b− a
b
Za
f (t) dt ≤ M
(b− a)2
12
.
(10)
(11)
4
Putting f (t) = t p−1 with p ∈ (0, 1] and a = 1 , b = x in the above inequalities, then we
have the desired results by a similar way to the proof of Theorem 1.
REMARK 4. The first inequality of (8) and the second inequality of (9) give tighter
bounds of Tsallis relative entropy Tp(AB) than those in the inequalities (3), because of
the following reasons.
(i) The first inequality of (10) gives tight lower bound more than the first inequality
of Hermite-Hadamard's inequality (2).
(ii) The first inequality of (11) gives tight upper bound more than the second inequal-
ity of Hermite-Hadamard's inequality (2).
3. Some reverse inequalities via Young type inequalities
The scalar Young's inequality says that if a, b > 0 , then
a1−pbp ≤ (1− p)a + pb,
p ∈ [0, 1] .
The following inequalities provide a refinement and a multiplicative reverse for Young's
inequality with Kantorovich constant:
Kr (h, 2) a1−pbp ≤ (1− p) a + pb ≤ KR (h, 2) a1−pbp,
(12)
where a, b > 0 , p ∈ [0, 1], r = min{p, 1− p}, R = max{p, 1− p}, K (h, 2) = (h+1)2
and h = b
a . Notice that the first inequality in (12) was obtained by Zou et al. in [21,
Corollary 3] while the second was obtained by Liao et al. [17, Corollary 2.2].
4h
In [14, 15], Kittaneh and Manasrah obtained another refinement and reverse of
Young's inequality:
r(cid:16)√a−
√b(cid:17)2
+ a1−pbp ≤ (1− p)a + pb ≤ R(cid:16)√a−
√b(cid:17)2
+ a1−pbp,
(13)
where a, b > 0 , p ∈ [0, 1] , r = min{p, 1− p}, R = max{p, 1− p}. Further refinements
and generalizations of Young's inequality have been obtained in [2, 7].
More interesting things happen when we apply these considerations to the opera-
tors. For instance, from the inequality (12) it follows that:
PROPOSITION 2. Let A, B be two invertible positive operators such that I < h′I ≤
2 BA− 1
2 BA− 1
A− 1
2 ≤ hI or 0 < hI ≤ A− 1
2 ≤ h′I < I , then
Kr(cid:0)h′, 2(cid:1) A#pB ≤ A∇pB ≤ KR (h, 2) A#pB,
where p ∈ [0, 1], r = min{p, 1− p}, R = max{p, 1− p}.
(14)
5
Proof. The choice a = 1 and b = x reduces (12) to the inequality
Kr (x, 2) xp ≤ (1− p) + px ≤ KR (x, 2) xp.
Whence
min
1<h′≤x≤h
Kr (x, 2) T p ≤ (1− p)I + pT ≤ max
1<h′≤x≤h
KR (x, 2) T p,
(15)
min
2 BA− 1
in (15) we get
Kr (x, 2)(cid:16)A− 1
for any positive operator T such that I < h′I ≤ T ≤ hI . On choosing T = A− 1
KR (x, 2)(cid:16)A− 1
2(cid:17)p
2(cid:17)p
≤ (1− p) I + pA− 1
1<h′≤x≤h
Since K (x, 2) is increasing for x > 1 we can write
2 ≤ KR (h, 2)(cid:16)A− 1
≤ (1− p)I + pA− 1
2 ≤ max
1<h′≤x≤h
Kr(cid:0)h′, 2(cid:1)(cid:16)A− 1
2(cid:17)p
2 BA− 1
2 BA− 1
2 BA− 1
Multiplying both sides by A
1
2 to inequality (16), we obtain the required inequality (14).
2 BA− 1
.
(16)
2 BA− 1
2
2 BA− 1
.
2(cid:17)p
Another case follows from the fact that K (x, 2) is decreasing for x < 1 .
Ando's inequality [3, Theorem 3] says that if A, B are positive operators and Φ is a
positive linear mapping, then
Concerning inequality (17), we have the following corollary:
Φ (A#pB) ≤ Φ (A) #pΦ (B) ,
p ∈ [0, 1].
(17)
COROLLARY 2. Let A, B be two invertible positive operators such that I < h′I ≤
2 BA− 1
2 ≤ h′I < I . Let Φ be positive linear map on
2 ≤ hI or 0 < hI ≤ A− 1
2 BA− 1
A− 1
B (H ) , then
Kr (h′, 2)
KR (h, 2)
Φ (A#pB) ≤
1
KR (h, 2)
Φ (A∇pB)
≤ Φ (A) #pΦ (B)
≤
1
Kr (h′, 2)
KR (h, 2)
Kr (h′, 2)
≤
Φ (A∇pB)
Φ (A#pB) ,
where p ∈ [0, 1], r = min{p, 1− p} and R = max{p, 1− p}.
Proof. If we apply positive linear map Φ in (14) we infer
On the other hand, if we take A = Φ (A) and B = Φ (B) in (14) we can write
Kr(cid:0)h′, 2(cid:1) Φ (A#pB) ≤ Φ (A∇pB) ≤ KR (h, 2) Φ (A#pB) .
Kr(cid:0)h′, 2(cid:1) Φ (A) #pΦ (B) ≤ Φ (A∇pB) ≤ KR (h, 2) Φ (A) #pΦ (B) .
Now, combining inequality (19) and inequality (20), we deduce the desired inequalities
(18).
(18)
(19)
(20)
6
REMARK 5. It is well-known that the generalized Kantorovich constant K (h, p)
[12, Definition 1] is defined by
K (h, p) :=
hp − h
(p− 1)(h− 1)(cid:18) p− 1
p
hp − h(cid:19)p
hp − 1
,
(21)
for all p ∈ R . By virtue of a generalized Kantorovich constant, in the matrix setting,
Bourin et al. in [4, Theorem 6] gave the following reverse of Ando's inequality for a
positive linear map:
Let A and B be positive operators such that mA ≤ B ≤ MA , and let Φ be a positive
linear map. Then
Φ (A) #pΦ (B) ≤
1
K (h, p)
Φ (A#pB) ,
p ∈ [0, 1]
(22)
where h = M
for p = 1
2 .
m . The above result naturally extends one proved in Lee [16, Theorem 4]
Of course the constant KR(h,2)
Kr(h′,2) is not better than
1
K( h
h′ ,p)
. Concerning the sharp-
ness of the estimate (22), see [4, Lemma 7]. However our bounds on Φ(A)#pΦ(B)
are calculated by the original Kantrovich constant K(h, 2) without the generalized one
K(h, p) . It is also interesting our bounds on Φ(A)#pΦ(B) are expressed by Φ(A∇pB)
with only one constant either h or h′ .
After discussion on inequalities related to the operator mean with positive linear map,
we give a result on Tsallis relative operator entropy (which is the main theme in this
paper) with a positive linear map. It is well-known that Tsallis relative operator entropy
has the following information monotonicity:
Φ (Tp (AB)) ≤ Tp (Φ (A)Φ (B)) .
(23)
Utilizing (13), we have the following counterpart of (23):
THEOREM 3. Let A, B be two invertible positive operators. Let Φ be normalized
positive linear map on B (H ) , then
(Φ (A∇B)− Φ (A) #Φ (B)) + Tp (Φ (A)Φ (B))
(24)
2r
p
≤ Φ (B− A)
≤
2R
p
(Φ (A∇B)− Φ (A#B)) + Φ (Tp (AB)) ,
where p ∈ (0, 1] , r = min{p, 1− p} and R = max{p, 1− p}.
7
Proof. Using the method of the proof of Proposition 2 and Corollary 2 for the
inequality (13), we can obtain that
2r (Φ (A∇B)− Φ (A) #Φ (B)) + Φ (A) #pΦ (B)
≤ Φ (A∇pB)
≤ 2R (Φ (A∇B)− Φ (A#B)) + Φ (A#pB) .
A simple calculation shows that
2r
p
≤ Φ (B− A)
≤
2R
p
(Φ (A∇B)− Φ (A) #Φ (B)) +
Φ (A) #pΦ (B)− Φ (A)
p
(Φ (A∇B)− Φ (A#B)) +
Φ (A#pB)− Φ (A)
p
.
Apparently, the above inequality is equivalent to inequality (24). The proof is com-
pleted.
The following example may render the above statement clearer.
EXAMPLE 1. Let Φ (X) = U∗XU (X ∈ M2) where U = √2
(cid:18) 2 −1
−1 1 (cid:19) and B =(cid:18)6 2
2 4(cid:19) and p = 1
4 we get
−
2
√2
2
√2
2
√2
2 ! .
If A =
2r
p
(Φ (A∇B)− Φ (A) #Φ (A)) + (Tp (Φ (A)Φ (A))) =(cid:18)0.486 0.443
0.443 5.638(cid:19)
Φ (B− A) =(cid:18)0.5 0.5
0.5 6.5(cid:19)
and
2R
p
(Φ (A∇B)− Φ (A#B)) + Φ (Tp (AB)) =(cid:18)0.562 0.951
0.951 13.521(cid:19)
which shows that inequality (24) is true.
Tsallis relative entropy Dp (AB) for two positive operators A and B is defined
by:
Dp (AB) :=
Tr [A]− Tr(cid:2)A1−pBp(cid:3)
p
,
p ∈ (0, 1].
In information theory, relative entropy (divergence) is usually defined for density oper-
ators which are positive operators with unit trace. However we consider Tsallis relative
entropy defined for positive operators to derive the relation with Tsallis relative operator
entropy. If A and B are positive operators, then
Tr [A− B] ≤ Dp (AB) ≤ −Tr [Tp (AB)] ,
p ∈ (0, 1].
(25)
8
Note that the first inequality of (25) is due to Furuta [13, Proposition F] and the second
inequality is due to Furuichi et al. [10, Theorem 2.2].
As a direct consequence of Theorem 3, we have the following interesting relation.
COROLLARY 3. Let A, B be two positive operators on a finite dimensional Hilbert
space H , then
Tr [A + B]
2R
2
p (cid:18)Tr [A#B]−
≤ Tr [A− B]
p (cid:18)pTr [A] Tr [B]−
≤
2r
(cid:19)− Tr [Tp (AB)]
Tr [A + B]
2
(cid:19) + Dp (AB) ,
where p ∈ (0, 1] , r = min{p, 1− p} and R = max{p, 1− p}.
Proof. Taking Φ (X) = 1
dimH Tr [X] in (18), we have
(Tr [A])1−p(Tr [B])p − Tr [A]
p
2r
2
p (cid:18) Tr [A + B]
≤ Tr [B− A]
p (cid:18) Tr [A + B]
≤
−pTr [A] Tr [B](cid:19) +
− Tr [A#B](cid:19) + Tr [Tp (AB)] .
2R
2
It is not too difficult to see that (27) can be also reformulated in the following way
(26)
(27)
(cid:19)− Tr [Tp (AB)]
Tr [A + B]
2R
2
p (cid:18)Tr [A#B]−
≤ Tr [A− B]
p (cid:18)pTr [A] Tr [B]−
≤
2r
Tr [A + B]
2
(cid:19) +
Tr [A]− (Tr [A])1−p(Tr [B]) p
p
.
Now, having in mind that Tr(cid:2)A1−pBp(cid:3) ≤ (Tr [A])1−p(Tr [B])p , we infer
Tr [A + B]
2R
(cid:19)− Tr [Tp (AB)]
2
p (cid:18)Tr [A#B]−
≤ Tr [A− B]
p (cid:18)pTr [A] Tr [B]−
≤
2r
which, in turn, leads to (26).
Tr [A + B]
2
(cid:19) +
Tr [A]− Tr(cid:2)A1−pBp(cid:3)
p
,
REMARK 6. If A and B are density operators, then the second inequality of (26)
2 , 1] , then
implies the non-negativity of Tsallis relative entropy, Dp(AB) ≥ 0 . If p ∈ [ 1
the first inequality of (26) implies
Tr[T1/2(AB)] ≤ Tr[Tp(AB)].
9
If p ∈ (0, 1
2 ] , then the first inequality of (26) also implies
(1− p)Tr[T1/2(AB)] + (2p− 1)Tr[B− A]≤ pTr[Tp(AB)].
R E F E R E N C E S
[1] S. Abe, Monotonic decrease of the quantum non-additive divergence by projective measurements,
Phys. Lett. A., 312(5-6) (2003), 336–338.
[2] Y. Al-Manasrah, F. Kittaneh, A generalization of two refined Young inequalities, Positivity., 19(4)
(2015), 757–768.
[3] T. Ando, Concavity of certain maps on positive definite matrices and applications to Hadamard prod-
ucts, Linear Algebra Appl., 26 (1979), 203–241.
[4] J.C. Bourin, E.-Y. Lee, M. Fujii, Y. Seo, A matrix reverse Holder inequality, Linear Algebra Appl.,
431(11) (2009), 2154–2159.
[5] J.I. Fujii, E. Kamei, Relative operator entropy in non-commutative information theory, Math. Japon.,
34 (1989), 341–348.
[6] S. Furuichi, Inequalities for Tsallis relative entropy and generalized skew information, Linear Multi-
linear Algebra., 59(10) (2011), 1143–1158.
[7] S. Furuichi, N. Minculete, Alternative reverse inequalities for Young's inequality, J. Math Inequal.,
5(4) (2011), 595–600.
[8] S. Furuichi, Precise estimates of bounds on relative operator entropies, Math. Inequal. Appl., 18(3)
(2015), 869–877.
[9] S. Furuichi, K. Yanagi and K. Kuriyama, A note on operator inequalities of Tsallis relative operator
entropy, Linear Algebra Appl., 407 (2005), 19–31.
[10] S. Furuichi, K. Yanagi, K. Kuriyama, Fundamental properties of Tsallis relative entropy, J. Math.
Phys., 45 (2004), 4868–4877.
[11] S. Furuichi, N. Minculete, F.C. Mitroi, Some inequalities on generalized entropies, J. Inequal. Appl.,
1 (2012), 226.
[12] T.Furuta, Basic properties of the generalized Kantorovich constant K (p) = hp−h
Acta Sci. Math (Szeged)., 70 (2004), 319–337.
(p−1)(h−1)(cid:16) p−1
p
hp−1
hp−h(cid:17)p
,
[13] T. Furuta, Two reverse inequalities associated with Tsallis relative operator entropy via generalized
Kantorovich constant and their applications, Linear Algebra Appl., 412(2) (2006), 526–537.
[14] F. Kittaneh and Y. Manasrah, Improved Young and Heinz inequalities for matrix, J. Math. Anal. Appl.,
361 (2010), 262–269.
[15] F. Kittaneh and Y. Manasrah, Reverse Young and Heinz inequalities for matrices, Linear Multilinear
Algebra., 59 (2011), 1031–1037.
[16] E.-Y. Lee, A matrix reverse Cauchy-schwarz inequality, Linear Algebra Appl., 430(2) (2009), 805–
810.
[17] W. Liao, J. Wu, J. Zhao, New versions of reverse Young and Heinz mean inequalities with the Kan-
torovich constant, Taiwanese J. Math., 19(2) (2015), 467–479.
[18] C.P. Niculescu, L.E. Persson. Old and new on the Hermite-Hadamard inequality, Real Anal. Ex-
change., 29(2) (2004), 663–686.
[19] K. Yanagi, K. Kuriyama, S. Furuichi, Generalized Shannon inequalities based on Tsallis relative
operator entropy, Linear Algebra Appl., 394 (2005), 109–118.
[20] L. Zou, Operator inequalities associated with Tsallis relative operator entropy, Math. Inequal. Appl.,
18(2) (2015), 401–406.
[21] H. Zuo, G. Shi, M. Fujii, Refined Young inequality with Kantorovich constant, J. Math. Inequal., 5(4)
(2011), 551–556.
10
Young Researchers and Elite Club, Mashhad Branch, Islamic Azad University, Mashhad, Iran.
e-mail: [email protected]
Department of Information Science, College of Humanities and Sciences, Nihon University, 3-25-40, Saku-
rajyousui, Setagaya-ku, Tokyo, 156-8550, Japan.
e-mail: [email protected]
Faculty of Mathematics and Computer Science, Transilvania University of Bras¸ov, Str. Iuliu Maniu, nr. 50,
Bras¸ov, 500091, Romania.
e-mail: [email protected]
Corresponding Author: Shigeru Furuichi
11
|
1103.0409 | 2 | 1103 | 2015-11-19T10:09:09 | The Pole Behaviour of the Phase Derivative of the Short-Time Fourier Transform | [
"math.FA"
] | The short-time Fourier transform (STFT) is a time-frequency representation widely used in applications, for example in audio signal processing. Recently it has been shown that not only the amplitude, but also the phase of this representation can be successfully exploited for improved analysis and processing. In this paper we describe a rather peculiar pole phenomenon in the phase derivative, a recurring pattern that appears in a characteristic way in the neighborhood around any of the zeros of the STFT, a negative peak followed by a positive one. We describe this phenomenon numerically and provide a complete analytical explanation. | math.FA | math |
THE POLE BEHAVIOUR OF THE PHASE
DERIVATIVE OF THE SHORT-TIME FOURIER
TRANSFORM
PETER BALAZSA), DOMINIK BAYERB), FLORENT JAILLETC) AND
PETER SØNDERGAARDD)
Abstract. The short-time Fourier transform (STFT) is a time-
frequency representation widely used in applications, for example
in audio signal processing. Recently it has been shown that not
only the amplitude, but also the phase of this representation can be
successfully exploited for improved analysis and processing. In this
paper we describe a rather peculiar pole phenomenon in the phase
derivative, a recurring pattern that appears in a characteristic way
in the neighborhood around any of the zeros of the STFT, a nega-
tive peak followed by a positive one. We describe this phenomenon
numerically and provide a complete analytical explanation.
1. Introduction
The short-time Fourier transform (STFT) [5, 13] is a time-frequency
representation widely used in audio signal processing. A common def-
inition of the STFT1 is
(cid:90)
(1)
V (f, g)(x, ω) =
f (t)g(t − x)e−2πiωt dt.
The STFT V (f, g)(x, ω) provides information about the frequency con-
tent of the signal f at time x and frequency ω. The analyzing window
g determines the resolution in time and frequency.
The interpretation of the modulus of the STFT is relatively easy, con-
sidering the fact that the spectrogram (defined as the square absolute
value of the STFT) can be interpreted as a time-frequency distribution
of the signal energy. This interpretation led to the important success of
A) Acoustics Research Institute, Austrian Academy of Sciences, Vienna A-1040,
Austria;
B) Corresponding author; Acoustics Research Institute, Austrian Academy of Sci-
ences, Vienna A-1040, Austria; (cid:84) +43 1 51581-2517 (cid:66) [email protected];
C) Institut de Neurosciences de la Timone UMR 7289, Aix Marseille Universit´e,
CNRS, 13385 cedex 5, Marseille, France;
D) Oticon A/S, 2765 Smørum, Denmark;
1This is the frequency-invariant STFT.
1
the STFT in signal processing. In particular, it has been widely used
for applications in speech processing and acoustics as a graphical tool
for signal analysis [21].
But the interpretation of the phase of the STFT is less obvious, and
was thus hardly considered in applications for some time.
The phase can be of particular interest for certain applications, as
illustrated by important applications such as phase vocoder [12, 7] or
reassignment [18, 2]. In digital image processing it is well known that
the phase information of the discrete Fourier transform is at least as
important as the amplitude information.
In [19] it is shown that as
long as the phase of the discrete Fourier transform of an image is re-
tained and the amplitude is set to 1, the image can still be recognized.
Similar effects can also be shown for acoustic signal depending on the
parameters of the STFT [4].
For applications modifying the STFT coefficients, phase information
is essential again. For these types of applications, in particular for the
applications using Gabor frame multipliers [10, 3] which motivated the
present study, better understanding of the structure of the phase is
necessary to improve the processing possibilities.
The phase of the STFT is usually not considered directly. In fact, it is
more interesting to consider the phase derivative over time or frequency.
Indeed, these quantities appear naturally in the context of reassignment
[2] and manipulations of phase derivative over time is the idea behind
the phase vocoder [7]. Their interpretation is easier, as the derivative
of phase over time can be interpreted as local instantaneous frequency
while the derivative of the phase over frequency can be interpreted as
a local group delay.
To numerically compute the local instantaneous frequency, an un-
wrapping of the phase is needed to avoid discontinuities. This is the
classical method used in [7, 18]. Another method was found in [2]:
(cid:32)
(cid:33)
,
(2)
∂
∂x
arg(V (f, g)(x, ω)) = Im
V (f, g(cid:48))(x, ω)V (f, g)(x, ω)
V (f, g)(x, ω)2
with g(cid:48)(t) = dg
dt (t). The benefit of this method is that is does not require
unwrapping, instead the phase derivative is computed by pointwise op-
erations using a second STFT based on the derivative of the window.
To understand the phase of the STFT more thoroughly, in particular
for applications dealing with multipliers, see for example [20, 22, 23],
we conducted related extensive numerical experiments. In the process
2
we observed a rather peculiar phenomenon in the phase derivative, a
recurring pattern that appears in a similar way in the neighborhood
around any of the zeros of the STFT. The behaviour of the phase
derivative close to the singularity always shows the same characteristic
shape, i.e., a negative peak followed by a positive one. We describe
this phenomenon and provide a complete analytical explanation.
This paper is organized as follows: In Section 2 we report the nu-
In Section 3 we give a short, instructive, analytical
In Section 4 we give the full analytical
merical results.
example for this behaviour.
results.
Results in this paper have partly been reported at a conference [17],
and a preprint of this paper has already been cited in [1].
2. Numerical Observations
For noise, naturally only statistical properties of the phase are ac-
cessible. Some interesting results for the phase derivative have been
shown in the context of reassignment.
In [8], the following result is
given: We consider a zero-mean Gaussian analytic white noise f such
that
E[Re(f (t)) · Re(f (s))] = E[Im(f (t)) · Im(f (s))] =
(3)
and E[f (t)f (s)] = 0 for any (t, s) ∈ R2, with its real and imaginary
parts a Hilbert transform pair. Using a Gaussian window given by
δ(t − s)
σ2
2
−π t2
g(t) = e
variable with distribution of the form:
2σ2 , the phase derivative over time of V (f, g) is a random
(4)
ρ(v) =
1
2(1 + v2) 3
2
.
This distribution is shown in Figure 1. As can be seen, it is a quite
"peaky" distribution, indicating that the values of the phase derivative
are mainly values close to zero, with some rare values with higher
absolute values.
The spatial distribution of the phase derivative seems to be difficult
to be solved analytically. Therefore we conducted systematic numerical
experiments to study this spatial distribution.
For this, we need to compute the derivative of the phase in discrete
settings. We used the expression (2) to compute the phase derivative.
3
Figure 1. Distribution of the values of the phase deriv-
ative over time of the STFT for a white Gaussian noise.
We see on this formula that we will face numerical difficulties when
the denominator V (f, g)(x, ω) is close to zero. But using double preci-
sion, these problems only appear for really small values of the modulus
(on the order of 10−13), which allows us to reliably observe the values of
the phase derivative even close to the zeros of the STFT. In the figures
of this paper, the phase derivative values are ignored and represented
as white at the points where the value of the modulus is too small.
The results of our experiments are illustrated by Figure 2. The time-
frequency distribution of the values appears to be highly structure, as
e.g. noted in [15]. The values of the phase derivative with high absolute
values are concentrated around several time-frequency points, which
can be identified as the zeros of the transform when looking at the
modulus. Furthermore, the shape of the phase derivative seems to be
very similar in the neighbourhood of the zeros, with a typical pattern
repeating at each zero, see Figure 2. When going from low to high
frequencies, it presents a negative peak followed by a positive one.
This phenomenon is related to the fact that the STFT of white noise
is a correlated process, with a correlation determined by the window
through the reproducing kernel of the transform (see part 6.2.1 of [5]).
It is thus interesting to study the influence of the window choice on the
observed structure of the phase derivative, as illustrated in Figure 3.
We can observe that narrowing the window results in similar patterns
around the zeros, but with a scaled shape: the resulting pattern is
narrower over time, but wider over frequency. Figure 3 also shows the
influence of the window type. The structure is more complicated for
windows with bad time-frequency concentration. On the representation
using a Hamming window, we still observe repeating patterns at the
zeros of the transform, but the variability of the shape of this pattern
seems higher, and the pattern orientation slightly varies, whereas it is
fixed in the case of a Gaussian window. For the case of the rectangular
4
−20−100102000.10.20.30.4 Figure 2. Observation for a Gaussian white noise, us-
ing a Gaussian window. Top: modulus of the STFT.
Bottom-left: derivative over time of the phase of the
STFT using the definition (1). Bottom-right: mesh plot
of the derivative over time of the phase in the neighbour-
hood of a zero of the STFT.
window, the zeros of the STFT form a more complicated, extended
structures. This leads to much more variable patterns. Yet, we still,
interestingly, observe that the values of the phase derivative with high
absolute values concentrate around the zeros of the transform, whereas
the phase derivative is close to zero in the regions of the STFT where
the modulus is high.
¿From the experiments above, we expect those properties to be
highly correlated with co-orbit properties regarding the STFT, i.e. in-
clusion in certain modulation spaces [9]. In particular, choosing win-
dows in the Feichtinger algebra S0 [11] should result in a phase deriv-
ative behavior comparable to the Gaussian window. We also expect
results to be valid as in Section 4, for windows in S0. The systematic
investigation of the phase derivative behaviour for modulation spaces
is beyond the scope of this paper, and will be investigate in future work.
The behaviour that we observe is not specific to noise signals. Indeed,
further experiments on other synthesized and recorded complex sounds
5
TimeAngular Frequency 2002202402602801.41.61.822.22.400.511.52TimeAngular Frequency 2002202402602801.41.61.822.22.4−2−10122202301.81.851.91.95−2−1012TimeAngular Frequency Figure 3. Influence of the window when analyzing a
frozen Gaussian white noise. For three different windows,
on the left, modulus of the STFT, on the right, derivative
over time of the phase of the STFT using the definition
(1). From top to bottom, the windows are: a narrower
Gaussian window, a Hamming window, a rectangular
window.
showed that the same characteristics can be observed for all signals: the
values of the phase derivative of high absolute value are concentrated in
the neighbourhood of the zeros of the STFT, and for "nice" windows,
a specific pattern appears in this neighbourhood.
6
TimeAngular Frequency 2002202402602801.41.61.822.22.400.511.522.5TimeAngular Frequency 2002202402602801.41.61.822.22.4−2−1012TimeAngular Frequency 2002202402602801.41.61.822.22.402468TimeAngular Frequency 2002202402602801.41.61.822.22.4−2−1012TimeAngular Frequency 2002202402602801.41.61.822.22.4024681012TimeAngular Frequency 2002202402602801.41.61.822.22.4−2−1012 3. A Simple Explicit Analytic Example
In this section we give a simple analytical example for which we can
explicitly compute the phase derivative.
Considering the signal given by
(5)
f (t) = e2πiω1t + e2πiω2t
and using a Gaussian window g(t) = e
the expression of the STFT, which results in the formula:
2σ2 , we can explicitly compute
−π t2
(6)
V (f, g)(x, ω) = e−2πix(ω−ω1)e−2πσ2(ω−ω1)2 + e−2πix(ω−ω2)e−2πσ2(ω−ω2)2.
The zeros of this STFT are the points of coordinates (xk, ωmid) in the
2(ω1−ω2) for k ∈ Z.
time-frequency plane, with ωmid = ω1+ω2
and xk = 1+2k
2
The expression of the phase derivative for this signal, given in part
VI-12 of [6], is:
(7)
∂
∂x
with δ = ω2−ω1
2
(cid:16)
ωmid − ω + δ tanh(s)
(cid:17)
1 + tan2(2πδx)
1 + tan2(2πδx) tanh2(s)
arg(V (f, g)(x, ω)) = 2π
and s = 4πσ2(ω − ωmid)δ.
The plot of this function around one of the zeros is visible in Figure 4.
We see the pattern that was already observed in the previous section.
For an extended treatment of this analytic example, see [1], where
the authors are already referring to a preprint of our present work.
4. Analytical Results
In this section, we denote by Mh the modulation operator Mh :
L2(R) → L2(R), f (t) (cid:55)→ Mhf (t) := e−2πihtf (t), and by Th the trans-
lation operator Th : L2(R) → L2(R), f (t) (cid:55)→ Thf (t) := f (t − h) (with
h ∈ R). The set S(R) is the Schwartz class of rapidly decaying func-
tions. The Fourier transform is denoted by F.
4.1. Regularity Properties of the STFT.
Definition 4.1. Define the (unbounded) operator P on L2(R) as the
multiplication operator
P f (t) := 2πit · f (t)
7
Figure 4. Observation for the signal defined in (5).
Top: modulus of the STFT. Bottom: derivative over
time of the phase of the STFT according to (7) repre-
sented as an image (left) and as a mesh (right).
with domain
Dom(P ) := {f ∈ L2(R) :
(cid:90)
R
t f (t)2 dt < ∞} ⊂ L2(R).
Further, define the (unbounded) operator
Q := F−1PF
(where F denotes the Fourier transform) with domain
Dom(Q) := {f ∈ L2(R) : Ff ∈ Dom(P )} ⊂ L2(R).
In quantum mechanics, these operators are essentially the momen-
tum and position operator, respectively. The operator P (and thus also
Q) are clearly densely-defined, since S(R) ⊂ Dom(P ) (and F−1S(R) =
S(R) ⊂ Dom(Q)). It can be shown that P and Q are closed unbounded
operators and that iP and iQ are self-adjoint.
We collect basic properties of these operators in the following lemma.
Lemma 4.2. The operators P and Q have the following properties:
(i) FQ = PF on Dom(Q), QF = −FP on Dom(P );
8
TimeAngular Frequency 99.5101.41.451.51.551.60.511.522.53TimeAngular Frequency 99.5101.41.451.51.551.6−20−100102099.5101.41.51.6−20−1001020TimeAngular Frequency (ii) Q is a (maximal extension of a) differential operator, more
precisely: if f ∈ S(R), then Qf (t) = d
dt f (t).
The next lemma is in essence a well-known result from the theory
of operator (semi-)groups; it gives the infinitesimal generators of the
modulation and translation group, respectively, see e.g. [14, 16]. The
version below needed in this manuscript can be proved in a straight-
forward way.
Lemma 4.3. Let f ∈ Dom(P ), then
(cid:107) 1
h
for h → 0.
Let f ∈ Dom(Q), then
(cid:107) 1
h
for h → 0.
(Mh − Id)f − P f(cid:107)L2 → 0
(Th − Id)f + Qf(cid:107)L2 → 0
We can now prove a regularity result for the short-time Fourier trans-
form.
Proposition 4.4. Let f, g ∈ L2(R).
(i) If f belongs to Dom(P ), then V (f, g) has a continuous partial
derivative with respect to the second argument ω, and we have
V (f, g)(x, ω) = −V (P f, g)(x, ω).
∂
∂ω
(ii) If g belongs to Dom(Q), then V (f, g) has a continuous partial
derivative with respect to the first argument x, and we have
V (f, g)(x, ω) = −V (f, Qg)(x, ω).
∂
∂x
Proof. Assume f ∈ Dom(P ). Then, by the preceding lemma,
1
h
(V (f, g)(x, ω + h) − V (f, g)(x, ω)) = (cid:104)f,
f, MωTxg(cid:105)
= (cid:104)M−h − Id
h→0−→ (cid:104)−P f, MωTxg(cid:105) = −V (P f, g)(x, ω),
h
Mh − Id
h
MωTxg(cid:105)
which is a continuous function on R2.
If g ∈ Dom(Q), then, analogously,
1
h
(V (f, g)(x + h, ω) − V (f, g)(x, ω)) = (cid:104)f, MωTx
g(cid:105)
h→0−→ (cid:104)f,−MωTxQg(cid:105) = −V (f, Qg)(x, ω).
h
Th − Id
9
(cid:3)
Using V (f, g)(x, ω) = e−2πixωV (g, f )(−x,−ω), the partial deriva-
tives of V (f, g) exist if and only if those of V (g, f ) exist. We may thus
change the roles of f and g.
Corollary 4.5. Let f, g ∈ L2(R).
(i) If g belongs to Dom(P ), then V (f, g) has a continuous partial
derivative with respect to the second argument ω, and we have
∂
∂ω
V (f, g)(x, ω) = V (f, P g)(x, ω) − 2πix V (f, g)(x, ω).
(ii) If f belongs to Dom(Q), then V (f, g) has a continuous partial
derivative with respect to the first argument x, and we have
∂
∂x
V (f, g)(x, ω) = V (Qf, g)(x, ω) − 2πiω V (f, g)(x, ω).
If f (resp. g) belongs to Schwartz class S(R), then P f, Qf (resp.
Iterated application of Proposition 4.4 or
P g, Qg) ∈ S(R), as well.
Corollary 4.5 gives the following smoothness result for the STFT.
Theorem 4.6. Let f, g ∈ L2(R), and at least one of them in Schwartz
class S(R). Then V (f, g)(x, ω) is infinitely partially differentiable in
both variables x and ω.
Although this result may be considered mathematical folklore, to
our knowledge it has not been stated and proved in the literature so
far. Note that this proves in particular that the STFT with Gaussian
window is smooth.
4.2. The Derivative of the Phase Around the Zeros of the
STFT. In this section we present an analytic explanation of the pe-
culiar behaviour of the phase derivatives of the STFT for a large class
of window functions. It turns out that the phenomenon is connected
to the smoothness and continuous differentiability of the STFT which,
as we have seen in the previous paragraph, is in turn connected to the
smoothness of the window.
Consider first the partial derivative of the phase of the STFT with
respect to the first variable (i.e., the 'time' variable). For convergence
along a vertical path, we have
Theorem 4.7 (Phase derivatives of the STFT, part I). Let f, g ∈
L2(R). Assume that
• V (f, g) = V = U + i · W ∈ C 2(R2)
10
(x0, ω0)
(x0, ω0)
∂W
∂x
(x0, ω0)
(x0, ω0)
∂U
∂ω
∂W
∂ω
• V (x0, ω0) = 0
• det JV (x0, ω0) < 0, where
∂U
∂x
JV (x0, ω0) =
(cid:40)
denotes the Jacobian matrix of V at the point (x0, ω0)
Then the phase ψ(x, ω) of V (f, g)(x, ω) satisfies
(x0, ω) −→
∂ψ
∂x
+∞,
−∞,
if ω ↑ ω0 from below
if ω ↓ ω0 from above.
Proof. We have
∂ψ
∂x
(x0, ω) =
U (x0, ω) · Wx(x0, ω) − W (x0, ω) · Ux(x0, ω)
U 2(x0, ω) + W 2(x0, ω)
.
Since Wx and Ux are continuous and thus remain bounded in a neigh-
bourhood of (x0, ω0), both numerator and denominator tend to zero
for ω → ω0. However, both functions are differentiable, since V ∈
C 2(R2, R2). So L'Hospital's Rule is applicable and yields
lim
ω→ω0
U (x0, ω) · Wx(x0, ω) − W (x0, ω) · Ux(x0, ω)
U 2(x0, ω) + W 2(x0, ω)
dω (U (x0, ω) · Wx(x0, ω) − W (x0, ω) · Ux(x0, ω))
d
L'Hosp.= lim
ω→ω0
d
dω (U 2(x0, ω) + W 2(x0, ω))
(UωWx + U Wxω − WωUx − W Uxω) (x0, ω)
= lim
ω→ω0
= lim
ω→ω0
(2U Uω + 2W Wω) (x0, ω)
(U Wxω − W Uxω) (x0, ω) + (UωWx − WωUx) (x0, ω)
(2U Uω + 2W Wω) (x0, ω)
.
Concerning this limit, we clearly have
(U Wxω − W Uxω) (x0, ω) → 0,
since U (x0, ω) → U (x0, ω0) = 0, W (x0, ω) → W (x0, ω0) = 0, and Wxω
and Uxω are continuous and thus remain bounded in a neighborhood
of (x0, ω0). Furthermore,
(UωWx − WωUx) (x0, ω) → (UωWx − WωUx) (x0, ω0)
= − det JV (x0, ω0) (cid:54)= 0,
by assumption. Hence the numerator tends to a nonzero number, in
this case (det JV (x0, ω0) < 0) a positive one. For the denominator, we
11
find
(2U Uω + 2W Wω) (x0, ω) =
=
d
dω
(cid:40)
(cid:0)U 2 + W 2(cid:1) (x0, ω)
< 0,
> 0,
if ω < ω0
if ω > ω0
since the function ω (cid:55)→ (2U Uω + 2W Wω) (x0, ω) has a strict local min-
imum in ω0. At the same time,
(2U Uω + 2W Wω) (x0, ω) → 0
for ω → ω0, hence the denominator goes to zero from below for ω ↑ ω0
and from above for ω ↓ ω0. This concludes the proof.
(cid:3)
Note that for simplicity we have only considered the case that the
Jacobian determinant det JV (x0, ω0) is negative; this case corresponds
to the examples we presented above. For positive Jacobian determi-
nant, the situation is completely analogous, although reversed in the
sense that the positive and negative singularities switch roles. Apart
from this, the general behaviour remains the same.
For convergence along a horizontal path, we need slightly more reg-
ularity:
Theorem 4.8 (Phase derivatives of the STFT, part II). Let f, g ∈
L2(R). Assume that
• V (f, g) = V = U + i · W ∈ C 3(R2)
• V (x0, ω0) = 0
• det JV (x0, ω0) < 0, where J is the Jacobian as in Theorem 4.7
Then there exists a number c ∈ R such that the phase ψ(x, ω) of
V (f, g)(x, ω) satisfies
lim
x→x0
∂ψ
∂x
(x, ω0) = c.
12
Proof. The assumptions allow us to apply L'Hospital's Rule twice, giv-
ing
lim
x→x0
∂ψ
∂x
(x, ω0)
= lim
x→x0
U (x, ω0) · Wx(x, ω0) − W (x, ω0) · Ux(x, ω0)
U 2(x, ω0) + W 2(x, ω0)
d
dx (U (x, ω0) · Wx(x, ω0) − W (x, ω0) · Ux(x, ω0))
L'Hosp.= lim
x→x0
d
dx (U 2(x, ω0) + W 2(x, ω0))
(UxWx + U Wxx − WxUx − W Uxx) (x, ω0)
= lim
x→x0
= lim
x→x0
(2U Ux + 2W Wx) (x, ω0)
(U Wxx − W Uxx) (x, ω0)
(2U Ux + 2W Wx) (x, ω0)
d
dx (U Wxx − W Uxx) (x, ω0)
d
dx (2U Ux + 2W Wx) (x, ω0)
L'Hosp.= lim
x→x0
(UxWxx + U Wxxx − WxUxx − W Uxxx) (x, ω0)
2 (U 2
x + U Uxx + W 2
x + W Wxx) (x, ω0)
= lim
x→x0
.
For the denominator,
(U Uxx + W Wxx) (x, ω0) → (U Uxx + W Wxx) (x0, ω0) = 0
for x → x0, but(cid:0)U 2
x + W 2
x
(cid:1) (x, ω0) →(cid:0)U 2
x + W 2
x
(cid:1) (x0, ω0) > 0
converges to a nonzero number, since not both Ux(x0, ω0) and Wx(x0, ω0
can be zero because of det JV (x0, ω0) = (UxWω − UωWx) (x0, ω0) (cid:54)= 0.
The numerator obviously converges:
(UxWxx + U Wxxx − WxUxx − W Uxxx) (x, ω0)
→ (UxWxx − WxUxx) (x0, ω0) ∈ R,
thus
lim
x→x0
∂ψ
∂x
(x, ω0) =
(UxWxx − WxUxx) (x0, ω0)
2 (U 2
x + W 2
x ) (x0, ω0)
=: c ∈ R.
(cid:3)
Concerning the partial derivatives of the phase of the STFT with
respect to the second variable (i.e., the 'frequency' variable), we can
argue almost identically and thus find the following analogous results:
Theorem 4.9 (Phase derivatives of the STFT, part III). Let f, g ∈
L2(R). Assume that
• V (x0, ω0) = 0
• det JV (x0, ω0) < 0
13
Let V (f, g) = V = U + i · W ∈ C 2(R2). Then the phase ψ(x, ω) of
V (f, g)(x, ω) satisfies
(cid:40)
lim
x→x0
∂ψ
∂ω
(x, ω0) =
+∞,
−∞,
if x → x0 from the left
if x0 ← x from the right.
Let V (f, g) = V = U + i · W ∈ C 3(R2), then
lim
ω→ω0
∂ψ
∂ω
(x0, ω) = c(cid:48) ∈ R,
if ω → ω0,
converges to some real number c(cid:48) ∈ R.
(cid:3)
Acknowledgements
The work on this paper was partly supported by the Austrian Sci-
ence Fund (FWF) START-project FLAME ('Frames and Linear Op-
erators for Acoustical Modeling and Parameter Estimation'; Y 551-
N13) and the Vienna Science and Technology Fund (WWTF) project
CHARMED ('Computational harmonic analysis of high-dimensional
biomedical data'; VRG12-009). The authors are thankful to the project
partners for fruitful discussions and valuable comments, in particular
to B. Torr´esani and M. Ehler. Particular thanks go to P. Flandrin
for his kind encouragement. The authors would also like to thank the
anonymous reviewer for valuable comments!
References
[1] F. Auger, r. Chassande-Mottin, and P. Flandrin. On phase-magnitude re-
lationships in the short-time fourier transform. IEEE Signal Process. Lett.,
19(5):267–270, 2012.
[2] F. Auger and P. Flandrin. Improving the readability of timefrequency and
time-scale representationsby the method of re-assigment. IEEE Trans. Signal
Process., 43(5):1068–1089, May 1995.
[3] P. Balazs. Basic definition and properties of Bessel multipliers. Journal of
Mathematical Analysis and Applications, 325(1):571–585, January 2007.
[4] P. Balazs, H. Waubke, and W. A. Deutsch. Phasenanalyse mit akustischen
Anwendungsbeispielen. In Proceedings DAGA 2003 - Fortschritte der Akustik,
March 2003.
[5] R. Carmona, W.-L. Hwang, and B. Torr´esani. Practical Time-Frequency Anal-
ysis. Academic Press San Diego, 1998.
[6] N. Delprat,
B. Escudie,
P. Guillemain, R. Kronland-Martinet,
P. Tchamitchian, and B. Torr´esani. Asymptotic wavelet and Gabor analysis:
extraction of instantaneous frequencies. IEEE Transactions on Information
Theory, 38(2):644–664, 1992.
[7] M. Dolson. The phase vocoder:
a tutorial. Computer Musical Journal,
10(4):11–27, 1986.
[8] F. A. E. Chassande-Mottin and P. Flandrin. On the statistics of spectrogram
reassignment vectors. Multidim. Syst. and Signal Proc., 9(4):355–362, 1998.
14
[9] H. G. Feichtinger. Modulation Spaces: Looking Back and Ahead. Sampl. The-
ory Signal Image Process., 5(2):109–140, 2006.
[10] H. G. Feichtinger and K. Nowak. A first survey of Gabor multipliers, chapter 5,
pages 99–128. Birkhauser Boston, 2003.
[11] H. G. Feichtinger and G. Zimmermann. A Banach space of test functions for
Gabor analysis, chapter 3, pages 123–170. Birkhauser Boston, 1998.
[12] J. Flanagan and R. M. Golden. Phase vocoder. Bell Syst. Tech., 45:1493 –
1509, 1966.
[13] P. Flandrin. Time-Frequency/Time-Scale Analysis. Academic Press, San
Diego, 1999.
[14] G. B. Folland and A. Sitaram. The uncertainty principle: A mathematical
survey. J. Fourier Anal. Appl., 3(3):207–238, 1997.
[15] T. Gardner and M. Magnasco. Sparse time-frequency representations. Proc.
Natl. Acad. Sci. USA, 103(16):6094–6099, April 2006.
[16] K. Grochenig. Foundations of Time-Frequency Analysis. Appl. Numer. Har-
mon. Anal. Birkhauser Boston, Boston, MA, 2001.
[17] F. Jaillet, P. Balazs, M. Dorfler, and N. Engelputzeder. On the structure of the
phase around the zeros of the short-time fourier transform. In M. M. Boone,
editor, NAG/DAGA 2009, pages 1584–1587, Rotterdam, 2009.
[18] K. Kodera, C. D. Villedary, and R. Gendrin. A new method for the numerical
analysis of nonstationary signals(magnetospheric ULF). Physics of the Earth
and Planetary Interiors, 12(2):142–150, 1976.
[19] A. V. Oppenheim and J. S. Lim. The importance of phase in signals. Proceed-
ings of the IEEE, 69(5):529–541, May 1981.
[20] N. Perraudin, P. Balazs, and P. Soendergaard. A fast Griffin-Lim algorithm.
In 2013 IEEE Workshop on Applications of Signal Processing to Audio and
Acoustics, New Paltz, NY, USA, 2013., 2013.
[21] T. Quatieri. Discrete-time speech signal processing: principles and practice.
Prentice Hall Press, Upper Saddle River, NJ, USA, 2001.
[22] D. T. Stoeva and P. Balazs. Invertibility of multipliers. Applied and Computa-
tional Harmonic Analysis, 33(2):292–299, 2012.
[23] D. T. Stoeva and P. Balazs. Canonical forms of unconditionally convergent
multipliers. Journal of Mathematical Analysis and Applications, 399:252–259,
2013.
15
|
1606.06857 | 1 | 1606 | 2016-06-22T09:00:46 | On Amalgamated Banach algebras | [
"math.FA"
] | Let $A$ and $B$ be Banach algebras, $\theta: A\to B$ be a continuous Banach algebra homomorphism and $I$ be a closed ideal in $B$. Then the direct sum of $A$ and $I$ with respect to $\theta$, denoted $A\bowtie^{\theta}I$, with a special product becomes a Banach algebra which is called the amalgamated Banach algebra. In this paper, among other things, we compute the topological centre of $A\bowtie^{\theta}I$ in terms of that of $A$ and $I$. Using this, we provide a characterization of the Arens regularity of $A\bowtie^{\theta}I$. Then we determine the character space of $A\bowtie^{\theta}I$ in terms of that of $A$ and $I$. Moreover, we study the weak amenability of $A\bowtie^{\theta}I$. | math.FA | math | ON AMALGAMATED BANACH ALGEBRAS
H. POURMAHMOOD AGHABABA AND N. SHIRMOHAMMADI
Abstract. Let A and B be Banach algebras, θ : A → B be a continuous Banach algebra
homomorphism and I be a closed ideal in B. Then the direct sum of A and I with respect
to θ, denoted A ⊲⊳ θ I, with a special product becomes a Banach algebra which is called the
amalgamated Banach algebra. In this paper, among other things, we compute the topological
centre of A ⊲⊳ θ I in terms of that of A and I. Using this, we provide a characterization of the
Arens regularity of A ⊲⊳ θ I. Then we determine the character space of A ⊲⊳ θ I in terms of that
of A and I. Moreover, we study the weak amenability of A ⊲⊳ θ I.
6
1
0
2
n
u
J
2
2
]
.
A
F
h
t
a
m
[
1
v
7
5
8
6
0
.
6
0
6
1
:
v
i
X
r
a
1. Introduction
Let A and B be Banach algebras, θ : A → B be a continuous Banach algebra homomorphism,
which without loss of generality we can assume that kθk ≤ 1, and let I be a closed ideal in B.
We consider the Banach algebra A ⊲⊳ θ I = {(a, i) : a ∈ A, i ∈ I}, the l1-direct sum of A and I,
with the following product formula:
(a, i) · (a′, i′) = (aa′, θ(a)i′ + iθ(a′) + ii′).
A ⊲⊳ θ I is called the amalgamation of A with B along I with respect to θ.
The algebraic version of amalgamated Banach algebras are studied by many algebraists, see
for example [7, 8, 9, 10, 20].
A special case of amalgamated Banach algebras, with I = B, is studied by some authors,
see [1, 3, 14], for example. To our knowledge there are no concrete Banach algebra with this
structure. While, as Example 2.1 shows, many classes of concrete Banach algebras can be
represented as amalgamated Banach algebras.
The organization of paper is as follows. In Section 2 of this paper, after presenting some ex-
amples of amalgamated Banach algebras we establish some primary properties of these algebras.
In Sections 3 and 4, we characterize the second dual and topological centres of A ⊲⊳ θ I as well
as its Arens regularity. In Section 5 we characterize the character space of A ⊲⊳ θ I. Finally,
Section 6 is devoted to the investigation of the weak amenability of A ⊲⊳ θ I.
The first author is supported by University of Tabriz.
2000 Mathematics Subject Classification. 46H25, 16E40, 13B02.
Key words and phrases. Banach modules, Topological centre, Weak amenability.
1
2
H. POURMAHMOOD AND N. SHIRMOHAMMADI
2. Some Examples and Primary Properties
We commence this section with listing a number of concrete Banach algebras that have the
amalgamated structure.
Example 2.1.
(i) If θ = 0, then A ⊲⊳ 0 I is nothing but the cartesian product of A and I.
(ii) Let A be a non-unital Banach algebra. Then the unitization of A, i.e. A# = C ⊕ A, is
the amalgamation of C with A# along A with respect to the homomorphism θ : C → A#
defined by θ(λ) = (λ, 0).
(iii) Let A be a Banach algebra and X be a Banach A-bimodule. Then the module extension
Banach algebra S = A ⊕ X is the amalgamation of A with S along X with respect to the
injection θ : A → S defined by θ(a) = (a, 0). Notice that the class of module extension
Banach algebras include the class of triangular Banach algebras.
(iv) Let A be a Banach algebra and φ be a nonzero character on A. Then A ⊲⊳ φ C is the
Banach algebra with the underlying Banach space A ⊕ C and with the product
(a, λ) · (a′, λ′) = (aa′, φ(a)λ′ + φ(a′)λ + λλ′).
(v) Let A and B be Banach algebras and let φ be a nonzero character on A. Then A ⊲⊳ θ B,
the amalgamation of A with B# along B with respect to the homomorphism θ : A → B#
defined by θ(a) = (φ(a), 0), is the Banach algebra with the underlying Banach space
A ⊕1 B, the l1-direct sum of A and B, and with the following product formula:
(a, b) · (a′, b′) = (aa′, φ(a)b′ + φ(a′)b + bb′).
This is a known Banach algebra denoted by A ⊕φ B, called the φ-Lau product of A and
B, see [21] for example. This class includes the class of Lau algebras introduced in [17].
(vi) One of the other interesting examples is the semidirect product of Banach algebras.
Indeed, let B be a Banach algebra, A be a closed subalgebra of B and I be a closed ideal
in B. If ι : A → B is the inclusion map, the amalgamated Banach algebra C = A ⊲⊳ ι I
is A ⋉ I, the semidirect product of A and I [6, Page 8] (as far as we know, the term
"semidirect product" in the theory of (commutative) Banach algebras is introduced and
studied by Thomas in [22]). We give an important class of Banach algebras which can be
recognized as a semidirect product. Let A be a dual Banach algebra with predual A∗ and
consider A∗∗, the second dual of A equipped with either first or second Arens product
(see Section 3 for definitions). It is shown in [6, Theorem 2.15] that A∗∗ = A ⋉ A⊥
∗ ,
∗ = {F ∈ A∗∗ : F = 0 on A∗}. We remark that every von Neumann algebra,
where A⊥
the measure algebra M (G) of a locally compact group G, and the second dual of an
Arens regular Banach algebra are examples of dual Banach algebras. Also the measure
algebra of a locally compact group G has a natural semidirect product structure. In fact
we have M (G) = l1(G) ⋉ Mc(G), where l1(G) and Mc(G) denote the space of discrete
measures and continuous measures in M (G), respectively.
ON AMALGAMATED BANACH ALGEBRAS
3
In the following proposition, we have collected some basic properties of the Banach algebra
A ⊲⊳ θ I.
Proposition 2.1. Let A ⊲⊳ θ I be the amalgamation of A with B along I with respect to θ.
(i) A ∼= A × {0} is a closed subalgebra of A ⊲⊳ θ I, I ∼= {0} × I is a closed ideal in A ⊲⊳ θ I
and A⊲⊳ θ I
I
∼= A.
(ii) A ⊲⊳ θ I is commutative if and only if A and θ(A) + I are commutative.
(iii) (a, i) is an identity for A ⊲⊳ θ I if and only if a = 1A, i2 = i, i ∈ AnnI (θ(A)) and
θ(a) + i = 1θ(A)+I , where AnnI (θ(A)) = {j ∈ I : jθ(a) = θ(a)j = 0 for all a ∈ A}.
(iv) ((aα, iα))α is a (bounded) left (right, or two-sided) approximate identity for A ⊲⊳ θ I if
and only if (aα)α is a (bounded) left (right, or two-sided) approximate identity for A,
(θ(aα) + iα)α is a (bounded) left (right, or two-sided) approximate identity for θ(A) + I
and iαθ(a) → 0 for all a ∈ A.
(v) If A ⊲⊳ θ I is commutative, then A ⊲⊳ θ I is regular if and only if both A and I are regular
(see [16, Definition 4.2.1]).
(vi) A ⊲⊳ θ I is amenable if and only if A and I are amenable (see [19, Definition 2.1.9]).
Proof. All of the parts (i)-(iv) can be easily checked. The part (v) follows from Theorems 4.2.6
∼= A. Finally, (vi) follows from Corollary 2.3.2, Theorem 2.3.7 and
and 4.3.8 of [16], since A⊲⊳ θ I
Theorem 2.3.10 of [19].
(cid:3)
I
Corollary 2.2. Let A ⊲⊳ id A be the amalgamation of A with A along A with respect to the
identity map id on A.
(i) A ⊲⊳ id A is commutative if and only if A is commutative.
(ii) (a, b) is an identity for A ⊲⊳ id A if and only if a = 1A and b = 0.
(iii) ((aα, bα))α is a (bounded) left (right, or two-sided) approximate identity for A ⊲⊳ id A if
and only if (aα)α is a (bounded) left (right, or two-sided) approximate identity for A and
bα → 0.
(v) If A is commutative, then A ⊲⊳ id A is regular if and only if A is regular.
(vi) A ⊲⊳ id A is amenable if and only if A is amenable.
3. The First and Second Arens Products on (A ⊲⊳ θ I)∗∗
Let A be a Banach algebra and A∗ and A∗∗ be the first and second duals of A, respectively.
Let a ∈ A, f ∈ A∗. Then a · f and f · a ∈ A∗ are defined by
ha · f, bi = hf, bai,
hf · a, bi = hf, abi
(b ∈ A),
making A∗ an A-bimodule. Similarly, A∗∗ is an A-bimodule.
There are two natural products on A∗∗, called the first and second Arens products, and are
denoted by (cid:3) and ♦, respectively. They were introduced by Arens [2] (for more details the
4
H. POURMAHMOOD AND N. SHIRMOHAMMADI
reader is refereed to [4]). We recall briefly the definitions. For f ∈ A∗ and F ∈ A∗∗ define
f · F ∈ A∗ and F · f ∈ A∗ by
hf · F, ai = hF, a · f i,
hF · f, ai = hF, f · ai
(a ∈ A).
Now, for F, G ∈ A∗∗, define F (cid:3)G ∈ A∗∗ and F ♦G ∈ A∗∗ by
hF (cid:3)G, f i = hF, G · f i,
hF ♦G, f i = hG, f · F i
(f ∈ A∗).
Then (A∗∗, (cid:3)) and (A∗∗, ♦) are Banach algebras containing A as a closed subalgebra.
Proposition 3.1. (A ⊲⊳ θ I)∗ is isometrically isomorphic to A∗ ⊕∞ I ∗ as Banach spaces. The
isomorphism Ψ : A∗ ⊕∞ I ∗ → (A ⊲⊳ θ I)∗ is given by
h(a, i), Ψ(f, g) i = f (a) + g(i)
((a, i) ∈ A ⊲⊳ θ I, (f, g) ∈ A∗ ⊕∞ I ∗).
Proof. The proof is straightforward and is omitted.
(cid:3)
Corollary 3.2. (A ⊲⊳ θ I)∗∗ is isometrically isomorphic to A∗∗ ⊕1 I ∗∗ as Banach spaces.
Now we explore the left and right module actions of A ⊲⊳ θ I on (A ⊲⊳ θ I)∗ in order to provide
a characterization of the first and second Arens product on (A ⊲⊳ θ I)∗∗.
Theorem 3.3. Let A ⊲⊳ θ I be the amalgamation of A with B along I with respect to θ. Then
((A ⊲⊳ θ I)∗∗, (cid:3)) = (A∗∗, (cid:3)) ⊲⊳ θ∗∗
(I ∗∗, (cid:3)),
where θ∗∗ is the second adjoint of θ.
Proof. Let a, b ∈ A, i, j ∈ I, f ∈ A∗ and g ∈ I ∗. Then (f, g)·(a, i) ∈ (A ⊲⊳ θ I)∗ can be calculated
as follows:
h(b, j), (f, g) · (a, i) i = h(a, i) · (b, j), (f, g) i
= h(ab, iθ(b) + θ(a)j + ij), (f, g) i
= h ab, f i + h θ(a)j, g i + h iθ(b), g i + h ij, g i
= h b, f · a i + h j, g · θ(a) i + h θ(b), g · i i + h j, g · i i
= h b, f · a i + h j, g · θ(a) i + h b, θ∗(g · i) i + h j, g · i i
= h b, f · a + θ∗(g · i) i + h j, g · (θ(a) + i) i,
and so
(3.1)
(f, g) · (a, i) = (f · a + θ∗(g · i), g · (θ(a) + i)).
ON AMALGAMATED BANACH ALGEBRAS
5
Further, let F1 ∈ A∗∗, F2 ∈ I ∗∗. Then, in order to calculate (F1, F2) · (f, g) ∈ (A ⊲⊳ θ I)∗, one
has
h(a, i), (F1, F2) · (f, g) i = h(f, g) · (a, i), (F1, F2) i
= h(f · a + θ∗(g · i), g · (θ(a) + i)), (F1, F2) i
= h f · a + θ∗(g · i), F1 i + h g · (θ(a) + i), F2 i
= h a, F1 · f i + h i, θ∗∗(F1) · g i + h θ(a), F2 · g i + h i, F2 · g i
= h a, F1 · f + θ∗(F2 · g) i + h i, F2 · g + θ∗∗(F1) · g i
= h(a, i), (F1 · f + θ∗(F2 · g), F2 · g + θ∗∗(F1) · g) i .
Thus
(3.2)
(F1, F2) · (f, g) = (F1 · f + θ∗(F2 · g), F2 · g + θ∗∗(F1) · g).
Now for (F1, F2), (G1, G2) ∈ (A ⊲⊳ θ I)∗∗ ∼= A∗∗ ⊕1 I ∗∗ and (f, g) ∈ (A ⊲⊳ θ I)∗ ∼= A∗ ⊕∞ I ∗, using
(3.1) and (3.2), we have
h(f, g), (F1, F2)(cid:3)(G1, G2) i = h(G1, G2) · (f, g), (F1, F2) i
= h(G1 · f + θ∗(G2 · g), G2 · g + θ∗∗(G1) · g), (F1, F2) i
= h G1 · f + θ∗(G2 · g), F1 i + h G2 · g + θ∗∗(G1) · g, F2 i
= h f, F1(cid:3)G1 i + h G2 · g, θ∗∗(F1) i + h g, F2(cid:3)G2 i + h θ∗∗(G1) · g, F2 i
= h f, F1(cid:3)G1 i + h g, θ∗∗(F1)(cid:3)G2 i + h g, F2(cid:3)G2 i + h g, F2(cid:3)θ∗∗(G1) i
= h f, F1(cid:3)G1 i + h g, θ∗∗(F1)(cid:3)G2 + F2(cid:3)G2 + F2(cid:3)θ∗∗(G1) i
= h(f, g), (F1(cid:3)G1, θ∗∗(F1(cid:3)G1) + θ∗∗(F1)(cid:3)G2 + F2(cid:3)G2 + F2(cid:3)θ∗∗(G1) i .
Therefore,
(F1, F2)(cid:3)(G1, G2) = (F1(cid:3)G1, θ∗∗(F1)(cid:3)G2 + F2(cid:3)G2 + F2(cid:3)θ∗∗(G1)).
This completes the proof.
(cid:3)
Similarly, as notation in the proof of Theorem 3.3, one can show that
(3.3)
(a, i) · (f, g) = (a · f + θ∗(i · g), (θ(a) + i) · g),
(f, g) · (F1, F2) = (f · F1 + θ∗(g · F2), g · F2 + g · θ∗∗(F1)),
(F1, F2)♦(G1, G2) = (F1♦G1, θ∗∗(F1)♦G2 + F2♦θ∗∗(G1) + F2♦G2).
Therefore,
((A ⊲⊳ θ I)∗∗, ♦) = (A∗∗, ♦) ⊲⊳ θ∗∗
(I ∗∗, ♦).
6
H. POURMAHMOOD AND N. SHIRMOHAMMADI
4. Topological Centres
Let A be a Banach algebra and X a Banach A-bimodule. Then X ∗∗ is canonically an (A∗∗, (cid:3))-
′′ : (A∗∗, (cid:3)) → X ∗∗
bimodule ((A∗∗, ♦)-bimodule), see [4, Page 248]. Let x′′ ∈ X ∗∗ and let Lx
be the left and right multiplication operators, respectively, i.e.
′′ , Rx
′′
Lx′′ (a
) = x
′′
′′
(cid:3)a
= lim
β
lim
α
xβaα
and Rx′′ (a
′′
) = a
′′
′′
(cid:3)x
= lim
α
lim
β
aαxβ
′′
(a
∈ A∗∗),
where (aα) and (xβ) are nets in A∗∗ and X ∗∗, respectively, in such a way that aα → a
w∗-topology of A∗∗ and xβ → x
Likewise let Lx′′ , Rx′′ : (A∗∗, ♦) → X ∗∗ be the left and right multiplication operators, respec-
tively, i.e.
in the w∗-topology of X ∗∗.
in the
′′
′′
Lx′′ (a
′′
) = x
′′
′′
♦a
= lim
α
lim
β
xβaα
and Rx′′ (a
′′
) = a
′′
′′
♦x
= lim
β
lim
α
aαxβ
′′
(a
∈ A∗∗).
The left and right topological centres, Z ℓ,t
A (X ∗∗) and Z r,t
A (X ∗∗) of X ∗∗ are
Z ℓ,t
A (X ∗∗) = {x′′ ∈ X ∗∗ : Lx′′ = Lx′′} = {x′′ ∈ X ∗∗ : x′′(cid:3)a′′ = x′′♦a′′, ∀a′′ ∈ A∗∗},
and
Z r,t
A (X ∗∗) = {x′′ ∈ X ∗∗ : Rx′′ = Rx′′} = {x′′ ∈ X ∗∗ : a′′(cid:3)x′′ = a′′♦x′′, ∀a′′ ∈ A∗∗},
respectively. Then we say that X is Arens regular (as an A-bimodule) or A acts regularly on X
if
A (X ∗∗) = X ∗∗,
Z ℓ,t
A (X ∗∗) = Z r,t
and X is left strongly Arens irregular if Z ℓ,t
Z r,t
A (X ∗∗) = X, right strongly Arens irregular if
A (X ∗∗) = X, and strongly Arens irregular if it is both left and right strongly Arens irregular.
If X = A, we will use the common notation Z ℓ
A (A∗∗) and
A (A∗∗), respectively.
Now, let B be a Banach algebra, I be a closed ideal in B and θ : A → B be a continuous
t (A∗∗) in place of Z ℓ,t
t (A∗∗) and Z r
Z r,t
Banach algebra homomorphism. Then we define
Z ℓ
θ∗∗(A∗∗) = {F ∈ Z ℓ
t (A∗∗) : θ∗∗(F ) ∈ Z ℓ
I (θ(A)∗∗)},
note that θ(A)∗∗ = θ∗∗(A∗∗) ([4, Page 251]), where
Z ℓ
I (θ(A)∗∗) = {F ∈ θ(A)∗∗ : LF = LF on I ∗∗} = {F ∈ θ(A)∗∗ : F (cid:3)G = F ♦G, ∀G ∈ I ∗∗}.
Theorem 4.1. With above notation and assumptions, one has
Z ℓ
t ((A ⊲⊳ θ I)∗∗) = Z ℓ
t (A∗∗ ⊲⊳ θ∗∗
I ∗∗) = Z ℓ
θ∗∗(A∗∗) ⊲⊳ θ∗∗
(Z ℓ
t (I ∗∗) ∩ Z ℓ,t
θ(A)(I ∗∗)).
ON AMALGAMATED BANACH ALGEBRAS
7
Proof. Let (F1, G1) ∈ Z ℓ
A∗∗ ⊲⊳ θ∗∗
I ∗∗,
t (A∗∗ ⊲⊳ θ∗∗
I ∗∗). Then L(F1,G1) = L(F1,G1) if and only if for all (F2, G2) ∈
if and only if
(F1, G1)(cid:3)(F2, G2) = (F1, G1)♦(F2, G2),
(F1(cid:3)F2, θ∗∗(F1)(cid:3)G2 + G1(cid:3)θ∗∗(F2) + G1(cid:3)G2) = (F1♦F2, θ∗∗(F1)♦G2 + G1♦θ∗∗(F2) + G1♦G2),
if and only if LF1 = LF1, Lθ∗∗(F1) = Lθ∗∗(F1) on I ∗∗, LG1 = LG1 on I ∗∗ and θ(A)∗∗. Hence,
(F1, G1) ∈ Z ℓ
θ∗∗(A∗∗) ⊲⊳ θ∗∗
(Z ℓ
t (I ∗∗) ∩ Z ℓ,t
θ(A)(I ∗∗)).
(cid:3)
Using above theorem we characterize Arens regularity and strong Arens irregularity of A ⊲⊳ θ I.
Corollary 4.2. A ⊲⊳ θ I is Arens regular if and only if A and I are Arens regular, θ(A) acts
regularly on I and I acts regularly on θ(A).
Corollary 4.3. A ⊲⊳ θ I is strongly Arens irregular if and only if A and I are strongly Arens
irregular, I acts strongly irregular on θ(A) and θ(A) acts strongly irregular on I.
In the following example we determine topological centres of some amalgamated Banach
algebras.
Example 4.1. Keep the notation of Example 2.1.
(i) If θ = 0, then
Z ℓ
t ((A ⊕ I)∗∗) = Z ℓ
t ((A ⊲⊳ 0 I)∗∗) = Z ℓ
t (A∗∗) ⊲⊳ 0 Z ℓ
t (I ∗∗) = Z ℓ
t (A∗∗) ⊕ Z ℓ
t (I ∗∗).
t (A∗∗) = Z ℓ
t ((A#)∗∗) = C ⊕ Z ℓ
(ii) Z ℓ
(iii) ([11]) Let S = A ⊕ X be the module extension Banach algebra corresponding A and X.
Then, by noting that θ is the canonical embedding of A into S, and I = X with X 2 = 0,
we have
t (A∗∗)#.
Z ℓ
θ∗∗(A∗∗) = {F ∈ Z ℓ
t (A∗∗) : F ∈ Z ℓ
X(A∗∗)} = Z ℓ
t (A∗∗) ∩ Z ℓ
X(A∗∗),
and
t (I ∗∗) ∩ Z ℓ,t
Z ℓ
θ(A)(I ∗∗) = Z ℓ
t (X ∗∗) ∩ Z ℓ,t
A (X ∗∗) = Z ℓ
t (X ∗∗) = X ∗∗.
Therefore,
Z ℓ
t (S ∗∗) = (Z ℓ
t (A∗∗) ∩ Z ℓ
X(A∗∗)) ⊲⊳θ∗∗
X ∗∗,
t (S ∗∗) is the module extension Banach algebra corresponding Z ℓ
t (A∗∗)∩Z ℓ
X(A∗∗)
that is, Z ℓ
and X ∗∗.
(iv) Z ℓ
t ((A ⊲⊳ φ C)∗∗) = Z ℓ
t (A∗∗) ⊲⊳ φ C. Details are similar to details of the next general case.
8
H. POURMAHMOOD AND N. SHIRMOHAMMADI
(v) ([21, Corollary 2.13]) For computing Z ℓ
t ((A ⊕φ B)∗∗) we note that since θ : A → B#
is defined by θ(a) = (φ(a), 0) = φ(a), one can easily check that θ∗∗ : A∗∗ → (B#)∗∗ =
C ⊕ B ∗∗ is given by θ∗∗(F ) = (F (φ), 0) = F (φ) = φ(F ). So
Z ℓ
θ∗∗(A∗∗) = Z ℓ
φ(A∗∗) = Z ℓ
t (A∗∗),
and
Whence
θ(A)(I ∗∗) = Z ℓ,t
Z ℓ,t
C (B ∗∗) = B ∗∗.
Z ℓ
t ((A ⊕φ B)∗∗) = Z ℓ
t (A∗∗) ⊲⊳θ∗∗
Z ℓ
t (B ∗∗) = Z ℓ
t (A∗∗) ⊕φ Z ℓ
t (B ∗∗).
(vi) If I = B and θ is surjective, then
Z ℓ
t ((A ⊲⊳ θ B)∗∗) = (Z ℓ
t (A∗∗) ∩ (θ∗∗)−1(Z ℓ
t (B ∗∗))) ⊲⊳id Z ℓ
t (B ∗∗).
(vii) Assume that B = A. Then
Z ℓ
t ((A ⊲⊳ id I)∗∗) = (Z ℓ
t (A∗∗) ∩ Z ℓ
I (A∗∗)) ⊲⊳id (Z ℓ
t (I ∗∗) ∩ Z ℓ,t
A (I ∗∗)).
t ((A ⊲⊳ id A)∗∗) = Z ℓ
(ix) Z ℓ
(x) Let B = I = A∗∗ and let ι : A → A∗∗ be the cononical injection. Then
t (A∗∗) ⊲⊳ id Z ℓ
t (A∗∗).
Z ℓ
A∗∗(ι(A)∗∗) = Z ℓ
A∗∗(A∗∗) = A∗∗,
and so Z ℓ
ι∗∗(A∗∗) = Z ℓ
t (A∗∗). Also Z ℓ,t
ι(A)((A∗∗)∗∗) = Z ℓ,t
A (A∗∗∗∗) = A∗∗∗∗, and thus
Z ℓ
t ((A ⊲⊳ ι A∗∗)∗∗) = Z ℓ
t (A∗∗) ⊲⊳ ι∗∗
Z ℓ
t (A∗∗∗∗).
Example 4.2. By Example 4.1 in mind we have the followings:
(i) The Banach algebra A ⊲⊳ 0 I is Arens regular if and only if A and I are Arens regular.
(ii) The unitization of A, A#, is Arens regular if and only if A is Arens regular.
(iii) The module extension Banach algebra S = A ⊕ X is Arens regular if and only if A is
Arens regular and A acts regularly on X.
(iv) A ⊲⊳ φ C is Arens regular if and only if A is Arens regular.
(v) A ⊕φ B is Arens regular if and only if A and B are Arens regular.
(vi) If I = B and θ is surjective, then A ⊲⊳ θ B is Arens regular if and only if A and B are
Arens regular.
(vii) If B = A, then A ⊲⊳ θ I is Arens regular if and only if A is Arens regular and A and I
act regularly on each other.
(ix) The Banach algebra A ⊲⊳ id A is Arens regular if and only if A is Arens regular.
(x) The Banach algebra A ⊲⊳ ι A∗∗ is Arens regular if and only if A and A∗∗ are Arens
regular.
ON AMALGAMATED BANACH ALGEBRAS
9
Example 4.3. Let G be an infinite locally compact group. Then by [18, Theorem 1] we have
Z ℓ
t (L1(G)∗∗) = L1(G), and so
Z ℓ
t (L1(G)∗∗) = L1(G) ⊲⊳ id L1(G).
t ((L1(G) ⊲⊳ id L1(G))∗∗) = Z ℓ
t (L1(G)∗∗) ⊲⊳ id Z ℓ
Therefore, L1(G) ⊲⊳ id L1(G) is strongly Arens irregular.
5. Characters of A ⊲⊳ θ I
In this section, first, we provide a characterization of character space of A ⊲⊳ θ I, and then we
calculate the Jacobson radical of A ⊲⊳ θ I when it is commutative.
Theorem 5.1. Let σ(A) 6= ∅ and θ(A)I ∪ Iθ(A) = I. Then σ(A ⊲⊳ θ I) = E ∪ F , where
E = {((i · ψ) ◦ θ, ψ) : ψ ∈ σ(I), i ∈ I, ψ(i) = 1}, F = {(φ, 0) : φ ∈ σ(A)}.
Moreover, E is open and F is closed in σ(A ⊲⊳ θ I).
Proof. Let (φ, ψ) ∈ σ(A ⊲⊳ θ I) and (a, i), (a′, i′) ∈ A ⊲⊳ θ I. Then
φ(aa′) + ψ(θ(a)i′ + iθ(a) + ii′) = h(φ, ψ), (aa′, θ(a)i′ + iθ(a) + ii′) i
(5.1)
= h(φ, ψ), ((a, i) · (a′, i′) i
= (φ(a) + ψ(i))(φ(a′) + ψ(i′))
= φ(a)φ(a′) + φ(a)ψ(i′) + ψ(i)φ(a′) + ψ(i)ψ(i′).
By taking a = a′ = 0 we see that ψ ∈ σ(I) ∪ {0}. Next by taking i = i′ = 0 it follows that
φ ∈ σ(A) ∪ {0}. But from (5.1), φ = 0 implies ψ(θ(a)i′) + ψ(iθ(a′)) = 0 for all a, a′ ∈ A and
i, i′ ∈ I, from which it follows that ψ = 0 on θ(A)I ∪ Iθ(A), and hence ψ = 0. But this is a
contradiction since (φ, ψ) ∈ σ(A ⊲⊳ θ I) and so (φ, ψ) 6= (0, 0). Now we have two cases:
Case I: If ψ = 0, then (φ, ψ) = (φ, 0).
Case II: If ψ 6= 0, then by (5.1) we have
ψ(θ(a)i′) + ψ(iθ(a′)) − ψ(i)φ(a′) − φ(a)ψ(i′) = 0,
which implies (take a′ = 0, i = 0),
ψ(θ(a)i′) = φ(a)ψ(i′)
(a ∈ A, i′ ∈ I).
Choose i′ ∈ I such that ψ(i′) = 1, then φ(a) = ψ(θ(a)i′) = (i′ · ψ) ◦ θ(a) for all a ∈ A. Therefore,
(φ, ψ) = ((i′ · ψ) ◦ θ, ψ) with ψ(i′) = 1.
Since the reverse inclusion is easy to check, so we omit its proof.
Now we show that E is open in the w∗-topology of σ(A ⊲⊳ θ I) induced from w∗-topology of
A∗ × I ∗. Let ((i · ψ0) ◦ θ, ψ0) ∈ σ(A ⊲⊳ θ I). Then there is i0 ∈ I in such a way that ψ(i0) 6= 0.
Let ε = ψ(i0). Then
U = {(φ, ψ) ∈ σ(A ⊲⊳ θ I) : (φ, ψ)(0, i0) − ((i · ψ0) ◦ θ, ψ0)(0, i0) < ε}
= {(φ, ψ) ∈ σ(A ⊲⊳ θ I) : ψ(i0) − ψ0(i0) < ε},
10
H. POURMAHMOOD AND N. SHIRMOHAMMADI
is a neighborhood of ((i · ψ0) ◦ θ, ψ0) in the w∗-topology of σ(A ⊲⊳ θ I). Since (φ, 0) ∈ U leads to
the contradiction ψ0(i0) < ε, it follows that U ⊆ E. Therefore, E is open and F is closed. (cid:3)
Corollary 5.2. ([21, Proposition 2.4]) Let σ(A) 6= ∅ and φ ∈ σ(A). Then
σ(A ⊕φ B) = {(ϕ, 0) : ϕ ∈ σ(A)} ∪ {(φ, ψ) : ψ ∈ σ(B)} = (σ(A) × {0}) ∪ ({φ} × σ(B)).
Proof. It is enough to note that (i · ψ) ◦ θ = θ if ψ(i) = 1.
(cid:3)
Let A be a commutative Banach algebra. The radical of A, rad A, is the intersection of the
kernels of all characters of A. Also A is called semisimple if rad A = {0}.
Theorem 5.3. Let A ⊲⊳ θ I be commutative, σ(A) 6= ∅ and θ(A)I = I. Then rad(A ⊲⊳ θ I) =
rad A ⊕ rad I.
Proof. Let (a, i) ∈ rad(A ⊲⊳ θ I). Then for each φ ∈ σ(A), φ(a) = (φ, 0)(a, i) = 0, that is,
a ∈ rad A. Now let ψ ∈ σ(I) and ψ(j) = 1 for some j ∈ I. Then (j · ψ) ◦ θ belongs to σ(A) and
so (j · ψ) ◦ θ(a) = 0. Hence
ψ(i) = ((j · ψ) ◦ θ, ψ)(a, i) − (j · ψ) ◦ θ(a) = 0,
and thus i ∈ rad I.
Conversely let a ∈ rad A and i ∈ rad I. Then for each φ ∈ σ(A), (φ, 0)(a, i) = φ(a) = 0 and
for each ψ ∈ σ(I),
((j · ψ) ◦ θ, ψ)(a, i) = (j · ψ) ◦ θ(a) + ψ(i) = 0.
Therefore, by Theorem 5.1, (a, i) ∈ rad(A ⊲⊳ θ I).
(cid:3)
Corollary 5.4. Let A ⊲⊳ θ I be commutative, σ(A) 6= ∅ and θ(A)I = I. Then A ⊲⊳ θ I is
semisimple if and only if both A and I are semisimple.
6. Weak Amenability
Let A be a Banach algebra and X a Banach A-bimodule. A derivation from A into X
is a bounded linear map satisfying D(ab) = a · D(b) + D(a) · b for all a, b ∈ A. For each
x ∈ X we denote by adx the derivation D(a) = a · x − x · a for all a ∈ A, called an inner
derivation. We denote by Z 1(A, X) the space of all derivations from A into X and by B1(A, X)
the space of all inner derivations from A into X. The first cohomology group of A with coefficients
in X is H1(A, X) = Z 1(A, X)/ B1(A, X). A Banach algebra A is called weakly amenable if
H1(A, A∗) = 0.
In [12], B. E. Forrest and L. W. Marcoux have investigated the weak amenability of triangular
Banach algebras, and Y. Zhang has studied the weak amenability of module extension Banach
algebras [23]. Motivated by these earlier investigations, in this section, we study the weak
amenability of amalgamated Banach algebra A ⊲⊳ θ I.
ON AMALGAMATED BANACH ALGEBRAS
11
Theorem 6.1. If A ⊲⊳ θ I is commutative, then A ⊲⊳ θ I is weakly amenable if and only if A and
I are weakly amenable.
Proof. This is immediate by Proposition 2.1, [4, Propositions 2.8.64 and 2.8.65(ii) and Theorem
2.8.69(i)] and noting that A⊲⊳ θ I
∼= A.
I
(cid:3)
Example 6.1. Let G be a locally compact abelian group and consider M (G) = l1(G) ⋉ Mc(G),
the semidirect product of l1(G) and Mc(G); see Example 2.1(iv). Since l1(G) is weakly amenable,
by above theorem, M (G) is weakly amenable if and only if Mc(G) is weakly amenable.
In general case, we have one direction of Theorem 6.1.
Proposition 6.2. If A and I are weakly amenable, then A ⊲⊳ θ I is also weakly amenable.
Proof. It follows immediately from Proposition 2.1 and [4, Proposition 2.8.65(ii)].
(cid:3)
The converse of above proposition does not hold in general. Indeed, it is shown in [15] that
the augmentation ideal I of L1(SL(2, R)) is not weakly amenable and that its unitization I # is
weakly amenable.
Example 6.2. Let G be a locally compact group. Since M (G) = l1(G) ⋉ Mc(G) and l1(G) is
always weakly amenable, by Proposition 6.2, M (G) is weakly amenable, provided that Mc(G)
is weakly amenable.
Example 6.3. Let G be a locally compact group. Then l1(G) ⋉ L1(G) is weakly amenable.
In order to prove a partial converse of Proposition 6.2 we first look at derivations from A to
A∗.
Proposition 6.3. H1(A, A∗) embeds in H1(A ⊲⊳ θ I, (A ⊲⊳ θ I)∗).
Proof. Every D ∈ Z 1(A, A∗) defines a derivation D : A ⊲⊳ θ I → (A ⊲⊳ θ I)∗ by D(a, i) =
(D(a), 0), and it can be easily checked that D is inner if and only if D is inner. It follows that
the mapping D 7→ D induces an embedding from H1(A, A∗) into H1(A ⊲⊳ θ I, (A ⊲⊳ θ I)∗).
(cid:3)
Corollary 6.4. If A ⊲⊳ θ I is weakly amenable, then so is A.
The following corollary has been obtained in [13] with a different method. In fact, we have
given a short proof for this result.
Corollary 6.5. ([13, Theorem 2.2]) Let A be a dual Banach algebra. If A∗∗ is weakly amenable,
then so is A.
Weak amenability of module extension Banach algebras is extensively studied in [23]. We are
going to characterize the weak amenability of Banach algebras A ⊲⊳ id A and A ⊕φ B.
12
H. POURMAHMOOD AND N. SHIRMOHAMMADI
6.1. Weak Amenability of A ⊲⊳ id A. Here, we focus on the special case A ⊲⊳ id A.
Proposition 6.6. Let A2 be dense in A. Then D ∈ Z 1(A ⊲⊳ id A, (A ⊲⊳ id A)∗) if and only if
D(a, b) = (D1(a) + D2(b), D2(a) + D2(b))
(a, b ∈ A),
for some D1, D2 ∈ Z 1(A, A∗). Moreover, D = ad(f,g) if and only if D1 = adf and D2 = adg,
where f, g ∈ A∗.
Proof. Let D : A ⊲⊳ id A → (A ⊲⊳ id A)∗ be a derivation. Then we may write
D(a, b) = (D1(a) + D2(b), D3(a) + D4(b))
(a, b ∈ A),
where Dk : A → A∗ (1 ≤ k ≤ 4) is a linear operator. If we use the derivation property of D
together with the equations (3.1) and (3.3), we get
(cid:0)D1(a1a2) + D2(a1b2 + b1a2 + b1b2), D3(a1a2) + D4(a1b2 + b1a2 + b1b2)(cid:1) =
(cid:0)a1D1(a2) + a1D2(b2) + b1D3(a2) + b1D4(b2) + D1(a1)a2 + D2(b1)a2 + D3(a1)b2 + D4(b1)b2,
(a1[D3(a2) + D4(b2)] + b1[D3(a2) + D4(b2)] + [D3(a1) + D4(b1)]a2 + [D3(a1) + D4(b1)]b2(cid:1).
By setting a1 = a2 = 0, we see that D1, D3 ∈ Z 1(A, A∗). By setting b1 = b2 = 0 and noting
that A2 is dense in A, we get D2 = D4 ∈ Z 1(A, A∗). Also, by choosing a2 = b1 = 0, we obtain
D2(a1b2) = a1 · D2(b2) + D3(a1) · b2,
which implies D2(a1) · b2 = D3(a1) · b2 for all a1, b2 ∈ A. Since A2 is dense in A, it follows that
D2 = D3. The claim about inner derivations can be easily verified.
(cid:3)
Theorem 6.7. Let A2 be dense in A. Then, as vector spaces, we have
H1(A ⊲⊳ id A, (A ⊲⊳ id A)∗) ∼= H1(A, A∗) ⊕ H1(A, A∗).
Proof. Define ϕ : Z 1(A, A∗) ⊕ Z1(A, A∗) → Z 1(A ⊲⊳ id A, (A ⊲⊳ id A)∗) by ϕ(D1, D2) = D, where
D(a, b) = (D1(a) + D2(b), D2(a) + D2(b))
(a, b ∈ A).
The Proposition 6.6 shows that ϕ is well defined and onto. Since D is inner if and only if D1
and D2 are inner, according to the Proposition 6.6, ϕ induces the desired isomorphism.
(cid:3)
Since every weakly amenable Banach algebra is square dense, we have the following corollary.
Corollary 6.8. The Banach algebra A ⊲⊳ id A is weakly amenable if and only if A is weakly
amenable.
Example 6.4. Let A be a C ∗-algebra or a group algebra of a locally compact group. Then
A ⊲⊳ id A is weakly amenable.
ON AMALGAMATED BANACH ALGEBRAS
13
6.2. Weak Amenability of A ⊕φ B. Let (a, b) ∈ A ⊕φ B and (f, g) ∈ (A ⊕φ B)∗. Then
(a, b) · (f, g), (f, g) · (a, b) ∈ (A ⊕φ B)∗ are given by
(6.1)
(6.2)
(a, b) · (f, g) = (a · f + g(b)φ, φ(a)g + b · g),
(f, g) · (a, b) = (f · a + g(b)φ, φ(a)g + g · b).
Proposition 6.9. Let B2 be dense in B. Then D ∈ Z 1(A ⊕φ B, (A ⊕φ B)∗) if and only if
D(a, b) = (D1(a) + D2(b), D4(b))
(a ∈ A, b ∈ B),
such that
(i) D1 ∈ Z 1(A, A∗),
(ii) D4 ∈ Z 1(B, B ∗),
(iii) D2 : B → A is a bounded linear map satisfying
(1) a · D2(b) = D2(b) · a = φ(a)D2(b) for all a ∈ A and b ∈ B,
(2) D2(bb′) = h b, D4(b′) i φ + h b′, D4(b) i φ for all b, b′ ∈ B.
Moreover, D = ad(f,g) if and only if D1 = adf , D2 = 0 and D4 = adg (f ∈ A∗, g ∈ B ∗).
Proof. Let D : A ⊕φ B → (A ⊕φ B)∗ ∼= A∗ ⊕∞ B ∗ be a derivation. Then D is of the form
D(a, b) = (D1(a) + D2(b), D3(a) + D4(b))
(a ∈ A, b ∈ B),
where D1 : A → A∗, D2 : B → A∗, D3 : A → B ∗ and D4 : B → B ∗ are linear operators. If we
use the derivation property of D together with the equations (6.1) and (6.2), we get
(cid:0)D1(aa′) + φ(a)D2(b′) + φ(a′)D2(b) + D2(bb′), D3(aa′) + φ(a)D4(b′) + φ(a′)D4(b) + D4(bb′)(cid:1) =
(cid:0)aD1(a′) + aD2(b′) + h b, D3(a′) i φ + h b, D4(b′) i φ, φ(a)D3(a′) + φ(a)D4(b′) + bD3(a′) + bD4(b′))+
(D1(a)a′ + D2(b)a′ + h b′, D3(a) i φ + h b′, D4(b) i φ, φ(a′)D3(a) + φ(a′)D4(b)] + D3(a)b′ + D4(b)b′(cid:1).
By setting b = b′ = 0 we see that D1 ∈ Z 1(A, A∗) and D3 ∈ Z 1
obtains D4 ∈ Z 1(B, B ∗) and D2(bb′) = h b, D4(b′) i φ + h b′, D4(b) i φ.
Now put a = b′ = 0. Then we get b · D3(a′) = 0 in B ∗ which implies D3 = 0 by density of B2 in
B. Hence D2(b) · a′ = φ(a′)D2(b). Similarly, choosing a′ = b = 0 gives a · D2(b′) = φ(a)D2(b′).
Using (6.1) and (6.2) one can easily see that D = ad(f,g) if and only if D1 = adf , D2 = 0 and
(cid:3)
φ(A, A∗). Letting a = a′ = 0 one
D4 = adg.
Let B be a Banach algebra. A derivation D : B → B ∗ is called cyclic if
h b, D(b′) i + h b′, D(b) i = 0
for all b, b′ ∈ B.
We denote by Z 1
cyclic cohomology group of B is H1
c (B, B ∗) the space of all cyclic derivations which includes B1(B, B ∗). The first
c (B, B ∗) = Z 1
c (B, B ∗)/B1(B, B ∗).
Theorem 6.10. H1(A, A∗) ⊕ H1
c (B, B ∗) embeds in H1(A ⊕φ B, (A ⊕φ B)∗).
14
H. POURMAHMOOD AND N. SHIRMOHAMMADI
Proof. Define ψ : Z 1(A, A∗) ⊕ Z 1
c (B, B ∗) −→ Z 1(A ⊕φ B, (A ⊕φ B)∗) by ψ(D1, D2) = D, where
D(a, b) = (D1(a), D4(b))
(a ∈ A, b ∈ B).
It follows from Proposition 6.9 that D is a derivation and it is inner if and only if D1 and D2 are
inner. So ψ induces an injective linear map from H1(A, A∗) ⊕ H1
c (B, B ∗) into H1(A ⊕φ B, (A ⊕φ
B)∗).
(cid:3)
Corollary 5.6 of [15] shows that in general H1(B, B ∗) does not embeds into H1(A ⊕φ B, (A ⊕φ
B)∗), and thus it seems that Theorem 6.10 be the best that one could expect.
Corollary 6.11. ([21, Theorem 2.11]) If A ⊕φ B is weakly amenable, then A is weakly amenable
and B is cyclicly amenable.
References
[1] F. Abtahi, A. Ghafarpanah and A. Rejali, Biprojectivity and biflatness of Lau product of Banach algebras
defined by a Banach algebra morphism, Bull. Aust. Math. Soc. 91 (2015), 134-144.
[2] R. Arens, The adjoint of a bilinear operator, Proc. Amer. Math. Soc. 2 (1951), 839-848.
[3] S. J. Bhatt and P. A. Dabhi, Arens regularity and amenability of Lau product of Banach algebras defined
by a Banach algebra morphism, Bull. Aust. Math. Soc. 87 (2013), 195-206.
[4] H. G. Dales, Banach algebras and automatic continuity, Clarendon Press, Oxford, 2000.
[5] H. G. Dales, F. Ghahramani and N. Grønbaek, Derivations into iterated duals of Banach algebras, Studia
Math. 128 (1998), 19-54.
[6] H. G. Dales and A. T.-M. Lau, The second duals of Beurling Banach algebras, Mem. Amer. Math. Soc.
177 (836) (2005).
[7] M. D'Anna, C. A. Finocchiaro and M. Fontana, Amalgamated algebras along an ideal, in: Commutative
Algebra and Applications, Proceedings of the Fifth International Fez Conference on Commutative Algebra
and Applications, Fez, Morocco (2008), W. de Gruyter Publisher, Berlin (2009), 155172.
[8] M. D'Anna, C. A. Finocchiaro and M. Fontana, Properties of chains of prime ideals in an amalgamated
algebra along an ideal, J. Pure Appl. Algebra, 214 (2010), 1633-1641.
[9] M. D'Anna, C. A. Finocchiaro and M. Fontana, New algebraic properties of an amalgamated algebra
along an ideal, to appear in Comm. Algebra.
[10] M. D'Anna and M. Fontana, An amalgamated duplication of a ring along an ideal: the basic properties, J.
Algebra Appl. 6 (2007), 443-459.
[11] M. Eshaghi Gordji and M. Filali, Arens regularity of module actions, Studia Math. 181 (2007), 237-254.
[12] B. E. Forrest and L. W. Marcoux, Weak amenability of triangular Banach algebras, Trans. Amer. Math.
Soc. 354 (2002), 1435-1452.
[13] F. Ghahramani and J. Laali, Amenability and topological centres of the second duals of Banach algebras,
Bull. Aust. Math. Soc. 65 (2002), 191-197.
[14] H. Javanshiri and M. Nemati, On a Certain product of Banach algebras and some of its properties, Proc.
Rom. Acad. Ser. A, 15 (2014), 219-227.
[15] B. E. Johnson and M. C. White, A non-weakly amenable augmentation ideal, Preprint.
[16] E. Kaniuth, A Course in Commutative Banach Algebras, Springer, New York, 2009.
[17] A. T.-M. Lau, Analysis on a class of Banach algebras with applications to harmonic analysis on locally
compact groups and semigroups, Fund. Math. 118 (1983), 161-175.
ON AMALGAMATED BANACH ALGEBRAS
15
[18] A. T.-M. Lau and V. Losert, On the second conjugate algebra of L1(G) of a locally compact group, J.
London Math. Soc. 37 (1988) 464-470.
[19] V. Runde, Lectures on amenability, Lecture Notes in Mathematics, Springer, Berlin, 2002.
[20] P. Sahandi, N. Shirmohammadi and S. Sohrabi, Cohen-Macaulay and Gorenstein properties under the
amalgamated construction, Comm. Algebra, 44 (3) (2016), 1096-1109.
[21] M. Sangani Monfared, On certain products of Banach algebras with applications to harmonic analysis,
Studia Math. 178 (3) (2007), 277-294.
[22] M. P. Thomas, Principal ideals and semi-direct products in commutative Banach algebras, J. Funct. Analysis
101 (1991), 312-328.
[23] Y. Zhang, Weak Amenability of Module Extensions of Banach Algebras, Trans. Amer. Math. Soc. 354 (2002),
4131-4151.
Department of Mathematics, University of Tabriz, Tabriz, Iran.
E-mail address: h p [email protected], h [email protected]
E-mail address: [email protected]
|
1302.1978 | 2 | 1302 | 2013-07-21T05:57:06 | Applications of Convex Analysis within Mathematics | [
"math.FA",
"math.OC"
] | In this paper, we study convex analysis and its theoretical applications. We first apply important tools of convex analysis to Optimization and to Analysis. We then show various deep applications of convex analysis and especially infimal convolution in Monotone Operator Theory. Among other things, we recapture the Minty surjectivity theorem in Hilbert space, and present a new proof of the sum theorem in reflexive spaces. More technically, we also discuss autoconjugate representers for maximally monotone operators. Finally, we consider various other applications in mathematical analysis. | math.FA | math |
Applications of Convex Analysis within Mathematics
Francisco J. Arag´on Artacho∗, Jonathan M. Borwein†, Victoria Mart´ın-M´arquez‡,
and Liangjin Yao§
July 19, 2013
Abstract
In this paper, we study convex analysis and its theoretical applications. We first apply
important tools of convex analysis to Optimization and to Analysis. We then show various
deep applications of convex analysis and especially infimal convolution in Monotone Operator
Theory. Among other things, we recapture the Minty surjectivity theorem in Hilbert space, and
present a new proof of the sum theorem in reflexive spaces. More technically, we also discuss
autoconjugate representers for maximally monotone operators. Finally, we consider various
other applications in mathematical analysis.
2010 Mathematics Sub ject Classification:
Primary 47N10, 90C25; Secondary 47H05, 47A06, 47B65
Keywords: Adjoint, Asplund averaging, autoconjugate representer, Banach limit, Chebyshev set,
convex functions, Fenchel duality, Fenchel conjugate, Fitzpatrick function, Hahn–Banach extension
theorem, infimal convolution, linear relation, Minty surjectivity theorem, maximally monotone
operator, monotone operator, Moreau’s decomposition, Moreau envelope, Moreau’s max formula,
Moreau–Rockafellar duality, normal cone operator, renorming, resolvent, Sandwich theorem, sub-
differential operator, sum theorem, Yosida approximation.
1
Introduction
While other articles in this collection look at the applications of Moreau’s seminal work, we have
opted to illustrate the power of his ideas theoretically within optimization theory and within math-
ematics more generally. Space constraints preclude being comprehensive, but we think the presen-
tation made shows how elegantly much of modern analysis can be presented thanks to the work of
Jean-Jacques Moreau and others.
∗Centre for Computer Assisted Research Mathematics and its Applications (CARMA), University of Newcastle,
Callaghan, NSW 2308, Australia. E-mail: [email protected]
†Centre for Computer Assisted Research Mathematics and its Applications (CARMA), University of Newcas-
tle, Callaghan, NSW 2308, Australia. E-mail: [email protected]. Laureate Professor at the
University of Newcastle and Distinguished Professor at King Abdul-Aziz University, Jeddah.
‡Departamento de An´alisis Matem´atico, Universidad de Sevilla, Spain. E-mail: [email protected]
§Centre for Computer Assisted Research Mathematics and its Applications (CARMA), University of Newcastle,
Callaghan, NSW 2308, Australia. E-mail: [email protected].
1
1.1 Preliminaries
Let X be a real Banach space with norm (cid:107) · (cid:107) and dual norm (cid:107) · (cid:107)∗ . When there is no ambiguity
we suppress the ∗. We write X ∗ and (cid:104) · , · (cid:105) for the real dual space of continuous linear functions
and the duality paring, respectively, and denote the closed unit ball by BX := {x ∈ X (cid:107)x(cid:107) ≤ 1}
and set N := {1, 2, 3, . . .}. We identify X with its canonical image in the bidual space X ∗∗ . A set
C ⊆ X is said to be convex if it contains all line segments between its members: λx + (1 − λ)y ∈ C
whenever x, y ∈ C and 0 ≤ λ ≤ 1.
Given a subset C of X , int C is the interior of C and C is the norm closure of C . For a set
D ⊆ X ∗ , D
is the weak∗ closure of D . The indicator function of C , written as ιC , is defined at
w*
(cid:40)
x ∈ X by
if x ∈ C ;
0,
+∞,
otherwise.
The support function of C , written as σC , is defined by σC (x∗ ) := supc∈C (cid:104)c, x∗ (cid:105). There is also a
naturally associated (metric) distance function, that is,
dC (x) := inf {(cid:107)x − y(cid:107) y ∈ C } .
ιC (x) :=
(1)
(2)
Distance functions play a central role in convex analysis, both theoretically and algorithmically.
Let f : X → ]−∞, +∞] be a function. Then dom f := f −1 (R) is the domain of f , and the lower
level sets of a function f : X → ]−∞, +∞] are the sets {x ∈ X f (x) ≤ α} where α ∈ R. The
epigraph of f is epi f := {(x, r) ∈ X × R f (x) ≤ r}. We will denote the set of points of continuity
of f by cont f . The function f is said to be convex if for any x, y ∈ dom f and any λ ∈ [0, 1], one
has
f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y).
We say f is proper if dom f (cid:54)= ∅. Let f be proper. The subdifferential of f is defined by
∂ f : X ⇒ X ∗ : x (cid:55)→ {x∗ ∈ X ∗ (cid:104)x∗ , y − x(cid:105) ≤ f (y) − f (x), for all y ∈ X }.
By the definition of ∂ f , even when x ∈ dom f , it is possible that ∂ f (x) may be empty. For example
∂ f (0) = ∅ for f (x) := −√
x whenever x ≥ 0 and f (x) := +∞ otherwise. If x∗ ∈ ∂ f (x) then x∗ is
said to be a subgradient of f at x. An important example of a subdifferential is the normal cone
to a convex set C ⊆ X at a point x ∈ C which is defined as NC (x) := ∂ ιC (x).
(cid:8)f (y) + g(x − y)(cid:9).
Let g : X → ]−∞, +∞]. Then the inf-convolution f (cid:3)g is the function defined on X by
f (cid:3)g : x (cid:55)→ inf
y∈X
(In [45] Moreau studied inf-convolution when X is an arbitrary commutative semigroup.) Notice
that, if both f and g are convex, so it is f (cid:3)g (see, e.g., [49, p. 17]).
We use the convention that (+∞) + (−∞) = +∞ and (+∞) − (+∞) = +∞. We will say a
function f : X → ]−∞, +∞] is Lipschitz on a subset D of dom f if there is a constant M ≥ 0 so
that f (x) − f (y) ≤ M (cid:107)x − y(cid:107) for all x, y ∈ D . In this case M is said to be a Lipschitz constant
for f on D . If for each x0 ∈ D , there is an open set U ⊆ D with x0 ∈ U and a constant M so that
2
f (x) − f (y) ≤ M (cid:107)x − y(cid:107) for all x, y ∈ U , we will say f is local ly Lipschitz on D . If D is the entire
space, we simply say f is Lipschitz or locally Lipschitz respectively.
Consider a function f : X → ]−∞, +∞]; we say f is lower-semicontinuous (lsc) if
lim inf x→ ¯x f (x) ≥ f ( ¯x) for all ¯x ∈ X , or equivalently, if epi f is closed. The function f is said
to be sequential ly weakly lower semi-continuous if for every ¯x ∈ X and every sequence (xn )n∈N
which is weakly convergent to ¯x, one has lim inf n→∞ f (xn ) ≥ f ( ¯x). This is a useful distinction
since there are infinite dimensional Banach spaces (Schur spaces such as (cid:96)1 ) in which weak and
norm convergence coincide for sequences, see [22, p. 384, esp. Thm 8.2.5].
1.2 Structure of this paper
The remainder of this paper is organized as follows. In Section 2, we describe results about Fenchel
conjugates and the subdifferential operator, such as Fenchel duality, the Sandwich theorem, etc.
We also look at some interesting convex functions and inequalities. In Section 3, we discuss the
Chebyshev problem from abstract approximation. In Section 4, we show applications of convex
analysis in Monotone Operator Theory. We reprise such results as the Minty surjectivity theorem,
and present a new proof of the sum theorem in reflexive spaces. We also discuss Fitzpatrick’s
problem on so called autoconjugate representers for maximally monotone operators. In Section 5
we discuss various other applications.
2 Subdifferential operators, conjugate functions & Fenchel duality
We begin with some fundamental properties of convex sets and convex functions. While many
results hold in all locally convex spaces, some of the most important such as (iv)(b) in the next
Fact do not.
Fact 2.1 (Basic properties [22, Ch. 2 and 4].) The fol lowing hold.
(i) The (lsc) convex functions form a convex cone closed under pointwise suprema: if fγ is convex
(and lsc) for each γ ∈ Γ then so is x (cid:55)→ supγ∈Γ fγ (x).
(ii) A function f is convex if and only if epi f is convex if and only if ιepi f is convex.
(iii) Global minima and local minima in the domain coincide for proper convex functions.
(iv) Let f be a proper convex function and let x ∈ dom f . (a) f is local ly Lipschitz at x if and
only f is continuous at x if and only if f is local ly bounded at x. (b) Additional ly, if f is
lower semicontinuous, then f is continuous at every point in int dom f .
(v) A proper lower semicontinuous and convex function is bounded from below by a continuous
affine function.
(vi) If C is a nonempty set, then dC (·) is non-expansive (i.e., is a Lipschitz function with constant
one). Additional ly, if C is convex, then dC (·) is a convex function.
(vii) If C is a convex set, then C is weakly closed if and only if it is norm closed.
3
(viii) Three-slope inequality: Suppose f : R →] − ∞, ∞] is convex and a < b < c. Then
f (b) − f (a)
≤ f (c) − f (a)
≤ f (c) − f (b)
c − b
c − a
b − a
.
The following trivial fact shows the fundamental significance of subgradients in optimization.
Proposition 2.2 (Subdifferential at optimality) Let f : X → ]−∞, +∞] be a proper convex
function. Then the point ¯x ∈ dom f is a (global) minimizer of f if and only if 0 ∈ ∂ f ( ¯x).
The directional derivative of f at ¯x ∈ dom f in the direction d is defined by
f ( ¯x + td) − f ( ¯x)
t
f (cid:48) ( ¯x; d) := lim
t→0+
if the limit exists. If f is convex, the directional derivative is everywhere finite at any point of
int dom f , and it turns out to be Lipschitz at cont f . We use the term directional derivative with
the understanding that it is actually a one-sided directional derivative.
If the directional derivative f (cid:48) ( ¯x, d) exists for all directions d and the operator f (cid:48) ( ¯x) defined
by (cid:104)f (cid:48) ( ¯x), · (cid:105) := f (cid:48) ( ¯x; · ) is linear and bounded, then we say that f is Gateaux differentiable at ¯x,
and f (cid:48) ( ¯x) is called the Gateaux derivative. Every function f : X → ]−∞, +∞] which is lower
semicontinuous, convex and Gateaux differentiable at x, it is continuous at x. Additionally, the
following properties are relevant for the existence and uniqueness of the subgradients.
Proposition 2.3 (See [22, Fact 4.2.4 and Corollary 4.2.5].) Suppose f : X → ]−∞, +∞] is convex.
(i) If f is Gateaux differentiable at ¯x, then f (cid:48) ( ¯x) ∈ ∂ f ( ¯x).
(ii) If f is continuous at ¯x, then f is Gateaux differentiable at ¯x if and only if ∂ f ( ¯x) is a singleton.
Example 2.4 We show that part (ii) in Proposition 2.3 is not always true in infinite dimensions
without continuity hypotheses.
(a) The indicator of the Hilbert cube C := {x = (x1 , x2 , . . .) ∈ (cid:96)2 : xn ≤ 1/n, ∀n ∈ N} at zero or
(b) Boltzmann-Shannon entropy x (cid:55)→ (cid:82) 1
any other non-support point has a unique subgradient but is nowhere Gateaux differentiable.
0 x(t) log(x(t))dt viewed as a lower semicontinuous and
convex function on L1 [0, 1] has unique subgradients at x(t) > 0 a.e. but is nowhere Gateaux
differentiable (which for a lower semicontinuous and convex function in Banach space implies
continuity).
That Gateaux differentiability of a convex and lower semicontinuous function implies continuity at
♦
the point is a consequence of the Baire category theorem.
The next result proved by Moreau in 1963 establishes the relationship between subgradients and
directional derivatives, see also [49, page 65]. Proofs can be also found in most of the books in
variational analysis, see e.g. [25, Theorem 4.2.7].
4
Theorem 2.5 (Moreau’s max formula [46]) Let f : X → ]−∞, +∞] be a convex function and
let d ∈ X . Suppose that f is continuous at ¯x. Then, ∂ f ( ¯x) (cid:54)= ∅ and
f (cid:48) ( ¯x; d) = max{(cid:104)x∗ , d(cid:105) x∗ ∈ ∂ f ( ¯x)}.
(3)
Let f : X → [−∞, +∞]. The Fenchel conjugate (also called the Legendre-Fenchel conjugate1 or
transform) of f is the function f ∗ : X ∗ → [−∞, +∞] defined by
{(cid:104)x∗ , x(cid:105) − f (x)}.
f ∗ (x∗ ) := sup
x∈X
We can also consider the conjugate of f ∗ called the biconjugate of f and denoted by f ∗∗ . This is a
convex function on X ∗∗ satisfying f ∗∗ X ≤ f . A useful and instructive example is σC = ι∗
C .
(cid:107)x∗ (cid:107)q∗
(cid:107)x(cid:107)p
Example 2.6 Let 1 < p < ∞ . If f (x) :=
for x ∈ X then f ∗ (x∗ ) =
(cid:27)
(cid:26)
(cid:27)
(cid:26)
Indeed, for any x∗ ∈ X ∗ , one has
p
q
(cid:104)x∗ , λx(cid:105) − (cid:107)λx(cid:107)p
f ∗ (x∗ ) = sup
λ∈R+
p
λ(cid:107)x∗(cid:107)∗ − λp
p
p + 1
, where 1
q = 1.
= sup
λ∈R+
sup
(cid:107)x(cid:107)=1
(cid:107)x∗(cid:107)q∗
q
.
=
♦
By direct construction and Fact 2.1 (i), for any function f , the conjugate function f ∗ is always
convex and lower semicontinuous, and if the domain of f is nonempty, then f ∗ never takes the value
−∞. The conjugate plays a role in convex analysis in many ways analogous to the role played by
the Fourier transform in harmonic analysis with infimal convolution, see below, replacing integral
convolution and sum replacing product [22, Chapter 2.].
2.1 Inequalities and their applications
An immediate consequence of the definition is that for f , g : X → [−∞, +∞], the inequality f ≥ g
implies f ∗ ≤ g∗ . An important result which is straightforward to prove is the following.
Proposition 2.7 (Fenchel–Young) Let f : X → ]−∞, +∞]. Al l points x∗ ∈ X ∗ and x ∈ dom f
satisfy the inequality
f (x) + f ∗ (x∗ ) ≥ (cid:104)x∗ , x(cid:105).
(4)
Equality holds if and only if x∗ ∈ ∂ f (x).
Example 2.8 (Young’s inequality) By taking f as in Example 2.6, one obtains directly from
Proposition 2.7
(cid:107)x(cid:107)p
(cid:107)x∗(cid:107)q∗
≥ (cid:104)x∗ , x(cid:105),
+
q
p
for all x ∈ X and x∗ ∈ X ∗ , where p > 1 and 1
q = 1. When X = R one recovers the original
p + 1
♦
Young inequality.
1Originally the connection was made between a monotone function on an interval and its inverse. The convex
functions then arise by integration.
5
This in turn leads to one of the workhorses of modern analysis:
Example 2.9 (Holder’s inequality) Let f and g be measurable on a measure space (X, µ).
(cid:90)
Then
X
f g dµ ≤ (cid:107)f (cid:107)p(cid:107)g(cid:107)q ,
(5)
where 1 < p < ∞ and 1
p + 1
q = 1. Indeed, by rescaling, we may assume without loss of generality
that (cid:107)f (cid:107)p = (cid:107)g(cid:107)q = 1. Then Young’s inequality in Example 2.8 yields
g(x)q
f (x)g(x) ≤ f (x)p
p
q
and (5) follows by integrating both sides. The result holds true in the limit for p = 1 or p = ∞. ♦
for x ∈ X,
+
We next take a brief excursion into special function theory and normed space geometry to
emphasize that “convex functions are everywhere.”
(cid:90) ∞
Example 2.10 (Bohr–Mollerup theorem) The Gamma function defined for x > 0 as
0
n! nx
x(x + 1) · · · (x + n)
e−t tx−1dt = lim
n→∞
Γ(x) :=
is the unique function f mapping the positive half-line to itself and such that (a) f (1) = 1, (b)
xf (x) = f (x + 1) and (c) log f is a convex function.
Indeed, clearly Γ(1) = 1, and it is easy to prove (b) for Γ by using integration by parts. In order
to show that log Γ is convex, pick any x, y > 0 and λ ∈ (0, 1) and apply Holder’s inequality (5)
with p = 1/λ to the functions t (cid:55)→ e−λt tλ(x−1) and t (cid:55)→ e−(1−λ)t t(1−λ)(y−1) . For the converse,
let g := log f . Then (a) and (b) imply g(n + 1 + x) = log [x(1 + x) . . . (n + x)f (x)] and thus
g(n + 1) = log(n!). Convexity of g together with the three-slope inequality, see Fact 2.1(viii),
implies that
g(n + 1) − g(n) ≤ g(n + 1 + x) − g(n + 1)
x
≤ g(n + 2 + x) − g(n + 1 + x),
and hence,
whence,
x log(n) ≤ log (x(x + 1) · · · (x + n)f (x)) − log(n!) ≤ x log(n + 1 + x);
(cid:19)
(cid:18)
(cid:19)
(cid:18)
n! nx
0 ≤ g(x) − log
x(x + 1) · · · (x + n)
Taking limits when n → ∞ we obtain
1 + x
n
≤ x log
1 +
.
f (x) = lim
n→∞
n! nx
x(x + 1) · · · (x + n)
= Γ(x).
6
As a bonus we recover a classical and important limit formula for Γ(x).
Application of the Bohr–Mollerup theorem is often automatable in a computer algebra system,
(cid:90) 1
as we now illustrate. Consider the beta function
0
tx−1 (1 − t)y−1 d t
β (x, y)
(6)
:=
for Re(x), Re(y) > 0. As is often established using polar coordinates and double integrals
(7)
β (x, y) =
Γ(x) Γ(y)
Γ(x + y)
.
We may use the Bohr–Mollerup theorem with
f := x → β (x, y) Γ(x + y)/Γ(y)
to prove (7) for real x, y .
Now (a) and (b) from Example 2.10 are easy to verify. For (c) we again use Holder’s inequality
♦
to show f is log-convex. Thus, f = Γ as required.
Example 2.11 (Blaschke–Santal´o theorem) The volume of a unit ball in the (cid:107) · (cid:107)p -norm, Vn (p)
is
(8)
Vn (p) = 2n
Γ(1 + 1
p )n
Γ(1 + n
p )
.
as was first determined by Dirichlet. When p = 2, this gives
Vn = 2n Γ( 3
2 )n
Γ(1 + n
2 )
=
Γ( 1
2 )n
Γ(1 + n
2 )
,
which is more concise than that usually recorded in texts.
Let C in Rn be a convex body which is symmetric around zero, that is, a closed bounded convex
set with nonempty interior. Denoting n-dimensional Euclidean volume of S ⊆ Rn by Vn (S ), the
Blaschke–Santal´o inequality says
Vn (C ) Vn (C ◦ ) ≤ Vn (E ) Vn (E ◦ ) = V 2
n (Bn (2))
(9)
where maximality holds (only) for any symmetric ellipsoid E and Bn (2) is the Euclidean unit ball.
It is conjectured the minimum is attained by the 1-ball and the ∞-ball. Here as always the polar
set is defined by C ◦ := {y ∈ Rn : (cid:104)y , x(cid:105) ≤ 1 for all x ∈ C }.
The p-ball case of (9) follows by proving the following convexity result:
(cid:18)
(cid:19)α
(cid:18)
(cid:19)
Theorem 2.12 (Harmonic-arithmetic log-concavity) The function
1 +
Vα (p) := 2αΓ
1 +
/Γ
α
p
1
p
7
satisfies
(10)
Vα (p)λ Vα (q)1−λ < Vα
(cid:32)
(cid:33)
,
1
p + 1−λ
λ
q
for all α > 1, if p, q > 1, p (cid:54)= q , and λ ∈ (0, 1).
q = 1 with λ = 1 − λ = 1/2 to recover the p−norm case of the Blaschke–Santal´o
Set α := n, 1
p + 1
inequality.
It is amusing to deduce the corresponding lower bound. This technique extends to
various substitution norms. Further details may be found in [16, §5.5]. Note that we may easily
♦
explore Vα (p) graphically.
2.2 The biconjugate and duality
The next result has been associated by different authors with the names of Legendre, Fenchel,
Moreau and Hormander; see, e.g., [22, Proposition 4.4.2].
Proposition 2.13 (Hormander2 )(See [66, Theorem 2.3.3] or [22, Proposition 4.4.2(a)].) Let
f : X → ]−∞, +∞] be a proper function. Then
f is convex and lower semicontinuous ⇔ f = f ∗∗ X .
i x
f (x, s) :=
Example 2.14 (Establishing convexity) (See [12, Theorem 1].) We may compute conjugates
by hand or using the software SCAT [20]. This is discussed further in Section 5.3. Consider
f (x) := ex . Then f ∗ (x) = x log(x) − x for x ≥ 0 (taken to be zero at zero) and is infinite for x < 0.
This establishes the convexity of x log(x) − x in a way that takes no knowledge of x log(x).
Ex. 13] which can be computed algorithmically: Given real α1 , α2 , . . . , αm > 0, define α := (cid:80)
A more challenging case is the following (slightly corrected) conjugation formula [21, p. 94,
i αi
µ−1sµ (cid:81)
and suppose a real µ satisfies µ > α + 1. Now define a function f : Rm × R (cid:55)→ ]−∞, +∞] by
−αi
if x ∈ Rm
++ , s ∈ R+ ;
i
0
+∞
ρν −1 tν (cid:81)
It transpires that
f ∗ (y , t) =
0
+∞
µ
µ − (α + 1)
ν :=
∀y := (yn )m
n=1 ∈ Rm , t ∈ R.
(cid:17)βi
(cid:16) αi
(cid:89)
µ
i
if y ∈ Rm−− , t ∈ R+
if y ∈ Rm− , t ∈ R−
otherwise
if ∃xi = 0, x ∈ Rm
+ , s = 0;
otherwise.
,
∀x := (xn )m
n=1 ∈ Rm , s ∈ R.
i (−yi )−βi
, βi :=
αi
µ − (α + 1)
, ρ :=
.
for constants
,
2Hormander first proved the case of support and indicator functions in [38] which led to discovery of general result.
8
We deduce that f = f ∗∗ , whence f (and f ∗ ) is (essentially strictly) convex. For attractive alterna-
♦
tive proof of convexity see [42]. Many other substantive examples are to be found in [21, 22].
The next theorem gives us a remarkable sufficient condition for convexity of functions in terms
of the Gateaux differentiability of the conjugate. There is a simpler analogue for the Fr´echet
derivative.
Theorem 2.15 (See [22, Corollary 4.5.2].) Suppose f : X → ]−∞, +∞] is such that f ∗∗ is
If f ∗ is Gateaux differentiable at al l x∗ ∈ dom ∂ f ∗ and f is sequential ly weakly lower
proper.
semicontinuous, then f is convex.
Let f : X → ]−∞, +∞]. We say f is coercive if lim(cid:107)x(cid:107)→∞ f (x) = +∞. We say f is supercoercive
if lim(cid:107)x(cid:107)→∞ f (x)(cid:107)x(cid:107) = +∞.
Fact 2.16 (See [22, Fact 4.4.8].) If f is proper convex and lower semicontinuous at some point in
its domain, then the fol lowing statements are equivalent.
(i) f is coercive.
(ii) There exist α > 0 and β ∈ R such that f ≥ α(cid:107) · (cid:107) + β .
(iii) lim inf (cid:107)x(cid:107)→∞ f (x)/(cid:107)x(cid:107) > 0.
(iv) f has bounded lower level sets.
Because a convex function is continuous at a point if and only if it is bounded above on a
neighborhood of that point (Fact 2.1(iv)), we get the following result; see also [38, Theorem 7] for
the case of the indicator function of a bounded convex set.
Theorem 2.17 (Hormander–Moreau–Rockafellar) Let f : X → ]−∞, +∞] be convex and
lower semicontinuous at some point in its domain, and let x∗ ∈ X ∗ . Then f − x∗ is coercive if and
only if f ∗ is continuous at x∗ .
Proof. “⇒”: By Fact 2.16, there exist α > 0 and β ∈ R such that f ≥ x∗ + α(cid:107) · (cid:107) + β . Then
f ∗ ≤ −β + ι{x∗+αBX ∗ } , from where x∗ + αBX ∗ ⊆ dom f ∗ . Therefore, f ∗ is continuous at x∗ by
Fact 2.1(iv).
“⇐”: By the assumption, there exists β ∈ R and δ > 0 such that
f ∗ (x∗ + z ∗ ) ≤ β ,
∀z ∗ ∈ δBX ∗ .
Thus, by Proposition 2.7,
(cid:104)x∗ + z ∗ , y(cid:105) − f (y) ≤ β ,
whence, taking the supremum with z ∗ ∈ δBX ∗ ,
δ(cid:107)y(cid:107) − β ≤ f (y) − (cid:104)x∗ , y(cid:105),
Then, by Fact 2.16, f − x∗ is coercive.
∀z ∗ ∈ δBX ∗ , ∀y ∈ X ;
∀y ∈ X.
(cid:4)
9
Example 2.18 Given a set C in X , recall that the negative polar cone of C is the convex cone
C − := {x∗ ∈ X ∗ sup(cid:104)x∗ , C (cid:105) ≤ 0}.
Suppose that X is reflexive and let K ⊆ X be a closed convex cone. Then K − is another nonempty
closed convex cone with K −− := (K − )− = K . Moreover, the indicator function of K and K − are
conjugate to each other. If we set f := ιK− , the indicator function of the negative polar cone of
K , Theorem 2.17 applies to get that
x ∈ int K if and only if the set {x∗ ∈ K − (cid:104)x∗ , x(cid:105) ≥ α} is bounded for any α ∈ R.
Indeed, since x ∈ int K = int dom ι∗
K− if and only if ι∗
K− is continuous at x, from Theorem 2.17 we
have that this is true if and only if the function ιK− − x is coercive. Now, Fact 2.16 assures us that
♦
coerciveness is equivalent to boundedness of the lower level sets, which implies the assertion.
Theorem 2.19 (Moreau–Rockafellar duality [47]) Let f : X → (−∞, +∞] be a lower semi-
continuous convex function. Then f is continuous at 0 if and only if f ∗ has weak∗ -compact lower
level sets.
Proof. Observe that f is continuous at 0 if and only if f ∗∗ is continuous at 0 ([22, Fact 4.4.4(b)])if
and only if f ∗ is coercive (Theorem 2.17) if and only if f ∗ has bounded lower level sets (Fact 2.16)
if and only if f ∗ has weak∗ -compact lower level sets by the Banach-Alaoglu theorem (see [59,
(cid:4)
Theorem 3.15]).
Theorem 2.20 (Conjugates of supercoercive functions) Suppose f : X → ]−∞, +∞] is a
lower semicontinuous and proper convex function. Then
(a) f is supercoercive if and only if f ∗ is bounded (above) on bounded sets.
(b) f is bounded (above) on bounded sets if and only if f ∗ is supercoercive.
Proof. (a) “⇒”: Given any α > 0, there exists M such that f (x) ≥ α(cid:107)x(cid:107) if (cid:107)x(cid:107) ≥ M . Now there
exists β ≥ 0 such that f (x) ≥ −β if (cid:107)x(cid:107) ≤ M by Fact 2.1(v). Therefore f ≥ α(cid:107) · (cid:107) + (−αM − β ).
Thus, it implies that f ∗ ≤ α((cid:107) · (cid:107))∗ ( ·
α ) + αM + β and hence f ∗ ≤ αM + β on αBX ∗ .
“⇐”: Let γ > 0. Now there exists K such that f ∗ ≤ K on γBX ∗ . Then f ≥ γ (cid:107) · (cid:107) − K and so
lim inf (cid:107)x(cid:107)→∞ f (x)(cid:107)x(cid:107) ≥ γ . Hence lim inf (cid:107)x(cid:107)→∞ f (x)(cid:107)x(cid:107) = +∞.
(b): According to (a), f ∗ is supercoercive if and only if f ∗∗ is bounded on bounded sets. By [22,
(cid:4)
Fact 4.4.4(a)] this holds if and only if f is bounded (above) on bounded sets.
We finish this subsection by recalling some properties of infimal convolutions. Some of their
many applications include smoothing techniques and approximation. We shall meet them again
in Section 4. Let f , g : X → ]−∞, +∞]. Geometrically, the infimal convolution of f and g is
the largest extended real-valued function whose epigraph contains the sum of epigraphs of f and
g (see example in Figure 1), consequently it is a convex function. The following is a useful result
concerning the conjugate of the infimal convolution.
10
Fact 2.21 (See [22, Lemma 4.4.15] and [49, pp. 37-38].) If f and g are proper functions on X ,
then (f (cid:3)g)∗ = f ∗ + g∗ . Additional ly, suppose f , g are convex and bounded below. If f : X → R
is continuous (resp. bounded on bounded sets, Lipschitz), then f (cid:3)g is a convex function that is
continuous (resp. bounded on bounded sets, Lipschitz).
Remark 2.22 Suppose C is a nonempty convex set. Then dC = (cid:107) · (cid:107)(cid:3)ιC , implying that dC is a
♦
Lipschitz convex function.
(cid:26) −√
Example 2.23 Consider f , g : R → ]−∞, +∞] given by
for − 1 ≤ x ≤ 1,
1 − x2 ,
+∞
otherwise,
(cid:40) −√
The infimal convolution of f and g is
1 − x2 , − √
2 ≤ x ≤ − √
x − √
2
2
2 ;
2,
otherwise.
and g(x) := x.
f (x) :=
(f (cid:3)g)(x) =
,
as shown in Figure 1.
♦
Figure 1: Infimal convolution of f (x) = −√
1 − x2 and g(x) = x.
2.3 The Hahn-Banach circle
Let T : X → Y be a linear mapping between two Banach spaces X and Y . The adjoint of T is the
linear mapping T ∗ : Y ∗ → X ∗ defined, for y∗ ∈ Y ∗ , by
for all x ∈ X.
(cid:104)T ∗y∗ , x(cid:105) = (cid:104)y∗ , T x(cid:105)
11
-1.5-1-0.50.511.5-1-0.50.511.5gffg•(f+g)epi A flexible modern version of Fenchel’s celebrated duality theorem is:
Theorem 2.24 (Fenchel duality) Let Y be another Banach space, let f : X → ]−∞, +∞] and
g : Y → ]−∞, +∞] be convex functions and let T : X → Y be a bounded linear operator. Define
the primal and dual values p, d ∈ [−∞, +∞] by solving the Fenchel problems
{f (x) + g(T x)}
{−f ∗ (T ∗y∗ ) − g∗ (−y∗ )}.
p := inf
x∈X
d := sup
y∗∈Y ∗
Then these values satisfy the weak duality inequality p ≥ d.
(cid:91)
Suppose further that f , g and T satisfy either
λ [dom g − T dom f ] = Y and both f and g are lower semicontinuous,
λ>0
(11)
(12)
or the condition
(13)
cont g ∩ T dom f (cid:54)= ∅.
Then p = d, and the supremum in the dual problem (11) is attained when finite. Moreover, the
perturbation function h(u) := inf x f (x) + g(T x + u) is convex and continuous at zero.
Generalizations of Fenchel duality Theorem can be found in [27, 26]. An easy consequence is:
Corollary 2.25 (Infimal convolution) Under the hypotheses of the Fenchel duality theorem 2.24
(f + g)∗ (x∗ ) = (f ∗(cid:3)g∗ )(x∗ ) with attainment when finite.
Another nice consequence of Fenchel duality is the ability to obtain primal solutions from dual
ones, as we now record.
Corollary 2.26 Suppose the conditions for equality in the Fenchel duality Theorem 2.24 hold, and
that ¯y∗ ∈ Y ∗ is an optimal dual solution. Then the point ¯x ∈ X is optimal for the primal problem
if and only if it satisfies the two conditions T ∗ ¯y∗ ∈ ∂ f ( ¯x) and − ¯y∗ ∈ ∂ g(T ¯x).
The regularity conditions in Fenchel duality theorem can be weakened when each function is
polyhedral, i.e., when their epigraph is polyhedral.
Theorem 2.27 (Polyhedral Fenchel duality) (See [21, Corollary 5.1.9].) Suppose that X is a
finite-dimensional space. The conclusions of the Fenchel duality Theorem 2.24 remain valid if the
regularity condition (12) is replaced by the assumption that the functions f and g are polyhedral
with
dom g ∩ T dom f (cid:54)= ∅.
Fenchel duality applied to a linear programming program yields the well-known Lagrangian
duality.
12
(14)
Corollary 2.28 (Linear programming duality) Given c ∈ Rn , b ∈ Rm and A an m × n real
matrix, one has
{cT x Ax ≤ b} ≥ sup
{−bT λ AT λ = −c},
+ := (cid:8)(x1 , x2 , · · · , xm ) xi ≥ 0, i = 1, 2, · · · , m(cid:9). Equality in (14) holds if b ∈ ran A + Rm
inf
x∈Rn
λ∈Rm
+
where Rm
+ .
Moreover, both extrema are obtained when finite.
Proof. Take f (x) := cT x, T := A and g(y) := ιb≥ (y) where b≥ := {y ∈ Rm y ≤ b}. Then apply
(cid:26) bT λ,
the polyhedral Fenchel duality Theorem 2.27 observing that f ∗ = ι{c} , and for any λ ∈ Rm ,
if λ ∈ Rm
g∗ (λ) = sup
+ ;
+∞, otherwise;
y≤b
and (14) follows, since dom g ∩ A dom f = {Ax ∈ Rm Ax ≤ b}.
(cid:4)
One can easily derive various relevant results from Fenchel duality, such as the Sandwich theorem,
the subdifferential sum rule, and the Hahn-Banach extension theorem, among many others.
yT λ =
Theorem 2.29 (Extended sandwich theorem) Let X and Y be Banach spaces and let T :
X → Y be a bounded linear mapping. Suppose that f : X → ]−∞, +∞], g : Y → ]−∞, +∞] are
proper convex functions which together with T satisfy either (12) or (13). Assume that f ≥ −g ◦ T .
Then there is an affine function α : X → R of the form α(x) = (cid:104)T ∗y∗ , x(cid:105) + r satisfying f ≥ α ≥
−g ◦ T . Moreover, for any ¯x satisfying f ( ¯x) = (−g ◦ T )( ¯x), we have −y∗ ∈ ∂ g(T ¯x).
Proof. With notation as in the Fenchel duality Theorem 2.24, we know d = p, and since p ≥ 0
because f (x) ≥ −g(T x), the supremum in d is attained. Therefore there exists y∗ ∈ Y ∗ such that
0 ≤ p = d = −f ∗ (T ∗y∗ ) − g∗ (−y∗ ).
Then, by Fenchel-Young inequality (4), we obtain
0 ≤ p ≤ f (x) − (cid:104)T ∗y∗ , x(cid:105) + g(y) + (cid:104)y∗ , y(cid:105),
(15)
for any x ∈ X and y ∈ Y . For any z ∈ X , setting y = T z in the previous inequality, we obtain
[−g(T z ) − (cid:104)T ∗y∗ , z (cid:105)] ≤ b := inf
[f (x) − (cid:104)T ∗y∗ , x(cid:105)]
a := sup
x∈X
z∈X
Now choose r ∈ [a, b]. The affine function α(x) := (cid:104)T ∗y∗ , x(cid:105) + r satisfies f ≥ α ≥ −g ◦ T , as claimed.
The last assertion follows from (15) simply by setting x = ¯x, where ¯x satisfies f ( ¯x) = (−g ◦ T )( ¯x).
Then we have supy∈Y {(cid:104)−y∗ , y(cid:105)−g(y)} ≤ (−g ◦T )( ¯x)− (cid:104)T ∗y∗ , ¯x(cid:105). Thus g∗ (−y∗ )+g(T ¯x) ≤ −(cid:104)y∗ , T ¯x(cid:105)
and hence −y∗ ∈ ∂ g(T ¯x).
(cid:4)
When X = Y and T is the identity we recover the classical Sandwich theorem. The next example
shows that without a constraint qualification, the sandwich theorem may fail.
(cid:26) −√
(cid:26) −√−x,
Example 2.30 Consider f , g : R → ]−∞, +∞] given by
for x ≥ 0,
for x ≤ 0,
x,
+∞ otherwise.
+∞
otherwise,
and g(x) :=
f (x) :=
13
In this case, (cid:83)
λ>0 λ [dom g − dom f ] = [0, +∞[ (cid:54)= R and it is not difficult to prove there is not any
affine function which separates f and −g , see Figure 2.
♦
The prior constraint qualifications are sufficient but not necessary for the sandwich theorem as
we illustrate in the next example.
(cid:26) 1
(cid:26) − 1
Example 2.31 Let f , g : R → ]−∞, +∞] be given by
for x < 0,
for x > 0,
x ,
x ,
Despite that (cid:83)
+∞ otherwise.
+∞ otherwise,
λ>0 λ [dom g − dom f ] = ]−∞, 0[ (cid:54)= R, the affine function α(x) := −x satisfies
f ≥ α ≥ −g , see Figure 2.
♦
and g(x) :=
f (x) :=
Figure 2: On the left we show the failure of the sandwich theorem in the absence of the constraint
qualification; of the right we show that the constraint qualification is not necessary.
Theorem 2.32 (Subdifferential sum rule) Let X and Y be Banach spaces, and let f : X →
]−∞, +∞] and g : Y → ]−∞, +∞] be convex functions and let T : X → Y be a bounded linear
mapping. Then at any point x ∈ X we have the sum rule
∂ (f + g ◦ T )(x) ⊇ ∂ f (x) + T ∗ (∂ g(T x))
with equality if (12) or (13) hold.
Proof. The inclusion is straightforward by using the definition of the subdifferential, so we prove
the reverse inclusion. Fix any x ∈ X and let x∗ ∈ ∂ (f + g ◦ T )(x). Then 0 ∈ ∂ (f − (cid:104)x∗ , · (cid:105) + g ◦ T )(x).
14
f−gf−g Conditions for the equality in Theorem 2.24 are satisfied for the functions f (·)− (cid:104)x∗ , · (cid:105) and g . Thus,
there exists y∗ ∈ Y ∗ such that
f (x) − (cid:104)x∗ , x(cid:105) + g(T x) = −f ∗ (T ∗y∗ + x∗ ) − g∗ (−y∗ ).
Now set z ∗ := T ∗y∗ + x∗ . Hence, by the Fenchel-Young inequality (4), one has
0 ≤ f (x) + f ∗ (z ∗ ) − (cid:104)z ∗ , x(cid:105) = −g(T x) − g∗ (−y∗ ) − (cid:104)T ∗y∗ , x(cid:105) ≤ 0;
whence,
f (x) + f ∗ (z ∗ ) = (cid:104)z ∗ , x(cid:105)
g(T x) + g∗ (−y∗ ) = (cid:104)−y∗ , T x(cid:105).
Therefore equality in Fenchel-Young occurs, and one has z ∗ ∈ ∂ f (x) and −y∗ ∈ ∂ g(T x), which
(cid:4)
completes the proof.
The subdifferential sum rule for two convex functions with a finite common point where one of
them is continuous was proved by Rockafellar in 1966 with an argumentation based on Fenchel
duality, see [55, Th. 3]. In an earlier work in 1963, Moreau [46] proved the subdifferential sum rule
for a pair of convex and lsc functions, in the case that infimal convolution of the conjugate functions
is achieved, see [49, p. 63] for more details. Moreau actually proved this result for functions which
are the supremum of a family of affine continuous linear functions, a set which agrees with the
convex and lsc functions when X is a locally convex vector space, see [44] or [49, p. 28]. See
also [36, 37, 27, 19] for more information about the subdifferential calculus rule.
Theorem 2.33 (Hahn–Banach extension) Let X be a Banach space and let f : X → R be a
continuous sublinear function with dom f = X . Suppose that L is a linear subspace of X and the
function h : L → R is linear and dominated by f , that is, f ≥ h on L. Then there exists x∗ ∈ X ∗ ,
dominated by f , such that
h(x) = (cid:104)x∗ , x(cid:105), for al l x ∈ L.
Proof. Take g := −h + ιL and apply Theorem 2.24 to f and g with T the identity mapping. Then,
there exists x∗ ∈ X ∗ such that
{f (x) − h(x) + ιL (x)}
0 ≤ inf
x∈X
= −f ∗ (x∗ ) − sup
{(cid:104)−x∗ , x(cid:105) + h(x) − ιL (x)}
x∈X
= −f ∗ (x∗ ) + inf
{(cid:104)x∗ , x(cid:105) − h(x)};
x∈L
(16)
whence,
f ∗ (x∗ ) ≤ (cid:104)x∗ , x(cid:105) − h(x),
for all x ∈ L.
Observe that f ∗ (x∗ ) ≥ 0 since f (0) = 0. Thus, being L a linear subspace, we deduce from the
above inequality that
h(x) = (cid:104)x∗ , x(cid:105),
for all x ∈ L.
15
Then (16) implies f ∗ (x∗ ) = 0, from where
f (x) ≥ (cid:104)x∗ , x(cid:105),
for all x ∈ X,
and we are done.
(cid:4)
Remark 2.34 (Moreau’s max formula, Theorem 2.5)—a true child of Cauchy’s principle of steep-
est descent—can be also derived from Fenchel duality. In fact, the non-emptiness of the subdiffer-
ential at a point of continuity, Moreau’s max formula, Fenchel duality, the Sandwich theorem, the
(cid:0)f (x) + g(Ax + u)(cid:1) and checks that ∂h(0) (cid:54)= ∅ implies the
subdifferential sum rule, and Hahn-Banach extension theorem are all equivalent, in the sense that
they are easily inter-derivable.
In outline, one considers h(u) := inf x
Fenchel and Lagrangian duality results; while condition (12) or (13) implies h is continuous at zero
and thus Theorem 2.5 finishes the proof. Likewise, the polyhedral calculus [21, §5.1] implies h is
polyhedral when f and g are and shows that polyhedral functions have dom h = dom ∂h. This
establishes Theorem 2.27. This also recovers abstract LP duality (e.g., semidefinite programming
♦
and conic duality) under condition (12). See [21, 22] for more details.
Let us turn to two illustrations of the power of convex analysis within functional analysis.
A Banach limit is a bounded linear functional Λ on the space of bounded sequences of real
numbers (cid:96)∞ such that
(ii) lim inf k xk ≤ Λ(cid:0)(xn )n∈N(cid:1) ≤ lim supk xk
(i) Λ((xn+1 )n∈N ) = Λ((xn )n∈N ) (so it only depends on the sequence’s tail),
where (xn )n∈N = (x1 , x2 , . . .) ∈ (cid:96)∞ and (xn+1 )n∈N = (x2 , x3 , . . .). Thus Λ agrees with the limit on
c, the subspace of sequences whose limit exists. Banach limits care peculiar ob jects!
The Hahn-Banach extension theorem can be used show the existence of Banach limits (see
Sucheston [65] or [22, Exercise 5.4.12]). Many of its earliest applications were to summability
theory and related fields. We sketch Sucheston’s proof as follows.
Theorem 2.35 (Banach limits) (See [65].) Banach limits exist.
(cid:32)
(cid:33)
Proof. Let c be the subspace of convergent sequences in (cid:96)∞ . Define f : (cid:96)∞ → R by
n(cid:88)
i=1
x := (xn )n∈N (cid:55)→ lim
n→∞
sup
j
xi+j
(17)
1
n
.
Then f is sublinear with full domain, since the limit in (17) always exists (see [65, p. 309]). Define
h on c by h := limn xn for every x := (xn )n∈N in c. Hence h is linear and agrees with f on c.
Applying the Hahn-Banach extension Theorem 2.33, there exists Λ ∈ ((cid:96)∞ )∗ , dominated by f , such
that Λ = h on c. Thus Λ extends the limit linearly from c to (cid:96)∞ . Let S denote the forward shift
(cid:33)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ lim
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) lim
(cid:32)
defined as S ((xn )n∈N ) := (xn+1 )n∈N . Note that f (S x − x) = 0, since
(xj+n+1 − xj+1 )
n→∞
n→∞
f (S x − x) =
xj = 0.
sup
j
sup
j
2
n
1
n
16
Thus, Λ(S x) − Λ(x) = Λ(S x − x) ≤ 0, and Λ(x) − Λ(S x) = Λ(x − S x) ≤ f (x − S x) = 0; that is, Λ
(cid:4)
is indeed a Banach limit.
Remark 2.36 One of the referees kindly pointed out that in the proof of Theorem 2.35, the
function h can be simply defined by h : {0} → R with h(0) = 0.
Theorem 2.37 (Principle of uniform boundedness) (See ([22, Example 1.4.8].) Let Y be
another Banach space and Tα : X → Y for α ∈ A be bounded linear operators. Assume that
supα∈A (cid:107)Tα (x)(cid:107) < +∞ for each x in X . Then supα∈A (cid:107)Tα(cid:107) < +∞.
Proof. Define a function fA by
(cid:107)Tα (x)(cid:107)
fA (x) := sup
α∈A
for each x in X . Then, as observed in Fact 2.1(i), fA is convex. It is also lower semicontinuous since
each mapping x (cid:55)→ (cid:107)Tα (x)(cid:107) is continuous. Hence fA is a finite, lower semicontinuous and convex
(actually sublinear) function. Now Fact 2.1(iv) ensures fA is continuous at the origin. Select ε > 0
with sup{fA (x) (cid:107)x(cid:107) ≤ ε} ≤ 1 + fA (0) = 1. It follows that
(cid:107)Tα(cid:107) = sup
(cid:107)Tα (x)(cid:107) =
1
1
α∈A
ε
ε
(cid:107)Tα (x)(cid:107) ≤ 1
ε
sup
α∈A
sup
(cid:107)x(cid:107)≤ε
sup
(cid:107)x(cid:107)≤ε
sup
α∈A
.
Thus, uniform boundedness is revealed to be continuity of fA .
(cid:4)
3 The Chebyshev problem
Let C be a nonempty subset of X . We define the nearest point mapping by
PC (x) := {v ∈ C (cid:107)v − x(cid:107) = dC (x)}.
A set C is said to be a Chebyshev set if PC (x) is a singleton for every x ∈ X . If PC (x) (cid:54)= ∅ for
every x ∈ X , then C is said to be proximal; the term proximinal is also used.
In 1961 Victor Klee [39] posed the following fundamental question: Is every Chebyshev set in a
Hilbert space convex? At this stage, it is known that the answer is affirmative for weakly closed
sets. In what follows we will present a proof of this fact via convex duality. To this end, we will
make use of the following fairly simple lemma.
Lemma 3.1 (See [22, Proposition 4.5.8].) Let C be a weakly closed Chebyshev subset of a Hilbert
space H . Then the nearest point mapping PC is continuous.
Theorem 3.2 Let C be a nonempty weakly closed subset of a Hilbert space H . Then C is convex
if and only if C is a Chebyshev set.
and g(z ) := σC (z ). Notice that (cid:83)
Proof. For the direct implication, we will begin by proving that C is proximal. We can and do
suppose that 0 ∈ C . Pick any x ∈ H . Consider the convex and lsc functions f (z ) := −(cid:104)x, z (cid:105)+ιBH (z )
λ>0 λ [dom g − dom f ] = H (in fact f is continuous at 0 ∈
dom f ∩ dom g). With the notation of Theorem 2.24, one has p = d, and the supremum of the dual
17
problem is attained if finite. Since f ∗ (y) = (cid:107)x + y(cid:107) and g∗ (y) = ιC (y), as C is closed, the dual
problem (11) takes the form
{−(cid:107)x + y(cid:107) − ιC (−y)} = − dC (x).
(18)
Indeed, for x ∈ H ,
d = sup
y∈H
Choose any c ∈ C . Observe that 0 ≤ dC (x) ≤ (cid:107)x − c(cid:107). Therefore the supremum must be attained,
and PC (x) (cid:54)= ∅. Uniqueness follows easily from the convexity of C .
2 (cid:107) · (cid:107)2 + ιC . We first show that
For the converse, consider the function f := 1
∂ f ∗ (x) = {PC (x)}, for all x ∈ H.
(cid:27)
(cid:26)
(cid:104)y , y(cid:105)
f ∗ (x) = sup
y∈C
(cid:104)x, x(cid:105) +
1
1
sup
y∈C
2
2
(cid:107)x(cid:107)2 − 1
(cid:107)x − y(cid:107)2 =
(cid:107)x(cid:107)2 − 1
1
1
d2
C (x)
inf
y∈C
2
2
2
2
(cid:107)x − PC (x)(cid:107)2 = (cid:104)x, PC (x)(cid:105) − 1
(cid:107)x(cid:107)2 − 1
(cid:107)PC (x)(cid:107)2
1
=
2
2
2
= (cid:104)x, PC (x)(cid:105) − f (PC (x)).
Consequently, by Proposition 2.7, PC (x) ∈ ∂ f ∗ (x) for x ∈ X . Now suppose y ∈ ∂ f ∗ (x), and define
n (y − PC (x)). Then xn → x, and hence PC (xn ) → PC (x) by Lemma 3.1. Using the
xn = x + 1
subdifferential inequality, we have
0 ≤ (cid:104)xn − x, PC (xn ) − y(cid:105) =
{−(cid:104)x, x(cid:105) + 2(cid:104)x, y(cid:105) − (cid:104)y , y(cid:105)}
(cid:104)y − PC (x), PC (xn ) − y(cid:105).
1
n
(cid:104)x, y(cid:105) − 1
2
=
=
This now implies:
0 ≤ lim
n→∞(cid:104)y − PC (x), PC (xn ) − y(cid:105) = −(cid:107)y − PC (x)(cid:107)2 .
Consequently, y = PC (x) and so (18) is established.
Since f ∗ is continuous and we just proved that ∂ f ∗ is a singleton, Proposition 2.3 implies that f ∗
is Gateaux differentiable. Now −∞ < f ∗∗ (x) ≤ f (x) = 1
2 (cid:107)x(cid:107)2 for all x ∈ C . Thus, f ∗∗ is a proper
function. One can easily check that f is sequentially weakly lsc, C being weakly closed. Therefore,
(cid:4)
Theorem 2.15 implies that f is convex; whence, dom f = C must be convex.
Observe that we have actually proved that every Chebyshev set with a continuous projection
mapping is convex (and closed). We finish the section by recalling a simple but powerful “hidden
convexity” result.
Remark 3.3 (See [5].) Let C be a closed subset of a Hilbert space H . Then there exists a
C (x) = (cid:107)x(cid:107)2 − f (x), ∀x ∈ H . Precisely,
continuous and convex function f defined on H such that d2
f can be taken as x (cid:55)→ supc∈C {2(cid:104)x, c(cid:105) − (cid:107)c(cid:107)2}.
18
4 Monotone operator theory
multifunction), i.e., for every x ∈ X , Ax ⊆ X ∗ , and let gra A := (cid:8)(x, x∗ ) ∈ X × X ∗ x∗ ∈ Ax(cid:9) be
Let A : X ⇒ X ∗ be a set-valued operator (also known as a relation, point-to-set mapping or
the graph of A. The domain of A is dom A := (cid:8)x ∈ X Ax (cid:54)= ∅(cid:9) and ran A := A(X ) is the range
of A. We say that A is monotone if
for all (x, x∗ ), (y , y∗ ) ∈ gra A,
(cid:104)x − y , x∗ − y∗ (cid:105) ≥ 0,
(19)
and maximal ly monotone if A is monotone and A has no proper monotone extension (in the sense
of graph inclusion). Given A monotone, we say that (x, x∗ ) ∈ X × X ∗ is monotonical ly related to
gra A if
(cid:104)x − y , x∗ − y∗ (cid:105) ≥ 0,
for all (y , y∗ ) ∈ gra A.
Monotone operators have frequently shown themselves to be a key class of ob jects in both modern
Optimization and Analysis; see, e.g., [13, 14, 15, 24], the books [7, 22, 28, 53, 61, 62, 58, 66, 67, 68]
and the references given therein.
Given sets S ⊆ X and D ⊆ X ∗ , we define S⊥ by S⊥ := {x∗ ∈ X ∗ (cid:104)x∗ , x(cid:105) = 0,
∀x ∈ S }
and D⊥ by D⊥ := {x ∈ X (cid:104)x, x∗ (cid:105) = 0,
∀x∗ ∈ D} [54]. Then the adjoint of A is the operator
gra A∗ := (cid:8)(x∗∗ , x∗ ) ∈ X ∗∗ × X ∗ (x∗ , −x∗∗ ) ∈ (gra A)⊥(cid:9).
A∗ : X ∗∗ ⇒ X ∗ such that
Note that the adjoint is always a linear relation, i.e. its graph is a linear subspace.
The Fitzpatrick function [33] associated with an operator A is the function FA : X × X ∗ →
(cid:16)(cid:104)x, a∗ (cid:105) + (cid:104)a, x∗ (cid:105) − (cid:104)a, a∗ (cid:105)(cid:17)
]−∞, +∞] defined by
Fitzpatrick functions have been proved to be an important tool in modern monotone operator
theory. One of the main reasons is shown in the following result.
FA (x, x∗ ) :=
(20)
sup
(a,a∗ )∈gra A
.
Fact 4.1 (Fitzpatrick) (See ([33, Propositions 3.2&4.2, Theorem 3.4 and Corollary 3.9].) Let
A : X ⇒ X ∗ be monotone with dom A (cid:54)= ∅. Then FA is proper lower semicontinuous in the
norm × weak∗ -topology ω(X ∗ , X ), convex, and FA = (cid:104)·, ·(cid:105) on gra A. Moreover, if A is maximal ly
monotone, for every (x, x∗ ) ∈ X × X ∗ , the inequality
(cid:104)x, x∗ (cid:105) ≤ FA (x, x∗ ) ≤ F ∗
A (x∗ , x)
is true, and the first equality holds if and only if (x, x∗ ) ∈ gra A.
The next result is central to maximal monotone operator theory and algorithmic analysis. Orig-
inally it was proved by more extended direct methods than the concise convex analysis argument
we present next.
19
Theorem 4.2 (Local boundedness) (See [53, Theorem 2.2.8].) Let A : X ⇒ X ∗ be monotone
with int dom A (cid:54)= ∅. Then A is local ly bounded at x ∈ int dom A, i.e., there exist δ > 0 and K > 0
such that
(cid:107)y∗(cid:107) ≤ K,
∀y ∈ x + δBX .
sup
y∗∈Ay
Proof. Let x ∈ int dom A. After translating the graphs if necessary, we can and do suppose that
x = 0 and (0, 0) ∈ gra A. Define f : X → ]−∞, +∞] by
(cid:104)y − a, a∗ (cid:105).
y (cid:55)→
sup
(a,a∗ )∈gra A, (cid:107)a(cid:107)≤1
By Fact 2.1(i), f is convex and lower semicontinuous. Since 0 ∈ int dom A, there exists δ1 > 0 such
that δ1BX ⊆ dom A. Now we show that δ1BX ⊆ dom f . Let y ∈ δ1BX and y∗ ∈ Ay . Thence, we
have
∀(a, a∗ ) ∈ gra A, (cid:107)a(cid:107) ≤ 1
(cid:104)y − a, y∗ − a∗ (cid:105) ≥ 0,
⇒ (cid:104)y − a, y∗ (cid:105) ≥ (cid:104)y − a, a∗ (cid:105),
∀(a, a∗ ) ∈ gra A, (cid:107)a(cid:107) ≤ 1
⇒ +∞ > ((cid:107)y(cid:107) + 1) · (cid:107)y∗(cid:107) ≥ (cid:104)y − a, a∗ (cid:105),
∀(a, a∗ ) ∈ gra A, (cid:107)a(cid:107) ≤ 1
⇒ f (y) < +∞ ⇒ y ∈ dom f .
Hence δ1BX ⊆ dom f and thus 0 ∈ int dom f . By Fact 2.1(iv), there is δ > 0 with δ ≤ min{ 1
2 δ1}
2 , 1
such that
∀y ∈ 2δBX .
f (y) ≤ f (0) + 1,
Now we show that f (0) = 0. Since (0, 0) ∈ gra A, then f (0) ≥ 0. On the other hand, by
the monotonicity of A, (cid:104)a, a∗ (cid:105) = (cid:104)a − 0, a∗ − 0(cid:105) ≥ 0 for every (a, a∗ ) ∈ gra A. Then we have
f (0) = sup(a,a∗ )∈gra A, (cid:107)a(cid:107)≤1 (cid:104)0 − a, a∗ (cid:105) ≤ 0. Thence f (0) = 0.
Thus,
∀y ∈ 2δBX , (a, a∗ ) ∈ gra A, (cid:107)a(cid:107) ≤ δ,
(cid:104)y , a∗ (cid:105) ≤ (cid:104)a, a∗ (cid:105) + 1,
whence, taking the supremum with y ∈ 2δBX ,
2δ(cid:107)a∗(cid:107) ≤ (cid:107)a(cid:107) · (cid:107)a∗(cid:107) + 1 ≤ δ(cid:107)a∗(cid:107) + 1,
∀(a, a∗ ) ∈ gra A, a ∈ δBX
⇒ (cid:107)a∗(cid:107) ≤ 1
∀(a, a∗ ) ∈ gra A, a ∈ δBX .
δ
,
Setting K := 1
δ , we get the desired result.
Generalizations of Theorem 4.2 can be found in [62, 18] and [23, Lemma 4.1].
(cid:4)
4.1 Sum theorem and Minty surjectivity theorem
In the early 1960s, Minty [43] presented an important characterization of maximally monotone
operators in a Hilbert space; which we now reestablish. The proof we give of Theorem 4.3 is due
to Simons and Zalinescu [63, Theorem 1.2]. We denote by Id the identity mapping from H to H .
20
Theorem 4.3 (Minty) Suppose that H is a Hilbert space. Let A : H ⇒ H be monotone. Then
A is maximal ly monotone if and only if ran(A + Id) = H .
Proof. “⇒”: Fix any x∗
0 ∈ H , and let B : H ⇒ H be given by gra B := gra A − {(0, x∗
0 )}. Then B
is maximally monotone. Define F : H × H → ]−∞, +∞] by
(x, x∗ ) (cid:55)→ FB (x, x∗ ) +
(cid:107)x2 +
(21)
1
2
(cid:107)x∗ 2 .
1
2
Fact 4.1 together with Fact 2.1(v) implies that F is coercive. By [66, Theorem 2.5.1(ii)], F has a
minimizer. Assume that (z , z ∗ ) ∈ H × H is a minimizer of F . Then we have (0, 0) ∈ ∂F (z , z ∗ ).
(cid:10)(−z , −z ∗ ), (b, b∗ ) − (z , z ∗ )(cid:11) ≤ FB (b, b∗ ) − FB (z , z ∗ ),
Thus, (0, 0) ∈ ∂FB (z , z ∗ ) + (z , z ∗ ) and (−z , −z ∗ ) ∈ ∂FB (z , z ∗ ). Then
∀(b, b∗ ) ∈ gra B ,
and by Fact 4.1, (cid:10)(−z , −z ∗ ), (b, b∗ ) − (z , z ∗ )(cid:11) ≤ (cid:104)b, b∗ (cid:105) − (cid:104)z , z ∗ (cid:105),
that is,
∀(b, b∗ ) ∈ gra B ;
∀(b, b∗ ) ∈ gra B .
0 ≤ (cid:104)b, b∗ (cid:105) − (cid:104)z , z ∗ (cid:105) + (cid:104)z , b(cid:105) + (cid:104)z ∗ , b∗ (cid:105) − (cid:107)z(cid:107)2 − (cid:107)z ∗(cid:107)2 ,
(22)
Hence,(cid:10)b + z ∗ , b∗ + z(cid:11) = (cid:104)b, b∗ (cid:105) + (cid:104)z , b(cid:105) + (cid:104)z ∗ , b∗ (cid:105) + (cid:104)z , z ∗ (cid:105) ≥ (cid:107)z + z ∗(cid:107)2 ≥ 0,
∀(b, b∗ ) ∈ gra B ,
which implies that (−z ∗ , −z ) ∈ gra B , since B is maximally monotone. This combined with (22)
implies 0 ≤ −2(cid:104)z , z ∗ (cid:105) − (cid:107)z(cid:107)2 − (cid:107)z ∗(cid:107)2 . Then we have z = −z ∗ , and (z , −z ) = (−z ∗ , −z ) ∈ gra B ,
whence (z , −z ) + (0, x∗
0 ) ∈ gra A. Therefor x∗
0 ∈ Az + z , which implies x∗
0 ∈ ran(A + Id).
“⇐”: Let (v , v∗ ) ∈ H × H be monotonically related to gra A. Since ran(A + Id) = H , there
−(cid:107)v − y(cid:107)2 = (cid:10)v − y , y∗ + y − v − y∗(cid:11) = (cid:10)v − y , v∗ − y∗(cid:11) ≥ 0.
exists (y , y∗ ) ∈ gra A such that v∗ + v = y∗ + y . Then we have
Hence v = y , which also implies v∗ = y∗ . Thus (v , v∗ ) ∈ gra A, and therefore A is maximally
(cid:4)
monotone.
Remark 4.4 The extension of Minty’s theorem to reflexive spaces (in which case it asserts the
surjectivity of A + JX for the normalized duality mapping JX defined below) was originally proved
by Rockafellar. The proof given in [22, Proposition 3.5.6, page 119] which uses Fenchel’s duality
theorem more directly than the one we gave here, is only slightly more complicated than that of
Theorem 4.3.
A + B : X ⇒ X ∗ : x (cid:55)→ Ax + Bx := (cid:8)a∗ + b∗ a∗ ∈ Ax and b∗ ∈ Bx(cid:9) is monotone. Rockafellar
Let A and B be maximally monotone operators from X to X ∗ . Clearly, the sum operator
established the following important result in 1970 [57], the so-called “sum theorem”: Suppose
If dom A ∩ int dom B (cid:54)= ∅, then A + B is maximally monotone. We can
that X is reflexive.
21
weaken this constraint qualification to be that (cid:83)
[4, 62, 64, 22, 2]).
We turn to a new proof of this generalized result. To this end, we need the following fact
along with the definition of the partial inf-convolution. Given two real Banach spaces X, Y and
F1 , F2 : X × Y → ]−∞, +∞], the partial inf-convolution F1(cid:3)2F2 is the function defined on X × Y
(cid:8)F1 (x, y − v) + F2 (x, v)(cid:9).
by
F1(cid:3)2F2 : (x, y) (cid:55)→ inf
v∈Y
λ>0 λ [dom A − dom B ] is a closed subspace (see
.
Fact 4.5 (Simons and Zalinescu) (See [64, Theorem 4.2] or [62, Theorem 16.4(a)].) Let X, Y
be real Banach spaces and F1 , F2 : X × Y → ]−∞, +∞] be proper lower semicontinuous and convex
bifunctionals. Assume that for every (x, y) ∈ X × Y ,
and that (cid:83)
(F1(cid:3)2F2 )(x, y) > −∞
λ>0 λ [PX dom F1 − PX dom F2 ] is a closed subspace of X . Then for every (x∗ , y∗ ) ∈
X ∗ × Y ∗ ,
u∗∈X ∗ {F ∗
1 (x∗ − u∗ , y∗ ) + F ∗
2 (u∗ , y∗ )} .
(F1(cid:3)2F2 )∗ (x∗ , y∗ ) = min
We denote by JX the duality map from X to X ∗ , which will be simply written as J , i.e., the
2 (cid:107) · (cid:107)2 . Let F : X × Y → ]−∞, +∞] be a bifunctional defined on
subdifferential of the function 1
two real Banach spaces. Following the notation by Penot [51] we set
: Y × X : (y , x) (cid:55)→ F (x, y).
(23)
F
monotone. Assume that (cid:83)
Theorem 4.6 (Sum theorem) Suppose that X is reflexive. Let A, B : X ⇒ X be maximal ly
λ>0 λ [dom A − dom B ] is a closed subspace. Then A + B is maximal ly
monotone.
Proof. Clearly, A + B is monotone. Assume that (z , z ∗ ) ∈ X × X ∗ is monotonically related to
1 . By [9, Lemma 5.8], (cid:83)
gra(A + B ).
∗(cid:124)
λ>0 λ [PX (dom FA ) − PX (dom FB )] is
Let F1 := FA(cid:3)2FB , and F2 := F
a closed subspace. Then Fact 4.5 implies that
A (x∗ − u∗ , x) + F ∗
u∗∈X ∗ {F ∗
B (u∗ , x)} ,
1 (x∗ , x) = min
F ∗
Set G : X × X ∗ → ]−∞, +∞] by
(x, x∗ ) (cid:55)→ F2 (x + z , x∗ + z ∗ ) − (cid:104)x, z ∗ (cid:105) − (cid:104)z , x∗ (cid:105) +
(cid:107)x∗(cid:107)2 .
(cid:107)x(cid:107)2 +
1
1
2
2
0 ) ∈ X ×X ∗ is a minimizer of G. ([66, Theorem 2.5.1(ii)] implies that minimizers
Assume that (x0 , x∗
exist since G is coercive). Then we have (0, 0) ∈ ∂G(x0 , x∗
0 ). Thus, there exists v∗ ∈ J x0 , v ∈ JX ∗ x∗
0
such that (0, 0) ∈ ∂F2 (x0 + z , x∗
0 + z ∗ ) + (v∗ , v) + (−z ∗ , −z ), and then
(z ∗ − v∗ , z − v) ∈ ∂F2 (x0 + z , x∗
0 + z ∗ ).
for all (x, x∗ ) ∈ X × X ∗ .
(24)
(cid:124)
22
Thence (cid:68)
(25)
(cid:69)
(z ∗ − v∗ , z − v), (x0 + z , x∗
0 + z ∗ )
Fact 4.1 and (24) show that
2 (z ∗ − v∗ , z − v).
0 + z ∗ ) + F ∗
= F2 (x0 + z , x∗
∗(cid:124)
2 = F1 ≥ (cid:104)·, ·(cid:105).
F2 ≥ (cid:104)·, ·(cid:105), F
Then by (25),(cid:68)
(cid:69)
≥ (cid:10)x0 + z , x∗
0 + z ∗(cid:11) + (cid:10)z ∗ − v∗ , z − v(cid:11).
(z ∗ − v∗ , z − v), (x0 + z , x∗
2 (z ∗ − v∗ , z − v)
0 + z ∗ )
= F2 (x0 + z , x∗
0 + z ∗ ) + F ∗
(26)
(cid:69) − (cid:10)x0 + z , x∗
(cid:68)
Thus, since v∗ ∈ J x0 , v ∈ JX ∗ x∗
0 + z ∗(cid:11) − (cid:10)z ∗ − v∗ , z − v(cid:11)
0 ,
= (cid:10) − x0 − v , x∗
0 + v∗(cid:11) = (cid:104)−x0 , x∗
0 ≤ δ :=
(z ∗ − v∗ , z − v), (x0 + z , x∗
0 + z ∗ )
0 (cid:105) − (cid:104)x0 , v∗ (cid:105) − (cid:104)v , x∗
0 (cid:105) − (cid:104)v , v∗ (cid:105)
0 (cid:105) − 1
= (cid:104)−x0 , x∗
(cid:107)v(cid:107)2 − (cid:104)v , v∗ (cid:105),
(cid:107)v∗(cid:107)2 − 1
(cid:107)x0(cid:107)2 − 1
0(cid:107)2 − 1
(cid:107)x∗
2
2
2
2
which implies
that is,
(cid:107)x∗
0(cid:107)2 +
(cid:107)x0(cid:107)2 = 0;
δ = 0 and (cid:104)x0 , x∗
0 (cid:105) +
1
1
2
2
0 + z ∗ ) = (cid:10)x0 + z , x∗
0 + z ∗(cid:11). By (24) and Fact 4.1,
0 ∈ −J x0 .
and x∗
(27)
δ = 0
Combining (26) and (27), we have F2 (x0 + z , x∗
0 + z ∗ ) ∈ gra(A + B ).
(x0 + z , x∗
(28)
(cid:11) = (cid:10)x0 + z − z , x∗
(cid:10)x0 , x∗
0 + z ∗ − z ∗(cid:11) ≥ 0,
Since (z , z ∗ ) is monotonically related to gra(A + B ), it follows from (28) that
0
−(cid:107)x0(cid:107)2 = −(cid:107)x∗
0(cid:107)2 ≥ 0,
0 ) = (0, 0). Finally, by (28), one deduces that (z , z ∗ ) ∈ gra(A + B ) and A + B is
whence (x0 , x∗
(cid:4)
maximally monotone.
It is still unknown whether the reflexivity condition can be omitted in Theorem 4.6 though many
partial results exist, see [14, 15] and [22, §9.7].
and then by (27),
23
(29)
4.2 Autoconjugate functions
Given F : X × X ∗ → ]−∞, +∞], we say that F is autoconjugate if F = F ∗(cid:124)
gra A = (cid:8)(x, x∗ ) ∈ X × X ∗ F (x, x∗ ) = (cid:104)x, x∗ (cid:105)(cid:9).
F is a representer for gra A if
Autoconjugate functions are the core of representer theory, which has been comprehensively studied
in Optimization and Partial Differential Equations (see [8, 9, 52, 62, 22, 34]).
Fitzpatrick posed the following question in [33, Problem 5.5]:
If A : X ⇒ X ∗ is maximal ly monotone, does there necessarily exist an autoconjugate
representer for A?
on X × X ∗ . We say
Bauschke and Wang gave an affirmative answer to the above question in reflexive spaces by con-
struction of the function BA in Fact 4.7. The first construction of an autoconjugate representer for
a maximally monotone operator satisfying a mild constraint qualification in a reflexive space was
provided by Penot and Zalinescu in [52]. This naturally raises a question:
Is BA still an autoconjugate representer for a maximally monotone operator A in a
general Banach space?
We give a negative answer to the above question in Example 4.12: in certain spaces, BA fails to be
autoconjugate.
Fact 4.7 (Bauschke and Wang) (See [8, Theorem 5.7].) Suppose that X is reflexive. Let
A : X ⇒ X ∗ be maximal ly monotone. Then
(cid:110) 1
2 (cid:107)y∗(cid:107)2(cid:111)
BA : X × X ∗ → ]−∞, +∞]
∗(cid:124)
A (x − y , x∗ − y∗ ) + 1
(x, x∗ ) (cid:55)→
2 (cid:107)y(cid:107)2 + 1
2 FA (x + y , x∗ + y∗ ) + 1
(30)
inf
2 F
(y ,y∗ )∈X×X ∗
is an autoconjugate representer for A.
We will make use of the following result to prove Theorem 4.11 below.
Fact 4.8 (Simons) (See [62, Corollary 10.4].) Let f1 , f2 , g : X → ]−∞, +∞] be proper convex.
(cid:8) 1
4 g(2z )(cid:9) > −∞,
Assume that g is continuous at a point of dom f1 − dom f2 . Suppose that
2 f2 (x − z ) + 1
2 f1 (x + z ) + 1
(cid:8) 1
4 g∗ (−2z ∗ )(cid:9) ,
2 (x∗ − z ∗ ) + 1
2 f ∗
1 (x∗ + z ∗ ) + 1
2 f ∗
h(x) := inf
z∈X
Then
∀x ∈ X.
∀x∗ ∈ X ∗ .
h∗ (x∗ ) = min
z∗∈X ∗
24
Let A : X ⇒ X ∗ be a linear relation. We say that A is skew if gra A ⊆ gra(−A∗ ); equivalently,
if (cid:104)x, x∗ (cid:105) = 0, ∀(x, x∗ ) ∈ gra A. Furthermore, A is symmetric if gra A ⊆ gra A∗ ; equivalently, if
(cid:104)x, y∗ (cid:105) = (cid:104)y , x∗ (cid:105), ∀(x, x∗ ), (y , y∗ ) ∈ gra A. We define the symmetric part and the skew part of A via
(31)
2 A − 1
2 A∗
2 A∗ ,
P := 1
and S := 1
2 A + 1
respectively. It is easy to check that P is symmetric and that S is skew.
Fact 4.9 (See [6, Theorem 3.7].) Let A : X ∗ → X ∗∗ be linear and continuous. Assume that
ran A ⊆ X and that there exists e ∈ X ∗∗\X such that
∀x∗ ∈ X ∗ .
(cid:104)Ax∗ , x∗ (cid:105) = (cid:104)e, x∗ (cid:105)2 ,
gra T := (cid:8)(−S x∗ , x∗ ) x∗ ∈ X ∗ , (cid:104)e, x∗ (cid:105) = 0(cid:9) = (cid:8)(−Ax∗ , x∗ ) x∗ ∈ X ∗ , (cid:104)e, x∗ (cid:105) = 0(cid:9).
Let P and S respectively be the symmetric part and skew part of A. Let T : X ⇒ X ∗ be defined by
Then the fol lowing hold.
(i) A is a maximal ly monotone operator on X ∗ .
(ii) P x∗ = (cid:104)x∗ , e(cid:105)e, ∀x∗ ∈ X ∗ .
(32)
(iii) T is maximal ly monotone and skew on X .
(iv) gra T ∗ = {(S x∗ + re, x∗ ) x∗ ∈ X ∗ , r ∈ R}.
(v) FT = ιC , where C := {(−Ax∗ , x∗ ) x∗ ∈ X ∗}.
(33)
We next give concrete examples of A, T as in Fact 4.9.
Example 4.10 (c0 ) (See [6, Example 4.1].) Let X := c0 , with norm (cid:107) · (cid:107)∞ so that X ∗ = (cid:96)1
with norm (cid:107) · (cid:107)1 , and X ∗∗ = (cid:96)∞ with its second dual norm (cid:107) · (cid:107)∗ (i.e., (cid:107)y(cid:107)∗ := supn∈N yn , ∀y :=
(cid:88)
(yn )n∈N ∈ (cid:96)∞ ). Fix α := (αn )n∈N ∈ (cid:96)∞ with lim sup αn (cid:54)= 0, and let Aα : (cid:96)1 → (cid:96)∞ be defined by
∀x∗ = (x∗
n )n∈N ∈ (cid:96)1 .
αnαix∗
nx∗
(Aαx∗ )n := α2
n + 2
i ,
i>n
Now let Pα and Sα respectively be the symmetric part and skew part of Aα . Let Tα : c0 ⇒ X ∗ be
gra Tα := (cid:8)(−Sαx∗ , x∗ ) x∗ ∈ X ∗ , (cid:104)α, x∗ (cid:105) = 0(cid:9) = (cid:8)(−Aαx∗ , x∗ ) x∗ ∈ X ∗ , (cid:104)α, x∗ (cid:105) = 0(cid:9)
defined by
(cid:88)
= (cid:8)(cid:0)(− (cid:88)
i )n∈N , x∗(cid:1) x∗ ∈ X ∗ , (cid:104)α, x∗ (cid:105) = 0(cid:9).
αnαix∗
αnαix∗
i +
i<n
i>n
(34)
Then
25
∀x∗ = (x∗
n )n∈N ∈ (cid:96)1 and (34) is well defined.
(i) (cid:104)Aαx∗ , x∗ (cid:105) = (cid:104)α, x∗ (cid:105)2 ,
(ii) Aα is a maximally monotone.
(iii) Tα is a maximally monotone operator.
i − (cid:88)
(cid:88)
(cid:0)G(x∗ )(cid:1)
(iv) Let G : (cid:96)1 → (cid:96)∞ be Gossez’s operator [35] defined by
x∗
x∗
i ,
i>n
i<n
Then Te : c0 ⇒ (cid:96)1 as defined by
gra Te := {(−G(x∗ ), x∗ ) x∗ ∈ (cid:96)1 , (cid:104)x∗ , e(cid:105) = 0}
∀(x∗
n )n∈N ∈ (cid:96)1 .
n :=
is a maximally monotone operator, where e := (1, 1, . . . , 1, . . .).
We may now show that BT need not be autoconjugate.
Theorem 4.11 Let A : X ∗ → X ∗∗ be linear and continuous. Assume that ran A ⊆ X and that
there exists e ∈ X ∗∗\X such that (cid:107)e(cid:107) < 1√
and
2
(cid:104)Ax∗ , x∗ (cid:105) = (cid:104)e, x∗ (cid:105)2 ,
∀x∗ ∈ X ∗ .
♦
Let P and S respectively be the symmetric part and skew part of A. Let T , C be defined as in
Fact 4.9. Then
BT (−Aa∗ , a∗ ) > B∗
T (a∗ , −Aa∗ ),
In consequence, BT is not autoconjugate.
Proof. First we claim that
∀a∗ /∈ {e}⊥ .
∗(cid:124)
C X×X ∗ = ιgra T .
(35)
ι
Clearly, if we set D := {(A∗x∗ , x∗ ) x∗ ∈ X ∗}, we have
∗(cid:124)
(cid:124)
(cid:124)
C = ι
C⊥ = ιD ,
C = σ
(36)
ι
where in the second equality we use the fact that C is a subspace. Additionally,
A∗x∗ ∈ X ⇔ (S + P )∗x∗ ∈ X ⇔ S ∗x∗ + P ∗x∗ ∈ X ⇔ −S x∗ + P x∗ ∈ X
⇔ −S x∗ − P x∗ + 2P x∗ ∈ X ⇔ 2P x∗ − Ax∗ ∈ X ⇔ P x∗ ∈ X (since ran A ⊆ X )
⇔ (cid:104)x∗ , e(cid:105)e ∈ X (by Fact 4.9(ii))
⇔ (cid:104)x∗ , e(cid:105) = 0 (since e /∈ X ).
(37)
Observe that P x∗ = 0 for all x∗ ∈ {e}⊥ by Fact 4.9(ii). Thus, A∗x∗ = −Ax∗ for all x∗ ∈ {e}⊥ .
Combining (36) and (37), we have
∗(cid:124)
C X×X ∗ = ιD∩(X×X ∗ ) = ιgra T ,
ι
26
(38)
Thus
(40)
=
min
(y∗ ,y∗∗ )∈X ∗×X ∗∗
BT (−Aa∗ , a∗ ) = inf
y=−Ay∗
and hence (35) holds.
Let a∗ /∈ {e}⊥ . Then (cid:104)a∗ , e(cid:105) (cid:54)= 0. Now we compute BT (−Aa∗ , a∗ ). By Fact 4.9(v) and (35),
2 (cid:107)y∗(cid:107)2(cid:9) .
(cid:8)ιC (−Aa∗ + y , a∗ + y∗ ) + ιgra T (−Aa∗ − y , a∗ − y∗ ) + 1
BT (−Aa∗ , a∗ )
2 (cid:107)y(cid:107)2 + 1
inf
=
(y ,y∗ )∈X×X ∗
2 (cid:107)y∗(cid:107)2(cid:9)
(cid:8)ιgra T (−Aa∗ − y , a∗ − y∗ ) + 1
(cid:8) 1
2 (cid:107)y∗(cid:107)2(cid:9) =
(cid:8) 1
2 (cid:107)y∗(cid:107)2(cid:9)
2 (cid:107)y(cid:107)2 + 1
2 (cid:107)Ay∗(cid:107)2 + 1
2 (cid:107)y(cid:107)2 + 1
inf
inf
=
y=−Ay∗ , (cid:104)a∗−y∗ ,e(cid:105)=0
(cid:104)a∗−y∗ ,e(cid:105)=0
≥
(cid:104)e, y∗ (cid:105)2
(cid:104)Ay∗ , y∗ (cid:105) =
inf
inf
(cid:104)a∗−y∗ ,e(cid:105)=0
(cid:104)a∗−y∗ ,e(cid:105)=0
= (cid:104)e, a∗ (cid:105)2 .
(39)
Next we will compute B∗
T (a∗ , −Aa∗ ). By Fact 4.8 and (38), we have
(cid:26) 1
(cid:27)
B∗
T (a∗ , −Aa∗ )
2 (cid:107)y∗(cid:107)2(cid:111)
(cid:110)
2 (cid:107)y∗∗(cid:107)2 + 1
gra T (a∗ − y∗ , −Aa∗ − y∗∗ ) + 1
C (a∗ + y∗ , −Aa∗ + y∗∗ ) +
2 (cid:107)y∗(cid:107)2
ι∗
ι∗
1
2
2
ιD (−Aa∗ + y∗∗ , a∗ + y∗ ) + ι(gra T )⊥ (a∗ − y∗ , −Aa∗ − y∗∗ ) + 1
2 (cid:107)y∗∗(cid:107)2 + 1
min
=
(y∗ ,y∗∗ )∈X ∗×X ∗∗
≤ ιD (−Aa∗ + 2P a∗ , a∗ ) + ι(gra T )⊥ (a∗ , −Aa∗ − 2P a∗ ) + 1
2 (cid:107)2P a∗(cid:107)2
(by taking y∗ = 0, y∗∗ = 2P a∗ )
= ιgra(−T ∗ ) (−Aa∗ − 2P a∗ , a∗ ) + 1
2 (cid:107)2P a∗(cid:107)2
2 (cid:107)2P a∗(cid:107)2
= 1
(by Fact 4.9(iv))
2 (cid:107)2(cid:104)a∗ , e(cid:105)e(cid:107)2
= 1
= 2(cid:104)a∗ , e(cid:105)2(cid:107)e(cid:107)2 .
This inequality along with (39), (cid:104)e, a∗ (cid:105) (cid:54)= 0 and (cid:107)e(cid:107) < 1√
, yield
2
BT (−Aa∗ , a∗ ) ≥ (cid:104)e, a∗ (cid:105)2 > 2(cid:104)a∗ , e(cid:105)2(cid:107)e(cid:107)2 ≥ B∗
T (a∗ , −Aa∗ ),
Hence BT is not autoconjugate.
(cid:4)
Example 4.12 (Example 4.10 revisited) Let X := c0 , with norm (cid:107) · (cid:107)∞ so that X ∗ = (cid:96)1 with
norm (cid:107) · (cid:107)1 , and X ∗∗ = (cid:96)∞ with its second dual norm (cid:107) · (cid:107)∗ . Fix α := (αn )n∈N ∈ (cid:96)∞ with
(cid:88)
lim sup αn (cid:54)= 0 and (cid:107)α(cid:107)∗ < 1√
, and let Aα : (cid:96)1 → (cid:96)∞ be defined by
2
∀x∗ = (x∗
n )n∈N ∈ (cid:96)1 .
αnαix∗
nx∗
(Aαx∗ )n := α2
n + 2
i ,
i>n
Now let Pα and Sα respectively be the symmetric part and skew part of Aα . Let Tα : c0 ⇒ X ∗ be
gra Tα := (cid:8)(−Sαx∗ , x∗ ) x∗ ∈ X ∗ , (cid:104)α, x∗ (cid:105) = 0(cid:9) = (cid:8)(−Aαx∗ , x∗ ) x∗ ∈ X ∗ , (cid:104)α, x∗ (cid:105) = 0(cid:9)
defined by
27
∀a∗ /∈ {e}⊥ .
(by Fact 4.9(ii))
(by (36))
(41)
αnαix∗
i +
= (cid:8)(cid:0)(− (cid:88)
i>n
Then, by Example 4.10 and Theorem 4.11,
BTα (−Aa∗ , a∗ ) > B∗
Tα (a∗ , −Aa∗ ),
In consequence, BTα is not autoconjugate.
i )n∈N , x∗(cid:1) x∗ ∈ X ∗ , (cid:104)α, x∗ (cid:105) = 0(cid:9).
αnαix∗
(cid:88)
i<n
∀a∗ /∈ {e}⊥ .
♦
The latter raises a very interesting question:
Problem 4.13 Is there a maximal ly monotone operator on some (resp. every) non-reflexive Ba-
nach space that has no autoconjugate representer?
4.3 The Fitzpatrick function and differentiability
The Fitzpatrick function introduced in [33] was discovered precisely to provide a more transparent
convex alternative to the earlier saddle function construction due to Krauss [22]—we have not
discussed saddle-functions but they produce interesting maximally monotone operators [57, §33
& §37]. At the time, Fitzpatrick’s interests were more centrally in the differentiation theory for
convex functions and monotone operators.
The search for results relating when a maximally monotone T is single-valued to differentiability
of FT did not yield fruit, and he put the function aside. This is still the one area where to the best
of our knowledge FT has proved of very little help—in part because generic properties of dom FT
and of dom(T ) seem poorly related.
That said, monotone operators often provide efficient ways to prove differentiability of convex
functions. The discussion of Mignot’s theorem in[22] is somewhat representative of how this works
as is the treatment in [53]. By contrast, as we have seen the Fitzpatrick function and its relatives
now provide the easiest access to a gamut of solvability and boundedness results.
5 Other results
5.1 Renorming results: Asplund averaging
Edgar Asplund [3] showed how to exploit convex analysis to provide remarkable results on the
existence of equivalent norms with nice properties. Most optimizers are unaware of his lovely idea
which we recast in the language of inf-convolution. Our development is a reworking of that in
Day [31]. Let us start with two equivalent norms (cid:107) · (cid:107)1 and (cid:107) · (cid:107)2 on a Banach space X . We consider
2/2, and average for n ≥ 0 by
1/2 and q0 := (cid:107) · (cid:107)2
the quadratic forms p0 := (cid:107) · (cid:107)2
(pn(cid:3)qn )(2x)
pn (x) + qn (x)
pn+1 (x) :=
(42)
.
and qn+1 (x) :=
2
2
Let C > 0 be such that q0 ≤ p0 ≤ (1 + C )q0 . By the construction of pn and qn , we have
qn ≤ pn ≤ (1 + 4−nC )qn ([3, Lemma]) and so the sequences (pn )n∈N , (qn )n∈N converge to a common
√
limit: a convex quadratic function p.
2p typically inherits the good properties of both (cid:107) · (cid:107)1
We shall show that the norm (cid:107) · (cid:107)3 :=
and (cid:107) · (cid:107)2 . This is based on the following fairly straightforward result.
28
Theorem 5.1 (Asplund) (See [3, Theorem 1].) If either p0 or q0 is strictly convex, so is p.
We make a very simple application in the case that X is reflexive. In [41], Lindenstrauss showed
that every reflexive Banach space has an equivalent strictly convex norm. The reader may consult
[22, Chapter 4] for more general results. Now take (cid:107) · (cid:107)1 to be an equivalent strictly convex norm
on X , and take (cid:107) · (cid:107)2 to be an equivalent smooth norm with its dual norm on X ∗ strictly convex.
Theorem 5.1 shows that p is strictly convex. We note that by Corollary 2.25 and Fact 2.21
n (x∗ ) + qn (x∗ )
q∗
(cid:3)q∗
(p∗
n )(2x∗ )
q∗
n+1 (x∗ ) :=
and p∗
n+1 (x∗ ) :=
n
2
2
so that Theorem 5.1 applies to p∗
0 . Hence p∗ is strictly convex (see also [30, Proof of
0 and q∗
√
√
Corollary 1, page 111]). Hence (cid:107) · (cid:107)3 (:=
2p∗ ) are equivalent strictly
2p) and its dual norm (:=
convex norms on X and X ∗ respectively.
Hence (cid:107) · (cid:107)3 is an equivalent strictly convex and smooth norm (since its dual is strictly convex).
The existence of such a norm was one ingredient of Rockafellar’s first proof of the Sum theorem.
5.2 Resolvents of maximally monotone operators and connection with convex
functions
It is well known since Minty, Rockafellar, and Bertsekas-Eckstein that in Hilbert spaces, monotone
operators can be analyzed from the alternative viewpoint of certain nonexpansive (and thus Lips-
chitz continuous) mappings, more precisely, the so-called resolvents. Given a Hilbert space H and
a set-valued operator A : H ⇒ H , the resolvent of A is
JA := (Id +A)−1 .
The history of this notion goes back to Minty [43] (in Hilbert spaces) and Brezis, Crandall and
Pazy [29] (in Banach spaces). There exist more general notions of resolvents based on different
tools, such as the normalized duality mapping, the Bregman distance or other maximally monotone
operators (see [40, 1, 11]). For more details on resolvents on Hilbert spaces see [7].
The Minty surjectivity theorem (Theorem 4.3 [43]) implies that a monotone operator is maximally
monotone if and only if the resolvent is single-valued with full domain. In fact, a classical result due
to Eckstein-Bertsekas [32] says even more. Recall that a mapping T : H → H is firmly nonexpansive
if for all x, y ∈ H , (cid:107)T x − T y(cid:107) ≤ (cid:104)T x − ty , x − y(cid:105).
Theorem 5.2 Let H be a Hilbert space. An operator A : H ⇒ H is (maximal) monotone if and
only if JA is firmly nonexpansive (with ful l domain).
Example 5.3 Given a closed convex set C ⊆ H , the normal cone operator of C , NC , is a maximally
monotone operator whose resolvent can be proved to be the metric pro jection onto C . Therefore,
♦
Theorem 5.2 implies the firm nonexpansivity of the metric pro jection.
In the particular case when A is the subdifferential of a possibly non-differentiable convex function
in a Hilbert space, whose maximal monotonicity was established by Moreau [48] (in Banach spaces
this is due to Rockafellar [56], see also [25, 22]), the resolvent turns into the proximal mapping in
29
the following sense of Moreau. If f : H → ]−∞, +∞] is a lower semicontinuous convex function
(cid:26)
(cid:27)
defined on a Hilbert space H , the proximal or proximity mapping is the operator proxf : H → H
defined by
(cid:107)x − y(cid:107)2
1
f (y) +
proxf (x) := argmin
.
y∈H
2
This mapping is well-defined because proxf (x) exists and is unique for all x ∈ H . Moreover, there
exists the following subdifferential characterization: u = proxf (x) if and only if x − u ∈ ∂ f (u).
Moreau’s decomposition in terms of the proximal mapping is a powerful nonlinear analysis tool
in the Hilbert setting that has been used in various areas of optimization and applied mathematics.
Moreau established his decomposition motivated by problems in unilateral mechanics. It can be
proved readily by using the conjugate and subdifferential.
Theorem 5.4 (Moreau decomposition) Given a lower semicontinuous convex function f :
H → ]−∞, +∞], for al l x ∈ H ,
x = proxf (x) + proxf ∗ (x).
Example 5.5 Note that for f := ιC , with C closed and convex, the proximal mapping turns into
the pro jection onto a closed and convex set C . Therefore, this result generalizes the decomposition
by orthogonal pro jection on subspaces. In particular, if K is a closed convex cone (thus ι∗
K = ιK− ,
see Example 2.18), Moreau’s decomposition provides a characterization of the pro jection onto K :
x = y + z with y ∈ K , z ∈ K − and (cid:104)y , z (cid:105) = 0 ⇔ y = PK x and z = PK− x.
This illustrates that in Hilbert space, the Moreau decomposition can be thought of as generalizing
the decomposition into positive and negative parts of a vector in a normed lattice [22, §6.7] to an
♦
arbitrary convex cone.
There is another notion associated to an operator A, which is strongly related to the resolvent.
That is the Yosida approximation of index λ > 0 or the Yosida λ-regularization :
(Id −JλA ).
Aλ := (λ Id +A−1 )−1 =
1
λ
If the operator A is maximally monotone, so is the Yosida approximation, and along with the
resolvent they provide the so-called Minty parametrization of the graph of A that is Lipschitz
continuous in both directions [58]:
(JλA (z ), Aλ (z )) = (x, y) ⇔ z = x + y , (x, y) ∈ gra A.
If A = ∂ f is the subdifferential of a proper lower semicontinuous convex function f , it turns out
that the Yosida approximation of A is the gradient of the Moreau envelope of f eλf , defined as the
(cid:27)
(cid:26)
infimal convolution of f and (cid:107) · (cid:107)2/2λ, that is,
eλf (x) := f (cid:3) (cid:107) · (cid:107)2
2λ
(cid:107)x − y(cid:107)2
1
2λ
= inf
y∈H
f (y) +
.
30
This justifies the alternative term Moreau-Yosida approximation for the mapping (∂ f )λ =
(λ Id +(∂ f )−1 )−1 . This allows to obtain a proof in Hilbert space of the connection between the
convexity of the function and the monotonicity of the subdifferential (see [58]): a proper lower
semicontinuous function is convex if and only its Clarke subdifferential is monotone.
It is worth mentioning that generally the role of the Moreau envelope is to approximate the
function, with a regularizing effect since it is finite and continuous even though the function may
not be so. This behavior has very useful implications in convex and variational analysis.
5.3 Symbolic convex analysis
The thesis work of Hamilton [20] has provided a conceptual and effective framework (the SCAT
Maple software) for computing conjugates, subdifferentials and infimal convolutions of functions of
several variables. Key to this is the notion of iterated conjugation (analogous to iterated integration)
and a good data structure.
(cid:18) sinh (3 x)
(cid:19)
As a first example, with some care, the convex conjugate of the function
f : x (cid:55)→ log
sinh x
(cid:33)
(cid:32)
(cid:32) (cid:112)16 − 3y2 − 2
y + (cid:112)16 − 3y2
can be symbolically nursed to obtain the result
4 − 2y
6
g : y (cid:55)→ y
2
· log
(cid:33)
,
+ log
with domain [−2, 2].
Since the conjugate of g is much more easily computed to be f , this produces a symbolic com-
putational proof that f and g are convex and are mutually conjugate.
Similarly, Maple produces the conjugate of x (cid:55)→ exp(exp(x)) as y (cid:55)→ y (log (y) − W (y) − 1/W (y))
in terms of the Lambert’s W function—the multi-valued inverse of z (cid:55)→ zez . This function is un-
y4 + O (cid:0)y5(cid:1) .
known to most humans but is built into both Maple and Mathematica. Thus Maple knows that to
order five
g(y) = −1 + (−1 + log y) y − 1
y3 − 3
1
2
3
8
Figure 3 shows the Maple -computed conjugate after the SCAT package is loaded: There is a
corresponding numerical program CCAT [20]. Current work is adding the capacity to symbolically
compute convex compositions—and so in principle Fenchel duality.
y2 +
5.4 Partial Fractions and Convexity
(cid:33) (cid:32) N(cid:88)
(cid:32) N(cid:89)
(cid:88)
We consider a network objective function pN given by
qσ(i)(cid:80)N
j=i qσ(j )
σ∈SN
i=1
i=1
pN (q) :=
(cid:33)
,
1(cid:80)N
j=i qσ(j )
31
Figure 3: The conjugate and subdifferential of exp exp.
(cid:33)
(cid:33) (cid:32) N(cid:88)
(cid:32) N(cid:89)
summed over al l N ! permutations; so a typical term is
1(cid:80)n
qi(cid:80)N
j=i qj
j=i qj
i=1
i=1
(cid:19) (cid:18) 1
(cid:19) (cid:18) 1
(cid:19) (cid:18)
q2 + q3
q3
(cid:18)
For example, with N = 3 this is
1
q1 + q2 + q3
1
q1 + q2 + q3
1
q2 + q3
q1 q2 q3
+
.
(cid:19)
.
+
1
q3
(43)
This arose as the ob jective function in research into coupon collection. The researcher, Ian Affleck,
wished to show pN was convex on the positive orthant.
First, we tried to simplify the expression for pN . The partial fraction decomposition gives:
1
x1
1
x1
1
x1
1
x1 + x2
−
1
x3
1
x1 + x2 + x3
1
x2 + x3
1
x1 + x2
1
x1 + x3
p3 (x1 , x2 , x3 ) =
p2 (x1 , x2 ) =
1
x2
1
x2
−
+
p1 (x1 ) =
,
+
.
−
−
+
+
,
32
> > > > > > > > > > (1)(1)(2)(2)(3)(3)(4)(4)restart:read("scat.mpl"):read("ccat.mpl"):with(SCAT);with(CCAT);"1. combinat,gfun,student,IntegerRelations,PolynomialTools loaded""2. p2s,s2p,r2p,f2p,pslq,find loaded""3. anim and binzetas loaded"f11:=convert(exp(exp(x)),PWF);g11:=Conj(f11,y);sdg11:=SubDiff(g11);Plot(sdg11,y=-1..1,view=[0..1,-5..0],axes=boxed,labels=["$y$",""]); − (cid:88)
N(cid:88)
In [60], the simplified expression of PN is given by
1
xi
1≤i<j≤N
i=1
− . . . + (−1)N −1
p(x1 , x2 , · · · , xN ) :=
(cid:88)
1≤i<j<k≤N
1
xi + xj
+
1
xi + xj + xk
1
x1 + x2 + . . . + xN
.
Partial fraction decompositions are another arena in which computer algebra systems are hugely
useful. The reader is invited to try performing the third case in (43) by hand. It is tempting to
predict the “same” pattern will hold for N = 4. This is easy to confirm (by computer if not by
hand) and so we are led to:
(cid:32)
(cid:33)
Conjecture 5.6 For each N ∈ N, the function
(cid:90) 1
1 − N(cid:89)
(1 − txi )
0
i=1
pN (x1 , · · · , xN ) =
dt
t
(44)
is convex; indeed 1/pN is concave.
One may check symbolically that this is true for N < 5 via a large Hessian computation. But this
is impractical for larger N . That said, it is easy to numerically sample the Hessian for much larger
N , and it is always positive definite. Unfortunately, while the integral is convex, the integrand is
not, or we would be done. Nonetheless, the process was already a success, as the researcher was
able to rederive his ob jective function in the form of (44).
A year after, Omar Hjab suggested re-expressing (44) as the joint expectation of Poisson distri-
butions.3 Explicitly, this leads to:
(cid:33) (cid:90)
(cid:32) n(cid:89)
(cid:33)
(cid:32)
Lemma 5.7 [17, §1.7] If x = (x1 , · · · , xn ) is a point in the positive orthant Rn
(cid:90) ∞
1 − n(cid:89)
++ , then
(1 − e−txi )
Rn
0
i=1
i=1
++
e−(cid:104)x,y(cid:105) max(y1 , · · · , yn ) dy ,
(45)
where (cid:104)x, y(cid:105) = x1y1 + · · · + xnyn is the Euclidean inner product.
(cid:18) y1
(cid:19)
(cid:90)
It follows from the lemma—which is proven in [17] with no recourse to probability theory—that
x1
RN
++
e−(y1+···+yN ) max
pN (x) =
dt =
xi
, · · · ,
yN
xN
dy ,
a+b ≤ √
and hence that pN is positive, decreasing, and convex, as is the integrand. To derive the stronger
result that 1/pN is concave we refer to [17, §1.7]. Observe that since 2ab
ab ≤ (a + b)/2, it
follows from (45) that pN is log-convex (and convex). A little more analysis of the integrand shows
3 See “Convex, II” SIAM Electronic Problems and Solutions at http://www.siam.org/journals/problems/
downloadfiles/99- 5sii.pdf.
33
pN is strictly convex on its domain. The same techniques apply when xk is replaced in (43) or (44)
by g(xk ) for a concave positive function g .
Though much nice related work is to found in [60], there is still no truly direct proof of the
convexity of pN . Surely there should be! This development neatly shows both the power of
computer assisted convex analysis and its current limitations.
(cid:12)(cid:12)(cid:12)(cid:12) sin x
(cid:12)(cid:12)(cid:12)(cid:12)p
(cid:90) ∞
Lest one think most results on the real line are easy, we challenge the reader to prove the empirical
observation that
p (cid:55)→ √
p
dx
x
0
is difference convex on (1, ∞), i.e. it can be written as a difference of two convex functions [5].
6 Concluding comments
All researchers and practitioners in convex analysis and optimization owe a great debt to Jean-
Jacques Moreau—whether they know so or not. We are delighted to help make his seminal role
more apparent to the current generation of scholars. For those who read French we urge them to
experience the pleasure of [44, 45, 46, 48] and especially [49]. For others, we highly recommend
[50], which follows [48] and of which Zuhair Nashed wrote in his Mathematical Review MR0217617:
“There is a great need for papers of this kind; the present paper serves as a model of clarity and
motivation.”
Acknowledgments The authors are grateful to the three anonymous referees for their pertinent
and constructive comments. The authors also thank Dr. Hristo S. Sendov for sending them the
manuscript [60]. The authors were all partially supported by various Australian Research Council
grants.
References
[1] Y. Alber and D. Butnariu, “Convergence of Bregman pro jection methods for solving consistent
convex feasibility problems in reflexive Banach spaces”, Journal of Optimization Theory and
Applications, vol. 92, pp. 33–61, 1997.
[2] M. Alimohammady and V. Dadashi, “Preserving maximal monotonicity with applications in
sum and composition rules”, Optimization Letters, vol. 7, pp. 511–517, 2013.
[3] E. Asplund, “Averaged norms”, Israel Journal of Mathematics vol. 5, pp. 227–233, 1967.
[4] H. Attouch, H. Riahi, and M. Thera, “Somme ponctuelle d’operateurs maximaux monotones”
[Pointwise sum of maximal monotone operators] Well-posedness and stability of variational
problems. Serdica. Mathematical Journal, vol. 22, pp. 165–190, 1996.
[5] M. Bac´ak and J.M. Borwein, “On difference convexity of locally Lipschitz functions”, Opti-
mization, pp. 961–978, 2011.
34
[6] H.H. Bauschke, J.M. Borwein, X. Wang, and L. Yao, “Construction of pathological maximally
monotone operators on non-reflexive Banach spaces”, Set-Valued and Variational Analysis,
vol. 20, pp. 387–415, 2012.
[7] H.H. Bauschke and P.L. Combettes, Convex Analysis and Monotone Operator Theory in
Hilbert Spaces, Springer, 2011.
[8] H.H. Bauschke and X. Wang, “The kernel average for two convex functions and its applications
to the extension and representation of monotone operators”, Transactions of the American
Mathematical Society, vol. 36, pp. 5947–5965, 2009.
[9] H.H. Bauschke, X. Wang, and L. Yao, “Monotone linear relations: maximality and Fitzpatrick
functions”, Journal of Convex Analysis, vol. 16, pp. 673–686, 2009.
[10] H.H. Bauschke, X. Wang, and L. Yao, “Autoconjugate representers for linear monotone oper-
ators”, Mathematical Programming (Series B), vol. 123, pp. 5-24, 2010.
[11] H.H. Bauschke, X. Wang, and L. Yao, “General resolvents for monotone operators: characteri-
zation and extension”, in Biomedical Mathematics: Promising Directions in Imaging, Therapy
Planning and Inverse Problems, Medical Physics Publishing, pp. 57–74, 2010.
[12] J.M. Borwein, “A generalization of Young’s (cid:96)p inequality”, Mathematical Inequalities & Ap-
plications, vol. 1, pp. 131–136, 1998.
[13] J.M. Borwein, “Maximal monotonicity via convex analysis”, Journal of Convex Analysis,
vol. 13, pp. 561–586, 2006.
[14] J.M. Borwein, “Maximality of sums of two maximal monotone operators in general Banach
space”, Proceedings of the American Mathematical Society, vol. 135, pp. 3917–3924, 2007.
[15] J.M. Borwein, “Fifty years of maximal monotonicity”, Optimization Letters, vol. 4, pp. 473–
490, 2010.
[16] J.M. Borwein and D.H. Bailey, Mathematics by Experiment: Plausible Reasoning in the 21st
Century, A.K. Peters Ltd, Second expanded edition, 2008.
[17] J.M. Borwein, D.H. Bailey and R. Girgensohn, Experimentation in Mathematics: Computa-
tional Paths to Discovery, A.K. Peters Ltd, 2004. ISBN: 1-56881-211-6.
[18] J.M. Borwein and S. Fitzpatrick, “Local boundedness of monotone operators under minimal
hypotheses”, Bul letin of the Australian Mathematical Society, vol. 39, pp. 439–441, 1989.
[19] J.M. Borwein, R.S Burachik, and L. Yao, “Conditions for zero duality gap in convex program-
ming”, Journal of Nonlinear and Convex Analysis, in press; http://arxiv.org/abs/1211.
4953v2.
[20] J.M. Borwein and C. Hamilton, “Symbolic Convex Analysis: Algorithms and Examples,”
Mathematical Programming, 116 (2009), 17–35. Maple packages SCAT and CCAT available
at http://carma.newcastle.edu.au/ConvexFunctions/SCAT.ZIP.
35
[21] J.M. Borwein and A.S. Lewis, Convex Analyis andd Nonsmooth Optimization, Second ex-
panded edition, Springer, 2005.
[22] J.M. Borwein and J.D. Vanderwerff, Convex Functions, Cambridge University Press, 2010.
[23] J.M. Borwein and L. Yao, “Structure theory for maximally monotone operators with points of
continuity”, Journal of Optimization Theory and Applications, vol 157, pp. 1–24, 2013 (Invited
paper).
[24] J.M. Borwein and L. Yao, “Recent progress on Monotone Operator Theory”, Infinite Products
of Operators and Their Applications, Contemporary Mathematics, in press; http://arxiv.
org/abs/1210.3401v2.
[25] J.M. Borwein and Q.J. Zhu, Techniques of variational analysis, CMS Books in Mathematic-
s/Ouvrages de Mathmatiques de la SMC, 20. Springer-Verlag, New York, 2005.
[26] R.I. Bot¸ S. Grad, and G. Wanka, Duality in Vector Optimization, Springer, 2009.
[27] R.I. Bot¸ and G. Wanka, “A weaker regularity condition for subdifferential calculus and Fenchel
duality in infinite dimensional spaces”, Nonlinear Analysis, vol. 64, pp. 2787–2804, 2006.
[28] R.S. Burachik and A.N. Iusem, Set-Valued Mappings and Enlargements of Monotone Opera-
tors, Springer, vol. 8, 2008.
[29] H. Brezis, G. Crandall and P. Pazy, Perturbations of nonlinear maximal monotone sets in
Banach spaces, Communications on Pure and Applied Mathematics, vol. 23, pp. 123–144,
1970.
[30] J. Diestel, Geometry of Banach spaces, Springer-Verlag, 1975
[31] M.M. Day, Normed linear spaces, Third edition, Springer-Verlag, New York-Heidelberg, 1973.
[32] J. Eckstein and D.P. Bertsekas, “On the Douglas–Rachford splitting method and the proxi-
mal point algorithm for maximal monotone operators”, Mathematical Programming, vol. 55,
pp. 293–318, 1992.
[33] S. Fitzpatrick, “Representing monotone operators by convex functions”, in Workshop/Mini-
conference on Functional Analysis and Optimization (Canberra 1988), Proceedings of the Cen-
tre for Mathematical Analysis, Australian National University, vol. 20, Canberra, Australia,
pp. 59–65, 1988.
[34] N. Ghoussoub, Self-dual partial differential systems and their variational principles. Springer
Monographs in Mathematics, Springer, 2009.
[35] J.-P. Gossez, “On the range of a coercive maximal monotone operator in a nonreflexive Banach
space”, Proceedings of the American Mathematical Society, vol. 35, pp. 88–92, 1972.
[36] J.-B. Hiriart-Urruty, M. Moussaoui, A. Seeger, and M. Volle, “Subdifferential calculus without
qualification conditions, using approximate subdifferentials: a survey”, Nonlinear Analysis,
vol. 24, pp. 1727–1754, 1995.
36
[37] J.-B. Hiriart-Urruty and R. Phelps, “Subdifferential Calculus Using ε-Subdifferentials”, Jour-
nal of Functional Analysis vol. 118, pp. 154–166, 1993.
[38] L. Hormander, “Sur la fonction d’appui des ensembles convexes dans un espace localement
convexe”, Arkiv for Matematik, vol. 3, pp. 181–186, 1955.
[39] V. Klee, “Convexity of Chebysev sets”, Mathematische Annalen, vol. 142, pp. 292–304, 1961.
[40] F. Kohsaka and W. Takahashi, “Existence and approximation of fixed points of firmly nonex-
pansivetype mappings in Banach spaces”, SIAM Journal on Optimization, vol. 19, pp. 824–835,
2008.
[41] J. Lindenstrauss, “On nonseparable reflexive Banach spaces”, Bul letin of the American Math-
ematical Society, vol. 72, pp. 967–970, 1966.
[42] P. Mar´echal, “A convexity theorem for multiplicative functions”, Optimization Letters, vol. 6,
pp. 357–362, 2012.
[43] G. Minty, “Monotone (nonlinear) operators in a Hilbert space”, Duke Mathematical Journal,
vol. 29, pp. 341–346, 1962.
[44] J.J. Moreau, “Fonctions convexes en dualit´e”, Facult´e des Sciences de Montpellier, S´eminaires
de Math´ematiques Universit´e de Montpellier, Montpellier, 1962.
[45] J.J. Moreau, “Fonctions `a valeurs dans [−∞, +∞]; notions alg´ebriques”, Facult´e des Sciences
de Montpel lier, S´eminaires de Math´ematiques, Universit´e de Montpellier, Montpellier, 1963.
[46] J.J. Moreau, “ ´Etude locale d’une fonctionnelle convexe”, Facult´e des Sciences de Montpellier,
S´eminaires de Math´ematiques Universit´e de Montpellier, Montpellier, 1963.
[47] J.J. Moreau, “Sur la function polaire d’une fonctionelle semi-continue sup´erieurement”,
Comptes Rendus de l’Acad´emie des Sciences, vol. 258, pp. 1128–1130, 1964.
[48] J.J. Moreau, “Proximit´e et dualit´e dans un espace hilbertien”, Bul letin de la Soci´et´e
Math´ematique de France, vol. 93, pp. 273–299, 1965.
[49] J.J. Moreau, Fonctionnel les convexes, S´eminaire Jean Leray, College de France, Paris,
pp. 1–108, 1966–1967. Available at http://carma.newcastle.edu.au/ConvexFunctions/
moreau66-67.pdf.
[50] J.J. Moreau, “Convexity and duality”, pp. 145–169 in Functional Analysis and Optimization,
Academic Press, New York, 1966.
[51] J.-P. Penot, “The relevance of convex analysis for the study of monotonicity”, Nonlinear
Analysis, vol. 58, pp. 855–871, 2004.
[52] J.-P. Penot and C. Zalinescu, “Some problems about the representation of monotone opera-
tors by convex functions”, The Australian New Zealand Industrial and Applied Mathematics
Journal, vol. 47, pp. 1–20, 2005.
37
[53] R.R. Phelps, Convex Functions, Monotone Operators and Differentiability, 2nd Edition,
Springer-Verlag, 1993.
[54] R.R. Phelps and S. Simons, “Unbounded linear monotone operators on nonreflexive Banach
spaces”, Journal of Nonlinear and Convex Analysis, vol. 5, pp. 303–328, 1998.
[55] R.T. Rockafellar, “Extension of Fenchel’s duality theorem for convex functions”, Duke Math-
ematical Journal, vol. 33, pp. 81–89, 1966.
[56] R.T. Rockafellar, “On the maximal monotonicity of subdifferential mappings”, Pacific Journal
of Mathematics, vol. 33, pp. 209–216, 1970.
[57] R.T. Rockafellar, “On the maximality of sums of nonlinear monotone operators”, Transactions
of the American Mathematical Society, vol. 149, pp. 75–88, 1970.
[58] R.T. Rockafellar and R.J-B Wets, Variational analysis. Grundlehren der Mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences], 317. Springer-Verlag, Berlin,
1998 (3rd Printing, 2009).
[59] R. Rudin, Functional Analysis, Second Edition, McGraw-Hill, 1991.
[60] H.S. Sendov and R. Zitikis, “The shape of the Borwein-Affleck-Girgensohn function gener-
ated by completely monotone and Bernstein functions”, Journal of Optimization Theory and
Applications, in press.
[61] S. Simons, Minimax and Monotonicity, Springer-Verlag, 1998.
[62] S. Simons, From Hahn-Banach to Monotonicity, Springer-Verlag, 2008.
[63] S. Simons and C. Zalinescu, “A new proof for Rockafellar’s characterization of maximal mono-
tone operators”, Proceedings of the American Mathematical Society, vol. 132, pp. 2969–2972,
2004.
[64] S. Simons and C. Zalinescu, “Fenchel duality, Fitzpatrick functions and maximal monotonic-
ity”, Journal of Nonlinear and Convex Analysis, vol. 6, pp. 1–22, 2005.
[65] L. Sucheston, “Banach limits”, American Mathematical Monthly, vol. 74, pp. 308–311, 1967.
[66] C. Zalinescu, Convex Analysis in General Vector Spaces, World Scientific Publishing, 2002.
[67] E. Zeidler, Nonlinear Functional Analysis and its Applications II/A: Linear Monotone Oper-
ators, Springer-Verlag, 1990.
[68] E. Zeidler, Nonlinear Functional Analysis and its Applications II/B: Nonlinear Monotone
Operators, Springer-Verlag, 1990.
38
|
1306.5116 | 2 | 1306 | 2015-05-04T10:43:18 | On the positive eigenvalues and eigenvectors of a non-negative matrix | [
"math.FA",
"math.OA",
"math.PR"
] | The paper develops the general theory for the items in the title, assuming that the matrix is countable and cofinal. | math.FA | math |
for all vertexes v, and
Bvwξw ≤ eβξv
Bvwξw = eβξv
(1.1)
(1.2)
Xw∈V
Xw∈V
ON THE POSITIVE EIGENVALUES AND EIGENVECTORS OF A
NON-NEGATIVE MATRIX
KLAUS THOMSEN
1. Introduction
Recent work on KMS states and weights on graph C ∗-algebras has uncovered an
intimate relation to positive eigenvalues and eigenvectors for non-negative matri-
ces naturally associated to the graph and the one-parameter action. Specifically,
for the gauge action on a graph C ∗-algebra there is a non-negative matrix B over
the vertexes V in the graph, such that KMS weights corresponding to the inverse
temperature β ∈ R are in bijective correspondence with the non-zero non-negative
vectors ξ with the properties that
when v is not a sink and does not emit infinitely many edges.
than a weight is sought for, one should in addition insist that Pv ξv = 1. The
same kind of equations, but with different matrices determine also the gauge invari-
ant KMS weights and states for more general actions on graph C ∗-algebras. See
[aHLRS],[CL],[Th].
If a state, rather
Finding the solutions to (1.1) and (1.2) is in general a highly non-trivial task.
The literature on positive eigenvalues and eigenfunctions of a non-negative matrix
is enormous, and for finite graphs there are in fact results available that can be
used to determine the possible values of β and for each β get a description of the
corresponding vectors ξ, albeit with some additional work, cf.
[CT]. For infinite
matrices this is no longer the case - very far from. The most fundamental questions
prompted by the connection to KMS states and weights are those that I guess come
to the mind of any mathematician:
a) For which β are there non-zero non-negative solutions to the equations ?
b) How does the structure of the solutions vary with β ?
From the theory of countable state Markov chains it is known that these questions are
often extremely hard to answer already when the matrix is irreducible and stochastic,
but also that the structure of the solutions can be very rich and interesting.
The concern here is that a general setting for an approach to the above problem is
missing, although the theory of harmonic and super-harmonic functions of countable
state Markov chains comes close. It is the purpose with the present paper to provide
a framework for the work on the problem when the graph C ∗-algebra is simple, or
more precisely when the graph is cofinal. As far as I know, no one has developed
Version: July 4, 2018.
1
2
KLAUS THOMSEN
the theory in this setting. It involves a possibly infinite matrix B as above for which
the underlying directed graph need not be strongly connected, and what is sought
are neither exactly the harmonic functions of e−βB nor exactly the super-harmonic
functions. But it is something in between and the cofinality of the associated graph
is a property which can substitute for the often assumed strong connectivity. What
I show is how the known methods, which typically deal with harmonic or super-
harmonic functions (or vectors) of a non-negative matrix, often stochastic or sub-
stochastic for which the underlying digraph is strongly connected, can be modified to
the yield the desired framework. As a consequence the paper is expository because
the ideas I present are known. The purpose is to show how the tools must be arranged
in order to address the problem above. When the graph is strongly connected with
finite out-degree at every vertex, everything I present can be obtained from the
theory of countable state Markov chains in combination with the work of Vere-
Jones, [V], although the translation may not be straightforward. In the more general
cofinal case, and in the presence of infinite emitters, some non-trivial adjustments
to the methods developed for Markov chains must be performed, and rather than
describing first the Markov chain results and then the adjustments, I have chosen to
give a self-contained account, requiring no knowledge of Markov chains or random
walks. I make no claim of originality for the underlying ideas, but I hope that the
presentation will be useful for mathematicians interested in KMS weights and states
on C ∗-algebras. Needless to say, I would be happy if workers from other fields of
mathematics also find it worthwhile. I have kept the list of references to an absolute
minimum by quoting only my own sources. The reason for this is that I am unable
to point to the original sources of the ideas presented, not for lack of good will, but
out of ignorance. I therefore choose to follow the principle 'none mentioned, none
forgotten'. I apologise to anyone offended by this.
Once the the right setup is in place it is easy to begin to harvest results from
the theory of Markov chains. As an illustration of this I use in the final section
the theorem on convergence to the boundary for a countable state Markov chain to
obtain the general and abstract description of extremal solutions to equations (1.1)
and (1.2) in the cofinal case.
2. The setting
Let V be a countable set and
V × V ∋ (v, w) 7→ Avw ∈ [0, ∞)
a non-negative matrix A over V . We can then consider the directed graph G with
vertexes V such that there is an arrow from v ∈ V to w ∈ V if and only if Avw 6= 0.
Let E denote the set of edges (or arrows) in G. When µ is an edge, or more
generally a finite path in G, we denote by s(µ) and r(µ) its initial and terminal
vertex, respectively. Let V∞ denote the union of the sinks and the infinite emitters
in V , i.e.
V∞ =(cid:8)v ∈ V : #s−1(v) ∈ {0, ∞}(cid:9) .
A subset H ⊆ V is hereditary when e ∈ E, s(e) ∈ H ⇒ r(e) ∈ H, and saturated
when
v ∈ V \V∞, r(s−1(v)) ⊆ H ⇒ v ∈ H.
We assume that A is cofinal in the sense that V does not contain any subsets that
are both hereditary and saturated other than ∅ and V .
POSITIVE EIGENVECTORS
3
Lemma 2.1. Assume that A is cofinal. Let e1e2e3 · · · be an infinite path in G. For
every v ∈ V there is an i ∈ N and a finite path µ in G such that s(µ) = v and
r(µ) = s(ei).
Proof. The set of vertexes v which do not have the stated property is hereditary and
saturated, and it is not all of V since it does not contain s(e1). It must therefore be
empty.
(cid:3)
A vertex v ∈ V is non-wandering when there is a finite path µ in G such that
v = s(µ) = r(µ). We denote by NW the set of non-wandering vertexes in V . The
non-wandering subgraph of G is the subgraph GN W consisting of the vertexes NW
It follows that GN W is strongly
and the edges emitted from any of its elements.
connected in the sense that for any pair v, w ∈ NW there is a finite path µ in G
such that s(µ) = v and r(µ) = w:
Lemma 2.2. NW is a (possibly empty) hereditary subset of V and the graph GN W
is strongly connected.
Proof. When v ∈ NW there is an infinite path in G which visits v infinitely often.
So when e ∈ E and s(e) ∈ NW the cofinality of G ensures that there is a path µ in
G connecting r(e) to s(e) by Lemma 2.1. Then eµ is a loop in G containing r(e),
proving that r(e) ∈ NW , and hence that NW is hereditary. The proof that GN W
is strongly connected is similar.
(cid:3)
Let β ∈ R. We are here looking for maps (or vectors) ξ : V → [0, ∞) such that
for all v ∈ V \V∞ and
Avwξw = eβξv
Avwξw ≤ eβξv
(2.1)
(2.2)
Xw∈V
Xw∈V
for v ∈ V∞. We say then that ξ is almost β-harmonic for A. When (2.1) holds for
all v ∈ V , and not only for v ∈ V \V∞, we say that ξ is β-harmonic for A.
Lemma 2.3. Let H be a non-empty hereditary subset of V and η : H → [0, ∞) a
function such that (2.1) holds for all v ∈ H\V∞ and (2.2) holds for all v ∈ H ∩ V∞.
There is a unique almost β-harmonic vector ξ such that ξv = ηv for all v ∈ H.
Proof. Set
H1 = {v ∈ V \V∞ : Avw 6= 0 ⇒ w ∈ H} ∪ H.
Then H1 is hereditary, contains H and there is a unique extension of η to H1 given by
we get a sequence H ⊆ H1 ⊆ H2 ⊆ H3 ⊆ · · · of subsets of V and a unique extension
the condition that eβηv =Pw∈V Avwηw for all v ∈ H1\H. Continuing by induction
ξ of η toSn Hn. This completes the proof sinceSn Hn is hereditary and saturated,
and hence equal to V .
(cid:3)
Lemma 2.4. Assume that G contains a sink (i.e. A contains a zero row).
It
follows that for all β ∈ R there is a non-zero almost β-harmonic vector, unique up
to multiplication by constants.
4
KLAUS THOMSEN
Proof. Note that a sink s constitutes a hereditary subset in itself. It follows therefore
from Lemma 2.3 that an almost β-harmonic vector is determined by its value at s.
To show that there exists a non-zero almost β-harmonic vector, set ηs = 1 and apply
Lemma 2.3.
(cid:3)
Lemma 2.5. Let ξ be a non-zero almost β-harmonic vector. It follows that ξv > 0
for all v ∈ V .
Proof. The set {v ∈ V : ξv = 0} is hereditary and saturated. Since ξ is not zero the
set is not all of V , and it must therefore be empty.
(cid:3)
We define the matrices An, n = 0, 1, 2, . . . , recursively such that A0 = I, where I
is the identity matrix,
A1 = A, and
0
otherwise,
Ivw =(1 when v = w
vw =Xu∈V
AvuAn
uw
An+1
when n ≥ 1. While the entries in A are finite by assumption, this need not be the
case for An, but since (2.1) and (2.2) imply that
An
vwξw ≤ enβξv
(2.3)
Xw∈V
for all v, n, the following conclusion follows from Lemma 2.5.
Lemma 2.6. Assume that there is a non-zero almost β-harmonic vector for A. It
follows that An
vw < ∞ for all n ∈ N and all v, w ∈ V .
In the following we therefore assume that all powers of A are finite. Since Lemma
2.4 contains all the information we seek when there is a sink present, there is also
nothing lost by assuming that there are no sinks in G. To summarise we assume in
the following that
i) G is cofinal,
ii) that there are no sinks in G (equivalently, there are no zero rows in A), and
iii) that An
In particular, from now on V∞ consists of the infinite emitters in G, corresponding
vw < ∞ for all n ∈ N and all v, w ∈ V .
to rows in A with infinitely many non-zero entries.
Lemma 2.7. No vertex v ∈ V \NW is an infinite emitter, i.e. V∞ ⊆ NW .
Proof. Let v ∈ V be an infinite emitter. Set
A = {w ∈ V : there is a finite path µ in G such that w = s(µ) and r(µ) = v}∪{v}.
Since V \A is hereditary and saturated,
r(s−1(v)) ⊆ A, which implies that v ∈ NW .
it follows that A = V .
In particular,
(cid:3)
Lemma 2.8. Let β ∈ R. Assume that there are vertexes v0, w0 ∈ V such that
n=0 An
vw0e−nβ = ∞ for all vertexes v ∈ V .
n=0 An
P∞
v0w0e−nβ = ∞. ThenP∞
POSITIVE EIGENVECTORS
5
Proof. Set
The equality
C =(v ∈ V :
NXn=0
An
An
∞Xn=0
vw0e−nβ < ∞) .
N +1Xn=0
An
Avu
uw0e−nβ = eβ
vw0e−nβ − eβIvw0
Xu∈V
shows that C is hereditary and saturated. Since v0 /∈ C it follows that C = ∅.
(2.4)
(cid:3)
When NW is not empty we take an element v ∈ NW and set
n
β0 = log(cid:18)lim sup
β0 = log(cid:18)lim sup
n
(An
vv)
(An
vw)
1
n(cid:19)
n(cid:19) ,
1
with the convention that log ∞ = ∞. Since GN W is strongly connected by Lemma
2.2 the value β0 is independent of the choice of vertex v ∈ NW , and in fact
vvξv ≤ enβξv for all n when ξ is an
for all v, w ∈ NW . We see from (2.3) that An
almost β-harmonic vector. Since ξv > 0 by Lemma 2.5 it follows that there can not
be an almost β-harmonic vector for A unless β ≥ β0. We will therefore also in the
following assume that
iv) β0 = log(cid:16)lim supn (An
when NW 6= ∅.
vv)
1
n(cid:17) < ∞ for all v ∈ NW ,
In this section we consider the case where NW 6= ∅ and where
3. The recurrent case
An
vve−nβ0 = ∞
(3.1)
∞Xn=0
for one and hence all v ∈ NW . We say that A is recurrent in this case. The main
result will be the following theorem. In the irreducible case, i.e. when G is strongly
connected, it is contained in Corollary 2 on page 371 in [V].
Theorem 3.1. Assume that A is recurrent. There is a non-zero almost β0-harmonic
vector which is unique up to multiplication by scalars, and it is β0-harmonic for A.
The proof will require some preparations, some of which will also play a role in
the following sections. Let β ≥ β0. If the sum
An
uue−nβ
(3.2)
is finite for some u ∈ NW , it will be finite for all u ∈ NW because GN W is strongly
connected by Lemma 2.2. Since
∞Xn=0
(v ∈ V :
∞Xn=0
vue−nβ < ∞)
An
6
KLAUS THOMSEN
is hereditary and saturated, we conclude that if the sum (3.2) is finite for some
u ∈ NW , the sums
An
vue−nβ,
(3.3)
∞Xn=0
where v ∈ V, u ∈ NW , will all be finite. Since V∞ ⊆ NW by Lemma 2.7, it follows
that the sums (3.3) are all finite when v ∈ V and u ∈ V∞, provided the sum (3.2) is
finite for one (and hence all) u ∈ NW .
Lemma 3.2. Let k : V∞ → [0, ∞) be a non-negative function on V∞ such that
It follows that
∞Xn=0
Xu∈V∞
∞Xn=0
kv = Xu∈V∞
e−nβAn
uuku < ∞.
e−nβAn
vuku < ∞
for all v ∈ V , and that k is an almost β-harmonic vector.
Proof. Straightforward.
We say that a non-negative function k : V∞ → [0, ∞) is β-summable when (3.4)
holds. When V∞ is a finite set and k 6= 0, this condition is equivalent to the finiteness
of the sum (3.2) for any u ∈ NW .
Lemma 3.3. (Riesz decomposition.) Let ψ an almost β-harmonic vector. There
is a unique pair φ, k, where φ is a β-harmonic vector and k : V∞ → [0, ∞) is
β-summable, such that
Proof. The arguments are standard, cf. e.g. 6.43 on page 170 in [Wo]. φ is defined
as the limit
(3.4)
(3.5)
(cid:3)
(3.6)
while
ψ = φ + k.
e−nβAn
vwψw,
e−βAuvψv.
φv = lim
n→∞Xw∈V
ku = ψu −Xv∈V
n→∞Xw∈V
v = limn→∞Pw∈V e−nβAn
kw = ψv −Xw∈V
e−nβAn
vw
lim
k′
w = 0
e−βAvwψw = k′
It is then easy to see that (3.6) holds. For uniqueness, assume that φ′ is β-harmonic,
that k′ : V∞ → [0, ∞) is β-summable and that ψ = φ′ + k′. Then
for all v ∈ V and hence φ′
and k = k′. It follows that
e−βAvw
kv = kv −Xw∈V
for all v ∈ V∞.
vwψw = φv for all v. Thus φ = φ′
v −Xw∈V
e−βAvw
k′
w = k′
v
(cid:3)
Corollary 3.4. Assume that A is recurrent. It follows that all almost β0-harmonic
vectors are β0-harmonic.
POSITIVE EIGENVECTORS
7
Proof. This follows from Lemma 3.3 since no non-zero function k : V∞ → [0, ∞) can
be β0-summable in the recurrent case.
(cid:3)
With Corollary 3.4 in place, the proof of Theorem 3.1 can be copied from the
work of Vere-Jones, [V]. We introduce for v, w ∈ V and n = 0, 1, 2, . . . , the numbers
rvw(n) such that rvw(0) = 0, rvw(1) = Avw and
rvw(n + 1) =Xu6=w
Avuruw(n)
Lemma 3.5. (Equation (4) in [V].) Assume NW 6= ∅ and that β > β0. Then
rvw(n)e−nβ! ∞Xn=0
wwe−nβ! .
An
when n ≥ 1.
An
∞Xn=0
for all v, w ∈ NW .
Proof. By using the product rule for power series the stated equality follows from
the observation that for n ≥ 1, An
(cid:3)
s=1 rvw(s)An−s
ww .
Lemma 3.6. (Lemma 4.1 in [V].) Assume ξ : V → [0, ∞) satisfies that
for all v. Assume ξv0 6= 0 for some vertex v0 ∈ V . It follows that
vwe−nβ = Ivw + ∞Xn=1
vw =Pn
Xw∈V
∞Xn=1
Avwξw ≤ eβξv
rvv0(n)e−nβ ≤
rvv0(n)e−nβ ≤
NXn=1
for all v.
Proof. We prove by induction in N that
ξv
ξv0
ξv
ξv0
(3.7)
for all N and all v. To start the induction note that
Avwξw ≥ e−βAvv0ξv0 = ξv0rvv0(1)e−β.
Assume then that (3.7) holds for all v. It follows that
ξv
ξv0
Avw
ξw
ξv0
= e−β Xw6=v0
Avw
ξw
ξv0
≥ e−β
Avwrwv0(n)e−nβ + e−βAvv0
=
rvv0(n + 1)e−(n+1)β + e−βrvv0(1) =
ξv ≥ e−βXw∈V
≥ e−βXw∈V
NXn=1 Xw6=v0
NXn=1
+ Avv0!
rvv0(n)e−nβ.
N +1Xn=1
(cid:3)
Now note that
(3.8)
(3.9)
8
KLAUS THOMSEN
Proof of Theorem 3.1: In view Corollary 3.4 we must prove the existence and
essential uniqueness of a non-zero β0-harmonic vector. Existence: Fix a vertex
w ∈ NW . It follows from Fatou's lemma that
Since
An
∞Xn=0
for all β > β0 by Lemma 3.5, it follows that
wwe−nβ = ∞.
rww(n)e−nβ! ∞Xn=0
wwe−nβ! .
An
By the monotone convergence theorem this leads to the conclusion that
An
lim
β↓β0
∞Xn=0
wwe−nβ = 1 + ∞Xn=1
∞Xn=1
∞Xn=1
ruw(n)e−nβ0!
lim
β↓β0
rww(n)e−nβ = 1.
rww(n)e−nβ0 = 1.
=
Avu NXn=1
Xu∈V
NXn=1Xu6=w
NXn=1
N +1Xn=1
= eβ0
=
Avuruw(n)e−nβ0 + Avw
rww(n)e−nβ0
NXn=1
rvw(n + 1)e−nβ0 + Avw
rww(n)e−nβ0
NXn=1
rvw(n)e−nβ0 + Avw NXn=1
(v ∈ V :
∞Xn=1
rvw(n)e−nβ0 < ∞)
rww(n)e−nβ0 − 1! .
It follows from (3.9) and (3.8) that
is both hereditary and saturated, and hence equal to V since G is cofinal. By letting
N tend to infinity in (3.9) we see that
defines a β0-harmonic vector ξ.
Uniqueness: Let ξ′ be a non-zero β0-harmonic vector for A such that ξ′
must show that ξ′ = ξ. It follows from Lemma 3.6 that ξ′
comparing this to the fact that
w = 1. We
v ≥ ξv for all v ∈ V . By
ξv =
∞Xn=1
rvw(n)e−nβ0
enβ0 =Xv∈V
An
wvξv =Xv∈V
An
wvξ′
v
POSITIVE EIGENVECTORS
9
for all n ∈ N, we conclude that ξ′
v = ξv for every vertex v ∈ V with the property
wv 6= 0 for some n. In particular, ξ and ξ′ agree on NW since w ∈ NW , and
that An
GN W is strongly connected by Lemma 2.2. As NW is also hereditary by the same
lemma, it follows from Lemma 2.3 that ξ′ = ξ.
(cid:3)
3.1. When NW is finite. In this case there are no infinite emitters in G and hence
all almost β-harmonic vectors are β-harmonic.
Lemma 3.7. Assume that NW is non-empty but finite. Then A is recurrent and
there are no non-zero β-harmonic vectors for A when β > β0.
Proof. Note that eβ0 is the spectral radius of AN W . It follows from linear algebra
that there is a non-zero vector ψ : NW → C and a complex number λ ∈ C, λ = 1,
vwψw = enβ0λnψv for all n ∈ N and all v ∈ NW . In particular,
vwe−nβ0 can not converge to zero for all v, w ∈ NW . It follows that
the sequence An
A must be recurrent. Assume that ξ is a non-zero β-harmonic vector. Let v ∈ NW .
Then
such thatPw∈N W An
enβξv = Xw∈N W
An
vwξw ≤ K max
w∈N W
An
vw,
where K = (#NW )(maxw∈N W ξw). There is therefore a vertex w ∈ NW such that
for an increasing sequence n1 < n2 < n3 < . . . of natural numbers. It follows that
eniβξv ≤ KAni
vw
eβ = lim
i (cid:0)eniβξv(cid:1) 1
ni ≤ lim sup
n
proving that β ≤ β0.
(KAn
vw)
1
n = eβ0,
(cid:3)
Corollary 3.8. Assume that NW is non-empty and finite. It follows that there
are no non-zero β-harmonic vectors for A unless β = β0. There is a non-zero
β0-harmonic vector which unique up to multiplication by scalars.
Proof. The first statement follows from Lemma 3.7, and the second from Lemma 3.7
and Theorem 3.1.
(cid:3)
In this section we consider the non-recurrent cases. Specifically, we assume that
4. The transient case
An
vwe−nβ < ∞
(4.1)
∞Xn=0
for all v, w ∈ V , and refer to this as the transient case. The following lemma shows
that the transient case covers all the non-recurrent cases.
It follows that (4.1) holds for all v, w ∈ V .
Lemma 4.1. Assume that NW = ∅ or thatP∞
Proof. If (4.1) fails for some v, w, it follows from Lemma 2.8 thatP∞
wwe−nβ =
In particular, w ∈ NW and β = β0. This is a recurrent case, contrary to
(cid:3)
n=0 An
n=0 An
uue−nβ < ∞ for some u ∈ NW .
∞.
assumption.
10
KLAUS THOMSEN
We say that A is row-finite when there are no infinite emitter in G, i.e. when
for all v ∈ V .
# {w ∈ V : Avw > 0} < ∞
Theorem 4.2. Assume that (4.1) holds for all v, w ∈ V , and let β ∈ R.
a) Assume that NW = ∅. Then A is row-finite and there is a non-zero β-
harmonic vector for A.
b) Assume that NW is non-empty but finite. There are no non-zero almost
β-harmonic vector for A.
c) Assume that NW is infinite. There is a non-zero almost β-harmonic vector
if and only if β ≥ β0.
Proof. In case a), it follows from Lemma 2.7 that A is row-finite. Case b) follows
from Lemma 3.7 and Corollary 3.8. It remains therefore only to show that there is
a non-zero almost β-harmonic vector in case a) and c). To this end, fix v0 ∈ V and
set
Consider a vertex v ∈ Hv0 and choose l ∈ N such that Al
v0v 6= 0. Then
Hv0 =(cid:8)w ∈ V : Al
v0w 6= 0 for some l ∈ N(cid:9) .
∞Xn=0
∞Xn=0
v0we−nβ ≤ elβ
∞Xn=0
vwe−nβ ≤
Al+n
An
An
v0we−nβ.
Al
v0v
It follows that
(4.2)
(4.3)
n=0 An
n=0 An
vwe−nβ
v0we−nβ ≤
elβ
Al
v0v
P∞
P∞
for all w ∈ Hv0. Note that Hv0 is infinite. When NW is infinite this follows since
NW ⊆ Hv0 by Lemma 2.1. When NW = ∅ it follows because there are no sinks by
assumption. It follows therefore from (4.3) that there is a sequence {wk} of distinct
elements in Hv0 such that the limit
ηv = lim
k→∞ P∞
P∞
n=0 An
n=0 An
vwk e−nβ
v0wke−nβ
exists for all v ∈ Hv0. Note that ηv0 = 1. By letting N tend to ∞ in (2.4) we find
that
Avu
Xu∈V
∞Xn=0
∞Xn=0
An
uwe−nβ = eβ
An
vwe−nβ − eβIvw.
(4.4)
It follows from (4.4) that Pu∈V Avuηu = eβηv for all v ∈ Hv0\V∞, while Fatou's
lemma shows thatPu∈V Avuηu ≤ eβηv for all v ∈ Hv0. The existence of a non-zero
almost β-harmonic vector for A follows then from Lemma 2.3.
(cid:3)
When G is strongly connected and A is row-finite, c) in Theorem 4.2 is a result
of Pruitt, [P]. When A is not row-finite it can happen, also when G is strongly
connected, that there are no non-zero β-harmonic vectors for any β ≥ β0 or that
they exist for some β ≥ β0 and not for others. See [Th] for such examples.
POSITIVE EIGENVECTORS
11
4.1. The structure of the positive eigenvectors. We denote the set of almost
β-harmonic vectors for A by E(A, β). Assume that E(A, β) 6= 0. For a given vertex
v0 ∈ V we set
E(A, β)v0 = {ξ ∈ E(A, β) : ξv0 = 1} .
Equipped with the product topology RV is a locally convex real vector space, and
E(A, β) is a closed convex cone in RV . It follows from Lemma 2.5 that E(A, β)v0 is a
base for E(A, β), and we aim now to show that E(A, β)v0 is a compact Choquet sim-
plex and to obtain an integral representation of the elements in E(A, β)v0, analogous
to the Poisson-Martin integral representation for the harmonic functions of a count-
able state Markov chain, cf. e.g. Theorem 7.45 in [Wo]. For this purpose we consider
the partial ordering ≥ in E(A, β) defined such that ξ ≥ µ ⇔ ξ − µ ∈ E(A, β).
Lemma 4.3. E(A, β) is a lattice cone; i.e. every pair of elements ξ, η ∈ E(A, β)
have a least upper bound ξ ∨ η ∈ E(A, β) and a greatest lower bound ξ ∧ η ∈ E(A, β)
for the order ≥.
Proof. To find the greatest lower bound ξ ∧ µ of ξ and µ, set νv = min{ξv, µv}. Then
Pw∈V e−βAvwνw ≤Pw∈V e−βAvwξw = ξv andPw∈V e−βAvwνw ≤Pw∈V e−βAvwµw =
µv, proving that
e−βAvwνw ≤ νv.
It follows by iteration that
Xw∈V
for all v and all n. We can therefore consider the limit
ψv = lim
e−nβAn
vwνw,
e−(n+1)βAn+1
e−nβAn
vwνw ≤ νv
Xw∈V
vw νw ≤Xw∈V
n→∞Xw∈V
and observe that ψ ∈ E(A, β) while ψ ≤ µ, ψ ≤ ξ, i.e. ψ is a lower bound for ξ and
µ in E(A, β). To see that it is the greatest such, consider ϕ ∈ E(A, β) such that
ϕ ≤ µ, ϕ ≤ ξ. Then ϕ ≤ ν and
ϕv =Xw∈V
e−nβAn
vwϕw ≤Xw∈V
e−nβAn
vwνw,
for all v, n, and hence ϕ ≤ ψ. This proves that ψ is the greatest lower bound for ξ
and µ, i.e. ψ = ξ ∧ µ.
The least upper bound ξ ∨ µ is then given by
ξ ∨ µ = ξ + µ − ξ ∧ µ.
Indeed, ξ ∨ µ is clearly an upper bound and if ψ is another such, we find that
ξ + µ ≤ (ψ + µ) ∧ (ψ + ξ) = ψ + ξ ∧ µ and hence ξ ∨ µ ≤ ψ. This shows that ξ ∨ µ
is the least upper bound, as claimed.
(cid:3)
Fix now a vertex v0 ∈ V . Set
and note that Hv0 is a hereditary set of vertexes. For every v ∈ V we consider the
function K β
v : Hv0 → [0, ∞[ defined by
Hv0 =(cid:8)w ∈ V : An
v (w) = P∞
P∞
K β
v0w 6= 0 for some n ∈ N(cid:9) ,
n=0 An
n=0 An
vwe−nβ
v0we−nβ .
12
KLAUS THOMSEN
Lemma 4.4. Let H ⊆ V be a non-empty hereditary subset of vertexes. For each
v ∈ V there is a mv ∈ N such that
Al
vw 6= 0, l ≥ mv ⇒ w ∈ H.
Proof. Define subsets Hi ⊆ V recursively such that H0 = H and
Hn+1 = {v ∈ V : Avw 6= 0 ⇒ w ∈ Hn} ∪ Hn.
Then H0 ⊆ H1 ⊆ H2 ⊆ · · ·
is a sequence of hereditary subsets, and the union
assumption. When v ∈ Hk we can use mv = k.
vwe−nβ < ∞ for all v, w ∈ V . For
every vertex v ∈ V there are positive numbers lv, Lv and a finite set Fv ⊆ Hv0 such
that K β
Sn Hn is both hereditary and saturated. It is therefore all of V since G is cofinal by
Lemma 4.5. Let β ∈ R and assume thatP∞
Nv =(cid:0)Al
v0v(cid:1)−1elβ and note that the calculation (4.2) gives the upper bound
Proof. Consider first a vertex v ∈ Hv0. There is an l ∈ N such that Al
v (w) ≤ Lv for all w ∈ Hv0 and 0 < lv ≤ K β
v (w) for all w ∈ Hv0\Fv.
K β
v (w) ≤ Nv
v0v 6= 0. Set
(4.5)
n=0 An
(cid:3)
for all w ∈ Hv0. Consider then a vertex v ∈ V \Hv0. By Lemma 4.4 there is an
mv ∈ N such that every path in G of length mv emitted from v terminates in Hv0.
Let Γ denote the set of finite paths µ in G starting at v and terminating in Hv0, and
such that r(µ) is the only vertex in µ which is in Hv0. Then µ ≤ mv for all µ ∈ Γ.
Now note that
V∞ ⊆ NW ⊆ Hv0,
where the first inclusion comes from Lemma 2.7 and the second follows from Lemma
2.1. Since V∞ ⊆ Hv0 and v /∈ Hv0, the set Γ has only finitely many elements. For
every µ = e1e2 · · · eµ ∈ Γ, set
W (µ) = As(e1)r(e1)As(e2)r(e2) · · · As(eµ)r(eµ)e−µβ.
Then
An−µ
r(µ)we−(n−µ)β
v0we−nβ!
W (µ) K β
vwe−nβ =Xµ∈Γ
W (µ) Xn≥µ
r(µ)(w) ∞Xn=0
v0we−nβ!Xµ∈Γ
An
An
An
∞Xn=0
=Xµ∈Γ
≤ ∞Xn=0
Pµ∈Γ W (µ) Nr(µ), otherwise.
for all w ∈ Hv0.
W (µ) Nr(µ)
(4.6)
It follows that we can use Lv = Nv when v ∈ Hv0 and Lv =
To establish the existence of lv and Fv, assume for a contradiction that for all
ǫ > 0 there are infinitely many elements w ∈ Hv0 such that
We can then construct a sequence {wk} of distinct elements in Hv0 such that
K β
v (w) ≤ ǫ.
K β
v (wk) ≤
1
k
(4.7)
POSITIVE EIGENVECTORS
13
for all k. The calculation (4.4) shows that
Av′uK β
u (wk) = eβK β
Xu∈V
v′(wk) − eβ ∞Xn=0
v0wke−nβ!−1
An
Iv′wk
(4.8)
It follows from (4.8) that a condensation point
for all v′ ∈ V and all k ∈ N.
ξ = (ξu)u∈V inQu∈V [0, Lu] of the sequence
(cid:0)K β
u (wk)(cid:1)u∈V , k ∈ N,
is an almost β-harmonic vector for A with ξv0 = 1. But (4.7) implies that ξv = 0
which is impossible by Lemma 2.5. This contradiction shows that there must be an
lv > 0 and a finite set Fv ⊆ Hv0 such that lv ≤ K β
v (w) for all w ∈ Hv0\Fv.
(cid:3)
It follows from Lemma 4.5 that K β
v ∈
l∞ (Hv0) for all v ∈ V . We denote by 1w the characteristic function of an element
w ∈ Hv0. Let Aβ be the C ∗-subalgebra of l∞ (Hv0) generated by K β
v , v ∈ V , and the
functions 1w, w ∈ Hv0, and let Bβ be the image of Aβ in the quotient algebra
v is a bounded function on Hv0, i.e. K β
l∞ (Hv0) /c0 (Hv0\V∞) .
Here c0 (Hv0\V∞) denotes the ideal in l∞ (Hv0) consisting of the elements f : Hv0 →
C with the property that for all ǫ > 0 there is a finite subset F ⊆ Hv0\V∞ such
that f (w) ≤ ǫ ∀w ∈ Hv0\F .
In particular, f (V∞) = {0} when V∞ 6= ∅ and
f ∈ c0 (Hv0\V∞).
Note that it follows from Lemma 4.5 that for every v ∈ V there is a finite subset
Fv ⊆ Hv0 such that
K β
v + Xw∈Fv
1w
is invertible in l∞(Hv0). Thus Aβ and Bβ are both unital C ∗-algebras. Since they
are also separable, the set Xβ of characters of Bβ is a compact metric space and
Bβ can be identified with C (Xβ) via the Gelfand transform. Since evaluation at a
vertex w ∈ V∞ annihilates c0 (Hv0\V∞), each element of V∞ gives rise to character
on Bβ and hence an element of Xβ. It follows that there is a canonical inclusion
V∞ ⊆ Xβ.
(4.9)
For each v ∈ V the function w 7→ K β
v (w) is an element of Aβ and its image in Bβ is
a continuous function on Xβ which we also denote by K β
v . Let M (Xβ) denote the
set of Borel probability measures on Xβ. We consider M (Xβ) as a compact convex
set in the weak*-topology obtained by considering the measures as elements of the
dual of C(Xβ).
also that P∞
Theorem 4.6. Let β ∈ R. Assume that NW = ∅ or that NW is infinite. Assume
vwe−nβ < ∞ for all v, w ∈ V . Then E(A, β)v0 is a non-empty
compact metrizable Choquet simplex and there is a continuous affine surjection I :
M(Xβ) → E(A, β)v0 defined such that
n=0 An
I(m)v =ZXβ
K β
v dm.
14
KLAUS THOMSEN
Proof. It follows from Theorem 4.2 that E(A, β)v0 6= ∅. Since E(A, β)v0 is a base
of E(A, β), which is a lattice cone by Lemma 4.3, to conclude that E(A, β)v0 is a
compact Choquet simplex we need only show that E(A, β)v0 is compact in RV , cf.
e.g. [BR]. Note that I : M(Xβ) → RV is continuous by definition of the topologies.
It suffices therefore to show that
I (M(Xβ)) = E(A, β)v0.
(4.10)
Let x ∈ Xβ. If x ∈ V∞, we find that
n=0 An
n=0 An
vxe−nβ
v0xe−nβ
K β
v (x) = P∞
P∞
Xu∈V
AvuK β
and it follows from (4.4) that
u (x) ≤ eβK β
v (x)
(4.11)
v (x) = limk→∞ K β
for all v ∈ V with equality when v 6= x. To show that (4.11) also holds when
x ∈ Xβ\V∞, note first that point-evaluations at points in Hv0 constitute a dense
subset of the character space of Aβ. It follows that there is a sequence {uk} ⊆ Hv0
such that K β
v (uk) for all v ∈ V . Since Bβ is the quotient of Aβ by
the ideal c0(Hv0\V∞), the sequence {uk} must eventually leave every finite subset of
Hv0\V∞, and since x /∈ V∞ it must also eventually leave every finite subset of V∞.
That is, limk→∞ uk = ∞ in the sense that for any finite set F of vertexes there is an
N ∈ N such that uk /∈ F ∀k ≥ N. It follows then from (4.4) that (4.11) holds for
all v ∈ V with equality when v /∈ V∞. This shows that the inequality
AvuK β
u ≤ eβK β
v
Xu∈V
holds point-wise on Xβ for all v ∈ V , and that equality holds globally on Xβ when
v /∈ V∞.
It follows therefore by integration that I(m) ∈ E(A, β)v0 for all m ∈
M(Xβ).
To obtain (4.10) it remains to show that E(A, β)v0 ⊆ I(M(Xβ)). For this we
modify the argument from the proof of Theorem 4.1 in [Sa]. Let ξ ∈ E(A, β)v0.
Fix an element x ∈ Xβ and let {wk} be a sequence of elements in Hv0 such that
limk→∞ K β
v (x) > 0 by Lemma
2.5 it follows that
v (x) for all v ∈ V . Observe that since K β
v (wk) = K β
for all v ∈ V . For each n ∈ N, set
lim
n→∞
nK β
v (wn) = ∞
Then
v (wn)(cid:9) .
vwξn
w = 0
ξn
Am
lim
m→∞
v = min(cid:8)ξv, nK β
e−mβXw∈V
w ≤ ∞Xj=0
wwne−jβ = nXj≥m
vwξn
Am
Aj
Aj
v0wne−jβ! ne−mβXw∈V
vwne−jβ.
for all n, v. To see this note that
v0wne−jβ! e−mβXw∈V
Aj
∞Xj=0
= ne−mβXw∈V
Am
vw
Aj
∞Xj=0
(4.12)
(4.13)
Am
vwK β
w(wn)
POSITIVE EIGENVECTORS
15
Hence (4.13) follows becauseP∞
kn(v) = ξn
j=0 Aj
vwne−jβ < ∞. Set
v − e−βXw∈V
Avwξn
w.
We claim that kn ≥ 0. To see this observe first that it follows from (4.8) that
w(wn) ≤ K β
e−βPw∈V AvwK β
plies that
Avwξn
proving the claim. Since
it follows from (4.13) that
e−βXw∈V
mXl=0
e−lβXw∈V
∞Xl=0
e−lβXw∈V
when v ∈ Hv0, where hn(w) = P∞
hn(w) = Xw∈Hv0
Xw∈Hv0
(4.14) that
ξn
v =
v (wn). Combined with e−βPw∈V Avwξw ≤ ξv this im-
w ≤ min(cid:8)ξv, nK β
v (wn)(cid:9) = ξn
v ,
Al
vwkn(w) = ξn
Am+1
vw ξn
w,
v − e−(m+1)βXw∈V
Al
vwkn(w) = Xw∈Hv0
K β
v (w)hn(w)
(4.14)
l=0 e−lβAl
v0wkn(w). In particular, it follows from
K β
v0(w)hn(w) = ξn
v0 ≤ ξv0 = 1.
We can therefore define a positive linear functional µn of norm ≤ 1 on Aβ such that
µn(g) = Xw∈Hv0
g(w)hn(w).
µ(g) = lim
l→∞
µnl(g)
µ(cid:0)K β
v(cid:1) = ξv
By compactness of the unit ball in the dual space of Aβ there is a strictly increasing
sequence {nl} in N and a positive linear functional µ on Aβ such that
for all g ∈ Aβ. Since liml→∞ ξnl
v = ξv by (4.12), it follows from (4.14) that
(4.15)
for all v ∈ Hv0. For any fixed w ∈ Hv0 we have that ξnl
w = ξw for all large l,
and hence also that knl(w) = 0 for all large l when w ∈ Hv0\V∞. It follows that
liml→∞ µnl(1w) = 0 for all w ∈ Hv0\V∞, which shows that µ factors through Bβ.
It follows therefore from (4.15) and the Riesz representation theorem that there is
a Borel probability measure m on Xβ such that I(m)v = ξv for all v ∈ Hv0. By
Lemma 2.3 this implies that I(m) = ξ.
(cid:3)
When A is sub-stochastic, irreducible and β = 0, it follows from the abstract
characterisation given in Theorem 7.13 of [Wo] that the spectrum of Aβ is the
Martin compactification of the associated Markov chain, cf. Definition 7.17 in [Wo],
while Xβ is the Martin boundary when A is also row-finite (has finite range in the
sense of [Wo]). When A is not row-finite Xβ consists of the Martin boundary and
the vertexes V∞.
Let δx denote the Dirac measure at a point x ∈ Xβ. Then
I(δx)v = K β
v (x)
16
KLAUS THOMSEN
for all v ∈ V . Thus I takes the extreme points in M(Xβ) to elements of the form
v 7→ K β
v (x) for some x ∈ Xβ. By definition of Xβ, when x ∈ Xβ\V∞, there is a
sequence {wk} of distinct elements in Hv0 such that
lim
k→∞
K β
v (wk) = K β
v (x).
Recall now that under a continuous affine surjection between compact convex sets,
the pre-image of an extremal point is a closed face and therefore contains an extremal
point.
In this way we obtain from Theorem 4.6 the following description of the
extreme points in E(A, β)v0.
Corollary 4.7. Let β ∈ R. Assume that NW = ∅ or that NW is infinite. Assume
vwe−nβ < ∞ for all v, w ∈ V . Let ξ be an extremal point of
n=0 An
E(A, β)v0. There is a sequence {wk} of distinct elements in Hv0 such that
also that P∞
for all v ∈ V , or there is a vertex w ∈ Hv0 ∩ V∞ such that
ξv = lim
k→∞ P∞
P∞
ξv = P∞
P∞
n=0 An
n=0 An
vwk e−nβ
v0wke−nβ
n=0 An
n=0 An
vwe−nβ
v0we−nβ
(4.16)
(4.17)
for all v ∈ V .
Every infinite emitter w ∈ V∞ gives rise to an extremal element in E(A, β)v0 via
the formula (4.17) (in the transient case we consider here). This follows from the
uniqueness part in the Riesz decomposition lemma, Lemma 3.3. The question about
which sequences of vertexes {wk} give rise to extremal β-harmonic by the formula
(4.16) is generally much more difficult to answer. But at least we shall show below
that the vertexes in {wk} can be chosen to be the vertexes in an infinite path in G
without repeated vertexes. It follows from this that there are cases where A is not
row-finite and where there are no non-zero β-harmonic vectors, for any β, because
there are no paths in G of this sort. This can occur also when all our standing
assumptions hold and G is strongly connected. See Example 2.9 in [Ru].
Let ∂E(A, β)v0 be the extreme boundary of the Choquet simplex E(A, β)v0. By
identifying an element x ∈ Xβ with the corresponding Dirac measure δx ∈ M(Xβ),
we have an inclusion Xβ ⊆ M(Xβ), and we set
∂Xβ = Xβ ∩ I −1(E(A, β)v0),
which is a Borel subset of Xβ, cf. Theorem 4.1.11 in [BR]. Thus ∂Xβ consists of the
elements x ∈ Xβ for which v 7→ K β
v (x) is extremal in E(A, β)v0.
Lemma 4.8. V∞ ⊆ ∂Xβ and I is injective on ∂Xβ.
Proof. Let u ∈ V∞. Then
I(u)v = K β
where k(u) : V∞ → [0, ∞) is the function
k(u)(w) =((cid:0)P∞
0
n=0 An
v (u) = dk(u)v,
v0ue−nβ(cid:1)−1 when w = u,
otherwise.
POSITIVE EIGENVECTORS
17
It follows therefore from the uniqueness part of the statement in Lemma 3.3 that
V∞ ⊆ ∂Xβ and that I is injective on V∞. Furthermore, ∂E(A, β)v0 is the disjoint
union
∂E(A, β)v0 = ∂H(A, β)v0 ⊔ {I(u) : u ∈ V∞} ,
where ∂H(A, β)v0 is the set of extremal β-harmonic elements of E(A, β)v0. It suffices
therefore now to show that I is injective on ∂Xβ ∩ I −1(∂H(A, β)v0). Since I(u) is
not β-harmonic when u ∈ V∞, it follows that
Xβ ∩ I −1(∂H(A, β)v0) ⊆ Xβ\V∞.
As Bβ is generated by the image of the functions K β
that I is injective on Xβ\V∞ and therefore also on ∂Xβ ∩ I −1(∂H(A, β)v0).
v , v ∈ V , and 1u, u ∈ V∞, it follows
(cid:3)
Let M(∂Xβ) denote the subset of M(Xβ) consisting of the elements m of M(Xβ)
with m(∂Xβ) = 1.
Theorem 4.9. The map I : M(∂Xβ) → E(A, β)v0 is an affine bijection.
Proof. Surjectivity: Let ψ ∈ E(A, β)v0. Since E(A, β)v0 is a metrizable Choquet
simplex there is a unique Borel probability measure ν on ∂E(A, β)v0 such that
in the sense that
ψ =Z∂E(A,β)v0
a(ψ) =Z∂E(A,β)v0
z ν(z),
a(z) ν(z),
for all continuous and affine functions a on E(A, β)v0, cf. Theorem 4.1.15 in [BR].
By Lemma 4.8 we can define a Borel probability measure m on ∂Xβ such that
m(B) = ν(I(B)). Extending m to a Borel probability measure on Xβ such that
m (Xβ\∂Xβ) = 0, we can consider I(m). For each v ∈ V , the evaluation map
evv(φ) = φv is continuous and affine on E(A, β)v0 so we find that
I(m)v =Z∂Xβ
=Z∂E(A,β)v0
K β
v (y) dm(y) =Z∂E(A,β)v0
evv (z) dν(z) = evv(ψ) = ψv.
K β
v (I −1(z)) dν(z)
Injectivity: Assume m1, m2 ∈ M(∂Xβ) and that I(m1) = I(m2). Using Lemma
4.8 again it follows that there are Borel probability measures νi on ∂E(A, β)v0 such
that mi = νi ◦ I on ∂Xβ, i = 1, 2. Note that
I(mi)v =Z∂Xβ
=Z∂E(A,β)v0
K β
v (y) dmi(y) =Z∂E(A,β)v0
evv(z) dνi(z) = evv Z∂E(A,β)v0
z dν1(z) =Z∂E(A,β)v0
Z∂E(A,β)v0
z dν2(z).
for all v ∈ V, i = 1, 2. As I(m1) = I(m2) it follows that
K β
v (I −1(z)) dνi(z)
z dνi(z)!
(4.18)
Since ∂E(A, β)v0 is a Choquet simplex this implies that ν1 = ν2, and hence m1 = m2.
(cid:3)
18
KLAUS THOMSEN
u (x) = eβK β
Corollary 4.10. ∂Xβ\V∞ =(cid:8)x ∈ ∂Xβ : Pu∈V AvuK β
that I(x) is β-harmonic, i.e. Pu∈V AvuK β
Proof. Let x ∈ ∂Xβ\V∞. Then I(x) ∈ ∂E(A, β)v0 and it follows from the Riesz
decomposition, Lemma 3.3, that I(x) is either β-harmonic or equal to I(w) for some
w ∈ V∞. Since x /∈ V∞ and I is injective on M(∂Xβ) by Theorem 4.9, it follows
(cid:3)
It follows from Theorem 4.9 and Corollary 4.10 that the map I of Theorem 4.6
restricts to an affine bijection between the β-harmonic elements of E(A, β)v0 and
the Borel probability measures m on Xβ with the property that m (∂Xβ\V∞) = 1.
v (x) ∀v ∈ V(cid:9).
v (x) for all v ∈ V .
u (x) = eβK β
4.2. An improvement. The purpose with this section is the use results from the
theory of Markov chains to improve the description of the extremal β-harmonic
vectors given in Corollary 4.7.
Let P (V ) denote the set
P (V ) =(cid:8)(vi)∞
i=1 ∈ V N : Avivi+1 > 0, i = 1, 2, 3, · · ·(cid:9) .
Since there are no multiple edges in G the elements of P (V ) are the infinite paths
in G. We consider P (V ) as a complete metric space whose topology is generated by
the cylinder sets
C(v1v2 · · · vn) = {(xi)∞
i=1 ∈ P (V ) : xi = vi, i = 1, 2, · · · , n} .
Lemma 4.11. Let ψ be a non-zero β-harmonic vector. There is a Borel measure
mψ on P (V ) such that
mψ (C(v1v2 · · · vn)) = Av1v2Av2v3 · · · Avn−1vne−(n−1)βψvn
for every cylinder set C(v1v2 · · · vn).
Proof. The matrix
(4.19)
It follows then from Theorem 1.12 in [Wo] that there is a Borel
v Avwψw
Bvw = e−βψ−1
is stochastic.
probability measure mv on C(v) for each v ∈ V such that
mv (C(vv2 · · · vn)) = ψ−1
v Avv2Av2v3 · · · Avn−1vne−(n−1)βψvn.
(4.20)
Define mψ such that
mψ(B) =Xv∈V
Theorem 4.12. Assume that P∞
vwe−nβ < ∞ for all v, w ∈ V . Let ψ be
an extremal non-zero β-harmonic vector with ψv0 = 1. There is an infinite path
t = (ti)∞
i=1 in P (V ) such that t1 = v0, i 6= j ⇒ ti 6= tj, and
n=0 An
ψvmv (C(v) ∩ B) .
(cid:3)
for all v ∈ V .
ψv = lim
k→∞ P∞
P∞
n=0 An
n=0 An
vtk e−nβ
v0tk e−nβ
(4.21)
Proof. By construction the measure mv on C(v) given by (4.20) is the measure Prv
from Theorem 1.12 in [Wo] and the measure Px, with x = v, from (2.2) in [Sa],
coming from the stochastic matrix (4.19). Since ψ is extremal in E(A, β)v0 the
constant vector 1 is minimal harmonic for the stochastic matrix B restricted to Hv0.
POSITIVE EIGENVECTORS
19
It follows therefore from Theorem 5.1 in [Sa] that with respect to the measure mv0
almost all elements in
{(xi) ∈ P (V ) : x1 = v0}
have the property that
i→∞ P∞
P∞
for all v ∈ Hv0. Let w ∈ Hv0. Since P∞
v0xi
n=0 Bn
from Theorems 3.2 and 3.4 in [Wo] that the set
n=0 Bn
vxi
n=0 Bn
lim
= 1
ww = P∞
(4.22)
(4.23)
n=0 An
wwe−nβ < ∞, it follows
{(xi) ∈ P (V ) : x1 = v0, xi = w for at most finitely many i}
has full mv0-measure. Since Hv0 is countable it follows that so has
{(xi) ∈ P (V ) : x1 = v0, xi = w for at most finitely many i} ,
\w∈Hv0
which is the same set as {(xi) ∈ P (V ) : x1 = v0, limi→∞ xi = ∞}. Hence the set of
elements (xi) from (4.22) for which (4.23) holds for all v ∈ Hv0 and at the same have
the property that limi→∞ xi = ∞, is also a set of full mv0-measure. Since
it follows that there is an element (t′
n=0 Bn
vxi
n=0 Bn
v0xi
P∞
P∞
= ψ−1
v P∞
P∞
n=0 An
n=0 An
vxie−nβ
v0xie−nβ ,
i) from (4.22) such that limi→∞ t′
n=0 An
vt′
i
n=0 An
e−nβ
e−nβ = ψv
v0t′
i
lim
i→∞ P∞
P∞
i = ∞ and
(4.24)
for all v ∈ Hv0. It follows then from the uniqueness part of the statement in Lemma
2.3 that (4.24) holds for all v ∈ V . Finally, since limi→∞ t′
i = ∞ it is easy to
construct from t′ a path (ti) ∈ P (V ) such that i 6= j ⇒ ti 6= tj, and such that there
is a sequence n1 < n2 < n3 < · · · in N with ti = t′
ni for all i. The sequence (ti) has
the stated properties.
In summary we have found that in the transient case the extremal elements of
E(A, β)v0 consist of the vectors arising from an element w ∈ Hv0 ∩V∞ by the formula
(4.17) when they are not β-harmonic, and are given by an infinite path (ti) ∈ P (V )
such that i 6= j ⇒ ti 6= tj via the formula (4.21) when they are β-harmonic.
(cid:3)
References
[BR]
O. Bratteli and D.W. Robinson, Operator Algebras and Quantum Statistical Mechanics
I + II, Texts and Monographs in Physics, Springer Verlag, New York, Heidelberg,
Berlin, 1979 and 1981.
T.M. Carlsen and N. Larsen, Partial actions and KMS states on relative graph C ∗-
algebras, arXiv:1311.0912.
[CT]
J. Christensen and K. Thomsen, Finite digraphs and KMS states, arXives May, 2015.
[aHLRS] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C ∗-algebras of
[CL]
[K]
[P]
reducible graphs, Ergodic Th. & Dynam. Syst., to appear.
B.P. Kitchens, Symbolic Dynamics, Springer Verlag, New York, Berlin, Heidelberg,
1998.
W.E. Pruitt, Eigenvalues of non-negative matrices, Ann. Math. Statist. 35 (1964),
1797-1800.
20
[Ru]
[Sa]
[Th]
[V]
[Wo]
KLAUS THOMSEN
S. Ruette, On the Vere-Jones classification and existence of maximal measure for count-
able topological Markov chains, Pac. J. Math. 209 (2003), 365-380.
S.A. Sawyer, Martin boundary and random walks, Harmonic functions on trees and
buildings (New York, 1995), 17-44, Contemp. Math. 206, Amer. Math. Soc., Provi-
dence, RI, 1997.
K. Thomsen, KMS weights on graph C ∗-algebras, arXiv:1409.3702
D. Vere-Jones, Ergodic properties of non-negative matrices I, Pacific J. Math. 22 (1967),
361-386.
W. Woess, Denumerable Markov Chains, EMS Textbooks in Mathematics, 2009.
E-mail address: [email protected]
Institut for Matematik, Aarhus University, Ny Munkegade, 8000 Aarhus C, Den-
mark
|
1211.4814 | 4 | 1211 | 2016-03-14T10:32:44 | Generic orbits and type isolation in the Gurarij space | [
"math.FA",
"math.LO"
] | We study the question of when the space of embeddings of a separable Banach space $E$ into the separable Gurarij space $\mathbf G$ admits a generic orbit under the action of the linear isometry group of $\mathbf G$. The question is recast in model-theoretic terms, namely type isolation and the existence of prime models. We characterise isolated types over $E$ using tools from convex analysis. We show that if the set of isolated types over $E$ is dense, then a dense $G\_\delta$ orbit exists, and otherwise all orbits are meagre. We then study some (families of) examples with respect to this dichotomy. We also point out that the class of Gurarij spaces is the class of models of an $\aleph\_0$-categorical theory with quantifier elimination, and calculate the density character of the space of types over $E$, answering a question of Avil{\'e}s et al. | math.FA | math | GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
ITAÏ BEN YAACOV AND C. WARD HENSON
ABSTRACT. We study the question of when the space of embeddings of a separable Banach space E into
the separable Gurarij space G admits a generic orbit under the action of the linear isometry group of G.
The question is recast in model-theoretic terms, namely type isolation and the existence of prime models.
We characterise isolated types over E using tools from convex analysis. We show that if the set of isolated
types over E is dense, then a dense Gδ orbit exists, and otherwise all orbits are meagre. We then study some
(families of) examples with respect to this dichotomy. We also point out that the class of Gurarij spaces is the
class of models of an ℵ0-categorical theory with quantifier elimination, and calculate the density character of
the space of types over E, answering a question of Avilés et al.
CONTENTS
Isolated types over one-dimensional spaces
Introduction
1. Quantifier-free types in Banach spaces
2. The Gurarij space
3.
4. The Legendre-Fenchel transformation of 1-types
5. Characterising isolated types over arbitrary spaces
6. Existence and non-existence results
7. Counting types
References
1
2
7
12
13
16
18
20
21
6
1
0
2
r
a
M
4
1
]
.
A
F
h
t
a
m
[
4
v
4
1
8
4
.
1
1
2
1
:
v
i
X
r
a
INTRODUCTION
In 1966, Gurarij [Gur66] defined what came to be known as the (separable) Gurarij space, and proved
that it is almost isometrically unique. The isometric uniqueness of the Gurarij space was proved in 1976
by Lusky [Lus76]. In the same paper, Lusky points out that his arguments could be modified to prove
also the isometric uniqueness of the separable Gurarij space equipped with a distinguished smooth
unit vector (namely, a unit that admits a unique norming linear functional, see Definition 3.1). In other
words, if G denotes the separable Gurarij space, then the set of smooth unit vectors in G forms an orbit
under the action of the linear isometry group Aut(G). By Mazur [Maz33], this orbit is moreover a dense
Gδ subset of the unit sphere.
These facts are strongly reminiscent of familiar model theoretic phenomena, and, as we show in this
paper, are indeed special cases thereof. It was observed some time ago by the second author that the
uniqueness of the Gurarij space can be accounted for by it being the unique separable model of an ℵ0-
categorical theory, which moreover eliminates quantifiers. Similarly, the Gurarij space is atomic over
a vector if and only if the latter is smooth, so Lusky's second uniqueness result is a special case of the
uniqueness of the prime model.
These observations serve as a starting point for the present paper, whose goals are threefold:
• Make the observations above precise, and generalise them to uniqueness results over a subset
other than the empty set or a singleton -- in other words, we study uniqueness and primeness
of the Gurarij space over a subspace E of dimension possibly greater than one. As we shall see,
this requires us to (define and) characterise when types over E are isolated.
2010 Mathematics Subject Classification. 46B04 ; 03C30 ; 03C50 ; 03C98.
Key words and phrases. Gurarij space ; Banach space ; isolated type ; atomic model ; prime model ; group action ; generic orbit.
Research supported by the Institut Universitaire de France, by ANR project GruPoLoCo (ANR-11-JS01-008) and by a grant
from the Simons Foundation (#202251 to the second author).
Part of the work was carried out during the Universality and Homogeneity programme at the Hausdorff Institute of the
University of Bonn (Fall 2013).
Revision 2822 of 11th March 2016.
1
2
ITAÏ BEN YAACOV AND C. WARD HENSON
• Present in a manner accessible to non-logicians, and without making use of formal logic, some
tools and techniques of model theory: types, type spaces, type isolation, the Tarski-Vaught
Criterion, the Omitting Types Theorem, atomic models, and the primeness and uniqueness of
atomic models.
• Present to model theorists, who are familiar with the tools mentioned in the previous item in
the context of classical logic, how these tools adapt to the metric setting.
In Section 1 we recall the definition of (quantifier-free) types and type spaces over a Banach space E,
and study their properties. The topometric structure of the type space, a fundamental notion of metric
model theory, is defined there, as well as (topometrically) isolated types, which are among the main
objects of study of this paper.
In Section 2 we start studying Gurarij spaces. At the technical level, we define and study Gurarij (and
other) spaces that are atomic over a fixed separable parameter space E, and prove the Omitting Types
Theorem (Theorem 2.12). We prove appropriate generalisations of the homogeneity and universality
properties of the Gurarij space to homogeneity and universality over E. In particular, we show that the
prime Gurarij spaces over E (see Corollary 2.13) are those Gurarij spaces that are separable and atomic
over E. If a prime Gurarij space over E exists then it is unique, up to an isometric isomorphism over E,
and is denoted G[E]. When G[E] exists, the set of embeddings of E in G over which G is prime forms a
dense Gδ orbit among all embeddings of E; otherwise, all orbits are meagre. We also give the standard
model theoretic criterion for the existence of G[E], namely that the isolated types over E are dense.
While this (re-)development of model-theoretic tools is carried out in a fairly specific context, we
present arguments that would be valid in the general case; these are sometimes followed by separate
results that improve the general ones in the specific context of the Gurarij space. The few results that
do make explicit use of formal logic (essentially, Proposition 1.21 and Theorem 2.3) serve mostly as
parenthetical remarks required for the sake of completeness, and are not used in any way in other parts
of the paper.
At this point we move on to the questions of when G[E] exists and how to characterise isolated
types in a fashion suitable to the Banach space context. In Section 3 we consider the special case where
dim E = 1, giving a model-theoretic account of Lusky's result about smooth points in G. Before consid-
ering the general case, we introduce an essential tool in Section 4, namely the presentation of 1-types
as convex Katetov functions (as per [Ben14]), and the Legendre-Fenchel transformation of those. In
Section 5 we characterise isolated 1-types in terms of their Legendre-Fenchel conjugate, which allows
us to give in Section 6 some sufficient conditions for the existence of G[E] for finite-dimensional spaces
E (e.g., smooth, polyhedral, or of dimension ≤ 3 -- see Theorem 6.4), as well as examples when G[E]
does not exist. The question of a satisfactory necessary and sufficient condition on E for the existence
of G[E] remains an open problem.
We conclude in Section 7 with a "counting types" result, showing that the space of types over E is
metrically separable if and only if E is finite-dimensional and polyhedral. This allows us to answer a
question of Avilés et al. [ACC+11].
Throughout, E, F and so on denote vector spaces over the real numbers -- normed spaces, most of
the time, although the Legendre-Fenchel duality in Section 4 is stated for general locally convex spaces.
An embedding (or isomorphism, automorphism) of normed spaces is always isometric.
The topological dual of a normed space E will be denoted E∗. We shall often use the notation B(E) for
the closed unit ball of E and ∂B(E) for the unit sphere (which, regardless of topology, is the boundary
of B(E) in the sense of convex geometry), and similarly for E∗ instead of E.
1. QUANTIFIER-FREE TYPES IN BANACH SPACES
Before we start, let us state the following basic amalgamation result, which we shall use many times,
quite often implicitly.
Fact 1.1. For any three Banach spaces E0, F0 and F1, and isometric embeddings fi : E0 → Fi, there is a fourth
Banach space E1 and isometric embeddings gi : Fi → E1 such that g0 f0 = g1 f1.
Proof. Equip the direct sum F0 ⊕ F1 with the semi-norm kv + uk = infw∈E0 kv + f0wk + ku − f1wk,
divide by the kernel and complete.
(cid:4)
We can now define the fundamental objects of study of this section and, to a large extent, the entire
paper.
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
3
Definition 1.2. Let E be a Banach space and X a sequence of distinct symbols (indexed by an arbitrary
on E(X) that extend the norm on E, calling it the space of types in X over E. We denote members of SX(E)
by ξ, ζ and so on, and the corresponding semi-norms by k·kξ, k·kζ and so on.
set I) that we call variables. We let E(X) = E ⊕Lx∈X Rx, and define SX(E) to consist of all semi-norms
When X = {xi}i∈I we may also write E(I) = E ⊕Li∈I Rxi instead of E(X), and similarly SI (E),
whose members are called I-types.
Definition 1.3. Given a Banach space extension E ⊆ F and an I-sequence ¯a = {ai}i∈I ⊆ F, we define
the type of ¯a over E, in symbols ξ = tp( ¯a/E) ∈ SI(E), to be the semi-norm kb + ∑ λixikξ = kb + ∑ λiaik,
and say that ¯a realises ξ. When a sequence ¯b generates E, we may also write tp( ¯a/¯b) for tp( ¯a/E).
Conversely, given a type ξ ∈ SI (E), we define the Banach space generated by ξ, in symbols E[ξ],
as the space obtained from(cid:0)E(I), k·kξ(cid:1) by dividing by the kernel and completing, together with the
distinguished generators {xi}i∈I ⊆ E[ξ].
Remark 1.4. A model-theorist will recognise types as we define them here as quantifier-free types, which
do not, in general, capture "all the pertinent information". However, by Fact 1.1, they do capture a
maximal existential type. Moreover, it follows from Lemma 1.16 below (and more specifically, from
the assertion that π ¯x : S ¯x,y(0) → S ¯x(0) is open) that being an existentially closed Banach space is an
elementary property, so the theory of Banach spaces admits a model companion. Then Fact 1.1 can be
understood to say that the model companion eliminates quantifiers, so quantifier-free types and types
are in practice the same -- see Proposition 1.21. As we shall see later, the model companion is separably
categorical, and its unique separable model is G, the separable Gurarij space.
Definition 1.5. We equip SI(E) with a topological structure as well as with a metric structure, which
may be distinct. The topology on SI (E) is the least one in which, for every member x ∈ E(I), the map
x : ξ 7→ kxkξ is continuous. Given ξ, ζ ∈ SI(E), we define the distance d(ξ, ζ) to be the infimum, over all
F extending E and over all realisations ¯a and ¯b of ξ and ζ, respectively, of supi kai − bik.
It is fairly clear that:
(i) The distance on SI(E) refines the topology.
(ii) While the distance need not agree with the topology (we shall see that unless the parameter
space E is trivial and I is finite, they are in fact distinct), it is lower semi-continuous.
In other words, SI (E), equipped with this double structure, is a topometric space in the sense of [Ben08b].
Lemma 1.6. Let E, F be Banach spaces, I an index set, and consider tuples ¯a = (ai)i∈I ∈ EI, ¯b ∈ FI and ¯ε ∈ RI.
Also let R(I) denote the set of all I-tuples in which all but finitely many entries vanish. The following conditions
are equivalent.
(i) There exists a semi-norm k·k on E ⊕ F extending the respective norms of E and F, such that for each
i ∈ I one has kai − bik ≤ εi.
(ii) For all ¯r ∈ R(I), one has
Proof. One direction being trivial, we prove the other. For c + d ∈ E ⊕ F define
≤ ∑ riεi.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)∑ riai(cid:13)(cid:13) −(cid:13)(cid:13)∑ ribi(cid:13)(cid:13)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
¯r∈R(I)(cid:13)(cid:13)c − ∑ riai(cid:13)(cid:13) +(cid:13)(cid:13)d + ∑ ribi(cid:13)(cid:13) + ∑ riεi.
kc + dk′ = inf
This is easily checked to be a semi-norm, with kck′ ≤ kck for c ∈ E. Now, for c ∈ E and ¯r ∈ R(I) we
have
(cid:13)(cid:13)c − ∑ riai(cid:13)(cid:13) +(cid:13)(cid:13)∑ ribi(cid:13)(cid:13) + ∑ riεi ≥(cid:13)(cid:13)c − ∑ riai(cid:13)(cid:13) +(cid:13)(cid:13)∑ riai(cid:13)(cid:13) ≥ kck.
Therefore kck′ = kck, and similarly kdk′ = kdk for d ∈ F, concluding the proof.
Proposition 1.7. Let ξ, ζ ∈ SI(E), and let E(I)1 consist of all a + ∑ λixi ∈ E(I) (where a ∈ E and all but
finitely many of the λi vanish) such that ∑ λi = 1. Then
(cid:4)
d(ξ, ζ) = sup
x∈E(I)1(cid:12)(cid:12)(cid:12)kxkξ − kxkζ(cid:12)(cid:12)(cid:12) .
Moreover, the infimum in the definition of distance between types is attained.
Proof. Immediate from Lemma 1.6.
(cid:4)
4
ITAÏ BEN YAACOV AND C. WARD HENSON
Convention 1.8. When referring to the topological or metric structure of SI(E), we shall follow the
convention that unqualified terms from the vocabulary of general topology (open, compact, and so on)
apply to the topological structure, while terms specific to metric spaces (bounded, complete, and so on)
refer to the metric structure.
Excluded from this convention is the notion of isolation which will be defined in a manner that takes
into account both the topology and the distance.
While this convention may seem confusing at first, it is quite convenient, as in the following.
Lemma 1.9.
(i) The space SI (E) is Hausdorff, and every closed and bounded set thereof is compact.
(ii) The distance on SI(E) is lower semi-continuous. In particular, the closure of a bounded set is bounded.
(iii) Assume that I is finite, say I = n = {0, 1, . . . , n − 1} ∈ N. Then every bounded set in Sn(E) is
contained in an open bounded set. It follows that the space Sn(E) is locally compact, and that a compact
subset is necessarily (closed and) bounded.
(iv) A subset X ⊆ Sn(E) is closed if and only if its intersection with every compact set is compact.
Proof. For the first item, clearly SI (E) is Hausdorff. If X ⊆ SI (E) is bounded, then for every x ∈ E(I)
there exists Mx such that kxkξ ≤ Mx for all ξ ∈ X. We can therefore identify X with a subset of
Y = ∏x[0, Mx], and if X is closed in SI (E), then it is closed in Y and therefore compact.
The second item follows from Proposition 1.7, and the third is immediate.
For the fourth item, assume that X ⊆ Sn(E) is not closed, let ξ ∈ X r X and let U be a bounded
(cid:4)
neighbourhood of ξ, in which case U ∩ X is not compact.
When I is infinite, the distance on SI (E) is somewhat badly behaved: it can be infinite, and parts of
Lemma 1.9 may fail. Using it will become even more problematic for finer notions considered below,
such as type isolation. Henceforth we shall only consider the distance between types when I is finite.
Definition 1.10. Let m ≤ n. The variable restriction map π : Sn(E) → Sm(E) is the natural one induced
by the inclusion E(x0, . . . , xm−1) ⊆ E(x0, . . . , xn−1), namely kykπξ = kykξ for y ∈ E(x0, . . . , xm−1).
Lemma 1.11. Let m ≤ n, and let π : Sn(E) → Sm(E) denote the variable restriction map. Then for every
ξ ∈ Sn(E) and ζ ∈ Sm(E) we have d(πξ, ζ) = d(ξ, π−1ζ). Moreover there exists ρ ∈ π−1ζ such that
d(πξ, ζ) = d(ξ, ρ) and kxikρ = kxikξ for all m ≤ i < n.
In particular, the map π is metrically open.
Proof. The inequality d(πξ, ζ) ≤ d(ξ, π−1ζ) is immediate. For the opposite inequality, assume that
d(πξ, ζ) < r. By definition, there exist an extension F ⊇ E and realisations ¯a of πξ and ¯b of ζ in F such
that kai − bik < r for i < m. By Fact 1.1, possibly extending F, there is ¯c ∈ Fn−m such that tp( ¯a ¯c) = ξ.
Then ρ = tp(¯b ¯c/E) is as desired for both the main assertion and the moreover part. It follows that
πB(ξ, r) ⊇ B(πξ, r), so π is metrically open.
(cid:4)
Definition 1.12. We say that a type ξ ∈ Sn(E) is isolated if the distance and the topology agree at ξ, i.e.,
if every metric neighbourhood of ξ is also a topological one.
This is the definition of isolation in a topometric space, taking into account both the metric struc-
ture and the topological structure. Ordinary topological spaces can be viewed as topometric spaces by
equipping them with the discrete 0/1 distance, in which case the notion of isolation as defined here
coincides with the usual one.
Many results regarding ordinary topological spaces still hold, when translated correctly, with the
topometric definitions. For example, the fact that a dense set must contain all isolated points becomes
the following. Notice that in Lemma 1.18 below we prove that the set of isolated types is itself metrically
closed.
Lemma 1.13. Let E be a Banach space, D ⊆ Sn(E) a dense, metrically closed set. Then D contains all isolated
types.
Proof. If ξ is isolated, then all metric neighbourhoods of ξ, which are also topological neighbourhoods,
must intersect D.
(cid:4)
For reasons that will become clearer in Section 2, one of our main goals in this paper is to characterise
isolated types. We start with the easiest situation.
Proposition 1.14. Let 0 denote the trivial Banach space. Then every type in Sn(0) is isolated. In other words,
the distance on Sn(0) agrees with the topology.
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
5
Proof. Given N ∈ N, let XN ⊆ 0(n)1 be the finite set consisting of all ∑ λixi where ∑ λi = 1 and each
λi is of the form k
N . For ξ ∈ Sn(0), let Uξ,N be its neighbourhood consisting of all ζ such that
∀x ∈ XN
kxkξ − 1/N < kxkζ < kxkξ + 1/N.
This means in particular that kxikζ < kxikξ + 1 for all i < n, and now an easy calculation together with
Proposition 1.7 yields that there exists a constant C(ξ) such that for all N, Uξ,N is contained in the ball
of radius C(ξ)/N around ξ, which is what we had to show.
(cid:4)
This already allows us to construct the following useful tool of variable change in a type.
Definition 1.15.
(i) Given a linear map ϕ : E( ¯y) → E( ¯x) extending idE, we define a pull-back map
ϕ∗ : S ¯x(E) → S ¯y(E), or ϕ∗ : Sn(E) → Sm(E), by
kwkϕ∗ξ = kϕwkξ,
w ∈ E( ¯y).
For A ⊆ S ¯y(E) we define ϕ∗A = (ϕ∗)−1(A) ⊆ S ¯x(E).
(ii) Given a tuple ¯z in E( ¯x), of the same length as ¯y, define ϕ : E( ¯y) → E( ¯x) to be the unique linear
map extending idE and sending yi 7→ zi. We then write ξ↾ ¯z = ϕ∗ξ, so
(In this notation E, ¯x and ¯y are assumed to be known from context.)
(cid:13)(cid:13)a + ∑ λiyi(cid:13)(cid:13)ξ↾ ¯z =(cid:13)(cid:13)a + ∑ λizi(cid:13)(cid:13)ξ .
Lemma 1.16. For a fixed tuple ¯y ∈ E( ¯x)m, possibly with repetitions, the restriction map Sn(E) → Sm(E),
ξ 7→ ξ↾ ¯y, is continuous and Lipschitz (here n = ¯x). If ¯y are linearly independent over E, then this map is also
topologically and metrically open. Moreover, the metric openness is "Lipschitz" as well, in the sense that there
exists a constant C = C( ¯y) such that for all ξ and all r > 0 we have
B(ξ, r)↾ ¯y ⊇ B(ξ↾ ¯y, Cr).
Proof. Continuity and the Lipschitz condition are easy. We therefore assume that ¯y are linearly inde-
pendent over E, in which case we may also view them as formal unknowns. This gives rise to an
inclusion map E( ¯y) ֒→ E( ¯x), and ξ 7→ ξ↾ ¯y may be viewed as a map S ¯x(E) → S ¯y(E). In the special case
where ¯y generate E( ¯x) over E we have E( ¯x) = E( ¯y), and the Lipschitz map S ¯y(E) → S ¯x(E), ξ 7→ ξ↾ ¯x
is the inverse of ξ 7→ ξ↾ ¯y giving the moreover part. In the special case where yi = xi for i < m, the
moreover part follows from Lemma 1.11. In the general case, we may complete ¯y to a basis for E( ¯x)
over E, and the moreover part follows as a composition of the two special cases.
For topological openness, we proceed as follows. In the case where E = 0, this follows from metric
openness and Proposition 1.14. Let us consider now the case where E is finite-dimensional. We fix a
basis ¯b for E and a corresponding tuple of variables ¯w. We may then identify E( ¯x) with 0( ¯w, ¯x), and
thus ¯y with its image in 0( ¯w, ¯x). We already know that ·↾ ¯w, ¯y : S ¯w, ¯x(0) → S ¯w, ¯y(0) is open. In addition, we
have a commutative diagram
S ¯w, ¯x(0)
·↾ ¯w, ¯y
S ¯w, ¯y(0)
❄❄❄❄❄❄❄❄❄❄❄❄
·↾ ¯w
⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
·↾ ¯w
S ¯w(0)
and the map ·↾ ¯y : S ¯x(E) → S ¯y(E) is homeomorphic to the fibre of the horizontal arrow over tp(¯b) ∈
S ¯w(0), so it is open as well. The infinite-dimensional case follows from the finite-dimensional one, since
any basic open set in S ¯x(E) can be defined using finitely many parameters in E.
(cid:4)
We leave it to the reader to check that if ¯y are not linearly independent over E, then ·↾ ¯y is not metric-
ally open, and a fortiori not topologically so (consider, for example ·↾x,x : S1(0) → S2(0)).
Lemma 1.17. Let U ⊆ Sn(E) be open and r > 0. Then B(U, r) = {ξ : d(ξ, U) < r} ⊆ Sn(E) is open as well.
Proof. Let ¯x and ¯y be two n-tuples of variables. Let us identify Sn(E) with S ¯x(E), and let W ⊆ S ¯x, ¯y(E)
consist of all ξ such that kxi − yikξ < r for i < n. Then W is open, and by Lemma 1.16 so is V =
(cid:4)
(cid:0)W ∩ (·↾ ¯x)−1(U)(cid:1)↾ ¯y ⊆ S ¯y(E). Identifying S ¯y(E) with Sn(E) as well, V = B(U, r).
Lemma 1.18. Let E be a Banach space.
/
/
6
ITAÏ BEN YAACOV AND C. WARD HENSON
(i) A type in Sn(E) is isolated if and only if all its metric neighbourhoods have non empty interior.
(ii) The set of isolated types in Sn(E) is metrically closed.
Proof. The first assertion follows easily from Lemma 1.17, and the second from the first.
(cid:4)
Another basic operation one can consider on types is the restriction of parameters Sn(F) → Sn(E) when
E ⊆ F.
Lemma 1.19. Let E ⊆ F be an isometric inclusion of Banach spaces. Then the natural type restriction map
θ : Sn(F) → Sn(E) is continuous, closed, and satisfies θB(ξ, r) = B(θξ, r).
In particular, θ is both topologically and metrically a quotient map.
Proof. It is clear that θ is continuous. To see that it is closed we use Lemma 1.9. Indeed, since closed
sets are exactly those that intersect compact sets on compact sets, it will be enough to show that if
K ⊆ Sn(E) is compact, then so is θ−1K, which follows from the characterisation of compact sets as
closed and bounded.
It is clear that θB(ξ, r) ⊆ B(θξ, r). Conversely, if ζ0 ∈ Sn(E), then using Fact 1.1, there exists ζ ∈ θ−1ζ0
(cid:4)
with d(ξ, ζ) ≤ d(θξ, ζ0), which proves that θB(ξ, r) = B(θξ, r).
We also obtain that the theory T of (unit balls of) Banach spaces admits a model completion
T∗, namely a companion with quantifier elimination, whose types are exactly those defined above.
Moreover, as we shall see in the following section, the models of T∗ are exactly the Gurarij spaces. Since
this is somewhat of an aside with respect to the rest of this paper, we shall allow ourselves to be brief,
and assume that the reader is familiar with continuous first order logic (see [BU10, BBHU08]), and, for
the part regarding Banach spaces as unbounded metric structures, also with unbounded continuous
logic (see [Ben08a]).
Lemma 1.20. Let T be an inductive theory in continuous first order logic, and for n ∈ N let Sqf
n (T) denote the
space of quantifier-free types consistent with T, equipped with the natural logic topology. Assume first, that every
two models of T amalgamate over a common substructure, and second, that for every n, the variable restriction
map Sqf
n (T) is open. Then T admits a model completion, namely a companion that eliminates
quantifiers.
n+1(T) → Sqf
(In fact, an approximate amalgamation property for models of T over a common finitely generated substructure
suffices.)
Proof. Let ϕ( ¯x, y) be a quantifier-free formula, inducing a continuous function ϕ : Sqf
has compact range, by compactness of Sqf
tion map, and define ρ : Sqf
n (T) → R as the infimum over the fibre:
n+1(T)). Let π : Sqf
n+1(T) → Sqf
n+1(T) → R (which
n (T) denote the variable restric-
ρ(q) = inf(cid:8) ϕ(p) : πp = q(cid:9).
Since π is continuous (automatically) and open (by hypothesis), ρ is continuous as well, and can there-
fore be expressed as a uniform limit of ψn : Sqf
n (T) → R, where ψn( ¯x) are quantifier-free formulae, say
kρ − ψnk ≤ 2−n for all n. One can now express that sup ¯x ψn( ¯x) − infy ϕ( ¯x, y) ≤ 2−n for all n by a set
of sentences.
Let T∗ consist of T together with all sentences constructed as above, for all possible quantifier-free
formulae ϕ( ¯x, y). Then every existentially closed model of T is easily checked to be a model of T∗ (using
our amalgamation hypothesis), so T and T∗ are companions. Moreover, by induction on quantifiers,
every formula is equivalent modulo T∗ to a uniform limit of quantifier-free formulae, so T∗ eliminates
quantifiers.
(cid:4)
Proposition 1.21. Consider Banach spaces either as metric structures in unbounded continuous logic, or as
bounded metric structures via their closed unit balls, as explained, say, in [Ben09]. Then (in either approach) the
theory of the class of Banach spaces is inductive, and admits a model completion T∗ which is moreover complete
and ℵ0-categorical.
When the entire Banach space is viewed as a structure, then the types over a subspace are as per Definition 1.2
and Definition 1.3, and if one only considers the unit ball, then the space of I-types over B(E) is S≤1
SI(E) : kxikξ ≤ 1 for all i ∈ I(cid:9).
Proof. Let us consider the theory T of unit balls of Banach spaces. It is clearly inductive, and it is fairly
easy to check that the space of quantifier-free I-types over a unit ball B(E) is the space S≤1
I (E) defined in
the statement. By the moreover part of Lemma 1.11, variable restriction S≤1
n (E) is metrically
n+1(E) → S≤1
I (E) =(cid:8)ξ ∈
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
7
n+1(T) → Sqf
open. For E = 0 this implies in particular that S≤1
n (0) is topologically open, but this latter is
just Sqf
n (T). This, together with Fact 1.1, fulfils the hypotheses of Lemma 1.20. By quantifier
elimination, Sn(T∗) = Sqf
n (0), so in particular, S0(T∗) is a singleton, whereby T∗ is complete.
Finally, T∗ is ℵ0-categorical by the Ryll-Nardzewski Theorem and the fact that all types over the trivial
space are isolated (see [BU07]).
n+1(0) → S≤1
n (T) = S≤1
The case of Banach spaces as unbounded structures follows via the bi-interpretability of the whole
(cid:4)
Banach space with its unit ball.
2. THE GURARIJ SPACE
Definition 2.1. We recall from, say, Lusky [Lus76] that a Gurarij space is a Banach space G having
the property that for any ε > 0, finite-dimensional Banach space E ⊆ F, and isometric embedding
ϕ : E → G, there is a linear map ψ : F → G extending ϕ such that in addition, for all x ∈ F, (1 − ε)kxk ≤
kψxk ≤ (1 + ε)kxk.
Some authors add the requirement that a Gurarij space be separable, but from our point of view it
seems more elegant to consider separability as a separate property.
Lemma 2.2. Let F be a Banach space. Then the following are equivalent:
(i) The space F is a Gurarij space.
(ii) For every n, the set of realised types tp( ¯a/F), as ¯a varies over Fn, is dense in Sn(F).
(iii) Same as (ii) for n = 1.
Proof. (i) =⇒ (iii). Let U ⊆ S1(F) be open and ξ ∈ U. We may assume that U is defined by a finite set of
conditions of the form(cid:12)(cid:12)kai + rixk − 1(cid:12)(cid:12) < ε, where kai + rixkξ = 1. Let E ⊆ F be the subspace generated
by the ai, and let E′ = E + Rx be the extension of E generated by the restriction of ξ to E. By hypothesis,
there is a linear embedding ψ : E′ → F extending the identity such that (1 − ε)kyk < kψyk < (1 + ε)kyk
for all y ∈ E′, and in particular for y = ai + rix, so tp(ψx/F) ∈ U.
(iii) =⇒ (ii). We prove this by induction on n, the case n = 0 being tautologically true. For the
induction step, let ∅ 6= U ⊆ S ¯x,y(F) be open, and let V = U↾ ¯x ⊆ S ¯x(F). By Lemma 1.16, V is open, and
by the induction hypothesis there are ¯b ∈ Fn such that tp(¯b/F) ∈ V. Now, consider the map θ : Sy(F) →
S ¯x,y(F), sending tp(a/F) 7→ tp(¯b, a/F).
It is continuous (in fact, it is a topological embedding), so
∅ 6= θ−1U ⊆ S1(F) is open. By hypothesis, there is c ∈ F such that tp(c/F) ∈ θ−1U, i.e., such that
tp(¯b, c/F) ∈ U, as desired.
(ii) =⇒ (i). Let E ⊆ E′ be finite-dimensional, with E ⊆ F, and let ε > 0. Let ¯a be a basis for E, and
let ¯a, ¯b be a basis for E′, say ¯a = n and ¯b = m. For N ∈ N, let UN ⊆ Sm(F) be defined by the (finitely
many) conditions of the form k ∑ siai + ∑ rjxjk ∈ (1 − ε, 1 + ε), where si and rj are of the form k
N and
k ∑ siai + ∑ rjbjk ∈ (1 − ε, 1 + ε). By hypothesis there is a tuple ¯c ∈ Fm such that tp( ¯c/F) ∈ UN, and we
may define ψ : E′ → F being the identity on E and sending ¯b 7→ ¯c. For N big enough, it follows from
(cid:4)
the construction that if y ∈ E′, kyk = 1, then(cid:12)(cid:12)kψyk − 1(cid:12)(cid:12) < 2ε, which is good enough.
Model theorists may find the second and third conditions of Lemma 2.2 reminiscent of a topological
formulation of the Tarski-Vaught Criterion: a metrically closed subset A of a structure is an elementary
substructure if and only if the set of types over A realised in A is dense. Indeed,
Theorem 2.3. Let T∗ be the model completion of the theory of Banach spaces, as per Proposition 1.21. Then its
models are exactly the Gurarij spaces. In particular, since T∗ is ℵ0-categorical, there exists a unique separable
Gurarij space (up to isometric isomorphism).
Proof. Let E be a Banach space, and embed it in a model F (cid:15) T∗. Then, first, by quantifier elimination, E
is a model of T∗ if and only if E (cid:22) F. Second, by the topological Tarski-Vaught Criterion evoked above,
E (cid:22) F if and only if the set of types over E, in the sense of Th(F) = T∗, realised in E, is dense.
By Proposition 1.21 the space of types over E (in the sense of T∗ = Th(E)) is S≤1
n (E) as defined there.
By a dilation argument, the set of types realised in E is dense in S1(E) if and only if the set of types
realised in B(E) is dense in S≤1
1 (E), and we conclude by Lemma 2.2 (or, if one works with the whole
space as an unbounded structure, the same holds without the dilation argument).
(cid:4)
As mentioned in the introduction, the isometric uniqueness of the separable Gurarij space was ori-
ginally proved by Lusky [Lus76] using the Lazar-Lindenstrauss matrix representation of L1 pre-duals.
The same was recently re-proved by Kubi´s and Solecki [KS13] using more elementary methods. Upon
careful reading, their argument essentially consists of showing that the separable Gurarij space is the
8
ITAÏ BEN YAACOV AND C. WARD HENSON
Fraïssé limit of the class of finite-dimensional Banach spaces. The first author points this out in [Ben15]
as an application of a general development of Fraïssé theory for metric structures (yielding yet another
uniqueness proof). From this point onward we shall leave continuous logic aside, and work entirely
within the formalism of type spaces as introduced in Section 1. As we shall see, uniqueness and exist-
ence of the separable Gurarij space also follow as easy corollaries from later results that do not depend
explicitly on any form of formal logic (Corollary 2.7 and Lemma 2.11).
Definition 2.4. Let E ⊆ F be Banach spaces. We say that F is atomic over E if the type over E of every
finite tuple in F is isolated.
By Proposition 1.14, every Banach space is atomic over 0.
Theorem 2.5. Let E ⊆ F0 ⊆ F1 be Banach spaces with dim F0/E finite and F1 separable and atomic over E.
Also let G ⊇ E be a Gurarij space, and let ϕ : F0 → G be an isometric embedding extending idE. Then for every
ε > 0 there exist an isometric embedding ψ : F1 → G extending idE with kψ↾F0 − ϕk < ε.
In particular, any separable Banach space atomic over E embeds isometrically over E in any Gurarij space
containing E.
Proof. It is enough to prove this in the case where dim F1/F0 = 1. We may then choose a basis ¯a ∈ Fn+1
for F1 over E, such that in addition a0, . . . , an−1 generate F0. By hypothesis, ξ = tp( ¯a/E) ∈ Sn+1(E) is
isolated. Let ρ : Sn+1(G) → Sn+1(E) be the parameter restriction map, and let K = ρ−1(ξ), observing
that for any ε > 0, we have that B(K, ε) = ρ−1B(ξ, ε) is a neighbourhood of K. Given r > 0, we construct
a sequence of tuples ¯ck ∈ Gn+1, each of which realises a type in B(ξ, 2−kr), as follows.
1
For k = 0, we let V ⊆ Sn+1(G) be the set of semi-norms defined by kxi − ϕaik < r for i < n,
which is open and intersects K. Then V ∩ B(K, r)◦ 6= ∅ (where ·◦ denotes topological interior), and we
choose ¯c0 to realise some type there. Given ¯ck, we let Uk ⊆ Sn+1(G) be the set of semi-norms defined by
kxi − ck,ik < 2−kr for i ≤ n, which is again open and intersects K, and we choose ¯ck+1 to realise a type
in Uk ∩ B(K, 2−n−1r)◦.
We obtain a Cauchy sequence ( ¯ck) converging to some ¯c ∈ Gn+1, whose type tp( ¯c/E), being the
7→ ci is an
metric limit of tp( ¯ck/E), must be ξ. Then the linear map ψ : F1 → G that extends idE by ai
isometric embedding.
Finally, reading through our construction, we have kϕai − cik < 3r for all i < n, and choosing r small
(cid:4)
enough, kψ↾F0 − ϕk is as small as desired.
In particular, any two separable Gurarij spaces atomic over E embed in one another, but we can do
better.
Theorem 2.6. Let Gi be separable Gurarij spaces atomic over E for i = 0, 1, and let E ⊆ F ⊆ G0 with dim F/E
finite. Also let ϕ : F → G1 be an isometric embedding extending idE. Then for any ε > 0 there exists an isometric
isomorphism ψ : G0 ∼= G1 extending idE with kψ↾F − ϕk < ε.
In particular, any two separable Gurarij spaces atomic over E are isometrically isomorphic over E.
Proof. Follows from Theorem 2.5 by a back-and-forth argument. Indeed, by induction on n, using The-
orem 2.5, we construct finite dimensional subspaces E ⊆ Fn ⊆ G1 and E ⊆ F′
n ⊆ G0, as well as isometric
embeddings ϕn : Fn ֒→ G0 and ϕ′
n ֒→ G1 extending idE, such that:
n : F′
(i) F0 = F and ϕ0 = ϕ.
n ⊆ F′
(ii) Fn ⊆ Fn+1, F′
(iii) ϕn(Fn) ⊆ F′
n and ϕ′
(iv) kϕ′
n ϕn − idFn k + kϕn+1 ϕ′
n+1.
n(F′
n) ⊆ Fn+1.
n − idF′
n k < 2−n−1ε.
(v) S Fn = G1,S F′
n = G0.
Once the construction is complete, we have
kϕn − ϕn+1k ≤ kϕn − ϕn+1 ϕ′
n ϕnk + kϕn+1 ϕ′
n ϕn − ϕn+1k < 2−n−1εkxk.
It follows that the sequence (ϕn) converges in norm to an isometric embedding ψ : S Fn ֒→ G0, which
extends uniquely to G1 ֒→ G0. We obtain ψ′ : G0 ֒→ G1 as a limit of ϕ′
ψ′ = ψ−1, and kψ↾F − ϕk < ε, as desired.
n similarly. Then ψ extends idE,
(cid:4)
Since every Banach space is atomic over 0, we obtain the uniqueness and universality of the separable
Gurarij space.
Corollary 2.7. Every two separable Gurarij spaces are isometrically isomorphic, and every separable Banach
space embeds isometrically in any Gurarij space (separable or not).
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
9
We also obtain that the Gurarij space is approximately homogeneous.
Corollary 2.8. Let G be a separable Gurarij space, let F ⊆ G be finite-dimensional, and let ϕ : F → G be
an isometric embedding. Then there exists an isometric automorphism ψ ∈ Aut(G) such that kψ↾F − ϕk is
arbitrarily small.
Moreover, if E ⊆ F is such that G is atomic over E, and ϕ↾E = id, then we may require that ψ↾E = id as
well.
Notation 2.9. We shall denote by G the unique separable Gurarij space. Similarly, for a separable
Banach space E, we let G[E] denote the unique atomic separable Gurarij space over E, if such an exten-
sion of E exists. Observe that since all types over 0 are isolated, G = G[0].
Let E be a separable Banach space. Let Emb(E, G) denote the space of linear isometric embeddings
E ֒→ G, on which Aut(G) acts by composition. Say that ϕ ∈ Emb(E, G) is an atomic embedding if G is
atomic over ϕE.
Corollary 2.10. Let E be a separable Banach space. Equip Emb(E, G) and Aut(G) with the topology of point-
wise convergence (the strong operator topology).
(i) The space Emb(E, G) is Polish, the action Aut(G) y Emb(E, G) is continuous and all its orbits are
dense.
(ii) If G[E] exists, then the set of atomic embeddings ϕ ∈ Emb(E, G) is a dense Gδ orbit under this action.
(iii) If G[E] does not exist, then there are no atomic embeddings and all orbits are meagre.
Proof. The first item is easy and left to the reader (density is by Corollary 2.8).
It follows from Theorem 2.6 that the set Z ⊆ Emb(E, G) of atomic embeddings forms a single orbit
under Aut(G). By definition, Z 6= ∅ if and only if G[E] exists. Let In ⊆ Sn(E) denote the set of isolated
types. For r > 0, we know that B(In, r) is a neighbourhood of In, so there exists an open set Un,r such
that In ⊆ Un,r ⊆ B(In, r) (in fact one can show that B(In, r) is open, but we shall not require this). For
each ¯b ∈ Gn, we define V¯b,r ⊆ Emb(E, G) to consist of all ϕ such that tp(¯b/ϕE) ∈ ϕUn,r. It is easy to see
that since Un,r is open, so is V¯b,r. Since the set of isolated types is metrically closed, we have
Z = \n,¯b∈Gn,r>0
V¯b,r = \n,¯b∈Gn
0,k
V¯b,2−k,
where G0 ⊆ G is any countable dense subset. Thus, if Z 6= ∅ it is a dense Gδ orbit.
Consider now a non atomic embedding ψ ∈ Emb(E, G). Non atomicity means that G realises some
non isolated type in Sn(ψE), which we may write as ψξ, where ψ is applied to the parameters of ξ, and
ξ ∈ Sn(E) is non isolated. By Lemma 1.18, for r > 0 small enough, the closed metric ball B(ξ, r) is
(topologically) closed with empty interior. For ¯b ∈ Gn, let V¯b ⊆ Emb(E, G) consist of all ϕ such that
tp(¯b/ϕE) /∈ B(ϕξ, r). Reasoning as above, each V¯b is a dense open set, and the set of ϕ ∈ Emb(E, G)
such that G omits ϕξ is co-meagre. Since this set is also disjoint from the orbit of ψ, we are done.
(cid:4)
We now turn to a criterion for the existence of G[E]. Say that a type ξ ∈ SN(E) is a Gurarij type if E[ξ],
the generated space in the sense of Definition 1.3, is a Gurarij space. By an abuse of notation, we shall
use X = {xi}i∈N to denote both the set of variables and the set of distinguished generators of E[ξ].
Lemma 2.11. Let E be a separable Banach space. Then the set of Gurarij types over E is co-meagre in SN(E).
Moreover, there exists a dense Gδ set Z ⊆ SN(E) such that if some ξ ∈ Z generates F = E[ξ], then F is Gurarij
and the set of generators {xi}i∈N ⊆ F is dense.
In particular, the separable Gurarij space G exists.
Proof. For k ∈ N, let {Wk,m}m∈N be a countable basis for the topology of Sk(E), which exists since E is
separable.
Whenever I ⊆ J we shall use πJ,I : SJ(E) → SI(E) to denote the variable restriction map, namely,
N,k(Wk,m) :
tp(cid:0)(ai)i∈J/E(cid:1) 7→ tp(cid:0)(ai)i∈I/E(cid:1), which is an open map by Lemma 1.16. Thus, for example,(cid:8)π−1
k, m ∈ N(cid:9) is a basis of open sets for SN(E).
Let us fix some n ∈ N, introduce yet another formal variable y, and consider an open set
U ⊆ Sn+1(E) = Sx0,...,xn−1,y(E). Let F = Sn(E) r πn+1,n(U), which is closed. For each k ∈ N, let
ϕ∗
k : SN(E) → Sn+1(E) be the variable change map given by sending xi → xi for i < n and y → xk, as
per Definition 1.15, namely, tp(a0, a1, . . . /E) 7→ tp(a0, . . . , an−1, ak/E). We then define
N,n(F) ∪[k
eU = π−1
(ϕ∗
k )−1(U).
10
ITAÏ BEN YAACOV AND C. WARD HENSON
This is the union of an open and a closed set in a second-countable compact space, so it is Gδ.
First, we claim that if k ≥ n then πN,k(eU) = Sk(E). Indeed, we have a commutative diagram
Sk(E)
SN(E)
πN,n
πN,k
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
❄❄❄❄❄❄❄❄❄❄❄❄
ϕ∗
k
πk,n
❄❄❄❄❄❄❄❄❄❄❄❄
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
πn+1,n
Sn(E)
Sn+1(E)
Given any type ξ ∈ Sk(E) there are two possibilities.
N,k(ξ) ⊆ π−1
• If πk,n(ξ) ∈ F then π−1
• If πk,n(ξ) /∈ F then there exists ζ ∈ U such that πn+1,n(ζ) = πk,n(ξ) = χ, say. Amalgamating
E[ξ] with E[ζ] over E[χ], as per Fact 1.1, we obtain a normed space F ⊇ E, a tuple ¯b ∈ Fk such
that tp(¯b/E) = ξ, and c ∈ F such that tp(b0, . . . , bn−1, c) = ζ. Let ρ = tp(¯b, c, c, c, . . . /E) ∈
SN(E). Then ϕ∗
N,n(F) ⊆ eU, so ξ ∈ πN,k(eU).
k (ρ) = ζ, and again ξ = πN,k(ρ) ∈ πN,k(eU).
N,k(Wk,m) 6= ∅. Since
Lastly, we claim that the desired set is
Second, we claim that eU is dense in SN(E). Indeed, consider a basic open set π−1
πN,k(eU) = Sk(E) ⊇ Wk,m, we have ∅ 6= π−1
N,k(Wk) ∩eU.
Z = \n,m
^Wn+1,m ⊆ SN(E).
It is indeed a dense Gδ set. Let ξ ∈ Z, let F = E[ξ] be the generated space, and ¯b = (bi)i∈N be the
generators. Then it will suffice to show that for any open ∅ 6= U ⊆ S1(F) there exists k such that
tp(bk/F) ∈ U: this clearly implies that ¯b is dense in F, and by Lemma 2.2, F is Gurarij. Let us fix some
ρ ∈ U, which we may always write as tp(c/F) where c lies in some F′ ⊇ F.
We define θ : S1(F) → SXy(E) to be the map sending tp(a/F) to tp(¯b, a/E), as in the proof of
Lemma 1.16 (working over E instead of 0). This is a topological embedding, so we may write
U = θ−1(W) with W ⊆ SX,y(E) open and θ(ρ) = tp(¯b, c/E) ∈ W, and possibly shrinking W (and U),
we may assume that W is defined using only finitely many variables, say ¯x, y = x0, . . . , xn−1, y. In other
words, we may assume that W = π−1
Xy, ¯xy(Wn+1,m) for some m, so πXy, ¯xy ◦ θ(ρ) = tp(b0, . . . , bn−1, c/E) ∈
Wn+1,m. Thus ρ provides us with a witness that
πN,n(ξ) = tp(b0, . . . , bn−1/E) ∈ πn+1,n(Wn+1,n).
On the other hand, we have ξ ∈ Z ⊆ ^Wn+1,m, so there must exist some k such that
πXy, ¯xy ◦ θ(cid:0)tp(bk/F)(cid:1) = tp(b0, . . . , bn−1, bk/E) = ϕ∗
k (ξ) ∈ Wn+1,m.
We conclude that tp(bk/F) ∈ (πXy, ¯xy ◦ θ)−1(Wn+1,m) = U, and the proof is complete.
Theorem 2.12 (Omitting Types Theorem for Gurarij spaces). Let E be a separable Banach space, and assume
we are given, for each n ∈ N, a metrically open and topologically meagre set Xn ⊆ Sn(E). Then there exists a
separable Gurarij space G ⊇ E such that in addition, for every n, no type in Xn is realised in G (we then say that
G omits all Xn). Moreover, the set of Gurarij types that generate such spaces is co-meagre.
Proof. Let Z ⊆ SN(E) be the set produced by Lemma 2.11. For each n, let [N]n = {s ⊆ N : s = n}.
Any s ∈ [N]n can be enumerated uniquely as an increasing sequence {k0, . . . , kn−1}, and we then define
7→ xki for i < n. Then [s]∗ : SN(E) → Sn(E) is continuous and open, so
[s] : E(n) → E(N) by xi
[s]∗Xn ⊆ SN(E) is meagre. Since everything is countable,
(cid:4)
Z1 = Z r \n,s∈[N]n
[s]∗Xn
is co-meagre as well. All we need to show is that if ξ ∈ Z1 generates G, then G omits Xn. Indeed,
assume that some ξ ∈ Xn is realised in G, say by ¯a. Since Xn is metrically open, there exists r > 0 such
that B(ξ, r) ⊆ Xn. Since the sequence {xi} is dense in G, and G has no isolated points, there exists an
?
?
/
/
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
11
increasing sequence k0 < . . . < kn−1 such that kxkj − ajk < r. But then tp(x¯k/E) ∈ Xn, so ξ ∈ [¯k]∗Xn,
contradicting the choice of ξ and completing the proof.
(cid:4)
Say that a Gurarij space G is prime over a separable subspace E if it embeds isometrically over E in
every Gurarij space containing E.
Corollary 2.13 (Criterion for primeness over E). Let G be a Gurarij space, and let E ⊆ G be a separable
subspace. Then the following are equivalent:
(i) The space G is prime over E.
(ii) The space G is separable and atomic over E, namely, G = G[E].
Proof. If G = G[E], then it is prime over E by Theorem 2.5. For the other direction, assume that G is
prime over E. Since E is separable, it embeds (by Theorem 2.5) in a separable Gurarij space, so G must
be separable as well. Finally, assume toward a contradiction that G realises some non isolated type
ξ. By Lemma 1.18 there exists r > 0 such that the closed metric ball B(ξ, r) has empty interior. Since
the distance is lower semi-continuous, the closed metric ball is topologically closed, and is therefore
meagre, as is the open ball B(ξ, r). By Theorem 2.12, there exists a separable Gurarij space G ⊇ E that
omits B(ξ, r). Thus G cannot embed over E in G, a contradiction.
(cid:4)
Proposition 2.14. Let E be a separable Banach space. Then G[E] exists if and only if, for each n, the set of
isolated types in Sn(E) is dense.
Proof. For a given n, let In be the set of isolated types in Sn(E), and assume that it is dense. Then
B(In, r) contains a dense open set, and Tr>0 B(In, r) = I n is co-meagre. By Lemma 1.18 we have
In = I n, so Sn(E) r In is meagre and metrically open. Therefore, by Theorem 2.12, if In is dense for
all n, then an atomic separable Gurarij space over E exists.
Conversely, assume that G[E] exists. Then the set of n-types over E realised in G[E] is dense (by
(cid:4)
Lemma 2.2), and they are all isolated.
Model theorists will recognise Proposition 2.14 as the usual criterion for the existence of an atomic
model, and as such it is in no way particular to Banach spaces. In the specific context of Banach spaces,
however, it can be improved, yielding Theorem 2.16.
Lemma 2.15. For a type ξ ∈ S ¯x(E) the following are equivalent
(i) The type ξ is isolated.
(ii) The type ξ↾ ¯y is isolated for every ¯y ∈ E( ¯x)m (and every m).
(iii) The type ξ↾y is isolated for every y ∈ E( ¯x).
Proof. (i) =⇒ (ii). When ¯y are linearly independent over E, this follows from Lemma 1.16. Hence, for
the general case, it is enough to consider the situation where ¯y, of length m, extends the original tuple
of variables ¯x, of length n. For j < m let us write yj = aj + ∑i<n λijxi. Given r > 0, there exists by
hypothesis an open set U such that ξ ∈ U ⊆ B(ξ, r), and let V = (·↾ ¯x)−1U ⊆ S ¯y(E). Intersecting V with
the open sets defined by kyj − ∑i<n λijyi − ajk < r we obtain an open set V′ with ξ↾ ¯y ∈ V′ ⊆ B(ξ↾ ¯y, r′)
for some r′ = r′(r, ¯y) that goes to zero with r.
(ii) =⇒ (iii). Immediate.
(iii) =⇒ (i). We repeat the proof of Proposition 1.14 (in fact, that result is merely a special case of the
present one, alongside the fact that types in S1(0) are trivially isolated). Indeed, for each N there exists
by hypothesis a neighbourhood UN ∋ ξ consisting of ζ such that
∀y ∈ XN
d(ζ↾y, ξ↾y) < 1/N.
Using Proposition 1.7 we conclude as for Proposition 1.14.
(cid:4)
Theorem 2.16. The following are equivalent for a separable Banach space E:
(i) The space G[E] exists.
(ii) For each n, the set of isolated types in Sn(E) is dense.
(iii) The set of isolated types in S1(E) is dense.
Proof. We only need to show that if the set of isolated 1-types is dense, then G[E] exists. Indeed, pro-
ceeding as in the proof of Proposition 2.14 there exists a separable Gurarij space G ⊇ E that only realises
isolated 1-types over E. By Lemma 2.15, G is atomic over E.
(cid:4)
12
ITAÏ BEN YAACOV AND C. WARD HENSON
3. ISOLATED TYPES OVER ONE-DIMENSIONAL SPACES
Recall that one of the goals of this paper is to characterise isolated types over arbitrary E. We start
with the next-easiest case after E = 0, namely when dim E = 1. Even though this case will be fully sub-
sumed in the general case, it is technically significantly simpler and deserves some specific comments,
so we chose to treat it separately.
Definition 3.1. A norming linear functional for v ∈ E r {0} is a continuous linear functional λ ∈ E∗ such
that kλk = 1 and λv = kvk.
By the Hahn-Banach Theorem, a norming linear functional always exists. We say that v is smooth in
E if the norming linear functional is unique.
Proposition 3.2. Let E be a Banach space, and let v ∈ E r {0}. Then E is atomic over v if and only if v is
smooth in E.
Proof. By Lemma 2.15, we may assume that E = hv, ui and show that tp(u/v) is isolated if and only if
v is smooth in E. Assume first that for some s, ε > 0 and D ∈ R we have
It follows by the triangle inequality that
kv ± suk < kvk ± sD + sε.
kvk ± tD − tε ≤ kv ± tuk < kvk ± tD + tε,
0 < t ≤ s,
or equivalently,
(cid:12)(cid:12)k ± rv + uk − rkvk ∓ D(cid:12)(cid:12) < ε,
r ≥ s−1.
If v is smooth, let λ be the unique norming functional, and let D = λu. Then for any ε > 0 there
exists s as above. Then ξ = tp(u/v) satisfies the open condition kv ± sxk < kvk ± sD + sε, which in
can ensure that that the same holds for all r, yielding an open set ξ ∈ U ⊆ B(ξ, 3ε), showing that ξ is
isolated.
turn implies that(cid:12)(cid:12)krv − uk − krv − xk(cid:12)(cid:12) ≤ 2ε for all r ≥ s−1. Finitely many additional open conditions
form(cid:12)(cid:12)kriv + xk − kriv + uk(cid:12)(cid:12) < ε. We can construct a Banach space E′ generated by {v, w}, with kvk as
Conversely, if v is not smooth, then there are norming functionals λ±, where D− = λ−u < D+ =
λ+u. Any neighbourhood of ξ contains an open set U that is defined by finitely many conditions of the
in E, such that ζ = tp(w/v) ∈ U and v is smooth in E′, with unique norming functional being defined
by µw = D−. This means that for r big enough we have
krv + wk ≈ rkvk + D− ≤ krv + uk + D− − D+,
so d(ξ, ζ) ≥ D+ − D−. Therefore B(ξ, D+ − D−) is not a topological neighbourhood of ξ, and ξ is not
isolated.
(cid:4)
We provided a fairly elementary argument to the "only if" part of Proposition 3.2. The machinery
developed above provides us with a conceptually different argument, which in a sense we find prefer-
able. First, let us recall that by Mazur [Maz33, Satz 2], the set of smooth points in the unit sphere of a
separable Banach space is a dense Gδ. Assume now that E is atomic over v, and without loss of gener-
ality, say that kvk = 1, and let u ∈ G be smooth of norm one. By Theorem 2.5 there exists an isometric
embedding of E in G sending v to u, so v must be smooth.
We obtain the following result of Lusky [Lus76].
Corollary 3.3. The smooth points in the unit sphere of G form a single dense Gδ orbit under isometric auto-
morphisms.
Proof. Immediate from Proposition 3.2 and Theorem 2.6.
(cid:4)
Corollary 3.4. The distance strictly refines the topology on Sn(F) for every F 6= 0.
Proof. It follows from Lemma 1.19 that if E ⊆ F, and the topology and distance agreed on Sn(F), then
they would also agree on Sn(E), and every type in Sn(E) would be isolated. However, by Proposi-
tion 3.2, not all types over a 1-dimensional E are isolated.
(cid:4)
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
13
4. THE LEGENDRE-FENCHEL TRANSFORMATION OF 1-TYPES
In this section we recall and develop a few technical tools that will be used later in order to char-
acterise isolated types over arbitrary E. We start with the Legendre-Fenchel transformation. This being a
duality construction, it will be convenient for us to put E and its dual E∗ on a more equal footing.
Recall that a locally convex topology on a vector space (for our purposes, only over R), is a vector
space topology admitting a basis of convex neighbourhoods for 0. Examples of such topologies include
the norm topology on a normed space E, as well as the weak topology w on E and the weak∗ topology
w∗ on E∗. Moreover, (E, w) and (E∗, w∗) are one another's dual in the locally convex category, yielding
the desired symmetry.
Convention 4.1. In the rest of the paper, unless a more restrictive hypothesis is stated, E will denote a
locally convex topological vector space over R. Its topological dual E∗ is the space of continuous linear
functionals, always equipped with the weak∗ topology, namely the least topology in which v : λ 7→ λv is
continuous for each v ∈ E.
This applies in particular when E is a normed space: we may refer to the dual norm via the sets
B(E∗) and ∂B(E∗), or the corresponding properties kλk ≤ 1 and kλk = 1, but the topology on E∗ is
always taken to be the weak∗ topology, so B(E∗) is compact for any normed E.
The weak∗ topology is again locally convex, and the bi-dual E∗∗ is canonically identified, as a set,
with E (which would not always be true if for a normed space E we calculated E∗∗ with respect to the
dual norm on E∗). This induces the weak topology on E, which may be weaker than the original one, but
gives rise to the same dual E∗ (= E∗∗∗). The Hahn-Banach Theorem tells us that the weak topology on E
agrees with the original one when it comes to closed convex sets. In other words, both topologies give
rise to the same notion of a closed convex function (defined below), which, for our purposes, is good
enough.
Fact 4.2 (Hahn-Banach Theorem, see Brezis [Bre83]). A closed convex subset of E (a locally convex vector
space) is the intersection of the closed half-spaces that contain it, and is therefore weakly closed (a half-space of E
is a set of the form {v : λv ≤ r} where λ ∈ E∗ and r ∈ R).
Definition 4.3. Following Rockafellar [Roc70], a proper convex function on E is a convex function f : E →
R ∪ {∞} that is not identically ∞. It is closed if it is lower semi-continuous (equivalently, by Fact 4.2,
lower semi-continuous in the weak topology). We then define its domain dom f = {v ∈ E : f (v) < ∞}.
A closed convex function is either a proper one or one of the constant functions f = ±∞ (an improper one).
For an arbitrary function f : E → [−∞, +∞] we define f ∗ : E∗ → [−∞, +∞] by
f ∗(λ) = sup
v∈E
λv − f (v).
If f is closed convex we call f ∗ its conjugate.
Fact 4.4. For any f : E → [−∞, +∞], the function f ∗ is closed convex. If f is closed and convex, then f = f ∗∗
under the canonical identification E = E∗∗, and f ∗ is proper if and only if f is.
Moreover, if g is another closed convex function, then k f − gk = k f ∗ − g∗k, where k·k denotes the supremum
norm, possibly infinite, and we agree that ± ∞ ∓ ∞ = 0.
Proof. For the finite-dimensional case, see Rockafellar [Roc70, Section 12]. The general case is proved
essentially in the same fashion, using Fact 4.2. The moreover part is easy to check directly.
(cid:4)
Lemma 4.5. Let X ⊆ E and suppose f : X → R ∪ {∞} is not identically ∞. Assume moreover that whenever
x ∈ X can be expressed as a limit of convex combinations ∑i<ℓk tk,ixk,i, where xk,i ∈ X, we have f (x) ≤
fk,i f (xk,i). Then extending f by ∞ outside X, we have that f ∗ is a proper closed convex function
lim infk ∑i<ℓk
on E∗ and f = f ∗∗↾X.
Proof. Let epi f = (cid:8)(v, s) : f (v) ≤ s} ⊆ X × R ⊆ E × R, the epigraph of f , let Y = co(epi f ) ⊆ E × R
be the closed convex hull and define g(v) = inf{t : (v, t) ∈ Y} ∈ R ∪ {∞}. Then g is a closed proper
convex function, g ≤ f , and the hypothesis implies that g agrees with f on X. Now g∗ ≥ f ∗, so f ∗
is in particular proper (it is automatically closed and convex), and g = g∗∗ ≤ f ∗∗ ≤ f . Therefore
f ∗∗↾X = f .
(cid:4)
We recall that if X ⊆ E is convex, then ∂X is defined as the set of all v such that, for some affine line
L, v is one of two distinct boundary points of L ∩ X in L. The relative interior, sometimes denoted ri(X),
is defined as X r ∂X: the set of all v ∈ X such that for every affine line L going though v, either L ∩ X
is a single point or contains v in its interior relative to L. When X generates a finite-dimensional affine
subspace, this agrees with the usual topological notions as calculated in that space.
14
ITAÏ BEN YAACOV AND C. WARD HENSON
Lemma 4.6. Let E be finite-dimensional and let X ⊆ E be a compact convex subset. Let f : X → R be closed
and convex, and assume that f ↾∂X is continuous. Then f is continuous.
Proof. We need to show that if xn → x in X and f (xn) → α ∈ [−∞, ∞], then f (x) = α. Since f is closed,
f (x) ≤ α, and let us assume that f (x) < α. We may then assume that f (xn) > f (x) + ε for some ε > 0
and all n. Then the ray Rn = xn + R≥0(xn − x) intersects ∂X at a single point, call it yn = xn + sn(xn − x),
and we may further assume that yn → y, where y is necessarily also on the boundary. Notice that
snx + yn
sn + 1
, so by convexity f (yn) ≥ (sn + 1) f (xn) − sn f (x) > f (x) + (sn + 1)ε ≥ f (x) + ε. Since f
xn =
is continuous on the boundary we must have y 6= x. But then kyn − xk is bounded away from zero, so
sn → ∞, so f is unbounded on the boundary, even though ∂X is compact and f is continuous there, a
contradiction.
(cid:4)
The relevance of convex conjugation to our context comes from the following alternative character-
isation of 1-types over a normed space E, introduced in [Ben14] (see also Katetov [Kat88] and Uspenskij
[Usp08]). From now on, E denotes a normed space.
Definition 4.7. Let X be an arbitrary metric space. A Katetov function on X is a function f : X → R satis-
fying f (x) ≤ f (y) + d(x, y) and d(x, y) ≤ f (x) + f (y) for all x, y ∈ X. The space of Katetov functions on
X is denoted K(X). As with type spaces, we equip K(X) with a double structure, the topology of point-
wise convergence and the distance of uniform convergence (i.e., the supremum distance). With this
topology and distance, K(X) is a topometric space (that is to say that the distance refines the topology,
and is lower semi-continuous).
If X is a normed space, or a convex subset thereof, we let KC(X) denote the space of convex Katetov
functions on X, with the induced topometric structure.
Fact 4.8. Let ξ ∈ Sx(E) be a 1-type over a normed space E, and let fξ (a) = kx − akξ for a ∈ E. Then
(i) The map ξ 7→ fξ defines a bijection between S1(E) and KC(E), whose inverse is given by
kαx − akξ =(kak
α = 0
α fξ (a/α) α 6= 0.
(ii) This bijection is a topological homeomorphism and a metric isometry.
Proof. The first item is [Ben14, Lemma 1.2]. For the second, that the bijection is homeomorphic (in the
respective topologies of point-wise convergence) follows easily from the characterisation of the inverse,
while the isometry is exactly Proposition 1.7 for 1-types.
(cid:4)
Consequently, from now on we shall identify KC(E) with S1(E).
Fact 4.9. Let X ⊆ Y be metric spaces, and for f ∈ K(X) and y ∈ Y define
f (y) = inf
x∈X
f (x) + d(x, y).
Then f ∈ K(Y) extends f , and the induced embedding K(X) ⊆ K(Y) is isometric. When Y = E is a normed
space, X ⊆ E is convex and f ∈ KC(X), the extension f is convex as well, inducing an isometric embedding
KC(X) ⊆ KC(E).
Proof. The first assertion goes back to Katetov [Kat88], and the second is [Ben14, Lemma 1.3(i)].
(cid:4)
Question 4.10. If X ⊆ E is convex and compact (or totally bounded), then the topology and distance
agree on KC(X), and it follows that the inclusion KC(X) ⊆ KC(E) is also continuous, and therefore
homeomorphic (since the restriction map is always continuous). At the other extremity, if X = E, then
the inclusion is homeomorphic as well. What about general convex X ⊆ E?
A closed proper convex function f : E → R ∪ {∞} is essentially the same thing as a closed convex
function f : X → R, with convex domain X, such that lim infv→u f (v) = ∞ for all u ∈ X r X. Indeed,
we can get one from the other by restricting to the finite domain in one direction, or by extending by ∞
in the other. A special case of the second form is when X ⊆ E is closed and convex and f ∈ KC(X). If
X is merely convex, every f ∈ KC(X), being 1-Lipschitz, admits a unique extension to f ∈ KC(X), so
requiring X to be closed is not truly a constraint.
Definition 4.11. Let f : E → R ∪ {∞} be a proper closed convex function, and let λ ∈ E∗.
• We shall say that f (or more precisely,
f ∗(−λ) ≤ 0.
f ∗) satisfies the antipode inequality at λ if f ∗(λ) +
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
15
• It satisfies the antipode identity at λ if f ∗(λ) + f ∗(−λ) = 0.
Lemma 4.12. Let X ⊆ E be closed and convex and let f ∈ KC(X). Then
(i) The domain dom f ∗ contains B(E∗), and if λ ∈ dom f ∗ with kλk > 1, then
f ∗(λ) = sup
v∈∂X
λv − f (v).
In particular, if X = E (so ∂X = ∅), then dom f ∗ is exactly the closed unit ball.
(ii) If X is bounded and kλk = 1, then f ∗(λ) = supv∈∂X λv − f (v).
(iii) Let f ∈ KC(E) be as per Fact 4.9. Then
f ∗(λ) =( f ∗(λ) kλk ≤ 1
kλk > 1
∞
and
f (v) = sup
kλk≤1
λv − f ∗(λ).
In addition, if v /∈ X, then f (v) = supkλk=1 λv − f ∗(λ).
(iv) When X = E, we have f ∈ KC(E) = S1(E). For λ ∈ ∂B(E∗), the least possible value of a norm-
preserving extension of λ at a realisation of f is f ∗(λ).
(v) Let g : E → R ∪ {∞} be any closed proper convex function. Then g ∈ KC(E) if and only if dom g∗ =
B(E∗) and the antipode inequality g∗(λ) + g∗(−λ) ≤ 0 holds for all λ ∈ ∂B(E∗), or, equivalently, for
all λ ∈ B(E∗).
(vi) Assume g ∈ KC(E) is such that the antipode identity g∗(λ) + g∗(−λ) = 0 holds at some λ ∈ B(E∗).
Then g∗ is continuous at λ.
Proof. For (i), first let kλk ≤ 1, and let u ∈ X be fixed. Then for all v ∈ X we have f (u) + kuk ≥ kv −
uk − f (v) + kuk ≥ λv − f (v), whereby f ∗(λ) ≤ f (u) + kuk < ∞. Now let kλk > 1, say λw > kwk, and
assume that λ ∈ dom f ∗. Then for each v ∈ dom f , the ray {v + αw : α ≥ 0} cannot be contained in X
(or else f ∗(λ) = ∞) and therefore intersects the boundary, say at v′. In this case λv − f (v) ≤ λv′ − f (v′),
proving our assertion. When kλk = 1 but X is assumed to be bounded, for every v ∈ X we can find
v′ ∈ ∂X with λv − f (v) ≤ λv′ − f (v′) + ε for ε arbitrarily small, whence (ii).
For (iii), we already know that dom f ∗ is exactly the closed unit ball. In addition, f ≤ f implies
f ∗ ≥ f ∗, and if kλk ≤ 1, then for every v ∈ E and u ∈ X:
λv − f (v) = λv − inf
u∈X(cid:2) f (u) + kv − uk(cid:3) ≤ sup
u∈X
λu − f (u) = f ∗(λ),
whereby f ∗(λ) ≤ f ∗(λ). This gives us the first identity, and then Fact 4.4 gives the second one. Now
assume that v /∈ X, so by Fact 4.2 there exists µ ∈ ∂B(E∗) such that µ↾X < µv. For any λ ∈ B(E∗) there
exists α ≥ 0 such that kλ + αµk = 1. Then f ∗(λ + αµ) ≤ f ∗(λ) + αµv, or equivalently, λv − f ∗(λ) ≤
(λ + αµ)v − f ∗(λ + αµ), whence it follows that f (v) = supkλk=1 λv − f ∗(λ).
Item (iv) is immediate. For (v), we have already seen that if g ∈ KC(E), then dom g∗ = B(E∗), and
the previous item implies that g∗(λ) + g∗(−λ) ≤ 0 for λ ∈ ∂B(E∗). Conversely, assume that dom g∗ =
B(E∗) and the antipode inequality holds. Then g = g∗∗ is necessarily 1-Lipschitz, so dom g = E. For
distinct v, u ∈ E, let λ ∈ ∂B(E∗) norm v − u. Then
g(v) + g(u) ≥ λv − g∗(λ) − λu − g∗(−λ) ≥ λ(v − u) = kv − uk,
as desired.
For (vi), let λα → λ. Then g∗(λα) ≤ −g∗(−λα) by the antipode inequality and g∗(λ) = −g∗(−λ) by
hypothesis. Since g∗ is lower semi-continuous,
g∗(λ) ≤ lim inf g∗(λα) ≤ lim sup g∗(λα) ≤ lim sup −g∗(−λα)
= − lim inf g∗(−λα) ≤ −g∗(−λ) = g∗(λ).
Therefore lim g∗(λα) = g∗(λ), as desired.
Remark 4.13. Let F ⊆ E be normed spaces and let g ∈ KC(F). Since F is convex in E, there may be
some ambiguity about g∗, so let g∗
E denote
the conjugate of the extension by infinity to E∗. Also let g ∈ KC(E) denote the canonical extension of
g. Then g∗
F(λ↾F) for λ ∈ E∗, and by Lemma 4.12(iii), if kλk ≤ 1, then this is further equal to
g∗(λ). Therefore, in what interests us, this ambiguity can never lead to any form of confusion.
F denote the conjugate as a convex function on F and let g∗
E(λ) = g∗
(cid:4)
We obtain a characterisation of the realised types.
Lemma 4.14. Let E be a Banach space, f ∈ KC(E). Then the following are equivalent:
(i) The type f is realised in E, i.e., there exists v ∈ E such that f (x) = kx − vk for all x ∈ E.
16
ITAÏ BEN YAACOV AND C. WARD HENSON
(ii) The conjugate f ∗ satisfies the antipode identity throughout B(E∗).
(iii) The conjugate f ∗ satisfies the antipode identity at some λ ∈ B(E∗) r ∂B(E∗).
(iv) We have f ∗(0) = 0.
Proof. (i) =⇒ (ii). We have f ∗(λ) = λv.
(ii) =⇒ (iii). Clear.
(iii) =⇒ (iv). Assume f ∗(0) 6= 0. Then necessarily f ∗(0) < 0 and λ 6= 0. Let α = kλk < 1 and µ =
λ/α, so λ = αµ + (1 − α)0. By convexity, α f ∗(µ) + α f ∗(−µ) + 2(1 − α) f ∗(0) ≥ f ∗(λ) + f ∗(−λ) = 0,
contradicting the antipode inequality at µ.
(iv) =⇒ (i). We have 0 = f ∗(0) = − inf f . Since f is Katetov, any sequence vk such that f (vk) → 0
must be Cauchy, say with limit v. It follows that f (v) = 0 and consequently that f is realised by v. (cid:4)
5. CHARACTERISING ISOLATED TYPES OVER ARBITRARY SPACES
In [Ben14] a special kind of "well behaved" convex Katetov functions is distinguished. These will
play a crucial role here as well, and admit a natural characterisation in terms of their conjugates.
Definition 5.1. We say that a function f ∈ KC(E) is local if there are fk ∈ KC(Xk), where each Xk ⊆ E
is convex and compact, such that fk → f uniformly. The set of local functions in KC(E) was denoted in
[Ben14] by KC,0(E).
Lemma 5.2. Let E be a normed space, and let f ∈ KC(E). Then f is local if and only if f ∗ is continuous on
B(E∗).
Proof. First let X ⊆ E be compact and let g ∈ KC(X). If X ⊆ Si<n B(vi, r), then g∗(λ) − g∗(µ) ≤
2rkλ − µk + maxi(λ − µ)vi, whence it follows that g∗ is continuous on every bounded subset of E∗, and
in particular on B(E∗). Since a uniform limit of continuous functions is continuous, if f is local, then f ∗
is continuous on B(E∗).
Conversely, assume that f ∗ is continuous on B(E∗). Since R is second countable, there exists a sep-
arable subspace F ⊆ E such that, for λ ∈ B(E∗), the value of f ∗(λ) only depends on λ↾F. In this case,
the map f ′ : B(F∗) → R, µ 7→ f ∗(µ′), where µ′ is any norm-preserving extension of µ, is continuous. By
Fact 4.4 there exists a proper closed convex function g : F → R ∪ {∞} such that g∗ = f ′. By Lemma 4.12
we then have g ∈ KC(F) and by Remark 4.13 we have f ∗ = g∗, so f = g. We may therefore assume that
E is separable, and choose an increasing sequence of compact convex subsets Xk ⊆ E such thatS Xk
is dense in E (take closed balls of increasing radius, of sub-spaces of increasing finite dimension). For
each k let fk = f ↾Xk. Then fk ց f point-wise, and for λ ∈ B(E∗) we have
f ∗
λv − fk(v) = sup
k (λ),
k
λv − f (v) = sup
v,k
f ∗(λ) = sup
v
k ր f ∗ point-wise on B(E∗). Since each f ∗
i.e., f ∗
is compact, this implies that f ∗
k is lower semi-continuous, f ∗ is continuous, and B(E∗)
k → f ∗ uniformly on B(E∗), whereby fk → f uniformly, and f is local. (cid:4)
A second ingredient is the following.
Definition 5.3. Let E be a normed space and let r ≥ 0. We say that f ∈ KC(E) is ∂-r-extreme (where
∂ could be pronounced boundary) if for every g ∈ KC(E) whenever g ≤ f (i.e., g∗ ≥ f ∗) we have
g∗ ≤ f ∗ + r on ∂B(E∗). If r = 0 we omit it and say that f is ∂-extreme.
Notice that Definition 5.3 would remain unchanged if we replaced the requirement that g∗ ≥ f ∗ on
B(E∗) with g∗ ≥ f ∗ on ∂B(E∗). Indeed, assume Definition 5.3 holds of f and g∗ ≥ f ∗ on ∂B(E∗). Then
h′ = max( f ∗, g∗) agrees with g∗ on ∂B(E∗) and is therefore of the form h∗ for some h ∈ KC(E), by
Lemma 4.12(v), so Definition 5.3 applies to h and yields g∗ = h∗ ≤ f ∗ + r on ∂B(E∗).
Lemma 5.4. Let f , g ∈ KC(E) with f ∂-r-extreme and g ≤ f . Then g is ∂-r-extreme as well.
Assume furthermore that f = gf ↾X for some convex X ⊆ E. Then outside X we have g ≥ f − r.
Proof. By hypothesis we have g∗ ≥ f ∗, and we clearly obtain the first assertion, as well as g∗(λ) ≤
f ∗(λ) + r for kλk = 1. By Lemma 4.12(iii), if v /∈ X, then
f (v) = sup
kλk=1
λv − f ∗(λ) ≤ sup
kλk=1
λv − g∗(λ) + r ≤ g(v) + r,
as claimed.
Lemma 5.5. Let E be a normed space, let f ∈ KC(E) be ∂-r-extreme, and let δ > 0. Then f + δ is ∂-(r + 2δ)-
extreme.
(cid:4)
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
17
Proof. Assume not, so let g ∈ KC(E) satisfy g ≤ f + δ (i.e., g∗ ≥ f ∗ − δ), and let λ ∈ ∂B(E∗) be such that
g∗(λ) > f ∗(λ) + r + δ. By Fact 4.4 there exists a closed convex h on E such that h∗ = max(cid:0) f ∗, (g∗ − δ)(cid:1).
For µ ∈ B(E∗) we have f ∗(µ) + f ∗(−µ) ≤ 0, g∗(µ) − δ + g∗(−µ) − δ ≤ −2δ < 0 and f ∗(µ) + g∗(−µ) −
δ ≤ g∗(µ) + f ∗(−µ) ≤ 0. Thus h∗(µ) + h∗(−µ) ≤ 0, and since the domain of h∗ is the unit ball,
h ∈ KC(E). Now, h∗ ≥ f ∗ implies h ≤ f , while on the other hand h∗(λ) ≥ g∗(λ) − δ > f ∗(λ) + r,
witnessing that f is not ∂-r-extreme.
(cid:4)
Lemma 5.6. Let E be a normed space, let f ∈ KC(E) be ∂-r-extreme and local, and let r′ > r. Then f admits a
neighbourhood f ∈ W ⊆ KC(E) such that diam W < r′ and every g ∈ W is ∂-r′-extreme.
Proof. Let δ > 0 be small enough. By locality, there exists a compact convex set X ⊆ E such that
Katetov functions are 1-Lipschitz, W is open, and we claim that it is as desired.
If g ∈ W, then g < f ′ < f + δ. By Lemma 5.5, f + δ is ∂-(r + 2δ)-extreme. By Lemma 5.4 so are f ′
f + δ > f ′ = gf ↾X. Let W ⊆ KC(E) consist of all g such that f − g < δ on X. Since X is compact and
and g, and since f ′ = gf ′↾X we have g ≥ f ′ − r − 2δ ≥ f − r − 2δ outside X. Since g ≥ f − δ inside X,
we conclude that f − r − 2δ ≤ g < f + δ throughout, which is enough.
(cid:4)
Theorem 5.7. Let f ∈ KC(E). Then f is isolated if and only if it is both local and ∂-extreme.
Proof. One direction follows directly from Lemma 5.6, so let us assume that f is isolated. We can then
construct a sequence of neighbourhoods Wk of f such that diam Wk → 0, each defined using finitely
many parameters. We let Xk be the (compact) convex hull of these parameters and fk = f ↾Xk. Then
fk ∈ Wk, so fk → f uniformly and f is local.
It remains to show that f is ∂-extreme, so let g ∈ KC(E) satisfy g ≤ f . Let W be a neighbourhood
of f of small diameter, say g ∈ W if and only if g(vi) − f (vi) < ε for some vi, i < n. We know that
f (v) = supkλk≤1 λv − f ∗(λ), and since a closed convex function in dimension one is continuous on
its domain, we have in fact f (v) = supkλk<1 λv − f ∗(λ). Therefore, for each i < n we may choose
λi ∈ E∗
kλik < 1 for each i, g′ agrees with g outside some ball. On the other hand, we have g′(vi) > f (vi) − ε
and g′ ≤ f , so g′ ∈ W. Therefore f − g ≤ diam W outside some ball. Since diam W can be taken
arbitrarily small, limkxk→∞ f (x) − g(x) = 0. It follows that f ∗ = g∗ on ∂B(E∗).
(cid:4)
Before stating a few more corollaries, let us recall a few definitions and facts.
<1 such that f (vi) − ε < λivi − f ∗(λi) ≤ f (vi). Let g′ = max(cid:18)g, maxi<n(cid:0)λi − f ∗(λi)(cid:1)(cid:19). Since
Definition 5.8. Let E be a locally convex space and let X ⊆ E be convex.
(i) A convex subset F ⊆ X is called a face of X if a member of F cannot be expressed as a proper
convex combination of two points in X that are not both in F. A proper face, i.e., a face F 6= X, is
always contained in the relative boundary ∂X.
(ii) A face consisting of a single point is called an extreme point. We shall denote the set of extreme
points of X by E (X). We shall also denote by E0(X) the set of v ∈ E (X) such that λv = sup λ↾X
for some λ ∈ E∗ r {0}.
By the Krein-Milman Theorem [Bou81, Chapitre II.7, Théorème 1], if X is compact and convex, then
{u ∈ X : λu > r}, where λ ∈ E∗ and λv > r ∈ R, forms a basis of neighbourhoods for v in X. Since
every such neighbourhood contains a member of E0(X) (any extreme point of {u ∈ X : λu = sup λ↾X}
X = co(cid:0)E (X)(cid:1). In addition, by [Bou81, Chapitre II.7, Proposition 2], if v ∈ E (X), then the family of sets
will do), we have E (X) ⊆ E0(X). In the special case where X = B(E∗), the set E0(cid:0)B(E∗)(cid:1) consists
exactly of those λ ∈ E(cid:0)B(E∗)(cid:1) for which some vector v 6= 0 is normed by λ.
In particular, the antipode identity f ∗(λ) + f ∗(−λ) = 0 holds at every λ ∈ E(cid:0)B(E∗)(cid:1).
Corollary 5.9. Let E be a Banach space, f ∈ KC(E) isolated. Let Fv = {λ ∈ B(E∗) : λv = 1} where kvk = 1.
Then f ∗↾Fv is the greatest closed convex function less than λ 7→ − f ∗(λ), i.e., f ∗(λ) = supu∈E infµ∈Fv (λ −
µ)u − f ∗(−µ) for λ ∈ Fv.
Proof. For u ∈ E and α > 0 define
hu(λ) = inf
µ∈Fv
(λ − µ)u − f ∗(−µ),
hu,α(λ) = α(λv − 1) + hu(λ).
Since Fv is closed, applying Fact 4.4 we have f ∗(λ) = supu∈E infµ∈Fv (λ − µ)u + f ∗(µ) ≤ supu∈E hu(λ).
For the converse inequality it will suffice to show that f ∗ ≥ hu on Fv. Notice that hu is linear and
continuous, and in addition, for every λ ∈ Fv we have h(λ) + f ∗(−λ) ≤ 0. Fixing ε > 0, there exists an
18
ITAÏ BEN YAACOV AND C. WARD HENSON
open set V ⊇ Fv such that h(λ) + f ∗(−λ) < ε for all λ ∈ V ∩ B(E∗). By compactness of B(E∗) r V, we
know that λ 7→ λv is bounded there below some r < 1. We may therefore assume that V = {λ : λv > r},
and that r > 0, so V ∩ −V = ∅. For α big enough we have hu,α(λ) ≤ inf f ∗ for all λ ∈ B(E∗) r V.
Having fixed such α, for λ ∈ B(E∗) ∩ V we have hu,α(λ) − ε + f ∗(−λ) ≤ hu(λ) + f ∗(−λ) − ε ≤ 0. It
g ∈ KC(E). But then g ≤ f , so g∗ = f ∗ on ∂B(E∗) and in particular on Fv. Thus f ∗ ≥ hu,α − ε = hu − ε
on Fv. Since ε was arbitrary, f ∗ ≥ hu on Fv, as desired.
follows that max(cid:0) f ∗, (hu,α − ε)(cid:1) satisfies the antipode inequality, and is therefore of the form g∗ for some
It follows that the antipode identity holds on E0(cid:0)B(E∗)(cid:1). By continuity of f ∗, it holds throughout
E(cid:0)B(E∗)(cid:1).
Thus isolated types satisfy the antipode identity at some boundary points (and recall that by
Lemma 4.14, the antipode identity at a non-boundary point amounts to the type being realised). When
dim E = 1 we recover the characterisation of isolated types given in Section 3.
Corollary 5.10. Assume dim E = 1 and let ∂B(E∗) = {±λ}. Then a type f ∈ KC(E) is isolated if and only if
it satisfies the antipode identity at λ. By Lemma 4.12(iv), this is equivalent to: any vector v ∈ E is smooth in the
generated extension E[ f ].
(cid:4)
Proof. If f is isolated then the antipode identity holds at λ by Corollary 5.9. Conversely, if the antipode
identity holds at ±λ, namely, on the entire boundary, then f is ∂-extreme. Since, in dimension one,
every closed convex function is continuous, f is isolated.
(cid:4)
6. EXISTENCE AND NON-EXISTENCE RESULTS
This section consists of examples of various cases where densely many isolated types are known to
exist or not to exist. We do not have a full characterisation of separable spaces E such that isolated types
over E are dense.
Definition 6.1. Let E be a Banach space. We say that f ∈ KC(E) is strongly ∂-extreme if
(i) The antipode identity f ∗(λ) + f ∗(−λ) = 0 holds on E (B(E∗)).
(ii) The values of f ∗ on ∂B(E∗) are maximal given the values of f ∗ at the extreme points and the
fact that f ∗ is convex.
Clearly, a strongly ∂-extreme type is in particular ∂-extreme.
Lemma 6.2. Let E be a normed space.
(i) Then the set of strongly ∂-extreme f ∈ KC(E) is dense.
(ii) Assume moreover that dim E < ∞ and f ∗↾∂B(E∗) is continuous whenever f ∈ KC(E) is strongly
∂-extreme. Then the isolated types over E are dense and G[E] exists.
Proof. In order to show that the strongly ∂-extreme types are dense, let U be open and f ∈ U. We may
assume that U consists of all g ∈ KC(E) such that g(vi) − f (vi) < ε for i < n. Let X = co(vi : i < n) ⊆
E, and replacing f withgf ↾X we may assume that f is local, i.e., that f ∗ is continuous. For each i < n fix
λi ∈ B(E∗) such that f (vi) + f ∗(λi) < λivi + ε, and we may require kλik < 1. Define
f (λ) =( f ∗(λ)− f ∗(−λ)
2
f ∗(λi)
λ ∈ E (B(E∗)),
λ = λi.
Then f satisfies the hypotheses of Lemma 4.5, and is therefore the restriction of g∗ : B(E∗) → R for some
g ∈ KC(E). We have g∗ ≥ f ∗, i.e., g ≤ f , and g(vi) ≥ λivi − g∗(λi) = λivi − f ∗(λi) > f (vi) − ε, so
g ∈ U. Also, g∗ is strongly ∂-extreme by construction.
Assume now that E is as in the second item. By Lemma 4.6, if f ∈ KC(E) is strongly ∂-extreme
then f ∗ is continuous, so f is isolated by Theorem 5.7. It follows that the isolated types are dense in
KC(E) = S1(E), so G[E] exists by Theorem 2.16.
(cid:4)
Remark 6.3. Over a reflexive Banach space E, every isolated type is strongly ∂-extreme, by Corollary 5.9
and the fact that every λ ∈ ∂B(E∗) norms some v 6= 0.
Theorem 6.4. Let E be a normed space of finite dimension. Then the isolated types in KC(E) are dense if either
of the following holds:
(i) Every face of B(E∗) of dimension at most dim E − 2 is a simplex. This holds in particular whenever
dim E ≤ 3. Special cases of this include:
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
19
(a) Every proper face of B(E∗) is a simplex. This holds in particular whenever dim E ≤ 2. In this case
G is atomic over E ⊆ G if and only if every λ ∈ E∗ admits a unique extension of the same norm
to G.
(b) The space E is smooth, i.e., every v ∈ E is smooth. In this case G is atomic over E ⊆ G if and only
if every v ∈ E is smooth in G.
(ii) The space E is polyhedral.
Proof. In each case we apply Lemma 6.2, so we assume throughout that f ∈ KC(E) is strongly ∂-extreme.
In the first case, let n = dim E and let X ⊆ ∂B(E∗) be the union of all faces of ∂B(E∗) that are
simplexes. Then ∂B(E∗) r X consists of the relative interiors of some faces of dimension n − 1, so X
is closed. On each face that is a simplex, f ∗ is affine, and since it satisfies the antipode identity at the
extreme points in satisfies it throughout X. By Lemma 4.12(vi), f ∗, as a function on B(E∗), is continuous
at every λ ∈ X. It follows by Lemma 4.6 that the restriction of f ∗ to any face (be it a simplex or not) is
continuous. Thus, if λk ∈ ∂B(E∗), λk → λ, then either λ belongs to the relative interior of some face of
dimension n − 1 or else belongs to X, and in any case f ∗(λk) → f ∗(λ).
In the first special case X = ∂B(E∗) and the characterisation of G[E] follows from Lemma 4.12(iv).
The second special case is clear.
If E is polyhedral, one prove by induction on m, using Lemma 4.6, that f ∗ is continuous on each
face of dimension m and therefore on the union of all such faces. For m = n − 1, this means that f ∗ is
continuous on ∂B(E∗).
(cid:4)
Notice that Proposition 3.2 fits all cases mentioned in Theorem 6.4. More generally, the case where
E = ℓ∞(n), namely Rn equipped with the supremum norm, fits cases (i)(a) and (ii). We next use this
to show that there are infinitely many distinct orbits in the action of Aut(G) y ∂B(G) (previously we
only knew there were at least two, since there are both smooth and non-smooth points).
simplex of dimension n. Consequently, the action Aut(G) y ∂B(G) admits infinitely many distinct orbits.
Corollary 6.5. For each n ∈ N there exists v ∈ ∂B(G) such that N(v) = (cid:8)λ ∈ ∂B(G∗) : λv = 1(cid:9) is a
Proof. By Theorem 6.4(i)(a), G(cid:2)ℓ∞(n + 1))(cid:3) exists, and letting v = (1, 1, . . .) ∈ ℓ(n + 1), the set N(v) is a
Say that two convex sets are isomorphic if there exists a homeomorphism between them respecting
convex combinations. Then the isomorphism type of N(v) is invariant under the action of Aut(G),
whence the existence of distinct orbits.
(cid:4)
simplex of dimension n.
Let us now give some examples in which the conclusion of Theorem 6.4 fails. This will show that
while Theorem 6.4 is not necessarily optimal, none of the hypotheses can be simply done away with.
Example 6.6. We construct an example of a space E of dimension 4, such that G[E] does not exist. Let
B0 = {±1}4 ⊆ R4, B1 = {(x, y, 0, 0) : x2 + y2 = 2 and x, y 6= ±1} and B = co(B0 ∪ B1). Then B is a
compact symmetric convex neighbourhood of 0, so we may take E∗ = R4 with B(E∗) = B. Moreover,
the set of extreme points in B is exactly B0 ∪ B1. It follows by Corollary 5.9 that if f ∈ KC(E) is isolated,
then f ∗ satisfies the antipode identity also at the four points (±1, ±1, 0, 0) (which are not extreme).
Let us construct a special
f ∈ KC(E) by constructing f ∗. At a point (x, y, z, w) ∈ B0 we let
f ∗(x, y, z, w) = xyz, on B1 we let f ∗ vanish, and the define f ∗ on B as the generated closed convex
function. This function satisfies the antipode inequality and therefore is indeed of the form f ∗ for some
f ∈ KC(E). In addition, a direct calculation reveals that f ∗(±1, ±1, 0, 0) = −1.
Fix ε > 0 (ε = 1/2 will do). At each point λ ∈ B0 there is some vλ ∈ E such that f ∗(λ) < λvλ −
f (vλ) + ε. Let U ⊆ KC(E) consist of all g such that g(vλ) − f (vλ) < ε for all λ ∈ B0. Then U is a
neighbourhood of f , and if g ∈ U, then at λ ∈ B0 we have
and therefore
g∗(λ) ≥ λvλ − g(vλ) > λvλ − f (vλ) − ε > f ∗(λ) − 2ε,
g∗(λ) ≤ −g(−λ) < − f ∗(−λ) + 2ε = f ∗(λ) + 2ε.
Thus g∗ < f ∗ + 2ε throughout the unit cube, and in particular g∗(±1, ±1, 0, 0) < 2ε − 1. Thus, for
ε = 1/2 or less, the antipode identity fails at (±1, ±1, 0, 0) for every g ∈ U, so U contains no isolated
points.
If we want counter-examples consisting of smooth spaces we need to move to infinite dimension. In
fact, we obtain a plethora of examples over which there are no isolated types other than the obvious
ones.
20
ITAÏ BEN YAACOV AND C. WARD HENSON
Proposition 6.7. Let E be a Banach space such that E (B(E∗)) * ∂B(E∗). Then the only isolated types over E
are the realised ones.
This holds in particular when E = c0, E∗ = ℓ1 or E = ℓp, E∗ = ℓq with 1 < p, q, 1
p + 1
q = 1, since λi → 0,
where λi consists of a single 1 at position i and 0 elsewhere.
Proof. Indeed, let f ∈ KC(E) be isolated. By Corollary 5.9, f ∗ satisfies the antipode identity on E (B(E∗))
and therefore (by continuity of f ∗) at some non-boundary point. By Lemma 4.14, f is realised.
(cid:4)
Over E = G the isolated types are again exactly the realised ones, but in this specific case they are
dense and G[E] = E.
Question 6.8. Is there any infinite-dimensional E other than G over which the isolated types are dense?
Is there any infinite-dimensional E over which there are unrealised isolated types? Specifically, what
happens in the case E = ℓ1, E∗ = ℓ∞, to which Proposition 6.7 does not apply?
7. COUNTING TYPES
We conclude with a calculation of the size of the type-space over a separable Banach space E. By
"size" we mean here the metric density character (since the cardinal Sn(E) is the continuum as soon
as n > 0 and E 6= 0).
Theorem 7.1. Let E be a separable Banach space.
(i) If E is finite-dimensional and polyhedral, then Sn(E) is metrically separable.
(ii) Otherwise, Sn(E) has metric density character equal to the continuum for every n ≥ 1.
Proof. Assume first that E is finite-dimensional and polyhedral. Then by Melleray [Mel07, Remarks
following Corollary 4.6], the space K(E) is separable, and a fortiori so is S1(E) = KC(E). The passage
from 1-types to n-types is done as in the proof of Lemma 2.15, and is left to the reader.
Now assume that E is not both finite-dimensional and polyhedral. Then by Lindenstrauss [Lin64,
Theorem 7.7] there exists a sequence {vn} ⊆ E such that for any n 6= m and choice of signs:
kvn ± vmk ≤ kvnk + kvmk − 1.
Embed E (isometrically) in ℓ∞, and for a sequence ¯ε ∈ {±1}N, consider the family of closed balls
B(εnvn, kvnk − 1
2 ). By hypothesis every two such balls intersect at a non empty set, and therefore there
exists v ∈ ℓ∞ that belongs to them all. In other words, there exists ξ ¯ε = tp(v/E) ∈ S1(E) such that
kx − εnvnkξ ¯ε ≤ kvnk − 1
2 . If ¯ε 6= ¯ε′, then d(ξ ¯ε, ξ ¯ε′ ) ≥ 1, so the density character of S1(E) is at least the
continuum. The same holds a fortiori for Sn(E), n ≥ 1.
(cid:4)
Remark 7.2. Lindenstrauss's argument is quite elementary and yields a quick proof for Theorem 7.1(ii)
that does not depend on the machinery developed in earlier sections. An argument closer to the spirit
of the present paper can also be given.
Let Ξ be the set of lower semi-continuous functions f0 : E (B(E∗)) → R that satisfy in addition
f0(λ) + f0(−λ) ≤ 0. Then E is not a finite-dimensional polyhedral space if and only if E (B(E∗)) is
infinite, in which case Ξ has density character continuum. If f0 ∈ Ξ and f = f ∗
0 as in Lemma 4.5, then
f ∗↾E (B(E∗)) = f0 and f ∗(λ) + f ∗(−λ) ≤ 0 throughout B(E∗), so f ∈ KC(E) and we are done. Notice
that this argument has the advantage of treating the two cases of "finite-dimensional, non polyhedral"
and "infinite-dimensional" in the same manner, while the proof of [Lin64, Theorem 7.7] treats them
separately, with the second one being significantly more involved.
Theorem 7.1(ii) answers Problem 2 of Avilés et al. [ACC+11, Section 4] in the negative (and we
thank Wiesław KUBI ´S for having pointed this out to us). They say that a Banach space G is of universal
disposition for finite-dimensional spaces if it satisfies a strengthening of Definition 2.1 with ψ being an
isometry.
Corollary 7.3. The density character of any space of universal disposition for finite-dimensional spaces is at least
the continuum. In other words, the answer to Problem 2 of [ACC+11, Section 4] is negative.
Proof. Assume that G is of universal disposition for finite-dimensional spaces. Then the Euclidean plane
E embeds isometrically in G, and all types over E are realised in G, so the density character of G must
be at least the metric density character of S1(E), namely the continuum.
(cid:4)
On the other hand, say that a Gurarij space G is strongly ℵ1-homogeneous if the following stronger
version of Corollary 2.8 holds in G:
GENERIC ORBITS AND TYPE ISOLATION IN THE GURARIJ SPACE
21
For every separable F ⊆ G and isometric embedding ϕ : F → G there exists an isomet-
ric automorphism ψ ∈ Aut(G) extending ϕ.
Clearly, a strongly ℵ1-homogeneous Gurarij space is of universal disposition for finite-dimensional (and
even separable) spaces. Moreover, there does exist such a space of density character the continuum.
This is merely a special case of a general model theoretic result: for any cardinal κ and structure M of
density character ≤ 2κ, in a language of cardinal ≤ κ, there exists an elementary extension M′ (cid:23) M of
density character still ≤ 2κ, which is moreover κ+-saturated and strongly κ+-homogeneous. Apply this
to M = G and κ = ℵ0.
REFERENCES
[ACC+11] Antonio AVILÉS, Félix CABELLO SÁNCHEZ, Jesús M. F. CASTILLO, Manuel GONZÁLEZ, and Yolanda MORENO, Banach
spaces of universal disposition, Journal of Functional Analysis 261 (2011), no. 9, 2347 -- 2361, doi:10.1016/j.jfa.2011.06.011.
[BBHU08] Itaï BEN YAACOV, Alexander BERENSTEIN, C. Ward HENSON, and Alexander USVYATSOV, Model theory for metric
structures, Model theory with applications to algebra and analysis. Vol. 2, London Math. Soc. Lecture Note Ser., vol.
350, Cambridge Univ. Press, Cambridge, 2008, pp. 315 -- 427, doi:10.1017/CBO9780511735219.011.
[Ben08a] Itaï BEN YAACOV, Continuous first order logic for unbounded metric structures, Journal of Mathematical Logic 8 (2008),
no. 2, 197 -- 223, doi:10.1142/S0219061308000737, arXiv:0903.4957.
[Ben08b]
, Topometric spaces and perturbations of metric structures, Logic and Analysis 1 (2008), no. 3 -- 4, 235 -- 272,
doi:10.1007/s11813-008-0009-x, arXiv:0802.4458.
[Ben09]
[Ben14]
[Ben15]
, Modular functionals and perturbations of Nakano spaces, Journal of Logic and Analysis 1 (2009), Paper 1, 42 pp.,
doi:10.4115/jla.2009.1.1, arXiv:0802.4285.
, The linear isometry group of the Gurarij space is universal, Proceedings of the American Mathematical Society 142
(2014), no. 7, 2459 -- 2467, doi:10.1090/S0002-9939-2014-11956-3, arXiv:1203.4915.
, Fraïssé limits of metric structures, Journal of Symbolic Logic 80 (2015), no. 1, 100 -- 115, doi:10.1017/jsl.2014.71,
arXiv:1203.4459.
[Bou81] Nicolas BOURBAKI, Espaces vectoriels topologiques. Chapitres 1 à 5, new ed., Masson, Paris, 1981, Éléments de math-
ématique.
[Bre83] Haïm BREZIS, Analyse fonctionnelle, Collection Mathématiques Appliquées pour la Maîtrise, Masson, Paris, 1983,
[BU07]
[BU10]
Théorie et applications.
Itaï BEN YAACOV and Alexander USVYATSOV, On d-finiteness in continuous structures, Fundamenta Mathematicae 194
(2007), 67 -- 88, doi:10.4064/fm194-1-4.
, Continuous first order logic and local stability, Transactions of the American Mathematical Society 362 (2010),
no. 10, 5213 -- 5259, doi:10.1090/S0002-9947-10-04837-3, arXiv:0801.4303.
[Gur66] Vladimir I. GURARIJ, Spaces of universal placement, isotropic spaces and a problem of Mazur on rotations of Banach spaces,
Akademija Nauk SSSR. Sibirskoe Otdelenie. Sibirskiı Matematiceskiı Žurnal 7 (1966), 1002 -- 1013.
[Kat88] Miroslav KAT ETOV, On universal metric spaces, General topology and its relations to modern analysis and algebra, VI
(Prague, 1986), Res. Exp. Math., vol. 16, Heldermann, Berlin, 1988, pp. 323 -- 330.
[KS13] Wiesław KUBI ´S and Sławomir SOLECKI, A proof of uniqueness of the Gurariı space, Israel Journal of Mathematics 195
(2013), no. 1, 449 -- 456, doi:10.1007/s11856-012-0134-9, arXiv:1110.0903.
Joram LINDENSTRAUSS, Extension of compact operators, Memoirs of the American Mathematical Society 48 (1964), 112.
[Lin64]
[Lus76] Wolfgang LUSKY, The Gurarij spaces are unique, Archiv der Mathematik 27 (1976), no. 6, 627 -- 635.
[Maz33] Stanisław MAZUR, Über konvexe mengen in linearen normierten räumen, Studia Mathematica 4 (1933), 70 -- 84.
[Mel07]
Julien MELLERAY, On the geometry of Urysohn's universal metric space, Topology and its Applications 154 (2007), no. 2,
384 -- 403, doi:10.1016/j.topol.2006.05.005.
[Roc70] R. Tyrrell ROCKAFELLAR, Convex analysis, Princeton Mathematical Series, No. 28, Princeton University Press, Princeton,
N.J., 1970.
[Usp08] Vladimir V. USPENSKIJ, On subgroups of minimal topological groups, Topology and its Applications 155 (2008), no. 14,
1580 -- 1606, doi:10.1016/j.topol.2008.03.001, arXiv:math/0004119.
ITAÏ BEN YAACOV, UNIVERSITÉ CLAUDE BERNARD -- LYON 1, INSTITUT CAMILLE JORDAN, CNRS UMR 5208, 43
BOULEVARD DU 11 NOVEMBRE 1918, 69622 VILLEURBANNE CEDEX, FRANCE
URL: http://math.univ-lyon1.fr/~begnac/
C. WARD HENSON, UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN, URBANA, ILLINOIS 61801, USA
URL: http://www.math.uiuc.edu/~henson/
|
1807.03780 | 1 | 1807 | 2018-07-10T09:39:56 | Characterizations of norm--parallelism in spaces of continuous functions | [
"math.FA"
] | In this paper, we consider the characterization of norm--parallelism problem in some classical Banach spaces. In particular, for two continuous functions $f, g$ on a compact Hausdorff space $K$, we show that $f$ is norm--parallel to $g$ if and only if there exists a probability measure (i.e. positive and of full measure equal to $1$) $\mu$ with its support contained in the norm attaining set $\{x\in K: \, |f(x)| = \|f\|\}$ such that $\big|\int_K \overline{f(x)}g(x)d\mu(x)\big| = \|f\|\,\|g\|$. | math.FA | math |
CHARACTERIZATIONS OF NORM–PARALLELISM IN SPACES
OF CONTINUOUS FUNCTIONS
ALI ZAMANI
Abstract. In this paper, we consider the characterization of norm–parallelism
problem in some classical Banach spaces.
In particular, for two continuous
functions f, g on a compact Hausdorff space K, we show that f is norm–
parallel to g if and only if there exists a probability measure (i.e. positive and
of full measure equal to 1) µ with its support contained in the norm attaining
set {x ∈ K : f (x) = kfk} such that (cid:12)(cid:12)RK f (x)g(x)dµ(x)(cid:12)(cid:12) = kfk kgk.
1. Introduction and preliminaries
Let (X,k · k) be a normed space and denote, as usual, by BX and SX its
closed unit ball and unit sphere, respectively, and denote the topological dual
of X by X ∗. Let Cb(Ω) and C(K) denote the Banach spaces of all bounded
continuous functions on a locally compact Hausdorff space Ω, with the usual norm
x∈Ωf (x) (f ∈ Cb(Ω)) and all continuous functions on a compact Hausdorff
kfk = sup
space K, with the usual norm kfk = max
x∈K f (x) (f ∈ C(K)), respectively. By
Cu(BX , X) we denote the space of all uniformly continuous X valued functions on
BX endowed with the supremum norm. Given a bounded function f : SX −→ X,
its numerical radius is
v(f ) := sup{x∗(f (x)) : kx∗k = x∗(x) = 1}.
Let us comment that for a bounded function f : BX −→ X, the above definitions
apply by just considering v(f ) := v(fSX ).
Recall that an element x ∈ X is said to be norm–parallel to another element
y ∈ X (see [12, 17]), in short x k y, if
kx + λyk = kxk + kyk
for some λ ∈ T.
Here, as usual, T is the unit circle of the complex plane. In the framework of inner
product spaces, the norm–parallel relation is exactly the usual vectorial parallel
relation, that is, x k y if and only if x and y are linearly dependent. In the setting
of normed linear spaces, two linearly dependent vectors are norm–parallel, but
the converse is false in general. Notice that the norm–parallelism is symmetric
and R-homogenous, but not transitive (i.e., x k y and y k z ; x k z; see [18,
Example 2.7], unless X is smooth at y; see [15, Theorem 3.1]).
2010 Mathematics Subject Classification. 47A30, 46B20, 46E15.
Key words and phrases. Norm–parallelism, Banach space of continuous functions, Probabil-
ity measure.
1
2
A. ZAMANI
In the context of continuous functions, the well-known Daugavet equation
kT + Idk = kTk + 1
is a particular case of parallelism. Here Id denotes, as usual, the identity function.
Applications of this equation arise in solving a variety of problems in approxima-
tion theory; see [14] and the references therein.
Some characterizations of the norm–parallelism for operators on various Banach
spaces and elements of an arbitrary Hilbert C ∗-module were given in [2, 3, 7, 9,
13, 15, 16, 17, 18].
In particular, for bounded linear operators T, S on a Hilbert space (H, [·,·]), it
was proved in [17, Theorem 3.3] that T k S if and only if there exists a sequence
of unit vectors {ξn} in H such that
lim
n→∞(cid:12)(cid:12)[T ξn, Sξn](cid:12)(cid:12) = kTk kSk.
(cid:12)(cid:12)[T ξ, Sξ](cid:12)(cid:12) = kTk kSk.
Further, for compact operators T, S it was obtained in [16, Theorem 2.10] that
T k S if and only if there exists a unit vector ξ ∈ H such that
In [18], the authors also considered the characterization of norm parallelism prob-
lem for operators when the operator norm is replaced by the Schatten p-norm
(1 < p < ∞). More precisely, it was proved in [18, Proposition 2.19] that T k S
in the Schatten p-norm if and only if
(cid:12)(cid:12)(cid:12)tr(Tp−1U ∗S)(cid:12)(cid:12)(cid:12) = kTkp
p−1kSkp,
where T = UT is the polar decomposition of T .
therein.
Some other related topics can be found in [1, 5, 6, 7, 10, 16], and the references
It is our aim in the next section to give characterizations of the norm–parallelism
in Cb(Ω) and C(K). More precisely, for f, g ∈ Cb(Ω) we prove that f k g if and
only if there exists a sequence of probability measures µn concentrated at the set
{x ∈ Ω : f (x) ≥ kfk − ε} (ε > 0) such that
n→∞ZΩ
(cid:12)(cid:12)(cid:12) lim
f (x)g(x)dµn(x)(cid:12)(cid:12)(cid:12) = kfk−1kgk lim
n→∞ZΩ f (x)2dµn(x).
Moreover, for f, g ∈ C(K) we show that f k g if and only if there exists a
probability measure µ with support contained in the norm attaining set {x ∈ K :
f (x) = kfk} such that
(cid:12)(cid:12)(cid:12)ZK
f (x)g(x)dµ(x)(cid:12)(cid:12)(cid:12) = kfk kgk.
Finally, in the next section, we state a characterization of the norm-parallelism
for uniformly continuous X valued functions on BX to the identity function.
Actually, we show that if X is a Banach space and f ∈ Cu(BX , X), then f k Id if
and only if kfk = v(f ).
NORM–PARALLELISM IN SPACES OF CONTINUOUS FUNCTIONS
3
2. Main results
We begin with the following results, which will be useful in other contexts as
well.
Lemma 2.1. [17, Theorem 4.1] Let X be a normed space. For x, y ∈ X the
following statements are equivalent:
(i) x k y.
(ii) There exists a norm one linear functional ϕ over X such that ϕ(x) = kxk
and ϕ(y) = kyk.
Lemma 2.2. [8, Theorem 3.1] Let Ω be a locally compact Hausdorff space and
let f, g ∈ Cb(Ω). Then
lim
t→0+
kf + tgk − kfk
t
= inf
ε>0
sup
x∈M ε
f
Re(cid:0)e−i arg(f (x))g(x)(cid:1),
where M ε
f = {x ∈ Ω : f (x) ≥ kfk − ε}.
We use some techniques of [8] to prove the following theorem. Recall that a
probability measure is a positive measure of total mass 1.
Theorem 2.3. Let Ω be a locally compact Hausdorff space and let f, g ∈ Cb(Ω).
Then the following statements are equivalent:
(i) For every ε > 0, there exists a sequence of probability measures µn con-
centrated at M ε
f such that
n→∞ZΩ
kfk(cid:12)(cid:12)(cid:12) lim
(ii) f k g.
f (x)g(x)dµn(x)(cid:12)(cid:12)(cid:12) = kgk lim
n→∞ZΩ f (x)2dµn(x).
Proof. (i)⇒(ii) Suppose (i) holds. Let ε > 0. So, there exist a sequence of
probability measures µn concentrated at M ε
f and λ ∈ T such that
n→∞ZΩ f (x)2dµn(x) = λkfk lim
n→∞ZΩ
f (x)g(x)dµn(x),
kgk lim
and hence
n→∞ZΩ
lim
f (x)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x) = 0.
(2.1)
Let us now define a linear functional ϕ : span(cid:8)f, (kgkf − λkfkg)(cid:9) −→ C by
setting
ϕ(cid:16)αf + β(cid:0)kgkf − λkfkg(cid:1)(cid:17) = αkfk
(α, β ∈ C).
4
We have
A. ZAMANI
(cid:13)(cid:13)(cid:13)αf +β(cid:0)kgkf − λkfkg(cid:1)(cid:13)(cid:13)(cid:13)
= sup
2
2
= sup
x∈Ω(cid:12)(cid:12)(cid:12)αf (x) + β(cid:0)kgkf (x) − λkfkg(x)(cid:1)(cid:12)(cid:12)(cid:12)
x∈Ω(cid:16)α2f (x)2 + 2Re(cid:2)αβf (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1)(cid:3)
≥(cid:12)(cid:12)(cid:12)ZΩ(cid:16)α2f (x)2 + 2Re(cid:2)αβf (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1)(cid:3)
≥ α2ZΩ f (x)2dµn(x) + 2Re(cid:2)αβZΩ
≥ α2(kfk − ε)2 + 2Re(cid:2)αβZΩ
+ β2(cid:12)(cid:12)kgkf (x) − λkfkg(x)(cid:12)(cid:12)2(cid:17)
+ β2(cid:12)(cid:12)kgkf (x) − λkfkg(x)(cid:12)(cid:12)2(cid:17)dµn(x)(cid:12)(cid:12)(cid:12)
f (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1)dµn(x)(cid:3)
dµn(x) ≥ 0(cid:1)
(cid:0)since µn is a probability measure concentrated at M ε
f(cid:1)
(cid:0)since RΩ β2(cid:12)(cid:12)kgkf (x) − λkfkg(x)(cid:12)(cid:12)2
f (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1)dµn(x)(cid:3).
Thus
2
(cid:13)(cid:13)(cid:13)αf +β(cid:0)kgkf − λkfkg(cid:1)(cid:13)(cid:13)(cid:13)
≥ α2(kfk − ε)2 + 2Re(cid:2)αβZΩ
f (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1)dµn(x)(cid:3).
(2.2)
Letting ε → 0+ and n −→ ∞ in (2.2), then by (2.1) we obtain
(cid:13)(cid:13)(cid:13)αf + β(cid:0)kgkf − λkfkg(cid:1)(cid:13)(cid:13)(cid:13) ≥ αkfk.
Hence (cid:12)(cid:12)(cid:12)ϕ(cid:16)αf + β(cid:0)kgkf − λkfkg(cid:1)(cid:17)(cid:12)(cid:12)(cid:12) ≤(cid:13)(cid:13)(cid:13)αf + β(cid:0)kgkf − λkfkg(cid:1)(cid:13)(cid:13)(cid:13) and ϕ(f ) = kfk.
Thus kϕk = 1. Therefore, the Hahn-Banach theorem extends ϕ to a linear
functional eϕ on Cb(Ω), with keϕk = 1. Since eϕ(f ) = kfk and eϕ(cid:0)kgkf−λkfkg(cid:1) = 0,
we get eϕ(λg) = kgk, hence eϕ(g) = kgk. Now, by Lemma 2.1, we conclude that
f k g.
(ii)⇒(i) Let f k g. By Lemma 2.1, there exists a norm one linear functional ϕ
over Cb(Ω) such that ϕ(f ) = kfk and ϕ(g) = kgk. So, there exists λ ∈ T such
NORM–PARALLELISM IN SPACES OF CONTINUOUS FUNCTIONS
5
that ϕ(λg) = kgk. From kϕk = 1, ϕ(f ) = kfk and ϕ(λg) = kgk it follows that
(cid:13)(cid:13)(cid:13)f + reiθ(cid:0)kgkf − λkfkg(cid:1)(cid:13)(cid:13)(cid:13) − kfk
r
r
≥
(cid:12)(cid:12)(cid:12)ϕ(cid:16)f + reiθ(cid:0)kgkf − λkfkg(cid:1)(cid:17)(cid:12)(cid:12)(cid:12) − kfk
= (cid:12)(cid:12)(cid:12)ϕ(f ) + reiθkgkϕ(f ) − reiθkfkϕ(λg)(cid:12)(cid:12)(cid:12) − kfk
= (cid:12)(cid:12)(cid:12)kfk + reiθkgkkfk − reiθkfkkgk(cid:12)(cid:12)(cid:12) − kfk
r
r
= 0
for all r > 0 and all θ ∈ [0, 2π). Hence, by Lemma 2.2, we get
inf
ε>0
sup
x∈M ε
f
Re(cid:16)eiθe−i arg(f (x))(cid:0)kgkf (x) − λkfkg(x)(cid:1)(cid:17) ≥ 0
for all θ ∈ [0, 2π). Thus for all ε > 0 the set
Kε := {e−i arg(f (x))(cid:0)kgkf (x) − λkfkg(x)(cid:1) : x ∈ M ε
f}
contains at least one element with nonnegative real part under all rotations
around the origin. Hence the values of the function kgkf − λkfkg on M ε
f are
not contained in an open half plane with boundary that contains the origin. So,
for all ε > 0 the closed convex hull of the set Kε contains the origin. The convex
hull of Kε consists of points of the formRΩ e−i arg(f (x))(cid:0)kgkf (x)−λkfkg(x)(cid:1)dµε(x),
where µε is a probability measure supported on a finite subset of M ε
f (see [11],
chap. 3). Let n0 ∈ N and 1
< ε. Thus for every n ≥ n0, there is a probability
measure µn concentrated at M ε
f such that
n0
(cid:12)(cid:12)(cid:12)ZΩ
e−i arg(f (x))(cid:0)kgkf (x) − λkfkg(x)(cid:1)dµn(x)(cid:12)(cid:12)(cid:12) <
1
n
.
We have
(cid:12)(cid:12)(cid:12)ZΩ
f (x)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)ZΩ(cid:16)f (x) − kfke−i arg(f (x))(cid:17)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x)
≤(cid:12)(cid:12)(cid:12)ZΩ(cid:16)f (x) − kfke−i arg(f (x))(cid:17)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x)(cid:12)(cid:12)(cid:12)
+ZΩ kfke−i arg(f (x))(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x)(cid:12)(cid:12)(cid:12)
(2.3)
(2.4)
+ kfk(cid:12)(cid:12)(cid:12)ZΩ
e−i arg(f (x))(cid:0)kgkf (x) − λkfkg(x)(cid:1)dµn(x)(cid:12)(cid:12)(cid:12).
6
A. ZAMANI
Furthermore, since kfk − 1
n ≤ f (x) ≤ kfk for all x ∈ M ε
(cid:12)(cid:12)(cid:12)ZΩ(cid:16)f (x) − kfke−i arg(f (x))(cid:17)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x)(cid:12)(cid:12)(cid:12)
f we have
≤
≤
1
1
nZΩ(cid:12)(cid:12)(cid:12)kgkf (x) − λkfkg(x)(cid:12)(cid:12)(cid:12)dµn(x)
n(cid:16)kgkZΩ f (x)dµn(x) + kfkZΩ g(x)dµn(x)(cid:17)
2
nkfkkgk.
(2.5)
≤
By (2.3), (2.4) and (2.5) we get
2
nkfkkgk + kfk
1
n
.
(2.6)
Taking lim
n→∞
and hence
(cid:12)(cid:12)(cid:12)ZΩ
in (2.6), we obtain
f (x)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x)(cid:12)(cid:12)(cid:12) ≤
n→∞ZΩ
n→∞ZΩ
kfk(cid:12)(cid:12)(cid:12) lim
f (x)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµn(x) = 0,
f (x)g(x)dµn(x)(cid:12)(cid:12)(cid:12) = kgk lim
lim
n→∞ZΩ f (x)2dµn(x).
(cid:3)
Next, we present a characterization of the norm–parallelism for continuous
functions on a compact Hausdorff space K. We will need the following lemma.
Lemma 2.4. [8, Theorem 3.1] Let K be a compact Hausdorff space and let f, g ∈
C(K). Then
lim
t→0+
kf + tgk − kfk
t
= max
x∈Mf
Re(cid:0)e−i arg(f (x))g(x)(cid:1),
where Mf = {x ∈ K : f (x) = kfk}.
Theorem 2.5. Let K be a compact Hausdorff space and let f, g ∈ C(K). Then
the following statements are equivalent:
(i) There exists a probability measure µ with support contained in the norm
attaining set Mf such that
(cid:12)(cid:12)(cid:12)ZK
f (x)g(x)dµ(x)(cid:12)(cid:12)(cid:12) = kfk kgk.
(ii) f k g.
Proof. We proceed as in the proof of Theorem 2.3.
(i)⇒(ii) Suppose (i) holds. So, there exists λ ∈ T such that
f (x)(cid:16)kgkf (x) − λkfkg(x)(cid:17)dµ(x) = 0.
ZK
NORM–PARALLELISM IN SPACES OF CONTINUOUS FUNCTIONS
7
Thus
(cid:13)(cid:13)(cid:13)αf +β(cid:0)kgkf − λkfkg(cid:1)(cid:13)(cid:13)(cid:13)
2
for any α, β ∈ C. Let ϕ : span(cid:8)f, (kgkf − λkfkg)(cid:9) −→ C be the linear functional
defined as
2
≥ α2kfk2 + β2ZMf (cid:12)(cid:12)(cid:12)kgkf (x) − λkfkg(x)(cid:12)(cid:12)(cid:12)
ϕ(cid:16)αf + β(cid:0)kgkf − λkfkg(cid:1)(cid:17) = αkfk
(α, β ∈ C).
dµ(x) ≥ α2kfk2,
Hence ϕ(f ) = kfk, ϕ(λg) = kgk and kϕk = 1. By the Hahn-Banach theorem, ϕ
extends to a linear functional eϕ on C(K), of the same norm. Since eϕ(f ) = kfk
and eϕ(g) = kgk, Lemma 2.1 yields f k g.
(ii)⇒(i) Let f k g. By Lemma 2.1, there exist λ ∈ T and a norm one linear
functional ϕ over C(K) such that ϕ(f ) = kfk and ϕ(λg) = kgk. Hence by Lemma
2.4, we get
max
x∈Mf
Re(cid:16)e−i arg(f (x))(cid:0)kgkf (x) − λkfkg(x)(cid:1)(cid:17) ≥ 0.
So, the convex hull of the set {f (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1) : x ∈ Mf} consists of
points of the form RK f (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1)dµ(x), where µ is a probability
measure supported on a finite subset of Mf . Then there is a sequence µn of
probability measures such that
n→∞ZK
lim
f (x)(cid:0)kgkf (x) − λkfkg(x)(cid:1)dµn(x) = 0.
By the Banach-Alaoglu compactness theorem in dual space, there is a probability
measure µ such that lim
µni = µ. Thus the support of µ is contained in Mf and
i→∞
we obtain
(cid:12)(cid:12)(cid:12)ZK
f (x)g(x)dµ(x)(cid:12)(cid:12)(cid:12) = kfk kgk.
(cid:3)
If
As a consequence of Theorem 2.5 we have the following result.
Corollary 2.6. Let K be a compact Hausdorff space and let f, g ∈ C(K).
Mf = {x0}, then the following statements are equivalent:
(i) f k g.
(ii) {x0} ⊆ Mg.
We closed this paper with the following equivalence theorem. More precisely,
we state a characterization of the norm–parallelism for uniformly continuous X
valued functions on BX to the identity function. Note that since BX is convex
and bounded, then every function in Cu(BX , X) is also bounded. Before stating
our result, let us quote a result from [4].
8
A. ZAMANI
Lemma 2.7. [4, Corollary 2.4] Let X be a Banach space. Let 0 < θ < 2 and
suppose y ∈ BX and y ∗ ∈ BX ∗ satisfy Rey ∗(y) > 1 − θ. Then, there are z ∈ SX
and z∗ ∈ SX ∗ such that
z∗(z) = 1,
and
ky ∗ − z∗k < √2θ.
ky − zk < √2θ
Theorem 2.8. Let X be a Banach space and let f ∈ Cu(BX , X). Then the
following statements are equivalent:
(i) v(f ) = kfk.
(ii) f k Id,
where Id stands for the identity function.
Proof. (i)⇒(ii) Let v(f ) = kfk. For every ε > 0, we may find x ∈ SX and x∗ ∈
with λ ∈ T. We have
SX ∗ such that x∗(x) = 1 and (cid:12)(cid:12)x∗(f (x))(cid:12)(cid:12) > kfk − ε. Let x∗(f (x)) = λ(cid:12)(cid:12)x∗(f (x))(cid:12)(cid:12)
1 + kfk ≥ kId + λfk ≥ kx + λf (x)k
≥(cid:12)(cid:12)(cid:12)x∗(cid:0)x + λf (x)(cid:1)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)x∗(x) + λx∗(cid:0)f (x)(cid:1)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)1 + λλ(cid:12)(cid:12)x∗(f (x))(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = 1 +(cid:12)(cid:12)x∗(f (x))(cid:12)(cid:12) > 1 + kfk − ε.
Thus
1 + kfk ≥ kId + λfk > 1 + kfk − ε.
Letting ε → 0+, we obtain kId + λfk = 1 + kfk, or equivalently, f k Id.
(ii)⇒(i) Let f k Id. So, there exists λ ∈ T such that
kId + λfk = 1 + kfk.
Fix 0 < ε < 1. Since f ∈ Cu(BX, X), there exists 0 < δ < ε such that
Since 1 + kfk = sup
ky − zk < δ =⇒ kf (y) − f (z)k < ε
(y, z ∈ BX ).
y∈BXky + λf (y)k, there exists y ∈ BX such that
ky + λf (y)k > 1 + kfk −
δ2
2
.
Then we may find y ∗ ∈ SX ∗ such that
Rey ∗(cid:0)y + Rey ∗(λf (y)(cid:1)(cid:1) > 1 + kfk −
δ2
2
,
(2.7)
(2.8)
(2.9)
(2.10)
which yields
and
Rey ∗(y) > 1 −
δ2
2
Rey ∗(λf (y)) > kfk −
δ2
2
.
By (2.8) and Lemma 2.7, there are z ∈ SX and z∗ ∈ SX ∗ such that
z∗(z) = 1,
ky − zk < δ
and ky ∗ − z∗k < δ.
NORM–PARALLELISM IN SPACES OF CONTINUOUS FUNCTIONS
9
So, by (2.10) and (2.7) we get
kf (y) − f (z)k < ε.
By (2.10) and (2.11) it follows that
(cid:12)(cid:12)(cid:12)Rez∗(λf (z)) − Rey ∗(λf (y))(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)Rez∗(cid:0)λf (z) − λf (y)(cid:1)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)Re(z∗ − y ∗)(λf (y))(cid:12)(cid:12)(cid:12)
≤ kλf (z) − λf (y)k + kz∗ − y ∗k ≤ ε + δ,
(2.11)
(2.12)
(cid:3)
whence
So, by (2.9) and (2.12) we obtain
(cid:12)(cid:12)(cid:12)Rez∗(λf (z)) − Rey ∗(λf (y))(cid:12)(cid:12)(cid:12) < ε + δ.
Rez∗(λf (z)) > kfk −
δ2
2 − (ε + δ) > kfk − 3ε.
This implies
kfk ≥ v(f ) ≥ Rez∗(λf (z)) > kfk − 3ε.
Letting ε → 0+, we conclude v(f ) = kfk.
As an immediate consequence of Theorem 2.8 we have the following result.
Corollary 2.9. Let X be a Banach space and let f ∈ Cu(BX , X). If f k Id, then
kfk = sup
x∈SXkf (x)k.
Acknowledgement. The author would like to thank the referees for their
careful reading of the manuscript and useful comments.
References
1. Lj. Arambasi´c and R. Raji´c, A strong version of the Birkhoff–James orthogonality in Hilbert
C ∗-modules, Ann. Funct. Anal. 5 (2014), no. 1, 109–120.
2. M. Barraa and M. Boumazgour, Inner derivations and norm equality, Proc. Amer. Math.
Soc. 130 (2002), no. 2, 471–476.
3. T. Bottazzi, C. Conde, M.S. Moslehian, P. W´ojcik and A. Zamani, Orthogonality and
parallelism of operators on various Banach spaces, J. Aust. Math. Soc. (to appear).
4. M. Chica, V. Kadets, M. Mart´ın, S. Moreno–Pulido, and F. Rambla–Barreno, Bishop-
Phelps-Bollob´as moduli of a Banach space, J. Math. Anal. Appl. 412 (2014), no. 2, 697-719.
5. J. Chmieli´nski, R. Lukasik and P. W´ojcik, On the stability of the orthogonality equation and
the orthogonality-preserving property with two unknown functions, Banach J. Math. Anal.
10 (2016), no. 4, 828–847.
6. P. Ghosh, D. Sain and K. Paul, On symmetry of Birkhoff-James orthogonality of linear
operators, Adv. Oper. Theory 2 (2017), no. 4, 428–434.
7. P. Grover, Orthogonality of matrices in the Ky Fan k-norms, Linear Multilinear Algebra,
65 (2017), no. 3, 496–509.
8. D. Kecki´c, Orthogonality and smooth points in C(K) and Cb(Ω), Eurasian Math. J. 3
(2012), no. 4, 44–52.
9. A. Mal, D. Sain and K. Paul, On some geometric properties of operator spaces, Banach J.
Math. Anal. (to appear).
10. M. S. Moslehian and A. Zamani, Mappings preserving approximate orthogonality in Hilbert
C ∗-modules, Math Scand. 122 (2018), 257–276.
11. W. Rudin, Functional Analysis, McGrawHill, Inc., New York, 1973.
10
A. ZAMANI
12. A. Seddik, Rank one operators and norm of elementary operators, Linear Algebra Appl.
424 (2007), 177–183.
13. A. Seddik, On the injective norm of Σn
i=1Ai ⊗ Bi and characterization of normaloid opera-
14. D. Werner, The Daugavet equation for operators on function spaces, J. Funct. Anal. 143
tors, Oper. Matrices 2 (2008), no. 1, 67–77.
(1997), 117–128.
15. P. W´ojcik, Norm–parallelism in classical M -ideals, Indag. Math. (N.S.) 28 (2017), no. 2,
287–293.
16. A. Zamani, The operator-valued parallelism, Linear Algebra Appl. 505 (2016), 282–295.
17. A. Zamani and M.S. Moslehian, Exact and approximate operator parallelism, Canad. Math.
Bull. 58(1) (2015), 207–224.
18. A. Zamani and M.S. Moslehian, Norm–parallelism in the geometry of Hilbert C ∗-modules,
Indag. Math. (N.S.) 27 (2016), no. 1, 266–281.
Department of Mathematics, Farhangian University, Tehran, Iran.
E-mail address: [email protected]
|
1605.04779 | 1 | 1605 | 2016-05-16T14:14:52 | Quasianalyticity in certain Banach function algebras | [
"math.FA"
] | Let $X$ be a perfect, compact subset of the complex plane. We consider algebras of those functions on $X$ which satisfy a generalised notion of differentiability, which we call $\mathcal{F}$-differentiability. In particular, we investigate a notion of quasianalyticity under this new notion of differentiability and provide some sufficient conditions for certain algebras to be quasianalytic. We give an application of our results in which we construct an essential, natural uniform algebra $A$ on a locally connected, compact Hausdorff space $X$ such that $A$ admits no non-trivial Jensen measures yet is not regular. This construction improves an example of the first author (2001). | math.FA | math |
Quasianalyticity in certain Banach function
algebras
J. F. Feinstein
S. Morley∗†
Abstract
Let X be a perfect, compact subset of the complex plane. We
consider algebras of those functions on X which satisfy a generalised
notion of differentiability, which we call F-differentiability. In partic-
ular, we investigate a notion of quasianalyticity under this new notion
of differentiability and provide some sufficient conditions for certain
algebras to be quasianalytic. We give an application of our results
in which we construct an essential, natural uniform algebra A on a
locally connected, compact Hausdorff space X such that A admits
no non-trivial Jensen measures yet is not regular. This construction
improves an example of the first author (2001).
Let X be a perfect, compact subset of the complex plane C. We consider
those normed algebras consisting of complex-valued, continuously complex-
differentiable functions on X, denoted D(1)(X). These algebras were intro-
duced by Dales and Davie in [9] and further investigated, for example, in [2]
and [10]. The algebra D(1)(X) need not be complete, and the completion of
D(1)(X) need not be a Banach function algebra in general.
Bland and the first author [2] introduced F -differentiation, which gener-
alises the usual complex-differentiation, and considered normed algebras of
F -differentiable functions, denoted D(1)
F (X) is com-
plete and D(1)(X) ⊆ D(1)
F (X). The algebra D(1)
F (X).
∗This paper contains work from the second author's PhD thesis.
†The second author was supported by EPSRC grant number EP/L50502X/1
Key words and phrases. Differentiable functions, Banach function algebra, Uniform
algebra, Quasianalyticity, Jensen measures, Swiss cheeses.
2010 Mathematics Subject Classification. Primary 46J10, 46J15; Secondary 46E25.
1
Dales and Davie ([9]) also considered those algebras of complex-valued
functions which have continuous complex-derivatives of all orders, and intro-
duced the Dales-Davie algebras D(X, M). They defined a notion of quasi-
analyticity for these algebras and gave sufficient conditions for the algebra
D(X, M) to be quasianalytic. (For the classical definition of quasianalytic
collections of functions, see [27, Chapter 19].)
In this paper, we define a notion of quasianalyticity for infinitely F -
differentiable functions (defined later) and give a sufficient condition for this
new notion of F -quasianalyticity. In certain cases, this sufficient condition
will improve that given by Dales and Davie in [9]. We conclude the paper
with the construction of a uniform algebra A on a locally connected, compact
Hausdorff space X such that A is essential and A does not admit any non-
trivial Jensen measures yet is not regular. (The relevant definitions are given
in Section 4.) This construction improves an example of the first author from
[13].
1 Definitions and Basic results
Throughout this paper we say compact plane set to mean a non-empty, com-
pact subset of the complex plane C. We denote the set of non-negative
integers by N0 and the set of positive integers by N. Let X be a compact
Hausdorff space. We denote the algebra (with pointwise operations) of all
continuous, complex-valued functions on X by C(X). For E ⊆ X, we set
f E := sup
x∈E
f (x)
(f ∈ C(X)).
With the norm · X, C(X) is a commutative, unital Banach algebra. Let S
be a subset of C(X). We say that S separates the points of X if, for each
x, y ∈ X with x 6= y, there exists f ∈ S such that f (x) 6= f (y).
Definition 1.1. Let X be a compact Hausdorff space. A normed function
algebra on X a normed algebra (A, k · k) such that A is a subalgebra of C(X),
A contains all constant functions and separates the points of X, and, for each
f ∈ A, kf k ≥ f X. A Banach function algebra on X is a normed function
algebra on X which is complete. A uniform algebra is a Banach function
algebra such that k · k = · X.
Let X be a compact Hausdorff space and let A be a Banach function
algebra on X. We say that A is natural on X if every character on A is given
by evaluation at some point of X.
2
We refer the reader to [8, Chapter 4] for further information on Banach
function algebras and uniform algebras.
We are particularly interested in Banach function algebras consisting of
continuous functions on a compact plane set which satisfy some notion of
differentiability.
Definition 1.2. Let X be a perfect compact plane set and let f : X → C
be a function. We say that f is complex differentiable at x ∈ X if the limit
f ′(x) := lim
z→x
z∈X
f (z) − f (x)
z − x
exists. We say that f is complex differentiable on X if f is complex differen-
tiable at each point x ∈ X and we call the function f ′ : X → C the derivative
of f . We say that f is continuously complex differentiable if f ′ is continuous.
In the remainder of this paper, we shall say differentiable and continu-
ously differentiable to mean complex differentiable and continuously complex
differentiable, respectively. We refer the reader to [7], for example, for results
from complex analysis.
Let X be a perfect compact plane set. We denote the algebra of all con-
tinuously differentiable functions on X by D(1)(X). For each n ∈ N, let
D(n)(X) denote the algebra of all n-times continuously differentiable func-
n=1 D(n)(X). Let n ∈ N
and let f ∈ D(n)(X). We denote the nth derivative of f by f (n), and we will
often write f (0) for f .
tions on X (defined inductively). Let D(∞)(X) :=T∞
Definition 1.3. Let M = (Mn)∞
n=0 be a sequence of positive real numbers.
We say that M is an algebra sequence if M0 = 1 and, for each j, k ∈ N0, we
have
(cid:18)j + k
k (cid:19) ≤
Mj+k
MjMk
.
We say that M is log-convex if, for each k ∈ N, we have M 2
k ≤ Mk−1Mk+1.
We conclude this section with a discussion of paths in C. For the remain-
der of this section, let a, b ∈ R with a < b.
Definition 1.4. A path in C is a continuous function γ : [a, b] → C. Let
γ : [a, b] → C be a path. The parameter interval of γ is the interval [a, b].
The endpoints of γ are the points γ(a) and γ(b), which we denote by γ− and
γ+, respectively. We denote by γ∗ the image γ([a, b]) of γ. A subpath of γ
is a path obtained by restricting γ to a non-degenerate, closed subinterval of
[a, b]. If X is a subset of C then we say that γ is a path in X if γ∗ ⊆ X.
3
Let γ : [a, b] → C be a path in C. We say that γ is a Jordan path
if γ is an injective function. We denote the length of γ, as defined in [1,
Chapter 6], by Λ(γ), and we say that γ is rectifiable if Λ(γ) < ∞ and γ is
non-rectifiable otherwise. We say that γ is closed if γ+ = γ−, and we say that
γ is a closed Jordan path if γ is closed and γ(s) = γ(t), where s, t ∈ [a, b],
implies that either s = t or s = a and t = b. We say that γ is admissible if
γ is rectifiable and has no constant subpaths. The reverse of γ is the path
−γ : [−b, −a] → γ∗ given by −γ(t) = γ(−t). It is standard that
Z−γ
f (z) dz = −Zγ
f (z) dz
(f ∈ C(γ∗)).
Now suppose that γ is non-constant and rectifiable. We define the path length
parametrisation γpl : [0, Λ(γ)] → C of γ to be the unique path satisfying
γpl(Λ(γ[a, t])) = γ(t) (t ∈ [a, b]); see, for example, [11, pp. 109-110] for
details. We define the normalised path length parametrisation γno : [0, 1] → C
of γ to be the path such that γno(t) = γpl(tΛ(γ)) for each t ∈ [0, 1]. It is clear
that γpl and γno are necessarily admissible paths and (γpl)∗ = (γno)∗ = γ∗.
It is not hard to show, using [11, Theorem 2.4.18], that
Zγ
f (z) dz =Zγpl
f (z) dz =Zγno
f (z) dz,
for all f ∈ C(γ∗). We shall use this fact implicitly throughout.
Definition 1.5. Let X be a perfect compact plane set and let F be a col-
lection of paths in X. Define F ∗ := {γ∗ : γ ∈ F }. We say that F is effective
if S F ∗ is dense in X, each path γ ∈ F is admissible, and each subpath of a
path in F belongs to F .
Let X be a compact plane set. We say that X is semi-rectifiable if the set
of all Jordan paths in X is an effective collection of paths in X. We say that
X is rectifiably connected if, for each x, y ∈ X, there exists a rectifiable path
γ ∈ F such that γ− = x and γ+ = y. We say that X is uniformly regular
if there exists a constant C > 0 such that for each x, y ∈ X there exists a
rectifiable path γ in X with γ− = x and γ+ = y such that Λ(γ) ≤ Cx − y.
We say that X is pointwise regular if for each x ∈ X there exists a constant
Cx > 0 such that, for each y ∈ X, there exists a path γ in X with γ− = x and
γ+ = y such that Λ(γ) ≤ Cxx − y. Note that each of the above conditions
on X imply that X is perfect.
4
2 F -derivatives
In this section we discuss algebras of F -differentiable functions as investigated
in [2] and [10].
Definition 2.1. Let X be a perfect compact plane set, let F be a collection
of rectifiable paths in X, and let f ∈ C(X). A function g ∈ C(X) is an
F -derivative for f if, for each γ ∈ F , we have
g(z) dz = f (γ+) − f (γ−).
Zγ
If f has an F -derivative on X then we say that f is F -differentiable on X.
The following proposition is a list of the elementary properties of F -
differentiable functions. Details can be found in [2] and [10].
Proposition 2.2. Let X be a semi-rectifiable compact plane set and let F
be an effective collection of paths in X.
(a) Let f, g, h ∈ C(X) be such that g and h are F -derivatives for f . Then
g = h.
(b) Let f ∈ D(1)(X). Then the usual complex derivative of f on X, f ′, is
an F -derivative for f .
(c) Let f1, f2, g1, g2 ∈ C(X) be such that g1 is an F -derivative for f1 and
g2 is an F -derivative for f2. Then f1g2 + g1f2 is an F -derivative for
f1f2.
(d) Let f1, f2, g1, g2 ∈ C(X) and α, β ∈ C be such that g1 is an F -derivative
for f1 and g2 is an F -derivative for f2. Then αg1 + βg2 is an F -
derivative for αf1 + βf2.
Let X be a semi-rectifiable compact plane set, and let F be an effective
collection of paths in X. By (a) of the above proposition, F -derivatives are
unique. So, in this setting, we write f [1] for the unique F -derivative of an
F -differentiable function. This will be the case considered throughout the
remainder of this paper. We will often write f [0] for f . We write D(1)
F (X)
for the algebra of all F -differentiable functions on X. We note that, with
the norm kf kF ,1 := f X + f [1]X (f ∈ D(1)
F (X) is a
Banach function algebra on X ([10, Theorem 5.6]).
F (X)), the algebra D(1)
For each n ∈ N, we define (inductively) the algebra
D(n)
F (X) := {f ∈ D(1)
F (X) : f [1] ∈ D(n−1)
(X)},
F
5
and, for each f ∈ D(n)
that, for each n ∈ N, D(n)
when given the norm
F (X), we write f [n] for the nth F -derivative for f . Note
F (X) is a Banach function algebra on X (see [2])
kf kF ,n :=
f [k]X
k!
n
Xk=0
(f ∈ D(n)
F (X)).
In addition, we define the algebra D(∞)
derivatives of all orders; that is, D(∞)
that, for each n ∈ N, we have D(n)(X) ⊆ D(n)
F (X) of all functions which have F -
F (X). It is easy to see
F (X).
We now describe a class of algebras of infinitely F -differentiable functions
F (X) and D(∞)(X) ⊆ D(∞)
n=1 D(n)
F (X) =T∞
analogous to Dales-Davie algebras as introduced in [2] (see also [4]).
Definition 2.3. Let X be a semi-rectifiable, compact plane set and let F
be an effective collection of paths in X. Let M = (Mn)∞
n=0 be an algebra
sequence. We define the normed algebra
DF (X, M) :=(f ∈ D(∞)
F (X) :
f [j]X
Mj
< ∞)
∞
Xj=0
with pointwise operations and the norm
kf k :=
f [j]X
Mj
∞
Xj=0
(f ∈ DF (X, M)).
3 F -quasianalyticity
In this section, we discuss an F -differentiability version of quasianalyticity,
and give a sufficient condition for a subalgebra of D(∞)
F (X) new notion of
quasianalyticity.
We now introduce the following notion of F -quasianalyticity.
Definition 3.1. Let X be a semi-rectifiable compact plane set and let F
be an effective collection of paths in X. Let A be a subalgebra of D(∞)
F (X).
Then A is F -quasianalytic if, for each γ ∈ F and z0 ∈ γ∗, the conditions
f ∈ A and f [k](z0) = 0 (k = 0, 1, . . . ),
together imply that f (z) = 0 for all z ∈ γ∗.
6
Let X be a semi-rectifiable compact plane set and let F be an effective
collection of paths in X.
We now aim to give some sufficient conditions for F -quasianalyticity for
the algebras DF (X, M). Our method will follow the proof in [6] of the tra-
ditional Denjoy-Carleman theorem.
For the remainder of the section we fix an admissible path Γ. We also
fix F to be the collection of all subpaths and reverses of subpaths of Γ. Let
M = (Mn)∞
n=0 be a sequence of positive real numbers satisfying
M −1/n
n
= +∞.
(1)
∞
Xn=1
We write ds for integrals with respect to the path length measure. Set
D := D(∞)
F (Γ∗) and, for each n ∈ N, set Dn := D(n)
F (Γ∗).
We will require the following lemmas. The first lemma is a summary
of the properties of the log-convex minorant of a sequence of positive real
numbers. We refer the reader to [25, Chapter IV] and [26, Chapter 1] for
details and properties of the log-convex minorants. The properties listed
below are from [6].
Lemma 3.2. Let (Mn)∞
lim inf n→∞ M 1/n
and a strictly increasing sequence (nj)∞
n = +∞. Then there exists a log-convex sequence (M c
n)∞
n=0
j=0 of integers with n0 = 0 such that:
n=0 be a sequence of positive real numbers such that
(a) M c
k ≤ Mk for all k ∈ N0;
(b) M c
nk = Mnk for all k ∈ N0;
(c) for each k ∈ N0, we have M c
j /M c
j+1 = M c
nk/M c
nk+1 for all j ∈ N0 with
nk ≤ j < nk+1.
It is easy to see that if a sequence M = (Mn)∞
n=0 of positive real num-
bers satisfies (1) then the log-convex minorant M c = (M c
n=0, as given by
Lemma 3.2, satisfies (1). Moreover, if M c satisfies (1) then M c satisfies
n)∞
n/M c
n=0 M c
We will require lemmas to prove the main result. The first lemma is
n+1 = ∞; see [25, Chapter IV] for details.
standard; see, for example, [7, Proposition 1.17].
P∞
Lemma 3.3. Let γ ∈ F and let f ∈ D1. Then
f (γ+) ≤ f (γ−) + Λ(γ)f [1]γ ∗.
Our next lemma is an F -differentiability analogue of [6, Lemma 2].
7
Lemma 3.4. Let γ ∈ F , let M > 0, let m ∈ N, and let σ = γpl. Let
s0, . . . , sm−1 ∈ [0, Λ(σ)] such that 0 ≤ s0 < s1 < · · · < sm−1 < Λ(σ). Let
g ∈ Dm and suppose that g[m](σ(t)) ≤ M for all t ∈ [s0, Λ(γ)]. Then, for
all s ∈ (sm−1, Λ(γ)], we have
g(σ(s)) ≤
m−1
Xp=0
g[p](σ(sm−p−1))(s − sm−p−1)p
p!
+
M
m!
(s − s0)m.
Proof. Note that, for each a, b ∈ [0, Λ(γ)] with a < b, Λ(σ[a, b]) = b − a. We
prove the result by induction on m. If m = 1 then, by Lemma 3.3 applied to
σ[s0, s], we have
g(z) ≤ g(z0) + MΛ(σ[s0, s]) = g(z0) + M(s − s0).
Let m ∈ N and assume that the result holds for m − 1. Since g[1] ∈ Dm−1, it
follows that, for each s ∈ [sm−2, Λ(γ)], we have
g[1](σ(s)) ≤
m−2
Xp=0
g[p+1](σ(sm−p−2))(s − sm−p−2)p
p!
+
M
(m − 1)!
(s − s0)m−1.
Fix t ∈ (sm−1, Λ(γ)]. For each j ∈ {1, . . . , m − 1} let σj := σ[sj, t]. Then
Zσm−1
g[1](σ(s)) ds ≤Zσm−1
m−2
Xp=0
+Zσm−1
g[p+1](σ(sm−p−2))(s − sm−p−2)p
p!
ds
M
(m − 1)!
(s − s0)m−1 ds.
(2)
For each j ∈ {1, . . . , m − 2}, we have
Zσm−1
(s − sm−j−2)j−1 ds ≤Zσm−j−1
(s − sm−j−2)j−1 ds =
(s − sm−j−2)j
j!
, (3)
and, by combining (3) and (2), we obtain
Zσm−1
g[1](σ(s)) ds ≤
m−1
Xp=1
g[p](σ(sm−p−1))(s − sm−p−1)p
p!
+
M
m!
(s−s0)m. (4)
But now
g(σ(t)) ≤ g(zm−1) +Zσm−1
g[1](σ(s)) ds,
and combining this with (4), we obtain the desired result.
8
We now check an easy special case of our result.
Lemma 3.5. Let M = (Mn)∞
that lim inf n→∞ M 1/n
and f [k](z0) = 0 for all k ∈ N0. Then f (z) = 0 for all z ∈ Γ∗.
n=0 be a sequence of positive real numbers such
n < ∞. Let z0 ∈ Γ∗, and let f ∈ D with f [k]Γ∗ ≤ Mk
Proof. Fix z ∈ Γ∗. Let γ ∈ F such that γ− = z0, γ+ = z, and γ = γpl. We
first claim that, for each k, n ∈ N0 with k ≤ n, we have
f [k](z) ≤ Mn
sn−k
(n − k)!
,
(5)
where s = Λ(γ).
Fix n ∈ N0. We prove our claim by induction on k ≤ n. First suppose
that k = n so that n − k = 0. Let z ∈ γ∗. We have
f [k−1](z) ≤ f [k−1](γ+) +Zγ
f [k]γ ∗ ds =Zγ
f [k]γ ∗ ds,
and applying the claim to f [k](ζ), we have
f [k−1](z) ≤Zγ
Mnsn−k
(n − k)!
ds =
Mnsn−k+1
(n − k + 1)!
.
This proves the claim.
We now see that f (z) ≤ Mnsn/n! for all n ∈ N0 and all z ∈ γ∗. Now
n < ∞, there exists R > 0 such that Mn < Rn for
k=1 be a strictly increasing sequence in N
since lim inf n→∞ M 1/n
infinitely many n ∈ N. Let (nk)∞
such that Mnk < Rnk for all k ∈ N. Then, for each z ∈ γ∗, we have
f (z) ≤
Mnksnk
nk!
<
(Rs)nk
nk!
→ 0
as k → ∞. This holds for all z ∈ Γ∗, so the result follows.
Let 0 < α < 1. As in [6], we define B(α)
follows. For each j ∈ N0, let B(α)
each j, k ∈ N0 with k > j, define B(α)
0,j = 0 and, for each j ∈ N, let B(α)
j,k , for j, k ∈ N0 with k ≥ j ≥ 0, as
j,j = 1. For
j+1,k+1 inductively by setting
B(α)
j+1,k.
j+1,k+1 = B(α)
j,k+1 + αB(α)
Our main tool in the proof of the main theorem is the following lemma,
which can be distilled from the proof of [6, Lemma 1] and Stirling's approx-
imation. We omit the proof.
9
Lemma 3.6. Let α ∈ (0, 1/4e). Then there exists a constant K > 0 such
that B(α)
j,k ≤ Kα < 1/2 for all j, k ∈ N with j < k. Moreover, for each n ∈ N,
we have
((j + 1)α)j
j!
+ 2
(nα)n
n!
< 2.
n−1
Xj=0
Note that, if α ∈ (0, 1/4e), then
B(α)
j+1,k+1 ≥ B(α)
j,k+1
(6)
for all j, k ∈ N0 with k ≥ j.
We now state and prove our main result. The proof is essentially the one
used in [6], adapted for F -differentiation, and including additional details for
the convenience of the reader.
Theorem 3.7. Let Γ : [a, b] → C be an admissible path and let F denote
the collection of all subpaths and reverses of subpaths of Γ. Let (Mn)∞
n=0 be a
= ∞. Suppose that
sequence of positive real numbers such that P∞
F (Γ∗) and z0 ∈ Γ∗ satisfy
f ∈ D(∞)
n=1 M −1/n
n
f [k]Γ∗ ≤ Mk
and
f [k](z0) = 0
(7)
for all k ∈ N0. Then f (z) = 0 for all z ∈ Γ∗.
Proof. If lim inf n→∞ M 1/n
suppose that lim inf n→∞ M 1/n
sequence (M c
(nj)∞
j=0 with n0 = 0 such that:
n)∞
n < ∞ then the result follows from Lemma 3.5, so
n = ∞. By Lemma 3.2, there exists a log convex
n=0 of positive real numbers and a strictly increasing sequence
(a) M c
k ≤ Mk for all k ∈ N0;
(b) M c
nk = Mnk for all k ∈ N0;
(c) for each k ∈ N0, we have M c
j /M c
j+1 = M c
nk/M c
nk+1 for all j ∈ N0 with
nk ≤ j < nk+1.
By the comments following Lemma 3.2, we have
M c
n
M c
n+1
∞
Xn=0
= ∞.
Let z ∈ Γ∗ and let γ ∈ F such that γ− = z0, γ+ = z and σ = γpl.
Fix n ∈ N such that n is an element in the sequence (np)∞
j=0 M c
j /M c
that Pn
j+1 > 4eΛ(σ). Let α := Λ(σ)/(Pn
10
j=0 M c
j /M c
p=0 and such
j+1). Then we
have α ∈ (0, 1/4e) and so, by Lemma 3.6, B(α)
v > u > 0.
u,v < 1/2 for all u, v ∈ N with
Define the points 0 = x0 < x1 < · · · < xn ≤ Λ(σ) such that
xj − xj−1 = αM c
n−j/M c
n−j+1
j = 1, 2, . . . , n.
For each j ∈ {0, . . . , n} we claim that, for each k ∈ N0, with k ≤ n−j +1,
f [k](σ(t)) ≤ B(α)
j,n−k+1M c
k
(t ∈ [0, xj]).
(∗j
k)
The proof of the claim is by induction on j. Since f [k](σ−) = 0 for each
k ∈ N0, (∗0
k) holds for all k ∈ N0 with k ≤ n + 1.
Fix j ∈ {1, . . . , n}. Assume now that (∗j ′
k′) holds for all j′, k′ ∈ N0 with
j′ < j and k′ ≤ n − j′ + 1. Set i := n − j + 1. We now prove (∗j
k) holds
for each k ∈ N0 with k ≤ i by backwards induction on k. We first check the
base case. Suppose that k = i. If i = nr for some r ∈ N0, then B(α)
j,j = 1 and
f [i](σ(t)) ≤ Mi = M c
i for all t ∈ [0, xj] by (7), and so (∗j
k) holds.
Otherwise, i 6= nr for all r ∈ N0, in which case there exists r ∈ N0 such
that nr < i < nr+1 ≤ n. For each s ∈ [0, xj−1], by (∗j−1
i
) and (6), we have
f [i](σ(s)) ≤ B(α)
j−1,jM c
i ≤ B(α)
j,j M c
i = M c
i ,
and so it remains to show that f [i](σ(s)) ≤ M c
As in [6], let m := nr+1 − i and let R := M c
i for all s ∈ [xj−1, xj].
nr /M c
nr+1. Note that
j − m = n − nr+1 + 1 ≥ 1
and, for each p ∈ N0 with p ≤ m, we have M c
ℓ ∈ N with nr ≤ ℓ < nr+1, we have M c
In particular, M c
for each p ∈ N with p ≤ m.
s ∈ [xj−1, xj], by Lemma 3.4 (applied with M = M c
points s0 = xj−m, . . . , sm−1 = xj−1), we have
i+pRp. Also, for each
ℓ+1 = R and so xj−p − xj−p−1 = αR
i = M c
nr+1Rm. For each
nr+1 = Mnr+1, g = f [i] and
i = M c
ℓ /M c
f [i](σ(s)) ≤
≤
m−1
m−1
Xp=0
Xp=0
f [i+p](σ(xj−p−1))(s − xj−p−1)p
p!
f [i+p](σ(xj−p−1))((p + 1)αR)p
p!
11
+
M c
nr+1
m!
+
M c
nr+1
m!
(s − xj−m)m
(mαR)m
and, by applying (∗j−p−1
i+p
) for each 0 ≤ p ≤ m − 1, we obtain
f [i](σ(s)) ≤
B(α)
j−1−p,n−(i+p)+1M c
p!
i+p((p + 1)αR)p
+
M c
nr+1
m!
(mαR)m
m−1
Xp=0
i m−1
Xp=0
= M c
B(α)
j−1−p,j−p((p + 1)α)p
p!
+
(mα)m
m! ! ,
Since B(α)
u,v < 1/2 for all u, v ∈ N0 with v > u > 0, we have
f [i](σ(s)) ≤
M c
i
2 m−1
Xp=0
((p + 1)α)p
p!
+ 2
(mα)m
m! ! ,
and so, by Lemma 3.6, we have f [i](σ(s)) < M c
of the base case k = i.
i . This concludes the proof
Now let k ∈ N0 with k ≤ i − 1 and assume that (∗j
k+1) holds, i.e.,
f [k+1](σ(t)) ≤ B(α)
j,n−(k+1)+1M c
k+1
for all t ∈ [0, xj]. Let s ∈ [0, xj]. If s ∈ [0, xj−1] then, by applying (∗j−1
(6), we have
k
) and
f [k](σ(s)) ≤ B(α)
j−1,n−k+1M c
k ≤ B(α)
j,n−k+1M c
k.
Thus we may assume that s ∈ [xj−1, xj]. By Lemma 3.3, we have
f [k](σ(s)) ≤ f [k](σ(xj−1)) + (xj − xj−1) sup{f [k+1](σ(t)) : t ∈ [xj−1, xj]}.
Applying (∗j−1
obtain
k
) to the first term and applying (∗j
k+1) to the second term we
f [k](σ(s)) ≤ B(α)
j−1,n−k+1M c
k + α
B(α)
j,n−kM c
k+1.
n−j
M c
M c
n−j+1
Since M c is log-convex and k ≤ n − j, we have M c
so we obtain
n−j/M c
n−j+1 ≤ M c
k /M c
k+1,
f [k](σ(s)) ≤ M c
j−1,n−k+1 + αB(α)
k(cid:16)B(α)
j,n−k(cid:17) = B(α)
j,n−k+1M c
k .
Thus (∗j
k) holds, and both inductions may now proceed.
Now, by Lemma 3.6, there exists a constant K > 0 such that, for all
u, v ∈ N with v > u, we have Bu,v < αK. Thus f (σ(t)) ≤ KαM0 for all
t ∈ [0, Λ(σ)]. It follows that f (σ(t)) = 0 for all t ∈ [0, Λ(σ)].In particular,
f (z) = 0 and hence f (z) = 0. Since z ∈ Γ∗ was arbitrary, the above holds
for all z ∈ Γ∗. This completes the proof.
12
In the remainder of this paper we adopt the following convention. Let
X be a semi-rectifiable compact plane set, let F be an effective collection of
paths in X, and let f ∈ D(∞)
F (X). If there exists k ∈ N0 such that f [k]X = 0
then we write
f [j]
−1/j
X = +∞.
∞
Xj=1
Our first corollary will be used in the next section.
Corollary 3.8. Let γ be an admissible path in C, let F be an effective col-
lection of paths in γ∗, and let f ∈ D(∞)
F (γ∗). Suppose that
−1/j
f [j]
= +∞.
(8)
∞
Xj=1
If there exists z0 ∈ γ∗ such that f [k](z0) = 0 for all k ∈ N0 then f is identically
zero on γ∗.
Proof. For each k ∈ N0, set Mk = f [k]γ ∗. By (8), the sequence M = (Mk)∞
k=0
satisfies (1) and certainly f [k]γ ∗ ≤ Mk for all k ∈ N0. Suppose that there
exists z0 ∈ γ∗ such that f [k](z0) = 0 for all k ∈ N0. Then, by Theorem 3.7,
we have f (γ∗) ⊆ {0}. This completes the proof.
Our next corollary asserts the existence of an F -quasianalytic algebra of
the form DF (X, M).
Corollary 3.9. Let X be a semi-rectifiable compact plane set, let F be an
effective collection of paths in X, and let M = (Mn)∞
n=0 be an algebra sequence
which satisfies (1). Then DF (X, M) is F -quasianalytic.
Proof. Fix f ∈ DF (X, M). Then there exists N ∈ N large enough so that
f [k]X ≤ Mk for all k ∈ N0 with k ≥ N.
If there exists k ∈ N0 such that
f [k]X = 0, then (8) holds, so suppose that f [k]X > 0 for all k ∈ N0. Then
we have
∞
Xj=1
f [j]
−1/j
X ≥
f [j]
−1/j
X ≥
∞
Xj=N
∞
M −1/j
j
= ∞.
Xj=N
Thus, in either case, f satisfies (8). Thus, by Corollary 3.8, if there exist
γ ∈ F and z0 ∈ γ∗ such that f [k](z0) = 0 for all k ∈ N0, then f (γ∗) ⊆ {0}.
It follows that DF (X, M) is F -quasianalytic.
Since D(∞)(X) ⊆ D(∞)
F (X), Theorem 3.7 generalises [9, Theorem 1.11] to
obtain the following.
13
Corollary 3.10. Let M = (Mn)∞
n=0 be a sequence of positive real numbers
= ∞. Let γ be an admissible path in C. Suppose that
f ∈ D(∞)(X), f (k)(γ−) = 0 and f (k) ≤ Mk for k = 0, 1, 2, . . . . Then f is
identically zero on γ∗.
such that P∞
n=1 M −1/n
n
We conclude this section with a note about F -analyticity, as introduced
in [4] (see also [5]). Let X be a semi-rectifiable compact plane set, let F be
an effective collection of paths in X, and let f ∈ D(∞)
F (X). We say that f is
F -analytic if
k→∞ f [k]
k! !1/k
lim sup
< ∞.
This is a generalisation of the term analytic used in [16, 17], and is used to
find sufficient conditions for maps to induce homomorphisms between the
algebras DF (X, M). (Note that, in [15], the term analytic was used for those
functions on X which extend to be analytic on a neighbourhood of X. This
condition is stronger than in [16, 17].)
Let X be a semi-rectifiable compact plane set, let F be an effective col-
lection of paths in X, and let f ∈ D(∞)
F (X). Using Theorem 3.7, we can
show that if f is F -analytic then, for each γ ∈ F and z ∈ γ∗, there exist
r > 0 and an analytic (in the usual sense) function g : B(z, r) → C such that
g(γ∗ ∩ B(z, r)) = f (γ∗ ∩ B(z, r)). From this, it follows that in fact, for each
γ ∈ F , there exist an open neighbourhood U of γ∗ and an analytic function
h : U → C such that hγ∗ = f γ∗. We wish to thank Prof. J. K. Langley and
Dr. D. A. Nicks for showing us how to prove the latter implication.
4 Trivial Jensen measures without regularity
We conclude the paper with an application of the results from the previous
sections. We construct a locally connected compact plane set X and an
essential uniform algebra A on X such that A does not admit any non-trivial
Jensen measures but is not regular. This example will improve an example
of the first author ([13]).
We begin with the relevant definitions.
Definition 4.1. Let X be a compact Hausdorff space, let A be a uniform
algebra on X, and let ϕ be a character on A. A probability measure µ on X
is a Jensen measure for ϕ (with respect to A) if
log ϕ(f ) ≤ZX
log f dµ
(f ∈ A).
14
We say that A is regular on X if, for each closed set E ⊆ X and each point
x ∈ X \ E, there exists f ∈ A such that f (x) = 0 and f (E) ⊆ {0}. We say
that A is regular if the Gelfand transform of A is regular on the character
space ΦA of A. We say that A is essential if there exists no proper closed
subset E of X such that A contains every f ∈ C(X) such that f (y) = 0 for
all y ∈ E.
In the above definition we adopt the convention that log(0) = −∞. Let
X, A, ϕ be as in the above definition. It is standard that every Jensen measure
for ϕ is a representing measure for ϕ. Moreover, for each ϕ ∈ ΦA, there is a
Jensen measure on X for ϕ. Note that, for x ∈ X, the point-mass measure δx
is a Jensen measure for εx, where (here, and for the remainder of the section)
εx is the evaluation character at x (with respect to A). We say that a Jensen
measure µ on X for εx is trivial if µ = δx.
Let X be a compact plane set. Let R0(X) denote the set of restrictions
to X of rational functions with no poles on X. Let R(X) denote the uniform
closure of R0(X) in C(X). It is standard that R0(X) is an algebra and that
R(X) is a natural uniform algebra on X. For the remainder of this section,
all Jensen measures will be with respect to R(X) unless otherwise specified.
For further details on uniform algebras, Jensen measures, and related
topics, see [3, 8, 22, 23, 28].
Let x ∈ X. Let Jx denote the ideal in R(X) of all functions which vanish
on a neighbourhood of x. Let Mx denote the ideal in R(X) of all functions
which vanish at x. Clearly Jx ⊆ Mx. We say that x is a point of continuity
(for R(X)) if, for all y ∈ X \ {x} we have Jy * Mx. We say that x is an
R-point if, for all y ∈ X \ {x}, we have Jx * My. For further information
see [12, 13, 18, 19]. (Note that in [12] points of continuity are referred to as
regularity points of type one and R-points are referred to as regularity points
of type two.)
It is standard that R(X) is regular if and only if every point of X is
a point of continuity, and this holds if and only if every point of X is an
R-point. It is also standard that if x is a point of continuity then the only
Jensen measure for εx is the point mass measure.
Let X be a topological space and let E be a subset of X. We denote
In
by intX (E) the interior of E with respect to the topological space X.
particular, if E ⊆ C then intC E coincides with the usual interior of E.
For the remainder of this section, we denote the set of non-negative real
numbers by R+. Let X be a metric space, let x ∈ X, and let r > 0. We
denote the open ball in X with centre x and radius r by BX(x, r). We the
denote the corresponding closed ball by ¯BX (x, r).
In the special case where X = C, for each a ∈ C and r > 0, we write
15
B(a, r) = BC(a, r) and ¯B(a, r) = ¯BC(a, r). For each a ∈ C, we set B(a, 0) = ∅
and ¯B(a, 0) = {a}.
Lemma 4.2. Let X, Y be compact plane sets with Y ⊆ X. Suppose that
R(Y ) is regular. Then each point y ∈ intX (Y ) is a point of continuity for
R(X).
Proof. Let y ∈ intX (Y ), and let x ∈ X with x 6= y. Then there exists
r > 0 such that ¯BX (y, r) ⊆ Y . Choose δ ∈ (0, r/3) such that y − x > 2δ.
Since R(Y ) is regular, it follows (from [14, lemma 1.6], for example) that
R(Y ∩ ¯B(y, r)) is regular. Set E := ¯BX (y, r) \ BX (y, δ). Then there exists a
function g ∈ R(Y ∩ ¯BX (y, r)) such that f (y) = 1 and f (x) = 0 for all x ∈ E.
Let f ∈ C(X) be given by f (z) = g(z) for all z ∈ X ∩ ¯BX(y, r) and f (z) = 0
for all z ∈ X \ ¯BX(y, r). It follows from [24, Corollary II.10.3] that f ∈ R(X)
and clearly f vanishes on a neighbourhood of x. Thus Jx * My (where these
are the ideals in R(X)). It follows that y is a point of continuity for R(X)
and so the proof is complete.
The following is effectively [19, Lemma 3.1].
Lemma 4.3. Let X be a compact plane set and let x ∈ X. Suppose that
there exists a neighbourhood U of x in X such that every point in U \ {x} is
a point of continuity. Then x is an R-point.
As in [21] (see also [20]), we define abstract Swiss cheeses as follows.
Definition 4.4. An abstract Swiss cheese is a sequence A = ((an, rn))∞
elements of C × R+. Let A = ((an, rn))∞
n=0 of
n=0 be an abstract Swiss cheese. Set
XA := ¯B(a0, r0) \ ∞
[n=1
B(an, rn)! ,
and set ρ(A) = P∞
and for all k ∈ N with rk > 0 the following hold:
n=1 rn. We say that A is classical if ρ(A) < ∞, r0 > 0
(a) ¯B(ak, rk) ⊆ B(a0, r0);
(b) whenever ℓ ∈ N with rℓ > 0 and ℓ 6= k, we have ¯B(ak, rk)∩ ¯B(aℓ, rℓ) = ∅.
We say that a compact plane set X is a Swiss cheese set if there exists
an abstract Swiss cheese A such that X = XA. We say that a Swiss cheese
set X is classical if there exists a classical abstract Swiss cheese A with
X = XA. If X is a classical Swiss cheese set then X is a uniformly regular
(see the proof of [10, Theorem 8.3]) and R(X) is essential (see [3, p. 167] or
16
[14, Theorem 1.8]). It follows that if A is classical then XA is also connected
and locally connected.
In [13], the first author gave an example of a non-classical Swiss cheese
set X such that R(X) has no non-trivial Jensen measures, but such that
R(X) is not regular. We shall show that there is a classical Swiss cheese set
with these properties; this is the content of the following theorem, which is
the main theorem of this section.
Theorem 4.5. There exists a classical abstract Swiss cheese A = ((an, rn))
such that R(XA) is not regular and does not admit any non-trivial Jensen
measures.
Most of the remainder of this section is devoted to the proof of this
theorem. We require some preliminary results. The following proposition is
[19, Lemma 4.1].
Proposition 4.6. Let X be a compact plane set, let Y be a non-empty closed
subset of X, and let x ∈ Y . Suppose that no bounded component of C \ Y
is contained in X, and that there exists a non-trivial representing measure µ
for εx with respect to R(X) such that supp µ ⊆ Y . Then µ is a non-trivial
representing measure for εx with respect to R(Y ), and R(Y ) 6= C(Y ).
Note that, if intC X = ∅ then the condition on bounded components of
C \ Y is automatically satisfied.
Let X be a compact plane set with intC X = ∅, let x ∈ X, and let µ
be a non-trivial representing measure for εx. Let Y = supp µ ∪ {x}, where
supp µ denotes the closed support of µ. Then, by the above, we must have
R(Y ) 6= C(Y ). In particular, as noted in [19], Y must have positive area.
(See also the Hartogs-Rosenthal theorem [22, Corollary II.8.4].) Combining
these observations with Proposition 4.6 gives the following corollary, which
we use below.
Corollary 4.7. Let X be a compact plane set with intC(X) = ∅, let E be a
closed subset of X, and let x ∈ E. Suppose that E has area 0, that µ is a
Jensen measure for εx with respect to R(X), and µ is supported on E. Then
µ is trivial.
We also require the following lemma, which is a special case of [19,
Lemma 2.1].
Lemma 4.8. Let X be a compact plane set and let x ∈ X. Suppose that
µ is a non-trivial Jensen measure for x, and let F be the closed support of
µ. Then, for all y ∈ F \ {x}, we have Jy ⊆ Mx. Thus x is not a point of
continuity and no point of F \ {x} is an R-point.
17
The following estimates on derivatives are standard. See, for example [19,
Lemma 4.4]. (This result also appears in [13] but with some typographical
errors.)
Lemma 4.9. Let A = ((an, rn))∞
n=1 be an abstract Swiss cheese, and let
z ∈ C. For each n ∈ N, let dn denote the distance from B(an, rn) to z. Let
d0 = r0 − z − a0. Suppose that dn > 0 for all n ∈ N0. Then z ∈ XA and,
for all f ∈ R0(XA) and k ∈ N0, we have
f (k)(z) ≤ k!
∞
Xj=0
rj
dk+1
j
f XA.
Our construction will use the following proposition, which is a combina-
tion of [21, Lemma 8.5] and, for example, [14, Example 2.9].
Proposition 4.10. Let a ∈ C, λ1 ≥ 0, λ0 > λ1, and ε > 0. Then there exists
a classical abstract Swiss cheese A = ((an, rn))∞
n=0 such that a0 = a1 = a,
r0 = λ0 and r1 = λ1 such that R(XA) is regular and P∞
Note that, since R(XA) is regular, we must have intC XA = ∅.
We now give the details of the construction.
n=2 rn < ε.
Lemma 4.11. Let 0 < r < 1 be given, let Cr denote the circle of radius r
centred at 0, and let ε > 0. Then there exists a classical abstract Swiss cheese
A = ((an, rn)) such that
(a) ρ(A) < ε, intC(XA) = ∅, and Cr ⊆ XA,
(b) there is a dense open subset U of XA such that XA \ U has area zero
and each point z ∈ U is a point of continuity for R(XA),
(c) for each f ∈ R(X), we have f Cr ∈ D(∞)(Cr) andP∞
Proof. Our abstract Swiss cheese A will be obtained by combining a certain
pair of sequences (An), (Bn) of abstract Swiss cheeses in a suitable way. We
first construct the sequences (An), (Bn).
k=1 f (k)
−1/k
Cr = ∞.
Choose a positive integer n0 large enough so that r + 21−n0 < 1 and
r − 21−n0 > 0. As in [19], choose a sequence (γn) of positive real numbers
such that, for each n, k ∈ N, we have
γn ≤
,
(21−n0−n)k+1(log (k + 3))k
2k
18
n=1 γn < ε. Thus, for each k ∈ N,
γn
(21−n0−n)k+1 ≤ (log (k + 3))k.
∞
Xn=1
and such that P∞
Let A1 = ((a(1)
n , r(1)
n , r(1)
n )) be the classical abstract Swiss cheese obtained from
Proposition 4.10 applied with λ0 = r − 21−n0, λ1 = 0, a = 0 and ε = γ1/2.
Let B1 = ((b(1)
n )) be the classical abstract Swiss cheese obtained from
Proposition 4.10 applied with a = 0, λ0 = 1 and λ1 = r + 21−n0 and ε = γ1/2.
n )) be the classical abstract
Swiss cheese obtained from Proposition 4.10 applied with λ0 = r − 22−n0−k,
λ1 = r − 23−n0−k, a = 0 and ε = γk/2; let Bk = ((b(k)
n )) be the classical
abstract Swiss cheese obtained from Proposition 4.10 applied with a = 0,
λ0 = r + 23−n0−k and λ1 = r + 22−n0−k and ε = γk/2.
For each k ∈ N with n ≥ 2: let Ak = ((a(k)
n , r(k)
n , r(k)
Let a0 = 0 and r0 = 1 and let ((an, rn))∞
n=1 be an enumeration (without
repeats) of the set
{(a(m)
n , r(m)
n ), (b(m)
n , s(m)
n )}.
[m,n∈N
n≥2
Then A = ((an, rn))∞
n=0 is an abstract Swiss cheese with Cr ⊆ XA. It is not
hard to see that A is classical and ρ(A) < ε. It is clear that intC(XA) = ∅.
Also, for each n ∈ N, let
0 , r(n)
0 ) ∪ ∂B(a(n)
1 ) ∪ ∂B(b(n)
En := ∂B(a(n)
0 ) ∪ ∂B(b(n)
1 , s(n)
1 ),
0 , s(n)
1 , r(n)
(where ∂S denotes the boundary of S ⊆ C) and set E := Cr ∪S∞
n=1 En. Set
X := XA and U := X \ E. It is easy to see that E is a closed set and that
U is a dense open subset of X. Moreover, E has area zero .
We claim that each point x ∈ U is a point of continuity for R(X). Let
x ∈ U. Then there exists a unique n ∈ N such that either x ∈ XAn or
x ∈ XBn. If x ∈ XAn, set Y = XAn, and if x ∈ XBn then set Y = XBn. By
our construction R(Y ) is regular and it is not hard to see that x ∈ intX Y .
Thus, by Lemma 4.2, x is a point of continuity for R(X). This proves the
claim.
It remains to show that (c) holds. We first consider functions in R0(X).
Let z ∈ Cr and let f ∈ R0(X). For each n ∈ N, let dn the distance from
¯B(an, rn) to Cr, and let d0 = 1 − r. Since each An and each Bn are classical,
and since Cr * XAn, XBn for all n ∈ N, it follows that dn > 0 for all n ∈ N0.
By Lemma 4.9, for each k ∈ N0, we have
∞
rj
dk+1
j
f X
f (k)(z) ≤ k!
Xj=0
19
Fix m ∈ N. Then there exists a unique n ∈ N such that there exists ℓ ∈ N
with (am, rm) = (a(n)
, s(n)
ℓ ).
In either case, since An and Bn are classical, we have dm > 23−n0−m. Thus
ℓ ) or there exists ℓ ∈ N with (am, rm) = (b(n)
, r(n)
ℓ
ℓ
k!
∞
Xj=0
rj
dk+1
j
f X ≤ k!f X 1
dk+1
0
+
γj
(21−n0−j)k+1! .
∞
Xj=1
for each k ∈ N0. It follows that, for each k ∈ N0, we have
f (k)Cr ≤ k!f X(cid:18) 1
dk+1
0
+ (log (k + 3))k(cid:19) .
From this we deduce that, for each f ∈ R(X) (not necessarily in R0(X)), we
have f Cr ∈ D(∞)(Cr) and the same estimates hold. (One way to see this is
to note that Cr is uniformly regular, and apply [10, Theorem 5.6].)
Now let f ∈ R(X). As in [19], choose N ∈ N large enough so that
(log (k + 3))k ≥ 1/dk+1
0
for all k ∈ N with k ≥ N. Then we have
∞
Xj=1
1
f (j)1/j
Cr
≥
∞
Xj=N
1
(2f X)1/jj log (j + 3)
= ∞,
and so (c) holds. This completes the proof.
We are now ready to prove Theorem 4.5.
Proof of Theorem 4.5. Apply Lemma 4.11 with r = 1/2 and ε = 1 to obtain
a classical abstract Swiss cheese A which satisfies properties (a) -- (c) described
in the statement of the lemma, and let X := XA. Then, for each f ∈ R(X),
we have f Cr ∈ D(∞)(X) and
f (n)
−1/n
Cr = ∞.
∞
Xn=1
So, by Corollary 3.8, if f ∈ R(X) for which f (k)(z) = 0 for some z ∈ Cr
and all k ∈ N0 then f is identically zero on Cr. It follows that no point of
Cr can be a point of continuity for R(X) and therefore R(X) is not regular.
It remains to see that R(X) has no non-trivial Jensen measures on X. By
Lemma 4.11(b) there is a dense open subset U of X for which every point
z ∈ U is a point of continuity for R(X) and such that X \ U has area 0. Set
E := X \ U. By Lemma 4.3, every point of U is also an R-point. Let x ∈ Cr
and let µ be a Jensen measure for εx. Then since x /∈ U, by Lemma 4.8,
supp µ ∩ U = ∅ and so supp µ ⊆ E. Since intC(X) = ∅ and the area of E is
0, it follows from Corollary 4.7 that µ must be trivial. This completes the
proof.
20
It is also possible to show that R(XA) admits no non-trivial Jensen mea-
sures by appealing to the theory of Jensen interior. (See, for example, [24,
p. 319].) This is the approach used in [13].
Our final corollary follows immediately from Theorem 4.5.
Corollary 4.12. There exists a locally connected compact plane set X such
that R(X) is essential, non-trivial and non-regular and yet R(X) admits no
non-trivial Jensen measures.
We conclude with some open questions.
Question 4.13. Is the uniform algebra R(XA) constructed in Theorem 4.5
necessarily antisymmetric? If not, can the construction be modified to yield
an example which has the properties in that theorem and is also antisym-
metric?
Question 4.14. Let A be a uniform algebra on a compact Hausdorff space X,
and let Mi (i ∈ I) be the decomposition of X into maximal A-antisymmetric
subsets.
(a) Suppose that AMi is regular on Mi for all i ∈ I. Must A be regular
on X? What if we assume the stronger condition that AMi is regular
(so natural and regular on Mi) for all i ∈ I?
(b) What is the answer to (a) in the special case where X is a compact
plane set and A = R(X)?
References
[1] T. M. Apostol. Mathematical analysis. Addison-Wesley Publishing Co.,
Reading, Mass.-London-Don Mills, Ont., second edition, 1974.
[2] W. J. Bland and J. F. Feinstein. Completions of normed algebras of
differentiable functions. Studia Math., 170(1):89 -- 111, 2005. ISSN 0039-
3223.
[3] A. Browder. Introduction to function algebras. W. A. Benjamin, Inc.,
New York-Amsterdam, 1969.
[4] T. Chaobankoh. Endomorphisms of Banach function algebras. PhD
thesis, University of Nottingham, 2012.
[5] T. Chaobankoh, J. F. Feinstein, and S. Morley. The chain rule for F -
differentiation. To appear in Irish Math. Soc. Bull.
21
[6] P. J. Cohen. A simple proof of the Denjoy-Carleman theorem. Amer.
Math. Monthly, 75:26 -- 31, 1968. ISSN 0002-9890.
[7] J. B. Conway. Functions of one complex variable, volume 11 of Graduate
Texts in Mathematics. Springer-Verlag, New York, second edition, 1978.
[8] H. G. Dales. Banach algebras and automatic continuity, volume 24 of
London Mathematical Society Monographs. New Series. The Clarendon
Press Oxford University Press, New York, 2000.
[9] H. G. Dales and A. M. Davie. Quasianalytic Banach function algebras.
J. Functional Analysis, 13:28 -- 50, 1973.
[10] H. G. Dales and J. F. Feinstein. Normed algebras of differentiable func-
tions on compact plane sets. Indian J. Pure Appl. Math., 41(1):153 -- 187,
2010.
[11] H. Federer. Geometric measure theory. Die Grundlehren der mathema-
tischen Wissenschaften, Band 153. Springer-Verlag New York Inc., New
York, 1969.
[12] J. Feinstein and R. Mortini. Partial regularity and t-analytic sets for
Banach function algebras. Math. Z., 271(1-2):139 -- 155, 2012.
[13] J. F. Feinstein. Trivial Jensen measures without regularity. Studia
Math., 148(1):67 -- 74, 2001. ISSN 0039-3223.
[14] J. F. Feinstein and M. J. Heath. Swiss cheeses, rational approximation
and universal plane curves. Studia Math., 196(3):289 -- 306, 2010.
[15] J. F. Feinstein and H. Kamowitz. Endomorphisms of Banach algebras
of infinitely differentiable functions on compact plane sets. J. Funct.
Anal., 173(1):61 -- 73, 2000. ISSN 0022-1236.
[16] J. F. Feinstein and H. Kamowitz. Compact endomorphisms of Banach
algebras of infinitely differentiable functions. J. London Math. Soc. (2),
69(2):489 -- 502, 2004. ISSN 0024-6107.
[17] J. F. Feinstein and H. Kamowitz. Compact homomorphisms between
Dales-Davie algebras. In Banach algebras and their applications, volume
363 of Contemp. Math., pages 81 -- 87. Amer. Math. Soc., Providence, RI,
2004.
[18] J. F. Feinstein and D. W. B. Somerset. Non-regularity for Banach func-
tion algebras. Studia Math., 141(1):53 -- 68, 2000. ISSN 0039-3223.
22
[19] J. F. Feinstein and H. Yang. Regularity points and jensen measures for
R(X). ArXiv e-prints: 1507.01779.
[20] J. F. Feinstein, S. Morley, and H. Yang. Swiss cheeses and their applica-
tions. In Function Spaces, volume 645 of Contemp. Math. Amer. Math.
Soc., Providence, RI, 2015.
[21] J. F. Feinstein, S. Morley, and H. Yang. Abstract Swiss cheese space and
classicalisation of Swiss cheeses. J. Math. Anal. Appl., 438(1):119 -- 141,
2016.
[22] T. W. Gamelin. Uniform algebras. Prentice-Hall Inc., Englewood Cliffs,
N. J., 1969.
[23] T. W. Gamelin. Uniform algebras and Jensen measures, volume 32 of
London Mathematical Society Lecture Note Series. Cambridge Univer-
sity Press, Cambridge-New York, 1978.
[24] T. W. Gamelin and T. J. Lyons. Jensen measures for R(K). J. London
Math. Soc. (2), 27(2):317 -- 330, 1983. ISSN 0024-6107.
[25] P. Koosis. The logarithmic integral. I, volume 12 of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, Cambridge, 1998.
Corrected reprint of the 1988 original.
[26] S. Mandelbrojt. S´eries adh´erentes, r´egularisation des suites, applica-
tions. Gauthier-Villars, Paris, 1952.
[27] W. Rudin. Real and complex analysis. McGraw-Hill Book Co., New
York, third edition, 1987.
[28] E. L. Stout. The theory of uniform algebras. Bogden & Quigley, Inc.,
Tarrytown-on-Hudson, N. Y., 1971.
School of Mathematical Sciences, The University of Nottingham, University
Park, Nottingham, NG7 2RD, UK
Email address: [email protected]
Email address: [email protected]
23
|
1208.2423 | 1 | 1208 | 2012-08-12T11:32:13 | Preliminaries on best proximity points in cyclic multivalued mappings | [
"math.FA"
] | This paper investigates the fixed points and best proximity points of multivalued cyclic self-mappings in metric spaces under a generalized contractive condition involving Hausdorff distances. | math.FA | math | PRELIMINARIES ON BEST PROXIMITY POINTS IN CYCLIC MULTIVALUED MAPPINGS
M. De la Sen
*Institute of Research and Development of Processes. University of Basque Country
Campus of Leioa (Bizkaia) - Aptdo. 644- Bilbao, 48080- Bilbao. SPAIN
email: [email protected]
Abstract: This paper investigates the fixed points and best proximity points of multivalued cyclic self-mappings in
metric spaces under a generalized contractive condition involving Hausdorff distances.
1. Introduction
Important attention is being devoted along the last years to investigate fixed point theory for multivalued
mappings concerning some relevant properties like, for instance, stability of the iterations, fixed points of
contractive and nonexpansive self-mappings and the existence of either common or coupled fixed points
of several multivalued mappings or operators. This paper investigates some properties of fixed point and
best proximity point results for multivalued cyclic self- mappings under a general contractive-type
condition based on the Hausdorff metric between subsets of a metric space and which includes a
particular case the contractive condition for contractive single-valued self-mappings, including the
problems related to cyclic self-mappings. This includes strict contractive cyclic self -mappings and Meir-
Keeler type cyclic contractions. Through this paper, we consider a metric space
dX ,
and a multivalued
BAT
BA
(simply referred to as a multivalued cyclic self-mapping in the
2-cyclic self-mapping
:
A
B
A
B
numbers
real
of
set
and
sequel) where A and B are nonempty closed subsets of X , so that
0
BA
subset
consider
us
. Let
,
R
R
R
R
R
0
0
disjunction (“or”) and conjunction (“and”), and define the functions:
AM
:
R,
B
10A
B
0
B
,
10A
B
, let the symbols
"" and
"" denote the logic
10
,
(2.1)
and
the
the
dist
of
:0
:0
D
z
z
A
z
z
T
T
,
,
,Ty,TxM,K,Ty,TxMmax
,
1
2
Ty,xd
Ty,yd
Tx,xd
/
21,
,
,
Tx,yd
Tx,xd
,
Ty,yd
(2.2)
;
y,x
A
B
B
,
,0
,0
:
1
1
as follows:
,
,K,Ty,TxM
yxd
Kxma
max
,
,K,y,x
,
1
1
1
1
1
if
1
TxM
TxM
Ty
KTy
,
,
,
,
,
,
,
1
2
2
1
Ty
KTy
TxM
TxM
if
,
,
,
,
,
,
3
2
1
TxM
Ty
KTy
TxM
,
if
,
,
,
,
,
4
2
1
K
TxM
/
TxM
Ty
0
21
,
,
2
1
Ty
TxM
TxM
/
K
,
,
21
1
1
2
10
,
,
for some real constants
, where
KTy
,
KTy
,
if
if
,
10
,K
,
,
,
,
K
A
(2.3)
,
:
,
1
1
1,210
,
/
,
:
3
1
;
,
:
,
1
2
1
2
1
-1
4
2
1,210
,
/
,
:
,
-1
2
1
1
1
(2.4)
10
A
B
,
,
Note that
is non-increasing since all its partial derivatives with
10A
B
A
,K
,
A
B
exist and are non-positive;
B
XCL
be the family of all nonempty and closed
HXCL
,
R
:
0
are disjoint by construction. Let
XCL
equipped with the Hausdorff metric
subsets of the vector space X . If
hyperspace of
dX ,
then we can define
XCLH
:
and note also that the subsets
being the generalized
respect to
;
i
,
,
,
4321
BA ,
y,x
i
sup
By
Ayd
,
(2.5)
max
sup
Ax
B,AH
,Bxd
,
The distance between A and B is
D
yxd
,
BA
,
dist
inf
By,Ax
inf
Ax
Bxd
,
Ayd
,
inf
By
(2.6)
2. Main results
The main result generalizes the contractive condition of multivalued self-mappings of [3] based on
Hausdorff´s generalized metric. It includes, in particular, a contractive condition for multi-valued self-
mappings. The main result follows:
dX ,
Theorem 2.1: Let
multivalued cyclic self-mapping, where
X
B,A are nonempty, closed and subject to the contractive
be a complete metric space and let
BAT
BA
:
be a, in general,
constraint through
BAT
BA
:
:
for some
,
,K,y,x
Tx,xd
R ,
K
y,xd
Ty,TxH
,
and
10 ,
;
D
,K,Ty,TxM
,
(2.7)
AB
BA
. Assume also
y,x
that
K
1
max
K
,
,
1
1
10
,
,
K
2
max
1
1
,
,
1
1
1
(2.8)
Then, the following properties hold:
(i) There is sequence nx
in
BA satisfying
xdD
,1
n
x
n
;
D
sup
lim
n
xd
x,
n
n
1
Ni
such that
Tx
x 1
,
i
i
DK
2
1 K
1
2
If A and B are bounded sets which intersect then
n
1
Nn
Tx
;
1
n
n
2
1 x,xd
bounded then the above property still holds if
then bounded of limit in
BA , with
x
.
x
xd
,1
n
n
for any given
and
nx
BA
x
1
is a Cauchy sequence,
. If A and B are not
n
n
1
1
x,
if
xd
D
x
BA
1
for any given
with the sequence
K
1
lim
K
n
2
A
then the sequence of sets
A
x
Tx
x 1
Tx n
BT
such that
converges asymptotically
. If
1
1
n
n
2
to a subset
Az
of best proximity points in A and the sequence of sets
B
Tx n
AT
2
B
to a subset
z
AT
zT
in B with
asymptotically
of best proximity points
A
BT
zT
.
B
A
BA
0D
, i.e. if
If
, and any sequence
being constructed
converges
then
0
nx
xd
x,
x,xd
nm
nx
1
A
B
z
n
n
lim
sup
nm
n
lim
n
to a fixed point
x
being iteratively generated as
1
n
n
BA
Tz
of
A and B are, in addition, convex and that
Tx
z
, for any
x
BAx
1
BAT
BA
. Assume that
:
Ni
,
21
;
i Tz
z
i
Tz
BA
and
Tz
;
j,i
,
i
j
BAT
BA
:
. Then,
z
i
z
j
BA
, is a Cauchy sequence which converges
, i.e.
, that
BA
0D
N,
....
are fixed points of
N,
i
N
,
21
, that
....
,
is, the image sets of any fixed points are identical .
(ii) Consider a uniformly convex Banach space
d,X
, so that
R
, and let A and B be nonempty, disjoint, convex and closed subsets of
XXd
induced metric
:
0
is a metric space for the norm -
,X
X with
BAT
BA
:
satisfying the contractive condition (2.7)-(2.8) with
with
x
1
Tx
2
n
2
2
n
xd
nx2
sequence
and a Cauchy sequence in B if
built such that
x
n Tx
n
2
2
1
B
x 1
so that
,
2
x
lim
n
D
;
and
x
BA
x
xd
lim
lim
xd
x
,
,
1
n
n
n
1
2
2
1
2
n
n
and
the sequences of sets
1
T n
2
1
TT
T
x
x
1
in A and B , respectively.
B Tz
A Tz
z
z
and
A
B
1
n
n
2
2
n
2
2
n
2
3
n
1
lim
n
x 1
A
. Then, a
1
K
1
K
2
is a Cauchy sequence in A if
0
xd
;
x
,
2
n
2
2
n
x 1
x
BA
1
A
,
2
x
and
Tx
1
. If
1
x
converge to best proximity points
then
AT
B
Proof: a) Take with no
in generality
loss
yxd
Tx,xd
Tx,xd
,K,y,x
,
,
then,
KTy
TxM
TxM
,
,
,
1
2
Ty
,
,
Ax and
since
Tx
y being arbitrary and assume
10 ,
,K,y,x
,
Assume
.
that
that
Ty,yd
TxM
2
,
Ty
K
max
Ty,ydyxd
,
,
D
Tx,xd
Ty,xd
2
,
Ty,yd
KD
max
,
TxMD
1
Ty,xd
2
,
yxd
,
D
D
(2.9)
Ty
3
since
K
10 ,
what also implies that
Ty,yd
Ty,yd
min
1
xd
,
K,Tx
max
yxd
,
,
1
1
Ty,xd
2
D
1
xT,xd
D
. Then,
min
since
xd
,
Tx
1
y,xd
xd
,
yxdK,Tx
,
max
K
,
1
yxd
,
D
1
(2.10)
;
y
Tx
.
b) Now, assume that
TxM
2
,
Ty
,
,
TxM
1
,
KTy
,
Ty,yd
TxM
1
,
Ty
K
TxM
2
,
Ty
max
Tx,xd
Ty,ydyxd
,
,
,
Ty,yd
D
so that
Ty,xd
2
K
max
yxd
,
(2.11)
D
,
Ty,xd
2
This implies also that
Ty,yd
1
Tx,xd
D
1
and again (2.10) holds. As a result,
,
1
K
;
max
D
1
yxd
,
Ty,yd
by interchanging the roles of A and B , one also gets by proceeding in a similar way:
D
1
ByAx
,
ByAx
,
Tx,xd
yxd
,
max
;
K
1
,
Thus,
Tx,xd
max
K
,
,
1
1
yxd
,
max
1
1
DKyxdKD
,
,
1
2
1
1
(2.12)
;
y,x
BA
BAT
,
where
K
1
max
K
,
,
1
1
10
,
and
. Note that, since
BAT
BA
:
is cyclic, then
Tx,y
B
if
Ax and
1
1
,
1
2
1
K
max
conversely.
Now,
construct
AT
Tx
x
2
n
1
2
n
sequence
a
B
,..,
x
2
n
1
nx
Tx
2
n
in
BT
BA as
A
follows
x
1
which satisfies:
x
A
,
x
2
Tx
1
AT
B
,….
xdD
,
x
n
xdK
1
n
,
x
n
1
n
KDK
1
2
1
x,xd
1
2
n
1
n
2
i
0
i
DKK
1
2
;
Nn
(2.13)
1
n
K
1
x,xd
2
1
Then,
D
sup
lim
n
xd
x,
n
n
1
Nn
1
;
1
DK
2
n
K
1
K
1
1
DK
. On the other hand,
2
1 K
1
j
n
1
xd
,
x
n
n
1
j
n
1
n
K
1
1
xd
,
x
1
2
1
n
K
1
1
K
1
1
DK
2
4
1
K
1
1
so that
1
j
xdK
1
,
x
1
2
j
n
1
1
1
n
K
1
DK
2
;
Nn
(2.14)
,
n
n
x
1
n
1
xd
xd
1
K
1
1
and we conclude that
nx
is a Cauchy sequence if
1
2 , x
xd
bounded or simply if
), since
n
1
n
K
1
1
x
1
2
,
1
DK
2
;
Nn
(2.15)
0D
x,xd
nm
(i.e. if A and B intersect provided that they are
0
xd
, which has a limit
x,
1
n
n
lim
sup
nm
n
lim
n
dX ,
z in X , since
B are both nonempty and closed since
is complete, which is also in
B
AT
and
BA which is nonempty and closed since A and
A
BT
. On the other hand, for any
distance
0D
between A and B :
xdD
2
n
3
,
x
2
n
2
2
K
1
n
1
x,xd
1
2
n
2
i
0
i
K
1
2
KD
1
n
1
x,xd
2
1
1
1
1
n
2
K
1
K
1
DK
2
;
Nn
(2.16)
xdD
,
x
2
n
1
n
2
x,xdK
1
1
2
2
n
2
n
2
1
i
0
i
K
1
n
2
x,xdKD
1
2
1
;
Nn
(2.17)
DK
2
n
2
K
1
1
K
1
1
DKK
2
1
3
2
n
n
n
x
K
1
2
K
1
1
1
1
2
K
1
1
K
1
1
are bounded if 1x and
K
1
1
K
2
DK
2
3
DKK
2
1
;
Nn
(2.18)
2 Tx
x
1
are such that
0
xd
,
x
2
n
1
,
x
n
2
2
xdK
1
1
,
x
2
n
2
n
2
K
1
n
2
3
3
xdK
1
n
2
x,xdK
3
1
1
Note that the sequences
n x,xd
,
,
2
2
2
n
n
x
1
1
2
xdKD
n
1
D
4
2
xdK
1
1
n
2
x,xdKD
1
1
3
K
n
1
2
i
0
1n
n x,xd
2n
and
i
K
1
2
1 x,xd
2
what is always guaranteed if A and B are bounded. If
then one gets from the
above relations that
xd
x
,
xd
D
,
2
2
2
3
2
2
2
n
n
n
n
x
x
A
A
A
B
1
Tx
Tx
and
and,
some
R
lim
n
x
,
lim
n
Tx
x
where
n
n
n
n
n
n
2
2
1
2
1
2
2
2
3
2
2
nx2
1
and
2 nx
contain asymptotically best proximity points of A and B , respectively, if
conversely, of B and A if
x 1
B
. This follows by contradiction since, if not, for each
, some subsequence
Nj
k
kjn
of
of natural numbers with
. Thus, any sequences of sets
x 1
, there exists
real
subsequences
related
D
xd
xd
1
1
0D
and consider separately the various cases in (2.3)-(2.4), by using the contractive
Now, assume
BA to which all sequences
z in
Tz
condition (2.7), subject to (2.2), to prove that there exists
with
nx
x,xd
lim
sup
D
lim
xd
x,
xn
BAz
converge by using
0
0
nm
nm
n
n
being a Cauchy sequence since
d,X
is complete and A and B are nonempty and closed.
jm ,
k
n
n
for
jk
mk
kjnx2
and
1
kjnx
2
is impossible.
so that
kn
numbers
Nk
some
such
and
that
as
n
kj
n
kj
D
n
k
n
k
1
x
x
n
n
2
2
2
2
2
2
,
,
5
Case 1:
,
Then,
Tz,zd
1
,K,y,x
,
Tz,zd
,
,
1
2
zd
,K,z,z
,
Tz,zd
,
Tz,zd
if
,
,
1
holds
0
contradiction
with
since
,
1
Tz,zd
Tz,zd
zd
Tz
z . Hence,
if
,
z if
is similar since
Tz
,
,
2
2
2 , and the fact that distances have the symmetry property.
z if
Tz
,
1
closed.
Tz is
and
Tz
If
Tz
KTz
,
,K,Tz,TzM
TzM
Tz
,
,
,
,
2
Tz,zd
if
.Thus, the
1
,
1
and
0
z . Hence,
Tz
Tz
z if
TzM
1
2
.
,
0
then
0 and
,0
so
1
0
. The proof that
1
that
from the definitions of the sets
1
Case 2:
Then,
1
,K,z,z
,
1
,
3
,
Tz,zd
,K,y,x
Tz,zd
,
TzM
,
,K,Tz,TzM
2
Tz
if
zd
,
3
Tz
KTz
TzM
,
,
,
1
Tz
z if
which fails for
.
,
,
,
and only if
1
. But
2
and then the inequality fails if
Case 3:
,K,z,z
,
1
1
,
:
3
Tz
z . Hence,
Tz
z since Tz is closed.
1,210
/
,
2
1
-1
so that
1
2
,
,
4
,
,K,Tz,TzM
TzM
2
,
Tz
,
,
TzM
1
,
KTz
,
.
Then,
1
1
Tz,zd
,
,K,z,z
Tz,zd
zd
,
Tz
if
,
4
which fails for
z if and only if
Tz
, equivalently if and only if
1
1
2
. But
1
1
1,210
,
/
2
,
-1
2
1
1
so that the inequality
1
,
:
Tz
z . Hence,
4
fails if
Case 4:
z since Tz is closed.
Tz
TzM
2
/
21
,
0
K
1
,K,y,x
,
,
Tz
,
,
,K,Tz,TzM
,
Tz,zd
2
TzM
1
,
Tz,zdK
KTz
,
Tz,zd
.
Then,
Tz,zd
,K,z,z
,
Tz,zd
K
,
max
zzd
,
Tz,zd
Tz
z since Tz is closed.
,
Tz
z . Hence,
which is a contradiction for any
Case 5:
,K,y,x
,
1
K
,
/
21
K
1
TzM
2
,
Tz
,
,
,
,K,Tz,TzM
TzM
1
,
KTz
,
. Then,
1
Tz,zdK
,K,z,z
,
Tz,zd
zd
,
Tz
,
K
Tz
max
zd
z . Hence,
Tz
Tz
z since Tz is closed. A combined result of Cases 1-5 is
which is a contradiction if
BA
BA
BA
x
Tz
xn
z
0D
and that
. Now, assume again that
for any
that
1
Tz,zdK
Tz,zd
Tz,zd
zzd
,
,
,
Tz,zd
2
6
z
there are two distinct fixed points
ny
and
nx
sequences
BAz
converge to
Nn
y,x
x
y
Ty
for
where
1
1
n
n
1
1
contractive condition (2.7), subject to (2.2)-(2.4), that:
and
BA
zT
z
x
y
y
BA to which the
Tz
necessary located in
x
z
BA
. Assume also that
Tz
Tq
q
, respectively, where
x
1
n
n
. One gets from the
Tx
,
Tzqd
,
K
1/2,
max
Tz,yd
,Tq,xd
Tz,qd,Tq,zdmax
sup
sup
y
Tq
Tz
x
zd
qzdmax
qzdK
,
,
Tz
z
z 1
q 1
sequences
construct
Thus,
1
1
n
n
n
n
n
n
Nn
q,zd
q,zd
z,qd
z,qd
BAqz
. Since
for
which is nonempty,
,
1
1
n
n
0
R
such that nx and nq are in
n
n
for any given
closed and convex, there exists
0
Tz
Tq
q
z
n
qn
z n
0n
BA
BA
Tz
z
q
Tq
with
and
and
BA contradicting the hypothesis that such sets are distinct. Property (i) has been
Hence,
in
with
qzd
,
and
BA for
. Then,
Tq
such
Tz
that
n
Tq
Tq
as
q
q
,
.
z
,
proven.
Property (ii) is proven by using, in addition, [Lemma 3.8, ref. 4], one gets:
D
xd
x
,
,
xd
2
xd
2
0
,
2
2
3
2
3
2
n
n
n
n
n
n
1
1
x
2
x
2
with
lim
n
lim
lim
n
n
dX ,
since
for any sequence nx
Tx
BA
x
1
1
n
n
A and B are nonempty and disjoint closed subsets of X and A is convex. Note that Lemma 3.8 of ref. 4
and its given proof remain fully valid for multivalued cyclic self-maps since only metric properties were
used in its proof. It turns out that
1
2 nx
is a Cauchy sequence, then bounded, with a limit Az in A ,
in A since
BA
BAT
which is also a best proximity point of
:
is a uniformly convex Banach space,
and
x
xd
xd
xd
xd
,
xd
,
D
,
,
2
,
2
2
2
2
2
2
2
2
2
n
n
n
n
n
n
n
z
z
z
A
A
A
x
x
B
1
D
Tz
and
lim
n
lim
n
lim
n
lim
n
, which is also a unique best proximity point in
A
is a uniformly convex Banach space and A and B are
. Also,
nx2
B
and B is convex
B
x 1
. Also, if
lim
n
nx2
and then
z
converges to some point in
B
Tz A ), since
dX ,
B
z
B Tz
B (then
A
nonempty closed and convex subsets of X . In the same way,
and
1
is bounded sequences since
nx2
2 nx
x
Tx
then the above result holds with
n
n
n
n
n
n
2
3
2
2
2
2
2
1
2
2
1
0D
for
, the reformulated five Cases 1-5 in the proof of Property (i) would lead to contradictions
z
D
A Tz
z
B Tz
z
zdD
if
or if
. From Proposition 3.2 of ref. 4, there exist
,
0
A
B
A Tz
zdD
Tz,
B Tz
z
z
zd
that
such
A
A
B
A
B
satisfying the contractive condition (2.7)-(2.8), where A and B are nonempty and closed subsets of a
nx2
and
, with convergent subsequences
1
d,X
complete metric space
in both A and B ,
2 nx
x
and in B and A , respectively, for any given
x
x
A
y
respectively, for any
1
1
BA
BAT
:
is bounded and
. Assume some
A
D
is bounded
is cyclic
B
and
. Now,
since
and
Tz,
Tx
Tx
Tz
B
B
B
A
A
,
x
x
B
A
z
7
Tz
Ty
A
with
with
y
in A of
x
Tx
n
n
1
2
2
. Assume also that there
, which
in A , being generated from some given
B
, in A generated from
given sequence
1
x 1
2 nx
A
z
BA
BAT
converges to the best proximity point
:
A
A
, distinct of
is some sequence
ny 2
nx2
x
y
1
n
1
1
2
BA
BAT
is a best proximity point in B of
A
B
z
Tz
z A 1
where
converges to
:
B
A
obtained by using the norm- induced metric in the Banach space
the metric space
d,X
Dy,xd
Ax and
for any
both spaces can be mutually identified to each other. Since
A
zdD
zd
zd
zd
in A and B . Hence,
, where Az
B Tz
Tz
z
and
1
A
A
A
B
any sequence converges to best proximity points . Hence Property (ii) has been proven. □
are best proximity points of
which
n
2
. Consider
By , it
so that
and Bz
and then
BAT
BA
:
ACKNOWLEDGMENTS
The author is grateful to the Spanish Ministry of Education for its partial support of this work through Grant
DPI2009-07197. He is also grateful to the Basque Government for its support through Grants IT378-10 and
SAIOTEK S-PE09UN12.
follows that
z
A
1
A
Tz
B
Tz,
B
z,
A
B
Tz,
Tz,
A
A
z
z,
B
,X
A
1
if
REFERENCES
[1] M. Kikkawa and T. Suzuki, “Three fixed point theorems for generalized contractions with constants in complete
metric spaces” , Nonlinear Analysis, Theory, Methods & Applications, Vol. 69, No. 9, pp. 2942-2949, 2008.
[2] Y. Enjouji, M. Nakanishi and T. Suzuki, “A generalization of Kannan´s fixed point theorem”, Fixed Point Theory
and Applications, Vol. 2009, Article ID 192872, 10 pages, doi:10.1155/2009/192872.
[3] D. Doric and R. Lazovic, “ Some Suzuki- type fixed point theorems for generalized multivalued mappings and
applications”, Fixed Point Theory and Applications, 2001 20011:40, doi.10.1186/1687-1812-2011-40.
[4] A.A. Eldred and P. Veeramani, “Existence and convergence of best proximity points”, Fixed Point Theory and
Applications, Vol. 323, pp. 1001-1106, 2006.
[5] M. De la Sen, “Linking contractive self-mappings and cyclic Meir-Keeler contractions with Kannan self-
mappings”, Fixed Point Theory and Applications, Vol. 2010, Article
ID 572057, 23 pages,
2010.doi:10.1155/2010/572057.
[6] L.J. Ciric, “Multi-valued nonlinear contraction mappings”, Nonlinear Analysis, vol. 71, pp. 2716-2723 (2009).
Doi:10.1016/j.na2009.01.116
[7] L.J. Ciric, “Fixed points for generalized multi-valued contractions”, Matematicki Vesnik, Vol. 9, No. 24, pp. 265-
272, 1972.
[8] M. De la Sen, “Stable iteration procedures in metric spaces which generalize a Picard-type iteration”, Fixed Point
Theory and Applications, Vol. 2010, Article ID 572057, 15 pages, 2010.doi:10.1155/2010/953091.
[9] S. Karpagam and S. Agrawal, “Best proximity point theorems for p-cyclic Meir-Keller contractions”, Fixed Point
Theory and Applications, Vol. 2009, Article ID 197308, 9 pages, 2009.doi:10.1155/2009/197308.
8
|
1212.2076 | 1 | 1212 | 2012-12-10T14:27:27 | On necessary and sufficient conditions for the variable exponent Hardy type inequality | [
"math.FA"
] | We derive a number of equivalent criterions for the variable exponent Hardy type inequality
|\frac{1}{x}\int_{0}^{x}f(t)dt|_{L^{p(.)}(0,1)}\leq C|f|_{L^{p(.)}(0,1)}; f\geq 0.
to hold, whenever the exponent $p:(0,1)\to (1,\infty)$ is increasing or decreasing near small neighborhood of the origin. | math.FA | math |
On necessary and sufficient conditions
for the variable exponent Hardy type
inequality
Institute Mathematics and Mechanics of Nat.Acad.Sci., Azerbaijan
Farman I. Mamedov
e-mail: [email protected]
December 20, 2012
We derive a number of equivalent criterions for the variable exponent
Hardy type inequality
Abstract
1
x Z x
0
(cid:13)(cid:13)(cid:13)(cid:13)
f (t)dt(cid:13)(cid:13)(cid:13)(cid:13)Lp(.)(0,1)
≤ C kf kLp(.)(0,1) ; f ≥ 0.
to hold, whenever the exponent p : (0, 1) → (1, ∞) is increasing or de-
creasing near small neighborhood of the origin.
Key words and phrases : Hardy operator, Hardy type inequality, variable exponent,
weighted inequality, necessary and sufficient condition.
2000 Mathematical Subject Classification: 42A05, 42B25, 26D10, 35A23
1 Introduction
We study Hardy's inequality
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤ C kf kLp(.)(0,1)
in the norms of variable exponent Lebesgue space Lp(.)(0, 1). Here Hf (x) =
0 f (t)dt is Hardy's operator and the constant C > 0 does not depend on arbi-
trary positive measurable function f. This subject has been studied by several
authors (see, e.g. [2], [4], [5], [7], [8], [9], [11], [12], [13], [14], [15], [16], [17]).
R x
(1.1)
1
There are several sufficient conditions on the function p : (0, 1) → (1, ∞) for
the inequality (1.1) to hold. They are expressed in terms of regularity conditions
for p at the origin. It follows from the results of works [4], [9], [15] ( see, also [2],
[12], [14]) that the inequality (1.1) holds if p− = inf p > 1, p+ = sup p(x) < ∞
and the condition
A := lim sup
p(x) − p(0) log
x→0
1
x
< ∞.
(1.2)
is satisfied.
One can think that the inequality (1.1) does not need for a condition type of
(1.2) at all. Since there exists an example of function p for which the inequality
(1.1) is violated by some sequence of functions {fk} (see, [9], [7]), we see that
the inequality (1.1) does not hold without restriction on p (Note, the p there is
not monotone and does not satisfy (1.2)). In [11] (see, also [7]), we had proved
that the condition
B := lim sup
x→0
hp(x) − p(cid:16) x
2(cid:17)i log
1
x
< ∞
(1.3)
is necessary for this case. Note that, condition (1.3) is strictly weaker than (1.2).
This condition is new and somewhat surprising. For example, it is satisfied by
p(x) = p(0) + C
x )α and 0 < α < 1, C > 0, whereas the condition (1.2) is not
(ln 1
satisfied. For the exponent, that is nondecreasing near the origin, the condition
(1.3) is also sufficient if the number B satisfies B < p(0) (p(0) − 1) (see, [11]).
Unfortunately, the good condition (1.3) is no longer sufficient for the inequality
(1.1) to hold if the condition on B be ignored. In this case, a necessary and
sufficient condition is still an open problem.
In Theorem 2.2, we prove that the condition
Z 1
a (cid:16)a
1
p′(a) x− 1
p′(x)(cid:17)p(x) dx
x
≤ C, 0 < a < 1
(1.4)
and several other equivalent conditions are necessary and sufficient for the in-
equality (1.1) to hold in the case of nondecreasing exponents.
Also, in Theorem 2.1, we prove that no condition is needed if the exponent
p is nonincreasing at small neighborhood of the origin.
We refer to the monograph [3] and references therein for a full description
of variable exponent Lebesgue spaces and boundedness of classical integral op-
erators there.
2 Main results and notation
As to the basic properties of spaces Lp(.) , we refer to [6], [18]. Throughout
this paper, it is assumed that p (x) is a measurable function in (0, 1) , taking its
values from the interval [1, ∞) with p+ = sup {p (x) : x ∈ (0, 1)} < ∞ . The
space of functions Lp(.) (0, 1) is introduced as the class of measurable functions
2
Lp(.) (0, 1) is given in the form
f (x) on (0, 1) which have a finite Ip(.) (f ) =R 1
kf k =(cid:26)λ > 0 : Ip(.)(cid:18) f
λ(cid:19) ≤ 1(cid:27) .
0 f p(x) dx modular. A norm in
For 1 < p−, p+ < ∞ the space Lp(.)(0, 1) is a reflexive Banach space.
The relation between modular and norm is expressed by the following in-
equalities (see, f.e. [18]):
kf kp+
Lp(.)(0,l) ≤ Ip (f ) ≤ kf kp−
Lp(.)(0,l) , 1 ≥ kf kp(.),
kf kp−
Lp(.)(0,l) ≤ Ip (f ) ≤ kf kp+
Lp(.)(0,l) , 1 ≤ kf kp(.) .
Such estimates alow us to perform our estimates in terms of a modular.
(2.1)
(2.2)
1
p(x) + 1
For the function 1 ≤ p(x) < ∞
p′(x) denotes the conjugate function of
p′(x) = 1 and p′ = ∞ if p = 1. We denote by C, C1, C2, ... various
p(x),
positive constants whose values may vary at each appearance. By χE we denote
the characteristic function of set E. We say the function f is almost increasing
(almost decreasing) on [0, 1] if f (x) ≤ Cf (y)
(f (y) ≤ Cf (x)) for all x ≤ y in
[0, 1] and C > 0.
Following main results are obtained in this paper.
Theorem 2.1 Let p : (0, 1) → [1, ∞) be a measurable function such that p is
nonincreasing on some interval (0, ǫ), ǫ > 0 and p+ < ∞. Then it holds the
inequality (1.1) for any positive measurable function f.
Theorem 2.2 Let p : (0, 1) → [1, ∞) be a nondecreasing function such that
p(1) < ∞. Then the following statements are equivalent:
1. There exists a constant C > 0 such that the inequality
holds for any positive measurable function f.
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤ C kf kLp(.)(0,1)
(2.3)
2. The condition
Z 1
a
is satisfied.
− 1
p′(x)
x
dx
x
≤ Ca
− 1
p′(a) , 0 < a < 1
(2.4)
3. There exists an ǫ > 0 such that the function x
− 1
p′ (x)
+ǫ is almost decreasing:
− 1
2
p′ (t2 )
t
+ǫ
≤ Ct
+ǫ
− 1
1
p′(t1 )
as
0 < t1 ≤ t2 < 1.
(2.5)
3
4. The condition (4.16) is satisfied.
5. The condition
kx−1kp(.);(a,1) ≤ Ca
− 1
p′(a) , 0 < a < 1.
(2.6)
is satisfied.
3 Proof of Theorem 2.1.
Let f (x) ≥ 0 be a measurable function such that kf kLp(.)(0,1) ≤ 1. Then it
follows from the inequality (2.1) that Ip(.) (f ) ≤ 1. In order to prove Theorem
2.1 we have to show that
To prove (3.1), we establish the estimate
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤ C1.
x (cid:19) ≤ C2.
Ip(.)(cid:18) Hf
(3.1)
Using triangle inequality for p(.)-norms and ǫ ∈ (0, 1), we have
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,ǫ) +(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(ǫ,1)
:= i1 + i2.
(3.2)
Taking into account
Hf (x)
0
x
f (tx)dt
=Z 1
f (. t)dt(cid:13)(cid:13)(cid:13)(cid:13)p(.); (0,ǫ)
f (xt)p(x)χf (xt)≥1dx +Z ǫ
≤Z 1
for x ∈ (0, ǫ). Therefore,
0
0
dx
f (tx)p(tx)χf (tx)≥1 = ǫ +
f (u)p(u)du.
1
t Z tǫ
0
Hf
i1 =(cid:13)(cid:13)(cid:13)(cid:13)
0
Z 1
≤(cid:13)(cid:13)(cid:13)(cid:13)
x (cid:13)(cid:13)(cid:13)(cid:13)p(.); (0,ǫ)
Z ǫ
f (xt)p(x)dx ≤Z ǫ
≤ ǫ +Z ǫ
Z ǫ
f (tx)p(x) ≤
0
0
0
0
Whence,
1
t
+ ǫ ≤
2
t
,
0 < t < 1.
4
and using Minkowskii's inequality for Lp(.) norms, it follows that (see, [6], [18])
Let us estimate the term kf (. t)kp(.); (0,ǫ) for 0 < t < 1. Since p is nonincreas-
ing on (0, ǫ), we have p(x) ≤ p(tx)
kf (. t)kp(.); (0,ǫ) dt.
(3.3)
This implies
Z ǫ
0 (cid:18) f (tx)
p− 2
− 1
t
p− (cid:19)p(x)
1
dx ≤ 1, 0 < t < 1.
Therefore and using the definition of p(.) -norms, we get
kf (· t)kp(.); (0,ǫ) ≤ 2
1
p− t
− 1
p− , 0 < t < 1.
Using (3.4) and (3.3) for the first summand in (3.2) we have the estimate
1
t
i1 ≤ 2
p− Z 1
Now we shall estimate the term(cid:13)(cid:13)(cid:13)
0
inequality, we get
− 1
p− dt ≤
p−
p− − 1
1
p− .
2
. For x ∈ (ǫ, 1) using Young's
Hf (.)
. (cid:13)(cid:13)(cid:13)p(.); (ǫ,1)
dt +Z 1
0
1
ǫp− +
p− ≤
p− − 1
f (tx)p(tx)
p(tx)
dt
p′(tx)
1
(p+)′ ≤ 1 +
1
ǫ
.
(3.4)
(3.5)
Therefore,
f (u)p(u)du +
0
0
0
1
f (tx)dt ≤Z 1
Z 1
xp− Z x
i2 =(cid:13)(cid:13)(cid:13)(cid:13)
≤(cid:18) 1
Hf (.)
ǫ
.
(cid:13)(cid:13)(cid:13)(cid:13)p(.); (ǫ,1)
+ 1(cid:19) k1kp(.); (ǫ,1) ≤ C.
Z 1
=(cid:13)(cid:13)(cid:13)(cid:13)
f (. t)dt(cid:13)(cid:13)(cid:13)(cid:13)p(.); (ǫ,1)
0
Inserting this estimate and (3.5) in (3.2) we complete the proof of Theorem 2.1.
4 Proof of Theorem 2.2.
To prove Theorem 2.2 we need several lemmas.
Lemma 4.1 Let p : (0, 1) → [1, ∞) be a monotone nondecreasing function such
that p(1) < ∞ and the condition (4.16) is satisfied. Then there exists a constant
C1 > 0 depending on C, p(1) such that the condition
is satisfied.
1
p′(2x)
−
(cid:12)(cid:12)(cid:12)(cid:12)
1
p′(x)(cid:12)(cid:12)(cid:12)(cid:12)
ln
1
x
≤ C1
Proof. From (4.16) it follows that
(4.1)
Z 4a
2a (cid:16)x− 1
p′(x) a
1
p′ (a)(cid:17)p(x) dx
x
≤ C.
5
Suppose a
Whence,
or
ln 2 ≥ 41−p(0) ln 2 a
1
p′(a)
− 1
p′(2a) .
1
p′(2a) a
p′(a)(cid:17)p(0)
C ≥(cid:16)(4a)− 1
a(cid:19)
(cid:18) 1
p′(a)(cid:19) ln
(cid:18) 1
p′(2a)
− 1
p′(2a)
−
1
1
p′ (a)
≤ 1 +
C4p(1)−1
ln 2
1
a
≤ ln(cid:18) C4p(1)−1
ln 2
+ 1(cid:19)
Since
1
p′(x) is monotone nondecreasing, we have
Z 4a
2a (cid:16)(4a)− 1
p′(2a) a
1
p′(a)(cid:17)p(x) dx
x
≤ C.
1
p′(a) (4a)− 1
p′ (2a) is greeter then 1. Then
This completes the proof of Lemma 4.1 with constant C1 = ln(cid:16) C4p(1)−1
ln 2 + 1(cid:17) .
Lemma 4.2 Let p : (0, 1) → [1, ∞) be a nondecreasing function satisfying the
condition (4.16) and p(1) < ∞. Then there exists a constant C1 > 0 depending
on C and p(0) such that for any x
2 ≤ y ≤ 2x, 0 < x < 1
4 the estimate
1
C1
φ(x) ≤ φ(y) ≤ C1φ(x)
(4.2)
holds, where the function φ(t) = t
− 1
p′(t) .
Proof. Since 1
p′ is nondecreasing it follows from Lemma 4.1 that
1
p′(2x)
− 1
p′(x)
x− 1
p′(x) 2
1
p′(1)
φ(y) ≤(cid:16) x
2(cid:17)− 1
p′(y) ≤(cid:18) 1
x(cid:19)
≤ 2(cid:16)C4p(1)−1 + 1(cid:17) φ(x).
By the same way,
1
p′(2y)
− 1
p′ (y)
y− 1
p′(y) 2
1
p′(1)
φ(x) ≤(cid:16) y
2(cid:17)− 1
p′(x) ≤(cid:18) 1
y(cid:19)
≤ 2(cid:16)C4p(1)−1 + 1(cid:17) φ(y).
Therefore, (4.2) is satisfied by the constant 2(cid:0)C4p(1)−1 + 1(cid:1) .
Lemma 4.3 Let p : (0, 1) → [1, ∞) be a nondecreasing function such that
p(1) < ∞ and the condition (4.14) is satisfied. Then there exists a constant
C1 > 0 depending on C such that the condition (4.1) is satisfied.
6
Proof. Using (4.14) we have
Ca− 1
p′(a) ≥Z 4a
2a
x− 1
p′(x)
dx
x
4a(cid:19)
≥(cid:18) 1
1
p′(2a)
ln 2 ≥ 4− 1
p′(1) ln 2(cid:18) 1
a(cid:19)
1
p′(2a)
;
that is,
1
p′ (2a)
− 1
p′(a)
≤
4C
ln 2
.
a(cid:19)
(cid:18) 1
This proves (4.1) with constant C1 = ln(cid:0) 4C
ln 2(cid:1) .
Lemma 4.4 Let p : (0, 1) → [1, ∞) be a nondecreasing function satisfying the
conditions (4.14) and p(1) < ∞. Then there exists a constant C1 such that
1
C1
φ(x) ≤ φ(y) ≤ C1φ(x),
for any x
2 < y < 2x, 0 < x < 1
4 , where the function φ(t) = t− 1
p′(t) .
Proof. To prove Lemma 4.4 it suffice to apply Lemma 4.3 as in Lemma 4.2.
Lemma 4.5 Let p : (0, 1) → [1, ∞) be a nondecreasing function such that
p(1) < ∞. Then the following two assertions are equivalent:
1) The condition (4.14) is satisfied.
2) There exists an ǫ > 0 such that the function xǫφ(x) is almost decreasing:
there exists a C1 > 0 such that
tǫ
2φ(t2) ≤ C1tǫ
1φ(t1), 0 < t1 ≤ t2 < 1.
(4.3)
Here the function φ(t) = t− 1
p′(t) .
Proof. Proof of 1) → 2). Denote g(x) =R 1
x φ(t) dt
t . Then
g′(x) = −
x
φ(x)
,
0 < x < 1.
Hence
g(x) ≤ −Cg′(x)x
or
1
C
1
x
≤
−g′(x)
g(x)
,
0 < x < 1.
Integrating this inequality in x over (t1, t2), we get
ln
g(t1)
g(t2)
≥
1
C
ln
t1
t2
or
7
g(t2)t
1
C
2 ≤ g(t1)t
1
C
1 .
Since
g(t2) =Z 1
t2
φ(x)
dx
x
≥Z 2t2
t2
φ(x)
dx
x
≥
1
C
φ(t2) ln 2,
using (4.14) and assertion of Lemma 4.4 we get
ln 2
C
φ(t2)t
1
C
2 ≤ Cφ(t1)t
1
C
1
Therefore, (4.3) is satisfied with ǫ = 1
C , C1 = C2.
Proof of 2) → 1). Estimating directly, we have
Z 1
a
φ(x)
dx
x
a
xǫφ(x)
=Z 1
= Caǫφ(a)Z 1
a
dx
x1+ǫ ≤ CZ 1
dx
x1+ǫ =
a
C
ǫ
φ(a).
aǫφ(a)
dx
x1+ǫ
The inequality (4.14) has been proved.
Lemma 4.6 Let p : (0, 1) → [1, ∞) be nondecreasing function such that p(1) <
∞. Then the condition (4.16) is necessary for the inequality (1.1) to hold.
Proof. Let a ∈ (0, 1) be a fixed number. Put a test function
f0(x) = x− 1
p(x) χ( a
2 ,a)(x),
0 < x < 1,
into the inequality (1.1). Then
= ln 2 ≤ 1,
a
2
Hf0
dx
x
Ip(.) (f0) =Z a
x (cid:13)(cid:13)(cid:13)p(.);(0,1)
x−p(x)dx ≥Z 1
a (cid:16) a
2
C2, whence
therefore, kf0kp(.) ≤ 1. Hence(cid:13)(cid:13)(cid:13)
p(t) dt!p(x)
t− 1
a Z a
C2 ≥Z 1
≥ 2−p+Z 1
a (cid:16)a
a
2
1
p′(a) x
− 1
p′ (x)(cid:17)p(x) dx
x
.
≤ C. This implies that Ip(.)(cid:16) Hf0
x (cid:17) ≤
a− 1
p(a)(cid:17)p(x)
x−p(x)dx
Hence
Z 1
a (cid:16)a
1
p′(a) x
− 1
p′(x)(cid:17)p(x) dx
x
≤ C3.
8
Lemma 4.7 Let p : (0, 1) → [1, ∞) be a nondecreasing function satisfying the
conditions p(1) < ∞ and (4.16). Then the function φ(x) = x− 1
p′(x) is almost
decreasing; that is for any 0 < t1 ≤ t2 < 1 we have
φ(t2) ≤ Cφ(t1)
Proof. Put t1 = a. Let 2k−1a ≤ t2 < 2ka, k ∈ N. Then using (4.16) and
Lemma 4.2, we have
C ≥
≥
+
1
∞
∞
p′(a) x
Xn=1Z 2na
2n−1a(cid:16)a
Xn=1,n∈N'Z 2na
2n−1a(cid:16)a
Xn=1,n∈N"Z 2na
2n−1a(cid:16)a
∞
− 1
p′(x)(cid:17)p(x) dx
x
1
p′ (a) (2na)− 1
1
p′(a) (2na)− 1
x
p′(2n a)(cid:17)p− dx
p′(2n a)(cid:17)p+ dx
x
,
n=1,n∈N'(...) means summing over n ∈ N such that a
n=1,n∈N"(...) means summing over n ∈ N such that a
1
p′(a) (2na)− 1
p′ (a) (2na)− 1
1
p′(2n a) ≤
p′(2n a) ≥
whereP∞
1 andP∞
1. Therefore,
1
p′(a) (2na)− 1
p′ (2n a) ≤ 1 +
a
C
C1 ln 2
,
n ∈ N.
(4.4)
Further using the Lemma 4.2, we deduce from (4.4)
hence by using Lemma 4.2, we have
a
1
p′(a) (cid:0)2ka(cid:1)− 1
p′ (2k a) ≤ C3,
1
p′ (a) t
a
− 1
2
p′(t2 )
≤ C4.
This completes the proof of Lemma 4.7.
Lemma 4.8 Let p : (0, 1) → [1, ∞) be a nondecreasing function satisfying the
conditions p(1) < ∞ and (4.16). Then the condition (4.14) is satisfied, more-
over, the function x− 1
+ǫ is almost decreasing by some ǫ > 0.
p′(x)
Proof. Using (4.16) and Lemma 4.7 we have the estimates
p′(a) x− 1
1
≥ C p−
C ≥Z 1
a (cid:16)a
4 Z 1
a (cid:18) 1
Z 1
a (cid:16)a
C p−−p+
4
C4
p′(x)(cid:17)p(x) dx
x
1
a
p′(a) x
x
− 1
p′(x)(cid:19)p(x) dx
p′(x)(cid:17)p+ dx
x
.
1
p′(a) x
− 1
9
This implies
Z 1
a
x− p+
p′(x)
dx
x
≤ C p(1)−1
4
a− p+
p′(a) ,
0 < a < 1.
(4.5)
Applying the approach of Lemmas 4.3 and 4.7, we find the function x− p+
p′(x) is
almost decreasing and satisfies the condition (4.5).
It follows from the Bari-
Stechkin theorem [1] (see, also [10]) that there exists an ǫ > 0 such that the
function x− p+
+ǫ1 is
almost decreasing. Again using Bari-Stechkin result [1] we deduce the function
x− 1
+ǫ is almost decreasing. This implies the function x− 1
p′(x) satisfies the condition (4.14)
Hence we have proved that (by using Lemmas 4.6 and 4.8 for the inequality
(1.1) to hold it is necessary the condition (4.14). Let us prove that the condition
(4.14) is also sufficient for (1.1).
p′(x)
p′(x)
Remark 4.1 It follows from Lemma 4.8 that the condition (4.14) for nonde-
creasing p : (0, 1) → [1, ∞) implies p(0) > 1. Hence the condition p(0) > 1 is
necessary (but not sufficient) for the inequality (1.1) to hold.
Lemma 4.9 Let p : (0, 1) → [1, ∞) be a nondecreasing function such that the
conditions (4.14) and p(1) < ∞ is satisfied. Then the inequality (1.1) holds.
Using Lemma 4.3 we infer that the function x
is almost de-
Proof.
creasing. Further, according to Lemma 4.5 the condition (4.14) implies that the
function x− 1
+ǫ is almost decreasing by some ǫ > 0.
p′(x)
p′(x)
− 1
Let us prove sufficiency of condition (4.14). It suffices to consider the case
when function f (x) ≥ 0 is a measurable function such that kf kLp(.)(0,1) ≤ 1
(see, [3]). Then Ip(.) (f ) ≤ 1. In order to prove Lemma 4.9 we have to prove
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤ C1. We shall derive this inequality from the estimate
Ip(.)(cid:0)x−1Hf(cid:1) ≤ C2.
By Minkowski inequality, for Lp(.) norms, we get the inequalities
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xn=0(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
x− 1
≤
∞
f (t)dt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp(.)(0,1)
∞
2−n−1x
Xn=0Z 2−nx
f (t)dt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp(.)(0,1)
(4.6)
x− 1
p(x) − 1
p(x)
p(x) − 1
p(x) Z 2−nx
2−n−1x
Denote Bx,n = (2−n−1x, 2−nx] and px,n = inf{p(t) : t ∈ Bx,n}; n = 1, 2, .... Put
p(t) . Since the condition (4.14) holds, it follows from Lemma 4.8 that
ϕ(t) = t
there exists an ǫ ∈ (0, 1) such that
1
ϕ(s)
sǫ ≤ C
ϕ(r)
rǫ
, 0 < s < r < 1.
(4.7)
10
Then by (4.7) we have
ϕ(t)
tǫ ≤ C
ϕ(x)
xǫ
,
(4.8)
where t is a point in Bx,n, 0 < x < 1 and the constant C does not depend on
n.
By using inequality (4.8) and 2−n−1x < t < 2−nx we have the estimates
1
p′ (t) = tǫt
1
p′ (t)
−ǫ ≤ Ctǫx
1
p′(x)
−ǫ ≤ C2−nǫx
1
p′(x) .
t
Hence
x− 1
p′(x) ≤ C2−nǫt− 1
p′ (t)
.
Therefore, and due to Holder's inequality, for x ∈ B(0, 1), we get
x− 1
p(x) − 1
p′(x)
f (t)dt
≤ C2−nǫx− 1
p(x) t− 1
f (t)dt
∞
2−n−1x
Xn=0Z 2−nx
p′(t) Z 2−nx
x,n dt!
f (t)p−
2−n−1x
≤ C2−nǫx− 1
p(x) t
− 1
p′(t) Z 2−nx
2−n−1x
1
−
x,n
p
1
−
x,n)′
(p
(cid:0)2−nx(cid:1)
It follows from Lemma 4.2 that
where C depends only p.
(cid:0)2−nx(cid:1)
1
−
x,n)′
(p
≤ 2
− 1
(p
−
x,n)′
1
−
x,n)′
(p
t
≤ C1t
1
p′(t) ,
(4.9)
(4.10)
Combining (4.9) and (4.10) we get
x− 1
p(x) − 1
p′(x)
∞
Xn=0Z 2−nx
2−n−1x
f (t)dt ≤ C2−nǫx− 1
p(x) Z 2−nx
2−n−1x
1
−
x,n
p
f (t)p−
x,n dt!
(4.11)
where 0 < x < 1, n = 1, 2, ...and the constant C2 does not depend on n, x.
Simultaneously,
Z 2−nx
2−n−1x
f (t)p−
x,n dt ≤Z 2−nx
2−n−1x
f (t)p(t)χ{f (t)≥1}dt +Z 2−nx
2−n−1x
dt ≤ 1 + 2−n ≤ C3.
By the last inequality and (4.11), we have
Ip(.) x
− 1
p(x) − 1
p′ (x) Z 2−nx
2−n−1x
f (t)dt! ≤ C42−nǫp−Z 1
0
x−1 Z 2−nx
2−n−1x
f (t)p−
x,n dt!
p(x)
p
−
x,n
dx
p+
p− −1
≤ C4C
3
2−nǫp−Z 1
0
x−1 Z 2−nx
2−n−1x(cid:16)f (t)p(t) + 1(cid:17) dt! dx
11
which, due to Fubini's theorem, yields
p+
p− −1
≤ C4C
3
2−nǫp−
ln 2Z 2−n
0
Z 2−nx
2−n−1x
x−1dx!(cid:16)f (t)p(t) + 1(cid:17) dt
= C52−nǫp−
ln 2
Therefore,
x− 1
p(x) − 1
p′(x) Z 2−nx
2−n−1x
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
By (4.12) and (4.6), we get
2−n
Z0 (cid:16)f (t)p(t) + 1(cid:17) dt ≤ C62−nǫp−
f (t)dt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp(.)(0,1)
Xn=0
≤ C2− nǫp−
2− nǫp−
p+ ≤ C1.
∞
p+
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤ C
This completes the proof of Lemma 4.9.
Proof of Theorem 2.2. Let 5) be satisfied, that is the condition (4.17).
Then by the definition,
.
(4.12)
Therefore, and using (4.17), we have
or
dx ≤ 1.
p(x)
dx ≤ 1
Z 1
a
x−1
(cid:13)(cid:13)(cid:13)
(.)−1χ{a,1}(.)(cid:13)(cid:13)(cid:13)p(.)
p′(a)(cid:19)p(x)
Z 1
a (cid:18) x−1
Z 1
p′(x)(cid:17)p(x) dx
a (cid:16)a
Ca− 1
p′(a) x
− 1
x
1
≤ C1.
This is the condition (4.16), that is 4) of Theorem 2.2. Hence 5) → 4) has been
proved. According to Lemmas 4.6, 4.7, 4.8, we have the implication 4) → 2).
The implication 2) → 3) follows from Lemma 4.5. The implication 3) → 1)
follows from Lemma 4.9. The implication 1) → 4) is proved in Lemma 4.6.
The implication 3) → 5) is direct: using the condition (4.15) we have
Z 1
a (cid:16)a
=Z
1
a
1
1
p′ (a) x− 1
C p(1)(cid:18) 1
x
p′(x)(cid:17)p(x) dx
t(cid:19)ǫp(at) dt
t
≤Z 1
a (cid:16)C(cid:16) a
≤ C p(1)Z ∞
1
x(cid:17)ǫ(cid:17)p(x) dx
x
dt
< C2.
t1+ǫp(0)
12
Rewriting the last inequality, we have
p(x)
dx ≤ 1,
a
Z 1
x−1
p′ (a)
1
p(1)
C
2
a
− 1
therefore, the condition (4.17) is satisfied.
This completes the proof of Theorem 2.2.
If the exponent function p in Theorem 2.2 is nondecreasing on not all the
interval (0, 1) but so is only near the origin the following assertion holds.
Remark 4.2 Let a measurable function p : [0, 1] → (1, ∞) be nondecreasing on
some interval (0, δ), 0 < δ < 1 and p+ < ∞; then the following statements are
equivalent:
a) There exists a constant C > 0 such that the inequality
holds for any positive measurable function f.
(cid:13)(cid:13)x−1Hf(cid:13)(cid:13)Lp(.)(0,1) ≤ C kf kLp(.)(0,1)
(4.13)
b) The condition
is satisfied.
Z δ
a
x− 1
p′(x)
dx
x
≤ Ca− 1
p′(a) , 0 < a < δ
(4.14)
c) There exists an ǫ > 0 such that the function x− 1
p′ (x)
+ǫ is almost decreasing:
− 1
2
p′(t2 )
t
+ǫ
≤ Ct
+ǫ
− 1
1
p′(t1 )
as
0 < t1 ≤ t2 < δ
(4.15)
d) The condition
is satisfied.
e) The condition
is satisfied.
Z δ
a (cid:16)a
1
p′(a) x− 1
p′(x)(cid:17)p(x) dx
x
≤ C, 0 < a < δ
(4.16)
kx−1kp(.);(a,δ) ≤ Ca− 1
p′(a) , 0 < a < δ.
(4.17)
13
References
[1] N.K. Bari and S.B. Stechkin. "Best approximations and differential proper-
ties of two conjugate functions" (in Russian). Proceedings of Moscow Math-
ematical Society, 5:483-522, 1956.
[2] D. Cruz.-Uribe, SFO and F. I. Mamedov, "On a general weighted Hardy
type inequality in the variable exponent Lebesgue spaces,"Revista Matem-
atica Complutense, vol. 25, no. 2, pp. 335-367, 2012.
[3] L. Diening, P. Harjulehto, P. Hasto and M. Ruzicka,: "Lebesgue and
Sobolev Spaces with Variable Exponents," Lecture Notes in Mathematics,
vol 2017, Springer, Heidelberg, Germany, 2011.
[4] L. Diening and S. Samko,"Hardy inequality in variable exponent Lebesgue
spaces," Fractional Calculus & Applied Analysis, vol. 10, no 1, pp. 1-17,
2007
[5] D. E. Edmunds, V. Kokilashvili and A. Meskhi, "On the boundedness and
compactness of the weighted Hardy operators in spaces," Georgian Math-
ematical Journal, vol. 12, no. 1, pp. 27-44, 2005.
[6] X. L. Fan and D. Zhao, "On the spaces Lp(x)(Ω) and W m,p(x)(Ω)," Journal
of Mathematical Analysis and Applications, vol. 263, no. 2, pp. 424-446,
2001.
[7] A. Harman, "On necessary condition for the variable exponent Hardy in-
equality," Journal of Function Spaces and Applications, vol. 2012, Article
ID 385925, 6 pages, doi:10.1155/2012/385925
[8] P. Harjulehto, P.Hasto and M. Koskinoja, "Hardy's inequality in variable
exponent Sobolev spaces," Georgian Mathematical Journal, vol. 12, no. 3,
pp. 431-442, 2005.
[9] A. Harman and F.I. Mamedov, "On boundedness of weighted Hardy oper-
ator in Lp(.) and regularity condition," Journal of Inequalities and Appli-
cations, vol. 2010, Article ID 837951, 14 pages, 2010.
[10] V. Kokilashvili, S. Samko and N. Samko, "The Maximal Operator in
Weighted Variable Spaces Lp(.)" J. Function Spaces Appl., vol. 5, no 3,
pp. 299-317, 2007.
[11] F. I. Mamedov, "On Hardy type inequality in variable exponent Lebesgue
space Lp(.)(0, 1)," Azerbaijan Journal of Mathematics, vol. 2, no. 1, pp.
90 -- 99, 2012.
[12] F.I. Mamedov and A. Harman, "On a weighted inequality of Hardy type
in spaces Lp(.)," Journal of Mathematical Analysis and Applications, vol.
353, no. 2, pp. 521-530, 2009.
14
[13] F.I. Mamedov and A. Harman, "On a Hardy type general weighted inequal-
ity in spaces Lp(.)," Integral Equations and Operator Theory, vol. 66, no.
4, pp. 565-592, 2010.
[14] F.I. Mamedov and Y. Zeren, "On equivalent conditions for the general
weighted Hardy type inequality in space Lp(.)," Zeitschrift fur Analysis
und ihre Anwendungen, vol. 31, no 1, pp. 55-74, 2012
[15] R. Mashiyev, B. Cekic, F. I. Mamedov and S. Ogrash, "Hardy's inequality
in power-type weighted Lp(.) spaces," Journal of Mathematical Analysis
and Applications, vol. 334, no. 1, pp. 289-298, 2007.
[16] H. Rafeiro and S. G. Samko, "Hardy inequality in variable Lebesgue
spaces," Annales Academiae Scientiarium Fennicae, vol. 34, no. 1, pp. 279-
289, 2009.
[17] S. G. Samko, "Hardy inequality in the generalized Lebesgue spaces," Frac-
tional Calculus & Applied Analysis, vol. 6, no. 4, pp. 355-362, 2003.
[18] S. G. Samko, "Convolution type operators in Lp(.)," Integral Transforms
and Special Functions, vol. 7, pp. 123-144, 1998.
15
|
1611.02979 | 1 | 1611 | 2016-11-09T15:27:54 | On the Iterations of a Sequence of Strongly Quasi-nonexpansive Mappings with Applications | [
"math.FA"
] | In this paper, we study $\Delta$- convergence of iterations for a sequence of strongly quasi-nonexpansive mappings as well as the strong convergence of the Halpern type regularization of them in Hadamard spaces. Then, we give some their applications in iterative methods, convex and pseudo-convex minimization(proximal point algorithm), fixed point theory and equilibrium problems. The results extend several new results in the literature and some of them seem new even in Hilbert spaces. | math.FA | math |
ON THE ITERATIONS OF A SEQUENCE OF STRONGLY
QUASI-NONEXPANSIVE MAPPINGS WITH APPLICATIONS
Hadi Khatibzadeh1 and Vahid Mohebbi2
1,2 Department of Mathematics, University of Zanjan, P. O. Box 45195-313, Zanjan, Iran.
Abstract. In this paper, we study ∆- convergence of iterations for a sequence
of strongly quasi-nonexpansive mappings as well as the strong convergence of
the Halpern type regularization of them in Hadamard spaces. Then, we give
some their applications in iterative methods, convex and pseudo-convex minimiza-
tion(proximal point algorithm), fixed point theory and equilibrium problems. The
results extend several new results in the literature (for example [5, 7, 13, 15, 16,
19, 20, 23, 28, 31, 35, 36, 38]) and some of them seem new even in Hilbert spaces.
1. Introduction and Preliminaries
Let (X, d) be a metric space. A geodesic from x to y is a map γ from the closed
interval [0, d(x, y)] ⊂ R to X such that γ(0) = x, γ(d(x, y)) = y and d(γ(t), γ(t′)) =
t − t′ for all t, t′ ∈ [0, d(x, y)]. The space (X, d) is said to be a geodesic space if
every two points of X are joined by a geodesic. The metric segment [x, y] contains
the images of all geodesics, which connect x to y. X is called unique geodesic iff
1E-mail: [email protected], [email protected].
1
2
[x, y] contains only one geodesic.
Let X be a unique geodesic metric space. For each x, y ∈ X and for each t ∈ [0, 1],
there exists a unique point z ∈ [x, y] such that d(x, z) = td(x, y) and d(y, z) =
(1 − t)d(x, y). We will use the notation (1 − t)x ⊕ ty for the unique point z satisfying
in the above statement.
In a unique geodesic metric space X, a set A ⊂ X is called convex iff for each
x, y ∈ A, [x, y] ⊂ A. A unique geodesic metric space X is called CAT(0) space if for
all x, y, z ∈ X and for each t ∈ [0, 1], we have the following inequality
d2((1 − t)x ⊕ ty, z) ≤ (1 − t)d2(x, z) + td2(y, z) − t(1 − t)d2(x, y).
A complete CAT(0) space is called a Hadamard space.
Berg and Nikolaev in [9, 10] have introduced the concept of quasi-linearization
along these lines (see also [1]). Let us formally denote a pair (a, b) ∈ X × X
→
ab and call it a vector. Then quasi-linearization is defined as a map h·, ·i :
by
(X × X) × (X × X) → R defined by
→
h
ab,
→
cdi =
1
2
{d2(a, d) + d2(b, c) − d2(a, c) − d2(b, d)}
(a, b, c, d ∈ X).
It is easily seen that h
→
ab,
→
abi = d2(a, b), h
→
ab,
→
cdi = h
→
cd,
→
abi, h
→
ab,
→
cdi = −h
→
ba,
→
cdi and
→
h
ax,
→
cdi + h
→
xb,
→
cdi = h
→
ab,
→
cdi for all a, b, c, d, x ∈ X. We say that X satisfies the
Cauchy-Schwartz inequality if h
→
ab,
→
cdi ≤ d(a, b)d(c, d) for all a, b, c, d ∈ X.
It is
known (Corollary 3 of [10]) that a geodesically connected metric space is a CAT(0)
space if and only if it satisfies the Cauchy-Schwartz inequality.
A kind of convergence introduced by Lim [26] in order to extend weak convergence
in CAT(0) setting. Let (X, d) be a Hadamard space, {xk} be a bounded sequence
in X and x ∈ X. Let r(x, {xk}) = lim sup d(x, xk). The asymptotic radius of {xk}
is given by r({xk}) = inf{r(x, {xk})x ∈ X} and the asymptotic center of {xk} is
the set A({xk}) = {x ∈ Xr(x, {xk}) = r({xk})}. It is known that in a Hadamard
space, A({xk}) consists exactly one point.
3
Definition 1.1. A sequence {xk} in a Hadamard space (X, d) △-converges to x ∈ X
if A({xkn}) = {x}, for each subsequence {xkn} of {xk}.
It is well-known that every bounded sequence in a Hadamard space has a ∆-
convergent subsequence (see [24]). We denote △-convergence in X by
△
−→ and
the metric convergence by →.
Let C ⊆ X be closed and convex. Suppose that T : C → C is a mapping
and F (T ) := {x ∈ C : T x = x}. T is said to be nonexpansive (resp. quasi-
nonexpansive) iff d(T x, T y) ≤ d(x, y), ∀x, y ∈ C (resp. F (T ) 6= ∅ and d(T x, q) ≤
d(x, q), ∀(x, q) ∈ C × F (T )). We recall the definitions of firmly nonexpansive and
quasi firmly nonexpansive mappings.
Definition 1.2. A mapping T : C → C is called firmly nonexpansive iff
→
h
xy,
−−−→
T xT yi ≥ d2(T x, T y), ∀x, y ∈ C
T is called quasi firmly nonexpansive if F (T ) 6= ∅ and
→
h
xp,
−−→
T xpi ≥ d2(T x, p), ∀(x, p) ∈ C × F (T ).
Our definitions of firmly nonexpansive and quasi firmly nonexpansive are exten-
sions of the definitions in Hilbert spaces but they seem different from the correspond-
ing definitions in the literature (see for example [3, 30]). We don't know the relation
4
between two definitions of firmly nonexpansive mappings but for quasi-firmly non-
expansive mappings which are more important in this paper, it is easy to check that
the usual definition in the literature implies our definition. Therefore our definition
is more general than the old definition.
Recently some authors considered the asymptotic behavior of iterations of a
(firmly) nonexpansive mapping in geodesic metric spaces specially in Hadamard
spaces (see [3, 30]). In the next section, we study the asymptotic behavior of itera-
tions of a sequence of quasi firmly nonexpansive mappings as well as the dynamical
behavior of their combination with Halperm iteration. We prove our results for more
general class of mappings that are strongly quasi nonexpansive sequence.
Following [21], we recall that T : C → C is strongly nonexpansive (resp. strongly
quasi nonexpansive) iff T is nonexpansive and d(xk, T xk)−d(yk, T yk) → 0, whenever
{xk} and {yk} are sequences in C such that d(xk, yk) is bounded and d(xk, yk) −
d(T xk, T yk) → 0 (resp. T is quasi-nonexpansive and d(xk, T xk) → 0, whenever {xk}
is a bounded sequence in C such that d(xk, q) − d(T xk, q) → 0, for some q ∈ F (T )).
We also recall strongly nonexpansive and strongly quasi-nonexpansive sequences that
play an essential role in this paper. The sequence {Tk} of nonexpansive mappings is
said to be strongly nonexpansive sequence iff d(xk, Tkxk)−d(yk, Tkyk) → 0, whenever
{xk} and {yk} are sequences in C such that d(xk, yk) is bounded and d(xk, yk) −
d(Tkxk, Tkyk) → 0 . The sequence {Tk} of quasi-nonexpansive mappings is said to
be strongly quasi-nonexpansive sequence iff Tk F (Tk) 6= ∅ and d(xk, Tkxk) → 0,
whenever {xk} is a bounded sequence in C such that d(xk, q) − d(Tkxk, q) → 0, for
some q ∈ Tk F (Tk). It is clear that a strongly nonexpansive sequence {Tk} with
Tk F (Tk) 6= ∅ is a strongly quasi-nonexpansive sequence.
5
Let T : C → C be a firmly nonexpansive mapping with F (T ) 6= ∅, where C is a
closed convex subset of a real Hilbert space H. A well-known result implies that the
orbit of an arbitrary point of H under T is convergent weakly to a fixed point of T .
This result recently has been improved to Hadamard spaces by Ariza-Ruiz et al. in
[3] and Nicolae in [30]. They showed that the sequence xk = T kx is ∆-convergent to
a fixed point of T . To achieve strong convergence we need some regularized methods
like Halpern regularization which was first used by Xu [38] in Hilbert spaces and
in Hadamard spaces by [33].
In this paper we consider the asymptotic behavior
of iterations of a sequence of quasi firmly nonexpansive mappings or more gener-
ally strongly quasi nonexpansive mappings as well as their Halpern regularization
and prove ∆- convergence and strong convergence of their iterations to a common
fixed point of the sequence. In Section 3 of the paper we consider the applications
of our results in iterative methods, convex and pseudo-convex minimization, fixed
point theory of quasi-nonexpansive mappings and equilibrium problems of pseudo-
monotone bifunctions.
2. Convergence of a Strongly Quasi-nonexpansive Sequence
To prove the convergence of the iterations T kx, where T is a firmly nonexpan-
sive mapping, demiclosedness of T is essential. T : C → C is called demiclosed iff
d(xk, T xk) → 0 and xk
△
−→ x imply x ∈ F (T ). Demiclosedness is satisfied for non-
expansive mappings but for quasi firmly nonexpansive mappings and strongly quasi
nonexpansive mappings we don't have this essential property even in Hilbert spaces,
therefore we must assume it. Since we intend to prove convergence for a sequence
6
of strongly quasi nonexpansive mappings we need the definition of demiclosedness
for a sequence of mappings.
A sequence Tk : C → C of strongly quasi-nonexpansive mappings with Tk F (Tk) 6=
∅ is called demiclosed iff
if {xkj } ⊂ {xk} and {Tkj } ⊂ {Tk} such that
xkj
△
−→ p ∈ C and lim d(xkj , Tkj xkj ) = 0, then p ∈ Tk F (Tk)
In this section, we obtain ∆- convergence of the sequence {xk} given by
xk+1 = Tkxk
(2.1)
(2.2)
to an element of Tk F (Tk) 6= ∅ as well as the strong convergence of the Halpern
type algorithm:
xk+1 = αku ⊕ (1 − αk)Tkxk,
(2.3)
to the element x∗ = ProjTk F (Tk)u, where u, x1 ∈ C and the sequence {αk} ⊂ (0, 1)
satisfies lim αk = 0 and P+∞
k=1 αk = +∞. The recent result extends the results of
[21] from Hilbert spaces to Hadamard spaces.
Theorem 2.1. Suppose that Tk : C → C is a sequence of strongly quasi-nonexpansive
mappings and x0 ∈ C. We define xk+1 = Tk · · · T1x0 such that {Tk} satisfies (2.1).
Then the sequence {xk}, ∆-converges to an element of Tk F (Tk).
Proof. Take x∗ ∈ Tk F (Tk), then we have d(xk+1, x∗) = d(Tkxk, x∗) ≤ d(xk, x∗).
Therefore lim d(xk, x∗) exists for all x∗ ∈ Tk F (Tk), also {xk} is bounded. Hence,
△
−→ p ∈ C. On the other hand,
there are {xkn} of {xk} and p ∈ C such that xkn
since {Tk} is a sequence of strongly quasi nonexpansive mappings and lim d(xk, x∗)
exists for all x∗ ∈ Tk F (Tk), hence lim d(xkn , Tkn xkn) = 0. Now, (2.1) shows that
p ∈ Tk F (Tk).
In the sequel, Opial lemma (see Lemma 2.1 in [22]) follows the
result.
(cid:3)
7
Remark 2.1. Suppose that {S0, S1, . . . , Sr−1} is a finite family of quasi nonexpan-
sive mappings which are demiclosed and define the sequence {Tk} by Tk = S[k] where
[k] = k (mod r). Therefore Theorem 2.1 extends Theorem 4.1 of [4] in Hadamard
space setting.
Lemma 2.2. [34] Let {sk} be a sequence of nonnegative real numbers, {ak} be a
sequence of real numbers in (0, 1) with P∞
k=1 ak = +∞ and {tk} be a sequence of
real numbers. Suppose that
sk+1 ≤ (1 − ak)sk + aktk,
∀k ∈ N
If lim sup tkn ≤ 0 for every subsequence {skn} of {sk} satisfying lim inf(skn+1−skn) ≥
0, then lim sk = 0.
Theorem 2.3. Suppose that Tk : C → C is a sequence of strongly quasi-nonexpansive
mappings such that (2.1) is satisfied, then the sequence {xk} generated by (2.3) con-
verges strongly to ProjTk F (Tk)u.
Proof. Since Tk F (Tk) is closed and convex, therefore we assume that x∗ = ProjTk F (Tk)u.
By (2.3), we have:
d(x∗, xk+1) ≤ αkd(x∗, u) + (1 − αk)d(x∗, Tkxk) ≤ αkd(x∗, u) + (1 − αk)d(x∗, xk) ≤
max{d(x∗, u), d(x∗, xk)} ≤ · · · ≤ max{d(x∗, u), d(x∗, x1)}.
Therefore {xk} is bounded. Now, by (2.3), we have:
d2(xk+1, x∗) ≤ (1 − αk)d2(Tkxk, x∗) + αkd2(u, x∗) − αk(1 − αk)d2(u, Tkxk).
8
Since Tk is quasi-nonexpansive, we have: d2(x∗, Tkxk) ≤ d2(x∗, xk), therefore we
have:
d2(xk+1, x∗) ≤ (1 − αk)d2(xk, x∗) + αkd2(u, x∗) − αk(1 − αk)d2(u, Tkxk).
(2.4)
In the sequel, we show d(xk+1, x∗) → 0. By Lemma 2.2, it suffices to show that
lim sup(d2(u, x∗) − (1 − αkn)d2(u, Tkn xkn)) ≤ 0 for every subsequence {d2(xkn , x∗)}
of {d2(xk, x∗)} satisfying lim inf(d2(xkn+1, x∗) − d2(xkn , x∗)) ≥ 0.
For this, suppose that {d2(xkn , x∗)} is a subsequence of {d2(xk, x∗)} such that
lim inf(d2(xkn+1, x∗) − d2(xkn , x∗)) ≥ 0. Then
0 ≤ lim inf(d2(x∗, xkn+1)−d2(x∗, xkn)) ≤ lim inf(αkn d2(x∗, u)+(1−αkn )d2(x∗, Tkn xkn)
−d2(x∗, xkn)) = lim inf(αkn(d2(x∗, u)−d2(x∗, Tkn xkn))+d2(x∗, Tkn xkn)−d2(x∗, xkn))
≤ lim sup αkn(d2(x∗, u) − d2(x∗, Tkn xkn)) + lim inf(d2(x∗, Tkn xkn) − d2(x∗, xkn))
= lim inf(d2(x∗, Tkn xkn) − d2(x∗, xkn)) ≤ lim sup(d2(x∗, Tkn xkn) − d2(x∗, xkn)) ≤ 0.
Therefore, we conclude that lim(d2(x∗, Tkn xkn) − d2(x∗, xkn)) = 0, hence by the def-
inition, we get lim d2(xkn , Tknxkn) = 0.
On the other hand, there are a subsequence {xkni
} of {xkn} and p ∈ C such that
△
−→ p and
xkni
lim sup(d2(u, x∗)−(1−αkn)d2(u, Tkn xkn)) = lim(d2(u, x∗)−(1−αkni
)d2(u, Tkni
xkni
))
Since xkni
△
−→ p and lim d(xkni
, Tkni
xkni
) = 0, by (2.1) we have p ∈ Tk F (Tk). On
the other hand, x∗ = ProjTk F (Tk)u, hence we have:
lim sup(d2(u, x∗) − (1 − αkn)d2(u, Tkn xkn)) ≤ d2(u, x∗) − d2(u, p) ≤ 0.
Therefore Lemma 2.2 shows that d(xk+1, x∗) → 0, i.e. xk → x∗ = ProjTk F (Tk)u. (cid:3)
3. Applications
9
In this section, we present some examples of strongly quasi nonexpansive se-
quences and give some applications of the main results in the previous section in
iterative methods, optimizatin, fixed point theory and equilibrium problems.
3.1. Application to Iterative Methods. Consider the following iteration which
is called Ishikawa iteration.
xk+1 = (1 − αk)xk ⊕ αkT ((1 − βk)xk ⊕ βkT xk),
(3.1)
where T is a quasi-nonexpansive mapping and αk, βk ∈ (0, 1) are two sequences with
suitable assumptions. Define
Tk := (1 − αk)I ⊕ αkT ((1 − βk)I ⊕ βkT ),
(3.2)
where I is the identity mapping. We will prove that {Tk} is a strongly quasi-
nonexpansive sequence and it satisfies (2.1). Then we apply the main results to
conclude ∆- convergence of Ishikawa iteration and the strong convergence of the
Halpern-Ishikawa iteration.
Lemma 3.1. Let T : C → C be a quasi-nonexpansive mapping. If αk ∈ (0, 1) be a
sequence such that lim sup αk < 1, then the sequence {Tk} defined by (3.2) is strongly
quasi-nonexpansive.
Proof. Take {xk} in C and p ∈ F (T ). Now, by definition of Tk, we have
d2(Tkxk, p) ≤ (1−αk)d2(xk, p)+αkd2(T (1−βk)xk⊕βkT xk), p)−αk(1−αk)d2(xk, T ((1−
βk)xk ⊕ βkT xk)) ≤ (1 − αk)d2(xk, p) + αkd2(xk, p) − 1−αk
αk
d2(xk, Tkxk).
10
Therefore
1 − αk
αk
d2(xk, Tkxk) ≤ d2(xk, p) − d2(Tkxk, p),
which shows Tk is strongly quasi-nonexpansive.
(cid:3)
Now, we show the sequence {Tk} satisfies (2.1).
Lemma 3.2. Let T : C → C be a demiclosed and quasi-nonexpansive mapping.
If αk, βk ∈ (0, 1) are two sequences such that 0 < lim inf αk ≤ lim sup αk < 1 and
βk → 0, then the sequence {Tk} defined by (3.2) satisfies (2.1).
Proof. Let {xk} be an arbitrary sequence such that xk
△
−→ p and d(xk, Tkxk) → 0.
We have to prove p ∈ Tk F (Tk). The definition of Tk together with d(xk, Tkxk) → 0
imply that αkd(xk, T ((1 − βk)xk ⊕ βkT xk)) → 0, hence we have d(xk, T ((1 − βk)xk ⊕
βkT xk)) → 0. Now, set yk = (1 − βk)xk ⊕ βkT xk. We show that d(yk, T yk) → 0.
On the other hand, since T quasi nonexpansive therefore F (T ) 6= ∅ and hence
d(T xk, p) ≤ d(xk, p) for all p ∈ F (T ), therefore {T xk} is bounded. Now, since
xk
△
−→ p and βk → 0, yk
△
−→ p. Note that d(xk, T yk) → 0, hence we have
d(yk, T yk) ≤ d(yk, xk) + d(xk, T yk) = βkd(xk, T xk) + d(xk, T yk) → 0.
Now, yk
△
−→ p and demiclosedness of T imply p ∈ F (T ), i.e. p ∈ Tk F (Tk).
(cid:3)
Remark 3.1. With assumptions of Lemma 3.2, if βk → 0, then F (T ) = ∩kF (Tk).
The following corollary implies that the generated sequence by (3.1) ∆-converges to
an element of F (T ).
Theorem 3.3. Let T : C → C be a demiclosed and quasi-nonexpansive mapping.
If αk, βk ∈ (0, 1) are two sequences such that lim sup αk < 1 and βk → 0, then the
sequence xk generated by (3.1) ∆-converges to an element of F (T ).
11
Proof. It is a consequence of Lemma 3.1, Lemma 3.2, Remark 3.1 and Theorem
2.1.
(cid:3)
Now, we prove the strong convergence of the generated sequence by (3.1) to an
element of F (T ).
Theorem 3.4. Let T : C → C be a demiclosed and quasi-nonexpansive mapping.
If αk, βk ∈ (0, 1) are two sequences such that lim sup αk < 1 and βk → 0, then the
sequence {xk} generated by
xk+1 = γku ⊕ (1 − γk)Tkxk,
where {Tk} is defined by (3.2), u, x1 ∈ C and the sequence γk ∈ (0, 1) satisfies
lim γk = 0 and P+∞
k=1 γk = +∞ converges strongly to ProjF (T )u.
Proof. {Tk} is a strongly quasi-nonexpansive sequence by Lemma 3.1. Also Lemma
3.2 shows that the sequence Tk satisfies (2.1). Now, Theorem 2.3 and Remark 3.1
imply that {xk} converges strongly to ProjF (T )u.
(cid:3)
If we take βk ≡ 0 in (3.1), then we gain the Mann iteration, i.e.
xk+1 = (1 − αk)xk ⊕ αkT xk,
(3.3)
Corollary 3.5. Let T : C → C be a demiclosed and quasi-nonexpansive mapping. If
αk ∈ (0, 1) is a sequence such that lim sup αk < 1, then the sequence {xk} generated
by (3.1) ∆-converges to an element of F (T ).
12
Proof. It is a consequence of Theorem 3.3.
(cid:3)
Corollary 3.6. Let T : C → C be a demiclosed and quasi-nonexpansive mapping. If
αk ∈ (0, 1) is a sequence such that lim sup αk < 1, then the sequence {xk} generated
by
xk+1 = γku ⊕ (1 − γk)((1 − αk)xk ⊕ αkT xk),
where u, x1 ∈ C and the sequence γk ∈ (0, 1) satisfies lim γk = 0 and P+∞
converges strongly to ProjF (T )u.
k=1 γk = +∞,
Proof. A consequence of Theorem 3.4.
(cid:3)
3.2. Applications to Proximal Point Algorithms. This section contains two
subsection. First we apply our main results to proximal point algorithm to approxi-
mate a minimizer of a convex or pseudo-convex function and in the second subsection
we consider a Lipschitz quasi-nonexpansive mapping to approximate a fixed point
of it by the proximal method. In the best of our knowledge some of the results in
this section are new even in Hilbert spaces.
3.2.1. Convex and Pseudo-convex Minimization. In this subsection, we show
some applications of our main results of Theorems 2.1 and 2.3 to convex and pseudo-
convex minimization.
A function f : X →] − ∞, +∞] is called
(i) convex iff
f (tx ⊕ (1 − t)y) ≤ tf (x) + (1 − t)f (y), ∀x, y ∈ X and ∀ 0 ≤ t ≤ 1
(ii) quasi convex iff
f (tx ⊕ (1 − t)y) ≤ max{f (x), f (y)}, ∀x, y ∈ X and ∀ 0 ≤ t ≤ 1
equivalently, for each r ∈ R, the sub-level set Lf
r := {x ∈ X : f (x) ≤ r} is a convex
13
subset of X.
(iii) α-weakly convex for some α > 0 iff
f (tx ⊕ (1 − t)y) ≤ tf (x) + (1 − t)f (y) + αt(1 − t)d2(x, y), ∀x, y ∈ X and ∀ 0 ≤ t ≤ 1
(iv) pseudo-convex iff
f (y) > f (x) implies that there exist β(x, y) > 0 and 0 < δ(x, y) ≤ 1 such that
f (y) − f (tx ⊕ (1 − t)y) ≥ tβ(x, y), ∀t ∈ (0, δ(x, y)).
Definition 3.7. Let f : X →] − ∞, +∞]. The domain of f is defined by D(f ) :=
{x ∈ X : f (x) < +∞}. f is proper iff D(f ) 6= ∅.
Definition 3.8. A function f : X →] − ∞, +∞] is called (∆-)lower semicontinuous
(shortly, lsc) at x ∈ D(f ) iff
lim inf
n→∞
f (yn) ≥ f (x)
for each sequence yn → x (yn
△
−→ x) as n → +∞. f is called (∆-)lower semicontinu-
ous iff it is (∆-)lower semicontinuous in each point of its domain. It is easy to see that
every lower semicontinuous and quasi-convex function is ∆-lower semicontinuous.
Let f : X →] − ∞, +∞] be a convex, proper and lower semicontinuous (shortly,
lsc) function where X is a Hadamard space. The resolvent of f of order λ > 0 is
defined at each point x ∈ X as follows:
J f
λ x := Argminy∈X {f (y) +
1
2λ
d2(y, x)}
14
Existence and uniqueness of J f
λ x for each x ∈ X and λ > 0 was proved by Jost
(see Lemma 2 in [17]). A similar argument shows the existence and uniqueness
of the resolvent for α-weakly convex function f when λ < 1
2α . The behavior of
iterations the resolvent on an arbitrary point of a Hadamard space (named the
proximal point algorithm) was proved by Bacak [5], which extends the corresponding
result proved by Martinet [29] in Hilbert spaces (see also Rockafellar [32]). In this
section we conclude ∆- convergence of the proximal point algorithm as a consequence
of Theorem 2.1. Also we prove the strong convergence of Halpern type proximal
point algorithm as a consequence of Theorem 2.3. The last result extends a result by
Cholamjiak [13]. First we prove the sequence J f
λk
of mappings satisfies the conditions
of Theorems 2.1 and 2.3.
Lemma 3.9. Let f : X →] − ∞, +∞] be a quasi-convex, α-weakly convex, proper
and lsc function. If Argminf 6= ∅, then J f
λ is a quasi firmly nonexpansive mapping
for each λ < 1
2α .
Proof. Taking x ∈ Argminf , y = tx ⊕ (1 − t)J f
λ x and using quasi-convexity of f , we
get
f (J f
λ x)+
1
2λ
d2(J f
λ x, x) ≤ f (J f
λ x)+
1
2λ
{td2(x, x)+(1−t)d2(J f
λ x, x)−t(1−t)d2(J f
λ x, x)}
By letting t → 0+, we receive to
d2(J f
λ x, x) − d2(x, x) + d2(J f
λ x, x) ≤ 0.
Therefore h
−−−→
J f
λ xx,
−−−→
J f
λ xxi ≤ 0 which implies that d2(J f
λ x, x) ≤ h
−−−→
J f
λ xx,
is quasi firmly nonexpansive mapping.
−→
xxi. Thus J f
λ
(cid:3)
Lemma 3.10. Let f : X →] − ∞, +∞] be a quasi-convex, α-weakly convex, proper
15
and lsc function. If λ ≤ 1
2α , then Argminf ⊆ F (J f
λ ), moreover, if f is pseudo-
convex, then Argminf = F (J f
λ ).
Proof. It is clear that Argminf ⊆ F (J f
λ ). Now, we assume that f is pseudo convex
and x ∈ F (J f
λ ), but x 6∈ Argminf . Therefore there is z ∈ X such that f (x) > f (z),
hence there are β(x, z) > 0 and 0 < δ(x, z) ≤ 1 such that
f (tz ⊕ (1 − t)x) + tβ(x, z) < f (x),
∀t ∈ (0, δ(x, z))
On the other hand, since x ∈ F (J f
λ ) we have
f (x) ≤ f (tz ⊕ (1 − t)x) +
1
2λ
d2(tz ⊕ (1 − t)x, x)
Therefore we obtain
tβ(x, z) <
1
2λ
d2(tz ⊕ (1 − t)x, x) =
t2
2λ
d2(x, z)
hence β(x, z) < t
2λ d2(x, z), thus when t → 0, we gain contradiction.
(cid:3)
Remark 3.2. In Lemma 3.2 if we define f : R →] − ∞, +∞], by f (x) = 3x4 −
16x3 + 24x2, then Argminf ⊂ F (J f
λ ).
Remark 3.3. In the previous lemma, if µ < λ, then F (J f
λ ) ⊆ F (J f
µ ). By definition
of resolvent
and
f (J f
µ x) +
f (J f
λ x) +
1
2µ
1
2λ
d2(J f
µ x, x) ≤ f (J f
λ x) +
d2(J f
λ x, x) ≤ f (J f
µ x) +
1
2µ
1
2λ
d2(J f
λ x, x)
d2(J f
µ x, x)
16
By summing the above two inequalities, we conclude that
(
1
2µ
−
1
2λ
)d2(J f
µ x, x) ≤ (
1
2µ
−
1
2λ
)d2(J f
λ x, x)
which implies that F (J f
λ ) ⊆ F (J f
µ ).
Lemma 3.11. Let f : X →] − ∞, +∞] be a convex, proper and lsc function. If
lim inf λk > 0, then J f
λk
satisfies (2.1).
Proof. Let {xk} be an arbitrary sequence such that xk
△
−→ p and d(xk, J f
λk
xk) → 0.
We want to prove p ∈ Tk F (J f
λk
). Note that
f (J f
λk
xk) +
1
2λk
d2(J f
λk
xk, xk) ≤ f (y) +
1
2λk
d2(y, xk)
Set y = tJ f
λk
xk ⊕ (1 − t)z, where t ∈ (0, 1) and z ∈ X, then we have
f (J f
λk
xk) +
1
2λk
d2(J f
λk
xk, xk)
≤ tf (J f
λk
xk)+(1−t)f (z)+
1
2λk
(td2(J f
λk
xk, xk)+(1−t)d2(z, xk)−t(1−t)d2(z, J f
λk
xk))
Therefore
f (J f
λk
xk) − f (z) ≤
1
2λk
(d2(z, xk) − d2(J f
λk
xk, xk) − td2(z, J f
λk
xk))
By taking t → 1−, we can conclude that
f (J f
λk
xk) − f (z) ≤
−−−−→
zJ f
h
xk,
λk
−−−−−→
J f
xkxki
λk
1
λk
Now, by Cauchy-Schwartz inequality, we have
f (J f
λk
xk) − f (z) ≤
1
λk
d(z, J f
λk
xk)d(J f
λk
xk, xk)
Since lim inf λk > 0, taking liminf and △-lower semicontinuity of f shows that f (p) ≤
f (z) for all z ∈ X. Hence p ∈ Argminf which implies that p ∈ Tk F (J f
λk
).
(cid:3)
It is valuable that the following theorem extends the results of [13].
17
Theorem 3.12. Let f : X →] − ∞, +∞] be a convex, proper and lsc function. If
lim inf λk > 0 and Argminf 6= ∅, then the sequence {xk} generated by
xk+1 = αku ⊕ (1 − αk)J f
λk
xk,
where u, x1 ∈ C and the sequence {αk} ⊂ (0, 1) satisfies lim αk = 0 and P+∞
+∞, converges strongly to ProjArgminf u.
k=1 αk =
Proof. Lemma 3.11 implies that J f
λk
satisfies (2.1). Also by Lemma 3.9 J f
λk
is a
quasi firmly nonexpansive sequence and therefore strongly quasi-nonexpansive se-
quence. Now, Theorem 2.3 and Lemma 3.8 imply that {xk} converges strongly to
ProjArgminf u.
(cid:3)
Lemma 3.13. Suppose that f : X →] − ∞, +∞] is proper, lsc and pseudo-convex
function and lim inf λk > 0, then the sequence J f
λk
is closed.
i.e.
if xk → p and
d(J f
λk
xk, xk) → 0 as k → +∞, then p ∈ F (J f
λk
) for each k ≥ 1.
Proof. Suppose that d(J f
λk
definition of J f
λk
xk, we get
xk, xk) → 0 and xk → p as k → +∞. Then by the
f (J f
λk
xk) +
1
2λk
d2(J f
λk
xk, xk) ≤ f (y) +
1
2λk
d2(y, xk),
∀y ∈ X
By letting k → +∞ and using lower semicontinuity of f , we get: f (p) ≤ f (y) +
1
2λ d2(y, p) where lim inf λk > λ > 0. Now, set y = J f
λ p, we get f (p) ≤ f (J f
λ p) +
1
2λ d2(p, J f
λ p). By the definition of J f
λ p we get: f (p) = f (J f
λ p) + 1
2λ d2(p, J f
λ p). If
p 6= J f
λ p, then f (J f
λ p) < f (p), then there exists β(J f
λ p, p) > 0 and 0 < δ(J f
λ p, p) ≤ 1
such that f (tJ f
λ p ⊕ (1 − t)p) + tβ(J f
λ p, p) < f (p) for all t ∈ (0, δ(J f
λ p, p)). On the
18
other hand by the definition of J f
λ p, we have f (J f
λ p) + 1
2λ d2(J f
λ p, p) ≤ f (tJ f
λ p ⊕ (1 −
t)p) + t2
2λ d2(p, J f
λ p). Therefore f (p) − t2
2λ d2(p, J f
λ p) + tβ(J f
λ p, p) < f (p) by letting
t → 0 we get β(J f
λ p, p) ≤ 0 which is a contradiction. Hence p ∈ F (J f
λ ).
(cid:3)
Theorem 3.14. Let f : X →] − ∞, +∞] be an α-weakly convex, pseudo-convex,
proper and lsc function where X is a locally compact Hadamard space. Suppose that
lim inf λk > 0 and Argminf 6= ∅. Then the sequence {xk} generated by xk+1 =
J f
λk
xk (proximal point algorithm) converges to an element of Argminf .
Proof. A consequence of Theorem 2.1, Lemmas 3.7, 3.8 and 3.13 and Proposition
4.4 of [2].
(cid:3)
3.2.2. Fixed Point of a Lipschitz Quasi-nonexpansive Mapping. In this sub-
section we apply our main results in Section 2 to approximate a fixed point of a
Lipschitz quasi-nonexpansive mapping by the proximal point algorithm. Similar to
the previous section we must prove the resolvent of a Lipschitz quasi-nonexpansive
mapping satisfies the conditions of Theorems 2.1 and 2.3. First we recall the defi-
nition as well as existence and uniqueness of the resolvent. The resolvent operator
J T
λ for a nonexpansive mapping T has been defined in the literature for Hadamard
spaces (see [7, 22]). The definition for a Lipschitz mapping is similar but it exists
only for some parameters λ.
Let C ⊆ X be closed and convex. Suppose that T : C → C is a mapping and
α > 1 such that d(T x, T y) ≤ αd(x, y). For λ > 0 and x ∈ C, we define T x
λ : C → C
as
T x
λ y =
1
1 + λ
x ⊕
λ
1 + λ
T y.
Now, take y1, y2 ∈ C, then note that
19
d(T x
λ y1, T x
λ y2) = d( 1
1+λ x⊕ λ
1+λ T y1,
1
1+λ x⊕ λ
1+λ T y2) ≤ λ
1+λ d(T y1, T y2) ≤ αλ
1+λ d(y1, y2).
In the sequel, if αλ
1+λ < 1, then T x
λ is a contraction, i.e.
if λ < 1
α−1 then T x
λ has a
unique fixed point which we denote it by J T
λ x and it is called the resolvent of T of
order λ > 0 at x. In fact, J T
λ x = F (T x
λ ). It is easy to see that F (J T
λ ) = F (T ). First,
suppose that J T
λ x = x therefore x = 1
1+λ x ⊕ λ
1+λ T x which implies that T x = x.
Now, suppose T x = x hence x = 1
1+λ x ⊕ λ
1+λ T x, therefore J T
λ x = x.
Remark 3.4. If X = H a Hilbert space and T and I are respectively a nonexpansive
and identify mappings, then the resolvent of the maximal monotone operator I − T
is exactly J T
λ which was defined above.
In the sequel, we will prove the ∆-convergence of generated sequence by (3.4).
Now, let T : C → C be aa α-Lipschitz and quasi-nonexpansive mapping, where C
is closed and convex and λk < 1
α−1 , we define J T
λk
: C → C as
xk+1 = J T
λk
xk =
1
1 + λk
xk ⊕
λk
1 + λk
T (J T
λk
xk).
(3.4)
Lemma 3.15. Let T : C → C be a quasi nonexpansive mapping, then F (T ) is
closed and convex.
Proof. Take p, q ∈ F (T ) and t ∈ [0, 1], we show that tp ⊕ (1 − t)q ∈ F (T ) or
equivalently d(tp ⊕ (1 − t)q, T (tp ⊕ (1 − t)q)) = 0. Note that
d2(tp ⊕ (1 − t)q, T (tp ⊕ (1 − t)q)) ≤ td2(p, T (tp ⊕ (1 − t)q)) + (1 − t)d2(q, T (tp ⊕ (1 −
t)q))−t(1−t)d2(p, q) ≤ td2(p, tp⊕(1−t)q)+(1−t)d2(q, tp⊕(1−t)q)−t(1−t)d2(p, q) =
t(1 − t)2d2(p, q) + t2(1 − t)d2(p, q) − t(1 − t)d2(p, q) = 0, i.e F (T ) is convex.
20
Now, take pk ∈ F (T ) such that pk → p. Note that d(pk, T p) ≤ d(pk, p) → 0, i.e.
p ∈ F (T ).
(cid:3)
Lemma 3.16. Let T : C → C be a quasi-nonexpansive mapping and α-Lipschitz
with α > 1. If {λk} is a positive sequence, then J T
λk
is a strongly quasi-nonexpansive
sequence.
Proof. Take {xk} in C and p ∈ F (T ). Now, by definition of J T
λk
, we have
d2(J T
λk
xk, p) = d2(
1
1+λk
xk⊕ λk
1+λk
T (J T
λk
xk), p) ≤ 1
1+λk
d2(xk, p)+ λk
1+λk
d2(T (J T
λk
xk), p)−
λk
(1+λk)2 d2(xk, T (J T
λk
xk)) ≤ 1
1+λk
d2(xk, p) + λk
1+λk
d2(J T
λk
xk, p) − 1
λk
d2(xk, J T
λk
xk).
Therefore
d2(xk, J T
λk
xk) ≤
λk
1 + λk
(d2(xk, p) − d2(J T
λk
xk, p))
which shows J T
λk
is strongly quasi nonexpansive.
(cid:3)
Lemma 3.17. Let T : C → C be an α-Lipschitz with α > 1, demiclosed and quasi-
nonexpansive mapping. If {λk} is a positive sequence such that lim inf λk > 0, then
J T
λk
satisfies (2.1).
Proof. Let {xk} be an arbitrary sequence such that xk
△
−→ p and d(xk, J T
λk
xk) → 0.
We want to prove that p ∈ Tk F (J T
λk
). Note that d(xk, J T
λk
xk) → 0 implies that
λk
1+λk
d(xk, T (J T
λk
xk)) → 0. Since lim inf λk > 0 hence d(xk, T (J T
λk
xk)) → 0. There-
fore we have d(J T
λk
xk, T (J T
λk
xk)) → 0. Now, since T is demiclosed and J T
λk
xk
we get p ∈ Tk F (Tk).
△
−→ p,
(cid:3)
Corollary 3.18. Let T : C → C be an α-Lipschitz with α > 1, demiclosed, quasi-
nonexpansive mapping and {λk} be a positive sequence such that lim inf λk > 0. If
we define xk+1 = J T
λk
xk such that x0 ∈ C, then the sequence {xk} ∆-converges to
an element of F (T ).
21
Proof. A consequence of Lemmas 3.16, 3.17 and Theorem 2.1.
(cid:3)
Theorem 3.19. Let T : C → C be an α-Lipschitz with α > 1, demiclosed and
quasi-nonexpansive. If lim inf λk > 0 and the sequence {xk} generated by
xk+1 = αku ⊕ (1 − αk)J T
λk
xk,
where u, x1 ∈ C and the sequence {αk} ⊂ (0, 1) satisfies lim αk = 0 and P+∞
+∞, then {xk} converges strongly to ProjF (T )u.
k=1 αk =
Proof. J T
λk
is strongly quasi-nonexpansive sequence by Lemma 3.16. Also Lemma
3.17 shows that the sequence {J T
λk
} satisfies (2.1). Now, Theorem 2.3 implies that
{xk} converges strongly to ProjTk F (J T
λk
)u. The result follows because F (J T
λk
) =
F (T ).
(cid:3)
3.3. Pseudo-monotone Equilibrium Problems. Let K ⊆ X be closed and con-
vex. Suppose that f : K × K → R is a bifunction. we recall the definitions of
pseudo-monotone and θ-under monotone bifunctions.
f is called pseudo-monotone, iff
Whenever f (x, y) ≥ 0 with x, y ∈ K it holds that f (y, x) ≤ 0.
f is called θ-under monotone, iff
There exists θ ≥ 0 such that f (x, y) + f (y, x) ≤ θd2(x, y), for all x, y ∈ K.
In [19] has been shown that for a given x ∈ X and λ > θ, there is a unique point
denoted by J f
λ x such that
f (J f
λ x, y) + λh
−−−→
xJ f
λ x,
−−−→
J f
λ xyi ≥ 0,
∀y ∈ K
(3.5)
22
J f
λ x is called the resolvent of f of order λ at x ∈ X. Take a sequence of regularization
parameters {λk} ⊂ (θ, ¯λ], for some ¯λ > θ and x0 ∈ X. The proximal point algo-
rithm for approximation of an equilibrium point of f proposed by xk+1 = J f
λk
xkthat
studied by Iusem and Sosa in [16] in Hilbert spaces. In this subsection we show that
∆- convergence of the proximal point algorithm and its Halpern version to an equi-
librium point of f is a consequence of the results of Section 2 by assuming existence
of a sequence that satisfies (3.5). The set of all equilibrium point of f is denoted by
S(f, K). In (3.5), it is obvious that F (J f
λ ) ⊆ S(f, K) and if f is pseudo-monotone,
then S(f, K) ⊆ F (J f
λ ).
Lemma 3.20. Let f : K × K → R be a pseudo-monotone and θ-under monotone
bifunction and suppose that f (x, x) = 0 and f (x, ·) is lsc and convex for all x ∈ K.
If S(f, K) 6= ∅ and f (·, y) is △-upper semicontinuous for all y ∈ K, then J f
λk
is
strongly quasi-nonexpansive sequence.
Proof. Take p ∈ S(f, K) and set y = p in (3.5), we obtain
f (J f
λk
xk, p) + λkh
−−−−−→
xkJ f
xk,
λk
−−−−→
J f
λk
xkpi ≥ 0
Since p is an equilibrium point and f is pseudo-monotone, therefore f (J f
λk
xk, p) ≤ 0.
Hence
which implies that
−−−−−→
xkJ f
h
xk,
λk
−−−−→
J f
λk
xkpi ≥ 0
d2(xk, J T
λk
xk) ≤ d2(xk, p) − d2(J T
λk
xk, p)
Therefore J T
λk
is strongly quasi-nonexpansive.
(cid:3)
23
Lemma 3.21. Let f : K × K → R be a pseudo-monotone and θ-under monotone
bifunction and suppose that f (x, x) = 0 and f (x, ·) is lsc and convex for all x ∈ K. If
S(f, K) 6= ∅ and f (·, y) is △-upper semicontinuous for all y ∈ K, then J f
λk
satisfies
(2.1).
Proof. Fix y ∈ K. Let {xk} be an arbitrary sequence such that xk
△
−→ p and
d(xk, J f
λk
0 ≤ f (J f
λk
xk) → 0. We want to prove p ∈ Tk F (J f
xk, y) + λkh
xkyi ≤ f (J f
λk
−−−−−→
xkJ f
xk,
λk
−−−−→
J f
λk
). Note that
λk
xk, y) + λkd(xk, J f
λk
xk)d(J f
λk
xk, y).
Since {λk} and {xk} are bounded and lim d(J f
λk
xk, xk) = 0, we have:
0 ≤ lim inf f (J f
λk
xk, y),
∀y ∈ K.
(3.6)
On the other hand, since lim d(J f
λk
xk, xk) = 0, therefore J f
λk
xk
△
−→ p. Now since
f (·, y) is △-upper semicontinuous for all y ∈ K, we have:
0 ≤ lim inf f (J f
λk
xk, y) ≤ lim sup f (J f
λk
xk, y) ≤ f (p, y)
for all y ∈ K. So that p ∈ S(f, K), i.e. p ∈ Tk F (J f
λk
).
(cid:3)
The following theorem is one of the consequences of Section 2 (to see an indepen-
dent proof, see [19]).
Theorem 3.22. Let f : K × K → R be a pseudo-monotone and θ-under monotone
bifunction and suppose that f (x, x) = 0 and f (x, ·) is lsc and convex for all x ∈ K.
If S(f, K) 6= ∅ and f (·, y) is △-upper semicontinuous for all y ∈ K, then the se-
quence {xk} generated by (3.5), is △-convergent to an element of S(f, K).
24
Proof. It is a consequence of Lemma 3.20, Lemma 3.21 and Theorem 2.1.
(cid:3)
Take a sequence of regularization parameters {λk} ⊂ (θ, ¯λ], for some ¯λ > θ
and x0 ∈ X. Consider the following Halpern regularization of the proximal point
algorithm for equilibrium problem:
f (J f
λk
xk, y) + λkh
−−−−−→
xkJ f
xk,
λk
−−−−→
J f
λk
xkyi ≥ 0,
∀y ∈ K,
(3.7)
xk+1 = αku ⊕ (1 − αk)J f
λk
xk,
where u ∈ X and the sequence {αk} ⊂ (0, 1) satisfies lim αk = 0 and P+∞
k=1 αk = +∞.
We will prove the strong convergence of the generated sequence by (3.7) to an
equilibrium point of f by assuming existence of a sequence that satisfies (3.7). In
fact, we prove xk → x∗ = P rojS(f,K)u.
Theorem 3.23. Let f : K × K → R be a pseudo-monotone and θ-under monotone
bifunction and suppose that f (x, x) = 0 and f (x, ·) is lsc and convex for all x ∈ K.
If S(f, K) 6= ∅ and f (·, y) is △-upper semicontinuous for all y ∈ K, then {xk}
generated by (3.7) converges strongly to ProjS(f,K)u.
Proof. A consequence of Lemmas 3.20, 3.21 and Theorem 2.3.
(cid:3)
References
[1] B. Ahmadi Kakavandi, M. Amini, Duality and subdifferential for convex function on CAT(0)
metric spaces, Nonlinear Anal., 73 (2010), 3450-3455.
[2] B. Ahmadi Kakavandi, Weak topologies in complete CAT(0) spaces, Proc. Amer. Math. Soc.,
141 (2013), 1029-1039.
[3] D. Ariza-Ruiz, L. Leustean, G. L´opez-Acedo, Firmly nonexpansive mappings in classes of
geodesic spaces. Trans. Amer. Math. Soc. 366 (2014), 4299-4322.
25
[4] D. Ariza-Ruiz, G. Genaro Lopez-Acedo and A. Nicolae, The asymptotic bihavior of the com-
position of firmly nonexpansive mappings, J. Optim. Theory Appl. 167 (2015), 409-429.
[5] M. Bacak, The proximal point algorithm in metric spaces, Israel J. Math. 194 (2013), 689-701.
[6] M. Bacak, Convex Analysis and Optimization in Hadamard Spaces, De Gruyter Series in
Nonlinear Analysis and Applications, 22. De Gruyter, Berlin, 2014.
[7] M. Bacak, Miroslav; S. Reich, The asymptotic behavior of a class of nonlinear semigroups in
Hadamard spaces. J. Fixed Point Theory Appl. 16 (2014), 189-202.
[8] G. C. Bento, O.P. Ferreira and P. R. Oliveira, Local convergence of the proximal point method
for a special class of nonconvex functions on Hadamard manifolds, Nonlinear Anal. 73 (2010),
564-572.
[9] I.D. Berg, I.G. Nikolaev, On a distance between directions in an Alexandrov space of curvature
≤ K, Michigan Math. J., 45 (1998), 275-289.
[10] I. D. Berg, I.G. Nikolaev, Quasilinearization and curvature of Alexandrov spaces, Geom. Ded-
icata, 133 (2008), 195-218.
[11] M. R. Bridson and A. Haefliger, Metric Spaces of Non-positive Curvature, Springer-Verlag,
Berlin, 1999.
[12] D. Burago, Y. Burago and S. Ivanov, A Course in Metric Geometry. Graduate Studies in
Mathematics, 33. American Mathematical Society, Providence, RI, 2001.
[13] P. Cholamjiak, The modified proximal point algorithm in CAT(0) spaces, Optim. Lett. 9 (2015)
1401-1410.
[14] H. Dehghan, J. Rooin, A characterization of metric projection in Hadamard spaces with ap-
plications, J. Nonlinear Convex Anal. (to appear)
[15] 0 . P. Ferreira, P. R. Oliveira, Proximal point algorithm on Riemannian manifolds, Optim.,
Vol. 51(2), (2002) 257-270.
[16] A. N. Iusem, W. Sosa, On the proximal point method for equilibrium problems in Hilbert
spaces. Optimization, 59 (2010), 1259-1274.
[17] J. Jost, Convex functionals and generalized harmonic maps into spaces of nonpositive curvature,
Comment. Math. Helv. 70 (1995), 659-673.
26
[18] J. Jost, Nonpositive Curvature: Geometric and Analytic Aspects, Lectures Math. ETH
ZNurich, BirkhNauser, Basel (1997).
[19] H. Khatibzadeh, V. Mohebbi, Monotone and pseudo-monotone equilibrium problems in CAT(0)
spaces, submitted.
[20] H. Khatibzadeh and S. Ranjbar, On the strong convergence of Halpern type proximal point
algorithm, J. Optim. Theory Appl., 158 (2013), 385-396.
[21] H. Khatibzadeh and S. Ranjbar, Halpern type iterations for strongly quasi-nonexpansive se-
quences and its applications. Taiwanese J. Math. 19 (2015), no. 5, 1561-1576.
[22] H. Khatibzadeh and S. Ranjbar, ∆-convergence and w-convergence of the modified Mann iter-
ation for a family of asymptotically nonexpansive type mappings in complete CAT(0) spaces,
Fixed Point Theory, 17 (2016), 151-158.
[23] H. Khatibzadeh and S. Ranjbar, A variational inequality in complete CAT(0) metric spaces,
J. Fixed Point Theory Appl. 17 (2015), 557-574.
[24] W. A. Kirk, B. Panyanak, A concept of convergence in geodesic spaces, Nonlinear Anal., 68
(2008), 3689-3696.
[25] W. A. Kirk, Geodesic geometry and fixed point theory. II, in International Conference on Fixed
Point Theory and Applications, pp. 113-142, Yokohama Publishers, Yokohama, Japan, 2004.
[26] T.C. Lim, Remarks on some fixed point theorems, Proc. Amer. Math. Soc., 60 (1976), 179-182.
[27] C. Li, G. Lopez, V. Martin-Marquez and J.H. Wang, Resolvent of set valued monotone vector
fields in Hadamard manifolds, Set-Valued Anal. 19 (2011), 361-383.
[28] C. Li, G. Lopez, V. Martin-Marquez, Monotone vector fields and the proximal point algorithm
on Hadamard manifolds, J. London Math. Soc. 79 (2009), 663-683.
[29] B. Martinet, R´egularisation d'in´equations variationnelles par approximations successives, Rev.
Fran¸caise Informat. Recherche Op´erationnelle 3 (1970), 154-158.
[30] A. Nicolae, Asymptotic behavior of averaged and firmly nonexpansive mappings in geodesic
spaces, Nonlinear Anal. 87 (2013), 102-115.
[31] E.A. Papa Quiroz, P. R. Oliveira, Proximal point method for minimizing quasiconvex locally
Lipschitz functions on Hadamard manifolds, Nonlinear Anal. 75 (2012), 5924-5932.
27
[32] R. T. Rockafellar, Monotone operators and the proximal point algorithm, SIAM J. Control
Optim. 14 (1976), 877-898.
[33] S. Saejung, Halpern's iteration in CAT(0) spaces, 2010, Article ID 471781,13 pp.
[34] S. Saejung, P. Yotkaew, Approximation of zeros of inverse strongly monotone operators in
Banach spaces. Nonlinear Anal. 75 (2012) 742-750.
[35] G. Tang, L. Zhou, and N. Huang, The proximal point algorithm for pseudomonotone variational
inequalities on Hadamard manifolds, Optim. Lett. 7 (2013), no. 4, 779-790.
[36] G. Tang and Y. Xiao, A note on the proximal point algorithm for pseudomonotone variational
inequalities on Hadamard manifolds, Adv. Nonlinear Var. Inequal. 18 (2015), 58-69.
[37] J.H. Wang, G. Lopez, V. Martin-Marquez and Li, Monotone and accretive vector fields on
Riemannian manifolds, J. Optim. Theory Appl. 146 (2010), 691-708.
[38] H.K. Xu, Iterative algorithms for nonlinear operators. J. London Math. Soc., 66 (2002), 240-
256.
|
0910.2454 | 2 | 0910 | 2013-11-25T16:04:37 | The quadratic Fock functor | [
"math.FA",
"math.PR"
] | We construct the quadratic analogue of the boson Fock functor. While in the first order case all contractions on the 1--particle space can be second quantized, the semigroup of contractions that admit a quadratic second quantization is much smaller due to the nonlinearity. Within this semigroup we characterize the unitary and the isometric elements. | math.FA | math |
The quadratic Fock functor
Luigi Accardi & Ameur Dhahri
Volterra Center, University of Roma Tor Vergata
Via Columbia 2, 00133 Roma, Italy
e-mail:[email protected]
[email protected]
Abstract
We construct the quadratic analogue of the boson Fock functor.
While in the first order (linear) case all contractions on the 1 -- particle
space can be second quantized, the semigroup of contractions that
admit a quadratic second quantization is much smaller due to the
nonlinearity. The encouraging fact is that it contains, as proper sub-
groups (i.e. the contractions), all the gauge transformations of second
invertible maps of Rd into itself leaving the
kind and all the a.e.
Lebesgue measure quasi-invariant (in particular all diffeomorphism
of Rd). This allows quadratic 2-d quantization of gauge theories, of
representations of the Witt group (in fact it continuous analogue), of
the Zamolodchikov hierarchy, and much more. . . . Within this semi-
group we characterize the unitary and the isometric elements and we
single out a class of natural contractions.
1
Introduction
The boson (this specification will be omitted in the following) Fock functor
has its origins in Heisenberg commutation relations. If H is a complex Hilbert
space the Heisenberg ∗ -- Lie algebra Heis(H) is defined by generators.
{Ag, A+
f , 1 (central element) : f ∈ H}
commutation relations
[Af , A+
g ] = hf, gi · 1
;
f, g ∈ H
1
(the omitted commutation relations are zero) and involution
(Af )∗ = A+
f
;
f ∈ H
On the universal enveloping algebra of Heis(H), denoted U(Heis(H)), there
is a unique state satisfying
ϕ(1) = 1
ϕ(xAg) = 0
;
∀x ∈ U(Heis(H)) ; ∀g ∈ H
Denoting Γ(H) the GNS space of U(Heis(H)) with respect to ϕ, the map
H 7→ Γ(H) is a functor defined on the category of Hilbert spaces, with mor-
phisms given by contractions to the category of infinite dimensional Hilbert
spaces with the same morphisms.
Γ(H) is called the Fock space over H and, if V is a contraction on H its
image Γ(V ) is called the Fock second quantization of V .
The domain of Γ is maximal in the sense that, if V is not a contraction on
H, then Γ(V ) cannot be a bounded operator on Γ(H).
Our goal in this paper is to extend the picture described above, from the
Heisenberg algebra, describing the white noise commutation relations, to the
algebra describing the commutation relations of the renormalized square of
white noise.
The algebra of the renormalized square of white noise (RSWN) with test
function algebra
A := L2(Rd) ∩ L∞(Rd)
is the ∗-Lie-algebra, with central element denoted 1, generators
{B+
f , Bh, Ng : f, g, h ∈ L2(Rd) ∩ L∞(Rd)}
involution
(B+
f )∗ = Bf
,
N ∗
f = N ¯f
and commutation relations
[Bf , B+
g ] = 2chf, gi + 4N ¯f g,
[Na, B+
f ] = 2B+
af , c > 0
[B+
f , B+
g ] = [Bf , Bg] = [Na, Na′] = 0
for all a, a′, f , g ∈ L2(Rd) ∩ L∞(Rd) (the theory can be developed for more
general Hilbert algebras, but we will deal only with this case). This is a
current algebra over sl(2, R) with test function algebra A. One can prove
2
that, on the universal enveloping algebra U(RSW N) of the RSW N algebra,
there exists a unique state ϕF such that
ϕF (1) = 1
ϕF (xBg) = ϕF (xNf ) = 0
;
∀f, g ∈ A ; ∀x ∈ U(RSW N)
By analogy with the Heisenberg algebra, it is natural to call this state the
quadratic Fock state and the associated GNS space, denoted Γ2(A), the
quadratic Fock space. The Fock representation of the RSWN is characterized
by a cyclic vector Φ, also called vacuum as in the first order case, satisfying
Bf Φ = NgΦ = 0
for all f, g ∈ L2(Rd) ∩ L∞(Rd).
We refer the interested reader to [4], [5] for more details.
The extensions, to the quadratic case, of the second quantization procedure
for linear operators on A requires the solution of the following two problems:
(1) when does a linear operator on A induce a linear operator on Γ2(A)?
(2) In the cases in which the answer to problem (1) is positive, when is the
induced operator bounded (a contraction, unitary, isometric, . . .)?
By inspection on the explicit form of the scalar product of the quadratic Fock
space (see Lemma 2 below) one is led to conjecture that two classes of linear
transformations of A should induce contractions on Γ2(A):
(i) ∗ -- endomorphisms of the Hilbert algebra A
(ii) generalized gauge transformations of the form
f 7→ eαf
;
eαf (x) := eα(x)f (x) ; x ∈ Rd
where α ∈ Rd → C is a complex valued Borel function with negative
real part (the −∞ value is allowed to include functions with non full
support).
One of our main results is that these are essentially all the linear operators
on A which admit a contractive second quantization on the quadratic Fock
space.
The scheme of the present paper is the following. In section 2, we recall some
properties on the quadratic exponential vectors. Moreover, we prove that
the quadratic Fock space is an interacting Fock space with scalar product
3
given explicitly.
In section 3, we characterize those operator on the one --
particle Hilbert algebra whose quadratic second quantization is isometric
(resp. unitary).
In section 4, we show with a counter -- example that even
very simple contractions have a second quantization that is not a contraction
and we give a sufficient condition for this to happen. We also introduce the
natural candidates for the role of quadratic analogue of the free Hamiltonian
evolution and of the Ornstein -- Uhlenbeck semigroup.
2 The quadratic Fock space
For n ∈ N the quadratic n -- particle space is the closed linear span of the set
{B+n
f Φ : f ∈ L2(Rd) ∩ L∞(Rd)}
where by definition B+0
f Φ = Φ, for all f ∈ L2(Rd) ∩ L∞(Rd). The quadratic
Fock space Γ2(L2(Rd) ∩ L∞(Rd)) is the orthogonal sum of all the quadratic
n -- particle spaces. The quadratic exponential vector with test function
f ∈ L2(Rd) ∩ L∞(Rd), if it exists, is defined by
where by definition
B+n
f Φ
n!
Ψ(f ) =Xn≥0
Ψ(0) = B+0
f Φ = Φ
(1)
(2)
The following theorem was proved in [2].
Theorem 1 The quadratic exponential vector Ψ(f ) exists if kf k∞ < 1
2, and
does not exists if kf k∞ > 1
2. The set of these vectors is linearly independent
and total in Γ2(L2(Rd) ∩ L∞(Rd)). Furthermore, the scalar product between
two exponential vectors, Ψ(f ) and Ψ(g), is given by
hΨ(f ), Ψ(g)i = e− c
2 RRd ln(1−4 ¯f (s)g(s))ds
(3)
The explicit form of the scalar product between two quadratic n -- particle
vectors is due to Barhoumi, Ouerdiane, Riahi [6]. Its proof, which we include
for completeness, one needs the following preliminary result which uses the
4
identity, proved in Proposition 1 of [2]. This identity will be frequently used
in the following:
B+m
f Φ2 = c
= c
22k+1 m!(m − 1)!
((m − k − 1)!)2 kf k+1k2
2kB+(m−k−1)
f
22k+1 m!(m − 1)!
((m − k − 1)!)2 kf k+1k2
2kB+(m−1)
Φk2
f
+2mckf k2
2kB+(m−k−1)
f
Φk2
Φk2
m−1
m−1
Xk=0
Xk=1
m−2
Xk=0
= c
22k+3
+2mckf k2
m!(m − 1)!
(((m − 1) − k − 1)!)2 kf k+2k2
2kB+(m−1)
Φk2
f
2kB+((m−1)−k−1)
f
Φk2
(4)
dn
dtn(cid:12)(cid:12)(cid:12)t=0
Lemma 1 For all f, g ∈ L2(Rd) ∩ L∞(Rd) such that kf k∞ < 1
one has
2, kgk∞ < 1
2,
hB+n
f Φ, B+n
g Φi = n!
hΨ(tf ), Ψ(g)i
(5)
Proof. Let f, g ∈ L2(Rd) ∩ L∞(Rd) such that kf k∞ < 1
0 ≤ t ≤ 1, one has
2, kgk∞ < 1
2 . For all
hΨ(tf ), Ψ(g)i = Xm≥0
tm
(m!)2 hB+m
f Φ, B+m
g Φi
We now prove that, for 0 ≤ t ≤ 1, the above series can be differentiated (in
t) term by term. For all m ≥ n, one has
dn
dtn(cid:16) tm
(m!)2 hB+m
f Φ, B+m
g Φi(cid:17) =
=
So that, for 0 ≤ t ≤ 1
m!tm−n
(m!)2(m − n)!
hB+m
f Φ, B+m
g Φi
tm−n
m!(m − n)!
hB+m
f Φ, B+m
g Φi
(m!)2 hB+m
f Φ, B+m
dn
dtn(cid:16) tm
(cid:12)(cid:12)(cid:12)
g Φi(cid:17)(cid:12)(cid:12)(cid:12)
≤ Um :=
1
m!(m − n)!
5
kB+m
f ΦkkB+m
g Φk
From the identity (4) it follows that
m−2
c
22k+3
Xk=0
≤(cid:16)4m(m − 1)kf k2
kB+((m−1)−k−1)
f
In conclusion
Therefore
kB+m
f ΦkkB+m
B+m
f Φ2 ≤h4m(m − 1)kf k2
g Φk ≤ q4m(m − 1)kf k2
q4m(m − 1)kgk2
The definition of Um then implies that
m!(m − 1)!
(((m − 1) − k − 1)!)2 kf k+2k2
2kB+((m−1)−k−1)
f
Φk2
(m − 1)!(m − 2)!
(((m − 1) − k − 1)!)2 kf k+1k2
2
m−2
22k+1
Xk=0
∞(cid:17)hc
Φk2i =(cid:16)4m(m − 1)kf k2
f Φk2
∞(cid:17)kB+m
∞ + 2mkf k2ikB+(m−1)
f
Φk2
∞ + 2mkf k2
2
∞ + 2mkgk2
2kB+(m−1)
f
ΦkkB+(m−1)
g
Φk
Um ≤ p4m(m − 1)kf k2
∞ + 2mkf k2
2p4m(m − 1)kgk2
m(m − n)
∞ + 2mkgk2
2
Um−1
If f and g are non-vanishing functions, then
lim
m→∞
Um
Um−1
≤ 4kf k∞kgk∞ < 1
because kf k∞ < 1
implies that
2, kgk∞ < 1
2. Hence, the series Pm Um converges. This
dn
dtn hΨ(tf ), Ψ(g)i = Xm≥n
tm−n
m!(m − n)!
hB+m
f Φ, B+m
g Φi
Evaluating the derivative at t = 0, one obtains (5).
(cid:3)
Lemma 2 For all f, g ∈ L2(Rd) ∩ L∞(Rd) the following identity holds
hB+n
f Φ, B+n
g Φi = Xi1+2i2+...+kik=n
(n!)222n−1ci1+...+ik
i1! . . . ik!2i2 . . . kik
hf, gii1hf 2, g2ii2 . . . hf k, gkiik (6)
6
Proof. The complex linearity of the map f 7→ B+
f
λ1, λ2 ∈ C,
hB+n
λ1f Φ, B+n
λ2gΦi = ¯λn
1 λn
2 hB+n
f Φ, B+n
g Φi
implies that, for all
Therefore it will be sufficient to prove the identity (6) for all f, g ∈ L2(Rd) ∩
L∞(Rd) such that kf k∞, kgk∞ < 1
hB+n
f Φ, B+n
g Φi = n!
= n!
where
hlog(1 − 4t ¯f g)i :=
2. In this case one has
dn
dn
hΨ(tf ), Ψ(g)i
dtn(cid:12)(cid:12)(cid:12)t=0
dtn(cid:12)(cid:12)(cid:12)t=0(cid:16) exp ( − hlog(1 − 4t ¯f g)i)(cid:17)
2ZRd
log (1 − 4t ¯f (s)g(s))ds
c
(7)
(8)
(9)
Denoting h(t, s) := log (1 − 4t ¯f (s)g(s)), its k -- th derivative (in t) is
h(k)(t, s) = 22k(k − 1)!( ¯f (s))k(g(s))k(1 − 4t ¯f (s)g(s))−k
Hence, uniformly for t ≤ 1
h(k)(t, s) ≤
22k(k − 1)!f (s)kg(s)k
(1 − 4kf k∞kgk∞)k
Thus, the left hand side of (8) is integrable in s and
hh(k)(t)i = 22k(k − 1)!ZRd
( ¯f (s))k(g(s))k
(1 − 4t ¯f (s)g(s))k
ds
Putting t = 0 one finds
hh(k)(0)i = 22k(k − 1)!hf k, gki
Combining the identity (cf. Refs [6], [7])
dn
dtn eϕ(t) = Xi1+2i2+...+kik=n
22nn!
i1! . . . ik!(cid:16) ϕ(1)(t)
1! (cid:17)i1
. . .(cid:16)ϕ(k)(t)
k! (cid:17)ik
eϕ(t)
(10)
with (7), (9) and (10) one obtains
hB+n
f Φ, B+n
g Φi = n!
hΨ(tf ), Ψ(g)i
dn
dtn(cid:12)(cid:12)(cid:12)t=0
Xi1+2i2+...+kik=n
=
n!22n−1n!ci1+...+ik
i1! . . . ik!2i2 . . . kik
hf, gii1hf 2, g2ii2 . . . hf k, gkiik
7
from which (6) follows.
(cid:3)
The following theorem is an immediate consequence of Lemma 2.
Theorem 2 There is a natural ismorphism between the quadratic Fock space
Γ2(L2(Rd)∩L∞(Rd)) and the interacting Fock space ⊕∞
with scalar products:
symm{L2(Rd), h·, ·in},
n=0⊗n
hf ⊗n, g⊗nin = Xi1+2i2+...+kik=n
22n−1(n!)2ci1+...+ik
i1! . . . ik!2i2 . . . kik
hf, gii1hf 2, g2ii2 . . . hf k, gkiik
3 Quadratic second quantization of contrac-
tions
Let T be a linear operator on L2(Rd) ∩ L∞(Rd). If the map
Ψ(f ) 7→ Ψ(T f )
(11)
is well defined for all quadratic exponential vectors then, by the linear inde-
pendence of these vectors, it admits a linear extension to a dense subspace of
Γ2(L2(Rd) ∩ L∞(Rd)), denoted Γ2(T ) and called the quadratic second quan-
tization of T .
From (2) and (11) it follows that, if Γ2(T ) exists then, whatever T is, it leaves
the quadratic vacuum invariant:
Γ2(T )Φ = Φ
Lemma 3 Let T be a linear operator on L2(Rd) ∩ L∞(Rd). Then Γ2(T ) is
well defined on the set of all the quadratic exponential vectors
1
2
f ∈ L2(Rd) ∩ L∞(Rd) s.t kf k∞ <
{Ψ(f )
,
}
if and only if T is a contraction on L2(Rd) ∩ L∞(Rd) equipped with the norm
k.k∞.
Proof. Sufficiency. If T : L∞(Rd) → L∞(Rd) is a contraction, then
kT f k∞ ≤ kf k∞ < 1/2 for any test function f ∈ L2(Rd) ∩ L∞(Rd) such that
kf k∞ < 1/2. Therefore Γ2(T )Ψ(f ) is well defined.
Necessity. If Γ2(T ) is well defined, then one has kT gk∞ ≤ 1
g ∈ L2(Rd) ∩ L∞(Rd) such that kgk∞ < 1
2. By linearity T maps the open
unit k.k∞ -- ball of L2(Rd) ∩ L∞(Rd) into the closed unit k.k∞ -- ball of L2(Rd) ∩
L∞(Rd) , i.e. it is a contraction.
(cid:3)
2, for any
8
3.1
Isometric and unitarity characterization of the quadratic
second quantization
Let us start by giving a sufficient condition on T , which ensures that Γ2(T )
is an isometry (resp. unitary operator).
A Hilbert algebra endomorphism (resp. automorphism) T of
L2(Rd) ∩ L∞(Rd) is said to be a ∗-endomorphism (resp. ∗-automorphism) if
T is an isometry (resp. a unitary operator) with respect to the pre-Hilbert
structure of L2(Rd) ∩ L∞(Rd), which satisfies
T (f g) = T (f )T (g), (T (f ))∗ = T ( ¯f ).
The following proposition is an immediate consequence of Lemma 2.
Proposition 1 If α : Rd → R is a Borel function, T1 is a ∗-endomorphism
of L2(Rd) ∩ L∞(Rd) and
T := eiαT1
then Γ2(T ) is an isometry. Moreover, if T1 is a ∗-automorphism of
L2(Rd) ∩ L∞(Rd), then Γ2(T ) is unitary.
Proof. To prove that Γ2(T ) is an isometry it is sufficient to prove that it
preserves the scalar product of two arbitray quadratic exponential vectors.
From (1) and the mutual orthogonality of different n -- particle spaces, it will
be sufficient to prove that, for each n ∈ N and f, g ∈ L2(Rd) ∩ L∞(Rd) one
has:
f Φ, B+n
and, because of Lemma 2, this identity follows from
T g Φi = hB+n
g Φi
hB+n
T f Φ, B+n
h(T f )k, (T g)ki = hf k, gki
;
∀k ∈ N
;
∀f, g ∈ L2(Rd)∩L∞(Rd)
But this identity holds because our assumptions on T imply that
h(T f )k, (T g)ki = heikα(T1f )k, eikα(T1g)ki = hT1(f k), T1(gk)i = hf k, gki
Thus Γ2(T ) is an isometry. If, in addition, T1 is a ∗-automorphism of
L2(Rd) ∩ L∞(Rd), then T is surjective. Hence the range of Γ2(T ), containing
all the quadratic exponential vectors, is the whole quadratic Fock space. The
thesis then follows because an isometry with full range is unitary.
(cid:3)
In the following our goal is to prove the converse of the above proposition.
9
Lemma 4
i) If Γ2(T ) is a unitary operator, then
h(T f )n, (T g)ni = hf n, gni
(12)
for all n ∈ N∗ and f, g ∈ L2(Rd) ∩ L∞(Rd).
ii) If Γ2(T ) is an isometry, then for all n ∈ N∗ and f ∈ L2(Rd) ∩ L∞(Rd)
k(T f )nk2 = kf nk2
Proof. Suppose that Γ2(T ) is a unitary operator. Let us fix two functions
f, g ∈ L2(Rd) ∩ L∞(Rd) such that kf k∞ < 1
2. Then, one has
2, kgk∞ < 1
hΨ(T f ), Ψ(T g)i = hΨ(f ), Ψ(g)i
It follows that
hΨ(tT f ), Ψ(T g)i = hΨ(tf ), Ψ(g)i
for all t such that t < 1. Therefore, Lemma 1 implies that
hB+n
T f Φ, B+n
T g Φi = hB+n
f Φ, B+n
g Φi
(13)
for all n ∈ N. Let us prove the statement i) by induction.
- For n = 1, we have
hB+
T f Φ, B+
T gΦi = hB+
f Φ, B+
g Φi
This gives
hT f, T gi = hf, gi
- Suppose that (12) holds for k ≤ n. Then, from (13) and the identity (4),
one obtains
hB+(n+1)
n
T f
= c
T g
Φi
Φ, B+(n+1)
22k+1 n!(n + 1)!
((n − k)!)2 h(T f )k+1, (T g)k+1ihB+(n−k)
T f
Φ, B+(n−k)
T g
Φi
= 22n+1n!(n + 1)!c h(T f )n+1, (T g)n+1i
Xk=0
n−1
Xk=0
+c
22k+1 n!(n + 1)!
((n − k)!)2 h(T f )k+1, (T g)k+1ihB+(n−k)
T f
Φ, B+(n−k)
T g
Φi
= 22n+1n!(n + 1)!c hf n+1, gn+1i
+c
n−1
Xk=0
22k+1 n!(n + 1)!
((n − k)!)2 hf k+1, gk+1ihB+(n−k)
f
Φ, B+(n−k)
g
Φi
10
By the induction assumption, one has
c
n−1
Xk=0
= c
22k+1 n!(n + 1)!
((n − k)!)2 h(T f )k+1, (T g)k+1ihB+(n−k)
T f
Φ, B+(n−k)
T g
Φi
22k+1 n!(n + 1)!
((n − k)!)2 hf k+1, gk+1ihB+(n−k)
f
Φ, B+(n−k)
g
Φi
n−1
Xk=0
which implies that
h(T f )n+1, (T g)n+1i = hf n+1, gn+1i
;
∀n ∈ N∗
Thus (12) holds for all n ∈ N∗.
The proof of statement ii) is obtained by replacing, in the above argument,
the test function g by f .
(cid:3)
Lemma 5 Suppose that Γ2(T ) is an isometry. Then, for any I ⊂ Rd such
that I < ∞, one has
T (χI)(x) = 1
on supp(T (χI)) a.e.
Proof. By assumption Γ2(T ) is an isometry, hence from Lemma 4, ∀n ∈ N:
h(T (χI))n, (T (χI))ni = h(χI)n, (χI)ni = hχI, χIi = I
(14)
for any subset I ⊂ Rd such that I < ∞. But, one has
h(T (χI))n, (T (χI))ni = {x ∈ Rd, T (χI)(x) = 1} +ZJ
T (χI)(x)2ndx (15)
where · denotes Lebesgue measure and
J := {x ∈ Rd, T (χI)(x) 6= 1 and T (χI)(x) > 0}
Since the identity (15) holds ∀n ∈ N, it follows that
ZJ
T (χI(x)2ndx =ZJ
T (χI(x)2(n+1)dx
;
∀n ∈ N
But it is not difficult to prove that this is impossible if J > 0.
(cid:3)
11
Lemma 6 If I ⊂ Rd such that I < ∞ and Γ2(T ) is an isometry, then there
exist a function αI : Rd → R and a subset τ (I) ⊂ Rd such that
T (χI) = eiαI χτ (I)
and I = τ (I). Moreover, if I1, I2 is an arbitrary partition of I, then
τ (I) = τ (I1) ∪ τ (I2)
,
a.e.
(16)
In particular, if I1 ⊂ I, then a.e. τ (I1) ⊂ τ (I).
Proof. Lemma 5 implies that there exist a function αI : Rd → R and a
subset τ (I) ⊂ Rd such that T (χI) = eiαI χτ (I). From (14) one has
τ (I) = hT (χI), T (χI)i = hχI, χIi = I
Let I1, I2 be a partition of I. From χI = χI1∪I2 = χI1 + χI2, it follows that
T (χI) = T (χI1) + T (χI2)
i.e.
eiαI χτ (I) = eiαI1 χτ (I1) + eiαI2 χτ (I2)
Multiplying both sides by χτ (I1)∪τ (I2), one finds
eiαI χτ (I)∩[τ (I1)∪τ (I2)] = eiαI1 χτ (I1) + eiαI2 χτ (I2) = eiαI χτ (I)
Therefore, one has τ (I) = τ (I1) ∪ τ (I2) a.e. Since the partition I1, I2 of I is
arbitrary, it follows that I1 ⊂ I implies that τ (I1) ⊂ τ (I).
(cid:3)
Lemma 7 If Γ2(T ) is an isometry and I1, I2 ⊂ Rd are such that
I1 < ∞, I2 < ∞ and I1 ∩ I2 = 0, then τ (I1) ∩ τ (I2) = 0.
Proof. Suppose that I1 ∩ I2 = 0. Then, from the identity
χI1∪I2 = χI1 + χI2 − χI1∩I2
χI1∪I2 = χI1 + χI2
it follows that, a.e.
and therefore also
T (χI1∪I2) = T (χI1) + T (χI2)
;
a.e
12
Applying (14) one then gets
I1 + I2 = hχI1∪I2, χI1∪I2i
= hT (χI1∪I2), T (χI1∪I2)i
= hT (χI1), T (χI1)i + hT (χI2), T (χI2)i
+hT (χI1), T (χI2)i + hT (χI2), T (χI1)i
= I1 + I2 +Zτ (I1)∩τ (I2)
ei(αI2
−αI1 )(x)dx
+Zτ (I1)∩τ (I2)
e−i(αI2
−αI1 )(x)dx
= I1 + I2 + 2Zτ (I1)∩τ (I2)
cos((αI2 − αI1)(x))dx
(17)
which implies that
Zτ (I1)∩τ (I2)
cos((αI2 − αI1)(x))dx = 0
(18)
Put I = I1 ∪ I2. From the identities
eiαI χτ (I) = eiαI1 χτ (I1) + eiαI2 χτ (I2)
τ (I) = τ (I1) ∪ τ (I2)
a.e
it follows that if x ∈ τ (I1) ∩ τ (I2), then
eiαI (x) = eiαI1 (x) + eiαI2 (x)
Thus, one obtains
This gives
ei(αI (x)−αI1 (x) = 1 + ei(αI2 (x)−αI1 (x))
1 = 1 + ei(αI2 (x)−αI1 (x))2 = 2 + 2cos(αI2(x) − αI1(x))
which yields that
cos(αI2(x) − αI1(x)) = −
1
2
This, together with (18) implies that τ (I1) ∩ τ (I2) = 0.
(cid:3)
13
Lemma 8 In the notations and assumptions of Lemma 6, for any I ⊂ Rd
such that I < ∞ and any I1 ⊂ I one has
eiαI1 χτ (I1) = eiαI χτ (I1)
for almost any x ∈ τ (I1).
Proof. Let I2 = I \ I1. Arguing as in the proof of of Lemma 6 one finds that
eiαI χτ (I) = eiαI1 χτ (I1) + eiαI2 χτ (I2)
Thus, if we multiply the two sides in the above identity by χτ (I1), then from
Lemmas 6, 7, it follows that
eiαI χτ (I1) = eiαI1 χτ (I1)
,
a.e
(cid:3)
Lemma 9 In the notations and assumptions of Lemma 6 there exists a func-
tion α : Rd → R such that for any I ⊂ Rd, with I < ∞
where τ (I) ⊂ Rd and τ (I) = I.
T (χI) = eiαχτ (I)
Proof. Let (In)n be an increasing sequence of subsets of Rd such that
In < ∞, ∀n ∈ N and Sn∈N In = Rd. Define the function α : Rd → R by
α(x) = αIn(x), for any n ∈ N such that x ∈ In, where αIn is defined as in
Lemma (6). Then α is well defined because, denoting
n(x) := min{n ∈ N, x ∈ In}
;
x ∈ Rd
Lemma 8 implies that, for any m, n ∈ N such that n(x) ≤ m ≤ n,
eiαIm χτ (Im) = eiαIn χτ (Im)
In particular, for any n ≥ n(x), one has
which implies that
eiαIn χτ (In(x)) = eiαIn(x) χτ (In(x))
αIn(x) = αIn(x), ∀n ≥ n(x)
This ends the proof of the above lemma.
(cid:3)
Using all together Proposition 1, Lemmas 4, 6 and 9, we prove the fol-
lowing.
14
Theorem 3 Γ2(T ) is an isometry (resp. unitary) if and only if there exist
a function α from Rd to R and a ∗-endomorphism (resp. ∗-automorphism)
T1 of L2(Rd) ∩ L∞(Rd) such that
T = eiαT1
Proof. Sufficiency has been proved in Proposition 1.
Necessity. Suppose that Γ2(T ) is an isometry. Then, from Lemma 4, T is an
isometry. Moreover, Lemma 9 implies that there exists a function α : Rd → R
such that for any I ⊂ Rd, I < ∞
T (χI) = eiαχτ (I)
where τ (I) ⊂ Rd and τ (I) = I. Define the map T1 by:
T1 : χI ∈ L2(Rd) ∩ L∞(Rd) → T1(χI) := χτ (I)
(19)
for all I ⊂ Rd such that I < ∞.
In order to prove that T1 extends,
by linearity and continuity, to a ∗-endomorphism of L2(Rd) ∩ L∞(Rd), it is
sufficient to prove that for all I, J ⊂ R with I < ∞, J < ∞
T1(χIχJ ) = T1(χI)T1(χJ ) = χτ (I)χτ (J) = χτ (I)∩τ (J)
,
a.e
(20)
But, by definition of T1 one has
T1(χIχJ ) = T1(χI∩J ) = χτ (I∩J)
therefore our thesis is equivalent to
τ (I) ∩ τ (I) = τ (I ∩ J)
,
a.e
(21)
Finally, since from Lemma 6 we know that τ (I ∩ J) ⊂ τ (I) ∩ τ (J), (21) will
follow if we prove that
τ (I) ∩ τ (J) = τ (I ∩ J)
(22)
To prove (22) notice that, since T , hence T1, is an isometry, one has
hT1(χI∪J ), T1(χI∪J )i = hχI∪J, χI∪J i = I + J − I ∩ J
(23)
15
On the other hand, from Lemma 6 we know that the map I 7→ τ (I) is finitely
addditive, hence monotone. Therefore, using linearity, (19) and the identity
χI∪J = χI + χJ − χI∩J , we find
hT1(χI∪J ), T1(χI∪J )i = hT1(χI) + T1(χJ ) − T1(χI∩J), T1(χI)
+T1(χJ ) − T1(χI∩J )i
= hT1(χI), T1(χI)i + hT1(χI), T1(χJ )i
−hT1(χI), T1(χI∩J )i + hT1(χJ ), T1(χI)i
+hT1(χJ ), T1(χJ )i − hT1(χJ ), T1(χI∩J )i
−hT1(χI∩J ), T1(χI)i − hT1(χI∩J ), T1(χJ )i
+hT1(χI∩J ), T1(χI∩J)i
= hχτ (I), χτ (I)i + hχτ (I), χτ (J)i − hχτ (I), χτ (I∩J)i
+hχτ (J), χτ (I)i + hχτ (J), χτ (J)i − hχτ (J), χτ (I∩J)i
−hχτ (I∩J), χτ (I)i − hχτ (I∩J), χτ (J)i
+hχτ (I∩J), χτ (I∩J)i
Using the isometry property and the fact that τ (I ∩ J) ⊆ τ (I) ∩ τ (J), we see
that this expression is equal to
I + τ (I) ∩ τ (J) − τ (I ∩ J) + τ (I) ∩ τ (J) + J − τ (I ∩ J)
−τ (I ∩ J) − τ (I ∩ J) + τ (I ∩ J)
= I + J + 2τ (I) ∩ τ (J) − 3τ (I ∩ J)
= I + J + 2τ (I) ∩ τ (J) − 3I ∩ J
Since this is equal to the right hand side of (23), we conclude that
−I ∩ J = 2τ (I) ∩ τ (J) − 3I ∩ J ⇔ τ (I) ∩ τ (J) = I ∩ J = τ (I ∩ J)
which is equivalent to (22) and therefore to (20).
Since a unitary operator is an isometry we conclude that, if Γ2(T ) is unitary,
then T1, defined by (19), is an invertible ∗-endomorphism of L2(Rd)∩L∞(Rd),
i.e. a ∗-automorphism.
(cid:3)
4 Quadratic second quantization of contrac-
tions
We will use the following remark.
Remark Let A = (aij)i,j, B = (bij)i,j, C = (cij)i,j and D = (dij)i,j be
16
matrices such that, in the operator order:
0 ≤ A ≤ B, and 0 ≤ C ≤ D
Then by Schur's Lemma
0 ≤ ((bij − aij)cij)i,j ⇔ (aijcij)i,j ≤ (bijcij)i,j
0 ≤ (bij(dij − cij))i,j ⇔ (bijcij)i,j ≤ (bijdij)i,j
Consequently one has
0 ≤ (aijcij)i,j ≤ (bijdij)i,j
(24)
Theorem 4 The set of all operators T such that Γ2(T ) is a contraction is a
multiplicative semigroups denoted Contr2(L2 ∩ L∞). Moreover
Γ2(S)Γ2(T ) = Γ2(ST )
; ∀S, T ∈ Contr2(L2 ∩ L∞)
(25)
Proof. Let S, T ∈ Contr2(L2∩L∞). Then Γ2(S), Γ2(T ) and hence Γ2(S)Γ2(T )
is a contraction on Γ2(L2 ∩ L∞). Therefore it is uniquely determined by its
value on the quadratic exponential vectors. If Ψ(f ) is such a vector, then
Γ2(S)Γ2(T )Ψ(f ) = Γ2(S)Ψ(T f ) = Ψ(ST f ) = Γ2(ST )Ψ(f )
Thus Γ2(ST ) is a contraction and (25) holds.
Now, we prove the following.
(cid:3)
Proposition 2 If T = MϕT1 is a contraction for L2(Rd) and L∞(Rd), where
ϕ ∈ L2(Rd) ∩ L∞(Rd) such that kϕk∞ ≤ 1 and T1 is an homomorphism of
L2(Rd) ∩ L∞(Rd), then Γ2(T ) is a contraction.
Proof. We have
kΓ2(T )(α1Ψ(f1) + . . . + αlΨ(fl))k2 = kα1Ψ(T f1) + . . . + αlΨ(T fl)k2
l
=
Xi,j=1
= Xn≥0
17
¯αiαjhΨ(T fi), Ψ(T fj)i
(26)
1
(n!)2h
l
Xi,j=1
¯αiαjhB+n
T fi
Φ, B+n
T fj
Φii.
Put
An,T =(cid:16)hB+n
T fi
Φ, B+n
T fj
Φi(cid:17)i,j
, An =(cid:16)hB+n
fi
Φ, B+n
fj
.
Φi(cid:17)i,j
Now, our purpose is to prove, under the assumptions of the above proposition,
that
0 ≤ An,T ≤ An,
(27)
for all n ∈ N.
Note that, for v = (α1, . . . , αl), one has
l
hv, An,T vi =
¯αiαjhB+n
T fi
Φ, B+n
T fj
Φi
Xi,j=1
= kα1B+n
T f1Φ + . . . + αlB+n
T fl
Φk2.
This implies that An,T is a positive matrix. Now, let us prove the second
inequality in (27) by induction on n.
- For n = 1, one has
l
hv, A1,T vi =
¯αiαjhB+
T fi
Φ, B+
T fj
Φi
Xi,j=1
Xi,j=1
l
= 2c
¯αiαjhT fi, T fji
= 2ckT (α1f1 + . . . + αlfl)k2
2
≤ 2ckα1f1 + . . . + αlflk2
2.
Because
2ckα1f1 + . . . + αlflk2
2 =
l
Xi,j=1
one obtains that A1,T ≤ A1.
¯αiαjhB+
fi
Φ, B+
fj
Φi = hv, A1vi,
- Let n ≥ 1 and suppose that An,T ≤ An. Note that for any f, g ∈
L2(Rd) ∩ L∞(Rd), Proposition 1 of [2] implies that
hB+(n+1)
f
Φ, B+(n+1)
g
Φi = c
22k+1 n!(n + 1)!
((n − k)!)2 hf k+1, gk+1i
Φ, B+(n−k)
Φi.
g
hB+(n−k)
f
n
Xk=0
18
Then, one gets
hv, An+1,T vi =
¯αiαjhB+(n+1)
T fi
Φ, B+(n+1)
T fj
Φi
= c
22k+1 (n + 1)!n!
((n − k)!)2
l
n
Xi,j=1
Xk=0
h
l
Xi,j=1
¯αiαjh(T fi)k+1, (T fj)k+1ihB+(n−k)
T fi
Φ, B+(n−k)
T fj
Φii.
Put
This gives
Mk = (hf k+1
i
, f k+1
j
i)i,j, Mk,T = (h(T fi)k+1, (T fj)k+1i)i,j.
l
hv, Mk,T vi =
¯αiαjh(T fi)k+1, (T fj)k+1i
Xi,j=1
= kα1(T f1)k+1 + . . . + αl(T fl)k+1k2
2
)k2
= kϕk+1T1(α1f k+1
2
≤ kϕkk
∞kT (α1f k+1
)k2
2
1 + . . . + αlf k+1
≤ kα1f k+1
1 + . . . + αlf k+1
1 + . . . + αlf k+1
k2
2 = hv, Mkvi.
l
l
l
This proves that
0 ≤ Mk,T ≤ Mk.
Note that by induction assumption
0 ≤ An−k,T ≤ An−k,
(28)
(29)
for all k = 0, . . . , n. Therefore, Lemma 24 and identies (28), (29) implies
that
An+1,T ≤ An+1.
Hence, we have proved that
hv, An,T vi =
l
Xi,j=1
¯αiαjhB+n
T fi
Φ, B+n
T fj
Φi ≤ hv, Anvi =
¯αiαjhB+n
fi
Φ, B+n
fj
Φi,
l
Xi,j=1
19
for all n. After using (26), it is easy to conclude that Γ2(T ) is a contraction.
(cid:3)
Remark
The contractions considered in Proposition 2 are very special, however they
are sufficient to prove the existence of the quadratic free Hamiltonian and
the quadratic Ornstein -- Uhlenbeck semigroup. In fact taking
T = T = ez1A
with Re(z) ≤ 0 where A = L2(Rd) ∩ L∞(Rd), Proposition 2 implies that
Γ2(ez1A) is a holomorphic semigroup which, by the remark done at the be-
ginning of section (3), leaves the vacuum vector Φ invariant. In particular,
for z = it, t ∈ R, the generator H0 of the strongly continuous 1 -- parameter
unitary group
Γ2(eit1A) = eitH0
is the quadratic analogue of the free Hamiltonian. By analytic continuation
one has
Γ2(ez1A) = ezH0
Moreover Lemma 2 shows that its action on the n -- particle space is the same
as the action of the number operator in the usual Fock space, i.e. it is reduced
to multiplication by
ezn
Thus H0 is the positive self -- adjoint operator characterized by the property
that, for any n ∈ N, the n -- particle space is the eigenspace of H0 correspond-
ing to the the eigenvalue n.
By considering the action of the number operators Nf , defined at the be-
ginning of section (1), one easily verifies that the definition of Nf can be
extended to the case in which f is a multiple of the identity function 1, so
that N1 is well defined. With this notation one has the identity
H0 =
1
2
N1
Using the functional realization of the quadratic Fock space given by Theorem
2 it is clear that the contraction semigroup
Γ2(e−t1A) = e−tH0
20
is positivity preserving and its explicit form gives that
Γ2(e−t1A)1 = e−t1 ≤ 1
(here we are extending in the obvious way the action of Γ2(e−t1A) to the mul-
tiples of the identity function which is not in L2(Rd) ∩ L∞(Rd)). This means
that the semigroup e−tH0 is sub -- Markovian. The above discussion shows
that e−tH0 is a natural candidate for the role of quadratic analogue of the
Ornstein -- Uhlenbeck semigroup. A more detailed analysis of this semigroup
and of its properties will be discussed elsewhere.
4.1 A counterexample
In this subsection, we discuss the behavior of contractions under quadratic
second quantization.
Lemma 10 Let T be a linear operator on L2(Rd) ∩ L∞(Rd). If T is a con-
traction on L2(Rd) and on L∞(Rd), then for any quadratic exponential vector
Ψ(f ) one has
kΓ2(T )Ψ(f )k ≤ kΨ(f )k
Proof. Recall that
kΓ2(T )Ψ(f )k2 = kΨ(T f )k2 =Xn≥0
kB+n
T f Φk2
(n!)2
(30)
(31)
and that, because of Lemma 2:
kB+n
T f Φk2 = Xi1+2i2+...+kik=n
22n−1(n!)2ci1+...+ik
i1! . . . ik!2i2 . . . kik
kT f ki1
2
2 k(T f )2ki2
2 . . . k(T f )kkik
If T is a contraction on L2(Rd) and on L∞(Rd) then by the Riesz -- Thorin
Theorem, for all p ≥ 2, T is also a contraction from Lp(Rd) into itself.
Therefore, for any p ≥ 1 and i ∈ N:
k(T f )pki
2 ="(cid:18)Z T f 2p(cid:19)1/2p#pi
= kT f kpi
2p ≤ 1
21
for all j = 1, . . . , k. This proves that for any n ∈ N
kB+n
T f Φk2 ≤ kB+n
f Φk2
and, in view of (31), this implies (30).
(cid:3)
From Lemma (10) it follows that the fact that T is a contraction for
L2(Rd) and for L∞(Rd) is a necessary condition for Γ2(T ) to be a contrac-
tion. The following counterexample shows that this condition is not sufficient.
Define the linear operator T : L2(R) ∩ L∞(R) → L2(R) ∩ L∞(R) by
T f =(cid:16)Z 1
0
f (t)dt(cid:17) χ[0,1]
It is easy to verify that T is a contraction in both L2 and L∞. Therefore,
from Lemma 10, one has
kΓ2(T )Ψ(f )k ≤ kΨ(f )k
In the following we will show that some linear combinations of quadratic
exponential vectors violate the inequality
kΓ2(T )(cid:16)Xi
αiΨ(fi)(cid:17)k ≤ kXi
αiΨ(fi)k
In fact taking
f1 := λχ[0, 1
2 ]
;
f2 := λχ[0,1]
;
λ ∈ R ;
λ <
1
2
one has
and
T f1 =
λ
2
χ[0,1]
;
T f2 = λχ[0,1]
kΓ2(T )(cid:16)α1Ψ(f1) + α2Ψ(f2)(cid:17)k2 = h(cid:18) α1
kα1Ψ(f1) + α2Ψ(f2)k2 = h(cid:18) α1
α2 (cid:19)i,
α2 (cid:19) , B(cid:18) α1
α2 (cid:19) , A(cid:18) α1
α2 (cid:19)i
where the matrices A, B are defined by:
A := (hΨ(fi), Ψ(fj)i)1≤i,j≤2
;
B := (hΨ(T fi), Ψ(T fj)i)1≤i,j≤2
22
The contraction condition
kΓ2(T )(cid:16)α1Ψ(f1) + α2Ψ(f2)(cid:17)k2 ≤ kα1Ψ(f1) + α2Ψ(f2)k2
is equivalent to say that B ≤ A. In the following we prove that this inequality
is not true. In fact recalling (3), i.e.
hΨ(f ), Ψ(g)i = e− c
2 RR ln(1−4 ¯f (x)g(x))dx
one finds
A =(cid:18) (
(
1
1−4λ2 )
1−4λ2 )
1
c
4
c
4
(
(
1
1−4λ2 )
1−4λ2 )
1
c
4
c
2 (cid:19) , B =(cid:18) ( 1
1−λ2 )
1−2λ2 )
(
1
c
2
c
2
( 1
1−2λ)
1
(
1−4λ2 )
c
2
c
2 (cid:19)
and a simple calculation proves that det(A − B) ≤ 0.
Acknowledgments
Ameur Dhahri gratefully acknowledges stimulating discussions with Uwe
Franz and Eric Ricard.
References
[1] L. Accardi, Y. G. Lu and I. V. Volovich: White noise approach to clas-
sical and quantum stochastic calculi, Centro Vito Volterra, Universit`a
di Roma "Tor Vergata", preprint 375, 1999.
[2] L. Accardi, A. Dhahri: The quadratic exponential vectors, J. Math.
Phys, to appear.
[3] L. Accardi, A. Dhahri and M. Skeide: Extension of quadratic expo-
nential vectors, submitted to: Proceedings of the 29-th Conference on
Quantum Probability and related topics , Hammamet (2008).
[3b] Accardi L., Skeide M.: On the relation of the Square of White Noise and
the Finite Difference Algebra, IDA -- QP (Infinite Dimensional Analysis,
Quantum Probability and Related Topics) 3 (2000) 185 -- 189 Volterra
Preprint N. 386 (1999)
23
[4] L. Accardi, U. Franz and M. Skeide: Renormalized squares of white noise
and other non-Gaussian noises as Levy processes on real Lie algebras,
Commun. Math. Phys. 228 (2002) 123-150.
[5] L. Accardi, G. Amosov and U. Franz: Second quantization automor-
phisms of the renormalized square of white noise (RSWN) algebra, Inf.
Dim. Anl, Quantum Probability and related topics, Vol. 7, No. 2 (2004)
183-194.
[6] A. Barhoumi, H. Ouerdiane, A. Riahi: Unitary representations of the
Witt and sl(2, R)-Algebras through Renormalized Powers of the Quan-
tum Pascal White Noise, Inf. Dim. Anl, Quantum Probability and related
topics, Vol. 11, No 3 (2008) 323-350.
[7] N. Bourbaki:
Fonctions d'une
variable R´eelle. El´ements de
Math´ematiques, Livre Vol. IV, Hermann et Cie, 1950.
[8]
I. Gradstein, I. Ryshik: Tables of series, Products and integrals, Vol. 1,
Verlag Harri Deutsch Thun 1981.
[9] K. R. Parthasarathy: An Introduction to Quantum Stochastic Calculus.
Birkhauser Verlag: Basel. Boston. Berlin.
24
|
1706.06043 | 1 | 1706 | 2017-06-19T16:33:16 | Path and quasihomotopy for Sobolev maps between manifolds | [
"math.FA"
] | We study the relationship between quasihomotopy and path homotopy for Sobolev maps between manifolds. We employ singular integrals on manifolds to show that, in the critical exponent case, path homotopy implies quasihomotopy - and observe the rather surprising fact that $n$-quasihomotopic maps need not be path homotopic. We also study the case where the target is an aspherical manifold, e.g. a manifold with nonpositive sectional curvature, and the contrasting case of the target being a sphere. | math.FA | math | PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS
BETWEEN MANIFOLDS
ELEFTERIOS SOULTANIS
Abstract. We study the relationship between quasihomotopy and path ho-
motopy for Sobolev maps between manifolds. We employ singular integrals on
manifolds to show that, in the critical exponent case, path homotopy implies
quasihomotopy -- and observe the rather surprising fact that n-quasihomotopic
maps need not be path homotopic. We also study the case where the target is
an aspherical manifold, e.g. a manifold with nonpositive sectional curvature,
and the contrasting case of the target being a sphere.
.
A
F
h
t
a
m
[
1
v
3
4
0
6
0
.
6
0
7
1
:
v
i
X
r
a
1. Introduction
Let M and N be compact Riemannian manifolds with n = dim M ≥ 2. The
study of harmonic and p-harmonic maps between M and N naturally leads to
questions about homotopies between finite energy Sobolev maps [10, 9, 8, 36, 29].
However classical homotopy is incompatible with Sobolev maps: on one hand
Sobolev maps need not be continuous, and on the other classical homototopy classes
are not stable under convergence in the Sobolev norm. Indeed, an easy example
by B. White [37] showed that the identity map S3 → S3 is homotopic to maps of
arbitrarily small energy, whilst not being homotopic to a constant map.
F. Burstall, in [5], studied energy minimization within classes of maps with
prescribed 1-homotopy class, and White [37] introduced the notion of d-homotopy
for an integer d ≤ n = dim M .
Two maps u, v ∈ W 1,p(M ; N ) are d-homotopic, d < p, if the restrictions of u and v
to a d-skeleton of a generic triangulation of M (which are continuous by the Sobolev
embedding theorem) are classically homotopic.
White proved [37, 38] that Sobolev maps u ∈ W 1,p(M ; N ) (p ≤ n) have a well
defined (⌊p⌋ − 1)-homotopy type (i.e.
the homotopy class of the restriction of
u does not depend on the generic (⌊p⌋ − 1)-dimensional skeleton) that is stable
under weak convergence in W 1,p(M ; N ), and therefore well suited for variational
minimization problems.
Connections of of d-homotopy with the topology of the Sobolev space W 1,p(M ; N )
are already visible in [37]. The notion of path homotopy, introduced by H. Brezis
and Y. Li in [3] utilizes this idea.
Two maps u, v ∈ W 1,p(M ; N ) (1 < p < ∞) are path homotopic if there exists a
continuous path h ∈ C([0, 1]; W 1,p(M ; N )) joining u and v.
Date: November 16, 2018.
Key words and phrases. Function spaces, Sobolev mappings, Riemannian manifolds,
Homotopy.
This research was conducted in the University of Jyvaskyla and at IMPAN, Warsaw.
1
2
ELEFTERIOS SOULTANIS
They proved [3, Theorem 0.2] that W 1,p(M ; N ) is always path connected when
1 < p < 2, while a deep result of Hang and Lin [15] states that, for 1 < p < n,
two maps u, v ∈ W 1,p(M ; N ) are path homotopic if and only if they are (⌊p⌋ − 1)-
homotopic.
When p = n this equivalence does not remain valid.
Instead, Sobolev maps
u ∈ W 1,n(M ; N ) have well-defined homotopy classes (due to the density of Sobolev
maps [30] and a result of White [37], see Theorem 2.1.) When p > n the Sobolev
embedding implies that Sobolev maps are continuous and indeed by results in Ap-
pendix A in [3] path homotopy is equivalent to classical homotopy.
With the emergence of analysis on metric spaces (see [14, 18, 17, 33] and the mono-
graphs [1, 19]) the study of energy minimization problems between more general
spaces has become viable. The first steps in this direction were taken by N. Kore-
vaar and R. Schoen [26] -- who studied the existence of minimizers of 2-energy in
homotopy classes of maps from a manifold to a nonpositively curved metric space
(see [4]) -- and J. Jost [20, 21, 22, 23] who studied the related problem of minimiz-
ing 2-energy in equivariance classes of maps from (1, 2)-Poincar´e space spaces to
nonpositively curved metric spaces.
In the more general setting both d-homotopy and path homotopy become prob-
lematic. The lack of triangulations in metric spaces on the one hand, and the fact
that the topology of Newton Sobolev spaces N 1,p(X; Y ) depends on the embedding
of Y into a Banach space (see [13]) on the other, make both notions of homotopy
difficult to work with.
In [35], for the purpose of studying minimizers of p-energy in homotopy classes
of maps from a (1, p)-Poincar´e space to a nonpositively curved metric space a third
notion, called p-quasihomotopy, was introduced. Here we state the definition for
manifolds. It is based on the known fact that Sobolev maps u ∈ W 1,p(M ; N ) have
p-quasicontinuous representatives, i.e. for every ε > 0 there is an open set E ⊂ M
with Capp(E) < ε so that uM\E is continuous. Quasicontinuity may be seen as a
refinement of the almost continuity of measurable maps.
Two quasicontinuous representatives u, v ∈ W 1,p(M ; N ) (1 < p < ∞) are p-
quasihomotopic if there is a map H : M × [0, 1] → N with the following property:
for any ε > 0 there is an open set E ⊂ M with Capp(E) < ε so that HM\E×[0,1]
is a (continuous) homotopy between uM\E and vM\E.
Capacity is a much finer measure of smallness than the Lebesgue measure; a set
E ⊂ M of zero p-capacity has Hausdorff dimension at most n − p, and sets of small
p-capacity have small Hausdorff content,
Capp(E) ≤ c(n, p, q)Hn−q
∞ (E) for any 1 < q < p
(Theorem 5.3 in [27].) Thus, while quasihomotopy allows for discontinuities, it does
so in a sense a minimal amount, preserving some amount of topology. For example,
a set of zero p-capacity, p > 1, does not separate a space, whereas a set of measure
zero may. There is also a p-quasicontinuous counterpart to the fact that if the
preimage of a point of a continuous function (from a connected space) is nonempty
and open, then the function must be constant (see Lemma 5.3 in [35]).
As such, p-quasihomotopy is a natural relaxation of classical homotopy to en-
compass Sobolev maps. Indeed, under the additional assumption that the target
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
3
space has hyperbolic universal cover there always exists minimizers of p-energy in
quasihomotopy classes in the metric setting, see Theorem 1.1. in [34].
When p > n the fact any nonempty set has p-capacity ≥ ε0 for some small number
ε0 implies that p-quasihomotopy coincides with classical homotopy, and thus with
path homotopy.
However when 1 < p < n the notion of p-quasihomotopy turns out to differ from
the other two. Theorem 1.4 in [35] states that when 1 < p < n, if u, v ∈ W 1,p(M ; N )
are p-quasihomotopic then they are path homotopic. The proof in fact yields more:
if 1 < p ≤ n and u, v ∈ W 1,p(M ; N ) are p-quasihomotopic then u and v are d-
homotopic, where d = ⌈p⌉−1 is the largest integer < p. Since ⌊p⌋−1 < ⌈p⌉−1 unless
p is an integer it is expected that path homotopic maps need not be quasihomotopic.
Indeed the constant map and
x 7→
x
x
∈ W 1,p(B2; S1), 1 < p < 2
are path homotopic but not p-quasihomotopic (see Section 4.2 in [35]).
The first main theorem in this paper considers the remaining case p = n.
Theorem 1.1. Let M and N be smooth compact Riemannian manifolds, with
n = dim M . If two maps f, g ∈ W 1,n(M ; N ) are path homotopic then they are
n-quasihomotopic.
The relationships between path-, quasi-, and d-homotopy are summarized in the
table below.
W 1,p(M ; N )
1 < p < n
p-quasihomotopy ⇒ ([p] − 1)-homotopy ⇔ path homotopy
p = n
p > n
path homotopy ⇒ p-quasihomotopy ⇒ (n − 1)-homotopy
p-quasihomotopy ⇔ homotopy
⇔ path homotopy
Surprisingly, the converse of Theorem 1.1 fails. Namely it can happen that two
maps f, g ∈ W 1,n(M ; N ) are n-homotopic but not path homotopic. An example to
this effect is given in Corollary 4.2. It is noteworthy that in the example the target
has the rational homology type of a sphere (in this case it is in fact a sphere) in
light of the discussion in [11] (see in particular Theorems 1.4 and 1.5 there). An
n-manifold M is a rational homology sphere if
H k
dR(M ) =(cid:26) 0
, k 6= 0, n
Z , k = 0, k = n,
where H k
dR(M ) denotes the de Rham cohomology of M .
For generic manifolds M, N , particularly rational homology sphere targets, the
implications between path- and quasihomotopy depend on p.
In contrast, for aspherical target manifolds the situation is simpler. An m-
manifold N is apsherical if the homotopy groups πk(N ) vanish for all k ≥ 2. Using
Whiteheads theorem (Theorem 4.5 in [16]) aspherical manifolds may be character-
ized as those with contractible universal cover. Aspherical manifolds include, as an
important subclass, manifolds of nonpositive sectional curvature.
For general p ∈ (1, ∞) we have the following theorem.
4
ELEFTERIOS SOULTANIS
Theorem 1.2. Suppose M and N are compact smooth Riemannian manifolds, N
aspherical and 1 < p < ∞. If two maps u, v ∈ W 1,p(M ; N ) are p-quasihomotopic
then they are path homotopic.
When p ≥ 2 we can say more.
Theorem 1.3. Let 2 ≤ p < ∞, M, N be smooth compact Riemannian manifolds,
N being aspherical. Then two maps f, g ∈ W 1,p(M ; N ) are path homotopic if and
only they are p-quasihomotopic.
The restriction p ≥ 2 is essential.
Indeed by Theorem 0.2 in [3] the space
W 1,p(M ; N ) is always path connected when 1 < p < 2, while there may exists
distinct p-quasihomotopy classes (see the example above).
Outline. The proof of Theorem 1.1 is based on approximating a given Sobolev map
with suitable mollified maps and showing the convergence is quasiuniform (Theorem
2.13). The second section is devoted to mollification and the use of singular integrals
to accomplish this.
Section 3 deals with the aspherical case. For nonpositively curved targets Propo-
sition 1.2 follows directly from Theorem 1.1 and Proposition 1.5 in [35] but the more
general case of aspherical targets requires somewhat different arguments and the
use of Theorem 2.13. Theorem 1.3 is an immediate consequence of Theorem 3.5,
presented in this Section.
The last Section is devoted to proving that W 1,p(M ; Sk) is p-quasiconnected,
i.e. any two maps in W 1,p(M ; Sk) are p-quasihomotopic, when p ≤ k (Proposition
4.1). Some of the auxiliary results (e.g. Proposition 4.3) may be interesting in
themselves. Proposition 4.1 serves as an example showing that sometimes -- though
not in general -- path homotopy - and p-quasihomotopyclasses coincide.
The paper is closed by remarking that W 1,p(Bk+1; Sk), while path connected
when p < k + 1, is not p-quasiconnected for k < p < k + 1.
2. Critical exponent case
The proof strategy of Theorem 1.1 utilizes Brian White's result.
Theorem 2.1 ([37], Theorem 0 and [2], Theorem 2). Two Lipschitz maps in
W 1,n(M ; N ) are path homotopic if and only if they are homotopic. Moreover for
each u ∈ W 1,n(M ; N ) there is a number ε > 0 so that if ku − vk1,n < ε then u and
v are path homotopic.
Coupled with the fact, due to Schoen-Uhlenbeck [30], that Lip(M ; N ) is dense
in W 1,n(M ; N ) the question, whether path homotopy implies n-quasihomotopy, is
reduced to the following statement. For every u ∈ W 1,n(M ; N ) and ε > 0 there is
a Lipschitz map uε with ku − uεk1,n < ε such that uε is n-quasihomotopic to u.
We will construct such functions by means of mollifying the original function.
2.1. Mollifiers. Suppose ψ : [0, ∞) → [0, 1] is a Lipschitz cut-off function with
spt ψ ⊂ [0, 1). Given r > 0 define ψr : M → R by
ψr(p) =ZM
ψ(cid:18) p − z
r (cid:19) dz.
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
5
Definition 2.2. Given u ∈ Lp(M ; Rν) and r > 0 set
ψr ∗ u(p) =
1
ψr(p)ZM
ψ(cid:18) p − z
r (cid:19) u(z)dz, p ∈ M
Lemma 2.3. For each u ∈ L1
Lipschitz continuous. Moreover
loc(M ; Rν) and r > 0 the map ψr ∗ u : M → Rν is
for almost every x ∈ M , with C, depending only on ψ, M and ν.
ψr ∗ u(x) ≤ CrMu(x)
Proof. For g ∈ L1
loc(M ) and arbitrary x, y ∈ M we have
ψ(cid:18) x − z
r (cid:19) g(z)dz −ZM
(cid:12)(cid:12)(cid:12)(cid:12)ZM
≤ Lip(ψ)ZB(x,r+d(x,y))(cid:12)(cid:12)(cid:12)(cid:12)
ZB(x,r+d(x,y))
≤ Lip(ψ)
d(x, y)
r
r
x − z − y − z
r (cid:19) g(z)dz(cid:12)(cid:12)(cid:12)(cid:12)
ψ(cid:18) y − z
(cid:12)(cid:12)(cid:12)(cid:12) g(z)dz
gdz.
(2.1)
The lipschitz continuity of ψr ∗ u follows from this by expressing the difference
ψr ∗ u(x) − ψr ∗ u(y) ,where d(x, y) < r, as
ψr(y) − ψr(x)
ψr(x)ψr(y) ZM
ψr(y)(cid:20)ZM
ψ(cid:18) x − z
ψ(cid:18) x − z
r (cid:19) u(z)dz
r (cid:19) u(z)dz −ZM
+
1
ψ(cid:18) y − z
r (cid:19) u(z)dz(cid:21)
and applying (2.1) and the doubling property of the measure.
The estimate in the claim follows by a standard decomposition of the integral
(cid:3)
into annular regions, see [19, 17].
Lemma 2.4. (Schoen-Uhlenbeck) Let u ∈ W 1,p(M ; N ). For r > 0 we have
dist(N, ϕr ∗ u(p)) . ZBr (p)
Dundz!1/n
for all p ∈ M . Consequently for each u ∈ W 1,n(M ; N ) there is r0 > 0 so that
dist(N, ϕr ∗ u(p)) < ε0
sup
p∈M
whenever r < r0.
Proof. Let p ∈ M . For a.e. z ∈ Br(p)
dist(N, ϕr ∗ u(p)) ≤ ku(z) − ϕr ∗ u(p)k.
Taking an average integral over Br(p) we obtain
dist(N, ϕr ∗ u(p)) ≤ −ZBr (p)
ku(z) − ϕr ∗ u(p)kdz.
6
ELEFTERIOS SOULTANIS
By the (1, n)-Poincare inequality (which every manifold of dimension n supports)
−ZBr (p)
ku(z) − ϕr ∗ u(p)kdz ≤
ϕ(cid:18) p − w
r (cid:19) ku(z) − u(w)kdzdw
ku(z) − u(w)kdzdw
1
−ZBr (p)
ϕr(p)ZBr (p)
−ZBr(p)
. −ZBr (p)
Dundz!1/n
. r −ZBr (p)
≃ ZBr (p)
Dundz!1/n
.
The implied constants in the estimates depend only on the data of M and on N .
The second assertion follows directly from the absolute continuity of the measure
Dudz.
(cid:3)
2.2. Singular integrals. Let us set some notation. Let ϕ : R → R be a smooth
cutoff function and define the kernel kr : (0, ∞) → R,
kr(t) =
ϕ(t/r)
tn−1 .
We abuse notation by writing
kr(p, q) = kr(p − q), p, q ∈ M
and finally, given g ∈ Lp(M ) (1 < p < ∞), we define the convolution
kr ∗ g(x) =ZM
kr(x, z)g(z)dz, x ∈ M.
By the compactness of M there exists r1 so that
expx : Bn(r1) → B(x, r1)
is a 2-bilipschitz diffeomorphism for all x ∈ M . Thus, when r < r1 we may use a
change of variables given by the exponential map and write the integral above
kr ∗ g(x) =ZBn(r)
kr(ξ)g(expx ξ)J expx ξdξ.
Lemma 2.5. Let 1 < p < ∞. Given g ∈ Lp(M ) the function kr ∗ g has distribu-
tional gradient
∇x(kr ∗ g)v = −P V ZM
k′
r(x − z)h∇xdz, vig(z)dz, v ∈ TxM.
Proof. We refer to [32, 6] for the existence and basic properties of singular integrals
on manifolds (see in particular Chapter IV in [25] and the example in [32, D].)
The distributional derivative is determined by the condition
ZM
h∇(kr ∗ g), V idx = −ZM
(kr ∗ g)divV dx
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
7
for all smooth vector fields V on M . We may write
k′
r(x − z)h∇xdz, Vxig(z)dzdx
k′
r(x − z)h∇xdz, Vxig(z)dzdx
P V ZM
ZM
δ→0ZMZM\Bδ (x)
g(z)ZM\Bδ (z)
δ→0ZM
= lim
= lim
(2.2)
k′
r(x − z)h∇xdz, Vxidxdz.
Note that when x 6= z the vector ∇xdz is the unit vector normal to ∂Bδ(z) at
x. Thus −∇xdz is the unit normal to ∂(M \ Bδ(z)) at x. The divergence theorem
gives
kr(y − z)h∇ydz, Vyidσ(y)
The second term is O(δ) since it may be estimated using again the divergence
theorem:
k′
r(x − z)h∇xdz, Vxidx
kr(x − z)divVxdx +Z∂Bδ (z)
kr(x − z)divVxdx + O(δ).
ZM\Bδ (z)
= −ZM\Bδ (z)
= −ZM\Bδ (z)
kr(y − z)h∇ydz, Vyidσ(y)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z∂Bδ(z)
P V ZM
ZM
g(z)ZM\Bδ (z)
δ→0ZM
δ→0ZMZM\Bδ (x)
= − lim
= − lim
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
kr(δ)ZBδ (z)
divVydy(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
δ→0ZM
kr(x − z)g(z)divVxdzdx = −ZM
kr(x − z)divVxdxdz + lim
k′
r(x − z)h∇xdz, Vxig(z)dzdx
Plugging (2.3) in (2.2) we obtain
. δ1−nδn.
O(δ)dz
(kr ∗ g)divV dx.
(2.3)
(2.4)
i.e.
(2.5)
Thus we are done.
(cid:3)
Lemma 2.6. The operators g 7→ kr ∗ g (r > 0) are uniformly bounded
Lp(M ) → W 1,p(M ),
gpdx,
g ∈ Lp(M ), for all 0 < r < r1.
Proof. For a.e. x ∈ M we have, v ∈ TxM and r > 0
ZM
kr ∗ g(x)pdx +ZM
∇x(kr ∗ g)v =(cid:12)(cid:12)(cid:12)(cid:12)P V ZM
≤vZM
∇x(kr ∗ g)pdx ≤ CZM
r(x − z)h∇xdz, vig(z)dz(cid:12)(cid:12)(cid:12)(cid:12)
rx − zn−1 g(z)dz +(cid:12)(cid:12)(cid:12)(cid:12)P V ZM
ϕ′(x − z/r)
ϕ(x − z/r)
x − zn
k′
h∇xdz, vig(z)dz(cid:12)(cid:12)(cid:12)(cid:12) .
8
ELEFTERIOS SOULTANIS
Using this and the estimate in Lemma 2.3 we obtain the estimate
kr ∗ g(x)p + ∇x(kr ∗ g)p ≤C(rp + 1)Mg(x)p
(2.6)
+ sup
v=1(cid:12)(cid:12)(cid:12)(cid:12)P V ZM
ϕ(x − z/r)
x − zn
p
h∇xdz, vig(z)dz(cid:12)(cid:12)(cid:12)(cid:12)
In light of (2.6) it suffices to demonstrate the (uniform) boundedness of Tj
Lp(M ) → Lp(M ) given by
:
Tjg(x) = P V ZM
ϕ(x − z/r)
x − zn
h∇xdz, ∂jig(z)dz
for each j = 1, . . . , dim M , when r < r1.
Sublemma 2.7. The operator Tj : Lp(M ) → Lp(M ) is bounded with norm inde-
pendent of r ∈ (0, r1).
Proof of sublemma. Since r < r1 and the integrand in Tj vanishes outside B(x, r)
which is bilipschitz diffeomorphic to Bn(r) through the exponential map expx :
Bn(r) → B(x, r), the operator Tj may be written
Tjg(x) = P V ZBn(r)
ϕ(ξ/r)
ξj
ξn+1 g(expx ξ)J expx(ξ)dξ.
By Definition 4 in [32, B] it is sufficient to prove the boundedness, uniformly in r,
for the Euclidean operator
eTj : Lp(Rn) → Lp(Rn) given by the same kernel:
ϕ(ξ/r)
ξj
eTjh(x) = P V ZRn
ξn+1 h(x − ξ)dξ.
By Theorem 5.4.1 in [12] (cf. Chapter 5, Theorem 5.1 in [7]) this is implied by
the following two conditions. Denote
Kr(y) = ϕ(y/r)
yj
yn+1 .
yn+1 .
(2) ∇Kr(y) ≤ B
(1) kcKrk∞ ≤ A, and
A change of variables implies cKr(ξ) = cK1(rξ) so that
kcKrk∞ ≤ kcK1k∞ := A.
ryn + kϕk∞∇(yj/yn+1)(cid:21) ≤
∇Kr(y) ≤ χBn(r)(y)(cid:20) kϕ′k∞
We may estimate
C(n, ϕ)
yn+1 .
Consequently both (1) and (2) are satisfied with constants independent of r. This
completes the proof of the sublemma.
(cid:3)
Having a bound kTjkLp(M)→Lp(M) ≤ C where C is independent of r we obtain
the estimate (2.5) with constant C independent of r. This proves Lemma 2.6. (cid:3)
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
9
2.3. The proof of Theorem 1.1. Using Lemma 2.4 we define a net of approxi-
mating maps with values in N .
Definition 2.8. Let u ∈ W 1,n(M ; N ), and let r0 be the constant in Lemma 2.4.
For 0 < r ≤ r0 set
Additionally, we set
ur(p) = π(ϕr ∗ u(p)), p ∈ M.
u0 = u.
For each r > 0 the maps ur : M → N are clearly Lipschitz. The resulting map
M × [0, r0] ∋ (p, t) 7→ ut(p) is a key component in the proof of Theorem 1.1.
Lemma 2.9. Let r > 0. Then us → ur uniformly as s → r.
Proposition 2.10. The maps ur converge n-quasiuniformly to u, i.e.
for each
ε > 0 there exists an open set U with Capn(U ) < ε such that (ur)M\U → uM\U
uniformly as r → 0.
Proof of Lemma 2.9. We will estimate the difference kur(p) − us(p)k by splitting it
into two parts. Let b be any vector in Rν. We will later choose it appropriately.
kur(p) − us(p)k ≤kϕr ∗ u(p) − ϕs ∗ u(p)k = kϕr ∗ [u − b](p) − ϕs ∗ [u − b](p)k
Let us estimate the two terms (2.7) and (2.8) separately, starting with the latter.
Throughout we assume that r − s < r, which implies that ϕr∨s(p) . ϕs(p) with
constant depending only on M .
1
1
−
ϕr(p)
1
+
≤(cid:12)(cid:12)(cid:12)(cid:12)
r (cid:19) ku(z) − bkdz
ϕs(p)(cid:12)(cid:12)(cid:12)(cid:12)ZBr (p)
ϕ(cid:18) p − z
ϕs(p)ZBs∨r(p)(cid:12)(cid:12)(cid:12)(cid:12)ϕ(cid:18) p − z
r (cid:19) − ϕ(cid:18) p − z
Lip(ϕ)p − z(cid:12)(cid:12)(cid:12)(cid:12)
ϕs(p)ZBs∨r (p)
. r/s − 1 ∨ s/r − 1−ZBr∨s(p)
s (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) ku(z) − bkdz
s(cid:12)(cid:12)(cid:12)(cid:12) ku(z) − bkdz
1
−
1
r
ku − bkdz.
(2.8) ≤
1
(2.7)
(2.8)
A similar computation yields the same bound for (2.7). Thus we arrive at
Now we choose b = uBs∨r(p) and use the (1, n)-Poincare inequality to estimate
kur(p) − us(p)k . r/s − 1 ∨ s/r − 1−ZBr∨s(p)
Dundz!1/n
−ZBr∨s(p)
ku − bkdz . ZBs∨r (p)
ku − bkdz.
. kDukLn(M).
Combining these we arrive at
kur(p) − us(p)k . (r/s − 1 ∨ s/r − 1)kDukLn(M)
for all p. Thus us → ur uniformly as s → r, as long as r 6= 0.
(cid:3)
Proposition 2.10 requires more work. We begin by estimating the difference of u
and ur by an expression which we study in more detail
10
ELEFTERIOS SOULTANIS
Lemma 2.11. Let u ∈ W 1,p(M ; N ). For p-q.e. x ∈ M we have
ku(x) − ur(x)k ≤ZM
ϕ(z − x/r)
z − xn−1 Du(z)dz.
Proof. The proof is similar to [17, p. 28, (4.5)].
(cid:3)
Lemma 2.12. Let 1 < p < ∞ and let (fk) ⊂ N 1,p(M ) be a bounded secuence
with 0 ≤ fk+1 ≤ fk pointwise and kfkkLp → 0 as k → ∞. Then fk → 0 p-
quasiuniformly.
Proof. Since N 1,p(M ) is reflexive we may pass to a subsequence converging weakly
to 0, and by the Mazur lemma a sequence of convex combinations converges to 0 in
norm. Passing to another subsequence if needed, we may assume that the sequence
of convex combinations,
hm = λm
1 fk1 + · · · + λm
NmfkNm , (k1 < · · · < kNm ),
converges to zero p-quasiuniformly. The monotonicity now imples
0 ≤ fkNm ≤ hm
so that a subsequence of (fk) converges p-quasiuniformly to zero. Since the sequence
is pointwise nonincreasing the whole sequence converges to zero p-quasiuniformly.
(cid:3)
These auxiliary results yield Proposition 2.10.
Proof of Proposition 2.10. By Lemma 2.11 we have
for p-quasievery x ∈ M . Choosing ϕ nonincreasing we get that
ku(x) − ur(x)k . kr ∗ Du(x)
pointwise whenever r < s, and further,
kr ∗ Du ≤ ks ∗ Du
kr ∗ Du Ln
−→ 0
as r → 0. By lemma 2.6 the functions kr ∗ Du have uniformly bounded W 1,n-
norms (in r) so by Lemma 2.12 we have that kr ∗ Du → 0 n-quasiuniformly.
Consequently ur → u n-quasiuniformly.
(cid:3)
Theorem 2.13. Let u ∈ W 1,n(M ; N ). The map M × [0, r0] → N given by
in 2.8 defines an n-quasihomotopy u ≃ ur0.
(p, r) 7→ ur(p)
Proof. Denote H(p, r) = ur(p) and suppose ε > 0 is given. Let U be the open set
satisfying the claim of Proposition 2.10. We claim that HM\U×[0,r0] is continuous.
For this it suffices to show that (us)M\U → (ur)M\U uniformly as s → r. This,
however, follows immediately from 2.9 and 2.10.
(cid:3)
We close this Section with the proof of Theorem 1.1.
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
11
Proof of Theorem 1.1. Suppose u, v ∈ W 1,n(M ; N ) are path homotopic. For small
enough ε we have, by Theorems 2.13 and 2.1, that uε is both n-quasihomotopic
and path homotopic to u. The same holds for v and vε.
It follows that uε and vε are path homotopic and since they are Lipschitz, ho-
motopic (Theorem 2.1).
Thus uε and vε are n-quasihomotopic. Consequently u and v are n-quasihomo-
(cid:3)
topic.
3. Aspherical targets
A topological space X is called aspherical if πi(X) = 0 for every i > 1. It is well
known that for smooth Riemannian manifolds the vanishing of higher homotopy
groups is equivalent to having contractible universal cover.
In particular mani-
folds with nonpositive sectional curvature are aspherical. The equivalence stated
in Theorem 1.3 can be seen as a Sobolev version of Whiteheads theorem [16].
Before turning our attention to Theorem 1.3 let us present a proof of Theorem
1.2.
Proof of Theorem 1.2. Suppose N is aspherical and let f, g ∈ W 1,p(M ; N ) be p-
quasihomotopic. We devide the proof into three cases:
(1) p < n: By Theorem 1.4 in [35] f and g are path homotopic.
(2) p > n: In this case path homotopy and p-quasihomotopy coincide, see the discus-
sion in the introduction.
(3) p = n: This is the only case that requires some work. By Theorem 2.13 f, g are
n-quasihomotopic to Lipschitz maps f0, g0 so we may assume that f and
g are themselves Lipschitz. Since N is aspherical it is path representable
[34, Proposition 3.4] and thus by [34, Theorem 1.2] (f, g) ∈ N 1,n(M ; N ) ∩
of N (see [34, Subsection 2.4]). Since gh = g(f,g) ≤ LIP(f ) + LIP(g) almost
everywhere (Lemma 4.3 in [34]) it follows that h is in fact Lipschitz. Thus
the continuous map (f, g) : M → N × N admits a (continuous) lift
Lip(M ; N ) has a lift h ∈ N 1,n(M ; bNdiag) where bNdiag is the diagonal cover
h : M → bNdiag. By Proposition 3.2 in [34] f and g are homotopic, hence
path homotopic in W 1,n(M ; N ).
(cid:3)
When p ≥ 2, a Sobolev map f ∈ W 1,p(M ; N ) induces a homorphism
u∗ : π(M, x0) → π(N, f (x0)) [31] (see also [38, 28]). For almost every x0 ∈ M an
induced homomorphism satisfies, for all [γ] ∈ π(M, x0):
• u∗[γ] = [u ◦ γ] if γ is such that u ◦ γ is continuous
• u∗[γ] = [u ◦ γ′] for some γ′ ∼ γ.
It is known that no such induced homomorphism need exist for a Sobolev map
f ∈ W 1,p(M ; N ) when 1 < p < 2.
To connect induced homomorphisms to p-quasihomotopies we recall the notion
of a fundamental system of loops from [34].
Given a p-quasicontinuous representative u ∈ W 1,p(M ; N ), an upper gradient g ∈
Lp(M ) and an exceptional path family Γ0 of curves in M , such that g is an upper
12
ELEFTERIOS SOULTANIS
gradient of u along any curve γ /∈ Γ0, and a basepoint x0 ∈ M with Mgg(x0) < ∞,
the collection of loops
Fx0(g, Γ0) = {αβ−1 : Γx0x \ Γ0, Mgp(x) < ∞}
is called the fundamental system of loops.
Recall the definition of sptp Γ0 of a negligible path family:
where the intersection is taken over all admissible metrics ρ ∈ Lp(M ) for which
sptp Γ0 =\{Mρp = ∞}
Zγ
ρ = ∞ for all γ ∈ Γ0.
Zγ
g = ∞, γ ∈ Γ0,
Lemma 3.1. There is a constant C with the following property. If Γ0 is a path
family and g ∈ Lp(M ) a nonnegative Borel function with
then for any x, y /∈ {Mgp = ∞} there exists a curve γ /∈ Γ0 joining x and y with
ℓ(γ) ≤ Cd(x, y).
Proof. By Lemma 4.5 in [34] and Theorem 2 (4) in [24] we have
d(x, y)1−p ≤ C Modp(Γxy \ Γg; µxy),
where
µxy(A) =ZA(cid:20)
d(x, z)
µ(B(x, d(x, z)))
+
d(y, z)
µ(B(y, d(y, z)))(cid:21) dµ(z), A ⊂ X.
In particular Γxy \ Γg is nonempty. Note that Γ0 ⊂ Γg.
If ℓ(γ) ≥ Dd(x, y) for all γ ∈ Γxy \ Γg ⊂ Γxy \ Γ0 then ρ = 1/(Dd(x, y)) is
admissible for Γxy \ Γg and thus
Modp(Γxy \ Γg; µxy) ≤ CD−pd(x, y)1−p.
Combining the two inequalitites yields the required bound on D.
(cid:3)
Lemma 3.2. Let p ≥ 2, and u ∈ W 1,p(M ; N ) be a quasicontinuous representative.
Given an upper gradient g of u, a path family Γ0 of zero p-modulus, and a point
x0 /∈ sptp Γ0 with Mgp(x0) < ∞, we have
u∗π(M, x0) = u♯Fx0(g, Γ0).
Proof. Let u ∈ W 1,p(M ; N ) and let g, Γ0 be as in the claim. Set
E = {x0 : Mgp(x0) = ∞} ∪ spt Γ0
and choose and arbitrary point x0 /∈ E. For any γ ∈ Fx0(g, Γ0) clearly [u ◦ γ] ∈
u∗π(M, x0). Thus we only need to prove the other inclusion.
To this end, fix a loop γ based on x0. Take a tubular neighbourhood T of
γ so that any loop in T is homotopic with γ. Take a finite chain of open balls
x0 ∈ B0, B1, . . . , Bk of radii r > 0 such that 2CBj ⊂ T , and Bj ∩ Bj+1 6= ∅, where
C is the constant in Lemma 3.1. Since E = 0 there exists, for each j, points
yj ∈ (Bj ∩ Bj+1) \ E (with the convention that y0 = x0 and yk ∈ (B0 ∩ Bk) \ E.)
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
13
By Lemma 3.1 there exists a curve γj /∈ Γ0 joining yj and yj+1 with ℓ(γj) ≤
Cd(yj, yj+1) (here yk+1 = x0). Hence γj ⊂ T . The loop γ′ = γ0 · · · γk+1 belongs
to Fx0(g, Γ0) and is contained in T , and therefore homotopic with γ.
It follows that [u ◦ γ′] = u∗[γ′] = u∗[γ] and since γ was arbitrary we obtain
(cid:3)
u∗π(M, x0) ≤ u♯Fx0(g, Γ0). The proof is complete.
Lemma 3.3. Let p ≥ 2. Two maps, u, v ∈ W 1,p(M ; N ), are p-quasihomotopic if
and only if u♯π(M ) and v♯π(M ) are conjugated subgroups of π(N ).
Proof. By [34, Theorem 1.2 and 1,3] the maps u, v are p-quasihomotopic if and only
if
(3.1)
(u, v)♯Fx0(g, Γ0) ≤ p∗π(bNdiag, [α])
cover of N which consists of homotopy classes of all paths in N (see [34] for the
precise construction). A modification of the proof of [34, Lemma 2.18] yields
for some [α] ∈ p−1(u(x0), v(x0)), and some x0 ∈ M . Here (p, bNdiag) is the diagonal
p∗π(bNdiag, [α]) = {([γ], [α−1γα]) : [γ] ∈ π(N, u(x0))} ≤ π(N, u(x0)) × π(N, v(x0)).
On the other hand by Lemma 3.2
(u, v)♯Fx0(g, Γ0) = (u, v)∗π(M, x0) = {(u∗[γ], v∗[γ]) : [γ] ∈ π(M, x0)}.
By these two identities (3.1) is equivalent to
u∗[γ] = [α]−1v∗[γ][α]
for all [γ] ∈ π(M, x0). Hence we are done.
(cid:3)
Lemma 3.4. If u, v ∈ W 1,p(M ; N ) are path homotopic (p ≥ 2) then for almost
every x0 ∈ M u∗π(M, x0) and v∗π(M, x0) are conjugated.
Proof. Suppose first that p < n. Then by [15, Theorem 1.1] u and v are [p − 1]-
homotopic and, since p ≥ 2, in particular 1-homotopic. Fix a 1-skeleton K of M
containing a point x0 ∈ {M(Dup + Dvp) < ∞}, and such that uK and vK are
(continuous and) homotopic by a homotopy h : K × [0, 1] → N .
To prove that the image subgroups of the homomorphisms are conjugated, take
a loop γ with basepoint x0. By [16, Section 4.1, Theorem 4.8] γ is homotopic to a
loop γ′ which lies in K. Thus the image loops u ◦ γ′ and v ◦ γ′ are conjugated by
H(s, t) = h(γ(s), t), t, s ∈ [0, 1]2.
Denoting by α the path t 7→ h(x0, t) we thus have
[u ◦ γ′] = [α−1(v ◦ γ′)α].
Consequently
u∗([γ]) = u∗([γ′]) = (v∗([γ′]))[α] = (v∗([γ]))[α], [γ] ∈ π(M, x0).
This proves the claim in the case p < n.
In case p ≥ n it follows from Theorem 1.1 and Theorem ?? that u and v are
(cid:3)
p-quasihomotopic. The claim now follows from Lemma 3.3 above.
Combining Proposition 1.2 and Lemmata 3.3 and 3.4 we obtain the following the-
orem, which directly implies Theorem 1.3.
Theorem 3.5. Let p ≥ 2, and N aspherical. Then two maps u, v ∈ W 1,p(M ; N )
are path homotopic if and only if the subgroups u∗π(M ) and v∗π(M ) are conjugated.
14
ELEFTERIOS SOULTANIS
Proof. Suppose u, v are path homotopic. Then Lemma 3.4 implies the claim. If,
conversely, u∗π(M ) and v∗π(M ) are conjugated, Lemma 3.3 implies that u and v
are p-quasihomotopic. By Proposition 1.2 u and v are path homotopic.
(cid:3)
4. Quasiconnectedness of W 1,p(M ; Sk)
In this section the following result is proven.
Proposition 4.1. Suppose M is a smooth compact riemannian manifold, possibly
with boundary, and 1 < p ≤ k. Then u ∈ W 1,p(M ; Sk) is p-quasiconnected, i.e.
every map is p-quasihomotopic to a constant.
We single out the following corollary.
Corollary 4.2. Suppose 2 ≤ k and 1 < p ≤ k. Then any two maps in W 1,p(Sk; Sk)
are p-quasihomotopic.
The proof of Theorem 4.1 is based on the example given in [2] after Theorem 3. We
begin by observing that that in a suitable range of p's points have small preimages
under Sobolev maps.
Lemma 4.3. Let f ∈ W 1,p(M ; N ) be a p-quasicontinuous representative, 1 < p ≤
dim N . Then for almost every y ∈ N we have
Capp(f −1(y)) = 0.
Proof. For y ∈ N , consider the function uk ∈ W 1,p(M ) given by
uk(x) = wk ◦ f,
where wk : N → R is defined by
Then ukf −1(y) ≡ 1 p-quasieverywhere and therefore
Capp(f −1(y)) ≤ lim inf
k→∞
kukkp
1,p.
We have the pointwise estimates
0 ≤ uk(x) ≤ χB(y,1/k)(f (x)),
∇uk(x) ≤ ∇wk(f (x))∇f (x) ≤ (log k)−1 χA(y,1/k2,1/k)(f (x))
f (x) − y
∇f (x)
almost everywhere. Thus
Capp(f −1(y)) ≤ lim inf
χB(y,1/k) ◦ f dx
k→∞ (cid:20)ZM
+(log k)−pZM
χA(y,1/k2,1/k)(f (x))
f (x) − yp
∇f pdx(cid:21) .
wk(z) =
1
(log k)−1 log(cid:16) 1/k
0
, z ∈ B(y, 1/k2)
z−y(cid:17) , z ∈ A(y, 1/k2, 1/k)
, z /∈ B(y, 1/k)
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
15
Integrating over y ∈ N and using Fatou and Fubini we obtain
(4.1)
≤ lim inf
Capp(f −1(y))dy
ZN
k→∞ ZMZN(cid:20)χB(y,1/k)(f (x)) + (log k)−p∇f p(x)
ZMZN
f (x) − yp
χA(y,1/k2,1/k)(f (x))
(cid:21) dydx
χB(f (x),1/k)(y)dy(cid:19) dx ≤ C/kdim N
Since
inequality (4.1) becomes
(4.2)
The righthand integral in turn may be written as
Capp(f −1(y))dy
χB(y,1/k)(f (x))dydx =ZM(cid:18)ZN
ZN
k→∞ ZMZN
(log k)−pZM
≤ lim inf
(log k)−p∇f p(x)
∇f p(x)(cid:18)ZN
dy . CZRdim N
tdim N −1−pdt ≤Z 1/k
Z 1/k
k→∞ ZM
1/k2
1/k2
ZN
For sufficiently large k ≥ 1 one may estimate
ZN
χA(f (x),1/k2,1/k)(y)
f (x) − yp
Since p ≤ dim N we obtain
Plugging all these inequalities into (4.2) we obtain
χA(y,1/k2,1/k)(f (x))
f (x) − yp
dydx.
χA(f (x),1/k2,1/k)(y)
f (x) − yp
dy(cid:19) dx.
yp ≃Z 1/k
1/k2
dy
t−1dt = log k.
χA(0,1/k2,1/k)(y)
tdim N −1−pdt.
Capp(f −1(y))dy ≤ C lim inf
(log k)1−p∇f p(x)dx = 0,
thus completing the proof.
(cid:3)
Corollary 4.4. Let 2 ≤ k and 1 < p ≤ k. For a p-quasicontinuous representative
f ∈ W 1,p(M ; Sk) the following holds for almost every y ∈ Sk.
lim
r→0
Capp(f −1B(y, r)) = 0.
Proof. Let ε > 0 be arbitrary and let U ⊂ M be open with Capp(U ) < ε and f M\U
continuous. We may estimate
Capp(f −1B(y, r)) ≤ Capp((f M\U )−1(B(y, r))) + Capp(U ).
The sets (f M\U )−1(B(y, r)) are compact and decrease to (f M\U )−1(y) as r > 0
decreases. By the monotonicity of capacity for compact sets therefore
Capp((f M\U )−1(B(y, r))) = Capp((f M\U )−1(y)).
lim sup
r→0
The latter quantity is zero for almost every y ∈ Sk by Lemma 4.3 above. Thus we
obtain
Capp(f −1B(y, r)) ≤ 0 + Capp(U ) < ε.
16
ELEFTERIOS SOULTANIS
Since ε > 0 was arbitrary the claim follows.
(cid:3)
Proof of Proposition 4.1. Suppose f ∈ W 1,p(M ; Sk). Choose y0 ∈ Sk so that the
claim of Corollary 4.4 holds for y = y0. Define h : Sk × [0, ∞] → Sk by
h(x, t) =(cid:26) x−ty0
x−ty0 ,
−y0,
0 ≤ t < ∞
t = ∞
Note that hSk\{x0}×[0,∞] is continuous. We claim that
H(x, t) = h(f (x), t), (x, t) ∈ M × [0, ∞]
is a p-quasihomotopy f ≃ −y0.
Given ε > 0 let U be an open set with Capp(U ) < ε/2 and f M\U continuous.
Further let r > 0 be small enough so that Capp(f −1B(y0, r)) < ε/2. Set E = U ∪
[(f −1B(x0, r)) \ U ]. Then E is open, Capp(E) < ε and HM\E×[0,∞] is continuous,
H(x, 0) =
f (x)
f (x)
= f (x), H(x, ∞) = −x0, x ∈ M \ E.
(cid:3)
Remark 4.5. A similar procedure yields a continuous path in W 1,p(Sn; Sn) between
f and a constant map when p < n (see [2]), but not when p = n. Indeed, in the
latter case it is not possible to connect every map to a constant path by a continuous
path ([2, Lemma 1"]) and so we see that the converse of Theorem 1.1 is not true.
In closing we remark that W 1,p(Bk+1; Sk), k < p < k + 1 provides another example
where path and p-quasihomotopy differ.
Consider the map g : (0, 1] × Sk → Bk+1 given by
g(t, y) = ty.
This is a p-quasihomotopy equivalence (p < k + 1) since the map h(x) = (x, x/x)
is p-quasicontinuous and g ◦ h = idBk+1, h◦ g = id(0,1]×Sk p-quasieverywhere. Thus,
postcomposition with g defines a continuous map
G : W 1,p(Bk+1; Sk) → W 1,p((0, 1] × Sk; Sk), Gf = f ◦ g,
which preserves p-quasihomotopy classes and is bijective (the map f 7→ f ◦ h is an
inverse to G).
It is known ([3], Proposition 0.2) that W 1,p((0, 1] × Sk; Sk) is path connected
when p < k + 1. However, when k < p < k + 1, the Sobolev space W 1,p((0, 1] ×
Sk; Sk) and consequently W 1,p(Bk+1; Sk) is not p-quasiconnected. (This easily seen
by noting that the map f (t, y) = y, (t, y) ∈ (0, 1] × Sk, is not p-quasihomotopic to
a constant map.)
Acknowledgements. I would like to thank Pekka Pankka for reading the manuscript
and making many valuable comments.
I also thank Pawel Goldstein for useful
discussions.
PATH AND QUASIHOMOTOPY FOR SOBOLEV MAPS BETWEEN MANIFOLDS
17
References
[1] Anders Bjorn and Jana Bjorn. Nonlinear potential theory on metric spaces, volume 17 of
EMS Tracts in Mathematics. European Mathematical Society (EMS), Zurich, 2011.
[2] Haım Brezis. The fascinating homotopy structure of Sobolev spaces. Atti Accad. Naz. Lincei
Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl., 14(3):207 -- 217 (2004), 2003. Renato
Caccioppoli and modern analysis.
[3] Haim Brezis and Yanyan Li. Topology and Sobolev spaces. J. Funct. Anal., 183(2):321 -- 369,
2001.
[4] Martin R. Bridson and Andr´e Haefliger. Metric spaces of non-positive curvature, volume 319
of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathemat-
ical Sciences]. Springer-Verlag, Berlin, 1999.
[5] Francis E. Burstall. Harmonic maps of finite energy from noncompact manifolds. J. London
Math. Soc. (2), 30(2):361 -- 370, 1984.
[6] A. P. Calder´on and A. Zygmund. On singular integrals. Amer. J. Math., 78:289 -- 309, 1956.
[7] Javier Duoandikoetxea. Fourier analysis, volume 29 of Graduate Studies in Mathematics.
American Mathematical Society, Providence, RI, 2001. Translated and revised from the 1995
Spanish original by David Cruz-Uribe.
[8] J. Eells and L. Lemaire. Another report on harmonic maps. Bull. London Math. Soc.,
20(5):385 -- 524, 1988.
[9] James Eells and Luc Lemaire. A report on harmonic maps. Bull. London Math. Soc., 10(1):1 --
68, 1978.
[10] James Eells, Jr. and Joseph H. Sampson. Harmonic mappings of Riemannian manifolds.
Amer. J. Math., 86:109 -- 160, 1964.
[11] Pawe lGoldstein and Piotr Haj l asz. Sobolev mappings, degree, homotopy classes and rational
homology spheres. J. Geom. Anal., 22(2):320 -- 338, 2012.
[12] Loukas Grafakos. Classical Fourier analysis, volume 249 of Graduate Texts in Mathematics.
Springer, New York, third edition, 2014.
[13] Piotr Haj l asz. Sobolev mappings: Lipschitz density is not a bi-Lipschitz invariant of the
target. Geom. Funct. Anal., 17(2):435 -- 467, 2007.
[14] Piotr Haj lasz. Sobolev spaces on an arbitrary metric space. Potential Anal., 5(4):403 -- 415,
1996.
[15] Fengbo Hang and Fanghua Lin. Topology of Sobolev mappings. II. Acta Math., 191(1):55 -- 107,
2003.
[16] Allen Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002.
[17] Juha Heinonen. Lectures on analysis on metric spaces. Universitext. Springer-Verlag, New
York, 2001.
[18] Juha Heinonen and Pekka Koskela. Quasiconformal maps in metric spaces with controlled
geometry. Acta Math., 181(1):1 -- 61, 1998.
[19] Juha Heinonen, Pekka Koskela, Nageswari Shanmugalingam, and Jeremy Tyson. Sobolev
spaces on metric measure spaces: an approach based on upper gradients. New Mathematical
Monographs. Cambridge University Press, United Kingdom, first edition, 2015.
[20] Jurgen Jost. Equilibrium maps between metric spaces. Calc. Var. Partial Differential Equa-
tions, 2(2):173 -- 204, 1994.
[21] Jurgen Jost. Convex functionals and generalized harmonic maps into spaces of nonpositive
curvature. Comment. Math. Helv., 70(4):659 -- 673, 1995.
[22] Jurgen Jost. Generalized harmonic maps between metric spaces. In Geometric analysis and
the calculus of variations, pages 143 -- 174. Int. Press, Cambridge, MA, 1996.
[23] Jurgen Jost. Generalized Dirichlet forms and harmonic maps. Calc. Var. Partial Differential
Equations, 5(1):1 -- 19, 1997.
[24] Stephen Keith. Modulus and the Poincar´e inequality on metric measure spaces. Math. Z.,
245(2):255 -- 292, 2003.
[25] J. J. Kohn and D. C. Spencer. Complex Neumann problems. Ann. of Math. (2), 66:89 -- 140,
1957.
[26] Nicholas J. Korevaar and Richard M. Schoen. Sobolev spaces and harmonic maps for metric
space targets. Comm. Anal. Geom., 1(3-4):561 -- 659, 1993.
[27] Jan Mal´y, David Swanson, and William P. Ziemer. The co-area formula for Sobolev mappings.
Trans. Amer. Math. Soc., 355(2):477 -- 492, 2003.
18
ELEFTERIOS SOULTANIS
[28] Nobumitsu Nakauchi. Homomorphism between homotopy groups induced by elements of the
Sobolev space L1,p(M, N ). Manuscripta Math., 78(1):1 -- 7, 1993.
[29] Stefano Pigola and Giona Veronelli. On the homotopy class of maps with finite p-energy into
non-positively curved manifolds. Geom. Dedicata, 143:109 -- 116, 2009.
[30] Richard Schoen and Karen Uhlenbeck. Boundary regularity and the Dirichlet problem for
harmonic maps. J. Differential Geom., 18(2):253 -- 268, 1983.
[31] Richard Schoen and Shing Tung Yau. Compact group actions and the topology of manifolds
with nonpositive curvature. Topology, 18(4):361 -- 380, 1979.
[32] R. T. Seeley. Singular integrals on compact manifolds. Amer. J. Math., 81:658 -- 690, 1959.
[33] Nageswari Shanmugalingam. Newtonian spaces: an extension of Sobolev spaces to metric
measure spaces. Rev. Mat. Iberoamericana, 16(2):243 -- 279, 2000.
[34] Elefterios Soultanis. Existence of p-energy minimizers in homotopy classes and lifts of New-
tonian maps. To appear in J. Anal. Math. arXiv:1506.07767 [math.DG].
[35] Elefterios Soultanis. Homotopy classes of Newtonian spaces. To appear in Rev. Mat.
Iberoamericana. arXiv:1309.6472 [math.MG].
[36] Giona Veronelli. A general comparison theorem for p-harmonic maps in homotopy class. J.
Math. Anal. Appl., 391(2):335 -- 349, 2012.
[37] Brian White. Infima of energy functionals in homotopy classes of mappings. J. Differential
Geom., 23(2):127 -- 142, 1986.
[38] Brian White. Homotopy classes in Sobolev spaces and the existence of energy minimizing
maps. Acta Math., 160(1-2):1 -- 17, 1988.
ul. ´Sniadeckich 8,
00-656 Warszawa
E-mail address: [email protected]
|
1711.08787 | 1 | 1711 | 2017-11-23T17:40:30 | Operator least squares problems and Moore-Penrose inverses in Krein spaces | [
"math.FA"
] | A Krein space H and bounded linear operators B, C on H are given. Then, some min and max problems about the operators (BX - C)^{#}(BX -C), where X runs over the space of all bounded linear operators on H, are discussed. In each case, a complete answer to the problem, including solvability conditions and characterization of the solutions, is presented. Also, an adequate decomposition of B is considered and the min-max problem is addressed. As a by-product the Moore-Penrose inverse of B is characterized as the only solution of a variational problem. Other generalized inverses are described in a similar fashion as well. | math.FA | math |
Integr. equ. oper. theory 99 (9999), 1 -- 23
DOI 10.1007/s00020-003-0000
c(cid:13) 2018 Birkhauser Verlag Basel/Switzerland
Integral Equations
and Operator Theory
Operator least squares problems and Moore-
Penrose inverses in Krein spaces
Maximiliano Contino, Alejandra Maestripieri and Stefania
Marcantognini
Abstract. A Krein space H and bounded linear operators B, C on H
are given. Then, some min and max problems about the operators
(BX − C)#(BX − C), where X runs over the space of all bounded
linear operators on H, are discussed. In each case, a complete answer
to the problem, including solvability conditions and characterization of
the solutions, is presented. Also, an adequate decomposition of B is
considered and the min-max problem is addressed. As a by-product the
Moore-Penrose inverse of B is characterized as the only solution of a
variational problem. Other generalized inverses are described in a simi-
lar fashion as well.
Mathematics Subject Classification (2010). 47A58, 47B50, 41A65.
Keywords. Operator approximation, Krein spaces, Moore-Penrose in-
verse.
1. Introduction
Several least squares problems, especially in connection with the search of
alternative H ∞ algorithms in system and control theory, have been placed in
the Krein space framework. Roughly speaking, the consideration of suitable
space models for the set of observations data have brought into play indefi-
nite metric spaces and, on the basis of the given information, least squares
problems on those spaces. Some references from the nineties are [16, 17, 18].
Commonly those least squares estimations are formulated and solved in terms
of vectors in Krein spaces and more often than not the vectors are set in the
(n + m)-dimensional Minkowski space.
We discuss, instead, least-squares problems for Krein space operators.
The approach we opt for is taken from [12, 13]. Several arguments we present
are adapted from [6, 7].
2
Contino, Maestripieri and Marcantognini
IEOT
Our aim is dual: to study abstract least-squares-type problems for Krein
space operators and to do so from a geometrical viewpoint. Pseudo-regularity
plays a key role, either as a technical tool to generalize some finite-dimensional
indefinite metric space arguments to the general Krein space framework or as
the natural assumption to grant solvability. In that regard we should mention
that every closed subspace of a Pontryagin space is pseudo-regular and, on
the other hand, that every pseudo-regular subspace of a Krein space is the
range of a normal projection. We may say so that pseudo-regular subspaces
lie somewhere in between closed subspaces and regular subspaces. For more
details on the subject see [14] and [21].
Recall that, in the Hilbert space case, the Moore-Penrose inverse B† of
a given bounded linear operator B satisfies the equations BB†B = B and
B†BB† = B†, and that both BB† and B†B are selfadjoint. If B is a closed
range operator then B† is known to be the unique minimal norm solution of
a linear equation. In the Krein space framework the analysis of the existence
of a generalized inverse B† such that BB† and B†B are selfadjoint -- with
respect to the indefinite inner product -- was carried on by X. Mary [22]. He
found out that a bounded linear operator B admits a unique bounded Moore-
Penrose inverse if and only if both the range and null space of B are regular
subspaces. His treatment is exhaustive but it fails to include the variational
characterization of the Moore-Penrose inverse.
Under the necessary and sufficient conditions given by Mary, we do
identify B† as the unique solution of a variational problem. Furthermore,
when the projections associated to the generalized inverse are only required
to be normal -- with respect to the indefinite inner products -- we prove that
pseudo-regularity of the range R(B) and null space N (B) are necessary and
sufficient conditions for a closed range B to admit such a sort of generalized
inverse. Matter-of-factly, if that is the case, there exists a whole family of
generalized inverses which is in one to one correspondence with the set of
pairs (Q, P ) with Q a normal projection onto R(B) and P a normal projection
onto N (B). Besides, we characterize the generalized inverses in a variational
way just as we do it for the Moore-Penrose inverse.
The paper comprises five sections, six if this introductory section is
included. Section 2 is a brief expository introduction to Krein spaces and op-
erators on them and serves to fix the notation. It presents also some results
that are needed in the following sections. In Section 3 the notion of indefi-
nite inverse of an operator is defined, generalizing the concept of W -inverse
introduced by Mitra and Rao for matrices in [23], and extended later for op-
erators acting on Hilbert spaces in [8]. In Section 4 we deal with the problem
of determining whether the
min
X∈L(H)
(BX − C)#(BX − C)1
exists for B a given closed range bounded operator and C either the identity
operator or any given bounded operator. The solutions to these problems are
1L(H) stands for the space of all the bounded linear operators from H to H.
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
3
characterized as the indefinite inverses of B. The results about this indefinite
minimization problem and their counterparts for the symmetric maximization
problem are applied in Section 5 where B is factorized as B = B+ + B− and
the min-max problem
min
max
X∈L(H)
Y ∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C)
is addressed. Section 6 contains the main results about the Moore-Penrose
inverse and the generalized inverses of a given Krein space operator.
2. Preliminaries
In the following all Hilbert spaces are complex and separable. If H and K
are Hilbert spaces, L(H, K) stands for the space of all the bounded linear
operators from H to K and CR(H, K) for the subset of L(H, K) comprising
all the operators with closed ranges. When H = K we write, for short, L(H)
and CR(H). The range and null space of any given A ∈ L(H, K) are denoted
by R(A) and N (A), respectively.
The direct sum of two closed subspaces M and N of H is represented
by M +N . If H is decomposed as H = M +N , the projection onto M with
null space N is denoted PM//N and abbreviated PM when N = M⊥. In
general, Q is used to indicate the subset of all the oblique projections in
L(H), namely, Q := {Q ∈ L(H) : Q2 = Q}.
The following is a well-known result about range inclusion and factor-
izations of operators. We will refer to it along the paper.
Theorem 2.1 (Douglas' Theorem [9]). Let Y, Z ∈ L(H). Then R(Z) ⊆ R(Y )
if and only if there exists D ∈ L(H) such that Z = Y D.
Krein Spaces
Although familiarity with operator theory on Krein spaces is presumed, we
hereafter include some basic notions. Standard references on Krein spaces
and operators on them are [1, 4, 5]. We also refer to [10, 11] as authoritative
accounts of the subject.
Consider a linear space H with an indefinite metric, i.e., a sesquilinear
]. A vector x ∈ H is said to be positive if [ x, x ] > 0.
Hermitian form [
A subspace S of H is positive if every x ∈ S, x 6= 0, is a positive vector.
Negative, nonnegative, nonpositive and neutral vectors and subspaces are
defined likewise.
,
We say that two closed subspaces M and N are orthogonal, and we
write M [⊥] N , if [ m, n ] = 0 for every m ∈ M and n ∈ N . We denote the
orthogonal direct sum of two closed subspaces M and N by M [∔] N .
Given any subspace S of H, the orthogonal companion of S in H, say
S[⊥], is defined as
S[⊥] := {x ∈ H : [ x, s ] = 0 for every s ∈ S}.
The isotropic part So := S ∩ S[⊥] can be a non-trivial subspace.
4
Contino, Maestripieri and Marcantognini
IEOT
An indefinite metric space (H, [
,
]) is a Krein space if H admits a
decomposition into an orthogonal direct sum in the form
H = H+ [∔] H−
(2.1)
where (H+, [
with these properties is called a fundamental decomposition of H.
]) are Hilbert spaces. Any decomposition
]) and (H−, −[ ,
,
Given a Krein space (H, [ ,
]) with a fundamental descomposition H =
H+ [∔] H−, the (orthogonal) direct sum of the Hilbert spaces (H+, [ ,
]) and
(H−, −[ ,
]) is a Hilbert space. It is denoted by (H, h , i). Notice that the
inner product h , i and the corresponding quadratic norm k k depend on the
fundamental decomposition. A subspace S of H is called uniformly positive
if, for some Hilbert space inner product h , i on H, there exists ε > 0 such
that [ s, s ] ≥ εksk2 for every s ∈ S. Uniformly negative subspaces are defined
in a similar fashion.
Every fundamental decomposition of H has an associated signature op-
erator, to wit, J := P+ − P− where P± := PH±//H∓ . The indefinite metric
and the inner product corresponding to a fundamental decomposition of H
with signature operator J are related to each other by
h x, y i = [ J x, y ] for every x, y ∈ H.
If H is a Krein space, L(H) stands for the vector space of all the linear
operators on H which are bounded in an associated Hilbert space (H, h , i).
Since the norms generated by different fundamental decompositions of a
Krein space H are equivalent, see, for instance, [4, Theorem 7.19], it comes
that L(H) does not depend on the chosen underlying Hilbert space.
Given T ∈ L(H), T # is the unique operator satisfying
[ T x, y ] = [ x, T #y ] for every x, y ∈ H.
An operator T ∈ L(H) is said to be selfadjoint if T = T #.
A positive operator T ∈ L(H) satisfies [ T x, x ] ≥ 0 for every x ∈ H.
The notation S ≤ T signifies that T − S is positive.
A (closed) subspace S of a Krein space H is a regular subspace if
H = S [∔] S[⊥]. Equivalently, S is a regular subspace if it is the range
of a selfadjoint projection, i.e., there exists Q ∈ Q such that Q = Q# and
R(Q) = S (see [4, Proposition 1.4.19]).
In [2, Theorem 2.3], T. Ando proved that any selfadjoint projection on
a Krein space can be decomposed as the sum of two selfadjoint projections
with uniformly definite ranges. See also [15, 20].
Theorem 2.2. Let (H, [
projection. Then Q can be written as
,
]) be a Krein space and let Q be a selfadjoint
Q = Q+ + Q−,
where Q+ and Q− are selfadjoint projections such that R(Q+) is uniformly
positive, R(Q−) is uniformly negative and Q+Q− = Q−Q+ = 0. Moreover,
each fundamental decomposition of H provides a (unique) decomposition of
Q in such a manner.
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
5
The next lemma shows that every closed subspace of a Krein space can
be decomposed as the orthogonal direct sum of a closed positive subspace and
a closed nonpositive subspace (see [4, Theorem 6.4], [5, Chapter V, Theorem
3.1]).
Lemma 2.3. Let (H, [ ,
]) be a Krein space with fundamental descomposition
H = H+ [∔] H− and corresponding Hilbert space inner product h , i. Let
S be a closed subspace of H. Then S can be represented uniquely as the or-
thogonal direct sum of a closed positive subspace S+ and a closed nonpositive
subspace S−, i.e.,
Furthermore, hS+, S−i = {0}.
S = S+ [∔] S−.
In [12, 13] least squares problems in the indefinite metric setting were
studied. From those references we recall the definition of indefinite least
squares solution.
Definition. Let (H, [ ,
u ∈ H is an indefinite least squares solution (ILSS) of Bz = x if
]) be a Krein space and let B ∈ CR(H). We say that
[ x − Bu, x − Bu ] ≤ [ x − Bz, x − Bz ]
for every x, z ∈ H.
We conclude this section by stating necessary and sufficient conditions
for the existence of an ILSS of the equation Bz = x. We refer to [5, Chapter
I, Theorem 8.4] where a proof of the result is given.
Lemma 2.4. Let (H, [ ,
]) be a Krein space and let B ∈ CR(H). Then u ∈ H
is an ILSS of the equation Bz = x if and only if R(B) is nonnegative and
x − Bu ∈ R(B)[⊥].
3. Indefinite inverses in Krein spaces
In [23] S. K. Mitra and C. R. Rao introduced the notion of the W -inverse
of a matrix for a given positive weight W. Later, in [8] and [6], the concept
was extended to Hilbert space operators, specifically, given a Hilbert space
(H, h , i), a positive operator W ∈ L(H) and an operator B ∈ CR(H), a
W -inverse of B is defined to be an operator X0 ∈ L(H) such that, for each
x ∈ H, X0x is a weighted least squares solution of Bz = x, i.e., so that
h W (BX0x − x), BX0x − x i ≤ h W (Bz − x), Bz − x i for every z ∈ H.
In [8]
it was proved that X0 is a W-inverse of B if and only if
B∗W (BX0 − I) = 0 or, equivalently, X0 satisfies the identities W (BX0)2 =
W BX0 = (BX0)∗W. The first equality means that BX0 is a W -projection
while the second says that BX0 is W -selfadjoint, see [8].
We extend the definition to Krein spaces in the following way. From
now on, (H, [ ,
]) stands for a Krein space.
6
Contino, Maestripieri and Marcantognini
IEOT
Definition. Let B ∈ CR(H). An operator X0 ∈ L(H) is an indefinite inverse
of B if X0 is a solution of
B#(BX − I) = 0.
Proposition 3.1. Let B ∈ CR(H). Then B admits an indefinite inverse if and
only if R(B) is regular.
Proof. Suppose that X0 is an indefinite inverse of B so that B#(BX0 − I) =
0. Then, for every x ∈ H, BX0x − x ∈ N (B#) = R(B)[⊥] and, therefore,
x ∈ R(B) + R(B)[⊥]. Whence H = R(B) + R(B)[⊥]. As, also, {0} = R(B) ∩
R(B)[⊥], it comes that H = R(B) [∔] R(B)[⊥] or, accordingly, that R(B) is
regular.
Conversely, if R(B) is regular then H = R(B)[∔]R(B)[⊥]. So, by ap-
plying B#, it results that R(B#) = R(B#B). From here and by Douglas'
Theorem (Theorem 2.1), it follows that the equation B#(BX −I) = 0 admits
(cid:3)
a solution or, equivalently, that B has an indefinite inverse.
It results from the proof of Proposition 3.1 that, for any B ∈ CR(H),
R(B) is regular if, and only if, R(B#) = R(B#B). In this case, N (B) =
N (B#B).
The next proposition characterizes the indefinite inverses of B ∈ L(H)
when R(B) is regular.
Proposition 3.2. Let B ∈ L(H). Assume that R(B) is regular. Then the
following conditions are equivalent:
i) X0 is an indefinite inverse of B,
ii) X0 is a solution of the equation BX = Q, where Q is the selfadjoint
projection onto R(B),
iii) X0 is an inner inverse of B, i.e., BX0B = B, and (BX0)# = BX0.
Moreover, if R(B) is also uniformly positive, conditions i), ii), iii) are also
equivalent to:
iv) For every x ∈ H, X0x is an ILSS of Bz = x.
A similar statement holds if R(B) is uniformly negative.
i) ⇒ ii) : Notice, first, that B# = B#Q, since B = QB and Q =
Proof.
Q#. Whence B#(BX0 − I) = 0 implies B#(BX0 − Q) = 0. Therefore,
R(BX0 − Q) ⊆ N (B#) ∩ R(Q) = N (Q) ∩ R(Q) = {0}. Thus BX0 = Q.
ii) ⇒ iii) : If BX0 = Q then BX0B = QB = B and (BX0)# = Q# =
Q = BX0.
iii) ⇒ i) : Suppose that BX0B = B and (BX0)# = BX0. Then
i) ⇔ iv) : X0 is an indefinite inverse of B if, and only if, B#(BX0 −I) =
0 if, and only if, for every x ∈ H, BX0x − x ∈ R(B)[⊥]. Since R(B) is
(cid:3)
nonnegative, Lemma 2.4 gives the equivalence.
B#(BX0 − I) = B#(X #
0 B# − I) = B#X #
0 B# − B# = 0.
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
7
We point out that, when R(B) is closed and uniformly definite, the
original definition of W -inverse for the indefinite metric is retrieved directly
from item iv) of the last proposition.
The more general concept of the indefinite inverse of B in the range of
C is given next.
Definition. Let B ∈ CR(H) and C ∈ L(H). An operator X0 ∈ L(H) is an
indefinite inverse of B in R(C) if X0 is a solution of
B#(BX − C) = 0.
Proposition 3.3. Let B ∈ CR(H) and C ∈ L(H). B has an indefinite inverse
in R(C) if and only if R(C) ⊆ R(B) + R(B)[⊥].
Proof.
Suppose that X0 is an indefinite inverse of B in R(C). Then
B#(BX0 − C) = 0. So, if x ∈ H then BX0x − Cx ∈ N (B#) = R(B)[⊥]
and, therefore, Cx ∈ R(B) + R(B)[⊥]. Thus, R(C) ⊆ R(B) + R(B)[⊥] and
the result follows.
Conversely, if R(C) ⊆ R(B) + R(B)[⊥] then
R(B#C) ⊆ R(B#B).
Here, as before in the proof of Proposition 3.1, Douglas' Theorem is applied to
grant that the equation B#(BX − C) = 0 admits a solution or, equivalently,
(cid:3)
that B has an indefinite inverse in R(C).
Corollary 3.4. Let B ∈ CR(H) and C ∈ L(H). If R(B) is regular then X0
is an indefinite inverse of B in R(C) if and only if X0 is a solution of the
equation BX = QC, with Q the selfadjoint projection onto R(B).
Proof. Suppose that R(B) is regular. Then R(B#) = R(B#B) or, equiva-
lently, N (B) = N (B#B).
If B#(BX0 − C) = 0 then B#(BX0 − QC) = 0, for B = QB, Q = Q#
and, consequently, B#C = B#QC. Hence R(BX0 − QC) ⊆ N (B#)∩ R(Q) =
N (Q) ∩ R(Q) = {0}. So BX0 = QC.
(cid:3)
From the last corollary we have that, when R(B) is regular, the set of
indefinite inverses of B in R(C) is the affine manifold
X0 + L(H, N (B)),
with X0 any solution of the equation BX = QC.
Proposition 3.5. Let B ∈ CR(H) and C ∈ L(H) satisfy that R(B) is non-
negative and R(C) ⊆ R(B) + R(B)[⊥]. Then X0 is an indefinite inverse of B
in R(C) if and only if, for every x ∈ H, X0x is an ILSS of Bz = Cx, i.e.,
[ Cx − BX0x, Cx − BX0x ] ≤ [ Cx − Bz, Cx − Bz ]
for every z ∈ H.
Proof. X0 is an indefinite inverse of B in R(C) if, and only if, B#(BX0−C) =
0 if, and only if, BX0x − Cx ∈ R(B)[⊥] for every x ∈ H. Since R(B) is
supposed to be nonnegative, Lemma 2.4 can be applied to get the equivalence.
(cid:3)
8
Contino, Maestripieri and Marcantognini
IEOT
4. Indefinite least squares problems
To state the next problems let us recall that the order is the one induced by
the positive operators in (H, [ ,
]): given two operators S, T ∈ L(H), S ≤ T
whenever T − S is positive.
Consider the following problem: given two operators B ∈ CR(H) and
C ∈ L(H), determine the existence of
min
X∈L(H)
(BX − C)#(BX − C).
(4.1)
Definition. Let B ∈ CR(H) and C ∈ L(H). We say that X0 ∈ L(H) is an
indefinite minimum solution (ImS) of BX − C = 0 if
(BX0 − C)#(BX0 − C) = min
(BX − C)#(BX − C).
X∈L(H)
(4.2)
In a similar fashion, the analogous maximization problem can be con-
sidered. From now on, we only address the problem related to the existence
of (4.1). The arguments we present in dealing with problem (4.1) can be
adapted to the maximum problem. In particular, each of the "min" results
we include in this section can be easily modified to get its "max" counterpart.
Theorem 4.1. Let B ∈ CR(H) and C ∈ L(H). Then there exists an ImS of
BX − C = 0 if and only if R(C) ⊆ R(B) + R(B)[⊥] and R(B) is nonnegative.
Proof. Suppose that X0 ∈ L(H) is an ImS of BX − C = 0, so that
[ (BX0 − C)x, (BX0 − C)x ] ≤ [ (BX − C)x, (BX − C)x ]
for every x ∈ H and every X ∈ L(H). Let z ∈ H be arbitrary. Then, for
every x ∈ H \ {0}, there exists X ∈ L(H) such that z = Xx. Therefore
[ (BX0 − C)x, (BX0 − C)x ] ≤ [ Bz − Cx, Bz − Cx ] for every x, z ∈ H.
So, for every x ∈ H, X0x is an ILSS of Bz = Cx. By Lemma 2.4, we get
that R(C) ⊆ R(B) + R(B)[⊥] and R(B) is nonnegative. Furthemore, by
Proposition 3.5, we have that X0 is an indefinite inverse of B in R(C).
Conversely, if R(C) ⊆ R(B) + R(B)[⊥] and R(B) is nonnegative then,
by Proposition 3.3, B admits an indefinite inverse in R(C). Now, if X0 is an
indefinite inverse of B in R(C) then, by Proposition 3.5,
[ (BX0 − C)x, (BX0 − C)x ] ≤ [ Bz − Cx, Bz − Cx ] for every x, z ∈ H.
Given x ∈ H and X ∈ L(H), set z = Xx, so that
[ (BX0 − C)x, (BX0 − C)x ] ≤ [ (BX − C)x, (BX − C)x ]
for every x ∈ H and every X ∈ L(H). Now it becomes clear that X0 is an
ImS of the equation BX − C = 0, as required to complete the proof.
(cid:3)
In the proof of Theorem 4.1 the X0's in (4.2), that is, the solutions of
the problem related to (4.1), were characterized. Indeed:
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
9
Corollary 4.2. Let B ∈ CR(H) and C ∈ L(H) satisfy that R(B) is nonnega-
tive and R(C) ⊆ R(B) + R(B)[⊥]. Then X0 is an ImS of BX − C = 0 if and
only if X0 is an indefinite inverse of B in R(C), i.e., X0 is solution of the
normal equation
B#(BX − C) = 0.
From the last result we have that the set of indefinite inverses of ImS
of BX − C = 0 is the affine manifold
X0 + L(H, N (B#B)),
with X0 any indefinite inverse of B in R(C).
The next two corollaries follow from Theorem 4.1 as well.
Corollary 4.3. Let B ∈ CR(H). Then there exists an ImS of BX − C = 0 for
every C ∈ L(H) if and only if R(B) is uniformly positive.
In this case,
min
X∈L(H)
(BX − C)#(BX − C) = C#(I − Q)C
where Q is the selfadjoint projection onto R(B).
Proof. Assume that, for every C ∈ L(H), there exists an ImS of BX −C = 0.
In particular, there exists an ImS of BX − I = 0. Then, by Theorem 4.1,
R(B) is regular and nonnegative. Whence R(B) is uniformly positive.
Conversely, if R(B) is uniformly positive then, for every C ∈ L(H), we
have that R(C) ⊆ H = R(B) + R(B)[⊥]. Hence, by Theorem 4.1, for every
C ∈ L(H), there exists an ImS X0 ∈ L(H) of BX − C = 0 or, equivalently,
X0 is a solution of the normal equation B#(BX − C) = 0 (see Corollary
4.2). In this case, since R(B) is regular, Corollary 3.4 gives that BX0 = QC.
Therefore,
min
X∈L(H)
(BX − C)#(BX − C) = (BX0 − C)#(BX0 − C) = C#(I − Q)C.
Corollary 4.4. Let B ∈ CR(H). Then there exists an ImS of BX − I = 0 if
and only if R(B) is uniformly positive.
In this case, the ImS of BX − I = 0 are the indefinite inverses of B and
(cid:3)
min
X∈L(H)
(BX − I)#(BX − I) = I − Q
where Q is the selfadjoint projection onto R(B).
Remark. By mimicking the arguments in the proof of Theorem 4.1, a similar
result can be proved for operators acting between different Krein spaces. More
precisely, let (H, [
]F ) be Krein spaces. Let
B ∈ CR(H, K) and C ∈ L(F , K). Then there exists X0 ∈ L(F , H) such that
]K) and (F , [
,
]H), (K, [
,
,
min
X∈L(F ,H)
(BX − C)#(BX − C) = (BX0 − C)#(BX0 − C)
if and only if R(C) ⊆ R(B) + R(B)[⊥]K and R(B) is nonnegative.
10
Contino, Maestripieri and Marcantognini
IEOT
4.1. Indefinite least squares problems: the pseudo-regular case
A (closed) subspace S of a Krein space H is called a pseudo-regular subspace
if the algebraic sum S + S[⊥] is closed. Observe that, this is equivalent to the
equality (S0)[⊥] = S + S[⊥], see [14]. Also, S is a pseudo-regular subspace if
and only if S is the range of a normal projection, i.e., there exists Q ∈ Q
such that QQ# = Q#Q and R(Q) = S (see [21, Theorem 4.3]). Unlike normal
projections in Hilbert spaces, a normal projection in a Krein space need not
be selfadjoint. In what follows QS stands for the set of normal projections
onto the pseudo-regular subspace S, i.e.,
QS := {Q ∈ L(H) : Q2 = Q, QQ# = Q#Q, R(Q) = S}.
The set QS has infinite elements, unless S is regular. See [21] for further
details on the subject.
Let B ∈ CR(H), the next results relate the pseudo-regularity of R(B)
to the indefinite inverse of B in R(C) and the ImS of BX − C = 0.
The next lemma, stated in [12, Remark 2.1], will be useful when dealing
with pseudo-regular ranges.
Lemma 4.5. Let S be a pseudo-regular subspace of H. If Q ∈ QS then
Q#(I − Q)y = 0 if and only if y ∈ S + S[⊥].
Proof. Let Q ∈ QS . If y ∈ S + S[⊥], by [12, Remark 2.1], we have that
Q#(I − Q)y = 0.
Conversely, if Q#(I − Q)y = 0, since Q#(I − Q) ∈ Q, y ∈ N (Q#(I −
(cid:3)
Q)) = S + S[⊥].
Proposition 4.6. Let B ∈ CR(H) and C ∈ L(H). If R(B) is pseudo-regular
and R(C) ⊆ R(B) + R(B)[⊥], then X0 is an indefinite inverse of B in R(C)
if and only if R(BX0 − QC) ⊆ R(B)o, for any Q ∈ QR(B).
Proof. Suppose that R(B) is pseudo-regular and pick any Q ∈ QR(B). By
Lemma 4.5, (I − Q)y ∈ N (Q#) = N (B#) for every y ∈ R(B) + R(B)[⊥].
Since R(C) ⊆ R(B) + R(B)[⊥], we have that B#(I − Q)C = 0. So X0
is a solution of B#(BX − C) = 0 if and only if B#(BX0 − QC) = 0 or
R(BX0 − QC) ⊆ R(B)o.
(cid:3)
Corollary 4.7. Let B ∈ CR(H). Then there exists an ImS of BX − C = 0
for every C ∈ L(H) such that R(C) ⊆ (R(B)o)[⊥] if and only if R(B) is a
pseudo-regular, nonnegative subspace of H.
In this case,
min
X∈L(H)
(BX − C)#(BX − C) = C#(I − Q)C,
for any Q ∈ QR(B).
Proof. Let C ∈ L(H). Note that there exists an ImS of the equation BX −
C = 0 if and only if R(C) ⊆ R(B) + R(B)[⊥] and R(B) is nonnegative (see
Theorem 4.1).
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
11
Suppose that R(B) is pseudo-regular and nonnegative. Then
(R(B)o)[⊥] = R(B) + R(B)[⊥],
and, therefore, there exists an ImS of the equation BX − C = 0 for every
C ∈ L(H) such that R(C) ⊆ (R(B)o)[⊥].
Conversely, suppose that there exists an ImS of the equation BX−C = 0
for every C ∈ L(H) such that R(C) ⊆ (R(B)o)[⊥]. Then pick a C such
that R(C) = (R(B)o)[⊥] = R(B) + R(B)[⊥]. It must happen that R(C) ⊆
R(B) + R(B)[⊥] and R(B) is nonnegative. That is, R(B) is to be pseudo-
regular and nonnegative. In this case, let X0 be an indefinite inverse of B in
R(C). By Corollary 4.2, X0 is an ImS of BX − C = 0. By Proposition 4.6,
R(BX0 − QC) ⊆ R(B)o. Then Lemma 4.5 with S = R(B) and the fact that
R(BX0 − QC) ⊆ R(B)o yield the result.
(cid:3)
5. Min-Max least squares problems
In this section a min-max problem is studied for operators with not necessarly
definite range. In order to pose the problem, choose a fundamental decom-
position H = H+ [∔] H− and fix the corresponding Hilbert space (H, h , i),
so that, for all x, y ∈ H, h x, y i = [ J x, y ] with J the signature operator
associated with the decomposition.
Let B ∈ CR(H). By Lemma 2.3, R(B) can be represented uniquely as
R(B) = S+ [∔] S−
(5.1)
with S+ a positive closed subspace of H, S− a nonpositive closed subspace
of H and hS+, S−i = {0}.
Consider P+ = PS+ and P− = PS− , the orthogonal projections from
the Hilbert space (H, h , i) onto S+ and S−, respectively. It readily follows
that P+ + P− = PR(B). Therefore, if B+ := P+B and B− := P−B then
B = B+ + B−, R(B+) = S+ and R(B−) = S−.
(5.2)
Since N (P #
± ) = S[⊥]
± , it holds that B#
+ B− = B#P #
+ B− = 0 and
B#
− B+ = 0.
Observe that if R(B) is regular then P+ and P− are the projections
given by Theorem 2.2.
Definition. Let C ∈ L(H). Let B in CR(H) be represented as in (5.2). An
operator Z0 ∈ L(H) is said to be an indefinite min-max solution (ImMS) of
BX − C = 0 (corresponding to the decomposition given by J) if
(BZ0 − C)#(BZ0 − C) =
= max
Y ∈L(H)
min
X∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C).
(5.3)
The following result shows that an ImMS of BX − C = 0 is indepen-
dent of the selected fundamental decomposition of H. Along the following
paragraphs, C denotes the cone of neutral vectors in H.
12
Contino, Maestripieri and Marcantognini
IEOT
Theorem 5.1. Let C ∈ L(H) and B ∈ CR(H). An operator Z0 ∈ L(H) is an
ImMS of BX −C = 0, for some (and, hence, any) fundamental decomposition
of H, if and only if
where Z1 is an indefinite inverse of B in R(C) and R(BZ2) ⊆ C.
Z0 = Z1 + Z2
Proof. Fix a fundamental decomposition H = H+ [∔] H−, and consider
B = B+ +B− as in (5.2). Suppose that Z0 ∈ L(H) is an ImMS of BX −C = 0
for that decomposition. Then Z0 verifies (5.3). So, for every fixed Y ∈ L(H),
there exists min
(B+X + B−Y − C)#(B+X + B−Y − C).
X∈L(H)
From Corollary 4.2 and by using that B#
+ B− = 0, we get that the
minimum is attained at X0(= X0(Y )) if and only if
0 = B#
+ (B+X0 − (C − B−Y )) = B#
+ (B+X0 − C).
The above says that X0 is an indefinite inverse of B+ in R(C) and, in par-
ticular, that X0 does not depend on Y . Hence, for every Y ∈ L(H),
(B+X0 + B−Y − C)#(B+X0 + B−Y − C) =
= min
X∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C)
and, since Z0 satisfies (5.3),
(BZ0 − C)#(BZ0 − C) = max
Y ∈L(H)
(B+X0 + B−Y − C)#(B+X0 + B−Y − C).
By the suitable version of Corollary 4.2 and using that B#
− B+ = 0, we
get that the maximum is attained at Y0 ∈ L(H) if and only if
0 = B#
− (B−Y0 − (C − B+X0)) = B#
− (B−Y0 − C).
Consequently,
(BZ0 − C)#(BZ0 − C) = (B+X0 + B−Y0 − C)#(B+X0 + B−Y0 − C). (5.4)
Let Z1 ∈ L(H) satisfy BZ1 = B+X0 + B−Y0 as in Douglas' Theorem,
so that, according with (5.4),
(BZ0 − C)#(BZ0 − C) = (BZ1 − C)#(BZ1 − C).
(5.5)
A straightforward computation gives that B#(BZ1 − C) = 0 and, in con-
sequence, that Z1 is an indefinite inverse of B in R(C). Now, as Z1 is an
indefinite inverse of B in R(C), it comes that
(BZ0 − C)#(BZ0 − C) = (BZ1 − C + BZ0 − BZ1)#(BZ1 − C + BZ0 − BZ1) =
= (BZ1 − C)#(BZ1 − C) + (B(Z0 − Z1))#B(Z0 − Z1).
Set Z2 := Z0 − Z1. By combining the above equation with (5.5) we conclude
that it must hold that (BZ2)#BZ2 = 0 or, equivalently, that R(BZ2) ⊆ C.
Clearly, Z0 = Z1 + Z2, with Z1 and Z2 as required.
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
13
Conversely, let Z1 ∈ L(H) be an indefinite inverse of B in R(C), and
R(BZ2) ⊆ C. If Z0 = Z1 + Z2 then
(BZ0 − C)#(BZ0 − C) = (BZ1 − C)#(BZ1 − C).
Write B = B+ + B− as in (5.2). Since B#(BZ1 − C) = 0, we have
− ).
+ B− = 0,
that R(BZ1 − C) ⊆ N (B#) = (R(B+) [∔] R(B−))[⊥] = N (B#
Therefore, B#
it readily follows that, for every X, Y ∈ L(H),
+ (B+Z1 − (C − B−Y )) = B#
+ (BZ1−C) = B#
− (BZ1−C) = 0. Then, as B#
− B+ = B#
+ (BZ1 − C) = 0
+ ) ∩ N (B#
B#
and
B#
− (B−Z1 − (C − B+X)) = B#
− (BZ1 − C) = 0.
So, by Corollary 4.2, we obtain that
(BZ0 − C)#(BZ0 − C) = (BZ1 − C)#(BZ1 − C) =
= (B+Z1 + B−Z1 − C)#(B+Z1 + B−Z1 − C)
= max
Y ∈L(H)
(B−Y − (C − B+Z1))#(B−Y − (C − B+Z1))
= max
Y ∈L(H)
min
X∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C).
Therefore, Z0 is an ImMS of BX − C = 0.
(cid:3)
The next remark follows from the proof of the last theorem.
Remark. Let C ∈ L(H) and B ∈ CR(H) such that B is represented as in
(5.2). Then
max
Y ∈L(H)
min
X∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C) =
= min
X∈L(H)
max
Y ∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C).
Indeed, if Z0 is an ImMS of BX − C = 0 then, as the last theorem
asserts, Z0 = Z1 + Z2 where Z1 is an indefinite inverse of B in R(C) and
R(BZ2) ⊆ C. In the proof of the theorem, on the other hand, we found out
that, for every X, Y ∈ L(H),
B#
+ (B+Z1 − (C − B−Y )) = B#
+ (BZ1 − C) = 0
and
B#
− (B−Z1 − (C − B+X)) = B#
− (BZ1 − C) = 0.
A direct application of both the Corollary 4.2 and its modified version gives
(BZ0 − C)#(BZ0 − C) = (BZ1 − C)#(BZ1 − C) =
= (B+Z1 + B−Z1 − C)#(B+Z1 + B−Z1 − C)
= min
X∈L(H)
(B+X − (C − B−Z1))#(B+X − (C − B−Z1))
= min
X∈L(H)
max
Y ∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C).
14
Contino, Maestripieri and Marcantognini
IEOT
Corollary 5.2. Let B ∈ CR(H) and C ∈ L(H). Then, there exists an ImMS
of BX − C = 0 if and only if R(C) ⊆ R(B) + R(B)[⊥].
Proof. Suppose that Z0 is an ImMS of BX − C = 0. Then, by Theorem 5.1,
Z0 = Z1 + Z2 where B#(BZ1 − C) = 0 and R(BZ2) ⊆ C. Therefore
R(C) ⊆ R(B) + R(B)[⊥].
Conversely, if R(C) ⊆ R(B) + R(B)[⊥] then, by Proposition 3.3, there
exists a solution of the normal equation B#(BX − C) = 0, say Z1 ∈ L(H). It
suffices to put Z2 = 0 and to apply Theorem 5.1 to get that Z1 is an ImMS
of BX − C = 0.
(cid:3)
Corollary 5.3. Let B ∈ CR(H). There exists an ImMS of BX − C = 0 for
every C ∈ L(H) if and only if R(B) is regular.
In this case, if B is represented with respect to a fixed (but arbitrary)
fundamental decomposition of H as in (5.2), then
max
Y ∈L(H)
min
X∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C) =
(B−Y − I)#(B−Y − I)]
(B+X − I)#(B+X − I)] C =
= C# [ max
Y ∈L(H)
[ min
X∈L(H)
= C#(I − Q)C,
where Q is the selfadjoint projection onto R(B).
If R(B) is regular then, for every C ∈ L(H), R(C) ⊆ R(B)[∔]R(B)[⊥]
Proof.
and, by Corollary 5.2, there exists an ImMS of BX − C = 0.
Conversely, assume that, for every C ∈ L(H), there exists an ImMS of
BX − C = 0. Set C = I and apply the corollary once again to get H =
R(I) ⊆ R(B) + R(B)[⊥] and R(B) regular.
In the case that R(B) is regular, given a fundamental decomposition
of H, Ando's Theorem (Theorem 2.2) provides unique selfadjoint projections
Q+, Q− ∈ L(H) such that Q = Q+ + Q− with R(Q+) uniformly positive and
R(Q−) uniformly negative. Then, as we already mentioned it, the subspaces
S± in the decomposition (5.1) of R(B) and the operators B± in (5.2) are
given by S± = R(Q±) and B± = Q±B.
Let Z0 ∈ L(H) be an ImMS of the equation BX − C = 0, so that, due
to Theorem 5.1, Z0 = Z1 + Z2 where Z1 is an indefinite inverse in R(C) and
R(BZ2) ⊆ C. On one hand, it holds that
max
Y ∈L(H)
min
X∈L(H)
(B+X + B−Y − C)#(B+X + B−Y − C)
= (BZ0 − C)#(BZ0 − C) = (BZ1 − C)#(BZ1 − C) = C#(I − Q)C,
for R(BZ2) ⊆ C and, by Corollary 3.4, BZ1 = QC. On the other hand,
Corollary 4.4 yields
C#(I − Q)C = C#(I − Q−)(I − Q+)C =
= C# [ max
Y ∈L(H)
(B−Y − I)#(B−Y − I)]
[ min
X∈L(H)
(B+X − I)#(B+X − I)] C.
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
15
By merging the above equations, the required identities are obtained and the
(cid:3)
proof is complete.
6. The Moore-Penrose inverse in Krein spaces
In [22, Theorem 2.16] X. Mary proved that, given B ∈ L(H), the range and
nullspace of B are regular subspaces of H if and only if B admits a (unique)
"Moore-Penrose inverse", in the sense that, there exists an operator B† ∈
L(H) such that BB†B = B, B†BB† = B†, (BB†)# = BB†, (B†B)# =
B†B.
Moreover, it was proven in [22, Corollary 2.13] that if Q is the selfadjoint
projection onto R(B) and P is the selfadjoint projection onto N (B)[⊥], then
BB† = Q and B†B = P.
In this section, we are interested in characterizing the Moore-Penrose
inverse in a variational way. To this end, we consider B ∈ CR(H) and C ∈
L(H) and analyze the following problem: find conditions for the existence of
an ImS X0 of BX − C = 0 such that X #
0 X0 ≤ Y #Y, for every ImS Y of
BX − C = 0.
By Theorem 4.1, the equation BX − C = 0 admits an ImS if and only
if R(C) ⊆ R(B) + R(B)[⊥] and R(B) is nonnegative. In this case, if MC is
the set of ImS of BX − C = 0, then the above problem becomes: determine
whether there exists
min
X∈MC
X #X
(6.1)
when MC 6= ∅.
We only address this problem. Alternatively, symmetric problems de-
pending on the signature of the involved subspaces can be adapted to solve
them.
Theorem 6.1. Let B ∈ CR(H) and C ∈ L(H). Then there exists a solution
of problem (6.1) if and only if R(B) and N (B#B) are nonnegative and
R(C) ⊆ B(N (B#B)[⊥]) + R(B)[⊥].
Proof. Suppose that there exists a solution of problem (6.1). By Corollary
4.2, the set MC can be described as
MC = {X = X0 + Y : Y ∈ L(H), R(Y ) ⊆ N (B#B)},
where X0 is any solution of the equation B#(BX − C) = 0.
Therefore, problem (6.1) can be rephrased as: analyze the existence of
min
Z∈L(H)
(RZ + X0)#(RZ + X0),
(6.2)
where R ∈ L(H) is such that R(R) = N (B#B).
16
Contino, Maestripieri and Marcantognini
IEOT
By Theorem 4.1, problem (6.2) has a solution if and only if N (B#B) is
nonnegative and R(X0) ⊆ N (B#B) + N (B#B)[⊥]. Applying B#B to both
sides of the inclusion, we have that
R(B#C) = R(B#BX0) ⊆ B#B(N (B#B)[⊥]).
Finally, applying (B#)−1 to both sides of the inclusion, we get
R(C) ⊆ B(N (B#B)[⊥]) + N (B#).
Conversely, suppose that R(B) and N (B#B) are nonnegative and
R(C) ⊆ B(N (B#B)[⊥]) + R(B)[⊥]. Clearly, R(C) ⊆ R(B) + R(B)[⊥], so,
by Theorem 4.1, there exists an ImS X0 of BX − C = 0, or equivalently,
B#(BX0 − C) = 0. On the other hand, since N (B#B) is nonnegative and
R(C) ⊆ B(N (B#B)[⊥]) + R(B)[⊥], it holds that
R(B#BX0) = R(B#C) ⊆ B#B(N (B#B)[⊥]).
Applying (B#B)−1 to both sides of the inclusion, it comes that
R(X0) ⊆ N (B#B) + N (B#B)[⊥].
Therefore, by Theorem 4.1, there exists a solution of (6.2) and hence, there
(cid:3)
exists a solution of problem (6.1).
It follows from the last theorem that, if B ∈ CR(H) and C ∈ L(H) are
such that R(B) and N (B#B) are nonnegative, then there exists a solution
of problem (6.1) if and only if MC 6= ∅, and for every X0 ∈ MC, R(X0) ⊆
N (B#B) + N (B#B)[⊥]. Moreover:
Lemma 6.2. Let B ∈ CR(H) and C ∈ L(H) such that R(B) and N (B#B) are
nonnegative and R(C) ⊆ B(N (B#B)[⊥]) + R(B)[⊥]. Then X1 is a solution
of (6.1) if and only if B#(BX1 − C) = 0 and R(X1) ⊆ N (B#B)[⊥].
Proof. Recall that X1 is a solution of problem (6.1) if and only if X1 =
RZ1 + X0, with R ∈ L(H) such that R(R) = N (B#B), X0 a solution of
B#(BX − C) = 0 and Z1 a solution of (6.2).
Since N (B#B) is nonnegative, by Theorem 4.1 and Corollary 4.3, Z1 ∈
L(H) is a solution of (6.2) if and only if Z1 is such that
that is,
or, equivalently,
R#(RZ1 + X0) = 0
R#X1 = 0
R(X1) ⊆ N (R#) = R(R)[⊥] = N (B#B)[⊥].
(cid:3)
As a corollary of Theorem 6.1, we have the following result.
Proposition 6.3. Let B ∈ CR(H). Then the following assertions are equiva-
lent:
i) There exists a solution of problem (6.1) for C = I,
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
17
ii) R(B) and N (B) are uniformly positive,
iii) there exists the Moore-Penrose inverse of B, B†, and R(B) and N (B)
are nonnegative.
i) ⇔ ii) : If there exists a solution of problem (6.1) for C = I, by
Proof.
Theorem 6.1, R(B) and N (B#B) are nonnegative, and
H ⊆ B(N (B#B)[⊥]) + R(B)[⊥].
(6.3)
Then, clearly, R(B) is regular and nonnegative, i.e., R(B) is uniformly posi-
tive. Since R(B) is regular then R(B#) = R(B#B) or, equivalently, N (B) =
N (B#B). Applying B# to both sides of (6.3), we have that R(B#) ⊆
B#B(N (B#B)[⊥]). Then,
R(B#B) = R(B#) ⊆ B#B(N (B#B)[⊥]) ⊆ R(B#B).
Therefore
and so,
R(B#B) = B#B(N (B#B)[⊥]),
H = N (B#B) + N (B#B)[⊥] = N (B) + N (B)[⊥].
Thus, N (B) is regular and nonnegative and therefore uniformly positive.
Conversely, if R(B) and N (B) are uniformly positive, then
H = R(B) + R(B)[⊥] = N (B#B) + N (B#B)[⊥],
where we used the fact that N (B) = N (B#B) since R(B) is regular. Ap-
plying B#B to both sides of the second equality, we get that R(B#B) =
B#B(N (B#B)[⊥]). Then, applying (B#)−1 to the left and right sides of the
last equality, the inclusion R(B)+R(B)[⊥] ⊆ B(N (B#B)[⊥])+R(B)[⊥] holds.
Therefore we get
H = B(N (B#B)[⊥]) + R(B)[⊥],
and, by Theorem 6.1, there exists a solution of problem (6.1) for C = I.
ii) ⇔ iii) : See [22, Theorem 2.6].
(cid:3)
Theorem 6.4. Let B ∈ CR(H) and suppose that N (B) and R(B) are uni-
formly positive. Then, the Moore-Penrose inverse of B, B†, is the unique ImS
X0 of BX − I = 0 such that X #
0 X0 ≤ Y #Y, for every ImS Y of BX − I = 0.
Proof. Since N (B) and R(B) are regular, the Moore-Penrose inverse of B,
B†, exists. Consider Q the selfadjoint projection onto R(B) and P the self-
adjoint projection onto N (B)[⊥], then
B#(BB† − I) = B#(Q − I) = 0.
On the other hand,
R(B†) = R(B†BB†) ⊆ R(B†B) = R(P ) = N (B)[⊥].
Hence, by Lemma 6.2, B† is a solution of problem (6.1), with C = I.
18
Contino, Maestripieri and Marcantognini
IEOT
Let X1 ∈ L(H) be any other solution of problem (6.1), with C = I. By
Lemma 6.2, X1 is an ImS of BX − I = 0 and R(X1) ⊆ N (B)[⊥]. Then, by
Corollary 3.4, BX1 = Q = BB†. Therefore,
X1 = P X1 = B†BX1 = B†BB† = B†.
(cid:3)
The next remark follows from the proofs of Proposition 6.3 and Theorem
6.4.
Remark. Let B ∈ CR(H) and C ∈ L(H), and suppose that N (B) and R(B)
are uniformly positive. Then problem (6.1) admits a unique solution, namely,
B†C.
6.1. The Moore-Penrose inverse: the pseudo-regular case
In [12, Proposition 5.1], a family of generalized inverses of a closed range
operator with pseudo-regular range and nullspace was given. In this case, the
associated projections turn out to be normal. In this section, we prove the
equivalence between the existence of this family of generalized inverses and
the pseudo-regularity of the range and nullspace of an operator B ∈ CR(H).
We also give a more general expression for these generalized inverses and we
characterize them in a variational way as we did in the last section with the
Moore-Penrose inverse.
Given B ∈ CR(H), recall that B is a {1, 2}-inverse of B if B is a solution
of the system
BXB = B, XBX = X.
If (H, h ·, · i) is a Hilbert space, every B ∈ CR(H) admits a {1, 2}-inverse, see
[3, Theorem 3.1]. Then, using any of the underlying Hilbert structures, the
same is true in the Krein space H. Observe that, if B is a {1, 2}-inverse of B,
then B B is a projection onto R(B) and BB is a projection with N ( BB) =
N (B).
Proposition 6.5. Let B ∈ CR(H). Then, there exists a solution of the system
BXB = B, XBX = X,
(BX)#(BX) = (BX)(BX)#,
(XB)#(XB) = (XB)(XB)#,
(6.4)
if and only if R(B) and N (B) are pseudo-regular subspaces of H.
In this case, D ∈ L(H) is a solution of (6.4) if and only if there exist
Q ∈ QR(B) and P ∈ QN (B) such that
D = (I − P ) BQ,
(6.5)
where B is any {1, 2}-inverse of B.
Proof. Suppose that R(B) and N (B) are pseudo-regular subspaces. Let Q ∈
QR(B) and P ∈ QN (B). Let B be any {1, 2}-inverse of B. Let D be defined
as in (6.5). From BP = 0 it follows immediately that
BD = Q, and DB = I − P.
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
19
So the last two equations of the system are satisfied. Also,
BDB = QB = B and DBD = (I − P )D = D.
Conversely, suppose that (6.4) admits a solution D. Let Q = BD and
P = I − DB, then P and Q are normal projections in L(H). Moreover,
R(Q) = R(BD) ⊆ R(B). On the other hand, R(Q) = R(BD) ⊇ R(BDB) =
R(B). Therefore, R(Q) = R(B) and R(B) is pseudo-regular. Also, N (B) ⊆
N (DB) = N (P ) ⊆ N (BDB) = N (B). So that N (B) = N (P ) and then
N (B) is pseudo-regular.
In this case, we have already proven that if D is as in (6.5), then D is
a solution of (6.4). Conversely, suppose that D ∈ L(H) is a solution of (6.4).
Note that Q := BD ∈ QR(B) and P := I − DB ∈ QN (B). Let B be any
{1, 2}-inverse of B. It is straightforward to check that D satisfies
(I − P ) BQ = D.
(cid:3)
Proposition 6.6. Let B ∈ CR(H), such that R(B) is pseudo-regular. Then
there exists a solution of problem (6.1) for every C ∈ L(H) such that R(C) ⊆
B(N (B#B)[⊥]) + R(B)[⊥] if and only if N (B#B) is pseudo-regular and
N (B#B) and R(B) are nonnegative.
Proof. Suppose that R(B) and N (B#B) are nonnegative and pseudoregular.
Then, by [12, Lemma 3.4], R(B#B) is closed. Since N (B#B) is pseudo-
regular, [19, Corollary 2.5] gives that R(B#BB#B) is closed too.
Suppose that R(C) ⊆ B(N (B#B)[⊥]) + R(B)[⊥]. Then
R(B#C) ⊆ B#[(B(R(B#B)) + R(B)[⊥])]
⊆ B#[(R(B) ∩ R(BB#B)[⊥])[⊥]] ⊆ [B−1(R(B) ∩ R(BB#B)[⊥])][⊥] =
= [B−1(R(BB#B)[⊥])][⊥] = [B#R(BB#B)][⊥] [⊥] = R(B#BB#B),
where we used the fact that B#(S[⊥]) ⊆ (B−1(S))[⊥], for any closed subspace
S ⊆ H and that R(B#BB#B) is closed. Then, applying (B#)−1 to both sides
of the inclusion, we have that
R(C) ⊆ (B#)−1(R(B#BB#B)) = N (B#) + B(N (B#B)[⊥]).
Whence, by Theorem 6.1, problem (6.1) admits a solution.
Conversely, suppose that there exists a solution of problem (6.1) for
every C ∈ L(H) such that R(C) ⊆ B(N (B#B)[⊥]) + R(B)[⊥]. Then pick C
such that
R(C) = B(N (B#B)[⊥]) + R(B)[⊥].
By Theorem 6.1, we have that N (B#B) and R(B) are nonnegative and
R(C) = B(N (B#B)[⊥]) + R(B)[⊥] ⊆ B(N (B#B)[⊥]) + R(B)[⊥]. Then the
subspace B(N (B#B)[⊥]) + R(B)[⊥] is closed. Hence,
B−1(B(N (B#B)[⊥]) + R(B)[⊥]) = N (B) + N (B#B)[⊥] + N (B#B) =
= N (B#B)[⊥] + N (B#B)
20
Contino, Maestripieri and Marcantognini
IEOT
is closed, so that N (B#B) is pseudo-regular.
(cid:3)
The next result is a corollary of Proposition 6.5. We will use it in the
proof of Theorem 6.8 in order to characterized the solutions of (6.1) in term
of pseudo-inverses when R(B) and N (B#B) are pseudo-regular.
Lemma 6.7. Let B ∈ CR(H) such that R(B) is a pseudo-regular subspace
of H. Given Q ∈ QR(B), let B′ = Q#B. Then there exists a solution of the
system
B′XB′ = B′, XB′X = X,
B′X = Q#Q,
(XB′)#(XB′) = (XB′)(XB′)#,
(6.6)
if and only if N (B#B) is a pseudo-regular subspace of H.
In this case, D ∈ L(H) is a solution of (6.6) if and only if there exists
P ∈ QN (B#B) such that
D = (I − P ) B′Q#Q,
where B′ is any {1, 2}-inverse of B′.
Proof. Note that
R(B′) = R(Q#Q) and N (B′) = N (B#B).
Then apply Proposition 6.5 to B′.
(cid:3)
Theorem 6.8. Let B ∈ CR(H) and C ∈ L(H). If R(B) and N (B#B) are non-
negative pseudo-regular subspaces and R(C) ⊆ B(N (B#B)[⊥]) + R(B)[⊥], set
X1 = DC, where D ∈ L(H) is a solution of (6.6), then X1 is a solution of
problem (6.1).
Proof. By the proof of Proposition 6.6, the set B(N (B#B)[⊥]) + R(B)[⊥] is
closed.
Given a solution D of (6.6), consider P ∈ QN (B#B), Q ∈ QR(B) and
any {1, 2}-inverse B′ of B′ such that
D = (I − P ) B′Q#Q.
Observe that
Q#BDC = B′DC = Q#QC = Q#C,
where we used the fact that R(C) ⊆ R(B) + R(B)[⊥] and Lemma 4.5. Then
R(BDC − C) ⊆ N (Q#) = N (B#) or, equivalently,
B#(BDC − C) = 0.
Then, by Proposition 4.6, DC is an ImS of BX − C = 0.
On the other hand,
R(B#BDC) = R(B#C) ⊆ B#(B(N (B#B)[⊥])),
so, by applying (B#B)−1 to both sides of the inclusion, we have that
R(DC) ⊆ N (B#B) + N (B#B)[⊥] = N (P #(I − P )).
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
21
Then
P #(I − P )DC = P #DC = 0.
Thus R(DC) ⊆ N (B#B)[⊥] and, by Lemma 6.2, X1 = DC is a solution of
(cid:3)
problem (6.1).
Remark. Under the same assumptions of the last theorem, by Proposition
6.6, there exists a solution of problem (6.1). Furthermore, if R(C) 6⊆ R(B)[⊥],
a converse of Theorem 6.8 holds: if X1 is a solution of problem (6.1) then
X1 = DC, where D ∈ L(H) is a solution of (6.6).
In fact, let X1 be a solution of problem (6.1), then by similar arguments
as those in [12, Theorem 3.5], there exists P ′ ∈ QN (B#B) such that
X1 = (I − P ′)X0,
where X0 is an ImS of BX − C = 0.
Let Q ∈ QR(B) and B′ be any {1, 2}-inverse of B′. Set
D = (I − P ′) B′Q#Q.
Then, by Lemma 6.7, we have that D is a solution of (6.6). Then, proceeding
as in the proof of the last theorem, we get that DC is an ImS of BX − C = 0.
Then, by Corollary 4.2, X0 = DC + Y, with R(Y ) ⊆ N (B#B). Hence,
X1 = (I − P ′)X0 = (I − P ′)DC = DC.
Acknowledgements
Maximiliano Contino was supported by CONICET PIP 0168. Alejandra
Maestripieri was partially supported by CONICET PIP 0168. The work of
Stefania Marcantognini was done during her stay at the Instituto Argentino
de Matem´atica with an appointment funded by the CONICET. She is greatly
grateful to the institute for the hospitality and to the CONICET for financing
her post.
References
[1] Ando T., Linear operators on Krein spaces, Hokkaido University, Sapporo,
Japan (1979).
[2] Ando T., Projections in Krein spaces, Linear Algebra Appl., 431 (2009), 2346-
2358.
[3] Arias M. L., Corach G., Gonzalez C., Generalized inverses and Douglas equa-
tions, Proc. Am. Math. Soc., 136 (2008), 3177-3183.
[4] Azizov T. Y., Iokhvidov I. S., Linear operators in spaces with and indefinite
metric, John Wiley and Sons, 1989.
[5] Bogn´ar J., Indefinite inner product spaces, Springer, Berlin (1974).
[6] Contino M., Giribet J. I., Maestripieri A., Weighted Procrustes problems, J.
Math. Anal. Appl., 445 (2017), 443-458.
22
Contino, Maestripieri and Marcantognini
IEOT
[7] Contino M., Giribet J. I., Maestripieri A., Weighted least square solutions of
the equation AXB − C = 0, Linear Algebra Appl., 518 (2017), 177-197.
[8] Corach G., Fongi G., Maestripieri A., Weighted projections into closed sub-
spaces, Studia Mathematica, 216 (2013), 131-148.
[9] Douglas R. G., On majorization, factorization and range inclusion of opera-
tors in Hilbert space, Proc. Amer. Math. Soc., 17 (1966), 413-416.
[10] Dritschel M. A., Rovnyak J., Extension theorems for contraction operators on
Krein spaces, Operator Theory: Adv. Appl., 47 (1990), 221-305.
[11] Dritschel M. A., Rovnyak J., Operators on indefinite inner product spaces,
Lectures on operator theory and its applications, 3 (1996), 141-232.
[12] Giribet J. I., Maestripieri A., Mart´ınez Per´ıa F., Indefinite least-squares prob-
lems and pseudo-regularity, J. Math. Anal. Appl., 430 (2015), 895-908.
[13] Giribet J. I., Maestripieri A., Mart´ınez Per´ıa F., A geometrical approach to
indefinite least squares problems, Acts Appl. Math, 111 (2010), 65-81.
[14] Gheondea A., On the Geometry of pseudo-regular subspaces of a Krein space,
Operator Theory: Advances and Applications, vol. 14, Birkhauser, Basel
(1984).
[15] Hassi S., Nordstrom K., On projections in a space with an indefinite metric,
Linear Algebra Appl., 208-209 (1974), 401-417.
[16] Hassibi B., Sayed A. H., Kailath T., Linear estimation in Krein spaces - part
I: theory, IEEE Trans. Automat. Control, 41 (1996), 18-33.
[17] Hassibi B., Sayed A. H., Kailath T., Linear estimation in Krein spaces - part
II: application, IEEE Trans. Automat. Control, 41 (1996), 33-49.
[18] Hassibi B., Sayed A. H., Kailath T., Indefinite-Quadratic Estimation and Con-
trol. A Unified Approach to H2 and H∞ Theories, Studies in Applied and
Numerical Mathematics, 1999.
[19] Izumino S., The product of operators with closed range and an extension of
the reverse order law, Tohoku Math. J., 34 (1982), 43-52.
[20] Maestripieri A., Mart´ınez Per´ıa F., Decomposition of selfadjoint projections
in Krein spaces, Acta Sci. Math.(Szeged), 72 (2006), 611-638.
[21] Maestripieri A., Mart´ınez Per´ıa F., Normal projections in Krein spaces, In-
tegral Equations Operator Theory, 76 (2013), 357-380.
[22] Mary X., Moore-Penrose inverse in Krein spaces, Integral Equations and
Operator Theory, 60 (2008), 419-433.
[23] Mitra S. K., Rao C. R., Projections under seminorms and generalized Moore
Penrose inverses and operator ranges, Linear Algebra Appl., 9 (1974), 155-
167.
Maximiliano Contino
Instituto Argentino de Matem´atica "Alberto P. Calder´on"
Saavedra 15, Piso 3 (1083) Buenos Aires, Argentina
and
Departamento de Matem´atica -- Facultad de Ingenier´ıa -- Universidad de Buenos
Aires
Paseo Col´on 850 (1063) Buenos Aires, Argentina
e-mail: [email protected]
Vol. 99 (9999)
Operator LSP and Moore-Penrose inverses
23
Alejandra Maestripieri
Instituto Argentino de Matem´atica "Alberto P. Calder´on"
Saavedra 15, Piso 3 (1083) Buenos Aires, Argentina
and
Departamento de Matem´atica -- Facultad de Ingenier´ıa -- Universidad de Buenos
Aires
Paseo Col´on 850 (1063) Buenos Aires, Argentina
e-mail: [email protected]
Stefania Marcantognini
Departamento de Matem´atica -- Instituto Venezolano de Investigaciones Cient´ıficas
Km 11 Carretera Panamericana Caracas, Venezuela
and
Instituto Argentino de Matem´atica "Alberto P. Calder´on"
Saavedra 15, Piso 3 (1083) Buenos Aires, Argentina
e-mail: [email protected]
|
1506.02474 | 1 | 1506 | 2015-06-08T13:05:27 | On the number of solutions of a quadratic equation in a normed space | [
"math.FA",
"math.AP"
] | We study an equation $Qu=g$, where $Q$ is a continuous quadratic operator acting from one normed space to another normed space. Obviously, if $u$ is a solution of such equation then $-u$ is also a solution. We find conditions implying that there are no other solutions and apply them to the study of the Dirichlet boundary value problem for the partial differential equation $u\Delta u =g$. | math.FA | math |
On the number of solutions of a quadratic equation
in a normed space
Victor Alexandrov
Abstract
We study an equation Qu = g, where Q is a continuous quadratic operator acting from one
normed space to another normed space. Obviously, if u is a solution of such equation then −u is
also a solution. We find conditions implying that there are no other solutions and apply them to
the study of the Dirichlet boundary value problem for the partial differential equation u∆u = g.
Mathematics Subject Classification (2010): 47H30; 39B52; 35Q60.
Key words: quadratic operator, quadratic functional equation, vector space
§ 1. Continuous quadratic operators
and continuous symmetric bilinear operators
Let L and L′ be normed spaces over the field of real numbers R.
Definition 1: A continuous mapping Q : L → L′ is called a continuous quadratic operator if
Q(u + v) + Q(u − v) = 2Q(u) + 2Q(v)
for all u, v ∈ L
and
Q(ku) = k2Q(u)
for all u ∈ L and all k ∈ R.
(1)
(2)
Definition 2: Given B : L × L → L′, a continuous symmetric bilinear mapping, the mapping
Q : L → L′, defined by the formula
Q(u) = B(u, u),
is called a continuous quadratic operator.
Theorem 1: Definitions 1 and 2 are equivalent.
Proof: Let Q be as in Definition 1. By definition, put
B(u, v) =
1
4(cid:0)Q(u + v) − Q(u − v)(cid:1).
(3)
(4)
Obviously, B is continuous, Q is generated by B according to the formula (3), and B is symmetric.
Less trivial is that B is bilinear. In order to prove the letter, we define mappings F : L × L × L → L′
and f : R × L × L → L′ by the formulas
and
F (u, v, w) = 4(cid:0)B(u + v, w) − B(u, w) − B(v, w)(cid:1)
f (k) = B(ku, v) − kB(u, v),
where B is defined by (4). Our aim is to prove that
F (u, v, w) = 0
for all u, v, w ∈ L
(5)
1
and
f (k, u, v) = 0 for all u, v ∈ L and all k ∈ R.
(6)
As soon as formulas (5) and (6) will be proved, we can conclude that B is bilinear and, thus,
Definition 2 follows from Definition 1.
In order to prove (5), we use (4) and rewrite F in the form
F (u, v, w) = Q(u + v + w) − Q(u + v − w) − Q(u + w) + Q(u − w) − Q(v + w) + Q(v − w).
(7)
It follows from (1) that
and
Q(u + v + w) = 2Q(u + w) + 2Q(v) − Q(u − v + w)
Q(u + v − w) = 2Q(u − w) + 2Q(v) − Q(u − v − w).
Substituting (8) and (9) to (7) yields
(8)
(9)
F (u, v, w) = Q(u − v − w) − Q(u − v + w) + Q(u + w) − Q(u − w) + Q(v − w) − Q(v + w).
(10)
Adding (10) to (7) and taking into account (2), we get 2F (u, v, w) = 0. Thus (5) is proved.
In order to prove (6), we use (4) and rewrite f in the form
f (k, u, v) =
1
4(cid:0)Q(ku + v) − Q(ku − v)(cid:1) −
1
4(cid:0)Q(u + v) − Q(u − v)(cid:1).
From (11), we find f (0, u, v) = 0 and, using (2),
Given any integer n 6= 0, we use (5) and (12) to obtain
f (−1, u, v) = 0.
(11)
(12)
f (n, u, v) = B(nu, v) − nB(u, v) = B(cid:0)sgn n(u + u + · · · + u), v(cid:1) − nB(u, v)
= sgn n(cid:0)B(u, v) + B(u, v) + · · · + B(u, v)(cid:1) − nB(u, v) = nB(u, v) − nB(u, v) = 0.
Here sgn n denotes the sign of n, i.e., sgn n = +1 for n > 0, sgn n = 0 for n = 0, and sgn n = −1 for
n < 0.
If k is a rational number, k = m/n, then
f (k, u, v) = B(cid:18) m
nB(cid:18) 1
u, v(cid:19) −
u, v(cid:19) −
m
n
m
n
=
n
n
B(u, v) = mB(cid:18) 1
n
u, v(cid:19) −
m
n
B(u, v) =
m
n
B(u, v) −
m
n
m
n
B(u, v)
B(u, v) = 0.
Thus, f (k, u, v) = 0 for all rational k and all u, v ∈ L. Since f is continuous, f (k, u, v) = 0 for all
k ∈ R and all u, v ∈ L. This completes the proof of (6). Hence, if Q satisfies Definition 2, it satisfies
Definition 1 also.
At last, observe that straightforward computations show that, if Q satisfies Definition 1, it satisfies
Definition 2 also. Theorem 1 is proved.
Theorem 1 shows that there is a one-to-one correspondence between continuous quadratic operators
Q : L → L′ and continuous symmetric bilinear operators B : L × L → L′. By BQ we denote the
unique continuous symmetric bilinear operator that corresponds to a continuous quadratic operator
Q. Thus, Q(u) = BQ(u, u) for all u ∈ L.
2
Quadratic operators were studied by many authors from different points of view. For more details
the reader is referred, for example, to [1] and references given therein. We are going to study a
quadratic equation Qu = g, where Q : L → L′ is a continuous quadratic operator acting from a
normed space L to another normed space L′. It follows from (2) that if u is a solution of a quadratic
equation then −u is also a solution. We are interested in the following question: given a continuous
quadratic operator Q, is it true that, for all g, the equation Qu = g has no more than two solutions?
The answer is given by the following
Theorem 2: Let L and L′ be normed spaces over the field R of real numbers, Q : L → L′ be a
continuous quadratic operator. Then the following conditions are equivalent:
(i) for all g ∈ L′, the equation Qu = g has no more then two solutions, i.e., Q(u) = Q(v) implies
v = ±u;
(ii) BQ is nondegenerate, i.e., for every u ∈ L, u 6= 0, the equality BQ(u, v) = 0 implies v = 0.
Proof: Suppose that condition (ii) is not satisfied. This means that, in L, there exist u 6= 0 and
v 6= 0 such that B(u, v) = 0. Then, according to (4), Q(u + v) − Q(u − v) = 4BQ(u, v) = 0. On the
other hand, u + v 6= u − v as well as u + v 6= −(u − v). Hence, condition (i) is not satisfied. In other
words, this means that condition (i) implies condition (ii).
Now suppose that condition (i) is not satisfied. This means that there are two vectors u, v ∈ L such
that Q(u) = Q(v) and neither v = u nor v = −u. By definition, put U = 1
2 (u − v).
Obviously, U 6= 0 and V 6= 0. Moreover, according to (4), 4BQ(U, V ) = Q(U + V ) − Q(U − V ) =
Q(u) − Q(v) = 0. Hence, condition (ii) is not satisfied. In other words, condition (ii) implies condition
(i). Theorem 2 is proved.
2 (u + v) and V = 1
§ 2. On the number of solutions to a boundary value problem
for the partial differential equation u∆u = g
In order to demonstrate the advantages of the replacement of the quadratic equation Qu = g by
the linear equation BQ(u, v) = 0 (g and v are prescribed functions) described in the previous section,
we consider the following Dirichlet problem:
u(x, y)(cid:18) ∂2u
∂x2 (x, y) +
∂2u
∂y2 (x, y)(cid:19) = g(x, y)
u(x, y) = 0
for (x, y) ∈ ∂D,
for (x, y) ∈ D,
(13)
(14)
where
D =(cid:26)(x, y) ∈ R2(cid:12)(cid:12)(cid:12)(cid:12) −
π
2
< x <
π
2
, −
π
2
< y <
π
2(cid:27)
is a square and ∂D is its boundary.
Obviously, if u is a solution to (13) -- (14) then −u is also a solution. We are interested if, for
every f , the problem (13) -- (14) has no more than two solutions. Equivalently, our problem can be
reformulated as follows: is it true that if u and v are two solutions to (13) -- (14) then v = ±u?
In this section we prove that the answer is negative.
The nonlinear PDE operator in the left-hand side of (13) is a simplification of the operator in the
equation u∆u + C1∇u2 = C2, where C1, C2 ∈ R, for which a strong minimum principle is proved
among other results in [2].
Let L be the linear space of functions u : D → R, each of which is of the class C∞ in some
(specific for every u) open set containing the closure of D and vanishes on ∂D, endowed with the
norm kukL = sup(x,y)∈D u(x, y) + sup(x,y)∈D ∇u(x, y), where
∇u(x, y) =(cid:18) ∂u
∂x
(x, y),
(x, y)(cid:19)
∂u
∂y
3
is the gradient of u. Let L′ be the linear space of functions which are continuous in the closure of D
and are endowed with the norm kukL′ = sup(x,y)∈D u(x, y). At last, let the operator Q : L → L′ be
defined by the formula Qu = u∆u or, equivalently,
(Qu)(x, y) = u(x, y)(cid:18) ∂2u
∂x2 (x, y) +
∂2u
∂y2 (x, y)(cid:19).
Obviously, Q is a continuous quadratic operator from L to L′ for which BQ(u, v) = u∆v + v∆u
and the equation Qu = g coincides with (13).
Theorem 3: Let D, L, Q, and BQ be as defined above in this section. Then there exist functions
u, v ∈ L that satisfy the equation BQ(u, v) = 0 and do not vanish identically in D.
Proof: Obviously, the equation u∆v + v∆u = 0 will be satisfied if there exists a function p such
that
for (x, y) ∈ D.
∆u(x, y) + p(x, y)u(x, y) = 0,
∆v(x, y) − p(x, y)v(x, y) = 0
(15)
(16)
We look for solutions to (15) -- (16) of a special form, namely, u(x, y) = U (x)eU (y) and v(x, y) =
V (x)eV (y). In order to satisfy the boundary condition (14), we assume that
2(cid:19) = 0.
2(cid:19) = V(cid:18)±
U(cid:18)±
(17)
π
Moreover, we assume that p is a function of a single variable x, i. e., that p is independent of y. Under
these assumptions, (15) -- (16) yield
π
2(cid:19) = eU(cid:18)±
π
π
2(cid:19) = eV(cid:18)±
U ′′(x)eU (y) + U (x)eU ′′(y) + p(x)U (x)eU (y) = 0,
V ′′(x)eV (y) + V (x)eV ′′(y) − p(x)V (x)eV (y) = 0,
or, after separation of variables,
U ′′(x)
U (x)
V ′′(x)
V (x)
+ p(x) = −eU ′′(y)
eU (y)
− p(x) = −eV ′′(y)
eV (y)
= λ,
= µ,
(18)
(19)
where λ and µ are some real constants. For our purpose it is sufficient to find at least one non-zero
solution to (18) and (19) satisfying the boundary condition (17). So, we put λ = µ = 1 for simplicity.
conditions (17).
Then we easily check that eU (y) = eV (y) = cos y satisfy both the equations (18) -- (19) and boundary
The problem of finding the functions U (x), V (x), and p(x) satisfying the equations (18) -- (19) and
boundary conditions (17) is more complicated. We put by definition a = −1 and p(x) = −2q cos 2x,
where q is a real number that will be specified later. Then the equation (18) implies that the function
U (x) satisfies the equation
Similarly, the equation (19) implies that the function V (x) satisfies the equation
z ′′(x) + (a − 2q cos 2x)z(x) = 0.
z ′′(x) + (a + 2q cos 2x)z(x) = 0,
4
(20)
(21)
which can be obtained from (20) by substituting −q instead of q.
The equation (20) is known as the Mathieu equation. Properties of its solutions are studied in
hundreds of articles and books among which we mention only three classical treatises [3], [4], and [5]
and the chapters [6], [7] in well-known handbooks.
Recall from [3] -- [7] that one of periodic solutions of the Mathieu equation (20) is usually denoted
by se2(x, q) and may be defined by the following expansion
se2(x, q) =
∞Xm=1
B2
2m sin(2mx),
(22)
2m satisfy the recurrence relations qB2
4 = (b2 − 4)B2
2 and qB2
2m+2 = (b2 −
where the coefficients B2
4m2)B2
2m − qB2
2m−2 for m > 2, as well as the following normalization condition P∞
Recall also that, for a given value of q, the function se2(x, q) is a solution of the Mathieu equation
(20) not for arbitrary values of the parameter a; the parameter a, referred to as an eigenvalue, depends
on q and, in fact, is a function of q. Usually, this function is denoted by a = b2(q). Though this function
is quiet complicated, it is known that b2(−q) = b2(q) and there is q∗ ≈ 8 such that b2(q∗) = −1.
2m)2 = 1.
m=1(B2
From these properties of the function se2(x, q) we immediately conclude that the function U (x) =
se2(x, q∗) satisfies the equation (20) and the boundary conditions (17). We conclude also that the
function V (x) = se2(x, −q∗) satisfies the equation (21) and the boundary conditions (17).
Hence, the functions u(x, y) = se2(x, q∗) cos y and v(x, y) = se2(x, −q∗) cos y satisfy the equation
u∆v + v∆u = 0 in D and vanish on ∂D. This completes the proof of Theorem 3.
Theorem 4: Let D, L, and Q be as defined at the beginning of this section. Then there exist
functions u, v ∈ L that satisfy the equation Q(u) = Q(v) and neither u = v, nor u = −v.
Proof follows immediately from Theorems 2 and 3.
References
[1] M. Adam, S. Czerwik, Quadratic operators and quadratic functional equation. In the book: P.M.
Pardalos (ed.) et al., Nonlinear analysis. Stability, approximation, and inequalities. New York:
Springer, 2012. P. 13 -- 37. MR2962629, Zbl 1248.39020.
[2] Xi-Nan Ma, Concavity estimates for a class of nonlinear elliptic equations in two dimensions. Math.
Z. 240 (2002), no.1, 1 -- 11. MR1906704 (2003h:35082), Zbl 1003.35047.
[3] M.J.O. Strutt, Lam´esche-, Mathieusche- und verwandte Funktionen in Physik und Technik. Berlin:
Springer, 1932. MR0350085 (50 #2578), Zbl 0005.16005.
[4] N.W. McLachlan, Theory and applications of Mathieu functions. Oxford: Clarendon Press, 1947.
MR0021158 (9,31b), Zbl 0029.02901.
[5] A. Angot, Compl´ements de math´ematiques `a l'usage des ing´enieurs de l'´electrotechnique et des
t´el´ecommunications. Paris: ´Editions de la Revue d'Optique, 1952. MR0049256 (14,145f), Zbl
0047.28401.
[6] G. Blanch, Mathieu functions. In the book: M. Abramowitz, I.A. Stegun (eds.) Handbook of math-
ematical functions with formulas, graphs and mathematical tables. Washington: U.S. Department
of Commerce, 1964. MR0167642 (29 #4914), Zbl 0171.38503.
[7] G. Wolf, Mathieu functions and Hill's equation. In the book: F.W.J. Olver, D.W. Lozier, R.F.
Boisvert, Ch.W. Clark (eds.), NIST handbook of mathematical functions. Cambridge: Cam-
bridge University Press, 2010. See also online companion of this book at http://dlmf.nist.gov .
MR2655368, Zbl 1198.00002.
5
Victor Alexandrov
Sobolev Institute of Mathematics
Koptyug ave., 4
Novosibirsk, 630090, Russia
and
Department of Physics
Novosibirsk State University
Pirogov str., 2
Novosibirsk, 630090, Russia
e-mail: [email protected]
Submitted: June 8, 2015
6
|
1302.4358 | 1 | 1302 | 2013-02-18T17:19:06 | Non-direct limit of simple dimension groups with finitely many pure traces | [
"math.FA"
] | There exist simple dimension groups which cannot be expressed as a direct limit of simple, or even approximately divisible dimension groups, each with finitely many pure traces, and we can specify its infinite-dimensional Choquet simplex of traces; a more drastic property is noted. On the other hand, a very easy argument shows that if $G$ is a $p$-divisible simple dimension group (for some integer $p>1$), then it can be expressed as such a direct limit. We also enlarge the class of initial objects for AF (and slightly more general) C*-algebras. | math.FA | math |
Non-direct limits of simple dimension groups with finitely many
pure traces
Abstract There exist simple dimension groups which cannot be
expressed as a direct limit of simple, or even approximately di-
visible dimension groups, each with finitely many pure traces,
and we can specify its infinite-dimensional Choquet simplex of
traces; a more drastic property is noted. On the other hand,
a very easy argument shows that if G is a p-divisible simple
dimension group (for some integer p > 1), then it can be ex-
pressed as such a direct limit. We also enlarge the class of
initial objects for AF (and slightly more general) C*-algebras.
David Handelman1
Thinking about properties of traces on dimension groups (see for example, [BeH]), especially simple
dimension groups, I realized that it would have been nice to be able to reduce to simple dimension
groups with finite pure trace space, or better, to simple dimension groups of finite rank (these
automatically have finitely many pure traces), or better still, to simple dimension groups finitely
generated as abelian groups.
This suggests three conjectures:
CONJECTURES Every noncyclic simple dimension group is a direct limit (that is, over
a directed set, and with positive maps) of simple dimension groups
(a) that are free of finite rank;
(b) that have finite rank;
(c) that have only finitely many pure traces.
Obvously, the truth of conjecture (a) would imply that of (b), which would in turn imply that of
(c). In fact, it turns out that (c) is false (so that (a) and (b) are too), via a reasonably well-known
example, and not only that, it is false in an extreme way: there is a noncyclic simple dimension
group R (in fact, an ordered ring) for which there are no nonzero positive group homomorphisms
α : G → R for any approximately divisible dimension group G with only finitely many pure traces.
(All noncyclic simple dimension groups are approximately divisible, so this is a drastic way of
showing R cannot be a direct limit of those with only finitely many pure traces.)
This contrasts with the corresponding (dual) question for Choquet simplices: a fundamental
result of Lazar & Lindenstrass [LL] is that every metrizable Choquet simplex is an inverse limit
(over the positive integers) of finite-dimensional simplices. As pointed out in [G], a quick proof of
this is derivable using dimension groups, from the main result of [EHS].
However, there is an easy result that yields many examples that are such direct limits, even
for case (b). Case (a) is somewhat problematic, but there is an example later, Z[√2][x] equipped
the strict ordering as functions on [1/3, 2/3], which is a direct limit for case (a) (the pure trace space
is a continuum, so the result is not entirely trivial).
All groups (and partially ordered groups) appearing here are torsion-free abelian; free means
free as an abelian group. Recall that the rank 2 of a (torsion-free) abelian group J is the dimension
over Q of the rational vector space J ⊗Z Q. All partially ordered groups appearing here are
unperforated.
1Supported in part by a Discovery grant from NSERC.
2Some authors have very unfortunately defined the rank of a dimension group to be the width of
the minimal Bratteli diagram realizing it. This is different from the rank of the underlying group,
and should be called the minimal width.
1
The first section gives fairly easy examples of simple dimension groups that can be expressed
as a direct limit of simple dimension groups with finitely many pure traces; this leads to a definition
of pro-finite dimensional (for a somewhat larger collection of dimension groups with order unit)
and the class consisting of them is denoted H. The second section is devoted to construction of
a single example of what we call (strongly) anti-finite dimensional (Anti-FD), a simple dimension
group R such that for all G ∈ H, the only positive group homomorphism φ : G → R is zero (R
has an even stronger property than this).
The third section shows that Anti-FD simple dimension groups can be constructed to have
any Choquet simplex as their trace space (and if the Choquet simplex is metrizable, the dimension
group can be chosen to be countable as well). Finally, the fourth section provides a characterization
of Anti-FD; it boils down to triviality of the infinitesimal subgroup and an elementary topological
constraint, that every finite rank subgroup be discrete with respect to the norm obtained from the
affine representation.
Section 6 deal with examples arising from actions of tori on UHF algebras. Sections 7 and 8
show that a small class of the latter are initial objects in the appropriate category, enlarging on
work of Elliott and Rørdam [ER].
1 Pro-finite dimensional dimension groups
The following formalizes a well-known construction. Let G be a partially ordered abelian group
(not necessarily a dimension group) having an order unit u. We define the simplification of G, Gs,
to be the same underlying group G, but with positive cone consisting of the set of order units of
G (denoted G++) together with 0. This puts a generally coarser ordering on the group, although
the identity map Gs → G induces a natural affine homeomorphism S(G, u) → S(Gs, u) (since
G+ + G++ = G++, and the order units of G and of Gs are identical). The simplification of an
ordered group is simple, and obviously, if G is already simple, it equals its simplification. Although
not formally named, this process occurs in many (well, some) papers on dimension groups.
In general, if G is a dimension group, Gs need not be: for any n > 1, take G = Zn with the
usual ordering, so that Gs is Zn with the strict ordering, that is, its positive cone consists of 0
together with the set of n-tuples, (k(1), . . . , k(n)) such that k(i) > 0 for all i. This admits discrete
traces, so cannot be a simple dimension group [G].
A dimension group G with order unit is approximately divisible if for all pure traces τ , τ (G)
is a dense subgroup of R. This is compatible with other definitions of approximately divisible
because of the following result.
THEOREM 1.1 [GH] A dimension group G with order unit is approximately divisible
iff its image in Aff S(G, u) is norm dense for one (hence for all) order unit(s) u. All
noncyclic simple dimension groups are approximately divisible.
In particular, if A is a unital AF C*-algebra, then K0(A) is approximately divisible if and only
if A has no finite-dimensional representations. This combined with the fact that a dense subgroup
of Aff K, equipped with the strict ordering, is a dimension group iff K is a (Choquet) simplex,
yields the following.
COROLLARY 1.2 Let (G, u) be a partially ordered abelian group with more than one
pure trace. Then its simplification is a dimension group if and only if (a) S(G, u) is a
Choquet simplex and (b) the image of G is norm dense in Aff S(G, u). In particular, if
G is an approximately divisible dimension group, then Gs is a dimension group.
Now suppose the partially ordered group G with order unit u can be expressed as a direct limit
of unperforated partially ordered groups, in particular, G = lim φn : Gn → Gn+1 (the following
remarks also apply when uncountable directed sets are used, however, not only does the notation
2
become tedious, but uncountable direct limits are not very important in this context), where φn
are order preserving (that is, positive group homomorphisms). Suppose each Gn has an order unit
un such that φn(un) is an order unit of Gn+1.
n ) ⊆ G++
Then φn(G++
iff there exists an integer k such that
un ≤ kx; hence kφn(x) ≥ φn(un), so that kφn(x) is an order unit in Gn+1, and by unperforation,
φn(x) is thus an order unit. Hence, the φn are positive homomorphisms φn : Gs
n+1 between
their simplifications.
n+1. To see this, note that x ∈ G++
n → Gs
n
n → Gs
n and v ∈ G++
The identity mapping sends lim φn : Gs
n+1 to G = lim φn : Gn → Gn+1, and is
obviously order preserving (since the positive cones in the simplified groups are contained in the
positive cones of the originals).
If we now assume that G itself is simple, then this map is an
order-isomorphism: if g is in G+ \{0}, then g ∈ G++, so there exists a positive integer k such that
kg ≥ u; hence there exists n such that g is represented by [h, n] and u is represented by [v, n] with
v, h ∈ G+
n , and kh ≥ v in Gn. Hence h is an order unit of Gn and [h, n] represents
g. Thus h is positive as an element of Gs
n, so everything that is positive in G is represented as the
n → Gs
image of a positive element in lim φn : Gs
LEMMA 1.3 Suppose that G = lim φn : Gn → Gn+1 is a simple partially ordered group,
each Gn is unperforated, and each Gn contains an order unit un such that φ(un) is an
order unit in Gs
n+1. Then G is also the ordered direct limit of simple unperforated
n → Gs
groups, lim φn : Gs
n+1. We have thus proved,
If we begin with a not necessarily simple ordered group G = lim φn : Gn → Gn+1 but with the
n+1 by essentially
same hypotheses on Gn and φn, then we quickly see that Gs = lim φn : Gs
the same argument.
n → Gs
n+1.
If G is a dimension group, we can write G = lim An : Zk(n) → Zk(n+1). To satisfy the order
unit condition, we telescope and delete rows if necessary so that the matrices An have no zero rows.
Then Gs = lim An : (Zk(n))s → (Zk(n+1))s. This is not very illuminating, since (Zk(n))s is not a
dimension group (unless k(n) = 1). However, if G is a Z[1/p]-module for some integer p > 1 (that
is, G is p-divisible), then G ∼= G ⊗ Z[1/p] (ordered tensor product over Z), and thus we can write
G ∼= lim An : Z[1/p]k(n) → Z[1/p]k(n+1). Now Z[1/p]k is dense in Rk, and so its simplification is
a dimension group, obviously simple.
A dimension group G is pro-finite dimensional (pro-fd) if it can be realized as a direct limit
(over a directed set) of positive group homomorphisms φα,β : Gα → Gβ where each Gα is an
approximately divisible dimension group with order unit, uα, such that φα,β(uα) is an order unit
in Gβ, and each ∂eS(Gα, uα) is finite (finite-dimensional trace space). Necessarily, a pro-FD
dimension group has an order unit. Most of the time, specifically when G is countable, the
index set can be taken to be the set of positive integers, in which case the notation simplifies to
: Gi → Gi+1. The class of pro-FD dimension groups will be denoted H. Rather surprisingly,
φi
not all simple noncyclic dimension groups belong to H.
We have seen that if G is simple and in H, then G can be represented as a limit of simple
dimension groups each with finite-dimensional trace space.
COROLLARY 1.4 If G is a dimension group that is p-divisible for some integer p > 1, then
Gs is a direct limit of simple dimension groups each with finitely many pure traces. In
particular, every simple p-divisible dimension group is a direct limit of simple dimension
groups, each having only finitely many pure traces, and is thus pro-fd.
A p-divisible dimension group is automatically approximately divisible. A similar construction
occurs if G can be represented as a limit of the form Rk(n) → Rk(n+1) where R is a noncyclic ordered
subring of the reals (for example, R = Z[√2]. If G is simple, G is again a direct limit of simple
dimension groups each with only finitely many pure traces.
3
COROLLARY 1.5 If G is a simple dimension group that is p-divisible for some prime p,
then G is a limit of simple dimension groups of finite rank.
This likely extends to groups satisfying the following property. Say an element g of a torsion-
free abelian group is i-divisible (i for infinitely) if there exist infinitely many positive integers n
such that the equation g = nxn can be solved (with xn in G). Now say the abelian group G is
wi-divisible if every element is a sum of i-divisible elements, that is, the i-divisible elements span
the group. It is plausible that p-divisible can be weakened to wi-divisible in the statement of the
result.
As an example, let Hn = ⊕k(n)
i=1 Ui,n where each Ui,n is a noncyclic subgroup of Q, and let
(An) be a sequence of strictly positive matrices, with An being of size k(n + 1) × k(n) and having
integer entries. We can take G = lim An : Hn → Hn+1. The outcome is a simple dimension
group that is wi-divisible; uninterestingly, because the Ai are strictly positive, there is a noncyclic
subgroup U of Q for which we can rewrite the terms as U k(n), so that the group is i-divisible. If
we impose the strict, rather than the ordinary direct sum ordering on each Hn, then each Hn is
now a simple dimension group, and the An are still positive homomorphisms between them (this
is why we insisted the An have no zero rows; if there were any zero rows, some positive elements
would be sent to a non-positive elements), and as in the argument above, simplicity of G entails
that the limit is just G with its original ordering. Hence G is a limit of simple dimension groups
of finite rank, specifically the Hn with the strict ordering.
On the other hand, G = Z[1/2] ⊕ Z[1/3] with the strict ordering (as a subgroup of R2) is
wi-divisible without being i-divisible.
2 Drastic example
At the opposite extreme to pro-fd, are two properties of dimension groups defined below. An
approximately divisible dimension group (G, u) with order unit is anti-finite dimensional (anti-fd)
if there are no nonzero positive group homomorphisms φ : H → G for any approximately divisible
dimension group H having finite-dimensional pure trace space (equivalently, for any H ∈ H).
This is a much stronger property than merely not being a limit of simple dimension groups with
finitely many pure traces. That simple anti-fd dimension groups even exist is somewhat surprising;
however, it turns out that for every Choquet simplex K, there exists one with trace space K (and if
K is metrizable, we can find a countable one). Actually we show there exist lots of approximately
divisible dimension groups with an even stronger property.
Let (G, u) be a partially order abelian group with order unit, and form the representation
G → Aff S(G, u), given by g 7→ bg, where bg(τ ) = τ (g) as usual. This induces a pseudo-norm on G
arising from the supremum norm on Aff S(G, u); it can also be characterized purely in terms of the
ordering on G. If the ordered group is unperforated, then the kernel of the representation consists
of the infinitesimals, Inf G. We say a group homomorphism φ : G → H between partially ordered
groups with order unit (but not necessarily sending order units to order units) is continuous if it
is continuous with respect to the pseudo-norms on G and H; equivalently, φ(Inf G) ⊂ Inf H and
the induced map φ : G/Inf G → H/Inf H is continuous with respect to to the norms on the groups
induced by their affine representations.
When G has finite-dimensional trace space, every continuous group homomorphism G → H is
bounded, hence will extend to a norm-continuous map between their completions. A continuous (or
boundedsee [G, section 7] for a discussion of bounded group homomorphisms on dimension groups)
group homomorphism from a dense subgroup of Rn to any Banach space automatically extends
to a bounded linear function from Rn to the Banach space (since continuous maps need not send
Cauchy sequences to Cauchy sequences, this is not completely obvious; however, if xn → x/k where
xn and x are in the dense subgroup (and k is an integer), then kxn → x, so continuity implies
α(kxn) → α(x), and this implies extension to a rational subspace, to which the standard method of
4
showing continuity implies boundedness for linear transformations applies.) Every positive group
homomorphism is automatically continuous, so the following even more drastic property implies
anti-finite dimensionality.
We say a dimension group with order unit R is strongly anti-finite dimensional (Anti-FD; we
use three upper case letters to distinguish it from anti-fd and Anti-fd, the latter being anti-fd at the
beginning of a sentence) if for every H in H, every continuous group homomorphism φ : H → R
is zero. This condition forces Inf R = {0} for trivial reasons. A few examples of simple anti-fd
dimension groups with no infinitesimals that are not Anti-FD are given in section 4.
While characterization of anti-fd dimension groups is a little complicated, the characterization
of Anti-FD dimension groups (within the class of approximately divisible dimension groups with
order unit) is easy: Inf R = {0} and every finite rank subgroup is discrete (in the norm topology),
In this section we show sufficiency, then give more examples (having as trace
Corollary 4.5.
space, all Choquet simplices) in the next, and finally show necessity in the last section. The
most interesting feature is that Anti-FD dimension groups exist. This means that H is far from
exhaustive, even for simple dimension groups.
Now we prepare to provide an example of a simple dimension group that cannot be obtained
as a direct limit of simple dimension groups, each with finitely many pure traces, in fact is Anti-FD,
and in particular, is a counter-example to (c).
For a pseudo-normed abelian group H, let ψ denote the map to its completion; that is, first
factor out the elements of norm zero, and complete the resulting normed abelian group. Of course,
the completion need not be a vector space (e.g., if the norm is discrete).
PROPOSITION 2.1 Suppose that H is a pseudo-normed group such that the completion
of ψ(H) a finite dimensional real vector space, and R is a normed abelian group with
the following properties:
(0) the completion of R is a Banach space
(a) every countable subgroup of R is free as an abelian group
(b) every finitely generated subgroup of R is discrete.
Then every continuous group homomorphism φ : H → R is zero.
Proof. We immediately reduce to the case that H is a dense subgroup of a finite dimensional
real vector space (so we can dispense with ψ). Now consider the subgroup φ(H) of R. If φ(H) is
nonzero and has finite rank (as an abelian group), it is countable, as a subgroup of a free group,
it is itself free; a free group of finite rank is finitely generated. Hence φ(H) is discrete. There thus
exists a continuous linear functional on the completion of R, ρ, such that ρ(φ(H)) = Z.
Then υ := ρ ◦ φ : H → R is a continuous group homomorphism. Since the completion of
Hactually pseudo-completion, but ψ kills the subgroup consisting of those elements of H with zero
norm, so we might as well assume H is dense in Rnis a finite dimensional vector space, υ extends
to a bounded linear functional on the completion. But its restriction to (the image of) H has
discrete range, which is impossible since H is dense in a vector space.
We are thus reduced to the case that φ(H) has infinite rank. Suppose the (real) vector space
dimension of the completion of H is n. Every abelian group of infinite rank contains a subgroup
of every finite rank. Hence there exists a rank n + 1 subgroup J of φ(H). As a subgroup of R, it is
free, and thus free on n + 1 generators, call them φ(hi). By (c), the group they generate is discrete,
and thus there exist continuous linear functionals αi on B (and thus continuous real-valued group
homomorphisms on R) such that αi(φ(hj)) = δij (Kronecker delta).
Then {υi := αi ◦ φ} is a collection of continuous additive real-valued group homomorphisms on
H, each of which extends to a bounded linear functional on the completion. However, υi(hj) = δij
entails that the set of bounded linear functionals is linearly independent (over the reals of course),
5
so that the dimension of the dual space of the completion, and thus of the completion itself, exceeds
•
n, a contradiction.
Properties (a) and (b) together can be restated as a single property (of normed abelian groups),
() every finite rank subgroup of R is discrete.
To see the equivalence, note that a finite rank discrete group is free [actually, a result due to
Steprans [S], generalizing a result of Lawrence [L], asserts that every discrete normed abelian
group is free]; by [Gr, Theorem 137, p 101], a countable group for which every finite rank subgroup
is free, is itself free. And of course, if R is a free abelian group, then every subgroup is free.
It is convenient, however, to separate (a) and (b), in view of the closely related examples we
obtain.
If {xi}n
i=1 is a finite and (real) linearly independent subset of a Banach space, then the abelian
group the set generates, P xiZ = ⊕xiZ not only is free (as an abelian group) on {xi} but is
discrete: the map xi 7→ ei (where ei run over the standard basis of Rn) extends to an automatically
continuous and real linear isomorphism (with continuous inverse) P xiR → Rn that sends P xiZ
to the usual copy of Zn ⊂ Rn.
EXAMPLE 2.2 A simple Anti-FD dimension group.
Set R = Z[x], the ring of polynomials with integer coefficients. Let I be a closed real interval (of
nonzero length) that contains no integers. Then a 1925 result due to Chlodovsky [C] (as cited in
[F]), says that R is dense in C(I, R) with respect to the supremum norm on I. In particular, if we
impose the strict ordering on R (that is, f ∈ R+ \ 0 iff f is strictly positive as a function on I),
then R is a simple dimension group. We note that R is an ordered ring with 1 as order unit, and
its pure trace space consists of the point evaluations at members of I, so its pure trace space is an
interval.
and its pure trace space is naturally homeomorphic to I).
i=0 xiZ for some n, so to prove the subgroup is discrete, it suffices to
with real coefficients that vanishes on I is zero), and thus L is discrete. So R is a simple dimension
group satisfying conditions (0, a, b) of Proposition 2.1, and is thus Anti-FD.
Moreover, R =P xiZ expresses R as free on(cid:8)xi(cid:9). Finally, every finitely generated subgroup
of R is contained in L = Pn
show L is discrete. However,(cid:8)xi(cid:9) is real linearly independent as a subset of C(I, R) (a polynomial
The affine representation of R (using 1 as an order unit) is simply the inclusion R ⊂ C(I, R)
•
If H is an approximately divisible dimension group with order unit, then the image of H is
dense in its affine representation. Hence if H is approximately divisible and has only finitely many
pure traces, then the natural pseudo-norm from the affine representation has Euclidean space as
its completion. The kernel of the affine repesentation consists of infinitesimals, which is exactly the
subset to be factored out in constructing the completion. Any positive homomorphism between
partially ordered abelian groups with order unit is automatically continuous with respect to the
pseudo-norm topologies on each.
An abelian group homomorphism φ : H → J where H is a pseudo-normed abelian group and
J is a normed abelian group is weakly continuous if for every continuous group homomorphism
ρ : J → R, the composition ρ◦φ : H → R is continuous. The proof of Proposition 2.1 only requires
that φ be weakly continuous. However, when the pseudo-completion of H is a finite-dimensional
real vector space, weak continuity implies continuity anyway. In the following result, dealing with
maps from members of H, we have two ways of proceeding: directly, using weak continuity of
the restriction to the image of something with finite dimensional completion, or using that the
restriction of a weakly continuous group homomorphism to one of the constituents in the direct
limit is automatically continuous.
With R the example of 2.2 (or any other anti-fd simple dimension group), we have the following.
6
Of course, R itself has the metric topology from the sup norm on the interval, and this coincides
with the metric (not just pseudo-metric) obtained from its affine representation (which would be
as a dense subring of AffM+(I) = C(I, R), M+ being the collection of probability measures on
I).
LEMMA 2.3 Let H ∈ H. If φ : H → R is a weakly continuous group homomorphism, then
φ = 0. This applies automatically if φ is an order-preserving group homomorphism.
This contrasts with Z[1/2][x] and Z[√2][x], each equipped with the strict ordering from restric-
tion to the interval; Z[1/2][x] is 2-divisible, so is even a direct limit of simple dimension groups of fi-
nite rank. And Z[√2][x] is order isomorphic to the ordered tensor product R⊗ZZ[√2] (where Z[√2]
is given the total ordering inherited from R: we can write R = lim An : Zk(n) → Zk(n+1) where the
An are strictly positive matrices (since R is simple, this can be arranged by telescoping), and since
R and thus R ⊗ Z[√2] is simple, the latter is the limit An ⊗ 1 : (cid:0)Z[√2]k(n)(cid:1)s
→ (cid:0)Z[√2]k(n+1)(cid:1)s
,
each direct sum equipped with the strict ordering. Every(cid:0)Z[√2]k(n)(cid:1)s
is a simple dimension group
(from being dense in Rk(n)). In this latter example, the underlying group is free, and it is a direct
limit of simple dimension groups that are finitely generated and free. In particular, both Z[1/2][x]
and Z[√2][x] equipped with the strict ordering from some closed interval in (0, 1) are simple dimen-
sion groups in H, but their intersection, R = Z[x] with the strict ordering on the same interval, is a
simple dimension group which has no nonzero incoming weakly continuous group homomorphisms
from any element of H.
The ring Z[√2][x] with the strict ordering satisfies (a) of Proposition 2.1, but not (b); on the
other hand, Z[1/2][x] satisfies (b) but not (a). Both satisfy (0).
We actually have uncountably many choices for I that give rise to nonisomorphic simple
dimension groups: for example, the unordered pair consisting of left and right endpoints of the
interval, I = [a, b], is topologically determined from the pure trace space (since they are the only two
points with no neighbourhoods homeomorphic to an open interval), and the corresponding value
groups, {Z[a], Z[b]}, viewed as subgroups of the reals, is an order-theoretic invariant of Z[x] with
the strict ordering coming from I (and there are uncountably many different order isomorphism
classes of Z[a]). In the next section, we will show that arbitrary Choquet simplices can be realized
as the trace space of simple Anti-FD dimension groups.
The argument in Proposition 2.1 was divided in two parts; first, dealing with subgroups of
finite rank, then with those of infinite rank. Had the following notion (not good enough to be a
conjecture) been true, the second argument would have been unnecessary.
(Notion) If G is a dense subgroup of Rn, there exists a finite rank subgroup of G that is also dense.
This is plausible but false; this phenomenon accounts for problems in characterizing anti-fd dimen-
sion groups, as discussed in section 4.
EXAMPLE 2.4 A countable dense subgroup of R2 which contains no dense subgroups of
finite rank. This yields a simple dimension group with exactly two pure traces to which
there are no positive homomorphisms from any noncyclic simple dimension group of
finite rank.
Let {α0 = 1, α1, α2, . . . } be an infinite set of real numbers that is linearly independent over the
rationals. For i = 0, 1, 2, . . . set xi = (αi, 2−i) ∈ R2. Let G = P xiZ ⊂ R2. By examining the
first coordinates, we see that G is a free abelian group on {xi}, obviously of infinite rank.
We observe that 2x1 − x0 = (2α1 − 1, 0) and 4x2 − x0 = (4α2 − 1, 0). Since {2α1 − 1, 4α2 − 1}
is linearly independent over the rationals, we see that R⊕ 0 is contained in the closure of G. Thus
for all i, (0, 2−i) is contained in the closure of G, so that 0 ⊕ R is also contained in the closure,
and thus, as the closure is a group, R2 is the closure. So G is dense in R2.
7
finite, it is finitely generated as an abelian group. Hence there exists n such that H ⊆Pn
Now suppose H is a subgroup of finite rank. Since G is free, H must be free, and as its rank is
i=1 xiZ.
The restriction of the second coordinate functional has values in 2−nZ, which is discrete.
In
particular, H has a trace with discrete range, so cannot be the image of a noncyclic simple dimension
group (even dropping the requirement that the image be dense in G).
Impose the strict ordering on G; it is the desired simple dimension group; all noncyclic simple
dimension groups have dense range in their affine representation.
As a special case, let α be a real transcendental number, and set αi = αi. The resulting G is
simply the ring Z[x] equipped with the ordering derived from the two ring homomorphisms x 7→ α
and x 7→ 1/2; the image is dense in R2, making Z[x] into a simple dimension group (and an ordered
ring), with these two homomorphisms as the only pure traces (up to normalization at 1). Since α
is transcendental, the map x 7→ α has zero kernel. This example is simply Z[x] with a much larger
•
positive cone than that of R in Example 2.2.
3 Other Choquet simplices
In the construction of the Anti-FD simple dimension group R = Z[x] with the strict ordering from
restriction to I, the set I is a closed interval; everything works just as well if we let I be any
infinite compact subset of the open unit interval (since a polynomial is determined by its values
on any infinite set). As we can embed a Cantor set in the open interval (for example, truncate the
usual Cantor set at both ends), we can even arrange that the pure trace space of Z[x] equipped
with the strict ordering be a Cantor set, hence totally disconnected; the result will still be a simple
dimension group satisfying (0), (a), and (b) of Proposition 2.1. However, we can go much farther.
We show that for every (metrizable) Choquet simplex K, there exists a (countable) simple
dimension group with order unit (G, u) whose trace space is K that satisfies conditions (a) and (b)
of the proposition, and therefore has the strong anti-finite dimensional property.
The following is a minor variation on well-known result.
LEMMA 3.1 Let B be a separable infinite-dimensional real Banach space, and let Q be a
countable linearly independent subset. Then there exists a countable dense set P such
that P ∩ Q = ∅ and P ∪ Q is linearly independent.
Proof. Since B is separable, there exists a countable dense set {pi}∞
i=1. Since B is complete and
not finite-dimensional, its dimension (as a real vector space) is uncountable. Select a basis for
B (as a real vector space), A = {eα}. Every element of B is uniquely representable in the form
b = P λαeα with λα ∈ R where all but finitely many λα are zero; define the support of b, given
by supp b = {eα ∈ A λα 6= 0}.
Let J0 be the union of the supports of all the elements of Q; this is countable. Let J1 be the
union of the supports of all the elements of {pi}. Since J := J1 ∪ J0 is countable, we can find in
A \ J , a countable family of pairwise disjoint subsets Vi each of which is itself countably infinite.
Index the members of Vi as Vi = {v(i, 1), v(i, 2), . . . }, where v(i, n) ∈ A \ J . For each pair (i, n),
set Pi,n = pi + 2−nv(i, n)/kv(i, n)k. Then we note that limn→∞ Pi,n = pi, so that the countable
set P := {Pi,n}N×N is dense.
Since the support of every element of Q lies in J0, P ∩ Q = ∅; since for any fixed pair (i′, n′),
the element v(i′, n′) appears exactly once in the supports of Pi,n as (i, n) varies, it easily follows
•
that P ∪ Q is linearly independent (using linear independence of Q).
We can ask whether
(*) for every real infinite-dimensional Banach space, there exists a dense linearly independent
subset.
This has been solved affirmatively in [BDHMP; Proposition 3.2] (I am indebted to Ilijas Farrah
8
for explaining their argument to me).
The argument in the separable case used the fact that the dimension (as a real vector space)
exceeded the cardinality of one of the dense subsets. However, this dimension property fails (when
CH holds) for all non-separable Banach spaces whose dimension is ℵ1 (and probably fails, in the
presence of GCH, for all non-separable Banach spaces). The argument of the separable case can
be adapted if B has a dense subspace of codimension at least as large as its dimension.
A weaker hypothesis that would still be enough for our purposes is the following:
whose Z-span (that is, P eαZ) is dense.
(**) for any real infinite dimensional Banach space, there exists a linearly independent set {eα}
Whenever this occurs, P eαZ is free (on {eα}) and has the property that every finitely generated
subgroup is discrete.
COROLLARY 3.2 Let K be a metrizable Choquet simplex. Then there exists a countable
dense subgroup G of Aff K that is free as an abelian group and for which every finitely
generated subgroup is discrete. Equipped with the strict ordering, this G is a countable
simple dimension group with trace space K, with the strong anti-finite dimensional
property.
Proof. The Banach space B = Aff K is separable (since K is metrizable), and thus there exists a
countable, dense, real linearly independent subset of B, {en}∞
n=1; we specify Q = {111} (consisting
of the constant function), so that if we take as order unit, u = 111, the affine representation agrees
with the inclusion of G in Aff K. Then G = P enZ is dense; linear independence over R implies
that G is free on {en}, and every finitely generated subgroup of G is contained inPm
i=1 eiZ, which
•
These examples have a property that our original R = Z[x] does not: the group is free on a
is discrete, and thus the subgroup is itself discrete. The rest is immediate.
set which is itself dense.
By the affirmative solution to (*) in [BDHMP], we also obtain immediately the following.
PROPOSITION 3.3 Let K be a Choquet simplex. Then there exists a dense subgroup G
of Aff K that is free as an abelian group and for which every finitely generated subgroup
is discrete. Equipped with the strict ordering, this G is a simple dimension group with
trace space K, with the strong anti-finite dimensional property.
4 Characterization of Anti-FD
The distinction between anti-finite dimensionality and its strong form is not large.
LEMMA 4.1 Let (R, v) be an anti-finite dimensional approximately divisible dimension
group with order unit such that Inf R = {0}. Let (G, u) be an approximately divisible
dimension group with order unit such that ∂eS(G, u) is finite. Then there exists no
continuous group homomorphism φ : G → R such that φ(G) contains an order unit of R.
Proof. Suppose φ(G) contains an order unit. We may replace G by its simplification Gs; the
topology is unchanged as is the set of infinitesimals, but now Gs is a simple dimension group, and
thus so is the quotient Gs/Inf (Gs). Since φ(Inf G) ⊂ Inf R (from continuity of φ with respect to
the pseudo-norms), φ induces a continuous map from Gs/Inf Gs to R.
Hence we are reduced to the case that G is a simple dimension group with Inf G = 0; the
latter implies we may regard G as a dense subgroup of Rn = Aff S(G, u). A continuous group
homomorphism from G is automatically bounded, hence sends Cauchy sequences to Cauchy se-
quences, so that φ extends to a map from the completion of G, Rn, to the completion of R, which
is Aff S(R, v); call it Φ : Rn → Aff S(R, v).
9
Now C := Φ−1(Aff S(R, v)++) is open; by hypothesis, it is nonempty. Obviously, C ∩−C = ∅,
C − C = Rn (since C, and therefore C − C, contains an open ball), and C + C ⊆ C. Thus
C ′ = C ∪{000} is a proper convex cone in Rn containing an open ball. Hence we may find a basis for
Rn, {a1, . . . , an} inside C. Let D be the convex cone spanned by this basis; then D is a simplicial
cone on Rn (that is, it is obtained from the original ordering by applying an invertible matrix to
everything in sight).
Density of G in Rn implies that if we impose the strict ordering on Rn given by the basis
(that is, if nonzero x = P λiai where λi ∈ R, then x is in the positive cone iff λi > 0 for all i),
then G becomes a simple dimension group with respect to this orderingcall it (as an ordered group)
G′. Since the positive cone is contained in D, which in turn is contained in C ′, it follows that φ is
positive as a group homomorphism from G′ to R. Since the trace space of G′ is finite-dimensional,
•
we reach a contradiction: φ must be zero, as R is anti-finite dimensional.
Now work toward completing the characterization of strong anti-finite dimensionality.
LEMMA 4.2 Let {fi} be a finite linearly independent subset of a Banach space B. Then
there exists K > 0 (depending only on {fi}) such that if ei are elements of B with
kei − fik ≤ K for all i, then {ei} is linearly independent.
Proof. Form the finite-dimensional vector spaceP fiR; since all norms on finite dimensional vector
spaces are equivalent, choosing the l1-norm, there exists K such that for all choices of λi ∈ R, we
have kP λifik > KPλi. If on the other hand, P λiei = 0, then
KXλi ≥Xλikfi − ek ≥(cid:13)(cid:13)(cid:13)X λi(fi − ei)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)X λifi(cid:13)(cid:13)(cid:13) > KX λi,
a contradiction.
•
LEMMA 4.3 Suppose that H is a finite rank subgroup of a Banach space B. Then we
may decompose H = K ⊕F where F is a discrete group, H = K ⊕F and K is the maximal
real subspace of H.
Proof. Since HQ (the rational vector space spanned by H inside B) is a finite dimensional
rational vector space, H is contained in a finite dimensional subspace V of B. Since all norms
are equivalent on finite dimensional spaces, we can write H = (P xjR) ⊕ (P fiZ) where the
complete set {xj}∪{fi} is linearly independent, and the maximal real subspace is the left summand.
Necessarily, P fiZ is discrete.
By density, we may find ei in H such that kei − fik is as small as we like; thus with xj = ej,
we obtain that {xj}∪{ei} is linearly independent; in particular,P eiZ is discrete. It easily follows
that H = (P xj R) ⊕ (P eiZ).
Now consider the map H ⊂ H → P eiZ; this is onto (since each ei ∈ H and the target is a
free abelian group), and so its kernel K splits, that is, H = K ⊕ (P eiZ) where K ⊆P xjR. Now
it is routine to verify that H = K ⊕ (P eiZ), and thus K must be the maximal real subspace of
•
Lawrence [L] has shown that any countable discrete normed abelian group (in particular, any
countable discrete subgroup of a Banach space) is free; this was extended by Steprans [S], dropping
countable. In any event, a finite rank discrete abelian group is free on a finite set, and this does
not require either of these results.
H.
COROLLARY 4.4 Let (R, v) be an approximately divisible dimension group with order
unit.
(a) If R is strongly anti-finite dimensional, then Inf R = 0 and all finite rank subgroups
of R are discrete. In particular, every countable subgroup of R is free.
10
(b) If R is anti-finite dimensional and Inf R = {0}, then for every finite rank subgroup
H of R whose closure is a vector space, H ∩ R++ = ∅.
Proof. First we prove that if R is strongly anti-finite dimensional, then Inf R = 0. This is trivial:
if L = Inf R, form the simple dimension group H = Q ⊕ L with positive cone (q, l) > 0 iff q > 0.
Then H is a simple noncyclic dimension group, and the map (q, l) 7→ l ∈ Inf R is automatically
pseudo-norm continuous, hence must be zero. So L is zero.
Let H be a finite rank subgroup of R that is not discrete. Taking the closure in the Banach
space Aff S(R, v), we have a split decomposition H = K ⊕ F where F is free and discrete and
the closure of K is the maximal real subspace of H; if H is not discrete, then K 6= 0. Since K
is dense in a finite dimensional vector space, we can equip it with the structure of a noncyclic
simple dimension group such that the affine representation of K is simply the embedding in K.
Necessarily, the identity map K → K ⊂ R is continuous, so must be zero, a contradiction to strong
anti-finite dimensionality.
If instead, R is only anti-finite dimensional, we can form Rs and factor out the infinitesimals;
then the hypotheses still apply, so we can reduce to the case that Inf R = 0. Now using K as in
the preceding paragraph, the hypotheses of lemma 4.1 above yield a positive map from a simple
•
noncyclic dimension group to R, again a contradiction.
COROLLARY 4.5 Let (R, v) be an approximately divisible dimension group with order
unit. Then R is Anti-FD if and only if Inf R = 0 and every finite rank subgroup of R is
discrete.
Proof. Follows from 2.1 and 4.4.
•
One would like a corresponding characterization of anti-fd, along the lines of, the simple
dimension group R is anti-fd if and only if every finite rank subgroup H of R/Inf R whose closure
is a vector space misses R++. However, there is a problem arising from the phenomenon illustrated
in Example 2.4, that dense subgroups of Rn need not contain any dense subgroups of finite rank.
Necessary and sufficient, in case R is simple, is that (after factoring out Inf R), if H is a subgroup
of R whose closure is a finite dimensional real vector space, then H ∩ R++ = {0}, but this is
practically tautological.
Finally we can give an example of a countable simple anti-fd with no infinitesimals that is not
Anti-FD. Let R = Z[x] + (1− 2x)Q (a subgroup of Q[x]) with the strict ordering from the interval
I = [1/3, 2/3]; this is a countable simple dimension group whose pure trace space can be identified
with I. The set (cid:8)1, 1 − 2x, x2, x3, . . .(cid:9) is linearly independent, and (1 − 2x)Q misses the positive
cone. Let φ : G → R be a positive group homomorphism, where G has finitely many pure traces.
If φ(G) ⊂ Z[x] + n−1(1 − 2x)Z ⊂ (1/n)Z[x] for some n, then as the latter is order isomorphic
to Z[x] with the strict ordering and is thus Anti-FD, and we obtain φ = 0.
Now φ(G) contains a real basis for its closure, say {p1, p2, . . . , pn}, where each pi is a polyno-
j=0 xiR for some m, and G is
mial; this basis is contained in the finite dimensional space V :=Pm
in R, that is, φ(G) ⊆ V ∩ R, and the latter is just Pi=0,2,3,...,m xiZ + (1 − 2x)Qwhich is of
finite rank. Hence φ(G) is of finite rank; the closure of Pi=0,2,3,...,m xiZ + (1 − 2x)Q is just
(1 − 2x)R +Pi=0,2,3,...,m xiZ; hence φ(G), being dense in a vector space, must map into the real
subspace, so φ(G) ⊆ φ(G) ⊆ (1 − 2x)R. But this is impossible, as (1 − 2x)R contains no positive
elements of R.
This example satisfies condition (b) of Proposition 2.1, but not (a). If instead we take R =
Z[x] + √2(1 − 2x)Z, the same argument shows that the resulting ordered group is again an anti-
contained in this closed subspace. Hence the closure of φ(G) is contained in V .
Since φ(G) thus consists of polynomials of degree less than or equal m and is contained
11
fd simple dimension group that is not Anti-FD, but satisfies condition (a) and not (b). And
R = Z[x] + (1 − 2x)Q[√2] is yet another example of a simple anti-fd but not Anti-FD dimension
group, this time satisfying neither (a) nor (b).
5 Consequences
By direct translation from well-known results (too well-known to bother referring to), we have
some consequences for unital AF C*-algebras and minimal actions on Cantor sets. Pro-fd simple
dimension groups are K0 of AF algebras which are direct limits of simple AF algebras each with
finitely many pure traces. On the other hand, if A is an AF algebra whose K0 group is anti-
fd, then A contains no simple infinite dimensional AF-subalgebra with finitely many pure traces
(more generally, the simple subalgebra need not be AF, but its K0 group should be dense in a
finite dimensional vector space).
If (X, T ) is a minimal action on a Cantor set, and its dimension group (ordered Cech coho-
mology) is anti-fd, then (X, T ) is not even orbit equivalent to a minimal system with a non-trivial
map to a minimal action on a Cantor set that has only finitely many ergodic measures.
6 More examples
Very often, a vaguely ring-like structure on dimension groups is enough to guarantee the condition
that every finite rank subgroup is discrete.
Let A = Z[x, x−1] equipped with the coordinatewise ordering. For a f ∈ A, define c(f ),
the content of f , as the greatest common divisor of the nonzero coefficients of f . Gauss' Lemma
implies that c(f g) = c(f )c(g) for f, g ∈ A. Let (Pi) be a sequence of elements of A+, and form
R = lim ×Pi : A → A as a partially ordered A-module. This is obviously a dimension group.
The order-theoretic properties of such dimension groups, beginning instead with A = R[x, x−1]
are studied in detail in [BH]. The determination of the positive cone and the pure trace (there
called state) space is independent of the choice of coefficient rings (Z or R), although of course
the topological and underlying group structures are dependent on it. When we quote a result from
[BH], it will apply to the pure trace space or to strong positivity.
i=1 Pi. For f ∈ A, define Log f =(cid:8)k ∈ Z(cid:12)(cid:12) (f, xk) 6= 0(cid:9), where (f, xk) denotes
Define Qj =Qj
Now let R be the order ideal of R generated by the constant function 1; since an order ideal
of a dimension group is again a dimension group, R is a dimension group. We can identify R with
the following subgroup of Z[x−1, p−1
the coefficient of xk in f .
],
i
R ≡ R(pi) =(cid:26) f
Qn (cid:12)(cid:12)(cid:12)(cid:12) n ∈ Z+; Log f ⊆ Log Qn(cid:27) .
Z(cid:17); this is a finitely generated subgroup of R.
Form Gn =Pi∈Log Qn(cid:16) xi
Qn
Now assume that c(pi) = 1 for almost all i. We may delete finitely many pi and so assume
that c(pi) = 1 for all i. We now show that every finite rank subgroup of R is discrete.
(i) If a ∈ R and ma ∈ Gn for some nonzero integer m, then a ∈ Gn. Write a = f /Qr for
some f ∈ A and r ≥ 1 such that Log f ⊆ Log Qr. Then there exists integers ti such that
mf /Qr = Pi∈Log Qn tixi/Qn. Let g be the numerator of the right side, so that Log g ⊆ Log Qn.
Set s = max {r, n}, and multiply the equation mf /Qr = g/Qn by Qs. Since Qs/Qn and Qs/Qr
are products of pi, each has content one. Thus c(g) = c(mf ) = mc(f ). Thus m divides c(g), so
that h = g/m is in A. Thus a = f /Qr = h/Qn ∈ Gn.
(ii) Gn is discrete. It suffices to show (cid:8)xi/Qn(cid:9) i ∈ Log Qn is linearly independent over the reals.
If {αi}i∈Log Qn are real numbers such that P αixi/Qn = 0, on multiplying by Qn, we obtain
P αixi = 0, forcing all αi = 0.
12
(iii) Any finite rank subgroup of R is a subgroup of some Gn. If S is a finite rank subgroup, we
may find a finite subset si of S such that S′ =P(siZ) has rank equalling that of S. There exists
n such that all si ∈ Gn. If s ∈ S, there exists a nonzero integer m such that ms ∈ S′ ⊂ Gn. By
(i), s ∈ Gn. Hence S ⊆ Gn.
(iv) Every finite rank subgroup of R is discrete. Follows from (ii) and (iii).
Thus we have proved most of the following bifurcation result.
PROPOSITION 6.1 Let pi be a family of elements of A+, and set R = R(pi). Then all
finite rank subgroups of R are discrete if and only if c(pi) = 1 for almost all i. If c(pi) > 1
for infinitely many i, then R ∈ H.
Proof. If c(pi) = 1 for almost all i, then we have just shown all finite rank subgroups are discrete.
Otherwise, c(pi) > 1 for infinitely many i. By telescoping, we may assume that di := c(pi) > 1 for
all i. Then we can write pi = diPi for some Pi ∈ A+, and it easily follows that if U = lim ×di :
Z → Z, then R(pi) ∼= U ⊗ R(Pi) as ordered abelian groups, hence we can write R(pi) as a limit of
dimension groups of the form U n(i), hence R(pi) ∈ H.
•
We can also often tell when the ordered groups R(pi) have dense range in their affine repre-
sentation (we normally use 1 as the order unit for the representation).
Density of the image of R(pi) in its affine representation with respect to the order unit 1 (hence
as this is a dimension group, with respect to any order unit) is equivalent to the nonexistence of
discrete pure traces. The kernel of a pure discrete trace of a dimension group is a maximal order
ideal (with quotient Z), by [GH, xxx]. Of course, we assume that all pi have at least two nonzero
terms.
There are two obvious maximal order ideals in R(pi), namely, the kernels of the point eval-
uations at 0, that is, τ0 : f /Qn 7→ limt↓0 f (t)/Qn(t) and τ∞ : f /Qn 7→ limt↑∞ f (t)/Qn(t).
It
is easy to evaluate the range of these two traces, in terms of leading and terminal coefficients.
By multiplying each pi by a suitable power of x (and converting the corresponding Pn), we may
assume 0 = min Log pi for all i, and let di = max Log pi, so Dn := max LogPn = Pi≤n di. Then
) : Z → Z.
It easily follows that τ0(G) = lim ×(pi, x0) : Z → Z and τ∞(G) = lim ×(pi, xdi
τ0(f /Pn) = (f, x0)/(Pn, x0) and τ∞(f /Pn) = (f, xDn)/(Pn, xD(n)).
Hence necessary and sufficient for both these traces to have dense image is that
(†)
for infinitely many i, (pi, x0) > 1
for infinitely many i, (pi, xdi ) > 1.
(Recall that we have altered the pi so the smallest exponent with nonzero coefficient is the constant
term.)
Hence, whenever ker τ0 and ker τ∞ are the only maximal order ideals of R(pi), these condi-
tions are also sufficient for density of the image of R(pi) in its affine representation, and thus for
approximate divisibility.
There are lots of situations in which these are the only maximal order ideals. We first make an
observation: if Log pi = Log qi, then there is an obvious order isomorphism between the lattices of
order ideals of R(pi) and those of R(qi). So to determine when they are the only maximal ideals,
we can replace all the nonzero coefficients of all the pi by 1. When we apply this process to pi, the
result will be denoted p′
i),
the lattice of order ideals does not.
i. Although the pure trace space changes when we go from R(pi) to R(p′
If (pi) is strongly positive (see [BH]); this is independent of the choice of coefficients], then the
pure traces are precisely the point evaluations (including τ0 and τ∞); the converse does not hold.
Since maximal order ideals are automatically contained in kernels of pure traces (not necessarily
13
discrete), it follows that if (pi) is strongly positive, then ker τ0 and ker τ∞ are the only maximal
order ideals. Hence, if there exists a strongly positive sequence (qi) such that Log pi = Log qi, then
R(pi) has only the two maximal order ideals.
If p is a polynomial written in the form p = P(p, xi)xi, we call an exponent i (or the cor-
responding monomial) an isolani* if (p, xi−1) and (p, xi+1) are both zero. Next we note that if
infinitely many pi have no leading or terminal isolani, then the sequence (p′
i) (obtained by replacing
each nonzero coefficient in pi by 1) is strongly positive (an easy consequence of the Superposition
Lemma of [BH]). Hence if infinitely many pi have no isolani, then R(pi) has dense range in its
affine representation if and only if (†) holds.
The pure trace space is similarly just the set of point evaluations when infinitely many pi are
equal to each other (and projectively faithful, that is Log pi − Log pi generates all of Z). Hence if
infinitely many Log pi are equal to each other and projectively faithful, then R(pi) has only the
two maximal order ideals, and so † is necessary and sufficient for density of R(pi) in its affine
representation.
If there is a bound on the number of isolani, more generally, if there is an integer N such
that infinitely many pi have at most N isolani and their diameter (distance from the smallest
isolani to the largest) is bounded, then I think the pure trace space of R(p′
i) again is the set of
point evaluations. The thesis by Alan Kelm [K] contains numerous results on strong positivity and
related ideas, improving some of those in [BH].
On the other hand, there exist lacunary choices for (pi) for which there are lots of maximal
order ideals, and even lots of discrete traces (the former depends only on (Log pi), the latter depends
on the actual coefficients). For example, let p′
; these aren't projectively faithful, but we
+ x5i
can also take p′
. Obtaining conditions on the coefficients in lacunary examples so
that density holds is not trivial.
i = 1 + x3i
i = 1 + x3i
Anyway, we have at least some results.
PROPOSITION 6.2 Suppose that pi ∈ A+ such that Log pi ≥ 2 for all i and (†) holds.
Then sufficient for R(pi) to have dense image in its affine representation is either
(a) for infinitely many i, pi has no leading or terminal isolani.
(b) there exists an infinite subset S of N such that for all i ∈ S, Log pi are equal to each
other.
If there is a bound on the degrees of the (nontrivial) polynomials pi, then (b) applies. If there
i = (1 + x)(1 + xi!),
i) is strongly positive because of the factor (1 + x) but all finite products will have gaps),
are no gaps, then there are no isolanis either (gaps can even persist, e.g., if p′
then (p′
and so the proposition applies to such sequences as well.
COROLLARY 6.3. Suppose that pi ∈ A+ such that Log pi ≥ 2 for all i.
following hold, then R(pi) is Anti-FD.
(i) for infinitely many n, the coefficient of the smallest degree term in pn exceeds 1;
(ii) for infinitely many n, the coefficient of the largest degree term in pn exceeds 1;
(iii) for almost all n, c(pn) = 1
(iv) either (a) or (b) of Proposition 6.2.
If all of the
fications of such R, so obtaining simple dimension groups that are Anti-FD.
(Conditions (i) and (ii) together are equivalent to (†) for R(pi).) Now we can take the simpli-
For example, if pi = 2x + 3 for all i, all the conditions hold, and the corresponding R (which
also happens to be a ring, resembling the GICAR dimension group, but with values Z[1/3] and
* based on Nimzovich's term for an isolated d-pawn in chess.
14
Z[1/2] instead of Z at the two endpoints. (Here R is the ring generated by Z = 1/(2x + 3) with
positive cone generated multiplicatively and additively by Z and xZ. If pi = 3 + x + 2x2, then the
corresponding ring is not singly generated, but all the results still apply.
We can actually extend this to several variables (all the relevant results proved in [H]), as
well as ordered subrings arising from actions of compact groups (since discreteness of finite rank
subgroups will be inherited, and in these cases, we know the pure trace space is typically the orbit
space with respect to the Weyl group of the pure trace space of the original).
Some of these are candidates to be initial objects for approximately divisible dimension groups.
7 Initial object preliminaries
The proof of the following is based on the proof of [H, Proposition 1.2, last few paragraphs]. It
was obtained en passent in the case of a partially ordered ring with u = 1; however, the proof is
identical, and does not require the ordered ring structure. The condition that bG contain buR in its
closure is equivalent to the stronger property of approximate divisibility when G is a dimension
group, but not for more general partially ordered groups.
LEMMA 7.1 Suppose (G, u) is a partially ordered unperforated abelian group with order
unit that the closure of bG contains bu·R. Let p and q be relatively prime positive integers;
then for all ǫ > 0 for all r ∈ (0, 1], there exist order units v and w such that u = pv + qw
and kbv − r111/pk < ǫ/p and kbw − (1 − r)111/qk < ǫ/q.
Proof. By interchanging p and q if necessary, we may assume that r ∈ [1/2, 1]. There exists a
positive integer pk ≡ 1 mod q, so that we can find integer t such that pk = qt + 1. Now form the
subgroup kuZ+qG of G; since this contains qG, its image in the affine representation (with respect
to u) is dense. Hence there exists z ∈ G such that 111·max{r − ǫ, 1/2} /p < d(ku − qz) < 111·r/p < 1/p.
Set v = ku − qz, so bv ≫ 0, and thus v is an order unit. Now u − pv = u − pku + pqz, so
u − pv = (1 − pk)u + qz = q(tu + z). Set w = tu + z; then u = pv + qw; moreover, bw = du − pv/q =
q−1(111 − pbv) ≫ 0, the latter since bv < p−1111. Hence w is an order unit.
It also follows that kbw − 111 · (1 − r)/qk < ǫ/q (just us the triangle inequality applied to
kqbw − (1 − r)kkk = k111 − pbv − (1 − r)k.
•
The following was proved by Perera and Rørdam [PR], in the context of homomorphisms from
finite dimensional algebras to real rank zero C*-algebras (it applies to AF algebras in particular).
Here and in the preceding, we do not require G to be a dimension group, merely to have dense
image in its affine representation. If G is a dimension group, then dense image is equivalent to all
pure traces not being discrete.
COROLLARY 7.2 Let (G, u) be a partially ordered abelian group with order unit such
that the image of G in its affine representation is dense. Suppose that {pi} is a finite set
of positive integers such that gcd{pi} = 1. Given an order unit U of G, there exist order
units vi such that U =P pivi.
Proof. Let q = gcd{p2, . . . , pn}, so that p1 and q are relatively prime. Given an order unit u, by
the preceding there exist order units v1 and w such that u = p1v1 + qw. Now qG has dense image
in whatever affine representation G has dense image. Since gcd{pi/q}n
i=2 = 1, so by induction
applied to qG, qw can be expressed as as Pi≥2 piq−1zi where zi are order units in qG. Then
zi = qvi for some vi in G, necessarily order units, and u =P pivi, completing the induction.
•
i=1 Mpi C is an initial object for approximately divis-
ible AF algebras. The paper by Perera and Rørdam [PR] proves more; real rank zero and no
representations which hits the compacts is sufficient to get this.
In particular, if gcd{pi} = 1, then Qn
There is a peculiarity, noted previously, in the characterization of Anti-FD partially ordered
15
groups. Another approach emphasizes the peculiarity.
Let (H, v) be an unperforated partially ordered abelian group with order unit v, and suppose
that Inf H = {0}. For m a positive integer, we say (H, v) (or just H) is Anti-FD(m) if all subgroups
H0 of rank m or less are discrete (with respect to the norm topology induced by the order unit v);
that is, cH0 is a discrete subgroup of Aff S(H, v) in the sup-norm).
By the characterization above (Corollary 4.5), (H, v) is Anti-FD(m) for all m iff (H, v) is
Anti-FD. Examples include the critical groups of [H2]: a subgroup G of Rm is critical if it is dense
and of rank k + 1; equipped with the strict ordering inherited from the vector space, these are
simple dimension groups with m pure traces, and every subgroup of rank m or less is discrete.
Hence Anti-FD(m) simple dimension groups which are not Anti-FD(m + 1) exist for every m ∈ N.
Anti-FD(m) partially ordered groups have the expected (relatively weak) property dealing
with continuous homomorphisms.
PROPOSITION 7.3 Suppose (G, u) is an approximately divisible unperforated partially
If (H, v) is
ordered abelian group, and m is an integer such that rank G/Inf G ≤ m.
Anti-FD(m) and Inf H = {0}, then any continuous group homomorphism φ : G → H is
zero.
We require a few elementary lemmas. The norm (or pseudo-norm) on a group (G, u) is
determined by the order unit u (changing the order unit changes the norm to an equivalent one,
but even for simple dimension groups, may actually change the Choquet simplex into one that is
not affinely homeomorphic to the original!).
LEMMA 7.4 Let (G, u) be an unperforated approximately divisible group, and n a positive
integer. Let Sn = {g ∈ G++ kbgk < 1/n}. Then Sn generates G as an abelian group.
Proof. It suffices to show G++ is contained in the group generated by Sn. Select an order unit w,
and a relatively prime pair of positive integers, p and q. By Lemma 7.1 above, there exist order
units y and z in G such that w = py + qz. Since w, y, and z are all in the positive cone, we have
is true for all p/q > δ, the result follows.
kbyk ≤ kbwk/p and kbzk ≤ kwk/q. If we choose p, q > nkbwk, we have y, z ∈ Sn, and we are done. •
LEMMA 7.5 Let (G, u) and (H, v) be unperforated partially ordered groups with order
unit, and let φ : G → H be a positive group homomorphism (not necessarily sending
order units to order units). Then dφ(g) ≤ kbgk · kdφ(u) (that is, kφk ≤ kdφ(u)k).
Proof. Suppose g ∈ G and kbgk = δ > 0. If p and q are positive integers such that δ < p/q, then
kbgk < p/q entails −q111 ≪ pbg ≪ q111, so by unperforation, −qu ≤ pg ≤ qu. Applying φ, we obtain
−qφ(u) ≤ pφ(g) ≤ qφ(u), and since φ(u) is positive, we deduce kdφ(g)k ≤ (p/q)kdφ(u)k. Since this
•
Proof. (Proposition 7.3). By definition, a continuous group homomorphism maps Inf G to Inf H =
{0}, so without loss of generality, we may factor out Inf G, and thus assume that G itself has rank
m or less. Then H0 := φ(G) has rank at most m, and so is discrete in the affine representation of
(H, v). There thus exists ǫ > 0 such that if kbhk < ǫ for h ∈ H0, thenbh = 0, and thus h = 0.
g ∈ G, kbgk < 1/n entails kdφ(g)k < ǫ. Thus if we define, as in 7.4, Sn = {g ∈ G}kbgk < 1/n, then
•
φ(Sn) = {0}, and since Sn generates G as an abelian group, φ(G) = 0.
Now we see the peculiarity. If we let m → ∞, we only obtain that there are no nontrivial
homomorphisms from finite rank simple dimension groups, whereas we know that we can replace
finite rank by finite pure trace space, and we also know that not all finite pure trace space dimension
groups can be written as a direct limit of finite rank approximately divisible groups.
From the continuity of φ at 0, given ǫ, there exists a positive integer n such that for all
16
8 Initial objects
Let (H, v) be a partially ordered unperforated abelian group with order unit u. We say that (H, v)
is an initial object if for all approximately divisible partially ordered abelian groups with order
unit, (G, u), there exists an order-preserving group homomorphism (H, v) → (G, u). Normally, H
will be a dimension group (G need not be, but in cases of interest, it will be). This is designed
to be compatible with initial objects in the class of unital C∗-algebras (typically AF), via their
ordered K0-groups. Obviously if v = 2x for some x ∈ H (and necessarily in H ++ as a consequence
of unperforation), then (H, v) cannot be an initial object, although it is still possible for (H, x) to
be one.
Initial objects in the category of dimension groups (which can be defined in many different,
and equally plausible ways) that are approximately divisible are automatically at least anti-fd (by
[ER], the pure trace space of an approximately divisible initial object is infinite). Some of the R(pi)
are initial objects; if pi = 1 + x for all i, this was shown in [ER]. We show by completely different
methods that if pi = ai + bix with ai, bi positive integers exceeding 1 such that gcd(ai, bi) = 1 and
one-point compactification of Z+. All R(pi) are the ordered Grothendieck group of the fixed point
algebras of product type actions of a circle on UHF C*-algebras.
Pi min{ai, bi} /(ai + bi) < ∞, then R(pi) is an initial object. Their pure trace spaces are all the
If g, h ∈ G and k is a positive integer, then we use either g ≡ h mod k or g ≡ h mod kG
to denote h − k ∈ kG.
LEMMA 8.1 Let (G, u) be a partially ordered abelian group which has dense image in
Aff S(G, u). Then for all integers k, all ǫ > 0, all g ∈ G there exists h ∈ G such that h ≡ g
mod k and kbhk < ǫ.
Proof. Since the image of G is dense, so is that of kG and therefore that of −g + kG. Hence
−g + kG contains elements of arbitrarily small norm.
•
i=0 be a set of order units such that for some δ, there
exists a positive rational c such that kbui − ck < δ for all i (in this case, c represents the constant
function with value c). Suppose that 1 < a < b are positive and relatively prime integers. Then
there exist elements {vi}N
We will show the following. Let {ui}N −1
i=0 of G such that
(cid:13)(cid:13)(cid:13)(cid:13)bvi −
c
a + b(cid:13)(cid:13)(cid:13)(cid:13) <
δ
b − a
for i = 0, 1, 2, . . . , N
ui = bvi + avi+1
for i = 0, 1, 2, . . . , N − 1.
A consequence is that if δ < c(b − a)/(a + b), then the vi are all order units, hence in G+.
We break this into three steps. We form G0 = G ⊗ Q, elements of which we regard as formal
fractions, g/q where g ∈ G and q ∈ Q++; it has the obvious positive cone. Since it is a vector
space over the rationals, solutions of linear systems (with integer coefficients) are relatively easy
to deal with.
We define the N × (N + 1) matrix A (with rows indexed 0, 1, 2, . . . , N − 1 and columns indexed
0, 1, 2, . . . , N ) via
The equations in the second line above can be compressed into the simple AV = U , where V = (vi)T
and U = (ui)T .
Aij =
b
if i = j
a if j = i + 1
0
else.
17
For each j = 0, 1, 2, . . . , N , define Tj = u0−ab−1u1+···+(ab−1)juj, that is, Tj =Pj
i=0(a/b)iui ∈
(indexed beginning at zero) to AX = U are
G0.
LEMMA 8.2 All solutions X = (xi)T ∈ GN +1
of the form given by
a (cid:19)i(cid:18)x0 −
xi =(cid:18)−b
0
Ti−1
b (cid:19) i = 1, 2, . . . , N .
Proof. We first describe all the solutions to the homogeneous equation AX = 0. If Z = (zi)T is a
solution, then bz0 + az1 = 0 implies z1 is a rational multiple of z0, and by induction, every zi is a
rational multiple of z0. Hence we may write Z = (1, q1, q2, . . . , qn)T z0 where all the qs are rational
numbers. We easily see from the equations that qi = (−b/a)qi−1, and thus zi = (−b/a)iz0. Let
w = (1,−b/a, b2/a2, . . . , (−b/a)N )T ∈ QN +1. We have just shown that all solutions to AZ = 0 are
of the form Z = wz0.
Next, define the vector S = (Si)T where Si = (−b/a)i−1Ti−1/a for 1 ≤ i ≤ N and S0 = 0. We
claim that AS = U , so S is a particular solution. The top coordinate of AS is aT0/a = T0 = u0.
We simply calculate the rest; for 1 ≤ j ≤ N − 1,
=(cid:18)−b
a (cid:19)j−1(cid:18) b
=(cid:18)−b
a (cid:19)j
a2 Tj(cid:19)
Tj−1 + −ba
(Tj − Tj−1) =(cid:18)−b
a (cid:19)j
·(cid:18)−a
b (cid:19)j
(AS)j = bSj + aSj+1
uj
a
= uj.
0
0
Now let X be any solution in GN +1
to AX = U such that
such that X = S + wz0, which is precisely the desired statement. (Note that x0 = z0.)
to AX = U . Then A(X − S) = 000, so there exists z0 ∈ G0
•
LEMMA 8.3 Suppose ui in G0 satisfy kbui − ck < δ for some positive constant c and
some δ > 0. Then there exists a solution X = (xi) ∈ GN +1
kbxi − c/(a + b)k < δ/(b − a) for all i = 0, 1, . . . , N.
Proof. Set Y = c111/(a + b) (the column of constants). Then AY = c111, which is close (in the infinity
norm) to U . Let M be the upper triangular N × N matrix (with coordinates indexed from 0 to
N − 1) with 1 directly above the diagonal and zeros elsewhere. Set B = bI + aM (so it constitutes
i=1 (−a/b)iM i(cid:17), and we
the first N columns of A). Then B−1 (with entries in Q) is b−1(cid:16)I +Pm−1
calculate B−1A has the identity as its first N columns, and has a((−a/b)N −1−i))T as the last
column (recalling the indexing begins at i = 0).
0 (where x is the
N −1, we have X0 + ax((−a/b)N −1−i) = B−1U .
final coordinate of X) and B−1U = (r0, r1, . . . , rT
We are free to vary x, and the value of x uniquely determines X. Set x = −cu/(a + b) (recall that
the affine representation of G is with respect to the order unit u, so bx is −c/(a + b), the constant
function.
The equation AX = U is equivalent to B−1AX = B−1U . Writing X = xX T
The rest of the coordinates of X = (xi) (xN = −u/(a + b)) are given by
1
b (cid:18)ui −
xi =
a
b
b(cid:17)2
ui+1 +(cid:16) a
ui+2 + ··· +(cid:16) a
b(cid:17)N −1−i
18
uN −1(cid:19) −
acu
a + b(cid:18)−a
b (cid:19)N −1−i
.
Now we approximate bxi.
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)bxi −
c(cid:0)1 − (a/b) + (a/b)2 + ··· + (−a/b)N −1−i(cid:1)
b
−
ac
a + b(cid:18)−a
cδ
b (cid:19)N −1−i(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)<
b (cid:18)1 +
cδ
b ·
cδ
b − a
a
b
=
<
.
b(cid:17)N −1−i(cid:19)
+
a2
b2 + ··· +(cid:16) a
1 − (a/b)N −i
b − a
b
Finally,
(1−(a/b)+(a/b)2+···+(−a/b)N−1−i)
b
a+b(cid:0) −a
− a
b (cid:1)N −1−i
= 1/(a + b).
•
to AX = U, then for all ǫ, there exists a
LEMMA 8.4 If X = (xi)T is a solution in GN +1
0
solution V = (vi) in GN +1 to AV = U such that kbxi −bvik < ǫ.
Proof. First, we may find y ∈ G such that kbx0 −by0k < ǫ/a. Now we perturb y0 by an element
of the form h =Pi=0 N hiai where kbhik < a−i−2 so that vi := (−b/a)i(y + h − Ti−1/b) ∈ G (and
v0 = y + h).
Set yi = (−b/a)i(y − Ti−1/b) ∈ G0. Multiply by bi, so we obtain
aiyi = (−b)iy − (−b)i−1Ti−1.
i=0 aihi, and find the hi inductively.
The term ti−1 := (−b)i−1T−1 = (−b)i−1u0 + (−b)i−2au1 + (−b)i−3a2u2 + ··· + ai−1ui−1 belongs
to G. We want to find h ∈ G with various properties such that (−b)i(y − h) ≡ −ti−1 mod ai for
i = 1, 2, . . . , N . We write h formally as PN −1
If i = 1, the equation reduces to −bh0 ≡ u0 − by mod a; since gcd{a, b} = 1, this has a
solution in G, and we may choose a solution h0 ∈ G such that such that kh0k < ǫ/a2. Now
suppose h0, . . . , hk−1 have been defined with khik < a−2i+2 for i = 0, 1, 2, . . . , k − 1 and if h(j) =
Pi
i=0 aihi, then (−b)i(y − h(i−1)) ≡ −ti−1 mod ai for all i ≤ k. We want to find hk such that
(−b)k+1(y − h(k) − akhk+1) ≡ (−b)kTk mod ak+1.
We have Tk = Tk−1 + (−a/b)kuk, so (−b)kTk = akuk + (−b)(−b)k−1Tk−1. Also, the in-
duction assumption implies (−b)k(y − h(k−1)) ≡ (−b)(−b)k−1Tk−1 mod ak; thus x = ((−b)k(y −
h(k−1))+(b)(−b)k−1Tk−1)/ak belongs to G. The equation thus reduces to ak(−b)(x−(−b)khk) ≡ 0
mod ak+1. This is equivalent to (−b)(x − (−b)khk) ≡ 0 mod 2. Since gcd{a, b} = 1, we can solve
•
this with arbitrarily small norm. This completes the induction.
If we assume b < a (and still have gcd{a, b} = 1) instead of a < b, then we index the ui in
reverse order; then the corresponding result (with b − a replacing b − a) applies in this case.
THEOREM 8.5 Let pn = bn + anx with 1 < an, bn, P min{an, bn} /(an + bn) < ∞, and
gcd{an, bn} = 1 for all n = 1, 2, . . . . Form R(pi) ≡ R = nf /Qi≤n pi(cid:12)(cid:12)(cid:12) Log f ⊆ Log Qi≤n pio,
the order ideal generated by 1 in lim ×pn : A → A. Then R is an initial object in the
class of approximately divisible dimension groups.
Proof. Let d = Q∞
of G such that(cid:13)(cid:13)bu1
the bu1
i=1 ai − bi/(an + bn); by hypothesis, d > 0. Let (G, u) be an approximately
0, u1
1
1. Since d < a1− b1/a1 + b1,
i are order units, and the process may be continued. Now
i − 1/(a1 + b1)(cid:13)(cid:13) < d/a1− b1 and u0 = a1u1
divisible dimension group with order unit. Begin with u0 = u. By 8.3, we can find order units u1
i are strictly positive, and thus u1
we proceed by induction.
0 + b1u1
19
such that uN −1
i
= anuN
i + bnuN
i+1 for
j=0
d
<
.
QN
t=1 at − bt
j are strictly positive, hence uN
j are order units,
i = 0, 1, 2, . . . , N − 1 and for all j = 0, 1, . . . , N ,
j (cid:9)N
After the N − 1st iteration, we can find (cid:8)uN
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
t=1 at − bt/(at + bt), it follows thatbuN
Since d <QN
homomorphism R → G, sending 1 to u.
Now the assignment xj/Qi≤n pi 7→ un
and we may continue to iterate this process.
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)buN
t=1(at + bt)
1
j −
QN
j (for 0 ≤ j ≤ n) extends to a well-defined, positive
•
We can replace approximately divisible dimension groups by partially ordered abelian groups
with dense image in their affine function space, since the constructions never use interpolation.
The sequence (pn = bn + anx) is either strongly positive (and more), in which case the pure
trace space is the two-point compactification of R++, or not strongly positive, in which case the
pure trace space the one-point compactification of Z+. Strong positivity is equivalent (this is a
negation of the hypothesis in 8.5. When the sum converges (as in the last result), the product
very special case of a result in [BH]) in this case to P min{an, bn} /(an + bn) = ∞, precisely the
Q pi/(ai + bi) converges to an entire function, and the Maclaurin series coefficients determine the
pure traces ([BH] again); the pure trace space is thus the one-point compactification of Z+, and
by 6.2, the condition that an, bn > 1 more than implies that R is approximately divisible.
Here is an amusing class of initial objects, easily proved to be so, and easily determined when
they have dense image in their affine representation. Moreover, they also provide examples of
R(pi) for which checking merely at the two end evaluations is not sufficient to determine density;
in fact, the collection of maximal order ideals is indexed by a Cantor space (in a natural way; in
fact, the indexing is topological, if we use the usual topology on maximal order ideals), and we
have to check density of the image at every one of the extremal traces.
Let pi be a family of polynomials in A+ such that Log pi ≥ 2 for all i. We say the sequence
(or the corresponding dimension group) (pi) is non-interactive if whenever xm ∈ Q pi (for some
finite product of distinct pi), there exist unique wi ∈ Log pi such that m = P wi. The simplest
for ai, bi ∈ N, but there are lots of others. The term comes
example arises when pi = ai + bix2i
from ergodic theory, actions of Z, analysed with respect to measure-theoretic equivalence.
PROPOSITION 8.6 Let (pi) be a non-interactive sequence of polynomials in A+ such that
c(pi) = 1 for almost all i. Then R(pi) is an initial object in the category of approximately
divisible unital partially ordered unperforated abelian groups. The range of every pure
trace of R(pi) is a subgroup of the rationals. Moreover, (R(pi), 1) has dense image in its
affine representation if and only if for infinitely many i, none of the coefficients appearing
pi is 1 (that is, (pi, xj) 6= 1 for all j).
Proof. Define Pn = Qi≤n pi. Non-interactive means the Bratteli diagram for R(pi) (and also for
S(pi), from which the former can be obtained by beginning at a single point in the top level) is non-
interactive, meaning if j 6= k ∈ Log Pn, then j +Log pn+1∩k +Log pn+1 = ∅. Since the set of entries
of Log pn+1 has greatest common divisor 1, for each vertex of the diagram, corresponding to xi/Pn
with i ∈ Log Pn, we can solve the equations U = P(pn+1, xj)uj in an arbitrary approximately
If we use the notation ui,n to denote the order unit corresponding to i ∈ Log Pn, we have
found order units vi+t,n+1 order units with t ∈ Log pn+1 such that ui,n = P(pn+1, xt)vi+t,n+1.
Then for s ∈ Log Pn+1 we define (us,n+1) to be vi+(s−i),n+1 where i is the unique element of Log Pn
divisible etc group G (no matter what the order unit U is) at each index, independently of the
other indices on the same level.
20
such that s ∈ i + Log pn+1 (so necessarily s − i ∈ Log pn+1). Then the map xi/Pn 7→ vi,n (for
i ∈ Log Pn) yields a well-defined, positive group homomorphism (R(pi), 1) 7→ (G, u0). Hence R(pi)
is an initial object.
Next, we show that the advertised condition is precisely what is needed to guarantee that
R(pi) has no (pure) discrete traces. It is practically tautological that the the pure trace space
can be identified with the path space of the reduced Bratteli diagram (obtained by collapsing all
multiple edges to single edges), which is a Cantor set. Along each path, the corresponding trace
divides the value by the coefficient. Hence if all the pi admitted 1 as a coefficient, there would be a
path corresponding to repeated division by 1, hence the image of the trace would be Z. Conversely,
if every path ran into infinitely many levels where all the nonzero coefficients of pn are not one, we
would obtain division by infinitely many integers exceeding 1, so that the range of the trace would
be a noncyclic subgroup of the rationals. In that case, no extreme trace is discrete, hence R(pi)
•
has dense range in its affine representation.
, then (pi) is non-interactive, and R(pi) is an initial object. It will
have dense range in its affine representation if and only if for infinitely many i, both ai, bi > 1 hold.
This is a stronger condition than the condition considered previously, which amounts to (in this
case), ai > 1 for infinitely many i and bj > 1 for infinitely many j. These two conditions merely
guarantee that the τ0 and τ∞ are not discrete. These traces correspond to the extreme left path
and the extreme right path in the Bratteli diagram, respectively. The case that ai = 2 and bi = 3
yields the ordered K0-group of the AF-algebra given as the infinite tensor product, ⊗(M2 ⊕ M3),
considered in [ER].
There is an easy generalization of a sequence of non-interactive polynomials, which also yields
So if we set pi = ai + bix2i
a family of initial objects.
plicities on the edges emanating from x.
Let X be an infinite tree, and let X 0 be the Bratteli diagram obtained from X by labelling
the edges with positive integers. We sometimes identify X with its path space. Let G be the
resulting direct limit. The pure traces are given by paths in X (not X 0) as follows. Let Xn be
the set of vertices at the nth level, so X0 = {x0} consists of the initial point, and X = ∪Xn. Let
p := (x0, x1, . . . ) be a path in X (where xi ∈ Xi), and suppose the multiplicity of the edge xi → xi+1
is m(i). Let f : Xn → Z; the trace τp is given by the map [f, n] → f (xn)/m(0)m(1)· . . .· m(n− 1).
It is easy to check that this is well-defined and a pure trace.
Associated to each vertex x ∈ X is its vector of multiplicities, v(x); this is the list of multi-
Moreover, ker τp is a maximal order ideal (the order ideals of C(X, Z) are in bijection with the
order ideals of G), and thus τp is pure. Moreover, {τp}p∈X is a compact set of pure traces (routine)
homeomorphic to the path space, X. To check that this is the pure trace space of G, we note that
the embedding ZXn → ZXn+1 (of which G is the direct limit) is an order embedding: [f, n] ≥ 0
iff f ≥ 0. But if f is nonnegative along every path, then its values at each of the points in Xn
are nonnegative. Hence {τp}p∈X is a compact set of pure traces that determines the ordering, and
thus its closed convex hull is the normalized trace space of X. It follows that {τp}p∈X is the pure
trace space of G.
Now G is an archimedean (in the strong sense) subgroup of C(X, R), and density is equivalent
to every τp having dense range. Sufficient for this is that there exist infinitely many n such that
for every x ∈ Xn, 1 6∈ v(x).
we have the following.
It is a direct consequence of 6.2 that G is an initial object if c(v(x)) = 1 for all x ∈ X. Thus
PROPOSITION 8.7 Let X be an infinite tree with root x0, and X0 the corresponding
Bratteli diagram obtained by attaching positive-integer-valued weights to each vertex,
21
and let (H, u := χ{x0}) be the resulting dimension group with order unit.
(i) Sufficient for (H, u) to be an initial object is that for all x ∈ X, c(v(x)) = 1.
(ii) Sufficient for H to be approximately divisible is that for infinitely many n, for all
x ∈ Xn, no entry of v(x) is 1.
(iii) The pure trace space of (H, u) is naturally homeomorphic to the path space of X.
A reasonable question is whether all Anti-FD dimension groups are initial objects.
If we
extend the definition of initial object to require a one to one map (as was proved for the Pascal's
triangle example in [ER]), then such initial objects must be countable anf free (as abelian groups).
So this stronger property excludes the non-Anti-FD but anti-fd example Z[x] + (2x− 1)Q discussed
at the end of section four.
References
[BDHMP] T Bartoszynski, M Dzamonja, L Halbeisen, E Martinov, A Plichko, On bases in Banach
spaces, Studia Mathematica, 170 (2005) 147 -- 171.
[BH] BM Baker & DE Handelman, Positive polynomials and time dependent integer-valued
random variables, Canadian J Math 44 (1992) 341.
[BeH] S Bezuglyi & D Handelman, Measures on Cantor set: the good, the ugly, the bad, Trans
Amer Math Soc (to appear).
[C] I Chlodovsky, Une rmarque sur la reprsentation des fonctions continues par des polyn-
mes coefficients entiers, Mat Sb 32 (1925) 472 -- 475.
[EHS] EG Effros, David Handelman, & Chao-Liang Shen, Dimension groups and their affine
representations, Amer J Math 102 (1980) 385 -- 407.
[ER] GA Elliott & M Rørdam, Perturbation of Hausdorff moment sequences and an applica-
tion to the theory of C*-Algebras of real rank zero, Operator Algebras Abel Symposia,
Volume 1 (2006) 97 -- 115.
[F] Le Baron O Ferguson, Approximation by polynomials with integral coefficients, Amer
Math Soc, Rhode Island, 1980.
[G] KR Goodearl, Partially ordered abelian groups with interpolation, Mathematical Sur-
veys and Monographs, 20, American Mathematical Society, Providence RI, 1986.
[GH] KR Goodearl & David Handelman, Metric completions of partially ordered abelian
groups, Indiana Univ J Math 29 (1980) 861 -- 895.
[Gr] P Griffith, Infinite abelian group theory, University of Chicago press, 1970.
[H] D Handelman, Iterated multiplication of characters of compact connected Lie groups,
J of Algebra 173 (1995) 67 -- 96.
[H2] D Handelman, Free rank n+1 dense subgroups of Rn and their endomorphisms, J Funct
Anal 46 (1982), no. 1, 1 -- 27.
[K] A Kelm, Strong positivity results for polynomials of bounded degree, PhD thesis, Uni-
versity of Ottawa (1993).
[L] J Lawrence, Countable abelian groups with a discrete norm are free, Proc Amer Math
Soc 90 (1984) 352 -- 354.
[LL] AJ Lazar and J Lindenstrauss, Banach spaces whose duals are L1 spaces and their
representing matrices, Acta Math 126 (1971) 165193.
22
[PR] F Perera and M Rørdam, AF-embeddings into C*-algebras of real rank zero, J Funct
Anal 217 (2004) 142 -- 170.
[S] J Steprans, A characterization of free abelian groups, Proc Amer Math Soc 93 (1985)
347 -- 349.
Mathematics Dept, University of Ottawa, Ottawa K1N 6N5 ON, Canada; [email protected]
23
|
1507.02698 | 3 | 1507 | 2016-12-01T11:35:24 | On the maximal Sobolev regularity of distributions supported by subsets of Euclidean space | [
"math.FA"
] | This paper concerns the following question: given a subset $E$ of $\mathbb{R}^n$ with empty interior and an integrability parameter $1<p<\infty$, what is the maximal regularity $s\in\mathbb{R}$ for which there exists a non-zero distribution in the Bessel potential Sobolev space $H^{s,p}(\mathbb{R}^n)$ that is supported in $E$? For sets of zero Lebesgue measure we apply well-known results on set capacities from potential theory to characterise the maximal regularity in terms of the Hausdorff dimension of $E$, sharpening previous results. Furthermore, we provide a full classification of all possible maximal regularities, as functions of $p$, together with the sets of values of $p$ for which the maximal regularity is attained, and construct concrete examples for each case. Regarding sets with positive measure, for which the maximal regularity is non-negative, we present new lower bounds on the maximal Sobolev regularity supported by certain fat Cantor sets, which we obtain both by capacity-theoretic arguments, and by direct estimation of the Sobolev norms of characteristic functions. We collect several results characterising the regularity that can be achieved on certain special classes of sets, such as $d$-sets, boundaries of open sets, and Cartesian products, of relevance for applications in differential and integral equations. | math.FA | math |
On the maximal Sobolev regularity
of distributions supported by subsets of Euclidean space
D. P. Hewett∗, A. Moiola†
December 2, 2016
Abstract
This paper concerns the following question: given a subset E of Rn with empty interior and
an integrability parameter 1 < p < ∞, what is the maximal regularity s ∈ R for which there
exists a non-zero distribution in the Bessel potential Sobolev space H s,p(Rn) that is supported
in E? For sets of zero Lebesgue measure we apply well-known results on set capacities from
potential theory to characterise the maximal regularity in terms of the Hausdorff dimension
of E, sharpening previous results. Furthermore, we provide a full classification of all possible
maximal regularities, as functions of p, together with the sets of values of p for which the maximal
regularity is attained, and construct concrete examples for each case. Regarding sets with positive
measure, for which the maximal regularity is non-negative, we present new lower bounds on the
maximal Sobolev regularity supported by certain fat Cantor sets, which we obtain both by
capacity-theoretic arguments, and by direct estimation of the Sobolev norms of characteristic
functions. We collect several results characterising the regularity that can be achieved on certain
special classes of sets, such as d-sets, boundaries of open sets, and Cartesian products, of relevance
for applications in differential and integral equations.
Keywords: Bessel potential Sobolev spaces, (s, p)-nullity, polar set, set of uniqueness, capacity,
Hausdorff dimension, Cantor sets.
Mathematical Subject Classification 2010: 46E35 (Primary), 28A78, 28A80, 31B15.
1
Introduction
This paper concerns the following question Q:
Given a subset E of Rn with empty interior, an integrability parameter 1 < p < ∞, and a regularity
parameter s ∈ R, does there exist a non-zero distribution in the Bessel potential Sobolev space
H s,p(Rn) which is supported in E?
This question has arisen repeatedly in the course of the first author's recent investigations [7–10]
into the analysis of acoustic scattering by planar screens with rough (e.g. fractal) boundaries. Indeed,
for such scattering problems one of the factors determining the unique solvability of the Helmholtz
equation boundary value problems (BVPs) with Dirichlet or Neumann boundary conditions, at least
as they are classically posed, and the associated boundary integral equation (BIE) formulations, is
whether the boundary of the screen (the screen being viewed as a relatively open subset of the plane)
can support non-zero elements of H±1/2,2(R2) [8, 9].
More generally, the question Q pertains to a number of other fundamental questions about
function spaces on subsets of Rn defined in terms of the spaces H s,p(Rn). We give a simple illustration
of this in Proposition 2.11 below, where we show how Q is related to the question of whether
H s,p
for closed sets F1 6= F2 ⊂ Rn. In [9, 12, 23], where our focus is on the case p = 2, we
F1
= H s,p
F2
∗Mathematical Institute, University of Oxford, Radcliffe Observatory Quarter, Woodstock Road, Oxford, OX2
6GG, UK. Current address: Department of Mathematics, University College London, Gower Street, London, WC1E
6BT, UK. E-mail: [email protected]
†Department of Mathematics and Statistics, University of Reading, Whiteknights PO Box 220, Reading RG6 6AX,
UK. Email: [email protected]
1
On the maximal Sobolev regularity of distributions
2
demonstrate the relevance of Q for understanding when H s,2
H s,2
Ω
sets Ω1 6= Ω2 ⊂ Rn. (Here, for closed F ⊂ Rn, H s,p
Ω ⊂ Rn, H s,p(Ω) := {uΩ : u ∈ H s,p(Rn)}, H s,p
0 (Ω) = H s,2(Ω) and when eH s,2(Ω) =
, for a given open set Ω ⊂ Rn, and also for understanding when eH s,2(Ω1) = eH s,2(Ω2) for open
F := {u ∈ H s,p(Rn) : supp u ⊂ F}, and for open
H s,p(Rn)
.)
Upon consulting the classical function space literature we found a number of disparate partial
results relevant to the question Q (in particular we note [1, 2, 6, 28–30, 34, 35, 39, 40]), but no single
convenient and up-to-date reference in which these results are collected in a form easily accessible
to applied and numerical analysts. The aim of this paper is to provide such a reference, which we
hope will be of use to those interested in problems involving PDEs and integral equations on rough
(i.e., non-Lipschitz) domains. But this is not simply a review paper. We also present a number of
apparently new results, along with a range of concrete examples and counterexamples illustrating
them. The key new results we contribute include:
and eH s,p(Ω) = C∞0 (Ω)
0 (Ω) = C∞0 (Ω)
H s,p(Ω)
• a sharpening of the relationship between maximal Sobolev regularity and fractal dimension
(cf. Theorem 2.12 and Remark 2.13);
• a complete characterisation of all possible maximal regularity behaviours for sets with zero
Lebesgue measure (cf. Corollary 2.15 and the concrete examples in Theorem 4.5);
• new results on the Sobolev regularity of the characteristic functions of certain fat Cantor sets
with positive Lebesgue measure (Propositions 4.9–4.10).
While the paper does not provide a definitive answer to Q in its full generality, we hope that the
results we provide, along with the open questions that we pose, will stimulate further research.
pq(Rn) and F s
pq(Rn) [2, 30, 38, 39], of which H s,p(Rn) = F s
Function space experts might correctly observe that the question Q could be posed in a much
more general setting, for instance in the context of the Besov and Triebel–Lizorkin spaces Bs
pq(Rn)
p2(Rn) is a special case. Our decision to restrict
and F s
attention to the classical Bessel potential Sobolev spaces H s,p(Rn) (sometimes referred to as "frac-
tional Sobolev spaces", "Liouville spaces" or "Lebesgue spaces") is made for two reasons. First, it
allows a relatively simple and accessible presentation: the proofs of many of our results make use of
classical nonlinear potential theoretic results on set capacities and Bessel potentials already avail-
able e.g. in [2, 30], allowing us to avoid any discussion of more intricate theories such as atomic and
quarkonial decompositions which are typically employed in the modern function space literature to
pq(Rn) [2,30,38,39]. Second, the spaces H s,p(Rn) are sufficient for
analyse the spaces Bs
a very large part of the study of linear elliptic BVPs and BIEs, which are the focus of attention for
example in the classic monographs [27] and [13] and in the much more recent book by McLean [32]
that has become the standard reference for the theory of BIE formulations of BVPs for strongly
elliptic systems. In such applications the focus is usually on the case p = 2, but since the potential
theoretic results we cite from [2, 30] are valid for any 1 < p < ∞, it seems natural to present results
for this general case wherever possible.
In §2 we review some basic definitions, introduce
the concepts of "(s, p)-nullity" and the "nullity threshold" of a set E ⊂ Rn (which will provide
a framework within which to study question Q), and state our main results. Sets with zero and
positive Lebesgue measure require different analyses, we study them in §2.3 and §2.4 respectively.
In §3 we collect a number of results relating to certain set capacities from nonlinear potential theory,
which we use to prove the results of §2. In §4 we provide concrete examples and counterexamples to
illustrate our general results. In §5 we offer some conclusions and highlight the key open problems
arising from our investigations.
The structure of the paper is as follows.
Contents
1 Introduction
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
3
On the maximal Sobolev regularity of distributions
2 Main results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1 Preliminaries
2.2
(s, p)-Nullity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 The case s < 0 (sets with zero measure) . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 The case s > 0 (sets with non-zero measure) . . . . . . . . . . . . . . . . . . . . . . .
3 Capacity
3.1 Sets of uniqueness and (s, p)-nullity . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Examples and counterexamples
4.1 Boundary regularity and Hausdorff dimension . . . . . . . . . . . . . . . . . . . . . .
4.2 Swiss-cheese and Cantor sets
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 The case s < 0 (sets with zero measure) . . . . . . . . . . . . . . . . . . . . .
4.2.2 The case s > 0 (sets with non-zero measure) . . . . . . . . . . . . . . . . . . .
5 Conclusion
A Tensor products and traces
B A result on the non-equality of capacities
2 Main results
3
3
3
5
8
10
12
16
16
16
17
18
19
23
24
27
2.1 Preliminaries
Before stating our main results we fix our notational conventions. Given n ∈ N, let D = D(Rn)
denote the space of compactly supported (real- or complex-valued) smooth test functions on Rn.
For any open set Ω ⊂ Rn let D(Ω) := {u ∈ D : supp u ⊂ Ω}, let D∗(Ω) denote the associated space
loc(Ω) ⊂ D∗(Ω) denote the
of distributions (anti-linear continuous functionals on D(Ω)), and let L1
space of locally integrable functions on Ω; for brevity we write D∗ = D∗(Rn) and L1
loc(Rn).
Similarly for 1 < p < ∞ we write Lp = Lp(Rn) and Lp
loc(Rn), and denote by p′ the Holder
conjugate of p, i.e. the number 1 < p′ < ∞ such that 1/p + 1/p′ = 1. For any set E ⊂ Rn we denote
the complement of E by Ec := Rn \ E, and the closure of E by E. Let ∅ denote the empty set.
Given x ∈ Rn and ε > 0 let Bε(x) denote the open ball of radius ε centred at x. Let S denote the
Schwartz space of rapidly decaying smooth test functions on Rn, and S∗ the dual space of tempered
distributions (anti-linear continuous functionals on S). For u ∈ S we define the Fourier transform
u = Fu ∈ S and its inverse u = F−1u ∈ S by
loc = Lp
loc = L1
u(ξ) :=
1
(2π)n/2ZRn
e−iξ·xu(x) dx, ξ ∈ Rn,
u(x) :=
eiξ·xu(ξ) dξ, x ∈ Rn.
1
(2π)n/2ZRn
We define the Bessel potential operator Js on S, for s ∈ R, by Js := F−1MsF, where Ms represents
multiplication by (1 + ξ2)s/2. We extend these definitions to S∗ in the usual way:
u(v) := u(v),
u(v) := u(v), Msu(v) := u(Msv),
(Jsu)(v) := u(Jsv),
u ∈ S∗, v ∈ S,
and note that for u ∈ S∗ it holds that dJsu = Ms u.
For s ∈ R and 1 < p < ∞ the Bessel potential Sobolev space H s,p(Rn) (abbreviated throughout
to H s,p, except in Appendix A where different dimensions n are considered) is defined by
H s,p := {u ∈ S∗ : Jsu ∈ Lp} , with kukH s,p := kJsukLp.
Note that in the special case p = 2, the norm kukH s,2 can be realised using Plancherel's theorem as
kukH s,2 =(cid:18)ZRn
(1 + ξ2)su(ξ)2 dξ(cid:19)1/2
.
(1)
On the maximal Sobolev regularity of distributions
4
Other commonly used notation for H s,p includes H s
p (cf. [30, 38]) and Ls,p (cf. [2]). In relation
to the wider function space literature we recall that (cf. e.g. [38, §2.5.6]) H s,p = F s
p2 with equivalent
norms, where F s
pq are the Triebel–Lizorkin spaces. For s ≥ 0, let W s,p ⊂ Lp be the classical Sobolev–
Slobodeckij–Gagliardo space defined in terms of weak derivatives (cf. e.g. [32, pp. 73–74]). Then for
s ∈ N0 it holds that H s,p = W s,p with equivalent norms [38, §2.3.5] (in particular, H 0,p = Lp with
equal norms). For p = 2 this result extends to all s ≥ 0 [32, Theorem 3.16]. For p 6= 2 and 0 < s /∈ N
it holds that W s,p = Bs
pq are the Besov spaces), so
that (by [38, §2.3.2] and [37, Theorem 2.12(c)]) W s,p $ H s,p for 1 < p < 2 and H s,p $ W s,p for
2 < p < ∞.
We recall some basic properties of H s,p that will be useful later. It is well known that D is dense
in H s,p, and that the following embeddings are continuous with dense image: [38, §2.7.1]
pp with equivalent norms [38, §2.2.2] (here Bs
H t,q ⊂ H s,p,
1 < q ≤ p < ∞,
t − s ≥ n(cid:18) 1
q −
1
p(cid:19) ≥ 0.
(2)
(3)
For distributions with compact support a more general embedding result holds. Given a closed set
F ⊂ Rn, define the closed subspace H s,p
H s,p
F ⊂ H s,p by
F :=(cid:8)u ∈ H s,p : supp(u) ⊂ F(cid:9),
where the support of a distribution u ∈ D is defined in the usual way, namely as the largest closed
subset Λ ⊂ Rn for which u(φ) = 0 for every φ ∈ D(Λc) (see e.g. [32, p. 66]). Then, since pointwise
multiplication by a fixed element of D defines a bounded linear operator from H s,q to H s,p for any
1 < p ≤ q < ∞ (see e.g. [37, Lemma 4.6.2]), for any compact K ⊂ Rn the following embedding is
continuous (in particular this holds for s = t and 1 < p ≤ q < ∞):
1
H t,q
K ⊂ H s,p
K ,
t − s ≥ max(cid:26)n(cid:18)1
q −
p(cid:19) , 0(cid:27) .
The dual space of H s,p can be isometrically realised as the space H−s,p′
by the duality pairing
which in the special case p = 2 can be realised using Plancherel's theorem as
hu, viH −s,p′×H s,p = hJ−su,JsviLp′×Lp,
hu, viH −s,2×H s,2 =ZRn
u(ξ)v(ξ) dξ.
When s > n/p, elements of H s,p are continuous functions by the Sobolev embedding theorem [2,
Theorem 1.2.4]. At the other extreme, for any x0 ∈ Rn, the Dirac delta function, defined as
δx0 (φ) = φ(x0) for φ ∈ D(Rn) to fit our convention of using anti-linear functionals, satisfies
δx0 ∈ H s,p
if and only if
s < −n/p′.
(4)
Finally, we note that part (d) of Theorem 1 in [37, §2.4.2] allows the spaces H s,p to be arranged in
interpolation scales. For s0, s1 ∈ R, 1 < p0, p1 < ∞ and 0 < θ < 1,
if
s = (1 − θ)s0 + θs1
and
1
p
=
1 − θ
p0
+
θ
p1
,
then H s,p = [H s0,p0, H s1,p1]θ,
(5)
where [·,·]θ denotes the space of exponent θ obtained with the complex interpolation method (see [37,
§1.9]), and equality of spaces holds with equivalent norms. Thus, if the spaces H s,p are represented
by points in the (1/p, s)-plane, then straight segments constitute interpolation scales.
On the maximal Sobolev regularity of distributions
5
2.2
(s, p)-Nullity
We now introduce the concept of (s, p)-nullity, which will be the focus of our studies.
Definition 2.1. Given 1 < p < ∞ and s ∈ R we say that a set E ⊂ Rn is (s, p)-null if H s,p
for every closed set F ⊂ E.
F = {0}
In other words, a set E ⊂ Rn is (s, p)-null if and only if there are no non-zero elements of H s,p
supported in E.
Remark 2.2. Clearly, if F is closed then F is (s, p)-null if and only if H s,p
F = {0}. Note also
that the Definition 2.1 can be equivalently stated with "closed" replaced by "compact". Indeed, if
0 6= u ∈ H s,p with supp u ⊂ E then 0 6= φu ∈ H s,p is compactly supported in E for any φ ∈ D such
that φ(x) 6= 0 for some x ∈ supp u (cf. the proof of Proposition 2.7(i)).
While our terminology "(s, p)-null" appears to be new, the concept it describes has been studied
previously, apparently first by Hormander and Lions in relation to properties of Sobolev spaces
normed by Dirichlet integrals [25], and then subsequently by a number of other authors in relation
to the removability of singularities for elliptic partial differential operators [29, 30], and to the
approximation of functions by solutions of the associated elliptic PDEs [34]. For integer s < 0 the
concept of (s, p)-nullity is referred to (in the special case p = 2) as (−s)-polarity in [25, D´efinition 2],
"p′-(−s) polarity" in [29] and "(p′,−s)-polarity" in [30, §13.2]. A related notion is discussed in the
more general context of the spaces Bs
pq in [39, §17] (see Remark 2.13). For s > 0, our notion
of (s, p)-nullity is closely related to the concept of "sets of uniqueness" considered in [2, §11.3]
and [30, p. 692] (for integer s); this relationship is discussed in §3.1. For s > 0 and E with empty
interior, the concept of nullity coincides with the concept of (s, p)-stability, discussed in [2, §11.5].
The reason why Maz'ya [30] uses two different terminologies (polarity and set of uniqueness) for
the positive and negative order spaces is not made clear in [30], but this may be due to the fact that
Maz'ya works primarily with the spaces W s,p, where the positive order spaces are defined using weak
derivatives, and the negative order spaces are defined by duality. By contrast, in the Bessel potential
framework of the current paper, the spaces H s,p are defined in the same way for all s ∈ R, so that
it seems natural to define the notion of "negligibility" in the same way for all s ∈ R. Our decision
to introduce the terminology "(s, p)-nullity" instead of using "(p′,−s)-polarity" was made simply
for clarity (personally we find it more natural to say that a set which does not support an H s,p
distribution is "(s, p)-null" rather than "(p′,−s)-polar"). But the difference is purely semantic, so
readers familiar with the concept of polarity may read "(p′,−s)- polar" for "(s, p)-null" throughout.
The following lemma collects some elementary facts about (s, p)-nullity.
Lemma 2.3. Let 1 < p, q < ∞, s, t ∈ R and E ⊂ Rn.
(i) If E is (s, p)-null and E′ ⊂ E then E′ is (s, p)-null.
(ii) If E is (s, p)-null and t ≥ s + max{n(1/q − 1/p), 0} then E is (t, q)-null.
(iii) If E is (s, p)-null then E has empty interior.
(iv) If s > n/p then E is (s, p)-null if and only if E has empty interior.
(v) E is (0, p)-null if and only if m(E) = 0, where m denotes inner Lebesgue measure (cf. Re-
mark 3.3).
(vi) For s < −n/p′ there are no non-empty (s, p)-null sets.
(vii) Let 1 < p0, p1 < ∞ and s0, s1 ∈ R. If there exists 0 6= u ∈ H s0,p0 ∩ H s1,p1 with supp u ⊂ E,
then E is not (s, p)-null for (s, p) defined as in (5), for every 0 < θ < 1.
On the maximal Sobolev regularity of distributions
6
Proof. (i) and (ii) follow straight from the definition of (s, p)-nullity, the standard embeddings (2)
and (3), and the boundedness on H s,p of pointwise multiplication by elements of D. (iii) If E has
non-empty interior then one can trivially construct a non-zero element of D ⊂ H s,p supported inside
E. (iv) follows from (iii) and the Sobolev embedding theorem. (v) follows from the fact that a closed
set supports a non-zero Lp function if and only if it has non-zero measure. (vi) follows from (4),
and (vii) follows from (5).
Lemma 2.3 immediately implies the following proposition.
Proposition 2.4. Fix 1 < p < ∞. For every E ⊂ Rn with empty interior there exists
sE(p) ∈ [−n/p′, n/p]
such that E is (s, p)-null for s > sE(p) and not (s, p)-null for s < sE(p). We call sE(p) the nullity
threshold of E for the integrability parameter p.
Our aim in this paper is to investigate the following three questions:
Q1: Given 1 < p < ∞ and E ⊂ Rn with empty interior, can we determine sE(p)?
Q2: For which functions f : (1,∞) → [−n, n] does there exist E ⊂ Rn such that f (p) = sE(p) for
all p ∈ (1,∞)?
Q3: Under what conditions on E and p is E "threshold null" (i.e. (sE(p), p)-null)?
Our (partial) answers to these questions are summarised in §5. To state some of our results it will
be useful to introduce the "nullity set" and "threshold nullity set" of a set E ⊂ Rn, defined by
Our attempts to answer questions Q1–Q3 will make extensive use of the relationship between
(s, p)-nullity and certain set capacities from classical potential theory. The following key theorem
is stated in [30, Theorem 13.2.2] for the case where s is a negative integer, but Maz'ya's proof in
fact works for all s ∈ R. We note that this result is actually a special case of a more general result
proved in [29, Lemma 1] (where the result is attributed to Grusin [21]). The inner capacity Cap
appearing in the theorem is defined in §3 below.
Theorem 2.5 ( [30, Theorem 13.2.2], [29, Lemma 1]). Let 1 < p < ∞ and s ∈ R. Then E ⊂ Rn is
(s, p)-null if and only if Cap−s,p′(E) = 0.
Maz'ya's proof goes via the following intermediate result, which we state for future reference,
since it provides another useful characterisation of (s, p)-nullity for closed sets.
Theorem 2.6 ( [30, Theorem 13.2.1]). Let 1 < p < ∞ and s ∈ R. Then a closed set F ⊂ Rn is
(s, p)-null if and only if D(F c) is dense in H−s,p′.
Theorem 2.5, combined with the classical potential theoretic results developed e.g. in [2, 30],
and summarised in §3 below, will underpin the proofs of many of our results about (s, p)-nullity,
including part (ii) of the following proposition, the proof of which is given in §3.
Proposition 2.7. Let 1 < p < ∞ and s ∈ R.
(i) If E, F ⊂ Rn are both (s, p)-null and F has no limit points in E \ F (which holds, for example,
if F is closed), then E ∪ F is (s, p)-null. In particular, a finite union of (s, p)-null closed sets
is (s, p)-null.
(ii) For s ≤ 0, a countable union of (s, p)-null Borel sets is (s, p)-null.
NE :=(cid:8)(s, p) ∈ R × (1,∞) s.t. E is (s, p)-null(cid:9),
TE :=(cid:8)p ∈ (1,∞) s.t. E is (sE(p), p)-null(cid:9).
(6)
(7)
On the maximal Sobolev regularity of distributions
7
Proof. (i) We argue by contrapositive. Suppose that E ∪ F is not (s, p)-null, i.e. there exists a
non-zero u ∈ H s,p with supp u ⊂ E ∪ F . Then if supp u ⊂ F , F is not (s, p)-null and we are done. If
not, there exists x ∈ supp u∩ (E \ F ), and, since F has no limit points in E \ F , ε := dist(x, F ) > 0.
Let φ ∈ D(Bε(x)) with φ(x) 6= 0. Then 0 6= φu ∈ H s,p with supp φu ⊂ E, so E is not (s, p)-null.
That φu 6= 0 follows from the fact that if u ∈ D∗ and φ ∈ D, and if there exists x ∈ supp u
such that φ(x) 6= 0, then φu 6= 0 as a distribution on Rn. To see this, let ε > 0 be such that φ is
non-zero in Bε(x). Then, since x ∈ supp u, uBε(x) 6= 0 and so u(ψ) 6= 0 for some ψ ∈ D(Bε(x)).
But then, defining ϕ ∈ D by ϕ(x) := ψ/φ, for x ∈ Bε(x), and ϕ(x) := 0 otherwise, we have
(φu)(ϕ) = u(ψ) 6= 0, so φu is non-zero as claimed.
Remark 2.8. Regarding part (i) of Proposition 2.7, it is natural to ask to what extent the assumption
on F can be weakened. Certainly the result does not extend to general Borel F when s > n/p. For
a simple counterexample, let E1 denote the elements of the open unit ball B = B1(0) which have at
least one rational coordinate, and let E2 = B \ E1. Then for s > n/p both E1 and E2 are (s, p)-null,
since they both have empty interior. But E1 ∪ E2 = B, which is not (s, p)-null for any s ∈ R, since
it has non-empty interior.
This example also shows that part (ii) of Proposition 2.7 does not hold for all s ∈ R. Determining
the maximal s ∈ [0, n/p] such that Proposition 2.7(ii) holds appears to be an open problem.
The following proposition gives bounds on the nullity threshold of Cartesian products, derived
from Propositions A.1 (the lower bound) and A.3 (the upper bound) in the Appendix. More general
results can be derived for Cartesian products of more than two sets, but we do not present them here.
The assumption that E1, E2 are Borel is needed only for the upper bound in the case m(E1×E2) = 0.
Proposition 2.9. Let n1, n2 ∈ N, 1 < p < ∞, and let E1 ⊂ Rn1 and E2 ⊂ Rn2 be Borel. Then the
nullity threshold of the Cartesian product E1 × E2 ⊂ Rn1+n2 satisfies:
s−(p) ≤ sE1×E2(p) ≤ s+(p),
(8)
where
s−(p) := min(cid:8)sE1(p), sE2(p), sE1(p) + sE2(p)(cid:9),
s+(p) :=(min(cid:8)sE1(p), sE2(p)(cid:9)
p , sE2(p) + n1
min{sE1(p) + n2
if m(E1 × E2) = 0,
if m(E1 × E2) > 0.
p(cid:9)
Moreover, if either p = 2 or s1, s2 ∈ N0, and if Ej are not (sEj (p), p)-null, j = 1, 2, then E1 × E2 is
not (s−, p)-null. If p ≤ 2, m(E1 × E2) > 0, and Ej are (sEj (p), p)-null, j = 1, 2, then E1 × E2 is
(s+, p)-null.
We also mention Proposition A.2, which states that tensor-product distributions cannot have
higher Sobolev regularity than their factors. In particular, we cannot directly use tensor-product
distributions to prove that the Cartesian product E1 × E2 ⊂ Rn1+n2 is not (s, p)-null for any s >
min{n1/p, n2/p}. From the results in [22] (discussed briefly in Appendix A) we might conjecture that
the upper bound in (8) in the case m(E1× E2) > 0 can be improved to sE1×E2(p) ≤ sE1(p) + sE2(p).
Remark 2.10. The bounds in (8) do not in general allow sE1×E2(p) to be computed from sE1(p) and
sE2(p) (unless sE1(p)·sE2(p) = 0 and m(E1×E2) = 0). That sE1 and sE2 do not in general determine
sE1×E2 is shown by the following examples. If E1 = E2 = {0} ⊂ R, then sE1(p) = sE2(p) = −1/p′
and sE1×E2(p) = −2/p′ = sE1(p) + sE2(p) = s−(p), so the lower bound in (8) is achieved.
If
E1, E2 ⊂ R are Borel sets with Hausdorff dimension zero, for which E1×E2 has Hausdorff dimension
one (cf. e.g. Example 7.8 of [18]), then by Theorem 2.12 below, sE1×E2(p) = −1/p′ = sE1(p) =
sE2(p) = s+(p), so the upper bound in (8) is achieved.
We end this section with a simple application of (s, p)-nullity to function spaces on subsets of Rn.
On the maximal Sobolev regularity of distributions
8
Proposition 2.11. Let 1 < p < ∞, s ∈ R, and let F1, F2 be closed subsets of Rn. Then the following
statements are equivalent:
(i) The symmetric difference F1 ⊖ F2 is (s, p)-null.
(ii) F1 \ F2 and F2 \ F1 are both (s, p)-null.
(iii) H s,p
F1
= H s,p
F2
.
Proof. That (i) ⇔ (ii) follows from Lemma 2.3(i) and Proposition 2.7(i). To show that (iii) ⇒ (ii)
we argue by contrapositive. Suppose without loss of generality that F1 \ F2 is not (s, p)-null. Then
there exists 0 6= u ∈ H s,p such that supp u ⊂ F1 \ F2, so that u ∈ H s,p
but u 6∈ H s,p
. To show that
(ii) ⇒ (iii) we also argue by contrapositive. Suppose that H s,p
F1 6= H s,p
. Without loss of generality,
we assume that there exists u ∈ H s,p
. Let x ∈ supp u ∩ (F1 \ F2), and (by the closedness of
F2), let ε > 0 be such that Bε(x) ∩ F2 is empty. Then, for any φ ∈ D(Bε(x)) such that φ(x) 6= 0,
it holds (cf. the proof of Proposition 2.7(i)) that 0 6= φu ∈ H s,p with supp(φu) ⊂ F1 \ F2, which
implies that F1 \ F2 is not (s, p)-null.
F1 \ H s,p
F2
F2
F1
F2
We now present our main theoretical results concerning (s, p)-nullity. Since Lemma 2.3(v) tells
us that a set is (0, p)-null if and only if its inner Lebesgue measure is zero (independently of p), it
makes sense to consider the cases s < 0 and s > 0 separately.
2.3 The case s < 0 (sets with zero measure)
The following theorem provides a partial characterisation of (s, p)-nullity for −n/p′ ≤ s < 0 in
terms of Hausdorff dimension dimH (defined, e.g., in [18, §3] or [2, §5.1]). It will be proved at the
end of §3 using standard results from [2, §5] connecting Hausdorff dimension and capacity1. We
note that Theorem 2.12 applies as a special case to regular submanifolds of Rn, and also to the
"multi-screens", relevant for acoustic and electromagnetic scattering, considered e.g. in [15].
Theorem 2.12. Let 1 < p < ∞ and E ⊂ Rn.
(i) For −n/p′ < s ≤ 0, if dimHE < n + p′s, then E is (s, p)-null.
(ii) For −n/p′ ≤ s < 0, if E is Borel and (s, p)-null, then dimHE ≤ n + p′s.
In particular, if E is Borel and m(E) = 0, then
sE(p) =
dimH E − n
p′
and
dimHE = inf(cid:8)d : E is(cid:0)(d − n)/p′, p(cid:1)-null(cid:9).
(9)
Remark 2.13. A link between (s, p)-nullity and fractal dimension was established previously in [29].
Specifically, [29, Theorem 4] implies part (i) of Theorem 2.12 for compact sets, with dimH replaced
by dimB, the lower box (or Minkowski) dimension2. Since dimH(E) ≤ dimB(E) for all bounded
E ⊂ Rn [18, Proposition 3.4], our result in part (i) is stronger than what is provided by [29,
Theorem 4]. Examples of sets for which dimH(E) < dimB(E) are easy to find: a particularly simple
example is the set E = {0} ∪ {1/n : n ∈ N} ⊂ R, for which dimH(E) = 0 but dimB(E) = 1/2
(cf. [18, Example 2.7]).
1We remark that the results in [2, Chapter 5] actually allow a slightly more precise characterisation of (s, p)-
nullity in terms of Hausdorff measure. But we shall not pursue such characterisations here, since doing so would
add considerable complexity with little gain in insight (indeed, [2, §5.6.4] implies that even Hausdorff measure is
not sufficient to provide a complete characterisation of (s, p)-nullity). In any case the results in Theorem 2.12 seem
sufficient for the applications of scattering by fractal screens [7–10] that motivate the current study.
2The definition of lower box dimension in [29] differs from the standard definition found e.g. in [18, Equation (2.5)].
log(r) = inf{d ≥ 0 : lim inf r→0+ N (r)rd = 0} for
The two definitions can be reconciled by noting that lim inf r→0+ − log(N(r))
any function N : (0, ∞) → [1, ∞).
On the maximal Sobolev regularity of distributions
9
Related results can also be found in [39, Theorem 17.8], where a formula similar to (9) is stated
pq. However, Triebel's result concerns a different notion of nullity to
F in Definition 2.1 he has the space
in the context of the spaces Bs
ours-in place of the space H s,p
:= {u ∈ Bs
Bs,F
pq
pq : u(ψ) = 0 for all ψ ∈ S for which ψ = 0 on F}.
As Triebel points out in [39, p. 126], while Bs,F
have equality here. Since Bs
Theorem 2.12, but part (i) of Theorem 2.12 is stronger than what is provided by Triebel's result.
pq : supp u ⊂ F}, in general we do not
p2 = H s,p for q ≤ min{p, 2}, Triebel's result implies part (ii) of
pq ⊂ {u ∈ Bs
pq ⊂ F s
Theorem 2.12 (specifically (9)) provides a simple characterisation of the nullity threshold sE(p)
for a Borel set E with dimHE < n. But it tells us nothing about "threshold nullity", i.e. whether or
not E is (sE(p), p)-null. A general result concerning threshold nullity is given by the next proposition,
which follows from Proposition 3.10, Theorem 3.15(ii) and Remark 2.2.
Proposition 2.14. Let E ⊂ Rn be Borel with 0 ≤ dimH E < n, and let 1 < q < p < ∞ and
−∞ < s < t < 0 satisfy tq′ = sp′ = dimH E − n. If E is (s, p)-null, then E is (t, q)-null.
The following corollary provides a full classification of all possible nullity and nullity threshold
sets NE and TE (defined as in (6)–(7)) that can arise when m(E) = 0. It is a simple consequence of
Proposition 2.14, Theorem 2.12 and Lemma 2.3(v). This result makes clear that the "gap" between
parts (i) and (ii) of Theorem 2.12 cannot in general be bridged: no complete characterisation of
(s, p)-null sets for −n/p′ ≤ s < 0 in terms of Hausdorff dimension is possible. The sharpness of our
classification is demonstrated in Theorem 4.5, which provides the nullity and nullity threshold sets
for a range of Cantor sets (defined in Definition 4.3), for which the question of threshold nullity can
be answered completely using Theorem 4.4.
Corollary 2.15. Let E ⊂ Rn be Borel with m(E) = 0, and set d = dimH(E) ∈ [0, n]. If d = n then
NE =(cid:8)(s, p) : s ≥ 0, 1 < p < ∞(cid:9) and hence TE = (1,∞). Otherwise, if 0 ≤ d < n then
(cid:8)(s, p) : s > (d − n)/p′(cid:9) ⊂ NE ⊂(cid:8)(s, p) : s ≥ (d − n)/p′(cid:9),
TE ∈n∅, (1,∞)o, or TE ∈n(1, p∗), (1, p∗]o, for some 1 < p∗ < ∞.
Moreover, Theorem 4.5 shows that this result cannot be improved: for every nullity set N ⊂ R ×
(1,∞) allowed by (10), there exists a Cantor set E(n) ⊂ Rn for which NE(n) = N .
and either
(10)
We now consider two other classes of sets for which it is possible to answer completely the
question of threshold nullity.
First, when E consists of a single point, any distribution supported by E is necessarily a linear
combination of the delta function and its derivatives [32, Theorem 3.9]. In this case it follows from
(4) that sE(p) = −n/p′, and moreover that E is (−n/p′, p)-null. Proposition 2.7 implies that the
same holds for all countable sets. (We note however that countability is not a necessary condition
for (−n/p′, p)-nullity; a counterexample is provided by the Cantor set F (n)
Corollary 2.16. A non-empty countable set is (s, p)-null if and only if s ≥ −n/p′.
0,∞ in Theorem 4.5).
Second, recall (e.g. [39, §3]) that for 0 ≤ d ≤ n a closed set F ⊂ Rn with dimH(F ) = d is called
a d-set if there exist constants c1, c2 > 0 such that
0 < c1rd ≤ Hd(Br(x) ∩ F ) ≤ c2rd < ∞,
for all x ∈ F, 0 < r < 1,
(11)
where Hd is the d-dimensional Hausdorff measure on Rn. (Note that this definition differs from
that used in the fractal geometry literature, e.g. [18, p. 48].) Condition (11) may be understood as
saying that d-sets are everywhere locally d-dimensional. Note that the definition of d-set includes
as a special case all Lipschitz d-dimensional manifolds, d ∈ {0, 1, . . . , n} (cf. also 2.18(iv) below).
By combining Theorem 2.6 with results due to Triebel on the density of test functions in function
spaces [40, Theorems 3 and 5] one can prove the following result.
On the maximal Sobolev regularity of distributions
10
Theorem 2.17. Let 1 < p < ∞, and 0 < d < n. Let F ⊂ Rn be either a compact d-set, or a
d-dimensional hyperplane (in which case d is assumed to be an integer). Then F is ((d − n)/p′, p)-
null.
Our final theorem in this section applies the results of Theorem 2.12 to the special case where the
set E is the boundary of an open set Ω ⊂ Rn. Its proof makes use of Lemma 4.1 in §4, which collects
a number of results concerning the relationship between the analytical regularity of the boundary
of a set and its fractal dimension. The proof of part (iv) of the theorem is postponed until §3.
Here and in what follows we shall say that a non-empty open set Ω is C 0 (respectively C 0,α,
0 < α < 1, respectively Lipschitz) if its boundary ∂Ω can be locally represented as the graph
(suitably rotated) of a C 0 (respectively C 0,α, respectively Lipschitz) function from Rn−1 to R, with
Ω lying only on one side of ∂Ω. For a more precise definition see [20, 1.2.1.1]. We note that for n = 1
there is no distinction between these definitions: we interpret them all to mean that Ω is a countable
union of open intervals whose closures are disjoint. We also point out that in the literature several
alternative definitions of Lipschitz open sets can be found (for a detailed discussion see e.g. [19,20]);
in particular, our definition includes Stein's "minimally smooth domains" [36, §VI.3.3].
Theorem 2.18. Let 1 < p < ∞ and let Ω ⊂ Rn be non-empty and open.
(i) If Ωc has non-empty interior then ∂Ω is not (s, p)-null for s < −1/p′. (In particular this holds
if Ω 6= Rn is C 0.)
(ii) If Ω is C 0 and s ≥ 0, then ∂Ω is (s, p)-null.
(iii) If Ω is C 0,α for some 0 < α < 1 and s > −α/p′, then ∂Ω is (s, p)-null.
(iv) If Ω is Lipschitz then ∂Ω is (s, p)-null if and only if s ≥ −1/p′.
Proof. The case n = 1 is covered by Corollary 2.16, so assume n ≥ 2. For (i), Lemma 4.1(i) states
that dimH∂Ω ≥ n− 1 and then Theorem 2.12(ii) implies that ∂Ω is not (s, p)-null for any s < −1/p′.
(ii) follows from Lemma 2.3(v) and Lemma 4.1(ii). (iii) follows from Theorem 2.12(i) and Lemma
4.1(iii). (iv) is proved in §3.
In §4 we provide concrete examples to demonstrate the sharpness of these results. In particular,
Lemma 4.1 implies that for n ≥ 2 there exists a bounded C 0,α open set whose boundary is not
(s, p)-null for any s < −α/p′, and a bounded C 0 open set whose boundary is not (s, p)-null for any
s < 0.
2.4 The case s > 0 (sets with non-zero measure)
As has been discussed above, questions of (s, p)-nullity for s < 0 can often be answered by appealing
to Theorem 2.5 and applying standard potential theoretic results on the capacity Capt,p′ with t =
−s > 0. When it comes to investigating (s, p)-nullity for s > 0, however, Theorem 2.5 appears to be
of little use because the properties of Capt,p′ for t < 0 do not seem to have been widely documented.
Certainly, if a set E is to have nullity threshold in (0, n/p] it must have non-zero measure and empty
interior (cf. Lemma 2.3(iii),(v)). But we are not aware of any general characterisations of (s, p)-
nullity for s ∈ (0, n/p] in terms of the geometrical properties of a set, analogous to that provided by
Hausdorff dimension for nullity thresholds in [−n/p′, 0].
The following theorem provides an alternative analytic characterisation in terms of the capacity
caps,p defined in §3. It is taken from [2, Theorem 11.3.2], where it is stated with part (i) replaced
by an equivalent statement in terms of sets of uniqueness (cf. §3.1) and the assumption that F be
closed relaxed to F being Borel. It generalises an earlier result presented in [34, Theorem 2.6].
Theorem 2.19 ( [2, Theorem 11.3.2]). Let 1 < p < ∞ and 0 < s ≤ n/p. Let F be closed with
empty interior. Then the following are equivalent:
On the maximal Sobolev regularity of distributions
11
(i) F is (s, p)-null;
(ii) caps,p(Ω \ F ) = caps,p(Ω) for all open Ω ⊂ Rn;
(iii) caps,p(Bδ(x) \ F ) = caps,p(Bδ(x)) for all open balls Bδ(x) ⊂ Rn;
(iv) For almost all x ∈ Rn (with respect to Lebesgue measure)
caps,p(Bδ(x) \ F )
lim sup
δ→0
δn
> 0.
As is pointed out in [2, p. 314], given 1 < p < ∞ and s ∈ (0, n/p], this characterisation allows
us to construct compact sets K ⊂ Rn with positive measure and empty interior which are not
(s, p)-null, by engineering the failure of condition (ii) above. The approach suggested in [2, p. 314]
(described in more detail in §4.2 below), is based on a standard "Swiss cheese" construction. One
starts with a bounded open set Ω ⊂ R and removes from Ω a countable sequence of open balls of
diminishing radius in such a way that the remaining compact set K ⊂ Ω has empty interior and
satisfies caps,p(Ω \ K) < caps,p(Ω). As we will explain in §4.2.2 (see in particular Theorem 4.6),
sufficient conditions to ensure the latter bound can be obtained using the countable sub-additivity
of capacity (Proposition 3.7) and standard estimates on the capacity of balls (Proposition 3.16).
In §4.2.2 we apply this methodology to derive sufficient conditions for the non-(s, p)-nullity of
In §4.2.2 we also prove similar but
certain fat Cantor sets (defined in Examples 4.7 and 4.8).
complementary results using a completely different methodology not involving capacity, adapted
from [26], which is based on direct estimates of the Sobolev norm of the characteristic function of
the set for the case p = 2, obtained via explicit bounds on its Fourier transform.
However, these two approaches (i.e., proving upper bounds on capacities, or on Sobolev norms)
do not in general allow us to calculate the nullity threshold sK(p) for the compact set K under
consideration; they only provide a lower bound on sK(p) (by proving the existence of some s ∈
(0, n/p] for which K is not (s, p)-null). Since K is assumed to have empty interior, all we can
deduce is that sK ∈ [s, n/p]. Only in the extreme case s = n/p does such a non-nullity result
specify the nullity threshold exactly. The existence of a compact set Kp with empty interior which
is not (n/p, p)-null appears to have been proved first by Polking in [35, Theorem 4]. Polking's
set Kp is a "Swiss cheese" set of the kind described above, but Polking's analysis does not make
use of the capacity-theoretic characterisation of Theorem 2.19; instead Polking provides an explicit
construction of a non-zero function fp ∈ H n/p,p
, appealing to Leibniz-type formulae for fractional
derivatives in order to prove that kfpkH n/p,p < ∞. As Polking remarks in [35], this result "illustrates
in a rather striking manner that the Sobolev embedding theorem is sharp".
Theorem 2.20 ( [35, Theorem 4], [2, p. 314]). Let 1 < p < ∞. There exists a compact set Kp ⊂ Rn
with empty interior which is not (n/p, p)-null. In particular, sKp(p) = n/p.
Kp
The set Kp whose existence is guaranteed by Theorem 2.20 is, at least a priori, p-dependent.
(Certainly the constructions in [35, Theorem 4] and [2, p. 314] are intrinsically p-dependent.) By
taking the closure of the union of a countable sequence of such sets Kp, suitably scaled and translated,
one can construct a compact set with empty interior that is not (n/p, p)-null for any 1 < p < ∞.
Corollary 2.21. There exists a compact set K ⊂ Rn with empty interior which is not (n/p, p)-null
for any 1 < p < ∞. In particular, sK(p) = n/p for every 1 < p < ∞.
Proof. For each j ∈ N let pj = 1 + 1/j, xj = (21−j, 0, . . . , 0) and let Kpj ⊂ [0, 1]n be a compact set
with empty interior which is not (n/pj, pj)-null. Then K = {0} ∪S∞j=1(xj + 2−jKpj ) is compact,
has empty interior, and is not (n/p, p)-null for any 1 < p < ∞.
the nullity threshold of certain non-compact sets, for example Rn \ Qn.
The following related result is a by-product of the arguments leading to Theorem 2.20, and gives
On the maximal Sobolev regularity of distributions
12
Theorem 2.22. Let A ⊂ Rn have non-empty interior and Q ⊂ A be countable and dense in the
interior of A. Then E := A \ Q is not (n/p, p)-null, and hence sE(p) = n/p, for all 1 < p < ∞.
Proof. Given A and Q as in the assertion and 1 < p < ∞, the Swiss-cheese construction of [35,
Theorem 4], applied to any non-empty bounded open subset of A, gives a set Kp ⊂ A \ Q that is
not (n/p, p)-null. Then A \ Q is not (n/p, p)-null for any p and has empty interior, so the theorem
follows.
Proving the existence of sets whose nullity threshold lies in the open interval (0, n/p) appears to
be an open problem-certainly we are not aware of any literature on this matter. The difficulty
here is that one would need to prove that a set with positive measure is (s, p)-null for some s ∈
(0, n/p). The possibility of doing this for a closed set F using the capacity-theoretic characterisation
of Theorem 2.19 seems remote. Using the conditions (ii) or (iii) from that theorem would require us
to prove equality of two capacities, which is difficult because capacity can usually only be estimated
rather than computed exactly. Condition (iv) could in principle be verified by proving a sufficiently
sharp lower bound on caps,p(Bδ(x) \ F ), but even this seems difficult in general as lower bounds
for caps,p appear to be available only for balls (by contrast, quite general upper bounds can be
obtained using countable sub-additivity, as we have mentioned above). We also remark that the
Fourier approach adopted in Proposition 4.10 is not promising in this regard: lower bounds for u
might be used to show that a particular u is not in a certain H s,p
F , but this would not rule out the
existence of other non-trivial functions in H s,p
F .
3 Capacity
In this section we provide proofs of Proposition 2.7(ii), Theorem 2.12, and Theorem 2.18(iv). Our
arguments rely on Theorem 2.5, which characterises s-null sets in terms of a set function called
capacity. The notion of capacity is central to potential theory, and we briefly review some of the
basic ideas here. Our presentation is based broadly on [2, 30], but other relevant references include
[1, 3, 6, 14, 28, 33, 34, 41]. We begin with a rather general definition of capacity before specialising to
the particular capacities of relevance to the problem in hand. Since (i) the literature is in places
highly technical; (ii) notational conventions are varied and sometimes conflicting (see the discussion
in Remark 3.2 below); and (iii) the concept of capacity may not be familiar to some readers, we
take care to clarify certain details that are not fully explained in [2, 30].
Definition 3.1. Let C comp be a set function defined on compact subsets of Rn, taking values in
[0,∞], such that C comp(∅) = 0 and C comp(K1) ≤ C comp(K2) for all compact K1 ⊂ K2 ⊂ Rn. From
C comp we define inner and outer capacities on arbitrary subsets E ⊂ Rn by
C(E) := sup
K⊂E
K compact
C comp(K),
C(E) := inf
Ω⊃E
Ω open
C(Ω).
Clearly C(E) ≤ C(E) for all E ⊂ Rn. If C(E) = C(E) then we say E is capacitable for C comp and
define the capacity of E to be C(E) := C(E) = C(E).
It follows straight from the definitions that open sets are capacitable, and that C(K) = C comp(K)
for all compact K ⊂ Rn. It is common practice [2,30] to denote the original set function from which
C and C(E) are defined simply by C, rather than C comp. No ambiguity arises from this abuse
of notation provided that compact sets are capacitable; that this is the case for the capacities of
interest to us will be demonstrated in Proposition 3.4 below.
We now define two particular capacities of relevance for the study of (s, p)-nullity. For 1 < p < ∞,
s ∈ R and K ⊂ Rn compact, we define
(K) := inf{kukp
(K) := inf{kukp
capcomp
Capcomp
s,p
s,p
H s,p : u ∈ D and u ≥ 1 in a neighbourhood of K},
H s,p : u ∈ D and u = 1 in a neighbourhood of K}.
(12)
(13)
On the maximal Sobolev regularity of distributions
13
Clearly capcomp
for s > 0, the reverse inequality also holds, up to a constant factor-see Theorem 3.13 below.
(K) for all compact sets K. It is a much deeper fact that, at least
(K) ≤ Capcomp
s,p
s,p
Remark 3.2. The capacities caps,p and Caps,p arising from (12) and (13) are classical and appear
throughout the potential theory literature. Our notation is adapted from [30, §10.4 and §13.1], where
p(Rn)). But many other
caps,p(·) and Caps,p(·) are respectively denoted cap(·, H s
conflicting notational conventions exist: for instance, {capcomp
s,p } are respectively denoted
{Cs,p, Ns,p} in [2, §2.2, §2.7], {Ns,p, Ms,p} in [28], and {Bs,p, Cs,p} in [3, 34]; capcomp
is denoted
Bs;p in [33], and B(n)
p(Rn)) and Cap(·, H s
, Capcomp
s,p in [1].
s,p
s,p
s,p
s,p
s,p
(K) ≤ capcomp
to prove. Let gcapcomp
be compact. Obviously gcapcomp
Navigating the literature is also complicated by the fact that caps,p and Caps,p can be defined
in a number of equivalent ways. Firstly, Definitions (12)–(13) are sometimes stated with the trial
functions u ranging over S (as in [2, pp. 19–20]) rather than D (as in [30, §13.1]). That these two
choices of trial space lead to the same set functions (and hence the same capacities) is straightforward
denote the set function defined by (12) using S instead of D, and let K ⊂ Rn
(K); for a bound in the opposite direction, consider
any u ∈ S with u ≥ 1 on an open neighbourhood Ω of K. Take R > 0 such that Ω ⊂ BR(0), and take
a cutoff χ ∈ D such that χ = 1 in BR(0). Then w := (1 − χ)u ∈ S, with support in the complement
O := Rn\ BR(0), so that given ε > 0 there exists ψε ∈ D(O) such that kw− ψεkH s,p < ε. This bound
implies that ηε := χu+ψε ∈ D satisfies ku−ηεkH s,p < ε, so that kηεkH s,p < kukH s,p +ε; note also that
(K),
(K), as claimed. The analogous result for Caps,p follows by a
similar argument with "≥ 1" replaced by "= 1" throughout.
Secondly, Definition (12) is sometimes stated (e.g. [2, §2.2] and [30, §13.1]) with "on K" instead
of "in a neighbourhood of K". Again it is easy to verify that the two definitions are equivalent. Let
denote the set function defined by (12) using "on K" instead of "in a neighbourhood of
(K) ≤ capcomp
(K); for a bound in the opposite direction, note that,
given α ∈ (0, 1), if u ≥ 1 on K then there exists a neighbourhood of K on which u ≥ α. Hence
capcomp
(K), and since this holds for α arbitrarily close to 1, we conclude that
ηε = χu+0 ≥ 1 on Ω. Since u and ε > 0 were arbitrary we conclude thatgcapcomp
and hence that gcapcomp
dcapcomp
K". Then clearly dcapcomp
(K) ≤ α−pdcapcomp
dcapcomp
in which the right-hand-side of (12)
is replaced by an infimum of kfkp
Lp over the non-negative f ∈ Lp for which J−sf ≥ 1 on K (cf.
e.g. [34, Definition 2.1]). That this definition is equivalent to (12) is proved in [2, Proposition
2.3.13].
s,p
Finally, for s > 0 some authors use a definition of capcomp
s,p
(K), as claimed.
(K) ≥ capcomp
(K) = capcomp
(K) = capcomp
s,p
s,p
s,p
s,p
s,p
s,p
s,p
s,p
s,p
s,p
Remark 3.3. For s = 0 one can show using standard measure-theoretic techniques that the capaci-
ties cap0,p and Cap0,p both coincide with the Lebesgue measure on Rn. Specifically, for E ⊂ Rn let
m(E) = sup{m(K) : E ⊃ K, K compact} and m(E) = inf{m(Ω) : E ⊂ Ω, Ω open} be the usual
inner and outer Lebesgue measures of E [5, Definitions 2.2–2.3]. Then for every 1 < p < ∞ it
holds that cap
(E) = m(E) and cap0,p(E) = Cap0,p(E) = m(E). Hence E is capac-
0,p
itable for capcomp
) if and only if E is Lebesgue measurable, in which case
cap0,p(E) = Cap0,p(E) = m(E), where m is the Lebesgue measure.
(equivalently for Capcomp
(E) = Cap
0,p
0,p
0,p
As promised, we now prove the capacitability of compact sets for capcomp
s,p
and Capcomp
s,p
.
s,p
and Capcomp
Proposition 3.4. Compact sets are capacitable for both capcomp
and s ∈ R.
Proof. The result for capcomp
same proof is in fact valid for all s ∈ R. To prove the result for Capcomp
given ε > 0, let u ∈ D satisfy u = 1 in a neighbourhood Ω of K, with kukp
Hence Capcomp
Caps,p(Ω) = Cap
conclude that K is capacitable for Capcomp
is stated and proved in [2, Proposition 2.2.3] for integer s > 0, but the
, let K ⊂ Rn be compact and,
(K) + ε.
(K) + ε for all compact K ⊂ Ω, which implies that
(K) + ε. Since ε was arbitrary, we
(K) ≤ Caps,p(K) ≤ Caps,p(Ω) ≤ Capcomp
, for all 1 < p < ∞
( K) ≤ kukp
H s,p < Capcomp
H s,p < Capcomp
s,p
s,p
s,p
s,p
s,p
s,p
s,p
s,p
.
s,p
On the maximal Sobolev regularity of distributions
14
The link between capacity and nullity was stated in Theorem 2.5 above: to repeat, a set E is
−s,p′(E) = 0. However, relatively little seems to be known about the
(s, p)-null if and only if Cap
capacity Caps,p; the capacity caps,p appears to be much better understood. In particular, there is a
class of capacities, known as "Choquet capacities" (cf. [2, Theorem 2.3.11] and Choquet's original
work [14]) for which all Suslin sets (defined in [2, §2.9, Notes to §2.3]), in particular all Borel sets,
are capacitable. The capacity caps,p is well-known to be of this class [2, §2.3] (and see also [33]), at
least for s ≥ 0. But it has been suggested that the same is probably not true of Caps,p [34, p. 1236]
(although of course it is true for Cap0,p, cf. Remark 3.3).
Proposition 3.5 ( [2, Propositions 2.3.12 and 2.3.13]). Borel sets are capacitable for capcomp
1 < p < ∞ and s ≥ 0.
Remark 3.6. In Theorem 3.15(ii), Proposition 2.7(ii), Proposition 2.9, Theorem 2.12(ii), Proposi-
tion 2.14, Theorem 4.5 and Proposition A.3 we require certain sets to be Borel. This assumption is
made solely to allow application of Proposition 3.5. Hence, if desired, throughout the paper "Borel"
may be substituted by "Suslin", or possibly by a more general class of capacitable sets.
for
s,p
The outer capacity caps,p is also known to be countably subadditive for s ≥ 0 [2, §2.3]. The
authors are not aware of a similar result for Caps,p, but the example in Remark 2.8, together with
Theorem 2.5, shows that Cap
is not subadditive (not even finitely) for s < −n/p′.
s,p
Proposition 3.7 ( [2, Propositions 2.3.6 and 2.3.13]). Let 1 < p < ∞, s ≥ 0 and let Ei ⊂ Rn,
i ∈ N. Then
caps,p ∞[i=1
Ei! ≤
∞Xi=1
caps,p(Ei).
The link between the analytical concept of capacity and the geometrical concept of fractal
dimension is provided by the following theorem, which provides a partial characterisation of the sets
of zero outer capacity caps,p(E) for 0 < s ≤ n/p in terms of Hausdorff dimension. The theorem,
which we state without proof, is essentially a rephrasing of the results in [2, §5.1] (specifically
Theorems 5.1.9 and 5.1.13). Similar results can be found e.g. in [30, §10.4.3], [33, §8] and [41,
Theorem 2.6.16]. For a historical background to these results the reader is referred to [2, §5.7].
Theorem 3.8 ( [2, Theorems 5.1.9 and 5.1.13]). Let 1 < p < ∞ and E ⊂ Rn.
(i) For 0 ≤ s < n/p, if dimH(E) < n − ps then caps,p(E) = 0.
(ii) For 0 ≤ s ≤ n/p, if caps,p(E) = 0 then dimH(E) ≤ n − ps.
The behaviour of caps,p under Lipschitz mappings is also understood [2, §5.2].
Theorem 3.9 ( [2, Theorem 5.2.1]). Let E ⊂ Rn, and let Φ : E → Rn be a Lipschitz map with
Lipschitz constant L. Then for 1 < p < ∞ and 0 ≤ s ≤ n/p there exists a constant a > 0, depending
only on n, p, s and L, such that
A further useful result on caps,p is the following.
caps,p (Φ(E)) ≤ a caps,p(E).
Proposition 3.10 ( [2, Theorem 5.5.1]). Let E ⊂ Rn be bounded, and let s, t ∈ R and 1 < p, q < ∞
be such that either 0 < tq < sp ≤ n or p < q and 0 < tq = sp ≤ n. Then, caps,p(E) = 0 implies that
capt,q(E) = 0.
On the maximal Sobolev regularity of distributions
15
Note that [2, Theorem 5.5.1] requires E to have diameter at most one, as that theorem deals
with the actual values of the capacities. Since here we are only concerned with the vanishing of the
same capacities, by affine scaling the result holds for any bounded set.
To prove Proposition 2.7(ii), Theorem 2.12, and Theorem 2.18(iv), which was the goal of this
section, we have to link the concept of nullity (which by Theorem 2.5 concerns Caps,p) with the
results of Proposition 3.5 and Theorem 3.9 (which concern caps,p). The link, as was hinted at just
before Remark 3.2, is that the two capacities caps,p and Caps,p are equivalent, at least for s ≥ 0, in
the sense of the following definition.
Definition 3.11. Let C comp and C comp be set functions satisfying Definition 3.1. The resulting
capacities C and C are said to be equivalent if there exist constants a, b > 0 such that
aC comp(K) ≤ C comp(K) ≤ bC comp(K),
for all compact K ⊂ Rn.
Proposition 3.12. If two capacities C and C are equivalent then, for any E ⊂ Rn,
aC(E) ≤ C(E) ≤ bC(E),
where a, b are the constants in Definition 3.11.
C(E) = 0 if and only if E is capacitable for C comp with C(E) = 0.
aC(E) ≤ C(E) ≤ bC(E),
In particular, E is capacitable for C comp with
Theorem 3.13 ( [2, eq. (2.7.4), Corollary 3.3.4, and Notes to §2.9 and §3.8]). For every 1 < p < ∞
and s ≥ 0 the capacities caps,p and Caps,p are equivalent. Specifically, for any s ≥ 0 there exists
b ≥ 1 such that, for all compact K,
capcomp
(K).
(K) ≤ b capcomp
s,p
s,p
(K) ≤ Capcomp
s,p
Remark 3.14. It is noted in [2, Notes to §2.7]) that results due to Deny [16, Th´eor`eme II:3, p. 144]
imply that for p = 2 and 0 < s ≤ 1, the constant b in (3.13) can be taken to be 1, so that caps,p and
Caps,p coincide (we have already noted this result for s = 0 in Remark 3.3). An interesting open
question concerns the extent to which this result generalises to p 6= 2 and/or s 6∈ [0, 1]. In Appendix
B we demonstrate that the result is certainly not true for p = 2 and s = 2, using an explicit formula
for the norm in the restriction space H 2,2(Ω) recently presented in [11].
Theorem 2.5, Proposition 3.5, Proposition 3.12 and Theorem 3.13 then provide the following key
result, which allows us to complete the proofs of the remaining results stated in §2.
Theorem 3.15. Let 1 < p < ∞, s ≤ 0 and E ⊂ Rn.
(i) If cap−s,p′(E) = 0 then E is (s, p)-null.
(ii) If E is Borel, then cap−s,p′(E) = cap−s,p′(E) = 0 if and only if E is (s, p)-null.
Proof of Proposition 2.7(ii). This follows immediately from Theorem 3.15 and Proposition 3.7.
Proof of Theorem 2.12. Part (i) follows from Theorem 3.15(i) and Theorem 3.8(i). Part (ii) follows
from Theorem 3.15(ii) and Theorem 3.8(ii).
Proof of Theorem 2.18(iv). By applying a suitable smooth cutoff and a coordinate rotation, it suf-
fices to consider the case where Ω = {x ∈ Rn : xn < φ(x1, . . . , xn−1)}, where φ : Rn−1 → R is
Lipschitz. Defining Rn
0 ⊂ Rn → ∂Ω ⊂ Rn is Lipschitz with
a Lipschitz inverse (given by the orthogonal projection of ∂Ω onto Rn
0 ). Hence by Theorem 3.9 and
Theorem 3.15(ii), the closed set ∂Ω is (s, p)-null if and only if the hyperplane Rn
0 is (s, p)-null, which
by Theorem 2.17 holds if and only if s ≥ −1/p′.
Capacities can rarely be computed exactly.
0 := {x ∈ Rn : xn = 0}, the map Φ : Rn
(An exception is provided by Appendix B.) But
estimates are available for the capacity of balls, which will be of use to us in §4.2.2.
On the maximal Sobolev regularity of distributions
16
Proposition 3.16 ( [2, Propositions 5.1.2–4]). Let 1 < p < ∞. Given 0 < s < n/p, there exist
constants 0 < As,p,n < Bs,p,n, depending on s, p and n, such that
As,p,nrn−sp ≤ caps,p(cid:0)Br(x)(cid:1) ≤ Bs,p,nrn−sp,
0 < r ≤ 1, x ∈ Rn.
For s = n/p, given c > 1 there exists a constant Cc,p,n > 1, depending on c, p and n, such that
1
Cc,p,n(cid:0) log(c/r)(cid:1)1−p ≤ capn/p,p(cid:0)Br(x)(cid:1) ≤ Cc,p,n(cid:0) log (c/r)(cid:1)1−p,
0 < r ≤ 1, x ∈ Rn.
3.1 Sets of uniqueness and (s, p)-nullity
We end this section by exploring the relationship between the concept of (s, p)-nullity and the
concept of sets of uniqueness considered in [2, 30]. Following [2, Definition 11.3.1] and [30, p. 692],
given 1 < p < ∞ and s > 0 we say that E ⊂ Rn is a (s, p)-set of uniqueness (abbreviated to
(s, p)-SOU) if
Note that if E is Borel then by Theorem 3.15(ii) this definition can be restated as
{u ∈ H s,p : caps,p(supp u ∩ Ec) = 0} = {0}.
{u ∈ H s,p : (supp u ∩ Ec) is (−s, p′)-null} = {0}.
Proposition 3.17. Let 1 < p < ∞ and s > 0.
(i) If E is a (s, p)-SOU then E is (s, p)-null.
(ii) If E is closed, then E is a (s, p)-SOU if and only if E is (s, p)-null.
Proof. (i) Suppose that u ∈ H s,p with supp u ⊂ E. Then caps,p(supp u ∩ Ec) = caps,p(∅) = 0, and
since E is a (s, p)-SOU it follows that u = 0. Hence E is (s, p)-null.
(ii) Suppose that u ∈ H s,p with caps,p(supp u ∩ Ec) = 0. Then m(supp u ∩ Ec) = 0 (this holds
e.g. by Theorem 3.8(ii), which gives dimH(supp u ∩ Ec) < n), hence supp u ∩ Ec is (s, p)-null. Since
E is closed and (s, p)-null, supp u ∩ E is also closed and (s, p)-null. Then Proposition 2.7(i) gives
that supp u is (s, p)-null, which implies that u = 0. Hence E is a (s, p)-SOU.
4 Examples and counterexamples
In this section we present examples and counterexamples to illustrate the results of §2.
4.1 Boundary regularity and Hausdorff dimension
The following lemma concerns the relationship between the analytical regularity of the boundary
of a set and its Hausdorff dimension. Its proof shows how to construct examples of C 0 open sets
whose boundaries have a given Hausdorff dimension, using the modified Weierstrass-type functions
analysed in [39, Theorem 16.2]. These results should be considered in the context of Theorem 2.18
above.
Lemma 4.1. Let Ω ⊂ Rn be an open set such that Ωc has non-empty interior. Then:
(i) n − 1 ≤ dimH(∂Ω) ≤ n.
(ii) If Ω is C 0 then m(∂Ω) = 0.
(iii) If Ω is C 0,α with 0 < α < 1, then n − 1 ≤ dimH(∂Ω) ≤ n − α.
(iv) If Ω is Lipschitz, then dimH(∂Ω) = n − 1.
On the maximal Sobolev regularity of distributions
17
(v) For n ≥ 2 and 0 < α < 1, there exists Ωα,n ⊂ Rn open, bounded and C 0,α such that
dimH(∂Ωα,n) = n − α.
(vi) For n ≥ 2, there exists Ω0,n ⊂ Rn open, bounded and C 0 such that dimH(∂Ω0,n) = n.
Proof. If n = 1, there is no distinction between C 0, Holder and Lipschitz open sets: they all mean
a countable union of open intervals with pairwise disjoint closures. Hence ∂Ω contains at most
countably many points, so it always has dimension 0 = n − 1. So we assume henceforth that n ≥ 2.
(i) The upper bound dimH(∂Ω) ≤ n is trivial. For the lower bound, since Ω is open and its
complement Ωc has non-empty interior we can take two disjoint balls of some radius ǫ > 0 such that
Bǫ(x) ⊂ Ω and Bǫ(y) ⊂ Ωc. After translation and rotation, without loss of generality we can assume
x = 0 and y = (y, 0, . . . , 0). For allez ∈ Rn−1 with ez < ǫ, the point (0,ez) lies in Ω and the point
(y,ez) lies in the interior of Ωc, so the segment [(0,ez), (y,ez)] contains at least one point in ∂Ω. The
the set {ez ∈ Rn−1 : ez < ǫ} which has Hausdorff dimension equal to n−1. Thus by [18, Corollary 2.4]
orthogonal projection P1 : Rn → Rn−1 defined by x 7→ (x2, . . . , xn) is Lipschitz and P1(∂Ω) contains
we have dimH(∂Ω) ≥ dimH(P1(∂Ω)) ≥ n − 1, as required. (ii) The graph of a continuous function
has zero Lebesgue measure (this follows from the translation invariance of Lebesgue measure-just
consider the measure of the union of infinitely many disjoint vertical translates of the graph). Since
the boundary of a C 0 open set can be covered by a countable number of C 0 graphs, the assertion
follows from the countable subadditivity of Lebesgue measure. (iii) If ∂Ω is the graph of a C 0,α
function, this follows from [39, Theorem 16.2(i)], which states that, for 0 < α < 1, the Hausdorff
dimension of the graph of a function in C 0,α([−1, 1]n−1) is at most n − α. The general case follows
for countably many
from the "countable stability" of the Hausdorff dimension (see [18, p. 49]):
subsets {Aj}j∈N of Rn we have dimHSj Aj = supj dimHFj (note that the open cover of ∂Ω given
by the definition of a C 0,α open set in [20, 1.2.1.1] allows a countable subcover, this can be seen
using the compactness of ∂Ω ∩ BR(0) for R > 0 and considering a countable number of balls, e.g.
{Bℓ(0)}ℓ∈N). (iv) Since a Lipschitz open set is C 0,α for every 0 < α < 1, the Hausdorff dimension of
its boundary satisfies n−1 ≤ dimH(∂Ω) ≤ n−α for all these values, from which the assertion follows.
(v) This follows from [39, Theorem 16.2(ii)] which provides a function fα,n : [−1, 1]n−1 → [−1, 1]
of class C 0,α (defined via a modification of the well-known Weierstrass function), whose graph
has Hausdorff dimension exactly equal to n − α (that fα,n can be taken with values in [−1, 1]
follows because affine transformations do not affect the Hausdorff dimension). We can immediately
define Ωα,n := {x ∈ (−1, 1)n−1 × R, −2 < xn < fα,n(x1, . . . , xn−1)}. (vi) We construct Ω0,n by
gluing together the functions from the previous step. We define f0,n : [−1, 1]n−1 → [−1, 1] as
f0,n(x1, . . . , xn−1) := f1/j,n(x1, . . . , 2j xn−1 − 1) if 2−j < xn−1 ≤ 2−j+1 and j ∈ N. To ensure global
continuity we assume that each f1/j,n vanishes at xn−1 = ±1 (if necessary we can ensure this by
multiplying f1/j,n by a suitable element of D((−1, 1))). We then construct the corresponding open
set as before. Its boundary has Hausdorff dimension not smaller than n − 1/j for all j ∈ N, from
which the assertion follows.
4.2 Swiss-cheese and Cantor sets
Our remaining examples belong to two classes of compact sets with empty interior: "Swiss-cheese"
sets and Cantor sets. When n = 1, Cantor sets are special cases of Swiss-cheese sets.
i ∈ N.
Definition 4.2. A Swiss-cheese set is a non-empty compact set K with empty interior, constructed
as K = Ω \ (S∞i=1 Bri(xi)), where Ω ⊂ Rn is a bounded open set, and xi ∈ Rn and ri > 0 for each
Given 1 < p < ∞ and s ∈ [−n/p′, n/p], one would like to engineer the (s, p)-nullity (or otherwise)
of K by choosing an appropriate sequence of ball centres {xi}∞i=1 and a sequence of radii {ri}∞i=1
which tends to zero at an appropriate rate as i → ∞.
In one dimension (n = 1), a well-studied family of examples of this construction is the classical
ternary Cantor set and its generalisations. These are Swiss-cheese sets for which for every i ∈ N
On the maximal Sobolev regularity of distributions
18
the centre xi of the ith subtracted ball is chosen as the centre of one of the largest connected
i′=1 Bri′ (xi′)). Following [2, §5.3], we consider Cantor sets in Rn (sometimes
components of Ω \ (Si−1
known as "Cantor dust" when n ≥ 2) defined in the following general way.
Definition 4.3. Let {lj}∞j=0 be a decreasing sequence of positive numbers such that 0 < lj+1 < lj/2
for j ≥ 0. Let E0 = [0, l0] and for j ≥ 0 let Ej+1 be constructed from Ej by removing an open interval
of length lj − 2lj+1 from the middle of each of the 2j subintervals making up Ej, each of which has
length lj. The Cantor set associated with the sequence {lj}∞j=1 is then defined to be E :=T∞j=0 Ej.
For n ≥ 1 we define the n-dimensional Cantor set E(n) ⊂ Rn to be the Cartesian product of n copies
of E.
The total measure of Ej is 2jlj, so the measure of E is limj→∞ 2jlj ∈ [0, l0). Since each Ej is
compact, E is also compact. Moreover, E has empty interior and is uncountable. Without loss of
generality we shall always take l0 = 1, so that E ⊂ [0, 1] and E(n) ⊂ [0, 1]n.
4.2.1 The case s < 0 (sets with zero measure)
For the case −n/p′ ≤ s < 0 the (s, p)-nullity of a Cantor set E(n) can be characterised precisely in
terms of the asymptotic behaviour of the sequence {lj}∞j=1. Theorem 5.3.2 in [2], which was first
proved in [31, Theorem 7.4] (see also [30, §10.4.3, Proposition 5]), provides necessary and sufficient
conditions for caps,p(E(n)) to vanish, for a given 0 < s ≤ n/p. Reinterpreting this result in terms of
(s, p)-nullity, using Theorem 3.15(ii), provides the following result.
Theorem 4.4. Let 1 < p < ∞ and −n/p′ ≤ s < 0. The Cantor set E(n) is (s, p)-null if and only if
(cid:17)p−1
= ∞,
for − n/p′ < s < 0,
j
∞Xj=0(cid:16)2−jnl−(sp′+n)
∞Xj=0
and if and only if
2−jn(p−1) log 1/lj = ∞,
for s = −n/p′.
Using Theorem 4.4 we can construct a zoo of Cantor sets, all with zero measure, which realise
all of the possible nullity sets permitted by Corollary 2.15.
Theorem 4.5. Let n ∈ N, 0 ≤ d < n and 1 < p∗ < ∞. Then there exists two Cantor sets
E(n)
d,p∗
= dimH F (n)
d,p∗
= d and
d,p∗
, F (n)
d,p∗ ⊂ Rn such that dimHE(n)
E(n)
d,p∗
F (n)
d,p∗
is (s, p)-null if and only if either s > (d − n)/p′,
is (s, p)-null if and only if either s > (d − n)/p′,
Furthermore, there exist F (n)
d,∞
n, and
, F (n)
d,1 , F (n)
n,∞ ⊂ Rn such that dimHF (n)
d,∞
or
or
s = (d − n)/p′ and p ≤ p∗,
s = (d − n)/p′ and p < p∗.
= dimH F (n)
d,1 = d, dimH F (n)
n,∞ =
F (n)
is (s, p)-null if and only if s ≥ (d − n)/p′,
d,∞
F (n)
d,1 is (s, p)-null if and only if s > (d − n)/p′,
0 ≤ d ≤ n,
0 ≤ d < n.
Moreover, for any Borel set E ⊂ Rn with m(E) = 0, (exactly) one of the n-dimensional Cantor sets
above is (s, p)-null precisely for the same values s, p for which E is (s, p)-null.
Proof. To construct each Cantor set, setting l0 = 1 as usual, we just need to find a sequence {lj}∞j=1
with 0 < l1 < 1/2 and 0 < lj+1 < lj/2 for j ≥ 1, such that the series in Theorem 4.4 diverge for the
desired set of values of s and p. The 9 different examples we require are defined in Table 1.
The convergence of the corresponding series can be verified using the fact that, for x, y > 0,
the geometric series Pj≥0 xj < ∞ ⇔ x < 1; its generalisation Pj≥0 xjy√j < ∞ ⇔ x < 1 or
On the maximal Sobolev regularity of distributions
19
Table 1: Cantor sets used in the proof of Theorem 4.5.
For
d = 0
d = 0
choosing l0 = 1 and, for j ≥ 1,
and
1 < p∗ < ∞ lj = 2−2(2jn(p∗ −1)−1)/(2n(p∗ −1)−1)
0 < d < n 1 < p∗ < ∞ lj = 2−jn/d(cid:0)1 + (j−1)
2−j2−1
d = 0
2
(2(n−d)(p∗−1) − 1)(cid:1)
1
d(p∗−1) ,
p∗ = 1
1 < p∗ < ∞ lj = 2−2(j+j0 )n(p∗−1)/(j+j0)2
d = 0
p∗ = ∞
0 < d < n p∗ = 1
j0 = min{j ∈ N s.t. 2(j+1)n(p∗ −1)
lj = 2−22j
lj = 2−jn/d 2(n/d−1)√j/2
(j+1)2
− 2jn(p∗ −1)
j2
> 1}
0 < d < n 1 < p∗ < ∞ lj = 2−jn/d(cid:0)(j + j0) log2(j + j0)(cid:1)
j0 = min(cid:8)2 ≤ j ∈ N s.t. (j+1) log2(j+1)
j log2 j
1
d(p∗−1)
and 2−jn(p∗−1)(j + 1) log2(j + 1) < 2(n−d)(p∗−1)(cid:9)
< 2(n−d)(p∗−1)
0 < d < n p∗ = ∞
p∗ = ∞
d = n
lj = 2−jn/d
lj = 2−j/(j + 1)
gives
E(n)
0,p∗
E(n)
d,p∗
F (n)
0,1
F (n)
0,p∗
F (n)
0,∞
F (n)
d,1
F (n)
d,p∗
F (n)
d,∞
F (n)
n,∞
j = 0. The final statement in the assertion follows from Corollary 2.15.
x = 1 and y < 0; the generalised harmonic seriesPj≥1 j−x < ∞ ⇔ x > 1; andPj≥2 j−x log−y j <
∞ ⇔ x > 1 or x = 1 and y > 1. To prove that F (n)
m(F (n)
n,∞) = limj→∞ 2njln
A few remarks are in order here. First, the standard one-third Cantor set with lj = 1/3j
corresponds to F (1)
0,∞ demon-
strates that Corollary 2.16 is not sharp: there exist uncountable sets which are (−n/p′, p)-null for all
1 < p < ∞. Third, Theorem 2.17 implies that none of the Cantor sets constructed in Theorem 4.5
are d-sets, except for F (n)
d,∞
in the table in the proof of Theorem 4.5. Second, the set F (n)
n,∞ is (0, p)-null, we compute the measure
, 0 ≤ d ≤ n.
log 2/ log 3,∞
4.2.2 The case s > 0 (sets with non-zero measure)
For a general Swiss-cheese set defined as in Definition 4.2 we can give sufficient conditions for
non-(s, p)-nullity, for 0 < s ≤ n/p, by combining Theorem 2.19 with capacity estimates from §3.
Theorem 4.6. Let K = Ω \ (S∞i=1 Bri(xi)) be a Swiss-cheese set defined as in Definition 4.2.
Suppose that 0 < ri ≤ 1 for each i ∈ N, and let 0 < r ≤ 1 be such that Br(x) ⊂ Ω for some x ∈ Ω.
Given 1 < p < ∞ and 0 < s ≤ n/p, let 0 < As,p,n ≤ Bs,p,n, c > 1 and Cc,p,n > 1 denote the
constants from Proposition 3.16. Then K is not (s, p)-null provided that
rn−sp
i
<
As,p,n
Bs,p,n
rn−sp,
for 0 < s < n/p,
(14)
∞Xi=1
and
,
for s = n/p.
(15)
Proof. By countable subadditivity (Proposition 3.7), we have that
c,p,n
C 2
∞Xi=1(cid:0) log(c/ri)(cid:1)1−p < (cid:0) log(c/r)(cid:1)1−p
caps,p(Br(x) \ K) ≤ caps,p ∞[i=1
Bri(xi)! ≤
∞Xi=1
caps,p(cid:0)Bri(xi)(cid:1).
On the maximal Sobolev regularity of distributions
20
The statement of the theorem then follows from Theorem 2.19 (in particular the equivalence
(i)⇔(iii)) and the estimates in Proposition 3.16.
As was alluded to in §2.4, given n ∈ N, 1 < p < ∞ and 0 < s ≤ n/p, Theorem 4.6 allows
us to construct non-empty compact Swiss-cheese sets K ⊂ Rn with empty interior, which are not
(s, p)-null. To ensure that K has empty interior, one can take the ball centres {xi}∞i=1 to be any
countable dense subset of Ω. To ensure that K is non-empty, one just needs to make sure that
ri → 0 sufficiently fast as i → ∞ so that P∞i=1 m(cid:0)Bri(xi)(cid:1) < m(Ω), from which it follows that
As further concrete examples we consider two families of Cantor sets in R, for which we prove
non-(s, p)-nullity for certain values of s > 0 in Proposition 4.9. Note that since Theorem 4.6 concerns
Swiss-cheese sets, it applies to Cantor sets only for n = 1.
m(K) > 0.
Example 4.7 ("Fat" Cantor set). Given 0 < α < 1/2 and 0 < β < 1 − 2α, denote by G(n)
α,β ⊂ Rn,
n ∈ N, the Cantor set constructed as in Definition 4.3 with lj+1 = 1/2(lj − βαj) for j ≥ 0. For
n = 1, the resulting set Gα,β is called a "generalised Smith–Volterra–Cantor" set in [17, §2.4-2.5],
where it is denoted SV C(α, β). The classical Smith–Volterra–Cantor set is obtained by the choice
α = β = 1/4. (We remark also that the "ǫ-Cantor" sets of [4, p. 140] are obtained by the choice
β = (1 − ǫ)/2, α = 1/4, for a given 0 < ǫ < 1.) Then
lj =
1
2j (cid:18)1 − β(cid:18) 1 − (2α)j
1 − 2α (cid:19)(cid:19) ,
so that
2jlj = 1 − β(cid:18)1 − (2α)j
1 − 2α (cid:19) → 1 −
β
1 − 2α
,
j → ∞,
from which we conclude that m(G(n)
Smith–Volterra–Cantor set we have m(G1/4,1/4) = 1/2.) Hence G(n)
and any 1 < p < ∞; on the other hand, since it has empty interior, G(n)
α,β) = (1 − β/(1 − 2α))n > 0. (In particular for the classical
α,β is not (s, p)-null for any s ≤ 0
α,β is (s, p)-null for s > n/p.
2 lj − lj+1 = 2−j−1(γδj − γδj+1
Example 4.8 ("Super-fat" Cantor set). To construct a Cantor set which is even fatter (in the
sense of nullity), choose lj = 2−j(1 − γ + γδj
) with 0 < γ < 1 and δ > 1. Then lj+1 < lj/2
for all j ≥ 0, and when constructing level j + 1 from level j ∈ N0 we subtract 2j intervals with
radii 1
, which decrease faster as j → ∞ than those in
Example 4.7. We denote the resulting Cantor set, which has Lebesgue measure 0 < (1 − γ)n < 1, by
K (n)
γ,δ .
Proposition 4.9. Given 1 < p < ∞ and 0 < s < 1/p, there exist 0 < α < 1/2 and 0 < β < 1 − 2α
such that the fat Cantor set Gα,β ⊂ R defined in Example 4.7 is not (s, p)-null.
) < 2−j−1γδj
Given 1 < p < ∞, there are parameters 0 < γ < 1 and δ > 2
super-fat Cantor set Kγ,δ ⊂ R of Example 4.8 is not (1/p, p)-null.
Proof. To analyse the fat Cantor set using Theorem 4.6 we note that in constructing Ej from Ej−1
we remove 2j−1 intervals of radius (β/2)αj−1. The intervals we remove are all disjoint, so
1
p−1 such that the corresponding
rn−sp
i
=
∞Xi=1
2j(cid:18) βαj
2 (cid:19)1−sp
∞Xj=0
=(cid:18) β
2(cid:19)1−sp ∞Xj=0
(2α1−sp)j =
(β/2)1−sp
1 − 2α1−sp ,
where for the sum to converge we need 2α1−sp < 1, i.e.
0 < α < 2−1/(1−sp),
or, equivalently,
s < sα,p :=
1
p(cid:18)1 +
log 2
log α(cid:19) .
Since l0 = 1 we can choose r = 1/2 in Theorem 4.6, then (14) will be satisfied (and hence Gα,β is
not (s, p)-null) provided that β lies in the range
0 < β <(cid:18) As,p,1
Bs,p,1(cid:19)1/(1−sp)
(1 − 2α1−sp)1/(1−sp).
On the maximal Sobolev regularity of distributions
21
To analyse the super-fat Cantor set one proceeds in a similar way using (15) instead of (14).
Possible parameters ensuring that Kγ,δ is not (1/p, p)-null are
δ > 2
1
p−1 and 0 < γ < (2c)(cid:0)−C−2
c,p,1(1−2δ1−p)(cid:1) 1
1−p
< 1.
By mimicking the approach of Theorem 4.5 (in particular the construction of the set F (n)
0,∞), one
might try to construct an "ultra-fat" Cantor set that is not (s, p)-null for all s ≤ 1/p and 1 < p < ∞.
We conjecture that such a set can be constructed by subtracting at every level j ∈ N0 an interval
with radius proportional to 1/222j
. However, to use Theorem 4.6 to prove non-nullity, one needs to
know the dependence on p of the constant Cc,p,1 from Proposition 3.16, which corresponds to the
constant A in [2, Proposition 5.1.3].
The proof of Proposition 4.9 implies that if 0 < α < 1/2 is fixed, then the fat Cantor set defined
in Example 4.7 is not (s, p)-null provided that s < sα,p and that β is sufficiently small. But the
permitted range of β appears to depend on s and p in a nontrivial way. The following proposition
provides a complementary result for the special case of (s, 2)-nullity, which removes the restriction
on β. Our proof, which is independent of the capacity-based approach of Theorem 4.6, is inspired
by the method used to prove similar results in [26, Lemma 2.5 and Theorem 3.1], and involves
estimating the local L2 norm of the Fourier transform of the characteristic function of E in terms of
the L2 norm of the difference between the characteristic function and a slightly shifted version of it.
Proposition 4.10. Let E ⊂ R be a Cantor set as in Definition 4.3. Let Gapj := lj−1 − 2lj, j ≥ 1,
denote the length of the 2j−1 gaps introduced in Ej−1 to construct Ej. Assume that Gapj < lj
(equivalently 3lj > lj−1) and that Gapj < Gapj−1 for all j ≥ 1. If
Xj≥2
Gap2−2s
j−1
Gap2
j (cid:16)Xk≥j
2kGapk(cid:17) < ∞
(16)
for s > 0, then the characteristic function χE of E belongs to H s,2 (and hence E is not (s, 2)-null).
In particular, for the fat Cantor set Gα,β with 0 < α < 1/2 and 0 < β < 1 − 2α, χGα,β ∈ H s,2
for all s < sα,2, where
sα,2 :=
1
2(cid:16)1 +
log 2
log α(cid:17) ∈ (0, 1/2).
−∞
Proof. To prove that χE ∈ H s,2 for the claimed range of s our strategy is to prove directly that
H s,2 =R ∞
(1+ξ2)s χE(ξ)2dξ < ∞ (cf. (1)). Since the characteristic function χE is compactly
kχEk2
supported, χE is analytic, and so it suffices to control the behaviour of the integrand at infinity.
Since {Gapj}∞j=1 is a non-increasing sequence, the gaps between the subintervals in Ej have mini-
mal length min1≤k≤j Gapk = Gapj. For any such Cantor set E, the symmetric difference between Ej
and its translation {Ej + Gapj} is composed of 2j+1 (some open, some semi-open) intervals of length
Gapj (in some cases touching each other at one extreme). So kχEj − χ{Ej+Gapj}k2
L2(R) = 2j+1Gapj.
L2(R) = Ej−E =Pk>j 2k−1Gapk and similarly for the
Comparing E and Ej we have kχEj − χEk2
translates. Using the triangle inequality we have
kχE − χ{E+Gapj}k2
L2(R) ≤Xk≥j
2k+2Gapk.
The intervals Ij := (Gap−1
Sj≥2 Ij = (Gap1,∞) with empty mutual intersections. We define the coefficients
j
], j ≥ 2, are a partition of a neighbourhood of infinity; explicitly,
j−1, Gap−1
ξ2s
ξ2s
zj := sup
ξ∈Ij
1 − cos(Gapjξ) ≤ sup
ξ∈Ij
c1
(Gapjξ)2 ≤ c1
Gap2−2s
j−1
Gap2
j
,
j ≥ 2,
≤ 2sXj≥2
≤ 2s−1Xj≥2
= 2s−1Xj≥2
≤ 2s+1Xj≥2
≤ 2s+1c1Xj≥2
zjZξ∈Ij(cid:0)1 − cos(Gapjξ)(cid:1) χE(ξ)2dξ
zjZR 1 − e−iξGapj2 χE(ξ)2dξ
zjkχE − χ{E+Gapj}k2
L2(R)
zj(cid:16)Xk≥j
2kGapk(cid:17)
j (cid:16)Xk≥j
Gap2−2s
j−1
Gap2
2kGapk(cid:17).
On the maximal Sobolev regularity of distributions
22
where c1 := 1/(1 − cos 1) ≈ 2.175. From Plancherel's Theorem and the above bounds we have
Zξ>Gap1
(1 + ξ2)s χE(ξ)2dξ ≤ 2sXj≥2Zξ∈Ij ξ2s χE(ξ)2dξ
Thus, if condition (16) holds, kχEk2
for each j ≥ 1. Hence
(1 + ξ2)s χE(ξ)2dξ < ∞ and the proof is complete.
In the fat Cantor case we have Gapj = βαj−1, so {Gapj}∞j=1 is strictly decreasing and Gapj < lj
Zξ>Gap1
(1 + ξ2)s χE(ξ)2dξ ≤
α3−4s(1 − 2α)Xj≥2
H s,2 =R ∞
α3−4s Xj≥2
α−2sj(cid:16)Xk≥j
(2α)k(cid:17) ≤
2s+1c1β1−2s
2s+1c1β1−2s
(2α1−2s)j,
−∞
which is bounded for s < sα,2.
Proposition 4.10 demonstrates the non-(s, 2)-nullity of the set Gα,β, for s < sα,2, by showing that
a specific function supported inside Gα,β (namely χGα,β ) belongs to H s,2. We note that the proof
relies on Plancherel's Theorem, and hence does not generalise easily to the case p 6= 2. However, we
can immediately deduce some simple consequences for the case p 6= 2. Since Gα,β is bounded, χGα,β
belongs to Lp for all 1 < p < ∞ and, by interpolation (5), it belongs to H s,p for 2 ≤ p < ∞ and
s < sα,p = 2sα,2/p. Hence the nullity threshold of Gα,β satisfies sGα,β (p) ≥ sα,p for all 2 ≤ p < ∞,
and all 0 < β < 1 − 2α. In the light of Proposition 4.9 we expect that this result also extends to
1 < p < 2, but we do not have a proof of this. (The only (weaker) β-independent result we have
for 1 < p < 2 is that sGα,β (p) ≥ sα,2 for all 1 < p < 2, which follows from the embedding (3).) One
might speculate that these inequalities are actually equalities, i.e. that sGα,β (p) = sα,p for all n ∈ N
and 1 < p < ∞. But we do not know of any techniques for obtaining an upper bound on sGα,β (p)
that would verify this. As far as we are aware, calculating the exact value of sGα,β (p) is an open
problem.
combined with Proposition 2.9, that s
Regarding the higher-dimensional fat Cantor sets G(n)
α,β, n ≥ 2, it follows from the above results,
(p) ≥ sα,2 for
all 1 < p < 2. The characteristic function of the one-dimensional fat Cantor set Gα,β does not
belong to H s,p(R) for any s > 1/p by the Sobolev embedding H s,p(R) ⊂ C 0(R) (or equivalently by
the (s, p)-nullity of Gα,β). Hence Proposition A.2 in the Appendix implies that for all n ≥ 2 the
α,β does not belong to H s,p(Rn) either, for
characteristic function χ
due to the Cartesian-
(p) ≥ sα,p for all 2 ≤ p < ∞, and s
s > 1/p, since it can be written as the tensor product χ
of the corresponding set G(n)
G(n)
α,β
G(n)
α,β
G(n)
α,β
= χ
G(n)
α,β
α,β ⊗ χ
G(1)
G(n−1)
α,β
product structure of Cantor sets. Thus we cannot hope to extend the proof of Proposition 4.10 to
show that G(n)
α,β is not (s, p)-null for some 1/p < s ≤ n/p. Indeed, for any Cantor set E(n) ⊂ Rn, if
1/p < s ≤ n/p then non-zero functions in H s,p
E(n) (if any exist) cannot be of tensor-product form.
On the maximal Sobolev regularity of distributions
23
5 Conclusion
In Definition 2.1 and Proposition 2.4 we introduced the concepts of "(s, p)-nullity" and the "nullity
threshold" sE(p) of a subset E ⊂ Rn, to describe the Sobolev regularity of the distributions sup-
ported by E. We now summarise our contributions to the questions Q1–Q3 posed in §2.2. (An
even more concise summary is given in tabular form in Table 2). We assume throughout this section
that E is Borel with empty interior. Our first observation is that the case m(E) = 0 is significantly
better understood than the case m(E) > 0.
Q1: Given 1 < p < ∞ and E ⊂ Rn with empty interior, can we determine sE(p)?
If m(E) = 0 then sE(p) is immediately computed in terms of Hausdorff dimension as sE(p) =
(dimHE − n)/p′, see Theorem 2.12. If m(E) > 0 then we know 0 ≤ sE(p) ≤ n/p, but the only
general result we know of that allows a more precise characterisation of sE(p) is Theorem 2.19,
which appears useful only for proving lower bounds on sE(p). Using a result from [35] we described
two types of set with maximum nullity threshold sE(p) = n/p for all 1 < p < ∞, see Corollary 2.21
and Theorem 2.22. We also derived lower bounds on sE(p) for some "fat" and "super-fat" Cantor
sets and Swiss cheese sets in §4.2.2 (the lower bound obtained for the fat Cantor set is represented
by the dotted line in Figure 1). But we are unaware of any viable techniques for obtaining nontrivial
upper bounds on sE(p) when m(E) > 0. In fact, we have no evidence whatsoever of the existence
of sets with sE(p) ∈ (0, n/p).
Open question: Given 1 < p < ∞, do there exist sets E for which sE(p) ∈ (0, n/p)?
Open question: What is the nullity threshold of the fat Cantor sets of Example 4.7?
Q2: For which functions f : (1,∞) → [−n, n] does there exist E ⊂ Rn such that f (p) = sE(p)
for all p ∈ (1,∞)? To describe the possible nullity-threshold functions, it is convenient to make
the change of variable p 7→ r = 1/p, and define, for E ⊂ Rn non-empty with empty interior,
0 < r < 1.
SE(r) := sE(1/r) = inf(cid:8)s ∈ R, such that E is (s, 1/r)-null(cid:9),
The function SE satisfies the following conditions:
• n(r − 1) ≤ SE(r) ≤ nr for 0 < r < 1, so the graph of SE(r) in the rs-plane lies in the shaded
parallelogram in Figure 1 (cf. Lemma 2.3(iv)–(vi)).
• SE is non-decreasing and Lipschitz-continuous with 0 ≤ S′E(r) ≤ n for a.e. 0 < r < 1 (cf.
Lemma 2.3(ii)).
• The graph of SE in the rs-plane cannot cross the line s = 0. Precisely, if SE(r∗) ≥ 0 (or
SE(r∗) < 0) for some r∗ ∈ (0, 1) then SE(r) ≥ 0 (or SE(r) < 0, respectively) for all r ∈ (0, 1)
(cf. Lemma 2.3(v)).
• If m(E) = 0 then SE(r) = (n − dimH E)(r − 1) (cf. Theorem 2.12). So if the graph of SE lies
in the lower half rs-plane in Figure 1, it is necessarily a straight line through (1, 0) with slope
equal to n − dimHE.
Hence the only functions F : (0, 1) → [−n, 0] which can be realised as F (r) = SE(r) for some
E ⊂ Rn are straight lines F (r) = λ(r− 1), λ ∈ [0, n], which are realised by any set E with m(E) = 0
and dimHE = n − λ (for example the Cantor sets of Theorem 4.5). By contrast, the only functions
F : (0, 1) → [0, n] that we can provably realise as F (r) = SE(r) for some E ⊂ Rn are the two
straight lines F (r) = 0 (already covered by the previous case), and the straight line F (r) = nr, with
E given either by the compact set of Corollary 2.21 or the non-compact set of Theorem 2.22.
On the maximal Sobolev regularity of distributions
24
s
n
sα,2
0
1
2
1
r = 1/p
d − n
−n
Figure 1: Schematic showing the region in the rs-plane in which the graph of SE(r) = sE(1/r) must
lie. The dashed line in the lower triangle is the graph of SE for a set E with Hausdorff dimension
0 < d < n, for instance a Cantor set from Theorem 4.5. The dotted line in the upper triangle
α,β from
represents a lower bound F (r) = min{2rsα,2, sα,2} for SE for a fat Cantor set E = G(n)
Example 4.7 and Proposition 4.10.
Q3: Under what conditions on E and p is E "threshold null" (i.e. (sE(p), p)-null)? For
the case m(E) = 0 we gave a complete classification of the possible threshold nullity behaviours in
Corollary 2.15, which we demonstrated was sharp by providing a zoo of examples in Theorem 4.5.
The only other general results we know of concerning threshold nullity are the following:
• If E is countable then it is (sE(p), p)-null for 1 < p < ∞, with sE(p) = −n/p′ (cf. Corol-
lary 2.16).
• If E = A \ Q, where A has non-empty interior in which the countable set Q is dense, then E
is not (sE(p), p)-null for 1 < p < ∞, with sE(p) = n/p (cf. Theorem 2.22).
• If E is a compact d-set or a d-dimensional hyperplane for 0 < d < n, then it is (sE(p), p)-null
for 1 < p < ∞, with sE(p) = (d − n)/p′ (cf. Theorem 2.17).
• If m(E) = 0 and dimH E = n, then E is (sE(p), p)-null for 1 < p < ∞, with sE(p) = 0.
• If sE(p) = n/p, as for the sets in Corollary 2.21 and Theorem 2.22, then if E is (sE(p0), p0)-null
for some 1 < p0 < ∞ then E is also (sE(p1), p1)-null for every 1 < p1 < p0.
Acknowledgements
The authors thank Simon Chandler-Wilde and Markus Hansen for helpful discussions.
A Tensor products and traces
In this section we collect some results on the Sobolev regularity of tensor-product distributions.
Propositions A.1 and A.2 give lower and a upper bounds, respectively, for the regularity of a tensor-
product distribution in terms of the regularity of the factors. They are simple consequences of the
results proved in [22] in the more general setting of Triebel–Lizorkin spaces and the so-called spaces
of dominating mixed smoothness. We improve them slightly in the Hilbert space case p = 2 by
exploiting Plancherel's theorem. The consequences of these results for (s, p)-nullity are summarised
On the maximal Sobolev regularity of distributions
25
Yes
Yes
Threshold-null?
Table 2: A summary of our results. The first part of the table reviews the dependence of sE(p) on
general properties of E, and the second part catalogues the examples introduced in §4.2 and §2.4.
Conditions on E ⊂ Rn (E 6= ∅, Borel) Nullity threshold sE(p)
sE(p) = +∞
E with non-empty interior
−n/p′ ≤ sE(p) ≤ n/p
E with empty interior
sE(p) = −n/p′
Countable E
sE(p) = (dimHE − n)/p′
0 ≤ dimHE < n
sE(p) = (d − n)/p′
d-set, Lipschitz d-dim. manifold
−1/p′ ≤ sE(p) ≤ 0
E = ∂Ω, with Ω open, int(Ωc) 6= ∅
−1/p′ ≤ sE(p) ≤ −α/p′
E = ∂Ω, with Ω open C 0,α
sE(p) = −1/p′
E = ∂Ω, with Ω open Lipschitz
dimH(E) = n, m(E) = 0
sE(p) = 0
sE(p) ≥ 0
m(E) > 0
E = open \ countable dense
sE(p) = n/p
Examples
Cantor set E(n)
d,p∗
Cantor set F (n)
d,p∗
Fat Cantor set G(n)
α,β
Super-fat Cantor set K (n)
γ,δ
Polking set Kp∗
Union of Polking sets
Rn \ Qn
sE(p) = (d − n)/p′
sE(p) = (d − n)/p′
sE(p) ≥ min{sα,p, sα,2}
No for p ≥ p∗
sE(p) ≥ min{1/p, 1/p∗}
sE(p) ≥ min{n/p, n/p∗} No for p ≥ p∗
sE(p) = n/p
No
No
sE(p) = n/p
Cor. 2.16
Th. 2.12
Th. 2.17
Th. 2.18
Th. 2.18
Th. 2.18
Th. 2.12
Lem. 2.3
Th. 2.22
Only if p ≤ p∗
Only if p < p∗
Prop. 4.9
Th. 2.20
Cor. 2.21
Th. 2.22
, 1 < p∗ < ∞
, 1 ≤ p∗ ≤ ∞
Th. 4.5
Th. 4.5
Yes
Yes
No
Prop. 4.10
in Proposition 2.9 and in the subsequent paragraph. In studying (s, p)-nullity of Cartesian-product
sets, we also make use of classical trace operators and standard bounds on the Hausdorff dimensions
of products: see Proposition A.3. Note that we do not consider "mixed" tensor products of Sobolev
distributions with different integrability parameter p, so we fix 1 < p < ∞ for the whole section.
Given n1, n2 ∈ N and two distributions u1 ∈ D∗(Rn1), u2 ∈ D∗(Rn2), we define the tensor
product u1 ⊗ u2 ∈ D∗(Rn1+n2) as in [24, Chapter V] or [22, Proposition 1.3.1]. This definition
immediately extends to tensor products of finitely many distributions u1⊗···⊗ uN ∈ D∗(Rn1+···nN ).
Proposition A.1. Let N ∈ N, sj ∈ R, nj ∈ N, and uj ∈ H sj ,p(Rnj ) for j = 1, . . . , N . Then
u1 ⊗ ··· ⊗ uN ∈ H s,p(Rn1+···+nN ),
for s < s∗,
(17)
where
s∗ := max(cid:8)0, min
j=1,...,N
sj(cid:9) +
NXj=1
min{0, sj} =
min
∅6=J⊂{1,...,N}Xj∈J
sj =
min
sj
j=1,...,N
Xj s.t. sj<0
sj
if sj ≥ 0 ∀j,
otherwise.
If either p = 2 or sj ∈ N0 for all 1 ≤ j ≤ N , then (17) holds also for s = s∗.
Proof. Assertion (17) is proved in [22, Propositions 2.3.8(ii) and 4.4.1]. The stronger result when
sj ∈ N0 for all 1 ≤ j ≤ N follows from the equivalence in this case between the norm in H sj,p(Rnj )
and norm in W sj,p(Rnj ) (involving Lp norms of weak derivatives), see [22, Proposition 2.3.8(i)].
To prove the stronger result for p = 2 it suffices to consider the case with N = 2 components.
H s∗,2(Rn1+n2 ) can be controlled
In this case, s∗ = min{s1, s2, s1 + s2} and the squared norm ku1 ⊗ u2k2
On the maximal Sobolev regularity of distributions
26
using the representation (1), the relation \u1 ⊗ u2 = u1 ⊗ u2, Fubini's theorem, and by bounding the
function (1 + ξ12 + ξ22)s∗ for all ξ1 ∈ Rn1, ξ2 ∈ Rn2, using the inequalities
(1 + ξ12 + ξ22)s1 ≤ (1 + ξ12)s1(1 + ξ22)s1 ≤ (1 + ξ12)s1(1 + ξ22)s2
(1 + ξ12 + ξ22)s1 ≤ (1 + ξ12)s1 ≤ (1 + ξ12)s1(1 + ξ22)s2
for 0 ≤ s1 ≤ s2,
for s1 < 0 ≤ s2,
(1 + ξ12 + ξ22)s1+s2 = (1 + ξ12 + ξ22)s1(1 + ξ12 + ξ22)s2
≤ (1 + ξ12)s1(1 + ξ22)s2
for s1, s2 < 0.
This gives ku1 ⊗ u2kH s∗ ,2(Rn1 +n1 ) ≤ ku1kH s1,2(Rn1 ) ku2kH s2,2(Rn2 ), from which the assertion follows.
Proposition A.2. Let N ∈ N, s ∈ R, nj ∈ N, and 0 6= uj ∈ S∗(Rnj ) for j = 1, . . . , N .
If
u1 /∈ H s,p(Rn1), then u1 ⊗···⊗ uN /∈ H t,p(Rn1+···+nN ) for any t > s. If either p = 2 or s ∈ N0, then
u1 ⊗ ··· ⊗ uN /∈ H s,p(Rn1+···+nN ).
Proof. Propositions 2.3.8(ii) and 4.4.1 in [22] show that, given sj ∈ N, j = 1, . . . , N , H t,p(Rn1+···+nN )
⊂ H s1,p(Rn1) ⊗αp ··· ⊗αp H sN ,p(RnN ) (defined as the completion of the algebraic tensor product
space under the norm αp of [22, Definition 1.3.1]) for t > s∗, where
sj =
max
j=1,...,N
sj
if sj ≤ 0 ∀j,
otherwise.
sj
max
j=1,...,N
NXj=1
sj(cid:9) +
max{0, sj} =
Xj s.t. sj>0
∅6=J⊂{1,...,N}Xj∈J
s∗ := min(cid:8)0, max
To prove the assertion it suffices to consider the case N = 2, in which case s∗ = max{s1, s2, s1 + s2}.
Setting s1 = s, it follows that if u1 ⊗ u2 ∈ H t,p(Rn1+n2) then u1 ⊗ u2 ∈ H s,p(Rn1) ⊗αp H s2,p(Rn2)
for sufficiently small s2 (i.e. s2 ≤ min{0, s}, so that s∗ = s). Since αp is a so-called crossnorm [22,
equation (1.3.2)], this in turn implies that u2 ∈ H s2,p(Rn2), and more importantly u1 ∈ H s,p(Rn1),
which proves the assertion by contrapositive. The stronger result for the case s ∈ N0 comes from [22,
Proposition 2.3.8(i)].
To prove the stronger assertion for the case p = 2, we note that since u2 ∈ H s2,2(Rn2) it holds
that u2 ∈ L2
loc(Rn2). Since u2 6= 0, there exist c0 > 0 and a bounded measurable set A ⊂ Rn2 with
positive measure m(A) such that u2(ξ2)2 ≥ c0 > 0 for a.e. ξ2 ∈ A. Then, if u1⊗u2 ∈ H s,2(Rn1+n2),
the Plancherel and Fubini theorems give the following contradiction:
∞ > ku1 ⊗ u2k2
H s,2(Rn1 +n2 ) =ZRn1ZRn2 u1(ξ1)2u2(ξ2)2(1 + ξ12 + ξ22)sdξ2dξ1
≥ c0ZRn1ZA u1(ξ1)2(1 + ξ12 + ξ22)sdξ2dξ1
c0ZRn1ZA u1(ξ1)2(1 + ξ12)sdξ2dξ1,
c0ZRn1 u1(ξ1)2(1 + ξ12)sdξ1ZA
c0 m(A) ku1k2
c0 m(A)(cid:16)1 + sup
ku1k2
ξ2∈Aξ22(cid:17)s
H s,2(Rn1 )
H s,2(Rn1 )
≥
≥
(1 + ξ22)sdξ2,
= ∞,
= ∞,
s ≥ 0,
s < 0,
s ≥ 0,
s < 0.
To better interpret these results, we define the "maximal Sobolev regularity" of a distribution:
mp,n(u) := sup(cid:8)s ∈ R, such that u ∈ H s,p(Rn)(cid:9) ∈ R ∪ {±∞},
u ∈ S∗(Rn), 1 < p < ∞, n ∈ N.
On the maximal Sobolev regularity of distributions
27
Then, Propositions A.1 and A.2 combine to give a precise characterisation of the maximal Sobolev
regularity of a tensor-product distribution if the maximal Sobolev regularity of at least one of the
two factors is non-negative:
nj ∈ N, max(cid:8)mp,n1(u1), mp,n2(u2)(cid:9) ≥ 0 ⇒ mp,n1+n2(u1 ⊗ u2) = min(cid:8)mp,n1(u1), mp,n2(u2)(cid:9).
Moreover, if p = 2 or mp,n1+n2(u1 ⊗ u2) ∈ N0, then u1 ⊗ u2 belongs to H mp,n1+n2 (u1⊗u2),p(Rn1+n2)
if and only if u1 ∈ H mp,n1+n2 (u1⊗u2),p(Rn1) and u2 ∈ H mp,n1+n2 (u1⊗u2),p(Rn2). A similar statement
holds for the tensor product of N distributions, if all except at most one have non-negative maximal
Sobolev regularity. If both u1 and u2 have negative maximal Sobolev regularity, the result is less
sharp:
nj ∈ N, max(cid:8)mp,n1(u1), mp,n2(u2)(cid:9) ≤ 0
⇒ mp,n1(u1) + mp,n2(u2) ≤ mp,n1+n2(u1 ⊗ u2) ≤ min(cid:8)mp,n1(u1), mp,n2(u2)(cid:9).
We point out that the lower bound here can be achieved. Indeed, if xj ∈ Rnj , j = 1, 2, then by (4)
mp,nj (δxj ) = −nj/p′, j = 1, 2, and mp,n1+n2(δx ⊗ δy) = −(n1 + n2)/p′ = mp,n1(δx) + mp,n2(δy).
So far we have related distributions defined on Euclidean spaces with different dimensions using
tensor products. To relate functions (with positive regularity exponent s) defined on an ambient
Euclidean space and on affine subspaces one can use traces. Using classical results on traces in
Triebel–Lizorkin spaces we can prove the following result.
Proposition A.3. Let n1, n2 ∈ N, 1 < p < ∞, s ∈ R, E1 ⊂ Rn1 and E2 ⊂ Rn2. If E1 is (s, p)-null
then E1 × E2 is (t, p)-null for
t ≥ s + n2
p
t > s + n2
p
t ≥ s
t > s
if s > 0, 1 < p ≤ 2,
if s > 0, 2 < p < ∞,
if s = 0,
if s < 0 and E1, E2 Borel.
Proof. The case s = 0 follows from Lemma 2.3(v), while the case s < 0 can easily be deduced from
the relation between nullity and Hausdorff dimension in (9) and the inequality dimH(E1 × E2) ≤
dimH(E1) + n2 [18, equation (7.7)].
For the case s > 0, for any y ∈ Rn2, §2.7.2 and §2.3.2 of [38] give continuity of the trace operator
Try : H t,p(Rn1+n2) → H s,p(Rn1),
s > 0, (t ≥ s + n2
p
t > s + n2
p
if 1 < p ≤ 2,
if 2 < p < ∞,
defined on D(Rn1+n2) as the pointwise trace onto the subspace {(x1, x2) ∈ Rn1+n2, x2 = y} (which
is canonically identified with Rn1), and then extended to H t,p(Rn1+n2) by density.
If 0 6= u ∈
H t,p(Rn1+n2) ⊂ Lp(Rn1+n2), by Fubini's theorem there exists at least one y ∈ Rn2 such that
Try(u) 6= 0. If moreover supp u ⊂ E1×E2, by convolution with a sequence of mollifiers in D(Rn1+n2)
with decreasing support it is straightforward to show that supp(Try(u)) ⊂ E1. This proves the
assertion by contrapositive.
B A result on the non-equality of capacities
In this appendix (due to Simon Chandler-Wilde) we give a concrete example of an open set Ω ⊂ R
for which cap2,2(Ω) < Cap2,2(Ω). Specifically, we consider an open interval Ω = (−a, a) ⊂ R, where
a > 0. We recall that the space H 2,2(Ω) ⊂ D∗(Ω) is defined by
H 2,2(Ω) :=(cid:8)u ∈ D∗(Ω) : u = UΩ for some U ∈ H 2,2(cid:9),
On the maximal Sobolev regularity of distributions
kukH 2,2(Ω) := inf
U∈H 2,2
UΩ=u
kUkH 2,2 .
It is straightforward to show that
cap2,2(Ω) = inf{kuk2
Cap2,2(Ω) = inf{kuk2
H 2,2(Ω) : u ∈ H 2,2(Ω), u ≥ 1 a.e. on Ω},
H 2,2(Ω) : u ∈ H 2,2(Ω), u = 1 a.e. on Ω} = k1k2
H 2,2(Ω).
28
(18)
(19)
An explicit formula for kukH 2,2(Ω) in the case where Ω is an open interval has been given recently
in [11, Equation (26)]. For even functions u ∈ H 2,2(Ω) this formula gives (note that we correct a
typographical error in [11, Equation (26)], replacing −φ′(a) in that formula by +φ′(a))
H 2,2(Ω) = 2(cid:18)u(a)2 + u′(a)2 + u(a) + u′(a)2 +Z a
kuk2
0
(u(t)2 + 2u′(t)2 + u′′(t)2) dt(cid:19) .
(20)
From (20) and (19) it follows that Cap2,2(Ω) = k1k2
a > 0 construct a function u ∈ H 2,2(Ω) which satisfies u ≥ 1 on Ω and has kuk2
This, in the light of (18), demonstrates that cap2,2(Ω) < Cap2,2(Ω) for this particular Ω.
H 2,2(Ω) = 4 + 2a. But using (20) we can for any
H 2,2(Ω) < 4 + 2a.
When a < √3 we consider the quadratic function u(t) = 1 + ε(a2 − t2), for some ε > 0 to be
specified. By (20) this function satisfies
H 2,2(Ω) = 4 + 2a − 8a(cid:18)1 −
kuk2
a2
3(cid:19) ε + O(cid:0)ε2(cid:1) ,
H 2,2(Ω) < 4 + 2a for sufficiently small ε, provided a < √3.
ε → 0,
When a > 1 we consider the function (again with ε > 0 to be specified)
so that kuk2
u(t) =(1,
1 + ε(a − t)(a − 1 − t)2, a − 1 ≤ t < a,
t ≤ a − 1,
which by (20) satisfies
kuk2
H 2,2(Ω) = 4 + 2a −
11
3
and again we have kuk2
H 2,2(Ω) < 4 + 2a for sufficiently small ε.
ε + O(cid:0)ε2(cid:1) ,
ε → 0,
References
[1] D. R. Adams, On the exceptional sets for spaces of potentials, Pacific J. Math., 52 (1974),
pp. 1–5.
[2] D. R. Adams and L. I. Hedberg, Function Spaces and Potential Theory, Springer, 1999.
corrected 2nd printing.
[3] D. R. Adams and J. C. Polking, The equivalence of two definitions of capacity, Proc. Am.
Math. Soc., 37 (1973), pp. 529–534.
[4] C. D. Aliprantis and O. Burkinshaw, Principles of Real Analysis, Academic Press, Inc.,
San Diego, CA, third ed., 1998.
[5] A. M. Bruckner, J. B. Bruckner, and B. S. Thomson, Real Analysis, Prentice Hall
(Pearson), 1997.
[6] L. Carleson, Selected Problems on Exceptional Sets, D. van Nostrand Company, Princeton,
1967.
On the maximal Sobolev regularity of distributions
29
[7] S. N. Chandler-Wilde, Scattering by arbitrary planar screens, in Computational Electro-
magnetism and Acoustics, Oberwolfach Report No. 03/2013, DOI: 10.4171/OWR/2013/03,
2013, pp. 154–157.
[8] S. N. Chandler-Wilde and D. P. Hewett, Well-posed PDE and integral equation formu-
lations for scattering by fractal screens, arXiv preprint (2016), arXiv:1611.09539.
[9]
, Acoustic scattering by fractal screens: mathematical formulations and wavenumber-
explicit continuity and coercivity estimates. Technical report (2013). University of Reading
preprint MPS-2013-17.
[10]
, Wavenumber-explicit continuity and coercivity estimates in acoustic scattering by planar
screens, Integr. Equat. Operat. Th., 82 (2015), pp. 423–449.
[11] S. N. Chandler-Wilde, D. P. Hewett, and A. Moiola, Interpolation of Hilbert and
Sobolev spaces: quantitative estimates and counterexamples, Mathematika, 61 (2015), pp. 414–
443.
[12]
, Sobolev spaces on non-Lipschitz subsets of Rn with application to boundary integral equa-
tions on fractal screens, arXiv preprint (2016), arXiv:1607.01994.
[13] J. Chazarain and A. Piriou, Introduction to the Theory of Linear Partial Differential Equa-
tions, North-Holland, 1982.
[14] G. Choquet, Theory of capacities, Annales de linstitut Fourier, 5 (1954), pp. 131–295.
[15] X. Claeys and R. Hiptmair, Integral equations on multi-screens, Integr. Equat. Oper. Th.,
77 (2013), pp. 167–197.
[16] J. Deny, Les potentiels d'´energie finie, Acta Math., 82 (1950), pp. 107–183.
[17] R. DiMartino and W. Urbina, On Cantor-like sets and Cantor-Lebesgue singular functions,
arXiv preprint (2014), arXiv:1403.6554.
[18] K. Falconer, Fractal Geometry: Mathematical Foundations and Applications, Wiley, 3rd ed.,
2014.
[19] L. E. Fraenkel, On regularity of the boundary in the theory of Sobolev spaces, Proc. London
Math. Soc., 3 (1979), pp. 385–427.
[20] P. Grisvard, Elliptic Problems in Nonsmooth Domains, SIAM Classics in Applied Mathemat-
ics, 2011.
[21] V. V. Grusin, A problem in the entire space for a certain class of partial differential equations,
Dokl. Akad. Nauk SSSR, 146 (1962), pp. 1251–1254.
[22] M. Hansen, Nonlinear Approximation and Function Spaces of Dominating Mixed Smoothness,
PhD thesis, Fakultat fur Mathematik und Informatik der Friedrich-Schiller-Universitat Jena,
2010. Available at http://suche.thulb.uni-jena.de/vufind/Record/644443235.
[23] D. P. Hewett and A. Moiola, A note on properties of the restriction operator on Sobolev
spaces, arXiv preprint (2016), arXiv:1607.01741.
[24] L. Hormander, The Analysis of Linear Partial Differential Operators I: Distribution Theory
and Fourier Analysis, Springer-Verlag, Berlin, second ed., 1990.
[25] L. Hormander and J. L. Lions, Sur la compl´etion par rapport `a une integrate de Dirichlet,
Math. Scand., 4 (1956), pp. 259–270.
On the maximal Sobolev regularity of distributions
30
[26] W. Lawton, The Feichtinger conjecture for exponentials, J. Nonlinear Anal. Optimiz., 2
(2011), pp. 123–132.
[27] J.-L. Lions and E. Magenes, Non-Homogeneous Boundary Value Problems and Applications
I, Springer-Verlag, 1972.
[28] W. Littman, A connection between α-capacity and m − p polarity, Bull. Am. Math. Soc., 73
(1967), pp. 862–866.
[29]
, Polar sets and removable singularities of partial differential equations, Ark. Mat., 7
(1967), pp. 1–9.
[30] V. G. Maz'ya, Sobolev Spaces with Applications to Elliptic Partial Differential Equations,
Springer, 2nd ed., 2011.
[31] V. G. Maz'ya and V. P. Havin, Nonlinear potential theory, Uspehi Mat. Nauk, 27 (1972),
pp. 67–138. English translation: Russian Math. Surveys 27 (1972), pp. 71–148.
[32] W. McLean, Strongly Elliptic Systems and Boundary Integral Equations, CUP, 2000.
[33] N. G. Meyers, A theory of capacities for potentials of functions in Lebesgue classes., Math.
Scand., 26 (1970), pp. 255–292.
[34] J. C. Polking, Approximation in Lp by solutions of elliptic partial differential equations, Am.
J. Math., 94 (1972), pp. 1231–1244.
[35]
, Leibniz formula for some differentiation operators of fractional order, Indiana U. Math.
J., 21 (1972), pp. 1019–1029.
[36] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Uni-
versity Press, 1970.
[37] H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, North Holland,
1978.
[38]
[39]
[40]
, Theory of Function Spaces, Birkhauser Verlag, 1983.
, Fractals and Spectra, Birkhauser Verlag, 1997.
, The dichotomy between traces on d-sets Γ in Rn and the density of D(Rn \ Γ) in function
spaces, Acta Math. Sin., 24 (2008), pp. 539–554.
[41] W. P. Ziemer, Weakly Differentiable Functions: Sobolev Spaces and Functions of Bounded
Variation, vol. 120 of Graduate Texts in Mathematics, Springer, 1989.
|
1006.3967 | 2 | 1006 | 2011-09-20T12:48:37 | Inversion Formula for the Windowed Fourier Transform | [
"math.FA"
] | In this paper, we study the inversion formula for recovering a function from its windowed Fourier transform. We give a rigorous proof for an inversion formula which is known in engineering. We show that the integral involved in the formula is convergent almost everywhere on $\bbR$ as well as in $L^p$ for all $1<p<\infty$ if the function to be reconstructed is. | math.FA | math |
Inversion Formula for the Windowed Fourier Transform∗
Department of Mathematics and LPMC, Nankai University, Tianjin 300071, China
Email: [email protected]
Wenchang Sun
Abstract
In this paper, we study the inversion formula for recovering a function from its
windowed Fourier transform. We give a rigorous proof for an inversion formula which is
known in engineering. We show that the integral involved in the formula is convergent
almost everywhere on R as well as in Lp for all 1 < p < ∞ if the function to be
reconstructed is.
Keywords. Fourier transforms; windowed Fourier transforms; inversion formula.
2000 Mathematics Subject Classification.
42A38.
1
Introduction and the Main Result
The Fourier transform is a very useful mathematical tool, which has been widely used in
characterization of function spaces as well as in signal and image processing [6, 11]. For a
function f ∈ L1(R), the Fourier transform of f is defined by
f (ω) =ZR
f (x)e−ixωdx.
To study local properties of functions (signals), the windowed Fourier transform, also
known as short-time Fourier transform, is introduced.
Given a window function g(x), the windowed Fourier transform of a function f with
respect to g is defined by
(Fgf )(t, ω) =ZR
f (x)g(x − t)e−ixωdx.
It is easy to see that Fgf is well defined if f ∈ Lp(R) and g ∈ Lp′
1/p + 1/p′ = 1.
(R), where p, p′ ≥ 1 and
Continuous and discrete windowed Fourier transforms have been discussed extensively
in the literature since they are widely used in communication theory, quantum mechanics,
and many other fields. We refer to [3, 4, 5, 7, 8] for an introduction to the windowed
Fourier transform.
∗This work was supported partially by the National Natural Science Foundation of China(10971105 and
10990012) and the Natural Science Foundation of Tianjin (09JCYBJC01000).
1
Finding a computationally efficient algorithm for the inversion of windowed Fourier
transforms is a fundamental topic in both theory and applications. The classical method
to recover f from its windowed Fourier transform is to use the following inversion formula,
f (x) =
1
2πkgk2
2 ZZR2
(Fgf )(t, ω)g(x − t)eixωdtdω,
(1.1)
where we assume that g ∈ L2(R). It can be shown that the convergence is in L2(R) as
well as in many other spaces if the function to be reconstructed is and g satisfies some
further conditions [7].
Since a double integral is involved in (1.1), it is obviously very complicated. An
alternate method is to use the filter-bank summation [1],
f (x) =
1
2πg(0)ZR
(Fgf )(x, ω)eixωdω,
(1.2)
where we assume that g(0) 6= 0. Note that (1.2) was presented in [1] in a discrete version
for compactly supported window functions and the authors stated that their results may
be equally well stated in a continuous time-domain setting.
Although (1.2) is well known in engineering, the convergence of the integral is not well
stated in literature. In this paper, we show that the integral in (1.2) is convergent in Lp(R)
for all 1 < p < ∞ if the function f is. Moreover, by applying the Carleson-Hunt theorem,
we also show that the convergence is almost everywhere on R.
Before stating our result, we introduce some definitions. Throughout this paper, x0 is
a fixed real number. For any A1, A2 > 0, define
(TA1,A2f )(x) =Z A2
−A1
Our main result is the following.
(Fgf )(x − x0, ω)eixωdω.
(1.3)
Theorem 1.1 Suppose that g is continuous and that g, g ∈ L1(R). Then for any f ∈
Lp(R), 1 < p < ∞, we have
and
lim
A1,A2→∞
kTA1,A2f − 2πg(x0)f kp = 0
lim
A→∞
(TAf )(x) = 2πg(x0)f (x),
a.e.,
(1.4)
(1.5)
where we use the shortcut TAf = TA,Af .
Remark 1.2 The reconstruction formula (1.3) is stable in the sense that for any f, f ∈
Lp(R),
kTA1,A2(f − f )kp ≤ 2Cpkgk1kf − f kp,
∀A1, A2 > 0,
where Cp is a constant depending only on p. For details, see the proof of Theorem 1.1.
In Section 2, we give the proof of Theorem1.1, which is based on the famous Carleson-
Hunt theorem [2, 9] for Fourier series and the extension to Fourier integrals by Kenig and
Tomas [10].
2
2 Proof of the Main Result
In this section, we give the proof of the main result.
We begin with a simple lemma on the Fourier transform, for which we omit the proof.
Lemma 2.1 For any f ∈ L2(R) with f ∈ L1(R), we have
f (x) =
1
2π ZR
f (ω)eixωdω,
a.e.
We also need the following formula on the windowed Fourier transform.
Proposition 2.2 ([8, Lemma 3.1.1]) For any f, g ∈ L2(R), we have
(Fgf )(x, ω) =
1
2π
(Fg f )(ω, −x)e−ixω.
Next, we show that for f ∈ L2(R) with f ∈ L1(R), TA1,A2f is convergent in L∞ norm.
Lemma 2.3 Suppose that g is continuous and that g, g ∈ L1(R). Then for any f ∈ L2(R)
with f ∈ L1(R), we have
lim
A1,A2→∞
kTA1,A2f − 2πg(x0)f k∞ = 0.
(2.1)
Proof. For any f ∈ L2(R), we see from Proposition 2.2 that
(Fgf )(x − x0, ω)eixωdω
(Fg f )(ω, x0 − x)e−i(x−x0)ωeixωdω
−A1
f (y)g(y − ω)e−iy(x0−x)dy
1
=
=
−A1
1
(TA1,A2f )(x) = Z A2
2π Z A2
2π Z A2
2π ZR
2π ZR
2π ZR
=
=
=
−A1
1
1
1
eix0ωdωZR
f (y)eiyxdyZ A2
f (y)eiyxdyZ y+A1
g(ω)e−ix0ωdωZ ω+A2
ω−A1
y−A2
−A1
g(y − ω)e−ix0(y−ω)dω
g(ω)e−ix0ωdω
f (y)eiyxdy,
(2.2)
where we use Fubini's theorem twice. By Lemma 2.1, for almost every x,
(TA1,A2f )(x) − 2πg(x0)f (x)
2π Z ω+A2
g(ω)e−ix0ωdω(cid:18) 1
= ZR
g(ω)e−ix0ωdωZ y<ω−A1
2π ZR
ω−A1
=
1
or y>ω+A2
f (y)eiyxdy − f (x)(cid:19)
f (y)eiyxdy.
3
Hence
kTA1,A2f − 2πg(x0)f k∞ ≤
1
2π ZR
g(ω)dωZ y<ω−A1
or y>ω+A2
f (y)dy.
By the dominated convergence theorem, we get
lim
A1,A2→∞
kTA1,A2f − 2πg(x0)f k∞ = 0.
This completes the proof.
(cid:3)
In the followings we prove the convergence in Lp(R). First, we show that TA1,A2 is well
defined on Lp(R).
Lemma 2.4 Suppose that g is continuous and that g, g ∈ L1(R). For any A1, A2 > 0, let
KA1,A2(x, y) = g(y − x + x0) ·(cid:18) sin A1(y − x) + sin A2(y − x)
(y − x)
2 sin (A2−A1)(y−x)
sin (A2+A1)(y−x)
2
2
(y − x)
(cid:19).
(2.3)
−i ·
Then we have
(TA1,A2f )(x) =ZR
f (y)KA1,A2(x, y)dy,
∀f ∈ Lp(R).
Proof. Since g, g ∈ L1(R), we have g ∈ Lp(R) for all 1 < p < ∞. Hence Fgf is well
defined for any f ∈ Lp(R). We have
(Fgf )(x − x0, ω)eixωdω
f (y)g(y − x + x0)e−iyωeixωdy
f (y)g(y − x + x0)e−i(y−x)ωdω
−A1
(TA1,A2f )(x) = Z A2
= Z A2
= ZR
= ZR
−A1
dωZR
dyZ A2
−A1
f (y)KA1,A2(x, y)dy,
where Fubini's theorem is used. This completes the proof.
(cid:3)
The pointwise convergence of Fourier series is a deep result in harmonic analysis.
Carleson proved that the Fourier series of a function in L2[−π, π] is convergent almost
everywhere [2]. Hunt [9] extended this result to Lp[−π, π] for 1 < p < ∞. And Kenig
and Tomas [10] proved the pointwise convergence of Fourier integral on Lp(R). For our
purpose, we cite the Carleson-Hunt theorem in the following form.
Proposition 2.5 For A > 0 and 1 < p < ∞, define
(SAf )(x) =ZR
f (y)
sin A(x − y)
π(x − y)
dy,
f ∈ Lp(R).
(2.4)
Then SA is a bounded linear operator on Lp(R) and there exists some constant Cp such
that
≤ Cpkf kp.
(cid:13)(cid:13)(cid:13)
sup
A>0
(SAf )(x)(cid:13)(cid:13)(cid:13)p
4
The Fourier multiplier is a useful tool in the study of Fourier transform. The following
result on the Fourier multiplier is useful in studying the convergence of TA1,A2.
Proposition 2.6 ([6, Corollary 3.8]) Suppose that h is a function of bounded variation
on R and that (T f ) = h · f for f ∈ L2(R). Then T can be extended to an operator on
Lp(R), 1 < p < ∞ and
kT f kp ≤ CpVhkf kp,
∀f ∈ Lp(R),
where Vh is the total variation of h on R and Cp is a constant depending only on p.
The following lemma shows that TA1,A2f converges to f in Lp(R) whenever f is in
c (R), the space of all continuous differentiable functions which are compactly supported.
C 1
Lemma 2.7 For any f ∈ C 1
c (R), we have
lim
A1,A2→∞
kTA1,A2f − 2πg(x0)f kp = 0,
1 < p ≤ ∞.
(2.5)
Proof. Fix some f ∈ C 1
c (R). Suppose that supp f ⊂ [−Ω, Ω] for some constant Ω > 0.
Since f, f ′ ∈ L2(R), we have f ∈ L1(R). By Lemma 2.3,
lim
A1,A2→∞
kTA1,A2f − 2πg(x0)f k∞ = 0.
Next we assume that 1 < p < ∞. By (2.6), we have
lim
A1,A2→∞
k(TA1,A2f − 2πg(x0)f ) · χ[−2Ω,2Ω]kp = 0.
(2.6)
(2.7)
On the other hand, put
K(x, y) =
4kgk∞
x − y
,
x 6= y.
By Minkovski's inequality, we have
Zx≥2Ω(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ Zy≤Ω
≤ 4kgk∞Zy≤Ω
Zy≤Ω
f (y) Zx≥2Ω
dx!1/p
K(x, y)f (y)dy(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
K(x, y)pdx!1/p
x − yp!1/p
f (y) Zx>2Ω
dy
p
dx
dy
= Mpkf k1
≤ Mp(2Ω)1/p′
kf kp,
where Mp is a constant and 1/p + 1/p′ = 1. Note that
(TA1,A2f )(x) ≤Zy≤Ω
KA1,A2(x, y)f (y)dy ≤Zy≤Ω
K(x, y)f (y)dy.
5
By the dominated convergence theorem, we have
lim
A→∞
k(TA1,A2f − 2πg(x0)f ) · χR\[−2Ω,2Ω]kp = 0.
Now the conclusion follows by combining (2.7) and (2.8).
We are now ready to give the proof of the main result.
Proof of Theorem 1.1. First, we prove the convergence in Lp(R).
For any f ∈ L2(R), by (2.2), we have
(2.8)
(cid:3)
(TA1,A2f )(x) =
1
2π ZR
f (y)eiyxdyZ y+A1
y−A2
g(ω)e−ix0ωdω.
(2.9)
Hence
where
(TA1,A2f )(y) = hA1,A2(y) f (y),
hA1,A2(y) =Z y+A1
y−A2
g(ω)e−ix0ωdω.
Obviously, hA1,A2 is of bounded variation on R and VhA1,A2
≤ 2kgk1.
By Lemma 2.4 and Proposition 2.6, TA1,A2 is a bounded linear operator on Lp(R) and
kTA1,A2f k ≤ 2Cpkgk1kf kp,
∀f ∈ Lp(R).
(2.10)
Fix some f ∈ Lp(R). For any ε > 0, there is some f ∈ C 1
Lemma 2.7, we can find some A0 > 0 such that for any A1, A2 > A0,
c (R) such that kf − f kp < ε. By
kTA1,A2
f − 2πg(x0) f kp < ε.
Consequently,
kTA1,A2f − 2πg(x0)f kp ≤ kTA1,A2(f − f )kp + kTA1,A2
+2πg(x0) · kf − f kp
f − 2πg(x0) f kp
≤ (2Cpkgk1 + 2πg(x0) + 1)ε.
Hence
lim
A1,A2→∞
kTA1,A2f − 2πg(x0)f kp = 0,
∀f ∈ Lp(R).
Next we consider the pointwise convergence. For A > 0, let SA be defined by (2.4).
Then SA is a bounded linear operator on Lp(R) and
(SAf )= f · χ[−A,A],
f ∈ L2(R).
For f ∈ L2T Lp(R), define
( TAf )(x) =
1
2π ZR
where the operator Mω is defined by
g(ω)e−ix0ω(M−ωSAMωf )(x)dω,
(Mωf )(x) = e−ixωf (x).
6
linear operator on Lp(R).
Since g ∈ L1(R) and L2T Lp(R) is dense in Lp(R), TA can be extended to a bounded
On the other hand, for any f ∈ L2T Lp(R), we see from (2.2) that
(TAf )(x) =
f (y)eiyxdy
f (y + ω)eiyxeixωdy
g(ω)e−ix0ωdωZ ω+A
g(ω)e−ix0ωdωZ A
ω−A
−A
1
1
2π ZR
2π ZR
2π ZR
1
=
=
= ( TAf )(x).
g(ω)e−ix0ω(M−ωSAMωf )(x)dω
Using the density of L2T Lp(R) again, we get that TA = TA on Lp(R). Hence
g(ω)e−ix0ω(M−ωSAMωf )(x)dω,
∀f ∈ Lp(R).
(TAf )(x) =
1
2π ZR
(2.11)
It follows that
(TAf )(x) ≤
sup
A>0
1
2π ZR
g(ω) · sup
A>0
(M−ωSAMωf )(x)dω.
By Minkovski's inequality and Proposition 2.5, we have
sup
A>0
(cid:13)(cid:13)(cid:13)(cid:13)
(TAf )(x)(cid:13)(cid:13)(cid:13)(cid:13)p
≤
=
≤
=
1
1
2π ZR
2π ZR
2π ZR
1
Cp
2π
dω
sup
A>0
sup
A>0
g(ω) ·(cid:13)(cid:13)(cid:13)(cid:13)
g(ω) ·(cid:13)(cid:13)(cid:13)(cid:13)
(M−ωSAMωf )(x)(cid:13)(cid:13)(cid:13)(cid:13)p
(SAMωf )(x)(cid:13)(cid:13)(cid:13)(cid:13)p
dω
g(ω) · CpkMωf kpdω
kgk1kf kp,
∀f ∈ Lp(R).
(2.12)
Fix some f ∈ Lp(R). For any ε > 0, we can find some f ∈ C 1
c (R) such that
By Lemma 2.7, we have
kf − f kp < ε.
kTA f − 2πg(x0) f k∞ = 0.
lim
A→∞
Note that TAf is continuous on R, thanks to Lemma 2.4. We have
Hence
lim
A→∞
sup
x∈R(cid:12)(cid:12)(cid:12)
(TA f )(x) − 2πg(x0) f (x)(cid:12)(cid:12)(cid:12)
= 0.
lim sup
A,A′→∞
(TA f )(x) − (TA′ f )(x) = 0,
∀x ∈ R.
7
It follows that
lim sup
A,A′→∞
lim sup
lim sup
(TAf )(x) − (TA′f )(x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(TA(f − f ))(x)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)(cid:13)p
A→∞ (cid:12)(cid:12)(cid:12)
A,A′→∞(cid:12)(cid:12)(cid:12)
(cid:13)(cid:13)(cid:13)(cid:13)p
+(cid:13)(cid:13)(cid:13)(cid:13)
(TA′(f − f ))(x)(cid:12)(cid:12)(cid:12)
A′→∞ (cid:12)(cid:12)(cid:12)
(TA(f − f ))(x)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)(cid:13)p
A>0(cid:12)(cid:12)(cid:12)
kgk1kf − f kp
(using (2.12))
lim sup
sup
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤ (cid:13)(cid:13)(cid:13)(cid:13)
≤ 2(cid:13)(cid:13)(cid:13)(cid:13)
≤
Cp
π
Cp
π
(TA f )(x) − (TA′ f )(x)(cid:12)(cid:12)(cid:12)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p
<
kgk1 · ε.
Since ε is arbitrary, we have
lim sup
A,A′→∞
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(TAf )(x) − (TA′f )(x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p
= 0.
Hence the limit limA→∞(TAf )(x) exists almost everywhere. Since TAf tends to 2πg(x0)f
in Lp(R), we have
lim
A→∞
(TAf )(x) = 2πg(x0)f (x),
a.e.
This completes the proof.
References
(cid:3)
[1] J.B. Allen and L.R. Rabiner, A unified approach to short-time Fourier analysis and
synthesis, Proc. IEEE, 65(1977), 1558 -- 1564.
[2] L. Carleson, On convergence and growth of partial sumas of Fourier series, Acta
Math., 116 (1966), 135 -- 157.
[3] O. Christensen, An Introduction to Frames and Riesz Bases, Birkhauser, Boston,
2003.
[4] C. Chui, An Introduction to Wavelets, Academic Press, Inc., Boston, MA, 1992.
[5] I. Daubechies, Ten Lectures on Wavelets, SIAM, 1990.
[6] J. Duoandikoetxea, Fourier analysis, Translated and revised from the 1995 Spanish
original by David Cruz-Uribe, American Mathematical Society, Providence, RI, 2001.
[7] H.G. Feichtinger and T. Strohmer, Eds., Gabor Analysis and Algorithms: Theory
and Applications, Birkhauser, Boston, 1998.
8
[8] K. Grochenig, Foundations of Time-Frequency Analysis, Birkhauser, Boston, 2001.
[9] R.A. Hunt, On the convergence of Fourier series, Orthogonal Expansions and Their
Continuous Analogues, Proc. Conf. Edwardsville, Ill. (1967), 235-255. Southern Illi-
nois Univ. Press, Carbondale, Ill. (1968).
[10] C. Kenig and P. Tomas, Maximal operators defined by Fourier multipliers, Studia
Math., 68 (1980), 79 -- 83.
[11] E.M. Stein, Harmonic Analysis, Princeton University Press, 1993.
9
|
1708.02576 | 4 | 1708 | 2018-06-22T16:24:59 | On approximation tools and its applications on compact homogeneous spaces | [
"math.FA"
] | We prove a characterization for the Peetre type $K$-functional on $\mathbb{M}$, a compact two-point homogeneous space, in terms the rate of approximation of a family of multipliers operator defined to this purpose. This extends the well known results on the spherical setting. The characterization is employed to show that an abstract H\"{o}lder condition or finite order of differentiability condition imposed on kernels generating certain operators implies a sharp decay rates for their eigenvalues sequences. The latest is employed to obtain estimates for the Kolmogorov $n$-width of unit balls in Reproducing Kernel Hilbert Space (RKHS). | math.FA | math |
On approximation tools and its applications on compact
homogeneous spaces
A. O. Carrijo∗ & T. Jordao †
Abstract. We prove a characterization for the Peetre type K-functional on M, a compact two-
point homogeneous space, in terms the rate of approximation of a family of multipliers operator
defined to this purpose. This extends the well known results on the spherical setting. The charac-
terization is employed to show that an abstract Holder condition or finite order of differentiability
condition imposed on kernels generating certain operators implies a sharp decay rates for their
eigenvalues sequences. The latest is employed to obtain estimates for the Kolmogorov n-width of
unit balls in Reproducing Kernel Hilbert Space (RKHS).
Keywords: K-funcional, multiplier (average) operators, eigenvalues sequences, Holder condition.
AMS Classification: 41A60, 41A10, 41A36, 45C05, 47A75.
1
Introduction
The basic framework of this paper refers to a compact two-point homogeneous space of dimension
m ≥ 1. Denoting by M this space which is both a Riemannian m-manifold and a compact symmetric
space of rank 1 for which there is a well-developed harmonic analysis structure on them. A very
large class of problems in approximation theory, harmonic analysis and functional analysis (as it
can be seen in the present paper) can be considered naturally on these spaces.
Each one of these manifolds M has an invariant Riemannian (geodesic) metric d(·,·) which
can be considered normalized so that all geodesics on M have the same length, namely, 2π. Also
M is endowed naturally with a measure dx induced by the normalized left Haar measure which
exists on a component of M seeing as a quotient. According to Wang [21], compact two-point
homogeneous spaces are: the unit spheres Sm, m = 1, 2, . . . ; the real projective spaces Pm(R),
m = 2, 3, . . . ; the complex projective spaces Pm(C), m = 4, 6, . . . ; the quaternion projective spaces
Pm(H), m = 8, 12, . . . and 16-dimensional Cayley's elliptic plane P16. These spaces have a very
similar geometry between them and we shall assume here that M 6= Sm since the results we will
present here already have their spherical version explored (see [6, 12] and references quoted there).
If we write B for Laplace-Beltrami operator on M it is known that its differential form depends
on a pair of index (α, β) varying according to the space. Namely, α = (σ + ρ − 1)/2 = (m − 2)/2
and β = (ρ − 1)/2 where: for Sm, σ = 0, ρ = m − 1; for Pm(R), σ = m − 1, ρ = 0; for Pm(C),
σ = m − 2, ρ = 1; for Pm(H), σ = m − 4, ρ = 3; and for P16, σ = 8, ρ = 7. We suggest [14, 17, 16]
and references therein for information above and details about these spaces.
∗E-mail: [email protected]
†E-mail: [email protected]. Partially supported by FAPESP, grant # 2016/02847-9
σmZM
bfk,j :=
Restricting ourselves to m ≥ 2 and 1 ≤ p ≤ ∞ we denote by p′ its exponent conjugate, i.e.,
1/p+1/p′ = 1. We write (Lp(M),k·kp) the usual Banach spaces of p-integrable complex functions on
M. In particular, L2(M) is a Hilbert space with the inner product h·,·i2 defined by the normalized
(by σm the volume of M) integral between square-integrable functions.
The Laplace-Beltrami operator on M has a discrete spectrum which arranged in an increasing
order is given by {k(k + α + β + 1) : k = 0, 1, . . .}. For each k the eigenspace Hm
k attached to
k(k + α + β + 1) has finite dimension dm
k and they are mutually orthogonal. If we write
k , then {Yk,j : k = 0, 1, . . . , j = 1, 2, . . . dm
{Yk,j : j = 1, 2, . . . , dm
k }
is an orthonormal basis of L2(M). This permits us to consider naturally Fourier expansions on
L2(M). On the sphere all those objects are the well known space of spherical harmonics in m + 1
variables and degree k ([7]).
k } for an orthonormal basis of Hm
k := dimHm
The Fourier coefficients of a function f ∈ Lp(M) are defined by
j = 1, 2, . . . , dm
k ,
f (y) Yk,j(y) dy,
1
k = 0, 1, . . . .
We write St(·) for shifting operator (see [4, 17]) on L2(M), which is defined by the average of a
function in a "ring" of M, namely for each x ∈ M the set is σx
t := {y ∈ M : d(x, y) = t}, 0 < t < π,
with the induced measure. Then the addition formula also available in this context implies the
following Fourier expansion of the shifting operator on L2(M):
St(f ) ∼
∞Xk=0
Q(α,β)
k
(cos t)Yk(f ),
f ∈ L2(M),
(1.1)
k
denotes the normalized Jacobi polynomial, it means Q(α,β)
(1) = 1, and Yk is the
k , k = 0, 1, . . .. All the facts mentioned above can be found constructed
where Q(α,β)
projection of L2(M) onto Hm
and/or explored in the cited references and [14, 16].
We write Br(f ) to denote the fractional derivative of order r which is defined on M in the
distributional sense and given by Br(f ) ∼ P∞
k=0(k(k + α + β + 1))r/2 Yk(f ), we are allowed to
consider the Sobolev class
k
p (M) := {f ∈ Lp(M) : Br(f ) ∈ Lp(M)} ,
W r
endowed which the usual norm k · kW r
p := k · kp + kBr(·)kp.
Let us consider r > 0, t > 0 and f ∈ Lp(M). We introduce the Peetre-type K-functional of
fractional order r given by
The r-th moduli of smoothness
And the generalized shifting operator
Kr(f, t)p := inf
g∈W r
p (M)nkf − gkp + trkgkW r
po .
ωr(f, t)p := supnk(I − Ss)r/2(f )kp : s ∈ (0, t]o .
r − j(cid:19)Sjt(f ),
(−1)j(cid:18) 2r
Sr,t(f ) := −2
∞Xj=1
(cid:0)2r
r(cid:1)
2
(1.2)
(1.3)
(1.4)
where for r, s real numbers
(cid:18)r
s(cid:19) =
Γ(r + 1)
Γ(s + 1)Γ(r − s + 1)
,
for s 6∈ Z−, (cid:18)r
0(cid:19) = r and (cid:18)r
s(cid:19) = 0,
for s ∈ Z−.
It is not difficult to see that this operator well defined and a bounded operator on Lp(M), 1 ≤ p ≤ ∞.
Platonov showed that the K-functional of fractional order and the moduli of smoothness are
related in a asymptotic sense. Notation A(t) ≍ B(t) stands for B(t) . A(t) and A(t) . B(t), while
A(t) . B(t) means that A(t) ≤ c B(t), for some constant c ≥ 0 not depending upon t.
Theorem 1.1. ([17, Theorem 1.2]) For 1 < p < ∞ and r ≥ 1 a natural number, it holds
K2r(f, t)p ≍ ω2r(f, t)p
f ∈ Lp(M), t > 0.
Our main interest on these tools is its relation with the decay of Fourier coefficients of functions
in terms of the rate of approximation of generalized shifting operator. It has shown extremely an
important and efficient tool to get good estimates for both Fourier coefficients of functions satisfying
a generalized Holder condition and eigenvalues sequences of positive integral integral operators with
Holderian kernels (see [12]). The relation we have stablished is the following.
Theorem 1.2. For 1 < p < ∞ and r ≥ 1 a natural number, it holds
K2r(f, t)p ≍ kSr,t(f ) − fkp,
f ∈ Lp(M), t > 0.
This extends the spherical version of it which can be found in [6]. A simple adaptation of
well-known results on spherical functions, namely the Hausdorff-Young inequality, on M play an
important role to show the following decay of Fourier coefficients. He fix the notation sk(f ) :=
k
Pdm
j=1 bfk,j2, k = 0, 1, . . ..
Theorem 1.3. Let r be a positive interger. If p ∈ (1, 2], then
k )(2−p′)/2p′
(min{1, tk})rq[sk(f )]p′/2)1/q
. kSr,t(f ) − fkp,
f ∈ Lp(M),
(dm
( ∞Xk=1
k≥0n(dm
sup
The inequality above becomes an equality in the case p = 2. And, if p = 1, then
k )−1/2(min{1, tk})rq[sk(f )]1/2o . kSr,t(f ) − fk1,
f ∈ L1(M).
Results above permit us to analyze the asymptotic behavior of eigenvalues sequences {λn}n of
certain integral operators generated by Holderian kernels having a Mercer-like series expansion. For
a historical review on the spherical setting we suggest see [12].
We will be dealing with integral operators LK(f ) = RM K(·, y)f (y) dy, having the generating
kernel K : M × M −→ C belonging to L2(M × M). It is easily seen that LK defines a compact
operator on L2(M). The study concerns to kernels on M × M of the form:
K(x, y) =
∞Xk=0
dm
kXj=1
ak,j Yk,j(x) Yk,j(y),
∞Xk=0
dm
kXj=1
ak,j < ∞, x, y ∈ M.
(1.5)
3
We work under two basic conditions: the first one, called positivity, means that the expansion
coefficients are non-negative, i.e., ak,j ≥ 0; and the second one, called monotonicity, means that the
expansion coefficients are monotone decreasing with respect to k, i.e., ak+1,j ≤ ak,j′, 1 ≤ j, j′ ≤ dm
k .
Recently, Berg and collaborators ([2]) showed the characterization given by Schoenberg ([19])
for continuous zonal positive definite kernels on the sphere as series expansion given by formula
(1.5) with coefficients do not depending on index j and satisfying the positivity definition above for
positive definite kernels, holds in a general setting, namely on products of compact Gelfand pairs
with locally compact groups. Therefore, assumptions made here on compact two-point homogeneous
spaces are very natural and expected in most of the applications.
The first application is continuation of the designed in [12]. We say that a kernel K on M
satisfies the (B, β)−Holder condition if there exist a β ∈ (0, 2] and a function B in L1(M) such that
(1.6)
St(K(y,·))(x) − K(y, x) ≤ B(y) tβ,
x, y ∈ M, t ∈ (0, π).
It is not hard to see that this definition is a generalized version the usual Holder condition.
Assumption of positivity on K implies self-adjointness of the integral operator, then the standard
spectral theorem is applicable and we obtain a sequence of nonnegative real numbers {λn(LK )}n
which is the eigenvalues sequence of LK.
Theorem 1.4. Let LK be the integral operator induced by a kernel K as in (1.5) and under
assumptions of positivity and monotonicity. If K satisfies the (B, β)-Holder condition, then it holds
λn(LK) = O(n−1−β/m),
as n → ∞.
For a positive real number r, we write Br,0K for the action of the fractional derivative operator
only applied to the first variable. Also, K y denotes the function · 7→ K(·, y), y ∈ M. We warn
the reader that the terminology "trace-class", common in operators theory appears in the next
corollary, it basically means that the trace of the operator is finite and independent of the choice
of basis. See [5], for details and applications.
Corollary 1.5. Let LK be the integral operator induced by a kernel K as in (1.5) and under
assumptions of positivity, monotonicity and such that for a fixed r > 0, all K y belong to W 2r
2 (M).
If the integral operator generated by B2r,0K is trace-class, then
λn(LK) = O(n−1−2r/m),
as n → ∞.
This extends both Theorem 2.5 in [5] and Theorem 3 in [12] for compact two-point homogeneous
spaces. We present it as consequence of previous theorem not because it is an immediate consequence
of it, but we apply similar techniques in order to prove it, though.
The paper is organized as following. Section 2 contains a short description of the Hausdorff-
Young type inequality its implication on the relation between the decay of Fourier coefficients of
a function and the proof of Theorem 1.3. Several technical lemmas are proved in order to present
the proof of Theorem 1.2. A technique involving relations between the decay of Fourier coefficients
and eigenvalues sequences of the operator is employed to prove 1.4 and 1.5. In Section 4 we give
a shortly background for Kolmogorov n-widths and apply our achievements to get sharp estimates
for the Kolmogorov n-width of RKHS of Holderian kernels. Finally, an example is given.
4
2 Decay of Fourier coefficients
In this section we present some background material in order to prove Theorema 1.3 and 1.2. They
include realization theorem, moduli of smoothness and the associated K-functional as well. Rela-
tions between these were proved recently by Dai, Ditzian and Platonov on two-point homogenous
spaces and play an important role here. References are [4, 9, 16, 17].
A linear operator T on Lp(M) is called a multiplier operator if there exists a sequence {µk}k of
complex numbers such that Yk(T f ) = µk Yk(f ), k = 0, 1, . . ., for any f ∈ Lp(M) and T is bounded.
In this case the sequence {µk}k is called the sequence of multipliers of T .
An important property involving the K-functional is the Realization Theorem for Kr(f, t)p
([9]), which is given by the relation below. In its statement, the multiplier operator ηt depends
upon a best approximation function η ∈ C ∞[0,∞) such that η = 1 in [0, 1], η = 0 in [2,∞) and
η(s) ≤ 1, s ∈ (1, 2). The operator ηt is defined by the formula ηt(f ) = P∞
k=1 η(tk)Yk(f ) for all
f ∈ Lp(M). For r > 0 and f ∈ Lp(M) Realization Theorem ([9]) assures that the K-functional
Kr(f, t)p assumes its infimum via the operator ηt as bellow:
kf − ηt(f )kp + tr kηt(f )kW r
p ≍ Kr(f, t)p,
t > 0.
(2.1)
Lemma 2.1. (Hausdorff-Young type inequality) Let q be the conjugate exponent of p. Then
and
( ∞Xk=1
(dm
k )(2−q)/2q [sk(f )]q/2)1/q
k≥0n(dm
sup
k )−1/2[sk(f )]1/2o . kfk1,
f ∈ L1(M).
. kfkp,
1 < p ≤ 2, f ∈ Lp(M);
The proof is based on the Riez-Thorin interpolation and nothing different from the spherical
setting (see [10], for example) that is why it is omitted here.
The following theorem relates the decay of the Fourier coefficients of a function to the rate of
approximation of operator defined in formula (1.4). In [11] a proof of similar result is presented for
a multiplier operator on the spherical setting and it is slightly different from below.
Proof of Theorem 1.3. Formula (1.1) implies
It means that Sr,t is a multiplier operator and its multiplier sequence {mr(k, t)}k is given by
(cos(jt)),
k = 0, 1 . . . , t > 0.
(2.2)
r − j(cid:19)Q(α,β)
k
(cos(jt)) Yk,
t > 0.
Sr,t =
∞Xk=0
mr(k, t) := −2
(−1)j(cid:18) 2r
rXj=1
−2
(cid:0)2r
r(cid:1)
(−1)j(cid:18) 2r
rXj=1
(cid:0)2r
r(cid:1)
r − j(cid:19)Q(α,β)
k
Fixing f ∈ Lp(M), the linearity of the orthogonal projections imply that
k = 0, 1 . . . ,
Yk(Sr,t(f ) − f ) = (mr(k, t) − 1)Yk(f ),
5
whence
dm
kXj=1
\(Sr,t(f ) − f )k,jYk,j = (mr(k, t) − 1)
fk,jYk,j,
k = 0, 1 . . . .
dm
kXj=1
Computing the L2-norm of both sides we obtain sk(Sr,t(f )− f ) = (mr(k, t)− 1)2sk(f ), k = 0, 1 . . . ,
that is,
k )(2−q)/2q [sk(Sr,t(f ) − f )]q/2 = (dm
(dm
k )(2−q)/2qmr(k, t) − 1q[sk(f )]q/2.
Taking in account that 1 − mr(k, t) ≍ (min{1, tk})r (this important and nontrivial equivalence is
obtained in Lemma 2.6) we have
k )(2−q)/2q [sk(Sr,t(f ) − f )]q/2 = (dm
(dm
≍ (dm
k )(2−q)/2qmr(k, t) − 1q[sk(f )]q/2
k )(2−q)/2q(min{1, tk})rq[sk(f )]q/2.
Finally Hausdorff-Young type formula reach us to the first inequality in the statement of the
theorem. As for the equality assumption in the case p = 2, it suffices to apply Parseval's identity
in the equality above. The inequality in the case p = 1 is settled in a similar fashion.
Theorem above and Theorem 1.2 permit us to choose the more convenient tool in order to study
the decay of Fourier coefficients. We can also relate the decay of the Fourier coefficients of a function
to the K-functional defined in (1.2). Ditzian [9] proved this theorem on the spherical setting and
for the special case in which r is a positive integer. He remarks that the same proof can be slightly
modified to fit for r been a real number. Since the proof is a direct application of Hausdorff-Young
type inequality (Lemma 2.1) we choose do not reproduce it here (it can be founded in [12, p. 9]on
the spherical setting).
Proposition 2.2. If f belongs to Lp(M), 1 ≤ p ≤ 2, q is the conjugate exponent of p and r > 0,
then
k )(2−q)/2q (min{1, tk})rq [sk(f )]q/2)1/q
(dm
( ∞Xk=1
. Kr(f, t)p,
t > 0.
(2.3)
2.1 Proof of Theorem 1.2
The technic employed here is to get good estimates for the multiplier sequence attached to averaged
operator and through an application of the Marcinkiewicz's Multiplier Theorem we prove Theorem
1.2. This proof is highly technical and we present it by steps.
The first technical lemma brings estimates for the difference operator applied to Jacobi poly-
for a sequence {bk}k, we set
nomials. The difference operator is defined inductively as follows:
△0bk = bk and △bk = bk+1 − bk from that if j is a positive integer △jbk = △(△j−1bk).
Lemma 2.3. If α ≥ β ≥ − 1
2 , then
k
(cid:12)(cid:12)(cid:12)△jQ(α,β)
(cos t)(cid:12)(cid:12)(cid:12) .(cid:26)
tj,
(kt)−(α+1/2),
kt ≤ 1
kt ≥ 1.
Its proof follows directly from [8, Lemma 2]. We apply it in order to obtain related estimates
for the sequence {mr(k, t)}k as follows.
6
Lemma 2.4. Let {mr(k, t)}k be the multiplier sequence of operator Sr,t. If 0 < t ≤ π
positive integer, then it holds
2r and j is a
tj,
tj(kt)−(α+1/2),
0 < kt ≤ 1
kt ≥ 1.
Proof. For each k the representation of mr(k, t), given in formula (2.2), implies the following
inequality
(cid:12)(cid:12)△jmr(k, t)(cid:12)(cid:12) .(cid:26)
(cid:12)(cid:12)△jmr(k, t)(cid:12)(cid:12) ≤
(cid:0)2r
r(cid:1)
2
rXj=1(cid:12)(cid:12)(cid:12)(cid:12)(cid:18) 2r
r − j(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)△jQ(α,β)
k
(cos(jt))(cid:12)(cid:12)(cid:12) .
An application of Lemma 2.3 is enough to complete the proof.
The next result asserts that we can represent the normalized Jacobi polynomial by a sum
of cosines with nonnegative coefficients (we warn the reader that it does not hold for all Jacobi
polynomial, see [1] for details). The proof for this fact can be found in [1, p. 63–66] and we just
need to observe that the one-dimensional unit sphere S1 is isometrically embedded in any M.
Lemma 2.5. If α and β are as described Section 1, then
Q(α,β)
k
(cos θ) =
kXv=0
[v/2]Xi=0
av,i (cos(v − 2i)θ),
k = 0, 1, . . . ,
where av,i ≥ 0, v = 0, 1, . . . , k and i = 0, 1, . . . , v.
Explicitly, a simple calculation for v = 0, 1, . . . , k and i = 0, 1, . . . , v, implies
av,i = bv,i
Γ(k + α + 1)Γ(α − β + 1)Γ(2α + v + 1)Γ(2α + 2v + 2)Γ(k + α + β + v + 1)
Γ(α − β − k + v + 1)Γ(k − v + 1)Γ(α + v + 1)Γ(2α + 2v + 1)Γ(k + 2α + v + 2)
,
where Γ(·) stands the Gamma function and bv,i = cv,iP (α,α)
v
Gegenbauer polynomial with index α representation ([20, p. 93]) in terms of cosine.
(1)/P (α,β)
k
(1) and cv,i are given by the
The main idea behind the next result is that if {mr(k, t)}k is the sequence of multipliers of Sr,t,
then 1 − mr(k, t) ≍ (min{1, tk})r.
Lemma 2.6. For t ∈ [0, π/2] it holds
0 < a ≤
1 − mr(k, t)
(kt)2r
≤ b < ∞,
for 0 < kt ≤ π,
where a and b are constants. Additionally, for any τ > 0 there exists vr,τ < 1 such that
mr(k, t) ≤ vr,τ ,
for kt ≥ τ > 0.
(2.4)
(2.5)
Proof. The main ideia of the proof is borrowed from [6, Lemma 4.4] but several considerations
are needed. By Lemma 2.5 we have
1 − mr(k, t) = 1 +
2
(cid:0)2r
r(cid:1)
(−1)j(cid:18) 2r
rXj=1
r − j(cid:19) kXv=0
[v/2]Xi=0
7
av,i (cos(v − 2i)jt).
From above the representation of a power of the sine function in terms of cosine implies
1 − mr(k, t) =
kXv=0
[v/2]Xi=0
4r
(cid:0)2r
r(cid:1)
av,i(cid:20)sin2r(cid:18)(v − 2i)
t
2(cid:19)(cid:21) .
If kt ≤ π, then sin2r(cid:16) (v−2i)t
2 (cid:17) ≤ (kt/2)2r. Which implies
av,i(cid:18) kt
2(cid:19)2r
4r
[v/2]Xi=0
kXv=0
(cid:0)2r
r(cid:1)
1 − mr(k, t) ≤
=
(kt)2r
(cid:0)2r
r(cid:1) ,
and the proof of the right-hand side of inequality (2.4) follows.
On the other hand, starting from
4r
(cid:0)2r
r(cid:1)
1 − mr(k, t) =
av,i(cid:20)sin2r(cid:18)(v − 2i)
[v/2]Xi=0
kXv=0
for each index i ≤ [v/4] it holds sin2r(cid:16) (v−2i)t
2 (cid:17) ≥ (vt/2π)2r. Then, we have for some positive
av,i (cid:18) vt
2π(cid:19)2r
[v/4]Xi=0
1 − mr(k, t) ≥
2(cid:19)(cid:21) ,
constant c that
kXv=1
≥ c
4r
t
.
4r
r(cid:1)(cid:18) kt
2π(cid:19)2r
(cid:0)2r
Which finishes the proof of inequalities in formula (2.4) in the statement of the lemma.
r(cid:1)
(cid:0)2r
Let any 0 < τ ≤ π and kt ≤ π, inequality (2.4) implies
0 < a ≤
1 − mr(k, t)
(kt)2r
≤
1 − mr(k, t)
τ 2r
.
And, therefore mr(k, t) ≤ 1 − aτ 2r, if choose vr,τ = 1 − aτ 2r < 1, it is (2.5)for 0 < τ ≤ kt.
Now if kt ≥ π we have
Writing c′ =Pk
v=0Pi∈I(k) av,i we have 1 − mr(k, t) ≥ 4r
Now we present some more properties of the difference operator.
(2r
r )
c′ And the lemma is proved.
Lemma 2.7. Let {ak}k, {bk}k sequences of real numbers and j a positive integer.
8
where
I(k) :=
1 − mr(k, t) ≥
4r
kXv=0 Xi∈I(k)
r(cid:1)
(cid:0)2r
2 ][l=0 (cid:26)i : 0 ≤ i ≤ [v/2];
π
4
[ vt
2π − 1
av,i(cid:20)sin2r(cid:18)(v − 2i)
t
2(cid:19)(cid:21) ,
+ lπ ≤ (v − 2i)
t
2 ≤
3π
4
+ lπ(cid:27) .
a) It holds
b) If the sequence {ak}k satisfies ak ≥ a > 0, k = 0, 1, . . ., then
i(cid:19)(△j−iak)(△ibk+j−i).
△j(akbk) =
jXi=0(cid:18)j
i(cid:19)△ia−1
k △j+iak+i ≤ c max
0≤i≤j △ia−1
k △j+iak+i,
△ja−1
k ≤
1
ak
where c = 2j/a.
j−1Xi=0(cid:18)j
The proof of item a) follows by mathematical induction. For the part b) note that a−1
k ak = 1,
choosing the sequences {a−1
Lemma 2.8. If t ∈ [0, π/2], then for any positive integer j and τ > 0 such that 0 < kt < τ the
following holds
k }k and {ak}k we apply item a) and the proof follows.
(cid:12)(cid:12)(cid:12)(cid:12)△j
1 − mr(k, t)
(k(k + α + β + 1)t2)r(cid:12)(cid:12)(cid:12)(cid:12) .(cid:2)k−j + k−j−1(cid:3) .
(2.6)
The proof is omitted since it can found in [7, p. 255-258]. We observe that inequality (2.6) holds
for the sequence of multiplicative inverse sated as well.
The very last result we need in oder to present the proof of Theorem 1.2 is the Marcinkiewicz's
Multiplier Theorem for the compact two-point homogeneous spaces. This result gives us a sufficient
condition such that a given operator constructed via sequences (multipliers) be bounded.
Theorem 2.9. [3, Theorem 7.1] Let M be a compact two-point homogeneous space of dimension
m and {µj}j a sequence of real numbers satisfying
i) supj{µj} ≤ M < ∞;
ii) supjn2j(s−1)P2j+1
l=2j △s µlo ≤ M < ∞, with s = (m + 1)/2 if s is odd and s = (m + 2)/2 if s
µk Yk(f )(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
∞Xk=0
where ap is a constant which does not depend on f .
f ∈ Lp(M), 1 < p < ∞,
≤ ap M kfkp,
is even. Then, it holds
Proof of Theorem 1.2. In order to prove the equivalence stated, it is enough to show that for
some positive a the following three inequalities hold
and
kf − ηatfkp . kf − Sr,tfkp,
t2r kB2r(ηatf )kp . kf − Sr,tfkp
kηatf − Sr,t(ηatf )kp . t2r kB2r(ηatf )kp.
9
(2.7)
(2.8)
(2.9)
Also, observe that (2.7) is assured if
kf − ηatf − (I + Sr,t + ··· + S4
r,t) (I − ηat) (f − Sr,tf )kp . kf − Sr,tfkp.
(2.10)
To obtain inequality above we need to show that
µk,1 := (1 − η(atk))
mr(k, t)4
1 − mr(k, t)
,
k = 0, 1, . . . ,
define sequence satisfying conditions in Theorem 2.9. We first note that if atk ≤ 1, η(atk) = 1 and
then, µk,1 = 0. If atk ≥ τ > 1, from Lemma 2.6 we have 1− mr(k, t) ≥ cτ,r > 0, also it is clear that
1 − η(atk) ≤ c for some constant c and since {(1/kt)4α+4/2−s}k is bounded for 4α + 4/2 − s ≥ 0,
Lemma 2.4 implies
△sµk,1 . △smr(k, t)4 .
ts
(kt)4α+4/2
From the inequality above, we have
sup
2j(s−1)
. sup
2j(s−1)
j
2j+1Xk=2j △s µk,1
j
ks .
1
ks .
.(cid:18) 1
kt(cid:19)4α+4/2−s 1
ks ≤ sup
1
j
2j+1Xk=2j
2j(s−1)
2j+1Xk=2j
1
(2j)s . 1.
And therefore Theorem 2.9 assures that {µk,1}k is a multiplier sequence and inequality (2.7) is
proved.
Heading to inequality (2.9), we proceed analogously as above but taking in account the conve-
nient multiplier sequence. For 0 < kt < τ , τ > 0 let
µk,2 :=
1 − mr(k, t)
(k(k + α + β + 1))rt2r η(atk),
k = 1, 2, . . .
be the sequence of multipliers to application involved in inequality (2.9). We write
ak :=
1 − mr(k, t)
(k(k + α + β + 1))rt2r
and
bk = η(atk),
k = 1, 2, . . . .
And we note that △iη(atk) . (at)i, for any i a positive integer. By Lemma 2.7 part a) and Lemma
2.8, respectively, we reach to
△sµk,2 .
We still need to verify that the following holds true
ks+1 +
1
ks(cid:19) .
ks+1 +
sXi=0(cid:18)s
i(cid:19)(cid:18) 1
j
2j(s−1)
sup
1
ks(cid:19) (kt)s .(cid:18) 1
2j+1Xk=2j △s µk,2
. c,
10
for some constant c. In fact, from previous estimates, we have
sup
2j(s−1)
j
2j+1Xk=2j △s µk,2
. sup
j
2j(s−1)
2j+1Xk=2j(cid:18) 1
ks+1 +
. sup
j (cid:26) 1
2j + 1(cid:27) .
1
ks(cid:19)
Thus, Theorem 2.9 implies that {µk,2}k is a multiplier sequence and inequality (2.9) holds.
Finally, we show that inequality (2.8) holds from we showing that for 0 < kt < τ and τ > 0,
µk,3 :=
(k(k + α + β + 1))rt2r
1 − mr(k, t)
η(atk),
k = 0, 1, . . . ,
is a multiplier sequence fitting in Theorem 2.9. We observe that in Lemmas 2.8 and 2.7, part a), we
can replace sequence {ak}k by {a−1
k }k. Also, Lemma 2.7, part b) fits in our context for the sequence
bellow and
ak :=
(k(k + α + β + 1))rt2r ≥ c > 0, k = 0, 1, . . . ,
1 − mr(k, t)
for some constant c by Lemma 2.6. More than that a−1
k ≥ b−1 > 0, k = 0, 1, . . ., where b is the
constant in formula (2.4). Taking in account remark right after the statement of Lemma 2.7 it is
not hard to see {µk,3}k is a multiplier sequence. The theorem is proved.
3 Application: decay of eigenvalues sequences
Our goal in this section is to prove both Theorem 1.4 and Corollary 1.5. To present them we will
first derive some additional technical results as following. We remind readers that the kernels K
we are dealing with satisfy all assumptions made Section 1.
We start noting that positivity of the kernel Kassures that the operator LK is positive and has
K whose kernel K1/2 has the following series expansion
a uniquely defined square root operator L1/2
kXj=1
K1/2(x, y) =
∞Xk=0
dm
a1/2
k,j Yk,j(x) Yk,j(y),
x, y ∈ M.
(3.1)
Both LK and L1/2
K are self-joint positive operators. The definition of the integral operator generated
by K makes easy to see that the spherical harmonics Yk,j, j = 1, 2, . . . , dm
k and k = 0, 1, . . ., are
all eigenvectors of the operator LK associated to the eigenvalues ak,j, respectively. Since we have
made a monotonicity assumption on coefficients of K it gives us an eigenvalue sequence ordering
that is suitable for our analysis.
For each y ∈ M, the Fourier coefficients of the function K y := K(·, y) are (cK y)k,j = ak,j Yk,j(y),
k and k = 0, 1, . . .. Considering the kernel K1/2 (formula (3.1)) in a similar way we
k and k = 0, 1, . . ., which implies
k,j Yk,j(y), j = 1, 2, . . . , dm
1/2)k,j = a1/2
j = 1, 2, . . . , dm
have its Fourier coefficients ([K y
that
ZM
sk(K y
1/2) dy =
ak,j,
k = 0, 1, . . . .
(3.2)
dm
kXj=1
11
The action of the fractional derivative on K y
1/2 is given by
Br(K y
1/2) ∼
a1/2
k,j (k(k + α + β + 1))r/2 Yk,j(y) Yk,j,
y ∈ M,
∞Xk=0
dm
kXj=1
ZM kSt(K y
1/2)(cid:13)(cid:13)(cid:13)
and it permits us, with simple calculation, to derive the following(cid:13)(cid:13)(cid:13)Br(K y
any y ∈ M. Also, if K is (B, β)-Holder (see [12]), then
2
2
= B2r,0K(y, y), for
1/2) − K y
1/2k2
2 dy . tβ,
y ∈ M.
(3.3)
Proof of Theorem 1.4 This proof can be found on the spherical setting in [12]. Since from this
point it is exactly the same one presented in this reference we just draw some steps of it.
By Proposition 2.2 for p = q = 2 and r = 2 and Theorem 1.2 we have
∞Xk=1
(min{1, tk})4 sk(K z
1/2) . kSt(K z
1/2) − K z
1/2k2
2,
Integrating both sides and making use of (3.3) we haveP∞
k=0(min{1, tk})4Pdm
∞Xk=n
this inequality (for t = 1/n) we get
an ≤ nβ+m−1
nβ+m an = nβ+m−1
2n−1Xk=n
ak ≤ C3, n = 1, 2, . . . ,
z ∈ M, t ∈ (0, π).
k
j=1 ak,j . tβ. Handling
or, equivalently, an = O(n−β−m), as n → ∞. Returning to our original notation for the eigenvalues
of LK and recalling that {λn(LK)}n decreases to 0, we have that an = λdm+1
(LK), n = 1, 2, . . .. In
particular,
n
Therefore, the decay in the statement of the theorem follows.
λdm+1
n
(LK ) = O(n−β−m), n → ∞.
Proof of Corollary 1.5 Most of steps in this proof are essentially repetitions of previous theorem
that is why we omitted it here. By Proposition 2.2 (p = q = 2 and the function K z
1/2) we have
∞Xk=0
(min{1, tk})2r sk(K z
1/2) .hωr(K z
1/2, t)2i2
,
z ∈ M, t ∈ (0, π).
Since K z
Then, we have
1/2 ∈ W 2r
2 , Proposition 4.2 in [17] asserts that ωr(K z
1/2, t)2 . t2r kBr(K z
1/2)k2, z ∈ M.
(min{1, tk})2r(cid:18)ZM
∞Xk=0
sk(K z
1/2) dz(cid:19) . t2rZM kBr(K z
1/2)k2
2 dz,
t ∈ (0, π).
Since B2r,0K is the kernel of a trace-class operator kBr(K z
lations analogous to before finishes the proof.
1/2)k2
2 is a nonnegative constant. Calcu-
12
4 Kernel-based spaces and example
The last application to be presented corroborates with Theorem 6 in [18]. There the authors work
on a context including just the Euclidean space and a bounded with smooth enough boundary
domains and they recover decay rates for sequences of eigenvalues of integral operators from the
decay of n-widths. Technique applied here is completely different from that since we make opposite
way.
The Kolmogorov n-width of a subset A of a Hilbert space (H,h·,·iH ) is defined as follows
dn(A; H) := inf
Vn⊂H
sup
f ∈A
fn∈Vn kf − fnkH ,
inf
(4.1)
where k · kH in the induced norm by the inner product in H and the first infimum above is taken
over all subspaces Vn having dimension n in H.
Additionally, under assumptions made here and continuity of the kernel K the classical Mercer's
Theorem assures that integral operator has a sequence of positive eigenvalues {λi} ordered in a
decreasing way, related to a sequence of eigenfunctions {ϕi}. More than that the kernel K can be
written as
K(x, y) =
λiϕi(x)ϕi(y),
x, y ∈ M,
(4.2)
∞Xi=1
where the sum is absolutely and uniformly convergent. This result permits us to characterize
(HK,h·,·iK ), which is the unique reproducing kernel Hilbert space (RKHS) attached to K, since
{√λi ϕi} is an orthonormal basis of it, (a complete reference for this basic facts are papers authored
by R. Schaback).
Let us denote by dn the n-width dn(S(HK ); L2(M)). Then we have ([15, Corollary 2.6])
dn = inf
Vn⊂L2
sup
f ∈S(HK ) kf − Pn(f )k2 =pλn+1,
(4.3)
where S(HK ) is unit ball in HK, and the projections Pn are defined by Pn(f ) :=Pn
n = 1, 2, . . ., where {hi
span {√λi ϕi : i = 1, . . . , n}, is the unique optimal space.
kernel K : M × M −→ R, the Kolmogorov n-width dn, defined in formula (4.1) is equivalent to
i=1hf, hiiK hi,
: i = 1, 2, . . . , n} is an orthonormal basis of Vn. Moreover, Hn =
According to Santin and Schaback ([18, p. 979]) if we consider a positive definite and symmetric
κn := inf
Vn⊂HK
sup
f ∈S(HK )kf − Pn(f )k2,
where the infimum above is taken over all subspaces Vn having dimension n in HK . We bring up
the following characterization for the Kolmogorov n-width, on this context.
Proposition 4.1. ([18, Theorem 2]) If Hn = span {√λi ϕi : i = 1, . . . , n}, the subspace of HK,
then κn =pλn+1. Moreover, Hn is the unique optimal space.
Theorem 4.2. Let K be a continuous, positive definite and symmetric kernel defined on M. If K
satisfies the (B, β) - Holder condition, then
The proof from a simple application of Proposition 4.1 and Theorem 1.4.
κn = O((n + 1)−1/2−β/2m), n → ∞.
13
4.1 A concrete case: example
The example bellow is a constructive way to consider a kernel to show the decay rates for the
integral operator generating for it fits into assumptions of our theorems.
Let ǫ > 0 be fixed and suppose mǫ > 1 and K is a kernel having expansion in the form
K(x, y) ∼ 1 +
where cos t = d(x, y), and
∞Xn=1
cn
nm(1+ǫ)+2r−1
P (α,β)
n
(cos t),
x, y ∈ M,
(4.1)
cn =
Γ(β + 1)(2n + α + β + 1)Γ(n + α + β + 1)
Γ(α + β + 2)Γ(n + β + 1)
.
The harmonic expansion of K is easily obtained with the help of the addition formula and it is not
hard to see that
1 +
∞Xn=1
cn
nm(1+ǫ)+2r−1 ≤ 1 + C
1
nmǫ+2r < ∞,
∞Xn=1
for some constant C depending on m. It means that the series expansion of K (4.1) converges
uniformly to K(x, y) and then K is continuous. Also its integral operator is positive, since K is
positive definite. The integral operator generated by B2r,0K is trace-class and K fits into Corollary
1.5. Therefore λn(LK ) = O(n−1−2r/m), as n → ∞. Also, Theorem 4.2 is applicable and then the
Kolmogorov n-width of Hn for example above decay as
κn = O((n + 1)−1/2−β/2m), n → ∞.
References
[1] R. Askey, Orthogonal polynomials and special functions. Society for Industrial and Applied
Mathematics, Philadelphia, Pa., 1975.
[2] C. Berg, A.P. Peron, E. Porcu, Orthogonal expansions related to compact Gelfand pairs. Ex-
positiones Mathematicae, 2017.
[3] A. Bonami, J.L. Clerc, Sommes de C`esaro et multiplicateurs des d´eveloppements en har-
moniques sph´eriques. Trans. Amer. Math. Soc. 183 (1973), 223–263.
[4] G. Brown, F. Dai, Approximation of smooth functions on compact two-point homogeneous
spaces. J. Funct. Anal. 220 (2005), no. 2, 401–423.
[5] M.H. Castro, V.A. Menegatto, Eigenvalue decay of positive integral operators on the sphere.
Math. Comp. 81 (2012), no. 280, 2303–2317.
[6] F. Dai, Z. Ditzian, Combinations of multivariate averages. J. Approx. Theory 131 (2004),
no. 2, 268–283.
14
[7] F. Dai, Y. Xu, Approximation theory and harmonic analysis on spheres and balls. Springer
Monographs in Mathematics. Springer, New York, 2013.
[8] F. Dai, K. Wang, C. Yu, On a conjecture of Ditzian and Runovskii, J. Approx. Theory 118
no. 2 (2002) 202–224.
[9] Z. Ditzian, Fractional derivatives and best approximation. Acta Math. Hungar. 81 (1998), no.
4, 323–348.
[10] Z. Ditzian, Relating smoothness to expressions involving Fourier coefficients or to a Fourier
transform. J. Approx. Theory 164 (2012), no. 10, 1369–1389.
[11] T. Jordao, V. A. Menegatto, Estimates for Fourier sums and eigenvalues of integral operators
via multipliers on the sphere. Proc. Amer. Math. Soc. 144 (2016), no. 1, 269–283.
[12] T. Jordao, V. A. Menegatto, X. Sun, Eigenvalue sequences of positive integral operators and
moduli of smoothness. Springer, Cham, 83 (2014), p. 239–254.
[13] T. Kuhn, Eigenvalues of integral operators with smooth positive definite kernels. Arch. Math.
(Basel) 49 (1987), no. 6, 525–534.
[14] A. Kushpel, S.A. Tozoni, Entropy and widths of multiplier operators on two-point homogeneous
spaces. Constr. Approx., 35 (2012), no.2, 137–180.
[15] A. Pinkus, n-Widths in Approximation Theory. Springer-Verlag, Berlin, 1985.
[16] S.S. Platonov, Approximations on compact symmetric spaces of rank 1. (Russian) Mat. Sb.
188 (1997), no. 5, 113–130; translation in Sb. Math. 188 (1997), no. 5, 753–769.
[17] S.S. Platonov, Some problems in the theory of the approximation of functions on compact
homogeneous manifolds. (Russian) Mat. Sb. 200 (2009), no. 6, 67–108; translation in Sb.
Math. 200 (2009), no. 5-6, 845–885.
[18] G. Santin, R. Schaback, Approximation of eigenfunctions in kernel-based spaces. Adv. Com-
put. Math. 42 (2016), no. 4, 973–993.
[19] I. J. Schoenberg, Positive definite functions on spheres. Duke Math. J. 9 (1942), 96–108.
[20] G. Szego, Orthogonal Polynomials. Amer. Math. Soc. New York, 1939.
[21] H-C. Wang, Two point homogeneous spaces. Ann. Math. (2) 55, (1952), 177–191.
A. O. Carrijo & T. Jordao
Departamento de Matem´atica,
ICMC - Universidade de Sao Paulo,
13566-590 - Sao Carlos - SP, Brazil.
15
|
1106.0709 | 2 | 1106 | 2011-10-24T14:15:42 | Asymmetric Covariance Estimates of Brascamp-Lieb Type and Related Inequalities for Log-concave Measures | [
"math.FA"
] | An inequality of Brascamp and Lieb provides a bound on the covariance of two functions with respect to log-concave measures. The bound estimates the covariance by the product of the $L^2$ norms of the gradients of the functions, where the magnitude of the gradient is computed using an inner product given by the inverse Hessian matrix of the potential of the log-concave measure. Menz and Otto \cite{OM} proved a variant of this with the two $L^2$ norms replaced by $L^1$ and $L^\infty$ norms, but only for $\R^1$. We prove a generalization of both by extending these inequalities to $L^p$ and $L^q$ norms and on $\R^n$, for any $n\geq 1$. We also prove an inequality for integrals of divided differences of functions in terms of integrals of their gradients. | math.FA | math |
Asymmetric Covariance Estimates of
Brascamp-Lieb Type and
Related Inequalities for Log-concave Measures
Eric A. Carlen1, Dario Cordero-Erausquin2 and Elliott H. Lieb3
1. Department of Mathematics, Hill Center,
Rutgers University, 110 Frelinghuysen Road Piscataway NJ 08854-8019 USA
2. Institut de Math´ematiques de Jussieu,
Universit´e Pierre et Marie Curie (Paris 6), 4 place Jussieu, 75252 Paris France
3. Departments of Mathematics and Physics, Jadwin Hall,
Princeton University, P. O. Box 708, Princeton, NJ 08542-0708
October, 2011
Abstract
An inequality of Brascamp and Lieb provides a bound on the covariance of two
functions with respect to log-concave measures. The bound estimates the covariance
by the product of the L2 norms of the gradients of the functions, where the magnitude
of the gradient is computed using an inner product given by the inverse Hessian
matrix of the potential of the log-concave measure. Menz and Otto [13] proved a
variant of this with the two L2 norms replaced by L1 and L∞ norms, but only for
R1. We prove a generalization of both by extending these inequalities to Lp and Lq
norms and on Rn, for any n ≥ 1. We also prove an inequality for integrals of divided
differences of functions in terms of integrals of their gradients.
Mathematics subject classification number: 26D10
Key Words: convexity, log-concavity, Poincar´e inequality
1
Introduction
Let f be a C 2 strictly convex function on Rn such that e−f is integrable. By strictly convex,
we mean that the Hessian matrix, Hessf , of f is everywhere positive.
1Work partially supported by U.S. National Science Foundation grant DMS 0901632.
2Work partially supported by U.S. National Science Foundation grant PHY 0965859.
c(cid:13) 2011 by the authors. This paper may be reproduced, in its entirety, for non-commercial purposes.
1
2
Adding a constant to f , we may suppose that
Let dµ denote the probability measure
e−f (x) dnx = 1 .
ZRn
dµ := e−f (x) dnx ,
(1.1)
For any two real-valued functions f, g ∈ L2(µ), the covariance of f and g is the quantity
and let k · kp denote the corresponding Lp(µ)-norm.
gh dµ −(cid:18)ZRn
cov(g, h) :=ZRn
g dµ(cid:19)(cid:18)ZRn
h dµ(cid:19) ,
(1.2)
(1.3)
and the variance of h is var(h) = cov(h, h).
The Brascamp-Lieb (BL) inequality [4] for the variance of h is
var(h) ≤ZRn
(∇h, Hess−1
f ∇h) dµ ,
where (x, y) denotes the inner product in Rn. (We shall also use x · y to denote this same
inner product in simpler expressions where it is more convenient.)
Since (cov(g, h))2 ≤ var(g)var(h), an immediate consequence of (1.3) is
f ∇h) dµ .
(∇h, Hess−1
(∇g, Hess−1
(cov(g, h))2 ≤ZRn
f ∇g) dµZRn
The one-dimensional variant of (1.4), due to Otto and Menz [13], is
cov(g, h) ≤ k∇gk1kHess−1
f ∇hk∞ = sup
x (cid:26)h′(x)
f ′′(x)(cid:27) ZR g′(x) dµ(x)
(1.4)
(1.5)
for functions g and h on R1. They call this an asymmetric Brascamp-Lieb inequality. Note
that it is asymmetric in two respects: One respect is to take an L1 norm of ∇g and an L∞
norm of ∇h, instead of L2 and L2. The second respect is that the L∞ norm is weighted
with the inverse Hessian -- which here is simply a number -- while the L1 norm is not
weighted.
Our first result is the following theorem, which generalizes both (1.4) and (1.5).
1.1 THEOREM (Assymetric BL inequality). Let dµ(x) be as in (1.1) and let λmin(x)
denote the least eigenvalue of Hessf (x). For any locally Lipschitz functions g and h on Rn
that are square integrable with respect to dµ, and for 2 ≤ p ≤ ∞, 1/p + 1/q = 1,
cov(g, h) ≤ (cid:13)(cid:13)Hess−1/p
f ∇g(cid:13)(cid:13)q (cid:13)(cid:13)λ(2−p)/p
min Hess−1/p
f ∇h(cid:13)(cid:13)p .
This is sharp in the sense that (1.6) cannot hold, generally, with a constant smaller than
1 on the right side.
(1.6)
For p = 2, (1.6) is (1.4). Note that (1.6) implies in particular that for Lipschitz functions
3
g, h on Rn,
For p = ∞ and q = 1, the latter is
min ∇g(cid:13)(cid:13)q(cid:13)(cid:13)λ−1/q
cov(g, h) ≤(cid:13)(cid:13)λ−1/p
min ∇h(cid:13)(cid:13)p .
cov(g, h) ≤(cid:13)(cid:13)∇g(cid:13)(cid:13)1(cid:13)(cid:13)λ−1
min∇h(cid:13)(cid:13)∞,
(1.7)
which for n = 1 reproduces exactly (1.5).
We also prove the following theorem. In addition to its intrinsic interest, it gives rise to
an alternative proof, which we give later, of Theorem 1.1 in the case p = ∞ (though this
proof only yields the sharp constant for R1, which is the original Otto-Menz case (1.5)).
1.2 THEOREM (Divided differences and gradients). Let µ be a probability measure with
log-concave density (1.1) . For any locally Lipschitz function h on Rn,
ZRnZRn
h(x) − h(y)
x − y
dµ(x) dµ(y) ≤ 2nZRn ∇h(x) dµ .
(1.8)
1.3 Remark. The constant 2n is not optimal, as indicated by the examples in Section 4
(we will actually briefly mention how to reach the constant 2n/2). We do not know whether
the correct constant grows with n (and then how), or is bounded uniformly in n. We do
know that for n = 1, the constant is at least 2 ln 2. We will return to this later.
The rest of the paper is organized as follows: Section 2 contains the proof of Theo-
rem 1.1, and Section 3 contains the proof of Theorem 1.2, as well as an explanation of
the connection between the two theorems. Section 4 contains comments and examples
concerning the constant and optimizers in Theorem 1.2. Section 5 contains a discussion
of an application that motivated Otto and Menz, and finally, Section 6 is an appendix
providing some additional details on the original proof of the Brascamp-Lieb inequalities,
which proceeds by induction on the dimension, and has an interesting connection with the
application discussed in Section 5.
We end this introduction by expressing our gratitude to D. Bakry and M. Ledoux
for fruitful exchanges on the preliminary version of our work. We originally proved (1.7)
with the constant n2n using Theorem 1.2, as explained in Section 3. Bakry and Ledoux
pointed out to us that using a stochastic representation of the gradient along the semi-
group associated to µ (sometimes referred to as the Bismut formula), one could derive
inequality (1.7) with the right constant 1. This provided evidence that something more
algebraic was at stake. It was confirmed by our general statement Theorem 1.1 and by its
proof below.
4
2 Bounds on Covariance
The starting point of the proof we now give for Theorem 1.1 is a classical dual representation
for the covariance which, in the somewhat parallel setting of plurisubharmonic potentials,
goes back to the work of Hormander. We shall then adapt to our Lp setting Hormander's
L2 approach [8] to spectral estimates.
Let g and h be smooth and compactly supported on Rn. Define the operator L by
L = ∆ − ∇f · ∇ ,
and note that
ZRn
g(x)Lh(x) dµ(x) = −ZRn ∇g(x) · ∇h(x) dµ(x) ,
so that L is self-adjoint on L2(µ). Let us (temporarily) add ǫx2 to f to make it uniformly
convex, so that the Hessian of f is invertible and so that the operator L has a spectral
gap. (Actually, L always has a spectral gap since µ is a log-concave probability measure,
as noted in [9, 1]. Our simple regularization makes our proof independent of these deep
results.)
Then provided
(2.1)
(2.2)
(2.3)
(2.4)
(2.5)
h(x) dµ(x) = 0 ,
etLh(x) dt
ZRn
u := −Z ∞
0
exists and is in the domain of L, and satisfies Lu = h.
Thus, assuming (2.3), and by standard approximation arguments,
cov(g, h) = ZRn
g(x)h(x) dµ(x) =ZRn
= −ZRn ∇g(x) · ∇u(x) dµ(x) .
g(x)Lu(x) dµ(x)
This representation for the covariance is the starting point of the proof we now give for
Theorem 1.1.
Proof of Theorem 1.1: Fix 2 ≤ p < ∞, and let q = p/(p− 1), as in the statement of the
theorem. Suppose h satisfies (2.3), and define u by (2.4) so that Lu = h. Then from (2.5),
cov(g, h) ≤ (cid:12)(cid:12)(cid:12)(cid:12)
ZRn ∇g(x) · ∇u(x) dµ(x)(cid:12)(cid:12)(cid:12)(cid:12)
≤ ZRn(cid:12)(cid:12)Hess−1/p
f ∇g(x) · Hess1/p
f ∇g(x)kq kHess1/p
≤ kHess−1/p
f ∇u(x)(cid:12)(cid:12) dµ(x)
f ∇u(x)kp .
(2.6)
5
estimate:
Thus, to prove (1.6) for 2 ≤ p < ∞, it suffices to prove the following W −1,p -- W 1,p type
(2.7)
min Hess−1/p
f ∇u(x)kp ≤ kλ(2−p)/p
f ∇hkp .
kHess1/p
Toward this end, we compute
L(∇up) = p∇up−2(L∇u) · ∇u
+p∇up−2Tr(Hess2
≥ p∇up−2(L∇u) · ∇u ,
u) + p(p − 2)∇up−4Hessu∇u2
(2.8)
where we have used the fact that p ≥ 2, and where the notation L(∇u) refers to the
coordinate-wise action (L∂1u, . . . , L∂nu) of L.
Then, using the commutation formula (see the remark below)
L(∇u) = ∇(Lu) + Hessf∇u ,
(2.9)
we obtain
0 =ZRn
L(∇up) dµ(x) ≥ pZRn ∇up−2∇u·∇h dµ(x) + pZRn ∇up−2∇u· Hessf∇u dµ(x) ,
and hence
ZRn ∇up−2Hess1/2
f ∇u2 dµ(x) ≤ZRn ∇up−2Hess1/p
f ∇uHess−1/p
f ∇h dµ(x) .
(2.10)
We now observe that for any positive n × n matrix and any vector v ∈ Rn,
A1/pvp ≤ vp−2A1/2v2 .
To see this, note that we may suppose v = 1. Then in the spectral representation of A,
by Jensen's inequality,
A1/pv = n
Xj=1
λ1/p
j v2
j!1/2
≤ n
Xj=1
λ1/2
j v2
j!1/p
.
Using this on the left side of (2.10), and using the obvious estimate
on the right, we have
∇u ≤ λ−1/p
min Hess1/p
f ∇u
kHess1/p
f ∇ukp
p ≤ZRn Hess1/p
f ∇up−1λ(2−p)/p
min Hess−1/p
f ∇h dµ(x) .
(2.11)
Then by Holder's inequality we obtain (2.7).
6
It is now obvious that we can take the limit in which ǫ tends to zero, so that we
obtain the inequality without any additional hypotheses on f . Our calculations so far have
required 2 ≤ p < ∞, however, having obtained the inequality for such p, by taking the
limit in which p goes to infinity, we obtain the p = ∞, q = 1 case of the theorem.
Finally, considering the case in which
dµ(x) = (2π)−n/2e−x2
dx ,
and g = h = x1, we have that Hessf = Id and so
λmin = Hess−1/p
f ∇g = Hess−1/p
f ∇h = 1
for all x, and so the constant is sharp, as claimed.
2.1 Remark. Many special cases and variants of the commutation relation (2.9) are well-
known under different names. Perhaps most directly relevant here is the case in which
f (x) = x2/2. Then ∂j and its adjoint in L2(µ), ∂∗
j = xj − ∂j, satisfy the canonical
commutation relations, and the operator L = −Pn
j=1 ∂∗
j ∂j is (minus) the Harmonic oscil-
lator Hamiltonian in the ground state representation. This special case of (2.9), in which
the Hessian on the right is the identity, is the basis of the standard determination of the
spectrum of the quantum harmonic oscillator using "raising and lowering operators".
In the setting of Riemannian manifolds, a commutation relation analogous to (2.9) in
which L is the Laplace-Beltrami operator and the Hessian is replaced by Ric, the Ricci
curvature tensor, is known as the Bochner-Lichnerowicz formula. Both the Hessian version
(2.9) and the Bochner-Lichnerowicz version have been used a number of times to prove
inequalities related to those we consider here, for instance in the work of Bakry and Emery
on logarithmic Sobolev inequalities.
We note that our proof immediately extends, word for word, to the Riemannian setting
if we use, in place of (2.9) the commutation satisfied by the operator L given by (2.1)
where f is a (smooth) potential on the manifold; That is, with some abuse of notation,
L(∇u) = ∇(Lu) + Hessf∇u + Ric∇u, or rather, more rigorously,
L(∇up) ≥ p∇up−2(cid:2)∇(Lu) · ∇u + Hessf∇u · ∇u + Ric∇u · ∇u(cid:3).
Thus, an analog of Theorem 1.1 holds on a Riemannian manifold M equipped with a
probability measure
dµ(x) = e−f (x) dvol(x)
where dvol is the Riemannian element of volume and f a smooth function on M, provided
Hessf at each point x is replaced in the statement by the symmetric operator
Hx = Hessf (x) + Ricx
defined on the tangent space. Of course, the convexity condition on f is accordingly
replaced by the assumption that Hx > 0 at every point x ∈ M.
3 Bounds on Differences
Proof of Theorem 1.2: Since h(x) − h(y) =R 1
h(x) − h(y) ≤ x − yZ 1
0 ∇h(xt) · (x − y) dt, we have
0 ∇h(xt) dt where xt := tx + (1 − t)y .
7
(3.1)
Next, by the convexity of f ,
e−f (x)e−f (y) = e−(1−t)f (x)e−tf (y)e−tf (x)e−(1−t)f (y) ≤ e−f (xt)e−(1−t)f (x)e−tf (y) .
(3.2)
Introduce the variables
w = tx + (1 − t)y
z = x − y .
(3.3)
A simple computation of the Jacobian shows that this change of variables is a measure
preserving transformation for all 0 ≤ t ≤ 1, and hence
ZRnZRn
h(x) − h(y)
dµ(x) dµ(y) ≤
Z 1
0 (cid:18)ZRnZRn ∇h(w)e−(1−t)f (w+(1−t)z)e−tf (w−tz) dz dµ(w)(cid:19) dt .
x − y
(3.4)
We estimate the right side of (3.4). By Holder's inequality,
ZRn
e−(1−t)f (w+(1−t)z)e−tf (w−tz) dnz ≤
(cid:18)ZRn
e−f (w+(1−t)z) dz(cid:19)1−t(cid:18)ZRn
e−f (w−tz) dz(cid:19)t
.
(3.5)
But
ZRn
ZRn
and finally, (1 − t)−n(1−t)t−nt = e−n(t log t+(1−t) log(1−t)) ≤ 2n.
e−f (w+(1−t)z) dz = (1 − t)−n
and
e−f (w−tz) dz = t−n ,
A corollary of Theorem 1.2 is a proof of Theorem 1.1 for the special case of q = 1 and
p = ∞. This proof is not only restricted to this case, it also has the defect that the constant
is not sharp, except in one-dimension. We give it, nevertheless, because it establishes a
link between the two theorems.
Alternative Proof of Theorem 1.1 for q = 1: We shall use the identity
cov(g, h) =
1
2ZRnZRn
[g(x) − g(y)][h(x) − h(y)] dµ(x) dµ(y) ,
(3.6)
8
and estimate the differences on the right in different ways.
Fix any x 6= y in Rn, and define the vector v := x − y, and for 0 ≤ t ≤ 1, define
xt = y + tv = tx + (1 − t)y. Then for any Lipschitz function h,
h(x) − h(y) =Z t
0
v · ∇h(xt) dt .
Now note that
d
dt
v · ∇f (xt) = (v, Hessf (xt)v) ≥ x − y2λmin(xt) > 0 .
Integrating this in t from 0 to 1, we obtain
(3.7)
(3.8)
(x − y,∇f (x) − ∇f (y)) =Z 1
0
(v, Hessf (xt)v) dt > 0 ,
(3.9)
which expresses the well-known monotonicity of gradients of convex functions.
Next, multiplying and dividing by (v, Hessf (xt)v) in (3.7), we obtain
0
0
Z 1
h(x) − h(y) = (cid:12)(cid:12)(cid:12)(cid:12)
(v, Hessf (xt)v)(v, Hessf (xt)v)−1v · ∇h(xt) dt(cid:12)(cid:12)(cid:12)(cid:12)
≤ Z 1
(v, Hessf (xt)v)(cid:12)(cid:12)(v, Hessf (xt)v)−1v · ∇h(xt)(cid:12)(cid:12) dt
≤ Z 1
(v, Hessf (xt)v)(cid:12)(cid:12)(λmin(xt))−1 x − y−2v · ∇h(xt)(cid:12)(cid:12) dt
λmin(z)(cid:27) x − y−1Z 1
z∈Rn(cid:26)∇h(z)
≤ sup
z∈Rn(cid:26)∇h(z)
λmin(z)(cid:27) x − y−1 (x − y,∇f (x) − ∇f (y)) .
(v, Hessf (xt)v) dt
= sup
0
0
(3.10)
Define
and use (3.10) in (3.6):
C := sup
z∈Rn(cid:26)∇h(z)
λmin(z)(cid:27) ,
cov(g, h) ≤
1
C
≤
2ZRnZRn g(x) − g(y)h(x) − h(y) dµ(x) dµ(y)
2 ZRnZRn g(x) − g(y)
2 ZRnZRn g(x) − g(y)
= −CZRnZRn g(x) − g(y)
x − y
1
x − y
x − y
=
C
1
1
(x − y) · [∇f (x) − ∇f (y)] dµ(x) dµ(y)
(x − y) ·(cid:2)∇ye−f (y)e−f (x) − ∇xe−f (x)e−f (y)(cid:3) dnx dny .
(x − y) · ∇xe−f (x)e−f (y) dnx dny ,
where, in the last line, we have used symmetry in x and y.
Now integrate by parts in x. Suppose first that n > 1. Then
div(cid:18) 1
z
z(cid:19) =
n − 1
z
,
9
and ∇xg(x) − g(y) = ∇xg(x) almost everywhere. Hence we obtain
cov(g, h) ≤ C(cid:18)ZRn ∇g(x) dµ(x) + (n − 1)ZRnZRn
For n = 1, div(cid:18) 1
z
dµ(x) dµ(y)(cid:19) . (3.11)
z(cid:19) = 2δ0(z) and (3.11) is still valid since g(x) − g(y)δ0(x − y) = 0.
Now, for n = 1, (3.11) reduces directly to (1.5). For n > 1, it reduces to (1.7) upon
g(x) − g(y)
x − y
application of Theorem 1.2, but with the constant n2n instead of 1.
4 Examples and Remarks on Optimizers in Theorem
1.2
Our first examples address the question of the importance of log-concavity.
(1.) Some restriction on µ is necessary: If a measure dµ(x) = F (x)dx on R has
F (a) = 0 for some a ∈ R, and F has positive mass to the left and right of a, then inequality
(1.8) cannot possibly hold with any constant. The choice of h to be the Heaviside step
function shows that (1.8) cannot hold with any constant for this µ.
(2.) Unimodality is not enough: Take dµ(x) = F (x)dx, with F (x) = 1/4ε on
(−ε, ε) and F (x) = 1/4(1 − ε) otherwise on the interval (−1, 1) and F (x) = 0 for x > 1.
Let g(x) = 1 for x < ε + δ and g(x) = 0 otherwise. When δ is positive but small,
while
ZRZR
ZR ∇g dµ(x) = 1/2(1 − ε)
g(x) − g(y)
x − y
dµ(x) dµ(y) = O(− ln(ǫ)) .
(3.) For n = 1, the best constant in (1.8) is at least 2 ln 2: Take dµ(x) = F (x)dx,
with F (x) = 1/2 on (−1, 1) and F (x) = 0 for x > 1. Let g(x) = 1 for x ≥ 0 and g(x) = 0
for x < 0. All integrals are easily computed.
(4.) The best constant is achieved for characteristic functions: When seeking
the best constant in (1.8), it suffices, by a standard truncation argument, to consider
bounded Lipschitz functions h. Then, since neither side of the inequality is affected if we
10
add a constant to h, it suffices to consider non-negative Lipschitz functions. We use the
layer-cake representation [11]:
h(x) =Z ∞
0
χ{h>t}(x) dt .
Then
ZRnZRn
h(x) − h(y)
x − y
dµ(x) dµ(y) ≤Z ∞
0 ZRnZRn
x − y
χ{h>t}(x) − χ{h>t}(y)
dµ(x) dµ(y) dt
(4.1)
Define Cn to be the best constant for characteristic functions of sets A and log-concave
measures µ:
Cn := sup
f,A (RRnRRn
χA(x)−χA(y)
x−y
dµ(x) dµ(y)
R∂A e−f (x) dHn−1(x)
)
(4.2)
where Hn−1 denotes n− 1 dimensional Hausdorff measure. Apply this to (4.1) to conclude
that
ZRnZRn
h(x) − h(y)
x − y
dµ(x) dµ(y) ≤ CnZ ∞
0 Z∂χ{h>t}
e−f (x) dHn−1(x) dt
= CnZRn ∇h(x) dµ(x) ,
(4.3)
where the co-area formula was used in the last line. Thus, inequality (1.8) holds with the
constant Cn; in short, it suffices to consider characteristic functions as trial functions. Note
that the argument is also valid at the level of each measure µ individually, although we are
interested here in uniform bounds.
With characteristic functions in mind, let us consider the case that g is the characteristic
function of a half-space in Rn. Without loss of generality let us take this to be {x : x1 < 0}.
Clearly, the left side of (1.8) is less than the integral with x− y−1 replaced by x1 − y1−1.
Since the marginal (obtained by integrating over x2, . . . , xn) of a log concave function is log
concave, we see that our inequality reduces to the one-dimensional case. In other words,
the constant Cn in (4.2) would equal C1, independent of n, if the supremum were restricted
to half-spaces instead of to arbitrary measurable sets.
(5.) Improved constants and geometry of log-concave measures: With addi-
tional assumptions on the measure one can see that the constant is not only bounded in
n, but of order 1/√n. We are grateful to F. Barthe and M. Ledoux for discussions and
improvements in particular cases concerning the constant in Theorem 1.2. This relies on
the Cheeger constant α(µ)−1 > 0 associated to the log-concave probability measure dµ,
which is defined to be the best constant in the inequality
∀A ⊂ Rn(regular enough),
µ(A)(1 − µ(A)) ≤ α(µ)Z∂A
e−f (x) dHn−1(x)
11
M. Ledoux suggested the following procedure. Split the function x − y−1 into two pieces
according to whether x − y is less than or greater than R, for some R > 0. With h being
the characteristic function of A, the contribution to the left side of (4.3) for x − y > R
is bounded above by 2R−1α(µ)R∂A e−f (x) dHn−1(x). The contribution for x − y ≤ R is
bounded above in the same manner as in the proof of Theorem 1.2, but this time we only
have to integrate z over the domain z ≤ R in each of the integrals in (3.5). Thus, our
bound 2n is improved by a factor, which is the dµ volume of the ball BR = {z ≤ R},
once we used the Brunn-Minkowski inequality for the bound
µ(cid:0)(1 − t)BR + w)(cid:1)1−t µ(cid:0)tBR + w(cid:1)t ≤ µ(cid:0)BR + w(cid:1) ≤ µ(BR) := sup
x
µ(BR + x).
The final step is to optimize the sum of the contributions of the two terms with respect to
R. Thus, if we denote Cn(µ) the best constant in the inequality (1.8) of Theorem 1.2 for
a fixed measure µ, we have
Cn(µ) ≤ inf
R>0(cid:8)2nµ(BR) + 2R−1α(µ)(cid:9) ≤ 2n.
(4.4)
Note that if µ is symmetric (i.e. if f is even), then the Brunn-Minkowski inequality ensures
that µ(BR) = µ(BR).
Unlike in (1.8), this improved bound depends on µ but there are situation where this
gives optimal estimates as pointed out to us by F. Barthe. As an example, consider the case
where µ is the standard Gaussian measure on Rn. Using the known value of the Cheeger
constant for this µ, and linear trial functions, one finds that the constant is bounded above
and below by a constant times n−1/2.
Actually, we can use (4.4) to improve the constant from 2n to 2n/2 for arbitrary measures
using some recent results from the geometry of log-concave measures. Without loss of
generality, we can assume, by translation of µ, thatR x dµ(x) = inf vR x+v dµ(x) =: Mµ.
It was proved in [9, 1] that for every log-concave measure on Rn,
α(µ) ≤ cMµ
where c > 0 is some numerical constant (meaning a possibly large, but computable, con-
stant, in particular independent of n and µ, of course). On the other hand, it was proved
by Gu´edon [7] that for every log-concave measure ν on Rn
ν(BR) ≤
R
C
R x dν
for some numerical constant C > 0. In the case µ is not symmetric, we pick v such that
µ(Br + v) = µ(BR), and then we apply the previous bound to ν(·) = µ(· + v) in order to
get that µ(BR) ≤ C
. Using these two estimates in (4.4) we see that
Mµ
Cn(µ) ≤ inf
s>0{C2ns + c/s} = κ 2n/2
12
for some numerical constant κ > 0.
The Brascamp-Lieb inequality (1.3), as well as inequality (1.8), have connections with
the geometry of convex bodies. It was observed in [2] that (1.3) can be deduced from the
Pr´ekopa-Leindler inequality (which is a functional form of the Brunn-Minkowski inequal-
ity). But the converse is also true: the Pr´ekopa theorem follows, by a local computation,
from the Brascamp-Lieb inequality (see [5] where the procedure is explained in the more
general complex setting). To sum up, the Brascamp-Lieb inequality (1.3) can be seen as
the local form of the Brunn-Minkowski inequality for convex bodies.
5 Application to Conditional Expectations
Otto and Menz were motivated to prove (1.5) for an application that involves a large
amount of additional structure that we cannot go into here. We shall however give an
application of Theorem 1.1 to a type of estimate that is related to one of the central
estimates in [13].
We use the notation in [4], which is adapted to working with a partitioned set of
variables. Write a point x ∈ Rn+m as x = (y, z) with y ∈ Rm and z ∈ Rn. For a function
A on Rn+m, let hAiz(y) denote the conditional expectation of A given y, with respect to
µ. For a function B of y alone, hBiy is the expected value of B, with respect to µ. As in
[4], a subscript y or z on a function denotes differentiation with respect to y or z, while a
subscript y or z on a bracket denotes integration. For instance, for a function g on Rn+m,
∂yi(cid:1)i≤n in Rn, and for i ≤ n, gyiz denotes the vector (cid:0) ∂ 2g
gy denotes the vector (cid:0) ∂g
∂yi∂zj(cid:1)j≤m in
Rm. Finally, (gyz) denotes the n × m matrix having the previous vectors as rows.
Let h be non-negative with hhix = 1 so that h(x) dµ(x) is a probability measure, and
so is hhiz(y) dν(y), where dν(y) is the marginal distribution of y under dµ(x).
A problem that frequently arises [3, 6, 10, 12, 13] is to estimate the Fisher information
of hhiz(y) dν(y) in terms of the Fisher information of h(x) dµ(x) by proving an estimate of
the form
(cid:28)(hhiz)y2
hhiz (cid:29)y ≤ C(cid:28)hx2
h (cid:29)x
.
(5.1)
Direct differentiation under the integral sign in the variable yi gives
(hhiz)yi = hhyiiz − covz(h, fyi) ,
where covz denotes the conditional covariance of h(y, z) and fyi(y, z), integrating in z for
each fixed y. Let u = (u1, . . . , um) be any unit vector in Rm. Then Hence, for each y,
(hhiz)y · u =
m
Xi=1
(hhiz)yiui =
m
m
Xi=1
hhyiizui −
Xi=1
= hhyiz · u − covz(h, fy · u) ,
covz(h, fyi)ui
and hence, choosing u to maximize the left hand side,
(hhiz)y2 ≤ 2hhyiz2 + 2 (covz(h, fy · u))2 .
(5.2)
By (1.6),
13
covz(h, fy · u) ≤ hhzizkλ−1
(5.3)
Note that the least eigenvalue of the n × n block fzz is at least as large as the least
eigenvalue λmin(y, z) of the full Hessian, by the variational principle. Hence, while we are
entitled to use the least eigenvalue of the n × n block fzz of the full (n + m) × (n + m)
Hessian matrix fxx, and this would be important in the application in the one dimensional
case made in [13], here, without any special structure to take advantage of, we simply use
the least eigenvalue of the full matrix in our bound.
min(fy · u)zk∞ .
Next note that
n
(fy · u)z2 ≤
m
Xi=1 n
Xj=1
(fyi,zj )2! u2
i ,
and that
(fyi,zj )2 is the i, i entry of f T
yzfyz where fyz denotes the upper right corner
Xj=1
block of the Hessian matrix. This number is no greater than the i, i entry of the square
of the full Hessian matrix. This, in turn, is no greater than λ2
max. Then, since u is a unit
vector, we have
Using this in (5.3), we obtain
(fy · u)z ≤ λmax .
covz(h, fy · u) ≤ hhzizkλmax/λmink∞ ,
and then from (5.2)
(hhiz)y2 ≤ 2hhyiz2 + 2kλmax/λmink2
∞hhzi2
z .
Then the Cauchy-Schwarz inequality yields
(hhziz)2 ≤(cid:28)hz2
h (cid:29)z hhiz .
(5.4)
(5.5)
(5.6)
Use this in (5.5), divide both sides by hhiz, and integrate in y. The joint convexity in A
and α > 0 of A2/α yields (5.1) with the constant C = 2kλmax/λmink2
∞.
The bound we have obtained becomes useful when λmax(x)/λmin(x) is bounded uni-
formly. Suppose that f (x) has the form f (x) = ϕ(x2). Then the eigenvalues of the
Hessian of f are 2ϕ′(x2), with multiplicity m + n − 1, and 4ϕ′′(x2)x2 + 2ϕ′(x2), with
multiplicity 1. Then both eigenvalues are positive, and the ratio is bounded, whenever ϕ′
is positive and, for some c < 1 < C < ∞,
−cϕ′(s) ≤ sϕ′′(s) ≤ Cϕ′(s) .
5.1 Remark (Other asymmetric variants of the BL inequality). Together, (5.3) and (5.6)
yield
14
(covz(h, fy · u))2
hhiz
≤(cid:28)hz2
h (cid:29)z kλ−1
min(fy · u)z)k2
∞ .
A weaker inequality is
(covz(h, fy · u))2
hhiz
≤(cid:28)hz2
h (cid:29)z kλ−1
mink2
∞k(fy · u)z)k2
∞ .
(5.7)
In the context of the application in [13], finiteness of k(fy · u)zk∞ limits f to quadratic
growth at infinity. A major contribution of [13] is to remove this limitation in applications
of (5.1). The success of this application of (1.5) depended on the full weight of the inverse
Hessian being allocated to the L∞ term.
Nonetheless, once the topic of asymmetric BL inequalities is raised, one might enquire
whether an inequality of the type
cov(g, h) ≤ Ck∇gk∞ kHess−1
f ∇hk1
(5.8)
can hold for any constant C. There is no such inequality, even in one dimension. To see
this, suppose that for some a ∈ R and some ǫ > 0, fxx > M on (a− ǫ, a + ǫ). Take h(x) = 1
for x > a and h(x) = 0 for x ≤ a. Take g(x) = x − a. Suppose that f is even about a.
Then cov(g, h) = R ∞
f ∇hk1 ≤ M −1, and f can be chosen
to make M arbitrarily large while keeping k∇gk∞ ≤ 1, and cov(g, h) bounded away from
zero.
a (x − a)e−f (x) dx, while kHess−1
6 Appendix
We recall that the original proof of (1.3), Theorem 4.1 of [4], used dimensional induction,
though interesting non-inductive proofs have since been provided [2].
The starting point for the inductive proof is that the proof for n = 1 is elementary. The
proof of the inductive step is more involved, and we take this opportunity to provide more
detail about the passage from eq. (4.9) of [4] to eq. (4.10) of [4]. There is an interesting
connection with the application discussed in the previous section, which also concerns
hhyiz − covz(h, fy). We continue using the notation introduced there, but now m = 1 (i.e.
y ∈ R).
Eq. (4.9) reads var(h) ≤ hBiy where
B = varz(h) +
[hhyiz − covz(h, fy)]2
hfyyiz − varzfy
.
(6.1)
Our goal is to prove
B ≤ h(hz, f −1
zz hz)iz + hhy − (hz, f −1
hfyy − (fyz, f −1
zz fyz)i2
zz fyz)iz
z
.
(6.2)
To do this, use the inductive hypothesis; i.e., for any H on Rn−1,
varz(H) ≤ hHz, f −1
zz Hziz .
15
(6.3)
Apply this to arbitrary linear combination H = λh + µfy to conclude the 2 × 2 matrix
inequality
varz(h)
covz(h, fy)
(cid:20)
covz(h, fy)
varz(fy) (cid:21) ≤(cid:20) h(hz, f −1
h(fyz, f −1
zz hz)iz
zz hz)iz
h(hz, f −1
h(fyz, f −1
zz fyz)iz (cid:21)
zz fyz)iz
Take the determinant of the difference to find that
[h(hz, f −1
h(fyz, f −1
zz hz)iz − varz(h) ≥
h(hz, f −1
zz fyz)iz − covz(h, fy)]2
zz fyz)iz − varz(fy)
.
(6.4)
Combine (6.1) and (6.4) to obtain
B ≤ h(hz, f −1
zz hz)iz +
[hhyiz − covz(h, fy)]2
hfyyiz − varz(fy) −
[h(hz, f −1
h(fyz, f −1
zz fyz)iz − covz(h, fy)]2
zz fyz)iz − varz(fy)
(6.5)
Since a2/α is jointly convex in a and α > 0, and is homogeneous of degree one, for all
α > β > 0 and all a and b,
a2
α ≤
b2
β
+
(a − b)2
α − β
.
That is, a2/α − b2/β ≤ (a − b)2/(α − β). Use this on the right side of (6.5) to obtain
(6.2), noting that the positivity of α− β = hfyyiz −h(fyz, f −1
zz fyz)iz is a consequence of the
positivity of the Hessian of f .
References
[1] S.G. Bobkov, Isoperimetric and analytic inequalities for log-concave probability measures,
Ann. Probab. 27, no. 4, 1903 -- 1921 (1999).
[2] S. Bobkov, M. Ledoux, From Brunn-Minkowski to Brascamp-Lieb and to logarithmic Sobolev
inequalities, Geom. Funct. Anal. 10, 102-1052 (2000).
[3] Th. Bodineau and B. Helffer, On log Sobolev inequalities for unbounded spin systems, J.
Funct. Anal. 166, 168-178 (1999).
[4] H. J.Brascamp and E. H. Lieb, On extensions of the Brunn-Minkovski and Pr´ekopa-Leindler
theorems, including inequalities for log-concave functions, and with an application to the
diffusion equation, J. Funct. Anal. 22, 366-389 (1976).
[5] D. Cordero-Erausquin, On Berndtsson's generalization of Pr´ekopa's theorem, Math. Z. 249,
no. 2, 401 -- 410 (2005).
16
[6] N. Grunewald, F. Otto, C. Villani and M. G. Westdickenberg, A two-scale approach to
logarithmic Sobolev inequalities and the hydrodynamic limit, Ann. Inst. H. Poincare Prob.
Stat. 45, 302-351 (2009).
[7] O. Gu´edon, Kahane-Khinchine type inequalities for negative exponent, Mathematika 46, no.
1, 165 -- 173 (1999).
[8] L. Hormander, L2 estimates and existence theorems for the ¯∂ operator, Acta Math. 113,
89 -- 152 (1965).
[9] R. Kannan, L. Lov´asz and M. Simonovits, Isoperimetric problems for convex bodies and a
localization lemma, Discrete Comput. Geom. 13, 541 -- 559 (1995).
[10] C. Landim, G. Panizo and H. T. Yau, Spectral gap and logarithmic Sobolev inequality for
unbounded conservative spin systems Ann. Inst. H. Poincar´e Prob. Stat. 38, 739-777 (2002).
[11] E. H. Lieb and M. Loss, Analysis, Second Edition Amer. Math. Soc., Providence RI (2001).
[12] F. Otto and M. G. Reznikoff, A new criterion for the logarithmic Sobolev inequality and two
applications, J. Funct. Anal. 243, 121-157 (2007).
[13] G. Menz and F. Otto, Uniform logarithmic Sobolev inequalities for conservative spin systems
with super-quadratic single-site potential, Leipzig Preprint no. 5 (2011).
|
1605.07071 | 1 | 1605 | 2016-05-23T16:04:58 | Strict positive definiteness on a product of compact two-point homogeneous spaces | [
"math.FA"
] | We present an explicit characterization for the real, continuous, isotropic and strictly positive definite kernels on a product of compact two-point homogeneous spaces, in the cases in which at least one of the spaces is a sphere of dimension greater than 1 and the other is not a circle. The result complements similar characterizations previously obtained for products of high dimensional spheres. | math.FA | math |
Strict positive definiteness on a product of
compact two-point homogeneous spaces
V. S. Barbosa and V. A. Menegatto
We present an explicit characterization for the real, continuous, isotropic and
strictly positive definite kernels on a product of compact two-point homoge-
neous spaces, in the cases in which at least one of the spaces is a sphere of
dimension greater than 1 and the other is not a circle. The result complements
similar characterizations previously obtained for products of high dimensional
spheres.
Mathematics Subject Classifications (2010): 22F30, 33C05, 33C45, 33C55, 41A63,
42C10
Keywords: two-point homogeneous spaces, strict positive definiteness, isotropy, antipo-
dal sets.
1
Introduction
Let Md denote a d-dimensional compact two-point homogeneous space. A real and con-
tinuous kernel K on Md is positive definite if it is a symmetric function, that is,
K(x, y) = K(y, x),
x, y ∈ Md,
and satisfies the following condition: if n is a positive integer and x1, x2, . . . , xn are distinct
points on Md, then the n × n matrix [K(xi, xj)] is nonnegative definite. This formulation
for the concept of positive definiteness is a refinement of the classical one described in [4]. A
positive definite kernel on Md is strictly positive definite if all the matrices in the previous
definition are in fact positive definite. Positive definite kernels and functions on metric
spaces are an important subject in many areas of classical and modern mathematics, such
as radial basis function interpolation and approximation, geomathematics, geostatistics,
Fourier analysis, etc. The interested reader may consider the references [6, 8, 11, 22] as a
starting point for ratifying that.
The continuity of a positive definite kernel on Md is attached to the usual (geodesic)
distance on Md, here denoted by xy, x, y ∈ Md. Throughout the paper, we will assume
the geodesic distance on Md fulfills the following requirement: all geodesics have the same
length 2π. The geodesic distance on Md allows the introduction of the notion of isotropy
of a positive definite kernel K on Md, in the same way I. J. Schoenberg did for positive
definite kernels on spheres ([19]). It demands that
K(x, y) = K d
i (cos xy/2),
x, y ∈ Md,
for some function K d
: [−1, 1] → R, here called the isotropic part of K. According to R.
i
Gangolli ([9]), a continuous and isotropic kernel K on Md is positive definite if and only
if
∞
K d
i (t) =
a(d−2)/2,β
k
P (d−2)/2,β
k
(t),
t ∈ [−1, 1],
(1.1)
Xk=0
k
k=0 a(d−2)/2,β
∈ [0, ∞), k ∈ Z+ and P∞
in which a(d−2)/2,β
(1) < ∞. Here, β =
(d − 2)/2, −1/2, 0, 1, 3, depending on the respective category Md belongs to, among the
following ones ([21]): the unit spheres Sd, d = 1, 2, . . ., the real projective spaces Pd(R),
d = 2, 3, . . ., the complex projective spaces Pd(C), d = 4, 6, . . ., the quaternionic projective
spaces Pd(H), d = 8, 12, . . ., and the Cayley projective plane Pd(Cay), d = 16. The
symbol P (d−2)/2,β
stands for the Jacobi polynomial of degree k associated with the pair
((d − 2)/2, β).
P (d−2)/2,β
k
k
k
This is the point where we can explain what the intentions in this paper are. The first
target is to present a characterization for the real, continuous and isotropic kernels which
are positive definite on a cartesian product of compact two-point homogeneous spaces. If
Md and Hd′ are the spaces, the isotropy of a kernel K : Md × Hd′ → R corresponds to
isotropy in both spaces, that is, K is of the form
K((x, w), (y, z)) = K d,d′
i
(cos(xy/2), cos(wz/2)),
x, y ∈ Md, w, z ∈ Hd′
,
i
for some function K d,d′
: [−1, 1]2 → R (the isotropic part of K). Since we believe such
a characterization can be deduced via classical results from harmonic analysis (see [2]),
the proposal here is to obtain the characterization using an alternative procedure in-
volving the Gauss hypergeometric function F2
1 and basic convergence arguments. The
characterization itself can be described as follows: if K is a real, continuous and isotropic
kernel on Md × Hd′, then it is positive definite if and only if its isotropic part has a series
representation in the form
K d,d′
i
(t, s) =
∞
Xk,l=0
ak,l(K d,d′
i
)P (d−2)/2,β
k
(t)P (d′−2)/2,β ′
l
(s),
t, s ∈ [−1, 1]2,
in which ak,l(K d,d′
(1) < ∞.
The numbers β and β′ have to agree, respectively, with d and d′ respecting Wang's
classification in [21]. The result will be deduced in Section 2.
) ≥ 0, k, l ∈ Z+ and P∞
k,l=0 ak,l(K d,d′
(1)P (d′−2)/2,β ′
)P (d−2)/2,β
k
i
i
l
2
As explained in [12], positive definiteness on a product of spaces allows intermedi-
ate notions of strict positive definiteness. In the case of a positive definite kernel K :
Md × Hd′ → R, one of them reads like this: K is DC-strictly positive definite if the strict
positive definiteness condition previously introduced holds for points of Md × Hd′ having
distinct components. Equivalently, all matrices of the form [K((xi, wi), (xj, wj))] are pos-
itive definite if the xi are distinct in Md and the wi are distinct in Hd′. Clearly, DC-strict
positive definiteness is a weaker notion in the sense that it demands plain strict positive
definiteness for just some of the distinct points in Md × Hd′. In the case Md = Sd and
Hd′ = Sd′, d, d′ ≥ 2, DC-strict positive definiteness of a real, continuous, isotropic and
positive definite kernel K was shown to be equivalent to the following condition ([14]):
the set {k + l : ak,l(K d,d′
) > 0} contains infinitely many even and infinitely many odd inte-
gers. In Section 3, we will complete this line of investigation, presenting a characterization
for DC-strict positive definiteness in all the other products involving compact two-point
homogeneous spaces, except when one of the spaces is a circle.
i
The third target in the paper is to present a characterization for plain strict positive
definiteness of a real, continuous, isotropic and strict positive definite kernel on Md × Hd′,
in the same cases mentioned at the end of the previous paragraph. That is a counterpart
of similar results for products of spheres previously obtained in [13, 14, 15]. The details
will appear in Section 4.
2 Positive definiteness
In this section, we will provide a characterization for the real, continuous, isotropic and
positive definite kernels on Md × Hd′. No additional assumption on either d or d′ will be
made. The characterization is an extension to all the compact two-point homogeneous
spaces of that one previously obtained in [12] in the case both spaces are spheres (see also
[4] for an alternative proof in that case).
Let us begin with some basics on Jacobi polynomials. For α, β > −1, the set {P α,β
:
k ∈ Z+} of Jacobi polynomials associated to the pair (α, β) is orthogonal on [−1, 1] in
the sense that
k
Z 1
−1
P α,β
k
(t)P α,β
l
(t)(1 − t)α(1 + t)βdt = δk,lhα,β
k
,
hα,β
k =
2α+β+1
2k + α + β + 1
Γ(k + α + 1)Γ(k + β + 1)
Γ(k + 1)Γ(k + α + β + 1)
,
k ∈ Z+.
where
Here, and in many other places in the paper, Γ will stand for the usual gamma function.
An immediate consequence is the orthogonality of the family
n(t, s) ∈ [−1, 1] 7→ P α,β
k
(t)P α′,β ′
l
(s) : k, l ∈ Z+o
3
with respect to the weight function
σd′
d (t, s) := (1 − t)α(1 + t)β(1 − s)α′
(1 + s)β ′
,
t, s ∈ [−1, 1].
There exists a generating formula for Jacobi polynomials via the Gauss hypergeometric
1 ([1, 7, 18, 20]). As a regular solution of the hypergeometric differential
function F2
equation, the hypergeometric function has a representation in the form
F2
1 (a, b; c; z) =
(a)n(b)nzn
(c)nn!
,
∞
Xn=0
in which a, b and c are generic parameters, z is a complex variable, and
(λ)n =
Γ(n + λ)
Γ(λ)
=(cid:26) λ(λ + 1) · · · (λ + n − 1),
1,
if n ≥ 1
if n = 0
is the Pochhammer symbol. The convergence holds for z < 1 if c is not a negative integer
and for z = 1 if Re (c − a − b) > 0 ([20, p. 63]).
For simplicity's sake, we will write
F (a, b; c; z) := F2
1 (a, b; c; z).
The hypergeometric function F is differentiable with respect to z in {z ∈ C : z < 1}
([23, p. 281]) and
d
dz
F (a, b; c; z) =
ab
c
F (a + 1, b + 1; c + 1; z).
In particular,
Z z2
z1
F (a, b; c; z)dz =
c − 1
(a − 1)(b − 1)
[F (a − 1, b − 1; c − 1; z2) − F (a − 1, b − 1; c − 1; z1)] .
A generating formula for Jacobi polynomials based on the function F is the content of
the following Poisson formula ([1, p. 21]):
∞
Xn=0
P α,β
n (1)P α,β
n (t)rn
hα,β
n
= Gα,β(r)F (cid:18)α + β + 2
2
,
α + β + 3
2
; β + 1;
2r(1 + t)
(1 + r)2 (cid:19) ,
t ∈ [−1, 1],
in which
Gα,β(r) :=
2−(α+β+1)Γ(α + β + 2)(1 − r)
Γ(α + 1)Γ(β + 1)(1 + r)α+β+2 .
We now detach a consequence of the results described above to be used ahead.
4
Lemma 2.1. Let α ≥ β ≥ −1/2 and r ∈ (−1, 1). The series
P α,β
n (1)P α,β
n (t)rn
hα,β
n
∞
Xn=0
is convergent for all t ∈ [−1, 1].
Proof. If r = 0, the result is obvious. Otherwise, taking into account the convergence of
the series representation for F , it is promptly seen that the series in the statement of the
lemma will be convergent as long as
1 + t <
(1 + r)2
2r
.
However, a simple calculation reveals that (1 + r)2 > 4r whenever r ∈ (−1, 1) \ {0}.
Thus, since t ∈ [−1, 1], the convergence follows in this case as well.
Lemma 2.2 is a critical step towards the desired characterization in this section. It
follows from the definition of positive definiteness along with Gangolli's characterization
for positive definiteness on a single compact two-point homogeneous space and the Schur
product theorem for nonnegative definite matrices. Once again, β and β′ have to agree
with Wang's classification for the compact two-point homogeneous spaces.
Lemma 2.2. For fixed nonnegative integers k and l, the function
(t, s) ∈ [−1, 1] 7→ P (d−2)/2,β
k
(t)P (d′−2)/2,β ′
l
(s)
is the isotropic part of a positive definite kernel on Md × Hd′.
In the next proposition, dx and dy will denote the volume elements on Md and Hd′,
respectively (dimensions will be omitted). We will require the expansion of a function f
from L1([−1, 1], σd′
d ):
f (t, s) ∼
∞
Xk,l=0
ak,l(f )P (d−2)/2,β
k
(t)P (d′−2)/2,β ′
l
(s)
in which
ak,l(f ) =
h(d−2)/2,β
k
1
h(d′−2)/2,β ′
l
Z[−1,1]2
f (t, s)P (d−2)/2,β
k
(t)P (d′−2)/2,β ′
l
(s)dσd′
d (t, s),
k, l ∈ Z+.
The dependence of ak,l upon d, d′, β, β′ will be omitted.
In the case f is the isotropic part of a kernel belonging to the setting of the paper,
the following formula holds.
5
Proposition 2.3. Let k and l be nonnegative integers. If K is a real, continuous and
isotropic kernel on Md × Hd′, then there exists a positive constant C, that depends upon
d, d′, β and β′, so that
ak,l(K d,d′
i
) = CZMd×Hd′(cid:20)ZMd×Hd′
K((x, y), (w, z))P (d−2)/2,β
k
(cos(xy/2))
×P (d′−2)/2,β ′
l
(cos(wz/2))dydzi dxdw.
Proof. The first step in the proof is to observe that the integral
ZMd×Hd′
f (cos(xy/2), cos(wz/2))P (d−2)/2,β
k
(cos(xy/2)P (d−2)/2,β ′
l
(cos(wz/2))dydz
equals to a positive multiple of ak,l(f ). Indeed, this follows after two applications of
the Funk-Hecke formula for harmonics on compact two-point homogeneous spaces (see
Proposition 2.8 and the Remark 2.9 in [17]) coupled with Fubini's theorem. The multiple
depends upon d, d′, β and β′. Calling it C and integrating leads to the formula in the
statement of the proposition.
If we introduce the positive definiteness of K as an assumption, we obtain the following
expected consequence.
Proposition 2.4. If K is a real, continuous, isotropic and positive definite kernel on
Md × Hd′, then
ak,l(K d,d′
i
) ≥ 0,
k, l ∈ Z+.
Proof. Due to Lemma 2.2, if K is positive definite on Md × Hd′, then the integrand
in the double integral appearing in the statement of the previous proposition defines a
real, continuous, isotropic and positive definite on Md × Hd′. Hence, the the proof of the
proposition resumes to showing that if K is a real, continuous, isotropic and positive
definite kernel on Md × Hd′, then the integral
I :=ZMd×Hd′(cid:20)ZMd×Hd′
K((x, y), (w, z))dydz(cid:21) dxdw
is nonnegative. In order to achieve that, we will fix ǫ > 0 and will show there exists a
number I = I(ǫ) ≥ 0 so that
I − I ≤ ǫVol.(Md × Hd′
)2.
Indeed, if I were negative, then the information above with a convenient choice for ǫ
would produce a contradiction. Since Md × Hd′ is a compact metric space, the kernel K
is actually uniformly continuous on Md × Hd′. In particular, we can select δ > 0 so that
6
K((x, y), (w, z)) − K((x′, y′), (w′, z′)) < ǫ whenever x, x′, y, y′ ∈ Md, w, w′, z, z′ ∈ Hd′,
xx′ < δ, yy′ < δ, ww′ < δ and zz′ < δ. Since the metric spaces Md and Hd′ are
totally bounded, we can cover them with finitely many open balls of radius δ/2. Likewise,
we can cover Md × Hd′ with finitely many open balls. Using the covering, we can partition
Md × Hd′ into finitely many Borel subsets
Md × Hd′
= ⊔p
j=1Bj
so that xx′ < δ and ww′ < δ whenever ((x, w), (x′, w′)) ∈ Bj, j = 1, 2, . . . , p. It is
now clear that if, ((x, y), (w, z)), ((x′, y′), (w′, z′)) ∈ Bj × Bk, for some pair (k, j), then
K((x, y), (w, z)) − K((x′, y′), (w′, z′)) < ǫ. To proceed, observe that
I =
p
Xl,k=1ZBjZBk
K((x, y), (w, z))dxdwdydz.
Next, for each j ∈ {1, 2, . . . , p}, choose (xj, wj) ∈ Bj and define λj := the volume of Bj.
Clearly, the number
p
is nonnegative due to the positive definiteness of the kernel K. On the other hand
I :=
λjλkK((xj, xk), (wj, wk))
Xj,k=1
[K((x, y), (w, z)) − K((xj, xk), (wj, wk))]dxdwdydz(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
p
I − I = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ ǫ
Xj,k=1ZBjZBk
Xj,k=1
λjλk,
p
that is, I − I ≤ ǫ Vol.(Md × Hd′)2.
Next, we move to convergence of double series defined by Jacobi polynomials.
Lemma 2.5. Let K be a real, continuous, isotropic and positive definite kernel on Md ×
Hd′. If r, ρ ∈ (−1, 1), then the double series
ak,l(K d,d′
i
)P (d−2)/2,β
k
(1)P (d′−2)/2,β ′
l
(1)rkρl
∞
Xk,l=0
converges.
Proof. We will prove the lemma in the case in which d 6= 1 and d′ 6= 1. The proof in the
cases in which either d = 1 or d′ = 1 can be adapted from the general proof presented
below. However, the computation referring to the the case in which the space is S1 needs
7
to be done directly without mentioning the hypergeometric function. The calculations
made in [12] are very similar to what is needed in these specials cases. Otherwise, the
general term of the series in the statement of the lemma is
(t)rk P (d′−2)/2,β ′
l
h(d′−2)/2,β ′
l
P (d−2)/2,β
k
h(d−2)/2,β
k
P (d′−2)/2,β ′
l
P (d−2)/2,β
(s)ρldσd′
d (t, s).
K d,d′
(t, s)
(1)
(1)
−1
k
i
Z 1
−1Z 1
Introducing the Gauss hypergeometric function in the above expression leaves the double
series in the form
Z 1
−1Z 1
−1(cid:20)K d,d′
i
(t, s)G(d−2)/2,β(r)F (cid:18)d + 2β + 2
4
,
d + 2β + 4
4
; β + 1;
2r(1 + t)
(1 + r)2 (cid:19)
(1 + ρ)2 (cid:19) dσd′
2ρ(1 + s)
d (t, s)(cid:21) .
in [−1, 1] × [−1, 1], we
× G(d′−2)/2,β ′
(ρ)F (cid:18) d′ + 2β′ + 2
4
,
d′ + 2β′ + 4
4
; β′ + 1;
Due to the positive definiteness of K and the continuity of K d,d′
can estimate the double integral above by
i
C r,ρZ 1
−1Z 1
−1(cid:20)F (cid:18) d + 2β + 2
4
,
d + 2β + 4
4
; β + 1;
2r(1 + t)
(1 + r)2 (cid:19)
× F (cid:18) d′ + 2β′ + 2
4
,
d′ + 2β′ + 4
4
; β′ + 1;
2ρ(1 + s)
(1 + ρ)2 (cid:19) dσd′
d (t, s)(cid:21) .
in which C r,ρ is a positive multiple of G(d−2)/2,β(r)G(d′−2)/2,β ′(ρ). The weight in the defi-
d can be bounded by 2β+β ′−2+(d+d′)/2. Introducing this bound and solving the
nition of dσd′
resulting integrals, we conclude that the double series is at most
C F (cid:18)d + 2β − 2
4
,
d + 2β
4
; β;
in which
4r
(1 + r)2(cid:19) F (cid:18)d′ + 2β′ − 2
4
,
d′ + 2β′
4
; β′;
4ρ
(1 + ρ)2(cid:19) ,
C = G(d−2)/2,β(r)G(d′−2)/2,β ′
(ρ)
ββ′2β+β ′+4+(d+d′)/2
(1 + r)2(1 + ρ)2
(d + 2β − 2)(d′ + 2β′ − 2)(d + β)(d′ + β)
rρ
.
The proof is complete.
Proposition 2.6. If K is a real, continuous, isotropic and positive definite kernel on
Md × Hd′. then the double series
ak,l(K d,d′
i
)P (d−2)/2,β
k
(t)P (d′−2)/2,β ′
l
(s)
∞
Xk,l=0
converges absolutely and uniformly for (t, s) ∈ [−1, 1]2.
8
Proof. Due to the Weierstrass M-test for double series, it suffices to show that
ak,l(K d,d′
i
)P (d−2)/2,β
k
(1)P (d′−2)/2,β ′
l
(1)
∞
Xk,l=0
converges. In order to do that, consider the sequence (sp,q)p,q∈Z+ given by the partial sums
sp,q :=
p
q
Xk=0
Xl=0
ak,l(K d,d′
i
)P (d−2)/2,β
k
(1)P (d′−2)/2,β ′
l
(1),
p, q ∈ Z+.
By Lemma 2.4, ak,l(K d,d′
and q ≤ q′. On the other hand, by the previous lemma,
i
) ≥ 0, for all k, l ∈ Z+. In particular, sp,q ≤ sp′,q′ when p ≤ p′
p
q
Xk=0
Xl=0
ak,l(K d,d′
i
)P (d−2)/2,β
k
(1)P (d′−2)/2,β ′
l
(1)rkρl ≤ C,
p, q ∈ Z+,
r, ρ ∈ (−1, 1).
for some C > 0. Applying the limits when r, ρ → 1+, we deduce the sequence (sp,q) is
bounded above. The convergence of (sp,q) follows.
The main result in the section is as follows.
Theorem 2.7. Let K be a real, continuous and isotropic kernel on Md × Hd′. It is positive
definite on Md × Hd′ if and only if its isotropic part K d,d′
has a representation in the form
i
K d,d′
i
(t, s) =
∞
Xk,l=0
ak,l(K d,d′
i
)P (d−2)/2,β
k
(t)P (d′−2)/2,β ′
l
(s),
t, s ∈ [−1, 1]2,
in which ak,l(K d,d′
i
) ≥ 0, k, l ∈ Z+ and P∞
Proof. Consider the function g defined by the Fourier expansion
k,l=0 ak,l(K d,d′
i
)P (d−2)/2,β
k
(1)P (d′−2)/2,β ′
l
(1) < ∞.
g(t, s) ∼
∞
Xk,l=0
ak,l(K d,d′
i
)P (d−2)/2,β
k
(t)P (d′−2)/2,β ′
l
(s),
t, s ∈ [−1, 1].
If K is positive definite, then Proposition 2.5 guarantees the convergence of the series for
t = s = 1. Proposition 2.6 implies convergence for all the other values of t and s while
Proposition 2.4 yields that all the coefficients in the expansion are nonnegative. Since g
is continuous and the Fourier coefficients of K d,d′
coincide with those of g, it follows that
K d,d′
i = g. This takes care of one implication in the theorem. As for the other, it follows
from Lemma 2.2 and the fact that the pointwise limit of positive definite kernels is itself
positive definite.
i
9
3 DC-strict positive definiteness
Either one of the concepts of strict positive definiteness we have introduced so far, demands
considering n × n matrices A = [Aµν] with
Aµν = K d,d′
i
(cos (xµxν/2), cos (wµwν/2)),
in which K d,d′
points in Md × Hd′. Analyzing the associated quadratic forms
is the isotropic part of the kernel and (xµ, wµ), µ = 1, 2, . . . , n, are distinct
i
ctAc :=
n
Xµ,ν=1
cµcνK((xµ, wµ), (xν, wν)),
cµ ∈ R, µ = 1, 2, . . . , n,
it is possible to obtain a quite more convenient formulation for either concept. We will
proceed discussing DC-strict positive definiteness and will just mention the formulation
for plain strict positive definiteness later. From now on, if K is a real, continuous, isotropic
and positive definite kernel K on Md × Hd′, we will use the following notation attached
to the series representation of its isotropic part:
At this point, we need the addition formula demonstrated by Gin´e ([10, 16]), that is,
JK :=n(k, l) : a(d−2)/2,β
k,l
> 0o .
δ(k,d)
Xj=1
cd,β
k
:=
where
Sd
k,j(x)Sd
k,j(y) = cd,β
k P (d−2)/2,β
k
(cos (xy/2)) ,
x, y ∈ Md,
Γ(β + 1)(2k + (d − 2)/2 + β + 1)Γ(k + (d − 2)/2 + β + 1)
Γ((d − 2)/2 + β + 2)Γ(k + β + 1)
.
The set {Sd
harmonics of degree k on Md.
k,2, . . . , Sd
k,1, Sd
k,δ(k,d)} denotes an orthonormal basis of the space Hd
k of spherical
If we consider the representation for K provided by Theorem 2.7 and the addition
formula above, then the equality ctAc = 0 corresponds to
∞
Xk,l=0
ak,l(K d,d′
cα,β
k cα′,β ′
i
l
)
δ(k,d)
Xi=1
In particular, ctAc = 0 if, and only if,
δ(l,d′)
n
Xj=1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xµ=1
cµSd
k,i(xµ)Sd′
2
l,j(wν)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
= 0.
cµSd
k,i(xµ)Sd′
l,j(wν) = 0, (k, l) ∈ JK, i ∈ {1, 2, . . . , δ(k, d)}, j ∈ {1, 2, . . . , δ(l, d′)}.
n
Xµ=1
10
Reintroducing the addition formula, now leaving a free variable (x, w) ∈ Md × Hd′, the
previous assertion implies that
cµP (d−2)/2,β
k
(cos (xµx/2))P (d′−2)/2,β ′
l
(cos (wµw/2)) = 0,
n
Xµ=1
for (x, w) ∈ Md × Hd′ and (k, l) ∈ JK. However, if this last assertion holds, it is promptly
seen that
cµ
n
Xµ=1
δ(k,d)
δ(l,d′)
Xi=1
Xj=1
Sd
k,i(xµ)Sd
k,i(x)Sd′
l,j(wµ)Sd′
l,j(w) = 0,
(x, w) ∈ Md × Hd′
,
(k, l) ∈ JK.
Using the fact that {Sd′
and Hd
k, respectively, we are reduced to
l,1, Sd′
l,2, . . . , Sd′
l,δ(l,d′)} and {Sd
k,1, Sd
k,2, . . . , Sd
k,δ(k,d)} are basis of Hd′
l
cµSd
k,i(xµ)Sd′
l,j(wν) = 0, (k, l) ∈ JK, i ∈ {1, 2, . . . , δ(k, d)}, j ∈ {1, 2, . . . , δ(l, d′)}.
n
Xµ=1
once again.
The discussion above justifies the following result.
Proposition 3.1. Let K be a real, continuous, isotropic and positive definite kernel on
Md × Hd′. The following assertions are equivalent:
(i) K is DC-strictly positive definite;
(ii) If n ≥ 1, x1, x2, . . . , xn are distinct points on Md and w1, w2, . . . , wn are distinct points
on Hd′, then the only solution of the system
k
Pn
µ=1 cµP (d−2)/2,β
(x, w) ∈ Md × Hd′,
(k, l) ∈ JK,
(cos (xµx/2))P (d′−2)/2,β ′
l
(cos (wµw/2)) = 0,
is the trivial one, that is, cµ = 0, µ = 1, 2, . . . , n.
In the lemma below, we use the symbol M1 ֒→ M2 to indicate the existence of an
isometric embedding of a metric space M1 into a metric space M2. The result is a classical
result in the theory of compact two-point homogeneous spaces (see [1]).
Lemma 3.2. There exists a chain of isometric embeddings as follows
S1 ֒→ P2(R) ֒→ Pd(R) ֒→ P2d(C) ֒→ P4d(H) ֒→ P8d(Cay),
d = 2, 3, . . . .
11
In particular, since P2(R) is isometrically isomorphic to S2, if the compact two-point
homogeneous space Hd′ is not a sphere, the lemma guarantees the existence of an integer
q ≥ 2 so that Sq ֒→ Hd′. On the other hand, this embedding justifies a decomposition of
the form
l
P (d′−2)/2,β ′
l
(s) =
with all coefficients bl
j positive.
jP (q−2)/2,(q−2)/2
bl
l−j
(s),
l = 0, 1, . . . ,
Xj=0
We will make use of the following normalized Jacobi polynomials
Rα,β
k =
P α,β
k
P α,β
k
(1)
and some of its properties listed in the lemma below (see [3, 20]).
(−t) = (−1)kP β,α
Lemma 3.3. The Jacobi polynomials have the following properties:
(i) P α,β
(ii) limk→∞ Rα,β
(iii) If α > β, then limk→∞ P β,α
(t) = 0, t ∈ (−1, 1);
(t), t ∈ [−1, 1];
(1)]−1 = 0.
(1)[P α,β
k
k
k
k
k
This is the first characterization for DC-strict positive definiteness we have found.
Theorem 3.4. Let K be a real, continuous, isotropic and positive definite kernel on
Sd × Hd′. Assume d ≥ 2 and that Hd′ is not a sphere. In order that K be DC-strictly
positive definite it is necessary and sufficient that either {l : (k, l) ∈ JK for some k} be
infinite or JK contain two sequences {(kr, l)} and {(ks, l′)} for which {kr + l} ⊂ 2Z+,
{ks + l′} ⊂ 2Z+ + 1, and limr→∞ kr = lims→∞ ks = ∞.
Proof. Assume K is DC-strictly positive definite. Recalling Lemma 3.2, it is easily seen
that K is DC-strictly positive definite on Sd×Sq for some q ≥ 2. Introducing the equalities
presented right after Lemma 3.2 into the series representation for K d,d′
and arranging leads
to
i
K d,d′
i
(t, s) =
∞
Xk,l=0 ∞
Xj=0
ak,l+j(K d,d′
i
In particular, the set
)bl+j
l ! P (d−2)/2,(d−2)/2
k
(t)P (q−2)/2,(q−2)/2
l
(s), t, s ∈ [−1, 1].
(k + l :
∞
Xj=0
ak,l+j(K d,d′
i
)bl+j
l > 0) =(k + l :
ak,l+j(K d,d′
i
) > 0)
∞
Xj=0
contains infinitely many even and infinitely many odd integers. However, it is not hard
to see that if the above condition holds and {l : ak,l(K d,d′
) > 0 for some k} is finite,
i
12
then JK must contain two sequences {(kr, l)} and {(ks, l′)} for which {kr + l} ⊂ 2Z+,
{ks + l′} ⊂ 2Z+ + 1, and limr→∞ kr = lims→∞ ks = ∞. Indeed, the inferring of this fact
demands to observe that if ak,l(K d,d′
) > 0 for some (k, l), then
i
k + 0, k + 1, . . . , k + l ∈(k + l :
ak,l+j(K d,d′
i
) > 0) .
∞
Xj=0
This shows the necessity of the condition. As for the sufficiency,
let n be a positive
integer, x1, x2, . . . , xn distinct points in Md and w1, w2, . . . , wn distinct points in Hd′. We
will show that, under the condition on JK mentioned in the statement of the theorem,
the only solution of the system
k
Pn
µ=1 cµP (d−2)/2,β
(x, w) ∈ Md × Hd′,
(k, l) ∈ JK,
(cos (xµx/2))P (d′−2)/2,β ′
l
(cos (wµw/2)) = 0,
is the trivial one. In order to achieve that, we will fix γ ∈ {1, 2, . . . , n} and will show that
cγ = 0. That will be done trough specific choices of points x ∈ Md and w ∈ Hd′ in the
equation defining the system. We also need to consider the antipodal index sets
Γxγ = {µ : cos (xµxγ/2) = −1} and Γwγ = {µ : cos (wµwγ/2) = −1}.
Due to the basic assumptions of the theorem, we know that d−2 = 2β and that Γxγ is uni-
tary, say, Γxγ = {δ}. The Jacobi polynomials P (d−2)/2,β
are then Gegenbauer polynomials
and, in particular, they are even functions when k is even and odd functions otherwise.
The equation defining the system, with the choice x = xδ and w = wδ, can be put into
the form
k
cγ + (−1)k+l P β ′,(d′−2)/2
P (d′−2)/2,β ′
l
l
+(−1)k Xµ∈{δ}\Γwγ
cµ
(1)
(1) Xµ∈{δ}∩Γwγ
cµR(d′−2)/2,β ′
l
(cos (wµwγ/2))
+(−1)l P β ′,(d′−2)/2
P (d′−2)/2,β ′
l
l
(1)
(1) Xµ∈Γwγ \{δ}
+ Xµ /∈{δ}∪Γwγ
cλR(d−2)/2,β
k
(cos (xµxγ/2))
cµR(d−2)/2,β
(cos (xµxγ/2))R(d′−2)/2,β ′
k
l
(cos (wµwγ/2)) = 0.
Obviously, some of the sets appearing in the sum decomposition above may be empty.
Also, the first two sums cannot co-exist, that is, just one of them can appear in the
13
expression. If JK contains a sequence (kr, lr) for which limr→∞ lr = ∞, we may conclude
that
lim
r→∞
due to Lemma 3.3-(iii), while
P β ′,(d′−2)/2
lr
P (d′−2)/2,β
lr
(1)
(1)
= 0,
lim
r→∞
R(d′−2)/2,β ′
lr
(cos (wµwγ/2)) = 0, µ 6∈ Γxγ ,
due to Lemma 3.3-(ii). Hence, the limit of each summand, but the first, in the previous
expression vanishes. In particular, cγ = 0. We now proceed assuming the existence of two
sequences {(kr, l)} and {(ks, l′)} in JK for which {kr + l} ⊂ 2Z+, {ks + l′} ⊂ 2Z+ + 1,
and limr→∞ kr = lims→∞ ks = ∞. If the second summand in the expression occurs, we
can employ these two sequences to deduce that
cγ +
P β ′,(d′−2)/2
l
P (d′−2)/2,β ′
l
(1)
(1)
cδ = cγ −
P β ′,(d′−2)/2
l′
P (d′−2)/2,β ′
l′
(1)
(1)
cδ = 0,
after letting r → ∞ and s → ∞. We observe that the limits of the two last summands
in the original equation are equal to 0 in this case. Now, if cδ 6= 0, the first equality
above provides a contradiction with the positivity of the gamma function in (0, ∞). Thus,
0 = cδ = cγ. Finally, if the third summand is the one occurring in the original expression,
we need an additional equation provided by a second choice of points in the equation
defining the system. Choosing x = xδ and w = wδ leads to
cδ + (−1)k+l P β ′,(d′−2)/2
P (d′−2)/2,β ′
(1)
(1)
cγ
l
l
+ (−1)k Xµ∈{γ}\Γwδ
cγR(d′−2)/2
l
(cos (wµwδ/2)
+(−1)l P β ′,(d′−2)/2
P (d′−2)/2,β ′
l
l
(1)
(1) Xµ∈Γwδ \{γ}
+ Xµ /∈Γwδ ∪{γ}
R(d−2)/2,β
k
(cos (xµxδ/2))
cµR(d−2)/2,β
(cos (xµxδ/2))R(d′−2)/2,β ′
k
l
(cos (wµwδ/2)) = 0.
Using just one of the sequences, say, {(kr, l)}, and letting r → ∞ in both equations, we
deduce that
cγ + Xµ∈{δ}\Γwγ
cµR(d′−2)/2
l
(cos (wµwγ/2) = cδ + Xµ∈{γ}\Γwδ
cµR(d′−2)/2
l
(cos (wµwδ/2) = 0.
14
But that corresponds to
cδ(cid:20)1 −(cid:16)R(d′−2)/2
l
(cos (wγwδ/2))(cid:17)2(cid:21) = 0,
with cos (wγwδ/2) 6= ±1. Thus, cδ = 0, and consequently, cγ = 0.
The next theorem takes care of the the remaining cases.
Theorem 3.5. Let K be a real, continuous, isotropic and positive definite kernel on
Md × Hd′. Assume neither Md nor Hd′ is a sphere. In order that K be DC-strictly positive
definite it is necessary and sufficient that {k + l : ak,l(K d,d′
) > 0} be infinite.
i
Proof. Since the proof is similar to the proof of the previous theorem, some details will
be omitted. The necessity part is similar to that in the proof of Theorem 3.4, using the
same trick twice. The resulting kernel is DC-strictly positive definite on some Sq × Sq
and the set of indices pertaining to the final argument takes the form
(k + l :
∞
∞
Xj=0
Xj ′=0
ak+j,l+j ′(K d,d′
i
)bk+j
k
cl+j ′
l > 0) =(k + l :
∞
∞
Xj=0
Xj ′=0
ak+j,l+j ′(K d,d′
i
) > 0) ,
k
cl+j ′
l
where all the constants bk+j
are positive. Since this set has infinitely many even and
infinitely many odd integers, it follows that {k + l : (k, l) ∈ JK} is infinite. The sufficiency
part follows the steps of the corresponding part in the previous theorem. Due to the
assumption on JK, we can select a sequence {(kr, lr)} in JK so that either limr→∞ kr = ∞
or limr→∞ lr = ∞. Choosing x = xγ and w = wγ in the equation defining the system, we
obtain
cγ + (−1)kr+lr
kr
P β,(d−2)/2
P (d−2)/2,β
kr
(1)
(1)
P β ′,(d′−2)/2
lr
P (d′−2)/2,β ′
lr
(1)
(1) Xµ∈Γxγ ∩Γwγ
cµ
+(−1)kr
kr
P β,(d−2)/2
P (d−2)/2,β
kr
(1)
(1) Xµ∈Γxγ \Γwγ
cµR(d′−2)/2,β ′
lr
(cos (wµwγ/2))
+(−1)lr
P β ′,(d′−2)/2
lr
P (d′−2)/2,β ′
lr
(1)
(1) Xµ∈Γwγ \Γxγ
+ Xµ /∈Γxγ ∪Γwγ
cλR(d−2)/2,β
kr
(cos (xµxγ/2))
cµR(d−2)/2,β
kr
(cos (xµxγ/2))R(d′−2)/2,β ′
lr
(cos (wµwγ/2)) = 0,
If limr→∞ kr = ∞, then the limit of each summand, but the first, vanishes. In particular,
cγ = 0. If limr→∞ lr = ∞, a similar analysis produces the same conclusion.
15
4 Strict positive definiteness
The strict positive definiteness of a real, continuous, isotropic and positive definite kernel
on a product of high dimensional spheres was completely characterized in [14] while the
characterization in the case of a product of circles was reached in [13]. Thus, just like
in the previous section, we will assume that at least one of the spaces involved is not a
sphere.
The section begins with the obvious counterpart of Proposition 4.1 for plain strict
positive definiteness on Md × Hd′.
Proposition 4.1. Let K be a real, continuous, isotropic and positive definite kernel on
Md × Hd′. The following assertions are equivalent:
(i) K is strictly positive definite;
(ii) If n ≥ 1 and (x1, w1), (x2, w2), . . . , (xn, wn) are distinct points on Md × Hd′, then the
only solution of the system
k
Pn
µ=1 cµP (d−2)/2,β
(x, w) ∈ Md × Hd′,
(k, l) ∈ JK,
(cos (xµx/2))P (d′−2)/2,β ′
l
(cos (wµw/2)) = 0,
is the trivial one, that is, cµ = 0, µ = 1, 2, . . . , n.
The characterization for strict positive definiteness in the case in which both spaces
are not spheres is as follows.
Theorem 4.2. Let K be a real, continuous, isotropic and positive definite kernel on
Md × Hd′. Assume that neither Md nor Hd′ is a sphere. In order that K be strictly positive
definite it is necessary and sufficient that the set JK contains a sequence {(kr, lr)} for
which limr→∞ kr = limr→∞ lr = ∞.
Proof. Assume there exists a sequence {(kr, lr)} as described in the statement of the
theorem. Let (x1, w1), (x2, w2), . . . , (xn, wn) be distinct points in Md × Hd′ and suppose
that
n
cµP (d−2)/2,β
k
(cos (xµw/2))P (d′−2)/2,β ′
l
(cos (wµw/2) = 0,
Xµ=1
for real scalars c1, c2, . . . , cn, (x, w) ∈ Md × Hd′ and (k, l) ∈ JK. For γ ∈ {1, 2, . . . , n} fixed,
let us put x = xγ and w = wγ in the previous equation and split it taking into account
the following index sets (recall the normalization we have adopted for the metric in the
spaces involved):
I1 = {µ : xµxγ = 2π = wµwγ},
I2 = {µ : xµxγ = 2π 6= wµwγ},
16
and
We observe that one or more of these sets may be empty. The outcome is
I3 = {µ : xµxγ 6= 2π = wµwγ}.
cγP (d−2)/2,β
k
(1)P (d′−2)/2,β ′
l
(1) + (−1)k+lP β,(d−2)/2
k
(1)P β ′,(d′−2)/2
l
cµ
(1)Xµ∈I1
+ (−1)kP β,(d−2)/2
k
cµP (d′−2)/2,β ′
l
(cos (wµwγ/2))
+ (−1)lP β ′,(d′−2)/2
l
cµP (d−2)/2,β
k
(cos (xµxγ/2))
(1)Xµ∈I2
(1)Xµ∈I3
+ Xµ /∈I1∪I2∪I3
cµP (d−2)/2,β
k
A small adjustment implies that
(cos (xµxγ/2))P (d′−2)/2,β ′
l
(cos (wµwγ/2)) = 0.
cγ + (−1)kr+lr
kr
P β,(d−2)/2
P (d−2)/2,β
kr
(1)
(1)
P β ′,(d′−2)/2
lr
P (d′−2)/2,β ′
lr
(1)
(1) Xµ∈I1
cµ
+(−1)kr
kr
P β,(d−2)/2
P (d−2)/2,β
kr
(1)
(1) Xµ∈I2
cµR(d′−2)/2,β ′
lr
(cos (wµwγ/2))
+(−1)lr
P β ′,(d′−2)/2
lr
P (d′−2)/2,β ′
lr
(1)
(1) Xµ∈I3
cµR(d−2)/2,β
kr
(cos (xµxγ/2))
cµR(d−2)/2,β
kr
(cos (xµxγ/2))R(d′−2)/2,β ′
lr
(cos (wµwγ/2)) = 0,
r = 1, 2, . . . .
+ Xµ /∈I1∪I2∪I3
We observe that in the last summand, if µ is fixed, either xγ 6= xµ or wγ 6= wµ. Since
α > β and α′ > β′, we may let r → ∞ and apply Lemma 3.3 to conclude that cγ = 0.
In view of the previous proposition, the sufficiency part is resolved. Going the other way
around, if K is strictly positive definite, we may repeat the procedure adopted in the first
half of the proof of Theorem 3.5. The index set of the resulting positive definite kernel on
Sq × Sq is
((k, l) :
∞
∞
Xj=0
Xj ′=0
ak+j,l+j ′(K d,d′
i
) > 0) .
Since the characterization for strict positive definiteness on Sq × Sq described in [14]
implies that the set above must contain at least one sequence (kr, lr) for which limr→∞ kr =
limr→∞ lr = ∞, the set JK must contain a sequence of this same type.
In the case in which Md = Sd, the following upgrade of the previous lemma will be
more favorable.
17
Lemma 4.3. Let K be a real, continuous, isotropic and positive definite kernel on Sd×Hd′,
in which d ≥ 2 and Hd′ is not a sphere. The following statements are equivalent:
(i) K is strictly positive definite on Sd × Hd′;
(ii) If n ≥ 1, (x1, w1), (x2, w2) . . . , (xn, wn) are distinct points on Sd × Hd′, and the set
{x1, x2, . . . , xn} does not contain any pair of antipodal points, then the only solution of
the system
µ + c′′
Pn
µ=1(cid:2)(−1)kc′
(x, w) ∈ Sd × Hd′
(k, l) ∈ JK,
k
µ(cid:3) P (d−2)/2,β
(xµ · x)P (d′−2)/2,β ′
l
(cos (wµw/2)) = 0,
is c′
µ = c′′
µ = 0, µ = 1, 2, . . . , n.
Proof. Assume (i) holds. Let (x1, w1), (x2, w2), . . . , (xn, wn) be distinct points in Sd × Hd′
and assume that {x1, x2, . . . , xn} does not contain pairs of antipodal points. Since 2β =
d − 2, the system described in (ii) can be written in the form
2n
cνP (d−2)/2,β
k
(cos (x′
νx/2))P (d′−2)/2,α′
l
(cos (w′
νw/2)) = 0,
x ∈ Sd, w ∈ Hd′
,
ν, w′
ν) = (xν, wν) and cν = c′
ν if ν ∈ {n + 1, 2, . . . , 2n}. Since the 2n points (x′
in which (x′
cν = c′′
implies that cν = 0, ν = 1, 2, . . . , 2n. In particular, c′
Conversely, if (i) does not hold, the previous proposition allows the selection of distinct
point (x1, w1), (x2, w2), . . . , (xn, wn) in Sd × Hd so that the system
ν) = (−xν, wν) and
ν) are distinct, Proposition 4.1
µ = 0, µ = 1, 2, . . . , n.
ν if ν ∈ {1, 2, . . . , n} and (x′
ν, w′
µ = c′′
ν, w′
Xν=1
k
Pn
µ=1 cµP (d−2)/2,β
(x, w) ∈ Sd × Hd′,
(k, l) ∈ JK,
(cos (xµx/2))P (d′−2)/2,α′
l
(cos (wµw/2)) = 0,
has a nontrivial solution cµ, µ = 1, 2, . . . , cn. We can select p (≤ n) distinct points
{(x′
p} con-
tains no pairs of antipodal points and
p)} in Sd × Hd′ in a such a way that {x′
2), . . . , (x′
2, . . . , x′
1), (x′
1, w′
p, w′
2, w′
1, x′
{(x1, w1), (x2, w2), . . . , (xn, wn) ⊆ {(±x′
1, w′
1), (±x′
2, w′
2), . . . , (±x′
p, w′
p)}.
However, it is an easy matter to verify that the system
n
µ + c′′
k
µ(cid:3) P (d−2)/2,β
Xµ=1(cid:2)(−1)kc′
(x, w) ∈ Sd × Hd′,
(k, l) ∈ JK,
(cos (x′
µx/2))P (d′−2)/2,α′
l
(cos (w′
µw/2)) = 0,
has a nontrivial solution as well. Thus, (ii) cannot hold.
18
The following proposition is an alternative to the previous lemma via the sets
J e
K := JK ∩ [2Z+ × Z+]
and J o
K := JK ∩ [(2Z+ + 1) × Z+].
Proposition 4.4. Let K be a real, continuous, isotropic and positive definite kernel on
Sd × Hd′, in which d ≥ 2 and Hd′ is not a sphere. The following statements are equivalent:
(i) K is strictly positive definite on Sd × Hd′;
(ii) If n ≥ 1, (x1, w1), (x2, w2), . . . , (xn, wn) are distinct points on Sd × Hd′, and the set
{x1, x2, . . . , xn} does not contain a pair of antipodal points, then the only solution of the
system
µP (d−2)/2,β
ce
k
(cos (xµx/2))P (d′−2)/2,β ′
l
(cos (wµw/2)) = 0,
(k, l) ∈ J e
K,
µP (d−2)/2,β
co
k′
(cos (xµx/2))P (d′−2)/2,β ′
l′
(cos (wµw/2)) = 0,
(k′, l′) ∈ J o
K,
Proof. If (ii) were not true, we could find distinct points (x1, w1), (x2, w2), . . . ,
(xn, wn) in Sd × Hd′, with {x1, x2, . . . , xn} containing no pair of antipodal points and
either
(cos (xµx/2))P (d′−2)/2,β ′
l
(cos (wµw/2)) = 0,
(k, l) ∈ J e
K,
(x, w) ∈ Sd × Hd′,
is ce
µ = co
µ = 0, µ = 1, 2, . . . , n.
n
n
Xµ=1
Xµ=1
or
( Pn
( Pn
µ=1 ce
µP (d−2)/2,β
(x, w) ∈ Sd × Hd′,
k
µ=1 co
µP (d−2)/2,β
(x, w) ∈ Sd × Hd′
k′
(cos (xµx/2))P (d′−2)/2,β ′
l′
(cos (wµw/2)) = 0,
(k′, l′) ∈ J o
K,
having a nontrivial solution. We proceed considering the first possibility that emerges
from the conclusion above, being the other case similar. For each µ ∈ {1, 2, . . . , n}, the
system
(cid:26) c′
µ + c′′
µ + c′′
µ = ce
µ
µ = 0
−c′
µ 6= 0 for at least one µ, then (c′
,
µ, c′′
µ. Since ce
has a unique solution c′
µ) 6= 0, for at least
one µ. It is now clear that the system in Lemma 4.1-(ii) would have a nontrivial solution
for the selection of points (x1, w1), (x2, w2), . . . , (xn, wn). Thus, (i) implies (ii).
The converse will be justified as long as we show that if (ii) holds, then Lemma 4.1-(ii)
holds. But, if ce
µ are known, the system
µ and co
µ, c′′
(cid:26) c′
−c′
µ + c′′
µ + c′′
µ = ce
µ
µ = co
µ
,
19
always has a unique solution. If ce
µ = 0 for all µ, then the solution of the corresponding
system vanishes. Thus, if the system in (ii) has the trivial solution only, the same will be
true of the system in Lemma 4.1-(ii).
µ = co
We are ready to state and prove the last main contribution of the paper.
r, l′
r)} so that {kr} ⊂ 2Z+, {k′
Theorem 4.5. Let K be a real, continuous, isotropic and positive definite kernel on
Sd × Hd′. Assume that d ≥ 2 and that Hd′ is not a sphere. In order that K be strictly
positive definite it is necessary and sufficient that the set JK contain sequences {(kr, lr)}
and {(k′
r =
limr→∞ lr = limr→∞ l′
Proof. Let us assume that JK contains sequences as described in the statement of the
theorem. We intend to use Proposition 4.4 in order to conclude that K is strictly positive
definite. Let (x1, w1), (x2, w2), . . . , (xn, wn) be distinct points in Sd × Hd′, assume that
{x1, x2, . . . , xn} does not contain any pairs of antipodal points and that the system in
Proposition 4.4-(ii) holds. Fixing γ, introducing x = xγ and w = wγ in the first equation
of the system and proceeding as in the proof of Theorem 4.2, we deduce that
r} ⊂ 2Z+ + 1, and limr→∞ kr = limr→∞ k′
r = ∞.
l
γ + (−1)l P β ′,(d′−2)/2
ce
P (d′−2)/2,β ′
+Xµ∈I2
l
(1)
(1) Xµ∈I1
µR(d−2)/2,β
ce
k
(cos (xµxγ/2))
µR(d−2)/2,β
ce
k
(cos (xµxγ/2))R(d′−2)/2,β ′
l
(cos (wµwγ/2)) = 0,
in which the index sets are now I1 = {µ : wµwγ = 2π} and I2 = {µ : wµwγ 6= 2π}.
Substituting the first double sequence guaranteed by our assumption in this equation and
letting r → ∞, Lemma 3.3 implies that ce
γ = 0. A similar procedure with the second
equation of the system and with the second double sequence from the assumption leads
to co
γ = 0. Since γ is arbitrary, the only solution of the system in Proposition 4.4-(ii) is
the trivial one. Therefore, K is strictly positive definite on Sd × Hd′ . In order to prove the
condition is necessary, we need to imitate the corresponding part in the proof of Theorem
3.4. We get a strictly positive definite kernel on Sd × Sq with corresponding index set
((k, l) :
ak,l+j(K d,d′
i
) > 0) .
∞
Xj=0
Once again, the characterization for strict positive definiteness on Sd × Sq described in
[14] implies that Jk must contain two sequences as quoted in the statement of the theorem
being proved.
As a final remark, we would like to observe that it is still an open problem to obtain
versions of the theorems proved in Sections 3 and 4 in the cases in which the spaces are
different and at least one of them is S1. As a matter of fact, for the theorems proved in
Sections 3, it is also open the case in which both spaces are S1.
20
References
[1] Askey, R., Orthogonal polynomials and special functions. Society for Industrial and
Applied Mathematics, Philadelphia, Pa., 1975.
[2] C. Bachoc, Semidefinite programming, harmonic analysis and coding theory. ArXiv
09094767, 2010.
[3] Barbosa, V. S.; Menegatto, V. A., Strictly positive definite kernels on two-point com-
pact homogeneous spaces, Math. Ineq. Appl., 19 (2016), no. 2, 743-756.
[4] C. Berg; E. Porcu, From Schoenberg coefficients to Schoenberg functions. Constr.
Approx., to appear.
[5] Chen, Debao; Menegatto, V. A.; Sun, Xingping, A necessary and sufficient condition
for strictly positive definite functions on spheres. Proc. Amer. Math. Soc. 131 (2003),
no. 9, 2733-2740.
[6] Cheney, E. W., Approximation using positive definite functions. Approximation theory
VIII, Vol. 1 (College Station, TX, 1995), 145-168, Ser. Approx. Decompos., 6, World
Sci. Publ., River Edge, NJ, 1995.
[7] Erd´elyi, A.; Magnus, W.; Oberhettinger, F.; Tricomi, F. G., Higher transcendental
functions. Vols. I, II. Based, in part, on notes left by Harry Bateman. McGraw-Hill
Book Company, Inc., New York-Toronto-London, 1953.
[8] Freeden, W.; Gervens, T.; Schreiner, M., Constructive approximation on the sphere.
With applications to geomathematics. Numerical Mathematics and Scientific Compu-
tation. The Clarendon Press, Oxford University Press, New York, 1998.
[9] Gangolli, R., Positive definite kernels on homogeneous spaces and certain stochastic
processes related to L´evy's Brownian motion of several parameters. Ann. Inst. H.
Poincar´e Sect. B (N.S.) 3 (1967), 121-226.
[10] Gin´e, E., The addition formula for the eigenfunctions of the Laplacian. Advances in
Math. 18 (1975), no. 1, 102-107.
[11] Gneiting, T., Strictly and non-strictly positive definite functions on spheres. Bernoulli
19 (2013), no. 4, 1327-1349.
[12] Guella, J. C.; Menegatto, V. A.; Peron, A. P., An extension of a theorem of Schoen-
berg to a product of spheres, Banach J. Math. Anal., to appear.
[13] Guella, J. C.; Menegatto, V. A.; Peron, A. P., Strictly positive definite kernels on a
product of circles, Positivity, to appear.
21
[14] Guella, J. C.; Menegatto, V. A.; Strictly positive definite kernels on a product of
spheres. J. Math. Anal. Appl. 435 (2016), no. 1, 286-301.
[15] Guella, J. C.; Menegatto, V. A.; Peron, A. P., Strictly positive definite kernels on a
product of spheres II. Preprint, 2016.
[16] Koornwinder, T., The addition formula for Jacobi polynomials and spherical har-
monics. Lie algebras: applications and computational methods (Conf., Drexel Univ.,
Philadelphia, Pa., 1972). SIAM J. Appl. Math. 25(1973), 236-246.
[17] Meaney, C., Localization of spherical harmonic expansions. Monatsh. Math. 98
(1984), no. 1, 65-74.
[18] Rainville, E. D., Special functions. The Macmillan Co., New York 1960.
[19] Schoenberg, I, J., Positive definite functions on spheres. Duke Math. J. 9, (1942),
96-108.
[20] Szego, G., Orthogonal polynomials. Fourth edition. American Mathematical Society,
Colloquium Publications, Vol. XXIII, American Mathematical Society, Providence,
R.I., 1975.
[21] Wang, Hsien-Chung, Two-point homogeneous spaces. Ann. Math. 55 (1952), no. 2,
177-191.
[22] Wendland, H., Scattered data approximation. Cambridge Monographs on Applied
and Computational Mathematics, 17. Cambridge University Press, Cambridge, 2005.
[23] Whittaker, E. T.; Watson, G. N., A course of modern analysis. An introduction to the
general theory of infinite processes and of analytic functions; with an account of the
principal transcendental functions. Reprint of the fourth (1927) edition. Cambridge
Mathematical Library. Cambridge University Press, Cambridge, 1996.
V. S. Barbosa and V. A. Menegatto
Departamento de Matem´atica,
ICMC-USP - Sao Carlos, Caixa Postal 668,
13560-970 Sao Carlos SP, Brasil
e-mails: [email protected]; [email protected]
22
|
1201.6226 | 1 | 1201 | 2012-01-25T15:34:35 | New integral inequalities via $(\alpha,m)$-convexity and quasi-convexity | [
"math.FA"
] | In this paper, we establish some new integral inequalities for $(\alpha, m)-$convex functions and quasi-convex functions, respectively. Our results in special cases recapture known results. | math.FA | math |
NEW INTEGRAL INEQUALITIES VIA (α, m)-CONVEXITY AND
QUASI-CONVEXITY
WENJUN LIU
Abstract. In this paper, we establish some new integral inequalities for (α, m)−convex functions and
quasi-convex functions, respectively. Our results in special cases recapture known results.
Let I be on interval in R. Then f : I → R is said to be convex (see [17, P.1]) if
f (tx + (1 − t) y) ≤ tf (x) + (1 − t) f (y)
1. INTRODUCTION
holds for all x, y ∈ I and t ∈ [0, 1].
In [27], Toader defined m-convexity as follows:
Definition 1. The function f : [0, b] → R, b > 0 is said to be m-convex, where m ∈ [0, 1], if
holds for all x, y ∈ [0, b] and t ∈ [0, 1] .We say that f is m−concave if −f is m−convex.
f (tx + m (1 − t) y) ≤ tf (x) + m (1 − t) f (y)
In [18], Mihe¸san defined (α, m) − convexity as follows:
Definition 2. The function f : [0, b] → R, b > 0, is said to be (α, m) − convex, where (α, m) ∈ [0, 1]2, if
holds for all x, y ∈ [0, b] and t ∈ [0, 1].
f (tx + m(1 − t)y) ≤ tαf (x) + m(1 − tα)f (y)
Denote by K α
m(b) the class of all (α, m) −convex functions on [0, b] for which f (0) ≤ 0.
It can be
easily seen that for (α, m) = (1, m) , (α, m) − convexity reduces to m− convexity and for (α, m) = (1, 1),
(α, m) − convexity reduces to the concept of usual convexity defined on [0, b], b > 0. For recent results
and generalizations concerning m−convex and (α, m) −convex functions see [4, 6, 10, 19, 21, 26].
We recall that the notion of quasi-convex functions generalizes the notion of convex functions. More
precisely, a function f : [a, b] → R is said to be quasi-convex on [a, b] if
f (λx + (1 − λ)y) ≤ max{f (x), f (y)}
holds for any x, y ∈ [a, b] and λ ∈ [0, 1]. Clearly, any convex function is a quasi-convex function.
Furthermore, there exist quasi-convex functions which are not convex (see [14]).
One of the most famous inequalities for convex functions is Hadamard's inequality. This double
inequality is stated as follows: Let f be a convex function on some nonempty interval [a, b] of real line
R, where a 6= b. Then
(1.1)
f (x)dx ≤
f (a) + f (b)
2
.
a
f(cid:18) a + b
2 (cid:19) ≤
1
b − aZ b
Hadamard's inequality for convex functions has received renewed attention in recent years and a re-
markable variety of refinements and generalizations have been found (see, for example, [1]-[19], [22]-[26],
[28]). In [4], Bakula et al. establish several Hadamard type inequalities for differentiable m−convex and
(α, m) −convex functions.
2000 Mathematics Subject Classification. 26D15, 26A51, 39B62.
Key words and phrases. Hermite's inequality, Holder's inequality, (α, m)-convexity, quasi-convexity.
1
2
W. J. LIU
Recently, Ion [14] established two estimates on the Hermite-Hadamard inequality for functions whose
first derivatives in absolute value are quasi-convex. Namely, he obtained the following results:
Theorem 1.1. Let f : I ⊂ R → R be a differentiable mapping on I, a, b ∈ I with a < b. If f ′ is
quasi-convex on [a, b], then the following inequality holds:
f (a) + f (b)
2
−
1
b − aZ b
a
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f (u)du(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤
b − a
4
{max f ′(a) , f ′(b)} .
Theorem 1.2. Let f : I ⊂ R → R be a differentiable mapping on I, a, b ∈ I with a < b and let p > 1. If
f ′
p−1 is quasi-convex on [a, b], then the following inequality holds:
p
f (a) + f (b)
2
−
1
b − aZ b
a
≤
b − a
2(p + 1)
1
p (cid:16)maxnf ′(a)
p
p−1 , f ′(b)
p−1
p
.
p
p−1o(cid:17)
In [2], Alomari et al. obtained the following result.
f (u)du(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Theorem 1.3. Let f : I ⊂ R → R be a differentiable mapping on I, a, b ∈ I with a < b and let q ≥ 1. If
f ′q is quasi-convex on [a, b], then the following inequality holds:
f (a) + f (b)
2
−
1
b − aZ b
a
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤
b − a
4
f (u)du(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:0)max(cid:8)f ′(a)q
, f ′(b)q(cid:9)(cid:1)
1
q .
In [20], Ozdemir et al. used the following lemma in order to establish several integral inequalities via
some kinds of convexity.
Lemma 1.1. Let f : [a, b] ⊂ [0, ∞) → R be continuous on [a, b] such that f ∈ L([a, b]), a < b. Then the
equality
(1.2)
Z b
a
(x − a)p(b − x)qf (x)dx = (b − a)p+q+1Z 1
0
holds for some fixed p, q > 0.
(1 − t)ptqf (ta + (1 − t)b)dt
Especially, Ozdemir et al. [20] discussed the following new results connecting with m−convex function
and quasi-convex function, respectively:
Theorem 1.4. Let f : [a, b] → R be continuous on [a, b] such that f ∈ L([a, b]), 0 ≤ a < b < ∞. If f is
m−convex on [a, b], for some fixed m ∈ (0, 1] and p, q > 0, then
a
(x − a)p(b − x)q f (x)dx
Z b
≤(b − a)p+q+1 min(cid:26)β(q + 2, p + 1)f (a) + mβ(q + 1, p + 2)f(cid:18) b
β(q + 1, p + 2)f (b) + mβ(q + 2, p + 1)f(cid:16) a
m(cid:17)o ,
m(cid:19) ,
(1.3)
where β(x, y) is the Euler Beta function.
Theorem 1.5. Let f : [a, b] → R be continuous on [a, b] such that f ∈ L([a, b]), 0 ≤ a < b < ∞. If f is
quasi-convex on [a, b], then for some fixed p, q > 0, we have
(1.4)
Z b
a
(x − a)p(b − x)q f (x)dx ≤ (b − a)p+q+1 max{f (a), f (b)}β(p + 1, q + 1).
The aim of this paper is to establish some new integral inequalities like those given in Theorems 1.4
and 1.5 for (α, m)−convex functions (Section 2) and quasi-convex functions (Section 3), respectively. Our
results in special cases recapture Theorems 1.4 and 1.5, respectively. That is, this study is a continuation
and generalization of [20].
NEW INTEGRAL INEQUALITIES VIA (α, m)-CONVEXITY AND QUASI-CONVEXITY
3
2. New integral inequalities for (α, m)− convex functions
Theorem 2.1. Let f : [a, b] → R be continuous on [a, b] such that f ∈ L([a, b]), 0 ≤ a < b < ∞. If f is
(α, m)−convex on [a, b], for some fixed (α, m) ∈ (0, 1]2 and p, q > 0, then
a
(x − a)p(b − x)q f (x)dx
Z b
≤(b − a)p+q+1 min(cid:26)β(q + α + 1, p + 1)f (a) + m[β(q + 1, p + 1) − β(q + α + 1, p + 1)]f(cid:18) b
β(q + 1, p + α + 1)f (b) + m[β(p + 1, q + 1) − β(q + 1, p + α + 1)]f(cid:16) a
m(cid:17)o ,
m(cid:19) ,
(2.1)
where β(x, y) is the Euler Beta function.
Proof. Since f is (α, m)−convex on [a, b], we know that for every t ∈ [0, 1]
(2.2)
f (ta + (1 − t)b) = f(cid:18)ta + m(1 − t)
b
m(cid:19) ≤ tαf (a) + m (1 − tα) f(cid:18) b
m(cid:19) .
Using Lemma 1.1, with x = ta + (1 − t)b, then we have
(x − a)p(b − x)q f (x)dx
a
Z b
≤(b − a)p+q+1Z 1
=(b − a)p+q+1(cid:20)f (a)Z 1
0
0
(1 − t)ptq(cid:18)tαf (a) + m (1 − tα) f(cid:18) b
m(cid:19)Z 1
(1 − t)ptq+αdt + mf(cid:18) b
m(cid:19)(cid:19) dt
(1 − t)ptq (1 − tα) dt(cid:21) .
0
Now, we will make use of the Beta function which is defined for x, y > 0 as
It is known that
Z 1
0
β(x, y) =Z 1
0
tx−1(1 − t)y−1dt.
tq+α(1 − t)pdt = β(q + α + 1, p + 1),
Z 1
0
(1 − t)ptq (1 − tα) dt =Z 1
0
tq(1 − t)pdt −Z 1
0
tq+α(1 − t)pdt
=β(q + 1, p + 1) − β(q + α + 1, p + 1)].
Combining all obtained equalities we get
a
(x − a)p(b − x)qf (x)dx
Z b
≤(b − a)p+q+1(cid:26)β(q + α + 1, p + 1)f (a) + m[β(q + 1, p + 1) − β(q + α + 1, p + 1)]f(cid:18) b
m(cid:19)(cid:27) .
(2.3)
If we choose x = tb + (1 − t)a, analogously we obtain
a
(x − a)p(b − x)q f (x)dx
Z b
≤(b − a)p+q+1nβ(q + 1, p + α + 1)f (b) + m[β(q + 1, p + 1) − β(q + 1, p + α + 1)]f(cid:16) a
m(cid:17)o .
Thus, by (2.3) and (2.4) we obtain (2.1), which completes the proof.
(cid:3)
(2.4)
4
W. J. LIU
Remark 1. As a special case of Theorem 2.1 for α = 1, that is for f be m−convex on [a, b], we recapture
Theorem 1.4 due to the fact that
β(q + 1, p + 1) − β(q + 2, p + 1) =β(q + 1, p + 1) −
q + 1
p + q + 2
β(q + 1, p + 1)
=
p + 1
p + q + 2
β(q + 1, p + 1) = β(q + 1, p + 2)
and
β(q + 1, p + 1) − β(q + 1, p + α + 1) = β(q + 2, p + 1).
Corollary 2.1. In Theorem 2.1, if p = q, then (2.1) reduces to
a
(x − a)p(b − x)pf (x)dx
Z b
≤(b − a)2p+1 min(cid:26)β(p + α + 1, p + 1)f (a) + m[β(p + 1, p + 1) − β(p + α + 1, p + 1)]f(cid:18) b
β(p + 1, p + α + 1)f (b) + m[β(p + 1, p + 1) − β(p + 1, p + α + 1)]f(cid:16) a
m(cid:17)o .
m(cid:19) ,
Theorem 2.2. Let f : [a, b] → R be continuous on [a, b] such that f ∈ L([a, b]), 0 ≤ a < b < ∞ and let
k > 1. If f
k−1 is (α, m)−convex on [a, b], for some fixed (α, m) ∈ (0, 1]2 and p, q > 0, then
k
(x − a)p(b − x)qf (x)dx
Z b
a
≤
(b − a)p+q+1
(α + 1)
k−1
k
[β(kp + 1, kq + 1)]
(2.5)
(cid:20)f (b)
k
k−1 + αm(cid:12)(cid:12)(cid:12)
k
k
k−1(cid:21)
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
k−1 =(cid:12)(cid:12)(cid:12)(cid:12)
k
f (ta + (1 − t)b)
k
k−1 + αm(cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
k−1
k
,
k
k−1#
k−1
1
k min
"f (a)
k
.
k
f(cid:18)ta + m(1 − t)
k
b
k−1
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
k−1 + m (1 − tα)(cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
≤tαf (a)
k
k−1
.
Proof. Since f
k−1 is (α, m)−convex on [a, b] we know that for every t ∈ [0, 1]
Using Lemma 1.1, with x = ta + (1 − t)b, then we have
(x − a)p(b − x)qf (x)dx
a
Z b
≤(b − a)p+q+1(cid:20)Z 1
0
(1 − t)kptkqdt(cid:21)
≤(b − a)p+q+1 [β(kq + 1, kp + 1)]
=(b − a)p+q+1 [β(kq + 1, kp + 1)]
f (ta + (1 − t)b)
1
1
0
k (cid:20)Z 1
k "Z 1
k " 1
0
1
tαf (a)
f (a)
α + 1
k−1
k
k
k
k−1 dt(cid:21)
k−1 dt + mZ 1
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(1 − tα)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
k−1#
α + 1(cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
k−1 + m
α
0
k
k
k
k−1
k−1
k
k
k−1
dt#
.
NEW INTEGRAL INEQUALITIES VIA (α, m)-CONVEXITY AND QUASI-CONVEXITY
5
If we choose x = tb + (1 − t)a, analogously we obtain
(x − a)p(b − x)qf (x)dx
Z b
a
≤(b − a)p+q+1 [β(kp + 1, kq + 1)]
which completes the proof.
1
k (cid:20) 1
α + 1
f (b)
k
k−1 + m
α
α + 1(cid:12)(cid:12)(cid:12)
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
k−1
k
,
k
k−1(cid:21)
(cid:3)
Corollary 2.2. In Theorem 2.2, if p = q, then (2.5) reduces to
(x − a)p(b − x)pf (x)dx
Z b
a
≤
(b − a)2p+1
(α + 1)
k−1
k
[β(kp + 1, kp + 1)]
(cid:20)f (b)
k
k−1 + αm(cid:12)(cid:12)(cid:12)
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
(x − a)p(b − x)qf (x)dx
Z b
a
≤
(b − a)p+q+1
k−1
k
2
[β(kp + 1, kq + 1)]
(cid:20)f (b)
k
k−1 + m(cid:12)(cid:12)(cid:12)
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
k
k−1(cid:21)
k−1
1
k min
k
.
"f (a)
k
k−1 + αm(cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
k−1
k
,
k
k−1#
1
k−1
k min
k
k−1(cid:21)
.
k
k
"f (a)
k
k−1 + m(cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
k−1
k
,
k
k−1#
Corollary 2.3. In Theorem 2.2, if α = 1, i.e., if f
k−1 is m−convex on [a, b], then (2.5) reduces to
Remark 2. As a special case of Corollary 2.3 for m = 1, that is for f
k
k−1 be convex on [a, b], we get
Z b
a
(x − a)p(b − x)qf (x)dx ≤
(b − a)p+q+1
k−1
k
2
[β(kp + 1, kq + 1)]
1
k hf (a)
k
k−1 + f (b)
k−1
k
.
k
k−1i
Theorem 2.3. Let f : [a, b] → R be continuous on [a, b] such that f ∈ L([a, b]), 0 ≤ a < b < ∞ and let
l ≥ 1. If f l is (α, m)−convex on [a, b], for some fixed (α, m) ∈ (0, 1]2 and p, q > 0, then
≤(b − a)p+q+1 [β(p + 1, q + 1)]
l
a
l−1
(x − a)p(b − x)qf (x)dx
Z b
× min
l#
"β(q + α + 1, p + 1)f (a)l + m[β(q + 1, p + 1) − β(q + α + 1, p + 1)](cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
l) .
(cid:20)β(q + 1, p + α + 1)f (b)l + m[β(q + 1, p + 1) − β(q + 1, p + α + 1)](cid:12)(cid:12)(cid:12)
l(cid:21)
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
≤ tαf (a)l + m (1 − tα)(cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f (ta + (1 − t)b)l =(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
f(cid:18)ta + m(1 − t)
b
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
.
l
l
1
1
l
,
Proof. Since f l is (α, m)−convex on [a, b], we know that for every t ∈ [0, 1]
(2.6)
6
W. J. LIU
Using Lemma 1.1, with x = ta + (1 − t)b, then we have
1
a
0
0
0
1
l
l−1
l−1
l−1
[(1 − t)ptq]
[(1 − t)ptq]
l f (ta + (1 − t)b)dt
l (cid:20)Z 1
(x − a)p(b − x)qf (x)dx
≤(b − a)p+q+1 [β(q + 1, p + 1)]
(1 − t)ptqdt(cid:21)
Z b
=(b − a)p+q+1Z 1
≤(b − a)p+q+1(cid:20)Z 1
(1 − t)ptqf (ta + (1 − t)b)ldt(cid:21)
×"β(q + α + 1, p + 1)f (a)l + m[β(q + 1, p + 1) − β(q + α + 1, p + 1)](cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
Z b
l(cid:21)
×(cid:20)β(q + 1, p + α + 1)f (b)l + m[β(q + 1, p + 1) − β(q + 1, p + α + 1)](cid:12)(cid:12)(cid:12)
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
≤(b − a)p+q+1 [β(p + 1, q + 1)]
(x − a)p(b − x)qf (x)dx
l−1
a
l
l
l
1
l
.
l#
If we choose x = tb + (1 − t)a, analogously we obtain
1
l
,
(cid:3)
which completes the proof.
Corollary 2.4. In Theorem 2.3, if p = q, then (2.6) reduces to
l
a
l−1
≤(b − a)2p+1 [β(p + 1, p + 1)]
(x − a)p(b − x)pf (x)dx
Z b
× min
"β(p + α + 1, p + 1)f (a)l + m[β(p + 1, p + 1) − β(p + α + 1, p + 1)](cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
l) .
(cid:20)β(p + 1, p + α + 1)f (b)l + m[β(p + 1, p + 1) − β(p + 1, p + α + 1)](cid:12)(cid:12)(cid:12)
l(cid:21)
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
Z b
(x − a)p(b − x)qf (x)dx
a
1
Corollary 2.5. In Theorem 2.3, if α = 1, i.e., if f l is m−convex on [a, b], then (2.6) reduces to
1
l
l#
,
l−1
≤(b − a)p+q+1 [β(p + 1, q + 1)]
l min
(cid:20)β(q + 1, p + 2)f (b)l + mβ(q + 2, p + 1)(cid:12)(cid:12)(cid:12)
"β(q + 2, p + 1)f (a)l + mβ(q + 1, p + 2)(cid:12)(cid:12)(cid:12)(cid:12)
m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:18) b
f(cid:16) a
m(cid:17)(cid:12)(cid:12)(cid:12)
Remark 3. As a special case of Corollary 2.5 for m = 1, that is for f l be convex on [a, b], we get
1
l
l#
,
l) .
l(cid:21)
1
(x − a)p(b − x)qf (x)dx
Z b
a
≤(b − a)p+q+1 [β(p + 1, q + 1)]
l−1
l hβ(q + 2, p + 1)f (a)l + β(q + 1, p + 2) f (b)li
1
l
.
NEW INTEGRAL INEQUALITIES VIA (α, m)-CONVEXITY AND QUASI-CONVEXITY
7
3. New integral inequalities for quasi-convex functions
Theorem 3.1. Let f : [a, b] → R be continuous on [a, b] such that f ∈ L([a, b]), 0 ≤ a < b < ∞ and let
k > 1. If f
k−1 is quasi-convex on [a, b], for some fixed p, q > 0, then
k
(3.1)
Z b
a
(x − a)p(b − x)q f (x)dx ≤ (b − a)p+q+1 [β(kp + 1, kq + 1)]
1
k (cid:16)maxnf (a)
k
k−1 , f (b)
k
k−1o(cid:17)
Proof. By Lemma 1.1, Holder's inequality, the definition of Beta function and the fact that f
quasi-convex on [a, b], we have
k−1
k
.
k
k−1 is
(x − a)p(b − x)q f (x)dx
a
Z b
≤(b − a)p+q+1(cid:20)Z 1
0
(1 − t)kptkqdt(cid:21)
≤(b − a)p+q+1 [β(kq + 1, kp + 1)]
=(b − a)p+q+1 [β(kq + 1, kp + 1)]
which completes the proof.
Corollary 3.1. Let f be as in Theorem 3.1. Additionally, if
(1) f is increasing, then we have
k−1
k
k
k−1 dt(cid:21)
k
k−1 , f (b)
k−1
k
k
k−1o dt(cid:21)
k−1
f (ta + (1 − t)b)
1
1
0
k (cid:20)Z 1
k (cid:20)Z 1
k hmaxnf (a)
0
1
maxnf (a)
k
k−1 , f (b)
k
k−1oi
k
,
(cid:3)
(x − a)p(b − x)qf (x)dx ≤ (b − a)p+q+1 [β(kp + 1, kq + 1)]
1
k f (b).
(2) f is decreasing, then we have
(x − a)p(b − x)qf (x)dx ≤ (b − a)p+q+1 [β(kp + 1, kq + 1)]
1
k f (a).
a
Z b
Z b
a
Theorem 3.2. Let f : [a, b] → R be continuous on [a, b] such that f ∈ L([a, b]), 0 ≤ a < b < ∞ and let
l ≥ 1. If f l is quasi-convex on [a, b], for some fixed p, q > 0, then
(3.2)
Z b
a
(x − a)p(b − x)qf (x)dx ≤ (b − a)p+q+1β(p + 1, q + 1)(cid:16)maxnf (a)l , f (b)lo(cid:17)
1
l
,
where β(x, y) is the Euler Beta function.
Proof. By Lemma 1.1, Holder's inequality, the definition of Beta function and the fact that f l is quasi-
convex on [a, b], we have
(x − a)p(b − x)qf (x)dx
a
Z b
=(b − a)p+q+1Z 1
≤(b − a)p+q+1(cid:20)Z 1
0
0
≤(b − a)p+q+1 [β(q + 1, p + 1)]
[(1 − t)ptq]
l−1
l
[(1 − t)ptq]
1
l f (ta + (1 − t)b)dt
(1 − t)ptqdt(cid:21)
l−1
1
l
0
l (cid:20)Z 1
(1 − t)ptqf (ta + (1 − t)b)ldt(cid:21)
l hmaxnf (a)l , f (b)lo β(q + 1, p + 1)i
l−1
=(b − a)p+q+1β(p + 1, q + 1)(cid:16)maxnf (a)l , f (b)lo(cid:17)
1
l
,
which completes the proof.
1
l
(cid:3)
8
W. J. LIU
Corollary 3.2. Let f be as in Theorem 3.2. Additionally, if
(1) f is increasing, then we have
(2) f is decreasing, then we have
a
Z b
Z b
a
(x − a)p(b − x)qf (x)dx ≤ (b − a)p+q+1β(p + 1, q + 1)f (b).
(x − a)p(b − x)q f (x)dx ≤ (b − a)p+q+1β(p + 1, q + 1)f (a).
References
[1] M. Alomari and M. Darus, On the Hadamard's inequality for log-convex functions on the coordinates, J. Inequal. Appl.
2009, Art. ID 283147 13 pp.
[2] M. Alomari, M. Darus and S.S. Dragomir, Inequalities of Hermite-Hadamard's type for functions whose derivatives
absolute values are quasi-convex, RGMIA Res. Rep. Coll., 12 (2009), Supp., No. 14.
[3] A. G. Azpeitia, Convex functions and the Hadamard inequality, Rev. Colombiana Mat. 28 (1994), no. 1, 7 -- 12.
[4] M. K. Bakula, M. E. Ozdemir and J. Pecari´c, Hadamard type inequalities for m-convex and (α, m)-convex functions,
JIPAM. J. Inequal. Pure Appl. Math. 9 (2008), no. 4, Article 96, 12 pp. (electronic).
[5] M. K. Bakula and J. Pecari´c, Note on some Hadamard-type inequalities, JIPAM. J. Inequal. Pure Appl. Math. 5 (2004),
no. 3, Article 74, 9 pp. (electronic).
[6] M. K. Bakula, J. Pecari´c and M. Ribici´c, Companion Inequalities to Jensen's Inequality for m-convex and (α, m)-convex
Functions, JIPAM. J. Inequal. Pure Appl. Math. 7 (2006), no. 5, Article 194, 15 pp. (electronic).
[7] C. Dinu, Hermite-Hadamard inequality on time scales, J. Inequal. Appl. 2008, Art. ID 287947, 24 pp.
[8] S.S. Dragomir and C.E.M. Pearce, Selected Topics on Hermite-Hadamard Inequalities and Applications, RGMIA Mono-
graphs, Victoria University, 2000.
[9] S. S. Dragomir and R. P. Agarwal, Two inequalities for differentiable mappings and applications to special means of
real numbers and to trapezoidal formula, Appl. Math. Lett. 11 (1998), no. 5, 91 -- 95.
[10] S. S. Dragomir, On some new inequalities of Hermite-Hadamard type for m-convex functions, Tamkang J. Math. 33
(2002), no. 1, 55 -- 65.
[11] S. S. Dragomir and S. Fitzpatrick, The Hadamard inequalities for s-convex functions in the second sense, Demonstratio
Math. 32 (1999), no. 4, 687 -- 696.
[12] P. M. Gill, C. E. M. Pearce and J. Pecari´c, Hadamard's inequality for r-convex functions, J. Math. Anal. Appl. 215
(1997), no. 2, 461 -- 470.
[13] V. N. Huy and N. T. Chung, Some generalizations of the Fej´er and Hermite-Hadamard inequalities in Holder spaces,
J. Appl. Math. Inform. 29 (2011), no. 3-4, 859 -- 868.
[14] D. A. Ion, Some estimates on the Hermite-Hadamard inequality through quasi-convex functions, An. Univ. Craiova
Ser. Mat. Inform. 34 (2007), 83 -- 88.
[15] U. S. Kirmaci et al., Hadamard-type inequalities for s-convex functions, Appl. Math. Comput. 193 (2007), no. 1, 26 -- 35.
[16] Z. Liu, Generalization and improvement of some Hadamard type inequalities for Lipschitzian mappings, J. Pure Appl.
Math. Adv. Appl. 1 (2009), no. 2, 175 -- 181.
[17] D. S. Mitrinovi´c, J. E. Pecari´c and A. M. Fink, Classical and new inequalities in analysis, Mathematics and its
Applications (East European Series), 61, Kluwer Acad. Publ., Dordrecht, 1993.
[18] V. G. Mihe¸san, A generalization of the convexity, Seminar on Functional Equations, Approx. and Convex., Cluj-
Napoca (Romania) (1993)
[19] M. E. Ozdemir, M. Avcı and E. Set, On some inequalities of Hermite-Hadamard type via m-convexity, Appl. Math.
Lett. 23 (2010), no. 9, 1065 -- 1070.
[20] M. E. Ozdemir, E. Set and M. Alomari, Integral inequalities via several kinds of convexity, Creat. Math. Inform. 20
(2011), no. 1, 62 -- 73.
[21] M. E. Ozdemir, E. Set and M. Z. Sarıkaya, Some new Hadamard type inequalities for co-ordinated m-convex and
(α, m)-convex functions, Hacet. J. Math. Stat. 40 (2011), no. 2, 219 -- 229.
[22] J. E. Pecari´c, F. Proschan and Y. L. Tong, Convex functions, partial orderings, and statistical applications, Mathe-
matics in Science and Engineering, 187, Academic Press, Boston, MA, 1992.
[23] M. Z. Sarikaya, E. Set and M. E. Ozdemir, On some new inequalities of Hadamard type involving h-convex functions,
Acta Math. Univ. Comenian. (N.S.) 79 (2010), no. 2, 265 -- 272.
[24] E. Set, M. E. Ozdemir and S. S. Dragomir, On the Hermite-Hadamard inequality and other integral inequalities
involving two functions, J. Inequal. Appl. 2010, Art. ID 148102, 9 pp.
[25] E. Set, M. E. Ozdemir and S. S. Dragomir, On Hadamard-type inequalities involving several kinds of convexity, J.
Inequal. Appl. 2010, Art. ID 286845, 12 pp.
[26] E. Set, M. Sardari, M. E. Ozdemir and J. Rooin, On generalizations of the Hadamard inequality for (α, m)-convex
functions, RGMIA Res. Rep. Coll., 12 (4) (2009), No. 4.
NEW INTEGRAL INEQUALITIES VIA (α, m)-CONVEXITY AND QUASI-CONVEXITY
9
[27] G. Toader, Some generalizations of the convexity, in Proceedings of the colloquium on approximation and optimization
(Cluj-Napoca, 1985), 329 -- 338, Univ. Cluj-Napoca, Cluj.
[28] K.-L. Tseng, S.-R. Hwang and S. S. Dragomir, New Hermite-Hadamard-type inequalities for convex functions (II),
Comput. Math. Appl. 62 (2011), no. 1, 401 -- 418.
(W. J. Liu) College of Mathematics and Statistics, Nanjing University of Information Science and Technol-
ogy, Nanjing 210044, China
E-mail address: [email protected]
|
1312.5587 | 1 | 1312 | 2013-12-19T15:29:02 | Commutators of vector-valued intrinsic square functions on vector-valued generalized weighted Morrey spaces | [
"math.FA"
] | In this paper, we will obtain the strong type and weak type estimates for vector-valued analogues of intrinsic square functions in the generalized weighted Morrey spaces $M^{\Phi,\varphi}_{w}(\mathbb{R}^n)$. We study the boundedness of intrinsic square functions including the Lusin area integral, Littlewood-Paley $\mathrm{g}$-function and $\mathrm{g}_{\lambda}^{*}$ -function and their commutators on vector-valued generalized weighted Morrey spaces $M^{\Phi,\varphi}_{w}(l_2)$. In all the cases the conditions for the boundedness are given either in terms of Zygmund-type integral inequalities on $\varphi(x,r)$ without assuming any monotonicity property of $\varphi(x,r)$ on $r$. | math.FA | math |
Commutators of vector-valued intrinsic square
functions on vector-valued generalized weighted
Morrey spaces
Vagif S. Guliyev, M.N. Omarova
Abstract
In this paper, we will obtain the strong type and weak type estimates
for vector-valued analogues of intrinsic square functions in the generalized
weighted Morrey spaces M Φ,ϕ
w (Rn). We study the boundedness of intrin-
sic square functions including the Lusin area integral, Littlewood-Paley
g-function and g∗
λ -function and their kth-order commutators on vector-
valued generalized weighted Morrey spaces M Φ,ϕ
w (l2). In all the cases the
conditions for the boundedness are given either in terms of Zygmund-type
integral inequalities on ϕ(x, r) without assuming any monotonicity prop-
erty of ϕ(x, r) on r.
AMS Mathematics Subject Classification:
Key words: Intrinsic square functions; Vector-valued generalized weighted Mor-
rey spaces; vector-valued inequalities; Ap weights; Commutators; BMO
42B25, 42B35
1
Introduction
It is well-known that the commutator is an important integral operator and it
plays a key role in harmonic analysis. In 1965, Calderon [2, 3] studied a kind
of commutators, appearing in Cauchy integral problems of Lip-line. Let K be a
Calder´on-Zygmund singular integral operator and b ∈ BM O(Rn). A well known
result of Coifman, Rochberg and Weiss [9] states that the commutator operator
[b, K]f = K(bf )−b Kf is bounded on Lp(Rn) for 1 < p < ∞. The commutator of
Calder´on-Zygmund operators plays an important role in studying the regularity of
solutions of elliptic partial differential equations of second order (see, for example,
[6]-[8], [5], [10], [11]).
The classical Morrey spaces were originally introduced by Morrey in [32] to
study the local behavior of solutions to second order elliptic partial differential
equations. For the properties and applications of classical Morrey spaces, we
refer the readers to [10, 11, 18, 32]. Recently, Komori and Shirai [29] first de-
fined the weighted Morrey spaces Lp,κ(w) and studied the boundedness of some
1
classical operators such as the Hardy-Littlewood maximal operator, the Calder´on-
Zygmund operator on these spaces. Also, Guliyev [21, 22] introduced the general-
ized weighted Morrey spaces M p,ϕ
w and studied the boundedness of the sublinear
operators and their higher order commutators generated by Calder´on-Zygmund
operators and Riesz potentials in these spaces (see, also [25, 27, 28, 35]).
The intrinsic square functions were first introduced by Wilson in [40, 41].
They are defined as follows. For 0 < α ≤ 1, let Cα be the family of functions
φ : Rn → R such that φ's support is contained in {x : x ≤ 1}, RRn φ(x)dx = 0,
and for x, x′ ∈ Rn,
φ(x) − φ(x′) ≤ x − x′α.
For (y, t) ∈ Rn+1
+
and f ∈ L1,loc(Rn) , set
Aαf (t, y) ≡ sup
φ∈Cα
f ∗ φt(y),
where φt(y) = t−nφ(
(intrinsic Lusin) function of f by the formula
y
t
) . Then we define the varying-aperture intrinsic square
Gα,β(f )(x) = Z ZΓβ (x)
tn+1!
(Aαf (t, y))2 dydt
1
2
,
where Γβ(x) = {(y, t) ∈ Rn+1
+
: x − y < βt}. Denote Gα,1(f ) = Gα(f ) .
This function is independent of any particular kernel, such as Poisson kernel.
It dominates pointwise the classical square function(Lusin area integral) and its
real-variable generalizations. Although the function Gα,β(f ) is depend of kernels
with uniform compact support, there is pointwise relation between Gα,β(f ) with
different β:
We can see details in [40].
Gα,β(f )(x) ≤ β
3n
2 +αGα(f )(x) .
The intrinsic Littlewood-Paley g-function and the intrinsic g∗
λ function are
defined respectively by
1
2
,
gαf (x) =(cid:18)Z ∞
0
t (cid:19)
(Aαf (y, t))2 dt
tn+1!
(Aαf (y, t))2 dydt
When we say that f maps into l2, we mean that ~f (x) = (cid:0)fj(cid:1)∞
λ,αf (x) = Z ZRn+1
+ (cid:18)
t + x − y(cid:19)nλ
fj is Lebesgue measurable and, for almost every x ∈ Rn
g∗
t
1
2
.
j=1, where each
k ~f (x)kl2 =(cid:18) ∞
Xj=1
fj(x)2(cid:19)1/2
.
2
Let ~f = (f1, f2, . . .) be a sequence of locally integrable functions on Rn. For
any x ∈ Rn, Wilson [41] also defined the vector-valued intrinsic square functions
of ~f by kGα
~f (x)kl2 and proved the following result.
Theorem A. Let 1 ≤ p < ∞, 0 < α ≤ 1 and w ∈ Ap. Then the operators
w(l2) to
w(l2) into itself for p > 1 and from L1
λ,α are bounded from Lp
Gα and g∗
W L1
w(l2).
Moreover, in [31], Lerner showed sharp Lp
w norm inequalities for the intrinsic
square functions in terms of the Ap characteristic constant of w for all 1 < p < ∞.
Also Huang and Liu [12] studied the boundedness of intrinsic square functions on
weighted Hardy spaces. Moreover, they characterized the weighted Hardy spaces
by intrinsic square functions. In [38] and [39], Wang and Liu obtained some weak
type estimates on weighted Hardy spaces. In [37], Wang considered intrinsic func-
tions and the commutators generated with BMO functions on weighted Morrey
spaces. Let b be a locally integrable function on Rn. Setting
Ak
α,bf (t, y) ≡ sup
ZRn
φ∈Cα(cid:12)(cid:12)(cid:12)(cid:12)
the kth-order commutators are defined by
,
[b(x) − b(z)]kφt(y − z)f (z)dz(cid:12)(cid:12)(cid:12)(cid:12)
tn+1(cid:19)
α,bf (t, y))2 dydt
t (cid:19)
α,bf (t, y))2 dt
(Ak
(Ak
1
2
1
2
,
[b, Gα]kf (x) =(cid:18)Z ZΓ(x)
[b, gα]kf (x) =(cid:18)Z ∞
0
and
[b, g∗
λ,α]kf (x) = Z ZRn+1
+ (cid:18)
t
t + x − y(cid:19)λn
(Ak
tn+1!
α,bf (t, y))2 dydt
1
2
.
A function b ∈ Lloc
1 (Rn) is said to be in BM O(Rn) if
kbk∗ = sup
x∈Rn, r>0
1
B(x, r)ZB(x,r)
b(y) − bB(x,r)dy < ∞,
where bB(x,r) = 1
B(x,r)RB(x,r) b(y)dy.
By the similar argument as in [14] and [37], we can get
Theorem B. Let 1 < p < ∞, 0 < α ≤ 1, w ∈ Ap and b ∈ BM O(Rn).
λ,α]k are bounded from
Then the kth-order commutator operators [b, Gα]k and [b, g∗
Lp
w(l2) into itself.
In this paper, we will consider the boundedness of the operators Gα, gα, g∗
λ,α
and their kth-order commutators on vector-valued generalized weighted Morrey
spaces. Let ϕ(x, r) be a positive measurable function on Rn × R+ and w be
3
n
p
ess inf
t<s<∞
ϕ1(x, s)s
n
p +1
t
Z ∞
r
dt ≤ C ϕ2(x, r).
(1.2)
non-negative measurable function on Rn. For any ~f ∈ Lp,loc
M p,ϕ
w (l2) the vector-valued generalized weighted Morrey spaces, if
w (l2) , we denote by
k ~f kM p,ϕ
w (l2) = sup
x∈Rn, r>0
ϕ(x, r)−1 w(B(x, r))− 1
p kk ~f (·)kl2kLp
w(B(x,r)) < ∞.
When w ≡ 1, then M p,ϕ
w (l2) coincide the vector-valued generalized Morrey spaces
M p,ϕ(l2). There are many papers discussed the conditions on ϕ(x, r) to obtain
the boundedness of operators on the generalized Morrey spaces. For example, in
[17] (see, also [18]), by Guliyev the following condition was imposed on the pair
(ϕ1, ϕ2) :
Z ∞
r
ϕ1(x, t)
dt
t
≤ Cϕ2(x, r).
(1.1)
where C > 0 does not depend on x and r. Under the above condition, they
obtained the boundedness of Calder´on-Zygmund singular integral operators from
M p,ϕ1(Rn) to M p,ϕ2(Rn). Also, in [1] and [20], Guliyev et. introduced a weaker
condition: If 1 ≤ p < ∞, there exits a constant C > 0, such that, for any x ∈ Rn
and r > 0,
If the pair (ϕ1, ϕ2) satisfies condition (1.1), then (ϕ1, ϕ2) satisfied condition (1.2).
But the opposite is not true. We can see remark 4.7 in [20] for details.
Recently, in [21, 22] (see, also [25, 28, 35]), Guliyev introduced a weighted
condition: If 1 ≤ p < ∞, there exits a constant C > 0, such that, for any x ∈ Rn
and t > 0,
ess inf
t<s<∞
ϕ1(x, s)w(B(x, s))
w(B(x, t))
1
p
Z ∞
r
1
p
dt
t
≤ C ϕ2(x, r),
(1.3)
In this paper, we will obtain the boundedness of the vector-valued intrin-
sic function, the intrinsic Littlewood-Paley g function, the intrinsic g∗
λ function
and their kth-order commutators on vector-valued generalized weighted Morrey
spaces when w ∈ Ap and the pair (ϕ1, ϕ2) satisfies condition (1.3) or the following
inequalities,
Z ∞
r
lnk(cid:16)e +
t
r(cid:17)
ess inf
t<s<∞
ϕ1(x, s)w(B(x, s))
w(B(x, t))
1
p
1
p
dt
t
≤ C ϕ2(x, r),
(1.4)
where C does not depend on x and r. Our main results in this paper are stated
as follows.
Theorem 1.1. Let 1 ≤ p < ∞, 0 < α ≤ 1, w ∈ Ap and (ϕ1, ϕ2) satisfies
condition (1.3). Then the operator Gα is bounded from M p,ϕ1
w (l2) for
p > 1 and from M 1,ϕ1
w (l2) to M p,ϕ2
w (l2) to W M 1,ϕ2
w (l2).
4
Theorem 1.2. Let 1 ≤ p < ∞, 0 < α ≤ 1, w ∈ Ap, λ > 3 +
satisfies condition (1.3). Then the operator g∗
M p,ϕ2
w (l2) for p > 1 and from M 1,ϕ1
λ,α is bounded from M p,ϕ1
w (l2).
w (l2) to W M 1,ϕ2
and (ϕ1, ϕ2)
w (l2) to
α
n
Theorem 1.3. Let 1 < p < ∞, 0 < α ≤ 1, w ∈ Ap, b ∈ BM O and (ϕ1, ϕ2)
satisfies condition (1.4). Then [b, Gα]k is bounded from M p,ϕ1
w (l2) .
w (l2) to M p,ϕ2
Theorem 1.4. Let 1 < p < ∞, 0 < α ≤ 1, w ∈ Ap, b ∈ BM O and (ϕ1, ϕ2)
satisfies condition (1.4), then for λ > 3 + α
w (l2)
to M p,ϕ2
λ,α]k is bounded from M p,ϕ1
n , [b, g∗
w (l2).
In [40], the author proved that the functions Gαf and gαf are pointwise
comparable. Thus, as a consequence of Theorem 1.1 and Theorem 1.3, we have
the following results.
Corollary 1.5. Let 1 ≤ p < ∞, 0 < α ≤ 1, w ∈ Ap and (ϕ1, ϕ2) satisfies
condition (1.3), then gα is bounded from M p,ϕ1
w (l2) for p > 1 and
from M 1,ϕ1
w (l2) to M p,ϕ2
w (l2) to W M 1,ϕ2
w (l2).
Corollary 1.6. Let 1 < p < ∞, 0 < α ≤ 1, w ∈ Ap, b ∈ BM O and (ϕ1, ϕ2)
satisfies condition (1.4), then [b, gα]k is bounded from M p,ϕ1
w (l2) to M p,ϕ2
w (l2).
Remark 1.7. Note that, in the scalar valued case the Theorems 1.1 - 1.4 and
Corollaries 1.5 - 1.6 was proved in [26] (w ≡ 1) and [27]. Also, in the scalar
p , 0 < κ < 1
valued case and w ≡ Ap and ϕ1(x, r) = ϕ2(x, r) ≡ w(B(x, r))
Theorems 1.1-1.4 and Corollaries 1.5-1.6 was proved by Wang in [37, 36]. How
p , then the vector-valued generalized weighed Morrey
as, if ϕ(x, r) ≡ w(B(x, r))
space M p,ϕ
w (l2) and the
pair (w(B(x, r))
p ) satisfies the both conditions (1.3) and (1.4).
Indeed, by Lemma 3.1 there exists C > 0 and δ > 0 such that for all x ∈ Rn and
t > r:
w (l2) coincide the vector-valued weighed Morrey space Lp,κ
p , w(B(x, r))
κ−1
κ−1
κ−1
κ−1
Then
w(B(x, t)) ≥ C(cid:16) t
r(cid:17)nδ
w(B(x, r)).
κ
p
dt
t
t
ess inf
t<s<∞
w(B(x, s))
r(cid:17)
r(cid:17) w(B(x, t))
w(B(x, t))1/p
dt
t
κ−1
t
p
ess inf
t<s<∞
w(B(x, s))
κ
p
w(B(x, t))1/p
dt
t
Z ∞
r
r
≤Z ∞
=Z ∞
r
lnk(cid:16)e +
lnk(cid:16)e +
5
.Z ∞
r
lnk(cid:16)e +
κ−1
= w(B(x, r))
= w(B(x, r))
≈ w(B(x, r))
κ−1
κ−1
p .
κ−1
p dt
t
p dt
t
t
r(cid:17)(cid:16)(cid:16) t
r(cid:17)nδ
w(B(x, r))(cid:17)
p Z ∞
r(cid:17)nδ κ−1
r(cid:17)(cid:16) t
lnk(cid:16)e +
p Z ∞
lnk(cid:16)e + τ(cid:17) τ nδ κ−1
dτ
τ
t
1
r
p
Throughout this paper, we use the notation A . B to mean that there is
a positive constant C independent of all essential variables such that A ≤ CB.
Moreover, C may be different from place to place.
2 Vector-valued generalized weighted Morrey spaces
The classical Morrey spaces M p,λ were originally introduced by Morrey in [32] to
study the local behavior of solutions to second order elliptic partial differential
equations. For the properties and applications of classical Morrey spaces, we refer
the readers to [15, 30].
We denote by M p,λ(l2) ≡ M p,λ(Rn, l2) the vector-valued Morrey space, the
space of all vector-valued functions ~f ∈ Lp,loc(l2) with finite quasinorm
= sup
x∈Rn, r>0
r− λ
p k ~f kLp(B(x,r),l2),
(cid:13)(cid:13)(cid:13)
~f(cid:13)(cid:13)(cid:13)M p,λ(l2)
where 1 ≤ p < ∞ and 0 ≤ λ ≤ n.
Note that M p,0(l2) = Lp(l2) and M p,n(l2) = L∞(l2). If λ < 0 or λ > n, then
M p,λ(l2) = Θ, where Θ is the set of all vector-valued functions equivalent to 0 on
Rn.
We define the vector-valued generalized weighed Morrey spaces as follows.
Definition 2.1. Let 1 ≤ p < ∞, ϕ be a positive measurable vector-valued
function on Rn × (0, ∞) and w be non-negative measurable function on Rn. We
denote by M p,ϕ
w (l2) the vector-valued generalized weighted Morrey space, the
space of all vector-valued functions ~f ∈ Lp,loc
w (l2) with finite norm
k ~f kM p,ϕ
w (l2) = sup
x∈Rn,r>0
ϕ(x, r)−1 w(B(x, r))− 1
p kf kLp
w(B(x,r),l2),
where Lp
functions f for which
w(B(x, r), l2) denotes the vector-valued weighted Lp-space of measurable
k ~f kLp
w(B(x,r)) ≡ k ~f χB(x,r)kLp
w(Rn) =(cid:18)ZB(x,r)
k ~f (y)kp
l2
1
p
.
w(y)dy(cid:19)
6
Furthermore, by W M p,ϕ
w (l2) we denote the vector-valued weak generalized
weighted Morrey space of all functions f ∈ W Lp,loc
w (l2) for which
k ~f kW M p,ϕ
w (l2) = sup
x∈Rn,r>0
ϕ(x, r)−1 w(B(x, r))− 1
p k ~f kW Lp
w(B(x,r),l2) < ∞,
where W Lp
which
w(B(x, r), l2) denotes the weak Lp
w-space of measurable functions f for
k ~f kW Lp
w(B(x,r),l2) ≡ k ~f χB(x,r)kW Lp
w(l2) = sup
t>0
t Z{y∈B(x,r): k ~f (y)kl2 >t}
w(y)dy!
1
p
.
If w ≡ 1, then M p,ϕ
1
(l2) = M p,ϕ(l2) is the vector-valued
Remark 2.2. (1)
generalized Morrey space.
(2) If ϕ(x, r) ≡ w(B(x, r))
weighted Morrey space.
κ−1
p , then M p,ϕ
w (l2) = Lp,κ
w (l2) is the vector-valued
(3)
If ϕ(x, r) ≡ v(B(x, r))
κ
p w(B(x, r))− 1
p , then M p,ϕ
w (l2) = Lp,κ
v,w(l2) is the
vector-valued two weighted Morrey space.
(4) If w ≡ 1 and ϕ(x, r) = r
the vector-valued Morrey space and W M p,ϕ
weak Morrey space.
λ−n
p with 0 < λ < n, then M p,ϕ
w (l2) = Lp,λ(l2) is
w (l2) = W Lp,λ(l2) is the vector-valued
(5)
If ϕ(x, r) ≡ w(B(x, r))− 1
p , then M p,ϕ
w (l2) = Lp
w(l2) is the vector-valued
weighted Lebesgue space.
3 Preliminaries and some lemmas
By a weight function, briefly weight, we mean a locally integrable function on Rn
which takes values in (0, ∞) almost everywhere. For a weight w and a measurable
set E, we define w(E) = RE w(x)dx, and denote the Lebesgue measure of E by
E and the characteristic function of E by χE . Given a weight w, we say that
w satisfies the doubling condition if there exists a constant D > 0 such that for
any ball B, we have w(2B) ≤ Dw(B). When w satisfies this condition, we write
brevity w ∈ ∆2.
If w is a weight function, we denote by Lp
w(l2) ≡ Lp
w(Rn, l2) the vector-valued
weighted Lebesgue space defined by finiteness of the norm
k ~f kLp
w(l2) =(cid:18)ZRn
k ~f (x)kp
l2
1
p
w(x)dx(cid:19)
< ∞,
if 1 ≤ p < ∞
and by k ~f kL∞
w (l2) = ess sup
x∈Rn
k ~f (x)kl2w(x) if p = ∞.
7
We recall that a weight function w is in the Muckenhoupt's class Ap [33],
1 < p < ∞, if
[w]Ap : = sup
B
[w]Ap(B)
= sup
B (cid:18) 1
BZB
w(x)dx(cid:19)(cid:18) 1
BZB
w(x)1−p′
dx(cid:19)p−1
< ∞,
where the sup is taken with respect to all the balls B and 1
for all balls B by Holder's inequality
p + 1
p′ = 1. Note that,
[w]1/p
Ap(B) = B−1kwk1/p
L1(B) kw−1/pkLp′ (B) ≥ 1.
M w(x)
For p = 1, the class A1 is defined by the condition M w(x) ≤ Cw(x) with [w]A1 =
sup
x∈Rn
w(x) , and for p = ∞ A∞ =S1≤p<∞ Ap and [w]A∞ = inf
Lemma 3.1. ([16]) (1) If w ∈ Ap for some 1 ≤ p < ∞, then w ∈ ∆2. Moreover,
for all λ > 1
[w]Ap.
1≤p<∞
w(λB) ≤ λnp[w]Apw(B).
(2) If w ∈ A∞, then w ∈ ∆2. Moreover, for all λ > 1
w(λB) ≤ 2λn
[w]A∞w(B).
(3)
If w ∈ Ap for some 1 ≤ p ≤ ∞, then there exit C > 0 and δ > 0 such
that for any ball B and a measurable set S ⊂ B,
We are going to use the following result on the boundedness of the Hardy
operator
(Hg)(t) :=
g(r)dµ(r), 0 < t < ∞,
≤ C(cid:16) S
B(cid:17)δ
.
w(S)
w(B)
1
t Z t
0
where µ is a non-negative Borel measure on (0, ∞).
Theorem 3.2. ([4]) The inequality
ess sup
ω(t)Hg(t) ≤ c ess sup
v(t)g(t)
t>0
t>0
holds for all functions g non-negative and non-increasing on (0, ∞) if and only if
A := sup
t>0
and c ≈ A.
ω(t)
t Z t
0
dµ(r)
ess sup
0<s<r
v(s)
< ∞,
8
We also need the following statement on the boundedness of the Hardy type
operator
(H1g)(t) :=
1
t Z t
0
where µ is a non-negative Borel measure on (0, ∞).
t
lnk(cid:16)e +
r(cid:17) g(r)dµ(r), 0 < t < ∞,
Theorem 3.3. The inequality
ess sup
t>0
ω(t)H1g(t) ≤ c ess sup
t>0
v(t)g(t)
holds for all functions g non-negative and non-increasing on (0, ∞) if and only if
A1 := sup
t>0
and c ≈ A1.
ω(t)
t Z t
0
lnk(cid:16)e +
t
r(cid:17)
dµ(r)
ess sup
0<s<r
v(s)
< ∞,
Note that, Theorem 3.3 can be proved analogously to Theorem 4.3 in [19].
Definition 3.4. BM O(Rn) is the Banach space modulo constants with the norm
k · k∗ defined by
kbk∗ = sup
x∈Rn,r>0
1
B(x, r)ZB(x,r)
b(y) − bB(x,r)dy < ∞,
where b ∈ Lloc
1 (Rn) and
bB(x,r) =
1
B(x, r)ZB(x,r)
b(y)dy.
Lemma 3.5. ([34], Theorem 5, p. 236) Let w ∈ A∞. Then the norm k · k∗ is
equivalent to the norm
kbk∗,w = sup
x∈Rn,r>0
where
1
w(B(x, r))ZB(x,r)
w(B(x, r))ZB(x,r)
1
b(y) − bB(x,r),ww(y)dy,
bB(x,r),w =
b(y)w(y)dy.
Remark 3.6. (1) The John-Nirenberg inequality : there are constants C1, C2 >
0, such that for all b ∈ BM O(Rn) and β > 0
{x ∈ B : b(x) − bB > β} ≤ C1Be−C2β/kbk∗,
∀B ⊂ Rn.
9
(2) For 1 ≤ p < ∞ the John-Nirenberg inequality implies that
kbk∗ ≈ sup
B (cid:18) 1
BZB
1
p
b(y) − bBpdy(cid:19)
and for 1 ≤ p < ∞ and w ∈ A∞
kbk∗ ≈ sup
B (cid:18) 1
w(B)ZB
b(y) − bBpw(y)dy(cid:19)
1
p
.
(3.1)
(3.2)
Note that, by the John-Nirenberg inequality and Lemma 3.1 (part 3) it follows
that
w({x ∈ B : b(x) − bB > β}) ≤ C δ
1 w(B)e−C2βδ/kbk∗
for some δ > 0. Hence
ZB
b(y) − bBpw(y)dy = pZ ∞
0
βp−1 w({x ∈ B : b(x) − bB > β})dβ
≤ pC δ
1 w(B) Z ∞
0
βp−1 e−C2βδ/kbk∗ dβ = C3w(B)kbkp
∗,
where C3 > 0 depends only on C δ
1, C2, p, and δ, which implies (3.2).
Also (3.1) is a particular case of (3.2) with w ≡ 1.
The following lemma was proved in [22].
Lemma 3.7. i) Let w ∈ A∞ and b ∈ BM O(Rn). Let also 1 ≤ p < ∞, x ∈ Rn,
k > 0 and r1, r2 > 0. Then
1
w(B(x, r1)) ZB(x,r1)
(cid:16)
b(y) − bB(x,r2),wkpw(y)dy(cid:17)
1
p ≤ C (cid:18)1 +(cid:12)(cid:12)(cid:12)
kbkk
∗,
ln
r1
r2(cid:12)(cid:12)(cid:12)(cid:19)k
where C > 0 is independent of f , w, x, r1, and r2.
ii) Let w ∈ Ap and b ∈ BM O(Rn). Let also 1 < p < ∞, x ∈ Rn, k > 0 and
r1, r2 > 0. Then
1
w1−p′(B(x, r1)) ZB(x,r1)
(cid:16)
b(y)−bB(x,r2),wkp′
w(y)1−p′
1
p′
dy(cid:17)
where C > 0 is independent of f , w, x, r1, and r2.
ln
≤ C (cid:18)1 +(cid:12)(cid:12)(cid:12)
r1
r2(cid:12)(cid:12)(cid:12)(cid:19)k
kbkk
∗,
10
4 Proofs of main theorems
Before proving the main theorems, we need the following lemmas.
Lemma 4.1. [37] For j ∈ Z+, denote
Gα,2j (f )(x) =(cid:18)Z ∞
0 Zx−y≤2jt
tn+1(cid:19)
(Aαf (y, t))2 dydt
1
2
Let 0 < α ≤ 1, 1 < p < ∞ and w ∈ Ap. Then any j ∈ Z+, we have
kGα,2j (f )kLp
w . 2j(cid:0) 3n
2 +α(cid:1) kGα(f )kLp
w.
This lemma is easy from the following inequality which is proved in [40].
Gα,β(f )(x) ≤ β
3n
2 +αGα(f )(x).
By the similar argument as in [3], we can get the following lemma.
Lemma 4.2. Let 1 < p < ∞, 0 < α ≤ 1 and w ∈ Ap, then the commutators
[b, Gα]k is bounded from Lp
w(l2) to itself whenever b ∈ BM O.
Now we are in a position to prove theorems.
Lemma 4.3. Let 1 ≤ p < ∞, 0 < α ≤ 1 and w ∈ Ap.
Then, for p > 1 the inequality
kGα
~f kLp
w(B,l2) .(cid:0)w(B)(cid:1)
1
p Z ∞
2r
k ~f k
Lp
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
p
dt
t
holds for any ball B = B(x0, r) and for all ~f ∈ Lp,loc
w (l2).
Moreover, for p = 1 the inequality
kGα
~f kW L1
w(B,l2) . w(B)Z ∞
2r
k ~f k
L1
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)−1 dt
t
,
holds for any ball B = B(x0, r) and for all ~f ∈ L1locl2.
Proof. The main ideas of these proofs come from [22]. For arbitrary x ∈ Rn,
set B = B(x0, r), 2B ≡ B(x0, 2r). We decompose ~f = ~f0 + ~f∞, where ~f0(y) =
~f (y)χ2B(y), ~f∞(y) = ~f (y) − ~f0(y). Then,
kGα
~f k
Lp
w(cid:0)B(x0,r),l2(cid:1) ≤ kGα
Lp
w(cid:0)B(x0,r),l2(cid:1) + kGα
~f0k
First, let us estimate I. By Theorem A, we can obtain that
~f∞k
Lp(cid:0)B(x0,r),l2(cid:1) := I + II.
I ≤ kGα
~f0kLp
w(l2) . k ~f0kLp
w(l2) = k ~f kLp
w(2B,l2).
(4.1)
11
On the other hand,
k ~f kLp
w(2B,l2)
≈ Bk ~f kLp
w(2B,l2)Z ∞
2r
dt
tn+1
dt
tn+1
dt
t
.
p
≤ BZ ∞
2r
k ~f k
Lp
. w(B)
. w(B)
1
. w(B)
dt
tn+1
dt
tn+1
2r
Lp
2r
2r
Lp
Lp
1
1
k ~f k
k ~f k
k ~f k
w(cid:0)B(x0,t),l2(cid:1)
p kw−1/pkLp′ (B) Z ∞
w(cid:0)B(x0,t),l2(cid:1)
p Z ∞
w(cid:0)B(x0,t),l2(cid:1) kw−1/pkLp′ (B(x0,t))
p Z ∞
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
) ~f∞(z)dz(cid:13)(cid:13)(cid:13)(cid:13)l2
≤ t−nZy−z≤t
y − z
dt
t
k ~f k
φ(
Lp
p
.
t
k ~f∞(z)kl2dz.
(4.2)
(4.3)
Therefore from (4.1) and (4.2) we get
I . w(B)
1
p Z ∞
2r
Then let us estimate II.
k ~f ∗ φt(y)kl2 =(cid:13)(cid:13)(cid:13)(cid:13)
t−nZy−z≤t
Since x ∈ B(x0, r), (y, t) ∈ Γ(x), we have z − x ≤ z − y + y − x ≤ 2t, and
r ≤ z − x0 − x0 − x ≤ x − z ≤ x − y + y − z ≤ 2t.
So, we obtain
(cid:13)(cid:13)Gα
~f∞(x)(cid:13)(cid:13)l2
≤ Z ZΓ(x)(cid:18)t−nZy−z≤t
≤ Zt>r/2Zx−y<t(cid:18)Zx−z≤2t
. Zt>r/2(cid:18)Zz−x≤2t
1
2
tn+1!
k ~f∞(z)kl2dz(cid:19)2 dydt
t3n+1!
k ~f∞(z)kl2dz(cid:19)2 dydt
t2n+1!
k ~f∞(z)kl2dz(cid:19)2 dt
1
2
.
By Minkowski and Holder's inequalities and z−x ≥ z−x0−x0−x ≥
12
1
2
1
2
z−x0,
we have
(cid:13)(cid:13)Gα
~f∞(x)(cid:13)(cid:13)l2
1
2
dt
t2n+1!
k ~f∞(z)kl2dz
k ~f (z)kl2
z − xn
dz .Zz−x0>2r
k ~f (z)kl2
z − x0n
dz
dt
tn+1
dz
k ~f (z)kl2Z +∞
Z2r<z−x0<t
z−x0
k ~f (z)kl2dz
dt
tn+1
2
.ZRn Zt> z−x
.Zz−x0>2r
=Zz−x0>2r
=Z +∞
.Z ∞
.Z ∞
k ~f k
Lp
2r
2r
2r
kk ~f (z)kl2kLp
w(B(x0,t)) kw−1kLp′ (B(x0,t))
dt
tn+1
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
p
dt
t
.
(4.4)
Thus,
kGα
~f∞kLp
w(B,l2) . w(B)
1
p Z ∞
2r
k ~f k
Lp
By combining (4.3) and (4.5), we have
kGα
~f kLp
w(B,l2) . w(B)
1
p Z ∞
2r
k ~f k
.
(4.5)
dt
t
p
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
Lp
p
dt
t
.
Proof of Theorem 1.1
By Lemma 4.3 and Theorem 3.2 we have for p > 1
kGα
~f kM
p,ϕ2
w
(l2) . sup
x0∈Rn,r>0
= sup
x0∈Rn,r>0
= sup
x0∈Rn,r>0
. sup
x0∈Rn,r>0
= sup
x0∈Rn,r>0
r
k ~f k
ϕ2(x0, r)−1 Z ∞
ϕ2(x0, r)−1Z r−1
r Z r
ϕ2(x0, r−1)−1 r
1
0
0
k ~f k
Lp
Lp
dt
t
p
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
w(cid:0)B(x0,t−1),l2(cid:1)(cid:0)w(B(x0, t−1))(cid:1)− 1
w(cid:0)B(x0,t−1),l2(cid:1)(cid:0)w(B(x0, t−1))(cid:1)− 1
dt
t
Lp
p k ~f k
p
p
k ~f k
dt
t
ϕ1(x0, r−1)−1(cid:0)w(B(x0, r−1))(cid:1)− 1
ϕ1(x0, r)−1(cid:0)w(B(x0, r))(cid:1)− 1
p k ~f k
Lp
w(cid:0)B(x0,r−1),l2(cid:1)
w(cid:0)B(x0,r),l2(cid:1) = k ~f kM
Lp
p,ϕ1
w
(l2)
13
and for p = 1
kGα
~f kW M
1,ϕ2
w
(l2) . sup
x0∈Rn,r>0
= sup
x0∈Rn,r>0
= sup
x0∈Rn,r>0
. sup
x0∈Rn,r>0
= sup
x0∈Rn,r>0
L1
r
k ~f k
ϕ2(x0, r)−1 Z ∞
ϕ2(x0, r)−1Z r−1
r Z r
ϕ2(x0, r−1)−1 r
1
0
0
k ~f k
t
L1
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)−1 dt
w(cid:0)B(x0,t−1),l2(cid:1)(cid:0)w(B(x0, t−1))(cid:1)−1 dt
w(cid:0)B(x0,t−1),l2(cid:1)(cid:0)w(B(x0, t−1))(cid:1)−1 dt
k ~f k
L1
t
t
ϕ1(x0, r−1)−1(cid:0)w(B(x0, r−1))(cid:1)−1 k ~f k
ϕ1(x0, r)−1(cid:0)w(B(x0, r))(cid:1)−1 k ~f k
L1
L1
w(cid:0)B(x0,r−1),l2(cid:1)
w(cid:0)B(x0,r),l2(cid:1) = k ~f kM
1,ϕ1
w
(l2).
Lemma 4.4. Let 1 ≤ p < ∞, 0 < α ≤ 1, λ > 3 +
p > 1 the inequality
α
n
and w ∈ Ap. Then, for
λ,α( ~f )(cid:13)(cid:13)Lp
(cid:13)(cid:13)g∗
w(cid:0)B,l2(cid:1) .(cid:0)w(B)(cid:1)
1
p Z ∞
2r
k ~f k
Lp
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
p
dt
t
holds for any ball B = B(x0, r) and for all ~f ∈ Lp,loc
w (l2).
Moreover, for p = 1 the inequality
λ,α( ~f )(cid:13)(cid:13)W L1
(cid:13)(cid:13)g∗
w(cid:0)B,l2(cid:1) . w(B)Z ∞
2r
k ~f k
L1
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)−1 dt
t
holds for any ball B = B(x0, r) and for all ~f ∈ L1locl2.
Proof. From the definition of g∗
λ,α(f ), we readily see that
λ,α( ~f )(x)(cid:13)(cid:13)l2
(cid:13)(cid:13)g∗
t
0 ZRn(cid:18)
0 Zx−y<t(cid:18)
0 Zx−y≥t(cid:18)
=(cid:13)(cid:13)(cid:13)(cid:16)Z ∞
≤(cid:13)(cid:13)(cid:13)(cid:16)Z ∞
+(cid:13)(cid:13)(cid:13)(cid:16)Z ∞
:= III + IV.
t + x − y(cid:19)nλ
(cid:16)Aα
t + x − y(cid:19)nλ
t + x − y(cid:19)nλ
t
t
~f (y, t)(cid:17)2 dydt
(cid:16)Aα
(cid:16)Aα
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
~f (y, t)(cid:17)2 dydt
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
~f (y, t)(cid:17)2 dydt
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
First, let us estimate III.
III ≤(cid:13)(cid:13)(cid:13)(cid:16)Z ∞
0 Zx−y<t(cid:18)
t
t + x − y(cid:19)nλ
(cid:16)Aα
14
~f (y, t)(cid:17)2 dydt
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
.
≤(cid:13)(cid:13)Gα
~f (x)(cid:13)(cid:13)l2
Now, let us estimate IV.
∞
∞
∞
∞
.
:=
0 Zx−y≤2jt(cid:16)Aα
Gα,2j ( ~f )(x)(cid:13)(cid:13)(cid:13)l2
Xj=1Z ∞
0 Z2j−1t≤x−y≤2jt(cid:18)
IV ≤(cid:13)(cid:13)(cid:13)(cid:16)
Xj=1Z ∞
0 Z2j−1t≤x−y≤2jt
.(cid:13)(cid:13)(cid:13)(cid:16)
2−jnλ(cid:13)(cid:13)(cid:13)(cid:16)Z ∞
Xj=1
2−jnλ(cid:13)(cid:13)(cid:13)
Xj=1
w(cid:0)B,l2(cid:1) ≤ kGα
w(cid:0)B,l2(cid:1) .(cid:0)w(B)(cid:1)
w(cid:0)B,l2(cid:1) +
p Z ∞
λ,α( ~f )k
kGα
kg∗
~f k
~f k
Lp
Lp
Lp
2r
.
1
k ~f kLp
In the following, we will estimate kGα,2j ( ~f )k
into two parts.
By Lemma 4.3, we have
Thus,
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
t
t + x − y(cid:19)nλ
2−jnλ(cid:16)Aα
~f (y, t)(cid:17)2 dydt
(cid:16)Aα
~f (y, t)(cid:17)2 dydt
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
~f (y, t)(cid:17)2 dydt
2− jnλ
2 kGα,2j ( ~f )k
∞
Xj=1
Lp
w(cid:0)B,l2(cid:1).
dt
t
p
.
Lp
w(B(x0,t))(cid:0)w(B(x0, t))(cid:1)− 1
w(cid:0)B,l2(cid:1). We divide kGα,2j ( ~f )k
w(cid:0)B,l2(cid:1),
w(cid:0)B,l2(cid:1) + kGα,2j ( ~f∞)k
Lp
Lp
(4.6)
(4.7)
Lp
w(cid:0)B,l2(cid:1)
(4.8)
.
dt
t
(4.9)
where ~f0(y) = ~f (y)χ2B(y), ~f∞(y) = ~f (y) − ~f∞(y). For the first part, by Lemma
4.1,
kGα,2j ( ~f0)k
Lp
2 +α)kGα( ~f0)kLp
w(l2) . 2j( 3n
2 +α)kf k
kGα,2j ( ~f )k
Lp
w(cid:0)B,l2(cid:1) ≤ kGα,2j ( ~f0)k
w(cid:0)B,l2(cid:1) . 2j( 3n
2 +α)w(B)
. 2j( 3n
1
p Z ∞
2r
k ~f k
Lp
w(cid:0)B,l2(cid:1)
p
Lp
w(cid:0)B(x0,t),l2)(cid:0)w(B(x0, t))(cid:1)− 1
~f (y, t)(cid:17)2 dydt
tn+1(cid:17)l/2(cid:13)(cid:13)(cid:13)l2
tn+1!
~f ∗ φt(y)(cid:19)2 dydt
(cid:13)(cid:13)(cid:13)l2
t3n+1!
k ~f∞(z)kl2dz(cid:19)2 dydt
1
2
1
2
.
=(cid:13)(cid:13)(cid:13)(cid:16)Z ∞
0 Zx−y≤2jt(cid:16)Aα
Z ∞
0 Zx−y≤2jt(cid:18) sup
=(cid:13)(cid:13)(cid:13)
≤ Z ∞
0 Zx−y≤2jt(cid:18)Zz−y≤t
φ∈Cα
15
For the second part.
(cid:13)(cid:13)(cid:13)
Gα,2j ( ~f∞)(x)(cid:13)(cid:13)(cid:13)l2
Since x − z ≤ y − z + x − y ≤ 2j+1t, we get
(cid:13)(cid:13)(cid:13)
Gα,2j ( ~f∞)(x)(cid:13)(cid:13)(cid:13)l2
t3n+1!
k ~f∞(z)kl2dz(cid:19)2 dydt
1
2
≤ Z ∞
≤ Z ∞
0 Zx−y≤2jt(cid:18)Zx−z≤2j+1t
0 (cid:18)Zz−x≤2j+1t
2 ZRn Zt≥ x−z
2 Zx0−z>2r
k ~f (z)kl2
x − zn
k ~f∞(z)k2
l2
dz.
2j+1
3jn
jn
≤ 2
≤ 2
t2n+1!
k ~f∞(z)kl2dz(cid:19)2 2jndt
1
2
1
2
dt
t2n+1!
dz
For z − x ≥ x0 − z − x − x0 ≥ x0 − z −
theorem and Holder's inequality, we obtain
1
2
x0 − z =
1
2
x0 − z, so by Fubini's
(cid:13)(cid:13)(cid:13)
Gα,2j ( ~f∞)(x)(cid:13)(cid:13)(cid:13)l2
≤ 2
= 2
≤ 2
≤ 2
≤ 2
≤ 2
So,
3jn
3jn
3jn
2 Zx0−z>2r
2 Zx0−z>2r
2 Z ∞
2r Zx0−z<t
2 Z ∞
2 Z ∞
2 Z ∞
k ~f k
2r
2r
2r
3jn
3jn
3jn
kGα,2j ( ~f∞)k
3jn
2 w(B)
Combining (4.8), (4.9) and (4.10), we have
Lp
w(cid:0)B,l2(cid:1) ≤ 2
w(cid:0)B,l2(cid:1) . 2j( 3n
Lp
kGα,2j ( ~f )k
Thus,
k ~f (z)kl2
x0 − zn
dz
k ~f (z)kl2Z ∞
x0−z
k ~f (z)kl2dz
dz
dt
tn+1
dt
tn+1
dt
tn+1
.
kk ~f (·)kl2kL1(B(x0,t))
dt
tn+1
dt
t
.
k ~f (·)kl2kLp
w(B(x0,t)) kw−1kLp′ (B(x0,t))
p
Lp
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
p Z ∞
k ~f k
Lp
2r
1
p
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
Lp
p
.
dt
t
(4.10)
dt
t
.
(4.11)
2 +α) w(B)
1
p Z ∞
2r
k ~f k
kg∗
λ,α( ~f )k
~f k
Lp
w(cid:0)B,l2(cid:1) ≤ kGα
Lp
w(cid:0)B,l2(cid:1) +
16
2− jnλ
2 kGα,2j ( ~f )k
∞
Xj=1
(4.12)
Lp
w(cid:0)B,l2(cid:1).
Since λ > 3 +
α
n
, by (4.7), (4.11) and (4.12), we have the desired lemma.
Proof of Theorem 1.2
From inequality (4.13) we have
kg∗
λ,α( ~f )kM
p,ϕ2
w
(l2) ≤ kGα
~f kM
p,ϕ2
w
(l2) +
∞
Xj=1
By Theorem 1.1, we have
2− jnλ
2 kGα,2j ( ~f )kM
p,ϕ2
w
(l2).
(4.13)
kGα
~f kM
(l2) . k ~f kM
p,ϕ1
w
(l2).
p,ϕ2
w
(4.14)
In the following, we will estimate kGα,2j ( ~f )kM
variables and Theorem 3.2, we get
p,ϕ2
w
(l2). Thus, by substitution of
kGα,2j ( ~f )kM
. 2j( 3n
2 +α)
p,ϕ2
w
(l2)
sup
x0∈Rn,r>0
= 2j( 3n
2 +α)
sup
x0∈Rn,r>0
. 2j( 3n
2 +α)
sup
x0∈Rn,r>0
p,ϕ1
w
(l2).
= 2j( 3n
2 +α)k ~f kM
α
n
ϕ2(x0, r)−1Z ∞
r
ϕ2(x0, r−1)−1 r
k ~f k
1
r Z r
0
dt
t
Lp
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
p
k ~f k
Lp
w(cid:0)B(x0,t−1),l2(cid:1)(cid:0)w(B(x0, t−1))(cid:1)− 1
p k ~f k
p
dt
t
Lp
w(cid:0)B(x0,r−1),l2(cid:1)
(4.15)
ϕ1(x0, r−1)−1(cid:0)w(B(x0, r−1))(cid:1)− 1
Since λ > 3 +
, by (4.13), (4.14) and (4.15), we have the desired theorem.
Lemma 4.5. Let 1 < p < ∞, 0 < α ≤ 1, w ∈ Ap and b ∈ BM O.
Then the inequality
k[b, Gα]k ~f k
Lp
w(cid:0)B,l2(cid:1) .(cid:0)w(B)(cid:1)
1
p Z ∞
2r
t
lnk(cid:16)e+
r(cid:17) k ~f k
Lp
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
p
dt
t
holds for any ball B = B(x0, r) and for all f ∈ Lp,loc
w (l2).
Proof. We decompose ~f = ~f0 + ~f∞, where ~f0 = ~f χ2B and ~f∞ = ~f − ~f0. Then
k[b, Gα]k ~f k
By Lemma 4.2, we have that
k[b, Gα]k ~f0k
Lp
Lp
w(cid:0)B,l2(cid:1) ≤ k[b, Gα]k ~f0k
w(cid:0)B,l2(cid:1) . kbkk
∗ k ~f0kLp
. kbkk
∗ w(B)
1
Lp
w(cid:0)B,l2(cid:1) + k[b, Gα]k ~f∞k
Lp
w(cid:0)B,l2(cid:1).
w(l2) = kbkk
∗ k ~f k
Lp
w(cid:0)2B,l2(cid:1)
p Z ∞
2r
k ~f k
Lp
w(cid:0)B(x0,t),l2(cid:1) (cid:0)w(B(x0, t))(cid:1)− 1
p
dt
t
.
17
[b(x) − b(z)]kφt(y − z) ~f∞(z)dz(cid:12)(cid:12)(cid:12)
[b(x) − bB,w]kφt(y − z) ~f∞(z)dz(cid:12)(cid:12)(cid:12)
[b(z) − bB,w]kφt(y − z) ~f∞(z)dz(cid:12)(cid:12)(cid:12)
2 dydt
2 dydt
1
1
2(cid:13)(cid:13)(cid:13)l2
tn+1(cid:17)
2(cid:13)(cid:13)(cid:13)l2
tn+1(cid:17)
2(cid:13)(cid:13)(cid:13)l2
tn+1(cid:17)
1
2 dydt
w(B) + kB(·)kLp
w(B).
φt(y − z) ~f∞(z)dz(cid:12)(cid:12)(cid:12)
2 dydt
tn+1(cid:17)
1
2(cid:13)(cid:13)(cid:13)l2
For the second part, we divide it into two parts.
Therefore
(cid:13)(cid:13)(cid:13)
[b, Gα]k ~f∞(x)(cid:13)(cid:13)(cid:13)l2
First, for A(x), we find that
Lp
sup
sup
sup
k[b, Gα]k ~f∞k
:= A(x) + B(x).
=(cid:13)(cid:13)(cid:13)(cid:16)Z ZΓ(x)
≤(cid:13)(cid:13)(cid:13)(cid:16)Z ZΓ(x)
+(cid:13)(cid:13)(cid:13)(cid:16)Z ZΓ(x)
φ∈Cα(cid:12)(cid:12)(cid:12)ZRn
φ∈Cα(cid:12)(cid:12)(cid:12)ZRn
φ∈Cα(cid:12)(cid:12)(cid:12)ZRn
w(cid:0)B,l2(cid:1) ≤ kA(·)kLp
φ∈Cα(cid:12)(cid:12)(cid:12)ZRn
A(x) = b(x) − bB,wk(cid:13)(cid:13)(cid:13)(cid:16)ZZΓ(x)
k(cid:13)(cid:13)Gα
~f∞(x)(cid:13)(cid:13)l2
b(x) − bB,wkp(cid:16)(cid:13)(cid:13)Gα
b(x) − bB,wkp w(x)dx(cid:19)
=(cid:12)(cid:12)b(x) − bB,w(cid:12)(cid:12)
w(B) =(cid:18)ZB
≤(cid:18)ZB
sup
.
kA(·)kLp
By Lemma 3.7 and from the inequality (4.4), we can get
For B(x), since y − x < t, we get x − z < 2t. Thus, by Minkowski's inequality,
1
p
w(x)dx(cid:19)
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
Lp
p
p
dt
t
dt
t
.
1
2r
Lp
k ~f k
~f∞(x)(cid:13)(cid:13)l2(cid:17)p
p Z ∞
w(cid:0)B(x0,t),l2(cid:1)(cid:0)w(B(x0, t))(cid:1)− 1
bB,w − b(z)k ~f∞(z)dz(cid:12)(cid:12)(cid:12)
dz(cid:12)(cid:12)(cid:12)
~f∞(z)(cid:13)(cid:13)l2
bB,w − b(z)k(cid:13)(cid:13)
~f (z)(cid:13)(cid:13)l2
x − zn
dz
≤ kbkk
∗w(B)
1
p Z ∞
2r
k ~f k
B(x) ≤(cid:13)(cid:13)(cid:13)(cid:16)Z ZΓ(x)(cid:12)(cid:12)(cid:12)Zx−z<2t
(cid:12)(cid:12)(cid:12)Zx−z<2t
bB,w − b(z)k(cid:13)(cid:13)
.(cid:16)Z ∞
≤Zx0−z>2r
0
1
2(cid:13)(cid:13)(cid:13)l2
2 dydt
t3n+1(cid:17)
t2n+1(cid:17)
2 dt
1
2
18
For B(x), using the inequality z − x ≥ 1
B(x) .Zx0−z>2r
.Zx0−z>2r
.Z ∞
2r Z2r≤x0−z≤t
b(z) − bB,wk(cid:13)(cid:13)
b(z) − bB,wk(cid:13)(cid:13)
Applying Holder's inequality and by Lemma 3.7, we get
dt
tn+1
dt
tn+1
.
2z − x0, we have
dz
dz
x0−z
b(z) − bB,wk(cid:13)(cid:13)
x0 − zn
~f (z)(cid:13)(cid:13)l2
~f (z)(cid:13)(cid:13)l2 Z ∞
~f (z)(cid:13)(cid:13)l2
dz(cid:19)
r(cid:17) kw−1/pkLp′ (B(x,t)) k ~f k
r(cid:17) k ~f k
w(z)1−p′
Lp
1
p′
t
b(z) − bB,wkp′
kB(·)kLp
w(B) . w(B)
. kbk∗ w(B)
. kbk∗w(B)
1
1
1
p Z ∞
2r (cid:18)ZB(x0,t)
p Z ∞
2r (cid:16)1 + lnk t
p Z ∞
lnk(cid:16)e +
p Z ∞
2r
2r
1
Thus,
(cid:13)(cid:13)[b, Gα]k ~f(cid:13)(cid:13)Lp
w(cid:0)B,l2(cid:1) . kbk∗ w(B)
lnk(cid:16)e+
kk ~f (·)kl2kLp
w(B(x0,t))
dt
tn+1
dt
tn+1
.
Lp
w(cid:0)B(x0,t),l2(cid:1)
w(cid:0)B(x0,t),l2(cid:1) w(B(x0, t))−1/p dt
r(cid:17) k ~f k
w(cid:0)B(x0,t),l2(cid:1) w(B(x0, t))−1/p dt
Lp
t
t
t
.
Proof of Theorem 1.3
By substitution of variables, we obtain
k[b, Gα]k ~f kM
p,ϕ2
w
(l2)
t
2r
ϕ2(x0, r)−1 Z ∞
lnk(cid:16)e +
ϕ2(x0, r)−1 Z r−1
lnk(cid:16)e +
r Z r
lnk(cid:16)e +
ϕ1(x0, r−1)−1w(B(x0, r−1))− 1
1
1
0
0
t
Lp
Lp
w(cid:0)B(x0,t),l2(cid:1) w(B(x0, t))−1/p dt
r(cid:17) k ~f k
tr(cid:17) k ~f k
w(cid:0)B(x0,t−1),l2(cid:1) w(B(x0, t−1))− 1
t(cid:17) k ~f k
w(cid:0)B(x0,t−1),l2(cid:1) w(B(x0, t−1))− 1
w(cid:0)B(x0,r−1),l2(cid:1)
p k ~f k
Lp
Lp
r
p
p
w(B(x0,r),l2)
dt
t
dt
t
ϕ1(x0, r)−1w(B(x0, r))− 1
= sup
x∈Rn, r>0
kbk∗ ϕ2(x0, r−1)−1 r
. kbk∗
sup
x0∈Rn,r>0
. kbk∗
sup
x0∈Rn,r>0
. kbk∗
= kbk∗
sup
x0∈Rn,r>0
sup
x0∈Rn,r>0
= kbk∗ k ~f kM
p,ϕ1
w
(l2).
~f(cid:13)(cid:13)Lp
p (cid:13)(cid:13)
By using the argument as similar as the above proofs and that of Theorem
1.2, we can also show the boundedness of [b, g∗
λ,α]k.
19
References
[1] Akbulut A., Guliyev V.S. and Mustafayev R.: On the boundedness of the
maximal operator and singular integral operators in generalized Morrey
spaces, Math. Bohem. 137 (1), 27-43 (2012).
[2] Calderon A.P.: Commutators of singular integral operators, Proc. Natl.
Acad. Sci. USA 53, 1092-1099 (1965).
[3] Calderon A.P.: Cauchy integrals on Lipschitz curves and related operators,
Proc. Natl. Acad. Sci. USA 74 (4), 1324-1327 (1977).
[4] Carro M., Pick L., Soria J., Stepanov V D.: On embeddings between classical
Lorentz spaces, Math. Inequal. Appl. 4, 397-428 (2001).
[5] Chen Y.: Regularity of solutions to elliptic equations with VMO coefficients,
Acta Math. Sin. (Engl. Ser.) 20, 1103-1118 (2004).
[6] Chiarenza F., Frasca M.: Morrey spaces and Hardy-Littlewood maximal
function, Rend Mat. 7, 273-279 (1987).
[7] Chiarenza F., Frasca M., Longo P.: Interior W 2,p-estimates for nondivergence
elliptic equations with discontinuous coefficients, Ricerche Mat. 40, 149-168
(1991).
[8] Chiarenza F., Frasca M., Longo P.: W 2,p-solvability of Dirichlet problem for
nondivergence elliptic equations with VMO coefficients, Trans. Amer. Math.
Soc. 336, 841-853 (1993).
[9] Coifman R., Rochberg R., Weiss G.: Factorization theorems for Hardy spaces
in several variables, Ann. of Math. 103 (2), 611-635 (1976).
[10] Fazio G. Di, Ragusa M.A.: Interior estimates in Morrey spaces for strong
solutions to nondivergence form equations with discontinuous coefficients, J.
Funct. Anal. 112, 241-256 (1993).
[11] Fan D., Lu S. and Yang D.: Boundedness of operators in Morrey spaces
on homogeneous spaces and its applications, Acta Math. Sinica (N. S.) 14,
suppl., 625-634 (1998).
[12] Huang J.Z., Liu Y.: Some characterizations of weighted Hardy spaces, J.
Math. Anal. Appl. 363, 121-127 (2010).
[13] Deringoz, F., Guliyev, V.S., Samko, S.: Boundedness of maximal and sin-
gular operators on generalized Orlicz-Morrey spaces. accepted in Advances
in Harmonic Analysis and Operator Theory, Series: Operator Theory: Ad-
vances and Applications, Vol. 235, 1-24 (2014).
20
[14] Ding Y., Lu S.Z. and Yabuta K.: On commutators of Marcinkiewicz integrals
with rough kernel, J. Math. Anal. Appl. 275, 60-68 (2002).
[15] Giaquinta M.: Multiple integrals in the calculus of variations and nonlinear
elliptic systems. Princeton Univ. Press, Princeton, NJ, 1983.
[16] Grafakos L.: Classical and Modern Fourier Analysis. Pearson Education,
Inc. Upper Saddle River, New Jersey, 2004.
[17] Guliyev, V.S.: Integral operators on function spaces on the homogeneous
groups and on domains in Rn. Doctor's degree dissertation, Mat. Inst.
Steklov, Moscow, 329 pp. (in Russian) (1994).
[18] Guliyev, V.S.: Boundedness of the maximal, potential and singular operators
in the generalized Morrey spaces, J. Inequal. Appl. Art. ID 503948 (2009).
20 pp.
[19] Guliyev, V.S., Aliyev S.S., Karaman T. : Boundedness of sublinear operators
and commutators on generalized Morrey spaces, Abstr. Appl. Anal. 2011,
Art. ID 356041, 18 pp.
[20] Guliyev, V.S., Aliyev, S.S., Karaman, T., Shukurov, P.S.: Boundedness of
sublinear operators and commutators on generalized Morrey Space. Int. Eq.
Op. Theory. 71 (3), 327-355 (2011).
[21] Guliyev, V.S.: Boundedness of classical operators and commutators of real
analysis in generalized weighted Morrey spaces. Some applications, Interna-
tional conference in honour of Professor V.I. Burenkov on the occasion of his
70th birthday to be held in Kirsehir, Turkey, from May 20 to may 27, 2011.
[22] Guliyev, V.S.: Generalized weighted Morrey spaces and higher order com-
mutators of sublinear operators, Eurasian Math. J. 3(3), 33-61 (2012).
[23] Guliyev, V.S., L. Softova, Global regularity in generalized Morrey spaces of
solutions to nondivergence elliptic equations with VMO coefficients, Poten-
tial Anal. 38 (4) 2013, 843-862.
[24] Guliyev, V.S., L. Softova, Generalized Morrey regularity for parabolic equa-
tions with discontinuity data, Proc. Edinb. Math. Soc. (in press).
[25] Guliyev, V.S., Karaman, T., Mustafayev R.Ch., Serbetci A., Commutators of
sublinear operators generated by Calder´on-Zygmund operator on generalized
weighted Morrey spaces, Czechoslovak Math. J. (in press).
[26] Guliyev, V.S., Shukurov, P.S. : Commutators of intrinsic square functions
on generalized Morrey spaces, Proceedings of IMM of NAS of Azerbaijan.
(in press).
21
[27] Guliyev, V.S., : Commutators of intrinsic square functions on generalized
weighted Morrey spaces, submitted.
[28] Karaman, T., Guliyev, V.S., Serbetci A., Boundedness of sublinear operators
generated by Calder´on-Zygmund operators on generalized weighted Morrey
spaces, Scientic Annals of "Al.I. Cuza" University of Iasi, 60 (1), 1-18 2014.
DOI: 10.2478/aicu-2013-0009
[29] Y. Komori and S. Shirai, Weighted Morrey spaces and a singular integral
operator, Math. Nachr. (2) 282 (2009), 219-231.
[30] Kufner A., John O. and Fu¸cik S.:Function Spaces. Noordhoff International
Publishing: Leyden, Publishing House Czechoslovak Academy of Sciences:
Prague, 1977.
[31] Lerner A.K.: Sharp weighted norm inequalities for Littlewood-Paley opera-
tors and singular integrals, Adv. Math. 226, 3912-3926 (2011).
[32] Morrey, C.B.: On the solutions of quasi-linear elliptic partial differential
equations. Trans. Amer. Math. Soc. 43, 126-166 (1938).
[33] Muckenhoupt B.: Weighted norm inequalities for the Hardy maximal func-
tion, Trans. Amer. Math. Soc. 165, 207-226 (1972).
[34] Muckenhoupt B. and Wheeden R.: Weighted norm inequalities for fractional
integrals, Trans. Amer. Math. Soc. 192, 261-274 (1974).
[35] Mustafayev R.Ch.: On boundedness of sublinear operators in weighted Mor-
rey spaces, Azerb. J. Math. 2 (1), 66-79 (2012).
[36] Wang, H.: Weak type estimates for intrinsic square functions on weighted
Morrey spaces. Anal. Theory Appl. 29 (2), 104-119 (2013).
[37] Wang, H.:
Intrinsic square functions on the weighted Morrey spaces. J.
Math. Anal. Appl. 396, 302-314 (2012).
[38] Wang, H.: Boundedness of intrinsic square functions on the weighted weak
Hardy spaces. Integr. Equ. Oper. Theory 75, 135-149 (2013).
[39] Wang, H., Liu, H. P.: Weak type estimates of intrinsic square functions on
the weighted Hardy spaces. Arch. Math. 97, 49-59 (2011).
[40] Wilson, M.: The intrinsic square function. Rev. Mat. Iberoam. 23, 771-791
(2007).
[41] Wilson, M.: Weighted Littlewood-Paley theory and Exponential-square in-
tegrability. Lecture Notes in Math. vol. 1924, Springer-Verlag (2007).
22
Vagif S. Guliyev
Ahi Evran University, Department of Mathematics
Kirsehir, Turkey and
Institute of Mathematics and Mechanics
Academy of Sciences of Azerbaijan
F. Agayev St. 9, Baku, AZ 1141, Azerbaijan
Mehriban N. Omarova
Baku State University
Baku, AZ 1148, Azerbaijan
23
|
1812.05435 | 3 | 1812 | 2019-09-24T07:32:02 | Multiplicities, invariant subspaces and an additive formula | [
"math.FA"
] | Let $T = (T_1, \ldots, T_n)$ be a commuting tuple of bounded linear operators on a Hilbert space $\mathcal{H}$. The multiplicity of $T$ is the cardinality of a minimal generating set with respect to $T$. In this paper, we establish an additive formula for multiplicities of a class of commuting tuples of operators. A special case of the main result states the following: Let $n \geq 2$, and let $\mathcal{Q}_i$, $i = 1, \ldots, n$, be a proper closed shift co-invariant subspaces of the Dirichlet space or the Hardy space over the unit disc in $\mathbb{C}$. If $\mathcal{Q}_i^{\bot}$, $i = 1, \ldots, n$, is a zero-based shift invariant subspace, then the multiplicity of the joint $M_{\boldsymbol{z}} = (M_{z_1}, \ldots, M_{z_n})$-invariant subspace $(\mathcal{Q}_1 \otimes \cdots \otimes \mathcal{Q}_n)^\perp$ of the Dirichlet space or the Hardy space over the unit polydisc in $\mathbb{C}^n$ is given by \[ \mbox{mult}_{M_{\boldsymbol z}|_{ (\mathcal{Q}_1 \otimes \cdots \otimes \mathcal{Q}_n)^\perp}} (\mathcal{Q}_1 \otimes \cdots \otimes \mathcal{Q}_n)^\perp = \sum_{i=1}^n (\mbox{mult}_{M_z|_{\mathcal{Q}_i^\perp}} (\mathcal{Q}_i^{\bot})) = n. \] A similar result holds for the Bergman space over the unit polydisc. | math.FA | math |
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE
FORMULA
ARUP CHATTOPADHYAY, JAYDEB SARKAR, AND SRIJAN SARKAR
Dedicated to Professor Kalyan Bidhan Sinha on the occasion of his 75th birthday
Abstract. Let T = (T1, . . . , Tn) be a commuting tuple of bounded linear operators on a
Hilbert space H. The multiplicity of T is the cardinality of a minimal generating set with
respect to T .
In this paper, we establish an additive formula for multiplicities of a class
of commuting tuples of operators. A special case of the main result states the following:
Let n ≥ 2, and let Qi, i = 1, . . . , n, be a proper closed shift co-invariant subspaces of the
Dirichlet space or the Hardy space over the unit disc in C. If Q⊥
i , i = 1, . . . , n, is a zero-based
shift invariant subspace, then the multiplicity of the joint Mz = (Mz1 , . . . , Mzn)-invariant
subspace (Q1 ⊗ · · · ⊗ Qn)⊥ of the Dirichlet space or the Hardy space over the unit polydisc
in Cn is given by
multMz(Q1 ⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ =
nX
i=1
(multMz Q⊥
i
(Q⊥
i )) = n.
A similar result holds for the Bergman space over the unit polydisc.
1. Introduction
This paper is concerned with an additive formula for a numerical invariant of commuting
tuples of bounded linear operators on Hilbert spaces. The additive formula arises naturally in
connection with a class of simple invariant subspaces of the two-variable Hardy space H 2(D2)
[4]. From function Hilbert space point of view, our additive formula is more refined for zero-
based invariant subspaces of the Dirichlet space, the Hardy space, the Bergman space and
the weighted Bergman spaces over the open unit polydisc Dn in Cn.
To be more specific, let us first define the numerical invariant. Given an n-tuple of com-
muting bounded linear operators T := (T1, . . . , Tn) on a Hilbert space H, we denote by
where
multT (H) = min{#G : [G]T = H, G ⊆ H},
[G]T = span{T k(G) : k ∈ Zn
+},
2010 Mathematics Subject Classification. 47A13, 47A15, 47A16, 47A80, 47B37, 47B38, 47M05, 46C99,
32A35, 32A36, 32A70.
Key words and phrases. Hardy space, Dirichlet space, Bergman and weighted Bergman spaces, polydisc,
rank, multiplicity, joint invariant subspaces, semi-invariant subspaces, zero-based invariant subspaces, tensor
product Hilbert spaces.
1
2
CHATTOPADHYAY, SARKAR, AND SARKAR
and T k = T k1
1
· · · T kn
n
for all k = (k1, . . . , kn) ∈ Zn
+. If
multT (H) = m < ∞,
then we say that the multiplicity of T is m. One also says that T is m-cyclic. If m = 1, then
we also say that T is cyclic, or simply cyclic. A subset G of H is said to be generating subset
with respect to T if [G]T = H.
We pause to note that the computation of multiplicities of (even concrete and simple)
bounded linear operators is a challenging problem (perhaps due to its inherent dynamical
nature). We refer Rudin [17] for concrete (as well as pathological) examples of invariant
subspaces of H 2(D2) of infinite multiplicities and [4, 5, 11, 12, 13] for some definite results on
computations of multiplicities (also see [7]).
The following example, as hinted above, illustrates the complexity of computations of the
multiplicities of general function Hilbert spaces. As a first step, we consider the Hardy
space H 2(D) over D (the space of all square summable analytic functions on D) and the
multiplication operator Mz by the coordinate function z. Let S be a closed Mz-invariant
subspace of H 2(D). Then Q = S ⊥ is a closed M ∗
It then
follows from Beurling that
z -invariant subspace of H 2(D).
multMzS (S) = 1,
that is, MzS on S is cyclic. Moreover, taking into account that multMz (H 2(D)) = 1, we
obtain (cf. Proposition 2.3)
multPQMzQ(Q) = 1,
where PQ denote the orthogonal projection of H 2(D) onto Q.
Now we consider the commuting pair of multiplication operators Mz = (Mz1, Mz2) on H 2(D2)
(the Hardy space over the bidisc). Observe that H 2(D2) ∼= H 2(D)⊗H 2(D). Let Q1 and Q2 be
two non-trivial closed M ∗
z2)-
invariant subspace of H 2(D2), and so (Q1 ⊗ Q2)⊥ is a joint (Mz1, Mz2)-invariant subspace
of H 2(D2). Set Mz(Q1⊗Q2)⊥ = (Mz1(Q1⊗Q2)⊥, Mz2(Q1⊗Q2)⊥). An equivalent reformulation of
Douglas and Yang's question (see page 220 in [6] and also [4]) then takes the following form:
Is
z -invariant subspaces of H 2(D). Then Q1 ⊗ Q2 is a joint (M ∗
z1
, M ∗
multMz(Q1⊗Q2)⊥ (Q1 ⊗ Q2)⊥ = 2?
The answer to this question is yes and was obtained by Das along with the first two authors
in [4]. This result immediately motivates (see page 1186, [4]) the following natural question:
Consider the joint Mz = (Mz1, . . . , Mzn)-invariant subspace (Q1 ⊗· · ·⊗Q2)⊥ of H 2(Dn) where
Q1, . . . , Qn are non-trivial closed M ∗
z -invariant subspaces of H 2(D). Is then
multMz(Q1⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ = n?
This can be reformulated more concretely as follows: Let Hi be the Dirichlet space, the
Hardy space, the Bergman space, or the weighted Bergman spaces over D (or, more generally,
a reproducing kernel Hilbert spaces of analytic functions on D for which the operator Mz of
multiplication by the coordinate function z on Hi is bounded), i = 1, . . . , n. Suppose Q⊥
is
i
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
3
an Mz-invariant closed subspace of Hi, i = 1, . . . , n. Is then
multMz(Q1⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ =
nX
i=1
(multMzQ⊥
i (cid:0)Q⊥
i ))?
In this paper, we aim to propose an approach to verify the above equality for a large class
of function Hilbert spaces over Dn. The methods and techniques used in this paper are
completely different from [4], and can also be applied for proving more powerful results in the
setting of general Hilbert spaces. There is indeed a more substantial answer, valid in a larger
context of tensor products of Hilbert spaces (see Theorem 4.3).
Let H ⊆ O(D) be a reproducing kernel Hilbert space (or, the Dirichlet space, the Hardy
space, the Bergman space, or the weighted Bergman spaces over D) and let the operator Mz
is bounded on H. Suppose S is a Mz-invariant closed subspace of H. We say that S is a
zero-based invariant subspace if there exists λ ∈ D such that f (λ) = 0 for all f ∈ S.
A particular case of our main theorem is the following: Let Hi be the Dirichlet space, the
Hardy space, the Bergman space, or the weighted Bergman spaces over D. Let Si be an M ∗
z -
invariant closed subspace of Hi, i = 1, . . . , n. Suppose Si := Q⊥
is a zero-based Mz-invariant
i
closed subspace of Hi such that
dim(Si ⊖ zSi) < ∞ and [Si ⊖ zSi]MzSi
= Si,
for all i = 1, . . . , n, then
multMz(Q1⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ =
nX
i=1
(multMzQ⊥
i (cid:0)Q⊥
i )) =
nX
i=1
dim(Si ⊖ zSi).
Note that the finite dimensional and generating subspace assumptions are automatically sat-
isfied if Hi is the Hardy space or the Dirichlet space. However, if S is an Mz-invariant closed
subspace of the Bergman space over D, then
dim(S ⊖ zS) ∈ N ∪ {∞}.
We refer the reader to [2, 8, 9] for more information. See also [10] for related results in the
setting of weighted Bergman spaces over D.
The proof of the above additivity formula uses generating wandering subspace property,
geometry of (tensor product) Hilbert spaces and subspace approximation technique.
The paper is organized as follows. In Section 2, we set up notation and prove some basic
results on weak multiplicity of (not necessarily commuting) n-tuples of operators on Hilbert
spaces. In Section 3, we study a lower bound multiplicity of joint invariant subspaces of a
class of commuting n-tuples of operators. The main theorem on additivity formula is proved
in Section 4. The paper is concluded in Section 5 with corollaries of the main theorem and
some general discussions.
2. Notation and basic results
In this section, we introduce the notion of weak multiplicities and describe some prepara-
tory results. This notion is not absolutely needed for the main results of this paper as we
shall mostly work in the setting of multiplicities. However, we believe that the idea of weak
4
CHATTOPADHYAY, SARKAR, AND SARKAR
multiplicities of (not necessary commuting) tuples of operators might be of independent in-
terest.
Throughout this paper the following notation will be adopted: Ti is a bounded linear operator
on a separable Hilbert space Hi, i = 1, . . . , n, and n ≥ 2. We set
and
where
H = H1 ⊗ · · · ⊗ Hn,
T = ( T1, . . . , Tn).
Ti = IH1 ⊗ · · · ⊗ IHi−1 ⊗ Ti ⊗ IHi+1 ⊗ · · · ⊗ IHn ∈ B( H),
for all i = 1, . . . , n. It is now clear that ( T1, . . . , Tn) is a doubly commuting tuple of operators
on H (that is, Ti Tj = Tj Ti and T ∗
p for all 1 ≤ i, j ≤ n and 1 ≤ p < q ≤ n).
Moreover, if multTi(Hi) = 1 for all i = 1, . . . , n, then mult T ( H) = 1. We denote by Dn the
unit polydisc in Cn and by z the element (z1, . . . , zn) in Cn.
Tq = Tq T ∗
p
The above notion of "tensor product of operators" is suggested by natural (and analytic)
examples of reproducing kernel Hilbert spaces over product domains in Cn. For instance, if
{α1, . . . , αn} ⊆ N, then
Kα(z, w) :=
nY
i=1
1
(1 − zi ¯wi)αi
(z, w ∈ Dn),
is a positive definite kernel over the polydisc Dn, and the multiplication operator tuple
(Mz1, . . . , Mzn) defines bounded linear operators on the corresponding reproducing kernel
α(Dn) (known as the weighted Bergman space over Dn with weight α =
Hilbert space L2
(α1, . . . , αn)). It follows that (cf. [19])
H = L2
α1(D) ⊗ · · · ⊗ L2
αn(D),
and
Mz = ( Mz1, . . . , Mzn),
αi(D), i = 1, . . . , n. In particular, if
where Mzi denotes the multiplication operator Mz on L2
α = (1, . . . , 1), then H = H 2(Dn) is the well known Hardy space over the unit polydisc. We
also refer the reader to Popescu [14, 15] for elegant and rich theory of "tensor product of
operators" in multivariable operator theory.
Let H be a Hilbert space, and let A = (A1, . . . , An) be an n-tuple (not necessarily com-
muting) of bounded linear operators on H. Let
w-multA(H) = min{#G : [G]A = H, G ⊆ H},
where
[G]A = span{Ak(G) : k ∈ Zn
+},
1 · · · Akn
and Ak = Ak1
+. If w-multA(H) = m < ∞, then we say that the weak
multiplicity of A is m. We say that A is weakly cyclic if w-multA(H) = 1. A subset G of H
is said to be weakly generating with respect to A if [G]A = H.
n for all k ∈ Zn
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
5
Now let L be a closed subspace of H. Then
WA(L) := L ⊖
nX
i=1
AiL,
is called the wandering subspace of L with respect to PLAL. If, in addition
L = _
k∈Zn
+
(PLAL)k(WA(L)),
then we say that PLAL satisfies the weakly generating wandering subspace property. Here
PLAL = (PLA1L, . . . , PLAnL) and
(PLAL)k = (PLA1L)k1 · · · (PLAnL)kn,
for all k ∈ Zn
+.
Note that if A is commuting and L is joint A-invariant subspace (that is, AiL ⊆ L for
all i = 1 . . . , n), then weakly generating wandering subspace property is commonly known as
generating wandering subspace property.
We now proceed to relate weak multiplicities and dimensions of weakly generating wander-
ing subspaces. Let A be an n-tuple of bounded linear operators on H, L be a joint A-invariant
subspace of H, and let M be a closed subspace of L. Then
since
PWA(L)([M]A) = PWA(L)(M),
PWA(L)(AkM) = 0 for all k ∈ Zn
+ \ {0}.
Now suppose that [M]A = L, that is, M is a weakly generating subspace of L with respect
to A. Then
Hence
WA(L) = PWA(L)(M).
w-multAL(L) ≥ dim WA(L).
Moreover, if L satisfies the weakly generating wandering subspace property, then
Therefore we have proved the following:
w-multA(L) = dim WA(L).
Proposition 2.1. Let L be a closed joint A-invariant subspace of H. If L satisfies the weakly
generating wandering subspace property with respect to AL, then w-multA(L) = dim WA(L).
We now proceed to a variation of Lemma 2.1 in [4] which relates the multiplicity of a
commuting tuple of operators with the weak-multiplicity of the compressed tuple to a semi-
invariant subspace.
Lemma 2.2. Let A be an n-tuple of bounded linear operators on a Hilbert space H. Let L1
and L2 be two joint A-invariant subspaces of H and L2 ⊆ L1. If L = L1 ⊖ L2, then
w-multPLAL(L) ≤ w-multAL1
(L1).
6
CHATTOPADHYAY, SARKAR, AND SARKAR
Proof. We have PLAjPL = PLAjPL1 − PLAjPL2 and thus by AjL2 ⊆ L2 we infer that
for all j = 1, . . . , n. Since AjL1 ⊆ L1, we have
PLAjPL = PLAjPL1,
(PLAiPL)(PLAjPL) = PLAiPL1AjPL1,
that is
for all i, j = 1, . . . , n, and so
(PLAiPL)(PLAjPL) = PL(AiAj)PL1,
(PLAPL)k = PLAkPL1,
for all k ∈ Zn
+. Clearly, if G is a minimal generating subset of L1 with respect to AL1,
then PLG is a generating subset of L with respect to PLAL, and thus w-multPLAL(L) ≤
w-multAL1
(cid:3)
(L1). This completes the proof of the lemma.
In particular, if L1 = H, then Q := H ⊖ L2 is a joint (A∗
n)-invariant subspace of
H. In this case, denote by Ci = PQAiQ the compression of Ai, i = 1, . . . , n, and define the
n-tuple on Q as
1, . . . , A∗
Then we have the following estimate:
CQ = (C1, . . . , Cn).
w-multCQ(Q) ≤ w-multA(H).
Moreover, we also have
Corollary 2.3. Let A = (A1, . . . , An) be a commuting tuple of bounded linear operators on a
Hilbert space H. If Q is a closed joint A∗-invariant subspace of H, then
multCQ(Q) ≤ multA(H).
This has the following immediate (and well-known) application: Suppose A is a commuting
tuple on H. If A is cyclic, then CQ on Q is also cyclic.
3. A lower bound for multiplicities
In this section, we first lay out the setting of joint invariant subspaces of our discussions
throughout the paper. Then we present a lower bound of multiplicities of those joint invariant
subspaces. We begin by recalling the following useful lemma (cf. Lemma 2.5, [18]):
Lemma 3.1. If {Ai}n
i=1 is a commuting set of orthogonal projections on a Hilbert space K,
then L =
ranAi is a closed subspace of K, and
nX
i=1
PL = I −
nY
i=1
(I − Ai)
= A1(I − A2) . . . (I − An) ⊕ A2(I − A3) . . . (I − An) ⊕ . . . + An−1(I − An) ⊕ An.
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
7
Next, we introduce the invariant subspaces of interest. Again, we continue to follow the
notations as introduced in Section 2.
Let Hi be a Hilbert space, Ti a bounded linear operator on Hi, and let Qi be a closed
T ∗
i -invariant subspace of Hi, i = 1, . . . , n. Set Si = Q⊥
i and
Pi = PSi
and Qi = IHi − PSi,
for all i = 1, . . . , n. Recall again that
Pi = IH1 ⊗ . . . ⊗ IHi−1 ⊗ PSi ⊗ IHi+1 ⊗ . . . ⊗ IHn ∈ B( H),
and
for all i, j = 1, . . . , n. By Lemma 3.1, it then follows that
Pi Pj = Pj Pi,
(3.1)
S =
i=1
is a joint T -invariant subspace of H. Moreover
nX
ran Pi,
S = (Q1 ⊗ · · · ⊗ Qn)⊥.
Our main goal is to compute the multiplicity of the commuting tuple T S = ( T1S , . . . , TnS)
on S.
For each i = 1, . . . , n, define Xi ∈ B( H) by
Then X 2
i = Xi = X ∗
i and
Xi = Pi Qi+1 . . . Qn.
XpXq = 0,
for all i = 1, . . . , n, and p 6= q. This implies that {Xi}n
with orthogonal ranges. Then, by virtue of (3.1) one can further rewrite S as
i=1 is a set of orthogonal projections
(3.2)
S =
i=1
and by Lemma 3.1 one represent PS as
nX
ran Pi =
nM
i=1
ranXi,
PS =
nM
i=1
Xi.
Define
(3.3)
F = ranX1 ⊕ ran( Q1X2) ⊕ · · · ⊕ ran( Q1 · · · Qn−1Xn).
Then, as easily seen
for all 1 ≤ i ≤ j and j = 1, . . . , n, it follows that
QiXj = Xj Qi,
ran( Q1 · · · QpXp+1) ⊆ ranXp+1,
8
CHATTOPADHYAY, SARKAR, AND SARKAR
for all p = 1, . . . , n − 1, and consequently
S ⊇ F .
Our first aim is to analyze the closed subspace F and to construct n − 1 nested (and suitable)
closed subspaces {Fi}n−1
i=1 such that
S ⊇ F1 ⊇ · · · ⊇ Fn−1 = F .
To this end, first set
F1 = ranX1 ⊕ ranX2 ⊕ · · · ⊕ ranXn−1 ⊕ ran( Qn−1Xn),
and define
F2 = ranX1 ⊕ ran( Q1X2) ⊕ · · · ⊕ ran( Q1Xn−1) ⊕ ran( Q1 Qn−1Xn).
We then proceed to define Fi, i = 2, . . . , n − 1, as
(3.4) Fi = ran(cid:16)X1 ⊕ Q1X2 ⊕ · · · ⊕ (
i−1Y
t=1
Qt)Xi ⊕ · · · ⊕ (
i−1Y
t=1
Qt)Xn−1 ⊕ (
Qt Qn−1)Xn(cid:17).
i−1Y
t=1
Therefore
(3.5)
PFi = X1 ⊕ Q1X2 ⊕ · · · ⊕ (
i−1Y
t=1
Qt)Xi ⊕ · · · ⊕ (
i−1Y
t=1
Qt)Xn−1 ⊕ (
i−1Y
t=1
Qt Qn−1)Xn,
for all i = 2, . . . , n − 1. Therefore, denoting
A = (
i−2Y
t=1
Qt) Pi−1,
we have
(3.6)
PFi−1⊖Fi = A(Xi ⊕ Xi+1 ⊕ · · · ⊕ Xn−1 ⊕ Qn−1Xn),
for all i = 2, . . . , n − 1. Since AXp = XpA for all p = i, . . . , n, the above formula yields
PFi−1⊖Fi = (Xi ⊕ Xi+1 ⊕ · · · ⊕ Xn−1 ⊕ Qn−1Xn)A.
Let i ∈ {2, . . . , n − 1} be a fixed natural number. We claim that Fi−1 ⊖ Fi is a joint
PFi−1
T PFi−1-invariant subspace, that is
or, equivalently
PFi−1
Tj(Fi−1 ⊖ Fi) ⊆ Fi−1 ⊖ Fi.
(PFi−1
TjPFi−1)PFi−1⊖Fi = PFi−1⊖Fi
TjFi−1⊖Fi,
for all j = 1, . . . , n. There are four cases:
Case I: If j > i, then one has TjA = A Tj and so
PFi−1⊖Fi
TjPFi−1⊖Fi = A(Xi ⊕ Xi+1 ⊕ · · · ⊕ Qn−1Xn) Tj(Xi ⊕ Xi+1 ⊕ · · · ⊕ Qn−1Xn).
On the other hand, since
PFi−1
TjPFi−1⊖Fi = PFi−1A Tj(Xi ⊕ · · · ⊕ Xj ⊕ · · · ⊕ Qn−1Xn),
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
9
and
PFi−1 = X1⊕( Q1X2) ⊕ · · · ⊕ (
i−2Y
t=1
QtXi−1) ⊕ (
i−2Y
t=1
QtXi)⊕
· · · ⊕ (
i−2Y
t=1
QtXn−1) ⊕ (
i−2Y
t=1
Qt Qn−1Xn),
it follows that
PFi−1
TjPFi−1⊖Fi = A(Xi−1 ⊕ Xi ⊕ · · · ⊕ Qn−1Xn) Tj(Xi ⊕ · · · ⊕ Qn−1Xn),
as XtA = 0 for all t = 1, . . . , i − 2, and
i−2Y
t=1
QtA = A. Moreover, since
Xi−1 Tj = ( Pi−1 Qi · · · Qj · · · Qn) Tj = Pi−1 Qi · · · ^QjTjQj · · · Qn,
it follows that
for all t = i, . . . , n. This leads to
Xi−1 TjXt = 0,
PFi−1
TjPFi−1⊖Fi = A(Xi ⊕ Xi+1 ⊕ · · · ⊕ Qn−1Xn) Tj(Xi ⊕ Xi+1 ⊕ · · · ⊕ Qn−1Xn).
Case II: If j = i, then
TiPFi−1⊖Fi = A((gTiPi Qi+1 · · · Qn) ⊕ TiXi+1 ⊕ · · · ⊕ Ti Qn−1Xn),
implies that
PFi−1
TiPFi−1⊖Fi = (
= (
t=1
i−2Y
i−2Y
t=1
Qt)(Xi−1 ⊕ Xi ⊕ · · · ⊕ Qn−1Xn) TiPFi−1⊖Fi
Qt)(Xi−1 ⊕ Xi ⊕ · · · ⊕ Qn−1Xn) TiA(Xi ⊕ Xi+1 ⊕ · · · ⊕ Qn−1Xn)
= A(Xi−1 ⊕ Xi ⊕ · · · ⊕ Qn−1Xn) Ti(Xi ⊕ Xi+1 ⊕ · · · ⊕ Qn−1Xn)
= PFi−1⊖Fi
TiPFi−1⊖Fi,
where the next-to-last equality follows from the fact again that A Ti = TiA, (
and Xi−1 TiXt = 0 for all t = i, . . . , n.
Case III: Let j = i − 1. Since
i−2Y
t=1
Qt)A = A
Ti−1A = (
i−2Y
t=1
Qt) ^Ti−1Pi−1 = A ^Ti−1Pi−1,
10
by setting
it follows that
CHATTOPADHYAY, SARKAR, AND SARKAR
A = (
i−2Y
t=1
Qt) ^Ti−1Pi−1,
Ti−1PFi−1⊖Fi = AXi ⊕ AXi+1 ⊕ · · · ⊕ AXn−1 ⊕ A Qn−1Xn.
Then Xp A = AXp for all p = i, . . . , n, and A A = A implies that
PFi−1
Ti−1PFi−1⊖Fi = A(Xi ⊕ Xi+1 ⊕ · · · ⊕ Xn−1 ⊕ Qn−1Xn)
= PFi−1⊖Fi
Ti−1PFi−1⊖Fi,
where the second equality follows from (3.6) and the fact that Ti−1Pi−1 = Pi−1Ti−1Pi−1.
Case IV: Let j < i − 1. Then it is clear that
TjPFi−1⊖Fi = A(Xi ⊕ Xi+1 ⊕ · · · ⊕ Xn−1 ⊕ Qn−1Xn),
where A = TjA, that is
A = Q1 · · · Qj−1
]TjQj Qj+1 · · · Qi−2 Pi−1.
Note that Xt A = AXt for all t = i, . . . , n, and
A A = Q1 · · · Qj−1
^QjTjQj Qj+1 · · · Qi−2 Pi−1.
Since XpXq = δpqXp for all p and q, it follows that
TjPFi−1⊖Fi = A A(Xi ⊕ Xi+1 ⊕ · · · ⊕ Xn−1 ⊕ Qn−1Xn).
On the other hand, the representation of TjPFi−1⊖Fi above and (3.6) yields
PFi−1
PFi−1⊖Fi
TjPFi−1⊖Fi = A A(Xi ⊕ Xi+1 ⊕ · · · ⊕ Xn−1 ⊕ Qn−1Xn),
and proves the claim.
We turn now to prove that (PFi
T1Fi, . . . , PFi
TnFi) is a commuting tuple for all i =
1, . . . , n − 1, that is
PFi
TsPFi
TtPFi
for all s, t = 1, . . . , n. Fix an i ∈ {1, . . . , n − 1} and let
TtPFi = PFi
TsPFi,
(3.7)
PFi = M1 ⊕ . . . ⊕ Mn,
where Mj, j = 1, . . . , n, denotes the j-th summand in the representation of PFi in (3.5).
Recalling the terms in (3.5), we see that Mj is a product of n distinct commuting orthogonal
projections of the form Pk, Ql and IHm, 1 ≤ k, l, m ≤ n. For each s = 1, . . . , n, we set
Mj = Mj,s Mj,s,
where Mj,s is the s-th factor of Mj and Mj,s is the product of the same factors of Mj, except
the s-th factor of Mj is replaced by IHs. Note again that Mj,s = Ps, Qs, or IHs. We first claim
that
(3.8)
Mj TsMk = 0,
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
11
for all j 6= k. Indeed, if Mj,s = Qs, then Mj TsMk = Mj,s Mj,s TsMk yields
Mj TsMk = Mj,s Ts Mj,sMk = Mj,s TsMj,s Mj,sMk = Mj,s TsMjMk = 0,
as Qs Ts Qs = Qs Ts. Similarly, if Mj,s = Ps, then
Mj TsMk = Mj Mk,s TsMk,s = Mj Mk,sMk,s TsMk,s = MjMk TsMk,s = 0,
as Ps Ts Ps = Ts Ps. The remaining case, Mj,s = IHs, follows from the fact that
This proves the claim. Hence the representation of PFi
TsPFi simplifies as
Mj TsMk = TsMjMk.
(3.9)
Thus
PFi
TsPFi = M1 TsM1 ⊕ · · · ⊕ Mn TsMn.
PFi
TsPFi
TtPFi = M1 TsM1 TtM1 ⊕ · · · ⊕ Mn TsMn TtMn.
Now if s 6= t, then for each j = 1, . . . , n, we have
Mj TsMj TtMj = Mj Mj,s TsMj,sMj,t Tt Mj,tMj
= (Mj Mj,sMj,t) Ts Tt(Mj,s Mj,tMj)
= Mj Ts TtMj,
and hence
TtPFi) = M1 Ts TtM1 ⊕ · · · ⊕ Mn Ts TtMn.
This completes the proof of the commutativity property of the tuple (PFi
i = 1, . . . , n − 1. Furthermore, if s = t, then
TsPFi)(PFi
(PFi
T1Fi, . . . , PFi
TnFi),
Indeed, if Mj,s = Qs, then Mj TsMj = Mj Ts Mj,s gives us
(Mj TsMj)2 = Mj T 2
s Mj.
Mj TsMj TsMj = Mj Ts Mj,s Ts Mj,sMj = Mj Ts Ts Mj,sMj = Mj T 2
s Mj.
Similarly, if Mj,s = Ps or IHs, then Mj TsMj = TsMj, and hence
Mj TsMj TsMj = Mj T 2
s Mj.
Hence we obtain
(3.10)
(PFi
TsPFi)(PFi
TtPFi) = M1 Ts TtM1 ⊕ · · · ⊕ Mn Ts TtMn,
for all s, t = 1, . . . , n.
Therefore, with the notations introduced above, we have proved the following:
Lemma 3.2. If S = (Q1 ⊗ · · · ⊗ Qn)⊥, then S is a joint T -invariant subspace of H and
S ⊇ F1 ⊇ · · · ⊇ Fn−1 = F ,
where F and Fi are defined as in (3.3) and (3.5), respectively. Moreover
PFi−1
T Fi−1 = (PFi−1
T1Fi−1, . . . , PFi−1
TnFi−1),
12
CHATTOPADHYAY, SARKAR, AND SARKAR
is a commuting tuple and
(cid:16)PFi−1
TjFi−1(cid:17)(Fi−1 ⊖ Fi) ⊆ Fi−1 ⊖ Fi,
for all i = 2, . . . , n − 1, and j = 1, . . . , n.
We now proceed to estimate a lower bound of mult T S (S). Note first that ran( Pn−1 Pn) is a
joint T -invariant subspace and
Then F1 is a T -semi invariant subspace, which, by Lemma 2.2, implies that
F1 = S ⊖ ran( Pn−1 Pn).
Now consider the commuting n-tuple PF1
Lemma 3.2 we infer that F1 ⊖ F2 is a joint PF1
F2 = F1 ⊖ (F1 ⊖ F2), it follows again by Lemma 2.2 that
T1F1, . . . , PF1
TnF1) on F1. Then by
T F1-invariant subspace of F1. But since
mult T S (S) ≥ multPF1
(F1).
T F1
T F1 = (PF1
In general, by virtue of Lemma 3.2, we have
multPF1
T F1
(F1) ≥ multPF2
T F2
(F2).
multPFi−1
T Fi−1
(Fi−1) ≥ multPFi
T Fi
(Fi),
for all i = 2, . . . , n − 1, and hence
mult T S (S) ≥ multPF1
T F1
(F1) ≥ . . . ≥ multPFn−1
T Fn−1
(Fn−1) = multPF T F (F ),
where (see (3.3))
F = ranX1 ⊕ ran( Q1X2) ⊕ · · · ⊕ ran( Q1 · · · Qn−1Xn),
and Xi = Pi Qi+1 · · · Qn, i = 1, . . . , n. We summarize the above discussion in the following
theorem:
Theorem 3.3. Let T1, . . . , Tn be bounded linear operators on Hilbert spaces H1, . . . , Hn, re-
spectively. If Qi is a T ∗
i -invariant closed subspace of Hi, i = 1, . . . , n, and
then
S = (Q1 ⊗ · · · ⊗ Qn)⊥,
mult T S (S) ≥ multPF T F (F ).
4. Additivity of multiplicities
We now proceed to prove the reverse inequality in Theorem 3.3. We start with a simple
but useful lemma.
Lemma 4.1. Let (A1, . . . , An) be an n-tuple of bounded linear operators on a Hilbert space
H. If G is a subset of H and (λ1, . . . , λn) ∈ Cn, then
[G](A1,...,An) = [G](A1−λ1IH,...,An−λnIH).
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
13
Proof. Note that, given p ∈ C[z1, . . . , zn] there exists q ∈ C[z1, . . . , zn] such that
p(A1, . . . , An) = p((A1 − λ1IH + λ1IH), . . . , (An − λnIH + λnIH))
= q((A1 − λ1IH), . . . , (An − λnIH)),
which implies that
[G](A1,...,An) ⊆ [G](A1−λ1IH,...,An−λnIH).
The reverse inclusion follows similarly, and hence the result follows.
(cid:3)
Now we return to the problem of rank computation of S as in Theorem 3.3. From now on,
we will use the setting and notations introduced in Section 3. Observe that, by (3.3), we have
where
F = M1 ⊕ · · · ⊕ Mn,
Mi = ran(cid:16) PiY
j6=i
Qj(cid:17).
By defining Mi = PMi, i = 1, . . . , n, one has (see (3.7))
PF = M1 ⊕ · · · ⊕ Mn.
Recall, by virtue of (3.9), that
(4.1)
for all s = 1, . . . , n. And, finally, recall that, by Lemma 3.2, PF T PF is a commuting tuple on
F . The equality in (4.1) implies that
PF TsPF = M1 TsM1 ⊕ · · · ⊕ Mn TsMn,
(PF TsPF )Mi ⊆ Mi
(s = 1, . . . , n),
that is, Mi is a joint PF T PF -invariant subspace of F for all i = 1, . . . , n. Then by virtue of
(3.10), we have
(PF T F )k =
nM
i=1
PMi
T kMi
(k ∈ Zn
+).
Now let G be a minimal generating subset of F with respect to PF T F . Then
F = span{(PF T F )k(G) : k ∈ Zn
+} ⊆
(cid:16)span{PMi
nM
i=1
T kMi(G) : k ∈ Zn
+}(cid:17) ⊆ F ,
and so
F =
(cid:16)span{PMi
nM
i=1
T kMi(G) : k ∈ Zn
+}(cid:17).
Now assume that the point spectrum σp(T ∗
subspace property, and
i Qi) 6= ∅, TiSi satisfies the generating wandering
dim(Si ⊖ TiSi) < ∞,
14
CHATTOPADHYAY, SARKAR, AND SARKAR
for all i = 1, . . . , n. If we then let ¯αi ∈ σp(T ∗
then
i Qi) and T ∗
i vi = ¯αivi for some non-zero vi ∈ Qi,
Ei := ran(cid:16) PSi⊖TiSiY
PCvj(cid:17) ⊆ Mi,
j6=i
and
for all i = 1, . . . , n. Thus, if we set
dimEi = dim(Si ⊖ TiSi) = multTiSi
(Si),
then E ⊆ F and
E = E1 ⊕ · · · ⊕ En,
dimE =
nX
i=1
multTiSi
(Si).
Fix i ∈ {1, . . . , n} and define (λ1, . . . , λn) ∈ Cn by λj = 0 if j = i and λj = αj if j 6= i. From
Lemma 4.1, it follows that
[PMiG]PMi
T Mi
= [PMiG](PMi
T1Mi −λ1IMi
,...,PMi
TnMi −λnIMi ).
For simplicity, we denote
Gi = [PMiG](PMi
T1Mi −λ1IMi
,...,PMi
TnMi −λnIMi ),
in the rest of this section. Also, notice that Cvj ⊥ ran(PQj TjQj − αjIQj ) for all j = 1, . . . , n,
and ranTiSi ⊥ Si ⊖ TiSi, so that
for all j = 1, . . . , n, and hence
On the other hand, since
PEi(PMi
TjMi − λjIMi) = 0,
PEiGi = PEi(span{G}).
j=1
and Ej ⊆ Mj for all j = 1, . . . , n, it follows that
PE =
PEj ,
nM
Hence
that is
and so
PEGi = PEiGi.
E = PEF = PE(cid:16) nM
i=1
[PMiG]PMi
T Mi(cid:17) =
nM
i=1
PEi[PMiG]PMi
T Mi
,
E =
nM
i=1
PEiGi =
nM
i=1
PEi(span{G}),
E = PE(span{G}).
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
15
From this it follows easily that
nX
i=1
dim(Si ⊖ TiSi) =
nX
i=1
multTiSi
(Si)
= dimE
≤ dim(span{G})
= dim(span{G})
= multPF T F (F ),
where the last equality follows from the minimality assumption on G. Therefore, Theorem
3.3 implies the following:
Theorem 4.2. Assume the setting of Theorem 3.3. If Si satisfies the generating wandering
subspace property with respect to TiSi and T ∗
i Qi has non-empty point spectrum for all i =
1, . . . , n, then
mult T S (S) ≥
nX
i=1
multTiSi
(Si) =
nX
i=1
dim(Si ⊖ TiSi).
To proceed further, we note, by Lemma 3.1 (or, more specifically (3.2)), that
S =
nX
i=1
ran Pi.
In addition, let us assume that multTi(Hi) = 1, i = 1, . . . , n. Then
mult T S (S) ≤
nX
i=1
multTiSi
(Si).
Therefore, by Theorem 4.2, we have the main theorem of this paper as:
Theorem 4.3. Let H1, . . . , Hn be Hilbert spaces, let Ti ∈ B(Hi), and let Qi be a T ∗
closed subspace of Hi, i = 1, . . . , n. Assume that TiQ⊥
wandering subspace property, T ∗
for all i = 1, . . . , n. Then
i -invariant
i ) satisfies the generating
i Qi has non-empty point spectrum and that multTi(Hi) = 1
∈ B(Q⊥
i
mult T (Q1⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ =
nX
i=1
multTiQ⊥
i
(Q⊥
i ).
5. Applications and Concluding Remarks
In this section, we complement the main theorem, Theorem 4.3, by some concrete examples
and final remarks.
We first explain the notion of zero-based invariant subspaces of reproducing kernel Hilbert
spaces. Let k : D × D → C be a positive definite kernel. For each fixed w ∈ D, let z 7→ k(z, w)
is analytic on D. Suppose Hk ⊆ O(D) is the reproducing kernel Hilbert space corresponding
16
CHATTOPADHYAY, SARKAR, AND SARKAR
to the kernel k and Mz, the multiplication operator by the coordinate function z, on Hk is
bounded. Let us further assume that
ker(M ∗
z − λIHk) = Ck(·, λ)
(λ ∈ D).
Here k(·, λ), for λ ∈ D, denotes the kernel function z 7→ k(z, λ) on D.
A reproducing kernel Hilbert space that satisfies all the properties listed above is called a
regular reproducing kernel Hilbert space.
It is easy to see that the Dirichlet space, the Hardy, the unweighted Bergman space and
the weighted Bergman spaces over D are regular reproducing kernel Hilbert spaces.
Suppose Hk is a regular reproducing kernel Hilbert space. A closed subspace S ⊆ Hk is
called zero-based invariant subspace if there exists λ ∈ D such that f (λ) = 0 for all f ∈ S
and zS ⊆ S.
Now let Hk be a regular reproducing kernel Hilbert space, and let Q be an M ∗
closed subspace of Hk. Suppose λ ∈ D. Then M ∗
if f = ck(·, λ) for some non-zero scalar c ∈ C. On the other hand, since
z -invariant
z f = ¯λf for some non-zero f ∈ Q if and only
hg, k(·, λ)i = g(λ)
(g ∈ Hk),
it follows that k(·, λ) ∈ Q if and only if g(λ) = 0 for all g ∈ Q⊥. We have therefore proved
the following:
Proposition 5.1. Let Hk be a regular reproducing kernel Hilbert space, and let Q be a closed
M ∗
z -invariant subspace of Hk. Then M ∗
z Q has non-empty point spectrum if and only if Q⊥ is
a zero-based invariant subspace of Hk.
As an immediate corollary of Theorem 4.3 we have now:
Corollary 5.2. Let Hki be a regular reproducing kernel Hilbert space, multMz (Hki) = 1, and
let Qi be a proper closed M ∗
is a zero-based
invariant subspace of Hki such that
z -invariant subspace of Hki, i = 1, . . . , n. If Q⊥
i
dim(Q⊥
i ⊖ zQ⊥
i ) < ∞,
for all i = 1, . . . , n, then
multMz(Q1⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ =
nX
i=1
(multMzQ⊥
i
(Q⊥
i )) =
nX
i=1
dim(Q⊥
i ⊖ zQ⊥
i ).
Now let Hki be the Hardy space or the Dirichlet space over D, and let Qi be a non-zero
shift co-invariant (that is, M ∗
satisfies
the generating wandering subspace property and the dimension of the generating wandering
subspace is one, that is
z -invariant) subspace of Hki. By [3] and [16], MzQ⊥
i
dim(Q⊥
i ⊖ zQ⊥
i ) = 1,
for all i = 1, . . . , n. Then, in view of Theorem 4.3 (and [19]) we have the following:
MULTIPLICITIES, INVARIANT SUBSPACES AND AN ADDITIVE FORMULA
17
Corollary 5.3. Let Hki, i = 1, . . . , n, denote either the Hardy space or the Dirichlet space
over D. Suppose Qi is a proper closed M ∗
is
a zero-based Mz-invariant subspace of Hki, i = 1, . . . , n, then,
z -invariant subspaces of Hki, i = 1, . . . , n. If Q⊥
i
multMz(Q1⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ = n.
A similar argument and the generating wandering subspace property of shift invariant
subspaces of the Bergman space [1] yields the following:
Corollary 5.4. Let Hki, i = 1, . . . , n, be the Dirichlet space, the Bergman space or the Hardy
space over D. Let Qi, i = 1, . . . , n, be proper closed shift co-invariant subspaces of Hki. If
Q⊥
i
is a zero based Mz-invariant subspace of Hki and
dim(Q⊥
i ⊖ zQ⊥
i ) < ∞,
for all i = 1, . . . , n, then
multMz(Q1⊗···⊗Qn)⊥ (Q1 ⊗ · · · ⊗ Qn)⊥ =
nX
i=1
(multMzQ⊥
i
(Q⊥
i )) =
nX
i=1
dim(Q⊥
i ⊖ zQ⊥
i ).
Note that the generating wandering subspace assumption in Corollary 5.4 ensures that (see
Proposition 2.1)
multMzQ⊥
i
(Q⊥
i ) < ∞,
for all i = 1, . . . , n. At present it is not very clear whether the generating wandering subspace
assumption can be replaced by finite multiplicity property. Our methods rely heavily on the
assumption that the invariant subspaces are zero-based and satisfies the generating wandering
subspace property.
Acknowledgement: The second author is supported in part by NBHM (National Board
of Higher Mathematics, India) grant NBHM/R.P.64/2014, and the Mathematical Research
Impact Centric Support (MATRICS) grant, File No : MTR/2017/000522, by the Science and
Engineering Research Board (SERB), Department of Science & Technology (DST), Govern-
ment of India. The third author is supported by the Department of Atomic Energy (DAE)
through the NBHM Postdoctoral fellowship and acknowledges Indian Statistical Institute,
Bangalore, for warm hospitality.
References
[1] A. Aleman, S. Richter and C. Sundberg, Beurling's theorem for the Bergman space, Acta Math. 177
(1996), 275-310.
[2] C. Apostol, H. Bercovici, C. Foias and C. Pearcy, Invariant subspaces, dilation theory, and the structure
of the predual of a dual algebra I, J. Funct. Anal. 63 (1985), 369404.
[3] A. Beurling, On two problems concerning linear transformations in Hilbert space, Acta Math. 81 (1949),
239-255.
[4] A. Chattopadhyay, B.K. Das and J. Sarkar, Rank of a co-doubly commuting submodule is 2, Proceedings
of American Math Society, 146 (2018), 1181-1187.
[5] A. Chattopadhyay, B.K. Das and J. Sarkar, Star-generating vectors of Rudin's quotient modules, J. Funct.
Anal. 267 (2014), 4341-4360.
18
CHATTOPADHYAY, SARKAR, AND SARKAR
[6] R. Douglas and R. Yang, Operator theory in the Hardy space over the bidisk (I), Integral Equations
Operator Theory 38 (2000), no. 2, 207-221.
[7] X. Fang, Additive invariants on the Hardy space over the polydisc, J. Funct. Anal. 253 (2007), 359-372.
[8] H. Hedenmalm, An invariant subspace of the Bergman space having the codimension two property, J.
reine angew. Math. 443 (1993), 19.
[9] H. Hedenmalm, B. Korenblum and K. Zhu, Theory of Bergman spaces, Grad. Texts Math. 199, Springer-
Verlag, New York 2000.
[10] H. Hedenmalm, S. Richter and K. Seip, Interpolating sequences and invariant subspaces of given index
in the Bergman spaces, J. reine angew. Math. 477 (1996), 1330.
[11] K. J. Izuchi, K. H. Izuchi and Y. Izuchi, Blaschke products and the rank of backward shift invariant
subspaces over the bidisk, J. Funct. Anal. 261 (2011), no. 6, 1457-1468.
[12] K. J. Izuchi, K. H. Izuchi and Y. Izuchi, Ranks of invariant subspaces of the Hardy space over the bidisk,
J. Reine Angew. Math. 659 (2011) 101-139.
[13] K. J. Izuchi, K. H. Izuchi and Y. Izuchi, Ranks of backward shift invariant subspaces of the Hardy space
over the bidisk, Math. Z. 274 (2013), 885-903.
[14] G. Popescu, Invariant subspaces and operator model theory on noncommutative varieties, Math. Ann.
372 (2018), 611-650.
[15] G. Popescu, Euler characteristic on noncommutative polyballs, J. Reine Angew. Math. 728 (2017), 195-
236.
[16] S. Richter, Invariant subspaces of the Dirichlet shift, J. Reine Angew. Math. 386 (1988), 205-220.
[17] W. Rudin, Function Theory in Polydiscs, Benjamin, New York 1969.
[18] J. Sarkar, Jordan blocks of H 2(Dn), J. Operator theory 72 (2014), 371-385. .
[19] A. Tomerlin, Products of Nevanlinna-Pick kernels and operator colligations, Integral Equations Operator
Theory 38 (2000), no. 3, 350-356.
Department of Mathematics, Indian Institute of Technology Guwahati, Guwahati, 781039,
India
E-mail address: [email protected], [email protected]
Indian Statistical Institute, Statistics and Mathematics Unit, 8th Mile, Mysore Road,
Bangalore, 560059, India
E-mail address: [email protected], [email protected]
Department of Mathematics, Indian Institute of Science, Bangalore, 560012, India
E-mail address: [email protected], [email protected]
|
1504.05841 | 2 | 1504 | 2015-08-15T10:05:15 | An example of a non-commutative uniform Banach group | [
"math.FA",
"math.GN",
"math.GR"
] | Benyamini and Lindenstrauss mention in their monograph \emph{Geometric nonlinear functional analysis Vol. 1., American Mathematical Society Colloquium Publications, 48. American Mathematical Society, Providence, RI, 2000} that there is no known example of a non-commutative uniform Banach group. Prassidis and Weston also asked whether there is a non-commutative example. We answer this problem affirmatively. We construct a non-commutative uniform Banach group which has the free group of countably many generators as a dense subgroup.
Moreover, we show that our example is a free one-generated uniform Banach group whose metric induced by the norm is bi-invariant. | math.FA | math |
AN EXAMPLE OF A NON-COMMUTATIVE UNIFORM
BANACH GROUP
MICHAL DOUCHA
Abstract. Benyamini and Lindenstrauss mention in their mono-
graph Geometric nonlinear functional analysis Vol. 1., Ameri-
can Mathematical Society Colloquium Publications, 48. American
Mathematical Society, Providence, RI, 2000 that there is no known
example of a non-commutative uniform Banach group. Prassidis
and Weston also asked whether there is a non-commutative ex-
ample. We answer this problem affirmatively. We construct a
non-commutative uniform Banach group which has the free group
of countably many generators as a dense subgroup.
Moreover, we show that our example is a free one-generated
uniform Banach group whose metric induced by the norm is bi-
invariant.
Introduction
A uniform Banach group is a Banach space equipped with an addi-
tional group structure so that the group unit coincides with the Banach
space zero and the group operations are uniformly continuous with re-
spect to norm. Uniform Banach groups were introduced and studied
by Enflo in [4], [5] with connection to the infinite-dimensional version
of the Hilbert's fifth problem. Typical example comes when we are
given two Banach spaces X and Y and a uniform homeomorphism
φ : X → Y between them such that φ(0) = 0. Then we can define a
(commutative) group operation · on X as follows: for x, y ∈ X we set
x · y = φ−1(φ(x) + φ(y)). Note that unless φ is linear there is no a priori
connection between the two group operations + (resp. +X) and ·.
A comprehensive source of information about uniform Banach groups
is Chapter 17 in [2]. As mentioned there, the following problem was
left open. Does there exist a non-commutative uniform Banach group?
This question was also asked by Prassidis and Weston in [9] and [10].
2010 Mathematics Subject Classification. 46B20,22A05.
Key words and phrases. uniform Banach group, free group, Hilbert's fifth prob-
lem, Lipschitz-free space.
The author was supported by IMPAN's international fellowship programme par-
tially sponsored by PCOFUND-GA-2012-600415.
1
2
M. DOUCHA
Here we give a positive answer to this question. The following is the
main result.
Theorem 0.1. There exists an infinite dimensional separable Banach
space (X, +, 0, k·k) equipped with an additional group structure (·, −1, 0)
whose unit coincides with the Banach space zero, the group multiplica-
tion · is invariant with respect to the norm k · k, and F∞, the free group
of countably many generators, is a dense subgroup of (X, ·, −1, 0).
In particular, there exists a non-commutative uniform Banach group.
1. Preliminaries
We assume the reader to know basic facts about uniform spaces. We
refer to Chapter 8 in [6] for more information.
Recall that for any topological group G there are two distinguished
compatible uniformities: the left uniformity UL generated by basic en-
tourages of the form {(g, h) : g−1h ∈ U}, where U is a basic neigh-
borhood of the identity in G; and the right uniformity UR which is
generated by basic entourages of the form {(g, h) : hg−1 ∈ U}, where
U is again a basic neighborhood of the identity in G.
We start with the following definition given by Enflo in [4].
Definition 1.1. Let G be a topological group. G is called uniform if
there exists a compatible uniformity on G such that the group multi-
plication is uniformly continuous with respect to that uniformity.
Below, we collect some basic facts about uniform groups.
Fact 1.2.
(1) (see Proposition 1.1.3. in [4]) A topological group G is uniform
if and only if the left and right uniformities coincide.
(2) (folklore) That is in turn equivalent with the fact that there ex-
ists a neighborhood basis of the unit of G consisting of open sets
closed under conjugation. Such groups are more often called
SIN (small invariant neighborhood) groups, or also balanced
groups.
(3) (folklore) In case that G is metrizable, i.e. the neighborhood ba-
sis can be taken countable, G is uniform, resp. SIN, if and only
if it admits a compatible bi-invariant metric; i.e. metric d such
that for any x, y, a, b ∈ G it holds that d(x, y) = d(axb, ayb)
(the same reasoning gives that if G is not metrizable then its
topology is given by a family of bi-invariant pseudometrics).
NON-COMMUTATIVE UNIFORM BANACH GROUP
3
Examples:
(a) All abelian and compact topological groups are uniform groups.
It is obvious for the former. For the latter, consider a compact
topological group G. Notice that by the continuity of the group
operations, for any open neighborhood U of the identity and for
any group element g there are open neighborhoods Vg of g and
Wg ⊆ U of the identity so that Vg · Wg · V −1
g ⊆ U. Then by
compactness one can find finitely many elements g1, . . . , gn ∈ G so
that Vg1, . . . , Vgn cover G. Take W = Ti≤n Wgi and notice that for
any g ∈ G we have g · W · g−1 ⊆ U. Clearly, Sg∈G g · W · g−1 is then
a conjugacy-invariant open neighborhood of the identity contained
in U.
(b) The Heisenberg group U T 3
3 (R) consisting of the upper triangular
3 × 3-matrices is not uniform. Note that the Heisenberg group
is very close to both being abelian and compact.
It is a locally
compact group of nilpotency class 2. Since U T 3
3 (R) is metrizable
it is sufficient to show that it does not admit a compatible bi-
invariant metric. Suppose for contradicition that d is a compatible
bi-invariant metric on U T 3
3 (R). Given matrices
1 a c
0 1 b
, a computation gives that
0 0 1
A′ =
1 a′
c′
0 1 b′
0 0
1
, A =
A−1A′A =
1 a′
0 1
0 0
c′ − ab′ + a′b
.
b′
1
Since d is compatible there exist r > 0 and R > 0 such that the
open ball of radius R with respect to d centred at I, the identity
matrix, is contained in the open set of matrices
1 x z
0 1 y
0 0 1
with
z < r. Since d is compatible, we can choose A′ with a′
6= 0
and d(A′, I) < R. Then by choosing b sufficiently large we can
guarantee that c′−ab′+a′b > r. It follows that d(A−1A′A, I) ≥ R.
So d(A′, I) 6= d(A−1A′A, I), and thus d is not bi-invariant.
(c) However, an example of a uniform group that is finite-dimensional
Euclidean is the group of unitary n×n-matrices. That follows either
from the fact that this group is compact, or one can check that
the Hilbert-Schmidt norm there induces a compatible bi-invariant
metric.
(d) Let G be a group that acts faithfully by isometries on a bounded
metric space (M, dM ). The following is then a bi-invariant metric
4
M. DOUCHA
(that makes G a topological uniform group): for g, h ∈ G set
d(g, h) = sup
x∈M
dM (gx, hx).
Every group with bi-invariant metric is of this form. Indeed, if
G is a group with bi-invariant metric dG, then G has a faithfull
action by isometries on itself induced by left translations. The
formula above defining a metric gives the original metric dG.
Definition 1.3 (Enflo, [5]). A uniform group G is Banach if it is
uniformly homeomorphic to some Banach space.
The preceding definition is readily checked to be equivalent with that
one mentioned in the introduction; i.e. a Banach space with additional
group structure, where the additional group operations are uniformly
continuous with respect to norm, and the additional group unit coin-
cides with the Banach space 0.
Recall the canonical example of a uniform Banach group given by
some uniform homeomorphism between Banach spaces that was also
mentioned in the introduction. Let us mention here the result of Enflo
which says that under some conditions the converse is true. We refer to
[4] and [5] for unexplained notions from the statement of the theorem.
Theorem 1.4 (Enflo, [5]). Let G be a commutative uniform Banach
group such that the corresponding Banach space has roundness p > 1
and such that moreover:
• G is uniformly dissipative,
• for every x1, x2, y ∈ G we have kx1 · y − x2 · yk = o(kx1 − x2k1/p)
uniformly in x1, x2, y as kx1 − x2k → 0.
Then G is isomorphic to the additive group of some Banach space.
1.1. Preliminary discussion of the proof. Before proceding to the
proof of Theorem 0.1, let us roughly explain the ideas behind it. We
shall construct a countable set X equipped with two group structures.
Under the first group structure, X is isomorphic to the free group of
countably many generators. Under the second one, X is isomorphic to
the minimal subgroup of the real vector space with countable Hamel
basis that contains the elements of this Hamel basis and is closed under
multiplication by scalars that are dyadic rationals. Moreover, X will
get equipped with a metric which is bi-invariant with respect to the
non-commutative group operation and which behaves like a norm with
respect to the latter group operation. In particular, the completion of
X with respect to this metric will become a real Banach space.
Let us mention two 'peculiarities' of the construction. First, X is
constructed so that there is no connection between these two algebraic
NON-COMMUTATIVE UNIFORM BANACH GROUP
5
structures (i.e. non-commutative group structure and commutative
group structure with dyadic scalar multiplication) in the sense that for
any x, y ∈ X, the elements x, y, x · y, x−1 are linearly independent (in
the commutative structure), and similarly, for any x, y ∈ X and dyadic
rationals α, β we have that αx + βy is a new free group generator (in
the non-commutative structure).
The second thing to mention is that the metric on X is constructed
by induction on finite fragments Xn, n ∈ N, of X. This will give us a
better control of the metric we are defining.
The construction is inspired and in some sense analogous to the
construction of Lipschitz-free Banach spaces (we refer the reader to
the book [11] as the main reference on this subject) and also to the
construction of free groups with the Graev metric (see [7]). Indeed, the
uniform Banach group we construct can be viewed as a free object in an
appropriate category and we refer the reader to the last section of the
paper where this is discussed. We also discuss there in more detail the
similarity between our Banach group and Lipschitz-free Banach spaces
and groups with the Graev metric.
1.2. Some notation regarding free groups and vector spaces.
Let S be some non-empty set and let w be a word over the alphabet
S ∪ S−1 ∪ {e}, where S−1 = {s−1 : s ∈ S} is a disjoint copy of S
interpreted as a set of inverses of elements from S and e is an element
not belonging to either S or S−1 which shall be interpreted as a group
unit. We say that w = w1 . . . wn is irreducible if either n = 1 and
w = w1 = e, or n > 1 and for every i ≤ n, wi 6= e and there is no
i < n such that wi = w−1
If w is a word that is not irreducible,
i+1.
then by w′ we shall denote the unique irreducible word obtained from
w by deleting each occurence of the letter e and each occurence of
neighbouring letters a and a−1 (if this procedure leads to an empty
word, then we set w′ to be e).
It is well-known and easy to observe that elements of F (S), the free
group of free generators coming from the set S with e as a unit, are
in one-to-one correspondence with irreducible words over the alphabet
S ∪ S−1 ∪ {e}.
Let n ∈ N. By Wn(S) we shall denote the set of all irreducible words
of length at most n over the alphabet S ∪ S−1 ∪ {e}.
Let now similarly B be some non-empty set not containing the dis-
tinguished element e. The vector space over some field F with B as
the maximal linearly independent set and e representing zero can be
viewed as a set of all functions from B to K that have finite support,
6
M. DOUCHA
where f has finite support if for all but finitely many b ∈ B we have
f (b) = 0.
In our case, we shall work with F = R (resp. Q), however since we
shall need to work with only finitely many vectors at any given time
we restrict to functions whose range is some specified finite subset of
R (also, the set B will be always finite).
Let K ⊆fin R some finite subset of reals. Then by VK(B) we shall
denote the set of all functions from B to K with finite support. The
requirement on finite support will be in our construction superfluous
since B will be at any given time finite as already mentioned.
2. The proof of Theorem 0.1
We start by describing the underlying countable set X mentioned
above in the preliminaries.
2.1. The underlying dense set. We now describe a countably infi-
nite set X, constructed as an increasing union of finite sets X0 ⊆ X1 ⊆
. . ., which will also carry a multiplicative group operation and the cor-
responding group inverse operation so that X will be isomorphic to a
free group of countably many generators, and it will also carry an addi-
tive (abelian) group operation together with multiplication by scalars
that are dyadic rationals so that it is a proper subgroup of an infinite
dimensional vector space over the rationals. Moreover, the unit for
addition and multiplication will be the same.
Later, we shall define a metric on X that will be invariant under
both addition and multiplication and will preserve scalar multiplica-
tion. The completion then will be a Banach space over the reals which
is also equipped with multiplication with free group as a dense part.
rationals a
all dyadic rational numbers.
Let n ≥ 1 be arbitrary. By Dn we shall denote the set of dyadic
2n , where a ∈ [−22n, 22n]. Clearly, D = Sn Dn is the set of
We set X0 = {e}. Let S1 = {x} be some singleton. We set
X1 = W1(S1) = {e, x, x−1}. Let B2 = X1 \ X0 = {x, x−1}. We set
X2 = VD1(B2) = {αx + βx−1 : α, β ∈ D1}.
Suppose now that X2n has been constructed. We need to construct
X2n+1 and X2n+2. Set S2n+1 to be S2n−1 ∪ (X2n \ X2n−1). Then we set
X2n+1 to be W2n+1(S2n+1).
Next we set B2n+2 to be B2n ∪ (X2n+1 \ X2n). Then we set X2n+2 to
be VDn+1(B2n+2).
NON-COMMUTATIVE UNIFORM BANACH GROUP
7
This finishes the inductive construction. Note that X = Sn Xn =
Sn W2n+1(S2n+1) = Sn VDn+1(B2n+2).
It follows that if we set S =
Sn S2n+1 and B = Sn B2n+2, then X is also naturally isomorphic to
F (S), the free group of countably many generators coming from the set
S, and it is a (additive) proper subgroup of the rational (or real) vector
space with B as the maximal linearly independent set - the minimal
subgroup that contains free abelian group with B as a set of generators
that is closed under multiplication by scalars from D.
We define inductively a rank function r : X → ω. For any x ∈ X we
set r(x) = 0 if neither x nor x−1 is possible to write as α1y1+. . .+αmym,
where m > 1, αi > 0, for every i, and α1 + . . . + αm = 1.
If x or x−1 is possible to write as α1y1 + . . . + αmym, where m > 1,
αi > 0, for every i, and α1+. . .+αm = 1, then we set r(x) = max{r(yi) :
i ≤ m} + 1.
2.2. Construction of the metric. We shall now define a metric ρ
and a norm k · k on X. Actually, the metric and the norm will be
one and the same in the sense that for any x, y ∈ X we shall have
ρ(x, y) = kx − yk and kxk = ρ(x, e). The distinguishing is done only
for practical notational reasons since we understand X as both a free
group and a subgroup of a vector space. In the former case, it is more
natural to consider a metric there, while in the latter to consider a
norm there. By induction, we shall define functions ρn : X 2
n → R (the
range will actually be a subset of non-negative rationals) for odd n
and functions k · kn : Xn → R for even n that satisfy the following
properties:
(1) for every odd n we have that ρn is a symmetric function that
is equal to zero only on diagonal, i.e. ρn(x, y) = 0 iff x = y for
x, y ∈ Xn; similarly, for every even n we have that kxkn = 0 iff
x = e for x ∈ Xn,
(2) for every even n, k · kn extends ρn−1, i.e. for every a, b ∈ Xn−1
we have
ka − bkn = ρn−1(a, b),
similarly for every odd n, ρn extends k · kn−1, i.e.
a, b ∈ Xn−1 such that a − b ∈ Xn−1 we have
for every
ρn(a, b) = ka − bkn−1,
(3) for every odd n, for any words (not necessarily irreducible)
w1, w2, v1, v2 over the alphabet Sn ∪ S−1
n ∪ {e} such that
8
M. DOUCHA
2, (w1w2)′, (v1v2)′ ∈ Xn, we have
w′
1, v′
1, w′
2, v′
ρn((w1w2)′, (v1v2)′) ≤ ρn(w′
and for every a, b ∈ Xn we have
1, v′
1) + ρn(w′
2, v′
2),
ρn(a, b) = ρn(a−1, b−1),
(4) for every odd n and for every a ∈ Xn and for every b ∈ Xn such
that b = α1c1 + . . . + αmcm, where m > 1, αi ≥ 0 and ci ∈ Bn−1,
for every i, and α1 + . . . + αm = 1, we have
ρn(a, b) ≤ α1 · ρn(a, c1) + . . . + αm · ρn(a, cm),
(5) for every even n and for every a ∈ Xn and any scalar α such
that αa ∈ Xn we have
kαakn = α · kakn,
and for every a, b ∈ Xn such that also a + b ∈ Xn we have
ka + bkn ≤ kakn + kbkn,
(6) for every odd n, for every a ∈ Xn and for every b ∈ Xn such
that b = (α1c1 + . . . + αmcm)−1, where, m > 1, αi ≥ 0 and
ci ∈ Bn−1, for every i, and α1 + . . . + αm = 1, we have
ρn(a, b) ≤ α1 · ρn(a, c−1
1 ) + . . . + αm · ρn(a, c−1
m ),
and similarly, for every even n, for every a ∈ Xn and for every
b ∈ Xn−1 such that b = (α1c1 + . . . + αmcm)−1, where m > 1,
αi ≥ 0 and ci ∈ Bn−2, for every i, and α1 + . . . + αm = 1, we
have
ka − bkn ≤ α1 · ka − c−1
1 kn + . . . + αm · ka − c−1
m kn.
We define ρ1 on X1 by setting ρ1(x, e) = ρ1(x−1, e) = 1 and ρ1(x, x−1) =
2. Obviously, this satisfies all the requirements.
Let us now define k · k2 on X2, i.e. we have to define kαx + βx−1k2
for every α, β ∈ D1. We set
kαx + βx−1k2 = min{γ1 · ρ1(x, e) + γ2 · ρ1(x−1, e) + γ3 · ρ1(x, x−1) :
αx + βx−1 = γ1x + γ2x−1 + γ3(x − x−1), γ1, γ2, γ3 ∈ R}.
Condition (6) is automatically satisfied as there is no b ∈ X1 of the
form (α1c1 + . . . + αmcm)−1 for appropriate α's and c's. Condition (1)
is also obvious. Thus we need to check the conditions (2) and (5).
Let us do the former. We need to check that k · k2 extends ρ1.
We shall check that kxk2 = ρ1(x, e). The other cases are similar.
Clearly, kxk2 ≤ ρ1(x, e). Suppose that kxk2 = γ1 · ρ1(x, e) + γ2 ·
NON-COMMUTATIVE UNIFORM BANACH GROUP
9
ρ1(x−1, e) + γ3 · ρ1(x, x−1), where x = γ1x + γ2x−1 + γ3(x − x−1). It
follows that necessarily γ2 = γ3 and γ1 + γ3 = 1. By triangle inequality
we have γ2 · ρ1(x−1, e) + γ2 = γ3 · ρ1(x, x−1) ≥ γ2 · ρ1(x, e), thus
γ1 · ρ1(x, e) + γ2 · ρ1(x−1, e) + γ3 · ρ1(x, x−1) ≥ γ1 · ρ1(x, e) + γ2 ·
ρ(x, e) ≥ γ1 + γ2 · ρ1(x, e) = ρ1(x, e).
We now check (5). Fix some a ∈ X2 and α such that αa ∈ X2.
Suppose that kak = γ1 · ρ1(x, e) + γ2 · ρ1(x−1, e) + γ3 · ρ1(x, x−1),
where we have a = γ1x + γ2x−1 + γ3(x − x−1). Then since αa =
α · γ1x+α · γ2x−1 +α · γ3(x−x−1) we have that kαak2 ≤ αγ1· ρ1(x, e)+
αγ2 · ρ1(x−1, e) + αγ3 · ρ1(x, x−1) = α · kak2. The other inequality
is analogous. By a similar argument one can also show that for any
a, b ∈ X2 such that also a + b ∈ X2 we have ka + bk2 ≤ kak2 + kbk2 and
we leave this to the reader. In fact, we note here that this definition
of k · k2 is equivalent with that one that says that k · k2 is the great-
est function that satisfies condition (4) and such that kxk2 ≤ ρ1(x, e),
kx−1k2 ≤ ρ1(x−1, e) and kx − x−1k2 ≤ ρ1(x, x−1).
Extending the metric.
Suppose we have defined k · kn for some even n ≥ 2. We now define
ρn+1 on Xn+1. First we inductively define an auxiliary function δ on
X 2
n+1. Fix a pair x, y ∈ Xn+1. If r(x) = r(y) = 0, then we set
δ(x, y) = min{ka1 − b1kn + . . . + kam − bmkn :
x = (a1 . . . am)′, y = (b1 . . . bm)′, ∀i ≤ m(ai, bi, ai − bi ∈ Xn)}.
Note that the minimum is indeed attained as Xn is finite. Note again
that for any z ∈ Xn+1 if r(z) > 0 we have that either z or z−1 belongs
to Xn. So if r(x) > 0, r(y) > 0, then we set
δ(x, y) =
kx − ykn
kx−1 − y−1kn
min{kx − zk + kz−1 − y−1k : z, z−1 ∈ Xn} if x, y−1 ∈ Xn
min{kx−1 − z−1k + kz − yk : z, z−1 ∈ Xn} if x−1, y ∈ Xn.
if x, y ∈ Xn
if x−1, y−1 ∈ Xn
Now we suppose that for one of the elements, say x, we have r(x) = 0,
and for the other one we have r(y) > 0. The following is done by
induction on r(y). First suppose that y = α1z1 + . . . + αmzm, where
m > 1, αi > 0, for all i, and α1 + . . . + αm = 1. Then we set
δ(x, y) = α1 · δ(x, z1) + . . . + αm · δ(x, zm).
Similarly, if y = (α1z1 + . . . + αmzm)−1, where m > 1, αi > 0, for all i,
and α1 + . . . + αm = 1, then we set
δ(x, y) = α1 · δ(x, z−1
1 ) + . . . + αm · δ(x, z−1
m ).
10
M. DOUCHA
We are now ready to define ρn+1. Thus fix now again a pair x, y ∈
Xn+1 and set
ρn+1(x, y) = min{δ(a1, b1) + . . . + δ(am, bm) :
x = (a1 . . . am)′, y = (b1 . . . bm)′, a1, . . . , am, b1, . . . , bm ∈ Xn+1}.
First notice that since Xn+1 is finite the minimum is attained, thus
in particular we have for x 6= y that ρn+1(x, y) > 0. Since ρn+1 is
clearly symmetric we get it satisfies the condition (1).
We claim that ρn+1 is the greatest function satisfying:
(a) ρn+1(x, y) ≤ kx − ykn for every x, y ∈ Xn such that x − y ∈ Xn,
(b) ρn+1(x, y) = ρn+1(x−1, y−1) for every x, y ∈ Xn+1,
(c) ρn+1(ab, cd) ≤ ρn+1(a, c) + ρn+1(b, d) for every a, b, c, d ∈ Xn+1 such
that ab, cd ∈ Xn+1,
(d) ρn+1(x, (α1z1+. . .+αmzm)) ≤ α1·ρn+1(x, z1)+. . .+αm·ρn+1(x, zm),
where m > 1, αi > 0, for all i, α1 + . . . + αm = 1 and α1z1 + . . . +
αmzm ∈ Xn+1,
(e) ρn+1(x, (α1z1 + . . . + αmzm)−1) ≤ α1 · ρn+1(x, z−1
1 ) + . . . + αm ·
m ), where m > 1, αi > 0, for all i, α1 + . . . + αm = 1 and
ρn+1(x, z−1
(α1z1 + . . . + αmzm)−1 ∈ Xn+1.
First of all, it is clear from the definitions of ρn+1 (and of δ) that
ρn+1 satisfies all these conditions. Thus in particular, we get that ρn+1
satisfies the conditions (3),(4) and (6). Next, if ξ is any other function
satisfying conditions (a)-(e), then it is readily checked that ξ ≤ δ and
because of (c) also ξ ≤ ρn+1.
n ⊆ Xn be such that for every x, y ∈ X ′
We shall now conclude from that that ρn+1 also satisfies (2). Indeed,
let X ′
n we have x − y ∈ Xn.
Then consider the metric ξ on X ′
n defined as ξ(x, y) = kx − ykn. We
claim it satisfies the conditions (a)-(e) above. Condition (a) is satisfied
since ξ(x, y) = kx − ykn for appropriate x, y. Take some x, y ∈ X ′
n
such that x−1, y−1 ∈ X ′
n. Necessarily x, y, x−1, y−1 ∈ Xn−1 and since
k · kn extends ρn−1 we get ξ(x, y) = ρn−1(x, y) = ρn−1(x−1, y−1) =
ξ(x−1, y−1). Take now some a, b, c, d ∈ X ′
n such that ab, cd ∈ X ′
n.
We again necessarily have that a, b, c, d, ab, cd ∈ Xn−1 and since k · kn
extends ρn−1 we again obtain ξ(ab, cd) = ρn−1(ab, cd) ≤ ρn−1(a, c) +
ρn−1(b, d) = ξ(a, c) + ξ(b, d). We have verified conditions (b) and (c).
Condition (d) follows since k·kn satisfies the condition (5) further above.
Finally, take some x, (α1z1 + . . .+ αmzm)−1 ∈ X ′
n, where m > 1, αi > 0,
for all i, α1 + . . . + αm = 1. We have ξ(x, (α1z1 + . . . + αmzm)−1) =
kx − (α1z1 + . . . + αmzm)−1kn ≤ α1 · kx − z−1
1 kn + . . . + αm · kx − zmkn =
NON-COMMUTATIVE UNIFORM BANACH GROUP
11
α1 · ξ(x, z1) + . . . + αm · ξ(x, zm), where the middle inequality follows
from the property (6) above.
We thus get kx − ykn = ξ(x, y) ≤ ρn+1(x, y) for every x, y ∈ Xn such
that x − y ∈ Xn. Since by assumption ρn+1(x, y) ≤ kx − ykn we get
ρn+1(x, y) = kx − ykn and we are done.
Extending the norm.
Now suppose we have defined ρn on Xn for some odd n > 2. We define
k · kn+1 on Xn+1. We again at first define, inductively, an auxiliary
function γ : Xn+1 → R. First, for every x, y ∈ Xn we set
γ(x − y) = ρn(x, y).
Next, for every x, y ∈ Xn+1 such that x − y ∈ Xn+1 and x ∈ Xn+1 \ Xn
we define γ(x − y) by induction on r(y). If r(y) = 0 then we set
γ(x − y) = min{β1 · γ(v1) + . . . + βi · γ(vi) :
x − y = β1v1 + . . . + βivi, ∀j ≤ i∃aj , bj ∈ Xn(vj = aj − bj)}.
If r(y) > 0 and y = (α1z1 + . . . + αmzm)−1, where m > 1, αi > 0, for
every i, and α1 + . . . + αm = 1, then we set
γ(x − y) = α1 · γ(x − z−1
Finally, for any x ∈ Xn+1 we set
1 ) + . . . + αm · γ(x − z−1
m ).
kxkn+1 = min{γ(y1) + . . . + γ(yi) : x = y1 + . . . + yi, y1, . . . , yi ∈ Xn+1}.
First thing to observe is again that for any x ∈ Xn+1 we have
It follows that condition (1) is satisfied for
kxkn+1 = 0 iff x = e.
k · kn+1.
Next we claim that k · kn+1 is the greatest function satisfying:
(a) kx − ykn+1 ≤ ρn(x, y) for every x, y ∈ Xn,
(b) kαx + βykn+1 ≤ kαxkn+1 + kβykn+1 = α · kxkn+1 + β · kykn+1
for x, y ∈ Xn+1 such that also αx + βy ∈ Xn+1,
(c) kx − (α1z1 + . . . + αmzm)−1kn+1 ≤ α1 · kx − z−1
1 kn+1 + . . . + αm · kx −
z−1
m kn+1, where m > 1, αi > 0, for every i, and α1 + . . . + αm = 1.
First, it follows from the definitions of γ and k · kn+1 that k · kn+1
satisfies all these conditions, thus we have that it satisfies the condi-
tions (5) and (6) further above. If ξ is any other functions satisfying
conditions (a)-(c) then it again follows that necessarily ξ ≤ γ and thus
also ξ ≤ k · kn+1.
We are ready to verify the remaining condition (2) for k · kn+1 that it
extends ρn. Let X ′
n+1 ⊆ Xn+1 be the set {x − y : x, y ∈ Xn}. Define ξ
on X ′
n+1. Then ξ
satisfies (a) since ξ(x−y) = ρn(x, y) for appropriate x, y. Next we check
n+1 as follows: ξ(x − y) = ρn(x, y) for every x − y ∈ X ′
12
M. DOUCHA
(b). Take some αx + βy. If α = 1 and β = −1 (or vice versa), then we
have ξ(x − y) = ρn(x, y) ≤ ρn(x, e) + ρn(e, y) = ξ(x) + ξ(y), where the
middle inequality follows from the condition (3) that ρn satisfies (which
implies the triangle inequality). Finally, for (α1z1 + . . . + αmzm)−1,
where m > 1, αi > 0, for every i, and α1 + . . . + αm = 1, we
have ξ(x − (α1z1 + . . . + αmzm)−1) = ρn(x, (α1z1 + . . . + αmzm)−1) ≤
α1·ρn(x, z−1
m ),
since ρn satisfies the condition (6), and we are done.
1 )+. . .+αm·ρn(x, z−1
m ) = α1·ξ(x−z−1
1 )+. . .+αm·ξ(x−z−1
Now we can define k · k on X by putting
k · k = [
i
k · ki
and analogously we can define ρ on X by putting
ρ = [
i
ρi.
By the condition (2) we have that for any x, y ∈ X we have kx−yk =
ρ(x, y). By (1) we have that for x 6= y ∈ X we have ρ(x, y) > 0 and
equivalently, for any x 6= e ∈ X we have kxk > 0.
Let us check that ρ is a bi-invariant metric on X when considered as
a (free) group. By (1) is symmetric. We use the following simple fact.
Fact 2.1. Let G be a group equipped with a symmetric function d :
G2 → R+
0 that is equal to 0 only on diagonal. Then d is a bi-invariant
metric if and only if for every x, y, v, w ∈ G we have d(x · y, v · w) ≤
d(x, v) + d(y, w).
Proof. If d is a bi-invariant metric then the inequality readily follows
from bi-invariance and using triangle inequality.
So suppose that d satisfies such an inequality for every x, y, v, w ∈ G.
Fix a, b, c ∈ G. Then d(a, c) = d(a · b−1 · b, b · b−1 · c) ≤ d(a, b) +
d(b−1, b−1)+d(b, c) = d(a, b)+d(b, c), so d is a metric. Now d(a·b, a·c) ≤
d(a, a)+d(b, c) = d(b, c) = d(a−1 ·a·b, a−1 ·a·c) ≤ d(a−1, a−1)+d(a·b, a·
c) = d(a · b, a · c) which shows the left-invariance. The right invariance
is done analogously.
(cid:3)
However, ρ does satisfy the condition from the statement of the fact
since it satisfies the condition (3).
Similarly, for every x, y and α, β ∈ D (recall that D denotes the
dyadic rationals) we have kαx+βyk ≤ kαxk+kβyk = α·kxk+β·kyk
since k · k satisfies the condition (5).
NON-COMMUTATIVE UNIFORM BANACH GROUP
13
Denote now by X the completion of X with respect to ρ, or equiva-
lently, with respect to k · k. Both the multiplicative and additive group
operations extend to the completion as well as the scalar multiplication
by dyadic rationals. Moreover, since the dyadic rationals are dense in
R, X has well-defined scalar multiplication by all the reals. Thus X is
a Banach space.
3. Concluding discussion and questions
Let us comment on similarities between the constructions of the
presented Banach group and Lipschitz-free Banach spaces. For any
pointed metric space (X, 0) there is a Banach space LF (X), called
Lipschitz-free space (or Arens-Eells space) over (X, 0), that has ele-
ments of X \ {0} as the Hamel basis and the point 0 as the Banach
space zero. LF (X) contains X isometrically and the norm of LF (X)
is uniquely described as the largest norm on the linear span of X that
extends the metric on X.
By similar means, one can define the largest bi-invariant metric on
FX\{0}, the free group having 0 as a neutral element and X \ {0} as the
set of free generators, that extends the metric on X.
The following generalization of the Lipschitz-free construction was
given in [1]. If Y is a Banach space and Y ⊆ X is a metric extension of
Y such that for every x ∈ X the distance function dist(x, ·) : Y → R is
convex, then there is a free space LFY (X) which has Y as a subspace,
contains X as a subset so that the elements of X \ Y are linearly
independent, and the norm on LFY (X) is the largest norm on such a
vector space that extends the metric on X.
An analogous generalization of the Graev metric on groups was given
for any group G with bi-invariant metric and any metric
in [3]; i.e.
extension G ⊆ X such that G is closed in X there exists the largest
bi-invariant metric on the free product G ∗ FX\G.
Roughly speaking, the idea behind our construction from the present
paper was to alternatively use those two constructions. That means,
to start with some pointed metric space (X, 0). Then consider the
Lipschitz-free space LF (X), then the free group FLF (X) with the Graev
metric, then LFLF (X)(FLF (X)), etc., and at the end to take the comple-
tion. The reason why this approach does not literally work is that for a
general point x in FLF (X), the distance function dist(x, ·) : LF (X) → R
is not convex. The remedy was to use 'finitary' versions of the con-
structions above where we were able to guarantee the convexity of the
appropriate distance functions.
14
M. DOUCHA
3.1. Freeness of the Banach group. We shall argue that although
X as a Banach space is likely not one of the "well-behaved" spaces it
is not completely random either. It can be uniquely characterized as
a free one-generated uniform Banach group whose metric induced by
the norm is bi-invariant. That can be described via certain universal
property. Let us at first define morphisms in the category of uniform
Banach groups. The natural definition seems to be a bounded linear
operator that is moreover a group homomorphism in the additional
group structure. Note that it is then automatically uniformly continu-
ous group homomorphism.
We claim that:
Theorem 3.1. X is the free uniform Banach group over one generator
whose metric induced by the norm is bi-invariant. That means, for
any uniform Banach group Y whose metric induced by the norm is bi-
invariant, and any element y ∈ Y there exists a unique uniform Banach
group morphism φ : X → Y such that φ(x) = y and kφk = kykY.
This property characterizes X uniquely up to isomorphism.
Remark 3.2. Note that every uniform Banach group Z admits a bi-
invariant metric that is uniformly continuous with respect to norm.
For any y, z ∈ Z set D(y, z) = supv,w∈Z kv · y · w − v · z · wkZ. It is
always satisfied that D(y, z) ≥ ky − zk and the inequality may be strict
in general.
Proof. If y = 0 then φ is the zero morphism and there is nothing to
prove. So we assume that y 6= 0.
Let us take a closer look on the countable dense set X ⊆ X. One
can see that each element of X is obtained in a unique way from the
element x (the only element of S1) using operations of addition +,
multiplication ·, additive and multiplicative inverses −, −1 and by scalar
multiplication by dyadic rationals D. It follows that there is a unique
map φ′ : X → Y that preserves those operations and satisfies φ′(x) = y.
We need to show that for every z ∈ X we have
(3.1)
kφ′(z)kY ≤ kykYkzkX.
Once this is proved we can (uniquely) extend φ′ to the completion of X
to obtain φ ⊇ φ′ : X → Y. φ is still linear and preserves the additional
group structure. By 3.1, we get that kφk ≤ kykY. On the other hand,
φ(x) = y, thus kφk = kykY.
To simplify the notation, we shall without loss of generality assume
that kykY = 1.
NON-COMMUTATIVE UNIFORM BANACH GROUP
15
We consider the following pseudometric σ, resp. pseudonorm k · k,
on X. For any y, z ∈ X we set σ(y, z) = kφ′(y) − φ′(z)kY; similarly, for
every y ∈ X we set kyk = kφ′(y)kY. For every odd n we denote by
σn the restriction of σ to Xn. Similarly, for every even n we denote by
k · kn the restriction of k · k to Xn. It suffices to show, by induction,
that for every odd n we have σn ≤ ρn and for every even n we have
k · kn ≤ k · kn.
Consider the case n = 1. We have (from bi-invariance) σ1(x, e) =
It follows from the triangle
σ1(x−1, e) = ρ1(x, e) = ρ1(x−1, e) = 1.
inequality that σ1(x, x−1) ≤ 2 = ρ1(x, x−1).
Consider now the case n = 2. Take any α, β ∈ D1. We have kαx +
βx−1k2 = min{γ1 · ρ1(x, e) + γ2 · ρ1(x−1, e) + γ3 · ρ1(x, x−1) : αx +
βx−1 = γ1x + γ2x−1 + γ3(x − x−1), γ1, γ2, γ3 ∈ R}. However, for any
γ1, γ2, γ3 ∈ R if αx + βx−1 = γ1x + γ2x−1 + γ3(x − x−1), then by the
subadditivity and homogeneity of the norm we must have
kαx + βx−1k2 =≤ γ1 · σ1(x, e) + γ2 · σ1(x−1, e) + γ3 · σ1(x, x−1) ≤
γ1 · ρ1(x, e) + γ2 · ρ1(x−1, e) + γ3 · ρ1(x, x−1).
Thus k · k2 ≤ k · k2.
Consider now some general odd n. Necessarily, σn+1 satisfies all the
following inequalities (recall the analogous inequalities for ρn+1)
(a) σn+1(x, y) ≤ kx − ykn for every x, y ∈ Xn such that x − y ∈ Xn,
(b) σn+1(x, y) = σn+1(x−1, y−1) for every x, y ∈ Xn+1,
(c) σn+1(ab, cd) ≤ σn+1(a, c)+σn+1(b, d) for every a, b, c, d ∈ Xn+1 such
that ab, cd ∈ Xn+1,
(d) σn+1(x, (α1z1+. . .+αmzm)) ≤ α1·σn+1(x, z1)+. . .+αm·σn+1(x, zm),
where m > 1, αi > 0, for all i, α1 + . . . + αm = 1 and α1z1 + . . . +
αmzm ∈ Xn+1,
(e) σn+1(x, (α1z1 + . . . + αmzm)−1) ≤ α1 · σn+1(x, z−1
1 ) + . . . + αm ·
m ), where m > 1, αi > 0, for all i, α1 + . . . + αm = 1 and
ρn+1(x, z−1
(α1z1 + . . . + αmzm)−1 ∈ Xn+1.
Inequalities (a),(b) and (c) are clear; (d) follows since σn+1(x, (α1z1 +
. . . + αmzm)) = kφ′(x − (α1z1 + . . . + αmzm))kY ≤ α1kφ′(x − z1)kY +
. . . + αmkφ′(x − zm)kY = α1 · σn+1(x, z1) + . . . + αm · σn+1(x, zm); (e)
follows analogously.
Since ρn+1 has been shown to be the greatest function satisfying the
analogous conditions (a)-(e), it follows from the induction hypothesis
that σn+1 ≤ ρn+1.
Finally, for a general even n, a completely analogous argumentation,
using the fact that k · kn+1 was the greatest function satisfying certain
16
M. DOUCHA
conditions, gives that k · kn+1 ≤ k · kn+1.
The uniqueness is a standard argument using the universality prop-
(cid:3)
erty.
As mentioned above, one can rightfully suspect that X is not one of
the well-behaved Banach spaces. Actually, in our opinion, it would not
be difficult to enhance the construction above so that X was isometric
to the Gurarij space ([8]). It thus seems natural to ask whether one
can get a non-commutative Banach group modeled on a 'reasonable'
space.
We mentioned in the preliminary section that the Heisenberg group
U T 3
3 (R), modeled on a three-dimensional Banach space, is not a uni-
form Banach group, although it is very closed to it (this is also the
content of Remark 5.2. in [9]). The following question thus arises.
Question 3.3. Does there exist a finite-dimensional non-commutative
uniform Banach group?
We expect the answer to the previous question to be negative. How-
ever, one can then ask:
Question 3.4. Does there exist a non-commutative uniform Banach
group modeled on a Hilbert space?
As already mentioned in the introduction, uniform Banach groups
were originally introduced with connection to the infinite-dimensional
Hilbert's fifth problem. The following question thus also seems to be
natural.
Question 3.5. Are the (non-commutative) group operations on X (Fr´echet)
differentiable? Does there exist a non-commutative uniform Banach
group that is a Banach-Lie group?
References
[1] I. Ben-Yaacov, The linear isometry group of the Gurarij space is universal,
Proc. Amer. Math. Soc. 142 (2014), no. 7, 2459 -- 2467
[2] Y. Benyamini, J. Lindenstrauss, Geometric nonlinear functional analysis Vol.
1., American Mathematical Society Colloquium Publications, 48. American
Mathematical Society, Providence, RI, 2000
[3] M. Doucha, Non-abelian group structure on the Urysohn space, Fund. Math.
228 (2015), 251 -- 263
[4] P. Enflo, Uniform structures and square roots in topological groups I, Israel J.
Math. 8 (1970), 230-252
[5] P. Enflo, Uniform structures and square roots in topological groups II, Israel J.
Math. 8 (1970), 253272
NON-COMMUTATIVE UNIFORM BANACH GROUP
17
[6] R. Engelking, General topology, Translated from the Polish by the author.
Second edition. Sigma Series in Pure Mathematics, 6. Heldermann Verlag,
Berlin, 1989
[7] M. I. Graev, Free topological groups, Amer. Math. Soc. Translation 1951 (1951),
no. 35, 61 pp.
[8] V.I. Gurarij, Spaces of universal placement, isotropic spaces and a problem of
Mazur on rotations of Banach spaces, Sibirsk. Mat. Zh. 7 (1966) 1002 -- 1013
(in Russian)
[9] S. Prassidis, A. Weston, Manifestations of nonlinear roundness in analysis,
discrete geometry and topology, Limits of graphs in group theory and computer
science, 141170, EPFL Press, Lausanne, 2009
[10] S. Prassidis, A. Weston, Uniform Banach groups and structures, C. R. Math.
Acad. Sci. Soc. R. Can. 26 (2004), no. 1, 25 -- 32
[11] N. Weaver, Lipschitz algebras, World Scientific Publishing Co., Inc., River
Edge, NJ, 1999
Institute of Mathematics, Polish Academy of Sciences, 00-656 Warszawa,
Poland
E-mail address: [email protected]
|
1905.03409 | 1 | 1905 | 2019-05-09T02:07:05 | Flexible versions of the Stone-Weierstrass Theorem in General and Applications to Probability Theory | [
"math.FA"
] | When applying the classical Stone-Weierstrass common version in Probability Theory for example, and in other fields as well, problems may arise if all points of the compact set are not separated. A solution may consist in going back to the proof and finding alternative versions. In this note, we did it and come back with two flexible versions which are easily used for the needs of classical Probability Theory. | math.FA | math | FLEXIBLE VERSIONS OF THE STONE-WEIERSTRASS THEOREM IN GENERAL
AND APPLICATIONS TO PROBABILITY THEORY
GANE SAMB LO
Abstract. When applying the classical Stone-Weierstrass common version in Probability
Theory for example, and in other fields as well, problems may arise if all points of the com-
pact set are not separated. A solution may consist in going back to the proof and finding
alternative versions. In this note, we did it and come back with two flexible versions which
are easily used for the needs of classical Probability Theory.
Gane Samb Lo.
LERSTAD, Gaston Berger University, Saint-Louis, S´en´egal (main affiliation).
LSTA, Pierre and Marie Curie University, Paris VI, France.
AUST - African University of Sciences and Technology, Abuja, Nigeria
[email protected], [email protected], [email protected]
Permanent address : 1178 Evanston Dr NW T3P 0J9,Calgary, Alberta, Canada.
9
1
0
2
y
a
M
9
]
.
A
F
h
t
a
m
[
1
v
9
0
4
3
0
.
5
0
9
1
:
v
i
X
r
a
1. Introduction
In Probability Theory, the theorem of Stone-Weierstrass is one of the pos-
sible tools to prove that the characteristic function serves as an effective
characterization of the probability law of a random vector (say, in Rk, k ≥ 1)
and that the weak convergence of Probability laws is equivalent to the con-
vergence of characteristics functions.
However, the application of such a theorem from the common version of
the theorem is not straightforward. An appeal for an extended version (see
Simmons (1963), page 165) is made. An example is available in Billingsley
(1968), but such a version is stated on locally Hausdorff Compact sets and
uses functions vanishing at infinity. At a such an earlier stage of Proba-
bility Theory, it is regrettable that a quite highly sophisticated approach is
required. Beyond this example, we think that the simple knowledge of the
common version only does not allow the researcher in Probability and Sta-
tistics to draw all the benefits of that extraordinary tool. More generally,
it seems that authors work from the consequences rather than the true
principle.
1
2
GANE SAMB LO
Here we want to show that simple versions can give more flexibility in ap-
plying that Theorem by highlighting the elements of the proofs and by
sticking more to the lines of the proof. In doing so, we just allow flexibility
in the application of the mentioned theorem.
The rest of the paper is organized as follows.
In the Section 2, we just
state the most common version of the Stone-Weierstrass theorem stated
In the same section, we explain
in the introductory books of Topology.
the reasons that led us to the current extended version.
In Section 3,
we provide remarks from the proof of the theorem in classical textbooks
and just propose versions preserving the classical conclusions. In Section
4, we show how to prove the Billingsley problem with the version of this
paper.
2. Common versions of the Stone Weierstrass Theorem and its
application
The theorem is stated as follows.
Theorem 1. (Complex version of Stone-Weierstrass's Theorem) Let K be
a non-singleton compactum (A Hausdorff space on which the Heine-Borel
holds) and A be a non-empty sub-algebra of C(K, C), the continuous func-
tions defined on K with values in ring C of complex numbers, such that the
following assertion :
(1) A separates the points of K, that is, for any distinct elements x and y of
K, there exists f ∈ A such that f (x) 6= f (y),
and one of the two assertions
(2) For any x ∈ K, there exists f ∈ A such that f (x) 6= 0.
(3) A contains all the constant functions,
and the additional assertion :
(4) For all f ∈ A, its conjugate function ¯f = R(f ) − iIm(f ) ∈ A,
hold. Then A is dense in C(K, C)
A = C(K, C).
FLEXIBLE VERSIONS OF THE STONE-WEIERSTRASS THEOREM
3
The case where K is a singleton is dismissed as an obvious thing.
This theorem is applied in Probability Theory and in many other fields like
Neural networks XXXX. Let us explain the following case study. Recall that
two probability measures P1 and P2 on a metric space (E, d) endowed with
the Borel σ-algebra B(E) are equal if and only if : for any f in the class
Cb(E) of bounded and continuous real-valued functions defined on E, we
have
ZE
f dP1 = ZE
f dP2. (EL)
(See Lo (2018), Chapter 3, Part III for a proof). If E = Rk, we define for any
probability measure P on (Rk, B(Rk), its characteristic function by
ΨP(u) = ZRk
exp(ihu, xi)P(x), u ∈ Rk.
As in classical books of Probability Theory, we want to show that the equal-
ity of the characteristic functions
ΨP1 = ΨP1,
(EC)
on Rk is equivalent to Formula (EL) and hence to the equality in law. In
fact, that (EL) implies (EC) is obvious. The main work is to prove that (EC)
implies (EL). The main ideas in the proof are :
(a) Fixing a real number a > 0 and considering Ka = [−a, a]k.
(b) Considering the family H of functions which are finite linear combina-
tion (with real coefficients) of the form
exp(iπhn, ui/a), u ∈ Rk.
where n = (n1, ..., nk)T ∈ Zk.
(c) Showing that the class of Ha of restrictions of elements of H on Ka is
dense in C(Ka, R) by the Stone-Weierstrass theorem, which implies that for
any ε > 0, for any f ∈ C(Rk), there exists h ∈ H such that
kf − hkKa = sup
x∈Ka
f (x) − h(x) ≤ ε,
4
GANE SAMB LO
The two following properties are important in the sequel :
(d) By periodicity, the uniform norm of h on Ka is equal to the uniform
norm of h on the whole space.
(e) The integral of h with respect to Pi, on Rk, is a finite linear combination
of the values of the characteristic function of Pi, i ∈ {1, 2}.
The conclusion is achieved through a smart combination of limit results
as a ↑ +∞. But the key tool is in Point (c).
Unfortunately, the subclass Ha possesses all the desired conditions to ap-
ply Stone-Weierstrass's theorem on the compact Ka except the separation
of the points of Ka. Indeed, each h ∈ H assigns the same values of all 2k
edge points of ka of the form (±a, ±a, · · · , ±a)T .
The version we are going to provide will solve this drawback.
Before we go further, it is worth-mentioning that in some special and par-
ticular cases, it might be possible to show that A = C(K, C) by explicitly
constructing for any f ∈ C(K, C) a sequence (fn)n≥0 ⊂ A such that fn con-
verges uniformly to f as n goes to infinity. But it is not reasonable to expect
this is non simple cases.
3. Analysis of the Proof of Stone-Weierstrass's theorem
The proof as in classical textbooks is based on the two following lemmas.
Lemma 1. Let K be a compactum (A Hausdorff space on which the Heine-
Borel holds) and A be a non-empty lattice subclass of C(K), that is A is
closed under finite minimum and maximum of its elements :
∀(f, g) ∈ A, min(f, g) ∈ A and max(f, g) ∈ A.
Let f ∈ C(K). Then f ∈ A if and only if for each two distinct elements x and
y of K, the restriction f{x,y} is limit of a restrictions of a sequence of elements
of A on {x, y} :
∀(x, y) ∈ K 2, ∃(f x,y
n )n≥0 ⊂ A, f x,y
n (t) → f (t) as n → +∞, f or t ∈ {x, y}.
FLEXIBLE VERSIONS OF THE STONE-WEIERSTRASS THEOREM
5
Lemma 2. Let A be a closed non-empty sub-algebra in C(K). Then A is a
lattice space.
As a consequence, a closed non-empty sub-algebra in C(K) is lattice.
By combining these two lemmas, proving A = C(K, C) is the same as prov-
ing that A1 = C(K, C) with A1 = A, which is a lattice sub-algebra of C(K, C)
whenever A is a sub-algebra of C(K, C). This leads to the following general
form
Theorem 2. (A first general version of Stone-Weierstrass's Theorem) Let K
be a compact space (A Hausdorff space on which the Heine-Borel holds) and
A be a non-empty sub-algebra of C(K, C), the continuous functions defined
on K with values in the ring C of complex numbers. We have A = C(K, C)
whenever we have for all f ∈ C(K, C),
(3.1)
∀(x, y) ∈ K 2, ∃(f x,y
n )n≥0 ⊂ A, f x,y
n (t) → f (t)asn → +∞, f or t ∈ {x, y}.
By stating this, we simply bring to the surface the most inner technical
tool in the proof of the classical Theorem.
Actually the usual version of the Stone-Weierstrass version seeks to get
this by requiring Assumptions (1) and (2) or (4) in Theorem 2 in the real
version. Assumption (4) is required to extend the real version to the com-
plex version. Let us remain in the real case. The proof in Choquet (1966)
and in almost many other books goes too far. Let us explain why.
Let x and y be two distinct elements of K. Let f ∈ C(K) with f (x) = α and
f (y) = β. In the classical proof, we combine Assumptions (1) and (2) to find
a function g ∈ A such that g(x) = α and g(y) = β. So, Formula 3.1 holds
but with a constant sequence f x,y
n = f x,y. If x = y, and since K is not a
singleton, the method is re-conducted for x and z with x 6= z and Formula
3.1 also holds.
As a conclusion, making happen Formula 3.1 with a constant sequence
is a high price to pay for getting the approximation. The following version
of the Stone-Weierstrass Theorem will be still valid. It is based on the fact
that we do not need that A separates all the points.
Corollary 1. (A second version of Stone-Weierstrass's Theorem) Let K be
a non-singleton compactum (A Hausdorff space on which the Heine-Borel
6
GANE SAMB LO
holds) and A be a non-empty sub-algebra of C(K, C), the continuous func-
tions defined on K with values in ring C of complex numbers. Suppose that
there exists K0 ⊂ K such that we have
∀(x, y) ∈ K 2
n (t) → f (t) as n → +∞, f or t ∈ {x, y}.
0 , ∃(f x,y
n )n≥0 ⊂ A, f x,y
(1) A separates the points of K \ K0 and separates any point of K0 from any
point of K,
and one of the two assertions
(2) For any x ∈ K \ K0, there exists f ∈ A such that f (x) 6= 0.
(3) A contains all the constant functions,
and the additional assertion :
(4) For all f ∈ A, its conjugate function ¯f = R(f ) − iIm(f ) ∈ A,
hold. Then A is dense in C(K, C) :
A = C(K, C).
Corollary 2. (A third version of Stone-Weierstrass's Theorem) Let K be a non-
singleton compactum (A Hausdorff space on which the Heine-Borel holds)
and A be a non-empty sub-algebra of C(K, C), the continuous functions de-
fined on K with values in ring C of complex numbers. Suppose that there
exists K0 ⊂ K such that K \ K0 has at least two elements and the following
assertions :
(0) We have : for all f ∈ C(K, C)
∃(fn)n≥0 ⊂ A, ∀x ∈ K0, fn(x) → f (x) as n → +∞, f or t ∈ {x, y}.
or
∃g ∈ A, ∀x ∈ K0, g(x) = f (x),
(1) A separates the points of K \ K0 and separates any point of K0 from any
point of K,
FLEXIBLE VERSIONS OF THE STONE-WEIERSTRASS THEOREM
7
and one of the two assertions
(2) For any x ∈ K \ K0, there exists f ∈ A such that f (x) 6= 0.
(3) A contains all the constant functions,
and the additional assertion :
(4) For all f ∈ A, its conjugate function ¯f = R(f ) − iIm(f ) belongs to A.
Then A is dense in C(K, C) :
A = C(K, C).
4. Return to the Application of SW Theorem in Probability Theory
Let us go back to Section 2, Point (b). We are going to apply Corollary ??.
Assume we have the same notation (regarding ]a, b[, r > 0 and Kr). Let us
denote Cb,0,r(Rd) the restrictions of elements of Cb,0(Rd) on Kr. We should
prove that Cb,0,r ⊂ Hr. For this, let us take f ∈ Cb,0,r(Rd). The general form
of an element of H is
h(x) = X1≤j≤p
ap exp(iπhnp, xi/r), x ∈ Rk,
where ap ∈ C, np ∈ Zk. Let us take
K0 = ∂Kr = {x ∈ Kr : ∀ j ∈ {1, · · · , k}, xi = −r or xi = r}.
We have f = 0 on K0. If x and y are two points in Kr such they are not among
the edging points in the border ∂Kr both, then there exists j0 ∈ {1, · · · , d}
such that 0 < xj0 − yj0 < 2r that is (xj0 − yj0)/r < 2 and the function
hr(x) = exp(iπxj0/r) separates x and y. By adding the other assumptions,
conclude that f is in the closure of Hr.
References
Billingsley, P. (1968). Convergence of Probability measures. John Wiley,
New-York.
Choquet G. (1964) Cours d'analyse II : Topologie. Masson.
Choquet G.(1966). Topology (translated by Amiel Feinstien). Academic
press, New-York and London.
8
GANE SAMB LO
Tichmarsh E. C.(1939) Theory of functions. Oxford University Press. 2nd
Edition. London.
Simmons G.F.(1963)
Introduction to topology and moder analysis.
McGraw-Hill Book Company Inc., New-York.
Munkress J.M.(1997) Topology : A first course. Prentice-Hall, Inc., Engle-
wood Cliffs, New-Jersey
Lo, G.S.(2018). Mathematical Foundation of Probability Theory.
:
SPAS Books Series. Saint-Louis, Senegal - Calgary, Canada. Doi
http://dx.doi.org/10.16929/sbs/2016.0008. Arxiv : 1808.01713
|
1311.2270 | 1 | 1311 | 2013-11-10T13:25:34 | On linear operators with s-nuclear adjoints: $0< s \le 1$ | [
"math.FA"
] | If $ s\in (0,1]$ and $ T$ is a linear operator with $ s$-nuclear adjoint from a Banach space $ X$ to a Banach space $ Y$ and if one of the spaces $ X^*$ or $ Y^{***}$ has the approximation property of order $s,$ $AP_s,$ then the operator $ T$ is nuclear. The result is in a sense exact. For example, it is shown that for each $r\in (2/3, 1]$ there exist a Banach space $Z_0$ and a non-nuclear operator $ T: Z_0^{**}\to Z_0$ so that $Z_0^{**}$ has a Schauder basis, $ Z_0^{***}$ has the $AP_s$ for every $s\in (0,r)$ and $T^*$ is $r$-nuclear. | math.FA | math |
ON LINEAR OPERATORS WITH
S-NUCLEAR ADJOINTS: 0 < S ≤ 1
O.I. Reinov
Abstract. If s ∈ (0, 1] and T is a linear operator with s-nuclear adjoint from a
Banach space X to a Banach space Y and if one of the spaces X ∗ or Y ∗∗∗ has the
approximation property of order s, APs, then the operator T is nuclear. The result
is in a sense exact. For example, it is shown that for each r ∈ (2/3, 1] there exist
a Banach space Z0 and a non-nuclear operator T : Z ∗∗
0 has a
Schauder basis, Z ∗∗∗
has the APs for every s ∈ (0, r) and T ∗ is r-nuclear.
0 → Z0 so that Z ∗∗
0
§1. Introduction
We will be interested in the following question from [6, Problem 10.1]: suppose
T is a (bounded linear) operator acting between Banach spaces X and Y, and let
s ∈ (0, 1). Is it true that if T ∗ is s-nuclear then T is s-nuclear too?
As well known, for s = 1 the negative answer was obtained for the first time by
Figiel and Johnson in [4]. For s ∈ (2/3, 1] the negative answer can be found, e.g.,
in [14]. Here, we are going to give some (partially) positive results in this direction
as well as to show the sharpness of them.
It is not difficult to see that if T ∗ ∈ Ns(Y ∗, X ∗) then T ∈ Np(X, Y ) with 1/s =
1/p + 1/2 (what is, surely, must be known; see, e.g., [13], [14]). This is the best
possible result in the scale of q-nuclear operators: the sharpness of the last assertion,
for s ∈ (2/3, 1], can be found, for instance, in [14].
Below, we consider a little bit different question: under which assumptions on
Banach spaces X and Y
(∗) an operator T ∈ L(X, Y ) is nuclear if its adjoint T ∗ is s-nuclear?
One of the possibilities for getting some positive answers to (∗) is to apply the
so-called approximation properties (the APs) of order s, s ∈ (2/3, 1] (we assume
that s > 2/3 since for s ≤ 2/3 the answer is evident).
‡ 2010 AMS Subject Classification: 47B10. Operators belonging to operator ideals (nuclear,
p-summing, in the Schatten-von Neumann classes, etc.)
Key words: s-nuclear operators, Schauder bases, approximation properties, tensor products.
Typeset by AMS-TEX
We will prove that (∗) is true if either X ∗ or Y ∗∗∗ has the APs (Theorem 1). Some
examples, given after Theorem 1, will show that those assumptions are, in a sense,
necessary (for example, it is not enough to assume that either X or Y ∗∗, or both of
the spaces have the approximation properties, even in the sense of A.Grothendieck).
Let us note that the case where s = 1 was firstly investigated in the paper [8] by
Eve Oja and Oleg Reinov.
§2. Notations and preliminaries
For the references of different things, one can see [9]. All the spaces under con-
siderations (X, Y, . . . ) are Banach, all linear mappings (operators) are continuous;
as usual, X ∗, X ∗∗, . . . are Banach duals (to X), and x′, x′′, . . . (or y ′, . . . ) are the
functionals on X, X ∗, . . . (or on Y, . . . ). By πY we denote the natural isometric
injection of Y into its second dual. If x ∈ X, x′ ∈ X ∗ then hx, x′i = hx′, xi = x′(x).
L(X, Y ) stands for the Banach space of all linear bounded operators from X to Y.
Recall that an operator T : X → Y is s-nuclear (0 < s ≤ 1) if it is of the form
T x =
hx′
k, xiyk
∞Xk=1
for x ∈ X, where (x′
ks yks < ∞. We use the notation
Ns(X, Y ) for the space of all such operators. Every s-nuclear operator is a canonical
k) ⊂ X ∗, (yk) ⊂ Y,Pk x′
s-projective tensor product of X ∗ and Y that is a subspace of the Grothendieck
image of an element of a projective tensor product, namely: denote by X ∗b⊗sY the
projective tensor product X ∗b⊗Y (= X ∗b⊗1Y ) consisting of all tensor elements z
which admit representations of kind
z =
∞Xk=1
x′
k ⊗ yk with
∞Xk=1
x′
ks yks < ∞.
Then every s-nuclear operator from X to Y is an image of an element of X ∗b⊗sY
via the canonical mappings
X ∗b⊗sY
js→ X ∗b⊗1Y
j
→ L(X, Y ).
If z ∈ X ∗b⊗Y then we denote the corresponding operator (from X to Y ) by z.
We say, following, e.g., [13] or [14], that a Banach space Y has the APs (the
approximation property of order s), if for every Banach space X the natural map jjs
2
(from above) is one-to-one (note that AP1 = AP of A. Grothendieck). It is the same
[5], see also [11], [13] or [14]).
hU, zi := trace U ◦ z
k ⊗ yk). So, the element
APt for 0 < t < s ≤ 1. Every Banach space has the AP2/3 (Grothendieck's Theorem
as to say that the natural map Y ∗b⊗sY → L(Y, Y ) is one-to-one. Therefore, if Y has
the APs, we can write X ∗b⊗sY = Ns(X, Y ) whichever X was. Clearly, APs =⇒
Recall that the dual space to X ∗b⊗1Y is L(X, Y ∗∗) and the duality is defined by
"trace": if z ∈ X ∗b⊗1Y and U ∈ L(Y, X ∗∗), then
(=Pkhx′
z ∈ X ∗b⊗sY is zero iff it it zero in the projective tensor product X ∗b⊗Y iff for every
U ∈ L(Y, X ∗∗) trace U ◦ z = 0. If z ∈ X ∗b⊗Y then the corresponding operator z is
k, U yki for a projective representation z =Pk x′
zero iff for every R ∈ Y ∗ ⊗ X we have: trace R ◦ z = 0 (evidently).
Examples of Banach spaces with APs : for s ∈ [2/3, 1] and 1/p + 1/2 = 1/s, every
quotient of any subspace of any Lp-space (and every subspace of any quotient of
any Lp′ -space) has the APs (as well as all their duals; see, e.g., [13] or [14]; for a
more general fact, see Lemma 3 below). Here 1/p + 1/p′ = 1. All the Banach spaces
have APs for s ∈ (0, 2/3]; but if 2/3 ≤ s1 < s2 ≤ 1 then APs2 =⇒ APs1 but APs1
does not imply APs2. It is known that for every p 6= 2, p ∈ [1, ∞], there exists a
subspace (a lot of ones) of lp without the Grothendieck approximation property;
thus, for example, AP1 6= APs if s ∈ (0, 1). Something (about "APs1 6= APs2") will
be possible to get from Examples or Theorems 2-3 below.
We will use later the following facts (surely, well known, but maybe not mentioned
in the literature):
Lemma 1. If T ∈ L(X, Y ∗∗) then T ∗πY ∗ (Y ∗) = T and (T ∗πY ∗ (Y ∗))∗X = T.
So, one can write L(X, Y ∗∗) = L(Y ∗, X ∗) (in Banach sense).
Lemma 2. If T ∈ L(X, Y ) then
1) πY T ∈ Ns(X, Y ∗∗) ⇐⇒ T ∗ ∈ Ns(Y ∗, X ∗);
2) T ∈ Ns(X, Y ) =⇒ T ∗ ∈ Ns(Y ∗, X ∗).
Lemma 3. If E is a Banach space of type 2 (respectively, of cotype 2) and of
cotype q0 (respectively, of type q′
0) then E has the APs where 1/s = 3/2 − 1/q0.
The proofs of Lemma 3 can be found in [13], [14].
3
§3. A positive result
Theorem 1. Let s ∈ (0, 1], T ∈ L(X, Y ) and either X ∗ ∈ APs or Y ∗∗∗ ∈ APs.
If T ∈ Ns(X, Y ∗∗), then T ∈ N1(X, Y ). In other words, under these conditions
from the s-nuclearity of the conjugate operator T ∗ it follows that the operator T
belongs to the space N1(X, Y ) (that is nuclear).
Proof. Suppose there exists an operator T ∈ L(X, Y ) such that T /∈ N1(X, Y ),
but πY T ∈ Ns(X, Y ∗∗). Since either X ∗ or Y ∗∗ has the APs, Ns(X, Y ∗∗) =
X ∗b⊗sY ∗∗. Therefore the operator πY T can be identified with the tensor element
t /∈ X ∗b⊗1Y (the
t ∈ X ∗b⊗sY ∗∗ ⊂ X ∗b⊗1Y ∗∗; in addition, by the choice of T,
space X ∗b⊗1Y is considered as a closed subspace of the space X ∗b⊗1Y ∗∗). Hence
there is an operator U ∈ L(Y ∗∗, X ∗∗) = (cid:0)X ∗b⊗1Y ∗∗(cid:1)∗
trace U ◦ t = trace (t∗ ◦ (U ∗X ∗ )) = 1 and trace U ◦ πY ◦ z = 0 for each z ∈ X ∗b⊗1Y.
From the last it follows that, in particular, U πY = 0 and π ∗
Y U ∗X ∗ = 0. In fact, if
with the properties that
x′ ∈ X ∗ and y ∈ Y, then
< U πY y, x′ >=< y, π ∗
Y U ∗X ∗ x′ >= trace U ◦ (x′ ⊗ πY (y)) = 0.
is equal identically to zero.
Evidently, the tensor element U ◦ t ∈ X ∗b⊗sX ∗∗ induces the operator U πY T, which
If X ∗ ∈ APs then X ∗b⊗sX ∗∗ = Ns(X, X ∗∗) and, therefore, this tensor element is
zero what contradicts to the equality trace U ◦ t = 1.
Let now Y ∗∗∗ ∈ APs. In this case
V := (U ∗X ∗ ) ◦ T ∗ ◦ π ∗
Y : Y ∗∗∗ → Y ∗ → X ∗ → Y ∗∗∗
n ⊗ y ′′
uniquely determines a tensor element t0 from the s-projective tensor product Y ∗∗∗∗b⊗sY ∗∗∗.
Let us take any representation t = P x′
X ∗b⊗sY ∗∗. Denoting for the brevity the operator U ∗X ∗ by U∗, we obtain:
Y y ′′′) = U∗ (cid:16)(X y ′′
= U∗ (cid:16)X < y ′′
n for t as an element of the space
V y ′′′ = U∗ (T ∗π ∗
Y y ′′′ > x′
n ⊗ x′
n) π ∗
n, π ∗
Y y ′′
n, y ′′′ > U∗x′
n.
Y y ′′′(cid:17) =
n(cid:17) =
=X < π ∗∗
So, the operator V (or the element t0) has in the space Y ∗∗∗∗b⊗sY ∗∗∗ the repre-
sentation
V =X π ∗∗
Y (y ′′
4
n) ⊗ U∗(x′
n).
Therefore,
trace t0 = trace V =X < π ∗∗
(since π ∗
Y U∗ = 0; see above).
On the other hand,
Y (y ′′
n), U∗(x′
n, π ∗
Y U∗x′
n) >=X < y ′′
n >=X 0 = 0
V y ′′′ = U∗ (πY T )∗ y ′′′ = U∗◦t∗(y ′′′) = U∗ (cid:16)X < y ′′
whence V =P y ′′
n >=X < U y ′′
trace t0 = trace V =X < y ′′
n). Therefore
n ⊗ U∗(x′
n, U∗x′
The obtained contradiction completes the proof of the theorem. (cid:4)
n, y ′′′ > x′
n, y ′′′ > U∗x′
n,
n(cid:17) =X < y ′′
n, x′
n >= trace U ◦ t = 1.
§4. Examples
We need two examples to show that all conditions, imposed on X and Y in
Theorem 1, are essential. Ideas of such examples are taken from the Reinov's work
[10] (not translated, as we know, from Russian).
Example 1. Let r ∈ (2/3, 1], q ∈ [2, ∞), 1/r = 3/2 − 1/q. There exist a separable
reflexive Banach space Y0 and a tensor element w ∈ Y ∗
0 b⊗rY0 so that w 6= 0, w = 0,
the space Y0 (as well as Y ∗
0 ) has the APs for every s < r (but, evidently, does not
have the APr). Moreover, Y0 is of type 2 and of cotype q0 for any q0 > q.
Proof. We will use a variant of Per Enflo's [3] example of a Banach space without
approximation property given in the book [9, 10.4.5]. Namely, it follows from
the constructions in [9] that there exist a Banach space X and a tensor element
z ∈ X ∗b⊗X so that trace z 6= 0, the operator z, generated by z, is identically zero
and z can be represented in the following form:
(1)
z =
∞XN=1
3·2NXn=1
N 1/22−3N/2x′
nN ⊗ xnN ,
where the sequences (x′
nN ) and (xnN ) are norm bounded by 1 (see [2] or [9, 10.4.5]).
Fix r ∈ (2/3, 1] and put 1/q = 3/2 − 1/r (thus, q ∈ [2, ∞)). Let {εN }∞
1 be a
number sequence such that P N −1−εN < +∞; γN = 2 + 3εN /2; qN is a number
qN (cid:17)lq
such that 1/q − 1/qN = N −1 log2 N γN (therefore, qN > q). Set Y =(cid:16)PN l3·2N
5
.
Denote by enN (and e′
nN ) the unit orths in Y and in Y ∗ respectively (N =
1, 2, . . . ; n = 1, 2, . . . , 3 · 2N ), and put
z1 =XN Xn
T =XN Xn
2−N/r N −(1+εN )/r e′
nN ⊗ xnN ;
2−(3/2−1/r)N N (1/2+1/r+εN /r) x′
nN ⊗ enN .
Let us show that z1 ∈ Y ∗b⊗rX, T ∈ L(X, Y ) (then, evidently, z = z1 ◦ T 1). The
first inclusion is evident (because of the choice of εN ).
If x 6 1 then
T x 6 3 XN (cid:16)N 1/2+1/r+εN /r/2(3/2−1/r−1/qN )N(cid:17)q!1/q
.
Since 3/2 − 1/r − 1/qN = 1/q − 1/qN = N −1 log2 N γN , we get from the last
inequality:
T 6 3(cid:16)X N (1/2+1/r+εN /r−γN )q(cid:17)1/q
= 3(cid:16)X N −1−εN(cid:17)1/q
< ∞.
Hence,
z = z1 ◦ T, z : X T→ Y
z1→ X.
Now, let Y0 := T (X) ⊂ Y, T0 : X → Y0 be induced by T and z0 := z1 ◦ j
where j : Y0 ֒→ Y is the natural embedding. Then T0 ∈ L(X, Y0), z0 ∈ Y ∗
z = z1 ◦ T = z0 ◦ T0, trace z0 ◦ T0 6= 0 (so, z0 6= 0) and z0 = 0.
Write z0 as z0 =P∞
∞. We get:
m=1 f ′
m⊗fm, where (f ′
m) ⊂ Y ∗
0 , (fm) ⊂ X andPm f ′
0 b⊗rX,
mrfmr <
trace z0 ◦ T0 =XhT ∗
0 f ′
m, fmi =Xhf ′
m, T0fmi = trace T0 ◦ z0.
Therefore, w := T0 ◦ z0 ∈ Y ∗
0 b⊗rY0, trace w 6= 0 and w = 0.
Since the space Y is of type 2 and of cotype q0 for every q0 > q (and Y ∗ is
of cotype 2 and of type q′
0), the space Y0 (respectively, Y ∗
0 ) has the APs, where
1/s = 3/2 − 1/q0, for every s < r (Lemma 3).
1If w ∈ Y ∗ b⊗X and U ∈ L(X, Y ) then w ◦ U denotes the image of w in X ∗ b⊗X under the map
m ⊗ ψm
m, Uψmi =
U ∗ ⊗ idX . So, if w = P ϕ′
is a representation of w ◦ U in X ∗ b⊗X. Then trace w ◦ U = PhU ∗ϕ′
trace U ◦ w, where U ◦ w is the image of w in Y ∗ b⊗Y under the map idY ∗ ⊗U.
m ⊗ ψm is a representation of w in Y ∗ b⊗X then w ◦ U = P U ∗ϕ′
m, ψmi = Phϕ′
6
Remark: We have a nice "by-product consequence" of Example 1. For q = 2
pj(cid:17)l2
(that is, r = 1), the space Y0 is a subspace of the space of type (cid:16)Pj lkj
pj ց 2 and kj ր ∞. Every such space is an asymptotically Hilbertian space (for
with
definitions and some discussion, see [1]). So, we got:
Corollary. There exists an asymptotically Hilbertian space without the Grothen-
dieck approximation property.
First example of such a space was constructed (by O. Reinov) in 1982 [10], where
A. Szankowski's results were used (let us note that in that time there was not yet
such notion as "asymptotically Hilbertian space"). Later, in 2000, by applying Per
Enflo's example in a version of A.M. Davie [2], P. G. Casazza, C. L. Garc´ıa and
W. B. Johnson [1] gave another example of an asymptotically Hilbertian space which
fails the approximation property. We here, not being searching for an example of
such a space, have got it (accidentally) by using the construction from [9].
Example 2. Let r ∈ [2/3, 1), q ∈ (2, ∞], 1/r = 3/2 − 1/q. There exist a subspace
Yq of the space lq and a tensor element wq ∈ Y ∗
s > r, wq 6= 0, wq = 0 and the space Yq (as well as Y ∗
q b⊗1Yq so that wq ∈ Y ∗
q b⊗sYq for each
q ) has the APr (but, evidently,
does not have the APs if 1 ≥ s > r). Clearly, Yq is of type 2 and of cotype q for
q < ∞.
Proof. We are going to follow the way indicated in the proof of the assertion from
Example 1. Let X and z be as in that proof, so that z has the form (1). Fix
r ∈ [2/3, 1). Now 1/q = 3/2 − 1/r, and we put εN = 0 and γN = 2 for all N ; all
qn's are equal to q. Let us fix also an α = α(q) > 0 (to be defined later). Consider
the space Y :=(cid:16)PN l3·2N
Denote by enN (and e′
q
(cid:17)lq
(in the case q = ∞ "lq" means "c0").
nN ) the unit vectors in Y and in Y ∗ respectively (N =
1, 2, . . . ; n = 1, 2, . . . , 3 · 2N ), and set this time
2−N/r N α e′
nN ⊗ xnN ;
2−(3/2−1/r)N N 1/2−α x′
nN ⊗ enN .
z1 =XN Xn
T =XN Xn
Then z1 ∈ Y ∗b⊗sX for every s > r. Indeed, if s ∈ (r, 1] then
XN Xn
3 · 2N · 2−Ns/r N α s =XN
[2−N/r N α]s =XN
7
3 · 2−ε0N N αs < ∞,
where ε0 = s/r − 1 > 0.
Show that T ∈ L(X, Y ) (clearly, then z = z1 ◦ T ). Indeed, if x 6 1 then, for
q < ∞ :
T xq
lq
≤XN Xn (cid:16)2−N/q N 1/2−α hx′
nN , xi(cid:17)q
≤XN
3·2N −(N/q)q N (1/2−α)q =XN
3·1·N α0,
where α0 = (1/2 − α)q. Now, take α > 0 such that α0 = −2. For q = ∞ take
α = α(∞) = 1.
Therefore,
z = z1 ◦ T, z : X T→ Y z1→ X;
As in the case of the previous proof (in Example 1), let Yq := T (X) ⊂ Y, Tq :
X → Yq be induced by T and zq := z1 ◦ j where j : Yq ֒→ Y is the natural
embedding. Then Tq ∈ L(X, Yq), zq ∈ Y ∗
trace zq ◦ Tq 6= 0 (so, zq 6= 0) and zq = 0.
Write zq as zq =P∞
∞. We get:
m=1 f ′
q b⊗sX for all s > r, z = z1 ◦ T = zq ◦ Tq,
q , (fm) ⊂ X andPm f ′
m fm <
m⊗fm, where (f ′
m) ⊂ Y ∗
trace zq ◦ Tq =XhT ∗
Therefore, wq := Tq ◦ zq ∈ Y ∗
q f ′
m, Tqfmi = trace Tq ◦ zq.
m, fmi =Xhf ′
q b⊗sYq for every s > r, trace wq 6= 0 and wq = 0.
Finally, Lemma 3 says that the space Yq has the APr if q < ∞. If q = ∞ then, as
we know, any Banach space has the AP2/3.
Remark: The space Y∞ from Example 2 not only does not have the APs for any
s ∈ (2/3, 1], but also does not have the APp (in the sense of the paper [12]) for all
p ∈ [1, 2) (this follows from some facts proved in [13]).
§4. Applications of Examples
Next two theorems show that the conditions "X ∗ has the APs" and "Y ∗∗∗ has
the APs" are essential in Theorem 1 and can not be replaces by weaker conditions
"X has the APs" (even by "X ∗ has the AP1") or "Y ∗∗ has the APs"; moreover,
even "both X and Y ∗∗ have the AP1" is not enough for the conclusion of Theorem
1 to be valid.
8
Theorem 2. Let r ∈ (2/3, 1], q ∈ [2, ∞), 1/r = 3/2 − 1/q. There exist a Banach
space Z0 and an operator T ∈ L(Z ∗∗
0 , Z0) so that
(1) Z ∗∗
0 has a Schauder basis (so, has the MAP);
(2) all the duals of Z0 are separable;
(3) Z ∗∗∗
0
has the APs for every s ∈ (0, r);
(4) πZ0T ∈ Nr(Z ∗∗
0 , Z ∗∗
0 );
(5) T /∈ N1(Z ∗∗
0 , Z0);
(6) Z ∗∗∗
0
does not have the APr.
Proof. Let us fix r ∈ (2/3, 1], q ∈ [2, ∞), 1/r = 3/2 − 1/q and take the pair (Y0, w)
from Example 1. Let Z0 be a separable space such that Z ∗∗
0 has a basis and there
exists a linear homomorphism ϕ from Z ∗∗
the subspace ϕ∗(Y ∗
0 ) in complemented in Z ∗∗∗
0
(see [7, Proof of Corollary 1]). Lift the tensor element w, lying in Y ∗
to an element2 w0 ∈ Y ∗
0 , so that ϕ ◦ w0 = w, and set T := w0 ◦ ϕ. Since
0 b⊗rZ ∗∗
trace w0 ◦ ϕ = trace ϕ ◦ w0 = trace w = 1 and Z ∗∗
Besides, the operator ^ϕ ◦ w0 : Y0 → Z ∗∗
is equal to zero. Therefore w0(Y0) ⊂ Ker ϕ = Z0 ⊂ Z ∗∗
0 onto Y0 with the kernel Z0 ⊂ Z ∗∗
0
∼= ϕ∗(Y ∗
and, moreover, Z ∗∗∗
0
so that
0 ) ⊕ Z ∗
0
0 b⊗rY0, up
0 has the AP, then fw0 = w0 6= 0.
0 , that is the operator w0 is
0 → Y0, associated with the tensor ϕ ◦ w0,
acted from Y0 into Z0.
Since the subspace ϕ∗(Y ∗
0 ) is complemented in Z ∗∗∗
0
, then w0 ◦ ϕ ∈ Z ∗∗∗
Nr(Z ∗∗
0 , Z0) iff w0 ∈ Y ∗
0 b⊗rZ0 = Nr(Y0, Z0).
If w0 ∈ Nr(Y0, Z0), then, for its arbitrary (nonzero!) Nr-representation of the
n ⊗ zn, the composition ϕ ◦ w0 is a zero tensor element in Y ∗
but this composition represents the element w, which, by its choice, can not be
form w0 =P y ′
zero. Thus, w0 /∈ Nr(Y0, Z0) and, thereby, w0 ◦ ϕ /∈ Z ∗∗∗
the other hand, certainly, w0 ◦ ϕ ∈ Z ∗∗∗
0 = Nr(Z ∗∗
0 b⊗rZ0 = Nr(Z ∗∗
0 , Z ∗∗
0 ).
∼= ϕ∗(Y ∗
0 ) ⊕ Z ∗
0 , one has that the space Z ∗∗∗
0
has the APs for
0 b⊗rZ ∗∗
Finally, since Z ∗∗∗
0
every s ∈ (0, r). (cid:4)
0 b⊗rZ0 =
0 b⊗rY0;
0 , Z0). On
Theorem 3. Let r ∈ [2/3, 1), q ∈ (2, ∞], 1/r = 3/2 − 1/q. There exist a Banach
space Zq and an operator T ∈ L(Z ∗∗
q , Zq) so that
(1) Z ∗∗
q has a Schauder basis (so, has the MAP);
2If w = P∞
k=1 y′
k ⊗ yk is any representation of w in Y ∗
a way that the last sequence is absolutely r-summing and ϕ(z′′
0 b⊗rY0, then we take {z′′
n) = yn for every n.
n} ⊂ Z ∗∗
0
in such
9
(2) if q < ∞ then all the duals of Zq are separable;
(3) Z ∗∗∗
q
has the APr;
(4) πZq T ∈ Ns(Z ∗∗
q , Z ∗∗
q ) for every s ∈ (r, 1];
(5) T /∈ N1(Z ∗∗
q , Zq);
(6) Z ∗∗∗
q
does not have the APs for any s ∈ (r, 1];
Proof. Let us fix r ∈ [2/3, 1), q ∈ (2, ∞], 1/r = 3/2 − 1/q and take the pair (Yq, wq)
from Example 2. Let Zq be a separable space such that Z ∗∗
q has a basis and there
exists a linear homomorphism ϕ from Z ∗∗
the subspace ϕ∗(Y ∗
q ) in complemented in Z ∗∗∗
q
q onto Yq with the kernel Zq ⊂ Z ∗∗
q
∼= ϕ∗(Y ∗
and, moreover, Z ∗∗∗
q
so that
q ) ⊕ Z ∗
q
(as in the proof of Theorem 2, see [7, Proof of Corollary 1]).
Construct w0 ∈ Y ∗
(following the way of the proof of Theorem 2) so that
q
q b⊗1Z ∗∗
q
q b⊗sZ ∗∗
w0 ∈ Y ∗
for every s ∈ (r, 1] and ϕ ◦ w0 = wq (it is possible to apply a
"simultaneous lifting" procedure -- see Footnote 2, -- since w has a form from the
proof of the assertion of Example 2). Set T := w0 ◦ ϕ. From this point, the proof
repeats the arguments of the proof of Theorem 2, and we have to mention only:
since Z ∗∗∗
q
∼= ϕ∗(Y ∗
q ) ⊕ Z ∗
q , one has that the space Z ∗∗∗
q
has the APr.
REFERENCES
1. Casazza P. G., Garc´ıa C. L., Johnson W. B.,, An example of an asymptoti-
cally Hilbertian space which fails the approximation property, Proceedings of
the American Mathematical Society 129 (2001), no. 10, 3017-3024.
2. Davie A.M., The approximation problem for Banach spaces, Bull. London Math.
Soc. 5 (1973), 261 -- 266.
3. Enflo P., A counterexample to the approximation property in Banach spaces,
Acta Math. 130 (1973), 309 -- 317.
4. Figiel T., Johnson W.B., The approximation property does not imply the bounded
approximation property, Proc. Amer. Math. Soc. 41 (1973), 197 -- 200.
5. Grothendieck A., Produits tensoriels topologiques et espases nucl´eaires, Mem.
Amer. Math. Soc. 16 (1955), 196 + 140.
6. Hinrichs A., Pietsch A., p-nuclear operators in the sense of Grothendieck, Math.
Nachr. 283 (2010), no. 2, 232 -- 261.
7. Lindenstrauss J., On James' paper "Separable Conjugate Spaces", Israel J.
Math. 9 (1971), 279 -- 284.
10
8. Oja E., Reinov O.I., Un contre-exemple `a une affirmation de A.Grothendieck,
C. R. Acad. Sc. Paris. -- Serie I 305 (1987), 121 -- 122.
9. Pietsch A., Operator ideals, North-Holland (1978), 536 p.
10. Reinov O. I., Banach spaces without approximation property, Functional Anal-
ysis and Its Applications 16 (1982), no. 4, 315-317.
11. Reinov O.I., A simple proof of two theorems of A. Grothendieck, Vestn. Leningr.
Univ. 7 (1983), 115-116.
12. Reinov O. I., Approximation properties of order p and the existence of non-
p-nuclear operators with p-nuclear second adjoints, Math. Nachr. 109 (1982),
125-134.
13. Reinov O.I., Disappearing tensor elements in the scale of p-nuclear operators,
in: Theory of Operators and Theory of Functions, Leningrad , LGU 1 (1983),
145-165.
14. Reinov O. I., Approximation properties APs and p-nuclear operators (the case
0 < s ≤ 1), Journal of Mathematical Sciences 115 (2003), no. 2, 2243-2250.
Oleg I. Reinov
Saint Petersburg State University
198504 St. Petersburg, Petrodvorets,
28 Universitetskii pr., Russia
E-mail address: [email protected]
11
|
1110.5706 | 1 | 1110 | 2011-10-26T05:50:48 | The Br\'ezis-Browder Theorem in a general Banach space | [
"math.FA"
] | During the 1970s Br\'ezis and Browder presented a now classical characterization of maximal monotonicity of monotone linear relations in reflexive spaces. In this paper, we extend and refine their result to a general Banach space. | math.FA | math | The Br´ezis-Browder Theorem in a general Banach space
Heinz H. Bauschke∗, Jonathan M. Borwein†, Xianfu Wang‡, and Liangjin Yao§
October 25, 2011
Abstract
1
1
0
2
t
c
O
6
2
]
.
A
F
h
t
a
m
[
1
v
6
0
7
5
.
0
1
1
1
:
v
i
X
r
a
During the 1970s Br´ezis and Browder presented a now classical characterization of maximal
monotonicity of monotone linear relations in reflexive spaces.
In this paper, we extend and
refine their result to a general Banach space.
2010 Mathematics Subject Classification:
Primary 47A06, 47H05; Secondary 47B65, 47N10, 90C25
Keywords: Adjoint, Br´ezis-Browder Theorem, Fenchel conjugate,
linear relation, maximally
monotone operator, monotone operator, operator of type (D), operator of type (FP), operator
of type (NI), set-valued operator, skew operator, symmetric operator.
1
Introduction
Throughout this paper, we assume that
X is a real Banach space with norm k · k,
that X ∗ is the continuous dual of X, and that X and X ∗ are paired by h·, ·i. The closed unit ball
in X is denoted by BX =(cid:8)x ∈ X kxk ≤ 1(cid:9), and N = {1, 2, 3, . . .}.
∗Mathematics, Irving K. Barber School, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail:
[email protected].
†CARMA, University
of Newcastle, Newcastle, New South Wales
2308, Australia.
E-mail:
[email protected]. Distinguished Professor King Abdulaziz University, Jeddah.
‡Mathematics, Irving K. Barber School, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail:
[email protected].
§Mathematics, Irving K. Barber School, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail:
[email protected].
1
We identify X with its canonical image in the bidual space X ∗∗. As always, X × X ∗ and
(X × X ∗)∗ = X ∗ × X ∗∗ are paired via
where (x, x∗) ∈ X × X ∗ and (y∗, y∗∗) ∈ X ∗ × X ∗∗.
h(x, x∗), (y∗, y∗∗)i = hx, y∗i + hx∗, y∗∗i,
Let A : X ⇒ X ∗ be a set-valued operator (also known as multifunction) from X to X ∗, i.e., for
every x ∈ X, Ax ⊆ X ∗, and let gra A = (cid:8)(x, x∗) ∈ X × X ∗ x∗ ∈ Ax(cid:9) be the graph of A. The
domain of A, written as dom A, is dom A = (cid:8)x ∈ X Ax 6= ∅(cid:9) and ran A = A(X) is the range
of A. We say A is a linear relation if gra A is a linear subspace. Now let U × V ⊆ X × X ∗.
We say that A is monotone with respect to U × V , if for every (x, x∗) ∈ (gra A) ∩ (U × V ) and
(y, y∗) ∈ (gra A) ∩ (U × V ), we have
(1)
hx − y, x∗ − y∗i ≥ 0.
Of course, by (classical) monotonicity we mean monotonicity with respect to X ×X ∗. Furthermore,
we say that A is maximally monotone with respect to U × V if A is monotone with respect to
U × V and for every operator B : X ⇒ X ∗ that is monotone with respect to U × V and such that
(gra A)∩(U ×V ) ⊆ (gra B)∩(U ×V ), we necessarily have (gra A)∩(U ×V ) = (gra B)∩(U ×V ). Thus,
(classical) maximal monotonicity corresponds to maximal monotonicity with respect to X × X ∗.
This slightly unusual presentation is required to state our main results; moreover, it yields a more
concise formulation of monotone operators of type (FP).
Now let A : X ⇒ X ∗ be monotone and (x, x∗) ∈ X × X ∗. We say (x, x∗) is monotonically related
to gra A if
hx − y, x∗ − y∗i ≥ 0,
∀(y, y∗) ∈ gra A.
If Z is a real Banach space with continuous dual Z ∗ and a subset S of Z, we denote S⊥ by
S⊥ = (cid:8)z∗ ∈ Z ∗ hz∗, si = 0,
hz, d∗i = 0,
∀s ∈ S(cid:9). Given a subset D of Z ∗, we define D⊥ by D⊥ = (cid:8)z ∈ Z
∀d∗ ∈ D(cid:9) = D⊥ ∩ Z.
The operator adjoint of A, written as A∗, is defined by
gra A∗ =(cid:8)(x∗∗, x∗) ∈ X ∗∗ × X ∗ (x∗, −x∗∗) ∈ (gra A)⊥(cid:9).
Note that the adjoint is always a linear relation with gra A∗ ⊆ X ∗∗ × X ∗ ⊆ X ∗∗ × X ∗∗∗. These
inclusions make it possible to consider monotonicity properties of A∗; however, care is required:
as a linear relation, gra A∗ ⊆ X ∗∗ × X ∗ while as a potential monotone operator we are led to
consider gra A∗ ⊆ X ∗∗ × X ∗∗∗. Now let A : X ⇒ X ∗ be a linear relation. We say that A is skew
if gra A ⊆ gra(−A∗); equivalently, if hx, x∗i = 0, ∀(x, x∗) ∈ gra A. Furthermore, A is symmetric if
gra A ⊆ gra A∗; equivalently, if hx, y∗i = hy, x∗i, ∀(x, x∗), (y, y∗) ∈ gra A.
We now recall three fundamental subclasses of maximally monotone operators.
Definition 1.1 Let A : X ⇒ X ∗ be maximally monotone. Then three key types of monotone
operators are defined as follows.
2
(i) A is of dense type or type (D) (1971, [22]) if for every (x∗∗, x∗) ∈ X ∗∗ × X ∗ with
inf
(a,a∗)∈gra A
ha − x∗∗, a∗ − x∗i ≥ 0,
there exist a bounded net (aα, a∗
to (x∗∗, x∗).
α)α∈Γ in gra A such that (aα, a∗
α)α∈Γ weak*×strong converges
(ii) A is of type negative infimum (NI) (1996, [32]) if
sup
(a,a∗)∈gra A(cid:0)ha, x∗i + ha∗, x∗∗i − ha, a∗i(cid:1) ≥ hx∗∗, x∗i,
∀(x∗∗, x∗) ∈ X ∗∗ × X ∗.
(iii) A is of type Fitzpatrick-Phelps (FP) (1992, [21]) if whenever V is an open convex subset of
X ∗ such that V ∩ ran A 6= ∅, it must follow that A is maximally monotone with respect to
X × V .
Fact 1.2 (See [33, 35, 15].) The following are maximally monotone of type (D), (NI), and (FP).
(i) ∂f , where f : X → ]−∞, +∞] is convex, lower semicontinuous, and proper;
(ii) A : X ⇒ X ∗, where A is maximally monotone and X is reflexive.
These and other relationships known amongst these and other monotonicity notions are described
in [15, Chapter 9]. As we see in [5] and [34, 32, 24], it is now known that the three classes coincide.
Monotone operators have proven to be a key class of objects in both modern Optimization and
Analysis; see, e.g., [12, 13, 14], the books [7, 15, 19, 27, 33, 35, 30, 39, 40, 41] and the references
therein.
Let us now precisely describe the aforementioned Br´ezis-Browder Theorem:
Theorem 1.3 (Br´ezis-Browder in reflexive Banach space [17, 18]) Suppose that X is re-
flexive. Let A : X ⇒ X ∗ be a monotone linear relation such that gra A is closed. Then A is
maximally monotone if and only if the adjoint A∗ is monotone.
In this paper, we generalize the Br´ezis-Browder Theorem to an arbitrary Banach space. (See [36]
for Simons' recent extension of the above result to symmetrically self-dual Banach spaces (SSDB)
spaces as defined in [35, §21].)
Our main result is the following.
Theorem 1.4 (Br´ezis-Browder in general Banach space) Let A : X ⇒ X ∗ be a monotone
linear relation such that gra A is closed. Then the following are equivalent.
(i) A is maximally monotone of type (D).
3
(ii) A is maximally monotone of type (NI).
(iii) A is maximally monotone of type (FP).
(iv) A∗ is monotone.
In Section 2, we collect auxiliary results for future reference and for the reader's convenience. In
Section 3, we provide the key technical step showing that when A∗ is monotone then A is of type
(D). Our central result, the generalized Br´ezis-Browder Theorem (Theorem 1.4), is then proved in
Section 4. Finally, in Section 5 with the necessary proviso that the domain be closed, we establish
further results such as Theorem 5.10 relating to the skew part of the operator. This was motivated
by and extends [2, Theorem 4.1] which studied the case of a bounded linear operator.
Finally, let us mention that we adopt standard convex analysis notation. Given a subset C of
is the weak∗
is the weak∗ closure of E in X ∗∗ with the topology induced by X ∗.
X, int C is the interior of C, C is the norm closure of C. For the set D ⊆ X ∗, D
closure of D. If E ⊆ X ∗∗, E
The indicator function of C, written as ιC, is defined at x ∈ X by
w*
w*
(2)
ιC(x) =(0,
+∞,
if x ∈ C;
otherwise.
For every x ∈ X, the normal cone operator of C at x is defined by NC(x) = (cid:8)x∗ ∈ X ∗
supc∈Chc − x, x∗i ≤ 0(cid:9), if x ∈ C; and NC (x) = ∅, if x /∈ C.
Let f : X → ]−∞, +∞]. Then dom f = f −1(R) is the domain of f , and f ∗ : X ∗ →
[−∞, +∞] : x∗ 7→ supx∈X(hx, x∗i − f (x)) is the Fenchel conjugate of f . The lower semicontinuous
hull of f is denoted by f . We say f is proper if dom f 6= ∅. Let f be proper. The subdifferential
of f is defined by
∂f : X ⇒ X ∗ : x 7→ {x∗ ∈ X ∗ (∀y ∈ X) hy − x, x∗i + f (x) ≤ f (y)}.
For ε ≥ 0,
the ε -- subdifferential of f is defined by ∂εf : X ⇒ X ∗ : x 7→ (cid:8)x∗ ∈ X ∗
(∀y ∈ X) hy − x, x∗i + f (x) ≤ f (y) + ε(cid:9). Note that ∂f = ∂0f . We denote by J := JX the du-
2 k · k2 mapping X to X ∗. For the properties of
ality map, i.e., the subdifferential of the function 1
J, see [27, Example 2.26].
Let (z, z∗) ∈ X × X ∗ and F : X × X ∗ → ]−∞, +∞]. Then F(z,z∗) : X × X ∗ → ]−∞, +∞] [25, 35]
is defined by
F(z,z∗)(x, x∗) = F (z + x, z∗ + x∗) −(cid:0)hx, z∗i + hz, x∗i + hz, z∗i(cid:1)
= F (z + x, z∗ + x∗) − hz + x, z∗ + x∗i + hx, x∗i,
(3)
∀(x, x∗) ∈ X × X ∗.
Let now Y be another real Banach space. We set PX : X×Y → X : (x, y) 7→ x. Let F1, F2 : X×Y →
]−∞, +∞]. Then the partial inf-convolution F1(cid:3)2F2 is the function defined on X × Y by
F1(cid:3)2F2 : (x, y) 7→ inf
v∈Y
F1(x, y − v) + F2(x, v).
4
2 Prerequisite results
Fact 2.1 (See [26, Proposition 2.6.6(c)] or [31, Theorem 4.7 and Theorem 3.12].) Let C be a
subspace of X, and D be a subspace of X ∗. Then
(C ⊥)⊥ = C and
(D⊥)⊥ = D
w*
.
Fact 2.2 (Rockafellar) (See [29, Theorem 3(b)], [35, Theorem 18.1] or [39, Theorem 2.8.7(iii)].)
Let f, g : X → ]−∞, +∞] be proper convex functions. Assume that there exists a point x0 ∈
dom f ∩ dom g such that g is continuous at x0. Then
∂(f + g) = ∂f + ∂g.
Fact 2.3 (Brøndsted-Rockafellar) (See [39, Theorem 3.1.2 or Theorem 3.1.4(ii)].) Let f : X →
]−∞, +∞] be a proper lower semicontinuous and convex function and x∗ ∈ dom f ∗. Then there
exists a sequence (xn, x∗
n)n∈N in gra ∂f such that x∗
n → x∗.
Fact 2.4 (Attouch-Br´ezis) (See [1, Theorem 1.1] or [35, Remark 15.2].) Let f, g : X →
]−∞, +∞] be proper lower semicontinuous and convex. Assume that Sλ>0 λ [dom f − dom g] is
a closed subspace of X. Then
(f + g)∗(z∗) = min
y∗∈X ∗
[f ∗(y∗) + g∗(z∗ − y∗)] ,
∀z∗ ∈ X ∗.
Fact 2.5 (Simons and Zalinescu) (See [37, Theorem 4.2] or [35, Theorem 16.4(a)].) Let Y be a
real Banach space and F1, F2 : X × Y → ]−∞, +∞] be proper, lower semicontinuous, and convex.
Assume that for every (x, y) ∈ X × Y ,
(F1(cid:3)2F2)(x, y) > −∞
and that Sλ>0 λ [PX dom F1 − PX dom F2] is a closed subspace of X. Then for every (x∗, y∗) ∈
X ∗ × Y ∗,
1 (x∗ − u∗, y∗) + F ∗
2 (u∗, y∗)] .
(F1(cid:3)2F2)∗(x∗, y∗) = min
u∗∈X ∗
[F ∗
The following result was first established in [11, Theorem 7.4]. Now we give a new proof.
Fact 2.6 (Borwein) Let A, B : X ⇒ X ∗ be linear relations such that gra A and gra B are closed.
Assume that dom A − dom B is closed. Then
Proof. We have
(4)
(A + B)∗ = A∗ + B∗.
ιgra(A+B) = ιgra A(cid:3)2ιgra B.
5
Let (x∗∗, x∗) ∈ X ∗∗ × X ∗. Since gra A and gra B are closed convex, ιgra A and ιgra B are proper
lower semicontinuous and convex. Then by Fact 2.5 and (4), there exists y∗ ∈ X ∗ such that
ιgra(A+B)∗ (x∗∗, x∗) = ι
(cid:0) gra(A+B)(cid:1)⊥(−x∗, x∗∗)
gra(A+B)(−x∗, x∗∗)
gra A(y∗, x∗∗) + ι∗
= ι∗
= ι∗
= ι(gra A)⊥(y∗, x∗∗) + ι(gra B)⊥ (−x∗ − y∗, x∗∗)
= ιgra A∗(x∗∗, −y∗) + ιgra B∗(x∗∗, x∗ + y∗)
= ιgra(A∗+B∗)(x∗∗, x∗).
(since gra(A + B) is a subspace)
gra B(−x∗ − y∗, x∗∗)
(5)
Then we have gra(A + B)∗ = gra(A∗ + B∗) and hence (A + B)∗ = A∗ + B∗.
(cid:4)
Fact 2.7 (Simons) (See [35, Lemma 19.7 and Section 22].) Let A : X ⇒ X ∗ be a monotone
operator such that gra A is convex with gra A 6= ∅. Then the function
(6)
g : X × X ∗ → ]−∞, +∞] : (x, x∗) 7→ hx, x∗i + ιgra A(x, x∗)
is proper and convex.
We also recall the somewhat more precise version of Theorem 1.3.
Fact 2.8 (Br´ezis and Browder) (See [18, Theorem 2], or [16, 17, 36, 38].) Suppose that X is
reflexive. Let A : X ⇒ X ∗ be a monotone linear relation such that gra A is closed. Then the
following are equivalent.
(i) A is maximally monotone.
(ii) A∗ is maximally monotone.
(iii) A∗ is monotone.
This has a recent non-reflexive counterpart:
Fact 2.9 (See [4, Theorem 3.1].) Let A : X ⇒ X ∗ be a maximally monotone linear relation. Then
the following are equivalent.
(i) A is of type (D).
(ii) A is of type (NI).
(iii) A is of type (FP).
(iv) A∗ is monotone
6
Comparing of Fact 2.9 and Fact 2.8, we observe that the hypothesis in the latter (maximality
of A) is more restrictive than in the former (closedness of the graph). In [4, Theorem 3.1] we were
unable to attack this issue. The result of the next section redresses our lacuna.
Now let us cite some basic properties of linear relations.
The following result appeared in Cross' book [20]. We give new proofs of (iv) -- (vi). The proof of
the (vi) below was adapted from [10, Remark 2.2].
Fact 2.10 Let A : X ⇒ X ∗ be a linear relation. Then the following hold.
(i) Ax = x∗ + A0,
∀x∗ ∈ Ax.
(ii) A(αx + βy) = αAx + βAy, ∀(α, β) ∈ R2 r {(0, 0)}, ∀x, y ∈ dom A.
(iii) hA∗x, yi = hx, Ayi is a singleton, ∀x ∈ dom A∗, ∀y ∈ dom A.
(iv) (dom A)⊥ = A∗0 is (weak∗) closed and dom A = (A∗0)⊥.
(v) If gra A is closed, then (dom A∗)⊥ = A0 and dom A∗w*
(vi) If dom A is closed, then dom A∗ = ( ¯A0)⊥ and thus dom A∗ is (weak∗) closed, where ¯A is the
= (A0)⊥.
linear relation whose graph is the closure of the graph of A.
Proof. (i): See [20, Proposition I.2.8(a)]. (ii): See [20, Corollary I.2.5]. (iii): See [20, Proposition
III.1.2].
(iv): We have
x∗ ∈ A∗0 ⇔ (x∗, 0) ∈ (gra A)⊥ ⇔ x∗ ∈ (dom A)⊥.
Hence (dom A)⊥ = A∗0 and thus A∗0 is weak∗ closed. By Fact 2.1, dom A = (A∗0)⊥.
(v): Using Fact 2.1,
x∗ ∈ A0 ⇔ (0, x∗) ∈ gra A =h(gra A)⊥i⊥
=(cid:2)gra −(A∗)−1(cid:3)⊥ ⇔ x∗ ∈ (dom A∗)⊥.
Hence (dom A∗)⊥ = A0 and thus, by Fact 2.1, dom A∗w*
= (A0)⊥.
(vi): Let ¯A be the linear relation whose graph is the closure of the graph of A. Then dom A =
dom ¯A and A∗ = ¯A∗. Then by Fact 2.4,
ιX ∗×( ¯A0)⊥ = ι∗
{0}× ¯A0 =(cid:0)ιgra ¯A + ι{0}×X ∗(cid:1)∗ = ιgra(− ¯A∗)−1 (cid:3) ιX ∗×{0} = ιX ∗×dom ¯A∗.
It is clear that dom A∗ = dom ¯A∗ = ( ¯A0)⊥ is weak∗ closed, hence closed.
(cid:4)
7
3 A key result
The proof of Proposition 3.1 below was partially inspired by that of [4, Theorem 3.1].
Proposition 3.1 Let A : X ⇒ X ∗ be a monotone linear relation such that gra A is closed and A∗
is monotone. Then A is maximally monotone of type (D).
Proof. By Fact 2.9, it suffices to show that A is maximally monotone. Let (z, z∗) ∈ X × X ∗.
Assume that
(7)
Define
(z, z∗) is monotonically related to gra A.
F : X × X ∗ → ]−∞, +∞] : (x, x∗) 7→ ιgra A(x, x∗) + hx, x∗i.
Fact 2.7 implies that F is convex and since gra A is closed, F is also proper, lower semicontinuous.
Recalling (3), note that
(8)
F(z,z∗) : (x, x∗) 7→ ιgra A(z + x, z∗ + x∗) + hx, x∗i
is proper, lower semicontinuous, and convex. Set
(9)
Then
(10)
G(x, x∗) := F(z,z∗)(x, x∗) + 1
2 kxk2 + 1
2 kx∗k2,
∀(x, x∗) ∈ X × X ∗.
inf G = −G∗(0, 0).
By (8), inf G ≥ 0. Then (0, 0) ∈ dom G∗. By Fact 2.3, there exists a sequence
(11)
such that
(12)
Thus,
(13)
(cid:0)(an, a∗
n), (y∗
n, y∗∗
n )(cid:1)n∈N in gra ∂G
(y∗
n, y∗∗
n ) → (0, 0).
By Fact 2.2 and (11), there exists (v∗
n, v∗∗
n ) ∈ Jan × JX ∗ a∗
n such that
Kn := max(cid:8)ky∗
nk, ky∗∗
n k(cid:9) → 0.
(14)
(y∗
n, y∗∗
n ) ∈ ∂F(z,z∗)(an, a∗
n) + (v∗
n, v∗∗
n ),
∀n ∈ N.
By (14), (8), and [39, Theorem 3.2.4(vi)&(ii)], there exists a sequence (z∗
that
n, z∗∗
n )n∈N in (gra A)⊥ such
(15)
(y∗
n, y∗∗
n ) = (a∗
n, an) + (z∗
n, z∗∗
n ) + (v∗
n, v∗∗
n ),
∀n ∈ N.
8
Since A∗ is monotone and (z∗∗
n) ∈ gra(−A∗), it follows from (15) that
n , z∗
n i + han, a∗
ni
hy∗
n, y∗∗
+ ha∗
−(cid:2)hy∗
n, ani + hy∗∗
n i + hv∗
n, v∗∗
n − a∗
n − v∗
n, y∗∗
n, z∗∗
n i ≤ 0,
ni(cid:3) −(cid:2)hy∗
n , a∗
n, v∗∗
n i + han, v∗
n, v∗∗
ni
n − an − v∗∗
n i
∀n ∈ N.
= hy∗
= hz∗
n i + hv∗
n, y∗∗
n i(cid:3)
(16)
Since (v∗
n, v∗∗
n ) ∈ Jan × JX ∗ a∗
n, by (16), we have
hy∗
n, y∗∗
n i + han, a∗
ni
(17)
−(cid:2)ky∗
+ ka∗
nk · kank + ky∗∗
nk2 − kank · ka∗
n k · ka∗
nk + kank2 ≤ 0,
nk(cid:3) −(cid:2)ky∗
∀n ∈ N.
nk · ka∗
nk + kank · ky∗∗
n k(cid:3)
Then by (17) and (13),
n + han, a∗
− K 2
+ 1
ni − Kn(cid:2)kank + ka∗
nk(cid:3) − Kn(cid:2)ka∗
∀n ∈ N.
nk + kank(cid:3)
(18)
Hence
nk2 + kank2(cid:3) ≤ 0,
2(cid:2)ka∗
ni − 2Kn(cid:2)kank + ka∗
− K 2
n + han, a∗
Set (xn, x∗
n) := (z + an, z∗ + a∗
n),
∀n ∈ N. Then by (8), we have
n) = ιgra A(z + an, z∗ + a∗
F(z,z∗)(an, a∗
= ιgra A(xn, x∗
n) + hxn − z, x∗
n) + han, a∗
ni
n − z∗i.
nk(cid:3) + 1
4(cid:2)ka∗
nk + kank(cid:3)2 ≤ 0,
∀n ∈ N.
(19)
(20)
By (14) and (20),
(21)
(xn, x∗
n) ∈ gra A,
∀n ∈ N.
Then by (21) and (7), we have
(22)
Combining (19) and (22),
han, a∗
ni = hxn − z, x∗
n − z∗i ≥ 0,
∀n ∈ N.
(23)
equivalently,
(24)
In view of (13),
(25)
1
4(cid:0)ka∗
n + 2Kn(cid:0)kank + ka∗
nk(cid:1),
nk + kank(cid:1)2 ≤ K 2
nk + kank − 4Kn(cid:1)2 ≤ 20K 2
(cid:0)ka∗
n,
∀n ∈ N.
∀n ∈ N;
kank + ka∗
nk → 0.
Thus (an, a∗
see (z, z∗) ∈ gra A. Therefore, A is maximally monotone.
n) → (0, 0) and hence (xn, x∗
n) → (z, z∗). Finally, by (21) and since gra A is closed, we
(cid:4)
9
Example 3.2 Let A : X ⇒ X ∗ be a monotone linear relation such that gra A is closed. We note
that we cannot guarantee the maximal monotonicity of A even if A is at most single-valued and
densely defined. To see this, suppose that X = ℓ2, and that A : ℓ2 ⇒ ℓ2 is given by
(26)
Ax := (cid:18)Pi<n xi −Pi>n xi(cid:19)n∈N
2 xn(cid:19)n∈N
where dom A := nx := (xn)n∈N ∈ ℓ2 Pi≥1 xi = 0,(cid:18)Pi≤n xi(cid:19)n∈N
single-valued linear relation. Now [9, Propositions 3.6] states that
=(cid:18)Xi<n
xi + 1
2
(27)
where
,
2 xn +Xi>n
A∗x =(cid:18) 1
xi(cid:19)n∈N
x = (xn)n∈N ∈ dom A∗ =(cid:26)x = (xn)n∈N ∈ ℓ2 (cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)Xi>n
,
∀x = (xn)n∈N ∈ dom A,
∈ ℓ2o. Then A is an at most
xi(cid:19)n∈N
∈ ℓ2(cid:27).
Moreover, [9, Propositions 3.2, 3.5, 3.6 and 3.8], [28, Theorem 2.5] and Fact 2.8 show that:
(i) A is maximally monotone and skew;
(ii) dom A is dense and dom A $ dom A∗;
(iii) A∗ is maximally monotone, but not skew;
(iv) −A is not maximally monotone.
Hence, −A is monotone with dense domain and gra(−A) is closed, but nonetheless −A is not
maximally monotone.
(cid:4)
4 The general Br´ezis-Browder theorem
We may now pack everything together. For ease we repeat Theorem 1.4:
Theorem 4.1 (Br´ezis-Browder in general Banach space) Let A : X ⇒ X ∗ be a monotone
linear relation such that gra A is closed. Then the following are equivalent.
(i) A is maximally monotone of type (D).
(ii) A is maximally monotone of type (NI).
(iii) A is maximally monotone of type (FP).
10
(iv) A∗ is monotone.
Proof. Directly combine Fact 2.9 and Proposition 3.1.
(cid:4)
The original Br´ezis and Browder result follows.
Corollary 4.2 (Br´ezis and Browder) Suppose that X is reflexive. Let A : X ⇒ X ∗ be a mono-
tone linear relation such that gra A is closed. Then the following are equivalent.
(i) A is maximally monotone.
(ii) A∗ is maximally monotone.
(iii) A∗ is monotone.
Proof. "(i)⇔(iii)": Apply Theorem 4.1 and Fact 1.2 directly.
"(ii)⇒(iii)": Clear.
"(iii)⇒(ii)": Since gra A is closed, (A∗)∗ = A. Apply Theorem 4.1 to A∗ .
(cid:4)
In the case of a skew operator we can add maximality of the adjoint and so we prefigure results
of the next section:
Corollary 4.3 Let A : X ⇒ X ∗ be a skew operator such that gra A is closed. Then the following
are equivalent.
(i) A is maximally monotone of type (D).
(ii) A∗ is monotone.
(iii) A∗ is maximally monotone with respect to X ∗∗ × X ∗.
Proof. By Theorem 4.1, it only remains to show
"(ii)⇒(iii)": Let (z∗∗, z∗) ∈ X ∗∗ × X ∗ be monotonically related to gra A∗. Since gra(−A) ⊆
gra A∗, (z∗∗, z∗) is monotonically related to gra(−A). Thus (z∗, z∗∗) ∈ [gra(−A)]⊥ since gra A is
linear. Hence (z∗∗, z∗) ∈ gra A∗. Hence A∗ is maximally monotone.
(cid:4)
Remark 4.4 We cannot say A∗ is maximally monotone with respect to X ∗∗ × X ∗∗∗ in Corol-
lary 4.3(iii): indeed, let A be defined by
gra A = {0} × X ∗.
Then gra A∗ = {0} × X ∗. If X is not reflexive, then X ∗ $ X ∗∗∗ and so gra A∗ is a proper subset
of {0} × X ∗∗∗. Hence A∗ is not maximally monotone with respect to X ∗∗ × X ∗∗∗ although A is
maximally monotone of type (D) (since A = N{0} by Fact 1.2).
In the next section, we turn to the question of how the skew part of the adjoint behaves.
11
5 Decomposition of monotone linear relations
Let us first gather some basic properties about monotone linear relations, and conditions for them
to be maximally monotone.
The next three propositions were proven in reflexive spaces in [8, Proposition 2.2]. We adjust
the proofs to a general Banach space setting.
Proposition 5.1 (Monotone linear relations) Let A : X ⇒ X ∗ be a linear relation. Then the
following hold.
(i) Suppose A is monotone. Then dom A ⊆ (A0)⊥ and A0 ⊆ (dom A)⊥; consequently, if gra A is
closed, then dom A ⊆ dom A∗w*
∩ X and A0 ⊆ A∗0.
(ii) (∀x ∈ dom A)(∀z ∈ (A0)⊥) hz, Axi is single-valued.
(iii) (∀z ∈ (A0)⊥) dom A → R : y 7→ hz, Ayi is linear.
(iv) If A is monotone, then (∀x ∈ dom A) hx, Axi is single-valued.
(v) A is monotone ⇔ (∀x ∈ dom A) infhx, Axi ≥ 0.
(vi) If (x, x∗) ∈ (dom A) × X ∗ is monotonically related to gra A and x∗
0 ∈ Ax, then x∗ − x∗
0 ∈
(dom A)⊥.
Proof. (i): Pick x ∈ dom A. Then there exists x∗ ∈ X ∗ such that (x, x∗) ∈ gra A. By monotonicity
of A and since {0} × A0 ⊆ gra A, we have hx, x∗i ≥ suphx, A0i. Since A0 is a linear subspace,
we obtain x⊥A0. This implies dom A ⊆ (A0)⊥ and A0 ⊆ (dom A)⊥.
If gra A is closed, then
Fact 2.10(v)&(iv) yields dom A ⊆ (A0)⊥ ⊆ (A0)⊥ = dom A∗w*
and A0 ⊆ A∗0.
(ii): Take x ∈ dom A, x∗ ∈ Ax, and z ∈ (A0)⊥. By Fact 2.10(i), hz, Axi = hz, x∗ + A0i = hz, x∗i.
(iii): Take z ∈ (A0)⊥. By (ii), (∀y ∈ dom A) hz, Ayi is single-valued. Now let x, y be in dom A,
and let α, β be in R. If (α, β) = (0, 0), then hz, A(αx + βy)i = hz, A0i = 0 = αhz, Axi + βhz, Ayi.
And if (α, β) 6= (0, 0), then Fact 2.10(ii) yields hz, A(αx + βy)i = hz, αAx + βAyi = αhz, Axi +
βhz, Ayi. This verifies linearity.
(iv): Apply (i)&(ii).
(v): "⇒": This follows from the fact that (0, 0) ∈ gra A. "⇐": If x and y belong to dom A, then
Fact 2.10(ii) yields hx − y, Ax − Ayi = hx − y, A(x − y)i ≥ 0.
(vi): Let (x, x∗) ∈ dom A × X ∗ be monotonically related to gra A, and take x∗
0 + v∗ ∈ A(x + v) (by Fact 2.10(ii)); hence, hx − (x + v), x∗ − (x∗
0 ∈ Ax. For every
0 + v∗)i ≥ 0
0i. Now take λ > 0 and replace (v, v∗) in the last inequality by
(v, v∗) ∈ gra A, we have x∗
and thus hv, v∗i ≥ hv, x∗ − x∗
12
(λv, λv∗). Then divide by λ and let λ → 0+ to see that 0 ≥ suphdom A, x∗ − x∗
linear, it follows that x∗ − x∗
0 ∈ (dom A)⊥.
0i. Since dom A is
(cid:4)
We define the symmetric part and the skew part of A via
(28)
A+ := 1
2 A + 1
2 A∗
and A◦ := 1
2 A − 1
2 A∗,
respectively. It is easy to check that A+ is symmetric and that A◦ is skew.
Proposition 5.2 (Maximally monotone linear relations) Let A : X ⇒ X ∗ be a monotone
linear relation. Then the following hold.
(i) If A is maximally monotone, then (dom A)⊥ = A0 and hence dom A = (A0)⊥.
(ii) If dom A is closed, then: A is maximally monotone ⇔ (dom A)⊥ = A0.
(iii) If A is maximally monotone, then dom A∗w*
A◦0 = (dom A)⊥ is (weak∗) closed.
∩ X = dom A = (A0)⊥, and A0 = A∗0 = A+0 =
(iv) If A is maximally monotone and dom A is closed, then dom A∗ ∩ X = dom A.
(v) If A is maximally monotone and dom A ⊆ dom A∗, then A = A+ + A◦, A+ = A − A◦, and
A◦ = A − A+.
(vi) If A is maximally monotone and dom A is closed, then both A+ and A◦ are maximally mono-
tone.
(vii) If A is maximally monotone and dom A is closed, then A∗ = (A+)∗ + (A◦)∗.
Proof. (i): Since A + Ndom A = A + (dom A)⊥ is a monotone extension of A and A is maximally
monotone, we must have A + (dom A)⊥ = A. Then A0 + (dom A)⊥ = A0. As 0 ∈ A0, (dom A)⊥ ⊆
A0. Combining with Proposition 5.1(i), we have (dom A)⊥ = A0. By Fact 2.1, dom A = (A0)⊥.
(ii): "⇒": Clear from (i). "⇐": The assumptions and Fact 2.1 imply that dom A = dom A =
(cid:2)(dom A)⊥(cid:3)⊥ = (A0)⊥. Let (x, x∗) be monotonically related to gra A. We have infhx−0, x∗ −A0i ≥
0. Then we have x ∈ (A0)⊥ and hence x ∈ dom A. Then by Proposition 5.1(vi) and Fact 2.10(i),
x∗ ∈ Ax. Hence A is maximally monotone.
(iii): By (i) and Fact 2.10(iv), A0 = (dom A)⊥ = A∗0 is weak∗ closed and thus A+0 = A◦0 =
A0 = (dom A)⊥. Then by Fact 2.10(v) and (i), dom A∗w*
∩ X = (A0)⊥ = dom A.
(iv): Combine (iii) with Fact 2.10(vi).
(v): We show only the proof of A = A+ + A◦ as the other two proofs are analogous. Clearly,
dom A+ = dom A◦ = dom A ∩ dom A∗ = dom A. Let x ∈ dom A, and x∗ ∈ Ax and y∗ ∈ A∗x.
We write x∗ = x∗+y∗
∈ (A+ + A◦)x. Then, by (iii) and Fact 2.10(i), Ax = x∗ + A0 =
x∗ + (A+ + A◦)0 = (A+ + A◦)x. Therefore, A = A+ + A◦.
2 + x∗−y∗
2
13
(vi): By (iv),
(29)
Hence, by (iii),
dom A+ = dom A◦ = dom A is closed.
(30)
A◦0 = A+0 = A0 = (dom A)⊥ = (dom A+)⊥ = (dom A◦)⊥.
Since A is monotone, so are A+ and A◦. Combining (29), (30), and (ii), we deduce that A+ and
A◦ are maximally monotone.
(vii): By (iv)&(v),
(31)
A = A+ + A◦.
Then by (vi), (iv), and Fact 2.6, A∗ = (A+)∗ + (A◦)∗.
(cid:4)
For a monotone linear relation A : X ⇒ X ∗ it will be convenient to define -- as in, e.g., [3] -- a
generalized quadratic form
(∀x ∈ X)
qA(x) =( 1
2 hx, Axi,
+∞,
if x ∈ dom A;
otherwise.
We write qA for the lower semicontinuous hull of qA.
Proposition 5.3 Let A : X ⇒ X ∗ be a monotone linear relation, let x and y be in dom A, and let
λ ∈ R. Then qA is single-valued, qA ≥ 0 and
(32)
λqA(x) + (1 − λ)qA(y) − qA(λx + (1 − λ)y) = λ(1 − λ)qA(x − y)
= 1
2 λ(1 − λ)hx − y, Ax − Ayi.
Consequently, qA is convex.
Proof. Proposition 5.1(iv)&(v) show that qA is single-valued and that qA ≥ 0. Combining with
Proposition 5.1(i)&(iii), we obtain (32). Therefore, qA is convex.
(cid:4)
As in the classical case, qA allows us to connect properties of A+ to those of A and A∗.
Proposition 5.4 Let A : X ⇒ X ∗ be a monotone linear relation. Then the following hold.
(i) qA + ιdom A+ = qA+ and thus qA+ is convex.
(ii) gra A+ ⊆ gra ∂qA. If A+ is maximally monotone, then A+ = ∂qA.
(iii) If A is maximally monotone and dom A is closed, then A+ = ∂qA.
(iv) If A is maximally monotone, then A∗X is monotone.
14
(v) If A is maximally monotone and dom A is closed, then A∗X is maximally monotone.
Proof. Let x ∈ dom A+.
(i): By Fact 2.10(iii) and Proposition 5.1(iv), qA+ = qAdom A+. Then by Proposition 5.3, qA+ is
convex. Let y ∈ dom A. Then by Fact 2.10(iii),
(33)
0 ≤ 1
2 hAx − Ay, x − yi = 1
2 hAy, yi + 1
2 hAx, xi − hA+x, yi,
we have qA(y) ≥ hA+x, yi − qA(x). Take the lower semicontinuous hull of qA at y to deduce that
qA(y) ≥ hA+x, yi − qA(x). For y = x, we have qA(x) ≥ qA(x). On the other hand, qA ≤ qA.
Altogether, qA(x) = qA(x) = qA+(x). Thus (i) holds.
(ii): Let y ∈ dom A. By (33) and (i),
(34)
qA(y) ≥ qA(x) + hA+x, y − xi = qA(x) + hA+x, y − xi.
Since dom qA ⊆ dom qA = dom A, by (34), qA(z) ≥ qA(x) + hA+x, z − xi,
A+x ⊆ ∂qA(x). If A+ is maximally monotone, then A+ = ∂qA. Thus (ii) holds.
∀z ∈ dom qA. Hence
(iii): Combine Proposition 5.2(vi) with (ii).
(iv): Suppose to the contrary that A∗X is not monotone. By Proposition 5.1(v), there exists
(x0, x∗
0) ∈ gra A∗ with x0 ∈ X such that hx0, x∗
0i < 0. Now we have
(35)
h−x0 − y, x∗
= −hx0, x∗
0 − y∗i = −hx0, x∗
0i + hy, y∗i > 0,
0i + hy, y∗i + hx0, y∗i − hy, x∗
0i
∀(y, y∗) ∈ gra A.
Thus, (−x0, x∗
gra A. Then h−x0 − (−x0), x∗
0) is monotonically related to gra A. By maximal monotonicity of A, (−x0, x∗
0) ∈
0 − x∗
0i = 0, which contradicts (35). Hence A∗X is monotone.
(v): By Fact 2.10(vi), dom A∗X = (A0)⊥ and thus dom A∗X is closed. By Fact 2.1 and
Proposition 5.2(i), (dom A∗X )⊥ = ((A0)⊥)⊥ = A0
= A0. Then by Proposition 5.2(iii),
(dom A∗X )⊥ = A∗0 = A∗X 0. Applying (iv) and Proposition 5.2(ii), we see that A∗X is maximally
monotone.
(cid:4)
w*
The proof of Proposition 5.4(iv) was borrowed from [18, Theorem 2]. Results very similar
to Proposition 5.4(i)&(ii) are verified in [38, Proposition 18.9]. The proof of the next Theo-
rem 5.5(i)⇒(ii) was partially inspired by that of [2, Theorem 4.1(v)⇒(vi)]. When the domain
of A is closed we can obtain additional information about the skew part of A.
Theorem 5.5 (Monotone relations with closed graph and domain) Let A : X ⇒ X ∗ be a
monotone linear relation such that gra A is closed and dom A is closed. Then the following are
equivalent.
(i) A is maximally monotone of type (D).
15
(ii) A◦ is maximally monotone of type (D) with respect to X × X ∗ and A∗0 = A0.
(iii) (A◦)∗ is maximally monotone with respect to X ∗∗ × X ∗ and A∗0 = A0.
(iv) (A◦)∗ is monotone and A∗0 = A0.
(v) A∗ is monotone.
(vi) A∗ is maximally monotone with respect to X ∗∗ × X ∗.
Proof. "(i)⇒(ii)": By Fact 2.9,
(36)
A∗ is monotone.
By Proposition 5.4(iii) and Fact 1.2,
(37)
By Fact 2.9,
(38)
Now we show that
(39)
A+ is maximally monotone of type (D).
(A+)∗ is monotone.
(A◦)∗ is monotone.
Proposition 5.2(vii) implies
(40)
A∗ = (A+)∗ + (A◦)∗.
Since A is maximally monotone and dom A is closed, Proposition 5.2(vi) implies that A◦ is maxi-
mally monotone. Hence gra(A◦) is closed. On the other hand, again since A is maximally monotone
and dom A is closed, Proposition 5.2(iv) yields dom(A◦) = dom A is closed. Altogether, and com-
bining with Fact 2.10(vi) applied to A◦, we obtain dom(A◦)∗ = (A◦0)⊥. Furthermore, since
A0 = A◦0 by Proposition 5.2(iii), we have (A0)⊥ = (A◦0)⊥. Moreover, applying Fact 2.10(vi) to
A, we deduce that dom A∗ = (A0)⊥. Therefore,
(41)
Similarly, we have
(42)
dom(A◦)∗ = (A◦0)⊥ = (A0)⊥ = dom A∗.
dom(A+)∗ = dom A∗.
Take (x∗∗, x∗) ∈ gra(A◦)∗. By (40) and (41), there exist a∗, b∗ ∈ X ∗ such that
(43)
and
(44)
(x∗∗, a∗) ∈ gra A∗, (x∗∗, b∗) ∈ gra(A+)∗
a∗ = b∗ + x∗.
16
Since A+ is symmetric, gra A+ ⊆ gra(A+)∗. Thus, by (38), (x∗∗, b∗) is monotonically related
to gra A+. By (37), there exist a bounded net (aα, b∗
α)α∈Γ
α) ∈ gra(A+)∗. By (42) and (40), there exist
weak*×strong converges to (x∗∗, b∗). Thus (aα, b∗
a∗
α ∈ A∗aα, c∗
α)α∈Γ in gra A+ such that (aα, b∗
α ∈ (A◦)∗aα such that
(45)
Thus by Fact 2.10(iii),
a∗
α = b∗
α + c∗
α,
∀α ∈ Γ.
(46)
haα, c∗
αi = hA◦aα, aαi = 0,
∀α ∈ Γ.
Hence for every α ∈ Γ, (−aα, c∗
α) is monotonically related to gra A◦. By Proposition 5.2(vi),
(47)
(−aα, c∗
α) ∈ gra A◦,
∀α ∈ Γ.
By (36) and (43), we have
0 ≤ hx∗∗ − aα, a∗ − a∗
= hx∗∗ − aα, a∗ − b∗
= hx∗∗ − aα, a∗ − b∗
= hx∗∗ − aα, a∗ − b∗
αi = hx∗∗ − aα, a∗ − b∗
α − c∗
αi
αi − hx∗∗, c∗
αi + haα, c∗
αi
αi − hx∗∗, c∗
(by (46))
αi
αi + hx∗, aαi
(by (47) and (x∗∗, x∗) ∈ gra(A◦)∗).
(by (45))
(48)
Taking the limit in (48) along with aα
w*
⇁ x∗∗ and b∗
α → b∗, we have
hx∗∗, x∗i ≥ 0.
Hence (A◦)∗ is monotone and thus (39) holds. Combining (39), Proposition 5.2(vi) and Fact 2.9,
we see that A◦ is of type (D).
"(ii)⇒(iii)⇒(iv)": Apply Corollary 4.3 to A◦.
"(iv)⇒(v)": By Fact 2.10(iv) and Proposition 5.2(ii), A is maximally monotone. Then by
Proposition 5.2(vii) and Proposition 5.4(iii), we have
(49)
A∗ = (A+)∗ + (A◦)∗ and A+ = ∂qA.
Then A+ is of type (D) by Fact 1.2, and hence (A+)∗ is monotone by Fact 2.9. Thus, by the
assumption and (49), we have A∗ is monotone.
"(v)⇒(vi)": By Proposition 3.1, A is maximally monotone. Then by Fact 2.10(vi) and Proposi-
tion 5.2(iii),
(50)
Then by Fact 2.1 and Fact 2.10(iv),
(51)
dom A∗ = (A∗0)⊥.
[dom A∗]⊥ = A∗0.
17
Let (x∗∗, x∗) ∈ X ∗∗ × X ∗ be monotonically related to gra A∗. Because {0} × A∗0 ⊆ gra A∗, we have
infhx∗∗, x∗ − A∗0i ≥ 0. Since A∗0 is a subspace, x∗∗ ∈ (A∗0)⊥. Then by (50),
(52)
x∗∗ ∈ dom A∗.
Take (x∗∗, x∗∗
hence (x∗∗ + λa∗∗, x∗
0 ) ∈ gra A∗ and λ > 0. For every (a∗∗, a∗) ∈ gra A∗, we have (λa∗∗, λa∗) ∈ gra A∗ and
0 + λa∗) ∈ gra A∗ (since gra A∗ is a subspace). Thus
λha∗∗, x∗
0 + λa∗ − x∗i = hx∗∗ + λa∗∗ − x∗∗, x∗
0 + λa∗ − x∗i ≥ 0.
0i. Then let λ → 0+ to see that 0 ≥
Now divide by λ to obtain λha∗∗, a∗i ≥ ha∗∗, x∗ − x∗
0 + A∗0 ⊆ A∗x∗∗ + A∗0.
suphdom A∗, x∗ − x∗
Then there exists (0, z∗) ∈ gra A∗ such that (x∗∗, x∗ − z∗) ∈ gra A∗. Since gra A∗ is s a subspace,
(x∗∗, x∗) = (0, z∗) + (x∗∗, x∗ − z∗) ∈ gra A∗. Hence A∗ is maximally monotone with respect to
X ∗∗ × X ∗.
0 ∈ (dom A∗)⊥. By (51), x∗ ∈ x∗
0i. Thus, x∗ − x∗
"(vi)⇒(i)": Apply Proposition 3.1 directly.
(cid:4)
The next three examples show the need for various of our auxiliary hypotheses.
Example 5.6 We cannot remove the condition that A∗0 = A0 in Theorem 5.5(iv). For example,
suppose that X = R2 and set e1 = (1, 0), e2 = (0, 1). We define A : X ⇒ X by
gra A = span{e1} × {0} so that gra A∗ = X × span{e2}.
Then A is monotone, dom A is closed, and gra A is closed. Thus
(53)
and so
gra A◦ = span{e1} × span{e2}
gra(A◦)∗ = span{e2} × span{e1}.
Hence (A◦)∗ is monotone, but A is not maximally monotone because gra A $ gra NX.
(cid:4)
Example 5.7 We cannot replace that "dom A is closed" by that "dom A is dense" in the statement
of Theorem 5.5. For example, let X, A be defined as in Example 3.2 and consider the operator
A∗. Example 3.2(iii)&(ii) state that A∗ is maximally monotone with dense domain; hence, gra A∗
is closed. Moreover, by Example 3.2(i),
(54)
Hence
(55)
(A∗)◦ = −A.
[(A∗)◦]∗ = −A∗.
Thus [(A∗)◦]∗ is not monotone by Example 3.2(iii); even though A∗ is a classically maximally
monotone and densely defined linear operator.
(cid:4)
18
Example 5.8 We cannot remove the condition that (A◦)∗ is monotone in Theorem 5.5(iv). For
It satisfies X = ℓ1, dom A = X,
example, consider the Gossez operator A (see [23] and [2]).
A◦ = A, A0 = {0} = A∗0, yet A∗ is not monotone.
(cid:4)
Remark 5.9 Let A : X ⇒ X ∗ be a maximally monotone linear relation.
(i) In general, (A∗)◦ 6= (A◦)∗. To see that, let X, A be as in Example 3.2 again. By Exam-
ple 3.2(i), we have
Hence (A∗)◦ 6= (A◦)∗ by Example 3.2(ii).
(A∗)◦ = −A and (A◦)∗ = A∗.
(ii) However, if X is finite-dimensional, we do have (A∗)◦ = (A◦)∗. Indeed, by Fact 2.6,
2 (cid:19)∗
(A◦)∗ =(cid:18) A − A∗
=
A∗ − A∗∗
2
= (A∗)◦.
We expect that (A∗)◦ = (A◦)∗ for all maximally monotone linear relations if and only if X
is finite-dimensional.
We are now able to present our main result relating monotonicity and adjoint properties of A
and those of its skew part A◦.
Theorem 5.10 (Adjoint characterizations of type (D)) Let A : X ⇒ X ∗ be a monotone lin-
ear relation such that gra A is closed and dom A is closed. Then the following are equivalent.
(i) A is maximally monotone of type (D).
(ii) A is maximally monotone of type (NI).
(iii) A is maximally monotone of type (FP).
(iv) A∗ is monotone.
(v) A∗ is maximally monotone with respect to X ∗∗ × X ∗.
(vi) A◦ is maximally monotone of type (D) and A∗0 = A0.
(vii) (A◦)∗ is maximally monotone with respect to X ∗∗ × X ∗ and A∗0 = A0.
(viii) (A◦)∗ is monotone and A∗0 = A0.
Proof. Apply Theorem 5.5 and Theorem 4.1.
(cid:4)
The work in [6] suggests that in every nonreflexive Banach space there is a maximally monotone
linear relation which is not of type (D).
19
When A is linear and continuous, Theorem 5.10 can also be deduced from [2, Theorem 4.1].
When X is reflexive and dom A is closed, Theorem 5.10 turns into the following refined version of
Fact 2.8:
Corollary 5.11 Suppose that X is reflexive and let A : X ⇒ X ∗ be a monotone linear relation.
such that gra A is closed and dom A is closed. Then the following are equivalent.
(i) A is maximally monotone.
(ii) A∗ is monotone.
(iii) A∗ is maximally monotone.
(iv) A0 = A∗0.
Proof. "(i)⇔(ii)⇔(iii)⇒(iv)": This follows from Theorem 5.10 and Fact 1.2(ii).
"(iv)⇒(i)": Fact 2.10(iv) implies that (dom A)⊥ = A∗0 = A0. By Proposition 5.2(ii), A is
(cid:4)
maximally monotone.
When X is finite-dimensional, the closure assumptions in the previous result are automatically
satisfied and we thus obtain the following:
Corollary 5.12 Suppose that X is finite-dimensional. Let A : X ⇒ X ∗ be a monotone linear
relation. Then the following are equivalent.
(i) A is maximally monotone.
(ii) A∗ is monotone.
(iii) A∗ is maximally monotone.
(iv) A0 = A∗0.
Acknowledgments
Heinz Bauschke was partially supported by the Natural Sciences and Engineering Research Council
of Canada and by the Canada Research Chair Program. Jonathan Borwein was partially supported
by the Australian Research Council. Xianfu Wang was partially supported by the Natural Sciences
and Engineering Research Council of Canada.
20
References
[1] H. Attouch and H. Br´ezis, "Duality for the sum of convex functions in general Banach spaces",
Aspects of Mathematics and its Applications, J. A. Barroso, ed., Elsevier Science Publishers,
pp. 125 -- 133, 1986.
[2] H.H. Bauschke and J.M. Borwein, "Maximal monotonicity of dense type, local maximal mono-
tonicity, and monotonicity of the conjugate are all the same for continuous linear operators",
Pacific Journal of Mathematics, vol. 189, pp. 1 -- 20, 1999.
[3] H.H. Bauschke, J.M. Borwein, and X. Wang, "Fitzpatrick functions and continuous linear
monotone operators", SIAM Journal on Optimization, vol. 18, pp. 789 -- 809, 2007.
[4] H.H. Bauschke, J.M. Borwein, X. Wang and L. Yao, "For maximally monotone linear relations,
dense type, negative-infimum type, and Fitzpatrick-Phelps type all coincide with monotonicity
of the adjoint", submitted; http://arxiv.org/abs/1103.6239v1, March 2011.
[5] H.H. Bauschke, J.M. Borwein, X. Wang, and L. Yao, "Every maximally monotone operator of
Fitzpatrick-Phelps type is actually of dense type", Optimization Letters, in press.
[6] H.H. Bauschke,
J.M. Borwein, X. Wang,
and L. Yao,
logical maximally monotone operators on non-reflexive Banach spaces",
http://arxiv.org/abs/1108.1463, August 2011.
"Construction of patho-
submitted;
[7] H.H. Bauschke and P.L. Combettes, Convex Analysis and Monotone Operator Theory in
Hilbert Spaces, Springer-Verlag, 2011.
[8] H.H. Bauschke, X. Wang, and L. Yao, "Monotone linear relations: maximality and Fitzpatrick
functions", Journal of Convex Analysis, vol. 16, pp. 673 -- 686, 2009.
[9] H.H. Bauschke, X. Wang, and L. Yao, "Examples of discontinuous maximal monotone linear
operators and the solution to a recent problem posed by B.F. Svaiter", Journal of Mathematical
Analysis and Applications, vol. 370, pp. 224-241, 2010.
[10] H.H. Bauschke, X. Wang, and L. Yao, "On Borwein-Wiersma decompositions of monotone
linear relations", SIAM Journal on Optimization, vol. 20, pp. 2636 -- 2652, 2010.
[11] J.M. Borwein, "Adjoint process duality", Mathematics of Operations Research, vol. 8, pp. 403 --
434, 1983.
[12] J.M. Borwein, "Maximal monotonicity via convex analysis", Journal of Convex Analysis,
vol. 13, pp. 561 -- 586, 2006.
[13] J.M. Borwein, "Maximality of sums of two maximal monotone operators in general Banach
space", Proceedings of the AMS, vol. 135, pp. 3917 -- 3924, 2007.
[14] J.M. Borwein, "Fifty years of maximal monotonicity", Optimization Letters, vol. 4, pp. 473 --
490, 2010.
21
[15] J.M. Borwein and J.D. Vanderwerff, Convex Functions, Cambridge University Press, 2010.
[16] H. Br´ezis, "On some degenerate nonlinear parabolic equations", in Nonlinear Functional Anal-
ysis (Proc. Sympos. Pure Math., vol. XVIII, part 1, Chicago, Ill., 1968), AMS, pp. 28 -- 38, 1970.
[17] H. Br´ezis and F.E. Browder, "Singular Hammerstein equations and maximal monotone oper-
ators", Bulletin of the AMS, vol. 82, pp. 623 -- 625, 1976.
[18] H. Br´ezis and F.E. Browder, "Linear maximal monotone operators and singular nonlinear
integral equations of Hammerstein type", in Nonlinear Analysis (collection of papers in honor
of Erich H. Rothe), Academic Press, pp. 31 -- 42, 1978.
[19] R.S. Burachik and A.N. Iusem, Set-Valued Mappings and Enlargements of Monotone Opera-
tors, Springer-Verlag, 2008.
[20] R. Cross, Multivalued Linear Operators, Marcel Dekker, Inc, New York, 1998.
[21] S. Fitzpatrick and R.R. Phelps, "Bounded approximants to monotone operators on Banach
spaces", Annales de l'Institut Henri Poincar´e. Analyse Non Lin´eaire, vol. 9, pp. 573 -- 595,
1992.
[22] J.-P. Gossez, "Op´erateurs monotones non lin´eaires dans les espaces de Banach non r´eflexifs",
Journal of Mathematical Analysis and Applications, vol. 34, pp. 371 -- 395, 1971.
[23] J.-P. Gossez, "On the range of a coercive maximal monotone operator in a nonreflexive Banach
space", Proceedings of the AMS, vol. 35, pp. 88 -- 92, 1972.
[24] M. Marques Alves and B.F. Svaiter, "On Gossez type (D) maximal monotone operators",
Journal of Convex Analysis, vol. 17, pp. 1077 -- 1088, 2010.
[25] J.-E. Mart´ınez-Legaz and B.F. Svaiter, "Monotone operators representable by l.s.c. convex
functions", Set-Valued Analysis, vol. 13, pp. 21 -- 46, 2005.
[26] R.E. Megginson, An Introduction to Banach Space Theory, Springer-Verlag, 1998.
[27] R.R. Phelps, Convex Functions, Monotone Operators and Differentiability, 2nd Edition,
Springer-Verlag, 1993.
[28] R.R. Phelps and S. Simons, "Unbounded linear monotone operators on nonreflexive Banach
spaces", Journal of Nonlinear and Convex Analysis, vol. 5, pp. 303 -- 328, 1998.
[29] R.T. Rockafellar, "Extension of Fenchel's duality theorem for convex functions", Duke Math-
ematical Journal, vol. 33, pp. 81 -- 89, 1966.
[30] R.T. Rockafellar and R.J-B Wets, Variational Analysis, 3rd Printing, Springer-Verlag, 2009.
[31] R. Rudin, Functional Analysis, Second Edition, McGraw-Hill, 1991.
[32] S. Simons, "The range of a monotone operator", Journal of Mathematical Analysis and Ap-
plications, vol. 199, pp. 176 -- 201, 1996.
22
[33] S. Simons, Minimax and Monotonicity, Springer-Verlag, 1998.
[34] S. Simons, "Five kinds of maximal monotonicity", Set-Valued and Variational Analysis, vol. 9,
pp. 391 -- 409, 2001.
[35] S. Simons, From Hahn-Banach to Monotonicity, Springer-Verlag, 2008.
[36] S. Simons,
"A Br´ezis-Browder
theorem for SSDB spaces", preprint. Available at
http://arxiv.org/abs/1004.4251v3, September 2010.
[37] S. Simons and C. Zalinescu, "Fenchel duality, Fitzpatrick functions and maximal monotonic-
ity," Journal of Nonlinear and Convex Analysis, vol. 6, pp. 1 -- 22, 2005.
[38] L. Yao, "The Br´ezis-Browder Theorem revisited and properties of Fitzpatrick functions of
order n", Fixed Point Theory for Inverse Problems in Science and Engineering (Banff 2009),
Springer-Verlag, vol. 49, pp. 391 -- 402, 2011.
[39] C. Zalinescu, Convex Analysis in General Vector Spaces, World Scientific Publishing, 2002.
[40] E. Zeidler, Nonlinear Functional Analysis and its Application, Vol II/A Linear Monotone
Operators, Springer-Verlag, New York-Berlin-Heidelberg, 1990.
[41] E. Zeidler, Nonlinear Functional Analysis and its Application, Vol II/B Nonlinear Monotone
Operators, Springer-Verlag, New York-Berlin-Heidelberg, 1990.
23
|
1305.5916 | 3 | 1305 | 2015-03-03T09:50:25 | Effective results on nonlinear ergodic averages in CAT$(\kappa)$ spaces | [
"math.FA",
"math.DS",
"math.LO"
] | In this paper we apply proof mining techniques to compute, in the setting of CAT$(\kappa)$ spaces (with $\kappa >0$), effective and highly uniform rates of asymptotic regularity and metastability for a nonlinear generalization of the ergodic averages, known as the Halpern iteration. In this way, we obtain a uniform quantitative version of a nonlinear extension of the classical von Neumann mean ergodic theorem. | math.FA | math |
Effective results on nonlinear ergodic averages in
CAT(κ) spaces
Laurent¸iu Leu¸stean1,2, Adriana Nicolae3,4
1 Faculty of Mathematics and Computer Science, University of Bucharest,
Academiei 14, P.O. Box 010014, Bucharest, Romania
2 Simion Stoilow Institute of Mathematics of the Romanian Academy,
P. O. Box 1-764, RO-014700 Bucharest, Romania
3 Department of Mathematics, Babe¸s-Bolyai University,
Kogalniceanu 1, 400084 Cluj-Napoca, Romania
4 Simion Stoilow Institute of Mathematics of the Romanian Academy,
Research group of the project PD-3-0152,
P. O. Box 1-764, RO-014700 Bucharest, Romania
E-mails: [email protected], [email protected]
Abstract
In this paper we apply proof mining techniques to compute, in the
setting of CAT(κ) spaces (with κ > 0), effective and highly uniform rates
of asymptotic regularity and metastability for a nonlinear generalization
of the ergodic averages, known as the Halpern iteration. In this way, we
obtain a uniform quantitative version of a nonlinear extension of the clas-
sical von Neumann mean ergodic theorem.
MSC: 47H25, 03F10, 47J25, 47H09.
Keywords: Proof mining, nonlinear ergodic averages, CAT(κ) spaces,
rates of metastabilty, Halpern iteration, asymptotic regularity.
1
Introduction
In this paper we apply methods from mathematical logic to obtain a uniform
quantitative version of a generalization of the classical von Neumann mean
ergodic theorem, giving effective rates of metastability for the so-called Halpern
iteration, a nonlinear generalization of the ergodic averages. Our results are a
contribution to the line of research known as proof mining, initiated in the 50's
by Kreisel under the name of unwinding of proofs and extensively developed
1
by Kohlenbach, beginning with the 90's. The idea of this research direction is
to extract new, effective information from mathematical proofs making use of
ineffective principles. Hence, it can be related to Terence Tao's proposal [32] of
hard analysis, based on finitary arguments, instead of the infinitary ones from
soft analysis. Proof mining has already been applied in approximation theory,
nonlinear analysis, ergodic theory, topological dynamics and Ramsey theory.
Related to these applications, general logical metatheorems were proved, having
the following form:
if certain statements satisfying general logical conditions
(e.g. ∀∃-sentences) are proved in some formal system associated to an abstract
space, then uniform finitary versions of these statements are guaranteed to hold
and, furthermore, one can transform the initial proof into a quantitative one for
the finitary version and, in this way, extract effective uniform bounds. We refer
to Kohlenbach's book [11] for an introduction to proof mining.
Our theorems guarantee under general logical conditions such strong uni-
form versions of non-uniform existence statements. Moreover, they provide
algorithms for actually extracting effective uniform bounds and transforming
the original proof into one for the stronger uniformity result.
Let us recall the Hilbert space formulation of the celebrated von Neumann
mean ergodic theorem.
Theorem 1.1. Let H be a Hilbert space and U : H → H be a unitary operator.
Then for all x ∈ H, the Ces`aro mean xn = 1
i=0 U ix converges strongly to
the projection of x onto the set of fixed points of U .
If X = (X,B, µ, T ) is a probability measure-preserving system, H = L2(X ) and
U = UT : L2(X ) → L2(X ), f (cid:55)→ f ◦ T is the induced operator, the Ces`aro mean
starting with f ∈ L2(X ) becomes the ergodic average Anf = 1
(cid:80)n−1
i=0 f ◦ T i.
(cid:80)n−1
The convergence of the ergodic averages can be arbitrarily slow, as shown by
Krengel [22]. Furthermore, one cannot expect, in general, to get effective rates
of convergence for the ergodic averages. Avigad, Gerhardy and Towsner [1]
applied methods of computable analysis on Hilbert spaces to obtain an example
of a computable Lebesgue measure-preserving transformation T on [0, 1] and
a computable characteristic function χA such that the limit of the sequence
AnχA is not a computable element of L2([0, 1]), which implies that there is no
computable bound on the rate of convergence of (AnχA).
n
n
Cauchy property of (xn):
However, one can consider the following equivalent reformulation of the
∀k ∈ N∀g : N → N∃N∀i, j ∈ [N, N + g(N )] (cid:0)(cid:107)xi − xj(cid:107) < 2−k(cid:1) .
(1)
This is known in logic as Kreisel's [20, 21] no-counterexample interpretation of
the Cauchy property and it was popularized in the last years under the name
of metastability by Tao [32, 33]. In [33], Tao generalized the mean ergodic the-
orem for multiple commuting measure-preserving transformations, by deducing
it from a finitary norm convergence result, expressed in terms of metastability.
Recently, Walsh [34] used again metastability to show the L2-convergence of
multiple polynomial ergodic averages arising from nilpotent groups of measure-
preserving transformations.
2
Logical metatheorems developed by Kohlenbach [13] show that, from wide
classes of mathematical proofs one can extract effective bounds on ∃N in (1).
Thus, taking ε > 0 instead of 2−k, we define a rate of metastability as a func-
tional Φ : (0,∞) × NN → N satisfying
∀ε > 0∀g : N → N∃N ≤ Φ(ε, g)∀i, j ∈ [N, N + g(N )] ((cid:107)xi − xj(cid:107) < ε) .
(2)
Thus, a natural direction of research is to obtain finitary, quantitative ver-
sions of convergence statements for sequences (xn) by providing effective rates
of metastability. A qualitative feature of these quantitative versions is that the
rates of metastability are highly uniform and independent or have only a weak
dependence on the input data. Furthermore, these quantitative versions can be
thereafter generalized to new structures, obtaining as an immediate consequence
the generalization of the initial (non-quantitative) Cauchy statement to these
structures. The main quantitative result of this paper is obtained in this way.
We refer to [15] for another example in the context of the asymptotic behaviour
of nonlinear iterations.
Avigad, Gerhardy and Towsner [1] computed for the first time explicit and
uniform rates of metastability for the ergodic averages, by a logical analysis of
Riesz' proof of the mean ergodic theorem. Their result was generalized, with
better bounds, to uniformly convex Banach spaces by Kohlenbach and the first
author [14], applying proof mining methods, but this time to a proof of Garrett
Birkhoff [3].
In fact, Avigad and Rute [2] realized that the computations in
[16] allow one to obtain an effective bound on the number of ε-fluctuations (i.e.
pairs (i, j) with i > j and (cid:107)xi− xj(cid:107) > ε). A very nice discussion on the different
types of quantitative information (metastability, effective learnability, bounds
on the number of oscillations) that can be extracted from convergence proofs is
done in a recent paper by Kohlenbach and Safarik [19].
In the important paper [35], Wittmann obtained the following nonlinear
generalization of the mean ergodic theorem.
Theorem 1.2. [35] Let C be a bounded closed convex subset of a Hilbert space
X, T : C → C a nonexpansive mapping and (λn)n≥1 a sequence in [0, 1]. For
any u ∈ C, define
x0 = u,
xn+1 = λn+1u + (1 − λn+1)T xn.
Assume that (λn) satisfies
lim
n→∞ λn = 0,
∞(cid:88)
n=1
λn+1 − λn < ∞ and
∞(cid:88)
n=1
λn = ∞
(3)
(4)
Then for any u ∈ C, (xn) converges to the projection PF ix(T )u of u onto the
(nonempty) set of fixed points F ix(T ).
One can easily see that (xn) coincides with the Ces`aro mean when T is linear
1
n + 1
and λn =
. The iteration (xn) is known as the Halpern iteration, as it was
3
introduced by Halpern [9] for the special case u = 0. We refer to [17, Section 3]
for a discussion on results in the literature on Halpern iterations, obtained by
considering different conditions on (λn) or more general spaces.
Kohlenbach's logical metatheorem for Hilbert spaces [13] guarantees also in
the case of Wittmann's theorem that from its proof one can extract a rate of
metastability Φ of (xn), uniform in the following sense:
it depends only on ε
and g, an upper bound on the diameter of C and moduli on (λn), given by the
quantitative version of (4). Thus, Φ is independent with respect to the starting
point u of the iteration, the nonexpansive mapping T , the Hilbert space X and
depends on C only via its diameter. Kohlenbach [12] computed such a uniform
rate of metastability by a logical analysis of Wittmann's proof.
Furthermore, Kohlenbach and the first author [16, 17, 18] extracted rates
of metastability from the proofs of two generalizations of Wittmann's theorem
given by Shioji and Takahashi [31] for a class of Banach spaces with a uniformly
Gateaux differentiable norm and by Saejung [29] for CAT(0) spaces. Both Sae-
jung's and Shioji-Takahashi's proofs use Banach limits (whose existence requires
the axiom of choice), inspired by Lorentz' seminal paper [25], introducing al-
most convergence. Our quantitative results were obtained by developing in [17]
a method to eliminate the use of Banach limits from these proofs and get, in this
way, elementary proofs to which general logical metatheorems for CAT(0) spaces
[13] and for uniformly smooth Banach spaces [16] can be applied to guarantee
the extractability of effective bounds. We point out that the use of Lorentz'
almost convergence (and hence, Banach limits) in nonlinear ergodic theory was
introduced by Reich [27], while Bruck and Reich [7] applied Banach limits for
the first time to the study of Halpern iterations (see also [8, Sections 12, 14]).
Geodesic spaces provide a suitable setting for extending the notion of sec-
tional curvature from Riemannian manifolds. An important class of geodesic
spaces of bounded curvature are CAT(κ) spaces, where geodesic triangles are in
some sense "thin". Such spaces enjoy nice properties inherited from the com-
parison with the model spaces and proved to be relevant in various problems
and aspects in geometry (see [4]).
Recently, Pi¸atek [26] extended Wittmann's result to the context of CAT(κ)
spaces with κ > 0. In this paper we extract an effective and uniform rate of
metastability for this generalization of Wittmann's theorem.
Our main quantitative result (Theorem 3.4) is obtained by generalizing to
CAT(κ) spaces the quantitative proof for CAT(0) spaces from [17]. Thus, we
apply again the general method developed in [17], together with the remark
that, in fact, our logical analysis of Saejung's proof for CAT(0) spaces results
in the elimination of any contribution of Banach limits, hence even the fini-
tary lemmas proved in [17, Section 8] are no longer needed (see [18]). Despite
this simplification, the proofs we give in this paper are much more involved,
since we work in the setting of CAT(κ) spaces. However, we still get a rate of
metastability having a form similar to the one described in [19].
As the first step in the convergence proof is to obtain the asymptotic regu-
larity, our first important result (Proposition 3.2) consists in the computation
of a uniform rate of asymptotic regularity.
4
For the rest of the paper N = {0, 1, 2, . . .} and Z+ = {1, 2, . . .}. Furthermore,
we consider CAT(κ) spaces with κ > 0.
2 CAT(κ) spaces
Let (X, d) be a metric space. A geodesic path from x to y is a mapping
[0, l] ⊆ R → X such that c(0) = x, c(l) = y and d (c(t), c(t(cid:48))) = t − t(cid:48) for
c :
every t, t(cid:48) ∈ [0, l]. The image c ([0, l]) of c forms a geodesic segment which joins
x and y. Note that a geodesic segment from x to y is not necessarily unique.
If no confusion arises, we use [x, y] to denote a geodesic segment joining x and
y. (X, d) is called a (uniquely) geodesic space if every two points x, y ∈ X can
be joined by a (unique) geodesic path. A point z ∈ X belongs to the geodesic
segment [x, y] if and only if there exists t ∈ [0, 1] such that d(z, x) = td(x, y) and
d(z, y) = (1 − t)d(x, y), and we write z = (1 − t)x + ty for simplicity. This, too,
may not be unique. A subset C of X is convex if C contains any geodesic seg-
ment that joins every two points in C. A geodesic triangle ∆(x1, x2, x3) consists
of three points x1, x2 and x3 in X (its vertices) and three geodesic segments
corresponding to each pair of points (its edges).
CAT(κ) spaces are defined in terms of comparisons with the model spaces
M n
κ . We focus here on CAT(κ) spaces with κ > 0. We give below the precise
definition and briefly describe some of their properties that play an essential
role in this work. For a detailed discussion on geodesic metric spaces and, in
particular, on CAT(κ) spaces, one may check, for example, [4].
The n-dimensional sphere Sn is the set {x ∈ Rn+1 : (x x) = 1}, where (· ·)
stands for the Euclidean scalar product. Consider the mapping d : Sn ×Sn → R
by assigning to each (x, y) ∈ Sn × Sn the unique number d(x, y) ∈ [0, π] such
that cos d(x, y) = (x y). Then, (Sn, d) is a metric space called the spherical
space. This space is also geodesic and, if d(x, y) < π, then there exists a unique
geodesic segment joining x and y. Moreover, open (resp. closed) balls of radius
≤ π/2 (resp. < π/2) are convex. The spherical law of cosines states that in a
spherical triangle with vertices x, y, z ∈ Sn and γ the spherical angle between
the geodesic segments [x, y] and [x, z] we have
cos d(y, z) = cos d(x, y) cos d(x, z) + sin d(x, y) sin d(x, z) cos γ.
Let κ > 0 and n ∈ N. The classical model spaces M n
√
κ are obtained from
the spherical space Sn by multiplying the spherical distance with 1/
κ. These
√
spaces inherit the geometrical properties from the spherical space. Thus, there is
a unique geodesic path joining x, y ∈ M n
√
κ. Further-
κ if and only if d(x, y) < π/
√
κ) are convex and we have a counterpart
more, closed balls of radius < π/(2
of the spherical law of cosines. We denote the diameter of M n
κ.
κ by Dκ = π/
For a geodesic triangle ∆=∆(x1, x2, x3), a κ-comparison triangle is a triangle
(¯xi, ¯xj) for i, j ∈ {1, 2, 3}.
¯∆ = ∆(¯x1, ¯x2, ¯x3) in M 2
For κ fixed, κ-comparison triangles of geodesic triangles (having perimeter less
than 2Dκ) always exist and are unique up to isometry.
κ such that d(xi, xj) = dM 2
κ
5
A geodesic triangle ∆ of perimeter less than 2Dκ satisfies the CAT(κ) in-
equality if for every κ-comparison triangle ¯∆ of ∆ and for every x, y ∈ ∆ we
have
d(x, y) ≤ dM 2
κ
(¯x, ¯y),
where ¯x, ¯y ∈ ¯∆ are the comparison points of x and y, i.e., if x = (1 − t)xi + txj
then ¯x = (1 − t)¯xi + t¯xj for i, j ∈ {1, 2, 3}.
A metric space is called a CAT(κ) space if every two points at distance
less than Dκ can be joined by a geodesic segment and every geodesic triangle
having perimeter less than 2Dκ satisfies the CAT(κ) inequality. CAT(0) spaces
are defined in a similar way considering the model space M 2
0 to be the Euclidean
plane of infinite diameter.
3 Main results
Let (X, d) be a geodesic space, C ⊆ X a convex subset, T : C → C a nonex-
pansive mapping and (λn) a sequence in [0, 1]. The Halpern iteration starting
at u ∈ C can be defined by
x0 = u,
xn+1 = λn+1u + (1 − λn+1)T xn.
(5)
The main purpose of our work is to prove a quantitative version of the
following generalization of Wittmann's theorem to CAT(κ) spaces, obtained
recently by Pi¸atek [26].
Theorem 3.1. Let X be a complete CAT(κ) space, C ⊆ X a bounded closed
and T : C → C a nonexpansive mapping.
convex subset with diameter dC <
Assume that (λn) satisfies (4). Then for any u ∈ C, the iteration (xn) starting
from u converges to the fixed point of T which is nearest to u.
Dκ
2
A first important result of this paper is the extraction of an effective rate of
asymptotic regularity for the Halpern iteration, that is, a rate of the convergence
of (d(xn, T xn)) towards 0. In order to state this result, we need to make the
hypotheses (4) on (λn) quantitative.
For brevity, we say that the sequence (λn) and the functions α : (0,∞) →
Z+, γ : (0,∞) → Z+ and θ : Z+ → Z+ satisfy (*) if the following conditions
hold:
lim
n→∞ λn+1 = 0 with rate of convergence α, i.e.,
λn+1 ≤ ε,
for all ε > 0 and all n ≥ α(ε);
λn+1 − λn converges with Cauchy modulus γ, i.e.,
(i)
(ii)
∞(cid:88)
n=1
λi+1 − λi ≤ ε,
for all ε > 0 and all n ∈ Z+;
γ(ε)+n(cid:88)
i=γ(ε)+1
6
∞(cid:88)
n=1
(iii)
λn+1 = ∞ with rate of divergence θ, i.e.,
θ(n)(cid:88)
k=1
λk+1 ≥ n,
for all n ∈ Z+.
Proposition 3.2. Let X be a CAT(κ) space, C ⊆ X a bounded convex subset,
T : C → C nonexpansive and M < Dκ
2 an upper bound on the finite diameter
dC of C. Assume furthermore that (λn), α, γ, θ satisfy (*).
Then lim
n→∞ d(xn, xn+1) = 0 with rate of convergence Φ given by
(cid:18)(cid:24)
Φ(ε, κ, M, γ, θ) = θ
1
cos(M
√
κ)
(cid:25)(cid:18)
(cid:16) ε
(cid:17)
γ
2M
(cid:110) Φ
(cid:16) ε
2
+ max
ln
(cid:26)(cid:24)
(cid:18) 2M
(cid:19)(cid:25)
(cid:27)(cid:19)(cid:19)
(cid:17)
ε
(cid:17)(cid:111)
.
(cid:16) ε
2M
, 1
(6)
(7)
, κ, M, γ, θ
, α
and lim
n→∞ d(xn, T xn) = 0 with rate of convergence Φ given by
Φ(ε, κ, M, γ, θ, α) = max
Proof. See Section 5.
If λn =
1
n + 1
one can easily obtain rates α, γ, θ:
(cid:24) 1
(cid:25)
ε
α(ε) = γ(ε) =
,
θ(n) = exp ((n + 1) ln 4) .
(8)
As an immediate consequence we get the following:
Corollary 3.3. Assume that λn =
1
n + 1
, n ≥ 1. Then
lim
n→∞ d(xn, xn+1) = lim
n→∞ d(xn, T xn) = 0
(cid:18)(cid:24)
(cid:25)(cid:24) 8M
ε
1
cos(M
√
κ)
(cid:25)
(cid:19)
+ 2
ln 4
,
(9)
with a common rate of convergence
Ψ(ε, κ, M ) = exp
which is exponential in
1
ε
.
We point out that exponential rates of asymptotic regularity for the Halpern
iteration were obtained by the first author for Banach spaces in [23] and for the
so-called W -hyperbolic spaces in [24]. Kohlenbach [12] remarked that the proof
in [23] can be simplified and, as a consequence, one gets quadratic rates in
Banach spaces. For CAT(0) spaces, Kohlenbach and the first author provide in
[17] a quantitative asymptotic regularity result for general (λn) by considering
7
∞(cid:88)
instead of
λn+1 = ∞ the equivalent condition
∞(cid:89)
(1 − λn+1) = 0. As a
n=1
n=1
corollary, one obtains again quadratic rates of asymptotic regularity. However,
the method used in [17] for CAT(0) spaces does not hold for CAT(κ) spaces.
The main result of the paper is the following quantitative version of Theorem
3.1, which provides an explicit uniform rate of metastability for the Halpern iter-
ation in CAT(κ) spaces. To get such a result we apply again the general method
developed by Kohlenbach and the first author in [17] for the Halpern iteration
in CAT(0) spaces and applied again in [16] for uniformly smooth Banach spaces
as well as in [30] for a modified Halpern iteration in CAT(0) spaces. As noticed
in [18], in the end we do not need the finitary Lemmas 8.3 and 8.4 from [17],
since, as a consequence of the proof mining methods applied to Saejung's proofs,
one gets a proof where no contributions of Banach limits can be traced.
Theorem 3.4. Let X be a complete CAT(κ) space, C ⊆ X a bounded closed
convex subset, T : C → C nonexpansive and M < Dκ
2 an upper bound on the
finite diameter dC of C. Assume furthermore that (λn), α, γ, θ satisfy (*). Then
for all ε ∈ (0, 2) and g : N → N,
∃N ≤ Σ(ε, g, κ, M, θ, α, γ) ∀m, n ∈ [N, N + g(N )] (d(xn, xm) ≤ ε),
(cid:18)
(cid:19)
√
4
κ
with N = ΘK0
sin2 ε
for some
(cid:24) 1
(cid:25)
ε0
≤ K0 ≤(cid:102)f∗Bε,κ,M
(cid:18)(cid:102)f∗Bε,κ,M
(0) +
(cid:25)
(cid:24) 1
(cid:25)(cid:19)
ε0
,
(0) +
(cid:24) 1
ε0
and
Σ(ε, g, κ, M, θ, α, γ) = Aε,κ,M,θ,α,γ
where the above constants and functionals are specified in Table 1.
Proof. We refer to Section 7 for the proof. We point here only the main steps:
(i) extract a rate of asymptotic regularity (this is done in Proposition 3.2);
(ii) obtain a quantitative Browder theorem (see Proposition 6.2);
(iii) define in an appropriate way an approximate fixed point sequence γt
n (see
(25));
(iv) apply Lemma 7.3, a quantitative lemma on sequences of real numbers.
Hence, we compute a rate of metastability which is uniform in the starting
point x0 of the iteration and the nonexpansive mapping T . Moreover, it depends
on the space X and the set C only via κ and the diameter dC of C. The
dependence on (λn) is through the rates α, γ, θ, which can be computed very
easily for the natural choice λn =
Furthermore, as in [16, 17] as well as in other case studies in proof mining,
the rate of metastability has the form described by Kohlenbach and Safarik [19].
1
.
n + 1
8
(cid:24) M
Bε,κ,M =
(cid:25)
,
κ)
√
√
κ tan(M
1 − cos(ε0)
(cid:18)(cid:24)
Aε,κ,M,θ,α,γ(n) = θ+
Γ(n) = max
(cid:38)
(cid:32)
S =
ln
1
cos(M
√
4
κ
√
(cid:19)
:
, χ∗
sin2 ε
(cid:33)(cid:39)
κ
(cid:26)
(cid:18) 1
χ∗
i
3
√
3 sin2 M
√
4
sin2 ε
κ
4
(cid:18)(cid:24)
χi(ε) = max
θ
1
cos(M
√
κ)
√
κ)
√
sin2 ε
4
κ
,
cos(M
36
(Γ(n) − 1 + max{S, 1})
(cid:19)
+ 1,
ε0 =
(cid:25)
(cid:24) 1
κ)
(cid:25)
(cid:27)
≤ i ≤ n
,
2
ε0
κ)
√
cos(M
i (ε) = χi
(cid:16) ε
(cid:17)
(cid:25)(cid:18)
(cid:18) 1
(cid:17) − 1 + max{T, 1}(cid:17)(cid:19)
(cid:16) ε
γ(Li) + max
(cid:26)(cid:24)
ln
Li
+ 1,
θ+(n) = max
1≤i≤n
θ(i),
√
√
cos(M
4M
κ)ε
κ(i + 1)
,
(cid:27)
, Li =
(cid:19)(cid:25)
(cid:27)(cid:19)(cid:19)
, 1
, α(2Li)
,
(cid:26)
(cid:18)(cid:24)
(cid:18) 3
ε
(cid:40)(cid:38)
Θi(ε) = θ
(cid:24)
T =
ln
√
1
cos(M
√
sin2 M
κ
2
(cid:25)(cid:16)
(cid:19)(cid:25)
κ)
√
χ∗
i
3
(cid:39)
(cid:41)
, ∆∗
i (ε, g) =
3Θi(ε) − 3χ∗
ε
i ( ε
3 ) + 3g (Θi(ε))
(cid:18)
(cid:24) 1
(cid:25)(cid:19)
ε0
+
,
(cid:24) 1
(cid:25)
ε0
− i,
f∗(i) = f
, i
i +
M
κ
√
κ
i (sin2 ε
4 , g)
∆∗
f (i) = max
and(cid:102)f∗(i) = i + f∗(i).
Table 1
and Bε,κ,M , and (cid:102)f∗(i) only uses g on one argument, Θi(sin2(ε
Thus, g does not appear at all in the definition of the mappings Aε,κ,M,θ,α,γ
κ/4)), which
itself does not depend on g. We refer to [19] for a logical explanation of this
phenomenon in terms of effective learnability and bounds on the number of
mind changes.
√
, with α, γ, θ given by (8), one can easily see that (χ∗
i )i is
If λn =
1
n + 1
nondecreasing. As a consequence we obtain the following:
Corollary 3.5. Assume that λn =
and g : N → N,
1
n + 1
for all n ≥ 1. Then for all ε ∈ (0, 2)
∃N ≤ Σ(ε, g, κ, M ) ∀m, n ∈ [N, N + g(N )] (d(xn, xm) ≤ ε),
9
(cid:18)(cid:102)f∗Bε,κ,M
(0) +
Σ(ε, g, κ, M ) = Aε,κ,M
(cid:24) 1
(cid:25)(cid:19)
ε0
Aε,κ,M (n) = exp
(Γ(n) − 1 + max{S, 1}) + 1
where
with
(cid:18)(cid:18)(cid:24)
(cid:18) 1
(cid:18)(cid:18)(cid:24)
(cid:18)(cid:18)(cid:24)
3
1
cos(M
√
κ
√
(cid:19)
sin2 ε
κ)
,
4
1
cos(M
1
cos(M
√
√
κ)
κ)
(cid:25)
(cid:25)(cid:18)(cid:24) 1
(cid:25)
(cid:25)(cid:16)
(cid:16) ε
Li
χ∗
i
3
Γ(n) = χ∗
n
χi(ε) = exp
Θi(ε) = exp
,
(cid:19)
(cid:19)
(cid:26)(cid:24)
(cid:19)(cid:25)
(cid:18) 1
(cid:17) − 1 + max{T, 1}(cid:17)
+ max
Li
ln
ln 4
+ 1,
(cid:27)(cid:19)
(cid:19)
, 1
(cid:19)
(cid:19)
+ 1
(cid:19)
ln 4
,
+ 1
ln 4
+ 1
and the other constants and functionals are defined as in Theorem 3.4.
4 Some technical lemmas
Throughout the paper, we shall use the following well-known facts:
(i) x ≥ sin x for all x ≥ 0.
(ii) sin(tx) ≥ t sin x for all x ∈ [0, π] and all t ∈ [0, 1].
(iii) The function f : (0, π) → (0, 1), f (x) =
sin x
is decreasing.
x
(iv) Given t ∈ [0, 1], the mapping f : (0, π) → (0,∞), f (x) =
creasing.
sin(tx)
sin x
is in-
The following very useful result is proved in [26] for κ = 1. The proof for
general κ > 0 is an immediate rescaling.
Lemma 4.1. Let ∆(x, y, z) be a triangle in X and M ≤ Dκ
on the lengths of the sides of ∆(x, y, z). Then for all t ∈ (0, 1),
d((1 − t)x + tz, (1 − t)y + tz) ≤ sin(cid:0)(1 − t)M
sin(cid:0)M
κ(cid:1)
κ(cid:1)
√
√
d(x, y) ≤ d(x, y).
2 be an upper bound
Let X be a CAT(κ) space. The next results gather some useful properties
which will be needed in the subsequent sections.
√
Lemma 4.2. Let ∆(x, y, z) be a triangle in X with perimeter < 2Dκ. Let
κ) ≥
√
w be a point on the segment joining x and z. Suppose that cos(d(y, z)
κ). Then d(x, w) ≤ d(x, y). Moreover, if ∆(¯x, ¯y, ¯z)
cos(d(y, w)
is a κ-comparison triangle for ∆(x, y, z), then ∠ ¯w(¯y, ¯x) ≥ π
2 .
√
κ) cos(d(w, z)
10
Proof. Let ∆(¯x, ¯y, ¯z) be a κ-comparison triangle for ∆(x, y, z) and α = ∠ ¯w(¯y, ¯z).
Suppose that α >
π
2
. Then
√
cos(d(y, z)
√
κ) = cos(d(¯y, ¯w)
√
+ sin(d(¯y, ¯w)
√
< cos(d(¯y, ¯w)
≤ cos(d(y, w)
√
κ) cos(d( ¯w, ¯z)
√
√
κ)
√
κ) sin(d( ¯w, ¯z)
√
κ) cos(d( ¯w, ¯z)
κ)
κ) cos(d(w, z)
κ),
κ) cos α
which contradicts the hypothesis. Thus, α ≤ π
follows that
2 and β = ∠ ¯w(¯y, ¯x) ≥ π
.
It
2
cos(d(¯x, ¯y)
√
√
√
κ) cos(d( ¯w, ¯y)
κ) = cos(d(¯x, ¯w)
√
+ sin(d(¯x, ¯w)
≤ cos(d(¯x, ¯w)
√
√
κ)
κ) sin(d( ¯w, ¯y)
κ) cos β
κ),
hence d(¯x, ¯y) ≥ d(¯x, ¯w). Thus, d(x, w) ≤ d(x, y).
Assume C ⊆ X is bounded with M < Dκ
2 an upper bound on its diameter.
In the sequel x, y, z are pairwise distinct points of C and w ∈ [x, y], v ∈ [x, z].
We shall use the following notation:
√
√
√
√
S1 = sin(d(x, w)
κ), S2 = sin(d(x, y)
κ) sin(d(x, z)
κ),
κ) sin(d(x, v)
√
√
√
√
κ) sin(d(x, z)
κ), S4 = sin(d(y, w)
S3 = sin(d(x, w)
κ) sin(d(x, z)
√
√
κ),
κ) sin(d(z, v)
S5 = sin(d(x, w)
√
√
√
κ) cos(d(x, v)
C1 = cos(d(x, w)
κ),
√
κ) cos(d(x, z)
κ), C2 = cos(d(x, y)
κ).
Lemma 4.3.
√
S2 − S3 ≤ S4 cos(d(x, w)
√
S3 − S1 ≤ S5 cos(d(x, v)
√
S2C1 − S1C2 = S4 cos(d(x, v)
κ),
(cid:18)
(cid:18)
κ),
√
κ) + S5 cos(d(x, y)
κ),
√
√
− sin2 d(x, w)
sin2 d(x, v)
2
2
√
− sin2 d(x, v)
sin2 d(x, y)
2
2
κ
√
(cid:19)
(cid:19)
κ
κ
κ
(10)
(11)
(12)
, (13)
.
(14)
√
S2 − S3 − S4 cos(d(x, v)
κ) ≤ 2S4
S3 − S1 − S5 cos(d(x, y)
√
κ) ≤ 2S5
11
Proof.
S2 − S3 =(cid:0) sin(d(x, y)
√
√
κ) − sin(d(x, w)
√
(d(x, y) − d(x, w))
κ
cos
(cid:18)(cid:18)
2
κ
cos
κ)(cid:1) sin(d(x, z)
(cid:19)√
d(y, w)
2
√
(cid:19)
(d(x, y) + d(x, w))
κ)
√
= 2 sin
= 2 sin
√
d(y, w)
√
d(w, y)
√
2
= sin(d(w, y)
≤ 2 sin
2
d(x, w) +
κ
√
cos(d(x, w)
√
κ) cos(d(x, w)
2
√
d(w, y)
√
2
κ) sin(d(x, z)
κ) cos
κ
√
sin(d(x, z)
κ)
√
sin(d(x, z)
√
sin(d(x, z)
κ)
κ)
κ
κ
√
κ)
√
κ)
κ) = S4 cos(d(x, w)
κ).
κ).
√
√
Similarly, one gets that S3 − S1 ≤ S5 cos(d(x, v)
√
√
√
S2C1 − S1C2 = sin(d(x, y)
κ) sin(d(x, z)
κ) cos(d(x, w)
κ) cos(d(x, v)
√
√
√
− sin(d(x, w)
κ) sin(d(x, v)
κ) cos(d(x, y)
√
√
√
κ) cos(d(x, v)
= sin(d(x, z)
√
√
κ) cos(d(x, y)
+ sin(d(x, w)
√
√
κ) cos(d(x, v)
κ) sin(d(y, w)
= sin(d(x, z)
√
√
κ) cos(d(x, y)
+ sin(d(x, w)
√
= S4 cos(d(x, v)
κ(cid:1)
κ) sin(cid:0)(d(x, y) − d(x, w))
κ) sin(cid:0)(d(x, z) − d(x, v))
κ(cid:1)
√
κ) + S5 cos(d(x, y)
κ) cos(d(x, z)
√
κ) sin(d(z, v)
κ)
√
√
κ).
κ)
Items (13) and (14) follow easily from (10) and (11), respectively.
Proposition 4.4.
√
sin2 d(w, v)
2
κ
√
sin2 d(y, z)
2
≤ S1
S2
κ
+
1
2
(1 − C1) − S1
2S2
(1 − C2).
(15)
Proof. Let ∆(¯x, ¯y, ¯z) be a κ-comparison triangle for ∆(x, y, z). Denote α =
∠¯x(¯y, ¯z) = ∠¯x( ¯w, ¯v). Using the cosine law we have
cos(d( ¯w, ¯v)
√
√
√
κ)
κ) cos(d(¯x, ¯v)
κ) = cos(d(¯x, ¯w)
√
√
κ) sin(d(¯x, ¯v)
+ sin(d(¯x, ¯w)
κ) cos α
and
Thus,
√
cos(d( ¯w, ¯v)
√
cos(d(¯y, ¯z)
√
κ) cos(d(¯x, ¯z)
κ) = cos(d(¯x, ¯y)
√
+ sin(d(¯x, ¯y)
√
κ)
√
κ) sin(d(¯x, ¯z)
κ) cos α.
√
√
√
√
κ) cos(d(¯x, ¯v)
κ) = cos(d(¯x, ¯w)
κ)
√
√
sin(d(¯x, ¯w)
κ) sin(d(¯x, ¯v)
κ)
sin(d(¯x, ¯y)
κ) sin(d(¯x, ¯z)
κ)
√
√
− cos(d(¯x, ¯y)
κ) cos(d(¯x, ¯z)
√
+
κ)
=
cos(d(y, z)
κ) + C1 − S1
S2
C2.
S1
S2
(cid:18)
(cid:19)
√
κ)
cos(d(¯y, ¯z)
12
It follows that
1 − cos(d(w, v)
√
κ)
2
≤ 1
2
+
S1
S2
(cid:18) 1 − cos(d(y, z)
√
2
κ)
− 1
2
(cid:19)
− 1
2
C1 +
S1
2S2
C2.
Hence,
√
sin2 d(w, v)
2
κ
≤ S1
S2
√
sin2 d(y, z)
2
κ
+
1
2
(1 − C1) − S1
2S2
(1 − C2).
Proposition 4.5.
√
sin2 d(w, v)
2
κ
(i)
√
√
≤ sin(d(x, w)
sin(d(x, y)
κ)
√
κ)
√
sin(d(y, w)
κ)
sin(d(x, y)
κ)
√
√
sin(d(z, v)
κ)
sin(d(x, z)
κ)
√
sin2 d(y, z)
2
κ
(cid:18)
sin2 d(x, v)
2
√
sin2 d(x, y)
2
κ
.
(cid:19)
√
κ
√
κ
− sin2 d(x, w)
2
+
+
√
(ii) Assume that v = sx + (1 − s)z, s ∈ [0, 1] and w = rx + (1 − r)y, r ∈ [0, 1].
Then,
sin2 d(w, v)
2
κ
√
√
sin(rM
κ)
sin(M
κ)
√
√
sin(sM
κ)
κ)
sin(M
+
+
≤ sin((1 − r)M
√
√
(cid:26)
√
κ)
sin2 d(y, z)
√
2
sin(M
sin2 d(x, v)
2
√
√
− sin2 d(x, w)
2
κ)
κ
κ
max
(cid:27)
, 0
κ
sin2 M
κ
.
2
Proof.
(i) We apply Proposition 4.4 to get that
√
κ
sin2 d(y, z)
2
√
≤ S1
sin2 d(w, v)
√
2
S2
S1
sin2 d(y, z)
√
S2
2
S2 − S1 − S4 cos(d(x, v)
2S2
=
+
κ
by (12)
≤ S1
S2
√
sin2 d(y, z)
2
κ
+
√
sin2 d(x, y)
2
by (13) and (14),
S5
S2
+
κ
S4
S2
κ
+
1
2
(1 − C1) − S1
2S2
(1 − C2)
√
κ) − S5 cos(d(x, y)
κ)
(cid:18)
√
sin2 d(x, v)
2
κ
√
− sin2 d(x, w)
2
κ
(cid:19)
which yields the desired inequality.
13
(ii) We have that
√
sin2 d(w, v)
2
√
κ
+
+
√
max
(cid:26)
≤ sin((1 − r)d(x, y)
√
√
sin(d(x, y)
√
sin(rd(x, y)
κ)
sin(d(x, y)
κ)
√
√
κ)
sin(sd(x, z)
√
κ)
sin(d(x, z)
≤ sin((1 − r)M
√
κ)
√
sin(M
κ)
√
sin(rM
κ)
sin(M
κ)
√
√
κ)
sin(sM
sin(M
κ)
sin2 d(x, y)
√
2
sin2 d(y, z)
2
sin2 d(x, v)
2
√
sin2 M
(cid:26)
max
+
+
κ
2
.
κ
√
κ
κ)
κ)
sin2 d(y, z)
√
2
sin2 d(x, v)
2
√
− sin2 d(x, w)
2
κ
(cid:27)
, 0
κ
κ
√
(cid:27)
√
κ
, 0
κ
− sin2 d(x, w)
2
For the rest of the section, we assume that v = sx + (1 − s)z, s ∈ (0, 1). We
use the additional notation
L1 =
S1
S3
=
Lemma 4.6.
√
√
sin(d(x, v)
sin(d(x, z)
κ)
κ)
, L2 =
=
√
√
κ)
sin(d(v, z)
sin(d(x, z)
κ)
.
κ)
14
S5
S3
√
κ),
√
.
κ)
≤
0 < 1 − L1 ≤ L2 cos(d(x, v)
1
s cos(M
√
κ) = sin(d(x, z)
L1
1 − L1
√
√
(d(x, z) − d(x, v))
(16)
(17)
κ)
√
κ) − sin(d(x, v)
√
(d(x, z) + d(x, v))
κ
2
√
cos(d(x, v)
κ)
Proof.
(1 − L1) sin(d(x, z)
= 2 sin
≤ 2 sin
√
d(z, v)
√
2
= sin(d(z, v)
Thus, 1 − L1 ≤ L2 cos(d(x, v)
√
sin(d(x, v)
sin(d(x, z)
L1
1 − L1
=
2
κ
κ
cos
κ
cos
√
d(z, v)
√
2
κ) cos(d(x, v)
√
κ).
√
κ).
κ)
κ) − sin(d(x, v)
√
sin(d(x, z)
κ)
(cid:16)(cid:16)
√
κ
cos
√
sin(d(x, z)
d(x, z) − d(z,v)
κ)
2
√
κ
√
cos(d(x, z)
(cid:17)√
(cid:17)
κ
≤
1
s cos(M
√
.
κ)
=
≤
2 sin d(z,v)
2
2 sin sd(x,z)
2
√
√
sin(d(x, z)
sin(d(x, z)
√
κ) − sin(d(x, v)
κ)
κ)
√
κ)
≤
Proposition 4.7.
(i)
√
sin2 d(y, v)
2
κ
√
≤ L1 sin2 d(y, z)
2
√
= L2 sin2 d(x, y)
2
κ
+
1 − L1
2
− 1
2
√
cos(d(x, y)
√
κ)L2
κ
+
1
2
(1 − L1 − L2) + L1 sin2 d(y, z)
2
κ
.
(ii) Let q ∈ C be such that d(q, z) ≤ d(y, v). Assume that
√
sin2 d(x, y)
2
κ
√
− sin2 d(x, v)
2
κ
≤ 0.
Then,
√
sin2 d(y, v)
2
κ
√
≤ sin2 d(x, y)
2
√
κ
κ)
κ
(cid:18)
√
− sin2 d(x, v)
√
2
sin2 d(y, q)
2
κ
+ sin
d(y, q)
2
+
1
s cos(M
√
≤ L1 sin2 d(y, z)
2
κ
Proof. (i) We apply Proposition 4.4 with w = y to get that
√
sin2 d(y, v)
2
√
(1 − cos(d(x, y)
1
+
2
√
√
L1(1 − cos(d(x, y)
κ) cos(d(x, z)
κ))
κ
− 1
2
κ) cos(d(x, v)
√
κ))
(18)
(cid:19)
.
√
κ
2
κ
+
√
1 − L1
(cid:0) cos(d(x, v)
κ)(cid:1)
√
= L1 sin2 d(y, z)
√
2
√
− cos(d(x, y)
κ)
√
2 sin(d(x, z)
κ)
− sin(d(x, v)
κ) cos(d(x, z)
√
1 − L1
= L1 sin2 d(y, z)
√
2
= L1 sin2 d(y, z)
2
1 − L1
− 1
2
√
+
+
κ
κ
2
2
√
κ) sin(d(x, z)
κ)−
− cos(d(x, y)
2 sin(d(x, z)
√
sin(d(v, z)
κ)
√
√
κ)
κ)
√
cos(d(x, y)
κ)L2.
(ii)
√
sin2 d(y, v)
2
κ
√
≤ L2 sin2 d(x, y)
2
√
≤ L2 sin2 d(x, y)
√
2
≤ L2 sin2 d(x, y)
√
2
≤ L2 sin2 d(x, y)
2
κ
κ
κ
(cid:18)
+ L1
since sin2 a + b
+
√
sin2 d(y, q)
2
≤ sin2 a
2
2
κ
+
+
κ
+
1
2
√
(1 − L1 − L2) + L1 sin2 d(y, z)
√
2
(1 − L1 − L2) + L1 sin2 (d(y, q) + d(q, z))
1
κ
√
2
(1 − L1 − L2) + L1 sin2 (d(y, q) + d(y, v))
1
2
(1 − L1 − L2)
1
√
2
κ
+ sin2 d(y, v)
2
sin(d(y, q)
(cid:19)
√
1
2
κ)
+
κ
κ
2
2
+ sin2 b
2
+
1
2
sin a for a, b ∈ [0, π] .
15
It follows that
√
sin2 d(y, v)
2
κ
√
(1 − L1) ≤ L2 sin2 d(x, y)
2
κ
(cid:18)
(cid:18)
(cid:18)
(cid:18)
+ L1
sin2 d(y, q)
2
√
≤ L2 sin2 d(x, y)
κ
2
+ L1
sin2 d(y, q)
2
√
κ
≤ L2 sin2 d(x, y)
2
+
√
1
2
κ
(1 − L1 − L2)
√
sin(d(y, q)
+
1
2
(cid:19)
κ)
+
√
1
2
κ
(1 − L1 − L2)
√
κ
+ sin
d(y, q)
2
L2(1 − cos(d(x, v)
√
κ))
(cid:19)
(cid:19)
√
κ
− 1
√
2
κ
+ L1
sin2 d(y, q)
2
+ sin
d(y, q)
2
by (16).
(cid:18)
√
sin2 d(x, y)
2
κ
√
sin2 d(y, q)
2
√
− sin2 d(x, v)
2
κ
√
d(y, q)
2
κ
κ
+ sin
L2
1 − L1
L1
1 − L1
+
(cid:19)
(cid:19)
.
Thus,
√
sin2 d(y, v)
2
κ
≤
√
By assumption, we have that sin2 d(x, y)
2
κ
≥ 1 and (17), it follows that
√
− sin2 d(x, v)
2
κ
≤ 0. Using the
fact that
L2
1 − L1
√
sin2 d(y, v)
2
κ
√
κ
≤ sin2 d(x, y)
2
√
1
s cos(M
+
κ)
κ
(cid:18)
√
− sin2 d(x, v)
√
2
sin2 d(y, q)
2
κ
√
d(y, q)
2
κ
+ sin
(cid:19)
.
5 Effective rates of asymptotic regularity
We assume the hypothesis of Proposition 3.2. As in [23, 16, 17], the main tool
in obtaining rates of asymptotic regularity is the following quantitative lemma,
which is a slight reformulation of [17, Lemma 1].
Lemma 5.1. Let (αn)n≥1 be a sequence in [0, 1] and (an)n≥1, (bn)n≥1 be se-
quences in R+ such that
an+1 ≤ (1 − αn+1)an + bn
for all n ∈ Z+.
(19)
16
∞(cid:88)
Assume that
bn is convergent with Cauchy modulus γ and
∞(cid:88)
n=1
αn+1 diverges
with rate of divergence θ.
n=1
Then,
lim
n→∞ an = 0 with rate of convergence Σ given by
(cid:18)
(cid:16) ε
(cid:17)
2
(cid:26)(cid:24)
(cid:18) 2P
(cid:19)(cid:25)
ε
(cid:27)(cid:19)
Σ(ε, P, γ, θ) = θ
γ
+ max
ln
, 1
+ 1,
(20)
where P > 0 is an upper bound on (an).
A second useful result, which is also needed in the metastability proof, is the
following:
Lemma 5.2. For all n ≥ 1, let
µn = 1 − sin ((1 − λn)M
√
sin(M
κ)
√
κ)
∈ (0, 1).
(21)
(ii)
λn+1 = ∞ with rate of divergence θ yields
(cid:18)(cid:24)
(cid:25)
(cid:19)
n
.
1
cos(M
√
κ)
∞(cid:88)
n=1
µn+1 = ∞ with rate of
Then
(i) µn ≥ λn cos (M
√
κ) for all n ≥ 1.
∞(cid:88)
n=1
divergence θ(n) = θ
Proof.
(i) One has
√
cos (2−λn)M
2
√
κ)
√
κ
√
κ
≥ 2 sin λnM
2
√
cos (M
sin (M
κ)
√
κ)
µn =
κ
2 sin λnM
2
√
sin (M
≥ λn cos(cid:0)M
κ(cid:1) .
(ii) Follows immediately from (i).
Lemma 5.3. For all n ∈ Z+
d(xn, xn+1) ≤ (1 − µn+1)d(xn−1, xn) + Mλn+1 − λn.
(22)
Proof. Let us denote for simplicity un = λn+1u + (1 − λn+1)T xn−1. Then,
d(xn, un) = λn+1 − λnd(u, T xn−1) ≤ Mλn+1 − λn and
d(un, xn+1) ≤ sin(cid:0)(1 − λn+1)M
sin(cid:0)M
κ(cid:1)
√
κ(cid:1)
√
d(xn−1, xn) by Lemma 4.1.
17
5.1 Proof of Proposition 3.2
Let Φ, Φ be given by (6) and (7). Apply Lemma 5.1 with
an = d(xn, xn−1),
bn = Mλn+1 − λn and αn = µn,
∞(cid:88)
(cid:16) ε
(cid:17)
M
and use Lemma 5.2.(ii) and the fact that
bn is convergent with Cauchy
modulus γ(ε) = γ
convergence Φ.
n=1
lim
n→∞ d(xn, xn+1) = 0 with rate of
Since d(xn, T xn) ≤ d(xn, xn+1) + M λn+1 for all n ≥ 1, it follows easily that
(cid:3)
Φ is a rate of asymptotic regularity.
to conclude that
6 A quantitative Browder theorem
Let X be a complete CAT(κ) space, C ⊆ X a bounded closed convex subset
with diameter dC <
and T : C → C be nonexpansive.
A very important step in the convergence proof for Halpern iterations is the
construction of a sequence of approximants converging strongly to a fixed point
of T . Given t ∈ (0, 1) and u ∈ C, Lemma 4.1 yields that the mapping
Dκ
2
: C → C, T u
t (y) = tu + (1 − t)T y
T u
t
is a contraction, hence it has a unique fixed point zu
t ∈ C. Thus,
t = tu + (1 − t)T zu
zu
t .
(23)
(24)
Pi¸atek [26] obtained the following generalization to CAT(κ) spaces of an
essential result due to Browder [5, 6].
Theorem 6.1. [26] In the above hypothesis,
of T .
lim
t→0+
zu
t exists and is a fixed point
In the setting of Hilbert spaces, Browder proved the result using weak se-
quential compactness and a projection argument (to the set of fixed points of
T ). A new and elementary proof of Browder's result was given by Halpern [9]
when C is the closed unit ball and the starting point is u = 0. Generalizations
of Browder's theorem were obtained by Reich [28] for uniformly smooth Ba-
nach spaces, Goebel and Reich [8] for the Hilbert ball and Kirk [10] for CAT(0)
spaces.
Kohlenbach [12] applied proof mining methods to both Browder's original
proof and the extension of Halpern's proof to bounded closed convex C and ar-
bitrary u ∈ C, obtaining in this way quantitative versions of Browder's theorem
with uniform effective rates of metastability. As pointed out in [12, Remark 1.4],
one cannot expect in general to get effective rates of convergence. Since Kirk's
18
proof of the generalization of Browder's theorem to CAT(0) spaces is obtained
by a slight change of Halpern's argument, Kohlenbach's quantitative result goes
through basically unchanged to CAT(0) spaces (see [17, Proposition 9.3]).
In this section we obtain a quantitative version of Theorem 6.1. As a con-
sequence of Halpern's proof, for any nonincreasing sequence (tn) in (0, 1), one
) converges strongly to some point z ∈ C, which is a fixed point
gets that (zu
tn
of T if
n→∞ tn = 0. Our quantitative result gives rates of metastability for such
lim
sequences (zu
tn
Proposition 6.2. Let X be a complete CAT(κ) space, C ⊆ X bounded closed
and T : C → C be nonexpansive. Assume that
convex with diameter dC <
(tn) ⊆ (0, 1) is nonincreasing. Then for every ε ∈ (0, 1) and g : N → N,
) and this suffices for the proof of our main Theorem 3.4.
Dκ
2
(cid:1),
, zu
tj
) ≤ ε√
κ
∃K0 ≤ K(ε, g, M ) ∀i, j ∈ [K0, K0 + g(K0)](cid:0)d(zu
(cid:25)(cid:19)
√
ti
κ)
(cid:18)(cid:24) M
K(ε, g, M ) =(cid:101)g
and(cid:101)g(n) = n + g(n).
√
κ tan(M
1 − cos ε
(0),
where
Dκ
2
with dC ≤ M <
Proof. Let ε ∈ (0, 1) and g : N → N. We assume without loss of generality that
i < j, hence tj ≤ ti. Denote ui,j = tju + (1 − tj)T zu
. Then,
ti
ti
ti
d(u, zu
ti
) = (1 − ti)d(u, T zu
) ≤ (1 − tj)d(u, T zu
∈ [u, ui,j]. It follows by Lemma 4.1 that d(zu
so zu
ti
, zu
d(zu
).
ti
tj
We can apply now Lemma 4.2 with x = u, y = zu
tj
) ≥ π/2. Since (d(u, zu
to
get that d(u, zu
, ¯ui,j) a κ-comparison triangle of
ti
∆(u, zu
))n is a nondecreasing
tj
sequence in [0, M ], by an application of [12, Lemma 4.1], there exists K0 ≤
K(ε, g, M ) such that
) ≤ d(u, zu
, ui,j), one has ∠¯zu
) and for ∆(¯u, ¯zu
tj
tj
(¯u, ¯zu
tj
) = d(u, ui,j),
, ui,j) ≤ d(T zu
, z = ui,j, w = zu
ti
) ≤
, T zu
ti
tn
tj
tj
ti
∀i, j ∈ [K0, K0 + g(K0)]
d(u, zu
tj
) − d(u, zu
ti
) ≤
√
1 − cos ε
√
κ tan(M
κ)
(cid:19)
.
(cid:18)
Let now i < j ∈ [K0, K0 + g(K0)]. Then,
cos(d(u, zu
ti
)
√
κ) − cos(d(u, zu
tj
√
)
κ) ≤ sin(M
√
κ)
1 − cos ε
√
√
κ)
tan(M
= (1 − cos ε) cos(M
κ).
) ≥ π
2 , we have that
√
)
(¯u, ¯zu
tj
κ) cos(d(¯zu
ti
, ¯zu
tj
κ).
√
)
ti
Furthermore, by the cosine law and the fact that ∠¯zu
√
cos(d(¯u, ¯zu
tj
)
κ) ≤ cos(d(¯u, ¯zu
ti
19
It follows that
Hence
cos(d(zu
tj
√
)
, zu
ti
Thus, d(zu
tj
cos(d(u, zu
ti
√
)
≤ cos(d(u, zu
≤ cos(d(¯u, ¯zu
≤ cos(d(u, zu
√
κ) − (1 − cos ε) cos(M
κ)
√
κ) = cos(d(¯u, ¯zu
)
√
tj
κ) cos(d(¯zu
)
√
tj
κ) cos(d(zu
)
tj
√
)
√
√
, ¯zu
ti
, zu
ti
κ).
κ)
tj
ti
ti
κ)
)
)
√
)
, zu
ti
κ) ≥ 1 − (1 − cos ε)
√
cos(M
√
κ)
cos(d(u, zu
ti)
≥ cos ε.
κ)
κ ≤ ε and the proof is complete.
7 Effective rates of metastability
In this section we shall prove the main result of our paper, Theorem 3.4, hence
we assume that its hypotheses are satisfied. We give first some technical results
that will be needed in the proof.
7.1 Some useful lemmas
As in [16, 17], one of the main ingredients of our proof is a sequence obtained by
combining the Halpern iteration (xn) and the points zu
t . However, in the setting
of CAT(κ) spaces, its definition and the proofs of the necessary properties are
based on the much more involved technical lemmas from Section 4.
an ≤ 0 with effective rate
If (an) is a real sequence, we say that lim sup
n→∞
(25)
(26)
(cid:19)
.
(27)
Ψ : (0,∞) → Z+ if
Let us define
Proposition 7.1.
∀ε > 0∀n ≥ Ψ(ε) (an ≤ ε).
√
n = sin2 d(u, zu
t )
γt
2
(i) For n ≥ 1, if γt
κ
κ
.
√
− sin2 d(u, xn+1)
n ≥ 0, then
− sin2 d(xn+1, zu
t )
√
κ
2
,
2
n ≤ an
γt
t
(cid:18)
where
an =
1
cos(M
√
κ)
√
sin2 d(xn+1, T xn+1)
2
κ
+ sin
√
d(xn+1, T xn+1)
κ
2
(ii) γt
n ≤ an
t
for all n ≥ 1.
20
n ≤ 0 with effective rate Ψ(ε, κ, M, t, γ, θ, α) given by
(cid:27)(cid:19)(cid:19)
(cid:18)(cid:24)
γt
(cid:25)(cid:18)
(cid:26)(cid:24)
(cid:19)(cid:25)
(cid:27)
γ(L) + max
ln
κ)
, 1
, α(2L)
, (28)
(cid:18) 1
L
(iii) lim sup
n→∞
(cid:26)
Ψ = max
θ
where L =
(iv) For n ≥ 1,
√
1
cos(M
√
√
cos(M
4M
κ)tε
κ
.
√
sin2 d(xn+1, zu
t )
κ
2
√
κ)
√
κ
sin2 d(xn, zu
t )
2
≤ sin ((1 − λn+1)M
√
√
κ)
sin(M
√
sin (λn+1M
√
sin(M
√
sin (tM
sin(M
κ)
κ)
+
+
κ)
max{γt
n, 0}
√
κ)
sin2 M
κ
.
2
Proof.
(i) Apply Proposition 4.7.(ii) with x = u, y = xn+1, z = T zu
t , v =
zu
t , s = t, q = T xn+1 and note that
√
− sin2 d(u, zu
t )
2
√
It follows that sin2 d(xn+1, zu
t )
sin2 d(u, xn+1)
≤ −γt
√
κ
κ
2
n +
2
κ
= −γt
n ≤ 0.
an
t
, hence (i).
(ii) Obviously, since
≥ 0.
an
t
(iii) Since an ≤
1
cos(M
√
κ)
d(xn+1, T xn+1)
√
κ and, by Proposition 3.2, the
sequence (d(xn, T xn)) converges to 0 with rate of convergence Φ given by
(7), we get that lim sup
n→∞
n ≤ 0 with effective rate
γt
Ψ(ε, κ, M, t, γ, θ, α) = Φ
κ)tε
, κ, M, γ, θ, α
.
(cid:18) cos(M
√
√
κ
(cid:19)
(iv) By Proposition 4.5.(ii) with x = u, y = T xn, z = T zu
t , w = xn+1, v =
zu
t , r = λn+1 and s = t.
In fact, it suffices for the proof of the main theorem to consider the case
ti =
1
i + 1
, i ≥ 0. Then (ti) converges towards 0 with rate
We shall denote γti
n. Furthermore, zu
ti
will be simply denoted by zu
i .
n with γi
√
n = sin2 d(u, zu
i )
γi
2
κ
√
− sin2 d(u, xn+1)
2
κ
.
(29)
Thus,
(cid:24) 1
(cid:25)
.
ε
21
Lemma 7.2. Assume that i, j ≥ 0 and δ ∈ (0, 1) are such that d(u, T zu
d(u, T zu
i ) −
√
√
+ 2 sin
κ
M
2(j + 1)
+ sin2 δ
2
+ 2 sin
M
sin
δ
2
2
κ
.
j ) ≤ δ√
n ≤ γj
γi
κ . Then,
n + sin2 M
√
κ
2(j + 1)
Proof. We have that
n = sin2
γi
i
√
i+1 d(u, T zu
i )
√
2
≤ sin2 d(u, T zu
i )
2
(cid:32)
κ
√
κ
sin
≤
d(u, T zu
j )
2
√
≤ sin2 d(u, T zu
j )
2
Note that
√
sin2 d(u, T zu
j )
2
κ
= sin2
≤ sin2
√
κ
κ
− sin2 d(u, xn+1)
√
2
− sin2 d(u, xn+1)
κ
(cid:33)2
+ sin
δ
2
2
√
− sin2 d(u, xn+1)
κ
2
√
κ
2
√
κ
− sin2 d(u, xn+1)
+ sin2 δ
2
+ 2 sin
M
sin
δ
2
√
2
κ
.
j
j+1 d(u, T zu
j )
κ + 1
√
κ
j
j+1 d(u, T zu
j )
2
κ
√
j+1 d(u, T zu
j )
√
2
+ sin2 M
κ
2(j + 1)
+ 2 sin
√
κ
M
2(j + 1)
.
Finally, let us recall the following slight reformulation of [17, Lemma 5.2].
Lemma 7.3. Let ε ∈ (0, 2), g : N → N, L > 0, θ : Z+ → Z+ and ψ : (0,∞) →
Z+. Define
(cid:18)
(cid:16) ε
(cid:17) − 1 + max
(cid:26)(cid:24)
(cid:18) 3L
(cid:19)(cid:25)
(cid:27)(cid:19)
Θ = Θ(ε, L, θ, ψ) = θ
ψ
ln
ε
, 1
+ 1,
3
ε
∆ = ∆(ε, g, L, θ, ψ) =
3gε(Θ − ψ(ε/3))
,
where gε(n) = n + g(n + ψ(ε/3)).
Assume that
(i) (αn) is a sequence in [0, 1] such that the series
of divergence θ;
∞(cid:88)
n=1
αn diverges with rate
(cid:16) ε
(cid:17)
.
3
(ii) (tn) is a sequence of real numbers such that tn ≤ ε
3
for all n ≥ ψ
Let (sn) be a bounded sequence of real numbers with upper bound L satisfying
sn+1 ≤ (1 − αn)sn + αntn + ∆
for all n ≥ 1.
(30)
Then sn ≤ ε for all n ∈ [Θ, Θ + g(Θ)].
22
7.2 Proof of Theorem 3.4
Let ε ∈ (0, 2) and g : N → N be fixed. For simplicity, we omit parameters
κ, M, Φ, θ, α, β for all functionals in this proof. Let us define h : (0, 1) → R+ by
≤ 6δ. (31)
h(δ) = sin
(cid:18)
(cid:19)
(cid:19)
(cid:18)
+ sin
√
√
√
sin
κ
κ
κ
δ
2
δ
sin
2
√
M
+ 2 sin
√
κ)
sin2 ε
cos(M
36
2
. Then, h(ε0) ≤ cos(M
6
κ
4
δM
2
δM
2
√
κ)
+ 2
√
sin2 ε
κ
.
4
Take ε0 =
Applying Proposition 6.2 for ti =
1
(cid:18)(cid:24) M
, ε0 and f∗, we get the existence of
i + 1
√
√
κ tan(M
1 − cos ε0
(cid:25)(cid:19)
κ)
K1 ≤ K(ε0, f∗) =(cid:102)f∗
j ) ≤ ε0√
κ
(cid:24) 1
for all i, j ∈ [K1,(cid:102)f∗(K1)].
(cid:25)
and J = K0 + f (K0) = (cid:102)f∗(K1).
(0)
It follows that
such that d(zu
i , zu
Let K0 = K1 +
d(zu
J , zu
K0
) ≤ ε0√
κ
ε0
, hence
d(u, T zu
J ) ≤ d(u, T zu
≤ d(u, T zu
) +
K0
ε0√
κ
.
) + d(T zu
K0
K0
, T zu
J ) ≤ d(u, T zu
K0
) + d(zu
K0
, zu
J )
√
An application of Lemma 7.2 with i = J, j = K0 and δ = ε0 gives us
n ≤ γK0
γJ
≤ γK0
+ sin2 ε0
√
2
κ
n + sin2 M
2(K0 + 1)
+ 2 sin
+ 2 sin
+ 2 sin
√
M
sin
√
κ
κ
√
κ
2
ε0
2
M ε0
2
κ
M
√
2(K0 + 1)
M
√
2
κ)
sin2 ε
κ
+ sin2 M ε0
2
√
κ
.
n + sin2 ε0
+ 2 sin
2
n + h(ε0) ≤ γK0
n +
sin
ε0
2
cos(M
6
= γK0
4
Applying now Proposition 7.1.(iv) with t = 1
J+1 and recalling the definition
(21) of (µn), it follows that for all n ≥ 1,
sin2 d(xn+1, zu
J )
2
≤ (1 − µn+1) sin2 d(xn, zu
J )
2
√
κ
√
κ
(cid:38)
√
κ)
√
κ)
(cid:17)
+
√
sin(λn+1M
sin(M
(cid:16) 1
sin
n , 0}
max{γJ
√
sin(M
.
2
κ
κ
κ)
J+1 M
√
sin2 M
(cid:39)
κ) − sin(cid:0)(1 − λn+1)M
and
κ(cid:1) ,
√
+
√
23
Since J = K0 + f (K0) ≥
√
cos(M
κ) sin(λn+1M
M
κ
√
∆∗
κ
(sin2 ε
4 , g)
√
√
κ) ≤ sin(M
K0
it follows that
√
sin2 d(xn+1, zu
J )
2
κ
1
√
≤ (1 − µn+1) sin2 d(xn, zu
J )
2
√
+µn+1 max
(cid:26)
, 0
κ
(cid:27)
γJ
n
cos(M
κ)
+ ∆∗
K0
(cid:18)
√
sin2 ε
4
(cid:19)
, g
.
κ
(cid:27)
(cid:25)
n
κ)
(cid:19)
Letting t =
in Proposition 7.1.(iii), we get that
K0 + 1
(cid:18) 1
3
sin2 ε
for all n ≥ χ∗
K0
4
n ≤ γK0
(cid:26)
γJ
n +
cos(M
6
√
n ≤ cos(M
(cid:19)
γK0
6
√
κ)
(cid:18) cos(M
sin2 ε
√
κ
κ)
(cid:19)
κ
√
√
4
κ
,
= χK0
√
κ)
√
6
κ
4
sin2 ε
√
sin2 ε
4
√
. Thus,
√
≤ cos(M
3
κ)
sin2 ε
κ
4
(cid:18) 1
3
(cid:19)
.
,
√
4
κ
and so, max
γJ
n
cos(M
√
κ)
, 0
≤ 1
3
sin2 ε
κ
for all n ≥ χ∗
K0
4
sin2 ε
Furthermore, by Lemma 5.2, we have that
µn+1 = ∞ with rate of diver-
∞(cid:88)
n=1
(cid:18)(cid:24)
(cid:26)
(cid:18)
gence θ(n) = θ
√
1
cos(M
√
sn = sin2 d(xn, zu
J )
2
(cid:18)
√
κ
κ
By letting N = ΘK0
,
tn = max
K0
4
(cid:19)
ε = sin2 ε
, ∆ = ∆∗
√
sin2 ε
√
sin2 d(xn, zu
J )
2
≤ sin2 ε
κ
κ
4
√
. Hence, we can apply Lemma 7.3 with
(cid:27)
(cid:19)
,
γJ
√
n
cos(M
√
sin2 ε
4
, 0
κ)
κ
, g
, αn = µn+1,
L = sin2 M
√
2
κ
.
, it follows that for all n ∈ [N, N + g(N )],
κ
,
and so d(xn, zu
J ) ≤ ε
2
.
4
Obviously, d(xn, xm) ≤ ε for all m, n ∈ [N, N + g(N )]. One can easily see that
N ≤ Σ(ε, g).
(cid:3)
Acknowledgements:
Laurent¸iu Leu¸stean was supported by a grant of the Romanian National Au-
thority for Scientific Research, CNCS - UEFISCDI, project number PN-II-ID-
PCE-2011-3-0383.
Adriana Nicolae was supported by a grant of the Romanian Ministry of Educa-
tion, CNCS - UEFISCDI, project number PN-II-RU-PD-2012-3-0152.
24
References
[1] J. Avigad, P. Gerhardy, H. Towsner, Local stability of ergodic averages,
Trans. Amer. Math. Soc. 362 (2010), 261-288.
[2] J. Avigad, J. Rute, Oscillation and the mean ergodic theorem for uni-
formly convex Banach spaces, Ergodic Theory Dynam. Systems, available
on CJO2014. doi:10.1017/etds.2013.90.
[3] G. Birkhoff, The mean ergodic theorem, Duke Math. J. 5 (1939), 19-20.
[4] M. Bridson, A. Haefliger, Metric spaces of non-positive curvature, Springer-
Verlag, Berlin, 1999.
[5] F.E. Browder, Existence and approximation of solutions of nonlinear varia-
tional inequalities, Proc. Nat. Acad. Sci. U.S.A. 56 (1966), 1080-1086.
[6] F.E. Browder, Convergence of approximants to fixed points of nonexpansive
nonlinear mappings in Banach spaces, Arch. Rational Mech. Anal. 24 (1967),
82-90.
[7] R.E. Bruck, S. Reich, Accretive operators, Banach limits, and dual ergodic
theorems, Bull. Acad. Polon. Sci. 29 (1981), 585 -- 589.
[8] K. Goebel, S. Reich, Uniform convexity, hyperbolic geometry, and nonex-
pansive mappings, Marcel Dekker, Inc., New York and Basel, 1984.
[9] B. Halpern, Fixed points of nonexpanding maps, Bull. Amer. Math. Soc. 73
(1967), 957-961.
[10] W.A. Kirk, Geodesic geometry and fixed point theory, Seminar of Math-
ematical Analysis (Malaga/Seville, 2002/2003), Colecc. Abierta, 64, Univ.
Seville Secr. Publ., Seville (2003), 195-225.
[11] U. Kohlenbach, Applied Proof Theory: Proof Interpretations and their Use
in Mathematics, Springer, Heidelberg-Berlin, 2008.
[12] U. Kohlenbach, On quantitative versions of theorems due to F.E. Browder
and R. Wittmann, Adv. Math. 226 (2011), 2764-2795.
[13] U. Kohlenbach, Some logical metatheorems with applications in functional
analysis, Trans. Amer. Math. Soc. 357 (2005), 89-128.
[14] U. Kohlenbach, L. Leu¸stean, A quantitative mean ergodic theorem for uni-
formly convex Banach spaces, Ergodic Theory Dynam. Systems 29 (2009),
1907-1915. Erratum: Ergodic Theory Dynam. Systems 29 (2009), 1995.
[15] U. Kohlenbach, L. Leu¸stean, Asymptotically nonexpansive mappings in
uniformly convex hyperbolic spaces, J. Eur. Math. Soc. 12 (2010), 71 -- 92.
25
[16] U. Kohlenbach, L. Leu¸stean, On the computational content of convergence
proofs via Banach limits, Phil. Trans. Royal Society A 370 (2012), No. 1971,
3449-3463.
[17] U. Kohlenbach, L. Leu¸stean, Effective metastability of Halpern iterates in
CAT(0) spaces, Adv. Math. 231 (2012), 2526-2556.
[18] U. Kohlenbach, L. Leu¸stean, Addendum to "Effective metastability of
Halpern iterates in CAT(0) spaces", Adv. Math. 250 (2014), 650-651.
[19] U. Kohlenbach, P. Safarik, Fluctuations, effective learnability and metasta-
bility in analysis, Ann. Pure and Applied Logic 165 (2014), 266-304.
[20] G. Kreisel, On the interpretation of non-finitist proofs, part I, J. Symbolic
Logic 16 (1951), 241-267.
[21] G. Kreisel, On the interpretation of non-finitist proofs, part II, J. Symbolic
Logic 17 (1952), 43-88.
[22] U. Krengel, On the speed of convergence in the ergodic theorem, Monatsh.
Math. 86 (1978/79), 3-6.
[23] L. Leu¸stean, Rates of asymptotic regularity for Halpern iterations of non-
expansive mappings, J. Universal Comp. Sci. 13 (2007), 1680-1691.
[24] L. Leu¸stean, Proof mining in metric fixed point theory and ergodic theory,
Habilitation thesis, Technische Universitat Darmstadt, 2009.
[25] G.G. Lorentz, A contribution to the theory of divergent series, Acta Math.
80 (1948), 167-190.
[26] B. Pi¸atek, Halpern iteration in CAT(κ) spaces, Acta Math. Sin. (Engl.
Ser.) 27 (2011), 635-646.
[27] S. Reich, Almost convergence and nonlinear ergodic theorems, J. Approx.
Theory 24 (1978), 269-272.
[28] S. Reich, Strong convergence theorems for resolvents of accretive operators
in Banach spaces, J. Math. Anal. Appl. 75 (1980), 287-292.
[29] S. Saejung, Halpern iterations in CAT(0) spaces, Fixed Point Theory Appl.
2010 (2010), Article ID 471781, 13pp.
[30] K. Schade, U. Kohlenbach, Effective metastability for modified Halpern
iterations in CAT(0) spaces, Fixed Point Theory Appl. 2012, 2012:191, 19pp.
[31] N. Shioji, W. Takahashi, Strong convergence of approximated sequences
for nonexpansive mappings in Banach spaces, Proc. Amer. Math. Soc. 125
(1997), 3641-3645.
26
[32] T. Tao, Soft analysis, hard analysis, and the finite convergence principle,
Essay posted May 23, 2007, appeared in: T. Tao, Structure and Randomness:
Pages from Year One of a Mathematical Blog. AMS, 298pp., 2008.
[33] T. Tao, Norm convergence of multiple ergodic averages for commuting
transformations, Ergodic Theory Dynam. Systems 28 (2008), 657-688.
[34] M. Walsh, Norm convergence of nilpotent ergodic averages, Ann. Math.
175 (2012), 1667-1688.
[35] R. Wittmann, Approximation of fixed points of nonexpansive mappings,
Arch. Math. 58 (1992), 486-491.
27
|
1501.04400 | 2 | 1501 | 2015-05-14T03:50:47 | A counterexample shows that not every locally $L^0$--convex topology is necessarily induced by a family of $L^0$--seminorms | [
"math.FA"
] | This paper constructs a counterexample showing that not every locally $L^0$--convex topology is necessarily induced by a family of $L^0$--seminorms. Random convex analysis is the analytic foundation for $L^0$--convex conditional risk measures, this counterexample, however, shows that a locally $L^0$--convex module is not a proper framework for random convex analysis. Further, this paper also gives a necessary and sufficient condition for a locally $L^0$--convex topology to be induced by a family of $L^0$--seminorms. Finally, we give some comments showing that based on random locally convex modules, we can establish a perfect random convex analysis to meet the needs of the study of $L^0$--convex conditional risk measures. | math.FA | math | A counterexample shows that not every locally L0 -- convex
topology is necessarily induced by a family of L0 -- seminorms
Mingzhi Wu∗ Tiexin Guo
School of Mathematics and Statistics
Central South University
Changsha 410083, China
Email: [email protected], [email protected]
5
1
0
2
y
a
M
4
1
]
.
A
F
h
t
a
m
[
2
v
0
0
4
4
0
.
1
0
5
1
:
v
i
X
r
a
Abstract
This paper constructs a counterexample showing that not every locally L0 -- convex topology is
necessarily induced by a family of L0 -- seminorms. Random convex analysis is the analytic foundation
for L0 -- convex conditional risk measures, this counterexample, however, shows that a locally L0 --
convex module is not a proper framework for random convex analysis. Further, this paper also
gives a necessary and sufficient condition for a locally L0 -- convex topology to be induced by a
family of L0 -- seminorms. Finally, we give some comments showing that based on random locally
convex modules, we can establish a perfect random convex analysis to meet the needs of the study
of L0 -- convex conditional risk measures.
Keywords. locally L0 -- convex module, locally L0 -- convex topology, L0 -- seminorm
1
Introduction
It is well known that classical convex analysis (see [4]) is the analytic foundation for convex risk measures,
cf.[1, 2, 6]. However, classical convex analysis is not well suited to the study of conditional convex (or,
L0 -- convex) conditional risk measures defined on the spaces of unbounded financial positions. It is to
overcome this obstacle that Filipovi´c, Kupper and Vogelpoth [5] presented the module approach to
conditional risk. The key point in this module approach is to establish random convex analysis as the
analytic foundation for L0 -- convex conditional risk measures. To this, they introduced the notion of
locally L0 -- convex modules, as a module analogue of locally convex spaces. In the theory of locally convex
spaces, it is a basic fact that every locally convex topology can be induced by a family of seminorms, it
is to establish the module analogue of the basic fact that Filipovi´c, Kupper and Vogelpoth [5] further
proved that the locally L0 -- convex topology of every locally L0 -- convex module can be induced by a
family of L0 -- seminorms, namely, Theorem 2.4 of [5] as the basis for their whole paper. Unfortunately,
this paper provides a counterexample to show that Theorem 2.4 of [5] is wrong. Besides, in particular
this paper gives a necessary and sufficient condition for a locally L0 -- convex topology to be induced
1
by a family of L0 -- seminorms. Finally, we give some comments showing that based on random locally
convex modules, we can establish a perfect random convex analysis to meet the needs of the study of
L0 -- convex conditional risk measures.
The remainder of this paper is organized as follows:
in Section 2 we first recall some necessary
terminology and notation; in Section 3 we construct the counterexample mentioned above and further
give a necessary and sufficient condition for a locally L0 -- convex topology to be induced by a family
of L0 -- seminorms; finally, in Section 4 we give some comments showing that based on random locally
convex modules, we can establish a perfect random convex analysis to meet the needs of the study of
L0 -- convex conditional risk measures.
2 Terminology and notation
Let (Ω, F , P ) be a probability space, K the scalar field R of real numbers or C of complex numbers and
L0(F , K) be the algebra of all equivalence classes of K -- valued F -- measurable random variables on Ω.
Specially, L0 = L0(F , R). As usual, L0 is partially ordered by ξ 6 η iff ξ0(ω) ≤ η0(ω) for P -- almost all
ω ∈ Ω, where ξ0 and η0 are arbitrarily chosen representatives of ξ and η, respectively. According to [3],
(L0, 6) is a conditionally complete lattice. For a subset A of L0 with an upper bound (a lower bound),
∨A (accordingly, ∧A) stands for the supremum (accordingly, infimum) of A. Let ξ and η be in L0, we
use "ξ < η (or ξ ≤ η) on A" for "ξ0(ω) < η0(ω) (resp., ξ0(ω) ≤ η0(ω)) for P -- almost all ω ∈ A", where
A ∈ F , ξ0 and η0 are a representative of ξ and η, respectively.
Denote L0
+ = {ξ ∈ L0 ξ > 0} and L0
++ = {ξ ∈ L0 ξ > 0 on Ω}.
IA always denotes the equivalence class of IA, where A ∈ F and IA is the characteristic function of
A. For any ξ ∈ L0(F , K), ξ denotes the equivalence class of ξ0 : Ω → R+ defined by ξ0(ω) = ξ0(ω),
where ξ0 is an arbitrarily chosen representative of ξ.
For any ε ∈ L0
++, denote Bε = {ξ ∈ L0(F , K) ξ 6 ε}. Let
Tc = {V ⊂ L0(F , K) for every y ∈ V there exists ε ∈ L0
++ such that y + Bε ⊂ V },
then Tc is a Hausdorff topology on L0(F , K) such that (L0(F , K), Tc) is a topological ring, namely
the addition and multiplication operations are jointly continuous. D. Filipovi´c, M. Kupper and N.
Vogelpoth first introduced in [5] this kind of topology and further pointed out that Tc is not necessarily
a linear topology since the scalar multiplication mapping: α 7→ αx (x is fixed) is no longer continuous in
general. These observations led them to the study of a class of topological modules over the topological
ring (L0(F , K), Tc).
Definition 2.1 (see [5]). A topological L0(F , K) -- module (E, T ) is an L0(F , K) -- module E endowed
with a topology T such that the addition and module multiplication operations:
2
(i) (E, T ) × (E, T ) → (E, T ), (x1, x2) 7→ x1 + x2 and
(ii) (L0(F , K), Tc) × (E, T ) → (E, T ), (ξ, x) 7→ ξx
are continuous w.r.t. the corresponding product topologies.
Locally L0 -- convex topologies are defined as follows.
Definition 2.2 (see [5]). For a topological L0(F , K) -- module (E, T ), the topology T is said to be locally
L0 -- convex if there is a neighborhood base U of 0 ∈ E for which every U ∈ U is:
(i) L0 -- convex: ξx1 + (1 − ξ)x2 ∈ U for all x1, x2 ∈ U and ξ ∈ L0 with 0 6 ξ 6 1,
(ii) L0 -- absorbent: for all x ∈ E there is ξ ∈ L0
++ such that x ∈ ξU ,
(iii) L0 -- balanced: ξx ∈ U for all x ∈ U and ξ ∈ L0(F , K) with ξ 6 1.
In this case, (E, T ) is called a locally L0 -- convex module.
Given an L0(F , K) -- module E, an easy way to construct a locally L0 -- convex topology on E is by
a family of L0 -- seminorms on E. The notions of L0 -- norms, L0 -- seminorms and random locally convex
modules were introduced by Guo before 2009 and random normed modules and random locally convex
modules have been deeply developed under the (ε, λ) -- topology, see Section 4 of this paper for the related
terminology. Let us first recall the notion of L0 -- seminorms.
Definition 2.3 Let E be an L0(F , K) -- module, a function k · k : E → L0
+ is called an L0 -- seminorm on
E if:
(i) kξxk = ξkxk for all ξ ∈ L0(F , K) and x ∈ E,
(ii) kx1 + x2k 6 kx1k + kx2k for all x1, x2 ∈ E.
Furthermore, an L0 -- seminorm k · k on E is called an L0 -- norm if kxk = 0 implies x = 0.
Proposition 2.4 (see [5]). Let E be an L0(F , K) -- module and P a family of L0-seminorms on E, for
finite Q ⊂ P and ε ∈ L0
++ we define
then
UQ, ε = {x ∈ E kxk 6 ε, ∀ k · k ∈ Q},
UP = {UQ, ε Q ⊂ P finite and ε ∈ L0
++}
forms a neighborhood base of 0, of some locally L0 -- convex topology on E, called the topology induced by
P.
Definition 2.5 (see [5]). Let E be an L0(F , K) -- module, the random gauge function pU : E → L0
+ of
an L0 -- absorbent set U ⊂ E is defined by
pU (x) = ∧{ξ ∈ L0
+ x ∈ ξU },
∀x ∈ E.
3
It is proved in [5] that: if U ⊂ E is L0 -- convex, L0 -- absorbent and L0 -- balanced, then the random
gauge function pU is an L0 -- seminorm on E and pU (x) = ∧{ξ ∈ L0
++ x ∈ ξU }, ∀x ∈ E.
For the subsequent use, we give two simple facts about random gauge function as follows.
Proposition 2.6 Let E be an L0(F , K) -- module, then we have the following:
(1). For any L0 -- seminorm p on E, let V = {x ∈ E p(x) 6 1}, then pV = p;
(2). For any finite family P of L0 -- seminorms on E and ε ∈ L0
++, let U = {x ∈ E p(x) 6 ε, ∀p ∈ P},
then {x ∈ E pU (x) 6 1} = U .
Proof. (1). For any given x ∈ E, we have pV (x) = ∧{ξ ∈ L0
++ x ∈ ξV } = ∧{ξ ∈ L0
++ ξ−1x ∈ V } =
∧{ξ ∈ L0
++ ξ−1p(x) 6 1} = ∧{ξ ∈ L0
++ p(x) 6 ξ} = p(x).
(2). The inclusion U ⊂ {x ∈ E pU (x) 6 1} is clear from the definition, so it only needs to show
the reverse inclusion. Let x be an element of E such that pU (x) 6 1, then, for any p ∈ P and δ ∈ L0
++
such that x ∈ δU , we have that p(x) 6 δ · ∨{p(y) : y ∈ U } 6 δε, therefore, p(x) 6 ε · ∧{δ ∈ L0
++ x ∈
δU } = εpU (x) 6 ε.
(cid:3)
3 A counterexample and a necessary and sufficient condition
for a locally L0 -- convex topology to be induced by a family
of L0 -- seminorms
To construct an example of a locally L0 -- convex topology which can not be induced by any family of
L0 -- seminorms, it is clear that we first need to find a new method (namely, not by use of L0 -- seminorms
as in Proposition 2.4) to construct a locally L0 -- convex topology. The following proposition is the basis
for our method.
Proposition 3.1 Let E be an L0(F , K)-module, U a family of L0 -- convex, L0 -- absorbent and L0 --
balanced subsets of E which satisfies the following three conditions:
(1). For any U1, U2 ∈ U, there exists U3 ∈ U such that U3 ⊂ U1 ∩ U2,
(2). For any U ∈ U, there exists V ∈ U such that V + V ⊂ U ,
(3). For any U ∈ U and ε ∈ L0
++, there exists V ∈ U such that εV ⊂ U .
Let T = {V ⊂ E for any y ∈ V , there is U ∈ U such that y + U ⊂ V }, then T is a locally L0 -- convex
topology on E and U is a neighborhood base of T at 0.
Proof.
It is easily seen that T is a topology on E.
It remains to show that (E, T ) is a topological
L0(F , K) -- module.
For any x, y ∈ E and U ∈ U, by (2), there exists V ∈ U such that V + V ⊂ U , it follows that
(x + V ) + (y + V ) = (x + y) + (V + V ) ⊂ (x + y) + U , thus the addition operation is continuous.
4
For any t ∈ L0(F , K), x ∈ E and U ∈ U, by (2), there exists V ∈ U such that V1 + V1 ⊂ U .
Since V1 is L0-absorbent, there exists ε ∈ L0
++ such that εx ∈ V1. Let δ = ε + t, according to (3),
there exists V2 ∈ U such that δV2 ⊂ V1. Noting that V1 is L0 -- balanced, we can further obtain that
Bεx := {ξx :
ξ 6 ε} ⊂ V1 and (t + Bε)V2 ⊂ BδV2 ⊂ V1, therefore,
(t + Bε)(x + V2) = tx + Bεx + (t + Bε)V2 ⊂ tx + V1 + V1 ⊂ tx + U,
which means that the module multiplication operation is continuous.
(cid:3)
Now we can give an example of a locally L0 -- convex module (E, T ) for which the topology T can
not be induced by any family of L0 -- seminorms on E.
For the sake of convenience, we first recall a notation here. For a subset C of an L0 -- module E, we
denote by
spanL0(C) :=( n
Xi=1
the L0 -- submodule of E generated by C.
xici (cid:12)(cid:12) ci ∈ C, xi ∈ L0, 1 ≤ i ≤ n, n ∈ N)
Example 3.2 Let Ω be the set of all positive integers, F the σ -- algebra of all the subsets of Ω, the
probability P on (Ω, F ) defined by P ({j}) = 2−j, ∀j ∈ Ω. Let E = L0, M = spanL0{ I{j} : j ∈ Ω},
++, recall that Bε = {x ∈ E x 6 ε}, then both M and Bε are L0 -- convex and
and for any ε ∈ L0
L0 -- balanced. In addition, Bε is L0-absorbent. Let Uε = M + Bε, then Uε is an L0 -- convex, L0-absorbent
and L0 -- balanced subset of E. Moreover, the family U = {Uε : ε ∈ L0
++} satisfies the three conditions in
Proposition 3.1, in fact, for any ε, δ ∈ L0
++, we have Uε∧δ ⊂ Uε ∩ Uδ, Uε/2 + Uε/2 ⊂ Uε and εUδ/ε ⊂ Uδ.
Therefore, U forms a neighborhood base of 0 of a locally L0 -- convex topology (denoted by T ) on E. We
have that the topology T can not be induced by any family of L0 -- seminorms on E.
Proof. Fix one ε ∈ L0
++, we calculate the random gauge function pUε . For each x ∈ E and each j ∈ Ω,
since
I{j}x ∈ spanL0{ I{j}} ⊂ M = δM ⊂ δUε, ∀δ ∈ L0
++,
we have that I{j}pUε(x) = pUε( I{j}x) 6 ∧L0
++ = 0, namely, pUε(x) = 0, ∀x ∈ E.
We show that T can not be induced by any family of L0 -- seminorms by contradiction. Assume that
P is a family of L0 -- seminorms on E which induces the topology T , then for any p ∈ P, there exists Uε
such that Uε ⊂ V := {x ∈ E : p(x) 6 1}, according to Proposition 2.6, we have that p = pV 6 pUε = 0,
thus p must also be zero. It follows that the family of L0 -- seminorms P is actually a singleton {0}.
Hence, the locally L0 -- convex topology induced by P is the trivial chaos topology which consists of ∅
and E. However, we claim that M is a proper T -- closed L0 -- submodule of E, which implies that T is
not trivial. This is a contradiction.
5
It remains to show the claim to complete the proof. Clearly, M is a proper subset of E, we only need
ε ∈ L0
to show that M is T -- closed. Since the T -- closure of M equalsT{M + Uε : ε ∈ L0
++}, we need to show that T{M + Bε : ε ∈ L0
for an arbitrarily given x ∈ E which is not in M , we need to find an ε ∈ L0
++} = T{M + Bε : ε ∈ L0
++} =T{M + M + Bε :
++} = M , that is to say,
++ such that x is not in
M + Bε. To this end, let S = {j ∈ Ω : x(j) 6= 0}, then S is an infinite set, otherwise, if S is finite, then
x ∈ spanL0{ I{j} : j ∈ S} ⊂ M . Define ε ∈ L0
++ by
ε(j) =( 1
2 x(j),
1,
j ∈ S;
j ∈ Ω \ S,
then x is not in M + Bε. In fact, if x = m + y for some m ∈ M and y ∈ Bε, then for each j ∈ S, we
have that m(j) = x(j) − y(j) ≥ x(j) − y(j) ≥ x(j) − ε(j) = 1
2 x(j) > 0. However, according to
(cid:3)
the definition of M , {j ∈ Ω : m(j) 6= 0} should be a finite set.
Remark 3.3 In the above example, one can easily see that the topology T is not Hausdorff since the
closure of {0} equals T{0 + Uε : ε ∈ L0
++} = M . By considering the
quotient L0 -- module E/M and the quotient topology (denoted by TΠ) on it, we get an example of a
locally L0 -- convex module (E/M, TΠ) which is Hausdorff but the topology cannot be induced by any
family of L0 -- seminorms.
++} = T{M + Bε : ε ∈ L0
Remark 3.4 In fact, for an arbitrarily given probability space (Ω, F , P ) which is not essentially gen-
erated by finitely many P -atoms, by making a slight modification to Example 3.2, we can always give
a locally L0 -- convex topology for L0 which cannot induced by any family of L0 -- seminorms. In fact, in
such a situation, there exists a countable partition {An : n ∈ N} of Ω to F such that each An has
positive probability, we can verify this fact as follows: first, (Ω, F , P ) has at most countably many dis-
joint P -atoms, which is denoted by {Bn : n = 1, 2, . . . , n0}, where n0 is a positive integer or n0 = +∞,
further, let Ω′ = Ω \ ∪n0
n=1Bn, then P (Ω′) > 0 and Ω′ does not include any P -atoms, and hence there
is a countable disjoint family {Cn ∈ F : n ∈ N} such that P (Cn) = P (Ω′)
2n
for each positive integer n,
now let {An : n ∈ N} = {Bn : n = 1, 2, . . . , n0} ∪ {Cn : n ∈ N}, it is clear that {An : n ∈ N} is a
countable partition of Ω to F . We only need to set M = spanL0{ IAn : n ∈ N} and still let Uε = M + Bε
++, then the locally L0 -- convex topology for L0 induced by the local neighborhood base
for each ε ∈ L0
++} meets our need, one can complete the verification only by replacing I{n} with IAn in
the proof of Example 3.2. Finally, we know from the referee's report that J. M. Zapata has also, inde-
{Uε : ε ∈ L0
pendently, obtained a similar result in [19], where J. M. Zapata only assumes that Ω has a countable
partition of positive probabilities, so he also obtained an enough general counterexample although [19] did
not give the details of the verification of his counterexample and he did not consider a counterexample
with the Hausdorff property.
6
Let (E, T ) be a locally L0 -- convex module. Assume that U is a neighborhood base of 0 ∈ E such
that each U ∈ U is L0 -- convex, L0 -- absorbent and L0 -- balanced. As pointed out in [15], if the inclusion
relation
{x ∈ E pU (x) < 1 on Ω} ⊂ U ⊂ {x ∈ E pU (x) 6 1}
holds true for each U ∈ U, in particular if the relation
U = {x ∈ E pU (x) 6 1}
(1)
holds true for each U ∈ U, then T can be induced by the family of L0 -- seminorms {pU : U ∈ U}.
Conversely, assume that P is a family of L0 -- seminorms on E which induces T , then for the neighborhood
base UP as in Proposition 2.4, according to Proposition 2.6, each U ∈ UP satisfies the relation (1). What
is more important is that the relation (1) helps us look for a necessary and sufficient condition for a
locally L0 -- convex topology to be induced by a family of L0 -- seminorms. In fact, a sufficient condition
was already given in [15], which requires that U has the countable concatenation property. Let us recall
this important notion as follows.
It should be pointed out that when Guo introduced the countable concatenation property for a
subset of an L0(F , K) -- module E, E is assumed to have the following property:
(C)
for any x, y ∈ E, if there is a countable partition {An, n ∈ N} of Ω to F such that IAn x = IAn y
for each n ∈ N , then x = y.
Guo proved in [9] that every random locally convex module has the above property (C), however, up
to now, no one has ever shown that every locally L0 -- convex module necessarily has the property (C),
in this paper for our purpose it is convenient to quit the requirement of the above property (C) and
rephrase the notion of the countable concatenation property in a slightly wide sense as follows.
Definition 3.5 Let E be an L0(F , K) -- module. A sequence {xn, n ∈ N} in E is countably concatenated
in E with respect to a countable partition {An, n ∈ N} of Ω to F if there is x ∈ E such that IAn x = IAn xn
for each n ∈ N, in which case we denote the set of all such x by P∞
n=1
∞
IAn xn, namely,
IAn xn = {x ∈ E x ∈ E such that IAn x = IAn xn for each n ∈ N}.
Xn=1
A subset G of E is said to have the countable concatenation property if each sequence {xn, n ∈ N} in
G is countably concatenated in E with respect to an arbitrary countable partition {An, n ∈ N} of Ω to
F and P∞
n=1
IAn xn ⊂ G.
In fact, Guo et al [15] already proved Proposition 3.6 and Corollary 3.7 below.
7
Proposition 3.6 (See [15]). Let (E, T ) be a locally L0 -- convex module and U an L0 -- convex, L0 --
absorbent and L0 -- balanced subset with the countable concatenation property. Then {x ∈ E pU (x) <
1 on Ω} ⊂ U ⊂ {x ∈ E pU (x) 6 1}.
Corollary 3.7 (See [15]). Let (E, T ) be a locally L0 -- convex module. If there exists a neighborhood base
U of 0 ∈ E such that each U ∈ U is an L0 -- convex, L0 -- absorbent and L0 -- balanced subset with the the
countable concatenation property, then T can be induced by the family of L0 -- seminorms {pU : U ∈ U}.
In order to give a necessary and sufficient condition for a locally L0 -- convex topology to be induced
by a family of L0 -- seminorms, we need Definition 3.8 below, which is based on Guo's earlier work [8],
where the notion of the relative countable concatenation property was first considered although the
terminology of the relative countable concatenation property did not occur in [8].
Definition 3.8 A subset G of an L0(F , K) -- module E is said to have the relative countable concate-
nation property if for a sequence {xn, n ∈ N} in G and a countable partition {An, n ∈ N} of Ω to F ,
IAn xn ⊂ G whenever {xn, n ∈ N} is countably concatenated in E with respect to
n=1
we always have P∞
{An, n ∈ N}.
Remark 3.9 For an arbitrary L0(F , K) -- module E, E need not have the countable concatenation prop-
erty, but it is clear that E as a subset of itself always has the relative countable concatenation property.
Furthermore, one can easily see that"G has the relative countable concatenation property" is the same
as "G has the countable concatenation property" for every subset G when E has the countable concate-
nation property. In Definition 3.8, the adjective "relative" means that whenever a sequence {xn, n ∈ N}
in G is countably concatenated "in E" with respect to a countable partition {An, n ∈ N}, then this
sequence must be countably concatenated "in G" with respect to the countable partition {An, n ∈ N}.
Theorem 3.10 Let (E, T ) be a topological L0(F , K) -- module, then the following two statements are
equivalent to each other:
(1). The topology T can be induced by a family of L0 -- seminorms on E;
(2). There exists a neighborhood base U of 0 ∈ E such that each element U ∈ U is an L0 -- convex,
L0 -- absorbent and L0 -- balanced subset with the relative countable concatenation property.
Proof. (1) ⇒ (2). Assume that the topology T can be induced by a family of L0 -- seminorms P on E.
For any ε ∈ L0
++ and finite Q ⊂ P, let UQ, ε = {x ∈ E kxk 6 ε, ∀ k · k ∈ Q}, then U := {UQ, ε Q ⊂
P finite and ε ∈ L0
++} is a neighborhood base of 0. We need only to show that each UQ, ε has the
relatively countable concatenation property. To this end, assume that {xn, n ∈ N} is a sequence in
UQ, ε which is countably concatenated in E with respect to a countable partition {An, n ∈ N} of Ω to
8
IAn xn, then for each k · k ∈ Q, we have that
F . If x ∈P∞
n=1
kxk = (
∞
∞
∞
Xn=1
IAn )kxk =
Xn=1
thus x ∈ UQ, ε and this in turn implies that P∞
IAnkxk =
Xn=1
k IAnxk =
∞
Xn=1
k IAn xnk 6
IAn ε = ε,
∞
Xn=1
IAn xn ⊂ UQ, ε.
n=1
(2) ⇒ (1). If (2) holds true, we will show that the topology T can be induced by the family of
L0 -- seminorms {pU : U ∈ U}.
It suffices to show that {x ∈ E pU (x) < 1 on Ω} ⊂ U . Let x
be a element of E such that pU (x) < 1 on Ω. Let A = {A ∈ F IAx ∈ U } and A = ess.supA,
according to [5, Proposition 2.25], pU (x) > 1 on Ac, thus P (Ac) = 0, namely, P (A) = 1. Since U
is L0 -- convex, A is directed upward and there exists an increasing sequence {An, n ∈ N} in A such
is a countable partition of Ω to F , and for each n we have that IBn x = IBn ( IAn x) ∈ U since U is
IBn x ⊂ U since U has the relative countable concatenation property.
that A = Sn∈N An. Let B1 = A1 ∪ Ac, B2 = A2 \ A1, . . . , Bn = An \ An−1, . . . , then {Bn, n ∈ N}
L0 -- balanced. Thus x ∈P∞
IBn
n=1
(cid:3)
Remark 3.11 Recently, we know from the referee's report that J.M.Zapata [19] also independently
presented the notion of being closed under the countable concatenation operation and further established
the characterization theorem for a locally L0 -- convex topology to be induced by a family of L0-seminorms,
see Page 6 and Theorem 2.1 in [19]. One can easily see that the relative countable concatenation property
in our paper is the same as being closed under the countable concatenation operation, so J.M.Zapata's
Theorem 2.1 is also the same as our Theorem 3.10.
4 Concluding remarks
Random convex analysis was first studied in [5], where a locally L0 -- convex module is chosen as the
space framework for random convex analysis. However, Example 3.2 and Remark 3.4 show that a locally
L0 -- convex module is not a proper space framework for random convex analysis. First, according to the
notion of a locally L0 -- convex module, only the locally L0 -- convex topology can be employed, whereas
for a random locally convex module (E, P), the two kinds of topologies, namely, the (ε, λ) -- topology
and the locally L0 -- convex topology can be used. Second, the locally L0 -- convex topology seems more
intuitive, but the (ε, λ) -- topology is more natural from probability theory, in particular a random locally
convex module absorbs both the advantages of this two kinds of topologies and the work of [15] shows
that a random locally convex module seems to be a more proper framework for random convex analysis.
It is to overcome the disadvantage of a locally L0 -- convex module that Guo et al [15] choose a random
locally convex module as a space framework to establish random convex analysis. Let us recall the
notion of a random locally convex module as follows.
9
Definition 4.1 (See [7, 13]). An ordered pair (E, P) is called a random locally convex module (briefly,
an RLC module) over K with base (Ω, F , P ) if E is an L0(F , K) -- module and P a family of L0 --
seminorms on E such that ∨{kxk : k · k ∈ P} = 0 iff x = θ (the null element of E).
Let P be a family of L0 -- seminorms on an L0(F , K) -- module E and Pf the family of all finite subsets
Q of P. For each Q ∈ Pf , the L0 -- seminorm k · kQ is defined by kxkQ = ∨{kxk : k · k ∈ Q}, ∀x ∈ E.
Definition 4.2 (See [7, 13].) Let (E, P) be an RLC module over K with base (Ω, F , P ). For any
positive numbers ε and λ with 0 < λ < 1 and Q ∈ Pf , let Nθ(Q, ε, λ) = {x ∈ E P {ω ∈ Ω kxkQ(ω) <
ε} > 1 − λ}, then {Nθ(Q, ε, λ) Q ∈ Pf , ε > 0, 0 < λ < 1} forms a local base at θ of some Hausdorff
linear topology on E, called the (ε, λ) -- topology induced by P.
Since Guo's paper [9], random metric theory has come into such a model that random locally convex
modules are developed by simultaneously considering the above two kinds of topologies, which makes
random metric theory deeply developed (see, e.g.
[10, 20, 12, 14, 17, 18, 11]) since the connection
between basic results derived from the two kinds of topologies has been established in [9]. Based on
these deep advances, a complete random convex analysis has been developed in [15] and some concrete
applications of random convex analysis to L0 -- convex conditional risk measures are also given in [16].
Acknowledgements.
The first author is supported by Central South University Postdoctoral Science Foundation. The
second author is supported by National Natural Science Foundation of China (Grant No. 11171015).
The authors would like to thank the referee for pointing out to us the reference [19], which makes us, for
the first time, see J.M.Zapata's work [19]. Although [19] appeared on arXiv on April 29, 2014, sooner
than the submitting time of our paper, one of the authors, namely, Prof.Guo told J.M.Zapata (on April
2, 2014) that we had obtained Counterexample 3.2 of our paper in an email to J.M.Zapata which did
not include any details of our paper, in fact we then had obtained all the results of our paper but
Remark 3.4(Remark 3.4 is added after reading [19]). Thus we can say that our results are independent
of [19].
References
[1] Artzner P, Delbaen F, Eber J M, Heath D. Coherent measures of risk. Math Finance, 1999, 9:
203 -- 228
[2] Delbaen F. Coherent risk measures on general probability spaces. In: Sandmann K, Schonbucher,
P J, eds. Advances in Finance and Stochastics. Berlin: Springer, 2002, 1-37
10
[3] Dunford N, Schwartz J T. Linear Operators (I). New York: Interscience, 1957
[4] Ekeland I, T´emam R. Convex Analysis and Variational Problems. New York: SIAM, 1999
[5] Filipovi´c D, Kupper M, Vogelpoth N. Separation and duality in locally L0−convex modules. J
Funct Anal, 2009, 256: 3996 -- 4029
[6] Filipovi´c D, Svindland G. Convex risk measures beyond bounded risks, or the canonical model
space for law-invariant convex risk measures is L1. Working paper. Vienna: Vienna Institute of
Finance, 2008
[7] Guo T X. Survey of recent developments of random metric theory and its applications in China
(II). Acta Anal Funct Appl, 2001, 3: 208 -- 230
[8] Guo T X. A comprehensive connection between the basic results and properties derived from two
kinds of topologies for a random locally convex module. 2009, arXiv:0908.1843
[9] Guo T X. Relations between some basic results derived from two kinds of topologies for a random
locally convex module. J Funct Anal, 2010, 258: 3024 -- 3047
[10] Guo T X. Recent progress in random metric theory and its applications to conditional risk measures.
Sci China Math, 2011, 54: 633 -- 660
[11] Guo T X. On some basic theorems of continuous module homomorphisms between random normed
modules. J Funct Spaces Appl, 2013, Article ID 989102, 13 pages.
[12] Guo T X, Shi G. The algebraic structure of finitely generated L0(F , K) -- modules and the Helly
theorem in random normed modules. J Math Anal Appl, 2011, 381: 833 -- 842
[13] Guo T X, Xiao H X, Chen X X. A basic strict separation theorem in random locally convex
modules. Nonlinear Anal, 2009, 71: 3794 -- 3804
[14] Guo T X, Yang Y J. Ekeland's variational principle for an ¯L0 -- valued function on a complete
random metric space. J Math Anal Appl, 2012, 389: 1 -- 14
[15] Guo T X, Zhao S E, Zeng X L. On random convex analysis -- the analytic foundation of the module
approach to conditional risk measures. 2012, arXiv:1210.1848
[16] Guo T X, Zhao S E, Zeng X L. The relations among the three kinds of conditional risk measures.
Sci China Math, 2014, 57: 1753 -- 1764
[17] Wu M Z. The Bishop-Phelps theorem in compete random normed modules endowed with the
(ε, λ) -- topology. J Math Anal Appl, 2012, 391: 648 -- 652
11
[18] Wu M Z. Farkas' lemma in random locally convex modules and Minkowski-Weyl type results in
L0(F , Rn). J Math Anal Appl, 2013, 404: 300 -- 309
[19] Zapata J M. On characterization of locally L0-convex topologies induced by a family of L0-
seminorms. 2014, arXiv:1404.0357v4
[20] Zhao S E, Guo T X. The random subreflexivity of complete random normed modules, Int J Math,
2012, 23: 1 -- 14
12
|
1509.01933 | 1 | 1509 | 2015-09-07T07:29:52 | Cohomological characterization of $T$-Lau product algebras | [
"math.FA"
] | Let $A$ and $B$ be Banach algebras and let $T$ be an algebra homomorphism from $B$ into $A$. The Cartesian product space $A\times B$ by $T$- Lau product and $\ell^{1}$- norm becomes a Banach algebra $A\times_{T}B$. We investigate the notions such as injectivity, projectivity and flatness for the Banach algebra $A\times_{T}B$. We also characterize Hochschild cohomology for the Banach algebra $A\times_{T}B$. | math.FA | math |
COHOMOLOGICAL CHARACTERIZATION OF T -LAU PRODUCT ALGEBRAS
N. RAZI AND A. POURABBAS
Abstract. Let A and B be Banach algebras and let T be an algebra homomorphism from B into A.
The Cartesian product space A × B by T - Lau product and ℓ1- norm becomes a Banach algebra A ×T B.
We investigate the notions such as injectivity, projectivity and flatness for the Banach algebra A ×T B.
We also characterize Hochschild cohomology for the Banach algebra A ×T B.
1. Introduction and Preliminaries
Suppose that A and B are Banach algebras and T : B → A is an algebra homomorphism. Then we
consider the Cartesian product space A × B with the following multiplication
(a, b) ×T (c, d) = (ac + T (b)c + aT (d), bd)
((a, b), (c, d) ∈ A × B),
which is denoted by A ×T B. Let kT k ≤ 1. Then we consider A ×T B with the following norm
k(a, b)k = kak + kbk
((a, b) ∈ A ×T B).
We note that A ×T B is a Banach algebra with this norm and it is called T -Lau product algebras
Whenever the Banach algebra A is commutative, Bhatt and Dabshi [1] have investigated the properties
of the Banach algebra A ×T B, such as Gelfand space, Arens regularity and amenability.
Whenever A is unital with unit element e and ϕ : B → C is a character on B, assume T : B → A is
defined by T (b) = φ(b)e. In this case the multiplication ×T corresponds with the product studied by Lau
[10]. Lau product was extended by Sangani Monfared for the general case and many basic properties of
this product are studied in[13].
In the definition of T -Lau product, we can replace condition kT k ≤ 1 with a bounded algebra homo-
morphism T , because if we consider the following norms
kakT = kT kkak (a ∈ A)
kbkT = kT kkbk (b ∈ B)
k(a, b)kT = kakT + kbkT
(a, b) ∈ A ×T B,
then all these norms are equivalent with the original norms. Clearly all results of this paper hold when
we consider these equivalent norms.
The authors in [12] for every Banach algebras A and B and for an algebra homomorphism T : B → A
with kT k ≤ 1 have investigated some homological properties of T - Lau product algebra A ×T B such as
2010 Mathematics Subject Classification. Primary: 46M10. Secondary: 46H25, 46M18.
Key words and phrases. T -Lau product, injectivity, projectivity, flatness, Hochschild cohomology.
1
2
N. RAZI AND A. POURABBAS
approximate amenability, pseudo amenability, φ-pseudo amenability, φ- biflatness and φ-biprojectivity
and have presented the characterization of the double centralizer algebra of A ×T B.
Following [12], in this paper we studied the homological notions such as injectivity, projectivity and
flatness for the Banach algebra A ×T B. We also characterize the Hochschild cohomology for the Banach
algebra A ×T B.
2. Injectivity, Flatness and Projectivity
Let A be a Banach algebra. In this paper, the category of Banach left A-modules and Banach right
A-modules is denoted by A-mod and mod- A, respectively. We denote by B(E, F ) the Banach space
of all bounded operators from E into F . In the category of A-mod, we denote the space of bounded
morphisms from E into F by AB(E, F ). A function S ∈ B(E, F ) is called admissible if there exists
S ′ ∈ B(F, E) such that S ◦ S ′ ◦ S = S.
A. Ya. Helemskii introduced the concepts of injectivity and flatness for Banach algebras [5] and these
concepts have been investigated for different classes of Banach modules in [3, 4, 11, 14].
Definition 2.1. A Banach left A-module K is called projective if for every admissible epimorphism
S : E → F in A-mod, the induced map SA : AB(K, E) → AB(K, F ) defined by
is surjective.
SA(RA) = RA ◦ S
(RA ∈A B(K, E))
Definition 2.2. A Banach left A-module K is called injective if for every admissible monomorphism
S : E → F in A-mod, the induced map SA : AB(F, K) → AB(E, K) defined by
is surjective.
SA(RA) = RA ◦ S
(RA ∈ AB(F, K))
Definition 2.3. A Banach left A-module K is called flat if the dual module K ∗ in mod-A is injective
with the action defined by
where a ∈ A, x ∈ K and f ∈ K ∗.
(f · a)(x) = f (a · x),
Let A, B, and C be Banach algebras and let T : B → A be an algebra homomorphism with kT k ≤ 1.
We note that if A is a Banach left C-module, then A ×T B is a Banach left C-module via the following
action
c · (a, b) = (c · a + c · T (b), 0)
((a, b) ∈ A ×T B, c ∈ C).
Similarly if B is a Banach left C-module, then A ×T B is a Banach left C-module via the following action
c · (a, b) = (−T (c · b), c · b).
Theorem 2.4. Suppose that A and B are Banach algebras and T : B → A is an algebra homomorphism
with kT k ≤ 1. Suppose that C is Banach algebra and A ×T B is injective as C-module. Then A and B
are injective as C-module.
COHOMOLOGICAL CHARACTERIZATION OF T -LAU PRODUCT ALGEBRAS
3
Proof. Let A be a Banach left C-module and let F, K ∈ C-mod. Suppose that S ∈ CB(F, K) is admissible
and monomorphism. We will show that the induced map SA : CB(K, A) → CB(F, A) is onto.
We conclude that A ×T B ∈ C-mod via the following action
c · (a, b) = (c · a + c · T (b), 0).
Since A ×T B is injective, the induced map SA×T B : CB(K, A ×T B) → CB(F, A ×T B) is onto. Let λ ∈
CB(F, A) and f ∈ F . We define λ : F → A ×T B by eλ(f ) = (λ(f ), 0). Hence we have λ ∈ CB(F, A ×T B).
Since SA×T B : CB(K, A ×T B) → CB(F, A ×T B) is onto, there exists RA×T B : K → A ×T B such that
RA×T B(T (f )) = λ(f ) = (λ(f ), 0). We define RA : K → A by RA = PA ◦RA×T B, where PA : A×T B → A
is defined by pA(a, b) = a. Clearly RA ∈ CB(K, A) and RA ◦ T ′ = λ. So A is injective.
For injectivity of B, suppose that B is a Banach left C-module, F and K ∈ C-mod and S ∈ CB(F, K)
is admissible and monomorphism. We must show that the induced map SB : CB(K, B) → CB(F, B) is
onto. We have A ×T B ∈ C-mod with the following actions
Since A ×T B is injective, the induced map SA×T B : CB(K, A ×T B) → CB(F, A ×T B) is onto. Let µ ∈
c · (a, b) = (−T (c · b), c · b).
Since SA×T B : CB(K, A ×T B) → CB(F, A ×T B) is onto, there exists RA×T B : K → A ×T B such that
CB(F, B) and f ∈ F . We define eµ : F → A ×T B by eµ(f ) = (0, µ(f )). Then we have µ ∈ CB(F, A ×T B).
RA×T B(T (f )) = eµ = (0, µ(f )). We define RA : K → A by RA = PA ◦ RA×T B. Clearly RA ∈ CB(K, B)
and RA ◦ S = µ. Hence B is injective. This completes the proof.
(cid:3)
Let A, B and C be Banach algebras and let T : B → A be an algebra homomorphism with kT k ≤ 1.
We note that if A ×T B is the Banach left C-module, then A and B can be the Banach left C-modules
with the following actions
c · a = c · (a, 0) and c · b = c · (0, b),
where c ∈ C, a ∈ A and b ∈ B.
Theorem 2.5. Suppose that A and B are Banach algebras and T : B → A is an algebra homomorphism
with kT k ≤ 1. Suppose that C is a Banach algebra and A and B are injective as C-mod. Then A ×T B
is injective as C-mod.
Proof. Let A ×T B be a Banach left C-module and F, K ∈ C-mod. Let S ∈ CB(F, K) such that S
is admissible and monomorphism. We must show that the induced map SA×T B : CB(K, A ×T B) →
CB(F, A ×T B) is onto. We have A, B ∈ C-mod with the following actions
c · a = c · (a, 0) and c · b = c · (0, b),
where c ∈ C, a ∈ A and b ∈ B. Since A and B are injective, the induced maps SA : C B(K, A) → CB(F, A)
and SB : CB(K, B) → CB(F, B) are onto. Suppose that λ ∈ CB(F, A ×T B) and (a, b) ∈ A ×T B such
that λ(f ) = (a, b) for f ∈ F .
We define eλ : F → A by eλ(f ) = a + T (b) and eµ : F → B by eµ(f ) = b. Hence we have eλ ∈ CB(F, A)
and eµ ∈ CB(F, B). Since SA : CB(k, A) → CB(F, A) and SB : CB(K, B) → CB(F, B) are onto, there
exist RA : K → A and RB : K → B such that RA ◦ S(f ) = eλ(f ) = a + T (b) and RB ◦ S(f ) = eµ(f ) = b.
4
N. RAZI AND A. POURABBAS
We define RA×T B : K → A ×T B by RA×T B = qA ◦ RA + ηB ◦ RB, where ηB : B → A ×T B such that
ηB(b) = (−T (b), b). Clearly RA×T B ∈C B(K, A ×T B) and RA×T B ◦ S = λ. So A ×T B is injective. (cid:3)
Let A, B and C be Banach algebras. We note that if A∗ × B∗ is a Banach left C-module, then A∗ and
B∗ can be consider as Banach left C-modules via the following actions
c · a∗ = c · (a∗, 0) and c · b∗ = c · (0, b∗),
where c ∈ C, a∗ ∈ A∗ and b∗ ∈ B∗.
Theorem 2.6. Suppose that A and B are Banach algebras and T : B → A is an algebra homomorphism
with kT k ≤ 1. Suppose that C is a Banach algebra. Then A ×T B is flat as mod- C if and only if A and
B are flat as mod- C.
Proof. Let A ×T B be flat as mod-C. With a simple argument we can show that (A ×T B)∗ ∼= A∗ × B∗.
Hence by similar argument as in Theorem 2.4, one can show that A and B are flat Banach algebras as
mod-C
Conversely, let A and B be flat Banach algebras as mod-C, let F, K ∈ C-mod and let S ∈ CB(F, K)
such that S is admissible and monomorphism. Then we show that the induced map SA∗×B∗ : CB(K, A∗ ×
B∗) → C B(F, A∗ × B∗) is onto. We have A∗, B∗ ∈ C-mod with the following actions
c · a∗ = c · (a∗, 0) and c · b∗ = c · (0, b∗),
where c ∈ C, a∗ ∈ A∗ and b∗ ∈ B∗.
Since A∗ and B∗ are injective as left C-module, the induced maps SA∗ : CB(K, A∗) → C B(F, A∗) and
SB∗ : CB(K, B∗) → CB(F, B∗) are onto. Suppose that λ∗ ∈ C B(F, A∗ × B∗) and (a∗, b∗) ∈ A∗ × B∗
such that λ∗(f ) = (a∗, b∗) for f ∈ F .
CB(K, A∗) → CB(F, A∗) and SB∗ :
We define fλ∗ : F → A∗ by fλ∗(f ) = a∗ and µ∗ : F → B∗ by fµ∗(f ) = b∗. Hence we have fλ∗ ∈ CB(F, A∗)
and fµ∗ ∈ CB(F, B∗). Since SA∗ :
CB(K, B∗) → CB(F, B∗)
are onto, there exist RA∗ : K → A∗ and RB∗ : K → B∗ such that RA∗ ◦ S(f ) = fλ∗(f ) = a∗ and
RB∗ ◦ S(f ) = fµ∗(f ) = b∗.
We define RA∗×B∗ : K → A∗ × B∗ by RA∗×B∗ = qA∗ ◦ RA∗ + qB∗ ◦ RB∗ , where qA∗ : A∗ → A∗ × B∗
B : B∗ → A∗ × B∗ are defined by qA∗ (a∗) = (a∗, 0) and qB∗ (b∗) = (0, b∗), respectively. Clearly
and q∗
RA∗×B∗ ∈ CB(K, A∗ × B∗) and RA∗×B∗ ◦ S = λ∗. So A∗ × B∗ is injective as left C-module. This
completes the proof.
(cid:3)
Let A, B and C be Banach algebras and let T : B → A be an algebra homomorphism with kT k ≤ 1. If
A ×T B is a Banach left C-module, as we have seen before, A and B are Banach left C-modules.
Theorem 2.7. Suppose that A and B are Banach algebras and T : B → A is an algebra homomorphism
with kT k ≤ 1. Suppose that C is a Banach algebra. Then A ×T B is projective as C-mod if and only if
A and B are projective as C-mod.
COHOMOLOGICAL CHARACTERIZATION OF T -LAU PRODUCT ALGEBRAS
5
Proof. Let A ×T B be projective as Banach left C-module and F, K ∈ C-mod. Let S ∈ CB(K, F ) be
admissible and epimorphism. We show that the induced map SA×T B : CB(A×T B, K) → CB(A×T B, F )
is onto.
Since A, B ∈ C-mod are projective, the induced map SA :
CB(A, K) → CB(A, F ) and SB :
CB(B, K) → C B(B, F ) are onto. Let λ ∈ CB(A ×T B, F ) and f1, f2 ∈ F such that λ(a, 0) = f1, λ(0, b) =
f2 for (a, 0), (0, b) ∈ A ×T B. We define eλ : A → F by eλ(a) = λ(a, 0) = f1 and eµ : B → F by
eµ(b) = λ(0, b) = f2. Hence we have eλ ∈ CB(A, F ) and eµ ∈ CB(B, F ). Since SA : CB(A, K) → CB(A, F )
that S ◦ RA(a) = eλ(a) = f1 and S ◦ RB(b) = eµ(b) = f2. We define RA×T B : A ×T B → K by
and SB : CB(B, K) → CB(B, F ) are onto, there exist RA ∈ C B(A, K) and RB ∈ CB(B, K) such
RA×T B = RA ◦ PA + RB ◦ PB. Clearly RA×T B ∈ CB(A ×T B, K) and T ′ ◦ RA×T B = λ. Hence A ×T B
is projective as left C-module.
Conversely, let A be a Banach left C-module and F, K ∈ C-mod and let S ∈ CB(K, F ) such that S be
admissible and epimorphism. We show that the induced map SA : CB(A, K) → CB(A, F ) is onto. We
have A ×T B ∈ C-mod with the following action
c · (a, b) = (c · a + c · T (b), 0).
Since A ×T B is projective, the induced map SA×T B : CB(A×T , K) → CB(A ×T B, F ) is onto. Let
λ ∈ CB(A, F ) and f ∈ F such that λ(a) = f for a ∈ A. We define eλ : A×T B → F by eλ(a, b) = λ(a+T (b)).
Hence we have eλ ∈ CB(A ×T B, F ). Since SA×T B : CB(A ×T B, K) → CB(A ×T B, F ) is onto, there
exists RA×T B ∈ B(A ×T B, K) such that S ◦ RA×T B = eλ. We define RA : A → K by RA = RA×T B ◦ qA,
where qA : A → A ×T B is defined by qA(a) = (a, 0). Clearly RA ∈ CB(A, K) and S ◦ RA = λ. Hence A
is projective as left C-module.
For the proof of projectivity of B, let B be a Banach left C-module and F, K ∈ C-mod and let S ∈
B(K, F ) such that S is admissible and epimorphism. We show that the induced map SB : CB(B, K) →
CB(B, F ) is onto. We have A ×T B ∈ C-mod with the following action
c · (a, b) = (−T (c · b), c · b).
Since A×T B is projective as left C-module, the induced map SA×T B : CB(A×T B, K) → CB(A×T B, F )
is onto. Let λ ∈ CB(B, F ). We define eλ : A ×T B → F by eλ(a, b) = λ(b) for (a, b) ∈ A ×T B.
Hence eλ ∈ C B(A ×T B, F ). Since SA×T B : CB(A ×T B, K) → CB(A ×T B, F ) is onto, there exists
RA×T B ∈ CB(A ×T B, K) such that S ◦ RA×T B(a, b) = eλ(a, b) = λ(b). We define RB : B → K
by RB = RA×T B ◦ qB, where qB : B → A ×T B is defined by qB(b) = (0, b) for b ∈ B. Clearly
RB ∈ CB(B, K) and S ◦ RB = λ. Hence B is projective as left C-module.
(cid:3)
3. Hochschild cohomology for the Banach algebra A ×T B
The concept of Hochschild cohomology for Banach algebras has been studied by Kamowitz[9], Johnson[7,
8] and others. Recall that let A be a Banach algebra and let X be a Banach A-bimodule. We denote
the space of bounded n-linear maps from A into X by C n(A, X). For T ∈ C n(A, X) we define the map
6
N. RAZI AND A. POURABBAS
δn : C n(A, X) → C n+1(A, X) by
(δnT )(a1, ..., an+1) =a1 · T (a2, ..., an+1)
+
nX
i=1
(−1)iT (a1, ..., aiai+1, ..., an+1)
=(−1)n+1T (a1, ..., an) · an+1.
T is called an n-cocycle if δnT = 0 and it is called n-coboundary if there exists S ∈ C n−1(A, X) such
that T = δn−1S. We denote the linear space of all n-cocycles by Z n(A, X) and the linear space of all n-
coboundaries by Bn(A, X). Clearly Z n(A, X) includes Bn(A, X). We also recall that the n-th Hochschild
cohomology group Hn(A, X) is defined by the following quotient,
Hn(A, X) =
Z n(A, X)
Bn(A, X)
,
for more details, see [8]. We remark that a left (right) Banach A-module X is called left (right) essential
if the linear span of A · X = {a · x : a ∈ A, x ∈ X} (X · A = {x · a : x ∈ X, a ∈ A}) is dense in X. A
Banach A-module X is called essential, if it is left and right essential.
Let A and B be Banach algebras and T : B → A be an algebra homomorphism with kT k ≤ 1. Let E
be a Banach A-bimodule. Then E is also a Banach B-bimodule and a Banach A ×T B-bimodule with
the following actions, respectively
b · x = T (b) · x and x · b = x · T (b),
where b ∈ B, x ∈ E and
(a, b) · x = T (b) · x and x · (a, b) = x · T (b),
where (a, b) ∈ A ×T B, x ∈ E.
Lemma 3.1. Let E be an essential Banach A×T B-bimodule. Then E is an essential Banach A-bimodule
and an essential Banach B-bimodule.
Theorem 3.2. Let A and B be Banach algebras with bounded approximate identity and let T : B → A
be an algebra homomorphism with kT k ≤ 1. Let E be an essential A ×T B-bimodule. Then
where ≃ denotes the vector space isomorphism.
H1(A ×T B, E ∗) ≃ H1(A, E ∗) × H1(B, E ∗),
Proof. By [2, Theorem 2.9.53] we have H1(A ×T B, E ∗) ≃ H1(M (A ×T B), E ∗), where M(A ×T B)
denotes the double centralizer algebra of A ×T B. But from [12, Theorem 4.3] we have
where ∼= denotes the algebra isomorphism. this implies that
M(A ×T B) ∼= M(A) × M(B),
H1(M (A ×T B), E ∗) ≃ H1(M(A) × M(B), E ∗),
where M(A) and M(B) denote the double centralizer algebra of A and B, respectively.
Hence we have
H1(A ×T B, E ∗) ≃ H1(M(A) × M(B), E ∗),
COHOMOLOGICAL CHARACTERIZATION OF T -LAU PRODUCT ALGEBRAS
7
Using [6, Theorem 4] we obtain H1(M(A) × M(B), E ∗) ≃ H1(M(A), E ∗) × H1(M(B), E ∗), thus we
have
H1(A ×T B, E ∗) ≃ H1(M(A), E ∗) × H1(M(B), E ∗).
Since E is an essential Banach A-bimodule and an essential Banach B-bimodule, we have
and
H1(M(A), E ∗) ≃ H1(A, E ∗)
H1(M(B), E ∗) ≃ H1(B, E ∗).
This completes the proof.
(cid:3)
We can extend the previous theorem for the n-th Hochschild cohomology for the Banach algebra A×T B.
Corollary 3.3. Let A and B be Banach algebras with bounded approximate identity and let T : B → A
be an algebra homomorphism with kT k ≤ 1. Let E be an essential A ×T B-bimodule. Then
Hn(A ×T B, E ∗) ≃ Hn(A, E ∗) × Hn(B, E ∗)
for every n ≥ 1.
Proof. By [12, Lemma 3.1] A and B have bounded approximate identities if and only if A ×T B has a
bounded approximate identity. Using [2, Theorem 2.9.54] one can show that if A ×T B has a bounded
approximate identity, then for every essential A×T B-bimodule E, we have Hn(A×T B, E ∗) ≃ Hn(M (A×T
B), E ∗). In [12, Theorem 4.3] the authors showed that M (A ×T B) ∼= M (A) × M (B) . On the other
hand Hochschild [6, Theorem 4] showed that Hn(M (A) × M (B), E ∗) ≃ Hn(M (A), E ∗) × Hn(M (B), E ∗).
Hence we have Hn(A ×T B, E ∗) ≃ Hn(A, E ∗) × Hn(B, E ∗) where n ≥ 1.
(cid:3)
Note that Bhatt and Dabshi in [1] showed that A×T B is amenable if and only if A and B are amenable.
In the essential case this is an immediate corollary of Theorem 3.2.
References
[1] S. J. Bhatt and P. A. Dabshi, Arens regularity and amenability of Lau product of Banach algebras defined by a Banach
algebra morphism, Bull. Aust. Math. Soc. 87 (2013), 195-206.
[2] H. G. Dales, Banach algebras and automatic continuity, London Mathematical Society Monographs 24, Clarendon
Press, Oxford, 2000.
[3] H. G. Dales, M. E. Polyakov, Homological properties of modules over group algebras, Proc. London Math. Soc. (3) 89
(2004), 390-426.
[4] A. Ya. Helemskii, A certain class of flat Banach modules and its applications, Vestnik. Moskov. Univ. Ser. Mat. Mekh.
27 (1972), 29-36
[5] A. Ya. Helemskii, The homology of Banach and topological algebras, Kluwer Academic Publishers Group, Dordrecht,
(1989).
[6] G. Hochschild, On the cohomology theory for associative algebras, Ann. of Math. (2) 47 (1946), 568-579.
[7] B. E. Johnson, The Wedderburn decomposition of Banach algebras with finite dimensional radical, Amer. J. Math. 90
(1968), 866-876.
[8] B. E. Johnson, Cohomology in Banach algebras, Mem. Amer. Math. Soc. 127 (1972).
[9] H. Kamowitz, Cohomology groups of commutative Banach algebras, Trans. Amer.Math. Soc. 102 (1962), 352-372.
8
N. RAZI AND A. POURABBAS
[10] A. T. Lau, Analysis on a class of Banach algebras with applications to harmonic analysis on locally compact groups
and semigroups, Fund. Math. 118 (3) (1983), 161-175.
[11] P. Ramsden, Homological properties of modules over semigroup algebras, J. Funct. Anal. 258 (2010) 3988-4009.
[12] N. Razi, A. Pourabbas, Some homological properties of T -Lau product algebras, Period. Math. Hungar. to appear.
[13] M. Sangani Monfared, On certain products of Banach algebras with applications to harmonic analysis, Studia Math.
178 (3) (2007), 277-294.
[14] M. C. White, Injective modules for uniform algebras, Proc. London Math. Soc. (3) 73 (1996), 155-184
Faculty of Mathematics and Computer Science, Amirkabir University of Technology, 424 Hafez Avenue,
Tehran 15914, Iran
E-mail address: [email protected]
E-mail address: [email protected]
|
1412.6165 | 3 | 1412 | 2015-09-22T10:57:25 | The convenient setting for ultradifferentiable mappings of Beurling- and Roumiue-type defined by a weight matrix | [
"math.FA"
] | We prove in a uniform way that all ultradifferentiable function classes of Roumieu- and of Beurling-type defined in terms of a weight matrix admit a convenient setting if the matrix satisfies some mild regularity conditions. We prove that these categories are cartesian closed and as special cases one obtains the classes defined by a weight sequence and by a weight function. | math.FA | math |
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE MAPPINGS
OF BEURLING- AND ROUMIEU-TYPE DEFINED BY A WEIGHT MATRIX
GERHARD SCHINDL
Abstract. We prove in a uniform way that all ultradifferentiable function classes E{M} of
Roumieu-type and E(M) of Beurling-type defined in terms of a weight matrix M admit a con-
venient setting if M satisfies some mild regularity conditions. For C denoting either E{M} or
E(M) the category C is cartesian closed, i.e. C(E × F, G) ∼= C(E, C(F, G)) for E, F, G convenient
vector spaces. As special cases one obtains the classes E{M } and E(M ) respectively E{ω} and E(ω)
defined by a weight sequence M respectively a weight function ω.
1. Introduction
Spaces of ultradifferentiable functions are subclasses of smooth functions with certain growth con-
ditions on all their derivatives.
In the literature two different approaches are considered, either
using a weight sequence M = (Mk)k or using a weight function ω. For compact K the set
(cid:26) f (k)(x)
hkMk
: x ∈ K, k ∈ N(cid:27)
respectively
(cid:26)
f (k)(x)
exp(1/lϕ∗ω(lk))
: x ∈ K, k ∈ N(cid:27)
should be bounded, where the positive real number h respectively l is subject to either a universal
or an existential quantifier and ϕ∗ω denotes the Young-conjugate of ϕω = ω ◦ exp. In the case of a
universal quantifier we call the class of Beurling-type, denoted by E(M) or E(ω), in the case of an
existential quantifier we call the class of Roumieu-type, denoted by E{M} or E{ω}. We write E[⋆] if
either E{⋆} or E(⋆) is considered.
That a class of mappings C admits a convenient setting means that one can extend the class to
admissible infinite dimensional vector spaces E, F, G such that C(E, F ) is again admissible and the
spaces C(E × F, G) and C(E,C(F, G)) are canonically C-diffeomorphic. This important property is
called the exponential law.
We recall now some facts, see [4] or the appendix in [5] for a short overview. The class E of all
smooth functions admits a convenient setting and for this approach one can test smoothness along
E-curves. The class Cω of all real-analytic mappings also admits a convenient setting. A mapping
is Cω if and only if it is E and in addition it is weakly Cω along (weakly) Cω-curves, i.e. curves
whose compositions with any bounded linear functional are Cω. It actually suffices to test along
affine lines.
In [5], [6] and finally in [7] A. Kriegl, P.W. Michor and A. Rainer were able to develop the conve-
nient setting for all reasonable classes E(M) and E{M}. In the first step in [5] they introduced the
Date: July 3, 2018.
2010 Mathematics Subject Classification. 46E10, 46T05, 46T10.
Key words and phrases. Ultradifferentiable functions, convenient setting.
GS was supported by FWF-Project P 23028-N13 and FWF-Project P 26735-N25.
1
2
G. SCHINDL
convenient setting for E{M} by testing with E{M}-curves for non-quasianalytic, strongly log-convex
weight sequences M of moderate growth. A function is E{M} if and only if it is E{M} along all
E{M}-curves. It was shown that moderate growth is really necessary for the exponential law and
non-quasianalyticity is needed for the existence of E{M}-partitions of unity.
Then, in [6], they succeeded to introduce the convenient setting for some quasianalytic classes E{M}.
In this case M has to satisfy again strong log-convexity, moderate growth and be such that E{M}
can be represented as the intersection of all larger non-quasianalytic classes E{L} with strongly
log-convex L. A mapping is E{M} if and only if it is E{L} along each E{L}-curve for each L ≥ M
which is strongly log-convex and non-quasianalytic. A family of explicit examples E{M} satisfying
the requested assumptions was constructed, but the approach does not cover the real analytic case
Cω and thus was not completely satisfactory.
Finally, in [7], it was shown that all classes E{M} and E(M) such that M is strongly log-convex and
has moderate growth admit a convenient setting, no matter if M is quasianalytic or not. Instead of
testing along curves the mappings are tested along Banach plots, i.e. mappings of the respective
weak class defined in open subsets of Banach spaces. A smooth mapping between convenient vector
spaces is E[M] if it maps E[M]-Banach-plots to E[M]-Banach-plots.
The aim of this work is to generalize the results of [7] to classes E[M] defined by (one-parameter)
weight matrices M := {M x : x ∈ R>0}. In [9] the classes E[M] and E[ω] were identified as particular
cases of E[M]. So using this new approach one is able to transfer results from one setting into the
other one. Moreover one is able to prove results for E[M] and E[ω] simultaneously and no longer
two separate proofs are necessary. We have also shown that there are classes E[M] which cannot be
described by a single M or ω, e.g. the class defined by the Gevrey-matrix G := {(p!s+1)p∈N : s > 0}.
To transfer the proofs of [7] we will assume for M among mild basic properties the so-called
generalized Faà-di-Bruno-property (M[FdB]) and the moderate growth condition (M[mg]).
After introducing the basic notation and definitions we recall the setting of Whitney jets between
Banach spaces. We introduce classes of ultradifferentiable functions defined by weight matrices, first
between Banach spaces and then between convenient vector spaces. This will be done in section 3.
In section 4 we are going to prove the most important and new tools in this work. We will develop
projective descriptions for the classes E[M] in order to get rid of both existence quantifiers in the
Roumieu-case (if M = {M} only one occurs). For this we have to use diagonal techniques and
to introduce several families of sequences of positive real numbers to generalize the results of [7].
These projective representations are needed in section 5 for the proof of Theorem 5.9 to show that
E[M] is a category and for cartesian closedness Theorem 6.2 in section 6.
Finally in section 7 we summarize some special cases. In 7.3 we revisit weight matrices as defined
by Beaugendre in [1] and Schmets and Valdivia in [13]. Put MΦ := {(p!mΦ
ap)p∈N : a > 0}, where
t = +∞, Φ(0) = 0. In the
Φ : [0, +∞) → R is a convex and increasing function with limt→∞
literature only the Beurling-type-class was studied. We will see that the results in this work can
also be applied to such classes.
Note that if M = {M} then the Faà-di-Bruno-property for M is sufficient to show closedness under
composition and is sufficient for the proofs in this work. But it is really weaker than strong log-
convexity as assumed always in the previous papers and proofs of Kriegl, Michor, Rainer, see [9,
3.3.] for an explicit (counter)-example. So our results are slightly more general than those of [7]
even in the single weight sequence case. In Lemma 6.6 we will show that (M{mg}) is necessary for
Φ(t)
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
3
cartesian closedness of E{M} and in Example 6.5 we will point out that there exist weight matrices
M such that no M x ∈ M has moderate growth but nevertheless (M{mg}) is valid. In particular
this holds if the matrix is associated to a weight function ω and such that E[ω] = E[M] does not
hold, see [2] and [9].
This paper contains some of the main results of the authors PhD-Thesis, see [12]. The author
thanks his advisor A. Kriegl, P.W. Michor and A. Rainer for the supervision and their helpful
ideas.
1 ··· ∂αn
f (x + tv)t=0.
dt(cid:1)k
vf (x) :=(cid:0) d
1.1. Basic notation. We denote by C the class of all continuous, by E the class of smooth functions
and Cω is the class of all real analytic functions. We will write N>0 = {1, 2, . . .}, N = N>0 ∪ {0}
and put R>0 := {x ∈ R : x > 0}. For α = (α1, . . . , αn) ∈ Nn we use the usual multi-index notation,
write α! := α1! . . . αn!, α := α1 + ··· + αn and for x = (x1, . . . , xn) ∈ Rn we set xα = xα1
1 ··· xαn
n .
We also put ∂α = ∂α1
n and denote by f (k) the k-th order Fréchet derivative of f . Iterated
uni-directional derivatives are defined by dk
Let E1, . . . , Ek and F be topological vector spaces, then L(E1, . . . , Ek, F ) is the space of all bounded
If E = Ei for i = 1, . . . , k, then we write Lk(E, F ).
k-linear mappings E1 × ··· × Ek → F .
Lk
→ F , so f (k) :
sym(E, F ) is the space of all symmetric k-linear bounded mappings E × ··· × E
}
U → Lk
the space of
all bounded linear functionals. If B ⊆ E is closed absolutely convex bounded, then EB denotes the
space generated by B with the Minkowski-functional k · kB.
Let E be a locally convex vector space, then the c∞-topology on E is the final topology w.r.t. all
smooth curves c : R → E. E is called convenient if E is c∞-complete which is equivalent for E to
be Mackey-complete and for EB to be a Banach space for every bounded absolutely convex subset
B of E. We refer to [4] or the appendix in [5] for more details and proofs.
Convention: Let ⋆ ∈ {M, ω,M}, then write E[⋆] if either E{⋆} or E(⋆) is considered, but not mixing
the cases if statements involve more than one E[⋆] symbol. The same notation will be used for the
conditions, so write (M[⋆]) for either (M{⋆}) or (M(⋆)).
sym(E, F ). E∗ denotes the space of all continuous linear functionals on E, E
k−times
′
{z
2. Basic definitions
2.1. Weight sequences and classes of ultradifferentiable functions E[M]. A weight sequence
is an arbitrary sequence of positive real numbers M = (Mk)k ∈ RN
>0. We introduce also m = (mk)k
defined by mk := Mk
(1) M is log-convex if
, µ0 := 1. M is called normalized if 1 = M0 ≤ M1 holds.
k! and µk := Mk
Mk−1
M is log-convex if and only if (µk)k is increasing. If M is log-convex and M0 = 1, then
(lc) :⇔ ∀ j ∈ N : M 2
j ≤ Mj−1Mj+1.
(alg) :⇔ ∃ C ≥ 1 ∀ j, k ∈ N : MjMk ≤ Cj+kMj+k
holds with C = 1 and the mapping j 7→ (Mj)1/j is increasing, see e.g. [11, Lemma 2.0.4, Lemma
2.0.6].
M is called strongly log-convex if
(slc) :⇔ ∀ j ∈ N : m2
j ≤ mj−1mj+1.
4
G. SCHINDL
This condition implies (lc) and was a basic assumptions for M in [5], [6] and [7]. It guarantees
all stability properties in [10, Theorems 5,6] for the case M = {M}, see also [9, Theorem 3.2.].
Related to this is the weaker condition
which is called the Faà-di-Bruno-property, see [9, 3.3.]. For m◦ = (m◦k)k we have put
(FdB) :⇔ ∃ D ≥ 1 ∀ k ∈ N : m◦k ≤ Dkmk,
m◦k := max{mjmα1 ··· mαj : αi ∈ N>0,
Strongly log-convexity is also related to
j
Xi=1
αi = k},
m◦0 := 1.
(rai) :⇔ ∃ C ≥ 1 ∀ 1 ≤ j ≤ k : (mj)1/j ≤ C(mk)1/k,
see [9] and [10].
(2) M has moderate growth if
This condition implies derivation closedness:
(mg) :⇔ ∃ C ≥ 1 ∀ j, k ∈ N : Mj+k ≤ Cj+kMjMk.
(dc) :⇔ ∃ C ≥ 1 ∀ j ∈ N : Mj+1 ≤ Cj+1Mj.
In both conditions one can replace the sequence M by m.
(3) For M = (Mp)p and N = (Np)p we write M ≤ N if and only if Mp ≤ Np for all p ∈ N. Moreover
we define
M (cid:22) N :⇔ ∃ C1, C2 ≥ 1 ∀ j ∈ N : Mj ≤ C2Cj
1Nj ⇐⇒ sup
p∈N>0(cid:18) Mp
Np(cid:19)1/p
< +∞
and we call the sequences equivalent if
We will write
M ≈ N :⇔ M(cid:22)N and N(cid:22)M.
M ⊳ N :⇔ ∀ h > 0 ∃ Ch ≥ 1 ∀ j ∈ N : Mj ≤ ChhjNj ⇐⇒ lim
= 0.
p→∞(cid:18) Mp
Np(cid:19)1/p
(Mk)1/k = +∞}.
For convenience we introduce the following set:
LC := {M ∈ RN
>0 : M is normalized, log-convex,
lim
k→∞
Let r, s ∈ N>0 and U ⊆ Rr be non-empty open. We introduce the ultradifferentiable class of
Roumieu-type by
E{M}(U, Rs) := {f ∈ E(U, Rs) : ∀ K ⊆ U compact ∃ h > 0 : kfkM,K,h < +∞},
and the class of Beurling-type by
E(M)(U, Rs) := {f ∈ E(U, Rs) : ∀ K ⊆ U compact ∀ h > 0 : kfkM,K,h < +∞},
where we have put
(2.1)
kfkM,K,h := sup
k∈N,x∈K
kf (k)(x)kLk(Rr,Rs)
hkMk
.
For compact sets K with smooth boundary
EM,h(K, Rs) := {f ∈ E(K, Rs) : kfkM,K,h < +∞}
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
5
is a Banach space and we have the topological vector space representations
(2.2)
and
(2.3)
E{M}(U, Rs) := lim←−K⊆U
lim−→h>0 EM,h(K, Rs) = lim←−K⊆U E{M}(K, Rs)
E(M)(U, Rs) := lim←−K⊆U
lim←−h>0 EM,h(K, Rs) = lim←−K⊆U E(M)(K, Rs).
We recall some facts for log-convex M :
(i) Put E global
{M}
functions
(U, Rs) := {f ∈ E(U, Rs) : ∃ h > 0 kfkM,U,h < +∞}. There exist characteristic
(chf) :⇔ ∃ θM ∈ E global
{M}
(R, C) with
and θM ∈ E global
{M}
∀ j ∈ N : θ(j)
M (0) = (√−1)jsj,
(2.4)
(R, R) : ∀ j ∈ N :(cid:12)(cid:12)(cid:12)θ(j)
M (0)(cid:12)(cid:12)(cid:12) ≥ Mj,
sj :=
Mk(2µk)j−k ≥ Mj,
∞Xk=0
θ(j)
hence(cid:12)(cid:12)(cid:12)
M (0)(cid:12)(cid:12)(cid:12) ≥ Mj for all j ∈ N, see [9, Lemma 2.9.] and [14, Theorem 1]. Note that the
(M) (R, R) cannot contain such θM , see [11, Proposition 3.1.2.].
(ii) If N is arbitrary, then M(cid:22)N ⇐⇒ E{M} ⊆ E{N} and M ⊳N ⇐⇒ E{M} ⊆ E(N ). If M ∈ LC,
(iii) Both classes E{M} and E(M) are closed under pointwise multiplication, see e.g. [11, Propo-
Beurling-class E global
then M(cid:22)N ⇐⇒ E[M] ⊆ E[N ].
sition 2.0.8].
2.2. Classes of ultra-differentiable functions defined by one parameter weight matrices
and basic definitions.
Definition 2.3. Let (Λ,≤) be a partially ordered set which is both up- and downward directed,
Λ = R>0 will be the most important example. A weight matrix M associated to Λ is a family of
weight sequences M := {M x ∈ RN
>0 : x ∈ Λ} such that
(M) :⇔ ∀ x ∈ Λ : M x is normalized, increasing, M x ≤ M y for x ≤ y.
We call M standard log-convex, if
Also mx
k := M x
k
k! and µx
k := M x
k
M x
k−1
, µx
0 := 1, will be used.
(Msc) :⇔ (M) and ∀ x ∈ Λ : M x ∈ LC.
We introduce ultradifferentiable classes of Roumieu- and Beurling-type defined by M as follows
(see also [9, 4.2.]):
Let r, s ∈ N>0, let U ⊆ Rr be non-empty and open. For all K ⊆ U compact we put
(2.5)
E{M x}(K, Rs)
E{M x}(K, Rs)
and
(2.6)
E{M}(K, Rs) := [x∈Λ
E(M)(K, Rs) := \x∈Λ
E{M}(U, Rs) := \K⊆U [x∈Λ
E(M)(U, Rs) := \x∈Λ
E(M x)(K, Rs)
E(M x)(U, Rs).
6
G. SCHINDL
For a compact set K ⊆ Rr (with smooth boundary) we have
and so for U ⊆ Rr non-empty open
(2.7)
E{M}(K, Rs) := lim−→x∈Λ
lim−→h>0 EM x,h(K, Rs),
E{M}(U, Rs) := lim←−K⊆U
lim−→x∈Λ
lim−→h>0 EM x,h(K, Rs),
and for the Beurling-case we get
(2.8)
E(M)(U, Rs) := lim←−K⊆U
lim←−x∈Λ
lim←−h>0 EM x,h(K, Rs).
lim−→h>0 EM x,h(K, Rs) = lim−→n∈N>0
Instead of compact sets K with smooth boundary one can also consider open K ⊆ U with K
compact in U , or one can work with Whitney jets on compact K.
If Λ = R>0 we can assume that all occurring limits are countable and so E(M)(U, Rs) is a Fréchet
EM n,n(K, Rs) is a Silva space, i.e. a countable
space. Moreover lim−→x∈Λ
inductive limit of Banach spaces with compact connecting mappings. For more details concerning
the locally convex topology on these spaces we refer to [9, 4.2.-4.4.].
2.4. Conditions for a weight matrix M = {M x : x ∈ Λ}. We are going to introduce now some
conditions on M which will be needed frequently, see also [9, 4.1.].
Roumieu-type-conditions
j+1 ≤ Cj+1M y
(M{dc}) ∀ x ∈ Λ ∃ C > 0 ∃ y ∈ Λ ∀ j ∈ N : M x
j M y2
j+k ≤ Cj+kM y1
(M{mg}) ∀ x ∈ Λ ∃ C > 0 ∃ y1, y2 ∈ Λ ∀ j, k ∈ N : M x
k ≤ Cj+kM y
j M x2
(M{alg}) ∀ x1, x2 ∈ Λ ∃ C > 0 ∃ y ∈ Λ ∀ j, k ∈ N : M x1
k ≤ DM y
(M{L}) ∀ C > 0 ∀ x ∈ Λ ∃ D > 0 ∃ y ∈ Λ ∀ k ∈ N : CkM x
j+k
k
k
j
k
M x
= +∞
k(cid:17)1/k
q )1/q ≤ H(my
supk∈N>0(cid:16) M y
(M{strict}) ∀ x ∈ Λ ∃ y ∈ Λ :
(M{FdB}) ∀ x ∈ Λ ∃ y ∈ Λ : (mx)◦(cid:22)my
(M{rai}) ∀ x ∈ Λ ∃ y ∈ Λ ∃ H > 0 : (mx
Beurling-type-conditions
(M(dc)) ∀ x ∈ Λ ∃ C > 0 ∃ y ∈ Λ ∀ j ∈ N : M y
(M(mg)) ∀ x1, x2 ∈ Λ ∃ C > 0 ∃ y ∈ Λ ∀ j, k ∈ N : M y
j+k ≤ Cj+kM x1
j M y2
(M(alg)) ∀ x ∈ Λ ∃ C > 0 ∃ y1, y2 ∈ Λ ∀ j, k ∈ N : M y1
k ≤ Cj+kM x
(M(L)) ∀ C > 0 ∀ x ∈ Λ ∃ D > 0 ∃ y ∈ Λ ∀ k ∈ N : CkM y
k ≤ DM x
p)1/p, 1 ≤ q ≤ p
j+1 ≤ Cj+1M x
k
j
j M x2
k
j+k
supk∈N>0(cid:16) M x
(M(strict)) ∀ x ∈ Λ ∃ y ∈ Λ :
(M(FdB)) ∀ x ∈ Λ ∃ y ∈ Λ : (my)◦(cid:22)mx
(M(rai)) ∀ x ∈ Λ ∃ y ∈ Λ ∃ H > 0 : (my
2.5. Inclusion relations of weight matrices. Let two matrices M = {M x : x ∈ Λ} and N =
{N x : x ∈ Λ′} be given, then we write
k(cid:17)1/k
q )1/q ≤ H(mx
p)1/p, 1 ≤ q ≤ p
= +∞
k
M y
and
M{(cid:22)}N :⇔ ∀ x ∈ Λ ∃ y ∈ Λ′ : M x(cid:22)N y
M((cid:22))N :⇔ ∀ y ∈ Λ′ ∃ x ∈ Λ : M x(cid:22)N y,
M{≈}N :⇔ M{(cid:22)}N and N{(cid:22)}M
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
7
respectively
By definition M[(cid:22)]N implies E[M] ⊆ E[N ]. Moreover write
M(≈)N :⇔ M((cid:22))N and N ((cid:22))M.
M ⊳ N :⇔ ∀ x ∈ Λ ∀ y ∈ Λ′ : M x⊳N y,
so M ⊳ N implies E{M} ⊆ E(N ). In [9, Proposition 4.6.] the above relations are characterized for
(Msc)-matrices with Λ = Λ′ = R>0. In this context we introduce
(M{Cω}) ∃ x ∈ Λ :
(MH)
∀ x ∈ Λ :
(M(Cω)) ∀ x ∈ Λ :
If (M{Cω}) holds, then Cω ⊆ E{M}, if (M(Cω )) then Cω ⊆ E(M). Finally if (MH), then the
restrictions of entire functions are contained in E(M), see [9, Proposition 4.6.].
Conventions:
lim inf k→∞(mx
lim inf k→∞(mx
limk→∞(mx
k)1/k > 0,
k)1/k > 0,
k)1/k = +∞.
(i) If Λ = R>0 or Λ = N>0, then these sets are always regarded with its natural order ≤.
(ii) We will call M constant if M = {M} or more generally if M x≈M y for all x, y ∈ Λ and
which violates both (M{strict}) and (M(strict)). Otherwise it will be called non-constant.
2.6. Weight functions and classes of ultradifferentiable functions E[ω]. A function ω :
[0,∞) → [0,∞) (sometimes ω is extended to C by ω(x) := ω(x)) is called a weight function if
(ω0) ω is continuous, on [0,∞) increasing, ω(x) = 0 for x ∈ [0, 1] (w.l.o.g.) and limx→∞ ω(x) =
+∞.
Moreover we consider the following conditions:
(ω1) ω(2t) = O(ω(t)) as t → +∞.
(ω2) ω(t) = O(t) as t → ∞.
(ω3) log(t) = o(ω(t)) as t → +∞ (⇔ limt→+∞
(ω4) ϕω : t 7→ ω(et) is a convex function on R.
(ω5) ω(t) = o(t) as t → +∞.
(ω6) ∃ H ≥ 1 ∀ t ≥ 0 : 2ω(t) ≤ ω(Ht) + H.
(ω1′ ) ∃ D > 0 : ∃ t0 > 0 : ∀ λ ≥ 1 : ∀ t ≥ t0 : ω(λt) ≤ Dλω(t).
An interesting example is ωs(t) := max{0, log(t)s}, s > 1, which satisfies all listed properties except
(ω6). For convenience we define the sets
ϕω(t) = 0).
t
W0 := {ω : [0,∞) → [0,∞) : ω has (ω0), (ω3), (ω4)},
For ω ∈ W0 we define the Legendre-Fenchel-Young-conjugate ϕ∗ω by
W := {ω ∈ W0 : ω has (ω1)}.
ϕ∗ω(x) := sup{xy − ϕω(y) : y ≥ 0},
It is a convex increasing function, ϕ∗ω(0) = 0, ϕ∗∗ω = ϕω, limx→∞
and x 7→ ϕ∗
For σ, τ ∈ W we write
are increasing on [0, +∞), see e.g. [3, Remark 1.3., Lemma 1.5.].
ω(x)
x
x
σ (cid:22) τ :⇔ τ (t) = O(σ(t)), as t → +∞
and call them equivalent if
x ≥ 0.
ω(x) = 0 and finally x 7→ ϕω(x)
ϕ∗
x
σ ∼ τ :⇔ σ(cid:22)τ and τ(cid:22)σ.
8
G. SCHINDL
Let r, s ∈ N>0, U ⊆ Rr be a non-empty open set and ω ∈ W0. The Roumieu-type space is defined
by
E{ω}(U, Rs) := {f ∈ E(U, Rs) : ∀ K ⊆ U compact ∃ l > 0 : kfkω,K,l < +∞}
and the Beurling-type space by
E(ω)(U, Rs) := {f ∈ E(U, Rs) : ∀ K ⊆ U compact ∀ l > 0 : kfkω,K,l < +∞},
where we have put
(2.9)
kfkω,K,l := sup
k∈N,x∈K
kf (k)(x)kLk(Rr ,Rs)
l ϕ∗ω(lk))
exp( 1
and f (k)(x) denotes the k-th order Fréchet derivative at x. For compact sets K with smooth
boundary
is a Banach space and we have the topological vector space representations
Eω,l(K, Rs) := {f ∈ E(K, Rs) : kfkω,K,l < +∞}
(2.10)
and
(2.11)
E{ω}(U, Rs) := lim←−K⊆U
lim−→l>0 Eω,l(K, Rs) = lim←−K⊆U E{ω}(K, Rs)
E(ω)(U, Rs) := lim←−K⊆U
lim←−l>0 Eω,l(K, Rs) = lim←−K⊆U E(ω)(K, Rs).
A new idea introduced in [9, Chapter 5] was the following:
j)j∈N : l > 0} by
(i) To each ω ∈ W we can associate a (Msc) weight matrix Ω = {Ωl = (Ωl
Ωl
j := exp(cid:0) 1
(ii) Ω has always (M{mg}) and (M(mg)), (M{L}) and (M(L)).
l ϕ∗ω(lj)(cid:1).
If ω is sub-additive, then
(M{FdB}) and (M(FdB)) hold, see [9, Lemma 6.1.]. Equivalent weight functions ω yield
equivalent weight matrices w.r.t. both (≈) and {≈}.
(iii) E[Ω] = E[ω] holds as locally convex vector spaces, so defining classes of ultradifferentiable
functions by weight matrices as in (2.5) and (2.6) is a common generalization of defining
them by using a single weight sequence M , i.e. a constant weight matrix, or a weight
function ω ∈ W. But one is also able to describe classes which cannot be described neither
by a weight function nor by a weight sequence, e.g. the class defined by the Gevrey-matrix
G := {(p!s+1)p∈N : s > 0}, see [9, 5.19.].
3. Basic definitions for the convenient setting
3.1. Whitney jets on Banach spaces. We recall the notation of [7, Chapter 3]. Let E, F be
Banach spaces, K ⊆ E compact and U ⊆ E open. Let f ∈ E(U, F ), then we introduce the jet
mapping j∞ : E(U, F ) → J∞(U, F ) :=Qk∈N C(U, Lk
sym(E, F )) defined by f 7→ j∞(f ) = (f (k))k∈N.
For an arbitrary subset X ⊆ E and an infinite jet f = (f k)k∈N we introduce the Taylor polynomial
y f )k : X → Lk
(T n
(T n
sym(E, F ) of order n at the point y as follows:
y f )k(x)(v1, . . . , vk) :=
f j+k(y)(x − y, . . . , x − y, v1, . . . , vk).
1
j!
n
y f )k(x) = (T n
x f )k(x) − (T n
y f )k(x)
The remainder is given by
Xj=0
y f )k(x) := f k(x) − (T n
(Rn
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
9
and so (Rn
y f )k(x) ∈ Lk
sym(E, F ). We put now
and
kfkk := supnkf k(x)kLk
y f )k(x)kLk
kx − ykn+1
kfkn,k := sup((n + 1)!k(Rn
sym(E,F ) : x ∈ Ko
sym(E,F )
: x, y ∈ K, x 6= y) .
We supply E(U, F ) with the seminorms f 7→ kj∞(f )Kkk, where K ⊆ U is a compact set and k ∈ N.
If K ⊆ E is compact and convex, then we introduce the space E(E ⊇ K, F ) of Whitney-jets on K
by
E(E ⊇ K, F ) :=(f = (f k)k∈N ∈ Yk∈N
C(K, Lk
sym(E, F )) : kfkn,k < +∞ ∀ n, k ∈ N)
and we supply these spaces with both seminorms kfkk and kfkn,k for k, n ∈ N. Finally recall [7,
Lemma 3.1.]:
Lemma 3.2. Let E and F be Banach spaces and K ⊆ E be a compact convex subset. Then
E(E ⊇ K, F ) is a Fréchet space.
3.3. Classes of ultra-differentiable mappings defined by a weight matrix. Let M := {M x :
x ∈ Λ} be (M), E and F be Banach spaces and K ⊆ E a compact subset. Then, as in [7, 4.1.], for
x ∈ Λ and h > 0 we define
where
C(K, Lj
sym(E, F )) : kfkJ
(f j)j ∈ Yj∈N
EM x,h(E ⊇ K, F ) :=
M x,h := max(cid:26)sup(cid:26) kfkk
kfkJ
EM x,K,h(U, F ) :=(cid:8)f ∈ E(U, F ) : j∞(f )(cid:12)(cid:12)K ∈ EM x,h(E ⊇ K, F )(cid:9) ,
M x,h < +∞
: k, n ∈ N(cid:27)(cid:27) .
: k ∈ N(cid:27) , sup(cid:26)
kfkn,k
hn+k−1M x
hkM x
k
n+k+1
,
For open U ⊆ E and compact K ⊆ U we introduce the space
J
M x,h
with semi-norm f 7→(cid:13)(cid:13)j∞(f )(cid:12)(cid:12)K(cid:13)(cid:13)
. It is not Hausdorff and for infinite dimensional E its Haus-
dorff quotient will not always be complete. Note that if K is assumed to be convex, then we can
take on EM x,K,h(U, F ) also the semi-norm
f 7→ sup(kf (n)(a)kLn
: a ∈ K, n ∈ N) =: kfkJ
M x,K,h.
hnM x
n
sym(E,F )
The bounded sets B in EM x,K,h(U, F ) are exactly those B ⊆ E(U, F ) such that (bm)m ∈ FM x,h with
FM x,h :=(cid:8)(fk)k ∈ RN
Thus we see that EM x,K,h(U, F ) =(cid:8)f ∈ E(U, F ) : (kj∞(f )(cid:12)(cid:12)Kkk)k ∈ FM x,h(cid:9) holds with
bm := sup(cid:8)kj∞(f )(cid:12)(cid:12)Kkm : f ∈ B(cid:9).
k(cid:9) .
>0 : ∃ C > 0 : ∀ k ∈ N : fk ≤ ChkM x
Let U ⊆ E be convex open and K ⊆ U be convex compact, then define
E(M)(E ⊇ K, F ) := lim←−x∈Λ,h>0 EM x,h(E ⊇ K, F )
10
G. SCHINDL
and finally
(3.1)
i.e.
E{M}(E ⊇ K, F ) := lim−→x∈Λ,h>0 EM x,h(E ⊇ K, F )
E[M](U, F ) := lim←−K⊆U E[M](E ⊇ K, F ),
where K runs through all compact and convex subsets of U .
If Λ = R>0, then we can restrict in both cases to the countable diagonal, see also [9, 4.2.-4.4.]. We
EM n,n(E ⊇
E[M](U, F ) :=nf ∈ E(U, F ) : ∀ K : (f (k)(cid:12)(cid:12)K) ∈ E[M](E ⊇ K, F )o ,
EM 1/n,1/n(E ⊇ K, F ) and E{M}(E ⊇ K, F ) = lim−→n∈N>0
have E(M)(E ⊇ K, F ) = lim←−n∈N>0
K, F ).
As already mentioned in [7, Proposition 4.1. (3)] the space E{M}(E ⊇ K, F ) is not a Silva space for
lim−→x∈Λ,h>0 EM x,h(E ⊇
infinite dimensional E, because the connecting mappings in the inductive limit
: kαk ≤ 1} is bounded in EM k,k(E ⊇ K, R)
K, F ) are not compact any more. The set B := {α ∈ E
for each k ≥ 1. We have kαk0 = sup{α(x) : x ∈ K} ≤ sup{kxk : x ∈ K}, kαk1 = kαk ≤ 1 and
kαkm = 0 for each m ≥ 2. Moreover (Rn
yα)0 = α(x − y). But B is
not relatively compact in any EM k,k(E ⊇ K, R), k ≥ 1, because it is not even pointwise relatively
compact in C(K, L(E, R)).
Moreover we define
y α)k = 0 for n + k ≥ 1 and (R0
′
E(M),K(U, F ) := lim←−x∈Λ,h>0 EM x,K,h(U, F )
E{M},K(U, F ) := lim−→x∈Λ,h>0 EM x,K,h(U, F )
and so
E(M),K(U, F ) =(cid:8)f ∈ E(U, F ) : (kj∞(f )(cid:12)(cid:12)Kkk)k ∈ F(M)(cid:9)
E{M},K(U, F ) =(cid:8)f ∈ E(U, F ) : (kj∞(f )(cid:12)(cid:12)Kkk)k ∈ F{M}(cid:9)
with F(M) =Tx∈Λ,h>0 FM x,h, F{M} =Sx∈Λ,h>0 FM x,h.
The bounded sets B ⊆ E[M],K(U, F ) are exactly those B ⊆ E(U, F ) for which the sequence (bm)m,
bm := sup(cid:8)kj∞(f )(cid:12)(cid:12)Kkm : f ∈ B(cid:9), belongs to F[M].
Finally we introduce
lim←−K⊆U E[M],K(U, F ) =(cid:8)f ∈ E(U, F ) : ∀ K : (kj∞(f )(cid:12)(cid:12)Kkm)m ∈ F[M](cid:9) .
The next result generalizes [7, Proposition 4.1.].
Proposition 3.4. Let M be (M) with Λ = R>0, then the following completeness properties are
valid:
(1) EM x,h(E ⊇ K, F ) is a Banach space.
(2) E(M)(E ⊇ K, F ) is a Fréchet space.
(3) E{M}(E ⊇ K, F ) is a compactly regular (LB)-space, i.e. compact subsets are contained and
(4) E(M)(U, F ) and E{M}(U, F ) are complete.
compact in some step and so (c∞)-complete, webbed and ultrabornological.
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
11
(5) As locally convex vector spaces we have
E(M)(U, F ) = lim←−K⊆U E(M)(E ⊇ K, F ) = lim←−K⊆U E(M),K(U, F )
and
E{M}(U, F ) = lim←−K⊆U E{M}(E ⊇ K, F ) = lim←−K⊆U E{M},K(U, F ).
M x1 ,h1 ≤ k · kJ
Proof. (1) This was already shown in [7, Proposition 4.1. (1)].
(2) Holds since Λ = R>0.
(3) We can restrict to Λ = N>0 and proceed analogously as in [7, Proposition 4.1. (3)]. To show
that the inductive limit is compactly regular it suffices to show that there exists a sequence of
increasing 0-neighborhoods Un ∈ EM n,n(E ⊇ K, F ) such that for each n ∈ N there exists l ∈ N with
l ≥ n and for which the topologies of EM l,l(E ⊇ K, F ) and of EM k,k(E ⊇ K, F ) coincide on Un for
all k ≥ l.
In general, for indices x1 ≥ x2 and positive real numbers h1 ≥ h2 we have clearly by definition
(f ) := {g : kg − fkJ
k · kJ
M x,h ≤ ε} in EM x,h(E ⊇
K, F ) and we restrict to the diagonal x = h = n and identify U n,n with U n.
We show that for arbitrary n ∈ N>0 and n2 > n1 := 2n, for each ε > 0 and f ∈ U n
1 (0) there exists
δ > 0 such that U n2
δ (f ) ∩ U n
By assumption f ∈ U n
a+b+1 for
all a, b ∈ N. Consider g ∈ U n2
a , kgka,b ≤
na+b+1M n
a+b+1 for all a, b ∈ N.
We estimate similarly as in [7, Proposition 4.1. (3)]. So for given ε > 0 consider N ∈ N (minimal)
with 1
ε (f ).
(0) we have kfka ≤ naM n
(f ) ∩ U n,n
a and kfka,b ≤ na+b+1M n
(0), then kgka ≤ naM n
1 (0) ⊆ U n1
δ (f ) ∩ U n
a , kg − fka,b ≤ δna+b+1
a+b+1 and moreover kg − fka ≤ δna
2 and put δ := ε(cid:16) n1
n2(cid:17)N−1
. Consider now the ε-Ball U x,h
1 (0) = U n2,n2
1 (0) = U n,n
δ
2M n2
2N < ε
M n2
M x2 ,h2
M n2
N
1
1
2
ε
1
.
For a ≥ N we have 1
2a ≤ 1
2N < ε
2 (⋆), so use triangle-inequality to get
kg − fka ≤ kgka + kfka ≤ 2naM n
a = 2na
1M n
a
εna
1M n
a ≤ εna
1M n1
a
1
2a <
{z}(⋆)
a ≤ M n1
a
and the last inequality holds since n1 = 2n > n and so M n
have
for all a ∈ N. For a < N we
2M n2
a ≤ εna
1
M n2
a
N ≤ εna
M n2
1 ≤ εna
1M n1
a ,
kg − fka ≤ δna
n2(cid:17)N−1
n2(cid:17)a
≤(cid:16) n1
N , (cid:16) n1
a ≤ M n
since a < N , n1
n2
because M n
Analogously we can use the same estimates for k · ka,b instead of k · ka for each a, b ∈ N.
(4) In the Beurling-case we have a projective limit of Fréchet spaces, in the Roumieu-case a pro-
jective limit of (LB)-spaces, which are all compactly regular by (3) and so complete, too. Since
projective limits of complete spaces are complete we are done.
(5) This holds precisely by the same proof as given in [7, Proposition 4.1. (5)]
< 1 and finally M n1
a ≥ 1.
(cid:3)
12
G. SCHINDL
Let E, F be convenient, U ⊆ E be c∞-open, then define
(M)(U, F ) :=nf ∈ E(U, F ) : ∀ B : ∀ K ⊆ U ∩ EB : ∀ x ∈ Λ ∀ h > 0 :
E b
hkM x
k
: k ∈ N, a ∈ K,kvikB ≤ 1ois bounded in Fo
n f (k)(a)(v1, . . . , vk)
=nf ∈ E(U, F ) : ∀ B : ∀ K ⊆ U ∩ EB : ∀ x ∈ Λ ∀ h > 0 :
n dk
: k ∈ N, a ∈ K,kvikB ≤ 1ois bounded in Fo.
vf (a)(v1, . . . , vk)
hkM x
k
and
{M}(U, F ) :=nf ∈ E(U, F ) : ∀ B : ∀ K ⊆ U ∩ EB : ∃ x ∈ Λ ∃ h > 0 :
E b
hkM x
k
: k ∈ N, a ∈ K,kvikB ≤ 1ois bounded in Fo
n f (k)(a)(v1, . . . , vk)
=nf ∈ E(U, F ) : ∀ B : ∀ K ⊆ U ∩ EB : ∃ x ∈ Λ ∃ h > 0 :
n dk
: k ∈ N, a ∈ K,kvikB ≤ 1ois bounded in Fo.
vf (a)(v1, . . . , vk)
hkM x
k
B runs through all closed absolutely convex bounded subsets in E, EB is the complete vector space
generated by B with the Minkowski-functional k · kB. Finally K runs through all sets in U ∩ EB
which are compact w.r.t. the norm k · kB. If E and F both are Banach spaces and U ⊆ E open we
have E b
Now we give the most important definition:
[M](U, F ) = E[M](U, F ), where the latter space is introduced in (3.1).
E[M](U, F ) :=(cid:8)f ∈ E(U, F ) : ∀ α ∈ F ∗ : ∀ B : α ◦ f ◦ iB ∈ E[M](UB, R)(cid:9) ,
where B is running again through all closed absolutely convex bounded subsets in E, the mapping
iB : EB → E denotes the inclusion of EB in E and we write UB := i−1
B (U ). The initial locally
convex structure is now induced by all linear mappings
E[M](iB, α) : E[M](U, F ) −→ E[M](UB, R),
f 7→ α ◦ f ◦ iB.
E[M](U, F ) ⊆Qα,B E[M](UB, R) are convenient vector spaces as c∞-closed subspaces in the product:
Smoothness can be tested by composing with inclusions EB → E and α ∈ F ∗ as mentioned in [4,
2.14.4, 1.8]. Hence we obtain the representation
(3.2)
E[M](U, F ) := {f ∈ F U : ∀ α ∈ F ∗ ∀ B : α ◦ f ◦ iB ∈ E[M](UB, R)}.
All definitions given here are clearly generalizations of the definitions in [7, 4.2.]
matrices.
for constant
4. Projective descriptions for E[M]
In this section we are going to study one of the most important new techniques in this work. Using
abstract families of sequences of positive real numbers we prove projective representations for the
Roumieu-class E{M}. This technique is very important since we want to get rid of both existence
quantifiers in the definitions of E{M} so we want to generalize [7, Lemma 4.6.]. Furthermore we are
going to prove analogous results for the Beurling-case E(M) and generalize [7, Lemma 4.5.]. To do
so we have to show variations and generalizations of [4, Lemma 9.2.] (for the Roumieu-case) and
of the Lemma between Lemma 4.5. and Lemma 4.6. in [7] (for the Beurling-case).
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
13
We will obtain different projective representations for E[M]. The choice of the appropriate represen-
tation depends on the application in the proofs. To show closedness under composition in section 5,
see Theorem 5.8 and Theorem 5.9, we will have to use the versions using the Faà-di-Bruno-property
(M[FdB]). For the exponential laws in section 6 the versions only assuming (M) or (Msc) for M
are sufficient.
First we have to introduce several classes of sequences of positive real numbers (rk)k and (sk)k. It
is no restriction to assume r0 = 1 resp. s0 = 1 (normalization) for all occurring sequences.
RRoum
RRoum,sub := {(rk)k ∈ RRoum : rj+k ≤ rkrj ∀ j, k ∈ N}
RBeur
RBeur,sub := {(rk)k ∈ RBeur : rj+k ≤ rkrj ∀ j, k ∈ N}
SMRoum
k ≤ Ck
x}
SMRoum
k ≤ Ck
x}
SMRoum,sub := {(sk)k ∈ SMRoum : ∃ D > 0 ∀ j, k ∈ N : sj+k ≤ Dj+ksjsk}
SMRoum,FdB := {(sk)k ∈ SMRoum : ∃ (sk)k ∈ SMRoum ∃ D > 0 ∀ k ∈ N : sk ≤ Dk(so)k}
SMBeur
k ≤ Ck
:= {(sk)k ∈ RN
x}
SMBeur
k ≤ Ck
:= {(sk)k ∈ RN
x}
SMBeur,sub
:= {(sk)k ∈ SMBeur : ∃ D > 0 ∀ j, k ∈ N : sj+k ≤ Dj+ksjsk}
SMBeur,FdB := {(sk)k ∈ SMBeur : ∃ (sk)k ∈ SMBeur ∃ D > 0 ∀ k ∈ N : sk ≤ Dk(so)k}
For (sk)k ∈ SMRoum,SMBeur we have put
>0 : rktk → 0 as k → ∞ for each t > 0}
>0 : rktk → 0 as k → ∞ for some t > 0}
>0 : ∀ x ∈ Λ ∃ Cx > 0 ∀ k ∈ N : skmx
>0 : ∀ x ∈ Λ ∃ Cx > 0 ∀ k ∈ N : skM x
:= {(rk)k ∈ RN
:= {(rk)k ∈ RN
:= {(sk)k ∈ RN
:= {(sk)k ∈ RN
>0 : ∃ x ∈ Λ ∃ Cx > 0 ∀ k ∈ N : skmx
>0 : ∃ x ∈ Λ ∃ Cx > 0 ∀ k ∈ N : skM x
(so)k := min{sjsα1 ··· sαj : αi ∈ N>0, α1 + ··· + αj = k},
(so)0 := 1.
By definition SMRoum ⊆ SMRoum and (sk)k ∈ SMRoum if and only if (k!sk)k ∈ SMRoum respectively for the
Beurling-case. If (sk)k ∈ SMBeur,SMBeur holds for x ∈ Λ, then also for all y ≤ x, too. All occurring
sets are stable w.r.t. (·k)k 7→ (Bk·k)k for arbitrary B > 0.
Using [7, Lemma 4.6.] directly we get:
Proposition 4.1. Let M = {M x : x ∈ Λ} be (M), E, F be Banach spaces, U ⊆ E open and
f : U → F a E-mapping. Then the following are equivalent:
(1) f is E{M} = E b
(2) For each compact K ⊆ U there exists x ∈ Λ such that for each (rk)k ∈ RRoum
{M}
.
(cid:26) f (k)(a)(v1, . . . , vk)rk
M x
k
: a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
is bounded in F .
(3) For each compact K ⊆ U there exists x ∈ Λ such that for each (rk)k ∈ RRoum,sub there
exists ε > 0 such that
(cid:26) f (k)(a)(v1, . . . , vk)rkεk
M x
k
: a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
is bounded in F .
Note that E{M} = E b
to get rid of the second existence quantifier.
{M}
holds by Lemma 5.4 below, but for our approach in this work we also have
14
G. SCHINDL
bk
k!mx
k
tk, so ax
, the following are equivalent:
4.2. Roumieu-case with (M{FdB}). We prove the following generalization of [4, Lemma 9.2.]:
Lemma 4.3. Let M = {M x : x ∈ Λ} be (Msc) with Λ = N>0 and (M{FdB}). For a formal power
ktk =Pk≥0
series Pk≥0 ax
(1) There exists x ∈ Λ such that Pk≥0 ax
(2) Pk≥0
(3) The sequence (cid:0) bkrksk
(4) For each (rk)k ∈ RRoum,sub and for each (sk)k ∈ SMRoum,FdB there exists ε > 0 such that
εk(cid:1)k is bounded.
(cid:0) bkrksk
k := bk
M x
k
ktk has positive radius of convergence.
converges absolutely for all (rk)k ∈ RRoum and (sk)k ∈ SMRoum.
(cid:1)k is bounded for all (rk)k ∈ RRoum and (sk)k ∈ SMRoum.
Proof. (1) ⇒ (2) For the given series (x ∈ Λ coming from (1)) and arbitrary (rk)k and (sk)k as
considered in (2) we have
bkrksk
k!
k!
k!
bkrksk
k!
Xk≥0
=Xk≥0
ax
kmx
krksk =Xk≥0
(ax
ktk) (skmx
k)
≤C k
{z }
x
rk
tk ≤Xk≥0
(ax
,
t (cid:19)k
ktk) rk(cid:18) Cx
}
{z
→0,as k→∞
hence the first sum converges for t > 0 sufficiently small.
(2) ⇒ (3) ⇒ (4) are clearly satisfied.
(4) ⇒ (1) Since (M{FdB}) is satisfied and mx ≤ my for x ≤ y we can associate to each x ∈ Λ the
index α(x) := min{y ∈ Λ : (mx)◦(cid:22)my}. Since (mx)◦ ≤ (my)◦ for x ≤ y we also have α(x) ≤ α(y)
for such indices and limx→∞ α(x) = +∞.
On the other hand for y ≥ α(1) we can define β(y) := max{x ∈ Λ : α(x) ≤ y} which is clearly
well-defined. So β(y1) ≤ β(y2) for y1 ≤ y2, limy→∞ β(y) = +∞ and finally by construction for
each x ∈ N>0, x ≥ α(1), there exist y ∈ N>0, y ≤ x, with (my)◦(cid:22)mx. Note that this does not
imply (M(FdB)). W.l.o.g. we could assume that α(x) = x + 1 and so β(y) = y − 1. If M has in
addition (M{Cω}), i.e. the real analytic functions are contained in E{M}, then we can take w.l.o.g.
M 1 = (p!)p∈N, so m1
We prove by contradiction. So assume that each Pk≥0 ax
ktk would have radius of convergence 0.
Then we would get Pk≥0 ax
= +∞ for each n ∈ N>0 and each x ∈ Λ = N>0. Consider
now n ∈ N>0 and x := n + α(1) and so we find an increasing sequence (kn)n≥0 with k0 = 1,
limn→∞ kn = +∞ such that
(4.1)
p = 1 for each p and α(1) = 1.
k(cid:0) 1
n2(cid:1)k
kn−1
∀ n ∈ N>0 :
Xk=kn−1
an+α(1)
k
(cid:18) 1
n2(cid:19)k
≥ 1.
We put now
n2(cid:19)k
rk :=(cid:18) 1
for kn−1 ≤ k ≤ kn − 1, n ∈ N>0,
1
and show (rk)k ∈ RRoum,sub. For kn−1 ≤ k ≤ kn − 1 by definition rktk =(cid:0) t
n2(cid:1)k
, and so rktk → 0
as k → ∞ and all t > 0. Clearly (rk)k is also log-sub-additive.
In addition one can see that
(√rk)k ∈ RRoum,sub and so for all ε > 0 there exists kε ∈ N such that for all k ≥ kε we have
√rk
No we define s := (sk)k. We put sk := 1
n ∈ N>0, and show (sk)k ∈ SMRoum,FdB.
, where γ(k) := n + α(1) for kn−1 ≤ k ≤ kn − 1,
εk ≤ 1 ⇔ √rk ≤ εk.
mγ(k)
k
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
15
So let x ∈ Λ be arbitrary (large) but fixed, then for kn−1 ≤ k ≤ kn−1 we get skmx
For all k ∈ N we can estimate M x
This proves (sk)k ∈ SMRoum. Define
1
= M x
M γ(k)
x with some constant Cx > 0, because limk→∞ γ(k) = +∞.
k = mx
mγ(k)
k ≤ Ck
M γ(k)
.
k
k
k
k
k
sk :=
mβ(γ(k))
k
for kn−1 ≤ k ≤ kn − 1, n ∈ N>0,
and similarly we find a constant Dx > 0 such that skmx
x for each x ∈ Λ and k ∈ N because
limk→∞ β(γ(k)) = +∞. This proves (sk)k ∈ SMRoum. For δ1 + ··· + δj = k we obtain for k ∈ N with
kn−1 ≤ k ≤ kn − 1, n ∈ N>0:
1
k ≤ Dk
1
sk =
≤ Chk
mγ(k)
k
mβ(γ(k))
j
mβ(γ(k))
δ1
1
··· mβ(γ(k))
δj
≤ Chk
mβ(γ(j))
j
mβ(γ(δ1))
δ1
··· mβ(γ(δj))
δj
= Chk sj sδ1 ··· sδj ,
which precisely shows s(cid:22)so. The first inequality holds by (M{FdB}) and by definition of β, the
second because j, δ1, . . . , δj ≤ k. So s is as desired.
Moreover
=Xn≥1
= an+α(1)
Xk=kn−1
for kn−1 ≤ k ≤ kn− 1 (note that n(k) = n + α(1)
n2(cid:19)k
(cid:18) 1
Xk=kn−1
≥Xn≥1
=Xn≥1
an+α(1)
1 = +∞,
bkrksk
bkrksk
Xk≥1
M n+α(1)
kn−1
kn−1
k!
k!
k
k
k
because by definition bkk! sk = bk
for k ∈ [kn−1, kn − 1]).
Finally we show that (cid:0) bk
bk
Xk≥1
k!
k! √rksk(2ε)k(cid:1)k cannot be bounded for any ε > 0. First we get
√rkskεk ≥ Xk≥kε
≥ Xk≥kε
rksk = +∞.
bk
k!
bk
k!
√rksk εk
{z}≥√rk
But if the sequence would be bounded for some ε, then for all k ∈ N we would get bk
hence Pk≥0 bkk! √rkskεk ≤Pk≥0
We use Lemma 4.3 to generalize [7, Lemma 4.6.].
Proposition 4.4. Let M = {M x : x ∈ Λ} be (Msc) with Λ = N>0 and (M{FdB}). Let E, F be
Banach spaces, U ⊆ E open and f : U → F a E-mapping. Then the following are equivalent:
C
2k = 2C, a contradiction.
k! √rkskεk ≤ C
2k ,
(cid:3)
(1) f is E{M} = E b
(2) For each compact K ⊆ U , for each (rk)k ∈ RRoum and each (sk)k ∈ SMRoum the set
{M}
.
(cid:26) f (k)(a)(v1, . . . , vk)
k!
rksk : a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
is bounded in F .
(3) For each compact K ⊆ U , for each (rk)k ∈ RRoum,sub and for each (sk)k ∈ SMRoum,FdB,
there exists ε > 0 such that the set
(cid:26) f (k)(a)(v1, . . . , vk)
k!
rkskεk : a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
16
G. SCHINDL
is bounded in F .
Proof. (1) ⇒ (2) Let f be E{M} and K ⊆ U compact, then estimate as follows (where we use
Lemma 5.4 below):
f (k)(a)
k!
(cid:13)(cid:13)(cid:13)(cid:13)
rksk(cid:13)(cid:13)(cid:13)(cid:13)Lk(E,F )
=(cid:13)(cid:13)(cid:13)(cid:13)
f (k)(a)
k!mx
khk(cid:13)(cid:13)(cid:13)(cid:13)Lk(E,F ) rkhk skmx
{z}≤C k
k
x
≤(cid:13)(cid:13)(cid:13)(cid:13)
f (k)(a)
hkM x
k (cid:13)(cid:13)(cid:13)(cid:13)Lk(E,F )(cid:12)(cid:12)rk(Cxh)k(cid:12)(cid:12)
}
{z
→0, as k→∞
for a ∈ K, x ∈ Λ and h > 0 large enough (depending on K and f ) and for arbitrary (rk)k and (sk)k
as considered in (2).
(2) ⇒ (3) Take ε = 1.
(3) ⇒ (1) We use (4) ⇒ (1) in Lemma 4.3. Let K ⊆ U be an arbitrary compact set but fixed and put
x hk < +∞,
bk := supa∈K(cid:13)(cid:13)f (k)(a)(cid:13)(cid:13)Lk(E,F ). Then there exists h > 0 and x ∈ Λ such that supk∈N
hence f is E{M}.
4.5. Roumieu-case without (M{FdB}).
Lemma 4.6. Let M = {M x : x ∈ Λ} be (Msc) with Λ = N>0. For a formal power series
Pk≥0 ax
ktk =Pk≥0
(1) There exists x ∈ Λ such that Pk≥0 ax
(2) Pk≥0 bkrksk converges absolutely for all (rk)k ∈ RRoum and (sk)k ∈ SMRoum.
(3) (bkrksk)k is bounded for all (rk)k ∈ RRoum and (sk)k ∈ SMRoum,sub.
(4) For each (rk)k ∈ RRoum,sub and for each (sk)k ∈ SMRoum,sub there exists ε > 0 such that
ktk has positive radius of convergence.
tk the following are equivalent:
bk
M k
bk
M x
k
(cid:3)
(bkrkskεk)k is bounded.
if kn−1 ≤ k ≤ kn − 1. If M is (Msc) then we have M x
If M is (M), then in (3) and (4) we replace SMRoum,sub by SMRoum.
Proof. (1) ⇒ (2) ⇒ (3) ⇒ (4) is the same as in Lemma 4.3. For (4) ⇒ (1) we prove again by
contradiction. In (4.1) consider x = n ∈ N>0, take the same r = (rk)k and for s = (sk)k we put
sk := 1
j+k for each j, k ∈ N and
M n
k
x ∈ Λ and M x ≤ M y for x ≤ y. This implies (sk)k ∈ SMRoum,sub. If M is (M), then (sk)k ∈ SMRoum
holds by definition.
So we can prove a new version of Proposition 4.4.
Proposition 4.7. Let M = {M x : x ∈ Λ} be (Msc) with Λ = N>0. Let E, F be Banach spaces,
U ⊆ E open and f : U → F a E-mapping, then the following are equivalent:
k ≤ M x
(1) f is E{M} = E b
(2) For each compact K ⊆ U , for each (rk)k ∈ RRoum and for each (sk)k ∈ SMRoum the set
j M x
{M}
(cid:3)
.
nf (k)(a)(v1, . . . , vk)rksk : a ∈ K, k ∈ N,kvikE ≤ 1o
is bounded in F .
(3) For each compact K ⊆ U , for each (rk)k ∈ RRoum,sub and for each (sk)k ∈ SMRoum,sub, there
exists ε > 0 such that the set
nf (k)(a)(v1, . . . , vk)rkskεk : a ∈ K, k ∈ N,kvikE ≤ 1o
is bounded in F .
If M is (M), then in (3) we replace SMRoum,sub by SMRoum.
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
17
Proof. Use precisely the same arguments as in Proposition 4.4, for (3) ⇒ (1) we use (4) ⇒ (1) in
Lemma 4.6.
(cid:3)
4.8. Beurling-case with (M(FdB)).
Lemma 4.9. Let M = {M x : x ∈ Λ} be (Msc) with Λ = R>0 and (M(FdB)). For a formal power
series Pk≥0 ax
ktk =Pk≥0
(1) The series Pk≥0 ax
(2) For each (rk)k ∈ RBeur,sub and for each (sk)k ∈ SMBeur,FdB the sequence (cid:0) bk
ktk has infinite radius of convergence for each x ∈ Λ.
k! rkskδk(cid:1)k is
, the following are equivalent:
bounded for each δ > 0.
k := bk
M x
k
tk, so ax
bk
M x
k
x
ax
k
≤C k
bk
k!
ksk)
ax
k (mx
Xk≥0
≤Xk≥0
t(cid:19)k
(rktk)(cid:18) δ
rkskδk =Xk≥0
Proof. (1) ⇒ (2) Let (rk)k and (sk)k be given as considered in (2), then
(rktk)
(cid:18) δCx
t (cid:19)k
{z }
{z }
is absolutely convergent for each δ > 0. The index x ∈ Λ was chosen such that skmx
x holds
for all k ∈ N and it is depending on (sk)k ∈ SMBeur. The real number t > 0 was chosen in such a
way that rktk → 0 as k → ∞. Hence (cid:0) bk
k! rkskδk(cid:1)k is bounded for each δ > 0.
(2) ⇒ (1) Assume that there would exist x ∈ Λ such that Pk≥0 ax
convergence. Then there would exist h > 0 such thatPk≥0 ax
rk := 1
Clearly (rk)k ∈ RBeur,sub holds.
Also (sk)k is as desired. By (M(FdB)) for all x ∈ Λ there exists y ∈ Λ and D > 0 such that for all
α1 + ··· + αj = k we get
ktk would have finite radius of
knk = +∞ for each n > h. Put now
nk for some n > h and sk := 1
mx
k
k ≤ Ck
→0, as k→∞
.
sk :=
1
k ≤ Dk
mx
my
j my
1
α1 ··· my
αj
=: Dk sj sα1 ··· sαj ,
is bounded in F .
where we have put sj := 1
. We have y ≤ x, since (my)◦ ≤ (mx)◦ for y ≤ x. Clearly (sk)k ∈ SMBeur,
my
j
hence (sk)k ∈ SMBeur,FdB and so both sequences are as considered in (2). But then there would exist
C > 0 such that for all k ∈ N:
C >
skrk(2n2)k =
rk(2n2)k = ax
krkn2k2k = ax
knk2k.
bk
k!
bk
k!mx
k
1
2k = 2C, a contradiction.
knk ≤ CPk≥0
Hence Pk≥0 ax
Using the previous result we can show:
Proposition 4.10. Let M = {M x : x ∈ Λ} be (Msc) with Λ = R>0 and (M(FdB)). Let E, F be
Banach spaces, U ⊆ E open and f : U → F a E-mapping, then the following are equivalent:
(1) f is E(M) = E b
(2) For each compact K ⊆ U , for each (rk)k ∈ RBeur and for each (sk)k ∈ SMBeur the set
(cid:3)
(M).
(cid:26) f (k)(a)(v1, . . . , vk)
k!
rksk : a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
18
G. SCHINDL
(3) For each compact K ⊆ U , for each (rk)k ∈ RBeur,sub and for each (sk)k ∈ SMBeur,FdB the set
(cid:26) f (k)(a)(v1, . . . , vk)
k!
rkskδk : a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
is bounded in F for each δ > 0.
Proof. (1) ⇒ (2) Let f be E(M) and (rk)k, (sk)k given by (2), then we can estimate as follows
(where we use Lemma 5.2 below):
f (k)(a)
k!
(cid:13)(cid:13)(cid:13)(cid:13)
rksk(cid:13)(cid:13)(cid:13)(cid:13)Lk(E,F )
=(cid:13)(cid:13)(cid:13)(cid:13)
f (k)(a)
k!mx
khk(cid:13)(cid:13)(cid:13)(cid:13)Lk(E,F ) rkhk skmx
{z}≤C k
k
x
≤(cid:13)(cid:13)(cid:13)(cid:13)
f (k)(a)
M x
k hk (cid:13)(cid:13)(cid:13)(cid:13)Lk(E,F )(cid:12)(cid:12)rk(Cxh)k(cid:12)(cid:12)
}
{z
→0, as k→∞
k ≤ Ck
bk
M k
x and h > 0
x hk < +∞, hence f is E(M).
for a ∈ K. We have chosen x ∈ Λ depending on (sk)k ∈ SMBeur such that skmx
depending on given (rk)k ∈ RBeur such that rk(Cxh)k → 0 as k → ∞.
(2) ⇒ (3) Replace in (2) the sequence (rk)k by (rkδk)k.
(3) ⇒ (1) Use (2) ⇒ (1) in Lemma 4.9. Let K ⊆ U be a compact set, arbitrary but fixed.
Then put bk := supa∈K(cid:13)(cid:13)f (k)(a)(cid:13)(cid:13)Lk(E,F ) and so for each h > 0 and each x ∈ Λ we have that
supk∈N
4.11. Beurling-case without (M(FdB)).
Lemma 4.12. Let M = {M x : x ∈ Λ} be (Msc) with Λ = R>0. For a formal power series
Pk≥0 ax
ktk =Pk≥0
(1) The series Pk≥0 ax
(2) For each (rk)k ∈ RBeur,sub and for each (sk)k ∈ SMBeur,sub the sequence (bkrkskδk)k is
k := bk
M x
k
ktk has infinite radius of convergence for each x ∈ Λ.
, the following are equivalent:
tk, ax
bk
M x
k
(cid:3)
bounded for each δ > 0.
If M is (M), then in (2) we replace SMBeur,sub by SMBeur.
Proof. Proceed as in Lemma 4.9: For (2) ⇒ (1) we put sk := 1
, where x ∈ Λ is the index arising
by the contradiction argument.
Hence (sk)k ∈ SMbeur,sub holds whenever M is (Msc) since each M x is log-convex. If M is (M),
then (sk)k ∈ SMBeur is clear.
So we are able to prove:
Proposition 4.13. Let M = {M x : x ∈ Λ} be (Msc) with Λ = R>0. Let E, F be Banach spaces,
U ⊆ E open and f : U → F a E-mapping, then the following are equivalent:
M x
k
(cid:3)
(1) f is E(M) = E b
(2) For each compact K ⊆ U , for each (rk)k ∈ RBeur and for each (sk)k ∈ SMBeur the set
(M).
nf (k)(a)(v1, . . . , vk)rksk : a ∈ K, k ∈ N,kvikE ≤ 1o
is bounded in F .
(3) For each compact K ⊆ U , for each (rk)k ∈ RBeur,sub and for each (sk)k ∈ SMBeur,sub the set
nf (k)(a)(v1, . . . , vk)rkskδk : a ∈ K, k ∈ N,kvikE ≤ 1o
is bounded in F for each δ > 0.
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
19
If M is (M), then in (3) we replace SMBeur,sub by SMBeur.
Proof. The proof is the same as for Proposition 4.10. For (3) ⇒ (1) we use (2) ⇒ (1) in Lemma
4.12.
(cid:3)
5. Closedness under composition
5.1. First observations. First we generalize [7, Lemma 4.2.]:
Lemma 5.2. Let M be (M), then E(M) = E b
Proof. Let E, F be convenient, U ⊆ E a c∞-open subset and let f : U → F be a E-mapping. Then
we obtain the following equivalences, where the set B runs through all closed absolutely convex
bounded subsets in E and K runs through all sets in UB which are compact w.r.t. the norm k·kB:
(M).
hkM x
k
f ∈ E(M)(U, F )
⇐⇒ ∀ α ∈ F ∗ ∀ B ∀ K ⊆ UB ∀ x ∈ Λ ∀ h > 0 :
(cid:26) (α ◦ f )(k)(a)(v1, . . . , vk)
⇐⇒ ∀ B ∀ K ⊆ UB ∀ x ∈ Λ ∀ h > 0 ∀ α ∈ F ∗ :
α(cid:18)(cid:26) f (k)(a)(v1, . . . , vk)
⇐⇒ ∀ B ∀ K ⊆ UB ∀ x ∈ Λ ∀ h > 0 :
(cid:26) f (k)(a)(v1, . . . , vk)
⇐⇒ f ∈ E b
(M)(U, F ).
: a ∈ K, k ∈ N,kvikB ≤ 1(cid:27) is bounded in R
: a ∈ K, k ∈ N,kvikB ≤ 1(cid:27)(cid:19) is bounded in R
: a ∈ K, k ∈ N,kvikB ≤ 1(cid:27) is bounded in R
hkM x
k
hkM x
k
(cid:3)
{M}
. To see this we show the following result; for the case
But in general we do not have E{M} = E b
M := {M} see [7, Example 4.4.].
Lemma 5.3. Let M = {M x : x ∈ Λ} be (Msc) with Λ = N>0.
Then there exists f : R2 → RN>0 which is E{M}, but there is no reasonable topology on E{M}(R, RN>0)
such that the associated mapping f∨ : R → E{M}(R, RN>0 ) is E b
For a "reasonable topology" on E{M}(R, RN>0 ) we assume only that all point-evaluations evt :
E{M}(R, RN>0 ) → RN>0 are bounded linear mappings.
Proof. Consider f : R2 → RN>0 defined by f (s, t) := (θx(st))x∈Λ, θx ∈ E global
(R, R), see (chf). f
{M x}
is clearly E{M} since each linear functional on RN>0 depends only on finitely many coordinates. If
f∨ : R → E{M}(R, RN>0) would be E b
, then there would exist h > 0 and some y ∈ Λ such that
the set
(cid:26) (f∨)(k)(0)
: k ∈ N(cid:27)
hkM y
k
.
{M}
{M}
20
G. SCHINDL
would be bounded in E{M}(R, RN>0 ). But if we apply the bounded linear function evt for t = 2h,
then
(f∨)(k)(0)(2h)
hkM y
k
= (2h)kθ(k)
x (0)
hkM y
k
!x∈Λ
≥(cid:18) 2kM x
k
M y
k (cid:19)x∈Λ
{M}
we have to assume additional assumptions, see [7, Lemma 4.3.]
and so the coordinates are unbounded as k → ∞ whenever x ≥ y.
To get E{M} = E b
constant case.
Lemma 5.4. Let M be (M), let E, F be convenient and let U ⊆ E be a c∞-open subset. Assume
that there exists a Baire-vector-space-topology on the dual F ∗ for which the point evaluations evx
are continuous for all x ∈ F . Then f : U → F is E{M} if and only if f is E b
Proof. (⇐) is clear.
(⇒) Let B a closed absolutely convex bounded subset of E, furthermore consider a compact set K
in UB (w.r.t. k · kB) and introduce the sets
for the
{M}
(cid:3)
.
Ax,h,C :=(cid:26)α ∈ F ∗ : (α ◦ f )(k)(a)(v1, . . . , vk)
hkM x
k
≤ C, ∀ k ∈ N, a ∈ K,kvikB ≤ 1(cid:27) .
These sets are closed in F ∗ for the Baire-topology andSx∈Λ,h,C>0 Ax,h,C = F ∗ holds. Then, by the
Baire-property of F ∗, there exist x0 ∈ Λ, h0, C0 > 0 such that the interior ◦Ax0,h0,C0 is non-empty.
◦Ax0,h0,C0 − α0 ⇔
Let α0 ∈
εα + α0 ∈
Thus for all a ∈ K, k ∈ N and kvikB ≤ 1 we get
◦Ax0,h0,C0, then for all α ∈ F ∗ there exists ε > 0, such that we get εα ∈
◦Ax0,h0,C0.
(α ◦ f )(k)(a)(v1, . . . , vk) ≤
1
ε(cid:16)((εα) + α0) ◦ f )(k)(a) + (α0 ◦ f )(k)(a)(cid:17) ≤
2C0
ε
0M x0
hk
k .
k
.
(cid:3)
{M}
0 M x0
hk
: k ∈ N, a ∈ K,kvikB ≤ 1o is weakly bounded (in F ), hence bounded.
if M is coming from ω ∈ W which
So the set n f (k)(a)(v1,...,vk)
Since B was arbitrary we get f ∈ E b
If the matrix is non-constant and has infinite index set, e.g.
does not have (ω6) - see [9, Section 5], then another phenomenon appears.
Proposition 5.5. Let M = {M x : x ∈ Λ = N>0} be (Msc) with (M{strict}).
Then there exist locally convex vector spaces E and E{M}-curves c : R → E that are not E{M x} for
any x ∈ Λ, i.e. E{M}(R, E) (Sx∈Λ E{M x}(R, E).
Proof. By (M{strict}) we have that for each x ∈ Λ we can find x1 ∈ Λ, x1 > x, such that
E{M x} ( E{M x1}. Iterating (M{strict}) we obtain a strictly increasing sequence (xi)i≥0 with x0 = x
and limi→∞ xi = +∞, w.l.o.g. one could assume that M = {M xi : i ∈ N}.
So let x ∈ Λ be arbitrary but from now on fixed and set E := RN. Consider a curve c : R → RN,
c(t) = (ci(t))i∈N = (c0(t), c1(t), . . . ), with the following property: c0 is E b
, and for each i ≥ 1
we assume ci ∈ E{M xi}\E{M xi−1}.
The curve c is E{M} since each α ∈ (RN)∗ = R(N) depends only on finitely many coordinates. Let i
be the maximal of these coordinates. Then α ◦ c ∈ E{M xi}(R, R), thus c ∈ E{M}(R, RN).
{M x0}
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
21
If there would exist some y ∈ Λ such that c is E{M y}, then for each α ∈ R(N) we would get that
α ◦ c ∈ E{M y}(R, R). According to this y we choose a linear functional α depending on at least
i0 + 1 many coordinates where xi0 > y.
(cid:3)
5.6. Closedness under composition of E[M].
Definition 5.7. Let E be a convenient vector space. A E[M]-Banach-plot in E is a mapping
c : D → E such that c ∈ E[M] and D denotes an open set in some Banach space F . It is sufficient
to consider the open unit ball D = oF .
Using the definitions and projective representations of section 4 we can generalize [7, Theorem 4.8.].
Theorem 5.8. Let M be (Msc) with Λ = R>0, let U ⊆ E be a c∞-open subset in a convenient
vector space E and F be a Banach space.
If M has (M[FdB]) and f : U → F , then f ∈ E[M] implies f ◦ c ∈ E[M] for all E[M]-Banach plots c.
The converse implication holds always by the definitions given in 3.3.
Proof. We follow the proof of [7, Theorem 4.8.] and apply Proposition 4.4 for the Roumieu- and
Proposition 4.10 for the Beurling-case.
(a) Beurling-case E(M).
We have to show that f ◦ c is E(M) for each E(M)-Banach-plot c : G ⊇ D → E, where D denotes
the open unit ball in an arbitrary Banach-space G. By (3) in Proposition 4.10 we have to prove
that for each compact K ⊆ D and for each (rk)k ∈ RBeur,sub, (sk)k ∈ SMBeur,FdB the set
(cid:26) (f ◦ c)(k)(a)(v1, . . . , vk)
k!
rkskδk : a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
is bounded in F for each δ > 0. So let δ > 0, the sequences (rk)k, (sk)k, and finally a compact
(w.l.o.g. convex) set K ⊆ D be given, arbitrary but from now on fixed. Then for each α ∈ E∗ by
assumption and by (2) in Proposition 4.10 applied to the sequence (rk(2Dδ)k)k and (sk)k ∈ SMBeur,
where the constant D is coming from sk ≤ Dk(so)k (since (sk)k ∈ SMBeur,FdB), the set
: a ∈ K, k ∈ N,kvikG ≤ 1(cid:27)
(cid:26) (α ◦ c)(k)(a)(v1, . . . , vk)rk sk(2Dδ)k
(5.1)
k!
is bounded in R. So the set
(cid:26) c(k)(a)(v1, . . . , vk)rk sk(2Dδ)k
k!
: a ∈ K, k ∈ N,kvikG ≤ 1(cid:27)
is contained in some closed absolutely convex bounded subset B of E, hence
(5.2)
kc(k)(a)kLk(G,EB)rk skδk
k!
1
(2D)k .
≤
We proceed now as in [7, Theorem 4.8.]. c(K) is compact in EB since the mapping c : K → EB is
Lipschitzian: For all a, b ∈ K we get c(a) − c(b) ∈ ka−bkG
2Dr1 s1δ B. Then we estimate for all δ > 0 and
αi!
kc(αi)(a)kLαi (G,EB )rαi sαiδαi
}
(2D)α1 ···
{z
αj = 1
≤ 1
1
(2D)
(2D)k
(1 + C1h)k−1
22
G. SCHINDL
k ∈ N>0 as follows:
rkskδk(cid:13)(cid:13)(cid:13)(cid:13)Lk(G,F )
k!
(f ◦ c)(k)(a)
(cid:13)(cid:13)(cid:13)(cid:13)
≤Xj≥0 Xα∈Nj
>0,Pj
i=1 αi=k
j
j!
Dk(cid:13)(cid:13)f (j)(c(a))(cid:13)(cid:13)Lj (EB ,F ) sj
Yi=1
(cid:13)(cid:13)f (j)(c(a))(cid:13)(cid:13)Lj(EB ,F )
{z }
}
2(cid:19)k
·Xj≥0(cid:18)k − 1
j − 1(cid:19)(hC1)j−1 = (ChC1)(cid:18) 1
(cid:19)k
(sjmx
j )
(⋆)≤Chj
{z
i=1 αi=k
j!mx
j
≤C j
.
1
>0,Pj
2(cid:19)k
≤(cid:18) 1
Xj≥0 Xα∈Nj
≤ (ChC1)(cid:18) 1
2(cid:19)k
≤ (ChC1)(cid:18) (1 + C1h)
2
We have to choose x ∈ Λ according to (sj )j ∈ SMBeur (arising in SMBeur,FdB) such that sjmx
1 for
some constant C1 > 0 and all j ∈ N. Since f ∈ E(M), we obtain the estimate (⋆) with this index x
and arbitrary h > 0 for a constant C = Cx,h and all j ∈ N. Finally we can choose h := 1
and so
the expression at the beginning is bounded by C = Cx,1/C1.
(b) Roumieu-case E{M}.
Use Proposition 4.4 and by (3) there it is sufficient to show that each compact K ⊆ D and for each
(rk)k ∈ RRoum,sub, (sk)k ∈ SMRoum,FdB, there exists ε > 0 such that the set
(5.3)
j ≤ Cj
C1
(cid:26) (f ◦ c)(k)(a)(v1, . . . , vk)
k!
rkskεk : a ∈ K, k ∈ N,kvikE ≤ 1(cid:27)
2
is bounded in F .
We use the same proof as above and replace in (2) in Proposition 4.4 the sequence (rk)k by
((2D)krk)k, where D is the constant arising in sk ≤ Dk(so)k (since (sk)k ∈ SMRoum,FdB and so
(sk)k ∈ SMRoum). Then we take δ = 1 in (5.1), in (5.2) and in the Lipschitz-argument. We can use
now precisely the same estimate as for the Beurling-case (for δ = 1) and so we have shown (5.3)
(1+C1h) . Note that f ∈ E{M}, hence we have to consider x ∈ Λ and h > 0 sufficiently
for ε =
large to obtain estimate (⋆) for some constant C. According to this chosen x ∈ Λ we can estimate
1 for a constant C1 and all j ∈ N, since (sj)j ∈ SMRoum.
sjmx
Using Theorem 5.8 we can generalize [7, Theorem 4.9.].
Theorem 5.9. Let M be (Msc) with Λ = R>0. Let E, F, G be convenient vector spaces, U ⊆ E
and V ⊆ F be c∞-open and f : U → F , g : V → G with f (U ) ⊆ V .
j ≤ Cj
(cid:3)
(a) If M(FdB), then f, g ∈ E(M) implies g ◦ f ∈ E(M).
(b) If M{FdB}, then f, g ∈ E{M} implies g ◦ f ∈ E{M}.
Proof. By definition of E[M] we have to show that for all closed absolutely convex bounded subsets
B ⊆ E and for all α ∈ G∗ the composite α◦ g◦ f ◦ iB : UB → R is E[M]. By assumption f ◦ iB ∈ E[M]
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
23
and α ◦ g ∈ E[M] hold, so we can use Theorem 5.8 to obtain the desired implication. Note that
f ◦ iB is a E[M]-Banach plot.
(cid:3)
6. Exponential laws for E[M]
We start with the generalization of [7, Lemma 5.1.].
Lemma 6.1. Let M be (M) or (Msc) with Λ = R>0, let E be Banach and U ⊆ E open. Let F be
convenient and B a family of bounded linear functionals on F which together detect bounded sets,
i.e. B ⊆ E is bounded in E if and only if α(B) is bounded in R for all α ∈ B. Then we have
f ∈ E[M](U, F ) ⇔ α ◦ f ∈ E[M](U, R)
∀ α ∈ B.
Proof. For E-curves this follows by [4, 2.1., 2.11.], and so by composing with such curves for
E-mappings f : U → F .
In the Roumieu-case we use (1) ⇔ (2) in Proposition 4.7. Hence for arbitrary α ∈ F ∗ the mapping
α ◦ f is E{M} if and only if for each compact K ⊆ U the set
n(α ◦ f )(k)(a)(v1, . . . , vk)rksk : a ∈ K, k ∈ N,kvikE ≤ 1o
is bounded in R for each (rk)k ∈ RRoum and for each (sk)k ∈ SMRoum. So the smooth mapping
f : U → F is E{M} if and only if the set
nf (k)(a)(v1, . . . , vk)rksk : a ∈ K, k ∈ N,kvikE ≤ 1o
is bounded in F , for each compact K ⊆ U , (rk)k ∈ RRoum and for each (sk)k ∈ SMRoum.
Because B detects bounded sets we can replace in the above equivalences F ∗ by B.
For the Beurling-case proceed analogously and use (1) ⇔ (2) in Proposition 4.13.
Now we are able to prove Cartesian closedness for classes E[M] and so generalize [7, Theorem 5.2.].
Theorem 6.2. Let M be (Msc) with Λ = R>0, let Ui ⊆ Ei be c∞-open subsets in convenient
vector spaces Ei for i = 1, 2 and moreover let F be also a convenient vector space. Then we obtain:
(cid:3)
(a) If (M{mg}), then
(b) If (M(mg)), then
f ∈ E{M}(U1 × U2, F ) ⇐⇒ f∨ ∈ E{M}(U1,E{M}(U2, F )).
f ∈ E(M)(U1 × U2, F ) ⇐⇒ f∨ ∈ E(M)(U1,E(M)(U2, F )).
Important remarks:
(i) In both cases (⇐=) holds also without (M{mg}) respectively (M(mg)).
(ii) To prove (⇐=) it is sufficient to assume that M is (M) and (M[alg]).
(iii) For the proof it is not necessary to assume that E{M} respectively E(M) is a category, i.e.
(iv) If M is (Msc) with Λ = R>0, (M[mg]) and (M[FdB]), then by Theorem 6.2 and Theorem
closedness under composition.
5.9 the category E[M] is cartesian closed.
Proof. The technique and methods are completely analogous to [7, Theorem 5.2.], for convenience
of the reader we give the full proof.
As shown in [4, 3.12.] we have E(U1 × U2, F ) ∼= E(U1,E(U2, F )). So we assume form now on that
all occurring mappings are smooth. Let B ⊆ E1 × E2 and Bi ⊆ Ei, i = 1, 2, where B, B1, B2 run
24
G. SCHINDL
through all closed absolutely convex bounded subsets. Similarly as shown in [7, Theorem 5.2.] we
get:
f ∈ E[M](U1 × U2, F )
⇔ ∀ α ∈ F ∗ ∀ B : α ◦ f ◦ iB ∈ E[M]((U1 × U2)B, R)
⇔ ∀ α ∈ F ∗ ∀ B1, B2 : α ◦ f ◦ (iB1 × iB2) ∈ E[M]((U1)B1 × (U2)B2 , R)
and
f∨ ∈ E[M](U1,E[M](U2, F ))
⇔ ∀ B1 : f∨ ◦ iB1 ∈ E[M]((U1)B1 ,E[M](U2, F ))
⇔ ∀ α ∈ F ∗ ∀ B1, B2 : E[M](iB2 , α) ◦ f∨ ◦ iB1 ∈ E[M]((U1)B1 ,E[M]((U2)B2 , R)),
where Lemma 6.1 is used and note that the linear mappings E[M](iB2 , α) generate the bornology.
With these preparations we are able to restrict ourselves to Ui ⊆ Ei open sets in Banach spaces Ei
and F = R. We start now with (=⇒) for both cases.
Let f ∈ E[M](U1 × U2, R), then clearly f∨ takes values in the space E[M](U2, R).
First we show that
Claim. f∨ : U1 → E[M](U2, R) is E with dj f∨ = (∂j
E[M](U2, R) are convenient vector spaces, hence by [4, 5.20.] it suffices to prove that the iterated
unidirectional derivatives dj
1f (x,·)(vj ), and are separately bounded for
x and v in compact subsets. For j = 1 and x, v, y fixed we consider the smooth curve c : t 7→
f (x + tv, y). Then, by the fundamental theorem of calculus, we obtain:
vf∨(x) exist, are equal to ∂j
1f )∨.
t
f∨(x + tv) − f∨(x)
= tZ 1
sZ 1
0
0
c(t) − c(0)
− c′(0)
∂2
1 f (x + tsrv, y)(v, v)drds.
t
(y) − (∂1f )∨(x)(y)(v) =
c′′(tsr)drds = tZ 1
0
sZ 1
0
1 f )∨(K1)(o(E1×E1)) is bounded in E[M](U2, R) and for each compact set K1 ⊆ U1 this expression
(∂2
is Mackey-convergent to 0 in E[M](U2, R) as t → 0. Hence dvf∨(x) exists an is equal to ∂1f (x,·)(v).
The induction argument is completely the same as in [7, Theorem 5.2.].
We distinguish now between the Roumieu- and the Beurling-case.
The Beurling-case.
We have to show that f∨ : U1 → E(M)(U2, R) is E(M).
By Lemma 6.1 it suffices to prove that f∨ : U1 → EM x,h(E2 ⊇ K2, R) is E b
(M) = E(M) for each
K2 ⊆ U2 compact, each h > 0 and x ∈ Λ = R>0. This holds, because each α ∈ (E(M)(U2, R))∗
factorizes over EM x,h(E2 ⊇ K2, R) for some K2, h and x.
So we have to show that for each compact sets K1 ⊆ U1, K2 ⊆ U2, each h1, h2 > 0 and each
x1, x2 ∈ Λ, the set
(6.1)
( dk1 f∨(a1)(v1
hk1
1 M x1
k1
1, . . . , v1
k1
)
: a1 ∈ K1, k1 ∈ N,kv1
jkE1 ≤ 1)
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
25
is bounded in the space EM x2 ,h2(E2 ⊇ K2, R). Equivalently, for all compact sets K1, K2, for all
h1, h1 > 0 and all x1, x2 ∈ Λ the set
(6.2)
2 ∂k1
( ∂k2
1 f (a1, a2)(v1
hk2
2 hk1
1, . . . , v1
k1
1 M x2
M x1
k2
k1
; v2
1, . . . , v2
k2
)
: ai ∈ Ki, ki ∈ N,kvi
jkEi ≤ 1; i = 1, 2)
is bounded in R.
Let a1 ∈ K1, k1 ∈ N, then we obtain the following estimate:
J
)
1, . . . , v1
k1
1 f (a1, a2)(v1
hk1
1 hk2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Ck1+k2(cid:12)(cid:12)(cid:12)∂k2
dk1 f∨(a1)(v1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
= sup
≤
{z}
(M(mg))
≤ sup
hk1
1 M x1
k1
2 ∂k1
(cid:12)(cid:12)(cid:12)∂k2
sup
(cid:12)(cid:12)(cid:12)∂k2
M x2 ,K2,h2
1, . . . , v1
k1
2 M x1
M x2
k1
k2
; v2
1, . . . , v2
k2
)(cid:12)(cid:12)(cid:12)
2 ∂k1
1 f (a1, a2)(v1
hk1
1 hk2
1, . . . , v1
k1
2 M y
k1+k2
; v2
1, . . . , v2
k2
)(cid:12)(cid:12)(cid:12)
: a2 ∈ K2, k2 ∈ N,kv2
jkE2 ≤ 1
: a2 ∈ K2, k2 ∈ N,kv2
jkE2 ≤ 1
2 ∂k1
1 f (a1, a2)(v1
1, . . . , v1
k1
hk1+k2 M y
k1+k2
; v2
1, . . . , v2
k2
)(cid:12)(cid:12)(cid:12)
: a2 ∈ K2, k2 ∈ N,kv2
< +∞,
jkE2 ≤ 1
lim−→h2>0EM x2 ,h2(E2 ⊇ K2, R).
lim−→h2>0EM x2 ,h2(E2 ⊇ K2, R) is E b
where we have put h := 1
C min{h1, h2}. Note that f is E(M) and so for arbitrary h1, h2 > 0 and
x1, x2 ∈ Λ we can find y ∈ Λ and h > 0 such that the last inequality is valid. This shows that f∨
is E(M).
The Roumieu-case.
By Lemma 6.1 it suffices to prove that f∨ : U1 → lim−→x2∈Λ
{M} ⊆ E{M}
for each compact set K2 ⊆ U2. This holds because each α ∈ (E{M}(U2, R))∗ factorizes over some
lim−→x2∈Λ
So we have to prove that for all K1 ⊆ U1, K2 ⊆ U2 compact there exist h1 > 0 and some x1 ∈ Λ
lim−→h2>0EM x2 ,h2(E2 ⊇ K2, R). Equivalently, we have to
such that the set in (6.1) is bounded in lim−→x2∈Λ
show that for all K1, K2 compact there exist h1, h2 > 0 and x1, x2 ∈ Λ such that the set in (6.2) is
bounded in R.
We can use now the same estimate as for the above Beurling-case and use (M{mg}). First, because
f is E{M} and by (3) in Proposition 3.4 we obtain that there exist some h > 0 and y ∈ Λ, such that
the last set
sup(cid:26) (cid:12)
2 ∂k1
∂k2
(cid:12)
(cid:12)
in the Beurling estimate is bounded. For this y ∈ Λ we obtain by (M{mg}) that there exist some
x1, x2 ∈ Λ and C > 0 such that M y
k holds for all j, k ∈ N. So we can put in the
estimate now hi := Ch for i = 1, 2 to get, that f∨ is E{M}.
Now we start with (⇐=) for both cases.
: a2 ∈ K2, k2 ∈ N,kv2
jkE2 ≤ 1(cid:27)
j+k ≤ Cj+kM x1
1 f (a1,a2)(v1
j M x2
hk1 +k2 M y
1 ,...,v2
k2
1,...,v1
k1
k1 +k2
)(cid:12)
(cid:12)
(cid:12)
;v2
26
G. SCHINDL
Let f∨ : U1 → E[M](U2, R) be E[M]. The mapping f∨ : U1 → E[M](U2, R) → E(U2, R) is E, hence it
remains to show that f ∈ E[M](U1 × U2, R).
The Beurling-case.
For each compact K2 ⊆ U2, each h2 > 0 and each x2 ∈ Λ, the mapping f∨ : U1 → EM x2 ,h2(E2 ⊇
K2, R) is E b
(M) = E(M). This means that for all compact K1 ⊆ U1, K2 ⊆ U2, each h1, h2 > 0
and each x1, x2 ∈ Λ the set in (6.1) is bounded in EM x2 ,h2(E2 ⊇ K2, R). Because it is contained
in the space EM x2 ,K2,h2(U2, R) := {f ∈ E(U2, R) : j∞(f )K2 ∈ EM x2 ,h2(E2 ⊇ K2, R)} with semi-
norm kfkJ
, it is also bounded in this space and so the set in (6.2) is
bounded in R.
k ≤ M x
By assumption each M x is log-convex and so M x
(M(alg)) would be sufficient. Let a1 ∈ K, k1 ∈ N and kv1
j+k for all j, k ∈ N. For the next estimate
jkE1 ≤ 1, then:
:= kj∞(f )K2kJ
M x2 ,K2,h2
j M x
M x2 ,h2
+∞ >(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
dk1 f∨(a1)(v1
1, . . . , v1
k1
)
J
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
hk1
1 M x1
k1
2 ∂k1
1 f (a1, a2)(v1
hk1
1 hk2
M x2 ,K2,h2
1, . . . , v1
k1
2 M x1
M x2
k1
k2
(cid:12)(cid:12)(cid:12)∂k2
(cid:12)(cid:12)(cid:12)∂k2
= sup
≥ sup
2 ∂k1
1 f (a1, a2)(v1
1, . . . , v1
k1
; v2
1, . . . , v2
k2
hk1+k2 M y
k1+k2
; v2
1, . . . , v2
k2
)(cid:12)(cid:12)(cid:12)
)(cid:12)(cid:12)(cid:12)
: a2 ∈ K2, k2 ∈ N,(cid:13)(cid:13)v2
: a2 ∈ K2, k2 ∈ N,(cid:13)(cid:13)v2
j(cid:13)(cid:13)E2 ≤ 1
j(cid:13)(cid:13)E2 ≤ 1
lim−→h2>0EM x2 ,h2(E2 ⊇ K2, R) is E{M}. By (3)
lim−→h2>0EM x2 ,h2(E2 ⊇ K2, R)(cid:1)∗ can be equipped with the Baire-
(EM x2 ,h2(E2 ⊇ K2, R))∗.
lim−→h2>0EM x2 ,h2 (E2 ⊇ K2, R)
where we have put y := max{x1, x2} and h := max{h1, h2} (put h := C max{h1, h2}, where y ∈ Λ
and C > 0 are coming from (M(alg))). So we have shown that f is E(M).
The Roumieu-case.
For each compact K2 ⊆ U2 the mapping f∨ : U1 → lim−→x2∈Λ
in Proposition 3.4 the dual space(cid:0) lim−→x2∈Λ
vector-space-topology of the countable limit of Banach spaces lim←−x2∈Λ
lim←−h2>0
Now we can use Lemma 5.4 to conclude that the mapping f∨ : U1 → lim−→x2∈Λ
is E b
By (3) in Proposition 3.4 this inductive limit is countable and compactly regular and so for each
compact K1 ⊆ U1 there exist h1 > 0 and x1 ∈ Λ such that the set in (6.1) is bounded in
EM x2 ,h2(E2 ⊇ K2, R) for some h2 > 0 and x2 ∈ Λ. Because it is contained EM x2 ,K2,h2(U2, R) := {f ∈
E(U2, R) : j∞(f )K2 ∈ EM x2 ,h2 (E2 ⊇ K2, R)} with semi-norm kfkJ
,
M x2 ,h2
it is also bounded in this space and so the set in (6.2) is bounded (in R) with those given h1, h2, x1, x2.
But now we can use the same estimate as in the above Beurling-case to conclude that f is E{M}.
Similarly (M{alg}) would be sufficient for this step.
Using Theorem 6.2 we can prove now the matrix generalization of [7, Corollary 5.5.]:
Corollary 6.3. Let M be a weight matrix as assumed in Theorem 6.2. Let E, F, Ei, Fi, G be
convenient vector spaces and let U and V be c∞-open subsets. Then we get
M x2 ,K2,h2 := kj∞(f )K2kJ
{M}
(cid:3)
.
(1) The exponential law
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
27
E[M](U,E[M](V, G)) ∼= E[M](U × V, G)
holds, it is a linear E[M]-diffeomorphism of convenient vector spaces.
The following mappings are E[M]:
(2) ev : E[M](U, F ) × U → F given by ev(f, x) = f (x).
(3) ins : E → E[M](F, E × F ) given by ins(x)(y) = (x, y).
(4) (·)∧ : E[M](U,E[M](V, G)) → E[M](U × V, G).
(5) (·)∨ : E[M](U × V, G) → E[M](U,E[M](V, G)).
(6) Q :Qi E[M](Ei, Fi) → E[M] (Qi Ei,Qi Fi).
(7) comp : E[M](F, G) × E[M](U, F ) → E[M](U, G).
(8) E[M](·,·) : E[M](F, F1) × E[M](E1, E) → E[M](E[M](E, F ),E[M](E1, F1)) which is given by
If M has also (M[FdB]), then we get
(f, g) 7→ (h 7→ f ◦ h ◦ g).
Remark: (7) proves the claim of [9, Remark 4.23.].
6.4. Comparison of conditions (mg) and (M{mg}). In [7, Example 5.4.]
it was shown that
cartesian closedness fails for M = {M} if M does not satisfy (mg). In the weight matrix case we
can prove the following (counter)-example:
Example 6.5. There exist (non-constant) (Msc) weight matrices M with (M{mg}) but such that
no M x ∈ M satisfies (mg).
Proof. Let M = Ω be coming from ω ∈ W such that (ω6) does not hold, see [9, 5.5., Corollary
5.8. (2)]. The weights ω(t) := max{0, log(t)s}, s > 1, are concrete examples, see also [2] for the
consequences of (ω6).
In the next step we generalize [7, Example 5.4.]. We show that (M{mg}) is necessary for Theorem
6.2.
Lemma 6.6. Let M be (Msc) with Λ = N>0 but such that (M{mg}) does not hold. Then there
exists f ∈ E{M}(R2, C) such that the associated mapping f∨ : R → E{M}(R, C) is not E{M}.
Proof. We follow the proof of [7, Example 5.4.]. The negation of (M{mg}) gives
(6.3)
For this x ∈ Λ and the choice C = y = n, n ∈ N>0, we obtain sequences (jn)n and (kn)n such that
(jn)n is increasing, jn → ∞, kn ≥ 1 for each n ∈ N>0 and with
∃ x ∈ Λ ∀ C > 0 ∀ y ∈ Λ ∃ j, k ∈ N : M x
j+k > Cj+kM y
j M y
k .
(cid:3)
M x
M n
jn+kn
jn M n
kn!1/(kn+jn)
≥ n.
Define a linear functional α : E{M}(R, C) → C by
α(f ) :=Xn≥1
(√−1)3jn f (jn)(0)
jn njn
M n
.
Claim. α is bounded. For given f ∈ E{M}(R, C) we choose h > 0 and l ∈ Λ large enough and
estimate
(√−1)3jn f (jn)(0)
Xn≥0
≤ kfkM l,[−1,1],hXn≥0
≤Xn≥0(cid:12)(cid:12)f (jn)(0)(cid:12)(cid:12)
n(cid:19)jn
jn (cid:18) h
n(cid:19)jn
jn (cid:18) h
< +∞.
M l
jn
M n
M l
jn
M n
hjn M l
jn
M n
jn
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
njn(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
28
G. SCHINDL
Note that M l ≤ M n for l ≤ n and Pn≥0(cid:0) h
We apply α to θx ∈ E global
{M x}
(0, 0) = (√−1)β1+β2sx
For s, t ∈ R define ψx(s, t) := θx(s+t) and so ψx ∈ E global
{M}
for all (β1, β2) ∈ N2.
Claim. α ◦ ψ∨x is not E{M}. Let h > 0 and l ∈ Λ be arbitrary (large) but fixed and estimate as
follows:
(R, C) (see (2.4)), where x ∈ Λ is the index from (6.3).
n(cid:1)jn < +∞ for each h > 0.
(R2, C) with ψ(β1,β2)
x
β1+β2
kα ◦ ψ∨xkM l,[−1,1],h =
sup
t∈[−1,1],k∈N
(α ◦ ψ∨x )(k)(t)
hkM l
k
≥ sup
k∈N
1
1
hkM l
hkM l
k (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(√−1)3jn
Xn≥1
k Xn≥1
= sup
k∈N
= sup
k∈N
≥ sup
n∈N>0
M n
kn
hkn njn M l
kn
(√−1)jn+ksx
jn njn
M n
jn+k
1
hkM l
= sup
k∈N
sup
n∈N>0
sx
jn+k
jn njn ≥
M n
{z}
M x
kn ≥ sup
M n
n∈N>0
jn+kn
jn M n
k=kn
1
hkn M l
kn
M n
kn
M n
kn
sx
jn+kn
M n
jn njn
njn+kn
hkn njn
M n
kn
M l
kn
= +∞.
(0, 0)
jn njn
hkM l
1
x
M n
k (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(√−1)3jn ψ(jn,k)
Xn≥1
k (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(√−1)kXn≥1
jn njn(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
sx
jn+k
M n
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:3)
7. Remarks and special cases
7.1. More results for E[M]. Let M be (M) with Λ = R>0. Using the closed graph theorem [4,
is valid
52.10] the matrix generalization of the uniform boundedness principle [7, Theorem 6.1.]
for E[M], see [12, Theorem 12.4.1.]. All further results from [7, Chapter 8] can be transferred to
the matrix-case, see [12, 12.4., 12.6., 12.7.]. For the generalization of [7, Theorem 2.2.] see [12,
Proposition 9.4.4.].
Let M be (M) and assume that
(i) M is (Msc) with Λ = R>0 and has
(ii) (M[mg]) (⇒ (M[dc]));
(iii) for the Roumieu-case (MH), for the Beurling-case (M(Cω ));
(iv) (M[FdB]) or equivalently (M[rai]) (see [10, Lemma 1]).
Using [10, Theorems 5,6], where we characterized the required stability properties for E[M], all
results from [7, Chapter 9] can be transferred to the E[M]-case, see [12, 12.8.] for full proofs. Note
that the characterization theorem for the Beurling-case shown in [12, Chapter 8] is weaker than
[10, Theorem 6].
7.2. Special cases M = {M} and M = Ω. To apply all previous results to the constant case
M = {M} we have to assume that
(i) M ∈ LC;
(ii) lim inf p→∞(mp)1/p > 0 in the Roumieu-, limp→∞(mp)1/p = +∞ in the Beurling-case;
(iii) M has (mg)(⇒ (dc)),
(iv) M has (FdB) or equivalently (rai) (see also [9, Chapter 3]).
If M = Ω = {(Ωl
j)j : l > 0} with Ωl
j := exp(1/lϕ∗ω(lj)), then we assume that ω ∈ W and
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
29
(see [9, Corollary 5.15.]);
(i) (ω2) in the Roumieu-, (ω5) in the Beurling-case to guarantee (MH) respectively (M(Cω ))
(ii) (ω1′), i.e. ω is equivalent w.r.t. ∼ to a sub-additive weight, see [10, Theorems 3,4] and [9,
Chapter 6].
7.3. Weight matrices in the sense of Beaugendre, Schmets and Valdivia. Beaugendre in
[1] and Schmets and Valdivia in [13] have considered weight matrices in the following sense: Let
t = +∞ and Φ(0) = 0
Φ : [0, +∞) → R be a convex and increasing function with limt→∞
(w.l.o.g. - replace Φ by Ψ(t) := Φ(t) − Φ(0), see [13, Definition 16.]). We introduce the following
weight matrix
Φ(t)
MΦ := {(p!mΦ
ap)p∈N : a > 0}
mΦ
ap := exp(Φ(ap)).
In the literature the Beurling-case E(MΦ) was considered. We summarize some properties:
(i) MΦ is (Msc) and (M(Cω )) holds.
(ii) (M{L}) and (M(L)) both are satisfied, compare this with [9, Lemma 5.9. (5.10)] where
condition (ω1) is needed. As shown in [13, Lemma 17] we get both
∀ a > 0 ∀ h > 0 ∃ b > 0 (b > a) ∃ D > 0 ∀ p ∈ N>0 :
log(h) −
log(D)
p
1
p
≤
(Φ(bp) − Φ(ap))
and
∀ b > 0 ∀ h > 0 ∃ a > 0 (a < b) ∃ D > 0 ∀ p ∈ N>0 :
log(h) −
log(D)
p
1
p
≤
(Φ(bp) − Φ(ap)) ,
since convexity of Φ yields
Φ(bp) − Φ(ap)
p(b − a)
Φ(bp)
pb ⇔
≥
Φ(bp) − Φ(ap)
p
Φ(bp)
pb
≥
(b − a) → ∞
(7.1)
∀ a, b > 0, b > a :
as p → ∞.
a, b > 0, then we would get
(iii) (7.1) implies also that all sequences are pairwise not equivalent. If (mΦ
ap)p≈(mΦ
bp)p for all
∀ a > 0 ∀ b > 0 ∃ C ≥ 1 ∀ p ∈ N : mΦ
but the left hand side tends to infinity as p → ∞ whenever b > a. So MΦ has both
(M{strict}) and (M(strict)).
2 Φ(2ap) +
2 Φ(2aq) ≤ Φ(2ap) + Φ(2aq) for all a > 0 and p, q ∈ N and so
(iv) MΦ has (M{mg}) and (M(mg)). By convexity of Φ we get Φ(ap + aq) ≤ 1
(Φ(bp) − Φ(ap)) ≤ log(C),
bp ≤ CpmΦ
ap ⇔
1
1
p
M Φ
a(p+q) ≤ M Φ
bp · M Φ
bq ⇔ Φ(a(p + q)) ≤ Φ(bp) + Φ(bq)
holds with b = 2a.
(v) (M{FdB}) and (M(FdB)) both are satisfied. This is clear since each (mΦ
ap)p is log-convex,
see e.g. [9, 2.2. Lemma (1)].
Thus also for E[MΦ] the exponential laws in Theorem 6.2 and the consequences in Lemma 6.3 are
valid. Moreover the characterizing results [10, Theorems 5,6] and all further generalizations of the
results from [7] hold.
As special case one may consider Φ = ϕ∗ω for ω ∈ W. Then on the one hand one has the matrix MΦ
p)p := exp(1/aϕ∗ω(ap)) : a > 0} as
as defined before, on the other hand the weight matrix Ω := {(Ωa
30
G. SCHINDL
the approach in [3]. By definition we have
mΦ
ap := exp(Φ(ap)) = exp(1/aϕ∗ω(ap))a = (Ωa
p)a.
As we have already pointed out the weights ωs := max{0, log(t)s}, s > 1, generate an infinite
non-constant weight matrix. We denote the associated matrices by MΦ
Lemma 7.4. For any s > 1 the matrices MΦ
Proof. Let s > 1 be arbitrary but fixed. For t ≥ 0 we get ϕωs(t) = ωs(exp(t)) = (log(exp(t)))s = ts,
hence ϕ∗ωs(x) = sup{xy − ys : y ≥ 0} =: sup{fx,s(y) : y ≥ 0} for all x ≥ 0. A straightforward
computation shows
s and Ωs are equivalent w.r.t. both {≈} and (≈).
s and Ωs and prove:
s−1
s(cid:17) s
−(cid:16) x
= x
s
s−1 (cid:18) 1
s
s−1 −
1
1
s−1(cid:19) =: x
s
s
s
s−1 R(s)
mΦ
lp = exp(ls/(s−1)ps/(s−1)R(s)).
and so
(7.2)
s−1
s−1(cid:19) = x(cid:16) x
ϕ∗ωs (x) = fx,s(cid:18)(cid:16) x
s(cid:17) 1
s(cid:17) 1
p = exp(cid:16)l1/(s−1)ps/(s−1)R(s)(cid:17)
p = (exp(lR(2)))p2
s . Let l ∈ N>0 (large) and get Ωl
Ωl
.
lp = (exp(l2/4))p2
s ((cid:22))Ωs holds, too.
and mΦ
p)lp! = p!mΦ
= (exp(l/4))p2
p ≤ (Ωl
p)l ≤ (Ωl
lp for each p ∈ N since Ωl
The case s = 2 gives Ωl
Ωs{(cid:22)}MΦ
p ≥ 1
for each l > 0, p ∈ N.
MΦ
s {(cid:22)}Ωs. Let l > 0, then we have to find n > l > 0 and C ≥ 1 such that for all p ∈ N we
p ⇔ p! exp(ls/(s−1)ps/(s−1)R(s)) ≤ Cp exp(n1/(s−1)ps/(s−1)R(s)). So the choice
get p!mΦ
lp ≤ CpΩn
n = 2s−1ls is sufficient and analogously MΦ
Ωs((cid:22))MΦ
s . For each l > 0 (small) there exists C ≥ 1 and n > 0 such that for all p ∈ N we get
lp ⇔ exp(n1/(s−1)ps/(s−1)R(s)) ≤ Cpp! exp(ls/(s−1)ps/(s−1)R(s)), so the choice n = ls
p ≤ Cpp!mΦ
Ωn
is sufficient.
If ω ∈ W, then Ω has always both (M{mg}) and (M(mg)). But Ωl≈Ωn for all l, n > 0 holds if and
only if (mg) for some/each Ωl and if and only if (ω6) for ω, see [9, Chapter 5].
For MΦ this is not true any more. As we have already seen the sequences in MΦ are always
pairwise not equivalent.
ap)p∈N is log-convex, (mg) holds for this sequence if and only if mΦ
On the other hand, since (mΦ
a2p ≤
2p Φ(2ap) − 1
C2p(mΦ
p Φ(ap) ≤ log(C) for a constant C ≥ 1 and all p ∈ N, see [8, Theorem 1,
(3) ⇒ (2)]. So if Φ satisfies
(7.3)
then each (mΦ
ap)p∈N has (mg). In [1] a weight with (7.3) is called a weight of moderate growth. (7.3)
is valid for Φ = ϕ∗ω if and only if ω ∈ W has (ω6). This holds by the proof of (5.11.) in [9, Lemma
5.9.] and by applying the conjugate operator to (7.3) (note that ϕ∗∗ω = ϕω).
Finally consider Φ(t) := t log(t) for t ≥ 1 and Φ(t) := 0 for 0 ≤ t < 1. Each (mΦ
2p Φ(2ap) − 1
show that this yields the Gevrey-matrix G and which should be compared with [9, 5.19.].
p Φ(ap) = a log(2). More precisely Stirling's formula and mΦ
∃ D ≥ 1 ∀ t ≥ 0 : Φ(2t) ≤ 2Φ(t) + Dt,
ap)p∈N has (mg), since
ap = exp(Φ(ap)) = (ap)ap
ap)2 ⇔ 1
(cid:3)
1
References
[1] P. Beaugendre. Extensions de jets dans des intersections de classes non quasi-analytiques. Ann. Polon. Math.,
76(3):213 -- 243, 2001.
[2] J. Bonet, R. Meise, and S. N. Melikhov. A comparison of two different ways to define classes of ultradifferentiable
functions. Bull. Belg. Math. Soc. Simon Stevin, 14:424 -- 444, 2007.
THE CONVENIENT SETTING FOR ULTRADIFFERENTIABLE FUNCTIONS
31
[3] R. W. Braun, R. Meise, and B. A. Taylor. Ultradifferentiable functions and Fourier analysis. Results Math.,
17(3-4):206 -- 237, 1990.
[4] A. Kriegl and P. W. Michor. The convenient setting of global analysis, volume 53 of Mathematical Surveys and
Monographs. American Mathematical Society, Providence, RI, 1997. http://www.ams.org/online_bks/surv53/.
[5] A. Kriegl, P. W. Michor, and A. Rainer. The convenient setting for non-quasianalytic Denjoy -- Carleman differ-
entiable mappings. J. Funct. Anal., 256:3510 -- 3544, 2009.
[6] A. Kriegl, P. W. Michor, and A. Rainer. The convenient setting for quasianalytic Denjoy -- Carleman differentiable
mappings. J. Funct. Anal., 261(7), 2011.
[7] A. Kriegl, P. W. Michor, and A. Rainer. The convenient setting for Denjoy -- Carleman differentiable mappings
of Beurling and Roumieu type. Rev. Mat. Complut., 28(3):549 -- 597, 2015.
[8] W. Matsumoto. Characterization of the separativity of ultradifferentiable classes. J. Math. Kyoto Univ.,
24(4):667 -- 678, 1984.
[9] A. Rainer and G. Schindl. Composition in ultradifferentiable classes. Studia Mathematica, 224(2):97 -- 131, 2014.
[10] A. Rainer and G. Schindl. Equivalence of stability properties for ultradifferentiable function classes, 2014. ac-
cepted for publication in Rev. R. Acad. Cienc. Exactas Fis. Nat. Ser. A Math. RACSAM, available online at
http://arxiv.org/pdf/1407.6673.pdf.
[11] G. Schindl. Spaces of smooth functions of Denjoy-Carleman-type, 2009. Diploma Thesis, Universität Wien,
available online at http://othes.univie.ac.at/7715/1/2009-11-18_0304518.pdf .
[12] G. Schindl. Exponential laws for classes of Denjoy-Carleman-differentiable mappings, 2014. PhD Thesis, Uni-
versität Wien, available online at http://othes.univie.ac.at/32755/1/2014-01-26_0304518.pdf .
[13] J. Schmets and M. Valdivia. Extension properties in intersections of non quasi-analytic classes. Note di Matem-
atica, 25(2):159 -- 185, 2006.
[14] V. Thilliez. On quasi-analytic local rings. Expo. Math., 26:1 -- 23, 2008.
G. Schindl: Fakultät für Mathematik, Universität Wien, Oskar-Morgenstern-Platz 1, A-1090 Wien,
Austria
E-mail address: [email protected]
|
1311.0153 | 1 | 1311 | 2013-11-01T11:57:02 | Higher-order Sobolev embeddings and isoperimetric inequalities | [
"math.FA"
] | Optimal higher-order Sobolev type embeddings are shown to follow via isoperimetric inequalities. This establishes a higher-order analogue of a well-known link between first-order Sobolev embeddings and isoperimetric inequalities. Sobolev type inequalities of any order, involving arbitrary rearrangement-invariant norms, on open sets in $\rn$, possibly endowed with a measure density, are reduced to much simpler one-dimensional inequalities for suitable integral operators depending on the isoperimetric function of the relevant sets.
As a consequence, the optimal target space in the relevant Sobolev embeddings can be determined both in standard and in non-standard classes of function spaces and underlying measure spaces. In particular, our results are applied to any-order Sobolev embeddings in regular (John) domains of the Euclidean space, in Maz'ya classes of (possibly irregular) Euclidean domains described in terms of their isoperimetric function, and in families of product probability spaces, of which the Gauss space is a classical instance. | math.FA | math |
HIGHER-ORDER SOBOLEV EMBEDDINGS
AND ISOPERIMETRIC INEQUALITIES
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Abstract. Optimal higher-order Sobolev type embeddings are shown to follow via isoperimetric in-
equalities. This establishes a higher-order analogue of a well-known link between first-order Sobolev
embeddings and isoperimetric inequalities. Sobolev type inequalities of any order, involving arbitrary
rearrangement-invariant norms, on open sets in Rn, possibly endowed with a measure density, are
reduced to much simpler one-dimensional inequalities for suitable integral operators depending on the
isoperimetric function of the relevant sets. As a consequence, the optimal target space in the relevant
Sobolev embeddings can be determined both in standard and in non-standard classes of function spaces
and underlying measure spaces. In particular, our results are applied to any-order Sobolev embeddings
in regular (John) domains of the Euclidean space, in Maz'ya classes of (possibly irregular) Euclidean
domains described in terms of their isoperimetric function, and in families of product probability spaces,
of which the Gauss space is a classical instance.
Contents
1.
Introduction
2. An overview
3. Spaces of measurable functions
4. Spaces of Sobolev type and the isoperimetric function
5. Main results
6. Euclidean -- Sobolev embeddings
7. Sobolev embeddings in product probability spaces
8. Optimal target function norms
9. Proofs of the main results
10. Proofs of the Euclidean Sobolev embeddings
11. Proofs of the Sobolev embeddings in product probability spaces
References
1
2
6
10
14
18
23
28
33
41
45
53
1. Introduction
Sobolev inequalities and isoperimetric inequalities had traditionally been investigated along indepen-
dent lines of research, which had led to the cornerstone results by Sobolev [80, 81], Gagliardo [44] and
Nirenberg [70] on the one hand, and by De Giorgi [35] on the other hand, until their intimate connection
was discovered some half a century ago. Such breakthrough goes back to the work of Maz'ya [65, 66],
Date: August 8, 2018.
2000 Mathematics Subject Classification. 46E35, 46E30.
Key words and phrases. Isoperimetric function, higher-order Sobolev embeddings, rearrangement-invariant spaces,
John domains, Maz'ya domains, product probability measures, Gaussian Sobolev inequalities, Hardy operator.
This research was partly supported by the the research project of MIUR (Italian Ministry of University) Prin 2008
"Geometric aspects of partial differential equations and related topics", by GNAMPA of the Italian INdAM (National
Institute of High Mathematics), and by the grants 201/08/0383 and P201/13/14743S of the Grant Agency of the Czech
Republic. The third named author was also partly supported by the grant SVV-2013-267316.
1
2
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
who proved that quite general Sobolev inequalities are equivalent to either isoperimetric or isocapac-
itary inequalities. Independently, Federer and Fleming [42] also exploited De Giorgi's isoperimetric
inequality to exhibit the best constant in the special case of the Sobolev inequality for functions whose
gradient is integrable with power one in Rn. These advances paved the way to an extensive research,
along diverse directions, on the interplay between isoperimetric and Sobolev inequalities, and to a num-
ber of remarkable applications, such as the classics by Moser [69], Talenti [85], Aubin [3], Br´ezis and
Lieb [14]. The contributions to this field now constitute the corpus of a vast literature, which includes
the papers [1, 4, 6, 10, 11, 15, 16, 20, 22, 24, 27, 31, 33, 39, 40, 45, 47, 49, 50, 52, 53, 54, 59, 60, 68, 89]
and the monographs [17, 19, 21, 48, 67, 78]. Needless to say, this list of references is by no means
exhaustive.
The strength of the approach to Sobolev embeddings via isoperimetric inequalities stems from the
fact that not only it applies to a broad range of situations, but also typically yields sharp results.
The available results, however, essentially deal with first-order Sobolev inequalities, apart from few
exceptions on quite specific issues concerning the higher-order case. Indeed, isoperimetric inequalities
are usually considered ineffectual in proving optimal higher-order Sobolev embeddings. Customary
techniques that are crucial in the derivation of first-order Sobolev inequalities from isoperimetric in-
equalities, such as symmetrization, or just truncation, cannot be adapted to the proof of higher-order
Sobolev inequalities. A major drawback is that these operations do not preserve higher-order (weak)
differentiability. A new approach to the sharp Sobolev inequality in Rn, based on mass transportation
techniques, has been introduced in [34], and has later been developed in various papers to attack other
Sobolev type inequalities, but still in the first-order case. On the other hand, methods which can be
employed to handle higher-order Sobolev inequalities, such as representation formulas, Fourier trans-
forms, atomic decomposition, are not flexible enough to produce sharp conclusions in full generality.
A paradigmatic instance in this connection is provided by the standard Sobolev embedding in Rn to
which we alluded above, whose original proof via representation formulas [80, 81] does not include the
borderline case when derivatives are just integrable with power one. This case was restored in [44] and
[70] through a completely different technique that rests upon one-dimensional integration combined
with a clever use of Holder's inequality.
One main purpose of the present paper is to show that, this notwithstanding, isoperimetric in-
equalities do imply optimal higher-order Sobolev embeddings in quite general frameworks. Sobolev
embeddings for functions defined on underlying domains in Rn, equipped with fairly general measures,
are included in our discussion. Also, Sobolev-type norms built upon any rearrangement-invariant Ba-
nach function norm are considered. The use of isoperimetric inequalities is shown to allow for a
unified approach to the relevant embeddings, which is based on the reduction to considerably sim-
pler one-dimensional inequalities. Such reduction principle is crucial in a characterization of the best
possible target for arbitrary-order Sobolev embeddings, in the class of all rearrangement-invariant
Banach function spaces. As a consequence, the optimal target in arbitrary-order Sobolev embeddings
involving various customary and non-standard underlying domains and norms can be exhibited. In
fact, establishing optimal higher-order Gaussian Sobolev embeddings, namely Sobolev embeddings in
Rn endowed with the Gauss measure, was our original motivation for the present research. Failure of
standard strategies in the solution of this problem led us to develop the general picture which is now
the subject of this paper.
A key step in our proofs amounts to the development of a sharp iteration method involving subse-
quent applications of optimal Sobolev embeddings. We consider this method of independent interest
for its possible use in different problems, where regularity properties of functions endowed with higher-
order derivatives are in question.
We shall deal with Sobolev inequalities in an open connected set -- briefly, a domain -- Ω in Rn,
n ≥ 1, equipped with a finite measure ν which is absolutely continuous with respect to the Lebesgue
2. An overview
HIGHER-ORDER EMBEDDINGS
3
measure, with density ω. Namely, dν(x) = ω(x) dx, where ω is a Borel function such that ω(x) > 0
a.e. in Ω. Throughout the paper, we assume, for simplicity of notation, that ν is normalized in such a
way that ν(Ω) = 1. The basic case when ν is the Lebesgue measure will be referred to as Euclidean.
Sobolev embeddings of arbitrary order for functions defined in Ω, with unconstrained values on ∂Ω,
will be considered. However, the even simpler case of functions vanishing (in the suitable sense) on
∂Ω together with their derivatives up to the order m − 1 could be included in our discussion.
The isoperimetric inequality relative to (Ω, ν) tells us that
Pν (E, Ω) ≥ IΩ,ν(ν(E)),
(2.1)
where E is any measurable subset of Ω, and Pν (E, Ω) stands for its perimeter in Ω with respect to ν.
Moreover, IΩ,ν denotes the largest non-decreasing function in [0, 1
2 ] for which (2.1) holds, called the
isoperimetric function (or isoperimetric profile) of (Ω, ν), which was introduced in [65].
In the Euclidean case, (Ω, ν) will be simply denoted by Ω, and IΩ,ν by IΩ. The isoperimetric function
IΩ,ν is known only in few special instances, e.g. when Ω is an Euclidean ball [67], or agrees with the
space Rn equipped with the Gauss measure [13]. However, the asymptotic behavior of IΩ,ν at 0 -- the
piece of information relevant in our applications -- can be evaluated for various classes of domains,
including Euclidean bounded domains whose boundary is locally a graph of a Lipschitz function [67],
or, more generally, has a prescribed modulus of continuity [23, 56]; Euclidean John domains, and even
s-John domains; the space Rn equipped with the Gauss measure [13], or with product probability
measures which generalize it [4, 5]. The literature on isoperimetric inequalities is very rich. Let us
limit ourselves to mentioning that, besides those quoted above, recent contributions on isoperimetric
problems in (domains in) Rn endowed with a measure ν include [18, 36, 43, 76].
Given a Banach function space X(Ω, ν) of measurable functions on Ω, and a positive integer m ∈ N,
the m-th order Sobolev type space built upon X(Ω, ν) is the normed linear space V mX(Ω, ν) of all
functions on Ω whose m-th order weak derivatives belong to X(Ω, ν), equipped with a natural norm
induced by X(Ω, ν).
A Sobolev embedding amounts to the boundedness of the identity operator from the Sobolev space
V mX(Ω, ν) into another function space Y (Ω, ν) and will be denoted by
(2.2)
V mX(Ω, ν) → Y (Ω, ν).
When m = 1, we refer to (2.2) as a first-order embedding; otherwise, we call it a higher-order embed-
ding.
Necessary and sufficient conditions for the validity of first-order Euclidean Sobolev embeddings with
X(Ω) = L1(Ω) and Y (Ω) = Lq(Ω) for some q ≥ 1 can be given through the isoperimetric function
IΩ. Sufficient conditions for first-order Sobolev embeddings when X(Ω) = Lp(Ω) for some p > 1 and
Y (Ω) = Lq(Ω), for some q ≥ 1 can also be provided in terms of IΩ. These results were established in
[65, 66], and are exposed in detail in [67, Section 6.4.3].
More recently, first-order Sobolev embeddings of the general form (2.2) (with m = 1), where X(Ω, ν)
and Y (Ω, ν) are Banach function spaces whose norm depends only on the measure of level sets of
functions, called rearrangement-invariant spaces in the literature, have been shown to follow from
one-dimensional inequalities for suitable Hardy type operators which depend on the isoperimetric
function IΩ,ν, and involve the representation function norms k · kX(0,1) and k · kY (0,1) of X(Ω, ν) and
Y (Ω, ν), respectively.
Although a reverse implication need not hold in very pathological settings (e.g. in Euclidean domains
of Nikod´ym type [67, Remark 6.5.2]), first-order Sobolev inequalities are known to be equivalent to the
associated one-dimensional Hardy inequalities in most situations of interest in applications. This is the
case, for instance, in the basic case when Ω is a regular Euclidean domain -- specifically, a John domain
in Rn, n ≥ 2 (see Section 6 for a definition). The class of John domains includes other more classical
families of domains, such as Lipschitz domains, and domains with the cone property. The John domains
arise in connection with the study of holomorphic dynamical systems and quasiconformal mappings.
John domains are known to support a first-order Sobolev inequality with the same exponents as in
4
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
the standard Sobolev inequality [12, 47, 52]. In fact, being a John domain is a necessary condition for
such a Sobolev inequality to hold in the class of two-dimensional simply connected open sets, and in
quite general classes of higher dimensional domains [15]. The isoperimetric function IΩ of any John
domain is known to satisfy
(2.3)
near 0, where n′ = n
n−1 . Here, and in what follows, the notation ≈ means that the two sides are
bounded by each other up to multiplicative constants independent of appropriate quantities. For
instance, in (2.3) such constants depend only on Ω.
IΩ(s) ≈ s
1
n′
As a consequence of (2.3), one can show that the first-order Sobolev embedding
holds if and only if the Hardy type inequality
(2.4)
(2.5)
V 1X(Ω) → Y (Ω)
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
t
f (s)s−1+ 1
n ds(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1) ≤ C kfkX(0,1)
holds for some constant C, and for every nonnegative f ∈ X(0, 1). Results of this kind, showing that
Sobolev embeddings follow from (and are possibly equivalent to) one-dimensional inequalities will be
referred to as reduction principles or reduction theorems. The equivalence of (2.4) and (2.5) is a key
tool in determining the optimal target Y (Ω) for V 1X(Ω) in (2.4) within families of rearrangement-
invariant function spaces, such as Lebesgue, Lorentz, Orlicz spaces, provided that such an optimal
target does exist [24, 26, 39]. An even more standard version of this reduction result, which holds
for functions vanishing on ∂Ω, and is called P´olya-Szego symmetrization principle, is a crucial step in
exhibiting the sharp constant in the classical Sobolev inequalities to which we alluded above [3, 14,
69, 85].
A version of this picture for higher-order Sobolev inequalities is exhibited in the present paper. We
show that any m-th order Sobolev embedding involving arbitrary rearrangement-invariant norms can
be reduced to a suitable one-dimensional inequalities for an integral operator, with a kernel depending
on IΩ,ν and m.
Just to give an idea of the conclusions which follow from our results, let us mention that, if, for
instance, Ω is an Euclidean John domain in Rn, n ≥ 2, then a full higher-order analogue of the
equivalence of (2.4) and (2.5) holds. Namely, the m-th order Sobolev embedding
holds if and only if the Hardy type inequality
(2.6)
V mX(Ω) → Y (Ω)
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
t
f (s)s−1+ m
n ds(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1) ≤ C kfkX(0,1)
holds for some constant C, and for every nonnegative f ∈ X(0, 1) (Theorem 6.1, Section 6).
Our approach to reduction principles for higher-order Sobolev embedding relies on the iteration of
first-order results. Loosely speaking, iteration is understood in the sense that, given a rearrangement-
invariant space and m ∈ N, a first order optimal Sobolev embedding is applied to show that the (m−1)-
th order derivatives of functions from the relevant Sobolev space belong to a suitable rearrangement-
invariant space. Another first-order optimal Sobolev embedding is then applied to show that the
(m − 2)-th order derivatives belong to another rearrangement-invariant space, and so on. Eventually,
m optimal first-order Sobolev embeddings are exploited to deduce that the functions themselves belong
to a certain space.
Let us warn that, although this strategy is quite natural in principle, its implementation is not
straightforward.
Indeed, even in the basic setting when Ω is an Euclidean domain with a smooth
boundary, and standard families of norms are considered, iteration of optimal first-order embeddings
need not lead to optimal higher-order counterparts.
To see this, recall, for instance, that, if Ω is a regular domain in R2, then
HIGHER-ORDER EMBEDDINGS
5
(2.7)
V 2L1(Ω) → L∞(Ω).
On the other hand, iterating twice the classical first-order Sobolev embedding only tells us that
(2.8)
for every q < ∞, and neither of the iterated embeddings can be improved in the framework of Lebesgue
space. This shows that subsequent applications of optimal first-order Sobolev embeddings in the class
of Lebesgue spaces do not necessarily yield optimal higher-order counterparts.
V 2L1(Ω) → V 1L2(Ω) → Lq(Ω)
One might relate the loss of optimality in the chain of embeddings (2.8) to the lack of an optimal
Lebesgue target space for the first-order Sobolev embedding of V 1L2(Ω) when n = 2. However, non-
optimal targets may appear after iteration even in situations where optimal first-order target spaces
do exist. Consider, for example, Euclidean Sobolev embeddings involving Orlicz spaces. The optimal
target in Sobolev embeddings of any order always exists in this class of spaces, and can be explicitly
determined [24, 28]. In particular, Orlicz spaces naturally arise in the borderline case of the Sobolev
embedding theorem. Indeed, if Ω is a regular domain in Rn and 1 ≤ m < n, then
(2.9)
[88, 75, 83]; see also [87] for m = 1. Here, exp Lα(Ω), with α > 0, denotes the Orlicz space associated
with the Young function given by etα
− 1 for t ≥ 0. Observe that the target space in (2.9) is actually
optimal in the class of all Orlicz spaces [24, 26]. Assume, for example, that n ≥ 3 and m = 2. Then
(2.9) reduces to
m (Ω) → exp L
n−m (Ω)
V mL
n
n
Via the iteration of optimal first-order embeddings, one gets
n
V 2L
2 (Ω) → exp L
n
n−2 (Ω).
n
V 2L
2 (Ω) → V 1Ln(Ω) → exp L
n
n−1 (Ω) % exp L
n
n−2 (Ω).
Thus, subsequent applications of optimal Sobolev embeddings even in the class of Orlicz spaces, where
optimal target spaces always exist, need not result in optimal higher-order Sobolev embeddings.
The underlying idea behind the method that we shall introduce is that such a loss of optimality of
the target space under iteration does not occur, provided that first-order (in fact, any-order) Sobolev
embeddings whose targets are optimal among all rearrangement-invariant spaces are iterated. We thus
proceed via a two-step argument, which can be outlined as follows. Firstly, given any function norm
k · kX(0,1) and the isoperimetric function IΩ,ν of (Ω, ν), the optimal target among all rearrangement-
invariant function norms for the first-order Sobolev space V 1X(Ω, ν) is characterized; secondly, first-
order Sobolev embeddings with an optimal target are iterated to derive optimal targets in arbitrary-
order Sobolev embeddings.
In order to grasp this procedure in a simple situation, observe that, when applied in the proof of
embedding (2.7), it amounts to strengthening the chain in (2.8) by
V 2L1(Ω) → V 1L2,1(Ω) → L∞(Ω),
(2.10)
where L2,1(Ω) denotes a Lorentz space (strictly contained in L2(Ω)). We refer to [72, 73, 51] for
standard Sobolev embeddings in Lorentz spaces. Note that both targets in the embeddings in (2.10)
are actually optimal among all rearrangement-invariant spaces.
As mentioned above, our reduction principle asserts that the Sobolev embedding (2.2) follows from
a suitable one-dimensional inequality for an integral operator depending on IΩ,ν, m, k · kX(0,1) and
k·kY (0,1). Interestingly, in contrast with the first-order case, the relevant integral operator is not just of
Hardy type, but involves a genuine kernel. The latter takes back the form of a basic (weighted) Hardy
operator only if, loosely speaking, the isoperimetric function IΩ,ν(s) does not decay too fast to 0 when
s tends to 0. This is the case, for instance, of (2.6). A major consequence of the reduction principle is
a characterization of a target space Y (Ω, ν) in embedding (2.2), depending on X(Ω, ν), m, and IΩ,ν,
which turns to be optimal among all rearrangement-invariant spaces whenever Sobolev embeddings
6
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
and associated one-dimensional inequalities in the reduction principle are actually equivalent. This
latter property depends on the geometry of (Ω, ν), and is fulfilled in most customary situations, to
some of which a consistent part of this paper is devoted.
Besides regular Euclidean domains, namely the John domains which we have already briefly dis-
cussed, the implementations of our results that will be presented concern Maz'ya classes of (possibly
irregular) Euclidean domains, and product probability spaces, of which the Gauss space and the
Boltzmann spaces are distinguished instances.
The Maz'ya classes are defined as families of domains whose isoperimetric function is bounded from
below by some fixed power. Sobolev embeddings in all domains from a class of this type take the same
form, and a worst, in a sense, domain from the relevant class can be singled out to demonstrate the
sharpness of the results.
The product probability spaces in Rn that are taken into account were analyzed in [4, 5], and
share common features with the Gauss space, namely Rn endowed with the probability measure
dγn(x) = (2π)− n
2 dx. In particular, the Boltzmann spaces can be handled via our approach.
2 e− x2
3. Spaces of measurable functions
In this section, we briefly recall some basic facts from the theory of rearrangement-invariant spaces.
For more details, a standard reference is [8].
Let (Ω, ν) be as in Section 2. Recall that we are assuming ν(Ω) = 1. The measure of any measurable
set E ⊂ Ω is thus given by
ν(E) =ZE
ω(x) dx.
We denote by M(Ω, ν) the set of all Lebesgue measurable (and hence ν-measurable) functions on Ω
whose values belong to [−∞,∞]. We also define M+(Ω, ν) = {u ∈ M(Ω, ν) : u ≥ 0}, and M0(Ω, ν) =
{u ∈ M(Ω, ν) : u is finite a.e. in Ω}.
The decreasing rearrangement u∗ : [0, 1] → [0,∞] of a function u ∈ M(Ω, ν) is defined as
u∗(s) = sup{t ∈ R : ν ({x ∈ Ω : u(x) > t}) > s}
for s ∈ [0, 1].
The operation u 7→ u∗ is monotone in the sense that
u ≤ v a.e. in Ω implies u∗ ≤ v∗ in (0, 1).
We also define u∗∗ : (0, 1] → [0,∞] as
u∗∗(s) =
for s ∈ (0, 1].
Note that u∗∗ is also non-increasing, and u∗ ≤ u∗∗ in (0, 1]. Moreover,
(3.1)
sZ s
u∗(r) dr +Z s
(u + v)∗(r) dr ≤Z s
Z s
u∗(r) dr
v∗(r) dr
0
1
0
0
0
for s ∈ (0, 1),
We say that a functional k· kX(0,1) : M+(0, 1) → [0,∞] is a function norm, if, for all f , g and {fj}j∈N
in M+(0, 1), and every λ ≥ 0, the following properties hold:
A basic property of rearrangements is the Hardy-Littlewood inequality, which tells us that, if u, v ∈
for every u, v ∈ M+(Ω, ν).
M(Ω, ν), then
(3.2)
A special case of (3.2) states that for every u ∈ M(Ω, ν) and every measurable set E ⊂ Ω,
ZΩ u(x)v(x)dν(x) ≤Z 1
ZE u(x)dν(x) ≤Z ν(E)
0
0
u∗(s)v∗(s)ds.
u∗(s) ds.
HIGHER-ORDER EMBEDDINGS
7
(P1)
(P2)
(P3)
(P4)
(P5)
(P6)
kfkX(0,1) = 0 if and only if f = 0; kλfkX(0,1) = λkfkX(0,1);
kf + gkX(0,1) ≤ kfkX(0,1) + kgkX(0,1);
f ≤ g a.e. implies kfkX(0,1) ≤ kgkX(0,1);
fj ր f a.e. implies kfjkX(0,1) ր kfkX(0,1);
k1kX(0,1) < ∞;
0 f (x) dx ≤ CkfkX(0,1) for some constant C independent of f .
R 1
kfkX(0,1) = kgkX(0,1) whenever f ∗ = g∗,
If, in addition,
we say that k · kX(0,1) is a rearrangement-invariant function norm.
M+(0, 1), denoted by k · kX ′(0,1), and defined, for g ∈ M+(0, 1), as
With any rearrangement-invariant function norm k · kX(0,1), it is associated another functional on
kgkX ′(0,1) =
kf kX(0,1)≤1Z 1
sup
f ≥0
0
f (s)g(s) ds.
It turns out that k·kX ′(0,1) is also a rearrangement-invariant function norm, which is called the associate
function norm of k · kX(0,1). Moreover, for every function norm k · kX(0,1) and every function f ∈
M+(0, 1), we have
(3.3)
kfkX(0,1) =
f (s)g(s) ds.
kgkX′(0,1)≤1Z 1
sup
g≥0
0
We also introduce yet another functional on M+(0, 1), denoted by k · kX ′
M+(0, 1), as
d(0,1), and defined, for g ∈
kgkX ′
d(0,1) =
kf kX(0,1)≤1Z 1
sup
0
f ∗(t)g(t) dt.
Clearly, one has that kgkX ′
non-increasing.
d(0,1) = kgkX ′(0,1) if g is
Given a rearrangement-invariant function norm k · kX(0,1), the space X(Ω, ν) is defined as the
d(0,1) ≤ kgkX ′(0,1) for every g ∈ M+(0, 1), and kgkX ′
collection of all functions u ∈ M(Ω, ν) such that the expression
kukX(Ω,ν) = ku∗kX(0,1)
is finite. Such expression defines a norm on X(Ω, ν), and the latter is a Banach space endowed with this
norm, called a rearrangement-invariant space. Moreover, X(Ω, ν) ⊂ M0(Ω, ν) for any rearrangement-
invariant space X(Ω, ν). The space X(0, 1) is called the representation space of X(Ω, ν).
We also denote by Xloc(Ω, ν) the space of all functions u ∈ M(Ω, ν) such that uχG ∈ X(Ω, ν) for
every compact set G ⊂ Ω. Here, χG denotes the characteristic function of G.
The rearrangement-invariant space X ′(Ω, ν) built upon the function norm k · kX ′(0,1) is called the
associate space of X(Ω, ν). It turns out that X ′′(Ω, ν) = X(Ω, ν). Furthermore, the Holder inequality
ZΩ u(x)v(x) dν(x) ≤ kukX(Ω,ν)kvkX ′(Ω,ν)
holds for every u ∈ X(Ω, ν) and v ∈ X ′(Ω, ν).
For any rearrangement-invariant spaces X(Ω, ν) and Y (Ω, ν), we have that
if and only if Y ′(Ω, ν) → X ′(Ω, ν),
X(Ω, ν) → Y (Ω, ν)
(3.4)
with the same embedding norms [8, Chapter 1, Proposition 2.10].
8
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Given any λ > 0, the dilation operator Eλ, defined at f ∈ M(0, 1) by
(Eλf )(s) =(f (λ−1s)
0
if 0 < s ≤ λ
if λ < s < 1,
is bounded on any rearrangement-invariant space X(0, 1), with norm not exceeding max{1, 1
λ}.
Hardy's Lemma tells us that if f1, f2 ∈ M+(0, 1) satisfy
Z s
0
0
f1(r)dr ≤Z s
Z 1
u∗(r) dr ≤Z s
0
0
Z s
0
f2(r)dr
for every s ∈ (0, 1),
f1(r)h(r)dr ≤Z 1
0
f2(r)h(r)dr
v∗(r) dr
for s ∈ (0, 1),
for every non-increasing function h : (0, 1) → [0,∞]. A consequence of this result is the Hardy --
Littlewood -- P´olya principle which asserts that if the functions u, v ∈ M(Ω, ν) satisfy
then
then
(3.7)
for every rearrangement-invariant space X(Ω, ν).
kukX(Ω,ν) ≤ kvkX(Ω,ν)
Let X(Ω, ν) and Y (Ω, ν) be rearrangement invariant spaces. By [8, Chapter 1, Theorem 1.8],
X(Ω, ν) ⊂ Y (Ω, ν)
if and only if
X(Ω, ν) → Y (Ω, ν).
For every rearrangement-invariant space X(Ω, ν), one has that
L∞(Ω, ν) → X(Ω, ν) → L1(Ω, ν).
(3.5)
An embedding of the form
where µ is a measure enjoying the same properties as ν, means that, for every compact set G ⊂ Ω,
there exists a constant C such that
Xloc(Ω, ν) → Yloc(Ω, µ),
kuχGkY (Ω,µ) ≤ CkuχGkX(Ω,ν),
for every u ∈ Xloc(Ω, ν).
Throughout, we use the convention that 1
A basic example of a function norm is the standard Lebesgue norm k · kLp(0,1), for p ∈ [1,∞], upon
The Lorentz spaces yield an extension of the Lebesgue spaces. Assume that 1 ≤ p, q ≤ ∞. We
which the Lebesgue spaces Lp(Ω, ν) are built.
∞ = 0, and 0 · ∞ = 0.
define the functionals k · kLp,q(0,1) and k · kL(p,q)(0,1) as
1
p − 1
kfkLp,q(0,1) =(cid:13)(cid:13)(cid:13)s
respectively, for f ∈ M+(0, 1). One can show that
(3.6)
Lp,q(Ω, ν) = L(p,q)(Ω, ν)
with equivalent norms. If one of the conditions
and kfkL(p,q)(0,1) =(cid:13)(cid:13)(cid:13)s
1
p − 1
q f ∗∗(s)(cid:13)(cid:13)(cid:13)Lq(0,1)
,
if 1 < p ≤ ∞ ,
q f ∗(s)(cid:13)(cid:13)(cid:13)Lq(0,1)
1 < p < ∞, 1 ≤ q ≤ ∞,
p = q = 1,
p = q = ∞,
is satisfied, then k · kLp,q(0,1) is equivalent to a rearrangement-invariant function norm. The corre-
sponding rearrangement-invariant space Lp,q(Ω, ν) is called a Lorentz space.
HIGHER-ORDER EMBEDDINGS
9
1
p − 1
q logα(cid:0) 2
s(cid:1) f ∗(s)(cid:13)(cid:13)(cid:13)Lq(0,1)
kfkLp,q;α,β(0,1) =(cid:13)(cid:13)(cid:13)s
p − 1
1
Let us recall that Lp,p(Ω, ν) = Lp(Ω, ν) for every p ∈ [1,∞] and that 1 ≤ q ≤ r ≤ ∞ implies
Assume now that 1 ≤ p, q ≤ ∞, and a third parameter α ∈ R is called into play. We define the
Lp,q(Ω, ν) → Lp,r(Ω, ν) with equality if and only if q = r.
functionals k · kLp,q;α(0,1) and k · kL(p,q;α)(0,1) as
kfkLp,q;α(0,1) =(cid:13)(cid:13)(cid:13)s
and kfkL(p,q;α)(0,1) =(cid:13)(cid:13)(cid:13)s
1
p − 1
q logα(cid:0) 2
s(cid:1) f ∗∗(s)(cid:13)(cid:13)(cid:13)Lq(0,1)
,
respectively, for f ∈ M+(0, 1). If one of the following conditions
1 < p < ∞, 1 ≤ q ≤ ∞, α ∈ R;
p = 1, q = 1, α ≥ 0;
p = ∞, q = ∞, α ≤ 0;
p = ∞, 1 ≤ q < ∞, α + 1
q < 0,
(3.8)
is satisfied, then k · kLp,q;α(0,1) is equivalent to a rearrangement-invariant function norm, called a
Lorentz -- Zygmund function norm. The corresponding rearrangement-invariant space Lp,q;α(Ω, ν) is
a Lorentz -- Zygmund space. At a few occasions, we shall need also the so-called generalized Lorentz --
Zygmund space Lp,q;α,β(Ω, ν), where p, q ∈ [1,∞] and α, β ∈ R. It is the space built upon the functional
given by
q logα(cid:0) 2
s(cid:1) logβ(1 + log(cid:0) 2
s(cid:1))f ∗(s)(cid:13)(cid:13)(cid:13)Lq(0,1)
for f ∈ M+(0, 1). The values of p, q, α and β, for which k · kLp,q;α,β(0,1) is actually equivalent to
a rearrangement-invariant function norm, are characterized in [41]. For more details on (generalized)
Lorentz -- Zygmund spaces, see e.g. [7, 41, 71]. Assume that one of the conditions in (3.8) is satisfied.
Then the associate space (Lp,q;α)′(Ω, ν) of the Lorentz -- Zygmund space Lp,q;α(Ω, ν) satisfies (up to
equivalent norms)
(cid:0)Lp,q;α(cid:1)′(Ω, ν) =
Lp′,q′;−α(Ω, ν)
L∞,∞;−α(Ω, ν)
L1,1;−α(Ω, ν)
L(1,q′;−α−1)(Ω, ν)
if 1 < p < ∞, 1 ≤ q ≤ ∞, α ∈ R;
if p = 1, q = 1, α ≥ 0;
if p = ∞, q = ∞, α ≤ 0;
if p = ∞, 1 ≤ q < ∞, α + 1
q < 0
[71, Theorems 6.11 and 6.12]. Moreover,
(3.9)
(3.10)
and
(3.11)
L(p,q;α)(Ω, ν) =(Lp,q;α(Ω, ν)
Lp(Ω, ν) → L(1,q)(Ω, ν)
L1,1;α+1(Ω, ν)
if 1 < p ≤ ∞;
if p = q = 1, α > −1,
for every 1 < p ≤ ∞, 1 ≤ q ≤ ∞
[71, Theorem 3.16 (i),(ii)].
A generalization of the Lebesgue spaces in a different direction is provided by the Orlicz spaces.
Let A : [0,∞) → [0,∞] be a Young function, namely a convex (non trivial), left-continuous function
vanishing at 0. Any such function takes the form
A(t) =Z t
0
a(τ )dτ
for t ≥ 0,
for some non-decreasing, left-continuous function a : [0,∞) → [0,∞] which is neither identically equal
to 0, nor to ∞. The Orlicz space LA(Ω, ν) is the rearrangement-invariant space associated with the
Luxemburg function norm defined as
kfkLA(0,1) = inf(cid:26)λ > 0 :Z 1
0
A(cid:18) f (s)
λ (cid:19) ds ≤ 1(cid:27)
10
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
for f ∈ M+(0, 1). In particular, LA(Ω, ν) = Lp(Ω, ν) if A(t) = tp for some p ∈ [1,∞), and LA(Ω, ν) =
L∞(Ω, ν) if A(t) = ∞χ(1,∞)(t).
A Young function A is said to dominate another Young function B near infinity if positive constants
c and t0 exist such that
The functions A and B are called equivalent near infinity if they dominate each other near infinity.
One has that
B(t) ≤ A(ct)
for
t ≥ t0 .
(3.12)
LA(Ω, ν) → LB(Ω, ν)
if and only if A dominates B near infinity .
We denote by Lp logα L(Ω, ν) the Orlicz space associated with a Young function equivalent to
tp(log t)α near infinity, where either p > 1 and α ∈ R, or p = 1 and α ≥ 0. The notation exp Lβ(Ω, ν)
will be used for the Orlicz space built upon a Young function equivalent to etβ
near infinity, where
β > 0. Also, exp exp Lβ(Ω, ν) stands for the Orlicz space associated with a Young function equivalent
to eetβ
near infinity.
The classes of Orlicz and (generalized) Lorentz-Zygmund spaces overlap, up to equivalent norms.
For instance, if 1 ≤ p < ∞ and α ∈ R, then
Lp,p;α(Ω, ν) = Lp(log L)pα(Ω, ν).
Moreover, if β > 0, then
and [41, Lemma 2.2]
L∞,∞;−β(Ω, ν) = exp L
1
β (Ω, ν)
L∞,∞;0,−β(Ω, ν) = exp exp L
1
β (Ω, ν).
A common extension of the Orlicz and Lorentz spaces is provided by a family of Orlicz-Lorentz
spaces defined as follows. Given p ∈ (1,∞), q ∈ [1,∞) and a Young function D such that
t1+p dt < ∞ ,
Z ∞ D(t)
kfkL(p,q,D)(0,1) =(cid:13)(cid:13)(cid:13)s− 1
p f ∗(s
.
1
q )(cid:13)(cid:13)(cid:13)LD(0,1)
we denote by L(p, q, D)(Ω, ν) the Orlicz-Lorentz space associated with the rearrangement-invariant
function norm defined, for f ∈ M+(0, 1), as
The fact that k · kL(p,q,D)(0,1) is actually a function norm follows via easy modifications in the proof
of [26, Proposition 2.1]. Observe that the class of the spaces L(p, q, D)(Ω, ν) actually includes (up to
equivalent norms) the Orlicz spaces and various instances of Lorentz and Lorentz-Zygmund spaces.
4. Spaces of Sobolev type and the isoperimetric function
Let (Ω, ν) be as in Section 2. Define the perimeter of a measurable set E in (Ω, ν) as
Pν(E, Ω) =ZΩ∩∂M E
ω(x)dHn−1(x),
where ∂M E denotes the essential boundary of E, in the sense of geometric measure theory [67, 90].
The isoperimetric function IΩ,ν : [0, 1] → [0,∞] of (Ω, ν) is then given by
2(cid:9)
IΩ,ν(s) = inf(cid:8)Pν(E, Ω) : E ⊂ Ω, s ≤ ν(E) ≤ 1
if s ∈ [0, 1
2 ],
and IΩ,ν(s) = IΩ,ν(1 − s) if s ∈ ( 1
consequence of this definition and of the fact that Pν (E, Ω) = Pν(Ω \ E, Ω) for every set E ⊂ Ω.
2 , 1]. The isoperimetric inequality (2.1) in (Ω, ν) is a straightforward
HIGHER-ORDER EMBEDDINGS
11
Let us observe that, actually, IΩ,ν(s) < ∞ for s ∈ [0, 1
2 ). To verify this fact, fix any x0 ∈ Ω, and let
2 . Here, BR(x0) denotes the ball, centered at x0, with radius R.
R > 0 be such that ν(Ω ∩ BR(x0)) = 1
By the polar-coordinates formula for integrals,
ω(x) dx =Z R
2 =ZΩ∩BR(x0)
0 ZΩ∩∂Bρ(x0)
(4.1)
1
ω(x) dHn−1(x) dρ =Z R
0
Pν(Ω ∩ Bρ(x0), Ω) dρ ,
whence Pν (Ω ∩ Bρ(x0), Ω) < ∞ for a.e. ρ ∈ (0, R). The finiteness of IΩ,ν in [0, 1
very definition.
2 ) now follows by its
The next result shows that the best possible behavior of an isoperimetric function at 0 is that given
1
n′ as s → 0, whatever (Ω, ν) is.
by (2.3), in the sense that IΩ,ν(s) cannot decay more slowly than s
Proposition 4.1. There exists a positive constant C = C(Ω, ν) such that
(4.2)
IΩ,ν(s) ≤ Cs
1
n′
near 0.
Proof. Let x0 be any Lebesgue point of ω, namely a point such that
(4.3)
lim
r→0+
1
Br(x0)ZBr(x0)
ω(x) dx
exists and is finite. Here, E denotes the Lebesgue measure of a set E ⊂ Rn. By (4.3), there exists
r0 > 0 and C > 0 such that
ω(x) dx ≤ Crn
if 0 < r < r0.
(4.4)
By an analogous chain as in (4.1),
ZBr(x0)
ω(x) dx =Z r
ZBr(x0)
n′ ≥ inf{Pν (Bρ(x0), Ω) : r
= inf{Pν (Bρ(x0), Ω) : 1
Thus, there exists a constant C such that
CBr(x0)
0
1
(4.5)
2 ≤ ρ ≤ r}
if 0 < r < r0. From (4.4) and (4.5) we deduce that there exists a constant C such that
2 inf{Pν (Bρ(x0), Ω) : r
Pν (Bρ(x0), Ω) dρ ≥ r
2 ≤ ρ ≤ r}
2nBr(x0) ≤ Bρ(x0) ≤ Br(x0)}
if 0 < r < r0.
(cid:3)
provided that s is sufficiently small, and hence (4.2) follows.
1
Cs
n′ ≥ inf{Pν (E, Ω) : s ≤ E ≤ 1
2},
Let m ∈ N and let X(Ω, ν) be a rearrangement-invariant space. We define the m-th order Sobolev
space V mX(Ω, ν) as
V mX(Ω, ν) =(cid:8)u : u is m-times weakly differentiable in Ω, and ∇mu ∈ X(Ω, ν)(cid:9).
Here, ∇mu denotes the vector of all m-th order weak derivatives of u. We shall also denote ∇0u = u.
Let us notice that in the definition of V mX(Ω, ν) it is only required that the derivatives of the highest
order m of u belong to X(Ω, ν). This assumption does not entail, in general, that also u and its
derivatives up to the order m − 1 belong to X(Ω, ν), and even to L1(Ω, ν). Thus, it may happen that
V mX(Ω, ν) * V kX(Ω, ν) for m > k. Such inclusion indeed fails, for instance, when (Ω, ν) = (Rn, γn),
the Gauss space, and k · kX(0,1) = k · kL∞(0,1) (or k · kX(0,1) = k · kexpLβ(0,1) for some β > 0). Examples
of Euclidean domains for which V mX(Ω) * L1(Ω) are those of Nykod´ym type, see, e.g., [67, Sections
5.2 and 5.4].
However, if IΩ,ν(s) does not decay at 0 faster than linearly, namely if there exists a positive constant
C such that
(4.6)
IΩ,ν(s) ≥ Cs
for s ∈ [0, 1
2 ],
12
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
then any function u ∈ V mX(Ω, ν) does at least belong to L1(Ω, ν), together with all its derivatives
up to the order m − 1. This is a consequence of the next result. Such result in the case when ν is
the Lebesgue measure is established in [67, Theorem 5.2.3]; the general case rests upon an analogous
argument. We provide a proof for completeness.
Proposition 4.2. [Condition for V 1L1(Ω, ν) ⊂ L1(Ω, ν)] Assume that (4.6) holds. Then V 1L1(Ω, ν) ⊂
L1(Ω, ν), and
(4.7)
C
2(cid:13)(cid:13)(cid:13)(cid:13)u −ZΩ
u dν(cid:13)(cid:13)(cid:13)(cid:13)L1(Ω,ν) ≤ k∇ukL1(Ω,ν)
for every u ∈ V 1L1(Ω, ν), where C is the same constant as in (4.6).
Proof. Let med(u) denote the median of a function u ∈ M(Ω, ν), given by
med(u) = sup{t ∈ R : ν({x ∈ Ω : u(x) > t}) > 1
2}.
We begin by showing that
(4.8)
for every u ∈ V 1L1(Ω, ν). On replacing, if necessary, u by u − med(u), we may assume, without loss
of generality, that med(u) = 0. Let us set u+ = 1
2 (u − u), the positive and the
negative parts of u, respectively. Thus,
Cku − med(u)kL1(Ω,ν) ≤ k∇ukL1(Ω,ν)
2 (u + u) and u− = 1
(4.9)
By (2.1) and (4.6),
ν({u± > t}) ≤ 1
2
for t > 0.
Therefore, owing to (4.9), and to the coarea formula, we have that
Pν ({u± > t}, Ω) ≥ Iν,Ω(ν({u± > t})) ≥ Cν({u± > t}).
Hence, (4.8) follows. In particular, (4.8) tells us that V 1L1(Ω, ν) ⊂ L1(Ω, ν). Inequality (4.7) is a
consequence of (4.8) and of the fact that
Pν ({u± > t}, Ω) dt
ω(x)dHn−1(x) dt =ZΩ ∇u±dν.
0
0
ν({u± > t}) dt ≤Z ∞
Cku±kL1(Ω,ν) = CZ ∞
=Z ∞
0 Z∂M {u±>t}∩Ω
(cid:13)(cid:13)(cid:13)(cid:13)u −ZΩ
u dν(cid:13)(cid:13)(cid:13)(cid:13)L1(Ω,ν) ≤ 2ku − med(u)kL1(Ω,ν)
for every u ∈ L1(Ω, ν).
Corollary 4.3. Assume that (4.6) holds. Let m ≥ 1. Let X(Ω, ν) be any rearrangement-invariant
space. Then V mX(Ω, ν) ⊂ V kL1(Ω, ν) for every k = 0, . . . , m − 1.
Proof. By property (P5) of rearrangement-invariant spaces, V mX(Ω, ν) → V mL1(Ω, ν). Thus, the
conclusion follows from an iterated use of Proposition 4.2.
(cid:3)
(cid:3)
Under (4.6), an assumption which will always be kept in force hereafter, V mX(Ω, ν) is easily seen
to be a normed linear space, equipped with the norm
kukV mX(Ω,ν) =
k∇kukL1(Ω,ν) + k∇mukX(Ω,ν).
m−1Xk=0
Standard arguments show that V mX(Ω, ν) is complete, and hence a Banach space, under the additional
assumption that
L1
loc(Ω, ν) → L1
loc(Ω).
HIGHER-ORDER EMBEDDINGS
13
We also define the subspace V m
⊥ X(Ω, ν) of V mX(Ω, ν) as
V m
⊥ X(Ω, ν) =(cid:26)u ∈ V mX(Ω, ν) :ZΩ ∇ku dν = 0, for k = 0, . . . , m − 1(cid:27).
The Sobolev embedding (2.2) turns out to be equivalent to a Poincar´e type inequality for functions
in V m
⊥ X(Ω, ν).
Proposition 4.4. [Equivalence of Sobolev and Poincar´e inequalities] Assume that (Ω, ν) fulfils
(4.6) and that m ≥ 1. Let k · kX(0,1) and k · kY (0,1) be rearrangement-invariant function norms. Then
(4.10)
V mX(Ω, ν) → Y (Ω, ν)
if and only if there exists a constant C such that
(4.11)
kukY (Ω,ν) ≤ Ck∇mukX(Ω,ν)
for every u ∈ V m
Proof. Assume that (4.10) holds. Thus, there exists a constant C such that
⊥ X(Ω, ν).
(4.12)
kukY (Ω,ν) ≤ C(cid:16) m−1Xk=0
k∇kukL1(Ω,ν) + k∇mukX(Ω,ν)(cid:17)
for every u ∈ V mX(Ω, ν). Iterating inequality (4.7) implies that there exist constants C1, . . . , Cm such
that
(4.13)
kukL1(Ω,ν) ≤ C1k∇ukL1(Ω,ν) ≤ C2k∇2ukL1(Ω,ν) ≤ ··· ≤ Cmk∇mukL1(Ω,ν)
for every u ∈ V m
⊥ X(Ω, ν). By property (P5) of rearrangement-invariant function norms, there exists
a constant C, independent of u, such that k∇mukL1(Ω,ν) ≤ Ck∇mukX(Ω,ν). Thus, (4.11) follows from
(4.12) and (4.13).
Suppose next that (4.11) holds. Given k ∈ N, denote by P k the space of polynomials whose
degree does not exceed k. Observe that P k ⊂ L1(Ω, ν) for every k ∈ N.
Indeed, ∇hP = 0 for
every P ∈ P k, provided that h > k, and hence P k ⊂ V hX(Ω, ν) for any rearrangement-invariant
space X(Ω, ν). The inclusion P k ⊂ L1(Ω, ν) thus follows via Corollary 4.3. Next, it is not difficult
to verify that, for each u ∈ V mX(Ω, ν), there exists a (unique) polynomial Pu ∈ P m−1 such that
u − Pu ∈ V m
⊥ X(Ω, ν). Moreover, the coefficients of Pu are linear combinations of the components of
RΩ ∇ku dν, for k = 0, . . . , m − 1, with coefficients depending on n, m and (Ω, ν). Now, we claim that
(4.14)
P m ⊂ Y (Ω, ν).
This inclusion is trivial in the case when Ω is bounded, owing to axioms (P2) and (P4) of the definition
of rearrangement-invariant function norms, since any polynomial is bounded in Ω. To verify (4.14) in
i ∈ P m. Let PQ ∈ P m−1 be
the general case, consider, for each i = 1, . . . , n, the polynomial Q(x) = xm
the polynomial associated with Q as above, such that Q− PQ ∈ V m
⊥ X(Ω, ν). Note that the polynomial
PQ also depends only on xi. From (4.11) applied with u = Q − PQ we deduce that Q − PQ ∈ Y (Ω, ν).
This inclusion and the inequality Q − PQ ≥ Cxim, which holds, for a suitable positive constant
C, if xi is sufficiently large, tell us, via axiom (P2) of the definition of rearrangement-invariant
function norms, that xim ∈ Y (Ω, ν) as well. Thus, xm ∈ Y (Ω, ν), and by axiom (P2) again, any
polynomial of degree not exceeding m also belongs to Y (Ω, ν). Hence, (4.14) follows. Thus, given any
14
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
u ∈ V mX(Ω, ν), we have that
kukY (Ω,ν) ≤ ku − PukY (Ω,ν) + kPukY (Ω,ν)
≤ Ck∇mukX(Ω,ν) +
CZΩ ∇ku dν Xα1+···+αn=k
m−1Xk=0
m−1Xk=0
≤ Ck∇mukX(Ω,ν) + C ′
k∇kukL1(Ω,ν),
kx1α1 ··· xnαnkY (Ω,ν)
for some constants C and C ′ independent of u. Hence, embedding (4.10) follows.
(cid:3)
Let us incidentally mention that more customary Sobolev type spaces W mX(Ω, ν) can be defined
as
W mX(Ω, ν) =(cid:8)u : u is m-times weakly differentiable in Ω,
and equipped with the norm
∇ku ∈ X(Ω, ν) for k = 0, . . . , m(cid:9),
The space W mX(Ω, ν) is a normed linear space, and it is a Banach space if
kukW mX(Ω,ν) =
k∇kukX(Ω,ν).
mXk=0
Xloc(Ω, ν) → L1
loc(Ω).
By the second embedding in (3.5),
(4.15)
W mX(Ω, ν) → V mX(Ω, ν)
for every (Ω, ν) fulfilling (4.6), but, in general, W mX(Ω, ν) $ V mX(Ω, ν). For instance, if (Ω, ν) =
(Rn, γn), the Gauss space, and k · kX(0,1) = k · kL∞(0,1) (or k · kX(0,1) = k · kexpLβ (0,1) for some β > 0),
then V mX(Ω, ν) 6= W mX(Ω, ν). However, the spaces W mX(Ω, ν) and V mX(Ω, ν) agree if condition
(4.6) is slightly strengthened to
(4.16)
Z0
ds
IΩ,ν(s)
< ∞.
Note that (4.16) indeed implies (4.6), since
1
IΩ,ν
is a non-increasing function.
Proposition 4.5. [Condition for W mX(Ω, ν) = V mX(Ω, ν)] Let (Ω, ν) be as above, and let m ∈ N.
Assume that (4.16) holds. Let k · kX(0,1) be a rearrangement-invariant function norm. Then
(4.17)
W mX(Ω, ν) = V mX(Ω, ν),
up to equivalent norms.
A proof of this proposition relies upon one of our main results, and can be found at the end of
Section 9.
5. Main results
The present section contains the main results of this paper, which link embeddings and Poincar´e
inequalities for Sobolev-type spaces of arbitrary order to isoperimetric inequalities. The relevant
results depend only on a lower bound for the isoperimetric function IΩ,ν of (Ω, ν) in terms of some
other non-decreasing function I : [0, 1] → [0,∞); precisely, on the existence of a positive constant c
such that
(5.1)
IΩ,ν(s) ≥ cI(cs)
for s ∈ [0, 1
2 ].
HIGHER-ORDER EMBEDDINGS
15
As mentioned in Proposition 4.2 and the preceding remarks, it is reasonable to suppose that the
function IΩ,ν satisfies the estimate (4.6). In the light of this fact, in what follows we shall assume that
(5.2)
inf
t∈(0,1)
I(t)
t
> 0.
Theorem 5.1. [Reduction principle] Assume that (Ω, ν) fulfils (5.1) for some non-decreasing func-
tion I satisfying (5.2). Let m ∈ N, and let k·kX(0,1) and k·kY (0,1) be rearrangement-invariant function
norms. If there exists a constant C1 such that
≤ C1 kfkX(0,1)
(5.3)
Z 1
t
f (s)
I(s)(cid:18)Z s
t
dr
I(r)(cid:19)m−1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
for every nonnegative f ∈ X(0, 1), then
(5.4)
V mX(Ω, ν) → Y (Ω, ν),
⊥ X(Ω, ν).
kukY (Ω,ν) ≤ C2 k∇mukX(Ω,ν)
and there exists a constant C2 such that
(5.5)
for every u ∈ V m
Remark 5.2. It turns out that inequality (5.3) holds for every nonnegative f ∈ X(0, 1) if and only
if it just holds for every nonnegative and non-increasing f ∈ X(0, 1). This fact will be proved in
Corollary 9.8, Section 9, and can be of use in concrete applications of Theorem 5.1.
Indeed, the
available criteria for the validity of one-dimensional inequalities for integral operators take, in general,
different forms according to whether trial functions are arbitrary, or just monotone.
As already stressed in Sections 1 and 2, the first-order case (m = 1) of Theorem 5.1 is already well
known; the novelty here amounts to the higher-order case when m > 1. To be more precise, when
m = 1, a version of Theorem 5.1 in the standard Euclidean case, for functions vanishing on ∂Ω, is
by now classical, and has been exploited in the proof of Sobolev inequalities with sharp constants,
including [3, 69, 85, 14]. An argument showing that (5.3) with m = 1 implies (5.4) and (5.5), for
functions with arbitrary boundary values, for Orlicz norms, on regular Euclidean domains, or, more
generally, on domains in Maz'ya classes, is presented [24, Proof of Theorem 2 and Remark 2]. A
proof for arbitrary rearrangement-invariant norms, in Gauss space, is given in [33]. The same proof
translates verbatim to general measure spaces (Ω, ν) as in Theorem (5.1) -- see e.g. [63].
A major feature of Theorem 5.1 is the difference occurring in (5.3) between the first-order case
(m = 1) and the higher-order case (m > 1). Indeed, the integral operator appearing in (5.3) when
m = 1 is just a weighted Hardy-type operator, namely a primitive of f times a weight, whereas, in
the higher-order case, a genuine kernel, with a more complicated structure, comes into play. In fact,
this seems to be the first known instance where such a kernel operator is needed in a reduction result
for Sobolev-type embeddings. Of course, this makes the proof of inequalities of the form (5.3) more
challenging, although several contributions on one-dimensional inequalities for kernel operators are
fortunately available in the literature (see e.g. the survey papers [55, 64, 82], and the monographs
[37, 38]).
Remark 5.3. As we shall see, the Sobolev embedding (5.4) (or the Poincar´e inequality (5.5)) and
inequality (5.3), in which the function I is equivalent to the isoperimetric function IΩ,ν on some
neighborhood of zero, are actually equivalent in customary families of measure spaces (Ω, ν), and
hence, Theorem 5.4 enables us to determine the optimal rearrangement-invariant target spaces in
Sobolev embeddings for these measure spaces. Incidentally, let us mention that when m = 1, this is
the case whenever the geometry of (Ω, ν) allows the construction of a family of trial functions u in
(5.4) or (5.5) characterized by the following properties: the level sets of u are isoperimetric (or almost
isoperimetric) in (Ω, ν); ∇u is constant (or almost constant) on the boundary of the level sets of u.
16
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
If m > 1, then the latter requirement has to be complemented by requiring that the derivatives of
u up to the order m restricted to the boundary of the level sets satisfy certain conditions depending
on I. The relevant conditions have, however, a technical nature, and it is not worth to state them
explicitly. In fact, heuristically speaking, properties (5.3), (5.5) and (5.4) turn out to be equivalent
for every m ≥ 1 on the same measure spaces (Ω, ν) as for m = 1. Such equivalence certainly holds in
any customary, non-pathological situation, including the three frameworks to which our results will
be applied, namely John domains, Euclidean domains from Maz'ya classes and product probability
spaces in Rn extending the Gauss space.
Now we are in a position to characterize the space which, in the situation discussed in Remark
5.3, is the optimal rearrangement-invariant target space in the Sobolev embedding (5.4). Such an
optimal space is the one associated with the rearrangement-invariant function norm k · kXm,I (0,1),
whose associate norm is defined as
kfkX ′
m,I (0,1) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1
I(t)Z t
0 (cid:18)Z t
s
dr
I(r)(cid:19)m−1
f ∗(s)ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
for f ∈ M+(0, 1).
Theorem 5.4. [Optimal target] Assume that (Ω, ν), m, I and k · kX(0,1) are as in Theorem 5.1.
m,I (0,1), given by (5.6), is a rearrangement-invariant function norm, whose
Then the functional k · kX ′
associate norm k · kXm,I (0,1) satisfies
(5.7)
V mX(Ω, ν) → Xm,I (Ω, ν),
and there exists a constant C such that
(5.8)
for every u ∈ V m
Moreover, if (Ω, ν) is such that (5.4), or equivalently (5.5), implies (5.3), and hence (5.3), (5.4)
and (5.5) are equivalent, then the function norm k · kXm,I (0,1) is optimal in (5.7) and (5.8) among all
rearrangement-invariant norms.
kukXm,I (Ω,ν) ≤ C k∇mukX(Ω,ν)
⊥ X(Ω, ν).
An important special case of Theorems 5.1 and 5.4 is enucleated in the following corollary.
Corollary 5.5. [Sobolev embeddings into L∞] Assume that (Ω, ν), m, I and k · kX(0,1) are as in
Theorem 5.1. If
(5.6)
(5.9)
then
(5.10)
(cid:13)(cid:13)(cid:13)(cid:13)
0
1
dr
I(r)(cid:19)m−1(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
I(s)(cid:18)Z s
V mX(Ω, ν) → L∞(Ω, ν),
< ∞ ,
and there exists a constant C such that
(5.11)
for every u ∈ V m
and (5.5) are equivalent, then (5.9) is necessary for (5.10) or (5.11) to hold.
kukL∞(Ω,ν) ≤ C k∇mukX(Ω,ν)
⊥ X(Ω, ν).
Moreover, if (Ω, ν) is such that (5.4), or equivalently (5.5), implies (5.3), and hence (5.3), (5.4)
Remark 5.6. If (Ω, ν) is such that (5.4), or equivalently (5.5), implies (5.3), and hence (5.3), (5.4)
and (5.5) are equivalent, then (5.10) cannot hold, whatever k · kX(0,1) is, if I decays so fast at 0 that
Z0
dr
I(r)
= ∞.
HIGHER-ORDER EMBEDDINGS
17
Our last main result concerns the preservation of optimality in targets among all rearrangement-
invariant spaces under iteration of Sobolev embeddings of arbitrary order.
Theorem 5.7. [Iteration principle] Assume that (Ω, ν), I and k · kX(0,1) are as in Theorem 5.1.
Let k, h ∈ N. Then
(Xk,I )h,I (Ω, ν) = Xk+h,I(Ω, ν),
up to equivalent norms.
We now focus on the case when
(5.12)
Z s
0
dr
I(r) ≈
s
I(s)
for s ∈ (0, 1).
If the function I satisfies (5.12), then the results of Theorems 5.1, 5.4 and 5.7 can be somewhat
simplified. This is the content of the next three corollaries. Let us preliminarily observe that, since
the right-hand side of (5.12) does not exceed its left-hand side for any non-decreasing function I, only
the estimate in the reverse direction is relevant in (5.12).
Corollary 5.8. [Reduction principle under (5.12)] Let (Ω, ν), m, I, k · kX(0,1) and k · kY (0,1) be
as in Theorem 5.1. Assume, in addition, that I fulfils (5.12). If there exists a constant C1 such that
(5.13)
for every nonnegative f ∈ X(0, 1), then
(5.14)
and there exists a constant C2 such that
(5.15)
for every u ∈ V m
⊥ X(Ω, ν).
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
t
f (s)
sm−1
I(s)m ds(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1) ≤ C1 kfkX(0,1)
V mX(Ω, ν) → Y (Ω, ν),
kukY (Ω,ν) ≤ C2 k∇mukX(Ω,ν)
Let us notice that a remark parallel to Remark 5.2 applies on the equivalence of the validity of
(5.13) for any f , or for any non-increasing f (cf. Proposition 8.6).
The next corollary tells us that, under the extra condition (5.12), the optimal rearrangement-
invariant target space takes a simplified form. Namely, it can be equivalently defined via the rearrangement-
invariant function norm k · kX ♯
m,I (0,1) obeying
(5.16)
kfk(X ♯
m,I )′(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)
tm−1
I(t)mZ t
0
f ∗(s) ds(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
for every f ∈ M+(0, 1).
Corollary 5.9. [Optimal target under (5.12)] Assume that (Ω, ν), m, I and k · kX(0,1) are as
in Corollary 5.8. Then the functional k · k(X ♯
m,I )′(0,1), given by (5.16), is a rearrangement-invariant
function norm, whose associate norm k · kX ♯
(5.17)
m,I (0,1) satisfies
V mX(Ω, ν) → X ♯
m,I (Ω, ν),
and there exists a constant C such that
(5.18)
for every u ∈ V m
hence (5.13), (5.14) and (5.15) are equivalent, then the function norm k·kX ♯
and (5.18) among all rearrangement-invariant norms.
Moreover, if (Ω, ν) is such that the validity of (5.14), or equivalently (5.15), implies (5.13), and
m,I (0,1) is optimal in (5.17)
m,I (Ω,ν) ≤ C k∇mukX(Ω,ν)
⊥ X(Ω, ν).
kukX ♯
18
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
We conclude this section with a stability result for the iterated embeddings under the additional
condition (5.12).
Corollary 5.10. [Iteration principle under (5.12)] Assume that (Ω, ν), I and k · kX(0,1) are as in
Corollary 5.8. Let k, h ∈ N. Then
up to equivalent norms.
h,I(Ω, ν) = X ♯
k+h,I(Ω, ν),
k,I(cid:1)♯
(cid:0)X ♯
6. Euclidean -- Sobolev embeddings
The main results of this section are reduction theorems and their consequences for Euclidean Sobolev
embeddings of arbitrary order m on John domains, and on domains from Maz'ya classes.
We begin with the reduction theorem for John domains. Recall that a bounded open set Ω in Rn
is called a John domain if there exist a constant c ∈ (0, 1) and a point x0 ∈ Ω such that for every
x ∈ Ω there exists a rectifiable curve : [0, l] → Ω, parameterized by arclength, such that (0) = x,
(l) = x0, and
dist ((r), ∂Ω) ≥ cr
for r ∈ [0, l].
Theorem 6.1. [Reduction principle for John domains] Let n ∈ N, n ≥ 2, and let m ∈ N.
Assume that Ω is a John domain in Rn. Let k · kX(0,1) and k · kY (0,1) be rearrangement-invariant
function norms. Then the following assertions are equivalent.
(i) The Hardy type inequality
(6.1)
holds for some constant C1, and for every nonnegative f ∈ X(0, 1).
(ii) The Sobolev embedding
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
t
f (s)s−1+ m
n ds(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1) ≤ C1 kfkX(0,1)
(6.2)
holds.
(iii) The Poincar´e inequality
V mX(Ω) → Y (Ω)
(6.3)
holds for some constant C2 and every u ∈ V m
⊥ X(Ω).
kukY (Ω) ≤ C2 k∇mukX(Ω)
Forerunners of Theorem 6.1 are known. The first order case (m = 1) on Lipschitz domains was
obtained in [39]. In the case when m = 2, and functions vanishing on ∂Ω are considered, the equivalence
of (6.1) and (6.3) was proved in [27], as a consequence of a non-standard rearrangement inequality
for second-order derivatives (see also [25] for a related one-dimensional second-order rearrangement
inequality). The equivalence of (6.1) and (6.2), when m ≤ n − 1 and Ω is a Lipschitz domain, was
established in [51] by a method relying upon interpolation techniques. Such a method does not carry
over to the more general setting of Theorem 6.1, since it requires that Ω be an extension domain.
Let us also warn that results reducing higher-order Sobolev embeddings to one-dimensional in-
equalities can be obtained via more standard methods, such as, for instance, representation formulas
of convolution type combined with O'Neil rearrangement estimates for convolutions, or plain iteration
of certain first-order pointwise rearrangement estimates [61]. However, these approaches lead to opti-
mal Sobolev embeddings only under additional technical assumptions on the involved rearrangement-
invariant function norms k · kX(0,1) and k · kY (0,1).
the rearrangement-invariant function norm, whose associate function norm is given by
Given a rearrangement-invariant function norm k · kX(0,1) and m ∈ N, we define k · kXm,John(0,1) as
(6.4)
kfkX ′
m,John(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)s−1+ m
n Z s
0
f ∗(r)dr(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
HIGHER-ORDER EMBEDDINGS
19
for f ∈ M+(0, 1). The function norm k · kXm,John(0,1) is optimal, as a target, for Sobolev embeddings
of V mX(Ω).
Theorem 6.2. [Optimal target for John domains] Let n, m, Ω and k · kX(0,1) be as in Theorem
m,John(0,1), given by (6.4), is a rearrangement-invariant function norm,
6.1. Then the functional k · kX ′
whose associate norm k · kXm,John(0,1) satisfies
(6.5)
V mX(Ω) → Xm,John(Ω),
and
(6.6)
for some constant C and every u ∈ V m
invariant norms.
kukXm,John(Ω) ≤ C k∇mukX(Ω)
⊥ X(Ω).
Moreover, the function norm k·kXm,John(0,1) is optimal in (6.5) and (6.6) among all rearrangement-
The iteration principle for optimal target norms in Sobolev embeddings on John domains reads as
follows.
Theorem 6.3. [Iteration principle for John domains] Let n ∈ N, Ω and k · kX(0,1) be as in
Theorem 6.1. Let k, h ∈ N. Then
up to equivalent norms.
(cid:0)Xk,John(cid:1)h,John(Ω) = Xk+h,John(Ω),
Let us now focus on Maz'ya classes of domains. Given α ∈ [ 1
class of all Euclidean domains Ω satisfying (5.1), with I(s) = sα for s ∈ [0, 1
Rn such that
IΩ(s) ≥ Csα
for s ∈ [0, 1
2 ],
n′ , 1], we denote by Jα the Maz'ya
2 ], namely domains Ω in
for some positive constant C. Thanks to (2.3), any John domain belongs to the class J 1
n′
.
The reduction theorem in the class Jα takes the following form.
Theorem 6.4. [Reduction principle for Maz'ya classes] Let n ∈ N, n ≥ 2, m ∈ N, and
α ∈ [ 1
n′ , 1]. Let k · kX(0,1) and k · kY (0,1) be rearrangement-invariant function norms. Assume that
either α ∈ [ 1
(6.7)
n′ , 1) and there exists a constant C1 such that
for every nonnegative f ∈ X(0, 1), or α = 1 and there exists a constant C1 such that
(6.8)
t
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Z 1
t
f (s)s−1+m(1−α) ds(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1) ≤ C1 kfkX(0,1)
s(cid:18) log
t(cid:19)m−1
≤ C1 kfkX(0,1)
1
s
f (s)
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
V mX(Ω) → Y (Ω)
for every nonnegative f ∈ X(0, 1). Then the Sobolev embedding
(6.9)
holds for every Ω ∈ Jα and, equivalently, the Poincar´e inequality
kukY (Ω) ≤ C2 k∇mukX(Ω)
(6.10)
holds for every Ω ∈ Jα, for some constant C2, depending on Ω, m, X and Y , and every u ∈ V m
for every Ω ∈ Jα, then either inequality (6.7), or (6.8) holds, according to whether α ∈ [ 1
.
⊥ X(Ω).
Conversely, if the Sobolev embedding (6.9), or, equivalently, the Poincar´e inequality (6.10), holds
n′ , 1) or α = 1
(6.11)
kfkX ′
m,α(0,1) =
(cid:13)(cid:13)s−1+m(1−α)R s
(cid:13)(cid:13)(cid:13) 1
sR s
0 (cid:0) log s
0 f ∗(r)dr(cid:13)(cid:13)X ′(0,1)
r(cid:1)m−1f ∗(r)dr(cid:13)(cid:13)(cid:13)X ′(0,1)
if α ∈ [ 1
n′ , 1),
if α = 1,
20
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
A major consequence of Theorem 6.4 is the identification of the optimal rearrangement-invariant
target space Y (Ω) associated with a given domain X(Ω) in embedding (6.9) as Ω is allowed to range
among all domains in the class Jα. This is the content of the next result. The rearrangement-invariant
function norm yielding such an optimal space will be denoted by k·kXm,α(0,1). Given a rearrangement-
invariant function norm k · kX(0,1), m ∈ N, and α ∈ [ 1
n′ , 1], it is characterized through its associate
function norm defined by
for f ∈ M+(0, 1).
Theorem 6.5. [Optimal target for Maz'ya classes] Let n ∈ N, n ≥ 2, m ∈ N, α and k · kX(0,1)
be as in Theorem 6.4. Then the functional k·kX ′
m,α(0,1), given by (6.11), is a rearrangement-invariant
function norm, whose associate norm k · kXm,α(0,1) satisfies
(6.12)
for every Ω ∈ Jα, and
(6.13)
for every Ω ∈ Jα, for some constant C, depending on Ω, m, X and Y , and every u ∈ V m
invariant norms, as Ω ranges in Jα.
Moreover, the function norm k·kXm,α(0,1) is optimal in (6.12) and (6.13) among all rearrangement-
kukXm,α(Ω) ≤ C k∇mukX(Ω)
V mX(Ω) → Xm,α(Ω)
⊥ X(Ω).
Theorem 6.5 is a straightforward consequence of Theorem 6.4, and either Corollary 5.9 or Theorem
n′ , 1) or α = 1.
5.4, according to whether α ∈ [ 1
The stability of the process of finding optimal rearrangement-invariant targets in Euclidean Sobolev
embeddings on Maz'ya domains under iteration is the object of the last main result of the present
section. This is the key ingredient which bridges the first-order case of Theorems 6.4 and 6.5 to their
higher-order versions.
Theorem 6.6. [Iteration principle for Maz'ya classes] Let n ∈ N, α ∈ [ 1
as in Theorem 6.4. Let k, h ∈ N. Assume that Ω ∈ Jα. Then,
n′ , 1] and k · kX(0,1) be
(Xk,α)h,α(Ω) = Xk+h,α(Ω),
up to equivalent norms.
Theorem 6.6 follows from a specialization of Corollary 5.10 (α ∈ [ 1
n′ , 1)), or Theorem 5.7 (α = 1).
Remark 6.7. Note that there is one important difference between the reduction and the optimal-
target theorem concerning John domains on the one hand, and their counterparts for general Maz'ya
domains on the other hand. Namely, the equivalence in Theorem 6.1 and the optimality result in The-
orem 6.2 are valid for each single John domain, whereas the necessity of condition (6.7) for (6.8)
(or (6.9)) in Theorem 6.4 as well as the optimality of the target space in Theorem 6.5 are valid in the
class of all Ω ∈ Jα. This is inevitable, since, of course, each class Jα contains all regular domains,
and for such domains Sobolev embeddings with stronger target norms hold.
The remaining part of this section is devoted to applications of Theorems 6.4 -- 6.6 to customary func-
tion norms. Consider first the case when Lebesgue or Lorentz norms are concerned. Our conclusions
take a different form, according to whether α ∈ [ 1
by the choice α = 1
n′ .
We begin by assuming that α ∈ [ 1
Sobolev embeddings involving usual Lebesgue norms are contained in the following theorem.
n′ , 1). Note that results for regular (i.e. John) domains are covered
n′ , 1), or α = 1.
Theorem 6.8. Let n ∈ N, n ≥ 2, and let Ω ∈ Jα for some α ∈ [ 1
Then
n′ , 1). Let m ∈ N and p ∈ [1,∞].
HIGHER-ORDER EMBEDDINGS
21
(6.14)
V mLp(Ω) →
p
1−mp(1−α) (Ω)
L
Lr(Ω)
L∞(Ω)
if m(1 − α) < 1 and 1 ≤ p <
for any r ∈ [1,∞), if m(1 − α) < 1 and p =
otherwise.
m(1−α) ,
1
1
m(1−α) ,
Moreover, in the first and the third cases, the target spaces in (6.14) are optimal among all Lebesgue
spaces, as Ω ranges in Jα.
Although the target spaces in (6.14) cannot be improved in the class of Lebesgue spaces, the
conclusions of (6.14) can be strengthened if more general rearrangement-invariant spaces are employed.
Such a strengthening can be obtained as a special case of a Sobolev embedding for Lorentz spaces
which reads as follows.
Theorem 6.9. Let n ∈ N, n ≥ 2, and let Ω ∈ Jα for some α ∈ [ 1
Assume that one of the conditions in (3.7) holds. Then
n′ , 1). Let m ∈ N and p, q ∈ [1,∞].
(6.15)
V mLp,q(Ω) →
p
1−mp(1−α) ,q(Ω)
L
L∞,q;−1(Ω)
L∞(Ω)
if m(1 − α) < 1 and 1 ≤ p <
if m(1 − α) < 1, p =
otherwise,
1
1
m(1−α) ,
m(1−α) and q > 1,
Moreover, the target spaces in (6.15) are optimal among all rearrangement-invariant spaces, as Ω
ranges in Jα.
The particular choice of parameters p = q, 1 ≤ p <
1
m(1−α) in Theorem 6.9 shows that
V mLp(Ω) → L
p
1−mp(1−α) ,p(Ω).
This is a non-trivial strengthening of the first embedding in (6.14), since L
Likewise, the choice m(1− α) < 1 and p = q =
can be in fact essentially improved by
1−mp(1−α) .
m(1−α) shows that also the second embedding in (6.14)
1
p
1−mp(1−α) ,p(Ω) $ L
p
Assume now that α = 1. The embedding theorem in Lebesgue spaces takes the following form.
V mLp(Ω) → L∞,p;−1(Ω).
Theorem 6.10. Let n ∈ N, n ≥ 2, and let Ω ∈ J1. Let m ∈ N and p ∈ [1,∞]. Then
V mLp(Ω) →(Lp(Ω)
Lr(Ω)
if 1 ≤ p < ∞,
for any r ∈ [1,∞), if p = ∞.
(6.16)
Moreover, in the former case of (6.16), the target space is optimal among all Lebesgue spaces, as Ω
ranges in J1.
Optimal embeddings for Lorentz-Sobolev spaces are provided in the next theorem.
Theorem 6.11. Let n ∈ N, n ≥ 2, and let Ω ∈ J1. Let m ∈ N and p, q ∈ [1,∞]. Assume that one of
the conditions in (3.7) holds. Then
(6.17)
V mLp,q(Ω) →(Lp,q(Ω)
exp L
1
m (Ω)
if 1 ≤ p < ∞,
if p = q = ∞.
The target spaces are optimal in (6.17) among all rearrangement-invariant spaces, as Ω ranges in J1.
22
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Our last application in this section concerns Orlicz-Sobolev spaces. Let n ∈ N, n ≥ 2, m ∈ N,
n′ , 1), and let A be a Young function. We may assume, without loss of generality, that m < 1
1−α
α ∈ [ 1
and
(6.18)
Z0(cid:18) t
A(t)(cid:19) m(1−α)
1−m(1−α)
dt < ∞.
Indeed, by (3.12), the function A can be modified near 0, if necessary, in such a way that (6.18) is
fulfilled, on leaving the space V mLA(Ω) unchanged (up to equivalent norms).
If m < 1
1−α and the integral
(6.19)
diverges, we define the function Hm,α : [0,∞) → [0,∞) as
1−m(1−α)
A(t)(cid:19) m(1−α)
dt
Z ∞(cid:18) t
A(t)(cid:19) m(1−α)
0 (cid:18) t
1−m(1−α)
dt(cid:19)1−m(1−α)
Hm,α(s) =(cid:18)Z s
for s ≥ 0,
and the Young function Am,α as
Am,α(t) = A(H −1
m,α(t))
for t ≥ 0.
Theorem 6.12. Assume that n ∈ N, n ≥ 2, m ∈ N, α ∈ [ 1
function fulfilling (6.18). Then
(6.20) V mLA(Ω) →(LAm,α(Ω)
if m < 1
if either m ≥ 1
1−α , or m < 1
L∞(Ω)
1−α , and the integral (6.19) diverges,
n′ , 1) and Ω ∈ Jα. Let A be a Young
1−α and the integral (6.19) converges.
Moreover, the target spaces in (6.20) are optimal among all Orlicz spaces, as Ω ranges in Jα.
Theorem 6.12 follows from Theorem 6.4, via [29, Theorem 4].
The first case of embedding (6.20) can be enhanced, on replacing the optimal Orlicz target spaces
with the optimal rearrangement-invariant target spaces. The latter turn out to belong to the family
of Orlicz-Lorentz spaces defined in Section 3.
Assume that m < 1
1−α , and the integral (6.19) diverges. Let a be the left-continuous function
appearing in (3.11), and let B be the Young function given by
b(τ )dτ
for t ≥ 0,
0
B(t) =Z t
a(t)(cid:19) m(1−α)
where b is the non-decreasing, left-continuous function in [0,∞) obeying
b−1(s) =(cid:18)Z ∞
a−1(s)(cid:18)Z τ
0 (cid:18) 1
1−m(1−α)
m(1−α)
dt(cid:19)− 1
dτ
1−m(1−α)(cid:19) m(1−α)
m(1−α)−1
1
a(τ )
for s ≥ 0 .
Here, a−1 and b−1 denote the (generalized) left-continuous inverses of a and b, respectively.
Recall from Section 3 that L(
1
m(1−α) , 1, B)(Ω) is the Orlicz-Lorentz space built upon the function
norm given by
kfkL(
m(1−α) ,1,B)(0,1) = ks−m(1−α)f ∗(s)kLB (0,1)
1
for f ∈ M+(0, 1).
Theorem 6.13. Assume that n ∈ N, n ≥ 2, m ∈ N, α ∈ [ 1
function fulfilling (6.18). Assume that m < 1
1−α , and the integral in (6.19) diverges. Then
n′ , 1) and Ω ∈ Jα. Let A be a Young
(6.21)
V mLA(Ω) → L(
1
m(1−α) , 1, B)(Ω),
HIGHER-ORDER EMBEDDINGS
23
and the target space in (6.21) is optimal among all rearrangement-invariant spaces, as Ω ranges in
Jα.
Embedding (6.21) is a consequence of Theorem 6.4, and of [26, inequality (3.1)].
Example 6.14. Consider the case when
A(t) ≈ tp(log t)β near infinity, where either p > 1 and β ∈ R, or p = 1 and β ≥ 0.
Hence, LA(Ω) = LplogβL(Ω). An application of Theorem 6.12 tells us that
p
β
1−pm(1−α) log
1−pm(1−α) L(Ω)
exp L
1
1−(1+β)m(1−α) (Ω)
1
1−m(1−α) (Ω)
exp exp L
L∞(Ω)
if mp(1 − α) < 1,
if mp(1 − α) = 1 and β < 1−m(1−α)
if mp(1 − α) = 1 and β = 1−m(1−α)
if either mp(1 − α) > 1,
or mp(1 − α) = 1 and β > 1−m(1−α)
m(1−α)
m(1−α)
m(1−α)
,
,
.
(6.22)
(6.23)
V mLplogβL(Ω) →
L
V mLplogβL(Ω) →
The first three embeddings in (6.22) can be improved on allowing more general rearrangement-
Moreover, the target spaces in (6.22) are optimal among all Orlicz spaces, as Ω ranges in Jα.
invariant target spaces. Indeed, we have that
1−pm(1−α) ,p; β
p
L
L∞,
L∞,
p (Ω)
1
m(1−α) ;m(1−α)β−1(Ω)
m(1−α) ;−m(1−α),−1(Ω)
1
if mp(1 − α) < 1,
if mp(1 − α) = 1 and β < 1−m(1−α)
if mp(1 − α) = 1 and β = 1−m(1−α)
m(1−α)
m(1−α)
,
,
the targets being optimal among all rearrangement-invariant spaces in (6.23) as Ω ranges among all
domains in Jα. This is a consequence of Theorem 6.13, and of the fact that the Orlicz-Lorentz spaces
m(1−α) , 1, B)(Ω) associated with the present choices of the function A agree (up to equivalent norms)
L(
with the (generalized) Lorentz-Zygmund spaces appearing on the right-hand side of (6.23).
1
7. Sobolev embeddings in product probability spaces
The class of product probability measures in Rn, n ≥ 1, which we consider in this section arises in
connection with the study of generalized hypercontractivity theory and integrability properties of the
associated heat semigroups. The isoperimetric problem in the corresponding probability spaces was
studied in [5] -- see also [4, 10, 57, 58].
Assume that Φ : [0,∞) → [0,∞) is a strictly increasing convex function on [0,∞), twice continuously
differentiable in (0,∞), such that √Φ is concave and Φ(0) = 0. Let µΦ be the probability measure on
R given by
(7.1)
where cΦ is a constant chosen in such a way that µΦ(R) = 1. The product measure µΦ,n on Rn, n ≥ 1,
generated by µΦ, is then defined as
dµΦ(x) = cΦe−Φ(x) dx,
(7.2)
µΦ,n = µΦ × ··· × µΦ
.
n−times
{z
}
Clearly, µΦ,1 = µΦ, and (Rn, µΦ,n) is a probability space for every n ∈ N.
The main example of a measure µΦ is obtained by taking
(7.3)
Φ(t) = 1
2 t2.
This choice yields µΦ,n = γn, the Gauss measure which obeys
2 e− x2
dγn(x) = (2π)− n
2 dx.
24
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
More generally, given any β ∈ [1, 2], the Boltzmann measure γn,β in Rn, associated with
(7.4)
Φ(t) = 1
β tβ,
satisfies the above assumptions.
Let H : R → (0, 1) be defined as
H(t) =Z ∞
t
cΦe−Φ(r) dr
for t ∈ R,
(7.5)
(7.6)
and
and let FΦ : [0, 1] → [0,∞) be given by
FΦ(s) = cΦe−Φ(H −1(s))
for s ∈ (0, 1),
and FΦ(0) = FΦ(1) = 0.
Since µΦ is a probability measure and µΦ,n is defined by (7.2), it is easily seen that, for each i = 1, . . . , n,
Hence, FΦ(s) agrees with the perimeter of any half-space of the form {xi > t}, whose measure is s.
µΦ,n({(x1, . . . , xn) : xi > t}) = H(t)
for t ∈ R,
PµΦ,n ({(x1, . . . , xn) : xi > t}, Rn) = cΦe−Φ(t) = −H ′(t)
LΦ(s) = sΦ′(cid:0)Φ−1(log(cid:0) 2
s(cid:1)(cid:1)(cid:1)
for s ∈ (0, 1],
Next, define LΦ : [0, 1] → [0,∞) as
(7.7)
Then the isoperimetric function of (Rn, µΦ,n) satisfies
(7.8)
I(Rn,µΦ,n)(s) ≈ FΦ(s) ≈ LΦ(s)
for s ∈ [0, 1
2 ]
for t ∈ R.
and LΦ(0) = 0.
(see [5, Proposition 13 and Theorem 15]; note that the second equivalence in (7.8) also relies upon
Lemma 11.1 (ii) of Section 11). Furthermore, half-spaces, whose boundary is orthogonal to a coordinate
axis, are "approximate solutions" to the isoperimetric problem in (Rn, µΦ,n) in the sense that there
exist constants C1 and C2, depending on n, such that, for every s ∈ (0, 1), any such half-space V with
measure s satisfies
C1PµΦ,n(V, Rn) ≤ I(Rn,µΦ,n)(s) ≤ C2PµΦ,n(V, Rn).
In the special case when µΦ,n = γn, the Gauss measure, equation (7.8) yields
Moreover, any half-space is, in fact, an exact minimizer in the isoperimetric inequality [13, 84].
s(cid:1) 1
I(Rn,γn)(s) ≈ s(cid:0) log 2
2
for s ∈ (0, 1
2 ].
Our reduction theorem for Sobolev embeddings in product probability spaces reads as follows.
Theorem 7.1. [Reduction principle for product probability spaces] Let n ∈ N, m ∈ N, let µΦ,n
be the probability measure defined by (7.2), and let k·kX(0,1) and k·kY (0,1) be rearrangement-invariant
function norms. Then the following facts are equivalent.
(i) The inequality
Φ−1(log 2
t !mZ 1
log 2
t )
t
f (s)
s (cid:16)log
s
t(cid:17)m−1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
(ii) The embedding
holds for some constant C1, and for every nonnegative f ∈ X(0, 1).
V mX(Rn, µΦ,n) → Y (Rn, µΦ,n)
(7.10)
(7.9)
holds.
≤ C1 kfkX(0,1)
(iii) The Poincar´e inequality
(7.11)
holds for some constant C2 and every u ∈ V m
⊥ X(Rn, µΦ,n).
kukY (Rn,µΦ,n) ≤ C2 k∇mukX(Rn,µΦ,n)
HIGHER-ORDER EMBEDDINGS
25
Let us notice that inequality (7.9) is not just a specialization of (5.3), but even a further simplifi-
cation of such specialization.
Let k · kX(0,1) be a rearrangement-invariant function norm, and let n, m ∈ N. The rearrangement-
invariant function norm k · kXm,Φ(0,1) which yields the optimal rearrangement-invariant target space
Y (Rn, µΦ,n) in embedding (7.10) is defined as follows. Consider the rearrangement-invariant function
norm k · keXm(0,1) whose associate norm fulfils
m(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)
sZ s
0 (cid:16)log
kfkXm,Φ(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
log 2
g∗(r) dr(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
r(cid:17)m−1
f ∗(s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eXm(0,1)
s )!m
for g ∈ M+(0, 1). Then k · kXm,Φ(0,1) is given by
kgkeX ′
Φ−1(log 2
for f ∈ M+(0, 1).
Remark 7.2. Note that if Φ(t) = t, and m ∈ N, we have that
s
s
(7.12)
1
eXm(0, 1) = Xm,Φ(0, 1)
for every r.i. norm k · kX(0,1).
Theorem 7.3. [Optimal target for product probability spaces] Let n, m, µΦ,n and k · kX(0,1)
be as in Theorem 7.1. Then the functional k·kXm,Φ (0,1), given by (7.12), is a rearrangement-invariant
function norm satisfying
(7.13)
V mX(Rn, µΦ,n) → Xm,Φ(Rn, µΦ,n),
and there exists a constant C such that
(7.14)
for every u ∈ V m
Moreover, the function norm k·kXm,Φ(0,1) is optimal in (7.13) and in (7.14) among all rearrangement-
invariant norms.
kukXm,Φ(Rn,µΦ,n) ≤ C k∇mukX(Rn,µΦ,n) ,
⊥ X(Rn, µΦ,n).
Remark 7.4. Let us emphasize that inequality (7.9) implies embedding (7.10) with a norm indepen-
dent of n, and the Poincar´e inequality (7.11) with constant C2 independent of n. The norm of the
optimal embedding (7.13), and the constant C in the corresponding Poincar´e inequality (7.14) are
independent of n as well.
For a broad class of rearrangement-invariant function norms k·kX(0,1) the expression of the associated
optimal Sobolev target norm k · kXm,Φ(0,1) can be substantially simplified, as observed in the next
proposition.
Proposition 7.5. Let m ∈ N and let Φ be as in (7.1). Suppose that k · kX(0,1) is a rearrangement-
invariant function norm such that the operator
is bounded on X ′(0, 1). Then
f 7→ f ∗∗
log 2
Φ−1(log 2
s
s )!m
f ∗(s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X(0,1)
kfkXm,Φ(0,1) ≈(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
up to multiplicative constants independent of f ∈ M+(0, 1).
26
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
The rearrangement-invariant spaces on which the operator "** " is bounded are fully characterized
in terms of their upper Boyd index. In particular, the assumptions of Proposition 7.5 are satisfied if
and only if the upper Boyd index of X ′(0, 1) is strictly smaller that 1 [8, Theorem 5.15].
The iteration principle for Sobolev embeddings on product probability measure spaces, on which
Theorem 7.1 rests, reads as follows.
Theorem 7.6. [Iteration principle for product probability spaces] Let n, µΦ,n and k · kX(0,1)
be as in Theorem 7.1, and let k, h ∈ N. Then,
up to equivalent norms.
(cid:0)Xk,Φ(cid:1)h,Φ(Rn, µΦ,n) = Xk+h,Φ(Rn, µΦ,n) ,
Specialization of Theorems 7.1, 7.3 and 7.6 to the case of (7.3) easily leads to the following results
for Gaussian Sobolev embeddings of any order.
Theorem 7.7. [Reduction principle in Gauss space] Let n ∈ N, m ∈ N, and let k · kX(0,1) and
k · kY (0,1) be rearrangement-invariant function norms. Then the following facts are equivalent.
(i) The inequality
≤ C1 kfkX(0,1)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1
2 Z 1
s(cid:1) m
(cid:0) log 2
s
f (r)
r (cid:16)log
r
s(cid:17)m−1
dr(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
holds for some constant C1, and for every nonnegative f ∈ X(0, 1).
(ii) The embedding
holds.
(iii) The Poincar´e inequality
V mX(Rn, γn) → Y (Rn, γn)
holds for some constant C2, and for every u ∈ V m
⊥ X(Rn, γn).
kukY (Rn,γn) ≤ C2 k∇mukX(Rn,γn)
Given n, m ∈ N, and a rearrangement-invariant function norm k·kX(0,1), define the rearrangement-
invariant function norm k · kXm,G(0,1) by
(7.15)
kfkXm,G(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:18)log
2
2
s(cid:19) m
f ∗(s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eXm(0,1)
for f ∈ M+(0, 1).
Theorem 7.8. [Optimal target in Gauss space] Let n ∈ N, m ∈ N, and let k · kX(0,1) be
a rearrangement-invariant function norm. Then the functional k · kXm,G(0,1), given by (7.15), is
a rearrangement-invariant function norm satisfying
(7.16)
and
V mX(Rn, γn) → Xm,G(Rn, γn)
(7.17)
for some constant C and every u ∈ V m
Moreover, the function norm k · kXm,G(0,1) is optimal in (7.16) and (7.17) among all rearrangement-
invariant norms.
kukXm,G(Rn,γn) ≤ C k∇mukX(Rn,γn)
⊥ X(Rn, γn).
Observe that, even for m = 1, Theorems 7.7 and 7.8 provide us with a characterization of Gaussian
Sobolev embeddings which somewhat simplifies earlier results in a similar direction [33, 63].
HIGHER-ORDER EMBEDDINGS
27
Theorem 7.9. [Iteration principle in Gauss space] Let n, k, h ∈ N, and let k · kX(0,1) be a
rearrangement-invariant function norm. Then,
up to equivalent norms.
(cid:0)Xk,G(cid:1)h,G(Rn, γn) = Xk+h,G(Rn, γn) ,
Of course, versions of Theorems 7.7 -- 7.9, with the Gauss measure replaced with the Boltzmann
measure, given by the choice (7.4), can similarly be deduced from Theorems 7.1, 7.3 and 7.6. The
reduction principle and the optimal target space then take the following form.
Theorem 7.10. [Reduction principle in Boltzmann spaces] Assume that n, m ∈ N, and β ∈
[1, 2]. Let k · kX(0,1) and k · kY (0,1) be rearrangement-invariant function norms. Then the following
facts are equivalent.
(i) The inequality
≤ C1 kfkX(0,1)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1
β
(cid:0) log 2
s(cid:1) m(β−1)
Z 1
s
f (r)
r (cid:16)log
r
s(cid:17)m−1
dr(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
(ii) The embedding
holds for some constant C1, and for every nonnegative f ∈ X(0, 1).
V mX(Rn, γn,β) → Y (Rn, γn,β)
holds.
(iii) The Poincar´e inequality
holds for some constant C2 and for every u ∈ V m
⊥ X(Rn, γn,β).
kukY (Rn,γn,β) ≤ C2 k∇mukX(Rn,γn,β )
Given n, m ∈ N, β ∈ [1, 2], and a rearrangement-invariant function norm k · kX(0,1), define the
rearrangement-invariant function norm k · kXm,B,β (0,1) by
(7.18)
kfkXm,B,β (0,1) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:18)log
β
2
s(cid:19) m(β−1)
f ∗(s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eXm(0,1)
for f ∈ M+(0, 1).
Theorem 7.11. [Optimal target in Boltzmann spaces] Let n, m ∈ N, and let k · kX(0,1) be
a rearrangement-invariant function norm. Then the functional k · kXm,B,β (0,1), given by (7.18), is
a rearrangement-invariant function norm satisfying
(7.19)
and
(7.20)
V mX(Rn, γn,β) → Xm,B,β(Rn, γn,β)
kukXm,B,β (Rn,γn,β) ≤ C k∇mukX(Rn,γn,β )
for some constant C and every u ∈ V m
invariant norms.
⊥ X(Rn, γn,β).
Moreover, the function norm k·kXm,B,β (0,1) is optimal in (7.19) and (7.20) among all rearrangement-
We present an application of the results of this section to the particular case when µΦ,n is a
Boltzmann measure, and the norms are of Lorentz -- Zygmund type.
28
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Theorem 7.12. Let n, m ∈ N, let β ∈ [1, 2] and let p, q ∈ [1,∞]
and α ∈ R be such that one of the conditions in (3.8) is satisfied. Then
if
if
V mLp,q;α(Rn, γn,β) →(Lp,q;α+ m(β−1)
L∞,q;α− m
(Rn, γn,β)
β (Rn, γn,β)
β
p < ∞;
p = ∞.
Moreover, in both cases, the target space is optimal among all rearrangement-invariant spaces.
When β = 2, Theorem 7.12 yields the following sharp Sobolev type embeddings in Gauss space.
Theorem 7.13. Let n, m ∈ N, and let p, q ∈ [1,∞] and α ∈ R be such that one of the conditions
in (3.8) is satisfied. Then
V mLp,q;α(Rn, γn) →(Lp,q;α+ m
L∞,q;α− m
2 (Rn, γn)
2 (Rn, γn)
if
if
p < ∞;
p = ∞.
Moreover, in both cases, the target space is optimal among all rearrangement-invariant spaces.
A further specialization of the indices p, q, α appearing in Theorem 7.13 leads to the following basic
embeddings.
Corollary 7.14. Let n, m ∈ N.
(i) Assume that p ∈ [1,∞). Then
V mLp(Rn, γn) → Lp(log L)
mp
2 (Rn, γn),
and the target space is optimal among all rearrangement-invariant spaces.
(ii) Assume that γ > 0. Then
and the target space is optimal among all rearrangement-invariant spaces.
V m exp Lγ(Rn, γn) → exp L
2γ
2+mγ (Rn, γn),
(iii)
and the target space is optimal among all rearrangement-invariant spaces.
V mL∞(Rn, γn) → exp L
2
m (Rn, γn),
Note that the target space in the second embedding of Theorem 7.13, and in the embeddings (ii)
and (iii) of Corollary 7.14 increases in m. This is caused by the fact that, V mL∞,q;α(Rn, γn) *
V kL∞,q;α(Rn, γn) if m > k.
8. Optimal target function norms
In this section we collect some basic properties about certain one-dimensional operators playing a
role in the proofs of our main results.
Let T : M+(0, 1) → M+(0, 1) be a sublinear operator, namely an operator such that
T (λf ) = λT f,
and T (f + g) ≤ C(T f + T g),
for some positive constant C, and for every λ ≥ 0 and f, g ∈ M+(0, 1).
X(0, 1) into Y (0, 1), and write
Given two rearrangement-invariant spaces X(0, 1) and Y (0, 1), we say that T is bounded from
(8.1)
if the quantity
T : X(0, 1) → Y (0, 1),
kTk = sup(cid:8)kT fkY (0,1); f ∈ X(0, 1) ∩ M+(0, 1), kfkX(0,1) ≤ 1(cid:9)
is finite. Such a quantity will be called the norm of T . The space Y (0, 1) will be called optimal, within
a certain class, in (8.1) if, whenever Z(0, 1) is another rearrangement-invariant space, from the same
HIGHER-ORDER EMBEDDINGS
29
class, such that T : X(0, 1) → Z(0, 1), we have that Y (0, 1) → Z(0, 1). Equivalently, the function
norm k · kY (0,1) will be said to be optimal in (8.1) in the relevant class.
Two operators T and T ′ from M+(0, 1) into M+(0, 1) will be called mutually associate if
Z 1
0
T f (s)g(s) ds =Z 1
0
f (s)T ′g(s) ds
for every f, g ∈ M+(0, 1).
Lemma 8.1. Let T and T ′ be mutually associate operators, and let X(0, 1) and Y (0, 1) be rearrangement-
invariant spaces. Then,
T : X(0, 1) → Y (0, 1)
if and only if T ′ : Y ′(0, 1) → X ′(0, 1),
and
Proof. The conclusion is a consequence of the following chain:
kTk = kT ′k.
kTk =
sup
f ≥0
kT fkY (0,1) =
sup
f ≥0
kf kX(0,1)≤1
kf kX(0,1)≤1
=
sup
g≥0
kgkY ′(0,1)≤1
kf kX(0,1)≤1Z 1
sup
f ≥0
0
kgkY ′(0,1)≤1Z 1
sup
g≥0
0
T f (s)g(s) ds
f (s)T ′g(s) ds =
sup
g≥0
kT ′gkX ′(0,1) = kT ′k.
kgkY ′(0,1)≤1
(cid:3)
Let I : [0, 1] → [0,∞) be a measurable function satisfying (5.2). We define the operators HI and
RI from M+(0, 1) into M+(0, 1) by
(8.2)
and
f (s)
I(s)
t
HI f (t) =Z 1
I(t)Z t
1
0
RIf (t) =
ds
for t ∈ (0, 1],
f (s) ds
for t ∈ (0, 1],
for f ∈ M+(0, 1). Moreover, given j ∈ N, we set
(8.3)
H j
I = HI ◦ HI ◦ ··· ◦ HI
j−times
{z
}
and Rj
I = RI ◦ RI ◦ ··· ◦ RI
.
j−times
{z
}
We also set H 0
I = R0
I = Id.
Remarks 8.2. (i) The operators HI and RI are mutually associate. Hence, H j
mutually associate for j ∈ N.
(ii) By the Hardy -- Littlewood inequality (3.2), we have, for everyf ∈ M+(0, 1),
I and Rj
I are also
RI f (t) ≤ RI f ∗(t)
for t ∈ (0, 1].
More generally, for every f ∈ M+(0, 1) and j ∈ N, one has that
(8.4)
I f ∗(t)
for t ∈ (0, 1].
(iii) For every j ∈ N and f ∈ M+(0, 1), we have that
dr
f (s)
(8.5)
H j
I f (t) =
I(s)(cid:18)Z s
t
I(r)(cid:19)j−1
I f (t) ≤ Rj
Rj
(j − 1)!Z 1
1
t
ds
for t ∈ (0, 1).
30
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Equation (8.5) holds for j = 1 by the very definition of HI. On the other hand, if (8.5) is assumed to
hold for some j ∈ N, then
H j+1
I
f (t) =Z 1
t
H j
I f (s)
I(s)
dτ
I(τ )(cid:19)j−1
I(τ )(cid:19)j−1
dτ
dr ds
ds dr
t
s
s
1
1
1
1
=
f (r)
f (r)
ds =
t
f (r)
I(r)(cid:18)Z r
(j − 1)!Z 1
I(s)Z 1
I(s)(cid:18)Z r
I(r)Z r
(j − 1)!Z 1
I(r)(cid:18)Z r
I(τ )(cid:19)j
j!Z 1
f (s)(cid:18)Z t
I(r)(cid:19)j−1
I(t)Z t
t
dτ
dr.
ds
dr
=
1
1
0
s
s
t
t
Hence, (8.5) follows by induction. Similarly, for every j ∈ N and f ∈ M+(0, 1), we also have that
(8.6)
Rj
If (t) =
1
(j − 1)!
for t ∈ (0, 1].
Given any j ∈ N and any rearrangement-invariant function norm k · kX(0,1), equation (8.6) implies
that
0,I
(8.7)
j,I (0,1) = (j − 1)!kRj
kfkX ′
j,I (0,1) is the functional introduced in (5.6). We also formally set
I f ∗kX ′(0,1),
= k · kX ′(0,1).
for f ∈ M+(0, 1), where k · kX ′
k · kX ′
Proposition 8.3. Let I : [0, 1] → [0,∞) be a measurable function satisfying (5.2). Let k · kX(0,1)
be a rearrangement-invariant function norm and let j ∈ N. Then the functional k · kX ′
j,I (0,1) defined
in (8.7) is a rearrangement-invariant function norm, whose associate norm k · kXj,I (0,1) fulfils
(8.8)
H j
I : X(0, 1) → Xj,I(0, 1).
Moreover, the space Xj,I (0, 1) is the optimal target in (8.8) among all rearrangement-invariant spaces.
Proof. We begin by showing that the functional k · kX ′
j,I (0,1) is a rearrangement-invariant function
norm. Let f, g ∈ M+(0, 1). By (3.1), R t
0 g∗(s) ds for t ∈ (0, 1).
Thus, by Hardy's lemma (see Section 3) applied, for each fixed t ∈ (0, 1), with f1(s) = (f + g)∗(s),
f2(s) = f ∗(s) + g∗(s) and h(s) = χ(0,t)(s)(cid:0)R t
0 (f + g)∗(s) ds ≤ R t
I(r)(cid:1)j−1, we obtain the triangle inequality
0 f ∗(s) ds +R t
dr
s
kf + gkX ′
j,I (0,1) ≤ kfkX ′
j,I (0,1) + kgkX ′
j,I (0,1).
Other properties in the axiom (P1) of the definition of rearrangement-invariant function norm, as well
as the axioms (P2), (P3) and (P6) are obviously satisfied. Next, it follows from (5.2) that there exists
a positive constant C such that
1
for t ∈ (0, 1). Therefore,
1
dr
≤ C j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
0 (cid:18)Z t
tZ t
= (j − 1)!C jk1kX ′(0,1),
s
r (cid:19)j−1
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
k1kX ′
t
1
I(t) ≤ C
j,I (0,1) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
0 (cid:18)Z t
I(t)Z t
dr
= C j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
0 (cid:18)log
tZ t
1
s
t
I(r)(cid:19)j−1
s(cid:19)j−1
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
f ∗(s) ds ≤ 2Z 1
Z 1
0
0
2
and (P4) follows. As far as (P5) is concerned, note that
f ∗(s) ds
HIGHER-ORDER EMBEDDINGS
31
for every f ∈ M+(0, 1). Thus, by (P5) for the norm k·kX ′ (0,1), there exists a positive constant C such
that, if f ∈ M+(0, 1), then
s
f ∗(s)(cid:18)Z t
f ∗(s)(cid:18)Z 1
s
dr
I(r)(cid:19)j−1
I(r)(cid:19)j
dr
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
j Z 1
C
1
2
0
1
dr
≥ CZ 1
I(t)Z t
I(r)!jZ 1
f ∗(s)(cid:18)Z t
I(r)(cid:19)j−1
f ∗(s) ds ≥ C ′kfkL1(0,1)
dr
0
0
s
2
ds ≥
ds dt
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=
0
1
C
I(t)Z t
j Z 1
2j (R 1
0
where C ′ = C
I(r) )j. Hence, property (P5) follows.
To prove (8.8), note that, by (8.4) and (8.7), we have
dr
1
2
kRj
I fkX ′(0,1) ≤ kRj
I f ∗kX ′(0,1) =
1
(j − 1)!kfkX ′
j,I (0,1)
for f ∈ M+(0, 1). Hence,
Rj
I : X ′
j,I (0, 1) → X ′(0, 1).
Since Rj
I and H j
I are mutually associate, equation (8.8) follows via Lemma 8.1.
It remains to prove that Xj,I(0, 1) is optimal in (8.8) among all rearrangement-invariant spaces. To
I : X(0, 1) →
this purpose, assume that Y (0, 1) is another rearrangement-invariant space such that H j
Y (0, 1). Then, by Lemma 8.1 again, Rj
I : Y ′(0, 1) → X ′(0, 1), namely
kRj
I fkX ′(0,1) ≤ CkfkY ′(0,1)
for some positive constant C, and every f ∈ M+(0, 1). Thus, in particular, by (8.7),
kfkX ′
j,I (0,1) = (j − 1)!kRj
I f ∗kX ′(0,1) ≤ (j − 1)!Ckf ∗kY ′(0,1) = (j − 1)!CkfkY ′(0,1)
for every f ∈ M+(0, 1). Hence, Y ′(0, 1) → X ′
shows that Xj,I(0, 1) is optimal in (8.8) among all rearrangement-invariant spaces.
j,I (0, 1), and, equivalently, Xj,I (0, 1) → Y (0, 1). This
(cid:3)
We introduce one more sequence of function norms, based on the iteration of the first-order function
1,I (0,1). Let I : [0, 1] → [0,∞) be a measurable function satisfying (5.2). Let k · kX(0,1) be
norm kfkX ′
a rearrangement-invariant function norm. Let j ∈ N∪{0}. We define k·kXj (0,1) as the rearrangement-
0(0,1) = k·kX ′(0,1),
invariant function norm whose associate norm k·kX ′
and, for j ≥ 1, by
(8.9)
j (0,1) is given, via iteration, by k·kX ′
kfkX ′
j (0,1) = kRI f ∗kX ′
j−1(0,1)
for f ∈ M+(0, 1). Note that
(8.10)
kfkX1(0,1) = kfkX1,I (0,1).
Remark 8.4. By Proposition 8.3, applied j times with j = 1, we obtain that, for every j ∈ N ∪ {0},
the functional k · kX ′
j (0,1) is actually a rearrangement-invariant function norm. Moreover, its associate
function norm k · kXj (0,1) fulfils
(8.11)
HI : Xj(0, 1) → Xj+1(0, 1),
and k·kXj+1(0,1) is the optimal target function norm in (8.11) among all rearrangement-invariant func-
tion norms. By Lemma 8.1, we also have
RI : X ′
j+1(0, 1) → X ′
j(0, 1).
32
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Remark 8.5. Note that, by the very definition of Xj(0, 1),
Xj(0, 1) = (. . . (X1,I )1,I . . . )1,I
(0, 1)
j−times
{z
}
(Xk)h(0, 1) = Xk+h(0, 1)
for j ∈ N. In particular,
(8.12)
for every k, h ∈ N.
We now turn our attention to the special situation when I satisfies, in addition, condition (5.12).
In this case, most of the results take a simpler form. We start with a result concerned with the
equivalence of two couples of functionals under (5.12).
Proposition 8.6. Let I : [0, 1] → [0,∞) be a non-decreasing function satisfying (5.12) and let k·kX(0,1)
be any rearrangement-invariant function norm. Then:
(i) For every j ∈ N, and f ∈ M+(0, 1),
f (s)
I(s)(cid:18)Z s
t
dr
I(r)(cid:19)j−1
(8.13)
up to multiplicative constants independent of k · kX(0,1) and f .
(ii) For every j ∈ N, and f ∈ M+(0, 1),
dr
f (t)(cid:18)Z s
t
I(r)(cid:19)j−1
t
Z 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
I(s)Z s
0
1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:18) s
I(s)(cid:19)j−1
t
,
0
f (s)
sj−1
sj−1
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X(0,1)
I(s)j ds(cid:13)(cid:13)(cid:13)(cid:13)X(0,1)
≈(cid:13)(cid:13)(cid:13)(cid:13)Z 1
dt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X(0,1)
≈(cid:13)(cid:13)(cid:13)(cid:13)
f (t)dt(cid:13)(cid:13)(cid:13)(cid:13)X(0,1)
I(s)j Z s
≤ 2j−1 Z s
dr!j−1
I(r)!j−1
I(r)!j−1
ds
dr
s
2
,
for s ∈ (0, 1).
f (s)
I(s)(cid:18)Z s
t
dr
I(r)(cid:19)j−1
ds
for t ∈ (0, 1
2 ].
up to multiplicative constants independent of k · kX(0,1) and f .
Proof. We first note that, owing to the monotonicity of I, we have, for every j ∈ N,
=
2j−1
I(s)j−1 Z s
s
2
Thus,
Z 1
2t
I(s)(cid:19)j−1
f (s)
I(s)(cid:18) s
≤ 2j−1Z 1
2t
ds ≤ 2j−1Z 1
I(r)(cid:19)j−1
dr
2t
f (s)
I(s)(cid:18)Z s
t
dr
f (s)
I(s) Z s
ds ≤ 2j−1Z 1
s
2
t
Hence, the right-hand side of (8.13) does not exceed a constant times its left-hand side, owing to
the boundedness of the dilation operator in rearrangement-invariant spaces. Note that this inequality
holds even without the assumption (5.12). On the other hand, (5.12) implies
Z 1
t
f (s)
I(s)(cid:18)Z s
t
dr
I(r)(cid:19)j−1
t
I(s)(cid:18)Z s
f (s)
ds ≤Z 1
≤ C j−1Z 1
dr
I(r)(cid:19)j−1
I(s)(cid:19)j−1
0
f (s)
I(s)(cid:18) s
t
ds
ds
for t ∈ (0, 1),
hence the converse inequality in (8.13) follows. This proves (i).
The proof of (ii) is similar.
(cid:3)
k · k(X ♯
j,I )′(0,1) by
(8.14)
kfk(X ♯
j,I )′(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)
tj−1
I(t)j Z t
0
f ∗(s) ds(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
Given j ∈ N and a rearrangement-invariant function norm k · kX(0,1), we define the functional
HIGHER-ORDER EMBEDDINGS
33
for f ∈ M+(0, 1).
Remark 8.7. It follows from Proposition 8.6 and its proof that for every rearrangement-invariant
norm k · kX(0,1) and every j ∈ N, we have
(X ♯
j,I )′(0, 1) → X ′
j,I(0, 1),
and if moreover (5.12) is satisfied, then, in fact,
j,I )′(0, 1) = X ′
This observation has a straightforward consequence.
(X ♯
j,I(0, 1).
Proposition 8.8. Let I : [0, 1] → [0,∞) be a non-decreasing function satisfying (5.12) and let k·kX(0,1)
be any rearrangement-invariant function norm. Then
Xj,I (0, 1) = X ♯
j,I(0, 1),
up to equivalent norms.
The following result is a counterpart of Proposition 8.3 under (5.12). It follows from Proposition
8.3, with j = 1 and I replaced with the function (0, 1) ∋ t 7→ I(t)j
Proposition 8.9. Let I : [0, 1] → [0,∞) be a non-decreasing function satisfying (5.12). Let X(0, 1)
be a rearrangement invariant space and let j ∈ N. Then the functional k · k(X ♯
j,I )′(0,1) defined in (8.14)
is a rearrangement-invariant function norm. Moreover,
H j
I : X(0, 1) → X ♯
tj−1 , which obviously satisfies (5.2).
j,I(0, 1),
and X ♯
j,I (0, 1) is optimal in (8.8) among all rearrangement-invariant spaces.
9. Proofs of the main results
Here we are concerned with the proof of the results of Section 5. In what follows, Rm
I denotes the
operator defined as in (8.3).
Lemma 9.1. Let I : [0, 1] → [0,∞) be a non-decreasing function fulfilling (5.2), and let m ∈ N ∪{0}.
Then, for every f ∈ M+(0, 1),
I f ∗(t) ≤ 2mRm
Rm
(9.1)
Consequently, for every f ∈ M+(0, 1),
I f ∗(d) ≤ 2m+1Z d
(9.2)
if 0 ≤ c < d ≤ 1.
2 ≤ s ≤ t ≤ 1.
(d − c)Rm
I f ∗(s) ds
I f ∗(s)
if 0 < t
Rm
c
Proof. We prove inequality (9.1) by induction. Fix any f ∈ M+(0, 1). If m = 0 then (9.1) is satisfied
thanks to the monotonicity of f ∗. Next, let m ≥ 1, and assume that (9.1) is fulfilled with m replaced
with m − 1. If 0 < t
2 ≤ s ≤ t ≤ 1, then
1
2m−1
2
Rm−1
f ∗(r) dr
I
Rm
I f ∗(t) =
Rm−1
f ∗(r) dr ≤
I(s) Z t
Rm−1
I
0
I f ∗(s),
f ∗(cid:16) r
2(cid:17) dr =
2m
I(s)Z t
0
Rm−1
I
f ∗(r) dr = 2mRm
I
I(t)Z t
I(s)Z s
0
2m
≤
0
34
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
where the first inequality holds according to the induction assumption and to the fact that I is non-
decreasing on [0, 1]. Inequality (9.1) follows.
Now, let 0 ≤ c < d ≤ 1, m ∈ N ∪ {0} and f ∈ M+(0, 1). Thanks to (9.1),
Z d
c
Rm
I f ∗(d) ds = 2Z d
2
c+d
Rm
I f ∗(d) ds ≤ 2m+1Z d
2
c+d
Rm
I f ∗(s) ds ≤ 2m+1Z d
c
This proves (9.2).
Rm
I f ∗(s) ds.
(cid:3)
Given m ∈ N and a non-decreasing function I : [0, 1] → [0,∞) fulfilling (5.2), we define the operator
Gm
I at every f ∈ M+(0, 1) by
Gm
I f (t) = sup
t≤s≤1
Rm
I f ∗(s)
for t ∈ (0, 1).
Note that, trivially, Rm
function, and hence (Rm
I f ∗ ≤ Gm
I f ∗)∗ ≤ Gm
I f for every f ∈ M+(0, 1). Moreover, Gm
I f as well.
The following lemma tells us that the operator Gm
I does not essentially change if I is replaced with
its left-continuous representative.
Lemma 9.2. Let m ∈ N, let I : [0, 1] → [0,∞) be a non-decreasing function fulfilling (5.2), and let
I0 : [0, 1] → [0,∞) be the left-continuous function which agrees with I a.e. in [0, 1]. Then, for every
f ∈ M+(0, 1),
I f is a non-increasing
Gm
I f = Gm
I0f
up to a countable subset of (0, 1).
Proof. Define M = {t ∈ (0, 1) : I(t) 6= I0(t)}. The set M is at most countable. We shall prove that,
for every g ∈ M+(0, 1),
sup
(9.3)
t≤s≤1
for t ∈ (0, 1) \ M .
g(r) dr = sup
t≤s≤1
g(r) dr
1
I0(s)Z s
0
The conclusion will then follow by applying (9.3) to the function g = Rm−1
1
f ∗, and by the fact that
f ∗)(r) dr for s ∈ (0, 1). Fix g ∈ M+(0, 1) and t ∈ (0, 1). Given
f ∗)(r) dr = 1
0 (Rm−1
0 (Rm−1
I0
I0
I
1
I(τ )(cid:17)Z s
0
g(r) dr = lim
τ →s−
1
I(τ )Z τ
0
g(r) dr ≤ sup
t<τ ≤1
1
I(τ )Z τ
0
g(r) dr.
On taking the supremum over all s ∈ (t, 1], we get that
g(r) dr ≤ sup
t<s≤1(cid:16) lim
τ →s−
g(r) dr ≤ sup
t<τ ≤1
1
I(τ )Z τ
0
g(r) dr.
I(s)R s
s ∈ (t, 1], we have that
1
I(s)Z s
0
0
1
I(s)Z s
I(s)R s
g(r) dr ≤(cid:16) lim
I(s)Z s
τ →s−
1
0
sup
t<s≤1
0
1
I(τ )(cid:17)Z s
I0(s)Z s
1
0
Hence, since I0(s) = limτ →s− I(τ ) for s ∈ (0, 1),
g(r) dr = sup
t<s≤1
sup
t<s≤1
1
I(s)Z s
0
g(r) dr
for t ∈ (0, 1).
This yields (9.3).
Proposition 9.3. Let m ∈ N, let I : [0, 1] → [0,∞) be a left-continuous non-decreasing function
fulfilling (5.2), and let f ∈ M+(0, 1). Define
(9.4)
Then E is an open subset of (0, 1). Hence, there exists an at most countable collection {(ck, dk)}k∈S
of pairwise disjoint open intervals in (0, 1) such that
E = {t ∈ (0, 1) : Rm
I f ∗(t) < Gm
I f (t)}.
(cid:3)
(9.5)
E = ∪k∈S(ck, dk).
Moreover,
(9.6)
and
(9.7)
HIGHER-ORDER EMBEDDINGS
35
Gm
I f (t) = Rm
I f ∗(t)
if t ∈ (0, 1) \ E,
Gm
I f ∗(dk)
I f (t) = Rm
I f (t) = ∞, then both functions Gm
if t ∈ (ck, dk) for some k ∈ S.
Proof. Fix t ∈ (0, 1). If Gm
and hence there is nothing to prove. Assume that Gm
is attained. This follows from the fact that the function Rm
I(s)Rm
since I is left-continuous and non-decreasing, and hence lower-semicontinuous.
I f ∗ are identically equal to ∞,
I f ∗(s)
I f ∗(s) is upper-semicontinuous, since
I(s) is upper-semicontinuous. Notice that this latter property holds
I f (t) < ∞. Then we claim that supt≤s≤1 Rm
I f ∗(s) is continuous, and 1
I f and Rm
Suppose now that t ∈ E. Then, due to the upper-semicontinuity of Rm
I f ∗, there exists δ > 0 such
that
(9.8)
Rm
I f ∗(r) < Gm
I f (t)
if r ∈ (t − δ, t + δ).
Let c ∈ [t, 1] be such that Rm
I f (t). Then, thanks to (9.8), c ∈ [t + δ, 1]. It easily follows that
Gm
I f (t) = Gm
I f (r) for every r ∈ (t − δ, t + δ), a piece of information that, combined with (9.8), yields
r ∈ E. This shows that E is an open set. Assertion (9.6) is trivial and (9.7) is an easy consequence
of the definition of Gm
I f ∗(c) = Gm
(cid:3)
I f .
Proposition 9.4. Let m ∈ N, let I : [0, 1] → [0,∞) be a left-continuous non-decreasing function
fulfilling (5.2), and let f ∈ M+(0, 1). Then
I GI f ≈ Gm+1
Gm
(9.9)
f,
I
up to multiplicative constants depending on m.
Proof. Fix any f ∈ M+(0, 1). Since RI f ∗ ≤ GI f , for every m ∈ N
(9.10)
Gm+1
Rm
f (t) = sup
t≤s≤1
I RI f ∗(s) ≤ sup
Rm
t≤s≤1
I
I GI f ∗(s) = Gm
I GI f (t)
for t ∈ (0, 1).
This shows that the right-hand side of (9.9) does not exceed the left-hand side. To show a converse
inequality, consider the set E defined as in (9.4), with m = 1. By Proposition 9.3, the set E is open.
Let {(ck, dk)}k∈S be open intervals as in (9.5). If t ∈ (ck, dk) for some k ∈ S, then, by (9.7) with
m = 1,
(9.11)
dk
I(dk)
f ∗∗(dk) = RI f ∗(dk) = GI f (t) ≥ RI f ∗(t) ≥
t
I(t)
f ∗∗(dk).
Observe that f ∗∗(dk) > 0. Indeed, if f ∗∗(dk) = 0, then RI f ∗(t) = RI f ∗(dk) = Gm
t /∈ E, a contradiction. Thus, we obtain from (9.11)
I f (t) = 0, and hence
(9.12)
dk
I(dk) ≥
t
I(t)
for t ∈ (ck, dk).
We shall now prove by induction that, given m ∈ N ∪ {0}, there exists a constant C = C(m) such
that
(9.13)
I GI f (t) ≤ C Rm+1
I
Rm
f ∗(t) +Xk∈S
χ(ck,dk)(t)Rm+1
I
f ∗(dk)! for t ∈ (0, 1).
36
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Let m = 0. Then (9.13) holds with C = 1, by (9.6) and (9.7) (with m = 1). Next, suppose that (9.13)
holds for some m ∈ N ∪ {0}. Fix any t ∈ (0, 1). Then
Rm+1
I GI f (t) =
0
1
Rm
I GI f (r) dr ≤
I(t)Z t
I(t) X{ℓ∈S:dℓ≤t}Z dℓ
C
cℓ
+
C
I(t)Z t
0
Rm+1
I
f ∗(dℓ) dr +
Rm+1
I
f ∗(r) dr
C
I(t)Xk∈S
χ(ck,dk)(t)Z t
ck
Rm+1
I
f ∗(dk) dr
2m+2C
I(t) X{ℓ∈S:dℓ≤t}Z dℓ
cℓ
Rm+1
I
f ∗(r) dr
χ(ck,dk)(t)Rm+1
I
f ∗(dk)
(by (9.2))
Rm+1
I
f ∗(r) dr
Rm+1
I
f ∗(dk)
(by (9.12))
f ∗(t) +
I
≤ CRm+2
t
+ C
I(t)Xk∈S
I
f ∗(t) +
χ(ck,dk)(t)
≤ CRm+2
+ CXk∈S
≤ (C + 2m+2C)Rm+2
+ C2m+2Xk∈S
= (C + 2m+2C)Rm+2
I
I
2m+2C
I(t) Z t
0
dk
I(dk)
f ∗(t)
≤ C ′ Rm+2
I
f ∗(t) +Xk∈S
χ(ck,dk)(t)
Rm+1
f ∗(r) dr
(by (9.2))
I
1
I(dk)Z dk
f ∗(t) + C2m+2Xk∈S
0
χ(ck,dk)(t)Rm+2
I
χ(ck,dk)(t)Rm+2
I
f ∗(dk)
f ∗(dk)! ,
where C ′ = C + 2m+2C. This proves (9.13).
Owing to (9.13), for every m ∈ N we have that
Gm
I GI f (t) = sup
t≤s≤1
I GI f (s) ≤ 2CGm+1
Rm
I
f (t)
for t ∈ (0, 1).
Combining this inequality with (9.10) yields (9.9).
Theorem 9.5. Let I : [0, 1] → [0,∞) be a non-decreasing function satisfying (5.2) and let k · kX(0,1)
be a rearrangement-invariant function norm. Let m ∈ N ∪ {0}. Then
(9.14)
for every f ∈ M+(0, 1), up to multiplicative constants depending on m.
Proof. We may assume, without loss of generality, that I is left-continuous. Indeed, equation (9.14)
is not affected by a replacement of I with its left-continuous representative, since the latter can differ
from I at most on a countable subset of [0, 1], and since Lemma 9.2 holds.
I ((RI f ∗)∗)kX ′(0,1) ≈ kGm+1
fkX ′(0,1) ≈ kRm+1
f ∗kX ′(0,1) ≈ kRm
kRm+1
f ∗kX ′
d(0,1)
(cid:3)
I
I
I
Fix any f ∈ M+(0, 1), and let m ≥ 1. By (8.4) and Proposition 9.4, there exists a constant
C = C(m) such that
Rm+1
I
Hence,
(9.15)
f ∗(t) ≤ Rm
I ((RI f ∗)∗)(t) ≤ Rm
I (GI f )(t) ≤ Gm
I GI f (t) ≤ CGm+1
I
f (t)
for t ∈ (0, 1).
kRm+1
I
f ∗kX ′(0,1) ≤ kRm
I ((RIf ∗)∗)kX ′(0,1) ≤ CkGm+1
I
fkX ′(0,1).
Observe that (9.15) trivially holds also when m = 0.
HIGHER-ORDER EMBEDDINGS
37
Let E be defined as in (9.4), with m replaced with m + 1, and let {(ck, dk)}k∈S be as in (9.5). For
every g ∈ X(0, 1), define
A(g) = χ(0,1)\Eg∗ +Xk∈S
χ(ck,dk)
1
dk − ckZ dk
ck
g∗(t) dt.
Then A(g) is non-increasing on (0, 1). Moreover, if kgkX(0,1) ≤ 1, then by [8, Theorem 4.8, Chapter
2],
(9.16)
Therefore,
kA(g)kX(0,1) ≤ kg∗kX(0,1) = kgkX(0,1) ≤ 1.
f ∗(t) dt +Xk∈SZ dk
ck
g∗(t)Rm+1
I
f ∗(dk) dt
g∗(t) dt(cid:19) (dk − ck)Rm+1
I
f ∗(dk)
A(g)(t)Rm+1
f ∗(t) dt
(by (9.2))
I
g∗(t)Rm+1
I
g∗(t)Rm+1
f (t) dt =Z(0,1)\E
dk − ck(cid:18)Z dk
f ∗(t) dt +Xk∈S
f ∗(t) dt + 2m+2Xk∈SZ dk
ck
ck
1
I
I
A(g)(t)Rm+1
A(g)(t)Rm+1
f ∗(t) dt
I
I
0
g∗(t)Gm+1
Z 1
=Z(0,1)\E
≤Z(0,1)\E
≤ 2m+2Z 1
≤ 2m+2
= 2m+2kRm+1
0
I
sup
khkX(0,1)≤1Z 1
f ∗kX ′
0
d(0,1).
h∗(t)Rm+1
I
f ∗(t) dt
(by (9.16))
On taking the supremum over all g from the unit ball of X(0, 1), we get
(9.17)
kGm+1
fkX ′
On the other hand, by the very definition of k · kX ′
(9.18)
fkX ′(0,1) = kGm+1
I
I
d(0,1) ≤ 2m+2kRm+1
d(0,1),
I
f ∗kX ′
d(0,1).
kRm+1
I
f ∗kX ′
d(0,1) ≤ kRm+1
I
f ∗kX ′(0,1).
Equation (9.14) follows from (9.15), (9.17) and (9.18).
(cid:3)
Corollary 9.6. Let I : [0, 1] → [0,∞) be a non-decreasing function satisfying (5.2), and let k · kX(0,1)
be a rearrangement-invariant function norm. Let m ∈ N. Then
(Xm,I )1(0, 1) = Xm+1,I (0, 1)
(9.19)
(up to equivalent norms).
Proof. By (8.9) and (8.7), if f ∈ M+(0, 1), then
kfk((Xm,I )1)′(0,1) = kRI f ∗kX ′
m,I (0,1) = (m − 1)!kRm
I ((RI f ∗)∗)kX ′(0,1),
and
kfkX ′
m+1,I (0,1) = m!kRm+1
I
f ∗kX ′(0,1).
Hence, it follows from Theorem 9.5 that
kfk((Xm,I )1)′(0,1) ≈ kfkX ′
m+1,I (0,1).
By (3.4), this establishes (9.19).
(cid:3)
38
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Theorem 9.7. Let I : [0, 1] → [0,∞) be a non-decreasing function satisfying (5.2) and let k · kX(0,1)
be a rearrangement-invariant function norm. Then, for every m ∈ N,
(9.20)
Xm,I (0, 1) = Xm(0, 1).
Proof. As noted in (8.10), we have X1(0, 1) = X1,I (0, 1). Assume now that (9.20) holds for some
m ∈ N. By (8.12), the induction assumption and (9.19),
Xm+1(0, 1) = (Xm)1(0, 1) = (Xm,I )1(0, 1) = Xm+1,I (0, 1).
The conclusion follows by induction.
(cid:3)
One consequence of Theorem 9.5, specifically of the equivalence of the leftmost and the rightmost
side of (9.14), is the following feature of inequality (5.3), which was already mentioned in Remark 5.2.
Corollary 9.8. Assume that (Ω, ν) fulfils (5.1) for some non-decreasing function I satisfying (5.2).
Let m ∈ N, and let k · kX(0,1) and k · kY (0,1) be rearrangement-invariant function norms. Then the
following two assertions are equivalent:
(i) There exists a constant C1 such that inequality (5.3) holds for every nonnegative f ∈ X(0, 1).
(ii) There exists a constant C ′
1 such that inequality (5.3) holds for every nonnegative non-increasing
f ∈ X(0, 1).
Proof. The fact that (i) implies (ii) is trivial. Conversely, assume that (ii) holds. Fix f ∈ M+(0, 1).
Equation (8.5) with j = m reads
(9.21)
Now, the function H m
Littlewood inequality (3.2) that
Z 1
t
f (s)
I(s)(cid:18)Z s
t
dr
I(r)(cid:19)m−1
ds = (m − 1)!H m
I f (t),
I f is non-increasing on (0, 1). Therefore, it follows from (3.3) and the Hardy --
kH m
I fkY (0,1) =
Consequently, by Fubini's theorem, we have
(9.22)
kH m
I fkY (0,1) =
0
sup
kgkY ′(0,1)≤1Z 1
kgkY ′(0,1)≤1Z 1
sup
0
g∗(t)H m
I f (t) dt.
f (t)Rm
I g∗(t) dt.
Owing to (9.21) and to the rearrangement-invariance of the norm k·kX(0,1), assertion (ii) tells us that
C ′
1 ≥ (m − 1)!
sup
kf kX(0,1) ≤1kH m
I f ∗kY (0,1).
Hence, on applying (9.22) with f replaced with f ∗, interchanging the suprema and recalling the
definition of the norm k · kX ′
d(0,1), we get
C ′
1 ≥ (m − 1)!
sup
kf kX(0,1) ≤1
= (m − 1)!
sup
kgkY ′(0,1)≤1
sup
kgkY ′(0,1)≤1Z 1
kf kX(0,1)≤1Z 1
sup
0
0
f ∗(t)Rm
I g∗(t) dt
f ∗(t)Rm
I g∗(t) dt = (m − 1)!
sup
kgkY ′(0,1)≤1kRm
I g∗kX ′
d(0,1).
It follows from the equivalence of the first and the last term in (9.14) that there exists a constant C
such that
Therefore,
kRm
I g∗kX ′(0,1) ≤ CkRm
d(0,1).
I g∗kX ′
kgkY ′(0,1)≤1kRm
sup
I g∗kX ′(0,1),
C ′
1C ≥ (m − 1)!
HIGHER-ORDER EMBEDDINGS
namely, by the definition of the norm k · kX ′(0,1),
C ′
1C ≥ (m − 1)!
sup
kgkY ′(0,1)≤1
kf kX(0,1)≤1Z 1
sup
0
f (t)Rm
I g∗(t) dt.
Interchanging suprema again and using Fubini's theorem and (9.22) yields
C ′
1C ≥ (m − 1)!
sup
kf kX(0,1) ≤1
kgkY ′(0,1)≤1Z 1
sup
0
g∗(t)H m
I f (t) dt = (m − 1)!
Hence, inequality (5.3), or equivalently assertion (i), follows.
sup
kf kX(0,1)≤1 kH m
I fkY (0,1).
39
(cid:3)
Proof of Theorem 5.1. As observed in Section 5, the case when m = 1 is already well-known, and
is in fact the point of departure of our approach. We thus focus on the case when m ≥ 2. On applying
Proposition 8.3 with j = 1, we get that
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
t
f (s)
I(s)
ds(cid:13)(cid:13)(cid:13)(cid:13)X1,I (0,1) ≤ kfkX(0,1)
for every f ∈ M+(0, 1). Thus, (5.3) holds with m = 1 and Y (0, 1) = X1,I (0, 1). Hence, by the result
for m = 1,
(9.23)
V 1X(Ω, ν) → X1(Ω, ν).
Note that here we have also made use of (8.10). By embedding (9.23) applied to each of the spaces
Xj(Ω, ν), for j = 0, . . . , m − 1, we get
V 1Xj(Ω, ν) → Xj+1(Ω, ν),
whence
(9.24)
V mX(Ω, ν) → V m−1X1(Ω, ν) → V m−2X2(Ω, ν) → ··· → V 1Xm−1(Ω, ν) → Xm(Ω, ν).
Inequality (5.3) tells us that
(9.25)
H m
I
: X(0, 1) → Y (0, 1).
The optimality of the space Xm,I (0, 1) as a target in (9.25), proved in Proposition 8.3, entails that
(9.26)
Xm,I (0, 1) → Y (0, 1).
A combination of (9.24), (9.20) and (9.26) yields
(9.27)
and (5.4) follows.
V mX(Ω, ν) → Xm(Ω, ν) = Xm,I (Ω, ν) → Y (Ω, ν),
Finally, (5.5) is equivalent to (5.4) by Proposition 4.4. Note that assumption (4.6) of that Propo-
(cid:3)
sition is satisfied, owing to (5.2).
Proof of Theorem 5.4. Embedding (5.7) is a straightforward consequence of (9.27). In turn, Propo-
sition 4.4 yields the Poincar´e inequality (5.8).
Assume now that the validity of (5.4) implies (5.3). Let k · kY (0,1) be any rearrangement-invariant
function norm such that (5.4) holds. Then, by our assumption, inequality (5.3) holds as well, namely
(9.28)
H m
I
: X(0, 1) → Y (0, 1).
Since, by Proposition 8.3, Xm,I (0, 1) is the optimal rearrangement-invariant target space in (9.28), we
necessarily have
Xm,I (0, 1) → Y (0, 1).
This implies the optimality of the norm k · kXm,I (0,1) in (5.7).
(cid:3)
40
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Proof of Corollary 5.5. Observe that
sup
f ≥0
kf kX(0,1)≤1(cid:13)(cid:13)(cid:13)(cid:13)Z 1
t
f (s)
I(s)(cid:18)Z s
t
dr
I(r)(cid:19)m−1
ds(cid:13)(cid:13)(cid:13)(cid:13)L∞(0,1)
0
0
=
dr
f (s)
sup
f ≥0
kf kX(0,1)≤1Z 1
=(cid:13)(cid:13)(cid:13)(cid:13)
I(s)(cid:18)Z s
I(s)(cid:18)Z s
I(r)(cid:19)m−1
I(r)(cid:19)m−1(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
dr
1
0
ds
.
Hence, (5.9) is equivalent to (5.3) with Y (0, 1) = L∞(0, 1). The assertion thus follows from Theo-
rem 5.1.
(cid:3)
Proof of Theorem 5.7. By Theorem 9.7 and (8.12),
(Xk,I )h,I(0, 1) = (Xk)h(0, 1) = Xk+h(0, 1) = Xk+h,I(0, 1),
and the claim follows.
(cid:3)
Corollaries 5.8, 5.9 and 5.10 follow from Theorems 5.1, 5.4 and 5.7, respectively (via Propositions
8.6 -- 8.9).
Proof of Proposition 4.5. Owing to (4.15), equation (4.17) will follow if we show that
(9.29)
V mX(Ω, ν) → W mX(Ω, ν).
The isoperimetric function IΩ,ν is non-decreasing on [0, 1
by
3 ] by definition. Let us define the function I
(9.30)
I(s) =(IΩ,ν(s)
IΩ,ν( 1
3 )
if s ∈ [0, 1
3 ],
if s ∈ [ 1
3 , 1].
Then I is non-decreasing on [0, 1]. Moreover, by (4.16), it satisfies (4.6). Let HI be the operator
defined as in (8.2), with I given by (9.30). Then,
kHI fkL1(0,1) ≤Z 1
0
f (t)
t
I(t)
dt ≤ CkfkL1(0,1),
and
kHI fkL∞(0,1) ≤Z 1
0
f (t)
I(t)
dt ≤Z 1
0
ds
I(s)kfkL∞(0,1) ≤ CkfkL∞(0,1),
for every f ∈ M+(0, 1). Thus, HI is well defined and bounded both on L1(0, 1) and on L∞(0, 1).
Owing to an interpolation theorem of Calder´on [8, Chapter 3, Theorem 2.12], the operator HI is
bounded on every r.i. space X(0, 1). Hence, from Theorem 5.1 applied with Y (0, 1) = X(0, 1) and
m = 1, we obtain that
(9.31)
V 1X(Ω, ν) → X(Ω, ν).
Iterating (9.31) tells us that there exists a constant C such that
(9.32)
k∇hukX(Ω,ν) ≤ C(cid:16) m−1Xk=h
k∇kukL1(Ω,ν) + k∇mukX(Ω,ν)(cid:17)
for every h = 0, . . . , m − 1, and u ∈ V mX(Ω, ν). Embedding (9.29) is a consequence of (9.32).
(cid:3)
HIGHER-ORDER EMBEDDINGS
41
10. Proofs of the Euclidean Sobolev embeddings
In what follows, we shall make use of the fact that the function I(t) = tα satisfies (5.12) if α ∈ (0, 1).
Proof of Theorem 6.1. If the one-dimensional inequality (6.1) holds, then the Sobolev embedding
(6.2) and the Poincar´e inequality (6.3) hold as well, owing to (2.3) and to Corollary 5.8. This shows
that (i) implies (ii) and (iii). The equivalence of (ii) and (iii) is a consequence of Proposition 4.4.
It thus only remains to prove that (ii) implies (i). Assume that the Sobolev embedding (6.2) holds.
If m ≥ n, then there is nothing to prove, because (6.1) holds for every rearrangement invariant spaces
X(0, 1) and Y (0, 1). Indeed,
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
t
f (s)s−1+ m
n ds(cid:13)(cid:13)(cid:13)(cid:13)L∞(0,1)
=Z 1
0
f (s)s−1+ m
n ds ≤ kfkL1(0,1)
for every nonnegative f ∈ L1(0, 1), and hence (6.1) follows from (3.5).
In the case when m ≤
n − 1, the validity of (6.1) was proved in [51, Theorem A]. Note that the proof is given in [51] for
Lipschitz domains, and with the space W mX(Ω) in the place of V mX(Ω). However, by Proposition 4.5,
W mX(Ω) = V mX(Ω) if Ω is a John domain, since (4.16) is fulfilled for any such domain. Moreover,
the Lipschitz property of the domain is immaterial, since the proof does not involve any property of
the boundary and hence applies, in fact, to any open set Ω.
(cid:3)
Proof of Theorem 6.2. By Theorem 6.1, every John domain has the property that (5.14) im-
plies (5.13). Consequently, the conclusion follows from Corollary 5.9.
(cid:3)
Proof of Theorem 6.3. The assertion is a consequence of Corollary 5.10.
(cid:3)
The following result provides us with model Euclidean domains of revolution in Rn in the class Jα.
It is an easy consequence of a special case of [67, Section 5.3.3]. In the statement, ωn−1 denotes the
Lebesgue measure of the unit ball in Rn−1.
Proposition 10.1. (i) Given α ∈ [ 1
n′ , 1), define ηα : [0,
1
ηα(r) = ω
n−1
− 1
n−1 (1 − (1 − α)r)
α
(1−α)(n−1)
1−α ] → [0,∞) as
for r ∈ [0,
1
1−α ].
Let Ω be the Euclidean domain in Rn given by
Ω = {(x′, xn) ∈ Rn : x′ ∈ Rn−1, 0 < xn < 1
1−α ,x′ < ηα(xn)}.
Then Ω = 1, and
(10.1)
IΩ(s) ≈ sα
for s ∈ [0, 1
2 ].
(ii) Define η1 : [0,∞) → [0,∞) as
η1(r) = ω
− 1
n−1 e− r
n−1
n−1
Let Ω be the Euclidean domain in Rn given by
for r ≥ 0.
Ω = {(x′, xn) ∈ Rn : x′ ∈ Rn−1, xn > 0,x′ < η1(xn)}.
Then Ω = 1, and
(10.2)
IΩ(s) ≈ s
for s ∈ [0, 1
2 ].
Proof of Theorem 6.4. The Sobolev embedding (6.9) and the Poincar´e inequality (6.10) are equiv-
alent, owing to Theorem 4.4. If α ∈ [ 1
n′ , 1), then inequality (6.7) implies (6.9) and (6.10), via Corollary
5.8, whereas if α = 1, then inequality (6.8) implies (6.9) and (6.10) via Theorem 5.1.
It thus remains to exhibit a domain Ω ∈ Jα such that the Sobolev embedding (6.9) implies either
(6.7), or (6.8), according to whether α ∈ [ 1
n′ , 1) or α = 1.
42
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
If α ∈ [ 1
n′ , 1), let Ω be the set given by Proposition 10.1, Part (i), whereas, if α = 1, let Ω be the set
given by Proposition 10.1, Part (ii). By either (10.1) or (10.2), one has that Ω ∈ Jα. Consequently,
embedding (6.9) entails that there exists a constant C such that
for every u ∈ V mX(Ω). Let us fix any nonnegative function f ∈ X(0, 1), and define u : Ω → [0,∞) as
(10.3)
where Mα is given by
kukY (Ω) ≤ C(cid:16)k∇mukX(Ω) +
1 Z 1
1
rα
f (rm)
rα
m
rm−1
r1
Mα(xn)
1
rα
2
u(x) =Z 1
. . .Z 1
Mα(r) =((1 − (1 − α)r)
e−r
1
1−α
m−1Xk=0
k∇kukL1(Ω)(cid:17)
drm drm−1 . . . dr1
for x ∈ Ω,
1
1−α ],
for r ∈ [0,
for r ∈ [0,∞),
if α ∈ [ 1
if α = 1.
n′ , 1),
The function u is m-times weakly differentiable in Ω, and, since −M ′
α = (Mα)α,
∇ku(x) =
∂ku
∂xk
n
(x) =Z 1
Mα(xn)
1
rα
k+1Z 1
rk+1
1
rα
k+2
. . .Z 1
rm−1
f (rm)
rα
m
drm drm−1 . . . drk+1
for x ∈ Ω,
for k = 1, . . . , m − 1, and
Moreover, on setting Lα = 1
n′ , 1), and Lα = ∞ if α = 1, we have that
(x) = f (Mα(xn))
for a.e. x ∈ Ω.
Thus,
(10.4)
(10.5)
(10.7)
∇mu(x) =
∂mu
∂xm
n
1−α if α ∈ [ 1
t
{(x′, xn) ∈ Ω : xn > t} = ωn−1Z Lα
=Z Lα
t −M ′
. . .Z 1
1
rα
2
1
rα
rα
m
rm−1
r1
s
f (rm)
u∗(s) =Z 1
∇ku∗(s) =Z 1
s
1 Z 1
k+1Z 1
1
rα
rk+1
ηα(r)n−1 dr =Z Lα
t
α(r) dr = Mα(t)
Mα(r)α dr
for t ∈ (0, Lα).
drm drm−1 . . . dr1
for s ∈ (0, 1),
1
rα
k+2
. . .Z 1
rm−1
f (rm)
rα
m
drm drm−1 . . . drk+1
for s ∈ (0, 1),
for 1 ≤ k ≤ m − 1, and
(10.6)
Equation (10.6) ensures that u ∈ V mX(Ω). On the other hand, by (10.3) and (10.4) -- (10.6),
∇mu∗(s) = f ∗(s)
for s ∈ (0, 1).
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
s
1
rα
1 Z 1
r1
1
rα
2
. . .Z 1
rm−1
≤ CkfkX(0,1) + C
rα
m
f (rm)
drm drm−1 . . . dr1(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
. . .Z 1
k+1Z 1
m−1Xk=0Z 1
0 Z 1
1
rα
k+2
1
rα
rk+1
s
f (rm)
rα
m
drm drm−1 . . . drk+1 ds.
rm−1
Subsequent applications of Fubini's Theorem tells us that
HIGHER-ORDER EMBEDDINGS
43
(10.8) Z 1
s
1
rα
k+1Z 1
rk+1
1
rα
k+2
. . .Z 1
rm−1
f (rm)
rα
m
drm drm−1 . . . drk+1 =
=
1
(m − k − 1)!Z 1
s
1
rα(cid:18)Z r
s
dt
tα(cid:19)m−k−1
f (r)dr
for s ∈ (0, 1).
By (10.8), (3.5) and (8.8) applied with I(t) = tα, one has that
(10.9)
drm drm−1 . . . drk+1 ds
Z 1
0 Z 1
s
rk+1
f (rm)
1
rα
1
rα
k+2
k+1Z 1
. . .Z 1
rα(cid:18)Z r
(m − k − 1)!Z 1
0 Z 1
≤ CkH m−k
tα(cid:19)m−k−1
fk(L1)m−k,I (0,1) ≤ C ′kfkL1(0,1).
rα
m
rm−1
dt
=
1
1
s
s
I
f (r)dr ds = kH m−k
I
fkL1(0,1)
for k = 0, . . . , m − 1, for some constants C and C ′.
(6.7) follows from (10.7), (10.8) and (10.9), via Proposition 8.6, Part (i).
When α = 1, inequality (6.8) follows from (10.7), (10.8) and (10.9). When α ∈ [ 1
n′ , 1), inequality
(cid:3)
Theorem 10.2. Let p, q ∈ [1,∞] and α ∈ R be such that one of the conditions in (3.8) is satisfied.
Let I : [0, 1] → [0,∞) be a non-decreasing function such that
t
I(t) is non-decreasing. Then
kRI f ∗k(Lp,q;α)′(0,1) ≈(cid:13)(cid:13)(cid:13)t
1
p′ − 1
q′(cid:0)log 2
t(cid:1)−α
RI f ∗(t)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
for every f ∈ M+(0, 1), up to multiplicative constants depending on p, q, α.
Proof. Fix f ∈ M+(0, 1). By Theorem 9.5, applied with m = 0 and X(0, 1) = Lp,q;α(0, 1), and
Holder's inequality, there exists a universal constant C such that
kRIf ∗(t)k(Lp,q;α)′(0,1) ≤ C kRI f ∗(t)k(Lp,q;α)′
kgkLp,q;α(0,1)≤1Z 1
q (log 2
g∗(t)t
t )αt
p′ − 1
p − 1
sup
0
1
1
= C
d(0,1) = C
q′ (log 2
sup
kgkLp,q;α(0,1)≤1Z 1
t )−αRI f ∗(t) dt ≤ Ckt
0
g∗(t)RI f ∗(t) dt
1
p′ − 1
q′ (log 2
t )−αRI f ∗(t)kLq′ (0,1).
In order to prove the reverse inequality, assume first that either 1 < p < ∞ or p = q = 1 and α ≥ 0
or p = q = ∞ and α ≤ 0. By Theorem 9.5 (with m = 0) and (3.9),
1
p′ − 1
q′ (log 2
t )−αRI f ∗(t)kLq′ (0,1) ≤ kt
kt
= kGI fkLp′,q′;−α(0,1) ≈ kGI fk(Lp,q;α)′(0,1) ≈ kRI f ∗k(Lp,q;α)′(0,1),
t )−α sup
q′ (log 2
t≤s≤1
RI f ∗(t)kLq′ (0,1)
1
p′ − 1
where the last but one equivalence holds up to multiplicative constants depending on p, q, α.
44
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
1
0
q < 0. We have that
It remains to consider the case when p = ∞, q ∈ [1,∞) and α + 1
kRI f ∗(t)k(Lp,q;α)′(0,1) ≈ kRI f ∗(t)kL(1,q′;−α−1)(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)t1− 1
(RI f ∗)∗ (s) ds(cid:13)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
tZ t
q′ (cid:0)log 2
t(cid:1)−α−1 1
≥(cid:13)(cid:13)(cid:13)(cid:13)t1− 1
dr(cid:13)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
RI f ∗(s) ds(cid:13)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
=(cid:13)(cid:13)(cid:13)(cid:13)t− 1
f ∗(r)Z t
t(cid:1)−α−1Z t
tZ t
t(cid:1)−α−1 1
q′ (cid:0)log 2
q′(cid:0)log 2
≥(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≥(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
dr(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
s(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
f ∗(r) drZ t
f ∗(r)Z t
t(cid:1)−α−1Z t2
t(cid:1)−α−1Z t2
q′ (cid:0)log 2
q′(cid:0)log 2
2(cid:13)(cid:13)(cid:13)(cid:13)χ(0, 1
t(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)Lq′ (0,1) ≥ 1
≥(cid:13)(cid:13)(cid:13)(cid:13)t2− 1
f ∗∗(t2)(cid:13)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
I(t2)(cid:18)log
t(cid:1)−α t2
t(cid:1)−α−1
q′(cid:0)log 2
q′ (cid:0)log 2
2 )(t)t2− 1
RI f ∗(t)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
≥ C(cid:13)(cid:13)(cid:13)t1− 1
q′ (cid:0)log 2
t(cid:1)−α
for some constant C = C(α, q). The proof is complete.
ds
I(s)
ds
I(s)
f ∗∗(t2)
t− 1
I(t2)
I(s)
ds
t2
(cid:3)
1
t2
s
−
t
0
0
0
0
,
r
r
Proof of Theorem 6.8. By Corollary 5.9,
for f ∈ M+(0, 1). If m(1 − α) < 1 and p <
1 < r < ∞), then this, (3.10) and (3.9) yield
kfk((Lp)m,α)′(0,1) = ksm(1−α)f ∗∗(s)kLp′ (0,1)
m(1−α) and r is given by 1
1
r = 1
p − m(1 − α) (note that
(Lp)m,α(0, 1) = (L(r′,p′))′(0, 1) = (Lr′,p′
)′(0, 1) = Lr,p(0, 1).
Since p < r, we have Lr,p(0, 1) → Lr(0, 1), and the claim follows. If m(1 − α) < 1 and p =
then, by (3.6), Lr′
1
m(1−α) ,
(0, 1) → L(1,p)(0, 1) for every r ∈ [1,∞), and hence
(Lp)m,α(0, 1) = (L(1,p′))′(0, 1) → Lr(0, 1).
Finally, if either m(1− α) ≥ 1, or m(1− α) < 1 and p >
thus follows from Corollary 5.5.
1
m(1−α) , then (5.9) is satisfied. The conclusion
(cid:3)
Proof of Theorem 6.9. First, assume that either m(1 − α) ≥ 1, or m(1 − α) < 1, p =
q = 1, or m(1 − α) < 1 and p >
I(t) = tα and X(0, 1) = Lp,q(0, 1). Hence, by Corollary 5.5, V mLp,q(Ω) → L∞(Ω).
J(t) = t−m(1−α)+1 for t ∈ [0, 1]. Then J is a non-decreasing function such that
on (0, 1). Given f ∈ M+(0, 1), by Theorem 10.2 (with α = 0),
Next, assume that m(1 − α) < 1, and either 1 ≤ p <
m(1−α) and
In each of these cases, condition (5.9) is satisfied with
m(1−α) and q > 1. Set
t
J(t) is non-decreasing
m(1−α) , or p =
m(1−α) .
1
1
1
1
kfk(Lp,q
m,α)′(0,1) =(cid:13)(cid:13)(cid:13)(cid:13)s−1+m(1−α)Z s
p′ − 1
0
1
f ∗(r) dr(cid:13)(cid:13)(cid:13)(cid:13)(Lp,q)′(0,1)
p′ − 1
1
q′ RJ f ∗(t)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
=(cid:13)(cid:13)(cid:13)t
≈(cid:13)(cid:13)(cid:13)t
= kRJ f ∗k(Lp,q)′(0,1)
q′ +m(1−α)f ∗∗(t)(cid:13)(cid:13)(cid:13)Lq′ (0,1)
where the equivalence holds up to constants depending on p and q, and r′ satisfies 1
p′ + m(1− α).
Owing to (3.10), (3.9) and (3.6), L(r′,q′)(0, 1) = (L
m(1−α) ,
and L(r′,q′)(0, 1) = (L∞,q;−1)′(0, 1) if m(1− α) < 1, p =
m(1−α) and q > 1. The conclusion follows. (cid:3)
Proof of Theorem 6.10. Since (5.1) holds with I(t) = t, in the case when 1 ≤ p < ∞, the assertion
follows from Theorem 7.12 applied with β = 1 (see Section 11 below for the proof of Theorem 7.12).
If p = ∞, Theorem 7.12 has to be combined with an appropriate relation from (3.10).
(cid:3)
1−mp(1−α) ,q)′(0, 1) if m(1−α) < 1 and 1 ≤ p <
r′ = 1
1
1
p
= kfkL(r′,q′)(0,1),
HIGHER-ORDER EMBEDDINGS
45
Proof of Theorem 6.11. Since (5.1) holds with I(t) = t, the conclusion is a consequence of Theo-
rem 7.12 applied with β = 1.
(cid:3)
11. Proofs of the Sobolev embeddings in product probability spaces
This final section is devoted to the proof of the results of Section 7.
Lemma 11.1. Let Φ be as in (7.1). Then:
(i) The function LΦ defined by (7.7) is nondecreasing on [0, 1];
(ii) The inequality
(11.1)
holds for every s ∈ (0, 1
2 ];
(iii) The inequality
sΦ′(cid:0)Φ−1(log(cid:0) 1
s(cid:1)(cid:1)(cid:1) ≤ LΦ(s) ≤ 2sΦ′(cid:0)Φ−1(log(cid:0) 1
s(cid:1)(cid:1)(cid:1)
(11.2)
Φ−1(s)
2s ≤
1
Φ′(Φ−1(s)) ≤
Φ−1(s) − Φ−1(t)
Φ−1(s)
s
≤
s − t
holds whenever 0 ≤ t < s < ∞.
Proof. (i) The convexity of Φ and the concavity of √Φ imply that
0 ≤ Φ′′(t) ≤
Φ′(t)2
2Φ(t)
for t > 0.
Therefore,
L′
Φ(s) = Φ′(Φ−1(log 2
s )) −
Φ′′(Φ−1(log 2
Φ′(Φ−1(log 2
s ))
s )) ≥
Hence, (i) follows.
Φ′′(Φ−1(log 2
Φ′(Φ−1(log 2
s ))
s )) (cid:0)2 log 2
s − 1(cid:1) > 0 for s ∈ (0, 1).
(ii) The first inequality in (11.1) trivially holds, since both Φ′ and Φ−1 are non-decreasing functions.
The latter inequality follows from (i) and from the fact that
2sΦ′(cid:0)Φ−1(log(cid:0) 1
s(cid:1)(cid:1)(cid:1) = LΦ(2s)
for s ∈ (0, 1
2 ].
(iii) Let 0 ≤ r1 < r2 < ∞. Owing to the convexity of Φ and the fact that Φ(0) = 0, we obtain that
Φ(r2)
r2 ≤
Φ(r2) − Φ(r1)
(11.3)
≤ Φ′(r2).
Furthermore, by the concavity of √Φ and the fact thatpΦ(0) = 0,
2pΦ(r2) ≤ pΦ(r2)
(√Φ)′(r2) =
and, therefore,
r2 − r1
Φ′(r2)
,
r2
(11.4)
Φ′(r2) ≤
2Φ(r2)
r2
.
Let 0 ≤ t < s < ∞.
inequalities (11.3) and (11.4) yield
If we set r1 = Φ−1(t), r2 = Φ−1(s), then 0 ≤ r1 < r2 < ∞. Hence,
s
Φ−1(s) ≤
s − t
Φ−1(s) − Φ−1(t) ≤ Φ′(Φ−1(s)) ≤
2s
Φ−1(s)
.
Assertion (11.2) follows.
(cid:3)
46
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Let m ∈ N. We define the operator P m
Φ from M+(0, 1) into M+(0, 1) by
Φ f (t) = Φ−1(log 2
t )
and for f ∈ M+(0, 1). Moreover, let H m
I = LΦ, namely
t !mZ 1
s (cid:16)log
t(cid:17)m−1
log 2
f (s)
P m
ds
LΦ
s
t
for t ∈ (0, 1),
be the operator defined as in (8.3) (see also (8.5)), with
for t ∈ (0, 1),
t
t
1
dr
ds
f (s)
H m
LΦf (t) =
LΦ(s)(cid:18)Z s
(m − 1)!Z 1
LΦ(r)(cid:19)m−1
and for f ∈ M+(0, 1). Observe that, by the change of variables τ 7→ Φ−1(cid:0)log 2
r(cid:1)(cid:1) Z r
tΦ′(cid:0)Φ−1(cid:0)log 2
s(cid:1) − Φ−1(cid:0)log 2
r(cid:1)(cid:1)m−1
r(cid:1)(cid:1)
rΦ′(cid:0)Φ−1(cid:0)log 2
rΦ′(cid:0)Φ−1(cid:0)log 2
= (cid:0)Φ−1(cid:0)log 2
LΦ(r)(cid:18)Z r
LΦ(t)(cid:19)m−1
(11.5)
dt
dt
=
1
1
s
s
In particular, this yields
(11.6)
t(cid:1), we have
t(cid:1)(cid:1)!m−1
for 0 < s ≤ r < 1.
H m
LΦf (t) =
1
(m − 1)!Z 1
t
f (s)
sΦ′(Φ−1(log 2
s ))(cid:18)Φ−1(cid:18)log
2
t(cid:19) − Φ−1(cid:18)log
2
s(cid:19)(cid:19)m−1
ds
for t ∈ (0, 1),
and f ∈ M+(0, 1).
A connection between the operators P m
Φ and H m
LΦ
is described in the following proposition.
Proposition 11.2. Suppose that Φ is as in (7.1), m ∈ N and f ∈ M+(0, 1). Then
(11.7)
1
P m
Φ f (t) ≤ H m
LΦf (t)
for t ∈ (0, 1).
2m(m − 1)!
Moreover, if f is nonincreasing on (0, 1), then
H m
LΦf (t) ≤
1
P m
Φ f (t)
for t ∈ (0, 1).
(11.8)
(m − 1)!
Proof. Let f ∈ M+(0, 1). Since the function s 7→
first inequality in (11.2) we obtain that
1
Φ′(Φ−1(log 2
s ))
is nondecreasing on (0, 1), from the
H m
LΦf (t) =
1
(m − 1)!Z 1
t
f (s)
sΦ′(Φ−1(log 2
dr
r ))!m−1
ds
t
s )) Z s
t ))(cid:1)mZ 1
t ))(cid:1)mZ 1
t !mZ 1
t )
t
t
t
for t ∈ (0, 1).
ds
f (s)
f (s)
rΦ′(Φ−1(log 2
t
s
dr
s (cid:18)Z s
s (cid:16)log
s (cid:16)log
r (cid:19)m−1
t(cid:17)m−1
t(cid:17)m−1
s
f (s)
ds
ds
≥
=
≥
=
1
(m − 1)!
1
(m − 1)!
1
(m − 1)!
1
(m − 1)!
1
1
(cid:0)Φ′(Φ−1(log 2
(cid:0)Φ′(Φ−1(log 2
2m Φ−1(log 2
1
2m P m
Φ f (t)
log 2
1
Now, assume that f is nonincreasing on (0, 1). In the special case when f is a characteristic function
HIGHER-ORDER EMBEDDINGS
47
of an open interval, namely, f = χ(0,b) for some b ∈ (0, 1], equation (11.6) tells us that
and
H m
LΦ(χ(0,b))(t) =
P m
Φ (χ(0,b))(t) = χ(0,b)(t)
2
1
1
m!
χ(0,b)(t)(cid:18)Φ−1(cid:18)log
m Φ−1(log 2
!m
t(cid:1) − Φ−1(cid:0)log 2
b(cid:1)
t − log 2
t(cid:19) − Φ−1(cid:18)log
t !m(cid:18)log
t !m
≤ Φ−1(log 2
log 2
log 2
t )
t )
b
2
t − log
2
b(cid:19)(cid:19)m
b(cid:19)m
2
for t ∈ (0, b).
for t ∈ (0, 1). By the last inequality in (11.2),
Φ−1(cid:0)log 2
log 2
Hence,
(11.9)
H m
LΦ(χ(0,b)) ≤
1
(m − 1)!
P m
Φ (χ(0,b)).
Assume next that f is a nonnegative non-increasing simple function on (0, 1). Then there exist k ∈
i=1 aiχ(0,bi)
N, nonnegative numbers a1, a2, . . . , ak ∈ R and 0 < b1 < b2 < ··· < bk ≤ 1 such that f =Pk
a.e. on (0, 1). Hence, owing to (11.9),
H m
LΦf (t) =
kXi=1
aiH m
LΦ(χ(0,bi))(t) ≤
1
(m − 1)!
aiP m
Φ (χ(0,bi))(t) =
1
(m − 1)!
P m
Φ f (t)
for t ∈ (0, 1).
kXi=1
Finally, if f ∈ M+(0, 1) is nonincreasing on (0, 1), then there exists a sequence fk of nontrivial
nonnegative nonincreasing simple functions on (0, 1) such that fn ↑ f . Clearly,
Φ f (t)
Φ fn(t) =
H m
P m
P m
1
1
LΦf (t) = lim
n→∞
H m
LΦfn(t) ≤
lim
n→∞
(m − 1)!
(m − 1)!
for t ∈ (0, 1),
(cid:3)
whence (11.8) follows.
Proposition 11.2 has an important consequence.
Proposition 11.3. Let Φ be as in (7.1), let m ∈ N and let k·kX(0,1) and k·kY (0,1) be rearrangement-
invariant function norms. Then
P m
Φ : X(0, 1) → Y (0, 1)
if and only if H m
LΦ : X(0, 1) → Y (0, 1).
Proof. By (11.7), the boundedness of the operator H m
LΦ
if P m
Φ is bounded from X(0, 1) into Y (0, 1) then, in particular, there exists a constant C such that
implies the boundedness of P m
Φ . Conversely,
for every nonnegative non-increasing function f ∈ X(0, 1). Combining this inequality with (11.8), we
obtain that
kP m
Φ fkY (0,1) ≤ CkfkX(0,1)
for every nonnegative non-increasing f ∈ X(0, 1). In view of Corollary 9.8, this is equivalent to the
boundedness of H m
LΦ
from X(0, 1) into Y (0, 1).
(cid:3)
kH m
LΦfkY (0,1) ≤ CkfkX(0,1)
Proof of Theorem 7.1. Properties (ii) and (iii) are equivalent, by Proposition 4.4. Let us show that
(i) and (ii) are equivalent as well. First, assume that (i) is satisfied. Owing to Proposition 11.3, there
exists a constant C such that
Z 1
t
f (s)
LΦ(s)(cid:18)Z s
t
dr
LΦ(r)(cid:19)m−1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
≤ CkfkX(0,1)
48
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
for every nonnegative f ∈ X(0, 1). By Lemma 11.1 (i), the function LΦ is non-decreasing on [0, 1].
Furthermore, condition (5.2) is clearly satisfied with I = LΦ. Thanks to these facts and to (7.8), the
assumptions of Theorem 5.1 are fulfilled with (Ω, ν) = (Rn, µΦ,n) and I = LΦ. Hence, (ii) follows.
It only remains to prove that (ii) implies (i). Assume that (ii) holds, namely, there exists a constant
C, such that
(11.10)
kukY (Rn,µΦ,n) ≤ C(cid:16)k∇mukX(Rn,µΦ,n) +
m−1Xk=0
k∇kukL1(Rn,µΦ,n)(cid:17)
Given any nonnegative function f ∈ X(0, 1) such that f (s) = 0 if s ∈ ( 1
2 , 1), consider the function
for every u ∈ V mX(Rn, µΦ,n).
u : Rn → R defined as
FΦ(r1)Z 1
u(x) =Z 1
(x) =Z 1
∇ku(x) =
∂ku
∂xk
1
H(x1)
H(x1)
r1
1
rm−1
1
FΦ(r2)
. . .Z 1
FΦ(rk+1)Z 1
1
rk+1
where H is given by (7.5). Note that, since H ′(t) = −FΦ(H(t)), then
for a.e. x ∈ Rn, for k = 1, . . . , m − 1, and
∂mu
∂xm
1
∇mu(x) =
(x) = f (H(x1))
for a.e. x ∈ Rn.
f (rm)
FΦ(rm)
drm drm−1 . . . dr1
for x ∈ Rn,
1
FΦ(rk+2)
. . .Z 1
rm−1
f (rm)
FΦ(rm)
drm drm−1 . . . drk+1
1
FΦ(rk+2)
. . .Z 1
rm−1
f (rm)
FΦ(rm)
drm drm−1 . . . drk+1
for s ∈ (0, 1),
Thus, by (7.6),
(11.11)
∇ku∗(s) =Z 1
s
for k = 0, . . . , m − 1, and
(11.12)
∗
∂mu
∂xm
1 (cid:12)(cid:12)(cid:12)(cid:12)
1
rk+1
FΦ(rk+1)Z 1
∇mu∗(s) =(cid:12)(cid:12)(cid:12)(cid:12)
. . .Z 1
m−1Xk=0Z 1
0 Z 1
rm−1
1
s
FΦ(r2)
(s) = f ∗(s)
for s ∈ (0, 1).
By (11.12), u ∈ V mX(Rn, µΦ,n). From (11.10), (11.11) and (11.12) we thus deduce that
(11.13)
1
FΦ(r1)Z 1
r1
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
s
≤ CkfkX(0,1) + C
f (rm)
FΦ(rm)
1
FΦ(rk+1)Z 1
drm drm−1 . . . dr1(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
. . .Z 1
FΦ(rk+2)
rm−1
rk+1
1
f (rm)
FΦ(rm)
drm drm−1 . . . drk+1 ds.
Owing to Fubini's Theorem, (7.8) and (11.5),
(11.14)
Z 1
s
1
FΦ(r2)
1
r1
FΦ(r1)Z 1
≈Z 1
s
f (r)
LΦ(r)(cid:18)Z r
s
rm−1
. . .Z 1
LΦ(t)(cid:19)m−1
dt
f (rm)
FΦ(rm)
dt
f (r)
drm drm−1 . . . dr1 ≈Z 1
f (r)(cid:0)Φ−1(cid:0)log 2
FΦ(t)(cid:19)m−1
FΦ(r)(cid:18)Z r
s(cid:1) − Φ−1(cid:0)log 2
r(cid:1)(cid:1)m−1
r(cid:1)(cid:1)
rΦ′(cid:0)Φ−1(cid:0)log 2
s
s
dr =Z 1
s
dr
dr for s ∈ (0, 1).
HIGHER-ORDER EMBEDDINGS
49
Note that the second equivalence makes use of the fact that f vanishes in ( 1
by (11.14) (with m replaced with m − k), (3.5), and (8.8) (with I replaced with LΦ),
(11.15)
drm drm−1 . . . drk+1 ds
1
2 , 1). On the other hand,
s
Z 1
0 Z 1
FΦ(rk+1)Z 1
≈Z 1
0 Z 1
≤ CkH m−k
f (r)
LΦ
s
rk+1
FΦ(rk+2)
1
. . .Z 1
LΦ(t)(cid:19)m−k−1
dt
rm−1
LΦ(r)(cid:18)Z r
fk(L1)m−k,LΦ (0,1) ≤ C ′kfkL1(0,1) ≤ C ′′kfkX(0,1)
dr ds ≈ kH m−k
LΦ
s
fkL1(0,1)
f (rm)
FΦ(rm)
for some constants C, C ′ and C ′′. From inequalities (11.13) -- (11.15), we deduce that there exists a
constant C such that
for every nonnegative function f ∈ X(0, 1) such that f (s) = 0 if s ∈ ( 1
each such function f we also have
2 , 1). By Proposition 11.2, for
(11.16)
(11.17)
(11.18)
Finally, assume that f is any nonnegative function from X(0, 1) (which need not vanish in ( 1
Then, by the boundedness of the dilation operator on Y (0, 1), there exists a constant C such that
2 , 1)).
≤ C kfkX(0,1)
≤ 2mCkfkX(0,1).
f (r)(cid:0)Φ−1(cid:0)log 2
t )
t !mZ 1
t
s
f (s)
dr(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
r(cid:1)(cid:1)m−1
s(cid:1) − Φ−1(cid:0)log 2
rΦ′(cid:0)Φ−1(cid:0)log 2
r(cid:1)(cid:1)
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
t(cid:17)m−1
s (cid:16)log
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
t(cid:17)m−1
s (cid:16)log
t !mZ 1
2t(cid:17)m−1
s (cid:16)log
f (s)
t )
2t
s
s
t )
f (s)
t !mZ 1
2 )(t) Φ−1(log 1
log 1
t
s
log 2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Z 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Φ−1(log 2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Φ−1(log 2
≤ C(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
χ(0, 1
log 2
.
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
Furthermore, since
Φ−1(log 1
t )
log 1
t
Φ−1(log 2
t )
log 1
t
≤
≤
2Φ−1(log 2
t )
log 2
t
for t ∈ (0, 1
2 ),
from inequality (11.16) with f replaced with χ(0, 1
operator, we obtain
2 )(t)f (2t), and the boundedness of the dilation
t )
f (s)
log 1
χ(0, 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2 )(t) Φ−1(log 1
≤ 2m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
t !mZ 1
2 )(t) Φ−1(log 2
s
2 )(t)f (2t)kX(0,1) ≤ C ′′kfkX(0,1)
s (cid:16)log
t !mZ 1
≤ C ′kχ(0, 1
χ(0, 1
log 2
t )
2t
t
2
f (2s)
s
2t(cid:17)m−1
(cid:16)log
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Y (0,1)
t(cid:17)m−1
s
for some constants C ′ and C ′′ independent of f . Coupling (11.17) with (11.18) yields (7.9).
Proof of Theorem 7.3. Set J(s) = s for s ∈ [0, 1]. Then condition (5.2) is obviously fulfilled with
I = J. The norm k · keXm,J (0,1) is thus well defined and, moreover, k · keXm(0,1) = k · kXm,J (0,1).
Therefore, Proposition 8.3 tells us that k · keXm(0,1) is a rearrangement-invariant function norm. We
shall now verify that k · kXm,Φ(0,1) is a rearrangement-invariant function norm as well. The first two
properties in (P1) and properties (P2) and (P3) are straightforward consequences of the corresponding
properties for k · keXm(0,1). To prove the triangle inequality, fix f , g ∈ M+(0, 1). By (3.1), R s
0 (f +
(cid:3)
50
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
s
Φ−1(log 2
yields that
0 (f ∗(r) + g∗(r)) dr for s ∈ (0, 1). We observe that for each t ∈ (0, 1), the function
is nonnegative and non-increasing on (0, 1). The Hardy's lemma therefore
g)∗(r) dr ≤ R s
s 7→ χ(0,t)(s)(cid:16) log 2
s )(cid:17)m
0 log 2
s )!m
Z t
(f + g)∗(s) ds ≤Z t
for t ∈ (0, 1). The triangle inequality now follows using the Hardy - Littlewood - P´olya principle and
the triangle inequality for k · keXm(0,1).
f ∗(s) + log 2
0 log 2
g∗(s)! ds
s )!m
s )!m
One has that
Φ−1(log 2
Φ−1(log 2
Φ−1(log 2
s
s
s
where the equality is a consequence of Theorem 6.11. Thus, there exists a constant C such that
1
Φ−1(log 2
s
s
exp L
Φ−1(log 2
k1kXm,Φ (0,1) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
log 2
≈ C(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Φ−1(log 2)(cid:19)m
kfkXm,Φ(0,1) ≥(cid:18) log 2
m (0, 1) = (L∞)m(0, 1) → eXm(0, 1),
≤ C(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
s )!m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eXm(0,1)
log 2
s ))m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L∞(0,1)
kf ∗keXm(0,1) ≥(cid:18) C log 2
(Φ−1(log 2
=
C
1
(Φ−1(log 2))m < ∞.
s )!m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)exp L
1
m (0,1)
Φ−1(log 2)(cid:19)mZ 1
0
f ∗(s) ds.
This proves (P4).
Finally, by property (P5) for k · keXm(0,1), there exists a positive constant C such that for all f ∈
M+(0, 1),
Therefore, k · kXm,Φ(0,1) satisfies (P5). Since the property (P6) holds trivially, k · kXm,Φ(0,1) is actually
a rearrangement-invariant norm.
It follows from the proof of Theorem 7.1 that the assumptions of Theorem 5.4 are fulfilled with
(Ω, ν) = (Rn, µΦ,n) and I = LΦ. Therefore, k · kXm,LΦ (0,1) is the optimal rearrangement-invariant
target function norm for k· kX(0,1) in the Sobolev embedding (7.10). Thus, the proof will be complete
if we show that Xm,Φ(0, 1) = Xm,LΦ(0, 1). We have that
(11.19)
(by Theorem 9.5)
kfkX ′
m,LΦ
LΦ f ∗kX ′(0,1) ≈ kRm
LΦf ∗(t) dt
g∗(t)Rm
LΦf ∗kX ′
d(0,1)
LΦg∗(t) dt
0
0
=
=
≈
sup
sup
sup
f ∗(t)P m
f ∗(t)H m
(0,1) = (m − 1)!kRm
kgkX(0,1)≤1Z 1
kgkX(0,1)≤1Z 1
kgkX(0,1)≤1Z 1
kgkX(0,1)≤1Z 1
kgkX(0,1)≤1Z 1
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
J f ∗(s) Φ−1(log 2
g∗(t)Rm
log 2
Rm
sup
sup
≈
=
0
0
0
Φ g∗(t) dt
(by Proposition 11.2)
f ∗(t) Φ−1(log 2
t !m
log 2
t )
H m
J g∗(t) dt
s )
s !m! (t) dt
for f ∈ L1(0, 1),
log 2
J f ∗(s) Φ−1(log 2
s !m!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′
d(0,1)
s )
up to multiplicative constants depending on m.
HIGHER-ORDER EMBEDDINGS
51
We now claim that, given f ∈ L1(0, 1), there exists a non-decreasing function I on [0, 1] fulfill-
ing (5.2) and a function h ∈ M+(0, 1) such that
up to multiplicative constants depending on m. Indeed, let s0 ∈ (0, 1) be chosen in such a way that
s )
log 2
≈ RI h∗(s)
s !m
f ∗(s) Φ−1(log 2
s(cid:1)m+1 is non-decreasing on (0, s0). Then we set
1
for s ∈ (0, 1),
I(s) =
f ∗(s)
for s ∈ (0, 1]
and I(0) = 0,
(11.20)
and
the function s 7→ s(cid:0) log 2
(Φ−1(log 2
(Φ−1(log 2
s ))m−1(cid:18)Φ−1(log 2
s ))m−1(cid:18)Φ−1(log 2
s(log 2
s )m+1
s )−
)m+1
s0(log 2
s0
log 2
s
s )−
Φ′(Φ−1(log 2
log 2
s
Φ′(Φ−1(log 2
s ))(cid:19)
s ))(cid:19)
,
,
s ∈ (0, s0]
s ∈ (s0, 1).
h(s) =
It follows from (11.2) that the function h is non-negative on (0, 1). To verify (11.20) we first show that
h is non-increasing on (0, 1). The function Φ−1 is clearly non-decreasing on (0,∞). Furthermore, we
deduce from the convexity of Φ that the function s 7→ Φ−1(s)−
Φ′(Φ−1(s)) is non-decreasing on (0,∞).
Altogether, this ensures that
s(cid:19) −
s 7→(cid:18)Φ−1(cid:18)log
s ))!
Φ′(Φ−1(log 2
log 2
s
2
2
s
is non-increasing on (0, 1). By the definition of s0, the function
s(cid:19)(cid:19)m−1 Φ−1(cid:18)log
s 7→(s(log 2
s0(log 2
s0
s )m+1,
)m+1,
s ∈ (0, s0]
s ∈ (s0, 1)
is non-decreasing (and continuous) on (0, 1), and therefore, in particular, h = h∗.
Consequently, we have
f ∗(s) Φ−1(log 2
s !m
log 2
s )
=
≈
m
I(s)Z s
I(s)Z s
1
0
0 (cid:0)Φ−1(log 2
r )(cid:1)m−1(cid:16)Φ−1(log 2
r(log 2
r )m+1
r ) −
log 2
r
Φ′(Φ−1(log 2
r ))(cid:17)
dr
h∗(r) dr = RIh∗(s)
for s ∈ (0, 1),
up to multiplicative constants depending on m. This proves (11.20). Furthermore, it can be easily
verified that the function I fulfils also the remaining required properties.
Coupling (11.19) with (11.20) entails that, for the fixed function f ,
(11.21)
kfkX ′
m,LΦ
(0,1) ≈ kRm
J RI h∗kX ′
d(0,1),
up to multiplicative constants depending on m.
Now, the same proof as that of Theorem 9.5 yields that
(11.22)
d(0,1) ≈ kRm
up to multiplicative constants still depending only on m.
J RI h∗kX ′
kRm
J (RI h∗)∗kX ′(0,1),
52
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
On combining (11.21), (11.22) and (11.20), we obtain that for every f ∈ L1(0, 1),
(11.23)
(·) )
log 2
J f ∗(·) Φ−1(log 2
(·) !m!∗
t !m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eX ′
f ∗(t) Φ−1(log 2
log 2
m(0,1)
t )
(t)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′(0,1)
m,LΦ
(0,1) ≤ 1(cid:27)
t !m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eX ′
f ∗(t) Φ−1(log 2
log 2
t )
,
≤ 1
m(0,1)
f ∗(s)g∗(s) ds : kfkX ′
f ∗(s)g∗(s) ds :(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
t )!m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eXm(0,1)
t
0
0
m,LΦ
Rm
kfkX ′
(0,1) ≈(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≈(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
kgkXm,LΦ (0,1) = sup(cid:26)Z 1
≈ sup
Z 1
≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
g∗(t) log 2
t )!m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)eXm(0,1)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
g∗(t) log 2
= sup(Z 1
g∗(t) log 2
t )!m
≈ sup
t )!m
g∗(t) log 2
Z 1
≤ kgkXm,LΦ (0,1),
Φ−1(log 2
Φ−1(log 2
Φ−1(log 2
Φ−1(log 2
0
0
t
t
t
up to mulplicative constants depending on m.
Conversely,
(11.24)
up to mulplicative constants depending on m. Consequently, by (11.23), we have that, for every
g ∈ M+(0, 1),
f ∗(t) dt : kfkeX ′
m(0,1) ≤ 1)
f ∗(t) log 2
t
Φ−1(log 2
f ∗(t) dt :(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
t )(cid:17)m
t )!m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X ′
m,LΦ
≤ 1
(0,1)
up to multiplicative constants depending on m. Note that the equivalence in (11.24) holds by (11.23)
and the fact that the function t 7→(cid:16) log 2
The proof is complete.
Proof of Proposition 7.5. Since the m-th iteration of the double-star operator g 7→ g∗∗ associates
a function g with 1
r )m−1g∗(r) dr for s ∈ (0, 1), we obtain from the boundedness of the double-
star operator on X ′(0, 1) that
is non-increasing. Hence, Xm,Φ(0, 1) = Xm,LΦ(0, 1).
(cid:3)
0 (log s
Φ−1(log 2
t
sR s
Thus, eXm(0, 1) = X(0, 1). Consequently, the assertion follows from (7.12).
Proof of Theorem 7.6. This is a consequence of Theorem 5.7 and of the fact that Xm,Φ(0, 1) =
Xm,LΦ(0, 1).
(cid:3)
(cid:3)
m(0,1) ≈ kgkX ′(0,1).
kgkeX ′
Proof of Theorem 7.12. Denote X(0, 1) = Lp,q;α(0, 1). We claim that
eXm(0, 1) =(Lp,q;α(0, 1)
L∞,q;α−m(0, 1)
if p < ∞,
if p = ∞.
HIGHER-ORDER EMBEDDINGS
53
Indeed, let p < ∞ and set Φ(t) = t, t ∈ [0,∞). Then, by Remark 7.2,
By (3.9) and (3.10), the operator f 7→ f ∗∗ is bounded on X ′(0, 1). Therefore, by Proposition 7.5,
eXm(0, 1) = Xm,Φ(0, 1).
Xm,Φ(0, 1) = X(0, 1) = Lp,q;α(0, 1).
Now, let p = ∞, and set I(s) = s, s ∈ [0, 1]. Then RI f ∗ = f ∗∗, whence, by Theorem 10.2,
t )−αf ∗∗(t)kLq′ (0,1) = kfkL(1,q′ ;−α)(0,1).
kfk(eX1)′(0,1) = kf ∗∗kX ′(0,1) ≈ kt1− 1
q′ (log 2
Owing to (3.9) and (3.10),
Thus,
(L(1,q′;−α))′(0, 1) = L∞,q;α−1(0, 1).
By making use of Theorem 7.6 combined with Remark 7.2, we obtain that
eX1(0, 1) = L∞,q;α−1(0, 1).
eXm(0, 1) = L∞,q;α−m(0, 1).
The conclusion is now a consequence of Theorem 7.11.
(cid:3)
References
[1] A. Alvino, V. Ferone and G. Trombetti, Moser-type inequalities in Lorentz spaces, Potential Anal. 5 (1996),
273 -- 299.
[2] M. Arino and B. Muckenhoupt, Maximal functions on classical Lorentz spaces and Hardy's inequality with weights
for non-increasing functions, Trans. Amer. Math. Soc. 320 (1990), 727 -- 735.
[3] T. Aubin, Probl`emes isop´erimetriques et espaces de Sobolev, J. Diff. Geom. 11 (1976), 573 -- 598.
[4] F. Barthe, P. Cattiaux and C. Roberto, Interpolated inequalities between exponential and Gaussian, Orlicz hyper-
contractivity and isoperimetry, Rev. Mat. Iberoam. 22 (2006), 993 -- 1067.
[5] F. Barthe, P. Cattiaux and C. Roberto, Isoperimetry between exponential and Gaussian, Electron. J. Probab. 12
(2007), 1212 -- 1237.
[6] T. Bartsch, T. Weth and M. Willem, A Sobolev inequality with remainder term and critical equations on domains
with topology for the polyharmonic operator, Calc. Var. Partial Differential Equations 18 (2003), 253 -- 268.
[7] C. Bennett and K. Rudnick, On Lorentz -- Zygmund spaces, Dissertationes Math. 175 (1980), 1-72.
[8] C. Bennett and R. Sharpley, Interpolation of Operators, Pure and Applied Mathematics Vol. 129, Academic Press,
Boston 1988.
[9] S.G. Bobkov and C. Houdr´e, Isoperimetric constants for product probability measures, Ann. Prob. 25 (1997),
184-205.
[10] S.G. Bobkov and C. Houdr´e, Some connections between isoperimetric and Sobolev-type inequalities, Mem. Am.
Math. Soc. 129 (1997), viii+111.
[11] S.G. Bobkov and M. Ledoux, From Brunn-Minkowski to sharp Sobolev inequalities, Ann. Mat. Pura Appl. 187
(2008), 389 -- 384.
[12] B. Bojarski, Remarks on Sobolev imbedding inequalities, in Proc. Conference on Complex Analysis, Joensuu 1987,
Lecture Notes in Math. 1351, Springer, 1988.
[13] C. Borell, The Brunn-Minkowski inequality in Gauss space, Invent. Math. 30 (1975), 207-216.
[14] H. Br´ezis and E. Lieb, Sobolev inequalities with remainder terms, J. Funct. Anal. 62 (1985), 73 -- 86.
[15] S. Buckley and P. Koskela, Sobolev-Poincar´e implies John, Math. Res. Lett. 2 (1995), 577 -- 593.
[16] S. Buckley and P. Koskela, Criteria for embeddings of Sobolev-Poincar´e type, Int. Math. Res. Not. 18 (1996),
881 -- 902.
[17] Yu.D. Burago and V.A. Zalgaller, Geometric inequalities, Springer, Berlin, 1988.
[18] A. Canete, M. Miranda Jr. and D. Vittone, Some isoperimetric problems in planes with density, J. Geom. Anal.
20 (2010), 243 -- 290.
[19] L. Capogna, D. Danielli, S.D. Pauls and J.T. Tyson, An introduction to the Heisenberg group and the sub-
Riemannian isoperimetric problem, Birkhauser, Basel, 2007.
[20] E.A. Carlen and C. Kerce, On the cases of equality in Bobkov's inequality and Gaussian rearrangement, Calc. Var.
Partial Differential Equations 13 (2001), 1 -- 18.
[21] I. Chavel, Isoperimetric inequalities: differential geometric aspects and analytic perspectives, Cambridge University
Press, Cambridge, 2001.
54
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
[22] J. Cheeger, A lower bound for the smallest eigevalue of the Laplacian, in Problems in analysis, 195 -- 199, Princeton
Univ. Press, Princeton, 1970.
[23] A. Cianchi, On relative isoperimetric inequalities in the plane, Boll. Un. Mat. Ital. 3-B (1989), 289 -- 326.
[24] A. Cianchi, A sharp embedding theorem for Orlicz -- Sobolev spaces, Indiana Univ. Math. J. 45 (1996), 39 -- 65.
[25] A. Cianchi, Second-order derivatives and rearrangements, Duke Math. J. 105 (2000), 355 -- 385.
[26] A. Cianchi, Optimal Orlicz-Sobolev embeddings, Rev. Mat. Iberoam. 20 (2004), 427 -- 474.
[27] A. Cianchi, Symmetrization and second-order Sobolev inequalities, Ann. Mat. Pura Appl. 183 (2004), 45 -- 77.
[28] A. Cianchi, Higher-order Sobolev and Poincar´e inequalities in Orlicz spaces, Forum Math. 18 (2006), 745 -- 767.
[29] A. Cianchi, Orlicz-Sobolev boundary trace embeddings, Math. Zeit. 266 (2010), 431 -- 449.
[30] A. Cianchi, D.E. Edmunds and P. Gurka, On weighted Poincar´e inequalities, Math. Nachr. 180 (1996), 15 -- 41.
[31] A. Cianchi, N. Fusco, F. Maggi and A. Pratelli, The sharp Sobolev inequality in quantitative form, J. Eur. Math.
Soc. 11 (2009), 1105 -- 1139.
[32] A. Cianchi, N. Fusco, F. Maggi and A. Pratelli, On the isoperimetric deficit in the Gauss space, Amer. J. Math.
133 (2011), 131 -- 186.
[33] A. Cianchi and L. Pick, Optimal Gaussian Sobolev embeddings, J. Funct. Anal. 256 (2009), 3588 -- 3642.
[34] D. Cordero-Erausquin, B. Nazaret and C. Villani, A mass-transportation approach to sharp Sobolev and Gagliardo-
Nirenberg inequalities, Adv. Math. 182 (2004), 307 -- 332.
[35] E. De Giorgi, Sulla propriet`a isoperimetrica dell'ipersfera, nella classe degli insiemi aventi frontiera orientata di
misura finita, Atti. Accad. Naz. Lincei Mem. Cl. Sci. Fis. Mat. Nat. 5 (1958), 33 -- 44.
[36] A. Diaz, N. Harman, S. Howe and D. Thompson, Isoperimetric problems in sectors with density, Adv. Geom., to
appear.
[37] D.E. Edmunds and W.D. Evans, Spectral theory and differential operators, Clarendon Press, Oxford 1987.
[38] D.E. Edmunds, V. Kokilashvili and A. Meskhi, Bounded and compact integral operators, Kluwer Academic Pub-
lishers, Dordrecht, 2002.
[39] D. E. Edmunds, R. Kerman and L. Pick, Optimal Sobolev imbeddings involving rearrangement-invariant quasinorms,
J. Funct. Anal. 170 (2000), 307 -- 355.
[40] L. Esposito, V. Ferone, B. Kawohl, C. Nitsch, and C. Trombetti, The longest shortest fence and sharp Poincar´e-
Sobolev inequalities, Arch. Ration. Mech. Anal. 206 (2012), 821 -- 851.
[41] W.D. Evans, B. Opic and L. Pick, Interpolation of integral operators on scales of generalized Lorentz -- Zygmund
spaces, Math. Nachr. 182 (1996), 127 -- 181.
[42] H. Federer and W. Fleming, Normal and integral currents, Annals of Math. 72 (1960), 458 -- 520.
[43] A. Figalli and F. Maggi, On the isoperimetric problem for radial log-convex densities, Calc. Var. Partial Differential
Equations, to appear.
[44] E. Gagliardo, Propriet`a di alcune classi di funzioni di pi`u variabili, Ric. Mat. 7 (1958), 102 -- 137.
[45] A. Grygor'yan, Isoperimetric inequalities and capacities on Riemannian manifolds, The Maz'ya Anniversary Col-
lection, vol. 1, Operator Theory, Advances and Applications 110, Birkhauser, Basel, 1999, 139 -- 153.
[46] L. Gross, Logarithmic Sobolev inequalities, Amer. J. Math. 97 (1975), 1061 -- 1083.
[47] P. Hai lasz and P. Koskela, Isoperimetric inequalites and imbedding theorems in irregular domains, J. London Math.
Soc. 58 (1998), 425 -- 450.
[48] E. Hebey, Analysis on manifolds: Sobolev spaces and inequalities, Courant Lecture Notes in Mathematics 5, AMS,
Providence, 1999.
[49] D. Hoffman and J. Spruck, Sobolev and isoperimetric inequalities for Riemannian submanifolds, Comm. Pure Appl.
Math. 27 (1974), 715 -- 727.
[50] D. Hoffmann and J. Spruck, A correction to: "Sobolev and isoperimetric inequalities for Riemannian submanifolds"
(Comm. Pure Appl. Math. 27 (1974), 715 -- 725), Comm. Pure Appl. Math. 28 (1975), 765 -- 766.
[51] R. Kerman and L. Pick, Optimal Sobolev imbeddings, Forum Math. 18, 4 (2006), 535 -- 570.
[52] T. Kilpelainen and J. Mal´y, Sobolev inequalities on sets with irregular boundaries, Z. Anal. Anwend. 19 (2000),
369 -- 380.
[53] V.S. Klimov, Imbedding theorems and geometric inequalities, Izv. Akad. Nauk SSSR 40 (1976), 645 -- 671 (Russian);
English translation: Math. USSR Izv. 10 (1976), 615 -- 638.
[54] V.I. Kolyada, Estimates on rearrangements and embedding theorems, Mt. Sb. 136 (1988), 3 -- 23 (Russian); English
translation: Math. USSR Sb. 64 (1989), 1 -- 21.
[55] M. Krbec, B. Opic, L. Pick and J. R´akosn´ık, Some recent results on Hardy type operators in weighted function
spaces and related topics, in: Function spaces, Differential Operators and Nonlinear Analysis (Friedrichroda, 1992),
Teubner-Texte Math. 133, Teubner, Stuttgart 1993, 158-184.
[56] D.A. Labutin, Embedding of Sobolev spaces on Holder domain, Proc. Steklov Inst. Math. 227 (1999), 163 -- 172.
[57] M. Ledoux, The concentration of measure phenomenon, Mathematical Surveys and Monographs 89, American
Mathematical Society, Providence, RI, 2001.
HIGHER-ORDER EMBEDDINGS
55
[58] M. Ledoux, Spectral gap, logarithmic Sobolev constant and geometric bounds, In Surveys in Differential Geometry,
Vol. IX, 219 -- 240, In. Press Somerville, MA, 2004.
[59] P.-L. Lions F. Pacella and M. Tricarico, Best constants in Sobolev inequalities for functions vanishing on some
part of the boundary and related questions, Indiana Univ. Math. J. 37 (1988), 301 -- 324.
[60] E. Lutwak D. Yang and G. Zhang, Sharp affine Lp Sobolev inequalities, J. Diff. Geom. 62 (2002), 17 -- 38.
[61] J. Mart´ın and M. Milman, Higher-order symmetrization inequalities and applications, J. Math. Anal. Appl. 330
(2007), 91-113.
[62] J. Mart´ın and M. Milman, Isoperimetry and symmetrization for logarithmic Sobolev inequalities, J. Funct. Anal.
256 (2009), 149 -- 178.
[63] J. Mart´ın and M. Milman, Pointwise symmetrization inequalities for Sobolev functions and applications, Adv.
Math. 225 (2010), 121 -- 199.
[64] F. Mart´ın-Reyes, Weights, one-sided operators, singular integrals and egodic theorems, in: Nonlinear Analysis,
Function Spaces and Applications, Vol. 5 (Prague, 1994), Prometheus, Prague, 1994, 103-137.
[65] V.G. Maz'ya, Classes of regions and imbedding theorems for function spaces, Dokl. Akad. Nauk. SSSR 133 (1960),
527 -- 530 (Russian); English translation: Soviet Math. Dokl. 1 (1960), 882 -- 885.
[66] V.G. Maz'ya, On p-conductivity and theorems on embedding certain functional spaces into a C-space, Dokl. Akad.
Nauk. SSSR 140 (1961), 299 -- 302 (Russian); English translation: Soviet Math. Dokl. 3 (1962).
[67] V.G. Maz'ya, Sobolev spaces with applications to elliptic partial differential equations, Springer, Berlin, 2011.
[68] E. Milman, On the role of convexity in functional and isoperimetric inequalities, Proc. London Math. Soc. 99
(2009), 32 -- 66.
[69] J. Moser, A sharp form of an inequality by Trudinger, Indiana Univ. Math. J. 20 (1971), 1077 -- 1092.
[70] L. Nirenberg, On elliptic partial differential equations, Ann. Sc. Norm. Sup. Pisa 13 (1959), 115 -- 162.
[71] B. Opic and L. Pick, On generalized Lorentz -- Zygmund spaces, Math. Inequal. Appl. 2 (1999), 391 -- 467.
[72] R. O'Neil, Convolution operators in L(p, q) spaces, Duke Math. J. 30 (1963), 129 -- 142.
[73] J. Peetre, Espaces d' interpolation et th´eor`eme de Soboleff, Ann. Inst. Fourier 16 (1966), 279 -- 317.
[74] E. Pelliccia and G. Talenti, A proof of a logarithmic Sobolev inequality, Calc. Var. Partial Differential Equations 1
(1993), 237 -- 242.
[75] S.I. Pohozaev, On the imbedding Sobolev theorem for pl = n, Doklady Conference, Section Math. Moscow Power
Inst. (1965), 158 -- 170 (Russian).
[76] C. Rosales A. Canete V. Bayle and F. Morgan, On the isoperimetric problem in Euclidean space with density,
Calc. Var. Partial Differential Equations 31 (2008), 27 -- 46.
[77] S. Saks, Theory of the integral, Dover Publ., Mineola, New York, 1964 and 2005.
[78] L. Saloff-Coste, Aspects of Sobolev-type inequalities, Cambridge University Press, Cambridge, 2002.
[79] E. Sawyer, Boundedness of classical operators on classical Lorentz spaces, Studia Math. 96 (1990), 145 -- 158.
[80] S.L. Sobolev, On some estimates relating to families of functions having derivatives that are square integrable,
Dokl. Akad. Nauk. SSSR 1 (1936), 267 -- 270 (Russian).
[81] S.L. Sobolev, On a theorem in functional analysis, Sb. Math. 4 (1938), 471 -- 497 (Russian); English translation:
Am. Math. Soc. Trans. 34 (1963), 39 -- 68.
[82] V.D. Stepanov, Weighted norm inequalities for integral operators and related topics, in: Nonlinear Analysis, Func-
tion Spaces and Applications, Vol. 5 (Prague, 1994), Prometheus, Prague, 1994, 139-175.
[83] R.S. Strichartz, A note on Trudinger' s extension of Sobolev' s inequality, Indiana Univ. Math. J. 21 (1972),
841 -- 842.
[84] V.N. Sudakov and B.S. Tsirel'son, Extremal properties of half-spaces for spherically invariant measures, Zap.
Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 41 (1974), 14 -- 24 (Russian); English translation: J.
Soviet Math. 9 (1978), 9 -- 18.
[85] G. Talenti, Best constant in Sobolev inequality, Ann. Mat. Pura Appl. 110 (1976), 353 -- 372.
[86] G. Talenti, Some inequalities of Sobolev type on two-dimensional spheres, General inequalities 5, (Oberwolfach,
1986), 401 -- 408, Internat. Schriftenreihe Numer. Math., 80, Birkha user, Basel, 1987.
[87] N.S. Trudinger, On imbeddings into Orlicz spaces and some applications, J. Mech. Anal. 17 (1967), 473 -- 483.
[88] V.I. Yudovich, Some estimates connected with integral operators and with solutions of elliptic equations, Soviet
Math. Dokl. 2 (1961), 746 -- 749 (Russian).
[89] G. Zhang, The affine Sobolev inequality, J. Diff. Geom. 53 (1999), 183 -- 202.
[90] W.P. Ziemer, Weakly differentiable functions, Graduate texts in Math. 120, Springer, Berlin, 1989.
Dipartimento di Matematica e Informatica "U. Dini", Universit`a di Firenze, Piazza Ghiberti 27, 50122
Firenze, Italy
E-mail address: [email protected]
56
ANDREA CIANCHI, LUBOS PICK AND LENKA SLAV´IKOV ´A
Department of Mathematical Analysis, Faculty of Mathematics and Physics, Charles University,
Sokolovsk´a 83, 186 75 Praha 8, Czech Republic
E-mail address: [email protected]
Department of Mathematical Analysis, Faculty of Mathematics and Physics, Charles University,
Sokolovsk´a 83, 186 75 Praha 8, Czech Republic
E-mail address: [email protected]
|
1807.08591 | 2 | 1807 | 2018-10-03T10:13:23 | On a class of non-Hermitian matrices with positive definite Schur complements | [
"math.FA"
] | Given a positive definite matrix $A\in \mathbb{C}^{n\times n}$ and a Hermitian matrix $D\in \mathbb{C}^{m\times m}$, we characterize under which conditions there exists a strictly contractive matrix $K\in \mathbb{C}^{n\times m}$ such that the non-Hermitian block-matrix \[ \left[ \begin{array}{cc} A & -AK \\ K^*A & D \end{array} \right] \] has a positive definite Schur complement with respect to its submatrix~$A$. Additionally, we show that~$K$ can be chosen such that diagonalizability of the block-matrix is guaranteed and we compute its spectrum. Moreover, we show a connection to the recently developed frame theory for Krein spaces. | math.FA | math |
ON A CLASS OF NON-HERMITIAN MATRICES WITH
POSITIVE DEFINITE SCHUR COMPLEMENTS
THOMAS BERGER, JUAN GIRIBET, FRANCISCO MART´INEZ PER´IA,
AND CARSTEN TRUNK
Abstract. Given Hermitian matrices A ∈ Cn×n and D ∈ Cm×m, and
κ > 0, we characterize under which conditions there exists a matrix
K ∈ Cn×m with kKk < κ such that the non-Hermitian block-matrix
(cid:20) A
K ∗A
−AK
D (cid:21)
has a positive (semi-)definite Schur complement with respect to its sub-
matrix A. Additionally, we show that K can be chosen such that di-
agonalizability of the block-matrix is guaranteed and we compute its
spectrum. Moreover, we show a connection to the recently developed
frame theory for Krein spaces.
1. Introduction
Given a matrix S ∈ C(n+m)×(n+m) assume it is partitioned as
S =(cid:20) A B
C D (cid:21) ,
where A ∈ Cn×n, B ∈ Cn×m, C ∈ Cm×n and D ∈ Cm×m. If A is invertible,
then the Schur complement of A in S is defined by
S/A := D − CA−1B.
This terminology is due to Haynsworth [12, 13], but the use of such a con-
struction goes back to Sylvester [18] and Schur [17]. The Schur complement
arises, for instance, in the following factorization of the block matrix S:
(1.1)
(cid:20) A B
C D (cid:21) =(cid:20)
In
CA−1
0
Im (cid:21)(cid:20) A
0 D − CA−1B (cid:21)(cid:20) In A−1B
Im (cid:21) ,
0
0
which is due to Aitken [1]; note that Ik denotes the identity matrix in Ck×k.
It is a common argument in the proof of the Schur determinant formula [3]:
(1.2)
det(S) = det(A) · det(S/A),
of the Guttman rank additivity formula [11], and of the Haynsworth inertia
additivity formula [14].
The Schur complement has been generalized for example to non-invertible
In this case, if A† is the Moore-Penrose inverse of A, then the Schur
A.
complement S/A is defined by S/A = D − CA†B. It is a key tool not only
2010 Mathematics Subject Classification. Primary 15A83; Secondary 15A23, 15B48.
1
2
T. BERGER, J. GIRIBET, F. MART´INEZ PER´IA, AND C. TRUNK
in matrix analysis but also in applied fields such as numerical analysis and
statistics. For further details see [19].
If A is invertible and S is a Hermitian matrix, then C = B∗ and the Schur
complement of A in S is S/A = D − B∗A−1B. Then (1.1) reads as
0 D − B∗A−1B (cid:21)(cid:20) In A−1B
Im (cid:21),
B∗ D (cid:21) =(cid:20) In A−1B
(cid:20) A B
Im (cid:21)∗(cid:20) A
0
0
0
which implies the following well-known criteria: S is positive definite if and
only if A and S/A are both positive definite. This equivalence is not true
for positive semidefinite matrices, but Albert [2] showed that S is positive
semidefinite if and only if A and S/A are both positive semidefinite and
R(B) ⊆ R(A), where R(X) stands for the range of a matrix X.
In this paper, given κ > 0, a Hermitian matrix A ∈ Cn×n with eigenvalues
λ1 ≥ . . . ≥ λk > 0 ≥ λk+1 ≥ . . . ≥ λn, and a Hermitian matrix D ∈ Cm×m
with eigenvalues µ1 ≤ . . . ≤ µr ≤ 0 < µr+1 ≤ . . . ≤ µm we investigate under
which conditions there exists a matrix K ∈ Cn×m with kKk < κ such that
(1.3)
S =(cid:20) A
K∗A
D (cid:21)
−AK
has a positive (semi-)definite Schur complement S/A with respect to the
submatrix A. Note that
S/A = D + K∗(AA†A)K = D + K∗AK.
Interest in such non-Hermitian block-matrices arises, for instance, in the
recently developed frame theory in Krein spaces, see [7, 9]. There, block-
matrices as in (1.3) with a positive definite A, a Hermitian D and a positive
definite S/A correspond to so-called J-frame operators, see Section 5.
In Theorem 3.3 below we show that this special structured matrix com-
pletion problem has a solution if and only if
r ≤ k and κ2λi + µi > 0 for all i = 1, . . . , r − p,
where p = dim (ker D); this condition may be slightly relaxed if only posi-
tive semidefinite S/A is required. We stress that S is not diagonalizable in
general, not even if S/A is positive definite. Under the above conditions,
we construct a particular matrix K, which depends on some parameters
ε1, . . . , εr. In Theorems 4.2 and 4.4 we compute the eigenvalues of the cor-
responding block matrix S in terms of the eigenvalues of A and D and the
parameters ε1, . . . , εr. A root locus analysis of the latter reveals that if
each εi is small enough, then S is diagonalizable and has only real eigenval-
ues, although S is non-Hermitian.
2. Preliminaries
Given Hermitian matrices A, B ∈ Cn×n, various different relations be-
tween the eigenvalues of A, B and A+ B can be obtained, see e.g. [4, 15, 16].
The following result was first proved by Weyl, see e.g. [4].
ON A CLASS OF NON-HERMITIAN MATRICES
3
Theorem 2.1. Let A, B ∈ Cn×n be Hermitian matrices. Then,
λ↓j (A + B) ≤ λ↓i (A) + λ↓j−i+1(B)
λ↓j (A + B) ≥ λ↓i (A) + λ↓j−i+n(B)
for i ≤ j;
for i ≥ j;
where λ↓j (C) denotes the j-th eigenvalue of C (counted with multiplicities)
if they are arranged in nonincreasing order.
For a rectangular matrix A ∈ Cm×n with rank (A) = r denote by
σ1(A) ≥ σ2(A) ≥ . . . ≥ σr(A) > 0
the singular values of A. Recall that σi(A) = λ↓i (A) for i = 1, . . . , r, where
A = (A∗A)1/2. In particular kAk = σ1(A) denotes the spectral norm of A.
Given A, B ∈ Cm×n, the following inequalities hold. If i ∈ {1, . . . , rank (A)}
and j ∈ {1, . . . , rank (B)} are such that i + j − 1 ≤ rank (AB∗), then
(2.1)
σi+j−1(AB∗) ≤ σi(A)σj(B),
see e.g. [16, Theorem 3.3.16]. As a consequence of these inequalities we have
the following well-known result; for completeness we include a short proof.
Proposition 2.2. Let A ∈ Cn×n be Hermitian with exactly k positive eigen-
values (counted with multiplicities) and let K ∈ Cn×m. Then,
λ↓j (K∗AK) ≤ kKk2λ↓j (A)
for j = 1, . . . , min{k, m, rank (K∗AK)}.
Proof. If K = 0, then the statement trivially holds, so assume that K 6= 0
and hence rank (K) ≥ 1. Then, for all j = 1, . . . , min{k, m, rank (K∗AK)}
λ↓j (K∗AK)≤ σj(K∗AK)≤ σj(K∗A)σ1(K∗)≤ σ1(K∗)2σj(A) =kKk2λ↓j (A),
because λ↓j (A) is positive for j = 1, . . . , k.
(cid:3)
3. Positive (semi-)definiteness of the Schur complement
Throughout this work we consider non-Hermitian block matrices S as
in (1.3), where A ∈ Cn×n and D ∈ Cm×m are Hermitian matrices and
K ∈ Cn×m.
In this section we characterize the existence of a matrix K
such that S in (1.3) has a positive definite (positive semidefinite) Schur
complement.
Assumption 3.1. Let λ1 ≥ λ2 ≥ . . . ≥ λk > 0 ≥ λk+1 ≥ . . . ≥ λn denote
the eigenvalues of A (counted with multiplicities) arranged in nonincreasing
order. Further, let µ1 ≤ µ2 ≤ . . . ≤ µr ≤ 0 < µr+1 ≤ . . . ≤ µm denote the
eigenvalues of D (counted with multiplicities) arranged in nondecreasing
order, and assume that dim (ker D) = p.
4
T. BERGER, J. GIRIBET, F. MART´INEZ PER´IA, AND C. TRUNK
Lemma 3.2. Let Assumption 3.1 hold. If r > k then there is no K ∈ Cn×m
such that D + K∗AK is positive definite. Moreover, if r − p > k then there
is no K ∈ Cn×m such that D + K∗AK is positive semidefinite.
Proof. Assume that r > k. Given K ∈ Cn×m let S1 = ker (K∗(A + A)K)
and consider the subspace S2 of Cm spanned by all eigenvectors of D corre-
sponding to non-positive eigenvalues. Observe that
dimS1 = m − rank (K∗(A + A)K) ≥ m − rank (A + A) = m − k.
By Assumption 3.1 we have that dimS2 = r and hence
dimS1 + dimS2 ≥ (m − k) + r = m + (r − k) > m.
Thus, S1 ∩ S2 6= {0} and for any non-trivial vector v ∈ S1 ∩ S2 we have
h(D + K∗AK)v, vi = hDv, vi − hK∗AKv, vi ≤ 0,
because K∗AKv = −K∗AKv. Therefore, D + K∗AK cannot be positive
definite.
Moreover, assume that r − p > k and consider the subspace S3 of Cm
spanned by all eigenvectors of D corresponding to negative eigenvalues.
Then, dimS3 = r− p and a similar argument shows that D + K∗AK cannot
be positive semidefinite.
(cid:3)
The next result characterizes under which conditions there exists a matrix
K ∈ Cn×m such that D + K∗AK is positive (semi-)definite.
Theorem 3.3. Let Assumption 3.1 hold. Given κ > 0, the following state-
ments hold.
(i) There exists K ∈ Cn×m with kKk < κ such that D + K∗AK is
positive definite if and only if
(3.1)
r ≤ k
and κ2λi + µi > 0
for all i = 1, . . . , r − p.
(ii) There exists K ∈ Cn×m with kKk ≤ κ such that D + K∗AK is
positive semidefinite if and only if
and κ2λi + µi ≥ 0
r − p ≤ k
(3.2)
Proof. We show (i). Assume that there exists a matrix K ∈ Cn×m with
kKk < κ such that D + K∗AK > 0. By Lemma 3.2, it is necessary that
r ≤ k. On the other hand, by Theorem 2.1,
for all i = 1, . . . , r − p.
0 < λ↓m(D + K∗AK) ≤ λ↓i (D) + λ↓m−i+1(K∗AK),
for i = 1, . . . , m. In particular, for i = m − r + p + 1, . . . , m we can combine
the above inequalities with Proposition 2.2 and obtain
0 < λ↓i (D) + kKk2λ↓m−i+1(A) < µm−i+1 + κ2λm−i+1.
Equivalently, we have that µj + κ2λj > 0 for j = 1, . . . , r − p.
ON A CLASS OF NON-HERMITIAN MATRICES
5
Conversely, assume that r ≤ k and κ2λi + µi > 0 for i = 1, . . . , r − p. For
each i = 1, . . . , r − p let 0 < εi < κ2 be such that εiλi + µi > 0, and for
j = r − p + 1, . . . , r let 0 < εj < κ2 be arbitrary. Then, define E ∈ Cn×m by
E =(cid:20) diag (√ε1, . . . ,√εr)
0n−r,r
0n−r,m−r (cid:21) ,
0r,m−r
where 0p,q is the null matrix in Cp×q. Further, let U ∈ Cn×n and V ∈ Cm×m
be unitary matrices such that A = U DλU∗ and D = V DµV ∗, where
Dλ = diag (λ1, . . . , λn) and Dµ = diag (µ1, . . . , µm).
Then, for
K := U EV ∗,
(3.3)
it is straightforward to observe that kKk < κ and
D + K∗AK = V (Dµ + E∗U∗AU E)V ∗ = V (Dµ + E∗DλE)V ∗
= V (cid:20) diag (ε1λ1 + µ1, . . . , εrλr + µr)
diag (µr+1, . . . , µm) (cid:21) V ∗
is a positive definite matrix because εi was chosen in such a way that εiλi +
µi > 0 for i = 1, . . . , r − p, and εj λj + µj = εj λj > 0 for j = r − p + 1, . . . , r.
If there is a matrix K ∈ Cn×m with
kKk ≤ κ such that D + K∗AK is positive semidefinite, then r − p ≤ k (see
Lemma 3.2) and following the same arguments as before it is easy to see
that κ2λi + µi ≥ 0 for i = 1, . . . , r − p. The converse can also be proved in a
similar way, but in this case εi may be equal to κ2 for some i = 1, . . . , r − p
(and εj can also be zero for j = r − p + 1, . . . , r). Therefore, kKk ≤ κ and
D + K∗AK is positive semidefinite.
The proof of (ii) is analogous.
0r,m−r
0m−r,r
(cid:3)
4. Spectrum of the block matrix
In the following, we consider the matrix K constructed in the proof of
Theorem 3.3 and investigate the location of the eigenvalues of S in (1.3).
The locations depend on the parameters ε1, . . . , εr and hence their study
resembles a root locus analysis. We start with a preliminary lemma.
Lemma 4.1. Let Assumption 3.1 and (3.2) hold and set
(4.1)
Then we have that
αi := (λi−µi)2
4λ2
i
,
i = 1, . . . , r − p.
0 < −µi
2 (cid:17)2
λi ≤ αi ≤(cid:16) κ2+1
,
for all i = 1, . . . , r − p.
Proof. Given i = 1, . . . , r − p it is straightforward that (λi − µi)2 ≥ −4µiλi.
If (3.2) holds, then λi > 0 for all i = 1, . . . , r − p and hence αi ≥ − µi
> 0.
Furthermore,
λi
λi − µi = (κ2 + 1)λi − (κ2λi + µi) ≤ (κ2 + 1)λi,
6
T. BERGER, J. GIRIBET, F. MART´INEZ PER´IA, AND C. TRUNK
2 (cid:1)2.
which implies that αi ≤(cid:0) κ2+1
In case that Assumption 3.1 and (3.2) hold, we describe the spectrum of
the block matrix S given in (1.3) for the matrix K defined in (3.3).
(cid:3)
Theorem 4.2. Let Assumption 3.1 hold. Given κ > 0, assume that (3.2)
also holds. For i = 1, . . . , r − p choose 0 < εi ≤ κ2 such that εiλi + µi ≥ 0,
and for j = r − p + 1, . . . , r set εj = 0.
If K ∈ Cn×m is as defined in (3.3), then kKk ≤ κ and the spectrum of the
block matrix S ∈ C(n+m)×(n+m) given in (1.3) consists of the real numbers
λr−p+1, . . . , λn, µr−p+1, . . . , µm, and
(4.2)
where αi is given by (4.1). Moreover, for i ∈ {1, . . . , r − p}, we have
2 ± λi√αi − εi,
i = 1, . . . , r − p,
η±i = λi+µi
, then λi > η+
i > 0 > η−i > µi;
λi ≤ εi < αi, then max{λi +µi, 0} ≥ η+
a) if 0 < εi < −µi
λi
b) if −µi
c) if αi < εi ≤ κ2, then η+
d) if εi = αi, then η+
i = η−i ∈ C \ R;
i = η−i = 1
i > η−i ≥ min(cid:8)λi +µi, 0(cid:9);
2 (λi + µi) and there exists a Jordan
chain of length 2 corresponding to this eigenvalue.
Additionally, if εi 6= αi for all i = 1, . . . , r − p, then S is diagonalizable.
Proof. First note that by Lemma 4.1 the range for εi in case a) is non-empty
independently of κ, but the same may not be true for cases b) and c). We
will discuss this later in Remark 4.3.
Using the notation from the proof of Theorem 3.3 we obtain
K∗A
V E∗DλU∗
0 V (cid:21)(cid:20) Dλ −B
where B ∈ Cn×m is given by
D (cid:21) =(cid:20) U DλU∗
−AK
0 V (cid:21)∗
B∗ Dµ (cid:21)(cid:20) U 0
S =(cid:20) A
=(cid:20) U 0
B := DλE =(cid:20) diag (λ1√ε1, . . . , λr−p√εr−p)
0n−r+p,r−p
−U DλEV ∗
(cid:21) =
V DµV ∗
= W(cid:20) Dλ −B
B∗ Dµ (cid:21) W ∗,
0n−r+p,m−r+p (cid:21) ,
0r−p,m−r+p
and W := (cid:2) U 0
notes the standard basis of Cn+m, it is easy to see that
0 V (cid:3) ∈ C(n+m)×(n+m) is unitary. Then, if {e1, . . . , en+m} de-
and S W ej = µj−n W ej
for i = r − p + 1, . . . , n,
for j = n + r − p + 1, . . . , n + m,
S W ei = λi W ei
(4.3)
which yields that λr−p+1, . . . , λn and µr−p+1, . . . , µm are eigenvalues of S.
Now, define the following (r − p) × (r − p) diagonal matrices:
Fλ := diag (λ1, . . . , λr−p),
G := diag (λ1√ε1, . . . , λr−p√εr−p),
Fµ := diag (µ1, . . . , µr−p),
ON A CLASS OF NON-HERMITIAN MATRICES
7
and observe that the remaining 2(r − p) eigenvalues of S coincide with the
spectrum of the submatrix S of W ∗SW given by
S :=(cid:20) Fλ −G
G Fµ (cid:21) .
In order to calculate the eigenvalues of S, consider the matrix Pσ ∈ C2(r−p)×2(r−p)
associated to the following permutation of the integers {1, 2, . . . , 2(r − p)}:
σ(j) =(2j − 1,
2(j − r + p),
for j = 1, . . . , r − p,
for j = r − p + 1, . . . , 2(r − p).
Then, we have that P 2
with r − p blocks of size 2 × 2 along the main diagonal:
σ = I2(r−p) and Pσ SPσ is a block diagonal matrix,
λj
λj√εj
(cid:20)
−λj√εj
µj
(cid:21) ,
j = 1, . . . , r − p.
Thus, the characteristic polynomial of S is given by
r−p
q(η) =
Yi=1(cid:0)(µi − η)(λi − η) + εiλ2
i(cid:1) ,
and η ∈ C is a root of q(η) if and only if
η2 − (λi + µi)η + λi(µi + εiλi) = 0
for some i ∈ {1, . . . , r − p}. This leads to the following eigenvalues of S:
(4.4)
i = λi+µi
η±i = λi+µi
2 ± 1
2q(λi − µi)2 − 4εiλ2
for i = 1, . . . , r − p. Hence, (4.2) follows and statement c) holds.
For statement a) we observe that if 0 < εi < −µi
λi
2 ± λi√αi − εi
, then √αi − εi > λi+µi
2λi
.
Therefore,
i > λi+µi
η+
2 + λi λi+µi
2λi ≥ 0 and η−i < λi+µi
2 − λi λi+µi
2λi ≤ 0.
Furthermore,
On the other hand, if −µi
i < λi+µi
η+
η−i > λi+µi
= λi,
λi−µi
2λi
λi−µi
2λi
2 + λi
2 − λi
2 + λi√αi = λi+µi
2 − λi√αi = λi+µi
λi ≤ εi < αi, then √αi − εi ≤ λi+µi
2 − λi λi+µi
= min{λi + µi, 0} ,
2λi
2 + λi λi+µi
= max {λi + µi, 0} ,
2λi
= µi.
2λi
η−i ≥ λi+µi
i ≤ λi+µi
η+
and, clearly, η+
i > λi+µi
2 > η−i , which proves b).
and
8
T. BERGER, J. GIRIBET, F. MART´INEZ PER´IA, AND C. TRUNK
To show d), assume that εi = αi for some i ∈ {1, . . . , r − p}. Since
η+
i = η−i = 1
, it is straightforward to compute
2 (λi + µi) and √εi = λi−µi
(cid:16) S − 1
2 (λi + µi)I2(r−p)(cid:17) (cid:16)1 + 2
2λi
fi
λi−µi(cid:17) fi
2 (λi + µi)I2r(cid:17)(cid:18)fi
! =(cid:18)fi
fi(cid:19) ,
fi(cid:19) = 0,
(cid:16) S − 1
using the standard basis {f1, . . . , fr−p} of Cr−p. The vectors above form a
Jordan chain of length 2 of S corresponding to the eigenvalue 1
2 (λi + µi).
Hence, a Jordan chain of S corresponding to the eigenvalue 1
2 (λi + µi) can
also be constructed.
Finally, assume that εi 6= αi for all i = 1, . . . , r − p.
In this case, the
space Cn+m has a basis consisting of eigenvectors of S. Indeed, this follows
from (4.3) together with
i I2(r−p)(cid:17)
(cid:16) S − η+
for i = 1, . . . , r − p.
fi
− λi√εi
µi−η+
i
fi! = 0, (cid:16) S − η−i I2(r−p)(cid:17)
fi
− λi√εi
µi−η−
i
fi! = 0
(cid:3)
We emphasize that if for all i = 1, . . . , r − p the parameter εi in Theo-
rem 4.2 is chosen such that a) or b) holds, then the block matrix S in (1.3)
is diagonalizable and has only real eigenvalues, cf. Lemma 4.1.
2 (cid:1)2 ≥ κ2 and equality holds
Remark 4.3. Given κ > 0, note that (cid:0) κ2+1
2 (cid:1)2 for some
if and only if κ = 1. Hence, if κ 6= 1 and κ2 < αi ≤ (cid:0) κ2+1
i ∈ {1, . . . , r − p}, then the corresponding eigenvalues η+
because the range of values for εi in case c) is empty.
For κ = 1, if there exists i ∈ {1, . . . , r − p} such that λi + µi > 0, then
i and η−i are real,
λi − µi = −(λi + µi) + 2λi < 2λi,
hence αi < 1 and we can choose the corresponding parameter εi such that S
has non-real eigenvalues.
Furthermore, if A is positive semidefinite, κ ≤ 1 and εi ≥ −µi
for each
i = 1, . . . , r−p, then λi +µi ≥ 0 and hence the eigenvalues of S are contained
in the (closed) complex right half-plane.
λi
In the remainder of this section, we calculate the eigenvalues of the block
matrix S under the assumption that its Schur complement is positive defi-
nite. Note that if Assumption 3.1 and (3.1) hold we may define αi as in (4.1)
for all i = 1, . . . , r. In this case, 0 < −µi
and αj = 1
2 (cid:1)2 for i = 1, . . . , r − p,
λi ≤ αi <(cid:0) κ2+1
4 for j = r − p + 1, . . . , r.
ON A CLASS OF NON-HERMITIAN MATRICES
9
Theorem 4.4. Let Assumption 3.1 hold. Given κ > 0, assume that (3.1)
also holds. For i = 1, . . . , r − p choose 0 < εi < κ2 such that εiλi + µi > 0,
and for j = r − p + 1, . . . , r let 0 ≤ εj < κ2 be arbitrary.
If K ∈ Cn×m is as defined in (3.3), then kKk < κ and the spectrum of the
block matrix S ∈ C(n+m)×(n+m) given in (1.3) consists of the real numbers
λr+1, . . . , λn, µr+1, . . . , µm, and
(4.5)
η±i = λi+µi
2 ± λi√αi − εi,
i = 1, . . . , r,
where αi is given by (4.1). Moreover, for i = 1, . . . , r, we have
2 (λi + µi) and there exists a Jordan
i > η−i ≥ min(cid:8)λi +µi, 0(cid:9);
, then λi > η+
i > 0 > η−i > µi;
λi ≤ εi < αi, then max{λi +µi, 0} ≥ η+
a) if 0 < εi < −µi
λi
b) if −µi
c) if αi < εi < κ2, then η+
d) if εi = αi, then η+
i = η−i = 1
e) if i > r − p and εi = 0, then η+
i = η−i ∈ C \ R;
chain of length 2 corresponding to this eigenvalue;
i = λi > 0 and η−i = µi = 0.
Additionally, if εi 6= αi for all i = 1, . . . , r, then S is diagonalizable.
Proof. The proof is analogous to the proof of Theorem 4.2, the main differ-
B∗ Dµi W ∗, where
ence is that in this case S = W h Dλ −B
0n−r,r
B := DλE =(cid:20) diag (λ1√ε1, . . . , λr√εr)
0n−r,m−r (cid:21) ∈ Cn×m,
0r,m−r
which yields that λr+1, . . . , λn and µr+1, . . . , µm are eigenvalues of S. The
remaining 2r eigenvalues of S can be calculated in the same way as before.
Also, the only difference in the characterization of the eigenvalues η±i appears
in the case in which i = r − p + 1, . . . , r and εi = 0. But the proof of this
last case is straightforward.
(cid:3)
Example 4.5. We illustrate Theorem 4.4 with a simple example. Let n =
m = 1, D = [0] and A = [a] with a > 0. Then r = 1 and choosing K as
in (3.3) with 0 < ε < 1 = κ2 gives K = [√ε]. In this case α = 1
4 .
By Theorem 4.4, for ε = 1
ing to the only eigenvalue a
4 there is a Jordan chain of length 2 correspond-
2 , and indeed we find that (cid:18) 1
a
−1
a (cid:19) ,(cid:18)1
1(cid:19) form a
Jordan chain of S, hence S is not diagonalizable.
a
On the other hand, for ε 6= 1
2 + aq 1
2 − aq 1
4 − ε and η− = a
are non-real if 1
4 the block matrix S has eigenvalues η+ =
4 − ε. They are positive if ε < 1
4 , and they
4 < ε < 1. In these last two cases S is diagonalizable.
5. Application to J-frame operators
In this section, we exploit Theorems 3.3 and 4.4 to investigate whether a
block matrix S as in (1.3) represents a so-called J-frame operator and when
it is similar to a Hermitian matrix. In the following we briefly recall the
10
T. BERGER, J. GIRIBET, F. MART´INEZ PER´IA, AND C. TRUNK
concept of J-frame operators, which arose in [7, 9] in the context of frame
theory in Krein spaces.
In a finite-dimensional setting, every indefinite inner product space is a
(finite-dimensional) Krein space, see [10]. A map [., .] : Ck × Ck → C is
called an indefinite inner product in Ck, if it is a non-degenerate Hermitian
sesquilinear form. The indefinite inner product allows a classification of
vectors: x ∈ Ck is called positive if [x, x] > 0, negative if [x, x] < 0 and
neutral if [x, x] = 0. Also, a subspace L of Ck is positive if every x ∈ L\{0}
is a positive vector. Negative and neutral subspaces are defined analogously.
A positive (negative) subspace of maximal dimension will be called maximal
positive (maximal negative, respectively).
It is well-known that there exists a Gramian (or Gram matrix) G ∈ Ck×k,
which is Hermitian, invertible and represents [., .] in terms of the usual inner
product in Ck, i.e., [x, y] = hGx, yi for all x, y ∈ Ck. The positive (resp.
negative) index of inertia of [., .] is the number of positive (resp. negative)
eigenvalues of the Gramian G, and it equals the dimension of any maximal
positive (resp. negative) subspace of Ck.
It is clear that the sum of the
inertia indices equals the dimension of the space.
A finite family of vectors F = {fi}q
span({fi}q
i=1 in Ck is a frame for Ck, if
i=1) = Ck
(see, e.g., [5]) or, equivalently, if there exist 0 < α ≤ β such that
for every f ∈ Ck.
2 ≤ βkfk2
αkfk2 ≤
q
The optimal set of constants 0 < α ≤ β (the biggest α and the smallest β)
are called the frame bounds of F. If
(5.1)
q
F : Ck → Ck, f 7→
hf, fii fi
Xi=1
Xi=1(cid:12)(cid:12)hf, fii(cid:12)(cid:12)
is the associated frame operator, then the frame bounds of F are
α = kF −1k−1 = λ↓k(F ) and β = kFk = λ↓1(F ),
see e.g. [5] and the references therein.
Roughly speaking, a J-frame is a frame which is compatible with the
indefinite inner product [., .].
Definition 5.1. Let (Ck, [., .]) be an indefinite inner product space. Then,
a frame F = {fi}q
i=1 in Ck is called a J-frame for Ck, if
M+ := span{f ∈ F [f, f ] ≥ 0} and M− := span{f ∈ F [f, f ] < 0}
are a maximal positive and a maximal negative subspace of Ck, respectively.
If [., .] is an indefinite inner product with positive and negative index
of inertia n and m, respectively, then the maximality of M+ and M− is
ON A CLASS OF NON-HERMITIAN MATRICES
11
Given a J-frame F = {fi}q
equivalent to dimM+ = n and dimM− = m. Note that if F is a J-frame
for Ck, then there are no (non-trivial) f ∈ F with [f, f ] = 0.
i=1 for Ck, its associated J-frame operator
S : Ck → Ck is defined by
Sf =
σi [f, fi] fi,
q
f ∈ Ck,
where σi = sgn [fi, fi] is the signature of the vector fi. S is an invertible
symmetric operator with respect to [., .], i.e.,
Xi=1
[Sf, g] = [f, Sg]
for all
f, g ∈ Ck.
Its relevance follows from the indefinite sampling-reconstruction formula:
Given an arbitrary f ∈ Ck,
Xi=1
σi(cid:2)f, S−1fi(cid:3) fi =
i.e., it plays a role analogous to the fame operator F in equation (5.1).
σi [f, fi] S−1fi,
Xi=1
f =
q
q
In the following, we aim to apply the results from Sections 3 and 4, hence
we restrict ourselves to the following inner product on Ck = Cn+m,
n
m
(5.2)
[(x1, . . . , xn+m), (y1, . . . , yn+m)] =
xn+j yn+j.
xiyi −
Xi=1
Xj=1
In [7, Theorem 3.1] a criterion was provided to determine if an (invertible)
symmetric operator is a J-frame operator. In our setting it says that an
invertible operator S in (Ck, [., .]), which is symmetric with respect to [., .],
is a J-frame operator if and only if there exists a basis of Ck such that S
can be represented as a block-matrix
(5.3)
S =(cid:20) A
K∗A
D (cid:21) ,
−AK
where A ∈ Cn×n is positive definite, K ∈ Cn×m is strictly contractive, and
D ∈ Cm×m is a Hermitian matrix such that D + K∗AK is also positive
definite. Any block-matrix S ∈ C(n+m)×(n+m) of the form (5.3), which
satisfies these conditions will be called J-frame matrix.
Throughout this section we consider the following hypothesis.
Assumption 5.2. Assume that A ∈ Cn×n is positive definite and D ∈
Cm×m is a Hermitian matrix. Let µ1 ≤ µ2 ≤ . . . ≤ µr ≤ 0 < µr+1 ≤ . . . ≤
µm denote the eigenvalues of D (counted with multiplicities) arranged in
nondecreasing order, and let λ1 ≥ λ2 ≥ . . . ≥ λn > 0 denote the eigenvalues
of A (counted with multiplicities) arranged in nonincreasing order.
Theorem 3.3 (for κ = 1) provides a criterion to determine whether there
exists a strictly contractive matrix K ∈ Cn×m (i.e., kKk < 1) such that S
as in (5.3) is a J-frame matrix.
12
T. BERGER, J. GIRIBET, F. MART´INEZ PER´IA, AND C. TRUNK
Theorem 5.3. Let Assumption 5.2 hold. Then there exists K ∈ Cn×m with
kKk < 1 such that S as in (5.3) is a J-frame matrix if and only if
(5.4)
for i = 1, . . . , r.
r ≤ n and λi + µi > 0
We mention that the study of the spectral properties of a J-frame operator
is quite recent, see [7, 8]. In the case of J-frame matrices, for given A and
D, we always find conditions such that a strictly contractive K exists which
turns S into a matrix similar to a Hermitian one. The following result is a
direct consequence of Theorem 4.4 and Lemma 4.1.
Theorem 5.4. Let Assumption 5.2 and (5.4) hold. Then, there exists a
strictly contractive matrix K such that the matrix S given in (5.3) is a
J-frame matrix which is similar to a Hermitian matrix. In this case, all
eigenvalues of S are positive and there exists a basis of Cn+m consisting of
eigenvectors of S.
In the next paragraphs we recall how to construct J-frames for Cn+m
with a prescribed J-frame matrix S. For K ∈ Cn×m with kKk < 1 define
(5.5)
K∗f(cid:19) (cid:12)(cid:12)(cid:12)(cid:12)
M− := {0} × Cm, M+ :=(cid:26) (cid:18) f
f ∈ Cn (cid:27) .
If Cn+m = Cn × Cm is endowed with the indefinite inner product given
in (5.2), then it is immediate that M− is a maximal negative subspace in
Cn+m and M+ is maximal positive in Cn+m. The contraction K ∈ Cn×m
represents the angle between the two subspaces M+ and M−.
Moreover, if K with kKk < 1 is such that the block matrix S given in (5.3)
is a J-frame matrix, consider S = S+ + S− with
(5.6)
S+ =(cid:20) A
K∗A −K∗AK (cid:21)
−AK
and S− =(cid:20) 0
0 D + K∗AK (cid:21) .
0
Then, the restriction of S+ to (M+, [., .]) is a positive definite matrix. In-
deed, if f ∈ Cn \ {0}, then
(cid:20)S+(cid:18) f
K∗f(cid:19)(cid:21) =(cid:20)(cid:18) A(I − KK∗)f
K∗f(cid:19) ,(cid:18) f
K∗A(I − KK∗)f(cid:19) ,(cid:18) f
K∗f(cid:19)(cid:21)
= hA(I − KK∗)f, fi − hKK∗A(I − KK∗)f, fi
= h(I − KK∗)A(I − KK∗)f, fi > 0.
(5.7)
On the other hand, it is evident that the restriction of S− to (M−,−[., .])
is just D + K∗AK, which is also a positive definite matrix.
Therefore, it is possible to construct frames F± for the (finite-dimensional)
Hilbert spaces (M±,±[., .]) with these matrices as frame operators, see [6].
Moreover, the family F+ ∪ F− is a J-frame for Cn+m with S as its J-frame
operator, see [9, Theorem 5.6].
ON A CLASS OF NON-HERMITIAN MATRICES
13
α− = λ↓m(D + K∗AK)
β− = λ↓1(D + K∗AK),
Proposition 5.5. Let Assumption 5.2 hold and let K ∈ Cn×m with kKk <
1 be such that S as in (5.3) is a J-frame matrix. Further, let M± be as
in (5.5) and let F± be frames for (M±,±[., .]) with frame operator S± given
in (5.6). Then, the frame bounds of F− are
(5.8)
and
and the frame bounds of F+ are the boundary values of the numerical range
of the positive definite matrix C := (I − KK∗)1/2A(I − KK∗)1/2 ∈ Cn×n,
(5.9)
β+ = λ↓1(C).
Proof. Recall g ∈ M+ if and only if g =(cid:16) f
K ∗f(cid:17) for some f ∈ Cn. Then,
K∗f(cid:19)(cid:21) = h(I − KK∗)f, fi = k(I − KK∗)1/2fk2.
kgk2 =(cid:20)(cid:18) f
On the other hand, if h = (I − KK∗)1/2f ∈ Cn, by (5.7) we have that
K∗f(cid:19) ,(cid:18) f
α+ = λ↓n(C)
and
[S+g, g] = h(I − KK∗)A(I − KK∗)f, fi
=DC(I − KK∗)1/2f, (I − KK∗)1/2fE = hCh, hi .
Since khk = kgk, it is immediate that the numerical ranges of S+ and C
coincide. Therefore, (5.9) holds.
On the other hand, the desired characterization of the frame bounds α−
and β− of F− has been obtained in [7, Proposition 4.1].
(cid:3)
Using Weyl's inequalities and the inequalities for the singular values of a
product of matrices presented in (2.1) we can obtain the following a priori
estimates for the frame bounds of F+ and F−.
Proposition 5.6. Let Assumption 5.2 and (5.4) hold and let K ∈ Cn×m
with kKk < 1 be such that S as in (5.3) is a J-frame matrix. Further, let
M± be as in (5.5) and let F± be frames for (M±,±[., .]) with frame operator
S± given in (5.6). If σ1 ≥ . . . ≥ σl > 0 are the singular values of K, then
the frame bounds of F− satisfy
0 < α− ≤ β− ≤ σ2
1λ1 + µm.
Furthermore, the frame bounds of F+ satisfy
(1 − σ2
1)λn ≤ α+ ≤ β+ ≤ (1 − σ2
l )λ1.
Proof. By Proposition 5.5 we have α− = λ↓m(D +K∗AK) > 0. Furthermore,
by Theorem 2.1 and Proposition 2.2 we have
β− = λ↓1(D + K∗AK) ≤ λ↓1(D) + λ↓1(K∗AK)
≤ λ↓1(D) + kKk2λ↓1(A) = µm + kKk2λ1 = σ2
1λ1 + µm.
14
T. BERGER, J. GIRIBET, F. MART´INEZ PER´IA, AND C. TRUNK
On the other hand, if C = (I − KK∗)1/2A(I − KK∗)1/2, then α+ = λ↓n(C)
and β+ = λ↓1(C). Hence, using (2.1) we obtain
α+ = λ↓n(C) = σn(A1/2(I − KK∗)1/2)2 = σ1(A−1/2)2
σ1(A−1/2)2 σn(A1/2(I − KK∗)1/2)2
≥ σn((I−KK ∗)1/2)2
σ1(A−1/2)2
= λ↓n(I − KK∗)λ↓n(A) = (1 − σ2
1)λn,
and further
β+ = λ↓1(C) = σ1(A1/2(I − KK∗)1/2)2 ≤ σ1(A1/2)2σ1((I − KK∗)1/2)2
= λ↓1(A)λ↓1(I − KK∗) = λ1(1 − σ2
l ),
which completes the proof.
(cid:3)
Finally, let A ∈ Cn×n and D ∈ Cm×m satisfy Assumption 5.2 and assume
that (5.4) holds. For i = 1, . . . , r choose 0 < εi < 1 such that εiλi + µi > 0.
If A = U DλU∗, D = V DµV ∗ and K ∈ Cn×m is given by (3.3) then kKk < 1,
C = (I − KK∗)1/2A(I − KK∗)1/2 =
= U(cid:20) diag ((1 − ε1)λ1, . . . , (1 − εr)λr)
0n−r,r
0r,m−r
diag (λr+1, . . . , λn) (cid:21) U∗,
and
0m−r,r
D + K∗AK = V (cid:20) diag (ε1λ1 + µ1, . . . , εrλr + µr)
Then, we can explicitly compute the frame bounds for F+ and F−:
• α+ = min{(1 − ε1)λ1, . . . , (1 − εr)λr, λn};
• β+ = max{(1 − ε1)λ1, . . . , (1 − εr)λr, λr+1};
• α− = min{ε1λ1 + µ1, . . . , εrλr + µr, µr+1};
• β− = max{ε1λ1 + µ1, . . . , εrλr + µr, µm}.
0r,m−r
diag (µr+1, . . . , µm) (cid:21) V ∗.
Observe that, since (1 − εi)λi < λi + µi and εiλi + µi < λi + µi for each
i = 1, . . . , r, we can obtain the following a priori estimates for the lower
frame bounds of F+ and F−:
and
α+ ≤ min{λ1 + µ1, . . . , λr + µr, λn},
α− ≤ min{λ1 + µ1, . . . , λr + µr, µr+1},
which are independent of the strictly contractive matrix K given in (3.3),
i.e. independent of the angle between the subspaces M+ and M−.
References
[1] A.C. Aitken, Studies in practical mathematics, I: The evaluation, with applications,
of a certain triple product matrix, Proceedings of the Royal Society of Edinburgh 57
(1937), 269 -- 304.
[2] A. Albert, Conditions for positive and nonnegative definiteness in terms of pseu-
doinverses, SIAM J. Appl. Math. 17 (1969) 434 -- 440.
ON A CLASS OF NON-HERMITIAN MATRICES
15
[3] T. Banachiewicz, Zur Berechnung der Determinanten, wie auch der Inversen, und
zur darauf basierten Auflosung der Systeme linearer Gleichungen, Acta Astronomica
Serie C, 3 (1937), 41 -- 67.
[4] R. Bhatia, Matrix Analysis, Springer-Verlag, New York, 1997.
[5] P. Casazza and G. Kutyniok, Finite Frames: Theory and Applications, Applied and
Numerical Harmonic Analysis, Birkhauser, Berlin, 2013.
[6] P. Casazza, and M. Leon, Existence and construction of finite frames with a given
frame operator, Int. J. Pure Appl. Math. 63 (2010), 149 -- 158.
[7] J.I. Giribet, M. Langer, L. Leben, A. Maestripieri, F. Mart´ınez Per´ıa, and C. Trunk,
Spectrum of J-frame operators, Opuscula Math. 38 (2018), 623 -- 649.
[8] J.I. Giribet, M. Langer, F. Mart´ınez Per´ıa, F. Philipp and C. Trunk, Spectral en-
closures for a class of block operator matrices, submitted.
[9] J.I. Giribet, A. Maestripieri, F. Mart´ınez Per´ıa, and P.G. Massey, On frames for
Krein spaces, J. Math. Anal. Appl. 393 (2012), 122 -- 137.
[10] I. Gohberg, P. Lancaster, and L. Rodman, Indefinite Linear Algebra and Applica-
tions, Birkhauser Verlag, Basel, 2005.
[11] L. Guttman, Enlargement methods for computing the inverse matrix, Annals of
Mathematical Statistics 17 (1946), 336 -- 343.
[12] E.V. Haynsworth, On the Schur complement, Basel Mathematical Notes #BMN 20,
1968.
[13] E.V. Haynsworth, Determination of the inertia of a partitioned Hermitian matrix,
Linear Algebra Appl. 1 (1968), 73 -- 81.
[14] E.V. Haynsworth and A.M. Ostrowski, On the inertia of some classes of partitioned
matrices, Linear Algebra Appl. 1 (1968), 299 -- 316.
[15] R. Horn and C. R. Johnson, Matrix Analysis, Second edition. Cambridge University
Press, Cambridge, 2013.
[16] R. Horn and C. R. Johnson, Topics in Matrix Analysis, Cambridge University Press,
Cambridge, 1991.
[17] I. Schur, Uber Potenzreihen, die im Innern des Einheitskreises beschrankt sind [I],
Journal fur die reine und angewandte Mathematik 147 (1917), 205 -- 232.
[18] J.J. Sylvester, On the relation between the minor determinants of linearly equivalent
quadratic functions, London, Edinburgh, and Dublin Philosophical Magazine and
Journal of Science, Fourth Series, 1 (1851), 295 -- 305.
[19] F. Zhang, The Schur Complement and Its Applications, Numerical Methods and
Algorithms 4, Springer, New York (2005).
Fachbereich Mathematik, Universitat Hamburg, Bundesstrasse 55, D-20146
Hamburg, Germany
E-mail address: [email protected]
Departamento de Ingenier´ıa Electr´onica y Matem´atica -- Universidad de
Buenos Aires and Instituto Argentino de Matem´atica "Alberto P. Calder´on"
(CONICET), Saavedra 15 (1083) Buenos Aires, Argentina
E-mail address: [email protected]
Centro de Matem´atica de La Plata (CeMaLP) -- FCE-UNLP, La Plata,
Argentina, and Instituto Argentino de Matem´atica "Alberto P. Calder´on"
(CONICET), Saavedra 15 (1083) Buenos Aires, Argentina
E-mail address: [email protected]
Institut fur Mathematik, Technische Universitat Ilmenau, Postfach 100565,
D-98684 Ilmenau, Germany, and Instituto Argentino de Matem´atica "Alberto
P. Calder´on" (CONICET), Saavedra 15 (1083) Buenos Aires, Argentina
E-mail address: [email protected]
|
1901.01537 | 2 | 1901 | 2019-01-09T02:01:18 | A note on $L^0$-convexly compact sets in random locally convex modules | [
"math.FA"
] | In this note, we study $L^0$-convexly compact sets in random locally convex modules. We show that an $L^0$-convexly compact set must be closed and almost surely bounded, and prove that an $L^0$-convexly compact set is also convexly compact. | math.FA | math |
A note on L0-convexly compact sets in random locally convex
modules
Mingzhi Wu1
Shien Zhao2
1. School of Mathematics and Physics, China University of Geosciences,
Wuhan 430074, China
Email: [email protected]
2.Elementary Educational College, Capital Normal University,
Beijing 100048, China
Email: [email protected]
Abstract
In this note, we study L0-convexly compact sets in random locally convex modules. We show
that an L0-convexly compact set must be closed and almost surely bounded, and prove that an
L0-convexly compact set is also convexly compact.
Key words. random locally convex module, L0-convexly compactness, convexly compactness
MSC2010: 46A16, 46A19, 46A50, 46H25
1
Introduction
For the recent several years, Guo, et.al. have been developing random convex analysis to meet the
need of analysis of conditional risk measures and the related variational and optimization problems
[9, 10, 11, 12]. The so-called random convex analysis is convex analysis on a Hausdorff topological
module over the topological algebra L0(F , K), where K is the scalar field of real or complex numbers and
L0(F , K) the algebra of equivalence classes of K-valued measurable functions defined on a probability
space (Ω, F , P ), endowed with the topology of convergence in probability. The most important kinds of
Hausdorff topological module over the topological algebra L0(F , K) are random normed modules and
random locally convex modules, which had been studied mainly by Guo since the early 1990s (see [4]
for a brief history of the development of random normed modules and random locally convex modules).
Since L0-convex (L0 is short for L0(F , R)) subsets rather than usual convex subsets have played crucial
roles in random convex analysis, in [7], Guo, et.al. introduce the notion of L0-convex compactness for
an L0-convex subset of a Hausdorff topological module over the topological algebra L0(F , K). An L0-
convex subset is said to be L0-convexly compact (or, is said to have L0-convex compactness), which is a
1
generalization of that of Zitkovi´c's convex compactness [13], if any family of nonempty closed L0-convex
subsets of it with the finite intersection property has a nonempty intersection.
As shown in [7, 8], L0-convex compactness is a proper substitution of classical weakly compactness.
As applications, Guo, et. al. successfully generalize some basic theorems of classical convex optimization
and variational inequalities from a convex function on a reflexive Banach space to an L0-convex function
on a random reflexive random normed module and establish the Kirk's fixed point theorem in a complete
random normed module.
Although many important properties have been established for L0-convexly compact sets in random
normed modules and random random locally convex modules, some subtle questions related to L0-
convexly compactness remain unclear. In Remark 2.6 in [8], Guo, et.al. pointed out that an L0-convexly
compact set in a random normed module must be convexly compact. However,they do not know whether
a closed L0-convexly compact set in a complete random locally convex module is necessarily convexly
compact. Besides, it is not clear whether a L0-convexly compact set in a random locally convex module
is closed and almost surely bounded.
In this note, we study L0-convexly compact sets in random locally convex modules. We prove that
an L0-convexly compact set must be closed and almost surely bounded, and show that an L0-convexly
compact set is also convexly compact.
2 The Basic Definitions and Properties
Throughout this paper, (Ω, F , P ) always denotes a given probability space, K the scalar field R of
real numbers or C of complex numbers, L0(F , K) the algebra of equivalence classes of K-valued F -
measurable random variables defined on (Ω, F , P ), in particular we simply write L0 for L0(F , R). Be-
sides, ¯L0 stands for the set of equivalence classes of extended real valued random variables on (Ω, F , P ).
Here, equivalence is understood as usual, namely two random variables are equivalent if they equals
P -almost surely. As usual, we denote by IA the equivalence class of IA for any A ∈ F , here, IA denotes
the characteristic function of A, namely IA(ω) = 1 for ω ∈ A and 0 otherwise. The partial order ≤ on
¯L0 is defined by ξ ≤ η iff ξ0(ω) ≤ η0(ω) for P -almost surely all ω ∈ Ω, where ξ0 and η0 are arbitrarily
chosen representatives of ξ and η, respectively.
Proposition 1 below can be regarded as a random version of the classical supremum principle.
Proposition 1 (see[2].) ( ¯L0, ≤) is a complete lattice, for any nonempty subset H of ¯L0, W H and V H
denote the supremum and infimum of H, respectively, and the following statements hold:
(1).There exists two sequences {an, n ∈ N } and {bn, n ∈ N } in H such that Wn≥1 an = W H and
Vn≥1 bn = V H.
(2).If H is directed upwards (downwards), namely there exists h3 ∈ H for any h1 and h2 ∈ H such
2
that h3 ≥ h1 W h2 (resp., h3 ≤ h1 V h2), then {an, n ∈ N } (resp., {bn, n ∈ N }) can be chosen as
nondecreasing (resp., nonincreasing).
(3).As a sublattice of ¯L0, L0 is conditionally complete, namely any nonempty subset with an upper
(resp., a lower) bound has a supremum (resp., an infimum).
Denote L0
+ = {ξ ∈ L0 ξ ≥ 0}. We recall the notion of a random locally convex module.
Definition 1 (see [4].) An ordered pair (E, P) is called a random locally convex module over K with
base (Ω, F , P ) if E is a left module over the algebra L0(F , K) (briefly, an L0(F , K) -- module) and P a
family of mappings from E to L0
+ such that the following three axioms are satisfied:
(i)∨{kxk : k · k ∈ P} = 0 iff x = θ (the null element of E);
(ii) kξxk = ξkxk for any ξ ∈ L0(F , K) and any x ∈ E;
(iii) kx + yk ≤ kxk + kyk for all x and y ∈ E;
Furthermore, a mapping k · k : E → L0
+ satisfying (ii) and (iii) is called an L0-seminorm; in addition,
if kxk = 0 also implies x = θ, then it is called an L0-norm, in which case (E, k · k) is called a random
normed module over K with base (Ω, F , P ), and is a special case of a random locally convex module
when P consists of a single L0-norm k · k.
The simplest example of random normed module is (L0(F , K), · ), where · is the absolute value
mapping.
In this note, any given random locally convex module is always endowed with the (ε, λ)-topology.
The (ε, λ)-topology for L0(F , K) is exactly the topology of convergence in probability. To introduce
the (ε, λ)-topology for a general random locally convex module, let (E, P) be a random locally convex
module with base (Ω, F , P ), for any finite nonempty subfamily Q of P, k · kQ : E → L0
+ defined
by kxkQ = W{kxk : k · k ∈ Q} for any x ∈ E is still an L0 -- seminorm on E. Furthermore, let
Uθ(Q, ε, λ) = {x ∈ E P {ω ∈ Ω kxkQ(ω) < ε} > 1 − λ} for any finite nonempty subfamily Q of P,
ε > 0 and 0 < λ < 1. Then we have the following:
Proposition 2 (see [4].) Let (E, P) be a random locally convex module over K with base (Ω, F , P ).
Then {Uθ(Q, ε, λ) Q is a finite nonempty subfamily of P, ε > 0, 0 < λ < 1} forms a local base at θ of
some Hausdorff linear topology for E, called the (ε, λ)-topology, denoted by Tε,λ. Furthermore, E is a
topological module over the topological algebra L0(F , K) when E and L0(F , K) are endowed with their
respective (ε, λ) -- topology.
Let E be an L0(F , K) -- module. A subset G of E is said to be L0 -- convex [5]: if ξx + (1 − ξ)y ∈ G
for all x and y ∈ G and ξ ∈ L0
+ such that 0 ≤ ξ ≤ 1.
Definition 2 (see [7].) Let (E, T ) be a topological module over the topological algebra (L0(F , K), Tε,λ)
and G an L0 -- convex subset of E. G is L0 -- convexly compact (or, is said to have L0 -- convex compactness)
3
if any family of closed L0 -- convex subsets of G has a nonempty intersection whenever this family has
the finite intersection property.
3 L0-convexly compactness and almost surely boundedness
According to Guo, et. al [7], a nonempty set G in a random locally convex module (E, P) is said to be
almost surely bounded if for each k·k ∈ P, there is some ξ ∈ L0
+ such that kgk ≤ ξ, ∀g ∈ G. Lemma 2.19
in [7] states that an L0-convex subset of a random locally convex module whose closure is L0-convexly
compact must be almost surely bounded. Since it is not clear whether the closure of an L0-convexly
compact set is still L0-convexly compact or not, a natural question arises: is an L0-convexly compact
set in a random locally convex module necessarily almost surely bounded? In this section, we answer
this question affirmatively.
Theorem 1 Let (E, P) be a random locally convex module over K with base (Ω, F , P ) and G is an
L0-convex subset of E. If G is L0-convexly compact, then G must be almost surely bounded.
Considering the definition of almost surely boundedness in random locally convex module, Theorem
1 will be obvious if we combine Proposition 3 and Proposition 4 below.
In the sequel, for a subset G of an L0(F , K) -- module, convL0 (G) always denotes the L0 -- convex hull
of G, namely, the smallest L0 -- convex subset containing G. For a subset G of a given topological space,
G always stands for the closure of G.
Proposition 3 Let G be a nonempty L0-convex subset of L0. If G is L0-convexly compact, then G
must be almost surely bounded and closed.
Proof. Let M = ∨G, m = ∧G, we will show that both M and m belong to G, and only give the
proof of M ∈ G, since the other is similar. Since the L0-convexity of G implies that G is directed
upwards, according to Proposition 1, there exists a nondecreasing sequence {ξn, n ∈ N } in G such
that {ξn, n ∈ N } converges to M almost surely. For each n, let Fn = convL0{ξk, k ≥ n} ∩ G, then Fn
is a closed L0-convex subset of G, and obviously the family {Fn : n ∈ N } has the finite intersection
property. Since {ξn, n ∈ N } is nondecreasing, each member in convL0{ξk, k ≥ n} must be greater
than ξn, implying each member in convL0{ξk, k ≥ n} must be greater than ξn. By the L0-convexly
compactness of G, there exists an ξ ∈ L0 such that ξ is in every Fn, which implies that ξ ∈ G and
ξ ≥ ξn for every n, thus ξ ≥ ∨{ξn : n ∈ N } = M , yielding ξ = M .
Now we have obtained that M ∈ G and m ∈ G, then by the L0-convexity of G, G must be the
random closed interval [m, M ].
(cid:3)
4
Proposition 4 Let (E, P) be a random locally convex module over K with base (Ω, F , P ) and G a
nonempty L0-convex subset of E. Then for each continuous L0-seminorm k · k on E, {kgk : g ∈ G} is
an L0-convex set in L0, furthermore, if G is L0-convexly compact, then {kgk : g ∈ G} is L0-convexly
compact in L0.
Proof. Denote L0[0, 1] = {ξ ∈ L0 0 ≤ ξ ≤ 1}. Given any two elements x, y in G, to show that
{kgk : g ∈ G} is an L0-convex set in L0, we need to find for each λ ∈ L0[0, 1] an element z ∈ G such that
kzk = λkxk + (1 − λ)kyk. Define f : L0[0, 1] → L0 by f (t) = k(1 − t)x + tyk for each t ∈ L0[0, 1]. Then,
we can check that f is continuous and L0-convex, and f (0) = kxk, f (1) = kyk, thus by the L0-valued
function's intermediate value theorem -- Theorem 1.6 of [6], for each λ ∈ L0[0, 1] there exists a t ∈ L0[0, 1]
such that f (t) = k(1 − t)x + tyk = λkxk + (1 − λ)kyk. Since G is L0-convex, z = (1 − t)x + ty ∈ G is
just the element we are looking for.
If G is L0-convexly compact, we need to show that {kgk : g ∈ G} is L0-convexly compact in L0.
Assume {Fa : a ∈ A} is a family of closed L0-convex subset of {kgk : g ∈ G} with the finite intersection
property, we need to show T{Fa : a ∈ A} 6= ∅. For each a ∈ A, let Ga = {g ∈ G : kgk ∈ Fa}.
{Fa : a ∈ A} has the the finite intersection property implies that {Ga : a ∈ A} has the finite intersection
property. Since k · k is a continuous L0-seminorm on E, each Ga is a closed L0-convex subset of G.
Therefore, from the L0-convexly compactness of G we obtain T{Ga : a ∈ A} 6= ∅, which in turn implies
(cid:3)
that T{Fa : a ∈ A} 6= ∅.
4 L0-convexly compactness and closedness
Our main result in this section is Theorem 2 which states that an L0-convex L0-convexly compact set
in a random locally convex module must be closed.
Theorem 2 is parallel to Lemma 1 of [1] which states that a convexly compact set in a locally convex
space must be closed. Since there does not exist any local base which consists of L0-convex and closed
neighborhoods in the case of the (ε, λ)-topology for random locally convex modules, we can not give a
proof of Theorem 2 by directly following the proof of Lemma 1 of [1]. This forces us to use the locally
L0-convex topology for random locally convex modules. The locally L0-convex topology was introduced
by Filipovi´c, et.al [3] in 2009. Since this note involves this topology only once, to save space, we do not
introduce this topology.
The bridge connects the (ε, λ)-topology and the locally L0-convex topology is the countable con-
catenation property. Let us recall the notion of the countable concatenation property, which was first
introduced by Guo in [4]: let E be an L0(F , K) -- module and G a subset of E. G is said to have the
countable concatenation property if there is g ∈ G for any sequence {gn : n ∈ N } in G and for any
countable partition {An : n ∈ N } of Ω to F such that IAn g = IAn gn for all n ∈ N .
5
We first show that an L0-convex L0-convexly compact set in a random locally convex module must
have the countable concatenation property.
Proposition 5 Let (E, P) be a random locally convex module over K with base (Ω, F , P ) and G a
nonempty L0-convex subset of E. If G is L0-convexly compact, then G must have the countable con-
catenation property.
Proof. We can, without loss of generality, assume the null element θ of E is in G, otherwise, we make
a translation. Let {xn : n ∈ N } be an arbitrary sequence in G and {An : n ∈ N } an arbitrary countable
partition of Ω to F . We need to show that there exists an x ∈ G such that: IAn x = IAn xn, ∀n ∈ N . For
each n, let yn = IA1 x1+ IA2x2+· · ·+ IAn xn, then by the L0-convexity of G and the assumption θ ∈ G, we
have yn ∈ G for every n. Further, for each n, let Gn = convL0{yk, k ≥ n} ∩ G. Then {Gn : n ∈ N } is a
sequence of closed L0-convex subset of G with the finite intersection property. Note by the construction,
each member z in convL0{yk, k ≥ n} must satisfy: IAk z = IAk xk, k = 1, 2, . . . , n, which in turn implies
that each member z in convL0{yk, k ≥ n} must satisfy: IAk z = IAk xk, k = 1, 2, . . . , n. By the L0-
convexly compactness of G, there exists some x ∈ E such that x is in every Gn. Thus x ∈ G and
IAk x = IAk xk for all k ∈ N , completing the proof.
(cid:3)
Now we state and prove the main result in this section.
Theorem 2 Let (E, P) be a random locally convex module over K with base (Ω, F , P ) and G a
nonempty L0-convex subset of E. If G is L0-convexly compact, then G must be closed.
Proof. Suppose that G is not closed and let x0 ∈ G \ G. Since G is L0-convexly compact, by Proposition
5, G has the countable concatenation property, then according to Theorem 3.12 in [4], G = Gc, where
Gc is the closure of G under the locally L0-convex topology. Let F in(P) denote the family of all finite
and nonempty subsets of P, for any Q ∈ F in(P) and any ξ ∈ L0
++
, {η ∈ L0 : P {η > 0} = 1},
let Uθ(Q, ξ) = {x ∈ E : kxkQ ≤ ξ}, then {Uθ(Q, ξ) : Q ∈ F in(P), ξ ∈ L0
++} is a local base of the
locally L0-convex topology. Since x0 ∈ Gc, each FQ,ξ := (x0 + Uθ(Q, ξ)) ∩ G is nonempty, which
in turn implies that the family {FQ,ξ : Q ∈ F in(P), ξ ∈ L0
++} has the finite intersection property.
Now each Uθ(Q, ξ) is L0-convex and closed under the (ε, λ)-topology, it follows that each FQ,ξ is an
L0-convex and closed subset of G, then by the assumption that G is L0-convexly compact, we obtain
T{FQ,ξ : Q ∈ F in(P), ξ ∈ L0
++} = {x0} ∩ G is nonempty, which is a contradiction.
(cid:3)
5 L0-convexly compactness and convexly compactness
In this section, we show that every L0-convex L0-convexly compact subset of a complete random locally
convex module must be convexly compact. This answers the question posed in Remark 2.6 in [8]. The
6
key idea of our proof is to embed a random locally convex module into the product space of a family of
random normed modules.
Proposition 6 Let (E, P) be a random locally convex module over K with base (Ω, F , P ), then there
exists a family of random normed modules {(Eq, k · kq), q ∈ Q} together with an L0(F , K)-module
homomorphism h : E → Q
q∈Q
Eq such that (E, Tε,λ) is topological homeomorphism to (h(E), T ), where
T is the product topology of Q
q∈Q
(Eq, Tε,λ).
Proof. Let Q be the family of continuous L0−seminorms on E. For each q ∈ Q, let Nq = {x ∈ E :
q(x) = 0}. Since q is a continuous L0−seminorm on E, Nq is a closed sub L0(F , K)-module of (E, Tε,λ).
Let Eq = E/Nq be the quotient L0(F , K)-module and πq denote the canonical quotient mapping from
E to Eq. Define kπq(x)kq = q(x), ∀x ∈ E, then it is easy to verify that (Eq, k · kq) is a random normed
module. Further, define h : E → Q
q∈Q
Eq by h(x)q = πq(x), ∀x ∈ E, q ∈ Q, where h(x)q stands for
the q-th coordinate of h(x). It is easy to verify that h is an L0(F , K)-module homomorphism and a
topological homeomorphism from (E, Tε,λ) to (h(E), T ).
(cid:3)
Remark 1 In Proposition 6, if (E, P) is complete with respect to Tε,λ, then h(E) must be closed in
Eq, T ). Also, each member (Eq, k · kq) of the family of random normed modules can be assumed to
( Q
q∈Q
be complete with respect to Tε,λ, otherwise we can replace it by its completion.
To prove Theorem 3 below, we need the help of abstract Lp space generated from a random
normed module. Assume (E, k · k) is a random normed module over K with base (Ω, F , P ),
let
L2(E) = {x ∈ E RΩ kxk2dP < +∞}, then (L2(E), k · k2) is an ordinary normed space, where
2 , ∀x ∈ L2(E), and if (E, k · k) is complete with respect to the (ε, λ)-topology, then
kxk2 = (RΩ kxkdP )
(L2(E), k · k2) is a Banach space.
1
Theorem 3 Let (E, P) be a complete random locally convex module over K with base (Ω, F , P ) and G
a closed L0-convex subset of E. If G is L0-convexly compact, then it is convexly compact.
Proof. According to Proposition 6, there exists a family of complete random normed modules {(Eq, k ·
kq), q ∈ Q} such that we can take (E, Tε,λ) as a closed sub L0(F , K)-module in the product space
Q
q∈Q
(Eq, Tε,λ). For each q ∈ Q, let πq : Q
r∈Q
Er → Eq be the canonical projection mapping. If an L0-
convex subset G of E is L0-convexly compact in (E, Tε,λ), then for each q ∈ Q, it is easily verified that
πq(G) is an L0-convexly compact subset of (Eq, Tε,λ), which in turn implies that ξq := ∨{kπq(x)kq : x ∈
G} + 1 ∈ L0
++ by Theorem 1. Let Gq = πq(G)
ξq
, then Gq ⊂ L2(Eq), and on Gq, the (ε, λ)-topology which
7
inherited from (Eq, Tε,λ) and the k · k2-topology which inherited from (L2(Eq), k · k2) coincide by the
Lebesgue dominance convergence theorem. Moreover, as shown in the proof of Theorem 2.5 in [7], Gq is
a weak compact subset in (L2(Eq), k·k2). If w indicates the weak topology, then by Tyhonoff's theorem,
Q
q∈Q
Gq is a compact subset in Q
q∈Q
(L2(Eq), w). If F is a subset of Q
q∈Q
Gq, then F is closed under Tε,λ
inherited from Q
q∈Q
Eq iff F is closed under the topology inherited from Q
q∈Q
(L2(Eq), k · k2), and if F
is convex, then F is closed in Q
q∈Q
(L2(Eq), k · k2) iff F is closed in ( Q
q∈Q
L2(Eq), w) = Q
q∈Q
(L2(Eq), w),
therefore by the definition of convexly compactness and from the fact that Q
q∈Q
Gq is a compact subset
in Q
q∈Q
(L2(Eq), w), we obtain: Q
q∈Q
Gq is a convexly compact set in Q
q∈Q
(Eq, Tε,λ). Finally, define f :
G → Q
q∈Q
Gq by πq(f (x)) = πq(x)
ξq
for every x ∈ G and each q ∈ Q, then (G, Tε,λ) is topological
homeomorphism to (f (G), Tε,λ) which is a convex closed subset of ( Q
q∈Q
Gq, Tε,λ), implying that G is
convexly compact.
(cid:3)
Acknowledgements. The work of the authors was supported by the Natural Science Foundations
of China (Grant No.11701531 and No.11401399) and the Fundamental Research Funds for the Central
Universities, China University of Geosciences (Wuhan)(Grant No. CUGL170820).
References
[1] M.Cristea, D.Dascalu and L.N.Nasta, Walrasian economy and some properties of convexly compact
sets, Universitatii "Ovidius" Constanta. Analele Stiintifice 25(1)(2017),69 -- 76.
[2] N.Dunford, J.T. Schwartz, Linear Operators (I): General Theory, John Wiley & Sons Inc., New
York, 1958.
[3] D.Filipovi´c, M. Kupper, N.Vogelpoth, Separation and duality in locally L0 -- convex modules, J.
Funct. Anal., 256 (2009), 3996 -- 4029.
[4] T. X. Guo, Relations between some basic results derived from two kinds of topologies for a random
locally convex module, J. Funct. Anal., 258(2010), 3024-3047.
[5] Tiexin Guo, Haixia Xiao, Xinxiang Chen, A basic strict separation theorem in random locally convex
modules, Nonlinear Anal. 71 (2009) 3794 -- 3804.
8
[6] T. X. Guo, X. L. Zeng, An L0(F , R)-valued function's intermediate value theorem and its applica-
tions to random uniform convexity, Acta Math Sin (Engl. Ser.), 28(2012), 909-924.
[7] T.X.Guo, E.X.Zhang, Y.C.Wang and M.Z.Wu, L0-convex compactness and its applications, arXiv:
1709.07137V2.
[8] T.X.Guo, E.X.Zhang, Y.C.Wang and G. Yuan, Kirk's fixed point theorem in complete random
locally convex module, arXiv: 1801.09341.
[9] T.X.Guo, E.X.Zhang, M.Z.Wu, B.X.Yang, G.Yuan and X.L.Zeng, On random convex analysis,
Journal of Nonlinear and Convex Analysis, 18(11) (2017), 1967-1996.
[10] T.X.Guo, S.E.Zhao and X.L.Zeng, The relations among the three kinds of conditional risk meaures,
Science China Mathematics 57(8) (2014), 1753-1764.
[11] T.X.Guo, S.E.Zhao and X.L.Zeng, Random convex analysis (I): separation and Fenchel- Moreau
duality in random locally convex modules (in Chinese), Scientia Sinica Mathematica, 45(12) (2015),
1960-1980 (see also arXiv: 1503.08695V3).
[12] T.X.Guo, S.E.Zhao and X.L.Zeng, Random convex analysis(II): continuity and subdifferentiability
in L0-pre-barreled random locally convex modules (in Chinese), Scientia Sinica Mathematica, 45(5)
(2015), 647-662 (see also arXiv: 1503.08637V2)
[13] G. Zitkovi´c, Convex compactness and its applications, Math. Finan. Eco. 3(1) (2010), 1-12.
9
|
1302.7207 | 1 | 1302 | 2013-02-28T14:34:52 | G-convergence of linear differential equations | [
"math.FA"
] | We discuss $G$-convergence of linear integro-differential-algebaric equations in Hilbert spaces. We show under which assumptions it is generic for the limit equation to exhibit memory effects. Moreover, we investigate which classes of equations are closed under the process of $G$-convergence. The results have applications to the theory of homogenization. As an example we treat Maxwell's equation with the Drude-Born-Fedorov constitutive relation. | math.FA | math | Als Typoskript gedruckt
Technische Universität Dresden
Herausgeber: Der Rektor
3
1
0
2
b
e
F
8
2
]
.
A
F
h
t
a
m
[
1
v
7
0
2
7
.
2
0
3
1
:
v
i
X
r
a
G-convergence of linear differential
equations.
Marcus Waurick
Institut für Analysis
MATH-AN-03-2013
G-convergence of linear differential
equations.
Marcus Waurick
Institut für Analysis, Fachrichtung Mathematik
Technische Universität Dresden
Germany
[email protected]
October 3, 2018
Abstract. We discuss G-convergence of linear integro-differential-algebaric
equations in Hilbert spaces. We show under which assumptions it is generic for the
limit equation to exhibit memory effects. Moreover, we investigate which classes
of equations are closed under the process of G-convergence. The results have ap-
plications to the theory of homogenization. As an example we treat Maxwell's
equation with the Drude-Born-Fedorov constitutive relation.
Keywords and phrases: G-convergence, integro-differential-algebraic equations, homogenization, integral
equations, Maxwell's equations
Mathematics subject classification 2010: 34L99 (Ordinary differential operators), 34E13 (Ordinary differ-
ential equations, multiple scale methods), 34A08 (Fractional differential equations), 35B27 (Partial differential
equations, homogenization; equations in media with periodic structure), 35Q61 (Maxwell's equations), 45A05
(Linear integral equations)
3
Contents
Contents
1 Introduction
2 Setting and main theorems
3 Time-independent coefficients
4 Time-translation invariant coeffcients
5 Time-dependent coefficients
6 Proof of the main theorems
References
5
6
10
15
19
23
26
4
1 Introduction
We discuss some issues occuring in the homogenization of linear integro-differential equations
in Hilbert spaces. Similar to [19, 18, 20, 21] we understand homogenization theory as the study
of limits of sequences of equations in the sense of G-convergence. Whereas in [19, 18] (non-
linear) ordinary differential equations in finite-dimensional space are considered, we choose
the perspective given in [14, 15, 20, 21, 1, 10, 11]. The abstract setting is the following.
Definition 1.1 (G-convergence, [28, p. 74], [25]). Let H be a Hilbert space. Let (An :
D(An) j H → H)n be a sequence of continuously invertible linear operators onto H and let
B : D(B) j H → H be linear and one-to-one. We say that (An)n G-converges to B if(cid:0)A−1
n (cid:1)n
converges in the weak operator topology to B−1, i.e., for all f ∈ H the sequence (A−1
n (f ))n
converges weakly to some u, which satisfies u ∈ D(B) and B(u) = f . B is called the1 G-limit
of (An)n and we write An
G−→ B.
Our starting point will be equations of the form
∂0Mu + N u = f,
where M,N are suitable operators in space-time and ∂0 is the time-derivative established in
In the usual framework of
a Hilbert space setting to be specified below (see also [16, 8]).
homogenization theory, one assumes M and N to be multiplication operators in space-time,
i.e., there are mappings a and b such that M = a(·) and N = b(·). Assuming well-posedness of
the above equation, i.e., existence, uniqueness and continuous dependence on the right-hand
side f in a suitable (Hilbert space) framework, one is interested in the sequence of equations
∂0Mnun + Nnun = f
(1)
with Mn = a(n·) and correspondingly for Nn yielding a sequence of solutions (un)n. The
question arises, whether the sequence (un)n converges and if so whether the respective limit
u satisfies an equation of similar form. A formal computation in (1) reveals that
un = (∂0Mn + Nn)−1 f.
Thus, if we show the convergence of (∂0Mn + Nn)−1 in the weak opertor topology to some
one-to-one mapping C =: B−1, we deduce the weak convergence of (un)n the limit of which
denoted by u satisfies
Bu = f.
In other words, (∂0Mn + Nn) G-converges to B.
In this article we think of (Mn)n and (Nn)n to be bounded sequences of bounded linear
operators in space-time. We want to discuss assumptions on these sequences guaranteeing a
compactness result with respect to G-convergence. Moreover, we outline possible assumptions
yielding the closedness under G-convergence and give examples for equations, where the asso-
ciated sequences of differential operators itself are G-convergent. We exemplify our findings
with examples from the literature [14, 20, 21, 10, 11], highlight possible connections and give
1Note that the G-limit is uniquely determined, cf. [25, Proposition 4.1].
5
2 Setting and main theorems
an example for a Drude-Born-Fedorov model in electro-magentism (see [7] and Example 3.11
below), where homogenization theorems are -- to the best of the author's knowledge -- not yet
available in the literature. We will also underscore the reason of the limit equation to exhibit
memory effects. An heuristic explanation is the lack of continuity of computing the inverse
with respect to the weak operator topology.
In Section 2 we introduce the functional analytic setting used for discussing integro-differential-
algebraic equations and state our main Theorems. We successively apply the results from
Section 2 to time-independent coefficients (Section 3), time-translation invariant coefficients
(Section 4) and time-dependent coefficients (Section 5). In each of the Sections 3, 4 and 5
we give examples and discuss whether particular classes of equations are closed under limits
with respect to G-convergence. In Section 6 we prove the main theorems of Section 2. The
respective proofs rely on elementary Hilbert space theory.
2 Setting and main theorems
The key fact giving way for computations is the possibility of establishing the time-derivative
as a continuously invertible normal operator in an exponentially weighted Hilbert space. For
ν > 0 we define the operator
∂0 : Hν,1(R) j L2
ν(R) → L2
ν(R), f 7→ f ′,
ν(R). We denote the scalar-product on L2
ν(R) := L2(R, exp(−2ν·)λ) is the space of square-integrable functions with respect to
where L2
the weighted Lebesgue measure exp(−2ν·)λ and Hν,1(R) is the space of L2
ν(R)-functions with
distributional derivative in L2
ν(R) by h·,·iν and the
induced norm by ·ν . Of course the operator ∂0 depends on the scalar ν. However, since it will
be obvious from the context, which value of ν is chosen, we will omit the explicit reference to
it in the notation of ∂0. It can be shown that ∂0 is continuously invertible ([16, Example 2.3]
or [8, Corollary 2.5]). The norm bound of the inverse is 1/ν. Of course the latter construction
can be extended to the Hilbert-space-valued case of L2
ν(R; H)-functions2. We will use the
same notation for the time-derivative. In order to formulate our main theorems related to the
theory of homogenization of ordinary differential equations, we need to introduce the following
notion.
Definition 2.1. Let H0, H1 be Hilbert spaces, ν0 > 0. We call a linear mapping
M : D(M ) j \ν>0
L2
ν(R; H0) → \ν≧ν0
L2
ν(R; H1)
(2)
evolutionary (at ν1 > 0)3 if D(M ) is dense in L2
bounded linear operator from L2
ν(R; H0) to L2
ν(R; H) for all ν ≧ ν1, if M extends to a
ν(R; H1) for all ν ≧ ν1 and is such that4
ν→∞ kMkL(L2
lim sup
ν (R;H0),L2
ν (R;H1)) < ∞.
2We will also use the notation h·, ·iν and ·ν for the scalar product and norm in L2
3The notion "evolutionary" is inspired by the considerations in [16, Definition 3.1.14, p. 91], where polynomial
ν(R; H), respectively.
expressions in partial differential operators are considered.
4For a linear operator A from L2
ν (R; H0) to L2
ν (R; H1) we denote its operator norm by kAkL(L2
ν (R;H0),L2
ν (R;H1)).
If the spaces H0 and H1 are clear from the context, we shortly write kAkL(L2
ν ).
6
The continuous extension of M to some L2
In particular, we
will not distinguish notationally between the different realizations of M as a bounded linear
operator for different ν as these realizations coincide on a dense subset. We define the set
ν will also be denoted by M .
Lev,ν1(H0, H1) := {M ; M is as in (2) and is evolutionary at ν1}.
We abbreviate Lev,ν1(H0) := Lev,ν1(H0, H0). A subset M j Lev,ν1(H0, H1) is called bounded
if lim supν→∞ supM ∈M kMkL(L2
ν ) < ∞. A family (Mι)ι∈I in Lev,ν1(H0, H1) is called bounded
if {Mι; ι ∈ I} is bounded.
Note that Lev,ν1(H0, H1) j Lev,ν2(H0, H1) for all ν1 ≦ ν2. We give some examples of evolu-
tionary mappings.
Example 2.2. Let H be a Hilbert space and M0 ∈ L(H). Then there is a canonical extension
M of M0 to L2
ν(R; H) and a.e.
t ∈ R. In that way M ∈ Tν>0 Lev,ν(H). Henceforth, we shall not distinguish notationally
ν(R; H)-functions such that (M φ) (t) := (M0φ(t)) for all φ ∈ L2
between M and M0.
Example 2.3. Let H be a Hilbert space and let L∞
measurable functions from R to L(H). For A ∈ L∞
tiplication operator on L2
s (R; L(H)) be the space of bounded strongly
s (R; L(H)) we denote the associated mul-
ν(R; H) by A(m0). Thus, also in this case, A(m0) ∈Tν>0 Lev,ν(H).
loc(R); g = 0 on R<0,RR g(t)e−νtdt <
Example 2.4. For ν0 > 0 let g ∈ L1
ν0(R>0) := {g ∈ L1
∞}. By Young's inequality or by Example 4.3 below, we deduce that g∗ ∈ Lev,ν0(C), where
g ∗ f denotes the convolution of some function f with g.
To formulate our main theorems, we denote the weak operator topology by τw. Convergence
within this topology is denoted by τw→. Limits within this topology are written as τw- lim.
We will extensively use the fact that for a separable Hilbert space H bounded subsets of
L(H), which are τw-closed, are τw-sequentially compact. Our main theorems concerning the
G-convergence of differential equations read as follows.
Theorem 2.5. Let H be a separable Hilbert space, ν0 > 0. Let (Mn)n, (Nn)n be bounded
sequences in Lev,ν0(H). Assume there exists c > 0 such that for all n ∈ N and ν ≧ ν0
RehMnφ, φiν ≧ chφ, φiν
(φ ∈ L2
ν(R; H)).
Then there exists ν ≧ ν0 and a subsequence (nk)k of (n)n such that
∂0Mnk + Nnk
G−→ ∂0M−1
hom,0 + ∂0
∞
Xj=1 −
∞
Xℓ=1
hom,0Mhom,ℓ!j
M−1
M−1
hom,0,
as k → ∞ in L2
ν(R; H), where
Mhom,0 = τw- lim
k→∞M−1
nk
7
2 Setting and main theorems
and
Mhom,ℓ = τw- lim
k→∞M−1
n (cid:0)−∂−1
n (cid:1)ℓ
0 NnM−1
.
Remark 2.6. (a) It should be noted that the positive-definiteness condition in Theorem 2.7 is
a well-posedness condition, i.e., a condition for ∂0Mnk +Nnk to be continuously invertible for
all ν sufficiently large. Indeed, for f ∈ L2
ν(R; H) and u ∈ L2
ν(R; H) with
we multiply by ∂−1
0
and get
The positive definiteness condition yields, see also Lemma 6.1, the invertibility of Mn. Hence,
we arrive at
(∂0Mn + Nn) u = f
(cid:0)Mn + ∂−1
0 Nn(cid:1) u = ∂−1
0 Nn(cid:1) u = M−1
(cid:0)1 + M−1
n ∂−1
0 f.
n ∂−1
0 f.
0 Nn(cid:1) is continu-
Choosing ν > 0 sufficiently large, we deduce that the operator (cid:0)1 + M−1
ously invertible with a Neumann series expression.
(b) If N = 0 in Theorem 2.5, then we deduce that equations of the form ∂0Mu = f are closed
under the process of G-convergence. If N 6= 0, then the above theorem suggests that this is
not true for equations of the form (∂0M + N ) u = f . However, if we consider ∂0M + N as
∂0(cid:0)M + ∂−1
0 N(cid:1), the equations under consideration in Theorem 2.5 are closed under G-limits.
Indeed, the limit may be represented by
n ∂−1
∂0
M−1
hom,0 +
∞
Xk=1 −
∞
Xℓ=1
hom,0Mhom,ℓ!k
M−1
hom,0
M−1
.
In the forthcoming sections we will further elaborate the aspect of closedness under G-limits.
In system or control theory one is interested in differential-algebraic systems, see e.g. [9]. We,
thus, formulate the analogous statement for (integro-differential-)algebraic systems.
n (cid:17)n
Theorem 2.7. Let H0, H1 be separable Hilbert spaces, ν0 > 0. Let (Mn)n ,(cid:16)N ij
be bounded
sequences in Lev,ν0(H0) and Lev,ν0(Hj, Hi), respectively (i, j ∈ {0, 1}). Assume there exists
c > 0 such that for all n ∈ N and ν ≧ ν0 we have for all (φ, ψ) ∈ L2
ν(R; H0 ⊕ H1)
n ψ, ψiν ≧ cψ2
ν .
RehMnφ, φiν ≧ cφ2
ν , RehN 11
Then there exists ν ≧ ν0 and a subsequence (nk)k of (n)n such that
0
nk
nk(cid:19)
∂0(cid:18) Mnk 0
0(cid:19) +(cid:18) N 00
nk N 01
nk N 11
N 10
0 1(cid:19)
hom,−1,11!
M−1
G−→(cid:18) ∂0 0
N −1
hom,−1,11!M(1)!ℓ M−1
Xℓ=1 − M−1
0
N −1
hom,0,00
hom,0,00
+
∞
0
0
0
hom,0,00
0
0
hom,−1,11!
,
N −1
8
where we put
as well as
and
n
n (cid:1)−1
n (cid:0)N 11
Nn := N 00
n − N 01
N 10
(n ∈ N)
ℓ=0 Mhom,ℓ,11(cid:19)
M(1) :=(cid:18)P∞
ℓ=1 Mhom,ℓ,00 P∞
ℓ=0 Mhom,ℓ,01
P∞
ℓ=0 Mhom,ℓ,10 P∞
nk(cid:1)ℓ
nk (cid:0)−∂−1
0 NnkM−1
k→∞M−1
nk(cid:1)−1
nk(cid:0)N 11
nk(cid:1)ℓ
nk (cid:0)−∂−1
∂−1
0 N 01
0 NnkM−1
k→∞−M−1
nk(cid:1)−1
nk(cid:1)ℓ
nk (cid:0)−∂−1
k→∞−(cid:0)N 11
0 NnkM−1
N 10
nkM−1
nk(cid:1)ℓ
nk (cid:0)−∂−1
nk(cid:1)−1
k→∞(cid:0)N 11
∂−1
0 N 01
0 NnkM−1
nkM−1
N 10
nk(cid:1)−1
k→∞(cid:0)N 11
.
,
,
Mhom,ℓ,00 = τw- lim
Mhom,ℓ,01 = τw- lim
Mhom,ℓ,10 = τw- lim
Mhom,ℓ,11 = τw- lim
Nhom,−1,11 = τw- lim
,
nk(cid:1)−1
nk(cid:0)N 11
,
Remark 2.8. (a) As in Theorem 2.5 the positive definiteness conditions in Theorem 2.7 serve
as well-posed conditions for the respective (integro-differential-algebraic) equations. We will
compute the respetive inverse in the proof of Theorem 2.7.
(b) Note that, by the definition of G-convergence, both the Theorems 2.5 and 2.7 implicitly
assert that the limit equations are well-posed, i.e., that the limit operator is continuously
invertible.
In fact it will be the strategy of the respective proofs to compute the limit of
the respective solution operators, which will be continuous linear operators and afterwards
inverting the limit.
(c) Assume that in Theorem 2.7 the expressions Mhom,ℓ,00, Mhom,ℓ,01, Mhom,ℓ,10, Mhom,ℓ,11,
Nhom,−1,11 can be computed without choosing subsequences. Then the sequence
(cid:18)∂0(cid:18) Mn 0
0 0(cid:19) +(cid:18) N 00
n N 01
N 10
n N 11
n (cid:19)(cid:19)n
n
is G-convergent. Indeed, the latter follows with a subsequence argument.
(c) Assuming H1 = {0} and, as a consequence, N ij = 0 for all i, j ∈ {0, 1} except i = j = 0,
we see that Theorem 2.7 is more general than Theorem 2.5. The generalization in Theorem
2.7 is also needed in the theory of homogenization of partial differential equations, see e.g. [25,
Theorem 4.4] for a more restrictive case. We give an example in the forthcoming sections.
For convenience, we include easy examples that show that the assumptions in the above
theorems are reasonable.
Example 2.9 (Uniform positive definiteness condition does not hold, [25]). Let H = C, ν > 0
and, for n ∈ N, let Mn = ∂−1
ν(R) \ {0}. For n ∈ N, let un ∈ L2
ν(R) be defined by
n , f ∈ L2
0
1
Then (un)n is not relatively weakly compact and contains no weakly convergent subsequence.
∂0Mnun =
1
n
un = f.
9
3 Time-independent coefficients
Example 2.10 (Boundedness assumption does not hold). Let H = C, ν > 0 and, for n ∈ N,
let Mn = ∂−1
ν(R). For n ∈ N, let un ∈ L2
0 n, f ∈ L2
ν(R) be defined by
∂0Mnun = nun = f.
Then (un)n converges to 0. Thus, a limit "equation" would be in fact the relation {0}×L2
L2
ν(R) ⊕ L2
We will now apply our main theorems to particular situations.
ν(R).
ν (R) j
3 Time-independent coefficients
In this section, we treat time-independent coefficients. That is to say, we assume that the
operators in the sequences under consideration only act on the "spatial" Hilbert spaces H0
and H1 in Theorem 2.7 or H in Theorem 2.5. More precisely and similar to Example 2.2, for
a bounded linear operator M ∈ L(H0, H1) there is a (canonical) extension to L2
ν-functions in
the way that (M φ)(t) := M (φ(t)) for φ ∈ L2
ν(R; H0) and a.e. t ∈ R. Thus M is evolutionary
(Example 2.2). We only state the specialization of this situation for Theorem 2.5. The result
reads as follows.
Theorem 3.1. Let H be a separable Hilbert space, ν0 > 0. Let (Mn)n, (Nn)n be bounded
sequences in L(H). Assume there exists c > 0 such that for all n ∈ N
Then there exists ν ≧ ν0 and a subsequence (nk)k of (n)n such that
RehMnφ, φiH ≥ chφ, φiH (φ ∈ H).
∂0Mnk + Nnk
G−→ ∂0M −1
hom,0 + ∂0
∞
Xj=1 −
∞
Xℓ=1
M −1
0 (cid:1)ℓ!j
hom,0Mhom,ℓ(cid:0)−∂−1
M −1
hom,0,
as k → ∞ in L2
ν(R; H), where
Mhom,0 = τw- lim
k→∞
M −1
nk
and
Mhom,ℓ = τw- lim
k→∞
.
M −1
nk (cid:0)NnK M −1
nK(cid:1)ℓ
0 = ∂−1
ν(R; H) and the extended M . Hence, the result follows from Theorem 2.5.
Proof. At first observe that M ∂−1
0 M for all bounded linear operators M ∈ L(H).
Moreover, the estimate RehM φ, φiH ≧ chφ, φiH for φ ∈ H also carries over to the analogous
one for φ ∈ L2
Remark 3.2. As it has already been observed in [14, 21], the class of equations treated in
Theorem 3.1 is not closed under G-convergence in general. The next example shows that this
effect only occurs if the Hilbert space H is infinite-dimensional and the convergence of (Mn)n
and (Nn)n is "weak enough" in a sense to be specified below.
10
Example 3.3. Assume that H is finite-dimensional. Then (Mn)n and (Nn)n are a mere
bounded sequences of matrices with constant coefficients. In particular, the weak operator
topology coincides with the topology induced by the operator norm. Hence, the processes of
computing the inverse and computing the limit interchange and multiplication is a continuous
process as well. Thus, assuming (Mn)n and (Nn)n to be convergent with the respective limits
M and N , we compute
M −1
nk (cid:0)Nnk M −1
nk(cid:1)ℓ(cid:17)(cid:0)−∂−1
0 (cid:1)ℓ!j
(cid:16)τw- lim
k→∞
M −1
nk(cid:17)−1
∂0M −1
hom,0 + ∂0
k→∞
∞
+ ∂0
= ∂0(cid:16)τw- lim
Xj=1 −
= ∂0(cid:16) lim
Xj=1 −
+ ∂0
k→∞
∞
= ∂0M + ∂0
= ∂0M + ∂0
∞
= ∂0 1 +
= ∂0(cid:16)(cid:0)1 + N M −1∂−1
M −1
hom,0
∞
∞
∞
∞
k→∞
k→∞
M −1
M −1
M −1
M −1
M −1
0 (cid:1)ℓ!j
nk(cid:17)−1(cid:16)τw- lim
hom,0Mhom,ℓ(cid:0)−∂−1
Xj=1 −
Xℓ=1
nk(cid:17)−1
Xℓ=1(cid:16)τw- lim
nk(cid:17)−1
Xℓ=1(cid:16) lim
Xj=1 −
Xℓ=1
0 (cid:1)ℓ!j
Xj=1 −
Xℓ=1(cid:0)N M −1(cid:1)ℓ(cid:0)−∂−1
0 (cid:1)ℓ!−1
M = ∂0 ∞
Xℓ=1(cid:0)N M −1(cid:1)ℓ(cid:0)−∂−1
0 (cid:1)−1(cid:17)−1
M = ∂0(cid:0)M + N ∂−1
nk(cid:17)−1(cid:16) lim
M(cid:16)M −1(cid:0)N M −1(cid:1)ℓ(cid:17)(cid:0)−∂−1
M −1
k→∞
k→∞
∞
∞
∞
∞
nk(cid:1)ℓ(cid:17)(cid:0)−∂−1
nk (cid:0)Nnk M −1
0 (cid:1)ℓ!j
M
M = ∂0
∞
∞
Xj=0 −
Xℓ=0(cid:0)−N M −1∂−1
0 (cid:1) = ∂0M + N.
0 (cid:1)ℓ!j
(cid:16) lim
k→∞
M −1
nk(cid:17)−1
0 (cid:1)ℓ!j
M
Xℓ=1(cid:0)N M −1(cid:1)ℓ(cid:0)−∂−1
0 (cid:1)ℓ!−1
M
Thus, in finite-dimensional spaces, the above theorem restates the continuous dependence of
the solution on the coefficients. Note that we only used that multiplication and computing
the inverse are continuous operations. Hence, the above calculation literally expresses the fact
of continuous dependence on the coefficients if H is infinite-dimensional and the sequences
(Mn)n and (Nn)n converge in the strong operator topology. Thus, one can only expect that
the limit expression differs from the one, which one might expect, if the actual convergence of
the operators involved is strictly weaker than in the strong operator topology.
We will turn to a more sophisticated example. For this we recall the concept of periodicity in
Rn, see e.g. [4].
Definition 3.4. Let a : Rn → Cm×m be bounded and measurable. a is called ]0, 1[n-periodic,
if for all x ∈ Rn and k ∈ Zn we have a(x + k) = a(x).
Moreover, recall the following well-known convergence result on periodic mappings, cf. e.g. [4,
Theorem 2.6].
Theorem 3.5. Let a : Rn → Cm×m be bounded and measurable and ]0, 1[n-periodic. Then
(a(k·))k converges in L∞(Rn)m×m ∗-weakly to the integral mean R[0,1]n a(y)dy.
11
3 Time-independent coefficients
∂0a(k·) + b(k·)
∞
∞
+ ∂0
ν(R; H).
Remark 3.6. For any bounded measurable function a : Rn → Cm×m one can associate the
corresponding multiplication operator in L2(Rn)m. Hence, Theorem 3.5 states the fact that
in case of periodic a the sequence of associated multiplication operators of a(k·) converges in
the weak operator topology to the operator of multiplying with the respective integral mean.
Indeed, this follows easily from L2(Rn) · L2(Rn) = L1(Rn). See also [5, 8.10].
Example 3.7. Let H = L2 (Rn)m and let a, b : Rn → Cm×m be bounded, measurable and
]0, 1[n-periodic. We assume Re a(x) ≧ c for all x ∈ Rn. Observe that any polynomial in a and
b is ]0, 1[n-periodic and so is a−1 :=(cid:0)x 7→ a(x)−1(cid:1) . Thus, by Theorem 3.1, we deduce that
G−→ ∂0 Z[0,1]n
Xj=1 −
as k → ∞ in L2
Remark 3.8. In [20, Theorem 1.2] or [21] the author considers the equation (∂0 + bk(·)) uk = f
with (bk)k being a [α, β]-valued (for some α, β ∈ R) sequence of bounded, measurable mappings
depending on one spatial variable. Also in that exposition a memory effect is derived. However,
the method uses the concept of Young measures. The reason for that is the representation of
the solution being a function of the oscillating coefficent. More precisely, the convergence of
a(y)−1dy!−1
Xℓ=1 Z[0,1]n
a(y)−1(cid:0)b(y)a(y)−1(cid:1)ℓ
0 (cid:1)ℓ!
dy(cid:0)−∂−1
a(y)−1dy!−1
a(y)−1dy!−1
,
j
Z[0,1]n
the sequence (cid:0)etb(k·)(cid:1)k is addressed. In order to let k tend to infinity in this expression one
needs a result on the (weak-∗) convergence of (continuous) functions of bounded functions.
This is where the Young-measures come into play, see e.g. [2, Section 2] and the references
therein or [21, p. 930]. The result used is the following. There exists a family of probabilty
measures (νx)x supported on [α, β] such that for (a subsequence of) (k)k and all real continuous
functions G we have
Z[0,1]n
G ◦ bk(·) → R ∋ x 7→Z[α,β]
G(λ)dνx(λ)!
as k → ∞ in L∞ (R) ∗-weakly. The family (νx)x is also called the Young-measure associated
to (bk)k. With the help of the family (νx)x a convolution kernel is computed such that the
respective limit equation can be written as
∂0u(t, x) + b0(x)u −Z t
0
K(x, t − s)u(x, s)ds = f (x, t),
where b0 is a weak-∗-limit of a subsequence of (bk)k and K(x, t) = RR>0
e−λtdνx(λ) for a.e.
t ∈ R>0 and x ∈ R. The relation to our above considerations is as follows. The resulting
limit equation within our approach can also be considered as an ordinary differential equation
perturbed by a convolution term. Denoting limits with respect to the σ(L∞, L1)-topology by
∗-lim, we realize that Theorem 3.1 in this particular situation states that the limit equation
admits the form
∂0 + ∂0
∞
Xk=1 −
∞
Xℓ=1
∗- lim
k→∞
0 (cid:1)ℓ!k
(bk)ℓ(cid:0)−∂−1
12
= ∂0 + b0 +
∞
Xℓ=2
∗- lim
k→∞
(bk)ℓ(cid:0)−∂−1
0 (cid:1)ℓ−1
+ ∂0
∞
Xk=2 −
∞
Xℓ=1
∗- lim
k→∞
0 (cid:1)ℓ!k
(bk)ℓ(cid:0)−∂−1
as k → ∞ in L2
Theorem 1.5.6 and Remark 1.5.7], we deduce that the term
ν(R; L2(R)). Indeed, using [16, 6.2.6. Memory Problems, (b) p. 448] or [22,
∞
Xℓ=2
∗- lim
k→∞
(bk)ℓ(cid:0)−∂−1
0 (cid:1)ℓ−1
+ ∂0
∞
Xk=2 −
∞
Xℓ=1
∗- lim
k→∞
0 (cid:1)ℓ!k
(bk)ℓ(cid:0)−∂−1
can be represented as a (temporal) convolution. Moreover, note that the choice of subsequences
is the same. Indeed, in the above rationale with the Young measure approach, by a density
argument, it suffices to choose a subsequence of (bk)k such that any polynomial of (bk)k
converges ∗-weakly. This choice of subsequences also suffices to deduce G-convergence of the
respective equations within the operator-theoretic perspective treated in this exposition.
In the next example, we consider a partial differential equation, which can be reformulated
as ordinary differential equation in an infinite-dimensional Hilbert space. More precisely, we
treat Maxwell's equations with the Drude-Born-Fedorov material model, see e.g. [7]. In order
to discuss this equation properly, we need to introduce several operators from vector analysis.
Definition 3.9. Let Ω j R3 be open. Then we define5
curlc : C∞,c(Ω)3 j L2(Ω)3 → L2(Ω)3
φ 7→
0 −∂3 ∂2
0 −∂1
∂3
−∂2 ∂1
0
φ,
where we denote by ∂i the partial derivative with respect to the i'th variable, i ∈ {1, 2, 3}.
Moreover, introduce
divc : C∞,c(Ω)3 j L2(Ω)3 → L2(Ω)
Xi=1
(φ1, φ2, φ3) 7→
3
∂iφi.
We define curl0 := curlc, div0 := divc. The 0 serves as a reminder for (the generalization of)
the electric and the Neumann boundary condition, respectively. If Ω is simply connected, we
also introduce
curl⋄ : D(curl⋄) j L2(Ω)3 → L2(Ω)3
φ 7→ curl φ,
where D(curl⋄) := {φ ∈ D(curl); curl φ ∈ D(div0)} .
Remark 3.10. It can be shown that if Ω is simply connected with finite measure, then curl⋄
is a selfadjoint operator, see [6, 7]. In that reference it is also stated that curl⋄ has, except 0,
only discrete spectrum. In particular, this means that the intersection of the resolvent set of
5We denote by C∞,c(Ω) the set of arbitrarily often differentiable functions with compact support in Ω.
13
3 Time-independent coefficients
curl⋄ with R is non-empty. For other geometric properties of Ω resulting in the selfadjointness
of curl⋄, we refer to [17].
We now treat a homogenization problem of the Drude-Born-Fedorov model as treated in [7].
Example 3.11. Assume that Ω j R3 is open, simply connected and has bounded Lebesgue
measure. Invoking Remark 3.10 and following [7, Theorem 2.1], the equation
(cid:18)∂0 (1 + η curl⋄)(cid:18) ε 0
0 µ(cid:19) +(cid:18) 0 − curl⋄
0 (cid:19)(cid:19)(cid:18) E
0(cid:19)
H(cid:19) =(cid:18) J
curl⋄
(3)
for η ∈ R such that − 1
ν(R; L2(Ω)3) and given ε, µ ∈ L(L2(Ω)3) being
strictly positive selfadjoint operators, admits a unique solution (E, H) ∈ Hν,1(R; L2(Ω)3).6
Indeed, multiplying (3) by (1 + η curl⋄)−1 , we get that
η ∈ (curl⋄), J ∈ L2
(cid:18)∂0(cid:18) ε 0
0 µ(cid:19) + curl⋄ (1 + η curl⋄)−1(cid:18) 0 −1
1 0 (cid:19)(cid:19)(cid:18) E
H(cid:19) = (1 + η curl⋄)−1(cid:18) J
0(cid:19) .
Realizing that curl⋄ (1 + η curl⋄)−1 is a bounded linear operator by the spectral theorem for
the selfadjoint operator curl⋄, we get that (E, H) ∈ Hν,1(R; L2(Ω)3) solves the above equation.
Note that the equation derived from (3) is a mere ordinary differential equation in an infinite-
dimensional Hilbert space. Assume we are given bounded sequences of selfadjoint operators
(εn)n and (µn)n satisfying εn ≧ c and µn ≧ c for some c > 0 and all n ∈ N. For n ∈ N we
consider the problem
(cid:18)∂0(cid:18) εn 0
0 µn(cid:19) + curl⋄ (1 + η curl⋄)−1(cid:18) 0 −1
1 0 (cid:19)(cid:19)(cid:18) En
Hn(cid:19) = (1 + η curl⋄)−1(cid:18) J
0(cid:19)
and address the question of G-convergence of (a subsequence of)
(DBFn)n :=(cid:18)∂0(cid:18) εn 0
0 µn(cid:19) + curl⋄ (1 + η curl⋄)−1(cid:18) 0 −1
1 0 (cid:19)(cid:19)n
.
Clearly, Theorem 3.1 applies and we get that (a subsequence of) (DBFn)n G-converges to
∂0M −1
hom,0 + ∂0
∞
Xk=1 −
∞
Xℓ=1
M −1
0 (cid:1)ℓ!k
hom,0Mhom,ℓ(cid:0)−∂−1
M −1
hom,0,
as k → ∞ in L2
ν(R; H), where
Mhom,0 = τw- lim
k→∞(cid:18) ε−1
0
nk
0 µ−1
nk (cid:19)
and
Mhom,ℓ = τw- lim
k→∞(cid:18) ε−1
0
nk
0 µ−1
nk (cid:19)(cid:18)curl⋄ (1 + η curl⋄)−1(cid:18) 0 −µ−1
0 (cid:19)(cid:19)ℓ
ε−1
nk
nk
.
6Note that for (E, H) ∈ Hν,1(R; L2(Ω)3) being a solution of (3) can only be true in the distributional sense,
which can be made more precise with the help of the extrapolation spaces of curl⋄. We shall, however, not
follow this reasoning here in more details and refer again to [7] or [16, Chapter 2].
14
We have seen that the class of problems discussed in Theorem 3.1 in this section is not closed
under the G-convergence, unless N = 0.
4 Time-translation invariant coeffcients
In Theorem 3.1, we have seen that the limit equation can be described as a power series expres-
sion in ∂−1
0 . A possible way to generalize this is the introduction of holomorphic functions in
∂−1
0 , see [16, Section 6.1, page 427]. To make this precise, we need the spectral representation
for ∂−1
0 , the Fourier-Laplace transform Lν, which is given as the unitary operator being the
closure of
C∞,c(R) j L2
ν(R)
ν(R) → L2
φ 7→(cid:18)x 7→
1
√2πZR
e−ixy−νyφ(y)dy(cid:19) .
Denoting by m : D(m) j L2(R) → L2(R), f 7→ (x 7→ xf (x)) the multiplication-by-argument-
operator with maximal domain D(m), we arrive at the representation
∂−1
0 = L∗
ν(cid:18) 1
im + ν(cid:19)Lν.
Thus, for bounded and analytic functions M : B(r, r) → C with r > 1
2ν we define
νM(cid:18) 1
im + ν(cid:19)Lν,
M(cid:0)∂−1
0 (cid:1) := L∗
it+ν(cid:17) φ(t) for φ ∈ L2(R) and a.e. t ∈ R. We canonically
im+ν(cid:17) φ(cid:17) (t) := M(cid:16) 1
where (cid:16)M(cid:16) 1
Hilbert space H . In this way, the definition of M(cid:0)∂−1
0 (cid:1) can be generalized to bounded and
operator-valued functions M : B(r, r) → L(H0, H1) for Hilbert spaces H0 and H1. We denote
extend the above definitions to the case of vector-valued functions L2
ν(R; H) with values in a
H∞(B(r, r); L(H0, H1)) := {M : B(r, r) → L(H0, H1); M bounded, analytic} .
A subset M j H∞(B(r, r); L(H0, H1)) is called bounded, if
sup{kM (z)k; z ∈ B(r, r), M ∈ M} < ∞.
A family (Mι)ι∈I in H∞(B(r, r); L(H0, H1)) is bounded, if {Mι; ι ∈ I} is bounded.
We will treat some examples for H∞-functions of ∂−1
0 below, see also [23]. In this reference,
a homogenization theorem of problems of the kind treated in Theorem 2.5 with (Mn)n =
0 (cid:1)(cid:1)n for a bounded sequence (Mn)n in H∞ has been presented, see [23, Theorem 5.2].
(cid:0)Mn(cid:0)∂−1
Moreover, in [25, Theorem 4.4] a special case of an analogous result of Theorem 2.7 has been
presented and used. In order to state a G-convergence theorem in a more general situation,
note that
(cid:8)M(cid:0)∂−1
0 (cid:1) ; M ∈ H∞(B(r, r); L(H0, H1))(cid:9) j \1
2r <ν
Lev,ν(H0, H1).
15
4 Time-translation invariant coeffcients
The theorem reads as follows.
n(cid:17)n
. Let (Mn)n ,(cid:16)N ij
Theorem 4.1. Let H0, H1 be separable Hilbert spaces, ν0 > 0, r > 1
2ν0
be bounded sequences in H∞(B(r, r); L(H0)) and H∞(B(r, r); L(Hj , Hi)), respectively (i, j ∈
{0, 1}). Assume there exists c > 0 such that for all n ∈ N we have for all (φ, ψ) ∈ H0 ⊕ H1
and z ∈ B(r, r)
RehMn(z)φ, φiH0 ≧ cφ2
H0 , RehN 11
n (z)ψ, ψiH1 ≧ cψ2
H1 .
Then there exists ν > ν0 and a subsequence (nk)k of (n)n such that
where we put
as well as
M (1)(cid:0)∂−1
and
0
N 10
nk(cid:0)∂−1
∂0(cid:18) Mnk (cid:0)∂−1
0 (cid:1) N 01
0(cid:19) +(cid:18) N 00
0 (cid:1) 0
nk(cid:0)∂−1
0 (cid:1) N 11
0 1(cid:19)
Mhom,0,00(cid:0)∂−1
0 (cid:1)−1
G−→(cid:18) ∂0 0
Xℓ=1 − Mhom,0,00(cid:0)∂−1
0 (cid:1)−1
+
0
∞
nk (cid:0)∂−1
0 (cid:1)
0 (cid:1)(cid:19)
nk (cid:0)∂−1
Nhom,−1,11(cid:0)∂−1
0
0
0
N 10
n
0 (cid:1)−1!
0 (cid:1)−1! M (1)(cid:0)∂−1
Nhom,−1,11(cid:0)∂−1
n (cid:0)N 11
n (cid:1)−1
n − N 01
Nn := N 00
0 (cid:1) := P∞
0 (cid:1)ℓ P∞
0 (cid:1)(cid:0)∂−1
ℓ=1 Mhom,ℓ,00(cid:0)∂−1
ℓ=0 Mhom,ℓ,10(cid:0)∂−1
0 (cid:1)(cid:0)∂−1
0 (cid:1)ℓ P∞
P∞
Mnk (z)−1(cid:0)−Nnk (z)Mnk (z)−1(cid:1)ℓ
k→∞−Mnk (z)−1(cid:0)−Nnk (z)Mnk (z)−1(cid:1)ℓ
k→∞−(cid:0)N 11
k→∞(cid:0)N 11
k→∞(cid:0)N 11
2ν1(cid:17) for some ν > ν1 ≧ ν0.
nk (z)(cid:1)−1
nk (z)(cid:1)−1
nk (z)(cid:1)−1
N 10
, 1
2ν1
,
,
Mhom,ℓ,00(z) = τw- lim
k→∞
Mhom,ℓ,01(z) = τw- lim
Mhom,ℓ,10(z) = τw- lim
Mhom,ℓ,11(z) = τw- lim
Nhom,−1,11(z) = τw- lim
for all z ∈ B(cid:16) 1
0 (cid:1)!ℓ Mhom,0,00(cid:0)∂−1
0 (cid:1)−1
0
0
Nhom,−1,11(cid:0)∂−1
0 (cid:1)−1!
,
(n ∈ N)
ℓ=0 Mhom,ℓ,01(cid:0)∂−1
ℓ=0 Mhom,ℓ,11(cid:0)∂−1
0 (cid:1)(cid:0)∂−1
0 (cid:1)(cid:0)∂−1
0 (cid:1)ℓ+1!
0 (cid:1)ℓ+1
N 01
N 10
nk (z)(cid:0)N 11
nk (z)(cid:1)−1
nk (z)Mnk (z)−1(cid:0)−Nnk (z)Mnk (z)−1(cid:1)ℓ
nk (z)Mnk (z)−1(cid:0)−Nnk (z)Mnk (z)−1(cid:1)ℓ
,
,
N 01
nk (z)(cid:0)N 11
nk (z)(cid:1)−1
,
Proof. Observe that bounded and analytic functions of ∂−1
0 . Note that
the only thing left to prove is that the operator-valued functions involved are indeed analytic
functions of ∂−1
0 . For this we need to introduce a topology on H∞(B(r, r); L(H0, H1)). Let τ
be the topology induced by the mappings
commute with ∂−1
0
H∞(B(r, r); L(H0, H1)) → H(B(r, r))
M 7→ hφ, M (·)ψi,
16
for all (φ, ψ) ∈ H1 ⊕ H0, where H(B(r, r)) is the set of analytic functions endowed with the
compact open topology. In [23, Theorem 3.4] or [26, Theorem 4.3] it is shown that closed
and bounded subsets of H∞(B(r, r); L(H0, H1)) are sequentially compact with respect to the
τ -topology. Furthermore, by [23, Lemma 3.5], we have that if a bounded sequence (Tn)n in
H∞(B(r, r); L(H0, H1)) converges in the τ -topology then the operator sequence (cid:0)Tn(cid:0)∂−1
0 (cid:1)(cid:1)n
converges in the weak operator topology of L(cid:0)L2
ν(R; H0 ⊕ H1)(cid:1). Putting all this together, we
deduce that the assertion follows from Theorem 2.7.
Remark 4.2. (a) Theorem 4.1 asserts that the time-translation invariant equations under con-
sideration are closed under G-convergence. Though the formulas may become a bit cluttered,
in principle, an iterated homogenization procedure is possible.
(b) In [25, Theorem 4.4] operator-valued functions that are analytic at 0 were treated. This
assumption can be lifted. Indeed, we only require analyticity of the operator-valued functions
under consideration on the open ball B(r, r) for some radius r > 0 and do not assume that
any of these functions have holomorphic extensions to 0.
We give several examples.
Example 4.3. Let ν0 > 0. In this example we treat integral equations of convolution type. Let
(gn)n be a bounded sequence in L1
ν0(R>0) such that there is h ∈ L1
ν0(R>0) with kg(t)k ≦ h(t)
for all n ∈ N and a.e. t ∈ R. For f ∈ C∞,c(R) consider the equation
un + gn ∗ un = f.
(4)
The latter equation fits into the scheme of Theorem 4.1 for H = C. Indeed, using that the
Fourier transform F translates convolutions into multiplication, we get for any g ∈ L1
ν(R>0)
and u ∈ L2
ν(R) for some ν > ν0 that
g ∗ u = √2πL∗
= √2πL∗
= √2πL∗
νLνg(m)Lν u
ν (Fg) (m − iν)Lνu
ν (Fg)(cid:18)−i
1
(im + ν)−1(cid:19)Lνu.
The support and integrability condition of g implies analyticity of Mg := √2π (Fg)(cid:16)−i 1
(·)(cid:17)
. The computation also shows that
on B(r, r) for 0 < r < 1
2ν0
g ∗ u2
1
ν =(cid:12)(cid:12)(cid:12)(cid:12)
√2π (Fg)(cid:18)−i
≦ 2π(cid:12)(cid:12)(cid:12)(cid:12)
(Fg)(cid:18)−i
(im + ν)−1(cid:19)Lνu(cid:12)(cid:12)(cid:12)(cid:12)
(i(·) + ν)−1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
≦ 2π (Fg) ((·) − iν)2
∞ u2
ν ,
1
2
∞ Lνu2
L2
2
L2
.
17
4 Time-translation invariant coeffcients
where
2π (Fg) ((·) − iν)2
∞ := sup
t∈R
2π (Fg) (t − iν)2 =(cid:12)(cid:12)(cid:12)(cid:12)
ZR
≦(cid:18)ZR
e−i(t−iν)yg(y)dy(cid:12)(cid:12)(cid:12)(cid:12)
e−νy g(y) dy(cid:19)2
2
,
which tends to zero, if ν → ∞. Thus, by our assumption on the sequence (gn)n having the
uniform majorizing function h, there exists ν1 > 0 such that we have
ε := sup
n∈Nkgn ∗ kL(L2
ν1
(R)) < 1.
Hence, we can reformulate (4) as follows
(cid:0)1 + Mgn(cid:0)∂−1
0 (cid:1)(cid:1) un = f,
Thus with H0 = {0}, H1 = H and N 11 = (1 + Mgn)n Theorem 4.1 is applicable. (Note that
Re N 11
n ≧ 1 − ε > 0 for all n ∈ N). The assertion states that, for a suitable subsequence for
which we will use the same notation, we have
with
(cid:0)1 + Mgn(cid:0)∂−1
0 (cid:1)−1
0 (cid:1)(cid:1) G−→ Nhom,−1,11(cid:0)∂−1
Nhom,−1,11(z) = τw- lim
n→∞
(1 + Mgn(z))−1
= τw- lim
n→∞
1 +
= τw- lim
n→∞
1 +
Mgn(z)ℓ
M(gn)∗ℓ(z)
∞
∞
Xℓ=1
Xℓ=1
= τw- lim
n→∞
1 + MP∞
ℓ=1(gn)∗ℓ(z)
2ν1
, 1
for all z ∈ B(cid:16) 1
function g by g∗ℓ, ℓ ∈ N.
In [27] we discussed the following variant of Example 3.7.
2ν1(cid:17) for some ν > ν1 ≧ ν0, where we denoted the ℓ-fold convolution with a
Example 4.4. In the situation of Example 3.7, we let (hk)k be a convergent sequence of
positive real numbers with limit h. Then Theorem 4.1 gives
∂0a(k·) + τ−hk b(k·)
G−→ ∂0 Z[0,1]n
Xk=1 −
a(y)−1dy!−1
Xℓ=1 Z[0,1]n
+ ∂0
∞
∞
18
a(y)−1dy!−1
Z[0,1]n
a(y)−1(cid:0)τ−hb(y)a(y)−1(cid:1)ℓ
0 (cid:1)ℓ!
dy(cid:0)−∂−1
k
Z[0,1]n
a(y)−1dy!−1
.
Indeed, it suffices to observe that τ−h = L∗
νe−h(im+ν)Lν.
Fractional differential equations are also admissible as the following example shows.
Example 4.5. Again in the situation of Example 3.7, let (αk)k and (βk)k be convergent
sequences in ]0, 1] and [−1, 0] with limits α and β, resp. Then Theorem 4.1 gives
∂αk
0 a(k·) + ∂βk
0 b(k·) = ∂0∂αk−1
0
a(k·) + ∂βk
0 b(k·)
G−→ ∂α
∞
∞
+ ∂α
0
a(y)−1dy!−1
Xℓ=1
0 Z[0,1]n
Xk=1 −
a(y)−1dy!−1
· Z[0,1]n
.
∂α−1
0 Z[0,1]n
a(y)−1dy!−1
Z[0,1]n
a(y)−1∂1−α
0
0
(cid:16)∂1+β−α
b(y)a(y)−1(cid:17)ℓ
k
0 (cid:1)ℓ!
dy(cid:0)−∂−1
Remark 4.6. Note that all the above theorems on homogenization of differential equations
straightforwardly apply to higher order equations. For example the equation
∂k
0 aku = f
n
Xk=0
can be reformulated as a first order system in the standard way. Another way is to integrate
n − 1 times, to get that
n
∂1+k−n
0
aku = ∂−(n−1)
f,
0
Xk=0
which is by setting M (∂−1
Theorem 4.1.
0 ) = an and N (∂−1
k=0 ∂1+k−n
0
0 ) = Pn−1
ak of the form treated in
5 Time-dependent coefficients
In this section we treat operators depending on temporal and spatial variables, which are,
in contrast to the previous section, not time-translation invariant. Thus, the structural hy-
pothesis of being analytic functions of ∂−1
has to be lifted. Consequently, the expressions
for the limit equations do not simplify in the manner as they did in the Theorems 3.1 and
4.1. Particular ((non-)linear) equations have been considered in [14, 20, 10, 11, 15]. The main
objective of this section is to give a sufficient criterion under which the choice of subsequences
in Theorem 2.7 is not required. We introduce the following notion.
0
Definition 5.1. Let H be a Hilbert space. A family ((Tn,ι)n∈N)ι∈I of sequences of linear
operators in L(H) is said to have the product-convergence property, if for all k ∈ N and
converges in the weak operator topology of L(H).
(ι1, . . . , ιk) ∈ I k the sequence(cid:16)Qk
i=1 Tn,ιi(cid:17)n
19
5 Time-dependent coefficients
Example 5.2. Let N, M ∈ N and denote P := {a : RN → CM ×M ; a is [0, 1]N -periodic}.
Theorem 3.5 asserts that the family(cid:0)(a(k·))k∈N(cid:1)a∈P has the product-convergence property in
L(L2(RN )M ).
We refer to the notion of homogenization algebras for other examples, see e.g. [12, 13]. The
main theorem of this section reads as follows. Recall from Example 2.3 the space L∞
s (R; L(H))
of strongly measurable bounded functions with values in L(H) endowed with the sup-norm.
Moreover, recall that for A ∈ L∞
s (R; L(H)) the associated multiplication operator A(m0) is
evolutionary at ν for every ν > 0.
Theorem 5.3. Let H be a Hilbert space, ν > 0. Let (cid:0)(Aι,n)n(cid:1)ι be a family of bounded se-
s (R; L(H)). Assume that the family(cid:0)(Aι,n(t))n(cid:1)ι,t∈R has the product-convergence
property. Then (cid:0)(Aι,n(m0))n ,(cid:0)∂−1
Remark 5.4. (a) With the latter result, it is possible to deduce that the choice of subsequences
in Theorem 2.7 is not needed. Indeed, assume that
has the product-convergence property.
0 (cid:1)n(cid:1)ι
quences in L∞
(cid:18) Mn 0
0 0(cid:19) +(cid:18) N 00
n N 01
N 10
n N 11
n (cid:19) =(cid:18) Mn(m0) 0
0(cid:19) +(cid:18) N 00
N 10
0
n
n (m0) N 01
n (m0) N 11
n (m0)
n (m0)(cid:19)
for some strongly measurable and bounded M 1
n, N 00
n , N 01
n , N 10
n , N 11
n and assume that the family
0
(cid:18)(cid:18) Mn(t) 0
0(cid:19)n
(cid:18) 0 N 01
0 (cid:19)n
,(cid:18) Mn(t)−1 0
0(cid:19)n
,(cid:18) 0
n (t) 0(cid:19)n
n (t)
N 10
0
0
0
,(cid:18) N 00
,(cid:18) 0
n (t) 0
0
0(cid:19)n
n (t)(cid:19)n
0
0 N 11
,
,(cid:18) 0
0
n (t)−1(cid:19)n(cid:19)t∈R
0 N 11
satisfies the product-convergence property. Then Theorem 5.3 ensures that the limit expres-
sions in Theorem 2.7 converge without choosing subsequences.
(b) The crucial fact in Theorem 5.3 is that powers of ∂−1
are involved.
a Hilbert space, ν > 0. Let (cid:0)(Aι,n)n(cid:1)ι be a family of bounded sequences in L∞
Assume that, for every t ∈ R, the family (cid:0)(Aι,n(t))n(cid:1)ι has the product-convergence property.
Then (cid:0)(Aι,n(m0))n(cid:1)ιhas the product-convergence property. Showing the assertion for two se-
quences (A1,n)n and (A2,n)n and using the boundedness of the sequence (A1,n(m0)A2,n(m0))n,
we deduce that it suffices to show weak convergence on a dense subset. For this to show let
K, L j R be bounded and measurable and φ, ψ ∈ H. We get for n ∈ N and ν > 0 that
0
Indeed, let H be
s (R; L(H)).
hχK φ, A1,n(m0)A2,n(m0)χLψiν =ZK∩Lhφ, A1,n(t)A2,n(t)ψie−2νtdt
→ZK∩L
n→∞hφ, A1,n(t)A2,n(t)ψie−2νtdt,
lim
by dominated convergence.
quences in L∞
Lemma 5.5. Let H be a Hilbert space, ν > 0. Let (cid:0)(Aι,n)n(cid:1)ι∈I be a family of bounded se-
s (R; L(H)). Assume that the family(cid:0)(Aι,n(t))n(cid:1)ι,t∈R has the product-convergence
20
j=1(cid:0)Aιj ,n(m0), ∂−1
0 (cid:1)ℓj converges in the weak operator toplogogy for all
property. Then7 Qk
k ∈ N, ℓ1, . . . , ℓk ∈ {0, 1} × N and ι1, . . . , ιk ∈ I.
Proof. Let k ∈ N, ℓ1, . . . , ℓk ∈ {0, 1} × N and ι1, . . . , ιk ∈ I. Moreover, take φ, ψ ∈ H and
K, L j R be bounded and measurable. For n ∈ N and ν > 0 we compute
1
−∞
0
0
ν
k
0
−∞
ℓ1,2−1
Aιk,n(sk−1
*χKφ,
=ZK*φ, Aι1,n(s0
0 (cid:1)ℓj χLψ+
Yj=1(cid:0)Aιj ,n(m0), ∂−1
···Z s1
0)ℓ1,1Z s0
−∞Z s1
)ℓk,1Z sk−1
···Z sk
−∞ Z sk
ℓk,2−1 ··· ds2
0 ··· ds2
0 ··· dsk
···Z s1
−∞Z s1
−∞Z s1
−∞Z s2
Dφ, Aι1,n(t)ℓ1,1Aι2,n(s0)ℓ2,1 ··· χL(sk
0 ··· ds2
0 ··· dsk
=ZKZ s0
ℓk,2−1 ··· ds2
ℓk,2−2dsk
ℓk,2−2dsk
dsk
dsk
ℓk,2−1
ℓ1,2−1
ℓ2,2−1
−∞
−∞
−∞
−∞
1
0
0
ℓ2,2−2ds2
···Z s2
0)ψE
ℓ2,2−2ds2
1
χL(sk
0)ψ+
Aι2,n(s1
0
0)ℓ2,1Z s1
−∞Z s2
−∞
ℓ2,2−1
1
···Z s2
−∞ ···
ℓ1,2−2ds1
ℓ1,2−1e−2νs0
0ds0
0
ℓ2,2−1ds1
0 ··· ds1
−∞ ···Z sk−1
−∞ Z sk
0
1
−∞
ℓk,2−1
1
···Z sk
−∞
ℓ2,2−1ds1
0 ··· ds1
ℓ1,2−2ds1
ℓ1,2−1e−2νs0
0ds0
0.
Using dominated convergence, we deduce the convergence of the latter expression.
Proof of Theorem 5.3. The proof follows easily with Lemma 5.5.
Theorem 5.3 serves as a possibility to deduce G-convergence of differential operators, where
the coefficients take values in, for example, periodic mappings as in Example 5.2. Another
instance is given in the following example.
Example 5.6. Let A, B ∈ L∞(R) be 1-periodic, f ∈ C∞,c(R). Assume that A ≧ c for some
c > 0. For n ∈ N and ν > 0 consider
(∂0A(n · m0) + B(n · m0)) un = f.
Recall that from Theorem 2.5, in order to compute the limit equation, we have to compute
expressions of the form
n→∞M−1
where Mn = A(n · m0) and Nn = B(n · m0), ℓ ∈ N.
7In what follows we adopt multiindex notation: For two operators A, B and k = (k1, k2) ∈ N2
Mhom,ℓ = τw- lim
n (cid:1)ℓ
0 NnM−1
n (cid:0)−∂−1
,
0 we denote
0, we denote its first and second component respectively by
(A, B)k := Ak1 Bk2 . If kj is a multiindex in N2
kj,1 and kj,2.
21
5 Time-dependent coefficients
In order to deduce G-convergence in the latter example we need the following theorem.
Theorem 5.7. Let A1, . . . , Ak ∈ L∞(R) be 1-periodic. Then for every ν > 0 we have
An := A1(n · m0)
Aj(y)dy ∈ L(cid:0)L2
0 Aj+1(n · m0)
−−−−→ (cid:0)∂−1
Yj=1Z 1
0 (cid:1)k−1
Yj=1
τw,n→∞
∂−1
k−1
0
k
ν(R)(cid:1) .
k
k
−∞
−∞
Proof. For n ∈ N and K, L j R bounded, measurable we compute
hχK ,AnχLiν =ZK
Ak(ntk)χL(tk)dtk ··· dt2e−2νt1 dt1
χL(tk)e−2νt1 dtk ··· dt1
χR>0(tj−1 − tj)
Yj=2
−∞ ···Z tk−1
A2(nt2)Z t2
A1(nt1)Z t1
Aj(ntj)
−∞
−∞ ···Z tk−1
−∞Z t2
=ZKZ t1
Yj=1
χK (t1)
Aj(ntj)
=ZR ···ZR
Yj=1
χL(tk)e−2νt1 dtk ··· dt1.
}
{z
Now, observe that (t1,··· , tk) 7→ χK(t1)(cid:16)Qk
j=2 χR>0(tj−1 − tj)(cid:17) χL(tk)e−2νt1 ∈ L1(Rk). More-
over, the mapping (t1,··· , tk) 7→ Qk
hχK,AnχLiν →*χK,(cid:0)∂−1
0 (cid:1)k−1
j=1 Aj(tj) is [0, 1]k-periodic. Thus, by Theorem 3.5, we
Aj(y)dyχL+
Yj=1Z 1
conclude that
k-times
as n → ∞ for all K, L j R bounded and measurable. A density argument concludes the
proof.
0
k
k
Example 5.8 (Example 5.6 continued). Thus, with the Theorems 5.7 and 2.5, we conclude
that (∂0A(n · m0) + B(n · m0)) G-converges to
∂0(cid:18)Z 1
A(y)
0
0
1
1
1
∞
∞
A(y)
+ ∂0
dy(cid:19)−1
Xk=1 −
Xℓ=1(cid:18)Z 1
dy(cid:19)−1
= ∂0(cid:18)Z 1
1 +
Xk=0 −
dy(cid:19)−1 ∞
= ∂0(cid:18)Z 1
dy(cid:19)−1 1 +
= ∂0(cid:18)Z 1
A(y)
A(y)
A(y)
1
1
0
0
0
0
0
0
1
∞
∞
A(y)
B(y)
A(y)
B(y)
A(y)
dy(cid:19)ℓ!k(cid:18)Z 1
dy(cid:19)−1Z 1
dy(cid:18)−∂−1
0 Z 1
dy(cid:19)ℓ!k
Xk=1 −
Xℓ=1(cid:18)−∂−1
0 Z 1
Xℓ=1(cid:18)−∂−1
0 Z 1
Xℓ=1(cid:18)−∂−1
0 Z 1
dy(cid:19)ℓ!k
dy(cid:19)ℓ!−1
B(y)
A(y)
B(y)
A(y)
∞
∞
0
0
0
1
A(y)
dy(cid:19)−1
22
0
0
0
1
1
1
1
0
0
A(y)
A(y)
A(y)
B(y)
A(y)
B(y)
A(y)
dy(cid:19)ℓ!−1
dy(cid:19)
dy(cid:19)−1Z 1
dy(cid:19)−1 ∞
Xℓ=0(cid:18)−∂−1
0 Z 1
dy(cid:19)−1(cid:18)1 + ∂−1
0 Z 1
dy(cid:19)−1
+(cid:18)Z 1
= ∂0(cid:18)Z 1
= ∂0(cid:18)Z 1
= ∂0(cid:18)Z 1
Remark 5.9. In [15], the authors consider an equation of the form (∂0 + an(m0)) un = f in
the space L2(R; L2(R)) with (an)n being a bounded sequence in L∞(R× R). Assuming weak-
∗-convergence of (an)n , the author shows weak convergence of (un)n. The limit equation is a
convolution equation involving the Young-measure associated to the sequence (an)n. Within
our reasoning, we cannot show that the whole sequence converges, unless any power of (an)n
converges in the weak-∗ topology of L∞. However, as we illustrated above (see e.g. Example
3.11) our approach has a wide range of applications, where the method involving Young-
measures might fail to work.
B(y)
A(y)
A(y)
dy.
0
0
6 Proof of the main theorems
We will finally prove our main theorems. The proof relies on elementary Hilbert space con-
cepts. We emphasize that the generality of the perspective hardly allows the introduction of
Young-measures, which have proven to be useful in particular cases (see the sections above for
a detailed discussion). Before we give a detailed account of the proofs of our main theorems,
we state the following auxilaury result, which we state without proof.
Lemma 6.1. Let H be a Hilbert space, T ∈ L(H). Assume that Re T ≧ c for some c > 0.
Then kT −1k ≦ 1
c and Re T −1 ≧ c
kT k2 .
Proof of Theorem 2.5. For f ∈ C∞,c(R; H) let un solve
(∂0Mn + Nn) un = f.
This yields
n
∞
∂−1
0 f
un = M−1
= M−1
= M−1
n (cid:1)−1
n (cid:0)1 + ∂−1
0 NnM−1
Xℓ=0(cid:0)−∂−1
n (cid:1)ℓ
0 NnM−1
Xℓ=1
n (cid:0)−∂−1
M−1
0 NnM−1
∂−1
0 f
n +
∞
n (cid:1)ℓ! ∂−1
0 f.
Hence, choosing an appropriate subsequence, we arrive at an expression of the form
u = Mhom,0 +
Mhom,ℓ! ∂−1
0 f.
∞
Xℓ=1
23
6 Proof of the main theorems
We remark here that due to the (standard) estimate kTk ≦ lim inf k→∞ kTkk for a sequence
(Tk)k of bounded linear operators in some Hilbert space converging to T , the seriesP∞
ℓ=1 Mhom,ℓ
converges with respect to the operator norm if ν is chosen large enough. Using the positive
definiteness of Mn for all n ∈ N and Lemma 6.1, we deduce that
ReM−1
n ≧
c
supn∈N kMnk2 .
By kM−1
n k ≦ 1
c , we conclude that
and
We arrive at
ReMhom,0 ≧
c
supn∈N kMnk2
ReM−1
hom,0 ≧
c3
supn∈N kMnk2 .
∞
∞
∞
= ∂0
hom,0u
M−1
hom,0Mhom,ℓ!−1
f = ∂0 1 +
Xℓ=1
M−1
Xk=0 −
hom,0Mhom,ℓ!k
Xℓ=1
M−1
M−1
= ∂0
hom,0Mhom,ℓ!k
Xk=1 −
Xℓ=1
M−1
1 +
Xk=1 −
Xℓ=1
= ∂0M−1
M−1
hom,0Mhom,ℓ!k
M−1
hom,0u + ∂0
hom,0u
∞
∞
∞
∞
hom,0u
M−1
hom,0u.
Proof of Theorem 2.7. We observe
n
n
(cid:18) ∂0Mn + N 00
n (cid:19)
n N 01
N 11
N 10
=(cid:18) 1 N 01
n (cid:0)N 11
n (cid:1)−1
Thus, with B :=(cid:16)∂0Mn + N 00
1
0
0
n − N 01
(cid:19)(cid:18) ∂0Mn + N 00
n (cid:0)N 11
n − N 01
n (cid:0)N 11
n (cid:1)−1 N 10
n (cid:17)−1
n (cid:1)−1 N 10
n
n (cid:19)(cid:18)
0
N 11
1
(cid:0)N 11
n (cid:1)−1 N 10
n
0
1(cid:19) .
n (cid:19)−1
n 1(cid:19)
0
1
N 10
n
n N 01
n
N 11
(cid:18) ∂0Mn + N 00
=(cid:18)
=(cid:18)
−(cid:0)N 11
−(cid:0)N 11
n (cid:1)−1 N 10
n (cid:1)−1 N 10
B
n B
24
n − N 01
0
(cid:16)∂0Mn + N 00
n (cid:1)−1(cid:19)(cid:18) 1 −N 01
(cid:0)N 11
0
0
n (cid:17)−1
n (cid:1)−1 N 10
n (cid:0)N 11
n (cid:1)−1
n (cid:0)N 11
(cid:19)
1
0
n (cid:1)−1
(cid:0)N 11
(cid:18) 1 −N 01
0
n (cid:1)−1
n (cid:0)N 11
1
(cid:19)
=
B
−(cid:0)N 11
n (cid:1)−1 N 10
n B
n (cid:1)−1 N 10
(cid:0)N 11
−BN 01
n BN 01
n (cid:1)−1
n (cid:0)N 11
n (cid:1)−1
n (cid:0)N 11
n (cid:1)−1! .
+(cid:0)N 11
With the Neumann series expression derived in the previous theorem, i.e.,
∞
Xℓ=1
n (cid:0)−∂−1
M−1
n (cid:1)ℓ
0 NnM−1
∂−1
0
n , we get that
n ∂−1
0 +
B = M−1
n (cid:0)N 11
n (cid:1)−1 N 10
n (cid:1)ℓ
0 NnM−1
n (cid:0)−∂−1
∂−1
0
0 NnM−1
∞
N 10
n
n − N 01
with Nn = N 00
n (cid:19)−1
n N 01
(cid:18) ∂0Mn + N 00
n
N 11
Xℓ=0
n (cid:0)−∂−1
n (cid:1)−1 N 10
−(cid:0)N 11
+(cid:18) 0
n (cid:1)−1(cid:19) .
0 (cid:0)N 11
n M−1
M−1
=
0
∂−1
0
−M−1
n (cid:0)−∂−1
n M−1
∂−1
0 N 01
0 NnM−1
0 NnM−1
n (cid:1)ℓ
n (cid:0)−∂−1
n (cid:1)ℓ
n (cid:1)−1
n (cid:0)N 11
n (cid:0)N 11
∂−1
0 N 01
n (cid:1)−1!
n (cid:1)−1 N 10
(cid:0)N 11
n (cid:1)ℓ
With Theorem 2.5, we deduce convergence of the top left corner in the latter matrix. Similarly,
we deduce convergence of the other expressions. Thus, for a suitable choice of subsequences,
we arrive at
∞
Xℓ=1(cid:18) Mhom,ℓ,00∂−1
Mhom,ℓ,10∂−1
0 Mhom,ℓ,11(cid:19) +(cid:18) Mhom,0,00∂−1
0 Mhom,ℓ,01
Mhom,0,10∂−1
0 Mhom,0,11 + Nhom,−1,11(cid:19) .
Mhom,0,01
0
We observe that
∞
Xℓ=1(cid:18) Mhom,ℓ,00∂−1
Mhom,ℓ,10∂−1
=(cid:18)M(1) +(cid:18) Mhom,0,00
0 Mhom,ℓ,11(cid:19) +(cid:18) Mhom,0,00∂−1
0 Mhom,ℓ,01
Mhom,0,10∂−1
0 1(cid:19) .
Nhom,−1,11(cid:19)(cid:19)(cid:18) ∂−1
0
0
0
0
0
0 Mhom,0,11 + Nhom,−1,11(cid:19)
Mhom,0,01
Moreover, note that the operator M(1) has norm arbitrarily small if ν was chosen large enough.
Hence, the operator
(cid:18)M(1) +(cid:18) Mhom,0,00
=(cid:18) Mhom,0,00
0
0
0
0
Nhom,−1,11(cid:19)(cid:19)
Nhom,−1,11(cid:19) (cid:18) Mhom,0,00
0
0
Nhom,−1,11(cid:19)−1
M(1) + 1!
is invertible. This gives
0
0
(cid:18)(cid:18)M(1) +(cid:18) Mhom,0,00
=(cid:18) ∂0 0
0 1(cid:19) ∞
Xℓ=0 −(cid:18) Mhom,0,00
Nhom,−1,11(cid:19)(cid:19)(cid:18) ∂−1
Nhom,−1,11(cid:19)−1
0
0
0
0
0 1(cid:19)(cid:19)−1
M(1)!ℓ M−1
0
hom,0,00
0
hom,−1,11!
N −1
25
References
=(cid:18) ∂0 0
+
hom,0,00
0
0 1(cid:19)
M−1
Xℓ=1 −(cid:18) Mhom,0,00
∞
0
0
hom,−1,11!
N −1
Nhom,−1,11(cid:19)−1
0
M(1)!ℓ M−1
0
hom,0,00
0
hom,−1,11!
.
N −1
References
[1] N. Antonić. Memory effects in homogenisation: Linear second-order equations. Arch.
Ration. Mech. Anal. 125(1): 1 -- 24, 1993.
[2] J.M. Ball. A Version of the fundamental Theorem for Young Measures PDE's and
Continuum Models of Phase Transition, Lecture Notes in Physics 344:207 -- 215, 1989.
[3] A. Bensoussan, J.L. Lions, and G. Papanicolaou. Asymptotic Analysis for Periodic Struc-
tures. North-Holland, Amsterdam, 1978.
[4] D. Cioranescu and P. Donato. An Introduction to Homogenization. Oxford University
Press, New York, 2010.
[5] J.B. Conway. A course in functional analysis.. 2nd edition. In Graduate Texts inMath-
ematics, 96, Vol. xvi. Springer-Verlag: New York etc. p. 399, 1990.
[6] N. Filonov. Spectral analysis of the selfadjoint operator curl in a region of finite measure.
St. Petersburg Math. J., 11(6):1085 -- 1095, 2000.
[7] H. Freymond and R. Picard. On electromagnetic waves in complex linear media in
nonsmooth domains Mathematical Methods in the Applied Sciences To appear. 2012.
[8] A. Kalauch, R. Picard, S. Siegmund, S. Trostorff, and M. Waurick. A Hilbert Space
Perspective on Ordinary Differential Equations with Memory Term. Technical Report,
TU Dresden, MATH-AN-015-2012, submitted. (http://arXiv:1204.2924)
[9] B. Jacob and H. J. Zwart. Linear Port-Hamiltonian systems on infinite-dimensional
spaces. Operator Theory: Advances and Applications 223. Basel: Birkhäuser, 2012.
[10] J.-S. Jiang, K.-H. Kuo, and C.-K. Lin On the Homogenization of Second Order Differ-
ential Equations Taiwanese Journal of Mathematics 9: 215 -- 236, 2005.
[11] J.-S. Jiang, K.-H. Kuo, and C.-K. Lin Homogenization of second order equation with
spatial dependent coefficient. Discrete Contin. Dyn. Syst. 12(2): 303 -- 313, 2005.
[12] G. Nguetseng. Homogenization structures and applications. I.
Z. Anal. Anwend.,
22(1):73 -- 107, 2003.
[13] G. Nguetseng. Homogenization structures and applications. II.
Z. Anal. Anwend.,
23(3):483 -- 508, 2004.
[14] M.-L. Mascarenhas A linear homogenization problem with time dependent coefficient.
Trans. Am. Math. Soc. 281: 179 -- 195, 1984.
[15] M. Petrini. Homogenization of linear and nonlinear ordinary differential equations with
time depending coefficients. Rend. Semin. Mat. Univ. Padova 99: 133 -- 159, 1998.
26
References
[16] R. Picard and D. McGhee. Partial Differential Equations: A unified Hilbert Space Ap-
proach, volume 55 of Expositions in Mathematics. DeGruyter, Berlin, 2011.
[17] R. Picard. On a selfadjoint realization of curl
in exterior domains.
Math. Z.,
229(2):319 -- 338, 1998.
[18] L.C. Piccinini
G-convergence for ordinary differential equations with Peano
phaenomenon. Rend. Sem. Mat. Univ. Padova 58: 65 -- 86, 1977.
[19] L.C. Piccinini Homogeneization problems for ordinary differential equations. Rend. Circ.
Mat. Palermo, II. Ser. 27: 95 -- 112, 1978.
[20] L. Tartar. Memory effects and homogenization. Archive for Rational Mechanics and
Analysis 111:121 -- 133, 1990.
[21] L. Tartar. Nonlocal effects induced by homogenization Partial Differential Equations
and the Calculus of Variations, Essays in Honor of Ennio De Giorgi 2:925 -- 938, 1989.
[22] M. Waurick.
Limiting
Processes
in
to Homogenization.
Hilbert
http://nbn-resolving.de/urn:nbn:de:bsz:14-qucosa-67442, 2011.
Space Approach
Evolutionary
- A
Dissertation, TU Dresden,
Equations
[23] M. Waurick. A Hilbert space approach to homogenization of linear ordinary differential
equations including delay and memory terms. Math. Methods Appl. Sci., 35(9): 1067 --
1077, 2012.
[24] M. Waurick. Homogenization of a class of linear partial differential equations. Asymptotic
Analysis, 2012. In press.
[25] M. Waurick. How far away is the harmonic mean from the homogenized matrix? Technical
report, TU Dresden, MATH-AN-07-2012, submitted. (arXiv:1204.3768v3).
[26] M. Waurick. Homogenization in fractional elasticity Technical report, TU Dresden,
MATH-AN-02-2013, submitted. (arxiv:1302.1731).
[27] M. Waurick and M. Kaliske. A Note on Homogenization of Ordinary Differential Equa-
tions with Delay Term. PAMM 11, 889-890, 2011.
[28] V.V. Zhikov, S.M. Kozlov, O.A. Oleinik, and K. T'en Ngoan. Averaging and G-
convergence of Differential Operators. Russian Mathematical Surveys, 34:69 -- 147, 1979.
27
|
1506.07346 | 3 | 1506 | 2016-08-30T07:39:00 | General coorbit space theory for quasi-Banach spaces and inhomogeneous function spaces with variable smoothness and integrability | [
"math.FA"
] | In this paper we propose a general coorbit space theory suitable to define coorbits of quasi-Banach spaces using an abstract continuous frame, indexed by a locally compact Hausdorff space, and an associated generalized voice transform. The proposed theory realizes a further step in the development of a universal abstract theory towards various function spaces and their atomic decompositions which has been initiated by Feichtinger and Gr{\"o}chenig in the late 1980ies. We combine the recent approaches in Rauhut, Ullrich and Rauhut to identify, in particular, various inhomogeneous (quasi-Banach) spaces of Besov-Lizorkin-Triebel type. To prove the potential of our new theory we apply it to spaces with variable smoothness and integrability which have attracted significant interest in the last 10 years. From the abstract discretization machinery we obtain atomic decompositions as well as wavelet frame isomorphisms for these spaces. | math.FA | math |
General coorbit space theory for quasi-Banach spaces
and inhomogeneous function spaces with variable smoothness
and integrability
Henning Kempka1,, Martin Schäfer2, Tino Ullrich3
1Faculty of Mathematics, Technical University Chemnitz
Reichenhainer Strasse 39, 09126 Chemnitz, Germany
2Mathematical Institute, Technical University of Berlin
Strasse des 17. Juni 136, 10623 Berlin, Germany
3Hausdorff Center for Mathematics & Institute for Numerical Simulation
Endenicher Allee 60, 53115 Bonn, Germany
October 2, 2018
Abstract
In this paper we propose a general coorbit space theory suitable to define coorbits of
quasi-Banach spaces using an abstract continuous frame, indexed by a locally compact
Hausdorff space, and an associated generalized voice transform. The proposed theory
realizes a further step in the development of a universal abstract theory towards various
function spaces and their atomic decompositions which has been initiated by Feichtinger
and Gröchenig in the late 1980ies. We combine the recent approaches in Rauhut, Ullrich [50]
and Rauhut [48] to identify, in particular, various inhomogeneous (quasi-Banach) spaces of
Besov-Lizorkin-Triebel type. To prove the potential of our new theory we apply it to spaces
with variable smoothness and integrability which have attracted significant interest in the
last 10 years. From the abstract discretization machinery we obtain atomic decompositions
as well as wavelet frame isomorphisms for these spaces.
Key Words: Coorbit space theory, continuous wavelet transform, Besov-Lizorkin-Triebel type
spaces, variable smoothness, variable integrability, 2-microlocal spaces, Peetre maximal func-
tion, atomic decomposition, wavelet bases
AMS Subject classification: 42B25, 42B35, 46E35, 46F05.
1 Introduction
The birth of coorbit theory dates back to the 1980ies, starting with a series of papers by
Feichtinger and Gröchenig [21, 31, 32]. The main intention was to characterize function spaces
Corresponding author, Email:
[email protected]
1
via an abstract transform, the so-called voice transform. In the original setup, this transform is
determined by an integrable irreducible representation of a locally compact group on a Hilbert
space H unifying e.g. the continuous wavelet transform, the short-time Fourier transform, and
the recent shearlet transform, to mention just a few. More recently, representations which are
not necessarily irreducible nor integrable have been considered [12]. They allow to treat, for
instance, Paley-Wiener spaces and spaces related to Shannon wavelets and Schrödingerlets.
Classical examples of coorbit spaces associated to the continuous wavelet transform on
the ax ` b-group are the homogeneous Besov-Lizorkin-Triebel spaces [55, 56, 57], identified
rigorously as coorbits in Ullrich [59]. What concerns further extensions of these spaces and
interpretations as coorbits we refer to Liang et al. [42, 43]. More general wavelet coorbit spaces
associated to a semidirect product G " Rd ¸ H, with a suitable subgroup H of GLpRdq as
dilation group, have been studied in [27, 28, 29] and could recently be identified with certain
decomposition spaces on the Fourier domain [30]. A specific example of this general setup is
the shearlet transform, where G is the shearlet group. The associated shearlet spaces have first
been studied in [9]. Other coorbit spaces, based on a voice transform different from the wavelet
transform, are e.g. modulation spaces [33, 19] and Bergman spaces [21].
Coorbit theory thus covers a great variety of different function spaces. The underlying group
structure however turns out to be a severe restriction for the theory since the identification of,
e.g., inhomogeneous spaces of the above type was long time not possible, however desirable. For
that reason the theory has evolved and several subsequent contributions have weakened among
others the assumption that the voice transform is supported on a locally compact group. For
instance, Dahlke, Steidl, and Teschke replaced it by a homogeneous space, i.e., a quotient of a
group with a subgroup, with the aim to treat functions on manifolds [10, 11, 8].
The starting point for the general coorbit space theory presented in this paper is the ap-
proach used by Fornasier and Rauhut [24], which was later revised and extended in [25] and
further expanded in [50]. There, the group structure is abandoned completely and the voice
transform is determined solely by an abstract continuous frame F " tϕxuxPX in H indexed
by a locally compact Hausdorff space X (not necessarily a group), i.e., X is equipped with a
Radon measure µ such that the map x ÞÑ ϕx is weakly measurable and that with constants
0 ă C1, C2 ă 8
C1}f H}2 ďżX
xf, ϕxy2dµpxq ď C2}f H}2
for all f P H .
(1.1)
(Note that weak measurability of x ÞÑ ϕx in H implies that the integral in (1.1) is well-defined.)
We combine the approach in [50] with ideas from [48] to define even coorbits
CopF , Y q :" tf : xf, ϕxy P Y u
of quasi-Banach spaces Y using the general voice transform associated to F. We thereby also
recall the relevant details of the existing theory, especially from [24, 50] and fix some earlier
inaccuracies. The developed theory yields noteworthy generalizations even for the Banach case,
e.g. some assumptions made in [24, 50] can be weakened, such as the uniform boundedness
of the analyzing frame F or some technical restrictions on the weights and the admissible
coverings. Most notably however, we can generalize the main results of the discretization
theory, which is possible since we take a different -- more direct -- route to establish them. It
turns out that the three essential Lemmas 2.36, 2.40, and 2.47 below constitute the technical
foundation for the proof of the general abstract discretization results in Theorems 2.48 and 2.50.
2
Putting these lemmas at the center of the exposition simplifies many arguments and allows for
a systematic approach towards new abstract discretization results. In fact, we obtain discrete
characterizations of coorbit spaces by "sampling" the function using a sub-sampled discrete
frame Fd " tϕxiuiPI on a suitable index set I. Of course, as usual in coorbit space theory,
there are several technical assumptions to check. However, a great advantage of the presented
discretization machinery is the fact that it provides a straight path towards discretization,
where matters essentially reduce to checking properties (associated to Y ) of the analyzing
frame F. This is in contrast to the usual approach where atomic decompositions and wavelet
characterizations, useful to study embeddings, s´numbers, interpolation properties etc., are
often developed from scratch for different related function spaces.
To prove the potential of the theory presented here, we apply it to identify spaces with
variable smoothness and integrability, so-called variable exponent spaces, as coorbits. Triebel-
Lizorkin spaces of this kind are defined via the quasi-norm
(1.2)
}f F w
ppq,qpqpRdq} "›››´ 8ÿj"0
wjpqpΦj f qpqqpq¯1{qpq
LppqpRdq››› ,
where the functions wj are weights and Φj are frequency filters corresponding to a dyadic
decomposition of the frequency plane. For the precise formulation see Definition 3.8 below.
The functions ppq, qpq represent certain integrability parameters, which may vary in the spa-
tial variable x of the space. The 2-microlocal weight sequence wjpq determines the variable
smoothness, see [39] for details. Function spaces with variable exponents are a fast developing
field thanks to its many applications in stochastics, fluid dynamics and image processing, see
[17] and [16] and references therein. The Lebesgue spaces LppqpRdq with variable integrability,
see Definition 3.1 below, were already used by Orlicz [46]. Recent contributions by Diening
[14] on the boundedness of the Hardy-Littlewood maximal operator on LppqpRdq make them
accessible for harmonic analysis issues.
Surprisingly, the spaces (1.2) can be handled within the generalized coorbit space theory
presented in this paper. In fact, due to unbounded left and right translation operators (within
the ax ` b-group) a coorbit characterization of homogeneous spaces of the above type already
seems to be rather impossible at first glance. However, we are able to identify them as coorbits
CopF , Y q of, what we call, Peetre-Wiener type spaces Y by using a suitable continuous frame
F " tϕxuxPX with the index set X " Rd rp0, 1q Y t8us. These spaces Y are solid quasi-
Banach function spaces (QBF) defined on X, see Section 4.1 below. Peetre-Wiener type spaces
can be seen as a mixture of the Peetre type spaces introduced in [59], and certain Wiener
amalgam spaces, see [22], [49]. They appear naturally when dealing with continuous local
mean characterizations, a strategy developed in [59] and [42]. In fact, we show in Subsection
3.3 below that with large enough a ą 0 the quantity
}f F w
ppq,qpqpRdq}3 " }wp, 8qxΦ
0 f yapqLppqpRdq}
t f yapqqpq dt
wp, tqxΦ
`›››´ż 1
0
represents an equivalent characterization for F w
ppq,qpqpRdq. Here
t¯1{qpq
(1.3)
LppqpRdq›››
xΦ
t f yapxq :"
sup
zPRd
t{2ďτ ď2t,τ ă1
pΦτ f qpx ` zq
p1 ` z{τ qa
3
denotes the corresponding maximal function, which is essentially a modification of the widely
used Peetre maximal function, see (3.7) below, and is used in the definition of the Peetre-Wiener
type spaces, see Definition 4.2. Now the representation (1.3) is actually the identification of
F w
ppq,qpqpRdq as a coorbit space of a Peetre-Wiener type space. Applying the abstract theory, in
particular Theorem 2.50, we obtain biorthogonal wavelet expansions [6] of the respective coorbit
spaces. We describe the application of the machinery for the rather simple (orthogonal) Meyer
wavelets, see Appendix A.2. Due to its generality, a straightforward modification of Theorem
2.50 leads to general (biorthogonal) wavelet expansions and other tight discrete wavelet frames.
Let us mention that the continuous local mean characterizations (1.3) of spaces with variable
exponents, see also Theorem 3.11, are new and interesting for their own sake. In fact, one has
to deal with additional difficulties since a version of the classical Fefferman-Stein maximal
inequality, a crucial tool in this respect, is in general not true in Lppqpℓqpqq if qpq is non-
constant.
Finally, the provided discretizations of such spaces are not entirely new. In [37] the au-
thor used a different technique in order to obtain discretizations with Meyer and Daubechies
wavelets. However, let us mention that the abstract Theorems 2.48, 2.50 below neither restrict
to orthonormal wavelets nor compactly supported atoms.
1.1 Outline
The paper is structured as follows. The abstract theory is established in Section 2. It generalizes
earlier contributions, especially [24, 50], and in particular now includes the quasi-Banach case.
In Section 3 we give a short introduction to variable exponent spaces, which will serve as our
demonstration object for a concrete application of the theory. We will utilize a new continuous
local means characterization in Section 4 to identify them as coorbits of a new scale of Peetre-
Wiener type spaces. The abstract theory then yields atomic decompositions as well as discrete
characterizations via wavelet frames. Some useful facts concerning the continuous and discrete
(orthogonal) wavelet transform are collected in the Appendix.
1.2 Notation
The symbols N, N0, Z, R, R`, and C denote the natural numbers, the natural numbers including
0, the integers, the real numbers, the non-negative real numbers, and the complex numbers.
For a real number t P R we put ptq` " maxtt, 0u and ptq´ " mintt, 0u. The conjugation of
z P C is denoted by z. Let us emphasize that Rd has the usual meaning and d P N0 is reserved
for its dimension. The symbol denotes the Euclidean norm on Rd and 1 the ℓ1-norm.
The space of all sequences with entries in some set M over some countable index set I is
denoted by M I and we write Λpiq for the i-th sequence element of a sequence Λ P M I.
For topological vector spaces Y and Z the class of linear continuous mappings from Y to Z
is denoted by LpY, Zq. The notation Φ : Y ãÑ Z indicates that Y is continuously embedded into
Z, i.e., Φ is an injective continuous linear map from Y into Z. If the embedding is canonical
we simply write Y ãÑ Z. If Y is equipped with a quasi-norm we use }f Y } for the quasi-norm
of f P Y . The operator quasi-norm of A P LpY, Zq is denoted by }AY Ñ Z}.
We use the notation a À b if there exists a constant c ą 0 (independent of the context
dependent relevant parameters) such that a ď c b.
If a À b and b À a we write a -- b.
Furthermore, we write Y -- Z for two quasi-normed spaces Y, Z which coincide as sets and
whose quasi-norms are equivalent.
4
2 General coorbit space theory
Let H be a separable Hilbert space and X a locally compact Hausdorff space endowed with a
positive Radon measure µ with supp µ " X. A family F " tϕxuxPX of vectors in H is called
a continuous frame (see [1]) if the assignment x ÞÑ ϕx is weakly measurable and if there exist
constants 0 ă C1, C2 ă 8 such that (1.1) is satisfied. Let us record an important property.
Lemma 2.1. Let F " tϕxuxPX be a continuous frame in H and N Ă X a set of measure zero.
Then tϕxuxPXzN is total in H.
Proof. Let us put X :" XzN. We have to show that V :" spantϕx : x P X u is dense in H.
Indeed, using the frame property of F, we can deduce for every f K V
}f H}2 "żX
xf, ϕxy2 dµpxq "żX
xf, ϕxy2 dµpxq " 0.
To avoid technicalities, we assume throughout this paper that X is σ-compact. We further
assume that the continuous frame is Parseval, i.e. C1 " C2 " 1, and note that -- apart from
minor changes -- the theory presented here is valid also for general tight frames where C1 " C2.
It is also possible to develop the theory in the setting of non-tight frames, where the associated
coorbit theory has been worked out in [24] -- at least to a significant extent.
For 0 ă p ă 8 we define the Lebesgue space LppXq :" LppX, µq as usual by
}F LppX, µq} :"´żX
F pxqp dµpxq¯1{p
ă 8.
A function F belongs to L8pXq :" L8pX, µq if and only if F is essentially bounded. The
corresponding sequence spaces ℓppIq are obtained by choosing X as a countable index set I,
equipped with the discrete topology and counting measure µ.
Associated to a continuous frame F is the voice transform VF : H Ñ L2pX, µq defined by
and its adjoint V
F : L2pX, µq Ñ H given in a weak sense by the integral
VF f pxq " xf, ϕxy ,
f P H, x P X,
V
F F "żX
F pyqϕy dµpyq .
(2.1)
Since we assume the frame F to be Parseval VF is an isometry and in particular injective.
F L2 Ñ H} " 1 and the associated frame operator SF :"
The adjoint V
F is surjective with }V
V
F VF is the identity. Hence we have
f "żX
VF f pyqϕy dµpyq
and VF f pxq "żX
VF f pyqxϕy, ϕxy dµpyq .
The second identity is the crucial reproducing formula RF pVF f q " VF f for f P H, where
RF px, yq " xϕy, ϕxy ,
x, y P X,
(2.2)
is an integral kernel (operator), referred to as the frame kernel associated to F. It acts as a self-
adjoint bounded operator RF " VF V
F : L2pXq Ñ L2pXq, which is an orthogonal projection
5
with RF pL2pXqq " VF pHq. The converse of the reproducing formula is also true, i.e.,
if
F P L2pXq satisfies RF pF q " F then there exists a unique element f P H such that VF f " F .
We remark that we use the same notation for the function RF : X X Ñ C given in
(2.2) and the associated operator RF : L2pXq Ñ L2pXq.
It is important to note that the
function RF is measurable. Indeed, utilizing an orthonormal basis pfnqnPN of H we can expand
RF px, yq "řnPNxϕy, fnyxfn, ϕxy as a point-wise limit of measurable functions.
The idea of coorbit theory is to measure "smoothness" of f via properties of the transform
VF f . Loosely speaking, the coorbit of a function space on X is its retract with respect to (a
suitably extended version of) the voice transform. The classical theory and its generalizations
have been developed for the case of certain Banach function spaces on X. In the classical setup,
where X is equipped with a group structure, the extension [48] deals with the quasi-Banach
case, and our aim is to extend the generalized theory from [24, 50] analogously.
2.1 Function spaces on X
We consider (quasi)-Banach function spaces, or shortly (Q)BF-spaces, which are linear spaces
of measurable functions on X, equipped with a quasi-norm under which they are complete.
Hereby, functions are identified when equal almost everywhere. Hence, when speaking of a
function one often actually refers to an equivalence class. In general, this inaccuracy of language
does not pose a problem. Only when it comes to point evaluations the precise meaning must
be made clear in the context.
Recall that a quasi-norm on a linear space Y generalizes the concept of a norm by replacing
the triangle inequality with the more general quasi-triangle inequality
}f ` g} ď CY p}f } ` }g}q,
f, g P Y,
with associated quasi-norm constant CY ě 1. Many aspects of the theory of normed spaces
carry over to the quasi-norm setting, e.g. boundedness and continuity coincide, all d-dimensional
quasi-norms are equivalent, etc.. An important exception is the Hahn-Banach theory concerned
with the dual spaces. Note that the (topological) dual Y 1 of a quasi-normed space Y , equipped
with the usual operator norm, is always a Banach space. Due to the possible non-convexity
of the quasi-norm however, it may not be sufficiently large for the Hahn-Banach theorems to
hold. In fact, Y 1 may even be trivial as the example of the Lp-spaces in the range 0 ă p ă 1
shows. This fact poses a serious problem for the theory.
An important tool for dealing with quasi-norms is the Aoki-Rolewicz theorem [3, 51], which
states that in every quasi-normed space Y there exists an equivalent r-norm -- in the sense of
an equal topology -- where an r-norm, 0 ă r ď 1, satisfies the r-triangle inequality
}f ` g}r ď }f }r ` }g}r,
f, g P Y,
and in particular is a quasi-norm with constant CY " 21{r´1. The exponent r " 1{plog2 CY ` 1q
of the equivalent r-norm is called the exponent of Y .
For a viable theory we need to further restrict the class of function spaces. A quasi-normed
function space Y on X is called solid, if the following condition is valid,
f µ-measurable, g P Y, f pxq ď gpxq a.e. ñ f P Y and }f Y } ď }gY }.
6
In a solid space Y we have the equality } f Y } " }f Y } for every f P Y . Moreover, there is a
useful criterion for a function f to belong to Y ,
f P Y ô f P Y and f µ-measurable.
A function space shall be called rich, if it contains the characteristic functions χK for all
compact subsets K Ă X. A rich solid quasi-normed function space on X then contains the
characteristic functions χU for all relatively compact, measurable subsets U Ă X.
We will subsequently develop coorbit theory mainly for rich solid QBF-spaces Y , that are
continuously embedded into Lloc
p pX, µq, 0 ă p ď 8,
consist of all functions F where }F χKLppXq} ă 8 for every compact subset K Ă X. The
case, where Y 6ãÑ Lloc
1 pXq, is shortly commented on at the end of Subsection 2.4
1 pXq. As usual, the spaces Lloc
p pXq :" Lloc
It is important to understand the relation between the quasi-norm convergence and the
pointwise convergence of a sequence of functions in Y . We have the following result.
Lemma 2.2. Let Y be a solid quasi-normed function space on X, and assume fn Ñ f in Y .
there is a subsequence pfnkqkPN, whose choice may depend on the particular x P X, such that
Then for arbitrary but fixed representing functions rfn,rf the following holds true. For a.e. x P X
rfnkpxq Ñ rf pxq as k Ñ 8.
Proof. Assume first that fn Ñ 0 in the quasi-norm of Y , which implies }fnY } Ñ 0. As
inf měn fm is a measurable function with inf měn fm ď fk for all k ě n we have inf měn fm P
Y with } inf měn fmY } ď }fkY } for all k ě n by solidity. It follows 0 ď } inf měn fmY } ď
pfn ´ f q Ñ 0 and by the previous argumentation for a.e. x P X there is a subsequence pfnkqkPN
inf měn }fmY } " 0, and hence inf měn rfmpxq " 0 for a.e. x P X. This implies that for
these x P X there is a subsequence pfnkqkPN such that rfnkpxq Ñ 0. Now let fn Ñ f . Then
such that rfnkpxq ´ rf pxq Ñ 0, whence rfnkpxq Ñ rf pxq.
Remark 2.3. A more thorough investigation of pointwise convergence in solid quasi-normed
function spaces is carried out in [54]. It turns out that Lemma 2.2 can be strengthened using
[54, Cor. 2.2.9] and the fact that X is σ-finite (see Step 1 in the proof of Lemma 2.14). In fact,
there is a subsequence pfnkqkPN, independent of x P X, with rfnkpxq Ñ rf pxq for a.e. x P X.
2.2 Associated sequence spaces
Let us take a look at sequence spaces associated with a function space Y on X. For this
we recall the notion of an admissible covering introduced in [24, 50]. We say that a covering
U " tUiuiPI of X is locally finite if every x P X possesses a neighborhood which intersects only
a finite number of the covering sets Ui.
Definition 2.4. A covering U " tUiuiPI of X is called admissible, if it is locally finite and if
it satisfies the following conditions:
(i) Each Ui is measurable, relatively compact and has non-void interior.
(ii) The intersection number σpUq :" supiPI 7tj : Ui X Uj ‰ Hu is finite.
A covering of a locally compact Hausdorff space is locally finite if and only if every compact
subset intersects only a finite number of the covering sets. Hence, every locally finite covering
of the σ-compact space X is countable. In particular, the following lemma holds true.
7
Lemma 2.5. Every admissible covering of the σ-compact space X has a countable index set.
Following [24, 50] we now define two types of sequence spaces associated to Y .
Definition 2.6. For a rich solid QBF-space Y on X and an admissible covering U " tUiuiPI
of X the sequence spaces Y 5 and Y 6 associated to Y and U are defined by
Y 5 " Y 5pUq :"!tλiuiPI
Y 6 " Y 6pUq :"!tλiuiPI
: }tλiuiPI Y 5} :"›››ÿiPI
: }tλiuiPI Y 6} :"›››ÿiPI
λiχUiY››› ă 8) ,
λiµpUiq´1χUiY››› ă 8) .
Note that due to Lemma 2.5 the index set I of these sequence spaces is necessarily countable.
Also observe that due to condition (i) of Definition 2.4 and supp µ " X we have µpUiq ą 0 for
every i P I, and in turn }χUiY } ą 0.
Viewing a sequence as a function on the index set I, equipped with the counting measure, we
subsequently use the terminology introduced above for function spaces. For better distinction,
we will speak of a quasi-Banach sequence space and use the abbreviation QBS-space.
Proposition 2.7. The sequence spaces Y 5pUq and Y 6pUq are rich solid QBS-spaces with the
same quasi-norm constant CY as Y .
Before we give the proof of this proposition let us establish some useful embedding results.
First observe that the mapping
I 6
5 : Y 5 Ñ Y 6, λi ÞÑ µpUiqλi
(2.3)
is an isometric isomorphism between Y 5 and Y 6, which allows to transfer statements from one
space to the other. Moreover, if inf iPI µpUiq ą 0 we have the embedding Y 5 ãÑ Y 6. Analogously,
supiPI µpUiq ă 8 implies Y 6 ãÑ Y 5. Consequently, Y 5 -- Y 6 if both conditions are fulfilled.
Let ν : I Ñ r0, 8q be a discrete weight and define }Λℓν
p} :" }Λνℓp} for 0 ă p ď 8 and
Λ P CI. The space ℓν
ppIq :" tΛ P CI : }Λℓν
p} ă 8u is a QBS-space with quasi-norm } ℓν
p}.
Lemma 2.8. Let 0 ă p ď 1 be the exponent of Y . We then have the continuous embeddings
ℓω5
p pIq ãÑ Y 5pUq ãÑ ℓω5
8 pIq
and ℓω6
p pIq ãÑ Y 6pUq ãÑ ℓω6
8 pIq
with weights defined by ω5piq :" }χUiY } and ω6piq :" µpUiq´1}χUiY } for i P I.
Proof. We have }tλiuiPI Y 5}p "››řiPI λiχUiY››p ÀřiPI λip}χUiY }p " }tλiuiPI ℓω5
p . If tλiuiPI P Y 5 we can estimate for every j P I
tλiuiPI P ℓω5
λjω5pjq " λj}χUj Y } " }λjχUj Y } ď›››ÿiPI
λiχUiY››› " }tλiuiPI Y 5}.
The embeddings for Y 6 follow with the isometry (2.3).
p }p for
(2.4)
The weights ω5 and ω6 also occur in the following result.
Corollary 2.9. For every j P I the evaluation Ej : tλiuiPI ÞÑ λj is a bounded functional on Y 5
and Y 6 with }EjY 5 Ñ C} ď pω5pjqq´1 and }EjY 6 Ñ C} ď pω6pjqq´1.
8
Proof. For Y 5 this follows directly from (2.4). The argument for Y 6 is similar.
Now we are ready to give the proof of Proposition 2.7.
Proof. [Proof of Proposition 2.7] We prove the completeness of Y 5. The result for Y 6 follows
then with the isometry (2.3). A Cauchy sequence pΛnqnPN in Y 5 is also a Cauchy sequence in
8 by Lemma 2.8. Let Λ be the limit in ℓω5
ℓω5
8 . We show that Λ P Y 5 and Λ " limnÑ8 Λn in the
which maps Λ P CI to a nonnegative measurable function on X.
For α P C and Λ, Λ1, Λ2 P CI we have ApαΛq " αApΛq and ApΛ1 ` Λ2q ď ApΛ1q ` ApΛ2q.
quasi-norm of Y 5. For this task let us introduce the auxiliary operator ApΛq :"řiPI ΛpiqχUi,
ApΛ1q ´ ApΛ2q ďÿiPIΛ1piq ´ Λ2piqχUi ďÿiPI
A sequence Λ P CI belongs to Y 5 if and only if ApΛq P Y , and we have the identity
We also have
Λ1piq ´ Λ2piqχUi " ApΛ1 ´ Λ2q.
(2.5)
}ΛY 5} " }ApΛqY }.
(2.6)
Since Λ is the limit of pΛnqnPN in ℓω5
the local finiteness of the sum in the definition of A it follows that
8 it holds limnÑ8 Λpiq´Λnpiq " 0 for all i P I. Considering
lim
nÑ8
ApΛ ´ Λnqpxq " 0
for all x P X.
(2.7)
The rest of the proof relies solely on Properties (2.5)-(2.7) of the operator A and the solidity
and completeness of Y . First we show ApΛq P Y which is equivalent to Λ P Y 5 according to
(2.6). The sequence pApΛnqqnPN is a Cauchy sequence in Y because with (2.5) we can estimate
}ApΛnq ´ ApΛmqY } ď }ApΛn ´ ΛmqY } " }Λn ´ ΛmY 5}. Furthermore, from (2.7) and (2.5) it
follows limnÑ8 ApΛnqpxq " ApΛqpxq for all x P X. Since Y is complete we can conclude with
Lemma 2.2 that ApΛnq Ñ ApΛq in Y and ApΛq P Y . Finally we show Λ " limnÑ8 Λn in Y 5.
The sequence pApΛn ´ ΛqqnPN is a Cauchy sequence in Y , because with (2.5) we get
}ApΛn ´ Λq ´ ApΛm ´ ΛqY } ď }ApΛn ´ ΛmqY } " }Λn ´ ΛmY 5}.
Using (2.7) and Lemma 2.2 we deduce ApΛn ´ Λq Ñ 0 in Y . In view of (2.6) this finishes the
proof.
We finally study sequence spaces where the finite sequences are a dense subset. Since Y 5
and Y 6 are isometrically isomorphic via the isometry I 6
5 is a bijection
on the sequences with finite support, these are dense in Y 5 if and only if they are dense in
Y 6. The next result occurs in [24, Thm. 5.2] in the context of Banach spaces. However, the
boundedness of the functions required there is not necessary.
5 from (2.3), and since I 6
Lemma 2.10. If the functions with compact support are dense in Y the finite sequences are
dense in Y 5pUq and Y 6pUq.
a function G P Y with compact support K such that }F ´ GY } ă ε. As the covering
U " tUiuiPI is locally finite, the index set J :" ti P I : Ui X K ‰ Hu is finite. Let Λ be the
Proof. Let Λ " tλiuiPI P Y 5 and fix ε ą 0. Then F :" řiPI λiχUi P Y and there exists
sequence which coincides with Λ on J and vanishes elsewhere. Then F :"řiPJ λiχUi P Y and
F ´ F ď F ´G. Using the solidity of Y we conclude }Λ´ ΛY 5} " }F ´ F Y } ď }F ´GY } ă ε.
For a countably infinite sequence Λ " tλiuiPI , a bijection σ : N Ñ I and n P N we define Λσ
n
as the sequence which coincides with Λ on σpt1, . . . , nuq and is zero elsewhere.
9
Lemma 2.11. Let U " tUiuiPI be an admissible covering and assume that there is a bijection
σ : N Ñ I. The finite sequences are dense in Y 5pUq if and only if for all Λ P Y 5pUq it holds
Λσ
n Ñ Λ in the quasi-norm of Y 5pUq for n Ñ 8.
Proof. Assume that the finite sequences are dense. For n P N we can then choose a finite
sequence Γn P Y 5 with }Γn ´ ΛY 5} ă 2´n. By solidity of Y we get for N ě 1 ` maxtk P
N ´ ΛY 5} ď }Γn ´ ΛY 5} ă
NΓnpσpkqq ‰ 0u, with the convention max H " 0, the estimate }Λσ
2´n. The other direction is trivial.
We end this paragraph with an illustration and examine the sequence spaces associated to
ppXq} :" }F νLppXq} ă 8, where ν is a
the weighted Lebesgue space Lν
weight and 0 ă p ď 8. In this special case we have a stronger statement than Lemma 2.8.
ppXq, defined by }F Lν
Proposition 2.12. Let U " tUiuiPI be an admissible covering of X, ν be a weight and 0 ă
p ď 8. Then for Y " Lν
by ν5
ν6
p
p pIq with weights given
ppXq we have Y 5pUq -- ℓ
ν5
p pIq and Y 6pUq -- ℓ
p
ppiq :" µpUiq´1ν5
ppiq :" }χUiLν
ppXq} and ν6
ppiq for i P I.
Proof. We give the proof for 0 ă p ă 8 and Y " Lν
ppXq. For tλiuiPI P CI we can estimate
p
}tλiuiPI Y 5}p "›››ÿiPI
λiχUiY›››
λipχUipyqpνpyqp dµpyq "ÿiPI
"żXÿiPI
λiχUipyqνpyq
λipżX
χUipyqpνpyqp dµpyq "ÿiPI
p
dµpyq
-- żXÿiPI
λipν5
ppiqp,
where we used that the intersection number σpUq is finite and the equivalence of the p-norm
and the 1-norm on CσpU q. Applying the isometry (2.3) yields the result for Y 6.
2.3 Voice transform extension
For the definition of the coorbit spaces, we need a sufficiently large reservoir for the voice
transform. Hence we extend it in this paragraph following [24]. For a weight ν : X Ñ r1, 8q
we introduce the space Hν
1pX, µqu . Since F is total in H by Lemma 2.1
it is easy to verify that }f Hν
1. Further, we define the
kernel algebra
1} constitutes a norm on Hν
1 :" tf P H : VF f P Lν
1} :" }VF f Lν
A1 :" tK : X X Ñ C : K is measurable and }KA1} ă 8u,
where
}KA1} :" max!ess sup
xPX şX Kpx, yqdµpyq , ess sup
yPX şX Kpx, yqdµpxq).
Associated to ν is a weight mν on X X given by
mνpx, yq " max! νpxq
νpyq
,
νpyq
νpxq) ,
The corresponding sub -- algebra Amν Ă A1 is defined as
x, y P X.
Amν :" tK : X X Ñ C : Kmν P A1u
(2.8)
and endowed with the norm }KAmν } :" }KmνA1}. Note that a kernel K P Amν operates
continuously on Lν
8 pXq} ď
}KAmν }. Technically, the theory rests upon (mapping) properties of certain kernel functions.
A first example of a typical result is given by the following lemma.
8 pXq with }KLν
1pXq}, }KL1{ν
1pXq and L1{ν
8 pXq Ñ L1{ν
1pXq Ñ Lν
10
F pxq "żX
Rpx, yqF pyq dµpyq ď C 2
BνpxqżX
F pyqνpyq dµpyq " C 2
Bνpxq}F Lν
1}.
Lemma 2.13. Assume that for a family G " tψxuxPX Ă H the Gramian kernel
GrG, Fspx, yq :" xϕy, ψxy
x, y P X,
(2.9)
is contained in Amν . Then ψx P Hν
1 with }ψxHν
Proof. We have }GrG, FsAmν } ěşX VF ψxpyq νpyq
1} ď }GrG, FsAmν }νpxq for a.e. x P X.
νpxq dµpyq " }ψxHν
for a.e. x P X.
νpxq
1 }
The theory in [24, 50] is developed under the global assumption that F is uniformly bounded,
i.e. }ϕx} ď CB for all x P X and some CB ą 0. This assumption can be weakened.
Lemma 2.14. Let ν ě 1 be a weight such that the analyzing frame F satisfies
(i) }ϕxH} ď CBνpxq for some constant CB ą 0 and all x P X,
(ii) RF P Amν .
(2.10)
1 is a Banach space and the canonical embedding Hν
Then Hν
Moreover, there is a subset X Ă X such that ϕx P Hν
The corresponding map Ψ : X Ñ Hν
1 , x ÞÑ ϕx is Bochner-measurable in Hν
1.
1 ãÑ H is continuous and dense.
1 for every x P X and µpXzX q " 0.
1, which converges to some F P Lν
Proof. A Cauchy sequence pfnqnPN Ă Hν
1 determines a Cauchy sequence pFn :" V fnqnPN in
Lν
1, the
equality Fn " RpFnq for n P N implies F " RpF q. Furthermore, because of }ϕxH} ď CBνpxq
it holds Rpx, yq ď C 2
1. Since the kernel R P Amν operates continuously on Lν
Bνpxqνpyq for all x, y P X and we can deduce
8 , and as L1{ν
1. Since }fn ´ f Hν
8 X Lν
1, which implies f P Hν
This shows F P L1{ν
yields f P H with V f " F P Lν
obtain fn Ñ f in Hν
we observe }hH}2 " }V hL2}2 ď }V hL1{ν
1 Ă L2 even F P L2. The reproducing formula on H
1} we
1. This proves the completeness. To prove the continuity of the embedding
8 } ď
νpxq }hH}) ď CB}hH}, where }ϕxH} ď CBνpxq was used, the continuity follows.
Due to Lemma 2.13, applied with G " F, there is a null-set N Ă X such that ϕx P Hν
x P X :" XzN. The density of Hν
in H, as stated by Lemma 2.1.
VF : Hν
1 for every
1 ãÑ H is thus a consequence of the totality of tϕxuxPX
It remains to prove the Bochner-measurability of Ψ. Since
1q is an isometric isomorphism, it suffices to confirm that
supxPX! }ϕxH}
1. Together with }V hL1{ν
1} " }Fn ´ F Lν
1} for h P Hν
1 Ñ VF pHν
8 }}hHν
rΨ :" VF Ψ : X Ñ Lν
1pXq, x ÞÑ VF ϕx
is Bochner-measurable in Lν
Step 1: Let us first construct an adequate partition of X. Since µ is a Radon measure, by
definition locally finite, all compact subsets of X have finite measure. As X is assumed to be
1pXq. The proof of this is divided into three steps.
finite measure. By subdividing each of these sets further into Ln,m :" tx P Ln : νpxq ď mu,
disjointifying these subdivided sets, and finally by renumbering the resulting countable family of
σ-compact, the measure µ is thus σ-finite. Hence X "ŤnPN Ln for certain subsets Ln Ă X of
sets, we obtain a sequence pKnqnPN of pairwise disjoint sets of finite measure with X "ŤnPN Kn
and such that νpxq ď Cn holds for all x P Kn and suitable constants Cn ą 0.
Step 2: We now show that for every n P N the function
1pXq, x ÞÑ VF ϕx χKn
rΨn : X Ñ Lν
11
1 for all ℓ P N. Such a basis exists since H is separable and Hν
1pXq. To this end, let pfℓqℓPN be an orthonormal basis of H with
1 is a dense subspace of H.
is Bochner-measurable in Lν
fℓ P Hν
Then we define the functions
Φℓ :" VF fℓ P Lν
1pXq
and Gn,ℓ :" VF fℓ χKn P Lν
1pXq.
Note that Φℓpxq " xϕx, fℓy is the ℓ-th expansion coefficient of ϕx with respect to pfℓqℓPN. Due
to the measurability of Φℓ P Lν
1pXq the function x ÞÑ ΦℓpxqGn,ℓ is clearly Bochner-measurable.
Since the pointwise limit of Bochner-measurable functions is again Bochner-measurable, Step 2
is finished if we can show that for every fixed x P X
This follows with Lebesgue's dominated convergence theorem: For every y P X we have
NÑ8
rΨnpxq " lim
VF´ Nÿℓ"1
in Lν
1pXq.
ΦℓpxqGn,ℓ
Nÿℓ"1
Φℓpxqfℓ¯pyq χKnpyq " VF ϕxpyq χKnpyq " rΨnpyq.
lim
NÑ8
ΦℓpxqGn,ℓpyq " lim
NÑ8
Nÿℓ"1
ℓ"1 Φℓpxqfℓ with convergence in H, and in general VF gN pxq Ñ VF gpxq
for fixed x P X if gN Ñ g in H. Finally, we estimate using }ϕxH} ď CBνpxq
Note here that ϕx "ř8
ΦℓpxqGn,ℓpyq ď´ Nÿℓ"1
Nÿℓ"1
Step 3: Similar to Step 2 the Bochner-measurability of rΨ is proved by showing for x P X
Since Kn is of finite measure this provides an integrable majorant (with respect to y).
ď }ϕxH}}ϕyH}χKnpyq ď CBνpyq}ϕxH}χKnpyq ď CBCn}ϕxH}χKnpyq.
Gn,ℓpyq2¯ 1
Φℓpxq2¯ 1
2´ Nÿℓ"1
2
in Lν
1pXq.
NÑ8
Nÿn"1rΨnpxq
rΨpxq " lim
Nÿn"1
rrΨpxqspyq " VF ϕxpyq " lim
NÑ8
χKnpyq VF ϕxpyq " lim
NÑ8
Nÿn"1
rrΨnpxqspyq.
The pointwise limit is obvious: For every y P X we clearly have
Using Lebesgue's dominated convergence theorem with majorant rΨpxq proves the claim.
Under the assumptions (2.10) we therefore have the chain of continuous embeddings
Hν
1
i
ãÑ H
i
ãÑ pHν
1qq,
1qq denotes the normed anti-dual of Hν
where pHν
1, which plays the role of the tempered distri-
butions in this abstract context. Moreover, there is a subset X Ă X with µpXzX q " 0 such
that ϕx P Hν
1 for x P X . Hence we may extend the transform VF : H Ñ L2pXq to pHν
1qq by
VF f pxq " xf, ϕxy ,
x P X , f P pHν
1qq,
(2.11)
where x, y denotes the duality product on pHν
product extends the scalar product of H.
1qq Hν
1. The anti-dual is used so that this
12
Lemma 2.15. Under the assumptions (2.10) the extension (2.11) is a well-defined continuous
mapping VF : pHν
1qq Ñ L1{ν
8 pXq.
1qq. The function VF f pxq " xf, ϕxy is well-defined for every x P X . It
Proof. Let f P pHν
determines a measurable function on X, in the sense of equivalence classes, due to the Bochner
measurability of x ÞÑ ϕx in Hν
1 proved in Lemma 2.14. Using Lemma 2.13 we can estimate
VF f pxq " xf, ϕxy ď }f pHν
1qq}}ϕxHν
1} ď }f pHν
1qq}}RF Amν }νpxq.
This shows VF f P L1{ν
8 pXq with }VF f L1{ν
8 } ď }f pHν
1qq}}RF Amν }.
Remark 2.16. The membership RF P Amν does not ensure F Ă Hν
1, wherefore the extended
voice transform (2.11) might not be defined at every point x P X. This detail has not been
accounted for in preceding papers, and fortunately it is negligible since functions on X are only
determined up to µ-equivalence classes. Therefore we -- as in [24, 50] -- will henceforth assume
F Ă Hν
1 to simplify the exposition.
We proceed to establish the injectivity of the extended voice transform. To this end, the
following characterization of the duality bracket x, ypHν
1 qqHν
1
will be useful.
Lemma 2.17. If F has properties (2.10), then for all f P pHν
1qq and g P Hν
1 it holds
xf, gypHν
1 qqHν
1
VF f pyqVF gpyq dµpyq ": xVF f, VF gy
L
1{ν
8 Lν
1
.
Proof. Let f P pHν
1qq and g P Hν
1. Then VF g P L2 X Lν
1 and we get
"żX
F VF gy "Bf,żX
xf, gy " xf, V
"żX
VF gpyqϕy dµpyqF
VF gpyqxf, ϕyy dµpyq " xVF f, VF gy
L
1{ν
8 Lν
1
.
For this equality, it is important that the duality product commutes with the integral. To
verify this, note that since G :" VF g P Lν
Bochner sense in Hν
Hν
1 the integral şX Gpyqϕy dµpyq also exists in the
1. Indeed, in view of Lemma 2.14 the integrand is Bochner-measurable in
1. Bochner-integrability follows then from the estimate
żX
Gpyq }ϕyHν
1} dµpyq ď }RF Amν }żX
Gpyqνpyq dµpyq " }RF Amν }}GLν
1},
where Lemma 2.13 was used. Moreover, the value of the Bochner integral h :"şX Gpyqϕy dµpyq
equals g since for every ζ P H
xg, ζy "żX
VF gpyq VF ζpyq dµpyq " xh, ζy.
Using Lemma 2.17 we can simplify the proof of [24, Lem. 3.2].
Lemma 2.18. Assume that the analyzing frame F has properties (2.10). Then the expression
}VF f L1{ν
8 } is an equivalent norm on pHν
1qq.
13
Proof. We already know from Lemma 2.15 that }VF f L1{ν
from below we argue with the help of Lemma 2.17
8 } À }f pHν
1qq}. For the estimate
}f pHν
1qq} " sup
}hHν
1 }"1
xf, hypHν
1 qqHν
1
" sup
}hHν
xVF f, VF hy
L
1{ν
8 Lν
1
ď
sup
xVF f, Hy
HPLν
1 ,}HLν
1 }ď1
L
1{ν
8 Lν
1
1 }"1
" }VF f L1{ν
8 }.
A direct consequence of this lemma is the injectivity of VF .
Corollary 2.19. The voice transform VF : pHν
1qq Ñ L1{ν
8 pXq is continuous and injective.
The injectivity of VF on pHν
1qq implies that F is total in Hν
1.
Corollary 2.20. Let N Ă X be a set of measure zero. Then tϕxuxPXzN is total in Hν
1.
Proof. If this is not the case, the closure C of spantϕx : x P XzN u in Hν
1 is a true subspace,
and the Hahn-Banach extension theorem yields f P pHν
1qq, f ‰ 0, with xf, ζy " 0 for all ζ P C.
Hence, VF f pxq " 0 for a.e. x P X and therefore f " 0 by injectivity of VF , which is true even
with respect to µ-equivalence classes in the image space. This is a contradiction.
The adjoint V
F : L1{ν
8 pXq Ñ pHν
1qq of the restriction VF : Hν
1 Ñ Lν
the adjoint of VF : H Ñ L2pXq due to the equality xF, VF ζy
1pXq naturally extends
" xF, VF ζyL2L2 in case
L
1{ν
8 Lν
1
1 and F P L1{ν
ζ P Hν
The relations
8 X L2, and it can also be represented by a weak integral of the form (2.1).
V
F VF f " f
and
VF V
F pF q " RpF q
(2.12)
F VF f, ζy " xVF f, VF ζy
remain valid for the extension, i.e., they hold for f P pHν
yields xV
xV
F F, ϕxy "şX F pyqxϕy, ϕxy dµpyq " RpF qpxq for all x P X.
" xf, ζy for all ζ P Hν
1{ν
8 Lν
L
1
formula extends to pHν
1qq, a result obtained differently in [24, Lemma 3.6].
An easy consequence of the relations (2.12) is the important fact that the reproducing
1qq and F P L1{ν
1. Further, we have VF V
8 . Indeed, Lemma 2.17
F F pxq "
Lemma 2.21. Let ν ě 1 be a weight on X and assume that the analyzing frame F satisfies
(2.10). Then VF f pxq " RpVF f qpxq for every f P pHν
8 pXq
satisfies F " RpF q then there is a unique f P pHν
1qq and x P X. Conversely, if F P L1{ν
1qq such that F " VF f .
Proof. According to (2.12) we have RpV f q " V V V f " V f for f P pHν
direction assume that F P L1{ν
has the property V V F " RpF q " F . It is unique since V is injective on pHν
8 satisfies F " RpF q. Then by (2.12) the element V F P pHν
1qq. For the opposite
1 qq
1qq.
Finally we state the correspondence between the weak*-convergence of a net pfiqiPI in pHν
1 qq
and the pointwise convergence of pVF fiqiPI (compare [24, Lem. 3.6]).
Lemma 2.22. Let pfiqiPI be a net in pHν
1qq.
(i) If pfiqiPI converges to some f P pHν
1qq in the weak*-topology of pHν
1qq, then pVF fiqiPI
converges pointwise to VF f everywhere.
14
(ii) If pVF fiqiPI converges pointwise a.e. to a function F : X Ñ C and if pfiqiPI is uniformly
1qq in the weak*-topology with
1qq, then pfiqiPI converges to some f P pHν
bounded in pHν
VF f " F a.e..
Proof. We give a proof for sequences pfnqnPN which extends straightforwardly to nets.
Part (i): The weak*-convergence implies xfn, ϕxy Ñ xf, ϕxy for n Ñ 8 and all x P X.
Part (ii): Let X Ă X denote the subset where the sequence pVF fnqnPN converges pointwise.
The space M " spantϕx : x P X u lies dense in Hν
1 by Corollary 2.20. We define a conjugate-
linear functional f on M by f phq :" limnÑ8xfn, hy for h P M. By assumption, there is C ą 0
so that }fnpHν
1} for all n P N
and shows that f is bounded on M with respect to } Hν
1}. Hence it can be uniquely extended
to some f P pHν
1} ă ε. We get
1 qq} ď C, which leads to xfn, hy ď }fnpHν
1} ď C}hHν
1qq}}hHν
1qq. For ε ą 0 and ζ P Hν
1 } }fn ´ f pHν
1 we choose h P M such that }h ´ ζHν
1qq} ` xfn ´ f, hy ď εpC ` }f pHν
xfn ´ f, ζy ď }ζ ´ hHν
1qq}q ` xfn ´ f, hy.
Letting n Ñ 8 it follows lim supnÑ8 xfn ´ f, ζy ď εpC ` }f pHν
1qq}q. This holds for all ε ą 0,
hence, limnÑ8 xfn ´ f, ζy " 0. This shows that fn Ñ f in the weak*-topology of pHν
1qq. As
a consequence VF f pxq " xf, ϕxy " limnÑ8xfn, ϕxy " limnÑ8 VF fnpxq " F pxq for all x P X .
A direct implication is the correspondence principle with respect to sums formulated below.
Corollary 2.23. IfřiPI fi converges unconditionally in the weak*-topology of pHν
series řiPI VF fipxq converges absolutely for all x P X. Conversely, if řiPI VF fipxq converges
absolutely for a.e. x P X and if the finite partial sums ofřiPI fi are uniformly bounded in pHν
thenřiPI fi converges unconditionally in the weak*-topology.
1qq then the
1 qq
2.4 Coorbit spaces
In this central part we introduce the notion of coorbit spaces, building upon the correspondence
between elements of pHν
1qq and functions on X as established by the transform VF . The idea is
to characterize f P pHν
1 qq by properties of the corresponding function VF f . For a viable theory
the analyzing frame F must fulfill certain suitability conditions with respect to Y .
Definition 2.24. Let ν ě 1 be a weight on X. We say that F has property F pν, Y q if it
satisfies condition (2.10) and if the following holds true,
(i) RF : Y Ñ Y acts continuously on Y ,
(ii) RF pY q ãÑ L1{ν
8 pXq.
Condition (2.10) ensures that the voice transform extends to pHν
and (ii) imply that RF F pxq "şX RF px, yqF pyq dµpyq is well-defined for a.e. x P X if F P Y .
In addition, also due to (i) and (ii), the operator RF : Y Ñ L1{ν
we have RpF q P L
8 pXq is continuous: For F P Y
1qq. Further, conditions (i)
1{ν
8 pXq and
}RpF qL1{ν
8 } À }RpF qY } ď }RY Ñ Y } }F Y }.
In view of Definition 2.24 it makes sense to introduce the following subalgebra of Amν from
(2.8)
BY,mν " tK : X X Ñ C : K P Amν and K is bounded from Y Ñ Y u ,
15
equipped with the quasi-norm }KBY,mν } :" maxt}KAmν }, }KY Ñ Y }u.
Now we are able to give the definition of the coorbit of a rich solid QBF-space Y .
Definition 2.25. Let Y be a rich solid QBF-space on X and assume that the analyzing frame
F " tϕxuxPX has property F pν, Y q for some weight ν : X Ñ r1, 8q. The coorbit of Y with
respect to F is defined by
Copν, F , Y q :" tf P pHν
1qq : VF f P Y u with quasi-norm }f Copν, F , Y q} :" }VF f Y } .
Since the coorbit is independent of the weight ν in the definition, as proved by the lemma
below, it is omitted in the notation and we simply write CopF , Y q :" Copν, F , Y q. Moreover, if
the analyzing frame F is fixed we may just write CopY q.
Lemma 2.26. The coorbit Copν, F , Y q does not depend on the particular weight ν chosen in the
definition in the following sense. If ν ě 1 is another weight such that F has property F pν, Y q
then we have Copν, F , Y q " Copν, F , Y q.
1qq
ãÑ pHω
Proof. If F has properties F pν, Y q and F pν, Y q it also has property F pω, Y q for ω " ν ` ν. We
show Copω, Y q " Copν, Y q. Since ω ě ν we have the continuous dense embedding Hω
1 ãÑ Hν
1
which implies pHν
1 qq and hence Copν, Y q Ă Copω, Y q. For the opposite inclusion let
f P Copω, Y q. Then f P pHω
8 by property F pν, Y q.
Since F " RpF q according to the reproducing formula on pHω
8 . The
1 qq with V f " F , which due
1 qq then yields f P pHν
inverse reproducing formula on pHν
1qq Ă pHω
to the injectivity of V is equal to f . This shows f P pHν
1 qq, and as V f P Y even f P Copν, Y q.
Finally note that the quasi-norms on Copω, Y q and Copν, Y q are equal. Analogously it follows
Copω, Y q " Copν, Y q.
1 qq and F :" V f P Y with RpF q P L1{ν
1 qq we thus have F P L1{ν
Remark 2.27. The claim of Lemma 2.26 has to be understood in the sense
f xϕx:xPXy : f P Copν, F , Y q( " f xϕx:xPXy : f P Copν, F , Y q(
since the two spaces are not strictly speaking equal. Further, the span xϕx : x P Xy is dense in
Hν
1, thus the notation Copν, F , Y q " Copν, F , Y q is justified.
1 and Hν
Regarding the applicability of the theory, it is important to decide whether a given analyzing
frame F " tϕxuxPX has property F pν, Y q. In the classical theory, where X is a group, the
frame is of the special form ϕx " πpxqg, where π is a group representation and g P H a suitable
vector. In this case properties of F break down to properties of the analyzing vector g, and
it suffices to check admissibility of g, see [21, 22, 32]. For the continuous wavelet transform
concrete conditions can be formulated in terms of smoothness, decay and vanishing moments,
generalized in [29] to wavelets over general dilation groups. In our general setup the algebras
Amν and BY,mν embody the concept of admissibility and for the (inhomogeneous) wavelet
transform utilized in Section 4 also concrete conditions can be deduced, see e.g. [50].
Concerning the independence of CopF , Y q on the reservoir pHν
1qq we state [50, Lem. 3.7],
whose proof carries over directly.
Lemma 2.28. Assume that the analyzing frame F satisfies F pν, Y q and let S be a topological
vector space such that F Ă S ãÑ Hν
1. In case F is total in S and the reproducing formula
VF f " RF pVF f q extends to all f P S q (the topological anti-dual of S) then
CopF , Y q " tf P S q : VF f P Y u .
16
We have the following result concerning the coincidence of the two spaces CopF , Y q and
CopG, Y q, where F and G are two different continuous frames.
Lemma 2.29. Assume that the frames G " tgxuxPX and F " tfxuxPX satisfy F pν, Y q. If the
Gramian kernels GrF , Gs and GrG, Fs defined in (2.9) are both contained in BY,mν , we have
CopF , Y q " CopG, Y q in the sense of equivalent quasi-norms.
Proof. This is a consequence of the relations VF " GrF , GsVG and VG " GrG, FsVF . In view
of Lemma 2.14 and Lemma 2.32 we have VGfx P Lν
8 pXq
for f P pHν
1pXq for a.e. x P X. Further, VGf P L1{ν
1qq and hence with Lemma 2.17
VF f pxq " xf, fxypHν
1 qqHν
1
" xVGf, VGfxy
L
1{ν
8 Lν
1
"żX
xf, gyyxgy, fxy dµpyq " GrF , GsVG f pxq.
This proves VF " GrF , GsVG , and by symmetry also VG " GrG, FsVF .
It is essential for the theory that the reproducing formula carries over to CopY q, which is
an immediate consequence of Lemma 2.21.
Lemma 2.30. A function F P Y is of the form V f for some f P CopY q if and only if F " RpF q.
The reproducing formula is the key to prove the main theorem of this section, which corre-
1 qq.
sponds to [24, Prop. 3.7]. We explicitly state the continuitiy of the embedding CopY q ãÑ pHν
Theorem 2.31.
(i) The space pCopY q, } CopY q}q is a quasi-Banach space with quasi-norm
constant CY , which is continuously embedded into pHν
1qq.
(ii) The map V : CopY q Ñ Y establishes an isometric isomorphism between CopY q and the
closed subspace RpY q of Y .
(iii) The map R : Y Ñ Y is a projection of Y onto RpY q " V pCopY qq.
Proof.
embedding CopY q ãÑ pHν
for f P CopY q, where Lemma 2.18 is used,
In general, we refer to the proof of [24, Prop. 3.7]. However, the continuity of the
1qq is not proved there. It is a consequence of the following estimate
}f pHν
1qq} -- }V f L1{ν
8 } ď }RY Ñ L1{ν
8 }}V f Y } " }RY Ñ L1{ν
8 }}f CopY q}.
Further, the proof of [24, Prop. 3.7] implicitly relies on the validity of R R " R on Y , which
a-priori is only clear for L2pXq. Therefore, we include a proof of this relation here. Let F P Y
and choose compact subsets pKnqnPN with X "ŤnPN Kn and Kn Ă Km for n ď m, which is
possible since X is σ-compact. Then we define the sets Un :" tx P Kn : F pxq ď nu, which
are relatively compact and thus of finite measure. As a consequence, Fn :" χUnF P L2pXq.
Moreover, Fn P Y since Fnpxq ď F pxq for every x P X. Since by assumption R : Y Ñ Y is
well-defined the assignment y ÞÑ Rpx, yqF pyq is integrable for a.e. x P X. As Fnpyq Ñ F pyq
pointwise, Lebesgue's dominated convergence theorem thus yields for these x P X
RFnpxq "żX
Rpx, yqFnpyq dµpyq ÑżX
Rpx, yqF pyq dµpyq " RF pxq.
Next, observe that the function Rpx, qmν px, q is integrable for a.e. x P X since R P Amν .
Further, due to RpY q ãÑ L1{ν
8 pXq the following estimate holds true for a.e. x, y P X
Rpx, yqRFnpyq ď CRpx, yqνpyq}FnY } ď CRpx, yqmν px, yqνpxq}F Y }.
17
Another application of Lebesgue's dominated convergence therefore yields for a.e. x P X
RpRFnqpxq "żX
Rpx, yqRFnpyq dµpyq ÑżX
Rpx, yqRF pyq dµpyq " RpRF qpxq.
Since Fn P L2pXq we have RFn " RpRFnq for every n P N. Altogether, we obtain
RpRF qpxq Ð RpRFnqpxq " RFnpxq Ñ RF pxq.
Let us finally provide some trivial examples, also given in [24, Cor. 3.8].
Lemma 2.32. If the analyzing frame F satisfies condition (2.10) for a weight ν ě 1, it has
properties F pν, L2q, F pν, L1{ν
8 q, F pν, Lν
1q, and it holds
pHν
1qq -- CopF , L1{ν
8 q,
Hν
1 " CopF , Lν
1q,
H " CopF , L2q.
Typically, the theory cannot be applied if the QBF-space Y is not embedded in Lloc
1 pXq,
since then the kernel conditions concerning operations on Y can usually not be fulfilled. Let
us close this paragraph with a short discussion of how to proceed in case Y 6ãÑ Lloc
1 pXq.
The case Y 6ãÑ Lloc
1 pXq
The main idea is to replace Y with a suitable subspace Z, which is embedded into Lloc
1 pXq and
fits into the existing theory. The basic observation behind this is that not all the information of
Y is used in the definition of the coorbit. In fact, the information about CopY q is fully contained
in the subspace RpY q, i.e., we have CopF , Y q " CopF , RpY qq. Thus, we can painlessly pass
over to a solid subspace Z of Y and regain the same coorbit if
RpY q ãÑ Z ãÑ Y.
This observation motivates the idea to substitute Y -- in case Y is not embedded into Lloc
1 pXq
itself -- by a suitable subspace Z of Y consisting of locally integrable functions, and then to
consider the coorbit of Z instead.
In the classical group setting [48] Wiener amalgams [18,
49] were used as suitable substitutes. Since Wiener amalgams rely on the underlying group
structure, they cannot be used in our general setup however. Instead, it is possible to resort
to the closely related decomposition spaces due to Feichtinger and Gröbner [20], which can be
viewed as discrete analoga of Wiener amalgams. This approach has been worked out in [53],
where the decomposition space DpY, Uq with local component L8 and global component Y is
used. It is defined as follows.
Definition 2.33 ([53]). The decomposition space DpY, Uq associated to a rich solid QBF-space
Y on X and an admissible covering U " tUiuiPI of X is defined by
DpY, Uq :"!f P Lloc
Note that the sumřiPI }f }L8pUiqχUi is locally finite and defines pointwise a function on X.
8 pXq : }f DpY, Uq} :"›››ÿiPI
}f L8pUiq}χUiY››› ă 8).
18
The space DpY, Uq is a subspace of Y , continuously embedded, and a rich solid QBF-space
with the same quasi-norm constant CY as Y . Moreover, it is contained in Lloc
1 pXq even if Y
8 pXq, where ω : X Ñ p0, 8q
itself is not.
defined by ωpxq :" maxi:xPUit}χUiY }´1u is a locally bounded weight. For a short proof,
let K Ă X be compact and tUiuiPJ the finite subfamily of sets in U intersecting K. Then
ωpxq ď maxiPJ t}χUiY }´1u for all x P K.
In fact, we have the embedding DpY, Uq ãÑ L1{ω
In the spirit of [48, Def. 4.1], we may therefore pass over to CopF , DpY, Uqq, the coorbit of
DpY, Uq. In general, one can only expect CopF , DpY, Uqq Ă tf P pHν
1qq : VF f P Y u and not
equality. In many applications however the equality can be proved by methods not available
in the abstract setting. Moreover, the choice DpY, Uq is consistent with the theory due to the
result below, which is analogous to a result obtained for Wiener amalgams [48, Thm. 6.1].
Theorem 2.34 ([53, Thm. 8.1]). Assume that Y is a rich solid QBF-space and that the ana-
lyzing frame F has property F pν, Y q. If U is an admissible covering of X such that the kernel
M
U rF , Fs (defined in (2.13) below) operates continuously on Y , then the frame F has
property F pν, DpY, Uqq and it holds
U " K
in the sense of equivalent quasi-norms.
CopF , DpY, Uqq -- CopF , Y q
Remark 2.35. The condition that the kernel M
instance in the important case when F has property Dpδ, ν, Y q (see Definition 2.43 below).
U operates continuously on Y is fulfilled for
In [53, Thm. 8.1] this theorem was formulated under the additional assumption that Y is
1 pXq. However, essential for the proof is only that the frame F
continuously embedded into Lloc
has property F pν, Y q, wherefore we chose to omit this assumption here.
2.5 Discretizations
A main feature of coorbit space theory is its general abstract discretization machinery. With
a coorbit characterization of a given function space at hand, the abstract framework (Theo-
rems 2.48 and 2.50 below) provides atomic decompositions of this space, i.e., a representation
of functions using "only" a countable number of atoms as building blocks.
Moreover, the function space can be characterized via an equivalent quasi-norm on an
associated sequence space.
The transition to sequence spaces bears many advantages, since those usually have a simpler,
more accessible structure than the original spaces. For example, the investigation of embedding
relations becomes much simpler by performing them on the associated sequence spaces.
In
addition, atomic decompositions naturally lend themselves to real world representations of the
considered functions: By truncation one obtains approximate expansions consisting only of a
finite number of atoms.
Our discretization results, Theorem 2.48 and Theorem 2.50, transfer the results from [50],
namely Theorem 3.11 and Theorem 3.14, to the general quasi-Banach setting. Applying a
different strategy for their proofs, however, we are able to strengthen these results significantly
even in the Banach space setting.
19
Preliminaries
Let us introduce the kernel functions KU rG, Fs and K
U rG, Fs, which are related by involution
and play a prominent role in the discretization theory. For a family G " tψxuxPX and an
admissible covering U " tUiuiPI they are defined by
KU rG, Fspx, yq :" sup
zPQy
xϕx, ψzy
and K
U rG, Fspx, yq :" KU rG, Fspy, xq
(2.13)
Ui for y P X. Their mapping properties are essential for two
where x, y P X and Qy :" Ťi : yPUi
central results, namely Lemmas 2.36 and 2.40, which together with Lemma 2.47 provide the
technical foundation for the proofs of Theorem 2.48 and Theorem 2.50.
We will subsequently use the symbol 9 9 for the operator quasi-norm } Y Ñ Y } on Y .
Lemma 2.36. Let Y be a rich solid QBF-space on X and let the analyzing frame F " tϕxuxPX
possess property F pν, Y q. Further, let G " tψxuxPX Ă Hν
1 be a family and U " tUiuiPI an
admissible covering such that K
U rG, Fs defines a bounded operator on Y . Then for
U :" K
sup
zPUi
›››ÿiPI
f P CopF , Y q the functionřiPI supzPUi VGf pzqχUi belongs to Y with the estimate
Note that the sumřiPI supzPUi VG f pzqχUi is locally finite and defined pointwise.
zPQxż GrG, Fspz, yqVF f pyq dµpyq
VGf pzqχUiY››› ď σpUq 9 K
Proof. Using VGf " GrG, FsVF f we can estimate for f P CopF , Y q and all x P X
VGf pzq " sup
zPQx
GrG, FsVF f pzq ď sup
U 9 }f CopF , Y q} .
sup
zPQx
GrG, Fspz, yqVF f pyq dµpyq
U rG, Fspx, yqVF f pyq dµpyq " K
U rG, FspVF f qpxq.
zPQx
ďż sup
"ż K
F pzq ďÿiPI
sup
zPQx
For functions F : X Ñ C we further have the estimate
sup
zPUi
F pzqχUi pxq ď σpUq sup
zPQx
F pzq,
(2.14)
where σpUq is the intersection number of U. Choosing F " VGf in (2.14), we can conclude
sup
zPUi
›››ÿiPI
VGf pzqχUiY››› ď σpUq 9 K
U 9 }VF f Y } " σpUq 9 K
U 9 }f CopF , Y q}.
We can immediately deduce an important result, which corresponds to [50, Lemma 3.12],
concerning the sampling of VGf .
Corollary 2.37. With the same assumptions as in the previous lemma let txiuiPI be a family
of points such that xi P Ui. Then tVGf pxiquiPI P Y 5pUq and it holds
}tVGf pxiquiPI Y 5} "›››ÿiPI
VGf pxiqχUiY››› ď σpUq 9 K
U 9 }f CopF , Y q}.
20
Let us turn to the synthesis side. Here the following lemma is a key result, which gener-
alizes [24, Lem. 5.10] and whose short direct proof is new and avoids technical difficulties. In
particular, it does not rely on [24, Lem. 5.4].
Lemma 2.38. Let Y be a rich solid QBF-space on X and let the analyzing frame F " tϕxuxPX
possess property F pν, Y q. Further, let G " tψxuxPX be a family in H and U " tUiuiPI an
admissible covering such that KU :" KU rG, Fs defines a bounded operator on Y . Then for
tλiuiPI P Y 6pUq and for points xi P Ui the series řiPI λiVF ψxipxq converges absolutely for a.e.
x P X defining a function in Y with
›››ÿiPI
λiVF ψxiY››› ď 9KU 9 }tλiuiPI Y 6}.
If the finite sequences are dense in Y 6pUq the series also converges unconditionally in the quasi-
norm of Y .
Proof. We have for every x P X the estimate
ÿiPI
λiVF ψxipxq ďÿiPI
µpUiq´1λiχUipyqKU px, yq dµpyq " KUÿiPI
"żXÿiPI
µpUiq´1λiżX
χUipyqKU px, yq dµpyq
µpUiq´1λiχUi¸ pxq,
where summation and integration can be interchanged due to monotone convergence. Since
tλiuiPI P Y 6 the sum řiPI µpUiq´1λiχUi defines pointwise a function in Y . By assumption
KU operates continuously on Y and hence also KU`řiPI µpUiq´1λiχUi P Y , which implies
KU`řiPI µpUiq´1λiχUi pxq ă 8 for a.e. x P X. It follows thatřiPI λiVF ψxipxq converges
absolutely at these points. As a consequence of the solidity of Y and the pointwise estimate
µpUiq´1λiχUi¸ P Y,
λiVF ψxi ď KUÿiPI
the measurable functionsřiPI λiVF ψxi andřiPI λiVF ψxi belong to Y with
λiVF ψxiY››› ď 9KU 9 }tλiuiPI Y 6}.
λiVF ψxi ďÿiPI
ÿiPI
λiVF ψxiY››› ď›››ÿiPI
›››ÿiPI
It remains to show thatřiPI λiVF ψxi converges unconditionally in Y to its pointwise limit, if
the finite sequences are dense in Y 6pUq. For this we fix an arbitrary bijection σ : N Ñ I and
obtain as in (2.16)
(2.15)
(2.16)
›››
8ÿm"n`1
λσpmqVF ψxσpmqY››› ď 9KU 9 }Λ ´ Λσ
nY 6},
(2.17)
where the sequence Λσ
of (2.17) tends to zero for n Ñ 8, which finishes the proof.
n is given as in Lemma 2.11. According to this lemma the right-hand side
Corollary 2.39. With the assumptions of the previous lemma G " tψxuxPX Ă CopF , Y q.
21
Proof. For every x P X there is an index i0 P I such that x P Ui0. Set xi0 :" x and choose
arbitrary points xi P Ui for i P Izti0u. Let δi0 denote the sequence, which has entry 1 at
position i0 and is zero elsewhere. Since Y is assumed to be rich δi0 P Y 6 and by the previous
lemma VF ψx "řiPI δi0
i VF ψxi P Y , whence ψx P CopF , Y q.
The correspondence principle allows to cast Lemma 2.38 in a different form, which corre-
sponds to [50, Lem. 3.11]. However, due to the different deduction the technical assumption
Y 6 ãÑ pL1{ν
8 q6 is not required any more.
Lemma 2.40. With the same assumptions as in Lemma 2.38 the series řiPI λiψxi converges
unconditionally in the weak*-topology of pHν
1qq to an element f P CopF , Y q with
and the estimate
VF f " VF´ÿiPI
}f CopF , Y q} "›››ÿiPI
λiVF ψxi
λiψxi¯ "ÿiPI
λiψxiCopF , Y q››› ď 9KU 9 }tλiuiPI Y 6}.
Moreover, if the finite sequences are dense in Y 6pUq the series also converges unconditionally
in the quasi-norm of CopF , Y q.
Proof.
If the subset J Ă I is finite we have VF´řiPJ λiψxi¯pxq " řiPJ λiVF ψxipxq for
all x P X. Moreover, we have proved in Lemma 2.38 that řiPI λiVF ψxi converges pointwise
it remains to verify that the sumsřiPJ λiψxi for finite subsets J Ă I are uniformly bounded in
absolutely a.e. to a function in Y . In order to apply the correspondence principle, Corollary 2.23,
1qq. With the continuous embedding CopY q ãÑ pHν
1 qq from Theorem 2.31 we can conclude
pHν
›››ÿiPJ
the sums are uniformly bounded in pHν
λiψxipHν
1qq››› À›››ÿiPJ
λiVF ψxiY›››
λiVF ψxiY››› ď›››ÿiPI
for every finite subset J Ă I, where we used that ψxi P CopY q for all i P I by Corollary 2.39.
λiψxiCopY q››› "›››ÿiPJ
We have shown in the proof of Lemma 2.38 that řiPI λiVF ψxi is a function in Y . Hence
convergence ofřiPI λiψxi to an element f P pHν
together with the previous lemma asserts that VF f "řiPI λiVF ψxi P Y .
It remains to show that řiPI λiψxi converges unconditionally in CopF , Y q, if the finite
so far -- the sum řiP I λiψxi converges in the weak*-topology to an element of CopY q and
VF´řiP I λiψxi¯ "řiP I λiVF ψxi . In view of (2.17) we conclude
sequences are dense in Y 6. For a subset I Ă I let Λ denote the sequence which coincides
with Λ on I and is trivial elsewhere. By solidity Λ P Y 6 and -- applying what we have proved
1qq and Corollary 2.23 implies the unconditional weak*-
1qq. Moreover, f P CopY q because Corollary 2.23
›››
8ÿm"n`1
λσpmqψxσpmqCopY q››› "›››
8ÿm"n`1
λσpmqVF ψxσpmqY››› Ñ 0 pn Ñ 8q,
for an arbitrary bijection σ : N Ñ I, which finishes the proof.
Atomic decompositions
Our first goal is to obtain atomic decompositions of the coorbit CopY q. Since CopY q is isomet-
rically isomorphic to the function space RpY q we initially focus on this space and recall from
22
Theorem 2.31 that for functions F P RpY q the reproducing formula holds, i.e.
F " RpF q "żX
F pyqRp, yq dµpyq,
F P RpY q.
This identity can be interpreted as a "continuous atomic decomposition" of F with atoms Rp, yq
indexed by y P X. The strategy is to discretize the integral, an approach which originates from
Feichtinger and Gröchenig [22] and was also used in subsequent papers e.g.
in [24, 50]. To
this end let U " tUiuiPI be an admissible covering of X and let Φ " tΦiuiPI be a U-PU, i.e. a
partition of unity subordinate to the covering U consisting of measurable functions Φi which
satisfy
(i) 0 ď Φipxq ď 1 for all x P X and all i P I,
(ii) supp Φi Ă Ui for all i P I,
(iii) řiPI Φipxq " 1 for all x P X.
We note that the construction of such a family Φ with respect to a locally finite covering is
standard, see e.g. [23, p. 127]. Using Φ the integral operator R can be written in the form
ΦipyqF pyqRpx, yq dµpyq.
RpF qpxq "ÿiPIżX
UΦF pxq :"ÿiPI
A formal discretization yields a discrete integral operator UΦ, called the discretization operator,
ciF pxiqRpx, xiq,
(2.18)
:" şX Φipyq dµpyq and the points txiuiPI are chosen such that xi P Ui. Here we
where ci
must give meaning to the point evaluations F pxiq since in general F P Y only determines
an equivalence class of functions where point evaluations are not well-defined. However, the
operator UΦ is only applied to elements F P RpY q and pointwise evaluation can be understood
in the sense
F pxiq " pRF qpxiq "żX
Rpxi, yqF pyq dµpyq .
Intuitively, UΦF approximates RpF q because the discretization resembles a Riemannian
sum of the integral. Hence we can hope to obtain an atomic decomposition from the relation
F " RpF q « UΦF "ÿiPI
ciF pxiqRp, xiq.
So far our considerations were just formal. To make the argument precise we have to impose
conditions on F so that UΦ is a well-defined operator. It turns out that here mapping prop-
erties of the kernels MU :" KU rF , Fs and M
U rF , Fs come into play. Recalling the
definition (2.13) of KU , K
U we have for x, y P X
U :" K
MU px, yq " sup
zPQy
xϕx, ϕzy and M
U px, yq " MU py, xq
(2.19)
with Qy " Ťi : yPUi
Ui for the covering U " tUiuiPI .
The lemma below provides definition (2.18) with a solid foundation.
23
Lemma 2.41. If MU and M
U given in (2.19) are bounded operators on Y the discretization
operator defined in (2.18) is a well-defined continuous operator UΦ : RpY q Ñ RpY q with op-
erator quasi-norm }UΦRpY q Ñ RpY q} ď σpUq 9 MU 9 9M
U 9. In general, the convergence
of the sum in (2.18) is pointwise absolutely a.e.. If the finite sequences are dense in Y 6 the
convergence is also in the quasi-norm of Y .
Proof. For F P RpY q Lemma 2.30 gives an element f P CopY q such that F pxq " V f pxq for
all x P X. Thus, using Corollary 2.37 with G " F, we can conclude tF pxiquiPI P Y 5 with
}tF pxiquiPI Y 5} ď σpUq 9 M
U 9 }F Y }. Since λi ÞÑ µpUiqλi is an isometry from Y 5 to Y 6
and since 0 ď ci ď µpUiq for all i P I it follows tciF pxiquiPI P Y 6pUq and }tciF pxiquiPI Y 6} ď
}tF pxiquiPI Y 5}. Therefore by Lemma 2.40 the sum řiPI ciF pxiqϕxi converges in the weak*-
topology to an element in CopY q and UΦF " V`řiPI ciF pxiqϕxi. As a consequence UΦF P
RpY q and again with Lemma 2.40
}UΦF Y } " }ÿiPI
ciF pxiqϕxiCopY q}
ď 9MU 9 }tF pxiquiPI Y 5} ď σpUq 9 MU 9 9M
U 9 }F Y }.
The operator UΦ is self-adjoint in a certain sense.
Lemma 2.42. Let U " tUiuiPI be an admissible covering and assume that the associated
maximal kernels MU and M
U of the analyzing frame F belong to Amν . Then UΦ is a well-
defined operator on RpL1{ν
1q and for every F P RpL1{ν
8 q and G P RpLν
8 q and RpLν
1q it holds
xUΦF, Gy
L
1{ν
8 Lν
1
" xF, UΦGy
L
1{ν
8 Lν
1
.
(2.20)
Proof. For F P RpL1{ν
of Lemma 2.41 for Y " L1{ν
by Lemma 2.38 and (2.15). Analogous statements hold for G P RpLν
8 q we have F pxq " xF, Rp, xqy
8 -- tciF pxiquiPI P pL1{ν
1{ν
8 Lν
L
1
8 q6. Therefore,řiPI ciF pxiqRp, xiq P L1{ν
1q. We conclude
8
and -- by arguments in the proof
xUΦF, Gy
L
1{ν
8 Lν
1
"ÿiPI
"ÿiPI
ciF pxiqxRp, xiq, Gy
L
1{ν
8 Lν
1
ciGpxiqxF, Rp, xiqy
L
1{ν
8 Lν
1
"ÿiPI
ciF pxiqGpxiq
" xF, UΦGy
L
1{ν
8 Lν
1
,
where Lebesgue's dominated convergence theorem was used.
Our next aim is to find suitable conditions on Φ and U such that the discretization operator
UΦ is invertible. The possible expansion
F " UΦU ´1
Φ F "ÿiPI
cipU ´1
Φ F qpxiqRp, xiq
then yields an atomic decomposition for F P RpY q.
Intuitively, for the invertibility of UΦ
the functions F P RpY q must be sufficiently "smooth", so that a discrete sampling is possible
without loss of information. Since RpY q is the isomorphic image of CopY q under the voice
transform, we have to ensure that the transforms VF f of elements f P CopY q are smooth
enough. An appropriate tool for the control of the smoothness are the oscillation kernels, a
24
concept originally due to Feichtinger and Gröchenig. We use the extended definition from [25],
utilizing a phase function Γ : X X Ñ S1 where S1 " tz P C : z " 1u, namely
oscU ,Γpx, yq :" sup
zPQy
RF px, yq ´ Γpy, zqRF px, zq
and
osc
U ,Γpx, yq :" oscU ,Γpy, xq
with x, y P X and Qy as in (2.13). The choice Γ " 1 yields the kernels used in [24, 50].
We can now formulate a condition on F which ensures invertibility of UΦ, but which is
weaker than the assumptions made in [24, 50] since we allow a larger class of coverings and
weights.
Definition 2.43. We say a tight continuous frame F " tϕxuxPX Ă H possesses property
Dpδ, ν, Y q for a weight ν ě 1 and some δ ą 0 if it has property F pν, Y q and if there exists an
admissible covering U and a phase function Γ : X X Ñ S1 so that
(i) RF , oscU ,Γ, osc
U ,Γ P BY,mν .
(ii) }oscU ,ΓBY,mν } ă δ and }osc
U ,ΓBY,mν } ă δ.
Remark 2.44. A frame F with property Dpδ, ν, Y q for a covering U and a phase function Γ
automatically possesses properties Dpδ, ν, L1{ν
1 q for the same covering U and the
same phase function Γ.
8 q and Dpδ, ν, Lν
Proof. Every K P Amν operates continuously on L
}KAmν } and }KLν
L1{ν
1{ν
8 and Lν
1} ď }KAmν }. Moreover, for Y " L1{ν
8 and the algebras BY,mν and Amν coincide with equal norms.
1 Ñ Lν
1{ν
8 Ñ L
1{ν
8 } ď
1 it holds RpY q ãÑ
1 with }KL
8 or Y " Lν
Note that for a measurable kernel function K : X X Ñ C the equality 9K9 " 9 K 9
does not hold in general. However, we have the following result.
Lemma 2.45. Let K, L : X X Ñ C be two measurable kernels and assume that K acts
continuously on Y . Then, if Lpx, yq ď Kpx, yq for almost all x, y P X, also L acts contin-
uously on Y with the estimate 9L9 ď 9 K 9. In particular, K acts continuously on Y with
9K9 ď 9 K 9.
Let us record an important consequence of the previous lemma.
Corollary 2.46. If the frame F has property Dpδ, ν, Y q the kernels RF , RF , oscU ,Γ, osc
MU , and M
U are continuous operators on Y .
U ,Γ,
Proof. For all x, y P X we have RF px, yq ď MU px, yq as well as the estimates
MU px, yq ď oscU ,Γpx, yq ` RF px, yq
and
oscU ,Γpx, yq ď MU px, yq ` RF px, yq.
The corresponding estimates for the involuted kernels also hold true. Hence Lemma 2.45 yields
the result.
The following lemma shows that UΦF approximates F P RpY q if the analyzing frame
possesses property Dpδ, ν, Y q for a suitably small δ ą 0. It corresponds to [24, Thm. 5.13] and
the proof is still valid in our setting -- with the triangle inequality replaced by the corresponding
quasi-triangle inequality.
25
Lemma 2.47. Suppose that the analyzing frame F possesses property Dpδ, ν, Y q for some δ ą 0
with associated covering U " tUiuiPI and phase function Γ. Then the discretization operator
UΦ for some U -PU Φ is a well-defined bounded operator UΦ : RpY q Ñ RpY q and it holds
}Id ´ UΦ RpY q Ñ RpY q} ď δp9R 9 ` 9 M
U 9qCY .
(2.21)
Proof. For F P RpY q there is f P CopY q with F " V f . By adapting the proof of Lemma 2.36,
U 9 }f CopF , Y q}.
The intersection number σpUq does not come into play here, since the inequality (2.14) can be
it can be shown that rH :" řiPI supzPUi V f pzqΦi P Y with }rHY } ď 9M
improved when using Φi instead of χUi. A solidity argument yields H :"řiPI F pxiqΦi P Y
and alsořiPI F pxiqΓp, xiqΦi P Y with respective quasi-norms dominated by }rHY }.
Let us introduce the auxiliary operator SΦ : RpY q Ñ RpY q, given pointwise for x P X by
Since F " RpF q we can estimate
F pxiqΓp, xiqΦipxq.
SΦF pxq :" RÿiPI
F pxiqΓp, xiqΦi¯Y››› ď 9R 9›››F ´ÿiPI
We further obtain for every x P X, because F pxq " RpF qpxq even pointwise,
F pxiqΓp, xiqΦiY›››.
}F ´ SΦF Y } "›››R´F ´ÿiPI
F pxiqΓpx, xiqΦipxq "ÿiPI`RpF qpxq ´ Γpx, xiqRpF qpxiqΦipxq
F pxq ´ÿiPI
ΦipxqżX
ďÿiPI
Rpy, xq ´ Γpx, xiqRpy, xiqF pyq dµpyq ď osc
U ,ΓpF qpxq.
We arrive at }F ´ SΦF Y } ď 9R 9 }osc
Let us now estimate the difference of UΦ and SΦ. First we see that for x P X
U ,ΓpF qY } ď 9R 9 9osc
U ,Γ 9 }F Y } ď δ 9 R 9 }F Y }.
SΦF pxq "żX
Rpx, yqÿiPI
F pxiqΓpy, xiqΦipyq dµpyq "ÿiPIżX
Rpx, yqF pxiqΓpy, xiqΦipyq dµpyq.
Here we used Lebesgue's dominated convergence theorem, which we use again to obtain
UΦF pxq ´ SΦF pxq "ÿiPIżX
F pxiqΦipyqoscU ,Γpx, yq dµpyq "żXÿiPI
ďÿiPIżX
where H "řiPI F pxiqΦi as above. We conclude
ΦipyqF pxiqpRpx, xiq ´ Γpy, xiqRpx, yqq dµpyq
F pxiqΦipyqoscU ,Γpx, yq dµpyq " oscU ,ΓpHqpxq,
}UΦF ´ SΦF Y } ď }oscU ,ΓpHqY } ď 9oscU ,Γ 9 }HY } ď δ 9 M
U 9 }F Y }.
Hence, altogether we have proved
}F ´ UΦF Y } ď CY p}F ´ SΦF Y } ` }SΦF ´ UΦF Y }q ď δCY }F Y }p9M
U 9 ` 9 R9q.
26
If the righthand side of (2.21) is less than one, UΦ : RpY q Ñ RpY q is boundedly invertible
n"0pId ´ UΦqn, which is still valid in the quasi-Banach
with the Neumann expansion U ´1
setting.
Finally, we are able to prove a cornerstone of the discretization theory, which generalizes
[24, Thm. 5.7] and [50, Thm. 3.11]. Note that the characterization via the sequence spaces is
a new result even in the Banach case and that we can drop many technical restrictions.
Φ "ř8
Theorem 2.48. Let Y be a rich solid QBF-space with quasi-norm constant CY and suppose that
the analyzing frame F " tϕxuxPX possesses property Dpδ, ν, Y q for the covering U " tUiuiPI
and a small enough δ ą 0 such that
Choosing arbitrary points xi P Ui, the sampled frame Fd :" tϕiuiPI :" tϕxiuiPI then possesses
δ`p1 ` CY q››RF BY,mν›› ` δCYCY ď 1.
1 X CopY q such that the following holds true:
(2.22)
a "dual family" xFd " tψiuiPI Ă Hν
(i) (Analysis) An element f P pHν
txf, ψiyuiPI P Y 6pUq) and we have the quasi-norm equivalences
1qq belongs to CopY q if and only if txf, ϕiyuiPI P Y 5pUq (or
}f CopY q} -- }txf, ϕiyuiPI Y 5pUq}
and
}f CopY q} -- }txf, ψiyuiPI Y 6pUq}.
}f CopY q} À }tλiuiPI Y 6pUq}. In general, the convergence of the sum is in the weak*-
topology induced by pHν
It is unconditional in the quasi-norm of CopY q, if the fi-
(ii) (Synthesis) For every sequence tλiuiPI P Y 6pUq it holds f " řiPI λiϕi P CopY q with
nite sequences are dense in Y 6. Similarly, f " řiPI λiψi P CopY q with }f CopY q} À
(iii) (Reconstruction) For all f P CopY q we have f "řiPI xf, ψiyϕi and f "řiPI xf, ϕiyψi.
}tλiuiPI Y 5pUq} in case tλiuiPI P Y 5pUq.
1qq.
Proof. According to Remark 2.44 the frame F has properties Dpδ, ν, Lν
1qq -- CopL1{ν
respect to the covering U, and by Lemma 2.32 it holds pHν
view of Theorem 2.31 the voice transform V : pHν
operator with isometric restrictions V : CopY q Ñ RpY q and V : Hν
1qq Ñ RpL1{ν
1q and Dpδ, ν, L1{ν
8 q and Hν
1 " CopLν
8 q with
1q. In
8 q is thus a boundedly invertible
Let us fix a U-PU Φ " tΦiuiPI and put ci :" şX Φipyq dµpyq. According to Lemma 2.42
the corresponding discretization operator UΦ is well-defined and bounded on RpL1{ν
8 q Ñ RpL1{ν
dition (2.22) on δ further implies that UΦ : RpL1{ν
consequence of Lemma 2.47. Indeed, using the estimates 9M
and 9RF 9 ď 9RF 9 together with the assumption 9osc
8 q. Con-
8 q is boundedly invertible as a
U 9 ď CY p9RF 9 ` 9 osc
U ,Γ9q
U ,Γ9 ă δ we can deduce
1 Ñ RpLν
1q.
δp9RF 9 ` 9 M
U 9qCY ď δpp1 ` CY q 9 RF 9 `CY 9 osc
U ,Γ9qCY
ă δpp1 ` CY q}RF BY,mν } ` CY δqCY ď 1.
Analogously, it follows that UΦ : RpLν
invertible.
1q Ñ RpLν
1q and UΦ : RpY q Ñ RpY q are boundedly
For the proof it is useful to note that the operator T :" V ´1U ´1
1 Ñ Hν
boundedly invertible isomorphism, whose restrictions T : Hν
Φ V : pHν
1qq is a
1 and T : CopY q Ñ CopY q
1qq Ñ pHν
27
are also boundedly invertible. Moreover, T is "self-adjoint". For this observe that relation (2.20)
also holds for the inverse U ´1
n"0pId ´ UΦqn. Consequently, for f P pHν
1qq and ζ P Hν
1
xf, T ζy " xf, V ´1U ´1
Φ V ζy " xV f, U ´1
Φ V ζy
L
1{ν
8 Lν
1
" xU ´1
Φ V f, V ζy
L
1{ν
8 Lν
1
" xT f, ζy.
Φ "ř8
It follows further that T is sequentially continuous with respect to the weak*-topology of pHν
1 qq.
To see this let fn Ñ f in the weak*-topology. Then xT fn, ζy " xfn, T ζy Ñ xf, T ζy " xT f, ζy for
every ζ P Hν
1 X CopY q.
Since T respects these subspaces we can define
1. By Lemma 2.39, Corollary 2.46 and Lemma 2.32 the atoms ϕxi lie in Hν
ψi :" ciT ϕi P Hν
1 X CopY q
and claim that xFd " tψiuiPI is the desired "dual" of Fd " tϕiuiPI .
After these preliminary considerations we now turn to the proof of the assertions.
If f P CopY q then txf, ϕiyuiPI " tV f pxiquiPI P Y 5 and }txf, ϕiyuiPI Y 5} À }f CopY q}
Step 1.
by Corollary 2.37. Furthermore, it holds T f P CopY q and Corollary 2.37 yields txf, ψiyuiPI "
tcixT f, ϕiyuiPI P Y 6 with the estimate }txf, ψiyuiPI Y 6} ď }txT f, ϕiyuiPI Y 5} À }T f CopY q} À
}f CopY q}.
are dense in Y 5 (or equivalently Y 6) the convergence is even in the quasi-norm of CopY q.
Step 2. If tλiuiPI P Y 6 then by Lemma 2.40 the sumřiPI λiϕi converges in the weak*-topology
to an element in CopY q with estimate }řiPI λiϕiCopY q} À }tλiuiPI Y 6}. If the finite sequences
tciλiuiPI P Y 6 and henceřiPI ciλiϕi converges in the weak*-topology to an element in CopY q.
A similar statement holds for the dual family tψiuiPI.
Since T is sequentially continuous it follows that
Indeed, for tλiuiPI P Y 5 we have
ÿiPI
λiψi "ÿiPI
ciλiT ϕi " TÿiPI
ciλiϕi¸ P CopY q
with weak*-convergence in the sums. The operator T is also continuous on CopY q, proving the
quasi-norm convergence if the finite sequences are dense. Moreover, we have the estimate
Step 3. In this step we prove the expansions in (iii). For f P pHν
1qq we have the identity
λiψiCopY q››› À›››ÿiPI
›››ÿiPI
Φ V f "ÿiPI
V f " UΦ`U ´1
ciλiϕiCopY q››› À }tciλiuiPI Y 6} ď }tλiuiPI Y 5}.
ci`U ´1
Φ V f pxiqRp, xiq "ÿiPI
xf, ψiyV ϕi
with pointwise absolute convergence a.e. in the sums. Since pHν
txf, ψiyuiPI belong to pL1{ν
8 q the coefficients
8 q6 according to Step 1. Hence, by Lemma 2.40 it holds V f "
V přiPI xf, ψiyϕiq with weak*-convergence of the sum. The injectivity of V finally yields
1qq -- CopL1{ν
(2.23)
xf, ψiyϕi.
f "ÿiPI
xT ´1f, ψiyT ϕi "ÿiPI
28
Using the sequential continuity of T with respect to the weak*-topology we can further deduce
f " T T ´1f "ÿiPI
xT ´1f, ciT ϕiyT ϕi "ÿiPI
xf, ϕiyψi.
(2.24)
In particular, these expansions are valid for f P CopY q with coefficients txf, ψiyuiPI P Y 6 and
txf, ϕiyuiPI P Y 5 by Step 1.
1qq and either txf, ϕiyuiPI P Y 5 or txf, ψiyuiPI P Y 6 we can conclude from the
Step 4. If f P pHν
expansions (2.23) and (2.24) together with Step 2 that f P CopY q. Moreover, }txf, ψiyuiPI Y 6}
and }txf, ϕiyuiPI Y 5} are equivalent quasi-norms on CopY q because using Steps 1 and 2
and
}f CopY q} "›››ÿiPI
}f CopY q} "›››ÿiPI
xf, ψiyϕiCopY q››› À }txf, ψiyuiPI Y 6} À }f CopY q}
xf, ϕiyψiCopY q››› À }txf, ϕiyuiPI Y 5} À }f CopY q}.
Remark 2.49. Properties (i)-(iii) in particular show that the discrete families Fd and xFd
both constitute atomic decompositions for CopY q, as well as quasi-Banach frames, compare e.g.
[50, 48].
Frame expansion
Now we come to another main discretization result, which allows to discretize the coorbit space
CopY q " CopF , Y q by samples of a frame G " tψxuxPX different from the analyzing frame F.
It is a generalization of [50, Thm. 3.14], whose original proof carries over to the quasi-Banach
setting based on Corollary 2.37 and Lemma 2.40. In contrast to Theorem 2.48, here we require
the additional property of the covering U " tUiuiPI that for some constant D ą 0
µpUiq ě D for all i P I.
(2.25)
Theorem 2.50. Let Y be a rich solid QBF-space on X and assume that the analyzing frame
F " tϕxuxPX has property F pν, Y q. For r P t1, . . . , nu let Gr " tψr
xuxPX be
families in H, and suppose that for some admissible covering U " tUiuiPI with the additional
U r Gr, Fs belong to BY,mν . Then, if
property (2.25) the kernels Kr :" KU rGr, Fs and K
every f P H has an expansion
xuxPX and Gr " t ψr
r :" K
f "
xf, ψr
xiyψr
xi
(2.26)
nÿr"1ÿiPI
r"1›››txf, ψr
with fixed points xi P Ui, this expansion extends to all f P CopY q " CopF , Y q. Furthermore,
xiyuiPI P Y 6pUq for each r P t1, . . . , nu, and
f P pH 1
ν qq belongs to CopY q if and only if txf, ψr
in this case we have }f CopY q} -- řn
in the quasi-norm of CopY q if the finite sequences are dense in Y 6pUq. In general, we have
weak*-convergence induced by pHν
1qq.
xiyuiPI Y 6pUq››› . The convergence in (2.26) is
Observe that the technical assumption Y 6 ãÑ pL1{v
8 q6 made in [50, Thm. 3.14] is not nec-
In fact,
r P Amν is a stronger condition than GrGr, Fs, Gr Gr, Fs P Amν and implies Gr, Gr Ă Hν
1.
In view of Lemma 2.13 it is further not necessary to require Gr, Gr Ă Hν
1.
essary.
Kr, K
3 Variable exponent spaces
In the remainder we give a demonstration of the theory. As an example we will show that
variable exponent spaces, which have caught some attention recently, fall into the framework
of coorbit theory and can be handled conveniently within the theory.
29
3.1 Spaces of variable integrability
The spaces of variable integrability LppqpRdq were first introduced by Orlicz [46] in 1931 as
a generalization of the Lebesgue spaces LppRdq. Before defining them let us introduce some
standard notation from [41]. For a measurable function p : Rd Ñ p0, 8s and a set Ω Ă Rd
we define the quantities p´
ppxq. Furthermore, we abbreviate
ppxq and p`
Ω " ess inf
xPΩ
Ω " ess sup
xPΩ
Rd and p` " p`
p´ " p´
p´ ą 0. Having an admissible exponent p P PpRdq we define the set Rd
and for every measurable function f : Rd Ñ C the modular
Rd and say that ppq belongs to the class of admissible exponents PpRdq if
8 " tx P Rd : ppxq " 8u
ppqpf q "żRdzRd
8
f pxqppxqdx ` ess sup
f pxq .
xPRd
8
Definition 3.1. The space LppqpRdq is the collection of all functions f such that there exists a
λ ą 0 with ppqpλf q ă 8. It is equipped with the Luxemburg quasi-norm
››› f LppqpRdq››› " inf"λ ą 0 : ppq f
λ ă 1* .
The spaces LppqpRdq share many properties with the constant exponent spaces LppRdq. Let
us mention a few; the proofs can be found in [41] and in [16]:
• If ppxq " p then LppqpRdq " LppRdq,
• if f pxq ě gpxq for a.e. x P Rd then ppqpf q ě ppqpgq and›› f LppqpRdq›› ě›› g LppqpRdq››,
• ppqpf q " 0 if and only if f " 0,
• for ppq ě 1 Hölder's inequality holds [41, Theorem 2.1]
żRd
f pxqgpxqdx ď 4››› f LppqpRdq›››››› g Lp1pqpRdq››› ,
where 1{ppq ` 1{p1pq " 1 pointwise.
There are also some properties of the usual constant exponent spaces which the LppqpRdq
spaces do not share. For example in general the LppqpRdq spaces are not translation invariant,
i.e. f P LppqpRdq does not automatically imply that f p ` hq belongs to LppqpRdq for h P Rd. As
a consequence also Young's convolution inequality is not valid (see again [41] for details).
The breakthrough for LppqpRdq spaces was made by Diening in [14] when he showed that
the Hardy-Littlewood maximal operator M is bounded on LppqpRdq under certain regularity
conditions on ppq. His result has been generalized in many cases (see [15],[45] and [7]) and it
turned out that logarithmic Hölder continuity classes are well adapted to the boundedness of
the maximal operator.
Definition 3.2. Let g P CpRdq. We say that g is locally log-Hölder continuous, abbreviated
g P C log
loc pRdq, if there exists clog ą 0 such that
clog
gpxq ´ gpyq ď
for all x, y P Rd.
logpe ` 1{x ´ yq
30
We say that g is globally log-Hölder continuous, abbreviated g P C logpRdq, if g is locally log-
Hölder continuous and there exists g8 P R such that
gpxq ´ g8 ď
clog
logpe ` xq
for all x P Rd.
With the help of the above logarithmic Hölder continuity the following result holds.
Lemma 3.3 ([15, Thm. 3.6]). Let p P PpRdq with 1 ă p´ ď p` ď 8. If 1
is bounded on LppqpRdq i.e., there exists c ą 0 such that for all f P LppqpRdq
p P C logpRdq, then M
›››Mf LppqpRdq››› ď c››› f LppqpRdq››› .
Since logarithmic Hölder continuous exponents play an essential role we introduce the class
P logpRdq of admissible exponents ppq with 1{p P C logpRdq and 0 ă p´ ď p` ď 8. As a
consequence of Lemma 3.3, for exponents p P P logpRdq the maximal operator M is bounded on
L ppq
t
pRdq for every 0 ă t ă p´.
3.2 2-microlocal function spaces with variable integrability
We proceed with spaces of Besov-Triebel-Lizorkin type featuring variable integrability and
smoothness. Spaces of the form F spq
ppq,qpRdq have been studied in [17, 2], where
s : Rd Ñ R with s P L8pRdq X C log
loc pRdq. A further generalization was pursued in [36, 37]
replacing the smoothness parameter spq by a more general weight function w. We make some
reasonable restrictions on w and use the class W α3
α1,α2 of admissible weights introduced in [36].
ppq,qpqpRdq and Bspq
Definition 3.4. For real numbers α3 ě 0 and α1 ď α2 a weight function w : X Ñ p0, 8q on the
index set X " Rd rp0, 1q Y t8us belongs to the class W α3
α1,α2 if and only if for x " px, tq P X,
wpx, sq
, s ě t
, s " 8 ,
,
,
t P p0, 1q
t " 8
for all y P Rd.
t ¯s1
t´s´1 ` x´x0
p1 ` x ´ x0qs1
,
,
t P p0, 1q
t " 8
.
t´α1 wpx, 8q ď wpx, tq ď t´α2 wpx, 8q
t¯α1
´ s
t¯α2
wpx, sq ď wpx, tq ď´ s
(W1)$'&'%
(W2) wpx, tq ď wpy, tq" p1 ` x ´ y{tqα3
ws,s1px, tq "$&%
p1 ` x ´ yqα3
Example 3.5. The main examples are weights of the form
where s, s1 P R. These weights are continuous versions of 2-microlocal weights, used to define
2-microlocal function spaces of Besov-Lizorkin-Triebel type, see [36, 38, 37].
By choosing s1 " 0 we get back to usual Besov-Lizorkin-Triebel spaces with smoothness s P R.
The special weights from this example are usually called 2-microlocal weights. Furthermore,
function spaces which are defined with admissible weights w P W α3
α1,α2 are usually called 2-
microlocal spaces. This term was coined by Bony [4] and Jaffard [35], who also introduced the
concept of 2-microlocal analysis to study local regularity of functions.
31
Remark 3.6. By the conditions on admissible weights w P W α3
estimates which will be useful later on:
α1,α2 we obtain the following
1. For s ď t we get from (W1)
t¯α2
´ s
2. For 0 ă c ă s{t we have from (W1) and (3.1)
t¯α1
wpx, sq ď wpx, tq ď´ s
t¯α2
ď maxt1, cα1´α2u´ s
wpx, tq
wpx, sq
wpx, sq.
.
.
(3.1)
(3.2)
(3.3)
3. For 0 ă c ă t{s we obtain similarly from (W1) and (3.1)
wpx, tq
wpx, sq
t¯α1
ď maxt1, cα1´α2u´ s
4. Consequently, we have for 0 ă c1 ă s{t ă c2 from (3.2) and (3.3)
wpx, tq -- wpx, sq
for all x P Rd.
5. Using (W2) and the inequalities (3.2) and (3.3) we can relate wpx, tq to wp0, 1{2q by
wp0, 1{2qt´α1 p1 ` xq´α3 À wpx, tq À wp0, 1{2qt´α2 p1 ` xqα3 .
A weight w P W α3
α1,α2 gives rise to a semi-discrete counterpart pwjqjPN0, corresponding to
an admissible weight sequence in the sense of [36, 38, 37], given by
wjpxq "" wpx, 2´j q
wpx, 8q
,
,
j P N ,
j " 0 .
(3.4)
loc pRdq or an admissible weight sequence stemming from w P W α3
In [37, Lemma 2.6] it was shown that it is equivalent to consider a smoothness function
s P L8pRdq X C log
α1,α2 if they are
connected by wjpxq " 2jspxq, see (3.4). But there exist weight sequences (Example 3.5 with
s1 ‰ 0) where it is not possible to find a smoothness function s : Rd Ñ R such that the above
relation holds.
Recently in [58] the concept of admissible weight sequences was extended to include more gen-
eral weights. We will not follow this generalization of admissible weights, but we remark that
by this definition we can have local Muckenhoupt weights as components in the sequence.
The spaces Bw
ppq,qpqpRdq are defined Fourier analytical as subspaces of the tem-
pered distributions S 1pRdq. As usual the Schwartz space SpRdq denotes the locally convex space
of rapidly decreasing infinitely differentiable functions on Rd. Its topology is generated by the
seminorms
ppq,qpRdq and F w
}ϕ}k,l " sup
xPRd
p1 ` xqk ÿβďl
Dβϕpxq
for every k, l P N0. Its topological dual, the space of tempered distributions on Rd, is denoted
by S 1pRdq. The Fourier transform and its inverse are defined on both SpRdq and S 1pRd) (see
32
Appendix A.1) and we denote them by f and f _. Finally, we introduce the subspace S0pRdq
of SpRdq by
0) .
S0pRdq :"!f P SpRdq : D ¯αpf p0q " 0 for every multi-index ¯α P Nd
ppq,qpRdq and F w
ppq,qpqpRdq relies on a dyadic decomposition of unity, see also
The definition of Bw
[55, 2.3.1].
Definition 3.7. Let ΠpRdq be the collection of all systems tϕj ujPN0 Ă SpRdq such that
(i) there is a function ϕ P SpRdq with ϕjpξq " ϕp2´j ξq , j P N ,
(ii) supp ϕ0 Ă tξ P Rd :
ξ ď 2u ,
supp ϕ Ă tξ P Rd : 1{2 ď ξ ď 2u ,
ϕjpξq " 1 for every ξ P Rd .
(iii)
8řj"0
Definition 3.8. Let tϕju8
associated weight sequence twjujPN0 defined as in (3.4).
j"0 P ΠpRdq and put pΦj " ϕj for j P N0. Let further w P W α3
α1,α2 with
(i) For p P PpRdq, q P p0, 8s, we define Bw
ppq,qpRdq "!f P S 1pRdq : }f Bw
}wjpqpΦj f qpqLppqpRdq}q¯1{q
.
ppq,qpRdq} ă 8) with
}f Bw
ppq,qpRdq} "´ 8ÿj"0
(ii) For p, q P PpRdq we define F w
ppq,qpqpRdq "!f P S 1pRdq : }f F w
wjpqpΦj f qpqqpq¯1{qpq
ppq,qpqpRdq} ă 8) with
LppqpRdq›››.
ppq,qpqpRdq} "›››´ 8ÿj"0
}f F w
Remark 3.9. It is also possible to consider Besov spaces Bw
ppq,qpqpRdq with variable index
qpq, which were introduced and studied in [2]. The definition of these spaces is very technical
since they require a new modular. Surprisingly it is much harder to work with Besov spaces
with variable indices ppq and qpq than to work with variable Triebel-Lizorkin spaces, in sharp
contrast to the constant exponent case. For example, Besov spaces with variable qpq are not
always normed spaces for mintppq, qpqu ě 1, even if ppq is a constant (see [40] for details).
So we restrict our studies on Besov spaces to the case were the index qpq remains a constant q
and we leave the fully variable case for further research.
Formally, the definition of F w
ppq,qpRdq depends on the chosen decomposition
of unity tϕj u8
j"0 P ΠpRdq. The following characterization by local means shows that under
certain regularity conditions on the indices ppq, qpq it is in fact independent, in the sense of
equivalent quasi-norms.
ppq,qpqpRdq and Bw
To get useful further characterizations of the spaces defined above we need a replacement
for the classical Fefferman-Stein maximal inequality since it does not hold in our case if qpq is
non-constant. We will use the following convolution inequality.
33
Lemma 3.10 (Theorem 3.2 in [17]). Let p, q P P logpRdq with 1 ă p´ ď p` ă 8 and 1 ă q´ ď
q` ă 8, then for m ą d there exists a constant c ą 0 such that
›››››› pην,m fνqνPN0 ℓqpq››› LppqpRdq››› ď c››››› pfνqνPN0 ℓqpq›› LppqpRdq››› ,
where ην,mpxq " 2νdp1 ` 2νxq´m.
3.3 Continuous local means characterization
For our purpose, it is more convenient to reformulate Definition 3.8 in terms of a continuous
characterization, where the discrete dilation parameter j P N0 is replaced by t ą 0 and the
sums become integrals over t. Characterizations of this type have some history and are usually
referred to as characterizations via (continuous) local means. For further references and some
historical facts we mainly refer to [57, 5, 52] and in particular to the recent contribution [59],
which provides a complete and self-contained reference.
The system tϕjujPN0 P ΠpRdq may be replaced by a more general one. Essential are func-
tions Φ0, Φ P SpRdq satisfying the so-called Tauberian conditions
on
on
pΦ0pξq ą 0
pΦpξq ą 0
DβpΦp0q " 0 for all
tξ ă 2εu ,
tε{2 ă ξ ă 2εu ,
β1 ď R .
(3.5)
(3.6)
for some ε ą 0, and -- for some R ` 1 P N0 -- the moment conditions
If R ` 1 " 0 the condition (3.6) is void. We will call the functions Φ0 and Φ kernels for local
means and use the notations Φk " 2kdΦp2kq, k P N, as well as Φt " DtΦ " t´dΦp{tq for t ą 0.
The associated Peetre maximal function
pΦ
t f qapxq " sup
yPRd
pΦt f qpx ` yq
p1 ` y{tqa
,
x P Rd , t ą 0 ,
(3.7)
was introduced in [47] for f P S 1pRdq and a ą 0. We also need the stronger version
xΦ
t f yapxq " sup
t
2 ďτ ď2t
τ ă1
pΦ
τ f qapxq
,
x P Rd , t ą 0 ,
(Convention: sup H " 0)
which we will refer to as Peetre-Wiener maximal function and which was utilized for the coorbit
characterization of the classical Besov-Lizorkin-Triebel-spaces in [53]. To adapt to the inhomo-
geneous setting we further put xΦ
0 f qa " ppΦ0q
0 f ya " pΦ
1 f qa.
Using these maximal functions we now state several different characterizations.
Theorem 3.11. Let w P W α3
α1,α2 and choose functions Φ0, Φ P SpRdq satisfying (3.5) and (3.6)
with R ` 1 ą α2. For x P Rd and t P p0, 1q define A1f px, tq :" pΦt f qpxq, A2f px, tq :"
pΦ
t f yapxq, a ą 0. Further, put A1f px, 8q :" pΦ0 f qpxq,
A2f px, 8q :" pΦ
t f qapxq, and A3f px, tq :" xΦ
0 f qapxq, and A3f px, 8q :" xΦ
0f yapxq.
34
and
}f Bw
ppq,qpRdq}4 " }wp, 8qpΦ
0
}wp, tqAif p, tqLppqpRdq}q dt
`´ż 1
`´ 8ÿj"1›››wjpqpΦ
t ¯1{q
2´j f qapqLppqpRdq›››
q¯1{q
0 f qapqLppqpRdq}
,
.
(i) If p P P logpRdq, 0 ă q ď 8, and a ą d
p´ ` α3 then
Bw
ppq,qpRdq " tf P S 1pRdq : }f Bw
ppq,qpRdq}i ă 8u
,
i " 1, 2, 3, 4,
where for i " 1, 2, 3
}f Bw
ppq,qpRdq}i " }wp, 8qAif p, 8qLppqpRdq}
Moreover, } Bw
ppq,qpRdq}i, i " 1, 2, 3, 4, are equivalent quasi-norms in Bw
ppq,qpRdq .
(ii) If p, q P P logpRdq with 0 ă q´ ď q` ă 8, 0 ă p´ ď p` ă 8, and a ą maxt d
p´ , d
q´ u ` α3
then
ppq,qpqpRdq " tf P S 1pRdq : }f F w
F w
ppq,qpqpRdq}i ă 8u
,
i " 1, 2, 3, 4,
where for i " 1, 2, 3
and
}f F w
ppq,qpqpRdq}4 " }wp, 8qpΦ
0 f qapqLppqpRdq}
}f F w
ppq,qpqpRdq}i " }wp, 8qAif p, 8qLppqpRdq}
wp, tqAif p, tqqpq dt
0
`›››´ż 1
`›››´ 8ÿj"1
t¯1{qpq
2´j f qapqqpq¯1{qpq
LppqpRdq››› ,
LppqpRdq›››.
wjpqpΦ
Moreover, } F w
ppq,qpqpRdq}i, i " 1, 2, 3, 4, are equivalent quasi-norms in F w
ppq,qpqpRdq .
Before we present a sketch of the proof recall an important convolution inequality from [36].
Lemma 3.12. Let 0 ă q ď 8, δ ą 0 and p, q P PpRdq. Let pgkqkPN0 be a sequence of non-
k"0 2´ℓ´kδgk for ℓ P N0. Then there
exist constants C1, C2 ě 0 such that
negative measurable functions on Rd and denote Gℓ "ř8
››tGℓuℓ ℓqpLppqq›› ď C1›› tgkuk ℓqpLppqq››
and
›› tGℓuℓ Lppqpℓqpqq›› ď C2›› tgkuk Lppqpℓqpqq›› .
Proof of Theorem 3.11. We only prove (ii) and comment afterwards briefly on the necessary
modifications for (i). The arguments are more or less the same as in the proofs of [59, Thm.
2.6] and [53, Thm. 9.6]. We remark that the equivalences } F w
ppq,qpqpRdq}4
and } Bw
ppq,qpqpRdq} -- } F w
ppq,qpRdq}4 are already known, see [36].
ppq,qpRdq} -- } Bw
Step 1. First, we prove a central estimate (3.9) between different start functions Φ and Ψ
incorporating the different types of Peetre maximal operators. The needed norm inequalities
in the theorem are consequences of this central estimate (3.9), and are subsequently deduced
in the following steps.
35
with supp λ0 Ă tζ P Rd : ζ ď 2εu and supp λ Ă tζ P Rd : ε{2 ď ζ ď 2εu such that
We use the special dyadic decomposition of unity given by η0ptq " 1 if t ď 4{3 and η0ptq " 0
k"1 ηk " 1 and
Let us put ϕ0 :"pΦ0 and ϕk :"pΦk for k P N. We can find a pair of functions λ0, λ P SpRdq
řkPN0 λkϕk " 1, where λk " λp2´kq for k P N. Let us shortly demonstrate how to do that.
if t ą 3{2. We put ηk :" η0p{2kq ´ η0p{2k´1q for k P N. Then clearly η0 `ř8
we obtainřkPN0 λkϕk " 1 by defining λk :" ηkp{εq{ϕk for k P N0 and λ :" λ1p2q.
The support of the function θ :" 1 ´řkPN λkϕk P C 8
0 pRdq is fully contained in M :" tx ď
3ε{2u. Due to the Tauberian conditions, ϕ0 is positive on M. Inverting ϕ0 on M and extending
appropriately outside, we can construct a function γ P C 8
0 pRdq, which coincides with 1{ϕ0 on
M. Since λ0ϕ0 " θ we thus have the factorization λ0 " γθ.
We now put λ0,upq :" γpqθpuq for u P r1, 2s, which gives
λ0,uϕ0 ` ÿkPN
λkpuqϕkpuq " 1.
We then define Ξ, Θ, Λ, Λ0,u, and Λk for k P N0, all elements of SpRdq, via inverse Fourier
transform of the functions γ, θ, λ, λ0,u, and λk, respectively. We get Λ0,u " Ξ Θu and it holds
g " Λ0,u Φ0 g `řkPN Λ2´ku Φ2´ku g for every g P S 1pRdq.
(3.6). Choosing g " Ψ2´ℓv f , where f P S 1pRdq, ℓ P N, and v P r1{2, 4s, we get
Let Ψ0, Ψ P SpRdq be another system which satisfies the Tauberian conditions (3.5) and
Ψ2´ℓv Λ2´ku Φ2´ku f ` Ψ2´ℓv Λ0,u Φ0 f.
(3.8)
Ψ2´ℓv f " ÿkPN
Defining Jℓ,k "şRd Ψ2´ℓv Λ2´kupzqp1 ` 2kz{uqa dz for k P N we have for y P Rd
Ψ2´ℓv Λ2´kupzqΦ2´ku f py ´ zq dz
pΨ2´ℓv Λ2´ku Φ2´ku f qpyq ďżRd
ď pΦ
2´kuf qapyqJℓ,k,
For k " 0 we get with Jℓ,0 "şRd Ψ2´ℓv Λ0,upzqp1 ` zqa dz
pΨ2´ℓv Λ0,u Φ0 f qpyq ďżRd
Ψ2´ℓv Λ0,upzqΦ0 f py ´ zq dz ď pΦ
0 f qapyqJℓ,0.
To estimate Jℓ,k the following identity for functions µ, ν P SpRdq is used,
pµu νvqpxq "
1
ud rµ νv{uspx{uq "
1
vd rµu{v νspx{vq,
valid for u, v ą 0 and x P Rd. In case ℓ ě k ą 0 we obtain
Jℓ,k "żRd
pΨ2k´ℓ v
u
Λqpzqp1 ` zqa dz À sup
zPRdpΨ2k´ℓ v
u
Λqpzqp1 ` zqa`d`1 À 2pk´ℓqpR`1q ,
where we used [52, Lemma 1] in the last step. In case 0 ă ℓ ă k we estimate similarly to obtain
Jℓ,k "żRd
pΨ Λ2´pk´ℓqu{vpzqqp1 ` 2k´ℓuz{vqa dz À 2pℓ´kqpL`1´aq,
36
where L can be chosen arbitrarily large since Λ P S0pRdq fulfills moment conditions for all
L P N0.
For ℓ ą k " 0 we estimate as follows, taking advantage of Ξ P SpRdq,
Jℓ,0 "żRd
À sup
pΨ2´ℓv Θuq Ξpzqp1 ` zqa dz
yPRdpΨ2´ℓv Θuqpyqp1 ` yqa`d`1żRdżRd
yPRdpΨ2´ℓv{u Θqpyqp1 ` yqa`d`1żRdżRd
À sup
Ξpz ´ yqp1 ` z ´ yqap1 ` yq´d´1 dzdy
p1 ` zq´d´1p1 ` yq´d´1 dzdy À 2´ℓpR`1q.
Using 1 ` tx ď maxt1, tup1 ` xq and 1 ` x ` y{t ď p1 ` y{tqp1 ` x{tq for t ą 0 and
x, y P Rd we further deduce for k P N
pΦ
2´kuf qapyq ď pΦ
À pΦ
2´kuf qapxqp1 ` 2kx ´ y{uqa
2´kuf qapxqp1 ` 2ℓx ´ y{vqa maxt1, 2pk´ℓqua.
and pΦ
0 f qapyq À pΦ
0 f qapxqp1 ` 2ℓx ´ y{vqa. Altogether, we arrive - for k ě 1 - at
pΨ2´ℓv Λ2´ku pΦ2´ku f qqpyq
p1 ` 2ℓx ´ y{vqa
sup
yPRd
À pΦ
2´kuf qapxq#2pk´ℓqpR`1q
2pℓ´kqpL`1´2aq
: ℓ ě k,
: ℓ ă k,
with an implicit constant independent of u P r1, 2s and v P r1{2, 4s. For k " 0 we obtain
pΨ2´ℓv Λ0,u Φ0 f qpyq
p1 ` 2ℓx ´ y{vqa
sup
yPRd
À pΦ
0 f qapxq2´ℓpR`1q.
We thus conclude from (3.8) that uniformly in t, u P r1, 2s
xΨ
2´ℓtf yapxq "
sup
t{2ďvď2t,vă1
pΨ
2´ℓvf qapxq
À pΦ
0 f qapxq2´ℓpR`1q ` ÿkPN
2´kuf qapxq#2pk´ℓqpR`1q
2pℓ´kqpL`1´2aq
pΦ
: ℓ ě k,
: ℓ ă k.
Writing wℓ,tpxq " wpx, 2´ℓtq for ℓ P N and w0,tpxq " wpx, 8q we have
wℓ,tpxq wk,upxq´1 À#2pℓ´kqα2
2pℓ´kqα1
ℓ ě k,
ℓ ă k,
as a consequence of pW 1q, (3.2), and (3.3). Multiplying both sides with wpx, 2´ℓtq we finally
derive with an implicit constant independent of t, u P r1, 2s
wpx, 2´ℓtqxΨ
2´ℓtf yapxq À wpx, 8qpΦ
0 f qapxq2´ℓpR`1´α2q
wpx, 2´kuqpΦ
2´kuf qapxq#2pk´ℓqpR`1´α2q
2pℓ´kqpL`1´2a`α1q
: ℓ ě k,
: ℓ ă k.
` ÿkPN
37
Choosing L ě 2a ´ α1 we have with 0 ă δ " mint1, R ` 1 ´ α2u the central estimate
xΨ
2´ℓtf yapxq À 2´ℓδ wpx, 8q
wpx, 2´ℓtq
pΦ
0 f qapxq ` ÿkPN
2´k´ℓδ wpx, 2´kuq
wpx, 2´ℓtq
pΦ
2´kuf qapxq.
(3.9)
ppq,qpqpRdq}1 -- }f F w
Step 2. We show }f F w
ppq,qpqpRdq}2,3 is obvious and it remains to verify }f F w
ppq,qpqpRdq}2,3. The direction }f F w
ppq,qpqpRdq}3 À }f F w
}f F w
We use (3.9) with Ψ " Φ. Choosing 0 ă δ ď δ we obtain for any r ą 0, using an embedding
argument if 0 ă r ď 1 and Hölder's inequality otherwise,
ppq,qpqpRdq}1 À
ppq,qpqpRdq}1.
xΦ
2´ℓtf yr
apxqwrpx, 2´ℓtq À 2´ℓδrwrpx, 8qpΦ
0 f qr
apxq ` ÿkPN
2´k´ℓδrwrpx, 2´kuqpΦ
2´kuf qr
apxq.
To estimate the sum on the right hand side we use (2.66) proved in Substep 1.3 of the proof of
[59, Thm. 2.6]. It states that for x P Rd, f P S 1pRdq, k P N, u P r1, 2q, r ą 0, and 0 ă a ď N for
some arbitrary but fixed N P N0
pΦ
2´kuf qapxqr ď CN ÿjPN0
2´jN r2pk`jqdżRd
pΦ2´pk`jqu f qpyqr
p1 ` 2kx ´ yqar dy ,
(3.10)
where the constant CN is independent of x, f, k, and u P r1, 2q, but may depend on r, a and
N. Taking into account pW 2q and (3.1), which give the relation wpx, 2´kuq À 2´jα1p1 ` 2kx ´
yqα3 wpy, 2´pj`kquq and p1 ` 2kzq´M ď 2jM p1 ` 2k`jzq´M , this leads to
xΦ
2´ℓtf yr
apxqwrpx, 2´ℓtq À 2´ℓδrwrpx, 8qpΦ
0 f qr
apxq
` ÿkPN
2´k´ℓδr ÿjPN0
2´jr N 2pk`jqdżRd
pΦ2´pk`jqu f qpyqwpy, 2´pj`kquqr
p1 ` 2k`jx ´ yqpa´α3qr
dy (3.11)
with N " N ´ a ` α1 ` α3 ą 0. Since x P Rd is fixed we can apply in t the Lqpxq{rpr1, 2q; dt
t q
norm with r ă mintp´, q´u. This changes only the constant and the left-hand side of (3.11).
The Lq´{rpr1, 2q; du
u q (quasi-)norm in the variable u only affects the right-hand side of (3.11).
With Minkowski's integral inequality we obtain
ż 2
1
xΦ
2´ℓtf yapxqwpx, 2´ℓtqqpxq dt
2´k´ℓδr ÿjPN0
À ÿkPN
À ÿkPN
2´k´ℓδr ÿjPN0
tr{qpxq
2´j´k Nr2jdżRd´ş2
2´j´k Nr«ηj,pa´α3qr ż 2
1
´ 2´ℓδrwrpx, 8qpΦ
0 f qr
apxq
1 pΦ2´j u f qpyqwpy, 2´j uqq´ du
p1 ` 2jx ´ yqpa´α3qr
u¯r{q´
dy
pΦ2´j u f qpqwp, 2´j uqq´ du
ur{q´ff pxq
with functions ην,mpxq " 2νdp1 ` 2νxq´m.
ă r ă mintp´, q´u, which is possible since a ą α3 `
Now we choose r ą 0 such that
mintp´,q´u , and N such that N ą 0. Applying the Lppq{rpℓqpq{rq norm with respect to x P Rd
a´α3
d
d
38
and ℓ P N and using Lemma 3.12 twice together with Lemma 3.10 (note pa ´ α3qr ą d) then
yields
›››››
1
ż 2
À›››››
À›››››
ż 2
1
xΦ
2´ℓtf yapqwp, 2´ℓtqqpq dt
«ηℓ,pa´α3qr ż 2
ż 2
1
1
pΦ2´ℓu f qpqwp, 2´ℓuqq´ du
pΦ2´ℓu f qpqwp, 2´ℓuqq´ du
t r{qpq
´ c››› wp, 8qpΦ
ur{q´ff pxq
Lppq{rpℓqpq{rq›››››
.
Lppq{rpℓqpq{rq›››››
ur{q´
ur{q´
pΦ2´ℓu f qpxqwpx, 2´ℓuqq´ du
r
0 f qapq LppqpRdq›››
Lppq{rpℓqpq{rq›››››
(3.12)
Finally, we use Hölder's inequality to estimate the integral in the last norm. We use 0 ă q´ ď
qpxq and get
1
ďż 2
ďż 2
1
pΦ2´ℓu f qpxqwpx, 2´ℓuqqpxq du
pΦ2´ℓu f qpxqwpx, 2´ℓuqqpxq du
ur{qpxqż 2
ur{qpxq
.
1
du
ur{q´
1
qpxq
q´ 1
Using this estimate we can reformulate (3.12) into
ż 1
0
›››››
xΦ
λf yapqwp, λqqpq dλ
À››› wp, 8qpΦ
0 f qapxqr À ÿkPN
pΦ
ż 1
0
Lppq›››››
λ1{qpq
0 f qapq LppqpRdq››› `›››››
2´kN r2kdżRd
pΦλ f qpqwp, λqqpq dλ
λ1{qpq
Lppq›››››
.
pΦ2´ku f qpyqr
p1 ` x ´ yqar dy `żRd
pΦ0 f qpyqr
p1 ` x ´ yqar dy.
The inhomogeneous term pΦ
ever, is analogous to the exposition before with (3.10) replaced by the inequality
0 f qapxq needs to be treated separately. The argumentation, how-
In the Besov space case we do not need the functions ην,m and one can work with the usual
maximal operator M together with Lemma 3.3, see [36] for details.
Step 3. In the third step we show }f F w
We immediately observe }f F w
Substep 3.1. To prove }f F w
ppq,qpqpRdq}2 -- }f F w
ppq,qpqpRdq}3.
ppq,qpqpRdq}4 we apply (3.9) with u " 1 and
Ψ " Φ . Since the inhomogeneous terms are identical, it suffices to estimate the homogeneous
part. Integration with respect to dt{t yields for ℓ P N
ppq,qpqpRdq}2 À }f F w
ppq,qpqpRdq}3 À }f F w
ppq,qpqpRdq}3 -- }f F w
ppq,qpqpRdq}4.
´ż 2
1
t¯ 1
2´ℓtf yapxqqpxq dt
wpx, 2´ℓtqxΦ
qpxq À 2´ℓδw0pxqpΦ
2´k´ℓδwkpxqpΦ
2´k f qapxq.
0 f qapxq ` ÿkPN
39
Let us denote the function on the right-hand side of the previous estimate by Gℓ. Applying the
vector-valued convolution inequality of Lemma 3.12 then proves
›››´ 8ÿℓ"1ż 2
1
wp, 2´ℓtqxΦ
2´ℓtf yapqqpq dt
À }w0pΦ
0 f qaLppq} ` }twkpΦ
t¯1{qpqLppq››› À }tGℓuℓPNLppqpℓqpqq}
2´k f qaukPNLppqpℓqpqq} " }f F w
ppq,qpq}4.
Substep 3.2: Let us finish by proving }f F w
ppq,qpqpRdq}2. Again it suffices
to estimate the homogeneous part. For this we let t " 1 and Ψ " Φ in (3.9). If qpxq ě 1 we
can use Minkowski's inequality to deduce
ppq,qpqpRdq}4 À }f F w
wℓpxqpΦ
2´ℓ f qapxq À 2´ℓδwpx, 8qpΦ
0 f qapxq
` ÿkPN
2´k´ℓδ´ż 2
1
wk,upxqpΦ
2´kuf qapxqqpxq du
u¯1{qpxq
.
řkPN0
Applying the ℓqpxq-norm on both sides, Young's convolution inequality then yields
ÿℓPN
wℓpxqqpxqpΦ
2´ℓf qapxqqpxq À pwpx, 8qpΦ
0 f qapxqqqpxq
` ÿkPNż 2
1
wk,upxqpΦ
2´kuf qapxqqpxq du
u
(3.13)
.
If qpxq ă 1 we use the qpxq-triangle inequality
´wℓpxqpΦ
2´ℓ f qapxq¯qpxq
À 2´ℓδqpxqpwpx, 8qpΦ
0 f qapxqqqpxq
` ÿkPN
2´k´ℓqpxqδż 2
1
wk,upxqpΦ
2´kuf qapxqqpxq du
.
u
Now we take on both sides the ℓ1-norm with respect to the index ℓ P N and take into account
2´kqpxqδ ď C. We thus arrive at the same estimate (3.13). Taking the Lppq-quasi-norm
of (3.13) finishes the proof of Substep 3.2 and hence Step 3.
Step 4: Relation (3.9) also immediately allows to change to a different system Ψ0, Ψ, however
in the discrete setting the change of systems has already been shown in [36].
Remark 3.13. The previous theorem ensures in particular the independence of Besov-Lizorkin-
Triebel type spaces with variable exponents from the chosen resolution of unity if p, q P P logpRdq
with p` ă 8, q` ă 8 in the F -case and p P P logpRdq, q P p0, 8s in the B-case.
4 Variable exponent spaces as coorbits
In order to treat the spaces Bw
ppq,qpqpRdq as coorbits we utilize an inhomogeneous
version of the continuous wavelet transform, which uses high scale wavelets together with a base
scale for the analysis. The corresponding index set is X " Rd rp0, 1q Y t8us, where 8 denotes
an isolated point, equipped with the Radon measure µ defined by
ppq,qpRdq and F w
żX
F pxqdµpxq "żRdż 1
0
F px, sq
40
ds
sd`1
dx `żRd
F px, 8qdx .
The wavelet transform is then given by VF f pxq " xf, ϕxy, x P X, for a continuous frame
F " tϕxuxPX on H " L2pRdq of the form
ϕpx,8q " TxΦ0 " Φ0p ´ xq
and ϕpx,tq " TxDL2
t Φ " t´d{2Φpp ´ xq{tq ,
(4.1)
with suitable functions Φ0, Φ P L2pRdq. Such a frame F " FpΦ0, Φq will in our context be
referred to as a continuous wavelet frame in L2pRdq.
Definition 4.1. A continuous wavelet frame F " FpΦ0, Φq is admissible if Φ0 P SpRdq and
Φ P S0pRdq are chosen such that they satisfy the Tauberian conditions (3.5), (3.6) and the
condition
" C
for a.e. ξ P Rd.
An admissible wavelet frame FpΦ0, Φq represents a tight continuous frame in the sense of
(1.1). To see this, apply Fubini's and Plancherel's theorem to get
0
t
pΦ0pξq2 `ż 1
pΦptξq2 dt
pf pξq2´pΦ0pξq2 `ż 1
0
C}f L2pRdq}2 "żRd
t¯ dξ " p2πq´dżX
pΦptξq2 dt
xf, ϕxy2dµpxq .
4.1 Peetre-Wiener type spaces on X
We intend to define two general scales of spaces on X, for which we need a Peetre type maximal
function, given for a measurable function F : X Ñ C by
P
a F px, tq :" ess sup
zPRd
,τ ă1
t
2 ďτ ď2t
F px ` z, τ q
p1 ` z{τ qa
,
x P Rd, 0 ă t ă 1,
P
a F px, 8q :" ess sup
zPRd
F px ` z, 8q
p1 ` zqa
,
x P Rd.
The operator P
take the supremum over t and was used e.g. in [50].
a is a stronger version of the usual Peetre maximal operator Pa, which does not
Definition 4.2. Let p, q P P logpRdq with 0 ă p´ ď p` ă 8 and 0 ă q´ ď q` ă 8 and let
0 ă q ď 8. Further, let a ą 0 and w P W α3
α1,α2. Then we define by
P w
ppq,qpq,apXq " tF : X Ñ C : }F P w
Lw
ppq,q,apXq " tF : X Ñ C : }F Lw
ppq,qpq,a} ă 8u ,
ppq,q,a} ă 8u
two scales of function spaces on X with respective quasi-norms
}F P w
}F Lw
ppq,qpq,a} :"›››wp, 8qP
a F p, 8qLppqpRdq›››
`›››´ż 1
a F p, tqıqpq dt
0 "wp, tqP
t ¯1{qpq
a F p, 8qLppqpRdq›››
ppq,q,a} :"›››wp, 8qP
`´ż 1
0 ›››wp, tqP
a F p, tqLppqpRdq›››
q dt
LppqpRdq›››,
t ¯1{q
.
41
It is not hard to verify that in case a ą d{p´ ` α3 these spaces are rich solid QBF-spaces as
defined and studied in Subsection 2.1. Moreover, the utilization of the Peetre-Wiener operator
P
a ensures that they are locally integrable, even in the quasi-Banach case in contrast to the
ordinary Peetre spaces where Pa is used instead of P
a . In fact, there is an associated locally
bounded weight function given by
νw,ppq,qpqpx, tq "" tα1´d{p´
p1 ` xqα3
p1 ` xqα3
, x P Rd, 0 ă t ă 1,
, x P Rd, t " 8,
(4.2)
such that the following lemma holds true.
Lemma 4.3. We have the continuous embeddings
P w
ppq,qpq,apXq ãÑ L
1{νw,ppq,qpq
8
pXq
and Lw
ppq,q,apXq ãÑ L
1{νw,ppq, q
8
pXq.
It is useful to interpret the component Rd p0, 1q of the index X as a subset of the
Proof.
ax ` b group G " Rd p0, 8q with multiplication px, tqpy, sq " px ` ty, tsq and px, tq´1 "
2 , 2s and define Upx,tq :" px, tqU ´1
p´x{t, 1{tq. Let U ´1 be the inversion of U :" r´2, 2sd r 1
and rUpx,tq :" px, tqU. Further put Qpx,tq :" x ` tr´1, 1sd and Upx,8q :" rUpx,8q :" Qpx,1q t8u.
Then we can estimate for F : X Ñ C and almost all px, tq P X at every fixed py, sq P X
F px, tq À P
a F py, sq.
For convenience, let us introduce
F px, tqχUpx,tqpy, sq ď ess sup
px,tqPXXrUpy,sq
ppq,qpq} :"›››wp, 8qF p, 8qLppq››› `›››´ż 1
}F M w
We obtain for almost all px, tq P X
0 wp, tqF p, tq
qpq dt
t¯1{qpq
Lppq›››.
F px, tq }χUpx,tqM w
ppq,qpq} À }P
a F M w
ppq,qpq} " }F P w
ppq,qpq,a}.
It remains to prove νw,ppq,qpqpx, tq Á }χUpx,tqM w
Upx,tq Ą Qpx,tq r t
2 , 2ts. If 0 ă t ă 1 it follows for x, y P Rd
ppq,qpq}´1. Since U ´1 Ą r´1, 1sd r 1
2 , 2s we have
´ż 1
0 "wpy, sqχUpx,tqpy, sq‰qpyq ds
s ¯1{qpyq
Á lnp4q1{qpyqwpy, tqχQpx,tqpyq Á wpy, tqχQpx,tqpyq
and χUpx,tqpy, 8q " 0. The properties (W1) and (W2) of w P W α3
α1,α2 further imply
wpy, tq Á wpx, tqp1 ` x ´ y{tq´α3 Á t´α1p1 ` xq´α3 p1 ` x ´ y{tq´α3 .
}χUpx,tqM w
This leads to
, Qpx,tq1{p´
Since }χQpx,tqLppq} ě 1
we obtain }χQpx,tqLppq} Á td{p´ and finally arrive at
2 mintQpx,tq1{p`
ppq,qpq} Á t´α1p1 ` xq´α3 }χQpx,tqpqp1 ` x ´ {tq´α3 Lppq}.
u by [16, Lemma 3.2.12] and Qpx,tq " p2tqd
}χUpx,tqM w
ppq,qpq} Á t´α1p1 ` xq´α3 }χQpx,tqLppqpRdq} Á pνw,ppq,qpqpx, tqq´1 ,
where χQpx,tqpyqp1 ` x ´ y{tq´α3 -- χQpx,tqpyq was used. If t " 8 we can argue analogously.
42
4.2 Coorbit identification
As the following lemma shows, every admissible wavelet frame F " FpΦ0, Φq in the sense of
Definition 4.1 is suitable for the definition of coorbits of Peetre-Wiener spaces.
Standing assumptions: For the rest of the paper the indices fulfill p, q P P logpRdq
with 0 ă p´ ď p` ă 8, 0 ă q´ ď q` ă 8. Further q P p0, 8s and w P W α3
α1,α2 for
arbitrary but fixed α2 ě α1 and α3 ě 0.
(4.3)
Lemma 4.4. An admissible continuous wavelet frame F in the sense of (4.1) with generators
Φ0 P SpRdq and Φ P S0pRdq has property F pν, Y q for Y " P w
ppq,q,apXq,
and where ν " νw,ppq,qpq is the corresponding weight from (4.2).
ppq,qpq,apXq and Y " Lw
Proof. The proof goes along the lines of [50, Lem. 4.18]. The kernel estimates in [50, Lem.
4.8, 4.24] have to be adapted to the Peetre-Wiener space. This is a straight-forward procedure
and allows for treating as well the quasi-Banach situation .
Now we are ready for the coorbit characterization of Bw
the weight w defined in (4.4) is an element of the class W α3
ppq,qpqpRdq. Note that
ppq,qpRdq and F w
α1`d{2,α2`d{2.
Theorem 4.5. Let ppq, qpq, q, w fulfill the standing assumptions (4.3). We choose an
admissible continuous wavelet frame F " FpΦ0, Φq according to Definition 4.1. Putting
wpx, tq :"" t´d{2wpx, tq
wpx, 8q
, 0 ă t ă 1 ,
,
t " 8 ,
(4.4)
we have Bw
a ą maxt d
ppq,qpRdq " CopF , L w
p´ , d
p´ ` α3 and F w
q´ u ` α3 in the sense of equivalent quasi-norms.
ppq,q,aq if a ą d
ppq,qpqpRdq " CopF , P w
ppq,qpq,aq if
Proof. By Lemma 4.4 the coorbits exist in accordance with the theory. Now, let f P SpRdq
and F px, tq :" VF f px, tq " xf, ϕpx,tqy with ϕpx,tq as in (4.1). According to [59, Lem. A.3]
VF f px, tq ď CN pf qGN px, tq with GN px, tq "#tN p1 ` xq´N
p1 ` xq´N
, 0 ă t ă 1,
, t " 8,
1pXq and thus F P Lν
where N P N is arbitrary but fixed and CN pf q ą 0 is a constant depending on N and f .
Choosing N large, we have GN P Lν
1}.
This proves f P Hν
1. Even more, given a sequence pfnqnPN Ă SpRdq we have CN pfnq Ñ 0 if
fn Ñ 0 in SpRdq. This is due to the fact that the constants CN pfnq can be estimated by the
Schwartz semi-norms of fn up to order N (see proof of [59, Lem. A.3]). Hence, F Ă SpRdq ãÑ Hν
1
and the voice transform VF extends to S 1pRdq. Moreover, by a straight-forward modification
of the argument in [34, Cor. 20.0.2], the reproducing formula is still valid on S 1pRdq. Therefore
we may apply Lemma 2.28 and use the larger reservoir S 1pRdq.
1} ď CN pf q}GN Lν
1pXq with }F Lν
To see that the coorbits coincide with Bw
ppq,qpqpRdq, note that the functions
Φ " Φp´q and Φ0 " Φ0p´q satisfy the Tauberian conditions (3.5), (3.6) and can thus be used
in the continuous characterization of Theorem 3.11. Recall the notation Φt " t´d Φp{tq. The
assertion is now a direct consequence of the possible reformulation pVF f qp, 8q " Φ0 f and
ppq,qpRdq and F w
pVF f qpx, tq "´DL2
t Φp´q f¯ pxq " td{2´ Φt f¯ pxq
, 0 ă t ă 1, x P Rd.
43
4.3 Atomic decompositions and quasi-Banach frames
Based on the coorbit characterizations of Theorem 4.5 we can now apply the abstract theory
from Section 2 in our concrete setup, in particular the discretization machinery. We will
subsequently use the following covering of the space X. For α ą 0 and β ą 1 we consider the
family U α,β " tUj,ku
jPN0,kPZd of subsets
U0,k " Q0,k t8u
Uj,k " Qj,k rβ´j , β´j`1q
,
k P Zd ,
,
j P N, k P Zd ,
where Qj,k " αβ´j k ` αβ´jr0, 1sd. Clearly, we have X ĂŤjPN0,kPZd Uj,k and U " U α,β is an
admissible covering of X.
The abstract Theorem 2.48 provides atomic decompositions for Bw
ppq,qpRdq and F w
α,β :" osc
To apply it we need to analyze the oscillation kernels oscα,β :" oscU ,Γ and osc
where we choose the trivial phase function Γ " 1. This goes along the lines of [50, Sect. 4.4]
ppq,qpqpRdq.
U ,Γ,
Proposition 4.6. Let F " FpΦ0, Φq be an admissible wavelet frame, Y " Lw
Y " P w
ppq,qpq,apXq, and ν " νw,ppq,qpq the associated weight (4.2).
ppq,q,apXq or
(i) The kernels oscα,β and osc
α,β are bounded operators on Y and belong to Amν .
(ii) If α Ó 0 and β Ó 1 then }oscα,βBY,mν } Ñ 0 and }osc
α,βBY,mν } Ñ 0.
Proof. The proof is a straight-forward modification of [50, Lem. 4.22]. Similar as in Lemma
4.4 above we have to adapt the kernel estimates to the Peetre-Wiener spaces.
Finally, Theorem 2.48 yields the following discretization result in our concrete setting, which
we only state for F w
ppq,qpqpRdq since for Bw
ppq,qpRdq it is essentially the same.
Theorem 4.7. Let ppq, qpq, w fulfill the standing assumptions (4.3), assume further a ą
maxtd{p´, d{q´u ` α3 and let w be given as in (4.4). For an admissible continuous wavelet
frame F " tϕxuxPX there exist α0 ą 0 and β0 ą 1, such that for all 0 ă α ď α0 and 1 ă β ď β0
the family Fd " tϕxj,k ujPN0,kPZd with xj,k " pαkβ´j , β´jq for j P N and x0,k " pαk, 8q is a
discrete wavelet frame with a corresponding dual frame Ed " tej,ku
jPN0,kPZd such that
(a) If f P F w
ppq,qpqpRdq we have the quasi-norm equivalence
}f F w
ppq,qpqpRdq} -- }txf, ϕxj,kyu
-- }txf, ej,kyu
jPN0,kPZdpP w
jPN0,kPZdpP w
ppq,qpq,aq5}
ppq,qpq,aq6} .
(b) For every f P F w
ppq,qpqpRdq the series f " ÿjPN0 ÿkPZd
xf, ej,kyϕxj,k " ÿjPN0 ÿkPZd
xf, ϕxj,kyej,k
converge unconditionally in the quasi-norm of F w
ppq,qpqpRdq.
Proof. The assertion is a consequence of the representation F w
ppq,qpq,aq
and Theorem 2.48. In fact, Proposition 4.6 proves that F has property Dpδ, ν, Y q and Dpδ, ν, L2q
for every δ ą 0. Also note that pP w
ppq,qpqpRdq " CopF , P w
ppq,qpq,aq6 with equivalent quasi-norms.
ppq,qpq,aq5 " pP w
44
4.4 Wavelet bases
According to Appendix A.2 we obtain a family of systems Gc, c P E :" t0, 1ud, whose union
constitutes a tensor wavelet system in L2pRdq. Our aim is now to apply the abstract result
ppq,qpqpRdq. We
in Theorem 2.50 to achieve wavelet basis characterizations of Bw
have to consider the Gramian cross kernels Kc " KU rGc, Fs and K
U rGc, Fs from (2.13)
in our concrete setup.
ppq,qpRdq and F w
c " K
ppq,q,apXq or Y " P w
Lemma 4.8. Let Y " Lw
ppq,qpq,apXq with associated weight ν " νw,ppq,qpq
given in (4.2). Assume that a ą 0 and ppq, qpq, q, w fulfill the standing assumptions (4.3).
Let further F " FpΦ0, Φq be an admissible wavelet frame, Gc be the systems from above, and
Kc " KU rGc, Fs, K
U rGc, Fs, c P E, the corresponding Gramian cross kernels. Then the
kernels Kc and K
c define bounded operators from Y to Y .
c " K
Proof. The proof is analogous to the treatment of the kernels osc in Proposition 4.6, see also
[50, Lem. 4.24].
Now we are ready for the discretization of Bw
ppq,qpRdq and F w
ppq,qpqpRdq in terms of orthonor-
mal wavelet bases. We again only state the result for F w
ppq,qpqpRdq for the sake of brevity.
Theorem 4.9. Let ppq, qpq, w P W α3
α1,α2 fulfill the standing assumptions (4.3), assume further
a ą maxtd{p´, d{q´u ` α3 and let w be given as in (4.4). Let ψ0, ψ1 P L2pRq be the Meyer
scaling function and associated wavelet. Then every f P F w
ppq,qpqpRdq has the decomposition
f "ÿcPE ÿkPZd
λc
0,kψcp ´ kq ` ÿcPEzt0uÿjPN ÿkPZd
λc
j,k2
jd
2 ψcp2j ´kq
with quasi-norm convergence in F w
ppq,qpqpRdq and sequences λc " tλc
j,ku
jPN0,kPZd defined by
λc
j,k " xf, 2
jd
2 ψcp2j ´kqyS 1S
,
j P N0, k P Zd ,
which belong to the sequence space pP w
pH1
qq belongs to F w
ν w,ppq,qpq
ppq,qpqpRdq if all sequences λcpf q belong to pP w
ppq,qpq,aq6.
ppq,qpq,aq6 for every c P E. Conversely, an element f P
Proof. The statement is a direct consequence of Theorem 4.5 and Theorem 2.50. The required
conditions of the kernels Kc, K
c , c P E, have been proved in Lemma 4.8.
A Appendix: Wavelet transforms
A.1 The continuous wavelet transform
As usual SpRdq denotes the locally convex space of rapidly decreasing infinitely differentiable
functions on Rd and its topological dual is denoted by S 1pRdq. The Fourier transform defined
on both SpRdq and S 1pRd) is given by pf pϕq :" f ppϕq, where f P S 1pRdq, ϕ P SpRdq, and
The Fourier transform is a bijection (in both cases) and its inverse is given by ϕ_ " pϕp´q.
pϕpξq :" p2πq´d{2żRd
e´ixξϕpxq dx.
45
Let us introduce the continuous wavelet transform. A general reference is provided by the
monograph [13, 2.4]. For x P Rd and t ą 0 we define the unitary dilation and translation
operators DL2
and Tx by
t
DL2
t g " t´d{2g´
t¯ and Txg " gp ´ xq
,
g P L2pRdq .
The vector g is said to be the analyzing vector for a function f P L2pRdq. The continuous
wavelet transform Wgf is then defined by
Wgf px, tq " xTxDL2
t g, f y
,
x P Rd, t ą 0 ,
where the bracket x, y denotes the inner product in L2pRdq. We call g an admissible wavelet if
cg :"żRd
ξd dξ ă 8 .
pgpξq2
If this is the case, then the family tTxDL2
L2pRq where C1 " C2 " cg.
t gu
tą0,xPRd represents a tight continuous frame in
Many consideration in this paper are based on decay results of the continuous wavelet
transform Wgf px, tq. This decay mainly depends on moment conditions of the analyzing vector
g as well as on the smoothness of g and the function f to be analyzed, see [59, Lem. A.3] which
is based on [52, Lem. 1]
A.2 Orthonormal wavelets
The Meyer wavelets
Meyer wavelets were introduced in [44] and are an important example of wavelets which belong
to the Schwartz class SpRq. The scaling function ψ0 P SpRq and the wavelet ψ1 P SpRq are real,
their Fourier transforms are compactly supported and they fulfill
ψ0p0q " p2πq´1{2
and supp ψ1 Ă„´
8
3
π, ´
2
3
π Y„ 2
3
π,
8
3
π .
Due to the support condition we have infinitely many moment conditions (3.6) on ψ1 and
both functions are fast decaying and infinitly often differentiable, see [60, Section 3.2] for more
properties.
Wavelets on Rd
In order to treat function spaces on Rd let us recall the construction of a d-variate wavelet basis
out of a resolution of unity in Rd, see for instance Wojtaszczyk [60]. It starts with a scaling
function ψ0 and a wavelet ψ1 belonging to L2pRq. For c P E " t0, 1ud the function ψc : Rd Ñ R
i"1 ψcipxiq, and we let
is then defined by the tensor product ψc " Âd
px,tqupx,tqPX be the system with
i"1 ψci, i.e., ψcpxq " śd
Gc " tψc
ψc
px,tq "" TxDL2
t ψc
Txψc
, 0 ă t ă 1 ,
,
t " 8 ,
if c ‰ 0 and ψ0
px,tq ""
0 , 0 ă t ă 1 ,
Txψ0
,
t " 8 .
46
Acknowledgement
The authors would like to thank Felix Voigtlaender for a careful reading of the manuscript and
many valuable comments and corrections. They would further like to thank the anonymous
referees for their careful proofreading and many valuable remarks. In particular, Lemma 2.14
should be pointed out, where now -- based on their input -- also questions of Bochner-measurability
are discussed. Furthermore, a serious technical issue in the proof of Theorem 3.11 has been
fixed. M.S. would like to thank Holger Rauhut for support during his diploma studies where
some ideas of this paper were developed.
References
[1] S. T. Ali, J.-P. Antoine, and J.-P. Gazeau. Continuous frames in Hilbert space. Ann.
Physics, 222(1):1 -- 37, 1993.
[2] A. Almeida and P. Hästö. Besov spaces with variable smoothness and integrability. J.
Funct. Anal., 258(5):1628 -- 1655, 2010.
[3] T. Aoki. Locally bounded linear topological spaces. Proc. Imp. Acad. Tokyo, 18:588 -- 594,
1942.
[4] J.-M. Bony. Second Microlocalization and Propagation of Singularities for Semi-Linear
Hyperbolic Equations. Taniguchi Symp. HERT. Katata, 11 -- 49, 1984.
[5] H.-Q. Bui, M. Paluszyński, and M. H. Taibleson. A maximal function characterization of
weighted Besov-Lipschitz and Triebel-Lizorkin spaces. Stud. Math., 119(3):219 -- 246, 1996.
[6] A. Cohen, I. Daubechies, and J.-C. Feauveau. Biorthogonal bases of compactly supported
wavelets. Comm. Pure Appl. Math., 45(5):485 -- 560, 1992.
[7] D. Cruz-Uribe, A. Fiorenza, and C. J. Neugebauer. The maximal function on variable Lp
spaces. Ann. Acad. Sci. Fenn. Math., 28(1):223 -- 238, 2003.
[8] S. Dahlke, M. Fornasier, H. Rauhut, G. Steidl, and G. Teschke. Generalized coorbit theory,
Banach frames, and the relation to alpha-modulation spaces. Proc. London Math. Soc.
(3), 96:464 -- 506, 2008.
[9] S. Dahlke, G. Kutyniok, G. Steidl, and G. Teschke. Shearlet coorbit spaces and associated
Banach frames. Appl. Comput. Harmon. Anal., 27(2):195 -- 214, 2009.
[10] S. Dahlke, G. Steidl, and G. Teschke. Coorbit spaces and Banach frames on homogeneous
spaces with applications to the sphere. Adv. Comput. Math., 21(1-2):147 -- 180, 2004.
[11] S. Dahlke, G. Steidl, and G. Teschke. Weighted coorbit spaces and Banach frames on
homogeneous spaces. J. Fourier Anal. Appl., 10(5):507 -- 539, 2004.
[12] S. Dahlke, F. De Mari, E. De Vito, D. Labate, G. Steidl, G. Teschke, S. Vigogna. Coorbit
spaces with voice in a Fréchet space. J. Fourier Anal. Appl., DOI 10.1007/s00041-016-
9466-x.
47
[13] I. Daubechies. Ten Lectures on Wavelets, volume 61 of CBMS-NSF Regional Conference
Series in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia, PA, 1992.
[14] L. Diening. Maximal function on generalized Lebesgue spaces Lppq. Math. Inequal. Appl.,
7(2):245 -- 253, 2004.
[15] L. Diening, P. Harjulehto, P. Hästö, Y. Mizuta, and T. Shimomura. Maximal functions in
limiting cases of the exponent. Ann. Acad. Sci. Fenn. Math.,
variable exponent spaces:
34(2):503 -- 522, 2009.
[16] L. Diening, P. Harjulehto, P. Hästö, and M. R užička. Lebesgue and Sobolev Spaces with
Variable Exponents. Number 2017 in Lecture Notes in Mathematics. Springer, 2011.
[17] L. Diening, P. Hästö, and S. Roudenko. Function spaces of variable smoothness and
integrability. J. Funct. Anal., 256(6):1731 -- 1768, 2009.
[18] H. G. Feichtinger. Banach convolution algebras of Wiener's type. In Functions, Series,
Operators; Proc. Conf. Budapest 1980, Vol. I, pages 509 -- 524, 1983.
[19] H. G. Feichtinger. Modulation spaces on locally compact Abelian groups. Technical report,
January 1983.
[20] H. G. Feichtinger and P. Gröbner. Banach spaces of distributions defined by decomposition
methods. I. Math. Nachr., 123:97 -- 120, 1985.
[21] H. G. Feichtinger and K. Gröchenig. A unified approach to atomic decompositions via
integrable group representations. In Function spaces and applications (Lund, 1986), volume
1302 of Lecture Notes in Math., pages 52 -- 73. Springer, Berlin, 1988.
[22] H. G. Feichtinger and K. Gröchenig. Banach spaces related to integrable group represen-
tations and their atomic decompositions, I. Journ. Funct. Anal., 21:307 -- 340, 1989.
[23] G. B. Folland. Real analysis. Pure and Applied Mathematics (New York). John Wi-
ley & Sons, Inc., New York, 1984. Modern techniques and their applications, A Wiley-
Interscience Publication.
[24] M. Fornasier and H. Rauhut. Continuous frames, function spaces, and the discretization
problem. J. Fourier Anal. Appl., 11(3):245 -- 287, 2005.
[25] P. Balazs and N. Holighaus.
frames,
ex-
tensions, annotations and errata for " Continuous
function spaces, and
the discretization problem" by M. Fornasier and H. Rauhut. Online available:
https://www.univie.ac.at/nonstatgab/warping/baho15.pdf
Discretization in generalized coorbit spaces:
[26] P. Balazs, N. Holighaus, and C. Wiesmeyr. Construction of warped time-frequency repre-
sentations on nonuniform frequency scales, part ii: Integral transforms, functions spaces,
atomic decompositions and Banach frames. Online available: arXiv: 1503.05439, 2005.
[27] H. Führ. Coorbit spaces and wavelet coefficient decay over general dilation groups. Trans.
Amer. Math. Soc., 367:7373 -- 7401, 2015.
48
[28] H. Führ. Vanishing moment conditions for wavelet atoms in higher dimensions. Adv.
Comput. Math., 42(1):127 -- 153, 2015.
[29] H. Führ and R. Raisi-Tousi. Simplified vanishing moment criteria for wavelets over general
dilation groups, with applications to abelian and shearlet dilation groups. ArXiv e-prints,
2015. arXiv:1407.0824 [math.FA].
[30] H. Führ and F. Voigtlaender. Wavelet coorbit spaces viewed as decomposition spaces. J.
Funct. Anal., 269:80-154, 2015.
[31] K. Gröchenig. Unconditional bases in translation and dilation invariant function spaces
on Rn. In Constructive theory of functions (Varna, 1987), pages 174 -- 183. Publ. House
Bulgar. Acad. Sci., Sofia, 1988.
[32] K. Gröchenig. Describing functions: atomic decompositions versus frames. Monatsh.
Mathem., 112:1 -- 41, 1991.
[33] K. Gröchenig. Foundations of Time-Frequency Analysis. Appl. Numer. Harmon. Anal.
Birkhäuser Boston, 2001.
[34] M. Holschneider. Wavelets, An Analysis Tool. Oxford University Press, 1995.
[35] S. Jaffard, Y. Meyer. Wavelet methods for pointwise regularity and local oscillations of
functions. Memoirs of the AMS, vol. 123, 1996.
[36] H. Kempka. 2-microlocal Besov and Triebel-Lizorkin spaces of variable integrability. Rev.
Mat. Complut., 22(1):227 -- 251, 2009.
[37] H. Kempka. Atomic, molecular and wavelet decomposition of 2-microlocal Besov and
Triebel-Lizorkin spaces with variable integrability. Funct. Appr., 43(2):171 -- 208, 2010.
[38] H. Kempka. Atomic, molecular and wavelet decomposition of generalized 2-microlocal
Besov spaces. Journ. Function Spaces Appl., 8:129 -- 165, 2010.
[39] H. Kempka and J. Vybíral. Spaces of variable smoothness and integrability: characteriza-
tions by local means and ball means of differences. J. Fourier Anal. Appl., 18(4):852 -- 891,
2012.
[40] H. Kempka and J. Vybíral. A note on the spaces of variable integrability and summability
of Almeida and Hästö. Proc. Amer. Math. Soc., 141(9):3207 -- 3212, 2013.
[41] O. Kováčik and J. Rákosník. On spaces Lppxq and W k,ppxq. Czechoslovak Math. J.,
41(116)(4):592 -- 618, 1991.
[42] Y. Liang, Y. Sawano, T. Ullrich, D. Yang, and W. Yuan. New characterizations of Besov-
Triebel-Lizorkin-Hausdorff spaces including coorbits and wavelets. J. Fourier Anal. Appl.,
18(5):1067 -- 1111, 2012.
[43] Y. Liang, Y. Sawano, T. Ullrich, D. Yang, and W. Yuan. A new framework for generalized
Besov-type and Triebel-Lizorkin-type spaces. Diss. Math., 489, 2013.
[44] Y. Meyer. Principe d'incertitude, bases hilbertiennes et algèbres d'opérateurs. Sém. Bour-
baki, 28:209 -- 223, 1985-1986.
49
[45] A. Nekvinda. Hardy-Littlewood maximal operator on LppxqpRq. Math. Inequal. Appl.,
7(2):255 -- 265, 2004.
[46] W. Orlicz. Über konjugierte Exponentenfolgen. Studia Math., 3:200 -- 212, 1931.
[47] J. Peetre. On spaces of Triebel-Lizorkin type. Ark. Mat., 13:123 -- 130, 1975.
[48] H. Rauhut. Coorbit space theory for quasi-Banach spaces. Studia Math., 180(3):237 -- 253,
2007.
[49] H. Rauhut. Wiener amalgam spaces with respect to quasi-Banach spaces. Colloq. Math.,
109(2):345 -- 362, 2007.
[50] H. Rauhut and T. Ullrich. Generalized coorbit space theory and inhomogeneous function
spaces of Besov-Lizorkin-Triebel type. J. Funct. Anal., 260(11):3299-3362, 2011.
[51] S. Rolewicz. On a certain class of linear metric spaces. Bull. Acad. Polon. Sci. Cl. III.,
5:471 -- 473, XL, 1957.
[52] V. S. Rychkov. On a theorem of Bui, Paluszyński and Taibleson. Proc. Steklov Inst.,
227:280 -- 292, 1999.
[53] M. Schäfer. Generalized coorbit space theory for quasi-banach spaces. Diplomarbeit,
Rheinische Friedrich-Wilhelms-Universität Bonn, Germany, 2012.
[54] F. Voigtlaender. Embedding theorems for decomposition spaces with applications to
wavelet coorbit spaces. PhD thesis, RWTH Aachen University, Germany, 2015. Online
available: https://publications.rwth-aachen.de/record/564979/files/564979.pdf
[55] H. Triebel. Theory of Function Spaces. Birkhäuser, Basel, 1983.
[56] H. Triebel. Characterizations of Besov-Hardy-Sobolev spaces: a unified approach. Journ.
of Approx. Theory, 52:162 -- 203, 1988.
[57] H. Triebel. Theory of Function Spaces II. Birkhäuser, Basel, 1992.
[58] A. I. Tyulenev. Some new function spaces of variable smoothness. Mat. Sb., 206(6):
85 -- 128, 2015.
[59] T. Ullrich. Continuous characterizations of Besov-Lizorkin-Triebel spaces and new inter-
pretations as coorbits. Journ. Funct. Spaces Appl., Article ID 163213, 2012.
[60] P. Wojtaszczyk. A Mathematical Introduction to Wavelets. Cambridge University Press,
1997.
50
|
1506.06177 | 1 | 1506 | 2015-06-19T23:04:00 | Minimal length curves in unitary orbits of a Hermitian compact operator | [
"math.FA",
"math.OA"
] | We study some examples of minimal length curves in homogeneous spaces of B(H) under a left action of a unitary group. Recent results relate these curves with the existence of minimal (with respect to a quotient norm) anti-Hermitian operators Z in the tangent space of the starting point. We show minimal curves that are not of this type but nevertheless can be approximated uniformly by those. | math.FA | math |
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN
COMPACT OPERATOR
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1
,
2
Abstract. We study some examples of minimal length curves in homogeneous spaces of B(H) under
a left action of a unitary group. Recent results relate these curves with the existence of minimal (with
respect to a quotient norm) anti-Hermitian operators Z in the tangent space of the starting point.
We show minimal curves that are not of this type but nevertheless can be approximated uniformly
by those.
1. Introduction
Let H be a separable Hilbert space and K(H) be the algebra of compact operators. In this work
we consider the orbit manifold of a self-adjoint compact operator A by a particular unitary group,
that is
OA = {uAu∗ : u unitary in B(H) and u − 1 ∈ K(H)}.
Given two points, x, y ∈ OA, the rectifiable distance between them is the infimum of the lengths of
all the smooth curves in OA that join x and y. Our purpose is to study the existence and properties
of some particular minimal length curves in OA.
The tangent space at any b ∈ OA is
(T OA)b = {zb − bz : z ∈ K(H), z∗ = −z}
endowed with the Finsler metric given by the usual operator norm k·k. If x ∈ (T OA)b, the existence
of a (not necessarily unique) minimal element z0 such that
kxkb = kz0k = inf {kzk : z ∈ K(H), z∗ = −z, zb − bz = x}
allows in [1] the description of minimal length curves of the manifold by the parametrization
γ(t) = etz0 b e−tz0, t ∈(cid:20)−
π
2 kz0k
,
π
2 kz0k(cid:21) .
These z0 can be described as i(C + D), with C ∈ K(H), C ∗ = C and D a real diagonal operator in
an orthonormal basis of eigenvectors of A.
If we consider B ⊂ A von Neumann algebras and a ∈ A, a∗ = a, there always exists an element
b0 in B such that ka + b0k ≤ ka + bk, for all b ∈ B (see [6]). The element a0 + b0 is called minimal
in the class [a0] of Ah/Bh. However, in the case of A = K(H), a C ∗-algebra which is not a von
Neumann algebra, and B ⊂ K(H) a subalgebra there is not always a minimal compact operator in
any class in K(Hh)/Bh. In [4] we exhibit an example of this fact. In this case, the existence of a
best approximant for C ∈ K(H), C ∗ = C is guaranteed when C, for example, has finite rank (see
Proposition 5.1 in[1]).
The above motivated us to study the following, among other issues, in the unitary orbit of a
Hermitian operator. Let b ∈ OA and x ∈ (T OA)b and suppose that there exists a uniparametric
2010 Mathematics Subject Classification. MSC: Primary: 22F30, 43A85, 47B15, 47A58, 53C22. Secondary: 47B07,
47B10, 47C15.
Key words and phrases. Unitary orbits, geodesic curves, minimal operators in quotient spaces, approximation of
minimal length curves.
1
2
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
curve ψ(t) = etZ be−tZ which is a minimal length curve among all the smooth curves joining b and
ψ(t) in OA for t ∈h− π
2kZk,
π
2kZki:
• Would Z be a compact minimal lifting of x (i.e x = Zb − bZ and kZk = kxkb)?
• Can ψ be approximated in OA by a sequence of minimal length curves of matrices?
The present work continues the analysis made in [1] of this homogeneous spaces and we use minimality
characterizations that we developed in [4].
The results in this paper are divided in three parts. In the first we describe and study minimal
length curves in the orbit of a particular compact Hermitian operator. In the second part we construct
a sequence of minimal length curves of matrices which converges uniformly to the minimal length
curves found in the first part. Finally, in the third part we study cases of anti-Hermitian compact
operators such that their best bounded diagonal approximants are not compact and the properties
of the minimal curves they determinate.
2. Preliminaries and notation
Let (H, h, i) be a separable Hilbert space. We denote by khk = hh, hi1/2 the norm for each h ∈ H.
Let B(H) denote the set of bounded operators (with the identity operator I) and K(H), the two-sided
closed ideal of compact operators on H. Given A ⊂ B(H), we use the superscript ah (resp. h) to
note the subset of anti-Hermitian (resp. Hermitian) elements of A.
We consider the group of unitary operators in B(H)
U(H) = {u ∈ B(H) : uu∗ = u∗u = I}
and the unitary Fredholm group, defined as
Uc(H) = {u ∈ U(H) : u − I ∈ K(H)}.
We denote with k·k the usual operator norm in B(H) and with [ , ] the commutator operator, that
is, for any T, S ∈ B(H)
[T, S] = T S − ST.
It should be clear from the context the use of the same notation k·k to refer to the operator norm or
the norm on H.
We define the unitary orbit of a fixed A ∈ K(H), A = A∗, as
(2.1)
OA is an homogeneous space if we consider the action πb : Uc(H) → OA, πb(u) = ubu∗. For each
b ∈ OA, the isotropy group Ib is
OA = {uAu∗ : u ∈ Uc(H)} ⊂ K(H).
Ib = {u ∈ Uc(H) : ubu∗ = b}.
Since for each u ∈ Uc(H) there always exists X ∈ K(H)ah such that u = eX (see Proposition 3), the
isotropy can be redefined by
Ib = {eX ∈ Uc(H) : X ∈ K(H)ah, [X, b] = 0}.
For each b ∈ OA, its tangent space is
(T OA)b = {Y b − bY : Y ∈ K(H)ah} ⊂ K(H)ah.
Consider a smooth curve (i.e. C 1 and with derivative non equal to zero) u : [0, 1] → Uc(H) such that
u(0) = 1 y u′(0) = Y , then the differential of the surjective map πb at 1 is
(dπb)1(Y ) =
d
dt
πb(u(t))t=0 = u′(0)b u∗(0) + u(0)b u′(0)∗
= Y b1∗ + 1bY ∗ = Y b − bY = [Y, b].
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN COMPACT OPERATOR 3
For every b ∈ OA we consider each tangent space as
(T OA)b
∼= (T Uc(H))1/(T Ib)1 ∼= K(H)ah/({b}′)ah,
being {b}′ the set of elements that commute with b in a C ∗-algebra A (in this particular case
A = K(H)). Let us consider the Finsler metric, defined for each x ∈ (T OA)b as
This metric can be expressed in terms of the projection to the quotient K(H)ah/({b}′)ah as
kxkb = inf{kY k : Y ∈ K(H)ah such that [Y, b] = x}
kY b − bY kb = k[Y ]k = inf
C∈({b}′)ah
kY + Ck
Uc(H).
for each class [Y ] = (cid:8)Y + C : C ∈ ({b}′)ah(cid:9). This Finsler norm is invariant under the action of
There always exists Z ∈ B(H)ah such that [Z, b] = x and kZk = kxkb. Such element Z is called
minimal lifting for x, and Z may not be compact and/or unique (see [4]). Consider piecewise smooth
curves β : [a, b] → OA. We define the rectifiable length of β as
and the rectifiable distance between two points of OA, named c1 and c2, as
L(β) =Z b
a
kβ′(t)kβ(t) dt,
dist(c1, c2) = inf{L(β) : β is smooth, β(a) = c1, β(b) = c2}.
If A is any C ∗-algebra of B(H) and {ek}∞
the set of diagonal operators with respect to this basis, that is
k=1 is a fixed orthonormal basis of H, we denote with D(A)
Given an operator Z ∈ A, if there exists an operator D1 ∈ D(A) such that
D(A) = {T ∈ A : hT ei, eji = 0 ,
for all i 6= j} .
kZ + D1k = dist (Z, D (A)) ,
we say that D1 is a best approximant of Z in D(A). In other terms, the operator Z + D1 verifies the
following inequality
kZ + D1k ≤ kZ + Dk
for all D ∈ D(A). In this sense, we call Z + D1 a minimal operator or similarly we say that D1 is
minimal for Z. If Z is anti-Hermitian it holds that
since kIm(X)k ≤ kXk for every X ∈ A.
dist (Z, D (A)) = dist(cid:0)Z, D(cid:0)Aah(cid:1)(cid:1) ,
Let T ∈ B(H) and consider the coefficients Tij = hT ei, eji for each i, j ∈ N, that define an infinite
matrix (Tij)i,j∈N. The jth-column and ith-row of T are the vectors in ℓ2 given by cj(T ) = (T1j, T2j, ...)
and fj(T ) = (Ti1, Ti2, ...), respectively.
We use σ(T ) and R(T ) to denote the spectrum and range of T ∈ B(H), respectively.
We define Φ : B(H) → D(B(H)), Φ(X) = Diag(X), that takes the main diagonal (i.e the elements
of the form {hXei, eii}i∈N) of an operator X and builds a diagonal operator in the chosen fixed
basis of H. For a given bounded sequence {dn}n∈N ⊂ C we denote with Diag(cid:0){dn}n∈N(cid:1) the diagonal
(infinite) matrix with {dn}n∈N in its diagonal and 0 elsewhere.
The following theorem is similar to Theorem 1 in [4] but this version only requires that T ∈ B(H)h,
instead of T ∈ K(H)h. The proof is exactly the same tha the one in the case where T is compact.
Theorem 1. Let T ∈ B(H)h described as an infinite matrix by (Tij)i,j∈N. Suppose that T satisfies:
(1) Tij ∈ R for each i, j ∈ N,
(2) there exists i0 ∈ N satisfying Ti0i0 = 0, with Ti0n 6= 0, for all n 6= i0,
4
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
(3) if T [i0] is the operator T with zero in its i0th-column and i0th-row then
(where kci0(T )k denotes the Hilbert norm of the i0th-column of T ), and
(4) if the Tnn's satisfy that, for each n ∈ N, n 6= i0:
kci0(T )k ≥(cid:13)(cid:13)T [i0](cid:13)(cid:13)
Tnn = −
hci0(T ), cn(T )i
Ti0n
.
Then T is minimal, that is
kT k = kci0(T )k =
inf
D∈D(B(H))
kT + Dk =
inf
D∈D(K(H)h)
kT + Dk
and moreover, D = Diag(cid:0){Tnn}n∈N(cid:1) is the unique bounded minimal diagonal operator for T .
3. The unitary Fredholm orbit of a Hermitian compact operator
In this section we consider the unitary Fredholm orbit OA of a particular case of a Hermitian
compact operator, that is: A ∈ K(H)h, A = uDiag ({λi}i∈N) u∗, with u ∈ Uc(H) and {λi}i∈N ⊂ R
such that λi 6= λj for each i 6= j. Consider OA as defined in section 2 and b = Diag ({λi}i∈N) ∈ OA.
The isotropy Ib is the set {ed : d ∈ D(K(H)ah)} and (T OA)b can be identified with the quotient
space K(H)ah/D(K(H)ah).
Proposition 2. Let b = Diag ({λi}i∈N) ∈ OA. For each x ∈ (T OA)b, if Z ∈ K(H)ah is such that
[Z, b] = x, then
(3.1)
kxkb =
inf
D∈D(K(H)ah)
kZ + Dk
Proof. If Y1, Y2 ∈ {Y ∈ K(H)ah : [Y, b] = x} then
Y1 − Y2 ∈ {D :
[D, b] = Db − bD = 0} = {b}′
and since b is a diagonal operator, then every D is diagonal. Thus
or equivalently: Y1 = Y2 + D, with D ∈ D(K(H)ah). Then,
Y1 − Y2 = D, with D diagonal
kxkb = inf{kY k : Y ∈ K(H)ah such that Y = Y2 + D, with D ∈ D(K(H)ah)}.
(cid:3)
Fix x = [Zr, b] = Zrb − bZr ∈ B(H)ah, where Zr is an anti-Hermitian operator defined as the
infinite matrix given by
(3.2)
Zr = i
· · ·
· · ·
· · ·
· · ·
. . .
0
0
0
d4
...
rγ rγ2 rγ3
0
γ2
d2
rγ
γ2
rγ2
γ
rγ3 γ2
d4
...
...
...
γ
d3
γ2
...
0
0
0 d2
0
0
0
0
...
...
· · ·
· · ·
· · ·
· · ·
. . .
}
+ i
0
0
d3
0
...
{z
D0
= Yr + D0.
· · ·
· · ·
· · ·
· · ·
. . .
}
(3.3)
= i
rγ rγ2 rγ3
0
γ2
0
rγ
γ2
rγ2
γ
rγ3 γ2
0
...
...
...
γ
0
γ2
...
{z
Yr
(3) r ≥
kY [1]+D0k
k=1 γ2k)1/2 , where Y [1] = Yr −
(P∞
1−γ2 . Notice that lim
j→∞
0
· · ·
· · ·
rγ
rγ2
· · ·
rγ3
· · ·
...
. . .
rγ rγ2 rγ3
0
0
0
0
0
0
...
...
0
0
0
...
.
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN COMPACT OPERATOR 5
The entries of the operator Zr are such that:
(1) γ ∈ R such that γ < 1.
(2) For each j ∈ N, j > 1: dj = − 1−γj−2
1−γ − γj
dj = 1
γ−1.
Observe that the definition of each dj is independent of the parameter r.
The operator −iZr fulfills the conditions of minimality stated in Theorem 1 stated in the Prelim-
inaries and has been studied in [4]. Therefore,
k[Yr]k =
inf
D∈D(K(H)ah)
kYr + Dk = kYr + D0k = kZrk .
Moreover, the diagonal operator D0 is the unique minimal diagonal (bounded, but non compact)
operator for Yr. Since D0b − bD0 = 0, then x = Yrb − bYr ∈ (T OA)b and
kxkb = kZrb − bZrkb =
inf
D∈D(K(H)h)
kYr + Dk = k[Yr]k = kZrk < kYr + Dk
for all D ∈ D(K(H)ah). In other words, there is no compact minimal lifting for x in this case.
The following Proposition is a characterization of the unitary Fredholm group in terms of operators
in K(H)ah.
Proposition 3. w ∈ Uc(H) if and only if there exists X ∈ K(H)ah such that w = eX .
Proof. Given w ∈ Uc(H), by Lemma 2.1 in [1] there exists X ∈ K(H)ah such that w = eX . On the
other hand, consider X ∈ K(H)ah and the series expansion of eX
eX = 1 + X +
1
2
X 2 +
1
3!
X 3 + ...
= 1 + X(cid:20)1 +
and therefore eX ∈ Uc(H).
1
2
X +
1
3!
X 2 + ...(cid:21) = 1 + K , K ∈ K(H),
(cid:3)
Remark 4. Even if Z /∈ K(H)ah, eZ may belong to Uc(H). Indeed, let X0 ∈ K(H)ah, then Z =
X0 + 2πiI /∈ K(H)ah but
For Zr as in (3.2) define the uniparametric curve β by
eX0+2πiI = eX0 ∈ Uc(H).
(3.4)
β(t) = etZr be−tZr , t ∈(cid:20)−
π
2 kZrk
,
π
2 kZrk(cid:21) .
To prove that β is a curve in OA, we introduce first the next result.
Lemma 5. Let Zr the operator defined in (3.2). Then for each t ∈ R, there exist zt ∈ C, zt = 1
and U(t) ∈ Uc(H) such that
etZr = ztU(t).
Proof. Let α = −i lim
n→∞
dn = i
1−γ . Then etZr+αIt = etZr etαI . Observe that etαI = etαI. Thus
etZr = e−tαetZr+tαI = e−tαetYr+tD0+tαI ,
6
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
with e−tα ∈ C, e−tα = 1 for every t ∈ R. Moreover, D0 + αI ∈ D(K(H)ah), since it is a bounded
diagonal and
(cid:12)(cid:12)(cid:12)(D0 + αI)jj(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)
−
1 − γj−2
1 − γ
−
γj
1 − γ2 +
γj−2
1 − γ
−
1
1 − γ(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)
→ 0
γj
1 − γ2(cid:12)(cid:12)(cid:12)(cid:12)
when j → ∞. Therefore, since tZr + tαI ∈ K(H)ah for every t ∈ R then U(t) = etZr+tαI ∈ Uc(H)
and
etZr = ztU(t), with zt = etα ∈ C.
(cid:3)
Remark 6. For any minimal lifting Z ∈ B(H)ah of x = [Y, b], the curve κ(t) = eZtbe−Zt has minimal
length over all the smooth curves in P = {uAu∗ : u ∈ U(H)} that join β(0) = b and β(t), with
t ≤ π
2kZrk (Theorem II in [6]). Since OA ⊆ P, then for each t0 ∈h− π
2kZki follows that
L(κ) = inf{L(χ) : χ ⊂ P, χ is smooth, χ(0) = b and χ(t0) = β(t0)}
2kZk,
π
≤ inf{L(χ) : χ ⊂ OA, χ is smooth, χ(0) = b and χ(t0) = β(t0)}
= dist(b, β(t0)).
Using the previous remark and Lemma 5 we can prove the following Theorem.
Theorem 7. Let A = uDiag ({λi}i∈N) u∗, with u ∈ Uc(H) and {λi}i∈N ⊂ R such that λi 6= λj for
each i 6= j. Let b = Diag ({λi}i∈N) ∈ OA and the parametric curve β defined in (3.4). Then β
satisfies:
1−γ I)be−t(Zr + i
(1) β(t) = et(Zr + i
(2) β′(0) = x = Yrb − bYr = Zrb − bZr ∈ (T OA)b.
(3) β has minimal length between all smooth curves in OA joining b with β(t0), for every t0 ∈
1−γ I), which means that β(t) ∈ OA for every t.
2kZrk,
h− π
π
2kZrki. That is
L(cid:16)β[0,t0](cid:17) = inf{L(χ) : χ is smooth, χ(0) = b and χ(t0) = β(t0)}
= dist(b, β(t0)).
(4) L(cid:16)β[0,t0](cid:17) = t0 kxkb, for each t0 ∈h− π
(1) By Lemma 5, if U(t) = etZr+t
i
2kZrk,
π
2kZrki.
Proof.
1−γ I, then β can be rewritten as
β(t) = ztU(t)b(ztU(t))∗ = ztztU(t)bU −1(t)
= U(t)bU −1(t) = et(Zr+ i
1−γ I)be−t(Zr + i
1−γ I)
and U(t) ∈ Uc(H) for each t ∈ R. Follows that β(t) ∈ OA for every t ∈ R.
(2) β′(0) = etZr [Zr, b] e−tZr(cid:12)(cid:12)t=0.
(4) Observe that L(β) =R t0
(3) Observe that kZrk = k[Yr]kB(H)ah/D(B(H))ah and Zr is (the unique) minimal lifting of x = [Yr, b]
in B(H). Then, the result is a direct consequence of Remark 6.
0 kβ′(t)kβ(t) dt = t0 kYrb − bYrkb. Indeed,
kβ′(t)kβ(t) =(cid:13)(cid:13)ZretZr be−tZr − etZr bZre−tZr(cid:13)(cid:13)β(t) =(cid:13)(cid:13)etZr [Zr, b] e−tZr(cid:13)(cid:13)β(t)
=(cid:13)(cid:13)zzU(t) [Zr, b] U −1(t)(cid:13)(cid:13)β(t) = z2(cid:13)(cid:13)U(t) [Zr, b] U −1(t)(cid:13)(cid:13)β(t)
=(cid:13)(cid:13)U(t) [Zr, b] U −1(t)(cid:13)(cid:13)U (t)bU −1(t) = kZrb − bZrkb
= kYrb − bYrkb = kxkb ,
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN COMPACT OPERATOR 7
where the equality kU(t) [Zr, b] U −1(t)kU (t)bU −1(t) = kZrb − bZrkb holds due to the unitary
invariance of the Finsler norm.
(cid:3)
Summarizing, if Zα = Zr + i
1−γ I ∈ K(H)ah, we obtained that the parametric curve given by
πb ◦ (etZα) = etZαbe−tZα
has minimal length between elements of OA. Nevertheless, the operator Zα is not a minimal element
in its class (recall that [Zr] = {Zr + D : D ∈ D(K(H)ah) = [Yr]}). On the other hand,
etZαbe−tZα = etZr be−tZr
and Zr is minimal, but it does not belong to K(H)ah. We conclude with the following comment.
Remark 8. Let b ∈ OA, b = Diag ({λi}i∈N) such that λi 6= λj for each i 6= j. Then, there exist
minimal length curves of the form ρ(t) = etZbe−tZ in OA such that they join b with other points of
the orbit, but with Z ∈ K(H)ah and kZk > k[Z]kK(H)ah/D(K(H)ah).
4. Approximation with minimal length curves of matrices
There are two main objectives in this section: the first is to build two sequences of minimal matrices
which approximate Zr and Zr + i
1−γ I in the strong operator topology (SOT) and in the operator
norm, respectively. The second objetive is to find a family of minimal length curves of matrices
which approximates the curve β defined in (3.4).
Let Yr be the anti-Hermitian compact operator defined in (3.3) and consider the following decom-
position
(4.1)
Yr = rL + Y [1], where L = i
γ γ2 γ3
· · ·
0
· · ·
0
γ
0
γ2 0
· · ·
0
γ3 0
· · ·
0
...
...
...
. . .
kY [1]+D0k
0
0
0
...
.
Let D0 be the diagonal bounded operator defined in (3.3). If r ≥
is minimal.
kc1(L)k
, then Zr = rL + Y [1] + D0
Let us consider for each n ∈ N the orthogonal projection Pn over the space generated by {e1, ..., en}.
We define the following finite range operators
(4.2)
Yn = rnPnLPn + PnY [1]Pn,
with rn ∈ R>0 for each n ∈ N. For each n ∈ N we define the diagonal operator Dn = iDiag(cid:16){d(n)
uniquely determined by the conditions:
k }k∈N(cid:17)
1 = 0;
(1) d(n)
(2) hc1(Yn + Dn), cj(Yn + Dn)i = 0, for each j ∈ N, j 6= 1;
(3) d(n)
k = 0, for every k > n.
Thus, if we use the convention that P2−3
j=0 γj = 0 = Pn−1
each n as
j=n γ2j−n+1 then each d(n)
k
is determined for
(4.3)
d(n)
k = −Pk−3
j=0 γj −Pn−1
d(n)
k = 0
j=i γ2j−k < 0
if
for all
k ≤ n
k > n.
The proof is by induction over the indices k for every n ∈ N. Observe that the choice of each d(n)
k
independent of the parameter rn.
is
8
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
The following lemma will be used to prove the minimality of each Yn + Dn for a fixed rn.
Lemma 9. Let Yn = rnPnLPn + PnY [1]Pn and Dn as defined in (4.2) and (4.3) for each n ∈ N,
respectively. Then
sup
Proof. Fix n ∈ N. Since supn∈N(cid:12)(cid:12)(cid:12)d(n)
1≤k≤n(cid:12)(cid:12)(cid:12)d(n)
k (cid:12)(cid:12)(cid:12)
(cid:13)(cid:13)PnY [1]Pn + Dn(cid:13)(cid:13) ≤(cid:13)(cid:13)PnY [1]Pn(cid:13)(cid:13) + kDnk ≤ kPnk2(cid:13)(cid:13)Y [1](cid:13)(cid:13) + sup
≤(cid:13)(cid:13)Y [1](cid:13)(cid:13) +(cid:12)(cid:12)d(n)
n (cid:12)(cid:12) ≤(cid:13)(cid:13)Y [1](cid:13)(cid:13) + kD0k < ∞.
n∈N(cid:13)(cid:13)PnY [1]Pn + Dn(cid:13)(cid:13) < ∞.
n (cid:12)(cid:12)(cid:12) ≤ kD0k, for D0 the diagonal operator defined in (3.3), then
n (cid:12)(cid:12) ≤(cid:13)(cid:13)Y [1](cid:13)(cid:13) + sup
n∈N(cid:13)(cid:13)PnY [1]Pn + Dn(cid:13)(cid:13) ,(cid:13)(cid:13)Y [1] + D0(cid:13)(cid:13)(cid:27) .
As a consequence of this lemma, there exists a constant M0 ∈ R>0 such that:
Now we can prove the minimality of each Yn + Dn for all n ∈ N.
M0 = max(cid:26)sup
n∈N(cid:12)(cid:12)d(n)
(4.4)
(cid:3)
Proposition 10. Let Yn = rnPnLPn +PnY [1]Pn and Dn as defined in (4.2) and (4.3) for each n ∈ N,
kc1(PnLPn)k . Then for each n ∈ N
respectively. Consider the constant M0 as in (4.4) and define rn =
the operator Yn + Dn is minimal in K(H)ah/D(K(H)ah), that is
M0
Proof. Fix n ∈ N. Without loss of generality, we can consider Yn + Dn as an n × n matrix. Then
k[Yn]k =
inf
D∈D(K(H)ah)(cid:13)(cid:13)(cid:13)Yn + D(cid:13)(cid:13)(cid:13) = kYn + Dnk = M0.
1 = 0;
• d(n)
• hc1(Yn + Dn), cj(Yn + Dn)i = 0, for each j ∈ N, 2 ≤ j ≤ n;
• kc1(Yn + Dn)k = rn kc1(PnLPn)k = M0 ≥(cid:13)(cid:13)PnY [1]Pn + Dn(cid:13)(cid:13).
As an n × n matrix, Dn is the unique minimal diagonal operator for Yn (see Theorem 8 in [8]). Since
inf
D∈D(K(H)ah)
kYn + Dk =
follows that
min
D∈D(Mn(C)ah)(cid:13)(cid:13)(cid:13)Yn + D(cid:13)(cid:13)(cid:13) ,
k[Yn]k = kYn + Dnk .
Observe that the norm of the minimal operator Yn + Dn is M0 for every n ∈ N.
Remark 11. For every n ∈ N
inf
D∈D(K(H)ah)
kYn + Dk =
min
D′∈D(Mn(C)ah)
kYn + D′k = kYn + Dnk ,
(cid:3)
but there is no uniqueness of the D′ ∈ D(K(H)ah) that attain the minimum. Moreover, every block
operator of the form
Cn =(cid:18)Dn
0 Dc(cid:19), with Dc diagonal and such that kDck ≤ kc1(Yn)k satisfies
0
kYn + Cnk = max {kYn + Dnk ; kDck} = kYn + Dnk = k[Yn]k .
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN COMPACT OPERATOR 9
Reconsider the operator Yr = rL + Y [1] fixing r = M0
kc1(L)k . Note that
(cid:13)(cid:13)Y [1] + D0(cid:13)(cid:13)
kc1(L)k
≤ r < ∞
where the last inequality holds due to Lemma 9. Then, Zr = Yr + D0 satisfies the hypothesis of
Theorem 1 and is a minimal operator with D0, the unique (non compact) bounded diagonal operator
such that
Moreover,
Therefore,
(4.5)
k[Yr]k =
inf
D∈D(K(H)ah)
kYr + Dk = kZrk .
k[Yr]k = kc1(Zr)k = kc1(Yr)k = M0.
k[Yr]k = k[Yn]k , for all n ∈ N.
The following result relates Yr with Yn.
Proposition 12. Let Yr be the operator defined in (4.1) and {Yn}∞
operators defined in (4.2).
rn =
kc1(PnLPn)k for each n ∈ N, are fixed. Then
M0
If M0 is the real constant defined in (4.4) such that r = M0
n=1 the family of finite range
kc1(L)k and
rn = r.
(1) lim
n→∞
(2) Yn → Yr when n → ∞ in the operator norm.
2 , follows that lim
n→∞
rn =
(cid:3)
k }n∈N defined in (4.3) converges to dk when n → ∞, for
Proof.
r.
(1) Since kc1(PnLPn)k =(cid:0)Pn−1
(2) kYr − Ynk =(cid:13)(cid:13)rL + Y [1] − rnPnLPn − PnY [1]Pn(cid:13)(cid:13)
i=1 γ2i(cid:1) 1
i=1 γ2i(cid:1) 1
2 and kc1(L)k =(cid:0)P∞
≤ krL ± rnL − rnPnLPnk +(cid:13)(cid:13)Y [1] − PnY [1]Pn(cid:13)(cid:13)
≤ r − rn kLk + rn kL − PnLPnk +(cid:13)(cid:13)Y [1] − PnY [1]Pn(cid:13)(cid:13) → 0
when n → ∞, since L and Y [1] are Hilbert-Schmidt operators and rn → r.
Observe that the numerical sequence {d(n)
each k ∈ N
d(n)
k ց −
γj −
γ2j−k = −
k−3
Xj=0
∞
Xj=k
γj −
γk
1 − γ2 = dk.
i−3
Xj=0
As a consequence, the sequence of diagonal operators {Dn}n∈N converges SOT to the unique best
approximant (non compact) diagonal D0 ∈ D(B(H)) for Yr.
Proposition 13. Let Yr be the operator defined in (3.3) and D0 the unique bounded diagonal operator
such that k[Yr]kK(H)ah/D(K(H)ah) = kYr + D0k. Let {Dn}n∈N be the sequence of finite range diagonal
operators defined in (4.3). Then
Dn → D0 SOT when n → ∞.
Proof. {Dn − D0}n∈N is a bounded family of B(H) and
(Dn − D0) (ek) = d(n)
k − dk → 0
when n → ∞ for every ek that belongs to the fixed orthonormal basis. Then standard arguments of
operator theory imply that Dn → D0 SOT when n → ∞ (see [5]).
(cid:3)
10
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
Observe that Propositions 12 and 13 imply that lim
n→∞
Yn + Dn = Zr SOT. Since Dn ∈ K(H)ah
for all n and D0 /∈ D(K(H)ah), the convergence can not be in the operator norm. To establish the
second main result of this section we prove first the convergence in the operator norm of Yn + DnαI
to Zr + αI, for a particular α ∈ R.
Proposition 14. Let Yr, D0, {Yn}n∈N, {Dn}n∈N, {Pn}n∈N be the operators and sequence of operators
defined previously in (4.1), (4.2) y (4.3). Then
Yn + Dn +
i
1 − γ
Pn → Yr + D0 +
i
1 − γ
I,
in the operator norm when n → ∞.
Proof. Let ǫ > 0, then
By Proposition 12, there exists n1 ∈ N such that kYr − Ynk < ǫ, for all n ≥ n1. Focus on the second
term. For each n ∈ N
Yr + D0 +
i
1 − γ
(cid:13)(cid:13)(cid:13)(cid:13)
D0 +
(cid:13)(cid:13)(cid:13)(cid:13)
I − Dn −
− d(n)
1 − γ
; sup
dk +
1
D +
I − Dn −
.
i
1 − γ
i
1 − γ
i
1 − γ
i
1 − γ
i
1 − γ
1
1 − γ
= sup
dk +
I − Yn − Dn −
Pn(cid:13)(cid:13)(cid:13)(cid:13)
Pn(cid:19)kk(cid:12)(cid:12)(cid:12)(cid:12)
≤ kYr − Ynk +(cid:13)(cid:13)(cid:13)(cid:13)
Pn(cid:13)(cid:13)(cid:13)(cid:13)
Pn(cid:13)(cid:13)(cid:13)(cid:13)
k∈N(cid:12)(cid:12)(cid:12)(cid:12)
γ2j−k(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1≤k≤n(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
= max( max
k>n(cid:12)(cid:12)(cid:12)(cid:12)
Xj=n
j=n γ2j−k(cid:12)(cid:12)(cid:12) and sup
k>n(cid:12)(cid:12)(cid:12)dk + 1
1−γ(cid:12)(cid:12)(cid:12) converges to 0 when n → ∞. Then, there
γ2j−k(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1≤k≤n(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
max( max
k>n(cid:12)(cid:12)(cid:12)(cid:12)
Xj=n
n ≥ n0 ⇒ (cid:13)(cid:13)(cid:13)(cid:13)
k −(cid:18) 1
) .
1 − γ(cid:12)(cid:12)(cid:12)(cid:12)
) < ǫ.
1 − γ(cid:12)(cid:12)(cid:12)(cid:12)
Pn(cid:13)(cid:13)(cid:13)(cid:13)
1−γ Pn converges to Yr + D0 + i
1−γ I when n → ∞ in the operator
(cid:3)
1≤k≤n(cid:12)(cid:12)(cid:12)P∞
I − Yn − Dn −
i
1 − γ
i
1 − γ
; sup
dk +
1
Yr + D0 +
∞
∞
< 2ǫ,
By Proposition 13, max
exists n2 ∈ N such that for each n ≥ n0
Finally, if n0 = max{n1; n2} follows that
which means that Yn + Dn + i
norm.
In the above proof we also obtained that {Dn + i
1−γ Pn}n∈N, which is a sequence of finite range
1−γ Pn
1−γ I are not minimal operators, they are useful to construct minimal length curves
1−γ Pn to construct a sequence
operators, converges in the operator norm to D0 + i
and Yr + D0 + i
in the unitary orbit of A. We will also use the operators Yn + Dn + i
of minimal length curves that converge to β defined in (3.4).
1−γ I ∈ D(K(H)ah). Even though Yn + Dn + i
The first result in this direction is the convergence of the sequence of exponential curves in OA.
Proposition 15. Let b ∈ OA and ςn(t) = etZnbe−tZn, a sequence of curves in OA with t ∈ R and
{Zn}n∈N ⊂ K(H)ah such that kZn − Zk → 0 when n → ∞. If we define ς(t) = etZbe−tZ , then
uniformly in the operator norm when n → ∞ for any interval [t1, t2] ⊂ R.
ςn → ς
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN COMPACT OPERATOR 11
Proof. Let ǫ > 0.
It is known that the exponential map exp : K(H)ah → Uc(H) is Lipschitz continuous in compact sets
of K(H), then there exists n0 ∈ N such that
kςn(t) − ς(t)k ≤(cid:13)(cid:13)etZnbe−tZn − etZ be−tZn(cid:13)(cid:13) +(cid:13)(cid:13)etZbe−tZn − etZbe−tZ(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:0)etZn − etZ(cid:1) be−tZn(cid:13)(cid:13) +(cid:13)(cid:13)etZ b(cid:0)e−tZn − e−tZ(cid:1)(cid:13)(cid:13)
≤(cid:0)(cid:13)(cid:13)etZn − etZ(cid:13)(cid:13) +(cid:13)(cid:13)e−tZn − e−tZ(cid:13)(cid:13)(cid:1) kbk .
for all n ≥ n0 ⇒ ( (cid:13)(cid:13)etZn − etZ(cid:13)(cid:13) < ǫ
(cid:13)(cid:13)e−tZn − e−tZ(cid:13)(cid:13) < ǫ
kbk
kbk,
for each t in a closed interval [t1, t2] ⊂ R. Therefore
for each n ≥ n0 and t ∈ [t1, t2], which implies that ςn → ς uniformly in the operator norm in that
interval.
(cid:3)
kςn(t) − ς(t)k < ǫ
If we consider the sequence {Yn + Dn + i
1−γ Pn}n∈N and use Proposition 14 then
i
i
Yn + Dn +
Pn → Yr + D0 +
1 − γ
I
1 − γ
in the operator norm when n → ∞. Define for each n ∈ N and t0 ∈ R the curves parametrized by
(4.6)
βn(t) = et(Yn+Dn+ i
1−γ Pn)be−t(Yn+Dn+ i
1−γ Pn), t ∈ [0, t0].
Observe that these can be considered as matricial type curves.
Theorem 16. Let A and b ∈ OA as in Theorem 7. Let {βn}n∈N be the sequence of curves defined in
(4.6), and β be the curve defined in (3.4). Then, for each n ∈ N
.
2k[Yn]k,
βn(0) = b
β′
n(0) = Ynb − bYn ∈ (T OA)b.
i
1−γ Pn commutes with Yn + Dn.
(2) βn(t) = et(Yn+Dn)be−t(Yn+Dn) for all t, since
, π
(1) (cid:26)
(3) For each t0 ∈h− π
(4) βn : [0, t0] → OA with t0 ∈h− π
Moreover, by Proposition 15, βn → β uniformly in the operator norm in the interval h− π
L(cid:16)βn[0,t0](cid:17) = t0 k[Yn]k = t0 M0 = L(cid:16)β[0,t0](cid:17) .
2M0i is a minimal length curve in OA.
n(0) → β′(0) in the norm k.kb of (T OA)b.
2M0i holds that
π
2k[Yn]ki =h− π
, π
2M0
(5) β′
2M0
2M0
, π
2M0i.
Proof. The proof of items (1), (2), (3) is analogous to the proof in Theorem 7. The equality k[Yn]k =
M0 = k[Yr]k is due to Proposition 10.
Since for each n ∈ N fixed Yn + Dn is a minimal compact operator, by Theorem I in [6] βn is a
2kYn+Dnk. Then
minimal length curve between all curves in OA joining βn(0) = b and βn(t) with t ≤
(4) is proved.
π
We proceed to prove (5): fix ǫ > 0. Then there exists n0 ∈ N such that if n ≥ n0 then kYn − Yrk < ǫ.
Therefore,
kβ′
n(0) − β′(0)kb = inf(cid:8)kZk : Z ∈ K(H)ah, [Z, b] = (Yn − Yr) b − b (Yn − Yr)(cid:9)
12
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
=
inf
D∈D(K(H)ah)
kYn − Yr + Dk ≤ kYn − Yrk < ǫ
for each n ≥ n0. Then kβ′
n(0) − β′(0)kb → 0 when n → ∞.
(cid:3)
Therefore, we obtained a minimal length curve β ⊂ OA that can be uniformly approximated by
minimal curves of matrices {βn}. Nevertheless, β does not have a minimal compact lifting, although
each βn has at least one minimal matricial lifting.
5. Bounded minimal operators Z + D with Z ∈ K(H) and non compact diagonal D
Let Yr, D0 be the operators defined in (3.3). To establish the equality β(t) = eYr+D0+ i
1−γ Ibe−(Yr+D0+ i
1−γ I)
in Theorem 7 the following properties were essential:
(1) D0 + i
(2)
1−γ I ∈ D(K(H)ah) and
i
1−γ I commutes with Zr and b but
i
1−γ I /∈ K(H).
This can be generalized.
Proposition 17. Let Z ∈ K(H)ah and suppose that there exists D1 ∈ D(B(H)ah) such that
k[Z]kK(H)ah/D(K(H)ah) = kZ + D1k
and D1 is not compact. If there exists λ ∈ iR such that limj→∞ (D1)jj = λ, then the curve
χ(t) = et(Z+D1−λI)be−t(Z+D1−λI)
has minimal length between all the smooth curves in OA joining b with χ(t0), for t0 ∈ [ π
2kZk,
π
2k[Z]k].
Proof. First observe that Re ((D1)jj) = 0 for each j ∈ N, since D1 ∈ D(B(H)ah). Then,
lim
j→∞
(D1)j = λ
and λ 6= 0 since D1 is not compact. Therefore, using functional calculus and Proposition 6 in [4]
kZ + D1 − λIk = max{− k[Z]k − λ ; k[Z]k − λ} > k[Z]k .
Also D1 −λI ∈ D(K(H)ah), since (D1 − λI)jj = (D1)jj − λ → 0, when j → ∞. Then, Z + D1 −λI
is not minimal in K(H)ah/D(K(H)ah) but the curve parameterized by
χ(t) = et(Z+D1−λI)be−t(Z+D1−λI) ∈ OA
has minimal length, as χ is equal to the curve δ(t) = et(Z+D1)be−t(Z+D1), which has minimal length
in the homogeneous space {uAu∗ : u ∈ U(H)} (Theorem II in [6]). Therefore χ has minimal length
in OA.
(cid:3)
Given Z ∈ K(H)ah, it is not true that every diagonal operator D1 such that Z + D1 is minimal
fulfills the condition
∃λ ∈ iR such that lim
j→∞
(D1)jj = λ.
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN COMPACT OPERATOR 13
Indeed, consider the following operator
(5.1)
Z0 = i
−δ
0
γ
γ −δ2
γ −δ2
0 −δ2
0
γ2
0
−δ
γ
−δ2 −δ2 −δ2
γ2
γ2
−δ3 −δ3 −δ3 −δ3 −δ3
γ3
γ3
...
...
γ2 −δ3 γ3
γ2 −δ3 γ3
γ2 −δ3 γ3
γ2 −δ3 γ3
0 −δ3 γ3
γ3
0
...
0
γ3
...
γ3
...
γ3
...
γ3
...
γ2
, with γ, δ ∈ (0, 1).
· · ·
· · ·
· · ·
· · ·
· · ·
· · ·
· · ·
. . .
It is easy to prove that Z0 is a Hilbert Schmidt operator.
Let D′
0 = iDiag(cid:0){d′
n}n∈N(cid:1) the unique bounded diagonal operator such that
hc1(Z0), cn(Z0 + D′
0)i = 0, ∀ n ∈ N.
(5.2)
Simple calculations show that the condition (5.2) implies that {d′
n}n∈N satisfies that
and
d′
2k = k−1
Xj=1
2k−1 = k−1
Xj=1
δj! − k−1
Xj=1
δj! − k−2
Xj=1
d′
γj! +
δk+2
δ (cid:19)k
1 − δ2 +(cid:18)γ2
1
1 − γ2
γj! −
γk+1
γ(cid:19)k
1 − γ2 −(cid:18)δ2
γ
1 − δ2
for each k ∈ N. If γ2 ≤ δ and δ2 ≤ γ both sequences,{d′
If Z [1]
0
is the operator Z0 defined in (5.1) but with zeros in its first column and row and r =
0 ) + Z [1]
then r(Z0 − Z [1]
minimal diagonal operator for r(Z0 − Z [1]
then
0 + D′
0 ) + Z [1]
0 is a minimal operator by Theorem 1 and D′
2k}k∈N and {d′
2k−1}k∈N, are convergent.
(cid:13)(cid:13)(cid:13)
Z [1]
0 +D′
c1(Z0)
0(cid:13)(cid:13)(cid:13)
0 is the unique bounded
0 . Also, if we fix the conditions γ2 = δ and δ2 < γ
lim
k→∞
d′
2k =
δ
1 − δ
−
γ
1 − γ
+
1
1 − γ2 and lim
k→∞
d′
2k−1 =
δ
1 − δ
−
γ
1 − γ
,
which implies that {(D′
the sequence {hDen, eni}n∈N has at least two different limits).
0)nn}n∈N has no limit. We call these diagonals "oscillant" in the sense that
Observe that an approximation to Z0 by matrices can be built as the one done in section 4.
Consider for each n ∈ N the orthogonal projection Pn and define the following finite range operators
(5.3)
Zn = rnPn(Z0 − Z [1]
0 )Pn + PnZ [1]
0 Pn,
with rn ∈ R>0 for each n ∈ N. For each n ∈ N we define a diagonal operator D′
uniquely determined for each n as
n = iDiag(cid:16){d(n)′
l
}l∈N(cid:17)
1 = 0;
(1) d(n)′
(2) hc1(Zn + D′
(3) d(n)′
n), cj(Zn + D′
l = 0, for every l > n.
n)i = 0, for each j ∈ N, j 6= 1;
14
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
Then, each d(n)′
l
is determined for each n as
(5.4)
d(n)′
2k =(cid:16)Pk−1
2k−1 =(cid:16)Pk−1
j=1 δj(cid:17) −(cid:16)Pk−1
j=1 δj(cid:17) −(cid:16)Pk−2
j=1 γj(cid:17) +(cid:18)P⌊ n−k
j=1 γj(cid:17) −(cid:18)P⌊ n−1
d(n)′
l = 0
j=k
j=1
d(n)′
2 ⌋
2 ⌋
j=1
δ2j+k(cid:19) + P
γ2j−k+1(cid:19) − P
⌊ n−2
2 ⌋
γ2j
δk
⌊ n
2 ⌋
j=1 δ2j
γk−1
if k < n
2
if k < n+1
2
for all l > n.
The proof is by induction over the indices k for every n ∈ N. Observe the folllowing:
k
d(n)′
(1) The choice of each d(n)′
(2) lim
n→∞
(3) lim
n→∞
2k =(cid:16)Pk−1
2k−1 =(cid:16)Pk−1
j=1 δj(cid:17) −(cid:16)Pk−1
j=1 δj(cid:17) −(cid:16)Pk−2
d(n)′
2k−1.
(4) For every k ∈ N and for each n:
is independent from the parameter rn.
j=1 γ2j
δk
j=1 γj(cid:17) +(cid:16)P∞
j=1 γj(cid:17) −(cid:16)P∞
j=1 δ2j+k(cid:17) + P∞
j=k γ2j−k+1(cid:17) − P∞
= d′
= d′
2k.
j=1 δ2j
γk−1
Then, kD′
(5) D′
n → D′
Diag(cid:0){d′
0k = sup
k∈N
{(cid:12)(cid:12)d′
2k−1(cid:12)(cid:12) ; d′
0 SOT, since Diag(cid:16){d(n)′
2k−1}k∈N(cid:1) SOT.
M1 = max(cid:26)sup
For any injective σ : N → N define the projection
(5.5)
(5.6)
With the previous properties, there exists M1 ∈ R>0 such that:
2k−1 ≤ d(n)′
2k ≤ d′
2k.
2k−1 ≤ d(n)′
d′
2k} ≥ kD′
nk.
2k−1}k∈N(cid:17) →
2k }k∈N(cid:17) → Diag ({d′
2k}k∈N) SOT and Diag(cid:16){d(n)′
n(cid:13)(cid:13)(cid:13) ,(cid:13)(cid:13)(cid:13)Z [1]
n∈N(cid:13)(cid:13)(cid:13)PnZ [1]
P σ =Xk∈N
0(cid:13)(cid:13)(cid:13)(cid:27) .
0 Pn + D′
eσ(k) ⊗ eσ(k).
0 + D′
Thus, the following result is a direct consequence of all previous remarks.
Theorem 18. Let Z0, D′
n be the operators defined in
(5.1), (5.2), (5.3) and (5.4) for each n ∈ N, respectively. Consider the real constant M1 as in (5.5)
and define rn =
for each n ∈ N and r =
0, Zn = rnPn(Z0 − Z [1]
0 )Pn + PnZ [1]
0 Pn and D′
. If
M1
M1
(cid:13)(cid:13)(cid:13)
c1(cid:16)Pn(Z0−Z [1]
0 )Pn(cid:17)(cid:13)(cid:13)(cid:13)
λ = lim
n→∞
d′
2k, µ = lim
n→∞
0 (cid:17)(cid:13)(cid:13)(cid:13)
c1(cid:16)Z0−Z [1]
(cid:13)(cid:13)(cid:13)
d′
2k−1,
then
(1) Zn + D′
n is minimal in K(H)ah/D(K(H)ah) and
k[Zn]k =
inf
D∈D(K(H)ah)
kZn + Dk = kZn + D′
nk = M1.
(2) If P σ1 and P σ2 are the projections defined in (5.6) for σ1(k) = 2k and σ2(k) = 2k − 1,
respectively, then
Zn + D′
n − λPnP σ1Pn − µPnP σ2Pn → r(Z0 − Z [1]
0 ) + Z [1]
0 + D′
0 − λP σ1 − µP σ2
in the operator norm when n → ∞.
MINIMAL LENGTH CURVES IN UNITARY ORBITS OF A HERMITIAN COMPACT OPERATOR 15
Proof.
(1) Observe that if D′
−i(Zn + D′
n is determined as in (5.4) the operator
n) fullfils the conditions stated in Theorem 1 and
nk = kc1(Zn + D′
kZn + Dk = kZn + D′
inf
n)k = M1.
D∈D(K(H)ah)
(2) Let ǫ > 0. Since Z0 is compact and rn → r then there exists n1 ∈ N such that
for each n ≥ n1. Similarly than in the case of diagonal with one limit point (see proof of
Proposition 14), for each n ∈ N:
(cid:13)(cid:13)(cid:13)Zn − r(Z0 − Z [1]
0 ) + Z [1]
ǫ
2
,
0 (cid:13)(cid:13)(cid:13) <
kD′
n − λPnP σ1Pn − µPnP σ2Pn − D′
0 + λP σ1 + µP σ2k
(5.7)
l − d(n)′
l
d′
2k − λ ; sup
= sup
l − λ (PnP σ1Pn)ll − µ (PnP σ2Pn)ll − d(n)′
l∈N(cid:12)(cid:12)(cid:12)d′
= max(cid:26) max
1≤l≤n(cid:12)(cid:12)(cid:12)d′
d(n)′
2k = d′
(cid:12)(cid:12)(cid:12) ; sup
k>n
Since lim
2k−1, lim
n→∞
n→∞
such that the last expression is smaller than ǫ
2k, lim
n→∞
d(n)′
2k−1 = d′
d′
2k = λ and lim
n→∞
2 for every n ≥ n2. Therefore, it holds that
d′
2k−1 = µ, there exists n2 ∈ N
l − λ (P σ1)ll − µ (P σ2)ll(cid:12)(cid:12)(cid:12)
k>n(cid:12)(cid:12)d′
2k−1 − µ(cid:12)(cid:12)(cid:27) .
0 − λP σ1 − µP σ2i(cid:13)(cid:13)(cid:13)
0 + D′
n − λPnP σ1Pn − µPnP σ2Pn −hr(Z0 − Z [1]
0 ) + Z [1]
(cid:13)(cid:13)(cid:13)Zn + D′
≤(cid:13)(cid:13)(cid:13)Zn − r(Z0 − Z [1]
for every n ≥ max{n1; n2}.
0 (cid:13)(cid:13)(cid:13) + kD′
0 ) + Z [1]
n − λPnP σ1Pn − µPnP σ2Pn − D′
0 + λP σ1 + µP σ2k < ǫ
0 ) + Z [1]
Remark 19. As r(Z0 − Z [1]
0 , with Z0 and r defined previously, there exist other compact
operators such that its best bounded diagonal approximant oscillates. Moreover, there exist examples
of minimal bounded operators in which the elements on the main diagonal are the union of m subse-
quences (m ∈ N) such that each one converges to a different limit. For those m−oscillant operators
an analogous result as that of Theorem 18 can be obtained with essentially the same arguments. Nev-
ertheless, the techniques used in Theorems 7 and 16 to prove that the curves constructed in (3.4)
and (4.6) belong to OA cannot be adapted to the case of an oscillant minimal diagonal for a compact
Z ∈ K(H)ah.
(cid:3)
References
[1] Andruchow, E. and Larotonda, G.: The rectifiable distance in the unitary Fredholm group. Studia Math. 196
(2010), no. 2, 151 -- 178.
[2] Andruchow, E., Larotonda, G., Recht, L. and Varela, A.: A characterization of minimal Hermitian matrices.
Linear algebra and its applications, vol. 436, no. 7, (2012).
[3] Andruchow, E. , Mata-Lorenzo, L., Mendoza, A., Recht, L. and Varela, A.: Minimal Hermitian matrices with
fixed entries outside the diagonal. Revista de la Uni´on Matem´atica Argentina, vol. 49, no. 2, (2008), 17-28.
[4] Bottazzi, T. and Varela, A.: Best approximation by diagonal compact operators. Linear algebra and its applica-
tions, Elsevier, no 439 (2013), p. 3044-3056.
[5] Conway, J.B.: A course in operator theory. AMS (2000).
[6] Dur´an, C., Mata-Lorenzo, L. and Recht, L.: Metric geometry in homogeneous spaces of the unitary group of a
C ∗-algebra. I. Minimal curves. Adv. Math. 184 (2004), no. 2, 342 -- 366.
[7] Dur´an, C., Mata-Lorenzo, L. and Recht, L.: Metric geometry in homogeneous spaces of the unitary group of a
C ∗-algebra. II. Geodesics joining fixed endpoints. Integral Equations Operator Theory 53 (2005), no. 1, 33 -- 50.
[8] Kloubouk, A. and Varela, A.: Minimal 3 × 3 Hermitian matrices. Publicaciones previas del Instituto Argentino
de Matem´atica, 455 (2012).
16
TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA1,2
1 Instituto Argentino de Matem´atica "Alberto P. Calder´on", Saavedra 15 3? piso, (C1083ACA)
Ciudad Aut´onoma de Buenos Aires, Argentina
2 Instituto de Ciencias, Universidad Nacional de General Sarmiento, J. M. Gutierrez 1150,
(B1613GSX) Los Polvorines, Pcia. de Buenos Aires, Argentina
E-mail address: [email protected], [email protected]
|
Subsets and Splits