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
|
|---|---|---|---|---|---|---|---|---|---|
1810.11680 | 1 | 1810 | 2018-10-27T17:18:04 | Numerical Range and Compressions of the Shift | [
"math.FA",
"math.CV"
] | The numerical range of a bounded, linear operator on a Hilbert space is a set in $\mathbb{C}$ that encodes important information about the operator. In this survey paper, we first consider numerical ranges of matrices and discuss several connections with envelopes of families of curves. We then turn to the shift operator, perhaps the most important operator on the Hardy space $H^2(\mathbb{D})$, and compressions of the shift operator to model spaces, i.e.~spaces of the form $H^2 \ominus \theta H^2$ where $\theta$ is inner. For these compressions of the shift operator, we provide a survey of results on the connection between their numerical ranges and the numerical ranges of their unitary dilations. We also discuss related results for compressed shift operators on the bidisk associated to rational inner functions and conclude the paper with a brief discussion of the Crouzeix conjecture. | math.FA | math |
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
KELLY BICKEL†, AND PAMELA GORKIN‡
Abstract. The numerical range of a bounded, linear operator on a Hilbert space is a
set in C that encodes important information about the operator. In this survey paper, we
first consider numerical ranges of matrices and discuss several connections with envelopes
of families of curves. We then turn to the shift operator, perhaps the most important
operator on the Hardy space H 2(D), and compressions of the shift operator to model
spaces, i.e. spaces of the form H 2 (cid:9) θH 2 where θ is inner. For these compressions of the
shift operator, we provide a survey of results on the connection between their numerical
ranges and the numerical ranges of their unitary dilations. We also discuss related results
for compressed shift operators on the bidisk associated to rational inner functions and
conclude the paper with a brief discussion of the Crouzeix conjecture.
1. Introduction
Let B(H) denote the set of bounded, linear operators on a Hilbert space H. Then for
A ∈ B(H), its numerical range or field of values is the subset of C defined as follows:
W (A) = {(cid:104)Ax, x(cid:105) : x ∈ H,(cid:107)x(cid:107) = 1}.
This crucial object encodes many properties of the operator A and is closely related to
the spectrum of A, denoted σ(A). Indeed, σ(A) is always contained in W (A) and the
convex hull of σ(A) can be recovered from the numerical ranges of operators similar to A
[45]. Typically, W (A) encodes significantly more information about A than the spectrum
does. For instance, if W (A) is contained in R, then A must be Hermitian. Similarly, if
H is finite dimensional, then W (A) is compact and the maximal elements of W (A) are
related to the combinatorial structure of A [54].
Due to these and many other such properties, numerical ranges and related objects have
found numerous applications in diverse areas including differential equations, numerical
analysis, and quantum computing, see for example [5, 28, 34, 36, 50, 51, 60]. As the topic
of numerical ranges is both natural and useful, it has been extensively studied and the
current body of research is quite vast. Thus this survey is not meant to be in any way
exhaustive. Instead we refer the interested readers to the books [39, 42, 46].
This survey primarily covers two topics: connections between numerical ranges and
envelopes and the numerical ranges of compressions of the shift. Section 2 presents several
relationships between envelopes of families of curves F and numerical ranges of matrices
W (A). Specifically, let F be continuously differentiable and let F denote the family of
curves of (x, y) points that satisfy F (x, y, t) = 0 for different values of t ranging over
an interval. Then, intuitively, an envelope of F is a curve that, at each of its points, is
Primary 47A12; Secondary 47A13, 30C15
Date: October 30, 2018.
Key words and phrases. Numerical Range, Envelopes, Compressions of Shifts, Unitary Dilations.
† Research supported in part by National Science Foundation DMS grant #1448846.
‡ Research supported in part by Simons Foundation Grant #243653.
1
2
KELLY BICKEL†, AND PAMELA GORKIN‡
∞(cid:88)
∞(cid:88)
an2 < ∞.
tangent to a member of the family. Envelopes have a number of applications and appear,
for example, in both economics and in robotics and gear construction, [53, 59]. They are
also connected to numerical ranges in several ways. First, in [48], Kippenhahn showed
that for any matrix A, the boundary ∂W (A) of the numerical range is -- after removing a
finite number of corners -- an envelope of the family of support lines of W (A). Similarly in
[27], Donoghue outlined a proof of the elliptical range theorem, which characterizes W (A)
for 2 × 2 matrices, that constructs ∂W (A) as an envelope of a family of circles.
Sections 3-5 concern compressions of the shift and their numerical ranges. To define
these classical operators, recall that the Hardy space on the unit disk H 2(D) consists of
functions of the form
(1)
f (z) =
anzn, where
n=0
n=0
A particularly important operator acting on H 2 is S, the (forward) shift operator defined
by [S(f )](z) = zf (z). To compress S, first recall that an inner function θ is a bounded
analytic function on D whose radial limits have modulus one almost everywhere. Then
for each inner θ, the space θH 2 is a subspace of H 2 and the model space Kθ and the
compression of the shift Sθ can be defined as follows:
Kθ := H 2 (cid:9) θH 2
and Sθ = PθSKθ,
√
Arguably the prettiest results concern finite Blaschke products B(z) = λ(cid:81)n
where Pθ is the orthogonal projection onto Kθ. The study of such spaces and operators
has been extensive and forms a key subarea of both classic operator theory and complex
analysis; for the main theory, we direct the readers to [30, 63]. Indeed, compressed shifts
represent a large class of operators. For a contraction T ∈ B(H), define the defect operator
1 − T (cid:63)T and the defect space DT = DTH. Then if T is a completely non-unitary
DT =
contraction with defect indices dimDT = dimDT (cid:63) = 1, the Nagy-Foias functional model
says that T is unitarily equivalent to a compressed shift, see [64].
z−aj
1−aj z
with λ = 1, which we discuss in Section 3. In this case, KB is finite dimensional and
SB has a nice (upper-triangular) matrix representation in terms of the zeros of B. This
allowed Gau and Wu to obtain a simple characterization of the unitary 1-dilations of SB
and show that
(2)
see [31, 32]. Their work -- and that of Mirman in [55] -- shows that each ∂W (SB) also satisfies
an elegant geometric condition called the Poncelet property.
W (SB) = ∩{W (U ) : U is a unitary 1-dilation of SB},
There are numerous ways to generalize or extend these investigations of W (SB). For
example, researchers have studied Sθ for infinite Blaschke products and general inner
functions, considered operators with higher defect indices, and studied compressions of
shifts in the bivariate setting [12, 8, 15]. While versions of (2) are true in some settings,
many open questions remain. For details about such generalizations, see Sections 4-5.
As shown by two previously-discussed topics, numerical ranges are at the heart of many
beautiful results and open questions in both operator and function theory. Perhaps the
most famous open question concerning numerical ranges is Crouzeix's conjecture [21],
which states:
j=1
3
Conjecture (2004): There is a constant C such that for any polynomial p ∈ C[z] and
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
n × n matrix A, the following inequality holds:
(cid:107)p(A)(cid:107) ≤ C maxp(z)z∈W (A).
The best constant should be C = 2.
Initially in [23], Crouzeix showed that 2 ≤ C ≤ 11.08. However, significant recent
progress has been made on improving C, proving special cases, and identifying other
questions that imply the conjecture, see [17, 22, 35, 61]. We include the details in Section
6.
2. Numerical Ranges and Envelopes
2.1. Preliminaries. To examine the connections between numerical ranges of matrices
and envelopes of families of curves, we need some well-known results about numerical
ranges. We state these for matrices, but many results have generalizations to bounded
linear operators on a Hilbert space. First, it is easy to show that numerical ranges are well
behaved with respect to operations like unitary conjugation and affine transformation:
Theorem 2.1. If A is an n × n matrix, then
a. For U an n × n unitary matrix, W (U (cid:63)AU ) = W (A).
b. For α, β ∈ C, W (αA + βI) = αW (A) + β := {αz + β : z ∈ W (A)}.
It is also easy to see that W (A) contains the eigenvalues of A; indeed if λ is an eigenvalue
of A with normalized eigenvector x, then
(cid:104)Ax, x(cid:105) = (cid:104)λx, x(cid:105) = λ(cid:104)x, x(cid:105) = λ.
One of the deepest results about numerical ranges follows from theorems of Toeplitz and
Hausdorff in [44, 67] and states:
Theorem 2.2 (Toeplitz-Hausdorff theorem). If A is an n × n matrix, then W (A) is
convex.
If A is normal, then this is the entire story. Indeed, the numerical range of a normal
matrix A is the convex hull of its eigenvalues. To see this, recall that A must be unitarily
equivalent to some diagonal matrix
λ1
...
,
D =
where λ1, . . . , λn are the eigenvalues of A. For each x ∈ Cn, we have (cid:104)Dx, x(cid:105) =(cid:80)n
j=1 λjxj2.
This implies that W (D), and hence W (A), is the convex hull of the eigenvalues of A. More
generally, the closure of the numerical range of a (bounded) normal operator is the convex
hull of its spectrum, see [42, pp. 112].
In contrast, non-normal matrices typically have more points in their numerical ranges.
λn
Consider A1 and A2 given below:
(3)
A1 =
and A2 =
(cid:21)
(cid:20) 0 0
0 0
(cid:20) 0 1
0 0
(cid:21)
.
4
KELLY BICKEL†, AND PAMELA GORKIN‡
They have the same eigenvalues with the same multiplicity, so the spectrum does not
distinguish between them. However, the numerical range does.
Indeed, the elliptical
range theorem says
Theorem 2.3. Let A be a 2 × 2 matrix with eigenvalues a and b. Then W (A) is an
elliptical disk with foci at a and b and minor axis given by (tr(A(cid:63)A) − a2 − b2)1/2.
This implies W (A1) = {0}, but W (A2) = {z : z ≤ 1/2}. More generally, the elliptical
range theorem is a key tool in several proofs of the Toeplitz-Hausdorff theorem, see for
example [39, pp. 4].
2.2. Envelopes. In what follows, we will study numerical ranges of matrices using en-
velopes of families of curves. To make this precise, let F be the family of curves given
by F (x, y, t) = 0, for some continuously differentiable function F . For each t in some
interval, let Γt denote the curve of (x, y) points satisfying F (x, y, t) = 0.
There is some historic vagueness concerning the definition of the envelope of a family of
curves, and many sources indicate at least three different ways to define it [14, 20, 47, 65].
Arguably, the most natural definition is the following:
Definition 2.4. A geometric envelope E1 of F is a curve so that each point on E1 is a
point of tangency to some member of the family Γt (and often, each Γt is touched by E1).
In practice, it can be hard to use Definition 2.4 to find exact formulas for a geometric
envelope. In contrast, one can often compute the following sets explicitly:
Definition 2.5. The limiting-position envelope E2 of F is the set of limits of intersections
points of nearby Γt; a point (x, y) ∈ E2 if there are sequences (tn) and (tn) converging to
some t, so that (x, y) is a limit of intersection points of Γtn and Γtn.
Definition 2.6. The discriminant envelope E3 of F is the set of points (x, y) for which
there is a value of t so that both F (x, y, t) = 0 and Ft(x, y, t) = 0.
In general, these definitions do not yield the same set of points. However it is known
that E1 and E2 are contained in E3, see [14, Propositions 1 and 2]. Moreover, if the curves
in E3 can be parameterized as (x(t), y(t)) and the relevant derivatives are nonvanishing
in the following sense:
F 2
x (x, y, t) + F 2
y (x, y, t) (cid:54)= 0 and x(cid:48)(t)2 + y(cid:48)(t)2 (cid:54)= 0,
then E1 = E3, see [20, pp. 173]. However, there are nice F for which nearby Γt never
intersect; these cases give examples where E1 and E3 may contain points, but E2 is empty.
Often, the easiest envelope to compute is the discriminant envelope E3. For simple F,
one can find E3 by setting F (x, y, t) = 0 and Ft(x, y, t) = 0 and then eliminating the
parameter t; this process is called the envelope algorithm. There are also connections
between the boundary of F and its envelope(s) and often, the boundary (or a piece of the
boundary) of F will correspond to an envelope. For example in [47], Kalman observes
that if the boundary is smoothly parameterized by t, then it is part of the geometric
envelope. However, such a condition is difficult to check. For more information about
these envelopes, additional definitions, and connections to boundaries, see [14, 20, 47, 65]
and the references therein.
Let us now consider two connections between numerical ranges and envelopes.
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
5
2.3. Finding the numerical range via Kippenhahn. Let A be an n×n matrix. Then
A can be decomposed as
A = (cid:60)(A) + i(cid:61)(A), where (cid:60)(A) = A+A(cid:63)
2
and (cid:61)(A) = A−A(cid:63)
.
2i
Using this decomposition, Kippenhahn developed a method that produces ∂W (A) as the
geometric envelope of a family of lines. Specifically, we say that a line is a support line
of W (A) if it touches ∂W (A) in either one point or along a line segment. The following
theorem, which can be found in Hochstenbach and Zachlin's translation of Kippenhahn's
paper [48, Theorem 9], allows us to identify support lines of W (A):
Theorem 2.7. If A = (cid:60)(A)+i(cid:61)(A) with α1 ≤ α2 ≤ ··· ≤ αn the eigenvalues of (cid:60)(A) and
β1 ≤ β2 ≤ ··· ≤ βn the eigenvalues of (cid:61)(A), then the points of W (A) lie in the interior
or on the boundary of the rectangle constructed by the line x = α1, x = αn; y = β1, y = βn
positioned parallel to the axes. The sides of the rectangle share either one point (possibly
with multiplicity > 1) or one closed interval with the boundary of W (A).
For a matrix A, let Me(A) denote the maximum eigenvalue of (cid:60)(A). Then Theorem
2.7 says that the vertical line
is a support line for W (A). To identify other support lines of W (A), fix γ ∈ (0, 2π) and
consider the rotated matrix e−iγA. As shown in the accompanying figure, the numerical
range W (e−iγA) is exactly the numerical range of W (A) rotated by the angle −γ. By
Theorem 2.7, it has vertical support line
x = Me(A)
x = Me(e−iγA).
Rotating this line by an angle γ gives the new line
(4)
which is a support line of W (A).
x cos γ + y sin γ = Me(e−iγA),
Figure 2.1. The Kippenhahn construction giving support lines of W (A).
Letting γ vary over [0, 2π] gives a family of support lines of W (A). Then the convexity
of W (A) implies that the intersections of ∂W (A) with these support lines must give the
entire boundary of W (A).
To connect this to envelopes, consider the family of lines F given in (4) for γ ∈ [0, 2π].
Then the differentiable components of ∂W (A) are geometric envelopes for F; this is
easy to see because each point of ∂W (A) lies on a line in (4) and as long as ∂W (A) is
6
KELLY BICKEL†, AND PAMELA GORKIN‡
differentiable at that point, it must be tangent to the line. Restricting to differentiable
components of ∂W (A) is reasonable because, as also proved by Kippenhahn, there are at
most a finite number of places where ∂W (A) is not differentiable. Moreover these singular
points must occur at the eigenvalues of A, see [48, Theorem 13].
Kippenhahn actually proved much more than this. In particular, he completely ana-
lyzed the boundary of the numerical range in the 3 × 3 setting. For more information
on this, see [48, Section 7]. For results about ∂W (A) for general A ∈ B(H), see Agler's
paper [1].
2.4. The elliptical range theorem. Recall that the elliptical range theorem, given in
Theorem 2.3, characterizes the numerical ranges of 2 × 2 matrices. Indeed, if A is a 2 × 2
matrix with eigenvalues a and b, then W (A) is an elliptical disk with foci a and b and
minor axis (tr(A(cid:63)A) − a2 − b2)1/2. C.-K. Li gave a simple computational proof of this
in [52] and other proofs can be found in [46, 56].
One can also use envelopes of families of circles to prove the elliptical range theorem.
The proof idea described here is due to Donoghue [27], but many of the details appear in
[13]. First observe that each A is unitarily equivalent to
B =
, where p = (tr(A(cid:63)A) − a2 − b2)1/2.
(cid:20)a p
(cid:21)
0 b
(cid:20)0 1
(cid:21)
T =
(cid:20)0 m
(cid:21)
(cid:20)
0
1
If A has a repeated eigenvalue a and J =
W (B) = pW (J) + a. A simple computation shows W (J) is the disk of radius 1
at (0, 0), which gives the result.
, then Theorem 2.1 gives W (A) =
2 centered
0 0
If A has distinct eigenvalues, we can define
where m = pb−a and apply Theorem 2.1 to show W (A) = W (B) = (b − a)W (T ) + a.
Indeed
Thus it suffices to study W (T ), which can be realized as a family of circles.
for some t ∈ [0, 1] and
each normalized z ∈ C2 can be written as z =
θ1, θ2 ∈ [0, 2π), which gives
√
teiθ1
1 − t2eiθ2
√
(cid:104)T z, z(cid:105) = (1 − t2) + mei(θ2−θ1)(t
√
(1 − t2, 0) and radius mt
F (x, y, t) = 0 for
This implies W (T ) is the union of circles (cid:83)
t∈[0,1] Ct, where Ct is the circle with center
1 − t2. Equivalently, W (T ) is the family of curves satisfying
1 − t2).
(cid:21)
F (x, y, t) := (x − (1 − t2))2 + y2 − m2t2(1 − t2)
and t ∈ [0, 1]. To find the discriminant envelope, we apply the envelope algorithm. Taking
the equations Ft(x, y, t) = 0 and F (x, y, t) = 0 and solving for x and y gives the curves
x(t) = (1 − t2) + m2
(5)
and the point (1, 0). It is easy to check that the curves in (5) give exactly the ellipse
m2(t2 − t4) − m4
4 (1 − 2t2)2
2 (1 − 2t2) and y(t) = ±(cid:113)
(6)
(x− 1
1+m2 + y2
2 )2
m2 = 1
4.
This leads to the question:
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
Do the envelope curves in (5) give the boundary of the union(cid:83)
t∈[0,1] Ct?
In general, the relationship between the boundary of a family of curves and its envelope
is murky. However in this case, the answer is yes. For the details about that and the fact
that W (T ) is the closed elliptical disk with boundary (5), see [13]. Then the elliptical
range theorem follows immediately from this result about W (T ).
7
Figure 2.2. W (T ) as a union of circles Ct with elliptical boundary from (6).1
f (z) = (cid:80)∞
n=0 anzn where (cid:80)∞
3. The numerical range of a compressed shift operator (single variable)
Recall that H 2 is the Hardy space on the unit circle T consisting of functions of the form
n=0 an2 < ∞ and an inner function is a bounded analytic
function on D with radial limits of modulus one almost everywhere. Perhaps the most
important operator acting on this space is S, the (forward) shift operator on H 2 defined
by [S(f )](z) = zf (z); the adjoint of S is the backward shift [S(cid:63)(f )](z) = (f (z) − f (0))/z.
In 1949, Beurling [9] proved the following theorem about (closed) nontrivial invariant
subspaces of the shift S, which has implications for the invariant subspaces of the backward
shift operator.
Theorem 3.1 (Beurling's theorem). The nontrivial invariant subspaces under S are
θH 2 = {θh : h ∈ H 2},
where θ is a (nonconstant) inner function.
Thus we see that the invariant subspaces for the adjoint S(cid:63) are Kθ := H 2(cid:9) θH 2. These
subspaces are called model spaces. The following description of Kθ is often helpful and it
is not difficult to prove.
Theorem 3.2. Let θ be inner. Then Kθ = H 2 ∩ θ zH 2.
Let aj < 1 for j = 1, . . . , n and consider KB where B(z) = (cid:81)n
z−aj
1−aj z is a finite
Blaschke product. The reproducing kernel corresponding to the point a ∈ D is defined
and it has the property that (cid:104)f, ga(cid:105) = f (a) for all f ∈ H 2. As a
by ga(z) =
consequence we see that (cid:104)Bh, gaj(cid:105) = B(aj)h(aj) = 0 for all h ∈ H 2. So if B is a Blaschke
products with zeros a1, . . . , an, then gaj ∈ KB for j = 1, 2, . . . , n. In fact, if the points aj
1This figure was created by Trung Tran. We thank him for his permission to use the figure in this paper.
1 − az
j=1
1
8
KELLY BICKEL†, AND PAMELA GORKIN‡
are distinct, KB = span{gaj : j = 1, . . . , n} and the reproducing kernels ga1, . . . , gan will
be linearly independent.
It is not really essential that the points be distinct, but certain adjustments must be
made if they are not. However, the representations for our matrices will not change; we
refer the reader to [31].
θ an inner function, we define Sθ : Kθ → Kθ by
The operators that we are interested in here are compressions of the shift operator: For
Sθ(f ) = Pθ(S(f ))
where Pθ is the orthogonal projection from H 2 onto Kθ. In this section, we are particularly
interested in the case in which θ = B is a finite Blaschke product.
In this case (and
precisely in this case) KB is finite dimensional and with the appropriate choice of an
orthonormal basis, we can analyze this operator. The basis that we choose is obtained
from applying the Gram-Schmidt process to the basis we obtained from the reproducing
kernels. In finite dimensions, this basis is called the Takenaka-Malmquist basis: Letting
ba(z) = z−a
1−az , we take it to be the following ordered basis:
(cid:32)(cid:112)1 − a12
1 − a1z
n(cid:89)
j=2
(cid:112)1 − a22
1 − a2z
n(cid:89)
j=3
(cid:112)1 − an−12
1 − an−1z
(cid:112)1 − an2
1 − anz
(cid:33)
.
baj ,
baj , . . . ,
ban,
We have chosen this ordered basis to yield an upper triangular matrix for SB. For example,
for two zeros a and b we obtain
(cid:20) a (cid:112)1 − a2(cid:112)1 − b2
(cid:21)
.
A =
0
b
length(cid:112)1 − a2(cid:112)1 − b2. When a = b we see that the numerical range is a circular disk
So A is the matrix representing SB when B has two zeros a and b. By the elliptical range
theorem, the numerical range is an elliptical disk with foci at a and b and minor axis of
with center at a and radius (1 − a2)/2. Thus, as we mentioned earlier, for the 2 × 2
Jordan block A2 that we met in (3), it follows that the numerical range is the closed disk
centered at the origin of radius 1/2. What about the n × n case?
Here things are more complicated, but we can still obtain a matrix representing SB: A
computation shows that the n × n matrix representing SB with respect to the Takenaka-
Malmquist basis is
(cid:112)1 − a12(cid:112)1 − a22
k=2(−ak))(cid:112)1 − a12(cid:112)1 − an2
((cid:81)n−1
k=3(−ak))(cid:112)1 − a22(cid:112)1 − an2
((cid:81)n−1
(7)
A =
.
a1
0
. . .
0
. . .
. . .
. . .
0
a2
. . .
0
. . .
an
9
Note that for each λ ∈ T, by adding only one row and one column, we can put A "inside" a
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
Aij
unitary matrix
k=1(−ak)(cid:1)(cid:112)1 − aj2
λ(cid:0)(cid:81)j−1
k=i+1(−ak)(cid:1)(cid:112)1 − ai2
(cid:0)(cid:81)n
λ(cid:81)n
k=1(−ak)
Example 3.3. Let B(z) = zn. Then
ij =
U λ
if 1 ≤ i, j ≤ n,
if i = n + 1 and 1 ≤ j ≤ n,
if j = n + 1 and 1 ≤ i ≤ n,
if i = j = n + 1.
and SB is represented (with respect to the Takenaka-Malmquist basis) by
KB = span(1, z, z2, . . . , zn−1)
0
···
0
0
0
0
···
0
1
.
.
1
0
···
0
0···
1···
···
0
0
0
···
0
0
0
1
0
1
0
···
0
0
√
0···
1···
···
0
0
(cid:18) A
0
0 0
0
0 0
··· 1 0
0 1
0
0 0
0
(cid:19)
U =
DA
DA(cid:63) −A(cid:63)
An example of a unitary 1-dilation of this matrix (the matrix with entries from SB in bold) is
3.1. Unitary dilations. Let A be a matrix with (cid:107)A(cid:107) ≤ 1. Then, following Halmos, we may
consider DA =
1 − A(cid:63)A. Halmos noted that
1 − AA(cid:63) and DA(cid:63) =
√
is a unitary dilation of A to a space twice as large as the original and he posed the following
question [43]: Is
(cid:92){W (U ) : U a unitary dilation of A}?
W (A) =
We note that Halmos considered operators on a Hilbert space H and posed this question for
operators on (possibly) infinite dimensional spaces. In particular, the closures of the numerical
ranges are required in this general setting. For finite Blaschke products, all numerical ranges in
question will be closed.
In this paper, we focus on unitary dilations of a compressed shift operator SB. We begin with
the case in which B is a finite Blaschke product (and we will see later that a similar result holds
for Sθ when θ is a general inner function). For these, we know that the unitary dilations can be
parametrized as a family {Uλ} for each λ ∈ T, where
, where we have added one row and one column.
(In this representation, each ∗λ can be determined once we have SB. This will be clear for SB
when B is finite and we discuss this parametrization later briefly for arbitrary inner functions.)
In fact, up to unitary equivalence, these are all the unitary 1-dilations of SB. Note also that
it makes sense that there is a unitary 1-dilation: Looking at the Halmos dilation, recalling that
BSB)1/2 = rank(I −
rank(I − S(cid:63)
B)1/2, we might expect that we need only add one row and one column to get to the
SBS(cid:63)
unitary dilation. To investigate this further, we mention some important connections between
B) and that this implies rank(I − S(cid:63)
BSB) = 1 = rank(I − SBS(cid:63)
(cid:20) SB ∗λ
(cid:21)
Uλ =
∗λ
∗λ
10
KELLY BICKEL†, AND PAMELA GORKIN‡
the Blaschke product B and the unitary dilations of SB. Before we do so, however, we note that
one half of Halmos's conjecture is easy: We show that
W (SB) ⊆(cid:92){W (Uλ) : λ ∈ T}.
(cid:20)x
(cid:21)
To see this, let V = [In, 0] be n × (n + 1) and λ ∈ T. Then SB = V UλV t and we see that for
, so (cid:107)V tx(cid:107) = 1. Suppose β ∈ W (SB). Then there exists a
x with (cid:107)x(cid:107) = 1 we have V tx =
unit vector x such that β = (cid:104)SBx, x(cid:105). Thus,
0
β = (cid:104)SBx, x(cid:105) = (cid:104)V UλV tx, x(cid:105) = (cid:104)UλV tx, V tx(cid:105).
Consequently, β ∈ W (Uλ). Since this holds for each λ ∈ T, we see that β lies in ∩λ∈TW (Uλ).
Thus containment holds because SB is a compression of Uλ and the same argument works in
greater generality. From this it is not difficult to see that it is the other containment that is
interesting.
Because we know that the numerical range of a unitary matrix is the convex hull of its eigenval-
ues, we first consider the eigenvalues of the unitary 1-dilation of SB where B is a finite Blaschke
product. In this case, it can be shown that the eigenvalues of Uλ are the values B(z) := zB(z)
maps to λ [31, 33]. Recalling that a finite Blaschke product is continuous on an open subset of
C containing the closed unit disk, maps the unit circle to itself, the open unit disk to itself, and
the complement of the closed unit disk to itself, we see that for λ ∈ T the only possible solutions
to this equation will lie on the unit circle. It is also well known (see, for example, [24]) that
the argument of a finite Blaschke product increases on the unit circle. As a consequence, the
solutions to B = λ will be distinct. If B has degree n, there will precisely n + 1 distinct solutions
to B = λ. Thus W (Uλ) is the convex hull of these n + 1 distinct points.
We are now ready to put all this together using a result of Gau and Wu [32] who studied
the class Sn of compressions of the shift to an n-dimensional space: These are operators that
have no eigenvalues of modulus 1, are contractions (completely non-unitary contractions) with
rank(I − T (cid:63)T ) = rank(I − T T (cid:63)) = 1 and they are compressions of the shift operator with finite
Blaschke product symbol:
(8)
where PB is defined by
and P− the orthogonal projection for L2 onto L2 (cid:9) H 2. Their result is the following (see Fig-
ure 3.1):
Theorem 3.4. [32] For T ∈ Sn and any point λ ∈ T, there is an (n + 1)-gon inscribed in T that
circumscribes the boundary of W (T ) and has λ as a vertex.
SB(f ) = PB(S(f )) where f ∈ KB, PB : H 2 → KB,
PB(g) = BP−(Bg) = B(I − P+)(Bg)
As a consequence of this result, Gau and Wu were able to prove the following:
Theorem 3.5. [32] Let B be a finite Blaschke product. Then
(cid:92){W (U ) : U a unitary 1-dilation of SB}.
W (SB) =
The authors note that, for compressions of the shift operator, this is the "most economical"
intersection; that is, we need only consider dilations of our operators to a space one dimension
larger. For general operators, Choi and Li answered Halmos's question in 2001:
Theorem 3.6 (General theorem, [16]). Let T be a contraction on a Hilbert space H. Then
(cid:92){W (U ) : U a unitary dilation of T on H ⊕ H}.
W (T ) =
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
11
Figure 3.1. Polygons intersecting to yield numerical range
In addition to answering Halmos's theorem for these operators, Gau and Wu's result has a
very beautiful geometric consequence that we investigate in the next subsection.
3.2. The Poncelet property. In this section, we discuss the connection to a famous theorem
from projective geometry known as Poncelet's theorem. Poncelet was born in Metz, France in
1788. He joined Napoleon's army as it was approaching Russia. On October 19 Napoleon ordered
the army to withdraw. The Russians then attacked the retreating French army and sources say
that Poncelet was left for dead on the battlefield. Poncelet was held as a prisoner in Saratov and
it was during this time that he discovered the following theorem, now bearing his name.
Theorem 3.7. (Poncelet's Theorem, 1813, ellipse version) Given one ellipse contained entirely
inside another, if there exists one circuminscribed (simultaneously inscribed in the outer and
circumscribed around the inner) n-gon, then every point on the boundary of the outer ellipse is
the vertex of some circuminscribed n-gon.
Poncelet's theorem says that if you shoot a ball starting at a point on the exterior ellipse,
shooting tangent to the smaller ellipse, and the path closes in n steps, then no matter where
you begin the path will close in n steps. There are now many proofs of Poncelet's theorem -- of
course there is one due to Poncelet [58], one due to Griffiths and Harris [38], and in 2015 for the
bicentennial of Poncelet's theorem, a proof due to Halbeisen and Hungerbühler appeared [41].
Though this version of Poncelet's theorem is about two ellipses, using an affine transformation
does not change the Poncelet nature of an inner ellipse. Thus we may assume that the outer
ellipse is the unit circle. Returning to Gau and Wu's theorem, we see that it says that the
boundary of the numerical range of a compression of the shift operator is a Poncelet curve; that
is, it is the case that given any point λ on the unit circle we can find a polygon with all vertices
on the unit circle that circumscribes the bounding curve. All of the polygons will have the same
number of sides. While the curves have this beautiful property, they are not usually ellipses and
we therefore call them Poncelet curves. For further study in this regard, we note that work of
Mirman [55] looks at this same property as well as packages of Poncelet curves; that is, what
if instead of connecting successive points, we connect every other point? What happens if we
connect every third point? Examples of higher-degree cases in which the numerical range is
elliptical can be found in [25, 29, 37].
Example 3.8. Consider the special case in which B(z) = zn, and the matrix representing SB is
the n× n Jordan block. Then the unitary 1-dilations are parametrized by the unit circle and for
each λ ∈ T the numerical range of Uλ is the convex hull of the points for which B(z) = zB(z) =
zn+1 = λ. By Theorem 3.4, the intersection of all W (Uλ) over λ ∈ T is the numerical range of
SB. It is now easy to see that the numerical range must be bounded by a circle. Looking at
12
KELLY BICKEL†, AND PAMELA GORKIN‡
the points on the unit circle that give rise to the real eigenvalues, we see that the radius of the
bounding circle must be cos(π/(n + 1)). (See also [40, 69] for this and related results.)
Thus, we have the following result.
Theorem 3.9. The numerical range of the n × n Jordan block is a circular disk of radius
cos(π/(n + 1)).
We note that this circular disk of radius cos(π/(n + 1)) is inscribed in the convex (n + 1)-gon
with vertices equally spaced on the unit circle; in other words, the boundary is a Poncelet circle.
For Poncelet ellipses inscribed in triangles, we refer to the paper [24]. For more on a Blaschke
product perspective of Poncelet's theorem, see also [25, 26].
4. Extensions: General inner functions and other defect indices
of the shift is defined as in (8).
Now we consider infinite Blaschke products as well as general inner functions. The compression
For an infinite Blaschke product, we require that the zeros an ∈ D satisfy the Blaschke
n=1(1 − an) < ∞ We recall ([8]) that if an operator T is a completely nonunitary
condition(cid:80)∞
contraction with a unitary 1-dilation, then
(1) every eigenvalue of T is in the interior of W (T );
(2) W (T ) has no corners in D.
4.1. Compressions of the shift with inner functions as symbol. To obtain the matrix
representation for our operators with inner function θ as symbol, we consider two orthogonal
decompositions of Kθ: The first decomposition will be
(9)
while the second will be
(10)
M1 = C(S(cid:63)θ) = {x(θ(z) − θ(0))/z : x ∈ C} and N1 = Kθ (cid:9) M1
M2 = C(θ θ(0) − 1) and N2 = Kθ (cid:9) M2.
Computations then show that
Sθ(xS(cid:63)θ + w) = x((θ θ(0) − 1)θ(0) + Sw
for x ∈ C and w ∈ N1. Thus, we get this matrix representation for unitary 1-dilations on
K = Kθ ⊕ C:
(cid:20)λ 0
(cid:21)
0 S
Sθ =
and Uαβ =
β(cid:112)1 − λ2
λ
0
0 α(cid:112)1 − λ2
S
0
0
−αβλ
.
Here α, β have modulus 1 and λ < 1. If θ(0) = 0, then λ = 0. It appears that there are several
free variables, but up to unitary equivalence there is only one free parameter and that is the
value of αβ. Thus, the unitary dilations may be parametrized by γ ∈ T. We have the celebrated
theorem of D. Clark.
Theorem 4.1. [18] If θ is inner and θ(0) = 0, then all unitary 1-dilations of Sθ are equivalent
to rank-1 perturbations of Szθ.
For compressions of the shift with a Blaschke product as symbol, we obtain the following:
Theorem 4.2. [15] Let B be an infinite Blaschke product. Then the closure of the numerical
range of SB satisfies
(cid:92)
γ∈T
W (SB) =
W (Uγ),
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
13
Figure 4.1. An approximation of an infinite Blaschke product with one singularity
where the Uγ are the unitary 1-dilations of SB (or, equivalently, the rank-1 Clark perturbations
of S B).
For some functions, we get an infinite version of Poncelet's theorem, see Figure 4.1.
To extend this theorem to arbitrary inner functions, the following well-known result of Frost-
man is useful.
Theorem 4.3 (Frostman's Theorem). Let I be an inner function. Let a ∈ D and ϕa(z) = z−a
1−az .
Then ϕa ◦ I is a Blaschke product for almost all a ∈ D.
Every inner function is, therefore, a uniform limit of Blaschke products. So, if θ is an arbitrary
inner function, we may find a sequence (Bn) of Blaschke products convering uniformly to θ. Since
Pθf = θP−(θf ) for f ∈ H 2,
where P− : L2(T) → L2(T) (cid:9) H 2 is the orthogonal projection, (cid:107)Bn − θ(cid:107)∞ → 0 implies that
(cid:107)PBn − Pθ(cid:107) → 0. We may use this to obtain W (Sθ) from W (SBn) where Bn is a Blaschke
product. For more details, see [15].
Combining results in [8] with Frostman's theorem tells us that Theorem 4.2 holds for arbitrary
inner functions.
4.2. Higher defect index. Thus far, our operators have defect index equal to 1. However, the
situation for more general defect index and an operator on a complex separable Hilbert space H
was studied by Bercovici and Timotin [8]. They considered n-dilations of contractions, which are
unitary dilations of T that act on H ⊕ Cn. In general, a contraction can be written as a direct
sum of a unitary operator and a completely nonunitary contraction. Since we understand the
numerical range of the unitary piece of the operator, most of the work in a proof focuses on the
completely nonunitary part of the contraction. To state the next result, we let DT = (I−T (cid:63)T )1/2
denote the defect operator and DT = DTH denote the defect space.
Theorem 4.4. Let T be a contraction with dim DT = dim DT (cid:63) = n < ∞. Then
(cid:92){W (U ) : U a unitary n − dilation of T}.
W (T ) =
Once again, we see that we can use the "most economical" dilations of T . Bercovici and
Timotin also show that if (cid:96) is a support line for the closure of the numerical range of T , then
there is a unitary n-dilation of T for which (cid:96) is a support line for W (U ). In some sense, then,
the geometry extends to this situation as well.
14
KELLY BICKEL†, AND PAMELA GORKIN‡
5. Compressed shifts on the bidisk
5.1. Two-variable Setup. While the earlier sections discuss many results on D, there is another
direction to pursue, namely compressions of shifts in several variables. To that end, let D2 denote
the unit bidisk and T2 its distinguished boundary
D2 = {(z1, z2) : z1,z2 < 1} and T2 = {(τ1, τ2) : τ1,τ2 = 1}.
In this setting, many one-variable objects generalize easily.
consists of functions of the form
Indeed, the Hardy space H 2(D2)
∞(cid:88)
∞(cid:88)
f (z) =
am,nzm
1 zn
2 , where
am,n2 < ∞.
m,n=0
m,n=0
Then there are two natural shift operators, Sz1 and Sz2, on H 2(D2), defined by [Szj (f )](z) =
zjf (z) for j = 1, 2. Similarly, a function Θ is inner if Θ ∈ Hol(D2) and limr(cid:37)1 Θ(rτ ) = 1 for
a.e. τ ∈ T2. Then ΘH 2(D2) is a shift-invariant subspace of the two-variable Hardy space and in
analogy with the one-variable setting, one can define two-variable model spaces as
KΘ = H 2(D2) (cid:9) ΘH 2(D2)
and the associated compressed shifts
Θ = PΘSz1KΘ
S1
and
Θ = PΘSz2KΘ,
S2
where PΘ is the orthogonal projection onto KΘ. The study of such compressed shifts is aided by
the existence of nice decompositions of KΘ into shift-invariant subspaces. Indeed, Ball, Sadosky,
and Vinnikov [7] showed that there are subspaces S1 and S2 such that
and Sz1S1 ⊆ S1, Sz2S2 ⊆ S2.
(11)
Such decompositions are called Agler decompositions and were introduced by J. Agler in a dif-
ferent form in [2]. For more information about Agler decompositions and their properties see
[3, 10, 11, 19, 49, 68] and the references therein. Then using basic properties of multiplication
operators, one can show (as in [12]):
Lemma 5.1. Assume KΘ = S1 ⊕ S2 gives an Agler decomposition as in (11). Then
KΘ = S1 ⊕ S2
a. If S1 (cid:54)= {0}, then W ( S1
b. If S1 = {0}, then W ( S1
Θ) = W (Sz1S1) = D.
ΘS1) = ∅.
Typically S1 (cid:54)= {0} and so Lemma 5.1 says that most compressed shifts have maximal numer-
ical ranges. This renders the standard numerical range questions trivial. To obtain interesting
questions, one can further compress the multiplication operators and define
S1
Θ := PS2
ΘS2 = PS2Sz1S2,
S1
where S2 is any subspace arising from an Agler decomposition of KΘ as in (11). The numerical
Θ have quite interesting properties, which will be discussed later. In what follows,
ranges of such S1
we always assume S2 (cid:54)= {0}.
5.2. Rational Inner Functions. First consider the two-variable analogues of finite Blaschke
products, called rational inner functions. Although more complicated than finite Blaschke prod-
ucts, rational inner functions still have fairly nice structures. Indeed, as shown in [4, 62], every
rational inner function Θ with deg Θ = (m, n) can be written as
Θ = λ
, where p(z) = zm
1 zn
2 p
, 1
¯z2
¯z1
p
p
(cid:16) 1
(cid:17)
,
and λ ∈ T. Here deg Θ = (m, n) means that, after canceling any common factors of the numerator
and denominator, m is the largest power of z1 appearing in Θ and n is the largest power of z2
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
15
appearing in Θ. Furthermore, one can choose p so that it has at most a finite number of zeros
on T2, is nonvanishing on D2 ∪ (D × T) ∪ (T × D), and shares no common factors with p. For
example, up to a unimodular constant, a general degree (1, 1) rational inner function has the
form
Θ(z) =
p(z)
p(z)
=
az1z2 + bz2 + cz1 + d
a + bz1 + cz2 + dz1z2
,
where p = a + bz1 + cz2 + dz1z2 satisfies the stated conditions on its zero set and shares no
common factors with p. For rational inner functions Θ, the associated model spaces KΘ have
particularly nice Agler decompositions. For example, the following result describes properties of
S2 from (11):
Lemma 5.2. Let Θ = p
dim H = m and if g ∈ H, then g = q
p be rational inner with degree (m, n) and let H = S2 (cid:9) Sz2S2. Then
p where q is a polynomial with deg q ≤ (m − 1, n).
The complete result appears in [12], but related results and ideas appeared earlier in [11, 68].
Rational inner functions also have close connections to one-variable finite Blaschke products. For
Θ a rational inner function with deg Θ = (m, n), define the exceptional set
EΘ = {τ ∈ T : p(τ1, τ ) = 0 for some τ1 ∈ T}.
Stating the main results requires some notation. Let H 2
Then if τ ∈ T \ EΘ, the function θτ (z) := Θ(z, τ ) is a finite Blaschke product with deg θτ = m.
In what follows, we let Kθτ denote the one-variable model space associated to each θτ .
5.3. Results. Let us restrict to rational inner Θ and consider the compressed shift S1
numerical range.
Θ and its
2 (D) denote the one-variable Hardy
2 (D)m := ⊕m
2 (D) denote the space of vector-valued functions
2 (D). If M is a bounded, m × m matrix-valued function, let TM
space with variable z2 and H 2
−→
f = (f1, . . . , fm) with fj ∈ H 2
denote the matrix-valued Toeplitz operator defined by
(12)
Then the following results and their corollaries appear in [12]. The proofs rely heavily on the
structure of Agler decompositions, as given in Lemma 5.2 and other results.
p be rational inner of degree (m, n) and let S2 be as in (11). There is
Theorem 5.3. Let Θ = p
an m × m matrix-valued function MΘ with entries continuous on D and rational in z2 such that
Θ is unitarily equivalent to TMΘ, the matrix-valued Toeplitz operator with symbol MΘ, defined
S1
as in (12).
−→
f = PH 2
2 (D)m(M
−→
f ).
j=1H 2
TM
The symbol MΘ generalizes the classical matrix associated to a compressed shift. Indeed, if
Θ(z) = B(z1) is a one-variable finite Blaschke product, then
MΘ = the constant matrix of SB on KB given in (7).
(cid:16) 2z1z2−z1−z2
2−z1−z2
(cid:17)(cid:16) 3z1z2−z1−2z2
3−2z1−z2
(cid:17)
In some more complicated situations, we can still compute MΘ. For example, let Θ(z) =
. Then an application of [12, Theorem 4.4] gives
MΘ(z2) =
1
2 − ¯z2
−√
6(1 − ¯z2)2
(2 − ¯z2)(3 − ¯z2)
0
2
3 − ¯z2
.
16
KELLY BICKEL†, AND PAMELA GORKIN‡
As in this example, Theorem 4.4 from [12] gives lower triangular matrices, rather than upper
triangular ones like in (7), because the proof orders the basis elements differently than in the
classical one-variable setup. This relationship between S1
Θ and the Toeplitz operator with symbol
MΘ gives information about the numerical range. Specifically,
Theorem 5.4. Let Θ = p
be as in Theorem 5.3. Then W (S1
p be rational inner of degree (m, n), let S2 be as in (11), and let MΘ
Θ) = the convex hull of
τ∈T W (MΘ(τ ))
(cid:17)
.
(cid:16)(cid:83)
(cid:17)(cid:16) 3z1z2−z1−2z2
(cid:17)
(cid:16) 2z1z2−z1−z2
2−z1−z2
For example, this means that if Θ(z) =
, then by the elliptical
Θ) is the convex hull of a union of elliptical disks. Moreover,
Θ) to the one-variable setting as follows.
3−2z1−z2
range theorem, the closure of W (S1
Theorem 5.4 allows us to connect the study of W (S1
Recall that the notation in the following theorem was defined after Lemma 5.2.
Theorem 5.5. Let Θ = p
W (S1
Θ) = the closure of the convex hull of
(cid:16) (cid:91)
W (Sθτ on Kθτ )
p be rational inner of degree (m, n) and let S2 be as in (11). Then
(cid:17)
.
τ∈T\EΘ
Thus, W (S1
the shift associated to Θ. As this formula no longer involves S2, it implies the following:
Corollary 5.6. Let Θ = p
W (PS2Sz1S2) = W (P(cid:101)S2
Θ) can also be obtained using the numerical ranges of one-variable compressions of
p be rational inner of degree (m, n) and let S2, S2 be as in (11). Then
Sz1(cid:101)S2
).
This is important because, in general, Agler decompositions are not unique and one would
Θ) would depend heavily on the choice of S2. Finally, one can use the connection
Θ has maximal numerical radius
expect that W (S1
to one-variable compressions of the shift to characterize when S1
Θ)}. Then:
w(S1
Θ), where the numerical radius is sup{z : z ∈ W (S1
p be rational inner of degree (m, n) and let S2 be as in (11). Then
Corollary 5.7. Let Θ = p
(cid:1) = 1 if and only if Θ has a singularity on T2.
w(cid:0)S1
Θ
This indicates, for example, that many ∂W (S1
Θ) cannot satisfy a Poncelet property because
they touch T.
To say more about the geometry of W (S1
Θ), we restrict to very simple rational inner functions,
p with p(z) = a − z1 + cz2 with a, c > 0 and p(1,−1) = 0. Then
namely Θ = θ2
deg Θ = (2, 2) and so each MΘ(τ ) is a 2 × 2 matrix. Indeed, each W (MΘ(τ )) is a circular disk
and so the numerical range looks like the convex hull of this:
1 where θ1 = p
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
17
Figure 5.1. A collection of boundaries of W (MΘ(τ )) associated to Θ(z) =
(cid:16) 2z1z2+z1−z2
2−z1+z2
(cid:17)2
.
In this case, taking the convex hull merely fills in the hole in this set. So, the boundary of
Θ) is precisely the outer boundary of the family of disks W (MΘ(τ )). This should bring to
W (S1
mind envelopes. Indeed, in this case, one can use the discriminant envelope from Definition 2.6
to obtain formulas for the boundaries of these numerical ranges:
Theorem 5.8. For Θ = θ2
1 given above, the boundary of W (S1
for t ∈ [0, 2π), where
Θ) is the curve E = (x(t), y(t))
x(t) =
y(t) =
a + c cos t
a + c
c sin t
a + c
+
ac(1 − cos t)
+
(a + c)2
ac(1 − cos t)
(cid:18)
(a + c)2
t − arcsin
sin
cos
t − arcsin
(cid:18)
(cid:19)(cid:19)
;
(cid:18) a
(cid:18) a
a + c
sin t
(cid:19)(cid:19)
sin t
.
a + c
For more details and additional geometric results about W (S1
Θ), see [12].
6. Open Questions
In [21] M. Crouzeix stated the following conjecture:
Conjecture (2004): There exists a constant C such that for any polynomial p ∈ C[z] and A an
n × n matrix, the inequality holds:
(cid:21)
(cid:20) 0 1
0 0
. Then
The best constant should be C = 2.
(cid:107)p(A)(cid:107) ≤ C maxp(z)z∈W (A).
First let us see why the constant must be at least 2. Let p(z) = z and A =
(cid:107)p(A)(cid:107) = (cid:107)A(cid:107) = 1 and maxp(z)z∈W (A) = max{z:z≤1/2} z = 1/2. So C ≥ 2.
Even though it is unclear that such a constant exists let alone is equal to 2, there is reason
to believe the conjecture is true. Crouzeix showed that in fact such a constant does exist and is
between 2 and 11.08. Okubo and Ando [57] showed that the conjecture is true if the numerical
range is a disk. Badea, Crouzeix and Delyon presented several approaches to this problem and
others in [6]. Recently, Glader, Kurula and Lindström [35] considered the problem for tridiagonal
3× 3 matrices with constant diagonal; this includes 3× 3 matrices with elliptical numerical range
and one eigenvalue at the center of the ellipse. Choi [17] showed the conjecture holds for 3 × 3
matrices that are "nearly" Jordan blocks. And recently, Crouzeix and Palencia [22] showed that
the best constant is between 2 and 1 +
√
2.
18
KELLY BICKEL†, AND PAMELA GORKIN‡
Crouzeix and Palencia's proof relies on a crucial lemma, which we reproduce below. In what
follows, we let Ω be a bounded open convex set with smooth boundary and let A(Ω) denote the
algebra of functions continuous on Ω and holomorphic on Ω.
Lemma 6.1 (Crouzeix and Palencia). Let T be a bounded Hilbert space operator and let Ω be
a bounded open set containing the spectrum of T . Suppose that for each f ∈ A(Ω) there exists
g ∈ A(Ω) such that
(cid:107)g(cid:107)Ω ≤ (cid:107)f(cid:107)Ω and (cid:107)f (T ) + g(T )(cid:63)(cid:107) ≤ 2(cid:107)f(cid:107)Ω.
Then
(cid:107)f (T )(cid:107) ≤ (1 +
√
2)(cid:107)f(cid:107)Ω, f ∈ A(Ω).
In proving their theorem, Crouzeix and Palencia apply the lemma with
dζ, z ∈ Ω.
g = Cf where (Cf )(z) =
1
2πi
f (ζ)
ζ − z
∂Ω
(cid:90)
That is, g is the Cauchy transform of f.
√
Ransford and Schwenninger [61] give a short proof of this and show that in this lemma, the
constant (1 +
2) is sharp. However, as they point out, this is not a counterexample to the
theorem; it just shows that the theorem will not be established "merely by adjusting the proof"
of the lemma. Taking g to be the Cauchy transform of f and considering the map from f to g,
we see that this is an antilinear map and it maps 1 to 1; we say that it is antilinear and unital.
The example appearing in [61] is antilinear, but sends 1 to −1. The authors of [61] suggest
considering the following question, as a positive answer to this would establish the Crouzeix
conjecture.
Question. [61] Let T be a bounded Hilbert space operator and let Ω be a bounded open set
containing the spectrum of T . Suppose that there exists a unital antilinear map α : A(Ω) → A(Ω)
such that for all f ∈ A(Ω)
(cid:107)α(f )(cid:107)Ω ≤ (cid:107)f(cid:107)Ω and (cid:107)f (T ) + (α(f ))(T )(cid:63)(cid:107) ≤ 2(cid:107)f(cid:107)Ω.
Does it follow that
(cid:107)f (T )(cid:107) ≤ 2(cid:107)f(cid:107)Ω?
References
[1] J. Agler, Geometric and topological properties of the numerical range. Indiana Univ. Math. J. 31
(1982), no. 6, 767 -- 777. 6
[2] J. Agler, On the representation of certain holomorphic functions defined on a polydisc. In Topics in
operator theory: Ernst D. Hellinger memorial volume, volume 48 of Oper. Theory Adv. Appl., pages
47 -- 66. Birkhäuser Verlag, Basel, 1990. 14
[3] J. Agler, J.E. McCarthy, Pick interpolation and Hilbert function spaces. Graduate Studies in Math-
ematics, 44. American Mathematical Society, Providence, RI, 2002. 14
[4] J. Agler, J.E. McCarthy, M. Stankus, Toral algebraic sets and function theory on polydisks. J. Geom.
Anal. 16 (2006), no. 4, 551 -- 562. 14
[5] O. Axelsson, H. Lu, and B. Polman, On the numerical radius of matrices and its application to
iterative solution methods. Special Issue: The numerical range and numerical radius. Linear and
Multilinear Algebra, 37 (1994), no. 1-3, 225 -- 238. 1
[6] C. Badea, M. Crouzeix, B. Delyon, Convex domains and K-spectral sets, Math. Z., 252 (2006),
345 -- 365. 17
[7] J.A. Ball, C. Sadosky, and V. Vinnikov, Scattering systems with several evolutions and multidimen-
sional input/state/output systems. Integral Equations Operator Theory, 52 (2005), 323 -- 393. 14
[8] H. Bercovici and D. Timotin, The numerical range of a contraction with finite defect numbers. J.
Math. Anal. Appl. 417 (2014), no. 1, 42 -- 56. 2, 12, 13
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
19
(2001), no. 2, 435 -- 447. 10
3247 -- 3257. 3, 17
12
767 -- 791. 14
[9] A. Beurling, On two problems concerning linear transformations in Hilbert spaces, Acta Math., 81
(1949), 239 -- 255. 7
2, 233 -- 257. 14
[10] K. Bickel, Fundamental Agler decompositions. Integral Equations Operator Theory, 74 (2012), no.
[11] K. Bickel and G. Knese, Inner functions on the bidisk and associated Hilbert spaces. J. Funct. Anal.
265 (2013) no. 11, 2753 -- 2790. 14, 15
[12] K. Bickel and P. Gorkin. Compressions of the shift on the bidisk and their numerical ranges. J.
Operator Theory 79 (2018), no. 1, 225 -- 265. 2, 14, 15, 16, 17
[13] K. Bickel, P. Gorkin, and T. Trung. Applications of Envelopes. In preparation. 6, 7
[14] J.W. Bruce and P.J. Giblin, What is an envelope? Math. Gaz. 65 (1981), no. 433, 186 -- 192. 4
[15] I. Chalendar, P. Gorkin, J.R. Partington, Numerical ranges of restricted shifts and unitary dilations.
Oper. Matrices 3 (2009), no. 2, 271 -- 281. 2, 12, 13
[16] M.-D. Choi, C.-K. Li, Constrained unitary dilations and numerical ranges. J. Operator Theory 46
[17] D. Choi, A proof of Crouzeix's conjecture for a class of matrices, Linear Algebra Appl. 438 (2013),
[18] D. N. Clark, One dimensional perturbations of restricted shifts. J. Analyse Math. 25 (1972), 169 -- 191.
[19] B.J. Cole and J. Wermer, Andô's theorem and sums of squares. Indiana Math. J. 48 (1999), no. 3,
[20] R. Courant, Differential and integral calculus. Vol. II. Translated from the German by E. J. McShane.
Reprint of the 1936 original. Wiley Classics Library. A Wiley-Interscience Publication, John Wiley
& Sons, Inc., New York, 1988. 4
[21] M. Crouzeix, Bounds for analytical functions of matrices, Integral Equations Operator Theory, 48
[22] M. Crouzeix, C. Palencia, The numerical range is a (1 +
2)-spectral set, SIAM J. Matrix Anal.
(2004), 461 -- 477. 2, 17
Appl., 38 (2017), 649 -- 655. 3, 17
no. 2, 668 -- 690. 3
(2002), no. 9, 785 -- 795. 10, 12
[23] M. Crouzeix, Numerical range and functional calculus in Hilbert space. J. Funct. Anal. 244 (2007),
[24] U. Daepp, P. Gorkin, R. Mortini, Ellipses and finite Blaschke products, Amer. Math. Monthly 109
[25] U. Daepp, P. Gorkin, A. Shaffer, K Voss, Möbius transformations and Blaschke products: the geo-
metric connection. Linear Algebra Appl. 516 (2017), 186 -- 211. 11, 12
[26] U. Daepp, P. Gorkin, A. Shaffer, K Voss, Finding Ellipses: What Blaschke Products, Poncelet's
Theorem, and the Numerical Range Know about Each Other, The Carus Mathematical Monographs,
AMS/MAA press, 34, 2018. 12
[27] W. F. Donoghue Jr., On the numerical range of a bounded operator. Michigan Math. J. 4 (1957)
√
[28] M. Eiermann, Fields of values and iterative methods. Linear Algebra Appl. 180 (1993), 167 -- 197. 1
[29] M. Fujimura, Inscribed ellipses and Blaschke products, Comput. Methods Funct. Theory 13 (2013),
[30] S.R. Garcia, W.T. Ross. Model spaces: a survey. Invariant subspaces of the shift operator, 197 -- 245,
Contemp. Math., 638, Centre Rech. Math. Proc., Amer. Math. Soc., Providence, RI, 2015. 2
[31] H.-L. Gau, Y.P. Wu, Numerical range and Poncelet property. Taiwanese J. Math. 7 (2003), no. 2,
[32] H.-L. Gau, Y.P. Wu, Numerical range of S(φ). Linear Multilinear Algebra. 45 (1998), no. 1, 49 -- 73.
[33] H.-L. Gau, Y.P. Wu, Numerical range circumscribed by two polygons. Linear Algebra Appl. 382
261 -- 263. 2, 6
no. 4, 557 -- 573. 11
173 -- 193. 2, 8, 10
2, 10
(2004), 155 -- 170. 10
versatile tool in the theory of quantum information. J. Math. Phys. 51 (2010), no. 10. 1
[34] P. Gawron, Z. Puchała, J.A. Miszczak, Ł. Skowronek, K. Życzkowski. Restricted numerical range: a
[35] C. Glader, M. Kurula, M. Lindström, Crouzeix's conjecture holds for tridiagonal 3× 3 matrices with
elliptic numerical range centered at an eigenvalue. SIAM J. Matrix Anal. Appl. 39 (2018), no. 1,
346 -- 364. 3, 17
20
KELLY BICKEL†, AND PAMELA GORKIN‡
[36] M. Goldberg and E. Tadmor, On the numerical radius and its applications. Linear Algebra Appl. 42
[37] P. Gorkin, N. Wagner, Ellipses and compositions of finite Blaschke products. J. Math. Anal. Appl.
[38] P. Griffiths, J. Harris, On Cayley's explicit solution to Poncelet's porism. Enseign. Math. (2) 24
(1982), 263 -- 284. 1
445 (2017), no. 2, 1354 -- 1366. 11
(1978), no. 1-2, 31 -- 40. 11
[39] K.E. Gustafson and D.K.M. Rao, Numerical range. The field of values of linear operators and ma-
trices. Universitext. Springer-Verlag, New York, 1997. 1, 4
[40] U. Haagerup, P. de la Harpe, The numerical radius of a nilpotent operator on a Hilbert space, Proc.
Amer. Math. Soc. 115 (1992) 371 -- 379. 12
[41] L. Halbeisen, N. Hungerbühler, A simple proof of Poncelet's theorem (on the occasion of its bicen-
tennial). Amer. Math. Monthly 122 (2015), no. 6, 537 -- 551. 11
[42] P.R. Halmos, A Hilbert Space Problem Book, 2nd ed., Springer-Verlag, New York, 1982. 1, 3
[43] P.R. Halmos, Numerical ranges and normal dilations, Acta Sci. Math. (Szeged) 25(1964), 1 -- 5. 9
[44] F. Hausdorff, Der Wertvorrat einer Bilinearform, Math. Z. 3 (1919), 314 -- 316. 3
[45] S. Hildebrandt, Über den Numerischen Wertebereich eines Operators, Math. Ann. 163 (1966), 230 --
[46] R. A. Horn, C.R. Johnson, Topics in matrix analysis. Corrected reprint of the 1991 original. Cam-
bridge University Press, Cambridge, 1994. 1, 6
[47] D. Kalman, Solving the ladder problem on the back of an envelope. Math. Mag. 80 (2007), no. 3,
247 1
163 -- 182. 4
[48] R. Kippenhahn, 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. 2, 5, 6
[49] G. Knese, Rational inner functions in the Schur-Agler class of the polydisk. Publ. Mat. 55 (2011),
no. 2, 343 -- 357. 14
[50] D. Kribs, A. Pasieka, M. Laforest, C. Ryan, and M.P. da Silva, Research problems on numerical
ranges in quantum computing. Linear Multilinear Algebra. 57 (2009), no. 5, 491 -- 502. 1
[51] C.-K. Li and Y.-T. Poon, Generalized numerical ranges and quantum error correction. J. Operator
[52] C.-K. Li, A simple proof of the elliptical range theorem. Proc. Amer. Math. Soc. 124 (1996), no. 7,
Theory. 66 (2011), no. 2, 335 -- 351. 1
1985 -- 1986. 6
[53] F. Maclachlan, Long-Run and Short-Run Cost Curves,Famous figures and diagrams in economics.
In Blaug, M., Lloyd, P. eds., Edward Elgar Publishing, (2010). 2
[54] J. Maroulas, P. Psarrakos, and M. Tsatsomeros, Perron-Frobenius type results on the numerical
range. Linear Algebra and Its Applications, 348 (2002), pp. 49 -- 62. 1
[55] B. Mirman, Numerical ranges and Poncelet curves. Linear Algebra Appl. 281 (1998), no. 1-3, 59 -- 85.
[56] F.D. Murnaghan, On the field of values of a square matrix. Proc. Natl. Acad. Sci. 18 (1932), no. 3,
[57] K. Okubo, T. Ando, Constants related to operators of class Cρ. Manuscripta Math. 16 (1975), no.
[58] J.-V. Poncelet, Traité des propriétés projectives des figures. Tome I. (French) [Treatise on the pro-
jective properties of figures. Vol. I] Reprint of the second (1865) edition. Les Grands Classiques
Gauthier-Villars. [Gauthier-Villars Great Classics] Éditions Jacques Gabay, Sceaux, 1995. 11
[59] H. Pottman, M. Peternell, Envelopes -- computational theory and applications. Proceedings of Spring
Conference in Computer Graphics, (2000) Budmerice, Slovakia, pp. 3 -- 23. 2
[60] P. Psarrakos and M. Tsatsomeros, On the stability radius of matrix polynomials. Linear Multilinear
Algebra. 50 (2002), no. 2, 151 -- 165. 1
√
[61] T. Ransford, F. Schwenninger, Remarks on the Crouzeix-Palencia proof that the numerical range is
a (1 +
2)-spectral set. SIAM J. Matrix Anal. Appl. 39 (2018), no. 1, 342 -- 345. 3, 18
[62] W. Rudin, Function theory in polydiscs. W. A. Benjamin, Inc., New York-Amsterdam, 1969. 14
[63] D. Sarason, Sub-Hardy Hilbert Spaces in the Unit Disk. University of Arkansas Lecture Notes, Wiley,
New York, 1994. 2
2, 11
246 -- 248. 6
4, 385 -- 394. 17
NUMERICAL RANGE AND COMPRESSIONS OF THE SHIFT
21
[64] B. Sz.-Nagy, C. Foias, H. Bercovici, L. Kérchy, Harmonic analysis of operators on Hilbert space,
second ed., Universitext, Springer, New York, 2010. 2
[65] J.W. Rutter, Geometry of curves. Chapman & Hall/CRC Mathematics. Chapman & Hall/CRC,
Boca Raton, FL, 2000. 4
[66] M. H. Stone, Linear transformations in Hilbert space and their applications to analysis, Amer. Math.
Colloq. Publ. Vol. 15, Amer. Math. Soc., Providence, R. I., 1932.
[67] O. Toeplitz, Das algebraische Analogon zu einem Satze von Fejér. Math. Z. 2 (1918), 187-197 3
[68] H.J. Woerdeman, A general Christoffel-Darboux type formula. Integral Equations Operator Theory.
[69] P. Y. Wu, A numerical range characterization of Jordan blocks, Linear and Multilinear Algebra, 43
67 (2010), no. 2, 203 -- 213. 14, 15
(1998), pp. 351 -- 361. 12
Kelly Bickel, Department of Mathematics, Bucknell University, 380 Olin Science
Building, Lewisburg, PA 17837, USA.
E-mail address: [email protected]
Pamela Gorkin, Department of Mathematics, Bucknell University, 380 Olin Science
Building, Lewisburg, PA 17837, USA.
E-mail address: [email protected]
|
1204.5062 | 1 | 1204 | 2012-04-23T13:36:09 | Sparsity and spectral properties of dual frames | [
"math.FA"
] | We study sparsity and spectral properties of dual frames of a given finite frame. We show that any finite frame has a dual with no more than $n^2$ non-vanishing entries, where $n$ denotes the ambient dimension, and that for most frames no sparser dual is possible. Moreover, we derive an expression for the exact sparsity level of the sparsest dual for any given finite frame using a generalized notion of spark. We then study the spectral properties of dual frames in terms of singular values of the synthesis operator. We provide a complete characterization for which spectral patterns of dual frames are possible for a fixed frame. For many cases, we provide simple explicit constructions for dual frames with a given spectrum, in particular, if the constraint on the dual is that it be tight. | math.FA | math | Sparsity and spectral properties of dual frames
Felix Krahmer∗, Gitta Kutyniok†, Jakob Lemvig‡
May 22, 2018
Abstract: We study sparsity and spectral properties of dual frames of a given
finite frame. We show that any finite frame has a dual with no more than n2
non-vanishing entries, where n denotes the ambient dimension, and that for most
frames no sparser dual is possible. Moreover, we derive an expression for the
exact sparsity level of the sparsest dual for any given finite frame using a general-
ized notion of spark. We then study the spectral properties of dual frames in terms
of singular values of the synthesis operator. We provide a complete characteriza-
tion for which spectral patterns of dual frames are possible for a fixed frame. For
many cases, we provide simple explicit constructions for dual frames with a given
spectrum, in particular, if the constraint on the dual is that it be tight.
1 Introduction
Redundant representations of finite dimensional data using finite frames have received consider-
able attention over the last years (see [10] for an extensive treatment including many references
concerning various aspects). Finite frames are overcomplete systems in a finite dimensional
Hilbert space. Similar to a basis expansion, frames can be used both to represent a signal in
terms of its inner products with the frame vectors, the frame coefficients, and to reconstruct
the signal from the frame coefficients. As for bases, the frames used for representation and
reconstruction depend on each other, but are not, in general, the same. In contrast to basis ex-
pansions, however, even once one of the two frames is fixed, the other, the dual frame, is not
unique. Hence the quality of a method for redundant representation will depend on both the
frame and the dual frame. There is, however, a standard choice for the dual frame, the so-called
canonical dual, corresponding to the Moore-Penrose pseudo-inverse, whose optimality in many
respects follows directly from its geometric interpretation. This is why most previous work con-
cerning the design of finite frames has focused primarily on the design of single frames rather
∗Georg-August-Universität Göttingen, Institut für Numerische und Angewandte Mathematik, 37083 Göttingen,
Germany, E-mail [email protected]
†Technische Universität Berlin, Institut für Mathematik, 10623 Berlin, Germany, E-mail: kutyniok@math.
2
1
0
2
r
p
A
3
2
]
.
A
F
h
t
a
m
[
1
v
2
6
0
5
.
4
0
2
1
:
v
i
X
r
a
‡Technical University of Denmark, Department of Mathematics, Matematiktorvet 303, 2800 Kgs. Lyngby,
Denmark, E-mail: [email protected]
2010 Mathematics Subject Classification. Primary 42C15; Secondary 15A09, 65F15, 65F20, 65F50
Key words and phrases. dual frames, frame theory, singular values, sparse duals, sparsity, spectrum, tight
tu-berlin.de
frames.
Page 1 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
than dual frame pairs, as it would be necessary to take into account all degrees of freedom in the
problem. In particular, design according to spectral properties [4] and sparsity [6] has recently
received attention.
In this paper, we will take a different viewpoint; we study the set of dual frames for a given
fixed finite frame. The guiding criteria, however, will again be sparsity and spectral properties.
This is motivated by work in the recent papers [8, 16], which ask the following question: If
the frame {φi}m
i=1 is given by the application at hand, e.g., by the way of measuring the data,
which dual for the synthesis process is the best to choose? The precise answer to this question
is, of course, dependent on the application, but universal desirable properties of the dual can,
nonetheless, be recognized. Among such desirable properties are fast and stable reconstruction.
Measures for the stability of the reconstructions resulting from a dual frame are extensively
discussed in [8]. It is shown that the computational properties of the dual frames are directly
linked to spectral properties of the dual frame matrix. In particular, the Frobenius norm and
the spectral norm play an important role in this context. The analysis in [8] focuses on stability
estimates when the frame coefficient vector is corrupted by random or uniform noise. The
papers [13, 14, 17] discuss an alternative scenario of corruptions through erasures. The fact that
the canonical dual is shown to be optimal in many cases also suggests connections to spectral
properties, but we will not explore this connection further in this paper, leaving it to potential
further work.
A fast reconstruction process is closely linked to fast matrix-vector multiplication properties
of the dual frame matrix. Matrices with these properties can be constructed in various ways.
However, many such constructions yield matrices with a specific structure, often linked to the
fast Fourier transform, which can be counteracting the constraint that the matrix should be a
dual to the given frame. The criterion of sparsity, while also connected to fast matrix-vector
multiplication properties, does not lead to such problems, but rather provides a way to ensure
both duality to a given frame and a fast matrix-vector multiply. The first paper on the topic [16]
constructs sparse duals of Gabor frames using (cid:96)1-minimization, and [8] also uses sparsity as a
measure for computational efficiency.
The focus of [8] has been on deriving criteria for comparing computational properties of
dual frames and empirically balancing them to find a good dual frame. This paper is meant to
be complementary to [8] as we analytically study the set of all dual frames of a given frame
with respect to sparsity and spectral properties. We ask how sparse a dual frame can be and
what we can say about the spectrum.
Our paper is organized as follows. In Section 2 we present more formal definitions of frames
and their duals as well as a review of results in frame theory, an overview of the notation we will
be using, and some basic observations. In Section 3 we then consider the sparsity properties of
the set of all dual frames. Finally, Section 4 focuses on the spectral picture of the set of all dual
frames.
2 Setup and basic observations
2.1 Frames and their duals
Throughout this paper, we are considering signals x in a n-dimensional Hilbert space Hn. We
will often identify Hn with Kn, where K denotes either R or C, using a fixed orthonormal basis
of Hn. We write this canonical identification as Hn ∼= Kn. Thus, Kn or Hn will be the signal
Page 2 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
space and Km the coefficient space. Then the precise definitions of frames and their dual frames
are the following.
Definition 2.1. A collection of vectors Φ = (φi)m
two constants 0 < A ≤ B such that
2 ≤ m
∑
i=1 ⊂ Hn is called a frame for Hn if there are
(cid:104)x,φi(cid:105)2 ≤ B(cid:107)x(cid:107)2
2 ,
for all x ∈ Hn.
A(cid:107)x(cid:107)2
(2.1)
i=1
If the frame bounds A and B are equal, the frame (φi)m
i=1 is called tight frame for Hn.
Let Φ = (φi)m
i=1 be a frame for Hn. Since we identify Hn ∼= Kn, we also write φi for the
i=1 ∈ Kn×m as a n× m matrix, and we say that
coefficient (column) vector of φi and Φ = [φi]m
j=1, j(cid:54)=k ∈ K(n−1)×m as
Φ is a frame for Kn. We write φ j for the jth row of Φ, and Φ(k) = [φ j]m
a (n− 1)× m submatrix of Φ with the kth row deleted; the jth column of Φ(k) is denoted by
j ∈ Kn−1. We also use the notation Φ = [φi, j] or [Φ]i, j = φi, j. For an index set I ⊂ {1, . . . ,m},
φ (k)
we denote the restriction of Φ to these column vectors by ΦI = [φi]i∈I.
Definition 2.2. Given a frame Φ, another frame Ψ = (ψi)m
of Φ if the following reproducing formula holds:
i=1 ⊂ Hn is said to be a dual frame
x =
m
∑
i=1
(cid:104)x,φi(cid:105)ψi
for all x ∈ Hn.
(2.2)
In matrix notation this definition reads
ΨΦ∗ = In
(2.3)
or, equivalently, ΦΨ∗ = In, where In is the n× n identity matrix. So the set of all duals of Φ is
the set of all left-inverses Ψ to Φ∗ (or the adjoints of all right-inverses to Φ).
If m > n, then the frame (φi)m
i=1 is redundant, i.e., it consists of more vectors than necessary
for the spanning property. For these frames there exist infinitely many dual frames.
Given a collection of m vectors Φ = (φi)m
i=1 in Hn, the synthesis operator is defined as the
mapping
The adjoint of the synthesis operator is called the analysis operator and is given by
F : Km → Hn,
F∗ : Hn → Km,
ciφi.
(ci)m
i=1
i=1 (cid:55)→ m
∑
x (cid:55)→(cid:0)(cid:104)x,φi(cid:105)(cid:1)m
i=1.
In applications the analysis operation typically represents the way data is measured, whereas
the synthesis procedure represents the reconstruction of the signal from the data.
The identification Hn ∼= Kn translates to identifying the synthesis mapping F with the matrix
Φ and F∗ with Φ∗. The frame operator S : Hn → Hn is defined as S = FF∗, that is,
S(x) =
(cid:104)x,φi(cid:105)φi;
m
∑
i=1
(2.4)
Page 3 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
in matrix form the definition reads S = ΦΦ∗. It is easy to see that Φ is a frame if and only if the
frame operator defined by (2.4) is positive definite. In this case, the following reconstruction
formula holds
x =
m
∑
(cid:104)x,S−1φi(cid:105)φi, for all x ∈ Hn,
(2.5)
hence, (cid:101)Ψ := (S−1φi)m
canonical dual frame. The canonical dual (cid:101)Ψ has frame bounds 1/B and 1/A, where A and B
are frame bounds of Φ. In terms of matrices, (cid:101)Ψ is the Moore-Penrose pseudo-inverse Φ† =
i=1 is a dual frame of Φ in the sense of Definition 2.2, it is called the
j=1
Φ∗(ΦΦ∗)−1 of the analysis matrix Φ∗.
2.2 Sparsity and spectrum
The measure we use for sparsity is the number of non-vanishing entries of the dual frame which
we denote, with a slight abuse of notation, by (cid:107)·(cid:107)0. This convention as well as the associated
name (cid:96)0-norm (although it is not a norm) is common practice for vectors and relates to the fact
that it can be interpreted as the limit of (cid:96)p-quasinorms, p → 0+. In our paper, the term norm
will also be used in a sense that includes the (cid:96)0-norm.
As for the spectral properties of a frame Φ, we will mostly work with singular values (σi)n
i=1
of the synthesis matrix Φ. While one often studies the eigenvalues (λi)n
i=1 of the frame operator
SΦ = ΦΦ∗ in related cases, the advantages of the singular value approach is that one then
has the frame directly accessible, whereby the construction of dual frames becomes, if not
straightforward, then at least tractable. Obviously, the values are related by λi = (σi)2 for all
i = 1, . . . ,n.
2.3 The set of all duals
By the classical result in [15], all duals to a frame Φ can be expressed as
(cid:40)
Φ φi + ηi − m
S−1
∑
k=1
(cid:10)S−1
Φ φi,φk
(cid:41)m
(cid:11)ηk
,
i=1
(2.6)
where ηi ∈ Hn for i = 1, . . . ,m is arbitrary; in matrix form it reads:
Φ Φ + E(Im − Φ∗S−1
S−1
Φ Φ),
Ψ + E(Im − Φ∗Ψ).
where E is a matrix representation of an arbitrary linear mapping from Km to Hn. The canonical
dual appearing in these formulas can be replaced by any other dual, that is, if Ψ is any dual,
then all duals are given by
(2.7)
The matrix Φ∗Ψ is the so-called cross Gramian of the dual frame pair. Here E : Km → Hn
can be an arbitrary linear mapping, but different choices of E can yield the same dual, so there
In fact, if one considers Φ as given and Ψ
are less degrees of freedoms than entries in E.
as our unknown, equation (2.3) corresponds to solving n independent linear systems, each of
which consists of n equations in m variables. Hence, the set of all duals Ψ to a frame Φ is an
n(m− n)-dimensional affine subspace of Mat(K,n× m).
A natural parametrization of this space is obtained using the singular value decomposition.
Let Φ = UΣΦV∗ be a full SVD of Φ, i.e., U ∈ Kn×n and V ∈ Km×m are unitary and ΣΦ ∈ Rn×m
Page 4 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
√
A > 0
is a diagonal matrix whose entries are the singular values
of Φ as its diagonal entries σ j ≥ 0 in a non-increasing order. Let Ψ be a frame and define
MΨ := U∗ΨV ∈ Kn×m, where U and V are the right and left singular vectors of Φ. Then Ψ
factors as Ψ = UMΨV∗. By ΦΨ∗ = In, we then see that
√
B = σ1 ≥ σ2 ≥ ··· ≥ σn =
In = U∗InU = U∗ΦΨ∗U = ΣΦM∗
Ψ.
Therefore, Ψ is a dual frame of Φ precisely when
ΣΦM∗
where Ψ = UMΨV∗. The solutions to (2.8) are given by
s1,1
s2,1
...
sn,1
...
. . . 1/σn
1/σ1
MΨ =
1/σ2
···
0
0
0
0
0
0
Ψ = In,
s1,r
s2,r
...
sn,r
s1,2
s2,2
...
sn,2
···
···
···
(2.8)
(2.9)
where si,k ∈ K for i = 1, . . . ,n and k = 1, . . . ,r = m− n. Note that the canonical dual frame is
obtained by taking si,k = 0 for all i = 1, . . . ,n and k = 1, . . . ,m− n.
3 Sparsity of duals
The goal of this section is to find the sparsest dual frames of a given frame Φ ∈ Kn×m. This
question can be formulated as the minimization problem
min(cid:107)Ψ(cid:107)0
s.t. ΦΨ∗ = In.
(3.1)
Definition 3.1. We call the solutions of (3.1) sparsest duals or optimal sparse duals.
Let us consider a small example illustrating that solutions to this minimization problem are
not unique.
Example 1. We consider duals of the frame Φ = (φ1,φ2,φ3) in R2 defined by
0
2 −1
.
(cid:21)
(cid:20)0 0 0
(cid:21)
(cid:21)
1
3 1
1
3
Φ =
(cid:20)1 −1
(cid:21)
1
1
3 1
3
0 0 0
(cid:20) 1
(cid:20) 2
3 − 1
0
(cid:35)
Ψ =
(cid:34) 2
3 − 1
(cid:20)0 −1 −2
(cid:21)
1
3
3 0
1
3 0
(3.2)
(3.3)
All duals Ψ are easily found by Gauss-Jordan elimination to be given by
+t1
+t2
t1,t2 ∈ R,
,
hence the sparsest duals are seen to be
Ψ1 =
, Ψ2 =
0
0 −1
(3.4)
which correspond to the parameter choices t2 = −1 and t1 = −2,0,1, respectively. These are the
only three duals with (cid:107)Ψ(cid:107)0 = 3. According to the measures derived in [8], the non-degenerate
dual Ψ2 without zero vectors is preferred. This reflects the fact that it is the only one of the
three duals for which one is not discarding frame coefficients in the reconstruction procedure.
,
,
and Ψ3 =
3 0
1
0 0 −1
0
3
0 −1
(cid:20) 2
(cid:21)
Page 5 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
Solving (3.1) of sizes much larger than the 2× 3 example in Example 1 is not feasible since
the problem is NP-complete. Hence to find sparse duals, one has to settle for approximate
solutions of (3.1), e.g., by convex relaxation or greedy strategies. For such approaches we refer
to [8, 16]. In the following subsections we will focus on analyzing the possible values of (cid:107)Ψ(cid:107)0,
rather than on how one finds minimizers. An exception is Section 3.3, where finding the sparsest
duals is tractable due to the specific structure of the frame.
3.1 Upper and lower bounds for the minimal sparsity
In this section we investigate the possible sparsity levels in the set of all dual frames. In other
words, we consider the possible objective function values of the minimization problem (3.1).
We start with a trivial upper bound for the sparsity level of sparse duals.
Lemma 3.1. Suppose Φ is a frame for Kn. Then there exists a dual frame Ψ of Φ with
(cid:107)Ψ(cid:107)0 ≤ n2.
(3.5)
Proof. Choose J ⊂ {1,2, . . . ,m} such that J = n and the vectors {φi}i∈J are linearly indepen-
dent. This is always possible, as otherwise, the φi do not span the space and hence do not form a
frame. Let {ψi}i∈J be the unique (bi-orthogonal) dual of {φi}i∈J, and set ψi = 0 for i /∈ J. Then
obviously (cid:107)Ψ(cid:107)0 ≤ n2 holds.
Similarly to Lemma 3.1 we see that the corresponding trivial lower bound on the dual frame
sparsity is n, that is, any dual frame satisfies (cid:107)Ψ(cid:107)0 ≥ n. However, we can give a much better
lower bound in terms of the spark of the Φ( j)'s. Here the spark of a matrix Φ ∈ Mat(K,n× m)
is defined as the smallest number of linear dependent columns of Φ and denoted by spark (Φ).
In case Φ is an invertible n× n matrix, one sets spark (Φ) = n + 1.
Theorem 3.2. Suppose Φ is a frame for Kn. Then any dual frame Ψ of Φ satisfies
(cid:107)Ψ(cid:107)0 ≥ n
∑
j=1
spark (Φ( j)).
(3.6)
Proof. Fix j = 1, . . . ,n. By equation (2.3), the jth row of Ψ, that is, ψ j, must be orthogonal to
φ k for k = 1, . . . , j− 1, j + 1, . . . ,n. Therefore, we have that
0n−1 = Φ( j)(ψ j)∗ = [φ ( j)
i
]i∈suppψ j ψ jsuppψ j,
(3.7)
: j ∈ suppψ j}
where 0n−1 is the (n−1)×1 zero column vector. Hence, the (cid:107)ψ j(cid:107)0 columns {φ ( j)
of Φ( j) are linearly dependent, which implies that (cid:107)ψ j(cid:107)0 ≥ spark (Φ( j)) and thus the result.
i
It is easy to see that the lower bound on the support given in Theorem 3.2 is not always
sharp. Indeed, if Φ has two linearly dependent columns, one has spark (Φ( j)) = 2 for all j,
but there is not necessarily a dual with only two entries in each row. In order to predict the
maximally possible sparsity, one needs a more refined notion that excludes this type of effects
arising from linearly dependent columns. More precisely, denote by spark j (Φ) the smallest
number of linearly dependent columns in Φ( j) such that the corresponding columns in Φ are
linear independent. Obviously, we have that spark j (Φ) ≥ spark (Φ( j)) for every j = 1, . . . ,n.
Using this refined notation of spark we can improve the bound obtained in Theorem 3.2 for
optimal sparse duals. Even more, the following result exactly describes the sparsity level of the
sparsest dual.
Page 6 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
Proposition 3.3. Suppose Φ is a frame for Kn. Then the sparsest dual frame Ψ of Φ satisfies
(cid:107)Ψ(cid:107)0 =
n
∑
j=1
spark j (Φ).
(3.8)
k
(cid:1)
Proof. Let Ψ be a sparsest dual of Φ. Then(cid:0)φ ( j)
to allow for Φ( j)(cid:0)ψ j(cid:1)∗
allow for the construction of ψ j with supp ψ j (cid:40) suppψ j such that Φ(cid:0) ψ j(cid:1)∗
k∈suppψ j must be linearly dependent in order
= 0n−1. However, the frame vectors (φk)k∈suppψ j must be linearly inde-
pendent, as otherwise one of these columns can be expressed through the others, which would
= e j. This in turn
would imply that the frame, whose synthesis matrix has rows ψ1, . . . ,ψ j−1, ψ j,ψ j+1, . . . ,ψn,
is also a dual frame of Φ, so Ψ is not the sparsest dual. Thus suppψ j ≥ spark j (Φ), which
implies that no dual frame can have less than ∑n
To show the existence of dual frames with this sparsity, let S be a set of size spark j (Φ)
such that (φk)k∈S is a set of independent columns of Φ such that the corresponding columns
of Φ( j) are linearly dependent. That is, there exist (λk)k∈S such that ∑k∈S λk(φ ( j))k = 0, but
(∑k∈S λkφk) j = a (cid:54)= 0. Hence the vector ψ j given by
j=1 spark j (Φ) non-zero entries.
(cid:40) λk
a
0
ψ j
k =
if k ∈ S,
else,
(3.9)
satisfies Φ(ψ j)∗ = e j, and the matrix Ψ with rows ψ j yields a dual with a frame matrix support
size of ∑n
j=1 spark j (Φ), as desired.
3.2 Frames with the lower sparsity bound n2
In Lemma 3.1 we saw that it is always possible to find a dual frames with sparsity level n2. In
this section we will show that for a large class of frames this is actually the best, or rather the
sparsest, one can achieve.
Recall that a n×m matrix is said to be in general position if any sub-collection of n (column)
vectors is linear independent, that is, if any n× n submatrix is invertible. The following result
states that if the projection of Φ onto any of the n coordinates hyperplanes of dimension n− 1
is in general position, then the frame has maximal dual sparsity level. This observation follows
from Theorem 3.2.
Theorem 3.4. Suppose Φ is a frame for Kn such that the submatrix Φ( j) is in general position
for every j = 1, . . . ,n. Then any dual frame Ψ of Φ satisfies
(cid:107)Ψ(cid:107)0 ≥ n2.
(3.10)
In particular, the sparsest dual satisfies (cid:107)Ψ(cid:107)0 = n2
Proof. Fix j = 1, . . . ,n, and let us consider c = (ci)m
chosen. Then
i=1 = (ψ j)∗ as a m× 1 column vector to be
m
∑
i=1
ciφ ( j)
i = 0n−1.
(3.11)
Page 7 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
Since Φ(k) is in general position, any collection of n − 1 vectors from (φ (k)
)m
i=1 is linearly
i
independent. Hence, equation (3.11) can only be satisfied if c has more than n − 1 nonzero
entries, that is, if {i : ci (cid:54)= 0} ≥ n. Hence, each row of Ψ must have at least n nonzero entries,
which yields a total of n2 nonzero entries.
We illustrate this result with a number of examples of frames which are well-known to be in
general position, and which thus do not allow for dual frames with less than n2 non-vanishing
entries.
Example 2. For any n,m ∈ N with m ≥ n, let ai > 0, i = 1, . . . ,m with ai (cid:54)= a j for all i (cid:54)= j, and
let b j > 0, j = 1, . . . ,n with b j (cid:54)= bi, j (cid:54)= i. We then define the generalized Vandermonde frames
as
(3.12)
ab1
...
ab1n
1
Φ =
.
ab2
1
...
ab2n
··· abm
1
...
··· abmn
It is not difficult to see that the submatrix Φ( j) is in general position for every j = 1, . . . ,m.
This yields a deterministic construction for a frame Φ ∈ Kn×m such that any dual frame Ψ of Φ
satisfies
(cid:107)Ψ(cid:107)0 ≥ n2.
Remark 1. We remark that any minor of a generalized Vandermonde frames Φ is non-zero,
which, in particular, implies that the submatrices Φ( j) are in general position for every j =
1, . . . ,m. Neither of these properties hold, in general, for standard Vandermonde matrices. A
standard Vandermonde frame Φ is obtained as the transpose of the matrix in (3.12) with b j =
j−1 for j = 1, . . . ,n and ai ∈ K for i = 1, . . . ,m. Such a standard Vandermonde frame is again in
general position, see also [1]. To see that this is not necessarily the case for all Φ( j), take n = 3
and let ai0 = −1 and ai1 = 1 for some i0 and i1 ∈ {1, . . . ,m}. Then the submatrix Φ(2) restricted
to the indices {i0,i1} is a 2-by-2 matrix of all ones, hence Φ(2) cannot be not in general position.
The generalized Vandemonde frames introduced above are almost never used in applications
due to the fact that for m (cid:29) n they are necessarily poorly conditioned, both in the traditional
sense and in the frame theory sense as introduced in [8]. The following example yields well-
conditioned, deterministic frames for which no dual frame with less than n2 entries can exist
either.
Example 3. Let n ∈ N, and let m be prime. Let Φ be a partial FFT matrix or a harmonic tight
frame matrix of size n × m. Then any Φ( j) is in general position, as the determinant of any
(n− 1)× (n− 1) submatrix of Φ is non-zero. This is a consequence of Chebotarev theorem
about roots of unity stating that any minor of an m× m DFT matrix is non-zero whenever m is
prime [19–21]. Thus again, the sparsest dual frame Ψ of Φ satisfies
(cid:107)Ψ(cid:107)0 = n2.
Example 4. Let n ∈ N be prime and let m = n2. It was shown in [11] that for almost every
a ∈ Cn, the Gabor frame generated by a has the property that any minor of its frame matrix is
non-zero. We conclude as before that the sparsest dual frame satisfies
(cid:107)Ψ(cid:107)0 = n2.
Page 8 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
In fact, the property of having a sparsest dual with sparsity n2 is a generic property. To
precisely formulate this observation, let F (n,m) be the set of all frames of m vectors in Hn, let
N (n,m) be the set of all frames of m vectors in Hn whose sparsest dual has n2 non-zero entries,
and let P(n,m) be the the set of all frames Φ which satisfy spark (Φ( j)) = n for all j = 1, . . . ,n.
We remark that the following result can be obtained using techniques from algebraic geometry
[1]; we include an elementary proof for illustrative purposes.
Lemma 3.5. Suppose m ≥ n. Then the following assertions hold:
(i) The set P(n,m) is open and dense in Kn×m, that is, P(n,m)c is closed and nowhere
dense.
(ii) The set P(n,m)c is of (induced Euclidean) measure zero.
Proof. (i) Let Φ0 ∈ Kn×m be a given. Pick Φ1 ∈ P(n,m), e.g., a generalized Vandermonde
frame. Define
Φt = tΦ1 + (1−t)Φ0
t ∈ K.
We claim that only finitely many of {Φt : t ∈ K} are not in P(n,m). To see this we introduce
the following polynomial in t:
P(t) =
n
∏
j=1
∏
I∈S
det([Φ( j)
t
]I),
where S is the collection of all subsets of {1, . . . ,m} of size n− 1 and [Φ( j)
t
the matrix Φt without row j and restricted to the columns in the index set I. Clearly, S =(cid:0) m
and P is therefore of degree n2(cid:0) m
(cid:1). By definition, we see that
]I ∈ K(n−1)×(n−1) is
n−1
(cid:1),
n−1
P(t) = 0
⇔ det([Φ( j)
t
⇔ spark (Φ( j)
]I) = 0
t ) < n
for some I ∈ S and j = 1, . . . ,n
for some j = 1, . . . ,n
the zero polynomial hence it has finitely many zeros, to be precise, it has at most n2(cid:0) m
Therefore, P(t) (cid:54)= 0 if and only if Φt ∈ P(n,m). Since P(1) (cid:54)= 0, the polynomial P(t) is not
Therefore, we can choose t arbitrarily small such that Φt ∈ P(n,m). Consequently, Φ is
arbitrarily close to a frame in P(n,m).
(cid:1) zeros.
n−1
Part (ii) follows in a similar setup as in the prove of (i) using an integration argument.
By Theorem 3.2 we see that P(n,m) ⊂ N (n,m) and thus F (n,m)\ N (n,m) ⊂ F (n,m)\
P(n,m). Hence, the following result immediately follows from Lemma 3.5.
Theorem 3.6. Any frame in F (n,m) is arbitrarily close to a frame in N (n,m). Moreover, the
set F (n,m)\ N (n,m) is of measure zero.
Another consequence of Lemma 3.5 is that for many randomly generated frames, the spars-
est dual has sparsity level (cid:107)Ψ(cid:107)0 = n2. As an example, this holds when the entries of Φ are
drawn independently at random from a standard normal distribution. Also frames obtained by
a small Gaussian random perturbations of a given frame will have this property.
Page 9 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
3.3 Sparse duals of sparse frames
We have seen in the previous sections that, while the sparsity of the sparsest dual of a frame
can lie between n and n2, generically, its value is n2. One can, however, by imposing structure,
obtain classes of frames that allow for considerably sparser duals. One example is a class of
very sparse frames, the so-called spectral tetris frames.
(cid:1)n
The spectral tetris algorithm is constructs unit norm frames of m vectors in Rm with pre-
scribed eigenvalues of the frame operator. It was developed in [5] and extended in [2]. The
j=1 that satisfy λ j ≥ 2 for all j and for
We will need to introduce some notation that we will use throughout this section. For a
j=1, we define ki and mi as follows. Set k0 = 0 and m0 = 0 and recursively
j=1 λ j ∈ N}. Naturally, the largest value µ that can be
construction works for any sequence of eigenvalues(cid:0)λ j
sequence (cid:0)λ j
Definition 3.2. For a sequence(cid:0)λ j
Tetris frame STF(n,m,(cid:8)λ j
choose ki = inf{N (cid:51) k > ki−1 : mi := ∑k
chosen for i is such that mµ = m and kµ = n. Furthermore, let K := {ki : i = 0,1, . . . , µ}.
(cid:9)) ∈ Rn×m is given by
j=1 satisfying the condition m := ∑n
j=1 λ j ∈ N, the Spectral
j=1 λ j is an integer.
which m := ∑n
(cid:1)n
(cid:1)n
,
··· 1 an−1
an−1
bn−1 −bn−1 1 ··· 1
mn times
(cid:124) (cid:123)(cid:122) (cid:125)
1 ··· 1 a1
a1
b1 −b1 1 ··· 1 a2
a2
b2 −b2 1 ···
(cid:124) (cid:123)(cid:122) (cid:125)
m1 times
(cid:124) (cid:123)(cid:122) (cid:125)
m2 times
(cid:113) r j
(cid:113)
where, for j /∈ K, a j :=
2 and b j :=
1− r j
λ j = m j + r j,
λ j = (2− r j−1) + m j + r j,
(cid:104) a j a j
2 , and r j ∈ [0,1) and m j ∈ N0 are defined by
(cid:105)
when j− 1 ∈ K,
otherwise.
(3.13)
(3.14)
b j −b j
is omitted.
If j ∈ K, the 2× 2-block matrix B j =
Proposition 3.7. Suppose(cid:0)λ j
Φ := STF(n,m,(cid:0)λ j
(cid:1)n
j=1). Define the set I,J ⊂ {1, . . . ,n} by
I = {0 ≤ j ≤ µ − 1 : k j+1 = k j + 1}
(cid:1)n
j=1 λ j ∈ N. Let
j=1 satisfies λ j ≥ 2, j = 1, . . . ,n, and m := ∑n
(cid:33)
(cid:32)(cid:36) j0−1
(cid:41)
(cid:37)
and
j0 ∈ (ki + 1,ki+1) : i = 0, . . . , µ and
(cid:40)
J =
j0∑
j=ki+1
λ j −
∑
j=ki+1
λ j
+ 2
≥ 1
,
and let k := 2µ +J−I, where µ is the maximal index i as above. Then the sparsest dual Ψ
satisfies
(cid:107)Ψ(cid:107)0 = k + 2(n− k).
Page 10 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
Proof. Let (e j)n
since Φ( j) contains a zero column. On the other hand, if e j (cid:54)∈ Φ, then spark j (Φ) = 2. Let
j=1 denote the standard orthonormal basis of Rn. If e j ∈ Φ, then spark j (Φ) = 1
k =(cid:12)(cid:12)(cid:8) j : e j ∈ Φ(cid:9)(cid:12)(cid:12) .
(3.15)
By Proposition 3.3, it then follows directly that the sparsest dual satisfies
(cid:107)Ψ(cid:107)0 =
n
∑
j=1
spark j (Φ) = k + 2(n− k).
To complete the proof, we need show that k always equals k = 2µ +J−I, i.e., that there are
exactly 2µ +J−I rows of Φ containing a 1.
For that, recall that each k j corresponds to a 2×2-block matrix B j =
is em j+1; each of the µ values of k j yields two rows that contain an e j. if j ∈(cid:0)K ∪ (K + 1)(cid:1)∩
being omitted
at the transition from one row to the next. Hence the k jth column is em j, the (k j + 1)st column
{1, . . . ,n}, then e j ∈ Φ. Correcting for the double counting resulting from the case k j+1 = k j +1,
there are 2µ −I values of (cid:96) such that e(cid:96) ∈ Φ resulting from this phenomenon.
In addition, there can be e(cid:96) ∈ Φ that do not correspond to such a transition, i.e., that appear
as the j0th column of the frame matrix, where j0 is such that ki + 1 < j0 < ki+1 for some i. For
such a j0 we calculate
(cid:104) a j a j
b j −b j
(cid:105)
m j0 + r j0 = λ j0 − 2 + r j0−1
= λ j0 − 2 + λ j0−1 − 2 + r j0−2 − m j0−1
= (λ j0 + λ j0−1)− (4 + m j0−1)− r j0−2
j0−1
= ··· =
∑
j=ki+1
λ j −(cid:16)
2( j0 − 1) +
j0∑
j=ki+1
(cid:17)
m j
By definition the j0th row contains a 1 if and only if m j0 ≥ 1. Furthermore, it can easily be
shown, e.g., by induction, that
2( j0 − ki − 1) +
m j =
λ j
+ 2.
(cid:37)
(cid:36) j0−1
∑
j=ki+1
(cid:33)
(cid:37)
j0−1
∑
j=ki+1
(cid:32)(cid:36) j0−1
∑
j=ki+1
j0∑
j=ki+1
λ j −
λ j
+ 2
≥ 1.
Therefore, the j0th row of Φ contains a 1, if and only if
This is just the defining condition of J, so one has an additional number of J vectors e(cid:96) ∈ Φ,
which proves the theorem.
Remark 2. Since µ ≥ 1, we see that k ≥ 2. Thus the sparsest dual of a Spectral Tetris frame
satisfies n ≤ (cid:107)Ψ(cid:107)0 = 2n− k ≤ 2n− 2. Furthermore, if λ j ≥ 3 for all j = 1, . . . ,n, we see that
(cid:107)Ψ(cid:107)0 = n since k = n.
Page 11 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
Remark 3. Recall that the spectral tetris algorithm is only guaranteed to work for redundancies
larger than two, see Definition 3.2. In [12] the proof of Proposition 3.7 is used to characterize
when the spectral tetris construction works for tight frames with redundancies below two. In
fact, a similar argument works for general frames without a tightness condition.
Remark 4. A construction scheme for optimal sparse duals of Spectral Tetris frames directly
follows from the proof of Proposition 3.7. The resulting dual will consist only of 1-sparse
vectors and 2-sparse vectors.
4 Spectral properties of duals
In this section we answer the question of which possible spectra the set of all dual frames
admits. A special case of this question has already been answered in [18], namely the question,
which frames admit tight duals. In linear algebra terms this boils down to the question whether
a full-rank matrix Φ has a right inverse with condition number one. The first step towards a
more general characterization will be a more quantitative version of their result. In particular,
we study which frame bounds tight duals can attain for a given frame. Furthermore, our proof
provides an explicit construction procedure for these duals. The understanding developed in the
proof will turn out useful in some more general cases.
It turns out that a frame always has a tight dual if the redundancy is two or larger. If the
redundancy is less than two, it will only be possible under certain assumptions on the singular
values of Φ. Note that in [18], the result is formulated in terms of the eigenvalues of the frame
operator, so even the existence part of the following result is formulated slightly different from
the result in [18].
Theorem 4.1. Let n,m ∈ N. Suppose Φ is a frame for Kn with m frame vectors and lower frame
bound A. Then the following assertions hold:
(i) If m ≥ 2n, then for every c ≥ 1√
, there exists a tight dual frame Ψ with frame bound c.
A
(ii) If m = 2n− 1, then there exists a tight dual frame Ψ; the only possible frame bound is 1√
.
A
(iii) Suppose m < 2n− 1. Then there exists a tight dual frame Ψ if and only if the smallest
2n− m ∈ {2, . . . ,n} singular values of Φ are equal.
Proof. Let Φ = UΣΦV∗ be a full SVD of Φ, and let Ψ be an arbitrary dual frame. Following
Section 2.3, we factor the dual frame as Ψ = UMΨV∗, where MΨ is given as in (2.9) with
si,k ∈ K for i = 1, . . . ,n and k = 1, . . . ,r = m− n. For Ψ to be tight, we need to choose si,k such
that the rows of MΨ are orthogonal and have equal norm. This follows from the fact that Ψ is
row orthogonal if and only if MΨ is row orthogonal.
As the diagonal block of MΨ is well-understood, the duality and tightness constraints trans-
late to conditions for the inner products of the si = (si,1, . . . ,si,r) ∈ Kr, i = 1, . . . ,n. Indeed, Ψ
is a tight dual frame with frame bound c if and only if, for all 1 ≤ i ≤ n, one has
1
σ 2
i
and, for all i (cid:54)= j = 1, . . . ,n, one has (cid:104)si,s j(cid:105) = 0.
Now assume that σn = σn−1 = ··· = σp+1 < σp for some p < n. The case of distinct singular
values corresponds to p = n− 1. As σp+1 < σi for all 1 ≤ i ≤ p, (4.1) implies that all si for
+(cid:107)si(cid:107)2
2 ,
c =
(4.1)
Page 12 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
j=1 is orthogonal, else the sequence (si)p
i = 1, . . . , p must be nonzero vectors even if sp+1, . . . ,sn are all zero. Furthermore, (4.1) also
determines the norms of s1, . . . ,sp as a function of (cid:107)sp+1(cid:107) = ··· = (cid:107)sn(cid:107). The second condition
implies that if sn (cid:54)= 0, the sequence (s j)n
If r ≥ n, that is, if m ≥ 2n, then any choice of sn allows for an orthogonal system with com-
patible norms, so tight dual frames with any frame bound above 1
exist and can be efficiently
σn
constructed. If r < n, then no n vectors can form an orthogonal system, one needs to have sn = 0
and hence also s j = 0 for all j > p. So no frame bound other than 1
is possible. The remaining
σn
vectors {s j}p
j=1 are all non-zero, so they must form an orthogonal system. For r ≥ n− p+1, this
is possible, and again a solution satisfying the norm constraints can be efficiently constructed.
For r ≤ n− p, no such system exists, hence there cannot be a tight dual.
i=1.
This completes the proof.
We remark that for tight frames Φ, hence p = 0, one has the following two cases: If m < 2n,
then Φ has only A−1Φ as tight dual frame. If m ≥ 2n, then Φ has infinitely many tight dual
frames, in particular, Φ has a C-tight dual frame for any C ≥ A−1.
We will now derive general conditions on which spectral patterns (now possibly consisting
of more than one point) can be achieved by a dual frame of a given frame. The reason that,
in the general framework, such an analysis is harder than in the context of tight duals is that
in that case, the frame operator is a multiple of the identity, hence diagonal in any basis. This
no longer holds true if we drop the tightness assumption, so when the orthogonality argument
of Theorem 4.1 fails, one cannot conclude that there is no dual with a given spectral pattern.
However, the orthogonality approach allows to choose a subset of the singular values of the dual
frame freely. In particular, if the redundancy of the frame Φ is larger than 2, it follows that for all
spectral patterns satisfying a set of lower bounds, which we will later show to be necessary (see
Theorem 4.6), a dual with that spectrum can be found using a constructive procedure analogous
to the proof of Theorem 4.1.
Theorem 4.2. Let n,m ∈ N, and let Φ be a frame for Kn with m frame vectors and singular
i=1. Suppose that r ≤ m− n and that Ir ⊂ {1, . . . ,n} with Ir = r. Then, for any
values (σi)n
sequence (qi)i∈Ir satisfying qi ≥ 1/σi for all i ∈ Ir, there exists a dual frame Ψ of Φ such that
{qi}i∈Ir is contained in the spectrum of Ψ. Furthermore, it can be found constructively using a
sequence of orthogonalization procedures.
Proof. The proof is just a slight modification of the proof of Theorem 4.1. Again, we choose
(si)i∈Ir to be orthogonal and the remaining si's to be the zero vector. The non-zero si vectors are
scaled to satisfy
q2
i =
1
σ 2
i
+(cid:107)si(cid:107)2
2 ,
where i∈ Ir. Hence, by this procedure we obtain a dual frame with spectrum {qi}i∈Ir ∪{σ−1
i }i /∈Ir.
As a corollary we obtain that using the same simple constructive procedure, one can find
dual frames with any frame bound that is possible.
Corollary 4.3. Let Φ be a redundant frame for Kn with singular values (σi)n
≥ AΨ ≥ 1
frame bound satisfying BΨ ≥ 1
σ 2
σ 2
n
1
and a lower frame bound
σ 2
m−n+1
1
i=1. Fix an upper
, where we use the
Page 13 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
convention
found constructively using a sequence of orthogonalization procedures.
= ∞ if m ≥ 2n. Then a dual frame Ψ of Φ with these frame bounds can be
σm−n+1
1
To obtain a complete characterization on the possible spectra for dual frames, we will need
the following two results by Thompson [23] on the interlacing properties of the spectrum of sub-
matrices. To simplify notation in the following proof, we use a simpler formulation than in [23],
which is technically speaking a special case of the setup given in [23], but up to permutation
captures the results in complete generality.
For our setup, let A ∈ Kn×m with singular values α1 ≥ . . .αmin(n,m), where m,n ∈ N. Fur-
thermore, for q, (cid:96) ∈ N, (cid:96) ≥ q, let Pq : K(cid:96) → Kq denote the restriction on the q first entries. Of
course, the Pq depends not only on q, but also on the input dimension (cid:96), but as this can usually
be inferred from the context, we suppress this dependence to simplify notation. Note that P∗
q Pq
is the projection onto span{ei : i = 1, . . . ,q}, i.e.,
(cid:40)
(P∗
q Pqx)i =
xi
0
if i ≤ q,
else,
q is the identity on Kq. Furthermore, we take p,q ∈ N such that p ≤ n and q ≤ m.
and PqP∗
Theorem 4.4 ( [23, Theorem 1]). Suppose β1 ≥ ··· ≥ βmin(p,q) are the singular values of B ∈
Kp×q given by B = PpAP∗
q . Then
αi ≥ βi
βi ≥ αi+(m−p)+(n−q)
(4.2)
(4.3)
Theorem 4.5 ( [23, Theorem 2]). Let β1 ≥ ··· ≥ βmin(p,q) satisfy (4.2) and (4.3). Then there
exist unitary matrices U ∈ Kn×n and V ∈ Km×m such that PpUAV∗P∗
q ∈ Kp×q has singular
values (βi)
1 ≤ i ≤ min{p,q} ,
i ≤ min{p + q− m, p + q− n} .
for
for
.
min{p,q}
i=1
These two results now allow for a complete characterization of the possible spectra of dual
frames.
Theorem 4.6. Let n,m ∈ N, and set r = m− n. Let Φ be a frame for Kn with singular values
(σi)n
i=1 (also arranged in a non-
increasing order). Then the following inequalities hold:
i=1. Suppose Ψ is any dual frame with singular values (σ Ψ
i )n
1
σn−i+1
1
σn−i+1
≤ σ Ψ
i
≤ σ Ψ
i ≤
1
σn−i+r+1
for i = 1, . . . ,r,
for i = r + 1, . . . ,n.
(4.4)
(4.5)
Furthermore, for every sequence (σ Ψ
with singular values (σ Ψ
i )n
i=1.
i )n
i=1 which satisfies (4.4) and (4.5), there is a dual Ψ of Φ
Proof. To show the necessity of the conditions, we will apply Theorem 4.4 for p = q = n. Note
that this entails that Pp = In. Let Φ = UΣΦV∗ be a full SVD of Φ. We factor dual frames of Φ as
Ψ =UMΨV∗, where MΨ is given as in (2.9) with si,k ∈ K for i = 1, . . . ,n and k = 1, . . . ,r = m−n.
Page 14 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
We let M(cid:101)Ψ denote the case where we choose si,k = 0 for all i,k; in this case(cid:101)Ψ = UM(cid:101)ΨV∗ is the
n Pn = M(cid:101)Ψ, namely
i=1. Furthermore, Ψ and MΨ have the same singular values (σ Ψ
i )n
i=1.
n ∈ Kn×n now shows (4.4) and
For the existence, we need to show that there are unitary matrices U ∈ Kn×n, V ∈ Km×m
canonical dual. Note that MΨP∗
singular values (1/σn−i+1)n
A direct application of Theorem 4.4 for A = MΨ and B = MΨP∗
(4.5).
such that for the diagonal matrix ΣΨ ∈ Rn×m with diagonal entries (σ Ψ
i )n
n ∈ Kn×n has the same spectrum as MΨP∗
i=1 one has
i )n
ΦVΣ∗
ΨU∗ = In.
For this we will apply Theorem 4.5 for A = Φ, (α)n
sequence (σ Ψ
ters, this is equivalent to having (βi)n
Theorem 4.5, there exist unitary matrices U1 ∈ Kn×n and V1 ∈ Km×m such that U1ΦV∗
hence ΦV∗
SVD, there exist unitary matrices U,V2 ∈ Kn×n such that U∗ΦV∗
i=1, and again p = q = n. Let a
i=1 which satisfies (4.4) and (4.5) be given. Note that for our choice of parame-
1 ) satisfy (4.2) and (4.3). Hence, by
n and
i=1 (in a non-decreasing order). By yet another
2 is an n× n diagonal ma-
i=1 (note that this time, the variant of the SVD with singular values
n have singular values(cid:0)(σ Ψ
trix with entries(cid:0)(σ Ψ
i )−1(cid:1)n
i )−1(cid:1)n
i=1 = (1/σ Ψ
n , . . . ,1/σ Ψ
i=1 = (σi)n
1 P∗
1 P∗
1 P∗
nV∗
(4.6)
1 P∗
2 PnΣ∗
U∗ΦV∗
and, noting that (Im − P∗
in a non-decreasing order is used). Thus, we have
nV∗
Ψ = 0, this leads to
ΦV∗
1 (Im − P∗
n Pn + P∗
2 Pn)Σ∗
A direct calculation using the fact that PnP∗
n = In shows that
1 (Im − P∗
n Pn + P∗
n Pn)Σ∗
V = V∗
Ψ = In
nV∗
nV∗
ΨU∗ = In.
2 Pn) ∈ Km×m
(4.7)
(4.8)
(4.9)
is unitary, so we obtain (4.6) as desired.
The inequalities (4.4) and (4.5), written in terms of the singular values (σ(cid:101)Ψ
ical dual frame(cid:101)Ψ := S−1Φ, have the following simple form:
i )n
i=1 of the canon-
σ(cid:101)Ψ
i ≤ σ Ψ
σ(cid:101)Ψ
i ≤ σ Ψ
i
i ≤ σ(cid:101)Ψ
i−r
for i = 1, . . . ,r,
for i = r + 1, . . . ,n.
We remark that the first part of Theorem 4.6 also follows from r applications of [7, Theorem
7.3.9] on the matrix MΨ defined in (2.9) or from the well-known interlacing inequalities for
Hermitian matrices by Weyl.
Also note that the proof of Theorem 4.5 and hence the proof of Theorem 4.6 involves exis-
tence results for unitary matrices from [22] and is, consequently, less intuitive and constructive
than our above proof for the special case in Theorem 4.2. For this reason, we decided to include
Theorem 4.2, even though it directly follows from the general existence result of Theorem 4.6.
In terms of eigenvalues of frame operators, Theorem 4.6 states that the spectra in the set of
(cid:110)
all duals exhaust the set Λ ⊂ Rn defined by
(λ1, . . . ,λn) ∈ Rn : λ(cid:101)Ψ
i ≤ λi ≤ λ(cid:101)Ψ
i−r for all i = 1, . . . ,n
(cid:111)
Λ =
,
Page 15 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
where λ(cid:101)Ψ
use the convention that λ(cid:101)Ψ
i = 1/λ Φ
i = ∞ for i ≤ 0. By considering the trace of MΨM∗
n−i+1 is the ith eigenvalue of the canonical dual frame operator; we again
Ψ, we see that
the canonical dual frame is the unique dual frame that minimizes the inequalities in Λ. Any
other spectrum in Λ will not be associated with a unique dual frame, in particular, if si =
(si,1, . . . ,si,r) (cid:54)= 0 in MΨ for some i = 1, . . . ,n, then replacing si by zsi for any z = 1 will
yield a dual frame with unchanged spectrum. The frame operator SΨ and the canonical dual
frame operator S(cid:101)Ψ are connected by SΨ = S(cid:101)Ψ +C, where C = W∗W , and W ∈ Kr×n is arbitrary.
matrix S(cid:101)Ψ with spectrum (λ(cid:101)Ψ
Hence, in linear algebra terms, Theorem 4.6 tells us that for a given positive definite, Hermitian
i=1 ∈ Λ, there exists a positive
i )n
semi-definite, Hermitian matrix C of rank at most min{m− n,n} such that the spectrum of
SΨ := S(cid:101)Ψ +C consists of (λi)n
i=1.
i=1 and for a given sequence (λi)n
For a better understanding of the more general framework where Theorem 4.2 does not yield
a complete characterization of the possible spectral patterns, we will continue by an extensive
discussion of the example of a frame of three vectors in R2.
Example 5. Suppose Φ is a frame in R2 with 3 frame vectors and frame bounds 0 < AΦ ≤ BΦ,
and let Φ = UΣΦV∗ be the SVD of Φ. Then all dual frames are given as Ψ = UMΨV∗, where
MΨ =
(cid:21)
0
1/σ2
s1
s2
0
(cid:20)1/σ1
(cid:20)1/σ 2
for s1,s2 ∈ R. Since the frame operator of the dual frame is given by SΨ = ΨΨ∗ = UMΨM∗
we can find the eigenvalues of SΨ by considering eigenvalues of
s1s2
2 + s2
2
S := MΨM∗
1 + s2
1
s1s2
1/σ 2
Ψ =
(cid:21)
.
ΨU∗,
These are given by
λ1,2 =
1
2
trS± 1
2
R, where R =
(cid:113)
(trS)2 − 4detS.
One easily sees that trS monotonically grows as a function of s2
term R grows as a function of s2
by the existence part of Theorem 4.6. A straightforward calculation shows that R + (s2
≥ 0, hence we see that
1
σ 2
2
2, whereas for fixed trS, the
2. This exactly yields the two degress of freedom predicted
2) ≥
1 − s2
− 1
σ 2
1
1 + s2
1 + s2
λ1 ≥ 1
σ 2
2
and
1
σ 2
2
≥ λ2 ≥ 1
σ 2
1
,
(4.10)
which is also the conclusion of the necessity part of Theorem 4.6. We remark that the two
eigenvalues depend only on quadratic terms of the form s2
2. Therefore, if s1 and s2
are non-zero, then the choices (±s1,±s2) yield four different dual frames having the same
eigenvalues. In this case the level sets of λ1 as a function of (s1,s2) are origin-centered ellipses
with major and minor axes in the s1 and s2 direction, respectively. Moreover, the semi-major
axis is always greater than s0 := (σ−2
1 )1/2. The level sets of λ2 are origin-centered,
East-West opening hyperbolas with semi-major axes greater than s0. In Figure 1 the possible
eigenvalues of the dual frame operator of the frame Φ defined by
2 − σ−2
1 and s2
(cid:20) 90 −12 −16
(cid:21)
120
9
6
Φ =
1
50
(4.11)
Page 16 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
are shown as a function of the two parameters s1 and s2; Figure 1b shows the level sets and the
four intersection points (±s1,±s2) for each allowed spectrum in the interior of Λ. Note that the
singular values are σ1 = 3 and σ2 = 1/2, hence BΦ = 9 and AΦ = 1/4.
(a) Graphs of λ1 and λ2
(b) Level curves of λ1 and λ2
s1 and s2. The two graphs in (a) meet at s1 = ±√
Figure 1: The lower and upper frame bounds of dual frames Ψ (to Φ defined in (4.11)) as a function of
35/3 and s2 = 0 which correspond to tight dual frames.
When the difference between the singular values of Ψ goes to zero, the ellipses degenerate
to a line segment (or even to a point if σ1 = σ2). The limiting case corresponds to tight dual
frames so Theorem 4.1(ii) applies, and we are forced to set s2 = 0 to achieve row orthogonality
of MΨ. We then need to pick s1 such that the two row norms of MΨ are equal, thus
(cid:115)
s1 =
1
σ 2
2
− 1
σ 2
1
=
(cid:114) 1
AΦ − 1
BΦ = s0,
which shows that the above lower bound for the semi-major axis is sharp.
Acknowledgments
The authors would like to thank Christian Henriksen for valuable discussions. G. Kutyniok
acknowledges support by the Einstein Foundation Berlin, by Deutsche Forschungsgemeinschaft
(DFG) Grant SPP-1324 KU 1446/13 and DFG Grant KU 1446/14, and by the DFG Research
Center MATHEON "Mathematics for key technologies" in Berlin.
References
[1] B. Alexeev, J. Cahill, D. G. Mixon, Full spark frames, preprint.
[2] R. Calderbank, P. G. Casazza, A. Heinecke, G. Kutyniok, A. Pezeshki, Sparse fusion
frames: existence and construction, Adv. Comput. Math. 35 (2011), 1–31.
Page 17 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
[3] P. G. Casazza, Custom building finite frames, in Wavelets, frames and operator theory,
61–86, Contemp. Math., 345, Amer. Math. Soc., Providence, RI, 2004.
[4] P. G. Casazza, M. Fickus, D. G. Mixon, M. J. Poteet, N. K. Strawn, Constructing finite
frames of a given spectrum and set of lengths, preprint, 2011.
[5] P. G. Casazza, M. Fickus, D. G. Mixon, Y. Wang, and Z. Zhou, Constructing tight fusion
frames, Appl. Comput. Harmon. Anal. 30 (2011), no. 2, 175–187.
[6] P. G. Casazza, A. Heinecke, F. Krahmer, G. Kutyniok, Optimally sparse frames, IEEE
Trans. Inform. Theory 57 (2011), 7279–7287.
[7] R. A. Horn, C. R. Johnson, Matrix analysis, Cambridge University Press, Cambridge,
1985.
[8] F. Krahmer, G. Kutyniok, J. Lemvig, Quantitative comparison of dual frames from the
computational viewpoint, preprint.
[9] M. Lammers, A. M. Powell, Ö. Yılmaz, Alternative dual frames for digital-to-analog
conversion in sigma-delta quantization, Adv. Comput. Math. 32 (2010), 73–102.
[10] P. Casazza, G. Kutyniok, Finite Frames: Theory and Applications, Springer, to appear.
[11] F. Krahmer, G. Pfander, P.Rashkov, Uncertainty in time-frequency representations on finite
abelian groups and applications, Appl. Comput. Harmon. Anal. 25 (2008), 209–225.
[12] J. Lemvig, C. Miller, K. A. Okoudjou, Prime tight frames, preprint.
[13] J. Leng and D. Han, Optimal dual frames for erasures II, Linear Algebra Appl. 435 (2011),
1464–1472.
[14] J. Leng, D. Han, T. Huang, Optimal Dual Frames for Communication Coding With Prob-
abilistic Erasures, IEEE Trans. Signal Proc. 59, 2011, 5380–5389.
[15] S. Li, On general frame decompositions, Numer. Funct. Anal. Optim. 16 (1995), 1181–
1191.
[16] S. Li, Y. Liua, T. Mi, Sparse Dual Frames and Dual Gabor Functions of the Smallest Time
and Frequency Supports, preprint.
[17] J. Lopez, D. Han, Optimal dual frames for erasures, Linear Algebra Appl. 432 (2010),
471–482.
[18] P. G. Massey, M. A. Ruiz, and D. Stojanoff, Optimal dual frames and frame completions
for majorization, preprint.
[19] F. Pakovich, A remark on the Chebotarev theorem about roots of unity, Integers 7 (2007),
A18.
[20] P. Stevenhagen, H.W. Lenstra, Chebotarëv and his density theorem, Math. Intelligencer
18 (1996), 26–37.
Page 18 of 19
date/time: 22-May-2018/3:06
F. Krahmer, G. Kutyniok, J. Lemvig
Sparsity and spectral properties of finite frames
[21] T. Tao, An uncertainty principle for cyclic groups of prime order, Math. Res. Lett. 12
(2005), 121–127.
[22] R. C. Thompson, Principal submatrices of normal and Hermitian matrices, Illinois J.
Math. 10 (1966), no. 2, 269–308
[23] R. C. Thompson, Principal Submatrices IX: Interlacing Inequalities for Singular Values
of Submatrices, Lin. Alg. Appl. 5 (1972), 1–12
Page 19 of 19
date/time: 22-May-2018/3:06
|
1701.04267 | 2 | 1701 | 2017-09-12T16:16:57 | A characterisation of isometries with respect to the L\'evy-Prokhorov metric | [
"math.FA"
] | According to the fundamental work of Yu.V. Prokhorov, the general theory of stochastic processes can be regarded as the theory of probability measures in complete separable metric spaces. Since stochastic processes depending upon a continuous parameter are basically probability measures on certain subspaces of the space of all functions of a real variable, a particularly important case of this theory is when the underlying metric space has a linear structure. Prokhorov also provided a concrete metrisation of the topology of weak convergence today known as the L{\'e}vy-Prokhorov distance. Motivated by these facts, the famous Banach-Stone theorem, and some recent works related to characterisations of onto isometries of spaces of Borel probability measures, here we give a complete description of surjective isometries with respect to the L{\'e}vy-Prokhorov metric in case when the underlying metric space is a separable Banach space. Our result can be considered as a generalisation of L. Moln\'ar's earlier Banach-Stone-type result which characterises onto isometries of the space of all probability distribution functions on the real line wit respect to the L\'evy distance. However, the present more general setting requires the development of an essentially new technique. | math.FA | math |
A CHARACTERISATION OF ISOMETRIES WITH RESPECT TO
THE L´EVY–PROKHOROV METRIC
GY ORGY P ´AL GEH´ER AND TAM ´AS TITKOS
Abstract. According to the fundamental work of Yu.V. Prokhorov, the gen-
eral theory of stochastic processes can be regarded as the theory of probability
measures in complete separable metric spaces. Since stochastic processes de-
pending upon a continuous parameter are basically probability measures on
certain subspaces of the space of all functions of a real variable, a particularly
important case of this theory is when the underlying metric space has a linear
structure. Prokhorov also provided a concrete metrisation of the topology of
weak convergence today known as the L´evy–Prokhorov distance. Motivated by
these facts, the famous Banach–Stone theorem, and some recent works related
to characterisations of onto isometries of spaces of Borel probability measures,
here we give a complete description of surjective isometries with respect to
the L´evy–Prokhorov metric in case when the underlying metric space is a sep-
arable Banach space. Our result can be considered as a generalisation of L.
Moln´ar's earlier Banach–Stone-type result which characterises onto isometries
of the space of all probability distribution functions on the real line wit respect
to the L´evy distance. However, the present more general setting requires the
development of an essentially new technique.
1. Introduction
There is a long history and vast literature of isometries (i.e. not necessarily
sujective distance preserving maps) on different kind of metric spaces. Two classical
results in the case of normed linear spaces are the Mazur–Ulam theorem which
states that every surjective isometry between real normed spaces is automatically
affine (i.e. linear up to translation), and the Banach–Stone theorem which provides
the structure of onto linear isometries between Banach spaces of continuous scalar-
valued functions on compact Hausdorff spaces. Since then several properties of
surjective linear isometries on different types of normed spaces have been explored,
see for instance the papers [2–5, 10, 13, 24] and the extensive books [16, 17]. The
reader can find similar results on non-linear spaces for example in [6,11,18,21,27,28].
2010 Mathematics Subject Classification. Primary: 46B04, 46E27, 47B49, 54E40, 60B10; Sec-
ondary: 28A33, 60A10, 60B05.
Key words and phrases. Borel probability measures; Weak convergence; L´evy–Prokhorov met-
ric; Isometries; Banach–Stone theorem.
Gyorgy P´al Geh´er was also supported by the "Lendulet" Program (LP2012-46/2012) of the
Hungarian Academy of Sciences, and by the Hungarian National Research, Development and
Innovation Office – NKFIH (grant no. K115383).
Tam´as Titkos was also supported by the "Lendulet" Program (LP2012-46/2012) of the Hungar-
ian Academy of Sciences, and by the Hungarian National Research, Development and Innovation
Office – NKFIH (grant no. K104206).
1
2
GY.P. GEH ´ER AND T. TITKOS
The starting point of our investigation is Moln´ar's paper [25] where a complete
description of surjective L´evy isometries of the non-linear space D(R) of all cumu-
lative distribution functions was given. If F, G ∈ D(R), then their L´evy distance is
defined by the following formula:
L(F, G) := inf(cid:8)ε > 0(cid:12)(cid:12) ∀ t ∈ R : F (t − ε) − ε ≤ G(t) ≤ F (t + ε) + ε(cid:9) .
The importance of this metric lies in the fact that it metrises the topology of weak
convergence on D(R). Moln´ar's result reads as follows (see [25, Theorem 1]):
let
Φ : D(R) → D(R) be a surjective L´evy isometry, i.e., a bijective map satisfying
L(F, G) = L(Φ(F ), Φ(G))
(∀ F, G ∈ D(R)).
Then there is a constant c ∈ R such that Φ is one of the following two forms:
Φ(F )(t) = F (t + c)
(∀ t ∈ R, F ∈ D(R)),
or
Φ(F )(t) = 1 − lim
s→t−
F (−s + c)
(∀ t ∈ R, F ∈ D(R)).
In other words, every surjective L´evy isometry is induced by an isometry of R with
respect to its usual norm (or equivalently, by a composition of a translation and a
reflection on R).
The investigation of surjective isometries on spaces of Borel probability mea-
sures was continued for example in [15, 26] for the Kolmogorov–Smirnov distance
which is important in the Kolmogorov–Smirnov statistic and test, and in [7, 8, 20]
with respect to the Wasserstein (or Kantorovich) metric which metrises the weak
convergence.
Let (X, d) be a complete and separable metric space. We will denote the σ-
algebra of Borel sets on X by BX and the set of all Borel probability measures
by PX . The L´evy distance gives a metrisation of weak convergence on D(R), or
equivalently on PR. In 1956 Prokhorov managed to metrise the weak convergence
of PX for general complete and separable metric spaces (X, d). The so-called L´evy–
Prokhorov distance which was introduced by him in [30] is defined by
For the details and elementary properties see e.g. [19, p. 27]. Let us point out that
in the special case when X = R this metric differs from the original L´evy distance.
Here arises the following very natural question:
What is the structure of onto isometries with respect to the L´evy–Prokhorov
metric on PX if X is a general separable real Banach space?
This paper is devoted to give an answer to this question. Namely, we will prove
that every such transformation is induced by an affine isometry of the underlying
space X.
There are some particularly important cases in our investigation which we em-
phasise now. Namely, since stochastic processes depending upon a continuous pa-
rameter are basically probability measures on certain subspaces of the space of
all functions of a real variable (see e.g. [1, 14]), one particularly interesting case
is when the underlying Banach space is C([0, 1]), i.e. the space of all continuous
(1.1)
where
π(µ, ν) := inf(cid:8)ε > 0(cid:12)(cid:12) ∀ A ∈ BX : µ(A) ≤ ν(Aε) + ε(cid:9) ,
Aε := [x∈A
Bε(x) := {z ∈ X d(x, z) < ε}.
Bε(x)
and
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC
3
real-valued functions on [0, 1] endowed with the uniform norm k · k∞. For details
see [30, Chapter 2] or [9, Chapter 2]. Further two important cases are when X
is a Euclidean space because of multivariate random variables, and when X is an
infinite dimensional, separable real Hilbert space because of the theory of random
elements in Hilbert spaces.
2. The setting and the statment of our main result
In this section we state the main result of the paper and collect some definitions
and well-known facts about weak convergence of Borel probability measures. For
more details the reader is referred to the textbooks of Billingsley [9], Huber [19]
and Parthasarathy [29].
Let (X, d) be a complete, separable metric space and denote by Cb(X; R) the
Banach space of all real-valued bounded continuous functions. Recall that BX is
the smallest σ-algebra with respect to each f ∈ Cb(X; R) is measurable. We say
that an element of PX is a Dirac measure if it is concentrated on one point, and
for an x ∈ X the symbol δx stands for the corresponding Dirac measure. The set of
all Dirac measures on X is denoted by ∆X . The collection of all finitely supported
measures is
FX :=(Xi∈I
λi = 1, λi > 0, xi ∈ X (∀ i ∈ I)) ,
λiδxi(cid:12)(cid:12)(cid:12) #I < ℵ0, Xi∈I
Sµ =(cid:8)x ∈ X(cid:12)(cid:12) ∀ r > 0 : µ(Br(x)) > 0(cid:9) .
Z f dµn →Z f dµ
(∀ f ∈ Cb(X; R)).
which is actually the convex hull of ∆X . The support (or spectrum) of µ ∈ PX is
the smallest d-closed set Sµ that satisfies µ(Sµ) = 1. Moreover, it is not hard to
verify the following equation:
The closure of a set H ⊆ X will be denoted by H.
We say that a sequence of measures {µn}∞
n=1 ⊂ PX converges weakly to a µ ∈ PX
if we have
This type of convergence is metrised by the L´evy–Prokhorov metric given by (1.1).
A map ϕ : PX → PX is called a π-isometry on PX if
π(µ, ν) = π(ϕ(µ), ϕ(ν))
(∀ µ, ν ∈ PX )
is satisfied.
Now, we are in the position to state the main result of this paper.
Main Theorem. Let (X, k · k) be a separable real Banach space and ϕ : PX → PX
be a surjective π-isometry. Then there exists a surjective affine isometry ψ : X → X
which induces ϕ, i.e. we have
(2.1)
(ϕ(µ)) (A) = µ(ψ−1[A])
(∀ A ∈ BX ),
where ψ−1[A] denotes the inverse-image set {ψ−1(a) a ∈ A}.
The converse of the above statement is trivial, namely, every transformation of
the form (2.1) is obviously an onto π-isometry. Note that our theorem can be re-
phrased in terms of push-forward measures. Namely, the action of ϕ is just the
push-forward with respect to the isometry ψ : X → X.
4
GY.P. GEH ´ER AND T. TITKOS
As we already mentioned, the L´evy–Prokhorov metric on PR differs from the
L´evy distance on PR. Therefore, in the special case when X = R, our hypotheses
are different from those given in [25, Theorem 1], although the conclusion is the
same.
Our proof is given in the next section where we will have four major steps. This
will be followed by some remarks in the final section, where we will also point out
that our Main Theorem still holds if we replace PX with an arbitrary weakly dense
subset S.
3. Proof
The proof is divided into four major steps. First, we will explore the action of ϕ
on ∆X . Then for finitely supported measures µ we will investigate the behaviour of
its image ϕ(µ) near to the vertices of the convex hull of Sµ. This will be followed by
providing a procedure which will allow us to obtain important information about
the "rest" of ϕ(µ). Finally, we close this section with the proof of the Main Theo-
rem. Note that although our main result deals with Borel probability measures on
separable Banach spaces, we state and prove some results in the context of complete
and separable metric spaces.
3.1. First major step: the action on Dirac measures. Here we will investigate
properties of the restricted map ϕDX . Namely, we will prove that ϕ maps ∆X
onto ∆X , furthermore, there is a surjective affine isometry of X which induces this
restriction. In order to do this first, we formulate the metric phrase "distance one"
by means of the supports of measures.
Proposition 3.1. Let (X, d) be a complete, separable metric space and µ, ν ∈ PX .
Then the following statements are equivalent:
(i) π(µ, ν) = 1,
(ii) d(Sµ, Sν) := inf {d(x, y) x ∈ Sµ, y ∈ Sν } ≥ 1,
(iii) Sν ∩ S1
(iv) Sµ ∩ S1
µ = ∅,
ν = ∅.
Proof. Observe that (ii) implies the following inequality for every 0 < ε < 1:
1 = µ(Sµ) > ν(Sε
µ) + ε = ε.
Consequently we have π(µ, ν) ≥ 1. But on the other hand, π(µ, ν) ≤ 1 holds for all
µ, ν ∈ PX , and therefore the (ii)⇒(i) part is complete.
To prove (i)⇒(ii) assume that := d(Sµ, Sν) < 1. In this case one can fix two
points x∗ ∈ Sµ and y∗ ∈ Sν, and a positive number r > 0 which satisfy both
≤ d(x∗, y∗) =: ′ < 1
and ′ + 2r < 1.
We also set t := min {µ (Br(x∗)) , ν (Br(y∗))} which is clearly positive by the very
definition of the support. We will show that ε := max{1 − t, ′ + 2r} < 1 is a
suitable choice to guarantee
µ(A) ≤ ν(Aε) + ε
(∀ A ∈ BX )
Indeed, if A ∈ BX satisfies µ(A) ≤ 1 − t, then
µ(A) ≤ 1 − t ≤ ν(A1−t) + 1 − t ≤ ν(Aε) + ε.
On the other hand, if µ(A) > 1 − t, then we observe that µ(A ∩ Br(x∗)) > 0,
and consequently A ∩ Br(x∗) is not empty. Let us fix a point z ∈ A ∩ Br(x∗).
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC
5
Using the triangle inequality we infer d(y∗, z) ≤ d(y∗, x∗) + d(x∗, z) < ′ + r and
Br(y∗) ⊆ B′+2r(z) ⊆ Aε. Therefore we conclude
µ(A) ≤ 1 ≤ t + ε ≤ ν(Br(y∗)) + ε ≤ ν(Aε) + ε,
which implies π(µ, ν) ≤ ε < 1.
The equivalence of (ii), (iii) and (iv) follows from the definitions.
(cid:3)
Next, let us define the unit distance set of a set of measures A ⊆ PX by
A =(cid:8)ν ∈ PX(cid:12)(cid:12) ∀ µ ∈ A : π(µ, ν) = 1(cid:9) .
(Remark that by definition we have ∅ = PX .) The following statement gives a
metric characterisation of Dirac measures when X is a separable real Banach space.
We point out that similar results were also crucial ideas in [15, 25, 26].
Proposition 3.2. Let (X, d) be a complete, separable metric space and µ ∈ PX be
an arbitrary Borel probability measure on it. Then the following three statements
are equivalent:
(i) ({µ} ) = {µ},
(ii) there exists an x ∈ X such that
(ii/a) µ = δx, and
(ii/b) B1(y) ⊆ B1(x) implies x = y for every y ∈ X,
(iii) # ({µ} ) = 1.
Proof. First, let us characterise the elements of ({µ} ) . It follows from Proposition
3.1 that
thus (ii/a) follows. On the other hand, if any y ∈ X satisfies
(cid:8)δz(cid:12)(cid:12) z ∈ Sµ(cid:9) ⊆ ({µ} ) ,
δy = B1(y) ⊆ B1(x) = S1
S1
δx ,
then again by (3.2) we infer x = y.
Finally, we show (ii)⇒(i). Assume that ϑ ∈ ({µ} ) . By (ii/b) we get that
S1
ϑ ⊆ S1
(cid:3)
δx holds if and only if Sϑ ⊆ {x}, which implies (i).
Remark 3.3. Note that if the diameter of the metric space X is less than 1, i.e.
there exists an 0 < r < 1 such that d(x, y) ≤ r (∀ x, y ∈ X), then π(µ, ν) ≤ r holds
for every µ, ν ∈ PX . In particular, {µ} = ∅ and thus ({µ} ) = PX for every
µ ∈ PX .
The following lemma describes the action of ϕ on Dirac measures.
Applying this observation twice, we easily see that
ν ∩ Sµ = ∅(cid:9) =(cid:8)ν ∈ PX(cid:12)(cid:12) Sν ∩ S1
{µ} =(cid:8)ν ∈ PX(cid:12)(cid:12) S1
({µ} ) = \ν∈{µ}
µ = ∅(cid:9) .
µ=∅(cid:8)ϑ ∈ PX(cid:12)(cid:12) Sν ∩ S1
(Note that if {µ} = ∅, thenTν∈∅{ν} = PX in (3.1) by definition.)
and therefore we obtain the following equivalence:
ϑ ∈ ({µ} )
⇐⇒
ϑ ⊆ S1
S1
µ.
We continue with proving (i)⇒(ii). Observe that (3.2) implies
{ν} =
\ν∈PX ,Sν ∩S1
(3.1)
(3.2)
ϑ = ∅(cid:9) ,
Now, since µ ∈ ({µ} ) always holds, the equivalence of (i) and (iii) is apparent.
6
GY.P. GEH ´ER AND T. TITKOS
Lemma 3.4. Let (X, k · k) be a separable real Banach space, and let ϕ : PX → PX
be a surjective π isometry. Then there exists a surjective affine isometry ψ : X → X
such that
(3.3)
ϕ(δx) = δψ(x)
(∀ x ∈ X).
Proof. Since ϕ is a bijective isometry, we have
ϕ(cid:0) ({µ} ) (cid:1) =(cid:0){ϕ(µ)} (cid:1)
(∀ µ ∈ PX ).
Thus an easy application of the previous proposition yields ϕ(∆X ) = ∆X . This also
means that there exists a bijective map ψ : X → X which induces the restriction
ϕ∆X , i.e.
(3.4)
ϕ(δx) := δψ(x)
(∀ x ∈ X).
We will show that ψ is an isometry. Observe that
π(δx1 , δx2) = min{1, kx1 − x2k}
(∀ x1, x2 ∈ X).
Therefore for all α ∈ (0, 1) we have
(3.5)
kψ(x1) − ψ(x2)k = α ⇐⇒ kx1 − x2k = α
(∀ x1, x2 ∈ X).
If X is one-dimensional, then it is rather easy to see that (3.5) implies the isometri-
ness of ψ. Now, assume that dim X ≥ 2. After suitable renorming (i.e. considering
α k · k), from a result of T.M. Rassias and P. Semrl [31, Theorem
the norm · := 1
1] we conclude that
kψ(x) − ψ(y)k = nα ⇐⇒ kx − yk = nα
(∀ α ∈ (0, 1), n ∈ N),
and therefore ψ is indeed an isometry. Finally, by the famous Mazur–Ulam theorem
we obtain that ψ is affine, which completes the proof.
(cid:3)
We remark that the last step of the proof (using the Rassias–Semrl theorem) can
be also done by the extension theorem of Mankiewicz [23].
In light of the above lemma, from now on we may and do assume without loss
of generality that ϕ acts identically on ∆X , i.e.,
(3.6)
ϕ(δx) = δx
(∀ x ∈ X),
and our aim will be to show that ϕ acts identically on the whole of PX . After we do
so, to obtain the result of our Main Theorem for general surjective π-isometries will
be straightforward. Namely, if (3.3) is fulfilled, then we can consider the following
modified transformation:
(3.7)
ϕψ : PX → PX ,
(ϕψ(µ)) (A) := (ϕ(µ)) (ψ[A])
(∀ µ ∈ PX , A ∈ B).
By our assumption, ϕψ fixes every element of ∆X and thus also of PX , which
implies (2.1).
Next, let us define the following continuous function for each µ ∈ PX :
Wµ : X → [0, 1], Wµ(x) := π(δx, µ)
which will be called the witness function of µ. The main advantage of the assump-
tion (3.6) is that the witness function becomes ϕ-invariant, i.e.
(3.8) Wµ(x) = π(δx, µ) = π(ϕ(δx), ϕ(µ)) = π(δx, ϕ(µ)) = Wϕ(µ)(x)
(∀ x ∈ X).
It is natural to expect that the shape of the witness function carries some informa-
tion about the measure. The last three major steps of the proof will be devoted to
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC
7
explore this for the ϕ-images of finitely supported measures in the setting of sepa-
rable Banach spaces. However, as demonstrated by the next example, the witness
function usually does not distinguish measures in general complete and separable
metric spaces.
Example 3.5. Consider the complete and separable metric space (X, d) with
X := {x1, x2, x3}
and
d(xi, xj) :=(cid:26) 1
3
0
if i 6= j
if i = j
.
Now, taking the limit of the right-hand side as h → 0+, and using that 0 < r < s
implies Br(x) ⊆ Bs(x), we obtain
ν k[j=1
BεA−δ(aj)! + εA − δ < µ ({a1, . . . , ak}) ≤ ν k[j=1
BεA (aj)! + εA
for every 0 < δ < εA, which proves (3.9).
(cid:3)
Note that the reason why we excluded the case when µ = ν in (3.9) is that then
for every A ⊆ Sµ and ε > 0 we have µ(A) ≤ ν(Aε) + ε, and for every ∅ 6= A ⊆ Sµ
we have µ(A) > 0 = ν(∅) + 0 = ν(A0) + 0. Of course if we had defined A0 to be
(3.9)
2 δx1 + 1
Let µ := 1
have π(δx, µ) = 1
2 δx2 and ν := 1
3 = π(δx, ν) for all x ∈ X and hence Wµ ≡ Wν .
2 δx2 + 1
2 δx3. An easy calculation shows that we
3.2. Second major step: isolated atoms on the vertices of the convex hull
of the support. Here we will prove that if µ is a finitely supported measure, and
x is a vertex of the convex hull of Sµ, then x is an isolated atom of ϕ(µ) and
We begin with a technical statement, which will be very useful in the sequel.
µ({x}) = (ϕ(µ))({x}).
Proposition 3.6. Let X be a separable real Banach space, and suppose that µ is
a finitely supported measure. Then for every ν ∈ PX, ν 6= µ we have
Therefore it is enough to show that for each subset A := {a1, . . . , ak} ⊆ Sµ if the
infimum
Proof. First, we observe that in (1.1) it is enough to consider Borel sets satisfying
A ⊆ Sµ. Furthermore, it is obvious that
εA := inf {ε > 0 µ(A) ≤ ν(Aε) + ε}
π(µ, ν) = min(cid:8)ε > 0 ∀ A ⊆ Sµ : µ(A) ≤ ν(Aε) + ε(cid:9) .
π(µ, ν) = max(cid:8)inf{ε > 0 µ(A) ≤ ν(Aε) + ε}(cid:12)(cid:12) A ⊆ Sµ(cid:9) .
Bε(aj)! + ε
= inf
ε > 0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
BεA+h(aj)! + εA + h.
µ({a1, . . . , ak}) ≤ ν k[j=1
BεA−h(aj)! + εA − h < µ ({a1, . . . , ak}) ≤ ν k[j=1
is positive, then it is actually a minimum if we take the closure of Aε instead of Aε.
The following line of inequalities holds for all 0 < h < εA by definition:
ν k[j=1
8
GY.P. GEH ´ER AND T. TITKOS
A, then (3.9) with ε ≥ 0 instead of ε > 0 would hold for the case µ = ν as well.
However, we prefer not to change usual notations.
The following proposition plays a key role in the proof. But before stating it we
introduce some notations. The convex hull of two points x and y will be denoted by
[x, y], and the symbol ]x, y[ will stand for the set [x, y] \ {x, y}. If f is a real valued
function on X and c ∈ R, then the sets {x ∈ X f (x) < c}, {x ∈ X f (x) = c}, and
{x ∈ X f (x) ≤ c} will be denoted by {f < c}, {f = c}, and {f ≤ c}, respectively.
Proposition 3.7. Let (X, k · k) be a separable real Banach space, µ ∈ FX \ ∆X and
K be the convex hull of Sµ. Assume that x is a vertex of K (which is a polytope)
and set λ := µ({x}) (for which we obviously have 0 < λ < 1). Then for every
ϑ ∈ PX with Sϑ ⊆ K the following two conditions are equivalent:
(i) ϑ = λδx + (1 − λ)eϑ where eϑ ∈ PX with Seϑ ⊆ K \ Br(x) for some r > 0,
(ii) there exist a number 0 < ρ ≤ 1 − λ and a half-line e starting from x such
that the restriction Wϑe is of the following form:
(3.10)
Wϑe(x) =
1
kx − xk
1 − λ
if
if
if
kx − xk ≥ 1,
1 − λ < kx − xk < 1,
1 − λ − ρ ≤ kx − xk ≤ 1 − λ.
Moreover, Sϕ(µ) ⊆ K and x is an isolated atom of ϕ(µ) with (ϕ(µ))({x}) = λ.
{f = c}
{f = c}
e
x
1
x
Y
Seϑ
K
Figure 1. Illustration for Proposition 3.7 on the finite dimen-
sional subspace Y . The support Sµ consists of the set of black
points in K.
Proof. First, we construct a half-line e which starts from x and satisfies
(3.11)
d ({x}, K) = kx − xk < kx − kk
(∀ x ∈ e, k ∈ K \ {x}).
Being the convex hull of a finite set, each vertex of K is strongly exposed, i.e. there
exists a continuous linear functional f ∈ X ∗ with
c := max{f (y) y ∈ K} = f (x) and K \ {x} ⊂ {f < c}.
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC
9
Let Y be the subspace generated by K. We fix an x ∈ Y such that x ∈ B1(x) and
B1(x) ∩ {f ≤ c} ∩ Y ⊆ {f = c}.
Note that as Y is finite dimensional, the existence of such an x is guaranteed.
Now, we define e to be the half-line starting from x and going through x.
It is
straightforward that e fulfils (3.11).
Next, we consider an arbitrary ϑ ∈ PX which satisfies (i). It is clear form the
compactness of Sµ and Sϑ, and the isolatedness of the point x in both Sµ and Sϑ,
that there is a number c < c such that
(3.12)
(Sµ ∪ Sϑ) \ {x} ⊆ {f < c}
holds. Consequently, for every x ∈ e we have
d ({x}, K) = kx − xk > d ({x}, {f ≤ c} ∩ Y ) > d ({x}, (Sµ ∪ Sϑ) \ {x}) .
Therefore, if x ∈ e and α := kx − xk > 0, then there exists a y ∈ ]x, x[ ⊆ e which
satisfies the following equations:
(3.13)
and
(3.14)
λ = ϑ({x}) = ϑ(Bα(x)) = ϑ(Bα(z))
(∀ z ∈ [x, y]).
λ = µ({x}) = µ(Bα(x)) = µ(Bα(z))
(∀ z ∈ [x, y]).
In fact, y can be chosen to be any point on ]x, x[ such that
kx − yk ≤ d ({f = c}, {f = c}) .
We proceed to prove the equivalence of (i) and (ii). Trivially, both (i) and (ii)
implies that ϑ /∈ ∆X , and therefore from now on we may and do assume that ϑ is
not a Dirac measure. Recall that according to Proposition 3.6 we have
(3.15)
Wϑ(x) = minnε > 0 1 ≤ ϑ(Bε(x)) + εo
(∀ x ∈ X).
If (i) holds, then by combining (3.15) with (3.11) and (3.13) we obtain that Wϑe
is of the form (3.10) with ρ := min(cid:16)1 − λ, d ({f = c}, {f = c})(cid:17).
Conversely, we suppose that ϑ ∈ PX , Sϑ ⊆ K and ϑ satisfies (ii). Let x1 and x2
be the points on e which satisfy kx1 − xk = 1 − λ and kx2 − xk = 1 − λ − ρ. By
(3.10) we have Wϑ(x1) = Wϑ(x2) = 1 − λ. Therefore on one hand, we obtain
1 ≤ ϑ(B1−λ(x1)) + 1 − λ = ϑ({x}) + 1 − λ,
from which λ ≤ ϑ({x}) follows. On the other hand,
1 > ϑ(B1−λ−δ(x2)) + 1 − λ − δ ≥ ϑ({x}) + 1 − λ − δ
(∀ 0 < δ < ρ)
is satisfied. Hence we infer ϑ({x}) < λ + δ for every δ > 0, and thus trivially
ϑ({x}) = λ holds. But we also observe the following:
1 > ϑ(cid:16)B1−λ−δ(x2)(cid:17) + 1 − λ − δ ≥ ϑ(cid:16)B1−λ−
ρ
2
(x2)(cid:17) + 1 − λ − δ
which implies
(cid:0)∀ 0 < δ < ρ
2(cid:1)
whence we conclude that x is indeed an isolated atom of ϑ.
λ ≥ ϑ(cid:16)B1−λ−
ρ
2
(x2)(cid:17) ≥ ϑ ({x}) = λ,
10
GY.P. GEH ´ER AND T. TITKOS
For the last statement first, by Proposition 3.1 we infer π(δx, µ) = 1 for every
x /∈ K 1. The ϕ-invariance of the witness function gives
1 = π(δx, ϕ(µ))
(x ∈ X \ K 1),
and hence, again by Proposition 3.1, we conclude
Sϕ(µ) ∩ B1(x) = ∅
(x ∈ X \ K 1).
Consequently, we obtain
Sϕ(µ) ⊆ X \ (X \ K 1)1 ⊆ K.
Finally, an application of the equivalence of (i) and (ii) gives the rest.
(cid:3)
We have the following consequence.
Corollary 3.8. Let (X, k · k) be a separable real Banach space. If µ ∈ FX such
that #Sµ ≤ 2, then ϕ(µ) = µ. Moreover, if ν ∈ PX with Wµ ≡ Wν , then µ and ν
coincide.
According to the above results, now we know that ϕ fixes every measure which
has at most two points in its support. Although we are expecting the same for all
µ ∈ FX , right now we only have some information about the behaviour of ϕ(µ)
near to the vertices of the convex hull of its support.
3.3. Third major step: the story beyond vertices. Here we show a procedure
how the behaviour of ϕ(µ) can be completely explored in case when µ is a finitely
supported measure. In order to do so, we need to introduce some technical nota-
tions. Let s > 0 be a positive parameter and define the s-L´evy–Prokhorov distance
πs : PX × PX → [0, 1] by the following formula:
(3.16)
πs(µ, ν) := inf {ε > 0 ∀ A ∈ BX : s · µ(A) ≤ s · ν(Aε) + ε} (∀ µ, ν ∈ PX ).
Note that although we do not know at this point whether πs defines a metric on
PX, we will see this later. The s-witness function (or modified witness function) of
µ ∈ PX is defined by
Ws,µ : X → R, Ws,µ(x) := πs(δx, µ).
Obviously, if we set s = 1, then we get the original L´evy–Prokhorov metric and
witness function.
In the next two lemmas we collect some properties of the s-L´evy–Prokhorov
distance analogous to those provided in the previous major step.
Lemma 3.9. Let (X, k · k) be a separable real Banach space and s > 0. Then
(PX , πs) is a metric space. Furthermore, for every µ ∈ FX and ν ∈ PX with ν 6= µ
we have
πs(µ, ν) = min(cid:8)ε > 0(cid:12)(cid:12) ∀ A ⊆ Sµ : s · µ(A) ≤ s · ν(Aε) + ε(cid:9) .
Proof. For the sake of clarity, let us use more detailed notations here. If k · k is a
norm on X, then denote by Aε,k·k and πs,k·k the open ε-neighborhood of A and the
s-L´evy–Prokhorov metric with respect to k · k, respectively. Observe that the Borel
σ-algebras of (X, k · k) and(cid:0)X, 1
elementary computation we have Asδ,k·k = Aδ, 1
s k · k(cid:1) coincide as the norms are equivalent. By an
s k·k, which yields
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC 11
(3.17)
πs,k·k(µ, ν) = infnε > 0(cid:12)(cid:12)(cid:12) ∀ A ∈ BX : s · µ(A) ≤ s · ν(cid:0)Aε,k·k(cid:1) + εo
= infnsδ > 0(cid:12)(cid:12)(cid:12) ∀ A ∈ BX : s · µ(A) ≤ s · ν(cid:0)Asδ,k·k(cid:1) + sδo
= s · infnδ > 0(cid:12)(cid:12)(cid:12) ∀ A ∈ BX : µ(A) ≤ ν(cid:0)Asδ,k·k(cid:1) + δo
= s · infnδ > 0(cid:12)(cid:12)(cid:12) ∀ A ∈ BX : µ(A) ≤ ν(cid:0)Aδ, 1
s k·k(cid:1) + δo
s k·k(µ, ν)
= s · π 1
for every µ, ν ∈ PX . In particular, πs is a metric on PX , and using the formula
(3.9) completes the proof.
(cid:3)
We omit the proof of the following lemma as it is a straightforward consequence
of (3.17).
Lemma 3.10. Let (X, k · k) be a separable real Banach space, µ ∈ FX \ ∆X and
s > 0. Let us denote the convex hull of Sµ by K, and assume that x is a vertex of
K. Set λ := µ({x}) ∈ (0, 1). Then for every ϑ ∈ PX with Sϑ ⊆ K the following
two conditions are equivalent:
that Ws,ϑe has the following form:
(ii) there exist a number 0 < ρ ≤ s(1 − λ) and a half-line e starting from x such
(i) ϑ = λδx + (1 − λ)eϑ where eϑ ∈ PX with Seϑ ⊆ K \ Br(x) for some r > 0,
Ws,ϑe(x) =
kx − xk ≥ s,
s(1 − λ) < kx − xk < s,
s(1 − λ) − ρ ≤ kx − xk ≤ s(1 − λ).
As a consequence we have that if µ ∈ FX, #Sµ ≤ 2, ν ∈ PX and Ws,µ ≡ Ws,ν ,
then µ = ν.
s
kx − xk
s(1 − λ)
if
if
if
Next, let us suppose for a moment that m ∈ N pieces of atoms of ϑ ∈ PX have
been already detected. (For instance by Lemma 3.7, if ϑ = ϕ(µ) with µ ∈ FX
then the atoms of ϑ in the vertices of the convex hull of Sϑ can be detected.) Our
aim with the forthcoming lemma is to describe a modified witness function of the
remaining part of ϑ in terms of the (original) L´evy–Prokhorov distances between
ϑ and some measures which are supported on at most m + 1 points. This will be
later utilised in order to explore the action of ϕ on FX.
Lemma 3.11. Let (X, k · k) be a separable real Banach space and ϑ ∈ PX. Let
x ∈ X and {yj,l 1 ≤ j ≤ k, 1 ≤ l ≤ dj} ⊂ X be some pairwise different points such
that
ρj := kx − yj,1k = kx − yj,lk
(∀ 1 ≤ l ≤ dj)
holds for every 1 ≤ j ≤ k,
ρj > ρj+1 > 0
(∀ 1 ≤ j ≤ k − 1),
and
We also set
wj,l := ϑ({yj,l}) > 0
(∀ 1 ≤ j ≤ k, 1 ≤ l ≤ dj).
wj :=
djXl=1
wj,l = ϑ({yj,1, . . . , yj,dj })
(∀ 1 ≤ j ≤ k),
12
GY.P. GEH ´ER AND T. TITKOS
wj
(∀ 0 ≤ r ≤ k).
and
(3.18)
ηr :=
rXj=1
djXl=1
kXj=1
ew := 1 −
wj(cid:19) · δx ∈ FX
wj,l · δyj,l +(cid:18)1 −
rXj=1
Furthermore, denote by eϑ ∈ PX the measure which satisfies
wj,l · δyj,l + ew ·eϑ.
Then the ew-witness function of eϑ can be expressed in terms of the L´evy–Prokhorov
ew,eϑ(x) =
if x is not (P1)
if x is (Pr) but not (Pr+1) with some 1 ≤ r < k
if x is (Pk)
distances of ϑ and ηr's in the following way:
(3.20)
where for every 1 ≤ r ≤ k the property (Pr) means
π(δx, ϑ)
π(ηr, ϑ)
π(ηk, ϑ)
ϑ =
kXj=1
djXl=1
(3.19)
W
(Pr)
π(ηr−1, ϑ) ≤ ρr.
y1,1
y1,2
y2,1
ρ1
y2,4
ρ3
ρ2
x
y3,1
y2,2
y3,3
y3,2
y2,3
Figure 2. An illustration when X = R2 with the ℓ∞-norm.
Remark 3.12. It is extremely important to observe that the subscripts in the
lemma above highly depend on the actual position of x. For instance on Figure 1
with that particular x we have k = 3. However, if x is moved slightly to the right,
then k becomes 7. In particular, this changes (Pr) and therefore (3.20) as well.
Proof of Lemma 3.11. We split our proof into five parts.
Part 1. First, we prove that for each 1 ≤ r ≤ k and 0 < ε < ρr we have
(3.21)
if and only if
(3.22)
ηr(A) ≤ ϑ(Aε) + ε
(∀ A ⊆ Sηr )
ηr({x}) ≤ ϑ(Bε(x)) + ε
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC 13
is satisfied. One direction is obvious. In order to see the reverse implication, observe
that (3.21) holds trivially if x /∈ A. On the other hand, if x ∈ A, then (3.22) yields
(3.21) for this A by the following estimation:
ηr(A) = ηr({x}) + ηr(A \ {x}) ≤ ϑ(Bε(x)) + ε + ϑ(A \ {x}) ≤ ϑ(Aε) + ε.
Part 2. Here we show that the right-hand side of (3.20) is well defined. First,
we observe that by Proposition 3.6 x is (Pr) if and only if
(3.23)
ηr−1(A) ≤ ϑ(Aρr ) + ρr
(∀ A ⊆ Sηr−1 ).
But by Part 1, this is equivalent to the following inequality:
(P ′
r)
1 −
r−1Xj=1
wj ≤ ϑ(Bρr (x)) + ρr.
Next, let 2 ≤ r ≤ k. In order to see the well-definedness, it is enough to show
that if x is (Pr), then x is also (Pr−1). So assume that x is (Pr). Since ρr < ρr−1
and
wr−1 =
dr−1Xl=1
wr−1,l = ϑ({yr−1,1, . . . yr−1,dr−1}) ≤ ϑ(cid:16)Bρr−1 (x) \ Bρr (x)(cid:17) ,
we obtain
(P ′
r−1)
1 −
r−2Xj=1
wj ≤ ϑ(Bρr−1 (x)) + ρr−1.
Therefore x is indeed (Pr−1).
Part 3. Next, we verify (3.20) in case when x is not (P1), i.e. π(δx, ϑ) > ρ1.
Observe that since
wi + ρ1,
kXi=1
1 > ϑ(Bρ1 (x)) + ρ1 = ew ·eϑ(Bρ1 (x)) +
ew > ew ·eϑ(Bρ1 (x)) + ρ1,
we have
and thus
follows. Using this fact we obtain
π ew(δx,eϑ) > ρ1
ε > ρ1(cid:12)(cid:12)(cid:12)(cid:12) 1 ≤
π ew(δx,eϑ) = minnε > ρ1(cid:12)(cid:12)ew ≤ ew ·eϑ(Bε(x)) + εo
= min
wj + ew ·eϑ(Bε(x)) + ε
= minnε > ρ1(cid:12)(cid:12) 1 ≤ ϑ(Bε(x)) + εo
= minnε > 0(cid:12)(cid:12) 1 ≤ ϑ(Bε(x)) + εo = π(δx, ϑ),
kXj=1
which completes this part.
14
GY.P. GEH ´ER AND T. TITKOS
Part 4. We proceed to show (3.20) in the case when x is (Pr) but not (Pr+1) with
some 1 ≤ r < k. As in the previous part, first we estimate the value of π ew(δx,eϑ).
According to the re-phrasing (P ′
r) and the assumption, we have
1 −
wj ≤ ϑ(Bρr (x)) + ρr =
r−1Xj=1
wj > ϑ(Bρr+1 (x)) + ρr+1 =
kXi=r
kXi=r+1
wi + ew ·eϑ(Bρr (x)) + ρr
wi + ew ·eϑ(Bρr+1 (x)) + ρr+1.
and
1 −
rXj=1
Observe that these inequalities are equivalent to
and
(3.24)
respectively. Thus we conclude that
π ew(δx,eϑ) = minnρr+1 < ε ≤ ρr(cid:12)(cid:12)ew ≤ ew ·eϑ(Bε(x)) + εo .
ew ≤ ew ·eϑ(Bρr (x)) + ρr
ew > ew ·eϑ(Bρr+1 (x)) + ρr+1,
ρr+1 < π ew(δx,eϑ) ≤ ρr.
In particular, eϑ is different from δx, and we have
From now on we consider two cases: (a) when ρr+1 < π ew(δx,eϑ) < ρr, and (b)
when π ew(δx,eϑ) = ρr. Assume first that (a) is fulfilled. Then (3.24) becomes
wi + ew ·eϑ(Bε(x)) + ε)
π ew(cid:16)δx,eϑ(cid:17) = min(ρr+1 < ε < ρr(cid:12)(cid:12)(cid:12) 1 −
wi ≤ ϑ(Bε(x)) + ε)
= min(ρr+1 < ε < ρr(cid:12)(cid:12)(cid:12) 1 −
= minnρr+1 < ε < ρr(cid:12)(cid:12) ηr({x}) ≤ ϑ({x}ε) + εo
= min(cid:8)ρr+1 < ε < ρr(cid:12)(cid:12) ∀ A ⊆ Sηr : ηr(A) ≤ ϑ(Aε) + ε(cid:9) (by Part 1)
= min(cid:8)ε > 0 ∀ A ⊆ Sηr : ηr(A) ≤ ϑ(Aε) + ε(cid:9)
rXi=1
rXi=1
kXi=r+1
= π(ηr, ϑ),
wi ≤
which is exactly the desired equation. Second, suppose that (b) is satisfied. Con-
sequently, we have
(∀ ρr+1 < ε < ρr),
1 −
kXj=1
whence
ηr({x}) = 1 −
wj = ew > ew ·eϑ(Bε(x)) + ε
rXi=1
kXi=r+1
wi >
π(ηr, ϑ) ≥ ρr.
follows for every ρr+1 < ε < ρr. In particular, we get
wk + ew ·eϑ(Bε(x)) + ε = ϑ(Bε(x)) + ε
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC 15
Finally, we verify that the converse inequality holds as well. Suppose indirectly
that there exists an A ⊆ Sηr such that
ηr(A) > ϑ(Aρr ) + ρr.
Clearly, x /∈ A contradicts the above inequality, thus x ∈ A follows. Therefore we
have
ηr(A) > ϑ(Aρr ) + ρr ≥ ϑ(A ∪ Bρr (x)) + ρr
≥ ϑ(A \ {x, yr,1, . . . , yr,dr }) + ϑ(Bρr (x)) + ρr
= ηr(A \ {x, yr,1, . . . , yr,dr}) + ϑ(Bρr (x)) + ρr.
Consequently,
(3.25)
1 −
r−1Xj=1
wj = ηr({x, yr,1, . . . , yr,dr})
≥ ηr(A) − ηr(A \ {x, yr,1, . . . , yr,dr}) > ϑ(Bρr (x)) + ρr,
which contradicts (Pr). This completes the present part.
Part 5. Finally, we prove (3.20) when x is (Pk), i.e. π(ηk−1, ϑ) ≤ ρk. We have
to show that π(ηk, ϑ) = π ew(δx,eϑ). Because of the assumption, we have
wj ≤ ϑ(Bρk (x)) + ρk = wk + ew ·eϑ(Bρk (x)) + ρk
k−1Xj=1
1 −
which implies ew ≤ ew ·eϑ(Bρk (x)) + ρk, and hence,
π ew(δx,eϑ) ≤ ρk.
We consider three cases: (a) when 0 < π ew(δx,eϑ) < ρk, (b) when π ew(δx,eϑ) = ρk,
and (c) when π ew(δx,eϑ) = 0. First, let us suppose (a). In this case we have
π ew(δx,eϑ) = minn0 < ε < ρk(cid:12)(cid:12)ew ≤ ew ·eϑ(Bε(x)) + εo
= minn0 < ε < ρk(cid:12)(cid:12) ηk({x}) ≤ ϑ(Bε(x)) + εo
= min(cid:8)0 < ε < ρk ∀ A ⊆ Sηk : ηk(A) ≤ ϑ(Aε) + ε(cid:9) (by Part 1)
= π(ηk, ϑ).
Second, we assume (b). Let us observe the following for every ε < ρk:
which implies π(ηk, ϑ) ≥ ρk. To show the converse inequality, i.e. π(ηk, ϑ) ≤ ρk,
assume indirectly that there exists an A ⊆ Sηk such that
ηk({x}) = ew > ew ·eϑ(Bε(x)) + ε = ϑ(Bε(x)) + ε,
ηk(A) > ϑ(Aρk ) + ρk.
Very similarly, as in the verification of (3.25), we conclude that this inequality
contradicts (Pk). Finally, the case (c) is trivial.
(cid:3)
Since the modified witness function is obviously continuous, we also know the
ew,eϑ(x) when x ∈ {yj,l 1 ≤ j ≤ k, 1 ≤ l ≤ dj}. Therefore if m pieces of
value of W
atoms of ϑ ∈ PX have been already detected, then a modified witness function of
16
GY.P. GEH ´ER AND T. TITKOS
the remaining part of ϑ can be calculated in terms of the L´evy–Prokhorov distances
between ϑ and some measures supported on a set of at most m + 1 points.
3.4. Final major step: the action on FX and PX . Now, we are in the position
to verify our main result.
Proof of Main Theorem. Recall that we assumed (3.6) and that our aim is to show
that ϕ is the identity map. Observe that it is enough to prove that ϕ acts identically
on FX , as FX is a weakly dense subset of PX and ϕ is continuous. In order to do
this we use induction on the cardinality of the support of µ ∈ FX . By Corollary 3.8
our map ϕ fixes all measures with an at most two-element support. Let k ∈ N, k ≥ 2
and assume that we had already proved the following:
(3.26)
ϕ(ν) = ν
(∀ ν ∈ PX , #Sν ≤ k).
Let us consider a measure
k+1Xi=1
where the xi's are pairwise different,Pk+1
µ =
also that for every 1 ≤ i ≤ k the point xi lies outside of the convex hull of {xj}k+1
Let us use the following notations in the sequel:
i=1 λi = 1 and each λi is positive. Assume
j=i+1.
λiδxi ∈ FX
ϑ := ϕ(µ)
and µ(i) :=
1
Pk+1
j=i+1 λj
·
k+1Xj=i+1
λj δxj
(0 ≤ i ≤ k).
By Proposition 3.7 we observe that the support of ϑ is contained in the convex hull
of Sµ = {xj}k+1
j=1 .
Now, we prove step by step that each xi is an atom of ϕ(µ) with the same weight
λi. By (3.8) we have Wµ ≡ Wϑ, thus an application of Proposition 3.7 gives
ϑ = λ1 · δx1 + (1 − λ1) · ϑ(1),
with a measure ϑ(1) ∈ PX such that x1 /∈ Sϑ(1) and Sϑ(1) lies in the convex hull of
Sµ = {xj}k+1
j=1 . Utilising Lemma 3.11 and (3.26) for measures with supports of at
most 2 elements we obtain
W1−λ1 ,µ(1) ≡ W1−λ1,ϑ(1) .
At this point, if k was 2, then #Sµ(1) = 2, thus by Lemma 3.10 the measures µ(1)
and ϑ(1) coincide, and therefore µ = ϕ(µ) is yielded. Otherwise, applying Lemma
3.10 for the measures µ(1) and ϑ(1) gives
ϑ = λ1 · δx1 + λ2 · δx2 + (1 − λ1 − λ2) · ϑ(2),
with a measure ϑ(2) ∈ PX such that x2 /∈ Sϑ(2) and Sϑ(2) lies in the convex hull of
Sµ(1) = {xj}k+1
j=2 . Using Lemma 3.11 and (3.26) for the case when the cardinality
of the support is at most 3, we obtain
Iterating this procedure, the conclusion of the (k − 2)nd step is the following:
W1−λ1−λ2,µ(2) ≡ W1−λ1−λ2,ϑ(2) .
(3.27)
W1−Pk−2
i=1 λi,µ(k−2) ≡ W1−Pk−2
i=1 λi,ϑ(k−2)
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC 17
where
ϑ =
k−2Xi=1
λiδxi +(cid:18)1 −
λi(cid:19) · ϑ(k−2)
k−2Xi=1
such that ϑ(k−2) ∈ PX , xk−2 /∈ Sϑ(k−2) and Sϑ(k−2)
lies in the convex hull of
Sµ(k−3) = {xk−2, xk−1, xk, xk+1}. Utilising Lemma 3.10 for the measures µ(k−2)
and ϑ(k−2) we get that
ϑ =
λiδxi +(cid:18)1 −
k−1Xi=1
λi(cid:19) · ϑ(k−1)
k−1Xi=1
with some ϑ(k−1) ∈ PX , xk−1 /∈ Sϑ(k−1) and Sϑ(k−1)
Sµ(k−2) = {xk−1, xk, xk+1}. Furthermore, by Lemma 3.11 and (3.26) we obtain
(3.28)
lies in the convex hull of
i=1 λi,µ(k−1) ≡ W1−Pk−1
i=1 λi,ϑ(k−1) .
W1−Pk−1
But since #Sµ(k−1) = 2, Lemma 3.10 and (3.28) imply µ(k−1) = ϑ(k−1), and there-
fore we conclude ϕ(µ) = µ, completing the proof.
(cid:3)
4. Concluding remarks
We noted at the end of Section 2 that it is possible to give a characterisation of
surjective π-isometries on certain subsets of PX . Namely, let S ⊂ PX be a weakly
dense subset (possibly disjoint from ∆X ), and assume that φ : S → S is onto and
satisfies
π(φ(µ), φ(ν)) = π(µ, ν)
(∀ µ, ν ∈ S).
Since (PX , π) is a complete metric space, there exists a unique isometric extension
ϕ : PX → PX , i.e. ϕPX = φ. Clearly, ϕ is a π-isometry which maps PX into PX .
Observe that ϕ[PX ] is closed in PX . On the other hand, as S = ϕ[S] ⊂ ϕ[PX ],
we infer ϕ[PX ] = PX . Therefore ϕ : PX → PX is induced by a surjective isometry
ψ : X → X, whence we conclude the same for φ : S → S, i.e.
(φ(µ)) (A) = µ(ψ−1[A])
(∀ µ ∈ S, A ∈ BX ).
We proceed to mention some typical examples of weakly dense subsets of PX
(for which the above statement holds). 1) The set of all discrete Borel probability
measures, which is the collection of those µ ∈ PX that are concentrated on a
countable subset of X. 2) The class of all continuous Borel probability measures,
i.e. those µ ∈ PX such that µ({x}) = 0 for every x ∈ X. 3) Let n ∈ N, X = Rn
and k · k be an arbitrary norm on Rn. Since any two norms on Rn are equivalent,
the Borel σ-algebra BRn does not depend on k · k. We say that µ ∈ PRn is an
absolutely continuous Borel probability measure if it is absolutely continuous with
respect to the usual Lebesgue measure on Rn. This set is clearly weakly dense in
PRn, as every element of FX can be approximated.
Next, as we have mentioned in the introduction, the most important special cases
of our result are the following: 1) when X is an infinite dimensional, separable real
Hilbert space; 2) when X is the real Banach space C([0, 1]); and 3) when X is
an n-dimensional Euclidean space (n ∈ N). We make some comments on how
our proof could be modified in these cases. In the first two cases the underlying
Banach spaces are of infinite dimension, hence the support of any µ ∈ FX lies in
a finite dimensional affine subspace. In case of 1) the equivalence in Proposition
3.7 can be done for every element x in Sµ by choosing a half-line e orthogonal to
18
GY.P. GEH ´ER AND T. TITKOS
that affine subspace. Therefore the proof becomes much simpler as we immediately
obtain that every µ ∈ FX is fixed by ϕ. A similar argument simplifies the proof
for general strictly convex infinite dimensional separable Banach spaces. In case of
2) the space is of infinite dimension but the norm is not strictly convex. Despite
of this obstacle the proof still can be shortened by utilising the Lindenstrauss–
Troyansky theorem [12, 22, 32]. Namely, if µ ∈ FX and Sµ is contained in the
kernel of a strongly exposing functional (for the definition see e.g. [12]), then the
equivalence part of Proposition 3.7 can be verified for every element x in Sµ. Since
by the Lindenstrauss–Troyansky theorem it is easy to see that every µ ∈ FX can be
weakly approximated by such measures, we easily complete the proof of the Main
Theorem in this case too. It seems that for finite dimensional spaces, even for the
case of 3), we really have to do the whole procedure presented in Section 3, or at
least we are not aware of any shortening possibilities.
Finally, we note that throughout Section 3 there were some parts where we
considered general complete and separable metric spaces. But later on most of our
techniques required that the underlying space had a linear structure. In our opinion
it would be interesting to find a characterisation of all surjective π-isometries in the
setting of other special (but still general enough) kinds of complete separable metric
spaces.
References
[1] W. Ambrose, On measurable stochastic processes, Trans. Amer. Math. Soc. 47 (1940),
66–79.
[2] J. Araujo, and J.J. Font, Linear isometries between subspaces of continuous functions,
Trans. Amer. Math. Soc. 349 (1997), 413–428.
[3] J. Araujo, and J.J. Font, Linear isometries on subalgebras of uniformly continuous functions,
Proc. Edinburgh Math. Soc. (2) 43 (2000), 139–147.
[4] J. Araujo, The noncompact Banach–Stone theorem, J. Operator Theory 55 (2006), 285–
294.
[5] M. Bachir, Remarks on isometries of products of linear spaces, Extracta Math. 30 (2015),
1–13.
[6] Ben Yaacov I., and Melleray J., Isometrisable group actions, Proc. Amer. Math. Soc. 144
(2016), 4081–4088.
[7] J. Bertrand, and B.R. Kloeckner, A geometric study of Wasserstein spaces: Hadamard
spaces, J. Topol. Anal. 4 (2012), 515–542.
[8] J. Bertrand, and B.R. Kloeckner, A geometric study of Wasserstein spaces: isometric rigid-
ity in negative curvature, Int. Math. Res. Not. IMRN (2016), 1368–1386.
[9] P. Billingsley, Convergence of Probability Measures, Second edition, John Wiley & Sons,
Inc., New York, 1999.
[10] D.P. Blecher, and L.E. Labuschagne, Logmodularity and isometries of operator algebras
Trans. Amer. Math. Soc. 355 (2003), 1621–1646 (electronic).
[11] F. Botelho, J. Jamison, and L. Moln´ar, Surjective isometries on Grassmann spaces, J. Funct.
Anal. 265 (2013), 2226–2238.
[12] J. Bourgain, Strongly exposed points in weakly compact convex sets in Banach spaces, Proc.
Amer. Math. Soc. 58 (1976), 197–200.
[13] Cheng L., Cheng Q., Tu K., and Zhang J., A universal theorem for stability of ε-isometries
of Banach spaces, J. Funct. Anal. 269 (2015), 199–214.
[14] J.L. Doob, Stochastic processes depending on a continuous parameter, Trans. Amer. Math.
Soc. 42 (1937), 107–140.
[15] G. Dolinar, and L. Moln´ar, Isometries of the space of distribution functions with respect to
the Kolmogorov–Smirnov metric, J. Math. Anal. Appl. 348 (2008), 494–498.
[16] R.J. Fleming, and J.E. Jamison, Isometries on Banach Spaces: Function Spaces, Chapman
& Hall/CRC, Boca Raton, FL, 2003.
A CHAR. OF ISOMETRIES WITH RESPECT TO THE L ´EVY–PROKHOROV METRIC 19
[17] R.J. Fleming, and J.E. Jamison, Isometries on Banach spaces. Vol. 2. Vector-valued func-
tion spaces, Chapman & Hall/CRC, Boca Raton, FL, 2008.
[18] Gy.P. Geh´er, and P. Semrl, Isometries of Grassmann spaces, J. Funct. Anal. 270 (2016),
1585–1601.
[19] P.J. Huber, Robust Statistics, John Wiley & Sons, Inc., New York, 1981.
[20] B. Kloeckner, A geometric study of Wasserstein spaces: Euclidean spaces, Ann. Sc. Norm.
Super. Pisa Cl. Sci. (5) 9 (2010), 297–323.
[21] E. Le Donne, and A. Ottazzi, Isometries of Carnot groups and sub-Finsler homogeneous
manifolds, J. Geom. Anal. 26 (2016), 330–345.
[22] J. Lindenstrauss, On operators which attain their norm, Israel J. Math. 1 (1963), 139–148.
[23] P. Mankiewicz, On extension of isometries in normed linear spaces, Bull. Acad. Polon. Sci.
S´er. Sci. Math. Astronom. Phys. 20 (1972), 367–371.
[24] M. Mart´ın, The group of isometries of a Banach space and duality, J. Funct. Anal. 255
(2008), 2966–2976.
[25] L. Moln´ar, L´evy isometries of the space of probability distribution functions, J. Math. Anal.
Appl. 380 (2011), 847–852.
[26] L. Moln´ar, Kolmogorov–Smirnov isometries and affine automorphisms of spaces of distri-
bution functions, Cent. Eur. J. Math. 9 (2011), 789–796.
[27] D. Monclair, Isometries of Lorentz surfaces and convergence groups, Math. Ann. 363 (2015),
101–141.
[28] P. Niemiec, Isometry groups of proper metric spaces, Trans. Amer. Math. Soc. 366 (2014),
2597–2623.
[29] K.R. Parthasarathy, Probability Measures on Metric Spaces, Academic Press, Inc., New
York-London 1967.
[30] Yu.V. Prokhorov, Convergence of random processes and limit theorems in probability the-
ory, Theory Probab. Appl. 1 (1956), 157–214.
[31] T.M. Rassias, and P. Semrl, On the Mazur–Ulam theorem and the Aleksandrov problem
for unit distance preserving mappings, Proc. Amer. Math. Soc. 118 (1993), 919–925.
[32] S.L. Troyanski, On locally uniformly convex and differentiable norms in certain non-
separable Banach spaces, Studia Math. 37 (1970/71), 173–180.
[33] C. Villani, Optimal Transport: Old and New, Springer-Verlag, Berlin, 2009.
Gyorgy P´al Geh´er
MTA-SZTE Analysis and Stochastics Research Group
Bolyai Institute, University of Szeged
Aradi v´ertan´uk tere 1., Szeged H-6720, Hungary
MTA-DE "Lendulet" Functional Analysis Research Group
Institute of Mathematics, University of Debrecen
P.O. Box 12, Debrecen H-4010, Hungary
E-mail address: [email protected] or [email protected]
Tam´as Titkos
Alfr´ed R´enyi Institute of Mathematics, Hungarian Academy of Sciences
Re´altanoda utca 13-15., Budapest H-1053, Hungary
E-mail address: [email protected]
|
1606.05736 | 4 | 1606 | 2019-04-09T04:44:56 | Absolutely minimum attaining closed operators | [
"math.FA",
"math.OA",
"math.SP"
] | We define and discuss properties of the class of unbounded operators which attain minimum modulus. We establish a relationship between this class and the class of norm attaining bounded operators and compare the properties of both. Also we define absolutely minimum attaining operators (possibly unbounded) and characterize injective absolutely minimum attaining operators as those with compact generalized inverse. We give several consequences, one of them is that every such operator has a non trivial hyperinvariant subspace. | math.FA | math |
ABSOLUTELY MINIMUM ATTAINING CLOSED OPERATORS
S. H. KULKARNI1 AND G. RAMESH 2∗
Abstract. We define and discuss properties of the class of unbounded op-
erators which attain minimum modulus. We establish a relationship between
this class and the class of norm attaining bounded operators and compare the
properties of both. Also we define absolutely minimum attaining operators
(possibly unbounded) and characterize injective absolutely minimum attaining
operators as those with compact generalized inverse. We give several con-
sequences, one of those is that every such operator has a non trivial hyper
invariant subspace.
1. Introduction
The class of norm attaining operators on Banach spaces is well studied by
several authors in the literature.
It is known that the class of norm attaining
operators is dense in the space of all bounded linear operators on a Hilbert space
with respect to the operator norm [18, Theorem 1]. For more details on norm
attaining operators on Banach spaces, we refer to [19, 2] and [14] and references
therein.
Every compact operator is norm attaining. In fact, restricted to any non zero
closed subspace of a Hilbert space, it remains as compact and hence norm attain-
ing. Motivated by this observation, Carvajal and Neves [12] introduced a class
of operators, called the absolutely norm attaining operators. Characterization of
such operators on separable Hilbert space, in a particular case is given in [21] and
a complete characterization on arbitrary Hilbert space is discussed in [28]. Many
properties of these operators resemble the properties of compact operators.
It is a natural question to ask what happens if the norm is replaced by the
minimum modulus. This leads to the definition of minimum attaining operators.
Analogously, we can define absolutely minimum attaining operators. In a recent
paper Carvajal and Neves [13], studied bounded operators between two different
Hilbert spaces having such property. The structure of positive absolutely mini-
mum attaining operators is described in [16]. This concept is also applicable to
linear operators, that are not bounded.
In this article, we introduce the minimum attaining property for densely defined
closed operators (possibly not bounded). We prove several characterizations of
such operators. We also prove the dual relation between the norm attaining
Date: April 10, 2019.
∗Corresponding author.
2000 Mathematics Subject Classification. 47A75, 47A05, 47A10, 47A15 .
Key words and phrases. closed operator, minimum modulus, absolutely minimum attaining
operator, invariant subspace, Lomonosov theorem, generalized inverse.
1
2
S. H. KULKARNI, G. RAMESH
bounded operators and the minimum attaining closed operators. Finally, we
introduce the absolutely minimum attaining operators and prove a representation
theorem for the injective absolutely minimum attaining operators. Furthermore,
we observe that this class is exactly the same as the class of densely defined
closed operators whose generalized inverse is compact. Finally, we show that
these operators possess a non trivial hyperinvariant subspace.
We organize the article as follows: In the second section we provide basic re-
sults which will be used throughout the article. In the third section, we define
minimum attaining property for densely defined closed operators and prove sev-
eral characterizations. Some of the results in this section generalize the existing
results of bounded operators and some of them are new. In the fourth section, we
define absolutely minimum attaining operators and show that all such operators
have a closed range.
In particular, we show that an injective densely defined
closed operator is absolutely minimum attaining if and only if its Moore-Penrose
inverse is compact. Using this result, we deduce several consequences. One of
the important consequences is that every such operator has a non trivial hyper
invariant subspace.
2. Preliminaries
In this section we introduce some basic notations, definitions and results that
are needed to prove our main results.
Throughout the article we consider infinite dimensional complex Hilbert spaces
which will be denoted by H, H1, H2 etc. The inner product and the induced
norm are denoted by h·i and ., respectively. Let T be a linear operator with
domain D(T ), a subspace of H1 and taking values in H2. If D(T ) is dense in
H1, then T is called a densely defined operator. The graph G(T ) of T is defined
by G(T ) := {(T x, x) : x ∈ D(T )} ⊆ H1 × H2. If G(T ) is closed, then T is called
a closed operator. Equivalently, T is closed if and only if (xn) is a sequence in
D(T ) such that xn → x ∈ H1 and T xn → y ∈ H2, then x ∈ D(T ) and T x = y.
By the closed graph Theorem [7], an everywhere defined closed operator is
bounded. Hence the domain of an unbounded closed operator is a proper subspace
of a Hilbert space.
The space of all bounded operators between H1 and H2 is denoted by B(H1, H2)
and the class of all closed operators between H1 and H2 is denoted by C(H1, H2).
We write B(H, H) = B(H) and C(H, H) = C(H).
If T ∈ C(H1, H2), then the null space and the range space of T are denoted by
N(T ) and R(T ) respectively and the space C(T ) := D(T ) ∩ N(T )⊥ is called the
carrier of T . In fact, D(T ) = N(T ) ⊕⊥ C(T ) [6, page 340].
For a densely defined operator, there exists a unique linear operator (in fact, a
closed operator) T ∗ : D(T ∗) → H1, with
D(T ∗) := {y ∈ H2 : x → hT x, yi for all x ∈ D(T ) is continuous} ⊆ H2
satisfying hT x, yi = hx, T ∗yi for all x ∈ D(T ) and y ∈ D(T ∗).
that T ∗ exists if and only if T is densely defined.
it is to be noted
MINIMUM ATTAINING CLOSED OPERATORS
3
If S and T are closed operators with the property that D(T ) ⊆ D(S) and
T x = Sx for all x ∈ D(T ), then S is called the restriction of T and T is called
an extension of S. Furthermore, S = T if and only if S ⊆ T and T ⊆ S.
If S ∈ B(H) and T ∈ C(H) is densely defined, the we say S and T are commut-
ing if ST ⊆ T S. That is, D(ST ) ⊆ D(T S) and ST x = T Sx for all x ∈ D(ST ).
A densely defined operator T ∈ C(H) is said to be normal if T ∗T = T T ∗,
self-adjoint if T = T ∗ and positive if hT x, xi ≥ 0 for all x ∈ D(T ).
If T is positive, then there exists a unique positive operator S such that T = S2.
The operator S is called the square root of T and it is denoted by S = T
1
2 .
If T ∈ C(H1, H2) is densely defined, then the operator T := (T ∗T )
2 is called
the modulus of T . There exists a unique partial isometry V : H1 → H2 with
initial space R(T ∗) and range R(T ) such that T = V T .
1
It can be verified that D(T ) = D(T ) and N(T ) = N(T ) and R(T ) = R(T ∗).
Let T ∈ C(H) be densely defined. The resolvent of T is defined by
ρ(T ) := {λ ∈ C : T − λI : D(T ) → H is invertible and (T − λI)−1 ∈ B(H)}
and
σ(T ) : = C \ ρ(T )
σp(T ) : = {λ ∈ C : T − λI : D(T ) → H is not one-to-one},
are called the spectrum and the point spectrum of T , respectively.
Let T ∈ C(H1, H2) be densely defined. A subspace D of D(T ) is called a core
for T if for any x ∈ D(T ), there exists a sequence (xn) ⊂ D such that lim
xn = x
n→∞
In other words, D is dense in the graph norm, which is
and lim
n→∞
defined by kxk := kxk + kT xk for all x ∈ D(T ). It is a well known fact that
D(T ∗T ) is a core for T (see [27, Proposition 3.18, page 47] for details).
T xn = T x.
If M is a closed subspace of a Hilbert space H, then PM denotes the orthogonal
projection PM : H → H with range M, and SM := {x ∈ M : kxk = 1} is the unit
sphere of M.
We refer [1, 4, 5, 7, 20, 27] for the above basics of unbounded operators.
Here we recall definition and properties of the Moore-Penrose inverse (or gener-
alized inverse) of a densely defined closed operator that we need for our purpose.
We refer [6] for more details on this topic.
Let T ∈ C(H1, H2) be densely defined. Then there exists a unique densely
defined operator T † ∈ C(H2, H1) with domain D(T †) = R(T ) ⊕⊥ R(T )⊥ and has
the following properties:
(1) T T †y = PR(T ) y, for all y ∈ D(T †)
(2) T †T x = PN (T )⊥ x, for all x ∈ D(T )
(3) N(T †) = R(T )⊥.
This unique operator T † is called the Moore-Penrose inverse of T .
The following property of T † is also well known.
For every y ∈ D(T †), let
L(y) := nx ∈ D(T ) : T x − y ≤ T u − y
for all u ∈ D(T )o.
4
S. H. KULKARNI, G. RAMESH
Here any u ∈ L(y) is called a least square solution of the operator equation
T x = y. The vector x = T †y ∈ L(y), T †y ≤ x
for all x ∈ L(y) and it is
called the least square solution of minimal norm. A different treatment of T † is
given in [6, Pages 314, 318-320], where it is called "the Maximal Tseng generalized
Inverse".
Here we recall some properties of T † that we will be using very frequently.
Theorem 2.1. [6, Page 320] Let T ∈ C(H1, H2) be densely defined. Then
(1) D(T †) = R(T ) ⊕⊥ R(T )⊥, N(T †) = R(T )⊥ = N(T ∗)
(2) R(T †) = C(T )
(3) T † is densely defined and T † ∈ C(H2, H1)
(4) T † is continuous if and only R(T ) is closed
(5) T †† = T
(6) T ∗† = T †∗
(7) N(T ∗†) = N(T )
(8) T ∗T and T †T ∗† are positive and (T ∗T )† = T †T ∗†
(9) T T ∗ and T ∗†T † are positive and (T T ∗)† = T ∗†T †.
3. Minimum attaining Operators
In this section first we discuss some important properties of minimum attaining
operators. These operators for the bounded case was discussed in [13] and the
unbounded case in [25]. It is proved that this class is dense in the class of densely
defined closed operators with respect to the gap topology.
Definition 3.1. [6, 5] Let T ∈ C(H1, H2) be densely defined. Then
m(T ) := inf {kT xk : x ∈ SD(T )}
γ(T ) := inf {kT xk : x ∈ SC(T )},
are called the minimum modulus and the reduced minimum modulus of T , respec-
tively. The operator T is said to be bounded below if and only if m(T ) > 0.
Remark 3.2. If T ∈ C(H1, H2) is densely defined, then
(a) By definition, we have m(T ) ≤ γ(T ). More over, if T is one-to-one,
m(T ) = γ(T ) since D(T ) = C(T )
(b) m(T ) > 0 if and only if R(T ) is closed and T is one-to-one
(c) Since D(T ) = D(T ) and kT xk = kT xk for all x ∈ D(T ), we can
conclude that m(T ) = m(T ) and γ(T ) = γ(T ).
Remark 3.3. If T ∈ C(H) is densely defined and R(T ) is closed, then γ(T ) =
1
kT †k.
Recall that T ∈ B(H1, H2) is said to be norm attaining if there exists x0 ∈ SH1
such that kT x0k = kT k. We denote the class of all norm attaining operators be-
tween H1 and H2 by N (H1, H2). In case H1 = H2 = H, we denote this by N (H).
In a similar way, we can define operators that attain minimum modulus. The
class of bounded operators that attain minimum modulus is defined and several
characterizations are proved in [12]. Here we discuss the same for unbounded
operators.
MINIMUM ATTAINING CLOSED OPERATORS
5
Definition 3.4. Let T ∈ C(H1, H2) be densely defined. If there exists x0 ∈ SD(T )
such that kT x0k = m(T ), then we call T to be minimum attaining.
We write
Mc(H1, H2) = {T ∈ C(H1, H2) : T is densely defined and minimum attaining}
and Mc(H, H) = Mc(H).
Remark 3.5. Let T ∈ C(H1, H2) be densely defined.
(1) If T is not one-to-one, then m(T ) = 0 and there exists a x0 ∈ SN (T ) such
that T x0 = 0. Hence T ∈ Mc(H1, H2).
(2) If T is one-to-one and R(T ) is not closed, then m(T ) = 0. But there does
not exists x0 ∈ D(T ) such that kT x0k = 0, since T is one-to-one. Thus
T /∈ Mc(H1, H2).
From the above two observations it is apparent that the injectivity of the
operator plays an important role in the minimum attaining property.
First, we establish some results related to the minimum modulus of a densely
defined closed operator, which are useful in discussing the minimum attaining
property.
Proposition 3.6. Let T ∈ C(H) be densely defined and normal. Then
(1) m(T ) = d(0, σ(T ))
(2) m(T n) = m(T )n.
Proof. If m(T ) = 0, then T is not invertible, so 0 ∈ σ(T ) and d(0, σ(T )) = 0. If
m(T ) > 0, then T −1 ∈ B(H). In this case, m(T ) = γ(T ) = 1
kT −1k. Therefore,
1
kT −1k
=
1
sup {µ : µ ∈ σ(T −1)}
1
=
sup { 1
λ : λ ∈ σ(T )}
= inf {λ : λ ∈ σ(T )}
Proof of (2): It is easy to verify that T n is normal. Hence, by (1) and the
spectral mapping theorem we can conclude that
= d(0, σ(T )).
(cid:3)
m(T n) = inf {µ : µ ∈ σ(T n)}
= inf {λn : λ ∈ σ(T )}
= inf {λn : λ ∈ σ(T )}
= m(T )n.
Corollary 3.7. If T ∈ C(H1, H2) is densely defined, then
(1) m(T ) = d(0, σ(T ))
(2) m(T ∗T ) = m(T )2.
Proof. We have m(T ) = m(T ) = d(0, σ(T )), by (1) of Proposition 3.6. Also,
m(T ∗T ) = m(T 2) = d(0, σ(T 2)) = d(0, σ(T ))2 = m(T )2 = m(T )2. Here we
have used both (1) and (2) of Proposition 3.6 to get the conclusion.
(cid:3)
6
S. H. KULKARNI, G. RAMESH
Proposition 3.8. Let T ∈ C(H) be densely defined and positive. Then
Proof. First note that D(T ) ⊆ D(T
2 ). Next,
m(T ) = inf (cid:8)hT x, xi : x ∈ SD(T )(cid:9) = mT .
1
mT = inf(cid:8)hT x, xi : x ∈ SD(T )(cid:9) = inf (cid:8)hT
≥ inf(cid:8)kT
1
1
1
2 x, T
1
2 xk2 : x ∈ D(T
2 xi : x ∈ SD(T )(cid:9)
2 )(cid:9)
1
= m(T
2 )2.
2 ) = inf {λ : λ ∈ σ(T
1
2 )} by Corollary 3.7. As σ(T ) = {λ2 : λ ∈ σ(T
1
2 )},
1
But m(T
we have that m(T
1
2 )2 = m(T ) and hence mT ≥ m(T ).
On the other hand, we have
mT ≤ hT x, xi for all x ∈ SD(T )
1
1
= hT
= kT
2 xi for all x ∈ SD(T )
2 x, T
2 xk2 for all x ∈ SD(T ).
1
Next, we claim that the above inequality holds for all x ∈ D(T
let x ∈ D(T
such that lim
n→∞
2 ). Since, D(T ) is a core for T
2 xn = T
2 ). To this end,
2 , there exists a sequence (xn) ⊂ D(T )
2 xnk2 ≥
2 x. Hence kT
xn → x and lim
n→∞
2 xk2 = lim
n→∞
kT
T
1
1
1
1
1
1
1
mT . As this is true for all x ∈ D(T
1
2 ), it follows that m(T ) ≥ mT .
By the above two observations the conclusion follows.
(cid:3)
Proposition 3.9. Let T ∈ C(H1, H2) be densely defined and m(T ) = m(T ∗).
Also, assume that R(T ) is closed. Then T ∈ Mc(H1, H2) if and only if T ∗ ∈
Mc(H2, H1).
Proof. Clearly, if m(T ) = m(T ∗) = 0, since R(T ) closed, both T and T ∗ are not
one-to-one. Hence both are minimum attaining. Now assume that m(T ) > 0.
It is sufficient to prove one implication, since T ∗∗ = T and m(T ) = m(T ∗). By
Proposition 3.13, T ∈ Mc(H1, H2) if and only if there exists a x0 ∈ SD(T ) such
that T x0 = m(T )x0. That is T ∗T x0 = m(T )T x0. Hence
kT ∗T x0k
kT x0k
= m(T )
kT x0k
kT x0k
= m(T ),
proving T ∗ ∈ Mc(H2, H1).
(cid:3)
Remark 3.10. Let T ∈ C(H) be densely defined and normal. Then D(T ) =
D(T ∗) and kT xk = kT ∗xk for all x ∈ D(T ). Hence T ∈ Mc(H) if and only if
T ∗ ∈ Mc(H). Clearly, in this case m(T ) = m(T ∗). Note that in this case we
don't have to assume that the range of T to be closed.
We recall that if T ∈ C(H1, H2) is densely defined, then the numerical range
W (T ) of T is defined by W (T ) = {hT x, xi : x ∈ SD(T )}.
Proposition 3.11. If T ∈ C(H) is positive, then the following are equivalent;
(1) T ∈ Mc(H)
MINIMUM ATTAINING CLOSED OPERATORS
7
(2) m(T ) ∈ σp(T )
(3) m(T ) is an extreme point of W (T ).
Proof. Proof of (1) ⇒ (2) : Choose x0 ∈ SD(T ) such that kT x0k = m(T ). Since,
T − m(T )I is positive, and by the Cauchy-Scwarz inequality, we get that
m(T ) ≤ hT x0, x0i ≤ kT x0k = m(T ),
or m(T ) = hT x0, x0i. Therefore,
kT x0 − m(T )x0k2 = kT x0k2 + m(T )2 − 2m(T )hT x0, x0i
= 2m(T )2 − 2m(T )2
= 0.
That is, T x0 = m(T )x0. Clearly, if m(T ) ∈ σp(T ), then T ∈ Mc(H).
Proof of (2) ⇒ (3) : Let x0 ∈ SD(T ) be such that T x0 = m(T )x0. Then
m(T ) = hT x0, x0i ∈ W (T ). Since, m(T ) = mT by Proposition 3.8, the conclusion
follows. The other way implication follows by the main theorem of [11].
(cid:3)
Using Proposition 3.11, we can prove the following.
Proposition 3.12. Let T ∈ C(H) be densely defined and positive. Then T ∈
Mc(H) if and only if T
2 ∈ Mc(H).
1
1
2 ∈ Mc(H), then m(T
Proof. If T
By Proposition 3.11, T ∈ Mc(H).
1
2 ) ∈ σp(T
1
2 ), which implies that m(T ) ∈ σp(T ).
Conversely, if T ∈ Mc(H), then m(T ) ∈ σp(T ), by Proposition 3.11. If m(T ) =
2 ∈ Mc(H). Next, assume that m(T ) > 0. Then
2 ) = 0 and hence T
1
1
0, then m(T
m(T
2 ) > 0 and
1
T − m(T )I = (cid:0)T
1
2 + m(T )
1
2 − m(T )
1
2 I(cid:1)(cid:0)T
1
2 I(cid:1).
1
1
2 + m(T )
As T
to-one. Hence m(T
2 I has a bounded inverse, we have that T
2 I is not one-
2 ). The conclusion follows by Proposition 3.11. (cid:3)
1
2 ) ∈ σp(T
2 − m(T )
1
1
1
Theorem 3.13. Let T ∈ C(H1, H2) be densely defined. Then the following state-
ments are equivalent:
(1) T ∈ Mc(H1, H2)
(2) T ∈ M(H1)
(3) T ∗T ∈ M(H1).
Proof. The equivalence of (1) and (2) follows by the observation that D(T ) =
D(T ) and kT xk = kT xk for all x ∈ D(T ). The equivalence of (2) and (3)
follows by the fact that T ∗T = T 2 and Proposition 3.12.
(cid:3)
Example 3.14. Let D = {(xn) ∈ ℓ2 :
∞
Xn=1
n2xn2 < ∞}. Define T : D → ℓ2 by
T ((x1, x2, x3, . . . , )) = ((0, x1, 2x2, 3x3, . . . )), for all (xn) ∈ D.
Clearly, T is densely defined closed operator. Note that T ∗T (xn) = (n2xn)
for all (xn) ∈ D(T ∗T ). It can be easily calculated that σ(T ∗T ) = σp(T ∗T ) =
8
S. H. KULKARNI, G. RAMESH
{n2 : n ∈ N}. Hence m(T ∗T ) = 1 and T ∗T ∈ M(ℓ2). By Theorem 3.13, we can
conclude that T ∈ M(ℓ2) and by Corollary 3.7 m(T ) = 1.
Proposition 3.15. Let T be densely defined and positive. Then T ∈ Mc(H) if
and only if T n ∈ Mc(H) for each n ≥ 1.
Proof. Let T ∈ Mc(H). Then by Proposition 3.11, there exists x0 ∈ D(T ) such
that T x0 = m(T )x0. Observe that x0 ∈ D(T 2). This implies that T 2x0 =
m(T )2x0 = m(T 2)x0, by Proposition 3.6. That is x0 ∈ D(T 4) ⊆ D(T 3). With
this, we have T 3x0 = m(T )3x0. By the induction argument we can show that
T nx0 = m(T )nx0. By Proposition 3.11, it follows that T n ∈ Mc(H).
To prove the converse, assume that n > 1 and T n ∈ Mc(H). Choose x0 ∈ SD(T )
such that T nx0 = m(T n)x0. As m(T n) = m(T )n, if m(T n) = 0, then m(T ) = 0.
In this case xo ∈ N(T n). That is T n−2x0 ∈ N(T 2) = N(T ). Hence x0 ∈ N(T n−1).
Proceeding in this we can conclude that x0 ∈ N(T ), proving T ∈ Mc(H).
Next assume that m(T ) > 0. Since T is positive, T −1 ∈ B(H). Hence T x0 =
m(T )x0 implies that
T n−1x0 = T −1T nx0 = m(T )nT −1x0 = m(T )n x0
m(T )
= m(T )n−1x0.
By proceeding in this way, we can conclude that T x0 = m(T )x0. Hence T ∈
Mc(H).
(cid:3)
Proposition 3.16. Let T ∈ Mc(H1, H2) be one-to-one. Then R(T ) is closed.
Proof. If R(T ) is not closed, then m(T ) = 0. Since T ∈ Mc(H1, H2), there exists
x0 ∈ SD(T ) such that kT x0k = 0, but contradicts T to be one-to-one. Thus R(T )
is closed.
(cid:3)
Remark 3.17. The condition one-one ness is not necessary in Proposition 3.16.
For example, let P be a bounded orthogonal projection. Then R(P ) is closed and
it is minimum attaining but not one-to-one.
Corollary 3.18. Let T ∈ Mc(H1, H2). Then T is one-to-one if and only T
bounded below.
Next, we will establish a relation between the minimum attaining property of
the operator and the norm attaining property of its generalized inverse. First we
prove a few results needed for this purpose.
Proposition 3.19. Let T ∈ C(H1, H2) be densely defined. Then
(1) T † = T ∗†
(2) (T †)∗ = T †.
Proof. Proof of (1): By definition of T †, and by Theorem 2.1,
1
2 = T ∗†.
1
1
1
T † = (cid:0)(T †)∗T †(cid:1)
2 = (cid:0)(T T ∗)†(cid:1)
2 = (cid:0)(T ∗2)†(cid:1)
2 = (cid:0)(T ∗)†(T ∗)†(cid:1)
The proof of (2) can be obtained by replacing T by T ∗ in (1) and observing that
(T ∗)† = (T ∗)† and (T ∗)∗ = T .
(cid:3)
Theorem 3.20. Let T ∈ C(H1, H2) be densely defined and one-to-one. Then the
following statements are equivalent;
MINIMUM ATTAINING CLOSED OPERATORS
9
(1) T ∈ Mc(H1, H2)
(2) R(T ) is closed and T † ∈ N (H2, H1).
Proof. First assume that T ∈ Mc(H1, H2). Note that R(T ) is closed, by Propo-
sition 3.16. As T is one-to-one, m(T ) > 0. Choose x0 ∈ D(T ) such that
T x0 = m(T )x0. Hence
m(T )T −1x0 = T −1T x0 = x0.
So
T −1x0 =
1
m(T )
x0.
(3.1)
By Proposition 3.19, we have T −1 = (T †)∗ = (T ∗)†. Since m(T ) = m(T ) =
, Equation 3.1, takes the form T †x0 = kT †kx0. Hence the
=
1
1
kT †k
conclusion follows.
kT †k
To prove the other implication, let T † ∈ N (H2, H1). Then clearly, R(T ) is
closed. By [12, Proposition 2.5], S := (T †)∗ ∈ N (H1, H2). Hence by Proposition
3.19, we have S = T †. Since T † is positive and norm attaining, there exists
x0 ∈ SH1 such that
Sx0 = kSkx0 = kT †kx0 =
1
m(T )
x0.
(3.2)
Note that x0 ∈ R(T †) = C(T ) = C(T ) ⊆ N(T )⊥. Premultiplying Equation
3.2 by T and noting that R(T ) = N(T )⊥, we have T x0 = m(T )x0, concluding
T ∈ M(H1). Hence T ∈ Mc(H1, H2) by Theorem 3.13.
(cid:3)
Next, we show that minimum attaining property of a closed densely defined
operator is related to the minimum attaining property of the corresponding
If T ∈ C(H1, H2) is densely defined, then the operator
bounded transform.
ZT := T (I + T ∗T )− 1
2 is called the bounded transform of T . More over, T =
T ZT )− 1
ZT (I − Z ∗
2 . We refer [27, section 7.3, page 142] for more details about
these operators.
Proposition 3.21. Let T ∈ C(H1, H2) be densely defined. Then
(1) m(ZT ) =
(2) m(T )2 =
m(T )
p1 + m(T )2
m(ZT )2
1 − m(ZT )2
(3) T ∈ Mc(H1, H2) if and only if ZT ∈ Mc(H1, H2).
Proof. Proof of (1): We have that
m(ZT )2 = m(Z ∗
T ZT ) = m(I−(I+T ∗T )−1) = 1−k(I+T ∗T )−1k = 1−
1
m(I + T ∗T )
,
10
S. H. KULKARNI, G. RAMESH
by Remark 3.3. Hence
m(ZT )2 = 1 −
1
1 + m(T )2
=
m(T )2
1 + m(T )2 .
Proof of (2): Note that T ∗T = Z ∗
T ZT )−1 − I. Thus
T ZT (I − Z ∗
T ZT )−1 = (I − Z ∗
m(T )2 = m(T ∗T ) = m(I − Z ∗
T ZT )−1 − 1 =
1
kI − Z ∗
T ZT k
− 1 =
1
1 − m(Z ∗
T ZT )
− 1,
by Remark 3.3. Hence
m(T )2 =
1
1 − m(ZT )2 − 1 =
m(ZT )2
1 − m(ZT )2 .
Proof of (3): In view of Theorem 3.13, it is enough to prove T ∗T ∈ M(H1) if
and only if Z ∗
T ZT ∈ M(H1). We know by (2) of Proposition 3.11, that T ∗T ∈
M(H1) if and only if m(T )2 ∈ σp(T ∗T ). Since, Z ∗
T ZT = T ∗T (I + T ∗T )−1 and
1+m(T )2 , it can be verified that m(T )2 ∈ σp(T ∗T ) if and only if
m(Z ∗
m(Z ∗
(cid:3)
T ZT ) = m(T )2
T ZT ) ∈ σp(Z ∗
T ZT ).
4. Absolutely minimum attaining operators
In this section, we define absolutely minimum attaining operators and describe
the structure of such operators.
Definition 4.1. Let T ∈ C(H1, H2) be densely defined. Then T is called absolutely
minimum attaining operator if T M : D(T ) ∩ M → H2 is minimum attaining for
each non- zero closed subspace M of H1. In other words, T is absolutely minimum
attaining if there exists x0 ∈ D(cid:0)T M(cid:1) with kx0k = 1 such that kT x0k = m(T M ).
Note that if T ∈ C(H1, H2) is densely defined and M is a closed subspace of
H, then the restriction operator T M : D(T ) ∩ M → H2 is a closed operator and
it is densely defined as D(T M ) is dense in the Hilbert space D(T M ).
We denote the set of all absolutely minimum attaining operators between H1
and H2 by AMc(H1, H2) and in case if H1 = H2 = H, this is denoted by
AMc(H). This class of operators was introduced and studied in detail by Car-
vajal and Neves in [13]. The structure of positive absolutely minimum attaining
bounded operators is studied in [16].
Proposition 4.2. Let T ∈ AMc(H1, H2). Then R(T ) is closed.
Proof. Since T ∈ AMc(H1, H2), we have T0 = T N (T )⊥ ∈ AMc(N(T )⊥, H2)
and one-to-one. Hence by Proposition 3.16, R(T0) is closed.
It is clear that
R(T ) = R(T0).
(cid:3)
Remark 4.3. The converse of Proposition 4.2 need not be true. Let P be a
bounded orthogonal projection with infinite dimensional null space and infinite
dimensional range space. Then R(P ) is closed but P is not absolutely minimum
attaining by [13, Lemma 3.2].
MINIMUM ATTAINING CLOSED OPERATORS
11
Let M be a closed subspace of H and T ∈ C(H) be densely defined. Then M
is said to be invariant under T , if T (M ∩ D(T )) ⊆ M.
Let P := PM . If P (D(T )) ⊆ D(T ) and (I − P )(D(T )) ⊆ D(T ), then
T = (cid:18) T11 T12
T21 T22 (cid:19) ,
where Tij = PiT PjMj (i, j = 1, 2). Here P1 = P and P2 = I − P . It is known
that M is invariant under T if and only if T21 = 0. Also, M reduces T if and only
if T21 = 0 = T12.
Remark 4.4. Let T ∈ C(H1, H2) be densely defined. Assume that M reduces T
and T1 = T M and T2 = T M ⊥. Then the following can be easily verified:
(1) m(T ) = min{m(T1), m(T2)}
(2) T ∈ Mc(H) if and only if the operator Tj with m(Tj) = m(T ), (j =
1 or 2), is minimum attaining.
Lemma 4.5. Let T ∈ C(H) be densely defined. If M reduces T , then T †M =
(T M )†.
Proof. Since M is reducing subspace, we have
T = (cid:18) T1
0 T2 (cid:19) ,
0
where T1 = T M and T2 = T M ⊥. Since R(T ) is closed, by Theorem [4, page 287,
V.5], R(Ti) is closed for i = 1, 2. Let S = (cid:18) T †
0
1
0 T †
2 (cid:19) . Note that S ∈ B(H) and
it can be verified that S satisfies all the conditions of the Moore-Penrose inverse.
Since T † is unique, it follows that S = T †. This proves the claim.
(cid:3)
Theorem 4.6. Let T ∈ C(H) be densely defined and have a bounded inverse. Let
M be a subspace of H. Then
(1) m(T M ) =
1
kT −1T (M ∩D(T ))k
(2) if M is closed, then T (M ∩ D(T )) is closed
(3) If N is any subspace of H, then m(T T −1(N )) =
1
.
kT −1N k
Furthermore, if N is closed and T ∈ B(H), then T −1(N) is closed.
(4) T ∈ AMc(H) if and only if T −1 ∈ AN (H).
12
S. H. KULKARNI, G. RAMESH
Proof. Proof of (1): First, note that as T is one-to-one, we have D(T ) = C(T )
and m(T ) = γ(T ). By definition,
=
kxk
m(T M ) = infn kT xk
supn kxk
supnkT −1yk
kT xk
=
kyk
1
=
kT −1T (M ∩D(T ))k
.
: x ∈ M ∩ D(T ), x 6= 0o
1
: x ∈ M ∩ C(T ), x 6= 0o
1
: y = T x ∈ T (M ∩ D(T )), x 6= 0o
Proof of (2): Let N := T (M ∩ D(T )) and let y ∈ N . Let (xn) ⊂ M ∩ D(T )
xn = T −1y.
be such that y = lim
n→∞
Since M is closed, we can conclude that T −1y ∈ M ∩ D(T ). Since R(T ) is closed,
y ∈ R(T ). Hence y = T (T −1y) ∈ T (M ∩ D(T )).
T xn. Since T −1 ∈ B(H), it follows that lim
n→∞
Proof of (3): This goes along the similar lines of (1) and (2).
Proof of (4): If M = H, then by Theorem 3.20, we have that T ∈ Mc(H) if and
only if T −1 ∈ N (H). Hence assume that {0} 6= M ⊂ H. Let T −1 ∈ AN (H). Let
X = T (M ∩ D(T )) and RX = T −1X. By (2), X is closed. Since RX ∈ N (X, H),
there exists y0 ∈ SX , such that kRXy0k = kRXk. This is equivalent to the fact
that R∗
XRX y0 = kRXk2y0. Let y0 = T x0 for some x0 ∈ M ∩ D(T ), we get
= m(T M ) kx0k by
X x0 = kRXk2T x0. Therefore kT x0k =
R∗
Xx0k
kR∗
kRXk2 ≤
kx0k
kRXk
(1). Writing z0 =
, we get that kT z0k ≤ m(T M ). But the other inequality
x0
kx0k
holds clearly. Hence T M ∈ Mc(M, H).
Conversely, assume that T ∈ AMc(H). Let N be a closed subspace of H
and let M := T −1(N) ⊆ C(T ). Since T ∈ AMc(H), we have TM := T M ∈
Mc(M, H). It can be easily verified that TM is closed, since T is closed. Since
D(T ) ∩ M is dense in D(T ) ∩ M, TM is densely defined operator. Hence T ∗
M :
D(T ∗
M ) → M exists. By Theorem 3.13 and Proposition 3.11, there exists x0 ∈
M TM ) such that
SD(T ∗
(4.1)
As T is bounded below, TM is bounded below and hence m(TM ) > 0. Let
x0 = T −1y0 for some y0 ∈ N. Then Equation 4.1 takes the form:
T ∗
M TM x0 = m(TM )2x0.
T ∗
M y0 = m(TM )2T −1y0.
(4.2)
First, observe that R(TM ) = TM (M ∩ D(T )) = T (D(T ) ∩ M) = T (M) =
T (T −1(N)) = N. Hence y0 ∈ N = R(TM ) = N((TM )∗)⊥. Taking norm both
sides of Equation 4.2, we get
kT −1y0k =
M y0k
kT ∗
m(TM )2 ≥
M )ky0k
γ(T ∗
m(TM )2 =
γ(TM )ky0k
m(TM )2 ≥
m(TM )ky0k
m(TM )2 =
ky0k
m(TM )
.
MINIMUM ATTAINING CLOSED OPERATORS
Hence z0 =
y0
ky0k
∈ SN and kT −1(z0)k = kT −1N k.
13
(cid:3)
Corollary 4.7. Let T ∈ B(H) be such that T −1 ∈ B(H). Then T ∈ AMc(H) if
and only if T −1 ∈ AN (H).
Theorem 4.8. Let T ∈ AMc(H), positive and not bounded. Assume that T
is one-to-one. Then there exists an unbounded (increasing) sequence {λn} of
eigenvalues of T with corresponding eigenvectors {φn} such that
(1)
D(T ) = {x ∈ H :
∞
Xn=1
λ2
n hx, φni2 < ∞} and
T x =
∞
Xn=1
λnhx, φniφn, for all x ∈ D(T ).
The series in the above representation converges in the strong operator
topology. Moreover, T −1 is compact.
(2) σ(T ) = {λn : n ∈ N}=σp(T )
(3) if µ ∈ σp(T ), then µ is an eigenvalue with finite multiplicity
(4) span{φn : n ∈ N} = H.
Proof. Proof of (1): First note that as T is one-to-one and R(T ) is closed, T is
bounded below. Since T ≥ 0, T −1 exists and bounded. If T ∈ AMc(H), then
T −1 ∈ AN (H) by Theorem 4.6. Hence by [3, Theorem 2.5], there exists unique
triple (K, F, α), where K ∈ K(H) is positive, F ∈ F (H) positive and α ≥ 0
such that KF = 0 = F K, F ≤ αI and T −1 = αI + K − F . If α = 0, then
F = 0 and hence T −1 = K ∈ K(H). Next, assume that α > 0. In this case,
R(T −1) = D(T ) is closed by [3, Proposition 2.8] is closed. Since T is densely
defined, we must have that D(T ) = H. By the closed graph theorem T must be
bounded, a contradiction. Hence α > 0 is not possible. This implies that α = 0
and hence T −1 ∈ K(H).
By the spectral theorem, there exists increasing sequence (µn) of positive eigen-
values of T −1 with corresponding eigenvectors {φn : n ∈ N} such that
T −1y =
∞
Xn=1
µnhy, φniφn, for all y ∈ H.
(4.3)
The sequence µn → 0 as n → ∞. More over, the above series converges to T −1
in the operator norm of B(H). We can also observe that the sequence (µn) is an
infinite sequence. Otherwise, T −1 is a finite rank operator and σ(T ) is bounded.
By [26] this implies that T is bounded which leads to a contradiction. Also,
since T −1 is compact, σ(T −1) = {µn : n ∈ N} = σp(T ) and each µn has finite
multiplicity. Also, µn+1 ≤ µn for each n ∈ N.
14
S. H. KULKARNI, G. RAMESH
Let λn := µ−1
n
for all n ∈ N. As T −1 is compact, by [10, Theorem 6.1, page
214], it follows that
D(T ) = {x ∈ H :
∞
Xn=1
λ2
n hx, φni2 < ∞} and
T x =
∞
Xn=1
λnhx, φniφn, for all x ∈ D(T ).
On the other hand, if T −1 is compact, by Theorem 4.6, T ∈ AMc(H).
Proof of statement (2) is clear. The statement (3) is proved in (1).
Proof of (4): Since T −1 is compact, R(T −1) = D(T ) is separable and by the
representation above, we have that H = D(T ) = R(T −1) = span{vn : n ∈ N}. (cid:3)
Remark 4.9. If T ∈ B(H), then the conclusion (1) of Theorem 4.8 is not true.
The unboundedness of the operator is used to get the inverse to be compact.
Theorem 4.10. Let T ∈ C(H) be densely defined and one-to-one but not bounded.
Then
(1) T ∈ AMc(H) if and only if T ∗T ∈ AMc(H)
(2) T ∈ AMc(H) if and only if T † ∈ K(H).
Proof. Proof of (1): If T ∈ AMc(H), then R(T ) is closed. As T is one-to-one, T
is bounded below. Also, since T and T ∗T are bounded below and positive, both
have bounded inverse. Hence
T ∈ AMc(H) ⇔ T ∈ AMc(H) ⇔ T −1 ∈ K(H) (by Theorem 4.8)
⇔ T −2 ∈ K(H)
⇔ (T ∗T )−1 ∈ K(H)
⇔ T ∗T ∈ AMc(H) (by Theorem 4.8).
Proof of (2): By (1),
T ∈ AMc(H) ⇔ T ∈ AMc(H) ⇔ T −1 ∈ K(H) (by Theorem 4.8)
⇒ T † ∈ K(H).
On the other hand, if T † ∈ K(H), then R(T ) is closed. As T is one-to-one, T
must be bounded below. This implies that T −1 ∈ B(H). Thus,
T † ∈ K(H) ⇔ (T ∗)† ∈ K(H) ⇔ (T ∗)† ∈ K(H)
⇔ T † = T −1 ∈ K(H) (by (2) of Proposition 3.19 )
⇔ T ∈ AMc(H)
⇔ T ∈ AMc(H).
(cid:3)
Theorem 4.11. Let T ∈ C(H) be densely defined unbounded and have a bounded
inverse. Then T ∈ AMc(H) if and only if T ∗ ∈ AMc(H).
MINIMUM ATTAINING CLOSED OPERATORS
15
Proof. First observe that both T and T ∗ are bounded below. We know that T ∈
AMc(H) if and only if T −1 ∈ K(H). This is true if and only if (T ∗)−1 ∈ K(H).
Now, by 2 of Theorem 4.10, this is equivalent to the fact that T ∗ ∈ AMc(H). (cid:3)
Definition 4.12 (Hyper invariant subspace). Let T ∈ C(H) be densely defined
and M be a closed subspace of H. Then M is said to be hyperinvariant subspace
of T if M is invariant under every S ∈ B(H) such that ST ⊆ T S.
Theorem 4.13. (Lomonosov )[22] Every operator that commutes with a non-zero
compact operator and is not a multiple of the identity has a non-trivial hyperin-
variant subspace.
Using Theorem 4.13 we will prove that every AM-operator has a non trivial
hyperinvariant subspace.
Theorem 4.14. Let T ∈ AMc(H), unbounded and T −1 ∈ B(H). Then T has a
non trivial hyper invariant subspace.
Proof. Let S ∈ B(H) be such that ST ⊆ T S. That is ST x = T Sx for all
x ∈ D(T ). Then it can be easily verified that T −1S = ST −1. But T −1 ∈ K(H)
by (2) of Theorem 4.10. Now, by Theorem 4.13, T −1 has a non trivial invariant
subspace, say M. Then M is invariant under S. Thus the conclusion follows. (cid:3)
Now, we can drop the condition that the operator to be one-to-one in Theorem
4.8 and prove the result.
Theorem 4.15. Let T ∈ AMc(H) be, positive but not bounded. Then
(1) T † is compact
(2) R(T ) is separable.
Proof. Since N(T ) reduces T , we can write T = (cid:18) T0
and T1 = TN (T )⊥. Then by Lemma 4.5, T † = (cid:18) T †
0
0
0 T −1
0
0 T1 (cid:19), where T0 = T N (T )
1 (cid:19). As T1 ∈ AM(N(T )⊥),
by Theorem 4.10, T −1
is compact. Note that T0 = 0 if N(T ) 6= {0} and T = T1
if N(T ) = {0}. Hence T † is compact. Also R(T1) is separable by (4) of Theorem
4.8. Now the conclusion follows as R(T ) = R(T1).
(cid:3)
1
Using Theorem 4.15, we can prove a more general result.
Theorem 4.16. Let T ∈ AMc(H), but not bounded. Then
(1) T † is compact
(2) N(T )⊥ and R(T ) are separable.
Proof. Proof of (1): We have T ∈ AMc(H) if and only if T ∈ AMc(H). Hence
T † ∈ K(H) by Theorem 4.15. But, by Proposition 3.19, T † = (T †)∗ and
hence (T †)∗ ∈ K(H). This implies that T † ∈ K(H).
Proof of (2): Since T † is compact, R(T †) = C(T ) is separable. Hence N(T )⊥ is
separable. Since R(T ) is closed, R(T ∗) must be closed and since N(T )⊥ = R(T ∗)
, R(T ∗) is separable. But, R(T ∗) is separable if and only if R(T ) is separable by
[23, Problem 11.4.6, page 362].
(cid:3)
16
S. H. KULKARNI, G. RAMESH
Question 4.17. If T ∈ C(H) is densely defined and T † ∈ K(H). Is it true that
T ∈ AMc(H).
Theorem 4.18. Let T be a densely defined and self-adjoint, one-to-one opera-
tor on an infinite dimensional Hilbert space H which is not bounded. Then the
following are equivalent:
(1) T ∈ AMc(H)
(2) T −1 ∈ K(H)
(3) there exists a real sequence (λn) and an orthonormal basis {vn : n ∈ N} of
H such that lim
n→∞
λn = ∞ and T vn = λnvn for each n ∈ N
(4) T has purely discrete spectrum
(5) the resolvent Rλ(T ) := (T − λI)−1 is compact for one, and hence for all
λ ∈ ρ(T )
(6) the embedding map JT : (D(T ), k · kT ) → H is compact (here kxkT =
(cid:0)kxk2 + kT xk2(cid:1)
1
2 , x ∈ D(T )).
Proof. If T ∈ AMc(H), then R(T ) is closed. As T is one-to-one, T must be
bounded below and hence T −1 ∈ B(H). Now, by (2) of Theorem 4.10, T −1 ∈
K(H). Again by (2) of Theorem 4.10, if T −1 ∈ K(H), then T ∈ AMc(H). Thus
(1) and (2) are equivalent. The equivalence of (2)-(6) follows by [27, Proposition
5.12, page 94].
(cid:3)
Next, we give an example of AM-operator.
Example 4.19. Let p, p′, q, w be continuous real valued functions defined on [a, b]
with a < b and w(t) > 0 for all t ∈ [a, b]. Consider the real Hilbert space
with the inner product
H := (cid:26)u : Z b
a
u2w < ∞(cid:27)
hu, vi := Z b
a
u(x)v(x)w(x)dx.
Let L be the Sturm-Liouville operator given by
Lu :=
1
w
[−(pu
′
′
)
+ qu]
with
D(L) =
u ∈ H : u ∈ C 2[a, b],
β1u(a) + γ1u
′
(a) = 0,
β2u(b) + γ2u
′
(b) = 0,
β1 + γ1 > 0,
β2 + γ2 > 0
.
Since D(L) contains continuous functions defined on [a, b] with compact support,
L is densely defined operator. Also L is symmetric (See [9, Chapter 7, section
5]. Let us assume that 0 /∈ σp(L). In this case it easy to see that L−1 is compact
and self adjoint. Let B := {φ1, φ2, . . .} is an orthonormal basis for H such that
MINIMUM ATTAINING CLOSED OPERATORS
17
Lvn = µnvn, where µ1, µ2, . . . is a sequence of real numbers which are eigenvalues
of L. In this case every u ∈ H can be expressed as
u =
∞
Xn=1
hu, vnivn.
∞
Xn=1
hu, vni2µ2
n < ∞
Lu =
∞
Xn=1
µnhu, vnivn.
∞
If u ∈ D(L), then
and
Note that L−1y =
µ−1
n hy, vnivn for all y ∈ H.
It is clear that L−1 is
Xn=1
compact and by Proposition 3.19, we have L−1 = L−1. Hence by Theorem
4.18, L ∈ AMc(H).
We end up this section with the following question:
Question 4.20. Does every bounded absolutely minimum attaining operator
have a non trivial (hyper) invariant subspace?
Author contributions All authors contributed equally and significantly in
this paper. All authors read and approved the final manuscript.
Compliance with ethical standards
Conflict of interest The authors declare that they have no conflict of interest.
Ethical approval: This article does not contain any studies with human partic-
ipants or animals performed by any of the authors.
References
[1] N. I. Akhiezer and I. M. Glazman, Theory of linear operators in Hilbert space, translated
from the Russian and with a preface by Merlynd Nestell, reprint of the 1961 and 1963
translations, Dover Publications, Inc., New York, 1993. MR1255973
[2] M. D. Acosta, R. M. Aron and F. J. Garc´ıa-Pacheco, The approximate hyperplane se-
ries property and related properties, Banach J. Math. Anal. 11 (2017), no. 2, 295 -- 310.
MR3598746
Venku
operators,
[3] D.
ing
Sci. Math.
(https://www.ias.ac.in/public/Volumes/pmsc/forthcoming/PMSC-D-18-00008.pdf)
norm attain-
Sci.
absolutely
Acad.
and
appear
Naidu
to
G.
Ramesh,
in
Proc.
On
Indian
[4] A. E. Taylor and D. C. Lay, Introduction to functional analysis, second edition, John Wiley
& Sons, New York, 1980. MR0564653 (81b:46001)
[5] S. Goldberg, Unbounded linear operators: Theory and applications, McGraw-Hill, New
York, 1966. MR0200692 (34 #580)
[6] A. Ben-Israel and T. N. E. Greville, Generalized inverses: theory and applications, Wiley-
Interscience, New York, 1974. MR0396607 (53 #469)
[7] W. Rudin, Functional analysis, second edition, International Series in Pure and Applied
Mathematics, McGraw-Hill, New York, 1991. MR1157815 (92k:46001)
[8] T. Kato, Perturbation theory for linear operators, second edition, Springer, Berlin, 1976.
MR0407617 (53 #11389)
18
S. H. KULKARNI, G. RAMESH
[9] A. W. Naylor and G. R. Sell, Linear operator theory in engineering and science, sec-
ond edition, Applied Mathematical Sciences, 40, Springer, New York, 1982. MR0672108
(83j:46001)
[10] I. Gohberg, S. Goldberg and M. A. Kaashoek, Basic classes of linear operators, Birkhauser,
Basel, 2003. MR2015498 (2005g:47001)
[11] Bernau, S. J. Extreme eigenvectors of a normal operator. Proc. Amer. Math. Soc. 18 1967
127 -- 128. MR0205066 (34 #4901)
[12] Carvajal, Xavier; Neves, Wladimir. Operators that achieve the norm. Integral Equations
Operator Theory 72 (2012), no. 2, 179 -- 195. MR2872473 (2012k:47044)
[13] X. Carvajal and W. Neves, Operators that attain their minima, Bull. Braz. Math. Soc.
(N.S.) 45 (2014), no. 2, 293 -- 312. MR3249529
[14] J. Falc´o et al., Spaceability in norm-attaining sets, Banach J. Math. Anal. 11 (2017), no. 1,
90 -- 107. MR3571146
[15] J. Ganesh, G. Ramesh and D. Sukumar, Perturbation of minimum attaining operators,
Adv. Oper. Theory 3 (2018), no. 3, 473 -- 490. MR3795095
[16] J. Ganesh, G. Ramesh and D. Sukumar, A characterization of absolutely minimum attain-
ing operators, J. Math. Anal. Appl. 468 (2018), no. 1, 567 -- 583. MR3849004
[17] G. Ramesh, Approximation methods for solving operator equations involving unbounded
operators, Ph. D Thesis, IIT Madras. (2008).
[18] P. Enflo, J. Kover and L. Smithies, Denseness for norm attaining operator-valued functions,
Linear Algebra Appl. 338 (2001), 139 -- 144. MR1861118 (2002g:47148)
[19] S. Shkarin, Norm attaining operators and pseudospectrum, Integral Equations Operator
Theory 64 (2009), no. 1, 115 -- 136. MR2501174 (2010c:47004)
[20] M. S. Birman and M. Z. Solomjak, Spectral theory of selfadjoint operators in Hilbert space
(Russian), Leningrad. Univ., Leningrad, 1980. MR0609148 (82k:47001)
[21] G. Ramesh, Structure theorem for AN -operators, J. Aust. Math. Soc. 96 (2014), no. 3,
386 -- 395. MR3217722
[22] H. Radjavi and P. Rosenthal, Invariant subspaces, second edition, Dover, Mineola, NY,
2003. MR2003221 (2004e:47010)
[23] Y. A. Abramovich and C. D. Aliprantis, An invitation to operator theory, Graduate Studies
in Mathematics, 50, Amer. Math. Soc., Providence, RI, 2002. MR1921782 (2003h:47072)
[24] S. H. Kulkarni and G. Ramesh, Projection methods for computing Moore-Penrose inverses
of unbounded operators, Indian J. Pure Appl. Math. 41 (2010), no. 5, 647 -- 662. MR2735209
(2012k:47001)
[25] S. H. Kulkarni and G. Ramesh, On the denseness of minimum attaining operators, Oper.
Matrices 12 (2018), no. 3, 699 -- 709. MR3853362
[26] S. H. Kulkarni and M. T. Nair, A characterization of closed range operators, Indian J.
Pure Appl. Math. 31 (2000), no. 4, 353 -- 361. MR1760936 (2001d:47009)
[27] Schmudgen, Konrad, Unbounded self-adjoint operators on Hilbert space, Graduate Texts
in Mathematics, 265, Springer, Dordrecht, 2012, xx+432, MR2953553
[28] S. K. Pandey and V. I. Paulsen, A spectral characterization of AN operators, J. Aust.
Math. Soc. 102 (2017), no. 3, 369 -- 391. MR3650963
1 Department of Mathematics, I. I. T. Madras, Chennai, Tamilnadu, India 600
036.
E-mail address: [email protected]
2Department of Mathematics, I. I. T. Hyderabad, Kandi(V), Sangareddy, Telan-
gana, India-502 285.
E-mail address: [email protected]
|
1211.3335 | 1 | 1211 | 2012-11-14T15:31:22 | The fixed point property in a Banach space isomorphic to $c_0$ | [
"math.FA"
] | We consider a Banach space, which comes naturally from c0 and it appears in the literature, and we prove that this space has the fixed point property for non-expansive mappings. | math.FA | math |
THE FIXED POINT PROPERTY IN A BANACH SPACE
ISOMORPHIC TO c0
COSTAS POULIOS
Abstract. We consider a Banach space, which comes naturally from c0 and
it appears in the literature, and we prove that this space has the fixed point
property for non-expansive mappings.
1. Introduction
Let K be a weakly compact, convex subset of a Banach space X. A mapping
T : K → K is called non-expansive if kT x − T yk ≤ kx − yk for any x, y ∈ K. In
the case where every non-expansive map T : K → K has a fixed point, we say that
K has the fixed point property. The space X is said to have the fixed point property
if every weakly compact, convex subset of X has the fixed point property.
A lot of Banach spaces are known to enjoy the aforementioned property. The
earlier results show that uniformly convex spaces have the fixed point property (see
[3]) and this is also true for the wider class of spaces with normal structure (see
[7]). The classical Banach spaces ℓp, Lp with 1 < p < ∞ are uniformly convex and
hence they have the fixed point property. On the contrary, the space L1 fails this
property (see [1]).
The proofs of many positive results depend on the notion of minimal invariant
sets. Suppose that K is a weakly compact, convex set, T : K → K is a non-
expansive mapping and C is a nonempty, weakly compact, convex subset of K such
that T (C) ⊆ C. The set C is called minimal for T if there is no strictly smaller
weakly compact, convex subset of C which is invariant under T . A straightforward
application of Zorn's lemma implies that K always contains minimal invariant sub-
sets. So, a standard approach in proving fixed points theorems is to first assume
that K itself is minimal for T and then use the geometrical properties of the space
to show that K must be a singleton. Therefore, T has a fixed point.
Although a non-expansive map T : K → K does not have to have fixed points, it
is well-known that T always has an approximate fixed point sequence. This means
that there is a sequence (xn) in K such that limn→∞ kxn − T xnk = 0. For such
sequences, the following result holds (see [6]).
Theorem 1.1. Let K be a weakly compact, convex set in a Banach space, T : K →
K a non-expansive map, such that K is T -minimal, and let (xn) be any approximate
fixed point sequence. Then, for all x ∈ K,
lim
n→∞
kx − xnk = diam(K).
2010 Mathematics Subject Classification. Primary 47H10, 47H09, 46B25.
Key words and phrases. Non-expansive mappings, fixed point property, Banach spaces isomor-
phic to c0.
1
2
COSTAS POULIOS
Although from the beginning of the theory it became clear that the classical
spaces ℓp, Lp, 1 < p < ∞ have the fixed point property, the case of c0 remained
unsolved for some period of time. The geometrical properties of this space are not
very nice, in the sense that c0 does not possess normal structure. However, it was
finally proved that the geometry of c0 is still good enough and and it does not allow
the existence of minimal sets with positive diameter, that is c0 has the fixed point
property. This was done by B. Maurey [8] (see also [4]) who also proved that every
reflexive subspace of L1 has the fixed point property.
Theorem 1.2. The space c0 has the fixed point property.
The proof of Theorem 1.2 is based on the fact that the set of approximate fixed
point sequences is convex in a natural sense. More precisely, we have the following
([8], [4]).
Theorem 1.3. Let K be a weakly compact, convex subset of a Banach space which
is minimal for a non-expansive map T : K → K. Let (xn) and (yn) be approximate
fixed point sequences for T such that limn→∞ kxn − ynk exists. Then there is an
approximate fixed point sequence (zn) in K such that
lim
n→∞
kxn − znk = lim
n→∞
kyn − znk =
1
2
lim
n→∞
kxn − ynk.
In the present paper, we define a Banach space X isomorphic to c0 and we prove
that this space has the fixed point property. Our interest on this space derives
from several reasons. Firstly, the space X comes from c0 in a natural way. In fact,
the Schauder basis of X is equivalent to the summing basis of c0. Secondly, the
space X is close to c0 in the sense that the Banach-Mazur distance between the two
spaces is equal to 2. It is worth mentioning that from the proof of Theorem 1.2 we
can conclude that whenever Y is a Banach space isomorphic to c0 and the Banach-
Mazur distance between Y and c0 is strictly less than 2, then Y has the fixed point
property. In our case, the Banach-Mazur distance is equal to 2, that is the space X
lies on the boundary of what is already known. This fact should also be compared
with the following question in metric fixed point theory: Find a nontrivial class of
Banach spaces invariant under isomorphism such that each member of the class has
the fixed point property (a trivial example is the class of spaces isomorphic to ℓ1).
We shall see that even for spaces close to c0, such as the space X, the situation is
quite complicated and this points out the difficulty of the aforementioned question.
Finally, the space X has been used in several places in the study of the geometry of
Banach spaces (for instance see [5], [2]). More precisely, the well-known Hagler Tree
space (HT ) [5] contains a plethora of subspaces isomorphic to X. Nevertheless, we
do not know if HT has the fixed point property.
2. Definition and basic properties
We consider the vector space c00 of all real-valued finitely supported sequences.
We let (en)n∈N stand for the usual unit vector basis of c00, that is en(i) = 1 if i = n
and en(i) = 0 if i 6= n. If S ⊂ N is any interval of integers and x = (xi) ∈ c00 then
we set S ∗(x) = Pi∈S xi. We now define the norm of x as follows
kxk = supS ∗(x)
where the supremum is taken over all finite intervals S ⊂ N. The space X is the
completion of the normed space we have just defined.
THE FIXED POINT PROPERTY IN A BANACH SPACE ISOMORPHIC TO c0
3
i=1 xiei.
i=1 xiei ∈ X we have e∗
It is easily verified that the sequence (en) is a normalized monotone Schauder ba-
sis for the space X. In the following, (e∗
n)n∈N denotes the sequence of the biorthog-
onal functionals and (Pn)n∈N denotes the sequence of the natural projections asso-
ciated to the basis (en). That is, for any x = P∞
n(x) = xn
and Pn(x) = Pn
Furthermore, if S ⊂ N is any interval of integers (not necessarily finite), we define
the functional S ∗ : X → R by S ∗(x) = S ∗(P∞
i=1 xiei) = Pi∈S xi. It is easy to see
that S ∗ is a bounded linear functional with kS ∗k = 1. In the special case where
S = N, the corresponding functional is denoted by B ∗ (instead of the confusing
N∗). Therefore, B ∗(x) = P∞
The following proposition provides some useful properties of the space X and
demonstrates the relation between X and c0. We remind that for any pair E, F of
isomorphic normed spaces, the Banach-Mazur distance between E and F is defined
as follows
i=1 xi for any x = P∞
i=1 xiei ∈ X.
d(E, F ) = inf{kT k · kT −1k T : E → F is an isomorphism from E onto F }.
Proposition 2.1. The following hold:
(1) The space X is isomorphic to c0 and in particular the basis of X is equiv-
alent to the summing basis of c0.
(2) The subspace of X ∗ generated by the sequence of the biorthogonal function-
als has codimension 1. More precisely, X ∗ = span{e∗
n}n∈N ⊕ hB ∗i.
(3) The Banach-Mazur distance d(X, c0) between X and c0 is equal to 2.
Proof. We define the linear operator
Φ : X → c0
x = (xi) 7→ (cid:16)
∞
X
i=1
xi,
∞
X
i=2
xi, . . .(cid:17).
It is easily verified that Φ is an isomorphism from X onto c0 with kΦk = 1, kΦ−1k =
2 and Φ maps the basis of X to the summing basis of c0. This proves the first
assertion. The second assertion is an immediate consequence of the relation between
X and c0 established above.
It remains to show that the Banach-Mazur distance d = d(X, c0) is equal to 2.
Firstly, we observe that the isomorphism Φ defined above implies that d ≤ 2. In
order to prove the reverse inequality we fix a real number ǫ > 0. Then there exists
an isomorphism T : X → c0 from X onto c0 such that kxk ≤ kT xkc0 ≤ (d + ǫ)kxk
for any x ∈ X. We now consider the normalized sequence (xn) in X where xn =
(xn(i))i∈N is defined by
xn(2n − 1) = −1, xn(2n) = 1, xn(i) = 0 otherwise.
The description of X ∗ given by the second assertion implies that any bounded
m(tn) → 0
sequence (tn)n∈N of elements of X converges weakly to 0 if and only if e∗
for every m ∈ N and B ∗(tn) → 0. It follows that the sequence (xn)n∈N defined
above is weakly null. Now we set yn = T (xn) for any n ∈ N and we have 1 ≤
kynkc0 ≤ d + ǫ and (yn)n∈N converges weakly to 0. Therefore, we find k1 ∈ N
such that the vectors y1 and yk1 have essentially disjoint supports. More precisely,
since y1 ∈ c0, there exists N1 ∈ N such that y1(i) < ǫ for any i > N1. Since
yn → 0 weakly, we find k1 so that yk1(i) < ǫ for any i ≤ N1.
It follows that
4
COSTAS POULIOS
ky1−yk1kc0 ≤ max{ky1kc0, kyk1kc0}+ǫ ≤ d+2ǫ. On the other hand, kx1 −xk1 k = 2.
Therefore,
2 = kx1 − xk1 k ≤ ky1 − yk1 kc0 ≤ d + 2ǫ.
If ǫ tends to 0, we obtain 2 ≤ d as we desire.
(cid:3)
3. The fixed point property
This section is entirely devoted to the proof of the fixed point property for the
space X. First we need to establish some notation. If S, S ′ ⊂ N are intervals we
write S < S ′ to mean that max S < min S ′. Moreover, if k ∈ N, we write k < S
(resp., S < k) to mean k < min S (resp., max S < k). Finally, for any x = (xi) ∈ X,
supp(x) = {i ∈ N xi 6= 0} denotes the support of x.
Theorem 3.1. The space X has the fixed point property.
Proof. We follow the standard approach. We assume that K is a weakly compact,
convex subset of X which is minimal for a non-expansive map T : K → K. Us-
ing the geometry of the space X, we have to show that K is a singleton, that is
diam(K) = 0. Let us suppose that diam(K) > 0 and now we have to reach a
contradiction. Without loss of generality we may assume that diam(K) = 1.
Let (xn)n∈N be an approximate fixed point sequence for the map T in the set
K. By passing to a subsequence and then using some translation, we may assume
that 0 ∈ K and (xn) converges weakly to 0. Theorem 1.1 implies that limn kxnk =
diam(K) = 1.
We next find a subsequence (xqn ) of (xn) and integers l0 = 0 < l1 < l2 < . . . such
that for every n ∈ N, kPln−1(xqn )k < 1/n and kxqn − Pln(xqn )k < 1/n. The desired
sequences (xqn ) and (ln) are constructed inductively. We start with xq1 = x1 and
l0 = 0. Suppose that q1 < q2 < . . . < qn and l0 < l1 < . . . < ln−1 have been
defined. Then there exists ln > ln−1 such that kxqn − Pln (xqn )k < 1/n. Since (xn)
is weakly null, it follows that Pm(xn) → 0 for every m ∈ N. Therefore, there exists
qn+1 > qn such that kPln (xqn+1 )k < 1
n+1 . The construction of (xqn ) and (ln) is
complete.
Consequently, passing to a subsequence, we may assume that for the original
sequence (xn) there are integers l0 = 0 < l1 < l2 < . . . such that for every n ∈ N,
kPln−1(xn)k <
1
n
and kxn − Pln (xn)k <
1
n
.
As a matter of fact, we can go one step further and suppose that for any n ∈ N,
Pln−1(xn) = 0 and xn − Pln (xn) = 0. Therefore, supp(xn) ⊂ (ln−1, ln], that is (xn)
is a block basis of (en). If we did not adopt this assumption, then in each inequality
written below we would have to add a term equal to O( 1
n ), which simply would
change nothing.
We next consider the subsequences (zn) = (x2n−1) and (yn) = (x2n) and we also
set l2n−1 = kn and l2n = mn for every n ∈ N. The properties of the sequence (xn)
imply that the following hold.
(1) (zn) and (yn) are approximate fixed point sequences for the map T . There-
fore, by Theorem 1.1, lim kznk = lim kynk = 1.
(2) (zn) and (yn) converge weakly to 0.
(3) supp(zn) ⊂ (mn−1, kn] and supp(yn) ⊂ (kn, mn] for every n ∈ N.
(4) lim kzn − ynk = 1.
THE FIXED POINT PROPERTY IN A BANACH SPACE ISOMORPHIC TO c0
5
In order to justify the fourth conclusion, we first observe that lim kzn − ynk ≤
diam(K) = 1. On the other hand, by the definition of the norm of the space X, for
every n ∈ N there exists a finite interval En ⊂ N such that kznk = E ∗
n(zn). Clearly
we may assume that En ⊂ (mn−1, kn]. Then kzn − ynk ≥ E ∗
n(zn − yn) = kznk and
therefore lim kzn − ynk ≥ lim kznk = 1.
We are ready now to apply Maurey's theorem (Theorem 1.3). To this end, we
fix a positive integer N ∈ N, which will be chosen properly at the end of the proof,
and we set ǫ = 2−N . After N iterated applications of Theorem 1.3 we find a
sequence (vn) in the set K such that: (vn) is an approximate fixed point sequence
for the map T (which implies that lim kvnk = 1) and further lim kvn − znk = ǫ and
lim kvn − ynk = 1 − ǫ. Therefore, for all sufficiently large n ∈ N the following hold:
(1) kvnk > 1 − ǫ
2 ;
(2) kvn − znk < 3ǫ/2 and kvn − ynk < 1 − ǫ
2 ;
(3) B ∗(zn) < ǫ/2 (since (zn) is weakly null).
We also set Sn = (mn−1, kn] so that we have S1 < S2 < . . .. Concerning the
sequence (vn) in the set K and the sequence of intervals (Sn) we prove the following
two claims.
Claim 1. For all sufficiently large n, the support of vn is essentially contained
in the interval Sn, in the sense that if S is any interval with S ∩ Sn = ∅ then
S ∗(vn) < 3ǫ/2.
Indeed, we know that supp(zn) ⊂ (mn−1, kn] = Sn. Therefore, if S is any
interval with S ∩ Sn = ∅ then S ∗(zn) = 0 and hence
S ∗(vn) = S ∗(vn − zn) ≤ kvn − znk <
3ǫ
2
.
Claim 2. For all sufficiently large n, there exist intervals Ln < Rn such that Sn =
Ln ∪ Rn and L∗
n(vn) < −1 + 7ǫ, R∗
n(vn) > 1 − 2ǫ.
We fix a sufficiently large positive integer n. Since kvnk > 1 − ǫ
2 , it follows that
there exists a finite interval Fn ⊂ N such that F ∗
2 . If kn < Fn, we know
n (vn) < 3ǫ/2, which is a contradiction. Moreover,
by the previous claim that F ∗
if we assume that Fn ≤ kn then Fn ∩ (kn, mn] = ∅ and the choice of (yn) implies
F ∗
n (vn) > 1− ǫ
n (yn) = 0. Thus,
F ∗
n (vn) = F ∗
n (vn − yn) ≤ kvn − ynk < 1 −
ǫ
2
,
which is also a contradiction. By this discussion it is clear that min Fn ≤ kn <
max Fn. Now we set Rn = Fn ∩ [1, kn] and we estimate
1 −
ǫ
2
< F ∗
n (vn) ≤ R∗
n(vn) + (Fn \ Rn)∗(vn) < R∗
n(vn) +
3ǫ
2
,
n(vn) > 1 − 2ǫ. Passing
where the last inequality follows by Claim 1. Therefore, R∗
n(vn) > 1 − 2ǫ for all sufficiently
to a subsequence, we may assume that either R∗
large n or R∗
n(vn) < −1 + 2ǫ for all sufficiently large n. We suppose that the
first possibility happens, as the second one is treated similarly (interchanging the
roles of Ln and Rn). Consequently, for the interval Rn we have max Rn = kn and
R∗
n(vn) > 1 − 2ǫ.
On the other hand, we observe that
B ∗(vn) ≤ B ∗(vn − zn) + B ∗(zn) ≤ kvn − znk +
ǫ
2
< 2ǫ.
6
COSTAS POULIOS
We note that the sequence (vn) is not necessarily weakly null. However, vn is close
to zn and hence B ∗(vn) is very small. We next set Gn = [1, min Rn) (possibly
empty) and Wn = (kn, +∞). Then,
2ǫ > B ∗(vn) = G∗
≥ R∗
n(vn) + R∗
n(vn) − G∗
n(vn) + W ∗
n(vn) − W ∗
n (vn)
n (vn)
> 1 − 2ǫ − G∗
n(vn) −
3ǫ
2
.
Therefore Gn is non-empty and G∗
then it would follow
n(vn) > 1 − 11ǫ
2 . However, if G∗
n(vn) > 1 − 11ǫ
2 ,
B ∗(vn) ≥ R∗
n(vn) + G∗
n (vn) ≥ 2 − 9ǫ,
n(vn) − W ∗
n(vn) < −1 + 11ǫ
which is a contradiction. Hence, G∗
2 . Further, we observe that we
can not have Gn < Sn, since in this case it would follow G∗
2 . Conse-
quently, max Gn > mn−1 which clearly implies min Rn > mn−1 + 1. Finally, we set
Ln = Gn ∩ (mn−1, kn] and we estimate
n(vn) < 3ǫ
−1 +
11ǫ
2
> G∗
n(vn) = L∗
n(vn) + (Gn \ Ln)∗(vn) ≥ L∗
n(vn) −
3ǫ
2
.
We deduce that L∗
following: Sn = Ln ∪ Rn, R∗
claim is now complete.
n(vn) < −1 + 7ǫ. Therefore, the intervals Ln < Rn satisfy the
n(vn) < −1 + 7ǫ. The proof of the
n(vn) > 1 − 2ǫ and L∗
Using the construction and the properties of the sequences (vn) and (Sn), we
can reach the final contradiction and finish the proof of the theorem. Indeed, we
fix a sufficiently large n ∈ N and we consider the intervals D = (kn, mn] and
S = Rn ∪ D ∪ Ln+1. Then, using Claim 1 and Claim 2 we have
S ∗(vn) = R∗
n(vn) + (D ∪ Ln+1)∗(vn) > 1 − 2ǫ −
3ǫ
2
= 1 −
7ǫ
2
S ∗(vn+1) = (Rn ∪ D)∗(vn+1) + L∗
n+1(vn+1) <
3ǫ
2
− 1 + 7ǫ = −1 +
17ǫ
2
.
Therefore,
kvn − vn+1k ≥ S ∗(vn − vn+1) = S ∗(vn) − S ∗(vn+1) ≥ 2 − 12ǫ.
If ǫ has been chosen small enough, then we have a contradiction, since kvn−vn+1k ≤
diam(K) = 1.
(cid:3)
References
1. D. Alspach, A fixed point free nonexpansive map, Proc. Amer. Math. Soc. 82 (1981), 423-424.
2. S. A. Argyros, I. Deliyanni and A. G. Tolias, Hereditarily indecomposable Banach Algebras of
diagonal operators, Israel J. Math. 181 (2011), 65-110.
3. F. E. Browder, Nonexpansive nonlinear operators in Banach spaces, Proc. Nat. Acad. Sci.
U.S.A. 54 (1965), 1041-1044.
4. J. Elton, P. Lin, E. Odell and S. Szarek, Remarks on the fixed point problem for nonexpansive
maps, Contemporary Math. 18 (1983), 87-119.
5. J. Hagler, A counterexample to several questions about Banach spaces, Studia Math. 60
(1977), 289-308.
6. L. A. Karlovitz, Existence of fixed points for nonexpansive mappings in a space without normal
structure, Pacific J. Math. 66 (1976), 153-159.
7. W. A. Kirk, A fixed point theorem for mappings which do not increase distances, Amer. Math.
Monthly 72 (1965), 1004-1006.
THE FIXED POINT PROPERTY IN A BANACH SPACE ISOMORPHIC TO c0
7
8. B. Maurey, Points fixes des contractions sur un convexe forme de L1, Seminaire d' Analyse
Fonctionelle 80-81, Ecole Polytechnique, Palaiseau.
Department of Mathematics, University of Athens, 15784, Athens, Greece
E-mail address: [email protected]
|
1305.7276 | 1 | 1305 | 2013-05-31T00:48:34 | An abstract result on Cohen strongly summing operators | [
"math.FA"
] | We present an abstract result that characterizes the coincidence of certain classes of linear operators with the class of Cohen strongly summing linear operators. Our argument is extended to multilinear operators and, as a consequence, we establish a few alternative characterizations for the class of Cohen strongly summing multilinear operators. | math.FA | math |
AN ABSTRACT RESULT ON COHEN STRONGLY SUMMING OPERATORS
JAMILSON R. CAMPOS
Abstract. We present an abstract result that characterizes the coincidence of certain classes of linear
operators with the class of Cohen strongly summing linear operators. Our argument is extended to
multilinear operators and, as a consequence, we establish a few alternative characterizations for the
class of Cohen strongly summing multilinear operators.
1. Introduction
With the work of A. Dvoretzky and C. Rogers [6], it became established that in infinite dimensional
spaces always exists unconditionally convergent series which does not converge absolutely. So, A.
Grothendieck [7] introduces the conceptual essence of absolutely summing operators as those that
improve the convergence of series, towards transforming an unconditionally convergent series in an
absolutely convergent one. After that, many research efforts are aimed to generalize the concept of
absolutely summing operators and make Grothendieck's ideas clearer, such as works of B. Mitiagin and
A. Pelczy´nski [11] and J. Lindenstrauss and A. Pelczy´nski [9]. We recommend the paper [12] for a
panorama on absolutely summing operators.
In 1967, Pietsch ([15], pg. 338) shows that the identity operator from l1 into l2 is absolutely 2-
summing while its conjugate, from l2 into l∞, is not absolutely 2-summing. Motivated by this fact, J. S.
Cohen [5] introduces the class of strongly p-summing linear operators which characterizes the conjugate
of the class of absolutely p∗-summing linear operators, with 1/p + 1/p∗ = 1. In his work, Cohen proves
several results for this new class such as inclusion relations and a Pietsch domination type theorem.
The concept of Cohen strongly summing multilinear operator was introduced and studied by D.
Achour and L. Mezrag [1] and Mezrag and K. Saadi [10] established some inclusion results between the
class of Cohen strongly summing multilinear operators and other classes of operators such as the class
of multiple summing operators. A related concept and new generalizations of concept of Cohen strongly
summing multilinear operators have been recently studied, such as the classes of almost Cohen strongly
multilinear operators [3] and the class of multiple Cohen strongly summing multilinear operators [4].
In this paper, we prove an abstract result derived from the Full General Pietsch Domination Theorem
([14], Theorem 4.6 page 1256) which has an immediate application regarding the class of Cohen strongly
summing linear operators. This result is also extended to the multilinear case and with that we can
establish alternative definitions for Cohen strongly summing multilinear operators. Through these
definitions, it is possible to characterize the class of Cohen strongly summing multilinear operators by
means of arbitrary sequences.
Key words and phrases. Cohen strongly p-summing operators; full general Pietsch Domination Theorem.
2010 Mathematics Subject Classification: 46G25, 47H60, 47L22.
1
2
JAMILSON R. CAMPOS
2. Notation and background
Let n ∈ N, E, E1, ..., En and F be Banach spaces over K = R or C. The space of all continuous linear
operators from E into F will be represented by L(E; F ). We will denote by L(E1, ..., En; F ) the space
of all continuous n-linear operators from E1, ..., En into F and if E1 = · · · = En we just write L(nE; F ).
i=1 in E with the
If 1 ≤ p ≤ ∞, we will denote by lp(E) and lw
p (E) the spaces of all sequences (xi)∞
norms
(xi)∞
i=1p = ∞Xi=1
′
xip!1/p
< ∞ and (xi)∞
i=1w,p = sup
ψ∈B
E
′ ∞Xi=1
ψ(xi)p!1/p
< ∞ ,
respectively, where E
represents the topological dual of E and BE denotes the closed unit ball of E.
Definition 2.1 (Cohen, [5]). A sequence (xi)∞
i=1 ϕi(xi) converges for all (ϕi)∞
the series P∞
of management, to replace the seriesP∞
use in our text.
We denote by lphEi the space of Cohen strongly p-summing sequences in E. It is possible, for reasons
i=1 ϕi(xi), which we
i=1 ϕi(xi) in Definition 2.1 by the seriesP∞
The space lphEi is a Banach space ([8], page 223) with the norm
i=1 in a Banach space E is Cohen strongly p-summing if
i=1 ∈ lw
p∗ (E
′
), with 1/p + 1/p∗ = 1.
(xi)∞
i=1C,p =
sup
i=1w,p∗ ≤1
(ϕi)∞
ϕi(xi) .
∞Xi=1
Definition 2.2 (Cohen, [5]). Let 1 < p ≤ ∞. An operator T ∈ L(E; F ) is Cohen strongly p-summing
if there exists a constant C > 0 such that for all m ∈ N, xi ∈ E and ϕi ∈ F
, i = 1, ..., m,
′
In an equivalent manner, an operator T ∈ L(E; F ) is Cohen strongly p-summing if the operator
ϕi(T (xi)) ≤ C (xi)m
mXi=1
bT : lp (E) → lphF i ; (xi)∞
i=1p(ϕi)m
i=1w,p∗.
i=1 7→ (T (xi))∞
i=1
(2.1)
(2.2)
is well-defined and continuous. The space of all Cohen strongly p-summing linear operators will be
denoted by Dp(E; F ). The smallest C such that (2.1) is satisfied defines a norm on Dp(E; F ), denoted
by dp(·).
Definition 2.3 (Achour-Mezrag, [1]). Let 1 < p ≤ ∞, Ej , F be Banach spaces, j = 1, ..., n. An
continuous n-linear operator T : E1 × ... × En → F is Cohen strongly p-summing if there exists a
constant C > 0 such that for all x(j)
m ∈ Ej and ϕ1, ..., ϕm ∈ F
and for all m ∈ N,
1 , ..., x(j)
′
mXi=1(cid:12)(cid:12)(cid:12)ϕi(cid:16)T(cid:16)x(1)
i
, ..., x(n)
i (cid:17)(cid:17)(cid:12)(cid:12)(cid:12) ≤ C
mXi=1
p
i (cid:13)(cid:13)(cid:13)
nYj=1(cid:13)(cid:13)(cid:13)x(j)
1/p
(ϕi)m
i=1w,p∗ .
The space of all Cohen strongly p-summing multilinear operators will be represented by
LCoh,p(E1, ..., En; F ). The smallest C such that (2.2) is satisfied defines a norm on LCoh,p(E1, ..., En; F ),
denoted by · Coh,p.
AN ABSTRACT RESULT ON COHEN STRONGLY SUMMING OPERATORS
3
An abstract very useful tool developed by Pellegrino et al. [14], the Full General Pietsch Domination
Theorem, allows to establish in a very simplified way various Pietsch Domination-type theorems for
many classes of operators. This approach was initiated in the works of G. Botelho, D. Pellegrino and
P. Rueda [2] and D. Pellegrino and J. Santos [13].
Let X1, ..., Xn, Y and E1, ..., Er be arbitrary non-empty sets, H be a family of operators from
X1 × · · · × Xn into Y . Let K1, ..., Kt be compact Hausdorff topological spaces, G1, ..., Gt be Banach
spaces and suppose that the mappings
(cid:26) Rj : Kj × E1 × · · · × Er × Gj → [0, ∞) , j = 1, ..., t ,
S : H × E1 × · · · × Er × G1 × · · · × Gt → [0, ∞)
have the following properties:
1) For each x(l) ∈ El and b ∈ Gj , with (j, l) ∈ {1, ..., t} × {1, ..., r}, the mapping
(Rj )x(1),...,x(r),b : Kj → [0, ∞) ,
2) The following inequalities hold:
defined by (Rj )x(1),...,x(r),b(ϕ) = Rj(cid:0)ϕ, x(1), ..., x(r), b(cid:1), is continuous;
(cid:26) Rj(cid:0)ϕ, x(1), ..., x(r), ηjb(j)(cid:1) ≤ ηjRj(cid:0)ϕ, x(1), ..., x(r), b(j)(cid:1)
S(cid:0)f, x(1), ..., x(r), α1b(1), ..., αtb(t)(cid:1) ≥ α1...αtS(cid:0)f, x(1), ..., x(r), b(1), ..., b(t)(cid:1) ,
for all ϕ ∈ Kj, x(l) ∈ El (with l = 1, ..., r), 0 ≤ ηj, αj ≤ 1, b(j) ∈ Gj, j = 1, ..., t and
f ∈ H.
Under these conditions, we have the following definition and theorem obtained from [14]:
Definition 2.4 (Pellegrino et al. [14]). Let 0 < p1, ..., pt, p0 < ∞, with 1
. An application
p0
f : X1 × · · · × Xn → Y belonging to H is R1, ..., Rt-S-abstract (p1, ..., pt)-summing if there is a constant
C > 0 such that
+· · ·+ 1
pt
= 1
p1
mXi=1(cid:16)S(cid:16)f, x(1)
i
, ..., x(r)
i
, b(1)
i
, ..., b(t)
i (cid:17)(cid:17)p0!1/p0
≤ C
ϕ∈Kk mXi=1
sup
tYk=1
Rk(cid:16)ϕ, x(1)
i
, ..., x(r)
i
, b(k)
i (cid:17)pk!1/pk
,
for all x(s)
1 , ..., x(s)
m ∈ Es, b(s)
1 , ..., b(s)
m ∈ Gl, m ∈ N and (s, l) ∈ {1, ..., r} × {1, ..., t}.
Theorem 2.5 (Full General Pietsch Domination Theorem, Pellegrino et al. ([14])). An application
f ∈ H is R1, ..., Rt-S-abstract (p1, ..., pt)-summing if and only if there exist a constant C > 0 and Borel
probability measures µk in Kk, k = 1, ..., t, such that
S(cid:16)f, x(1), ..., x(r), b(1), ..., b(t)(cid:17) ≤ C
tYk=1(cid:18)ZKk
Rk(cid:16)ϕ, x(1), ..., x(r), b(k)(cid:17)pk
dµk(cid:19)1/pk
,
for all x(l) ∈ El, l = 1, ..., r and b(k) ∈ Gk, with k = 1, ..., t.
We will denote by C(q0,q1;p)(E; F ) the class of all operators T ∈ L (E; F ) such that exists a constant
C > 0 satisfying
Γ =(cid:26)(q0, q1) ∈ [1, ∞ ) × (1, ∞) :
ϕi (T (xi))q0
≤ C k(xi)m
1/q0
′
mXj=1
1
q0
=
1
q1
+
1
p∗(cid:27) .
i=1kq1
k(ϕi)m
i=1kw,p∗ ,
4
JAMILSON R. CAMPOS
3. The linear case
We show that the class of Cohen strongly p-summing linear operators can be characterized by means
of different inequalities. This means a coincidence result between classes of linear operators.
Let p∗ ∈ (1, ∞) with 1 = 1
p + 1
p∗ and
for all positive integers m and all xi ∈ E, ϕi ∈ F
It follows immediately from Definition 2.1 that
, i = 1, ..., m.
Dp(E; F ) ⊂ C(q0,q1;p)(E; F ),
for all (q0, q1) ∈ Γ.
Another noteworthy fact is that the class C(q0,q1;p)(E; F ) is trivial if p < q1. The proof of this fact is
given by the following proposition:
Proposition 3.1. Let 1 < p < ∞, with 1 = 1/p + 1/p∗, and (q0, q1) ∈ Γ, that is 1/q0 = 1/q1 + 1/p∗. If
p < q1, then C(q0,q1;p)(E; F ) = {0}.
Proof. The case E = {0} or F = {0} is immediate. Let us suppose that E 6= {0} and F 6= {0}. Let
(3.1)
(λi)∞
i=1 = (αiβi)∞
i=1 /∈ lq0 with (αi)∞
i=1 ∈ lp∗ and (βi)∞
i=1 ∈ lq1.
If T ∈ C(q0,q1;p)(E; F ), 0 6= x ∈ E and 0 6= ϕ ∈ F
′
, then, for all m ∈ N,
mXi=1
ϕ(T (λix))q0!1/q0
= mXi=1
= mXi=1
ϕ(T (αiβix))q0!1/q0
(αiϕ)(T (βix))q0!1/q0
≤ C(βix)m
i=1q1 (αiϕ)m
i=1w,p∗ .
Thus,
ϕ(T (x)) mXi=1
λiq0!1/q0
≤ Cx (βi)m
i=1q1
sup
ψ∈B
F
ψ(αiϕ)p∗!1/p∗
αip∗!1/p∗
′′ mXi=1
ψ(ϕ) mXi=1
αip∗!1/p∗
.
′′
= Cx (βi)m
i=1q1
sup
ψ∈B
F
= Cx (βi)m
i=1q1 ϕ mXi=1
AN ABSTRACT RESULT ON COHEN STRONGLY SUMMING OPERATORS
5
Taking thereafter the supϕ≤1 and supx≤1 in the above inequality, it follows that
T mXi=1
λiq0!1/q0
≤ C (βi)m
i=1q1 mXi=1
αip∗!1/p∗
.
Therefore we conclude that if T 6= 0 then (λi)∞
i=1 ∈ lq0 , which contradicts the assertion (3.1).
(cid:3)
Under all conditions and notations given above we will establish the following theorem. For our case,
will be sufficient to consider r = 1 and t = 2 in Definition 2.4 and Theorem 2.5:
Theorem 3.2. Let f : X → Y be an application belonging to H and let
such that
0 < q0, q1, p0, p1, p∗ < ∞ ,
1
q0
=
1
q1
+
1
p∗
e
1
p0
=
1
p1
+
1
p∗ .
If (R1)(x,b)(·) is constant, for each x and for each b, then the following statements are equivalent:
(i) f is R1, R2-S-abstract (q1, p∗)-summing;
(ii) f is R1, R2-S-abstract (p1, p∗)-summing.
Proof. By Theorem 2.5, f is R1, R2-S-abstract (q1, p∗)-summing if and only if there exist a constant
C > 0 and Borel probability measures µi in Ki, i = 1, 2, such that
S(f, x, b1, b2) ≤ C
2Yi=1(cid:18)ZKi
Ri(ϕ, x, bi)pi dµi(cid:19)1/pi
.
In our case, f is R1, R2-S-abstract (q1, p∗)-summing if and only if there exist a constant C > 0 and a
Borel probability measure µ in K2 such that
(3.2)
since, by hypothesis, for any fixed ϕ ∈ K1,
S(f, x, b1, b2) ≤ C (R1(ϕ, x, b1)) ·(cid:18)ZK2
R1(ϕ, x, b1)q1 dµ1(cid:19)1/q1
(cid:18)ZK1
R2(ϕ, x, b2)p∗
dµ(cid:19)1/p∗
,
= R1(ϕ, x, b1).
On the other hand, the same reasoning shows that f is R1, R2-S-abstract (p1, p∗)-summing if and only
if there exist a constant C > 0 and a Borel probability measure µ in K2 such that
S(f, x, b1, b2) ≤ C (R1(ϕ, x, b1)) ·(cid:18)ZK2
R2(ϕ, x, b2)p∗
dµ(cid:19)1/p∗
,
and this expression corresponds exactly to the one given by (3.2).
(cid:3)
The consequence of the above theorem is the coincidence Dp(E; F ) = C(q0,q1;p)(E; F ), which implies
several ways to characterize the class of Cohen strongly summing linear operators. This is shown by
the following corollary:
6
JAMILSON R. CAMPOS
Corollary 3.3. For all (q0, q1) , (p0.p1) ∈ Γ,
In particular,
for all (q0, q1) ∈ Γ.
C(q0,q1;p)(E; F ) = C(p0,p1;p)(E; F ).
C(q0,q1;p)(E; F ) = C(1,p;p)(E; F ) = Dp(E; F ) ,
Proof. Let T ∈ C(q0,q1;p)(E; F ). Since 1
q0
= 1
q1
+ 1
p∗ , with the choices
t = 2, r = 1
E1 = {0}, X1 = E, Y = F
f = T, H = L(E; F )
K1 = {0} and K2 = BF ′′
G1 = E, and G2 = F
p0 = q0, p1 = q1, and p2 = p∗
S (T, 0, x, ϕ) = ϕ (T (x))
R1 (γ, 0, x) = kxk
R2(ψ, 0, ϕ) = ψ(ϕ)
′
(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)(cid:12)(cid:12)(cid:12)
it follows that
and also
mXi=1(cid:16)S(cid:16)T, xi, b(1)
i
, b(2)
i (cid:17)(cid:17)p0!1/p0
(S (T, 0, xi, ϕi))q0!1/q0
ϕi (T (xi))q0!1/q0
,
= mXi=1
= mXi=1
i (cid:17)pk!1/pk
sup
ϕ∈Kk mXi=1
2Yk=1
γ∈K1 mXi=1
= mXi=1
= sup
= k(xi)m
i=1kq1
Rk(cid:16)ϕ, xi, b(k)
R1 (γ, 0, xi)q1!1/q1
′′ mXi=1
sup
ψ∈B
·
Y
(ϕi)m
i=1w,p∗ .
·
kxikq1!1/q1
R2(ψ, 0, ϕi)p∗!1/p∗
· sup
ψ∈K2 mXi=1
ψ (ϕi) p∗!1/p∗
Thus, by Definition 2.4, it follows that T is R1, R2-S-abstract (q1, p∗)-summing and, by Theorem 3.2,
T is R1, R2-S-abstract (p1, p∗)-summing. Therefore there is a constant C > 0 such that
mXj=1
ϕi (T (xi))p0
1/p0
≤ C k(xi)m
i=1kp1
k(ϕi)m
i=1kw,p∗ ,
for all positive integers m and for all xi ∈ E, ϕi ∈ F
′
, i = 1, ..., m. So T ∈ C(p0,p1;p)(E; F ).
The other inclusion is obtained by the same argument.
(cid:3)
AN ABSTRACT RESULT ON COHEN STRONGLY SUMMING OPERATORS
7
4. The multilinear case and applications
In our recent work [4], we establish an alternative definition for the concept of Cohen strongly
summing multilinear operators given by Achour-Mezrag (Definition 2.3):
Theorem 4.1 (Campos, [4]). Let 1 < p ≤ ∞, with 1/p + 1/p∗ = 1. For T ∈ L(X1, ..., Xn; Y ), the
following statements are equivalent:
(i) There is a constant C > 0 such that
for all m ∈ N, x(j)
(ii) There is a constant C > 0 such that
i ∈ Xj, ϕi ∈ Y
′
mXi=1(cid:12)(cid:12)(cid:12)ϕi(cid:16)T(cid:16)x(1)
i
1/p
(ϕi)m
, ..., x(n)
mXi=1
p
i (cid:17)(cid:17)(cid:12)(cid:12)(cid:12) ≤ C
i (cid:13)(cid:13)(cid:13)
nYj=1(cid:13)(cid:13)(cid:13)x(j)
np!1/np
np!1/np
... mXi=1(cid:13)(cid:13)(cid:13)x(n)
i (cid:17)(cid:17)(cid:12)(cid:12)(cid:12) ≤ C mXi=1(cid:13)(cid:13)(cid:13)x(1)
i (cid:13)(cid:13)(cid:13)
i (cid:13)(cid:13)(cid:13)
, i = 1, ..., m , j = 1, ..., n ;
, i = 1, ..., m , j = 1, ..., n.
mXi=1(cid:12)(cid:12)(cid:12)ϕi(cid:16)T(cid:16)x(1)
i
, ..., x(n)
for all m ∈ N, x(j)
i ∈ Xj, ϕi ∈ Y
′
i=1w,p∗ ,
(ϕi)m
i=1w,p∗ ,
As a consequence of Theorem 4.1 and the following theorem, similarly to the linear case, we can
characterize the class of Cohen strongly summing multilinear operators by means of several inequalities.
To do this, we generalize the abstract result of Theorem 3.2.
Theorem 4.2. Let f : X1 × · · · × Xn → Y be an application belonging to H and let
0 < p∗, p0, q0, p1, ..., pt−1, q1, ..., qt−1 < ∞ ,
such that
1
p0
=
1
p1
+ · · · +
1
pt−1
+
1
p∗ and
1
q0
=
1
q1
+ · · · +
1
qt−1
+
1
p∗ .
If Rk(x1 ,...,xr ,b) (·) is constant for all x1, ..., xr, b and for all 1 ≤ k ≤ t − 1, then the following statements
are equivalent:
(i) f is R1, ..., Rt-S-abstract (p1, ..., pt−1, p∗)-summing;
(ii) f is R1, ..., Rt-S-abstract (q1, ..., qt−1, p∗)-summing.
Proof. Similarly to Theorem 3.2, it follows that f is R1, ..., Rt-S-abstract (p1, ..., pt−1, p∗)-summing if
and only if there exist a constant C > 0 and a Borel probability measure µ in Kt such that
Ri(ϕ, x1, ..., xr, bi)! ·(cid:18)ZKt
Rt(ϕ, x1, ..., xr, bt)p∗
dµ(cid:19)1/p∗
.
S(f, x1, ..., xr, b1, ..., bt) ≤ C t−1Yi=1
S(f, x1, ..., xr, b1, ..., bt) ≤ C t−1Yi=1
An analogous reasoning shows that f is R1, ..., Rt-S-abstract (q1, ..., qt−1, p∗)-summing if and only if
there exist a constant C > 0 and a Borel probability measure µ in Kt such that
Ri(ϕ, x1, ..., xr, bi)! ·(cid:18)ZKt
Rt(ϕ, x1, ..., xr, bt)p∗
dµ(cid:19)1/p∗
.
(cid:3)
8
JAMILSON R. CAMPOS
Corollary 4.3. For T ∈ L(E1, ..., En; F ), 1 = 1
statements are equivalent:
p + 1
p∗ and 1
q0
= 1
q1
+ · · · + 1
qn
+ 1
p∗ , the following
(i) There exists a constant C > 0 such that
mXi=1(cid:12)(cid:12)(cid:12)ϕi(cid:16)T(cid:16)x(1)
i
for all m ∈ N, x(j)
(ii) There exists a constant C > 0 such that
i ∈ Ej, ϕi ∈ F
′
np!1/np
... mXi=1(cid:13)(cid:13)(cid:13)x(n)
i (cid:13)(cid:13)(cid:13)
, i = 1, ..., m , j = 1, ..., n;
, ..., x(n)
np!1/np
i (cid:17)(cid:17)(cid:12)(cid:12)(cid:12) ≤ C mXi=1(cid:13)(cid:13)(cid:13)x(1)
i (cid:13)(cid:13)(cid:13)
q0!1/q0
mXi=1(cid:12)(cid:12)(cid:12)ϕi(cid:16)T(cid:16)x(1)
i (cid:17)(cid:17)(cid:12)(cid:12)(cid:12)
qn!1/qn
... mXi=1(cid:13)(cid:13)(cid:13)x(n)
q1!1/q1
≤ C mXi=1(cid:13)(cid:13)(cid:13)x(1)
i (cid:13)(cid:13)(cid:13)
i (cid:13)(cid:13)(cid:13)
, ..., x(n)
i ∈ Ej, ϕi ∈ F
, i = 1, ..., m , j = 1, ..., n .
i
′
(ϕi)m
i=1w,p∗ ,
(4.1)
for all m ∈ N, x(j)
(ϕi)m
i=1w,p∗ ,
Proof. With adequate choices and using Definition 2.4 (see [4], Theorem 3.8), if T satisfies (i) then T
is R1, ..., Rn+1-S-abstract (np, , ..., np, p∗)-summing. Using the same argument it can be shown that if
T satisfies (ii) then T is R1, ..., Rn+1-S-abstract (q1, ..., qn, p∗)-summing. So, by Theorem 4.2, we finish
the proof.
(cid:3)
From Corollary 4.3 it is also possible to establish more characterizations for the class of Cohen strongly
summing multilinear operators by means of sequences and, equivalently, by means of an inequality
similar to that given by (4.1), but adding the index i to infinity:
Let 1 < p ≤ ∞, Ej, F be Banach spaces, j = 1, ..., n and 1
q0
= 1
q1
+ · · · + 1
qn
+ 1
p∗ . An operator
T ∈ L(E1, ..., En; F ) is Cohen strongly p-summing if
This last definition amounts to saying that an operator T ∈ L(E1, ..., En; F ) is Cohen strongly p-
summing if There exists a constant C > 0 such that
, ..., x(n)
i (cid:17)(cid:17)(cid:17)∞
i=1
i
(cid:16)ϕi(cid:16)T(cid:16)x(1)
∞Xi=1(cid:12)(cid:12)(cid:12)ϕi(cid:16)T(cid:16)x(1)
i (cid:17)∞
whenever(cid:16)x(j)
i=1
i
q0!1/q0
, ..., x(n)
i (cid:17)(cid:17)(cid:12)(cid:12)(cid:12)
∈ lqj (Ej) , j = 1, ..., n .
i=1
i (cid:17)∞
∈ lq0 whenever (cid:16)x(j)
qn!1/qn
... ∞Xi=1(cid:13)(cid:13)(cid:13)x(n)
q1!1/q1
≤ C ∞Xi=1(cid:13)(cid:13)(cid:13)x(1)
i (cid:13)(cid:13)(cid:13)
i (cid:13)(cid:13)(cid:13)
′
∈ lqj (Ej) , j = 1, ..., n and (ϕi)∞
i=1 ∈ lw
p∗ (F
).
(ϕi)∞
i=1w,p∗ ,
These kinds of characterizations for the class of Cohen strongly summing multilinear operators appear
in our work [4] but without this abstract focus.
References
[1] D. Achour, L. Mezrag. On the Cohen strongly p-summing multilinear operators, Journal of Mathematical Analysis
and Applications, 327 (2007), 550 -- 563.
[2] G. Botelho, D. Pellegrino, P. Rueda. A unified Pietsch Domination Theorem, Journal of Mathematical Analysis and
Applications. 365 (2010), 269 -- 276.
[3] Q. Bu, Z. Shi. On Cohen almost summing multilinear operators, Journal of Mathematical Analysis and Applications,
401 (2013), 174 -- 181.
AN ABSTRACT RESULT ON COHEN STRONGLY SUMMING OPERATORS
9
[4] J. R. Campos. Cohen and Multiple Cohen strongly summing multilinear operators, Linear and Multilinear Algebra,
On-line version (2013), doi:10.1080/03081087.2013.779270.
[5] J. S. Cohen. Absolutely p-summing, p-nuclear operators and their conjugates, Mathematische Annalen, 201 (1973),
177 -- 200.
[6] A. Dvoretzky, C Rogers. Absolutely and unconditional convergence in normed linear spaces, Proceedings of the
National Academy of Sciences of USA, 36 (1950), 192 -- 197.
[7] A. Grothendieck. Sur certaines classes de suites dans les espaces de Banach et le thorme de Dvoretzky-Rogers, Boletim
da Sociedade Matemtica de So Paulo 8 (1956), 81 -- 110.
[8] R. Khalil. On some Banach space sequences, Bulletin of the Australian Mathematical Society, 25 (1982), 231 -- 241.
[9] J. Lindenstrauss, A. Pelczy´nski. Absolutely summing operator in Lp spaces and their applications, Studia
Mathematica, 29 (1968), 275 -- 326.
[10] L. Mezrag, K. Saadi. Inclusion theorems for Cohen strongly summing multilinear operators, Bulletin of the Belgian
Mathematical Society Simon Stevin, 16 (2009), 1 -- 11.
[11] B. Mitiagin, A. Pelczy´nski. Inclusion theorems for Cohen strongly summing multilinear operators, Bulletin of the
Belgian Mathematical Society Simon Stevin, 16 (2009), 1 -- 11.
[12] D. Pellegrino, J. Santos. Absolutely summing multilinear operators: a panorama, Quaestiones Mathematicae, 34
(2011), 447 -- 478.
[13] D. Pellegrino, J. Santos. A general Pietsch Domination Theorem, Journal of Mathematical Analysis and Applications,
375 (2011), 371 -- 374.
[14] D. Pellegrino, J. Santos, J. B. Seoane-Sep´ulveda. Some techniques on nonlinear analysis and applications, Advances
in Mathematics, 229 (2012), 1235 -- 1265.
[15] A. Pietsch. Absolut p-summierende Abbildungen in normiertn Rumen, Studia Mathematica, 28 (1967), 333 -- 353.
Departamento de Ciencias Exatas,
Universidade Federal da Para´ıba - Campus IV,
R. da Mangueira,
s/n, Centro,
Rio Tinto, 58.297-000, Brazil.
E-mail address: [email protected] and [email protected]
|
1805.03277 | 2 | 1805 | 2019-11-14T03:48:00 | On quasinilpotent operators and the invariant subspace problem | [
"math.FA"
] | We show that a bounded quasinilpotent operator $T$ acting on an infinite dimensional Banach space has an invariant subspace if and only if there exists a rank one operator $F$ and a scalar $\alpha\in\mathbb{C}$, $\alpha\neq 0$, $\alpha\neq 1$, such that $T+F$ and $T+\alpha F$ are also quasinilpotent. We also prove that for any fixed rank-one operator $F$, almost all perturbations $T+\alpha F$ have invariant subspaces of infinite dimension and codimension. | math.FA | math |
ON QUASINILPOTENT OPERATORS AND THE INVARIANT
SUBSPACE PROBLEM
ADI TCACIUC
Abstract. We show that a bounded quasinilpotent operator T acting on an infinite
dimensional Banach space has an invariant subspace if and only if there exists a rank
one operator F and a scalar α ∈ C, α 6= 0, α 6= 1, such that T + F and T + αF are
also quasinilpotent. We also prove that for any fixed rank-one operator F , almost all
perturbations T +αF have invariant subspaces of infinite dimension and codimension.
1. Introduction
One of the most important unsolved problem in Operator Theory is the Invariant
Subspace Problem: Does every bounded operator on an infinite dimensional, separable,
complex Hilbert space have a non-trivial invariant closed subspace? Von Neumann
proved the existence of such subspaces for compact operators acting on a separable
Hilbert space, a result which was extended by Aronszajn and Smith [5] to separable
Banach spaces. Lomonosov [18] greatly increases the class of operators with invariant
subspaces by showing that every operator commuting with a compact operator has
an invariant subspace. Enflo ([8, 9], see also [6]) constructed the first example of a
bounded operator on a (non-reflexive) Banach space which has no non-trivial invariant
subspaces, followed by a construction by Read [23]. Later Read produced several such
examples: strictly singular operators, quasinilpotent operators, and operators acting
on l1 (see [24], [26],[25]). All these examples are on non-reflexive Banach spaces, and
the Invariant Subspace Problem is still open for general reflexive Banach spaces. For
an overview of the Invariant Subspace Problem see the monographs by Radjavi and
Rosenthal [22] or the more recent book by Chalendar and Partington [7].
A very important special case for which the Invariant Subspace Problem is still open
is that of quasinilpotent operators on Hilbert spaces, or, more generally, on reflexive
Banach spaces. An operator T is called quasinilpotent if σ(T ) = {0}, where by σ(T )
we denote the spectrum of T . Substantial work has been devoted over the years to
ISP for quasinilpotent operators, in particular on Hilbert spaces. We mention several
2010 Mathematics Subject Classification. Primary: 47A15. Secondary: 47A55.
Key words and phrases. Operator, invariant subspace, finite rank, perturbation.
1
2
A. TCACIUC
important papers, without attempting to provide an exhaustive list: Apostol and
Voiculescu [4], Herrero [13], Foia¸s and Pearcy [10], Foia¸s, Jung, Ko, and Pearcy,[14,
11, 12].
In Section 2 we develop a method of investigating invariant subspaces for quasinilpo-
tent operators on complex Banach spaces by examining the resolvent function. In our
main result in this section, Theorem 2.3, we prove a necessary and sufficient condition
for a quasinilpotent operator to have invariant subspaces, a condition which is related
to the stability of the spectrum under rank-one perturbations.
Next we examine the existence of invariant half-spaces for rank-one perturbations
of quasinilpotent operators. By a half-space we understand a closed subspace which is
both infinite dimensional and infinite codimensional. A method of examining invari-
ant half-spaces for finite rank perturbations was introduced by Androulakis, Popov,
Tcaciuc, and Troitsky in [3], where the authors showed that certain classes of bounded
operators have rank-one perturbations which admit invariant half-spaces. In [21] Popov
and Tcaciuc showed that every bounded operator T acting on a reflexive Banach space
can be perturbed by a rank-one operator F such that T +F has an invariant half-space.
Moreover, when a certain spectral condition is satisfied, F can be chosen to have ar-
bitrarily small norm. Recently these results were extended to general Banach spaces
in [28]. In this line of investigation, Jung, Ko, and Pearcy [15, 16] adapted this theory
to operators on Hilbert spaces, where the presence of additional structure and specific
Hilbert space methods allowed them to prove important results regarding the matricial
structure of arbitrary operators on Hilbert spaces. For algebras of operators this type
of problems have been studied in [20], [19], and [27]. More control on the construction
of rank-one perturbation that have invariant half-spaces was achieved in [29]. In that
paper the authors showed that for any bounded operators T with countable spectrum
acting on a Banach space X, and for any non-zero x ∈ X, one can find a rank one
operator with range span{x} such that T + F has an invariant subspace.
In Section 3, we refine the method developed in the previous section to show that al-
most all (in a sense that is made precise in Theorem 3.3 below ) rank-one perturbations
of quasinilpotent operators have invariant half-spaces.
2. Invariant subspaces for quasinilpotent operators
For a Banach space X, we denote by B(X) the algebra of all (bounded linear)
operators on X. When T ∈ B(X), we write σ(T ), σp(T ),σess(T ), and ρ(T ) for the
spectrum of T , point spectrum of T , the essential point spectrum of T , and the resolvent
R(z) = (zI − T )−1 =
T i
zi+1 .
Xi=0
ON QUASINILPOTENT OPERATORS AND THE INVARIANT SUBSPACE PROBLEM
3
set of T , respectively. The closed span of a set {xn}n of vectors in X is denoted by
[xn].
For T ∈ B(X), the resolvent of T is the function R : ρ(T ) → B(X) defined by
R(z) = (zI − T )−1. When z > r(T ), where r(T ) is the spectral radius of T , the
resolvent is given by the Neumann series expansion
∞
In particular, when T is quasinilpotent this expansion holds for all complex numbers
z 6= 0. The resolvent R is analytic on ρ(T ), hence on C \ {0} when T is quasinilpotent.
We first prove a simple lemma which gives sufficient and necessary conditions for
λ ∈ ρ(T ) to be an eigenvalue for some fixed rank-one perturbation.
Lemma 2.1. Let X be a separable Banach space, T ∈ B(X), and F := e∗ ⊗ f a rank
one operator. Fix λ ∈ ρ(T ) and α ∈ C \ {0}. Then the following are equivalent:
(i) e∗((R(λ)f ) = α−1.
(ii) λ ∈ σp(T + αF ).
Proof. i)⇒ ii) We are going to show that y := R(λ)f is an eigenvector for T + αF ,
corresponding to the eigenvalue λ. Note that T y = λy − f . Then:
(T + αF )y = T y + αe∗(y)f = λy − f + αe∗(y)f = λy − f + αα−1f = λy.
ii)⇒ i)
Let y ∈ X be an eigenvector for T + αF corresponding to λ. Hence T y + αe∗(y)f =
λy. Note that since λ ∈ ρ(T ), it follows that e∗(y) 6= 0. We have:
T y + αe∗(y)f = λy ⇔ (λI − T )y = αe∗(y)f ⇔ y = αe∗(y)(λI − T )−1f.
Applying e∗ to both sides of the last equality, we get that
e∗(y) = αe∗(y)e∗((λI − T )−1f ),
and since e∗(y) 6= 0 it follows that e∗(R(λ)f ) = α−1.
(cid:3)
Remark 2.2. Note that when e∗(R(λ)f ) = α−1, from the proof of the previous lemma
it follows that R(λ)f is an eigenvector for T + αF corresponding to the eigenvalue λ.
We are now ready to prove the main theorem of this section.
Theorem 2.3. Let X be a separable Banach space and T ∈ B(X) a quasinilpotent
operator. Then the following are equivalent:
4
A. TCACIUC
(i) T has an invariant subspace.
(ii) There exists a rank one operator F such that for any α ∈ C, T + αF is
quasinilpotent.
(iii) There exists a rank one operator F and α ∈ C, α 6= 0, α 6= 1, such that T + F
and T + αF are quasinilpotent.
Proof. Note first that since σess(T ) = {0}, and the essential spectrum is stable under
compact perturbations, it follows that for any α ∈ C and any rank-one operator
F , σess(T + αF ) = {0}. Therefore σ(T + αF ) is at most countable with 0 the only
accumulation point, and any λ ∈ σ(T +αF )\{0} is an eigenvalue((see e.g. [1], Corollary
7.49 and 7.50). Hence, the condition that T + αF is quasinilpotent, is equivalent to
σp(T + αF ) \ {0} = ∅.
i) ⇒ ii)
Suppose Y is a non-trivial invariant subspace for T . Pick f ∈ Y , and e∗ ∈ X ∗ such
that e∗(Y ) = 0. Let F be the rank one operator defined by F := e∗ ⊗ f . Then, since
Y is T -invariant and f ∈ Y , we have that the orbit (T nf ) is contained in Y , hence for
all n ∈ N, e∗(T nf ) = 0. It follows that, for any z ∈ C \ {0} we have:
(1)
e∗(R(z)f ) = e∗ ∞
Xi=0
T if
zi+1! =
∞
Xi=0
e∗(T if )
zi+1 = 0.
Fix α 6= 0, arbitrary. From (1) and Lemma 2.1 it now follows that for any z ∈ C\{0}
we have that z /∈ σp(T +αF ). Therefore, for any α 6= 0 we have that σp(T +αF )\{0} =
∅, hence T + αF is quasinilpotent.
ii) ⇒ iii) obvious
iii) ⇒ i) We argue by contradiction. Assume that T has no invariant subspaces, and
fix F := e∗ ⊗ f an arbitrary rank one operator. Since T has no invariant subspaces it
follows that e∗(T nf ) 6= 0 for infinitely many values of n. Indeed, otherwise there exist
k ∈ N such that e∗(T jf ) = 0 for all j ≥ k. However this means that the closed span
of (T jf )j≥k is contained in the kernel of e∗, thus it would be a non-trivial T -invariant
subspace, contradicting the assumption.
To simplify the notation, we denote by g : C \ {0} → C the analytic function defined
by g(z) = e∗(R(z)f ). Note that g has an isolated singularity at z = 0 and its Laurent
series about z = 0 is
g(z) =
1
zi+1 e∗(T if ).
∞
Xi=0
ON QUASINILPOTENT OPERATORS AND THE INVARIANT SUBSPACE PROBLEM
5
Since e∗(T nf ) 6= 0 for infinitely many values of n, it follows that the Laurent expan-
sion of g will have infinitely many non-zero terms of the form 1
zi+1 e∗(T if ). Therefore
z = 0 is an isolated essential singularity for g. From Picard's Great Theorem it fol-
lows that g attains any value, with possibly one exception, infinitely often, in any
neighbourhood of z = 0. Hence, for all α 6= 0, with possibly one exception, the set
{z ∈ C : g(z) = α−1} is infinite. Note that this set is in fact countably infinite, as it
is easy to see that in Picard's Theorem the values can be attained at most countably
many times. Therefore σp(T + αF ) \ {0} = {z ∈ C : g(z) = α−1} is countably infinite
for all α, with possibly one exception, so T + αF is quasinilpotent for at most one
non-zero value α. Since F was arbitrary, this contradicts iii), and the implication is
proved.
(cid:3)
The techniques employed in the proof of the previous theorem also gives the following
characterization of the spectrum of rank-one perturbation of quasinilpotent operators.
Proposition 2.4. Let X be a Banach space, T ∈ B(X) a quasinilpotent operator, and
F := e∗ ⊗ f a rank one operator. Then exactly one of the following three possibilities
holds:
(i) For all α ∈ C, T + αF is quasinilpotent.
(ii) For all non-zero α ∈ C, with possibly one exception, σp(T + αF ) is countably
infinite.
(iii) There exists K ∈ N such that for all non-zero α ∈ C, 0 < σp(T + αF ) \ {0} <
K.
Proof. From the proof of Theorem 2.3 the options (i) and (ii) hold when e∗(T nf ) = 0
for all n, and when e∗(T nf ) 6= 0 for infinitely many values of n, respectively. It remains
to examine the case when e∗(T nf ) 6= 0 for finitely, non-zero, values of n. Let k > 0
be the smallest natural number such that e∗(T kf ) 6= 0 and e∗(T jf ) = 0 for all j > k.
With the notations from Theorem 2.3 it follows that:
g(z) = e∗(R(z)f ) =
1
zi+1 e∗(T if ).
k
Xi=0
Therefore z = 0 is a pole of order k + 1 for g. In this case it is easy to see that for any
α 6= 0, the equation g(z) = α−1 has at most k + 1 solutions, hence the cardinality of
the non-empty set σp(T + αF ) \ {0} is at most k + 1, and (iii) holds.
(cid:3)
6
A. TCACIUC
We will show in the next example that the second option in Proposition 2.4 can
indeed hold, and that the one exception is in general unavoidable.
Example. Let H be a separable Hilbert space, denote by (en)n an orthonormal basis,
and define T ∈ B(H) to be the weighted shift defined by
T en =
1
n
en+1,
for n = 1, 2, . . .
operator F ∈ B(H) defined by F (x) := hx, f ie1, where f = P∞
It is easy to see that T is a compact quasinilpotent operator. Consider the rank one
n en. We are going
to show that T − F is quasinilpotent, and that for any α 6= −1 we have σp(T + αF ) is
countably infinite. From the previous considerations this is equivalent to showing that
the function g(z) := hR(z)e1, f i is analytic on C \ {0}, has an essential singularity at
0, and g(z) 6= −1 for all z ∈ C \ {0}. We have
n=1
1
R(z)e1 =
1
zn+1 T ne1 =
∞
Xn=0
1
zn+1
1
n!
ei+1.
∞
Xi=0
Therefore
g(z) := hR(z)e1, f i = h
1
zn = exp(1/z) − 1.
Clearly g has an essential singularity at z = 0 and g(z) 6= −1 for any z ∈ C \ {0}.
Xn=0
Xn=1
eni =
en+1,
1
n!
1
n
∞
Xn=1
∞
1
zn+1
1
n!
∞
3. Invariant half-spaces for rank one perturbations
We next turn our attention to the study of invariant half-spaces of rank-one perturba-
tions of quasinilpotent operators. First recall some standard notations and definitions.
A sequence (xn)∞
n=1 in X is called a basic sequence if any x ∈ [xn] can be written
n=1 anxn, where the convergence is in norm (see [17, section 1.a]
for background on Schauder bases and basic sequences). As [x2n] ∩ [x2n+1] = {0} it
is immediate that [x2n] is of both infinite dimension and infinite codimension in [xn],
thus a half-space, and since every Banach space contains a basic sequence, it follows
uniquely as x = P∞
that every infinite dimensional Banach space contains a half-space.
An important tool that we are going to use is the following criterion of Kadets
and Pe lczy´nski for a subset of Banach space to contain a basic sequence (see, e.g., [2,
Theorem 1.5.6])
Theorem 3.1 (Kadets, Pe lczy´nski). Let S be a bounded subset of a Banach space X
such that 0 does not belong to the norm closure of S. Then the following are equivalent:
(i) S fails to contain a basic sequence,
ON QUASINILPOTENT OPERATORS AND THE INVARIANT SUBSPACE PROBLEM
7
(ii) The weak closure of S is weakly compact and fails to contain 0.
As mentioned in the introduction, Tcaciuc and Wallis proved in [29] the following
theorem:
Proposition 3.2. [29, Proposition 2.11] Let X be an infinite-dimensional complex
Banach space and T ∈ B(X) a bounded operator such that σ(T ) is countable and
σp(T ) = ∅. Then for any nonzero x ∈ X and any ε > 0 there exists F ∈ B(X) with
kF k < ε and Range(F ) = [x], and such that T + F admits an invariant half-space.
When X is reflexive, a companion result, [29, Proposition 2.12] allows one to (sep-
arately) control the kernel of the perturbation. Our main result in this section shows
that for quasinilpotent operators we can control both the range and the kernel at the
same time, in a very strong way: with at most two exceptions, all perturbations by
scalar multiples of a fixed rank-one operator have invariant half-spaces. Also note that
this results holds in general Banach spaces. No reflexivity condition is needed.
Theorem 3.3. Let X be a separable Banach space and T ∈ B(X) a quasinilpotent
operator such that σp(T ) = σp(T ∗) = ∅. Then for any rank one operator F , and any
non-zero α ∈ C, with possibly two exceptions, T + αF has an invariant half-space.
Proof. It is easy to check that σp(T ) = ∅ if and only if T has no non-trivial finite
dimensional invariant subspaces, and that σp(T ∗) = ∅ if and only if T has no non-
trivial finite codimensional invariant subspaces. Therefore we can conclude from the
hypotheses that any non-trivial invariant subspace of T must be a half-space.
Let F = e∗ ⊗ f be a rank-one operator, and consider the orbit (T nf ). If there exists
k ∈ N such that e∗(T nf ) = 0 for all n > k, then Y := [T nf ]n>k is an invariant subspace
for T , contained in the kernel of F . Therefore Y is a T -invariant half-space, and it is
also invariant for T + αF , for any α ∈ C.
There remains to consider the situation when e∗(T nf ) 6= 0 for infinitely many values
of n. In this case it follows from the proof of Theorem 2.3 that for all non-zero α ∈ C,
with possibly one exception, σp(T + αF ) is countably infinite. Moreover, 0 is the only
accumulation point for σp(T + αF ). Denote by C0 the set of all these values α; in other
words, C0 does not contain 0, and at most one more other value, depending on F .
For any α ∈ C0, define the set Sα as
Sα := {R(z)f : z ∈ σp(T + αF ) \ {0}} ⊆ X.
Note from Remark 2.2 that Sα is a set of (linearly independent) eigenvectors corre-
sponding to all distinct eigenvalues from σp(T +αF )\{0}. For any z ∈ σp(T +αF )\{0}
8
A. TCACIUC
we have that e∗(R(z)f ) = α−1, hence kR(z)f k ≥ (αke∗k)−1. That is, for any α ∈ C0,
Sα is bounded away from zero, therefore 0 does not belong to the norm closure of Sα.
Define the following sets:
A := {α ∈ C0 : Sα is not bounded}
B := {α ∈ C0 : Sα is bounded and Sα
C := {α ∈ C0 : Sα is bounded and Sα
w
w
is not w-compact}
is w-compact}.
Clearly C0 = A ∪ B ∪ C, and the union is disjoint. We are going to show that for any
α ∈ A ∪ B, T + αF has an invariant half-space, and that C ≤ 1.
Let first α ∈ A. Denote σp(T + αF ) = (λn)n, and note that we have that λn → 0.
We are going to show that Sα contains a basic sequence. Since Sα is not bounded, by
passing to a subsequence we may assume that kR(λn)f k → ∞. For any n ∈ N, denote
by xn := R(λn)f /kR(λn)f k, and put Wα := {xn : n ∈ N}.
w
If Wα
The set Wα is bounded.
is not weakly-compact, we can apply Kadets-
Pe lczy´nski criterion (Theorem 3.1) to conclude that Wα contains a basic sequence.
Therefore, by passing to a subsequence, we can assume that (xn) is a basic sequence
in X. Then Y := [x2n] is a half-space which is invariant for T + αF .
w
If Wα
is weakly compact, then it is weakly sequentially compact by the Eberlein-
w−→ x ∈ X.
Smulian theorem, and by passing to a subsequence we can assume that xn
It is easy to see that
(2)
1
T xn = λnxn −
kR(λn)f k
−→ 0, and kR(λn)f k w∗
w∗
f.
w∗
−→ T x, λnxn
Since T xn
−→ ∞, it follows from (2) that
T x = 0. However 0 is not an eigenvalue for T , so we must have x = 0. Hence
0 ∈ Wα
, and again by the Kadets-Pe lczy´nski criterion we have that Wα contains a
basic sequence, and we finish up as in the case when Wα
is not weakly-compact.
w
w
When α ∈ B, therefore Sα is bounded and Sα
is not weakly compact, we can again
apply the Kadets-Pe lczy´nski criterion to conclude that Sα contains a basic sequence,
and again finish up as before. Therefore, we have shown that for α ∈ A ∪ B, T + αF
has an invariant half-space. There remains to show that C ≤ 1.
w
Assume towards a contradiction that there exist α 6= β in C. Denote by (λn) and
by (µn) the eigenvalues in σp(T + αF ) \ {0} and σp(T + βF ) \ {0}, respectively, and
note that both (λn) and (µn) converge to 0. For each n ∈ N, we define hn := R(λn)f ,
and by kn := R(µn)f . We have
(3)
T hn = λnhn − f
and
T kn = µnkn − f.
ON QUASINILPOTENT OPERATORS AND THE INVARIANT SUBSPACE PROBLEM
9
Since Sα
w
w
and Sβ
that hn
w−→ h and kn
are weakly compact, we can assume, by passing to subsequences,
w−→ k. Note that for any n ∈ N we have that
e∗(hn) = g(λn) = α−1
and
e∗(hn) = g(µn) = β −1.
Therefore, e∗(h) = α−1 and e∗(k) = β −1, and since α 6= β, it follows that h 6= k.
Taking weak limits in (3), and taking into account that λn → 0 and µn → 0, we get
that T h = −f and T k = −f . Therefore T (h − k) = 0, and since h − k 6= 0 it follows
that 0 is an eigenvalue for T , which is a contradiction since σp(T ) = ∅. It follows that
C ≤ 1, and this completes the proof.
(cid:3)
While not explicitly stated in the previous theorem, note that, in particular, we
can obtain rank-one perturbations of arbitrarily small norms that have invariant half-
spaces.
Indeed, since -- for a fixed rank-one F -- almost all perturbations T + αF
have invariant half-spaces, for any given ε > 0 we can take a "good" α < ε/kF k. We
summarize this in the following corollary:
Corollary 3.4. Let X be a separable Banach space and T ∈ B(X) a quasinilpotent
operator such that σp(T ) = σp(T ∗) = ∅. Then for any non-zero f ∈ X, e∗ ∈ X ∗,
and ε > 0, we can find rank-one F ∈ B(X) with Range(F ) = [f ], ker F = ker e∗, and
kF k < ε such that T + F has an invariant half-space.
In the Hilbert space setting we get more specific information about the structure of
a quasinilpotent operator.
Corollary 3.5. Let H be a separable Hilbert space and T ∈ B(H) a quasinilpotent
operator such that σp(T ) = σp(T ∗) = ∅. Then for any rank one operator F , and any
non-zero α ∈ C, with possibly two exceptions, there exists an orthogonal projection of
infinite rank and co-rank such that P ⊥T P = αP ⊥F P .
Proof. Fix a rank one operator F ∈ B(H). From Theorem 3.3 we have that for all
non-zero α ∈ C, with possibly two exceptions, T − αF has an invariant half-space. Fix
such an α ∈ C, let Y be an invariant half-space for T − αF , and let P ∈ B(H) be the
orthogonal projection onto Y (which clearly has infinite rank and co-rank). Since Y
is invariant for T − αF it is easy to see that P ⊥(T − αF )P = 0, and the conclusion
follows.
(cid:3)
4. Open questions
In light of Theorem 2.3, the behaviour of the spectrum of a quasinilpotent operator
under rank one perturbations is related to a solution for the invariant Subspace Problem
10
A. TCACIUC
for quasinilpotent operators. This suggests a natural and very important open question
in this direction.
Question 1: Given T ∈ B(X) a quasinilpotent operator acting on an infinite
dimensional, separable, complex Banach space, can we find a rank-one operator F
such that T + F is quasinilpotent?
Note that a positive answer to this question is not sufficient to conclude a positive
solution to the Invariant Subspace Problem for quasinilpotent operators by applying
Theorem 2.3. Indeed, from iii) in Theorem 2.3 we would need one more perturbation
by a scalar multiple of the same rank one operator that is also quasinilpotent. On
the other hand, a negative answer to the Question 1 will provide a counterexample to
ISP for quasinilpotent operators. As we mentioned before, Read [26] already provided
such a counterexample on l1, therefore it would be a natural starting point to examine
Question 1 for Read's operator.
The requirement that F has rank one is very important, as we can always find F
of rank two such that T + F is quasinilpotent. Indeed, let N be a rank one nilpotent
operator (thus N 2 = 0) and put S := (I − N)T (I + N). Then S is similar to T ,
therefore S is also quasinilpotent, and an easy calculation shows that T − S has rank
at most two.
In Section 2 we constructed an example of a quasinilpotent operator T ∈ B(l2), and a
rank one operator F such that T +F is quasinilpotent, but T +αF is not, for all α 6= 0,
α 6= 1. Therefore F does not satisfy iii) in Theorem 2.3, but of course there are other
rank one operators that do, as T in that example has plenty of invariant subspaces.
Whether a positive solution to Question 1 already implies a positive solution to ISP
for quasinilpotent operators is another important open question:
Question 2: If T ∈ B(X) is a quasinilpotent operator with the property that there
exists F ∈ B(X) rank one such that T + F is quasinilpotent, does T have an invariant
subspace?
Note that if Question 1 has a positive solution for Read's operator, then Read's
operator provides a negative answer to Question 2.
Acknowledgments. It is my pleasure to acknowledge the helpful conversations and
feedback provided to me by Heydar Radjavi. This research was supported in part by
the Natural Sciences and Engineering Research Council of Canada.
ON QUASINILPOTENT OPERATORS AND THE INVARIANT SUBSPACE PROBLEM
11
References
[1] Y.A. Abramovich and C.D. Aliprantis. An Invitation to Operator Theory (2000), ISBN 0-8219-
2146-6.
[2] F. Albiac, N. Kalton, Topics in Banach space theory Graduate Texts in Mathematics, 233.
Springer, New York, 2006.
[3] George Androulakis, Alexey I Popov, Adi Tcaciuc and Vladimir G. Troitsky. Almost Invariant
Half-spaces of Operators on Banach Spaces Integr.Equ.Oper.Theory 65 (2009), 473 -- 484.
[4] C. Apostol, D.V. Voiculescu, On a problem of Halmos Rev. Roumaine Math. Pures Appl. 19
(1974) 283 -- 284
[5] N. Aronszajn, K.T. Smith, Invariant subspaces of completely continuous operators Ann. of Math
(2), 60 (1954) 345-350
[6] B. Beauzamy, Un oprateur sans sous-espace invariant: simplification de l'exemple de P. Enflo
Integr.Equ.Oper.Theory 8 (1985), no.3, 314 -- 384.
[7] I. Chalendar, J.R. Partington, Modern approaches to the invariant-subspace problem Cambridge
Tracts in Mathematics, 188. Cambridge University Press, Cambridge, 2011.
[8] P. Enflo, On the invariant subspace problem in Banach spaces Seminaire Maurey-Schwartz (1975-
1976) Espaces Lp, applications radonifiantes et g`eom`etrie des espaces de Banach, Exp. Nos. 14-15,
7 pp. Centre Math., ´Ecole Polytech., Palaiseau, 1976.
[9] P. Enflo, On the invariant subspace problem for Banach spaces Acta Math., 158 (34) (1987)
213 -- 313.
[10] C. Foia¸s, C. Pearcy, A model for quasinilpotent operators Michigan Math J. 21 (1974) 399 -- 404
[11] C. Foia¸s, I. Jung, E. Ko, C. Pearcy, On quasinilpotent operators II J. Austral. Math. Soc.
77(2004), 349 -- 356
[12] C. Foia¸s, I. Jung, E. Ko, C. Pearcy, On quasinilpotent operators III J. Operator Theory
54:2(2005), 401 -- 414
[13] D. Herrero, Almost every quasinilpotent Hilbert space operator is a universal quasinilpotent Proc.
Amer. Math. Soc. 13 (1978) 212 -- 216
[14] I. Jung, E. Ko, C. Pearcy, On quasinilpotent operators Proc. Amer. Math. Soc. 131(2003), 2121 --
2127
[15] I. Jung, E. Ko, C. Pearcy, Almost invariant half-spaces for operators on Hilbert space Bull. Aust.
Math. Soc. doi:10.1017/S0004972717000533
[16] I. Jung, E. Ko, C. Pearcy, Almost invariant half-spaces for operators on Hilbert space II: operator
matrices arXiv:1710.09937 [math.FA]
[17] J. Lindenstrauss, L. Tzafriri, Classical Banach spaces. I. Sequence spaces. Ergebnisse der Math-
ematik und ihrer Grenzgebiete, Vol. 92. Springer-Verlag, Berlin-New York, 1977.
[18] V.I. Lomonosov, Invariant subspaces of the family of operators that commute with a completly
continuous operator. Functional. Anal. i Prilozhen, 7 (1973), no.3, 55-56
[19] Laurent W. Marcoux, Alexey I. Popov and Heydar Radjavi, On Almost-invariant Subspaces and
Approximate Commutation, J. Funct. Anal. 264(4) (2013), 10881111.
[20] Alexey I. Popov Almost invariant half-spaces of algebras of operators Integr.Equ.Oper.Theory
67(2) (2010), 247-256
[21] Alexey I. Popov and Adi Tcaciuc. Every Operator has Almost-invariant Subspaces, J. Funct.
Anal. 265(2) (2013), 257265.
[22] H. Radjavi, P. Rosenthal, Invariant subspaces, Second edition. Dover Publications, Inc., Mineola,
NY, 2003.
[23] Charles J. Read A solution to the invariant subspace problem, Bull. Lond. Math. Soc. 16(4)
(1984), 337-401.
[24] Charles J. Read A solution to the invariant subspace problem on the space l1, Bull. Lond. Math.
Soc. 17(4) (1985), 305-317
[25] Charles J. Read Strictly singular operators and the invariant subspace problem, Studia Math.
132(3) (1991), 203-226
12
A. TCACIUC
[26] Charles J. Read Quasinilpotent operators and the invariant subspace problem, J. London Math.
Soc. 56(2) (1997), 595-606
[27] Gleb Sirotkin and Ben Wallis. Almost-invariant and essentially-invariant halfspaces Linear Al-
gebra Appl. 507 (2016), 399-413
[28] Adi Tcaciuc The almost-invariant subspace problem for Banach spaces arXiv:1707.07836
[math.FA]
[29] Adi Tcaciuc, Ben Wallis. Controlling almost-invariant halfspaces in both real and complex setting
Integr.Equ.Oper.Theory 87(1) (2017), 117-137.
(A. Tcaciuc) Mathematics and Statistics Department, Grant MacEwan University,
Edmonton, Alberta, Canada T5J P2P, Canada
E-mail address: [email protected]
|
1409.2212 | 24 | 1409 | 2017-07-07T19:37:19 | Product Theorem for Singular Integrals of Nonconvolution Type | [
"math.FA"
] | A new class of symbols is investigated. These symbols satisfy a differential inequality which has a mixture of homogeneities on the product space. We show that pseudo differential operators with symbols of order zero in the class are L^p -bounded for 1 < p < infinity. Moreover, they form an algebra under compositions. | math.FA | math |
An Algebra of Singular Integrals
of Non-convolution Type
Zipeng Wang
Cambridge Centre for Analysis, University of Cambridge,
Cambridge, CB3 0WA, United Kingdom.
e-mail: [email protected]
December 9, 2014
Abstract
We find a class of pseudo differential operators classified by their symbols. These
symbols satisfy a differential inequality with a mixture of homogeneities in the frequency
space. When taking the singular integral realization, the class of operators can be equiv-
alently defined by their kernels. We prove that the pseudo differential operators in this
class are Lp continuous, admitted in a weighted norm inequality for 1 < p < ∞. Moreover,
the class itself forms an algebra.
1 Introduction
In this paper, we prove an Lp theorem for a class of pseudo differential operators. By taking
singular integral realization, they are singular integrals of non-convolution type. The class
hereby investigated, classified either by the symbols, or equivalently by the kernels, forms
an algebra under the composition of operators. We show that the operators consisted are
Lp continuous, for 1 < p < ∞. With clarity on some more definitions, these operators are
admitted in a weighted norm inequality.
Our class of operators first arises when studying some sub-elliptic problems. These include
the Grushin type operators, oblique derivative problems and Cauchy-Szeg o projections. We
give some examples in the end.
Let f ∈ S be a Schwartz function. A pseudo differential operator Tσ is defined by
(cid:16)Tσ f(cid:17)(x) = Z σ(x, ξ)e2πix·ξbf (ξ)dξ
where σ denotes the symbol.
Let N = N1 + N2 + ··· + Nn be the dimension of a product space. We write
x = (x1, x2,
. . . , xn) ∈ RN1 × RN2 × ··· × RNn
(1. 1)
1
with Ni the homogeneous dimension of the i-th subspace, for i = 1, 2, . . . , n.
Definition of Sρ : Let 0 < ρ < 1. A symbol σ belongs to the class Sρ if σ(x, ξ) ∈ C∞(cid:16)RN × RN(cid:17)
and satisfies the differential inequality
∂α
∂ξα
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂β
∂xβ σ(x, ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,β
nYi=1
1
1 + ξi + ξρ!αi
(1 + ξ)ρβ
(1. 2)
for every multi-index α = (α1, α2, . . . , αn) and β.
Observe that Sρ forms a sub class of the exotic class S0
introduction. Pseudo differential operators with their symbols consisted in S0
on L2-space.
ρ,ρ. See chapter VII of [7] for an
ρ,ρ are bounded
In the other hand, σ in (1. 2) satisfies a variant of Marcinkiewicz theorem. Namely,
If Tσ is translation invariant, i.e: σ = σ(ξ), then it is Lp continuous for
uniformly in x.
1 < p < ∞. See [4]-[6] for references.
By taking singular integral realization, we write Tσ f (x) in the form of
∂
∂ξ!α
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ξ
σ(x, ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα
Z f (x − y)Ω(x, y)dy.
(1. 3)
(1. 4)
(1. 5)
Its kernel Ω, is a distribution Ω(x,·) that coincides with a smooth function
e2πiy·ξσ(x, ξ)dξ
Ω(x, y) = ZRN
Ω(x, y)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,β
for y , 0. We will show that σ ∈ Sρ if and only if Ω: has a size estimate of
ρ!Ni+αi+ρβ
∂α
∂yα
∂β
∂xβ
1
1
nYi=1
yi + y
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for y , 0, and decays rapidly as y −→ ∞.
Moreover, recall an normalized bump function ϕ is smooth, supported on the unit ball,
with all its derivatives uniformly bounded by 1 upon to a sufficiently large order. For
I ∪ J = {1, 2, . . . , n}. Let ϕi be an normalized bump function for every i ∈ I and y† is the
projection of y on the subspaceL RNj for j ∈ J. Ω carries the cancellation property
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Yj∈J
∂α j
α j
∂y
j
∂xβ
!Z ···ZRNi ∂β
≤ Aα,βYj∈J
1
Ω(x, y)! ϕi(Riyi)dyi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ρ!Nj +α j+ρβ 1 +Xi∈I
1
yj + y†
for y† , 0, and decays rapidly as y† −→ ∞.
2
(1. 6)
Ri!ρβ
Lastly, we introduce the following class of weight functions defined on the product space.
Definition of Ap : Let λ be an non-negative, locally integrable function on RN. Then, λ ∈ Ap
for 1 < p < ∞ if
1
QZQ
λ(x)dx · 1
λ(x)! q
QZQ 1
p
q
dx! p
≤ A
(1. 7)
with 1/p + 1/q = 1, for all Q = B1 × B2 × ··· × Bn. Each Bi is a standard ball in RNi, for
i = 1, 2, . . . , n. The smallest constant A, denoted by Ap(λ), is taking to be the Ap norm of λ.
Given dµ to be an nonnegative Borel measure. Our main result is the following:
Theorem 1 Suppose σ ∈ Sρ. Then the operator Tσ, initially defined on S, extends to a bounded
operator on Lp(cid:16)RN, dµ(cid:17) such that
ZRN(cid:12)(cid:12)(cid:12)Tσ f (x)(cid:12)(cid:12)(cid:12)p
dµ(x) ≤ Ap,ρZRN(cid:12)(cid:12)(cid:12) f (x)(cid:12)(cid:12)(cid:12)p
dµ(x)
if and only if dµ(x) = λ(x)dx is absolutely continuous, with λ ∈ Ap for 1 < p < ∞. Moreover, Sρ
forms an algebra under the composition of operators.
We develop our Lp estimate in the framework of Littlewood-Paley projections. Such method-
ology in general has been applied to several earlier works, for example in [3] and [4] for
Fourier multiplier operators and translation invariant singular integrals; in [8] for pseudo
differential operators. The key investigation here would be the combinatorial estimate on the
Dyadic decomposition of the frequency space, where the symbol in Sρ satisfies a desired size
estimate, in the sense of differential inequality (1. 2). It appears that such estimate is crucial,
as it is required in each major step of our analysis. In particular, we show that every partial
sum operator in the framework is bounded by the strong maximal function, which satisfies
the weighted Lp-norm inequality with respect to the measure λ(x)dx for λ ∈ Ap.
The paper is organized as follows: We derive a lemma in section 2 which plays a principal role
in later estimation. As a direct consequence, Sρ forms an algebra under the composition of
operators. In section 3, we give the combinatorial estimate. In section 4, we show that every
partial sum operator of Littlewood-Paley projections is bounded by the strong maximal
function. The Lp estimate is developed in section 5. We show that Tσ admits the desired
weighted inequality in section 6, based on some earlier results proved in [3]. In section 7,
we estimate the kernel of Tσ with σ ∈ Sρ. By carrying out a mollified estimation in analogue
to the first section of [1], we show that our class of operators can be equivalently defined by
their kernels. The last section is devoted for applications.
Abbreviations:
⋄: Except otherwise indicated, we writeZ =ZRN
⋄: We write Lp as Lp(RN, dx), and Lp(λ) as Lp(RN, λ(x)dx) with λ ∈ Ap, for 1 < p < ∞.
⋄: We always write A as a positive, generic constant with a subindex indicating its dependence.
and " = "RN×RN
.
3
2 Principal Lemma
The main objective of this section is to prove a lemma which is principal for later estimation.
The original framework of the relevant investigation was developed by Beals and C. Fefferman
[10] in 1974, also by Boutet de Monvel [9] in the same year. It has later been refined by Nagel
and Stein in [2] where our estimate follows.
Let ϕ = ϕ(x, y, ξ, η) to be any smooth function. Its norm is bounded uniformly by a norm
function ϑ = ϑ(ξ, η) which has polynomial growth in ξ and η. Moreover, ϕ satisfies
differential inequality
∂α
∂ξα
∂β
∂ηβ
∂µ
∂xµ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂ν
∂yν ϕ(x, y, ξ, η)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
nYi=1 (cid:18)
≤ Aα,β,µ,ν ϑ(ξ, η)
1
1 + ξi + ξρ(cid:19)αi nYj=1
1
1 + ηi + ηρ!βj
We consider the integral of the type
" e2πi(x−y)·(ξ−η)ϕ(x, y, ξ, η)dydη.
(1 + ξ + η)ρ(µ+ν).
(2. 1)
(2. 2)
Let φ ∈ C∞o (Rn) withφ ≤ 1, is constantly 1 in the unit ball and zero outside a compact support.
Define φε(y) = φ(εy) and ϕε = ϕφε. We have ϕε satisfies inequality (2. 1) uniformly in
0 < ε ≤ 1.
Define the differential operator
and respectively
4π2(cid:19)
D = I −(cid:18) 1
Q = 1 +
1
1
1 + ξ2
+
1
1 + η2!ρ
1 + η2!ρ
1
+
1 + ξ2
∆y −(cid:18) 1
4π2(cid:19)(cid:16)1 + ξ2 + η2(cid:17)ρ
∆η
ξ − η2 +(cid:16)1 + ξ2 + η2(cid:17)ρ
x − y2.
(2. 3)
(2. 4)
Denote L = Q−1D. We have the identity
(2. 5)
for every N ≥ 1. By replacing ϕ with ϕε inside the integral (2. 2), integrating by parts with
respect to y and η gives
LN(cid:16)e2πi(x−y)·(ξ−η)(cid:17) = e2πi(x−y)·(ξ−η)
" (cid:16)tL(cid:17)N(cid:16)ϕε(x, y, ξ, η)(cid:17)e2πi(x−y)·(ξ−η)dydη
(2. 6)
where tL = tDQ−1. One major estimate is to show that
Lemma 2.1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:16)tL(cid:17)N
ϕε(x, y, ξ, η)(cid:12)(cid:12)(cid:12)(cid:12) ≤ AN 1
uniformly in 0 < ε ≤ 1 and for every N ≥ 1.
Q!N 1 + ξ2
1 + η2
+ η2
1 + ξ2!ρN
ϑ(ξ, η)
(2. 7)
4
.
(2. 8)
Proof : Let
We have
b =(cid:16)1 + ξ2 + η2(cid:17)ρ
Q!N
ϕε(x, y, ξ, η).
b
a =
+
1
1
and
1 + η2!ρ
1 + ξ2
ϕε(x, y, ξ, η) = 1
(cid:16)tL(cid:17)N
a(cid:12)(cid:12)(cid:12)(cid:12)∆yϕε(x, y, ξ, η)(cid:12)(cid:12)(cid:12)(cid:12) . a(cid:16)1 + ξ2 + η2(cid:17)ρ
a
Q − ∆η
Q − ∆y
If all differentiations fall on ϕ, we have
(2. 9)
(2. 10)
(2. 11)
(2. 12)
(2. 13)
ϑ(ξ, η)
. 1 + ξ2
1 + η2
+ η2
1 + ξ2!ρ
ϑ(ξ, η)
b
(cid:16)1 + η2(cid:17)ρ ϑ(ξ, η)
. 1 + ξ2
+ η2
1 + η2
1 + ξ2!ρ
Q 1 + ξ2
1 + η2
1 + ξ2!ρ
+ η2
1
1 + η!α 1
Q 1 + ξ2
1 + η2
1
ϑ(ξ, η).
+ η2
1 + ξ2!ρ
Turn to the general settings. From (2. 4), direct computation shows
and
b(cid:12)(cid:12)(cid:12)(cid:12)∆ηϕε(x, y, ξ, η)(cid:12)(cid:12)(cid:12)(cid:12) .
∂α
∂ηα
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
and
Q ≤
1
1
1
∂β
∂yβ
Q(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) .
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Q!No ∂α1
b
1 + η!α 1
Q(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Q!N1 ∂α2
∂ηα1
1
Q ≤
∂ηα2
.
b
for every multi-index α and β.
From (2. 8)-(2. 9), it is suffice to estimate the term
b
Q!N2
··· ∂αk
∂ηαk
b
Q!Nk
∂β
ηϕε(x, y, ξ, η)
(2. 14)
where
for every 1 ≤ k ≤ N.
From (2. 8), we have
No + N1 + N2 + ··· + Nk
α1N1 + α2N2 + ··· + αkNk
= N,
+ β = 2N
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂αb
∂ηα(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα
1
1 + η!α
b.
5
(2. 15)
(2. 16)
The estimates in (2. 12)-(2. 13), together with (2. 16) imply that
∂α
∂ηα
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1 + η!γ 1
Q
b
1
1
b
Q(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα Xβ+γ=α
≤ Aα
1 + η!α b
1 + η!β
Q 1 + ξ2
1 + η2
1
+ η2
1 + ξ2!ρ
.
(2. 17)
By bringing in the estimate (2. 17) into the combinatorial settings in (2. 15), we obtain
differential inequality (2. 7).
(cid:3)
Lemma 2.2 Suppose ϕ = ϕ(x, y, ξ, η) is smooth, absolutely bounded and satisfies the differential
inequality (2. 1). Let
Φ(x, ξ) = " e2πi(x−y)·(ξ−η)ϕ(x, y, ξ, η)dydη.
Then, Φ(x, ξ) is absolutely bounded and satisfies the differential inequality
(2. 18)
(2. 19)
∂α
∂ξα
∂β
∂xβ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Φ(x, ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,β,ρ
nYi=1 (cid:18)
1
1 + ξi + ξρ(cid:19)αi
(1 + ξ)ρβ.
Proof: 1. We momentarily assume that ϕ has compact support in y by replacing ϕ with ϕε as
in the previous lemma. The restriction will be removed by taking approximation in the sense
of distributions.
Integration by parts with respect to y in (2. 18) implies
∂α
∂ξα
∂β
∂xβ
Φ(x, ξ)
(2. 20)
consists of the terms
dydη
(2. 21)
" e2πi(x−y)·(ξ−η)
i
j
i
∂β1
β1
j
∂α1
α1
i
i
nYi=1
nYj=1
j + β2
j = β j for every i and j.
∂α2
α2
i
∂η
i
∂ξ
∂x
j
j
∂β2
β2
j
∂y
j
ϕε(x, y, ξ, η)
i + α2
i = αi and β1
with α1
Since ϕ is absolutely bounded, from differential inequality (2. 1), the integrant inside (2. 21)
has norm bounded by
Aα,β
nYi=1 (cid:18)
1
i
1 + ξi + ξρ(cid:19)α1
1
1 + ηi + ηρ!α2
i nYj=1
(1 + ξ)ρβ1
j(1 + η)ρβ2
j
(2. 22)
for every multi-index α = (α1, α2, . . . , αn) and β = (β1, β2, . . . , βn).
2. When ξ − η ≤ 1
21) has its norm bounded by
2 (1 + ξ + η), indeed we have ξ ∼ η. By lemma 2.1, the integral in (2.
Aα,β,N " Q−Ndydη! nYi=1 (cid:18)
1
1 + ξi + ξρ(cid:19)αi
(1 + ξ)ρβ
(2. 23)
6
for every N ≥ 1.
When ξ − η ≥ 1
21) gains a factor of
2 (1 + ξ + η). An N-fold of integration by parts with respect to y inside (2.
1
1 + ξ + η!(1−ρ)N
.
(2. 24)
It is clear that for N sufficiently large, depending on α, β and ρ, the integral in (2. 21) has its
norm bounded by (2. 23).
3. From (2. 4), we have
Q ≥ 1 +
1
1 + ξ2!ρ
ξ − η2 +(cid:16)1 + ξ2(cid:17)ρ
x − y2.
By changing variables
η −→ (cid:16)1 + ξ2(cid:17)ρ/2
η
and
y −→
1
1 + ξ2!ρ/2
y,
we have
" Q−Ndydη . "
dydη
(cid:16)1 + η2 + y2(cid:17)N
(2. 25)
(2. 26)
(2. 27)
which is absolutely convergent provided that N is sufficiently large. Lastly, ϕε −→ ϕ point-
wisely as ε −→ 0. Taking
Φ(x, ξ) = lim
ε−→0" e2πi(x−y)·(ξ−η)ϕε(x, y, ξ, η)dydη.
Since all estimates hold uniformly in 0 < ε ≤ 1, the lemma is proved.
As a consequence from Lemma 2.2, the symbol class Sρ forms an algebra under the composi-
tion of operators Tσ1 ◦ Tσ2
= Tσ1◦σ2.
(2. 28)
(cid:3)
Theorem 2.1 Suppose σ1 ∈ Sρ and σ2 ∈ Sρ. Then, σ1 ◦ σ2 ∈ Sρ.
Proof: From direct computations, we have
σ1 ◦ σ2(x, ξ) = " e2πi(x−y)·(ξ−η)σ1(x, η)σ2(y, ξ)dydη.
The function σ1(x, η)σ2(y, ξ) is absolutely bounded and satisfies differential inequality (2. 1).
Lemma 2.2 implies that σ1 ◦ σ2 satisfies differential inequality (1. 2).
(cid:3)
3 The Combinatorial Estimate
Let φ ∈ C∞o (RN), such that φ ≡ 1 when ξ ≤ 1 and φ ≡ 0 when ξ > 2. Define
ψ(ξ) = φ(ξ) − φ(2ξ)
7
(3. 1)
In the other hand, we define the n-tuples
t = (cid:16)2−t1, 2−t2,
. . . , 2−tn(cid:17) ,
and the dilations
With respect to the dilations (3. 3)-(3. 5), we let
. . . , 2tn(cid:17)
t−1 = (cid:16)2t1, 2t2,
. . . , 2−tn ξn(cid:17) ,
. . . , 2tn ξn(cid:17) .
ψ (tiξ) .
tξ = (cid:16)2−t1 ξ1, 2−t2 ξ2,
t−1ξ = (cid:16)2t1 ξ1, 2t2 ξ2,
nYi=1
δt(ξ) =
(3. 2)
(3. 3)
(3. 4)
(3. 5)
(3. 6)
(3. 7)
so that ψ is supported on a spherical shell 1
2 ≤ ξ ≤ 2.
For computational purpose, we write q = 1/ρ hence that 1 < q < ∞. Let ti be integers for
every i = 1, 2, . . . , n. We define the n-tuples
ti = (cid:16)2−qti,
. . . , 2−qti, 2−ti, 2−qti,
. . . , 2−qti(cid:17)
where 2−ti is located in the i-th component, and the non-isotropic dilations
tiξ = (cid:16)2−qti ξ1,
. . . , 2−qti ξi−1, 2−ti ξi, 2−qti ξi+1,
. . . , 2−qti ξn(cid:17)
i = 1, 2, . . . , n.
The partial sum operator ∆t is defined by
(cid:16)d∆t f(cid:17)(ξ) = δt (ξ)bf (ξ)
=
nYi=1
ψ(cid:16)2−qti ξ1, . . . , 2−qti ξi−1, 2−ti ξi, 2−qti ξi+1, . . . , 2−qti ξn(cid:17) bf (ξ).
The support of δt(ξ) lies inside the intersection of n elliptical shells with different homo-
geneities of given dilations.
We aim to prove the following lemma:
and
(cid:16)ti1
Lemma 3.1 Let I ∪ J = {1, 2, . . . , n} with 1 ≤ I ≤ n such that
+ 2 + log2 n(cid:17) /q < ti2 < qti1 −(cid:16)2 + log2 n(cid:17)
qt j −(cid:16)2 + log2 n(cid:17) ≤ tı
q − 1(cid:16)2 + 2 log2 n(cid:17)
By assuming that at least one
ti >
1
The support of δt(ξ) is non-empty if and only if
for some ı ∈ I and all
j ∈ J.
for some i ∈ {1, 2, . . . , n}
for all i1, i2 ∈ I,
(3. 8)
(3. 9)
(3. 10)
ξi ∼ 2ti
ξ j . 2tj
for every i ∈ I,
for every j ∈ J
and
ξı ∼ 2tı ∼ 2qtj
for some ı ∈ I.
8
Remark 3.1 A direct outcome from Lemma 3.1 is that we have ti . qt j for every pair of i,
{1, 2, . . . , n}, provided that t satisfies (3. 10) with the support of δt(ξ) is non-empty.
j ∈
Proof: First, by writing down
supp δt(ξ) = supp
nYi=1
ψ(cid:16)2−qti ξ1, . . . , 2−qtiξi−1, 2−ti ξi, 2−qti ξi+1, . . . , 2−qti ξn(cid:17) ,
(3. 11)
the definition of ψ in (3. 1) implies that ξ satisfies
ξi < 2ti +1,
ξi < 2qtj+1,
i = 1, 2, . . . , n,
for j , i
(3. 12)
whenever ξ ∈ suppδt(ξ). Next, we write ξ = (cid:16)ξi, ξ′i(cid:17) ∈ RNi × RN−Ni with ξ′i denotes the
complement of ξi in RN. Our estimation will be obtained in several steps.
1: Suppose that for every i, j ∈ {1, 2, . . . , n},
We shall have
(cid:16)t j + 2 + log2 n(cid:17) /q < ti < qt j −(cid:16)2 + log2 n(cid:17) .
ξi >
2ti−1
√2
or
(cid:12)(cid:12)(cid:12)ξ′i(cid:12)(cid:12)(cid:12) >
2qti−1
√2
for every i = 1, 2, . . . , n. If ξi does not satisfies the estimate in (3. 20) for some i, there at least
exists one ξ j for some j , i such that
(3. 13)
(3. 14)
(3. 15)
(3. 16)
(3. 17)
ξ j >
2qti−1
√2
1
√n − 1
.
In the other hand, from the first inequality in (3. 12), we have
From (3. 15) and (3. 16), the support in (3. 11) is empty provided that
ξ j < 2tj +1.
t j < qti − 2 −
log2 2(n − 1).
1
2
From (3. 13), we have t j < qti−(cid:16)2 + log2 n(cid:17) for every i and j . Inequality (3. 17) is true provided
by (1/2) log2 2(n − 1) < log2 n for every n ≥ 2. We therefore have
ξi ∼ 2ti,
i = 1, 2, . . . , n.
2: Suppose that there exists ı ∈ (1, 2, . . . , n) such that
qt j −(cid:16)2 + log2 n(cid:17) ≤ tı
ξı >
We shall have
or
2tı−1
√2
for all
j , ı.
(cid:12)(cid:12)(cid:12)ξ′ı(cid:12)(cid:12)(cid:12) >
2qtı−1
√2
.
9
(3. 18)
(3. 19)
(3. 20)
Suppose ξı does not satisfies the estimate in (3. 14), then alternatively we must at least have
one ξ j for some l , ı such that
In the other hand, from the first inequality in (3. 12), we have
ξ j >
2qtı−1
√2
1
.
√n − 1
From (3. 21) and (3. 22), the support in (3. 11) is empty provided that
ξ j < 2tj +1.
qtı − t j − 2 −
1
2
log2 2(n − 1) > 0.
Recall that qt j ≤ tı + 2 + log2 n from (3. 13). We have
log2 2(n − 1)
q(cid:17) tı −(cid:16)1 + 1
> (cid:16)q − 1
qtı − t j − 2 −
1
2
q(cid:17)(cid:16)2 + log2 n(cid:17)
(3. 21)
(3. 22)
(3. 23)
(3. 24)
provided that tı satisfies (3. 10) and the fact that (1/2) log2 2(n − 1) < log2 n for every n ≥ 2.
Otherwise, suppose t j satisfies (3. 10) for some j , ı. Then, (3. 13) implies
> 0
tı >
=
q
q − 1(cid:16)2 + log2 n(cid:17) −(cid:16)2 + log2 n(cid:17)
q − 1(cid:16)2 + log2 n(cid:17)
1
(3. 25)
which again implies tı satisfies (3. 10). All together with the second inequality in (3. 12), we
must have
ξı ∼ 2tı ∼ 2qtj
and
ξ j . 2tj
for all j , ı.
3: The lemma will be concluded by combining the estimations in the two previous steps.
Write the support in (3. 11) as
suppYi∈I
ψ(cid:16)2−ti ξi, 2−qti ξ′i(cid:17) ∩ suppYj∈J
ψ(cid:16)2−tj ξ j, 2−qtj ξ′j(cid:17)
(3. 26)
whereas 1 ≤ I ≤ n.
Since there exists at least one i ∈ {1, 2, . . . , n} such that ti satisfies (3. 10). By (3. 9), either i ∈ I
or there exists ı ∈ I such that tı satisfies (3. 10). By carrying out the same estimation in step 1,
we obtain
(3. 27)
In the other hand, suppose ti satisfies (3. 10) for some i ∈ I. For i , ı, we either have ti > tı so
that ti ≥ qt j −(cid:16)2 + log2 n(cid:17) for all j ∈ J, or we have ti < tı so that tı satisfies (3. 10). By carrying
out the same estimation in step 2, we obtain
for every i ∈ I.
ξi ∼ 2ti
ξ j . 2tj
for every j ∈ J and qt j ∼ tı for some ı ∈ I.
(3. 28)
(cid:3)
10
Lemma 3.2 Let t and s satisfy (3. 10). There exists a constant const, such that the intersection of
is non-empty only when
suppδt(ξ) ∩ suppδs(ξ)
nXi=1
ti − si < const.
Proof: We set I1 ∪ J1 = I2 ∪ J2 = {1, 2, . . . , n} for which I1, I2 take the same meaning of I and
J1, J2 take the same meaning of J in Lemma 3.1. Consider ti for i ∈ I1 ∪ J1 and si for i ∈ I2 ∪ J2
respectively. From Lemma 3.1, we have ξi ∼ 2ti for every i ∈ I1 and ξ j . 2tj for every j ∈ J1.
Similarly, ξi ∼ 2si for every i ∈ I2 and ξ j . 2s j for every j ∈ J2.
Therefore, for any i ∈ I1, we need si ≥ ti in order to have the intersection of supports to
be non-empty. Suppose i ∈ I2, we have ξi ∼ 2si so that ti − si must be then bounded by
some constant. Otherwise, suppose i ∈ J2. Lemma 3.1 implies there exists an ı ∈ I2 such that
ξı ∼ 2sı ∼ 2qsi. But again, we need tı ≥ sı = qsi ≥ qti. However, we have t j ≤ qti for every
other j , i by Remark 3.1. Thus that ti − si must be bounded by some constant for every
i ∈ I1. The same argument goes for every i ∈ I2.
Next, from Lemma 3.1, there exists ı ∈ I1 such that
qt j −(cid:16)2 + log2 n(cid:17) ≤ tı
ξı ∼ 2tı ∼ 2qtj
and
for all j ∈ J1. It is necessary that tı ≥ ti for all other i ∈ I1. Therefore, since ti − si is bounded
for every i ∈ I1 ∪ I2, we must have t j − sj bounded for every j ∈ J1 ∩ J2.
(cid:3)
4 Maximal Function Bound
A strong maximal function operator is defined by
M f (x) = sup
Q⊂RN
1
QZQ f (x − y)dy
(4. 1)
where Q = B1 × B2 × ··· × Bn. Each Bi ⊂ RNi here denotes a standard ball centered at xi = 0.
We are going to show that for σ ∈ Sρ, then
whenever t satisfies (3. 10). A direct computation shows
(cid:12)(cid:12)(cid:12)(cid:12)(cid:16)∆tTσ f(cid:17)(x)(cid:12)(cid:12)(cid:12)(cid:12) . M f (x)
(cid:16)∆tTσ f(cid:17)(x) = Z f (x − y)Ωt(x, y)dy
= Z f (x − y)(Z e2πiy·ξδt(ξ)ω(x − y, ξ)dξ) dy
where
ω(x, ξ) = " e2πi(x−y)·(ξ−η)σ(y, η)dydη.
11
(4. 2)
(4. 3)
(4. 4)
Since σ(x, ξ) satisfies the differential inequality (2. 1). Lemma 3.2 implies ω(x, ξ) is absolutely
bounded and satisfies the differential inequality
for every multi-index α.
By changing dilations with respected to ξ = t−1η and y = tz, (4. 3) cab be rewritten as
1
∂α
nYi=1 (cid:18)
1 + ξi + ξρ(cid:19)αi
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂ξα ω(x, ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,q
Z f (x − tz)(Z e2πiz·ηδt(cid:16)t−1η(cid:17) ω(x − tz, t−1η)dη) dz.
Observe that by definition (3. 1),
supp δt(cid:16)t−1η(cid:17) = supp
= supp
nYi=1
nYi=1
ψ(cid:16)t−1tiξ(cid:17)
ψ(cid:16)2t1−qti ξ1, . . . , 2ti−1−qti ξi−1, ξi, 2ti+1−qti ξi+1, . . . , 2tn−qtiξn(cid:17)
is inside a ball with radius 2. Moreover, from Remark 3.1, we have ti . qt j for every pair of
i, j ∈ {1, 2, . . . , n} provided that the support is non-empty. Therefore, the dilation t−1ti has all
its components bounded, for every i. Indeed, we have 2tj−qti . 1 for every i and j.
Let I ∪ J = {1, 2, . . . , n} as in Lemma 3.1. Decompose ξ = (ξI, ξJ) where ξI consists all
components ξi for i ∈ I and ξJ consists of all components ξ j for j ∈ J. We have ηi ∼ 1
for every i ∈ I, by Lemma 3.1.
Integrating by parts with respect to η in (4. 6) gives that the integral inside the bracket equals
to
∂ηα δt(cid:16)t−1η(cid:17) ω(x − tz, t−1η)) dη
for every multi index α. From differential inequality (4. 5) and Lemma 3.1, we have
(4. 5)
(4. 6)
(4. 7)
(4. 8)
(4. 9)
∂α
∂ηα
I
∂β
β
∂η
J
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2πiz(cid:19)αZ e2πiz·η( ∂α
(cid:18) 1
ω(x − tz, t−1η)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
. Yi∈I (cid:16)2ti(cid:17)αi
. Yi∈I (cid:16)2ti(cid:17)αi
ηi!αi
1
. Yi∈I
1 + 2ti(cid:12)(cid:12)(cid:12)ηi(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)t−1η(cid:12)(cid:12)(cid:12)ρ
αiYj∈J (cid:16)2tj(cid:17)βj
1 + 2ti(cid:12)(cid:12)(cid:12)ηi(cid:12)(cid:12)(cid:12)
·
(cid:12)(cid:12)(cid:12)ηı(cid:12)(cid:12)(cid:12)
ρβ
1
1
1
≤ Aα,β,ρ
12
βj
1
1 + 2tj(cid:12)(cid:12)(cid:12)ηj(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)t−1η(cid:12)(cid:12)(cid:12)ρ
αiYj∈J (cid:16)2tj(cid:17)βj
1 + (2tı)ρ(cid:12)(cid:12)(cid:12)ηı(cid:12)(cid:12)(cid:12)ρ
1
βj
provided that there exists ı ∈ I such that 2tı ∼ 2qtj for every j ∈ J, from Lemma 3.1.
The estimates in (4. 8)-(4. 9) imply that
(cid:12)(cid:12)(cid:12)(cid:12)(cid:16)∆tTσ f(cid:17)(x)(cid:12)(cid:12)(cid:12)(cid:12) ≤ AN,qZ f (x − tz)
1
1 + zN! dz
(4. 10)
can be uniformly approximated by the absolutely
ball centered at the origin with radius k. We then have
for every N ≥ 1. The function(cid:16)1 + zN(cid:17)−1
convergent seriesX akχB(k) where each ak is positive constant and B(k) ⊂ RN is the standard
akχB(k) dz .
Z f (x − tz)
QZQ f (x − y)dy
akB(k) sup
Q⊂RN
∞Xk=1
∞Xk=1
(4. 11)
1
. M f (x).
We thus obtain the result in (4. 2).
The Lp continuity of M follows that
(cid:16)M f(cid:17)(x) ≤ (cid:16)Me1(cid:16)Me2(cid:16)··· Men f(cid:17)(cid:17)(cid:17)(x)
(4. 12)
as was investigated in [5] with each Mei the standard maximal function operator on RNi for
i = 1, 2, . . . , n.
It is clear that for all t that (3. 10) does not hold, the summation of δt(ξ) has a compact support.
Before moving forward to the Lp estimate, we hereby give a remark on the following fact:
where the summations are taking over all t s such that (3. 10) does not hold, or in other words,
Remark 4.1
(cid:12)(cid:12)(cid:12)(cid:12)X ∆tTσ f(cid:12)(cid:12)(cid:12)(cid:12) . M f,
q − 1(cid:16)2 + log2 n(cid:17) ,
1
ti ≤
(cid:12)(cid:12)(cid:12)(cid:12)X Tσ∆t f(cid:12)(cid:12)(cid:12)(cid:12) . M f
for all i = 1, 2, . . . , n.
(4. 13)
Let ϕ = ϕ(ξ) be a smooth function supported on a ball centered at the origin. We decompose
Tσ = Tσϕ + Tσ(1−ϕ). In the other hand, define the operatoreTσ by
(cid:16)deTσ f(cid:17)(ξ) = (1 − ϕ(ξ)) d(cid:16)Tσ(1−ϕ) f(cid:17)(ξ).
By adjusting the size of the ball, we can assume that
eTσ f = Xt,s
∆tTσ∆s f
(4. 14)
(4. 15)
where the summation is taking over all t and s such that (3. 10) is satisfied. By Remark 4.1, it
is suffice to estimate the Lp continuity ofeTσ. For the sake of brevity, we will take Tσ =eTσ for
the next section.
13
5 Lp Estimate
In this section, we prove the Lp continuity of pseudo differential operator Tσ with σ ∈ Sρ.
The estimation will be developed in the framework of Littlewood-Paley projections. If Tσ is
translation invariant where σ = σ(ξ), then ∆tTσ∆s vanishes when t and s away, in the sense of
that
nXi=1
ti − si > const
by Lemma 3.2. In general, we have the following estimate of decaying:
Lemma 5.1
for some ε = ε(ρ) > 0.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:16)∆tTσ∆s f(cid:17)(x)(cid:12)(cid:12)(cid:12)(cid:12) .
nYi=1
2−εti−si(cid:16)M f(cid:17)(x)
Proof: 1. Recall the formulae in (4. 3)-(4. 4). By direct computations, we have
where
and
(cid:16)∆tTσ∆s f(cid:17)(x) = Z f (x − y)Ωt,s(x, y)dy
Ωt,s(x, y) = Z e2πiy·ξδt(ξ)ωs(x − y, ξ)dξ
ωs(x, ξ) = " e2πi(x−y)·(ξ−η)δs(η)σ(y, η)dydη.
(5. 1)
(5. 2)
(5. 3)
(5. 4)
Observe that δs(η)σ(y, η) satisfies differential inequality (2. 1). By Lemma 2.2, ωs(x, ξ) is
absolutely bounded and satisfies differential inequality (2. 19).
2. We momentarily assume that σ(x, ξ) has x-compact support.
inside ωs implies
Integration by parts in y
ωs(x, ξ) = (cid:18) 1
2πi(cid:19)2" e2πi(x−y)·(ξ−η)δs(η)
1
ξ − η!2
∆yσ(y, η)dydη.
(5. 5)
q
η
Again, we have δs(η)∆yσ(y, η) satisfies differential inequality (2. 1). By Lemma 2.2, ωs(x, ξ) is
bounded by
ξ − ηq! 2
for which ξ ∈ suppδt(ξ) and η ∈ suppδs(η).
3. We set I1 ∪ J1 = I2 ∪ J2 = {1, 2, . . . , n} for which I1, I2 take the same meaning of I and J1, J2
take the same meaning of J in Lemma 3.1. Consider ti for i ∈ I1 ∪ J1 and si for i ∈ I2 ∪ J2
respectively.
Let ηı to be the largest component of η so that η ∼ ηı ∼ 2sı.
(5. 6)
14
Assume ti , si for some i ∈ {1, 2, . . . , n}. We first consider the case for ti < si and qti < sı.
From Remark 3.1, we have t j ≤ qti for every i, j ∈ {1, 2, . . . , n} whenever the support of δt(ξ)
is non-empty. By Lemma 3.1, we must have ξ . η. Therefore,
η
ξ − ηq
. 2−(q−1)sı ≤ 2−(q−1)si = 2−(q−1)ti 2−(q−1)(si−ti),
ti < si.
(5. 7)
The rest of possible cases will be estimated separately.
Case 1: Suppose that i ∈ I1 ∩ I2. By Lemma 3.1, we have ξi ∼ 2ti and ηi ∼ 2si such that
(cid:12)(cid:12)(cid:12)ξi − ηi(cid:12)(cid:12)(cid:12) ∼
2ti = 2si2(ti−si),
ti > si,
2si = 2ti2(si−ti),
ti < si,
qti ≥ sı.
(5. 8)
(5. 9)
Since qsi ≥ sı for every i from Remark 3.1. (5. 8) implies
. 2−qti−si.
η
ξ − ηq
Case 2: Suppose i ∈ I1 ∩ J2. By Lemma 3.1, we have ξi ∼ 2ti and ηi . 2si.
If si < ti, we are in the first case of (5. 8). Let si > ti. By Lemma 3.1, there exists an ∈ I2 such
that s ∼ qsi and η ∼ 2s ∼ 2qsi. In the other hand, we have t < qti since i ∈ I1. Therefore,
ξı − ηı ∼ 2sı with sı > tı. If qt < sı, we go back to the estimate in (5. 7). If qt > sı, we are in
the second case of (5. 8). The same argument applies to i ∈ J1 ∩ I2.
Lastly, suppose i ∈ J1 ∩ J2. By Lemma 3.1, there exists ı ∈ I1 such that ξı ∼ 2tı ∼ 2qti. If ı ∈ I2,
we are in Case 1. If ı ∈ J2, we are in Case 2.
4. By putting together all estimates (5. 7)-(5. 9), one easily generalizes that
δt(ξ)ωs(x, ξ) ≤
nYi=1
2−εti−si,
ε = ε(q) > 0.
(5. 10)
Let y = tz. By carrying out the same estimations from (4. 6)-(4. 10), with Ωt(x, y) replaced by
Ωt,s(x, y), we have
(cid:16)∆tTσ∆s f(cid:17)(x) ≤ AN,q
nYi=1
2−εti−si Z f (x − tz)(cid:18)
1
1 + z(cid:19)N
dz
(5. 11)
for ε = ε(q) > 0 and every N ≥ 1.
Since all estimates are independent from the size of the x-support of σ(x, ξ), the compactness
can be removed by taking approximation in the sense of distribution. (5. 11) indeed implies
the decaying of ∆tTσ∆s in (5. 1), by following the argument in the end of the previous section.
(cid:3)
Let hi be integers for every i = 1, 2, . . . , n. We define the n-tuples (t + h)i and the non-isotropic
dilations (t + h)iξ simply by replacing ti with ti + hi respectively in (3. 2) and (3. 3). We then
have
δt+h(ξ) =
nYi=1
ψ(cid:16)(t + h)iξ(cid:17)
15
as in (3. 6). The partial sum operator ∆
Define the operator
t+h is defined by(cid:16) d∆
t+h f(cid:17)(ξ) = δt+h (ξ)bf (ξ) as in (3. 7).
(5. 12)
Λ
t,h (cid:17) ∆tTσ∆
t+h.
We have the following results of almost orthogonality:
Lemma 5.2
for some ε = ε(ρ) > 0.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:16)Λ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:16)Λ
t,h∗Λ
Λ
t,h
s,h f(cid:17)(x)(cid:12)(cid:12)(cid:12)(cid:12) .
s,h∗ f(cid:17)(x)(cid:12)(cid:12)(cid:12)(cid:12) .
nYi=1
nYi=1
2−εti−si ×
2−εti−si ×
nYi=1
nYi=1
2−2εhi(cid:16)M f(cid:17)(x)
2−2εhi(cid:16)M f(cid:17)(x)
(5. 13)
Proof: Recall from (5. 2)-(5. 4). We have(cid:16)∆tTσ∆s(cid:17)∗ = ∆sTσ∗∆t. By symmetry, it is suffice to
estimate one of the inequalities in (5. 13).
A direct computation shows
(cid:16)Λ
t,h∗Λ
s,h f(cid:17)(x) = Z f (x − y)(cid:16)Ω
t,t+h(x, y)Ω
s,s+h(x, y)(cid:17)dy
= Z f (x − y)(Z e2πiy·ξδt+h(ξ)ωt,s(x − y, ξ)dξ) dy
(5. 14)
(5. 15)
where
ωt,s(x, ξ) = " e2πi(x−y)·(η−ξ)δs+h(η)ωt(y, ξ)ωs(y, η)dydη.
From Lemma 5.1, both ωt(x, ξ) and ωs(x, η) in (5. 15) have their norm bounded byQ 2εhi for
some ε = ε(ρ) > 0. In the other hand, we have ωt(x, ξ)ωs(x, η) satisfies differential inequality
(2. 1). By Lemma 2.2, we have ωt,s(x, ξ) satisfies differential inequality (2. 19) again, with its
norm bounded by
nYi=1
2−2εhi,
ε = ε(ρ) > 0.
(5. 16)
By carrying out the same estimation in the proof of Lemma 5.1, with ωs(x, ξ) replaced by
ωt,s(x, ξ), the desired result follows.
(cid:3)
We now start to conclude the Lp continuity of Tσ. Let f ∈ L2 ∩ Lp and g ∈ L2 ∩ Lq with
1/p + 1/q = 1. Consider the identity
Z (cid:16)Tσ f(cid:17)gdx = Z Xt
∆sg dx.
Xs
only upon to a finitely many of s, such thatPti − si < const. Write s = t + m and denote
m =Pti − si. Plancherel theorem implies that (5. 17) equals to
By Lemma 3.2, for each t fixed, the intersection of suppδt(ξ) and suppδs(ξ) is non-empty
∆tTσ f
(5. 17)
∆tTσ f ∆t+mg dx.
(5. 18)
Xm<const
Z Xt
16
In the other hand, for each t fixed, we write f =Ph
all n-tuples h = (h1, h2, . . . , hn) of integers. We further write (5. 18) as
∆
t+h f where the summation is taking over
By applying Schwarz inequality and then H older inequality, we have
Xm<const
Z (cid:16)Tσ f(cid:17)gdx ≤ const ×Xh
≤ const ×Xh
Z Xt
Xh
Z Xt
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xt
∆tTσ∆
∆tTσ∆
∆tTσ∆
1
.
t+h f ∆t+mg dx
2Xt
∆tg2
t+h f2
(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)(cid:13)(cid:13)(cid:13)(cid:13)Lp
Xt
∆tg2
t+h f2
1
2
1
dx
1
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq
(5. 19)
(5. 20)
.
The restriction of L2 boundedness can be removed by taking a sequence of functions f j ∈ L2∩Lp
converging to f ∈ Lp as j −→ ∞ in Lp space, then using the inequalities in (5. 20).
By taking the supremum of all g with kgkLq = 1, (5. 20) implies that
.
(5. 21)
Recall from (4. 2), we have ∆tTσ f bounded by M f for every t. Let h be fixed. By using
the vector-valued inequality of strong maximal function in [3], and then Littlewood-Paley
inequality, we have
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp
1
1
∆tTσ∆
Xt
(cid:13)(cid:13)(cid:13)Tσ f(cid:13)(cid:13)(cid:13)Lp . Xh
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lr
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
t+h f2
. (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lr
Xt (cid:16)M∆
t+h f2(cid:17) 1
. (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lr
Xt (cid:16)∆
t+h f2(cid:17) 1
∼ (cid:13)(cid:13)(cid:13) f(cid:13)(cid:13)(cid:13)Lr ,
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L2
t+h f2
nYi=1
2−εhi(cid:13)(cid:13)(cid:13) f(cid:13)(cid:13)(cid:13)L2 ,
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp
Xt
t+h f2
nYi=1
∆tTσ∆
.
.
1
1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2−ε′hi(cid:13)(cid:13)(cid:13) f(cid:13)(cid:13)(cid:13)Lp
Xt
∆tTσ∆
t+h f2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xt
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
∆tTσ∆
(5. 22)
1 < r < ∞.
ε = ε(ρ) > 0.
(5. 23)
In the other hand, recall from Lemma 5.2. By applying Cotlar-Stein Lemma in [7], together
with Lemma 3.2 and Plancherel theorem, we have
Let p ∈ (r, 2] and p ∈ [2, r). By Riesz interpolation theorem in [13], we have
with ε′ > 0 depending on p and ρ.
Lastly, by summing over all the hi s in the summand of (5. 21), we obtain the desired result.
Namely, let σ ∈ Sρ with 0 < ρ < 1, we have
for 1 < p < ∞.
(cid:13)(cid:13)(cid:13)Tσ f(cid:13)(cid:13)(cid:13)Lp ≤ Ap,ρ (cid:13)(cid:13)(cid:13) f(cid:13)(cid:13)(cid:13)Lp
17
(5. 24)
(5. 25)
6 Weighted Norm Inequality
The studies of weighted norm inequalities for maximal functions and singular integrals can
be traced back to the earlier results in 1972 by Muckenhoupt [11], and then in 1974 by Coifman
and C. Fefferman [12].
Write x = (xi, x′i ) ∈ RNi × RN−Ni where x′i denotes the complement of xi in RN. We first
introduce the following class of weight functions defined in [3].
Definition of A∗p : Let λ be an non-negative, locally integrable function on RNi for every
i = 1, 2, . . . , n. Then, λ ∈ A∗p for 1 < p < ∞ if
1
BiZBi
λ(xi, x′i )dxi ·
1
BiZBi
1
λ(xi, x′i )! q
p
p
q
≤ A,
(6. 1)
dxi
i = 1, 2, . . . , n
with 1/p + 1/q = 1, for all Bi ⊂ RNi uniformly in x′i ∈ RN−Ni.
In below, we recall the two results proved in [3] by R. Fefferman and Stein.
(∗) Given dµ to be an nonnegative Borel measure.
dµ(x)
(6. 2)
Z (cid:12)(cid:12)(cid:12)M f (x)(cid:12)(cid:12)(cid:12)p
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xt
dµ(x) ≤ ApZ (cid:12)(cid:12)(cid:12) f (x)(cid:12)(cid:12)(cid:12)p
2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp(λ)
(cid:12)(cid:12)(cid:12)∆t f (x)(cid:12)(cid:12)(cid:12)2
∼ (cid:13)(cid:13)(cid:13) f(cid:13)(cid:13)(cid:13)Lp(λ)
if and only if dµ(x) = λ(x)dx is absolutely continuous, with λ ∈ A∗p , for 1 < p < ∞.
(∗∗)
1
(6. 3)
with λ ∈ A∗p for 1 < p < ∞.
Noted that the second result (∗∗) stated hereby is proved in [3] with the language of square
functions. We are going to show that the two classes Ap and A∗p are equivalent, with a
mollification on a set of measure zero. Without losing of generality, we assume n = 2 such
that (x, y) ∈ RN1 × RN2.
First, suppose λ ∈ Ap. By H older inequality, we have
1
B2ZB2 1
B1ZB1
≤
p
1
q
1
1
p
λ(cid:0)x, y(cid:1) dx! 1
B1ZB1
λ(cid:0)x, y(cid:1)! q
p
λ(cid:0)x, y(cid:1) dxdy! 1
B1B2 "B1×B2
dx
B1B2 "B1×B2
dy
1
1
p
1
λ(cid:0)x, y(cid:1)! q
dxdy
1
q
≤ A.
(6. 4)
Suppose the integrant on the left of inequality (6. 4) is strictly greater than A, on any subset
of RN2, with non-empty interior. Then, taking B2 to be contained in that set will reach a
18
contradiction. A more direct argument by applying Lesbegue differentiation theorem was
given in [16].
Turn to the converse. Suppose λ ∈ A∗p. By (∗) in (6. 2), choosing f to be non-negative, we have
λ(x, y)dxdy ≤ Ap " (cid:16) f (x, y)(cid:17)p
Let B1 × B2 ∈ RN1 × RN2. It is clear that inequality (6. 5) holds for
" (cid:16)M f (x, y)(cid:17)p
λ(x, y)dxdy.
(6. 5)
" (cid:16)M f (x, y)(cid:17)p
λ(x, y)χB1×B2dxdy
where χB1×B2 is the characteristic function on B1 × B2. For any x ∈ B1, leteB1(x) denotes the ball
centered on x with twice the radius of B1. Certainly B1 ⊂eB1(x) for every x ∈ B1. The same is
true for any y ∈ B2. Therefore, we shall have
f (u, v)dudv ≤ M f (x, y),
(x, y) ∈ B1 × B2.
2−(N1 +N2)
≤
1
B1B2 "B1×B2
"
(cid:12)(cid:12)(cid:12)eB1(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)eB2(y)(cid:12)(cid:12)(cid:12)
1
f (u, v)dudv
eB1(x)×eB2(y)
Let N = N1 + N2, the estimates (6. 5)-(6. 6) imply
1
B1B2 "B1×B2
λ(x, y)dxdy!
f (x, y)dxdy!p "B1×B2
≤ 2NpAp" (cid:16) f (x, y)(cid:17)p
λ(x, y)dxdy.
By choosing f = (1/λ + ε)
q
p for some ε > 0 and taking into account that
q
p
= q − 1
and
p
q
= p − 1,
(6. 6)
(6. 7)
inequality (6. 7) implies
1
B1B2 "B1×B2
λ(x, y)dxdy!
1
B1B2 "B1×B2
λ(x, y) + ε! q
1
p
p
q
≤ A
(6. 8)
dxdy
where the constant A = 2NpAp depends only on p and λ. By letting ε −→ 0, we have λ ∈ Ap.
To conclude that Tσ f with σ ∈ Sρ admits the desired weighted norm inequality, we simply
carry out the same estimation in section 5, with the Lp norm replaced by Lp(λ) norm.
In
particular, (5. 20) holds provided by λ ∈ Ap if and only if (1/λ)q/p ∈ Aq with 1/p + 1/q = 1. See
chapter 3 of [3] and chapter V of [7]. In (5. 22), we need apply the vector-valued weighted
norm inequality for M f , proved in chapter 3 of [3], together with (∗∗), which are now valid
for λ ∈ Ap.
19
7 Estimation on Kernel
In this section, we are going to show that the class of operator Tσ with σ ∈ Sρ can be
equivalently classified by
Z f (x − y)Ω(x, y)dy
where Ω satisfies differential inequality (1. 5) and cancellation property (1. 6). This will be
showed by obtaining the following theorem.
Theorem 7.1 The symbol σ ∈ Sρ if and only if the kernel Ω(x, y) with y , 0, satisfies differential
inequality (1. 5) and the cancellation property (1. 6).
We divided the proof into two major parts.
from symbol to kernel: Recall that the kernel Ω is a distribution Ω(x,·) that coincides with a
smooth function
for y , 0. Suppose σ(x, ξ) satisfies differential inequality (1. 2). We write
∂α
∂yα
∂β
∂xβ
e2πiy·ξσ(x, ξ)dξ
Ω(x, y) = ZRN
i ∂β
∂xβ σ(x, ξ)! dξ
2πi(cid:19)−αZ e2πiy·ξ
nYi=1
∂xβ δt(ξ)σ(x, ξ)! dξ
i ∂β
2πi(cid:19)−αXt Z e2πiy·ξ
nYi=1
ξαi
ξαi
Ω(x, y) = (cid:18) 1
= (cid:18) 1
(7. 1)
(7. 2)
,
i = 1, 2, . . . , n
(7. 3)
where the summation is taking over all t satisfying (3. 10).
Let ξ ∈ suppδt(ξ). For each component ξi, we have
∂α
∂ξα
i
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
σ(x, ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
.
1
1 + ξi + ξρ!α
∂β
for every multi-index α. From Lemma 3.1, we either have ξi ∼ 2ti or otherwise there exists
ξı such that ξı ∼ 2tı ∼ 2ti/ρ. Thus that the derivative in (7. 3) is bounded by 2ti, for every i.
In the other hand, by differential inequality (1. 2), we have
. (1 + ξ)ρβ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂xβ σ(x, ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for every multi-index β. Clearly, from Lemma 3.1, (7. 4) is further bounded by constant
multiple of 2ρβti, for some i ∈ {1, 2, . . . , n}.
Now, consider the norm of
σ(x, ξ)! dξi
i ∂β
j (δt(ξ)ξαi
e2πiy·ξ ∂µi
∂ξ
(yi)µi(yj)νiZRNi
= (cid:18) −1
2πi(cid:19)µi+νiZRNi
σ(x, ξ)!) dξi,
e2πiy·ξδt(ξ)ξαi
i ∂β
∂xβ
(7. 5)
∂νi
ξνi
µi
i
(7. 4)
j , i.
∂xβ
20
By previous estimates in (7. 3) and (7. 4), we have the norm of (7. 5) bounded by a constant
multiple of
By Lemma 3.1, we have ti ≤ t j/ρ for every pair of i, j ∈ {1, 2, . . . , n}. Hence that the norm of
(7. 5) is further bounded by a constant multiple of
2−µiti × 2−νitj × 2(Ni+αi+ρβ)ti.
2−(µi+ρνi)ti × 2(Ni+αi+ρβ)ti.
(7. 6)
Let νi fixed, we consider the summations separately for
X2ti≤yi−1
and
X2ti >yi−1
.
For 2ti ≤ yi−1, we choose µi = 0 so that the summation
2(Ni+αi+ρβ−ρνi)ti . 1
yi!Ni+αi+ρβ−ρνi
.
(7. 7)
For 2ti > yi−1, we choose µi > Ni + αi + ρβ − ρνi so that
2(Ni +αi+ρβ−ρνi−µi)ti . 1
yi!Ni+αi+ρβ−ρνi
.
(7. 8)
X2ti≤yi−1
1
yi!µi X2ti >yi−1
In the other hand, let µi fixed, we consider
X2ti≤y′i−1/ρ
and
X2ti >y′i−1/ρ
.
For 2ti ≤ y′i−1/ρ, we choose νi = 0 so that the summation
X2ti≤y′i−1/ρ
2(Ni +αi+ρβ−µi)ti . 1
y′i!(Ni+αi+ρβ−µi)/ρ
.
(7. 9)
For 2ti > y′i−1/ρ, we choose νi >(cid:0)Ni + αi + ρβ − µi(cid:1) /ρ so that
y′i!νi X2ti >y′i−1/ρ
1
2(Ni+αi+ρβ−µi−ρνi)ti . 1
y′i!(Ni +αi+ρβ−µi)/ρ
By summing all together, we shall have
nYi=1
yiµiy′iνi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂α
∂yα
∂β
∂xβ
Ω(x, y)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aµ,ν
for every multi-index µ = (µ1, µ2, . . . , µn) and ν = (ν1, ν2, . . . , νn), provide that
µi + ρνi ≥ Ni + αi + ρβ
i = 1, 2, . . . , n
21
.
(7. 10)
(7. 11)
(7. 12)
We can then verify differential inequality (1. 5). It is also clear that Ω(x, y) in (7. 1) decays
rapidly as y −→ ∞.
To show the cancellation property, let ϕi ∈ C∞(RNi) be an normalized bump function, and
Ri > 0 for every i = 1, 2, . . . , n. Observe that
We have (7. 13) equals to
where
We rewrite (7. 14) as
ZRNi ∂α
∂xα
i bϕi(cid:16)−R−1
Ω(x, y)! ϕi(Ri yi)dyi
= Z e2πiy′i·ξ′i ∂α
e2πiy′i·ξ′i ∂α
ZRN−Ni
∂xα σ(cid:16)x, ξ′i(cid:17) = ZRNi ∂α
∂xα σ(x, ξ)! R−Ni
∂xα σ(cid:16)x, ξ′i(cid:17)! dξ′i
∂xα σ(x, ξ)! R−Ni
i bϕi(cid:16)−R−1
(1 + ξ)ρα ∼ ξiρα +(cid:16)1 + ξ′i(cid:17)ρα (cid:17) ϑ(ξ).
∂xα σ(x, ξ)ϑ−1(ξ)! R−Ni
∂α
ZRNi ξiρα ∂α
i ξi(cid:17) dξ.
i ξi(cid:17) dξi.
+(cid:16)1 + ξ′i(cid:17)ραZRNi ∂α
i bϕi(cid:16)−R−1
i ξi(cid:17) dξi
∂xα σ(x, ξ)ϑ−1(ξ)! R−Ni
i bϕi(cid:16)−R−1
i ξi(cid:17) dξi.
(7. 13)
(7. 14)
(7. 15)
(7. 16)
(7. 17)
Observe that ∂α
x σ(x, ξ) satisfies differential inequality (1. 2) with norm bounded by
Now, ∂α
x σ(x, ξ)ϑ−1(ξ) is absolutely bounded and satisfies differential inequality (1. 2). The
sufficient smoothness of ϕi implies the L1 norm of bϕi is bounded. By changing dilations
ξi = Riηi, so that 0 < ηi < 1, it follows that σ(x, ξ′i ) satisfies differential inequality (1. 2)
in RN−Ni with its norm bounded by (1 + Ri)ρα. An analogue of estimation also valid on
L RNi, whenever i ∈ I as a subset of {1, 2, . . . , n}. By carrying out the induction, we obtain
the cancellation property (1. 6).
from kernel to symbol
Suppose Ω(x, y) satisfies differential inequality (1. 5) and cancellation property (1. 6). We
have
σ(x, ξ) = Z Ω(x, y)e−2πiy·ξdy.
Let
In order to prove σ(x, ξ) satisfies differential inequality (1. 2), we are going to show that
ρi(ξ) = 1 + ξi + ξρ.
∂ξαZ ∂β
∂xβ
Ω(x, y)! e−2πiy·ξdy
ρi(ξ)αi ∂α
nYi=1
22
is absolutely bounded by
Aα,β(1 + ξ)ρβ
(7. 18)
for every multi-index α and β.
Differentiations with respect to ξ inside (7. 17) imply that it is equal to (−1)α(2πi)α times
∂xβ
Z nYi=1 (cid:16)ρi(ξ)yi(cid:17)αi ∂β
Ω(x, y)! e−2πiy·ξ
nYi=1
Ω(x, y)! e−2πiy·ξdy.
ϕi(cid:16)ρi(ξ)yi(cid:17)
i ϕi(yi)exp(−2πi yi · ξi
ρi(ξ)!)
nYi=1 (cid:16)ρi(ξ)yi(cid:17)αi
yαi
(7. 19)
dy.
(7. 20)
(7. 21)
Let ϕi for i = 1, 2, . . . , n be normalized bump functions. Consider
Z ∂β
∂xβ
Observe that the function
is a product of n normalized bump functions of yi for i = 1, 2, . . . , n. By cancellation property
(1. 6), we have (7. 20) is bounded by
Aα,β(cid:16)1 + ρ1(ξ) + ρ2(ξ) + ··· + ρn(ξ)(cid:17)ρβ
which is further bounded by (7. 18).
In the other hand, we have a remaining term of
Z ∂β
∂xβ
Ω(x, y)! e−2πiy·ξ
The integration is taking over
nYi=1 (cid:16)ρi(ξ)yi(cid:17)αi 1 −
supp 1 −
ϕi!.
nYi=1
nYi=1
ϕi(cid:16)ρi(ξ)yi(cid:17)!dy.
(7. 22)
(7. 23)
By definition of ϕi, i = 1, 2, . . . , n, we have yi > ρi(ξ) in the support (7. 23). Since Ω(x, y) has
rapidly decaying as y −→ ∞, the integral converges absolutely.
8 Applications
In this last section, we give three examples. These are the Grushin type operators, oblique
derivative problems and Cauchy-Szeg o projections. For further detail of regarding estimates
, please see chapter VII of [2], chapter IV, XII of [7] and [14].
Example 1: Let x = (t, y) ∈ R × Rn−1 with its dual variable ξ = (τ, η) ∈ R × Rn−1. Consider the
Grushin type operator
(8. 1)
+ t2∆y.
∂2
t
23
The inverse of (8. 1) can be carried out by using Hermite functions, as showed in [15]. We
have the symbol of the inverse operator equals to
where M denotes the Mehler kernel which has the explicit form
0
1
2
eiτt
σ(t, τ, η) = −
η Z 2π
s(cid:19) 1
(cid:18) 1
2 exp
M(r, u, v) = (cid:16)π(1 − r2)(cid:17)− 1
1
2
s
1
2 t,
2πi
,η
τ
η
2 ds.
M
4ruv −(cid:16)u2 + v2(cid:17)(cid:16)1 + r2(cid:17)
2(cid:16)1 − r2(cid:17)
(8. 2)
.
(8. 3)
Recall from § 14, chapter IV of [2]. Let ρ = ρ(t, τ, η) be a distance function such that
ρ(t, τ, η) ∼ 1 + τ + tη + η
1
2 .
(8. 4)
The symbol σ(t, τ, η) defined in (8. 2)-(8. 3) has its norm bounded by ρ(t, τ, η)−2 and satisfies
the differential inequality
∂α
∂τα
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂β
∂ηβ σ(t, τ, η)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,β
1
ρ(t, τ, η)!α+2β
.
(8. 5)
In the other hand, the t-derivatives of σ(t, τ, η) satisfy a variant of differential inequalities
1
defined inductively on the order of derivatives. Each one gains at most a factor of (1+τ+η)
2 .
Observe that
1 + τ + η
1
2 . ρ(t, τ, η)
and 1 + τ2 + η . ρ(t, τ, η)2
(8. 6)
uniformly in t. (8. 5)-(8. 6) implies that σ(t, τ, η) satisfies differential inequality (1. 2) with
ρ = 1/2.
Let σi denote the symbol in (8. 2) with t chosen to be xi so that τ = ξi respectively for
i = 1, 2, . . . , n. We have σi ∈ S 1
. Moreover, let (σ1 ◦ ··· ◦ σn) denotes the symbol of the
composition operator Tσ1 ◦ ··· ◦ Tσn. We have
2
∂α
∂ξα
∂β
∂xβ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(σ1 ◦ ··· ◦ σn) (x, ξ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,β
nYi=1
2!αi
1
1
1 + ξi + ξ
for every multi-index α = (α1, α2, . . . , αn) and β. Hence that (σ1 ◦ ··· ◦ σn) ∈ S 1
2
.
β
2
(1 + ξ)
(8. 7)
Example 2: Let S be a smooth bounded, open set in Rn+1, and X be a given smooth real
vector field on its boundary ∂S. Consider the oblique derivative problem as follows:
∆u = 0
on S,
Xu = f
on ∂S.
(8. 8)
The question can be reduced to that of inverting a pseudo differential operator of order one
on ∂S. An introduction of further background can be found in [17] and [18]. If X is transverse
to ∂S, then it is the classical Neumann problem and the boundary operator has a symbol
24
which is elliptic. In general, we decompose X as XT + αbn where XT is tangential to ∂S and
bn is the outer normal field on ∂S. Noted that α = α(x) is considered here to be emergent type,
as a necessary condition for the existence of solution. This means that α never changes its
sign on the integral curve along XT. In other words, X can change its orientation only if ∂S
is non-orientable. See [19] for references. In the other hand, we say α is in finite type if there
exists an integer k and a constant δ > 0, such that
α + XTα + ··· + Xk
Tα ≥ δ.
(8. 9)
An investigation of general oblique derivative problems with k ≥ 1 in (8. 9) can be found
in [20]. In the present example, for simplicity we only consider k = 1 of which the relevant
estimates also hold for higher value k s.
By appropriately choosing the local coordinates (x1, x2, . . . , xn), we have X is transverse to the
manifold M = {xn = 0} which has co-dimension one to ∂S. The boundary operator involved
has a symbol essentially equals to
iξn ± q(x, ξ)xn,
q(x, ξ) = Xi, j
aij(x)ξiξ j! 1
2
(8. 10)
with {aij(x)} smoothly varying, real positive definite symmetric matrix.
It is well known
that iξn + q(x, ξ)xn can only have a right parametrix and iξn − q(x, ξ)xn can only have a left
parametrix. The two problems are adjoint to each other. Write x = (t, y) with t = xn and its
dual ξ = (τ, η). For the sake of brevity, in what follows we seek only for the right parametrix
of
(8. 11)
Recall the strategy given in § 15, chapter IV of [2]. By inverting the ordinary differential
operator ∂t + tη, where the inverse is taking to be L2-orthogonal to the non-trivial null space
e−t2η/2, our desired right parametrix is
iτ + tη,
t > 0.
σ(t, τ, η) = rη
π Z ∞
−∞
e−ηz2(Z t
z
e−η(t2−s2)e2πi(t−s)τds) dz.
(8. 12)
The symbol in (8. 12) has its norm bounded by ρ(t, τ, η)−1 defined in (8. 4), and satisfies
Its t-derivatives satisfy a variant of differential inequalities
differential inequality (8. 5).
1
defined inductively on the order of derivatives. Each one gains at most a factor of (1+τ+η)
2 .
. Moreover, let σi denote the symbol in (8. 2) with t chosen to be xi so
We thus have σ ∈ S 1
that τ = ξi respectively for i = 1, 2, . . . , n. The symbol (σ1 ◦ ··· ◦ σn) belongs to S 1
.
2
2
Example 3: Our last example is the Cauchy-Szeg o projections on the Heisenberg groups. Let
x = (z, t) ∈ Cn × R and y = (w, s) ∈ Cn × R. Define the multiplication law ⊗ by
x ⊗ y = (cid:16)z + w , t + s + 2Im (z · w)(cid:17).
(8. 13)
An inverse element is taking to be x−1 = (−z,−t) and the identity is taking to be the origin
(0, 0). The multiplication law ⊗ defined in (8. 13) turns the space Cn × R into the Heisenberg
group, denoted by Hn.
25
Essentially, a Cauchy-Szeg o kernel is a distribution, represented by the function
K(t, z) =
1
t + iz2!n+1
(8. 14)
away from the origin. As showed in [14], its Fourier transform behaves like e−η2/2τ for τ > 0
and is zero for τ ≤ 0. We shall be looking at the local version of the Cauchy-Szeg o projection,
defined by
ZHn
ϕ1(x)K(cid:16)y−1 ⊗ x(cid:17) ϕ2(y) f (y)dy
(8. 15)
where ϕ1, ϕ2 are some smooth cut-off functions.
Let ξ = (ζ, τ) be the dual variable of x = (z, t) on the space Cn × R, implemented by the inner
product x · ξ = Re(cid:16)z · ζ(cid:17) + tτ. The symbol corresponding to the Cauchy-Szeg o projections is
derived from
σ(ζ, τ) =
ψ(ζ, τ)e−ζ2/2τ
0
τ > 0,
τ ≤ 0
(8. 16)
where ψ is a smooth cut-off function that vanishes near the origin and equals 1 for large (ζ, τ).
See chapter VII of [7].
Let ρ = ρ(ζ, τ) be a distance function such that
ρ(ζ, τ) ∼ 1 + ζ + τ
The symbol σ in (8. 16) satisfies the differential inequality
1
2 .
∂α
∂ζα
∂β
∂τβ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for every multi-index α and β.
σ(ζ, τ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,β
1
ρ(ζ, τ)!α+2β
(8. 17)
(8. 18)
In the other hand, the kernel defined bybΩ(ζ, τ) = σ(ζ, τ), is a distribution smooth away from
the origin, satisfies the differential inequality
and decays rapidly as (z, t) −→ ∞.
We thus can verify the size of Cauchy-Szeg o kernel given by (8. 14), for which
∂β
∂tβ
∂α
∂zα
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Ω(z, t) .
Ω(z, t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Aα,β
2!2n
1
1
z + t
2!2n+α
1
1
t + z2!1+β
1
z + t
1
t + z2! ∼
1
t + z2!n+1
.
(8. 19)
(8. 20)
In fact, from (8. 15), it is not restricted to allow the kernel has rapidly decaying at the
infinity. As showed in 7.15, chapter XII of [7], Cauchy-Szeg o projection given by (8. 15) can
be considered as a pseudo differential operator Ta.
26
Its symbol has a form of
a(x, ξ) = ϕ(x)σ(cid:16)Lxξ(cid:17)
(8. 21)
where ϕ is a smooth cut-off function and Lx is a linear transformation, which actually depends
linearly on x. To be precise, recall the multiplication law ⊗ defined in (8. 13). We write
y−1 ⊗ x = Mx(x − y) where Mx is a linear transformation with determinant 1, and depends
linearly on x. Lx is taking to be the transpose inverse of Mx.
In what follows, we take n = 1 for brevity. Accordingly, we have
Lx =
1
0
0
0
2z2
1 − 2z1
1
0
.
(8. 22)
Then, the symbol a is written explicitly as
a(x, ξ) = ϕ(x)ψ (Lxξ) exp (ζ1 + 2z2τ)2 + (ζ2 − 2z1τ)2
−2τ
! ,
τ > 0
(8. 23)
and is zero for τ ≤ 0. Recall from chapter VII, XII of [7], we have a(x, ξ) in (8. 23) belongs to
the exotic class S0
. In the following, we aim to show that a ∈ S 1
2 , 1
1
2
.
2
1
Suppose ζ & τ
varying of z2, we can have z2τ ∼ ζ1 and ζ1 + 2z2τ ∼ τ
2 . Without losing of generality, let ζ ∼ ζ1 ∼ τ
1
+δ for some 0 ≤ δ ≤ 1
2 . By
2 . Therefore,
2−ε, for some δ ≤ ε ≤ 1
1
2
1
2 , the quotient in (8. 24) is dominated by τ− 1
1
2 . We thus have each ζ-derivative
2 )−1.
When ζ . τ
gains a factor of (1 + ζ + τ
Turning to the τ-derivatives. For ζ . τ
implies z ∼ τ− 1
have ζ1 + 2z2τ ∼ τ
z2 ∼ τ− 1
2−ε for some δ ≤ ε ≤ 1
2−δ; For τ
1
1
1
2
2 . ζ . τ. Let ζ1 ∼ τ
1
2 . Then ζ1 ∼ z2τ ∼ τ
2−δ for some 0 ≤ δ ≤ 1
2
2 . In order to
2 , we necessarily have z2τ ∼ ζ1 which implies
+δ for some 0 ≤ δ ≤ 1
1
2
+δ. Therefore, in all cases for ζ . τ, each τ-derivative is dominated by
z2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ζ1 + 2z2τ
ζ1 + 2z2τ
τ
+δτ− 1
The same estimate goes to ζ2 − 2z1τ. Together with that a ∈ S0
2 , 1
1
2
conclude that a ∈ S 1
Acknowledgement:
τ2
.
2
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . τ− 1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) + (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2−δ ∼ τ−1.
(8. 25)
showed in [7] and [14], we
I would like to express my strongest appreciation to professor Elias M. Stein. During
the academic year 2012-2013, I was visiting Princeton under his direction. A series of
invaluable talks were held in our weekly meetings. The result in this paper, could not
been obtained without his help and encouragement.
27
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ζ1 + 2z2τ
τ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∼ τ−( 1
2
+ε) . ζ1−1 ∼ ζ−1.
(8. 24)
References
[1] A. Nagel, F. Ricci and E. M. Stein, Singular Integrals with Flag Kernels and Analysis on
Quadratic CR Manifolds, Journal of Functional Analysis, 181, 29-118, 2001.
[2] A. Nagel and E. M. Stein, Lectures on Pseudo Differential Operators: Regularity Theorems
and Applications, Princeton University Press, 1979.
[3] R. Fefferman and E. M. Stein, Singular Integrals on Product Spaces, Advances in Mathe-
matics, 45,117-143, 1982.
[4] D. M uller, F. Ricci, E. M. Stein, Marcinkiewicz Multipliers and Multi-parameter structures
on Heisenberg (-type) group, I, Inventiones Mathematicae, 119, 199-233, 1995.
[5] F. Ricci and E. M. Stein, Multiparameter Singular Integrals and Maximal Functions, Annales
de L’ institute Fourier, 42, 637-670, 1992.
[6] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Uni-
versity Press, 1970.
[7] E. M. Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality and Oscillatory Inte-
grals, Princeton University Press, 1993.
[8] A. Carbery and A. Seeger, Conditionally Convergent Series of Linear Operators on Lp-Spaces
and Lp-Estimates for Pseudodifferential Operators, Proceeding of London Mathematical
Society, 57(3), 481-510, 1988.
[9] L. Boutet de Monvel, Hypoelliptic Operators with Double Characteristics and related Pseudo
Differential Operators, Communication in Pure and Applied Mathematics, 27, 585-639,
1974.
[10] R. Beals and C. Fefferman, Spatially Inhomogeneous Pseudo Differential Operators, Com-
munication in Pure and Applied Mathematics, 27, 1.24, 1974.
[11] B. Muckenhoupt, Weighted Norm Inequalities for the Hardy Littlewood Maximal Function,
Transaction of American Mathematical Society, 156, 207-226, 1972.
[12] R. R. Coifman and C. Fefferman, Weighted Norm Inequalities for Maximal Functions and
Singular Integrals, Studia Mathematica, 51, 241-250, 1974.
[13] E. M. Stein and G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, Princeton
University Press, 1971.
[14] P. C. Greiner and E. M. Stein, Estimates on ∂-Neumann problems, Mathematics Notes 19,
Princeton University Press, 1977.
28
[15] N. Wiener, The Fourier Integrals and Certain of its Applications, Cambridge University
Press, 1989.
[16] D. S. Kurtz, Littlewood-Paley and Multiplier Theorems on Weighted Lp Spaces, Transactions
of the American Mathematical Society, 259, 235-254, 1980.
[17] L. H omander, Pseudodifferential Operators and Non-elliptic Boundary Problems, Annals of
Mathematics, 83, 129-209, 1966.
[18] J. Sj ostrand, Operators of Principal Type with Interior Boundary Conditions, Acta Mathe-
matica, 137, 247-320, 1976.
[19] Ju. V. Egorov and V. A. Kondrat´ev, The Oblique Derivative Problem, Mathematics of the
USSR Sbornik, 7, 148-176, 1969.
[20] H. Smith, The Subelliptic Oblique Derivative Problem, Communications in Partial Differ-
ential Equations, 15, 97-137, 1990.
29
|
1709.04233 | 1 | 1709 | 2017-09-13T10:14:44 | Cone unrectifiable sets and non-differentiability of Lipschitz functions | [
"math.FA"
] | We provide sufficient conditions for a set $E\subset\mathbb{R}^n$ to be a non-universal differentiability set, i.e. to be contained in the set of points of non-differentiability of a real-valued Lipschitz function. These conditions are motivated by a description of the ideal generated by sets of non-differentiability of Lipschitz self-maps of $\mathbb{R}^n$ given by Alberti, Cs\"ornyei and Preiss, which eventually led to the result of Jones and Cs\"ornyei that for every Lebesgue null set $E$ in $\mathbb{R}^n$ there is a Lipschitz map $f:\mathbb{R}^n\to\mathbb{R}^n$ not differentiable at any point of $E$, even though for $n>1$ and for Lipschitz functions from $\mathbb{R}^n$ to $\mathbb{R}$ there exist Lebesgue null universal differentiability sets. | math.FA | math |
Cone unrectifiable sets and non-differentiability of
Lipschitz functions∗
Olga Maleva and David Preiss
Abstract
We provide sufficient conditions for a set E ⊂ Rn to be a non-universal
differentiability set, i.e. to be contained in the set of points of non-
differentiability of a real-valued Lipschitz function. These conditions
are motivated by a description of the ideal generated by sets of non-
differentiability of Lipschitz self-maps of Rn given by Alberti, Csornyei
and Preiss, which eventually led to the result of Jones and Csornyei
that for every Lebesgue null set E in Rn there is a Lipschitz map
f ∶ Rn → Rn not differentiable at any point of E, even though for n > 1
and for Lipschitz functions from Rn to R there exist Lebesgue null
universal differentiability sets.
1
Introduction and main results
A recent surge of interest in validity of Rademacher's theorem on almost
everywhere differentiability of Lipschitz maps of Rn to Rm arose from several
new results and approaches. For infinite-dimensional Banach spaces there
were successful attempts to obtain its analogues for the notion of Gateaux
derivative, for results and references see [6, Chapter 6], and some results for
the stronger notion of Fr´echet derivative to which a recent monograph [17]
is devoted. In another direction, Pansu [19] obtained an almost everywhere
result for Lipschitz maps between Carnot groups, and Cheeger [7] generalised
Rademacher's theorem to Lipschitz functions on metric measure spaces.
Here we contribute to this research in the direction started by a result of
[21] that, in terms of the size of differentiability sets of real-valued Lipschitz
∗The research leading to these results has received funding from the European Research
Council / ERC Grant Agreement n. 291497. The first named author also acknowledges
the support of the EPSRC grant EP/N027531/1 and of the National Science Foundation
under Grant No. DMS-1440140 while the author was in residence at the Mathematical
Sciences Research Institute in Berkeley, California, during the Fall 2017 semester.
1
functions on R2, Rademacher's theorem is not sharp: there is a Lebesgue null
set in R2 containing points of differentiability of any real-valued Lipschitz
function on R2. Following [12, 13], where it was shown how unexpectedly
small such sets may be, they are now called universal differentiability sets.
The analogues of universal differentiability sets were recently introduced and
investigated in the Heisenberg group [20].
The non-differentiability sets of Lipschitz maps Rn → Rm, m ≥ n were
first completely described in geometric measure theoretical terms in [3] (see
[1, 2] for a published less formal description), and then [8] showed that this
description gives precisely the Lebesgue null sets in Rn. Hence Rademacher's
theorem is sharp for maps into spaces of the same or higher dimension. This
result was complemented in [23] where it was proved that whenever m < n,
there is a Lebesgue null set in Rn containing points of differentiability of
any Lipschitz map Rn → Rm. We will return to the description originally
introduced in [3] later as it forms the main starting point of what we do in
the present paper.
The problem we address in this paper stems from the above results: can
one give a geometric measure theoretical description of non-differentiability
sets of Lipschitz maps of Rn to R? Notice that this is a question about
sets and not about measures: if we are given a σ-finite Borel measure µ in
Rn that is singular with respect to the Lebesgue measure, we may use [3]
and [8] to find a Lipschitz µ-almost everywhere non-differentiable mapping
f = (f1, . . . , fn) ∶ Rn → Rn and observe that for a random choice of αi ∈ (0, 1)
the real-valued function ∑n
i=1 αifi is Lipschitz and µ-almost everywhere non-
differentiable. This argument appears both in [3] and [4], and moreover [4]
simplifies the general arguments of [3] in the special case of differentiability
µ-almost everywhere; notice also that in this case even the results of [8] may
be demonstrated by a more accessible proof given in [11] (which is based on
different ideas).
We will now state our results and explain them in more detail. Their
proofs will be given in Section 3. The short Section 4 contains two examples
whose meaning will also be discussed here.
The main concept that we use is based on the notion of width that has
been, together with several variants, introduced in [3].
Definition 1.1. Suppose e ∈ Rn ∖ {0} and α ∈ (0, 1]. We let Ce,α be the
cone {x ∈ Rn ∶ ⟨x, e⟩ ≥ αxe} and Γe,α the set of Lipschitz curves such
that γ′(t) ∈ Ce,α for almost every t ∈ R. The (e, α)-width of an open set
G ⊂ Rn is defined by
we,α(G) = sup{H1(G ∩ γ(−∞, ∞)) ∶ γ ∈ Γe,α},
(1.1)
2
and of any E ⊂ Rn by
we,α(E) = inf{we,α(G) ∶ G ⊃ E, G is open}.
(1.2)
For the sake of completeness, when e = 0 or α > 1 we let we,α(E) = 0 for
every E ⊂ Rn. Of course, these cases have no bearing on what we do.
Notice that, as [3] points out, for constructions of Lipschitz functions,
where values of we,α matter only for arbitrarily small α, the number α in
Definition 1.1 may be replaced by any quantity or function that may attain
arbitrarily small positive values. For example [4] replaces it by cos α, which
is a bound on the angle between γ′(t) and e and so is geometrically natural,
but for us has the disadvantage that values of α that matter, namely those
for which the cone Ce,α is close to a half-space, are close to π~2 rather than
to zero.
Many variants of Definition 1.1 that are easily seen or shown to be
equivalent to the one given here may be found in [22, Definition 1.1 and
Remark 1.2]. A rather useful variant is that Γe,α may be defined as the col-
lection of γ ∈ C 1(R, Rn) satisfying γ′(t)) = 1 and γ′(t) ∈ Ce,α for every t.
Perhaps the most interesting modification of Definition 1.1 comes from
a so far unpublished result of M´ath´e and allows taking Borel sets G both in
(1.1) and (1.2). It is not exactly equivalent with ours, but has the properties
that a set of (e, β) width zero according to M´ath´e has (e, α) width zero
according to the above definition for every α > β, and a set of (e, α) width
zero according to the above definition has (e, α) width zero according to
M´ath´e. We have not used this, since our constructions, like that of [3], use
that width is determined by open sets, and so the only difference would
be that an appropriate version of Definition 1.1 would be called M´ath´e's
Theorem.
Terms like "cone null" have been used for sets that are defined with
the help of the notion of width. We follow this trend in our main notion,
introduced in Definition 1.6. Before coming to it, we recall the main starting
motivation behind what we do, namely the following definition from [2] and
a special case of their result (proved in [3]) which is most relevant for us.
Definition 1.2 (see [2, Definition 1.11]). A map τ of a subset E of Rn to
the Grassmanian G(k, n) is said to be a k-dimensional tangent field of E if
we,α{x ∈ E ∶ τ(x) ∩ Ce,α = {0}} = 0 for every e ∈ Rn and α > 0.
(1.3)
Obviously, if E is a k-dimensional embedded C 1 submanifold of Rn, its
tangent field τ(x) satisfies (1.3). However, the following theorem proved
3
in [2, 3] shows that many non-smooth sets admit a k-dimensional tangent
field. Before stating it, we notice that Definition 1.2 uses the value α in
two different meanings and so it is sensitive on the choice of the notion of
width. As a more detailed discussion of this will appear in [3], we just point
out that the use of M´ath´e's width and the width from Definition 1.1 are
equivalent. The only case to treat is when Definition 1.1 holds in M´ath´e's
sense. Assuming we,αk{x ∈ E ∶ τ(x) ∩ Ce,αk = {0}} = 0 in M´ath´e's sense for
all k ≥ 1, where 0 < αk < α < 1 and αk → α, we conclude that in the sense
of Definition 1.1 we have we,α{x ∈ E ∶ τ(x) ∩ Ce,αk = {0}} = 0 for all k ≥ 1.
Hence writing {x ∈ E ∶ τ(x) ∩ Ce,α = {0}} = ⋃∞
k=1{x ∈ E ∶ τ(x) ∩ Ce,αk = {0}},
we get we,α{x ∈ E ∶ τ(x) ∩ Ce,α = {0}} = 0.
Theorem 1.3 (see [2, Theorem 1.12]). A set E ⊂ Rn is contained in a non-
differentiability set of a Lipschitz map Rn → Rm for some, or any, m ≥ n
if and only if it admits an (n − 1)-dimensional tangent field. If n = 2, this
holds if and only of E is Lebesgue null.
As already mentioned, in the last assertion of Theorem 1.3 the assump-
tion n = 2 was removed in [8]. Notice also that the general case of Theo-
rem 1.3 says that the existence of k-dimensional tangent fields is similarly
related to the existence of functions that at every point of the set can be
differentiable in the direction of linear subspaces of dimension at most k
only.
Based on these results, we conjecture that sets of non-differentiability of
real-valued Lipschitz functions may be described as those for which there
is an (n − 1)-dimensional tangent field satisfying conditions that make it in
some sense closer to being "genuinely" (n − 1)-dimensional. We do not have
a more precise conjecture, but a simple consequence of our main result, The-
orem 1.9, is that sets for which there exists a continuous (n− 1)-dimensional
tangent field are indeed sets of non-differentiability of real-valued Lipschitz
functions.
Since for the real-valued case only the tangent fields of codimension one
are relevant, we base our approach on an obvious variant of Definition 1.2
that uses the normal fields instead of tangent fields. More interestingly,
having in mind conditions similar to continuity of the normal field, we can
define the normal vectors pointwise, while in general no pointwise definition
of tangent fields from Definition 1.2 is known. A highly interesting exception
to this is the special case when we are interested in measures supported by a
set admitting a k-dimensional tangent field where [4] provides an interesting
pointwise definition of the tangent field at almost every point.
4
Since our "normal vectors" are not exactly those orthogonal to the tan-
gent field from Definition 1.2, we do not actually call them "normal vectors"
and instead use just notation N (E, x) for their collection.
Definition 1.4. For x ∈ E ⊂ Rn let
N (E, x) ∶= {e ∈ Rn ∶ (∀ε > 0)(∃r > 0)we,ε(B(x, r) ∩ E) = 0}.
Remark 1.5. Although we will not use it directly, we remark that N (E, x)
is a linear subspace of Rn for any x ∈ E. This follows from results on
"joining cones" in [3], but we describe a quick argument based on the result
of M´ath´e. Since λN (E, x) = N (E, x) for each λ ∈ R, we conclude that every
nonzero e from the linear span of N (E, x) can be written as e = ∑k
i=1 ei
where ei ∈ N (E, x) ∖ {0}. Suppose ε > 0 is fixed and γ ∈ Γe,ε belongs to
C 1(R) and satisfies γ′(t) = 1 for all t ∈ R (cf. remarks after Definition 1.1).
2 εe~ ∑i ei and find δ > 0 such that wei,α(E ∩ B(x, δ)) = 0 for
Let α = 1
each i. From Definition 1.1 we see that there is a Borel (in fact Gδ) set
G ⊃ E such that wei,α(G ∩ B(x, δ)) = 0 for every i. Fix now any t ∈ R and
notice that there exists an i such that ⟨γ′(t), ei⟩ ≥ 2αei. By continuity of
γ′ there is a τ > 0 such that for this particular i we have ⟨γ′(s), ei⟩ > αei
whenever s ∈ (t − τ, t + τ). Hence wei,α(G ∩ B(x, δ)) = 0 for this i implies
γ−1(G ∩ B(x, δ)) ∩ (t − τ, t + τ) = 0. Finally, existence of such τ > 0 for each
t ∈ R allows us to conclude that γ−1(G ∩ B(x, δ)) = 0. As this holds for
every γ ∈ Γe,ε, we get we,ε(G ∩ B(x, δ)) = 0.
Definition 1.6. A set E ⊂ Rn satisfying N (E, x) ≠ {0} for every x ∈ E is
said to be cone unrectifiable.
Remark 1.7. Of course any cone unrectifiable set is Lebesgue null. A basic
example of cone unrectifiable sets E ⊂ Rn is provided by those for which
N (E, x) = Rn for every x ∈ E. Such sets are called uniformly purely unrec-
tifiable. By the result of Andr´as M´ath´e alluded to above these are precisely
those sets that are contained in a Borel 1-purely unrectifiable set, i.e., in a
Borel set B whose intersection with any C 1 curve has one-dimensional Haus-
dorff measure zero. The arguments used to prove Remark 1.5 simplify their
definition in another way: E is uniformly purely unrectifiable if and only if
there is 0 < η < 1 such that we,η(E) = 0 for every unit vector e (this is used
as a definition of a uniformly purely unrectifiable set in [9]). In Example 4.4
we point out that a similar simplification of the notion of cone unrectifiable
sets is false: given any e ≠ 0 and η ∈ (0, 1), we construct a set E which does
not satisfy the conclusions of the following Theorem 1.8 but is of Ce,η-width
zero.
5
We are now ready to state the main results of this paper.
Theorem 1.8. If E ⊂ Rn is cone unrectifiable, there is a Lipschitz function
f ∶ Rn → R that is non-differentiable at any point of E.
There are various ways of stating that non-differentiability of a function
f at a given point x is rather strong. The most usual one is by comparing
the upper and lower directional derivatives of f at x defined by
and
D+f(x; y) ∶= lim sup
t↘0
f(x + ty) − f(x)
t
D+f(x; y) ∶= lim inf
t↘0
f(x + ty) − f(x)
t
,
respectively. An even stronger non-differentiability statement is obtained
by showing that close to x, f may be approximated by many linear func-
tions. Our next result shows that the non-differentiability statement of
Theorem 1.8 may be strengthened in this way.
Theorem 1.9. For every cone unrectifiable set E ⊂ Rn and every ε > 0 there
are a Lipschitz function f ∶ Rn → R with Lip(f) ≤ 1 + ε and a continuous
function u ∶ E → {z ∈ Rn ∶ z ≤ ε} such that
lim inf
r↘0
sup
y≤r
f(x + y) − f(x) − ⟨e + u(x), y⟩
r
= 0
(1.4)
whenever x ∈ E and e ∈ N (E, x) has e ≤ 1. In particular,
D+f(x; y) − D+f(x; y) ≥ 2 sup⟨e, y⟩ ∶ e ∈ N (E, x), e ≤ 1
whenever x ∈ E and y ∈ Rn.
Additionally, if E is contained in {x ∶ ω(x) > 0}, where ω ∶ Rn → [0, ∞)
is continuous, then f may be chosen in such a way that f ≤ ω.
For a set E admitting an (n − 1)-dimensional continuous tangent we
obviously have N(E, x) ⊃ τ(x)⊥ ≠ {0}. Hence such sets are cone unrectifi-
able and so are sets of non-differentiability of a real valued Lipschitz func-
tion. More interestingly, having countably many cone unrectifiable sets, we
may try to add the functions obtained from Theorem 1.9 to get a function
non-differentiable at the points of the union. However, addition of non-
differentiable functions may create new points of differentiability. To solve
6
this problem we employ the idea that Zahorski [25] used in his precise de-
scription of non-differentiability sets of Lipschitz functions on the real line as
Gδσ-sets of measure zero; it is based on the simple observation that the sum
of a differentiable and a non-differentiable function is non-differentiable. For
this we need the function f obtained in Theorem 1.9 to be differentiable out-
side E, in other words, to have that E coincides with the set of points where
f is not differentiable. This is however not possible in general as shown
in Example 4.2. In the following two Corollaries we solve this difficulty by
making a special assumption that the sets we consider are Fσ.
Corollary 1.10. Suppose E = ⋃k Ek ⊂ Rn, where Ek are disjoint cone
unrectifiable Fσ sets, and let Nx ∶= N(Ek, x) ∩ B(0, 1) when x ∈ Ek. Then
there is a Lipschitz f ∶ Rn → R such that
• f is differentiable at every x ∈ Rn ∖ E;
• for every x ∈ E there is c > 0 such that for every y ∈ Rn,
D+f(x; y) − D+f(x; y) ≥ c sup
e∈Nx
⟨e, y⟩,
so, in particular, f is not differentiable at x.
If we are not interested in estimates of the difference of the upper and
lower derivatives, Corollary 1.10 gives the following more naturally sounding
statement.
Corollary 1.11. For any E ⊂ Rn that is a countable union of cone unrectifi-
able Fσ sets there is a Lipschitz function f ∶ Rn → R that is non-differentiable
at any point of E and is differentiable at any point of its complement Ec.
The next Corollary 1.12 contains the constructions of µ-almost every-
where non-differentiable functions from [3] and [4, Theorem 4.1]. Given a
Radon measure µ in Rn, these authors assign to µ-a.a. points a linear sub-
space T (x) of Rn that in certain sense represents the directions of curves on
which µ is "seen". For [3], the definition of "seen" is exactly the assumption
of Corollary 1.12 while [4] bases the definition on a related but different
property and shows in [4, Lemma 7.5] that the assumption of Corollary 1.12
is satisfied. It is, however, important to point out that both these references
define the linear space T (x) which is in a natural sense smallest, and this
allows them to obtain also a counterpart to Corollary 1.12 that every Lips-
chitz function is µ-a.e. differentiable in the direction of T (x). We also notice
that the constructions of µ-almost everywhere non-differentiable Lipschitz
7
functions have been strengthened in a different direction by [18] where the
authors produce functions that µ-a.e. admit any blow-up behaviour permit-
ted by the differentiability results.
Corollary 1.12. Let µ be a Radon measure on Rn and T a µ-measurable
m=0 G(n, m) such that for every unit vector e and 0 < α < 1,
map of Rn to ⋃n
the set {x ∶ Ce,α ∩ T (x) = {0}} is the union of a µ-null set and a set E
with we,α(E) = 0. Then there is a real valued Lipschitz function f on Rn
such that for µ-a.e. x ∈ Rn there is c > 0 such that D+f(x, v) − D+f(x, v) ≥
cdist(v, T (x)) for every v ∈ Rn.
Our final result deals with the uniformly purely unrectifiable sets intro-
duced in Remark 1.7. For such sets the statement of Theorem 1.9 concern-
ing upper and lower derivatives is shown in [3]. We prove a stronger result,
namely that these sets are non-universal differentiability sets in the strongest
possible sense, which corresponds to having ε = 0 in Theorem 1.9. However,
in Example 4.1 we demonstrate that such an improvement is specific to the
case of uniformly purely unrectifiable sets even when E ⊂ R2 is compact, for
each x ∈ E the set N (E, x) is a one-dimensional linear subspace of R2 and
the map x ↦ N (E, x) is continuous.
Theorem 1.13. For every uniformly purely unrectifiable set E ⊂ Rn there
is a real valued 1-Lipschitz function f on Rn such that
lim inf
r↘0
sup
y≤r
f(x + y) − f(x) − ⟨e, y⟩
r
= 0
(1.5)
for every x ∈ E and e ∈ Rn with e ≤ 1. In particular, D+f(x; y) = y and
D+f(x; y) = −y for every x ∈ E and y ∈ Rn.
2 Technical lemmas
We will work in the space Rn equipped with the Euclidean norm ⋅ . Most
of the notation we use is standard; the open and closed balls will be denoted
by B(x, r) and B(x, r), respectively. Since we will often need to use the
distance of a point to the complement of an open set, we will simplify the
notation for it: for an open G ⊂ Rn we let
ρG(x) ∶= dist(x, Rn ∖ G).
(2.1)
As usual, the Lipschitz constant of a real-valued function f defined on
a set E ⊂ Rn is the smallest constant Lip(f, E) ∈ [0, ∞], or just Lip(f)
8
when E = Rn, such that f(y) − f(x) ≤ Lip(f, E)y − x for all x, y ∈ E,
and functions with Lip(f) ≤ c will be termed c-Lipschitz. The space of
Lipschitz functions on Rn, those for which Lip(f) < ∞, will be denoted
by Lip(Rn). We will also use the pointwise Lipschitz constants defined by
Lipx(f) ∶= lim supy→x f(y) − f(x) ~y − x and often use the following well
known fact.
Lemma 2.1. For any f ∶ Rn → R, it holds that Lip(f) = supx∈Rn Lipx(f).
Proof. It suffices to handle the case n = 1 when it follows, for example, from
the considerably more general Theorem 4.5 in [24, Chapter IX].
The following lemma allows us to modify the functions we are construct-
ing so that they become smooth on suitable subsets of Rn. A similar lemma
is proved in [3], and in [4], although the authors of the latter paper could
have used more general [5, Theorem 1] or [16, Corollary 16]. We need a
slightly more technical variant of results from these references.
Recall that for any collection B of open sets in Rn there is a C ∞ partition
of unity of order n subordinated to it, that is a collection of C ∞ functions
ϕk ∶ Rn → [0, 1], k = 1, 2, . . . , such that
• each spt(ϕk) is a compact subset of some B ∈ B,
• ∑k ϕk(x) = 1 for every x ∈ ⋃B,
• for each x ∈ ⋃B there is r > 0 such that B(x, r) ∩ spt(ϕk) ≠ ∅ for at most
n + 1 values of k.
Lemma 2.2. Suppose H ⊂ Rn is open, g ∶ Rn → R is Lipschitz, Φ ∶ H → Rn
and ξ ∶ H → [0, ∞) are continuous and bounded, and g′(x) − Φ(x) ≤ ξ(x)
for almost every x ∈ H. Then for every continuous ω ∶ H → (0, ∞) there is
a Lipschitz function f ∶ Rn → R such that
(i) f(x) = g(x) for x ∉ H ∩{ξ > 0} and f(x) − g(x) ≤ ω(x) for x ∈ H;
(ii) f ∈ C 1(H) and f ′(x) − Φ(x) ≤ ξ(x)(1 + ω(x)) for x ∈ H;
(iii) Lip(f) ≤ max(Lip(g), supx∈H(Φ(x) + ξ(x)(1 + ω(x)))).
Proof. Let U ∶= H ∩ {ξ > 0}, extend ξ and ω to (possibly discontinuous)
functions defined on all of Rn by letting ξ(x) = ω(x) = 0 for x ∉ H and
U(x)). Let B be the family of balls
let ω0(x) ∶= 1
B(x, r) such that x ∈ U and r < ω0(x). Choose (ϕk)k≥1 forming a locally
2 min(1, ξ(x)ω(x), ω(x), ρ2
9
finite C ∞ partition of unity on U subordinate to B, and denote mk = 1 +
ϕ′
k∞.
As, for example, in [14, Appendix C.4], let η be the standard C ∞-smooth
mollifier in Rn and define ηs(x) ∶= η(x~s)~sn. For each k choose sk > 0 small
enough so that the convolution fk = g ∗ ηsk satisfies for every x ∈ spt(ϕk),
• fk(x) − g(x) ≤ 2−k−1m−1
• f ′
k ω0(x);
k(x) − Φ(x) ≤ ξ(x) + ω0(x).
Define f ∶ Rn → R by f(x) = ∑k fk(x)ϕk(x) for x ∈ U and f(x) = g(x)
for x ∉ U . Since each fkϕk is in C 1(U), we have f ∈ C 1(U). Also, for all
x ∈ Rn,
(2.2)
(2.3)
f(x) − g(x) ≤ ω0(x)
since for x ∉ U both sides are zero, and for x ∈ U ,
k fk(x) − g(x) ϕk(x) ≤ Q
Since ω0 ≤ ω and ω0(x) = 0 for x ∉ U , (i) holds.
f(x) − g(x) ≤ Q
k
We show that f is differentiable at every x ∈ H and
f ′(x) − Φ(x) ≤ ξ(x) + 2ω0(x).
To see this for x ∈ U , we use ∑k ϕk(x) = 1 and ∑k ϕ′
ω0(x)ϕk(x) ≤ ω0(x).
k(x) = 0 to infer that
f ′(x) − Φ(x) = Q
k (f ′
k(x) − Φ(x))ϕk(x) + Q
k (fk(x) − g(x))ϕ′
k(x),
hence
f ′(x) − Φ(x) ≤ Q
k(x) − Φ(x)ϕk(x) + Q
k f ′
k (ξ(x) + ω0(x))ϕk(x) + Q
≤ Q
≤ ξ(x) + 2ω0(x).
k
k fk(x) − g(x) ϕ′
2−k−1ω0(x)
k(x)
To see (2.3) for x ∈ H ∖ U , we infer from the assumptions on g, Φ and ξ that
g is differentiable at x and g′(x) = Φ(x). Since (f − g)′(x) = 0 because (2.2)
gives f(y) − g(y) ≤ ω0(y) ≤ ρ2
U(y) ≤ y − x2 for all y ∈ Rn, we get that f is
differentiable at x and f ′(x) = g′(x) = Φ(x).
Clearly, (2.3) and the inequality 2ω0(x) ≤ ξ(x)ω(x) show the second
statement of (ii).
10
To prove (iii), we infer from (2.2) that Lipx(f) ≤ Lip(g) for x ∈ Rn ∖ U ,
and from (2.3) that
Lipx(f) ≤ sup
y∈U(Φ(y) + ξ(y) + 2ω0(y)) ≤ sup
y∈H(Φ(y) + ξ(y)(1 + ω(y)))
for x ∈ U . Thus (iii) holds by Lemma 2.1 and, since its right side is finite,
we also see that f is Lipschitz.
We already know that f is differentiable at every y ∈ H and f ′ is contin-
uous at every y ∈ U . If y ∈ H ∖ U , (2.3) shows that limx→y(f ′(x) − Φ(x)) = 0.
Since Φ is continuous at y, it follows that f ′ is continuous at y. Hence
f ∈ C 1(H), which is the last statement we needed to prove.
The next simple Lemma is used to show that the functions we construct
may be approximated by linear ones in the way required in equation (1.4)
of our main result, Theorem 1.9.
Lemma 2.3. Suppose that H ⊂ Rn is open, g ∶ Rn → R belongs to C 1(H),
ω ∶ Rn → [0, ∞) is continuous and strictly positive on H, and η ∈ (0, 1].
Then there is a function ξ ∶ Rn → [0, ∞) such that
(i) ξ ∈ C(Rn,[0, ∞)) ∩ C(H,(0, ∞)) and ξ ≤ 1
2 ω;
(ii) if x ∈ H and h ∶ Rn → R satisfies h − g ≤ 2ξ, there is 0 < r < ω(x) such
that h(x + y) − h(x) − ⟨g′(x), y⟩ ≤ ηr whenever y ≤ r.
2 min(ρH , ω, 1), and g′(y) − g′(z) ≤ 1
Proof. Let Ψ be the set of functions ψ ∶ Rn → [0, ∞) satisfying Lip(ψ) ≤ 1,
2 η whenever x ∈ H and
0 ≤ ψ ≤ 1
max(y − x,z − x) < ψ(x). Since 0 ∈ Ψ, ϕ(x) ∶= sup{ψ(x) ∶ ψ ∈ Ψ} is
well-defined. We also have ϕ ∈ Ψ since for any x, y, z satisfying x ∈ H and
max(y − x,z − x) < ϕ(x) there is ψ ∈ Ψ such that max(y − x,z − x) <
ψ(x) and hence g′(y) − g′(z) ≤ 1
2 η.
Let x ∈ H. Since both ρH and ω are continuous and strictly positive
2 min(ρH , ω, 1) > ε on B(x, ε). Then the
at x, there is ε > 0 such that 1
function ψε,x(y) ∶= max(0, ε −y − x) satisfies ψε,x = 0 outside B(x, ε) and
2 min(ρH(y), ω(y), 1) for y ∈ B(x, ε). Hence ψε,x belongs
0 ≤ ψε,x(y) ≤ ε ≤ 1
to Ψ and we infer that ϕ(x) ≥ ψε,x(x) = ε > 0. Consequently, ϕ is strictly
positive on H. Furthermore,
g(x + y) − g(x) − ⟨g′(x), y⟩ ≤ y
sup
z∈B(x,y)
g′(z) − g′(x) ≤ 1
2 ηy
whenever x ∈ H and y < ϕ(x).
11
Letting ξ(x) ∶= 1
we let r ∶= ϕ(x), observe that 0 < r < ω(x) and use that Lip(ξ) ≤ 1
ξ(x) = 1
12 ηϕ(x), we see that (i) holds. To prove (ii), given x ∈ H,
12 η and
12 ηr to estimate
h(x + y) − h(x) − ⟨g′(x), y⟩
≤ 2ξ(x + y) + 2ξ(x) + g(x + y) − g(x) − ⟨g′(x), y⟩
≤ 4ξ(x) + 2Lip(ξ)y + 1
≤ 1
2 ηy
2 ηy ≤ ηr
3 ηr + 1
6 ηy + 1
whenever y < r = ϕ(x), and so whenever y ≤ ξ(x).
The following Lemmas 2.4 and 2.6 modify corresponding lemmas from [3]
in a way suitable for our applications. A special version of Lemma 2.4, which
does not suffice for our purposes, can be found also in [4, Lemmas 4.12 -- 4.14].
Since [3] is not yet available, we provide full proofs.
Lemma 2.4. Given ε > 0 there is ϑ ∈ (0, 1) such that the following holds.
For every E ⊂ Rn, every unit vector e ∈ Rn such that we,ϑ(E) = 0 and
every continuous ω ∶ Rn → [0, ∞) which is strictly positive on E, there is a
Lipschitz function g ∶ Rn → R such that 0 ≤ g ≤ ω, Lip(g) ≤ 1 + ε and there
is an open set H ⊃ E contained in {ω > 0} such that g′(x) − e ≤ ε for
Lebesgue almost all x ∈ H.
Proof. Let ϑ = sin β, where 0 < β < π~2, be such that tan β < ε~2. Denote
G ∶= {x ∶ ω(x) > 0} and choose ϕk ∈ C ∞(Rn), k ≥ 1, with compact support
contained in G that form a locally finite partition of unity on G. Let εk > 0
G(x), ω(x)) for each
be such that ∑k εkϕ′
k ≥ 1 and all x ∈ Rn.
Using values εk which we have just defined, find open sets Gk such that
k < ε~2 and εkϕk(x) ≤ 2−k min(1, ρ2
G ⊃ Gk ⊃ E and we,ϑ(Gk) < εk. For each x ∈ Rn we put
gk(x) ∶= supH1Gk ∩ γ(−∞, b] − s ∶ γ ∈ Γe,ϑ, s ≥ 0, γ(b) = x + se
(2.4)
and show that
(i) 0 ≤ gk(x) ≤ εk;
(ii) gk(x + y) − gk(x) ≤ y tan β when y is perpendicular to e;
(iii) gk(x) ≤ gk(x + re) ≤ gk(x) + r for every r > 0;
(iv) gk(x + re) = gk(x) + r when [x, x + re] ⊂ Gk;
12
(v) gk is a Lipschitz function and Lip(gk) ≤ 1 + tan β;
(vi) g′
k(x) − e ≤ tan β for almost every x ∈ Gk.
The first inequality in (i) is obvious by considering in (2.4), s = 0 and
any γ ∈ Γe,ϑ with γ(b) = x, and the second is immediate from we,ϑ(Gk) < εk.
If y ≠ 0 is orthogonal to e, and γ, b, s come from (2.4), we let r ∶= y
and y ∶= y~r and redefine γ on (b, ∞) by γ(b + t) = γ(b) +(t cot β)y + te for
t > 0. Using (2.4) for gk(x + y) with b′ ∶= b + r tan β and s′ ∶= s + r tan β, we
gk(x + y) ≥ gk(x) − r tan β = gk(x) −y tan β.
To get a lower estimate for gk(x) apply the above to the vector −y added to
get
x + y:
gk(x) = gk(x + y − y) ≥ gk(x + y) −y tan β.
This verifies (ii).
Now consider x′ = x + re where r > 0. Since any γ used for x′ may be
used for x with γ(b) = x +(r + s)e, we get gk(x) ≥ gk(x′) − r. For the rest of
letting γ(b + t) = x + se + te for t ≥ 0, we get
(iii) and for (iv), note that as any γ used in (2.4) for x may be redefined by
gk(x′) ≥ H1(Gk ∩ γ(−∞, b + r]) − s ≥ H1(Gk ∩ γ(−∞, b]) − s
for all γ satisfying (2.4), so gk(x′) ≥ gk(x), and this verifies (iii). If [x, x′] =
[x, x + re] ⊂ Gk and r ≤ s, the same argument shows that
gk(x′) ≥ H1(Gk ∩ γ(−∞, b + s]) −(s − r) ≥ H1(Gk ∩ γ(−∞, b]) − s + r,
and if r > s, then
gk(x′) ≥ H1(Gk ∩ γ(−∞, b + r]) = H1(Gk ∩ γ(−∞, b + s]) +(r − s)
≥ H1(Gk ∩ γ(−∞, b]) − s + r
(iii), implies equality in (iv).
for all such γ. Hence in both cases gk(x′) ≥ gk(x) + r, which, together with
The statements (ii) -- (iv) imply that gk is Lipschitz and for almost every x,
0 ≤ Dgk(x; e) ≤ 1, the equality Dgk(x; e) = 1 is satisfied for x ∈ Gk and
Dgk(x; y) ≤ y tan β for y perpendicular to e. This gives both (v) and (vi).
G, ω) for
every k ≥ 1, we conclude that 0 ≤ g ≤ ω and Lipx(g) = 0 for x ∉ G. Since the
gkϕk. Since by (i) one has 0 ≤ gkϕk ≤ 2−k min(1, ρ2
Let g ∶=
∞
Q
k=1
13
k g′
gk(x)ϕ′
k
k(x)ϕk(x) + Q
k(x) ≤ 1 + tan β + Q
sum defining g is locally finite, g is locally Lipschitz on G and by (v) and
(i) for almost every x ∈ G,
g′(x) ≤ Q
k ≤ 1 + ε.
Hence Lipx(g) ≤ 1 + ε for every x ∈ G, and we infer from Lemma 2.1 that
Lip(g) ≤ 1 + ε.
Let H ∶= ⋂k Uk, where Uk ∶= (G ∖ spt(ϕk)) ∪ Gk are open. Then E ⊂
⋂k Gk ⊂ H ⊂ G and H is open because the complements of the Uk in G are
closed in G and their collection is locally finite in G since G ∖ Uk ⊂ spt(ϕk).
Finally, by (vi) for almost every x ∈ H,
g′(x)−e ≤ Q
k(x)−eϕk(x)+Q
gk(x)ϕ′
k(x) ≤ tan β +Q
εkϕ′
k < ε.
εkϕ′
k
k g′
k
k
Definition 2.5. Since we will need to use Lemma 2.4 for several values of
ε at the same time, we introduce a function ϑ ∶ (0, ∞) → (0, ∞) such that
ϑ(σ) is the value of ϑ from Lemma 2.4 for ε = 1
7 σ.
Lemma 2.6. Suppose E ⊂ Rn, the functions ω ∶ Rn → [0, ∞) and ϕ ∶ Rn →
[0, 1] are continuous, ω > 0 on E, e ∈ Rn, σ > 0 and we,ϑ(σ)(E ∩{ϕ > 0}) = 0.
Then there exist functions f ∶ Rn → R and ψ ∶ Rn → [0, 1] and an open set
H ⊂ Rn such that
(i) E ⊂ H ⊂ {x ∶ ω(x) > 0} and f ∈ Lip(Rn) ∩ C 1(H);
(ii) f(x) ≤ ω(x)e for all x ∈ Rn and f(x) = 0 when ϕ(x) = 0;
(iii) f ′(x) − ψ(x)e ≤ σ1{ω>0}(x)1{ϕ>0}(x)e for almost all x ∈ Rn;
(iv) 0 ≤ ψ(x) ≤ ϕ(x)1{ω>0}(x) for x ∈ Rn and ψ(x) = ϕ(x) for x ∈ H.
Proof. If e = 0 or σ ≥ 1, it suffices to let f ∶= 0, ψ ∶= ϕ and H ∶= {ω > 0}. So
we assume e = 1 and σ < 1, let ε ∶= σ~7 and pick an integer k ∈ [6~σ, 7~σ].
2 min(1, ω), G0 ∶= {ω > 0}, H0 ∶= G0 ∩ {ϕ > 0} and, whenever
Hi−1 has been defined for some i = 1, . . . , k, let Gi ∶= Hi−1 ∩ {ϕ > i~k} and
Gi), where ρGi is defined
use Lemma 2.4 with continuous ωi(x) = 1
by (2.1), to find a Lipschitz function gi ∶ Rn → R and nested open sets
Hi ⊂ Gi ⊂ Hi−1 such that for each 1 ≤ i ≤ k,
2 min(ω, ρ2
Gi);
(a) Lip(gi) ≤ 1 + ε and gi ≤ 1
(b) Gi ⊃ Hi ⊃ Gi ∩ E and g′
i(x) − e ≤ ε for a.e. x ∈ Hi.
2 min(ω, ρ2
Let ω0 ∶= 1
14
k ∑k
Let g ∶= 1
i=1 gi. Then by (a), Lip(g) ≤ 1+ε and g ≤ 1
G1). For
any x ∈ G0 find the biggest j = j(x) ∈ {0, 1, . . . , k} with x ∈ Gj; since Gk = ∅,
we have j(x) ≤ k − 1. Define ψ(x) = min((j(x) + 2)~k, ϕ(x)); and for x ∉ G0
let ψ(x) = 0. Clearly, 0 ≤ ψ ≤ ϕ1G0 on Rn, which is the first statement of (iv).
(cid:6), i.e. 0 < ψ(x) − j(x)~k ≤ 2~k.
For any x ∈ H0 it holds ψ(x) ∈ j(x)
2 min(ω, ρ2
k , j(x)+2
k
Define now
H =
k
j=0{x ∈ Hj ∶ ϕ(x) < (j + 2)~k}
k
and notice that ψ(x) = ϕ(x) whenever x ∈ H. Indeed, if x ∈ Hj is such that
k > ϕ(x),
ϕ(x) < (j + 2)~k, then Hj ⊂ Gj implies j(x) ≥ j, so j(x)+2
hence by definition ψ(x) = ϕ(x), and this verifies (iv). Also, E ⊂ H since
E ⊂ ⋃k−1
j=0(Hj ∖ Gj+1) from (b), and for x ∈ Hj ∖ Gj+1 we have j(x) = j and
so ϕ(x) ≤ (j + 1)~k < (j + 2)~k. Since it is clear that H is open and ω > 0 on
H because H ⊂ G0, we conclude that the first part of (i) is satisfied for E,
H and ω. We are now left to define the Lipschitz function f and verify the
remaining part of (i), and also (ii) and (iii).
≥ j+2
Note that for almost all x ∈ G1 (where ϕ > 1~k), all gi are differentiable
at x and the estimate in (b) is satisfied whenever x ∈ Hi and 1 ≤ i ≤ k.
Consider any such x ∈ G1. To estimate g′(x), notice that for such x we have
j = j(x) ≥ 1 and
• if 1 ≤ i < j, then x ∈ Hi and so g′
• if i ≥ j + 1, then x ∉ Gi and so g′
i(x) − e ≤ ε by (b);
i(x) = 0 by (a).
Hence, for almost all x ∈ G1
g′(x) − ψ(x)e ≤ g′(x) − j−1
k e + 3
ke
i(x) − e +g′
k ≤ 5ϕ(x).
k(
≤ 1
≤ ε + 4
j−1
i=1 g′
Q
k ≤ 5
j(x)) + 3
k
k , ϕ(x)) for almost all x ∈ Rn.
Since g′(x) = 0 outside G1, we get g′(x) − ψ(x)e = ψ(x) for x ∉ G1.
Using that ψ = 0 outside H0 and ψ ≤ ϕ ≤ 1
k for x ∈ H0 ∖ G1 we infer that
k , ϕ(x)) outside G1 and conclude g′(x) − ψ(x)e ≤
g′(x) − ψ(x)e ≤ min( 1
5 min( 1
k , ϕ(x)) and ω(x) =
5 min(ω(x), ϕ(x), ρ2
H), we find Lipschitz f ∶ Rn → R such that f ∈ C 1(H),
f(x) −g(x) ≤ ω(x) and f ′(x) −ψ(x)e ≤ ξ(x)(1 + ω(x)) for all x ∈ H. Since
f ∈ C 1(H), the remaining condition of (i) is satisfied. Finally, the conditions
Using Lemma 2.2 with Φ(x) ∶= ψ(x)e, ξ(x) ∶= 5 min( 1
1
15
(ii) and (iii) hold since f ≤ f −g + g ≤ 1
and f ′(x) − ψ(x)e ≤ 6 min( 1
outside G0 = {ω > 0}.
G1) ≤ ω1{ϕ>0},
k , ϕ(x)) ≤ σ1{ϕ>0}(x) and f ′ = g′ = 0 and ψ = 0
5 min(ω, ϕ)+ 1
2 min(ω, ρ2
In a rather straightforward way, we will use Lemma 2.6 recursively to
obtain the main tool for our construction of a function non-differentiable at
points of a given set E.
Lemma 2.7. Suppose E ⊂ H0 ⊂ Rn, H0 is open, f0 ∈ Lip(Rn) ∩ C 1(H0) and
ω0 ∈ C(Rn,[0, ∞)) ∩ C(H0,(0, ∞)). Suppose further that for k ≥ 1 we are
given vectors ek ∈ B(0, 1), functions ϕk ∈ C(Rn,[0, 1]) and σk > 0 such that
wek,ϑ(σk)(E ∩{ϕk > 0}) = 0. Then for each j ≥ 1 there are sets Hj ⊂ Rn and
functions fj, ωj , ψj ∶ Rn → R such that
(i) Hj is open, E ⊂ Hj ⊂ Hj−1 and fj ∈ Lip(Rn) ∩ C 1(Hj);
(ii) ωj ∈ C(Rn,[0, ∞)) ∩ C(Hj,(0, ∞)) and ωj ≤ 1
(iii) fj − fj−1 ≤ ωj−1 and fj(x) = fj−1(x) when ϕj(x) = 0;
(iv) if h ∶ Rn → R and h − fj ≤ 2ωj then for every x ∈ Hj one may find
j(x), y⟩ ≤ σjr;
2 min(1, ωj−1, ρ2
Hj);
0 < r < ωj−1(x) such that supy≤r h(x + y) − h(x) − ⟨f ′
(v) ψj ∶ Rn → [0, 1], 0 ≤ ψj ≤ ϕj 1Hj−1 and ψj = ϕj on Hj;
(vi) f ′
(vii) f ′
j−1(x) − ψj(x)ej ≤ σj 1{ϕj >0}(x) for every x ∈ E;
j(x) − f ′
j(x)−z ≤ f ′
and a.e. x ∈ Rn.
i=1 ψi(x)ei −z+∑j
0(x)+∑j
i=1 σi1{ϕi>0}(x) for any z ∈ Rn
2 min(1, ω0, ρ2
H0) if necessary, we may and will as-
Proof. Replacing ω0 by 1
H0) and observe that then H0 = {ω0 > 0}. Assume
2 min(1, ρ2
sume that ω0 ≤ 1
j ≥ 1 and an open set Hj−1 ⊃ E, a function fj−1 ∈ Lip(Rn) ∩ C 1(Hj−1), and
a function ωj−1 ∈ C(Rn,[0, ∞)) ∩ C(E,(0, ∞)) such that Hj−1 = {ωj−1 > 0},
have been already defined; this is certainly the case for j = 1. We will
now explain how to construct functions fj, ωj , ψj and sets Hj such that
conditions (i) -- (vi) of the present lemma are satisfied. Notice that once
we construct these objects, we have an open set Hj ⊃ E and a function fj ∈
Lip(Rn) ∩ C 1(Hj) from (i), and a function ωj ∈ C(Rn,[0, ∞)) ∩ C(E,(0, ∞))
satisfying ωj ≤ min(1, ρ2
Hj) for all x ∈ Rn from (ii). This will allow us re-
cursively to construct all required objects so that (i) -- (vi) hold, and then we
will finish the proof by showing that (vii) holds as well.
By Lemma 2.6 find gj , ψj and Hj ⊂ Rn such that
16
(a) Hj is open, E ⊂ Hj ⊂ Hj−1 and gj ∈ Lip(Rn) ∩ C 1(Hj);
(b) gj(x) ≤ ωj−1(x)ej for all x ∈ Rn and gj(x) = 0 when ϕj(x) = 0;
(c) g′
j(x) − ψj(x)ej ≤ σj 1Hj−1(x)1{ϕj >0}(x)ej for almost all x ∈ Rn;
(d) 0 ≤ ψj(x) ≤ ϕj(x)1Hj−1(x) for x ∈ Rn and ψj(x) = ϕj(x) for x ∈ Hj.
Here we used that Hj−1 = {ωj−1 > 0} to obtain conditions (a) -- (d) directly
Let fj ∶= fj−1 + gj, then (a) and (b) imply (i) and (iii), respectively.
By Lemma 2.3 we may find ξj ∈ C(Rn,[0, ∞)) ∩ C(Hj,(0, ∞)) having the
property that whenever x ∈ Hj and h ∶ Rn → R satisfies h − fj ≤ ξj, there is
0 < r < ωj−1(x) such that h(x + y) − h(x) −⟨f ′
j (x), y⟩ ≤ ηj r whenever y ≤ r.
Letting ωj ∶= 1
Hj ), we have (ii) and (iv). Clearly, (v) is the
same as (d), and (c) implies that
from conditions (i) -- (iv) of Lemma 2.6.
2 min(ωj , ξj , ρ2
f ′
j(x) − f ′
j−1(x) − ψj(x)ej ≤ σj 1{ϕj>0}(x)
(2.5)
for almost every x ∈ Rn. From this, since f ′
on the open set Hj ⊃ E, we have (vi).
j, f ′
j−1 and ψj = ϕj are continuous
By the recursive use of the above construction we have defined Hj , fj, ωj
and ψj such that (i) -- (vi) hold. The last required statement (vii) follows by
using (2.5) to estimate, for almost every x ∈ Rn,
f ′
j(x) − z ≤ f ′
0(x) +
≤ f ′
0(x) +
j
Q
i=1
j
Q
i=1
ψi(x)ei − z +
ψi(x)ei − z +
j
Q
i=1
j
Q
i=1
f ′
i(x) − f ′
i−1(x) − ψi(x)ei
σi1{ϕi>0}(x).
We will use Lemma 2.7 to prove the two key results, Theorem 1.9 and
Theorem 1.13. To prove the former, we will choose the objects required in
Lemma 2.7 using the following combination of suitable partitions of unity.
Lemma 2.8. Suppose E ⊂ Rn is cone unrectifiable and ε > 0. Then there
exist sequences of positive numbers σl > 0, vectors el ∈ B(0, 1) and continuous
functions ϕl ∶ Rn → [0, 1], such that
(i) ∑l≥1 σl1spt(ϕl) ≤ ε;
(ii) wel,ϑ(σl)(E ∩{ϕl > 0}) = 0 for each l ≥ 1;
(iii) if x ∈ E, e ∈ N(E, x) and e ≤ 1, then for every η > 0 there are
arbitrarily large l such that σl < η, e − el < η and ϕl(x) = 1.
17
Proof. For x ∈ E, e ∈ N(E, x) and any σ > 0 there exists, by definition of the
cone unrectifiable set, a radius δ(x, e, σ) > 0 such that we,ϑ(σ)(E ∩Bx,e,σ) = 0,
where Bx,e,σ = B(x, δ(x, e, σ)).
We may suppose ε = 1~p for some p ∈ N (so that 1~ε is a positive integer).
(n+1)−1. For each pair of i ≥ 1
For each i ≥ 1 we let εi ∶= 2−iε and τi ∶= 3−nεn+1
choose ei,j ∈ B(0, 1) such that B(0, 1) ⊂ ⋃j B(ei,j, εi)
and j = 1, . . . , 3nε−n
for every fixed i ≥ 1. Let
i
i
Ei,j ∶= {x ∈ E ∶ (∃e ∈ N(E, x))e − ei,j < εi},
so that of course ⋃j Ei0,j = E for each fixed i0 ≥ 1. For each pair (i0, j0) find
a partition of unity {ϕi0,j0,k ∶ k ≥ 1} of order n subordinated to
{By,u,σ ∶ y ∈ Ei0,j0, u ∈ N(E, y), u − ei0,j0 < εi0 , σ = τi0}.
∶= min1,(n + 1)ϕi(l),j(l),k(l) and σl
Order the triples (i, j, k) into a single sequence (i(l), j(l), k(l)), and
∶= τi(l). Also, observing that
let ϕl
spt(ϕl) = spt(ϕi(l),j(l),k(l)), find yl ∈ Ei(l),j(l) and el ∈ N(E, yl) such that
spt(ϕl) ⊂ Byl,el,σl. Notice for future reference that el − ei(l),j(l) < εi(l).
We show that the Lemma holds with the σl, el and ϕl defined above.
To prove (i), observe that for each fixed i0 ≥ 1 and x0 ∈ Rn there are at
i0 (n + 1) pairs (j, k) for which x0 ∈ spt(ϕi0,j,k). Notice also that σl
most 3nε−n
is constant and equal τi0 over all l with the same value of i(l) = i0. Hence
Q
l
σl1spt(ϕl)(x0) ≤ Q
i
3nε−n
i (n + 1)τi ≤ Q
i
εi ≤ ε.
The statement (ii) is immediate from wel,ϑ(σl)(E ∩ Byl,el,σl) = 0 and the
inclusion spt(ϕl) ⊂ Byl,el,σl.
Finally, suppose x ∈ E, e ∈ N(E, x), e ≤ 1, η > 0 and l0 ∈ N. Let
i0 > max{i(l); l ≤ l0} be such that εi0 < η~2. For any i > i0 there is j such
that e − ei,j < εi < εi0 < η~2. Then x ∈ Ei,j and since the partition of unity
{ϕi,j,k ∶ k ≥ 1} is of order n, there is k such that ϕi,j,k(x) ≥ 1~(n + 1). This
implies ϕl(x) = 1 for l satisfying (i, j, k) = (i(l), j(l), k(l)). Then l > l0 and
σl = τi < εi, so e − el ≤ e − ei,j + ei,j − el < 2εi < η, so (iii) holds as
well.
Our second use of Lemma 2.7, to prove Theorem 1.13, will be more
straightforward: we use it to construct functions that will approximate the
required function.
Lemma 2.9. Suppose E ⊂ H ⊂ Rn, E is uniformly purely unrectifiable, H is
open, ω ∈ C(Rn,[0, ∞)) ∩ C(H,(0, ∞)) and f ∈ Lip(Rn) ∩ C 1(H). Then for
18
every e ∈ Rn and η > 0 there are g, ξ ∶ Rn → R and an open set U ⊂ Rn such
that
(i) E ⊂ U ⊂ H, ξ ∈ C(Rn,[0, ∞)) ∩ C(U,(0, ∞)) and ξ ≤ 1
2 ω;
(ii) g − f ≤ ω, Lip(g) ≤ max(Lip(f),e) + η and g ∈ C 1(U);
(iii) if x ∈ E and a function h ∶ Rn → R satisfies h − g ≤ 2ξ, there is
0 < r < ω(x) such that supy≤r h(x + y) − h(x) − ⟨e, y⟩ ≤ ηr.
Proof. Let σ = η~8(n + 1). Since f ∈ C 1(H) and E ⊂ H, for each x ∈ E there
is δx > 0 such that f ′(y) − f ′(z) < 1
4 η for y, z ∈ Bx ∶= B(x, δx). Find a
partition of unity {γk ∶ k ≥ 1} of order n subordinated to {Bx ∶ x ∈ E} and
choose xk ∈ E such that spt(γk) ⊂ Bxk .
2 ω, f0 = f , σk = σ, e2k−1 = −f ′(xk) ∈ N (E, xk),
e2k = e ∈ N (E, xk), and ϕ2k−1 = ϕ2k = γk. Since E is uniformly purely
unrectifiable, the hypothesis of Lemma 2.7 is satisfied, and so find fk, ωk,
Hk and ψk, k ≥ 1, such that the statements (i) -- (vii) of Lemma 2.7 hold (we
leave out (iv) and (vi) as we do not use them here):
Set H0 = H, ω0 = 1
(a) Hk is open, E ⊂ Hk ⊂ Hk−1 and fk ∈ Lip(Rn) ∩ C 1(Hk);
(b) ωk ∈ C(Rn,[0, ∞)) ∩ C(Hk,(0, ∞)) and ωk ≤ 1
(c) fk − fk−1 ≤ ωk−1 and fk(x) = fk−1(x) when ϕk(x) = 0;
(d) ψk ∶ Rn → [0, 1], 0 ≤ ψk ≤ ϕk 1Hk−1 and ψk = ϕk on Hk;
(e) f ′
i=1 ψi(x)ei − z + ∑k
k(x) − z ≤ f ′(x) + ∑k
and a.e. x ∈ Rn.
2 min(1, ωk−1, ρ2
Hk);
i=1 σ1{ϕi>0}(x) for all z ∈ Rn
By (b) and (c), the sequence of Lipschitz functions (fk) converges to a
function g ∶ Rn → R and g − f ≤ ω. For every x at which f ′(x) exists write
f ′(x) +
2k
Q
i=1
ψi(x)ei = af ′(x) + be + v,
i=1 ψ2i(x), v = ∑k
(2.6)
where a = 1−∑k
Using ∑i γi ≤ 1 as it is a partition of unity, and (d) to get
i=1 ψ2i−1(x), b = ∑k
i=1 ψ2i−1(x)(f ′(x)−f ′(xi)).
0 ≤ ψ2i ≤ ϕ2i1H2i−1 = ϕ2i−11H2i−1 ≤ ψ2i−1 ≤ ϕ2i−1 = γi,
(2.7)
i=1(ψ2i(x) − ψ2i−1)(x) ≤ 1, and v ≤
we see that a, b ≥ 0, a + b = 1 + ∑k
∑i∶x∈spt(γi) γi(x)f ′(x) − f ′(xi). Recall that spt(γi) ⊂ Bxi, and by the def-
4 η for x ∈ Bxi, hence
inition of the ball Bxi we have f ′(x) − f ′(xi) < 1
19
v < 1
k ≥ 1
4 η. Thus we conclude from (2.6) that for almost all x ∈ Rn and all
f ′(x) +
2k
Q
i=1
ψi(x)ei ≤ max(Lip(f),e) + η~4.
(2.8)
Since for every x there are at most 2(n + 1) values of i with ϕi(x) ≠ 0, we
see that ∑2k
4 η for any k ≥ 1, and infer from (e)
with z = 0 and (2.8) that for a.e. x,
i=1 σ1{ϕi>0}(x) ≤ 2(n + 1)σ = 1
2k
Q
i=1
2k
Q
i=1
ψi(x)ei +
for each k, and so (ii) holds.
σ1{ϕi>0}(x) ≤ max(Lip(f),e) + 1
2 η.
2k(x) ≤ f ′(x) +
f ′
Since, by (a), f2k is Lipschitz, we conclude Lip(f2k) < max(Lip(f),e) + η
For each x ∈ E there is a neighbourhood where all but a finite number
of the functions ϕk's are zero, so we can find rx > 0 and kx ∈ N such that
B(x, rx) ∩ sptϕk = ∅ for k ≥ kx. Let Ux ∶= B(x, rx) ∩ Hkx, where Hkx ⊃ E ∋ x
is defined in (a), and define an open set U ∶= ⋃x∈E Ux. As x ∈ Ux ⊂ Hkx ⊂
H0 = H for any x ∈ E, we conclude that E ⊂ U ⊂ H, this verifies the first two
statements of (i). By (c), g = fk on B(x, rx) ⊃ Ux for every k ≥ kx; hence
g ∈ C 1(Ux) by (a) as Ux ⊂ Hkx, and so g ∈ C 1(U). Thus Lemma 2.3 applied
2 η provides a continuous function ξ ∶ Rn → [0, ∞) such that
to U, g, ω and 1
(i) holds and for every x ∈ E ⊂ U and h ∶ Rn → R satisfying h − g ≤ 2ξ, there
is 0 < r < ω(x) such that
sup
y≤r h(x + y) − h(x) − ⟨g′(x), y⟩ ≤ 1
2 ηr.
(2.9)
i=1 ψ2i−1(x) = ∑k
Observe now that for x ∈ E we have x ∈ Hi for any i ≥ 1, hence ψi(x) =
ϕi(x) for any i ≥ 1 by (d). Together with definition of kx this implies
that ∑k
i=1 γi(x) = ∑i≥1 γi(x) = 1 for any
k ≥ kx, hence for such k the constants a, b from (2.6) satisfy a = 0 and,
similarly, b = 1. Using equation (2.6) and recalling that v ≤ 1
4 η, we get
f ′(x) + ∑2k
4 η for any k ≥ kx. With k = kx we have
g = f2k on Ux, hence using (e) with z = e it follows
i=1 ψi(x)ei − e = v ≤ 1
i=1 ϕ2i−1(x) = ∑k
g′(x) − e = f ′
2k(x) − e ≤ f ′(x) +
2k
Q
i=1
ψi(x)ei − e + σ
2k
Q
i=1
1{ϕi>0}(x) ≤ 1
2 η,
and by combining this with (2.9), we obtain (iii).
20
3 Proofs of main results
Proof of Theorem 1.9. Recall that we are given a cone unrectifiable set E ⊂
Rn. We are also given ε > 0 and a continuous function ω ≥ 0 such that
E ⊂ {x ∶ ω(x) > 0}; if ω is not given, we set ω = 1 everywhere on Rn.
We begin by finding numbers σk > 0, vectors ek ∈ B(0, 1) and continuous
functions ϕk ∶ Rn → [0, 1], k = 1, 2, . . . , such that
(A) ∑k σk1spt(ϕk) ≤ ε;
(B) wek,ϑ(σk)(E ∩{ϕk > 0}) = 0;
(C) if x ∈ E, e ∈ N(E, x) and e ≤ 1, then for every η > 0 there are
arbitrarily large k such that σ2k−1 < η, e−e2k−1 < η and ϕ2k−1(x) = 1;
(D) for every k ≥ 1, ϕ2k = ϕ2k−1 and e2k = −e2k−1.
2 min(1, ω, ρ2
H0) and use Lemma 2.7
For this, it suffices to take σl, el and ϕl from Lemma 2.8 with ε replaced by
ε~2 and let σ2l−1 = σ2l ∶= σl, ϕ2l−1 = ϕ2l ∶= ϕl, e2l−1 ∶= el and e2l ∶= −el.
We set f0 ∶= 0, H0 ∶= {ω > 0}, ω0 ∶= 1
to find fj, ωj , Hj , ψj, j = 1, 2, . . . such that
(E) Hj is open, E ⊂ Hj ⊂ Hj−1 and fj ∈ Lip(Rn) ∩ C 1(Hj);
(F) ωj ∈ C(Rn,[0, ∞)) ∩ C(Hj,(0, ∞)) and ωj ≤ 1
(G) fj − fj−1 ≤ ωj−1 and fj(x) = fj−1(x) when ϕj(x) = 0;
(H) if h ∶ Rn → R and h − fj ≤ 2ωj then for every x ∈ Hj one may find
j(x), y⟩ ≤ σjr;
2 min(1, ωj−1, ρ2
Hj);
0 < r < ωj−1(x) such that supy≤r h(x + y) − h(x) − ⟨f ′
(I) ψj ∶ Rn → [0, 1], 0 ≤ ψj ≤ ϕj 1Hj−1 and ψj = ϕj on Hj;
(J) f ′
(K) f ′
j−1(x) − ψj(x)ej ≤ σj 1{ϕj >0}(x) for every x ∈ E;
i=1 ψi(x)ei − z +∑j
j(x) − f ′
j(x) − z ≤ f ′
and a.e. x ∈ Rn.
0(x) +∑j
i=1 σi1{ϕi>0}(x) for all z ∈ Rn
Notice that (F) implies ωj ≤ 2i−j ωj for j ≥ i, and so also ωj ≤ 2−j.
Consequently, by (G), fj converge uniformly to a function f ∶ Rn → R and
f − fj ≤ ∑∞
i=j ωi ≤ 2ωj. We show that f has the required properties.
Notice that (I) and (D) imply that
ψ2i−1(x)e2i−1 + ψ2i(x)e2i = −(ψ2i−1(x) − ψ2i(x))1H2i−2∖H2i(x)e2i,
21
and this vector has norm at most 1H2i−2∖H2i(x), as condition (I) implies
0 ≤ ψ2i ≤ ϕ2i1H2i−1 = ϕ2i−11H2i−1 ≤ ψ2i−1 ≤ ϕ2i−1 ≤ 1 (cf. (2.7)). Hence (K)
with z = 0 and (A) give
f ′
2k(x) =
k
i=1(ψ2i(x)e2i + ψ2i−1(x)e2i−1) +
Q
k
1H2i−2∖H2i(x) +
Q
i=1
2k
Q
i=1
σi1{ϕi>0}(x) ≤ 1 + ε
2k
Q
i=1
σi1{ϕi>0}(x)
≤
for almost every x. Since (E) shows that f2k is Lipschitz, Lip(f2k) ≤ 1 + ε,
and we conclude that Lip(f) ≤ 1 + ε.
For every i ≥ 1 and x ∈ E ⊂ H2i ⊂ H2i−1, (I), (D) and (J) imply
f ′
2i(x) − f ′
= (f ′
2i−1(x) − ϕ2i(x)e2i) +(f ′
≤ σ2i1{ϕ2i>0}(x) + σ2i−11{ϕ2i−1>0}(x).
2i−2(x)
2i(x) − f ′
2i−1(x) − f ′
2i−2(x) − ϕ2i−1(x)e2i−1)
2k(x) − u(x) < 1
4 η, e − e2k+1 < 1
Since ∑j σj 1{ϕj >0}(x) ≤ ε by (A), the restrictions of f ′
2k to E converge point-
wise to a function u ∶ E → Rn and u(x) ≤ ε for x ∈ E.
Suppose x ∈ E, e ∈ N(E, x), e ≤ 1 and η > 0. By (C) there is k such that
4 η and ϕ2k+1(x) = 1.
2−2k < η, f ′
4 η, σ2k+1 < 1
Since x ∈ E ⊂ H2k+1, the latter immediately implies ψ2k+1(x) = 1 by (I).
Since f − f2k+1 ≤ 2ω2k+1 and (J) gives f ′
2k(x) + e2k+1) ≤ σ2k+1,
we conclude that (H) provides 0 < r < ω2k(x) ≤ 2−2kω0 < η such that for
every y ≤ r,
f(x+y) − f(x) − ⟨u(x) + e, y⟩
≤ f(x + y) − f(x) − ⟨f ′
2k+1(x) −(f ′
2k+1(x) − (f ′
2k(x) + e2k+1)y
2k+1(x), y⟩ + f ′
2k(x) − u(x)y + e2k+1 − ey
+ f ′
< (σ2k+1 + σ2k+1 + η~4 + η~4)r < ηr.
Since η > 0 may be arbitrarily small,
lim inf
r↘0
sup
y≤r
f(x + y) − f(x) − ⟨e + u(x), y⟩
r
= 0,
(3.1)
which is the main statement we wished to prove. The estimate of the lower
and upper derivatives is an immediate consequence:
e ≤ 1, we use (3.1) for e and −e to infer
if e ∈ N (E, x) and
D+f(x; y) − D+f(x; y) ≥ ⟨e + u(x), y⟩ − ⟨−e + u(x), y⟩ = 2⟨e, y⟩.
22
Proof of Corollary 1.10. We are given E = ⋃k≥1 Ek ⊂ Rn where Ek are dis-
joint cone unrectifiable Fσ sets, and Nx = N (Ek, x) ∩ B(0, 1) for x ∈ Ek.
Write Ek = ⋃j≥1 Hk,j where Hk,j are closed cone unrectifiable sets, and
let Fk,j ∶= ⋃i<j Hk,i and Ek,j ∶= Hk,j ∖ Fk,j, so that Ek,j are pairwise disjoint
over all (k, j). Let ck,j ∶= 2−k−j and ωk,j(x) ∶= ck,j min(1, dist2(x, Fk,j)). By
Theorem 1.9 there are Lipschitz functions fk,j ∶ Rn → R such that Lip(fk,j) <
2, fk,j ≤ ωk,j and
D+fk,j(x; y) − D+fk,j(x; y) ≥ 2 sup{⟨e, y⟩ ∶ e ∈ N (Ek,j, x), e ≤ 1}
≥ 2 sup
e∈Nx
⟨e, y⟩
for x ∈ Hk,j and y ∈ Rn; the last inequality follows from Nx ⊂ N (Ek,j, x).
Apply Lemma 2.2 to ω = ωk,j+1, H = {ωk,j+1 > 0}, g = fk,j, Φ = 0 and
ξ = 2 to find Lipschitz functions gk,j ∶ Rn → R such that gk,j ∈ C 1{ωk,j+1 > 0},
gk,j −fk,j ≤ ωk,j+1 and Lip(gk,j) ≤ 3. We observe that gk,j is differentiable at
every x ∉ Hk,j. Indeed, for such an x, if ωk,j(x) = 0, i.e. x ∈ Fk,j ⊂ Fk,j+1, then
gk,j(x) = fk,j(x) = 0 as ωk,j(x) = ωk,j+1(x) = 0, and gk,j(y) ≤ 2ck,jy − x2 ≤
y − x2, using upper estimates for gk,j − fk,j and fk,j , and x ∈ Fk,j ⊂ Fk,j+1;
hence g′
k,j(x) = 0. If, however, x ∉ Hk,j and ωk,j(x) > 0, then x ∉ Ek,j ∪ Fk,j,
hence ωk,j+1(x) > 0 and so it follows that gk,j is C 1 on a neighbourhood of x.
We also observe that for every x ∈ Hk,j and y ∈ Rn, we have x ∈ Fk,j+1, and
therefore gk,j(y) − fk,j(y) ≤ ck,j+1y − x2 and hence gk,j(x) = fk,j(x) and
D+gk,j(x; y) − D+gk,j(x; y) = D+fk,j(x; y) − D+fk,j(x; y) ≥ 2 sup
e∈Nx
⟨e, y⟩. (3.2)
Summarising, gk,j is differentiable at every x ~∈ Hk,j and is not differentiable
at any x ∈ Hk,j, moreover, it satisfies (3.2) at such points x.
cs,tgs,t and hk,j ∶= Q
cs,tgs,t. Since for any (s, t),
We let f ∶= Q
(s,t)
(s,t)≠(k,j)
if x ∉ Hs,t, then the function gs,t is differentiable at x, and since we have
∑s,t Lip(cs,tgs,t) < ∞, we infer that f is differentiable at any x ∉
Hs,t = E
(s,t)
and hk,j is differentiable at any x ∈ Hk,j ∪ (Rn ∖ E).
Let x ∈ Ek and find j such that x ∈ Hk,j. Then for every y ∈ Rn,
D+gk,j(x; y) − D+gk,j(x; y) ≥ 2 supe∈Nx⟨e, y⟩ by (3.2), and so, since f =
ck,jgk,j + hk,j and hk,j is differentiable at x, we conclude that
D+f(x; y) − D+f(x; y) ≥ 2ck,j sup
e∈Nx
⟨e, y⟩.
Proof of Corollary 1.11. We are given a set E ⊂ Rn that is a countable
union of (not necessarily disjoint) cone unrectifiable Fσ sets. Since each
23
of these Fσ sets is a countable union of closed cone unrectifiable sets, we
can write E = ⋃∞
k=1 Fk where Fk are closed and cone unrectifiable. Hence
E = ⋃∞
k=1 Ek where Ek ∶= Fk ∖ ⋃j<k Fj are disjoint cone unrectifiable Fσ sets,
and it suffices to take the function f obtained from Corollary 1.10 used with
these sets Ek.
will have all the required properties.
Proof of Corollary 1.12. We are given a Radon measure µ on Rn and a µ-
measurable map T ∶ Rn → ⋃n
m=0 G(n, m) such that for every unit vector e and
α ∈ (0, 1), the set {x ∶ Ce,α ∩ T (x) = {0}}, where Ce,α ∶= {u ∶ ⟨u, e⟩ ≥ αu},
is the union of a µ-null set and a set E with we,α(E) = 0. We show that
there are cone unrectifiable Fσ sets Ek such that µ(Rn ∖ ⋃k Ek) = 0 and
T (x)⊥ ⊂ N (x, Ek) for every x ∈ Ek. Then the function f from Corollary 1.10
By Lusin's Theorem, µ-almost all of Rn is covered by the union of disjoint
closed sets Fk such that for each k, the restriction of T to Fk is continuous.
For every rational α ∈ (0, 1) and u from a countable dense subset Q of
the unit sphere in Rn write {x ∶ Cu,α ∩ T (x) = {0}} = Zu,α ∪ Eu,α, where
µ(Zu,α) = 0 and wu,α(Eu,α) = 0. Letting Ek be Fσ subsets of Fk ∖ ⋃u,α Zu,α
satisfying µ(Fk ∖ Ek) = 0, we just need to show that T (x)⊥ ⊂ N (x, Ek)
for x ∈ Ek. For this, assume x ∈ Ek, e ∈ T (x)⊥ and ε ∈ (0, 1), and choose
u ∈ Q and rational α ∈ (0, 1) so that Ce,ε ⊂ Cu,α and Cu,α ∩ T (x) = {0}.
By continuity of T on Fk, there is r > 0 such that Cu,α ∩ T (y) = {0} for
every y ∈ B(x, r) ∩ Fk. Hence B(x, r) ∩ Ek ⊂ Eu,α and we,ε(B(x, r) ∩ Ek) ≤
wu,α(Eu,α) = 0.
Proof of Theorem 1.13. Let E be the given uniformly purely unrectifiable
set. Pick a sequence ek dense in the unit ball of Rn such that ek ≤ 1 − 2−k.
Let f0 = 0, H0 = Rn, ω0 = 1 and ηk = 2−k−1. When fk−1, Hk−1 and ωk−1
have been defined, we use Lemma 2.9 to find fk, Hk and ωk ∶= ξ such that
(a) E ⊂ Hk ⊂ Hk−1, ωk ∈ C(Rn,[0, ∞)) ∩ C(U,(0, ∞)) and ωk ≤ 1
(b) fk −fk−1 ≤ ωk−1, Lip(fk) ≤ max(Lip(fk−1),ek)+ηk and fk ∈ C 1(Hk);
(c) if x ∈ E and h ∶ Rn → R satisfies h − fk ≤ 2ωk, there is 0 < r < ωk−1(x)
2 ωk−1;
such that supy≤r h(x + y) − h(x) − ⟨ek, y⟩ ≤ ηkr.
Notice that ω0 = 1 and the last inequality in (a) imply ωj ≤ 2j−kωk and
ωk ≤ 2−k for j ≥ k ≥ 0. From (b) we see by induction that Lip(fk) ≤ 1 − 2−k−1.
Hence the inequality fk − fk−1 ≤ ωk−1 ≤ 2−k+1 implies that fk converge to
some f ∶ Rn → R with Lip(f) ≤ 1.
24
j=k ωj ≤ ∑∞
Given any x ∈ E, e ∈ Rn with e ≤ 1, and ε > 0, there are arbitrarily
large k such that ek − e < ε and ηk < ε. Inferring from (b) that f − fk ≤
∑∞
j=k 2j−kωk ≤ 2ωk, we use (c) to find 0 < r < ωk−1(x) ≤ 2−k+1 such
that supy≤r f(x + y) − f(x) − ⟨ek, y⟩ ≤ ηkr < εr. Since ek − e < ε, we
conclude that supy≤r f(x + y) − f(x) − ⟨e, y⟩ < 2εr. As ε > 0 is arbitrary
and k may be arbitrarily large,
lim inf
r↘0
sup
y≤r
f(x + y) − f(x) − ⟨e, y⟩
r
= 0,
which is the statement (1.5) of the Theorem. The estimate of upper and
lower derivatives follows by using this with e = y~y and e = −y~y to get
D+f(x; y) ≥ y and D+f(x; y) ≤ −y, respectively.
4 Examples
The argument behind our first example has already been used many times,
starting with [21], to find points of differentiability or almost differentiability
of Lipschitz functions. See, e.g., [10, 13] or [4, Example 4.7] for an example
showing that in Corollary 1.12 the constant c = c(x) cannot be bounded
away from zero.
Example 4.1. There is a compact set E ⊂ R2 and a continuous mapping
x ∈ E → ex ∈ {e ∈ R2 ∶ e = 1} such that N (E, x) = {tex ∶ t ∈ R} for every
x ∈ E and whenever f ∶ R2 → R has Lip(f) ≤ 1, there is x ∈ E such that
Df(x, ex) < 1. Consequently, in Theorem 1.9 we cannot take ε = 0.
Proof. Let ϕ ∶ R → R be a C 1 function such that ϕ(−1) = ϕ(1) = 0, ϕ′(−1) =
ϕ′(1) = 0 and ϕ(s) > 0 for s ≠ ±1. Denote ϕ0 = 0 and ϕk = ϕ~k, and let
E ∶= {(s, ϕk(s)) ∶ s ∈ [−1, 1], k = 0, ±1, ±2, . . .}.
For x ∈ E, x = (s, ϕk(s)) let ux and ex denote the unit vectors in the
directions of (1, ϕ′
k(s), 1), respectively. Then ex ∈ N(E, x)
and, since ϕ′(−1) = ϕ′(1) = 0, the map x ∈ E → ex is continuous.
Suppose f ∶ R2 → R has Lip(f) ≤ 1. Consider any x ∈ E ∖{(−1, 0),(1, 0)}
such that a ∶= f ′(x, ux) exists and Df(x; ex) = 1. Then
k(s)) and (−ϕ′
= lim sup
t→0
f(x + tex) − f(x)
t
− lim
t→0
f(x − atux) − f(x)
t
= 1 + a2.
25
lim sup
t→0
f(x + tex) − f(x − atux)
t
Hence
1 = Lip(f) ≥ lim sup
t→0+
f(x + tex) − f(x − atux)
(x + tex) −(x − atux) =
a2 + 1√a2 + 1
,
which gives f ′(x, ux) = 0.
If f is a function satisfying the conclusion of Theorem 1.9 with ε = 0, then
for every k, x = (s, ϕk(s)) satisfies the above assumptions for a.e. s ∈ (−1, 1).
Since f is Lipschitz, we infer that s → f(s, ϕk(s)) is constant on [−1, 1], and
hence f is constant on E. Consequently, when s ∈ (−1, 1) and x = (s, ϕ0(s)),
ex = (0, 1) and so limt→0 (f(x +tex)−f(x))~t ≤ limt→0 dist(x +tex, E)~ t = 0,
as dist(x + tex, E)~ t ≤ (k + 1)~2k(k + 1) = 1~(2k) when t is between
ϕ(x)~(k + 1) and ϕ(x)~k. This contradicts Df(x; ex) = 1.
Our second example is related to Zahorski's description of non-differen-
tiability sets of real-valued functions of a real variable which was already
mentioned in the introductory remarks to Corollaries 1.10 and 1.11. Recall
first that the set of points of non-differentiability of any real-valued function
f ∶ Rn → R is easily seen to be of the type Gδσ: just write it as
ε>0
e∈Rn{x ∶ (∃r > 0)(∃u, v ∈ B(x, r)) f(x + u) − f(x + v) − ⟨e, u − v⟩ > εr}
where ε runs over positive rational numbers and e over elements of a dense
countable subset of Rn. The main argument in Zahorski's [25] proof of
the converse when n = 1 (both in the general and in the Lipschitz case)
constructs, for a given Gδ Lebesgue null set E ⊂ R, a function f ∶ R → R
with Lip(f) = 1 which is differentiable at every point of R ∖ E and at every
point of E has upper derivative 1 and lower derivative −1.
(For a more
modern treatment of this construction see [15].)
While it is not clear what an exact analogy of Zahorski's result for n > 1
should be, one may at least hope that its analogy holds for uniformly purely
unrectifiable sets, namely that for every uniformly purely unrectifiable Gδσ
set E ⊂ Rn there is a Lipschitz function f ∶ Rn → R such that E is precisely
the set of points at which f in non-differentiable in any direction. We do not
know whether this is true or not, but the following example shows that in this
situation the argument based on uniform discrepancy between upper and
lower derivatives fails in a very strong sense. Recalling that every uniformly
purely unrectifiable set is contained in a Gδ uniformly purely unrectifiable
set, the example provides a Gδ uniformly purely unrectifiable set such that
not only for it, but even for any bigger Gδ uniformly purely unrectifiable set
there is no function analogous to the one from Zahorski's main argument.
26
Example 4.2. There is a uniformly purely unrectifiable set A ⊂ R2 such that
for any set E ⊃ A and any c > 0 there is no Lipschitz function f ∶ R2 → R
such that
(a) D+f(x; y) − D+f(x; y) ≥ cy for every x ∈ E and y ∈ R2;
(b) f is differentiable at every point x of R2 ∖ E.
Proof. By [9] there is a universal differentiability set D ⊂ R2, i.e., a set such
that every real-valued Lipschitz function on R2 has a point of differentiability
belonging to D, such that there is a Lipschitz h ∶ R2 → R for which the set
A of points x ∈ D such that h is differentiable at x, is uniformly purely
unrectifiable. Suppose E ⊃ A and Lipschitz f ∶ R2 → R satisfy (a) and (b).
For a small ε ∈ 0, c~(4Lip(h)) consider the function g ∶= f + εh. If x ∈ E,
(a) shows that for some y ∈ R, D+g(x; y) − D+g(x; y) ≥ (c − 2εLip(h))y > 0.
If x ∈ D ∖ E, g is the sum of the function f that is differentiable at x and of
the function εh that is non-differentiable at x; hence it is non-differentiable
at x. Consequently, the Lipschitz function g has no point of differentiability
at D, contradicting that D is a universal differentiability set.
Remark 4.3. The reason for considering a uniform non-differentiability
condition such as (a) was explained in the text before the Example. Notice
that, if (a) were replaced just by non-differentiability of f at every point of E,
the statement of the Example would be false: we would use Theorem 1.13
to find a function g that is non-differentiable at every point of A and define
E as the non-differentiability set of g. On the other hand, it is easy to find
uniformly purely unrectifiable sets E ⊃ A for which there is no Lipschitz
function non-differentiable exactly at points of E, as such E need not be Gδσ.
For the set A from [9] which was used in the proof of the Example 4.2 we
can take E = A as it is not difficult to see that A is not Gδσ, although it is
Fσδ since A is the intersection of D with the set of points of differentiability
of h and D used in [9] is Gδ. It may be of interest to notice that the fact
that A is not a non-differentiability set of any Lipschitz function f may
be seen directly from the properties of A, D and h:
for any such f the
Lipschitz function f + h would be non-differentiable at any x ∈ D ∖ A as f is
differentiable and h is not differentiable at such x; and f + h would be non-
differentiable at any x ∈ A as f is not differentiable and h is differentiable
at such x. As in the proof of the Example 4.2, this a contradiction as D is
a universal differentiability set.
Our final example is related to the already pointed out fact that E
is uniformly purely unrectifiable if and only if there is 0 < η < 1 such
27
that we,η(E) = 0 for every unit vector e. When considering general non-
differentiability sets, a natural analogy of this statement would say that for
any set E ⊂ Rn satisfying we,η(E) = 0 for some unit vector e and some
0 < η < 1 there is a real-valued Lipschitz function f on Rn that is non-
differentiable at any point of E. We show here that this is false; recall
however that [3] shows (directly, not using [8]) that for any such set E there
is an Rn-valued Lipschitz function f on Rn that is non-differentiable at any
point of E.
Example 4.4. For every η ∈ (0, 1) and a unit vector e ∈ R2 there is a
universal differentiability set E ⊂ R2 such that we,η(E) = 0.
Proof. Let Lj be an enumeration of all rational lines in Rn, J the set of those
indexes j for which the direction u of Lj satisfies ⟨u, e⟩ < 1
2 η and εi,j > 0 such
that ∑i,j εi,j < ∞. It is easy to see that E ∶= ⋂j ⋃j∈J{x ∶ dist(x, Lj) < εi,j}
satisfies we,η(E) = 0. The fact that E is a universal differentiability set has
been often mentioned, but does not seem to be documented in the literature.
We therefore explain the argument.
Recall from [12], [13] or [21] that, given any Lipschitz g ∶ Rn → R, a
procedure leading to a point of differentiability of g may be described as
follows. One starts with an arbitrary δ0 > 0 and (x0, e0) from the set D
of pairs (x, u) where x ∈ Rn, u is a unit vector, and there is j = j(x, u)
such that x ∈ Lj and u is the direction of L. Recursively, when (xk, ek)
has been defined, one first chooses an arbitrarily small δk+1 > 0 and then
(xk+1, ek+1) ∈ D satisfying rather delicate conditions about which we need
to know only that xk+1 ∈ B(xk, δk+1), Dg(xk+1, ek+1) ≥ Dg(xk, ek) and that
they imply that the sequence xk converges to a point of differentiability of g.
Returning to our set E, given any Lipschitz f ∶ Rn → R, choose (x0, e0) ∈
D so that ⟨e0, e⟩ < 1
4 η and let g(x) ∶= f(x) + c⟨x, e0⟩ with c > 64Lip(f)~η2;
the choice of such large c guarantees that Dg(x, u) ≥ Dg(x0; e0) implies
0 ≤ 1−⟨u, e0⟩ ≤ 1
32 η2, so that u−e0 ≤
1
4 η, hence ⟨u, e⟩ ≤ u − e0 + ⟨e0, e⟩ < 1
2 η. This will imply that in the
recursive construction jk ∶= j(xk, ek) ∈ J, and so we can choose δk+1 such
that B(xk, δk+1) ⊂ B(Ljk , εk,jk) ∩B(xk, δk). Hence the limit of the xk, which
c (Df(x; u)−Df(x0; e0)) ≤ 2Lip(f)~c ≤ 1
is a differentiability point of g and so of f , belongs to E.
References
[1] Giovanni Alberti, Marianna Csornyei, and David Preiss. Structure of
null sets in the plane and applications. In European Congress of Math-
ematics, 3 -- 22. Eur. Math. Soc., Zurich, 2005.
28
[2] Giovanni Alberti, Marianna Csornyei, and David Preiss. Differentiabil-
ity of Lipschitz functions, structure of null sets, and other problems. In
Proceedings of the International Congress of Mathematicians. Volume
III, 1379 -- 1394. Hindustan Book Agency, New Delhi, 2010.
[3] Giovanni Alberti, Marianna Csornyei, and David Preiss. Structure of
null sets, differentiability of Lipschitz functions, and other problems
(provisional title). Paper in preparation, 2017.
[4] Giovanni Alberti and Andrea Marchese. On the differentiability of Lip-
schitz functions with respect to measures in the Euclidean space. Geom.
Funct. Anal., 26(1), 1 -- 66, 2016.
[5] Daniel Azagra, Juan Ferrera, Fernando L´opez-Mesas, and Yenni
Rangel. Smooth approximation of Lipschitz functions on Riemannian
manifolds. J. Math. Anal. Appl., 326(2), 1370 -- 1378, 2007.
[6] Yoav Benyamini and Joram Lindenstrauss. Geometric nonlinear func-
tional analysis. Vol. 1, volume 48 of American Mathematical Society
Colloquium Publications. American Mathematical Society, Providence,
RI, 2000.
[7] Jeff Cheeger. Differentiability of Lipschitz functions on metric measure
spaces. Geom. Funct. Anal., 9(3), 428 -- 517, 1999.
[8] Marianna Csornyei and Peter Jones. Product formulas for measures
and applications to analysis and geometry. Lecture of Peter Jones, see
http://www.math.sunysb.edu/Videos/dfest/PDFs/38-Jones.pdf.
[9] Marianna Csornyei, David Preiss, and Jaroslav Tiser. Lipschitz func-
tions with unexpectedly large sets of nondifferentiability points. Abstr.
Appl. Anal., (4), 361 -- 373, 2005.
[10] Thierry de Pauw and Petri Huovinen. Points of ε-differentiability of
Lipschitz functions from Rn to Rn−1. Bull. London Math. Soc., 34(5),
539 -- 550, 2002.
[11] Guido De Philippis and Filip Rindler. On the structure of A-free mea-
sures and applications. Ann. of Math. (2), 184(3), 1017 -- 1039, 2016.
[12] Michael Dor´e and Olga Maleva. A universal differentiability set in
Banach spaces with separable dual. J. Funct. Anal., 261(6), 1674 -- 1710,
2011.
29
[13] Michael Dymond and Olga Maleva. Differentiability inside sets with
Minkowski dimension one. Michigan Math. J., 65(3), 613 -- 636, 2016.
[14] Lawrence Craig Evans. Partial differential equations, volume 19 of
Graduate Studies in Mathematics. American Mathematical Society,
Providence, RI, 1998.
[15] Thomas Fowler and David Preiss. A simple proof of Zahorski's de-
scription of non-differentiability sets of Lipschitz functions. Real Anal.
Exchange, 34(1), 127 -- 138, 2009.
[16] Petr H´ajek and Michal Johanis. Smooth approximations. J. Funct.
Anal., 259(3), 561 -- 582, 2010.
[17] Joram Lindenstrauss, David Preiss, and Jaroslav Tiser. Fr´echet dif-
ferentiability of Lipschitz functions and porous sets in Banach spaces,
volume 179 of Annals of Mathematics Studies. Princeton University
Press, Princeton, NJ, 2012.
[18] Andrea Marchese and Andrea Schioppa. Lipschitz functions with pre-
scribed blowups at many points. arXiv:1612.05280.
[19] Pierre Pansu. M´etriques de Carnot-Carath´eodory et quasiisom´etries
des espaces sym´etriques de rang un. Ann. of Math. (2), 129(1), 1 -- 60,
1989.
[20] Andrea Pinamonti and Gareth Speight. A measure zero universal differ-
entiability set in the heisenberg group. Mathematische Annalen, 368(1-
2), 233 -- 278, 2017.
[21] David Preiss. Differentiability of Lipschitz functions on Banach spaces.
J. Funct. Anal., 91(2), 312 -- 345, 1990.
[22] David Preiss. Gateaux differentiability of cone-monotone and pointwise
Lipschitz functions. Israel J. Math., 203(1), 501 -- 534, 2014.
[23] David Preiss and Gareth Speight. Differentiability of Lipschitz func-
tions in Lebesgue null sets. Invent. Math., 199(2), 517 -- 559, 2015.
[24] Stanis law Saks. Theory of the integral. 1937.
[25] Zygmunt Zahorski. Sur l'ensemble des points de non-d´erivabilit´e d'une
fonction continue. Bull. Soc. Math. France, 74, 147 -- 178, 1946.
30
|
1301.2213 | 1 | 1301 | 2013-01-10T18:46:20 | A real seminorm with square property is submultiplicative | [
"math.FA"
] | A seminorm with square property on a real associative algebra is submultiplicative | math.FA | math | A REAL SEMINORM WITH SQUARE PROPERTY IS SUBMULTIPLICATIVE
M.El Azhari
Ecole Normale Supérieure, Avenue Oued Akreuch, Takaddoum,
BP 5118, Rabat, Morocco
E-mail: [email protected]
A seminorm with square property on a real associative algebra is submultiplicative.
Key words: real seminorm, seminorm with square property, submultiplicative seminorm.
1. INTRODUCTION
It was proved in [3] that every complex seminorm with square property on a commutative algebra is
submultiplicative, and it was posed the problem whether this result holds in a noncommutative
algebra. This problem was answered in the particular case of Banach algebras [4] and fully resolved in
[2], [5] and [7]. The result of [3] holds in the real case. But the results of [2], [4], [5] and [7] don’t hold
in the real case since we use the Hirschfeld-Zelazko Theorem [6], or its locally bounded version [2],
which are not valid in the real case. Using a functional representation theorem [1, Theorem 1], we
show that every real seminorm with square property is submultiplicative.
2. PRELIMINARIES
Let A be an associative algebra over the field K =R or C. A seminorm on A is a function p: A→[0, ∞)
satisfying p(a+b) ≤ p(a)+p(b) for all a, b in A and p(ka)=│k│p(a) for all a in A and k in K. p is a complex
seminorm if K = C and p is a real seminorm if K=R. p is submultiplicative if p(ab) ≤ p(a)p(b) for all a, b
in A. p satisfies the square property if p(a²)=p(a)² for al l a in A. Let A be a real algebra and let a be any
element of A. The spectrum sp (a) of a in A is defined to be equal to the spectrum of a as an element
of the complexification of A. If A is unital, then sp (a) ={s + it ϵ C, (a - se) ² + t²e ∉ A⁻¹} for all a in A,
where e is the unit of A and A⁻¹ is the set of all invertible elements in A. Let (A , ǁ.ǁ) be a real normed
algebra, the limit r (a) =lim ǁaⁿǁ ¹/ⁿ exists for each a in A . If A is complete,
then r (a) =sup {│v│, v ϵ sp (a)} for every a in A.
-2-
3. RESULTS
Let (A, ǁ.ǁ) be a real Banach algebra with unit such that ǁaǁ ≤ m r (a) for some positive constant m and
all a in A. Let X (A) be the set of all nonzero multiplicative linear functionals from A into the
noncommutative algebra H of quaternions. For a in A and x in X (A), put J (a) (x) = x (a). For a in A,
J(a): X(A) → H is a map from X(A) into H. X(A) is endowed with the topology generated by J(a), a ϵ A,
that is the weakest topology such that all the functions J(a), a ϵ A, are continuous. By [1, Theorem 1],
X(A) is a nonempty compact space and the map J: A → C(X(A),H), a → J(a), is an isomorphism (into),
where C(X(A),H) is the real algebra of all continuous functions from X(A) into H.
Proposition 3.1. (1) An element a is invertible in A if and only if J (a) is invertible in C(X (A), H), and
(2) sp (a) = sp (J (a)) for all a in A.
Proof. (1) The direct implication is obvious. Conversely, there exists g in C(X(A),H) such that
J(a) g = g J(a)=1 i.e. x(a) g(x) = g(x) x(a) = 1 for all x in X(A). Let T be a nonzero irreducible
representation of A, by the proof of [1, Theorem 1] there exists S: T(A) → H an isomorphism (into).
Since SoT ϵ X(A) and 0ǂ SoT(a) = S(T(a)), it follows that T(a) ǂ 0. If aA ǂ A, there exists a maximal
right ideal M containing aA. Let L be the canonical representation of A on A/M which is nonzero and
irreducible, also L(a) = 0 since aA is included in M, contradiction. Then aA = A and by the same
Aa = A. There exist b, c in A such that ab = ca = e ( e is the unit of A). We have c = c(ab) = (ca)b = b, so
a is invertible in A.
(2) s + it ϵ sp(a) iff (a - se)2 + t2e ∉ A 1
iff J((a – se)2 +t2e) ∉ C(X(A),H) 1 by (1)
iff (J(a) – s J(e))2 + t2 J(e) ∉ C(X(A),H) 1
iff s + it ϵ sp(J(a)).
Theorem 3.2. Let A be a real associative algebra. Then every seminorm with square property on A is
submultiplicative.
Proof. If A is commutative, see [3, Theorem 1]. If A is noncommutative, let p be a seminorm with
square property on A. By [5] or [7], there exists m > 0 such that p(ab) ≤ mp(a)p(b) for all a, b in A. Ker(p)
is a two sided ideal in A , the norm │.│ on the quotient algebra A/Ker(p), defined by │a + Ker(p)│= p(a)
is a norm with square property on A/Ker(p). Define ǁa + Ker(p)ǁ = m│a + Ker(p)│ for all a in A . Let a, b
in A, ǁab +Ker(p)ǁ = m│ab + Ker(p)│ ≤ m2 │a + Ker(p)││b + Ker(p)│ = ǁa + Ker(p)ǁ ǁb + Ker(p)ǁ.
(A/Ker(p), ǁ.ǁ) is a real normed algebra. Let a in A, ǁa2 + Ker(p)ǁ = m│a2 + Ker(p)│= m│a + Ker(p)│2 =
m 1 ( m│a + Ker(p)│)2 = m 1 ǁa + Ker(p)ǁ2. The completion B of ( A/Ker(p), ǁ.ǁ) satisfies also the
property ǁb2ǁ = m 1 ǁbǁ2 for all b in B, and consequently ǁb2ⁿǁ2¯ⁿ = m2¯ⁿ 1 ǁbǁ for all b in B and n in
N*, then r(b) = m 1 ǁbǁ i.e. ǁbǁ = m r(b). We consider two cases.
B is unital: By [1, Theorem 1], X(B) is a nonempty compact space and the map J: B → C(X(B),H) is an
isomorphism ( into). C(X(B),H) is a real Banach algebra with unit under the supnorm ǁ.ǁ s . By
-3-
Proposition 3.1, r(b) = r(J(b)) for all b in B. Let b in B, ǁbǁ = m r(b) = m r(J(b)) = m ǁJ(b)ǁ s since the
supnorm satisfies the square property. Then ǁJ(b)ǁs = m 1 ǁbǁ = │b│ for all b in A/Ker(p), so │.│ is
submultiplicative on A/Ker(p) i.e. p is submultiplicative.
B is not unital: Let B1 be the algebra obtained from B by adjoining the unit. By the same proof of
[6, Lemma 2], there exists a norm N on B1 such that
(i)
(ii)
(iii)
(B1 , N) is a real Banach algebra with unit
N(b) ≤ m3 rB₁ (b) for all b in B1
N and ǁ.ǁ are equivalent on B.
By [1, Theorem 1], X(B1) is a nonempty compact space and the map J: B1 → C(X(B1),H) is an
isomorphism ( into). Let b in B, ǁbǁ = m rB(b) = m rB₁(b) by (iii)
= m r(J(b)) by Proposition 3.1
= m ǁJ(b)ǁs by the square property of the supnorm.
Then ǁJ(b)ǁs = m 1 ǁbǁ = │b│ for all b in A/Ker(p), so │.│ is submultiplicative on A/Ker(p)
i.e. p is submultiplicative.
REFERENCES
1. M.Abel and K.Jarosz, Noncommutative uniform algebras, Studia Math., 162 (2004), 213-218.
2. J.Arhippainen, On locally pseudoconvex square algebras, Publicacions Matemàtiques,
39 (1995), 89-93.
3. S.J.Bhatt and D.J.Karia, Uniqueness of the uniform norm with an application to topological
algebras, Proc. Amer. Math. Soc., 116 (1992), 499-504.
4. S.J.Bhatt, A seminorm with square property on a Banach algebra is submultiplicative, Proc.
Amer. Math. Soc., 117 (1993), 435-438.
5. H.V.Dedania, A seminorm with square property is automatically submultiplicative, Proc.
Indian. Acad. Sci ( Math. Sci.), 108 ( 1998), 51-53.
6. R.A.Hirschfeld and W.Zelazko, On spectral norm Banach algebras, Bull. Acad. Polon. Sci. Sér.
Sci. Math. Astronom. Phys., 16 (1968), 195-199.
7. Z.Sebestyén, A seminorm with square property on a complex associative algebra is
submultiplicative, Proc. Amer. Math. Soc., 130 (2001), 1993-1996.
|
1006.3092 | 1 | 1006 | 2010-06-15T20:53:06 | Topological classification of closed convex sets in Frechet spaces | [
"math.FA",
"math.GN",
"math.GT"
] | We prove that each non-separable completely metrizable convex subset of a Frechet space is homeomorphic to a Hilbert space. This resolves an old (more than 30 years) problem of infinite-dimensional topology. Combined with the topological classification of separable convex sets due to Klee, Dobrowoslki and Torunczyk, this result implies that each closed convex subset of a Frechet space is homemorphic to $[0,1]^n\times [0,1)^m\times l_2(k)$ for some cardinals $0\le n\le\omega$, $0\le m\le 1$ and $k\ge 0$. | math.FA | math |
TOPOLOGICAL CLASSIFICATION OF
CLOSED CONVEX SETS IN FR ´ECHET SPACES
TARAS BANAKH AND ROBERT CAUTY
Abstract. We prove that each non-separable completely metrizable convex subset of a Fr´echet space
is homeomorphic to a Hilbert space. This resolves an old (more than 30 years) problem of infinite-
dimensional topology. Combined with the topological classification of separable convex sets due to
Klee, Dobrowoslki and Toru´nczyk, this result implies that each closed convex subset of a Fr´echet space
is homeomorphic to [0, 1]n × [0, 1)m × ℓ2(κ) for some cardinals 0 ≤ n ≤ ω, 0 ≤ m ≤ 1 and κ ≥ 0.
The problem of topological classification of convex sets in linear metric spaces traces its history back
to founders of functional analysis S.Banach and M.Fr´echet. For separable closed convex sets in Fr´echet
spaces this problem was resolved by combined efforts of V.Klee [8] (see [3, III.7.1]), Dobrowolski and
Toru´nczyk [4], [5]:
Theorem 1 (Klee-Dobrowolski-Toru´nczyk). Each separable closed convex subset C of a Fr´echet space
is homeomorphic to [0, 1]n × [0, 1)m × (0, 1)k for some cardinals 0 ≤ n, k ≤ ω and 0 ≤ m ≤ 1. In
particular, C is homeomorphic to the separable Hilbert space l2 if and only if C is not locally compact.
By a Fr´echet space we mean a locally convex complete linear metric space. A linear metric space is
a linear topological space endowed with an invariant metric that generates its topology. A topological
space is called completely metrizable if its topology is generated by a complete metric.
In this paper we study the topological structure of non-separable (completely metrizable) convex
sets in Fr´echet spaces and prove the following theorem that answers an old problem LS10 posed in
Geoghegan's list [7] and then repeated in [11] and [2].
Theorem 2. Each non-separable completely metrizable convex subset of a Fr´echet space is homeo-
morphic to a Hilbert space.
Theorems 1 and 2 imply the following topological classification of closed convex subset in Fr´echet
spaces.
Theorem 3. Each closed convex subset C of a Fr´echet space is homeomorphic to [0, 1]n×[0, 1)m×ℓ2(κ)
for some cardinals 0 ≤ n ≤ ω, 0 ≤ m ≤ 1 and κ ≥ 0. In particular, C is homeomorphic to an infinite-
dimensional Hilbert space if and only if C is not locally compact.
Here ℓ2(κ) stands for the Hilbert space that has an orthonormal basis of cardinality κ. The topology
of infinite-dimensional Hilbert spaces was characterized by Toru´nczyk [9], [10]. This characterization
was used in the proof of the following criterion from [2] which is our main tool for the proof of
Theorem 2.
Theorem 4 (Banakh-Zarichnyy). A convex subset C in a linear metric space is homeomorphic to
an infinite-dimensional Hilbert space if and only if C is a completely metrizable absolute retract with
LFAP.
A topological space X is defined to have the locally finite approximation property (briefly, LFAP),
if for each open cover U of X there is a sequence of maps fn : X → X, n ∈ ω, such that each fn
1991 Mathematics Subject Classification. 57N17; 46A04.
Key words and phrases. Convex set, Fr´echet space, non-separable Hilbert space.
1
2
TARAS BANAKH AND ROBERT CAUTY
is U -near to the identity idX : X → X and the family (cid:0)fn(X)(cid:1)n∈ω is locally finite in X. The latter
means that each point x ∈ X has a neighborhood O(x) ⊂ X that meets only finitely many sets
fn(X), n ∈ ω.
Theorem 2 follows immediately from Theorem 4, the Borsuk-Dugundji Theorem [3, II.3.1] (saying
that convex subsets of Fr´echet spaces are absolute retracts) and the following theorem that will be
proved in Section 3.
Theorem 5. Each non-separable convex subset of a Fr´echet space has LFAP.
1. Separated Approximation Property
Theorem 5 establishing LFAP in non-separable convex sets will be proved with help of the metric
counterpart of LFAP, called SAP.
A metric space (X, d) is defined to have the separated approximation property (briefly, SAP) if for
each ε > 0 there is a sequence of maps fn : X → X, n ∈ ω, such that each fn is ε-homotopic to idX
and the family (fn(X))n∈ω is separated in the sense that inf n6=m d(fn(X), fm(X)) > 0.
Here for two non-empty subsets A, B ⊂ X we put d(A, B) = inf{d(a, b) : a ∈ A, b ∈ B}. Two
maps f, g : A → X are called ε-homotopic if they can be linked by a homotopy (ht)t∈I : A → X such
that h0 = f , h1 = g and diam{ht(a) : t ∈ I} ≤ ε for all a ∈ A. By I we denote the unit interval [0, 1].
The following lemma is proved by analogy with Lemma 1 of [5] and Lemma 5.2 of [2].
Lemma 1. Each metric space with SAP satisfies LFAP.
Proof. Assume that a metric space (X, d) has SAP. To show that X has LFAP, fix an open cover U
of X and find a non-expanding function ε : X → (0, 1) such that the cover {Bd(x, ε(x)) : x ∈ X}
refines the cover U . For every k ∈ ω consider the closed subset Xk = {x ∈ X : ε(x) ≥ 2−k} of X. Put
εk = 1/4k+2 for k ≤ 1 and let f0 : X × ω → X, f0 : (x, n) 7→ x, be the projection. By induction we
shall construct a sequence (εk)k∈ω of positive real numbers and a sequence of maps fk : X × ω → X,
k ∈ ω, such that the following conditions are satisfied:
4k+2 ;
4 εk−1 ≤ 1
(1) εk ≤ 1
(2) fk is εk-homotopic to fk−1;
(3) fkXk−3 × ω = fk−1Xk−3 × ω;
(4) fk(X \ Xk+1) × ω = f0(X \ Xk+1) × ω;
(5) inf n6=m d(fk(Xk × {n}), fk(Xk × {m})) ≥ 4εk+1.
Assume that maps fi : X × ω → X and numbers εi+1 satisfying the conditions (1) -- (4) have been
constructed for all i < k. By SAP, there is an εk-homotopy (ht)t∈I : X ×ω → X such that h0 = f0 and
δ = inf n6=m d(h1(X × {n}), h1(X × {m}) > 0. Choose a continuous function λ : X → [0, 1] such that
Xk−3 ∪ (X \ Xk+1) ⊂ λ−1(0) and Xk \ Xk−2 ⊂ λ−1(1). Take any positive number εk+1 ≤ 1
4 min{δ, εk}
and define a function fk : X × ω → X by
fk(x, n) = hλ(x)(fk−1(x, n), n).
It is clear that the conditions (1) -- (4) are satisfied. The condition (5) will follow as soon as we check
that d(fk(x, n), fk(y, m)) ≥ 4εk+1 for any points x, y ∈ Xk and distinct numbers n 6= m.
Find unique numbers i, j ≤ k such that x ∈ Xi \ Xi−1 and y ∈ Xj \ Yj−1. If i, j < k, then
d(fk(x, n), fk(y, m)) ≥ d(fk−1(x, n), fk−1(y, m)) − 2εk ≥ 4εk − 2εk = 2εk ≥ 4εk+1.
It remains to consider the case max{i, j} = k. We lose no generality assuming that i = k.
If
j ≥ k − 1, then
d(fk(x, n), fk(y, m)) = d(h1(fk−1(x, n), n), h1(fk−1(y, m), m)) ≥ δ ≥ 4εk+1.
Next, assume that j ≤ k − 2. In this case k ≥ j + 2 ≥ 3. Then
ε(x) < 2−i+1 = 2−k+1 < 2−k+2 ≤ 2−j ≤ ε(y)
TOPOLOGICAL CLASSIFICATION OF CLOSED CONVEX SETS IN FR´ECHET SPACES
3
and the non-expanding property of ε imply that d(x, y) ≥ ε(x) − ε(y) ≥ 2−j − 2−k+1 ≥ 2−j−1. It
follows from (4) and (2) that
d(x, fk(x, n)) = d(fi−2(x, n), fk(x, n)) = d(fk−2(x, n), fk(x, n)) ≤ εk−1 + εk ≤ 2εk−1 ≤
2
4k+1
and
Then
d(y, fk(y, m)) = d(fj−2(y, m), fk(y, m)) ≤ εj−1 + · · · + εk ≤ 2εj−1 ≤
2
4j+1 .
d(fk(x, n), fk(y, m)) ≥ d(x, y) − d(x, fk(x, n)) − d(y, fk(y, m))
1
2j+1 −
This completes the inductive step.
≥
2
4k+1 −
2
4j+1 ≥
1
2j+1 −
2
4j+3 −
2
4j+1 ≥
4
4j+5 ≥
4
4k+3 ≥ 4εk+1.
After completing the inductive construction, let f∞ = limk→∞ fk : X × ω → X. The conditions
(1)-(3) guarantee that the limit function f∞ is well-defined and continuous. Let us show that f∞ is
ε-near to f0. Given any point (x, n) ∈ X × ω, find a unique number i ∈ N such that x ∈ Xi \ Xi−1.
By (3) and (4), f∞(x, n) = fi+2(x, n) and f0(x, n) = fi−2(x, n). Then
d(f∞(x, n), x) = d(f∞(x, n), f0(x, n)) = d(fi+2(x, n), fi−2(x, n))
≤ ei+2 + · · · + εi−1 ≤ 2εi−1 ≤
2
4i <
1
2i ≤ ε(x).
The choice of the function ε guarantees that f∞ is U -near to the projection f0 : X × ω → X.
It remains to prove that the family (f∞(X × {n})n∈ω is discrete in X. Given any point x ∈ X, let
i ∈ N be the unique number such that x ∈ Xi \ Xi−1. Consider the ball B(x; 1/2i+2) = {x′ ∈ X :
d(x, x′) < 1/2i+2} centered at x.
Claim 1. B(x; 1/2i+2) ∩ f∞(X × ω) ⊂ f∞(Xi+1 × ω).
Proof. Assume conversely that f∞(y, m) ∈ O(x) for some y ∈ X \ Xi+1 and m ∈ ω. Let j ∈ ω be a
unique number with y ∈ Xj \ Xj−1. It follows from y /∈ Xi+1 that j ≥ i + 2. Since
d(f∞(y, m), y) = d(fj+3(y, m), fj−2(y, m)) ≤ 2εj−1 ≤
2
4j+1 ≤
1
2i+2 ,
and ε(y) < 1
2j−1 < 1
1
2i+1 ≤
1
2i −
2i ≤ ε(x), by the non-expanding property of ε, we get a contradiction:
1
2i+2 =
1
2j−1 ≤ ε(x) − ε(y) ≤ d(x, y) ≤ d(x, f∞(y, m)) + d(f∞(y, m), y) <
1
2i+2 +
1
2i+1 .
(cid:3)
Now the condition (5) and the inequality εi+2 ≤ 1
2i+2 implies that the ball B(x; εi+2) meets
at most one set f∞(Xi+1 × {n}) and hence at most one set f∞(X × {n}), which means that the family
(cid:0)f∞(X × {n})(cid:1)n∈ω is discrete in X and hence X has LFAP.
(cid:3)
4i+4 ≤ 1
2. SAP in non-separable convex cones
In this section we shall prove that non-separable convex cones in Fr´echet spaces have SAP.
A subset C of a linear metric space (L, d) is called a convex cone if it is convex and R+ · C = C
where R+ = [0, ∞). The principal result of this section is
Lemma 2. Each non-separable convex cone C in a Fr´echet space L has SAP.
For the proof of this lemma we shall use an operator version of Josefson-Nissenzweig Theorem
proved in [1]:
4
TARAS BANAKH AND ROBERT CAUTY
Lemma 3. For any dense continuous non-compact linear operator S : X → Y between normed spaces
there is a linear continuous operator T : Y → c0 such that the operator T S : X → c0 is not compact.
Let us recall that an operator T : X → Y between linear topological spaces is
• dense if T X is dense in Y ;
• compact if the image T (U ) of some open neighborhood U ⊂ X of zero is totally bounded in
Y .
A subset B of a linear topological space Y is totally bounded if for each open neighborhood V ⊂ Y of
zero there is a finite subset F ⊂ Y such that B ⊂ V + F .
Proof of Lemma 2. Assume that C is a non-separable convex cone in a Fr´echet space L. By [3, I.6.4],
the topology of the Fr´echet space L is generated by an invariant metric dL such that for every ε > 0
the ε-ball BL(ε) = {x ∈ L : dL(x, 0) < ε} centered at the origin is convex. We lose no generality
assuming that the linear subspace C − C is dense in the Fr´echet space L.
Given any ε > 0, we need to construct maps fk : C → C, k ∈ ω, such that each fk is ε-homotopic
to idC and inf k6=n d(fk(C), fn(C)) > 0. Since the metric d has convex balls, any two ε-near maps into
C are ε-homotopic.
Claim 2. There is a linear continuous operator R : L → Y onto a normed space Y such that the
image R(C) is not separable and R−1( ¯BY ) ⊂ BL(ε/2) where ¯BY = {y ∈ Y : kyk ≤ 1} is the closed
unit ball in the normed space Y .
Proof. By [3, I.6.4], the Fr´echet space L can be identified with a closed linear subspace of the countable
product Qi∈ω Xi of Banach spaces. For every n ∈ ω let Yn = Qi<n Xi and prn : L → Xn be the
natural projection. Since C is non-separable, there is n ∈ ω such that for every m ≥ n the image
prm(C) ⊂ Ym is not separable. We can take m ≥ n so large that the ball BL(ε/2) contains the
preimage pr−1
m (U ) of some open neighborhood U ⊂ Ym of the origin. The neighborhood U contains
the closed r-ball ¯BYm(r) of the Banach space Ym for some r > 0. Finally, consider the linear space
Y = prm(L) ⊂ Ym endowed with the norm kyk = 1
r kykm where k · km is the norm of the Banach
space Ym. Then the operator R = prm : L → Y has the desired properties.
(cid:3)
In the convex cone C consider the convex subset BC = C ∩ R−1( ¯BY ) and observe that C = R+ · C1
and hence T (C) = R+ · T (BC). Since the space T (C) is not separable, T (BC) is not separable
too. Consider the convex bounded symmetric subset D = T (BC) − T (BC) ⊂ Y and observe that
R · D = R(C) − R(C) = R(C − C). Then the Minkowski functional
kxkZ = inf{λ > 0 : x ∈ λD}
is a well-defined norm on the linear space Z = R · D = R(C − C) and the identity inclusion I :
Z → Y is a bounded linear operator from the normed space (Z, k · kZ ) to the Banach space Y . Since
I(Z) = Z is non-separable, the operator I is not compact. By Lemma 3, there is a bounded operator
T : Y → c0 such that the composition T I : Z → c0 is not compact. The latter means that the image
T (D) = T R(BC) − T R(BC) is not totally bounded in c0 and hence the bounded set T R(BC) is not
totally bounded in c0.
Consequently, there is δ ∈ (0, 1] such that for every n ∈ ω
(1)
T R(BC) 6⊂ {(xi)i∈ω ∈ c0 : max
i≥n
xi < δ}.
n ∈ c∗
0, e∗
For every n ∈ ω let e∗
n = (T R)∗(e∗
z∗
Claim 3. There are an increasing number sequence (mk)k∈ω and a sequence (zk)k∈ω ⊂ BC such that
for every k ∈ ω;
n : (xi)i∈ω 7→ xn, be the nth coordinate functional of c0 and let
n) ∈ L∗.
(1) z∗
mk (zk) ≥ δ;
TOPOLOGICAL CLASSIFICATION OF CLOSED CONVEX SETS IN FR´ECHET SPACES
5
(2) z∗
mi(zk) < δ3/100 for all i > k.
Proof. The sequences (mk) and (zk) will be constructed by induction. By (1) there are a point
z0 ∈ BC and a number m0 ∈ ω such that e∗
m0 (z0) ≥ δ. Now assume that for some k ∈ ω points
z0, . . . , zk and numbers m0 < m1 < · · · < mk have been constructed. Since the points T R(zi), i ≤ k,
belong to the Banach space c0, there is a number m > mk so large that e∗
n(T R(zi)) < δ3/100 for
all n ≥ m and i ≤ k. By (1), there are a point zk+1 ∈ BC and a number mk+1 ≥ m such that
z∗
(cid:3)
mk (zk+1) = e∗
Divide ω into the countable union ω = Sk∈ω Nk of pairwise disjoint infinite subsets and by induction
define a function ξ : ω × ω → ω such that ξ(i, k) ∈ Nk and ξ(i + 1, k) > ξ(i) > i for all i, k ∈ ω. For
any numbers i, k ∈ ω let
mk (T R(zk+1)) ≥ δ. This complete the inductive step.
zi,k := zξ(i,k) and z∗
i,k := z∗
mξ(i,k)
= (T R)∗(e∗
mξ(i,k)
),
where (zi)i∈ω and (mk)k∈ω are given by Claim 3. It follows that the double sequences (zi,k)i,k∈ω and
(z∗
i,k)i,k∈ω have the following properties (that will be used in the proof of Claim 8 below):
Claim 4. If (i, k), (j, n) ∈ ω × ω, then
(1) z∗
(2) z∗
(3) z∗
i,k(zi,k) ≥ δ;
j,k(zi,n) < δ3/100 provided ξ(j, k) > ξ(i, n);
i,k(z) ≤ 1 for any z ∈ BC .
Claim 5. There is a map f : C → C such that d(f, id) < ε/2 and each point x ∈ C has a neighborhood
O(x) whose image f (O(x)) lies in the convex hull conv(Fx) of some finite subset Fx ⊂ C.
Proof. Using the paracompactness of the metrizable space C, find a locally finite open cover U of
X that refines the cover of C by open ε
In each set U ∈ U pick up a point cU ∈ U . Let
{λU : C → [0, 1]}U ∈U be a partition of the unity, subordinated to the cover U in the sense that
λ−1
U ((0, 1]) ⊂ U for all U ∈ U . Finally, define a map f : C → C by the formula
4 -balls.
f (x) = X
U ∈U
λU (x)cU .
It is standard to check that f has the desired property.
(cid:3)
For every k ∈ Z by Ck denote the set of points x ∈ C that have neighborhood O(x) ⊂ C such that
mf (x′) < δ3/100. It is clear
for each point x′ ∈ O(x) and a non-negative number m ≥ k we get z∗
that each set Ck is open in C and lies in Ck+1.
nT R(z) < δ3/100 for all z ∈ conv(F ), in particular, e∗
Claim 6. C = Sk∈ω Ck.
Proof. By Claim 5, each point x ∈ C has a neighborhood O(x) ⊂ C such that f (O(x)) ⊂ conv(F )
for some finite subset F ⊂ C. Taking into account that T R(F ) is a finite subset of the Banach space
nT R(z) < δ3/100 for all n ≥ m and all z ∈ F . Then
c0, we can find a number m ∈ ω such that e∗
nT Rf (x′) < δ3/100 for any x′ ∈ O(x).
also e∗
This means that x ∈ Cm by the definition of the set Cm.
(cid:3)
Claim 7. There is an open cover (Uk)k∈ω of the space C such that Uk ⊂ ¯Uk ⊂ Ck−1 ∩ Uk+1 for all
k ∈ ω.
Proof. By Theorem 5.2.3 of [6], there is an open cover (Vk)k∈ω of X such that ¯Vk ⊂ Ck ∩ Vk+1 for all
k ∈ ω. For each x ∈ X find the smallest number k ∈ ω with x ∈ Ck and the largest number n ≤ k
with x /∈ ¯Vn and put O(x) = Ck \ ¯Vn. Consider the open cover W0 = {O(x) : x ∈ C} and observe that
St( ¯Vk, W0) ⊂ Ck for every k ∈ ω. Here St(A, W0) = ∪{W ∈ W0 : W ∩ A 6= ∅} for a subset A ⊂ C.
6
TARAS BANAKH AND ROBERT CAUTY
Using the paracompactness of the space C, for every n ∈ ω by induction find an open cover Wn
of C whose star St(Wn+1) = {St(W, Wn+1) : W ∈ Wn+1} is inscribed into the cover Wn. Then the
open sets
Uk = St( ¯Vk−1 ∪ Uk−1, Wk+1), k ∈ ω,
have the required property: Uk ⊂ ¯Uk ⊂ Ck−1 ∩ Uk+1 for all k ∈ ω.
(cid:3)
By Theorem 5.1.9 of [6] there us a partition of unity {λk : C → [0, 1])k∈ω, subordinated to the
k (0, 1] ⊂ Uk+1 \ ¯Uk−1 for all k ∈ ω (here we assume
cover {Uk+1 \ ¯Uk−1}k∈ω of C in the sense that λ−1
that Uk = ∅ for k < 0).
Now, for every k ∈ ω define a map fk : C → C by the formula
fk(x) = f (x) + X
i∈ω
λi(x)zi,k = f (x) + λi(x)zi,k + (1 − λi(x))zi+1,k,
where i is the unique number such that x ∈ Ui+1 \ Ui. Since fk(x) − f (x) ∈ BC ⊂ Bd(ε/2), we
conclude that d(f (x), fk(x)) < ε/2 and hence
d(x, fk(x)) ≤ d(x, f (x)) + d(f (x), fk(x)) <
ε
2
+
ε
2
= ε
for all x ∈ C. So, each function fk : C → C is ε-near and ε-homotopic to the identity idC : C → C.
Claim 8. The family (fk(C))k∈ω is separated.
Proof. By the continuity of the operator T R : L → c0, there is η > 0 such that T R(BL(η)) ⊂
Bc0(δ3/20). We claim that inf n6=k d(fn(C), fk(C)) ≥ η.
Fix any distinct numbers n, k ∈ ω and points x, y ∈ C. By the choice of η, the inequality
d(fk(x), fn(y)) ≥ η will follow as soon as we check that kT R(fk(x) − fn(y))k > δ3/20. The latter
mT R(fk(x) − fn(y)) > δ3/20. Since
inequality will follow as soon as we find m ∈ ω such that e∗
m(fk(x) − fn(y)) > δ3/20 for some m ∈ ω.
mT R(z) = z∗
e∗
Since C = Si∈ω Ui+1\Ui, there are unique numbers i, j ∈ ω such that x ∈ Ui+1\Ui and y ∈ Uj+1\Uj.
m(z) for all z ∈ L, it suffices to show that z∗
Then
fk(x) = f (x) + λi(x)zi,k + λi+1(x)zi+1,k,
fn(y) = f (y) + λj(y)zj,n + λj+1(y)zj+1,n.
Without loss of generality, ξ(i + 1, k) < ξ(j + 1, n).
Since x, y ∈ Umax{i,j}+1 ⊂ Cmax{i,j}, we conclude that
δ3
100
m(f (x)), z∗
m(f (y))} <
max{z∗
(2)
according to the definition of the set Cmax{i,j}.
for all m ≥ max{i, j}
We shall consider five cases.
1) λj+1(y) > δ2/10. In this case, put m = mξ(j+1,n) and observe that z∗
m(zj+1,n) = z∗
j+1,n(zj+1,n) ≥
δ. Since max{ξ(j, n), ξ(i + 1, k), ξ(i, k)} < ξ(j + 1, k), we conclude that
m(zi,k)} < δ3/100
by Claim 5. It follows from (2) and max{i, j} ≤ max{ξ(i, n), ξ(j, k)} that
m(zi+1,k), z∗
m(zj,n), z∗
max{z∗
max{z∗
m(f (x)), z∗
m(f (y))} <
δ3
100
.
Now we see that
z∗
m(fn(y) − fk(x)) = z∗
m(λj+1(y)zj+1,n + λj(y)zj,n + f (y) − f (x) − λi+1(x)zi+1,k − λi(x)zi,k)
≥ λj+1(y)z∗
m(zj+1,n) − z∗
δ3
100
δ3
20
δ2
10
δ − 5
≥
>
.
m(λj(y)zj,n + f (y) − f (x) − λi(x)zi,k + λi+1(x))zi+1,k)
TOPOLOGICAL CLASSIFICATION OF CLOSED CONVEX SETS IN FR´ECHET SPACES
7
2) λj+1(y) ≤ δ2/10 and ξ(j, n) > ξ(i + 1, k). In this case put m = ξ(j, n). Arguing as in the preced-
m(zi,k} <
m(f (y))} < δ3/100 and max{z∗
m(zi+1,k), z∗
m(f (x)), z∗
ing case, we can show that max{z∗
δ3/100. Then
e∗
m(fn(y) − fk(x)) = e∗
m(λj(y)zj,n + λj+1(y)zj+1,n + f (y) − f (x) − λi+1(x)zi+1,k − λi(x)zi,k)
≥ λj(y)z∗
m(zj,n) − λj+1(y)z∗
δ3
100
δ2
10
− 4
≥ (1 − λj+1(y))δ −
≥ (1 −
δ2
10
)δ −
δ2
10
−
δ3
25
>
δ3
20
.
m(zj+1,n) − z∗
m(f (y) − f (x) − λi(x)zi,k + λi+1(x)zi+1,k)
3) λj+1(y) ≤ δ2/10, ξ(j, n) < ξ(i + 1, k), and λi+1(x) > δ/4. In this case put m = ξ(i + 1, n) and
observe that
z∗
m(fk(x) − fn(y)) ≥ λi+1(x)z∗
m(zi+1,k) − λj+1(y)z∗
m(zj+1,n) − z∗
m(f (x) + λi(x)zi,k − f (y) − λj(y)zj,n)
>
δ
4
δ −
δ2
10
− 4
δ3
100
>
δ3
20
.
4) λj+1(y) ≤ δ2/10, ξ(j, n) < ξ(i + 1, k), λi+1(x) ≤ δ/4, and ξ(i, k) < ξ(j, n). In this case put
m = mξ(j,n) and observe that λj(y) = (1 − λj+1(y)) > 1 − δ2/10 ≥ 9/10 and thus
z∗
m(zj,n) − λj+1(y)z∗
m(fk(x) − fn(y)) ≥ λj(y)z∗
m(zj+1,n) − λi+1(x)z∗
δ3
δ2
20
10
δ3
100
9
10
− 3
δ −
δ
4
−
≥
>
.
m(zi+1,k) − z∗
m(f (x) − f (y) − λi(x)zi,k)
5) λj+1(y) ≤ δ2/10, ξ(j, n) < ξ(i + 1, k), λi+1(x) ≤ δ/4, and ξ(i, k) > ξ(j, n). In this case put
m = mξ(i,k) and observe that λi(x) = 1 − λi+1(x) ≥ 1 − δ/4 ≥ 3/4. Then
z∗
m(fk(x) − fn(y))
m(zi,n) − λi+1(x)z∗
≥ λi(x)z∗
δ3
δ
20
4
δ3
100
δ2
10
− 3
δ −
3
4
−
≥
>
.
m(zi+1,k) − λj+1(y)z∗
m(zj+1,n) − z∗
m(f (x) − f (y) − λj(y)zj,k)
3. Proof of Theorem 5
Given a non-separable convex set X is a Fr´echet space L, consider the convex cone
in L × R with base X × {1} which will be identified with X.
C = {(tx, t) : x ∈ X, t ∈ [0, +∞)} ⊂ L × R
(cid:3)
By pr : C → R+, pr : (x, t) 7→ t, we denote the projection onto the second coordinate. Observe
that the map r : C \ {0} → X, r : (x, t) 7→ x/t, determines a retraction of C \ {0} onto X. This
retraction restricted to the set C[ 1
3 , 3]) is a perfect map.
3 ,3] = pr−1([ 1
To prove that X has LFAP, fix an open cover U of X. For each open set U ∈ U consider the set
U = {(tx, t) : x ∈ U, 1
3 < t < 3}. Then U = {pr−1(R \ [ 1
2 , 2]), U : U ∈ U} is an open cover of C.
By Lemma 2, the convex cone C has SAP and by Lemma 1, C has LFAP. Consequently, there is
a map f : C × ω → C such that f is U -near to the projection f0 : C × ω → C, f0 : (x, n) 7→ x, and
the family (cid:0)f (C × {n})(cid:1)n∈ω is locally finite in C. Let f = f X × ω and f0 = f0X × ω. It follows
3 ,3] and the map g = r ◦ f : X × ω → X is U -near
from the choice of the cover U that f (X × ω) ⊂ C[ 1
to the projection f0 : X × ω → X.
Since the family ( f (X × {n})n∈ω is locally finite in C 1
3 ,3] and the map r : C[ 1
the family (r ◦ f (X × {n})n∈ω is locally finite in X, witnessing that X has LFAP.
3 ,3] → X is perfect,
8
TARAS BANAKH AND ROBERT CAUTY
4. Open Problems
The proof of Theorem 5 heavily exploits the machinery of Banach space theory and does not work
in the non-locally convex case. This leaves the following problem open:
Problem 1. Is each non-separable completely metrizable convex AR-subset of a linear metric space
homeomorphic to a Hilbert space?
Even a weaker problem seems to be open:
Problem 2. Is each complete linear metric AR-space homeomorphic to a Hilbert space?
This is true in the separable case, see [4], [5].
References
[1] I. Banakh, T. Banakh, Constructing non-compact operators into c0, preprint.
[2] T. Banakh, I. Zarichnyy, Topological groups and convex sets homeomorphic to non-separable Hilbert spaces, Cent.
Eur. J. Math. 6:1 (2008), 77 -- 86.
[3] C. Bessaga, A. Pelczynski, Selected topics in infinite-dimensional topology, PWN, Warsaw, 1975.
[4] T. Dobrowolski, H. Torunczyk, On metric linear spaces homeomorphic to l2 and compact convex sets homeomorphic
to Q, Bull. Acad. Polon. Sci. Ser. Sci. Math. 27:11-12 (1979), 883 -- 887.
[5] T. Dobrowolski, H. Toru´nczyk, Separable complete ANR's admitting a group structure are Hilbert manifolds, Topol-
ogy Appl. 12 (1981), 229 -- 235.
[6] R.Engelking, General topology, Heldermann Verlag, Berlin, 1989.
[7] R.Geoghegan, Open problems in infinite-dimensional topology, Topology Proc. 4:1 (1979), 287 -- 338.
[8] V. Klee, Some topological properties of convex sets, Trans. Amer. Math. Soc. 78 (1955), 30 -- 45.
[9] H. Torunczyk, Characterizing Hilbert space topology, Fund. Math. 111:3 (1981), 247 -- 262.
[10] H. Torunczyk, A correction of two papers concerning Hilbert manifolds, Fund. Math. 125:1 (1985), 89 -- 93.
[11] J.West, Open problems in infinite-dimensional topology, in: Open problems in topology, North-Holland, Amsterdam,
1990, P. 523 -- 597.
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]
Universit´e Paris VI (France)
E-mail address: [email protected]
|
1701.01131 | 3 | 1701 | 2017-01-13T04:31:21 | A note on spaces of continuous functions on compact scattered spaces | [
"math.FA"
] | In 1959, Pelczynski and Semadeni proved a theorem in which they gave some equivalent conditions for a compact Hausdorff space to be scattered. The purpose of the current note is that to clarify the meaning of the subtle term "conditionally weakly sequentially compact" they used as the basis for the proof of their theorem. Unfortunately, the term now is taken over by a similar but subtle concept that may cause a serious problem. | math.FA | math |
A NOTE ON SPACES OF CONTINUOUS FUNCTIONS ON COMPACT
SCATTERED SPACES
FOUAD NADERI
Abstract. In 1959, Pelczynski and Semadeni proved a theorem in which they gave
some equivalent conditions for a compact Hausdorff space to be scattered. The purpose
of the current note is that to clarify the meaning of the subtle term "conditionally
weakly sequentially compact" they used as the basis for the proof of their theorem.
Unfortunately, the term now is taken over by a similar but subtle concept that may
cause a serious problem.
1. The main result
A locally compact Hausdorff topological space Ω is scattered if Ω does not contain any
non-empty perfect subset (i.e., a closed non-empty subset Π of Ω such that each point of
Π is an accumulation point of Π). Equivalently, any non-empty subset of Ω contains at
least one isolated point. Some authors use the term dispersed instead of scattered. For
more details on scattered spaces see [2], [3] and [4].
Example 1.1. Consider N with its usual topology and its one point compactification
N∗ = N ∪ {∞}. Both N and N∗ are scattered. Also, { 1
n : n ∈ N} ∪ {0} is another compact
It can be shown that a compact metric space is
scattered space which is not discrete.
scattered if and only if it is countable [2, p.737]. A scattered space is always totally
disconnected. The converse of this is not true as seen by the set Q of rational numbers.
Consider the following two definitions for the term conditionally weakly sequentially
compact.
Definition 1.2. Let Ω be a compact Hausdorff space. We say that C(Ω) has condi-
tionally weakly sequentially compact property in the sense of ES in the sense that for
every bounded sequence (xn) of elements of C(Ω) there exists a subsequence (xnk ) and a
member x0 ∈ C(Ω) such that for every bounded linear functional ξ on C(Ω) the sequence
of numbers ξ(xnk ) converges to ξ(x0). In other words, if S is a bounded subset of C(Ω),
then S must be conditionally (=relatively) weakly sequentially compact in the modern
language.
Definition 1.3. Let Ω be a compact Hausdorff space. We say that C(Ω) has conditionally
weakly sequentially compact property in the sense of PS if for every bounded sequence
(xn) of elements of C(Ω) there exists a subsequence (xnk ) such that for every bounded
linear functional ξ on C(Ω) the sequence of numbers ξ(xnk ) is convergent.
Remark 1.4. The way Definition 1.3 seems is now clear, but it was after a correspondence
with Professor Semadeni that I could write down it this way. In the Definition 1.2, the
space is weakly sequentially complete while in the second one we do not need such
2000 Mathematics Subject Classification. Primary 46A03; Secondary 46A22.
Key words and phrases. Compact Hausdorff scattered space, Separable Banach space, Weakly compact
set.
1
2
F. NADERI
a strong condition (not to the point x0). Meanwhile, we use Definition 1.2 in Eberlin-
Smulian theorem to assure weakly compactness of a given set.
Theorem 1.5. Let Ω be a compact Hausdorff space and C(Ω) have conditionally weakly
sequentially compact property in the sense of ES. Then Ω is finite.
Proof. Suppose to the contrary that Ω is infinite. Then, the Banach space of continuous
functions C(Ω) is infinite dimensional. It is well-known that the weak closure of the unit
sphere of an infinite dimensional Banach space is the closed unit ball of the space [1, p.128].
Therefore, if S is the unit sphere of C(Ω), then S
is equal to the unit ball B of C(Ω).
Since C(Ω) has propert ES, S
= B is weakly sequentially compact. According to
Eberlin-Smulian Theorem [1, p.163], B must be weakly compact. Therefore by [1, Theorem
4.2, p.132], C(Ω) must be reflexive. By [1, p.90], Ω must be finite. But, this contradicts
our assumption that Ω was infinite. Hence, Ω can only be a finite set.(cid:4)
wk
wk
Remark 1.6. Suppose Ω is a compact Hausdorff space and the main theorem of [3, p.214]
holds. Condition (0) and (9) of the latter theorem assert that Ω is scattered if and only
if C(Ω) has conditionally weakly sequentially compact property in the sense of PS. But,
if one wants to classify scatteredness of Ω in the sense of EP, (s)he would always end up
with a finite set! which is not always the case as Example 1.5 indicates.
Acknowledgments. The author would like to thank Professor Z. Semadeni for his
careful comments. He would also thanks Professor W. Zelazko who made this corespon-
dents possible.
References
[1] J. B. Conway, A course in functional analysis, Springer-Verlag, New York, 1990.
[2] V. Montesinos, P. Zizler and V. Zizler, An Introduction to Modern Analysis, Springer International
Publishing, Switzerland, 2015.
[3] A. Pelczinsky and Z. Semadeni, Spaces of continuous functions III, Studia. Math. 18 (1959), 211-222.
[4] Z. Semadeni, Banach spaces of continuous functions, Vol. I. Monografie Matematyczne, PWN, Warsaw,
1971.
Department of Mathematical and Statistical Sciences, University of Alberta, Edmonton,
Alberta, T6G 2G1, Canada.
E-mail address: [email protected]
|
1507.07291 | 1 | 1507 | 2015-07-27T03:42:00 | Sampling time-frequency localized functions and constructing localized time-frequency frames | [
"math.FA"
] | We study functions whose time-frequency content are concentrated in a compact region in phase space using time-frequency localization operators as a main tool. We obtain approximation inequalities for such functions using a finite linear combination of eigenfunctions of these operators, as well as a local Gabor system covering the region of interest. These would allow the construction of modified time-frequency dictionaries concentrated in the region. | math.FA | math |
SAMPLING TIME-FREQUENCY LOCALIZED FUNCTIONS AND
CONSTRUCTING LOCALIZED TIME-FREQUENCY FRAMES
MONIKA D ORFLER AND GINO ANGELO VELASCO
Abstract. We study functions whose time-frequency content are concentrated in a
compact region in phase space using time-frequency localization operators as a main
tool. We obtain approximation inequalities for such functions using a finite linear com-
bination of eigenfunctions of these operators, as well as a local Gabor system covering
the region of interest. These would allow the construction of modified time-frequency
dictionaries concentrated in the region.
1. Introduction
When processing audio signals such as music or speech, one sometimes strives for
a meticulous separation of signal components in time and frequency. However, it is
known that no nonzero function can be compactly supported simultaneously in time and
frequency, so that good concentration in one domain usually has to be paid for with
increased leakage in the other domain. Due to this trade-off, window design has been an
important issue in signal processing. As opposed to traditional approaches, we investigate
the optimization of concentration simultaneously in time and frequency. More precisely,
we investigate functions that exhibit good concentration in a compact region in the time-
frequency plane. Our method is related to the approach introduced by Landau, Slepian,
and Pollak, who considered operators composed of consecutive time- and bandlimiting
steps, cf.
[25, 18, 19]. The resulting operators yield the well-known prolate spheroidal
functions as eigenfunctions. These functions satisfy some optimality in concentration in
a rectangular region in the time-frequency domain.
In [4], Daubechies introduced time-frequency localization operators obtained by re-
stricting the synthesis from time-frequency coefficients to a desired region of interest.
Here, we make use of time-frequency localization operators to describe a function's lo-
cal time-frequency content in regions more general than the rectangles considered in
[25, 18, 19]. As in the case of the prolate spheroidal wave functions, the eigenfunctions
of the time-frequency localization operators are maximally time-frequency-concentrated
in the region of interest. We will use these eigenfunctions to characterize a function's
time-frequency localization.
Using Gabor frames, we show, to which extent time-frequency localized functions can
be approximated using only a finite number of Gabor coefficients, namely, those which
are inside some larger cover of the given region. This is influenced by the approximation
result formulated by Daubechies in a seminal paper [5]. Similar estimates were also
established in [22].
Using the obtained results, we construct global time-frequency frames consisting of
atoms which are optimally concentrated in small regions corresponding to a prescribed
The first author was supported by the Vienna Science and Technology Fund (WWTF) through project
MA14-018. The second author was partially supported by the FWF Einzelprojekte P 27773.
1
2
M. D ORFLER AND G. VELASCO
lattice. Each locally concentrated system is constructed by projecting the local Ga-
bor atoms onto the subspace spanned by a finite number of eigenfunctions. Then, by
considering a family of such locally concentrated time-frequency dictionaries, we obtain
an adaptive frame for all L2(R). The resulting frames are useful for processing tasks
in which, as explained above, certain signal components have to be processed separately
with minimum distortion of close-by signals parts. We will give an application example in
the experiments Section 5. Note that a similar construction involving prolate spheroidal
functions has recently been studied in [16].
In the next section, we recall the necessary tools from time-frequency analysis, namely,
the short-time Fourier transform and Gabor frames. In Section 3, we review some prop-
erties of time-frequency localization operators, and we prove characterization and ap-
proximation results concerning time-frequency localized functions and eigenfunctions of
time-frequency localization operators. Section 4 deals with approximation using a local
Gabor system and the construction of the new time-frequency dictionaries. Finally, in
Section 5, we present numerical experiments concerning our results and conclude with
some perspectives in Section 6.
2. Preliminaries
In this section we recall some definitions and properties about the short-time Fourier
transform and Gabor frames. For a thorough introduction to the field of time-frequency
analysis, we refer the reader to [14].
Fix a window function ϕ ∈ L2(R) with norm kϕk2 = 1. The short-time Fourier
transform (STFT) of f ∈ L2(R) with respect ϕ is given by
(1)
Vϕf (z) = Vϕf (x, ω) =ZR
f (t) ϕ(t − x) e−2πiω·tdt = hf, π(z)ϕi,
where z = (x, ω) ∈ R2 and π(z) is the time-frequency shift operator given by π(z)f =
f (t−x) e2πiω·t. The STFT is an isometry from L2(R) to L2(R2), i.e. kVϕfk2 = kϕk2kfk2,
and inversion is realized using the formula
(2)
f = V∗ϕ Vϕf =ZZR2 Vϕf (z)π(z)ϕ dz.
The membership of the STFT in Lp(R2) provides a definition of a class of function
spaces called modulation spaces. In particular, for a fixed non-zero window function ϕ ∈
S(R), the modulation space M p(R) is defined as the space of all tempered distributions
f ∈ S′(R) such that Vϕf ∈ Lp(R2).
It is a Banach space equipped with the norm
kfkM p(R) := kVϕfkLp(R2), where a different window function ϕ would yield an equivalent
norm. Note that for ϕ ∈ L2(R) with kϕk2 = 1, the isometry of the STFT implies that
M 2(R) = L2(R). The space S0(R) = M 1(R) is also known as Feichtinger's algebra. It
is the smallest Banach space isometrically invariant under time-frequency shifts and the
Fourier transform, cf. [13].
Discretization of the time-frequency representation via the STFT leads to the theory of
Gabor frames. We consider a Gabor system G(g, Λ) with window function g ∈ L2(R) and
a countable set of points Λ in R2, consisting of time-frequency shifted copies of a function
g, i.e. G(g, Λ) := {gλ := π(λ)g}λ∈Λ. We say that G(g, Λ) is a frame for f ∈ L2(R) if
there exist constants A, B > 0 such that for all f ∈ L2(R)
hf, gλi2 ≤ Bkfk2
2.
(3)
Akfk2
2 ≤Xλ∈Λ
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
3
If A = B, then we say that G(g, Λ) is a tight frame.
Associated with the frame G(g, Λ) is the frame operator S given by
Sf =Xλ∈Λ
hf, gλigλ.
The frame conditions (3) are equivalent to the invertibility of S, and reconstruction
from the coefficients {hf, gλi}λ∈Λ is possible because of the existence of a dual frame
{ gλ}λ∈Λ, the canonical one being {S−1gλ}λ∈Λ, having frame bounds B−1 and A−1. Every
f ∈ L2(R) will then have the expansion
f =Xλ∈Λ
(4)
hf, gλi gλ =Xλ∈Λ
hf, gλigλ,
A Sf .
where both series converge unconditionally in L2(R). If G(g, Λ) is a tight frame, then
Sf = Af so f = 1
Moreover, if G(g, Λ) is a frame, then the analysis operator U given by U f = {hf, gλi}λ∈Λ
and its adjoint U∗, called the synthesis operator, given by U∗c =Pλ∈Λ cλgλ, c = {cλ}λ∈Λ,
are bounded from L2(R) into ℓ2(Λ) and ℓ2(Λ) into L2(R), respectively, with operator
norms kUkOp = kU∗kOp ≤ √B. For the dual frame { gλ}λ∈Λ, the associated analysis
and synthesis operators are also bounded operators with operator norms not exceeding
1/√A.
3. Time-frequency concentration via the STFT
Time-frequency localization operators as introduced by Daubechies in [4] are built
by restricting the integral in the inversion formula (2) to a subset of R2. Its properties,
connections with other mathematical topics, and applications have been topics in various
works, e.g. [23, 11, 6, 29, 3, 1, 15, 8, 9].
Let Ω be a compact set in R2 and ϕ a window function in L2(R), with kϕk2 = 1. The
time-frequency localization operator H is defined by
(5)
Hf =ZZΩ Vϕf (z)π(z)ϕ dz = V∗ϕ χΩ Vϕf.
Note that while we denote a time-frequency localization only by H, we emphasize that
it is dependent on the window ϕ and the region Ω.
The above integral can be interpreted as the portion of the function f that is essen-
tially contained in Ω. Moreover, the following inner product involving H measures the
function's energy inside Ω:
hHf, fi =ZZΩ Vϕf (z)hπ(z)ϕ, fidz =ZZΩ Vϕf (z)2dz.
We will say that a function f ∈ L2(R) is (ε, ϕ)-concentrated inside Ω if hHf, fi ≥
(1 − ε)kfk2
The time-frequency localization operator H is a compact and self-adjoint operator so
we can consider the spectral decomposition
2 or equivalently h(I − H)f, fi ≤ εkfk2
2, where I is the identity operator.
(6)
Hf =
∞Xk=1
αkhf, ψkiψk,
where {αk}∞k=1 are the positive eigenvalues arranged in a non-increasing order and {ψk}∞k=1
are the corresponding eigenfunctions. By the min-max theorem for compact, self-adjoint
4
M. D ORFLER AND G. VELASCO
operators, the first eigenfunction, has optimal time-frequency concentration inside Ω in
the sense of (3), i.e.
ZZΩ Vϕψ1(z)2dz = max
kfk2=1ZZΩ Vϕf (z)2dz.
In general, the first N eigenfunctions ψ1, ψ2, . . . , ψN form an orthonormal set in L2(R)
having optimal cumulative time-frequency concentration inside Ω:
NXk=1ZZΩ Vϕψk(z)2dz =
NXk=1ZZΩ Vϕφk(z)2dz.
If we let VN be the span of the first N eigenfunctions and if f ∈ VN so f =PN
k=1orthonormal
{φk}N
then
max
(7)
hHf, fi =
αkhf, ψki2 ≥ αN
hf, ψki2 = αNkfk2
2.
NXk=1
NXk=1
k=1hf, ψkiψk,
By contrast, functions which are (1 − αN , ϕ)-concentrated on Ω need not lie in VN .
This implies that a function f in VN is at least (1 − αN , ϕ)-concentrated on Ω, and
any other orthonormal, N -dimensional subspace cannot be better concentrated than
(1 − αN , ϕ)-concentrated on Ω.
The following proposition characterizes a function that is (ε, ϕ)-concentrated on Ω.
Proposition 3.1. Let ϕ, Ω and ε be given and let N0 be the integer such that αN0 ≥ 1−ε
and αN0+1 < 1 − ε. Furthermore, let fker denote the orthogonal projection of f onto the
kernel ker(H) of H. A function f in L2(R) is (ε, ϕ)-concentrated on Ω if and only if
N0Xk=1
(αk + ε − 1)hf, ψki2 ≥
∞Xk=N0+1
(1 − ε − αk)hf, ψki2 + (1 − ε)kfkerk2
2.
Proof : The eigenfunctions {ψk}k form an orthonormal subset in L2(R), possibly incom-
plete if ker(H) 6= {0}; hence, we can write f = P∞j=1hf, ψjiψj + fker and, as in (7),
hHf, fi =P∞k=1 αkhf, ψki2. So the function f is (ε, ϕ)-concentrated on Ω if and only if
αkhf, ψki2 ≥ (1 − ε) ∞Xk=1
∞Xk=1
2! ,
hf, ψki2 + kfkerk2
and the conclusion follows.
∞Pk=N +1
(αN − αk)hf, ψki2 + αNkfkerk2
2.
Remark 3.2. A function f in L2(R2) is (1 − αN , ϕ)-concentrated on Ω if and only if
N−1Pk=1
(αk − αN )hf, ψki2 ≥
In [5, Theorem 3.1], Daubechies bounded the error of a function's local approximation
using a finite number of Gabor atoms by means of an estimate based on the function's
and its Fourier transform's projection onto bounded intervals.
In order to achieve a
similar bound, but for more general regions, in Proposition 4.4 in the next section, we
consider the projection of a function f onto the best-concentrated eigenfunctions of a
localization operator and derive the following estimate. We note that approximations
of bandlimited functions via projections onto eigenspaces of approximately time- and
bandlimited functions were presented in [27, 17].
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
5
Proposition 3.3. Let f be (ε, ϕ)-concentrated on Ω ⊂ R2. For fixed c > 1, let ψk, k =
1, . . . , N , be all eigenfunctions of H corresponding to eigenvalues αk > c−1
c . Then
Proof : Without loss of generality, we assume that kfk2 = 1. We have, by assumption:
2
2
< cεkfk2
2.
hf, ψkiψk(cid:13)(cid:13)(cid:13)(cid:13)
NXk=1
(cid:13)(cid:13)(cid:13)(cid:13)f −
αkhf, ψki2 =ZZΩ Vϕf (z)2dz ≥ (1 − ε)kfk2
2
hHf, fi =
∞Xk=1
Furthermore
We argue by contradiction; to this end, assume that PN
hf, ψki2 + kfkerk2
2 ,
2 = 1 =
kfk2
∞Xk=1
k=1 hf, ψki2 = K < 1 − cε.
hence
We then have
such that
∞Xk=1
∞Xk=N +1
αkhf, ψki2 <
∞Xk=N +1
hf, ψki2 = 1 − K − kfkerk2
2.
c − 1
c
· (1 − K − kfkerk2
2)
αkhf, ψki2 < K +
· (1 − K − kfkerk2
2)
c − 1
c
c − 1 + K
=
< 1 +
−
c
1 − cε − 1
c
2
c − 1
c kfkerk2
c − 1
c kfkerk2
−
2 < 1 − ε,
which is a contradiction. Hence,PN
k=1 hf, ψki2 must be greater than or equal to 1− cε.
We note that while Hf is interpreted as the part of f in Ω, the uncertainty principle
prohibits its STFT to have nonzero values only in Ω, cf. [28], and there will always be
points z ∈ R2 \ Ω at which VϕHf (z) 6= 0. It can be shown, however, that VϕHf (z)
decays fast with respect to the distance of z from Ω. Daubechies proved this result in [4]
for the case where the window function is the Gaussian ϕ0(t) = e−πt2
, showing that the
pointwise magnitude of the STFT decays exponentially.
Lemma 3.4 (Daubechies, [4]). Let H be a time-frequency localization operator over the
region Ω with the Gaussian as its window function, i.e. ϕ = ϕ0. For any δ between 0 and
1, and f ∈ L2(R), one has
hHf, π(z)ϕ0i ≤ 1√2
δ−
1
2kfk2 exp[− π
2 (1 − δ) dist(z, Ω)2].
A similar result involving windows with milder decay conditions is the following.
6
M. D ORFLER AND G. VELASCO
Lemma 3.5. Let ϕ, g ∈ L2(R) such that kϕk2 = 1 and Vϕg(z) ≤ C(1 + z2s)−1, for
some C > 0 and s > 1, for all z ∈ R2. For any δ between 0 and 1, one has
VϕHf (z) = hHf, π(z)gi ≤ Csδ−
where Cs = C√2 π
√s sin(π/s)
.
1
2skfk2(1 + (1 − δ)dist(z, Ω)s)−1,
Remark 3.6. An example of the inequality Vϕg(z) ≤ C(1 + z2s)−1 being satisfied for
all z ∈ R2 is when ϕ and g are in the Schwartz space S(R). Moreover, in that case, for
every s > 0, there is a C for which the inequality is satisfied.
Proof : If z, z′ ∈ R2, then hπ(z′)ϕ, π(z)gi = hϕ, π(z − z′)gi ≤ C′s(1 + z − z′2s)−1.
For 0 < δ < 1,
hHf, π(z)gi ≤ZZΩ hf, π(z′)ϕihπ(z′)ϕ, π(z)gi dz′
1 + z − z′2s dz′
p1 + δz − z′2s
z ′∈Ωz − z′s(cid:18)ZZR2
≤ CZZΩ hf, π(z′)ϕi
≤ CZZΩ hf, π(z′)ϕi
≤ C√2
1 + (1 − δ) inf
1
1
1
1
p1 + (1 − δ)z − z′2s
1 + δz − z′2s dz′(cid:19) 1
1
2
dz′
·
(cid:18)ZZR2 hf, π(z′)ϕi2 dz′(cid:19) 1
2
=
C√2 π
ps sin(π/s)
1
δ−
2s (1 + (1 − δ) inf
z ′∈Ω z − z′s)−1kϕk2kfk2,
and the conclusion follows.
4. Local Gabor approximation and new TF dictionaries
We shall make use of the results concerning time-frequency localization to obtain an
approximation of a function using a finite Gabor expansion. We expect that if the
function is well localized on a region Ω in the time-frequency plane, then f can be
approximated with good accuracy using only the Gabor coefficients on a larger region
covering Ω. From the local Gabor systems, we will obtain frames for the eigenspace, the
collection of which forms a frame for L2(R).
4.1. Time-frequency localization and local Gabor approximation. In this sec-
tion, we consider a given region Ω in R2 and let VN be the N -dimensional subspace
spanned by the first N eigenfunctions of the corresponding localization operator H.
Here, the eigenvalues are assumed to be arranged in descending order. Furthermore, we
let g be a window function in L2(R) such that kgk2 = 1 and Vϕg(z) ≤ C(1 +z2s)−1 for
some C > 0 and s > 1. We then consider the Gabor system G(g, Λ), assume that it forms
a frame with lower and upper frame bounds A and B, respectively, and let { gλ : λ ∈ Λ}
be its dual frame.
We first have the following approximation of a function in VN local Gabor atoms.
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
7
2
≤ εkfk2,
Proposition 4.1. For any ε > 0, there exists an Ω∗ ⊃ Ω such that
(8)
for all f ∈ VN .
Proof : Let ε > 0 and f ∈ VN . We first observe that for any Ω∗ ⊃ Ω,
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)f − Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)
f − Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
hf, ψki(cid:18) Xλ /∈Λ∩Ω∗hψk, gλi gλ(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)
NXk=1
hf, ψki(cid:13)(cid:13)(cid:13)(cid:13) Xλ /∈Λ∩Ω∗hψk, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)2(cid:19)2
≤(cid:18) NXk=1
NXk=1(cid:13)(cid:13)(cid:13)(cid:13) Xλ /∈Λ∩Ω∗hψk, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)
NXk=1
NXk=1 Xλ /∈Λ∩Ω∗ hψk, gλi2.
≤ A−1kfk2
hf, ψki2
≤
2
2
2
2
2
2
We consider hψk, gλi and note that hψk, gλi = 1
C(1 + z2s)−1, it follows from Lemma 3.5 that hψk, gλi ≤ 1
where δ is taken to be 1
2 , which gives us
αk
αkhHψk, gλi. Since g satisfies Vϕg(z) ≤
2 dist(λ, Ω)s)−1,
2s (1 + 1
Cs2
1
2
2 ≤ A−1kfk2
2C 2
s 2
1
s(cid:18) NXk=1
1
α2
k(cid:19) Xλ /∈Λ∩Ω∗
(1 + 1
2 dist(λ, Ω)s)−2.
The right-hand side of the above inequality approaches 0 as Ω∗ gets larger. In particular,
given ε > 0, one can choose Ω∗ so that the sum
(cid:13)(cid:13)(cid:13)(cid:13)f − Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)
Xλ /∈Λ∩Ω∗
(1 + 1
2 dist(λ, Ω)s)−2 < ε2/(cid:18)A−1C 2
s 2
1
α2
k(cid:19)
1
s
NXk=1
which gives the conclusion of the proposition.
We show an example for the case where the window function is the Gaussian ϕ0 and the
region Ω is the disk B(O, R) with center at the origin O and with radius R. We will make
use of the decay of the STFT of H in Lemma 3.4. First, we prove the following lemma that
gives an estimate on the decay of the tail of the sum of samples of the two-dimensional
Gaussian outside the disk B(O, R∗). Let Q(j) = [j1 − 1
2 ), j =
(j1, j2) ∈ Z2.
Lemma 4.2. Let Λ be a relatively separated set of points in R2 with supz∈R2 #(Λ ∩
Q(z)) =: NΛ < ∞. Fix R > 0. If R∗ > R, then
(9)
2 ) × [j2 − 1
2 , j2 + 1
2 , j1 + 1
exp(− π
2 (λ − R)2) ≤ CΛ exp(cid:0) − π
4(cid:0) (R∗)2
4 − R2(cid:1)(cid:1),
Xλ∈Λ, λ>R∗
where CΛ = 8 exp( 5π
Proof : Let R∗ > R and define the sets
4 )NΛ.
JR∗ = {j ∈ Z2 : Q(j) ∩ (R2 \ B(O, R∗)) 6= ∅} and
8
M. D ORFLER AND G. VELASCO
2
(10)
have
e− π
2 (λ−R)2
≤ e− π
exp(− π
exp(− π
estimate (10) as follows:
Xλ∈Λ, λ>R∗
ΛR∗,j = {λ ∈ Λ : λ > R∗, λ ∈ Q(j)}.
We are then able to rewrite the left-hand side of (9) as
2 (λ − R)2) = Xj∈JR∗ Xλ∈ΛR∗,j
2 (λ − R)2).
If λ, z ∈ Q(j), then λ ≥ z − √2. And since −(z − R − √2)2 ≤ −(cid:0) (z−R)2
2 (z−R−√2)2
4(cid:0)z − R(cid:1)2(cid:1).
≤ eπ exp(cid:0) − π
2 + R2 and −(R∗ − √2)2 ≤ − (R∗)2
Using the inequalities −(z − R)2 ≤ −z2
2 (λ − R)2) ≤ Xj∈JR∗ Xλ∈ΛR∗ ,jZZQ(j)
Xj∈JR∗ Xλ∈ΛR∗,j
exp(cid:0) − π
4(cid:0)z − R(cid:1)2(cid:1) dz
≤ NΛ eπZZR2\B(O,R∗−√2)
exp(cid:0) − π
4(cid:0)z − R(cid:1)2(cid:1) dz
4 ZZz>R∗−√2
8 (cid:1) dz
exp(cid:0) − πz2
≤ NΛ eπe
4 exp(cid:0) − π(R∗−√2)2
(cid:1)
= 8NΛ eπe
4 − R2)(cid:1).
4 exp(cid:0) − π
≤ 8NΛ e
4 ( (R∗)2
4 , we get the conclusion of the lemma.
By taking CΛ = 8NΛ e
exp(− π
πR
πR
5π
2
2
5π
8
− 2(cid:1), we
2 + 2, we
: λ ∈ Λ} forms a
Example 4.3. Suppose that the Gabor system {ϕ0,λ := π(λ)ϕ0
frame having upper and lower frame bounds A and B, respectively, and let the system
{gϕ0,λ : λ ∈ Λ} be its dual frame. Then, for any ε between 0 and 1, there exists Rε > 0
such that if Ω is a disk centered at the origin with radius R, the following inequality
holds for all f ∈ VN :
(11)
hf, ϕ0,λigϕ0,λ(cid:13)(cid:13)(cid:13)(cid:13)2 ≤ εkfk2.
π ln(ε2/A−1CΛPN
NXk=1 Xλ /∈Λ∩Ω∗
(cid:13)(cid:13)(cid:13)(cid:13)f − Xλ≤R+Rε
Here, we can take Rε ≥ −R +q4R2 − 16
Proof : Following the proof of Proposition 4.1, we have for Ω∗ ⊃ Ω,
(cid:13)(cid:13)(cid:13)(cid:13)f − Xλ∈Λ∩Ω∗hf, ϕ0,λigϕ0,λ(cid:13)(cid:13)(cid:13)(cid:13)
NXk=1 Xλ /∈Λ∩Ω∗
use Lemma 3.4 (with δ = 1
side as follows:
1
k hHψk, ϕ0,λi2 ≤
α2
2 ≤ A−1kfk2
exp(− π
1
α2
1
α2
k
k=1
2
2
k
We can take Ω∗ to be a disk centered at the origin with radius R∗ := R + Rε > R. We
2 ) and Lemma 4.2 to estimate the double sum on the right
), where CΛ = 8NΛ e
5π
4 .
1
k hHΩ,ϕ0ψk, ϕ0,λi2.
α2
NXk=1 Xλ /∈Λ∩Ω∗
NXk=1 Xλ>R∗
=
2 dist(λ, Ω)2)kψkk2
2
1
α2
k
exp(− π
2 (λ − R)2)
(12)
Proof : Since
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
f − Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
f − Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
A(cid:17)kf − fNk2 + εkfk2.
fN − Xλ∈Λ∩Ω∗hfN , gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
≤(cid:16)1 +q B
≤ kf − fNk2 +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xλ∈Λ∩Ω∗hfN − f, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
9
1
α2
k(cid:19)CΛ exp(cid:0) − π
=(cid:18) NXk=1
4(cid:0) (R∗)2
4 − R2(cid:1)(cid:1) ≤ ε2/A−1CΛPN
k(cid:17), and the conclusion follows.
4(cid:0) (R∗)2
π ln(cid:16)ε2/A−1CΛPN
1
α2
k
1
α2
k=1
4 − R2(cid:1)(cid:1).
Now, exp(cid:0)− π
tor4R2 − 16
k=1
whenever R∗ is greater than or equal
The following proposition gives a localization result by means of a given Gabor frame
whose atoms are known to have sufficient TF-localization, provided by the condition
Vϕg(z) ≤ C(1 + z2s)−1.
Proposition 4.4. For all N > 0 and all ε > 0, there exists a set Ω∗ ⊃ Ω in R2, such that
for all f ∈ L2(R) with corresponding orthogonal projection fN onto the TF-localization
subspace VN , the following estimate holds:
the result follows from Proposition 4.1 and the boundedness of the associated analysis
and synthesis operators.
As a corollary, we obtain the following result for local approximation of functions with
known time-frequency concentration in a given set Ω by Gabor frame elements.
Corollary 4.5. For fixed c > 1, let ψk, k = 1, . . . , N , be the eigenfunctions of H cor-
responding to eigenvalues αk > c−1
c . For N and ε > 0, choose a set Ω∗ ⊃ Ω as in
Proposition 4.4. Then the following approximation holds for all functions f which are
(ε, ϕ)-concentrated on Ω:
(13)
f − Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤(cid:16)1 +r B
A(cid:17) · (√cε + ε)kfk2.
Proof : The result follows immediately from Proposition 3.3 and Proposition 4.4 .
4.2. Local and global frames with TF-localization. The next results deal with the
construction of frames for the subspace VN of eigenfunctions of H, and the whole of
L2(R), respectively. Denote by PN the orthogonal projection operator onto the subspace
VN .
Proposition 4.6. If ε < 1 and inequality (8) is satisfied, then for all f ∈ VN ,
(14)
A(1 − ε)2kfk2
2 ≤ Xλ∈Λ∩Ω∗ hf, gλi2 ≤ Bkfk2
2,
10
M. D ORFLER AND G. VELASCO
where A and B are lower and upper frame bounds, respectively, for G(g, Λ). This im-
plies that the system {PN gλ}λ∈Λ∩Ω∗ forms a frame for VN . More generally, the sys-
tem {π(ν)PN π(λ)g}λ∈Λ∩Ω∗ , where ν ∈ R2, forms a frame for the subspace π(ν)VN :=
{π(ν)f : f ∈ VN}.
Proof : From Proposition 4.1, we get
≤ εkfk2.
And we obtain
kfk2 −(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(1 − ε)2kfk2
≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
f − Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
2 ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xλ∈Λ∩Ω∗hf, gλi gλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
A Xλ∈Λ∩Ω∗ hf, gλi2
B
Akfk2
2.
≤
≤
1
2
2
For the subspace π(ν)VN , we first note that
kfk2 = kπ(ν)fk2 and hf,PN gλi = hπ(ν)f, π(ν)PN gλi.
The inequality in (14) can then be reformulated as
A(1 − ε)2kπ(ν)fk2
2 ≤ Xλ∈Λ∩Ω∗ hπ(ν)f, π(ν)PN gλi2 ≤ Bkπ(ν)fk2
2,
for all f ∈ VN , or π(ν)f ∈ π(ν)VN .
It follows from Proposition 4.6 that any function f ∈ VN can be completely recon-
structed from the samples {hf, gλi}λ∈Λ∩Ω∗ . In [13, Theorem 3.6.16], a reconstruction pro-
cedure was presented where a function on a closed subspace can be reconstructed from re-
stricted Gabor coefficients. Following its approach, we apply the following iterative recon-
struction for functions in VN from the local STFT samples, where Ulocf = {hf, gλi}λ∈Λ∩Ω∗
and U∗locc =Pλ∈Λ∩Ω∗ cλ gλ, c = {cλ}λ∈Λ∩Ω∗ : let f0 = 0 and define recursively
fn = fn−1 + PN U∗loc(hf, gλi − Ulocfn−1).
(15)
We implement the above reconstruction procedure in Section 5, observing the dependence
of the performance of the algorithm on the choice of Ω∗.
Remark 4.7.
(a) Since {PN gλ : λ ∈ Λ ∩ Ω∗} is a frame for VN , in the language of [2], the system
{gλ : λ ∈ Λ ∩ Ω∗} is also called an outer frame for VN . Related terminologies,
e.g. atomic system, resp. pseudoframe for the subspace VN , appear in [12, 20]. In
particular, by Proposition 4.6 and [20, Theorems 2 and 3], the sequence { ]gλ,VN},
where ]gλ,VN = PN (UPN )†gλ is called a dual pseudoframe sequence for VN with
respect to {gλ}. This sequence coincides with a dual frame to the frame {PN gλ}
for VN .
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
11
approximate projection operator HN defined as: HN f :=PN
(b) It may sometimes be more natural to use, instead of the projection PN , the
k=1 αkhf, ψkiψk. Ob-
viously, since we use a finite sequence of positive weights √αk, we obtain an
equivalent frame for the subspace VN , if PN is replaced by HN in Proposition 4.6.
We now consider a family of time-frequency localization operators H µ, µ ∈ Λ over the
region Ωµ with a common window function ϕ. In [9, Theorem 5.10], Dorfler and Romero
showed that under certain conditions on H µ, one can choose Nµ such that
(16)
2 ≍Xµ∈ Λ
kfk2
NµXk=1
hf, ψµ
ki2, f ∈ L2(R),
such that
where the functions {ψµ
k} are eigenfunctions of H µ. We use this result to obtain a frame
for L2(R) consisting of local frame elements on the time-frequency localized subspaces.
Theorem 4.8. Let {Ωµ}µ∈ Λ be a family of compact regions in R2
2 < inf µ∈ Λ Ωµ ≤ supµ∈ Λ Ωµ < ∞ and Pµ∈ Λ χΩµ ≍ 1, and let ϕ ∈ S0(R) such that
kϕk2 = 1.
Corresponding to each µ ∈ Λ, choose lattices Λµ and windows gµ ∈ L2(R) with kgµk2 =
1, Vϕgµ(z) ≤ Cµ(1+z2sµ )−1 for some Cµ > 0 and sµ > 1 such that for all µ the system
G(gµ, Λµ) is a frame for L2(R) with frame bounds Aµ and Bµ.
k}Nµ
Denote by VNµ the span of the first Nµ eigenfunctions {ψµ
k=1 of H µ corresponding
to the Nµ largest eigenvalues, where each Nµ is chosen so that (16) holds.
If 0 < εµ < 1 such that 0 < infµ∈ Λ Aµ(1 − εµ)2 ≤ supµ∈ Λ Bµ < ∞, then there exist
Ω∗µ ⊃ Ωµ such that the global systemSµ∈ Λ{PNµ π(λ)gµ}λ∈Λµ∩Ω∗
is a frame for L2(R).
Proof : Let f ∈ L2(R) and µ ∈ Λ. We first note that the conditions on the regions Ωµ
and ϕ ensure that (16) holds, cf. [9]. Since hf,PNµπ(λ)gµi = hPNµf, π(λ)gµi, it follows
from Proposition 4.6 that for every εµ ∈ (0, 1), there exists Ω∗µ ⊃ Ωµ such that
µ
Aµ(1 − εµ)2
NµXk=1
hf, ψµ
ki2 ≤ Xλ∈Λµ∩Ω∗
µ
hf,PNµπ(λ)gµi2 ≤ Bµ
NµXk=1
hf, ψµ
ki2.
By the assumption that 0 < A := infµ∈ Λ Aµ(1 − εµ)2 ≤ B := supµ∈ Λ Bµ < ∞ and the
equivalence in (16), we get
ki2 ≤Xµ∈ Λ Xλ∈Λµ∩Ω∗
µ
hf,PNµ π(λ)gµi2 ≤ BXµ∈ Λ
NµXk=1
hf, ψµ
ki2,
NµXk=1
AXµ∈ Λ
hf, ψµ
and finally Pµ∈ Λ Pλ∈Λµ∩Ω∗
µ
hf,PNµπ(λ)gµi2 ≍ kfk2
2.
Remark 4.9.
(a) For the special case where each Ωµ is just the translated region µ + Ω, if {ψk}k∈N
is an orthonormal system of eigenfunctions of H, then one can choose N ∈ N
frame, for L2(R).
and Ω∗ ⊃ Ω such thatSλ∈Λ∩Ω∗ G(PN gλ, Λ) is a frame, or a multi-window Gabor
12
M. D ORFLER AND G. VELASCO
(b) This global system forming a frame obtained from local systems is comparable
to quilted Gabor frames introduced by Dorfler in [7], the difference being the the
projection of the time-frequency dictionary elements onto the time-frequency lo-
calized subspaces. In [24], Romero proved results concerning frames for general
spline-type spaces from portions of given frames which provide existence condi-
tions for quilted Gabor frames.
(c) IfPµ∈ Λ χΩµ ≡ 1, each Λµ is a separable lattice, i.e. Λµ = aµZ × bµZ, aµ, bµ > 0,
and the samples {hf, π(λ)gµi}λ∈Λµ∩Ω∗
µ are given, then f can be recovered com-
pletely from the samples if the set of sampling functions G = ∪µ∈ ΛG(gµ, Λµ ∩ Ωµ)
form a quilted Gabor frame. In [10], the authors presented an approximate re-
construction of f from the given samples using approximate projection H µ
in
Nµ
Remark 4.7(b). In particular, the following error
(17)
(cid:13)(cid:13)(cid:13)(cid:13)f −Xµ∈ Λ Xλ∈Λµ∩Ω∗
µ
hf, π(λ)gµiH µ
Nµ
π(λ)gµ(cid:13)(cid:13)(cid:13)(cid:13)2
was estimated, and numerical experiments were performed in comparison with
the method presented in [21] that used truncated Gabor expansions with weighted
coefficients. Through the numerical experiments, it was illustrated how (17) can
be decreased by having a larger region Ω∗µ or a larger number Nµ of eigenfunctions.
5. Numerical Examples
In this section, we consider examples in the finite discrete case (CL, L = 480) that
illustrate the results in the previous sections. The experiments were done in MATLAB
using the NuHAG Matlab toolbox available in the following website:
http://www.univie.ac.at/nuhag-php/mmodule/.
Each point set Λ that we will use for a Gabor frame is a separable lattice aZL × bZL,
where a and b are divisors of L and also called lattice parameters of Λ. The redundancy
of Λ is given by L
. For more details on Gabor analysis in the finite discrete setting, the
ab
reader is referred to the [26].
5.1. Experiment 1. We first examine the approximation of time-frequency localized
signals by a local Gabor system, in particular, functions lying in the N -dimensional
subspace VN of eigenfunctions of H, as shown in Proposition 4.1. In this example, we
take Ω to be a disk centered at the origin with radius 80 and ϕ to be a normalized
Gaussian.
Figure 1 shows the STFT of a signal in VN and the sample points taken over circular
regions with varying radii, each containing Ω.
In each case, the sampling points are
obtained by restricting a lattice with parameters a = b = 20 (redundancy 1.2) over
the circular region. The error of the approximation kPN − SlocPNkOp, where Sloc is a
truncated tight frame operator, is shown in Table 1 below.
We saw in Proposition 4.6 that if ε < 1, corresponding to the operator norm kPN −
SlocPNkOp being less than 1, then the local Gabor system projected into VN forms a
frame for VN so perfect reconstruction is possible by the reconstruction algorithm (15).
The performance of the reconstruction algorithm is shown in Figure 2. As expected, the
larger the covering region, the faster the convergence.
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
13
Cover radius 80
Cover radius 100
200
100
0
−100
−200
200
100
0
−100
−200
−200
0
200
−200
0
200
Cover radius 120
Cover radius 140
200
100
0
−100
−200
200
100
0
−100
−200
−200
0
200
−200
0
200
Figure 1. Sampling points over various enlargements of the covering region.
Cover radius No. of samp. pts. Op. norm error
80
100
120
140
45
77
109
145
0.9650
0.1105
0.0194
0.0031
Table 1. Error kPN − SlocPNkOp over varying radii for the disk Ω
r
o
r
r
e
e
v
i
t
l
a
e
R
Convergence of reconstruction algorithm
radius 80
radius 100
radius 120
radius 140
100
10−2
10−4
10−6
10−8
10−10
10−12
10−14
10−16
2
4
Number of iterations
6
8
10
Figure 2. Convergence of the reconstruction algorithm from the local
samples with the same lattice parameters but with varying radii of the
covering regions.
14
M. D ORFLER AND G. VELASCO
5.2. Experiment 2. In this next experiment, we look at an example of how the collection
of local Gabor systems can form a frame given that the sum of the characteristic functions
over the regions is bounded above and below by a positive number. Figure 3 shows ten
regions in the TF-plane and Figure 4 shows its sum.
Region 1
Region 2
Region 3
Region 4
Region 5
200
100
0
−100
−200
200
100
0
−100
−200
200
100
0
−100
−200
200
100
0
−100
−200
200
100
0
−100
−200
−200
0
200
−200
0
200
−200
0
200
−200
0
200
−200
0
200
Region 6
Region 7
Region 8
Region 9
Region 10
200
100
0
−100
−200
200
100
0
−100
−200
200
100
0
−100
−200
200
100
0
−100
−200
200
100
0
−100
−200
−200
0
200
−200
0
200
−200
0
200
−200
0
200
−200
0
200
Figure 3. Ten regions that partition the time-frequency plane.
Sum of the characteristic functions
4
−200
−150
−100
−50
0
50
100
150
200
−200
−100
0
100
200
3.5
3
2.5
2
1.5
1
Figure 4. Sum of the characteristic functions over the ten regions.
Sample points are then taken over sets that contain each region, where different lattices
are used for each set. The lattice parameters assigned to each set are summarized in Table
2, and the sample points are depicted in Figure 5. The left image shows sample points
obtained by restricting each lattice over the regions themselves, while the samples in the
right image are obtained from the restriction over larger sets containing each region, thus
producing more overlap. Tight windows are used corresponding to each set of restricted
lattice points.
We form a quilted Gabor frame from the collection of local Gabor systems. And
by projecting each local Gabor system onto the local subspace corresponding to each
region, we likewise obtain a global frame as in Theorem 4.8. The average of the relative
error kf − Sif k2
when the frame operators S1 and S2, corresponding to the quilted Gabor
frame (i.e. without projection) and the global frame (i.e. with projection), respectively,
are applied to a random signal f are shown in Table 3.
kf k2
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
15
Region
(a, b)
Region
(a, b)
1
2
3
4
5
(20, 20)
(16, 20)
(20, 16)
(16, 16)
(15, 16)
6
7
8
9
10
(15, 15)
(12, 15)
(12, 12)
(10, 12)
(10, 10)
Table 2. Lattice parameters over the different regions.
Sampling points, less overlap
Sampling points, more overlap
200
150
100
50
0
−50
−100
−150
−200
200
150
100
50
0
−50
−100
−150
−200
−200
−100
0
100
200
−200
−100
0
100
200
Figure 5. Sampling points on the different local patches.
without projection with projection
Less overlap
More overlap
0.2610
0.5840
0.1687
0.1709
Table 3. Average of the error in applying the frame operator to a random
signal (average of 1000 attempts).
In both cases of less and more overlap, projecting onto the TF-localized subspaces
decreases the relative error between the signal and the approximation by the frame oper-
ator. Note that in both quilted Gabor frame and the global frame with projection, having
more overlap increases the relative error since we are just comparing f with Sif . Since
we are dealing with frames, perfect reconstruction (up to numerical error) is possible via
the frame algorithm cf. [14, Algorithm 5.1.1].
We first compare the respective condition numbers of the frame operators for the
cases of less and more overlap. The values are shown in Table 4. Once again, in both
quilted Gabor frame and the global frame with projection, having more overlap improves
the condition number. Note that the large condition number for the frame operator
corresponding to the global frame with less overlap can be attributed to the lower frame
bound in Theorem 4.8, which is related to the set Ω∗µ that covers the region Ωµ - a smaller
region Ω∗µ implies a smaller lower frame bound.
Figure 6 compares the convergence of the frame algorithm for the four cases considered.
16
M. D ORFLER AND G. VELASCO
without projection with projection
Less overlap
More overlap
5.1429
3.5472
16.0406
1.9845
Table 4. Condition numbers of the resulting frame operators.
Convergence of the frame algorithm
Less Overlap, No Proj.
Less Overlap, With Proj.
More Overlap, No Proj.
More Overlap, With Proj.
100
10−2
r
o
r
r
e
e
v
i
t
l
a
e
R
10−4
10−6
10−8
10−10
20
40
60
Number of iterations
80
100
Figure 6. Convergence of the frame algorithm.
5.3. Experiment 3. In this final experiment, we illustrate the approximate reconstruc-
tion of a signal from analysis coefficients obtained from a union of tight Gabor systems,
each restricted on an enlarged region covering a given region of interest, forming a quilted
Gabor frame (see Remark 4.9(c)).
Similar to the experiment in [10] we consider four rectangular regions and associate
tight Gabor frames to each one:
1. G(g1, 20, 8) on the region corresponding to lower frequency and time t ≤ L/2;
2. G(g2, 24, 10) on the region corresponding to lower frequency and time t > L/2;
3. G(g3, 12, 24) on the region corresponding to higher frequency and time t ≤ L/2;
4. G(g4, 15, 20) on the region corresponding to higher frequency and time t > L/2.
The sample points on the four regions are depicted in Figure 7.
We apply an approximate projection onto the subspaces of eigenfunctions of the time-
frequency localization operators on the regions and compute the relative error from the
approximate reconstruction. We compare the relative errors over varying overlap b (the
amount of increase in the length of a side of the rectangular region) in Figure 8 for three
different eigenspace dimensions (nEV). We also include the relative error obtained from
re-synthesizing with the same quilted Gabor frame elements.
For the case of quilted Gabor frames, an overlap would initially decrease the relative
error in the approximate reconstruction but since we are just essentially getting the
error from applying the frame operator on the signal, more overlap in the regions would
eventually lead to an increase in the relative error. For the cases with approximate
projection, we see the decrease in the relative error as the overlap amount increases.
Moreover, the relative error improves as the eigenspace dimension is increased.
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
17
Sampling points
200
150
100
50
0
−50
−100
−150
−200
50
100
150
200
250
300
350
400
450
Figure 7. Sampling points over the four regions.
Approximation error vs. amount of overlap
via quilted Gabor frames
via approx. projection, nEV = 140
via approx. projection, nEV = 144
via approx. projection, nEV = 147
100
10−1
10−2
r
o
r
r
e
S
M
R
10−3
0
20
40
60
Amount of overlap b
80
100
120
Figure 8. Approximation error vs. amount of overlap.
6. Conclusions and Perspectives
In this paper we investigated the representation of time-frequency localized functions
by means of sampling in the time-frequency domain. Motivated by the problem of pro-
viding dictionaries with good concentration within prescribed regions in time-frequency,
we constructed frames of localized time-frequency atoms. The improved localization is
obtained by means of projections onto eigenspaces corresponding to time-frequency lo-
calization operators. Numerical experiments illustrated the promising potential of the
proposed method: providing good reconstruction/approximation quality while preserving
the good localization property. However, for applications to real signals, the proposed
method would entail having to compute for eigenvalues and eigenfunctions of large ma-
trices, which may be numerically cumbersome. The study of efficient numerical methods
for the evaluation/approximation of eigenvalues and eigenfunctions of time-frequency
localization operators would be a topic of future work.
18
M. D ORFLER AND G. VELASCO
References
[1] L. D. Abreu and M. Dorfler. An inverse problem for localization operators. Inverse Problems,
28(11):115001, 16, 2012.
[2] A. Aldroubi, C. A. Cabrelli, and U. Molter. Wavelets on irregular grids with arbitrary dilation
matrices, and frame atoms for L2(Rd). Appl. Comput. Harmon. Anal., Special Issue on Frames
II.:119–140, 2004.
[3] E. Cordero and K. Grochenig. Time-frequency analysis of localization operators. J. Funct. Anal.,
205(1):107–131, 2003.
[4] I. Daubechies. Time-frequency localization operators: a geometric phase space approach. IEEE
Trans. Inform. Theory, 34(4):605–612, July 1988.
[5] I. Daubechies. The wavelet transform, time-frequency localization and signal analysis. IEEE Trans.
Inform. Theory, 36(5):961–1005, 1990.
[6] F. DeMari, H. G. Feichtinger, and K. Nowak. Uniform eigenvalue estimates for time-frequency
localization operators. J. London Math. Soc., 65(3):720–732, 2002.
[7] M. Dorfler. Quilted Gabor frames - A new concept for adaptive time-frequency representation. Adv.
in Appl. Math., 47(4):668–687, Oct. 2011.
[8] M. Dorfler and J. L. Romero. Frames of eigenfunctions and localization of signal components. In Pro-
ceedings of the 10th International Conference on Sampling Theory and Applications (SampTA2013),
Bremen, July 2013.
[9] M. Dorfler and J. L. Romero. Frames adapted to a phase-space cover. Constr. Approx., 39(3):445–
484, 2014.
[10] M. Dorfler and G. Velasco. Adaptive Gabor frames by projection onto time-frequency subspaces. In
Proceedings of the 39th IEEE International Conference on Acoustics, Speech, and Signal Processing
(ICASSP 2014), pages 3097–3101, 2014.
[11] H. G. Feichtinger and K. Nowak. A Szego-type theorem for Gabor-Toeplitz localization operators.
Michigan Math. J., 49(1):13–21, 2001.
[12] H. G. Feichtinger and T. Werther. Atomic systems for subspaces. In Proceedings SampTA, volume
2001, pages 163–165, Orlando, 2001.
[13] H. G. Feichtinger and G. Zimmermann. A Banach space of test functions for Gabor analysis. In H. G.
Feichtinger and T. Strohmer, editors, Gabor Analysis and Algorithms: Theory and Applications,
Applied and Numerical Harmonic Analysis, pages 123–170, Boston, MA, 1998. Birkhauser Boston.
[14] K. Grochenig. Foundations of Time-Frequency Analysis. Appl. Numer. Harmon. Anal. Birkhauser
Boston, Boston, MA, 2001.
[15] K. Grochenig and J. Toft. The range of localization operators and lifting theorems for modulation
and Bargmann-Fock spaces. Trans. Amer. Math. Soc., 365:4475–4496, 2013.
[16] J. Hogan and J. Lakey. Frame properties of shifts of prolate spheroidal wave functions. Appl. Comput.
Harmon. Anal., 39(1):21 – 32, 2015.
[17] J. A. Hogan, S. Izu, and J. D. Lakey. Sampling approximations for time- and bandlimiting. Sampl.
Theory Signal Image Process., 9(1-3):91–117, 2010.
[18] H. J. Landau and H. O. Pollak. Prolate spheroidal wave functions, Fourier analysis and uncertainty
II. Bell System Tech. J., 40:65–84, 1961.
[19] H. J. Landau and H. O. Pollak. Prolate spheroidal wave functions, Fourier analysis and uncer-
tainty,III: The dimension of the space of essentially time- and band-limited signals. Bell System
Tech. J., 41:1295–1336, 1962.
[20] S. Li and H. Ogawa. Pseudoframes for subspaces with applications. J. Fourier Anal. Appl., 10(4):409–
431, 2004.
[21] M. Liuni, A. Robel, E. Matusiak, M. Romito, and X. Rodet. Automatic Adaptation of the Time-
Frequency Resolution for Sound Analysis and Re-Synthesis. Audio, Speech, and Language Processing,
IEEE Transactions on, 21(5):959–970, 2013.
[22] E. Matusiak and Y. C. Eldar. Sub-Nyquist sampling of short pulses. IEEE Trans. Signal Process.,
60(3):1134–1148, March 2012.
[23] J. Ramanathan and P. Topiwala. Time-frequency localization and the spectrogram. Appl. Comput.
Harmon. Anal., 1(2):209–215, 1994.
[24] J. L. Romero. Surgery of spline-type and molecular frames. J. Fourier Anal. Appl., 17:135–174, 2011.
[25] D. Slepian and H. O. Pollak. Prolate Spheroidal Wave Functions, Fourier Analysis and Uncertainty
–I. Bell System Tech. J., 40(1):43–63, 1961.
SAMP. TF-LOC. FUNCTIONS AND CONST. LOC. TF FRAMES
19
[26] T. Strohmer. Numerical algorithms for discrete Gabor expansions. In H. G. Feichtinger and
T. Strohmer, editors, Gabor Analysis and Algorithms: Theory and Applications, pages 267–294.
Birkhauser Boston, Boston, 1998.
[27] G. G. Walter and X. A. Shen. Sampling with prolate spheroidal wave functions. Sampl. Theory
Signal Image Process., 2(1):25–52, 2003.
[28] E. Wilczok. New uncertainty principles for the continuous Gabor transform and the continuous
wavelet transform. Doc. Math., J. DMV, 5:201–226, 2000.
[29] M.-W. Wong. Wavelet Transforms and Localization Operators. Operator Theory: Advances and
Applications. 136. Basel: Birkhauser, 2002.
Numerical Harmonic Analysis Group, Faculty of Mathematics, University of Vienna,
Oskar-Morgenstern-Platz 1, 1090 Vienna, Austria
E-mail address, Monika Dorfler: [email protected]
Institute of Mathematics, University of the Philippines, Diliman, Quezon City 1101,
Philippines
E-mail address, Gino Angelo Velasco: [email protected]
|
1707.04387 | 1 | 1707 | 2017-07-14T05:23:43 | The Ritt property of subordinated operators in the group case | [
"math.FA"
] | Let $G$ be a locally compact abelian group, let $\nu$ be a regular probability measure on $G$, let $X$ be a Banach space, let $\pi\colon G\to B(X)$ be a bounded strongly continuous representation. Consider the average (or subordinated) operator $S(\pi,\nu) = \int_{G} \pi(t)\,d\nu(t)\,\colon X\to X$. We show that if $X$ is a UMD Banach lattice and $\nu$ has bounded angular ratio, then $S(\pi,\nu)$ is a Ritt operator with a bounded $H^\infty$ functional calculus. Next we show that if $\nu$ is the square of a symmetric probability measure and $X$ is $K$-convex, then $S(\pi,\nu)$ is a Ritt operator. We further show that this assertion is false on any non $K$-convex space $X$. | math.FA | math |
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE
GROUP CASE
FLORENCE LANCIEN AND CHRISTIAN LE MERDY
Abstract. Let G be a locally compact abelian group, let ν be a regular probability measure
on G, let X be a Banach space, let π : G → B(X) be a bounded strongly continuous represen-
tation. Consider the average (or subordinated) operator S(π, ν) = RG π(t) dν(t) : X → X.
We show that if X is a UMD Banach lattice and ν has bounded angular ratio, then S(π, ν) is
a Ritt operator with a bounded H ∞ functional calculus. Next we show that if ν is the square
of a symmetric probability measure and X is K-convex, then S(π, ν) is a Ritt operator. We
further show that this assertion is false on any non K-convex space X.
2000 Mathematics Subject Classification : 47A60, 47A80
1. Introduction
Let G be a locally compact abelian group and let M(G) denote the Banach algebra
of all bounded regular Borel measures on G. Let X be a complex Banach space and let
B(X) denote the Banach algebra of all bounded operators on X. Let π : G → B(X) be a
representation, that is, π(t + s) = π(t)π(s) for any t, s in G, and π(e) = IX (where e and IX
denote the unit of G and the identity operator on X, respectively). Assume further that π
is bounded, that is supt∈G kπ(t)k < ∞ , and that π is strongly continuous, that is, for any
x ∈ X, the mapping t 7→ π(t)x is continuous from G into X. To any probability measure
ν ∈ M(G), one can associate the average operator
(1)
π(t) dν(t) ∈ B(X),
S(π, ν) =ZG
where the integral is defined in the strong sense.
In this paper we are interested in the following two questions.
Q.1 When is S(π, ν) a Ritt operator?
Q.2 When does S(π, ν) admit a bounded H ∞ functional calculus?
Background on Ritt operators and their H ∞ functional calculus (along with references) will
be given in Section 2 below.
Average operators appear in various contexts, notably in ergodic theory. When X is a
function space, the behaviour of the norm of the powers of S(π, ν), the almost everywhere
convergence of these powers and various maximal and oscillation inequalities were widely
studied, see [11, 20, 21, 29] and the references therein. Recent papers on these topics [8,
Date: November 14, 2018.
1
2
F. LANCIEN AND C. LE MERDY
10, 27, 28] show the importance of the Ritt property and H ∞ functional calculus in the
behaviour of the powers of operators S(π, ν). This is the source of motivation for this paper.
Let U ∈ B(X) be an invertible operator such that supk∈Z kU kk < ∞. Then the mapping
π : Z → B(X) defined by π(k) = U k for any k ∈ Z is a representation and any bounded
representation of Z on X has this form. A probability measure on Z is given by a sequence
ν = (ck)k∈Z of nonnegative real numbers such that Pk ck = 1. In this case, we have
S(ν, π) =
ckU k.
∞Xk=−∞
Thus S(ν, π) is subordinated to U in the sense of [13]. The special case G = Z therefore
indicates that average operators (1) may be regarded as subordinated operators in the context
of group representations. Subordination operators induced by probability measures on the
semigroup N were extensively studied recently [4, 13, 14, 15]. In this context a major question
n=0 ckT k is a Ritt operator for a nonnegative sequence (ck)k≥0 with
is to determine when P∞
Pk ck = 1 and a power bounded T : X → X. Question Q.1 in the present paper should be
considered as its analogue in the group case.
Our results in Sections 3-5 emphasize the role of Banach space geometry in these issues. In
Section 3 we recall the so-called bounded angular ratio (BAR) condition and extend some of
the results in [8]. We show that if ν has BAR and X is a UMD Banach lattice, then S(π, ν)
is a Ritt operator and it admits a bounded H ∞ functional calculus for any π as above.
Section 4 deals with the case when X is a K-convex Banach space and ν is the square
of a symmetric probability measure. In this case we show (see Theorem 11) that for any
bounded strongly continuous representation π : G → B(X), S(π, ν) is a Ritt operator. In the
case when ν = λX
p is the regular representation on Lp(G; X) (see (8) for the definition), this
result is a discrete analogue of Pisier's Theorem [35] showing the analyticity of the tensor
extension of any convolution semigroup associated with a family of symmetric probability
measures.
Section 5 provides examples of pairs (π, ν) for which Q.1 (and hence Q.2) has a negative
answer. These examples further show that the K-convexity assumption is unavoidable in
Theorem 11.
We conclude this introduction with a few notations and conventions. Unless otherwise
specified, G denotes an arbitrary locally compact abelian group, equipped with a fixed Haar
measure dt. For any 1 ≤ p ≤ ∞, we let Lp(G) denote the Lp-space associated to this measure.
We let bG denote the dual group of G and, for any ν ∈ M(G), we let bν : bG → C denote the
Fourier transform of ν. For any measurable subset V ⊂ G, we let V and χV : G → R
denote the Haar measure of V and the characteristic function of V , respectively.
For any t ∈ G, let λp(t) : Lp(G) → Lp(G) be the translation operator defined by [λp(t)f ](s) =
f (s − t) for any f ∈ Lp(G). We say that an operator T : Lp(G) → Lp(G) is a Fourier multi-
plier if T λp(t) = λp(t)T for any t ∈ G.
Let (Ω, µ) be a measure space. For any 1 ≤ p ≤ ∞ and for any Banach space X, we let
Lp(Ω; X) be the Bochner space of measurable functions f : Ω → X (defined up to almost
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
3
everywhere zero functions) such that the norm function kf (· )k belongs to Lp(Ω) (see e.g.
[12, Chapter II]). For p 6= ∞, the algebraic tensor product Lp(Ω) ⊗ X is dense in Lp(Ω; X).
Let 1 ≤ p < ∞, let T : Lp(Ω1) → Lp(Ω2) and let S : X → X be bounded operators.
If T ⊗ S : Lp(Ω1) ⊗ X → Lp(Ω2) ⊗ X extends to a bounded operator from Lp(Ω1; X) into
Lp(Ω2; X), then we let
T ⊗S : Lp(Ω1; X) −→ Lp(Ω2; X)
denote this extension. It is well-known that if T is positive (i.e. T (f ) ≥ 0 for any f ≥ 0),
then T ⊗ S has a bounded extension for any S : X → X.
2. Ritt operators and their H ∞ functional calculus
An operator T : X → X is called power bounded if there exists a constant C0 > 0 such
that
∀ n ≥ 0,
kT nk ≤ C0.
Then a power bounded T is called a Ritt operator if there exists a constant C1 > 0 such that
(2)
∀ n ≥ 1,
nkT n − T n−1k ≤ C1.
Ritt operators can be characterized by a spectral condition, as follows. Let
D = {z ∈ C : z < 1}
be the open unit disc. For any T ∈ B(X), let σ(T ) denote the spectrum of T . Then T is a
Ritt operator if and only if σ(T ) ⊂ D and there exists a constant K > 0 such that
(3)
∀ z ∈ C \ D,
k(z − T )−1k ≤
K
z − 1
.
This result goes back to [31, 33, 34], see also [3, 42] for complements.
Let us now turn to functional calculus. For any angle γ ∈ (0, π
2 ), consider the so-called
Stolz domain Bγ as sketched in Figure 1. In analytic terms, Bγ is defined as the interior of
the convex hull of 1 and the disc {z ∈ C : z < sin γ}.
γBγ
0
1
Figure 1. Stolz domain
4
F. LANCIEN AND C. LE MERDY
It turns out that if T is a Ritt operator, then σ(T ) ⊂ Bγ for some γ ∈ (0, π
2 ). In particular
we have σ(T ) ⊂ D ∪ {1}.
One important feature of the above Slolz domains is that for any γ ∈ (0, π
2 ) there exists a
constant Cγ > 0 such that
(4)
∀ z ∈ Bγ,
1 − z ≤ Cγ(1 − z).
Let P be the algebra of all complex polynomials. Let γ ∈ (0, π
2 ). Following [26] (to which
we refer for more information), we say that an operator T : X → X has a bounded H ∞(Bγ)
functional calculus if there exists a constant C ≥ 1 such that
(5)
∀ ϕ ∈ P,
A routine argument shows that this condition implies
kϕ(T )k ≤ C sup(cid:8)ϕ(z) : z ∈ Bγ(cid:9).
Further if T satisfies (5), then T is a Ritt operator. Indeed (5) applied to z 7→ zn immediately
implies that kT nk ≤ C for any n ≥ 0. Next define ϕn ∈ P by
σ(T ) ⊂ Bγ.
for any n ≥ 1. Then for any z ∈ Bγ, we have
ϕn(z) ≤ Cγ nzn−1(1 − z)
ϕn(z) = n(zn − zn−1)
by (4). An elementary computation shows that
We deduce that
∀ t ∈ (0, 1),
ntn−1(1 − t) ≤ (cid:16)1 −
n(cid:17)n−1
sup(cid:8)ϕn(z) : z ∈ Bγ, n ≥ 1(cid:9) < ∞.
1
.
Thus (5) implies the boundedness of the sequence (ϕn(T ))n≥1, hence an estimate (2). We
record this simple fact for further use.
Lemma 1. Let T : X → X be an operator satisfying (5) for some γ ∈ (0, π
Ritt operator.
2 ). Then T is
Condition (5) is close but different from the notion of polynomial boundedness. Recall that
an operator T : X → X is called polynomially bounded if there exists a constant C ≥ 1 such
that kϕ(T )k ≤ C sup{ϕ(z) : z ∈ D}. Clearly if T admits a bounded H ∞(Bγ) functional
calculus, then it is polynomially bounded. However there exist polynomially bounded Ritt
operators which do not admit any bounded H ∞(Bγ) functional calculus [25].
Definition 2. We say that a Ritt operator T ∈ B(X) has a bounded H ∞ functional calculus
if it admits a bounded H ∞(Bγ) functional calculus for some γ ∈ (0, π
2 ).
Recent papers show the relevance of this notion for the study of Ritt operators. It is proved
in [16, 26] that for a large class of Banach spaces X, a Ritt operator T : X → X has a bounded
H ∞ functional calculus in the above sense if and only if T and its adjoint T ∗ : X ∗ → X ∗
satisfy certain square functions estimates which naturally arise in the harmonic analysis of
the discrete semigroup (Tn)n≥0. We refer the reader to the papers [27, 28] for applications
of this characterization of bounded H ∞ functional calculus.
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
5
We now briefly discuss sectorial operators, which will be used in Section 3. For any angle
ω ∈ (0, π), set
Σω =(cid:8)λ ∈ C∗ : Arg(λ) < ω(cid:9).
Recall that a closed operator A : D(A) → X with dense domain D(A) ⊂ X is called sectorial
of type ω if its spectrum is included in Σω and for any α ∈ (ω, π) there exists a constant
Kα > 0 such that
(6)
∀ λ ∈ C \ Σα,
k(λ − A)−1k ≤
Kα
λ
.
A simple connection between the Ritt condition and sectoriality is that an operator T ∈ B(X)
is a Ritt operator if and only if A = IX − T is sectorial of type < π
2 and σ(T ) ⊂ D ∪ {1}.
This follows from comparing (3) and (6), see [33] for details.
Let R be the algebra of all rational functions with nonpositive degree and poles in the
closed half-line R−. For any φ ∈ R and any sectorial operator A, φ(A) is a well-defined
bounded operator on X. For any α ∈ (0, π), we say that a sectorial operator A has a
bounded H ∞(Σα) functional calculus if there exists a constant C ≥ 1 such that
∀ φ ∈ R,
kφ(A)k ≤ C sup(cid:8)φ(λ) : λ ∈ Σα(cid:9).
This definition is equivalent to other classical ones that the interested reader will find e.g.
in [17, 24].
A strong connection between H ∞ calculus for sectorial and Ritt operators is given by the
following statement.
Theorem 3. ([26, Proposition 4.1]) Let T ∈ B(X) be a Ritt operator and let A = IX − T .
Then T has a bounded H ∞ functional calculus (in the sense of Definition 2) if and only if
there exists α ∈ (0, π
2 ) such that A has a bounded H ∞(Σα) functional calculus.
3. The BAR condition and UMD Banach lattices
Let X be an arbitrary Banach space. We adopt the notation (1) for any ν ∈ M(G) and
any bounded strongly continuous representation π : G → B(X).
A straightforward application of Fubini's Theorem shows that for any ν1, ν2 ∈ M(G), we
have
S(π, ν2)S(π, ν1) = S(π, ν2 ∗ ν1).
This implies that for any ϕ ∈ P and for any ν ∈ M(G),
(7)
ϕ(S(π, ν)) = S(π, νϕ),
where νϕ := ϕ(ν) ∈ M(G) is obtained by applying polynomial functional calculus in the
Banach algebra M(G).
For any 1 ≤ p ≤ ∞, let
be the convolution operator defined by setting C X
Lp(G; X). Note that if p 6= ∞ and λX
defined by
C X
ν,p : Lp(G; X) −→ Lp(G; X)
ν,p(f ) = ν ∗ f =RG f (· −t) dν(t) for any f in
p : G → B(Lp(G; X)) denotes the regular representation
(8)
f ∈ Lp(G; X), s, t ∈ G,
p (t)f(cid:3)(s) = f (s − t),
(cid:2)λX
6
F. LANCIEN AND C. LE MERDY
then we have
(9)
C X
ν,p = S(λX
p , ν).
For convenience we will write Cν,p instead of C X
notation introduced at the end of Section 1, we have
ν,p when X = C. Observe that with the
(10)
for any 1 ≤ p < ∞.
C X
ν,p = Cν,p⊗IX
We will use the following well-known transference principle, which is a variant of [9, The-
orem 2.4]. We include a proof for the sake of completeness.
Proposition 4. Let π : G → B(X) be a bounded strongly continuous representation and set
kπk = sup
t∈G
kπ(t)k.
Then for any 1 < p < ∞, we have
(11)
(cid:13)(cid:13)S(π, ν) : X → X(cid:13)(cid:13) ≤ kπk2(cid:13)(cid:13)C X
ν,p : Lp(G; X) → Lp(G; X)(cid:13)(cid:13).
Proof. We first assume that ν has support in a compact subset K ⊂ G. Let V be an arbitrary
open set, with 0 < V < ∞. Let x ∈ X and define f : G → X by setting
f (t) = χV −K(t)π(−t)x,
t ∈ G.
The assumptions imply that V − K has a finite Haar measure, hence f belongs to Lp(G; X)
with
kf kp
p ≤ V − Kkπkpkxkp.
For any s ∈ V , we may write
S(π, ν)x = π(s)π(−s)S(π, ν)x = π(s)ZG
π(t − s)x dν(t),
which yields
Integrating over V , we deduce
By definition,
Since ν has support in K, RG π(t − s)x dν(t) is equal to RG χV −K(s − t)π(t − s)x dν(t) for
any s ∈ V . Consequently,
kS(π, ν)xk ≤ kπk(cid:13)(cid:13)(cid:13)ZG
π(t − s)x dν(t)(cid:13)(cid:13)(cid:13).
χV (s)(cid:13)(cid:13)(cid:13)ZG
V kS(π, ν)xkp ≤ kπkpZG
π(t − s)x dν(t)(cid:13)(cid:13)(cid:13)
χV −K(s − t)π(t − s)x dν(t)(cid:13)(cid:13)(cid:13)
p
V kS(π, ν)xkp ≤ kπkpZG(cid:13)(cid:13)(cid:13)ZG
ν,p(f )(cid:3)(s) =ZG
(cid:2)C X
χV −K(s − t)π(t − s)x dν(t)
ds .
p
ds .
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
7
for a.e. s ∈ G. Hence we obtain
V kS(π, ν)xkp ≤ kπkpkC X
ν,pkpkf kp
p
We deduce that
≤ kC X
ν,pkpV − Kkπk2pkxkp.
V (cid:19) 1
kS(π, ν)xk ≤(cid:18)V − K
p
kπk2kCν,pkkxk.
The group G is abelian, hence amenable. Thus according to Folner's condition (see e.g.
is arbitrary close to 1. Hence the above
[9, Chapter 2]), we can choose V such that V −K
V
inequality shows that ν satisfies (11), in the case when ν has compact support.
Now consider an arbitrary ν ∈ M(G) (without any assumption on its support). Since this
measure is regular, there is a sequence (Kn)n≥1 of compact subsets of G such that
ν(G) = lim
n→∞
ν(Kn).
Define νn = νKn for any n ≥ 1. Then kνn − νkM (G) → 0 hence kC X
ν,pk → 0 when
n → ∞. Further for any x ∈ X, S(π, νn)x → S(π, ν)x when n → ∞, by Lebesgue's
Theorem. By the first part of the proof, each νn satisfies (11), hence ν satisfies (11) as
well.
(cid:3)
νn,p − C X
Following [8] we say that a probability measure ν ∈ M(G) has bounded angular ratio
(BAR in short) if there exists a constant K ≥ 1 such that
(cid:12)(cid:12)1 −bν(ξ)(cid:12)(cid:12) ≤ K(cid:0)1 − bν(ξ)(cid:1),
ξ ∈ bG.
It is well-known that this holds true if and only if there exists an angle γ ∈ (0, π
2 ) such that
bν(ξ) ∈ Bγ for any ξ ∈ bG. The relevance of BAR for the study of Q.1 and Q.2 is shown by
the following elementary result.
Lemma 5. Let ν ∈ M(G) be a probability measure and let H be a Hilbert space. Then C H
ν,2
is a Ritt operator if and only if ν has BAR. In this case, the operator C H
ν,2 admits a bounded
H ∞ functional calculus.
Proof. For any T ∈ B(L2(G)), T ⊗ IH extends to a bounded operator on L2(G; H) with
kT ⊗IHk = kT k. This immediately implies that σ(C H
ν,2 is a normal
operator on L2(G; H), hence
ν,2) = σ(Cν,2). Further C H
(12)
∀ ϕ ∈ P,
Applying Fourier transform, we have
kϕ(C H
ν,2)k = sup(cid:8)ϕ(z) : z ∈ σ(Cν,2)(cid:9).
hence ν has BAR if and only if there exists γ ∈ (0, π
this equivalence with (12) yields the result.
2 ) such that σ(Cν,2) ⊂ Bγ. Combining
(cid:3)
σ(Cν,2) =(cid:8)bν(ξ) : ξ ∈ bG(cid:9),
8
F. LANCIEN AND C. LE MERDY
It follows from (9) and Lemma 5 that if a probability measure ν ∈ M(G) is such that
S(π, ν) is a Ritt operator for all bounded strongly continuous representations π acting on
Hilbert space, then ν necessarily has BAR. The next theorem shows that if we consider
representations acting on UMD Banach lattices, this necessary condition is also sufficient.
We refer the reader to [7, 19] for background and information on the UMD property and
to [38] for a more specific study for Banach lattices. We merely recall that any UMD Banach
space is reflexive, that Hilbert spaces and Lp-spaces for 1 < p < ∞ are UMD and that the
UMD property is stable under taking subspaces and quotients.
Theorem 6. Let ν ∈ M(G) be a probability measure with BAR. Let X be a UMD Banach
lattice. For any bounded strongly continuous representation π : G → B(X), S(π, ν) is a Ritt
operator with a bounded H ∞ functional calculus.
This theorem is close in spirit to [43]. Furthermore it extends some of the results of [8].
Indeed it is shown in [8, Proposition 5.2] that a probability measure ν has BAR if and only
if Cν,p is a Ritt operator for any 1 < p < ∞. Further [8, Theorem 5.6] says that if X is
an Lp-space for some 1 < p < ∞ and ν has BAR, then S(π, ν) is a Ritt operator for any
bounded strongly continuous representation π : G → B(X). H ∞ functional calculus is not
discussed in [8].
For the proof of Theorem 6, will use complex interpolation, for which we refer to [5, 23].
Given any compatible couple (X0, X1) of Banach spaces and any θ ∈ [0, 1], we let [X0, X1]θ
denote the interpolation space defined by [23, Section 4]. The interpolation theorem ensures
that if T : X0 + X1 → X0 + X1 is a linear operator such that T : X0 → X0 and T : X1 → X1
boundedly, then T : [X0, X1]θ → [X0, X1]θ boundedly for any θ ∈ [0, 1]. In the context of
Ritt operators, we have the following.
Proposition 7. ([6]) Let (X0, X1) be a compatible couple of Banach spaces and let T : X0 +
X1 → X0 +X1 be a linear operator such that T : X0 → X0 is power bounded and T : X1 → X1
is a Ritt operator. Then for any θ ∈ (0, 1], T : [X0, X1]θ → [X0, X1]θ is a Ritt operator.
This result was established by Blunck [6] in the case when X0, X1 are Lp-spaces. However
the proof works as well in the more general setting of interpolation couples so we omit it.
Proof of Theorem 6. Let ν ∈ M(G) be a probability measure with BAR. According to Def-
inition 2 and Lemma 1, it suffices to show the existence of γ ∈ (0, π
2 ) and C ≥ 1 such
that
By (7) and Proposition 4, we have
∀ ϕ ∈ P,
kϕ(S(π, ν))k ≤ C sup(cid:8)ϕ(z) : z ∈ Bγ(cid:9).
kϕ(S(π, ν))k ≤ kπk2kϕ(C X
ν,2)k
for any ϕ ∈ P. Hence it suffices to show that T = C X
ν,2 satisfies (5) for some γ ∈ (0, π
2 ).
Since X is a UMD Banach lattice, it follows from [38] that there exist a compatible couple
(Y, H) and some θ ∈ (0, 1] such that H is a Hilbert space, Y is a UMD Banach space and
X = [Y, H]θ isometrically. By [5, Theorem 5.1.2], this implies that
(13)
L2(G; X) = [L2(G; Y ), L2(G; H)]θ
isometrically.
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
9
By Lemma 5, C H
ν,2 is a Ritt operator. Further C Y
ν,2 is a contraction hence it follows from
(13) and Lemma 7 that C X
ν,2 is a Ritt operator.
Now we set
and we similarly define AY on L2(G; Y ) and AH on L2(G; H) . Then we consider
AX = IL2(G;X) − C X
ν,2
Tt = e−tetCν,2,
t ≥ 0.
It is plain that (Tt)t≥0 is c0-semigroup of contractions on L2(G). Its negative generator is
equal to IL2 − Cν,2. Moreover Tt is a positive operator (in the lattice sense) for any t ≥ 0.
Since Y is UMD, [18, Theorem 6] asserts that the negative generator of (cid:0)Tt⊗IY(cid:1)t≥0 has a
2 , π). By construction, this negative
bounded H ∞(Σα) functional calculus for any α ∈ ( π
generator is (IL2 − Cν,2)⊗IY , which is equal to AY .
On the other hand, C H
ν,2 has a bounded H ∞ functional calculus by Lemma 5 hence by
Theorem 3, there exists β < π
2 such that AH admits a bounded H ∞(Σβ) functional calculus.
Applying the interpolation theorem for H ∞ functional calculus [22, proposition 4.9], we
deduce that AX admits a bounded H ∞(Σθβ+(1−θ)α) functional calculus for any α ∈ ( π
2 , π).
Choosing α sufficiently close to π
2 we obtain that AX admits a bounded H ∞(Σρ) functional
calculus for some ρ < π
ν,2 has a bounded H ∞
functional calculus.
(cid:3)
2 . By Theorem 3 we finally obtain that C X
In the next corollary we focus on the simple case G = Z.
Corollary 8. Let (ck)k∈Z be a sequence on nonnegative real numbers such thatP∞
and there exists a constant K ≥ 1 such that
k=−∞ ck = 1
θ ∈ R.
∞Xk=−∞
(cid:12)(cid:12)(cid:12)1 −
ckeikθ(cid:12)(cid:12)(cid:12) ≤ K(cid:16)1 −(cid:12)(cid:12)(cid:12)
∞Xk=−∞
ckeikθ(cid:12)(cid:12)(cid:12)(cid:17),
(1) Let X be a UMD Banach lattice and let U : X → X be an invertible operator such
that
Then
kU kk < ∞ .
sup
k∈Z
V =
∞Xk=−∞
ckU k
is a Ritt operator with a bounded H ∞ functional calculus.
(2) Assume that ck = 0 for any k ≤ −1. Let (Ω, µ) be a measure space, let 1 < p < ∞
and let T : Lp(Ω) → Lp(Ω) be a positive operator with kT k ≤ 1. Then
S =
∞Xk=0
ckT k
is a Ritt operator with a bounded H ∞ functional calculus.
10
F. LANCIEN AND C. LE MERDY
Proof. Part (1) is the translation of Theorem 6 in the case G = Z.
To prove part (2), we apply the Akcoglu-Sucheston dilation theorem for positive contrac-
tions [1]: there exist a measure space (Ω′, µ′), a surjective isometry U : Lp(Ω′) → Lp(Ω′) and
two contractions J : Lp(Ω) → Lp(Ω′) and Q : Lp(Ω′) → Lp(Ω) such that T k = QU kJ for any
k ≥ 0.
Let S =Pk≥0 ckT k and V =Pk≥0 ckU k. For any integer n ≥ 1, we have
c(n)kT k
and
c(n)kU k,
Sn =Xk≥0
V n =Xk≥0
where c(n) ∈ ℓ1 is defined by c(1) = (ck)k≥0 and c(n) = c(1) ∗ · · · ∗ c(1) (n times). Then the
above dilation property implies that Sn = QV nJ for any n ≥ 0. Thus ϕ(S) = Qϕ(V )J for
any ϕ ∈ P and hence
∀ ϕ ∈ P,
kϕ(S)k ≤ kϕ(V )k.
The operator V has a bounded H ∞ functional calculus by part (1). By the above inequality,
S also has a bounded H ∞ functional calculus.
(cid:3)
Remark 9. Let ν ∈ M(G) be a probability measure with BAR and let 1 < p < ∞. According
to Theorem 6, C ν,p is a Ritt operator with a bounded H ∞ functional calculus.
Let X be an SQp space, that is, a quotient of a subspace of an Lp-space. For any T in
B(Lp(G)), T ⊗ IX extends to a bounded operator on Lp(G; X) with kT ⊗IXk = kT k. This
implies that
∀ ϕ ∈ P,
kϕ(C X
ν,p)k = kϕ(Cν,p)k.
The proof of Theorem 6 therefore shows that for any bounded strongly continuous represen-
tation π : G → B(X), S(π, ν) is a Ritt operator with a bounded H ∞ functional calculus.
It would be interesting to characterize the class of all Banach spaces X with this property.
4. Subordination on K-convex spaces
Let Ω0 denote the compact group {−1, 1}N equipped with its normalized Haar measure
and for any n ≥ 1, let εn : Ω0 → R denote the Rademacher function defined by setting
εn(Θ) = θn for any Θ = (θi)i≥1 ∈ Ω0. Let Rad2 ⊂ L2(Ω0) denote the closed linear span of
the εn and let Q : L2(Ω0) → L2(Ω0) be the orthogonal projection with range equal to Rad2.
A Banach space X is called K-convex if Q ⊗ IX extends to a bounded operator on
L2(Ω0; X). In this case, we let
KX =(cid:13)(cid:13)Q⊗IX : L2(Ω0; X) −→ L2(Ω0; X)(cid:13)(cid:13).
This number is called the K-convexity constant of X. This property plays a fundamental role
in Banach space theory, in relation with the notions of type and cotype. In his fundamental
paper [35], Pisier showed that X is K-convex if and only if it admits a non trivial Rademacher
type. We refer to [32, 35, 36] for more information on these topics.
Lp-spaces are K-convex whenever 1 < p < ∞. More generally we have the following
well-known fact.
Lemma 10. Let 1 < p < ∞ and let X be a K-convex Banach space. There exists a constant
C > 0 such that whenever (Ω, µ) is a measure space, the space Lp(Ω; X) is K-convex and
KLp(Ω;X) ≤ C.
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
11
Proof. It follows from the Khintchine-Kahane inequalities (see e.g.
[30, Theorem 1.e.13])
that Q extends to a bounded projection Qp : Lp(Ω0) → Lp(Ω0), that X is K-convex if and
only if Qp ⊗ IX extends to a bounded operator on Lp(Ω0; X), and that in this case,
ApKX ≤(cid:13)(cid:13)Qp⊗IX : Lp(Ω0; X) −→ Lp(Ω0; X)(cid:13)(cid:13) ≤ BpKX
for some universal constants 0 < Ap < Bp. By Fubini's Theorem, the boundedness of Qp ⊗IX
on Lp(Ω0; X) implies that Qp ⊗ ILp(Ω;X) is bounded on Lp(Ω0; Lp(Ω; X)), with
(cid:13)(cid:13)Qp⊗ILp(Ω;X) : Lp(Ω0; Lp(Ω; X)) → Lp(Ω0; Lp(Ω; X))(cid:13)(cid:13) =(cid:13)(cid:13)Qp⊗IX : Lp(Ω0; X) → Lp(Ω0; X)(cid:13)(cid:13).
The result follows at once.
(cid:3)
Let ν ∈ M(G) be a probability measure. We say that ν is symmetric provided that
ν(−V ) = ν(V ) for any measurable V ⊂ G. It is clear that ν is symmetric if and only if bν is
real valued. In this case, bν is actually valued in [−1, 1].
We will say that such a measure ν is a square if there exists another symmetric probability
measure η such that
ν = η ∗ η.
In this case, bν = bη2 is valued in [0, 1], hence ν has BAR.
Our main result is the following.
Theorem 11. Let ν ∈ M(G) be a symmetric probability measure. Assume that ν is a square
and let X be a K-convex Banach space.
(1) For any 1 < p < ∞, the convolution operator C X
ν,p = ν∗ · : Lp(G; X) → Lp(G; X) is
a Ritt operator.
(2) For any bounded strongly continuous representation π : G → B(X), S(π, ν) is a Ritt
operator.
If we think of Ritt operators as the discrete analogues of bounded analytic semigroups,
part (1) of the above theorem should be regarded as a discrete version of [35, Theorem 1.2].
Its proof will make crucial use of the following result of Pisier.
Theorem 12. ([35, Theorem 3.1]) Let Z be a K-convex Banach space. There exists a
constant C > 0 only depending on the K-convexity constant KZ such that for any integer
n ≥ 1 and for any n-tuple (P1, . . . , Pn) of mutually commuting contractive projections on Z,
(IZ − Pk) Y1≤j6=k≤n
(cid:13)(cid:13)(cid:13)(cid:13)
nXk=1
Pj(cid:13)(cid:13)(cid:13)(cid:13) ≤ C.
For any integer n ≥ 2, the Haar measure on Gn can be defined as the n-fold product of
the Haar measure on G. Then for any 1 ≤ p < ∞, we regard the n-fold tensor product
Lp(G) ⊗ · · · ⊗ Lp(G) as a subspace of Lp(Gn) is the usual way. We recall that
is a dense subspace.
Lp(G) ⊗ · · · ⊗ Lp(G) ⊂ Lp(Gn)
12
F. LANCIEN AND C. LE MERDY
Lemma 13. Let 1 < p < ∞. Let n ≥ 2 and m ≥ 1 be two integers and let (cid:8)νij(cid:9)1≤i≤n, 1≤j≤m
be a family in M(G). For any i, j as above, let
be the corresponding convolution operator. Then
Sij = Cνij ,p = νij∗ · : Lp(G) −→ Lp(G)
(14)
(cid:13)(cid:13)(cid:13)
mXj=1
Snj · · · S1j⊗IX(cid:13)(cid:13)(cid:13)B(Lp(G;X))
≤ (cid:13)(cid:13)(cid:13)
mXj=1
Snj⊗ · · · ⊗S1j⊗IX(cid:13)(cid:13)(cid:13)B(Lp(Gn;X))
.
Proof. In the first part of the proof, we assume that all the measures νij have a compact
support. Taking their union, we obtain a compact set K ⊂ G such that νij has support
included in K for any 1 ≤ i ≤ n, 1 ≤ j ≤ m. Let V ⊂ G be an arbitrary open subset with
0 < V < ∞. We set W = V − K for convenience.
We fix some f ∈ Lp(G; X) and introduce F : Gn → X defined by
F (s1, s2, . . . , sn) = χW (s2) · · · χW (sn)f (s1 + · · · + sn).
It is plain that F belongs to Lp(Gn; X), with
kF kp
p = W n−1kf kp
p.
We claim that for any j, for a.e. s1 ∈ G and for a.e. s2, . . . , sn in V , we have
(15)
Indeed applying S2j⊗IX, . . . , Snj⊗IX successively to S1j⊗IX(f ), we have
(cid:2)Snj · · · S1j⊗IX (f )(cid:3)(s1 + s2 + · · · + sn) =(cid:2)(cid:0)Snj⊗ · · · ⊗S1j⊗IX(cid:1)(F )(cid:3)(s1, . . . , sn).
(cid:2)Snj · · · S2jS1j⊗IX(f )(cid:3)(s1 + s2 + · · · + sn)
=ZGn−1(cid:2)S1j(f )(cid:3)(s1 + s2 + · · · + sn − t2 − · · · − tn)dν2j(t2) · · · dνnj(tn)
for a.e. (s1, s2, . . . , sn) in Gn. It follows from the fact that ν2j, . . . , νnj have support in K
that whenever s2, . . . , sn belong to V , this is equal to
ZGn−1
and hence to
χW (s2 − t2) · · · χW (sn − tn)(cid:18)ZG
χW (s2 − t2) · · · χW (sn − tn)f(cid:16) nXi=1
χW (s2 − t2) · · · χW (sn − tn)(cid:2)S1j(f )(cid:3)(s1 + s2 + · · ·+ sn − t2 − · · · − tn) dν2j(t2) · · · dνnj(tn),
(si − ti)(cid:17) dν1j(t1)(cid:19) dν2j(t2) · · · dνnj(tn)
ZGn−1
=ZGn
=ZGn
=(cid:2)(cid:0)Snj⊗ · · · ⊗S1j⊗IX(cid:1)(F )(cid:3)(s1, . . . , sn)
f(cid:16) nXi=1
(si − ti)(cid:17) dν1j(t1)dν2j(t2) · · · dνnj(tn)
F (s1 − t1, s2 − t2, . . . , sn − tn) dν1j(t1)dν2j(t2) · · · dνnj(tn)
as claimed.
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
13
To derive the estimate (14) from this identity, we note that by translation invariance,
p
p
=ZG(cid:13)(cid:13)(cid:13)Xj (cid:2)Snj · · · S1j⊗IX(f )(cid:3)(s1 + z)(cid:13)(cid:13)(cid:13)
p
ds1 .
p
p
ds1 · ds2 · · · dsn .
ds1 · ds2 · · · dsn
According to (15), this implies that
for any z ∈ G. Hence integrating over V n−1,
p
p
(cid:13)(cid:13)(cid:13)Xj
V n−1(cid:13)(cid:13)(cid:13)Xj
Snj · · · S1j⊗IX (f )(cid:13)(cid:13)(cid:13)
Snj · · · S1j⊗IX(f )(cid:13)(cid:13)(cid:13)
=ZV n−1ZG(cid:13)(cid:13)(cid:13)Xj (cid:2)Snj · · · S1j⊗IX (f )(cid:3)(s1 + s2 + · · · + sn)(cid:13)(cid:13)(cid:13)
V n−1(cid:13)(cid:13)(cid:13)Xj
Snj · · · S1j⊗IX(f )(cid:13)(cid:13)(cid:13)
=ZV n−1ZG(cid:13)(cid:13)(cid:13)h(cid:16)Xj
Snj⊗ · · · ⊗S1j⊗IX(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:13)Xj
≤(cid:13)(cid:13)(cid:13)Xj
Snj⊗ · · · ⊗S1j⊗IX(cid:13)(cid:13)(cid:13)
V (cid:19) n−1
Snj · · · S1j⊗IX(cid:13)(cid:13)(cid:13) ≤(cid:18) V − K
(cid:13)(cid:13)(cid:13)Xj
Snj⊗ · · · ⊗S1j⊗IX(cid:17)(F )i(s1, . . . , sn)(cid:13)(cid:13)(cid:13)
p (cid:13)(cid:13)(cid:13)Xj
V − Kn−1kf kp.
kF kp
p
p
p
p
This shows that
Snj⊗ · · · ⊗S1j⊗IX(cid:13)(cid:13)(cid:13).
As indicated in the proof of Proposition 4, we can choose V such that V −K
is arbitrary
close to 1. Consequently the above estimate implies (14) under the assumption that all νij
have compact support.
V
The argument at the end of Proposition 4 can be easily adapted to deduce the general
(cid:3)
result from this special case. Details are left to the reader.
Proof of Theorem 11. Let π be as in part (2) and let S = S(π, ν) ∈ B(X). It follows from
(7) and Proposition 4 that for any integer n ≥ 1,
(cid:13)(cid:13)Sn − Sn−1(cid:13)(cid:13) ≤ kπk2(cid:13)(cid:13)(C X
ν,2)n − (C X
ν,2)n−1(cid:13)(cid:13).
Hence part (2) is a consequence of part (1). We now aim at proving part (1).
It is plain that C X
ν,p is a contraction, hence a power bounded operator (this does not require
the K-convexity assumption).
We fix 1 < p < ∞ and set T = Cν,p : Lp(G) → Lp(G). By assumption, there exists a
symmetric probability measure η ∈ M(G) such that T is the square of Cη,p.
For any 1 ≤ q ≤ ∞, Cη,q = ν∗ · is a positive contraction on Lq(G). Since η is symmetric,
the operator Cη,2 is selfadjoint on the Hilbert space L2(G). Further Cη,∞(1) = 1, because η
is a probability measure. It therefore follows from Rota's dilation Theorem (see e.g. [41, p.
14
F. LANCIEN AND C. LE MERDY
106]) that there exist a measure space (Ω, µ), two positive contractions J : Lp(G) → Lp(Ω)
and Q : Lp(Ω) → Lp(G), as well as a conditional expectation E : Lp(Ω) → Lp(Ω) such that
(16)
T = QEJ
and
ILp(G) = QJ.
In the sequel we will make essential use of the fact that E is a positive contraction and
E2 = E.
Let n ≥ 1 be an integer. We may write
−n(T n − T n−1) = nT n−1(ILp − T ) =
nXj=1
T n−j(ILp − T )T j−1.
We apply Lemma 13 with m = n and the family {Sij}1≤i,j≤n of convolution operators on
Lp(G) defined by
Sjj = ILp − T
and
Sij = T if i 6= j.
This yields
(17)
n(cid:13)(cid:13)(T n − T n−1)⊗IX(cid:13)(cid:13)B(Lp(G;X)) ≤
T ⊗ · · · ⊗T
⊗(ILp − T )⊗ T ⊗ · · · ⊗T
(cid:13)(cid:13)(cid:13)
nXj=1
n−j
{z
}
⊗IX(cid:13)(cid:13)(cid:13)B(Lp(Gn;X))
.
j−1
{z
}
For convenience we let
Y = Lp(Ω)
and
Zn = Lp(Ωn; X).
According to (16), we have
T ⊗ · · · ⊗ T
n−j
{z
=
j−1
⊗ (ILp − T ) ⊗ T ⊗ · · · ⊗ T
}
}
(cid:16)Q ⊗ · · · ⊗ Q
(cid:17)(cid:16)E ⊗ · · · ⊗ E
}
}
{z
{z
{z
n−j
n
⊗(IY − E) ⊗ E ⊗ · · · ⊗ E
(cid:17)(cid:16)J ⊗ · · · ⊗ J
(cid:17)
}
}
{z
n
j−1
{z
for any j = 1, . . . , n. Thus the operator in (17) is the composition of the following three
operators:
(i) the operator J⊗ · · · ⊗J⊗IX, which is a well-defined contraction from Lp(Gn; X) into
Lp(Ωn; X), because J is a positive contraction;
(ii) the operator
E⊗ · · · ⊗E
⊗(IY − E)⊗ E⊗ · · · ⊗E
⊗IX : Zn −→ Zn,
nXj=1
n−j
{z
}
j−1
{z
}
which is well-defined because E is positive.
(iii) the operator Q⊗ · · · ⊗Q⊗IX , which is a well-defined contraction from Lp(Ωn; X) into
Lp(Gn; X), because Q is a positive contraction.
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
15
We deduce that
⊗(IY − E)⊗ E⊗ · · · ⊗E
⊗IX(cid:13)(cid:13)(cid:13)B(Zn)
.
}
j−1
{z
Now for any j = 1, . . . , n, define Pj : Zn → Zn by
E⊗ · · · ⊗E
n(cid:13)(cid:13)(T n − T n−1)⊗IX(cid:13)(cid:13)B(Lp(G;X)) ≤(cid:13)(cid:13)(cid:13)
nXj=1
{z
{z
nXk=1
(IZn − Pk) Y1≤j6=k≤n
Pj = IY ⊗ · · · ⊗IY
}
n−j
{z
(18)
}
n−j
j−1
}
Pj.
⊗E⊗ IY ⊗ · · · ⊗IY
⊗IX.
Then the operator in the right-hand side of the above inequality is equal to
Since E is a positive contraction, each Pj is a contraction. Moreover the Pj mutually com-
mute. Further Zn is K-convex and supn≥1 KZn < ∞ by Lemma 10. It therefore follows from
Theorem 12 that the operators in (18) are uniformly bounded. Consequently,
sup
n≥1
n(cid:13)(cid:13)(T n − T n−1)⊗IX(cid:13)(cid:13)B(Lp(G;X)) < ∞.
ν,p = T ⊗IX is a Ritt operator.
According to (10), this proves that C X
(cid:3)
We will show in Proposition 14 below that if a symmetric probability ν has a density we
do not need to assume that ν is a square in the statement of Theorem 11.
h,p, S(π, h) instead of C ν,p, C X
We identify any h ∈ L1(G) with the measure ν with density h and use the notations
Ch,p, C X
ν,p, S(π, ν). With this convention, h is a probability when
h ∈ L1(h)+ and khk1 = 1 and h is symmetric when h(t) = h(−t) for a.e. t ∈ G. Note that a
symmetric probability h ∈ L1(G) has BAR if and only if there exists a ∈ (−1, 1] such that
A classical result asserts that for any h ∈ L1(G), we have σ(Ch,p) = σ(Ch,2) for any
1 ≤ p < ∞ (see e.g. [44]). This implies that
∀ ξ ∈ bG,
a ≤bh(ξ) ≤ 1.
(19)
σ(C X
h,p) = σ(Ch,2)
for any 1 ≤ p < ∞ and any Banach space X. Indeed let λ ∈ C\σ(Ch,1). The operator λIL1 −
Ch,1 is a Fourier multiplier hence its inverse is a Fourier multiplier on L1(G). Consequently
there exists ρ ∈ M(G) such that (λIL1 − Ch,1)−1 = Cρ,1 = ρ∗ · : L1(G) → L1(G). This
implies that λILp(X) − C X
h,p) is included
in σ(Ch,1) = σ(Ch,2). The reverse inclusion is clear.
Proposition 14. Let h ∈ L1(G)+ with khk1 = 1. Assume that h is symmetric. If h has
BAR and X is a K-convex Banach space, then C X
h,p is a Ritt operator for any 1 < p < ∞.
Furthermore for any bounded strongly continuous representation π : G → B(X), S(π, h) is a
Ritt operator.
h,p is invertible, with inverse equal to C X
ρ,p. Hence σ(C X
Proof. As in the proof of Theorem 11 it suffices to prove the first assertion.
We let T = C X
h,p. By the BAR assumption and Lemma 5, Ch,2 is a Ritt operator hence
−1 /∈ σ(Ch,2). Applying (19) we deduce that I + T is invertible.
16
F. LANCIEN AND C. LE MERDY
By Theorem 11, T 2 is a Ritt operator. Hence the sequence (cid:0)n(T 2n − T 2(n−1))(cid:1)n≥1 is
bounded. Writing
T 2n − T 2(n−1) = T 2(n−1)(T 2 − I) = T 2(n−1)(T − I)(T + I),
we deduce that the sequence (cid:0)nT 2(n−1)(T − I)(cid:1)n≥1 is bounded. This immediately implies
that the sequence (cid:0)nT n(T − I)(cid:1)n≥1 itself is bounded, that is, T is a Ritt operator.
Applying the above result in the case G = Z, we derive the following.
(cid:3)
Corollary 15. Let (ck)k∈Z be a sequence on nonnegative real numbers such that
∀ θ ∈ R,
0 ≤
∞Xk=−∞
ckeikθ ≤ 1.
Let X be a K-convex Banach space and let U : X → X be an invertible operator such that
supk∈Z kU kk < ∞. Then
S =
∞Xk=−∞
ckU k
is a Ritt operator.
5. A family of counterexamples
In Theorem 17 below we provide examples which show two facts. First BAR does not
imply that Q.1. has an affirmative answer on general Banach spaces, although Theorem 6
says that this is the case on UMD Banach lattices. Second, the K-convexity assumption in
Theorem 11 is optimal.
In the sequel we let ℓ1
We give some background on the so-called regular operators, which will be used in the proof
of Theorem 17. Let (Ω, µ) be a measure space, let 1 < p < ∞ and let T : Lp(Ω) → Lp(Ω) be
a bounded operator. We say that T is regular if there exists a constant C ≥ 0 such that
n ) denote the n-dimensional L1-space (resp. L∞-space).
n (resp. ℓ∞
(cid:13)(cid:13)(cid:13) sup
1≤k≤n(cid:12)(cid:12)T (fk)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)Lp
≤ C(cid:13)(cid:13)(cid:13) sup
1≤k≤n(cid:12)(cid:12)fk(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)Lp
for any integer n ≥ 1 and any f1, . . . , fn in Lp(Ω).
In this case, we let kT kr denote the
smallest C for which this property holds; this is called the regular norm of T . We mention
(although we will not use it) that an operator T is regular if and only if it is a linear
combination of positive bounded operators Lp(Ω) → Lp(Ω).
This definition can be reformulated as follows: a bounded operator T on Lp(Ω) is regular
if and only if the tensor extensions T ⊗ Iℓ∞
n ) are uniformly bounded.
It is not hard to deduce that T is regular if and only if T ⊗ IX extends to a bounded operator
on Lp(Ω; X) for any Banach space X and that kT ⊗IX : Lp(Ω; X) → Lp(Ω; X)k ≤ kT kr in
this case. This immediately implies that a bounded operator T : Lp(Ω) → Lp(Ω) is regular
if and only if its adjoint T ∗ : Lp′(Ω) → Lp′(Ω) is regular, and that
n ) → Lp(Ω; ℓ∞
n : Lp(Ω; ℓ∞
(20)
kT ∗kr = kT kr
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
17
in this case. (Here p′ = p/(p − 1) denotes the conjugate of p.) We refer to [37, 40] for details
and complements.
We say that a Banach space X contains the ℓ1
n uniformly if there exists a constant C ≥ 1
such that for any n ≥ 1, there exist x1, x2, . . . , xn in X such that
nXi=1
αi ≤(cid:13)(cid:13)(cid:13)
nXi=1
αixi(cid:13)(cid:13)(cid:13) ≤ C
nXi=1
αi
for any complex numbers α1, α2, . . . , αn. This is equivalent to the existence, for any n ≥ 1,
of an n-dimensional subspace of X which is C-isomorphic to ℓ1
n. It is shown in [35] that a
Banach space contains the ℓ1
n uniformly if and only if it is not K-convex. This leads to the
following.
Lemma 16. Let X be a Banach space and assume that X is not K-convex. Let T : Lp(Ω) →
Lp(Ω) be a bounded operator such that T ⊗ IX extends to a bounded operator
Then T is regular.
T ⊗IX : Lp(Ω; X) −→ Lp(Ω; X).
Proof. Since X is not K-convex, there exist a constant C ≥ 1 such that for any n ≥ 1,
X contains an n-dimensional subspace C-isomorphic to ℓ1
n. As a consequence of the tensor
extension assumption, we obtain that for any n ≥ 1,
(cid:13)(cid:13)T ⊗ Iℓ1
(cid:13)(cid:13)T ∗ ⊗ Iℓ∞
n : Lp(Ω; ℓ1
n) −→ Lp(Ω; ℓ1
n : Lp′
(Ω; ℓ∞
n ) −→ Lp′
(Ω; ℓ∞
n)(cid:13)(cid:13) ≤ CkT ⊗Xk.
n )(cid:13)(cid:13) ≤ CkT ⊗Xk.
Since Lp(Ω; ℓ1
n)∗ = Lp′(Ω; ℓ∞
n ) isometrically, this implies that for any n ≥ 1,
This shows that T ∗ is regular. By (20), the operator T is regular as well.
(cid:3)
Theorem 17. Assume that G is a non discrete locally compact abelian group and that X is
a non K-convex Banach space. Then there exists a symmetric probability measure η on G
such that for any 1 < p < ∞, C X
η2,p : Lp(G; X) → Lp(G; X) is not a Ritt operator.
Proof. Let δe denote the Dirac measure at point e (the unit of G). Since G is non discrete,
there exists a symmetric probability measure η on G such that
τ = δe + η2
is not invertible in the Banach algebra M(G). We refer to the proof of [39, Theorem 5.3.4]
for this fact, which is a generalization of the Wiener-Pitt Theorem (see in particular the last
three lines of p. 107 in the latter reference).
We fix some 1 < p < ∞. By [8, Proposition 5.2] (or by Theorem 6), Cη2,p : Lp(G) → Lp(G)
is a Ritt operator. In particular, −1 ∈ σ(Cη2,p). Thus the operator
is invertible.
S = ILp + Cη2,p
18
F. LANCIEN AND C. LE MERDY
Assume now that C X
η2,p : Lp(G; X) → Lp(G; X) is a Ritt operator. Then similarly, the
η2,p = S⊗IX hence the invertibility
operator ILp(X) + C X
of this operator implies that S−1 ⊗ X extends to a bounded operator
η2,p is invertible. By (10), ILp(X) + C X
S−1⊗IX : Lp(G; X) −→ Lp(G; X).
Since X is not K-convex, it follows from Lemma 16 that S−1 is regular.
The operator S−1 is a Fourier multiplier. Hence according to Arendt's description of
regular Fourier multipliers [2, Proposition 3.3], there exists a measure ρ ∈ M(G) such that
S−1(f ) = ρ ∗ f for any f ∈ Lp(G). By construction, S(f ) = τ ∗ f for any f ∈ Lp(G). Hence
we have ρ ∗ τ ∗ f = f for any f ∈ Lp(G). This yields ρ ∗ τ = δe and contradicts the fact that
τ is not invertible in the Banach algebra M(G).
(cid:3)
Acknowledgements. The two authors were supported by the French "Investissements
d'Avenir" program, project ISITE-BFC (contract ANR-15-IDEX-03).
References
[1] M. Akcoglu, L. Sucheston, Dilations of positive contractions on Lp spaces, Canadian Math. Bulletin 20
(1977), 285-292.
[2] W. Arendt, On the o-spectrum of regular operators and the spectrum of measures, Math. Z. 178 (1981),
271-287.
[3] C. Arhancet, C. Le Merdy, Dilation of Ritt operators on Lp-spaces, Israel J. Math. 201 (2014), no. 1,
373-414.
[4] C. Batty, A. Gomilko, Y. Tomilov, Resolvent representations for functions of sectorial operators, Adv.
Math. 308 (2017), 896-940.
[5] J. Bergh, J. Lofstrom, Interpolation spaces, an introduction, Grundlehren der Mathematischen Wis-
senschaften, No. 223. Springer-Verlag, Berlin-New York, 1976. x+207 pp.
[6] S. Blunck, Analyticity and discrete maximal regularity on Lp-spaces, J. Funct. Anal. 183 (2001), no. 1,
211-230.
[7] D. Burkholder, Martingales and singular integrals in Banach spaces, Handbook of the geometry of
Banach spaces, Vol. I, 233-269, North-Holland, Amsterdam, 2001.
[8] G. Cohen, C. Cuny, M. Lin, Almost everywhere convergence of powers of some positive Lp contractions,
J. Math. Analysis App. 420 (2014), 1129-1153.
[9] R. Coifman, G. Weiss, Transference methods in analysis, Conference Board of the Mathematical Sciences
Regional Conference Series in Mathematics, No. 31. American Mathematical Society, Providence, R.I.,
1976. ii+59 pp.
[10] C. Cuny, On the Ritt property and weak type maximal inequalities for convolution powers on ℓ1(Z),
Preprint 2016, arXiv:1601.05618.
[11] Y. Derriennic, M. Lin, Convergence of iterates of averages of certain operator representations and of
convolution powers, J. Funct. Anal. 85 (1989), no. 1, 86-102.
[12] J. Diestel, J. Uhl, Vector measures, Mathematical Surveys, No. 15. American Mathematical Society,
Providence, R.I., 1977. xiii+322 pp.
[13] N. Dungey, Subordinated discrete semigroups of operators, Trans. Amer. Math. Soc. 363 (2011), no. 4,
1721-1741.
[14] A. Gomilko, Y. Tomilov, On subordination of holomorphic semigroups, Adv. Math. 283 (2015), 155-194.
[15] A. Gomilko, Y. Tomilov, On discrete subordination of power bounded and Ritt operators, Indiana Univ.
Math. J., to appear.
[16] B. Haak, M. Haase, Square function estimates and functional calculi, Preprint 2013, arXiv:1311.0453.
THE RITT PROPERTY OF SUBORDINATED OPERATORS IN THE GROUP CASE
19
[17] M. Haase, The functional calculus for sectorial operators, Operator Theory: Advances and Applications,
169. Birkhauser Verlag, Basel, 2006. xiv+392 pp.
[18] M. Hieber, J. Pruss, Functional calculi for linear operators in vector-valued Lp-spaces via the transference
principle, Adv. Differential Equations 3 (1998), no. 6, 847-872.
[19] T. Hytonen, J. van Neerven, M. Veraar, L. Weis, Analysis in Banach spaces, Vol. I, A Series of Modern
Surveys in Mathematics 63. Springer, Cham, 2016. xvi+614 pp.
[20] R. Jones, K. Reinhold, Oscillation and variation inequalities for convolution powers, Ergodic Theory
Dynam. Systems 21 (2001), no. 6, 1809-1829.
[21] R. Jones, J. Rosenblatt, A. Tempelman, Ergodic theorems for convolutions of a measure on a group,
Illinois J. Math. 38 (1994), 521-553.
[22] N. Kalton, P. Kunstmann, L. Weis, Perturbation and interpolation theorems for the H ∞-calculus with
applications to differential operators, Math. Ann. 336 (2006), no. 4, 747-801.
[23] N. Kalton, S. Montgomery-Smith, Interpolation of Banach spaces, Handbook of the geometry of Banach
spaces, Vol. 2, 1131-1175, North-Holland, Amsterdam, 2003.
[24] N. Kalton, L. Weis, The H ∞-functional calculus and square function estimates, Nigel J. Kalton Selecta,
715-771, Birkhauser Mathematics 2015, Contemporary mathematicians, F. Gesztesy, G. Godefroy, L.
Grafakos, I. Verbitsky (Eds.).
[25] F. Lancien, C. Le Merdy, On functional calculus properties of Ritt operators, Proc. Roy. Soc. Edinburgh
Sect. A 145 (2015), 1239-1250.
[26] C. Le Merdy, H ∞-functional calculus and square function estimates for Ritt operators, Rev. Mat.
Iberoam. 30 (2014), 1149-1190.
[27] C. Le Merdy, Q. Xu, Maximal theorems and square functions for analytic operators on Lp-spaces, J.
Lond. Math. Soc. (2) 86 (2012), no. 2, 343-365.
[28] C. Le Merdy, Q. Xu, Strong q-variation inequalities for analytic semigroups, Ann. Inst. Fourier (Greno-
ble) 62 (2012), no. 6, 2069-2097 (2013).
[29] M. Lin, R. Wittmann, Ergodic sequences of averages of group representations, Ergodic Theory Dynam.
Systems 14 (1994), no. 1, 181-196.
[30] J. Lindenstrauss, L. Tzafriri, Classical Banach spaces II, Springer-Verlag, Berlin-New York, 1979. x+243
pp.
[31] Y. Lyubich, Spectral localization, power boundedness and invariant subspaces under Ritt's type condition,
Studia Math. 134 (1999), no. 2, 153-167.
[32] B. Maurey, Type, cotype and K-convexity, Handbook of the geometry of Banach spaces, Vol. 2, 1299-
1332, North-Holland, Amsterdam, 2003.
[33] B. Nagy, J. Zemanek, A resolvent condition implying power boundedness Studia Math. 134 (1999), no.
2, 143-151.
[34] O. Nevanlinna, Convergence of iterations for linear equations, Lectures in Mathematics ETH Zurich.
Birkhauser Verlag, Basel, 1993. viii+177 pp.
[35] G. Pisier, Holomorphic semigroups and the geometry of Banach spaces, Annals of Math. 115 (1982),
375-392.
[36] G. Pisier, Factorization of linear operators and geometry of Banach spaces, CBMS Regional Conference
Series in Mathematics 60, American Mathematical Society, Providence, RI, 1986. x+154 pp.
[37] G. Pisier, Complex interpolation and regular operators between Banach spaces, Arch. Math. 62 (1994),
261-269.
[38] J. L. Rubio de Francia, Martingale and integral transforms of Banach space valued functions, Probability
and Banach spaces (Zaragoza, 1985), 195-222, Lecture Notes in Math., 1221, Springer, Berlin, 1986.
[39] W. Rudin, Fourier analysis on groups, Interscience Tracts in Pure and Applied Mathematics, No. 12
Interscience Publishers, New York-London 1962 ix+285 pp.
[40] H. Schaefer, Banach lattices and positive operators, Die Grundlehren der mathematischen Wis-
senschaften, Band 215. Springer-Verlag, New York-Heidelberg, 1974. xi+376 pp.
[41] E. Stein, Topics in harmonic analysis, related to the Littlewood-Payley theory, Annals of Mathematics
Studies, Vol. 63, Princeton University Press, Princeton, NJ, 1970.
20
F. LANCIEN AND C. LE MERDY
[42] P. Vitse, A band limited and Besov class functional calculus for Tadmor-Ritt operators, Arch. Math.
(Basel) 85 (2005), no. 4, 374-385.
[43] Q. Xu, H ∞ functional calculus and maximal inequalities for semigroups of contractions of vector-valued
Lp-spaces, Int. Math. Res. Not. IMRN 2015, no. 14, 5715-5732.
[44] M. Zafran, Spectra of convolution operators on Lp-spaces, Proc. Nat. Acad. Sci. U.S.A. 72 (1975), no.
9, 3285-3286.
E-mail address: [email protected]
E-mail address: [email protected]
Laboratoire de Mathmatiques de Besanc¸on, UMR 6623, CNRS, Universit´e Bourgogne
Franche-Comt´e, 25030 Besanc¸on Cedex, FRANCE
|
1612.06270 | 1 | 1612 | 2016-12-19T17:11:53 | Amenable semigroups of nonexpansive mappings on weakly compact convex sets | [
"math.FA"
] | We show a few fixed point theorems for semigroups acting on weakly compact convex subsets of Banach spaces when $LUC(S), AP(S), WAP(S)$ or $WAP(S)\cap LUC(S)$ have a left invariant mean. In particular, we give a characterization of semitopological semigroups that have a left invariant mean on the space of weakly almost periodic functions in terms of a fixed point property for nonexpansive mappings. It answers, in the case of Banach spaces, Question 4 of [A.T.-M. Lau, Y. Zhang, J. Funct. Anal. 263 (2012), 2949--2977] in affirmative. We also extend the fixed point theorem of R. Hsu from left reversible discrete semigroups to left amenable semitopological semigroups in Banach spaces. | math.FA | math |
AMENABLE SEMIGROUPS OF NONEXPANSIVE
MAPPINGS ON WEAKLY COMPACT CONVEX SETS
ANDRZEJ WI´SNICKI
Dedicated to Professor Tom´as Dom´ınguez Benavides
on the occasion of his 65th birthday
Abstract. We show a few fixed point theorems for semigroups act-
ing on weakly compact convex subsets of Banach spaces when LU C(S),
AP (S), W AP (S) or W AP (S) ∩ LU C(S) have a left invariant mean.
In particular, we give a characterization of semitopological semigroups
that have a left invariant mean on the space of weakly almost periodic
functions in terms of a fixed point property for nonexpansive mappings.
It answers, in the case of Banach spaces, Question 4 of [A.T.-M. Lau,
Y. Zhang, J. Funct. Anal. 263 (2012), 2949–2977] in affirmative. We
also extend in Banach spaces the fixed point theorem of R. Hsu from
left reversible discrete semigroups to left amenable semitopological semi-
groups.
1. Introduction
A well-known characterization of amenable semigroups is given by the
following Day fixed point theorem: a semigroup S is (left) amenable if and
only if whenever S acts affinely (from the left) on a nonempty compact
convex subset K of a locally convex space, there is a common fixed point
of S in K. This characterization was extended to topological groups by N.
Rickert and to semitopological semigroups by T. Mitchell.
A natural question arises whether a similar characterization can be given
in terms of nonexpansive mappings. The first results in this direction were
given by W. Takahashi, T. Mitchell and A. T.-M. Lau. The following fixed
point property is a corollary of [6, Theorem 4.1] (where a more general case
of locally convex spaces was considered).
Theorem 1.1. Let S be a semitopological semigroup. Then AP (S), the
space of continuous almost periodic functions on S, has a LIM (left invariant
mean) if and only if S has the following fixed point property:
(D): Whenever the action of S on a compact convex subset K of a Banach
space is separately continuous and nonexpansive, there is a common
fixed point of S in K.
2010 Mathematics Subject Classification. Primary 47H10; Secondary 20M30, 28C10,
37C25, 43A07 47H09, 47H20, 54H25.
Key words and phrases. Amenable semigroup, Topological semigroup, Fixed point
property, Semigroup action, Nonexpansive mapping, Invariant mean, Almost periodic
function, Weakly compact set, Radon measure.
1
2
This article is motivated by the recent papers [10, 11], where the similar
results were obtained for W AP (S), the space of continuous weakly almost
periodic functions on S, and LU C(S), the space of left uniformly continuous
functions on S, under the additional assumption that S is separable (or at
least the set K on which S acts is separable). We show that it is possible to
omit the separability assumption when S acts on a weakly compact convex
subset of a Banach space. Thus we give a complete characterization of
a sentence "W AP (S) has a LIM" in terms of a fixed point property for
nonexpansive mappings that answers, in the case of Banach spaces, Question
4 in [11] in affirmative. A similar method allows us to strengthen Theorem
1.1 above. Moreover, we obtain the fixed point theorems for semigroups of
nonexpansive mappings when LU C(S) or W AP (S) ∩ LU C(S) have a left
invariant mean. In particular, the fixed point theorem of Hsu [5] (see also
[11, Theorem 3.10]) is extended from left reversible discrete semigroups to
left amenable semitopological semigroups acting on weakly compact convex
subsets of Banach spaces.
2. Preliminaries
Let S be a semitopological semigroup, i.e., a semigroup with a Hausdorff
topology such that the mappings S ∋ s → ts and S ∋ s → st are continuous
for each t ∈ S. Notice that every semigroup can be equipped with the
discrete topology and then it is called a discrete semigroup. Let ℓ∞(S)
be the Banach space of bounded complex-valued functions on S with the
supremum norm. For s ∈ S and f ∈ ℓ∞(S), we define the left and right
translations of f in ℓ∞(S) by
lsf (t) = f (st) and rsf (t) = f (st)
for every t ∈ S. Let X be a closed linear subspace of ℓ∞(S) containing
constants and invariant under translations, i.e., ls(X) ⊂ X and rs(X) ⊂ X.
Then a linear functional m ∈ X ∗ is called a left invariant mean on X (LIM,
for short), if kµk = µ(1) = 1 and
µ(lsf ) = µ(f )
for each s ∈ S and f ∈ X. Similarly, we can define a right invariant mean.
Denote by C(S) the closed subalgebra of ℓ∞(S) consisting of continuous
functions and let LU C(S) be the space of left uniformly continuous func-
tions on S, i.e., all f ∈ C(S) such that the mapping S ∋ s → lsf from S
to C(S) is continuous when C(S) has the sup norm topology. A semigroup
S is called left amenable if there exists a left invariant mean on LU C(S).
Left uniformly continuous functions are often called in the literature "right
uniformly continuous" and vice-versa.
Two other subspaces of C(S) are very important in this context. A
bounded continuous function f on S is called almost periodic if {lsf : s ∈ S}
(equivalently, {rsf : s ∈ S}) is relatively compact in the norm topology
of C(S). A bounded continuous function f on S is called weakly almost
periodic if {lsf : s ∈ S} (equivalently, {rsf : s ∈ S}) is relatively compact
in the weak topology of C(S). The space of almost periodic (resp., weakly
3
almost periodic) functions on S is denoted by AP (S) (resp., W AP (S)). In
general,
AP (S) ⊂ LU C(S) and AP (S) ⊂ W AP (S),
and if S is discrete, then
AP (S) ⊂ W AP (S) ⊂ LU C(S) = ℓ∞(S).
Let K be a topological space. A semigroup S is said to act on K (from
the left) if there is a map
such that
ψ : S × K → K
ψ (s, ψ (s′, x)) = ψ (ss′, x)
for all s, s′ ∈ S and x ∈ K. We write ψ (s, x) = s · x = Tsx. The action is
said to be separately continuous if it is continuous in each of the variables
when the other is fixed. If K is a subset of a Banach space E, the action ·
is called nonexpansive if
ks · x − s · yk ≤ kx − yk
for every x, y ∈ K and s ∈ S. Given an action · of S, an element x ∈ K is
called a common fixed point for S in K if s · x = x for every s ∈ S.
If X is a closed linear subspace of ℓ∞(S) invariant under translations then
X is said to be (left) introverted if for any ϕ ∈ X ∗ and f ∈ X, the function
h(s) = ϕ(lsf ), s ∈ S, belongs to X. It is known that AP (S), W AP (S) and
LU C(S) are introverted subspaces of ℓ∞(S) (see, e.g., [14, Prop. 2.11]).
Since we work in Banach spaces rather than in locally convex topological
spaces we shall need the following theorem of Granirer and Lau (compare
[3, Theorem 1], [11, Remark 6.4], and, in the case of locally compact groups,
[14, Theorem 2.13]).
Theorem 2.1. Let X be an introverted subspace of ℓ∞(S) with 1 ∈ X. Then
X has a LIM if and only if cop{rsf : s ∈ S}, the pointwise closure of the
convex hull of the right orbit of f , contains a constant function for every
f ∈ X.
Furthermore, X (containing constants) is introverted if and only cop{rsf :
s ∈ S} ⊂ X (see [3, Lemma 2]).
Let µ be a (locally finite, positive) Radon measure on a topological space
K, i.e., a Borel measure which is inner regular: µ(A) = sup{µ(C) : C ⊂ A, C
compact} for any measurable set A. Recall that the support of µ is defined
as the complement of the set of points that have neighborhoods of measure
0.
In general, the support may be empty but if K is (at least) locally
compact and µ a finite Radon measure, then µ(supp(µ)) = µ(K). The
crucial observation for this paper is the following theorem of Grothendieck
[4].
Theorem 2.2. Every finite Radon measure on a weakly compact set in a
Banach space has a norm-separable support.
4
Proof. We sketch the proof following Todorcevi´c [17, Theorem 9.1], see,
e.g., [12, Theorem 4.3] for two detailed proofs. Since S = supp(µ) satisfies
the countable chain condition and is weakly compact, it follows from [15,
Theorem 1.4] that every weakly compact subset of C(S) is separable. Hence
C(S) is separable as a weakly compactly generated space by [1, Prop. 1].
It follows that S is also separable.
(cid:3)
3. Fixed point theorems
Let S be a semitopological semigroup. An action · : S × K → K on a
Hausdorff space K is called equicontinuous if the family of functions {K ∋
x → s · x ∈ K}s∈S is equicontinuous in the usual sense. It is called quasi-
equicontinuous if ¯Sp, the closure of {K ∋ x → s · x ∈ K}s∈S in K K with the
product topology, consists of only continuous mappings. If K is a subset
of a Banach space E, then the action on K is called weakly equicontinuous
(resp., weakly quasi-equicontinuous) if it is equicontinuous (resp., quasi-
equicontinuous) when K is equipped with the weak topology of E.
The aim of this section is to prove the following fixed point properties.
Theorem 3.1. W AP (S) has a LIM if and only if S has the following fixed
point property:
(F ): Whenever S acts on a weakly compact convex subset K of a Banach
space and the action is weakly separately continuous (i.e., separately
continuous when K is equipped with the weak topology), weakly quasi-
equicontinuous and nonexpansive, then K contains a common fixed
point for S.
In the case of Banach spaces, the above theorem drops the separability
assumption from [10, Theorem 3.4] and answers in affirmative [11, Question
4] in that case.
Theorem 3.2. AP (S) has a LIM if and only if S has the following fixed
point property:
(E): Whenever S acts on a weakly compact convex subset K of a Ba-
nach space and the action is weakly separately continuous, weakly
equicontinuous and nonexpansive, then K contains a common fixed
point for S.
It removes the separability assumption from [10, Theorem 3.6] in the case
of Banach spaces and thus strengthen Theorem 1.1 which was alluded to in
the introduction.
In addition to the above characterization theorems, we obtain the follow-
ing fixed point theorems.
Theorem 3.3. If W AP (S) ∩ LU C(S) has a LIM, then S has the following
fixed point property:
(F ∗): Whenever S acts on a weakly compact convex subset K of a Banach
space and the action is (jointly) weakly continuous, weakly quasi-
equicontinuous and nonexpansive, then K contains a common fixed
point for S.
5
It drops the separability assumption from the part of [10, Theorem 5.1].
It would be interesting to prove also the reverse implication as it was done
there in the case of locally convex spaces.
Theorem 3.4. If LU C(S) has a LIM, then S has the following fixed point
property:
(G∗): Whenever S acts on a weakly compact convex subset K of a Banach
space and the action is (jointly) weakly continuous and nonexpansive,
then K contains a common fixed point for S.
Note that Hsu [5] proved that in the more general case of locally con-
vex spaces, left reversible discrete semigroups have property (G∗) and Lau
and Zhang [10, Theorem 5.4] generalized Hsu's result to left reversible,
metrizable semitopological semigroup. Theorem 3.4 extends it also to left
amenable semigroups (in Banach spaces).
The following diagram summarizes the known relations among the fixed
point properties of semitopological semigroups acting on weakly compact
convex subsets of Banach spaces (compare the diagram on p. 2553 in [10]):
LU C(S)
has LIM
⇓
S is left reversible
& metrizable
⇒
(G∗)
⇒
(F ∗)
⇐ W AP (S) ∩ LU C(S) has LIM
⇑
(G)
⇒
⇑
(F )
m
W AP (S)
has LIM.
⇒ (E) ⇔ (D) ⇔ AP (S) has LIM
We begin with a general lemma which is patterned after [6, Theorem 4.1]
and [8, Lemma 5.1]. We will denote by C(K) the space of weakly continuous
complex-valued functions defined on a weakly compact set K.
Lemma 3.5. Let S be a semitopological semigroup and X a closed linear
subspace of ℓ∞(S) containing constants and invariant under translations.
Suppose that S acts on a weakly compact subset K of a Banach space E so
that the action is weakly separately continuous and there exists y ∈ K such
that for every f ∈ C(K), the function S ∋ s → fy(s) = f (s·y) belongs to X.
If X has a left invariant mean, then there exists a nonempty weakly compact
and norm-separable subset K0 of K such that sK0 = {s · x : x ∈ K0} = K0
for every s ∈ S.
Proof. Let m be a left invariant mean on X and define a positive functional
Φ on C(K) by
Φ(f ) = m(fy)
6
for every f ∈ C(K). Notice that Φ is well-defined since, by assumption,
fy ∈ X, and kΦk = 1. Define
tf (x) = f (t · x), x ∈ K,
for every t ∈ S and f ∈ C(K). Then tf : K → C is weakly continuous since
the action · is weakly separately continuous and
Φ(f ) = Φ(tf ).
Let µ be the probability Radon measure on K corresponding to Φ. Then
µ(A) = µ(s−1A) for every Borel subset A of K (with the weak topology)
and s ∈ S, where as usual, s−1A = {x ∈ K : s · x ∈ A} . Define K0 =
supp(µ) and notice that µ(s−1K0) = µ(K0) = 1. Hence K0 ⊂ s−1K0 since
s−1K0 is weakly closed. Similarly, µ(sK0) = µ(s−1(sK0)) = µ(K0) = 1
and consequently K0 ⊂ sK0 (notice that sK0 is weakly compact). Thus
sK0 = K0 for every s ∈ S. Moreover, K0 is nonempty, weakly compact and
separable by Theorem 2.2.
(cid:3)
We can now state a general fixed point theorem for a semitopological
semigroup S. The proof follows [8, Theorem 5.3]. We sketch it for the
convenience of the reader.
Theorem 3.6. Let S be a semitopological semigroup and X a closed linear
subspace of ℓ∞(S) containing constants and invariant under translations.
Suppose that S acts on a weakly compact subset K of a Banach space E
so that the action is weakly separately continuous, nonexpansive and the
function S ∋ s → fy(s) = f (s · y) belongs to X for every y ∈ K and every
f ∈ C(K). If X has a left invariant mean then there is a common fixed
point of S in K.
Proof. By Kuratowski-Zorn's lemma, there exists a nonempty minimal weakly
compact and convex subset C of K which is invariant under S. Let F be a
nonempty minimal weakly compact subset of C which is invariant under S.
Fix y ∈ F and notice that if f ∈ C(F ), then fy = ¯fy ∈ X, where ¯f : K → C
is an extension of f to the whole set K. Applying Lemma 3.5 there exists a
weakly compact and norm-separable subset K0 of F such that sK0 = K0 for
every s ∈ S. From minimality of F , K0 = F is separable and {s · x : s ∈ S}
is weakly dense in F for every x ∈ F. Thus F is norm-compact by [8, Lemma
5.2] (which in turn follows the ideas of Hsu [5] related to "fragmentability",
see [13]). Suppose that r = diam F > 0. Then by [2, Lemma 1], there is
u ∈ co F ⊂ C such that r0 = sup{ku − yk : y ∈ F } < r. Let
C0 = {x ∈ C : kx − yk ≤ r0 for all y ∈ F }.
Then u ∈ C0 and C0 is a weakly compact convex proper subset of C. Since
the action is nonexpansive and sF = F it follows that sC0 ⊂ C0 for all s ∈ S
which contradicts the minimality of C. Thus diam F = 0 and F consists of
a single point which is a common fixed point of S in K.
(cid:3)
Theorem 3.6, together with Theorem 2.1 and some results from [10, 11],
allows proving theorems stated at the beginning of this section in a compact
way. The remainder of this section will be devoted to their proofs.
7
Proof of Theorem 3.1. Assume that W AP (S) has a LIM. Since the action
is weakly separately continuous and weakly quasi-equicontinuous, it follows
from [10, Lemma 3.2] that the function fy(s) = f (s · y), s ∈ S, belongs
to W AP (S) for every f ∈ C(K) and y ∈ K. Thus, the assumptions of
Theorem 3.6 are satisfied with X = W AP (S) and we obtain a common
fixed point of S in K.
To prove the reverse implication, we follow [11, Prop. 6.5]. Assume that
the semigroup S is a monoid. It does not lose the generality, see [11, Lemma
6.3]. Fix f ∈ W AP (S) and let Kf = co{rsf : s ∈ S} be the norm closure
of the convex hull of the right orbit of f. Then Kf is weakly compact, and
notice that the mapping (s, g) → rsg, s ∈ S, g ∈ Kf , defines a left action of
S on Kf . Indeed,
rs′rsg(t) = rsg(ts′) = g(ts′s) = rs′sg(t)
for every s, s′, t ∈ S, g ∈ Kf , and consequently,
rs(α1rs1f + ... + αnrsnf ) = α1rss1f + ... + αnrssnf
so that rs′(co{rsf : s ∈ S}) ⊂ co{rsf : s ∈ S} for every s′ ∈ S. Furthermore,
the action is affine and nonexpansive:
sup
t∈S
krsg(t) − rsh(t)k = sup
t∈S
kg(ts) − h(ts)k ≤ sup
t∈S
kg(t) − h(t)k .
Thus rs(Kf ) ⊂ Kf and rs is weakly continuous on Kf for every s ∈ S. To
prove that the mapping S ∋ s → rsg ∈ Kf is also continuous for every fixed
g ∈ Kf , choose a net (sα) ⊂ S converging to s in the topology of S. Then
rsg(t) = g(t lim
α
sα) = lim
α
g(tsα) = lim
α
rsαg(t), t ∈ S,
and since Kf is weakly compact, the pointwise limit limα rsαg(·) must agree
with the weak limit w-limα rsαg. It follows that the action S ×Kf ∋ (s, g) →
rsg ∈ Kf is weakly separately continuous.
The next step is to show that this action is also weakly quasi-equicontinuous.
We follow the argument in [10, Theorem 3.4]. Choose a net (sα) ⊂ S
such that w-limα rsαg = T (g) for each g ∈ Kf . We have to show that
T : Kf → Kf is weakly continuous. Suppose on the contrary that there
exists a net (gβ) ⊂ Kf such that w-limβ gβ = g ∈ Kf but (T (gβ))β does not
converge weakly to T (g). Then there exists ε > 0, a functional ϕ ∈ E ∗ and
a subnet of (gβ) (still denoted by (gβ)) such that Re(hϕ, T (gβ) − T (g)i > ε
for all β. By Mazur's lemma, there is a net (hλ) ⊂ co(gβ) which con-
verges in norm to g. But T is affine since all rsα are affine and hence
Re(hϕ, T (hλ) − T (g)i > ε for all λ. On the other hand,
hϕ, T (hλ) − T (g)i ≤ kϕk
(cid:13)
(cid:13)
(cid:13)
w- lim
α
rsαhλ − w- lim
α
rsαg
(cid:13)
(cid:13)
(cid:13)
≤ kϕk lim sup
krsαhλ − rsαgk ≤ kϕk khλ − gk λ−→ 0,
α
and we obtain a contradiction which shows that the action is weakly quasi-
equicontinuous. By the fixed point property (F ), for every f ∈ W AP (S)
there exists g ∈ Kf such that rsg = g for every s ∈ S. It follows that
g(ts) = g(t) for every t ∈ S and taking t = e, the identity of S, we have
8
that g(s) = g(e) is a constant function. Since W AP (S) is introverted and
Kf = cop{rsf : s ∈ S} contains a constant function for every f ∈ W AP (S),
the conclusion follows from Theorem 2.1.
(cid:3)
Proof of Theorem 3.2. Assume that AP (S) has a LIM. Since the action is
weakly separately continuous and weakly equicontinuous, fy ∈ AP (S) for
every f ∈ C(K) and y ∈ K by [6, Lemma 3.1]. Applying Theorem 3.6 with
X = AP (S) we get a common fixed point of S in K.
To prove the converse, assume that S is a monoid. Fix f ∈ AP (S) and
let Kf = co{rsf : s ∈ S}. Now Kf is compact and define, as before, a left
action of S on Kf by
S × Kf ∋ (s, g) → rsg ∈ Kf .
The action is weakly separately continuous and nonexpansive, and hence
weakly equicontinuous since the weak topology coincides with the norm
topology on Kf . By property (E), for every f ∈ AP (S) there exists g ∈ Kf
such that rsg = g for every s ∈ S. The rest of the proof runs as before. (cid:3)
Proof of Theorem 3.3. By assumption, the action is jointly weakly con-
tinuous and weakly quasi-equicontinuous, hence fy ∈ W AP (S) for every
f ∈ C(K) and y ∈ K as in the proof of Theorem 3.1. Furthermore,
fy ∈ LU C(S) by [8, Lemma 5.1] (joint weak continuity is enough here).
Therefore, if W AP (S) ∩ LU C(S) has a LIM, we can apply Theorem 3.6
with X = W AP (S) ∩ LU C(S) to obtain a common fixed point of S in
K.
(cid:3)
Proof of Theorem 3.4. If the action is jointly weakly continuous then fy ∈
LU C(S) for every f ∈ C(K) and y ∈ K by [8, Lemma 5.1]. Since the
action is also nonexpansive and LU C(S) has a left invariant mean, the
result follows from Theorem 3.6.
(cid:3)
Recall that a bounded convex subset C of a Banach space has normal
structure if r(C) < diam C, where r(C) = inf x∈C supx∈C kx − yk is the
Chebyshev radius of C. Clearly, Theorem 3.2 shows that if AP (S) has a
left invariant mean then a semitopological semigroup S has the following
property:
(E ′): Whenever S acts on a weakly compact convex subset K of a Banach
space with normal structure and the action is weakly separately con-
tinuous, weakly equicontinuous and nonexpansive, then K contains
a common fixed point for S.
Notice that since norm-compact convex subsets of Banach spaces have
normal structure, we can follow the proof of Theorem 3.2 and show that
the reverse implication is also true. Thus property (E ′) is equivalent to
property (E). A similar result was proved in [10] for separable semigroups
in the case of locally convex spaces.
The results of this paper have a natural generalization to semigroups
acting on weak-star compact convex subsets of a dual Banach space with the
9
Radon-Nikodym property and, more generally, on norm-fragmented spaces.
This problem will be studied in a subsequent publication.
References
[1] D. Amir, J. Lindenstrauss, The structure of weakly compact sets in Banach spaces,
Ann. of Math. (2) 88 (1968), 35–46.
[2] R. DeMarr, Common fixed points for commuting contraction mappings, Pacific J.
Math. 13 (1963), 1139–1141.
[3] E. Granirer, A.T.-M. Lau, Invariant means on locally compact groups, Illinois J.
Math. 15 (1971), 249–257.
[4] A. Grothendieck, Sur les applications lin´eaires faiblement compactes d'espaces du
type C(K), Canadian J. Math. 5 (1953), 129–173.
[5] R. Hsu, Topics on weakly almost periodic functions, PhD thesis, SUNY at Buffalo,
1985.
[6] A. T.-M. Lau, Invariant means on almost periodic functions and fixed point prop-
erties, Rocky Mountain J. Math. 3 (1973), 69–76.
[7] A. T.-M. Lau, Amenability and fixed point property for semigroup of nonexpansive
mappings, in: Fixed Point Theory and Applications, M.A. Thera, J.B. Baillon (eds.),
Longman Sci. Tech., Harlow, 1991, 303–313.
[8] A. T.-M. Lau, W. Takahashi, Invariant means and fixed point properties for non-
expansive representations of topological semigroups, Topol. Methods Nonlinear
Anal. 5 (1995), 39–57.
[9] A. T.-M. Lau, W. Takahashi, Fixed point and non-linear ergodic theorems for semi-
groups of non-linear mappings, in: Handbook of Metric Fixed Point Theory, W. A.
Kirk, B. Sims (eds.), Kluwer Academic Publishers, Dordrecht, 2001, 515–553.
[10] A.T.-M. Lau, Y. Zhang, Fixed point properties of semigroups of non-expansive map-
pings, J. Funct. Anal. 254 (2008), no. 10, 2534–2554.
[11] A.T.-M. Lau, Y. Zhang, Fixed point properties for semigroups of nonlinear mappings
and amenability, J. Funct. Anal. 263 (2012), 2949–2977.
[12] J. Lindenstrauss, Weakly compact sets – their topological properties and the Ba-
nach spaces they generate, in: Symposium on Infinite-Dimensional Topology, R. D.
Anderson (ed.), Ann. of Math. Studies 69, Princeton Univ. Press, Princeton, N. J.,
1972, pp. 235–273.
[13] I. Namioka, Fragmentability in Banach spaces:
interaction of topologies, Rev. R.
Acad. Cienc. Exactas F´ıs. Nat. Ser. A Math. RACSAM 104 (2010) 283–308.
[14] A. L. T. Paterson, Amenability, American Mathematical Society, Providence, RI,
1988.
[15] H. P. Rosenthal, On injective Banach spaces and the spaces C(S), Bull. Amer. Math.
Soc. 75 (1969), 824–828.
[16] W. Takahashi, Fixed point theorem for amenable semigroup of nonexpansive map-
pings, K¯odai Math. Sem. Rep. 21 (1969), 383–386.
[17] S. Todorcevi´c, Chain-condition methods in topology, Topology Appl. 101 (2000),
45–82.
Andrzej Wi´snicki, Department of Mathematics, Rzesz´ow University of
Technology, Al. Powsta´nc´ow Warszawy 12, 35-959 Rzesz´ow, Poland
E-mail address: [email protected]
|
1711.09785 | 1 | 1711 | 2017-11-27T15:51:22 | On compactness in $L^0$-modules | [
"math.FA"
] | Several results in functional analysis are extended to the setting of $L^0$-modules, where $L^0$ denotes the ring of all measurable functions $x\colon \Omega\to \mathbb{R}$. The focus is on results involving compactness. To this end, a notion of stable compactness is introduced, and it is argued that the conventional notion of compactness does not allow to establish a functional analytic discourse in $L^0$-modules. Several characterizations of stable compactness are discussed, and its importance in applications is highlighted. | math.FA | math |
ON COMPACTNESS IN L0-MODULES
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
Abstract. Several results in functional analysis are extended to the setting of L0-
modules, where L0 denotes the ring of all measurable functions x : Ω → R. The focus
is on results involving compactness. To this end, a notion of stable compactness is
introduced, and it is argued that the conventional notion of compactness does not
allow to establish a functional analytic discourse in L0-modules. Several character-
izations of stable compactness are discussed, and its importance in applications is
highlighted.
Key words and phrases: compactness, L0-modules, conditional analysis, probabilistic
analysis, measurable selections, random set theory.
1. Introduction
i=1 x2
form {x ∈ (L0)d : kxk < r a.e.} where kxk := (Pd
Let (Ω, F, P) be a probability space. Let L0 = L0(Ω, F, P) denote the space of
all real-valued Borel measurable functions modulo almost everywhere identity. The
locally L0-convex topology on (L0)d, d > 1, is defined through neighbourhoods of the
i )1/2 is an L0-valued norm and
r ∈ L0 with r > 0 a.e., see [7, 14]. The space (L0)d is a topological L0-module of
rank d. By verifying that (L0)d is anti-compact1 in general, we confirm the observation
in the literature ([7, 10, 18]) that there is no hope to extend theorems in analysis
to (L0)d with a classical notion of compactness, for example a Heine-Borel theorem.
More generally, we prove anti-compactness for arbitrary stable topological L0-modules.
In [10], a notion of conditional compactness is introduced within the conditional set
theory, and it was shown that it allows to extend results to the setting of conditional
locally convex topological vector spaces. The interpretation of conditional compactness
in L0-modules is called stable compactness which seems to be a suitable substitute of
classical compactness in modules, as it permits to extend a thorough functional analytic
discourse to the setting of L0-modules which is illustrated in this article. Roughly
speaking, a subset of a stable topological L0-module is said to be stable compact if
for every open covering O of S there exists a measurable countable partition (Ak) of
Ω such that S is covered by finitely many elements of O on each Ak.
In the finite
dimensional case (L0)d characterizations of stable compactness exist in the literature,
see [7], which are applied to solve stochastic control problems in [6, 5, 31] and to
establish a Brouwer fixed point theorem in [13]. We will discuss further equivalent
characterizations of stable compactness. One of them builds a unique correspondence
to measurable selection theory, and thus closes one gap between L0-module theory and
probabilistic analysis theory (see e.g. [41]) and random set theory (see e.g. [37]).
2010 Mathematics Subject Classification. 46H25,54D30,03C90.
The authors would like to thank Michael Kupper for advice.
Asgar Jamneshan is supported by DFG project KU-2740/2-1. Jos´e M. Zapata is supported by the
grant MINECO MTM2014-57838-C2-1-P.
1A topological space X is called anti-compact if any compact set in X is finite.
1
2
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
For an arbitrary L0-module E, a locally L0-convex topology can be defined by a
family P of L0-seminorms, see [14]. Similarly as in topological vector spaces, one in-
troduces neighbourhoods of the form {x ∈ E : supp∈N p(x) < r} where N ⊂ P is finite.
As shown in [14], the locally L0-convex topology allows to extend the Hahn-Banach ex-
tension and separation theorems and the Fenchel-Moreau theorem to L0-modules. The
L0-scalar multiplication defines a function F × E → E given by (A, x) 7→ 1Ax. It is said
that E is stable under countable pastings, or stable for short, if for every measurable
countable partition (Ak) of Ω and every sequence (xk) in E there exists a unique x ∈ E
such that 1Ak x = 1Ak xk for all k. We call a module topology T on E stable whenever
E is stable and T admits a topological base consisting of a stable collection of stable
subsets of E, see [10]. The hyperplane separation theorem in [14] is proved under the
assumption that the inducing family P of L0-seminorms is closed under finite suprema
and countable pastings, which is equivalent to the property that the locally L0-convex
topology on E induced by P is in fact a stable topology, see the discussion in Section
4 for details. The abstraction of stability leads to the notion of a conditional set in
[10]. As a result, all objects in conditional set theory satisfy stability under pastings.2
Conditional set theory provides a formalism to establish systematically conditional ver-
sions of results in linear algebra, topology, measure theory, functional analysis, etc., see
[10, 28, 30, 32, 39, 43]. It follows from the definition of a conditional set that there is a
close connection between classical structures and corresponding conditional structures.
This connection is systematically studied in [10], and is reflected in all definitions,
propositions and theorems there. In particular, a one-to-one correspondence between
conditional topological base, conditional convergence, and conditional continuity and
respective classical structures is established, see [10, Section 3], and more specifically
[10, Propositions 3.5, 3.11, 3.22]. Moreover, in the context of the associated measure
algebra, an isometric isomorphism between conditional real numbers and L0 is proved
in [10, Theorem 4.4], which in particular implies that it follows from the definition of
a conditional vector space that its underlying classical structure is a stable L0-module,
see [10]. We verify that this correspondence is an equivalence of categories, which en-
ables to study the classical meaning of results in conditional set theory in L0-modules.
We start by studying algebraic properties in Section 3, where we argue that the con-
ventional notion of an algebraic basis is limited to finitely ranked L0-modules. We
suggest a notion of stable basis, obtained as an interpretation of a vector space basis
in conditional set theory, to work with in infinite dimensional settings.
A different type of topology which is well studied in the context of L0-modules is the
(ǫ, λ)-topology or L0-topology which can be viewed as a generalization of the topology
of convergence in probability, see [21, 26]. Several results in functional analysis were
extended to the setting of L0-modules endowed with an (ǫ, λ)-topology, see [18, 20, 21]
and the references therein. Using the conventional notion of compactness, it is proved
in [18] that the Banach-Alaoglu theorem holds in an L0-module E endowed with an
(ǫ, λ)-topology if and only if E has an essentially atomic support. We prove that if E
is endowed with a stable topology then it is anti-compact if and only if the support
of E is atomless. However, stable compactness allows to establish an extension of the
the Banach-Alaoglu theorem to L0-modules in full generality. Moreover, we provide,
among others, an L0-module version of the Heine-Borel theorem, Tychonoff's theorem,
the Eberlein-Smulian theorem, and the James' compactness theorem. Further results
in functional analysis are extended to L0-modules as well, e.g. the Bipolar theorem,
2Conditional set theory is formalized in the context of arbitrary complete Boolean algebras, so that
pastings might be uncountable.
STABLE COMPACTNESS
3
the Baire category theorem, the uniform boundedness principle, Banach's fixed point
theorem, and the `Arzela-Ascoli theorem. In [19], connections between the locally L0-
convex topology and the (ǫ, λ)-topology are established, see also [22, 23]. In Section
4 we provide an analysis of the relations between the stable topology, the locally L0-
convex topology and the (ǫ, λ)-topology, which combined with the results in [19], allow
to derive extensions of functional analytic results with respect to all three types of
topologies in the last two sections.
Related literature. The L0-topology in a module setting is introduced in [26, Section
5] to study the local behaviour of Bochner-Lebesgue spaces. Motivated by problems in
probabilistic analysis, the (ǫ, λ)-topology is introduced in L0-modules, see the survey
in [22, 23] for a historical background, and see [41] for an introduction to probabilistic
analysis. The locally L0-convex topology is introduced in [14], motivated by prob-
lems in mathematical finance, see [7, 13, 33] for further results in this direction. The
stable topology is a result of conditional set theory and the connections between clas-
sical and conditional structures in [10]. Conditional set theory is closely related to
Scott-Solovay-Vopenka's Boolean-valued models of set theory, see [3] for an introduc-
tion, and to toposes of sheaves on complete Boolean algebras, see [36] for an intro-
duction. In [28, 29], the connection between conditional sets and toposes of sheaves
is studied. The term Boolean-valued analysis refers to the application of Boolean-
valued models to the standard model of ZFC, see for a thorough account the excellent
monographs [34, 35] and their extensive list of references. In [34] the connection be-
tween vector spaces in a Boolean-valued model and modules in the standard model
are investigated systematically. Conditional analysis is applied to stochastic control
problems, dynamic risk sharing, representation of preferences, duality of conditional
risk measures, equilibrium pricing, Principal-Agent problems and vector duality in
[2, 4, 5, 6, 8, 9, 12, 15, 16, 31, 38, 39].
The rest of this paper is organized as follows. Section 2 fixes the notations and intro-
duces the preliminaries. In Section 3 algebraic aspects of L0-modules are discussed. In
Section 4 the connection among the three types of L0-module topologies and conditional
locally convex topological vector spaces is analyzed. In Section 5 stable compactness
is introduced and characterized. In Section 6 several functional analytic results are ex-
tended to L0-modules and stable metric spaces for different types of module topologies.
2. Preliminaries
We denote by boldface letters objects in conditional set theory and with non-boldface
ones those in classical set theory. Throughout we fix a probability space (Ω, F, P),
unless we mention otherwise. Measurable sets in Ω are denoted by the upper case
latin letters A, B, C and in their indexed forms. We will always identify two functions
on Ω or measurable sets in Ω if they are equal almost everywhere. In the latter case
this identification leads to the associated measure algebra which is denoted by A =
(A, ∧, ∨, c, 1, 0), see [17, Chapter 31] for an introduction to measure algebras. We
denote the equivalence classes in A by the lower case latin letters a, b, c and in their
indexed forms.
If we use both notations A and a, B and b, etc., then it is always
understood that A is a representative of a, B is a representative of b, etc. Let p(a),
a ∈ A, denote the set of all partitions of a, i.e. a family (ai) consisting of pairwise
disjoint elements in A with ∨ai = a. Recall the following two facts which are crucial
in the following analysis.
(i) A is a complete Boolean algebra.
4
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
(ii) A satisfies the countable chain condition, that is, all partitions are at most
countable.
An important property will be a stability property w.r.t. countable gluings which we
will impose on sets, functions, sequences, modules, topologies, filters, etc. We name
this property stable under countable concatenations, or stable for short, e.g. a stable
set, a stable module, a stable topology, etc. The essence of this stability property is a
conditional set, see [10] for an introduction to conditional set theory. In the following,
basic concepts in conditional set theory are summerized.
Definition 2.1. Let X be a non-empty set. A collection X of objects xa for a ∈ A
and x ∈ X is said to be a conditional set with carrier X, if
(i) xa = yb implies a = b,
(ii) Consistency: xa = ya and b 6 a imply xb = yb,3
(iii) Stability: for all (ak) ∈ p(1) and every countable family (xk) in X there exists
a unique x ∈ X such that xak = xkak for all k. We call x the concatenation
of (xk) along (ak), and write the formal expression x =P xkak.
Example. Let X be a non-empty set. We denote by L0
s(X) the set of all step functions
x = Pk 1Ak xk : Ω → X where (ak) ∈ p(1) and Pk 1Ak xk denotes the function with
value xk on Ak for all k. Identify on L0
s(X) × A two pairs (x, a) and (y, b) whenever
a = b and x(ω) = y(ω) for almost all ω ∈ A. Denote by xa the equivalence class of
(x, a). Then the collection Xs of all xa, a ∈ A and x ∈ L0
s(X), is a conditional set
with carrier L0
s(N) is the carrier of the conditional
⋄
natural numbers Ns.
Let X be a conditional set with carrier X. A subset Y ⊂ X is said to be stable, if
s(X). The set of step functions L0
Y 6= ∅ and P ykak ∈ Y for all (yk) ⊂ Y and (ak) ∈ p(1). A stable subset Y ⊂ X
induces the conditional set Y := {ya : y ∈ Y, a ∈ A} which is called a conditional
subset of X. Let Y ⊂ X be non-empty. Then
st(Y ) :=nX ykak : (yk) ⊂ Y, (ak) ∈ p(1)o .
is a stable set which is called the stable hull of Y . For a countable family (Yk) of stable
subsets of X and (ak) ∈ p(1), define the stable set
A collection Y of stable subsets of X is said to be a stable collection of stable subsets, if
X Ykak :=nX ykak : yk ∈ Yk for each ko .
Y 6= ∅ and P Ykak ∈ Y for all (Yk) ⊂ Y and (ak) ∈ p(1). Similarly as above, a stable
collection Y of stable sets induces a conditional collection of conditional sets, see [10,
Section 2]. For a non-empty collection Y of subsets of X, we form its stable hull by
st(Y ) :=nX st(Yk)ak : (Yk) ⊂ Y , (ak) ∈ p(1)o .
family (Xi) of carriers defines in a natural way a conditional set which is called the
conditional Cartesian product of the family (Xi). Let X and Y be conditional sets. A
Given a non-empty family (Xi) of conditional sets, the direct product Q Xi of the
relation S ⊂ X × Y is said to be stable, if (P xkak,P ykak) ∈ S whenever (xk, yk) ∈ S
stable relation S. A function f : X → Y is said to be stable, if f (P xkak) =P f (xk)ak
for all (xk) ⊂ X and (ak) ∈ p(1). A conditional function f : X → Y is the conditional
set Gf induced by the graph Gf of a stable function f : X → Y . A family (xi) = (xi)i∈I
for all k and (ak) ∈ p(1). A conditional relation S is the conditional set induced by a
3In every Boolean algebra, b 6 a is defined by b = b ∧ a.
STABLE COMPACTNESS
5
of elements in X is said to be stable, if it is parametrized by a stable function I → X.
Let (R, +, ·, 6) be the totally ordered field of real numbers. Let L0 denote the space
of all real-valued Borel measurable functions on (Ω, F, P). Consider on L0 the order of
almost everywhere dominance. We will always understand equalities and inequalities
between measurable functions in the almost everywhere sense. Further, let L0
+ :=
{r ∈ L0 : r > 0} and L0
++ := {r ∈ L0 : r > 0}. Recall that (L0, +, ·, 6) is a
Dedekind complete Riesz algebra. For a subset X ⊂ L0, we write sup X = ess sup X
and inf X = ess inf X whenever these quantities exist. Define on L0 × A the equivalence
relation (r, a) ∼ (t, b), if a = b and 1Ar = 1At, and denote by ra the equivalence class
of (r, a). Then the collection R of all equivalence classes ra is a conditional set with
carrier L0, where the concatenation of (xk) along (ak) corresponds to Pk 1Ak xk :=
1A1x1 + 1A2x2 + . . .. Addition and multiplication are stable functions which define the
structure of a conditional field on R, and the order of almost everywhere dominance
is a stable relation which defines a conditional total order on R such that (R, +, ·, 6)
is a conditional totally ordered field, see [10, Section 4].
+ and
L0
++ correspond to the conditional set of positive and strictly positive conditional real
numbers, respectively.
In this context, L0
Recall that L0
as a subset of L0
parametrized by a stable function L0
n ∈ L0
s(N) is the carrier of the conditional natural numbers. We understand N
s(N) via the embedding n 7→ 1Ωn. A stable family (xn) = (xn)n∈L0
s(N)
s(N) → X is called a stable sequence. For each
s(N), define the stable set
{1 6 m 6 n} := {m ∈ L0
s(N) : 1 6 m 6 n}.
A stable set X is said to be stable finite, if there exists a stable bijection {1 6 m 6
n} → X for a unique n ∈ L0
s(N). If X is stable finite, then there exist (ak) ∈ p(1),
(nk) ⊂ N and a countable family (Xk) of subsets of X with the cardinality of Xk being
equal to nk for all k such that X =P st(Xk)ak.
3. Stable L0-modules and conditional vector spaces
In this section, we establish an equivalence of categories between the category of
stable L0-modules and the category of conditional real vector spaces. We show that a
stable L0-module is free if and only if it is the direct sum of finitely many copies of L0.
We construct a stable algebraic basis which can serve as a useful replacement in cases
where we lack a classical basis. We use the upper case latin letters E and F to denote
L0-modules, and the boldface upper case latin letters E and F to denote conditional
vector spaces.
An L0-module E is an additive group on which the commutative ring L0 acts by
scalar multiplication. In particular, this implies that the associated measure algebra
acts on E by scalar multiplication with indicator functions which makes the following
definition plausible.
Definition 3.1. An L0-module E is said to be stable, if for all (xk) ⊂ E and (ak) ∈ p(1)
there exists a unique element x ∈ E such that 1Ak x = 1Ak xk for all k. x is called the
concatenation of (xk) along (Ak), and denoted formally by x =Pk 1Ak xk.
In [14, Example 2.12] the existence of concatenations is not satisfied, and in [42,
Example 1.1] their uniqueness fails. Thus an L0-module is not a priori stable. However,
we can always stabilize an L0-module as follows.
6
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
Remark 3.2. Let E be an L0-module. Let Es be the collection of all (xk, ak) ⊂ E × A
where (ak) ∈ p(1). Identify (xk, ak) and (yk, bk) whenever 1Ak∩Bhxk = 1Ak∩Bhyh for all
h, k. Defining (xk, ak) + (yk, bk) := (xk + yh, ak ∧ bh) and r(xk, ak) = (rxk, ak) provides
Es with the structure of a stable L0-module in which E can be included via x 7→ (x, 1).
Definition 3.3. A triplet (E, +, ·) consisting of a conditional set and two conditional
functions is said to be a conditional vector space, if the carrier structure (E, +, ·) is an
L0-module.4
Now every stable L0-module E induces also a conditional vector space. Indeed, define
on E × A the equivalence relation (x, a) ∼ (y, b) whenever a = b and 1Ax = 1By, and
let xa denote the equivalence class of (x, a). Then the collection E of all equivalence
classes xa is a conditional set with carrier E. Since addition and scalar multiplication
in E are stable functions, E inherits the structure of a conditional vector space.
Recall that a function f : E → F is a module homomorphism, if it is L0-linear.
Observe that every module homomorphism satisfies 1Af (1Ax) = 1Af (x) for all a ∈ A
and x ∈ E. It follows that every module homomorphism is a stable function whenever
E and F are stable modules. A conditional function f : E → F is conditional linear, if
the stable function f : E → F is a module homomorphism.
Let CVectR denote the category whose objects are conditional vector spaces and
whose morphisms are conditional linear functions. Let Mods
L0 be the category whose
objects are stable L0-modules and whose morphisms are module homomorphisms. The
next equivalence readily follows from the previous considerations.
Theorem 3.4. The functor mapping a conditional vector space E to the L0-module E
and a conditional linear function f to the module homomorphism f is an equivalence
of categories between CVectR and Mods
L0 .
Since L0 is a commutative ring, every free L0-module has a rank, see [27, Theorem
2.1].
Proposition 3.5. Suppose there exists an infinite partition in p(1) and let E be a free
L0-module. Then E is stable if and only if E has finite rank.
Proof. If E has finite rank, then E is finitely generated, say by {v1, . . . , vn}. Let
m ∈ L0,
(ak) ∈ p(1) and (xk) ⊂ E . Each xk is of the form rk
m for m = 1, . . . , n and x := r1v1 + . . . + rnvn.
Then 1Ak x = 1Ak xk for all k. Now, let y ∈ E be satisfying 1Ak y = 1Ak xk for all k.
If we prove that x = y, then we conclude that E is stable. Indeed, y is of the form
s1v1 + . . . + snvn for some sm ∈ L0, m = 1, . . . , n. Then, for each k, 0 = 1Ak (x − y) =
1Ak (r1 − s1)v1 + . . . + 1Ak (rn − sn)vn. This means that 1Ak (rm − sm) = 0 for all k, m,
and thus rm = sm for each m. We conclude that x = y.
m = 1, . . . , n. Put rm := Pk 1Ak rk
nvn for some rk
1 v1 + . . . + rk
Conversely, suppose that E is stable. By contradiction, assume that E has not finite
rank. Since E is free, E has a basis (xi)i∈I of determined length. Let (ak) ∈ p(1) be
an infinite partition, (ik) be an injective infinite subfamily of I and x := Pk 1Ak xik .
Then there exists a finite collection J ⊂ I and (ri)i∈J ⊂ L0 such that x = Pi∈J rixi.
Since J is finite, there exists ik ∈ I \ J for some k. One has
0 = 1Ak x − 1Ak x =Xi∈J
ri1Ak xi − 1Ak xik .
4Suppose that E is a conditional vector space. For (xk) ⊂ E and (ak) ∈ p(1) let x := P xkak. Then
the stability of the scalar product allows to prove that 1Ak x = 1Ak xk for each k. This means that E is
stable in the modular sense and P 1Ak xk = P xkak, that is, the abstract conditioning ·a is consistent
with the modular conditioning 1A·.
STABLE COMPACTNESS
Then necessarily 1Ak = 0, but this contradicts P(Ak) > 0.
This motivates the following definition.
7
(cid:3)
Definition 3.6. Let E be a conditional vector space, (xm)16m6n a stable finite family
in E and (sm)16m6n a stable finite family in L0. A stable linear combination is defined
as5
nk
X16m6n
rmxm :=X(
Xm=1
rmxm)ak,
where n =Pk 1Ak nk. The stable linear span st-span(S) of a non-empty subset S ⊂ E
is the stable hull of all stable linear combinations of the elements in S.
A stable set S ⊂ E is said to be a stable basis, if the following conditions are satisfied:
stable linear independence: If (xm)16m6n is a stable finite subset of S and
(rm)16m6n is a stable finite family in L0, then
(i)
rmxm = 0
implies
rm = 0
for all
1 6 m 6 n.
X16m6n
(ii) stable generating: for every x ∈ E there exist a stable finite family (xm)16m6n
in S and a stable finite family (rm)16m6n in L0 such that
x = X16m6n
rmxm.
Theorem 3.7. Every conditional vector space has a stable basis.
Proof. Let E be a conditional vector space. Let H be the collection of all stable
linearly independent sets in E. Then H is non-empty since it contains every one-
element family. Order H by inclusion. Let (Si) be a chain in H . Then st(∪Si) is a
member of H . By Zorn's lemma, H has a maximal element which we denote by S∗.
By contradiction, suppose st-span(S∗) 6= E. In other words,
0 < a∗ := ∨{a ∈ A : there is x ∈ E such that 1Bx 6∈ 1Bst-span(S∗) for all 0 < b 6 a}.
By stability of S∗ and E, a∗ is attained by some x∗ ∈ E due to an exhaustion argument.
Assume w.l.o.g. that a∗ = 1. Then S∗∗ := st-span(S∗ ∪ {x∗}) ∈ H and S∗ ( S∗∗ which
contradicts the maximality of S∗.
(cid:3)
One can adapt Definition 3.6 to provide a definition of a stable basis for stable
L0-modules. It then follows readily from Proposition 3.4:
Corollary 3.8. Every stable L0-module has a stable basis.
Let n ∈ L0
s(N) and (L0)n denote the set of all the stable functions {1 6
⋄
Example.
m 6 n} → L0. Then {1{m=n} : 1 6 m 6 n} is a stable basis of (L0)n.
Definition 3.9. An L0-module E is said to be finitely generated if it has a stable finite
subset S which stably generates E.
The following result is an extension of [7, Theorem 2.8].
Corollary 3.10. Let E be a stable finitely generated L0-module. Then E admits a
stable finite basis. Moreover, there exist unique (ak) ∈ p(1) and (nk) ⊂ N such that
1Ak E is a free 1Ak L0-module of rank nk.
5In the definition of a stable finite sum, we use the fact that every stable finite family can be written
as the concatenation of the stable hulls of finite subfamilies along (ak).
8
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
In the setting of the previous corollary, we call n =P 1Ak nk the stable dimension of
E. As in the classical case, we have the following characterization.
Corollary 3.11. If E is stable finitely generated, then there exist n ∈ L0
module isomorphism E → (L0)n.
s(N) and a
One can apply Proposition 3.4 to derive the Hahn-Banach extension theorem6 for
stable L0-modules, see e.g. [14, Theorem 2.14], from the conditional Hahn-Banach
extension theorem [10, Theorem 5.3]. Recall that a function f : E → L0 is said to be
L0-convex, if f (rx + (1 − r)y) 6 rf (x) + (1 − r)f (y) for all r ∈ L0 with 0 6 r 6 1 and
every x, y ∈ E.
Theorem 3.12. Let E be a stable L0-module. Let p : E → L0 be L0-convex, F ⊂ E a
sub-module and f : F → L0 L0-linear and such that f (x) 6 p(x) for all x ∈ F . Then
there exists an L0-linear function f : E → L0 such that f F = f and f (x) 6 p(x) for
all x ∈ E.
4. Topological L0-modules and conditional topological vector spaces
We know of three types of module topologies on L0-module. The first kind of topol-
ogy is the (ǫ, λ)-topolog which is introduced in [26] in the case of a single L0-norm
and for families of L0-seminorms, see [21]. The second type of topology are locally
L0-convex topologies introduced in [14]. The last type of topology which we shall call
stable topologies are intimately connected to conditional linear topologies introduced
in [10]. In this section, we focus on the relations among these three kind of topolo-
gies and their connection with conditional topological vector spaces which will be the
foundation for the results in the following sections. Unless we mention otherwise, an
L0-module is stable.
Definition 4.1. Let E be an L0-module. A function p : E → L0
L0-seminorm, if
+ is said to be an
(i) p(rx) = rp(x) for all r ∈ L0, x ∈ E,
(ii) p(x + y) 6 p(x) + p(y) for all x, y ∈ E.
In addition, if p(x) = 0 implies x = 0, then p is said to be an L0-norm.
Let E be a conditional vector space. A conditional function p : E → R+ is said to
+ is an L0-seminorm.7 A
be a conditional seminorm, if the carrier function p : E → L0
conditional seminorm is a conditional norm, if the carrier function is an L0-norm.
We recall the definition of the (ǫ, λ)-topology induced by a family of L0-seminorms.
Definition 4.2. Let E be an L0-module and P a non-empty collection of L0-seminorms.
For x ∈ E, let Uǫ,λ(x) denote the set of all
Ux,N,ǫ,λ := {y ∈ E : P(sup
p∈N
p(x − y) < ǫ) > 1 − λ},
where ǫ, λ ∈ R+ with λ < 1 and N ⊂ P is finite. The topology on E generated by the
base ∪x∈X Uǫ,λ(x) is denoted by Tǫ,λ = Tǫ,λ(P).
6The Hahn-Banach extension theorem can be proved for modules without the stability assumption, see
the discussion and the references in [14].
7Notice that (i) implies that p is stable.
STABLE COMPACTNESS
9
Notice that a neighbourhood Ux,N,ǫ,λ is not necessarily a stable set, nor is Uǫ,λ(x) a
stable collection. The absolute value · : L0 → L0
+ induces the topology of convergence
in probability which is a linear as well as a module topology on L0. More generally,
given an L0-module E and any non-empty collection P of L0-seminorms on E, the set
Uǫ,λ(0) is a neighbourhood base of 0 for a module topology where we assume that L0 is
endowed with the topology of convergence in probability. Notice that an (ǫ, λ)-topology
is always a linear topology as well, for more details see [19, Proposition 2.6].
Let (pk) be a countable family of L0-seminorms and (ak) ∈ p(1). Then x 7→
Pk 1Ak pk(x) is an L0-seminorm. A collection P of L0-seminorms is stable, if P 6= ∅
and Pk 1Ak pk ∈ P for all (pk) ⊂ P and every (ak) ∈ p(1). For a non-empty col-
lection P of L0-seminorms, we denote by st(P) its stable hull and by st-sup(P) the
stable hull of the collection of all supp∈N p, where N ⊂ P is finite. Next, we recall the
definition of a locally L0-convex topology induced by a family of L0-seminorms.
Definition 4.3. Let E be an L0-module and P a non-empty collection of L0-seminorms.
For x ∈ E, let U0(x) be the collection of all
Ux,N,r := {y ∈ E : sup
p∈N
p(x − y) < r},
where r ∈ L0
∪x∈X U0(x) is denoted by T0 = T0(P).
++ and N ⊂ P is finite. The topology on E generated by the base
Notice that a neighbourhood Ux,N,r is a stable set, however U0(x) is not necessarily
a stable collection. The condition P = st-sup(P) guarantees the stability of U0(x),
see [14, Lemma 2.18] and the discussion following it. The locally L0-convex topology
induced by the absolute value · : L0 → L0
+ is the interval topology w.r.t. the order of
a.e. dominance.
Definition 4.4. Let E be an L0-module and P a stable collection of L0-seminorms.
For x ∈ E, let Us(x) denote the set of all
Ux,N ,r := {y ∈ E : sup
p∈N
p(x − y) < r},
where r ∈ L0
∪x∈E Us(x) is denoted by Ts = Ts(P).
++ and N ⊂ P is stable finite. The topology generated by the base
By construction, Us(x) is a stable collection of stable sets. Moreover it holds
Us(Xk
1Ak xk) =Xk
1Ak
Us(xk)
for all (xk) ⊂ E and (ak) ∈ p(1). Since B := ∪x∈E Us(x) is a stable collection of stable
sets and a topological base, it follows from [10, Proposition 3.5] that the conditional
collection B induced by B is the conditional topological base of a conditional topology
on E. The conditional topology induced by the absolute value on L0 is isomorphic
to the conditional Euclidean topology on R, see [10, Theorem 4.4] and the dicussion
preceding it.
Remark 4.5. All three types of topologies are Hausdorff whenever supp∈P p(x) = 0 if
and only if x = 0. In this case we say that P is separated. We will assume the latter
property throughout this article.
Next, we focus on the connection to conditional locally convex topological vector
spaces.
10
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
Definition 4.6. Let E be an L0-module. A subset S ⊂ E is said to be L0-convex, if
rx + (1 − r)y ∈ S for all r ∈ L0 with 0 6 r 6 1 and every x, y ∈ S.
Let E be conditional vector space. A conditional subset S of E is said to be condi-
tional convex, if its carrier is L0-convex.
Definition 4.7. A topological L0-module (E, T ) is said to be locally L0-convex, if there
exists a neighbourhood base U of 0 ∈ E consisting of stable L0-convex sets in E. In this
case, T is a locally L0-convex topology. A conditional topological vector space (E, T )
is said to be conditional locally convex, if there exists a conditional neighbourhood base
of 0 in E consisting of conditional convex sets.
It is proved in [14, 44] that a topological L0-module (E, T ) is locally L0-convex if
and only if T = T0(P) for some collection P of L0-seminorms. The Hahn-Banach
separation theorem [14] is proved under the additional assumption that the collection of
L0-seminorms satisfies P = st-sup(P). This amounts to the identity U0(0) = Us(0),
which makes U0(0) a stable collection of stable subsets. In addition, the existence of
a stable neighbourhood base of 0 ∈ E consisting of stable subsets is key to establish
a categorical equivalence between locally L0-convex modules and conditional locally
convex topological vector spaces. This leads us to the following definition:
Definition 4.8. A locally L0-convex module (E, T ) is called stable if there exists a
neighbourhood base U of 0 ∈ E, which is a stable collection of stable L0-convex sets in
E. In this case, T is a stable topology.
Suppose that T = Ts(P) for some collection P of L0-seminorms on E, then it is
clear that T is a stable locally L0-convex topology. Conversely, by an adaptation of a
classical argument, it can be proven that for any stable locally L0-convex topology T
there is a stable family of L0-seminorms P such that T = Ts(P), see [39].
Let CLCS denote the category whose objects are conditional locally convex topo-
logical vector spaces and whose morphisms are conditional continuous linear functions.
Let LCSs
L0 denote the category whose objects are stable locally L0-convex modules
and whose morphisms are continuous L0-linear functions. By [10, Proposition 3.5] and
[10, Proposition 3.11], we obtain the following equivalence of categories established in
[39, Theorem 1.2].8
Theorem 4.9. The functor mapping
• a conditionally locally convex topological vector space (E, T ) to the stable locally
L0-convex module (E, T ), where E is the carrier of E and T is the topology
generated by the set of carriers of conditional open sets.
• and a conditional linear continuous function f : E → F to the L0-linear con-
tinuous function f : E → F ,
is an equivalence of categories between CLCS and LCSs
L0 .
Next, we will compare the different types of classical and conditional topologies. We
need two preliminary results.
Proposition 4.10. Let P be a non-empty collection of L0-seminorms on an L0-module
E. Then
Tǫ,λ(P) = Tǫ,λ(st(P)) = Tǫ,λ(st-sup(P)).
8The statement in [39] contains a mistake. There T is identified with the set of carriers of conditional
open sets. As stated here, T must be the topology generated by the latter set, which is a topological
base for T . The use of [10, Proposition 3.11] is not mentioned in the proof of [39, Theorem 1.2].
STABLE COMPACTNESS
11
Proof. Since P ⊂ st(P) ⊂ st-sup(P), it suffices to show that Tǫ,λ(P) = Tǫ,λ(st-sup(P)).
Since P ⊂ st-sup(P), then Tǫ,λ(P) ⊂ Tǫ,λ(st-sup(P)).
Conversely, let (ak) ∈ p(1), (Sk) be a countable family of finite subcollections of P
such that P(∪k>mAk) < λ/2. Then we have
and ǫ, λ ∈ R+ with λ < 1. Put q = Pk 1Ak supp∈Sk p. Choose m ∈ N large enough
U0,{q},ǫ,λ = {x ∈ E : P([k>1
[k=1
(x) < ǫ} ∩ Ak) > 1 − λ}
(x) < ǫ} ∩ Ak) > 1 − λ}
⊃ {x ∈ E : P(
{ sup
p∈Sk
{ sup
p∈Sk
m
p(x) < ǫ} ∩ ∪m
k=1Ak) > 1 − λ}
⊃ {x ∈ E : P({
sup
k=1Sk
p∈∪m
⊃ {x ∈ E : P(
⊃ {x ∈ E : P(
p∈∪m
sup
k=1Sk
sup
k=1Sk
p∈∪m
p(x) < ǫ) > 1 − λ + P({
sup
k=1Sk
p(x) < ǫ) > 1 − λ + P(∪k>mAk)}
p∈∪m
p(x) < ǫ} ∩ ∪k>mAk)}
p(x) < ǫ) > 1 −
λ
2
},
= {x ∈ E : P(
sup
k=1Sk
p∈∪m
where the last set is a neighbourhood w.r.t. P.
(cid:3)
Proposition 4.11. Let P1, P2 be collections of L0-seminorms on an L0-module E.
If T0(P1) = T0(P2), then Tǫ,λ(P1) = Tǫ,λ(P2).
Proof. For ǫ, λ ∈ R+ with λ < 1 and finite N ⊂ P1, choose finite N ′ ⊂ P2 and
s ∈ L0
++ such that
{x ∈ E : sup
p∈N ′
p(x) < s} ⊂ {x ∈ E : sup
p∈N
p(x) < ǫ}.
Let (ak) ∈ p(1) and (ǫk) ⊂ R with 0 < Pk 1Ak ǫk 6 s. Choose m ∈ N large enough so
that P(∪k>mAk) < λ/2. Then one has
{x ∈ E : P(sup
p∈N
p(x) < ǫ) > 1 − λ}
⊃ {x ∈ E : P(∪m
⊃ {x ∈ E : P(∪k∈N{ sup
p∈N ′
k=1{ sup
p∈N ′
p(x) < min
16k6m
⊃ {x ∈ E : P({ sup
p∈N ′
p(x) < ǫk} ∩ Ak) > 1 − λ}
p(x) < ǫk} ∩ Ak) > 1 − λ}
ǫk} ∩ ∪m
k=1Ak) > 1 − λ}
⊃ {x ∈ E : P( sup
p∈N ′
p(x) < min
16k6m
ǫk) > 1 − λ + P(∪k>mAk)}
⊃ {x ∈ E : P( sup
p∈N ′
p(x) < min
16k6m
ǫk) > 1 −
λ
2
}.
This proves that Tǫ,λ(P1) ⊂ Tǫ,λ(P2). By interchanging the roles of P1 and P2, it
follows the claim.
(cid:3)
In the following remark, we compare the three topologies.
Remark 4.12. Given a locally L0-convex module (E, T ), there exists a collection
P of L0-seminorms such that T = T0(P), see [14, 44]. Proposition 4.11 tells us
that the topology Tǫ,λ does not depend on P. One can verify that Ts = Ts(st(P)) =
12
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
T0(st-sup(P)) is the coarsest stable topology such that T ⊂ Ts. Therefore, Ts does not
depend on P. Proposition 4.10 shows that Ts does not produce a new (ǫ, λ)-topology.
Consequently, T uniquely defines Tǫ,λ and Ts independent of the choice of the inducing
family of L0-seminorms.
For S ⊂ E stable, we have clǫ,λ(S) = cl0(S) = cls(S) due to [19, Theorem 3.12].
Therefore, by [10, Proposition 3.5] and [10, Proposition 3.7], the classical closure cl0(S)
is the carrier of the conditional closure of S w.r.t. the conditional topology T .
ǫ,λ = E∗
ǫ,λ = E∗
0 and E∗
Denote by E∗
s . By [22, Corollary 3.6], if T if stable, then E∗
ǫ,λ the space of L0-linear and Tǫ,λ-continuous functions f : E → L0.
Similarly, define the dual modules E∗
s . By [19, Proposition 2.14], we have
E∗
0 , and it follows from
ǫ,λ is the carrier of the conditional dual space E∗ w.r.t. the
[10, Proposition 3.11] that E∗
corresponding conditional topology. In this case, since we have the same dual module
for the three topologies, we will denote them by E∗ = E∗[T ]. If T is not stable, then
E∗
0 is not necessarily stable as the next example shows. However, by [22, Theorem 3.6]
ǫ,λ, where we denote by Es the stabilization of E
and stabilization, we have (E∗
as defined in Section 3.
0 )s = E∗
Example. Let E = (L0)N be the set of all sequences in L0. We consider the family of
L0-seminorms P := {pn : n ∈ N}, where pk((xn)) := xk, (xn) ∈ E. We claim that
E∗
0 = {(xn) ∈ E : ∃m ∈ N such that xn = 0 for all n > m},
ǫ,λ = E∗
E∗
s := {(xn) ∈ E : ∃m ∈ L0
s(N) such that xn = 0 for all n > m},
where xn := Pk 1Ak xnk whenever n = Pk 1Ak nk, (ak) ∈ p(1) and (nk) ⊂ N. Notice
that the latter space is the stable hull of the former. Observe also that these spaces
are different whenever there exists an infinite (ak) ∈ p(1). By Remark 4.12, it suffices
to prove the first equality. Any x := (xn) with xn = 0 for all n > m defines via
fx((yn)) :=Xn>1
xnyn
a T0-continuous L0-linear function.
Conversely, suppose that f ∈ E∗
0 . By continuity, there exists s ∈ L0
++ and m ∈ N
such that f (x) < 1 whenever x = (xn) satisfies xk < s for all k = 1, . . . , m. Let
ek ∈ E denote the sequence such that its kth-coordinate is one and zero otherwise.
Then f (nek) = nf (ek) 6 1 for every k > m and n ∈ N. This implies that f (ek) = 0
whenever k > m. Now define x = (xn) by xk := f (ek) for k 6 m and xk := 0 otherwise.
⋄
Then we have fx = f .
Finally, we collect some examples of stable locally L0-convex modules.
Examples.
(i) Let (X, k · k) be a normed vector space, L0(X) the space of all strongly
measurable functions modulo a.e. equality, and define k · k : L0(X) → L0 by
kxk(ω) := kx(ω)k a.e. ω ∈ Ω. Then L0(X) is a stable L0-module and k · k
is an L0-norm, see [26, Section 5.1] and [11, Section 4]. The space L0(X)
can be constructed by a conditional completion, and thus is the carrier of a
conditional Banach space, see [11].
(ii) We call (E, F, h·, ·i) a stable dual pair,
if E, F are stable L0-modules and
h·, ·i : E × F → L0 is an L0-bilinear form such that hx, yi = 0 for all y ∈ F
implies x = 0, and hx, yi = 0 for all x ∈ E implies y = 0, see [19] as well. From
L0-linearity follows stability, and therefore the associated conditional structure
STABLE COMPACTNESS
13
is a conditional dual pair in the sense of [10, Definition 5.6]. It can be checked
that {h·, yi : y ∈ F } is a stable collection of L0-seminorms.
(iii) Let (X, Y, h·, ·i) be a dual pair of Banach spaces such that
(a) hx, yi 6 kxkkyk for all x ∈ X and y ∈ Y ,
(b) both norm-closed unit balls are weakly closed.
Then h·, ·i extends to an L0-duality pairing on L0(X) × L0(Y ), see [12, Lemma
3.1].
(iv) Let (E, T ) be a locally L0-convex module. Consider the canonical duality
pairing between E and E∗
s ) and σs(E, E∗
s )
the weak (ǫ, λ)-topology, the weak L0-topology and the weak stable topology
on E induced by the stable collection of L0-seminorms {h·, x∗i : x∗ ∈ E∗
s },
respectively. Similarly, we can introduce the three different types of weak∗-
topologies σǫ,λ(E∗
s , and denote by σǫ,λ(E, E∗
s ), σ0(E, E∗
s , E) and σs(E∗
(v) Let E ⊂ F be a sub-σ-algebra and 1 6 p 6 ∞. Let ¯L0(E) denote the space of
equivalence classes of R ∪ {+∞}-valued random variables. Let k · kp : L0(F) →
¯L0(E) be defined by
s , E), σ0(E∗
s , E).
kxkp :=(limn→∞ E[xp ∧ nE]1/p,
inf{y ∈ ¯L0(E) : y > x},
if p < ∞,
if p = ∞.
E (F) := {x ∈ L0(F) : kxkp ∈ L0(E)}. Then (Lp
Let Lp
L0(E)-normed module for which one has the representation Lp
For Holder conjugates (p, q) where 1 6 p < ∞, the function h·, ·i : Lp
Lq
E (F) → L0(E) defined by (x, y) 7→ E[xyE] is an L0(E)-duality pairing.9
E (F), k · kp) is a stable
E (F) = L0(E)Lp(F).
E (F) ×
5. Classical and conditional compactness
In this section, we study to which extent basic compactness results in functional
analysis can be established in topological L0-modules. The classical notion of compact-
ness does not bear much fruits. Indeed, it is proved in [18] that the Banach-Alaoglu
and the Banach-Bourbaki-Kakutani-Smulian theorem for the (ǫ, λ)-topology are ful-
filled if and only if the support of E is essentially atomic. Since the corresponding L0-
and stable topology are finer than the (ǫ, λ)-topology, there is also no hope to prove
these theorems in full generality for stable and L0-type topological modules. We will
complement the previous finding by verifying that a stable locally L0-convex module
is anti-compact if and only if its support is atomless. This motivates the concept of
stable compactness for stable locally L0-convex modules which we will derive from con-
ditional compactness. In the remainder of this article, we then proceed to show that all
basic compactness results in functional analysis can be established in full generality for
stable locally L0-convex modules using stable compactness. In the finite dimensional
case, we provide characterizations of stable compactness in terms of a.e. convergence
and compact set-valued mappings, respectively, which are relevant in applications.
The support of an L0-module E is defined as
supp(E) := ∧ {a ∈ A : 1Ac E = {0}} .
Lemma 5.1. Let (rn) ⊂ L0
is atomless. Then there exists r ∈ L0
+ be with rn 6= 0 for all n ∈ N. Suppose that {supn∈N rn > 0}
++ such that P(rn > r) > 0 for all n ∈ N.
9For p = 2, the L2-type module Lp
representation result.
E (F) was introduced in [25], see [33] for other values of p and the
14
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
Proof. For each n ∈ N, let An := {rn > 0}. Since rn 6= 0, one has that P(An) > 0 for
all n ∈ N. Let B1 := A1. Since A2 ⊂ {supn rn > 0} and the latter is atomless, we can
choose B2 ⊂ A2 with 0 < P(B2) < 1
2 P(B1). By induction, for each n > 1, we can find
Bn+1 ⊂ An+1 with 0 < P(Bn+1) < 1
2 P(Bn). For fixed n ∈ N, it thus holds
1
2k−n P(Bn)
for all k > n.
P(Bk) <
Now, for each n ∈ N, put Cn := Bn − Sk>n
Bk. Then,
P(Cn) > P(Bn) −Xk>n
P(Bk) > P(Bn) −
∞
Xk>n
Note that (cn) ∈ p(∨cn). Define r := Pn∈N 1Cnrn/2 + 1∩C c
conclude
P(rn > r) > P(Cn) > 0
for all n ∈ N.
1
2k−n P(Bn) = 0.
n. Then r ∈ L0
++, and we
(cid:3)
Proposition 5.2. Let (E, T ) be a stable locally L0-convex module with supp(E) > 0.
Then E is anti-compact if and only if supp(E) is atomless.
Proof. Let P be a stable collection of L0-seminorms which induce Ts. For the sake
of contradiction, suppose that K ⊂ E is an infinite compact set. Then there exists an
injective sequence (xn) in K which has a cluster point x0 ∈ K. We assume w.l.o.g. that
x0 6= xn for all n ∈ N. Since P is separated, for every n ∈ N there is pn ∈ P such that
pn(x0 − xn) 6= 0.
Now, define ρ : E × E → L0 by
ρ(x, y) := Xn∈N
1
2n
pn(x − y)
1 + pn(x − y)
.
Then ρ(xn, x0) 6= 0 for all n ∈ N and {supn ρ(xn, x0) > 0} ⊂ supp(E). By Lemma 5.1,
there is r ∈ L0
++ such that P(ρ(xn, x0) > r) > 0 for all n ∈ N. Let
Vr(x0) := {y ∈ E : ρ(x0, y) < r} ,
and notice that xn /∈ Vr(x0) for all n ∈ N. If we would show that Vr(x0) is a neighbour-
hood of x0, then we obtain the desired contradiction. First, observe that there exist
(ak) ∈ p(1) and (nk) ⊂ N such that
Xk∈N
1
2n
pn(x0 − y)
1 + pn(x0 − y)
<
r
2
for all y ∈ E.
1Ak Xn>nk
++ be small enough such that if 1Ak pn(x0 − y) 6 1Ak s for n < nk,
Second, let s ∈ L0
then
Xk∈N
1Ak X16n6nk
1
2n
pn(x0 − y)
1 + pn(x0 − y)
<
r
2
.
Put n :=Pk 1Ak nk and N := {pm : 1 6 m 6 n} ⊂ P. Then Ux0,N ,s ⊂ Vr(x0).
As for the converse, suppose that there exists an atom A ⊂ supp(E). Notice that
from stability of an L0-seminorm p it follows that 1Bp(1Bx) = p(1Bx) for all x ∈ E
and B ∈ F. Thus, since A is an atom, for all p ∈ P and x ∈ E there exists ηx,p ∈ R+
such that p(1Ax) = 1Aηx,p. Let p : E → R be defined by p(x) := E[p(1Ax)] for each
p ∈ P. Since P is separated, there exists x0 ∈ E and p0 ∈ P such that p0(x0) > 0.
STABLE COMPACTNESS
15
Then P := {p : p ∈ P} is a set of real-valued seminorms which induces on the 1-
dimensional vector space E0 := {1Arx0 : r ∈ L0} a locally convex Hausdorff topology.
By the Heine-Borel theorem, the unit ball K of E0 is infinite compact. Since P and
P induce the same topology on E0, K is also compact in E.
(cid:3)
Remark 5.3. Next, we introduce the notions of stable topology, stable filter, stable
compactness and stable metric on a stable subset of an L0-module. These definitions
make also sense on stable sets of functions as defined in [30], or more generally on the
carrier of a conditional set. A stable structure can be viewed as an interpretation of a
corresponding conditional structure in classical set theory. The definitions of a condi-
tional filter, conditional ultrafilter, conditional compactness and conditional metric are
introduced in [10], where the relation between classical and conditional topology is stud-
ied systematically, in which the following definitions of stable structures are implicitly
applied.
Definition 5.4. A topology on a stable subset S of an L0-module is said to be stable
if it admits a topological base which is a stable collection of stable sets.10
Definition 5.5. A filter on a stable subset S of an L0-module is said to be stable if it
admits a filter base which is a stable collection of stable sets. A stable filter is called a
stable ultrafilter if it is a maximal element in the set of all stable filters on S.
Definition 5.6. Let T be a stable topology on S. Then S is said to be stable compact
if every stable filter F on S has a cluster point in S.
Similarly to classical compactness, we can prove the following characterizations of
stable compactness.
Proposition 5.7. Let T be a stable topology on S. Then the following are equivalent.
(i) S is stable compact.
(ii) Every stable ultrafilter on S has a cluster point.
(iii) For every stable collection O of stable and open sets with S = ∪O there exists
a stable finite subcollection O ⊂ O such that S = ∪ O.
(iv) Every stable collection C of stable and closed sets in S with ∩ C 6= ∅ for every
stable finite subcollection C ⊂ C satisfies ∩C 6= ∅.
Proof. The claim can be established similarly to [10, Proposition 3.25] by applying [10,
Proposition 3.25] and a straightforward adaptation of [10, Theorem 3.16] to the present
context.
(cid:3)
To be able to prove a Tychonoff's theorem, one needs a new notion of product topol-
ogy since the product of a family of stable topologies is not necessarily a stable topology,
see [39, Example 1.1]. This motivates the following construction.
many U k
Example. Let (Si, Ti) be a non-empty family of stable topological spaces with base
i = Si for all but finitely
i ∈ Bi and (ak) ∈ p(1), is a stable collection of stable sets and the base
If B is the
Bi for each i. The collection B of all Pk 1AkQ U k
for a stable topology on Q Si, referred to as the stable product topology.
conditional collection of conditional subsets of Qi Si induced by stable sets in B, then
the conditional topology generated by B is the conditional product topology on Qi Si,
i , where U k
for a definition see [10, Examples 3.9(2)].
⋄
10It can straightforwardly be checked that if T is a stable topology, then T is a stable collection of
not necessarily stable sets. Notice that, in general, the union of stable sets in not stable.
16
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
The conditional Tychonoff's theorem [10, Theorem 3.28] yields the following
Theorem 5.8. Let (Si, Ti)i∈I be a non-empty family of stable topologies. Then the
Cartesian product Qi Si is stable compact w.r.t. the stable product topology if and only
if Si is stable compact for all i ∈ I.
Next, we study compactness in stable metric spaces.
Definition 5.9. Let E be an L0-module and S ⊂ E stable. A stable metric is a stable
function d : S × S → L0
+ such that for all x, y, z ∈ S it holds
(i) d(x, y) = 0 if and only if x = y,
(ii) d(x, y) = d(y, x),
(iii) d(x, z) 6 d(x, y) + d(y, z).
The collection of all Br(x) := {y ∈ S : d(x, y) < r}, where r ∈ L0
++ and x ∈ S, forms
a stable collection of stable sets and the base for a stable topology on S. We introduce
the following definitions.
• A stable sequence (xn) in stable metric space is said to be Cauchy if for every
r ∈ L0
++ there exists n0 ∈ L0
s(N) such that d(xn, xm) 6 r for all n, m > n0.
• A stable metric space is said to be
-- stable complete if every Cauchy stable sequence is convergent;11
-- stable sequentially compact if for every stable sequence there exists a con-
vergent stable subsequence;12
-- bounded if sup{d(x0, x) : x ∈ S} < ∞ for some x0 ∈ S;
-- stable totally bounded if for every r ∈ L0
++ there exists a stable finite
subset S ⊂ S such that S = ∪x∈ SBr(x);
-- stable separable if there exists a stable sequence (xn) in S such that for
every x ∈ S there exists a stable subsequence (xm) of (xn) such that
x = lim xm.
We have the Heine-Borel theorem for stable metric spaces whose proof can be carried
out similarly to the proof of [10, Theorem 4.6].
Theorem 5.10. Let S be a stable metric space. Then the following are equivalent.
(i) S is stable compact.
(ii) If (Rn) is a decreasing stable family of stable and closed sets in S, then ∩Rn 6=
∅.
(iii) S is stable sequentially compact.
(iv) S is stable totally bounded and stable complete.
If the underlying L0-module E is stable finitely generated, then we have:
Corollary 5.11. Let (E, k · k) be a stable finitely generated L0-normed module. Then
the stable unit ball {x ∈ E : kxk 6 1} is stable compact.
endowed with the Euclidean L0-norm kxk = (Pd
An example of a finitely generated L0-normed module is (L0)d, d ∈ N, which is
i )1/2. A subset S ⊂ (L0)d is said
to be sequentially closed if x ∈ S whenever (xn) ⊂ S and limn xn = x a.e. In this case,
we have the following characterizations of stable compactness.
Theorem 5.12. Let K be a subset of (L0)d. Then the following are equivalent.
i=1 x2
11Here we understand convergence as convergence of (xn) as a net.
12A stable subsequence (yn) is said to be a stable subsequence of (xn) if there exists an increasing
stable sequence (nk) with yk = xnk for all k ∈ L0
s(N).
STABLE COMPACTNESS
17
(i) K is stable compact.
(ii) K is stable, closed and bounded.
(iii) K is stable, sequentially closed and bounded.13
(iv) K is the set of measurable selectors of a non-empty measurable compact-valued
mapping SK : Ω ⇒ Rd.14
In this case, the mapping SK : Ω ⇒ Rd is unique modulo a.e. equality.
Proof. (i) ⇔ (ii): Theorem 5.11.
(ii) ⇔ (iii): Follows from the fact that a stable set K ⊂ (L0)d is sequentially closed
if and only if it is closed in the stable topology.
(iv) ⇒ (i): Let S : Ω ⇒ Rd be a measurable compact-valued mapping with dom(S) =
Ω. Then K = {x ∈ (L0)d : x(ω) ∈ S(ω) for all ω ∈ Ω} is non-empty due to the
Kuratowski-Ryll-Nardzewski theorem (see e.g. [1, Theorem 18.13]). By inspection, K
is stable. Since the Euclidean norm k · k : Rd → R is continuous, by the Measurable
Maximum Theorem (see e.g. [1, Theorem 18.19]), the function
r : Ω → R,
r(ω) := max
x∈S(ω)
kxk
is an element of L0. This shows that K ⊂ Br(0) which means boundedness. To show
that K is closed, let x ∈ cl(K). There is xn ∈ K with kx − xnk 6 1
n for all n ∈ N.
Let xn be a representative of xn each n ∈ N. Then x has a representative x such that
kx(ω) − xn(ω)k 6 1
n for all ω ∈ Ω. Therefore, x(ω) ∈ cl(S(ω)) = S(ω) for all ω ∈ Ω
which implies x ∈ K.
(i) ⇒ (iv): Suppose that K ⊂ (L0)d is stable compact. Enumerate Qd := {q1, q2, . . .}.
Notice that Qd can be considered as a subset of L0
s(Qd) via q 7→ 1Ωq. For each
n ∈ N, let rn := inf{kqn − xk : x ∈ K}. For each m ∈ N, since K is stable, we
find x(n, m) ∈ K with kqn − x(n, m)k 6 rn + 1
s(N) of the form
3m . For n, m ∈ L0
n = Pk 1Ak nk and m = Ph 1Bhmh, define rn := Pk 1Ak rnk , qn := Pk 1Ak qnk and
x(n, m) :=Pk,h 1Ak∩Bhxnk,mh. Then,
kqn − x(n, m)k 6 rn +
for all n, m ∈ L0
s(N).
1
3m
(5.1)
Put H := {x(n, m) : n, m ∈ L0
in K. Indeed, for any x ∈ K and m ∈ L0
qn ∈ L0
The previous two inequalities together with (5.1) yield
s(N)}. By stability, H ⊂ K. We claim that H is dense
s(Qd) is dense in (L0)d, there exists
s(Qd) such that kqn − xk 6 1/3m. Moreover, note that rn 6 kqn − xk 6 1/3m.
s(N), since L0
kx − x(n, m)k 6 kx − qnk + kqn − x(n, m)k 6
1
3m
Now enumerate N × N := {(nh, mh) : h ∈ N}. For each h ∈ L0
1
3m
+
+
1
3m
6
1
m
.
Pk 1Ak hk, let yh := Pk 1Ak x(nhk , mhk ). Considering N as a subset of L0
13In the context of the conditional analysis on Rd, this equivalent characterization of stable compactness
is applied to prove that a stable sequentially lower semi-continuous function attains its minimum on a
stable, sequentially closed and stable bounded set, see [7, Theorem 4.4]. This result was subsequently
applied in e.g. [5, 31] to solve stochastic control problems.
14In [31, Theorem 4.2], a characterization of measurable closed-valued mappings S : Ω ⇒ Rd in terms
of stable sequentially closed subsets of (L0)d is proved, which according to the analysis in this article
yields a one-to-one correspondence between conditional closed subsets of a conditional Euclidean space
and measurable closed-valued maps in Rd.
s(N) of the form h =
s(N) via
18
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
n 7→ 1Ωn, we thus obtain a sequence (yn). Since K is stable bounded, we can choose a
representative yn of yn such that {yn(ω) : n ∈ N} is bounded for all ω ∈ Ω. We define
SK(ω) := cl({yn(ω) : n ∈ N}), ω ∈ Ω.
By Castaing's representation theorem (see e.g. [1, Corollarly 18.14]), SK is a mea-
surable. Moreover, SK is non-empty compact-valued. By inspection, K is the set of
equivalence classes of all measurable selectors of SK.
Finally, let us show that if K is the set of equivalence classes of all measurable
selectors of two measurable non-empty compact-valued mappings S and S′, then S(ω) =
S′(ω) a.e. Indeed, since S and S′ are measurable, they are graph-measurable due to
[1, Theorem 18.6]. Thus Ax := {ω ∈ Ω : x ∈ S(ω)} and Bx := {ω ∈ Ω : x ∈ S′(ω)} are
measurable for all x ∈ Rd.
It suffices to prove that P(Ax∆Bx) = 0 for all x ∈ Rd,
where ∆ denotes symmetric difference. By contradiction, suppose that for instance
P(Ax \ Bx) > 0 for some x ∈ Rd. Let C := Ax \ Bx and choose x0 ∈ K. Since K
is stable, z := 1C x + 1C c x0 ∈ K. Hence z is a measurable selector of K, and thus
z(ω) ∈ S′(ω) for all ω ∈ Ω for a representative z of z, which is a contradiction since
P(C) > 0.
(cid:3)
Let (S, T ) be a stable topological space. A function f : S → ¯L0 is said to be lower
semi-continuous if {x ∈ S : f (x) 6 η} is closed for all η ∈ ¯L0. The following result is
an extension of [7, Theorem 4.4] to the infinite dimensional case.
Theorem 5.13. Let (K, T ) be a stable topological space and f : K → L0 stable and
If K is stable compact, then there exists x0 ∈ K such that
lower semi-continuous.
f (x0) = minx∈K f (x).
Proof. Define r := inf y∈K f (y) and
Cn :=(cid:26)x ∈ K : f (x) 6 r +
s(N).
1
n(cid:27) , n ∈ L0
By inspection, {Cn : n ∈ L0
s(N)} is a stable collection of stable and closed subsets of K.
In fact, since Cn ∩ Cm = Cn∧m, the collection {Cn : n ∈ L0
s(N)} is a stable filter base
on K. Denoting by F the induced stable filter, by stable compactness, F has a cluster
point x0. In particular, x0 ∈ cl(Cn) = Cn for all n ∈ L0
(cid:3)
s(N), and thus f (x0) = r.
6. Functional analysis in L0-modules
In the remainder of this article, we will establish module variants of results in func-
tional analysis. In doing so, the comparison results in Section 4 are key which we recall
in the following. Given a locally L0-convex module (E, T ), according to the discussion
in Remark 4.12, it uniquely defines two topologies Tǫ,λ and Ts with Tǫ,λ ⊂ T ⊂ Ts.
Moreover, for a stable subset S ⊂ E, we have clǫ,λ(S) = cl0(S) = cls(S), and there-
fore we just write cl(S). The stable topology Ts is uniquely related with a conditional
locally convex topology T on E, and cl(S) is the carrier of the conditional closure of
S. Further, recall that if T is stable collection, that is, if T = Ts, then it holds the
s for the topological L0-dual spaces. In this case, we just write
identity E∗
E∗. If T is non-stable, then we only have E∗
s since the first space may be
non-stable. We will make frequent use of these results without further comment.
ǫ,λ = E∗
0 = E∗
0 ⊂ E∗
ǫ,λ = E∗
Throughout, we will suppose that any L0-module considered E has support 1. For
x ∈ E, we define supp(x) := ∧{a ∈ A : 1Acx = 0}. We start with the hyperplane
separation theorem. For a discussion of variants of this theorem w.r.t. the L0- and
STABLE COMPACTNESS
19
(ǫ, λ)-topology, we refer to the comparison in [19]. We state the locally L0-convex
version obtained in [14].
Theorem 6.1. Let (E, T ) be a locally L0-convex module and S1, S2 ⊂ E non-empty
and L0-convex with S1 open. If 1AS1 ∩ 1AS2 = ∅ for all a ∈ A with a > 0, then there
exists f ∈ E∗
0 such that
f (y) 6 f (z)
for all y ∈ S1, z ∈ S2.
With the notion of stable compactness, we can strengthen the previous statement
and obtain the following strong separation result which has not been available before.
Theorem 6.2. Let (E, T ) be a stable locally L0-convex module. Let S1, S2 ⊂ E be
stable L0-convex with S1 stable compact and S2 closed.
If 1AS1 ∩ 1AS2 = ∅ for all
a ∈ A with a > 0, then there exists f ∈ E∗ and r ∈ L0
++ such that
f (x) + r < f (y)
for all x ∈ S1, y ∈ S2.
Proof. The claim follows from [10, Theorem 5.5(ii)].
(cid:3)
A Fenchel-Moreau theorem for stable locally L0-convex modules is proved in [14,
Theorem 3.8]. It was shown in [22] that this result is valid for the L0-topology (see [22,
Theorem 5.5]) and the (ǫ, λ)-topology (see [22, Theorem 5.3]). Recall that a function
f : E → ¯L0 is said to be proper if f (x) > −∞ for all x ∈ E and f (x) ∈ L0 for some
x ∈ E. Let (E, T0) be a locally L0-convex module and f : E → ¯L0 a proper function.
We can consider the following three module conjugates of f :
ǫ,λ : E∗
f ∗
ǫ,λ → ¯L0, f ∗
ǫ,λ(g) := sup
x∈E
(g(x) − f (x)),
f ∗
0 : E∗
0 → ¯L0, f ∗
f ∗
s : E∗
s → ¯L0, f ∗
0 (g) := sup
x∈E
s (g) := sup
x∈E
(g(x) − f (x)),
(g(x) − f (x)).
ǫ,λ = f ∗
ǫ,λ = E∗
s , we have f ∗
Since E∗
of f .
Theorem 6.3. Let (E, T0) be a locally L0-convex module. Let f : E → ¯L0 be proper,
L0-convex and lower semi-continuous. Then
s . Similarly, we can define three module bi-conjugates
f (x) = f ∗∗
0 (x) = f ∗∗
s (x) = f ∗∗
ǫ,λ(x),
x ∈ E.
Proof. Since f is stable, if it is lower semi-continuous w.r.t. T0, then it is also lower
semi-continuous w.r.t. Ts and Tǫ,λ. By [14, Theorem 3.8], we have f ∗∗
s = f . Since
E∗
s , and thus, for fixed
x ∈ E, one has
s = f . Further it holds (E∗
ǫ,λ, one has f ∗∗
ǫ,λ = f ∗∗
0 )s = E∗
s = E∗
f ∗∗
s (x) = sup
g∈E ∗
s
= sup
g∈E ∗
0
(g(x) − f ∗
(g(x) − f ∗
s (g)) = sup
g∈(E ∗
s (x)) = sup
g∈E ∗
0
(g(x) − f ∗
s (g))
0 )s
(g(x) − f ∗
0 (x)) = f ∗∗
0 (g).
In [21, Corollary 3.4], a module variant of Mazur's lemma for the (ǫ, λ)-topology is
proved. The conditional version of this result is obtained in [43, Proposition 3.3]. We
have the following statement.
(cid:3)
20
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
Proposition 6.4. Let (E, T ) be a locally L0-convex module and S ⊂ E stable L0-
convex. Then it holds
cl(S) = clσ(E,E ∗
s )(S) = clσ(E,E ∗
ǫ,λ)(S) = clσ(E,E ∗
0 )(S).
6.5. Stable dual pairs. Given a stable dual pair (E, F, h·, ·i), one has (E, σ0(E, F ))∗ =
F by [23, Theorem 3.4]. For a conditional dual pair (E, F, h·, ·i), it holds (E, σ(E, F))∗ =
F by [32, Corollary 4.48]. We can derive from the latter the following extension of [23,
Theorem 3.4].
Proposition 6.6. Let (E, F, h·, ·i) be a stable dual pair. Then
(E, σ0(E, F ))∗ = (E, σs(E, F ))∗ = (E, σǫ,λ(E, F ))∗ = F.
Proof. Since the induced conditional structure (E, F, h·, ·i) is a conditional dual pair,
by [32, Corollary 4.48], one has (E, σs(E, F ))∗ = F . The claim thus follows from [23,
Theorem 3.4].
(cid:3)
Next, by an adaptation of the proof of the conditional Banach-Alaoglu theorem (see
[10, Theorem 5.10]), we can derive the following module variant of it.
Theorem 6.7. Let (E, T ) be a stable locally L0-convex module. Then
U ◦ := {f ∈ E∗ : f (x) 6 1 for all x ∈ U }
is stable σs(E∗, E)-compact.
Corollary 6.8. Let (E, k · k) be an L0-normed module. Then {f ∈ E∗ : kf k 6 1} is
stable σs(E∗, E)-compact.
For a stable set S, its stable L0-convex hull cos(S) is defined as the collection of all
stable finite sumsP16n6m rnxn, where (xn)16n6m ⊂ S is stable finite and (rn)16n6m ⊂
L0 is stable finite with rn > 0 and P16n6m rn = 1. A bipolar theorem for stable dual
pairs w.r.t. the L0-topology is proved in [23, Theorem 3.4]. Its conditional version is
provided in [10, Theorem 5.9] from which we obtain:
Proposition 6.9. Let (E, F, h·, ·i) be a stable dual pair and S ⊂ E stable. Then
S◦◦ = cl(cos(S ∪ {0})).
6.10. L0-normed modules. Given L0-normed modules (E1, k · k1) and (E2, k · k2), we
denote by L (E1, E2) the set of all continuous L0-linear functions f : E1 → E2. Then
L (E1, E2) has the structure of a stable L0-module. Define the L0-operator norm by
kf k := sup{kf (x)k2 : x ∈ E1, kxk1 6 1}.
We have the following version of the Krein-Smulian theorem, which can be proved by
an adaptation of its conditional version [10, Theorem 5.12].
Theorem 6.11. Let (E, k · k) be a stable complete L0-normed module and S ⊂ E∗
stable L0-convex. Then S is σ(E∗, E)-closed if and only if S ∩ {f ∈ E∗ : kf k 6 r} is
σ(E∗, E)-closed for each r ∈ L0
++.
The conditional Eberlein-Smulian Theorem [43, Theorem 4.7] provides the following
module variant.
Theorem 6.12. Let (E, k · k) be an L0-normed and S ⊂ E stable. Then the following
are equivalent.
(i) S is stable σs(E, E∗)-compact.
STABLE COMPACTNESS
21
(ii) For every stable sequence (xn) ⊂ S there exists a stable subsequence (xnk )
which converges in S w.r.t. the σ0(E, E∗)-topology.
(iii) For every stable sequence (xn) ⊂ S there exists a stable subsequence (xnk )
which converges in S w.r.t. the σs(E, E∗)-topology.
Proof. The equivalence (i) ⇔ (iii) follows from a rountine adaptation of the proof
(iii) ⇒ (ii) is trivially satisfied since σs(E, E∗) is finer than
of [43, Theorem 4.7].
σ0(E, E∗).
(ii) ⇒ (iii): Let (xn) be a stable sequence in S and (xnk ) a stable subsequence
which converges to x ∈ S w.r.t. the σ0(E, E∗)-topology. Any U ∈ Us(x) is of the form
U =Pk 1Ak Uk with (ak) ∈ p(1) and Uk ∈ U0(x) for all k ∈ N. For each k ∈ N, choose
s(N) such that xnk ∈ Uk for all k > mk. Let m := Pk 1Ak mk. Then, xnk ∈ U
mk ∈ L0
for each k > m.
(cid:3)
Finally, a transcription of [39, Theorem 2.5] yields the following module version of a
variant of the James compactness theorem.
Theorem 6.13. Let (E, k · k) be a stable complete L0-normed module such that {f ∈
E∗ : kf k 6 1} is stable sequentially compact. Let S ⊂ E be stable and σs(E, E∗)-closed.
Then S is stable σs(E, E∗)-compact if and only if for each f ∈ E∗ there exists y ∈ E
such that f (y) = supx∈S f (x).
6.14. Stable metric spaces. The following statement is a translation of the condi-
tional version of the Baire category theorem (see [43, Theorem 3.3]).
Theorem 6.15. Let (S, d) be a stable complete metric space. Suppose that (En) is a
stable sequence of stable closed subsets of S with S = ∪En. Then there exist x ∈ S,
r ∈ L0
s(N) such that Br(x) ⊂ En.
++ and n ∈ L0
We have the following module variant of the uniform boundedness principle as a
consequence of [43, Theorem 3.4].
Theorem 6.16. Let (E1, k · k1) and (E2, k · k2) be L0-normed modules and E1 stable
complete. Suppose that S is a stable subset of L (E1, E2) such that for every x ∈ E1
there exists rx ∈ L0 with kf (x)k2 6 rx for all f ∈ S. Then there exists r ∈ L0
++ such
that kf k 6 r for all f ∈ S.
Next, we extend Banach's fixed point theorem to L0-modules.
Theorem 6.17. Let (S, d) be a stable complete metric space and T : S → S a stable
continuous function such that there exists r ∈ L0 with 0 6 r < 1 and such that
d(T (x), T (y)) 6 rd(x, y)
for all x, y ∈ S.
Then there exists a unique z ∈ S satisfying T (z) = z. Furthermore, for any x1 ∈ S, the
stable sequence (xn) defined by xn := Pk∈N 1{n=k}xk, where xn+1 := T (xn) for n > 2,
converges to z.
Proof. Our proof is based on a simple proof of the classical contraction principle given
in [40]. For n ∈ L0
s(N) with n =P nkak, we define
T n(x) :=X T nk (x)ak,
s ∈ S.
From stability we obtain
d(T n(x1), T n(x2)) 6 rnd(x1, x2)
for all x1, x2 ∈ S.
(6.1)
22
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
Now for fixed x1, x2 ∈ S, by triangle inequality,
d(x1, x2) 6 d(x1, T (x1)) + d(T (x1), T (x2)) + d(T (x2), x2),
from which it follows that
1
1 − r
d(x1, x2) 6
(d(x1, T (x1)) + d(x2, T (x2)))
for all x1, x2 ∈ S.
(6.2)
As for uniqueness, if z1, z2 satisfy z1 = T (z1) and z2 = T (z2), then from (6.2) one has
d(z1, z2) = 0. Now fix x ∈ S. For n, m ∈ L0
s(N), if we replace x1 and x2 by T n(x) and
T m(x) in (6.2), then we obtain from (6.1)
d(T n(x), T m(x)) 6
6
1
1 − r
rn + rm
1 − r
(d(T n(x), T n(T (x))) + d(T m(x), T m(T (x))))
d(x, T (x)).
Since r < 1, we obtain limn,m d(T n(x), T m(x)) = 0. Since S is conditional complete,
the conditional sequence (T n(x)) has a conditional limit z ∈ S. By letting first m and
second n tend to ∞, we have T (z) = z.
(cid:3)
Finally, we prove a conditional version of the `Arzela-Ascoli theorem which is based on
an adaptation of a proof of the classical result in [24]. Let (S, T ) be a stable topological
space. Denote by C (S, L0) the set of all stable continuous functions f : S → L0. Notice
that C (S, L0) is a stable L0-module. By Theorem 5.13, for stable compact S,
d∞(f, g) := max {f (x) − g(x) : x ∈ S} ,
f, g ∈ C (S, L0),
is a well-defined stable metric.
Definition 6.18. A stable subset M ⊂ C (S, L0) is stable equicontinuous if for every
x ∈ S and r ∈ L0
++ there exists a neighbourhood V of x such that f (x) − f (y) 6 r for
all y ∈ V and f ∈ M .
Finally, we state a module variant of the Arzel`a-Ascoli theorem.
Theorem 6.19. Let (S, T ) be a stable compact topological space. A stable subset M of
C (S, L0) is stable totally bounded in the stable metric d∞ if and only if it is pointwise
stable bounded and stable equicontinuous.
Lemma 6.20. Let (S, d) be a stable metric space. Suppose that for every r ∈ L0
there exist sr ∈ L0
function fr : S → Sr so that fr(S) is stable totally bounded and for x, y ∈ S
++
++, a stable metric space (Sr, dr) in some L0-module Er, and a stable
dr(fr(x), fr(y)) 6 sr implies d(x, y) 6 r.
(6.3)
Then S is stable totally bounded.
++. Choose sr ∈ L0
Proof. Let r ∈ L0
++, a stable metric space (Sr, dr) and fr : S → Sr
according to the assumptions of the statement above. Since fr(S) is stable totally
bounded, there exists a stable finite subset N ⊂ S such that fr(S) = ∪
Bsr (fr(x)).
x∈S
By (6.3), we have S = ∪
x∈S
Br(x). Thus S is stable totally bounded.
(cid:3)
Theorem 6.21. Let (K, T ) be a stable compact topological space. Then a stable subset
M of C (K, L0) is stable totally bounded with respect to d∞ if and only if
(i) Mx := {f (x) : f ∈ M } is stable bounded for each x ∈ K, and
(ii) M is conditional equicontinuous.
STABLE COMPACTNESS
23
Proof. Suppose that M is stable equicontinuous and Mx is stable bounded for all x ∈ K.
Using that K is stable compact and M is stable equicontinuous, one can find a stable
finite sequence (xn)16n6m ⊂ K, a stable family (Vn)16n6m, where Vn is a neighbourhood
of xn for all 1 6 n 6 m, such that K = ∪16n6mVn and f (xn) − f (y) 6 r whenever
1 6 n 6 m, y ∈ Vn and f ∈ M .
Let g : M → (L0)m be defined by g(f ) := (f (xn))16n6m, which is a stable function.
Note that k(rn)16n6mk∞ := max
16n6m
rn defines an L0-norm on (L0)m.
Since Mxn is stable bounded for each x ∈ K, we can choose rn ∈ L0
++ such that
f (xn) < rn for all f ∈ M . Then, one has that
kg(f )k∞ := max
16n6m
f (xn) 6 max
16n6m
rn < +∞ for all f ∈ M.
This means that g(M ) is stable bounded in ((L0)m, k · k∞).
Now, suppose that f1, f2 ∈ M with kg(f1) − g(f2)k∞ 6 r and y ∈ K. Since K is
covered by (Vn)16n6m, we can pick n, 1 6 n 6 m, such that y ∈ Vn. Then, one has
f1(y) − f2(y) 6 f1(y) − f1(xn) + f1(xn) − f2(xn) + f2(xn) − f2(z) 6 3r.
Since y ∈ K is arbitrary, this implies that d∞(f1, f2) 6 3r. Lemma 6.20 yields the
result.
Conversely, suppose that M ⊂ C (K, L0) is stable totally bounded. In particular, it
is stable bounded and, therefore, Mx is stable bounded for each x ∈ K.
Let us show that M is stable equicontinuous. Indeed, for fixed x ∈ K and r ∈ L0
++,
since M is stable totally bounded, one can pick a stable finite sequence (fn)16n6m ⊂ M
so that M = ∪
Br(fn).
16n6m
For any n ∈ N, we define hn := 1Anfn where an := ∨ {b ∈ A : 1Bn 6 1Bm}. For
each n ∈ N, since hn is continuous at x, we can take a neighbourhood Vn of x so that
hn(h)−hn(y) 6 r for all y ∈ Vn. Now, for any n 6 m, which is of the form n =P nkbk,
we define Vn =P 1Bk Vnk . Then (Vn)16n6m is a stable family of neighbourhood s of x
such that
hn(x) − hn(y) = fn(x) − fn(y) 6 r
for all y ∈ Vn, n 6 m.
Let V := ∪
Vn. If f ∈ M , one has that f ∈ Br(fn) for some n 6 m. Thus, for any
16n6m
y ∈ V it holds
f (y) − f (x) 6 f (y) − fn(x) + fn(x) − fn(y) + fn(y) − f (x) 6 3r.
This proves that M is stable equicontinuous.
(cid:3)
References
[1] C. D. Aliprantis and K. C. Border. Infinite Dimensional Analysis: a Hitchhiker's Guide. Springer,
2006.
[2] J. Backhoff and U. Horst. Conditional analysis and a Principal-Agent problem. SIAM J. Financial
Math., 7(1):477 -- 507, 2016.
[3] J. L. Bell. Set Theory: Boolean-Valued Models and Independence Proofs. Oxford Logic Guides.
Clarendon Press, 2005.
[4] T. Bielecki, I. Cialenco, S. Drapeau, and M. Karliczek. Dynamic Assessement Indices. Stochastics,
88(1):1 -- 44, 2016.
[5] P. Cheridito, U. Horst, M. Kupper, and T. Pirvu. Equilibrium pricing in incomplete markets under
translation invariant preferences. Math. Oper. Res., 41(1):174 -- 195, 2016.
[6] P. Cheridito and Y. Hu. Optimal consumption and investment in incomplete markets with general
constraints. Stoch. Dyn., 11(2):283 -- 299, 2011.
24
ASGAR JAMNESHAN AND JOS´E M. ZAPATA
[7] P. Cheridito, M. Kupper, and N. Vogelpoth. Conditional analysis on Rd. Set Optimization and
Applications, Proceedings in Mathematics & Statistics, 151:179 -- 211, 2015.
[8] P. Cheridito and M. Stadje. BS∆Es and BSDEs with non-Lipschitz drivers: comparison, conver-
gence and robustness. Bernoulli, 19(3):1047 -- 1085, 2013.
[9] S. Drapeau and A. Jamneshan. Conditional preferences and their numerical representations.
J. Math. Econom., 63:106 -- 118, 2016.
[10] S. Drapeau, A. Jamneshan, M. Karliczek, and M. Kupper. The algebra of conditional sets, and the
concepts of conditional topology and compactness. J. Math. Anal. Appl., 437(1):561 -- 589, 2016.
[11] S. Drapeau, A. Jamneshan, and M. Kupper. Vector duality via conditional extension of dual pairs.
Preprint available at arXiv:1608.08709, 2016.
[12] S. Drapeau, A. Jamneshan, and M. Kupper. A Fenchel-Moreau theorem for ¯L0-valued functions.
Preprint available at arXiv:1708.03127, 2017.
[13] S. Drapeau, M. Karliczek, M. Kupper, and M. Streckfuss. Brouwer fixed point theorem in (L0)d.
J. Fixed Point Theory Appl., 301(1), 2013.
[14] D. Filipovi´c, M. Kupper, and N. Vogelpoth. Separation and duality in locally L0-convex modules.
J. Funct. Anal., 256:3996 -- 4029, 2009.
[15] D. Filipovi´c, M. Kupper, and N. Vogelpoth. Approaches to conditional risk. SIAM J. Financial
Math., 3(1):402 -- 432, 2012.
[16] M. Frittelli and M. Maggis. Complete duality for quasiconvex dynamic risk measures on modules
of the Lp-type. Stat. Risk Model., 31(1):103 -- 128, 2014.
[17] S. R. Givant and P. R. Halmos. Introduction to Boolean algebras. Undergraduate texts in mathe-
matics. Springer, 2009.
[18] T. Guo. The relation of Banach-Alaoglu theorem and Banach-Bourbaki-Kakutani-Smulian theo-
rem in complete random normed modules to stratification structure. Science in China Series A
Mathematics, 51:1651 -- 1663, 2008.
[19] T. Guo. Relations between some basic results derived from two kinds of topologies for a random
locally convex module. J. Funct. Anal., 258:3024 -- 3047, 2010.
[20] T. Guo and G. Shi. The algebraic structure of finitely generated L0(F, K)-modules and the Helly
theorem in random normed modules. J. Math. Anal. Appl., 381(2):833 -- 842, 2011.
[21] T. Guo, H. Xiao, and X. Chen. A basic strict separation theorem in random locally convex modules.
Nonlinear Anal. Real World Appl., 71:3794 -- 3804, 2009.
[22] T. Guo, S. Zhao, and X. Zeng. Random convex analysis (I): Separation and Fenchel-Moreau duality
in random locally convex modules. Scientia Sinica Mathematica, 45(12):1961 -- 80, 2015.
[23] T. Guo, S. Zhao, and X. Zeng. Random convex analysis (II): continuity and subdifferentiability the-
orems in L0-pre-barreled random locally convex modules. Preprint available at arXiv:1503.08637,
2015.
[24] H. Hanche-Olsen and H. Holden. The Kolmogorov-Riesz compactness theorem. Preprint available
at arXiv:0906.4883, 2009.
[25] L. P. Hansen and S. F. Richard. The Role of Conditioning Information in Deducing Testable
Restrictions Implied by Dynamic Asset Pricing Models. Econometrica, 55:587 -- 613, 1987.
[26] R. Haydon, M. Levy, and Y. Raynaud. Randomly Normed Spaces. Hermann, 1991.
[27] T. Hungerford. Algebra. Graduate Texts in Mathematics. Springer New York, 2003.
[28] A. Jamneshan. A theory of conditional sets. PhD thesis, Humboldt-Universitat zu Berlin, 2014.
[29] A. Jamneshan. Sheaves and conditional sets. Preprint available at arXiv:1403.5405, 2014.
[30] A. Jamneshan, M. Kupper, and M. Streckfuss. Measures and integrals in conditional set theory.
Preprint available at arXiv:1701.02661, 2017.
[31] A. Jamneshan, M. Kupper, and J. M. Zapata. Parameter-dependent stochastic optimal control in
finite discrete time. Preprint available at arXiv: 1705.02374, 2017.
[32] M. Karliczek. Elements of conditional optimization and their applications to order theory. PhD
thesis, Humboldt-Universitat zu Berlin, 2014.
[33] M. Kupper and N. Vogelpoth. Complete L0-Normed Modules and Automatic Continuity of Mono-
tone Convex Functions. Working paper, 2008.
[34] A. G. Kusraev and S. S. Kutateladze. Boolean Valued Analysis. Mathematics and Its Applications.
Springer Netherlands, 2012.
[35] S. S. Kutateladze. Nonstandard Analysis and Vector Lattices. Mathematics and Its Applications.
Springer Netherlands, 2012.
[36] S. MacLane and I. Moerdijk. Sheaves in Geometry and Logic: A First Introduction to Topos
Theory. Universitext. Springer New York, 2012.
STABLE COMPACTNESS
25
[37] I. Molchanov. Theory of Random Sets. Probability and Its Applications. Springer London, 2005.
[38] J. Orihuela and J. M. Zapata. A note on conditional risk measures of Orlicz spaces and Orlicz-type
modules. Preprint available at arXiv:1612.03680, 2016.
[39] J. Orihuela and J. M. Zapata. Stability in locally L0-convex modules and a conditional version of
James' compactness theorem. J. Math. Anal. Appl., 452(2):1101 -- 1127, 2017.
[40] R. S. Palais. A simple proof of the Banach contraction principle. J. Fixed Point Theory Appl.,
2:221 -- 223, 2007.
[41] B. Schweizer and A. Sklar. Probabilistic Metric Spaces. Dover Publications, 2011.
[42] J. M. Zapata. Randomized versions of Mazur lemma and Krein-Smulian theorem. To appear in J.
Convex Anal.
[43] J. M. Zapata. Versions of Eberlein-Smulian and Amir-Lindenstrauss theorems in the framework
of conditional sets. Appl. Anal. Discrete Math., 10(2):231261, 2016.
[44] J. M. Zapata. On the Characterization of Locally L0-Convex Topologies Induced by a Family of
L0-Seminorms. J. Convex Anal., 24(2):383 -- 391, 2017.
Department of Mathematics and Statistics, University of Konstanz
E-mail address: [email protected]
Departamento de Matem´aticas, Universidad de Murcia
E-mail address: [email protected]
|
1607.05933 | 1 | 1607 | 2016-07-20T12:25:46 | On the structure of universal differentiability sets | [
"math.FA"
] | We prove that universal differentiability sets in Euclidean spaces possess distinctive structural properties. Namely, we show that any universal differentiability set contains a `kernel' in which the points of differentiability of each Lipschitz function are dense. We further prove that no universal differentiability set may be decomposed as a countable union of relatively closed, non-universal differentiability sets. The sharpness of this result, with respect to existing decomposibility results of the opposite nature, is discussed. | math.FA | math |
On the structure of
universal differentiability sets.∗
Michael Dymond
7th September 2018
We prove that universal differentiability sets in Euclidean spaces possess
distinctive structural properties. Namely, we show that any universal differ-
entiability set contains a 'kernel' in which the points of differentiability of
each Lipschitz function are dense. We further prove that no universal differ-
entiability set may be decomposed as a countable union of relatively closed,
non-universal differentiability sets. The sharpness of this result, with respect
to existing decomposibility results of the opposite nature, is discussed.
1 Introduction.
Subsets of Rd containing a point of differentiability of every Lipschitz function f : Rd → R
form a complex and still somewhat mysterious class of sets, despite significant modern
progress. Such sets are called universal differentiability sets (or UDSs), a term introduced
in [5]. The classical Rademacher's Theorem states that Lipschitz functions on Euclidean
spaces are differentiable almost everywhere with respect to the Lebesgue measure. Thus,
every set of positive measure is a universal differentiability set. Whilst one may char-
acterise universal differentiability sets in R as sets of positive Lebesgue measure (see
[11]), this description fails in all Euclidean spaces of higher dimension. Preiss proves,
in [8], that R2 contains a dense, Gδ universal differentiability set of Lebesgue measure
zero. In [3], Doré and Maleva verify the existence of compact universal differentiability
sets of Lebesgue measure zero. Recent work in [5] and [6] establishes the existence of
universal differentiability sets of an exceptional nature with respect to the Hausdorff and
Minkowski dimensions.
In order to better understand the nature of universal differentiability sets, it is helpful
to study the class of porous and σ-porous sets. A subset P ∈ Rd is called porous if there
exists c ∈ (0, 1) such that for every x ∈ P and ε > 0, there exists h ∈ B(x, ε) such that
∗The research presented in this paper was completed when the author was a PhD student, supervised
by Dr. Olga Maleva at University of Birmingham, UK. The paper is based on part of the author's
PhD thesis. The author was supported by EPSRC funding.
1
B(h, c kh − xk) ∩ P = ∅. A subset F ⊆ Rd is called σ-porous if F may be expressed as
a countable union of porous sets. σ-porous sets in Rd are negliglible in several senses.
Any σ-porous set is of the first category with respect to the Baire Category Theorem.
Moreover, unlike sets of the first category, σ-porous subsets of Rn must have Lebesgue
measure zero, a consequence of the Lebesgue Density Theorem. For a survey on porous
and σ-porous sets we refer to [12].
Porous and σ-porous sets are extremely relevant in the study of differentiability of
Lipschitz functions because they admit Lipschitz functions which are nowhere differen-
tiable inside them. Given a porous set P ⊆ Rd the distance function x 7→ dist(x, P ) is
Lipschitz and fails to be differentiable at any point of P . Moreover, the paper [10] proves
that every σ-porous subset of Rd is a non-universal differentiability set. Universal dif-
ferentiability sets are therefore necessarily non-σ-porous and further there appears to be
an intimate relationship between these two classes of sets. We note, however, that non-
σ-porosity is not sufficient for the universal differentiability property: Zajicek constructs
in [12] a non-σ-porous subset of the interval [0, 1] with Lebesgue measure zero. This is
a non-universal differentiability set in R and its pre-image under the first co-ordinate
projection map p1 : Rd → R gives a non-σ-porous, non-universal differentiability set in
Rd with d > 1.
A fundamental property of non-σ-porous sets is that they cannot by decomposed:
whenever a non-σ-porous set G is expressed as a countable union of sets Gi, at least one
of the sets Gi has to be non-σ-porous. The research presented in this paper began with the
question of whether universal differentiability sets exhibit the same robustness. That is,
does every countable decomposition of a universal differentiability set necessarily contain
a universal differentiability set? After a short time, we found an example to show that
the answer is negative, using a result of Alberti, Csörnyei and Preiss in [1] (see Section 2).
Nevertheless, it became clear that the universal differentiability property imposes strict
demands on the nature of sets possessing it.
In this paper we examine the structural properties of universal differentiability sets
in Euclidean spaces. Our approach is motivated by the work [13], of Zelený and Pelant
on the structure of non-σ-porous sets, which provides a level of insight into the nature
of non-σ-porous sets not yet available for universal differentiability sets. In particular,
we prove that, like non-σ-porous sets, universal differentiability sets contain a 'kernel',
which in some sense captures the core or essence of the set. In the papers [3], [5] and
[4], Doré and Maleva observe that the universal differentiability sets constructed possess
the property that the differentiability points of each Lipschitz function form a dense
subset. We verify that this is, broadly speaking, an intrinsic property of universal dif-
ferentiability sets. We go on to establish that no universal differentiability set can be
decomposed as a countable union of relatively closed, non-universal differentiability sets.
Our main results are stated in Section 2 and proved in Section 3. Finally, in Section 4
we give an application to a question of Godefroy relating to the existence of exceptional
universal differentiability sets.
2
2 Main Results
In this section we present our main results and discuss their connections to the existing
theory. Our two main theorems are based on the following lemma:
Lemma 2.1. Let F ⊆ Rd be a universal differentiability set and suppose that A is a
relatively closed subset of F . Then either A or F \ A is a universal differentiability set.
In general, a universal differentiability set in Rd may be decomposed as a countable
union of non-universal differentiability sets. Indeed, let S be a universal differentiability
set in R2 with Lebesgue measure zero (such a set is given in [8]). By a result of Alberti,
Csörnyei and Preiss in [1], there exist Lipschitz functions f, g : R2 → R such that f and
g have no common points of differentiability inside S. Writing Df for the set of points
of differentiability of f , we get that
S = (S \ Df ) ∪ (S ∩ Df )
is a decomposition of S as a union of two non-universal differentiabilty sets.
The above discussion indicates that questions about decomposability of universal diff-
ferentiability sets are closely related to questions about simultaneous differentiability of
Lipschitz functions. The latter is an active area of current research: In [7], Lindenstrauss,
Preiss and Tišer prove that every pair of Lipschitz functions on a Hilbert space have a
common point of differentiability; the corresponding question for a triple of Lipschtiz
functions remains open.
Gδ sets arise naturally in the theory of universal differentiability sets because they
admit an equivalent metric with respect to which they are complete, see [8] and [9].
We therefore considered the question of whether it is possible to weaken the assumption
on A in Lemma 2.1 to be Gδ rather than closed.
It turns out that it is not possible
to improve Lemma 2.1 in this manner: Csörnyei, Preiss and Tišer construct, in [2], a
universal differentiability set S in R2 which may be decomposed as the union of two
non-universal differentiability sets U and V , where U is uniformly purely unrectifiable.
Since any uniformly purely unrectifiable set is both a non-universal differentiability set
and contained in a Gδ uniformly purely unrectifiable set (see [2]), it follows that the set
U may be taken to be a Gδ subset of S.
We define the kernel, ker(S), of a set S ⊆ Rd with respect to universal differentiability
similarly to the kernel of S with respect to non-σ-porosity, see [13, Definition 3.2].
Definition 2.2.
Given a set S ⊆ Rd, we let
ker(S) = S \ {x ∈ S : ∃ε > 0 such that B(x, ε) ∩ S is a non-UDS} .
Note that ker(S) is closed as a subset of S. Through an application of Lemma 2.1 we
obtain the following theorem, which shows that the kernel of a universal differentiability
set can be thought of as the core of the set. We remark that universal differentiability
sets behave similarly to non-σ-porous sets in this respect - see [13, Lemma 3.4].
3
Theorem 2.3. Suppose F ⊆ Rd is a universal differentiability set. Then,
(i) ker(F ) ⊆ F is a universal differentiability set.
(ii) ker(ker(F )) = ker(F ) and F \ ker(F ) is a non-universal differentiability set. In
particular, for each Lipschitz function f : Rd → R, the differentiability points of f
in ker(F ) form a dense subset of ker(F ).
An iterative construction based on the proof of Lemma 2.1 leads to our second main
theorem:
Theorem 2.4. Suppose that E ⊆ Rd is a universal differentiability set and that (Ai)∞
i=1
i=1 Ai. Then at least one
is a collection of relatively closed subsets of E satisfying E =S∞
of the sets Ai is a universal differentiability set.
We finish this section with a sketch of the proof of Lemma 2.1. Suppose that the
contrary holds for some universal differentiability set F and a relatively closed, non-
universal differentiability set A ⊆ F . We aim to construct a Lipschitz function f : Rd →
R which is nowhere differentiable in F . This provides the desired contradiction. The set
F \ A is an open subset of F and a non-universal differentiability set. Accordingly, we
may cover F \ A by a countable collection of boxes Ui with pairwise disjoint interiors,
such that F ∩ Ui is a non-universal differentiability set for each i. On each box Ui we
construct the function f so that it is nowhere differentiable in F ∩ (Ui \ ∂Ui) and satisfies
f = h on ∂Ui, where h : Rd → R is a fixed Lipschitz function nowhere differentiable on
A. Outside of the boxes Ui we simply set f = h. The obtained Lipschitz function f is
clearly nowhere differentiable in F ∩ (Ui \ ∂Ui) for each i, whilst the non-differentiability
of f at all points of A is achieved by controlling the size of the boxes Ui. Using the
closedness of A, we ensure that the diameter of the boxes Ui goes to zero with their
distance from A. Therefore, around 'difficult' points in A, the boundaries of the boxes
Ui become increasingly concentrated. Since f = h on these boundaries, we get that the
differentiability of f coincides with that of h at points of A.
3 Construction.
In this section we will prove the main results stated in Section 2. We begin with a
summary of the notation that we will use: We fix an integer d ≥ 2 and let e1, e2, . . . , ed
denote the standard basis of Rd. For a point x ∈ Rd and ε > 0, we let B(x, ε) (respectively
B(x, ε)) denote the open (respectively closed) ball with centre x and radius ε. The
corresponding norm k−k is the standard Euclidean norm on Rd. Given a set S ⊆ Rd, we
let Int(S) denote the interior, Clos(S) denote the closure and ∂S denote the boundary
of S. For non-empty subsets A and B of Rd we let
diam(A) = sup(cid:8)(cid:13)(cid:13)a′ − a(cid:13)(cid:13) : a, a′ ∈ A(cid:9) and dist(A, B) = inf {kb − ak : a ∈ A, b ∈ B} .
When A = {a} is a singleton, we will just write dist(a, B) rather than dist({a} , B). We
also adopt the convention dist(A, ∅) = 1 for all A ⊆ Rd. We write Lip(f ) for the Lipschitz
4
constant of a Lipschitz function f . Moreover, given e ∈ Sd−1 and function f : Rd → R
we let f ′(x, e) := limt→0+
denote the one-sided directional derivative of f ,
provided that the limit exists. The restriction of f to a set S is denoted by f S and the
support of f by supp(f ).
f (x+te)−f (x)
t
A subset U of Rd is called a box if U = I1 × I2 × . . . × Id for some sequence of
closed, bounded intervals I1, . . . , Id ⊆ R. Writing Ik = [ak, bk] for each k, we call a
set Y ⊆ ∂U a face of U if there exist m ∈ {1, . . . , d} and y ∈ {am, bm} such that
Y = I1 × . . . × Im−1 × {y} × Im+1 × . . . × Id. Note that each face of U is a subset
of a (d − 1)-dimensional hyperplane which is orthogonal to exactly one of the vectors
e1, . . . , ed.
Lemma 3.1. Suppose that A is a relatively closed subset of F ⊆ Rd. Then, there exists
a collection {Ui}∞
i=1 of boxes with pairwise disjoint interiors such that
F ∩
∞[i=1
Ui = F \ A and
diam(Ui)
dist(Ui, A)
→ 0 as i → ∞.
Proof. Let {Si}∞
Furthermore, if {Vi}∞
i=1 is a collection of boxes with pairwise disjoint interiors satisfying
i=1 can be chosen so that for each index i
i=1 Vi, then the collection {Ui}∞
there exists an index j such that Ui ⊆ Vj.
F \ A ⊆ F ∩S∞
F ∩S∞
relatively closed in F . For each i, k ≥ 1 we let Si,k = Si∩Vk and note that F ∩S∞
i=1 be a collection of boxes with pairwise disjoint interiors such that
i=1 Si = F \ A. Such a collection is easy to construct using the fact that A is
i,k=1 Si,k =
F \A. Moreover, W = {Si,k}i,k≥1 is a countable collection of boxes with pairwise disjoint
interiors. After relabelling, we can write W = {Wi}∞
i=1.
Set p0 = 0. For each i ≥ 1, partition the box Wi into a finite number of boxes
Upi−1+1, . . . , Upi with pairwise disjoint interiors such that
diam(Uj)
dist(Uj, A)
≤ 2−i for pi−1 + 1 ≤ j ≤ pi.
The assertions of the Lemma are now readily verified.
The proof of Theorem 2.4 will involve constructing a sequence of Lipschitz functions
which converge to a Lipschitz function nowhere differentiable on the set E. We will need
the next Lemma in order to pick, at each step of the construction, a Lipschitz function
nowhere differentiable on part of the set.
Lemma 3.2. Let Ω be a bounded subset of Rd. A set E ⊆ Rd is a non-universal differ-
entiability set if and only if for every ε > 0, there exists a Lipschitz function f : Rd → R
such that kf Ωk∞ ≤ ε, Lip(f ) ≤ ε and f is nowhere differentiable in E.
Proof. We focus only on the non-trivial direction. Suppose that there exists ε > 0 such
that every Lipschitz function f : Rd → R, with kfΩk∞ and Lip(f ) ≤ ε, has a point of
differentiability inside E. We may assume that ε < 1.
5
Fix a Lipschitz function g : Rd → R. We show that g has a point of differentiability
inside E. We note that kgΩk∞ + Lip(g) is finite, since Ω is bounded. Further, we may
assume that kgΩk∞ + Lip(g) 6= 0. Let
h(x) =
ε
2
+
ε
4(kgΩk∞ + Lip(g))
g(x)
for all x ∈ Rd.
Note that (cid:13)(cid:13)h(x) − ε
4 for all x ∈ Ω. Hence khΩk∞ < ε. Moreover, we have
h : Rd → R with Lip(h) ≤ ε
4 < ε. Hence, there exists a point x ∈ E such that h is
differentiable at x. But then g is also differentiable at x because g = αh + β where
α, β ∈ R are fixed constants. Since g : Rd → R was arbitrary, we deduce that E is a
universal differentiability set.
2(cid:13)(cid:13) ≤ ε
Using the next Lemma, we will later be able to ignore points lying in the boundaries
of boxes.
Lemma 3.3. Suppose E is a subset of Rd and {Ui}∞
i=1 is a collection of boxes in Rd
i=1 ∂Ui is a non-universal differentiability set. Then E is a non-universal
differentiability set.
such that E \S∞
are orthogonal to ej Then S∞
Proof. For j = 1, . . . , d, let Hj denote the union of all faces of the boxes Ui which
j=1 Hj. Moreover, writing pj for the j-th co-
ordinate projection map on Rd, we have that pj(Hj) is a subset of R with one-dimensional
Lebesgue measure zero, (in fact it is a countable set) for each j = 1, . . . , d. The result
now follows from [6, Lemma 2.1].
i=1 ∂Ui = Sd
Lemma 3.4. Let E ⊆ Rd, η > 0 and {Ui}∞
i=1 be a collection of boxes with pairwise
disjoint interiors such that and E ∩ Int(Ui) is a non-universal differentiability set for
each i. Then there exists a function g : Rd → R such that
kgk∞ ≤ η and Lip(g) ≤ η,
g is nowhere differentiable in
E ∩ Int(Ui), and
∞[i=1
whenever x ∈ Rd \ ∞[i=1
Int(Ui)! .
g(x) = 0
(3.1)
(3.2)
(3.3)
Proof. For x ∈ Rd \ (S∞
i=1 Int(Ui)) we define g(x) according to (3.3). Given i ∈ N we
define g on Int(Ui) as follows: Let ϕ = ϕi ∈ C ∞(Rd) be a smooth function satisfying
ϕ(x) > 0 for all x ∈ Int(Ui), kϕk∞ ≤ 1, Lip(ϕ) ≤ 1 and ϕ(x) = 0 for all x ∈ Rd \ Int(Ui).
By Lemma 3.2, there exists a Lipschitz function h = hi : Rd → R such that h is nowhere
differentiable in E ∩ Int(Ui), khUik∞ ≤ η/2 and Lip(h) ≤ η/2. We define g on Int(Ui) by
g(x) = ϕ(x)h(x). The smoothness of ϕ and the fact that ϕ > 0 on Int(Ui) ensure that g
inherits all non-differentiability points of h in Int(Ui). The assertions of the lemma are
now readily verified.
6
The previous two lemmas admit the following corollary:
Corollary 3.5. A set E ⊆ Rd is a non-universal differentiability set if and only if
for every x ∈ E, there exists ε = εx > 0 such that B(x, ε) ∩ E is a non-universal
differentiability set.
In the next lemma, we show that, given a Lipschitz function h and a collection of
pairwise disjoint boxes, we can slightly modify the function h so that it becomes dif-
ferentiable everywhere in the interiors of the boxes and remains unchanged everywhere
else.
Lemma 3.6. Let {Ui}∞
i=1 be a collection of boxes in Rd with pairwise disjoint interiors,
h : Rd → R be a Lipschitz function and σ > 0. Then, there exists a Lipschitz function
bh : Rd → R such that
(3.4)
(3.5)
(3.6)
Int(Ui), and
for all x ∈ Rd \
Int(Ui).
≤ σ and Lip(bh) ≤ Lip(h) + σ,
∞[i=1
(cid:13)(cid:13)(cid:13)bh − h(cid:13)(cid:13)(cid:13)∞
bh is everywhere differentiable inside
∞[i=1
bh(x) = h(x)
Proof. We may assume that Lip(h) > 0. Outside ofS∞
bh according to (3.6). Given i ∈ N we definebh on Int(Ui) as follows: Fix a C ∞ function
π : Rd → R such that π(z) ≥ 0 for all z ∈ Rd, supp(π) ⊆ B(0, 1) and RRd π(z)dz = 1.
γ(x) = 0 for all x ∈ Rd \ Int(Ui) and Lip(γ) ≤ σ/Lip(h). We now definebh on Int(Ui) by
γ(x)dRRd h(x − z)π(z/γ(x))dz for all x ∈ Int(Ui).
bh(x) = 1
We presently verify statement (3.4) forbh. Fix i ∈ N and x ∈ Int(Ui). Then
(cid:13)(cid:13)(cid:13)bh(x) − h(x)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)
π(z/γ(x))dz(cid:13)(cid:13)(cid:13)(cid:13)
Choose a C ∞ function γ = γi : Rd → R such that 0 < γ(x) ≤ σ/Lip(h) for all x ∈ Int(Ui),
kh(x − z) − h(x)k π(z/γ(x))dz ≤ γ(x)Lip(h) ≤ σ.
i=1 Int(Ui), we define the function
h(x − z)π(z/γ(x))dz − h(x)
1
γ(x)dZRd
1
γ(x)dZRd
γ(x)dZRd
1
≤
This establishes the first part of (3.4). The penultimate inequality also proves that h is
continuous on Rd, since γ(x) = 0 for all x ∈ ∂Ui. Thus, for the second part of (3.4), it
suffices to check the Lipschitz condition for points x, y ∈ Int(Ui). For such x, y we have
1
(cid:13)(cid:13)(cid:13)bh(y) −bh(x)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)
γ(y)dZRd
≤ZB(0,1)
h(y − z)π(z/γ(y))dz −
1
γ(x)dZRd
kh(y − γ(y)v) − h(x − γ(x)v)k π(v)dv
h(x − z)π(z/γ(x))dz(cid:13)(cid:13)(cid:13)(cid:13)
≤ Lip(h)(1 + Lip(γ)) ky − xk ≤ (Lip(h) + σ) ky − xk ,
7
as required. Finally, let us verify (3.5). Fix i ∈ N and define for each u ∈ Rd a map
Φu : Int(Ui) → Rd by
Φu(x) = π(x − u/γ(x)).
Note that each Φu is a differentiable function and for x ∈ Int(Ui) we can writebh(x) =
eh(x)/γ(x)d where
prove thateh is every differentiable in Int(Ui). We show that
Since the function Int(Ui) → R, x 7→ 1/γ(x)d is everywhere differentiable, it suffices to
eh(x) =ZRd
for all x ∈ Int(Ui), e ∈ Sd−1.
h(u)Φu(x)du
u(x, e)du
h(u)Φ′
Fix x ∈ Int(Ui) and e ∈ Sd−1. Next, choose t0 small enough so that for all t ∈ (0, t0)
we have
x + te ∈ Int(Ui),
1/2 ≤ γ(x + te)/γ(x) ≤ 2,
t/γ(x + te) ≤ 1/2.
(3.7)
eh′(x, e) =ZRd
Observe that
eh(x + te) −eh(x)
t
=ZRd
h(u)(cid:20) Φu(x + te) − Φu(x)
t
(cid:21) du,
and for each u we have
Therefore, using the Dominated Convergence Theorem, we only need to show that there
exists ϕ ∈ L1(Rd) such that
(cid:21) = h(u)Φ′
u(x, e).
for all
t ∈ (0, t0), u ∈ Rd.
lim
t→0
t
h(u)(cid:20) Φu(x + te) − Φu(x)
(cid:21)(cid:13)(cid:13)(cid:13)(cid:13) ≤ ϕ(u)
t
(cid:13)(cid:13)(cid:13)(cid:13)h(u)(cid:20) Φu(x + te) − Φu(x)
γ(x + te)(cid:13)(cid:13)(cid:13)(cid:13) ≥
x + te − u
(cid:13)(cid:13)(cid:13)(cid:13)
Let t ∈ (0, t0). Suppose ku − xk > 4γ(x). Then, from (3.7),
4γ(x)
t
−
γ(x + te)
γ(x + te)
≥ 2 −
1
2
> 1.
t
It follows that π(x + te − u/γ(x + te)) = 0. Similarly, we have π(x − u/γ(x)) = 0. Hence,
Φu(x + te) − Φu(x)/t = 0 for all u ∈ Rd \ B(0, 4γ(x)) and t ∈ (0, t0).
B(x, 4γ(x)) and t ∈ (0, t0). Fix u ∈ B(x, 4γ(x)) and t ∈ (0, t0). Then, using (3.7) we
obtain
It now suffices to show that (cid:13)(cid:13)(cid:13)h(u)h Φu(x+te)−Φu(x)
(cid:21)(cid:13)(cid:13)(cid:13)(cid:13) ≤ khk∞ Lip(π)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)h(u)(cid:20) Φu(x + te) − Φu(x)
γ(x)2(cid:18)(cid:13)(cid:13)(cid:13)(cid:13)
≤(cid:13)(cid:13)(cid:13)hB(x,4γ(x))(cid:13)(cid:13)(cid:13)∞
≤(cid:13)(cid:13)(cid:13)hB(x,4γ(x))(cid:13)(cid:13)(cid:13)∞
Lip(π)
Lip(π)
γ(x)2 (Lip(γ)4γ(x) + γ(x)).
γ(x + te) − γ(x)
i(cid:13)(cid:13)(cid:13) is uniformly bounded for u ∈
tγ(x)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13) kx − uk + γ(x)(cid:19)
x + te − u
tγ(x + te)
x − u
−
t
2
2
t
8
In our main construction, we are faced with a situation where we would like to
slightly modify a Lipschitz function h to obtain a new function f , whilst preserving
non-differentiability points. Using the previous lemma, we ensure that f coincides with
h on the boundaries of boxes. The application of the following lemma, is to show that if
these boundaries become increasingly concentrated around a point x, then the differen-
tiability of f and h at x coincide.
Lemma 3.7. Let x ∈ Rd, e ∈ Sd−1, f, h : Rd → R be Lipschitz functions and ∆ > 0.
Suppose that {(tk,1, tk,2)}N
k=1, where 0 < tk,1 ≤ tk,2 and N ∈ N ∪ {∞}, is a finite or
countable collection of open, possibly degenerate intervals inside (0, ∞) such that the
following conditions hold:
If N = ∞ then
tk,2 − tk,1
tk,1
f (x + te) = h(x + te)
→ 0 as k → ∞.
∀t ∈ [0, ∆) \
(tk,1, tk,2)! ∪ N[k=1
N[k=1
(3.8)
(3.9)
{tk,1, tk,2}! .
Then the directional derivative f ′(x, e) exists if and only if the directional derivative h′(x, e)
exists and f ′(x, e) = h′(x, e).
Proof. Note that the statement is symmetric with respect to f and h. We may assume
that Lip(f ) + Lip(h) > 0.
Suppose that the directional derivative f ′(x, e) exists. Fix ε > 0 and choose δ > 0 so
that
If N < ∞ we set K = N. Otherwise, using (3.8), we may pick K ≥ 1 large enough so
that
for every k ≥ K.
(3.11)
(cid:12)(cid:12)f (x + te) − f (x) − tf ′(x, e)(cid:12)(cid:12) ≤
tk,2 − tk,1
≤
ε
tk,1
3(Lip(f ) + Lip(h))
ε
3
t whenever
0 < t < δ.
(3.10)
Let t ∈ (0, min {δ, ∆, t1,1, . . . , tK,1}. We distinguish two cases:
First assume that t ∈ (0, ∞) \(cid:16)SN
t ∈ (0, min {δ, ∆}) we get
k=1(tk,1, tk,2)(cid:17). Then, combining (3.9), (3.10) and
(cid:12)(cid:12)h(x + te) − h(x) − tf ′(x, e)(cid:12)(cid:12) =(cid:12)(cid:12)f (x + te) − f (x) − tf ′(x, e)(cid:12)(cid:12) ≤
In the remaining case there exists k such that t ∈ (tk,1, tk,2). Moreover, since t < tl,1 for
1 ≤ l ≤ K, we must have k > K. Note, in particular, that in this case N = ∞. Using
(3.9), (3.11) and (3.10) we deduce
ε
3
t.
(cid:12)(cid:12)h(x + te) − h(x) − tf ′(x, e)(cid:12)(cid:12) ≤ h(x + te) − h(x + tk,1e)
+ f (x + tk,1e) − f (x + te) +(cid:12)(cid:12)f (x + te) − f (x) − tf ′(x, e)(cid:12)(cid:12)
≤ (Lip(h) + Lip(f ))(t − tk,1) +
t ≤
(Lip(h) + Lip(f ))(tk,2 − tk,1)
ε
3
tk,1
t +
ε
3
t ≤ εt.
We have now established that the directional derivative h′(x, e) exists and equals f ′(x, e).
9
Lemma 3.8. Let F ⊆ Rd and let A be a closed subset of F . Let h : Rd → R be a
Lipschitz function and let {Ui}∞
i=1 be a collection of boxes with pairwise disjoint interiors
such that
Ui = F \ A and
diam(Ui)
dist(Ui, A)
→ 0
as
i → ∞.
(3.12)
F ∩
∞[i=1
Then the following two statements hold:
1. Suppose that F \ A is a non-universal differentiability set and let ε > 0. Then there
exists a Lipschitz function f : Rd → R with the following properties:
f is nowhere differentiable in (F \ A) \ ∞[i=1
and Lip(f ) ≤ Lip(h) + ε,
kf − hk∞ ≤ ε
f (y) = h(y) whenever
y ∈ Rd \
Int(Ui).
∞[i=1
∂Ui! ,
(3.13)
(3.14)
(3.15)
2. Let x ∈ A and suppose f : Rd → R is any Lipschitz function satisfying the condition
(3.15). Then f is differentiable at x if and only if h is differentiable at x.
Proof. Let us first verify statement 1. Suppose F \ A is a non-universal differentiability
set and let ε > 0. Note that the boxes {Ui}∞
i=1, the function h and σ = ε/2 satisfy the
Lemma 3.6.
The conditions of Lemma 3.4 are satisfied for E = F \ A, the collection {Ui}∞
conditions of Lemma 3.6. Let bh : Rd → R be the function given by the conclusion of
η = ε/2. Let g : Rd → R be given by the conclusion of Lemma 3.4. We define f =eh + g.
The assertions (3.13), (3.14) and (3.15) now follow easily from the properties ofeh and g.
Finally, we prove statement 2. Suppose f : Rd → R satisfies (3.15) and x ∈ A. Let
e ∈ Sd−1 be any direction. We show that f ′(x, e) exists if and only if h′(x, e) exists and
f ′(x, e) = h′(x, e). Since e ∈ Sd−1 is arbitrary, this suffices.
i=1 and
Let {Uik}N
k=1, where N ∈ N ∪ {∞}, be the collection of all boxes Ui which intersect
i=1 Ui. For
the line x + [0, ∞)e. Since x ∈ A ⊆ F and (3.12) holds, we have that x /∈S∞
each k ≥ 1 we write
Uik ∩ (x + (0, ∞)e) = [x + tk,1e, x + tk,2e]
(3.16)
where tk,1 ≤ tk,2 are strictly positive real numbers. Choose ∆ > 0 small enough so that
B(x, ∆) ⊆ O. We note that the following condition is satisfied:
x + te ∈(∂Uik
Rd \ (S∞
i=1 Int(Ui))
if t = tk,j, j = 1, 2,
if t ∈(cid:16)[0, ∆) \SN
k=1(tk,1, tk,2)(cid:17) ∪SN
k=1 {tk,1, tk,2} .
(3.17)
Let us verify that the conditions of Lemma 3.7 hold for x, e, f , h, ∆ and the intervals
{(tk,1, tk,2)}N
k=1. In view of the conclusion of Lemma 3.7, this completes the proof.
10
If N = ∞, then using (3.16), we get that
0 ≤
tk,2 − tk,1
tk,1
≤
diam(Uik )
dist(Uik , A)
→ 0 as k → ∞.
This proves (3.8). Finally, we note that (3.9) follows from (3.17) and (3.15).
We are now ready to combine the results of the present section in proofs of our main
results:
Proof of Lemma 2.1. Suppose the contrary for some A ⊆ F ⊆ Rd. Then there exists
a Lipschitz function h : Rd → R such that h is nowhere differentiable in A. Applying
Lemma 3.1, we find a collection {Ui}∞
i=1 of boxes with pairwise disjoint interiors such
that (3.12) holds. The conditions of Lemma 3.8 are satisfied for F , A, h and {Ui}∞
i=1.
Further, the hypothesis of Lemma 3.8 part 1 is satisfied for F , A and arbitrary ε > 0.
Let f : Rd → R be the Lipschitz function given by the conclusion of Lemma 3.8, part 1.
By Lemma 3.8, part 2 the differentiability of f and h coincides at all points of A. Thus f
i=1 ∂Ui
i=1 ∂Ui is a non-universal differentiability set and Lemma 3.3
is nowhere differentiable in A. Moreover, f is nowhere differentiable in (F \ A) \S∞
by (3.13). Hence F \S∞
asserts that F is also a non-universal differentiability set.
Proof of Theorem 2.3. Note that ker(F ) is a closed subset of F and F \ ker(F ) is a non-
universal differentiability set by Corollary 3.5. Therefore, we may apply Lemma 2.1 with
A = ker(F ) to deduce that ker(F ) is a universal differentiability set. This proves (i).
For (ii), it only remains to check that ker(ker(F )) = ker(F ). Let x ∈ ker(F ) and ε > 0.
Then we observe that
B(x, ε) ∩ ker(F ) = ker(B(x, ε) ∩ F ),
and the latter set is a universal differentiability set by part (i) and x ∈ ker(F ).
Proof of Theorem 2.4. Suppose that the contrary holds for some universal differentiab-
ility set E ⊆ Rd. This means that there exist relatively closed subsets Ai of E such that
i=1 Ai and each Ai is a non-universal differentiability set. By Lemma 2.1, we may
assume that Ak ⊆ Ak+1 for each k ≥ 1. We will obtain a contradiction, by proving that
E is a non-universal differentiability set.
E =S∞
We begin the construction by using Lemma 3.1 to find a collection of boxes {Ui,1}∞
i=1
with pairwise disjoint interiors such that
E ∩
∞[i=1
Ui,1 = E \ A1
and
diam(Ui,1)
dist(Ui,1, A1)
→ 0 as i → ∞.
Choose a Lipschitz function f1 : Rd → R such that f1 is nowhere differentiable in A1.
11
i=1 of
(3.18)
Suppose n ≥ 1, the Lipschitz function fn : Rd → R and the collections {Ui,l}∞
boxes with pairwise disjoint interiors are defined for l = 1, . . . , n such that
fn is nowhere differentiable in the set An \ n−1[l=1
∞[i=1
Ui,n = E \ An
dist(Ui,n, An)
diam(Ui,n)
and
E ∩
∂Ui,l! ,
∞[i=1
→ 0 as
i → ∞.
(3.19)
Let the Lipschitz function fn+1 : Rd → R be given by the conclusion of Lemma 3.8,
part 1 when we take A = An, F = An+1, h = fn, Ui = Ui,n and ε = 2−(n+1). Then fn+1
i=1 ∂Ui,n. From part 2 of Lemma 3.8, the
differentiability of fn+1 and fn coincides at all points of An. Hence, using (3.18), fn+1
is nowhere differentiable in (An+1 \ An) \S∞
is nowhere differentiable in the set An+1 \ (Sn
Let the collection of boxes {Ui,n+1}∞
i=1 be given by the conclusion of Lemma 3.1 when
we take F = E, A = An+1 and Vi = Ui,n. This ensures the validity of (3.19) with n
replaced by n + 1.
l=1S∞
i=1 ∂Ui,n).
We have defined, for each integer n ≥ 1, a Lipschitz function fn : Rd → R and a
i=1 with pairwise disjoint interiors. In addition to (3.18) and
collection of boxes {Ui,n}∞
(3.19). the construction ensures that the following conditions hold for each n ≥ 2:
kfn − fn−1k∞ ≤ 2−n and Lip(fn) ≤ Lip(fn−1) + 2−n,
fn(y) = fn−1(y) whenever y ∈ Rd \ ∞[i=1
Int(Ui,n−1)! ,
For each index i, there exists an index j such that Ui,n ⊆ Uj,n−1.
This follows from (3.21) and (3.22). By (3.20) the sequence {fn}∞
n=1 converges uniformly
to a Lipschitz function f : Rd → R. Using (3.23), we deduce that the function f satisfies
For the sake of future reference we point out that
fm(y) = fn(y) whenever m ≥ n and y ∈ Rd \ ∞[i=1
y ∈ Rd \ ∞[i=1
k=1S∞
Lemma 3.3, it is sufficient to show that E′ = E \ (S∞
f (y) = fn(y) whenever
Int(Ui,n)! .
Int(Ui,n)! .
We are now ready to prove that E is a non-universal differentiability set.
In view of
i=1 ∂Ui,k) is a non-universal
differentiability set. We will prove that f is nowhere differentiable in E′.
Fix x ∈ E′ and choose n such that x ∈ An. The condition (3.19) ensures that the
conditions of Lemma 3.8 are satisfied for F = E, A = An, h = fn and Ui = Ui,n. Further,
from (3.24), the hypothesis of Lemma 3.8, part 2 is satisfied for the function f : Rd → R.
Therefore, the differentiability of f at x coincides with that of fn at x and, by (3.18),
the proof is complete.
(3.20)
(3.21)
(3.22)
(3.23)
(3.24)
12
4 Differentiability inside sets of positive measure.
In this section we give an application of Theorem 2.4 to differentiability inside sets of
positive Lebesgue measure.
Theorem 4.1. Let d ≥ 2 and suppose P1, P2, . . . , Pd ⊆ R are sets of positive one-
dimensional Lebesgue measure. Then P1 × . . . × Pd contains a compact universal differ-
entiability set with Lebesgue measure zero.
Proof. We may assume that each set Pi is closed. For k = 0, 1, . . . , d, let Πk be the
statement that P1 × P2 × . . . × Pk × Rd−k contains a compact universal differentiability
set Ck with Lebesgue measure zero. The statement Π0 is proved in [3]. Suppose now
that 0 < k ≤ d and that the statement Πk−1 holds. Let us prove the statement Πk and
thus, by induction, Theorem 4.1.
Let {rn}∞
n=1 be a countable dense subset of R and consider the set
Fk =
(Rk−1 × (Pk + rn) × Rd−k).
∞[n=1
Writing Fk,n = Rk−1 × (Pk + rn) × Rd−k for each n, we have Fk =S∞
n=1 Fk,n and each set
Fk,n is closed. Further, observe that pk(Rd \ Fk) is a subset of R with one-dimensional
Lebesgue measure zero. We can write
Ck−1 = (Ck−1 ∩ Fk) ∪ (Ck−1 ∩ (Rd \ Fk)).
Since Rd \ Fk projects to a set of one-dimensional Lebesgue measure zero, we may apply
[6, Lemma 2.1] to conclude that Ck−1 ∩ Fk is a universal differentiability set. Next,
using Theorem 2.4, we deduce that there exists n such that Ck−1 ∩ Fk,n is a universal
differentiability set. Setting
Ck = (Ck−1 + −rnek) ∩ (Rk−1 × Pk × Rd−k),
we observe that
Ck = (Ck−1 ∩ Fk,n) − rnek.
Ck is a universial differentiability set, due to the easily verified fact that any translate of a
universal differentiability set is a universal differentiability set. Note that (Ck−1 + λn) ⊆
P1 × . . . × Pk−1 × Rd−k+1. Hence, Ck ⊆ P1 × . . . × Pk × Rd−k and the proof of statement
Πk is complete.
The above Theorem 4.1 provides a partial answer to the following question of Godefroy:
Does every subset of Rd with positive Lebesgue measure contain a universal differentiab-
ility set of Lebesgue measure zero? This question was asked following a talk of Maleva
at the 2012 conference 'Geometry of Banach spaces' in CIRM, Luminy, and remains
open. Theorem 4.1 also builds on an observation of Doré and Maleva: A consequence of
Lemma 3.5 in [4] is that every set of the form P × Rd−1 ⊆ Rd, where P ⊆ R is a set of
positive Lebesgue measure, contains a Lebesgue null universal differentiability set.
Acknowledgement. The author wishes to thank Olga Maleva for helpful discussions.
13
References
[1] G. Alberti, M. Csörnyei, and D. Preiss. Structure of null sets in the plane and
applications. In European Congress of Mathematics, pages 3 -- 22, 2005.
[2] M. Csörnyei, D. Preiss, and J. Tišer. Lipschitz functions with unexpectedly large
sets of nondifferentiability points. In Abstract and Applied Analysis, volume 2005,
pages 361 -- 373. Hindawi Publishing Corporation, 2005.
[3] M. Doré and O. Maleva. A compact null set containing a differentiability point of
every Lipschitz function. Mathematische Annalen, 351(3):633 -- 663, 2009.
[4] M. Doré and O. Maleva. A universal differentiability set in Banach spaces with
separable dual. Journal of Functional Analysis, 261(6):1674 -- 1710, 2011.
[5] M. Doré and O. Maleva. A compact universal differentiability set with Hausdorff
dimension one. Israel Journal of Mathematics, 191(2):889 -- 900, 2012.
[6] M. Dymond and O. Maleva. Differentiability inside sets with Minkowksi dimension
one.
[7] J. Lindenstrauss, D. Preiss, and J. Tišer. Frechet Differentiability of Lipschitz Func-
tions and Porous Sets in Banach Spaces. Princeton University Press, 2012.
[8] D. Preiss. Differentiability of Lipschitz functions on Banach spaces. Journal of
Functional Analysis, 91(2):312 -- 345, 1990.
[9] D. Preiss and G. Speight. Differentiability of Lipschitz functions in Lebesgue null
sets. Inventiones mathematicae, 199(2):517 -- 559, 2014.
[10] D. Preiss and L. Zajíček. Directional derivatives of Lipschitz functions. Israel Journal
of Mathematics, 125(1):1 -- 27, 2001.
[11] L. Zahorski. Sur l'ensemble des points de non-derivabilite d'une fonction continue.
Bull. Soc. Math. France.
[12] Z. Zajíček. Sets of σ-porosity and sets of σ-porosity (q). Časopis pro pěstováni
matematiky, 1976.
[13] M. Zelen`y and J. Pelant. The structure of the σ-ideal of σ-porous sets. Comment.
Math. Univ. Carolin, 45(1):37 -- 72, 2004.
Michael Dymond
Institut für Mathematik
Universität Innsbruck
Technikerstrasse 13, 6020 Innsbruck, Österreich (Austria)
[email protected]
14
|
1207.4719 | 1 | 1207 | 2012-07-19T16:10:31 | Embeddings of M\"{u}ntz Spaces: Composition Operators | [
"math.FA"
] | Given a strictly increasing sequence $\Lambda=(\lambda_n)$ of nonegative real numbers, with $\sum_{n=1}^\infty \frac{1}{\lambda_n}<\infty$, the M\"untz spaces $M_\Lambda^p$ are defined as the closure in $L^p([0,1])$ of the monomials $x^{\lambda_n}$. We discuss how properties of the embedding $M_\Lambda^2\subset L^2(\mu)$, where $\mu$ is a finite positive Borel measure on the interval $[0,1]$, have immediate consequences for composition operators on $M^2_\Lambda$. We give criteria for composition operators to be bounded, compact, or to belong to the Schatten--von Neumann ideals. | math.FA | math |
Embeddings of Muntz Spaces: Composition
Operators
S.Waleed Noor
n=1
1
λn
< ∞, the Muntz spaces M p
Abstract. Given a strictly increasing sequence Λ = (λn) of nonegative
real numbers, with P∞
Λ are defined as
the closure in Lp([0, 1]) of the monomials xλn. We discuss how proper-
ties of the embedding M 2
Λ ⊂ L2(µ), where µ is a finite positive Borel
measure on the interval [0, 1], have immediate consequences for compo-
sition operators on M 2
Λ. We give criteria for composition operators to be
bounded, compact, or to belong to the Schatten -- von Neumann ideals.
Mathematics Subject Classification (2010). 46E15, 46E20, 46E35.
Keywords. Muntz space, embedding measure, lacunary sequence, Schatten --
von Neumann classes, composition operators.
Introduction
The Muntz -- Szasz Theorem states that, if 0 = λ0 < λ1 < ··· < λn < . . . is an
increasing sequence of nonnegative real numbers, then the linear span of xλn
is dense in C([0, 1]) if and only if P∞
= ∞. When P∞
< ∞, the
closed linear span of the monomials (xλn )∞
n=0 in Lp([0, 1]) for 1 ≤ p < +∞
is a proper subspace of Lp([0, 1]). These spaces, called Muntz spaces and
denoted M p
Λ, exhibit interesting properties that have not been very much
investigated. We refer principally to the monographies [3, 5]; recent results
appear in [1, 2, 8, 4].
1
λn
1
λn
n=1
n=1
In the paper [6], of which this work is a sequel, we investigated various
properties and necessary conditions that allowed us to embed the Hilbert
Muntz space M 2
Λ into the Lebesgue space L2(µ) for some positive measure
µ on [0, 1]. The boundedness, compactness and Schatten ideal properties of
this embedding were studied.
The purpose of this paper is to provide applications of the theory in-
troduced in [6] to composition operators on M 2
Λ. The plan of the paper is
the following. After a section of preliminaries, we show in Section 2 that M p
Λ
.
2
S.Waleed Noor
is not an invariant subspace of composition operators in general. It is then
natural to study composition operators as mapping M p
Λ into Lp([0, 1]). This
is done in the sequel: sufficient conditions for composition operators to be
bounded, compact or belong to Schatten ideals are obtained in Section 3,
and necessary conditions in Section 4.
1. Preliminaries
We denote by m the Lebesgue measure on [0, 1]. Lp(µ) shall be used to denote
the space of Lebesgue integrable functions of order p ∈ [1,∞] with respect
to the measure µ on [0, 1]. We will frequently use Lp to mean Lp(m), and
denote by k·kp and · Lp(µ) the norms in Lp(m) and Lp(µ) respectively.
Let us denote, for a set S of nonnegative real numbers, the subspace
When clear from the context, we shall denote by LS the space Lp
S.
Lp
S = closed span{xt : t ∈ S} ⊂ Lp.
1
Definition 1.1. Let Λ be an increasing sequence of nonnegative real numbers
Λ is defined to be the space Lp
Λ.
λ < ∞. The Muntz space M p
with Pλ∈Λ
tive real numbers withPλ∈Λ
In this paper, Λ shall always denote an increasing sequence of nonnega-
Λ are continuous on
[0, 1) and real analytic in (0,1). A feature of the Muntz monomials (xλ)λ∈Λ
is that they form a minimal system in M p
Λ, which means that for any λ′ ∈ Λ
λ < ∞. The functions in M p
1
dist (xλ′
, LΛ\{λ′}) =
g∈LΛ\{λ′} xλ′
inf
− gLp > 0.
This can easily be extended to show that if Λ′ ⊂ Λ is a finite subset, then
LΛ′ ∩ LΛ\Λ′ = {0}.
(1.1)
The monograph [5] may be consulted for a discussion on the minimality of
Muntz monomials.
We shall need the Clarkson-Erdos Theorem from [5]:
1
λk
< ∞ and inf k(λk+1 − λk) > 0. If f ∈ M p
Λ
Theorem 1.2. Assume that Pk
then there exist bk ∈ R such that
Xk=1
f (x) =
∞
bkxλk
for x ∈ [0, 1),
where the series converges uniformly on compact subsets of [0, 1). Also, for
any ε > 0, there is a constant M > 0 such that
bk.xλkLp ≤ (1 + ε)λkfLp
(1.2)
A sequence Λ is called lacunary if for some γ > 1 we have λn+1/λn ≥ γ
for n ≥ 1. More generally, Λ is called quasilacunary if for some increasing
sequence {nk} of integers with N := supk(nk+1 − nk) < ∞ and some γ > 1
we have λnk+1 /λnk ≥ γ. The main feature of lacunarity is that the monomials
if k ≥ M.
3
λ1/p
n xλn form a basis in each of the spaces M p
(λ1/2
n xλn )n≥1 forms a Riesz basis in M 2
Λ.
Λ. In particular, the sequence
If T : E → F is a bounded operator between Banach spaces, we define
by kTke = inf K kT +Kk the essential norm of an operator, where the infimum
is taken over all compact operators K : E → F . This norm measures how far
an operator is from being compact. In particular, T is compact if and only if
kTke = 0.
The Schatten -- Von Neumann class Sq(H1,H2) is formed by the compact
Hilbert space operators T : H1 → H2 such that T = √T ∗T : H1 → H1 has
a family of eigenvalues {sn(T )}∞
n=1 ∈ ℓq. If we define
sn(T )q!1/q
,
kTkq = ∞
Xn=1
then we obtain a quasinorm for 0 < q < 1 and a norm for q ≥ 1, with
respect to which Sq(H1,H2) is complete. It is immediate that kTkq ≥ kTkq′
for q ≤ q′, hence Sq ⊂ Sq′ .
We now define Λ-embedding measures which were previously studied in
[4] and [6]:
Definition 1.3. A positive measure µ on [0, 1] is called Λp -embedding, if there
is a constant C > 0 such that
for all polynomials g ∈ M p
remove subscript p and use the notation Λ-embedding.
Λ. Whenever p is clear from the context, we will
kgkLp(µ) ≤ C kgkp
(1.3)
Λ: if µ is Λp-embedding, then M p
It follows easily from the definition (see [4]) that a Λp-embedding mea-
sure µ has to satisfy µ(1) = 0. Therefore, as in Remark 2.5 of [4], we may
extend the embedding to all f ∈ M p
Λ ⊂ Lp(µ)
and kfkLp(µ) ≤ C kfkp for all f ∈ M p
Λ. For a Λp-embedding µ we denote by
µ : M p
Λ ֒→ Lp(µ), which is bounded. If 0 < ε < 1,
ip
µ the embedding operator ip
then the interval [1 − ε, 1] will be denoted by Jε.
The next result is proved in [4] for p = 1, but the extension to all p ≥ 1
is straightforward.
Proposition 1.4. Let M p
Λ be a Muntz space, and suppose there exists δ >
0 such that dµJδ = h dmJδ for some bounded measurable function h with
limt→1 h(t) = a. Then ip
µ is bounded and ip
µe = a1/p.
A new class of measures called sublinear measures was introduced in [4].
Λ ֒→ L1(µ)
There they were used to characterize embedding operators iµ : M 1
for the class of quasilacunary sequences Λ.
Definition 1.5. A measure µ is called sublinear if there is a constant C >
0 such that for any 0 < ε < 1 we have µ(Jε) ≤ Cε. The smallest such
C will be denoted by kµkS. The measure µ is called vanishing sublinear if
ε = 0. Furthermore, a measure µ is called α-sublinear if µ(Jε) ≤
limε→0
Cεα for some α > 1.
µ(Jε)
4
S.Waleed Noor
The main embedding results in [6] are contained in the next two theo-
rems:
Theorem 1.6. Let Λ be lacunary and µ a positive measure on [0, 1]. Then
(i) i2
(ii) i2
µ is bounded if µ is sublinear.
µ is compact if µ is vanishing sublinear.
The above results are shown in [6] to be true, after an interpolation
argument, for all embeddings ip
µ for 1 ≤ p ≤ 2.
In [6], we also investigated conditions for measures that enabled the
embedding i2
µ to belong to Sq. We shall need the main results therein:
µ ∈ Sq(M 2
Theorem 1.7. Let µ be a positive measure on [0, 1]. Then i2
for all q > 0 if either of the following is true
(i) µ has compact support in [0, 1),
(ii) Λ is quasilacunary and µ is α-sublinear.
Λ, L2(µ))
Our goal is to apply these embedding results to composition operators.
Recall that the pullback of a measure ν by φ is the measure φ∗ν on [0, 1]
defined by
φ∗ν(E) = ν(φ−1(E))
for any Borel set E. If g is a positive measurable function, then the formula
Z 1
0
g(φ(x))dx =Z[0,1]
g d(φ∗m)
is easily checked on characteristic functions, hence the usual argument ex-
tends it to all positive Borel functions on [0, 1]. In particular, if we de-
fine µ = φ∗m and choose g = fp for some f ∈ Lp(µ), then the map
J : Lp(µ) −→ Lp defined by J(f ) = f ◦ φ is an isometry.
Let φ be a Borel function on [0, 1] such that φ([0, 1]) ⊂ [0, 1]. The
composition operator Cφ is defined as
Cφ(g) = g ◦ φ
for all polynomials g ∈ M p
to all f ∈ M p
composition operators.
µ, we can extend Cφ = J ◦ ip
µ
Λ. Since J is an isometry, we obtain the following results for
Λ. Just as we did for ip
Lemma 1.8. Define the measure µ = φ∗m. Then
(i) Cφ is bounded from M p
(ii) Cφ is compact from M p
(iii) Cφ ∈ Sq(M 2
Λ, L2) if and only if i2
µ ∈ Sq(M 2
Λ to Lp if and only if µ is a Λp-embedding measure.
Λ to Lp if and only if ip
µ is compact.
Λ, L2(µ)).
5
2. Muntz Spaces are not Invariant to Most Composition
Operators
It has already appeared above that we study composition operators defined
on M p
Λ, but whose range space is Lp. The reason is that Muntz spaces are
usually not invariant with respect to composition. This has already been
noticed by Al Alam [2], in the case of Muntz space M ∞
Λ , i.e. the closure of
the span of monomials xλn in L∞, and operators Cφ with continuous φ. The
following result was proved therein.
Proposition 2.1. Let Λ = (λk)k ⊂ N and Pk
Λ if φ = αxm + βxn with α, β 6= 0 and m, n ∈ N.
Λ if φ is a polynomial with positive coefficients and more
< ∞. Then
1
λk
(i) CφM ∞
(ii) CφM ∞
than one term.
Λ * M ∞
Λ * M ∞
In this section we will significantly extend these results to other values of
p ≥ 1 and functions φ. We prove in Theorem 2.5 that CφM p
Λ whenever
φ is a function of the form c1xs1 + . . . + clxsl with ci ∈ R and si ∈ R+. These
functions will be called real-exponent polynomials. This generalizes Proposi-
tion 2.1 and Λ may not even satisfy the gap condition inf k(λk+1 − λk) > 0.
If we assume the gap condition, then Theorem 2.8 generalizes Proposition
2.1(i) for arbitrary Λ ⊂ R+.
Λ * M p
We start with a result of A. Schinzel [7]:
Lemma 2.2. If φ is a polynomial with at least two terms and λ ∈ N, then φλ
has at least λ + 1 terms.
The next result is an analog of Lemma 2.2 for real-exponent polynomi-
als.
Lemma 2.3. If φ is a real-exponent polynomial with at least two terms and
λ ∈ N, then φλ has at least λ + 1 terms.
Proof. Let φ(x) = c1xs1 + . . . + clxsl with ci ∈ R\{0} and si ∈ R+. Consid-
ering R as a vector space over the rationals Q, choose a basis r1, . . . , rτ > 0
for the space spanned by s1, . . . , sl where τ ≤ l. Therefore
τ
aijrj
for i = 1, . . . , l
si =
Xj=1
where aij ∈ Q. We may assume that aij ∈ Z by adjusting the rj suitably. We
note that for any positive real number N , φλ has the same number of terms
as (xN φ)λ. So by choosing N = b1r1 + . . . + bτ rτ with integers bj > aij for
i = 1, . . . , l and j = 1, . . . , τ , we may also assume that each aij rj > 0 hence
aij ∈ N. We then obtain
Xi=1
ci(xr1 )ai1 . . . (xrτ )aiτ .
Xi=1
cixsi =
φ(x) =
l
l
6
S.Waleed Noor
We define a polynomial ψ in τ variables by
ψ(Y1, . . . , Yτ ) =
ciY ai1
1
. . . Y aiτ
τ
.
l
Xi=1
Define Φ to be the collection of monomial terms in φλ after reduction and can-
celation, and Ψ similarly for ψλ. Hence our goal is to prove that cardΦ ≥ λ+ 1.
Since both φ and ψ each have l distinct monomial terms, the total number
of possible products while computing φλ or ψλ is lλ.
1
1
. . . Y m′
1 . . . (xrτ )m′
We claim that whenever two such products p(x) = k.(xr1 )m1 . . . (xrτ )mτ
and q(x) = k′.(xr1 )m′
τ reduce (respectively cancel) in φλ, the cor-
responding products pψ(Y1, . . . , Yτ ) = k.Y m1
and qψ(Y1, . . . , Yτ ) =
k′.Y m′
j ∈ N. Indeed, it
is obvious that p and q combine (resp. cancel) if and only if m1r1+. . .+mτ rτ =
m′
τ rτ . Since r1, . . . , rτ are linearly independent over Q, this is
possible if and only if mj = m′
j for j = 1, . . . , τ . And this is equivalent to the
reducing (resp. cancelling) of pψ and qψ. This proves that cardΦ = cardΨ.
also reduce (resp. cancel) in ψλ, where mj, m′
1r1 + . . . + m′
1
. . . Y mτ
τ
τ
τ
Note that ψ has at least two terms because φ has at least two terms. This
implies that for some 1 ≤ j′ ≤ τ , ψ as a polynomial in Yj′ has at least two
terms. Applying Lemma 2.2 to ψ′(Yj′ ) := ψ(1, . . . , Yj′ , . . . , 1) =Pl
,
we see that (ψ′)λ has at least λ + 1 terms. Therefore ψλ has at least λ + 1
terms and cardΨ ≥ λ + 1. Therefore cardΦ ≥ λ + 1.
(cid:3)
aij′
i=1 ciY
j′
The next lemma is a consequence of formula (1.1).
1
λk
Lemma 2.4. Let Λ = (λk)k and Pk
c1xs1 + . . . + clxsl belongs to M p
Proof. Given c1xs1 + . . . + clxsl ∈ M p
Λ = LΛ, suppose on the contrary that
some subset Λ′ = {sk1, . . . , skm} ⊂ {s1, . . . , sl} does not belong to Λ and
{s1, . . . , sl}\Λ′ ⊂ Λ. Then
< ∞. If a real-exponent polynomial
Λ , then s1, . . . , sl ∈ Λ.
p(x) = (c1xs1 + . . . + clxsl ) − (ck1 xsk1 + . . . + ckm xskm ) ∈ LΛ.
This implies that ck1 xsk1 +. . .+ckmxskm = c1xs1 +. . .+clxsl−p(x) ∈ LΛ′∩LΛ.
But LΛ′ ∩ LΛ = {0} by (1.1), a contradiction.
Theorem 2.5. Suppose Λ = (λk)k ⊂ N with Pk
exponent polynomial with more than one term, then CφM p
Proof. Let φ(x) = c1xs1 + . . . + clxsl with ci ∈ R\{0} and si ∈ R+. Then for
any λ ∈ Λ, we get Cφ(xλ) = φλ which has at least λ + 1 terms by Lemma
2.3. We may assume that these λ + 1 terms are nonzero multiples of
< ∞. If φ is a real-
Λ * M p
Λ.
1
λk
(cid:3)
xs1λ, xt1 , . . . , xtλ−1 , xslλ where s1λ < t1 < . . . < tλ−1 < slλ.
Suppose that CφM p
Λ, then Theorem 2.4 gives us s1λ, t1, . . . , tλ−1, slλ ∈
Λ. We construct a subsequence (λkj )j of Λ as follows: Let λk1 = λ1 and
Λ ⊂ M p
inductively choose λkj such that s1λkj > slλkj−1 for j ≥ 2. Then the sequence
∞
7
Λ∗ :=
[j=1
{s1λkj , t1, . . . , tλkj −1, slλkj}
is increasing and has distinct elements; moreover, Λ∗ ⊂ Λ. So
∞
Xk=1
1
λk ≥ Xs∈Λ∗
1
s ≥
∞
Xj=1
λkj +1
Xi=1
1
slλkj ≥
Λ * M p
Λ.
1
sl
∞
Xj=1
= ∞
(cid:3)
and hence the contradiction implies CφM p
Λ ⊂ M p
Λ
Λ → M p
φ (xλ) = Cm−1
φ
Λ is a bounded operator.
Corollary 2.6. Let Λ ⊂ N and φ be a real-exponent polynomial. Then the
following are equivalent:
(i) CφM p
(ii) φ(x) = αxη and Λ = Λ.{1, η, η2, . . .} for some 0 ≤ α ≤ 1 and η ∈ R+
(iii) Cφ : M p
Proof. (i) ⇒ (ii). Theorem 2.5 implies that φ(x) = αxη for η ∈ R+ and
(αλxλη) = . . . =
0 ≤ α ≤ 1 because φ([0, 1]) ⊂ [0, 1]. Then Cm
Kxληm
∈ M p
Λ for any λ ∈ Λ, m ∈ N and some constant K. Hence ληm ∈ Λ
for all λ ∈ Λ and m ∈ N by Lemma 2.4. Therefore Λ = ∪λ∈Λλ.{1, η, η2, . . .} =
Λ.{1, η, η2, . . .}.
(ii) ⇒ (iii). Suppose φ(x) = αxη with 0 ≤ α ≤ 1 and η ∈ R+. If α < 1,
then µ = φ∗m is supported on [0, α] and dµJ1−α = 0. Hence ip
µe = 0 by
µ is compact. For α = 1, the measure µ = φ∗m
Proposition 1.4 and Cφ = J ◦ ip
satisfies
ZJδ
for any continuous f and 0 < δ < 1. Therefore dµJδ = h dmJδ where
η −1 is bounded on Jδ, and hence Cφ is bounded by Proposition
h(x) = η−1x
1.4. Moreover, for any λ ∈ Λ we see that Cφxλ = αλxλη ∈ M p
Λ. Hence by the
density of linear span of monomials xλ in M p
Λ and continuity of Cφ, we get
CφM p
(cid:3)
Λ ⊂ M p
It is easy to see that Theorem 2.5 and Corollary 2.6 can be extended to
the case when Λ * N, but contains a subsequence of integers. To go beyond
this case, we need some preparation about real-exponent power series.
Λ. The last part (iii) ⇒ (i) is trivial.
f · (φ−1)′dm =Z 1
f dµ =Zφ−1(Jδ )
f ◦ φ dm =ZJδ
f (x) η−1x
1
η −1dx
1−δ
1
Lemma 2.7. Suppose f (x) =Pk akxsk is a series such that (sk)k ⊂ R+ is the
finite union of sequences that satisfy the gap condition. Then f is uniformly
convergent on some interval [0, ρ] if L := lim supk ak1/sk < ∞. Furthermore,
if f ≡ 0 on [0, ρ0] for ρ0 ≤ ρ then ak = 0 for all k.
Proof. It is sufficient to prove the first part for the case when (sk)k itself
satisfies the gap condition; in the general case, we can write f as a finite sum
of uniformly convergent series.
8
S.Waleed Noor
Since akxsk1/sk = ak1/skx, we get lim supk akxsk1/sk < 1 if and
only if x < L−1 (taking L−1 = ∞ if L = 0). So, for Lx < 1, we get
lim supk akxsk1/sk < r < 1 for some r and hence there exists a positive
integer N such that akxsk1/sk < r for k ≥ N . Therefore
∞
Xk≥N
akxsk ≤
rsk < ∞
∞
Xk≥N
where the convergence follows from the ratio test and the gap condition
because
lim
k→∞
rsk+1
rsk
= lim
k→∞
rsk+1−sk ≤ rinf k(sk+1−sk) < 1.
So f (x) converges absolutely for Lx < 1, and in particular converges uni-
formly on [0, ρ] for some ρ > 0.
For the second part, suppose on the contrary that a1 is the first non-zero
coefficient. We see that
f (x) =Xk≥1
akxsk = a1xs1 (1 +Xk>1
ak
a1
xsk−s1 )
where (sk − s1)k is again a union of finitely many series satisfying the gap
condition and
lim sup
1
sk −s1
(cid:12)(cid:12)(cid:12)(cid:12)
ak
a1(cid:12)(cid:12)(cid:12)(cid:12)
≤ lim sup
(ak
1
sk )
sk
sk −s1 . lim sup
1
sk −s1
(cid:18) 1
a1(cid:19)
= L < ∞
k
k
ak
a1
k
xsk−s1 converges uniformly on some interval [0, ρ1].
hence g(x) = 1+Pk>1
So f (x) = a1xs1 g(x) = 0 on [0, r], where r = min{ρ0, ρ1}. Therefore g = 0 on
(0, r] and hence on [0, r] by continuity. A contradiction, since g(0) = 1. (cid:3)
Theorem 2.8. Suppose Λ ⊂ R+ with Pk
< ∞ satisfies the gap condition
inf k(λk+1 − λk) > 0. If φ = αxζ1 + βxζ2 with α, β 6= 0 and ζ1 < ζ2 ∈ R+,
then CφM p
Proof. If Λ ⊂ N, then Theorem 2.5 proves the result. So we assume Λ * N,
Λ ⊂ M p
hence there exists λ ∈ Λ that is not an integer. Suppose that CφM p
Λ;
then
Λ * M p
Λ.
1
λk
Cφ(xλ) = (αxζ1 + βxζ2 )λ = αλxλζ1 (1 +
Hence by the binomial series we can represent Cφ(xλ) as
Cφ(xλ)(t) = αλtλζ1
∞
Xk=0
aktk(ζ2−ζ1) = αλ
β
α
xζ2−ζ1 )λ ∈ M p
Λ.
aktλζ1+k(ζ2−ζ1)
∞
Xk=0
where the series converges for t < α
β
ζ2−ζ1 , in particular on [0, η] for some
η < 1. The sequence of exponents (λζ1 + k(ζ2 − ζ1))k clearly satisfies the gap
condition, while the coefficients
1
ak =(cid:18) β
α(cid:19)k λ(λ − 1)(λ − 2) . . . (λ − k + 1)
k!
satisfy
L1 := lim sup
k→∞ ak1/λζ1+k(ζ2−ζ1) < ∞.
Similarly, by Theorem 1.2 there exists a sequence of scalars bk ∈ R such that
∞
9
Cφ(xλ)(t) =
bktλk
Xk=1
and the series converges uniformly on compact subsets of [0, 1). By (1.2), the
coefficients (bk)k satisfy
L2 := lim sup
] < ∞.
Since both series representations coincide on [0, η], the series defined by
[(1 + ε)(2λk + 1)1/2λkf1/λk
k→∞ bk1/λk ≤ lim sup
k→∞
L2
∞
∞
f (t) =
bktλk − αλ
Xk=1
Xk=0
aktλζ1+k(ζ2−ζ1) =Xk
γktsk
vanishes on [0, η]. Since (sk)k is the union of two series satisfying the gap
condition and lim supk γk1/sk ≤ L1 + L2 < ∞, by Lemma 2.7 we get γk = 0
for all k. Since λ is not an integer, all the ak are non-zero; this implies that
λζ1 + k(ζ2 − ζ1) ∈ Λ for all k. This contradicts the fact that Pk
< ∞ and
hence CφM p
Λ * M p
Λ.
1
λk
(cid:3)
3. Composition Operators on M 2
Λ: direct results
The next result is essentially contained in the work of Chalendar, Fricain and
Timotin [4]:
Proposition 3.1. Suppose the Borel function φ : [0, 1] −→ [0, 1] satisfies the
following:
(a) φ−1(1) = {x1, . . . , xk} is finite.
(b) There exists ǫ > 0 such that, for each i = 1, ..., k, φ is continuous on
(xi − ǫ, xi + ǫ), φ ∈ C1((xi − ǫ, xi)) and φ ∈ C1((xi, xi + ǫ)).
(c) φ′
(φ′
may be infinite).
(d) There exists α < 1 such that, if x /∈ ∪k
+(x) denote the left and right derivatives at x respectively, which
i=1(xi − ǫ, xi + ǫ), then φ(x) < α.
+(xi) < 0 for all i = 1, . . . , k.
−(xi) > 0 and φ′
−(x) and φ′
Then Cφ : M 2
i=1 L(xi), where
1
1
1
Λ −→ L2 is bounded and Cφe =Pk
if xi ∈ (0, 1),
L(xi) =
if xi = 1,
if xi = 0.
−(xi) +
φ′
φ′
−(xi)
−(xi) = ∞ and φ′
+(xi)
+(xi)
φ′
φ′
1
In particular, if φ′
is compact.
+(xi) = −∞ for all i = 1, . . . , k, then Cφ
We intend to go beyond the regularity assumptions in Proposition 3.1.
10
S.Waleed Noor
Definition 3.2. If φ : [0, 1] → [0, 1] is a Borel function and α = ess sup[0,1] φ,
then a point x ∈ [0, 1] is an essential point of maximum for φ if ess supE φ = α
for every neighborhood E of x. Denote by Mφ the set of all essential points
of maxima of φ, and by Vε the neighborhood of Mφ defined for each ε > 0
by
Vε = {x ∈ [0, 1] : dist(x, Mφ) < ε}.
Lemma 3.3. The following statements are true:
(i) Mφ is non-empty and closed,
(ii) ess sup φ[0,1]\Vε < α for all ε > 0,
(iii) for every ε > 0 there exists a δ0 > 0 such that φ−1([α − δ, α]) ⊂ Vε
almost everywhere whenever 0 < δ < δ0.
Proof. (i). If Mφ were empty, then every point x ∈ [0, 1] would have a neigh-
borhood Nx such that ess supNx φ < α and all such Nx would cover [0, 1].
Choosing a finite subcover so that ∪m
φ = max
k=1Nxk = [0, 1], we see that
k {ess sup
Nxk
ess sup
[0,1]
φ} < α.
The contradiction yields Mφ 6= ∅. To prove that Mφ is closed, consider the
set S := ∪x∈[0,1]\MφNx, where Nx again represents a neighborhood of x on
which ess supNx φ < α. So clearly S is open, and S ∩ Mφ = ∅ since otherwise
some Nx′ for x′ ∈ [0, 1]\Mφ would contain an essential point of maximum.
Hence S = [0, 1]\Mφ and Mφ is closed.
For (ii), suppose that ess sup φ[0,1]\Vε′ = α for some ε′ > 0. Then the
argument in the proof of (i) applied to the compact set [0, 1]\Vε′, shows that
it contains an essential point of maximum.
Finally for (iii), it follows from (ii) that for every ε > 0 there exists a
δ0 > 0 such that ess sup φ[0,1]\Vε < α − δ0 < α and hence
φ−1([α − δ, α]) = {x ∈ [0, 1] : α − δ ≤ φ(x) ≤ α} ⊂ Vε
except possibly for a subset of measure 0, whenever 0 < δ < δ0.
(cid:3)
We recall that the left and right derivatives of φ at the point y are
defined as
Di
−(y) = lim inf
t→y−
Di
+(y) = lim inf
t→y+
Ds
−(y) = lim sup
t→y−
Ds
+(y) = lim sup
t→y+
φ(y) − φ(t)
y − t
φ(y) − φ(t)
y − t
φ(y) − φ(t)
y − t
φ(y) − φ(t)
y − t
respectively.
11
Suppose φ : [0, 1] → [0, 1] is a Borel function such that α = ess sup[0,1] φ <
1. Then it is easy to show that the measure defined by µ = φ∗m has support
in [0, α]. In fact
µ((α, 1]) =Z(α,1]
d(φ∗m) =Zφ−1((α,1])
dm = m(φ−1(α, 1]) = 0.
Hence in this case i2
1.8, so from here onwards we assume that α = ess sup[0,1] φ = 1.
µ ∈ Sq by Theorem 1.7(i). Therefore Cφ ∈ Sq by Lemma
Since changing the values of φ on a set of measure zero does not effect
µ = φ∗m, whenever m(Mφ) = 0, one may take φ ≡ 1 on Mφ. This will be
assumed in the rest of the paper.
Lemma 3.4. Suppose φ is a Borel function with Mφ = {x1, . . . , xk} and
µ = φ∗m. If for some s ≥ 1 there exists an ε > 0 and a constant c > 0
such that
x − xi ≤ cφ(x) − 1s whenever x − xi < ε
for all i = 1, . . . , k, then there exists a δ0 > 0 such that µ(Jδ) ≤ 2kcδs
whenever 0 < δ < δ0.
Proof. By Lemma 3.3(iii), there exists a δ0 > 0 such that φ−1(Jδ) ⊂ Vε
almost everywhere whenever 0 < δ < δ0. Since supφ−1(Jδ) φ(x) − 1 ≤ δ, we
get
k
k
m(φ−1(Jδ)) ≤
m(φ−1(Jδ) ∩ {dist(x, xi) < ε}) ≤ 2
φ−1(Jδ )∩{x−xi<ε}x − xi
sup
Xi=1
Xi=1
≤ 2
k
Xi=1
cφ(x) − 1s ≤ 2kcδs.
sup
φ−1(Jδ)∩{x−xi<ε}
Therefore we get
µ(Jδ) =ZJδ
dµ =ZJδ
d(φ∗m) =Zφ−1(Jδ)
dm = m(φ−1(Jδ)) ≤ 2kcδs
whenever 0 < δ < δ0.
(cid:3)
We arrive at the main theorem that gives necessary conditions for com-
position operators on M 2
Λ to be bounded, compact or in Sq.
Theorem 3.5. Let Λ be lacunary and Mφ = {x1, . . . , xk}.
(i) If Di
(ii) If Di
+ = −∞ on Mφ, then Cφ : M 2
(iii) If for some ε > 0, β > 1 and constant c we have
− > 0 and Ds
− = +∞ and Ds
+ < 0 on Mφ, then Cφ : M 2
Λ → L2 is bounded.
Λ → L2 is compact.
x − xi ≤ cφ(x) − 1β ∀ x − xi < ε
(3.1)
for i = 1, . . . , k, then Cφ ∈ Sq(M 2
Λ, L2) ∀ q > 0.
12
S.Waleed Noor
Proof. (i) The hypothesis about the derivatives implies that for some constant
M > 0 there exists an ε > 0 such that
φ(x) − 1
x − xi ≥ M > 0 ⇐⇒ x − xi ≤ M −1φ(x) − 1
whenever x − xi < ε for all i = 1, . . . , k. Hence by Lemma 3.4, we get
µ(Jδ) ≤ 2kM −1δ for 0 < δ < δ0 . Therefore µ is sublinear and i2
µ is bounded
Λ → L2 is bounded.
by Theorem 1.6 (i). So Lemma 1.8 implies that Cφ : M 2
(ii) By our hypothesis, for any M > 0 there exists an ε > 0 such that
φ(x) − 1
x − xi ≥ M ⇐⇒ x − xi ≤ M −1φ(x) − 1
whenever x−xi < ε. And for every such ε > 0 there exists a δ0 > 0 such that
µ(Jδ) ≤ 2kM −1δ whenever 0 < δ < δ0 by Lemma 3.4. Therefore µ(Jδ )
δ → 0
as δ → 0. So the measure µ defined above is a vanishing sublinear measure
hence i2
µ : M 2
Λ → L2(µ) is compact by Theorem 1.6 (ii), and so is Cφ = J ◦ i2
µ.
(iii) Applying Lemma 3.4 directly to condition (3.1), we get µ(Jδ) ≤
Λ, L2(µ))
(cid:3)
2kcδβ whenever 0 < δ < δ0. Hence by Theorem 1.7(ii), iµ ∈ Sq(M 2
for all q > 0. So Cφ ∈ Sq(M 2
Remark 3.6. If ψ ∈ L∞ then these results still hold true for the weighted
composition operator Mψ ◦ Cφ where Mψ is the multiplication operator with
symbol ψ, which is a bounded operator on L2.
Λ, L2) for all q > 0.
4. Composition Operators on M 2
Λ: Inverse results
We conclude by presenting some results that serve as converses to the bound-
edness and compactness theorems given above for composition operators on
M 2
Λ. We shall need the following two lemmas.
µ(Jδ)
µ is bounded, then lim inf δ→0
µ is compact, then lim inf δ→0
Lemma 4.1. Let µ be a positive measure on [0, 1]. Then the following hold:
(i) If i2
(ii) If i2
Proof. (i) Suppose µ is Λ2-embedding. Since limn→∞(1 − 1
)λn = 1
exists an integer N such that, for all n ≥ N and for all x ∈ [1 − 1
have xλn ≥ 1
δ < ∞
δ = 0 .
µ(Jδ )
e , there
, 1], we
λn
λn
3 . It follows that for all n ≥ N
x2λn dµ ≤ i2
32 µ(J1/λn ) ≤ZJ1/λn
1
µ2Z 1
0
x2λn dx = i2
µ2
2λn + 1
.
Therefore for all n ≥ N , we have
µ2
32i2
λn ⇐⇒
δ < ∞.
This implies that lim inf δ→0
µ(J1/λn ) ≤
µ(Jδ)
µ(J1/λn )
1/λn ≤ 9i2
µ2.
(ii) Choosing fn(x) = λ1/2
n xλn , we see that
hfn , xλki =Z[0,1]
λ1/2
n xλn+λk dx =
λ1/2
n
λn + λk + 1 −→ 0
13
µ is compact, this implies that (i2
as n → ∞ for all k ∈ N. Noting that fnL2 is bounded and the linear span
Λ, it follows that fn → 0 weakly in M 2
of the sequence (xλk )k is dense in M 2
Λ,
as n → ∞. If i2
µ fn)n converges strongly to
0 in L2(µ) and hence fnL2(µ) → 0 as n → 0. Therefore
λnx2λn dµ ≥ (1 −
fn2
µ(J1/λn )
λnx2λn dµ ≥ZJ1/λn
L2(µ) =Z[0,1]
1
λn
1/λn
)2λn
.
Since (1 − 1
λn
)2λn → e−2 as n → ∞, we get
µ(J1/λn )
1/λn −→ 0 as n → ∞
and the result follows.
(cid:3)
The next lemma might be compared to Lemma 3.4.
Lemma 4.2. Suppose φ : [0, 1] → [0, 1] is a Borel function and µ = φ∗m. If
for some x0 ∈ [0, 1] with φ(x0) = 1 and η > 0, there exists an ε > 0 such that
x0 − x >
1
η
(1 − φ(x)) whenever 0 < x0 − x < ε,
η for 0 < δ < η ε.
then µ(Jδ) ≥ δ
Proof. Since there exists an ε > 0 such that for 0 < x0 − x < ε, we have
1 − φ(x)
x0 − x
< η ⇐⇒ 1 − φ(x) < η(x0 − x).
η then 1−φ(x) < η δ
Then suppose 0 < δ < δ0 = η ε. If 0 < x0−x < δ
δ which implies φ(x) > 1 − δ. So φ−1(Jδ) contains the interval (x0 − δ
of Lebesgue measure δ
ε = δ
η =
η , x0)
δ0
η . Therefore
m(φ−1(Jδ)) ≥
δ
η ⇒ µ(Jδ) ≥
δ
η ⇒
µ(Jδ)
δ ≥
1
η
.
(cid:3)
For the partial converses to parts (i) and (ii) of Theorem 3.5, we need
neither lacunarity nor any assumption on Mφ:
Theorem 4.3. Suppose φ : [0, 1] → [0, 1] is a Borel function, and φ(x0) = 1
for some x0 ∈ [0, 1].
(i) If Cφ is bounded, then Ds
(ii) If Cφ is compact, then Ds
−(x0) > 0 and Di
−(x0) = +∞ and Di
+(x0) = −∞.
+(x0) < 0.
14
S.Waleed Noor
Proof. (i) Suppose on the contrary that either Ds
+(x0) = 0. We
shall deduce a contradiction for one of these cases since both are analogous.
So suppose Ds
−(x0) = 0 or Di
−(x0) = 0, that is
lim
x→x0−
1 − φ(x)
x0 − x
= 0.
< η.
1 − φ(x)
x0 − x
δ ≥ 1
For each η > 0 there exists an ε > 0 such that for 0 < x0 − x < ε, we have
Therefore by Lemma 4.2 we get µ(Jδ )
δ → +∞
as δ → 0. Since Cφ is bounded we get that µ is Λ2-embedding by Lemma 1.8.
This leads to a contradiction since Lemma 4.1 gives lim inf δ→0
η whenever δ < η ε. So µ(Jδ )
µ(Jδ )
For (ii), suppose to the contrary that either Ds
+(x0) >
−∞ for some x0 ∈ Mφ. Again due to similarities we shall deal with one case.
So suppose Ds
−(x0) < +∞ or Di
δ < ∞.
−(x0) < ∞, that is
lim
x→x0−
1 − φ(x)
x0 − x
< ∞.
So there exists a ζ > 0 and an ε > 0 such that for 0 < x0 − x < ε, we have
1 − φ(x)
x0 − x
Therefore by Lemma 4.2 we get µ(Jδ )
Lemma 4.1 because i2
< ζ.
µ is compact by Lemma 1.8.
δ ≥ 1
ζ for δ < ζ ε. This contradicts
(cid:3)
Corollary 4.4. Suppose φ is a polynomial with φ−1(1) non-empty. Then Cφ
is not compact, and if it is bounded then φ−1(1) ⊂ {0, 1}.
Proof. If Cφ is bounded and some x0 ∈ φ−1(1) is an interior point of [0, 1],
then clearly x0 must be a local maximum and hence φ′(x0) = 0. This con-
tradicts Theorem 4.3 (i) and hence φ−1(1) ⊂ {0, 1}. Similarly, by part (ii) of
the theorem we get the conclusion that Cφ can never be compact because φ
is differentiable everywhere.
(cid:3)
Acknowledgments: The author wishes to thank his supervisor Professor
Dan Timotin for the many interesting ideas he shared and discussions we
had.
References
1. I.Al Alam. G´eometrie des espaces de Muntz et op´erateurs de composition `a
poids. PhD thesis. Universit´e Lille 1, 2008.
2. I.Al Alam. Essential norms of weighted composition operators on Muntz spaces.
J. Math. Anal. Appl., 358(2), 2009.
3. P.Borwein and T.Erdelyi. Polynomials and polynomial inequalities, volume 161
of Graduate Texts in Mathematics. Springer-Verlag, New York, 1995.
15
4. I.Chalendar, E.Fricain, and Dan Timotin. Embedding Theorems for Muntz
Spaces, Annales de l'Institut Fourier, To Appear.
http://math.univ-lyon1.fr/∼fricain/arxiv-muntz.pdf
5. V.I.Gurariy and W.Lusky. Geometry of Muntz spaces and related questions,
volume 1870 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 2005
6. S. W. Noor and D. Timotin. Embeddings of Muntz Spaces: The Hilbertian Case,
Proc. Amer. Math. Soc., To appear 2012. http://arxiv.org/abs/1110.5422
7. A. Schinzel. On the number of terms of a power of a polynomial, Acta Arith.
49 (1987), no. 1, 5570.
8. A. Spalsbury. Perturbations in Muntz's theorem. J. Approx. Theory, 150(1):48 --
68, 2008.
S.Waleed Noor
Abdus Salam School of Mathematical Sciences
New Muslim Town, Lahore, Pakistan
e-mail: waleed [email protected]
|
1502.07863 | 1 | 1502 | 2015-02-27T10:57:17 | The spectrum of Volterra operators on weighted spaces of entire functions | [
"math.FA"
] | We investigate the spectrum of the Volterra operator $V_g$ with symbol an entire function $g$, when it acts on weighted Banach spaces $H_v^{\infty}(\mathbb{C})$ of entire functions with sup-norms and when it acts on H\"ormander algebras $A_p$ or $A^0_p$. | math.FA | math |
The spectrum of Volterra operators on weighted spaces of
entire functions
Jos´e Bonet
Abstract
We investigate the spectrum of the Volterra operator Vg with symbol an entire
v (C) of entire functions with
function g, when it acts on weighted Banach spaces H ∞
sup-norms and when it acts on Hormander algebras Ap or A0
p.
1
Introduction, notation and preliminaries
The aim of this note is to investigate the spectrum of the Volterra operator when it acts
continuously on a weighted Banach space of entire functions H ∞
v (C). Our main result is
Theorem 2.3. The characterizations of continuous Volterra operators on H ∞
v (C) obtained
by Taskinen and the author in [13] play an important role. The present research originated
with a question of A. Aleman in a meeting celebrated in Granada in February 2015 in
which the results in [13] were presented. In Section 3 we study the spectrum of Volterra
operators acting on a (DFN) Hormander algebra Ap in Theorem 3.4, and on a Fr´echet
Hormander algebra A0
p in Theorem 3.5.
In what follows H(C) denotes the space of entire functions. The space H(C) is a
Fr´echet space endowed with the compact open topology. The differentiation operator
0 f (ζ)dζ and the multiplication operator
Df (z) = f ′(z), the integration operator Jf (z) = R z
Mh(f ) = hf, h ∈ H(C), are continuous on H(C).
Given a non-constant entire function g ∈ H(C) with g(0) = 0, the Volterra operator
Vg with symbol g is defined on H(C) by
Vg(f )(z) := Z z
0
f (ζ)g′(ζ)dζ
(z ∈ C).
For g(z) = z this reduces to the integration operator. Clearly Vg defines a continuous
operator on H(C). The Volterra operator for holomorphic functions on the unit disc
was introduced by Pommerenke [28] and he proved that Vg is bounded on the Hardy
space H 2, if and only if g ∈ BM OA. Aleman and Siskakis [3] extended this result for
H p, 1 ≤ p < ∞, and they considered later in [4] the case of weighted Bergman spaces; see
also [26]. We refer the reader to the memoir by Pel´aez and Rattya [27] and the references
therein. Volterra operators on weighted Banach spaces of holomorphic functions on the
disc of type H ∞ have been investigated recently in [5]. Constantin started in [15] the study
of the Volterra operator on spaces of entire functions. She characterized the continuity
of Vg on the classical Fock spaces and investigated its spectrum. Constantin and Pel´aez
2010 Mathematics Subject Classification. Primary: 47G10, secondary: 30D15, 30D20, 46E15, 47B07,
47B37, 47B38.
Key words and phrases. Spectrum; integral operator; Volterra operator; entire functions; growth
conditions; weighted Banach spaces of entire functions; Hormander algebras
1
[16] characterize the entire functions g ∈ H(C) such that Vg is bounded or compact on a
large class of Fock spaces induced by smooth radial weights. See also [13]. Aleman and
Constantin [1] and Aleman and Pel´aez [2] investigate the spectra of Volterra operators on
several spaces of of holomorphic functions on the disc.
A weight v is a continuous function v : [0, ∞[→]0, ∞[, which is non-increasing on [0, ∞[
and satisfies limr→∞ rmv(r) = 0 for each m ∈ N. We extend v to C by v(z) := v(z). For
such a weight, the weighted Banach space of entire functions is defined by
H ∞
v (C) := {f ∈ H(C) kf kv := supz∈C v(z)f (z) < ∞},
and it is endowed with the weighted sup norm k · kv. Changing the value of v on a compact
interval does not change the space and gives an equivalent norm. Spaces of this type
appear in the study of growth conditions of analytic functions and have been investigated
in various articles, see e.g. [10, 11, 22] and the references therein.
For an entire function f ∈ H(C), we denote M (f, r) := max{f (z) z = r}. Using
the notation O and o of Landau, f ∈ H ∞
v (C) if and only if M (f, r) = O(1/v(r)), r → ∞.
By C, C ′, c etc. we denote positive constants, the value of which may vary from place to
place.
Let T : X → X be a continuous operator on a space X. The resolvent of T on X is
the set ρ(T, X) of all λ ∈ C such that T − λI : X → X is bijective and has a continuous
inverse. Here I stands for the identity operator on X. The spectrum σ(T, X) of T is the
complement in C of the resolvent. The point spectrum is the set σpt(T, X) of those λ ∈ C
such that T − λI is not injective. Observe that in this paper we consider operators defined
not only on Banach spaces, but also on more general spaces: H(C) and the Hormander
algebra A0
p are Fr´echet spaces and the Hormander algebra Ap is the dual of a Fr´echet
space. Accordingly, when we refer to a space, we mean a Hausdorff locally convex space.
We refer the reader to [25] for results and terminology about functional analysis.
At this point we include some preliminary results, in particular about the spectrum
of the Volterra operator on H(C) that are inspired by [1] and [2]. Let g ∈ H(C) be a
0 f (ζ)g′(ζ)dζ, z ∈ C,
non-constant entire function such that g(0) = 0 and let Vgf (z) = R z
denote the Volterra operator associated with g, that acts continuously on H(C).
Proposition 1.1 The operator Vg − λI : H(C) → H(C) is injective for each λ ∈ C. In
particular σpt(Vg, H(C)) = ∅. Moreover, 0 ∈ σ(Vg, H(C)).
Proof. If (Vg − λI)f = 0, then λf (0) = 0 and f (z)g′(z) = λf ′(z), z ∈ C,. If λ = 0, we
get f = 0, since g is not constant. In case λ 6= 0, we get f (z) = C exp(g(z)/λ), which
implies f = 0, since f (0) = 0.
The operator Vg is not surjective on H(C) because Vgf (0) = 0 for each f ∈ H(C).
✷
Thus 0 ∈ σ(Vg, H(C)).
Lemma 1.2 Given λ ∈ C, λ 6= 0, and h ∈ H(C), the equation f − (1/λ)Vgf = h has a
unique solution given by
f (z) = Rλ,gh(z) = h(0)e
g(z)
λ + e
g(z)
λ Z z
0
e− g(ζ)
λ h′(ζ)dζ,
z ∈ C.
[1, p. 200] or [2, p. 2]. The uniqueness follows
Proof. This is well known; see e.g.
from Proposition 1.1.
It is enough to substitute in the equation to check the result.
Alternatively, the equation implies f (0) = h(0) and that f is the solution of the equation
y′ − (g′(z)/λ)y = h′(z). The result is also obtained solving this linear equation.
✷
2
Proposition 1.3 Let g ∈ H(C) be a non-constant entire function such that g(0) = 0.
The Volterra operator Vg satisfies σ(Vg, H(C)) = {0} and σpt(Vg, H(C)) = ∅.
Proof. This is a direct consequence of Proposition 1.1 and the continuity of the operator
Rλ,g : H(C) → H(C), λ 6= 0.
✷
Lemma 1.4 Let X ⊂ H(C) be a locally convex space that contains the constants and such
that the inclusion X ⊂ H(C) is continuous. Assume that Vg : X → X is continuous for
some non-constant entire function g such that g(0) = 0. Then
{0} ∪ {λ ∈ C \ {0} e
g
λ /∈ X} ⊂ σ(Vg, X).
If X is a Banach space, then
{0} ∪ {λ ∈ C \ {0} e
g
λ /∈ X} ⊂ σ(Vg, X).
Proof. Since X contains the constants, Vg is not surjective and 0 ∈ σ(Vg, X).
If λ /∈
σ(Vg, X), then λ 6= 0 and the operator Rλ,g : X → X defined in Lemma 1.2 is continuous.
λ ∈ X. This implies the desired inclusion. Recall that in case
In particular, Rλ,g(1) = e
X is a Banach space, σ(Vg, X) is compact.
✷
g
Lemma 1.5 Let X ⊂ H(C) be a locally convex space that contains the constants and
such that the inclusion X → H(C) is continuous. Assume that Vg : X → X is continuous
for some non-constant entire function g such that g(0) = 0. The following conditions are
equivalent:
(i) λ ∈ ρ(Vg, X).
(ii) Rλ,g : X → X is continuous.
(iii) (a) e
g
λ ∈ X, and
(b) Sλ,g : X0 → X0, Sλ,gh(z) := e
subspace X0 of X of all the functions h ∈ X with h(0) = 0.
0 h′(ζ)e− g(ζ)
g(z)
λ R z
λ dζ, z ∈ C, is continuous on the
Proof. This follows directly from the definitions.
✷
Lemma 1.6 Let X ⊂ H(C) be a locally convex space that contains the constants and
such that the inclusion X → H(C) is continuous. Let X0 be the subspace of X of all the
functions h ∈ X with h(0) = 0. The following conditions are equivalent for λ ∈ C \ {0}.
(i) Sλ,g : X0 → X0, Sλ,gh(z) := e
λ dζ, z ∈ C, is continuous.
g(z)
0 h′(ζ)e− g(ζ)
λ R z
0 h(ζ)g′(ζ)e− g(ζ)
g(z)
λ R z
(ii) T : X0 → X0, T h(z) := e
λ dζ, z ∈ C, is continuous.
Proof. This is a direct consequence of the identity
e
g(z)
λ Z z
0
h′(ζ)e− g(ζ)
λ dζ = h(z) + (1/λ)e
g(z)
λ Z z
0
h(ζ)g′(ζ)e− g(ζ)
λ dζ,
valid for h ∈ H(C), h(0) = 0, that can be seen integrating by parts.
✷
3
2 Spectra of Volterra operators on H ∞
v (C)
We concentrate our attention in the Volterra operator acting on the Banach space H ∞
v (C),
with v(r) = exp(−αrp), where α, p > 0. According to [13, Corollaries 3.12 and 3.13], we
have the following result.
Proposition 2.1 Assume that v(r) = exp(−αrp), α > 0, p > 0.
(i) Vg : H ∞
v (C) → H ∞
v (C) is continuous if and only if g is a polynomial of degree less
than or equal to the integer part of p.
(ii) Vg : H ∞
v (C) → H ∞
v (C) is compact if and only if g is a polynomial of degree less than
or equal to the integer part of p − 1.
Lemma 2.2 Let v be a weight such that v(r)eαrn
r0 > 0, α > 0 and n ∈ N. The operator T : H ∞
is non-increasing on [r0, ∞[ for some
v (C) → H ∞
v (C) defined by
Tγh(z) := eγzn Z z
0
ζ n−1h(ζ)e−γζ n
dζ,
z ∈ C,
is continuous if γ < α.
Proof. Changing the value of v(r) on a compact interval, we may assume that v(r)eαrn
is non-increasing on [0, ∞[. Fix z ∈ C, z 6= 0. For each 0 ≤ t ≤ 1 we have
v(z) ≤ v(tz)eαzn(tn−1).
Therefore we can estimate
Z 1
0
v(z)Tγ h(z) = v(z)(cid:12)(cid:12)(cid:12)(cid:12)
≤ znZ 1
0
zntn−1h(tz)eγzn(1−tn)dt(cid:12)(cid:12)(cid:12)(cid:12)
≤
tn−1h(tz)v(tz)eγzn(1−tn)eαzn(tn−1)dt ≤
≤ hvznZ 1
0
tn−1e(α−γ)zn(tn−1)dt =
(1/n)hv (α − γ)−1(cid:0)1 − e−(α−γ)zn(cid:1) ≤ (1/n)(α − γ)−1hv,
since α − γ > 0.
✷
Theorem 2.3 Assume that v(r) = exp(−αrp), α > 0, p > 0. Let g be a polynomial of
degree n less than or equal to the integer part of p with g(0) = 0.
(i) If the degree n of g satisfies n < p, then σ(Vg, H ∞
v (C)) = {0}.
(ii) If p = n ∈ N and g(z) = βzn + k(z), k a polynomial of degree strictly less than n,
then σ(Vg, H ∞
v (C)) = {λ ∈ C λ ≤ β/α}.
Moreover, we have σ(Vg, H ∞
v (C)) = {0} ∪ {λ ∈ C \ {0} e
4
g
λ /∈ H ∞
v (C)}.
Proof. (i) If n is less than or equal to the integer part of p − 1, then Vg : H ∞
H ∞
by Proposition 1.1.
v (C) →
v (C)) = {0}, since Vg has no eigenvalues
v (C) is compact. This implies that σ(Vg, H ∞
Assume now that p − 1 < n < p. For each λ 6= 0, e
v (C) as it is easy to check.
Suppose first that g(z) = βzn for some β 6= 0. For λ 6= 0, take α > β/λ. Clearly
is non-increasing on [r0, ∞[ for some r0 > 0. We can apply Lemma 2.2 to conclude
g
λ ∈ H ∞
v(r)eαrn
that
Tβ/λh(z) := e
βzn
λ Z z
0
ζ n−1h(ζ)e− βζn
λ dζ,
z ∈ C,
is continuous on the subspace of H ∞
the operator
v (C) of the functions vanishing at 0. By Lemma 1.6,
Sλ,gh(z) := e
h′(ζ)e− g(ζ)
λ dζ, z ∈ C,
g(z)
λ Z z
0
is continuous in the same Banach space. Finally, Lemma 1.5 implies λ ∈ ρ(Vg, H ∞
This completes the proof in this case.
v (C)).
Suppose now that g(z) = βzn + k(z) for some β 6= 0 and some polynomial k of degree
strictly less than n. Setting g1(z) := βzn, we have Vg = Vg1 + Vk, and Vk is a compact
injective operator on H ∞
v (C). If λ 6= 0, we have Vg − λI = (Vg1 − λI) + Vk, and Vg − λI
is bijective if and only if Vg1 − λI is bijective. This is a consequence e.g. of [20, Corollary
34.14] keeping in mind that both Vg − λI and Vg1 − λI are injective by Proposition 1.1.
Therefore, we conclude σ(Vg, H ∞
v (C)) = σ(Vg1, H ∞
v (C)) = {0}.
(ii) We suppose now that v(r) = exp(−αrn), α > 0, and that g is a polynomial of
degree exactly n.
Again we consider first the case g(z) = βzn. For λ ∈ C \ {0}, we have e
v (C) if
and only if β/λ ≤ α. Therefore, we can apply Lemma 1.4 to conclude that {λ λ ≤
β/α} ⊂ σ(Vg, H ∞
v (C)). Now take λ ∈ C with λ > β/α. Since v(r) exp(αrn) = 1, we
apply Lemma 2.2 and Lemma 1.6 to get that the operator
λ ∈ H ∞
g
Sλ,gh(z) := e
g(z)
λ Z z
0
h′(ζ)e− g(ζ)
λ dζ, z ∈ C,
is continuous on the subspace of H ∞
λ ∈ ρ(Vg, H ∞
the present case.
v (C)) by Lemma 1.5, which yields σ(Vg, H ∞
v (C) of the functions vanishing at 0. Consequently
v (C)) = {λ ∈ C λ ≤ β/α} in
In the general case g(z) = βzn + k(z), β 6= 0 and some polynomial k of degree
v (C)) =
strictly less than n, we proceed as in the proof of part (i) to conclude σ(Vg, H ∞
σ(Vg1, H ∞
v (C)), g1(z) = βzn.
Given a weight v, one defines the following closed subspace of H ∞
v (C)):
H 0
v (C) := {f ∈ H(C)
lim
z→∞
v(z)f (z) = 0}.
✷
The polynomials are contained and dense in H 0
v (C) is isometrically isomorphic to H ∞
v (C). By [12, Ex 2.2], the bidual
of H 0
v (C). By [13, Corollaries 3.8 and 3.12] Vg is
v (C) when v(r) = exp(−αrp), α, p > 0.
bounded on H 0
Now, proceeding as in the proof of [5, Lemma 1] or [6, Lemma 2.1] one can show, for
v(r) = exp(−αrp), that if Vg is bounded on H ∞
v (C), then it coincides with the bi-transpose
(Vg)′′ of Vg : H 0
v (C) if and only if it is bounded on H ∞
v (C). Accordingly, σ(Vg, H 0
v (C)) = σ(Vg, H ∞
v (C) → H 0
v (C)).
5
3 Spectra of Volterra operators on Hormander algebras
A function p : C → [0, ∞[ is called a growth condition if it is continuous, subharmonic,
radial, increases with z and satisfies:
(α) log(1 + z2) = o(p(z)) as z → ∞,
(β) p(2z) = O(p(z)) as z → ∞.
Given a growth condition p, we define the following weighted spaces of entire functions
(see e.g. [8] and [9]):
Ap := {f ∈ H(C) there is A > 0 : sup
z∈C
f (z) exp(−Ap(z)) < ∞},
endowed with the inductive limit topology, for which it is a (DFN)-algebra (cf. [23]). We
also define
A0
p := {f ∈ H(C) for all ε > 0 : sup
z∈C
f (z) exp(−εp(z)) < ∞},
endowed with the projective limit topology, for which it is a nuclear Fr´echet algebra (cf.
[24]).
Clearly A0
p ⊂ Ap. Condition (α) implies that the polynomials are dense in Ap and
in A0
p. Condition (β) implies that both spaces are stable under differentiation. By the
closed graph theorem, the differentiation operator D is continuous on Ap and on A0
p. It
was observed in [7, Lemma 4.1] that the integration operator J is continuous on both
spaces.
Weighted algebras of entire functions of this type, usually known as Hormander alge-
bras, have been considered since the work of Berenstein and Taylor [9] by many authors;
see e.g. [8] and the references therein. Braun, Meise and Taylor studied in [14], [23] and [24]
the structure of (complemented) ideals in these algebras. As an example, when p(z) = za,
then Ap consists of all entire functions of order a and finite type or order less than a, and
A0
p is the space of all entire functions of order at most a and type 0. For a = 1, Ap is the
space of all entire functions of exponential type, and A0
p is the space of entire functions of
infraexponential type.
The following Lemma, that is a consequence of condition (β) for the growth condition
p, is well known.
Lemma 3.1 If p : C → [0, ∞[ is a growth condition, then there are M > 0 and s > 0
such that p(r) ≤ M rs + M for each r ∈ [0, ∞[.
Proposition 3.2 Let g be an entire function.
(i) Vg : Ap → Ap is continuous if and only if g ∈ Ap.
(ii) Vg : A0
p → A0
p is continuous if and only if g ∈ A0
p.
Proof. We show (i), the proof of (ii) being similar. The identities D ◦ Vg = Mg′, Vg =
J ◦ Mg′ and the continuity of D and J on Ap imply that Vg : Ap → Ap is continuous if and
only the operator of multiplication Mg′ : Ap → Ap is continuous. Since Ap is an algebra
containing the constants, this holds if and only if g′ ∈ Ap. We can apply again that D
and J are continuous on Ap to conclude that this statement is equivalent to g ∈ Ap. ✷
6
The spectrum of the integration operator J on Ap and A0
p was investigated in [7,
Proposition 4.7]. This result, that corresponds to Vg for g(z) = z, is extended in this
section.
Lemma 3.3 Let p : C → [0, ∞[ be a growth condition and let h be an entire function.
(i) The function eh belongs to Ap if and only if M (h, r) = O(p(r)) as r → ∞. If this is
the case, then h is a polynomial.
(ii) The function eh belongs to A0
p if and only if M (h, r) = o(p(r)) as r → ∞. If this is
the case, then h is a polynomial.
Proof. (i) If M (h, r) = O(p(r)) as r → ∞, then eh clearly belongs to Ap. To prove the
converse, denote by A(h, r) the maximum of the real part of h in the circle z ≤ r for
r > 0. As eh ∈ Ap, there is C > 0 such that A(h, r) ≤ Cp(r) + C for each r > 0. We apply
Carath´eodory inequality for the circle z ≤ 2r (c.f. [21, Theorem 6.8]) and property (β)
of p to get
M (h, r) ≤ 2(A(h, 2r) − Reh(0)) + h(0) ≤ 2Cp(2r) + 2(C + h(0)) ≤ C ′p(r) + C ′.
Now Lemma 3.1 and a standard argument imply that h is a polynomial.
(ii) By [24, Lemma 2.3], eh ∈ A0
p if and only if there is a growth condition q such that
eh ∈ Aq and q(r) = o(p(r)), r → ∞. By part (i), M (h, r) = O(q(r)) = o(p(r)), r → ∞.
The reverse implication is trivial.
✷
Theorem 3.4 Let p : C → [0, ∞[ be a growth condition and let g ∈ Ap be non-constant.
(i) If M (g, r) = O(p(r)), r → ∞, is not satisfied (which happens in particular when
p(r) = o(r), r → ∞), then σ(Vg, Ap) = C.
(ii) If M (g, r) = O(p(r)), r → ∞, then σ(Vg, Ap) = {0}. In this case g is a polynomial
and r = O(p(r)), r → ∞.
Moreover, in both cases we have σ(Vg, Ap) = {0} ∪ {λ ∈ C \ {0} e
g
λ /∈ Ap}.
Proof. First observe that M (g/λ, r) = (1/λ)M (g, r) for each λ ∈ C \ {0} and r > 0.
λ ∈ Ap for some (all) λ 6= 0, if and only if
Therefore, it follows from Lemma 3.3 (i) that e
eg ∈ Ap, that is equivalent to M (g, r) = O(p(r)) as r → ∞.
(i) If M (g, r) = O(p(r)) as r → ∞ is not satisfied, then e
λ /∈ Ap for each λ 6= 0 by our
g
g
remarks above. We can apply Lemma 1.4 to conclude σ(Vg, Ap) = C.
Observe that in case p(r) = o(r), r → ∞, then M (g, r) = O(p(r)) as r → ∞ is not
satisfied, since otherwise we would have M (g, r) = O(p(r)) = o(r) as r → ∞, that implies
that g is constant; a contradiction.
g
g
(ii) If M (g, r) = O(p(r)), r → ∞, then e
λ ∈ Ap for each λ 6= 0. Given λ ∈ C \ {0},
λ , the operator Sλ,g of Lemma 1.5 (iii) satisfies Sλ,g = MG ◦ J ◦
setting G = e
M1/G ◦D. These four operators are continuous on the algebra Ap. Therefore λ ∈ ρ(Vg, Ap),
and σ(Vg, Ap) = {0}.
λ , 1/G = e− g
In this case, since g must be a non constant polynomial, the assumption in (ii) implies
r = O(p(r)), r → ∞.
✷
7
Theorem 3.5 Let p : C → [0, ∞[ be a growth condition and let g ∈ A0
p be non-constant.
(i) If M (g, r) = o(p(r)), r → ∞, is not satisfied (which happens in case p(r) = O(r), r →
∞), then σ(Vg, A0
p) = C.
(ii) If M (g, r) = o(p(r)), r → ∞, then σ(Vg, A0
p) = {0}. In this case g is a polynomial
and r = o(p(r)), r → ∞.
Moreover, in both cases, we have σ(Vg, A0
p) = {0} ∪ {λ ∈ C \ {0} e
g
λ /∈ A0
p}.
Proof. Since M (g/λ, r) = (1/λ)M (g, r) for each λ ∈ C \ {0} and r > 0, Lemma 3.3 (ii)
implies that e
p for some (or all) λ 6= 0 if and only if eg ∈ A0
p.
λ /∈ A0
p for each λ 6= 0 by our
(i) If M (g, r) = o(p(r)) as r → ∞ is not satisfied, then e
λ ∈ A0
g
g
comments above. By Lemma 1.4 we have σ(Vg, Ap) = C.
Observe that in case p(r) = O(r), r → ∞, then M (g, r) = o(p(r)) as r → ∞ does not
hold. Otherwise, M (g, r) = o(p(r)) = O(r) as r → ∞, which implies that g is constant; a
contradiction.
g
λ ∈ A0
(ii) If M (g, r) = o(p(r)), r → ∞, then e
above. Given λ ∈ C \ {0}, setting G = e
Sλ,g = MG ◦ J ◦ M1/G ◦ D and is continuous on the algebra A0
p for each λ 6= 0 by our comments
λ , the operator Sλ,g of Lemma 1.5 (iii) satisfies
p).
Since g must be a non constant polynomial, the assumption in (ii) yields r = o(p(r))
p. Therefore λ ∈ ρ(Vg, A0
g
as r → ∞.
✷
Acknowledgement. This research was partially supported by MINECO Project
MTM2013-43540-P and by GV Project Prometeo II/2013/013.
References
[1] A. Aleman, O. Constantin, Spectra of integration operators on weighted Bergman
spaces, J. Anal. Math. 109 (2009), 199-231.
[2] A. Aleman, O. Constantin, Spectra of integration operators and weighted square func-
tions, Indiana Univ. Math. J. 61 (2012), 1-19.
[3] A. Aleman, A.G. Siskakis, An integral operator on H p, Complex Var. Theory Appl.
28 (1995), 149-158.
[4] A. Aleman, A.G. Siskakis, Integration operators on Bergman spaces, Indiana Univ.
Math. J. 46 (1997), 337-356.
[5] M. Basallote, M.D. Contreras, C. Hern´andez-Mancera, M.J. Mart´ın, P.J. Pa´ul,
Volterra operators and semigroups in weighted Banach spaces of analytic functions,
Collect. Math. 65 (2014), 233-249.
[6] M.J. Beltr´an, J. Bonet, C. Fern´andez, Classical operators on weighted Banach spaces
of entire functions, Proc. Amer. Math. Soc. 141 (2013), 4293 -- 4303.
[7] M.J. Beltr´an, J. Bonet, C. Fern´andez, Classical operators on the Hormander algebras,
Discr. Cont. Dynam. Syst. 35 (2015), 637-652.
[8] C. A. Berenstein and R. Gay. Complex analysis and special topics in harmonic anal-
ysis. Springer-Verlag, New York, 1995.
8
[9] C. Berenstein and B. A. Taylor, A new look at interpolation theory for entire functions
of one variable, Adv. in Math. 33 (1979), 109 -- 143.
[10] K.D. Bierstedt, J. Bonet, A. Galbis, Weighted spaces of holomorphic functions on
balanced domains, Michigan Math. J. 40 (1993), no. 2, 271-297.
[11] K.D. Bierstedt, J. Bonet, J. Taskinen, Associated weights and spaces of holomorphic
functions, Studia Math. 127 (1998), 137-168.
[12] K.D. Bierstedt, W. H. Summers, Biduals of weighted Banach spaces of analytic func-
tions, J. Austral. Math. Soc. Ser. A 54 (1993), no. 1, 70-79.
[13] J. Bonet, J. Taskinen, A note about Volterra operators on weighted Banach spaces of
entire functions, Math. Nachr. (to appear).
[14] R. W. Braun, Weighted algebras of entire functions in which each closed ideal admits
two algebraic generators, Michigan Math. J. 34 (1987), 441 -- 450.
[15] O. Constantin, A Volterra-type integration operator on Fock spaces, Proc. Amer.
Math. Soc. 140 (2012), 4247-4257.
[16] O. Constantin, J.A. Pel´aez, Integral operators, embedding theorems and a Littlewood-
Paley formula on weighted Fock spaces, arXiv:1304.7501v1.
[17] O. Constantin, J.A. Pel´aez, Boundedness of the Bergman projection on Lp spaces with
exponential weights, arXiv: 1309.6071v2.
[18] M. Contreras, A.G. Hern´andez-D´ıaz, Weighted composition operators in weighted Ba-
nach spacs of analytic functions, J. Austral. Math. Soc. (Series A) 69 (2000), 41-60.
[19] A. Harutyunyan, W. Lusky, On the boundedness of the differentiation operator be-
tween weighted spaces of holomorphic functions, Studia Math. 184 (2008), 233-247.
[20] G.J.O. Jameson, Topology and Normed Spaces, Chapman Hall, London, 1974.
[21] B. Ya. Levin, Distribution of zeros of entire functions. Translated from the Russian
by R. P. Boas, J. M. Danskin, F. M. Goodspeed, J. Korevaar, A. L. Shields and H. P.
Thielman. Revised edition. Translations of Mathematical Monographs, 5. American
Mathematical Society, Providence, R.I., 1980.
[22] W. Lusky, On the isomorphism classes of weighted spaces of harmonic and holomophic
functions, Studia Math. 75 (2006) 19-45.
[23] R. Meise, Sequence space representations for (DFN)-algebras of entire functions mod-
ulo closed ideals, J. Reine Angew. Math. 363 (1985), 59 -- 95.
[24] R. Meise and B. A. Taylor, Sequence space representations for (FN)-algebras of entire
functions modulo closed ideals, Studia Math. 85 (1987), 203 -- 227.
[25] R. Meise and D. Vogt. Introduction to functional analysis, volume 2 of Oxford Gradu-
ate Texts in Mathematics. The Clarendon Press Oxford University Press, New York,
1997.
[26] J. Pau, J.A. Pel´aez, Embedding theorems and integration operators on Bergman spaces
with rapidly decreasing weights, J. Funct. Anal. 259 (2010), 2727-2756.
9
[27] J.A. Pel´aez, J. Rattya, Weighted Bergman Spaces induced by rapidly decreasing
weights, Memoirs Amer. Math. Soc. 227, 2014.
[28] Ch. Pommerenke, Schlichte Funktionen un analytische Functionen von beschrnkter
mittlerer Oszilation, Comment. Math. Helv. 52 (1977), 591-602.
[29] J.H. Shapiro, Composition Operators and Classical Function Theorey, Springer, New
York, 1993.
Author's address:
Jos´e Bonet: Instituto Universitario de Matem´atica Pura y Aplicada IUMPA, Univer-
sitat Polit`ecnica de Val`encia, E-46071 Valencia, Spain
email: [email protected]
10
|
1703.02218 | 2 | 1703 | 2017-09-08T03:52:29 | Strongly quasinonexpansive mappings, II | [
"math.FA"
] | This paper is devoted to the study of strongly quasinonexpansive mappings in an abstract space and a Banach space. | math.FA | math |
STRONGLY QUASINONEXPANSIVE MAPPINGS, II
KOJI AOYAMA AND KEI ZEMBAYASHI
Abstract. This paper is devoted to the study of strongly quasinonexpansive
mappings in an abstract space and a Banach space.
1. Introduction
One of the authors introduced the notion of strongly quasinonexpansive map-
pings in a metric space in [3]. The notion is analogous to strong nonexpansiveness
introduced in Bruck and Reich [10]. Indeed, a strongly nonexpansive mapping in
the sense of [10] with a fixed point is strongly quasinonexpansive in the sense of [3]
in the framework of Banach space.
On the other hand, a mapping of type (sr) was studied in [5, 7, 8] in a Banach
space. Such a mapping is also analogous to a strongly nonexpansive mapping in [10];
see also Reich [12]. However, a mapping of type (sr) is different from a strongly
quasinonexpansive mapping in the sense of [3].
In this paper, in order to unify strong quasinonexpansiveness as above, we intro-
duce and study a quasinonexpansive mapping, a strictly quasinonexpansive map-
ping, and a strongly quasinonexpansive mapping in an abstract space. In particu-
lar, we give some characterizations and basic properties of such quasinonexpansive
mappings. Then, using these results, we obtain characterizations and properties of
mappings of type (sr) in the sense of [5, 7, 8].
2. Preliminaries
Throughout this paper, R+ denotes the set of nonnegative real numbers, N the
set of positive integers, E a real Banach space, E∗ the dual of E, k · k the norms of
E and E∗, hx, x∗i the value of x∗ ∈ E∗ at x ∈ E, and J the duality mapping of E
into 2E ∗
, that is, Jx = (cid:8)x∗ ∈ E∗ : hx, x∗i = kxk2 = kx∗k2(cid:9) for x ∈ E.
It is known that the duality mapping J is single-valued if E is smooth; J is
injective if E is strictly convex, that is, Jx ∩ Jy = ∅ for all x, y ∈ E with x 6= y; see
[13] for more details. It is also known that every uniformly convex Banach space is
strictly convex.
Let X be a metric space with metric d and C a nonempty subset of X. A
mapping T : C → X is said to be quasinonexpansive if F(T ) is nonempty and
d(z, T x) ≤ d(z, x) for all z ∈ F(T ) and x ∈ C, where F(T ) is the fixed point set of
T , that is, F(T ) = {z ∈ C : z = T z}; T is said to be strictly quasinonexpansive
if F(T ) is nonempty and d(z, T x) < d(z, x) for all z ∈ F(T ) and x ∈ C \ F(T );
2010 Mathematics Subject Classification. 47H09, 47H10, 41A65.
Key words and phrases. Strongly quasinonexpansive mapping, strictly quasinonexpansive
mapping, quasinonexpansive mapping, fixed point.
1
2
KOJI AOYAMA AND KEI ZEMBAYASHI
T is said to be strongly quasinonexpansive [3] if F(T ) is nonempty and for any
z ∈ F (T ), M > 0, and ǫ > 0 there exists δ > 0 such that
x ∈ C, d(z, x) ≤ M, d(z, x) − d(z, T x) < δ ⇒ d(T x, x) < ǫ.
3. Lemmas
In this section, we deal with two lemmas, which play a central role in providing
characterizations of strongly quasinonexpansive mappings in the next section.
Throughout this section, we assume that D is a nonempty set, both f and g are
bounded (above) functions of D into R+, and α = sup{f (x) : x ∈ D}.
Lemma 3.1. Suppose that α > 0 and γ(t) is defined by
(3.1)
γ(t) = inf{g(x) : x ∈ D, f (x) ≥ t}
for t ∈ [0, α). Then the following hold:
(1) γ is a nondecreasing bounded function of [0, α) into R+;
(2) if x ∈ D and f (x) ∈ [0, α), then γ(cid:0)f (x)(cid:1) ≤ g(x);
(3) limt↑α γ(t) = sup{γ(t) : t ∈ [0, α)} < ∞.
Proof. Let t ∈ [0, α) be fixed. Then, by the definition of α, there exists x ∈ D
such that f (x) > t. Thus {g(x) : x ∈ D, f (x) ≥ t} is nonempty, and hence γ is a
real-valued function defined on [0, α). The conclusions (1) and (2) follow from the
definition of γ; (3) follows from (1). This completes the proof.
(cid:3)
Using Lemma 3.1, we obtain the following:
Lemma 3.2. The following are equivalent:
(1) For any ǫ > 0 there exists δ > 0 such that
x ∈ D, g(x) < δ ⇒ f (x) < ǫ;
(2) f (xn) → 0 whenever {xn} is a sequence in D and g(xn) → 0;
(3) there exists a nondecreasing bounded function γ of [0, α] into R+ such that
γ(cid:0)f (x)(cid:1) ≤ g(x) for all x ∈ D and γ(t) > 0 for all t ∈ (0, α].
Proof. We first show that (1) implies (3). Suppose that α = 0. In this case, the
conclusion is clear by setting γ(0) = 0. Next we suppose that α > 0. Let γ be
a function of [0, α] into R+ defined by (3.1) for t ∈ [0, α) and γ(α) = limt↑α γ(t).
Then it follows from Lemma 3.1 that γ is well defined and nondecreasing, and
moreover, γ(cid:0)f (x)(cid:1) ≤ g(x) for all x ∈ D with f (x) ∈ [0, α). If x ∈ D and f (x) = α,
then it is obvious that γ(t) ≤ g(x) for all t ∈ [0, α) and hence
γ(α) = lim
t↑α
γ(t) = sup{γ(t) : t ∈ [0, α)} ≤ g(x).
Therefore, γ(cid:0)f (x)(cid:1) ≤ g(x) for all x ∈ D. We finally show that γ(t) > 0 for all
t ∈ (0, α]. Suppose that there exists t ∈ (0, α] such that γ(t) = 0. Without loss
of generality, we may assume that t 6= α. By the definition of γ, there exists a
sequence {yn} in D such that f (yn) ≥ t and g(yn) < 1/n for all n ∈ N. On the
other hand, by assumption, there exists δ > 0 such that
x ∈ D, g(x) < δ ⇒ f (x) < t/2.
Choosing m ∈ N with 1/m < δ, we have t ≤ f (ym) ≤ t/2, which is a contradiction.
Therefore, γ(t) > 0 for all t ∈ (0, α].
STRONGLY QUASINONEXPANSIVE MAPPINGS, II
3
We next show that (3) implies (2). Let {xn} be a sequence in D and suppose
that g(xn) → 0 and f (xn) 6→ 0. Then there exist ǫ > 0 and a subsequence {xni} of
{xn} such that f (xni ) ≥ ǫ for all i ∈ N. Thus, by assumption, it follows that
0 < γ(ǫ) ≤ γ(cid:0)f (xni )(cid:1) ≤ g(xni ) → 0
as i → ∞, which is a contradiction. Therefore, f (xn) → 0.
We finally show that (2) implies (1). Suppose that the conclusion does not hold.
Then there exist ǫ > 0 and a sequence {xn} in D such that g(xn) < 1/n and
f (xn) ≥ ǫ for all n ∈ N. Thus g(xn) → 0 and hence, by assumption, f (xn) → 0,
which is a contradiction.
(cid:3)
4. Strongly quasinonexpansive mappings in an abstract space
In this section, we introduce and study a quasinonexpansive mapping, a strictly
quasinonexpansive mapping, and a strongly quasinonexpansive mapping in an ab-
stract space, which is a generalization of a metric space.
Throughout this section, X denotes a nonempty set, σ a function of X × X into
R+, and ¯B(z, M ) a subset of X defined by
¯B(z, M ) = {x ∈ X : σ(z, x) ≤ M }
for z ∈ X and M > 0.
We say that the pair (X, σ) satisfies the condition (S) if
x 6= y ⇔ σ(x, y) > 0
for all x, y ∈ X; (X, σ) satisfies the condition (B) if
(4.1)
sup{σ(x, y) : x, y ∈ ¯B(z, M )} < ∞
for all z ∈ X and M > 0; (X, σ) satisfies the condition (T) if for any u ∈ X, M > 0,
and ǫ > 0 there exists η > 0 such that
x, y, z ∈ ¯B(u, M ), σ(x, y) < η, σ(y, z) < η ⇒ σ(x, z) < ǫ.
(4.2)
It is clear that if (X, σ) satisfies the condition (S), then z ∈ ¯B(z, M ), and hence
¯B(z, M ) is nonempty for all z ∈ X and M > 0. It is also clear that (X, σ) satisfies
the condition (T) if and only if σ(xn, zn) → 0 whenever {xn}, {yn}, and {zn} are
sequences in ¯B(u, M ) for some u ∈ X and M > 0 such that σ(xn, yn) → 0 and
σ(yn, zn) → 0.
Remark 4.1. Suppose that σ is a metric on X, that is, (X, σ) is a metric space.
Then it is obvious that (X, σ) satisfies the conditions (S), (B), and (T).
Let C be a nonempty subset of X, T a mapping of C into X, and F(T ) the fixed
point set of T .
Inspired by [3, 6 -- 9], we introduce the following: T is said to be
quasinonexpansive (with respect to σ) if F(T ) is nonempty and σ(z, T x) ≤ σ(z, x)
for all z ∈ F(T ) and x ∈ C; T is said to be strictly quasinonexpansive (with respect
to σ) if F(T ) is nonempty and σ(z, T x) < σ(z, x) for all z ∈ F(T ) and x ∈ C \ F(T );
T is said to be strongly quasinonexpansive (with respect to σ) if F(T ) is nonempty
and for any z ∈ F(T ), M > 0, and ǫ > 0 there exists δ > 0 such that
x ∈ C ∩ ¯B(z, M ), σ(z, x) − σ(z, T x) < δ ⇒ σ(T x, x) < ǫ.
(4.3)
It is clear from the definitions that every strictly quasinonexpansive mapping is
quasinonexpansive, and moreover, we obtain the following:
4
KOJI AOYAMA AND KEI ZEMBAYASHI
Lemma 4.2. Let C be a nonempty subset of X and T a strongly quasinonexpansive
mapping of C into X. Suppose that (X, σ) satisfies the condition (S). Then T is
(strictly) quasinonexpansive.
Proof. Suppose that T is not strictly quasinonexpansive, that is, there exist z ∈
F(T ) and y ∈ C \ F(T ) such that σ(z, T y) ≥ σ(z, y). Then clearly z 6= y and
T y 6= y. Set M = σ(z, y) and ǫ = σ(T y, y). Taking into account the condition (S),
we see that M > 0 and ǫ > 0. Since T is strongly quasinonexpansive, there exists
δ > 0 such that (4.3) holds. Therefore, σ(T y, y) < ǫ, which is a contradiction. (cid:3)
In order to prove the next theorem, we need the following lemma:
Lemma 4.3. Let C be a nonempty subset of X, T a quasinonexpansive mapping
of C into X, z ∈ F(T ), and M > 0. Suppose that (X, σ) satisfies the condition (B).
Let f and g be functions defined by
f (x) = σ(T x, x) and g(x) = σ(z, x) − σ(z, T x)
(4.4)
for x ∈ D, where D = C ∩ ¯B(z, M ). Then f and g are bounded functions of D into
R+.
Proof. Since T is quasinonexpansive, it follows that 0 ≤ g(x) ≤ σ(z, x) ≤ M and
T x ∈ ¯B(z, M ) for all x ∈ D. Hence g is a bounded function of D into R+. Moreover,
by the condition (B), we have
sup{f (x) : x ∈ D} ≤ sup{σ(x, y) : x, y ∈ ¯B(z, M )} < ∞.
Thus f is a bounded function of D into R+.
(cid:3)
We obtain the following characterization of strongly quasinonexpansive map-
pings.
Theorem 4.4. Let C be a nonempty subset of X and T a mapping of C into X.
Suppose that F(T ) is nonempty and (X, σ) satisfies the conditions (S) and (B).
Then the following are equivalent:
(1) T is strongly quasinonexpansive;
(2) T is quasinonexpansive and σ(T xn, xn) → 0 whenever {xn} is a sequence in
C ∩ ¯B(z, M ) and σ(z, xn) − σ(z, T xn) → 0 for some z ∈ F(T ) and M > 0.
Proof. We first show that (1) implies (2). Suppose that T is strongly quasinonex-
pansive. Then Lemma 4.2 implies that T is quasinonexpansive. Moreover, suppose
that {xn} is a sequence in C ∩ ¯B(z, M ), and σ(z, xn) − σ(z, T xn) → 0 for some
z ∈ F(T ) and M > 0. Set D = C ∩ ¯B(z, M ). Let f and g be functions of D
into R+ defined by (4.4) for x ∈ D. Lemma 4.3 shows that f and g are bounded
functions of D into R+. By the implication (1) ⇒ (2) of Lemma 3.2, we conclude
that σ(T xn, xn) = f (xn) → 0 and hence (2) holds.
We next show that (2) implies (1). Suppose that (2) holds. Let z ∈ F(T ),
M > 0, and ǫ > 0 be given. Set D = C ∩ ¯B(z, M ). Let f and g be functions defined
by (4.4) for x ∈ D. Lemma 4.3 shows that f and g are bounded functions of D
into R+. By the implication (2) ⇒ (1) of Lemma 3.2, we see that (1) holds.
(cid:3)
As a direct consequence of Theorem 4.4, we obtain the following:
Corollary 4.5 ([3, Lemma 3.3]). Let X be a metric space with metric d, C a
nonempty subset of X, and T a mapping of C into X. Then the following are
equivalent:
STRONGLY QUASINONEXPANSIVE MAPPINGS, II
5
(1) T is strongly quasinonexpansive;
(2) T is quasinonexpansive and d(T xn, xn) → 0 whenever {xn} is a bounded
sequence in C and d(z, xn) − d(z, T xn) → 0 for some z ∈ F(T ).
We provide another characterization of strongly quasinonexpansive mappings as
follows:
Theorem 4.6. Let C be a nonempty subset of X and T a mapping of C into X.
Suppose that F(T ) is nonempty and (X, σ) satisfies the conditions (S) and (B).
Then the following are equivalent:
(1) T is strongly quasinonexpansive;
(2) for any z ∈ F(T ) and M > 0 there exists a nondecreasing bounded function
γ of [0, α] into R+ such that γ(t) > 0 for all t ∈ (0, α] and
γ(cid:0)σ(T x, x)(cid:1) ≤ σ(z, x) − σ(z, T x)
for all x ∈ D, where D = C ∩ ¯B(z, M ) and α = sup{σ(T x, x) : x ∈ D}.
Proof. We first show that (1) implies (2). Suppose that T is strongly quasinonex-
pansive. As in the proof of Theorem 4.4, we see that T is quasinonexpansive. Let
z ∈ F(T ) and M > 0 be given. Set D = C ∩ ¯B(z, M ). Let f and g be functions
defined by (4.4) for x ∈ D. Then Lemma 4.3 shows that f and g are bounded
functions of D into R+. Using Lemma 3.2, we know that (2) holds.
We next show that (2) implies (1). Suppose that (2) holds. Then we see that T
is quasinonexpansive. Let z ∈ F(T ) and M > 0 be given. Let f and g be functions
defined by (4.4) for x ∈ D = C ∩ ¯B(z, M ). Thus Lemmas 4.3 and 3.2, we conclude
that (1) holds.
(cid:3)
As a direct consequence of Theorem 4.6, we obtain the following:
Corollary 4.7. Let X be a metric space with metric d, C a nonempty subset of X,
and T a mapping of C into X. Suppose that F(T ) is nonempty. Then the following
are equivalent:
(1) T is strongly quasinonexpansive;
(2) for any z ∈ F(T ) and M > 0 there exists a nondecreasing bounded function
γ of [0, α] into R+ such that γ(t) > 0 for all t ∈ (0, α] and
γ(cid:0)d(T x, x)(cid:1) ≤ d(z, x) − d(z, T x)
for all x ∈ C with d(z, x) ≤ M , where α = sup{d(T x, x) : x ∈ C, d(z, x) ≤
M }.
Remark 4.8. Corollary 4.7 is almost the same as [3, Theorem 3.7], which is a precise
characterization of strongly quasinonexpansive mappings in a metric space.
We know that the class of strongly quasinonexpansive mappings in a metric
space is closed under composition [3, Theorem 3.6]; see also [10, Proposition 1.1].
The class of strongly quasinonexpansive mappings with respect to σ has a similar
property as follows:
Theorem 4.9. Let C and D be nonempty subsets of X, and S : C → X and
T : D → X quasinonexpansive mappings such that T (D) ⊂ C and F(S) ∩ F(T ) is
nonempty. Then the following hold:
(1) If S or T is strictly quasinonexpansive, then F(S) ∩ F(T ) = F(ST ) and ST
is quasinonexpansive;
6
KOJI AOYAMA AND KEI ZEMBAYASHI
(2) if both S and T are strongly quasinonexpansive and (X, σ) satisfies the
conditions (S) and (T), then ST is also strongly quasinonexpansive.
Proof. We first prove (1). It is clear that F(S) ∩ F(T ) ⊂ F(ST ) and hence F(ST ) is
nonempty. We show that F(S)∩F(T ) ⊃ F(ST ). Let z ∈ F(ST ) and w ∈ F(S)∩F(T )
be given. Since both S and T are quasinonexpansive, we have
σ(w, z) = σ(w, ST z) ≤ σ(w, T z) ≤ σ(w, z).
This shows that
(4.5)
σ(w, z) = σ(w, T z) = σ(w, ST z).
Now suppose that T is strictly quasinonexpansive. Then T z = z from (4.5). Hence
z = ST z = Sz. Therefore we conclude that z ∈ F(S) ∩ F(T ). On the other
hand, suppose that S is strictly quasinonexpansive. Then it follows from (4.5) that
ST z = T z. Thus z = T z and hence z ∈ F(S) ∩ F(T ). Consequently we know that
F(S) ∩ F(T ) = F(ST ). This implies that σ(z, ST x) ≤ σ(z, T x) ≤ σ(z, x) for all
x ∈ D because S and T are quasinonexpansive. Thus ST is also quasinonexpansive.
We next show (2). Lemma 4.2 implies that S and T are strictly quasinonexpan-
sive. Thus it follows from (1) that F(ST ) = F(S) ∩ F(T ) 6= ∅. Let u ∈ F(ST ),
M > 0, and ǫ > 0 be given. By the condition (T), there exists η > 0 such that (4.2)
holds. Since u ∈ F(S) ∩ F(T ) and both S and T are strongly quasinonexpansive,
there exists δ > 0 such that
x ∈ C ∩ ¯B(u, M ), σ(u, x) − σ(u, Sx) < δ ⇒ σ(Sx, x) < η
and
x ∈ D ∩ ¯B(u, M ), σ(u, x) − σ(u, T x) < δ ⇒ σ(T x, x) < η.
Suppose that y ∈ D ∩ ¯B(u, M ) and σ(u, y) − σ(u, ST y) < δ. Since S and T are
quasinonexpansive and u ∈ F(S) ∩ F(T ), we have
σ(u, ST y) ≤ σ(u, T y) ≤ σ(u, y) ≤ M,
σ(u, y) − σ(u, T y) < δ, and σ(u, T y) − σ(u, ST y) < δ.
Thus y, T y, ST y ∈ ¯B(u, M ), σ(ST y, T y) < η, and σ(T y, y) < η. Therefore it
follows from (4.2) that (ST y, y) < ǫ and hence ST is strongly quasinonexpansive.
(cid:3)
As a direct consequence of Theorem 4.9, we obtain the following:
Corollary 4.10 ([3, Lemma 3.5 and Theorem 3.6]). Let C and D be nonempty
subsets of a metric space X, and S : C → X and T : D → X quasinonexpansive
mappings such that T (D) ⊂ C and F (S) ∩ F (T ) is nonempty. Then the following
hold:
(1) If S or T is strictly quasinonexpansive, then F (S) ∩ F (T ) = F (ST ) and
ST is quasinonexpansive;
(2) if S and T are strongly quasinonexpansive, then ST is also strongly quasi-
nonexpansive.
STRONGLY QUASINONEXPANSIVE MAPPINGS, II
7
5. Strongly quasinonexpansive mappings in a Banach space
In this section, we give some characterizations of mappings of type (sr) in the
sense of [5, 7, 8] and show a fundamental property of such mappings by using the
results in the previous section.
Throughout this section, let E be a smooth Banach space, φ a function of E × E
into R+ defined by
for x, y ∈ E, and ¯B(z, M ) a subset of E defined by
φ(x, y) = kxk2 − 2 hx, Jyi + kyk2
¯B(z, M ) = {x ∈ E : φ(z, x) ≤ M }
for z ∈ E and M > 0; see [1] for the function φ.
By definition, it is clear that
x = y ⇒ φ(x, y) = 0
and
(5.1)
(kxk − kyk)2 ≤ φ(x, y) ≤ (kxk + kyk)2
for all x, y ∈ E. It is also clear that if E is strictly convex, then
(5.2)
φ(x, y) = 0 ⇒ x = y
for all x, y ∈ E. Furthermore, we know the following:
Lemma 5.1 ([11]). Let E be a smooth and uniformly convex Banach space and
both {xn} and {yn} bounded sequences in E. If φ(xn, yn) → 0, then kxn − ynk → 0.
Under the appropriate assumptions, we know that the pair (E, φ) satisfies the
conditions (B), (S), and (T) as stated in the previous section:
Lemma 5.2. Let E be a smooth Banach space. The the following hold:
(1) (E, φ) satisfies the condition (B);
(2) if E is strictly convex, then (E, φ) satisfies the condition (S);
(3) if E is uniformly convex, then (E, φ) satisfies the condition (T).
Proof. We first show (1). Let z ∈ E and M > 0. Then it is clear from (5.1) that
sup{φ(x, y) : x, y ∈ ¯B(z, M )} ≤ supn(kxk + kyk)2 : x, y ∈ ¯B(z, M )o
≤ (2M + kzk)2 < ∞.
(2) immediately follows from (5.2).
Lastly, we show (3). Let {xn}, {yn}, and {zn} be sequences in ¯B(u, M ) for some
u ∈ E and M > 0. Suppose that φ(xn, yn) → 0 and φ(yn, zn) → 0. Since {xn},
{yn}, and {zn} are bounded, it follows from Lemma 5.1 that kxn − ynk → 0 and
kyn − znk → 0. Thus kxn − znk → 0. Therefore we conclude that
φ(xn, zn) = kxnk2 − kznk2 − 2 hxn − zn, Jzni
≤ (kxnk − kznk)(kxnk + kznk) + 2 kxn − znk kznk → 0
as n → 0. This completes the proof.
(cid:3)
8
KOJI AOYAMA AND KEI ZEMBAYASHI
Let C be a nonempty subset of E and T a mapping of C into E. Recall that
a mapping T is said to be quasinonexpansive with respect to φ if F(T ) 6= ∅ and
φ(z, T x) ≤ φ(z, x) for all z ∈ F(T ) and x ∈ C; T is said to be strictly quasinon-
expansive with respect to φ if F(T ) 6= ∅ and φ(z, T x) < φ(z, x) for all z ∈ F(T )
and x ∈ C\F(T ); T is said to be strongly quasinonexpansive with respect to φ if
F(T ) 6= ∅ and for any z ∈ F(T ), M > 0, and ǫ > 0 there exists δ > 0 such that
x ∈ C ∩ ¯B(z, M ), φ(z, x) − φ(z, T x) < δ ⇒ φ(T x, x) < ǫ.
Using Theorems 4.4, 4.6, and Lemma 5.2, we obtain the following characteriza-
tions of strongly quasinonexpansive mappings with respect to φ.
Theorem 5.3. Let C be a nonempty subset of a smooth and strictly convex Banach
space E and T a mapping of C into E. Suppose that F(T ) is nonempty. Then the
following are equivalent:
(1) T is strongly quasinonexpansive with respect to φ;
(2) T is quasinonexpansive with respect to φ and φ(T xn, xn) → 0 whenever
{xn} is a sequence in C ∩ ¯B(z, M ) and φ(z, xn) − φ(z, T xn) → 0 for some
z ∈ F(T ) and M > 0;
(3) for any z ∈ F(T ) and M > 0 there exists a nondecreasing bounded function
γ of [0, α] into R+ such that γ(t) > 0 for all t ∈ (0, α] and
γ(cid:0)φ(T x, x)(cid:1) ≤ φ(z, x) − φ(z, T x)
for all x ∈ D, where D = C ∩ ¯B(z, M ) and α = sup{φ(T x, x) : x ∈ D}.
Remark 5.4. A quasinonexpansive mapping with respect to φ is called a mapping of
type (r) in [2, 4, 5, 7, 8]; a mapping which satisfies the condition (2) in Theorem 5.3
is called a mapping of type (sr) in [5, 7, 8].
Using Lemmas 4.2, 5.2, and Theorem 4.9, we obtain the following theorem, which
is a generalization of [7, Lemma 3.2]; see also [7, Lemma 3.3].
Theorem 5.5. Let C and D be nonempty subsets of a smooth Banach space E,
and S : C → E and T : D → E quasinonexpansive mappings with respect to φ such
that T (D) ⊂ C and F(S) ∩ F(T ) is nonempty. Then following hold:
(1) If S or T is strictly quasinonexpansive with respect to φ, then F(S)∩F(T ) =
F(ST ) and ST is quasinonexpansive with respect to φ;
(2) if E is uniformly convex and both S and T are strongly quasinonexpansive
with respect to φ, then ST is strongly quasinonexpansive with respect to φ.
References
[1] Y. I. Alber, Metric and generalized projection operators in Banach spaces: properties and
applications, Theory and applications of nonlinear operators of accretive and monotone type,
1996, pp. 15 -- 50. MR1386667
[2] K. Aoyama, Asymptotic fixed points of sequences of quasi-nonexpansive type mappings, Ba-
nach and function spaces III (ISBFS 2009), 2011, pp. 343 -- 350. MR3013135
[3]
, Strongly quasinonexpansive mappings, Nonlinear analysis and convex analysis, 2016,
pp. 19 -- 27.
[4] K. Aoyama, Y. Kimura, and F. Kohsaka, Strong convergence theorems for strongly rela-
tively nonexpansive sequences and applications, J. Nonlinear Anal. Optim. 3 (2012), 67 -- 77.
MR2960628
[5] K. Aoyama and F. Kohsaka, Strongly relatively nonexpansive sequences generated by firmly
nonexpansive-like mappings, Fixed Point Theory Appl. (2014), 2014:95, 13. MR3258665
STRONGLY QUASINONEXPANSIVE MAPPINGS, II
9
[6]
, Viscosity approximation process for a sequence of quasinonexpansive mappings,
Fixed Point Theory Appl. (2014), 2014:17, 11. MR3213161
[7] K. Aoyama, F. Kohsaka, and W. Takahashi, Strong convergence theorems by shrinking and
hybrid projection methods for relatively nonexpansive mappings in Banach spaces, Nonlinear
analysis and convex analysis, 2009, pp. 7 -- 26. MR2884574 (2012k:47099)
[8]
, Strongly relatively nonexpansive sequences in Banach spaces and applications, J.
Fixed Point Theory Appl. 5 (2009), 201 -- 224. MR2529497 (2010h:47076)
[9] R. E. Bruck, Random products of contractions in metric and Banach spaces, J. Math. Anal.
Appl. 88 (1982), 319 -- 332. MR667060 (84a:47075)
[10] R. E. Bruck and S. Reich, Nonexpansive projections and resolvents of accretive operators in
Banach spaces, Houston J. Math. 3 (1977), 459 -- 470. MR0470761 (57 #10507)
[11] S. Kamimura and W. Takahashi, Strong convergence of a proximal-type algorithm in
a Banach space, SIAM J. Optim. 13 (2002), 938 -- 945 (electronic) (2003). MR1972223
(2004c:90096)
[12] S. Reich, A weak convergence theorem for the alternating method with Bregman distances,
Theory and applications of nonlinear operators of accretive and monotone type, 1996, pp. 313 --
318. MR1386686
[13] W. Takahashi, Nonlinear functional analysis, Yokohama Publishers, Yokohama, 2000. Fixed
point theory and its applications. MR1864294 (2002k:47001)
(K. Aoyama) Department of Economics, Chiba University, Yayoi-cho, Inage-ku, Chi-
ba-shi, Chiba, 263-8522 Japan
E-mail address: [email protected]
(K. Zembayashi) Keio Girls Senior High School, Mita, Minato-ku, Tokyo, 108-0073
Japan
E-mail address: [email protected]
|
1807.00922 | 1 | 1807 | 2018-07-02T23:22:20 | Positivity, complex FIOs, and Toeplitz operators | [
"math.FA",
"math.AP"
] | We establish a characterization of complex linear canonical transformations that are positive with respect to a pair of strictly plurisubharmonic quadratic weights. As an application, we show that the boundedness of a class of Toeplitz operators on the Bargmann space is implied by the boundedness of their Weyl symbols. | math.FA | math | Positivity, complex FIOs, and Toeplitz operators
8
1
0
2
l
u
J
2
]
Lewis A. Coburn
Michael Hitrik
Department of Mathematics
Department of Mathematics
SUNY at Buffalo
University of California
Buffalo
Los Angeles
NY 14260, USA
CA 90095-1555, USA
[email protected]
[email protected]
.
A
F
h
t
a
m
[
1
v
2
2
9
0
0
.
7
0
8
1
:
v
i
X
r
a
Johannes Sjostrand
IMB, Universit´e de Bourgogne
9, Av. A. Savary, BP 47870
FR-21078 Dijon, France
and UMR 5584 CNRS
[email protected]
Abstract: We establish a characterization of complex linear canonical transforma-
tions that are positive with respect to a pair of strictly plurisubharmonic quadratic
weights. As an application, we show that the boundedness of a class of Toeplitz
operators on the Bargmann space is implied by the boundedness of their Weyl sym-
bols.
Keywords and Phrases: Positive Lagrangian plane, positive canonical transfor-
mation, strictly plurisubharmonic quadratic form, Fourier integral operator in the
complex domain, Toeplitz operator.
Mathematics Subject Classification 2000: 32U05, 32W25, 35S30, 47B35, 70H15
1
Contents
1 Introduction and statement of results
2 Positive Lagrangian planes and positive canonical transformations
in the HΦ–setting
3 Positivity and Fourier integral operators
4 Applications to Toeplitz operators
A Schwartz kernel theorem in the HΦ–setting
B Positivity and Weyl quantization
1
Introduction and statement of results
2
6
15
25
31
34
The notion of a positive complex Lagrangian manifold, introduced by Horman-
der [15], has long played an important role in microlocal analysis and spectral theory.
Restricting the attention to the linear case, relevant for this work, let us recall that
a complex Lagrangian plane Λ ⊂ C2n is said to be positive if we have
1
i
σ(ρ,C(ρ)) ≥ 0,
ρ ∈ Λ.
(1.1)
Here σ is the complex symplectic form on C2n and C : C2n → C2n is the antilin-
ear map of complex conjugation. Let us mention here several familiar problems,
where considerations of positive Lagrangian manifolds are essential. These include
the spectral analysis and resolvent estimates for elliptic quadratic differential oper-
ators [23], [13], the study of spectral instability and pseudospectra for semiclassical
non-normal operators [14], [6], as well as the construction of Gaussian beam quasi-
modes for semiclassical selfadjoint operators of principal type, associated with closed
elliptic trajectories [22], [2].
The work [24] by the third named author introduced and developed the notion of
positivity of a complex Lagrangian space relative to a strictly plurisubharmonic
quadratic weight, which is the starting point for the present work. To recall this
notion, we let Φ0 be a strictly plurisubharmonic quadratic form on Cn and let us
introduce the real linear subspace
ΛΦ0 =(cid:26)(cid:18)x,
2
i
∂Φ0
∂x
(x)(cid:19) , x ∈ Cn(cid:27) ⊂ C2n.
(1.2)
2
We can view ΛΦ0 as the image of the real phase space R2n under a suitable complex
linear canonical transformation, and in particular we notice that ΛΦ0 is maximally
totally real.
In analogy with the discussion above, we say that a complex linear
Lagrangian space Λ ⊂ C2n is positive relative to ΛΦ0 provided that the natural
analog of (1.1) holds,
1
i
σ(ρ, ιΦ0(ρ)) ≥ 0,
ρ ∈ Λ.
(1.3)
Here the map of complex conjugation C has been replaced by the unique antilinear
involution ιΦ0 : C2n → C2n such that ιΦ0ΛΦ0
= 1. A result of [24] establishes a
complete characterization of complex Lagrangians that are positive relative to ΛΦ0
- see also Theorem 2.1 below.
In this work, we shall be mainly concerned with positive complex canonical transfor-
mations. Indeed, the main goal of the present work is to provide a characterization
of positive complex linear canonical transformations relative to plurisubharmonic
weights, and to consider Fourier integral operators (FIOs) in the complex domain
associated to positive canonical transformations, establishing a link between such
operators and Toeplitz operators. In particular, it seems that the point of view of
complex FIOs allows us to shed some new light on some basic questions in the theory
of Toeplitz operators. We would like to emphasize here that the original motivation
for attempting to establish a link between FIOs in the complex domain and Toeplitz
operators came from a talk delivered by the first named author at the conference
"Complex and functional analysis and their interactions with harmonic analysis",
at the Mathematical Research and Conference Center, B¸edlewo, June 2017.
We shall now proceed to define the notion of a complex linear canonical transforma-
tion which is positive relative to a strictly plurisubharmonic quadratic weight, and
to state our main results. In fact, proceeding in the spirit of the discussion above,
it will be more transparent to introduce the notion of positivity relative to a pair
of strictly plurisubharmonic quadratic forms rather than relative to a single one.
Thus, let Φ1, Φ2 be two strictly plurisubharmonic quadratic forms on Cn with the
corresponding antilinear involutions ιΦ1, ιΦ2. Let κ : C2n → C2n be a complex linear
canonical transformation, κ∗σ = σ. We say that κ is positive relative to (ΛΦ1, ΛΦ2)
provided that
1
i(cid:18)σ(κ(ρ), ιΦ1κ(ρ)) − σ(ρ, ιΦ2(ρ))(cid:19) ≥ 0,
ρ ∈ C2n.
(1.4)
The positivity of κ relative to (ΛΦ1, ΛΦ2) is said to be strict provided that the
inequality in (1.4) is strict for all 0 6= ρ ∈ C2n. Let us remark that in the case when
3
the positivity is taken relative to the real phase space R2n, see (1.1), such canonical
transformations were studied in [16], [17], see also the recent works [21], [1].
We can now state the first main result of this work.
Theorem 1.1 Let κ : C2n → C2n be a complex linear canonical transformation
and let Φ1, Φ2 be strictly plurisubharmonic quadratic forms on Cn. The canonical
transformation κ is positive relative to (ΛΦ1, ΛΦ2) precisely when we have
κ(ΛΦ2) = ΛΦ,
(1.5)
where Φ is a strictly plurisubharmonic quadratic form such that Φ ≤ Φ1.
Remark. The definition (1.4) of a positive canonical transformation is a direct
adaptation of the corresponding notion of positivity due to Hormander [16], [17],
to the weighted setting. One advantage of the consideration of the general case
of a pair of weights Φ1, Φ2, is that we can let κ be the identity in (1.4) and get
an invariant notion of the positivity of one plurisubharmonic weight compared to
another, in view of Theorem 1.1.
Our second main result is concerned with applications of Theorem 1.1 to the study
of Toeplitz operators in the Bargmann space
HΦ0(Cn) = L2(Cn, e−2Φ0L(dx)) ∩ Hol(Cn),
where Φ0 is a strictly plurisubharmonic quadratic form on Cn. Specifically, we shall
be concerned with the continuity properties of (in general unbounded) Toeplitz
operators of the form
Top(e2q) = ΠΦ0 ◦ e2q ◦ ΠΦ0 : HΦ0(Cn) → HΦ0(Cn),
(1.6)
where q is a complex-valued quadratic form on Cn and
ΠΦ0 : L2(Cn, e−2Φ0L(dx)) → HΦ0(Cn)
is the orthogonal projection. Sufficient conditions for the boundedness of Top(e2q)
are provided in the following result.
Theorem 1.2 Let Φ0 be a strictly plurisubharmonic quadratic form on Cn and let
q be a quadratic form on Cn such that
2Re q(x) < Φherm(x) := (1/2) (Φ0(x) + Φ0(ix)) ,
x 6= 0
(1.7)
4
and
(1.8)
Let a ∈ C ∞(ΛΦ0) be the Weyl symbol of the Toeplitz operator Top(e2q). Assume that
a ∈ L∞(ΛΦ0). Then the Toeplitz operator
∂x∂x (Φ0 − q) 6= 0.
Top(e2q) : HΦ0(Cn) → HΦ0(Cn)
is bounded.
Remark. Let us remark that Theorem 1.2 is closely related to the conjecture
of [3], [5], stating that a Toeplitz operator is bounded on HΦ0(Cn) precisely when its
Weyl symbol is bounded on ΛΦ0. Theorem 1.2 can therefore be regarded as estab-
lishing the sufficiency part of the conjecture in the special case when the Toeplitz
symbol is of the form exp (2q), where q is a complex valued quadratic form on Cn,
satisfying (1.7), (1.8).
Remark. As we shall see in Section 4, the condition (1.7) guarantees that the
operator Top(e2q) is densely defined. Notice also that the Hermitian form Φherm in
(1.7) is positive definite on Cn, thanks to the strict plurisubharmonicity of Φ0.
The plan of the paper is as follows. In Section 2, we establish the necessity part
of Theorem 1.1, by means of direct geometric arguments, relying on some general
results of [24], see also [4], [12]. The proof of Theorem 1.1 is completed in Section 3,
where we have found it convenient to introduce explicitly a Fourier integral operator
in the complex domain quantizing the canonical transformation κ satisfying (1.5),
when verifying the positivity of κ. Applications to Toeplitz operators are given in
Section 4, where Theorem 1.2 is established. Appendix A is devoted to some elemen-
tary remarks concerning integral representations for linear continuous maps between
weighted spaces of holomorphic functions, which can be regarded as a version of the
Schwartz kernel theorem in this setting. These representations are to be applied in
the main text when deriving a Bergman type representation for our complex FIOs.
Finally, Appendix B, for the use in Section 4, characterizes boundedness properties
of operators given as Weyl quantizations of symbols of the form eiF (x,ξ), where F is
a holomorphic quadratic form on C2n.
Acknowledgements. The second named author would like to express his sincere
and profound gratitude to the Institut de Math´ematiques de Bourgogne at the Uni-
versit´e de Bourgogne for the kind hospitality in August-September 2017, where part
of this project was conducted.
5
2 Positive Lagrangian planes and positive canon-
ical transformations in the HΦ–setting
Let Φ0 be a strictly plurisubharmonic quadratic form on Cn. Associated to Φ0 is
the I-Lagrangian R-symplectic linear manifold ΛΦ0, given by
ΛΦ0 =(cid:26)(cid:18)x,
2
i
∂Φ0
∂x
(x)(cid:19) ; x ∈ Cn(cid:27) ⊂ C2n.
The linear manifold ΛΦ0 is maximally totally real, and we let ιΦ0 be the unique
antilinear involution
such that the restriction of ιΦ0 to ΛΦ0 is the identity. For future reference, we may
recall the explicit description of the involution ιΦ0 given in [12],
ιΦ0 : C2n → C2n,
(2.1)
(2.2)
(2.3)
(2.4)
(cid:18)y,
We also have
0,xxy + Φ′′
2
i (cid:0)Φ′′
ιΦ0 :(cid:18)y,
2
0,x¯x ¯x(cid:1)(cid:19) 7→(cid:18)x,
∂yΨ0(x, y)(cid:19) 7→(cid:18)x,
0,xxx + Φ′′
i (cid:0)Φ′′
∂xΨ0(x, y)(cid:19) ,
2
i
0,x¯x ¯y(cid:1)(cid:19) .
2
i
x × Cn
y , such that Ψ0(x, x) = Φ0(x).
where Ψ0(x, y) is the polarization of Φ0, i.e., the unique holomorphic quadratic form
on Cn
Let Λ ⊂ C2n be a C-Lagrangian space, i.e. a complex linear subspace such that
dimCΛ = n and σΛ = 0. Here σ is the standard symplectic form on C2n. Let us
consider the Hermitian form
b(ν, µ) =
1
i
σ(ν, ιΦ0(µ)),
ν, µ ∈ C2n.
(2.5)
We say that Λ is positive relative to ΛΦ0 if the Hermitian form (2.5) is positive
semidefinite when restricted to Λ,
b(µ, µ) ≥ 0, µ ∈ Λ.
(2.6)
The positivity is said to be strict if the form b in (2.5) is positive definite along Λ. As
remarked in the introduction, this notion is a direct adaptation of the corresponding
notion of positivity due to Hormander [15] where in place of (ΛΦ0, ιΦ0) we have
(R2n,C), with C being the antilinear map of complex conjugation.
Remark. It is easy to see and is established in [4], [12] that the Hermitian form b is
non-degenerate along Λ precisely when Λ and ΛΦ0 are transversal.
Our starting point is the following well known result, see [24], [4], [12].
6
Theorem 2.1 A C-Lagrangian space Λ is positive relative to ΛΦ0 if and only if
Λ = ΛΨ, where Ψ is a pluriharmonic quadratic form such that Ψ ≤ Φ0.
The proof of Theorem 2.1 given in [24], [4], [12] discusses the case of strictly pos-
itive Lagrangian planes only and depends on the general fact that the set of all
C-Lagrangian spaces which are strictly positive relative to ΛΦ0 is a connected com-
ponent in the set of all C-Lagrangian spaces that are transversal to ΛΦ0. Here we
shall give a more direct proof, using the explicit description of the involution ιΦ0,
given in (2.3), (2.4). Let Λ ⊂ C2n be C-Lagrangian, positive relative to ΛΦ0. It
follows from (2.3), as explained in [24], [12], that the fiber {(0, ξ); ξ ∈ Cn} is strictly
negative relative to ΛΦ0, in the sense that the Hermitian form b in (2.5) is negative
definite along the fiber, and therefore Λ is necessarily of the form ξ = ∂xϕ(x), where
ϕ is a holomorphic quadratic form on Cn. It follows that
Λ = ΛΨ,
(2.7)
where Ψ = −Im ϕ is pluriharmonic quadratic. We shall now see that Ψ ≤ Φ0, and
to this end, let us consider the decomposition,
where
Φ0 = Φherm + Φplh,
is positive definite Hermitian and
Φherm(x) = Φ′′
Φplh(x) = Re (cid:0)Φ′′
0,xxx · x
0,xxx · x(cid:1)
(2.8)
(2.9)
(2.10)
(2.11)
(2.12)
is pluriharmonic. Let
A =
2
i
(Φplh)′′
xx =
2
i
(Φ0)′′
xx ,
and let us consider the complex linear "vertical" canonical transformation
We have
κA(y, η) = (y, η + Ay).
κA(ΛΦherm) = ΛΦ0,
and letting ιΦherm be the antilinear involution associated to ΛΦherm, it is then clear
that
ιΦherm = κ−1
A ◦ ιΦ0 ◦ κA.
(2.13)
7
It follows that Λ is positive relative to ΛΦ0 precisely when
κ−1
A (Λ) = ΛΨ−Φplh
is positive relative to ΛΦherm, and when proving Theorem 2.1 we may assume therefore
that the pluriharmonic part of Φ0 vanishes. In this discussion, we are also allowed to
perform complex linear changes of variables in Cn, which correspond to canonical
transformations of the form κC : (y, η) 7→ (C −1y, C tη), where C is an invertible
complex n × n matrix. We have κC(ΛΦ0) = ΛΦ1, Φ1(x) = Φ0(Cx), and it follows
therefore that when establishing Theorem 2.1 it suffices to consider the model case
when
Φ0(x) = x2
2
.
(2.14)
(2.15)
An application of (2.3) shows that the involution ιΦ0 is then given by
(y, η) 7→ (
1
i
η,
1
i
y),
and therefore
b(µ, µ) =
1
i
σ(µ, ιΦ0(µ)) = x2 − ξ2 , µ = (x, ξ) ∈ C2n.
(2.16)
When µ ∈ Λ = ΛΨ, we write ξ = (2/i)∂xΨ(x) = ∂xϕ(x), Ψ(x) = −Im ϕ, where ϕ
is a quadratic holomorphic form, and therefore if Λ is positive relative to ΛΦ0, then
(2.16) shows that
ϕ′′
xxx ≤ x ,
x ∈ Cn ⇐⇒ ϕ′′
xx ≤ 1.
We get
Ψ(x) = −Im ϕ(x) ≤ ϕ′′
xxx · x
2
≤ x2
2
= Φ0(x),
x ∈ Cn.
(2.17)
(2.18)
Conversely, let Λ be C-Lagrangian of the form Λ = ΛΨ, where Ψ is pluriharmonic
quadratic such that Ψ ≤ Φ0. Let us write Ψ = −Im ϕ, where ϕ is a holomorphic
quadratic form. We shall now see that ΛΨ is positive relative to ΛΦ0, and it follows
from the remarks above that it suffices to verify the positivity in the model case
when Φ0 is given by (2.14), so that we have
Writing
Ψ(x) = −Im ϕ(x) ≤ Φ0(x) = x2
2
.
− Im ϕ′′
xxx · x ≤ x2 ,
(2.19)
(2.20)
8
replacing x by eiθx and varying θ ∈ R, we get
xxx · x ≤ x2 ,
ϕ′′
x ∈ Cn.
Next, writing
ϕ′′
xxx · y =
1
4
we get, using (2.21),
(ϕ′′
xx(x + y) · (x + y) − ϕ′′
xx(x − y) · (x − y)) ,
1
ϕ′′
xxx · y ≤
4(cid:0)x + y2 + x − y2(cid:1) =
Replacing x 7→ λ1/2x, y 7→ λ−1/2y, λ > 0, we get
2(cid:18)λ x2 +
ϕ′′
xxx · y ≤
1
1
2(cid:0)x2 + y2(cid:1) .
λ y2(cid:19) ,
1
(2.21)
(2.22)
(2.23)
and choosing λ = y / x, assuming for simplicity that x 6= 0, y 6= 0, we obtain that
ϕ′′
xxx · y ≤ x y .
Hence, ϕ′′
(2.17). The proof of Theorem 2.1 is complete.
xx ≤ 1 and the positivity of ΛΨ relative to ΛΦ0 follows from (2.16),
Remark. Closely related to the proof of Theorem 2.1 given above is the normal
form for strictly plurisubharmonic quadratic forms, given in Lemma 5.1 of [18], see
also [8].
Let Φ1, Φ2 be two strictly plurisubharmonic quadratic forms on Cn and let κ :
C2n → C2n be a complex linear canonical transformation which is positive relative
to (ΛΦ1, ΛΦ2), in the sense of (1.4). In the remainder of this section, we shall establish
the necessity part of Theorem 1.1, while the sufficiency is discussed in Section 3.
To this end, let us observe first that the linear I-Lagrangian R-symplectic manifold
κ(ΛΦ2) is transversal to the fiber {(0, ξ); ξ ∈ Cn}. Indeed, we have in view of (1.4),
(2.24)
1
i
σ(ρ, ιΦ1(ρ)) ≥ 0,
ρ ∈ κ(ΛΦ2),
while, as recalled above, we know from [24], [12] that the fiber is strictly negative
relative to ΛΦ1. It follows that κ(ΛΦ2) = ΛΦ, where Φ is a real quadratic form such
that the Levi form ∂∂Φ is non-degenerate. When verifying that Φ is (necessarily
9
strictly) plurisubharmonic, we claim that it suffices to do so when the pluriharmonic
part of Φ2 vanishes. Indeed, introducing the decomposition (2.8), with the quadratic
form Φ2 in place of Φ0 and considering the canonical transformation κA given in
(2.11), we see, using also (2.13), that κ is positive relative to (ΛΦ1, ΛΦ2) precisely
when κ−1
A ◦ κ◦ κA is positive relative to (ΛΦ1−Φ2,plh, ΛΦ2,herm). Here Φ2,plh and Φ2,herm
are the pluriharmonic and the Hermitian parts of Φ2, respectively. Here it is also
helpful to notice that
ιΦ1−Φ2,plh = κ−1
A ◦ ιΦ1 ◦ κA.
To summarize, if we know that the generating function of the linear I-Lagrangian
R-symplectic manifold
κ−1
A ◦ κ ◦ κA(ΛΦ2,herm)
is plurisubharmonic, then the same property is also enjoyed by the generating func-
tion of κ(ΛΦ2). In what follows we shall assume therefore that
Φ2,xx = Φ2,x x = 0.
(2.25)
As above, in this discussion, we are also allowed to perform complex linear changes
of variables in Cn, which correspond to canonical transformations of the form
(y, η) 7→ (C −1y, C tη), where C is an invertible complex n×n matrix. Such canonical
transformations preserve the plurisubharmonicity of the generating functions, and
similarly to the proof of Theorem 2.1, it suffices therefore to consider the case when
Φ2(x) = x2
2
.
(2.26)
Theorem 2.1 then shows that the C-Lagrangian plane given by {(x, ξ) ∈ C2n; ξ = 0}
is strictly positive relative to ΛΦ2, and therefore κ({(x, ξ) ∈ C2n; ξ = 0}) is strictly
positive relative to ΛΦ1, in view of the positivity of κ. Another application of
Theorem 2.1 gives that
κ({(x, ξ) ∈ C2n; ξ = 0}) = ΛΨ,
(2.27)
where the quadratic form Ψ is pluriharmonic, with Ψ ≤ Φ1.
Let φ(x, y, θ) be a holomorphic quadratic form on Cn
x × Cn
θ , which is a non-
degenerate phase function in the sense of Hormander, generating the graph of κ. It
follows from (2.27), as explained in [4], that the quadratic form
y × CN
Cn × CN ∋ (y, θ) 7→ −Im φ(0, y, θ)
(2.28)
10
is non-degenerate, and since it is pluriharmonic, the signature is necessarily (n +
N, n + N). Recalling that
we see, using [4], that the quadratic form
κ(ΛΦ2) = ΛΦ,
(y, θ) 7→ −Im φ(0, y, θ) + Φ2(y)
(2.29)
(2.30)
is non-degenerate as well. We would like to conclude that the signature of the
quadratic form in (2.30) is also (n + N, n + N), and to that end, we follow [24] and
consider the continuous deformation
[0, 1] ∋ t 7→ −Im φ(0, y, θ) + tΦ2(y).
Using (2.16) we see that
1
i
σ(µ, ιΦ2(µ)) ≥ 0, µ ∈ ΛtΦ2,
0 ≤ t ≤ 1.
(2.31)
(2.32)
It follows as before that the I-Lagrangian manifold κ(ΛtΦ2) is transversal to the fiber,
0 ≤ t ≤ 1, and therefore we conclude that the non-degeneracy of the quadratic forms
in (2.31) is maintained along the deformation 0 ≤ t ≤ 1. Recalling that the set of
non-degenerate quadratic forms of a fixed given signature is a connected component
in the set of all non-degenerate quadratic forms, we conclude that the signature of
the quadratic form in (2.30) is (n+N, n+N). Now, as explained in [4], the quadratic
form Φ in (2.29) is given by
Φ(x) = vcy,θ(−Im φ(x, y, θ) + Φ2(y))
(2.33)
where vcy,θ stands for the critical value with respect to y, θ, and we conclude by the
fundamental lemma of [24] that Φ is plurisubharmonic. (As already observed, the
plurisubharmonicity of Φ is necessarily strict.)
We shall next see that Φ ≤ Φ1, and when doing so it will be convenient the discuss
the following auxiliary result first, which may be of some independent interest.
Proposition 2.2 Let κ : C2n → C2n be a complex linear canonical transforma-
tion which is positive relative to (ΛΦ1, ΛΦ2). If Φ2 is strictly convex then κ has a
generating function ϕ(x, η) which is a holomorphic quadratic form such that
κ : (ϕ′
η(x, η), η) 7→ (x, ϕ′
x(x, η)).
(2.34)
11
Proof: It suffices to show that the map
π : graph(κ) ∋ (x, ξ; y, η) 7→ (x, η) ∈ C2n
is bijective, i.e. injective. Let (0, ξ; y, 0) ∈ Ker(π) so that κ : (y, 0) 7→ (0, ξ). Let us
consider the Hermitian forms,
bj(ν, µ) =
1
i
σ(ν, ιΦj (µ)),
j = 1, 2.
The strict convexity of Φ2 together with Theorem 2.1 implies that
b2((y, 0), (y, 0)) ≍ y2 ,
y ∈ Cn,
(2.35)
and the strict negativity of the fiber with respect to ΛΦ1 gives,
b1((0, ξ), (0, ξ)) ≍ −ξ2 ,
ξ ∈ Cn.
Hence by the positivity of κ, we get
0 ≤ b1((0, ξ), (0, ξ)) − b2((y, 0), (y, 0)) ≍ −(cid:0)ξ2 + y2(cid:1) .
It follows that (y, ξ) = 0 and we conclude that π is injective.
✷
Remark. Assume that the assumptions of Proposition 2.2 hold. The holomorphic
quadratic form ϕ(x, θ)− y · θ is then a non-degenerate phase function generating the
graph of κ.
Let us now turn to the proof of the fact that
Φ ≤ Φ1.
(2.36)
It follows from the remarks above that it suffices to verify (2.36) when the plurihar-
monic part of Φ2 vanishes, and since we are again allowed to perform complex linear
changes of variables in Cn, as before, we conclude that it suffices to consider the
case when Φ2 is given by (2.26). Proposition 2.2 applies and there exists therefore
a holomorphic quadratic form ϕ(x, θ) such that
κ : (ϕ′
θ(x, θ), θ) 7→ (x, ϕ′
x(x, θ)).
(2.37)
We shall now express the positivity of κ relative to (ΛΦ1, ΛΦ2) in terms of the gen-
erating function ϕ. To this end, we shall first obtain an explicit expression for the
Hermitian form
1
i
σ((y, η), ιΦ1(y, η)),
(y, η) ∈ C2n,
12
where we write
Φ1(x) =
1
2
Lx · x + Re (Ax · x), L = 2Φ′′
1,xx, A = Φ′′
1,xx.
(2.38)
Here L is Hermitian positive definite and performing a unitary transformation, we
may assume, for simplicity, that L is diagonal, with real positive diagonal elements.
A simple computation using (2.3) shows that
1
i
σ((y, η), ιΦ1(y, η)) = Ly · y + (2Ay − iη) · x,
(2.39)
where
and therefore we get
Lx = iη − 2Ay,
1
i
σ((y, η), ιΦ1(y, η)) = Ly · y − L−1(2iAy + η) · (2iAy + η).
(2.40)
Using also (2.37), we conclude that κ is positive relative to (ΛΦ1, ΛΦ2) precisely when
L−1(ϕ′
x + 2iAx) · (ϕ′
x + 2iAx) + ϕ′
θ(x, θ)2 ≤ Lx · x + θ2 ,
(x, θ) ∈ C2n.
(2.41)
It is now easy to conclude the proof of the necessity part of Theorem 1.1, using
(2.41). It follows from (2.33) that we can write
Φ(x) = vcy,θ (−Im (ϕ(x, θ) − y · θ) + Φ2(y)) .
At the unique critical point (y(x), θ(x)), we have
y = ϕ′
θ(x, θ),
2
i
∂Φ2
∂y
(y) = θ ⇐⇒ θ =
1
i
y.
(2.42)
(2.43)
(2.44)
Injecting (2.44) into (2.42), we get
Φ(x) = −Im ϕ(x, θ) − θ2
2
,
θ = θ(x),
(2.45)
and in view of (2.38), it suffices therefore to establish the inequality
− 2Im ϕ(x, θ) ≤ Lx · x + θ2 + 2Re(Ax · x),
(x, θ) ∈ C2n.
(2.46)
13
When verifying (2.46), we write, using the Euler homogeneity relation,
2ϕ(x, θ) = ϕ′
x(x, θ) · x + ϕ′
θ(x, θ) · θ,
and therefore,
− 2Im ϕ(x, θ) = −Im ((ϕ′
x(x, θ) + 2iAx) · x + ϕ′
θ(x, θ) · θ) + 2Re(Ax · x).
(2.47)
(2.48)
An application of the Cauchy-Schwarz inequality with respect to the positive definite
Hermitian forms (x, y) 7→ L−1x · y, (x, y) 7→ x · y together with the inequality
ab ≤ a2/2 + b2/2 allows us to conclude that the first term in the right hand side of
(2.48) does not exceed
1
2(cid:16)L−1(ϕ′
x + 2iAx) · (ϕ′
x + 2iAx) + Lx · x + ϕ′
θ(x, θ)2 + θ2(cid:17) .
The inequality (2.46) follows, in view of (2.41). The proof of the necessity part of
Theorem 1.1 is complete.
Remark. In the context of Theorem 1.1, assume that Φ1 = Φ2 =: Φ0 and let us
write
Φ0(x) = supy∈Rn (−Im ϕ(x, y)) ,
where ϕ(x, y) is a holomorphic quadratic form on Cn
and Im ϕ′′
yy > 0. In the special case when Φ0 is given by (2.26), we can take
y , such that det ϕ′′
(2.49)
xy 6= 0
ϕ(x, y) = i(cid:18) x2
2
+ √2x · y +
The complex canonical transformation
x × Cn
2 (cid:19) .
y2
κϕ : C2n ∋ (y,−ϕ′
y(x, y)) 7→ (x, ϕ′
x(x, y)) ∈ C2n
(2.50)
maps R2n bijectively onto ΛΦ0, see [12], and it exchanges the complex conjugation
map C and the involution ιΦ0. Setting
eκ = κ−1
we see that the complex linear canonical transformation eκ is positive in the sense
ϕ ◦ κ ◦ κϕ,
of [17],
(2.51)
1
i(cid:18)σ(eκ(ρ),Ceκ(ρ)) − σ(ρ,C(ρ))(cid:19) ≥ 0,
ρ ∈ C2n.
(2.52)
14
An application of Proposition 5.10 of [17] allows us to conclude therefore that the
map eκ enjoys the following factorization,
where eκ1 and eκ3 are real linear canonical maps and the map eκ2 is of the form
where eq is a quadratic form with Reeq ≥ 0 on R2n - see also the discussion in the
eκ = eκ1 ◦ eκ2 ◦ eκ3,
eκ2 = exp (−iHeq)
proof of Proposition 5.12 of [17]. We obtain the factorization
(2.53)
(2.54)
where we have
and
κ = κ1 ◦ κ2 ◦ κ3,
κj : ΛΦ0 → ΛΦ0,
j = 1, 3,
(2.55)
(2.56)
(2.57)
where q is a holomorphic quadratic form on C2n such that Re q ≥ 0 along ΛΦ0. The
representation (2.55) can be used to give an alternative proof of the basic inequality
Φ ≤ Φ0 in Theorem 1.1, in this special case.
κ2 = exp (−iHq),
3 Positivity and Fourier integral operators
The purpose of this section is to establish the sufficiency part of Theorem 1.1. To
this end, let Φ1, Φ2 be two strictly plurisubharmonic quadratic forms on Cn and let
κ : C2n → C2n be a complex linear canonical transformation. Assume that
where Φ is a strictly plurisubharmonic quadratic form such that
κ(ΛΦ2) = ΛΦ,
Φ ≤ Φ1.
(3.1)
(3.2)
We shall establish the positivity of κ relative to (ΛΦ1, ΛΦ2) by making a judicious
choice of a non-degenerate phase function generating the graph of κ, and to this end,
it will be convenient to consider a metaplectic Fourier integral operator associated
y × CN
to κ. Let therefore ϕ(x, y, θ) be a holomorphic quadratic form on Cn
θ ,
x × Cn
15
which is a non-degenerate phase function in the sense of Hormander, generating the
graph of κ. It follows from [4] that the plurisubharmonic quadratic form
Cn × CN ∋ (y, θ) 7→ −Im ϕ(0, y, θ) + Φ2(y)
(3.3)
is non-degenerate of signature (n + N, n + N). We conclude, following [24], [4] that
the Fourier integral operator
Au(x) =ZZ eiϕ(x,y,θ)au(y) dy dθ,
a ∈ C,
(3.4)
quantizing κ, can be realized by means of a good contour and we obtain a bounded
linear map,
Here
A : HΦ2(Cn) → HΦ(Cn).
(3.5)
HΦ2(Cn) = Hol(Cn) ∩ L2(Cn, e−2Φ2L(dx)),
with HΦ(Cn) having an analogous definition.
We shall now discuss a Bergman type representation of the bounded operator in
(3.5), see also [19] for a related discussion. To this end, let us recall from Theorem
A.1 that we can write
(3.6)
Here the kernel KA(x, z) is holomorphic on Cn
Au(x) =Z KA(x, y)u(y) e−2Φ2(y) L(dy) =: eAu(x).
x × Cn
y 7→ K(x, y) ∈ HΦ2(Cn),
z , with
uniquely determined by (3.6). If u ∈ L2
HΦ2(Cn), we see from (3.6) that eAu = 0. Hence the operator eA in (3.6) is a well
defined linear continuous map
Φ2(Cn) = L2(Cn, e−2Φ2L(dx)) is orthogonal
Φ2(Cn) → HΦ2(Cn).
eA : L2
KA(x, y)e−2Φ2(y) =(cid:16)eAδy(cid:17) (x),
Furthermore, eA extends to a map: E ′(Cn) → Hol(Cn) and we have
Φ2(Cn) → HΦ2(C) be
where δy ∈ E ′(Cn) is the delta function at y. Let next Π2 : L2
the orthogonal projection and let us recall from [12] that the operator Π2 is given
by
(3.7)
Π2u(x) = a2Z e2Ψ2(x,y)−Φ2(y)u(y) L(dy),
a2 > 0.
(3.8)
16
Here Ψ2 is the polarization of Φ2, i.e. a holomorphic quadratic form on C2n
x,y such
that Ψ2(x, x) = Φ2(x). We get eAδy = eAΠ2δy = AΠ2δy, and it follows from (3.7)
that
KA(x, y) = A(a2e2Ψ2(·,y))(x).
(3.9)
From [12], let us recall the basic property,
2Re Ψ2(x, y) − Φ2(x) − Φ2(y) ∼ −x − y2 ,
on Cn
x × Cn
y , and in particular we have
It follows that
2Re Ψ2(x, y) ≤ Φ2(x) + Φ2(y).
(3.10)
(3.11)
− Im ϕ(0,ey, θ) + 2Re Ψ2(ey, 0) ≤ −Im ϕ(0,ey, θ) + Φ2(ey).
Here, as observed in (3.3), the right hand side is a non-degenerate plurisubharmonic
quadratic form of signature (n + N, n + N), and since the left hand side is pluri-
harmonic, we conclude that it is also non-degenerate of signature (n + N, n + N).
Writing
−Im ϕ(0,ey, θ) + 2Re Ψ2(ey, 0) = Re (iϕ(0,ey, θ) + 2Ψ2(ey, 0)) ,
we conclude that the holomorphic quadratic form
is non-degenerate. It follows that the holomorphic function
Cn × CN ∋ (ey, θ) 7→ iϕ(0,ey, θ) + 2Ψ2(ey, 0)
Cn × CN ∋ (ey, θ) 7→ iϕ(x,ey, θ) + 2Ψ2(ey, z)
has a unique critical point which is non-degenerate, for each (x, z) ∈ Cn × Cn. An
application of exact (quadratic) stationary phase allows us therefore to conclude
that
Here Ψ(x, z) is a holomorphic quadratic form on C2n given by
KA(x, y) =bae2Ψ(x,y),
ba ∈ C.
2Ψ(x, z) = vcey,θ (iϕ(x,ey, θ) + 2Ψ2(ey, z)) .
Let us now make the following basic observation.
17
(3.12)
(3.13)
Proposition 3.1 The holomorphic quadratic form Ψ(x, z) given in (3.13) satisfies
2Re Ψ(x, y) ≤ Φ(x) + Φ2(y),
(x, y) ∈ Cn
x × Cn
y .
Proof: It will be more convenient to verify that
2Re Ψ(x, y) ≤ Φ(x) + Φ∗
2(y),
(x, y) ∈ Cn
x × Cn
y ,
where Φ∗
2(y) = Φ2(y). A direct calculation shows that
or equivalently,
2
i
∂yΦ∗
2(y) = −
2
i
(∂yΦ2)(y),
2
i
∂y(Φ∗
2)(y) = −
2
i
(∂yΦ2)(y).
It follows that the antilinear involution
Γ : C2n ∋ (y, η) 7→ (y,−η) ∈ C2n
maps ΛΦ2 bijectively onto ΛΦ∗
2. We conclude in view of (3.1) that
κ ◦ Γ : ΛΦ∗
2 → ΛΦ,
(3.14)
(3.15)
(3.16)
(3.17)
(3.18)
and let us consider the graph of the map in (3.17), Graph(κ◦ Γ)∩(cid:0)ΛΦ × ΛΦ∗
ΛΦ×ΛΦ∗
form
2(y) is I-Lagrangian and R-symplectic for the standard symplectic
2(cid:1). Here
2 = ΛΦ(x)+Φ∗
dξ ∧ dx + dη ∧ dy,
x,ξ × C2n
y,η and we claim that Graph(κ ◦ Γ) ∩(cid:0)ΛΦ × ΛΦ∗
on C2n
the symplectic form in (3.18), restricted to ΛΦ × ΛΦ∗
computation: when (t, s) ∈ ΛΦ∗
form on C2n,
2×ΛΦ∗
2. This can be seen by a direct
2 we have, writing σ for the standard symplectic
2(cid:1) is Lagrangian for
σ(κ(Γ(t)), κ(Γ(s))) + σ(t, s) = σ(Γ(t), Γ(s)) + σ(t, s) = −σ(t, s) + σ(t, s) = 0,
since σ(t, s) is real. Here we have also used that, by a straightforward computation,
σ(Γt, Γs) = −σ(t, s).
(3.19)
It is then well known that πx,y(cid:0)Graph(κ ◦ Γ) ∩(cid:0)ΛΦ × ΛΦ∗
Γ) ∩ (ΛΦ × ΛΦ∗
x,y, is maximally totally real, see [19].
2) in C2n
2(cid:1)(cid:1), the projection of Graph(κ◦
18
We now come to check (3.15). To this end, we observe that (3.13) gives
and
where
2∂xΨ(x, y) = i∂xϕ(x,ey, θ)
2∂yΨ(x, y) = 2∂yΨ2(ey, y),
2
i
(3.20)
(3.21)
(3.22)
∂θϕ(x,ey, θ) = 0,
∂eyϕ(x,ey, θ) +
∂eyΨ2(ey, y) = 0.
We shall consider (3.20), (3.21) at the points (x, y) ∈ πx,y(cid:0)Graph(κ ◦ Γ) ∩ ΛΦ × ΛΦ∗
2(cid:1),
which corresponds to ey = y in (3.22). Using (3.22) together with the fact that
and (3.21) together with the fact that
∂eyΨ2(ey,ey) = ∂eyΦ2(ey),
(∂yΨ2) (y, y) = ∂yΦ∗
2(y),
we conclude that at the points
(x, y) ∈ πx,y(cid:0)Graph(κ ◦ Γ) ∩ ΛΦ × ΛΦ∗
2(cid:1) ,
the following equalities hold,
∂xΨ(x, y) = ∂xΦ(x),
∂yΨ(x, y) = ∂yΦ∗
2(y).
(3.23)
In other words,
∂x (Φ(x) − 2Re Ψ(x, y)) = ∂y (Φ∗
2(y) − 2Re Ψ(x, y)) = 0,
tion
along πx,y(cid:0)Graph(κ ◦ Γ) ∩ ΛΦ × ΛΦ∗
vanishes on πx,y(cid:0)Graph(κ ◦ Γ) ∩ ΛΦ × ΛΦ∗
2(cid:1), and thus the gradient of the real valued func-
2(y) − 2Re Ψ(x, y)
2(cid:1). It follows that the strictly plurisubhar-
monic quadratic form F (x, y) vanishes to the second order along
F (x, y) = Φ(x) + Φ∗
(3.24)
πx,y(cid:0)Graph(κ ◦ Γ) ∩ ΛΦ × ΛΦ∗
2(cid:1) ,
(3.25)
and since the latter is maximally totally real, we get F ≥ 0, thus implying (3.15).
✷
19
Remark. The strictly plurisubharmonic quadratic form F (x, y) in (3.24) vanishes to
the second order along the maximally totally real subspace (3.25), and therefore the
conclusion that F ≥ 0 can be strengthened to the following,
F (x, y) ≍ dist(cid:0)(x, y), πx,y(cid:0)Graph(κ ◦ Γ) ∩ ΛΦ × ΛΦ∗
2(cid:1)(cid:1)2 .
Let us now return to the Bergman type representation of the Fourier integral oper-
ator A in (3.4) quantizing κ. Combining (3.6) and (3.12), we get
Au(x) =ZZ ae2(Ψ(x,y)−Φ2(y))u(y)dydy,
for some a ∈ C. This can be viewed as a Fourier integral operator
Au(x) =ZZ ae2(Ψ(x,θ)−Ψ2(y,θ))u(y)dy dθ,
where we take the integration contour θ = y in (3.27).
Since ∂y∂θΨ2(y, θ) is non-degenerate, the phase function
φ(x, y, θ) =
2
i
(Ψ(x, θ) − Ψ2(y, θ))
(3.26)
(3.27)
(3.28)
is non-degenerate in the sense of Hormander, and the canonical transformation κ
takes the form
κ : (cid:18)y,
∂yΨ2(y, θ)(cid:19) 7→(cid:18)x,
2
i
∂xΨ(x, θ)(cid:19) , with ∂θΨ(x, θ) = ∂θΨ2(y, θ).
2
i
(3.29)
We may also notice here that if we define
κΨ : (θ,−(2/i)∂θΨ(y, θ)) 7→ (y, (2/i)∂yΨ(y, θ))
and κΨ2 similarly, then κ = κΨ ◦ κ−1
Ψ2.
The discussion so far shows that the canonical transformation κ enjoying the map-
ping properties (3.1), (3.2), admits a non-degenerate phase function of the form
(3.28), where the quadratic form Ψ satisfies
2Re Ψ(x, y) ≤ Φ1(x) + Φ2(y),
(x, y) ∈ Cn
x × Cn
y .
(3.30)
The positivity of κ relative to (ΛΦ1, ΛΦ2) is then implied by the following general
result.
20
Proposition 3.2 Let κ be a canonical transformation satisfying (3.1) and let us
consider a metaplectic Fourier integral operator of the form (3.26), or equivalently
(3.27), associated to κ. Then the following conditions are equivalent:
(i) κ is positive relative to (ΛΦ1, ΛΦ2) in the sense of (1.4),
1
i
σ(t1, ιΦ1t1) −
1
i
σ(t2, ιΦ2t2) ≥ 0, whenever t1 = κ(t2),
t2 ∈ C2n.
(3.31)
(ii) Λ2Re Ψ(x,y) is positive relative to ΛΦ1(x)+Φ2(y),
x × Cn
(iii) 2Re Ψ(x, y) − Φ1(x) − Φ2(y) ≤ 0 on Cn
y .
Proof: The equivalence (ii)⇔(iii) follows from Theorem 2.1, so it suffices to show
the equivalence (i)⇔(ii).
Clearly, (iii) is equivalent to
2Re Ψ(x, y) − Φ1(x) − Φ∗
2(y) ≤ 0 on C2n
x,y,
(3.32)
where Φ∗
2(y) = Φ2(y)(= Φ2(y)), and by Theorem 2.1 (ii) is equivalent to
Λ2Re Ψ(x,y) is positive relative to ΛΦ1(x)+Φ∗
2(y).
(3.33)
We have,
Λ2Re Ψ = {(x, (2/i)∂x2Re Ψ(x, y); y, (2/i)∂y2Re Ψ(x, y)}
= {(x, (2/i)∂xΨ(x, y); y, (2/i)∂yΨ(x, y)},
(3.34)
and (3.33) means that
1
i
σ(t1, ιΦ1t1) +
1
i
σ(t2, ιΦ∗
2t2) ≥ 0, for all (t1, t2) ∈ Λ2Re Ψ.
(3.35)
Here, we shall relate the involutions ιΦ∗
given by
2 and ιΦ2. From (2.4) let us recall that ιΦ2 is
ιΦ2 : (cid:16)y, (2/i)∂yΨ2(x, y)(cid:17) 7→ (x, (2/i)∂xΨ2(x, y)) .
(3.36)
We also know that the antilinear involution Γ, given in (3.16), maps ΛΦ2 bijectively
onto ΛΦ∗
2 are the unique antilinear maps equal to the identity on
ΛΦ2 and ΛΦ∗
2 respectively, it follows that
2, and since ιΦ2, ιΦ∗
ιΦ∗
2 = ΓιΦ2Γ.
(3.37)
21
From (3.19), let us recall that
1
i
σ(Γt, Γs) =
1
i
σ(t, s),
so using (3.37), we find that the second term in (3.35) is equal to
1
i
σ(t2, ΓιΦ2Γt2) =
1
i
σ(Γt2, ιΦ2Γt2) =
1
i
σ(Γt2, ιΦ2Γt2) = −
1
i
σ(ιΦ2Γt2, Γt2),
where we also used the fact that (1/i)σ(t, ιΦ2t) is real. Hence (3.33) is equivalent,
via (3.35), to
1
i
σ(t1, ιΦ1t1) −
1
i
σ(ιΦ2Γt2, Γt2) ≥ 0, ∀(t1, t2) ∈ Λ2Re Ψ.
(3.38)
From (3.36), we get
i.e.
ιΦ2Γ : (cid:18)y,−
ιΦ2Γ : (cid:18)θ,
2
i
2
i
∂yΨ2(x, y)(cid:19) 7→(cid:18)x,
∂θΨ2(y, θ)(cid:19) 7→(cid:18)y,
2
i
∂xΨ2(x, y)(cid:19) ,
∂yΨ2(y, θ)(cid:19) ,
2
i
(3.39)
where we changed the notation slightly for convenience.
Write,
Λ2Re Ψ ∋ (t1, t2) =(cid:18)x,
2
i
∂xΨ(x, θ); θ,
∂θΨ(x, θ)(cid:19) ,
2
i
and put t3 = ιΦ2Γt2, so that by (3.39),
∂yΨ2(y, θ)(cid:19) ,
2
i
t3 =(cid:18)y,
∂θΨ(x, θ)(cid:19) =(cid:18)θ,
∂θΨ2(y, θ)(cid:19) .
2
i
where
(cid:18)θ,
2
i
Comparing with (3.29), we see that t1 = κ(t3). Since Γt2 = ι2
that (3.38) is equivalent to
Φ2Γt2 = ιΦ2t3, we see
1
i
σ(t1, ιΦ1t1) −
1
i
σ(t3, ιΦ2t3) ≥ 0, when t1 = κ(t3),
(3.40)
22
which is precisely (3.31) up to a change of notation. This completes the proof of the
equivalence (i)⇔(ii) and of the proposition.
✷
Combining Proposition 3.1 and Proposition 3.2, we see that the proof of the suffi-
ciency part of Theorem 1.1 is now complete.
Remark: Let κ : C2n → C2n be a complex linear canonical transformation, such
that (3.1) holds, where Φ2, Φ are strictly plurisubharmonic. It follows from (3.23)
that the holomorphic quadratic form Ψ(x, y) depends only on κ and on the weights
Φ2, Φ, but not on the choice of a non-degenerate phase function ϕ(x, y, θ), (x, y, θ) ∈
Cn
x × Cn
θ , such that
y × CN
Λ′
ϕ = Graph(κ),
where
Λ′
ϕ = {(x, ϕ′
x(x, y, θ); y,−ϕ′
It follows that if ψ(x, y, w), (x, y, w) ∈ Cn
phase function such that
y(x, y, θ)); ϕ′
y × CN ′
x × Cn
θ(x, y, θ) = 0}.
w , is a second non-degenerate
Λ′
ϕ = Λ′
ψ = Graph(κ),
then both ϕ and ψ give rise to the same Fourier integral operators, realized as
bounded linear maps: HΦ2(Cn) → HΦ(Cn).
We shall finish this section by making some remarks concerning metaplectic Fourier
integral operators in the complex domain, associated to canonical transformations
that are strictly positive relative to (ΛΦ1, ΛΦ2). Let
κ : C2n → C2n
(3.41)
be a complex linear canonical transformation which is strictly positive relative to
(ΛΦ1, ΛΦ2). According to Theorem 1.1, we then have
κ(ΛΦ2) = ΛΦ,
where Φ is a strictly plurisubharmonic quadratic form on Cn such that
Let
Φ1(x) − Φ(x) ≍ x2 ,
x ∈ Cn.
T u(x) =ZZ eiφ(x,y,θ)au(y) dy dθ,
a ∈ C,
(3.42)
(3.43)
23
be a Fourier integral operator associated to κ. As discussed above,
it follows
from [4], [24] that the operator T can be realized by means of a suitable good
contour and we then obtain a bounded operator
T : HΦ2(Cn) → HΦ(Cn).
(3.44)
It follows from (3.43) that the inclusion map HΦ(Cn) → HΦ1(C) is compact, and the
operator T : HΦ2(Cn) → HΦ1(Cn) is therefore compact. The following sharpening
is essentially well known, see [1].
Proposition 3.3 The operator
is of trace class, with the singular values sj(T ) satisfying
T : HΦ2(Cn) → HΦ1(Cn)
sj(T ) = O(j−∞).
(3.45)
Proof: Let q be a holomorphic quadratic form on C2n such that its restriction to ΛΦ1
is real positive definite. Let us introduce the Weyl quantization of q, the operator
Q = qw(x, Dx). The quadratic differential operator Q is selfadjoint on HΦ1(Cn) with
discrete spectrum, and let us consider the metaplectic Fourier integral operator etQ,
0 ≤ t ≤ t0 ≪ 1, acting on the space HΦ(Cn). Using some well known arguments,
explained in detail in [9], [10], [11], we see that for t ∈ [0, t0] with t0 > 0 small
enough, the operator etQ is bounded,
etQ : HΦ(Cn) → HΦt(Cn),
(3.46)
where Φt is a strictly plurisubharmonic quadratic form on Cn, depending smoothly
on t ≥ 0 small enough, such that
Φt(x) = Φ(x) + O(t) x2 .
(3.47)
Combining this observation with (3.43) we conclude that there exists δ > 0 small
enough such that the operator
is bounded. Writing
eδQT : HΦ2(Cn) → HΦ1(Cn)
T = e−δQeδQT,
(3.48)
(3.49)
24
and applying the Ky Fan inequalities, we get
sj(T ) ≤ sj(e−δQ) eδQT L(HΦ2 ,HΦ1 ) = O(j−∞).
Here we have also used the fact that the singular values of the compact positive
selfadjoint operator e−δQ on HΦ1(Cn) satisfy
It follows that T is of trace class and the proof of the proposition is complete. ✷
sj(e−δQ) = O(j−∞).
4 Applications to Toeplitz operators
The purpose of this section is to apply the point of view of Fourier integral operators
in the complex domain, developed in the previous sections, to the study of Toeplitz
operators in the Bargmann space, establishing Theorem 1.2.
Let Φ0 be a strictly plurisubharmonic quadratic form on Cn and let p : Cn → C be
measurable. Associated to p is the Toeplitz operator
Top(p) = ΠΦ0 ◦ p ◦ ΠΦ0 : HΦ0(Cn) → HΦ0(Cn).
(4.1)
Here
ΠΦ0 : L2(Cn, e−2Φ0L(dx)) → HΦ0(Cn)
is the orthogonal projection. We shall always assume that when equipped with the
natural domain
D(Top(p)) = {u ∈ HΦ0(Cn); pu ∈ L2(Cn, e−2Φ0L(dx))},
(4.2)
the operator Top(p) becomes densely defined.
For future reference, let us recall the link between the Toeplitz and Weyl quantiza-
tions on Cn. Let p ∈ L∞(Cn), say. Then we have
Top(p) = aw(x, Dx),
(4.3)
where a ∈ C ∞(ΛΦ0) is given by
a(cid:18)x,
2
i
∂Φ0
∂x
(x)(cid:19) =(cid:18)exp (cid:18)1
4(cid:0)Φ′′
0,xx(cid:1)−1 ∂x · ∂x(cid:19) p(cid:19) (x),
x ∈ Cn.
(4.4)
25
operator on Cn whose symbol is the positive definite quadratic form
∂x·∂x is a constant coefficient second order differential
0,xx(cid:1)−1
See [7], [25]. Here −(cid:0)Φ′′
0,xx(cid:1)−1
4(cid:0)Φ′′
1
ξ · ξ > 0,
0 6= ξ ∈ Cn ≃ R2n,
and therefore the operator in (4.4) can be regarded as the forward heat flow acting
on p.
In this section we shall be concerned with the question of when an operator of the
form Top(p) is bounded,
Top(p) ∈ L(HΦ0(Cn), HΦ0(Cn)),
and following [3], in doing so we shall only consider Toeplitz symbols of the form
p = e2q,
(4.5)
where q is a complex-valued quadratic form on Cn. Let us first proceed to give an
explicit criterion, guaranteeing that when equipped with the domain (4.2), the oper-
ator Top(e2q) is densely defined. Recalling the decomposition (2.8) and considering
the unitary map
HΦ0(Cn) ∋ u 7→ ue−f ∈ HΦherm(Cn),
f (x) = Φ′′
0,xxx · x,
we may observe that the space efP(Cn) = {ef p; p ∈ P(Cn)} is dense in HΦ0(Cn).
Here P(Cn) is the space of holomorphic polynomials on Cn. It follows that
so that Top(e2q) is densely defined, provided that
efP(Cn) ⊂ D(Top(e2q)),
2Re q(x) < Φherm(x),
(4.6)
in the sense of quadratic forms on Cn.
Recalling (3.8), we may write
Top(e2q)u(x) = CZ e2(Ψ0(x,y)−Φ0(y))e2q(y,y)u(y) dy dy,
u ∈ D(Top(e2q)).
(4.7)
Here C > 0 and Ψ0 is the polarization of Φ0. Similarly to (3.27), we get
Top(e2q)u(x) = CZZΓ
e2(Ψ0(x,θ)−Ψ0(y,θ)+q(y,θ))u(y) dy dθ,
(4.8)
26
where Γ is the contour in C2n, given by θ = y. Here the holomorphic quadratic form
F (x, y, θ) =
2
i
(Ψ0(x, θ) − Ψ0(y, θ) + q(y, θ))
(4.9)
is a non-degenerate phase function in the sense of Hormander, in view of the fact
that det Ψ′′
0,xθ 6= 0, and therefore the operator Top(e2q) in (4.8) can be viewed as
a metaplectic Fourier integral operator associated to a suitable canonical relation
⊂ C2n × C2n. We have the formal factorization
Top(e2q) = AB,
where
Av(x) =Z e2Ψ0(x,θ)v(θ) dθ, Bu(θ) =Z e−2 eΨ0(y,θ)u(y) dy,
(4.10)
and where we have written eΨ0(y, θ) = Ψ0(y, θ) − q(y, θ). Here the operator A,
formally, is an elliptic Fourier integral operator associated to the canonical transfor-
mation
(θ,−
∂θΨ0(x, θ)) 7→ (x,
2
i
2
i
∂xΨ0(x, θ)).
It follows that the canonical relation associated to Top(e2q) is the graph of a canon-
ical transformation if and only if this is the case for the Fourier integral operator
B. We conclude that the operator Top(e2q) in (4.8) is associated to a canonical
transformation precisely when
The condition (4.11) is equivalent to the assumption (1.8) in Theorem 1.2. The
canonical transformation is then given by
κ : (y,−∂yF (x, y, θ)) 7→ (x, ∂xF (x, y, θ)),
∂θF (x, y, θ) = 0.
(4.12)
Example. In the following discussion, we shall revisit the family of examples dis-
cussed in Section 6 of [3] and show how the point of view of Fourier integral operators
in the complex domain, developed above, allows one to recover the main findings of
Section 6 in [3], obtained there by means of a direct computation.
Let Φ0(x) = (1/2) x2 and q = (λ/2) y2, λ ∈ C with Re λ < 1/2. Here the
restriction on Re λ implies that (4.6) holds, so that the operator Top(e2q) is densely
defined in HΦ0(Cn). We have
Ψ0(x, y) =
1
2
x · y,
27
∂y∂θeΨ0 6= 0.
(4.11)
and the phase function F in (4.9) is given by
F (x, y, θ) =
2
i (cid:18) 1
2
x · θ −(cid:18)1 − λ
2 (cid:19) y · θ(cid:19) .
(4.13)
In particular, the condition (4.11) is satisfied and we may then compute the canonical
transformation κ associated to the corresponding Fourier integral operator Top(e2q)
in (4.8).
The critical set CF of the phase F is given by ∂θF = 0 ⇐⇒ x = (1 − λ)y, and the
corresponding canonical transformation κ is of the form
κ : (y,−∂yF (x, y, θ)) 7→ (x, ∂xF (x, y, θ)),
(x, y, θ) ∈ CF .
(4.14)
It follows that κ is given by
κ : (y, η) 7→(cid:18)(1 − λ)y,
η
1 − λ(cid:19) .
(4.15)
We shall now determine when the canonical transformation κ is positive relative
to ΛΦ0, which can be done by a direct computation: it follows from (2.4) that the
involution ιΦ0 is given by
ιΦ0 : (y, η) 7→(cid:18)1
i
η,
y(cid:19) ,
1
i
and therefore, we may compute,
1
i
σ(κ(y, η), ιΦ0κ(y, η)) =
σ(cid:18)((1 − λ)y,
1
i
η
1 − λ
η
,
), (
1
i
1
i
(1 − λ)y)(cid:19)
1 − λ
= 1 − λ2 y2 − η2
1 − λ2 .
Similarly, we have
1
i
σ((y, η), ιΦ0(y, η)) = y2 − η2 .
(4.16)
(4.17)
(4.18)
Combining (4.17), (4.18) we see that the κ is positive relative to ΛΦ0 if and only if
1 − λ ≥ 1.
(4.19)
This condition occurs in [3], pp. 581, 582 (with the inessential difference that in the
discussion in [3] one considers Φ0(x) = x2 /4), where it is verified that the operator
Top(e2q) ∈ L(HΦ0(Cn), HΦ0(Cn)) precisely when (4.19) holds.
28
In the case when the strict inequality holds in (4.19), the canonical transformation
κ in (4.15) is strictly positive relative to ΛΦ0 and it follows from Proposition 3.3 that
the Toeplitz operator Top(e2q) is of trace class on HΦ0(Cn).
We shall now proceed to discuss the "boundary" case when
1 − λ = 1.
(4.20)
In this case, using (4.15) we immediately see that κ(ΛΦ0) = ΛΦ0, and therefore we
conclude, in view of [4], [24], that the operator
Top(e2q) : HΦ0(Cn) → HΦ0(Cn)
(4.21)
is bounded, with a bounded two-sided inverse.
We claim next that the operator in (4.21) is in fact unitary when (4.20) holds, and
when verifying the unitarity, it will be convenient to pass to the Weyl quantization,
computing the Weyl symbol of Top(e2q). It follows from (4.4) that
2
i
∂Φ0
∂x
(x)(cid:19) =(cid:18)exp (cid:18)1
a(cid:18)x,
Here ∆ is the Laplacian on Cn ≃ R2n. Computing the Gaussian integral in (4.22)
by the exact version of stationary phase, we get, see also [3],
∆(cid:19) e2q(cid:19) (x) =(cid:18) 2
π(cid:19)nZCn
L(dy). (4.22)
e−2x−y2
eλy2
8
a(cid:18)x,
2
i
∂Φ0
∂x
Here we may notice that
(x)(cid:19) =(cid:18) 2
Re(cid:18) 2λ
2 − λ(cid:19)n
2 − λ(cid:19) = 0,
exp (cid:18) 2λ
2 − λ x2(cid:19) .
(4.23)
when (4.20) holds, reflecting the fact that the associated canonical transformation
in (4.15) is "real" in this case. We conclude that the Weyl symbol of the Toeplitz
operator Top(e2q) is given by
so that
exp (iF (x, ξ)), F (x, ξ) =
a(x, ξ) =(cid:18) 2
2 − λ(cid:19)n
Top(e2q) =(cid:18) 2
2 − λ(cid:19)n
2λ
2 − λ
x · ξ,
(4.24)
(4.25)
(exp (iF ))w .
29
= 0 and an application of Proposition 5.11 of [17] together with
We have (Im F )ΛΦ0
the metaplectic invariance of the Weyl quantization allows us to conclude that the
operator
pdet(I − F /2) (exp (iF ))w : HΦ0(Cn) → HΦ0(Cn)
(4.26)
is unitary. Here F is the Hamilton map of F , i.e. the matrix of the (linear) Hamilton
field HF , and it remains therefore to check that
pdet(I − F /2) =(cid:18) 2
2 − λ(cid:19)n
eiθ,
θ ∈ R.
(4.27)
To this end, we compute using (4.24),
1
2F =
λ
2 − λ(cid:18)1
0 −1(cid:19) ,
0
1
2F =
I −
2
1(cid:19) ,
2 − λ(cid:18)1 − λ 0
0
and (4.27) follows, thanks to (4.20). We conclude therefore that the Toeplitz op-
erator Top(e2q) is unitary on HΦ0(Cn), when Re λ < 1/2 and (4.20) holds. The
unitarity property has also been observed in [3].
Remark. In the case when Re λ < 1/2, 1 − λ > 1, we observed that the operator
Top(e2q) is of trace class on HΦ0(Cn), and we get, using (4.24) and the metaplectic
invariance of the Weyl quantization,
tr Top(e2q) =
1
(2π)nZZΛΦ0
a
(σΛΦ0
n!
)n
,
where a is given in (4.24).
We are now ready to discuss the proof of Theorem 1.2. It follows from Theorem
1.1 and the discussion in this section that it suffices to check that the canonical
transformation (4.12) associated to the operator Top(e2q) is positive relative to ΛΦ0.
To this end, let us consider the Weyl symbol of Top(e2q), given by (4.4),
a(x, ξ) =(cid:18)exp (cid:18)1
4(cid:0)Φ′′
0,xx(cid:1)−1
∂x · ∂x(cid:19) e2q(cid:19) (x),
(x, ξ) ∈ ΛΦ0.
(4.28)
A simple computation of the inverse Fourier transform of a real Gaussian shows that
a(x, ξ) = CΦ0ZCn
exp (−4Φherm(x − y))e2q(y) L(dy), CΦ0 6= 0.
(4.29)
30
Here the convergence of the integral in (4.29) is guaranteed by (4.6). In view of the
exact version of stationary phase, it is therefore clear that
a(x, ξ) = Cexp (iF (x, ξ)),
(x, ξ) ∈ ΛΦ0,
(4.30)
for some constant C 6= 0, where F is a holomorphic quadratic form on C2n. Propo-
sition B.1 shows that the positivity of κ in (4.12) relative to ΛΦ0 is equivalent to
the fact that the Weyl symbol in (4.30) is such that Im FΛΦ0 ≥ 0 ⇐⇒ exp (iF ) ∈
L∞(ΛΦ0). The proof of Theorem 1.2 is complete.
A Schwartz kernel theorem in the HΦ–setting
In this appendix we shall make some elementary remarks concerning integral rep-
resentations for linear continuous maps between weighted spaces of holomorphic
functions. Such observations are essentially well known, see for instance [20].
Let Ωj ⊂ Cnj be open, j = 1, 2, and let Φj ∈ C(Ωj; R). We introduce the weighted
spaces
(A.1)
where L(dyj) is the Lebesgue measure on Cnj . When viewed as closed subspaces of
L2(Ωj, e−2Φj L(dyj)), the spaces HΦj (Ωj) are separable complex Hilbert spaces and
the natural embeddings HΦj (Ωj) → Hol(Ωj) are continuous. Here the space Hol(Ωj)
is equipped with its natural Fr´echet space topology of locally uniform convergence.
Let
HΦj (Ωj) = Hol(Ωj) ∩ L2(Ωj, e−2Φj L(dyj)),
j = 1, 2,
T : HΦ1(Ω1) → HΦ2(Ω2)
(A.2)
be a linear continuous map. Let us also write Ω1 = {z ∈ Cn1; z ∈ Ω1}.
Theorem A.1 There exists a unique function K(x, z) ∈ Hol(Ω2 × Ω1) such that
Ω1 ∋ y 7→ K(x, y) ∈ HΦ1(Ω1),
K(x, y)f (y)e−2Φ1(y) L(dy),
f ∈ HΦ1(Ω1).
Ω2 ∋ x 7→ K(x, z) ∈ HΦ2(Ω2),
(A.3)
(A.4)
(A.5)
for each x ∈ Ω2, and
T f (x) =ZΩ1
We also have
for each z ∈ Ω1.
31
When proving Theorem A.1, we observe that it follows from the remarks above that
for each x ∈ Ω2, the linear form
HΦ1(Ω1) ∋ f 7→ (T f ) (x) ∈ C
(A.6)
is continuous, and there exists therefore a unique element kx ∈ HΦ1(Ω1) such that
for all f ∈ HΦ1(Ω1), we have
T f (x) = (f, kx)Φ1,
x ∈ Ω2.
(A.7)
Here and in what follows (·,·)Φj stands for the scalar product in the space HΦj (Ωj),
j = 1, 2.
Letting (ej) be an orthonormal basis for HΦ1(Ω1), we may write with convergence
in HΦ1(Ω1), for each x ∈ Ω2 fixed,
∞Xj=1
(kx, ej)Φ1ej =
∞Xj=1
kx =
T ej(x)ej.
(A.8)
By Parseval's formula we get
Here we know that
kx 2
Φ1 =
∞Xj=1
T ej(x)2 ,
x ∈ Ω2.
kx Φ1 = sup
f Φ1 ≤1T f (x) ,
(A.9)
(A.10)
and it follows that the function Ω2 ∋ x 7→ kx Φ1 is locally bounded. Let us now
make the following elementary observation: let Ω ⊂ Cn be open and let fn ∈ Hol(Ω)
be such that the series
∞Xn=1
fn(z)2
(A.11)
converges for each z ∈ Ω, with the sum being locally integrable in Ω. Then the
series converges locally uniformly in Ω. Indeed, let us write
∞Xn=1
fn(z)2 =: F (z) ∈ L1
loc(Ω).
Let K ⊂ Ω be compact and let ω be an open neighborhood of K such that K ⊂
ω ⊂⊂ Ω. Then by Cauchy's integral formula and the Cauchy-Schwarz inequality we
have
K fn2 ≤ OK,ω(1) fn 2
sup
L2(ω).
32
We get therefore the uniform bound
NXn=1
K fn2 ≤ OK,ω(1) F L1(ω), N = 1, 2, . . . ,
sup
implying the locally uniform convergence of (A.11).
It follows that (A.9) holds with locally uniform convergence in x ∈ Ω2, and in
particular the function Ω2 ∋ x 7→ kx 2
Φ1 is continuous plurisubharmonic. We may
therefore conclude that the series in (A.8) converges locally uniformly in Ω1 × Ω2.
Letting
K(x, z) := kx(z) =
∞Xj=1
T ej(x)ej(z),
(A.12)
we conclude that K ∈ Hol(Ω2 × Ω1) is such that (A.3) and (A.4) hold, and these
properties characterize the kernel K uniquely.
When verifying (A.5), we let ekx ∈ HΦ2(Ω2) be the reproducing kernel for HΦ2(Ω2).
We may then write, when f ∈ HΦ1(Ω1), x ∈ Ω2,
and therefore,
Here
T f (x) = (T f,ekx)Φ2 = (f, T ∗ekx)Φ1,
kx = T ∗ekx.
T ∗ : HΦ2(Ω2) → HΦ1(Ω1)
(A.13)
(A.14)
is the adjoint of T . Letting (fj) be an orthonormal basis for HΦ2(Ω2) and recalling
that
fj(x)fj,
∞Xj=1
ekx =
∞Xj=1
(A.15)
(A.16)
we get
Therefore,
kx(y) =
fj(x)T ∗fj(y),
K(x, y) =
∞Xj=1
fj(x)T ∗fj(y).
33
and we see that (A.5) follows. We also get
K(·, y) 2
Φ2 =
∞Xj=1
T ∗fj(y)2 .
(A.17)
Remark. It follows from (A.9) that T ∈ L(HΦ1(Ω1), HΦ2(Ω2)) is of Hilbert-Schmidt
class precisely when
ZZΩ1×Ω2 K(x, y)2 e−2(Φ1(y)+Φ2(x)) L(dy) L(dx) < ∞.
Remark. An alternative proof of Theorem A.1 can be obtained by applying the
Schwartz kernel theorem directly to the linear continuous map
ΠΦ2T ΠΦ1 : L2(Ω1, e−2Φ1L(dy1)) → L2(Ω2, e−2Φ2L(dy2)).
Here
ΠΦj : L2(Ωj, e−2Φj L(dyj)) → HΦj (Ωj)
is the orthogonal projection. Writing the Schwartz kernel of ΠΦ2T ΠΦ1 in the form
K(x, y)e−2Φ1(y), we see that K should satisfy ∂xK(x, y) = 0. Now the distribu-
tion kernel of the adjoint ΠΦ1T ∗ΠΦ2 is given by K(y, x)e−2Φ2(y), and it follows
that ∂x(cid:16)K(y, x)(cid:17) = 0. We get ∂x (K(y, x)) = 0, so that (∂yK) (y, x) = 0 ⇐⇒
∂yK(x, y) = 0. We conclude that K(x, y) is holomorphic in (x, y).
B Positivity and Weyl quantization
The purpose of this appendix is to characterize the boundedness of the Weyl quan-
tization of a symbol of the form exp (iF (x, ξ)), where F a complex quadratic form,
in the HΦ-setting. See also [17] for a related discussion in the context of L2–
boundedness.
Let F = F (x, ξ) be a complex valued holomorphic quadratic form on C2n and let
us consider formally the Weyl quantization of eiF (x,ξ),
Au(x) = Opw(eiF )u(x) =
1
(2π)nZZ ei((x−y)·θ+F ((x+y)/2,θ))u(y)dydθ.
(B.1)
34
The holomorphic quadratic form (x − y) · θ + F ((x + y)/2, θ) is a non-degenerate
phase function in the sense of Hormander and generates a canonical relation
given by
κ : (y, η) 7→ (x, ξ),
x =
y =
x + y
2 −
+
x + y
2
ξ = θ +
η = θ −
1
2
1
2
F ′
x(
F ′
x(
F ′
ξ(
1
2
1
F ′
ξ(
2
x + y
2
x + y
2
x + y
2
x + y
2
, θ),
, θ).
, θ),
, θ),
(B.2)
(B.3)
The graph is parametrized by ρ = ((x + y)/2, θ) ∈ C2n and (B.2), (B.3) take the
form
(B.4)
κ : ρ +
HF (ρ),
1
2
HF (ρ) 7→ ρ −
1
2
ξ(ρ),−F ′
where HF (ρ) = (F ′
We shall now give a criterion for when κ in (B.4) is a canonical transformation.
Recall that HF (ρ) = F ρ, where
x(ρ)) is the Hamilton field of F at ρ.
F =(cid:18) F ′′
−F ′′
xξ(cid:19)
F ′′
ξξ
xx −F ′′
ξx
is the fundamental matrix of F (usually appearing as the linearization of a Hamilton
vector field, which in our case is already linear). We have
ξξ(cid:19) ,
and we notice that J 2 = −1, J t = −J. Then (B.4) is the relation
−1 0(cid:19) , F ′′ =(cid:18)F ′′
F = JF ′′, J =(cid:18) 0
xx F ′′
xξ
F ′′
ξx F ′′
1
(1 + F /2) ρ 7→ (1 − F /2) ρ
(B.5)
Now F is antisymmetric with respect to the bilinear form σ(ν, µ) = Jν · µ, hence
1−F /2 is bijective if and only if its transpose 1 +F /2 with respect to σ is bijective.
We conclude that the following three statements are equivalent:
(i) κ is a canonical transformation,
(ii) 1 − F /2 is bijective,
(iii) 1 + F /2 is bijective.
(B.6)
35
In what follows, we shall assume that (B.6) holds.
Let Φ0 be a strictly plurisubharmonic quadratic form on Cn and let ιΦ0 : C2n → C2n
be the corresponding antilinear involution, i.e. the unique antilinear map which is
equal to the identity on ΛΦ0. We shall now proceed to characterize the positivity of
the canonical transformation κ in (B.4) relative to ΛΦ0. Let
[µ, ν] =
1
2
b(µ, ν),
where b(µ, ν) has been defined in (2.5). It is a Hermitian form and that κ is positive
relative to ΛΦ0 precisely when
[ν, ν] ≥ [µ, µ], for all ν, µ with ν = κ(µ).
(B.7)
By (B.4) this is equivalent to
[ρ − (1/2)HF (ρ), ρ − (1/2)HF (ρ)] ≥ [ρ + (1/2)HF (ρ), ρ + (1/2)HF (ρ)], ρ ∈ C2n,
or equivalently,
Re [HF (ρ), ρ] ≤ 0, ρ ∈ C2n.
(B.8)
To simplify the following discussion, we shall make use of the invariance (exact
Egorov theorem) under conjugation of A in (B.1) with a unitary metaplectic Fourier
integral operator U : L2(Rn) → HΦ0(Cn) with the associated canonical transforma-
tion κU , mapping R2n onto ΛΦ0. The operator B = U −1AU is the Weyl quantization
of eiG, where G = F ◦κU . Also ιΦ0 = κUCκ−1
U , where C is the involution associated to
R2n, which is just the map of ordinary complex conjugation. By abuse of notation
we write F also for the pull back F ◦ κU and we continue the discussion in the case
when ΛΦ0 has been replaced with R2n and ιΦ0 with C, C(ρ) = ρ. In this setting,
(B.8) becomes
Im σ(F ′
ξ(ρ),−F ′
x(ρ); x, ξ) ≤ 0, ∀ρ = (x, ξ) ∈ C2n,
i.e.
Im (F ′
x(x, ξ) · x + F ′
ξ(x, ξ) · ξ) ≥ 0, (x, ξ) ∈ C2n,
or even more simply,
Writing ρ = µ + iν, µ, ν ∈ R2n we see that the last inequality is equivalent to
Im (F ′′
ρρρ · ρ) ≥ 0.
Im F ′′µ · µ + Im F ′′ν · ν ≥ 0,
36
i.e.
i.e.
Im F ′′ ≥ 0,
Im F ≥ 0 on R2n.
By the metaplectic invariance it follows that the positivity condition (B.7) is equiv-
alent to
Im F ≥ 0 on ΛΦ0,
(B.9)
now with the original F .
Remark. The condition (B.9) is quite natural since we know that for ordinary
symbols instead of eiF , the natural contour of integration in (B.1) should be θ =
(2/i)∂xΦ((x + y)/2), see [25], [12].
We summarize the discussion in this section in the following result.
Proposition B.1 Let F be a holomorphic quadratic form on C2n such that the
fundamental matrix of F does not have the eigenvalues ±2. Let Φ0 be a strictly
plurisubharmonic quadratic form on Cn. The canonical transformation associated
to the Fourier integral operator Opw(eiF ) is positive relative to ΛΦ0 precisely when
Im FΛΦ0 ≥ 0.
In particular, if (B.10) holds, then the operator
(B.10)
Opw(eiF ) : HΦ0(Cn) → HΦ0(Cn)
is bounded.
References
[1] A. Aleman and J. Viola, On weak and strong solution operators for evolution
equations coming from quadratic operators, J. Spectral Theory 8 (2018), 33-121.
[2] V. M. Babich and V. S. Buldyrev, Short-wavelength diffraction theory. Asymp-
totic methods, Springer Series on Wave Phenomena, 4. Springer-Verlag, Berlin,
1991. xi+445 pp.
[3] C. A. Berger and L. A. Coburn, Heat flow and Berezin-Toeplitz estimates, Amer.
J. Math. 116 (1994), 563-590.
37
[4] E. Caliceti, S. Graffi, M. Hitrik, and J. Sjostrand, Quadratic PT-symmetric
operators with real spectrum and similarity to selfadjoint operators, Journal of
Physics A, 2012.
[5] L. Coburn, Fock space, the Heisenberg group, heat flow and Toeplitz operators,
book chapter in preparation, 2018.
[6] N. Dencker, J. Sjostrand, and M. Zworski, Pseudospectra of semiclassical
(pseudo-)differential operators, Comm. Pure Appl. Math. 57 (2004), 384-415.
[7] V. Guillemin, Toeplitz operators in n dimensions, Integral Equations Operator
Theory 7 (1984), 145-205.
[8] R. F. Harvey and R. O. Wells, Jr. Zero sets of non-negative strictly plurisub-
harmonic functions, Math. Ann. 201 (1973), 165-170.
[9] F. H´erau, J. Sjostrand, and C. Stolk, Semiclassical analysis for the Kramers-
Fokker-Planck equation, Comm. Partial Differential Equations 30 (2005), 689-
760.
[10] M. Hitrik and K. Pravda-Starov, Spectra and semigroup smoothing for non-
elliptic quadratic operators, Math. Ann. 344 (2009), 801-846.
[11] M. Hitrik, K. Pravda-Starov, and J. Viola, From semigroups to subelliptic esti-
mates for quadratic operators, Trans. Amer. Math. Soc., to appear.
[12] M. Hitrik and J. Sjostrand, Two minicourses on analytic microlocal analysis,
"Algebraic and Analytic Microlocal Analysis", Springer, to appear.
[13] M. Hitrik, J. Sjostrand, and J. Viola, Resolvent estimates for elliptic quadratic
differential operators, Anal. PDE 6 (2013), 181-196.
[14] L. Hormander, Differential equations without solutions, Math. Ann. 140 (1960),
169-173.
[15] L. Hormander, On the existence and the regularity of solutions of linear pseudo-
differential equations, Enseignement Math. 17 (1971), 99-163.
[16] L. Hormander, L2 estimates for Fourier integral operators with complex phase,
Ark. Mat. 21 (1983), 283-307.
[17] L. Hormander, Symplectic classification of quadratic forms, and general Mehler
formulas, Math. Z. 219 (1995), 413-449.
38
[18] L. Hormander, On the Legendre and Laplace transformations, Ann. Scuola
Norm. Sup. Pisa Cl. Sci. 25 (1997), 517-568.
[19] A. Melin and J. Sjostrand, Bohr-Sommerfeld quantization condition for non-
selfadjoint operators in dimension 2, Ast´erisque, 284 (2003), 181-244.
[20] J. Peetre, The Berezin transform and Ha-plitz operators, J. Operator Theory
24 (1990), 165–186.
[21] K. Pravda-Starov, L. Rodino, and P. Wahlberg, Propagation of Gabor singu-
larities for Schrodinger equations with quadratic Hamiltonians, Math. Nachr.
291 (2018), 128–159.
[22] J. Ralston, On the construction of quasimodes associated with stable periodic
orbits, Comm. Math. Phys. 51 (1976), 219-242.
[23] J. Sjostrand, Parametrices for pseudodifferential operators with multiple char-
acteristics, Ark. Mat. 12 (1974), 85-130.
[24] J. Sjostrand, Singularit´es analytiques microlocales, Ast´ersique, 1982.
[25] J. Sjostrand, Function spaces associated to global I-Lagrangian manifolds, Struc-
ture of solutions of differential equations (Katata/Kyoto, 1995), 369423, World
Sci. Publ., River Edge, NJ, 1996.
39
|
1503.00144 | 1 | 1503 | 2015-02-28T15:35:38 | Entropy numbers of embedding operators of weighted Sobolev spaces with weights that are functions of distance from some h-set | [
"math.FA",
"math.CA"
] | In this paper order estimates for entropy numbers of embeddings of weighted Sobolev spaces on a John domain are obtained. In addition, we obtain order estimates for entropy numbers of summation operators on trees. | math.FA | math |
Entropy numbers of embedding operators of
weighted Sobolev spaces with weights that are
functions of distance from some h-set
A.A. Vasil'eva
1 Introduction
In this paper we obtain order estimates for entropy numbers of embedding operator
of weighted Sobolev spaces on a John domain into weighted Lebesgue space. Estimates
for n-widths of such embeddings were recently obtained in [42, 45].
Definition 1. Let X, Y be normed spaces, and let T : X → Y be a linear continuous
operator. Entropy numbers of T are defined by
ek(T ) = infnε > 0 : ∃y1, . . . , y2k−1 ∈ Y : T (BX ) ⊂ ∪2k−1
i=1 (yi + εBY )o ,
For properties of entropy numbers, we refer the reader to the books [4, 7, 32].
Kolmogorov, Tikhomirov, Birman and Solomyak [2, 15, 36] studied properties of ε-
entropy (this magnitude is related to entropy numbers of embedding operators).
k ∈ N.
Estimates for entropy numbers of the embedding operator of lm
p
obtained in the paper of Schutt [35] (see also [7]). Here lm
Rm with the norm
q were
p (1 6 p 6 ∞) is the space
into lm
k(x1, . . . , xm)kq ≡ k(x1, . . . , xm)klm
p =(cid:26) (x1p + · · · + xmp)1/p, if p < ∞,
max{x1, . . . , xm}, if p = ∞.
Later Edmunds and Netrusov [5], [6] generalized this result for vector-valued sequence
spaces (in particular, for sequence spaces with mixed norm). Haroske, Triebel, Kuhn,
Leopold, Sickel, Skrzypczak [8 -- 12, 14, 16 -- 23] studied the problem of estimating
entropy numbers of embeddings of weighted sequence spaces or weighted Besov and
Triebel -- Lizorkin spaces.
Lifshits and Linde [25] obtained estimates for entropy numbers of two-weighted
Hardy-type operators on a semiaxis (under some conditions on weights). The similar
problem for one-weighted Riemann-Liouville operators was considered in the paper
of Lomakina and Stepanov [29]. In addition, Lifshits and Linde [26 -- 28] studied the
problem of estimating entropy numbers of two-weighted summation operators on a
tree.
1
Triebel [37] and Mieth [31] studied the problem of estimating entropy numbers
of embedding operators of weighted Sobolev spaces on a ball with weights that have
singularity at the origin.
Estimates of entropy numbers of weighted function spaces are applied in spectral
theory of some degenerate elliptic operators (see, e.g., [8, 9, 13, 14, 19, 20, 22]) and in
estimating the probability of small deviation of Gaussian random functions (see,
e.g., [24, 25, 27]).
The paper is organized as follows. In this section we introduce notations and some
basic definitions, and we conclude this section with main result about estimates for
entropy numbers of embeddings of weighted Sobolev spaces. In §2 we formulate
some known results which will be required in the sequel. In §3 we obtain upper
estimates for entropy numbers of embedding operators of some function spaces on a
set with tree-like structure (the similar results for n-widths are obtained in [42]). In
§4 we prove Theorems 1, 2 and 3 about estimates for entropy numbers of embedding
operators of weighted Sobolev spaces. In §5 we obtain estimates of entropy numbers
of two-weighted summation operators on a tree.
Let us give the definition of a John domain.
We denote by AC[t0, t1] the space of absolutely continuous functions on an
interval [t0, t1].
Let Ba(x) be the closed euclidean ball of radius a in Rd centered at the point x.
Definition 2. Let Ω ⊂ Rd be a bounded domain, and let a > 0. We say that
Ω ∈ FC(a) if there exists a point x∗ ∈ Ω such that for any x ∈ Ω there exist
T (x) > 0 and a curve γx : [0, T (x)] → Ω with the following properties:
1. γx ∈ AC[0, T (x)], (cid:12)(cid:12)(cid:12)
2. γx(0) = x, γx(T (x)) = x∗,
(cid:12)(cid:12)(cid:12)
dγx(t)
dt
= 1 a.e.,
3. Bat(γx(t)) ⊂ Ω for any t ∈ [0, T (x)].
Definition 3. We say that Ω satisfies the John condition (and call Ω a John domain)
if Ω ∈ FC(a) for some a > 0.
For a bounded domain the John condition is equivalent to the flexible cone
condition (see the definition in [1]). As examples of such domains we can take
1. domains with Lipschitz boundary;
2. the interior of the Koch snowflake;
3. domains Ω = ∪06t6T Bct(γ(t)), where γ : [0, T ] → Rd is a curve with natural
parametrization and c > 0.
2
Domains with zero inner angles do not satisfy the John condition.
Reshetnyak [33,34] found the integral representation for smooth functions defined
on a John domain Ω in terms of their derivatives of order r. It follows from this
integral representation that, for p > 1, 1 6 q < ∞ and r
p > 0,
respectively) the class W r
p (Ω) is continuously (respectively, compactly) embedded in
the space Lq(Ω) (i.e., the conditions of the continuous and compact embedding are
the same as for Ω = [0, 1]d).
p > 0 ( r
d + 1
d + 1
q − 1
q − 1
Introduce the notion of h-set according to [3].
Denote by H the set of all nondecreasing positive functions defined on (0, 1].
Definition 4. Let Γ ⊂ Rd be a non-empty compact set, and let h ∈ H. We say that
Γ is an h-set if there are a constant c∗ > 1 and a finite countably additive measure
µ on Rd such that supp µ = Γ and for any x ∈ Γ, t ∈ (0, 1]
c−1
∗ h(t) 6 µ(Bt(x)) 6 c∗h(t).
(1)
Example 1. Let Γ ⊂ Rd be a Lipschitz manifold of dimension k, 0 6 k < d. Then
Γ is an h-set with h(t) = tk.
Example 2. Let Γ ⊂ R2 the the Koch snowflake. Then Γ is an h-set with h(t) =
tlog 4/ log 3 (see [30, p. 66 -- 68]).
Let us formulate the main result of this paper.
Everywhere below, we use the notation log x = log2 x.
Let · be a norm on Rd, and let E, E′ ⊂ Rd, x ∈ Rd. We set
diam· E = sup{y − z : y, z ∈ E}, dist· (x, E) = inf{x − y : y ∈ E}.
Let Ω ∈ FC(a) be a bounded domain, and let Γ ⊂ ∂Ω be an h-set. Further we
suppose that in some neighborhood of zero the function h ∈ H is defined by
h(t) = tθ log tγτ ( log t),
0 6 θ < d,
where τ : (0, +∞) → (0, +∞) is an absolutely continuous function such that
tτ ′(t)
τ (t)
→
t→+∞
0.
(2)
(3)
Let 1 < p 6 ∞, 1 6 q < ∞, r ∈ N, δ := r + d
q − d
p > 0, βg, βv ∈ R,
g(x) = ϕg(dist·(x, Γ)), v(x) = ϕv(dist·(x, Γ)),
ϕg(t) = t−βg log t−αg ρg( log t), ϕv(t) = t−βv log t−αv ρv( log t),
(4)
where ρg and ρv are absolutely continuous functions,
tρ′
g(t)
ρg(t)
→
t→+∞
0,
tρ′
v(t)
ρv(t)
→
t→+∞
0.
(5)
3
In addition, we suppose that
βv <
d − θ
q
or βv =
d − θ
q
, αv >
1 − γ
q
.
(6)
Without loss of generality we may assume that Ω ⊂(cid:0)− 1
We set β = βg+βv, α = αg+αv, ρ(y) = ρg(y)ρv(y), Z = (r, d, p, q, g, v, h, a, c∗),
Z∗ = (Z, R), where c∗ is the constant from Definition 4 and R = diam Ω.
2 , 1
2(cid:1)d.
We use the following notations for order inequalities. Let X, Y be sets, and let
f1(x, y)) if for
f1, f2 : X × Y → R+. We write f1(x, y) .
y
f2(x, y) (or f2(x, y) &
y
any y ∈ Y there exists c(y) > 0 such that f1(x, y) 6 c(y)f2(x, y) for any x ∈ X;
f1(x, y) ≍
y
f2(x, y) and f2(x, y) .
y
f2(x, y) if f1(x, y) .
y
f1(x, y).
Denote by Pr−1(Rd) the space of polynomials on Rd of degree not exceeding r−1.
For a measurable set E ⊂ Rd we set
Pr−1(E) = {f E : f ∈ Pr−1(Rd)}.
Notice that W r
p,g(Ω) ⊃ Pr−1(Ω).
In Theorems 1, 2, 3 the conditions on weights are such that W r
p,g(Ω) ⊂ Lq,v(Ω)
and there exist M > 0 and a linear continuous operator P : Lq,v(Ω) → Pr−1(Ω) such
that for any function f ∈ W r
p,g(Ω)
kf − P f kLq,v(Ω) 6 M(cid:13)(cid:13)(cid:13)(cid:13)
∇rf
g (cid:13)(cid:13)(cid:13)(cid:13)Lp(Ω)
(7)
(see [39, 40, 44, 45]).
Remark 1. Let x∗ be the point from Definition 2, and let R′ = dist·(x∗, ∂Ω). The
operator P defined in [40] (see also [41]) has the following property: there exists
s0 = s0(Z) ∈ (0, 1) such that, for any function f ∈ C ∞(Ω) ∩ Lq,v(Ω) satisfying the
condition f Bs0R′ (x∗) = 0, the equality P f = 0 holds.
We set W r
p,g(Ω) = {f − P f : f ∈ W r
p,g(Ω)}. Let W r
Theorem 1. Let (2), (3), (4), (5), (6) hold and 0 < θ < d.
equipped with norm kf k W r
embedding operator. From (7) it follows that I is continuous.
. Denote by I : W r
∇rf
g (cid:13)(cid:13)(cid:13)Lp(Ω)
p,g(Ω) :=(cid:13)(cid:13)(cid:13)
1. Suppose that β −δ < −θ(cid:16) 1
p(cid:17)+
q . We also suppose that δ
q − 1
. We set α0 = α for βv < d−θ
for βv = d−θ
and
d 6= δ−β
θ . Denote σ∗(n) = 1 for δ
q and α0 = α− 1
d < δ−β
θ
q
p,g(Ω) = span W r
p,g(Ω) be
p,g(Ω) → Lq,v(Ω) the
σ∗(n) = (log n)−α0+ (β−δ)γ
θ ρ(log n)τ
β−δ
θ (log n)
for δ
d > δ−β
θ . Then
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
4
n− min{ δ
d , δ−β
θ }+ 1
q − 1
p σ∗(n).
2. Suppose that β − δ = −θ(cid:16) 1
q − 1
.
p(cid:17)+
(a) Let p > q and α0 := α − (1 − γ)(cid:16) 1
p(cid:17) > 0 for βv = d−θ
α − 1 − (1 − γ)(cid:16) 1
q − 1
p(cid:17) > 0 for βv < d−θ
q − 1
q . Then
q , α0 :=
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
(log n)−α0 ρ(log n)τ − 1
q + 1
p (log n).
(b) Let p < q and α0 := α > 0 for βv < d−θ
q , α0 := α − 1
q > 0 for βv = d−θ
q .
Suppose that α0 6= 1
p − 1
q . Then
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
n
for α0 > 1
p − 1
q ,
1
q − 1
p (log n)−α0− 1
q + 1
p ρ(log n)
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
n−α0ρ(n)
for α0 < 1
p − 1
q .
Now we consider the case θ = 0.
Theorem 2. Let (2), (3), (4), (5) hold and θ = 0, β − δ < 0, βv < d
q . Then
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
n− r
d .
In the case θ = 0, β − δ = 0 we suppose that ρg(t) = log t−λg, ρv = log t−λv,
τ (t) = log tν (in the general case the estimates in assertion 1 of Theorem 3 can be
obtained similarly). Denote λ = λg + λv.
Theorem 3. Suppose that (2), (3), (4), (5) hold and θ = 0, β − δ = 0, βv < d
q .
1. Let α − (1 − γ)(cid:16) 1
δ
q − 1
p(cid:17)+
d < α
1−γ and σ∗(n) = (log n)−λ− αν
> 0. Suppose that
d > α
1−γ for δ
1−γ 6= δ
1−γ . Then
α
d. We set σ∗(n) = 1 for
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
n− min{ δ
d , α
1−γ }+ 1
q − 1
p σ∗(n).
2. Suppose that α − (1 − γ)(cid:16) 1
(a) Let p > q. Then
q − 1
p(cid:17)+
= 0, λ > (1 − ν)(cid:16) 1
q − 1
.
p(cid:17)+
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
5
(log n)−λ+(1−ν)( 1
q − 1
p ).
(b) Let p < q, λ 6= 1
p − 1
q . Then
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
n
1
q − 1
p (log n)−λ+ 1
p − 1
q
for λ > 1
p − 1
p ,
for λ < 1
p − 1
q .
en(I : W r
p,g(Ω) → Lq,v(Ω)) ≍
Z∗
n−λ
If Γ is a singleton, then estimates of entropy numbers are given by formulas from
Theorem 3 with γ = 0, τ ≡ 1. The proof is the same as for Theorem 3. These
estimates are the generalization of the result of Triebel [37] (in [37] the case p = q
was considered).
Without loss of generality we may assume that (x1, . . . , xd) = max16i6d xi.
Further we shall denote dist := dist·, diam := diam·.
2 Preliminaries
The following properties of entropy numbers are well-known (see, e.g., [7]):
1. if T : X → Y , S : Y → Z are linear continuous operators, then ek+l−1(ST ) 6
ek(S)el(T );
2. if T, S : X → Y are linear continuous operators, then
ek+l−1(S + T ) 6 ek(S) + el(T ).
From property 1 it follows that
ek(ST ) 6 kSkek(T ),
ek(ST ) 6 kT kek(S).
Further we denote by Iν the identity operator on Rν.
Theorem A. (see [7, 35]). Let 1 6 p 6 q 6 ∞. Then
(8)
(9)
ek(Iν : lν
p → lν
q ) ≍
p,q
Let 1 6 q < p 6 ∞. Then
1,
1 6 k 6 log ν,
k )
(cid:18) log(1+ ν
1
2− k
ν ν
k
q − 1
p ,
(cid:19)
1
p − 1
q
,
log ν 6 k 6 ν,
ν 6 k.
ek(Iν : lν
p → lν
q ) ≍
p,q
2− k
ν ν
1
q − 1
p ,
k ∈ N.
6
Remark 2. In estimates from [7] the value 2ν was taken instead of ν since the
spaces lν
q were considered as spaces over C.
p, lν
In the paper of Kuhn [23] the order estimates for entropy numbers od diagonal
operators Dσ : lp → lq were obtained for p > q.
Theorem B. (see [23]). Let 0 < q < p 6 ∞, σ = (σk)k∈N ∈ l pq
p−q
Suppose that there exists C > 0 such that ωn 6 Cω2n for any n. We define the
operator Dσ : lp → lq by Dσ(xk)k∈N = (σkxk)k∈N. Then en(Dσ : lp → lq) ≍
C,p,q
ωn.
, ωn =(cid:18) ∞
Pk=n
1
q − 1
p
.
σ
pq
p−q
k (cid:19)
The following result was proved by Lifshits [26].
Theorem C. (see [26]). Let X, Y be normed spaces, and let V ∈ L(X, Y ), {Vν}ν∈N ⊂
L(X, Y ). Then for any n ∈ N
en+[log2 N ]+1(V ) 6 sup
ν∈N
en(Vν) + sup
x∈BX
inf
ν∈N
kV x − VνxkY .
3 Estimates for entropy numbers of function classes
on a set with tree-like structure
First we give some notations.
Let (Ω, Σ, mes) be a measure space. We say that sets A, B ⊂ Ω are disjoint if
mes(A ∩ B) = 0. Let E, E1, . . . , Em ⊂ Ω be measurable sets, and let m ∈ N ∪ {∞}.
We say that {Ei}m
i=1 is a partition of E if the sets Ei are pairwise disjoint and
mes ((∪m
i=1Ei) △ E) = 0.
Denote by χE(·) the indicator function of a set E.
Let G be a graph containing at most countable number of vertices. We shall
denote by V(G) and by E(G) the set of vertices and the set of edges of G, respectively.
Two vertices are called adjacent if there is an edge between them. Let ξi ∈ V(G),
1 6 i 6 n. The sequence (ξ1, . . . , ξn) is called a path if the vertices ξi and ξi+1 are
adjacent for any i = 1, . . . , n − 1. If all the vertices ξi are distinct, then such a path
is called simple.
Let (T , ξ0) be a tree with a distinguished vertex (or a root) ξ0. We introduce
a partial order on V(T ) as follows: we say that ξ′ > ξ if there exists a simple
path (ξ0, ξ1, . . . , ξn, ξ′) such that ξ = ξk for some k ∈ 0, n. In this case, we set
ρT (ξ, ξ′) = ρT (ξ′, ξ) = n + 1 − k. In addition, we denote ρT (ξ, ξ) = 0. If ξ′ > ξ or
ξ′ = ξ, then we write ξ′ > ξ. This partial order on T induces a partial order on its
subtree.
Given j ∈ Z+, ξ ∈ V(T ), we denote
Vj(ξ) := VT
j (ξ) := {ξ′ > ξ : ρT (ξ, ξ′) = j}.
7
For ξ ∈ V(T ) we denote by Tξ = (Tξ, ξ) the subtree in T with vertex set
{ξ′ ∈ V(T ) : ξ′ > ξ}.
(10)
Let G be a subgraph in T . Denote by Vmax(G) and Vmin(G) the sets of maximal
and minimal vertices in G, respectively.
Let W ⊂ V(T ). We say that G ⊂ T is a maximal subgraph on the set of vertices
W if V(G) = W and any two vertices ξ′, ξ′′ ∈ W adjacent in T are also adjacent
in G.
Let {Tj}j∈N be a family of subtrees in T such that V(Tj) ∩ V(Tj′) = ∅ for j 6= j′
and ∪j∈NV(Tj) = V(T ). Then {Tj}j∈N is called a partition of the tree T . Let ξj
be the minimal vertex of Tj. We say that the tree Ts succeeds the tree Tj (or Tj
precedes the tree Ts) if ξj < ξs and
{ξ ∈ T : ξj 6 ξ < ξs} ⊂ V(Tj).
We consider the function spaces on sets with tree-like structure from [42].
Let (Ω, Σ, mes) be a measure space, let Θ be a countable partition of Ω into
measurable subsets, let A be a tree with a root such that
∃c1 > 1 :
card V1(ξ) 6 c1,
ξ ∈ V(A),
(11)
and let F : V(A) → Θ be a bijective mapping.
Throughout we consider at most countable partitions into measurable subsets.
Let 1 < p 6 ∞, 1 6 q < ∞ be arbitrary numbers. We suppose that, for any
measurable subset E ⊂ Ω, the following spaces are defined:
• the space Xp(E) with seminorm k · kXp(E),
• the space Yq(E) with seminorm k · kYq(E),
which all satisfy the following conditions:
1. Xp(Ω) ⊂ Yq(Ω);
2. Xp(E) = {f E : f ∈ Xp(Ω)}, Yq(E) = {f E : f ∈ Yq(Ω)};
3. if mes E = 0, then dim Yq(E) = dim Xp(E) = 0;
4. if E ⊂ Ω, Ej ⊂ Ω (j ∈ N) are measurable subsets, E = ⊔j∈NEj, then
kf kXp(E) =(cid:13)(cid:13)(cid:13)(cid:8)kf Ej kXp(Ej )(cid:9)j∈N(cid:13)(cid:13)(cid:13)lp
kf kYq(E) =(cid:13)(cid:13)(cid:13)(cid:8)kf Ej kYq(Ej )(cid:9)j∈N(cid:13)(cid:13)(cid:13)lq
8
,
,
f ∈ Xp(E),
(12)
f ∈ Yq(E);
(13)
5. if E ∈ Σ, f ∈ Yq(Ω), then f · χE ∈ Yq(Ω).
Let P(Ω) ⊂ Xp(Ω) be a subspace of finite dimension r0 and let kf kXp(Ω) = 0 for
any f ∈ P(Ω). For each measurable subset E ⊂ Ω we write P(E) = {P E : P ∈
P(Ω)}. Let G ⊂ Ω be a measurable subset and let T be a partition of G. We set
ST (Ω) = {f : Ω → R : f E ∈ P(E), f Ω\G = 0}.
(14)
If T is finite, then ST (Ω) ⊂ Yq(Ω) (see property 5).
For any finite partition T = {Ej}n
j=1 of the set E and for each function f ∈ Yq(Ω)
we put
kf kp,q,T = n
Xj=1
kf Ejkσp,q
Yq(Ej )!
1
σp,q
,
where σp,q = min{p, q}. Denote by Yp,q,T (E) the space Yq(E) with the norm k· kp,q,T .
Notice that k · kYq(E) 6 k · kp,q,T .
For each subtree A′ ⊂ A we set ΩA′ = ∪ξ∈V(A′) F (ξ).
Assumption 1. There is a function w∗ : V(A) → (0, ∞) with the following
property: for any ξ ∈ V(A) there exists a linear continuous operator PΩA ξ
: Yq(Ω) →
P(Ω) such that for any function f ∈ Xp(Ω) and any subtree A′ ⊂ A rooted at ξ
kf − PΩA ξ
f kYq(ΩA′ ) 6 w∗( ξ)kf kXp(ΩA′ ).
(15)
Assumption 2. There exist a function w∗ : V(A) → (0, ∞) and numbers δ∗ > 0,
c2 > 1 such that for each vertex ξ ∈ V(A) and for any n ∈ N, m ∈ Z+ there is a
partition Tm,n(G) of the set G = F (ξ) with the following properties:
1. card Tm,n(G) 6 c2 · 2mn.
2. For any E ∈ Tm,n(G) there exists a linear continuous operator PE : Yq(Ω) →
P(E) such that for any function f ∈ Xp(Ω)
kf − PEf kYq(E) 6 (2mn)−δ∗ w∗(ξ)kf kXp(E).
(16)
3. For any E ∈ Tm,n(G)
card {E′ ∈ Tm±1,n(G) : mes(E ∩ E′) > 0} 6 c2.
(17)
Assumption 3. There exist k∗ ∈ N, λ∗ > 0,
µ∗ > λ∗,
(18)
9
γ∗ > 0, absolutely continuous functions u∗ : (0, ∞) → (0, ∞) and ψ∗ : (0, ∞) →
yu′
∗(y)
(0, ∞), c3 > 1, t0 ∈ N, a partition {At,i}t>t0, i∈ Jt of the tree A such that lim
u∗(y) =
y→∞
0, lim
y→∞
yψ′
∗(y)
ψ∗(y) = 0,
c−1
3 2−λ∗k∗tu∗(2k∗t) 6 w∗(ξ) 6 c3 · 2−λ∗k∗tu∗(2k∗t),
ξ ∈ V(At,i),
3 2−µ∗k∗tu∗(2k∗t) 6 w∗(ξ) 6 c3 · 2−µ∗k∗tu∗(2k∗t),
c−1
ξ ∈ V(At,i),
and for νt := Pi∈ Jt
or
card V(At,i) one of the following estimates holds:
νt 6 c3 · 2γ∗k∗tψ∗(2k∗t) =: c3νt,
t > t0,
k∗ = 1,
νt 6 c3 · 2γ∗2t
ψ∗(22t
) =: c3ν t,
t > t0.
In addition, we assume that the following assertions hold.
1. If p > q, then
2−λ∗k∗t(card Jt)
2. Let t, t′ ∈ Z+. Then
1
q − 1
p 6 c3 · 2−µ∗k∗tν
1
q − 1
t
p
.
2−λ∗k∗t′
u∗(2k∗t′
) 6 c3 · 2−λ∗k∗tu∗(2k∗t)
if
t′ > t,
2−µ∗k∗t′
p
1
q − 1
u∗(2k∗t′)ν
t′ 6
if
q − 1
t
1
6 c3 · 2−µ∗k∗tu∗(2k∗t)ν
p
t′ > t,
p > q.
(19)
(20)
(21)
(22)
(23)
(24)
(25)
3. If the tree At′,i′ succeeds the tree At,i, then t′ = t + 1.
Remark 3. If ξ ∈ V(At,i), ξ′ ∈ V(At′,i′), ξ′ > ξ, then t′ > t.
Remark 4. If p > q, then from (25) it follows that (22) cannot hold.
We introduce some more notation.
• ξt,i is the minimal vertex of the tree At,i.
• Γt is the maximal subgraph in A on the set of vertices ∪i∈ Jt
for 1 6 t < t0 we put Γt = ∅, Jt = ∅.
V(At,i), t > t0;
• Gt = ∪ξ∈V(Γt) F (ξ) = ∪i∈ JtΩAt,i.
10
• Γt is the maximal subgraph on the set of vertices ∪j>tV(Γj), t ∈ N.
• { At,i}i∈J t is the set of connected components of the graph Γt.
• Ut,i = ∪ξ∈V( At,i)
F (ξ).
• Ut = ∪i∈J t
Ut,i = ∪ξ∈V(Γt)
F (ξ).
If t > t0, then
Vmin(Γt) = Vmin(Γt) = { ξt,i}i∈ Jt
(see [42, p. 30]), and we may assume that
J t = Jt,
t > t0.
(26)
(27)
The set Jt0 is a singleton. Denote {i0} = Jt0.
We set Z0 = (p, q, r0, w∗, w∗, δ∗, k∗, λ∗, µ∗, γ∗, ψ∗, u∗, c1, c2, c3).
From Assumption 1 it follows that there exists a linear continuous operator
P : Yq(Ω) → P(Ω) such that for any function f ∈ Xp(Ω)
kf − P f kYq(Ω) .
Z0
kf kXp(Ω)
(28)
(we take as P the operator PΩA ξt0,i0
). We set
Xp(Ω) = {f − P f : f ∈ Xp(Ω)}.
Then Xp(Ω) ⊂ Xp(Ω). Moreover, since Yq(Ω) is a normed space and kf kXp(Ω) =
kf − P f kXp(Ω) (by the property of P(Ω)), then ( Xp(Ω), k · kXp(Ω)) is a normed space.
Indeed, if kf − P f kXp(Ω) = 0, then kf − P f kYq(Ω)
(28)
= 0 and f − P f = 0.
Denote by B Xp(Ω) the unit ball of Xp(Ω). Let I : Xp(Ω) → Yq(Ω) be the
embedding operator.
The following assertions were proved in [42, p. 30, 32].
Lemma 1. There exists x0 ∈ (0, ∞) such that for any x > x0 the equation
yγ∗ψ∗(y) = x has a unique solution y(x). Moreover, y(x) = xβ∗ϕ∗(x), where β∗ = 1
γ∗
and ϕ∗ is an absolutely continuous function such that limx→+∞
∗(x)
xϕ′
ϕ∗(x) = 0.
Lemma 2. Let γ∗ > 0, ψ∗(y) = log yα∗ρ∗( log y), where ρ∗ : (0, ∞) → (0, ∞)
yρ′
∗(y)
ρ∗(y) = 0. Let ϕ∗ be such as in
is an absolutely continuous function such that lim
y→∞
Lemma 1. Then for sufficiently large x > 1
ϕ∗(x) ≍
γ∗,α∗,ρ∗
(log x)− α∗
γ∗ [ρ∗(log x)]− 1
γ∗ .
11
Theorem 4. Let 1 < p 6 ∞, 1 6 q < ∞, let assumptions 1, 2 and 3 with (21)
hold, and let
δ∗ >(cid:18) 1
q
−
1
p(cid:19)+
.
(29)
Suppose that δ∗ 6= λ∗β∗ for p 6 q and δ∗ 6= µ∗β∗ for p > q. Then there exists
n0 = n0(Z0) such that for any n > n0 the following estimates hold.
• Let p 6 q. We set
σ∗(n) =(cid:26) 1
for
δ∗ < λ∗β∗,
u∗(nβ∗ϕ∗(n))ϕ−λ∗
∗
(n)
for
δ∗ > λ∗β∗.
Then
en(I : Xp(Ω) → Yq(Ω)) .
Z0
• Let p > q. We set
n− min(δ∗, λ∗β∗)+ 1
q − 1
p σ∗(n).
σ∗(n) =(cid:26) 1
for
δ∗ < µ∗β∗,
u∗(nβ∗ϕ∗(n))ϕ−µ∗
∗
(n)
for
δ∗ > µ∗β∗.
Then
en(I : Xp(Ω) → Yq(Ω)) .
Z0
n− min(δ∗, µ∗β∗)+ 1
q − 1
p σ∗(n).
(30)
(31)
Theorem 5. Let 1 < p < q < ∞, let assumptions 1, 2 and 3 with (22) hold, and
let δ∗ >(cid:16) 1
q − 1
p(cid:17)+
. Then
en(I : Xp(Ω) → Yq(Ω)) .
Z0
1
q − 1
p (log n)−λ∗− 1
q + 1
p u∗(log n)
n
if λ∗ > 1
p − 1
q , and
en(I : Xp(Ω) → Yq(Ω)) .
Z0
n−λ∗u∗(n)
(32)
(33)
if λ∗ < 1
p − 1
q .
First we prove some auxiliary assertions.
Definition of operators Qt. By assumption 1 and (19), for any t > t0, i ∈
(27)
= J t there exists a linear continuous operator Pt,i : Yq(Ω) → P(Ω) such that
Jt
for any function f ∈ Xp(Ω) and for any subtree A′ ⊂ A rooted at ξt,i
kf − Pt,if kYq(ΩA′ ) .
Z0
2−λ∗k∗tu∗(2k∗t)kf kXp(ΩA′ );
(34)
12
in particular,
kf − Pt,if kYq( Ut,i) .
Z0
2−λ∗k∗tu∗(2k∗t)kf kXp( Ut,i).
(35)
Moreover, from the definitions of the space Xp(Ω) and of the operator P it follows
that we can set Pt0,i0 = 0. Let 1 6 t < t0. Then J t = {i0}, Ut,i0 = Ut0,i0. We set
Pt,i0 = 0. By (24) we get that (35) holds.
Let
Qtf (x) = Pt,if (x)
for x ∈ Ut,i,
i ∈ J t, Qtf (x) = 0 for x ∈ Ω\ Ut,
(36)
Tt = { Ut+1,i}i∈J t+1
.
(37)
Then (Qt+1f − Qtf )χ Ut+1 ∈ STt(Ω), and for p 6 q we have
kf − Qtf kYq( Ut) 6 kf − Qtf kYp,q,Tt−1 ( Ut)
(35)
.
Z0
2−λ∗k∗tu∗(2k∗t).
(38)
Notice that if t < t0, then Qtf = Qt+1f = 0 (since Pt,i0 = 0 for t 6 t0 by definition).
Definition of operators Pt,m. For t > t0 we set
mt = ⌈log νt⌉.
(39)
In [42] for each m ∈ Z+ the set Gm,t ⊂ Gt, the partition Tt,m of Gm,t and the linear
continuous operator Pt,m : Yq(Ω) → S Tt,m(Ω) were constructed with the following
properties:
1. Gm,t ⊂ Gm+1,t, Gmt,t = Gt;
2. for any m ∈ Z+
card Tt,m .
Z0
2m;
(40)
3. for any function f ∈ Xp(Ω) and for any set E ∈ Tt,m
kf − Pt,mf kYq(E) .
Z0
2−λ∗k∗tu∗(2k∗t)kf kXp(E), m 6 mt,
(41)
kf − Pt,mf kYq(E) .
Z0
2−µ∗k∗tu∗(2k∗t) · 2−δ∗(m−mt)kf kXp(E), m > mt;
(42)
4. for any set E ∈ Tt,m
card {E′ ∈ Tt,m±1 : mes(E ∩ E′) > 0} .
Z0
1.
(43)
13
Moreover, we may assume that Tmt,t = { F (ξ)}ξ∈V(Γt).
Remark 5. In [42] the relation (42) was proved with λ∗ instead of µ∗; however it
follows from the construction and from assumption 2 that the estimate with µ∗ is
also correct (see [42, proof of formula (80)]).
Let
t∗(n) = min{t ∈ N : ν t > n},
t∗∗(n) =(cid:26) t∗(n),
if p > q,
min{t ∈ N : ν t > 2n},
if p < q.
(44)
(45)
The following lemma is proved in [43, formula (60)].
Lemma 3. Let Λ∗ : (0, +∞) → (0, +∞) be an absolutely continuous function such
that
yΛ′
Λ∗(y) = 0. Then for any ε > 0
∗(y)
lim
y→+∞
t−ε .
ε,Λ∗
Λ∗(ty)
Λ∗(y)
tε,
.
ε,Λ∗
1 6 y < ∞, 1 6 t < ∞.
(46)
Proposision 1. If (21) holds, then
ν t∗(n) ≍
Z
n,
2k∗t∗(n) ≍
Z
nβ∗ϕ∗(n),
and in the case p < q we have
ν t∗∗(n) ≍
Z
2n,
2k∗t∗∗(n) ≍
Z
2β∗nϕ∗(2n).
If (22) holds, then for sufficiently large n ∈ N
2t∗(n) ≍
Z
log n;
if
p < q,
then 2t∗∗(n) ≍
Z
n.
(47)
(48)
(49)
Proof. Estimates (47) follow from [42, formula (51)]. The relations (48) can be
proved similarly.
Let us prove the first estimate in (49); the second relation is proved similarly.
We have
therefore, 2t∗(n) &
Z0
(22),(44)
6 2γ∗·2t∗(n)
n
ψ∗(22t∗(n)
)
(46)
.
Z0
22γ∗·2t∗(n)
;
log n for sufficiently large n ∈ N. Further,
(22),(44)
>
n
2γ∗·2t∗(n)−1
ψ∗(22t∗(n)−1
)
(46)
&
Z0
2γ∗·2t∗(n)−2
,
and 2t∗(n) .
Z0
log n for sufficiently large n ∈ N.
14
It is proved in [42, p. 37 -- 39] that for f ∈ Xp(Ω)
f =
t∗∗(n)−1
Pj=t0
(Qj+1f − Qjf )χ Uj+1 +
t∗∗(n)−1
Pj=t0
t∗∗(n)−1
t∗(n)−1
(f − Qjf )χGj + (f − Qt∗∗(n)f )χ Ut∗∗(n)
,
(f − Qtf )χGt =
(Pt,m+1f − Pt,mf )χGm,t+
Pt=t0
∞
Pm=0
(Pt,m+1f − Pt,mf )χGt +
t∗∗(n)−1
Pt=t0
Pt=t∗(n)
+
∞
Pm=mt
(Pt,mtf − Qtf )χGt,
t∗∗(n)−1
Pt=t∗(n)
t∗∗(n)−1
t∗(n)−1
∞
(f − Qtf )χGt =
(Pt,m+1f − Pt,mf )χGm,t+
(50)
(51)
(Pt,mt+1f − Qtf )χGt +
(Pt,m+1f − Pt,mf )χGt+
(52)
t∗(n)−1
Pt=t0
Pt=t0
+
Lemma 4.
Pt=t0
t∗∗(n)−1
Pm=mt+1
Pt=t∗(n)
∞
Pm=mt
(Pt,mtf − Qtf )χGt.
+
t∗∗(n)−1
Pt=t∗(n)
1. Let (21) hold. Then for any f ∈ B Xp(Ω)
kf − Qt∗∗(n)f kYq( Ut∗∗(n)) .
Z0
n−µ∗β∗+ 1
q − 1
p ϕ−µ∗
∗
(n)u∗(nβ∗ϕ∗(n))
if
p > q, (53)
kf − Qt∗∗(n)f kYq( Ut∗∗(n)) .
Z0
n−λ∗β∗ϕ−λ∗
∗
(n)u∗(nβ∗ϕ∗(n))
if
p = q,
(54)
kf − Qt∗∗(n)f kYq( Ut∗∗(n)) .
Z0
2−λ∗β∗nϕ−λ∗
∗
(2n)u∗(2β∗nϕ∗(2n))
if
p < q.
(55)
2. Let (22) hold (by Remark 4, this case is possible only for p 6 q). Then for
f ∈ B Xp(Ω)
kf − Qt∗∗(n)f kYq( Ut∗∗(n)) .
Z0
(log n)−λ∗u∗(log n)
if
p = q,
(56)
kf − Qt∗∗(n)f kYq( Ut∗∗(n)) .
Z0
n−λ∗u∗(n)
if
p < q.
(57)
Proof. Estimates from assertion 1 follow from (38), (46), (47), (48). Estimates from
assertion 2 follow from (38), (46) and (49).
Recall the notation σp,q = min{p, q}.
15
Lemma 5. (see [42]). Let T be a finite partition of a measurable subset G ⊂ Ω,
ν = dim ST (Ω) (see (14)). Then there exists a linear isomorphism A : ST (Ω) → Rν
such that kAkYp,q,T (G)→lν
1, kA−1klν
1.
.
q →Yq(G) .
q, r0
σp,q
σp,q, r0
Lemma 6. There exists a sequence {kt}t∗∗(n)−1
t=t0
⊂ N such that
t∗∗(n)−1
Xt=t0
(kt − 1) .
Z0
n
(58)
and the following assertions hold.
1. Suppose that (21) holds. Then for p > q
t∗∗(n)−1
Xt=t0
ekt(Qt+1 − Qt : Xp(Ω) → Yq( Ut+1)) .
Z0
and for p 6 q
t∗∗(n)−1
Xt=t0
ekt(Qt+1 − Qt : Xp(Ω) → Yq( Ut+1)) .
Z0
n−µ∗β∗+ 1
q − 1
p ϕ−µ∗
∗
n−λ∗β∗+ 1
q − 1
p ϕ−λ∗
∗
(n)u∗(nβ∗ϕ∗(n)),
(59)
(n)u∗(nβ∗ϕ∗(n)).
(60)
2. Suppose that (22) holds (by Remark 4, this case is possible only for p 6 q). If
p = q, then
t∗∗(n)−1
Xt=t0
ekt(Qt+1 − Qt : Xp(Ω) → Yq( Ut+1)) .
Z0
(log n)−λ∗u∗(log n);
(61)
if p < q and λ∗ > 1
p − 1
q , then
t∗∗(n)−1
Xt=t0
ekt(Qt+1 − Qt : Xp(Ω) → Yq( Ut+1)) .
Z0
n
if p < q and λ∗ < 1
p − 1
q , then
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n);
(62)
t∗∗(n)−1
Xt=t0
ekt(Qt+1 − Qt : Xp(Ω) → Yq( Ut+1)) .
Z0
n−λ∗u∗(n).
(63)
16
Proof. We set st = dim STt(Ω). By (21), (22) and (37), there exists C = C(Z0) > 1
such that
By Lemma 5, there exists an operator At : STt(Ω) → Rst such that
st 6 Cνt =: st.
kAtkYp,q,Tt →lst
σp,q
.
1,
σp,q,r0
By (9), (64) and (65),
t∗∗(n)−1
kA−1
t klst
q →Yq(Ω) .
q,r0
1.
(64)
(65)
t∗∗(n)−1
ekt(Qt+1 − Qt : Xp(Ω) → Yq( Ut+1)) .
Z0
Xt=t0
Xt=t0
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω)ekt(Ist : lst
.
σp,q → lst
q ).
Let p > q. Then k · kp,q,Tt = k · kYq( Ut+1). It was proved in [42, Step 3 in the proof
of Theorem 2] that for f ∈ B Xp(Ω)
kf − Qtf kYq( Ut) .
Z0
2−µ∗k∗tu∗(2k∗t)ν
Hence,
1
q − 1
t
p
.
(66)
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω) .
Z0
2−µ∗k∗tu∗(2k∗t)ν
1
q − 1
t
p
,
p > q.
(67)
Let p 6 q. It was proved in [42, Step 3 in the proof of Theorem 2] that for
f ∈ B Xp(Ω)
kf − Qtf kp,q,Tt .
Z0
2−λ∗k∗tu∗(2k∗t),
which implies
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω) .
Z0
2−λ∗k∗tu∗(2k∗t),
p 6 q.
(68)
Let ε > 0 (it will be chosen later by Z0). Denote
t(n) = t∗(n),
if (21) holds,
or if (22) holds and λ∗ >
1
p
−
1
q
,
and t(n) = t∗∗(n), if (22) holds and λ∗ < 1
p − 1
q .
We set
kt =(cid:26) ⌈n · 2−ε(t∗(n)−t)⌉
⌈n · 2−εt−t(n)⌉
for
for
t < t∗(n),
t∗(n) 6 t < t∗∗(n).
(69)
17
Then (58) holds.
By (21), (22), (64) and (69), there exists l∗ = l∗(Z0) ∈ N such that for sufficiently
decreases not
small ε > 0 and for any 0 6 r < l∗ the sequence n kl∗t+r
slower than some geometric progression. Moreover, for t < t∗(n) we have st
kt
sl∗t+rol∗t+r6t∗(n)−1
1.
.
Z0
Therefore, by Theorem A,
ekt(Ist : lst
σp,q → lst
q ) ≍
p,q
2− kt
st s
1
q − 1
σp,q
t
,
t < t∗(n).
Let p > q. By Remark 4, we have (21). Hence,
t∗(n)−1
Xt=t0
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω)ekt(Ist : lst
σp,q → lst
q )
(67)
.
Z0
. 2−µ∗k∗t∗(n)u∗(2k∗t∗(n))ν
1
q − 1
p
t∗(n)
(47)
≍
Z0
If p 6 q and (21) holds, then
n−µ∗β∗+ 1
q − 1
p ϕ−µ∗
∗
(n)u∗(nβ∗ϕ∗(n)).
t∗(n)−1
Xt=t0
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω)ekt(Ist : lst
σp,q → lst
q )
(68)
.
Z0
. 2−λ∗k∗t∗(n)u∗(2k∗t∗(n))ν
1
q − 1
p
t∗(n)
(47)
≍
Z0
If p 6 q and (22) holds, then
n−λ∗β∗+ 1
q − 1
p ϕ−λ∗
∗
(n)u∗(nβ∗ϕ∗(n)).
t∗(n)−1
Xt=t0
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω)ekt(Ist : lst
σp,q → lst
q )
(68)
.
Z0
. 2−λ∗k∗t∗(n)u∗(2k∗t∗(n)) · max
06r6l∗
kt∗(n)−1−r
st∗(n)−1−r s
−
2
p
1
q − 1
t∗(n)−1−r .
Z0
. 2−λ∗k∗t∗(n)u∗(2k∗t∗(n))k
1
q − 1
t∗(n)−1
p
(49)
.
Z0
n
1
q − 1
p (log n)−λ∗u∗(log n).
If p > q, we get the desired estimates in Lemma since t∗(n) = t∗∗(n).
Let p < q (then σp,q = p). We apply Theorem A. For t∗(n) 6 t < t∗∗(n) we have
ekt(Ist : lst
σp,q → lst
q ) .
p,q
k
1
q − 1
t
p
(cid:18)log(cid:18)1 +
ν t
kt(cid:19)(cid:19)
1
p − 1
q
k
.
Z0
1
q − 1
t
p
If (21) holds, then for sufficiently small ε > 0
t∗∗(n)−1
Xt=t∗(n)
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω)ekt(Ist : lst
σp,q → lst
q )
18
(log νt)
1
p − 1
q .
(68),(69)
.
Z0
.
t∗∗(n)−1
Xt=t∗(n)
2−λ∗k∗tu∗(2k∗t)n
1
q − 1
p 2ε( 1
p − 1
q )(t−t∗(n))(cid:18)log(cid:18)1 +
1
p − 1
q (47),(69)
.
Z0
ν t
kt(cid:19)(cid:19)
. n−λ∗β∗+ 1
q − 1
p ϕ−λ∗
∗
(n)u∗(nβ∗ϕ∗(n)).
If (22) holds, then
t∗∗(n)−1
Xt=t∗(n)
kQt+1 − Qtk Xp(Ω)→Yp,q,Tt (Ω)ekt(Ist : lst
σp,q → lst
q )
(22),(46),(68),(69)
.
Z0
.
t∗∗(n)−1
Xt=t∗(n)
2−λ∗tu∗(2t)n
1
q − 1
p 2ε( 1
p − 1
q )t−t(n)2t( 1
p − 1
q ) =: S.
If λ∗ > 1
p − 1
q , then S
(49)
.
Z0
n
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n). If λ∗ < 1
p − 1
q , then S
(49)
.
Z0
n−λ∗u∗(n). This completes the proof.
For t > t0, m ∈ Z+ we set
Tt,m = {E ∩ E′ : E ∈ Tt,m, E′ ∈ Tt,m+1, mes(E ∩ E′) > 0}.
(70)
Let
st,m = dim S Tt,m(Ω),
st,m = dim S Tt,m(Ω).
From (40) and (43) it follows that there exists C1(Z0) > 1 such that
st,m 6 C1 · 2m,
st,m 6 C1 · 2m.
(71)
(72)
By Lemma 5, there exists a linear isomorphism At,m : S Tt,m(Ω) → Rst,m such that
kAt,mkYp,q, Tt,m
(Gm,t)→l
st,m
σp,q
.
1,
σp,q,r0
kA−1
t,mkl
st,m
q →Yq(Gm,t) .
q,r0
Lemma 7. There exists a sequence {kt,m}t06t<t∗(n),m∈Z+ ⊂ N such that
n and the following assertions hold.
1) .
Z0
1.
(73)
Pt06t<t∗(n),m∈Z+
(kt,m−
1. Let (21) hold, let the sequence {kt}t>t0 be defined by (69), and let σ∗(n) be such
as in Theorem 4. Then for p > q
t∗(n)−1
∞
Xt=t0
Xm=mt+1
ekt,m(Pt,m+1 − Pt,m : Xp(Ω) → Yq(Gt))+
19
+
t∗(n)−1
Xt=t0
and for p 6 q
ekt(Pt,mt+1 − Qt : Xp(Ω) → Yq(Gt)) .
Z0
n− min{δ∗, µ∗β∗}+ 1
q − 1
p σ∗(n),
t∗(n)−1
∞
Xt=t0
Xm=0
ekt,m(Pt,m+1 − Pt,m : Xp(Ω) → Yq(Gm,t)) .
Z0
2. Suppose that (22) holds. Then
n− min{δ∗, λ∗β∗}+ 1
q − 1
p σ∗(n).
t∗(n)−1
∞
Xt=t0
Xm=0
ekt,m(Pt,m+1−Pt,m : Xp(Ω) → Yq(Gm,t)) .
Z0
n
Proof. By (9) and (73), it suffices to estimate
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n).
kPt,m+1 − Pt,mk Xp(Ω)→Yq(Gt)ekt,m(Ist,m : lst,m
q → lst,m
q
)+
(74)
kPt,mt+1 − Qtk Xp(Ω)→Yq(Gt)ekt(Ist,mt+1 : l
st,mt+1
q
→ l
st,mt+1
q
) =: S1
t∗(n)−1
∞
Pm=mt+1
t∗(n)−1
+
Pt=t0
Pt=t0
for p > q and
kPt,m+1 − Pt,mk Xp(Ω)→Yp,q, Tt,m
t∗(n)−1
∞
Xt=t0
Xm=0
for p 6 q.
Let
(Ω)ekt,m(Ist,m : lst,m
p → lst,m
q
) =: S2
(75)
st,m = ⌈C1 · 2m⌉.
We define the number t1(n) as follows. In assertion 1 of Lemma we set
for p > q and
t1(n) =(cid:26) t0
if
t∗(n),
δ∗ < µ∗β∗,
if
δ∗ > µ∗β∗
t1(n) =(cid:26) t0,
if
t∗(n),
δ∗ < λ∗β∗,
if
δ∗ > λ∗β∗
for p 6 q. In assertion 2 we set
t1(n) = t∗(n).
20
(76)
(77)
(78)
(79)
Denote
m∗
t = ⌈log(n · 2−εt−t1(n))⌉,
kt,m = ⌈n · 2−ε(t−t1(n)+m−m∗
t )⌉.
(80)
Then kt,m ∈ N and
t∗(n)−1
Pt=t0
∞
Pm=0
(kt,m − 1) .
Z0,ε
n.
Case p > q. From Holder's inequality it follows that for f ∈ B Xp(Ω), m > mt
kPt,m+1f − Pt,mf kYq(Gm,t)
(42)
.
Z0
. 2−µ∗k∗t−δ∗(m−mt)u∗(2k∗t)(card Tt,m)
1
q − 1
p
(40),(43)
.
Z0
2−µ∗k∗t−δ∗(m−mt)u∗(2k∗t)·2m( 1
p )
q − 1
(39)
.
Z0
· 2m(−δ∗+ 1
Further, for f ∈ B Xp(Ω) by Holder's inequality we het
. 2−µ∗k∗tu∗(2k∗t) · ν δ∗
t
q − 1
p ).
kPt,mt+1f − Qtf kYq(Gt) 6 kf − Pt,mt+1f kYq(Gt) + kf − Qtf kYq(Gt)
(40),(42),(66)
.
Z0
. 2−µ∗k∗tu∗(2k∗t) · 2mt( 1
q − 1
p ) + 2−µ∗k∗tu∗(2k∗t)ν
1
q − 1
t
p
(39)
.
Z0
2−µ∗k∗tu∗(2k∗t)ν
1
q − 1
t
p
Thus,
kPt,m+1f − Pt,mf kYq(Gm,t) .
Z0
2−µ∗k∗tu∗(2k∗t) · ν δ∗
t
· 2m(−δ∗+ 1
q − 1
p ),
kPt,mt+1f − Qtf kYq(Gt) .
Z0
2−µ∗k∗tu∗(2k∗t)ν
1
q − 1
t
p
.
.
(81)
(82)
From (76) we get that st,mt+1 = ⌈C1 · 2mt+1⌉
(39)
6 ⌈4C1ν t⌉. Similarly as in Lemma
6 we prove that
t∗(n)−1
Pt=t0
6
2−µ∗k∗tu∗(2k∗t)ν
1
q − 1
t
p
ekt(Ist,mt+1 : l
st,mt+1
q
→ l
st,mt+1
q
)
(72),(76)
6
t∗(n)−1
Pt=t0
2−µ∗k∗tu∗(2k∗t)ν
1
q − 1
t
p
ekt(Ist,mt+1 : l
st,mt+1
q
→ l
st,mt+1
q
) .
Z0
(83)
. n−µ∗β∗+ 1
q − 1
p ϕ−µ∗
∗
(n)u∗(nβ∗ϕ∗(n)).
Let ε > 0 be sufficiently small. From Theorem A it follows that for m 6 m∗
t
ekt,m(Ist,m : lst,m
q → lst,m
q
)
(72),(76)
6 ekt,m(Ist,m : lst,m
q → lst,m
q
(76),(80)
)
.
p,q
−
2
kt,m
st,m .
21
In addition, there exists l∗ = l∗(Z0) such that for any t0 6 t < t∗(n), 0 6 r < l∗ the
decreases not slower than some geometric progression.
sequence n kt,l∗m+r
st,l∗m+rol∗m+r6m∗
If m > m∗
t , then
t
ekt,m(Ist,m : lst,m
q → lst,m
q
)
By Remark 4, (21) holds. Hence,
(72),(76)
6 ekt,m(Ist,m : lst,m
q → lst,m
q
1.
) .
p,q
(74),(81),(82)
S1
.
Z0
t∗(n)−1
∞
Xt=t0
Xm=mt+1
2−µ∗k∗tu∗(2k∗t)·νδ∗
t ·2m(−δ∗+ 1
q − 1
p )ekt,m(Ist,m : lst,m
q → lst,m
q
)+
+
t∗(n)−1
Xt=t0
2−µ∗k∗tu∗(2k∗t)ν
1
q − 1
t
p
ekt(Ist,mt+1 : l
st,mt+1
σp,q → l
st,mt+1
q
)
(29),(83)
.
Z0,ε
.
t∗(n)−1
Xt=t0
2−µ∗k∗tu∗(2k∗t)·νδ∗
t ·2m∗
t (−δ∗+ 1
q − 1
p )+n−µ∗β∗+ 1
q − 1
p ϕ−µ∗
∗
(n)u∗(nβ∗ϕ∗(n))
(21),(80)
.
Z0,ε
.
t∗(n)−1
Xt=t0
2−µ∗k∗tu∗(2k∗t) · 2δ∗γ∗k∗tψδ∗
∗ (2k∗t) ·(cid:0)n · 2−εt−t1(n)(cid:1)−δ∗+ 1
q − 1
p +
+n−µ∗β∗+ 1
q − 1
p ϕ−µ∗
∗
(n)u∗(nβ∗ϕ∗(n)) =: S′
1.
Recall that β∗ = 1
γ∗
. If δ∗ < µ∗β∗, then for sufficiently small ε > 0 we have S′
1
n−δ∗+ 1
q − 1
p . If δ∗ > µ∗β∗, then for sufficiently small ε > 0 we get
S′
1
(77)
.
Z0
2−µ∗k∗t∗(n)u∗(2k∗t∗(n))ν δ∗
t∗(n) · n−δ∗+ 1
q − 1
p + n−µ∗β∗+ 1
q − 1
p ϕ−µ∗
∗
(n)u∗(nβ∗ϕ∗(n))
(77)
.
Z0
(47)
.
Z0
. n−µ∗β∗+ 1
q − 1
p σ∗(n).
This completes the proof for p > q.
Case p 6 q. Let f ∈ B Xp(Ω). Then for m 6 mt
kPt,m+1f − Pt,mf kp,q, Tt,m
(41)
.
Z0
and for m > mt
2−λ∗k∗tu∗(2k∗t) 6 2−λ∗k∗tu∗(2k∗t) · 2−δ∗(m−mt),
(84)
kPt,m+1f − Pt,mf kp,q, Tt,m
(18),(42)
.
Z0
22
2−λ∗k∗tu∗(2k∗t) · 2−δ∗(m−mt).
(85)
By Theorem A, we have for m 6 m∗
t
ekt,m(Ist,m : lst,m
p → lst,m
q
)
(72),(76)
6 ekt,m(Ist,m : lst,m
p → lst,m
q
(76),(80)
)
.
p,q
−
2
kt,m
st,m s
1
q − 1
t,m
p
(moreover, there exists l∗ = l∗(Z0) ∈ N such that for any 0 6 r < l∗, t0 6 t <
decreases not slower than some geometric
t∗(n) the sequence n kt,l∗m+r
progression), and for m > m∗
st,l∗m+ro06l∗m+r6m∗
t we have
t
ekt,m(Ist,m : lst,m
p → lst,m
q
This implies that
) .
p,q
min
log(cid:16)1 + st,m
kt,m(cid:17)
kt,m
, 1
1
p − 1
q
.
(75),(84),(85)
S2
.
Z0
t∗(n)−1
Xt=t0
m∗
t
Xm=0
2−λ∗k∗tu∗(2k∗t) · 2−δ∗(m−mt)2
−
kt,m
st,m s
1
q − 1
t,m +
p
+
t∗(n)−1
∞
Xt=t0
Xm=m∗
t +1
2−λ∗k∗tu∗(2k∗t)·2−δ∗(m−mt) min
t∗(n)−1
kt,m(cid:17)
log(cid:16)1 + st,m
kt,m
1
p − 1
q
(39),(76),(80)
.
Z0
, 1
2−λ∗k∗tu∗(2k∗t) · 2−δ∗m∗
t νδ∗
t k
1
q − 1
p
t,m∗
t
(80)
.
Z0
2−λ∗k∗tu∗(2k∗t)(n · 2−εt−t1(n))−δ∗+ 1
q − 1
p ν δ∗
t =: S′
2.
.
Xt=t0
.
t∗(n)−1
Xt=t0
If (21) holds, then νt = 2γ∗k∗tψ∗(2k∗t). Hence, for δ∗ < λ∗β∗ we have S′
2
and for δ∗ > λ∗β∗ we get
(78)
.
Z0
n−δ∗+ 1
q − 1
p ,
S′
2
(78)
.
Z0
2−λ∗k∗t∗(n)u∗(2k∗t∗(n))n−δ∗+ 1
q − 1
p νδ∗
t∗(n)
(47)
.
Z0
n−λ∗β∗+ 1
q − 1
p σ∗(n).
If (22) holds, then νt = 2γ∗2t
ψ∗(22t), k∗ = 1 and
S′
2
(44),(49),(79)
.
Z0
n
1
q − 1
p (log n)−λ∗u∗(log n) 6 n
This completes the proof.
23
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n).
Theorems 4 and 5 for p > q follow from (8), (45), (50), (51), (52) and Lemmas
4, 6, 7.
Consider the case p < q.
We set
t2(n) =(cid:26) t∗(n),
t∗∗(n),
if (21) holds, or if (22) holds and λ∗ > 1
if (22) holds and λ∗ < 1
p − 1
q .
p − 1
q ,
(86)
Lemma 8. Let p < q. We set
kt,m = ⌈n · 2−εt−t2(n)−ε(m−mt)⌉,
t∗(n) 6 t < t∗∗(n), m > mt.
(87)
Then
Xt∗(n)6t<t∗∗(n), m>mt
(kt,m − 1) .
Z0,ε
n
(88)
and for sufficiently small ε > 0 the following assertions hold.
1. Let (21) hold. Then
t∗∗(n)−1
∞
Xt=t∗(n)
Xm=mt
ekt,m(Pt,m+1−Pt,m : Xp(Ω) → Yq(Gt)) .
Z0
2. Let (22) hold. Then
n−λ∗β∗+ 1
q − 1
p ϕ−λ∗
∗
(n)u∗(nβ∗ϕ∗(n)).
t∗∗(n)−1
∞
Xt=t∗(n)
Xm=mt
ekt,m(Pt,m+1−Pt,m : Xp(Ω) → Yq(Gt)) .
Z0
n
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n)
for λ∗ > 1
p − 1
q ,
t∗∗(n)−1
∞
Xt=t∗(n)
Xm=mt
for λ∗ < 1
p − 1
q .
ekt,m(Pt,m+1 − Pt,m : Xp(Ω) → Yq(Gt)) .
Z0
n−λ∗u∗(n)
Proof. The relation (88) follows from (87).
Let Tt,m, st,m be defined by (70), (71). From (72) it follows that st,m 6 C1 · 2m.
We set st,m = ⌈C1 · 2m⌉. Then
st,mt
(39)
≍
Z0
νt.
24
(89)
By (9) and (73), it suffices to estimate
t∗∗(n)−1
∞
Xt=t∗(n)
Xm=mt
kPt,m+1 − Pt,mk Xp(Ω)→Yp,q, Tt,m
From (18), (41), (42) it follows that
(Gt)ekt,m(Ist,m : lst,m
p → lst,m
q
) =: S.
kPt,m+1 − Pt,mk Xp(Ω)→Yp,q, Tt,m
(Gt) .
Z0
2−λ∗k∗tu∗(2k∗t) · 2−δ∗(m−mt).
(76)
> C1·2mt
(39)
&
Z0
νt
(21),(22)
&
Z0
Since st,m
A that
ekt,m(Ist,m : lst,m
p → lst,m
q
) .
Z0
Hence, for sufficiently small ε > 0 we have
S .
Z0
t∗∗(n)−1
Xt=t∗(n)
2−λ∗k∗tu∗(2k∗t)
2−λ∗k∗tu∗(2k∗t) · n
1
q − 1
p · 2εt−t2(n)( 1
p − 1
.
t∗∗(n)−1
Xt=t∗(n)
(44)
> n and kt,m
(87)
6 n, we get by Theorem
νt∗(n)
1
p − 1
q
.
, 1
log(cid:16)1 + st,m
kt,m(cid:17)
kt,m
min
log(cid:16)1 + st,mt
kt,mt(cid:17)
kt,mt
q )(cid:18)log(cid:18)1 +
1
p − 1
q
(87),(89)
.
Z0
νt
⌈n · 2−εt−t2(n)⌉(cid:19)(cid:19)
1
p − 1
q
.
This together with (44), (47), (49), (86) yields the desired estimates for sufficiently
small ε > 0.
Lemma 9. Let p < q. Then there exists a sequence {kt}t∗(n)6t<t∗∗(n) ⊂ N such that
t∗∗(n)−1
Xt=t∗(n)
(kt − 1) .
Z0
n
and the following assertions hold.
1. Let (21) hold. Then
t∗∗(n)−1
Xt=t∗(n)
ekt(Pt,mt − Qt : Xp(Ω) → Yq(Gt)) .
Z0
25
n−λ∗β∗+ 1
q − 1
p u∗(nβ∗ϕ∗(n))ϕ−λ∗
∗
(90)
(n).
2. Let (22) hold. Then for λ∗ > 1
p − 1
q we have
t∗∗(n)−1
Xt=t∗(n)
ekt(Pt,mt − Qt : Xp(Ω) → Yq(Gt)) .
Z0
n
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n),
and for λ∗ < 1
p − 1
q we have
t∗∗(n)−1
Xt=t∗(n)
ekt(Pt,mt − Qt : Xp(Ω) → Yq(Gt)) .
Z0
n−λ∗u∗(n).
We shall use Theorem C and the following assertion.
Lemma 10. (see [38]). Let (T , ξ∗) be a tree with finite vertex set, let
card V1(ξ) 6 k, ξ ∈ V(T ),
(91)
and let the mapping Φ : 2V(T ) → R+ satisfy the following condition:
Φ(V1 ∪ V2) > Φ(V1) + Φ(V2), V1, V2 ⊂ V(T ), V1 ∩ V2 = ∅,
(92)
Φ(V(T )) > 0. Then there is a number C(k) > 0 such that for any n ∈ N there
exists a partition Sn of the tree T into at most C(k)n subtrees Tj, which satisfies
the following conditions:
1. Φ(V(Tj)) 6 (k+2)Φ(V(T ))
n
for any j such that card V(Tj) > 2;
2. if m 6 2n, then each element of Sn intersects with at most C(k) elements of
Sm.
Proof of Lemma 9. Step 1. We define the numbers t2(n) by (86); ε = ε(Z0) will
be chosen later. The sequence kt(n) is defined so that
kt(n) − 1 .
Z0
n · 2−εt−t2(n)
if (21) holds,
kt(n) − 1 .
Z0
max(cid:8)n · 2−εt−t2(n), 2t(cid:9) if (22) holds.
This together with (49) implies (90).
(93)
(94)
Further we consider t∗(n) 6 t < t∗∗(n).
Step 2. Let T be a partition of Γt into subtrees. Then T = {At,i,s}i∈ Jt, s∈It,i
,
where {At,i,s}s∈It,i (cid:22) is a partition of At,i. Denote by ξt,i,s the root of At,i,s.
Define the operator PT : Xp(Ω) → Yq(Ω) as follows. We set PTf Ω\Gt = 0,
PTf ΩAt,i,s
= 0,
if
ξt,i,s = ξt,i
(95)
26
PTf ΩAt,i,s
= (Pt,mtf − Qtf )ΩAt,i,s
if V(At,i,s) = { ξt,i,s};
in other cases we set
PTf ΩAt,i,s
= (PΩA ξt,i,s
f − Qtf )ΩAt,i,s
(see Assumption 1). Let T = {ΩA′}A′∈T. Notice that
PTf ∈ ST (Ω).
Let f ∈ B Xp(Ω). If ξt,i,s = ξt,i, then
(96)
(97)
(98)
kPt,mt f − Qtf − PTf kYq(ΩAt,i,s )
(34),(36),(41),(95)
.
Z0
2−λ∗k∗tu∗(2k∗t)kf kXp(ΩAt,i,s ).
(99)
If V(At,i,s) = { ξt,i,s}, then
kPt,mtf − Qtf − PTf kYq(ΩAt,i,s )
(96)
= 0.
(100)
In other cases
kPt,mtf − Qtf − PTf kYq(Ω)
(15),(19),(41),(97)
.
Z0
2−λ∗k∗tu∗(2k∗t)kf kXp(ΩAt,i,s ).
(101)
We set T′ = {A′ ∈ T : card V(A′) > 2}. Then for any f ∈ B Xp(Ω)
kPt,mt f − Qtf − PTf kYq(Gt)
(99),(100),(101)
.
Z0
2−λ∗k∗tu∗(2k∗t) XA′∈T′
Step 3. Let r ∈ N,
r 6
νt
4
.
kf kq
Xp(ΩA′ )!1/q
.
(102)
(103)
Denote by Nr the family of partitions T of the graph Γt into subtrees such that
card {At,i,s ∈ T : ξt,i,s 6= ξt,i} 6 r. The number Nr can be estimated from above
by the number of choices of sets of vertices ξt,i,s 6= ξt,i in V(Γt). Therefore,
Nr 6
r
Xm=0
C m
V(Γt)
(21),(22)
6
C m
⌈c3νt⌉
r
Xm=0
If (21) holds, then we set
(103)
. C r
r (cid:19)r
⌈c3νt⌉ 6(cid:18) e⌈c3νt⌉
.
(104)
rt = ⌈n · 2−2εt−t2(n)−c⌉, n > N(Z0),
(105)
27
where c = c(Z0), N(Z0) are such that rt 6 νt
n · 2−2εt−t2(n)−c > 1 we have
4 for n > N(Z0). Then in the case
log Nrt−1 6 log Nrt
(104)
.
Z0
rt log
c3eν t
rt
(86),(105)
.
Z0
. n · 2−2ε(t−t∗(n)) log(cid:18)
n · 2−2ε(t−t∗(n))−c ·
.
Z0
. n · 2−2ε(t−t∗(n)) log(22ε(t−t∗(n)) · 22γ∗k∗(t−t∗(n)) · e) .
Z0
νt
νt∗(n)(cid:19) (21),(46),(47)
c3eν t∗(n)
. n · 2−2ε(t−t∗(n)) · (2γ∗k∗ + 2ε)(t − t∗(n) + 1) .
Z0,ε
n · 2−ε(t−t∗(n));
if n · 2−2εt−t2(n)−c < 1, then Nrt−1
(104)
= C 0
V(Γt) = 1 and log Nrt−1 = 0. Therefore,
log Nrt−1 .
Z0,ε
n · 2−ε(t−t∗(n)).
If (22) holds, then we set
rt =(cid:24) n · 2−εt−t2(n)
2t+c
(cid:25) ,
where c = c(Z0) is such that rt 6 ν t
4 . Then
log Nrt−1 6 log Nrt
(104)
.
Z0
rt log
ec3νt
rt
6 rt log(c3eν t)
(22),(46)
.
Z0
(106)
(107)
(108)
. rt · 2t .
Z0
i.e.,
max(cid:8)n · 2−εt−t2(n), 2t(cid:9) ;
max(cid:8)n · 2−εt−t2(n), 2t(cid:9) .
log Nrt−1 .
Z0,ε
Step 4. Consider the tree At with vertex set V(A)\V(Γt+1). For f ∈ B Xp(Ω)
we define the function Φf : 2V( At) → R+ by
Φf (W) = Xξ∈W∩V(Γt)
kf kp
Xp( F (ξ))
.
(109)
Then Φf (W1 ⊔ W2) = Φf (W1) + Φf (W2). From (11) and Lemma 10 it follows that
there exists a sequence of partitions {Tf,t,l}06l6log rt of the tree At such that
card Tf,t,l 6 rt · 2−l,
(110)
28
card {A′′ ∈ Tf,t,l±1 : V(A′′) ∩ V(A′) 6= ∅} .
Z0
1, A′ ∈ Tf,t,l,
(111)
and for any subtree A′ ∈ Tf,t,l such that card V(A′) > 2 the following estimate
holds:
Here we may assume that
Denote T∗
f,t,l := Tf,t,lΓt.
We have
Φf (V(A′)) .
Z0
2l
rt
.
card Tf,t,⌊log rt⌋ = 1.
T∗
f,t,0 = Tf,t,0Γt ∈ Nrt−1,
(112)
(113)
(114)
sup
f ∈B Xp(Ω)
kPt,mt f − Qtf − PT∗
f,t,0
f kYq(Gt)
(102),(109),(110),(112)
.
Z0
2−λ∗k∗tu∗(2k∗t)r
1
q − 1
t
p
=: At.
Step 5. Let (21) hold. Then for sufficiently small ε > 0 we get
t∗∗(n)−1
Xt=t∗(n)
At
(86),(105),(115)
.
Z0
t∗∗(n)−1
Xt=t∗(n)
2−λ∗k∗tu∗(2k∗t)(n · 2−2ε(t−t∗(n)))
1
q − 1
p
(115)
(47)
.
Z0
. n−λ∗β∗+ 1
q − 1
p u∗(nβ∗ϕ∗(n))ϕ−λ∗
∗
(n);
i.e.,
t∗∗(n)−1
Xt=t∗(n)
At .
Z0
Let (22) hold. Then
n−λ∗β∗+ 1
q − 1
p u∗(nβ∗ϕ∗(n))ϕ−λ∗
∗
(n).
(116)
t∗∗(n)−1
Xt=t∗(n)
At
(107),(115)
.
Z0
t∗∗(n)−1
Xt=t∗(n)
2−λ∗tu∗(2t)(n · 2−εt−t2(n)−t)
1
q − 1
p =: A.
In λ∗ > 1
p − 1
q , then for sufficiently small ε > 0 we have
A
(86)
.
Z0
2−λ∗t∗(n)u∗(2t∗(n)) · n
1
q − 1
p · 2( 1
p − 1
q )t∗(n)
(49)
.
Z0
n
29
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n);
i.e.,
t∗∗(n)−1
Xt=t∗(n)
n
At .
Z0
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n).
If λ∗ < 1
p − 1
q , then for sufficiently small ε > 0
A
(86)
.
Z0
2−λ∗t∗∗(n)u∗(2t∗∗(n)) · n
i.e.,
1
q − 1
p · 2( 1
p − 1
q )t∗∗(n)
(49)
.
Z0
n−λ∗u∗(n);
t∗∗(n)−1
Xt=t∗(n)
n−λ∗u∗(n).
At .
Z0
(117)
(118)
Step 6. Let 0 6 l 6 log rt, Tt,l = T∗
f ,t,l
for some function f ∈ B Xp(Ω). We set
kt,l = ⌈n · 2−ε(t−t2(n)+l)⌉.
(119)
Let us estimate the sum
t∗∗(n)−1
Xt=t∗(n) X06l6log rt−1
ekt,l
( PTt,l − PTt,l+1 : Xp(Ω) → Yq(Gt))
(notice that PT
t,⌊log rt⌋
(95)
= 0 since card T f ,t,⌊log rt⌋
(113)
= 1).
We set Tt,l = {ΩA′}A′∈Tt,l,
Tt,l = {E ∩ E′ : E ∈ Tt,l, E′ ∈ Tt,l+1, mes (E ∩ E′) > 0}.
By construction, for any function f ∈ Xp(Ω) we have PTt,lf − PTt,l+1f
(Ω). Let s′
(98)
∈
(Ω). From (110) and (111) it follows that there exists
S Tt,l
C(Z0) > 1 such that
t,l = dim S Tt,l
s′
t,l 6 ⌈C(Z0)rt · 2−l⌉ =: s′′
t,l.
By Lemma 5, there exists an isomorphism At,l : S Tt,l
(Ω) → Rs′
t,l such that
kAt,lk
s′
t,l
(Gt)→l
p
1,
.
Z0
Yp,q, Tt,l
kA
−1
t,l k
s′
t,l
q →Yq(Gt)
l
1.
.
Z0
(120)
(121)
Hence, by (9) it suffices to estimate the sum
t∗∗(n)−1
Xt=t∗(n) X06l6log rt−1
k PTt,l − PTt,l+1k Xp(Ω)→Yp,q, Tt,l
(Ω)ekt,l
(Is′′
t,l
s′′
t,l
p → l
s′′
t,l
q ) =: S.
: l
30
Let us estimate k PTt,l − PTt,l+1k Xp(Ω)→Yp,q, Tt,l
(Ω). Consider a function f ∈ B Xp(Ω).
Then
XE′∈Tt,l
i.e.,
k PTt,lf − PTt,l+1f kp,q, Tt,l
k PTt,lf − PTt,l+1f kp
kPt,mtf − Qtf − PTt,lf kp
kPt,mtf − Qtf − PTt,l+1f kp
=
XE∈ Tt,l
Yq(E′) + XE′′∈Tt,l+1
(99),(100),(101)
.
Z0
2−λ∗k∗tu∗(2k∗t);
1/p
(111)
.
Z0
Yq(E)
1/p
Yq(E′′)
k PTt,l − PTt,l+1k Xp(Ω)→Yp,q, Tt,l
(Ω) .
Z0
2−λ∗k∗tu∗(2k∗t).
(122)
Since for 0 6 l 6 log rt and small ε > 0
s′′
t,l
kt,l
(119),(120)
≍
Z0
rt · 2−l
⌈n · 2−ε(t−t2(n)+l)⌉
(105),(107)
.
Z0
1,
by Theorem A we get
ekt,l
(Is′′
t,l
s′′
t,l
p → l
: l
s′′
t,l
q ) .
Z0
(s′′
t,l)
1
q − 1
p · 2
−
kt,l
s′′
t,l .
Moreover, there exists m∗ = m∗(Z0) such that for 0 6 ν < m∗ the sequence
increases not slower than some geometric progression. This
s′′
n kt,m∗l+ν
t,m∗l+νo06m∗l+ν6log rt
together with (122) yields that
S .
Z0
t∗∗(n)−1
Xt=t∗(n)
2−λ∗k∗tu∗(2k∗t) max
06ν<m∗
(s′′
t,ν)
1
q − 1
p · 2
−
kt,ν
s′′
t,ν .
Z0
.
t∗∗(n)−1
Xt=t∗(n)
2−λ∗k∗tu∗(2k∗t)(kt,0)
1
q − 1
p =: S′.
If (21) holds, then
S′
(119)
.
Z0
2−λ∗k∗t∗(n)u∗(2k∗t∗(n))(kt∗(n),0)
1
q − 1
p
(47),(86)
.
Z0
31
n−λ∗β∗+ 1
q − 1
p ϕ−λ∗
∗
(n)u∗(nβ∗ϕ∗(n)).
Hence,
t∗∗(n)−1
Xt=t∗(n) X06l6log rt−1
ekt,l
( PTt,l − PTt,l+1 : Xp(Ω) → Yq(Gt)) .
Z0
n−λ∗β∗+ 1
q − 1
p ϕ−λ∗
∗
(n)u∗(nβ∗ϕ∗(n)).
(123)
Let (22) hold. Then for λ∗ > 1
p − 1
q
S′
(119)
.
Z0
2−λ∗k∗t∗(n)u∗(2k∗t∗(n))(kt∗(n),0)
1
q − 1
p
(49),(86)
n
.
Z0
1
q − 1
p (log n)−λ∗u∗(log n).
ekt,l
( PTt,l − PTt,l+1 : Xp(Ω) → Yq(Gt)) .
Z0
n
1
q − 1
p (log n)−λ∗+ 1
p − 1
q u∗(log n).
(119)
> n · 2−εt∗∗(n)
(49)
≍
Z0
n1−ε. Therefore, for sufficiently small
(124)
2−λ∗k∗t∗(n)u∗(2k∗t∗(n))(kt∗(n),0)
1
q − 1
p
(49)
.
Z0
Hence,
t∗∗(n)−1
Xt=t∗(n) X06l6log rt−1
p − 1
q . Then kt,0
Let λ∗ < 1
ε > 0
S′ .
Z0
. n( 1
q − 1
p )(1−ε)(log n)−λ∗u∗(log n) .
n−λ∗u∗(n);
Z0
i.e.,
t∗∗(n)−1
Xt=t∗(n) X06l6log rt−1
ekt,l
( PTt,l − PTt,l+1 : Xp(Ω) → Yq(Gt)) .
Z0
n−λ∗u∗(n).
(125)
Step 7. Let N ′
t = {T f ,t,0Γt : f ∈ B Xp(Ω)}. Then N ′
t
(114)
⊂ Nrt−1, and
log N ′
t
(86),(106)
.
Z0,ε
n · 2−εt−t2(n)
if
(21) holds,
(126)
log N ′
t
(108)
.
Z0,ε
max{n · 2−εt−t2(n), 2t} if
(22) holds.
(127)
Let kt,l be defined by (119). We set
kt = X06l6log rt−1
(kt,l − 1) + 1 .
Z0,ε
n · 2−εt−t2(n).
(128)
32
By Theorem C,
ekt+[log N ′
t ]+1(Pt,mt − Qt : Xp(Ω) → Yq(Gt)) 6
6 supT∈N ′
t
ekt( PT : Xp(Ω) → Yq(Gt))+
+ supf ∈B Xp(Ω) inf T∈N ′
: Xp(Ω) → Yq(Gt))+
6 sup f ∈B Xp(Ω) ekt
+ supf ∈B Xp(Ω) kPt,mtf − Qtf − PTf,t,0f kYq(Gt).
t kPt,mtf − Qtf − PTf kYq(Gt) 6
( PT f ,t,0
(129)
Let kt = kt(n) = kt + [log N ′
t ] + 1. Then
kt − 1
(126),(128)
.
Z0,ε
n · 2−εt−t2(n)
for
(21),
kt − 1
(127),(128)
.
Z0,ε
max{n · 2−εt−t2(n), 2t} for
(22);
i.e., (93), (94) hold. Further,
t∗∗(n)−1
t∗∗(n)−1
Xt=t∗(n)
Xt=t∗(n)
Xt=t∗(n)
t∗∗(n)−1
6
+
ekt(Pt,mt − Qt : Xp(Ω) → Yq(Gt))
(129)
6
sup
f ∈B Xp(Ω)
ekt
( PT f ,t,0
: Xp(Ω) → Yq(Gt))+
sup
f ∈B Xp(Ω)
kPt,mtf − Qtf − PTf,t,0f kYq(Gt).
We apply (8) and (123), (124), (125) to estimate the first summand, and we use
(115), (116), (117) and (118) to estimate the second summand.
This completes the proof.
The relations (8), (50), (51) and Lemmas 4, 6, 7, 8, 9 yield Theorems 4 and 5
for p < q.
Remark 6. Suppose that Assumptions 1 and 3 hold, and Assumption 2 is substituted
by the following condition: for any ξ ∈ V(A) the set F (ξ) is the atom of mes. Then
the assertions of Theorems 4 and 5 hold with δ∗ = +∞.
33
4 Estimates of entropy numbers of weighted Sobolev
spaces
Let Ω ∈ FC(a), and let Γ ⊂ ∂Ω be an h-set. Suppose that there exists c0 > c∗ such
that
h(t)
h(s)
6 c0,
j ∈ N,
t,
s ∈ [2−j−1, 2−j+1]
(130)
(here c∗ is the constant from Definition 4). In [39, 40] there were constructed a
tree A with vertex set {ηj,i}j>jmin,i∈ Ij , a number s = s(a, d) ∈ N and a partition
{Ω[ηj,i]}j>jmin,i∈ Ij of the domain Ω with following properties:
1. ηjmin,1 is the minimal vertex in A, and for any j > jmin the family of sets
{VA
1 (ηj,i)}i∈ Ij form the partition of {ηj+1,t}t∈ Ij+1.
2. For any j > jmin, i ∈ Ij we have Ω[ηj,i] ∈ FC(b∗) with b∗ = b∗(a, d) > 0.
3. diam Ω[ηj,i] ≍
a,d,c0
2−sj.
4. For any x ∈ Ω[ηj,i] we have dist (x, Γ) ≍
a,d,c0
2−sj.
5. For any j > jmin, i ∈ Ij, j′ > j we have card VA
j′−j(ηj,i) .
a,d,c0
particular,
card Ij .
a,d,c0
h(2−sjmin)
h(2−sj)
.
h(2−sj )
h(2−sj′ )
; in
(131)
Suppose that conditions of Theorem 1, 2 or 3 hold (then we have (130) with
c0 = c0(Z)). We define weight functions u, w : V(A) → (0, ∞) as follows:
u(ηj,i) = uj = g(2−sj) · 2−(r− d
p )sj, w(ηj,i) = wj = v(2−sj) · 2− d
q sj.
(132)
For each subtree D ⊂ A we denote Ω[D] = ∪ξ∈V(D)Ω[ξ]. It was proved in [40, 44, 45]
that for any j0 > jmin, i0 ∈ Ij0 and for any vertex ηj0,i0 there exists a linear continuous
: Lq,v(Ω) → Pr−1(Ω) such that for any subtree D with minimal vertex
operator Pηj0 ,i0
ηj0,i0 and for any function f ∈ W r
p,g(Ω)
kf − Pηj0 ,i0
f kLq,v(Ω[D]) .
Z
here C(j0) is defined as follows.
C(j0)(cid:13)(cid:13)(cid:13)(cid:13)
∇rf
g (cid:13)(cid:13)(cid:13)(cid:13)Lp(Ω[D])
;
(133)
34
1. Let βv < d−θ
q . Then (see [40])
C(j0) = sup
j>j0
ujwj
for p 6 q,
C(j0) = Xj>j0
(ujwj)
pq
p−q
h(2−sj0)
h(2−sj)!
1
q − 1
p
2. Let θ > 0, βv = d−θ
q . Then (see [45])
C(j0) = 2−(δ−β)sj0(sj0)−α+ 1
q ρ(sj0),
for p > q.
if p < q
or p > q,
β − δ < −θ(cid:16) 1
q − 1
p(cid:17) ,
C(j0) = 2−θ( 1
q − 1
p )sj0(sj0)−α+1+ 1
q − 1
p ρ(sj0)
for p > q, β − δ = −θ(cid:18)1
q
−
(134)
1
p(cid:19) .
(135)
The magnitude C(j0) in Case 1 is estimated as follows. If β − δ < −θ(cid:16) 1
q − 1
,
p(cid:17)+
then
C(j0) .
Z
2(β−δ)sj0(sj0)−αρ(sj0),
(136)
if β − δ = −θ(cid:16) 1
q − 1
p(cid:17)+
and α > (1 − γ)(cid:16) 1
q − 1
p(cid:17)+
, then
C(j0) .
Z
2
−θ( 1
q − 1
p )+
−α+( 1
sj0j
0
q − 1
p )+
ρ(j0)
(137)
(these estimates are proved in [40]). Suppose that conditions of assertion 2 of
Theorem 3 hold: i.e., θ = 0,
β − δ = 0, α = (1 − γ)(cid:18)1
q
−
1
p(cid:19)+
,
λ > (1 − ν)(cid:18)1
q
−
1
p(cid:19)+
.
Then
(4),(132)
C(j0)
.
Z
(log(sj0))−λ,
p 6 q,
(138)
(139)
(2),(4),(132),(138)
C(j0)
.
Z
(sj0)γ( 1
q − 1
p )[log(sj0)]ν( 1
q − 1
(sj0)γ( 1
q − 1
p )[log(sj0)]−λ+ 1
.
Z
p ) Pj>j0
q − 1
p ,
(sj)−1(log(sj))−λ pq
p−q −ν!
p > q.
1
q − 1
p
.
Z
In proofs of lower estimates of entropy numbers we use the following assertions.
(140)
35
Lemma 11. Let Ω ⊂ Rd be a domain, let W r
be pairwise non-overlapping sets, and let ψ1, . . . , ψm ∈ W r
supp ψj ⊂ Gj,
p,g(Ω) ⊂ Lq,v(Ω), let G1, . . . , Gm ⊂ Ω
= 1,
∇rψj
p,g(Ω), (cid:13)(cid:13)(cid:13)
g (cid:13)(cid:13)(cid:13)Lp(Ω)
kψjkLq,v(Gj ) > M,
1 6 j 6 m.
(141)
Let X = span {ψj}m
Lq,v(Ω) be the embedding operator. Then for any n ∈ {0, . . . , m}
j=1 be equipped with norm kf kX =(cid:13)(cid:13)(cid:13)
∇rf
g (cid:13)(cid:13)(cid:13)Lp(Ω)
en(id : X → Lq,v(Ω)) > M · en(Im : lm
p → lm
q ).
, and let id : X →
In particular, if X ⊂ W r
p,g(Ω), then
en(I : W r
p,g(Ω) → Lq,v(Ω)) > M · en(Im : lm
p → lm
q ).
This lemma is proved similarly as the lower estimate of n-widths in [47]. Similarly
from Theorem B we obtain
Corollary 1. Let Ω ⊂ Rd be a domain, let W r
pairwise non-overlapping sets, and let ψj ∈ W r
p,g(Ω) ⊂ Lq,v(Ω), let Gj ⊂ Ω, j ∈ N, be
= 1, supp ψj ⊂ Gj,
∇rψj
p,g(Ω),(cid:13)(cid:13)(cid:13)
j ∈ N.
g (cid:13)(cid:13)(cid:13)Lp(Ω)
kψjkLq,v(Gj ) > Mj,
(142)
j=1 ∩ span W r
Let X = span {ψj}∞
p,g(Ω) be equipped with norm kf kX = (cid:13)(cid:13)(cid:13)
and let id : X → Lq,v(Ω) be the embedding operator. Denote ωn = ∞
Pj=n
Suppose that there exists C > 1 such that ωn 6 Cω2n for any n ∈ N. Then for any
n ∈ N we have
g (cid:13)(cid:13)(cid:13)Lp(Ω)
j !
q − 1
,
.
pq
p−q
∇rf
M
1
p
en(id : X → Lq,v(Ω)) &
p,q,C
ωn.
In particular, if X ⊂ W r
p,g(Ω), then
en(I : W r
p,g(Ω) → Lq,v(Ω)) &
p,q,C
ωn.
Proof of Theorems 1, 2, 3. The upper estimate. We set Θ = {Ω[ηj,i]}j>jmin,i∈ Ij ,
F (ηj,i) = Ω[ηj,i], Xp(Ω) = span W r
p,g(Ω), Yq(Ω) = Lq,v(Ω), P(Ω) =
Pr−1(Ω).
p,g(Ω), Xp(Ω) = W r
Similarly as in [42, p. 49] we can prove that Assumption 2 holds with δ∗ = δ
d and
w∗(ηj,i) .
Z
ujwj.
Consider the following cases.
36
1. Suppose that one of the following conditions holds:
,
q − 1
• β − δ < −θ(cid:16) 1
• β − δ = −θ(cid:16) 1
• θ = 0, β − δ = 0, α > (1 − γ)(cid:16) 1
p(cid:17)+
p(cid:17)+
, p > q,
q − 1
q − 1
.
p(cid:17)+
Then the partition {At,i}t>t0,i∈ Jt is constructed similarly as in [42, p. 49 -- 50],
and by properties 1 -- 5 of the tree A and (133), (134), (135), (136), (137) we
get that Assumptions 1 and 3 hold. Here λ∗, γ∗ and ψ∗ are the same as in
[42] (see Cases 1, 3, 4, and Case 2 for p > q; in Cases 1 and 4 we take
Λ(x) = log xγτ ( log x)). If βv < d−θ
q , then the function u∗ is the same as
in [42] (in Cases 1 and 4 we take Ψ(x) = log x−αρ( log x)). If βv = d−θ
q ,
then u∗(y) = (log y)−α+ 1
, and
q − 1
u∗(y) = (log y)−α+1+ 1
that for βv = d−θ
Applying Theorem 4 and Lemma 2, we get the upper estimate in assertions 1
and 2a of Theorem 1, in Theorem 2 and in assertion 1 of Theorem 3.
q ρ(log y) for p < q or β − δ < −θ(cid:16) 1
p ρ(log y) for β − δ = −θ(cid:16) 1
q we consider only the case θ > 0).
and p > q (recall
p(cid:17)+
p(cid:17)+
q − 1
q − 1
2. Let θ > 0, β − δ = −θ(cid:16) 1
in A on the vertex set
q − 1
p(cid:17)+
, p < q. Denote by Γt the maximal subgraph
{ηj,l : 2t−1 < sj 6 2t,
l ∈ Ij},
and by {At,i}i∈ Jt we denote the set of connected components of Γt. We set t0 =
2θ·2t2−γtτ −1(2t).
min{t ∈ Z+ : V(Γt) 6= ∅}. By (2) and (131), card V(Γt) .
Z0
This together with (133), (134) and (137) implies that Assumptions 1 and 3
hold with λ∗ = −α0 (see assertion 2b of Theorem 1), u∗(y) = ρ(y), γ∗ = θ,
ψ∗(y) = (log y)−γτ −1(log y) in (22). Applying Theorem 5, we get the upper
estimate in assertion 2b of Theorem 1.
3. Let θ = 0, β − δ = 0, α = (1 − γ)(cid:16) 1
q − 1
.
p(cid:17)+
(a) Suppose that p > q. We define the partition {At,i}t>t0,i∈ Jt similarly
as in [42, p. 50] (see Case 3). Then Assumptions 1 and 3 hold with
q − 1
γ∗ = 1 − γ, λ∗ = α + 1
ψ∗(x) = (log x)−ν, u∗(x) = (log x)−λ+ 1
p(cid:17), µ∗ = α = (1 − γ)(cid:16) 1
q = −γ(cid:16) 1
p(cid:17),
p − 1
q − 1
q − 1
p (see (139), (140)). Hence,
(131)
β∗ = 1
1−γ , ϕ∗(x) = (log x)
ν
1−γ (see Lemma 2). In addition, card Jt
.
Z
37
2−γtt−ν. The relations (23) and (25) follow from the conditions α =
q − 1
q − 1
q − 1
Indeed, 2−λ∗k∗tu∗(2k∗t) = 2γ( 1
p . Since the function h is non-
decreasing, we have γ 6 0; moreover, ν 6 0 for γ = 0. If γ < 0 and p > q,
p(cid:17) and λ > (1 − ν)(cid:16) 1
(1 − γ)(cid:16) 1
then (24) follows from the inequality γ(cid:16) 1
q − 1
p )tt−λ+ 1
q − 1
then (24) follows from the inequality −λ + 1
p = q, then the assertion follows from the inequality λ > 0.
Notice that for p = q we have λ∗ = µ∗. Since µ∗β∗ = 1
q − 1
get by Theorem 4 that
p < ν(cid:16) 1
p(cid:17). Let us check that (24) holds.
p(cid:17) < 0. If γ = 0 and p > q,
p(cid:17) 6 0. If
q − 1
q − 1
d = δ∗, we
p < δ
en(I : W r
p,g(Ω) → Lq,v(Ω)) .
Z
u∗(nβ∗ϕ∗(n))ϕ−µ∗
∗
(n) ≍
Z
(log n)−λ+ 1
q − 1
p −ν( 1
q − 1
p ).
Thus, we obtain the upper estimate in assertion 2a of Theorem 3.
(b) Let p < q. We set V(Γt) = {ηj,i :
22t−1
< j 6 22t} and denote by
{At,i}i∈ Jt the set of connected components of the graph Γt. Then
card V(Γt) .
Z X22t−1 +16sj622t
(sj)−γ(log(sj))−ν .
2(1−γ)2t
2−νt =: ν t
Z
(i.e., (22) holds). This together with (139) implies that Assumptions 1
and 3 hold with λ∗ = µ∗ = λ, u∗ ≡ 1. Applying Theorem 5, we get the
upper estimate in assertion 2b of Theorem 3.
The lower estimate. Similarly as in [42, p. 50] we can prove that
en(I : W r
p,g(Ω) → Lq,v(Ω)) &
Z∗
en(I : W r
p ([0, 1]d) → Lq([0, 1]d)) ≍
p,q,r,d
n− δ
d + 1
q − 1
p .
This gives the desired lower estimates in Theorem 2, in assertion 1 of Theorem
1 for δ
d < δ−β
θ and in assertion 1 of Theorem 3 for δ
d < α
1−γ .
In other cases we apply Lemma 11 or Corollary 1.
In [42, p. 50], [40] and [45] the number k∗∗ = k∗∗(Z∗) ∈ N is defined and the
functions {ψt,j}j∈Jt ∈ C ∞(Rd) are constructed with the following properties:
card Jt &
Z∗
2θk∗∗t(k∗∗t)−γτ −1(k∗∗t),
(143)
= 1. Moreover, kψt,jkLq,v(Ω) is estimated from below as follows.
(cid:13)(cid:13)(cid:13)
1. If βv < d−θ
q , then
∇rψt,j
g (cid:13)(cid:13)(cid:13)Lp(Ω)
kψt,jkLq,v(Ω) &
Z∗
2k∗∗t(β−δ)(k∗∗t)−αρ(k∗∗t).
(144)
38
2. Let θ > 0, βv = d−θ
q ; in addition, we suppose that p < q or p > q, β − δ <
−θ(cid:16) 1
q − 1
p(cid:17)+
. Then
kψt,jkLq,v(Ω) &
Z∗
2k∗∗t(β−δ)(k∗∗t)−α+ 1
q ρ(k∗∗t).
(145)
3. If θ > 0, p > q, βv = d−θ
q and β − δ = −θ(cid:16) 1
q − 1
p(cid:17), then
kψt,jkLq,v(Ω) &
Z∗
2−θ( 1
q − 1
p )k∗∗t(k∗∗t)−α+ 1
q +1− 1
p ρ(k∗∗t).
(146)
In addition, in Case 1 the supports of ψt,j do not overlap pairwise for different (t, j);
in Cases 2, 3 for any t the supports of ψt,j do not overlap pairwise for different j.
Moreover, it follows from the construction of functions ψt,i that for any x ∈
supp ψt,j
dist (x, Γ) .
Z∗
2−k∗∗t.
Hence, by Remark 1, there exists t = t(Z∗) ∈ N such that for t > t we have P ψt,i = 0
and ψt,i ∈ W r
p,g(Ω).
1. Suppose that θ > 0.
(a) Let β − δ < −θ(cid:16) 1
q − 1
1. We take tn such that
p(cid:17)+
. Let α0 be as defined in assertion 1 of Theorem
n 6 2θk∗∗tn(k∗∗tn)−γτ −1(k∗∗tn) .
Z∗
n.
Then
(see Lemma 2) and
2k∗∗tn ≍
Z∗
1
θ (log n)
γ
θ τ
1
θ (log n)
n
k∗∗tn ≍
Z∗
log n.
This together with Theorem A and Lemma 11 yields that
(147)
(148)
(149)
en(I : W r
p,g(Ω) → Lq,v(Ω))
(144),(145)
&
Z∗
n
1
q − 1
p ·2k∗∗tn(β−δ)(k∗∗tn)−α0ρ(k∗∗tn) ≍
Z∗
≍ n
β−δ
θ + 1
q − 1
p (log n)
(β−δ)γ
θ −α0ρ(log n)τ
β−δ
θ (log n).
This implies the lower estimate in assertion 1 of Theorem 1.
39
(b) Let β − δ = −θ(cid:16) 1
If βv < d−θ
q − 1
p(cid:17), p > q. We take tn such that (147) holds.
q and p = q, then Theorem A and Lemma 11 yield that
en(I : W r
p,g(Ω) → Lq,v(Ω))
(144),(149)
&
Z∗
(log n)−αρ(log n).
Let βv < d−θ
m0 = m0(Z∗) that
q , p > q. We apply Corollary 1 and get that there exists such
en(I : W r
p,g(Ω) → Lq,v(Ω))
(144)
&
Z∗
& Xt>tn+m0Xi∈Jt
& Xt>tn+m0
(k∗∗t)−γτ −1(k∗∗t)(k∗∗t)
& (k∗∗tn)−α+(1−γ)( 1
q − 1
p )ρ(k∗∗tn)τ
2−k∗∗tθ(k∗∗t)
pq
p−q αρ
pq
p−q (k∗∗t)!
1
q − 1
p (143)
&
Z∗
1
q − 1
p
&
Z∗
pq
p−q αρ
pq
p−q (k∗∗t)!
1
p − 1
q (k∗∗tn)
(149)
&
Z∗
& (log n)−α+(1−γ)( 1
q − 1
p )ρ(log n)τ
1
p − 1
q (log n).
Let βv = d−θ
q . Then Theorem A and Lemma 11 imply that
en(I : W r
p,g(Ω) → Lq,v(Ω))
(146)
&
Z∗
2−θ( 1
q − 1
p )k∗∗tn(k∗∗tn)−α+ 1
q +1− 1
p ρ(k∗∗tn)n
1
q − 1
p
(148),(149)
&
Z∗
& (log n)−γ( 1
q − 1
p )−α+ 1
q +1− 1
p ρ(log n)τ
1
p − 1
q (log n).
Thus, we obtain the desired estimate in assertion 2a of Theorem 1.
(c) Let β − δ = 0, p < q. We take tn such that
n2 6 2θk∗∗tn(k∗∗tn)−γτ −1(k∗∗tn) .
Z∗
n2.
(150)
Then k∗∗tn ≍
Z∗
log n. Applying Theorem A and Lemma 11, we obtain that
en(I : W r
p,g(Ω) → Lq,v(Ω))
(143),(144),(145),(150)
&
Z∗
& (k∗∗tn)−α0ρ(k∗∗tn)n
1
q − 1
p log
1
p − 1
q (cid:18)1 +
n2
n(cid:19) ≍
Z∗
n
40
1
q − 1
p (log n)−α0+ 1
p − 1
q ρ(log n).
Now we take tn such that
2n 6 2θk∗∗tn(k∗∗tn)−γτ −1(k∗∗tn) .
Z∗
2n.
Then k∗∗tn ≍
Z∗
n. Applying Theorem A and Lemma 11, we obtain that
en(I : W r
p,g(Ω) → Lq,v(Ω))
(144),(145)
&
Z∗
(k∗∗tn)−α0ρ(k∗∗tn) ≍
Z∗
n−α0ρ(n).
Thus, we get desired estimates in assertion 2b of Theorem 1.
2. Let θ = 0 and conditions of Theorem 3 hold. Then
kψt,jkLq,v(Ω)
(144)
&
Z∗
(k∗∗t)−α log−λ(k∗∗t),
card Jt
(143)
&
Z∗
(k∗∗t)−γ log−ν(k∗∗t).
Hence, for 2m 6 k∗∗t < 2m+1 and sufficiently large m ∈ N
kψt,jkLq,v(Ω) &
Z∗
2−αmm−λ,
card (∪2m6k∗t<2m+1 Jt) &
Z∗
2m(1−γ)m−ν.
(151)
(a) Let α − (1 − γ)(cid:16) 1
q − 1
p(cid:17)+
> 0. We take mn ∈ N such that
n 6 2mn(1−γ)m−ν
n.
(152)
n .
Z∗
Then 2mn ≍
Z∗
1
1−γ (log n)
n
ν
A and Lemma 11, we get that
1−γ (see Lemma 2), mn ≍
Z∗
log n. Applying Theorem
en(I : W r
p,g(Ω) → Lq,v(Ω))
(151)
&
Z∗
2−αmnm−λ
n n
1
q − 1
p ≍
Z∗
≍ n− α
1−γ + 1
q − 1
p (log n)− αν
1−γ −λ.
Thus, we obtain the desired estimate in assertion 1 of Theorem 3.
(b) Let p > q, α = (1 − γ)(cid:16) 1
have α = 0 and
q − 1
p(cid:17). We define mn by (152). For p = q we
en(I : W r
p,g(Ω) → Lq,v(Ω))
(151),(152)
&
Z∗
(log n)−λ.
If p > q, then we apply Corollary 1 and get that there exists m = m(Z∗)
such that
en(I : W r
p,g(Ω) → Lq,v(Ω))
41
(151)
&
Z∗
& ∞
Xm=mn+ m
2m(1−γ)m−ν · 2−mα pq
p−q m−λ pq
p−q!
1
q − 1
p
≍
Z∗
−λ+(1−ν)( 1
≍ m
n
p )
q − 1
(log n)−λ+(1−ν)( 1
q − 1
p ).
≍
Z∗
Thus, we get the desired estimate in assertion 2a of Theorem 3.
(c) Let p < q, α = 0. First we take mn such that
2(1−γ)mn m−ν
n ≍
Z∗
n2.
(153)
Then mn ≍
Z∗
log n. Applying Theorem A and Lemma 11, we get
en(I : W r
p,g(Ω) → Lq,v(Ω))
(151),(153)
&
Z∗
& m−λ
n n
1
q − 1
p log
1
p − 1
q (cid:18)1 +
n2
n(cid:19) ≍
Z∗
n
1
q − 1
p (log n)−λ+ 1
p − 1
q .
Now we take mn such that 2(1−γ)mn m−ν
Theorem A and Lemma 11, we get
n ≍
Z∗
2n. Then mn ≍
Z∗
n. Applying
en(I : W r
p,g(Ω) → Lq,v(Ω))
(151)
&
Z∗
m−λ
n ≍
Z∗
n−λ.
Thus, we obtain the desired estimate in assertion 2b of Theorem 3.
This completes the proof of Theorems 1, 2, 3.
5 Estimates of entropy numbers of weighted summation
operators on a tree
Let A be a tree, and let f : V(A) → R. We set
1/p
kf klp(A) =
Xξ∈V(A)
f (ξ)p
,
if
1 6 p < ∞,
kf kl∞(A) = sup
ξ∈V(A)
f (ξ).
Denote by lp(A) the space of functions f : V(A) → R with finite norm kf klp(A).
Let u, w : V(A) → [0, ∞) be weight functions.
Define the summation operator Su,w,A by
Su,w,Af (ξ) = w(ξ)Xξ′6ξ
u(ξ′)f (ξ′),
ξ ∈ V(A),
f : V(A) → R.
42
In papers of Lifshits and Linde [26 -- 28] estimates for entropy numbers of the
operator Su,w,A : lp(A) → l∞(A) or its dual were obtained under some conditions on
u, w. Here we obtain order estimates for entropy numbers of the operator Su,w,A :
lp(A) → lq(A) for 1 < p 6 ∞, 1 6 q < ∞.
Let V(A) = {ηj,i : j ∈ Z+,
i ∈ Ij}. In addition, we suppose that the family
1 (ηj,i)}i∈Ij forms the partition of the set {ηj+1,t}t∈Ij+1. Suppose that for
of sets {VA
some c∗ > 1, m∗ ∈ N
c−1
∗
h(2−m∗j)
h(2−m∗(j+l))
6 VA
l (ηj,i) 6 c∗
h(2−m∗j)
h(2−m∗(j+l))
,
j, l ∈ Z+.
Here the function h is defined by (2), (3) in some neighborhood of zero. Let u,
w : V(A) → (0, ∞), u(ηj,i) = uj, w(ηj,i) = wj, j ∈ Z+, i ∈ Ij,
uj = 2−κum∗j(m∗j + 1)−αuρu(m∗j + 1), wj = 2−κwm∗j(m∗j + 1)−αw ρw(m∗j + 1),
where ρu, ρw : (0, ∞) → (0, ∞) are absolutely continuous functions,
lim
y→∞
yρ′
w(y)
ρw(y) = 0. Moreover, we suppose that 1 < p 6 ∞, 1 6 q < ∞,
lim
y→∞
yρ′
u(y)
ρu(y) =
κw >
θ
q
or κw =
θ
q
, αw >
1 − γ
q
.
(154)
We set κ = κu + κw, α = αu + αw, ρ(y) = ρu(y)ρw(y), Z = (p, q, u, w, h, m∗, c∗).
Let D be a subtree in A. Denote by Sp,q
D,u,w the operator norm of Su,w,D : lp(D) →
lq(D). Applying the results of [40], [46], we get that for j > 2 and for any i ∈ Ij we
have Sp,q
C(j), where C(j) is defined as follows.
Aηj,i ,u,w ≍
Z
1. Let κw > θ
q − 1
. Then C(j) = 2−κm∗j(m∗j)−αρ(m∗j).
2. Let κw > θ
q , κ > θ(cid:16) 1
q , κ = θ(cid:16) 1
p(cid:17)+
p(cid:17)+
3. Let θ > 0, κw = θ
α > 1
q , p < q. Then C(j) = 2−κm∗j(m∗j)−α+ 1
C(j) = 2−θ( 1
q − 1
p )+
q − 1
. Then
q − 1
q − 1
, α > (1 − γ)(cid:16) 1
p(cid:17)+
m∗j(m∗j)−α+( 1
p )+ρ(m∗j).
p(cid:17)+
q . Suppose that either κ > θ(cid:16) 1
q , κ = θ(cid:16) 1
, p > q, α > 1 + (1 − γ)(cid:16) 1
q − 1
q ρ(m∗j).
p(cid:17)+
q − 1
p )m∗j(m∗j)−α+1+ 1
or κ = θ(cid:16) 1
p(cid:17)+
q − 1
q − 1
p ρ(m∗j).
q − 1
,
p(cid:17)+
. Then
4. Let θ > 0, κw = θ
C(j) = 2−θ( 1
q − 1
Applying 6, we get results similar to Theorems 1, 3.
Theorem 6. Let θ > 0.
43
1. Suppose that κ > θ(cid:16) 1
κw = θ
q . Then
p(cid:17)+
q − 1
. We set α0 = α for κw > θ
q and α0 = α − 1
q for
en(Su,w,A : lp(A) → lq(A)) ≍
Z
n− κ
θ + 1
q − 1
p (log n)−α0− κγ
θ ρ(log n)τ − κ
θ (log n).
2. Suppose that κ = θ(cid:16) 1
q − 1
.
p(cid:17)+
(a) Let p > q and α0 := α − (1 − γ)(cid:16) 1
p(cid:17) > 0 for κw = θ
α − 1 − (1 − γ)(cid:16) 1
q − 1
p(cid:17) > 0 for κw > θ
q − 1
q . Then
q , α0 :=
en(Su,w,A : lp(A) → lq(A)) ≍
Z
(log n)−α0ρ(log n)τ − 1
q + 1
p (log n).
(b) Let p < q, α0 := α > 0 for κw > θ
p − 1
addition, suppose that α0 6= 1
q and α0 := α − 1
q . Then
q > 0 for κw = θ
q . In
en(Su,w,A : lp(A) → lq(A)) ≍
Z
n
1
q − 1
p (log n)−α0− 1
q + 1
p ρ(log n)
if α0 > 1
p − 1
q ,
if α0 < 1
p − 1
q .
en(Su,w,A : lp(A) → lq(A)) ≍
Z
n−α0ρ(n)
If θ = 0, κ = 0, then we suppose that ρu(t) = log(t+2)−λu, ρw = log(t+2)−λw,
> 0,
Aηj,i ,u,w are already
, then C(j) for
τ (t) = log(t + 2)ν. Denote λ = λu + λw. If κw > 0, α − (1 − γ)(cid:16) 1
p(cid:17)+
obtained. If κw > 0, α − (1 − γ)(cid:16) 1
then the bounds for C(j) from the sharp two-sided estimate of Sp,q
q − 1
j > 2 is defined as follows. If p 6 q, then C(j) = [log(m∗j)]−λ; if p > q, then
= 0, λ > (1 − ν)(cid:16) 1
p(cid:17)+
p(cid:17)+
q − 1
q − 1
C(j) = (m∗j)γ( 1
q − 1
p )[log(m∗j)]−λ+ 1
q − 1
p .
It follows from estimates in [40], [46].
Theorem 7. Suppose that θ = 0, κ = 0, κw > 0.
1. Let α − (1 − γ)(cid:16) 1
q − 1
p(cid:17)+
> 0. Then
en(Su,w,A : lp(A) → lq(A)) ≍
Z
44
n− α
1−γ + 1
q − 1
p (log n)−λ− αν
1−γ .
2. Let α − (1 − γ)(cid:16) 1
(a) Suppose that p > q. Then
q − 1
p(cid:17)+
= 0, λ > (1 − ν)(cid:16) 1
q − 1
.
p(cid:17)+
en(Su,w,A : lp(A) → lq(A)) ≍
Z
(b) Let p < q, λ 6= 1
p − 1
q . Then
(log n)−λ+(1−ν)( 1
q − 1
p ).
en(Su,w,A : lp(A) → lq(A)) ≍
Z
n
1
q − 1
p (log n)−λ+ 1
p − 1
q
for λ > 1
p − 1
p ,
for λ < 1
p − 1
q .
REFERENCES
en(Su,w,A : lp(A) → lq(A)) ≍
Z
n−λ
[1] O.V. Besov, V.P. Il'in, S.M. Nikol'skii, Integral representations of functions,
and imbedding theorems. "Nauka", Moscow, 1996. [Winston, Washington DC;
Wiley, New York, 1979].
[2] M.Sh. Birman and M.Z. Solomyak, "Piecewise polynomial approximations of
functions of classes W α
p ", Mat. Sb. 73:3 (1967), 331- -- 355.
[3] M. Bricchi,
"Existence and properties of h-sets", Georgian Mathematical
Journal, 9:1 (2002), 13(cid:22)32.
[4] B. Carl,
I. Stephani, Entropy, Compactness, and the Approximation of
Operators. Cambridge Tracts in Mathematics, V. 98. Cambridge: Cambridge
University Press, 1990.
[5] D.E. Edmunds, Yu.V. Netrusov, "Entropy numbers of operators acting between
vector-valued sequence spaces", Math. Nachr., 286:5 -- 6 (2013), 614 -- 630.
[6] D.E. Edmunds, Yu.V. Netrusov, "Schutt's theorem for vector-valued sequence
spaces", J. Approx. Theory, 178 (2014), 13 -- 21.
[7] D.E. Edmunds, H. Triebel, Function spaces, entropy numbers, differential
in Mathematics, 120 (1996). Cambridge
operators. Cambridge Tracts
University Press.
[8] D.D. Haroske, "Entropy numbers in weighted function spaces and eigenvalue
distributions of some degenerate pseudodifferential operators. I", Math. Nachr.,
167 (1994), 131 -- 156.
[9] D.D. Haroske, "Entropy numbers in weighted function spaces and eigenvalue
distributions of some degenerate pseudodifferential operators. II", Math. Nachr.,
168 (1994), 109 -- 137.
45
[10] D.D. Haroske, L. Skrzypczak, "Entropy and approximation numbers of function
spaces with Muckenhoupt weights", Rev. Mat. Complut., 21:1 (2008), 135 -- 177.
"Entropy and approximation numbers of
embeddings of function spaces with Muckenhoupt weights, II. General weights",
Ann. Acad. Sci. Fenn. Math., 36:1 (2011), 111 -- 138.
[11] D.D. Haroske, L. Skrzypczak,
[12] D.D. Haroske, L. Skrzypczak, "Entropy numbers of embeddings of function
spaces with Muckenhoupt weights, III. Some limiting cases," J. Funct. Spaces
Appl. 9:2 (2011), 129 -- 178.
[13] D.D. Haroske, L. Skrzypczak, "Spectral theory of some degenerate elliptic
operators with local singularities", J. Math. An. Appl., 371:1 (2010), 282 -- 299.
[14] D.D. Haroske, H. Triebel, "Wavelet bases and entropy numbers in weighted
function spaces", Math. Nachr., 278:1 -- 2 (2005), 108 -- 132.
[15] A.N. Kolmogorov, V.M. Tikhomirov, "ε-entropy and ε-capacity of sets in
function spaces" (Russian) Uspehi Mat. Nauk, 14:2(86) (1959), 3 -- 86.
[16] T. Kuhn, "Entropy numbers of diagonal operators of logarithmic type", Georgian
Math. J. 8:2 (2001), 307-318.
[17] T. Kuhn, "A lower estimate for entropy numbers", J. Appr. Theory, 110 (2001),
120 -- 124.
[18] T. Kuhn, "Entropy numbers of general diagonal operators", Rev. Mat. Complut.,
18:2 (2005), 479 -- 491.
[19] Th. Kuhn, H.-G. Leopold, W. Sickel, and L. Skrzypczak. "Entropy numbers of
embeddings of weighted Besov spaces", Constr. Approx., 23 (2006), 61 -- 77.
[20] Th. Kuhn, H.-G. Leopold, W. Sickel, L. Skrzypczak, "Entropy numbers of
embeddings of weighted Besov spaces II", Proc. Edinburgh Math. Soc. (2) 49
(2006), 331 -- 359.
[21] T. Kuhn, "Entropy Numbers in Weighted Function Spaces. The Case of
Intermediate Weights", Proc. Steklov Inst. Math., 255 (2006), 159 -- 168.
[22] Th. Kuhn, H.-G. Leopold, W. Sickel, and L. Skrzypczak, "Entropy numbers of
embeddings of weighted Besov spaces III. Weights of logarithmic type", Math.
Z., 255:1 (2007), 1 -- 15.
[23] T. Kuhn, "Entropy numbers in sequence spaces with an application to weighted
function spaces", J. Appr. Theory, 153 (2008), 40 -- 52.
[24] W.V. Li, W. Linde, "Approximation, metric entropy and small ball estimates
for Gaussian measures", Ann. Probab., 27:3 (1999), 1556 -- 1578.
[25] M.A. Lifshits, W. Linde, "Approximation and entropy numbers of Volterra
operators with application to Brownian motion", Mem. Amer. Math. Soc.,
157:745, Amer. Math. Soc., Providence, RI, 2002.
[26] M.A. Lifshits, "Bounds for entropy numbers for some critical operators", Trans.
Amer. Math. Soc., 364:4 (2012), 1797 -- 1813.
46
[27] M.A. Lifshits, W. Linde, "Compactness properties of weighted summation
operators on trees", Studia Math., 202:1 (2011), 17 -- 47.
[28] M.A. Lifshits, W. Linde, "Compactness properties of weighted summation
operators on trees (cid:22) the critical case", Studia Math., 206:1 (2011), 75 -- 96.
[29] E.N. Lomakina, V. D. Stepanov, "Asymptotic estimates for the approximation
and entropy numbers of the one-weight Riemann -- Liouville operator", Mat. Tr.,
9:1 (2006), 52 -- 100 [Siberian Adv. Math., 17:1 (2007), 1 -- 36].
[30] P. Mattila, Geometry of sets and measures in Euclidean spaces. Cambridge
Univ. Press, 1995.
[31] T. Mieth, "Entropy and approximation numbers of embeddings of weighted
Sobolev spaces", J. Appr. Theory, 192 (2015), 250 -- 272.
[32] A. Pietsch, Operator ideals. Mathematische Monographien [Mathematical
Monographs], 16. Berlin, 1978. 451 pp.
[33] Yu.G. Reshetnyak,
"Integral representations of differentiable functions in
domains with a nonsmooth boundary", Sibirsk. Mat. Zh., 21:6 (1980), 108 -- 116
(in Russian).
[34] Yu.G. Reshetnyak, "A remark on integral representations of differentiable
functions of several variables", Sibirsk. Mat. Zh., 25:5 (1984), 198 -- 200 (in
Russian).
[35] C. Schutt, "Entropy numbers of diagonal operators between symmetric Banach
spaces", J. Appr. Theory, 40 (1984), 121 -- 128.
[36] V.M. Tikhomirov, "The ε-entropy of certain classes of periodic functions"
(Russian) Uspehi Mat. Nauk 17:6 (108) (1962), 163 -- 169.
[37] H. Triebel, "Entropy and approximation numbers of limiting embeddings, an
approach via Hardy inequalities and quadratic forms", J. Approx. Theory, 164:1
(2012), 31 -- 46.
[38] A.A. Vasil'eva, "Widths of weighted Sobolev classes on a John domain", Proc.
Steklov Inst. Math., 280 (2013), 91 -- 119.
[39] A.A. Vasil'eva, "Embedding theorem for weighted Sobolev classes on a John
domain with weights that are functions of the distance to some h-set", Russ. J.
Math. Phys., 20:3 (2013), 360 -- 373.
[40] A.A. Vasil'eva, "Embedding theorem for weighted Sobolev classes on a John
domain with weights that are functions of the distance to some h-set", Russ. J.
Math. Phys. 21:1 (2014), 112 -- 122.
[41] A.A. Vasil'eva, "Widths of weighted Sobolev classes on a John domain: strong
singularity at a point", Rev. Mat. Compl., 27:1 (2014), 167 -- 212.
[42] A.A. Vasil'eva, "Widths of function classes on sets with tree-like structure", J.
Appr. Theory, 192 (2015), 19 -- 59.
[43] A.A. Vasil'eva, "Kolmogorov and linear widths of the weighted Besov classes
with singularity at the origin", J. Appr. Theory, 167 (2013), 1 -- 41.
47
[44] A.A. Vasil'eva, "Some sufficient conditions for embedding a weighted Sobolev
class on a John domain", Sib. Mat. Zh., 56:1 (2015), 65 -- 81.
[45] A.A. Vasil'eva, "Widths of weighted Sobolev classes with weights that are
functions of distance to some h-set: some limiting cases", arXiv:1502.01014v1.
[46] A.A. Vasil'eva, "Estimates for norms of two-weighted summation operators on
a tree under some restrictions on weights", Math. Nachr. 1 -- 24 (2015), /DOI
10.1002/mana.201300355.
[47] A.A. Vasil'eva, "Widths of weighted Sobolev classes on a domain with a peak:
some limiting cases", arXiv:1312.0081.
48
|
1002.3123 | 2 | 1002 | 2010-02-17T13:34:40 | Local behavior of traces of Besov functions: Prevalent results | [
"math.FA",
"math-ph",
"math-ph"
] | Let $1 \leq d < D$ and $(p,q,s)$ satisfying $0 < p < \infty$, $0 < q \leq \infty$, $0 < s-d/p < \infty$. In this article we study the global and local regularity properties of traces, on affine subsets of $\R^D$, of functions belonging to the Besov space $B^{s}_{p,q}(\R^D)$. Given a $d$-dimensional subspace $H \subset \R^D$, for almost all functions in $B^{s}_{p,q}(\R^D)$ (in the sense of prevalence), we are able to compute the singularity spectrum of the traces $f_a$ of $f$ on affine subspaces of the form $a+H$, for Lebesgue-almost every $a \in \R^{D-d}$. In particular, we prove that for Lebesgue-almost every $a \in \R^{D-d}$, these traces $f_a$ are more regular than what could be expected from standard trace theorems, and that $f_a$ enjoys a multifractal behavior. | math.FA | math |
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS:
PREVALENT RESULTS
JEAN-MARIE AUBRY, DELPHINE MAMAN, AND STÉPHANE SEURET
Abstract. Let 1 ≤ d < D and (p, q, s) satisfying 0 < p < ∞, 0 < q ≤
∞, 0 < s − d/p < ∞. In this article we study the global and local regu-
larity properties of traces, on affine subsets of RD, of functions belonging
p,q(RD). Given a d-dimensional subspace H ⊂ RD,
to the Besov space Bs
p,q(RD) (in the sense of prevalence), we are
for almost all functions in Bs
able to compute the singularity spectrum of the traces fa of f on affine
subspaces of the form a + H, for Lebesgue-almost every a ∈ RD−d. In
particular, we prove that for Lebesgue-almost every a ∈ RD−d, these
traces fa are more regular than what could be expected from standard
trace theorems, and that fa enjoys a multifractal behavior.
1. Introduction
Investigating regularity properties of traces of functions belonging to some
Besov or Sobolev spaces is a longstanding issue. For instance, such questions
arise from PDE's theory, where the Dirichlet condition imposes some regu-
larity properties of the trace of the solution on the frontier of the domain.
In this article, we study the local behavior of traces of functions belonging
to the Besov space Bs
p,q(RD) on d-dimensional affine subspaces of RD.
Not only concerned with global smoothness properties (i.e.
to which
Sobolev and Besov spaces the traces belong), we will especially focus on
the local behavior of such traces. The notion of pointwise regularity we dis-
cuss in the sequel is the following. Given a real function f ∈ L∞
loc(RD) and
x0 ∈ RD, f is said to belong to Cα(x0), for some α ≥ 0, if there exists a
polynomial P of degree at most (cid:98)α(cid:99) and a constant C > 0 such that locally
around x0 :
(1)
The local regularity of f at x0 is measured by the pointwise Hölder exponent :
f (x) − P (x − x0) ≤ Cx − x0α.
hf (x0) = sup{α ≥ 0 : f ∈ Cα(x0)}.
As will be observed soon, this exponent hf (x0) may vary rather erratically
with x0, and the relevant information is then provided by the spectrum of
singularities df of f, which is the function :
df : h ∈ [0,∞] (cid:55)−→ dimH Ef (h), where Ef (h) := {x0 ∈ RD : hf (x0) = h}.
2000 Mathematics Subject Classification. Prim. 46E35; Second. 26B35, 28A80, 37C20.
Key words and phrases. Besov space; Trace theorem; Pointwise regularity; Hausdorff
dimension and measures; Prevalence; Wavelets.
1
2
J.-M. AUBRY, D. MAMAN, AND S. SEURET
Here dimH stands for the Hausdorff dimension. We adopt the convention
that dimH ∅ = −∞. The spectrum of singularities df describes the geomet-
rical repartition of the singularities of f.
This spectrum and its relevance in physics, especially in fluid mechanics,
goes back to the 1980's. At this time, physicists have been able to measure
one coordinate of the velocity of a turbulent fluid, and they observed that
their signals exhibited very different local behaviors at different times. This
variability was proposed by Frisch and Parisi as a possible explanation for
the concavity of the scaling function associated with the velocity (see [8] and
several references on the subject). These works are very intimately related
to our questions, since only the trace of the fluid's velocity is measured in
practice. Hence, to infer some results on the regularity properties of the
three-dimensional velocity, it is key to investigate the possible local behavior
of traces of Sobolev or Besov functions.
Precise results on the pointwise regularity of functions belonging to clas-
p,∞(RD) spaces have recently been obtained [1,
sical spaces such as Besov Bs
7, 11, 12]. These results are of two kinds: universal upper bounds for the
spectrum of singularities (valid for any element of the space) and almost-sure
spectrum (valid for a "large" subset of the space, in the sense of prevalence
or Baire categories). We detail these results, as well as ours, now.
In all that follows, 0 < d < D are two fixed integers. Let d(cid:48) := D − d and
= RD. For a ∈ Rd(cid:48) we shall denote by Ha := {(x, a)} the
(x, x(cid:48)) ∈ Rd × Rd(cid:48)
d-dimensional affine subspace of RD.
Let f be a continuous function on RD. Its trace on Ha is
fa := fHa : Rd −→ R
x (cid:55)−→ f (x, a)
(cid:0)F−1(cid:0)1ξ≤NFf(cid:1)(cid:1)
a
If f is not continuous, its trace can be defined by Fourier regularization: we
shall again write fa for lim
, whenever that limit exists.
N→∞
Standard trace theorems inevitably involve a loss of regularity, for in-
stance it is well known that when s > 1/2, the trace of f ∈ H s(R2) on
any one-dimensional subspace belongs to H s−1/2(R). Similar results hold
for Besov spaces (see § 2.4): it can easily be shown that the trace operator
f (cid:55)→ fa maps Bs
p,∞ (Rd). However, one may expect better
regularity properties for most of the traces fa. Indeed, according to a result
of Jaffard [10]:
Theorem 1.1. Let 0 < p, s < ∞. If f ∈ Bs
almost all a ∈ Rd(cid:48)
p,∞(RD), then for Lebesgue-
p,∞(RD) to Bs−d(cid:48)/p
fa ∈ (cid:84)
Bs(cid:48)
p,∞(Rd).
s(cid:48)<s
In particular, when s > d/p, the inclusion Bs−ε
p,∞(Rd) (cid:44)→ Cs−ε−d/p(Rd) for
any ε > 0 small enough implies that Lebesgue-almost all traces fa exist and
are uniform Hölder functions. For the sake of integrity, we present a short
and independent proof of Theorem 1.1 in Appendix A.
By another result of Jaffard [11], belonging to a Besov space yields an
upper bound on the spectrum of singularities:
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
Theorem 1.2. Let 0 < p < ∞ and d/p < s < ∞. For any g ∈ Bs
for all h ≥ s − d/p,
3
p,∞(Rd),
dg(h) ≤ min(d, d + (h − s)p),
and Ef (h) = ∅ if h < s − d/p.
Remark 1.3. The results so far have been stated for Besov spaces Bs
p,q with
q = ∞ but it is clear from classical Besov embeddings (see equation (10)
below) that they hold identically for any q > 0.
Not only is Theorem 1.2 optimal, the upper bound is actually an almost
p,q(RD) (Theorem 1.4) in the sense of prevalence, as ex-
sure equality in Bs
plained below.
Prevalence theory is used to supersede the notion of Lebesgue measure in
any real or complex topological vector space E. This notion was proposed by
Christensen [4] and independently by Hunt et al. [9]. The space E is endowed
with its Borel σ-algebra B(E) and all Borel measures µ on (E,B(E)) will be
automatically completed, that is we put µ(A) := µ(B) if B ∈ B(E) and the
symmetric difference A∆B is included in some D ∈ B(E) with µ(D) = 0. A
set is said to be universally measurable if it is measurable for any (completed)
Borel measure.
Definition 1. A universally measurable set A ⊂ E is called shy if there
exists a Borel measure µ that is positive on some compact subset K of E and
such that
for every x ∈ E, µ(A + x) = 0.
More generally, a set that is included in a shy universally measurable set is
also called shy.
any x ∈ E and if (Bn)n∈N is a sequence of prevalent sets then so is(cid:84)
Finally, the complement in E of a shy subset is called prevalent.
The measure µ used to show that some subset is shy or prevalent is called
a probe. It can be for instance the Lebesgue measure carried by some finite-
dimensional subspace of E: this is the technique that will be used in § 4.2.
When a set B is prevalent, it is dense in E, B + x is also prevalent for
n∈N Bn.
Finally, when E has finite dimension, B is prevalent in E if and only if it has
full Lebesgue measure. This justifies that a prevalent set B is a "large" set
in E and extends reasonably the notion of full Lebesgue measure to infinite
dimensional spaces.
From now on, without any possible confusion, the term "almost all" will
be indiscriminately used to describe elements in a prevalent subset of an
infinite-dimensional space, or in a subset having full Lebesgue measure in a
finite-dimensional space.
In this setting, the following was proved by Fraysse and Jaffard [7]:
Theorem 1.4. Let 0 < p < ∞, 0 < q ≤ ∞ and 0 < s − D/p < ∞. For
almost all g ∈ Bs
p,q(RD),
(cid:40)
dg(h) =
D + (h − s)p
−∞
if h ∈ [s − D/p, s]
else
and for x in a set of full Lebesgue measure in RD, hg(x) = s.
4
J.-M. AUBRY, D. MAMAN, AND S. SEURET
Remark 1.5. Another notion of genericity is given by Baire's theory: a
property is said to be quasi-sure in a complete metric space E if this property
is realized on a residual (comeagre) set in E. We choose to work within the
prevalence framework, but Baire's genericity is also worthy of interest and
will be studied in a subsequent paper.
In this paper we prove the following result on the singularity spectrum of
traces of almost all Besov functions.
Theorem 1.6. Let 0 < p < ∞, 0 < q ≤ ∞ and 0 < s − d/p < +∞.
p,q(RD), for Lebesgue-almost all a ∈ Rd(cid:48), the following
For almost all f in Bs
holds:
(cid:40)
(i) the spectrum of singularities of fa is
(2)
dfa(h) =
d + (h − s)p
−∞
if h ∈ [s − d/p, s]
else.
(ii) for every open set Ω ⊂ Rd, the level set Efa(s)∩ Ω has full Lebesgue
measure in Ω.
Let us make some remarks on Theorem 1.6:
• In a given Besov space Bs
p,q(RD) (Theorem 1.4), as well as in Cα(RD) [13]
or for Borel measures supported by [0, 1]D [2], the almost-sure regularity is
often the "worst possible", i.e. the upper bound on the spectrum valid for all
elements of the considered space turns out to be an equality for almost all
functions or measures. This is not the case in Theorem 1.6, for which the
almost sure spectrum does not coincide with the a priori upper bound, and
thus the traces are more regular than what could be expected a priori.
• Observe that the singularities with Hölder exponent h less than s − d/p
are "not seen" by Lebesgue-almost every traces fa. This corresponds to the
level sets Ef (h) of Hausdorff dimension less than d(cid:48). B. Mandelbrot referred
to this phenomenon as negative dimensions: By this, he means that almost
p,q(RD) possesses singularities with exponent s−D/p ≤
every function f ∈ Bs
h < s − d/p, but these singularities form a set of too small a dimension to
intersect a large quantity among the hyperplanes Ha of dimension d(cid:48) = D−d.
Remark 1.7. Theorem 1.6 and the above remark are reminiscent of classi-
cal results of P. Mattila [15] on the Hausdorff dimensions of intersection of
fractal subset of RD with Lebesgue-almost all d-dimensional hyperplanes, or
of sliced measures [15, 14].
In this theorem, all the hyperplanes Ha on which the traces are taken are
parallel (to the d first coordinates axes). Since the Besov spaces are invari-
ant by unitary transformation of the coordinates, the result remains valid
in any fixed direction. Thanks to the stability of prevalence by countable
intersection, we thus obtain:
Corollary 1.8. Let ∆ be a countable subset of the Grassmannian Grd(D).
p,q(RD), for any
Under the same hypotheses on p, q, s, for almost all f in Bs
H ∈ ∆, for Lebesgue-almost all a ∈ H⊥, the trace of f on H + a has the
properties stated in Theorem 1.6.
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
5
Figure 1. Singularity spectrum of almost all f ∈ Bs
and its trace fa for Lebesgue almost every a ∈ Rd(cid:48).
p,q(RD)
Unfortunately no Fubini theorem holds for prevalence, so we cannot di-
rectly deduce from this the natural generalization below, which we leave for
subsequent studies.
Conjecture 1.9. Consider the Grassmannian Grd(D) and its Haar measure
p,q(RD), for µd,D-almost all H ∈ Grd(D), for
µd,D. For almost all f in Bs
Lebesgue-almost all a ∈ H⊥, the trace of f on H + a has the properties stated
in Theorem 1.6.
The paper is organized as follows. Our method is based on wavelets, and
requires various notions of real and functional analysis. Section 2 provides
all the definitions and important results needed to complete the proof of
Theorem 2.1. In Section 3, we prove the upper bound for the singularity
p,q(RD), and the lower bound for all functions
spectrum for all functions in Bs
in a set that we call F. Then, in Section 4, we show that this set F is
prevalent, the main difficulties lying in the measurability properties of F.
Appendix A contains a shorter proof of Theorem 1.1, and Appendix B deals
with the universal measurability of F in the case q = +∞ (which differs
from the case q < +∞ since Bs
p,∞(RD) is not separable).
2. Preliminaries
2.1. Dimensions.
Two notions of dimensions of sets in Rd will be used below: the Hausdorff
dimension and the upper box dimension. We recall them quickly.
Let E be a bounded set in Rd. For every ε > 0, denote by Nε(E) the
minimal number of cubes of size ε needed to cover the set E. The upper box
dimension of E, denoted by dimB(E), is the real number ∈ [0, d] defined as
(3)
.
dimB(E) = lim sup
ε→0+
log Nε(E)
− log ε
as−d/ps−D/psh0Dd−d'd (h)fd (h)f 6
J.-M. AUBRY, D. MAMAN, AND S. SEURET
For the reader's convenience we also recall the definition of the Hausdorff
dimension.
Definition 2. Let s ≥ 0. The s-dimensional Hausdorff measure of a set E,
Hs(E), is defined as
Hs(E) = lim
r(cid:38)0
Hs
r(E), with Hs
r(E) = inf
Ei ≤ r and E ⊂(cid:83)
defined as
the infimum being taken over all the countable families of sets Ei such that
i Ei. Then, the Hausdorff dimension of E, dimH E, is
(cid:110)(cid:88)
Eis(cid:111)
,
i
dimH E = inf{s ≥ 0 : Hs(E) = 0} = sup{s ≥ 0 : Hs(E) = +∞}.
For a bounded set E ⊂ Rd, we have
0 ≤ dimH(E) ≤ dimB(E) ≤ d.
2.2. Wavelets.
vanishing moments (i.e. (cid:82)
We recall very briefly the basics of multiresolution wavelet analysis (for
details see for instance [5]). For an arbitrary integer N ≥ 1 one can construct
compactly supported functions Ψ0 ∈ CN (R) (called the scaling function)
and Ψ1 ∈ CN (R) (called the mother wavelet), with Ψ1 having at least N + 1
R xnΨ1(x)dx = 0 for n ∈ {0, . . . , N}), and such
j,k : x (cid:55)→ Ψ1(2jx − k)
Ψ1
that the set of functions
for j ∈ Z, k ∈ Z form an orthogonal basis of L2(R) (note that we choose the
L∞ normalization, not L2). In this case, the wavelet is said to be N-regular.
Let us introduce the notations
0d := (0, 0,··· , 0),
An orthogonal basis of L2(Rd) is then obtained by tensorization. For every
1d := (1, 1,··· , 1),
Ld := {0, 1}d\0d.
λ := (j, k, l) ∈ Z × Zd × {0, 1}d, let us define the tensorized wavelet
Ψλ(x) :=
Ψli
j,ki
(xi),
d(cid:89)
i=1
(cid:88)
(cid:90)
Rd
Any function f ∈ L2(Rd) can be written (the inequality being true in
with obvious notations that k = (k1, k2,··· , kd) and l = (l1, l2,··· , ld).
L2(Rd))
(4)
f =
cλΨλ(x),
λ=(j,k,l): j∈Z, k∈Zd, l∈Ld
where
(5)
cλ := 2jd
f (x)Ψλ(x) dx.
It is implicit in (5) that the wavelet coefficients depend on f. Observe that
in the wavelet decomposition (4), no wavelet Ψλ such that l = 0d (where
λ = (j, k, l)) appears.
Similar notations (e.g. λD := (j, (k, k(cid:48)), (l, l(cid:48))) ∈ Z×ZD×{0, 1}D) with the
straightforward modifications will produce an orthogonal basis of L2(RD).
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
7
Figure 2. Plot of the periodized Daubechies wavelet with
eight vanishing moments, and its derivative. The approxima-
tion is precise enough to ensure that G and its derivative do
not vanish at the same time.
The wavelets and the corresponding wavelet coefficients in L2(RD) will be
denoted respectively by ΨλD and cλD.
In 4.2 we shall need to consider the 1-periodic function
k∈Z
and make the technical hypothesis on Ψ1:
G : t ∈ R (cid:55)−→(cid:88)
(i) Ψ1 is N-regular,
(cid:48)
(6)
(HN)
(ii) The set Z := G−1({0}) ∩ [0, 1] is finite,
(iii) For every t ∈ Z, G
(t) > 0 .
Ψ1(t − k).
This condition is very reasonable for a given wavelet Ψ1. Numerical sim-
ulations (see figure 2) indicate that (HN ) is verified for suitable choices of
regular wavelets, including in particular Daubechies's compactly supported
wavelets [5]. In Figure 2, the simulations of Ψ1 and (Ψ1)(cid:48) (computed using
the associated wavelet filters) are precise enough to guarantee that G(cid:48) does
not vanish around the zeros of G.
2.3. Localization of the problem.
We will be first focusing on the local behavior of traces on (0, 1)d×{a}, a ∈
(0, 1)d(cid:48). As Proposition 2.3 shows, if f is written as (4), only the coefficients
λ such that j ≥ 0 and (k2−j, k(cid:48)2−j) ∈ [0, 1]D can play a role in the value of
cD
the pointwise exponent hfa(x). For our purpose, we can identify functions
λ when (k2−j, k(cid:48)2−j) ∈ [0, 1]D.
that have the same wavelet coefficients cD
Hence we will consider functions f of the form
(cid:88)
(7)
f =
cλD ΨλD (x),
λD∈ΛD×LD
8
where
J.-M. AUBRY, D. MAMAN, AND S. SEURET
for j ≥ 1, Zj = {0, 1,··· , 2j − 1} and ΛD
j = {j} × ZD
j
(cid:91)
j≥1
ΛD =
ΛD
j .
If we prove Theorem 1.6 on [0, 1]D instead of RD, then by dilation it will
be true on any cube [−N, N ]D. Prevalence results being stable by countable
intersection on N ∈ N, Theorem 1.6 will thus be obtained.
We shall present our results in this framework, and we will effectively
prove the following:
Theorem 2.1. Let 0 < p < ∞, 0 < q ≤ ∞. Assuming that s > d/p, for
p,q([0, 1]D), for almost all a ∈ [0, 1]d(cid:48), the following holds:
almost all f in Bs
(i) the spectrum of singularities of f is
(8)
dfa(h) =
d + (h − s)p
−∞
if h ∈ [s − d/p, s]
else.
(cid:40)
(ii) the level set Efa(s) has full Lebesgue measure in [0, 1]d.
2.4. Characterization of local and global regularity properties.
Let 0 < s < ∞, 0 < p, q ≤ ∞. Assume that the wavelet Ψ is at least
p,q([0, 1]D) Besov norm (quasi-norm when p < 1 or
[s + 1]-regular. The Bs
q < 1) of a function f on [0, 1]D having wavelet coefficients cλD is defined as
(cid:18)(cid:88)
(cid:18)
2(sp−D)j (cid:88)
j≥1
(k,k(cid:48))∈ZD
j
p(cid:19) 1
(cid:19) q
q
cλDp
(9)
(cid:107)f(cid:107)Bs
p,q
=
The following standard embeddings are easy to deduce from (9): For any
with the obvious modifications when p = ∞ or q = ∞. The Besov space
p,q([0, 1]D) is naturally the set of functions with finite (quasi-)norm. It is a
Bs
complete metrizable space, normed when p and q ≥ 1, separable when both
are finite.
0 < s < ∞, 0 < p ≤ ∞, 0 < q < q(cid:48) ≤ ∞, ε > 0,
(10)
Remark 2.2. In contrast with Theorems 1.1 and 1.2, the prevalence result
for a given q < ∞ cannot simply be deduced from the result for q = ∞ (nor
the other way round). Indeed it can be shown that in (10) each included space
is shy in the next one.
p,q(cid:48)([0, 1]D) (cid:44)→ Bs−ε
p,q([0, 1]D) (cid:44)→ Bs
Bs
p,q ([0, 1]D)
Let us finally recall the fundamental result linking pointwise regularity
and the size of wavelet coefficients, which justifies our approach.
Proposition 2.3. Suppose that γ > 0 and the wavelet Ψ is at least [γ +
1]-regular. Let f : [0, 1]d → R be a locally bounded function with wavelet
coefficients {cλ}, and let x ∈ [0, 1]d.
If f ∈ Cγ(x), then there exists a constant M < ∞ such that for all
λ = (j, k, l) ∈ Λd × Ld,
(11)
cλ ≤ M(cid:0)2−j +(cid:12)(cid:12)x − k2−j(cid:12)(cid:12)(cid:1)γ
= M 2−jγ(1 +(cid:12)(cid:12)2jx − k(cid:12)(cid:12))γ
Reciprocally, if (11) holds true and if f ∈(cid:83)
for every η > 0.
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
9
ε>0 Cε([0, 1]d), then f ∈ Cγ−η(x),
Finally, the notion of cone of influence will be needed later.
Definition 3. Let L > 0. The cone of influence of width L above x ∈ Rd is
the set of cubes (j, k, l) ∈ Λd such that
x − k 2−j ≤ L2−j.
2.5. Traces.
Recall that for a ∈ [0, 1]d(cid:48)
and f continuous on [0, 1]D, the function fa is
simply defined as fa(x) := f (x, a). Moreover, recall that λ = (j, k, l) with
j ∈ N∗, k ∈ Zd
j and l ∈ {0, 1}d and that λD = (j, (k, k(cid:48)), (l, l(cid:48))) with j ∈ N∗,
k ∈ Zd
j , k(cid:48) ∈ Zd(cid:48)
j , l ∈ {0, 1}d and l(cid:48) ∈ {0, 1}d(cid:48). Using the expansion (7) of f
in the tensorized wavelet basis {ΨλD}, we have
(cid:88)
d(cid:89)
d(cid:48)(cid:89)
cλD
Ψli
j,ki
(xi)
i=1
i=1
Ψl(cid:48)
j,k(cid:48)
i
i
(ai)
fa(x) =
λD∈ΛD×LD
= Ga(x) + Fa(x)
(12)
where
(13)
(14)
Ga(x)
:=
Fa(x)
:=
dλ(a)Ψλ(x)
dλ(a)Ψλ(x)
(cid:88)
(cid:88)
λ∈Λd×0d
λ∈Λd×Ld
(cid:88)
(cid:88)
λD=(j,(k,k(cid:48)),(0d,l(cid:48))):
k(cid:48)∈Zd(cid:48)
j , l(cid:48)∈Ld(cid:48)
λD=(j,(k,k(cid:48)),(l,l(cid:48))):
j , l(cid:48)∈{0,1}d(cid:48)
k(cid:48)∈Zd(cid:48)
d(cid:48)(cid:89)
d(cid:48)(cid:89)
i=1
i=1
cλD
Ψl(cid:48)
j,k(cid:48)
i
i
(ai).
cλD
Ψl(cid:48)
j,k(cid:48)
i
(ai).
i
and for λ = (j, k, l) ∈ Λd × {0, 1}d,
(15)
if l = 0d,
dλ(a) :=
(16)
if l ∈ Ld,
dλ(a) :=
Formula (14) indeed yields a wavelet decomposition of the function Fa, since
the wavelets appearing in (14) form a wavelet basis of L2([0, 1]d) (if completed
by the function Ψ0,0d,0d). This is not the case for the function Ga with
formula (13), since only the scaling function Ψ0 is used in this decomposition.
Fortunately, we have the following standard result for the Besov properties
of a function Ga defined through a formula like (13).
Proposition 2.4. If s0 > 0, and g(x) = (cid:80)
λ∈Λd×0d dλΨλ(x) with {dλ}
satisfying
(cid:88)
λ=(j,k,l): k∈Zd
j , l=0d
< +∞,
dλp
(17)
2j(p0s0−d)
sup
j≥1
10
then g ∈ Bs0
p0,∞([0, 1]d).
J.-M. AUBRY, D. MAMAN, AND S. SEURET
Proposition 2.4 entails that the same Besov characterization as (9) when
one considers only scaling functions. The proof of Proposition 2.4, that we
do not reproduce here, consists of decomposing each scaling function Ψλ,
for λ = (j, k, 0d) on the wavelets of smaller frequencies, i.e. on Ψ(cid:101)λ with
(cid:101)λ = ((cid:101)j,(cid:101)k,(cid:101)l) such that(cid:101)j ≤ j and(cid:101)l ∈ Ld.
(cid:88)
As a conclusion, the trace fa can be written
(18)
fa =
dλ(a)Ψλ(x)
λ∈Λd×{0,1}d
Recalling now Theorem 1.1 (proved in Appendix A), fa ∈(cid:84)
where for λ = (j, k, l) ∈ Λd × {0, 1}d, dλ(a) is given by (15) and (16).
For such a decomposition, the Besov characterization (9) holds true, the
difference with (4) is that the sum over λ ∈ Ld is replaced by λ ∈ {0, 1}d.
p,∞([0, 1]d)
for Lebesgue-almost every a ∈ [0, 1]d. Hence, still for almost every a, we can
consider the effective wavelet decomposition of fa on the wavelet basis pro-
vided by (4), and we write
ε>0 Bs−ε
(19)
fa = c0,0d,0dΨ0,0d,0d(x) +
cλ(a)Ψλ(x).
(cid:88)
λ∈Λd×Ld
We will use both forms (18) and (19).
2.6. Dyadic approximation.
Let B(x, r) denote the closed l∞ ball of radius r around x in [0, 1]d. For
α ≥ 1 and j ∈ N, let
(cid:91)
k∈Zd
j
X α
j
:=
B(k2−j, 2−jα)
and
X α := lim sup
j→∞
X α
j
(cid:12)(cid:12)x − Kn2−Jn(cid:12)(cid:12) ≤ 2−αJn.
The set X α is constituted by points in [0, 1]d that are approached at rate
at least α by dyadics. In other words, x ∈ X α if and only if there exists a
sequence (Jn, Kn)n≥1 ∈ Λd such that Jn → +∞ and for all n ∈ N
(20)
Observe that X 1 = [0, 1]d and if α ≤ α(cid:48) then X α(cid:48) ⊂ X α. Observe also that
if x ∈ X α is not itself a dyadic, then the sequence (Jn, Kn) can be chosen
so that for every n the fraction Kn
2Jn is irreducible. We call (Jn, Kn)n≥1 an
irreducible sequence.
About the dimension of X α, a well know result (for instance proved in [6])
states:
Theorem 2.5. There exists a positive σ-finite measure mα carried by X α
and such that any set E having Hausdorff dimension dimH(E) < d
α has
measure mα(E) = 0.
In particular, mα(X α) > 0 and dimH X α = d/α.
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
11
2.7. Prevalence, universal measurability, analytic sets.
In the Definition 1 of the prevalence in a complete metric space E, a set
B ⊂ E needs to be universally measurable to be shy or prevalent (this in-
cludes the Borel sets). One main difficulty occurring in the proof of Theorem
1.6 lies in the universal measurability property of subsets of E for which we
aim to prove a prevalence property. Indeed, these sets will be defined through
complicated formulas, not easily tractable. In particular, these subsets of E
can often be viewed as continuous images of Borel sets.
p,q(RD) when q < ∞),
When E is a Polish space (this is the case for Bs
When E is not Polish (in our context, when E = Bs
such sets are called analytic, and we have the following theorem [3].
Theorem 2.6. Every analytic set in a Polish space is universally measurable.
p,∞(RD)), continuous
images of Borel sets need not be universally measurable. Hence, in order to
obtain the universal measurability for our specific sets, the definition of an
analytic set has to be modified and is more complicated (see § B.1). Once
this second definition is adopted, the same result as Theorem 2.6 holds, i.e.
analyticity implies universal measurability. The fact that the sets we will
meet indeed satisfy this second definition of analytic set is proved in § B.2.
3. Proof of Theorem 2.1
3.1. A first property of the wavelet.
Recall the definition (6) of the periodized wavelet G and let us introduce
its d(cid:48)-dimensional version Gd(cid:48) defined as
(21)
Proposition 3.1. If Ψ satisfies (HN), then the set
Gd(cid:48) : x ∈ Rd(cid:48) (cid:55)−→ G(x1) · G(x2)··· G(xd(cid:48)).
: ∃ ja ∈ N such that ∀j ≥ ja,(cid:12)(cid:12)Gd(cid:48)(2ja)(cid:12)(cid:12) > j−2d(cid:48)(cid:111)
a ∈ [0, 1]d(cid:48)
A1 :=
(cid:110)
has full Lebesgue measure.
Remark that Proposition 3.1 holds in fact for any function G ∈ CN (R)
satisfying assumptions (ii) and (iii) of (HN).
Proof. Obviously, if we are able to prove that the set
1 :=(cid:8)a1 ∈ [0, 1] : ∃ ja ∈ N such that ∀j ≥ ja,(cid:12)(cid:12)G(2ja1)(cid:12)(cid:12) > j−2(cid:9)
A(cid:48)
1)d(cid:48) ⊂ A1.
has full Lebesgue measure in [0, 1], then Proposition 3.1 will be proved since
we have the inclusion (A(cid:48)
By (HN), there is a finite number, say y1, y2, ··· , yp, of zeros of G on
the interval [0, 1], and G(cid:48) does not vanish at these real numbers. Let M =
min(G(cid:48)(y1),G(cid:48)(y2),··· ,G(cid:48)(yp)). For each yi there is a small interval
[yi − ri, yi + ri] around yi on which G(a) ≥ M/2a − yi.
Let us denote by m the minimum of G on the compact set [0, 1]\(cid:83)p
every a /∈(cid:83)p
i=1(yi−
The above construction guarantees that for every integer n ≥ n1, for
M·n ], we have G(a) ≥ 1/n. In other words,
r, yi + r). We now choose an integer n1 such that 1/n1 ≤ min(m, r).
Let r = min(ri : i = 1, ..., p).
i=1[yi − 2
M·n , yi + 2
12
G(a) < 1/n except on a set of at most Lebesgue measure(cid:80)p
J.-M. AUBRY, D. MAMAN, AND S. SEURET
M·n = C/n,
for some constant C > 0. This immediately implies that for every j large
enough, the set
i=1 2 2
has a Lebesgue measure less than Cj−2.
(cid:101)A(j) := {a ∈ [0, 1] : G(a) ≤ j−2}
A(cid:48)(j) := {a ∈ [0, 1] :(cid:12)(cid:12)G(2ja)(cid:12)(cid:12) ≤ j−2}
is also equal Cj−2 (the same as that of (cid:101)A(j)). Obviously,
of the set
Remarking the 1-periodicity of G, we deduce that the Lebesgue measure
L(A(cid:48)(j)) < +∞.
Thus, applying the Borel-Cantelli lemma to the sets A(cid:48)(j), we deduce that
the limsup set
(cid:88)
(cid:92)
j≥1
(cid:91)
A(cid:48)(j)
J≥1
j≥J
has zero Lebesgue measure. This set is the complement of the set A(cid:48)
by deduction is of full Lebesgue measure in [0, 1].
3.2. Prevalence property of an ancillary set.
1, which
(cid:3)
The key result to obtain the prevalence of the singularity spectrum of
Theorem 2.1 is the following theorem.
Theorem 3.2. Suppose that 0 < s − D/p < ∞ and 0 < q ≤ ∞. Let α ≥ 1
and let us defined the exponent
H(α) := s − d
p
+
d
αp
.
(22)
The set
Fα :=
(cid:110)
f ∈ Bs
(cid:111)
p,q([0, 1]D) : ∃A(f ) of full Lebesgue measure such that
a ∈ A(f ) =⇒ ∀x ∈ X α, hfa(x) ≤ H(α)
is prevalent in Bs
p,q([0, 1]D).
The proof of Theorem 3.2 is postponed to § 4. We admit it for the moment,
and we explain how we conclude once Theorem 3.2 is proved. Our main
result, Theorem 2.1, is a direct consequence of Propositions 3.4 -- 3.6 below.
From now on, let (αn)n∈N be a dense sequence in [1,∞) such that α0 = 1.
Using the fact that a countable intersection of prevalent (resp. full Lebesgue
measure) sets is prevalent (resp. of full Lebesgue measure), it follows imme-
diately that:
Corollary 3.3. The set
F :=
p,q([0, 1]D) : ∃A(f ) of full Lebesgue measure such that
(cid:111)
a ∈ A(f ) ⇒ ∀n ∈ N, ∀x ∈ X αn, hfa(x) ≤ H(αn)
f ∈ Bs
(cid:110)
is prevalent in Bs
p,q([0, 1]D).
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
13
3.3. Prevalent upper bound.
We first find an upper bound for the singularity spectrum of Lebesgue
p,q([0, 1]D).
p,q([0, 1]D), for almost all a ∈ [0, 1]d(cid:48)
almost traces of f, for every f ∈ Bs
Proposition 3.4. For every f ∈ Bs
have
(23)
Proof. Let f ∈ Bs
p,q(RD). By Theorem 1.1, there is a set A(f ) of full
Lebesgue measure in [0, 1]d(cid:48) such that for every a ∈ A(f ), the trace fa
p,∞(Rd). Then, by Theorem 1.2, for every h ≥ s − d/p,
Bs−ε
belongs to (cid:84)
dfa(h) ≤ min(d, d + (h − s)p).
for every h ≥ s − d/p,
, we
s−ε<s
for every ε > 0,
dfa(h) ≤ min(d, d + (h − s)p + εp).
p,q([0, 1]D),
Moreover, for every ε > 0, since fa ∈ Bs−ε
no point x ∈ [0, 1]d such that hfa(x) < s − d/p − ε.
Letting ε > 0 yields exactly the upper bound (23).
One can obtain more precise informations for almost all f ∈ Bs
p,∞(Rd) for every a ∈ A(f ), there is
(cid:3)
p,q([0, 1]D), for Lebesgue-almost all
i.e. on a prevalent set in f ∈ Bs
p,q([0, 1]D).
Proposition 3.5. For almost all f ∈ Bs
a ∈ [0, 1]d(cid:48)
, for all x ∈ [0, 1]d, hfa(x) ≤ s.
Proof. We apply Corollary 3.3 with αn = α0 = 1:
if f belongs to the
prevalent set F, then for any a ∈ A(f ), for any x ∈ X α0 = X 1 = [0, 1]d,
hfa(x) ≤ H(α0) = s.
(cid:3)
3.4. Prevalent lower bound.
p,q([0, 1]D), for almost all a ∈ [0, 1]d(cid:48)
Proposition 3.6. For almost all f ∈ Bs
,
for any h ∈ [s − d/p, s], dfa(h) ≥ d + (h − s)p and furthermore, Efa(s) has
full Lebesgue measure.
Proof. Consider a function f in the prevalent set F. Let h ∈ (s − d/p, s].
This exponent can be written
(24)
for some given α ≥ 1.
h = H(α) = s − d
p
+
d
αp
X α ⊂(cid:84)
Consider a subsequence (αφ(n))n∈N of (αn)n∈N which is nondecreasing and
converges to α (for α = 1 this would just be φ = 0).
Let us first assume that α > 1, i.e. H(α) ∈ (s − d/p, s). Remark that
n≥1 X αφ(n). Since f ∈ F, it follows that for all a ∈ A(f ) and x ∈ X α,
hfa(x) ≤ H(α). Hence X α ⊂ {x : hfa(x) ≤ H(α)}.
Recall that Theorem 2.5 provides us with a measure mα which is supported
by X α, and which gives measure 0 to every set of dimension strictly less than
d/α.
14
Let us introduce the set Y α :=(cid:8)x : hfa(x) < H(α)(cid:9). Clearly,
J.-M. AUBRY, D. MAMAN, AND S. SEURET
Y α =
{x : hfa(x) ≤ H(α) − 1/n}.
(cid:91)
n≥1
By (23), each set {x : hfa(x) ≤ H(α) − 1/n} has Hausdorff dimension
strictly less than d/α. The scaling properties and the σ-additivity of the
measure mα yield that mα(Y α) = 0.
Remembering that mα(X α) > 0, we have mα(X α\Y α) > 0. This means
equivalently that mα({x ∈ X α : hfa(x) = H(α)}) > 0. This implies that the
set {x ∈ X α : hfa(x) = H(α)} has Hausdorff dimension greater than d/α,
and thus
dfa(h) = dfa(H(α)) = dimH{x : hfa(x) = H(α)} ≥ d/α = p(h − s) + d,
the last equality following from (24).
When α = 1, the same reasoning using the d -dimensional Lebesgue mea-
sure Ld instead of mα yields Efa(s) ⊃ [0, 1]d\Y 1 with Ld(Y 1) = 0. Hence
Ld(Efa(s)) = 1.
α≥1 X α = (cid:84)
x of X ∞ := (cid:84)
Finally, it remains us to treat the case of the smallest exponent h =
s − d/p. Remembering the definition of F, observe that at any element
n≥1 X αn, one necessarily has hf (x) ≤ s − d/p.
Since the converse inequality holds true for any x, we have proved that
X ∞ ⊂ Efa(s− d/p). We conclude by noting that X ∞ is certainly not empty
(cid:3)
(and uncountable), since it is a dense Gδ set in Rd.
Theorem 2.1 is now proved, provided that we can establish Theorem 3.2.
4. Proof of Theorem 3.2: prevalence of Fα
We simplify the problem by including the complement of Fα in a countable
union of simpler ancillary sets. Let N be an integer, α > 1, γ > H(α) and
Oγ,N :=
f ∈ Bs
where
Aγ,N (f ) =
(cid:26)
a ∈ [0, 1]d(cid:48)
:
(cid:110)
(cid:111)
p,q([0, 1]D) : Ld(cid:48)(Aγ,N (f )) > 0
cλ(a) ≤ N 2−γj(1 +(cid:12)(cid:12)2jx − k(cid:12)(cid:12))γ
∃x ∈ X α, ∀λ = (j, k, l) ∈ Λd × Ld,
(cid:27)
.
(cid:110)
definition of Aγ,N (f ) implies that fa has exponent greater than γ at x.
Remark that the conditions on the wavelet coefficients that appear in the
Recall the definition of Fα
Fα :=
(cid:111)
p,q([0, 1]D) : ∃A(f ) of full Lebesgue measure such that
a ∈ A(f ) =⇒ ∀x ∈ X α, hfa(x) ≤ H(α)
f ∈ Bs
.
Proposition 4.1. For any sequence (γn)n∈N strictly decreasing to H(α), we
have
p,q([0, 1]D)\Fα ⊂ (cid:91)
Bs
Oγn,N
n,N∈N
15
(cid:27)
.
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
(cid:26)
Proof. We write that Fα =(cid:84)
(cid:110)
f ∈ Bs
n≥1 Fα,γn, where for any γ > H(α) we put
∃Aγ(f ) of full Lebesgue measure such that
(cid:111)
a ∈ [0, 1]d(cid:48)
: ∃x ∈ X α, fa ∈ Cγ(x)
a ∈ Aγ(f ) ⇒ ∀x ∈ X α, fa /∈ Cγ(x)
Fα,γ =
p,q([0, 1]D) :
When f (cid:54)∈ Fα,γ, the set
has positive
Lebesgue measure. But by (11) of Proposition 2.3 which gives the charac-
terization of Cγ(x) in terms of wavelet coefficients, this last set is included
in Aγ,N (f ), for some N ≥ 1. Hence Bs
N∈N∗ Oγ,N and
(cid:3)
the conclusion follows.
To prove Theorem 3.2, it suffices now to show that each set Oγ,N is uni-
From now on we fix N ∈ N∗, α > 1 and γ > H(α).
p,q([0, 1]D)\Fα,γ ⊂(cid:83)
versally measurable (Proposition 4.2) and shy (Proposition 4.7).
4.1. Measurability.
First we deal here only with the case q < ∞, that is when Bs
Polish space. The case q = ∞ is proved in Appendix B, Proposition B.1.
Proposition 4.2. The set Oγ,N is universally measurable in Bs
p,q([0, 1]D).
Proof. Let
Φλ(f, a, x) := N 2−γj(1 +(cid:12)(cid:12)2jx − k(cid:12)(cid:12))γ − cλ(a)
p,q(RD) is a
Φ(f, a, x) := inf
Φλ(f, a, x)
λ∈Λd×Ld
and
(cid:101)Φ(f ) := Ld(cid:48)
: ∃x ∈ X α, Φ(f, a, x) ≥ 0
(cid:111)(cid:17)
a ∈ [0, 1]d(cid:48)
(cid:16)(cid:110)
Oγ,N =(cid:101)Φ−1(cid:0)(0, +∞)(cid:1).
so that
To obtain Proposition 4.2, we just have to prove that (cid:101)Φ is universally mea-
p,q([0, 1]D) −→ R+. For this, let us fix a complete Borel
surable as a map : Bs
measure µ on Bs
p,q([0, 1]D).
First observe that each Φλ is continuous on Bs
p,q([0, 1]D)× [0, 1]d(cid:48) × [0, 1]d.
This follows from the fact that the dependence of the wavelet coefficients
λ on f is continuous (the Besov topology induced on the space of wavelet
cD
coefficients by (9) is stronger than the product topology) and the dependence
of cλ(a) on the variables a and cD
λ is also continuous (from their definitions
(15), (16) and (19)). As a countable infimum of continuous functions, Φ is
Borel on the Polish space Bs
p,q([0, 1]D) × [0, 1]d(cid:48) × [0, 1]d.
p,q([0, 1]D) × [0, 1]d(cid:48) × X α(cid:17)
Clearly X α ∈ B([0, 1]d), so the set
Bs
(cid:101)T := Φ−1([0,∞)) ∩(cid:16)
(cid:110)
Π((cid:101)T ) :=
is also Borel and its projection along the third coordinate
: ∃x ∈ X α, (f, a, x) ∈ (cid:101)T(cid:111)
the universal measurability Theorem 2.6, Π((cid:101)T ) is then µ ⊗ Ld(cid:48)-measurable.
p,q([0, 1]D) × [0, 1]d(cid:48)
p,q([0, 1]D)× [0, 1]d(cid:48)
p,q([0, 1]D)× [0, 1]d(cid:48)
is analytic in the space (Bs
(f, a) ∈ Bs
,B(Bs
)). By
16
J.-M. AUBRY, D. MAMAN, AND S. SEURET
To conclude, we notice that (cid:101)Φ can be written as
(cid:90)
[0,1]d(cid:48) 1Π((cid:101)T )(f, a) da.
Since Π((cid:101)T ) is µ ⊗ Ld(cid:48)-measurable, we can apply Fubini's theorem, so that
we conclude that (cid:101)Φ is µ-measurable, for any complete Borel measure µ on
(cid:101)Φ : f (cid:55)→
(cid:3)
Bs
p,q([0, 1]D).
4.2. Probe space.
In this section q ∈ (0,∞), with the obvious modifications when q = ∞.
We use the following notation: to each (j, k) ∈ Λd we associate the unique
(J, K), J ∈ N and K ∈ Zd
J satisfying K2−J = k2−j (K2−J is the
irreducible version of the dyadic point k2−j). Obviously, with the preceding
notations, J ≤ j.
Proposition 4.3. Let us define, for every λD = (j, (k, k(cid:48)), (l, l(cid:48))) ∈ ΛD
J\2Zd
(cid:40)
The function g := (cid:80)
eλD :=
(25)
j
0
− q+2
qp 2( d
p−s)j2
− d
p J
if
if
,
l (cid:54)= 0d and l(cid:48) = 1d(cid:48)
l = 0d or l(cid:48) (cid:54)= 1d(cid:48)
.
p,q([0, 1]D).
eλD ψλD belongs to Bs
λD∈ΛD
(cid:88)
Proof. Observe that eλD does not depend on l ∈ Ld. Using the wavelet
characterization (9) of Bs
p,q([0, 1]D), the proof boils down to studying for all
integers j ≥ 1 the quantity
Aj := 2j(sp−D)
eλDp.
λD=(j,(k,k(cid:48)),(l,l(cid:48))): (k,k(cid:48))∈ZD
j , (l,l(cid:48))∈{0,1}D
By construction, eλD is the same for all k(cid:48) and all l, and equals zero except
when l(cid:48) = 1d(cid:48) and l (cid:54)= 0d. Thus,
Aj = (2d − 1)2j(sp−D+d(cid:48)) (cid:88)
eλDp ≤ 2d+j(sp−d) (cid:88)
−p q+2
qp 2(d−sp)j2−dJ
j
k∈Zd
j
k∈Zd
j
≤ 2dj
− q+2
2−dJ .
q (cid:88)
k∈Zd
j
where one should not forget that J depends on k. For a given integer 1 ≤
J ≤ j, the number of multi-integers k ∈ Zd
j such that its irreducible version
can be written K2−J (for some K) is exactly 2d(J−1). Hence
Aj ≤ 2dj
− q+2
q
2d(J−1)−dJ = j
− 2
q
j(cid:88)
J=1
which is an lq sequence.
Remark 4.4. Although we did not prove it here, the singularity spectrum of g
(and of the functions g(i) below) can be explicitly computed: for every h ∈ [s−
d/p, s], dg(h) = ph−ps+D, and dg(h) = −∞ else. It is noticeable that g does
p,q([0, 1]D) (the generic spectrum has the
not enjoy the generic spectrum in Bs
(cid:3)
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
17
same formula but the range of h is [s−D/p, s], not [s−d/p, s]). Nevertheless
its traces will be shown to have the typical spectrum in Bs
p,q([0, 1]d).
Let J0 ≥ 1 to be fixed later and d1 := 2dJ0. For each d-dimensional dyadic
cube λ ∈ Λd at scale j, we enumerate in an arbitrary fashion λ(1), . . . , λ(d1)
its d1 sub-cubes at scale j + J0.
Definition 4. We set the probe space P to be the d1-dimensional subspace
p,q([0, 1]D) spanned by the functions g(i), whose wavelet coefficients e(i)
of Bs
λD
are defined in the following way: for each λD = (j, (k, k(cid:48)), (l, l(cid:48))) ∈ ΛD, let
λ := (j, k, l) and
if λ =(cid:101)λ(i) for some(cid:101)λD := (j − J0, (k, k(cid:48)), (l,l(cid:48)))
(cid:40)
e(cid:101)λD
else.
In the definition above, (cid:101)λ(i) is the sub-cube associated with (cid:101)λ = (j −
J0, k,l) (which is the restriction to [0, 1]d of(cid:101)λD).
e(i)
λD =
(26)
0
In particular, recalling (25), as soon as l(cid:48) (cid:54)= 1d(cid:48), e(i)
λD = 0, and this coeffi-
j . By the same proof as Proposition 4.3, each
cient is the same for all k(cid:48) ∈ Zd(cid:48)
p,q([0, 1]D).
g(i) also belongs to Bs
Heuristically, the wavelet coefficients of g at generation j are dispatched in
wavelet coefficients at generation j +J0 for the functions g(i), the distribution
being organized so that for any cube λD, there is only one g(i) such that
λD (cid:54)= 0.
e(i)
Let us now consider their traces g(i)
a on the affine subspace Ha.
Lemma 4.5. For every i ∈ {1,··· , d1}, for every j ≥ 1, for every λ =
(j, k, l) with k ∈ Zd
j and l (cid:54)= 0d, we have the formula
Gd(cid:48)(2ja)
e(i)
λ (a) = e(i)
),(1d,1d(cid:48)
(j,(k,1d(cid:48)
))
(27)
where Gd(cid:48) was defined in (21).
Moreover, if l = 0d, then e(i)
λ (a) = 0.
Proof. Following (15) and (16), the wavelet coefficients of these traces are:
for all j ≥ 1, for all λ = (j, k, l) ∈ {j} × Zd
Ψl(cid:48)
j,k(cid:48)
i
i
(ai).
(28)
if l = 0d,
e(i)
λ (a) :=
(29)
if l ∈ Ld,
e(i)
λ (a) :=
j × {0, 1}d,
(cid:88)
(cid:88)
j , l(cid:48)∈Ld(cid:48)
e(i)
λD
d(cid:48)(cid:89)
d(cid:48)(cid:89)
i=1
λD=(j,(k,k(cid:48)),(0d,l(cid:48))):
k(cid:48)∈Zd(cid:48)
λD=(j,(k,k(cid:48)),(l,l(cid:48))):
k(cid:48)∈Zd(cid:48)
j , l(cid:48)∈{0,1}d(cid:48)
Ψl(cid:48)
j,k(cid:48)
i
(ai).
i
e(i)
λD
i=1
By definition of e(i)
by construction e(i)
λD, the coefficients (28) are all zero. Now, remember that
λD does not depend on k(cid:48), nor on l ∈ Ld, and that they all
18
J.-M. AUBRY, D. MAMAN, AND S. SEURET
have the same values as one of them, say the one with l = 1d. Thus, as soon
as l (cid:54)= 0d, formula (29) can be simplified into
e(i)
λ (a) = e(i)
(j,k,1d)(a) =
= e(i)
(j,(k,1d(cid:48)
),(1d,1d(cid:48)
))
(cid:88)
d(cid:48)(cid:89)
λD=(j,(k,k(cid:48)),(1d,1d(cid:48)
(cid:88)
e(i)
λD
): k(cid:48)∈Zd(cid:48)
j
Ψ1(2jai − k(cid:48)
i).
d(cid:48)(cid:89)
i=1
Ψ1
j,k(cid:48)
i
(ai)
Since Ψ1 has compact support, for a given a ∈ (0, 1)d(cid:48), when j is large
k(cid:48)∈Zd(cid:48)
j
i=1
(cid:88)
d(cid:48)(cid:89)
k(cid:48)∈Zd(cid:48)
i=1
enough, we have(cid:88)
d(cid:48)(cid:89)
k(cid:48)∈Zd(cid:48)
j
i=1
This yields (27).
Ψ1(2jai − k(cid:48)
i) =
Ψ1(2jai − k(cid:48)
i) = Gd(cid:48)(2ja).
(cid:3)
Let x ∈ X α not a dyadic element of [0, 1]d, and consider the irreducible
sequence (Jn, Kn) associated to x as in (20), i.e.
x − Kn2−Jn ≤ 2−αJn.
Let a ∈ A1 and let ja be the associated integer constructed in Proposition 3.1.
Let n be such that jn := [αJn] ≥ ja and such that (27) holds. Let us denote
by λn := (jn, kn, l) the unique dyadic node (unique in the sense that l varies
in Ld) such that Kn2−Jn = kn2−jn. With each λn can be associated its
sub-cubes λ(i)
Lemma 4.6. For all 1 ≤ i ≤ d1, λ(i)
width 2J0+2 above x, and
n lies within the cone of influence of
n , i ∈ {1,··· , d1}.
(cid:12)(cid:12)(cid:12)e(i)
λ(i)
n
(cid:12)(cid:12)(cid:12) ≥ Cj
(a)
−(2d(cid:48)+ q+2
qp )
n
2−H(α)jn
(30)
n can be written λ(i)
n = (jn + J0, k(i)
n , l) for some
the constant C depending only on J0.
Proof. Remark that λ(i)
n ∈ Zd
integer k(i)
. By construction,
k(i)
n 2−(jn+J0) − Kn2−Jn = k(i)
jn+J0
Using (20) we deduce that
(cid:12)(cid:12)x − k(i)
n 2−(jn+J0)(cid:12)(cid:12) ≤ (cid:12)(cid:12)x − Kn2−Jn(cid:12)(cid:12) +(cid:12)(cid:12)k(i)
n 2−(jn+J0) − Kn2−Jn(cid:12)(cid:12)
n 2−(jn+J0) − kn2−jn ≤ 2−jn.
≤ 2−αJn + 2−jn ≤ 32−jn ≤ (2J0+2)2−(jn+J0).
This shows the first part of Lemma 4.6.
Recall now Proposition 3.1. The fact that a ∈ A1 guarantees that Gd(cid:48)(2ja) ≥
j−2d(cid:48) as soon as j ≥ ja. Combining this with (27), we get
(cid:12)(cid:12)(cid:12)e(i)
(λn)(i)(a)
(cid:12)(cid:12)(cid:12) ≥ (jn + J0)−2d(cid:48)(cid:12)(cid:12)(cid:12)(cid:12)e(i)
(jn+J0,(k(i)
n ,1d(cid:48)
),(1d,1d(cid:48)
))
(cid:12)(cid:12)(cid:12)(cid:12).
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
19
Remembering now how we chose the coefficients of g(i) in (26), we see that
(cid:12)(cid:12)(cid:12)e(i)
(λn)(i)(a)
(cid:12)(cid:12)(cid:12) ≥ (jn + J0)−2d(cid:48) ·(cid:12)(cid:12)(cid:12)e(jn,(kn,1d(cid:48)
≥ (jn + J0)−2d(cid:48) · (jn)
≥ Cj
−(2d(cid:48)+ q+2
qp )
n
2−H(α)jn
− q+2
qp 2( d
(cid:12)(cid:12)(cid:12)
))
),(1d,1d(cid:48)
p−s)jn− d
p Jn
(cid:3)
where we used that αJn ≤ jn + 1.
4.3. Shyness of Oγ,N.
Recall that γ > H(α).
Take an arbitrary f ∈ Bs
each β ∈ Rd1 define
p,q([0, 1]D) with wavelet coefficients cD
λ , and for
d1(cid:88)
i=1
f β := f +
βig(i).
a will denote its trace at level x(cid:48) = a and cβ
As usual now, f β
coefficients of that trace. Now we choose J0 large enough so that
(31)
Our goal is to prove:
Proposition 4.7. For any f ∈ Bs
has d1-dimensional Lebesgue measure Ld1 equal to 0.
p,q([0, 1]D), the set(cid:8)β ∈ Rd1 : f β ∈ Oγ,N
d − d1(γ − H(α)) < 0
λ(a) the wavelet
(cid:9)
This will show that Oγ,N is shy. Let us quickly explain this fact.
Let us denote by µ the measure Ld1 carried by P. Assume that Proposition
p,q([0, 1]D). For µ-almost F ∈ P, we know
4.7 holds true, and fix any f ∈ Bs
that f + F /∈ Oγ,N. Hence the set {f + Oγ,N} ∩ P has a µ-measure equal to
0, i.e.
µ({f + Oγ,N}) = 0.
(cid:41)
(cid:40)
Since this is true for any f ∈ Bs
shy.
p,q([0, 1]D), by Definition 1, the set Oγ,N is
Ba :=
β ∈ Rd1 :
Before that, two intermediary lemmas are necessary. Let us introduce
(cid:12)(cid:12)(cid:12)cβ
λ(a)
∃ xβ ∈ X α, ∀ λ ∈ Λd,
(cid:12)(cid:12)(cid:12) ≤ N 2−γj(1 +(cid:12)(cid:12)2jxβ − k(cid:12)(cid:12))γ
N 2−γj(1 +(cid:12)(cid:12)2jx − k(cid:12)(cid:12))γ −(cid:12)(cid:12)(cid:12)cβ
Lemma 4.8. The application (a, β) (cid:55)→ 1Ba(β) is Lebesgue-measurable as an
application from [0, 1]d(cid:48) × Rd1 to R.
Proof. Let φ : (a, β, x) (cid:55)→ inf
λ∈Λd
argument similar to the one used in proving Proposition 4.2, φ is Borel on
[0, 1]d(cid:48) × Rd1 × [0, 1]d. Remark then that 1Ba(β) can be written as 1Ba(β) =
1G(a, β), where
G :=
(a, β) ∈ [0, 1]d(cid:48) × Rd1 : ∃x ∈ X α, φ(a, β, x) ≥ 0
(cid:12)(cid:12)(cid:12). By an
(cid:110)
(cid:111)
λ(a)
.
(cid:16)
(cid:40)
20
J.-M. AUBRY, D. MAMAN, AND S. SEURET
This set can be written as
(32)
φ−1([0,∞))
π
(cid:92)(cid:16)
[0, 1]d(cid:48) × Rd1 × X α(cid:17)(cid:17)
,
where π(a, β, x) = (a, β) is the (continuous) canonical projection on the two
first coordinates. Since the set between brackets in (32) is clearly a Borel set,
G is analytic and in particular, by Theorem 2.6, it is Lebesgue-measurable.
(cid:3)
By a Fubini argument, we deduce Lemma 4.8.
Lemma 4.9. For each a ∈ A1, the set Ba has Lebesgue measure 0.
Proof. For any λ0 := (j0, k0, l0) ∈ Λd we put
Ba,λ0 :=
β ∈ Rd1 :
so that
Ba = lim sup
j0→∞
(cid:41)
(cid:12)(cid:12)(cid:12)cβ
(cid:91)
∃ xβ ∈ B(k02−j0, 2−αj0),∀λ ∈ Λd,
λ(a)
(cid:12)(cid:12)(cid:12) ≤ N 2−γj(1 +(cid:12)(cid:12)2jxβ − k(cid:12)(cid:12))γ
(cid:92)
(cid:91)
(cid:91)
Ba,λ0 =
Ba,λ0
k0∈Zd
j0
j≥1
j0≥j
k0∈Zd
j0
Suppose that β and (cid:101)β both belong to some Ba,λ0, where j0 is large enough
We want to show that Ld1(Ba) = 0 by bounding by above each Ld1(Ba,λ0).
so that j1 := (cid:98)αj0(cid:99) ≥ ja (cf. Proposition 3.1).
i ≤ d1,
Applying Lemma 4.6, there exists λ1 = (j1, k1, l1) such that for all 1 ≤
is in the cone of influence of width 2J0+1 above xβ and x(cid:101)β,
(i) λ(i)
1
(cid:12)(cid:12)(cid:12)(cid:12)e(i)
λ(i)
1
(ii)
(a)
From (i) we deduce that
−(qd(cid:48) + q+2
qp )
1
2−H(α)j1,
Recall that γ > H(α). Combining this with (ii), we deduce that
(a)
λ(i)
1
(cid:12)(cid:12)(cid:12)(cid:12)c
(cid:12)(cid:12)(cid:12)(cid:12) =
(cid:12)(cid:12)(cid:12)(cid:12) ≥ Cj
(cid:12)(cid:12)(cid:12)(cid:12)cβ
(cid:12)(cid:12)(cid:12)(cid:12) =
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)c
(cid:12)(cid:12)(cid:12)βi −(cid:101)βi
(a)
λ(i)
1
≤
≤ C2−γj1.
d1(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12) ≤ C
2
2−γj1
(a) +
βie(i)
λ(i)
1
(a)
i=1
λ(i)
1
construction of the functions g(i), we have e(i(cid:48))
λ(i)
1
(C depending on N, J0, γ) and the same for (cid:101)β. On the other hand, by
(cid:12)(cid:12)(cid:12)(cid:12)(βi −(cid:101)βi)ei
(a) = 0 for any i (cid:54)= i(cid:48). Thus
c
λ(i)
1
(a) +
βiei
λ(i)
1
d1(cid:88)
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:101)βiei
d1(cid:88)
i=1
(a)
λ(i)
1
(cid:101)βiei
(a)
λ(i)
1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(a) +
c
λ(i)
1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)c
i=1
(a) +
λ(i)
1
i=1
(a)
−
λ(i)
1
d1(cid:88)
(cid:19)
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) +
(cid:12)(cid:12)(cid:12) ≤ C2−(γ−H(α))j1j
d1(cid:88)
βiei
λ(i)
1
(a)
i=1
(a) +
qd(cid:48) + q+2
1
qp
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
21
hence
Ld1(Ba,λ0) ≤ C2−d1(γ−H(α))j1j
d1(qd(cid:48) + q+2
qp )
1
Summing over all 2dj0 nodes λ0 at scale j0 we conclude that
qp )d1
≤ C2dj0−d1(γ−H(α))j1j
(qd(cid:48) + q+2
1
Ba,λ0
Ld1
(cid:19)
(cid:18) (cid:91)
k0∈Zd
j0
≤ C2j0(d−d1α(γ−H(α)))(αj0)(qd(cid:48) + q+2
qp )d1
whose series converges because of our choice (31) for J0. The Borel-Cantelli
lemma implies that Ld1(Ba) = 0.
(cid:3)
(cid:90)
Proof of Proposition 4.7. We can rewrite the result of Lemma 4.9 as
Applying Fubini's theorem (measurability being guaranteed by Lemma 4.8),
[0,1]d(cid:48)
1Ba(β) dβ da = 0
Rd1
(cid:90)
[0,1]d(cid:48) 1Ba(β) da dβ = 0
Rd1
(cid:90)
(cid:90)
(cid:110)
In other words, for almost all β ∈ Rd1, there exists a set Aβ(f ) of full
Lebesgue measure in [0, 1]d(cid:48)
such that a ∈ Aβ(f ) implies β (cid:54)∈ Ba. This in
turn implies f β (cid:54)∈ Oγ,N and the announced result follows by complementarity
(cid:3)
in Rd1.
Our proof is now complete.
(cid:111)
Appendix A. Direct proof of Theorem 1.1
It suffices to show that for any ε > 0, the set
a ∈ [0, 1]d(cid:48)
p,∞([0, 1]d)
: fa ∈ Bs−ε
. Then we will get the announced result
has full Lebesgue measure in [0, 1]d(cid:48)
by considering a sequence εn decreasing to 0.
Recall that fa = Fa + Ga, where Fa and Ga are defined in (13) and (14).
Recall also that the wavelet coefficients of fa are denoted by dλ(a) in its
λ .
wavelet decomposition (18), while those of f are denoted by cD
First, we are going to apply Proposition 2.4. Recalling the definition (15)
of the wavelet coefficients of Ga, we need to bound by above the sum
(cid:88)
dλ(a)p =
(cid:88)
λ∈Λd
j×0d
j ,l(cid:48)∈Ld(cid:48)
k(cid:48)∈Zd(cid:48)
(cid:12)(cid:12)(cid:12)(cid:12) (cid:88)
(cid:88)
(cid:88)
cλD
Ψl(cid:48)
j,k(cid:48)
i
i
(ai)
cλDp
Ψl(cid:48)
j,k(cid:48)
i
i
(ai)
i=1
d(cid:48)(cid:89)
(cid:12)(cid:12)(cid:12)(cid:12) d(cid:48)(cid:89)
(cid:12)(cid:12)(cid:12)(cid:12) d(cid:48)(cid:89)
i=1
i=1
cλDp
Ψl(cid:48)
j,k(cid:48)
i
i
(ai)
(cid:12)(cid:12)(cid:12)(cid:12)p
(cid:12)(cid:12)(cid:12)(cid:12)p
(cid:12)(cid:12)(cid:12)(cid:12)p
λD∈ΛD
j ×{0d×Ld(cid:48)}
λD∈ΛD
j ×{0d×{0,1}d(cid:48)}
≤ C
≤ C
λ∈Λd
j×0d
(33)
22
J.-M. AUBRY, D. MAMAN, AND S. SEURET
(cid:88)
(cid:88)
for some constant C that depends only on Ψ0 and Ψ1. Indeed, since Ψ0 and
Ψ1 are compactly supported, for each a only a finite number (independent
of j or a) of terms in the second sum are non zero.
Now, using definition (16) for the wavelet coefficients of Fa, we find that
dλ(a)p =
λ∈Λd
j×Ld
λ∈Λd
j×Ld
j ,l(cid:48)∈{0,1}d(cid:48)
(cid:88)
(cid:12)(cid:12)(cid:12)(cid:12) (cid:88)
(cid:88)
k(cid:48)∈Zd(cid:48)
λD∈ΛD
j ×{Ld×{0,1}d(cid:48)}
d(cid:48)(cid:89)
(cid:12)(cid:12)(cid:12)(cid:12) d(cid:48)(cid:89)
i=1
i=1
cλD
Ψl(cid:48)
j,k(cid:48)
i
i
(ai)
cλDp
Ψl(cid:48)
j,k(cid:48)
i
i
(ai)
(cid:12)(cid:12)(cid:12)(cid:12)p
(cid:12)(cid:12)(cid:12)(cid:12)p
(34)
≤ C
again for some constant C that depends only on Ψ0 and Ψ1.
it over a ∈ [0, 1]d(cid:48)
we get
We now consider the sum of all wavelet coefficients of fa, and we integrate
. Using (33) and (34), and recalling that Ld ∪ 0d = {0, 1}d,
(cid:88)
(cid:90)
(cid:12)(cid:12)(cid:12)(cid:12) d(cid:48)(cid:89)
i=1
cλDp
[0,1]d(cid:48)
Ψl(cid:48)
j,k(cid:48)
i
i
(ai)
da.
(cid:12)(cid:12)(cid:12)(cid:12)p
λD∈ΛD
j ×{0,1}D
dλ(a)pda ≤ C
[0,1]d(cid:48)
λ∈Λd
j×{0,1}d
(cid:90)
The attentive reader has noticed that there is no wavelet coefficient associ-
ated with the index lD = 0D for the function f, hence the sum of λD over
j × {0, 1}D is the same as the sum of λD over ΛD
ΛD
Since the functions Ψl(cid:48)
j,k(cid:48)
(cid:88)
(cid:88)
support width ≤ K2−j,
are bounded uniformly in j, k(cid:48) and l(cid:48), and has
j × LD.
(cid:90)
i
i
dλ(a)pda ≤ C
cλDp2−jd(cid:48)
.
[0,1]d(cid:48)
λ∈Λd
j×{0,1}d
j ×LD
Using the definition of Besov norm (9), we see that
dλ(a)pda ≤ C(cid:107)f(cid:107)Bs
(cid:88)
λD∈ΛD
(cid:90)
[0,1]d(cid:48)
λ∈Λd
j×{0,1}d
p,∞2(d−sp)j
for some other constant C that still depends only on Ψ.
(cid:26)
Let us now define
a ∈ [0, 1]d(cid:48)
Aj :=
:
(cid:88)
λ∈Λd
j×{0,1}d
dλ(a)p > C(cid:107)f(cid:107)Bs
p,∞([0,1]D)2−((s−ε)p−d)j
(cid:27)
By Markov's inequality, it follows that
p,∞([0,1]D)2(d−sp)j
Ld(cid:48)(Aj) ≤ C(cid:107)f(cid:107)Bs
C(cid:107)f(cid:107)Bs
p,∞([0,1]D)2−((s−ε)p−d)j
= 2−jεp
thus, applying the Borel-Cantelli lemma, Ld(cid:48)(lim supj Aj) = 0.
Finally, by construction, for any a ∈ [0, 1]d(cid:48)\ lim supj Aj, there exists j0
such that j ≥ j0 implies(cid:88)
dλ(a)p2(sp−d)j ≤ C(cid:107)f(cid:107)Bs
p,∞([0,1]D),
λ∈Λd
j×{0,1}d
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
hence fa ∈ Bs
p,∞([0, 1]d).
23
(cid:3)
Appendix B. Proof of Proposition 4.2, case q = ∞
The only serious difference is due to the fact that Bs
p,∞([0, 1]D) is no longer
separable, so the argument for universal measurability of the ancillary set
(cid:111)
p,∞([0, 1]D) : Ld(cid:48)(Aγ,N (f )) > 0
Oγ,N :=
f ∈ Bs
(cid:110)
has to use a different definition of analyticity.
B.1. Analytic sets in non-Polish spaces.
Analytic sets were previously defined in Polish spaces as continuous im-
ages of Borel sets: this cannot apply to Bs
p,∞([0, 1]D). However we can use
the following more general definition, adapted from [3], for any Hausdorff
topological space X endowed with its Borel σ-algebra B(X). First, for a
compact topological space K we write K the collection of its closed subsets
and (B(X)×K)σδ the collection of countable intersection of countable unions
of sets that are Cartesian products of a Borel set in X and a closed set in K
and π : X × K → X is the canonical projection map π(x, y) = x.
Definition 5. A set A ⊂ X is said to be analytic if there exists a compact
space K and S ∈ (B(X) × K)σδ such that
A = π(S).
It is easy to check that this definition coincides with the previous one
when X is Polish. Furthermore, in this framework, Theorem 2.6 (based on
Choquet's capacitability theorem) now holds in any Hausdorff topological
space (see [3]).
B.2. Measurability.
Proposition B.1. The set Oγ,N is universally measurable in Bs
Proof. We use the same notations Φλ, Φ, (cid:101)T and Π as in the proof of Propo-
sition 4.2. We will show that the set Π((cid:101)T ) is analytic in the sense of Defini-
p,∞([0, 1]D).
tion 5, with X := Bs
put ∆j := Zd(cid:48)
j × {0, 1}d(cid:48). Given n ∈ N, m ∈ Z∆j and m(cid:48) ∈ Zd(cid:48) we write
and K := [0, 1]d. For short let us
p,∞([0, 1]D) × [0, 1]d(cid:48)
Q(n, m) := 2−n (cid:89)
Q(cid:48)(n, m(cid:48)) := 2−n (cid:89)
(k(cid:48),l(cid:48))∈∆j
i, m(cid:48)
Q(n, m, m(cid:48)) := Q(n, m) × Q(cid:48)(n, m(cid:48))
1≤i≤d(cid:48)
[m(cid:48)
i + 1]
[mk(cid:48),l(cid:48), mk(cid:48),l(cid:48) + 1]
Having fixed λ = (j, k, l) ∈ Λd×Ld and considering λD = (j, (k, k(cid:48)), (l, l(cid:48))),
p,∞([0, 1]D) induces a map sλ : (k(cid:48), l(cid:48)) (cid:55)→ cλD that we identify to
any f ∈ Bs
an element of R∆j. Then we define
Θ(sλ, a, x) := N 2−γj(1 + 2jx − k)γ − cλ(a)
J.-M. AUBRY, D. MAMAN, AND S. SEURET
24
as well as
(cid:26)
(cid:27)
Xλ(n, m, m(cid:48)) :=
x ∈ [0, 1]d :
sup
(sλ,a)∈Q(n,m,m(cid:48))
Θ(sλ, a, x) ≥ 0
(cid:92)
(cid:110)
(cid:91)
The dependency of cλ(a) on sλ and a is given in (15), (16) and (19) and it
is continuous. So is the function Θ. Since Q(n, m, m(cid:48)) is compact, it follows
that Xλ(n, m, m(cid:48)) is closed (Lemma B.2). Furthermore, if we put
(cid:111)
p,∞([0, 1]D) : sλ ∈ Q(n, m)
Fλ(n, m) :=
f ∈ Bs
then it is clear by continuity of Φλ that
j
(m,m(cid:48))∈Z∆j×Zd(cid:48)
Fλ(m, n) × Q(cid:48)(m(cid:48), n) × Xλ(m, m(cid:48), n)
Φ−1
λ ([0,∞)) =
n∈N
(cid:84)
λ ([0,∞)) and the set (cid:101)T belong to (B(X) × K)σδ as well because
This proves that Φ−1
λ ([0,∞)) ∈ (B(X)×K)σδ. We deduce that Φ−1([0,∞)) =
p,∞([0, 1]D)×[0, 1]d(cid:48)×X α is obviously in (B(X)×K)σδ. Its projection Π((cid:101)T )
λ Φ−1
Bs
is thus analytic and we conclude in the same way as for Proposition 4.2. (cid:3)
Lemma B.2. Let A and B be topological spaces, A compact and B locally
compact. If f is continuous: A × B → R, then fs : b (cid:55)→ supa∈A f (a, b) is
continuous on B.
Proof. Recall that a Hausdorff space-valued function defined on a compact
set is continuous if and only if its graph is compact. Continuity being a local
property, we can suppose without loss of generality that B is also compact.
The graph Γ of f is then compact and so is its image by the projection
: (a, b, y) (cid:55)→ (b, y). As a supremum of continuous functions, fs is lower
semi-continuous, so its epigraph E is closed. But the graph of fs is precisely
E ∩ (Γ), so it is compact; it follows that fs is continuous.
(cid:3)
Acknowledgments. The authors are thankful to Basarab Matei for his
simulations and pictures of Daubechies' wavelets.
References
[1] Aubry, J.-M., Bastin, F., and Dispa, S. Prevalence of multifractal functions in
Sν spaces. J. Fourier Anal. Appl. 13, 2 (2007), 175 -- 185.
[2] Buczolich, Z. and Seuret, S. Typical measures satisfy a multifractal formalism.
Preprint (2009).
[3] Choquet, G. Theory of capacities. Ann. Inst. Fourier (Grenoble) 5 (1954), 131 -- 295.
[4] Christensen, J. P. R. On sets of Haar measure zero in Abelian Polish groups.
Israel J. Math. 13 (1972), 255 -- 260.
[5] Daubechies, I. Ten Lectures on Wavelets, vol. 61 of CBMS-NSF regional conference
series in applied mathematics. SIAM, 1992.
[6] Falconer, K. J., Fractal Geometry, John Wiley & Sons, (1990).
[7] Fraysse, A., and Jaffard, S. How smooth is almost every function in a Sobolev
space? Rev. Mat. Iberoamericana 22, 2 (2006), 663 -- 683.
[8] Frisch, U. and Parisi, G., Fully developped turbulence and intermittency, Proc.
International Summer school Phys., Enrico Fermi, 84-88, North Holland, 1985
[9] Hunt, B. R., Sauer, T., and Yorke, J. A. Prevalence: a translation-invariant
"almost every" on infinite-dimensional spaces. Bull. Amer. Math. Soc. (N.S.) 27, 2
(Oct. 1992), 217 -- 238.
LOCAL BEHAVIOR OF TRACES OF BESOV FUNCTIONS
25
[10] Jaffard, S. Théorèmes de trace et dimensions négatives . C. R. Acad. Sci. Paris
Sér. I 320 (1995), 409 -- 413.
[11] Jaffard, S. Multifractal formalism for functions Part I : results valid for all func-
tions. SIAM J. Math. Anal. 28, 4 (July 1997), 944 -- 970.
[12] Jaffard, S. and Meyer, Y. On the pointwise regularity of functions in critical
Besov spaces. J. Funct. Anal. 175 (2000), 415 -- 434.
[13] Jaffard, S. Wavelet tools in multifractal analysis Fractal Geometry and Applica-
tions: A Jubilee of Benoit Mandelbrot, Proc. Symp. Pure Math. (AMS) Vol. 72 Part
2, pp. 91-151 (2004).
[14] Järvenpää, E., Järvenpää, M. and Llorente M., Local dimensions of sliced
measures and stability of packing dimensions of sections of sets, Adv. Maths (2004)
183(1) 127-154 .
[15] Mattila, P. Hausdorff dimension, orthogonal projections and intersections with
planes. Ann. Acad. Sci. Fenn. Ser. A I Math. 1 (1975), 227 -- 244.
Laboratoire d'Analyse et Mathématiques Appliquées (CNRS UMR 8050),
Université Paris-Est- Créteil - Val-de-Marne, 61 av. du Général de Gaulle,
94010 Créteil, France
E-mail address: [email protected]
E-mail address: [email protected]
E-mail address: [email protected]
|
1201.5996 | 1 | 1201 | 2012-01-28T20:54:40 | Uniform Algebras Over Complete Valued Fields | [
"math.FA"
] | UNIFORM algebras have been extensively investigated because of their importance in the theory of uniform approximation and as examples of complex Banach algebras. An interesting question is whether analogous algebras exist when a complete valued field other than the complex numbers is used as the underlying field of the algebra. In the Archimedean setting, this generalisation is given by the theory of real function algebras introduced by S. H. Kulkarni and B. V. Limaye in the 1980s. This thesis establishes a broader theory accommodating any complete valued field as the underlying field by involving Galois automorphisms and using non-Archimedean analysis. The approach taken keeps close to the original definitions from the Archimedean setting. Basic function algebras are defined and generalise real function algebras to all complete valued fields. Several examples are provided. Each basic function algebra is shown to have a lattice of basic extensions related to the field structure. In the non-Archimedean setting it is shown that certain basic function algebras have residue algebras that are also basic function algebras. A representation theorem is established. Commutative unital Banach F-algebras with square preserving norm and finite basic dimension are shown to be isometrically F-isomorphic to some subalgebra of a Basic function algebra. The theory of non-commutative real function algebras was established by K. Jarosz in 2008. The possibility of their generalisation to the non-Archimedean setting is established in this thesis. In the context of complex uniform algebras, a new proof is given using transfinite induction of the Feinstein-Heath Swiss cheese "Classicalisation" theorem. | math.FA | math | Uniform Algebras Over Complete
Valued Fields
Jonathan W. Mason, MMath.
Thesis submitted to The University of Nottingham
for the degree of Doctor of Philosophy
March 2012
2
1
0
2
n
a
J
8
2
]
.
A
F
h
t
a
m
[
1
v
6
9
9
5
.
1
0
2
1
:
v
i
X
r
a
For Olesya
Follow the Romany patteran
West to the sinking sun,
Till the junk-sails lift through the houseless drift.
And the east and west are one.1
1From Rudyard Kipling's poem The Gipsy Trail.
i
Abstract
UNIFORM algebras have been extensively investigated because of their importance in
the theory of uniform approximation and as examples of complex Banach algebras. An
interesting question is whether analogous algebras exist when a complete valued field
other than the complex numbers is used as the underlying field of the algebra. In the
Archimedean setting, this generalisation is given by the theory of real function alge-
bras introduced by S. H. Kulkarni and B. V. Limaye in the 1980s. This thesis establishes
a broader theory accommodating any complete valued field as the underlying field by
involving Galois automorphisms and using non-Archimedean analysis. The approach
taken keeps close to the original definitions from the Archimedean setting.
Basic function algebras are defined and generalise real function algebras to all complete
valued fields whilst retaining the obligatory properties of uniform algebras.
Several examples are provided. A basic function algebra is constructed in the non-
Archimedean setting on a p-adic ball such that the only globally analytic elements of
the algebra are constants.
Each basic function algebra is shown to have a lattice of basic extensions related to the
field structure. In the non-Archimedean setting it is shown that certain basic function
algebras have residue algebras that are also basic function algebras.
A representation theorem is established. Commutative unital Banach F-algebras with
square preserving norm and finite basic dimension are shown to be isometrically F-
isomorphic to some subalgebra of a Basic function algebra. The condition of finite
basic dimension is always satisfied in the Archimedean setting by the Gel'fand-Mazur
Theorem. The spectrum of an element is considered.
The theory of non-commutative real function algebras was established by K. Jarosz in
2008. The possibility of their generalisation to the non-Archimedean setting is estab-
lished in this thesis and also appeared in a paper by J. W. Mason in 2011.
In the context of complex uniform algebras, a new proof is given using transfinite
induction of the Feinstein-Heath Swiss cheese "Classicalisation" theorem. This new
proof also appeared in a paper by J. W. Mason in 2010.
ii
Acknowledgements
I would particular like to thank my supervisor J. F. Feinstein for his guidance and en-
thusiasm over the last four years. Through his expert knowledge of Banach algebra
theory he has helped me to identify several productive lines of research whilst always
allowing me the freedom required to make the work my own.
In addition to my supervisor, I. B. Fesenko also positively influenced the direction of
my research. During my doctoral training I undertook a postgraduate training module
on the theory of local fields given by I. B. Fesenko. With extra reading, this enabled
me to work both in the Archimedean and non-Archimedean settings as implicitly sug-
gested by my thesis title.
It was a pleasure to know my friends in the algebra and analysis group at Nottingham
and I thank them for their interest in my work and hospitality.
I am grateful to the School of Mathematical sciences at the University of Nottingham
for providing funds in support of my conference participation and visits.
Similarly I appreciate the support given to me by the EPSRC through a Doctoral Train-
ing Grant.
This PhD thesis was examined by A. G. O'Farrell and J. Zacharias who I thank for their
time and interest in my work.
iii
Contents
1
Introduction
1.1 Background and Overview . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Summary .
.
.
.
.
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Complete valued fields
2.1
Introduction .
.
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1
Series expansions of elements of valued fields . . . . . . . . . . .
2.1.2 Examples of complete valued fields . . . . . . . . . . . . . . . . .
2.1.3 Topological properties of complete valued fields . . . . . . . . . .
2.2 Extending complete valued fields . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Extensions .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2 Galois theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3 Functions and algebras
3.1 Functional analysis over complete valued fields
. . . . . . . . . . . . . .
3.1.1 Analytic functions
. . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Banach F-algebras .
. . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
3.2.1
Spectrum of an element
. . . . . . . . . . . . . . . . . . . . . . . .
4 Uniform algebras
4.1 Complex uniform algebras . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1
Swiss cheese sets in the complex plane
. . . . . . . . . . . . . . .
4.1.2 Classicalisation theorem . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Non-complex analogs of uniform algebras . . . . . . . . . . . . . . . . . .
1
1
3
5
5
7
9
12
14
15
20
22
22
24
28
29
34
34
36
39
48
iv
CONTENTS
4.2.1 Real function algebras . . . . . . . . . . . . . . . . . . . . . . . . .
50
5 Commutative generalisation over complete valued fields
5.1 Main definitions .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Generalisation theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Examples .
.
.
.
.
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Non-Archimedean new basic function algebras from old . . . . . . . . .
5.4.1 Basic extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.2 Residue algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6 Representation theory
6.1 Further Banach rings and Banach F-algebras
. . . . . . . . . . . . . . . .
6.2 Representations .
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Established theorems . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.2 Motivation .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.3 Representation under finite basic dimension . . . . . . . . . . . .
7 Non-commutative generalisation and open questions
7.1 Non-commutative generalisation . . . . . . . . . . . . . . . . . . . . . . .
7.1.1 Non-commutative real function algebras . . . . . . . . . . . . . .
53
53
55
59
65
65
67
79
79
88
88
90
91
98
98
98
7.1.2 Non-commutative non-Archimedean analogs
. . . . . . . . . . . 100
7.2 Open questions .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
References
105
v
CHAPTER 1
Introduction
This short chapter provides an informal overview of the material in this thesis. Justifi-
cation of the statements made in this chapter can therefore be found in the main body
of the thesis which starts at Chapter 2.
1.1 Background and Overview
Complex uniform algebras have been extensively investigated because of their impor-
tance in the theory of uniform approximation and as examples of complex Banach alge-
bras. Let CC(X) denote the complex Banach algebra of all continuous complex-valued
functions defined on a compact Hausdorff space X. A complex uniform algebra A is a
subalgebra of CC(X) that is complete with respect to the sup norm, contains the con-
stant functions making it a unital complex Banach algebra and separates the points of X
in the sense that for all x1, x2 ∈ X with x1 (cid:54)= x2 there is f ∈ A satisfying f (x1) (cid:54)= f (x2).
Attempting to generalise this definition to other complete valued fields simply by re-
placing C with some other complete valued field L produces very limited results. This
is because the various versions of the Stone-Weierstrass theorem restricts our attention
to CL(X) in this case.
However the theory of real function algebras introduced by S. H. Kulkarni and B. V.
Limaye in the 1980s does provide an interesting generalisation of complex uniform
algebras. One important departure in the definition of these algebras from that of com-
plex uniform algebras is that they are real Banach algebras of continuous complex-
valued functions. Similarly the elements of the algebras introduced in this thesis are
also continuous functions that take values in some complete valued field or division
ring extending the field of scalars over which the algebra is a vector space.
A prominent aspect of the emerging theory is that it has a lot to do with representation.
As a very simple example the field of complex numbers itself is isometrically isomor-
phic to a real function algebra, all be it on a two point space.
1
CHAPTER 1: INTRODUCTION
When considering the generalisation of complex uniform algebras over all complete
valued fields I naturally wanted the complex uniform algebras and real function alge-
bras to appear directly as instances of the new theory. This resulted in the definition
of basic function algebras involving the use of a Galois automorphism and homeomor-
phic endofunction that interact in a useful way, see Definition 5.1.2. In retrospect these
particular algebras should more appropriately be referred to as cyclic basic function
algebras since the functions involved take values in some cyclic extension of the un-
derlying field of scalars of the algebra.
Necessarily this thesis starts by surveying complete valued fields and their properties.
The transition from the Archimedean setting to the non-Archimedean setting preserves
in places several of the nice properties that complete Archimedean fields have. How-
ever all complete non-Archimedean fields are totally disconnected, some of them are
not locally compact and there is no non-Archimedean analog of the Gel'fand-Mazur
Theorem.
On the other hand some complete non-Archimedean fields have interesting properties
that only appear in the non-Archimedean setting. Consider for example the closed unit
disc of the complex plane. It is closed under multiplication but not with respect to ad-
dition. In the non-Archimedean setting the closed unit ball OF, of a complete valued
field F, is a ring since in this case the valuation involved observes the strong version of
the triangle inequality, see Definition 2.1.1. The set MF = {a ∈ F : aF < 1} is a max-
imal ideal of OF from which the residue field F = OF/MF is obtained. The residue
field is of great importance in the study of such fields.
Similarly in the non-Archimedean setting we will see that certain basic function al-
gebras have residue algebras that are also basic function algebras. In the process of
proving this result an interesting fact is shown concerning a large class of complete
non-Archimedean fields. For such a field F and every finite extension L of F, extending
F as a valued field, it is shown that for each Galois automorphism g ∈ Gal(L/F) there
exists a set RL,g ⊆ OL of residue class representatives such that the restriction of g to
RL,g is an endofunction, i.e. a self map, on RL,g. This fact is probably known to certain
number theorists.
This thesis also includes several examples of basic function algebras and these are con-
sidered at depth. A new proof of an existing theorem in the setting of complex uniform
algebras is given and theory in the non-commutative setting is also considered.
With respect to commutative Banach algebra theory, Chapter 6 presents an interest-
ing new Gel'fand representation result extending those of the Archimedean setting. In
particular we have the following theorem where the condition of finite basic dimen-
sion is automatically satisfied in the Archimedean setting and compensates for the lack
of a Gel'fand-Mazur Theorem in the non-Archimedean setting. See Chapter 6 for full
2
CHAPTER 1: INTRODUCTION
details.
Theorem 1.1.1. Let F be a locally compact complete valued field with nontrivial valuation. Let
A be a commutative unital Banach F-algebra with (cid:107)a2(cid:107)A = (cid:107)a(cid:107)2
A for all a ∈ A and finite basic
dimension. Then:
(i) if F is the field of complex numbers then A is isometrically F-isomorphic to a complex
uniform algebra on some compact Hausdorff space X;
(ii) if F is the field of real numbers then A is isometrically F-isomorphic to a real function
algebra on some compact Hausdorff space X;
(iii) if F is non-Archimedean then A is isometrically F-isomorphic to a non-Archimedean
analog of the real function algebras on some Stone space X where by a Stone space we
mean a totally disconnected compact Hausdorff space.
In particular A is isometrically F-isomorphic to some subalgebra A of a basic function algebra
and A separates the points of X.
Note that (i) and (ii) of Theorem 1.1.1 are the well known results from the Archimedean
setting. This brings us to the following summary.
1.2 Summary
Chapter 2: The relevant background concerning complete valued fields is provided. Several
examples are given and the topological properties of complete valued fields are
compared and discussed. A particularly useful and well known way of express-
ing the extension of a valuation is considered and the relevant Galois theory is
introduced.
Chapter 3: Some background concerning functional analysis over complete valued fields is
given. Analytic functions are discussed. Banach F-algebras are introduced and
the spectrum of an element is considered.
Chapter 4: Complex uniform algebras are introduced. In the context of complex uniform
algebras, a new proof is given using transfinite induction of the Feinstein-Heath
Swiss cheese "Classicalisation" theorem. This new proof also appeared in a paper
by J. W. Mason in 2010. This is followed by a preliminary discussion concerning
non-complex analogs of uniform algebras. Real function algebras are introduced.
Chapter 5: Basic function algebras are defined providing the required generalisation of real
function algebras to all complete valued fields. A generalisation theorem proves
3
CHAPTER 1: INTRODUCTION
that Basic function algebras have the obligatory properties of uniform algebras.
Several examples are provided. Complex uniform algebras and real function al-
gebras now appear as instances of the new theory. A basic function algebra is
constructed in the non-Archimedean setting on a p-adic ball such that the only
globally analytic elements of the algebra are constants.
Each basic function algebra is shown to have a lattice of basic extensions related
to the field structure. Further, in the non-Archimedean setting it is shown that
certain basic function algebras have residue algebras that are also basic function
algebras. To prove this each Galois automorphism, for certain field extensions, is
shown to restrict to an endofunction on some set of residue class representatives.
Chapter 6: A representation theorem is established in the context of locally compact com-
plete fields with nontrivial valuation. For such a field F, commutative unital
Banach F-algebras with square preserving norm and finite basic dimension are
shown to be isometrically F-isomorphic to some subalgebra of a Basic function
algebra. The condition of finite basic dimension is automatically satisfied in the
Archimedean setting by the Gel'fand-Mazur Theorem.
Chapter 7: The theory of non-commutative real function algebras was established by K.
Jarosz in 2008. The possibility of their generalisation to the non-Archimedean
setting is established in this thesis having been originally pointed out in a pa-
per by J. W. Mason in 2011. The thesis concludes with a list of open questions
highlighting the potential for further interesting developments of this theory.
4
CHAPTER 2
Complete valued fields
In this chapter we survey some of the basic facts and definitions concerning complete
valued fields. Whilst also providing a background, most of the material presented here
is required by later chapters and has been selected accordingly.
2.1 Introduction
We begin with some definitions.
Definition 2.1.1. We adopt the following terminology:
(i) Let F be a field. We will call a multiplicative norm · F : F → R a valuation on F
and F together with · F a valued field.
(ii) Let F be a valued field. If the valuation on F satisfies the strong triangle inequality,
a − bF ≤ max(aF,bF) for all a, b ∈ F,
then we call · F a non-Archimedean valuation and F a non-Archimedean field. Else
we call · F an Archimedean valuation and F an Archimedean field.
(iii) If a valued field is complete with respect to the metric obtained from its valuation
then we call it a complete valued field. Similarly we have complete valuation and
complete non-Archimedean field etc.
(iv) More generally, a metric space (X, d) is called an ultrametric space if the metric d
satisfies the strong triangle inequality,
d(x, z) ≤ max(d(x, y), d(y, z)) for all x, y, z ∈ X.
The following theorem is a characterisation of non-Archimedean fields, courtesy of
[Sch06, p18].
5
CHAPTER 2: COMPLETE VALUED FIELDS
Theorem 2.1.2. Let F be a valued field. Then F is a non-Archimedean field if and only if
2F ≤ 1.
Remark 2.1.3. Whilst it is clear from the definition of the strong triangle inequality
that an Archimedean field can't be extended as a valued field to a non-Archimedean
field, Theorem 2.1.2 also shows that a non-Archimedean field can't be extended to an
Archimedean field.
Theorem 2.1.4. Let F be a valued field. Let C, with pointwise operations, be the ring of Cauchy
sequences of elements of F and let N denote its ideal of null sequences. Then the completion
C/N of F with the function
(an) + NC/N := lim
n→∞ anF,
for (an) + N ∈ C/N, is a complete valued field extending F as a valued field.
Proof. We will only highlight one important part of the proof since further details can
be found in [McC66, p80]. We first note that since the valuation · F is multiplicative
we have a−1F = a−1
for all units a ∈ F×. Let (an) be a Cauchy sequence taking
values in F but not a null sequence. Then there exists δ > 0 and N ∈ N such that for all
n > N we have anF > δ. If (an) takes the value 0 then a null sequence can be added
to (an) such that the resulting sequence (bn) does not takes the value 0 and (bn) agrees
with (an) for all n > N. Hence for all m > N and n > N we have
F
1
δ2bn − bmF
n F = b−1
m Fb−1
n Fbn − bmF <
n ) is also a Cauchy sequence. This shows that the ideal of null
b−1
m − b−1
and so the sequence (b−1
sequences N is maximal and C/N is therefore a field opposed to merely a ring.
Definition 2.1.5. Let F be a valued field. We will call a function ν : F → R ∪ {∞} a
valuation logarithm if and only if for an appropriate fixed r > 1 we have aF = r−ν(a)
for all a ∈ F.
Remark 2.1.6. We have the following basic facts.
(i) With reference to Definition 2.1.1, a valuation logarithm ν on a non-Archimedean
field F has the following properties. For a, b ∈ F we have:
(1) ν(a + b) ≥ min(ν(a), ν(b));
(2) ν(ab) = ν(a) + ν(b);
(3) ν(1) = 0 and ν(a) = ∞ if and only if a = 0.
6
CHAPTER 2: COMPLETE VALUED FIELDS
(ii) Every valued field F has a valuation logarithm since we can take r = e, where
e := exp(1), and for a ∈ F define
(cid:40) − logaF
ν(a) :=
∞
a (cid:54)= 0
a = 0.
if
if
(iii) If ν is a valuation logarithm on a valued field F then so is cν for any c ∈ R
with c > 0. However there will sometimes be a preferred choice. For example a
valuation logarithm of rank 1 is such that ν(F×) = Z.
Lemma 2.1.7. Let F be a non-Archimedean field with valuation logarithm ν. If a, b ∈ F are
such that ν(a) < ν(b) then ν(a + b) = ν(a).
Proof. Given that ν(a) < ν(b) we have ν(a + b) ≥ min(ν(a), ν(b)) = ν(a). Moreover
0 = ν(1) = ν((−1)(−1)) = 2ν(−1) therefore giving ν(−b) = ν(−1) + ν(b) = ν(b).
Hence ν(a) ≥ min(ν(a + b), ν(−b)) = min(ν(a + b), ν(b)). But ν(a) < ν(b) and so
ν(a) ≥ ν(a + b) giving ν(a + b) = ν(a).
Before looking at specific examples of complete valued fields we first consider some of
the theory concerning series representations of elements.
2.1.1 Series expansions of elements of valued fields
Definition 2.1.8. Let F be a valued field. If 1 is an isolated point of F×F, equivalently
0 is an isolated point of ν(F×) for ν a valuation logarithm on F, then the valuation on F
is said to be discrete, else it is said to be dense.
Lemma 2.1.9. If a valued field F has a discrete valuation then ν(F×) is a discrete subset of R
for ν a valuation logarithm on F.
Proof. We show the contrapositive. Suppose there is a sequence (an) of elements of F×
such that ν(an) converges to a point of R with ν(an) (cid:54)= limm→∞ ν(am) for all n ∈ N.
We can take (an) to be such that ν(an) (cid:54)= ν(am) for n (cid:54)= m. Then setting bn := ana−1
defines a sequence (bn) such that
ν(bn) = ν(ana−1
n+1) = ν(an) − ν(an+1)
n+1) = ν(an) + ν(a−1
n+1
which converges to 0.
The following standard definitions are particularly important.
Definition 2.1.10. For F a non-Archimedean field with valuation logarithm ν, Define:
7
CHAPTER 2: COMPLETE VALUED FIELDS
(i) OF := {a ∈ F : ν(a) ≥ 0, equivalently aF ≤ 1} the ring of integers of F noting
that this is a ring by the strong triangle inequality;
F := {a ∈ F : ν(a) = 0, equivalently aF = 1} the units of OF;
(ii) O×
(iii) MF := {a ∈ F : ν(a) > 0, equivalently aF < 1} the maximal ideal of OF of
elements without inverses in OF;
(iv) F := OF/MF the residue field of F of residue classes.
Definition 2.1.11. Let F be a field with a discrete valuation and valuation logarithm ν.
(i) If FF = {0, 1}, equivalently ν(F) = {0, ∞}, then the valuation is called trivial.
(ii) If · F is not trivial then an element π ∈ F× such that ν(π) = min ν(F×) ∩ (0, ∞)
is called a prime element since π (cid:54)= ab for all a, b ∈ OF\O×
F given above.
Remark 2.1.12. For a field F, as in part (ii) of Definition 2.1.11, it follows easily from
Lemma 2.1.9 that F has a prime element π and from Remark 2.1.6 that ν(F×) = ν(π)Z
which we call the value group. Moreover for a ∈ F× we have
(cid:17)−ν(a)/ν(π)
(cid:16)π−1
F
aF = r−ν(a) = eν(π) log(r)(−ν(a)/ν(π)) = elog(π−1
F )(−ν(a)/ν(π)) =
giving a rank 1 valuation logarithm 1
ν(π) ν noting that π−1
F > 1 since r > 1.
Theorem 2.1.13. Let F be a valued field with a non-trivial, discrete valuation. Let π be a prime
element of F and let R ⊆ O×
F ∪ {0} be a set of residue class representatives with 0 representing
¯0 = MF. Then every element a ∈ F× has a unique series expansion over R of the form
a =
∞
∑
i=n
aiπi
for some n ∈ Z with an (cid:54)= 0.
Moreover if F is complete then every series over R of the above form defines an element of F×.
Remark 2.1.14. Concerning Theorem 2.1.13.
(i) A proof is given in [Sch06, p28], in fact a generalisation of Theorem 2.1.13 is also
given that can be applied to non-Archimedean fields with a dense valuation.
(ii) For a = ∑∞
i=n aiπi as in Theorem 2.1.13 and using the rank 1 valuation logarithm
of Remark 2.1.12 we have, for each i ≥ n, ν(aiπi) = ν(ai) + iν(π) = i if aiπi (cid:54)= 0
and ν(aiπi) = ∞ otherwise. Further by Lemma 2.1.7 bm := ∑m
i=n aiπi defines a
Cauchy sequence in F, with respect to · F, and its limit is a. Hence, since for
each m > n we have aF − bmF ≤ a − bmF, bmF converges in R to aF. But
ν(bm) = n for all m > n by Lemma 2.1.7 and so ν(a) = n.
We will now look at some examples of complete valued fields and consider the avail-
ability of such structures in the Archimedean and non-Archimedean settings.
8
CHAPTER 2: COMPLETE VALUED FIELDS
2.1.2 Examples of complete valued fields
Example 2.1.15. Here are some non-Archimedean examples.
(i) Let F be any field. Then F with the trivial valuation is a non-Archimedean field.
It is complete noting that in this case each Cauchy sequences will be constant
after some finite number of initial values. The trivial valuation induces the triv-
ial topology on F where every subset of F is clopen i.e. both open and closed.
Furthermore F will coincide with its own residue field.
(ii) There are examples of complete non-Archimedean fields of non-zero character-
istic with non-trivial valuation. For each there is a prime p such that the field is
a transcendental extension of the finite field Fp of p elements. The reason why
such a field is not an algebraic extension of Fp follows easily from the fact that
the only valuation on a finite field is the trivial valuation. One example of this
sort is the valued field of formal Laurent series Fp{{T}} in one variable over Fp
with termwise addition,
∑n∈Z anTn + ∑n∈Z bnTn := ∑n∈Z(an + bn)Tn,
multiplication in the form of the Cauchy product,
(∑n∈Z anTn)(∑n∈Z bnTn) := ∑n∈Z(∑i∈Z aibn−i)Tn,
and valuation given at zero by 0T := 0 and on the units Fp{{T}}× by,
∑n∈Z anTnT := r−min{n:an(cid:54)=0} for any fixed r > 1.
The valuation on Fp{{T}} is discrete and its residue field is isomorphic to Fp.
The above construction also gives a complete non-Archimedean field if we re-
place Fp with any other field F, see [Sch06, p288].
(iii) On the other hand complete valued fields of characteristic zero necessarily con-
tain one of the completions of the rational numbers Q. The Levi-Civita field R
is such a valued field, see [SB10]. Each element a ∈ R can be represented as a
formal power series of the form
a = ∑q∈Q aqTq with aq ∈ R for all q ∈ Q
such that for each q ∈ Q there are at most finitely many q(cid:48) < q with aq(cid:48)
(cid:54)= 0.
Moreover addition, multiplication and the valuation for R can all be obtained
by analogy with example (ii) above. A total order can be put on the Levi-Civita
field such that the order topology agrees with the topology induced by the field's
9
CHAPTER 2: COMPLETE VALUED FIELDS
valuation which is non-trivial. To verify this one shows that the order topology
sub-base of open rays topologically generates the valuation topology sub-base of
open balls and vice versa. This might be useful to those interested in generalising
the theory of C*-algebras to new fields where there is a need to define positive
elements. The completion of Q that the Levi-Civita field contains is in fact Q
itself since the valuation when restricted to Q is trivial.
We consider what examples of complete Archimedean fields there are. Since the only
valuation on a finite field is the trivial valuation, it follows from Remark 2.1.3 that
every Archimedean field is of characteristic zero. Moreover every non-trivial valuation
on the rational numbers is given by Ostrowski's Theorem, see [FV02, p2][Sch06, p22].
Theorem 2.1.16. A non-trivial valuation on Q is either a power of the absolute valuation · c∞,
p for some prime p ∈ N with positive
with 0 < c ≤ 1, or a power of the p-adic valuation · c
c ∈ R.
Remark 2.1.17. We will look at the p-adic valuations on Q and the p-adic numbers in
Example 2.1.18. We note that any two of the valuations mentioned in Theorem 2.1.16
that are not the same up to a positive power will also not be equivalent as norms. Fur-
ther, since all of the p-adic valuations are non-Archimedean, Theorem 2.1.16 implies
that every complete Archimedean field contains R, with a positive power of the ab-
solute valuation, as a valued sub-field. It turns out that almost all complete valued
fields are non-Archimedean with R and C being the only two Archimedean exceptions
up to isomorphism as topological fields, see [Sch06, p36]. This in part follows from
the Gel'fand-Mazur Theorem which depends on spectral analysis involving Liouville's
Theorem and the Hahn-Banach Theorem in the complex setting. We will return to these
issues in the more general setting of Banach F-algebras.
Example 2.1.18. Let p ∈ N be a prime. Then with reference to Remark 2.1.6, for n ∈ Z,
(cid:40)
νp(n) :=
max{i ∈ N0 : pin} if n (cid:54)= 0
∞
if n = 0
, N0 := N ∪ {0},
extends uniquely to Q under the properties of a valuation logarithm. Indeed for n ∈ N
we have
0 = νp(1) = νp(n/n) = νp(n) + νp(1/n)
giving νp(1/n) = −νp(n) etc. The standard p-adic valuation of a ∈ Q is then given
by ap := p−νp(a). This is a discrete valuation on Q with respect to which p is a prime
element in the sense of Definition 2.1.11. Moreover Rp := {0, 1,· · · , p − 1} is one
choice of a set of residue class representatives for Q. This is because for m, n ∈ N with
10
CHAPTER 2: COMPLETE VALUED FIELDS
p (cid:45) m and p (cid:45) n we have that m, using the Division algorithm, can be expressed as
m = a1 + pb1 and 1/n, using the extended Euclidean algorithm, can be expressed as
1/n = a2 + pb2/n with a1, a2 ∈ {1,· · · , p − 1} and b1, b2 ∈ Z. Hence, with reference to
Definition 2.1.10, m/n can be expressed as m/n = a3 + pb3/n with a3 ∈ {1,· · · , p − 1}
and pb3/n ∈ Mp as required. With these details in place we can apply Theorem 2.1.13
so that every element a ∈ Q× has a unique series expansion over Rp of the form
a =
∞
∑
i=n
ai pi
for some n ∈ Z with an (cid:54)= 0.
The completion of Q with respect to · p is the field of p-adic numbers denoted Qp.
The elements of Q×
p are all of the series of the above form when using the expansion
over Rp. Further, with reference to Remark 2.1.14, for such an element a = ∑∞
i=n ai pi
with an (cid:54)= 0 we have νp(a) = n. As an example of such expansions for p = 5 we have,
= 3 · 50 + 2 · 5 + 2 · 52 + 2 · 53 + 2 · 54 + · · · .
1
2
More generally the residue field of Qp is the finite field Fp of p elements. Each non-zero
element of Fp has a lift to a p − 1 root of unity in Qp, see [FV02, p37]. These roots of
unity together with 0 also constitute a set of residue class representatives for Qp. The
ring that they generate embeds as a ring into the complex numbers, e.g. see Figure 2.1.
Moreover as a field, rather than as a valued field, Qp has an embedding into C. The
Figure 2.1: Part of a ring in C. The points are labeled with the first coefficient of their
corresponding 5-adic expansion over R5 under a ring isomorphism.
p-adic valuation on Qp can then be extended to a complete valuation on the complex
numbers which in this case as a valued field we denote as Cp, see [Sch06, 46][Roq84].
11
CHAPTER 2: COMPLETE VALUED FIELDS
Finally it is interesting to note that the different standard valuations on Q, when re-
stricted to the units Q×, are related by the equation · 0 · 2 · 3 · 5 · · · · ∞ = 1 where
· 0 denotes the trivial valuation and · ∞ the absolute value function. See [FV02, p3].
2.1.3 Topological properties of complete valued fields
In this subsection we consider the connectedness and local compactness of complete
valued fields.
Definition 2.1.19. Let X be a topological space and Y ⊆ X.
(i) If Y cannot be expressed as the disjoint union of two non-empty clopen subsets
with respect to the relative topology then Y is said to be a connected subset of X.
(ii) If the only non-empty connected subsets of X are singletons then X is said to be
totally disconnected.
(iii) If for each pair of points x, y ∈ X there exists a continuous map f
I := [0, 1] ⊆ R, such that f (0) = x and f (1) = y then X is path-connected.
: I → X,
(iv) A neighborhood base Bx at a point x ∈ X is a collection of neighborhoods of x such
that for every neighborhood U of x there is V ∈ Bx with V ⊆ U.
(v) We call X locally compact if and only if each point in X has a neighborhood base
consisting of compact sets.
The following lemma is well known however I have provided a proof for the reader's
convenience.
Lemma 2.1.20. Let F be a non-Archimedean field. Then F is totally disconnected.
Proof. Let F be a non-Archimedean field and r ∈ R with r > 0. For a, b ∈ F, a ∼ b
if and only if a − bF < r defines an equivalence relation on F by the strong triangle
inequality noting that for transitivity if a ∼ b and b ∼ c then
a − cF ≤ max(a − bF,b − cF) < r.
Hence for a ∈ F the F-ball Br(a) is an equivalence class and so every element of Br(a) is
at its center because every element is an equivalence class representative. In particular
if b ∈ Br(a) then Br(b) = Br(a) but for b /∈ Br(a) we have Br(b) ∩ Br(a) = ∅, showing
that Br(a) is clopen. Since this holds for every r > 0, a has a neighborhood base of
clopen balls. Hence since F is Hausdorff, {a} is the only connected subset of F with a
as an element and so F is totally disconnected.
12
CHAPTER 2: COMPLETE VALUED FIELDS
Remark 2.1.21. We make the following observations.
(i) Every complete Archimedean field is path-connected whereas every complete
non-Archimedean field is totally disconnected.
(ii) In general a valued field being totally disconnected is not the same as it being
discrete. For example Q with the absolute valuation is a totally disconnected
Archimedean field but it is obviously neither discrete nor complete. Also it is
easy to show that a valued field admits a non-constant path if and only if it is
path-connected, see [Wil04, p197] for the standard definitions used here.
(iii) With reference to the proof of Lemma 2.1.20, a ∼ b if and only if a − bF ≤ r,
noting the change from the strict inequality, is again an equivalence relation on F.
Hence every ball of positive radius in a non-Archimedean field is clopen although
a ball ¯Br(a) := {b ∈ F : b − aF ≤ r} may contain elements in addition to those in
Br(a) depending on whether r ∈ F×F. To clarify then, in the non-Archimedean
setting ¯Br(a) does not denote the closure of Br(a) with respect to the valuation.
(iv) In section 4.1.1 concerning complex uniform algebras we will look at Swiss cheese
sets. For a non-Archimedean field F if a, b ∈ F and r1, r2 ∈ R with r1 ≥ r2 > 0
then either Br2(b) ⊆ Br1(a) or Br2(b) ∩ Br1(a) = ∅ since either Br1(b) = Br1(a) or
Br1(b) ∩ Br1(a) = ∅. Further if S is an F-ball or the complement of an F-ball then
the closure of S with respect to · F coincides with S since F-balls are clopen.
Hence a Swiss cheese set X ⊆ F will be classical exactly when there exists a
such that each element of D is a subset of ∆ and X = ∆\(cid:83) D. It follows that such
countable or finite collection D of F-balls, with finite radius sum, and an F-ball ∆
a set X can be empty in the non-Archimedean setting.
Theorem 2.1.22. Let X be a Hausdorff space. Then X is locally compact if and only if each
point in X has a compact neighborhood.
Theorem 2.1.23. Let F be a complete non-Archimedean field that is not simultaneously both
infinite and with the trivial valuation. Then the following are equivalent:
(i) F is locally compact;
(ii) the residue field F is finite and the valuation on F is discrete;
(iii) each bounded sequence in F has a convergent subsequence;
(iv) each infinite bounded subset of F has an accumulation point in F;
(v) each closed and bounded subset of F is compact.
13
CHAPTER 2: COMPLETE VALUED FIELDS
Proofs of Theorem 2.1.22 and Theorem 2.1.23 can be found in [Wil04, p130] and [Sch06,
p29,p57] respectively.
Remark 2.1.24. Concerning Theorem 2.1.23.
(i) Let F be an infinite field with the trivial valuation. Then F is locally compact
since for a ∈ F it follows that {{a}} is a neighborhood base of compact sets for
a. However F does not have any of the other properties given in Theorem 2.1.23.
For example the residue field F is F and so it is not finite and F itself is closed and
bounded but not compact etc.
(ii) We will call a complete non-Archimedean field F that satisfies (ii) of Theorem
2.1.23 a local field. Some authors weaken the condition on the residue field when
defining local fields so that the residue field needs only to be of prime charac-
teristic for some prime p and perfect, that is the Frobenius endomorphism on F,
¯a → ¯ap, is an automorphism.
(iii) Since the only complete Archimedean fields are R and C all but property (ii) of
Theorem 2.1.23 hold for complete Archimedean fields by the Heine-Borel Theo-
rem and Bolzano-Weierstrass Theorem etc. In fact this provides one way to prove
Theorem 2.1.23 since if F is a local field then there is a homeomorphic embedding
of F into C as a closed unbounded subset.
(iv) By the details given in Example 2.1.18 it is immediate that for each prime p the
field of p-adic numbers Qp is a local field. However Cp is not a local field since
its valuation is dense and its residue field is infinite, see [Sch06, p45].
2.2 Extending complete valued fields
In this section and later chapters we will adopt the following notation.
Notation
If F is a field and L is a field extending F then we will denote the Galois group of
F-automorphisms on L, that is automorphisms on L that fix the elements of F, by
Gal(L/F). Further we will denote fixed fields by:
(i) Lg := {x ∈ L : g(x) = x}, for g ∈ Gal(L/F);
(ii) LG :=(cid:84)
g∈G Lg, for a subgroup G (cid:54) Gal(L/F).
14
CHAPTER 2: COMPLETE VALUED FIELDS
More generally if S is a set and G is a group of self maps g : S → S, with group law
composition, then we will denote:
(1) ord(g) := min{n ∈ N : g(n) = id}, the order of an element g ∈ G with finite
order;
(2) ord(g, s) := min{n ∈ N : g(n)(s) = s}, the order at an element s ∈ S, when finite,
of an element g ∈ G;
(3) ord(g, S) := {ord(g, s) : s ∈ S}, the order set of an element g ∈ G with finite
order.
In the rest of this section we will look mainly at extensions of valued fields, including
their valuations, as well as some Galois theory used in later chapters.
2.2.1 Extensions
The first theorem below is rather general in scope.
Theorem 2.2.1. Let F be a complete non-Archimedean field. All non-Archimedean norms, that
is norms that observe the strong triangle inequality, on a finite dimensional F-vector space E
are equivalent. Further E is a Banach space, i.e. complete normed space, with respect to each
norm.
Theorem 2.2.1 also holds for complete Archimedean fields and in the Archimedean set-
ting proofs often make use of the underlying field being locally compact. However the
complete non-Archimedean field F in Theorem 2.2.1 is not assumed to be locally com-
pact and so a proof of the theorem from [Sch06] has been included below for interest.
Proof of Theorem 2.2.1. We use induction on n := dim E. The base case n = 1 is imme-
diate. Suppose Theorem 2.2.1 holds for (n − 1)-dimensional spaces and let E be such
that dim E = n. We choose a base e1,· · · , en for E and define
(cid:107)x(cid:107)∞ =: max
i
aiF
for x =
n∑
aiei ∈ E.
i=1
Note that (cid:107) · (cid:107)∞ is a non-Archimedean norm on E by · F being a non-Archimedean
valuation. Now let (cid:107) · (cid:107) be any other non-Archimedean norm on E. We show that (cid:107) · (cid:107)
is equivalent to (cid:107) · (cid:107)∞. For x = ∑n
i=1 aiei ∈ E we have
(cid:107)x(cid:107) ≤ max
aiF(cid:107)ei(cid:107) ≤ M(cid:107)x(cid:107)∞
i
where M := maxi (cid:107)ei(cid:107). Hence it remains to show that there is a positive constant N
such that for all x ∈ E we have (cid:107)x(cid:107) ≥ N(cid:107)x(cid:107)∞. Let D be the linear subspace generated
15
CHAPTER 2: COMPLETE VALUED FIELDS
by e1,· · · , en−1. By the inductive hypothesis there is c > 0 such that for all x ∈ D we
have (cid:107)x(cid:107) ≥ c(cid:107)x(cid:107)∞. Further D is complete and hence closed in E with respect to (cid:107) · (cid:107).
Hence for
c(cid:48) := (cid:107)en(cid:107)−1 inf{(cid:107)en − y(cid:107) : y ∈ D}
we have 0 < c(cid:48) ≤ 1. Now set N := min(c(cid:48)c, c(cid:48)(cid:107)en(cid:107)) and let x ∈ E. Then x = y + anen
for some y ∈ D and an ∈ F. If an (cid:54)= 0 then (cid:107)x(cid:107) = anF(cid:107)en + a−1
n y(cid:107) ≥ anF(cid:107)en(cid:107)c(cid:48) =
c(cid:48)(cid:107)anen(cid:107). But then we also get
(cid:107)x(cid:107) ≥ max(c(cid:48)(cid:107)x(cid:107), c(cid:48)(cid:107)anen(cid:107)) ≥ c(cid:48)(cid:107)x − anen(cid:107) = c(cid:48)(cid:107)y(cid:107)
and this inequality also holds for an = 0 since 0 < c(cid:48) ≤ 1. We get
(cid:107)x(cid:107) ≥ c(cid:48) max((cid:107)y(cid:107),(cid:107)anen(cid:107)) ≥ c(cid:48) max(c(cid:107)y(cid:107)∞,(cid:107)en(cid:107)anF) ≥ N max((cid:107)y(cid:107)∞,anF)
where N max((cid:107)y(cid:107)∞,anF) = N(cid:107)x(cid:107)∞. Hence (cid:107) · (cid:107) and (cid:107) · (cid:107)∞ are equivalent. Finally
we note that a sequence in E is a Cauchy sequence with respect to (cid:107) · (cid:107)∞ if and only if
each of its coordinate sequences is a Cauchy sequence with respect to · F. Hence E is a
Banach space with respect to each norm by the equivalence of norms and completeness
of F.
Remark 2.2.2. Let L be a non-Archimedean field and let F be a complete subfield of L.
If L is a finite extension of F then L is also complete by Theorem 2.2.1. Now suppose L
is such a complete finite extension of F, then viewing L as a finite dimensional F-vector
space we note that convergence in L is coordinate-wise since · L is equivalent to (cid:107) · (cid:107)∞
by Theorem 2.2.1. Hence each element g ∈ Gal(L/F) is continuous since being linear
over F. We will see later in this section that, for such complete finite extensions, each
element of Gal(L/F) is in fact an isometry. Finally in all cases if L is complete and g
is continuous then the fixed field Lg is also complete. To see this let (an) be a Cauchy
sequence in Lg and let a be its limit in L. For id the identity map on L, note that g − id
is also continuous on L and so Lg = (g − id)−1(0) is a closed subset of L. In particular
we have a ∈ Lg as required.
The following is Krull's extension theorem, a proof can be found in [Sch06, p34].
Theorem 2.2.3. Let F be a subfield of a field L and let · F be a non-Archimedean valuation
on F. Then there exists a non-Archimedean valuation on L that extends · F.
The following corollary to Theorem 2.2.3, which also uses Theorem 2.1.4, contrasts with
the Archimedean setting.
Corollary 2.2.4. For every complete non-Archimedean field F there exists a proper extension
L of F for which the complete valuation on F extends to a complete valuation on L.
16
CHAPTER 2: COMPLETE VALUED FIELDS
Moreover an extension of a valuation is often unique.
Theorem 2.2.5. Let F be a complete non-Archimedean field, let L be an algebraic extension of
F and let a ∈ L. Then:
(i) there is a unique valuation · L on L that extends the valuation on F;
(ii) if (cid:107) · (cid:107) is an arbitrary norm on the F-vector space L then aL = limn→∞ n(cid:112)(cid:107)an(cid:107).
Remark 2.2.6. We make the following observations.
(i) Part (i) of Theorem 2.2.5 follows easily from Theorem 2.2.3 and Theorem 2.2.1
applied respectively noting that if a ∈ L then a is also an element of a finite
extension of F. See [Sch06, p39] for the rest of the proof.
(ii) It is worth emphasizing that Theorem 2.2.1, Theorem 2.2.3 and Theorem 2.2.5 all
hold for the case where the valuation on F is trivial.
(iii) Now for L and F conforming to the conditions of Theorem 2.2.5 we have that
each g ∈ Gal(L/F) is indeed an isometry on L since a(cid:48) := g(a)L, for a ∈ L, is a
valuation on L giving g(a)L = aL by uniqueness.
The following theory will often allow us to express the extension of a valuation in a
particularly useful form. We begin with a standard theorem.
Theorem 2.2.7. Let F be a field, let L be an algebraic extension of F and let a ∈ L. Then there
is a unique monic irreducible polynomial IrrF,a(x) ∈ F[x] such that IrrF,a(a) = 0. Moreover,
for the simple extension F(a), we have [F(a), F] = degIrrF,a(x) where [F(a), F] denotes the
dimension of F(a) as an F-vector space.
Definition 2.2.8. Let F be a field and let L be an algebraic extension of F.
(i) An element a ∈ L is said to be separable over F if a is not a repeated root of its own
irreducible polynomial IrrF,a(x).
(ii) We call Lsc := {a ∈ L : a is separable over F} the separable closure of F in L.
(iii) The extension L is said to be a separable extension of F if L = Lsc.
(iv) Let f (x) ∈ F[x]. Then L is called a splitting field of f (x) over F if f (x) splits
completely in L[x] as a product of linear factors but not over any proper subfield
of L containing F.
17
CHAPTER 2: COMPLETE VALUED FIELDS
(v) We will call L a normal extension of F if L is the splitting field over F of some
polynomial in F[x].
(vi) The field L is called a Galois extension of F if LG = F for G := Gal(L/F).
Remark 2.2.9. Following Definition 2.2.8 we note that the separable closure Lsc of F in
L is a field with F ⊆ Lsc ⊆ L. Moreover if F is of characteristic zero then L is a separable
extension of F.
For proofs of the following two theorems and Remark 2.2.9 see [McC66, p13-p19,p36].
Theorem 2.2.10. Let F be a field and let L be a finite extension of F. Then there is a normal
extension Lne of F which contains L and which is the smallest such extension in the sense that
if K is a normal extension of F which contains L then there is a L-monomorphism of Lne into K,
i.e. an embedding of Lne into K that fixes L.
Theorem 2.2.11. Let F be a field and let L be a finite extension of F. Then, with reference to
Theorem 2.2.10 and Definition 2.2.8, there are exactly [Lsc : F] distinct F-isomorphisms of L
onto subfields of Lne. Further if L = Lne then #Gal(L/F) = [Lsc : F]. Moreover L is a Galois
extension of F if and only if Lsc = L = Lne in which case #Gal(L/F) = [L : F].
Definition 2.2.12. Let F be a field, let L be a finite extension of F and let n0 := [Lsc : F].
By Theorem 2.2.11 there are exactly n0 distinct F-isomorphisms g1,· · · , gn0 of L onto
subfields of Lne. The norm map NL/F : L → F is defined as
(cid:33)[L:Lsc]
(cid:32) n0∏
i=1
NL/F (a) :=
gi(a)
for a ∈ L.
A proof showing that the norm map only takes values in the ground field can be found
in [McC66, p23,p24]. Using the preceding theory we can now state and prove a theorem
that will often allow us to express the extension of a valuation in a particularly useful
form. The theorem is in the literature. However, having set out the preceding theory,
the proof presented here is more immediate than the sources I have seen.
Theorem 2.2.13. Let F be a complete non-Archimedean field with valuation aF = r−ν(a), for
a ∈ F, where ν is a valuation logarithm on F. Let L be a finite extension of F as a field. Then,
with reference to Theorem 2.2.5 and Theorem 2.2.1, the unique extension of · F to a complete
valuation · L on L is given by
aL = n(cid:113)NL/F (a)F = r−ω(a)
for a ∈ L,
n ν ◦ NL/F is the corresponding extension of ν to L. If in addition
where n = [L : F] and ω := 1
the valuation · F is discrete then · L is also discrete. If further · F is non-trivial and ν is
the rank 1 valuation logarithm of remark 2.1.12 then eω(L×) = Z for some e ∈ N.
18
CHAPTER 2: COMPLETE VALUED FIELDS
Proof. Let Lne be the normal extension of F containing L of Theorem 2.2.10. Since Lne
is the splitting field of some polynomial in F[x] it is a finite extension of F and so also
of L. Hence by Theorem 2.2.5 the valuation · L extends uniquely to a valuation · Lne
on Lne. Let n0 := [Lsc : F] and let g1,· · · , gn0 be the n0 distinct F-isomorphisms of L
onto subfields of Lne as given by Theorem 2.2.11. Then for each i ∈ {1,· · · , n0} we
have that ai := gi(a)Lne, for a ∈ L, is a valuation on L extending · F. Hence, by the
uniqueness of · L as an extension of · F to L, each of g1,· · · , gn0 is an isometry from
L onto a subfield of Lne with respect to · Lne. Hence setting n := [L : F], n1 := [L : Lsc]
and noting that the norm map NL/F takes values in F, we have for all a ∈ L
(cid:33)n1
(cid:118)(cid:117)(cid:117)(cid:116)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:32) n0∏
i=1
= n
gi(a)
(cid:33)n1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Lne
= n(cid:113)NL/F (a)F.
n(cid:113)
(cid:118)(cid:117)(cid:117)(cid:116)(cid:32) n0∏
i=1
aL =
a[L:Lsc][Lsc:F]
L
= n
gi(a)Lne
n ν(an) = 1
n ν ◦ NL/F (a) = 1
n ν ◦ NL/F (L×) ⊆ 1
n ν(F×) and so if · F is a discrete
Therefore we also have ω(L×) = 1
valuation then so is · L. Moreover ω is indeed an extension of ν since for a ∈ F we
have ω(a) = 1
n nν(a) = ν(a). Now suppose that ν is a rank
1 valuation logarithm so that ν(F×) = Z and ω(L×) ⊆ 1
Z. Then there are at most
n elements in ω(L×) ∩ (0, 1] but also at least 1 element since there is π ∈ F× that is
prime with respect to ν giving ω(π) = ν(π) = 1. Hence let e(cid:48) := min ω(L×) ∩ (0, 1]
and a ∈ L× such that ω(a) = e(cid:48). We show that ω(L×) = e(cid:48)Z. Let b ∈ L× giving
ω(b) = ke(cid:48) + ε for some 0 ≤ ε < e(cid:48) and k ∈ Z. Then since ak, a−k ∈ L× we have
ω(ba−k) = ω(b) − kω(a) = ke(cid:48) + ε − ke(cid:48) = ε giving ε = 0 by the definition of e(cid:48).
Hence ω(L×) ⊆ e(cid:48)Z. On the other hand for k ∈ Z we have ω(ak) = kω(a) = ke(cid:48) so
e(cid:48)Z ⊆ ω(L×) giving ω(L×) = e(cid:48)Z. Finally since ω(π) = 1 we have 1 ∈ e(cid:48)Z and so
there is e ∈ N such that e(cid:48)e = 1 giving eω(L×) = Z which completes the proof.
n
Remark 2.2.14. Let F and L be as in Theorem 2.2.13 with non-trivial discrete valuations.
Let ν be the rank 1 valuation logarithm on F and let ω be the extension of ν to L.
(i) With group law addition, ω(F×) and ω(L×) are groups. It is immediate from
Theorem 2.2.13 that e = [ω(L×) : ω(F×)], the index of ω(F×) in ω(L×).
(ii) If e = 1 then L is called an unramified extension of F. If e = [L : F] then L is
called a totally ramified extension of F. Other classifications are also in use in the
literature.
(iii) The value of e has implications for the degree of the extension L of the residue
field F. For F and L as specified in these remarks we have [L : F] = e[L : F],
see [McC66, p107,p108] for details. Hence in this case, with reference to Theorem
2.1.23, if F is locally compact then L is locally compact.
19
CHAPTER 2: COMPLETE VALUED FIELDS
2.2.2 Galois theory
The following is the fundamental theorem of Galois theory, see [McC66, p36].
Theorem 2.2.15. Let F and E be fields such that E is a finite Galois extension of F, that is
EG = F for G := Gal(E/F). Then we have the following one-one correspondence
{G(cid:48) : G(cid:48) (cid:54) G is a subgroup} ↔ {E(cid:48) : E(cid:48) is a field with F ⊆ E(cid:48) ⊆ E}
given by the inverse maps G(cid:48) (cid:55)→ EG(cid:48)
and E(cid:48) (cid:55)→ Gal(E/E(cid:48) ).
Corollary 2.2.16. Let F and L be fields such that L is a finite extension of F and let G :=
Gal(L/F). Then L is a finite Galois extension of LG and so for L and LG Theorem 2.2.15 is
applicable.
Proof. We show that L is a Galois extension of LG. For g ∈ Gal(L/F) we have g(a) = a
for all a ∈ LG and so g ∈ Gal(L/LG ). On the other hand for g ∈ Gal(L/LG ) we have
g(a) = a for all a ∈ F since F ⊆ LG and so g ∈ Gal(L/F). Therefore Gal(L/F) =
Gal(L/LG ) and so setting G(cid:48) := Gal(L/LG ) gives LG(cid:48)
= LG as required.
The following group theory result must be known. However we will provide a proof
in lieu of a reference.
Lemma 2.2.17. Let (G, +) be a group and g ∈ Aut(G) be a group automorphism on G. If
a, b ∈ G are such that gcd(ord(g, a), ord(g, b)) = 1 then ord(g, a + b) = ord(g, a)ord(g, b).
Proof. We assume the conditions of Lemma 2.2.17 and note that the result is imme-
diate if one or more of ord(g, a) and ord(g, b) is equal to 1. So assuming otherwise,
let pk1
j be the prime decompositions of ord(g, a) and ord(g, b)
respectively. For n := ord(g, a)ord(g, b) we have
2 · · · pki
i and ql1
2 · · · qlj
1 pk2
1 ql2
g(n)(a + b) = g(n)(a) + g(n)(b) = a + b.
Therefore ord(g, a + b)n. Suppose towards a contradiction that ord(g, a + b) < n.
r for some r ∈ {p1, p2,· · · , pi, q1, q2,· · · , qj}. If r = pm for some
Then ord(g, a + b) n
m ∈ {1, 2,· · · , i} then a + b = g( n
r )(a) + b giving, by
right cancellation of b, g( n
r )(a + b) = g( n
r )(b) = g( n
r )(a) + g( n
r )(a) = a. It then follows that
2 · · · pkm−1
m · · · pki
i pk1
1 pk2
2 · · · pkm
m
2 · · · qlj
j
· · · pki
i ql1
1 ql2
pk1
1 pk2
2 · · · qlj
1 ql2
giving pmql1
j which is a contradiction since gcd(ord(g, a), ord(g, b)) = 1. A
similar contradiction occurs for r = qm with m ∈ {1, 2,· · · , j}. Hence ord(g, a + b) = n
as required.
20
CHAPTER 2: COMPLETE VALUED FIELDS
Lemma 2.2.18. Let F be a field with finite extension L and let g ∈ Gal(L/F). If n ∈ N is such
that nord(g) then n ∈ ord(g, L).
Proof. Suppose towards a contradiction that there is n ∈ N such that nord(g) but
n /∈ ord(g, L). We can take n to be the least such element and note that n (cid:54)= 1 since
1 ∈ ord(g, L). Express n as n = pkr where p is a prime, k, r ∈ N and p (cid:45) r. We thus have
the following two cases.
Case: r (cid:54)= 1. In this case by the definition of n we have pk, r ∈ ord(g, L) and so there
are a, b ∈ L with ord(g, a) = pk, ord(g, b) = r and
gcd(ord(g, a), ord(g, b)) = 1.
Then by Lemma 2.2.17 we have ord(g, a + b) = ord(g, a)ord(g, b) which contradicts
our assumption that n /∈ ord(g, L).
Case: r = 1. In this case n = pk and note that ord(g) = nm for some m ∈ N. Hence we
have the following subgroups of G:
p )(cid:105) := ({id, g( n
(i) (cid:104)g(n)(cid:105) := ({id, g(n), g(2n),· · · , g((m−1)n)},◦) < G;
p )},◦) (cid:54) G.
(ii) (cid:104)g( n
p ), g(2 n
Therefore #(cid:104)g(n)(cid:105) = m, #(cid:104)g( n
Hence by Corollary 2.2.16 we have the following tower of fields
p ),· · · , g((mp−1) n
p )(cid:105) = mp and (cid:104)g(n)(cid:105) is a proper normal subgroup of (cid:104)g( n
p )(cid:105).
LG ⊆ L(cid:104)g(n/p)(cid:105) (cid:36) L(cid:104)g(n)(cid:105) ⊆ L.
Now it is immediate that Lg(n) = L(cid:104)g(n)(cid:105) and Lg(n/p) = L(cid:104)g(n/p)(cid:105) and so there is some
a ∈ Lg(n)\Lg(n/p) with ord(g, a)n but ord(g, a) (cid:45) n
p. Therefore ord(g, a) = pk = n which
again contradicts our assumption that n /∈ ord(g, L). In particular the lemma holds.
The following lemma is well known but we will provide a proof in lieu of a reference.
Lemma 2.2.19. Let F be a field, let L be an algebraic extension of F and let a ∈ L. For the
simple extension F(a) of F and F[X] the ring of polynomials over F we have F(a) = F[a].
p(a)
Proof. It is immediate that F[a] ⊆ F(a). Now by Theorem 2.2.7 there is a unique monic
irreducible polynomial IrrF,a(x) ∈ F[X] such that IrrF,a(a) = 0. Further for any element
q(a) ∈ F(a), given by p(x), q(x) ∈ F[X], we have q(a) (cid:54)= 0. Hence IrrF,a(x) and q(x) are
relatively prime, that is we have gcd(IrrF,a, q) = 1. Therefore by Bezout's identity there
are s(x), t(x) ∈ F[X] such that s(x)q(x) + t(x)IrrF,a(x) = 1 giving q(x) = 1−t(x)IrrF,a(x)
.
Finally then we have q(a) = 1
q(a) = p(a)s(a) which is an element of F[a] as
required.
s(a) giving p(a)
s(x)
21
CHAPTER 3
Functions and algebras
In this chapter we build upon some of the basic facts and analysis of complete valued
fields surveyed in Chapter 2. The first section establishes particular facts in functional
analysis over complete valued fields that will be used in later chapters. However it is
not the purpose of the first section to provide an extensive introduction to the subject.
The second section provides background on Banach F-algebras, Banach algebras over a
complete valued field F. Whilst some of the details are included purely as background
others also support the discussion from Remark 2.1.17 of Chapter 2.
3.1 Functional analysis over complete valued fields
We begin with the following lemma.
Lemma 3.1.1. Let F be a non-Archimedean field and let (an) be a sequence of elements of F.
(i) If limn→∞ an = a for some a ∈ F× then there exists N ∈ N such that for all n ≥ N we
have anF = aF. We will call this convergence from the side, opposed to from above
or below.
(ii) If F is also complete then ∑ an converges if and only if limn→∞ an = 0, in sharp contrast
to the Archimedean case. Further if ∑ an does converge then
(cid:12)(cid:12)∑ an
(cid:12)(cid:12)F ≤ max{anF : n ∈ N}.
Proof. For (i), since a (cid:54)= 0 we have aF > 0 and so there is some N ∈ N such that,
for all n ≥ N, an − aF < aF. Hence for all n ≥ N we have by Lemma 2.1.7 that
anF = (an − a) + aF = aF.
For (ii), suppose limn→∞ an = 0 and let ε > 0. Then there is N ∈ N such that for all
22
CHAPTER 3: FUNCTIONS AND ALGEBRAS
n ≥ N we have anF < ε. Hence for n1, n2 ∈ N with N < n1 < n2 we have
ai − n1∑
i=1
ai
=
n2∑
i=n1+1
ai
≤ max{an1+1F,· · · ,an2F} < ε.
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)F
Hence the sequence of partial sums is a Cauchy sequence in F and so converges. The
converse is immediate. Further suppose ∑ an does converge. For ∑ an (cid:54)= 0 we have by
(i) that there is N ∈ N such that for all n ≥ N
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n2∑
i=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∞
∑
i=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)F
ai
=
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n∑
i=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)F
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)F
ai
≤ max{a1F,· · · ,anF} ≤ max{aiF : i ∈ N}.
On the other hand for ∑ an = 0 the result is immediate.
The following theorem appears in [Sch06, p59] without proof.
Theorem 3.1.2. Let F be a complete non-Archimedean field and let a0, a1, a2,· · · be a sequence
of elements of F. Define the radius of convergence by
ρ :=
1
lim supn→∞
n(cid:112)anF
where by convention 0−1 = ∞ and ∞−1 = 0.
Then the power series ∑ anxn, x ∈ F, converges if xF < ρ and diverges if xF > ρ. Further-
more for each t ∈ (0, ∞), t < ρ the convergence is uniform on ¯Bt(0) := {a ∈ F : aF ≤ t}.
Proof. Note that the following equalities hold, except for when xF = ρ = 0,
n(cid:113)anxnF = lim sup
n(cid:113)anFxn
n→∞
lim sup
n→∞
xF
ρ
.
(3.1.1)
F =
ε. Hence since limm→∞ m√
Suppose ∑ anxn is divergent. Then by part (ii) of Lemma 3.1.1, limn→∞ anxn is not 0.
Therefore there is some ε ∈ (0, 1] such that for each m ∈ N there is n > m with
ε = 1 we have
anxnF ≥ ε, in particular n(cid:112)anxnF ≥ n√
n(cid:112)anxnF ≥ 1. Therefore in this case xF ≥ ρ by (3.1.1). In particular for
ε ≥ m√
lim supn→∞
cases where xF < ρ the series ∑ anxn converges.
n(cid:113) 1
n(cid:112)anxnF ≤ lim supn→∞
On the other hand suppose ∑ anxn converges. Then since limn→∞ anxnF = 0 we have
2 = 1. Therefore in this case xF ≤ ρ by (3.1.1).
lim supn→∞
In particular for cases where xF > ρ the series ∑ anxn diverges.
Now suppose there is t ∈ (0, ∞) with t < ρ and let ε > 0.
If the valuation on F
is dense then F×F is dense in the positive reals and so there is some x0 ∈ F× with
t < x0F < ρ. Alternatively, if the valuation on F is discrete, there is x0 ∈ ¯Bt(0) with
x0F = max{aF : a ∈ ¯Bt(0)}. In either case, since x0F < ρ, ∑ anxn
0 converges and
0 = 0. Hence there is some N ∈ N such that for all n > N we have
so limn→∞ anxn
23
CHAPTER 3: FUNCTIONS AND ALGEBRAS
0F < ε. Now by the last part of Lemma 3.1.1 we have for all m > N and x ∈ ¯Bt(0)
anxn
that
∞
∑
n=1
anxn − m∑
But anxnF = anFxn
and the convergence is uniform on ¯Bt(0).
anxnF = ∞
∑
n=m+1
F = anxn
F ≤ anFx0n
n=1
anxnF ≤ max{anxnF : n ≥ m + 1}.
0F and so max{anxnF : n ≥ m + 1} < ε
Remark 3.1.3. With reference to Theorem 3.1.2.
(i) We note that the radius of convergence as defined in Theorem 3.1.2 is the same as
that used in the Archimedean setting when replacing F with the complex num-
bers. However, unlike in the complex setting, if the valuation on F is discrete
then a power series ∑ anxn may not have a unique choice for the definition of its
radius of convergence since F×F is discrete in this case.
(ii) We need to be careful when considering convergence of power series. Let · ∞
denote the absolute valuation on R and let · 0 denote the trivial valuation on
R. All power series are convergent on B1(0) := {a ∈ R : a0 < 1} = {0} with
respect to · 0. Whereas the only power series that are convergent at a point a ∈
R× with respect to · 0 are polynomials. On the other hand exp(x) := ∑∞
xn
n!
converges everywhere on R with respect to · ∞. The function exp(x) defined
with respect to · ∞ is a continuous function on all of R with respect to · 0 but
does not have a power series representation on R with respect to · 0. Similarly
∑∞
xn
n! does not converge everywhere on the p-adic numbers Qp with respect to
· p, see [Sch06, p70] for details in this case.
n=1
n=1
(iii) Under the conditions of Theorem 3.1.2, suppose that the ball Bρ(0) is without
isolated points where ρ is the radius of convergence of f (x) := ∑ anxn. Then,
with differentiation defined as in the Archimedean setting, the derivative of f
exists on Bρ(0) and it is f (cid:48)(x) = ∑ nanxn−1. We will not consider this in depth
but note, for x ∈ Bρ(0), the series ∑ anxn converges giving limn→∞ cn = 0 for
cn := anxn by Lemma 3.1.1. Hence for all n ∈ N, since nF = 11 + · · · + 1nF ≤
max{11F,· · · ,1nF} = 1, we have for x (cid:54)= 0 that
nanxn−1F = nFx−1FanxnF ≤ x−1FcnF.
Therefore the series ∑ nanxn−1 also converges on Bρ(0) by Lemma 3.1.1.
3.1.1 Analytic functions
Let F be a complete valued field. In this subsection we consider F valued functions
that are analytic on the interior of some subset of F that is without isolated points. In
24
CHAPTER 3: FUNCTIONS AND ALGEBRAS
particular the situation concerning such analytic functions is some what different in the
non-Archimedean setting to that in the Archimedean one, even though the standard re-
sults of differentiation such as the chain rule and Leibniz rule are the same, see [Sch06,
p59]. Recall that if a complex valued function f is analytic on an open disc Dr(a) ⊆ C
then f can be represented by the convergent power series
f (z) =
∞
∑
n=0
f (n)(a)
n!
(z − a)n
for z ∈ Dr(a),
known as the Taylor expansion of f about a, where f (n)(a) is the nth derivative of f at a.
Moreover if b ∈ Dr(a) then f can also be expanded about b. However this expansion
need not be convergent on all of Dr(a) merely on the largest open disc centered at b
contained in Dr(a) since a lack of differentiability of f at points on, or outside, the
boundary of Dr(a) will restrict the radius of convergence of such an expansion, see
[Apo74, p449,p450].
Now for F a complete non-Archimedean field the same scenario in this case is such
that if f is analytic on a ball Br(a) ⊆ F and can be represented by a Taylor expansion
about a on all of Br(a) then f can be represented by a Taylor expansion about any other
point b ∈ Br(a) and this expansion will also be valid on all of Br(a), see [Sch06, p68].
This is closely related to the fact that every point of Br(a) is at its center, see the proof
of Lemma 2.1.20. However in general a function f analytic on Br(a) ⊆ F need not have
a Taylor expansion about a that is valid on all of Br(a). This is because Br(a) can be
decomposed as a disjoint union of clopen balls, see Remark 2.1.21, upon each of which
f can independently be defined. This leads to the following definitions.
Definition 3.1.4. Let F be a complete valued field with non-trivial valuation.
(i) We will call a subset X ⊆ F strongly convex if X is either F, the empty set ∅, a ball
or a singleton set.
(ii) Let X be an open strongly convex subset of F and let f : X → F be a continuous
F-valued function on X. If f can be represented by a single Taylor expansion that
is valid on all of X then we say that f is globally analytic on X.
(iii) Let X be an open subset of F and let f : X → F be a continuous F-valued function
on X. If for each a ∈ X there is an open strongly convex neighborhood V ⊆ X of
a such that fV is globally analytic on V then we say that f is locally analytic on X.
(iv) Let X and f be as in (iii). As usual, if the derivative
f (x) − f (a)
x − a
exists at every a ∈ X then we say that f is analytic on X.
f (cid:48)
(a) := lim
x→a
25
CHAPTER 3: FUNCTIONS AND ALGEBRAS
(v) In the case where X = F we similarly define globally entire, locally entire and entire
functions on X.
Remark 3.1.5. Note that the condition in Definition 3.1.4 that F has a non-trivial valua-
tion is there because it does not make sense to talk about analytic functions defined on
a space without accumulation points. We also note that (ii), (iii) and (iv) of Definition
3.1.4 are equivalent in the complex setting for X an open strongly convex subset of C,
see [Apo74, p450].
Now let F be a complete non-Archimedean field and let ( fn) be a sequence of F-valued
functions analytic on ¯B1(0) and converging uniformly on ¯B1(0) to a function f . We
ask whether f will also be analytic on ¯B1(0) in this case? It is very well known that the
answer to the analog of this question involving the complex numbers is yes although in
this case the functions are required to be continuous on ¯B1(0) and analytic only on the
interior of ¯B1(0) since ¯B1(0) will not be clopen. In the case involving the real numbers
the answer to the question is of course no since for example a function with a chevron
shaped graph in R2 can be uniformly approximated by differentiable functions. In the
non-Archimedean setting the following theorem provides insight for when F is not
locally compact and also gives a maximum principle result, see [Sch06, p122] for proof.
Theorem 3.1.6. Let F be a complete non-Archimedean field that is not locally compact and let
r ∈ F×F.
(i) If f1, f2,· · · are globally analytic functions on ¯Br(0) and if f := limn→∞ fn uniformly
on ¯Br(0) then f is also globally analytic on ¯Br(0).
(ii) Let f be a globally analytic function on ¯Br(0) with power series f (x) = ∑∞
n=0 anxn.
If the valuation · F is dense then
sup{ f (x)F : xF ≤ r} = sup{ f (x)F : xF < r} = max{anFrn : n ≥ 0} < ∞.
If the residue field F is infinite then
max{ f (x)F : xF ≤ r} = max{ f (x)F : xF = r} = max{anFrn : n ≥ 0} < ∞.
Remark 3.1.7. In Theorem 3.1.6 ¯Br(0) is not compact since F is not locally compact. In
fact every ball of positive radius is not compact in this case and this follows from Theo-
rem 2.1.23 noting that translations and non-zero scalings in F are homeomorphisms on
F. Now since we are progressing towards a study of uniform algebras and their gen-
eralisation over complete valued fields we note that in order to use the uniform norm,
see Remark 4.1.2, on such algebras of continuous functions we need the functions to be
bounded. Hence to avoid imposing boundedness directly it is convenient to work on
compact spaces.
26
CHAPTER 3: FUNCTIONS AND ALGEBRAS
For ¯B1(0) compact, i.e. in the F locally compact case, I provide the following example to
show that in this case the uniform limit of locally analytic, and hence analytic, functions
on ¯B1(0) need not be analytic.
Example 3.1.8. Let F be a locally compact, complete non-Archimedean field with non-
trivial valuation. Then we have the following sequence of functions on ¯B1(0) ⊆ F,
fn(x) :=
πν(x)
0
if
if
ν(x) < n
ν(x) ≥ n
for x ∈ ¯B1(0)
where π is a prime element of F and ν is the rank 1 valuation logarithm. For each n ∈
N, fn is a locally constant function since convergence in F is from the side, see Lemma
3.1.1, and so fn is locally analytic on ¯B1(0). Moreover the sequence ( fn) converges
uniformly on ¯B1(0) to the continuous function
(cid:40)
(cid:40)
f (x) :=
πν(x)
0
if
if
x (cid:54)= 0
x = 0
for x ∈ ¯B1(0)
with limx→0 f (x) = 0 since f (x)F = xF for all x ∈ ¯B1(0). We now show that f is not
differentiable at zero. let a1, a2,· · · and b1, b2,· · · be sequences in F given by an := πn
and bn := −πn. Both of these sequences tend to zero as n tends to ∞. But then
f (an) − f (0)
an
= (πn − 0)π
−n = 1
and
f (bn) − f (0)
bn
= (πn − 0)(−π
−n) = −1
for all n ∈ N so that the limit limx→0
we can obtain a similar example by redefining f as
f (x)− f (0)
x
does not exist as required. Alternatively
π
π
0
f (x) :=
1
2 ν(x)
2 (ν(x)−1)
1
if
if
if
ν(x) is even
ν(x) is odd
x = 0
for x ∈ ¯B1(0).
f (x)− f (0)
In this case limx→0
quence c1, c2,· · · with cn := π2n.
x
blows up with respect to · F as demonstrated by the se-
Later when we look at non-complex analogs of uniform algebras we will see, from Ka-
plansky's non-Archimedean generalisation of the Stone-Weierstrass theorem, that the
continuous functions in Example 3.1.8 can be uniformly approximated by polynomials
on ¯B1(0) given that ¯B1(0) is compact in this case. Hence Example 3.1.8 also shows that,
for F locally compact, the uniform limit of globally analytic functions on ¯B1(0) need
not be analytic in contrast to Theorem 3.1.6.
In anticipation of topics in the next section we now consider Liouville's theorem. It is
immediate that the standard Liouville theorem never holds in the non-Archimedean
27
CHAPTER 3: FUNCTIONS AND ALGEBRAS
setting since for a complete non-Archimedean field F with non-trivial valuation the in-
dicator function χB for B := ¯B1(0) is a non-constant bounded locally analytic function
from F to F noting that ¯B1(0) is a clopen subset of F. However the following is called
the ultrametric Liouville theorem.
Theorem 3.1.9. Let F be a complete non-Archimedean field with non-trivial valuation. Then
every bounded globally analytic function from F to F is constant if and only if F is not locally
compact.
Proof. See [Sch06, p124,p125] for a full proof of Theorem 3.1.9. However proof in the if
n=0 anxn, for x ∈ F, be as in Theorem 3.1.9. Since f
direction is as follows. Let f (x) = ∑∞
is bounded there is M < ∞ such that f (x)F ≤ M for all x ∈ F. Let m ∈ N. Since F is
not locally compact we can apply (ii) of Theorem 3.1.6 so that for r ∈ F×F we have
amFrm ≤ max{anFrn : n ≥ 0} = sup{ f (x)F : xF ≤ r} ≤ M.
This holds for every r ∈ F×F and so am = 0 leaving f = a0 for all x ∈ F.
For a field F with the trivial valuation we note that F is locally compact and that there
are bounded non-constant polynomials from F to F, where we take polynomials to be
the analog of globally analytic functions in this case.
3.2 Banach F-algebras
We begin this section with the following definitions.
Definition 3.2.1. Let F be a complete valued field.
(i) A general Banach ring is a normed ring R that is complete with respect to its norm
which is required to be sub-multiplicative, i.e.
(cid:107)ab(cid:107)R ≤ (cid:107)a(cid:107)R(cid:107)b(cid:107)R
for all a, b ∈ R.
We do not assume that R has a multiplicative identity or that its multiplication is
commutative, we merely assume it is associative.
(ii) A Banach ring is a general Banach ring R that has a left/right multiplicative iden-
tity satisfying (cid:107)1R(cid:107)R = 1 = (cid:107) − 1R(cid:107)R.
(iii) A Banach F-algebra is a general Banach ring A that is also a normed vector space
over F, with respect to the ring's addition operation and norm, and such that the
ring's multiplication operation is a bilinear map over F, i.e. respectively
(cid:107)αa(cid:107)A = αF(cid:107)a(cid:107)A and (αa)b = a(αb) = α(ab)
for all a, b ∈ A and α ∈ F.
28
CHAPTER 3: FUNCTIONS AND ALGEBRAS
(iv) A unital Banach F-algebra is a Banach F-algebra that is also a Banach ring opposed
to being merely a general Banach ring.
(v) By commutative general Banach ring and commutative Banach F-algebra etc. we mean
that the multiplication is commutative in these cases. By F-algebra we mean the
structure of a Banach F-algebra but without the requirement of a norm.
Remark 3.2.2. In Definition 3.2.1 we always require a multiplicative identity to be dif-
ferent to the additive identity. As standard we will usually dispense with the sub-
script when denoting elements of the structures defined in Definition 3.2.1 and in the
Archimedean setting we will call a Banach C-algebra a complex Banach algebra and a
Banach R-algebra a real Banach algebra.
3.2.1 Spectrum of an element
The following discussion concerns the spectrum of an element.
Definition 3.2.3. Let F be a complete valued field and let A be a unital Banach F-
algebra. Then for a ∈ A we call the set
Sp(a) := {λ ∈ F : λ − a is not invertible in A}
the spectrum of a.
Theorem 3.2.4. Every element of every unital complex Banach algebra has non-empty spec-
trum.
Theorem 3.2.4 is very well known. A proof can be found in [Sto71, p11] and relies
on Liouville's theorem and the Hahn-Banach theorem in the complex setting. We will
confirm that this result is unique among unital Banach F-algebras and I will give details
of where the proof from the complex setting fails for other complete valued fields.
Let us first recall the Gelfand-Mazur theorem which demonstrates the importance of
Theorem 3.2.4 in the complex setting and supports Remark 2.1.17 of Chapter 2.
Theorem 3.2.5. A unital complex Banach algebra that is also a division ring is isometrically
isomorphic to the complex numbers.
Proof. Let A be a unital complex Banach algebra that is also a division ring and let
a ∈ A. Since in this case Sp(a) is non-empty, there is some λ ∈ Sp(a). Hence because
A is a division ring λ − a = 0 giving a = λ. More accurately we have a = λ1A but
because A is unital we have (cid:107)a(cid:107)A = (cid:107)λ1A(cid:107)A = λ∞(cid:107)1A(cid:107)A = λ∞ and so the map
from A onto C given by λ1A (cid:55)→ λ is an isometric isomorphism.
29
CHAPTER 3: FUNCTIONS AND ALGEBRAS
Remark 3.2.6. In the Archimedean setting it follows immediately from Theorem 3.2.5
that any complete valued field containing the complex numbers as a valued subfield
will coincide with the complex numbers. Note that the proof of Theorem 3.2.5 is very
well known.
In contrast to Theorem 3.2.4 we have the following lemma. The result is certainly
known but we give full details in lieu of a reference.
Lemma 3.2.7. Let F be a complete valued field other than the complex numbers. Then there
exists a unital Banach F-algebra A such that Sp(a) = ∅ for some a ∈ A.
Proof. Let F be a complete valued field other that the complex numbers. By Corollary
2.2.4 in the non-Archimedean setting, and since R is the only complete valued field
other than C in the Archimedean setting, we can always find a complete valued field L
that is a proper extension of F. Let a ∈ L\F and note that L is a unital Banach F-algebra.
Then for every λ ∈ F we have λ − a (cid:54)= 0 and so λ − a is invertible in L since L is a field.
Hence Sp(a) = ∅.
Whilst not considering every case, we now consider where the proof of Theorem 3.2.4
fails when applying it to unital Banach F-algebras with F (cid:54)= C. For F = R the Hahn-
Banach theorem holds but Liouville's theorem does not with the trigonometric sin func-
tion restricted to R as an example of a non-constant, bounded, analytic function from
R to R. In the non-Archimedean setting we do have the ultrametric Liouville theorem,
Theorem 3.1.9 for F not locally compact, and there is also an ultrametric Hahn-Banach
theorem for spherically complete fields, as follows.
Definition 3.2.8. An ultrametric space, see Definition 2.1.1, is spherically complete if each
nested sequence of balls has a non-empty intersection.
Theorem 3.2.9. Let F be a spherically complete non-Archimedean field and let V be an F-vector
space, s a seminorm on V and V0 ⊆ V a vector subspace. Then for every linear functional
(cid:96)0 : V0 → F such that (cid:96)0(v)F ≤ s(v) for all v ∈ V0 there is a linear functional (cid:96) : V → F
such that (cid:96)V0 = (cid:96)0 and (cid:96)(v)F ≤ s(v) for all v ∈ V.
Remark 3.2.10. We note that Theorem 3.2.9 is exactly the same as the Hahn-Banach
theorem from the Archimedean setting, see [Sto71, p472], except with R and C replaced
by any spherically complete non-Archimedean field. A proof can be found in both
[Sch02, p51] and [Sch06, p288] the latter of which further states that Theorem 3.2.9
becomes a falsity if F is replaced by a non-spherically complete field. It is immediate
that spherically complete ultrametric spaces are complete.
A proof of the following lemma can be found in [Sch02, p6].
30
CHAPTER 3: FUNCTIONS AND ALGEBRAS
Lemma 3.2.11. All complete non-Archimedean fields with a discrete valuation are spherically
complete. In particular if F is a complete non-Archimedean fields that is locally compact then F
is spherically complete.
From the above details we see that for both Theorem 3.1.9 and Theorem 3.2.9 to be ap-
plicable we need a non-locally compact, spherically complete, non-Archimedean field.
This restricts the possibilities since for example, for any prime p, a finite extension of
Qp is locally compact and Cp whilst not locally compact is also not spherically com-
plete, see [Sch02, p5]. However, with reference to (ii) of Example 2.1.15, the complete
non-Archimedean field C{{T}} is not locally compact since having an infinite residue
field and it is spherically complete since its valuation is discrete. Moreover the totally
T) is a unital Banach C{{T}}-
ramified, see Remark 2.2.14, simple extension C{{T}}(
algebra with complete valuation given by Theorem 2.2.13. But by the proof of Lemma
T) = ∅. So let's briefly review how the proof of Theorem 3.2.4 from
3.2.7 we have Sp(
[Sto71, p11] works and then consider where it fails for C{{T}}(
Let A be a unital complex Banach algebra and let a ∈ A. Suppose towards a contra-
diction that Sp(a) = ∅. Then λ − a is invertible for all λ ∈ C. In particular a−1 exists
in A and the map (cid:96)0 : Ca−1 → C, given by (cid:96)0(λa−1) := λα for a fixed α ∈ C with
0 < α∞ ≤ (cid:107)a−1(cid:107)A, is a continuous linear functional on the subspace Ca−1 of A to
which the Hahn-Banach theorem can be applied directly. Hence there exists a continu-
ous linear functional (cid:96) : A → C such that (cid:96)(−a−1) = −α (cid:54)= 0. On the other hand for
any continuous linear functional ϕ : A → C we can define a function f ϕ : C → C by
T).
√
√
√
f ϕ(λ) := ϕ((λ − a)
−1).
The proof then shows that f ϕ is differentiable at every point of C and is therefore an
entire function. Moreover limλ→∞ f ϕ(λ) = 0 since
≤ 1
λ∞
(cid:107)ϕ(cid:107)op(cid:107)(1 − λ
f ϕ(λ)∞ =
ϕ((1 − λ
(cid:12)(cid:12)(cid:12)(cid:12) 1
−1(cid:107)A,
(cid:12)(cid:12)(cid:12)(cid:12)∞
−1a)
λ
−1a)
−1)
√
where (cid:107) · (cid:107)op is the standard operator norm. Hence, by Liouville theorem in the com-
plex setting, f ϕ is the zero function. But we have f(cid:96)(0) = −α (cid:54)= 0, a contradiction, and
so Sp(a) (cid:54)= ∅ as required. Note however that the function f ϕ is defined on C\Sp(a).
T) → C{{T}} given by
Now for C{{T}}(
T) := α, where α, β ∈ C{{T}}, is a continuous linear functional analogous
P(α + β
√
to an evaluation functional noting that convergence in C{{T}}(
T) is coordinate-wise
over C{{T}} by Remark 2.2.2. Hence we can define a function fP : C{{T}} → C{{T}}
given by
T) the coordinate projection P : C{{T}}(
√
√
√
−1) = P((λ +
T)
31
fP(λ) := P((λ −
√
T)(λ2 − T)
−1) = λ(λ2 − T)
−1.
CHAPTER 3: FUNCTIONS AND ALGEBRAS
P (λ) = (λ2 − T)
f (1)
P (0) = − 1
(cid:18) λ
(cid:19)
The function fP is defined on all of C{{T}} since the roots of λ2 − T are
T.
Furthermore fP is not constant and so it is the relative weakness of the ultrametric
Liouville theorem in the non-Archimedean setting that allows the argument used in
the proof of Theorem 3.2.4 to fails in this case. Indeed we will now show that fP is not
globally analytic on all of C{{T}}. The first derivative of fP is
−1 − 2λ2(λ2 − T)
−2
√
T and −√
and so fP(0) = 0 and f (1)
expansion of fP about zero as
T . Continuing in this way we obtain the Taylor
fP(λ) =
∞
∑
n=0
αnλn = −
λ3
T2 +
λ5
T3 +
λ7
T4 + · · ·
T +
for λT < ρ,
,
where αn :=
n! = − 1−(−1)n
f (n)
P (0)
2
T− 1
2 (1+n) ∈ C{{T}} and
ρ =
1
lim supn→∞
n(cid:112)αnT
is the radius of convergence of the Taylor series expansion. Hence we show that ρ
is finite. Using the rank 1 valuation logarithm, for ∑n∈Z anTn ∈ C{{T}}× we have
∑n∈Z anTnT = r− min{n:an(cid:54)=0} for some fixed r > 1. Hence, noting that α2n = 0 and
α2n−1 = −T−n for n ∈ N, we have
n(cid:113)αnT = lim
n→∞
2n−1(cid:113)α2n−1T = lim
n→∞
2n−1√
lim sup
n→∞
r < 1 since r > 1. In particular fP is only locally analytic on C{{T}}
Hence ρ = 1√
and not globally analytic, consistent with the ultrametric Liouville theorem not being
applicable to fP as required.
rn = lim
n→∞ r n
2n−1 = r 1
2 .
Definition 3.2.12. Let F be a complete valued field and let A be a unital Banach F-
algebra. Define F (A) as the set of all complete valued fields L contained inside A over
which A is also a unital Banach L-algebra.
Remark 3.2.13. Concerning the spectrum of an element.
(i) It is tempting to conjecture that a generalisation of Theorem 3.2.4 might hold for
every complete valued field F provided that, given F, we restrict the statement
to those unital Banach F-algebras A for which F is a maximal element of F (A).
This conjecture is false in both the non-commutative and commutative settings
by Lemma 3.2.14 below. However for a more general version of the conjecture
one could permit the elements of F (A) to be complete normed division rings.
32
CHAPTER 3: FUNCTIONS AND ALGEBRAS
(ii) Let A be a unital real Banach algebra. In order to avoid an element a ∈ A having
empty spectrum Kaplansky gave the following alternative definition in this case,
SpK(a) := {α + iβ ∈ C : (a − α)2 + β2 is not invertible in A}.
We won't investigate this definition here but for more details see [KL92, p6].
Lemma 3.2.14. In both the non-commutative and commutative algebra settings one can find
a complete valued field F, a unital Banach F-algebra A and an element a ∈ A such that F is a
maximal element of F (A) and Sp(a) = ∅.
Proof. Hamilton's real quaternions, H, are an example of a non-commutative complete
Archimedean division ring and unital real Banach algebra. Viewing H as a real vec-
tor space, the valuation on H is the Euclidean norm which is complete, Archimedean
and indeed a valuation since being multiplicative on H, see [Lam05, p56,p57]. By the
Gelfand-Mazur theorem, Theorem 3.2.5, H is not a unital complex Banach algebra since
being different to C and so R is maximal in F (H). Moreover for a ∈ H\R it is imme-
diate that we have Sp(a) = ∅.
In the commutative setting consider the field of complex numbers C with the absolute
valuation replaced by the L1-norm as it applies to the real vector space R2, that is for
a = α + iβ ∈ C we have (cid:107)a(cid:107)1 := α∞ + β∞. Then C is complete with respect to (cid:107) · (cid:107)1
by the equivalence of norms on finite dimensional R-vector spaces. Expressing com-
plex numbers in their coordinate form it is easy to show that (cid:107) · (cid:107)1 is sub-multiplicative
and so (C,(cid:107) · (cid:107)1) is a unital real Banach algebra. However (cid:107) · (cid:107)1 is not multiplicative
since (cid:107)(1 + i)(1 − i)(cid:107)1 = (cid:107)2(cid:107)1 = 2 < 4 = (cid:107)1 + i(cid:107)1(cid:107)1 − i(cid:107)1 and so (cid:107) · (cid:107)1 is not a valua-
tion on C. Consequently R is maximal in F ((C,(cid:107) · (cid:107)1)) and over R, Sp(i) = ∅ which
completes the proof.
33
CHAPTER 4
Uniform algebras
In the first section of this chapter we survey some of the basic facts about complex
uniform algebras and recall the close connection with the study of compact Hausdorff
spaces, such as Swiss cheese sets, upon which such algebras of functions are defined.
An inductive proof by the author of the Feinstein-Heath Swiss cheese "classicalisa-
tion" theorem is then presented. An article containing this proof has been published
by the American Mathematical Society, see [Mas10]. In the second section of this chap-
ter we turn our attention to non-complex analogs of uniform algebras. The constraints
imposed by the various generalisations of the Stone-Weierstrass theorem are consid-
ered and the theory of real function algebras developed by Kulkarni and Limaye is
introduced. We will establish the topological requirements of the spaces upon which
algebras of functions in the non-Archimedean setting can be defined whilst qualifying
as non-complex analogs of uniform algebras. These observations together with some
of the details and examples from other chapters have been gathered together by the
author into a survey paper which was subsequently accepted for publication by the
American Mathematical Society, see [Mas11].
4.1 Complex uniform algebras
Definition 4.1.1. Let CC(X) be the unital complex Banach algebra of all continuous
complex valued functions, defined on a compact Hausdorff space X, with pointwise
operations and the sup norm given by
(cid:107) f(cid:107)∞ := sup
x∈X
f (x)∞ for all f ∈ CC(X).
A uniform algebra, A, is a subalgebra of CC(X) that is complete with respect to the sup
norm, contains the constant functions making it a unital complex Banach algebra and
separates the points of X in the sense that for all x1, x2 ∈ X with x1 (cid:54)= x2 there is f ∈ A
satisfying f (x1) (cid:54)= f (x2).
34
CHAPTER 4: UNIFORM ALGEBRAS
Remark 4.1.2. Introductions to uniform algebras can be found in [Bro69], [Gam69] and
[Sto71]. Some authors take Definition 4.1.1 to be a representation of uniform algebras
and take a uniform algebra A to be a unital complex Banach algebra with a square
preserving norm, that is (cid:107)a2(cid:107) = (cid:107)a(cid:107)2 for all a ∈ A, which they sometimes then referred
to as a uniform norm. This is quite legitimate since, as we will discuss at depth in
the section on representation theory, the Gelfand transform shows us that every such
algebra is isometrically isomorphic to an algebra conforming to Definition 4.1.1. In this
thesis we mainly introduce generalisations of Definition 4.1.1 over complete valued
fields and then investigate the important representation theory results. Hence for us
by uniform norm we will mean the sup norm.
It is very well known that in the complex setting, for suitable X, there exist uniform
algebras that are proper subalgebras of CC(X). However if A is such a uniform algebra
then A is not self-adjoint, that is there is f ∈ A with ¯f /∈ A where ¯f denotes the complex
conjugate of f . This result is the complex Stone-Weierstrass theorem, generalisations of
which we will meet in Section 4.2. We will also meet several analogs of the following
example.
Example 4.1.3. A standard example is the disc algebra A(∆) ⊆ CC(∆), of continuous
functions analytic on the interior of ∆ := {z ∈ C : z ≤ 1}, which is as far from
being self-adjoint as possible since if both f and ¯f are in A(∆) then f is constant, see
[KL92, p47]. Also P(∆) = A(∆) where P(∆) is the uniform algebra of all functions on
∆ that can be uniformly approximated by polynomials restricted to ∆ with complex
coefficients. This largely follows from Remark 3.1.5, see [Bro69, p5] or [Sto71, p2].
For a compact Hausdorff space X let R(X) denote the uniform algebra of all functions
on X that can be uniformly approximated by rational functions from CC(X). We also
generalise to X the uniform algebras introduced in Example 4.1.3 giving A(X) and
P(X). In the theory of uniform approximation it is standard to ask for which X is one
or more of the following inclusions non-trivial
P(X) ⊆ R(X) ⊆ A(X) ⊆ CC(X).
Whilst not always the case, this often only depend on X up to homeomorphism. In
particular many properties of uniform algebras are topological properties of the spaces
upon which they are defined. Hence there is a strong connection between the study of
uniform algebras and that of compact Hausdorff spaces. Therefore, in addition to being
of interest in their own right, uniform algebras are important in the theory of uniform
approximation; as examples of complex Banach algebras; in representation theory and
in the study of compact Hausdorff space. With respect to the latter, we now turn our
attention to the compact plane sets known as Swiss cheese sets.
35
CHAPTER 4: UNIFORM ALGEBRAS
4.1.1 Swiss cheese sets in the complex plane
Throughout subsections 4.1.1 and 4.1.2, all discs in the complex plane are required to
have finite positive radius. More generally let N0 := N ∪ {0} from here on throughout
the thesis. We begin with the following definitions taken from [FH10].
Definition 4.1.4. For a disc D in the plane let r(D) denote the radius of D.
(i) A Swiss cheese is a pair D := (∆,D) for which ∆ is a closed disc and D is a count-
able or finite collection of open discs. A Swiss cheese D = (∆,D) is classical
if the closures of the discs in D intersect neither one another nor C\int∆, and
∑D∈D r(D) < ∞.
∆\(cid:83) D.
(ii) The associated Swiss cheese set of a Swiss cheese D = (∆,D) is the plane set XD :=
A classical Swiss cheese set is a plane set X for which there exists a classical Swiss
cheese D = (∆,D) such that X = XD.
(iii) For a Swiss cheese D = (∆,D), we define δ(D) := r(∆) − ∑D∈D r(D) so that
δ(D) > −∞ if and only if ∑D∈D r(D) < ∞.
Figure 4.1 provides an example for (ii) of Definition 4.1.4. Swiss cheese sets are used
Figure 4.1: A classical Swiss cheese set.
extensively in the theory of uniform algebras since they provide many examples of uni-
form algebras with particular properties. For examples see [Fei04], [Gam69, Ch2] and
[Rot38]. In particular [FH10] includes a survey of the use of Swiss cheese constructions
in the theory of uniform algebras. The following example is from [Rot38].
36
CHAPTER 4: UNIFORM ALGEBRAS
Lemma 4.1.5. For X ⊆ C non-empty and compact, let A1 ⊆ A2 ⊆ CC(X) be uniform
algebras with A0 uniformly dense in A1. Suppose we can find a continuous linear functional
ϕ : CC(X) → C such that ϕ(A0) = {0} and ϕ( f ) = a (cid:54)= 0 for some f ∈ A2. Then A1 (cid:54)= A2.
Proof. Let q ∈ A0. Then
0 < a∞ = ϕ( f ) − ϕ(q)∞ = ϕ( f − q)∞ ≤ (cid:107) f − q(cid:107)∞(cid:107)ϕ(cid:107)op
op > 0 for all q ∈ A0. Hence f can not be uniformly
giving (cid:107) f − q(cid:107)∞ ≥ a∞(cid:107)ϕ(cid:107)−1
approximated by elements of A0. More simply, ϕ(A1) = {0} by continuity.
Example 4.1.6. It is possible to have R(X) (cid:54)= A(X). Let D0 be a closed disc and let
D = (D0,D) be a classical Swiss cheese with δ(D) > 0 and D infinite. Let (Dn) be
a sequence of open discs such that the map n (cid:55)→ Dn is a bijection from N to D. For
n ∈ N0 define γn : [0, 1] → C as the circular path
γn(x) := rn exp(2πix) + an
around the boundary ∂Dn. Now for a rational function q ∈ CC(XD) we note that on C
the finitely many poles of q lie in the open complement of XD and so XD is a subset of
an open subset of C upon which q is analytic. Hence by Cauchy's theorem, see [Rud87,
p218], we have ϕ(q) = 0 for ϕ : CC(XD) → C defined by
ϕ( f ) :=
f dz − ∞
∑
n=1
γ0
for f ∈ CC(XD).
f dz
γn
We now check that ϕ is a bounded linear functional on CC(XD). The following uses the
fundamental estimate. Let f ∈ CC(XD) with (cid:107) f(cid:107)∞ ≤ 1. Then,
ϕ( f )∞ =
f (z)dz − ∞
∑
n=1
γ0
f (z)dz
γn
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)∞
(cid:90)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:90)
≤ ∞
∑
n=0
≤ ∞
∑
n=0
≤ ∞
∑
n=0
∞
∑
n=0
=
(cid:90)
(cid:90)
(cid:12)(cid:12)(cid:12)(cid:12)∞
(cid:12)(cid:12)(cid:12)(cid:12)(cid:90)
(cid:107) f(cid:107)∞
(cid:90) 1
γn
0
f (z)dz
(cid:90) 1
0
γ
n(x)∞dx
(cid:48)
rn2πi exp(2πix)∞dx
(cid:33)
(cid:32) ∞
(cid:90) 1
rn2π
dx = 2π
rn
< 4πr0
0
∑
n=0
where ∑∞
n=0 rn < 2r0 since δ(D) > 0. Now 4πr0 is an upper bound for the series
of absolute terms, hence we have absolute convergence. Since absolute convergence
37
CHAPTER 4: UNIFORM ALGEBRAS
implies convergence we have, for all f ∈ CC(XD),
(cid:19)
(cid:18)(cid:107) f(cid:107)∞
(cid:107) f(cid:107)∞
ϕ( f ) = ϕ
= (cid:107) f(cid:107)∞ ϕ
f
(cid:18) f
(cid:107) f(cid:107)∞
(cid:19)
= (cid:107) f(cid:107)∞a f ∈ C
for some a f ∈ C. Moreover, our calculation shows that ϕ is bounded with (cid:107)ϕ(cid:107)op <
4πr0. The linearity of ϕ follows from the linearity of integrating over a sum of terms
and so ϕ is a continuous linear functional on CC(XD). Next we note that the function
g : z (cid:55)→ ¯z on XD given by complex conjugation is an element of CC(XD). Cauchy's
theorem does not imply that ϕ(g) will be zero since g is not analytic on any non-empty
open subset of C. We have
(cid:90)
(cid:32)
(cid:33)
g(z)dz = 2πi
γn
0 − ∞
∑
r2
n=1
r2
n
,
since for each n ∈ N0
(cid:90)
γ0
ϕ(g) =
(cid:90)
γn
g(z)dz =
g(z)dz − ∞
∑
n=1
(cid:90) 1
(cid:90) 1
0
=
0
= 2πirn
= 2πirn
rn
(cid:48)
ndx
g(γn)γ
(rn exp(−2πix) + ¯an)rn2πi exp(2πix)dx
(cid:90) 1
(cid:18)
(cid:19)
(cid:32)
(rn + ¯an exp(2πix))dx
exp(2πix)dx
dx + ¯an
(cid:90) 1
(cid:90) 1
(cid:33)
0
0
0
exp(2πix)
(cid:20) 1
2πi
(cid:21)1
0
= 2πirn
rn + ¯an
= 2πir2
n.
n ≤ (∑∞
n=1 r2
0 since ∑∞
n=1 rn)2 < r2
Furthermore ∑∞
n=1 rn < r0, by δ(D) > 0, and so
ϕ(g) (cid:54)= 0. Therefore by Lemma 4.1.5 we have R(XD) (cid:54)= CC(XD) and g (cid:54)∈ R(XD).
Certainly this result is immediate in the case where XD has interior since then g will
not be an element of A(XD). However this is one of the occasions where the usefulness
of Swiss cheese set constructions becomes evident since, with some consideration, it is
straightforward to construct a classical Swiss cheese D = (D0,D) with δ(D) > 0 such
that XD has empty interior. Since A(XD) is the uniform algebra of all elements from
CC(XD) that are analytic on the interior of XD, in this case we have A(XD) = CC(XD)
and so by the above R(XD) (cid:54)= A(XD) which completes this example.
Let D be a Swiss cheese as specified in Example 4.1.6 such that XD has empty interior.
The following subsection shows that in this case there is actually no need to require D
to be classical in order that R(XD) (cid:54)= A(XD).
38
CHAPTER 4: UNIFORM ALGEBRAS
4.1.2 Classicalisation theorem
(cid:48)
(cid:48)
with δ(D
In this subsection we give a new proof of an existing theorem by J. F. Feinstein and M.
J. Heath, see [FH10]. The theorem states that any Swiss cheese set defined by a Swiss
cheese D with δ(D) > 0 contains a Swiss cheese set as a subset defined by a classical
) ≥ δ(D). Feinstein and Heath begin their proof by devel-
Swiss cheese D
oping a theory of allocation maps connected to such sets. A partial order on a family
of these allocation maps is then introduced and Zorn's lemma applied. We take a more
direct approach by using transfinite induction, cardinality and disc assignment func-
tions, where a disc assignment function is a kind of labeled Swiss cheese that defines a
Swiss cheese set. An explicit theory of allocation maps is no longer required although
we are still using them implicitly. In this regard we will discuss the connections with
the original proof of Feinstein and Heath. See [Kel75, p266] and [Dal00, p9] for useful
introductions to ordinals and transfinite induction which has been used in this subsec-
tion. We begin with the following definitions.
Definition 4.1.7. Let O be the set of all open discs and complements of closed discs in
the complex plane.
(i) A disc assignment function d : S → O is a map from a subset S ⊆ N0, with 0 ∈ S,
into O such that Dd := (C\d(0), d(S\{0})) is a Swiss cheese. We allow S\{0} to
be empty since a Swiss cheese D = (∆,D) can have D = ∅.
(ii) For a disc assignment function d : S → O and i ∈ S we let
¯d(i) denote the
closure of d(i) in C, that is ¯d(i) := d(i). A disc assignment function d : S → O
is said to be classical if for all (i, j) ∈ S2 with i (cid:54)= j we have ¯d(i) ∩ ¯d(j) = ∅ and
∑n∈S\{0} r(d(n)) < ∞.
(iii) For a disc assignment function d : S → O we let Xd denote the associated Swiss
cheese set of the Swiss cheese Dd.
(iv) A disc assignment function d : S → O is said to have the Feinstein-Heath condition
when ∑n∈S\{0} r(d(n)) < r(C\d(0)).
(v) Define H as the set of all disc assignment functions with the Feinstein-Heath con-
dition.
For h ∈ H, h : S → O, define δh := r(C\h(0)) − ∑n∈S\{0} r(h(n)) > 0.
Here is the Feinstein-Heath Swiss cheese "Classicalisation" theorem as it appears in
[FH10].
39
CHAPTER 4: UNIFORM ALGEBRAS
(cid:48)
) ≥ δ(D).
(cid:48)
(cid:48) ⊆ XD and δ(D
Theorem 4.1.8. For every Swiss cheese D with δ(D) > 0, there is a classical Swiss cheese D
with XD
From Definition 4.1.7 we note that if a disc assignment function d : S → O is classical
then the Swiss cheese Dd will also be classical. Similarly if d has the Feinstein-Heath
condition then δ(Dd) > 0. The converse of each of these implications will not hold in
general because d need not be injective. However it is immediate that for every Swiss
cheese D = (∆,D) with δ(D) > 0 there exists an injective disc assignment function
h ∈ H such that Dh = D. We note that every disc assignment function h ∈ H has
δ(Dh) ≥ δh with equality if and only if h is injective and that classical disc assignment
functions are always injective. With these observations it easily follows that Theorem
4.1.8 is equivalent to the following theorem involving disc assignment function.
Theorem 4.1.9. For every disc assignment function h ∈ H there is a classical disc assignment
function h
(cid:48) ∈ H with Xh(cid:48) ⊆ Xh and δh(cid:48) ≥ δh.
Several lemmas from [FH10] and [Hea08, §2.4.1] will be used in the proof of Theorem
4.1.9 and we consider them now.
Lemma 4.1.10. Let D1 and D2 be open discs in C with radii r(D1) and r(D2) respectively
such that ¯D1 ∩ ¯D2 (cid:54)= ∅. Then there is an open disc D with D1 ∪ D2 ⊆ D and with radius
r(D) ≤ r(D1) + r(D2).
Figure 4.2, Example 1 exemplifies the application of Lemma 4.1.10.
Lemma 4.1.11. Let D be an open disc and ∆ be a closed disc such that ¯D (cid:54)⊆ int∆ and ∆ (cid:54)⊆ ¯D.
Then there is a closed disc ∆(cid:48) ⊆ ∆ with D ∩ ∆(cid:48)
) ≥ r(∆) − r(D).
= ∅ and r(∆(cid:48)
Figure 4.2, Example 2 exemplifies the application of Lemma 4.1.11.
Figure 4.2: Examples for lemmas 4.1.10 and 4.1.11.
40
D1D2DExample1∆∆0DExample2 CHAPTER 4: UNIFORM ALGEBRAS
Lemma 4.1.12. Let F be a non-empty, nested collection of open discs in C, such that we have
sup{r(E) : E ∈ F} < ∞. Then(cid:83) F is an open disc D. Further, for F ordered by inclusion,
inf{r(E) : E ∈ F} > 0. Then (cid:84) F is a closed disc ∆. Further, for F ordered by reverse
r(D) = limE∈F r(E) = supE∈F r(E).
Lemma 4.1.13. Let F be a non-empty, nested collection of closed discs in C, such that we have
inclusion, r(∆) = limE∈F r(E) = infE∈F r(E).
Proof of Theorem 4.1.9. At the heart of the proof of Theorem 4.1.9 is a completely defined
map f : H → H which we now define case by case.
Definition 4.1.14. Let f : H → H be the self map with the following construction.
Case 1: If h ∈ H is a classical disc assignment function then define f (h) := h.
Case 2: If h ∈ H is not classical then for h : S → O let
Ih := {(i, j) ∈ S2 : ¯h(i) ∩ ¯h(j) (cid:54)= ∅, i (cid:54)= j}.
We then have lexicographic ordering on Ih given by
(i, j) (cid:46) (i
(cid:48)
(cid:48)
, j
) if and only if i < i
(cid:48)
or (i = i
(cid:48)
and j ≤ j
(cid:48)
).
Since this is a well-ordering on Ih, let (n, m) be the minimum element of Ih and hence
note that m (cid:54)= 0 since m > n. We proceed toward defining f (h) : S
(cid:48) → O.
Define S
(cid:48)
:= S\{m} and for i ∈ S
(cid:48)\{n} we define f (h)(i) := h(i).
It remains for the definition of f (h)(n) to be given and to this end we have the follow-
ing two cases.
Case 2.1: n (cid:54)= 0.
In this case, by Definition 4.1.7, we note that both h(m) and h(n)
are open discs. Associating h(m) and h(n) with D1 and D2 of Lemma 4.1.10 we de-
fine f (h)(n) to be the open disc satisfying the properties of D of the lemma. Note in
particular that,
h(m) ∪ h(n) ⊆ f (h)(n) with n < m.
(4.1.1)
Case 2.2: n = 0. In this case, by Definition 4.1.7, we note that h(m) is an open disc and
h(0) is the complement of a closed disc. Associate h(m) with D from Lemma 4.1.11 and
put ∆ := C\h(0). Since (0, m) ∈ Ih we have ¯h(0) ∩ ¯h(m) (cid:54)= ∅ and so ¯h(m) (cid:54)⊆ int∆,
noting int∆ = C\¯h(0). Further, since h ∈ H we have r(h(m)) < r(∆) and so ∆ (cid:54)⊆ ¯h(m).
Therefore the conditions of Lemma 4.1.11 are satisfied for h(m) and ∆. Hence we define
f (h)(0) to be the complement of the closed disc satisfying the properties of ∆(cid:48)
of Lemma
4.1.11. Note in particular that,
h(m) ∪ h(0) ⊆ f (h)(0) with 0 < m.
(4.1.2)
41
CHAPTER 4: UNIFORM ALGEBRAS
For this definition of the map f we have yet to show that f maps into H. We now show
this together with certain other useful properties of f .
Lemma 4.1.15. Let h ∈ H, then the following hold:
(i) f (h) ∈ H with δf (h) ≥ δh;
(ii) For (h : S → O) (cid:55)→ ( f (h) : S
classical. Otherwise S
(cid:48)
= S\{m} for some m ∈ S\{0};
(cid:48) → O) we have S
(cid:48) ⊆ S with equality if and only if h is
(iii) X f (h) ⊆ Xh;
(iv) For all i ∈ S
(cid:48)
, h(i) ⊆ f (h)(i).
Proof. We need only check (i) and (iii) for cases 2.1 and 2.2 of the definition of f , as
everything else is immediate. Let h ∈ H.
(i) It is clear that f (h) is a disc assignment function. It remains to check that δf (h) ≥ δh.
For Case 2.1 we have, by Lemma 4.1.10,
δh = r(C\h(0)) − (r(h(m)) + r(h(n))) − ∑
≤ r(C\h(0)) − r( f (h)(n)) − ∑
i∈S\{0,m,n}
For Case 2.2 we have, by Lemma 4.1.11,
r(h(i))
i∈S\{0,m,n}
r(h(i)) = δf (h).
δh = r(C\h(0)) − r(h(m)) − ∑
≤ r(C\ f (h)(0)) − ∑
r(h(i))
i∈S\{0,m}
r(h(i)) = δf (h).
i∈S(cid:48) f (h)(i).
i∈S(cid:48) f (h)(i).
i∈S\{0,m}
i∈S h(i) ⊆(cid:83)
i∈S h(i) we require(cid:83)
For Case 2.1 we have by Lemma 4.1.10 that h(m) ∪ h(n) ⊆ f (h)(n), as shown at (4.1.1),
:= C\ f (h)(0). We have by Lemma 4.1.11 that
= ∅. Hence h(0) ∪ h(m) ⊆ f (h)(0), as shown at (4.1.2), and so
(iii) Since Xh = C\(cid:83)
i∈S h(i) ⊆(cid:83)
giving(cid:83)
For Case 2.2 put ∆ := C\h(0) and ∆(cid:48)
i∈S h(i) ⊆(cid:83)
(cid:83)
∆(cid:48) ⊆ ∆ and h(m) ∩ ∆(cid:48)
We will use f : H → H to construct an ordinal sequence of disc assignment functions
and then apply a cardinality argument to show that this ordinal sequence must stabilise
at a classical disc assignment function. We construct the ordinal sequence so that it has
the right properties.
Definition 4.1.16. Let h ∈ H.
i∈S(cid:48) f (h)(i) as required.
42
CHAPTER 4: UNIFORM ALGEBRAS
(a) Define h0 : S0 → D by h0 := h.
Now let α > 0 be an ordinal for which we have defined hβ ∈ H for all β < α.
(b) If α is a successor ordinal then define hα : Sα → O by hα := f (hα−1).
(c) If α is a limit ordinal then define hα : Sα → O as follows.
(cid:92)
Set Sα :=
Sβ. Then for n ∈ Sα define hα(n) :=
hβ(n).
(cid:91)
β<α
β<α
Suppose that for every ordinal α for which Definition 4.1.16 can be applied we have
hα ∈ H. Then Definition 4.1.16 can be applied for every ordinal α by transfinite induc-
tion and therefore defines an ordinal sequence of disc assignment function. We will use
transfinite induction to prove Lemma 4.1.17 below which asserts that hα is an element
of H as well as other useful properties of hα.
Lemma 4.1.17. Let α be an ordinal number and let h ∈ H. Then the following hold:
(α,1) hα ∈ H with δhα ≥ δh;
(α,1.1) 0 ∈ Sα;
(α,1.2) hα(0) is the complement of a closed disc and
hα(n) is an open disc for all n ∈ Sα\{0};
(α,1.3) ∑n∈Sα\{0} r(hα(n)) ≤ r(C\hα(0)) − δh;
(α,2) For all β ≤ α we have Sα ⊆ Sβ;
(α,3) For all β ≤ α we have Xhα ⊆ Xhβ;
(α,4) For all n ∈ Sα, {hβ(n) : β ≤ α} is a nested increasing family of open sets.
Proof. We will use transfinite induction.
For α an ordinal number let P(α) be the proposition, Lemma 4.1.17 holds at α.
The base case P(0) is immediate and our inductive hypothesis is that for all β < α, P(β)
holds.
Now for α a successor ordinal we have hα = f (hα−1) and so P(α) is immediate by the
Sα := (cid:84)
inductive hypothesis and Lemma 4.1.15. Now suppose α is a limit ordinal. We have
β<α Sβ giving, for all β ≤ α, Sα ⊆ Sβ. Hence (α,2) holds. Also for all β < α we
have 0 ∈ Sβ by (β,1.1). So 0 ∈ Sα showing that (α,1.1) holds. To show (α,1.2) we will
use lemmas 4.1.12 and 4.1.13.
43
CHAPTER 4: UNIFORM ALGEBRAS
(i) Now for all n ∈ Sα\{0}, {hβ(n) : β < α} is a nested increasing family of open
discs by (β,1.2) and (β,4).
(ii) Further, {C\hβ(0) : β < α} is a nested decreasing family of closed discs by (β,1.2)
and (β,4).
(iii) Now for n ∈ Sα\{0} and β < α we have
r(hβ(n)) ≤ ∑m∈Sβ\{0} r(hβ(m)) = r(C\hβ(0)) − δhβ ≤ r(C\h(0)) − δh, by (β,1)
and (ii). Hence sup{r(hβ(n)) : β < α} ≤ r(C\h(0)) − δh. So by (i) and Lemma
4.1.12 we have for n ∈ Sα\{0} that
(cid:91)
hα(n) :=
hβ(n)
is an open disc with,
β<α
r(hα(n)) = sup
β<α
r(hβ(n)) ≤ r(C\h(0)) − δh.
(iv) Now for β < α we have r(C\hβ(0)) ≥ δh by (β,1.3).
(cid:92)
Hence inf{r(C\hβ(0)) : β < α} ≥ δh. So by De Morgan, (ii) and Lemma 4.1.13 we
have
C\hα(0) := C\ (cid:91)
C\hβ(0)
hβ(0) =
is a closed disc with,
β<α
β<α
r(C\hα(0)) = inf
β<α
r(C\hβ(0)) ≥ δh.
Hence hα(0) is the complement of a closed disc and so (α,1.2) holds.
a nested increasing family of open sets. We also have hα(n) = (cid:83)
We now show that (α,4) holds. By (β,4) we have, for all n ∈ Sα, {hβ(n) : β < α} is
β<α hβ(n) so, for all
β ≤ α, hβ(n) ⊆ hα(n) and hα(n) is an open set since (α,1.2) holds. Hence (α,4) holds.
We will now show that (α,1.3) holds. We first prove that, for all λ < α, we have
∑
m∈Sα\{0}
r(hα(m)) ≤ r(C\hλ(0)) − δh.
Let λ < α, and suppose, towards a contradiction, that
∑
m∈Sα\{0}
r(hα(m)) > r(C\hλ(0)) − δh,
(4.1.3)
(4.1.4)
noting that the right hand side of (4.1.4) is non-negative by (λ,1.3).
Set
(cid:33)
(cid:32)
ε :=
1
2
∑
m∈Sα\{0}
r(hα(m)) − (r(C\hλ(0)) − δh)
> 0.
44
CHAPTER 4: UNIFORM ALGEBRAS
Then there exists n ∈ Sα\{0} such that for Sαn
1 := {m ∈ Sα\{0} : m ≤ n} we have
r(hα(m)) > r(C\hλ(0)) − δh + ε > 0.
(4.1.5)
1
∑
m∈Sαn
Further for each m ∈ Sαn
each m ∈ Sαn
Sαn
maximum over a finite set of elements since Sαn
max{λ, λ
∑
(cid:48)} ≤ γ < α we have,
r(hγ(m)) ≥ ∑
1, k (cid:54)= 0 by (4.1.5). Let λ
r(hγ(m))
(cid:48)
m∈Sγ\{0}
m∈Sα\{0}
1 there exists βm < α such that r(hβm (m)) ≥ r(hα(m)) − 1
1 we have, by (iii), r(hα(m)) = supβ<α r(hβ(m)). Hence for
2k ε, for k :=
1} < α and note that this is a
1 ⊆ N is finite. Now for any γ with
:= max{βm : m ∈ Sαn
(since Sα ⊆ Sγ)
1
≥ ∑
m∈Sαn
≥ ∑
m∈Sαn
≥ ∑
m∈Sαn
1
1
r(hγ(m))
r(hβm (m))
(by (γ,4))
(r(hα(m)) − ε
2k )
(by the above)
> r(C\hλ(0)) − δh + ε − k ε
2k
> r(C\hλ(0)) − δh
≥ r(C\hγ(0)) − δh
(by (4.1.5) and k := Sαn
1)
(by (ii)).
This contradicts (γ,1.3). Hence we have shown that, for all λ < α, (4.1.3) holds.
Now by (iv) we have r(C\hα(0)) = infλ<α r(C\hλ(0)).
Hence we have ∑m∈Sα\{0} r(hα(m)) ≤ r(C\hα(0)) − δh and so (α,1.3) holds.
i∈Sβ hβ(i) ⊆(cid:83)
(cid:83)
We now show that (α,3) holds. We will show that for all ordinals β < α,
i∈Sα hα(i). Let β < α and z ∈(cid:83)
i∈Sβ hβ(i). Define,
m := min{i ∈ N0 : there exists λ < α with i ∈ Sλ and z ∈ hλ(i)}.
By the definition of m there exists ζ < α with m ∈ Sζ and z ∈ hζ(m). We claim that the
set {λ < α : m (cid:54)∈ Sλ} is empty. To prove this suppose towards a contradiction that we
can define,
(cid:48)
λ
:= min{λ < α : m (cid:54)∈ Sλ}.
(cid:84)
Then λ
> 0 since, by (ζ,2), Sζ ⊆ S0 with m ∈ Sζ. If λ
(cid:48)
(cid:48) Sγ giving m (cid:54)∈ Sγ, for some γ < λ
= f (hλ
γ<λ
(cid:48)
is a successor ordinal then hλ
λ
(cid:48)
(cid:48)
m (cid:54)∈ Sλ
(cid:48) and Definition 4.1.14 of f : H → H, hλ
and (4.1.2) of Definition 4.1.14 there is n ∈ Sλ
Further for all λ with λ
(cid:48) ≤ λ < α we have m (cid:54)∈ Sλ since m (cid:54)∈ Sλ
(cid:48)
is a limit ordinal then m (cid:54)∈ Sλ
, and this contradicts the definition of λ
(cid:48)−1) with m ∈ Sλ
(cid:48)
(cid:48)−1 by the definition of λ
(cid:48)−1(m) ⊆ hλ
(cid:48) =
. If
. By
(cid:48)−1 is not classical. Therefore by (4.1.1)
(n).
(cid:48) with n < m and hλ
(cid:48) and, by (λ,2), Sλ ⊆ Sλ
(cid:48) .
(cid:48)
(cid:48)
45
CHAPTER 4: UNIFORM ALGEBRAS
(cid:48)
(cid:48)
(cid:48) − 1, 4), {hγ(m) : γ ≤ λ
(cid:48)−1(m) ⊆ hλ
(n) with n ∈ Sλ
λ<α hλ(m) = hα(m) ⊆ (cid:83)
(cid:48) − 1} is a nested increasing
Hence we have ζ < λ
. Now, by (λ
family of sets giving z ∈ hζ(m) ⊆ hλ
(cid:48) . This contradicts the
giving m ∈ Sα = (cid:84)
definition of m since n < m. Hence we have shown that {λ < α : m (cid:54)∈ Sλ} is empty
have z ∈ hζ(m) ⊆ (cid:83)
λ<α Sλ. Therefore, by Definition 4.1.16 and the definition of ζ, we
i∈Sα hα(i) as required. Hence (α,3) holds.
Therefore we have shown, by the principal of transfinite induction, that P(α) holds and
this concludes the proof of Lemma 4.1.17.
Recall that our aim is to prove that for every h ∈ H there is a classical disc assignment
(cid:48) ∈ H with Xh(cid:48) ⊆ Xh and δh(cid:48) ≥ δh. We have the following closing argument
function h
using cardinality. By (α,2) of Lemma 4.1.17 we obtain a nested ordinal sequence of
domains (Sα),
N0 ⊇ S ⊇ S1 ⊇ S2 ⊇ · · · ⊇ Sω ⊇ Sω+1 ⊇ · · · ⊇ {0}.
Now setting Sc
∅ ⊆ Sc ⊆ Sc
Lemma 4.1.18. For the disc assignment function hβ we have,
hβ is classical if and only if (Sα) has stabilised at β, i.e. Sβ+1 = Sβ.
α := N0\Sα gives a nested ordinal sequence (Sc
α),
ω+1 ⊆ · · · ⊆ N.
1 ⊆ Sc
2 ⊆ · · · ⊆ Sc
ω ⊆ Sc
Proof. The proof follows directly from (ii) of Lemma 4.1.15.
β+1 but nβ+1 (cid:54)∈ Sc
Now let ω1 be the first uncountable ordinal. Suppose towards a contradiction that,
for all β < ω1, (Sα) has not stabilised at β. Then for each β < ω1 there exists some
nβ+1 ∈ N such that nβ+1 ∈ Sc
α for all α ≤ β. Hence since there are
ω1 ⊆ N, a contradiction.
uncountably many β < ω1 we have Sc
Therefore there exists β < ω1 such that (Sα) has stabilised at β and so, by Lemma 4.1.18,
hβ is classical. Now by (β,1) of Lemma 4.1.17 we have hβ ∈ H with δhβ ≥ δh and by
(β,3) we have Xhβ ⊆ Xh. In particular this completes the proof of Theorem 4.1.9 and
the Feinstein-Heath Swiss cheese "Classicalisation" theorem.
ω1 uncountable with Sc
The proof of Theorem 4.1.8 as presented above proceeded without reference to a theory
of allocation maps.
In the original proof of Feinstein and Heath, [FH10], allocation
maps play a central role. We will recover a key allocation map from the original proof
using the map f : H → H of Definition 4.1.14. Here is the definition of an allocation
map as it appears in [FH10].
Definition 4.1.19. Let D = (∆,D) be a Swiss cheese. We define
Now let E = (E,E ) be a second Swiss cheese, and let f : (cid:101)D → (cid:101)E. We define G( f ) =
f −1(C\E) ∩ D. We say that f is an allocation map if the following hold:
(cid:101)D = D ∪ {C\∆}.
46
CHAPTER 4: UNIFORM ALGEBRAS
(A1) for each U ∈ (cid:101)D, U ⊆ f (U);
(A2)
(A3) for each E ∈ E,
r(D) ≥ r(∆) − r(E);
∑
D∈G( f )
∑
D∈ f −1(E)
r(D) ≥ r(E).
maps defined on (cid:101)D. In [FH10] a partial order is applied to S(D) and subsequently a
Let D be the Swiss cheese of Theorem 4.1.8 and let S(D) be the family of allocation
(cid:48)
maximal element fmax is obtained using Zorn's lemma. The connection between al-
location maps and Swiss cheeses is then exploited. Towards a contradiction the non-
existence of the desired classical Swiss cheese D
of Theorem 4.1.8 is assumed. This
assumption implies the existence of an allocation map f (cid:48) ∈ S(D) that is higher in the
partial order applied to S(D) than fmax, a contradiction. The result follows. It is at the
last stage of the original proof where a connection to the new version can be found. In
the construction of Feinstein and Heath the allocation map f (cid:48) factorizes as f (cid:48) = g ◦ fmax
where g is also an allocation map. Let E = (E,E ) be a non-classical Swiss cheese with
g, an allocation map gE defined on(cid:101)E can be obtained without contradiction. Clearly
δ(E) > 0. Using the same method of construction that Feinstein and Heath use for
(cid:101)E (cid:54)= fmax((cid:101)D). We will obtain gE using the map f
: H → H of Definition 4.1.14. Let
h ∈ H, h : S → O, be an injective disc assignment function such that Dh = E and recall
(cid:48) → O has S
= S\{m} where (n, m) is the minimum
from Definition 4.1.14 that f (h) : S
element of Ih. Set E(cid:48) := D f (h). By Definitions 4.1.7 and 4.1.19 we have
Now define a map ι :(cid:101)E → S
for U ∈(cid:101)E,
(cid:101)E = (cid:102)Dh = h(S) and (cid:101)E(cid:48) = (cid:93)D f (h) = f (h)(S
ι(U) :=
by,
).
(cid:48)
(cid:48)
(cid:48)
h−1(U)
(cid:101)E(cid:48)
(cid:101)E
gE
n
h−1(U) (cid:54)= m
h−1(U) = m
,
if
if
and note that this is well defined since h is injective. The commutative diagram in Fig-
ι
(cid:48)
S
f (h)
Figure 4.3: gE = f (h) ◦ ι.
ure 4.3 show how gE is obtained using f : H → H. The construction of f in Definition
47
/
/
?
?
CHAPTER 4: UNIFORM ALGEBRAS
4.1.14 was developed from the construction that Feinstein and Heath used for g. The
method of combining discs in Lemma 4.1.10 also appears in [Zha93].
Remark 4.1.20. Concerning classicalisation.
(i) Interestingly, as Heath points out in [FH10], every classical Swiss cheese set in
C with empty interior is homeomorphic to the Sierpi ´nski carpet. Hence up to
homeomorphism there is only 1 Swiss cheese set of this type.
In particular if
XD is a Swiss cheese contained in C with empty interior then either one of the
conditions XD is classical or δ(D) > 0 is enough for R(XD) (cid:54)= A(XD).
(ii) I also anticipate the possibility of an analog of Theorem 4.1.9 on the sphere. Let
S ⊆ R3 be a sphere of finite positive radius rs and center c ∈ R3. For a, b ∈ S let
ds(a, b) := rs∠acb be the length of the geodesic path in S from a to b. Now ds is a
metric with respect to which we will define open and closed S-discs contained in
S. With analogy to Definition 4.1.4 let Ds := (∆,D) be a Swiss cheese on S. Then
s := (S,D ∪ {S\∆}) is a Swiss cheese on S, since S\∆ is an open
either ∆ = S or D
S-disc, for which XD
= XDs in S. Further we have
(cid:48)
(cid:48)
s
δ(Ds) := r(∆) − ∑
D∈D
r(D) = πrs − r(S\∆) − ∑
D∈D
r(D) = δ(D
(cid:48)
s)
and so for our choice of metric on S we note that δ is independent of whether we
(cid:48)
use Ds or D
s. Hence the situation for the sphere is a little simpler than that for the
plane since we can allow all Swiss cheeses on the sphere to have the form Ds :=
(S,D) and avoid the need to handle a special closed S-disc ∆. Therefore on the
sphere analogs of lemmas 4.1.11 and 4.1.13 are not required. We will not establish
here whether the condition δ(Ds) > 0 is sufficient for the analog of Theorem 4.1.9
on S to hold since the next step in generalising this theorem should be to establish
the class of all metric spaces for which a general version of the theorem holds.
However the sphere is of particular interest in the context of uniform algebras
since, less one point, the sphere is homeomorphic to the plane allowing many
examples of uniform algebras to be defined on subsets of the sphere.
4.2 Non-complex analogs of uniform algebras
The most obvious non-complex analog of Definition 4.1.1 is obtained by simply replac-
ing the complex numbers in the definition by some other complete valued field F. In
this case, whilst CF(X) will be complete and contain the constants, we need to take
care concerning the topology on X when F is non-Archimedean, e.g. CQp ([0, 1]) only
contains the constant.
48
CHAPTER 4: UNIFORM ALGEBRAS
Theorem 4.2.1. Let F be a complete non-Archimedean valued field and let CF(X) be the unital
Banach F-algebra of all continuous F-valued functions defined on a compact, Hausdorff space
X. Then CF(X) separates the points of X if and only if X is totally disconnected.
Before giving a proof of Theorem 4.2.1 we have the following version of Urysohn's
lemma which will certainly already be known in some form because of its simplicity.
Lemma 4.2.2. Let X be a totally disconnected, compact, Hausdorff space with finite subset
{x, y1, y2, y3,· · · , yn} ⊆ X, x (cid:54)= yi for all i ∈ {1,· · · , n} where n ∈ N. Let L be any non-
empty topological space and a, b ∈ L. Then there exists a continuous map h : X −→ L such
that h(x) = a and h(y1) = h(y2) = h(y3) = · · · = h(yn) = b.
subsets Ui and Vi of X with x ∈ Ui and yi ∈ Vi. Hence U := (cid:84)
Proof. Since X is a Hausdorff space, for each i ∈ {1,· · · , n} there are disjoint open
i∈{1,··· ,n} Ui is an open
subset of X with x ∈ U and U ∩ Vi = ∅ for all i ∈ {1,· · · , n}. Now since X is a totally
disconnected, compact, Hausdorff space, x has a neighborhood base of clopen sets, see
[Wil04, Theorem 29.7] noting that X is locally compact by Theorem 2.1.22. Hence there
is a clopen subset W of X with x ∈ W ⊆ U. The function h : X −→ L given by
h(W) := {a} and h(X\W) := {b} is continuous as required.
We now give the proof of Theorem 4.2.1.
Proof. With reference to Lemma 4.2.2 it remains to show that CF(X) separates the points
of X only if X is totally disconnected. Let X be a compact, Hausdorff space such that
CF(X) separates the points of X. Let U be a non-empty connected subset of X and
let f ∈ CF(X). We note that f (U) is a connected subset of F since f is continuous.
Now, since F is non-Archimedean it is totally disconnected i.e. its connected subsets
are singletons. Hence f (U) is a singleton and so f is constant on U. Therefore, since
CF(X) separates the points of X, U is a singleton and X is totally disconnected.
We next consider the constraints on CF(X) revealed by generalisations of the Stone-
Weierstrass theorem. In the real case the Stone-Weierstrass theorem for CR(X) says
that for every compact Hausdorff space X, CR(X) is without a proper subalgebra that
is uniformly closed, contains the real numbers and separates the points of X. A proof
can be found in [KL92, p50]. The non-Archimedean case is given by a theorem of
Kaplansky, see [Ber90, p157] or [Kap50].
Theorem 4.2.3. Let F be a complete non-Archimedean valued field, let X be a totally dis-
connected compact Hausdorff space, and let A be a F-subalgebra of CF(X) which satisfies the
following conditions:
49
CHAPTER 4: UNIFORM ALGEBRAS
(i) the elements of A separate the points of X;
(ii) for each x ∈ X there exists f ∈ A with f (x) (cid:54)= 0.
Then A is everywhere dense in CF(X).
Note that, in Theorem 4.2.3, A being a F-subalgebra of CF(X) means that A is a sub-
algebra of CF(X) and a vector space over F. If we take A to be unital then condition
(ii) in Theorem 4.2.3 is automatically satisfied and the theorem is analogous to the real
version of the Stone-Weierstrass theorem. In subsection 4.2.1 we will see that real func-
tion algebras are a useful example when considering non-complex analogs of uniform
algebras with qualifying subalgebras.
4.2.1 Real function algebras
Real function algebras were introduced by Kulkarni and Limaye in a paper from 1981,
see [KL81]. For a thorough text on the theory see [KL92]. The following definition is a
little more general than what we need in this subsection.
Definition 4.2.4. Let X be a topological space and let τ : X → X be a homeomorphism.
(i) We will call τ a topological involution on X if τ ◦ τ = id on X.
(ii) We will call τ a topological element of finite order on X if τ has finite order but with
ord(τ) > 2.
Let F and L be complete valued fields such that L is a finite extension of F as a valued
field and let g ∈ Gal(L/F). Let A either be an F-algebra or an L-algebra for which
σ : A → A is a map satisfying ord(σ) = ord(g) and for all a, b ∈ A and scalars α:
σ(a + b) = σ(a) + σ(b);
σ(ab) = σ(b)σ(a);
σ(αa) = g(α)σ(a).
(iii) We will call σ a algebraic involution on A if σ ◦ σ = id on A.
(iv) We will call σ a algebraic element of finite order on A if σ has ord(σ) > 2.
Note, in Definition 4.2.4 the requirement that τ be a homeomorphism is satisfied if τ is
continuous.
50
CHAPTER 4: UNIFORM ALGEBRAS
Definition 4.2.5. Let X be a compact Hausdorff space and τ a topological involution
on X. A real function algebra on (X, τ) is a real subalgebra A of
C(X, τ) := { f ∈ CC(X) : f (τ(x)) = ¯f (x) for all x ∈ X}
that is complete with respect to the sup norm, contains the real numbers and separates
the points of X.
Remark 4.2.6. Concerning real function algebras.
(i) Later, Theorem 5.2.1 will confirm that C(X, τ) in Definition 4.2.5 is itself always a
real function algebra on (X, τ) and in some sense it is to real function algebras as
CC(X) is to complex uniform algebras.
(ii) Let X be a compact Hausdorff space and Y a closed non-empty subset of X. Then
CY := { f ∈ CC(X) :
f (Y) ⊆ R} is a commutative real Banach algebra. As
pointed out in [KL92, p2], every such CY can be transformed into a real function
algebra but the converse of this is false. Hence Definition 4.2.5 is a more general
object.
(iii) With reference to Definition 4.2.5 we have C(X, τ) = { f ∈ CC(X) : σ( f ) = f}
where σ( f ) := ¯f ◦ τ. Moreover σ is an algebraic involution on CC(X) and every
algebraic involution on CC(X) arises from a topological involution on X in this
way, see [KL92, p29] for a proof.
The following example is a useful standard.
Example 4.2.7. Recall from Example 4.1.3 the disc algebra A(∆) on the closed unit disc
and let τ : ∆ −→ ∆ be the map τ(z) := ¯z given by complex conjugation, which we note
is a Galois automorphism on C. Now let
B(∆) := A(∆) ∩ C(∆, τ).
We see that B(∆) is complete since both A(∆) and C(∆, τ) are, and similarly B(∆)
contains the real numbers. Further by the definition of C(∆, τ) and the fact that A(∆) =
P(∆) we have that B(∆) is the R-algebra of all uniform limits of polynomials on ∆ with
real coefficients. Hence B(∆) separates the points of ∆ since it contains the function
f (z) := z. However whilst τ is in C(∆, τ) it is not an element of A(∆). Therefore B(∆)
is a real function algebra on (∆, τ) and a proper subalgebra of C(∆, τ). It is referred to
as the real disc algebra.
Finally for each compact Hausdorff space X, CC(X) can be put into the form of a real
function algebra as the following example shows. In particular C can be expressed as
a real function algebra on a two point set.
51
CHAPTER 4: UNIFORM ALGEBRAS
Example 4.2.8. Let X be a compact Hausdorff space, let Y := {i,−i} ⊆ C have the
trivial topology and give X × Y the product topology. We note that the subspace given
:= {(x, y) ∈ X × Y : y = i} is homeomorphic to X and similarly so is X−i.
by Xi
Define a topological involution τ : X × Y → X × Y by (x, y) (cid:55)→ (x, ¯y). Then CC(Xi) is
isometrically isomorphic to C(X × Y, τ) by way of the mapping f (cid:55)→ h f where
f (z)
¯f (τ(z))
h f (z) :=
so that for z ∈ Xi we have
if
if
z ∈ Xi
z ∈ X−i
,
for f ∈ CC(Xi),
and for z ∈ X−i
h f (τ(z)) = ¯f (τ(τ(z))) = ¯f (z) = ¯h f (z)
h f (τ(z)) = f (τ(z)) = ¯¯f (τ(z)) = ¯h f (z)
showing that h f ∈ C(X × Y, τ). The inverse mapping from C(X × Y, τ) to CC(Xi) is
given by the restriction map h (cid:55)→ hXi. One might suspect that such a mapping exists
for every C(Z, τ) by restricting its elements to a compact subspace of equivalence class
representatives for the forward orbits of τ. But this is not the case in general since
there can be z ∈ Z with ord(τ, z) = 1 forcing all of the functions to be real valued at z
preventing the representation of the complex constants in C(Z, τ).
52
CHAPTER 5
Commutative generalisation over
complete valued fields
If J is a maximal ideal of a commutative unital complex Banach algebra A then J has
codimension one since A/J with the quotient norm is isometrically isomorphic to the
complex numbers. This follows from the Gelfand-Mazur theorem noting that A/J with
the quotient norm is unital since J is closed as a subset of A and J is different to A, see
Lemma 6.1.7 and [Sto71, p16].
In contrast, for a complete non-Archimedean field F, if I is a maximal ideal of a com-
mutative unital Banach F-algebra then I may have large finite or infinite codimension,
note Corollary 2.2.4. Hence, with Gelfand transform theory in mind, it makes sense
to consider non-Archimedean analogs of uniform algebras in the form suggested by
real function algebras where the functions take values in a complete extension of the
underlying field of scalars. Moreover when there is a lattice of intermediate fields then
these fields provide a way for a lattice of extensions of the algebra to occur. See [Ber90,
Ch1] and [Esc03, Ch15] for one form of the Gelfand transform in the non-Archimedean
setting. This chapter introduces the main definitions of interest in the thesis. We will
generalise the definitions made by Kulkarni and Limaye to all complete valued fields,
show that the algebras obtained all qualify as generalisations of uniform algebras and
that restricting attention to the Archimedean setting recovers the complex uniform al-
gebras and real function algebras. Non-Archimedean examples and residue algebras
are also introduced.
5.1 Main definitions
The following definition gives the requirements for those commutative algebras that
are to be considered as generalisations of uniform algebras.
53
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
Definition 5.1.1. Let F and L be complete valued fields such that L is an extension of
F as a valued field. Let X be a compact Hausdorff space and let CL(X) be the unital
Banach L-algebra of all continuous L-valued functions on X with pointwise operations
and the sup norm. If a subset A of CL(X) satisfies:
(i) A is an F-algebra under pointwise operations;
(ii) A is complete with respect to (cid:107) · (cid:107)∞;
(iii) F ⊆ A;
(iv) A separates the points of X,
then we will call A an L/F uniform algebra or just a uniform algebra when convenient.
In the language of Definition 5.1.1, an L/F uniform algebra is a Banach F-algebra of
L-valued functions, also every L/L uniform algebra is an L/F uniform algebra. We now
generalise, in two parts, Kulkarni and Limaye's definition of a real function algebra.
Definition 5.1.2. Let F and L be complete valued fields such that L is a finite extension
of F as a valued field. Let X be a compact Hausdorff space and totally disconnected if
F is non-Archimedean. Define C(X, τ, g) ⊆ CL(X) as the subset of elements f ∈ CL(X)
for which the diagram in Figure 5.1 commutes.
X
τ
X
f
f
L
/ L
(i) g ∈ Gal(L/F);
g
Where:
(ii) τ : X → X with ord(τ)ord(g);
(iii) g and τ are continuous.
Figure 5.1: Commutative diagram for f ∈ C(X, τ, g).
We will call C(X, τ, g) := { f ∈ CL(X) : f (τ(x)) = g( f (x)) for all x ∈ X} the basic L/Lg
function algebra on (X, τ, g) or just a basic function algebra when convenient.
Definition 5.1.3. Let F and L be complete valued fields such that L is a finite extension
of F as a valued field. Let (X, τ, g) conform to the conditions of Definition 5.1.2 and let
A be a subset of the basic L/Lg function algebra on (X, τ, g). If A is also an L/Lg uniform
algebra then we will call A an L/Lg function algebra on (X, τ, g).
Remark 5.1.4. In definitions 5.1.2 and 5.1.3 the valued field Lg is complete by Remark
2.2.2. The continuity of g in Definition 5.1.2 is only an observation since g is an isometry
on L by Remark 2.2.6. In fact g also acts as an isometric automorphism on C(X, τ, g).
54
/
/
/
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
5.2 Generalisation theorems
With Definition 5.1.3 in mind the following theorem, which is the main theorem of this
chapter, clarifies why an algebra conforming to the conditions of Definition 5.1.2 is to
be called a basic L/Lg function algebra on (X, τ, g).
Theorem 5.2.1. Let (X, τ, g) conform to the conditions of Definition 5.1.2. Then the basic
L/Lg function algebra on (X, τ, g) is always an L/Lg uniform algebra.
Remark 5.2.2. We will see in the proof of Theorem 5.2.1 that ord(τ)ord(g) is an op-
timum condition in Definition 5.1.2 since if we do not include it in the definition then
C(X, τ, g) separates the points of X if and only if ord(τ)ord(g) as per Figure 5.2.
ord(τ)ord(g)
1
3
ord(τ, X) ⊆ ord(g, L)
2
3 C(X, τ, g) separates X
Figure 5.2: Equivalence diagram for Definition 5.1.2.
Proof of Theorem 5.2.1. Let (X, τ, g) conform to the conditions of Definition 5.1.2. It is
immediate that C(X, τ, g) is a ring under pointwise operations and Lg ⊆ C(X, τ, g).
We now show that C(X, τ, g) is complete with respect to the sup norm. First note that
C(X, τ, g) = { f ∈ CL(X) : σ( f ) = f}
where σ( f ) := g(ord(g)−1) ◦ f ◦ τ is an isometry on CL(X) since τ is surjective and g
is an isometry on L by Remark 2.2.6. Further σ is either an algebraic involution or
a algebraic element of finite order on CL(X). Hence since CL(X) is commutative σ
is in fact an isometric automorphism on CL(X). Now let ( fn) be a Cauchy sequence
in C(X, τ, g) and let f be its limit in CL(X). Then for each ε > 0 there exists N ∈
N such that for all n > N we have (cid:107) f − fn(cid:107)∞ = (cid:107)σ( fn) − σ( f )(cid:107)∞ < ε
2. But then
(cid:107) f − σ( f )(cid:107)∞ = (cid:107) f − σ( fn) + σ( fn) − σ( f )(cid:107)∞ ≤ (cid:107) f − fn(cid:107)∞ + (cid:107)σ( fn) − σ( f )(cid:107)∞ < ε.
Hence (cid:107) f − σ( f )(cid:107)∞ = 0 giving σ( f ) = f and so f ∈ C(X, τ, g). Hence C(X, τ, g) is
complete. It remains to show that C(X, τ, g) separates the points of X and to this end
we now show each of the implications in Figure 5.2.
1: Let n ∈ ord(τ, X). It is immediate that nord(τ) and since ord(τ)ord(g) we have
n ∈ ord(g, L) by Lemma 2.2.18. We also note that the converse is immediate since for
each n ∈ ord(g, L) we have nord(g) and so ord(τ)ord(g).
55
q
y
+
e
m
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
2: Note that ord(σ) = ord(g) and so, like a norm map, for every h ∈ CL(X) we have
hσ(h)σ(2)(h) · · · σ(ord(g)−1)(h) ∈ C(X, τ, g)
h + σ(h) + σ(2)(h) + · · · + σ(ord(g)−1)(h) ∈ C(X, τ, g).
and
Now if g = id is the identity then C(X, τ, g) = CL(X) which separates the points of X
when L is Archimedean by Urysohn's lemma, since X is locally compact, and when L
is non-Archimedean by Theorem 4.2.1 since we required X to be totally disconnected
in this case. So now suppose ord(g) > 1 and let x, y ∈ X with x (cid:54)= y. We need to check
two cases.
Case 1: In this case y (cid:54)= τ(n)(x) for all n ∈ N with n ≤ ord(τ, x). By Urysohn's lemma in
the Archimedean setting or Lemma 4.2.2 otherwise, there is h ∈ CL(X) with h(y) = 0
and h(τ(n)(x)) = 1 for all n ∈ N0. Let f
:= hσ(h)σ(2)(h) · · · σ(ord(g)−1)(h) so that
f ∈ C(X, τ, g) with f (y) = 0, by construction, and f (x) = 1 by construction given that
g(1) = 1. Then in this case x and y are separated by f .
Case 2: In this case y = τ(n)(x) for some n ∈ N with n < ord(τ, x). Let m := ord(g) and
k := ord(τ, x) and note therefore that we have 1 ≤ n ≤ k − 1, since y (cid:54)= x, and m = km(cid:48)
for some m(cid:48) ∈ N. Further since ord(τ, X) ⊆ ord(g, L) there is a ∈ L with ord(g, a) = k.
By Urysohn's lemma in the Archimedean setting or Lemma 4.2.2 otherwise, there is
h ∈ CL(X) with h(x) = a and h(τ(i)(x)) = 0 for 1 ≤ i ≤ k − 1. We will now check two
sub-cases.
Case 2.1: The characteristic of the field L is zero, i.e. char(L) = 0.
Let f := h + σ(h) + σ(2)(h) + · · · + σ(ord(g)−1)(h) so that we have f ∈ C(X, τ, g) with
f = h + g(m−1) ◦ h ◦ τ + g(m−2) ◦ h ◦ τ(2) + · · · + g ◦ h ◦ τ(m−1). This gives
f (x) =h(x) + g(m−k) ◦ h(τ(k)(x)) + g(m−2k) ◦ h(τ(2k)(x)) + · · ·
· · · + g(m−(m(cid:48)−1)k) ◦ h(τ((m(cid:48)−1)k)(x))
=h(x) + g((m(cid:48)−1)k) ◦ h(τ(k)(x)) + g((m(cid:48)−2)k) ◦ h(τ(2k)(x)) + · · ·
· · · + g(k) ◦ h(τ((m(cid:48)−1)k)(x))
=a + a + a + · · · + a, m(cid:48) times,
=m(cid:48)a
and
f (y) = f (τ(n)(x))
=g(m−(k−n)) ◦ h(τ(k)(x)) + g(m−(2k−n)) ◦ h(τ(2k)(x)) + · · ·
· · · + g(m−((m(cid:48)−1)k−n)) ◦ h(τ((m(cid:48)−1)k)(x)) + g(m−(m(cid:48)k−n)) ◦ h(τ(m(cid:48)k)(x))
=g((m(cid:48)−1)k+n) ◦ h(τ(k)(x)) + g((m(cid:48)−2)k+n) ◦ h(τ(2k)(x)) + · · ·
· · · + g(k+n) ◦ h(τ((m(cid:48)−1)k)(x)) + g(n) ◦ h(τ(m)(x))
=m(cid:48)g(n)(a) with 1 ≤ n ≤ k − 1.
56
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
Hence since ord(g, a) = k we have f (x) (cid:54)= f (y).
Case 2.2: The characteristic of L is p, i.e. char(L) = p, for some prime p ∈ N. In this
case the proof for Case 2.1 breaks down when m(cid:48) = pts with s, t ∈ N, p (cid:45) s. So with
respect to such circumstances define f (cid:48) := hσ(sk)(h)σ(2sk)(h) · · · σ((pt−1)sk)(h) and
f := f (cid:48)
+ σ( f (cid:48)
) + σ(2)( f (cid:48)
) + · · · + σ(sk−1)( f (cid:48)
).
We will now show that σ( f ) = f so that f ∈ C(X, τ, g) and note that this is satisfied if
σ(sk)( f (cid:48)) = f (cid:48) since σ( f ) = σ( f (cid:48)) + σ(2)( f (cid:48)) + σ(3)( f (cid:48)) + · · · + σ(sk)( f (cid:48)). Indeed we have
that σ(sk)( f (cid:48)) = σ(sk)(h)σ(2sk)(h)σ(3sk)(h) · · · σ(ptsk)(h) with σ(ptsk)(h) = σ(m)(h) = h so
that σ(sk)( f (cid:48)) = f (cid:48) giving f ∈ C(X, τ, g). Now, for 0 ≤ i ≤ sk − 1, we have
σ(i)( f (cid:48)
)(x) =σ(i)(h)(x)σ(sk+i)(h)(x)σ(2sk+i)(h)(x) · · · σ((pt−1)sk+i)(h)(x)
apt
0
=
ki
k (cid:45) i
if
if
since σ(kj)(h)(x) = g(m−kj) ◦ h(τ(kj)(x)) = g((pts−j)k) ◦ h(x) = a for kj < m, j ∈ N0, and
σ(j)(h)(x) = g(m−j) ◦ h(τ(j)(x)) = g(m−j)(0) = 0 for j < m, k (cid:45) j. Hence
(x) + σ( f (cid:48)
(x) + σ(k)( f (cid:48)
f (x) = f (cid:48)
= f (cid:48)
=sapt.
)(x) + σ(2)( f (cid:48)
)(x) + σ(2k)( f (cid:48)
)(x) + · · · + σ(sk−1)( f (cid:48)
)(x)
)(x) + · · · + σ((s−1)k)( f (cid:48)
)(x)
But for 0 ≤ i ≤ sk − 1 we also have
σ(i)( f (cid:48)
)(y) =σ(i)(h)(y)σ(sk+i)(h)(y)σ(2sk+i)(h)(y) · · · σ((pt−1)sk+i)(h)(y)
g(n)(a)pt
0
=
k(i + n)
k (cid:45) (i + n)
if
if
since if k(i + n) then i has the form kj − n and the exponents of σ therefore also have
the form kj − n < m, j ∈ N, giving
σ(kj−n)(h)(y) =g(m−(kj−n)) ◦ h(τ(kj−n)(y))
=g((pts−j)k+n) ◦ h(τ(kj−n)(τ(n)(x)))
=g((pts−j)k+n) ◦ h(x) = g(n)(a)
and if k (cid:45) (i + n) then for j < m an exponents of σ we also have k (cid:45) (j + n) giving
σ(j)(h)(y) =g(m−j) ◦ h(τ(j)(y))
=g(m−j) ◦ h(τ(j+n)(x))
=g(m−j)(0) = 0.
57
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
Hence
f (y) = f (cid:48)
(y) + σ( f (cid:48)
)(y) + σ(2)( f (cid:48)
)(y) + σ(2k−n)( f (cid:48)
)(y) + · · · + σ(sk−1)( f (cid:48)
)(y) + · · · + σ(sk−n)( f (cid:48)
)(y)
)(y)
=σ(k−n)( f (cid:48)
=sg(n)(a)pt.
Now since p (cid:45) s we have s ∈ L×. Furthermore recall that since ord(g, a) = k and
(cid:54)= apt
1 ≤ n ≤ k − 1 we have g(n)(a) (cid:54)= a. Therefore it remains to show that g(n)(a)pt
in order to conclude that f (y) (cid:54)= f (x). Recall that p = char(L) ∈ N is a prime. For
b ∈ L the Frobenius Frob : L → L, Frob(b) := bp, is an injective endomorphism on L.
We show that the Frobenius is injective on L. Let b1, b2 ∈ L with bp
The case p > 2 gives (b1 − b2)p = bp
The case p = 2 gives (b1 − b2)2 = b2
In each case L is an integral domain and so b1 − b2 = 0 giving b1 = b2 as required.
(cid:54)= apt since
Therefore Frob(t) : L → L, Frob(t)(b) := bpt, is also injective giving g(n)(a)pt
g(n)(a) (cid:54)= a and this finishes the proof of implications 2 in Figure 5.2.
3: We now show implication 3 in Figure 5.2 by showing the contrapositive. Suppose
ord(τ) (cid:45) ord(g). Then there exists some x ∈ X such that τ(ord(g))(x) (cid:54)= x. Let y :=
τ(ord(g))(x). Now for all f ∈ C(X, τ, g) we have for all i ∈ N that
1 − bp
1 + b2
2 = 0.
2 = 2b2
1 = bp
2.
1 = 0.
f ◦ τ(i) = f ◦ τ ◦ τ(i−1) = g ◦ f ◦ τ(i−1) = · · · = g(i) ◦ f .
Therefore f (y) = f ◦ τ(ord(g))(x) = g(ord(g)) ◦ f (x) = f (x). Hence for all f ∈
C(X, τ, g) we have f (x) = f (y) as required.
Hence having shown each of the implications in Figure 5.2, ord(τ)ord(g) is a neces-
sary and sufficient condition in Definition 5.1.2 in order that C(X, τ, g) separates the
points of X and this completes the proof of Theorem 5.2.1.
Remark 5.2.3. It's worth noting that for char(L) = p the Frobenius is also a endomor-
phism on C(X, τ, g). Moreover for L of any characteristic we have seen in the proof of
Theorem 5.2.1 that σ, given by σ( f ) := g(ord(g)−1) ◦ f ◦ τ, is an isometric automorphism
on CL(X) with fixed elements C(X, τ, g).
With reference to the complex, real and non-Archimedean Stone-Weierstrass theorems
from Chapter 4, the following combined Stone-Weierstrass theorem is immediate.
Theorem 5.2.4. Let L be a complete valued field. Let X conform to Definition 5.1.2 and let A
be an L/L function algebra on (X,id,id). Then either A = CL(X) or L = C and A is not self
adjoint, that is there is f ∈ A with ¯f /∈ A.
58
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
5.3 Examples
Our first example considers L/Lg function algebras in the Archimedean setting.
Example 5.3.1. Let F = R, L = C and X be a compact Hausdorff space. We have
Gal(C/R) = (cid:104)id, ¯z(cid:105).
Setting g = id in Definition 5.1.2 forces τ to be the identity on X.
In this case it's
immediate that C(X, τ, g) = CC(X) and each L/Lg function algebra on (X, τ, g) is a
complex uniform algebra.
On the other hand, setting g = ¯z forces τ to be a topological involution on X. In this
case the L/Lg function algebras on (X, τ, g) are precisely the real function algebras of
Kulkarni and Limaye.
Our first non-Archimedean example is very straightforward involving the trivial valu-
ation.
Example 5.3.2. Let F = Q, but with the trivial valuation instead of the absolute val-
uation, and let L = Q(a) with the trivial valuation where a = exp( 1
102πi). With ref-
erence to Theorem 2.2.7, and having factorised x10 − 1 in F[x], we have IrrF,a(x) =
x4 − x3 + x2 − x + 1 which gives [L, F] = degIrrF,a(x) = 4. The roots of IrrF,a(x) are
the elements of S := {a, a3, a7, a9} and so, with reference to Definition 2.2.8, L is a
normal extension of F. Moreover with reference to Remark 2.2.9 L is a separable ex-
tension of F and so L is also a Galois extension of F with #Gal(L/F) = [L, F] = 4 by
Theorem 2.2.11. Now for g ∈ Gal(L/F) we must have g : S → S since, for b ∈ S,
0 = g(0) = g(IrrF,a(b)) = IrrF,a(g(b)). Putting g(a) := a3 makes g a generator of
Gal(L/F) and we have
g(a) = a3,
g(2)(a) = a9,
g(3)(a) = a7 and g(4)(a) = a.
Hence L is a cyclic extension of F meaning that Gal(L/F) is a cyclic group. Moreover
√
giving a + a9 − a3 − a7 =
√
g(
5 noting that the real part of each term is positive. Further
(a + a9 − a3 − a7)2 =4 − a2 − a4 − a6 − a8
=4 − (a4 − a3 + a2 − a)
=4 − (IrrF,a(a) − 1) = 5
5) =g(a + a9 − a3 − a7)
=g(a + g(2)(a) − g(a) − g(3)(a))
=g(a) + g(3)(a) − g(2)(a) − a = −
√
5
59
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
√
5,−√
5). Now let S1 ⊆ N × {1},
and so we have the intermediate field Q(a)(cid:104)g(2)(cid:105) = Q(
S2 ⊆ N × {√
5}, S3 ⊆ N × {a, a3, a7, a9} and X := S1 ∪ S2 ∪ S3 all be non-empty
finite sets such that for (x, y) ∈ X we have (x, g(y)) ∈ X. Put the trivial topology on
X so that X is a totally disconnected compact Hausdorff space noting that a set with
the trivial topology is compact if and only if it is finite. Define a topological element
of finite order τ on X by τ((x, y)) := (x, g(y)) and note that for our choice of topology
every self map on X is continuous as is every map from X to L. Hence CL(X) is the
Q(a)/Q uniform algebra of all functions from X to L and we also have ord(τ)ord(g) by
construction. Hence with reference to Definition 5.1.2 we have C(X, τ, g) as an example
of a basic Q(a)/Q function algebra. For z ∈ X each f ∈ C(X, τ, g) is such that
f (z) ∈ Q(a)
√
f (z) ∈ Q(
5)
f (z) ∈ Q
if z ∈ S3,
if z ∈ S2 and
if z ∈ S1 since
f (z) = f (τ(ord(τ,z))(z)) = g(ord(τ,z))( f (z)) giving ord(g, f (z))ord(τ, z). Furthermore
C(X, τ, g) extends to C(X, τ(2), g(2)) which is a basic Q(a)/Q(
5) function algebra. We
will look at such extensions in the next section. Finally note that in general if we use
the trivial valuation on L then for every totally disconnected compact Hausdorff space
X the sup norm (cid:107) · (cid:107)∞ is the trivial norm on CL(X).
√
We now look at some non-Archimedean examples involving an order two extension of
the 5-adic numbers.
Example 5.3.3. Let F := Q5 and L := Q5(
2). Suppose towards a contradiction that
√
2 is already an element of Q5. With reference to Chapter 2, we would have 1 = 25 =
√
√
225 = √
2 would have a 5-adic expansion over
R5 := {0, 1,· · · , 4} of the form ∑∞
5 giving √
22
25 = 1. But then
√
i=0 ai5i with a0 (cid:54)= 0. Hence
(cid:33)2
(cid:32) ∞
ai5i
∑
i=1
a2
0 + 2a0
ai5i +
∞
∑
i=1
(5.3.1)
should be equal to 2. In particular a2
0 should have the form 2 + b where b is a natural
number, with a factor of 5, that cancels with the remaining terms of (5.3.1). But since
0 ∈ {1, 4, 4 + 5, 1 + 3 · 5}, a contradiction. Therefore we have
a0 ∈ {1, 2, 3, 4} we have a2
√
IrrF,
2Q5 as a Q5-vector
space. It is immediate that L is a Galois extension of F with Gal(L/F) = (cid:104)id, g(cid:105) where
√
g(
2. The complete valuation on F has a unique extension to a complete
valuation on L, see Theorem 2.2.13. Explicitly we have, for all a ∈ L,
2(x) = x2 − 2 giving [L, F] = 2 and so Q5(
2) = −√
2) = Q5 ⊕ √
√
(cid:113)aLg(a)L =
(cid:113)ag(a)L =
(cid:113)ag(a)5,
aL =
60
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
(cid:112)ag(a)5 = 5− 1
noting that ag(a) ∈ F. Moreover in terms of the valuation logarithm ν5 on F we have
2 ν5(ag(a)) for
2 ν5(a2 − 2b2).
2 ν5(ag(a)) and so the valuation logarithm for L is ω(a) := 1
2b) = 1
2b ∈ L×, with a, b ∈ F, we have ω(a +
√
√
a ∈ L. Hence for a +
We show that this in fact gives
√
2b) = min{ν5(a), ν5(b)}.
(5.3.2)
ω(a +
√
√
2b) = 1
2 ν5(a2) = 1
22ν5(a) = ν5(a).
22ν5(b) = ν5(b), noting
2 (ν5(−2) + 2ν5(b)) = 1
2 ν5(−2b2) = 1
i=1 4 · 5i.
First recall that ν5(0) = ∞. If b = 0 then ω(a) = 1
If a = 0 then ω(
that ν5(−2) = 0 since −2 = 3 + ∑∞
If a, b ∈ F× and ν5(a) (cid:54)= ν5(b) then by the above ν5(a2) (cid:54)= ν5(−2b2). Hence, by Lemma
2 min{ν5(a2), ν5(−2b2)} = min{ν5(a), ν5(b)}.
2.1.7, ω(a +
If a, b ∈ F× and ν5(a) = ν5(b) = n for some n ∈ Z then the expansion a = ∑∞
over R5 has an ∈ {1, 2, 3, 4} and so the expansion a2 = ∑∞
2n = a2
residue field F = F5 giving a(cid:48)
2 ν5(a2 − 2b2) = 1
bn ∈ {1, 2, 3, 4} and so −2b2 =(cid:0)3 + ∑∞
2n ∈ {1, 4}. Similarly the expansion b = ∑∞
i5i has b(cid:48)
= ∑∞
i=2n ci5i has c2n = a(cid:48)
2n ∈ {2, 3}. Hence the expansion a2 − 2b2 = ∑∞
i=n ai5i
n in the
i=n bi5i has
2n = 3b2
n
2n + b(cid:48)
2n
i=1 4 · 5i(cid:1)(cid:0)∑∞
i=n bi5i(cid:1)2
i5i has a(cid:48)
i=2n b(cid:48)
i=2n a(cid:48)
2b) = 1
in F giving b(cid:48)
in F giving c2n ∈ {1, 2, 3, 4}. In particular c2n (cid:54)= 0 and so
∞
∑
i=2n
2 ν5(a2 − 2b2) =
1
2 ν5(
ω(a +
2b) =
√
1
ci5i) =
1
2
2n = n
√
and this completes the proof of (5.3.2). With reference to Remark 2.2.14 it follows that L
is an unramified extension of F with a +
2bL = max{aL,bL} = max{aF,bF} for
a, b ∈ F. Further it follows easily from (5.3.2) that RL := {a +
2b : a, b ∈ {0,· · · , 4}}
is a set of representatives in L of the elements in the residue field L. Hence L = F25
since #RL = 25 and [L, F] = 2. Therefore, by Theorem 2.1.23, L is locally compact and
the unit ball ∆L := {x ∈ L : xL ≤ 1} = {x ∈ L : ω(x) ≥ 0} is a totally disconnected
compact Hausdorff space with respect to · L. Further if we take τ1 to be the restriction
of g to ∆L then, since g is an isometry on L, τ1 is a topological involution on ∆L and the
basic L/F function algebra
√
C(∆L, τ1, g) = { f ∈ CL(∆L) : f (τ1(x)) = g( f (x)) for all x ∈ ∆L}
is a non-Archimedean analog of the real disc algebra. Now let f (x) = ∑∞
power series in C(∆L, τ1, g). Then for x ∈ ∆L and σ from Remark 5.2.3 we have
n=0 anxn be a
∞
∑
n=0
anxn = f (x) = σ( f )(x) = g
ang(x)n
=
∞
∑
n=0
g(an)xn
(cid:33)
(cid:32) ∞
∑
n=0
where the last equality follows because the two series have identical sequences of
partial sums. Hence similarly we have, for x ∈ ∆L, ∑∞
In
the general case of such circumstance we can not immediately assume that all the
n=0(an − g(an))xn = 0.
61
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
n=m(an − g(an))xn−m = 0 on ∆L and ∑∞
pairs of coefficients an and g(an) are equal since ∆L could be a set of roots of the se-
n=0(an − g(an))xn whilst there being an element of L in the region of conver-
ries ∑∞
gence of the series that is not a root. However since 0 ∈ ∆L we have a0 = g(a0).
Now let m ∈ N be such that for all i ∈ N0 with i < m we have ai = g(ai). Then
n=m(an − g(an))xn−m = 0 on ∆L\{0}. Let
xm ∑∞
n=m(an − g(an))xn−m. Then with reference to The-
ρ be the radius of convergence of ∑∞
orem 3.1.2, since 1 ∈ ∆L\{0} with 1L = 1 and ∑∞
n=m(an − g(an))xn−m converges on
n=m(an − g(an))xn−m converges uniformly on the ball
∆L\{0} we have ρ ≥ 1. Hence ∑∞
n=m(an − g(an))xn−m is con-
¯B 1
tinuous on ¯B 1
(0). Hence am = g(am)
and by induction we have shown that an = g(an) for all n ∈ N0. In particular all the
power series in C(∆L, τ1, g) only have F valued coefficients. However since ∆L (cid:54)⊆ F
these functions take values in L.
(0) = {x ∈ L : ω(x) ≥ 1} by Theorem 3.1.2. Therefore ∑∞
5
(0) and so ∑∞
n=m(an − g(an))xn−m = 0 at 0 ∈ ¯B 1
5
5
Whilst the last example of a basic function algebra included many globally analytic
functions the only globally analytic functions in the following example are constants.
However many locally analytic functions are included.
Example 5.3.4. Let F, L, ∆L, ω and g be as in Example 5.3.3 and therefore note that
ω∆L : ∆L → N0 ∪ {∞}. Define τ2(0) := 0 and for x ∈ ∆L\{0},
(cid:40)
τ2(x) :=
5x
5−1x
if
if
2 ω(x)
2 (cid:45) ω(x).
(5.3.3)
Let x ∈ ∆L with ω(x) ∈ N0. Then ω(τ2(x)) = ω(5x) = ω(x) + ω(5) = ω(x) + 1 if
2ω(x). Similarly ω(τ2(x)) = ω(x) − 1 if 2 (cid:45) ω(x). Hence when ω(x) ∈ N0 we have
ω(τ2(x)) ∈ N0 giving τ2(x) ∈ ∆L. Further τ2 : ∆L → ∆L since τ2(0) = 0. Moreover
ord(τ2) = 2 and so to show that τ2 is a topological involution on ∆L it remains to show
that τ2 is continuous. Let x ∈ ∆L and (xn) be a sequence in ∆L such that limn→∞ xn = x.
Let ε > 0. For x (cid:54)= 0 there exists N ∈ N such that for all n ≥ N we have ω(xn) = ω(x)
since convergence in L is from the side, see Lemma 3.1.1. With reference to (5.3.3) this
gives, for all n ≥ N, τ2(xn) = τ2(xn)x−1
n xn = τ2(x)x−1xn. Further since limn→∞ xn = x
there exists M ∈ N such that, for all m ≥ M, τ2(x)x−1Lx − xmL < ε. Hence for all
n ≥ max{M, N} we have
τ2(x) − τ2(xn)L = τ2(x)x−1(x − xn)L = τ2(x)x−1Lx − xnL < ε.
On the other hand for x = 0 note that ω(τ2(xn)) ≥ ω(xn) − 1 for all n ∈ N. In this case
since limn→∞ xn = 0 there exists N(cid:48) ∈ N such that for all n ≥ N(cid:48) we have 5xnL < ε
giving
τ2(xn)L = 5−ω(τ2(xn)) ≤ 5−(ω(xn)−1) = 5xnL < ε
62
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
as required. Hence τ2 is a topological involution on ∆L. We now consider the basic L/F
function algebra
C(∆L, τ2, g) = { f ∈ CL(∆L) : f (τ2(x)) = g( f (x)) for all x ∈ ∆L}.
We begin by proving that the only power series in C(∆L, τ2, g) are the constants be-
longing to F. Let f (x) := ∑∞
n=0 anxn be a power series in C(∆L, τ2, g). Since τ2(0) = 0
and f (τ2(0)) = g( f (0)) we have a0 = g(a0) giving a0 ∈ F and so a0 ∈ C(∆L, τ2, g).
Hence h := f − a0 is also in C(∆L, τ2, g). Suppose towards a contradiction that h is
not identically zero on ∆L. Since 1 ∈ ∆L, ∑∞
n=1 an converges and so by Lemma 3.1.1
we have limn→∞ ω(an) = ∞. Hence we can define M := min{ω(an) : n ∈ N}. Also
let m := min{n ∈ N : an (cid:54)= 0}. Now since ∆L = {x ∈ L : ω(x) ≥ 0} we can find
y ∈ ∆L\{0} such that 2ω(y) and M + ω(y) > ω(am). Hence for every n > m we have
ω(amym) = ω(am) + mω(y) < M + ω(y) + mω(y) ≤ ω(an) + nω(y) = ω(anyn).
So, by Lemma 3.1.1, ω (∑∞
ω (∑∞
every n > m we have
n=m anyn) = ω (amym + ∑∞
n=m+1 anyn) ≥ min{ω(anyn) : n ≥ m + 1} > ω(amym). Hence
n=m+1 anyn) = ω(amym) by Lemma 2.1.7. Similarly for
ω(am5mym) =ω(am) + m(ω(y) + 1)
<M + ω(y) + 1 + m(ω(y) + 1)
≤ω(an) + n(ω(y) + 1) = ω(an5nyn),
giving ω (∑∞
n=m an5nyn) = ω(am5mym) = ω(5m) + ω(amym) = m + ω(amym). Now
h(τ2(y)) = g(h(y)) and 2ω(y) hence ∑∞
n=m anyn). But, since g is
an isometry, ω (g (∑∞
n=m an5nyn) = m + ω(amym) with
m ∈ N which is a contradiction. Therefore h is identically zero on ∆L as required.
However, whilst the only power series in C(∆L, τ2, g) are constants belonging to F, it is
easy to construct locally analytic elements of C(∆L, τ2, g) using power series. Define
n=m anyn)) = ω(amym) and ω (∑∞
n=m an5nyn = g (∑∞
C(n) := {x ∈ ∆L : ω(x) = n} for n ∈ ω(∆L),
let (en)n∈N be the even sequence in ω(∆L) given by en := 2(n − 1) and let a ∈ F. Now
let ( fn)n∈N be a sequence of power series with the following properties:
(i) for all n ∈ N the coefficients of fn are elements of L;
(ii) for all n ∈ N we have 5−en < ρn where ρn is the radius of convergence of fn so
that fn is convergent on C(en);
(iii) we have limn→∞ inf{ω( fn(x) − a) : x ∈ C(en)} = ∞.
63
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
We then define f : ∆L → L by
f (x) :=
fn(x)
g( fn(τ2(x)))
a
if
if
if
x ∈ C(en)
x ∈ C(en + 1)
x = 0.
We show that f is continuous. Let x ∈ ∆L and let (xn) be a sequence in ∆L such that
limn→∞ xn = x.
If x (cid:54)= 0 then by Lemma 3.1.1 there is N ∈ N such that for all n ≥ N we have ω(xn) =
ω(x). If for some m ∈ N we have x ∈ C(em) then f (x) = fm(x) and for all n ≥ N we
have f (xn) = fm(xn) since xn ∈ C(em). Hence by the continuity of fm on C(em) we have
limn→∞ f (xn) = f (x).
If for some m ∈ N we have x ∈ C(em + 1) then f (x) = g( fm(τ2(x))), with τ2(x) ∈
C(em), and for all n ≥ N we have f (xn) = g( fm(τ2(xn))), since xn ∈ C(em + 1), with
τ2(xn) ∈ C(em). Now τ2 is continuous on ∆L, fm is continuous on C(em) and g is an
isometry on L. Hence again we have limn→∞ f (xn) = f (x).
If x = 0 then by the definition of f we have f (x) = a. Let ε < ∞. We need to show
that there is N ∈ N such that for all n ≥ N we have ω( f (xn) − a) > ε. By property
(iii) given in the construction of f there is M ∈ N such that for all m ≥ M we have
inf{ω( fm(y) − a) : y ∈ C(em)} > ε. Since limn→∞ ω(xn) = ∞ there is N ∈ N such
that for all n ≥ N we have ω(xn) ≥ eM. So let n ≥ N. Then either xn = 0, noting that
ω(0) = ∞, or there is m ≥ M with either xn ∈ C(em) or xn ∈ C(em + 1).
For xn = 0 we have ω( f (0) − a) = ω(a − a) = ∞ > ε.
For xn ∈ C(em) we have ω( f (xn) − a) = ω( fm(xn) − a) > ε since m ≥ M.
For xn ∈ C(em + 1) define y := τ2(xn) and note that y ∈ C(em). Then
ω( f (xn) − a) =ω(g( fm(τ2(xn))) − a)
=ω(g( fm(y) − a))
=ω( fm(y) − a)
>ε
(since a ∈ F)
(since g is an isometry on L)
(since m ≥ M and y ∈ C(em)).
Hence f is continuous. We now show that f ∈ C(∆L, τ2, g). Let x ∈ ∆L.
For x = 0 we have f (τ2(0)) = f (0) = a = g(a) = g( f (0)) since a ∈ F.
For x ∈ C(en), for some n ∈ N, we have f (x) = fn(x). Define y := τ2(x) giving
y ∈ C(en + 1). Then we have
f (τ2(x)) = f (y) = g( fn(τ2(y))) = g( fn(τ2(τ2(x)))) = g( fn(x)) = g( f (x)).
For x ∈ C(en + 1), for some n ∈ N, we have f (x) = g( fn(τ2(x))). Put y := τ2(x) giving
y ∈ C(en). Then we have
f (τ2(x)) = f (y) = fn(y) = fn(τ2(x)) = g(g( fn(τ2(x)))) = g( f (x)).
64
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
Hence f ∈ C(∆L, τ2, g) as required. Now suppose there is N ∈ N such that for all
n ≥ N we have fn = a. Then f will be locally analytic on ∆L noting that convergence
in ∆L is from the side, in particular see the proof of Lemma 3.1.1.
Remark 5.3.5. Concerning examples 5.3.3 and 5.3.4.
(i) Since g is an isometry on ∆L and 5 ∈ F we note that τ1 ◦ τ2 is also a topological
involution on ∆L with ord(τ1 ◦ τ2) = 2. For f ∈ C(∆L, τ1, g) ∩ C(∆L, τ2, g) and
x ∈ ∆L we have f (τ1 ◦ τ2(x)) = g( f (τ2(x))) = g(g( f (x))) = f (x) which gives
f ∈ C(∆L, τ1 ◦ τ2, id). But, with reference to Figure 5.2, C(∆L, τ1 ◦ τ2, id) is not
a basic function algebra since it fails to separate the points of ∆L noting that we
have ord(τ1 ◦ τ2) (cid:45) ord(id). Hence C(∆L, τ1, g) ∩ C(∆L, τ2, g) is not a L/F function
algebra on (∆L, τ1, g). However C(∆L, τ1 ◦ τ2, g) is a basic function algebra.
(ii) We note that by Theorem 5.2.4 every element f ∈ C(∆L, τ2, g) can be uniformly
approximated by polynomials belonging to CL(∆L). However, apart from the
elements of F, none of these polynomials belong to C(∆L, τ2, g).
In the next section we look at ways of obtaining more basic function algebras.
5.4 Non-Archimedean new basic function algebras from old
5.4.1 Basic extensions
The following theorem concerns extensions of basic function algebras resulting from
the field structure involved.
Theorem 5.4.1. Let the basic L/Lg function algebra on (X, τ, g) be such that Gal(L/Lg ) and
(cid:104)id(cid:105) are respectively at the top and bottom of a lattice of groups with intermediate elements.
Then CL(X) and C(X, τ, g) are respectively at the top and bottom of a particular lattice of basic
function algebras with intermediate elements and there is a one-one correspondence between the
subgroups of Gal(L/Lg ) and the elements of this lattice which we will call the lattice of basic
extensions of C(X, τ, g).
Proof. With reference to Remark 5.2.3, the automorphism σ( f ) = g(ord(g)−1) ◦ f ◦ τ, for
f ∈ CL(X), is such that CL(X)(cid:104)σ(cid:105) = C(X, τ, g) where
CL(X)
(cid:104)σ(cid:105) := { f ∈ CL(X) : σ( f ) = f}.
Now by the fundamental theorem of Galois theory we have Gal(L/Lg ) = (cid:104)g(cid:105) and so
Gal(L/Lg ) is a cyclic group. Moreover we have ord(σ) = ord(g) giving (cid:104)σ(cid:105) ∼= (cid:104)g(cid:105) as
65
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
cyclic groups. It is standard from group theory that a subgroup of a cyclic group is
cyclic. In particular, for nord(σ), (cid:104)σ(n)(cid:105) is the unique cyclic subgroup of (cid:104)σ(cid:105) of size
#(cid:104)σ(n)(cid:105) = ord(σ(n)) = ord(σ)
#G ∈ N,
#G we have (cid:104)σ(n)(cid:105) = G with nord(σ).
by Lagrange's theorem, and so for n = ord(σ)
Hence we define a map ς : {(cid:104)σ(n)(cid:105) : nord(σ)} → {CL(X)(cid:104)σ(n)(cid:105) : nord(σ)} by
. Moreover for G a subgroup of (cid:104)σ(cid:105) we have ord(σ)
n
ς((cid:104)σ(n)(cid:105)) := CL(X)
(cid:104)σ(n)(cid:105) := { f ∈ CL(X) : σ(n)( f ) = f} = C(X, τ(n), g(n)).
Now let nord(σ). Since ord(τ)ord(g) we also have ord(τ, X) ⊆ ord(g, L), see Figure
5.2. Hence ord(τ(n), X) ⊆ ord(g(n), L) giving ord(τ(n))ord(g(n)) and so ς((cid:104)σ(n)(cid:105)) is a
basic function algebra. Moreover since
C(X, τ(n), g(n)) = { f ∈ CL(X) : f (τ(n)(x)) = g(n)( f (x)) for all x ∈ X}
the constants in ς((cid:104)σ(n)(cid:105)) are the elements of the field L(cid:104)g(n)(cid:105) and so ς is injective by
the fundamental theorem of Galois theory. Finally it is immediate that the elements of
{CL(X)(cid:104)σ(n)(cid:105) : nord(σ)} form a lattice as described in the theorem and this completes
the proof.
Example 5.4.2. Let F = Q and let L = Q(a) where a = exp( 1
142πi). Having factorised
x14 − 1 in F[x], we have IrrF,a(x) = x6 − x5 + x4 − x3 + x2 − x + 1 with roots S :=
{a, a3, a5, a9, a11, a13}. Hence L is the splitting field of IrrF,a(x) over F and so L is a
Galois extension of F with #Gal(L/F) = [L, F] = 6 by Theorem 2.2.11. In fact putting
g(a) := a3 makes g a generator of Gal(L/F) and so L is a cyclic extension of F and
F = Lg since L is a Galois extension. We can take C(X, τ, g) to be a basic L/Lg function
algebra constructed by analogy with Example 5.3.2 where X ⊆ N × L is non-empty
and finite with τ((x, y)) = (x, g(y)) ∈ X for all (x, y) ∈ X. In this case Figure 5.3 shows
the lattice of basic extensions of C(X, τ, g) as given by Theorem 5.4.1. Finally we note
CL(X)(cid:104)id(cid:105)
ς
CL(X)(cid:104)σ(2)(cid:105)
CL(X)(cid:104)σ(3)(cid:105)
(cid:104)σ(3)(cid:105)
CL(X)(cid:104)σ(cid:105)
(cid:104)σ(2)(cid:105)
(cid:104)id(cid:105)
(cid:104)σ(cid:105)
Figure 5.3: Lattice of basic extensions.
that in this example F = Q, requiring the trivial valuation, was chosen just to keep
things simple.
66
3
3
y
y
'
'
8
8
f
f
3
3
&
&
3
3
w
w
f
f
8
8
3
3
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
5.4.2 Residue algebras
We begin with an analog of Definition 2.1.10.
Definition 5.4.3. For C(X, τ, g) a basic L/Lg function algebra in the non-Archimedean
setting, with valuation logarithm ω on L, we define:
(i) O(X, τ, g) := { f ∈ C(X, τ, g) : infx∈X ω( f (x)) ≥ 0, equivalently (cid:107) f(cid:107)∞ ≤ 1};
(ii) O×(X, τ, g) := { f ∈ C(X, τ, g) : ω( f (x)) = 0, equivalently f (x)L = 1,∀x ∈ X};
(iii) J (X, τ, g) := { f ∈ C(X, τ, g) : infx∈X ω( f (x)) > 0, equivalently (cid:107) f(cid:107)∞ < 1};
(iv) My(X, τ, g) := { f ∈ O(X, τ, g) : ω( f (y)) > 0, equivalently f (y)L < 1} for
y ∈ X.
In this subsection we will mainly be interested in the following two theorems and their
proofs. The main theorem is Theorem 5.4.5 which concerns the residue algebra of par-
ticular basic function algebras. Before proving these theorems we will need to prove
several other results that are also of interest in their own right.
Theorem 5.4.4. If C(X, τ, g) is a basic L/Lg function algebra in the non-Archimedean setting
then:
(i) O(X, τ, g) is a ring;
(ii) O×(X, τ, g) is the multiplicative group of units of O(X, τ, g);
(iii) J (X, τ, g) is an ideal of O(X, τ, g);
(iv) My(X, τ, g) is a maximal ideal of O(X, τ, g) for each y ∈ X.
Theorem 5.4.5. Let F be a locally compact complete non-Archimedean field of characteristic
zero with nontrivial valuation. Let L be a finite unramified extension of F with Lg = F for some
g ∈ Gal(L/F) and let C(X, τ, g) be a basic L/F function algebra. Then there is an isometric
isomorphism
O(X, τ, g)/J (X, τ, g) ∼= C(X, τ, ¯g)
where C(X, τ, ¯g) is the basic L/F function algebra on (X, τ, ¯g). Here F and L are respectively
the residue field of F and L whilst ¯g is the residue automorphism on L induced by g. More
generally L need not be an unramified extension of F for the above to hold provided that we
impose the condition ord(τ)ord( ¯g) directly.
Remark 5.4.6. Concerning Theorem 5.4.5.
67
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
(i) The conditions in Theorem 5.4.5 imply that F contains a p-adic field, up to a posi-
tive exponent of the valuation, since by Theorem 2.1.23 the residue field F is finite
and so the valuation on F when restricted to Q can not be trivial.
(ii) Since L is finite the valuation on L is the trivial valuation. In general the quotient
norm on a residue field is the trivial valuation. In particular for ¯a ∈ L with residue
class representative a ∈ OL we have, by Lemma 2.1.7, that
min{a − bL : b ∈ ML} =
(cid:40)
1
0
if
if
a /∈ ML
a ∈ ML.
(iii) To be thorough for the reader we show that ¯g is well defined, although this is
covered in [FV02, p52]. Let ¯a ∈ L with residue class representative a ∈ OL. For
g ∈ Gal(L/F) we obtain ¯g ∈ Gal(L/F) by ¯g(¯a) := g(a). Now let b ∈ OL with
b (cid:54)= a but ¯b = ¯a so that a − b ∈ ML. By Remark 2.2.6, g is an isometry on L and
so g(a) − g(b) = g(a − b) ∈ ML giving ¯g(¯b) = g(b) = g(a) = ¯g(¯a) and so ¯g is
well defined.
(iv) The map g (cid:55)→ ¯g is a homomorphism from Gal(L/F) to Gal(L/F). Indeed for ¯a ∈ L
and g1, g2 ∈ Gal(L/F) we have
g1 ◦ g2(¯a) = g1 ◦ g2(a) = g1(g2(a)) = ¯g1
g2(a)
= ¯g1( ¯g2(¯a)) = ¯g1 ◦ ¯g2(¯a).
(cid:16)
(cid:17)
Under the conditions of Theorem 5.4.5 this homomorphism becomes an isomor-
phism as per Lemma 5.4.7 below, see [FV02, p52]. In particular this ensures that
C(X, τ, ¯g), in Theorem 5.4.5, is a basic function algebra since ord( ¯g) = ord(g)
gives ord(τ)ord( ¯g).
Lemma 5.4.7. Let F be a local field, as per Remark 2.1.24, and let L be a finite unramified Galois
extension of F. Then Gal(L/F) ∼= Gal(L/F) giving ord( ¯g) = ord(g) for all g ∈ Gal(L/F).
The following definition and lemma will be useful when proving Theorem 5.4.4. Note
that the first part of Lemma 5.4.9 makes sense even though we have yet to show that
O(X, τ, g) is a ring.
Definition 5.4.8. Let L be a complete valued field with valuation logarithm ω and let
X be a totally disconnected compact Hausdorff space.
(i) We call a map ι : X → ω(OL) a value level function.
(ii) We place a partial order on the set of all value level function by setting
ι1 ≥ ι2 if and only if for all x ∈ X we have ι1(x) ≥ ι2(x).
68
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
Lemma 5.4.9. Let C(X, τ, g) be a basic L/Lg function algebra in the non-Archimedean setting
with valuation logarithm ω on L and let ι : X → ω(OL) be a value level function. Then:
(i) Mι(X, τ, g) := { f ∈ C(X, τ, g) : ω( f (x)) ≥ ι(x) for all x ∈ X} is an ideal of
O(X, τ, g);
(ii) for ι(cid:48) another value level function with ι ≥ ι(cid:48) we have Mι(X, τ, g) ⊆ Mι(cid:48) (X, τ, g).
Proof. For (i), let f1, f2 ∈ Mι(X, τ, g) and f ∈ O(X, τ, g). Then for each x ∈ X we have
ω( f1(x) + f2(x)) ≥ min{ω( f1(x)), ω( f2(x))} ≥ ι(x) giving f1 + f2 ∈ Mι(X, τ, g) and
ω( f1(x) f (x)) = ω( f1(x)) + ω( f (x)) ≥ ω( f1(x)) ≥ ι(x) giving f1 f ∈ Mι(X, τ, g) as
required. For (ii), this is immediate.
Note, the mapping ι
Theorem 5.4.4.
(cid:55)→ Mι(X, τ, g) is not assumed to be injective. We now prove
Proof of Theorem 5.4.4. For (i), note that since ω(1) = 0, ω(0) = ∞ and 1, 0 ∈ F we
have 1, 0 ∈ O(X, τ, g). Further O(X, τ, g) is closed under multiplication and addi-
tion by Lemma 5.4.9 since for the value level function that is constantly zero we have
O(X, τ, g) = M0(X, τ, g). Hence O(X, τ, g) is a ring.
For (ii), we need to show that O×(X, τ, g) = O(X, τ, g)×. Let f ∈ O(X, τ, g)×. Then
for all x ∈ X we have ω( f (x)) ≥ 0 and ω( f −1(x)) ≥ 0 since f , f −1 ∈ O(X, τ, g)×
but we also have ω( f −1(x)) = ω(( f (x))−1) = −ω( f (x)) giving ω( f (x)) = 0. Hence
O(X, τ, g)× ⊆ O×(X, τ, g). Now let f ∈ O×(X, τ, g). We have
ω( f −1(x)) = −ω( f (x)) = 0
(5.4.1)
for all x ∈ X and so it remains to show that f −1 is an element of C(X, τ, g). We have
1 = g(1) = g( f −1 f ) = g( f −1)g( f ) giving g( f −1) = (g( f ))−1 and so
f −1(τ) = ( f (τ))
−1 = (g( f ))
−1 = g( f −1).
For continuity let x ∈ X and (xn) be a sequence in X with limn→∞ xn = x in X. Then
by (5.4.1) we have
ω( f −1(xn) − f −1(x)) =ω( f −1(xn) − f −1(x)) + ω( f (x))
=ω(( f −1(xn) − f −1(x)) f (x))
=ω( f −1(xn) f (x) − 1)
=ω( f −1(xn)( f (x) − f (xn)))
=ω( f −1(xn)) + ω( f (x) − f (xn))
=ω( f (x) − f (xn)).
69
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
n
(X, τ, g) for all n ∈ N and that a union of nested ideals is an ideal.
n
Therefore by the continuity of f we have limn→∞ f −1(xn) = f −1(x) from which it fol-
lows that O×(X, τ, g) ⊆ O(X, τ, g)× as required.
For (iii), taking into account that the valuation on L could be dense, J (X, τ, g) is an
(X, τ, g) noting that
ideal of O(X, τ, g) by Lemma 5.4.9 since J (X, τ, g) = (cid:83)
1 (cid:54)∈ M 1
(cid:83)
For (iv), My(X, τ, g) is an ideal of O(X, τ, g) by Lemma 5.4.9 since My(X, τ, g) =
n∈N M 1
n χ{y} (X, τ, g) where χ{y} is the indicator function. Also by Lemma 5.4.9, since
n ≥ 1
n χ{y} for all n ∈ N, we have J (X, τ, g) ⊆ My(X, τ, g). We now show that
My(X, τ, g) is a maximal ideal of O(X, τ, g). Let I(X, τ, g) be a, not necessarily proper,
ideal of O(X, τ, g) with My(X, τ, g) (cid:36) I(X, τ, g). Then there is f ∈ I(X, τ, g) with
ω( f (y)) = 0. Define on X
n∈N M 1
1
(cid:40)
(cid:40)
f (cid:48)
(x) :=
0
1
if ω( f (x)) = 0
if ω( f (x)) > 0.
We show that f (cid:48) is an element of My(X, τ, g) and so f (cid:48) ∈ I(X, τ, g). For continuity let
x ∈ X and (xn) be a sequence of elements of X with limn→∞ xn = x in X. Since f is
continuous we have limn→∞ f (xn) = f (x) with respect to ω. Hence if f (x) = 0 then
there exists N ∈ N such that for all n ≥ N we have ω( f (xn)) = ω( f (xn) − f (x)) > 0.
If f (x) (cid:54)= 0 then since convergence in L is from the side, see Lemma 3.1.1, there exists
N ∈ N such that for all n ≥ N we have ω( f (xn)) = ω( f (x)). Hence in every case
there exists N ∈ N such that for all n ≥ N we have f (cid:48)(xn) = f (cid:48)(x) and so f (cid:48) is
continuous. We need to show that f (cid:48)(τ(x)) = g( f (cid:48)(x)). Since g is an isometry on L we
have ω( f (τ(x))) = ω(g( f (x))) = ω( f (x)) giving
f (cid:48)
(τ(x)) =
0
1
if ω( f (τ(x))) = 0
if ω( f (τ(x))) > 0
= f (cid:48)
(x)
=g( f (cid:48)
(x))
noting that f (cid:48) takes values only in {0, 1} ⊆ F. Now ω(1) = 0 and ω(0) = ∞ so that for
all x ∈ X we have ω( f (cid:48)(x)) ≥ 0 giving f (cid:48) ∈ O(X, τ, g). Further since ω( f (y)) = 0 we
have ω( f (cid:48)(y)) = ω(0) = ∞ and so we have shown that f (cid:48) ∈ My(X, τ, g) (cid:36) I(X, τ, g).
Now since I(X, τ, g) is an ideal we have f + f (cid:48) ∈ I(X, τ, g). Moreover by the definition
of f (cid:48), for each x ∈ X, if ω( f (x)) = 0 then ω( f (cid:48)(x)) = ω(0) = ∞ and if ω( f (x)) > 0
then ω( f (cid:48)(x)) = ω(1) = 0. Hence for all x ∈ X, ω( f (x) + f (cid:48)(x)) = 0 by Lemma 2.1.7
and so f + f (cid:48) ∈ O×(X, τ, g) giving I(X, τ, g) = O(X, τ, g). Therefore My(X, τ, g) is a
maximal ideal of O(X, τ, g) and this completes the proof of Theorem 5.4.4.
The following two lemmas will be used in the proof of Theorem 5.4.5. The first of these,
Lemma 5.4.10, will be known but we provide a proof in the absence of a reference. The
70
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
second, Lemma 5.4.12, may be new, since I have not seen it in the literature, however it
could be known to some number theorists.
Lemma 5.4.10. Let F be a complete non-Archimedean field with a nontrivial, discrete valu-
ation and valuation logarithm ν. Let π be a prime element and R be a set of residue class
representatives for F, as shown in Theorem 2.1.13. Then, for X a compact Hausdorff space, each
f ∈ CF(X) has a unique expansion, as a series of locally constant R-valued functions, of the
form
f =
∞
∑
i=n
fiπi,
for some n ∈ Z.
Moreover, for j ≥ n and x, y ∈ X with ν( f (x) − f (y)) > jν(π), we have fi(x) = fi(y) for
all i in the interval n ≤ i ≤ j.
Proof. Let f ∈ CF(X) and note that since X is compact, f is bounded. Hence there is
n ∈ Z such that, for all x ∈ X, ν( f (x)) ≥ nν(π). Therefore by allowing terms to be
zero where necessary and by using the unique π-power series expansion over R for
elements of F×, as shown in Theorem 2.1.13, we have for each x ∈ X
f (x) =
fi(x)πi ∈ F.
∞
∑
i=n
Hence for each i ≥ n we have obtained a function fi : X → R and the resulting ex-
i=n fiπi is unique. Now for j ≥ n let x, y ∈ X be such that we have
pansion f = ∑∞
ν( f (x) − f (y)) > jν(π). If we do not have fk(x) = fk(y) for all k ≥ n then let k ≥ n
be the first integer for which fk(x) (cid:54)= fk(y). Therefore fk(x) and fk(y) are representa-
tives in OF of two different residue classes. Hence fk(x) − fk(y) (cid:54)∈ MF showing that
ν( fk(x) − fk(y)) = 0. Therefore by Lemma 2.1.7 and the definition of k we have
kν(π) =ν( fk(x) − fk(y)) + ν(πk)
(cid:32)
=ν(( fk(x) − fk(y))πk)
( fk(x) − fk(y))πk +
(cid:32) ∞
=ν
∑
i=n
fi(x)πi − ∞
∑
i=n
=ν
=ν( f (x) − f (y)) > jν(π).
∞
∑
i=k+1
(cid:33)
fi(y)πi
(cid:33)
fi(x)πi − ∞
∑
i=k+1
fi(y)πi
Hence k > j giving fi(x) = fi(y) for all i in the interval n ≤ i ≤ j. Finally we show, for
all j ≥ n, that fj is a locally constant function. For x ∈ X define the following ball in F
Bjν(π)( f (x)) := {a ∈ F : ν( f (x) − a) > jν(π)}.
Then since f is continuous there exists an open subset U of X with x ∈ U such that
f (U) ⊆ Bjν(π)( f (x)). Hence for each y ∈ U we have ν( f (x) − f (y)) > jν(π) and so
71
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
fj(x) = fj(y). In particular fj is constant on U and this completes the proof of Lemma
5.4.10.
Lemma 5.4.10 has the following corollary which, for locally constant functions, goes
slightly further than Theorem 5.2.4 since it does not assume that X is totally discon-
nected.
Corollary 5.4.11. Let F and X be as in Lemma 5.4.10 and let LCF(X) be the set of all locally
constant F valued functions defined on X. Then LCF(X) is uniformly dense in CF(X).
Proof. For f ∈ CF(X) let f = ∑∞
i=n fiπi be the expansion from Lemma 5.4.10. We
note that a finite sum of locally constant functions is locally constant. Hence, for each
m ≥ n, fi≤m := ∑m
ν(π). Then we
have infx∈X ν( f (x) − fi≤m(x)) = infx∈X ν(∑∞
i=m+1 fi(x)πi) ≥ ν(πm+1) > mν(π) > ε as
required.
i=n fiπi is an element of LCF(X). Let ε < ∞ and m > ε
Here is the second of the two lemmas that will be used in the proof of Theorem 5.4.5.
Lemma 5.4.12. Let F and L be non-Archimedean fields with L a finite extension of F as a
valued field such that the following holds:
(i) we have Q ⊆ F and the valuation logarithm ν on F when restricted to Q is a p-adic
valuation logarithm;
(ii) the residue field F is finite and so L is also finite;
(iii) the elements of Gal(L/F) are isometric on L, noting that this is automatically satisfied if
F is complete.
L ∪ {0} of residue class representatives
Then for each g ∈ Gal(L/F) there exists a set RL,g ⊆ O×
for L such that the restriction of g to RL,g is an endofunction gRL,g : RL,g → RL,g.
Proof. Let ω be the extension of ν to L and let RL with 0 ∈ RL be an arbitrary set of
residue class representatives for L. Fix g ∈ Gal(L/F) and for a ∈ O×
L denote the orbit
of a with respect to g by
(cid:104)g(cid:105)(a) := {g(n)(a) : n ∈ {1,· · · , ord(g, a)}}.
Also denote (cid:104)g(cid:105)(a) := {g(n)(a) : n ∈ {1,· · · , ord(g, a)}} = (cid:104) ¯g(cid:105)(¯a) ⊆ L. We will show
that RL,g ⊆ O×
L ∪ {0} can be constructed from RL. Clearly we can let 0 represent ¯0 and
so include 0 in RL,g. More generally we need to make sure that:
72
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
(1) for each a(cid:48) ∈ RL there is precisely one element a ∈ RL,g such that ¯a = a(cid:48), that is
a = a(cid:48) + b for some b ∈ L with ω(b) > 0;
(2) for each a ∈ RL,g we have g(a) ∈ RL,g.
To this end we will show that the following useful facts hold for Lemma 5.4.12.
(a) For a1, a2 ∈ O×
L either (cid:104)g(cid:105)(a1) ∩ (cid:104)g(cid:105)(a2) = ∅ or (cid:104)g(cid:105)(a1) = (cid:104)g(cid:105)(a2). Clearly, since
ord(g) is finite, if (cid:104)g(cid:105)(a1) ∩ (cid:104)g(cid:105)(a2) (cid:54)= ∅ then (cid:104)g(cid:105)(a1) = (cid:104)g(cid:105)(a2).
(b) Let a(cid:48) ∈ RL\{0}. Then there exists a ∈ O×
L with ¯a = a(cid:48) such that if a1, a2 ∈ (cid:104)g(cid:105)(a)
with a1 (cid:54)= a2 then a1 (cid:54)= a2. Further since g is an isometry we have ω(a1) = 0 for
all a1 ∈ (cid:104)g(cid:105)(a). This ensures that every residue class that has a representative in
(cid:104)g(cid:105)(a) has only one representative in (cid:104)g(cid:105)(a).
√
√
Hence by applying (a) and (b) above we obtain RL,g as a disjoint union of the orbits of
finitely many elements form O×
L ∪ {0}. Note if RF is a set of residue class represen-
tatives for F and RF ⊆ RL then with the above construction we can choose to have
RF ⊆ RL,g since g restricts to the identity map on RF. Also note that (b) is not in gen-
eral satisfied for all a ∈ O×
L with ¯a = a(cid:48). Indeed, in the case of Example 5.3.3 where
√
2 we have g(a) = 1 − 5
2 (cid:54)= a and yet g(a) = ¯a = ¯1. We
L = Q5(
2), for a = 1 + 5
will now prove that (a) and (b) above hold.
For (a) it is enough to confirm that (cid:104)g(cid:105)(a) = (cid:104) ¯g(cid:105)(¯a) for all a ∈ O×
L . Since g is an isom-
etry, (iii) and (iv) of Remark 5.4.6 are applicable and so we have for each n ∈ N that
g(n)(a) = g(n)(¯a) = ¯g(n)(¯a). Hence the result follows.
For (b) we first note that, for each a(cid:48) ∈ RL\{0}, g maps residue class to residue class.
That is g restricts to a bijection ga(cid:48)
: a(cid:48) → g(a(cid:48)) and g also restricts to a bijection
gg(a(cid:48)) : g(a(cid:48)) → g(2)(a(cid:48)) and so forth. This is because g restricts to a bijective endo-
function on ML since g has finite order and is an isometry. Now for (b) to hold we
need to check that for each a(cid:48) ∈ RL\{0} there is an a ∈ a(cid:48) such that when the forward
orbit of a, with respect to g, returns to a residue class it has visited before then it returns
to the same element of that residue class. Let a(cid:48) ∈ RL\{0} and let n be the first element
of {1, 2,· · · , ord(g, a(cid:48))} such that there exists an i ∈ {0, 1,· · · , n − 1} with g(i)(a(cid:48)) in
the same residue class as g(n)(a(cid:48)). Hence ω(g(i)(a(cid:48)) − g(n)(a(cid:48))) > 0 and since g is an
isometry we have ω(a(cid:48) − g(n−i)(a(cid:48))) > 0 giving i = 0 by the definition of n. Therefore
g(n)(a(cid:48)) = a(cid:48) + b for some b ∈ ML and g(n) restricts to g(n)a(cid:48) : a(cid:48) → a(cid:48).
Hence for (b) to hold it is enough to show that there is a ∈ a(cid:48) which is a fixed point
with respect to g(n). To this end we more generally show that for each g ∈ Gal(L/F)
with ga(cid:48)
: a(cid:48) → a(cid:48) there is a fixed point a ∈ a(cid:48) of g. So for such a g ∈ Gal(L/F) let
m := ord(g, a(cid:48)). Now recall that we have Q ⊆ F and that ν on F when restricted to Q
73
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
is a p-adic valuation logarithm for some prime p. Hence we have two cases, p (cid:45) m and
pm.
Suppose p (cid:45) m. We have g(a(cid:48)) = a(cid:48) + b for some b ∈ ML. Further then we have
g(2)(a(cid:48)
g(3)(a(cid:48)
) =g(a(cid:48)
) =g(a(cid:48)
) + g(b) = a(cid:48)
+ b) = g(a(cid:48)
+ b + g(b)) = g(a(cid:48)
+ b + g(b),
) + g(b) + g(2)(b) = a(cid:48)
+ b + g(b) + g(2)(b),
...
) =a(cid:48)
g(m−1)(a(cid:48)
Hence consider
+ b + g(b) + g(2)(b) + · · · + g(m−2)(b).
=
))
1
1
a :=
) + g(2)(a(cid:48)
) + · · · + g(m−1)(a(cid:48)
m ) = 1
m giving g(a) = a. Moreover since p (cid:45) m we have
Since Q ⊆ F we have g( 1
ω(m−1) = ν(m−1) = −ν(m) = 0. Therefore
ω(a − a(cid:48)
+ g(a(cid:48)
+ (m − 1)b + (m − 2)g(b) + · · · + (m − (m − 1))g(m−2)(b)).
m (a(cid:48)
m (ma(cid:48)
(cid:18) 1
m ((m − 1)b + (m − 2)g(b) + (m − 3)g(2)(b) + · · · + g(m−2)(b))
) =ω
=0 + ω((m − 1)b + (m − 2)g(b) + (m − 3)g(2)(b) + · · · + g(m−2)(b))
≥ min{ω((m − 1)b), ω((m − 2)g(b)), ω((m − 3)g(2)(b)),· · · , ω(g(m−2)(b))}
=ω(g(m−2)(b)) = ω(b) > 0.
(cid:19)
Hence a is an element of a(cid:48) with g(a) = a as required.
Suppose pm. Then there is n, m(cid:48) ∈ N such that m = pnm(cid:48) with p (cid:45) m(cid:48). Hence
ord(g(pn−1m(cid:48)), a(cid:48)) = p. Now suppose that the following holds.
(b2) Let a0 ∈ O×
L . Then for each g(cid:48) ∈ Gal(L/F) with g(cid:48)a0 : a0 → a0 and ord(g(cid:48), a0) = p
there is a fixed point a1 ∈ a0 of g(cid:48).
Then by applying (b2) there is a1 ∈ a(cid:48) which is a fixed point of g(pn−1m(cid:48)) and so we have
ord(g, a1)pn−1m(cid:48). By repeated application of (b2) we can obtain an element an ∈ a(cid:48)
such that ord(g, an)m(cid:48). Now, since the set RL was an arbitrary set of residue class
representatives for L and an represents a(cid:48), we can apply the case p (cid:45) m for an to obtain
a ∈ an = a(cid:48) with g(a) = a as required.
It remains to show that (b2) holds. So let a0 ∈ O×
conditions of (b2). Hence for some b ∈ ML we have
L and g(cid:48) ∈ Gal(L/F) satisfy the
g(cid:48)
(a0) =a0 + b,
g(cid:48)(2)(a0) =a0 + b + g(cid:48)
(b),
...
g(cid:48)(p−1)(a0) =a0 + b + g(cid:48)
(b) + · · · + g(cid:48)(p−2)(b).
74
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
Define b1 := b, b2 := b + g(cid:48)(b), · · · , bp−1 := b + g(cid:48)(b) + · · · + g(cid:48)(p−2)(b) and note that
for all i ∈ {1,· · · , p − 1} we have
ω(bi) ≥ min{ω(b), ω(g(cid:48)
(5.4.2)
Now since ωQ is the p-adic valuation logarithm νp we have Fp ⊆ L. Therefore since L
is a finite field we have #L = pk for some k ∈ N. Hence we consider
(b)),· · · , ω(g(cid:48)(p−2)(b))} = ω(b) > 0.
a1 :=(a0g(cid:48)
(a0)g(cid:48)(2)(a0) · · · g(cid:48)(p−1)(a0))pk−1
=(a0(a0 + b1)(a0 + b2) · · · (a0 + bp−1))pk−1
0 + a0b1(a0 + b2) · · · (a0 + bp−1) + a2
=(ap
· · · + ap−1
bp−1)pk−1.
0
0b2(a0 + b3) · · · (a0 + bp−1) + · · ·
Now, by Lemma 2.1.7 and (5.4.2), we have
ω(a0b1(a0 + b2) · · · (a0 + bp−1)) =ω(a0) + ω(b1) + ω(a0 + b2) + · · ·
· · · + ω(a0 + bp−1)
=0 + ω(b1) + 0 + · · · + 0
≥ω(b) > 0.
The same inequality holds for later terms in the above expansion on a1, hence for
0b2(a0 + b3) · · · (a0 + bp−1) + · · · + ap−1
c := a0b1(a0 + b2) · · · (a0 + bp−1) + a2
bp−1
0
we have ω(c) > 0. This gives
(cid:19)
ap(pk−1−i)
0
ci
a1 = (ap
= apk
0 +
0 + c)pk−1
(cid:19)
ap(pk−1−i)
ci
0
=ω
i
pk−1
∑
i=1
(cid:18)pk−1
(cid:19)(cid:19)
(cid:19)(cid:19)
(cid:18)(cid:18)pk−1
(cid:18)(cid:18)pk−1
i
such that for each i ∈ {1,· · · , pk−1} we have
(cid:18)(cid:18)pk−1
(cid:19)
ω
i
+ p(pk−1 − i)ω(a0) + iω(c)
i
=ω
+ 0 + iω(c) > 0
(cid:17) ≥ 0 since ωQ is the p-adic valuation logarithm νp. Hence for
(cid:16)
(pk−1
noting that ω
i )
i )ap(pk−1−i)
c(cid:48) := ∑pk−1
i=1 (pk−1
= pk − 1 we have
pk−1 = ¯1 by Lagrange's theorem. In particular a1 = apk
pk
+ ¯0 = a0 giving
a0
a1 ∈ a0 and since a1 = (a0g(cid:48)(a0)g(cid:48)(2)(a0) · · · g(cid:48)(p−1)(a0))pk−1 with ord(g(cid:48), a0) = p we
have g(cid:48)(a1) = a1 as required. This completes the proof of Lemma 5.4.12.
×
ci we have ω(c(cid:48)) > 0. Further since #L
0 + c(cid:48) = a0
0
We will now prove Theorem 5.4.5.
75
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
Proof of Theorem 5.4.5. For f ∈ O(X, τ, g) let f
:= f + J (X, τ, g) denote the quotient
class to which f belongs and let π be a prime element of L. Note, O(X, τ, g)/J (X, τ, g)
is endowed with the usual quotient operations and the quotient norm which in this case
gives the trivial valuation. We begin by establishing a set R(X, τ, g) ⊆ O(X, τ, g) of
quotient class representatives for O(X, τ, g)/J (X, τ, g). By Lemma 5.4.12 there is a set
RL,g of residue class representatives for L such that gRL,g : RL,g → RL,g. Furthermore
by Lemma 5.4.10 every f ∈ CL(X) has a unique expansion of the form
f =
∞
∑
i=n
fiπi,
for some n ∈ Z,
(5.4.3)
where, for each i ≥ n, fi
expansion (5.4.3), for f ∈ O(X, τ, g) we have
: X → RL,g is a locally constant function. Hence, using
f0 ◦ τ + h ◦ τ = f ◦ τ = g ◦ f = g ◦ f0 + g ◦ h
i=1 fiπi with ω(h(x)) > 0 for all x ∈ X. Now note that τ : X → X and
where h := ∑∞
gRL,g : RL,g → RL,g give f0 ◦ τ : X → RL,g and g ◦ f0 : X → RL,g. Further since g is
an isometry on L we have ω(h ◦ τ(x)) > 0 and ω(g ◦ h(x)) > 0 for all x ∈ X. Hence
since the expansion of f ◦ τ in the for of (5.4.3) is unique we have f0 ◦ τ = g ◦ f0 and
h ◦ τ = g ◦ h. Moreover f0 is continuous since locally constant and, for x ∈ X,
(cid:40) ∞ if
0
if
f0(x) = 0
f0(x) (cid:54)= 0.
ω( f0(x)) =
In particular we have f0 ∈ O(X, τ, g). Hence we also have h = f − f0 ∈ O(X, τ, g) since
O(X, τ, g) is a ring. But since ω(h(x)) > 0 for all x ∈ X we in fact have h ∈ J (X, τ, g)
giving
Now by the uniqueness of expansions in the form of (5.4.3) and since J (X, τ, g) is an
ideal we have for any other element f (cid:48) = f (cid:48)
f (cid:48)
0 = f0. Hence using expansion (5.4.3) we define
f = (cid:101)f0.
0 + h(cid:48) ∈ O(X, τ, g) that (cid:101)f (cid:48) = f if and only if
(cid:40)
(cid:41)
R(X, τ, g) :=
f0 : f =
∞
∑
i=0
fiπi ∈ O(X, τ, g)
noting that 0 ∈ R(X, τ, g) since 0 ∈ RL,g and 0 ∈ O(X, τ, g). We now define a map
φ : O(X, τ, g)/J (X, τ, g) → C(X, τ, ¯g) by
φ( f ) = φ((cid:101)f0) = φ( f0 + J (X, τ, g)) := f0
where for x ∈ X we define f0(x) := f0(x) = f0(x) + ML. We show that φ is a ring
isomorphism by checking that:
76
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
(i) for all f ∈ O(X, τ, g)/J (X, τ, g) we have f0 ∈ C(X, τ, ¯g);
(ii) φ is multiplicative, linear and φ(1) = ¯1;
(iii) ker(φ) = {0} ensuring that φ is injective;
(cid:16)
(iv) φ is surjective.
For (i), since f0 is a locally constant function on X we have f0 ∈ CL(X). Furthermore
we have already shown above that f0 ◦ τ = g ◦ f0. Hence for each x ∈ X we have
For (ii), let f ,(cid:101)f (cid:48) ∈ O(X, τ, g)/J (X, τ, g). We show that φ is multiplicative. Set h := f0 f (cid:48)
f0(τ(x)) = f0(τ(x)) = g( f0(x)) = ¯g
0, h0 ∈ R(X, τ, g). Hence for each
giving f0 f (cid:48)
x ∈ X we have h(cid:48)(x) ∈ ML. Therefore for each x ∈ X we have
0 = h0 + h(cid:48) with h(cid:48) ∈ J (X, τ, g) and f0, f (cid:48)
φ( f(cid:101)f (cid:48))(x) = φ((cid:101)f0(cid:101)f (cid:48)
and so f0 ∈ C(X, τ, ¯g).
f0(x)
f0(x)
(cid:17)
(cid:16)
(cid:17)
= ¯g
0
0)(x) =φ((cid:103)f0 f (cid:48)
=φ((cid:101)h0)(x)
0)(x)
=h0(x)
=h0(x)
=h0(x) + h(cid:48)(x)
= f0(x) f (cid:48)
= f0(x) f (cid:48)
=(φ((cid:101)f0)φ((cid:101)f (cid:48)
0(x)
0(x)
0))(x) = (φ( f )φ((cid:101)f (cid:48)))(x).
Linearity, φ( f + (cid:101)f (cid:48)) = φ( f ) + φ((cid:101)f (cid:48)), is shown in much the same way. Showing that
10 ∈ R(X, τ, g) giving φ(1) = φ((cid:101)10) = 10 = ¯1 as required. In fact we can always choose
φ(1) = ¯1 is almost immediate. Let 10 be the representative in RL,g of ¯1. Then we have
If for all x ∈ X we have φ((cid:101)f0)(x) = f0(x) = ¯0 then
RL,g such that 10 = 1.
ω( f0(x)) > 0 for all x ∈ X giving f0 ∈ J (X, τ, g). Hence (cid:101)f0 = 0 and so ker(φ) = {0}.
For (iii), let f ∈ O(X, τ, g).
In fact since f0 is an element of R(X, τ, g) we have f0 = 0 in this case.
For (iv), given ¯f ∈ C(X, τ, ¯g) and x ∈ X we have ¯f (x) = a0(x) + ML for some element
a0(x) of RL,g since RL,g is a set of residue class representatives for L. Since the valuation
on L is the trivial valuation, ¯f is a locally constant function and hence, when viewed as
a function on X, so is a0. Therefore a0 is a continuous OL valued function noting that
RL,g ⊆ OL. Further since ¯f ∈ C(X, τ, ¯g) we have for each x ∈ X that
a0(τ(x)) = ¯f (τ(x)) = ¯g( ¯f (x)) = ¯g
a0(x)
= g(a0(x)).
Now because gRL,g : RL,g → RL,g we have g(a0(x)) ∈ RL,g giving a0(τ(x)) = g(a0(x))
for all x ∈ X. Hence, as a function on X, a0 ∈ O(X, τ, g) and so a0 ∈ R(X, τ, g) with
77
(cid:16)
(cid:17)
CHAPTER 5: COMMUTATIVE GENERALISATION OVER COMPLETE VALUED FIELDS
φ((cid:101)a0) = a0 = ¯f as required. Finally, since the valuation on L is the trivial valuation,
the sup norm on C(X, τ, ¯g) is the trivial norm. Therefore it is immediate that φ is an
isometry and this completes the proof of Theorem 5.4.5.
The last result of this section follows easily from the preceding results.
Corollary 5.4.13. Let F, L and g ∈ Gal(L/F) conform to Lemma 5.4.12 with F and L having
complete nontrivial discrete valuations and let C(X, τ, g) be a basic L/Lg function algebra.
Further let R(X, τ, g) ⊆ O(X, τ, g) be the subset of all locally constant RL,g valued functions.
If there is a prime element π of L such that g(π) = π then each f ∈ C(X, τ, g)\{0} has a
unique series expansion of the form
f =
∞
∑
i=n
fiπi,
for some n ∈ Z,
where for each i ≥ n we have fi ∈ R(X, τ, g). In particular the subset of all locally constant
functions, LC(X, τ, g) ⊆ C(X, τ, g), is uniformly dense in C(X, τ, g).
Remark 5.4.14. For L an unramified extension of F, every prime element π ∈ F is a
prime element of L with g(π) = π. In particular Corollary 5.4.13 holds when L is a
finite unramified extension of Qp as is the case for examples 5.3.3 and 5.3.4.
Proof of Corollary 5.4.13. Let f be an element of C(X, τ, g)\{0} and let π be a prime
element of L with g(π) = π. By Lemma 5.4.10 f has a unique series expansion of the
form
f =
fiπi,
∞
∑
i=n
for some n ∈ Z,
with fn (cid:54)= 0 and fi : X → RL,g a locally constant function for all i ≥ n. Hence
∞
∑
i=n
fi ◦ τπi = f ◦ τ = g ◦ f =
(g ◦ fi)(g(π))i =
∞
∑
i=n
g ◦ fiπi.
∞
∑
i=n
Therefore since g restricts to an endofunction on RL,g and by the uniqueness of the
expansion we have fi ◦ τ = g◦ fi for all i ≥ n. Hence fi is an element of R(X, τ, g) for all
i ≥ n and for each m ∈ N we have ∑n+m−1
m
is a sequence of locally constant functions which converges uniformly to f as required.
fiπi ∈ C(X, τ, g). Finally
fiπi(cid:17)
(cid:16)∑n+m−1
i=n
i=n
This brings us to the end of Chapter 5. In the next chapter we will see that L/Lg function
algebras have a part to play in representation theory.
78
CHAPTER 6
Representation theory
The first section of this chapter introduces several results from the theory of Banach
rings and Banach F-algebras that we will use later in the chapter. These results have
been taken from [Ber90, Ch1]. However I have provided a thorough proof of each result
in order to give significantly more detail than [Ber90] since some of them may not be
widely known. The second section begins by recalling which Banach F-algebras can be
represented by complex uniform algebras or real function algebras in the Archimedean
setting and one such result in the non-Archimedean setting provided by [Ber90] is also
noted. We then develop this theory further by identifying a large class of Banach F-
algebras that can be represented by L/Lg function algebras. The resulting representa-
tion theorem is the main result of interest in this chapter and the rest of the chapter is
given over to the proof of the theorem.
6.1 Further Banach rings and Banach F-algebras
Since the definition of a Banach ring was given in Definition 3.2.1 we begin with the
first lemma.
Lemma 6.1.1. Let R be a Banach ring and let r ∈ R be positive. Define
R(cid:104)r−1T(cid:105) := { f =
aiTi : ai ∈ R and
∞
∑
i=0
(cid:107)ai(cid:107)Rri < ∞}.
∞
∑
i=0
Then with the Cauchy product and usual addition:
(i) we have that R(cid:104)r−1T(cid:105) is a Banach ring with respect to the norm
(cid:107) f(cid:107)R,r :=
(cid:107)ai(cid:107)Rri;
∞
∑
i=0
(ii) for a ∈ R we have 1 − aT invertible in R(cid:104)r−1T(cid:105) if and only if ∑∞
i=0 (cid:107)ai(cid:107)Rri < ∞.
79
CHAPTER 6: REPRESENTATION THEORY
Proof. For (i), let f1 = ∑∞
i=0 biTi be elements of R(cid:104)r−1T(cid:105). Then
i=0 aiTi and f2 = ∑∞
(cid:107) f1 + f2(cid:107)R,r =
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
(ai + bi)Ti
(cid:107)ai + bi(cid:107)Rri
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
=
∑
i=0
∞
∑
i=0
≤ ∞
∑
i=0
∞
∑
i=0
((cid:107)ai(cid:107)R + (cid:107)bi(cid:107)R) ri
(cid:107)ai(cid:107)Rri +
(cid:107)bi(cid:107)Rri
=
=(cid:107) f1(cid:107)R,r + (cid:107) f2(cid:107)R,r < ∞
∞
∑
i=0
showing that R(cid:104)r−1T(cid:105) is closed under addition and that the triangle inequality holds
for (cid:107) · (cid:107)R,r. Clearly (cid:107) f(cid:107)R,r = 0 if and only if f = 0. Further
(cid:107) f1 f2(cid:107)R,r =
Ti
akbi−k
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
(cid:33)
(cid:33)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R
(cid:107)ak(cid:107)R(cid:107)bi−k(cid:107)R
(cid:33)(cid:32) ∞
ri
akbi−k
k=0
=
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
(cid:32) i∑
∑
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) i∑
i=0
∞
∑
(cid:32) i∑
i=0
≤ ∞
∑
(cid:32) ∞
i=0
∑
i=0
k=0
k=0
(cid:107)ai(cid:107)Rri
∑
=
i=0
=(cid:107) f1(cid:107)R,r(cid:107) f2(cid:107)R,r < ∞
ri
(cid:33)
, by Mertens' Theorem, see [Apo74, p204]
(cid:107)bi(cid:107)Rri
showing that R(cid:104)r−1T(cid:105) is closed under multiplication and (cid:107) · (cid:107)R,r is sub-multiplicative.
Furthermore we have 1R,r = 1RT0 which gives (cid:107)1R,r(cid:107)R,r = (cid:107)1R(cid:107)Rr0 = 1 and similarly
(cid:107) − 1R,r(cid:107)R,r = (cid:107) − 1R(cid:107)Rr0 = 1.
We now show that R(cid:104)r−1T(cid:105) is complete. Let(cid:0)∑∞
i=0 ai,nTi(cid:1)
i=0 ai,mTi − ∑∞
n be a Cauchy sequence in
R(cid:104)r−1T(cid:105). Then for k ∈ N0 and ε > 0 there exists M ∈ N such that for all m, m(cid:48) ≥ M
we have (cid:107) ∑∞
i=0 (cid:107)ai,m − ai,m(cid:48)(cid:107)Rri < εrk. Hence for all
m, m(cid:48) ≥ M we have (cid:107)ak,m − ak,m(cid:48)(cid:107)R < ε and so for all k ∈ N0, (ak,n)n is a Cauchy
sequence in R. Since R is a Banach ring (ak,n)n converges to some bk ∈ R.
We show that ∑∞
such that for all m ≥ M we have
i=0 biTi is an element of R(cid:104)r−1T(cid:105). Let ε0 > 0. Then there exists M ∈ N
i=0 ai,m(cid:48)Ti(cid:107)R,r = ∑∞
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
∑
i=0
ai,mTi
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
∑
i=0
≤
<ε0 +
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
∑
i=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
ai,MTi
+
ai,MTi
ai,mTi − ∞
∑
i=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
∑
i=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
ai,MTi
80
< ∞.
CHAPTER 6: REPRESENTATION THEORY
Now let N ∈ N0 and ε > 0. Since for each k ∈ N0, (ak,n)n is a Cauchy sequence in R
with limit bk, there is M(cid:48) ∈ N such that for all m(cid:48) ≥ M(cid:48) we have ∑N
i=0 (cid:107)bi − ai,m(cid:48)(cid:107)Rri < ε.
Hence letting m0 ≥ max{M, M(cid:48)} gives
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N∑
i=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
biTi
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N∑
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N∑
i=0
i=0
=
≤
ai,m0Ti +
ai,m0Ti
biTi − N∑
i=0
(bi − ai,m0)Ti
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
∑
i=0
N∑
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
i=0
∑
i=0
+
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
.
ai,m0Ti
<ε + ε0 +
ai,MTi
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
i=0 ai,MTi(cid:107)R,r. Since
i=0 ai,MTi(cid:107)R,r giving
Since ε > 0 was arbitrary we have (cid:107) ∑N
this holds for each N ∈ N0 we have (cid:107) ∑∞
i=0 biTi ∈ R(cid:104)r−1T(cid:105) as required.
∑∞
Let ε > 0. We will show, for large enough n ∈ N, that (cid:107) ∑∞
and so R(cid:104)r−1T(cid:105) is complete. Since (cid:0)∑∞
i=0 ai,nTi(cid:1)
i=0 biTi(cid:107)R,r ≤ ε0 + (cid:107) ∑∞
i=0 biTi(cid:107)R,r ≤ ε0 + (cid:107) ∑∞
i=0 ai,nTi − ∑∞
i=0 biTi(cid:107)R,r < ε
n is a Cauchy sequence there exists
i=0(ai,n − ai,m)Ti(cid:107)R,r < ε/4.
i=0 ai,nTi and ∑∞
i=N+1 ai,nTi(cid:107)R,r < ε/4 and (cid:107) ∑∞
i=0 biTi are elements of R(cid:104)r−1T(cid:105) there exists N ∈ N
M1 ∈ N such that for all m, n ≥ M1 we have (cid:107) ∑∞
Let n ≥ M1. Since ∑∞
such that (cid:107) ∑∞
Since for each i ∈ N0, (ai,m)m is a Cauchy sequence in R with limit bi, there is M2 ∈ N
such that for all m ≥ M2 we have (cid:107) ∑N
i=0 (cid:107)ai,m − bi(cid:107)Rri < ε/4.
Let m = max{M1, M2} and define cn := (cid:107) ∑∞
(ai,n − ai,m)Ti +
i=0(ai,m − bi)Ti(cid:107)R,r = ∑N
i=N+1 biTi(cid:107)R,r < ε/4.
i=0 biTi(cid:107)R,r, then
i=0 ai,nTi − ∑∞
ai,nTi +
cn =
biTi
N∑
N∑
(ai,m − bi)Ti − ∞
∑
i=N+1
i=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
∑
i=N+1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R,r
i=0
<ε/4 + ε/4 + ε/4 + ε/4 = ε as required.
For (ii), the result is obvious for a = 0 and so suppose a (cid:54)= 0. If ∑∞
i=0 (cid:107)ai(cid:107)Rri < ∞ then
i=0 aiTi is an element of R(cid:104)r−1T(cid:105) and by the definition of the Cauchy product we have
∑∞
(cid:33)
(cid:32) ∞
∑
i=0
aiTi
(1 − aT) =(a01R)T0 +
(ai1R + ai−1(−a))Ti
∞
∑
i=1
ai−1(a + (−a))Ti
=1RT0 +
∞
∑
i=1
Similarly this holds for (1 − aT)(∑∞
Now conversely if 1− aT is invertible in R(cid:104)r−1T(cid:105) then for ∑∞
in R(cid:104)r−1T(cid:105) we have by the definition of the Cauchy product
=1RT0 = 1R,r = 1.
i=0 aiTi) and so 1 − aT is invertible.
i=0 biTi the inverse of 1− aT
(cid:33)
(cid:32) ∞
∑
i=0
1R,r =
biTi
(1 − aT) = (b01R)T0 +
(bi1R + bi−1(−a))Ti.
∞
∑
i=1
81
CHAPTER 6: REPRESENTATION THEORY
Hence b0 = 1R = a0 and, for each i ∈ N, 0 = bi + bi−1(−a) giving
bi−1a = bi + bi−1a + bi−1(−a) = bi + bi−1(a + (−a)) = bi.
i=0 biTi is an element of R(cid:104)r−1T(cid:105) and so ∑∞
Therefore for each i ∈ N, bi = bi−1a with b0 = 1R giving bi = ai by induction. Hence
∑∞
i=0 aiTi = ∑∞
Remark 6.1.2. Since R(cid:104)r−1T(cid:105) extends R as a ring and by the definition of the Cauchy
product it is immediate that R(cid:104)r−1T(cid:105) is commutative if and only if R is commutative.
Similarly by the definition of the norm (cid:107) · (cid:107)A,r and the Cauchy product, if A is a unital
Banach F-algebra then A(cid:104)r−1T(cid:105) is also a unital Banach F-algebra. These details are
easily checked.
i=0 (cid:107)ai(cid:107)Rri < ∞ as required.
The following definitions will be used many times in this chapter.
Definition 6.1.3. Let R be a Banach ring. A bounded multiplicative seminorm on R is a
map · : R → R taking non-negative values that is:
(1) bounded, a ≤ (cid:107)a(cid:107)R for all a ∈ R, but not constantly zero on R;
(2) multiplicative, ab = ab for all a, b ∈ R and hence 1R = 1 by setting a = 1R
and b (cid:54)∈ ker( · );
(3) a seminorm and so · also satisfies the triangle inequality and 0 ∈ ker( · ) but
the kernel is not assumed to be a singleton.
Definition 6.1.4. Let F be a complete non-Archimedean field and let A be a commuta-
tive unital Banach F-algebra. In this chapter M0(A) will denote the set of all proper
closed prime ideals of A that are the kernels of bounded multiplicative seminorms on
A. For x0 ∈ M0(A), or any proper closed ideal of A, we will denote the quotient norm
on A/x0 by · x0 that is a + x0x0 := inf{(cid:107)a + b(cid:107)A : b ∈ x0} for a ∈ A.
We now proceed with a number of Lemmas. In particular towards the end of Section
6.1 it will be show that M0(A) is always nonempty.
Lemma 6.1.5. Let A be a unital Banach F-algebra. If · is a bounded multiplicative seminorm
on A as a Banach ring then we have α = αF for all α ∈ F. Hence since · is multiplicative
it is also a vector space seminorm, that is αa = αFa for all a ∈ A and α ∈ F.
Proof. For α ∈ F× we first note that 1 = 1A = αα−1 = αα−1 and so α (cid:54)= 0 and
α−1 = α−1. Similarly α−1F = α−1
since · F is a valuation. Moreover since · is
bounded we have α ≤ (cid:107)α1A(cid:107)A = αF(cid:107)1A(cid:107)A = αF. Since this holds for all α ∈ F×
we also have α−1 ≤ α−1F giving αF ≤ α and so α = αF for all α ∈ F.
F
82
CHAPTER 6: REPRESENTATION THEORY
Lemma 6.1.6. Let F and A be as in Definition 6.1.4. For x0 ∈ M0(A), or any proper closed
ideal of A, the following holds:
(i) the quotient ring A/x0 has F ⊆ A/x0 and is an integral domain if x0 is prime;
(ii) the quotient norm is such that α + x0x0 = αF for all α ∈ F;
(iii) the quotient norm · x0 is an F-vector space norm on A/x0, opposed to being merely a
seminorm, and it is sub-multiplicative;
(iv) if (cid:107) · (cid:107)A is square preserving, that is (cid:107)a2(cid:107)A = (cid:107)a(cid:107)2
A for all a ∈ A, then both (cid:107) · (cid:107)A and
· x0 observe the strong triangle inequality noting that F is non-Archimedean;
(v) by way of the map a (cid:55)→ a + x0x0, as a seminorm on A, · x0 is bounded.
Proof. For (i), if a1 + x0, a2 + x0 ∈ A/x0 with (a1 + x0)(a2 + x0) = a1a2 + x0 = 0 + x0
then we have a1a2 ∈ x0. Hence if x0 is a prime ideal of A then at least one of a1 + x0
and a2 + x0 is equal to 0 + x0 and so A/x0 is an integral domain. It is immediate that
A/x0 has a subset that is an isomorphic copy of F since x0 is a proper ideal of A.
For (ii), we first show that 1 + x0x0 = 1. Note that 1 + x0x0 ≤ (cid:107)1(cid:107)A = 1 since 0 ∈ x0.
So now suppose toward a contradiction that there is b ∈ x0 such that (cid:107)1 + b(cid:107)A < 1.
We have for all n ∈ N that bn := −((1 + b)n − 1) is an element of x0 since x0 is an
ideal of A. But (cid:107)1 − bn(cid:107)A = (cid:107)(1 + b)n(cid:107)A ≤ (cid:107)1 + b(cid:107)n
A = 0 and
so 1 is an element of x0 since x0 is closed which contradicts x0 being a proper ideal
of A. We conclude that (cid:107)1 + b(cid:107)A ≥ 1 for all b ∈ x0 and so 1 + x0x0 ≥ 1. Hence
1 + x0x0 = 1 by the above. Now for α ∈ F× we have x0 = αx0 since α is invertible
where αx0 := {αb : b ∈ x0}. Hence
A with limn→∞ (cid:107)1 + b(cid:107)n
α + x0x0 = inf{(cid:107)α + b(cid:107)A : b ∈ x0}
= inf{(cid:107)α + αb(cid:107)A : b ∈ x0}
= inf{αF(cid:107)1 + b(cid:107)A : b ∈ x0}
=αF1 + x0x0 = αF
as required. In a similar way for a ∈ A one shows that αa + x0x0 = αFa + x0x0.
For (iii), we note that · x0 is a norm on A/x0 because x0 is closed as a subset of A so
that for a ∈ A\x0 there is ε > 0 with (cid:107)a + b(cid:107)A ≥ ε for all b ∈ x0 giving a + x0x0 ≥ ε.
We now show that · x0 is sub-multiplicative. For a1, a2 ∈ A we have
a1a2 + x0x0 = inf{(cid:107)a1a2 + b(cid:107)A : b ∈ x0}
≤ inf{(cid:107)a1a2 + a1b2 + a2b1 + b1b2(cid:107)A : b1, b2 ∈ x0}
≤ inf{(cid:107)a1 + b1(cid:107)A(cid:107)a2 + b2(cid:107)A : b1, b2 ∈ x0}
=a1 + x0x0a2 + x0x0.
83
CHAPTER 6: REPRESENTATION THEORY
For (iv), suppose (cid:107) · (cid:107)A is square preserving. In this case the proof of Theorem 2.1.2
also works for A and so (cid:107) · (cid:107)A observes the strong triangle inequality, see [Sch06, p18]
for details. Hence for a1, a2 ∈ A we also have
a1 + a2 + x0x0 = inf{(cid:107)a1 + a2 + b(cid:107)A : b ∈ x0}
= inf{(cid:107)a1 + b1 + a2 + b2(cid:107)A : b1, b2 ∈ x0}
≤ inf{max{(cid:107)a1 + b1(cid:107)A,(cid:107)a2 + b2(cid:107)A} : b1, b2 ∈ x0}
= max{inf{(cid:107)a1 + b1(cid:107)A : b1 ∈ x0}, inf{(cid:107)a2 + b2(cid:107)A : b2 ∈ x0}}
= max{a1 + x0x0,a2 + x0x0}.
For (v), since we have 0 ∈ x0 it is immediate that a + x0x0 ≤ (cid:107)a(cid:107)A for all a ∈ A.
Lemma 6.1.7. Let R be a commutative Banach ring. Then:
(i) if a ∈ R has (cid:107)1 − a(cid:107)R < 1 then a is invertible in R;
(ii) for I a proper ideal of R the closure J of I, as a subset of R, is a proper ideal of R;
(iii) each non-invertible element of R is an element of some maximal ideal of R. The maximal
ideals of R are proper, closed and prime.
Proof. For (i), for a ∈ R with (cid:107)1 − a(cid:107)R < 1 let δ > 0 be such that (cid:107)1 − a(cid:107)R < δ < 1.
Then setting b := 1 − a gives (cid:107)bn(cid:107)R ≤ (cid:107)b(cid:107)n
R < δn < 1 for all n ∈ N. Therefore we have
n=0 bn ∈ R since R is complete.
n=0 (cid:107)bn(cid:107)R < ∑∞
∑m
Moreover
1−δ for each m ∈ N and so ∑∞
n=0 δn = 1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R
∞
∑
n=0
bn − (1 − b)
(cid:33)
bn + (1 − b)
m∑
n=0
bn − 1
m∑
n=0
(cid:107)bn(cid:107)R
(cid:33)
(cid:107)1 − b(cid:107)R + (cid:107)bm+1(cid:107)R.
(cid:33)
(cid:107)bn(cid:107)R
(cid:107)1 − b(cid:107)R + (cid:107)bm+1(cid:107)R
= 0.
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(1 − b)
bn − 1
∞
∑
n=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(1 − b)
(cid:32) ∞
∑
n=m+1
=
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)R
(cid:32)(cid:32) ∞
≤
∑
n=m+1
lim
m→∞
Hence a is invertible in R since 1 − b = 1 − (1 − a) = a and
For (ii), let I be a proper ideal of R and J its closure as a subset. For a, b ∈ J there are
sequences (an), (bn) of elements of I converging to a and b respectively with respect to
(cid:107) · (cid:107)R. Hence for a(cid:48) ∈ R, (a(cid:48)an) is a sequence in I with (cid:107)a(cid:48)a − a(cid:48)an(cid:107)R ≤ (cid:107)a(cid:48)(cid:107)R(cid:107)a − an(cid:107)R
for each n ∈ N and so a(cid:48)a ∈ J since limn→∞ (cid:107)a − an(cid:107)R = 0. Similarly (an + bn) is a
sequence in I with (cid:107)(a + b) − (an + bn)(cid:107)R ≤ (cid:107)a − an(cid:107)R + (cid:107)b − bn(cid:107)R for each n ∈ N and
so a + b ∈ J. Hence J is an ideal of R. Now since I is a proper ideal of R each element an
of the sequence (an) is not invertible and so (cid:107)1 − an(cid:107)R ≥ 1 for all n ∈ N by (i). Hence
84
CHAPTER 6: REPRESENTATION THEORY
1 ≤ (cid:107)1 − a + a − an(cid:107)R ≤ (cid:107)1 − a(cid:107)R + (cid:107)a − an(cid:107)R for all n ∈ N and so (cid:107)1 − a(cid:107)R ≥ 1 for
all a ∈ J giving 1 (cid:54)∈ J. Hence J is proper.
For (iii), let a be a non-invertible element of R noting that we can always take a = 0.
The principal ideal Ia := aR is proper since for all b ∈ R, ab (cid:54)= 1. By Zorn's lemma Ia is
a subset of some maximal ideal Ja of R. Every maximal ideal J of R is proper and prime,
noting that R/J is a field or by other means, and closed as a subset of R by (ii).
Remark 6.1.8. We note that if A is a commutative unital Banach F-algebra then Lemma
6.1.7 applies to A and A(cid:104)r−1T(cid:105) for each r > 0.
Lemma 6.1.9. Let F be a complete non-Archimedean field and let A be a commutative unital
Banach F-algebra with maximal ideal m0. Let S(A) denote the set of all norms on the field
A/m0 that are also unital bounded seminorms on A as a Banach ring. That is if · is an
element of S(A) then 1 = 1 and · conforms to Definition 6.1.3 except it need not be
multiplicative merely sub-multiplicative. It follows that:
(i) the set S(A) is non-empty;
(ii) for · ∈ S(A), A/m0 the completion of A/m0 with respect to ·, r > 0 and a ∈ A/m0,
if a − T is non-invertible in A/m0(cid:104)r−1T(cid:105) then there is · (cid:48) ∈ S(A) with a(cid:48) ≤ r and
b(cid:48) ≤ b for all b ∈ A/m0.
Proof. For (i), we note that the quotient norm · m0 is an element of S(A) since (iii) of
Lemma 6.1.7 shows that (ii), (iii) and (v) of Lemma 6.1.6 apply to · m0.
For (ii), suppose that a − T is non-invertible in A/m0(cid:104)r−1T(cid:105). Then a − T is an element
of some maximal ideal J of A/m0(cid:104)r−1T(cid:105) by Lemma 6.1.7. Hence the quotient norm · J
on A/m0(cid:104)r−1T(cid:105)/J is an element of S(A/m0(cid:104)r−1T(cid:105)) by Lemma 6.1.6. Therefore since J
is closed and A/m0 is a field, a(cid:48)(cid:48) := a(cid:48) + JJ, for a(cid:48) ∈ A/m0, defines a norm on A/m0.
Since · J is an element of S(A/m0(cid:104)r−1T(cid:105)) we have that · (cid:48) is unital as a seminorm
on A. Similarly since · J is bounded as a seminorm on A/m0(cid:104)r−1T(cid:105) we have for all
a(cid:48) ∈ A that
a(cid:48)
(6.1.1)
noting that · is an element of S(A). Hence · (cid:48) is bounded as a seminorm on A and
so · (cid:48) is an element of S(A). Further since a − T is an element of J we have
+ m0(cid:107)A/m0,r = a(cid:48)
+ m0 ≤ (cid:107)a(cid:48)(cid:107)A
+ m0(cid:48)
= (a(cid:48)
+ m0) + JJ ≤ (cid:107)a(cid:48)
a(cid:48)
= a + JJ = T + JJ ≤ (cid:107)T(cid:107)A/m0,r = r.
Finally we have b(cid:48) ≤ b for all b ∈ A/m0 by (6.1.1).
The following Lemma will be particularly useful in Section 6.2.
85
CHAPTER 6: REPRESENTATION THEORY
Lemma 6.1.10. Let F be a complete non-Archimedean field and let A be a commutative unital
Banach F-algebra. With reference to Definition 6.1.4 the following holds:
(i) the set M0(A) is non-empty since every maximal ideal of A is an element of M0(A);
(ii) an element a ∈ A is invertible if and only if a + x0 (cid:54)= 0 + x0 for all x0 ∈ M0(A).
Proof. Whilst this proof provides more detail, much of the following has been taken
from [Ber90, Ch1]. For (i), let m0 be a maximal ideal of A. Hence the quotient ring
A/m0 is a field. Let S(A) be as in Lemma 6.1.9 and note therefore that S(A) is non-
empty. We put a partial order on S(A) by · (cid:46) · (cid:48) if and only if a + m0 ≤ a + m0(cid:48)
for all a + m0 ∈ A/m0. Now let E be a chain in S(A), that is E is a subset of S(A) such
that (cid:46) restricts to a total order on E. Define a map · E : A/m0 → R by
a + m0E := inf{a + m0 : · ∈ E}.
We will show that · E is a lower bound for E in S(A). It is immediate from the defini-
tion of · E that it is unital and bounded since all of the elements of E are. Hence it suf-
fices to show that · E is a sub-multiplicative norm on A/m0. Clearly 0 + m0E = 0 so,
0 . We show that aE (cid:54)= 0.
simplifying our notation slightly, let a be an element of A/m×
Let · be an element of E and suppose towards a contradiction that there is · (cid:48) ∈ E
such that a(cid:48) < min{a,a−1−1}. Then
1 = 1(cid:48)
= aa−1(cid:48) ≤ a(cid:48)a−1(cid:48)
< a−1−1a−1(cid:48).
Hence by the above we have a−1 < a−1(cid:48) and a(cid:48) < a giving ·(cid:48) (cid:54)(cid:46) · and · (cid:54)(cid:46) ·(cid:48)
which contradicts both · and · (cid:48) being elements of E. Therefore for all · (cid:48) ∈ E we
have a(cid:48) ≥ min{a,a−1−1}. In particular aE (cid:54)= 0. Now for a, b ∈ A/m0 we have
a + bE = inf{a + b : · ∈ E}
≤ inf{a + b : · ∈ E}
= inf{a + b(cid:48) : · , · (cid:48) ∈ E},
=aE + bE,
†
where line † follows from the line above it because if · (cid:46) ·(cid:48) then a +b ≤ a +b(cid:48).
Hence the triangle inequality holds for · E. Similarly we have abE ≤ aEbE and so
· E is sub-multiplicative as required. Hence · E is a lower bound for E in S(A).
Therefore by Zorn's lemma there exists a minimal element of S(A) with respect to (cid:46).
Let · be a minimal element of S(A) and denote by A/m0 the completion of A/m0
with respect to · . We will show that · is multiplicative on A/m0 and hence satisfies
(i) of Lemma 6.1.10. Note that for now we should only take A/m0 to be an integral
86
CHAPTER 6: REPRESENTATION THEORY
domain and not a field since we can't apply Theorem 2.1.4.
Now since A/m0 is a field · will be multiplicative if a−1 = a−1 for all a ∈ A/m×
since for a, b ∈ A/m×
0 with a−1 = a−1 we have b = baa−1 ≤ baa−1 = baa−1
giving ab ≤ ab and since · is sub-multiplicative we have ab = ab. Hence we
will show that a−1 = a−1 for all a ∈ A/m×
0 . To this end we first show that · is
power multiplicative that is an = an for all a ∈ A/m0 and n ∈ N. Suppose towards
a contradiction that there is a ∈ A/m0 with an < an for some n > 1. We claim that
a − T is non-invertible in the Banach ring A/m0(cid:104)r−1T(cid:105) with r := n(cid:112)an. By Lemma
i=0 a−iri does not converge. Expressing i as
6.1.1 it suffices to show that the series ∑∞
i = pn + q, for some q ∈ {0,· · · , n − 1}, we have ai ≤ anpaq and ai−1 ≤ a−i
since 1 = aia−i ≤ aia−i. Therefore
0
a−iri ≥ ai−1anp+
q
n ≥ anpan q
anpaq =
n
n
an q
aq .
Hence a−iri ≥ min{an q
n
aq
is non-invertible in A/m0(cid:104)r−1T(cid:105) with r := n(cid:112)an. Now by Lemma 6.1.9 there exists
gives a(cid:48) ≤ r = n(cid:112)an < a which contradicts · being a minimal element of S(A).
· (cid:48) ∈ S(A) such that a(cid:48) ≤ r and b(cid:48) ≤ b for all b ∈ A/m0. But, since an < an, this
: q ∈ {0,· · · , n − 1}} > 0 for all i ≥ 0. Therefore a − T
Hence we have shown that an = an for all a ∈ A/m0 and n ∈ N.
Now suppose towards a contradiction that there exists an element a ∈ A/m×
0 with
a−1 < a−1. We claim that a − T is non-invertible in A/m0(cid:104)r−1T(cid:105) with r := a−1−1.
i=0 a−iri does not converge.
Again by Lemma 6.1.1 it suffices to show that the series ∑∞
Indeed since · is power multiplicative we have
a−iri = (a−1)iri = a−1i(a−1−1)i = a−10 = 1.
Hence a − T is non-invertible in A/m0(cid:104)r−1T(cid:105) with r := a−1−1. Now again by Lemma
6.1.9 there exists · (cid:48) ∈ S(A) such that a(cid:48) ≤ r and b(cid:48) ≤ b for all b ∈ A/m0. But,
since a−1 < a−1, this gives a(cid:48) ≤ r = a−1−1 < a which contradicts · being a
minimal element of S(A). Hence we have shown that a−1 = a−1 for all a ∈ A/m×
and so · is multiplicative. Finally m0 is the kernel of · and as a maximal ideal of A
it is proper, closed and prime by Lemma 6.1.7. In particular since m0 was an arbitrary
maximal ideal of A every maximal ideal of A is an element of M0(A).
For (ii), for a an invertible element of A we have a (cid:54)∈ x0 for all x0 ∈ M0(A) since x0 is
a proper ideal of A. Hence a + x0 (cid:54)= 0 + x0 in A/x0 for all x0 ∈ M0(A). On the other
hand for a a non-invertible element of A we have by Lemma 6.1.7 that a is an element
of a maximal ideal Ja of A. By (i) above, Ja is an element of M0(A) and a + Ja = 0 + Ja
in A/Ja. Therefore for a a non-invertible element of A we do not have a + x0 (cid:54)= 0 + x0
in A/x0 for all x0 ∈ M0(A).
0
87
CHAPTER 6: REPRESENTATION THEORY
With the preceding theory in place we can now turn our attention to the main topic of
this chapter.
6.2 Representations
6.2.1 Established theorems
The particular well known representation theorems in the Archimedean setting that we
will find an analog of in the non-Archimedean setting are as follows. See [KL92, p35]
for details of Theorem 6.2.2.
Theorem 6.2.1. Let A be a commutative unital complex Banach algebra with (cid:107)a2(cid:107)A = (cid:107)a(cid:107)2
for all a ∈ A. Then A is isometrically isomorphic to a uniform algebra on some compact
Hausdorff space X, in other words a C/C function algebra on (X, id, id).
Theorem 6.2.2. Let A be a commutative unital real Banach algebra with (cid:107)a2(cid:107)A = (cid:107)a(cid:107)2
A for
all a ∈ A. Then A is isometrically isomorphic to a real function algebra on some compact
Hausdorff space X with topological involution τ on X, in other words a C/R function algebra
on (X, τ, ¯z).
A
We will now recall some of the theory behind Theorem 6.2.1. For more details see
[Sto71, p29] or [Gam69, p4,p11]. The space X is the character space Car(A) which as a
set is the set of all non-zero, complex-valued, multiplicative C-linear functionals on A.
It turns out that the characters on A are all automatically continuous. Note, in the case
of Theorem 6.2.2 the functionals are complex-valued but R-linear and τ maps each such
functional to its complex conjugate. For a commutative unital complex Banach algebra
A the Gelfand transform is a homomorphism from A to a space of complex valued
functions A defined by a (cid:55)→ a where a(ϕ) := ϕ(a) for all a ∈ A and ϕ ∈ Car(A). The
topology on Car(A) is the initial topology given by the family of functions A. Known
in this case as the Gelfand topology it is the weakest topology on Car(A) such that all
the elements of A are continuous giving A ⊆ CC(Car(A)). The norm given to A is the
sup norm.
Now for a commutative unital complex Banach algebra A the set of maximal ideals of
A and the set of kernels of the elements of Car(A) agree. In Theorem 6.2.1, (cid:107) · (cid:107)A being
square preserving ensures that A is semisimple, that is that the Jacobson radical of A
is {0} where the Jacobson radical is the intersection of all maximal ideals of A and so
the intersection of all the kernels of elements of Car(A). Forcing A to be semisimple
ensures that the Gelfand transform is injective since if A is semisimple then the kernel
of the Gelfand transform is {0}. Similarly to confirm that the Gelfand transform is in-
jective it is enough to show that it is an isometry. Given Theorem 6.2.1 it is immediate
88
CHAPTER 6: REPRESENTATION THEORY
that a commutative unital complex Banach algebra A is isometrically isomorphic to a
uniform algebra if and only if its norm is square preserving since the sup norm has this
property. Hence Theorem 6.2.1 provides a characterisation of uniform algebras.
Now in the non-Archimedean setting Berkovich, the author of [Ber90], takes the fol-
lowing approach involving Definition 6.2.3.
Definition 6.2.3. Let F be a complete non-Archimedean field and let A be a commuta-
tive unital Banach F-algebra. Define M1(A) to be the set of all bounded multiplicative
seminorms on A. Further a character on A is a non-zero, multiplicative F-linear func-
tional on A that takes values in some complete field extending F as a valued field.
For an appropriate topology, M1(A) plays the role for A in Definition 6.2.3 that the
maximal ideal space, equivalently the character space, plays in the Archimedean set-
ting. For · ∈ M1(A) let x0 := ker( · ). Then x0 is a proper closed prime ideal of A.
Hence the quotient ring A/x0 is an integral domain. Lemma 6.2.4 is useful here.
Lemma 6.2.4. Let F be a complete valued field and let A be a commutative unital Banach F-
algebra. For · a bounded multiplicative seminorm on A with kernel x0 the value a of a ∈ A
only depends on the quotient class in A/x0 to which a belongs. Hence · is well defined when
used as a valuation on A/x0 by setting a + x0 := a. Further x0 is a closed subset of A.
Proof. For a ∈ A and b ∈ x0 we have a = a + b − b ≤ a + b + b = a + b and
a + b ≤ a + b = a hence a + b = a as required. Furthermore this also gives an
easy way of seeing that x0 is a closed subset of A. Let a be an element of A\x0 then for
all b ∈ x0 we have a = a − b ≤ (cid:107)a − b(cid:107)A since · is bounded and so x0 is closed.
Now by Lemma 6.2.4 we can take · to be a valuation on A/x0 and hence extend it
to a valuation on the field of fractions Frac(A/x0). Hence an element · ∈ M1(A)
defines a character on A by sending the elements of A to their image in the completion
of Frac(A/x0) with respect to · . With these details in place we have the following
theorem by Berkovich, see [Ber90, p157].
Theorem 6.2.5. Let F be a complete non-Archimedean field. Let A be a commutative unital
A for all a ∈ A. Suppose that all of the characters of A
Banach F-algebra with (cid:107)a2(cid:107)A = (cid:107)a(cid:107)2
take values in F. Then:
(i) the space M1(A) is totally disconnected;
(ii) the Gelfand transform gives an isomorphism from A to CF(M1(A)).
As we move on to the next subsection it's worth pointing out that the Gelfand theory
presented in [Ber90] does not make use of any definition such as that of L/Lg function
algebras.
89
CHAPTER 6: REPRESENTATION THEORY
6.2.2 Motivation
For A a commutative unital complex Banach algebra it is straightforward to confirm
that there is a one-one correspondence between the elements of Car(A) and the ele-
ments of the maximal ideal space. Since A is unital the complex constants are elements
of A and for ϕ ∈ Car(A), ϕ restricts to the identity on C. Hence by the first isomor-
phism theorem for rings we have
A/ker(ϕ) ∼= ϕ(A) = C
(6.2.1)
showing that ker(ϕ) is a maximal ideal of A. Therefore, by also noting the prelude
to Chapter 5, the set of maximal ideals of A and the set of kernels of the elements of
Car(A) do indeed agree. It remains to show that no two characters on A have the same
kernel and this marks an important difference with the theory we are about to present.
First though let ϕ, φ be elements of Car(A) with ker(ϕ) = ker(φ). We note that for
each a ∈ A there is a unique α ∈ C representing the quotient class a + ker(ϕ) by (6.2.1).
Hence for some b ∈ ker(ϕ) we have a + b = α giving
ϕ(a) = ϕ(a) + ϕ(b) = ϕ(a + b) = ϕ(α) = α = φ(α) = φ(a + b) = φ(a) + φ(b) = φ(a)
and so no two characters on A have the same kernel.
Now let F be a complete non-Archimedean field. We wish to identify sufficient condi-
tions for a commutative unital Banach F-algebra to be represented by some L/F func-
tion algebra. In this respect the following lemma is informative and motivates an ap-
propriate choice of character space in Subsection 6.2.3.
Lemma 6.2.6. For A an L/Lg function algebra on (X, τ, g), where L can be Archimedean or
non-Archimedean and A is not assumed to be basic, define a family of maps on A by
fA,x := f (x)L
for x ∈ X and f ∈ A.
Then for each x ∈ X:
(i) the map · A,x is a bounded multiplicative seminorm on A;
(ii) the kernel ker( · A,x), which is the same as ker( x) where x is the evaluation character
x(f):=f(x) on A, is not only a proper closed prime ideal of A but it is also a maximal ideal;
(cid:16)(cid:100)τ(x)
(cid:17)
(iii) we have ker
= ker( x) even if τ is not the identity and in general different
evaluation characters can have the same kernel.
Proof. For (i), it is immediate that · A,x is a bounded multiplicative seminorm on A
since the norm on A is the sup norm and · L is a valuation on L.
90
CHAPTER 6: REPRESENTATION THEORY
For (ii), it is immediate that ker( · A,x) is a proper ideal of A noting that 1A,x =
1L = 1. It remains to show that ker( · A,x) is a maximal ideal of A noting Lemma
6.1.7. To this end we show that the quotient ring A/ker( x) is a field. We first note that
Lg ⊆ x(A) ⊆ L and that x(A) is a ring and so an integral domain. Further by the first
isomorphism theorem for rings we have A/ker( x) ∼= x(A) and so A/ker( x) contains
an embedding of Lg and each element a ∈ A/ker( x) is an element of an algebraic
extension of Lg since L is a finite extension of Lg. Therefore for a ∈ A/ker( x) with
a (cid:54)= 0 we have by Lemma 2.2.19 that Lg(a) = Lg[a] where Lg(a) is a simple extension
of Lg and Lg[X] is the ring of polynomials over Lg. Hence, since Lg[a] ⊆ A/ker( x), the
inverse a−1 is an element of A/ker( x) which is therefore a field as required.
For (iii), we note that for all f ∈ A and x ∈ X we have f (τ(x)) = g( f (x)) since f is an
element of C(X, τ, g). Further since g ∈ Gal(L/F) we have g( f (x)) = 0 if and only if
f (x) = 0 and so ker
= ker( x). However in general f (x) need not be equal to
g( f (x)) and so different evaluation characters can have the same kernel.
(cid:16)(cid:100)τ(x)
(cid:17)
6.2.3 Representation under finite basic dimension
This subsection will involve the use of Definition 6.2.7.
Definition 6.2.7. Suppose F1 and F2 are extensions of a field F such that there exists an
isomorphism ϕ : F1 → F2 with ϕ(a) = a for all a ∈ F. Then ϕ is called an F-isomorphism
and F1 and F2 are called F-isomorphic or with the same meaning F-conjugate. Similarly
if F is complete then we can talk of F-isomorphic Banach F-algebras etc.
The following definition and theorem will be the focus of attention for the rest of this
chapter.
Definition 6.2.8. Let F be a complete valued field and let A be a commutative unital Ba-
nach F-algebra. We say that A has finite basic dimension if there exists a finite extension
L of F extending F as a valued field such that:
(i) for each proper closed prime ideal J of A, that is the kernel of a bounded mul-
tiplicative seminorm on A, the field of fractions Frac(A/J) is F-isomorphic to a
subfield of L;
(ii) there is g ∈ Gal(L/F) with Lg = F.
Cases where L = F are allowed.
The purpose of Definition 6.2.8 is to generalise to the non-Archimedean setting con-
ditions that are innately present in the Archimedean case due to the Gelfand Mazur
theorem. We will discuss this in Remark 6.2.10.
91
CHAPTER 6: REPRESENTATION THEORY
Theorem 6.2.9. Let F be a locally compact complete non-Archimedean field with nontrivial
A for all a ∈ A
valuation. Let A be a commutative unital Banach F-algebra with (cid:107)a2(cid:107)A = (cid:107)a(cid:107)2
and finite basic dimension. Then:
(i) for some finite extension L of F extending F as a valued field, a character space M(A) of
L valued, multiplicative F-linear functionals can be defined;
(ii) the space M(A) is a totally disconnected compact Hausdorff space;
(iii) A is isometrically F-isomorphic to a L/F function algebra on (M(A), g, g) for some
g ∈ Gal(L/F).
Remark 6.2.10. Concerning the condition of finite basic dimension.
(i) We first note that all commutative unital complex Banach algebras and commu-
tative unital real Banach algebras have finite basic dimension. To see this let A
be such an algebra and let J be a proper closed prime ideal of A such that J is
the kernel of a bounded multiplicative seminorm · on A. Then, by Lemma
6.1.5 and Lemma 6.2.4, · extends the absolute valuation on R to a valuation on
the integral domain A/J. Extending · to a valuation on Frac(A/J) gives either
R or C by the Gelfand Mazur theorem and noting Theorem 2.1.4. Finally with
consideration of Gal(C/R) the result follows. Hence we note that with little mod-
ification Theorem 6.2.1, Theorem 6.2.2 and Theorem 6.2.9 could be combined into
a single theorem.
(ii) Now the argument in (i) was deliberately a little naive noting that for every com-
mutative unital Banach F-algebra A with finite basic dimension the kernel of ev-
ery bounded multiplicative seminorm on A is a maximal ideal of A. This follows
easily from Lemma 2.2.19 since such a kernel J is a proper closed prime ideal of
A and the elements of the quotient ring A/J are algebraic over F and so A/J is a
field.
(iii) Finally if A is a commutative unital Banach F-algebra then in general the set of
maximal ideals of A is a subset of the set of kernels of bounded multiplicative
seminorms on A by Lemma 6.1.10. Hence Theorem 6.2.9 might be strengthened
if we can find a proof that accepts changing (i) in Definition 6.2.8 to the condition
that for each maximal ideal J of A the field A/J is F-isomorphic to a subfield of
L. This is something for the future. The change only makes a difference for cases
where there is a bounded multiplicative seminorm on A with kernel J such that
A/J has elements that are transcendental over F since otherwise J is a maximal
ideal of A.
92
CHAPTER 6: REPRESENTATION THEORY
Proof of Theorem 6.2.9. Let M0(A) be as in Definition 6.1.4. Now A has finite basic di-
mension so for each x0 ∈ M0(A) the quotient ring A/x0 is a field by Remark 6.2.10.
Further there is a finite extension L of F extending F as a valued field such that for
all x0 ∈ M0(A) the field A/x0 is F-isomorphic to a subfield of L. Moreover for · a
bounded multiplicative seminorm on A with kernel x0 the map a + x0A/x0 := a, for
a ∈ A, defines a valuation on A/x0 extending the valuation on F by Lemma 6.2.4 and
Lemma 6.1.5. We note that since L and A/x0 are both finite extensions of F they are
complete valued fields. Further since · A/x0 is defined by a bounded multiplicative
seminorm on A we have
a + x0A/x0 ≤ (cid:107)a(cid:107)A for all a ∈ A.
(6.2.2)
We now progress towards defining the character space of A. Define M(A) as the set
of all pairs x := (x0, ϕ) where x0 ∈ M0(A) and ϕ is an F-isomorphism from A/x0
to a subfield of L extending F. Then to each x = (x0, ϕ) ∈ M(A) we associated a
map x : A → L given by x(a) := ϕ(a + x0) for all a ∈ A. Note that for each element
x = (x0, ϕ) ∈ M(A) we have
a + x0A/x0 = x(a)L
for all a ∈ A
(6.2.3)
by the uniqueness of the valuation on A/x0 extending the valuation on F, see Theorem
2.2.5. In particular each F-isomorphism from A/x0 to a subfield of L extending F is an
isometry and similarly we recall that each element of Gal(L/F) is isometric. Now for
the element g ∈ Gal(L/F) with Lg = F, or indeed any other element of Gal(L/F), we
g((x0, ϕ)) := (x0, g ◦ ϕ). In particular for x = (x0, ϕ1) ∈ M(A) we have g ◦ x = (cid:100)g(x)
note that g can be considered as a map of finite order g : M(A) → M(A) given by
and so there is y = (y0, ϕ2) ∈ M(A) with y0 = x0 such that the diagram in Figure 6.1
commutes. Note that in the case of Figure 6.1 the fields x(A) and y(A) are F-conjugate
L
g
ϕ2
ϕ1
L
y
x
A/x0
q
A
Figure 6.1: Commutative diagram for the characters associated to x and y.
and could actually be the same subfield of L if the restriction g x(A) is an element of
Gal( x(A)/F). Now by construction for each x ∈ M(A) the map x is a non-zero, L-
valued, multiplicative F-linear functional on A. Hence x is continuous since we have
x(a)L ≤ (cid:107)a(cid:107)A for all a ∈ A
(6.2.4)
93
o
o
<
<
b
b
E
E
O
O
Y
Y
CHAPTER 6: REPRESENTATION THEORY
by (6.2.3) and (6.2.2). We now set up the Gelfand transform in the usual manner by
defining a map
(cid:98)· : A → A,
a (cid:55)→ a,
A are the functions a : M(A) → L given by a(x) := x(a). We
where the elements of
equip A with the binary operations of pointwise addition and multiplication and put
the sup norm
(cid:107) a(cid:107)∞ := sup
x∈M(A)
a(x)L
for all a ∈ A
on A making A a commutative unital normed F-algebra. Note that with these binary
operations it is immediate that the Gelfand transform is an F-homomorphism and so
A is closed under addition and multiplication. Later we will show that(cid:98)· : A → A is an
isometry and so it is also injective. It then follows that A is a Banach F-algebra since A
and A are isometrically F-isomorphic.
Now we equip M(A) with the Gelfand topology which is the initial topology of
A.
A are continuous L-valued functions on the space M(A). We
Hence the elements of
A separates the points of M(A) and that M(A) is a compact Hausdorff
show that
space. Let x and y be elements of M(A) with x = (x0, ϕ), y = (y0, φ) and x (cid:54)= y.
If x0 (cid:54)= y0 then there is a ∈ x0 ∪ y0 such that a (cid:54)∈ x0 ∩ y0 for which precisely one of
a(x) = x(a) and a(y) = y(a) is zero. If x0 = y0 then ϕ (cid:54)= φ on A/x0. Hence there is
some a ∈ A such that ϕ(a + x0) (cid:54)= φ(a + x0) giving
a(x) = x(a) = ϕ(a + x0) (cid:54)= φ(a + x0) = y(a) = a(y)
and so A separates the points of M(A). We now show that M(A) is Hausdorff, note
in fact that the proof is standard. Let x and y be elements of M(A) with x (cid:54)= y. Since
A separates the points of M(A) there is a ∈ A such that a(x) (cid:54)= a(y). Further L is
Hausdorff and so there are disjoint open subsets U1 and U2 of L such that a(x) ∈ U1
and a(y) ∈ U2. Since the topology on M(A) is the initial topology of A the preimage
a−1(U1) is an open neighborhood of x in M(A) and the preimage a−1(U2) is an open
neighborhood of y in M(A). Moreover a−1(U1) and a−1(U2) are disjoint because U1
and U2 are, as required.
The following, showing that M(A) is compact, is an adaptation of part of the proof of
Theorem 6.2.2 from [KL92, p23]. For each a ∈ A define La := {α ∈ L : αL ≤ (cid:107)a(cid:107)A}
and LA := ∏a∈A La with the product topology. Each La is compact by Theorem 2.1.23
noting that L is locally compact by Remark 2.2.14. Hence LA is compact by Tychonoff's
Theorem. Now by (6.2.4) we have x(a)L ≤ (cid:107)a(cid:107)A for all x ∈ M(A) and a ∈ A.
Therefore for each x ∈ M(A) we have x(a) ∈ La and so x is a point of LA and M(A)
can be considered as a subset of LA. Now the product topology on LA is the initial
topology of the family of coordinate projections Pa : LA → La, a ∈ A. Since we have
94
CHAPTER 6: REPRESENTATION THEORY
PaM(A) = a the Gelfand topology on M(A) is the initial topology of the family of
coordinate projections restricted to M(A). Hence the topology on M(A) is the relative
topology of M(A) as a subspace of LA. Since LA is compact, any subspace of LA that
is closed as a subset is also compact. Hence it remains to show that M(A) is a closed
subset of LA. Let ϕ ∈ LA be in the closure of M(A). Hence we have ϕ(a)L ≤ (cid:107)a(cid:107)A
for all a ∈ A and there is a net (xλ) in M(A) converging to ϕ. Now since LA has the
product topology, convergence in LA is coordinate-wise, see [Wil04, §8]. Therefore for
a, b ∈ A we have
ϕ(a + b) = lim xλ(a + b) = lim( xλ(a) + xλ(b)) = ϕ(a) + ϕ(b).
Similarly, ϕ(ab) = ϕ(a)ϕ(b) and ϕ(α) = α for all a, b ∈ A and α ∈ F. Now since ϕ
takes values in L and L is a finite extension of F we have that ϕ(A) is a subfield of L
extending F by Lemma 2.2.19. Hence since A/ker(ϕ) ∼= ϕ(A), by the first isomorphism
theorem for rings, the kernel of ϕ is a maximal ideal of A. Therefore ker(ϕ) is an
element of M0(A). Further ϕ defines an F-isomorphism from A/ker(ϕ) to a subfield
of L extending F by ϕ(cid:48)(a + ker(ϕ)) := ϕ(a). Hence we have obtained y := (ker(ϕ), ϕ(cid:48))
which is an element of M(A) with y = ϕ and so M(A) is closed as a subset on LA.
We will now show that g : M(A) → M(A) is continuous. The set of preimages
S := { a−1(U) : a ∈ A and U ⊆ L is open}
is a sub-base for the Gelfand topology on M(A). To show that g : M(A) → M(A)
is continuous it is enough to show that for each V ∈ S the preimage g−1(V) is also an
element of S. We note that g : M(A) → M(A) is a bijection since g has finite order.
So let V = a−1(U) be an element of S for some a ∈ A and open U ⊆ L. We have
x = (x0, ϕ) ∈ M(A) an element of V if and only if a(x) = x(a) = ϕ(a + x0) is an
element of U. Now consider the elements of the preimage g−1(V) and note that they
are the elements y = (y0, φ) ∈ M(A) such that g(y) = (y0, g ◦ φ) ∈ V. These are
precisely the elements of M(A) such that
(cid:17)
a(y) = y(a) = φ(a + y0) ∈ g(ord(g)−1)(U).
g(ord(g)−1)(U)
And so g−1(V) = a−1(cid:16)
and since g is an isometry on L we note that
g(ord(g)−1)(U) is an open subset of L. Hence g−1(V) is an element of S as required.
We now show that the Gelfand transform is an isometry. Note that the following adapts
material that can be found in [Ber90, Ch1]. Let a be an element of A. By (6.2.4) we have
a(x)L = x(a)L ≤ (cid:107)a(cid:107)A for all x ∈ M(A) and so (cid:107) a(cid:107)∞ ≤ (cid:107)a(cid:107)A. For the reverse
inequality let ε > 0 and set r := (cid:107) a(cid:107)∞ + ε. Then for all x0 ∈ M0(A) we have
a + x0A/x0 = x(a)L = a(x)L ≤ (cid:107) a(cid:107)∞ < r
(6.2.5)
95
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞
∑
i=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)A,r−1
aiTi
=
∞
∑
i=0
(cid:107)ai(cid:107)A(r−1)i =
i=0 aiTi with
∞
(cid:107)ai(cid:107)Ar−i < ∞
∑
i=0
CHAPTER 6: REPRESENTATION THEORY
for some x = (x0, ϕ) ∈ M(A) by (6.2.3) and noting that A has finite basic dimension.
Now consider the commutative unital Banach F-algebra A(cid:104)rT(cid:105). Let M0(A(cid:104)rT(cid:105)) be the
set of all proper closed prime ideals of A(cid:104)rT(cid:105) that are the kernels of bounded multi-
plicative seminorms on A(cid:104)rT(cid:105). Note that M0(A(cid:104)rT(cid:105)) is non-empty by Lemma 6.1.10.
We recall that the elements of A(cid:104)rT(cid:105) are of the form ∑∞
and ai ∈ A for all i ∈ N0. Hence A is a subring of A(cid:104)rT(cid:105) since for each b ∈ A we
have b = bT0 an element of A(cid:104)rT(cid:105). Now for y0 ∈ M0(A(cid:104)rT(cid:105)) let · be a bounded
multiplicative seminorm on A(cid:104)rT(cid:105) with y0 = ker( · ). Since · is bounded we have
T ≤ (cid:107)T(cid:107)A,r−1 = r−1.
(6.2.6)
Moreover since for b ∈ A we have (cid:107)bT0(cid:107)A,r−1 = (cid:107)b(cid:107)A(r−1)0 = (cid:107)b(cid:107)A, the restriction
· A is a bounded multiplicative seminorm on A. Further m0 := ker( · A) is closed
as a subset of A by Lemma 6.2.4 and so m0 is an element of M0(A). Hence m0 is a
maximal ideal of A by remark 6.2.10. In particular b + m0A/m0 := b, for b ∈ A, is the
unique valuation on A/m0 extending the valuation on F as we have seen earlier in this
proof for other elements of M0(A). Therefore by (6.2.5) and (6.2.6) we have
aT = aT ≤ a + m0A/m0r−1 < rr−1 = 1.
Furthermore 1 = 1 ≤ 1 − aT + aT and so we have 1 − aT ≥ 1 − aT > 0.
Therefore 1 − aT is not an element of y0 since y0 is the kernel of · . Since y0 was any
element of M0(A(cid:104)rT(cid:105)) we have 1− aT (cid:54)∈ y0 for all y0 ∈ M0(A(cid:104)rT(cid:105)). Hence by Lemma
6.1.10 we note that 1 − aT is invertible in A(cid:104)rT(cid:105). Therefore by Lemma 6.1.1 the series
i=0 (cid:107)ai(cid:107)Ar−i converges. In particular we can find N ∈ N such that for all n > N we
∑∞
2 since (cid:107) · (cid:107)A is square preserving. Hence
have (cid:107)a2n(cid:107)Ar−2n
(cid:107)a(cid:107)A < r = (cid:107) a(cid:107)∞ + ε and since ε > 0 was arbitrary we have (cid:107)a(cid:107)A ≤ (cid:107) a(cid:107)∞ and so
(cid:107)a(cid:107)A = (cid:107) a(cid:107)∞ as required.
What remains to be shown is that the elements of A are also elements of C(M(A), g, g)
and that M(A) is totally disconnected. For a ∈ A and x = (x0, ϕ) ∈ M(A) we have
2 giving ((cid:107)a(cid:107)Ar−1)2n
< 1
< 1
a(g(x)) = a((x0, g ◦ ϕ)) =
(cid:92)
(x0, g ◦ ϕ)(a)
=g ◦ ϕ(a + x0)
(cid:17)
=g(ϕ(a + x0))
(x0, ϕ)(a)
=g
(cid:16)(cid:92)
= g( x(a)) = g(a(x))
and so a is an element of C(M(A), g, g). Finally it is immediate that M(A) is totally
disconnected since A separates the points of M(A), the elements of A are continuous
96
CHAPTER 6: REPRESENTATION THEORY
functions from M(A) to L, the image of a connected component is connected for con-
tinuous functions and L is totally disconnected. In particular see the proof of Theorem
4.2.1. This completes the proof of Theorem 6.2.9.
In the next chapter we will survey some existing results in the Archimedean non-
commutative setting and also consider the possibility of their generalisation to the
non-Archimedean setting. We will then finish by noting some of the open questions
arising from the Thesis.
97
CHAPTER 7
Non-commutative generalisation and open
questions
In recent years a theory of non-commutative real function algebras has been developed
by Jarosz, see [Jar08] and [AJ04]. In the first section of this short chapter we survey
and remark upon some of this non-commutative Archimedean theory and consider
the possibility of non-commutative non-Archimedean analogs. In the second section
we note some of the open questions arising from the thesis.
7.1 Non-commutative generalisation
7.1.1 Non-commutative real function algebras
In the recent theory of non-commutative real function algebras the continuous func-
tions involved take values in Hamilton's real quaternions, H, which are an example
of a non-commutative complete Archimedean division ring and R-algebra. Viewing
H as a real vector space, the valuation on H is the Euclidean norm which is complete,
Archimedean and indeed a valuation since being multiplicative on H. To put H into
context, as in the case of complete Archimedean fields, there are very few unital divi-
sion algebras over the reals with the Euclidean norm as a valuation. Up to isomorphism
they are R, C, H and the octonions O. We note that the octonions are non-associative.
The proof that there are no other unital division algebras over the reals with the Eu-
clidean norm as a valuation is given by Hurwitz's 1, 2, 4, 8 Theorem, see [Sha00, Ch1]
and [Lew06]. In particular for such an algebra A the square of the Euclidean norm is
a regular quadratic form on A and since for A the Euclidean norm is a valuation it is
multiplicative. This shows that A is a real composition algebra to which Hurwitz's 1,
2, 4, 8 Theorem can be applied.
Here we only briefly consider non-commutative real function algebras and hence the
98
CHAPTER 7: NON-COMMUTATIVE GENERALISATION AND OPEN QUESTIONS
reader is also referred to [Jar08]. Note I am unaware of any such developments involv-
ing the octonions. Here is Jarosz's analog of C(X, τ) from Definition 4.2.5.
Definition 7.1.1. Let Gal(H/R) be the group of all automorphisms on H that are the
identity on R. Let X be a compact space and Hom(X) be the group of homeomor-
phisms on X. For a group homomorphism Φ : Gal(H/R) → Hom(X), Φ(T) = ΦT, we
define
CH(X, Φ) := { f ∈ CH(X) : f (ΦT(x)) = T( f (x)) for all x ∈ X and T ∈ Gal(H/R)}.
Remark 7.1.2. Concerning Definition 7.1.1.
(i) The groups Gal(H/R) and Hom(X) in Definition 7.1.1 have composition as their
group operation. We note that the map ∗ : Gal(H/R) × CH(X) → CH(X) given
by T ∗ f := T−1( f (ΦT(x))) is similar to a group action on CH(X) only with the
usual associativity replaced by T1 ◦ T2 ∗ f = T2 ∗ T1 ∗ f .
(ii) There is an interesting similarity between Definition 7.1.1 and Definition 5.1.2
of Basic function algebras. Let X be a compact Hausdorff space, F a complete
valued field and L a finite extension of F. Further let (cid:104)g(cid:105) be the cyclic group gen-
erated by some g ∈ Gal(L/F) and similarly let (cid:104)τ(cid:105) be the cyclic group generated
by some homeomorphism τ : X → X. Then there exists a surjective group ho-
momorphism Φ : (cid:104)g(cid:105) → (cid:104)τ(cid:105) if and only if ord(τ)ord(g). To see this suppose
such a surjective group homomorphism exists. Then there are m, n ∈ N such that
Φ(g(m)) = id and Φ(g(n)) = τ. This gives
τ(ord(g)) = Φ(g(n))(ord(g)) =Φ(g(nord(g)))
=Φ(id)
=id ◦ Φ(id)
=Φ(g(m)) ◦ Φ(id)
=Φ(g(m) ◦ id) = Φ(g(m)) = id
and so ord(τ)ord(g). Conversely if ord(τ)ord(g) then Φ defined by Φ(g) := τ
will do. It is an interesting question then whether the definition of basic function
algebras can be further generalised by utilizing group homomorphisms as Defi-
nition 7.1.1 suggests noting that Φ is onto for some subgroup of Hom(X). In par-
ticular, with reference to Definition 5.1.2, we have considered basic L/Lg function
algebras where g is an element of Gal(L/F). We note that L is a cyclic extension of
Lg by the fundamental theorem of Galois theory. Therefore it is interesting to con-
sider the possibility of basic L/F function algebras where L is a Galois extension
of F but not necessarily a cyclic extension. Such group homomorphisms might
also be useful in cases involving infinite extensions of F.
99
CHAPTER 7: NON-COMMUTATIVE GENERALISATION AND OPEN QUESTIONS
(iii) Turning our attention back to the non-commutative setting, as a conjecture I sug-
gests that Definition 7.1.1 may also be useful if Gal(H/R) is replaced by a sub-
group, particularly when considering extensions of the algebra.
Definition 7.1.1 has been used by Jarosz in the representation of non-commutative real
Banach algebras with square preserving norm as follows.
Definition 7.1.3. A real algebra A is fully non-commutative if every nonzero multiplica-
tive, linear functional ϕ : A → H is surjective.
Theorem 7.1.4. Let A be a non-commutative real Banach algebra with (cid:107)a2(cid:107)A = (cid:107)a(cid:107)2
A for all
a ∈ A. Then there is a compact set X and an isomorphism Φ : Gal(H/R) → Hom(X) such
that A is isometrically isomorphic with a subalgebra A of CH(X, Φ). Furthermore a ∈ A is
invertible if and only if the corresponding element a ∈ A does not vanish on X. If A is fully
non-commutative then A = CH(X, Φ).
Jarosz also gives the following Stone-Weierstrass theorem type result.
Theorem 7.1.5. Let X be a compact Hausdorff space and let A be a fully non-commutative
closed subalgebra of CH(X). Then A = CH(X) if and only if A strongly separates the points
of X, that is for all x1, x2 ∈ X with x1 (cid:54)= x2 there is f ∈ A satisfying f (x1) (cid:54)= f (x2) = 0.
7.1.2 Non-commutative non-Archimedean analogs
Non-commutative, non-Archimedean analogs of uniform algebras have yet to be seen.
Hence in this subsection we give an example of a non-commutative extension of a
complete non-Archimedean field which would be appropriate when considering such
analogs of uniform algebras. We first have the following definition from the general
theory of quaternion algebras. The main reference for this subsection is [Lam05, Ch3]
but [Lew06] is also useful.
Definition 7.1.6. Let F be a field, with characteristic not equal to 2, and s, t ∈ F× where
s = t is allowed. We define the quaternion F-algebra ( s,t
F ) as follows. As a 4-dimensional
vector space over F we define
(cid:18) s, t
(cid:19)
F
:= {a + bi + cj + dk : a, b, c, d ∈ F}
with {1, i, j, k} as a natural basis giving the standard coordinate-wise addition and
scalar multiplication. As an F-algebra, multiplication in ( s,t
k2 = ij = −ji
F ) is given by
i2 = s,
j2 = t,
together with the usual distributive law and multiplication in F.
100
CHAPTER 7: NON-COMMUTATIVE GENERALISATION AND OPEN QUESTIONS
−1,−1
Hamilton's real quaternions, H := (
R ) with the Euclidean norm, is an example of
a non-commutative, complete valued, Archimedean division algebra over R. It is not
the case that every quaternion algebra ( s,t
F ) will be a division algebra, although there
are many examples that are. For our purposes we have the following example.
Example 7.1.7. Using Q5, the complete non-Archimedean field of 5-adic numbers, de-
fine
Then for q, r ∈ H5, q = a + bi + cj + dk, the conjugation on H5 given by
(cid:18)5, 2
(cid:19)
.
Q5
H5 :=
¯q := a − bi − cj − dk
(cid:113) ¯qq5
qH5 :=
is such that q + r = ¯q + ¯r, qr = ¯r ¯q, ¯qq = q ¯q = a2 − 5b2 − 2c2 + 10d2 with ¯qq ∈ Q5.
Further
is a complete non-Archimedean valuation on H5, where · 5 is the 5-adic valuation on
Q5. In particular H5, together with · H5, is an example of a non-commutative, com-
plete valued, non-Archimedean division algebra over Q5. When showing this directly
it is useful to know that for a, b, c, d ∈ Q5 we have
ν5(a2 − 5b2 − 2c2 + 10d2) = min{ν5(a2), ν5(5b2), ν5(2c2), ν5(10d2)}
where ν5 is the 5-adic valuation logarithm as defined in Example 2.1.18. Given the
above, we will confirm that · H5 is multiplicative. For more details please see the
suggested references [Lam05, Ch3] and [Lew06]. Let q, r ∈ H5 and note that we have
¯r ¯qqr = ¯qq¯rr since ¯qq is an element of Q5. Therefore
(cid:113)qrqr5 =
(cid:113)¯r ¯qqr5
(cid:113) ¯qq¯rr5
(cid:113) ¯qq5¯rr5 =
=
=
qrH5 =
as required.
(cid:113) ¯qq5
(cid:113)¯rr5 = qH5rH5
More generally for the p-adic field Qp the quaternion algebra ( p,u
algebra as long as u is a unit of {a ∈ Qp : ap ≤ 1}, i.e. up = 1, and Qp(
quadratic extension of Qp.
Qp ) will be a division
u) is a
√
7.2 Open questions
There are many open questions related to the content of this thesis and I had intended
to investigate more of them but there was no time. Many of these questions come
101
CHAPTER 7: NON-COMMUTATIVE GENERALISATION AND OPEN QUESTIONS
from the need to generalise established Archimedean results whilst others arise from
the developing theory itself. We now consider some of these questions and note that
several of them appear to be quite accessible.
(Q1) J. Wermer gave the following theorem in 1963.
Theorem 7.2.1. Let X be a compact Hausdorff space, A ⊆ CC(X) a complex uniform
algebra and (cid:60)(A) := {(cid:60)( f ) : f = (cid:60)( f ) + i(cid:61)( f ) ∈ A} the set of the real components
of the functions in A. If (cid:60)(A) is a ring then A = CC(X).
The following analog of Theorem 7.2.1 for real function algebras was given by S.
H. Kulkarni and N. Srinivasan in [KS90], although I have not used their notation.
Theorem 7.2.2. Let X be a compact Hausdorff space, τ a topological involution on X
and A a C/R function algebra on (X, τ, ¯z), i.e. a real function algebra. If (cid:60)(A) is a ring
then A = C(X, τ, ¯z).
It is interesting to know whether Theorem 7.2.2 can be generalised to all L/Lg
function algebras on (X, τ, g). Of course the result would be trivial if, in the non-
Archimedean setting, the basic L/Lg function algebra on (X, τ, g) is the only L/Lg
function algebra on (X, τ, g). With analogy to (cid:60)(A) above, in this case we should
ask whether the set of Lg components of the functions in C(X, τ, g) form a ring.
(Q2) As alluded to in (Q1) we have not given an example in the non-Archimedean
setting of a L/Lg function algebra on (X, τ, g) that is not basic. We need to know
whether the basic function algebras are the only such examples. Theorem 4.2.3,
Kaplansky's version of the Stone Weierstrass Theorem, may be important here.
Further even if in the non-Archimedean setting there is a L/Lg function algebra on
(X, τ, g) that is not basic, such an algebra might still be isometrically isomorphic
to some Basic function algebras.
(Q3) With reference to Theorem 6.2.9 we note that there are plenty of examples of
commutative, unital Banach F-algebras with finite basic dimension in the non-
Archimedean setting. Indeed if K is not only a finite Galois extension of F but
also a cyclic extension then taking A := K gives such an algebra. In this case
the character space M(A) will be finite with each element given by an element
of Gal(K/F), see the proof of Theorem 6.2.9 for details. However such examples
are not particularly interesting and it would be good to know whether all L/Lg
function algebras on (X, τ, g) have finite basic dimension so that Theorem 6.2.9
becomes closer to a characterisation result. We recall that all commutative uni-
tal complex Banach algebras and commutative unital real Banach algebras have
finite basic dimension, see Remark 6.2.10.
102
CHAPTER 7: NON-COMMUTATIVE GENERALISATION AND OPEN QUESTIONS
(Q4) With reference to Definition 5.1.2 of the basic L/Lg function algebra on (X, τ, g)
the map σ( f ) = g(ord(g)−1) ◦ f ◦ τ on CL(X) is such that each f ∈ CL(X) is an
element of C(X, τ, g) if and only if σ( f ) = f . We have seen that σ is either an
algebraic involution on CL(X) or a algebraic element of finite order on CL(X). It
should be established whether every such involution and element of finite order
on CL(X) has the form of σ for some g and τ. This is the case for real function
algebras, see [KL92, p29].
(Q5) As described in Remark 7.1.2 it might be possible to generalise the definition of
Basic function algebras by involving a group homomorphism in the definition.
The algebras currently given by Definition 5.1.2 could then appropriately be re-
ferred to as cyclic basic function algebras given that the group Gal(L/Lg ) is cyclic.
Further the possibility of generalising the definition of Basic function algebras to
the case where the functions take values in some infinite extension of the under-
lying field over which the algebra is a vector space should also be considered.
The involvement of a group homomorphism might also be useful in this case as
well as some more of the theory from [Ber90].
(Q6) As seen in Subsection 7.1.2 the general theory of quaternion algebras provides the
necessary structures for generalising the theory of non-commutative real function
algebras to the non-Archimedean setting. Further with reference to Subsection
5.4.1 it would be interesting to see what sort of lattice of basic extensions the
non-commutative real function algebras have. We can also look at this in the
non-Archimedean setting along with the residue algebra.
(Q7) A proof of the following theorem can be found in [KL92, p18].
Theorem 7.2.3. Let A be a unital Banach algebra in the Archimedean setting satisfying
one of the following conditions:
(i) the algebra A is a complex algebra and there exists some positive constant c such
that (cid:107)a(cid:107)2
A ≤ c(cid:107)a2(cid:107)A for all a ∈ A;
(ii) the algebra A is a real algebra and there exists some positive constant c such that
(cid:107)a(cid:107)2
A ≤ c(cid:107)a2 + b2(cid:107)A for all a, b ∈ A with ab = ba.
Then A is commutative.
It would be interesting to establish whether there is such a theorem for all unital
Banach F-algebras.
If not then perhaps some special cases are possible in the
non-Archimedean setting. The proof of theorem 7.2.3 uses Liouville's theorem
and some spectral theory in the Archimedean setting. Both of these are different
in the non-Archimedean setting, see Theorem 3.1.9 and Subsection 3.2.1.
103
CHAPTER 7: NON-COMMUTATIVE GENERALISATION AND OPEN QUESTIONS
(Q8) It might be interesting to investigate the isomorphism classes of basic function
algebras. That is for a given basic function algebra A are there other basic function
algebras that are isometrically isomorphic to A.
(Q9) It is interesting to consider whether the Kaplansky spectrum of Remark 3.2.13 can
be used for some cases in the non-Archimedean setting and, if so, whether it is
one such definition in some larger family of definitions of spectrum applicable in
the non-Archimedean setting.
(Q10) More broadly the established theory of Banach algebras provides a large supply
of topics that can be considered for generalisation over complete valued fields. In
addition to several of the other references included in this thesis [Dal00] will be
of much interest when considering such possibilities. One obvious example is the
generalisation of automatic continuity results. That is what conditions on a Ba-
nach F-algebra force homomorphisms from, or to, that algebra to be continuous.
There is one such result in this thesis noting that in Theorem 6.2.9 the elements
of M(A) are automatically continuous. Further [BBN73] may also be of interest
concerning function algebras.
(Q11) As mentioned in Remark 4.1.20 there is a possible generalisation of the Swiss
cheese classicalisation theorem to the Riemann sphere and possibly to a more
general class of metric spaces.
(Q12) It might be interesting to consider generalising over all complete valued fields
the theory of algebraic extensions of commutative unital normed algebras. See
the survey paper [Daw03] for details.
(Q13) The possibility of generalising C∗-Algebras over complete valued fields is inter-
esting but perhaps not straightforward. The Levi-Civita field might be of interest
here since it is totally ordered such that the order topology agrees with the val-
uation topology. Hence it might be possible to define positive elements in this
case. Perhaps the algebraic elements of finite order mentioned in (Q4) are rele-
vant. Also there is a monograph by Goodearl from 1982 that considers real C∗-
Algebras that might be of use. The possibility of a non-Archimedean theory of
Von Neumann algebras might also be a good place to start.
104
References
[AJ04] Mati Abel and Krzysztof Jarosz. Noncommutative uniform algebras. Studia
Math., 162(3):213 -- 218, 2004.
[Apo74] Tom M. Apostol. Mathematical analysis. Addison-Wesley Publishing Co.,
Reading, Mass.-London-Don Mills, Ont., second edition, 1974.
[BBN73] G. Bachman, E. Beckenstein, and L. Narici. Function algebras over valued
fields. Pacific J. Math., 44:45 -- 58, 1973.
[Ber90] Vladimir G. Berkovich.
Spectral theory and analytic geometry over non-
Archimedean fields, volume 33 of Mathematical Surveys and Monographs. Amer-
ican Mathematical Society, Providence, RI, 1990.
[Bro69] Andrew Browder. Introduction to function algebras. W. A. Benjamin, Inc., New
York-Amsterdam, 1969.
[Dal00] 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. Oxford Science Publications.
[Daw03] Thomas Dawson. A survey of algebraic extensions of commutative, unital
In Function spaces (Edwardsville, IL, 2002), volume 328 of
normed algebras.
Contemp. Math., pages 157 -- 170. Amer. Math. Soc., Providence, RI, 2003.
[Esc03] Alain Escassut. Ultrametric Banach algebras. World Scientific Publishing Co.
Inc., River Edge, NJ, 2003.
[Fei04] J. F. Feinstein. A counterexample to a conjecture of S. E. Morris. Proc. Amer.
Math. Soc., 132(8):2389 -- 2397 (electronic), 2004.
[FH10] J. F. Feinstein and M. J. Heath. Swiss cheeses, rational approximation and
universal plane curves. Studia Math., 196(3):289 -- 306, 2010.
[FV02] I. B. Fesenko and S. V. Vostokov. Local fields and their extensions, volume 121
of Translations of Mathematical Monographs. American Mathematical Society,
Providence, RI, second edition, 2002. With a foreword by I. R. Shafarevich.
105
REFERENCES
[Gam69] Theodore W. Gamelin. Uniform algebras. Prentice-Hall Inc., Englewood Cliffs,
N. J., 1969.
[Hea08] M. J. Heath. Bounded derivations from Banach algebras. Ph.d. thesis, University
of Nottingham, February 2008.
[Jar08] Krzysztof Jarosz. Function representation of a noncommutative uniform al-
gebra. Proc. Amer. Math. Soc., 136(2):605 -- 611 (electronic), 2008.
[Kap50] Irving Kaplansky. The Weierstrass theorem in fields with valuations. Proc.
Amer. Math. Soc., 1:356 -- 357, 1950.
[Kel75] John L. Kelley. General topology. Springer-Verlag, New York, 1975. Reprint of
the 1955 edition [Van Nostrand, Toronto, Ont.], Graduate Texts in Mathemat-
ics, No. 27.
[KL81] S. H. Kulkarni and B. V. Limaye. Gleason parts of real function algebras.
Canad. J. Math., 33(1):181 -- 200, 1981.
[KL92] S. H. Kulkarni and B. V. Limaye. Real function algebras, volume 168 of Mono-
graphs and Textbooks in Pure and Applied Mathematics. Marcel Dekker Inc., New
York, 1992.
[KS90] S. H. Kulkarni and N. Srinivasan. An analogue of Wermer's theorem for a
real function algebra. Math. Today, 8:33 -- 42, 1990.
[Lam05] T. Y. Lam.
Introduction to quadratic forms over fields, volume 67 of Graduate
Studies in Mathematics. American Mathematical Society, Providence, RI, 2005.
[Lew06] David W. Lewis. Quaternion algebras and the algebraic legacy of Hamilton's
quaternions. Irish Math. Soc. Bull., (57):41 -- 64, 2006.
[Mas10] J. W. D. Mason. An inductive proof of the Feinstein-Heath Swiss cheese "clas-
sicalisation" theorem. Proc. Amer. Math. Soc., 138(12):4423 -- 4432, 2010.
[Mas11] J. W. D. Mason. A survey of non-complex analogs of uniform algebras. In
Function spaces, volume 547 of Contemp. Math. Amer. Math. Soc., Providence,
RI, 2011.
[McC66] Paul J. McCarthy. Algebraic extensions of fields. Blaisdell Publishing Co. Ginn
and Co. Waltham, Mass.-Toronto, Ont.-London, 1966.
[Roq84] Peter Roquette. Some tendencies in contemporary algebra. In Perspectives in
mathematics, pages 393 -- 422. Birkhäuser, Basel, 1984.
106
REFERENCES
[Rot38] Alice Roth. Approximationseigenschaften und Strahlengrenzwerte mero-
morpher und ganzer Funktionen. Comment. Math. Helv., 11(1):77 -- 125, 1938.
[Rud87] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New York,
third edition, 1987.
[SB10] Khodr Shamseddine and Martin Berz. Analysis on the Levi-Civita field, a
In Advances in p-adic and non-Archimedean analysis, volume
brief overview.
508 of Contemp. Math., pages 215 -- 237. Amer. Math. Soc., Providence, RI, 2010.
[Sch02] Peter Schneider. Nonarchimedean functional analysis. Springer Monographs in
Mathematics. Springer-Verlag, Berlin, 2002.
[Sch06] W. H. Schikhof. Ultrametric calculus, volume 4 of Cambridge Studies in Ad-
vanced Mathematics. Cambridge University Press, Cambridge, 2006. An intro-
duction to p-adic analysis, Reprint of the 1984 original [MR0791759].
[Sha00] Daniel B. Shapiro. Compositions of quadratic forms, volume 33 of de Gruyter
Expositions in Mathematics. Walter de Gruyter & Co., Berlin, 2000.
[Sto71] Edgar Lee Stout. The theory of uniform algebras. Bogden & Quigley, Inc.,
Tarrytown-on-Hudson, N. Y., 1971.
[Wil04] Stephen Willard. General topology. Dover Publications Inc., Mineola, NY, 2004.
Reprint of the 1970 original [Addison-Wesley, Reading, MA; MR0264581].
[Zha93] Guan Hou Zhang. Theory of entire and meromorphic functions, volume 122
of Translations of Mathematical Monographs. American Mathematical Society,
Providence, RI, 1993. Deficient and asymptotic values and singular direc-
tions, Translated from the Chinese by Chung-chun Yang.
107
|
1305.5365 | 2 | 1305 | 2016-03-12T12:08:24 | Fine scales of decay of operator semigroups | [
"math.FA",
"math.AP",
"math.DS"
] | Motivated by potential applications to partial differential equations, we develop a theory of fine scales of decay rates for operator semigroups. The theory contains, unifies, and extends several notable results in the literature on decay of operator semigroups and yields a number of new ones. Its core is a new operator-theoretical method of deriving rates of decay combining ingredients from functional calculus, and complex, real and harmonic analysis. It also leads to several results of independent interest. | math.FA | math |
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Abstract. Motivated by potential applications to partial differential
equations, we develop a theory of fine scales of decay rates for operator
semigroups. The theory contains, unifies, and extends several notable
results in the literature on decay of operator semigroups and yields a
number of new ones. Its core is a new operator-theoretical method of
deriving rates of decay combining ingredients from functional calculus,
and complex, real and harmonic analysis. It also leads to several results
of independent interest.
Contents
Introduction
1.
2. Preliminaries on some classes of functions
3. Functional calculus of sectorial operators
4. Some estimates for semigroup asymptotics
5. Singularity at infinity
6. Singularity at zero: general results
7. Singularity at zero: polynomial and regularly varying rates
8. Singularities at both zero and infinity
References
1
11
19
27
32
49
65
72
76
1. Introduction
1.1. Historical background, motivation and sample results. The study
of stability of solutions of the abstract Cauchy problem
( u(t) + Au(t) = 0,
u(0) = x,
t ≥ 0,
x ∈ X,
where −A is the generator of a C0-semigroup (T (t))t≥0 on a Banach space
X, is a classical subject of functional analysis having numerous applications
to partial differential equations. Namely, asymptotic stability (or simply:
Date: October 29, 2018.
1991 Mathematics Subject Classification. Primary 47D06; Secondary 34D05 34G10.
The research described in this paper was supported by the Leverhulme Trust and by the
EPSRC grant EP/J010723/1. The third author was also partially supported by the NCN
grant DEC-2014/13/B/ST1/03153 and the EU grant "AOS", FP7-PEOPLE-2012-IRSES,
No 318910.
1
2
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
stability), that is convergence of all orbits of (T (t))t≥0 to zero, and expo-
nential stability, that is, convergence of all orbits of (T (t))t≥0 to zero with
exponential rate, are two of the main building blocks of stability theory.
Since the resolvent of the generator is often easier to compute than the
semigroup, it is traditional and efficient to use the resolvent when dealing
with both kinds of stability, and actually a number of resolvent criteria for
stability are known in this context (see [6, Chapter 5], [20] and [56]).
Recently, the study of various PDEs revealed that the resolvent can also
be used successfully in order to treat intermediate rates of convergence, thus
distinguishing and quantifying fine modes of convergence to zero. This has
become especially transparent for the damped wave equation
wave
wave
(1.1)
utt + a(x)ut − ∆u = 0 in R+ × M,
u = 0 in R+ × ∂M,
u(0,·) = u0 in M,
ut(0,·) = u1 in M,
which is one of the basic models in control theory. Here M is a smooth,
compact, connected Riemannian manifold with boundary, a ∈ L∞(M ) and
a ≥ 0. The wave equation can be rewritten as a first order Cauchy problem
in the Hilbert space X = H 1
rater
rater
0 (M ) × L2(M ) with A defined by
dom(A) = (H 2(M ) ∩ H 1
A = (cid:18) O −I
−∆ a (cid:19) .
0 (M )) × H 1
0 (M ),
The operator A is invertible, and −A generates a non-analytic contraction
semigroup (T (t))t≥0. Since the natural norm on the Hilbert space X corre-
sponds to the energy of the system, any estimate for the rate of decay of the
semigroup is an estimate for the rate of decay of the energy of the system.
Recall that a semigroup is exponentially stable if and only if kT (t)k =
O(r(t)) with limt→∞ r(t) = 0. Hence, one cannot expect a uniform rate of
decay for all solutions if the semigroup is only stable but not exponentially
stable. Nevertheless, one may have uniform estimates of rates of decay for
a dense set of initial values. The study of the damped wave equation and
various similar PDEs has revealed that the resolvent can be used successfully
to obtain quantitative rates of convergence of the form
t → ∞,
kT (t)A−1k = O (r(t)) ,
that is, uniform rates for classical solutions.
(1.2)
In his pioneering work [51], Lebeau discovered that the spectrum of A is
contained in the strip {λ ∈ C : 0 < Re λ ≤ 2kak∞}, the resolvent (λ + A)−1
admits an exponential estimate on the imaginary axis and this forces the
energy of solutions of (1.1) to decay at least with rate
r(t) =
log(3 + log(3 + t))
log(3 + t)
,
t → ∞.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
3
His method was extended subsequently by Burq [18] and Lebeau and Rob-
biano [52] to cover similar situations involving local energy decay for (1.1)
where a = 0 and M is a non-compact manifold ("scattering in an exterior
domain"), and global energy decay for (1.1) with Neumann boundary con-
ditions, respectively. Moreover Lebeau's estimate was improved in [18] and
[52] to 1/ log(3 + t). These results gave rise to a number of papers treating
the rates of decay of solutions to PDEs by abstract semigroup methods.
Among them, we would like to mention [9], [16], [19], [21], [22], [53], and
especially the very recent paper [4] as samples. This last paper includes
a complete historical account, and a detailed discussion of damped wave
equations in the resolvent context.
The study of rates was put into the setting of Tauberian theorems for
Laplace transforms in [13], by regarding the resolvent as the Laplace trans-
form of the semigroup and partially inverting the Laplace transform. This
approach by pure complex analysis unified a number of known results (in-
cluding those of Burq, Lebeau and others), improved some of them and
applied to arbitrary growth of resolvents. In particular, the following result
was obtained.
chdu Theorem 1.1. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
X with generator −A. Assume that σ(A) ∩ iR is empty, and define
M (s) = sup{k(ir + A)−1k : r ≤ s},
Mlog(s) = M (s) (log(1 + M (s)) + log(1 + s)) .
s ≥ 1,
Then there exist positive constants C, C′, c, c′ such that
genestintro
genestintro
(1.3)
c′
M−1(C′t) ≤ kT (t)A−1k ≤
C
M−1
log (ct)
for all sufficiently large t.
Note that by [13, Proposition 1.3] any decay of kT (t)A−1k to zero as
t → ∞ implies that σ(A) ∩ iR is empty. However, in many situations the
spectrum approaches the imaginary axis asymptotically forcing the function
M to grow at infinity. So we shall say (slightly loosely) that if M (s) → ∞ as
s → ∞, then the resolvent, restricted to the imaginary axis, has a singularity
at infinity.
The upper estimate in (1.3) was established by examining a function-
theoretic method first used in unquantified form in [5] (see also [11]). The
gap between the lower and upper estimates is, in general, of "logarithmic
size" and any improvement of (1.3) should fall within that gap.
(Note,
however, that if M grows exponentially, then the lower and upper estimates
are of the same order.) It was conjectured in [13] that the gap could be
bridged in the case of Hilbert spaces, but one cannot expect rates better
log (ct)(cid:1)−1 for general Banach spaces. This conjecture was partially
than(cid:0)M−1
settled in [17]. It was proved in [17] that if M grows polynomially then it is
possible to obtain a characterization of decay of kT (t)A−1k which is optimal
4
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
in the sense that M−1
theorem holds.
log in (1.3) is replaced by M−1. Namely the following
borto Theorem 1.2. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X, with generator −A. Assume that σ(A) ∩ iR is empty, and fix α > 0.
Then the following statements are equivalent:
(i) k(is + A)−1k = O(sα),
(ii) kT (t)A−1k = O(t−1/α),
Moreover, it was shown in [17] that the logarithmic gap in (1.3) is unavoid-
able for semigroups on Banach spaces, even if the growth of the resolvent is
polynomial as in Theorem 1.2(i).
s → ∞.
t → ∞.
The result above has already found a number of applications to the study
of concrete PDEs, see e.g. [1], [2], [3], [8], [26], [29], [30], [32], [33], [34], [37],
[58], [59], [63], [66]. In view of its importance, Theorem 1.2 (and Theorem
1.1) raised the natural and important problem of obtaining similar results
for growth scales which are finer than polynomial ones, or showing that for
other scales results of this kind cannot be true.
Recall that the optimal function-theoretical analogue of Theorem 1.1 has
a logarithmic correction term as in the right-hand side of (1.3) [17, Theorem
3.8]. Thus, one cannot expect, in general, that a purely function-theoretical
approach could produce a version of Theorem 1.2 for more general resolvent
bounds, and the need for an alternative approach to the problem of sharp-
ening (1.3) becomes apparent. Such an approach based on operator theory
is proposed in this paper. Note that Theorem 1.2 itself is proved in [17] by
means of an auxiliary operator construction. The framework of the present
paper is different, more general, and in a sense more transparent.
The polynomial scale has a number of special (e.g. algebraic) properties
making it comparatively amenable to an operator-theoretical approach. For
other scales where these properties are not available, our approach must be
much more involved and the task of finding appropriate scales is rather non-
trivial. The scale which is closest to the polynomial one is the finer scale of
regularly varying functions, which are products of polynomials and slowly
varying functions. Such scales have been widely used as natural refinements
of polynomial scales in various areas of analysis, including number theory,
complex analysis and probability theory (see e.g. [15, Sections 6-8]). This
suggests that if there is a generalization of Theorem 1.2 to more general
classes of resolvent bounds, the scale of regularly varying functions should
be one of the first candidates for such a class.
In this paper, we develop a general operator-theoretical approach to the
study of optimal decay rates of operator semigroups within the fine scales
of regularly varying functions. Since our primary motivation comes from
the study of rates in (1.2) (the case of singularity at infinity), we first for-
mulate several partial cases of our results obtained below for the setting of
(1.2). We use the fact that the function (log s)β is slowly varying with de
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
5
Bruijn conjugate (log s)−β (see Theorem 5.6, Corollary 5.7 and Theorem
5.12 below).
regvarinf Theorem 1.3. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X, with generator −A. Assume that σ(A) ∩ iR is empty, and fix α > 0 and
β ≥ 0.
(a) The following statements are equivalent:
rvii1
(b) If
(i) k(is + A)−1k = O(cid:0)sα(log s)−β(cid:1) ,
(ii) kT (t)A−1k = O(cid:0)t−1/α(log t)−β/α(cid:1) ,
k(is + A)−1k = O(cid:16)sα(log s)β(cid:17) ,
kT (t)A−1k = O(cid:16)t−1/α(log t)ε+β/α(cid:17) ,
then, for every ε > 0,
s → ∞.
t → ∞.
s → ∞,
t → ∞.
rvii2
ratesatzero
ratesatzero
We do not know whether (b) is true for ε = 0. If so, the implication would
become an equivalence.
It is apparent that our methodology goes beyond clarifying (1.2) and it
allows one to deal with rates of decay of several other operator families stem-
ming from bounded C0-semigroups on Hilbert spaces. Thus it is significant
from the point of view of both abstract operator theory and applications.
Observe that if (T (t))t≥0 is a bounded C0-semigroup on a Hilbert space
X with generator −A, then for each x ∈ X the function −T (·)A−1x is a
primitive of T (·)x. Now instead of considering asymptotics of the primitives
of T (·)x, we may study asymptotics of the derivatives of T (·)x. In other
words, we may consider the orbits of the form T (·)Ax for x ∈ dom(A),
and so study long-time regularity of (T (t))t≥0. This type of asymptotic
behaviour of semigroups has not been treated systematically in the literature
so far. Some related but very partial results pertaining to orbits of analytic
semigroups were obtained in [25]. More results are available in the discrete
setting, see e.g. [24], [44], [45]and [57]. In this paper, the decay of T (·)x for
x ∈ ran(A) is studied systematically and the resulting structure appears to
be very similar to the one established in the study of decay rates for the
orbits (T (t))t≥0 starting from the domain of A and discussed above.
Since we are interested in a decay of T (t)Ax to zero that is uniform
with respect to x ∈ dom(A), the problem which we address in this case is to
quantify the decay of kT (t)A(I +A)−1k. Thus our task is to identify spectral
conditions on A and a corresponding "optimal" function r such that
kT (t)A(I + A)−1k = O(r(t)),
(1.4)
Remark that if kT (t)A(I + A)−1k → 0 as t → ∞, then the spectrum of A
does not meet iR \ {0} and the resolvent of A is bounded on i(R \ (−1, 1)),
see Theorem 6.10 below. Therefore the only singularity of the resolvent on
the imaginary axis may be at zero and we have a spectral situation which is
in a sense opposite to the situation considered above in the study of (1.2).
t → ∞.
6
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
We describe this situation by saying that the resolvent restricted to the
imaginary axis has a singularity at zero. Basic examples in the framework
of a singularity at zero are provided, in particular, by generators of bounded
eventually differentiable semigroups, for example those arising in the study
of delay differential equations [12], [27, Section VI.6].
To treat (1.4), it is natural to assume the boundedness of (λ+A)−1 outside
a neighbourhood of zero in iR, and to relate the decay of kT (t)A(I + A)−1k
to the growth of (λ+ A)−1 near zero. The first problem that we encountered
on this way was that the mere convergence of kT (t)A(I + A)−1k to zero was
wide open. This type of convergence can be considered as an extension of the
famous Katznelson-Tzafriri theorem (see Theorem 6.1) with L1-functions
replaced by certain bounded measures on the real half-line. Using a new
technique, we obtain an interesting generalization of the Katznelson-Tzafriri
theorem for a class of bounded measures. Moreover, we are able to derive
a partial analogue of Theorem 1.1 for the case of a singularity at zero thus
equipping our version of the Katznelson-Tzafriri theorem with rates. These
results are given below, see also Theorems 6.14 and 6.15.
katzintro Theorem 1.4. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X with generator −A. Assume that E := −iσ(A)∩ R is compact and of spec-
tral synthesis, and moreover the resolvent of A is bounded on i(R \ (−η, η))
for some η > 0.
(a) If µ is a finite measure on R+ such that its Fourier transform vanishes
on E, then
lim
t→∞kT (t)µ(T )k = 0,
where µ(T )x :=R ∞0 T (s)x dµ(s) for x ∈ X.
(b) Assume that (λ + A)−1 has a singularity at zero (so E = {0}), let
m(s) := sup{k(ir + A)−1k : r ≥ s},
s > 0,
and assume that lims→0+ sm(s) = ∞. Let ε ∈ (0, 1). Then there exist
positive constants c, C, Cε such that
cm−1(Ct) ≤ kT (t)A(I + A)−1k ≤ Cεm−1(t1−ε),
for all sufficiently large t.
Our technique of employing fine scales of regularly varying functions to
the study of orbits decay proves to be efficient also in the situation of a
singularity at zero. In particular, we obtain a counterpart of Theorem 1.3(a)
in that case, given in Theorem 1.5 (see also Theorem 7.7). Unfortunately,
Theorem 1.4(b) is not quite as strong as Theorem 1.1, because the correction
term involves tε instead of a logarithmic term. As a result, the theorem below
covers only the case when the resolvent grows slightly slower than a power
of s−1 (in analogy with Theorem 1.3(a)). The problem of characterizing
the decay of kT (t)A(I + A)−1k in the other case when the resolvent grows
slightly faster than a power of s−1 remains open.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
7
rateszerointro Theorem 1.5. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space,
with generator −A. Assume that σ(A) ∩ iR ⊂ {0}, and let α > 1, β ≥ 0.
The following statements are equivalent.
O(1),
s → 0,
s → ∞.
t → ∞.
(i) k(is + A)−1k =(O(cid:0)s−α(log(1/s))−β(cid:1) ,
(ii) kT (t)A(I + A)−1k = O(cid:0)t−1/α(log t)−β/α(cid:1) ,
Note that Theorem 1.5 for β = 0 provides a resolvent characterization
of polynomial rates of decay for kT (t)A(I + A)−1k. This characterization
holds for α = 1, too. The case when 0 < α < 1 does not arise since
k(is + A)−1k ≥ s−1 if 0 ∈ σ(A). For other results revealing properties of
decay rates for kT (t)A(I + A)−1k we refer to Sections 6 and 7.
The operator-theoretical approach of the present paper can also be used
successfully to treat decay rates of orbits of (T (t))t≥0 starting in dom(A) ∩
ran(A), combining the situations of a singularity at infinity and a singular-
ity at zero considered above. Since dom(A) ∩ ran(A) = ran(cid:0)A(I + A)−2(cid:1)
(Proposition 3.10) and since we are interested in uniform rates of decay, the
task of characterizing such rates can be considered as the task of character-
izing the property
ratesboth
ratesboth
(1.5)
kT (t)A(I + A)−2k = O(r(t)),
t → ∞,
in resolvent terms.
While (1.5) implies that σ(A) ∩ iR ⊂ {0}, it does not exclude growth of
the resolvent near zero and near infinity (along the imaginary axis). At the
same time, (1.5) imposes certain restrictions on the resolvent growth (see
Theorem 8.1) which might serve as a starting point for obtaining optimal
decay rates in (1.5). The following result (see Theorem 8.4) illustrates that
point. It is a generalization of Theorem 1.2 above.
Theorem 1.6. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X with generator −A, and assume that σ(A)∩iR = {0}. If there exist α ≥ 1
and β > 0 such that
resboth
resboth
(1.6)
then
and
k(is + A)−1k =(O(s−α),
O(sβ),
s → 0,
s → ∞,
kT (t)Aα(I + A)−(α+β)k = O(t−1),
t → ∞,
decayboth
decayboth
(1.7)
kT (t)A(I + A)−2k = O(t−1/γ),
where γ = max(α, β).
Conversely, if (1.7) holds for some γ > 0, then (1.6) holds for α =
max(1, γ) and β = γ.
8
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
After this review of the main results of this paper, we note that we derive
as by-products a number of results of independent interest. These include
abstract converses to interpolation inequalities (Theorem 4.3), estimates of
decay rates for semigroup orbits with bounded local resolvents (Theorem
4.7), an extension of the Katznelson-Tzafriri theorem to the setting of mea-
sure algebras (Theorem 6.14), and lower bounds for orbit decay in the Ba-
nach space setting (Corollaries 6.11 and 8.2).
1.2. Strategy. One of the main novelties of the paper is its operator-
theoretical method for deriving estimates for rates of decay which in many
cases happen to be sharp. Let us describe the method in some more detail.
If (T (t))t≥0 is a bounded C0-semigroup on a Hilbert space X with gen-
erator −A, then we look for decay rates of T (t)B for a bounded operator
B which takes one of the three forms: A−1, A(I + A)−1 and A(I + A)−2.
Given a (upper) bound M for the resolvent on the imaginary axis, we start
by establishing lower bounds for decay rates of T (t)B in terms of M . Such
bounds are known in the case of a singularity at infinity (see (1.3)), and
they are obtained in this paper in the case of a singularity at zero and in
the case of singularities at both zero and infinity (Corollaries 6.11 and 8.2).
To obtain an upper bound for decay rates of kT (t)Bk matching the lower
bound mentioned above we then proceed in three steps, of which only the
second is restricted to Hilbert spaces. First, we show that if the resolvent
(λ + A)−1 grows regularly on the imaginary axis, then it is bounded in the
right half-plane when restricted to the range of an associated operator W
involving a fractional power and a Bernstein function f of A, see Theorems
5.5 and 7.3. Second, in Theorem 4.7, we prove that if k(λ + A)−1Wk is
bounded in the right half-plane then kT (t)Wk decays like t−1 as t → ∞.
Third, using new abstract converses to interpolation inequalities for Bern-
stein functions (Theorem 4.3), we deduce that if kT (t)Wk decays like t−1,
then on the domain of A, or the range of A, or the intersection of the two, the
decay of (T (t))t≥0 is expressed in terms of f . In the context of Theorem 1.2
this step reduces to an application of the moment inequality for fractional
powers. For the finer scales of rates of decay, the argument is much more
subtle and its effect is to improve some a priori bounds for the semigroup.
The decay obtained in this way matches the lower bound in many cases and
then it is optimal (apart from Theorem 5.12 where the upper bound differs
from the lower bound by an arbitrarily small power of a logarithm).
Operator Bernstein functions were used successfully in [31] to deal with
rates in mean ergodic theorems for bounded C0-semigroups on Banach spaces.
While there are formal similarities between [31] and the present paper, the
problems treated here are much more involved and technically and ideo-
logically demanding. While the mean ergodic theorem for C0-semigroups
is a comparatively simple statement, the majority of stability conditions
for C0-semigroups are deep results with tricky proofs. This extends to the
framework of the study of rates.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
9
Our approach establishes a certain structure for dealing with rates in three
cases, when the resolvent of the semigroup generator has singularities at
infinity, or at the origin, or at both of them. While the structure has many
elements in common between the three cases, there are several essential
differences between them which have to be addressed separately.
In this paper the approach is applied to the class of regularly varying
rates which are close to polynomial rates. When the resolvent grows rapidly
and fairly regularly, the upper and lower bounds in Theorem 1.1 are of the
same order. However there are some rates which grow regularly but are
intermediate between polynomial and exponential, where the optimal rate
of decay of kT (t)A−1k in Theorem 1.1 is not established for semigroups on
Hilbert space. For very irregular (but arbitrarily fast) rates M of growth of
the resolvent it is not possible to improve Theorem 1.1 by replacing M−1
log
by M−1, even for semigroups of normal operators on Hilbert spaces (see
Propositions 5.1 and 6.13).
1.3. Notation and conventions. In this paper, X will be a complex Ba-
nach space, and will often be specified to be a Hilbert space. We let L(X)
denote the space of all bounded linear operators on Banach space X, and
the identity operator will be denoted by I. If A is a linear operator on X, we
denote the domain of A by dom(A), the range of A by ran(A), the spectrum
of A by σ(A) and the resolvent set by ρ(A). If A is closable, its closure is
written as A.
If B is another linear operator on X, then we take A + B and AB to be
the operators with
dom(A + B) = dom(A) ∩ dom(B),
dom(AB) = {x ∈ dom(B) : Bx ∈ dom(A)}.
A complex variable may be denoted by either z or λ. We shall use the
symbol ι to denote the identity function on domains in C. The closure of a
subset E of C will be denoted by E.
For ϕ ∈ (0, π] we shall let Σϕ := {λ ∈ C : arg λ < ϕ} be the sector of
angle ϕ in C. Note that Σπ is the slit plane C \ (−∞, 0]. We may write
C+ := Σπ/2 = {λ ∈ C : Re λ > 0},
R+ := [0,∞).
We shall consider integrals of functions f with respect to positive Radon
measures µ over (0,∞) or [0,∞). We shall write such integrals as
f (s) dµ(s) or
f (s) dµ(s),
respectively. If g is an increasing right-continuous function on (0,∞) and µ
is the associated Lebesgue-Stieltjes measure, we shall write
Z ∞
0+
Z ∞
0
Z ∞
0+
f (s) dg(s)
10
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
instead of
Z ∞
0+
f (s) dµ(s).
Let M b(R) denote the space of all complex Borel measures of bounded vari-
ation on R. For a ∈ R, we let δa be the Dirac measure at a.
We consider the spaces L1(R+) and M b(R+) as (closed) subspaces of
L1(R) and M b(R), respectively, by extending functions or measures on R+
by 0 on (−∞, 0). In addition, we view L1(R) as a closed subspace of M b(R).
The standard convolution of two measures µ1, µ2 ∈ M b(R) will be denoted
by µ1 ∗ µ2.
The Fourier transform of a function f ∈ L1(R; X) is defined by
Ff (ξ) :=ZR
e−iξtf (t) dt,
ξ ∈ R.
This definition of the Fourier transform holds in particular for f in the
Schwartz space S(R) of scalar-valued test functions and it then induces a
Fourier transform for all vector-valued, tempered distributions. The Fourier
transform of µ ∈ M b(R) is therefore given by
e−iξt dµ(t),
ξ ∈ R.
Fµ(ξ) :=ZR
When X is a Hilbert space, we shall also use the symbol F to denote the
Fourier transform induced on the Hilbert space L2(R; X), so that (2π)−1/2F
is a unitary operator.
We are interested in asymptotic properties of functions defined on inter-
vals of the form [a,∞) for some a > 0, with values in (0,∞). We shall
say that f and g are asymptotically equivalent, and we write f ∼ g, or
f (s) ∼ g(s), if
f (s)
g(s)
lim
s→∞
= 1.
This defines an equivalence relation on such functions and we shall in effect
be working with equivalence classes of functions. This viewpoint provides
a justification for sometimes not including precise statements about the
domains of our functions (provided that each domain contains some interval
of the form (a,∞)) or repeatedly saying that an inequality holds for all
sufficiently large s.
Where we use the notation f (s) ∼ g(s), it will mean asymptotic equiv-
alence as s → ∞ unless otherwise specified. Occasionally we shall use the
corresponding notation as s → 0+, but then it will be specified. Thus, for
positive functions f and g defined on (0, a], the notation
means
f (s) ∼ g(s),
s → 0+,
f (s)
g(s)
lim
s→0+
= 1.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
11
For an increasing function f : [a,∞) → (0,∞) such that lims→∞ f (s) =
∞, the notation f−1 may denote the inverse function, or more generally a
right inverse, of f , defined on the range of f , so that f (f−1(t)) = t for all t
in the range. It may also denote an asymptotic inverse of f , defined on an
interval [b,∞), such that
(1.8)
f−1(f (s)) ∼ s,
f (f−1(s)) ∼ s.
It should be clear from the context which notion of inverse function is in-
volved.
We shall occasionally use the notation f α to denote function s 7→ f (s)α
when α 6= −1, but we shall use 1/f to denote the reciprocal function of f , in
order to avoid confusion with any inverse function. Similarly f g or f.g will
denote a pointwise product of two functions f and g, and f ◦ g will denote
composition.
We shall use C and c to denote (strictly) positive constants, whose values
may change from place to place.
2. Preliminaries on some classes of functions
In this section we review various classes of functions on (0,∞) with em-
phasis on the properties that we shall need.
2.1. Bernstein functions, complete Bernstein functions and Stielt-
jes functions. In this subsection, we recall the definitions and some prop-
erties of complete Bernstein functions and Stieltjes functions. Most of this
material can be found in [64].
In Section 3 we shall review the operator
functional calculus associated with these functions, and that will be used in
later sections of the paper.
Recall that a function f ∈ C∞(0,∞) is completely monotone if
for every n ∈ N ∪ {0}, λ ∈ (0,∞).
(−1)nf (n)(λ) ≥ 0
By Bernstein's theorem [64, Theorem 1.4], every completely monotone func-
tion f is the Laplace transform of a positive Radon measure on R+, and
f extends to a holomorphic function in the right half-plane. A function
f ∈ C∞(0,∞) is called a Bernstein function if
f ≥ 0 and f′ is completely monotone.
Clearly, every Bernstein function also extends to a holomorphic function in
the right half-plane. By the L´evy-Khintchine representation theorem [64,
Theorem 3.2], a function f is a Bernstein function if and only if there exist
constants a, b ≥ 0 and a positive Radon measure µLK on (0,∞) such that
dµLK(s) < ∞,
s
0+
s + 1
Z ∞
f (λ) = a + bλ +Z ∞
and
asyinv
asyinv
subsectfun
hpfc.e.bf
hpfc.e.bf
(2.1)
(1 − e−λs) dµLK(s),
λ > 0.
0+
12
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
The triple (a, b, µLK) is uniquely determined by the corresponding Bernstein
function f and is called the L´evy-Khintchine triple of f .
The class of Bernstein functions is rather large and to ensure good al-
gebraic and function-theoretic properties of Bernstein functions it is conve-
nient, and also sufficient for many purposes, to consider the subclass con-
sisting of complete Bernstein functions. A function f ∈ C∞(0,∞) is called
a complete Bernstein function if it is a Bernstein function and the measure
µLK in the L´evy-Khintchine triple has a completely monotone density with
respect to Lebesgue measure [64, Definition 6.1]. By [64, Theorem 6.2],
every complete Bernstein function admits a representation of the form
compbern
compbern
(2.2)
f (λ) = a + bλ +Z ∞
0+
λ
s + λ
dµ(s),
λ > 0,
for some constants a, b ≥ 0 and some positive Radon measure µ on (0,∞)
satisfying
mmummu
(2.3)
Z ∞
0+
dµ(s)
s + 1
< ∞.
berviahirsch Remark 2.1. Complete Bernstein functions admit various representations
In particular, the following formula is used in some
different from (2.2).
papers related to Bernstein functions (for example in [40], [41], [42], [43],
[62]):
diffrepr
diffrepr
(2.4)
f (λ) = a +Z ∞
0
λ
1 + λt
where ν is a positive Radon measure on R+ satisfying
λ
1 + λt
dν(t),
dν(t) = a + ν({0})λ +Z ∞
Z ∞
dν(t)
1 + t
< ∞,
0+
0
and the pair (a, ν) is unique.
The representations (2.4) and (2.2) are equivalent by the change of vari-
able s = 1/t, with ν being the push-forward measure of µ combined with an
atom of mass b at 0, and vice versa.
There are striking dualities between complete Bernstein functions and
another class known as Stieltjes functions. A function h ∈ C∞(0,∞) is
a Stieltjes function if there exist constants a, b ≥ 0 and a positive Radon
measure µ on (0,∞) satisfying (2.3) such that
dµ(s)
,
(2.5)
s + λ
h(λ) =
λ > 0.
a
λ
+ b +Z ∞
0+
The representation formulas (2.2) (for complete Bernstein functions) and
(2.5) (for Stieltjes functions) are unique, and they are called the Stieltjes
representations for f and h, respectively; see e.g. [64, Chapter 2]. We write
hpfc.e.stieltjes
hpfc.e.stieltjes
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
13
f ∼ (a, b, µ) and h ∼ (a, b, µ), and we call (a, b, µ) the Stieltjes triple, and µ
the Stieltjes measure for f and h, respectively. Note that
a = lim
λ→0+
f (λ) = lim
λ→0+
λ h(λ),
b = lim
λ→∞
f (λ)
λ
h(λ).
= lim
λ→∞
We shall be particularly interested in the Stieltjes (and complete Bernstein)
functions with Stieltjes representations of the form (0, 0, µ). When µ is the
Lebesgue-Stieltjes measure associated with an increasing right-continuous
function g we shall denote the Stieltjes function with representation (0, 0, µ)
by Sg, and we shall call it the Stieltjes function associated with g.
Note that every (complete) Bernstein function f is increasing and every
Stieltjes function is decreasing. Comparison of (2.2) and (2.5) shows that
if h ∈ C∞(0,∞) is a Stieltjes function then f (λ) := λh(λ) is a complete
Bernstein function, and conversely if f ∈ C∞(0,∞) is a complete Bernstein
function then h(λ) := f (λ)/λ is a Stieltjes function. The classes of complete
Bernstein functions and Stieltjes functions are preserved under various op-
erations [64, Theorem 6.2, Theorem 7.3, Corollary 7.4, Corollary 7.6]. We
present here some which will be used in the paper.
sti-char Theorem 2.2. Let f be a non-zero function on (0,∞).
(a) f is a complete Bernstein function if and only if 1/f (λ) is a Stieltjes
function.
(b) If f is a complete Bernstein function, then λ/f (λ) and λf (1/λ) are
complete Bernstein functions. Conversely, if λ/f (λ) or λf (1/λ) is a
complete Bernstein function, then f is a complete Bernstein function.
compcbf
(c) If f and g are both complete Bernstein functions or both Stieltjes func-
tions, then g ◦ f is a complete Bernstein function.
Many examples of complete Bernstein functions, and hence of Stieltjes
functions, are given in [64, Chapter 15]. We give here a few of the most
elementary examples that will be relevant in this paper.
contex Example 2.3. (a) The function h(λ) := λ−γ (γ ∈ (0, 1)) is a Stieltjes func-
tion with the Stieltjes representation
h(λ) =
sin πγ
π
1
s + λ
ds
sγ ,
λ > 0.
Z ∞
0
Accordingly, f (λ) = λ1−γ = λh(λ) is a complete Bernstein function.
(b) For α ∈ (0, 1), the function f (λ) := λ(λ+1)−α is a complete Bernstein
function with the Stieltjes representation
f (λ) =
sin πα
π
Z ∞
1
λ
s + λ
ds
(s − 1)α .
Moreover, λ(λ + 1)−1 is a complete Bernstein function with Stieltjes triple
(0, 0, δ1).
(c) It follows from (a) and (b), together with Theorem 2.2(c), that f (λ) :=
λα(1 + λ)−β is a complete Bernstein function whenever 0 ≤ β ≤ α ≤ 1.
14
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Any Stieltjes function or complete Bernstein function can be extended to
a holomorphic function on the slit plane Σπ by means of the formula (2.5)
or (2.2), respectively. We shall often regard the functions as being defined
on the slit plane in this way, without explicit comment. The rate of decay or
growth of such functions at infinity in sectors is determined by the rate on
(0,∞), as shown by the following known fact [61, Lemma 2], [49, Proposition
2.21 (c)].
prop.asymp Proposition 2.4. Let g be either a Stieltjes function or a complete Bern-
stein function with Stieltjes representation (0, 0, µ). Let λ ∈ Σπ and ϕ =
arg λ. Then
cos(ϕ/2) g(λ) ≤ g(λ) ≤ sec(ϕ/2) g(λ).
Proof. The second inequality follows from (2.5) or (2.2) and the elementary
inequality:
s + λ2 ≥
For the first inequality, we can assume that g 6= 0. By Theorem 2.2, 1/g is
a complete Bernstein function or a Stieltjes function. Hence
(s + λ)2 =(cid:0) cos(ϕ/2) (s + λ)(cid:1)2,
s ∈ (0,∞).
(1 + cos ϕ)
2
1
g(λ) ≤
1
cos(ϕ/2) g(λ)
.
(cid:3)
2.2. Slowly and regularly varying functions. Most of the material in
this subsection is standard (see [15, Chapter 1], [47, Sections IV.1-9] or [65]).
slowvar Definition 2.5. Let ℓ be a strictly positive measurable function defined on
[a,∞) for some a ∈ R and satisfying
ℓ(λs)
ℓ(s)
lim
s→∞
= 1
for every λ > 0. Then ℓ is said to be slowly varying.
Clearly the value of a is not important in this definition. By defining
ℓ(s) = ℓ(a) for s ∈ [0, a], we may assume that a = 0.
Example 2.6. (a) Standard examples of slowly varying functions include
logk(s) := log ... log s,
iterated logarithms
exp {(log s)α1(log2(s))α2 ...(logk(s))αk )} ,
exp(cid:18) log s
log log s(cid:19).
k ∈ N,
αi ∈ (0, 1), 1 ≤ i ≤ k,
(b) It is a straightforward consequence of Definition 2.5 that the sum and
product of two slowly varying functions is slowly varying. Moreover, if ℓ is
slowly varying then the following are also slowly varying.
ℓα : s 7→ ℓ(s)α,
ℓα : s 7→ ℓ(sα),
ℓ. log : s 7→ ℓ(s) log s.
α ∈ R,
α > 0,
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
15
A proof of the following representation theorem, originally due to Kara-
mata for continuous functions ℓ, may be found in [15, Theorem 1.3.1] or [47,
Section IV.3].
slowvarrep Theorem 2.7. The function ℓ is slowly varying if and only if it is of the
form
ℓ(s) = c(s) exp(cid:26)Z s
a
ε(t)
t
dt(cid:27) ,
s ≥ a,
for some a > 0, where c and ε are measurable functions, c(s) → c > 0 and
ε(s) → 0 as s → ∞.
The following corollary is easily deduced from Theorem 2.7; see [15, The-
orem 1.5.6], [65, p.18].
cor.sv Corollary 2.8. Let ℓ be a slowly varying function, and γ > 0.
(a) There are positive constants C, c such that
t(cid:17)γ
c(cid:16) s
≤
ℓ(t)
ℓ(s) ≤ C(cid:18) t
s(cid:19)γ
for all sufficiently large s, t with t ≥ s.
(b) As s → ∞,
(2.6)
svfsvf
sγℓ(s) → ∞,
s−γℓ(s) → 0.
On the other hand there are slowly varying functions such that
ℓ(s) = 0,
lim inf
s→∞
lim sup
s→∞
ℓ(s) = ∞.
regvar Definition 2.9. A positive function f is called regularly varying with index
α ∈ R if there is a slowly varying function ℓ such that
f (s) = sαℓ(s),
s ≥ a.
Such a function has a representation
f (s) = sα c(s) exp(cid:26)Z s
ε(t)
t
dt(cid:27) ,
s ≥ a,
a
where c and ε are as in Theorem 2.7.
If f is regularly varying of index α > 0, there is a strictly increasing, regu-
larly varying function g which is asymptotically equivalent to f [15, Theorem
1.5.3]. One can also arrange that g is smooth [15, Theorem 1.8.2]. More-
over, f has an asymptotic inverse in the sense of (1.8). For example, one
may take the asymptotic inverse of f to be the usual inverse of a strictly in-
creasing, continuous, function which is asymptotically equivalent to f . The
asymptotic inverse is regularly varying, it depends only on the asymptotic
equivalence class of f , and it is unique up to asymptotic equivalence. In-
deed, the asymptotic equivalence classes of regularly varying functions with
positive index form a group under composition; they also form a semigroup
under pointwise multiplication [15, Theorem 1.8.7].
16
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
A convenient way to handle asymptotic inverses of regularly varying func-
tions involves the de Bruijn conjugate ℓ# of the slowly varying function ℓ
[15, Section 1.5.7]. This is a slowly varying function ℓ# such that
ℓ(s)ℓ#(sℓ(s)) → 1
One can take
and
ℓ#(s)ℓ(sℓ#(s))) → 1
(ι.ℓ)−1(s)
,
s
ℓ#(s) =
as s → ∞.
where (ι.l)(s) = sℓ(s). For this choice of ℓ#, it is easy to see that if ℓ
is increasing (resp., decreasing), ℓ# is decreasing (resp., increasing). The
group structure of the asymptotic equivalence classes of regularly varying
functions immediately implies that one-sided asymptotic inverses are unique
up to asymptotic equivalence. Hence, if k is a slowly varying function and
either ℓ(s)k(sℓ(s)) → 1 or k(s)ℓ(sk(s))) → 1 as s → ∞, then k ∼ ℓ#.
ex.dBinv Example 2.10. A method for finding many de Bruijn conjugates is given
in [15, Appendix 5], including the following cases.
1. If ℓ(s) = (log s)β where β ∈ R, then ℓ#(s) ∼ (log s)−β.
2. Let ℓ(s) = exp(cid:0)(log s)β(cid:1) where 0 < β < 1.
(a) If 0 < β ≤ 1/2, then ℓ#(s) ∼ exp(cid:0)−(log s)β(cid:1).
2 ≤ β < 2
(b) If 1
3 , then
ℓ#(s) ∼ exp(cid:16)−(log s)β + β(log s)2β−1(cid:17) ,
(1/ℓ)#(s) ∼ exp(cid:16)(log s)β + β(log s)2β−1(cid:17) .
For values of β between 2/3 and 1, there are longer formulas of this type.
regvarinv Proposition 2.11. Let ℓ be slowly varying, and let α > 0. Then
rvi1
rvi2
rvi3
(a) ℓ## ∼ ℓ.
(b) If f (s) ∼ sαℓ(sα), then f−1(s) ∼ s1/αℓ#(s)1/α.
(c) If g : (0, a] → (0,∞) and g(s) ∼ sα/ℓ(s−α) as s → 0+, then g−1(s) ∼
s1/α/ℓ#(1/s)1/α as s → 0+.
Proof. The statements (a) and (b) are in [15, Section 1.5] and [65, Section
1.6]. For (c), note that g(s) ∼ 1/f (1/s) as s → 0+, where f is as (b). Using
the regular variation, one can easily deduce that g−1(s) ∼ 1/f−1(1/s) as
s → 0+.
(cid:3)
The following lemma describes a common situation in which ℓ# has a
particularly simple form. Parts of the lemma appear in [15, Section 1.5] and
[65, Section 1.6].
regvarinv2 Lemma 2.12. Let ℓ be a slowly varying function. The following are equiv-
alent:
dB1
dB2
dB3
(i) ℓ# ∼ 1/ℓ.
(ii) ℓ (sℓ(s)) ∼ ℓ(s).
(iii) ℓ (s/ℓ(s)) ∼ ℓ(s).
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
17
If ℓ is monotonic and α > 0, these properties are equivalent to each of the
following:
(iv) ℓ (sℓ(s)α) ∼ ℓ(s).
(v) (ℓα)# ∼ 1/ℓα, where ℓα(s) = ℓ(sα).
Proof. The statements (ii) and (iii) are equivalent to the statements
ℓ(s)k(sℓ(s)) → 1,
k(s)ℓ(sk(s)) → 1
respectively, when k = 1/ℓ. So their equivalence to (i) follows from the
one-sided uniqueness properties of ℓ# set out before Example 2.10.
Now assume that (ii) holds and ℓ is monotonic. Replacing s by sℓ(s) gives
Moreover, by (ii),
so
ℓ(cid:16)sℓ(s)ℓ(cid:0)sℓ(s)(cid:1)(cid:17) ∼ ℓ(cid:0)sℓ(s)(cid:1) ∼ ℓ(s).
ℓ(cid:0)sℓ(s)2(cid:1) ∼ ℓ(cid:16)sℓ(s)ℓ(cid:0)sℓ(s)(cid:1)(cid:17) ∼ ℓ(s)
sℓ(s)2 ∼ sℓ(s)ℓ(cid:0)sℓ(s)(cid:1),
ℓ (sℓ(s)α) ∼ ℓ(s)
by the Uniform Convergence Theorem [15, Theorem 1.2.1]. Iterating this
argument gives
whenever α is a power of 2. Then the monotonicity of ℓ gives it for all α > 0.
Thus (ii) implies (iv). The converse follows by replacing ℓ(s) by ℓ(s)1/α.
Finally, let kα = 1/ℓα. Assume that (iv) holds. Replacing s by sα in (iv)
gives
ℓα(s)kα(sℓα(s)) =
ℓ(sα)
ℓ (sαℓ(sα)α) → 1.
Then (v) follows from the one-sided uniqueness property of (ℓα)#. Replacing
α by 1/α shows that (v) implies (iv).
(cid:3)
We shall say that a monotonic, slowly varying, function ℓ is dB-symmetric
when the conditions of Lemma 2.12 are satisfied. Example 2.10 shows that
the following functions ℓ are dB-symmetric:
If ℓ is dB-symmetric, it is clear from Lemma 2.12 that the following functions
are also dB-symmetric:
ℓ(s) = (log s)β, for any β ∈ R,
ℓ(s) = exp(cid:0)(log s)β(cid:1), if 0 < β < 1
ℓα : s 7→ ℓ(sα), for any α > 0,
ℓα : s 7→ ℓ(s)α, for any α ∈ R.
2 .
It is not difficult to show that the product of two dB-symmetric functions
is dB-symmetric.
For a regularly varying function f , let f.log be the following regularly
varying function:
We shall need the following relations.
(f.log)(s) = f (s) log s.
18
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
lem.logpert Lemma 2.13. Let f (s) = sαℓ(s) be a regularly varying function with α > 0,
logpert1
logpert2
and let δ > 1/α. The following hold for some constant C.
(log f (s))δ(cid:19) ≤ Cf (s).
(i) (f.log)(cid:18)
(ii) f−1(s) ≤ C(f.log)−1(s)(log s)δ.
When ℓ is increasing, these statements are also true for δ = 1/α.
s
Proof. (i). By Corollary 2.8, log f (s) ∼ α log s. Also,
(f.log)(cid:18)
s
(log f (s))δ(cid:19) = f (s)
s
(log f (s))αδ
log(cid:16)
(log f (s))δ(cid:17)
ℓ(cid:16)
≤ Cf (s)(log s)1−αδ
≤ Cf (s)(log s)1−αδ+δγ
s
ℓ(cid:16)
(log f (s))δ(cid:17)
(log f (s))δ(cid:17)
ℓ(s)
ℓ(s)
s
for any γ > 0 by Corollary 2.8, and for γ = 0 if ℓ is increasing.
(ii). Replacing s by f−1(s) in (i) gives
(f.log)(cid:18) f−1(s)
(log s)δ(cid:19) ≤ Cs.
The claim follows since (f.log)−1 is increasing and regularly varying.
(cid:3)
Finally in this section, we consider Stieltjes functions associated with
regularly varying functions.
stfng Example 2.14. Let ℓ be a slowly varying function on R+, α ≥ 0, and
assume that g(s) := sαℓ(s) is increasing. The associated Stieltjes function
Sg(λ) :=Z ∞
0+
dg(s)
s + λ
ℓ(s)
s+1 ds is finite.
=Z ∞
0
sαℓ(s)
(s + λ)2 ds,
or if α = 1 andR ∞0
is defined if either integral is finite [68, p.7]. This occurs if α < 1, by (2.6),
We shall need the following abelian/Tauberian theorem of Karamata. For
the main results we shall need only the abelian parts (i)⇒(ii) and (iii)⇒(iv).
karamata Theorem 2.15 (Karamata). Let g be an increasing function on R+, and
let Sg be the associated Stieltjes function. Let 0 < σ ≤ 1, and ℓ be slowly
varying on R+.
(a) The following are equivalent:
kar1
kar2
(b) The following are equivalent:
s → ∞;
(i) g(s) ∼ s1−σℓ(s)
(ii) Sg(λ) ∼ Γ(σ)Γ(2 − σ)λ−σℓ(λ),
(iii) g(s) ∼ s1−σℓ(1/s),
s → 0+;
(iv) Sg(λ) ∼ Γ(σ)Γ(2 − σ)λ−σℓ(1/λ),
λ → ∞.
λ → 0+.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
19
Proof. The first statement is proved in [15, Theorem 1.7.4] (the Tauberian
implication is proved in [65, Theorem 2.5]). The proof of the second state-
ment is very similar, using the same preliminary results from [15, Sections
1.5,1.7].
(cid:3)
s.fc
3. Functional calculus of sectorial operators
In this section we recall basic properties of functional calculus of sectorial
operators based on complete Bernstein and Stieltjes functions, originally
due to Hirsch and extended to include fractional powers. We concentrate on
those properties which are needed for our main purposes in the later sections
of this paper. A much fuller account of the calculus for Bernstein functions
can be found in the monograph [64], and of the extended calculus in [36].
3.1. Sectorial operators. Many parts of the paper will involve the notion
of a sectorial operator which we recall now.
Definition 3.1. A densely defined, linear operator A on a Banach space X
is called sectorial if (−∞, 0) ⊂ ρ(A) and there exists C > 0 such that
(3.1)
λ > 0.
kλ(λ + A)−1k ≤ C,
Some authors require a sectorial operator to be injective, and some do
not require it to be densely defined. Some of the operators that we consider
will not be injective.
Note that any sectorial operator A is closed and by the Neumann series
expansion, (λ + A)−1 is defined and (3.1) holds on the sector Σϕ, for some
ϕ ∈ (0, π]. Moreover,
(3.2)
λ(λ + A)−1x → x, λ → ∞,
x ∈ X,
sector
sector
approxid
approxid
[36, Proposition 2.1.1,d)].
If −A generates a bounded C0-semigroup, then A is sectorial. The follow-
ing standard lemma establishes that certain auxiliary operators which play
important roles in this paper are also sectorial even though they may not
be negative generators of bounded semigroups.
sectops Lemma 3.2. Let A be a sectorial operator on a Banach space X. Then the
operators A−1 (if A is invertible), A(I + A)−1 and A(I + A)−2 are sectorial.
Proof. Sectoriality of A−1 and A(I + A)−1 follows from the identities
1
λ + 1
+
1
=
(λ + A−1)−1 = λ−1 − λ−2(λ−1 + A)−1,
(λ + 1)2(cid:18) λ
which hold for λ > 0 (see [36, Proposition 2.1.1,b)] and [60, Lemma 3.1]).
To prove sectoriality of A(I + A)−2, we note the identity
(cid:0)λ + A(I + A)−1(cid:1)−1
λ(λ + A(I + A)−2)−1 = I − λ−1(µ + A)−1(cid:16)I − µ−1(cid:0)µ−1 + A(cid:1)−1(cid:17) ,
+ A(cid:19)−1
λ + 1
,
20
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
where µ > 1 and µ−1 are the roots of µ2−(λ−1 +2)µ+1 = 0. Since µ > λ−1,
sectoriality of A implies that kλ−1(µ + A)−1k and kµ−1(cid:0)µ−1 + A(cid:1)−1 k are
bounded for λ > 0, hence A(I + A)−2 is sectorial.
(cid:3)
Other information about sectorial operators may be found in [36] and [55]
(in the latter sectorial operators are called non-negative operators).
3.2. Hirsch functional calculus. We now define complete Bernstein func-
tions of sectorial operators and review those basic properties that we need
in the sequel. There are several different definitions of functional calculus of
sectorial operators, and we shall describe some properties which cross over
between the different definitions. We try to present the ideas of functional
calculus in a way which reveals the heuristics of our subsequent arguments,
and to give the later proofs in ways which do not rely on any unjustified
assumptions about compatibility of different definitions.
Let f be a complete Bernstein function with Stieltjes representation (a, b, µ),
and let A be a sectorial operator on a Banach space X. The next definition
was essentially given in [41, p. 255].
defoperbern Definition 3.3. Define an operator f0(A) : dom(A) → X by
defbernop
defbernop
(3.3)
f0(A)x = ax + bAx +Z ∞
0+
A(λ + A)−1x dµ(λ),
x ∈ dom(A).
By (2.3), this integral is absolutely convergent and f0(A)(I + A)−1 is a
bounded operator on X, extending (I + A)−1f0(A). Hence f0(A) is closable
as an operator on X. Define
hirschdef
hirschdef
We call f (A) a complete Bernstein function of A.
f (A) = f0(A).
Actually Hirsch used the representation (2.4) for a complete Bernstein
function f and then defined the corresponding operator f (A) as
f H(A) = f H
0 (A),
where f H
0
is defined on dom(A) by
0 (A)x := a+ν({0})Ax+Z ∞
0+
(3.4) f H
A(I +λA)−1x dν(λ),
x ∈ dom(A).
By Remark 2.1 the representations (2.2) and (2.4) describe exactly the same
classes of complete Bernstein functions. To see that the resulting operators
f (A) and f H(A) coincide, it suffices to change variables in the same way
as in Remark 2.1, considering vector-valued integrals instead of scalar ones.
Thus we can use Hirsch's results even if we think of complete Bernstein
functions f as being represented by (2.2).
When −A generates a bounded C0-semigroup, one can define the op-
erators f (A) for arbitrary Bernstein functions f by adapting the L´evy-
Khintchine formula (2.1). A detailed discussion of this approach as well
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
21
as of various properties of f (A) can be found in [64, Section 12] (where A
denotes the generator of the semigroup). One can prove that the definition
of complete Bernstein functions of semigroup generators in [64] is consistent
with Hirsch's definition (Definition 3.3), but we shall not go into details.
we shall refer to some results in [64] for complete Bernstein functions even
though they start from the L´evy-Khintchine representation.
By Definition 3.3, dom(A) is a core for f (A). Using (3.2) and the fact
that (λ + A)−1 commutes with (I + A)f0(A)(I + A)−1, we infer that dom(A)
is a core for (I + A)f0(A)(I + A)−1 as well. Since (I + A)f0(A)(I + A)−1 is
closed as a product of a closed operator and a bounded operator, it follows
that
f (A) = (I + A)f0(A)(I + A)−1.
boundedcbf Remark 3.4. If f is a bounded complete Bernstein function then, in the
Stieltjes representation of f , b = 0 and Fatou's lemma implies that the
measure µ is finite (see [64, Corollary 3.7(v)]). In this case, f0(A) is bounded
on dom(A), and therefore f (A) is a bounded operator on X.
It is also
straightforward to see that f (A) is bounded for any complete Bernstein
function f if A is bounded (see [64, Corollary 12.7]).
Remark 3.5. In the sequel, (3.3) will usually be used for measures µ which
are Lebesgue-Stieltjes measures associated with increasing, regularly varying
functions g. Then we may think of the integrals as being vector-valued
Riemann-Stieltjes integrals rather than Bochner integrals. The theory of
vector-valued Riemann-Stieljes integrals is presented in [6, Section 1.9], [39,
Section III.3.3], [68, Section 1]. One can define complete Bernstein functions
and operator Bernstein functions initially by means of Riemann-Stieltjes
integrals. However, in the relevant literature, including [41], [43], [62] and
[64], (2.2) and (3.3), or (2.4) and (3.4), are standard ways to define complete
Bernstein functions, and we have chosen to follow an established route.
Complete Bernstein functions of sectorial operators possess a number of
properties which allow one to create a partial functional calculus for A.
However the set of complete Bernstein functions is not closed under point-
wise multiplication, so the multiplicative properties of this process are re-
stricted. The subject was thoroughly investigated by Hirsch in the 1970s,
and he proved the following properties of operator Bernstein functions in
[41, Th´eor`eme 1-3] and [43, Th´eor`eme 1] (see also [43, pp. 200-201], [62] and
[40]).
hirsch Theorem 3.6. Let A be a sectorial operator on a Banach space X, and let
f and g be complete Bernstein functions. Then the following statements
hold.
hiri
hirii
(i) The operators f (A) and g(A) are sectorial.
(ii) The composition rule holds:
f (g(A)) = (f ◦ g)(A).
22
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
hiriii
(iii) If f g is also a complete Bernstein function, then the product rule holds:
f (A)g(A) = (f g)(A).
For α ∈ (0, 1), Example 2.3(a) shows that zα is a complete Bernstein
function, and we write Aα for the corresponding complete Bernstein function
of A. These fractional powers coincide with the standard fractional powers
which are extensively studied in [55]. If A is sectorial and α, β ∈ (0, 1) then
Theorem 3.6 implies that (Aα)β = Aαβ, and moreover AαAβ = Aα+β if
α + β ≤ 1. In this paper we shall need these properties for a larger range of
α and β. This is standard theory, but we put it in a broader context here.
To this aim, we shall use an extended holomorphic functional calculus,
which has become a standard tool to deal with functions of sectorial op-
erators. To keep the presentation within reasonable limits we give only a
very brief sketch of part of the extended holomorphic calculus and refer for
further details to [36, Chapters 1,2] and [48, Sections II.9, II.15].
Let H be the algebra of functions which are holomorphic in Σπ = C \
(−∞, 0]. Let H stand for the set of all f ∈ H such that for any ϕ ∈ (0, π)
there exist c ∈ C and C > 0, α > 0 (both depending on ϕ) satisfying
(3.5)
f (z) − c ≤ Czα,
z ∈ Σϕ.
Let eH denote the set of all f ∈ H such that for any ϕ ∈ (0, π) one has
f (z) ≤ C max(zα,z−α),
z ∈ Σϕ,
(3.6)
infinf
zeroinf
zeroinf
for some C, α > 0.
contains all complete Bernstein functions and all Stieltjes functions, and
one can set α = 1 for all ϕ ∈ (0, π).
X.
It follows easily from (2.2), (2.3) and (2.5) that eH
Let C(X) denote the set of all closed, densely defined, linear operators on
calculus Theorem 3.7. Let A be a sectorial operator on a Banach space X. Then
there exists a well-defined mapping
called an extended holomorphic functional calculus, such that
H 7→ C(X),
f
7→ f (A),
calci
calcii
calciii
calciv
calcv
(i) 1(A) = I and z(A) = A;
(ii) If T ∈ L(X) and T A ⊂ AT , then T f (A) ⊂ f (A)T ;
(iii) f (A) + g(A) = (f + g)(A), if g(A) ∈ L(X);
(iv) f (A)g(A) = (f g)(A), if g(A) ∈ L(X);
(v) if g ∈ H is such that g(A) is sectorial and there exists ϑ : (0, π) → (0, π)
such that limϕ→π− ϑ(ϕ) = π and g(Σϑ(ϕ)) ⊂ Σϕ for all ϕ ∈ (0, π), then
(f ◦ g)(A) = f (g(A)).
If A is injective then there exists a mapping eH 7→ C(X), f 7→ f (A), satisfying
the properties (i)-(v) above.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
23
calcrem Remark 3.8. We note the following facts about the extended holomorphic
calcremciii
calcremci
calcremcii
functional calculus described in Theorem 3.7.
(i) If f (A) ∈ L(X) and λ ∈ ρ(A), then f (A)(λ + A)−1 = (λ + A)−1f (A).
(ii) If f is a rational function whose poles all lie in (−∞, 0], then f (A) as
defined in the extended holomorphic functional calculus of Theorem
3.7 agrees with the natural definition of f (A).
(iii) If f ∈ H is a complete Bernstein function, then f (A) as in the calculus
of Theorem 3.7 agrees with f (A) as defined in Definition 3.3. This is
shown in [14, Theorem 4.12] for injective A. The proof in the general
case is the same up to replacement of the regulariser (z/(1 + z)2)n by
the regulariser (1 + z)−2.
(iv) The composition rule Theorem 3.7(v) applies in particular when g(z) =
z−1 (if A is invertible), z(1+z)−1, or z(1+z)−2. In each case, it suffices
to note Lemma 3.2 and to use the fact that the mapping z 7→ 1/z
preserves sectors.
(v) In particular, if A is invertible, g ∼ (0, 0, ν) is a Stieltjes function and
f is the complete Bernstein function given by f (z) = g(1/z), then
g(A) =Z ∞
0+
(λ + A)−1dν(λ) = f (A−1).
It is easy to see that {zα : α > 0} ⊂ H and {zα : α ∈ R} ⊂ H. Thus if A
is sectorial, then the "fractional powers" Aα for α > 0 (and for all α ∈ R,
if A is invertible) are well-defined by means of the extended holomorphic
functional calculus and, as we shall see below, they behave very similarly to
the scalar functions zα. These fractional powers coincide with the fractional
powers considered in [36, Chapters 3,4], [48, Section II.15] and [54, Chapter
2] where all properties of fractional powers of operators needed in this paper
can be found. For the reader's convenience we recall and summarize them
in the following statement.
fractional Theorem 3.9. Let A be a sectorial operator on a Banach space X. Then
sgp
shiftdom
powers
for any α > 0 and β > 0 the following statements hold.
(i) If T ∈ L(X) and T A ⊂ AT then T Aα ⊂ AαT .
(ii) The semigroup law holds: Aα+β = AαAβ.
(iii) dom(Aβ) = dom((I + A)β) = ran((I + A)−β).
(iv) If in addition α ∈ (0, 1), then Aα is sectorial, and the composition law
If A is invertible then A−α = (A−1)α for α > 0, (ii) is true for α, β ∈ R,
and (iv) holds for α ∈ (−1, 1) and β ∈ R.
holds: (Aα)β = Aαβ.
The following facts are elementary when α and β are integers (Lemma
3.2 and [48, Proposition 9.4(d)]) and they may be known for fractional
powers, but we were unable to find them in the literature. The fact that
(−∞, 0) is contained in the resolvent set of Aα(I + A)−1 when 0 < α < 1
follows from [35, Theorem 4.1], but the method given there does not establish
24
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
the sectoriality estimate. We are very grateful to Alexander Gomilko for
providing the proof of that estimate.
prop.aab Proposition 3.10. Let A be a sectorial operator on a Banach space X, and
domim
aabsect
let α, β ∈ [0, 1]. The following statements hold:
(i) ran(Aα(I + A)−(α+β)) = ran(Aα) ∩ dom(Aβ).
(ii) Aα(I + A)−β is sectorial.
Proof. To prove (i), we shall use Theorem 3.9 and the consequential facts
that Aα(I + A)−α and (I + A)−β are bounded commuting operators, and
ran(Aα) = ran(Aα(I + A)−α). Hence
ran(Aα) ∩ dom(Aβ) = ran(Aα(I + A)−α) ∩ ran((I + A)−β)
⊃ ran(cid:0)Aα(I + A)−(α+β)(cid:1).
Conversely, let x ∈ ran(Aα)∩dom(Aβ). Then x = Aαy1 and x = (I +A)−βy2
for some y1 ∈ dom(Aα), y2 ∈ X. Since Aαy1 = (I + A)−βy2 ∈ dom(Aβ),
one has y1 ∈ dom(Aα+β) = dom((I + A)α+β). Let y3 = (I + A)α+βy1. Then
x = Aαy1 = Aα(I + A)−(α+β)y3.
For (ii), first consider the case when 0 < β ≤ α ≤ 1. By Example 2.3(b),
f (λ) := λα(1 + λ)−β is a complete Bernstein function. By Theorem 3.6
(i),(ii), Aα(I + A)−β = f (A) which is sectorial.
Now consider the case when 0 < α < β = 1. Let λ > 0, and
fα(z) =
zα
1 + z
,
and
gα,λ(z) =
1
λ −
1
fα(z) + λ
=
fα(z)
λ(fα(z) + λ)
=
zα
λ (zα + λ(1 + z))
,
for z ∈ Σπ. For ϕ ∈ (0, π), let γϕ be the contour given by
γϕ := {te−iϕ : t ≥ 0} ∪ {teiϕ : t ≥ 0},
taken in the downward direction. Then
galint
galint
(3.7)
and
gα,λ(µ) =
1
galform
galform
Z0Z0
Letting ϕ → π−, we obtain
sin πα
(3.8)
gα,λ(µ) =
π
0
Letting µ → 0+ we also have
(3.9)
Z ∞
0
dz,
< ∞,
z
gα,λ(z)
µ − z
Zγϕ gα,λ(z)dz
2πiZγϕ
Z ∞
eπiαtα + λ(1 − t)2 (µ + t)−1 dt.
(1 − t)tα−1
arg µ < ϕ.
(1 − t)tα
eπiαtα + λ(1 − t)2 dt = 0.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
25
By the product rule of Theorem 3.7(iv), fα(A) = Aα(I + A)−1. From
(3.7) and [36, Section 2.3],
gα,λ(A) =
1
2πiZγϕ
gα,λ(z)(z + A)−1 dz,
for ϕ ∈ (0, π) sufficiently large. Here the integral is absolutely convergent,
so gα,λ(A) ∈ L(X). Since
the product rule implies that
galinv
galinv
(3.10)
(λ + fα(z))(cid:0)λ−1 − gα,λ(z)(cid:1) = 1,
λ−1 − gα,λ(A) =(cid:0)λ + Aα(I + A)−1(cid:1)−1
.
The same arguments as for (3.8) show that
(1 − t)tα
gα,λ(A) =
sin πα
Z ∞
0
π
From this, (3.1) and (3.9) we obtain
eπiαtα + λ(1 − t)2 (t + A)−1 dt.
Z ∞
Z 1
eπiαtα + λ(1 − t)2 dt
eπiαtα + λ(1 − t)2 dt.
(1 − t)tα−1
1 − ttα−1
0
0
π
π
C sin πα
2C sin πα
kgα,λ(A)k ≤
=
For t ∈ (0, 1),
for some cα > 0. Hence
Now
and
eπiαtα + λ(1 − t)2 = t2α + (1 − t)2λ2 + 2 cos(πα)λtα(1 − t)
0
tα−1
≥ cα(cid:0)t2α + (1 − t)2λ2(cid:1) ,
(1 − t)tα−1
t2α + (1 − t)2λ2 dt.
kgα,λ(A)k ≤ CαZ 1
t2α + (1 − t)2λ2 dt ≤ Z 1/2
(1 − t)tα−1
αZ ∞
t2α + (1 − t)2λ2 dt ≤ 21−αZ 1
= 21−αZ 1/2
2αλ2 log 1 +(cid:18) 2αλ
(1 − t)tα−1
0
1
≤
dτ
1/2
=
1
0
0
Z 1/2
0
Z 1
1/2
t2α + (λ/2)2 dt
τ 2 + (λ/2)2 =
π
αλ
,
(1 − t)
2−2α + (1 − t)2λ2 dt
2−2α + τ 2λ2 dτ
τ
2 (cid:19)2! ≤
1
λ
,
26
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
since
Thus
log(1 + s2) ≤ 2 log(1 + s) ≤ 2s,
s > 0,
It follows from (3.10) that Aα(I + A)−1 is sectorial when 0 < α < 1.
kλgα,λ(A)k ≤ Cα, λ > 0.
When 0 ≤ α < β < 1, we have
Aα(I + A)−β =(cid:16)Aα/β(I + A)−1(cid:17)β
,
and this is sectorial, by the previous case together with Theorem 3.9 (iv). (cid:3)
Let A be a sectorial operator and f be a complete Bernstein function.
Then one has
ineqbern
ineqbern
(3.11)
kf (A)xk ≤ Ckxkf(cid:18)kAxk
kxk (cid:19) ,
x ∈ dom(A), x 6= 0,
where C is a constant independent of x (and f ). This is shown in [62]
for sectorial operators, and in [64, Corollary 12.8] where −A is assumed to
be the generator of a bounded C0-semigroup but f may be any Bernstein
function.
If A is invertible, we also have
ineqbern1
ineqbern1
(3.12)
kxk (cid:19) ,
kf (A−1)xk ≤ Ckxkf(cid:18)kA−1xk
x ∈ X, x 6= 0.
This follows since A−1 is also sectorial. Alternatively, one can easily pass
between (3.11) and (3.12) by considering ϕ(z) := zf (1/z). Then ϕ is also
a complete Bernstein function by Theorem 2.2, and ϕ(A) = Af (A−1) by
the product rule and composition rule for g(z) = z−1, in Theorem 3.7 and
Remark 3.8(v). Applying (3.11) with f replaced by ϕ gives
kf (A−1)Axk ≤ CkAxkf(cid:18) kxk
kAxk(cid:19) ,
x ∈ dom(A), x 6= 0.
Setting Ax = y we obtain (3.12). Conversely, applying (3.12) with f replaced
by ϕ gives
kf (A)A−1xk ≤ CkA−1xkf(cid:18) kxk
kA−1xk(cid:19) ,
x ∈ X, x 6= 0.
Setting A−1x = y (so y ∈ dom(A)) we obtain (3.11).
for fractional powers in the forms
When f (z) = zα (0 < α < 1), we recover the classical moment inequality
momineq0
momineq1
(3.13)
(3.14)
kAαxk ≤ Ckxk1−αkAxkα,
kA−αxk ≤ Ckxk1−αkA−1xkα,
x ∈ dom(A),
x ∈ X.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
27
4. Some estimates for semigroup asymptotics
In this section we give some results relating different types of asymptotic
estimates for semigroups.
In Subsection 4.1 we present some inequalities
which are related to (3.12) and (3.13) but apply to generators of bounded
semigroups. In Subsection 4.2 we show how certain resolvent estimates can
be transferred to semigroup estimates in the case of bounded semigroups on
Hilbert space.
ss.inter
4.1. Moment and interpolation inequalities. We start by recalling the
full moment inequality for sectorial operators, extending (3.13) and (3.14).
prop.momineq Proposition 4.1. Let B be a sectorial operator, and 0 ≤ α < β < γ. There
is a constant C such that
momineq4
momineq4
momineq3
momineq3
kBβxk ≤ CkBαxk(γ−β)/(γ−α)kBγxk(β−α)/(γ−α),
x ∈ dom(Bγ).
(4.1)
Hence if S : X → dom(Bγ) is a linear operator and BγS ∈ L(X), then
(4.2)
kBβSk ≤ CkBαSk(γ−β)/(γ−α)kBγSk(β−α)/(γ−α).
If B is invertible, then (4.1) and (4.2) hold whenever α < β < γ.
Proof. For α = 0, (4.1) is the standard inequality (3.13) [48, Theorem 15.14],
[54, Corollary 5.1.13]. The more general cases follow by replacing β by β−α,
γ by γ − α and x by Bαx. Then (4.2) follows on replacing x by Sx. When
B is invertible the range of the inequalities can be extended by replacing x
or S by B−nx or B−nS.
(cid:3)
Next we deduce an inequality of interpolation type which was proved
in a slightly less general form (and with a slightly different proof) in [9,
Proposition 3.1].
interpoldec Lemma 4.2. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
X, and let B ∈ L(X) be a sectorial operator commuting with (T (t))t≥0. Let
γ, δ > 0. Then there exist positive constants C, c such that
momineq
momineq
(4.3)
ckT (Ct)Bγkδ ≤ kT (t)Bδkγ ≤ CkT (ct)Bγkδ,
In particular, the following statements are equivalent:
t > 0.
(i) kT (t)Bγk = O(t−1),
(ii) kT (t)Bk = O(t−1/γ),
t → ∞.
t → ∞.
Proof. Take n ∈ N such that nγ ≥ δ, and apply Proposition 4.1 with α = 0,
β = δ/n and S = T (t/n). Then
kBδ/nT (t/n)k ≤ CkT (t/n)k1−δ/(nγ)kBγT (t/n)kδ/(nγ) ≤ CkBγT (t/n)kδ/(nγ).
Now
kT (t)Bδkγ ≤ kT (t/n)Bδ/nknγ ≤ CkT (t/n)Bγkδ.
This gives the second inequality in (4.3), and the first follows by interchang-
ing γ and δ. The final statement follows by taking δ = 1.
(cid:3)
28
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Our next result gives more interpolation properties for the generator of a
bounded C0-semigroup. We shall need them for our main results in Sections
5 and 7. To simplify the presentation here and in Section 7 we introduce
the shorthand notation:
B(A) := A(I + A)−1
when A is a sectorial operator. By Lemma 3.2, B(A) is sectorial. Thus the
fractional powers B(A)α, α > 0, are well-defined, and by the product and
composition rules in Theorem 3.7(iv) and Remark 3.8(iv),
powerrules
powerrules
(4.4)
B(A)α = Aα(I + A)−α.
interpol2 Theorem 4.3. Let −A be the generator of a bounded C0-semigroup (T (t))t≥0
on a Banach space X, and let B(A) = A(I + A)−1. Assume that T (t)A 6= 0
for each t > 0. There exists a constant c > 0 such that the following hold.
(a) If A is invertible, f is a complete Bernstein function, γ ≤ 1, and
int2a
Aγf (A−1) is a bounded operator, then
intineq
intineq
(4.5)
int2b
(b) If f is a bounded complete Bernstein function and γ > 0, then
for all t1, t2 ≥ 0.
(cid:13)(cid:13)T (t1)Aγf (A−1)(cid:13)(cid:13) ≥ ckT (t1 + t2)Aγ−1k
kT (t1)B(A)γf (A)k ≥ ckT (t1 + t2)B(A)γ+1k
kT (t2)A−1k
kT (t2)B(A)k
for all t1, t2 ≥ 0.
f (kT (t2)A−1k)
f (kT (t2)B(A)k)
Remark 4.4. In our applications of Theorem 4.3, we shall take t1 = t2 and
choose specific values of γ > 0, but the applications are rather delicate. For
example, the ratio
kT (2t)Aγ−1k
kT (t)A−1k
tends to 0 as t → ∞ in all the cases in which we are interested in Section 5.
The inequality (4.5) will be used to improve other estimates for the rate of
decay of kT (t)A−1k.
Proof. (a) We can assume that f is non-zero, and we let ϕ be the complete
Bernstein function given by ϕ(z) = z/f (z), z > 0 (Theorem 2.2). Since f ,
ϕ and the identity function z are complete Bernstein functions, the product
rule (Theorem 3.6(iii)) yields
Then
f (A−1)ϕ(A−1) = A−1.
kT (t1 + t2)Aγ−1k = kT (t1 + t2)Aγf (A−1)ϕ(A−1)k
≤ kT (t1)Aγf (A−1)kkT (t2)ϕ(A−1)k.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
29
Thus
interpol3
interpol3
(4.6)
kT (t1)Aγf (A−1)k ≥ kT (t1 + t2)Aγ−1k
kT (t2)ϕ(A−1)k
.
We now estimate kT (t)ϕ(A−1)k from above for t > 0. Let ϕ have Stieltjes
representation (a, b, µ). By (3.3),
T (t)ϕ(A−1) = aT (t) + bT (t)A−1 +Z ∞
T (t)A−1(cid:0)λ + A−1(cid:1)−1 dµ(λ).
Let τ = kT (t)A−1k. Then
0+
dµ(λ)
0+
0+
τ +
since A−1 is sectorial and (T (t))t≥0 is bounded. Moreover,
(cid:13)(cid:13)(cid:13)(cid:13)Z τ
T (t)A−1(cid:0)λ + A−1(cid:1)−1 dµ(λ)(cid:13)(cid:13)(cid:13)(cid:13) ≤ CZ τ
T (t)A−1(cid:0)λ + A−1(cid:1)−1 dµ(λ)(cid:13)(cid:13)(cid:13)(cid:13) ≤ τZ ∞
≤ CτZ ∞
dµ(λ) + CτZ ∞
(cid:13)(cid:13)(cid:13)(cid:13)Z ∞
kT (t)ϕ(A−1)k ≤ aK + bτ + CZ τ
≤ aK + bτ + 2CZ τ
Let K = supt≥0 kT (t)k. Then
λ + τ
τ +
λ
τ +
dµ(λ) + 2CZ ∞
τ +
.
λ
dµ(λ)
0+
0+
τ
τ + (cid:13)(cid:13)(cid:0)λ + A−1(cid:1)−1(cid:13)(cid:13) dµ(λ)
dµ(λ)
τ
λ + τ
dµ(λ)
≤ 2Cϕ(τ ).
Putting t = t2, (4.6) gives
kT (t1)Aγf (A−1)k ≥ ckT (t1 + t2)Aγ−1k
ϕ(kT (t2)A−1k)
= ckT (t1 + t2)Aγ−1k
kT (t2)A−1k
f (kT (t2)A−1k).
(b) The proof is similar to (a). We now use the product rule in the form
f (A)ϕ(A) = A.
Observe also that f (A) is bounded (Remark 3.4), ϕ(A) is closed and dom(A) ⊂
dom(ϕ(A)) (Definition 3.3), and hence the operator ϕ(A)(I+A)−1 is bounded.
So
kT (t1 + t2)B(A)γ+1k = kT (t1 + t2)B(A)γf (A)ϕ(A)(I + A)−1k
≤ kT (t1)B(A)γ f (A)kkT (t2)ϕ(A)(I + A)−1k.
Thus
kT (t1)B(A)γf (A)k ≥ kT (t1 + t2)B(A)γ+1k
kT (t2)ϕ(A)(I + A)−1k
.
30
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Now
T (t)ϕ(A)(I + A)−1
= aT (t)(I +A)−1 +bT (t)A(I +A)−1 +Z ∞
Let τ = kT (t)A(I + A)−1k = kT (t)B(A)k. Estimating as in a) gives
T (t)A(λ+A)−1(I +A)−1 dµ(λ).
0+
kT (t)ϕ(A)(I + A)−1k ≤ 2Cϕ(τ ).
The claim follows as in (a).
(cid:3)
Remark 4.5. The inequality (3.12) easily implies that
kf (A−1)k ≤ Cf (kA−1k),
for a constant C independent of the complete Bernstein function f . On the
other hand, taking t1 = t2 = γ = 0 in (4.5) we obtain a reversed inequality
kf (A−1)k ≥ cf (kA−1k).
The two inequalities together form an operator counterpart to Proposition
2.4.
Remark 4.6. Theorem 4.3 can be generalized in various ways. For example,
if B is any bounded sectorial operator commuting with T (t) for all t > 0 (in
particular, if B = B(A)), then the following version of (4.5) holds:
kT (t1)Bγf (B)k ≥ ckT (t1 + t2)Bγ+1k
kT (t2)Bk
f (kT (t2)Bk)
for γ > 0. It also holds for γ ≥ −1 if B is injective and Bγf (B) is a bounded
operator (in particular, if B = A−δ for 0 < δ ≤ 1).
4.2. Transference from resolvents to semigroups. The final estimate
of this section is for bounded semigroups on Hilbert space. The following
result shows how the effect of cancelling resolvent growth can be transferred
to an estimate for the semigroup itself. When B = A−α, the result was
obtained in [17, Theorem 2.4]. In our applications the operator B will be a
function of the generator A, such as the operator Wα,β,ℓ(A) of Subsection
5.2.
ss.transfer
thm.CRbound Theorem 4.7. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X, with generator −A, and let B : dom(A) → X be a linear operator which
is bounded for the graph norm on dom(A), and such that T (t)Bx = BT (t)x
for all t ≥ 0 and x ∈ dom(A). Assume that
(4.7)
sup(cid:8)kB(λ + A)−1k : λ ∈ C+(cid:9) < ∞.
Then T (t)B extends to a bounded linear operator (also denoted by T (t)B)
on X for each t > 0, and kT (t)Bk = O(t−1) as t → ∞.
CRbnd
CRbnd
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
31
Proof. Let x ∈ dom(A). For a fixed τ > 0 define
fτ (t) =(T (t)x, 0 ≤ t ≤ τ,
t > τ,
0,
and
ϕτ (t) = (BT ∗ fτ )(t) =(tBT (t)x,
τ BT (t)x,
0 ≤ t ≤ τ,
t > τ.
0
C 2Z τ
0
e−2atkT (t)xk2 dt,
Let bfτ and bϕτ be the Laplace transforms of these functions, so that bϕτ (λ) =
B(λ + A)−1bfτ (λ) for λ ∈ C+. By Plancherel's theorem, for a > 0,
e−2atkϕτ (t)k2 dt
t2e−2atkBT (t)xk2 dt.
ZR kbϕτ (a + is)k2 ds = 2πZR
≥ 2πZ τ
ZR kbϕτ (a + is)k2 ds = ZR kB(a + is + A)−1bfτ (a + is)k2 ds
Letting C be the finite supremum in (4.7), we have
again by Plancherel's theorem. These two inequalities imply that
≤ C 2ZR kbfτ (a + is)k2 ds
= C 22πZ τ
e−2atkT (t)xk2 dt ≥Z τ
0 kT (t)xk2 dt ≥Z τ
C 2Z τ
τ Z τ
C 2K 2kxk2 ≥
hτ T (τ )Bx, yi = (cid:12)(cid:12)(cid:12)(cid:12)
thT (t)Bx, T ∗(τ − t)yi dt(cid:12)(cid:12)(cid:12)(cid:12)
τ Z τ
t2kT (t)Bxk2 dt(cid:27) 1
0 kT ∗(τ − t)yk2 dt(cid:27) 1
2(cid:26) 2
≤ (cid:26) 2
τ Z τ
τ Z τ
t2kT (t)Bxk2 dt(cid:27) 1
≤ (cid:26) 2
τ Z τ
2 √2Kkyk
≤ 2CK 2kxkkyk.
Let K = supt≥0 kT (t)k. Then
(4.8)
Letting a → 0+ one gets
t2e−2atkBT (t)xk2 dt.
Hence, for any y ∈ X,
t2kBT (t)xk2 dt.
0
t2kBT (t)xk2 dt.
0
0
0
2
0
0
0
CRces
CRces
1
2
aZR+
α
πZR(cid:13)(cid:13)B(α + is + A)−2x(cid:13)(cid:13)2 ds,
t2e−atkBT (t)xk2 dt =
αZR kB(α + is + A)−2xk2 ds ≤ Ckxk2,
by Plancherel's theorem. Thus the assumption (4.7) can be replaced by
a = 2α > 0,
α > 0, x ∈ X.
32
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
This implies that T (τ )B has a bounded extension to X with norm at most
2CK 2/τ .
(cid:3)
Remark 4.8. We do not know whether there is a converse of Theorem 4.7,
for example whether (4.7) holds whenever B is a bounded operator on X,
commuting with T (t) and satisfying kT (t)Bk = O(t−1) as t → ∞. There is
a result in function theory [38, Lemma 2.5] which says that under these as-
sumptions the boundary function of hB(·+A)−1x, yi on iR lies in BMO(iR).
The crux of the proof of Theorem 4.7 is the estimate (4.8) showing that
the Ces´aro means of the scalar function t 7→ t2kBT (t)xk2 are bounded. For
a positive measurable function boundedness of its Abel means is equivalent
to boundedness of its Ces´aro means. On Hilbert space, the Abel means of
this function are
On Hilbert space, the bounded operator-valued function B(· + A)−1 is an
L2(R+; X)-Laplace multiplier, as shown by means of Plancherel's Theorem.
Theorem 4.7 holds for a C0-semigroup on a Banach space X provided that
B(· + A)−1 is an Lp(R+; X)-Laplace multiplier for some p ∈ [1, +∞), in the
sense that the convolution operator f 7→ BT ∗ f is bounded on Lp(R+, X).
The proof is as in Theorem 4.7 with an application of Holder's inequality
replacing the Cauchy-Schwarz inequality.
s.infinity
5. Singularity at infinity
In this section we shall consider a bounded C0-semigroup (T (t))t≥0, with
generator −A, on a Hilbert space X under the assumption that σ(A)∩ iR is
empty. First we recall and elaborate Theorem 1.1, where X is any Banach
space but the other assumptions are the same.
The spectral assumption that σ(A) ∩ iR is empty is equivalent to the
property that
TtATtA
(5.1)
lim
t→∞kT (t)(ω + A)−1k = 0
mm
The rate of decay in (5.1) is closely related to the growth of the resolvent
for any ω ∈ ρ(−A) [6, Theorem 4.4.14]. We choose to take ω = 0.
of A on iR. Let M be a function such that
k(is + A)−1k ≤ M (s),
s ∈ R.
(5.2)
We shall always make the natural assumption that M is even, so we shall
consider M as being a function on R+. It is also natural to assume that M
is increasing and continuous.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
33
Define also
defMlog
defMlog
(5.3)
s ≥ 0.
It is shown in [13, Theorem 1.5] (see also [6, Theorem 4.4.14]) that
Mlog(s) = M (s)(cid:0) log(1 + M (s)) + log(1 + s)(cid:1),
genest+
genest+
(5.4)
kT (t)A−1k = O(cid:16)
1
M−1
log (ct)(cid:17),
t → ∞,
for any c ∈ (0, 1).
given by
The smallest function M satisfying (5.2) and our other assumptions is
defM
defM
(5.5)
M (s) = sup(cid:8)k(ir + A)−1k : r ≤ s(cid:9) ,
s ≥ 0.
For this choice of M it is a simple consequence of [13, Proposition 1.3] (see
also [6, Theorem 4.4.14]) that there exist constants c, C > 0 such that
genest-
genest-
(5.6)
kT (t)A−1k ≥
c
M−1(Ct)
for all sufficiently large t. Here we assume that lims→∞ M (s) = ∞ and M−1
may be any right inverse of M .
The estimates (5.4) and (5.6) are both valid for bounded semigroups on
any Banach space. They raise the question whether, or when, it is possible
to improve (5.4) to
genest++
genest++
(5.7)
kT (t)A−1k = O(cid:16)
1
M−1(ct)(cid:17),
t → ∞.
In some cases, for example if M (s) = eαs for α > 0, (5.4) and (5.7) are
equivalent. In many cases, each estimate is independent of c. In the case
when M (s) = sα for α > 0, the two estimates differ by a logarithmic factor.
In this case, (5.4) is optimal for arbitrary Banach spaces, but (5.7) holds
when X is a Hilbert space [17]. However, for some M one cannot make
this improvement even for normal operators on Hilbert space [6, Example
4.4.15]. We give a more detailed analysis of normal semigroups in Subsection
5.1.
In later subsections, we consider cases when X is a Hilbert space and M
is regularly varying. In Subsection 5.2 we shall show that if −A generates a
bounded C0-semigroup on a Banach space X and σ(A) ∩ iR is empty, then
for regularly varying functions M , the property (5.2) is equivalent to
cancelled
cancelled
(5.8)
k(λ + A)−1fM (A)k ≤ C,
λ ∈ C+,
where fM (A) is defined by the extended functional calculus of Theorem 3.7
for an appropriate function fM related to the classes of Bernstein functions
and Stieltjes functions discussed in Subsection 2.1. In Subsections 5.3 and
5.4, we pass from (5.8) towards (5.7).
ss.normal
34
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
5.1. Normal semigroups. The following result gives the precise condition
on M which is both necessary (if M is defined by (5.5)) and sufficient for
(5.7) to be valid for a semigroup of normal operators on Hilbert space. In
fact, it holds more generally for any bounded C0-semigroup for which the
norms of all the associated bounded operators are determined by σ(A), that
is,
kT (t)r(A)k = sup(cid:8)e−λtr(λ) : λ ∈ σ(A)(cid:9)
for every rational function r whose poles are outside σ(A) and which is
bounded at infinity. This includes multiplication semigroups on Lp-spaces
and spaces of continuous functions. We call such C0-semigroups quasi-
multiplication semigroups.
normalinf Proposition 5.1. Let (T (t))t≥0 be a quasi-multiplication semigroup on a
Banach space X with generator −A. Assume that σ(A) ⊂ C+ and inf{Re λ :
λ ∈ σ(A)} = 0. Let M be defined by (5.5) and let c > 0. The following are
equivalent:
normalcond1
(i) There exists C such that
normalest1
normalest1
normalcond2
normalres
normalres
(5.9)
kT (t)A−1k ≤
(ii) There exists B such that
M (s) ≥ c log(cid:16) τ
M (τ )
(5.10)
s(cid:17) − B,
C
M−1(ct)
,
t ≥ c−1M (0);
τ > 0, s ≥ 1.
Proof. Note first that kT (t)k = 1 for all t ≥ 0, and
M (s)−1 = min{µ − ir : µ ∈ σ(A),r ≤ s} → 0,
s → ∞.
Now (5.9) is equivalent to
e−tα
µ ≤
M−1(ct)
This may be rewritten as
C
normalest2
normalest2
(5.11)
for all such µ and t.
,
µ = α + iβ ∈ σ(A), t ≥ c−1M (0).
Cµ (cid:19)
tα ≥ log(cid:18) M−1(ct)
Assume that (i) holds. Let t ≥ c−1M (0), and put τ = M−1(ct). From
(5.11),
M (τ ) ≥
c
α
log(cid:18) τ
Cµ(cid:19) .
Given s ≥ 1, take µ = α + iβ ∈ σ(A) such that M (s)−1 = µ − ir for some
r ≤ s. Then α ≤ k, where k = M (0)−1, and β ≤ s + k. So
M (τ )
M (s) ≥
cα + i(β − r)
α
log(cid:18)
τ
C(s + 2k)(cid:19) ≥ c log(cid:18)
τ
C(s + 2k)(cid:19) .
By (5.10),
τ = M−1(ct),
s = β.
ct
M (β) ≥ c log(cid:18) M−1(ct)
β (cid:19) − B.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
35
Since log(C(s + 2k))− log s ≤ log((2k + 1)C) for s ≥ 1, it follows that (5.10)
holds whenever τ is in the range of M−1. For other values of τ one can
apply the above with τ replaced by τn := M−1(M (τ ) + n−1) > τ , and let
n → ∞.
with β > 1, take
Now assume that (ii) holds. Given t ≥ c−1M (0) and µ = α + iβ ∈ σ(A)
Rearranging this, using αM (β) ≥ 1 and µ ≥ β, and putting C =
exp(B/c) gives (5.11), provided that β ≥ 1.
If there exist µ = α + iβ ∈ σ(A) with β ≤ 1, let
Putting s = √τ and then s = 1 in (5.10) shows that
ε = inf {α : α + iβ ∈ σ(A),β ≤ 1} > 0.
normalest3
normalest3
(5.12) M (τ ) ≥(cid:16) c
2
log τ − B(cid:17) M (√τ ) ≥(cid:16) c
2
log τ − B(cid:17)2
for all sufficiently large τ . Putting τ = M−1(ct) shows that
M (1) ≥
c2(log τ )2
5ε
c2(log M−1(ct))2
5ε
,
ct ≥
and hence
for all sufficiently large t. We can then choose C sufficiently large that
εt ≥ log M−1(ct)
Cε (cid:19)
εt ≥ log(cid:18) M−1(ct)
whenever t ≥ c−1M (0). Then (5.11) holds for µ = α + iβ ∈ σ(A) with
β ≤ 1 and t ≥ c−1M (0). Hence (5.9) holds.
(cid:3)
It is clear that (5.10) implies that M grows at least logarithmically. This
corresponds to the elementary fact that, for a quasi-multiplication semi-
group, kT (t)A−1k cannot decrease faster than exponentially. The estimate
(5.12) shows that (5.10) implies that M (τ ) grows at least as fast as (log τ )2.
More generally, any slowly varying function M fails to satisfy (5.10). Given
B and c > 0, choose λ = e(B+2)/c. Then (5.10) implies that
M (λs)
M (s) ≥ 2,
s ≥ 1,
so M is not slowly varying. In particular, the rate of decay of kT (t)A−1k for
a quasi-multiplication semigroup cannot be given by (5.9) unless the rate is
slower than exp(−ct1/n) for all n.
when M grows logarithmically.
The following example shows the rate of decay for normal semigroups
36
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
log-growth Example 5.2. Let (T (t))t≥0 be a quasi-multiplication semigroup on a Ba-
nach space X, with generator −A such that
Then
σ(A) =(cid:26) 1
+ is : s ≥ 2(cid:27) .
log s ≤ k(is − A)−1k = M (s) ≤ log(s + 1),
log s
1
However,
M−1(ct) ∼ e−ct.
kT (t)A−1k = sup(cid:26) exp(−t/ log s)
s
: s ≥ 2(cid:27) = e−2√t,
s ≥ 2,
t ≥ (log 2)2.
ss.cancel
If M (s) = (1 + s)α for some α > 0, then (5.10) holds. More generally, if
M is regularly varying with index α > 0, then M satisfies (5.10).
On the other hand, rapid growth of M does not on its own imply that
(5.10) holds. See [6, Example 4.4.15].
5.2. Cancelling resolvent growth. Here we shall show how regularly
varying growth of k(is + A)−1k as s → ∞ can be cancelled by restrict-
ing to the range of a suitable operator. In the case of purely polynomial
growth this was achieved by taking a (negative) fractional power of A [50,
Lemma 3.2], [17, Lemma 2.3], but we shall need a more complicated function
of A.
The following lemma of Phragm´en-Lindelof type may be known, but we
have not been able to trace it in the literature. The formulation stated here
is more general than is needed for this section, but we shall need the stronger
form, or variants of it, when we consider singularities at zero in Sections 7
and 8.
phragmen Lemma 5.3. Let Y be a Banach space and f : C+ \ {0} → Y be a function
which is continuous on C+\{0}, holomorphic in C+, and bounded on iR\{0}.
If there exists C > 0 such that
kf (z)k ≤
C
Re z
,
z ∈ C+,
then f is bounded in C+ \ {0}.
Proof. Let K = sup{kf (z)k : z ∈ iR, z 6= 0}. For fixed r, R with 0 < r <
1 < R, let
Consider
Ar,R = {z ∈ C+ : r ≤ z ≤ R}.
gr,R(z) :=
z
1 + z(cid:18)1 +
z2
R2(cid:19)(cid:18)1 +
r2
z2(cid:19) f (z),
z ∈ Ar,R.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
37
Let
so that
γr = {z ∈ C+ : z = r}, ΓR = {z ∈ C+ : z = R},
Jr,R = {is : r ≤ s ≤ R},
If z ∈ γr then
so that
∂Ar,R = γr ∪ ΓR ∪ Jr,R.
z
1 + z(cid:12)(cid:12)(cid:12)(cid:12) ≤
(cid:12)(cid:12)(cid:12)(cid:12)
1 + z(cid:12)(cid:12)(cid:12)(cid:12) ≤
(cid:12)(cid:12)(cid:12)(cid:12)
z
r
1 − r
,
kgr,R(z)k ≤
,
R − 1
(cid:12)(cid:12)(cid:12)(cid:12)1 +
=
C
Re z
2 Re z
,
R
r
(cid:12)(cid:12)(cid:12)(cid:12)1 +
(cid:12)(cid:12)(cid:12)(cid:12)1 +
r
z2
4 Re z
R2(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2,
R2(cid:12)(cid:12)(cid:12)(cid:12) =
z2
1 − r
,
.
r
r2
2 Re z
4C
1 − r
z2(cid:12)(cid:12)(cid:12)(cid:12) =
z2(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2,
(cid:12)(cid:12)(cid:12)(cid:12)1 +
r2
Similarly, if z ∈ ΓR then
R
thus
Finally, if z ∈ Jr,R then
Applying the maximum principle to gr,R over Ar,R, letting r → 0+ and
R → ∞ and setting C′ = 4 max(C, K), we infer that
In particular, f is bounded on {z ∈ C+ : z ≥ 1} and
.
kgr,R(z)k ≤
4C
R − 1
kgr,R(z)k ≤ 4K.
z
1 + z
sup
z∈C+(cid:13)(cid:13)(cid:13)(cid:13)
f (z)(cid:13)(cid:13)(cid:13)(cid:13) ≤ C′.
kf (z)k ≤
,
z ≤ 1.
2C′
z
Now, by applying a standard Phragm´en-Lindelof principle for half-planes
[23, Corollary VI.4.2] to h(z) := f (z−1) for z ∈ C+, we conclude that f is
bounded on {z ∈ C+ : z ≤ 1} and this completes the proof.
Definition 5.4. Let −A be the generator of a bounded C0-semigroup, and
assume that A is invertible. Let β ∈ (0, 1], and let ℓ be a slowly varying
function such that g : s 7→ s1−βℓ(s) is increasing on R+. Let Sg be the
Stieltjes function associated with g. For α ≥ β, define
(5.13) Wα,β,ℓ(A) := A−(α−β)Z ∞
(s + A)−1 d(cid:0)s1−βℓ(s)(cid:1) = A−(α−β)Sg(A).
See Example 2.14 and Remark 3.8, (iii) and (v), for the definition of Sg, the
convergence of the integral and compatibility with the extended holomorphic
functional calculus.
0+
(cid:3)
stw2
stw2
38
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
For 0 < α < β, define
stw3
(5.14)
Wα,β,ℓ(A)
stw4
(5.15)
:= Z ∞
= Z ∞
0+
0
Aβ−α(s + A)−1 d(cid:0)s1−βℓ(s)(cid:1)
Aβ−α(s + A)−2s1−βℓ(s) ds.
Since A is sectorial and invertible,
C
k(s + A)−1k ≤
1 + s
The moment inequality (4.2) gives
,
kA(s + A)−1k ≤ C.
kAβ−α(s + A)−1k ≤
C
(1 + s)1+α−β .
Then convergence of the integrals in (5.14) and (5.15) follows from the dis-
cussion of Example 2.14. In particular, Wα,β,ℓ(A) is a bounded operator on
X, and A−(β−α)Wα,β,ℓ(A) = Sg(A), by the product rule (Theorem 3.7(iv)).
When ℓ(s) = 1 for all s > 0, Wα,β,ℓ(A) = A−α for all β ∈ (0, 1). The
following result is already known in that special case [50, Lemma 3.2], [17,
Lemma 2.3].
thm.bounded Theorem 5.5. Let −A be the generator of a bounded C0-semigroup (T (t))t≥0
on a Banach space X such that iR ⊂ ρ(A). Let ℓ be a slowly varying func-
tion on R+, let α > 0 and β ∈ (0, 1]. Assume that g : s 7→ s1−βℓ(s) is
increasing. The following statements are equivalent:
ℓ(s)(cid:19) ,
(i) k(is + A)−1k = O(cid:18) sα
z∈C+ k(z + A)−1Wα,β,ℓ(A)k < ∞.
s → ∞;
(ii) sup
boundedd1
boundedd3
resident3
resident3
bounded-ses
bounded-ses
1
(5.16)
(z + A)−1A−k =
Proof. For z ∈ ρ(−A) and k ∈ N, observe that
k−1Xi=0
(−z)k (z + A)−1 −
kA−(α−β)(is + A)−1k = O(cid:18) sβ
ℓ(s)(cid:19) ,
We shall show that (i) is equivalent to
(5.17)
1
(−z)i+1 A−(k−i).
s → ∞.
Write α − β = m + γ where m ∈ Z, m ≥ −1 and 0 ≤ γ < 1.
Assume that (i) holds. Then
kA(is + A)−1k = kI − is(is + A)−1k = O(cid:18)sα+1
ℓ(s)(cid:19) ,
By the moment inequality (4.2),
s → ∞.
ℓ(s) (cid:19) ,
kA1−γ(is + A)−1k = O(cid:18)sα+1−γ
s → ∞.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
39
Using (5.16) for k = m + 1, it follows that
kA−(α−β)(is + A)−1k = kA1−γ A−(m+1)(is + A)−1k
!
= O sα+1−γ−(m+1)
= O(cid:18) sβ
ℓ(s)(cid:19) ,
ℓ(s)
s → ∞.
Conversely, if α ≥ β, (5.17) implies that
ℓ(s)(cid:19) ,
kA−(α−β−1)(is + A)−1k = O(cid:18)sβ+1
s → ∞.
Since m = (1 − γ)(α − β) + γ(α − β − 1) and (1 − γ)β + γ(β + 1) = α − m,
the moment inequality (4.2) gives
kA−m(is + A)−1k = O(cid:18)sα−m
ℓ(s) (cid:19) ,
s → ∞.
Then (5.16) for k = m gives (i).
If α < β, then m = −1 and (5.17) implies that
kA−(α−β+1)(is + A)−1k = (cid:13)(cid:13)A−(α−β+1) − A−(α−β)(is + A)−1(cid:13)(cid:13)
= O(cid:18)sβ−1
ℓ(s)(cid:19) ,
s
s → ∞.
Since 0 = γ(α − β) + (1 − γ)(α − β + 1) and γβ + (1 − γ)(β − 1) = α,
the moment inequality (4.2) gives (i). This completes the proof that (i) is
equivalent to (5.17).
Next, we consider the Stieltjes function Sg associated with g, defined as
in Example 2.14. By Karamata's Theorem 2.15(a) with σ = β,
stasymp
stasymp
(5.18)
By Proposition 2.4,
Sg(λ) = O(cid:16)λ−βℓ(λ)(cid:17) ,
λ → ∞.
equiv
equiv
(5.19)
Sg(−is) = O(cid:16)s−βℓ(s)(cid:17) ,
s → ∞.
40
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Now we observe that
(is + A)−1Wα,β,ℓ(A) = Z ∞
0+
= A−(α−β)(is + A)−1Z ∞
0+
d(cid:0)λ1−βℓ(λ)(cid:1)
(is + A)−1A−(α−β)(λ + A)−1 d(cid:0)λ1−βℓ(λ)(cid:1)
−Z ∞
−Z ∞
A−(α−β)(λ + A)−1 d(cid:0)λ1−βℓ(λ)(cid:1)
A−(α−β)(λ + A)−1 d(cid:0)λ1−βℓ(λ)(cid:1).
λ − is
= Sg(−is)A−(α−β)(is + A)−1
λ − is
λ − is
1
1
0+
0+
formula
(5.20)
The last integral is bounded, uniformly for s ≥ 1, by the same argument
as for (5.14) together with the fact that λ − is ≥ 1.
Now assume that (i) holds, so (5.17) holds. By (5.17) and (5.19),
ss.slower
reslest
reslest
kSg(−is)A−(α−β)(is + A)−1k ≤ C,
s ≥ 1.
Hence k(is + A)−1Wα,β,ℓ(A)k is bounded for s ≥ 1, and then for all real s.
We now apply Lemma 5.3, with
f (z) := (z + A)−1Wα,β,ℓ(A),
and we deduce that (ii) holds.
Conversely, assume that (ii) holds. Letting z → is and using (5.20) shows
that kSg(−is)A−(α−β)(is + A)−1k is bounded for s ∈ R, s > 1. Proposition
2.4 and Theorem 2.15(a) show that (5.17) holds. Then (i) follows.
(cid:3)
5.3. Resolvent growth slower than sα. In this and the next subsection
we consider cases when −A generates a bounded C0-semigroup (T (t))t≥0 on
a Hilbert space X, σ(A) ∩ iR is empty and
(5.21)
k(is + A)−1k = O(cid:18) sα
ℓ(s)(cid:19) ,
s → ∞,
where ℓ is slowly varying and monotonic. The upper bound for kT (t)A−1k
in (5.4) is then valid for M (s) = Csα/ℓ(s). Then
Mlog(s) ∼ C(1 + α)
sα log s
ℓ(s)
= C(1 + α)(M.log)(s),
so we may replace Mlog by M.log in (5.4).
If we put k(s) = 1/ℓ(s1/α), then Proposition 2.11 gives
(M.log)−1(t) ∼ t1/α(k. log)#(t)1/α.
Thus we obtain that
klogest
klogest
(5.22)
kT (t)A−1k = O(cid:16)(cid:0)t(k. log)#(t)(cid:1)−1/α(cid:17) ,
t → ∞.
kest
kest
(5.24)
lest
lest
(5.25)
or, assuming that ℓ is dB-symmetric,
kT (t)A−1k = O(cid:16)(cid:0)tk#(t)(cid:1)−1/α(cid:17) ,
kT (t)A−1k = O(cid:16)(cid:0)tℓ(t1/α)(cid:1)−1/α(cid:17) ,
t → ∞,
t → ∞.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
41
llogest
llogest
(5.23)
When k. log is dB-symmetric, this becomes
kT (t)A−1k = O (log t)1/α
tℓ(t1/α)1/α! ,
t → ∞.
On the other hand, if we have
k(is + A)−1k ≥ c sα
ℓ(s)
for large s, then (5.6) and Proposition 2.11(b) give
.
c
kT (t)A−1k ≥
tk#(t1/α)1/α
Thus, assuming (5.21), the optimal result would be to establish that
We now give one of our main results showing that (5.25) holds when ℓ is
increasing, i.e., k(is + A)−1k grows slightly slower than sα. The case when
ℓ is decreasing will be considered in Subsection 5.4.
thm.main Theorem 5.6. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X, with generator −A. Assume that σ(A) ∩ iR is empty, and that
s → ∞,
where α > 0 and ℓ is increasing and slowly varying. Then
ℓ(s)(cid:19) ,
k(is + A)−1k = O(cid:18) sα
kT (t)A−1k = O(cid:16)(cid:0)tℓ(t1/α)(cid:1)−1/α(cid:17) ,
t → ∞.
Proof. We can assume that T (t) 6= 0 for each t > 0. First we note a known
upper bound for kT (t)A−1k. Since k(is + A)−1k = O(sα), [17, Theorem
2.4] gives
mainest
mainest
(5.26)
BTest
BTest
(5.27)
By Theorem 5.5,
kT (t)A−1k ≤
C
t1/α
.
k(λ + A)−1Wα,1,ℓ(A)k ≤ C,
λ ∈ C+.
By Theorem 4.7,
kT (t)Wα,1,ℓ(A)k ≤
C
t
,
t > 0.
Let Sℓ be the Stieltjes function associated with ℓ (Example 2.14) and let fℓ
be the complete Bernstein function defined by
fℓ(s) = Sℓ(1/s),
s > 0.
42
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Using (5.13) and Remark 3.8(v), we have
kT (t)A−(α−1)fℓ(A−1)k ≤
C
t
,
t > 0.
By Theorem 4.3(a), with γ = 1 − α,
fℓ(kT (t)A−1k) ≤
Letting fα,ℓ(s) = sα−1fℓ(s), we have
CkT (t)A−1k
tkT (2t)A−αk
.
fatest2
fatest2
(5.28)
fα,ℓ(kT (t)A−1k) ≤
CkT (t)A−1kα
tkT (2t)A−αk
.
By Theorem 2.15(a) with g = ℓ and σ = 1, Sℓ(s) ∼ s−1ℓ(s) as s → ∞.
Hence fα,ℓ(s) ∼ sαℓ(1/s) as s → 0+. Let k(s) = 1/ℓ(s1/α). Then
fα,ℓ(s) ∼
By Proposition 2.11(c),
sα
k(s−α)
,
s → 0+,
f−1
α,ℓ (s) ∼(cid:18)
s
k#(1/s)(cid:19)1/α
,
s → 0 + .
If the right-hand side of (5.28) is sufficiently small, it follows that
Lest
Lest
(5.29)
where
bigL
bigL
(5.30)
kT (t)A−1k ≤
CkT (t)A−1k
(tL(t)kT (2t)A−αk)1/α
,
kT (t)A−1kα(cid:19) .
L(t) = k#(cid:18)tkT (2t)A−αk
Let ψ(s) = (sk#(s))1/α. Since ψ is regularly varying with positive index,
we can choose k# so that ψ is strictly increasing and continuous (see the
remarks following Definition 2.9; in fact, we can choose k# to be increasing,
since k is decreasing). Then
psiest
(5.31)
1
kT (t)A−1k ≥ ct1/α kT (2t)A−αk1/α
kT (t)A−1k
kT (t)A−1kα(cid:19) .
= c ψ(cid:18)tkT (2t)A−αk
k#(cid:18)tkT (2t)A−αk
kT (t)A−1kα(cid:19)1/α
If the right-hand side of (5.28) is bounded away from zero, then (5.31) also
holds for some c > 0, since kT (t)A−1k is bounded and ψ is bounded on
bounded intervals. Hence (5.31) holds for all t > 0, for some c > 0.
Since ψ is (asymptotically equivalent to) an increasing function,
tkT (2t)A−αk
kT (t)A−1kα ≤ ψ−1(cid:18)
C
kT (t)A−1k(cid:19) .
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
43
By Proposition 2.11(b),
ψ−1(s) ∼ sαk##(sα) ∼
sα
ℓ(s)
,
s → ∞,
so
impest
impest
(5.32)
kT (2t)A−αk ≤
C
tℓ(cid:0)kT (t)A−1k−1(cid:1) ,
for large t and then for all t > 0. Then (5.27) gives
kT (2t)A−αk ≤
C
tℓ(t1/α))
.
Applying Lemma 4.2 with B = A−1, γ = α and δ = 1, it follows that
kT (t)A−1k ≤ CkT (ct)A−αk1/α ≤
C
(tℓ(t1/α))1/α
.
(cid:3)
cor.main Corollary 5.7. In addition to the assumptions of Theorem 5.6, assume that
ℓ is dB-symmetric. Then
kT (t)A−1k = O(cid:16)
1
M−1(t)(cid:17),
t → ∞,
where M−1 is any asymptotic inverse of sα/ℓ(s).
Proof. By Proposition 2.11(b),
M−1(t) ∼ t1/αk#(t)1/α,
where k(t) = 1/ℓ(t1/α). By Lemma 2.12, k#(t) ∼ ℓ(t1/α) if (and only if) ℓ
is dB-symmetric.
(cid:3)
iterest Remark 5.8. Corollary 5.7 establishes the optimal estimate (5.7) when ℓ is
dB-symmetric. However the upper bound (5.26) is not as sharp as (5.7)
when ℓ is not dB-symmetric. Under the assumptions of Theorem 5.6 the
proof takes the upper bound (5.27) from [17] and improves it to the upper
bound (5.26). Given an estimate
genest4
genest4
(5.33)
kT (t)A−1k = O(cid:16)(tm(t))−1/α(cid:17) ,
t → ∞,
where m is increasing and slowly varying, under the same assumptions the
same argument shows that
kT (t)A−1k = O(cid:18)(cid:16)tℓ(t1/αm(t)1/α)(cid:17)−1/α(cid:19) ,
t → ∞.
This process can be iterated, so one obtains (5.33) for each of the following
functions m:
m(t) = 1, ℓ(t1/α), ℓ(t1/αℓ(t1/α)1/α), ......
In some cases this process stabilises (up to asymptotic equivalence) after a
finite number of iterations at the optimal estimate in Corollary 5.7. This is
44
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
analogous to B´ek´essy's method of finding the de Bruijn conjugate of many
slowly varying functions [15, Proposition 2.3.5] (see Example 5.10).
altstart Remark 5.9. An alternative to the known upper bound (5.27) is the estimate
(5.22). In the context of Theorem 5.6, Lemma 2.13(ii) shows that this gives
kT (t)A−1k = O (cid:18) log t
tk#(t)(cid:19)1/α! ,
t → ∞,
where k(s) = 1/ℓ(s1/α). This is (5.33) with
k#(t)
log t
m(t) =
.
Following Remark 5.8, we obtain
kT (t)A−1k = O tℓ (cid:18) tk#(t)
log t (cid:19)1/α!!−1/α ,
t → ∞.
In many cases, this is asymptotically equivalent to the optimal estimate
(5.24).
explog Example 5.10. We take α = 1 (for simplicity of presentation) and ℓ(s) =
exp(cid:0)(log s)β(cid:1) where 0 < β < 1 (see Example 2.10). In this case, the process
in Remark 5.8, starting from m(t) = 1, stabilises at the optimal estimate
optest
optest
(5.34)
kT (t)A−1k ≤
C
M−1(t) ∼
C
t(1/ℓ)#(t)
.
When 1/2 < β < 3/4, the process stabilises after two iterations at the
optimal estimate:
kT (t)A−1k ≤
C
t
exp(cid:16)−(cid:16)(log t)β + β(log t)2β−1(cid:17)(cid:17) .
When (n − 1)/n < β < n/(n + 1), n iterations are needed.
Starting from
m(t) =
(1/ℓ)#(t)
log t
,
as in Remark 5.9, the process stabilises at the optimal estimate (5.34) after
one iteration for every β ∈ (0, 1).
easier Remark 5.11. If (5.6) and (5.7) both hold for some regularly varying func-
tion M of index α > 0, then the ratio
ratio
ratio
(5.35)
kT (2t)A−1k
kT (t)A−1k
is bounded away from zero (it is bounded above, for any bounded semi-
group).
On the other hand, if (5.35) is bounded away from zero, then so is
kT (2t)A−αk
kT (t)A−1kα ,
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
45
by Lemma 4.2. Then the function L in the proof of Theorem 5.6 is asymp-
totically equivalent to k#, and one can then pass easily from (5.30) to (5.7).
This argument would not require the assumption that ℓ is dB-symmetric,
and it would not use Theorem 4.3 or (5.4). However we do not see any way
to prove directly that (5.35) is bounded away from zero.
The assumptions of Theorem 5.6 are not sufficient to ensure that (5.35)
is bounded away from zero. However, we may assume in addition that
resbd-
resbd-
(5.36)
k(is + A)−1k ≥ cM (s).
If one can prove that (5.7) holds under this additional assumption, then it
can be deduced that (5.7) holds without the additional assumption. This
follows by means of a direct sum argument which is used in the proof of
Theorem 5.12 below.
Under the assumptions of Theorem 5.6 and (5.36), we can obtain from
(5.4) and (5.6) that
kT (2t)A−αk
kT (t)A−1kα ≥
It then follows from (5.29) that
c
(log t)α .
kT (t)A−1k = O (cid:18)tk#(cid:18)
t
(log t)α(cid:19)(cid:19)−1/α! ,
t → ∞.
The direct sum argument can then be used to show that this estimate
holds under the assumptions of Theorem 5.6, without (5.36).
In general
this estimate is slightly worse than (5.26), but in many cases, for example
ℓ(s) = (log s)β, one can recover the estimate of Corollary 5.7.
5.4. Resolvent growth faster than sα. Now we consider the case when
(5.21) holds with ℓ decreasing so k(is + A)−1k grows slightly faster than sα.
Unfortunately, our result is not quite optimal in this case as a logarithmic
term still appears in (5.37) below. However it does improve (5.22) and
(5.23) as the logarithm has an arbitrarily small power. If ℓ is dB-symmetric,
then k#(t) = ℓ(t1/α) (see the proof of Corollary 5.7), so (5.37) is directly
comparable with (5.23) and (5.25).
ss.faster
thm.faster Theorem 5.12. Let (T (t)t≥0) be a bounded C0-semigroup on a Hilbert space
X, with generator −A. Assume that σ(A) ∩ iR is empty, and that
mainest2
mainest2
(5.37)
where k(t) = 1/ℓ(t1/α) and k# is the de Bruijn conjugate of k.
where α > 0 and ℓ is decreasing and slowly varying. Then, for any ε > 0,
ℓ(s)(cid:19) ,
k(is + A)−1k = O(cid:18) sα
(tk#(t))1/α(cid:19) ,
kT (t)A−1k = O(cid:18) (log t)ε
s → ∞,
t → ∞,
46
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Proof. Much of the proof is similar to Theorem 5.6. We can assume that
kT (t)k 6= 0 for each t > 0. Given ε ∈ (0, α−2), let β = 1 − εα2 ∈ (0, 1).
By replacing ℓ by an asymptotically equivalent, decreasing function, we may
assume, without loss, that g : s 7→ s1−βℓ(s) is increasing on R+. Let Sℓ be
the Stieltjes function associated with ℓ, and let
fℓ(s) = Sℓ(1/s),
s > 0.
Then fℓ is a complete Bernstein function, and
Wα,β,ℓ(A) = A−(α−β)fℓ(A−1).
By Theorem 5.5,
By Theorem 4.7,
sup
z∈C+ k(z + A)−1Wα,β,ℓ(A)k < ∞.
kT (t)A−(α−β)fℓ(A−1)k ≤
C
t
,
t > 0.
By Theorem 4.3(b), with γ = β − α,
fℓ(kT (t)A−1k) ≤
Letting fα,ℓ(s) = sα−βfℓ(s), we have
CkT (t)A−1k
tkT (2t)A−(α−β+1)k
.
fatest
fatest
(5.38)
fα,ℓ(kT (t)A−1k) ≤
CkT (t)A−1kα−β+1
tkT (2t)A−(α−β+1)k
.
By Theorem 2.15(a) with g(λ) = λ1−βℓ(λ) and σ = β, Sℓ(s) ∼ s−βℓ(s) as
s → ∞. Hence
fα,ℓ(s) ∼ sαℓ(1/s) =
sα
k(s−α)
,
s → 0 + .
By Proposition 2.11(c),
f−1
s
,
s → 0 + .
α,ℓ (s) ∼(cid:18)
k#(1/s)(cid:19)1/α
If the right-hand side of (5.38) is small, it follows that
CkT (t)A−1k1+(1−β)/α
kT (t)A−1k ≤
kT (t)A−1kα−β+1 ! .
L(t) = k# tkT (2t)A−(α−β+1)k
(tL(t)kT (2t)A−(α−β+1)k)1/α
where
,
Let ψ(s) = (sk#(s))1/α. Since ψ is regularly varying with positive index, we
can choose k# so that ψ is strictly increasing and continuous (see the remarks
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
47
following Definition 2.9; in fact, we can also choose k# to be decreasing, since
k is increasing). Then
psiest2
(5.39)
1
kT (t)A−1k ≥ ct1/α kT (2t)A−(α−β+1)k1/α
kT (t)A−1k1+(1−β)/α
= c ψ tkT (2t)A−(α−β+1)k
kT (t)A−1kα−β+1 ! .
L(t)1/α
If the right-hand side of (5.38) is bounded away from zero, then (5.39) also
holds for some c > 0, since kT (t)A−1k is bounded and ψ is bounded on
bounded intervals. Hence (5.39) holds for all t > 0, for some c > 0.
Since ψ is strictly increasing,
By Proposition 2.11(b),
tkT (2t)A−(α−β+1)k
kT (t)A−1kα−β+1 ≤ ψ−1(cid:18)
ψ−1(s) ∼ sαk##(sα) ∼
sα
ℓ(s)
,
C
kT (t)A−1k(cid:19) .
so
impest2
impest2
(5.40)
kT (2t)A−(α−β+1)k ≤
CkT (t)A−1k1−β
tℓ(cid:0)kT (t)A−1k−1(cid:1) ,
for large t. By Lemma 4.2 with B = A−1, γ = 1 and δ = α − β + 1,
for some K. Hence
kT (2Kt)A−1k ≤
kT (2Kt)A−1k1+(1−β)/α ≤ CkT (2t)A−(α−β+1)k1/α
kT (2Kt)A−1k(cid:19) 1−β
(tℓ(kT (t)A−1k−1))1/α(cid:18) kT (t)A−1k
cM (s) ≤ k(is + A)−1k ≤ M (s),
C
α
.
Now we shall temporarily assume that
tempass
tempass
(5.41)
where M (s) = sα/ℓ(s) for large s. We know from (5.4) and (5.6) that
c
M−1(Ct) ≤ kT (t)A−1k ≤
C
(M.log)−1(ct)
.
Since ℓ is decreasing and slowly varying and M−1 is regularly varying,
1
ℓ(kT (t)A−1k−1) ≤
1
ℓ(c−1M−1(Ct)) ∼
1
ℓ(M−1(t))
.
Hence
kT (2Kt)A−1k ≤
C
(tℓ(M−1(t)))1/α(cid:18) M−1(Ct)
(M.log)−1(ct)(cid:19) 1−β
α
for large t. Since M is regularly varying of index α, M−1 and (M.log)−1
are both regularly varying of index 1/α. Since ℓ is decreasing, we can apply
48
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Lemma 2.13(ii), with ℓ replaced by 1/ℓ and δ = α, to obtain that for a, b > 0
there exists ca,b > 0 such that
(M.log)−1(at)
M−1(bt)
ca,b
(log t)1/α
≥
for large t. Moreover,
M−1(t) ∼ t1/αk#(t)1/α,
and k(tk#(t)) ∼ (k#(t))−1, by definition of k#, so
ℓ(M−1(t)) ∼ ℓ(t1/αk#(t)1/α) ∼ k#(t).
Hence
kT (2Kt)A−1k ≤
C(log t)ε
(tk#(t))1/α
for large t. Replacing t by t/(2K) and changing the value of C we obtain
kT (t)A−1k ≤
C(log t)ε
(tk#(t))1/α
for large t.
Now we no longer assume (5.41), and we obtain the same result by means
of an artificial device. Let (S(t))t≥0 be a bounded C0-semigroup on a Hilbert
space Y , with generator −B which does satisfy (5.41). For example, we can
take a normal semigroup on a Hilbert space with
We then consider the C0-semigroup
σ(B) =(cid:8)2M (s)−1 + is : s ≥ 2(cid:9) .
T (t) ⊕ S(t)
on X ⊕ Y . This satisfies the assumption (5.41), so we find that
C(log t)ε
(tk#(t))1/α .
kT (t)A−1k ≤(cid:13)(cid:13)T (t)A−1 ⊕ S(t)B−1(cid:13)(cid:13) ≤
(cid:3)
Remark 5.13. A variation of the proof above proceeds from (5.32) by the
method of Remark 5.9. This uses the estimate (5.22) in the form
kT (t)A−1k−1 ≥ ct1/α(k. log)#(t)1/α.
Since the function s 7→ s1−βℓ(s) is asymptotically equivalent to an increasing
function,
kT (t)A−1k1−β ≥ ct(1−β)/α(k. log)#(t)(1−β)/αℓ(cid:16)t1/α(k. log)#(t)1/α(cid:17) .
ℓ(kT (t)A−1k−1)
By (5.40),
kT (2t)A−(α−β+1)k ≤
t(α−β+1)/α(k. log)#(t)(1−β)/αℓ(cid:0)t1/α(k. log)#(t)1/α(cid:1) .
C
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
49
Applying Lemma 4.2 with B = A−1, γ = α − β + 1 and taking β arbitrarily
close to 1, it follows that
kT (t)A−1k ≤
Cε
ht(k. log)#(t)εℓ(cid:0)t1/α(k. log)#(t)1/α(cid:1)1+εi1/α
.
We do not expect that this gives better estimates than the simpler (5.37).
ss.background
s.zero1
6. Singularity at zero: general results
6.1. Preliminary background. In this section we treat the rates of decay
of the derivatives −T (t)Ax for x ∈ dom(A), where (T (t))t≥0 is a bounded
C0-semigroup on a Banach space or a Hilbert space X, with generator −A.
In other words, we study the orbits of (T (t))t≥0 starting at points y = Ax
in the range of A. We shall see that such decay corresponds to properties of
the resolvent of A permitting a singularity at 0 but requiring boundedness
at infinity on the imaginary axis.
If the decay of such orbits is uniform for kxk ≤ 1, the semigroup must
be eventually differentiable, i.e., for sufficiently large t, T (t) maps X into
dom(A) and AT (t) ∈ L(X). For an eventually differentiable, bounded C0-
semigroup on a Banach space, Arendt and Pruss [7, Theorem 3.10] (see [6,
Theorem 4.4.16]) showed that
uniform1
uniform1
(6.1)
if and only if
specsub0
specsub0
(6.2)
lim
t→∞kAT (t)k = 0
σ(A) ∩ iR ⊂ {0}.
In fact the Arendt-Pruss theorem follows from (the corollary of) the
Katznelson-Tzafriri theorem for C0-semigroups which we recall below (see
also [67], [28], [10] and the survey [20]). Let E be a closed subset of R. A
function f ∈ L1(R) is said to be of spectral synthesis with respect to E if
there is a sequence (fn) in L1(R) such that limn→∞ kf − fnkL1 = 0 and, for
each n, Ffn vanishes in a neighbourhood of E.
The closed subset E ⊂ R is said to be of spectral synthesis if every function
f ∈ L1(R) whose Fourier transform Ff vanishes on E is of spectral synthesis
with respect to E. Any countable closed subset of R is of spectral synthesis
[46, p. 230].
KT Theorem 6.1. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
X, with generator −A. Let f ∈ L1(R+) be of spectral synthesis with respect
to E := iσ(−A) ∩ R. Then
(6.3)
KT1KT1
lim
t→∞kT (t) f (T )k = 0,
where
f (T )x =Z ∞
0
f (t)T (t)x dt,
x ∈ X.
Conversely, if f ∈ L1(R+) and (6.3) holds, then Ff = 0 on E.
50
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
convtozero Corollary 6.2. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
KTcori
KTcorii
X with generator −A. The following statements are equivalent.
(i) limt→∞ kT (t)A(I + A)−2k = 0;
(ii) σ(A) ∩ iR ⊂ {0}.
Proof. This follows from applying the Katznelson-Tzafriri theorem, with
Then
f (t) = e−t − te−t,
t ≥ 0.
f (T ) = A(I + A)−2.
Ff (s) = is(1 + is)−2,
Since Ff (0) = 0 and {0} is a set of spectral synthesis, f is of spectral
synthesis with respect to {0}.
(cid:3)
Remark 6.3. If (T (t))t≥0 is eventually differentiable, then (1 + A)2T (τ ) is a
bounded operator for some τ > 0. Moreover,
s ∈ R;
AT (t + τ ) = (I + A)2T (τ )T (t)A(I + A)−2,
so Corollary 6.2(i) is equivalent to (6.1) in this case. Thus the Arendt-Pruss
theorem follows from the Katznelson-Tzafriri theorem.
The range of A(I + A)−2 is ran(A) ∩ dom(A) (Proposition 3.10(i)), and
Corollary 6.2 describes decay of orbits on that space in an appropriate uni-
form sense. We defer consideration of the rate of convergence in Corollary
6.2 (i) until Section 8. In this and the next section, we consider instead the
question whether the decay of orbits is uniform with respect to the graph
norm on dom(A), i.e., whether
lim
t→∞
sup(cid:8)kT (t)Axk : x ∈ dom(A),kxkdom(A) = 1(cid:9) = 0.
This can be reformulated as
uniform0
uniform0
(6.4)
lim
t→∞kT (t)A(ω + A)−1k = 0,
for any ω ∈ ρ(−A). This property is intermediate between (6.1) and the
statements of Corollary 6.2, and it is independent of the choice of ω. When
A is invertible, (6.4) is equivalent to the much studied notion of exponential
stability, i.e., limt→∞ kT (t)k = 0, for which the rate of decay is always at
least exponential (see [6, Chapter 5]). Thus we are interested only in cases
when 0 ∈ σ(A).
Our goal in this section and Section 7 is to develop a framework for decay
of the form (6.4) in terms of the spectrum and resolvent of the generator in
similar form to the one from Section 5. However, the decay given by (6.4) is
much less studied in the literature. Thus we shall prove several statements
clarifying the limitations imposed by (6.4) and the consequences following
from the decay of orbits in this sense.
In the later subsections, we shall assume that the semigroup is bounded,
but first it is instructive to make a few remarks relating to that assumption.
We begin by observing that (6.4) implies (6.2), without the assumption of
boundedness.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
51
spec0 Proposition 6.4. Let (T (t))t≥0 be a C0-semigroup on a Banach space X,
with generator −A. Assume that
lim
t→∞kT (t)A(ω + A)−1k = 0
for some ω ∈ ρ(−A). Then, for each η > 0,
In particular,
inf{Re λ : λ ∈ σ(A),λ > η} > 0.
σ(A) ∩ {λ ∈ C : Re λ ≤ 0} ⊂ {0}.
Proof. We may assume that ω > 0 and kT (t)k ≤ M e(ω−1)t for all t > 0.
Define ft ∈ L1(R+) ⊂ M b(R+) and µt ∈ M b(R+) by
s ≥ t,
0 ≤ s < t,
ft(s) = (ωe−ω(s−t),
0
µt = δt − ft.
Here δt is the unit mass at t and ft ∈ L1(R+) is regarded as an absolutely
continuous measure. Then
T (t)A(ω + A)−1 =Z ∞
0
T (s) dµt(s).
By the spectral mapping (inclusion) theorem for the Hille-Phillips functional
calculus [39, Theorem 16.3.5],
specinc
specinc
(6.5)
and then
(cid:8)e−λtλ(ω + λ)−1 : λ ∈ σ(A)(cid:9) ⊂ σ(cid:0)T (t)A(ω + A)−1(cid:1) ,
kT (t)A(ω + A)−1k ≥ sup(cid:8)e−Re λtλω + λ−1 : λ ∈ σ(A)(cid:9).
If λ ∈ σ(A) and λ ≥ η > 0, then λω + λ−1 ≥ η(ω + η)−1. Taking t such
that kT (t)A(ω + A)−1k ≤ η/(e(ω + η)), it follows that e− Re λt ≤ 1/e and
Re λ ≥ 1/t for all such λ.
(cid:3)
We shall see in Theorem 6.10 that k(is + A)−1k is bounded for s ≥ η > 0
if the assumptions of Proposition 6.4 are satisfied and (T (t))t≥0 is bounded.
The following simple example shows that (6.4) does not imply that the
semigroup is bounded.
Example 6.5. Let X = L2([0, 1]). Define a C0-semigroup (T (t))t≥0 on X
by
(T (t)f )(s) = e−(s3/2+is)tf (s),
t ≥ 0, f ∈ L2([0, 1]), s ∈ [0, 1].
The semigroup (T (t))t≥0 has a bounded generator −A given by (−Af )(s) =
−(s3/2 + is)f (s) for s ∈ [0, 1], and it is easy to check that
kAT (t)k = O(t−2/3),
t → ∞.
52
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Consider then the semigroup (T (t))t≥0, with generator −A, on the space
X ⊕ X given by the operator matrix
Then a simple calculation shows that there exist c, C > 0 such that
0
T (t) =(cid:18)T (t) −tAT (t)
T (t) (cid:19) ,
kAT (t)k ≤ Ct−1/3,
t ≥ 0.
and
kT (t)k ≥ ct1/3
t > 0.
For semigroups of normal operators on Hilbert space, (6.4) does imply
boundedness. We omit the easy proof of the following proposition.
normal0 Proposition 6.6. Let (T (t))t≥0 be a C0-semigroup of normal operators on
sing1
sing2
sing0
a Hilbert space X, with generator −A. The following are equivalent:
(i) For some/all ω ∈ ρ(−A), limt→∞ kT (t)A(ω + A)−1k = 0.
(ii) σ(A) ⊂ C+ ∪ {0} and inf{Re λ : λ ∈ σ(A),λ > η} > 0 for some/all
(iii) (T (t))t≥0 is bounded, σ(A) ∩ iR ⊂ {0} and sups≥η k(is + A)−1k < ∞
η > 0.
for some/all η > 0.
Proposition 6.6 shows that there are bounded semigroups of normal op-
erators on Hilbert spaces, with σ(A) ∩ iR ⊂ {0}, for which property (i)
of Proposition 6.6 does not hold. The next example shows that there are
bounded semigroups on Hilbert space which satisfy property (ii), but not
property (i).
counterex0 Example 6.7. Let X0 and X1 be Hilbert spaces and consider the Hilbert
space X = X0 ⊕ X1. Let (T (t))t≥0 be a bounded C0-semigroup on X given
by T (t) = T0(t) ⊕ I, where (T0(t))t≥0 is a bounded, but not exponentially
stable, C0-semigroup on X0 with generator −A0 such that σ(A0) ⊂ C+. If
−A is the generator of (T (t))t≥0 then σ(A)∩ iR = {0}. Moreover, (6.4) does
not hold; if it did, since A0 is invertible it would follow that kT0(t)k → 0, a
contradiction.
We may take (T0(t))t≥0 to be a bounded semigroup on a Hilbert space,
which is not exponentially stable, with inf{Re λ : λ ∈ σ(A)} > 0 (see [6,
Example 5.1.10]). Then (T (t))t≥0 satisfies (ii) of Proposition 6.6, but (i)
does not hold.
ss.decay0
6.2. Rates of decay. For the rest of Section 6 and throughout Section 7,
we shall consider bounded semigroups. For simplicity of notation, we shall
set ω = 1 in (6.4).
Our main objective is to deduce (6.4) as a consequence of suitable spectral
assumptions. We shall need to assume strong conditions which are consis-
tent with Proposition 6.6 and which also exclude examples such as Example
6.7. In fact, we shall show that the property (iii) of Proposition 6.6 implies
(6.4) for all semigroups on Hilbert space, and not only for normal semi-
groups. This will be deduced from Theorem 6.14 which is a version of the
Katznelson-Tzafriri Theorem for semigroups on Hilbert spaces for certain
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
53
measures which are not absolutely continuous. This extension is of indepen-
dent interest, and it will be the subject of Subsection 6.3. In this subsection
we shall consider other aspects of the rates of decay in (6.4).
Remark 6.8. For bounded semigroups on Banach spaces the property (6.2)
implies that limt→∞ T (t)A(I + A)−1 = 0 in the strong operator topology.
This was explicitly shown in [10, Example, p.802] using a Tauberian theorem
for vector-valued Laplace-Stieltjes transforms. It follows from Corollary 6.2
using the uniform boundedness of T (t)A(I + A)−1 and the density of the
range of (I + A)−1.
We consider a bounded C0-semigroup (T (t))t≥0, with generator −A, on
a Banach space or Hilbert space X. We assume that σ(A) ∩ iR = {0} and
that, for some/all η > 0, sups≥η k(is + A)−1k < ∞. We aim to exhibit
the possible rates of decay of kT (t)A(I + A)−1k in terms of the growth of
k(is + A)−1k as s → 0. For this purpose, let m and N be decreasing
functions on (0,∞) such that
(6.6)
mbound
Nbound
k(is + A)−1k ≤ m(s),
kT (t)A(I + A)−1k ≤ N (t),
The smallest possible functions are given by
(6.7)
s 6= 0,
t > 0.
defm
defN
(6.8)
(6.9)
m(s) = sup{k(ir + A)−1k : r ≥ s},
N (t) = sup{kT (τ )A(I + A)−1k : τ ≥ t},
s > 0,
t > 0.
This function m is continuous, and we shall always assume continuity of m
so that m has a right inverse m−1 defined on an interval of the form [a,∞).
The function N defined by (6.9) may not be continuous, but it is lower
semicontinuous and right-continuous. Assuming that limt→∞ N (t) = 0, we
define
defN-1
defN-1
s > 0,
N∗(s) = min{t ≥ 0 : N (t) ≤ s},
(6.10)
so that N (N∗(s)) ≤ s for all s > 0, and N (N∗(s)) = s if s is in the range of
N .
Since we assume that 0 ∈ σ(A), elementary theory of resolvents implies
that m(s) ≥ s−1 for all s > 0. The hypothesis that there might be a
corresponding lower bound for N is too naive, because examples of the type
considered in Example 6.7 show that N can decay arbitrarily slowly even
for semigroups of normal operators on Hilbert space. On the other hand the
next result shows that examples where N decays faster than t−1 must be of
the form considered in Example 6.7.
split Theorem 6.9. Let (T (t))t≥0 be a C0-semigroup on a Banach space X with
generator −A. If 0 ∈ σ(A), then at least one of the following two properties
holds:
(i) lim supt→∞ tkT (t)A(I + A)−1k > 0;
(ii) There are closed T -invariant subspaces X0, X1 of X such that
spliti
splitii
54
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
(a) X = X0 ⊕ X1,
(b) T (t)x = x for all t ≥ 0, x ∈ X1, and
(c) the generator −A0 of the restriction of T to X0 is invertible.
Proof. Assume first that 0 is a limit point of σ(A). Then there exists {λn :
n ≥ 1} ⊂ σ(A) \ {0} such that λn → 0 as n → ∞. Setting tn = (Re λn)−1,
and using (6.5), we infer that
tkT (t)A(I + A)−1k ≥ lim
n→∞(cid:18) tne− Re λntn Re λn
1 + λn
(cid:19) =
1
e
.
lim sup
t→∞
So (i) holds.
Now assume that 0 is not a limit point of σ(A). Then X can be decom-
posed into the direct sum of T -invariant subspaces X = X0 ⊕ X1 such
that A1 := A↾X1∈ L(X1), σ(A1) = {0}, and for A0 := A↾X0 one has
σ(A0) = σ(A) \ {0}. If (i) is false, then
lim inf
t→∞
tkA1e−tA1k ≤ lim sup
t→∞
≤ lim sup
t→∞
tkA1e−tA1(I + A1)−1kkI + A1k
tkT (t)A(I + A)−1kkI + A1k = 0.
By [44, Theorem 2.1], A1 = 0. Thus (ii) holds.
(cid:3)
Thus the rates are of interest only in the case when N (t) decreases no
faster than t−1 as t → ∞, and m(s) increases at least as fast as s−1 as
s → 0+. In that case the bound in (6.12) below gives m(s) = O(N∗(cs)) for
small s > 0.
resbound0 Theorem 6.10. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach
resbd0ass
resbd0ass
space X, with generator −A. Assume that
(6.11)
lim
t→∞kT (t)A(I + A)−1k = 0.
Let N be a decreasing function such that (6.7) holds and limt→∞ N (t) = 0,
and let N∗ denote any function such that N (N∗(s)) ≤ s for all s ∈ (0, 1).
Then σ(A) ∩ iR ⊂ {0} and, for any c ∈ (0, 1),
k(is + A)−1k =(O(cid:0)N∗(cs) + s−1(cid:1) ,
O(1),
s → 0,
s → ∞.
resest0
resest0
(6.12)
Proof. Proposition 6.4 shows that σ(A)∩ iR ⊂ {0}. Alternatively, the argu-
ments which follow show that A has no approximate eigenvalues in iR\{0}.
Since −A generates a bounded semigroup, σ(A) ⊂ C+ and σ(A)∩iR consists
only of approximate eigenvalues.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
55
Let K = supt≥0 kT (t)k. Let s ∈ R \ {0}, t > 0 and x ∈ dom(A). We use
the formula
iseistx = iseistZ t
= iseistZ t
0
0
e−isτ T (τ )(is + A)x dτ + isT (t)x
e−isτ T (τ )(is + A)x dτ + T (t)(I + A)−1(is + A)x
− (1 − is)T (t)A(I + A)−1x.
Since kT (t)(I + A)−1k ≤ K, this gives
skxk ≤ K(st + 1)k(is + A)xk + 1 − isN (t)kxk.
Hence
Nest
Nest
(6.13)
(s − 1 − isN (t))kxk ≤ K(st + 1)k(is + A)xk.
Set t = N∗(cs). For s sufficiently small,
s − 1 − isN (t) ≥ s(1 − 1 − isc) > 0.
For any K′ > K, (6.13) gives
K (sN∗(cs) + 1)
s(1 − 1 − isc) ≤
K′
1 − c(cid:0)N∗(cs) + s−1(cid:1) ,
Nest2
Nest2
(6.14)
k(is + A)−1k ≤
for s sufficiently small.
ciently large,
Since limt→∞ N (t) = 0, we may set t = τ with N (τ ) < 1. For s suffi-
so (6.13) gives
s − 1 − isN (τ ) > 1,
k(is + A)−1k ≤
K(sτ + 1)
s − 1 − isN (τ )
= O(1),
s → ∞.
(cid:3)
Theorem 6.10 is analogous to [13, Proposition 1.3]. It allows one to get
lower bounds for the decay of kT (t)A(I + A)−1k in terms of the growth of
the resolvent near the origin, analogous to (5.6).
addasump
addasump
lowbound0 Corollary 6.11. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach
space X, with generator −A. Assume that σ(A) ∩ iR = {0} and that
(6.15)
max(cid:0)ks(is + A)−1k,ks(−is + A)−1k(cid:1) = ∞.
Let m be the function defined by (6.8) and m−1 be any right inverse for m.
Then there exist c > 0 and c′ > 0 such that
lim
s→0
kT (t)A(I + A)−1k ≥ cm−1(c′t)
for all sufficiently large t.
56
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Proof. We may assume that (6.11) holds. Let N and N∗ be defined by (6.9)
and (6.10) respectively. By Theorem 6.10,
m(s) ≤ C(cid:18)N∗(cs) +
1
s(cid:19) ,
0 < s ≤ 1.
Rearranging this and using the assumption (6.15),
N∗(cs) >
m(s)
2C
for all sufficiently small s > 0. For t sufficiently large, put s = c−1N (t).
Then N∗(cs) ≤ t. Hence
m(m−1(2Ct)) = 2Ct ≥ 2CN∗(cs) > m(s).
Since m is decreasing,
m−1(2Ct) < s = c−1N (t) ≤ c−1KkT (t)A(I + A)−1k,
for all sufficiently large t.
(cid:3)
remlb0 Remark 6.12. The assumption (6.15) in Corollary 6.11 cannot be omitted,
because of the trivial semigroup where A = 0, for example. For a C0-
semigroup of contractions, it can be weakened to the assumption that
lim inf
s→0
max(cid:0)ks(is + A)−1k,ks(−is + A)−1k(cid:1) > 1.
This follows from (6.14).
The ideal counterpart to Corollary 6.11 would be to show that if (6.6)
holds, then
genest0
genest0
(6.16)
kT (t)A(I + A)−1k = O(cid:0)m−1(ct)(cid:1) ,
t → ∞,
for some c > 0. However for some m it is not possible to get this sharper
estimate, even for semigroups of normal operators on Hilbert space, as the
following proposition shows. The proposition gives the necessary (when m
is defined by (6.8)) and sufficient condition on m for (6.16) to be true for
semigroups of normal operators on Hilbert spaces. It is presented in the more
general context of quasi-multiplication semigroups introduced in Section 5
and it should be compared to Proposition 5.1. The conditions (5.10) for M
and (6.18) for m are clearly dual to each other: M satisfies (5.10) if and
only if m(s) := M (1/s) satisfies (6.18). Later we shall consider more general
semigroups on Hilbert space, and we shall establish (6.16) when m(s) = s−α
for some α ≥ 1, and weaker estimates for more general m, see Theorems
7.6, 7.7 and 6.15.
normalinf0 Proposition 6.13. Let (T (t))t≥0 be a quasi-multiplication semigroup on a
Banach space X with generator −A. Assume that
0 ∈ σ(A) ⊂ C+ ∪ {0}
and
k(is + A)−1k = O(1),
s → ∞.
Let c > 0, m be defined by (6.8), and m−1 be any right inverse for m. Then
the following are equivalent:
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
57
normalcond01
(i) There exists C such that
normalest01
normalest01
(6.17)
normalcond02
(ii) There exists B such that
kT (t)A(I + A)−1k ≤ Cm−1(ct),
τ(cid:17) − B,
m(s) ≥ c log(cid:16) s
m(τ )
t ≥ c−1m(1);
0 < τ, s ≤ 1.
normalres0
normalres0
(6.18)
Proof. Note first that the assumptions imply that kT (t)k = 1 for all t ≥ 0,
m(s)−1 = min{µ + ir : µ ∈ σ(A),r ≥ s} ≤ s,
s > 0,
and there exists ε > 0 such that Re µ ≥ ε for all µ ∈ σ(A) with µ ≥ 1.
Moreover (6.17) is equivalent to
µ = α + iβ ∈ σ(A) \ {0}, t ≥ c−1m(1),
and hence to
e−tα(cid:12)(cid:12)(cid:12)(cid:12)
µ
1 + µ(cid:12)(cid:12)(cid:12)(cid:12) ≤ Cm−1(ct),
tα ≥ log(cid:18)
for all such µ and t.
normalest02
normalest02
(6.19)
Assume that (i) holds. Let t ≥ c−1m(1), and put τ = m−1(ct). From
(6.19),
1
Cm−1(ct)(cid:12)(cid:12)(cid:12)(cid:12)
Cτ1 + µ(cid:19) ≥
µ
µ
1 + µ(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)
log(cid:18) µ
2Cτ(cid:19) ,
c
α
m(τ ) ≥
c
α
log(cid:18)
if µ = α + iβ ∈ σ(A) and µ ≤ 1.
Given s > 0 with s < ε, take µ = α+iβ ∈ σ(A) such that m(s)−1 = µ+ir
for some r ≥ s. Then α ≤ µ + ir ≤ s < ε, so µ ≤ 1. If µ ≥ s/2, then
m(τ )
m(s) ≥
cα + i(β + r)
α
If µ < s/2, then m(s)−1 ≥ r − µ ≥ s/2. Since m(τ ) ≥ 1/τ ,
τ(cid:17) − c log(4C).
log(cid:16) s
2τ ≥ c log(cid:16) s
4Cτ(cid:17) ≥ c log(cid:16) s
τ(cid:17) − B,
s
m(τ )
m(s) ≥
for some B ≥ c log(4C).
It follows that (6.18) holds whenever 0 ≤ s < ε and τ is in the range
of m−1. For other values of τ one can apply the above with τ replaced by
m−1(m(τ ) − n−1), and let n → ∞. If ε ≤ s ≤ 1, one can use
m(τ )
m(s) ≥
m(τ )
m(ε) ≥ c log(cid:16) ε
τ(cid:17) − B ≥ c log(cid:16) s
τ(cid:17) − B + c log ε.
Now assume that (ii) holds. Given t ≥ c−1m(1) and µ = α + iβ ∈ σ(A)
with β ≤ 1, take
By (6.18),
τ = m−1(ct),
s = β.
ct
m(β) ≥ c log(cid:18) β
m−1(ct)(cid:19) − B.
58
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Rearranging this, using αm(β) ≥ 1, and putting C = 2 exp(B/c) gives
(6.19), provided that µ/2 ≤ β ≤ 1. If β < µ/2, then α > (√3/2)µ,
and
tα =
α
c
m(m−1(ct)) >
2cm−1(ct) ≥ log √3µ
√3µ
2cm−1(ct)! .
So we establish (6.19) in this case also with C = 2c/√3. Taking the maxi-
mum of two values of C we have established (6.19) whenever β ≤ 1.
Now consider µ = α + iβ ∈ σ(A) with µ ≥ 1, so α ≥ ε. Proceeding in
a similar way to the last part of the proof of Proposition 5.1 (or applying
the same arguments to the function M (s) := m(1/s)), one sees that there
exists C such that
εt ≥ log(cid:18)
Cm−1(ct)(cid:12)(cid:12)(cid:12)(cid:12)
Cm−1(ct)(cid:19) ≥ log(cid:18)
whenever t ≥ c−1m(1). Then (6.19) holds for µ = α + iβ ∈ σ(A) with
β ≥ 1 and t ≥ c−1m(1). Hence (i) holds.
6.3. An extension of the Katznelson-Tzafriri Theorem. The follow-
ing result is an extension of the Katznelson-Tzafriri Theorem 6.1 in the case
of semigroups on Hilbert spaces to some measures which are not absolutely
continuous.
1 + µ(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)
(cid:3)
1
µ
1
s.KT+
KT+ Theorem 6.14. Let −A be the generator of a bounded C0-semigroup (T (t))t≥0
on a Hilbert space X. Assume that E := iσ(−A)∩ R is compact and of spec-
tral synthesis. Assume in addition that, for some η ≥ 0,
(6.20)
nosinginf
nosinginf
sup
s≥η k(is + A)−1k < +∞.
Let µ ∈ M b(R+) be such that Fµ vanishes on E. Then
where
lim
t→∞kT (t)µ(T )k = 0,
µ(T )x =Z ∞
0
T (s)x dµ(s),
x ∈ X.
Proof. Let ϕ ∈ S(R) be such that Fϕ has compact support and such that
Fϕ = 1 on a neighbourhood of E. We decompose the measure µ as follows
µ = µ ∗ ϕ + µ ∗ (δ0 − ϕ) =: µ0 + µ1,
and we note that the measure µ0 is absolutely continuous with respect to
Lebesgue measure: let f ∈ L1(R) be the density function for µ0.
Consider the function
F (t) :=ZR
T (t + s) dµ(s),
t ∈ R,
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
59
where we have extended the semigroup T by 0 on (−∞, 0) and the integral
is convergent in the strong operator topology. Since µ is supported in R+,
we have
Following the decomposition of µ, we define also the functions
F (t) = T (t)µ(T ),
t ≥ 0.
and
F0(t) :=ZR
F1(t) :=ZR
T (t + s)f (s) ds,
T (t + s) dµ1(s),
t ∈ R,
t ∈ R,
so that F = F0 + F1. In these two definitions, we have also extended the
semigroup T by 0 on (−∞, 0).
Note that
Fµ0(ξ) = Ff (ξ) = Fµ(ξ)Fϕ(ξ),
ξ ∈ R.
Since Fµ vanishes on E, the Fourier transform Ff also vanishes on E.
Since E is of spectral synthesis, there exists a sequence (fn)n≥2 ⊂ L1(R)
such that each Fourier transform Ffn vanishes on a neighbourhood of E
and such that limn→∞ kfn − fkL1 = 0. Furthermore, we may assume that
each Fourier transform Ffn has compact support. Set
We have, by Parseval's formula,
T (t + s)fn(s) ds,
n ≥ 2, t ∈ R.
Fn(t) :=ZR
α→0+Z ∞
2πZR
2πZR
(iξ + A)−1Ffn(−ξ)eiξt dξ,
e−αsT (s)fn(s − t) ds
(α + iξ + A)−1Ffn(−ξ)eiξt dξ
t ∈ R.
= lim
α→0+
1
=
Fn(t) = lim
0
1
Here the function under the integral is well defined since Ffn vanishes on
a neighbourhood of E. Moreover, the function ξ 7→ (iξ + A)−1Ffn(−ξ)
is continuous and has compact support. Hence, by the Riemann-Lebesgue
theorem,
rieleb
rieleb
(6.21)
Since
we therefore obtain
F00F00
(6.22)
lim
t→∞kFn(t)k = 0,
n ≥ 2.
lim
n→∞
sup
t∈R kFn(t) − F0(t)k = 0,
lim
t→∞kF0(t)k = 0.
60
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Let us now examine the function F1. Take x ∈ X. The function t 7→
e−tT (t)x is in L2(R+; X). Extending this function by zero on (−∞, 0),
Plancherel's theorem implies that
ZR k(1 + iξ + A)−1xk2 dξ ≤ C 2kxk2.
For α ∈ (0, 1), let
F1,α,x(t) =ZR
e−α(t+s)T (t + s)x dµ1(s),
t ∈ R.
f11f11
(6.23)
Then
f13f13
(6.24)
F1,α,x(t) = F1(t)x.
lim
α→0+
Moreover, F1,α,x ∈ L1(R; X) and its Fourier transform is
(FF1,α,x)(ξ)
= Fµ(ξ)(1 − Fϕ(−ξ))(α + iξ + A)−1x
= Fµ(ξ)(1 − Fϕ(−ξ))(cid:0)I + (1 − α)(α + iξ + A)−1(cid:1) (1 + iξ + A)−1x,
by the resolvent identity. The assumption (6.20) extends by the Neumann
series to boundedness of k(α + iξ + A)−1k for small α > 0 and ξ ≥ R,
and then for all α > 0 and ξ ≥ R since the semigroup is bounded. Since
Fϕ = 1 in a neighbourhood of E, there is a constant C (independent of α,
ξ and x) such that
f12f12
(6.25)
k(FF1,α,x)(ξ)k ≤ Ck(1 + iξ + A)−1xk,
0 < α < 1, ξ ∈ R.
Moreover,
lim
α→0+
(FF1,α,x)(ξ)
= (Fµ(ξ)(1 − Fϕ(−ξ))(iξ + A)−1x if Fϕ(−ξ) 6= 1,
otherwise.
0
=: G1,x(ξ).
Using (6.23), (6.25), and the Dominated Convergence Theorem, it follows
that G1,x ∈ L2(R; X), kG1,xkL2(R;X) ≤ Ckxk and
lim
α→0+kFF1,α,x − G1,xkL2(R;X) = 0.
Using Plancherel's theorem again shows that
lim
α→0+kF1,α,x − F−1G1,xkL2(R;X) = 0,
where F−1G1,x is inverse L2-Fourier transform of G1,x. From this and (6.24),
we deduce that
f135
f135
(6.26)
F1(t)x = (F−1G1,x)(t)
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
61
for almost all t > 0. By Plancherel's theorem once more,
ZR kF1(t)xk2 dt ≤ C 2 kxk2,
x ∈ X.
Let t ≥ 0 and x ∈ X. We compute:
T (s + r)x dµ1(r) ds
0
0
T (t − s)F1(s)x ds
Z t
T (t − s)ZR
= Z t
T (t + r)x dµ1(r) ds −Z t
= Z t
0 ZR
= t F1(t)x +Z 0
tZ t
F1(t)x =
−t
rT (t + r)x dµ1(r),
1
T (t − s)F1(s)x ds −Z 0
0 Z (−s)−
−t
−t
We estimate the first term on the right-hand side:
0
T (t + r)x dµ1(r) ds
r
t
T (t + r)x dµ1(r).
so that
f14f14
(6.27)
f15f15
(6.28)
where K = supt≥0 kT (t)k. For the second term on the right-hand side of
(6.27), we note that
K kF1(s)xk ds ≤ KC
1
√t kxk,
1
tZ t
0
(cid:13)(cid:13)(cid:13)
0
1
T (t − s)F1(s)x ds(cid:13)(cid:13)(cid:13) ≤
t→∞Z 0
−t(cid:13)(cid:13)(cid:13)
tZ t
T (t + τ )(cid:13)(cid:13)(cid:13) dµ1(τ ) = 0,
lim
τ
t
by the Bounded Convergence Theorem for the bounded measure µ1. Thus
we have
From this and (6.22), we obtain
lim
t→∞kF1(t)k = 0.
lim
t→∞kF (t)k = lim
t→∞kT (t)µ(T )k = 0,
which is the claim.
(cid:3)
Now we return to the situation of Subsection 6.2. We assume that σ(A)∩
iR = {0} and that sups≥η k(is + A)−1k < +∞ for some η > 0. Since
one-point sets are of spectral synthesis, we can apply Theorem 6.14 with
µ ∈ M b(R+) given by
(6.29)
Then Fµ(0) = 0 and µ(T ) = I − (I + A)−1 = A(I + A)−1, so the conclusion
of Theorem 6.14 is that
µ = δ0 − e1,
e1(s) = e−s,
s ≥ 0.
mufor0
mufor0
lim
t→∞kT (t)A(I + A)−1k = 0.
62
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
The following result includes an estimate of the rate of decay of kT (t)A(I +
A)−1k in terms of the growth of k(is + A)−1k as s → 0.
mlog Theorem 6.15. Let (T (t))t≥0 be a C0-semigroup on a Hilbert space X with
generator −A, and assume that σ(A) ∩ iR = {0} and that sups≥η k(is +
A)−1k < +∞ for some η > 0. Then
(6.30)
lim
uniform00
uniform00
t→∞kT (t)A(I + A)−1k = 0.
More precisely, let m : (0, 1) → (1,∞) be a continuous increasing function
such that
k(is + A)−1k = O(m(s))
s → 0.
Let ε ∈ (0, 1). Then
(6.31)
polydec01
polydec01
kT (t)A(I + A)−1k = O(cid:0)m−1(t1−ε)(cid:1) ,
t → ∞.
Proof. We have already shown how (6.30) follows from Theorem 6.14. To
establish (6.31), we follow the proof of that theorem with µ given by (6.29)
and E = {0}. We take ϕ ∈ S(R) so that Fϕ has compact support and
Fϕ = 1 on [−1, 1]. Let
Then µ0 is absolutely continuous with density function
µ0 = µ ∗ ϕ,
µ1 = µ − µ0.
f (s) = ϕ(s) −Z ∞
0
ϕ(s − τ )e−τ dτ,
and
Ff (ξ) =
iξ
1 + iξ Fϕ(ξ).
Take a C∞-function ψ such that 0 ≤ ψ ≤ 1, ψ = 1 on [−1, 1] and the
support of ψ is contained in [−2, 2]. For 0 < r ≤ 1/2, let gr be the Schwartz
function such that
Fgr(ξ) = Ff (ξ)ψr(ξ) =(cid:18) iξ
1 + iξ(cid:19) ψ(cid:18) ξ
r(cid:19) .
The last equality holds because ψ(ξ/r) = 0 if ξ ≥ 1 and (Fϕ)(ξ) = 1 if
ξ ≤ 1. Then Fgr(ξ) = 0 if ξ ≥ 2r, and, for ξ ≤ 2r and j ≥ 1,
(Fgr)(j)(ξ) =
(−1)j−n+1ij−nj!
(1 + iξ)j−n+1n!
r(cid:19) +(cid:18) iξ
1 + iξ(cid:19) 1
rj ψ(j)(cid:18) ξ
r(cid:19) .
1
rn ψ(n)(cid:18)ξ
j−1Xn=0
Hence
Fgr(ξ) ≤ ξ,
Cj
(Fgr)(j)(ξ) ≤
rj−1(cid:18)1 + ξ
r (cid:19) ,
kFgrkL1 ≤ 4r2,
8Cj
rj−2 .
(cid:13)(cid:13)(Fgr)(j)(cid:13)(cid:13)L1 ≤
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
63
Now, for s ∈ R,
1
−2r ξ dξ =
1
1
so
−2r
−2r Fgr(ξ)eiξs dξ(cid:12)(cid:12)(cid:12)(cid:12) ≤
gr(s) = (cid:12)(cid:12)(cid:12)(cid:12)
2πZ 2r
2πZ 2r
(Fgr)′′ (ξ)eiξs dξ(cid:12)(cid:12)(cid:12)(cid:12) ≤ C,
s2gr(s) = (cid:12)(cid:12)(cid:12)(cid:12)
2πZ 2r
kgrkL1 ≤Zs≤1/r
ds +Zs≥1/r
k(Ffr)(j)kL1 ≤
C
s2 ds ≤ Cr.
Let fr = f − gr. Then fr ∈ L1(R), Ffr = 0 on [−r, r], and
Cj
Cj
rj−2 ,
kf − frkL1 ≤ Cr,
kFfrkL1 ≤ C,
2r2
π
2r2
π
,
j = 0, 1,
j ≥ 2.
These functions fr (as r → 0+) replace the functions fn (as n → ∞) in the
proof of Theorem 6.14. Accordingly, we define
Fr(t) =ZR
T (t + s)fr(s) ds,
0 < r < 1/2, t > 0.
Instead of applying the Riemann-Lebesgue theorem as we did in the proof
of Theorem 6.14 to obtain (6.21), we integrate by parts k times, and we
obtain that
ik
j(cid:19) ik−j(k − j)!(iξ + A)−(k−j+1)(Ffr)(j)(−ξ)eiξt dξ.
1
2πtkZR(cid:18) d
dξ(cid:19)k(cid:0)(iξ + A)−1Ffr(−ξ)(cid:1) eiξt dξ
kXj=0(cid:18)k
2πiktkZR
rj−2 ≤
tkm(r)k+1 + m(r)k +
kF0(t)k ≤ kFr(t)k + Kkf − frkL1 ≤ C(cid:18) m(r)k+1
m(r)k−j+1
kXj=2
tk
C
Cm(r)k+1
tk
.
+ r(cid:19) ,
Since m(r) ≥ r−1 ≥ 2, we have
Fr(t) =
=
kFr(t)k ≤
Now
where K = supt≥0 kT (t)k.
Now we want to show that
oneovert
oneovert
(6.32)
kF1(t)k = O(t−1),
t → ∞.
The estimate (6.28) is not adequate for this purpose, and we use a different
argument.
Let x ∈ X. First note that the resolvent identity yields
k(iξ + A)−1xk ≤ (C + 1)k(iξ + 1 + A)−1xk,
(6.33)
in1in1
ξ ≥ 1,
64
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
in2in2
where C := supξ≥1 k(iξ + A)−1k. Similarly,
(6.34)
k(−iξ + A∗)−1xk ≤ (C + 1)k(−iξ + 1 + A∗)−1xk,
1 + iξ(cid:0)1 − Fϕ(−ξ)(cid:1),
g(ξ) := Fµ(ξ)(1 − Fϕ(−ξ)) =
iξ
Set
ξ ≥ 1.
ξ ∈ R,
and note that g is zero on (−1, 1) and g′ ∈ L1(R) ∩ L∞(R). Let x ∈ X.
It follows from (6.26) that F1(·)x is the inverse Fourier transform (in the
L2-sense) of the function G1,x which has derivative
g1x'
g1x'
G′1,x(ξ) = −g(ξ)(iξ + A)−2x + g′(ξ)(iξ + A)−1x.
(6.35)
Thus G′1,x ∈ L2(R; X), and G1,x belongs to the first-order vector-valued
Hilbert-Sobolev space H 1(R; X). Hence
tF1(t)x = −iF−1(G′1,x)(t).
Since g′ ∈ L1(R), the second term on the right-hand side of (6.35) is in
L1(R, X), and its inverse Fourier transform is bounded by (2π)−1kg′k1Ckxk.
To handle the first term, take y ∈ X. By the Cauchy-Schwarz inequality,
(6.33) and (6.34) and Plancherel's theorem,
1
2πZR(cid:12)(cid:12)hg(ξ)(iξ + A)−2x, yi(cid:12)(cid:12) dξ
≤ kgk∞
2π ZR\(−1,1) k(iξ + A)−1xk2 dξ!1/2 ZR\(−1,1) k(−iξ + A∗)−1yk2 dξ!1/2
(cid:18)ZR k(1 + iξ + A)−1xk2 dξ(cid:19)1/2(cid:18)ZR k(1 − iξ + A∗)−1yk2 dξ(cid:19)1/2
(C + 1)2kgk∞
2π
≤
ke−τ T (τ )xk2 dτ(cid:19)1/2(cid:18)Z ∞
0
ke−τ T ∗(τ )yk2 dτ(cid:19)1/2
= (C + 1)2kgk∞(cid:18)Z ∞
≤ (C + 1)2kgk∞K 2kxkkyk.
It then follows that
0
thF1(t)x, yi ≤ (C + 1)2kgk∞K 2kxkkyk + kg′k1C
2π kxkkyk
for almost all t. Since F1 is continuous, this estimate holds for all t, so (6.32)
follows.
Overall we have that
kT (t)A(1 + A)−1k = kF0(t) + F1(t)k ≤ C(cid:18) m(r)k+1
tk
+
1
t
+ r(cid:19) .
For a given ε ∈ (0, 1), we take r = m−1(t1−ε), for sufficiently large t. Then
we obtain
kT (t)A(1 + A)−1k ≤ C(cid:16)t1−(k+1)ε + t−1 + m−1(t1−ε)(cid:17) .
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
65
We may choose k so that (k + 1)ε ≥ 2. Since m(r) ≥ r−1, we have
m−1(t1−ε) ≥ tε−1 ≥ t−1. Then we obtain (6.31).
(cid:3)
ex.KT+ Example 6.16. (a) Consider the case when m(s) = s−α where α ≥ 1. Then
(6.31) gives
kT (t)A(I + A)−1k = O(cid:0)t−γ(cid:1) ,
t → ∞,
for any γ < 1/α. We shall see in Theorem 7.6 that this also holds with
γ = 1/α.
(b) Now consider the case when m(s) = eα/s where α > 0. Then
m−1(t1−ε) =
α
(1 − ε) log t
,
so
kT (t)A(I + A)−1k = O(cid:0)(log t)−1(cid:1) = O(cid:0)m−1(t)(cid:1) ,
This rate of decay is sharp in this case (Corollary 6.11).
t → ∞.
s.zero2
7. Singularity at zero: polynomial and regularly varying rates
In this section we consider the situation of Subsection 6.2 and Theorem
6.15 in cases when X is a Hilbert space and the function m(1/s) is regularly
varying. For simplicity of presentation, we shall write B(A) = A(I + A)−1.
Let (T (t))t≥0 be a bounded C0-semigroup with generator −A such that
σ(A) ∩ iR = {0} and
O(1),
O(cid:18)
k(is + A)−1k =
kT (t)B(A)k = O(cid:18)
1
sαℓ(1/s)(cid:19) ,
s → 0,
s → ∞,
1
(tℓ(t1/α))1/α(cid:19) ,
t → ∞.
where ℓ is slowly varying and monotonic. Then the optimal estimate (6.16)
on the decay of kT (t)B(A)k under the assumption of dB-symmetry of ℓ
would be
We shall show in Theorem 7.7 that this estimate does hold if ℓ is increasing,
i.e., k(is+A)−1k grows slightly slower than s−α as s → 0. Its proof requires
a series of steps similar to those used in the case of a singularity at infinity
in Section 5.
7.1. Cancelling resolvent growth.
dfm Definition 7.1. Let β ∈ (0, 1], and let ℓ be a slowly varying function such
that g : s 7→ s1−βℓ(s) is increasing on R+. Let Sg be the Stieltjes function
associated with g (Example 2.14), so
Sg(λ) =Z ∞
0+
d(cid:0)s1−βℓ(s)(cid:1)
s + λ
,
λ > 0.
66
Let
fg-def
(7.1)
fm
(7.2)
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
fg(λ)
fm(λ)
:= Sg(1/λ),
fg(λ)
:=
1 + fg(λ)
λ > 0,
,
λ > 0.
Since fg is a complete Bernstein function, 1/fg is Stieltjes by Theorem
2.2, and then fm = 1/(1 + 1/fg) is also a complete Bernstein function.
Thus the operator fm(A) is well-defined, either by Definition 3.3 or by the
extended functional calculus of Theorem 3.7 (see Remark 3.8(iii)). Moreover
fm(A) = fg(A)(I + fg(A))−1, by the composition rule in Theorem 3.6(ii).
In particular fm(A) is bounded (see Remark 3.4 and [64, Corollary 12.7]).
Now we make a definition analogous to (5.13).
dvabl Definition 7.2. For α ≥ 1 and β ∈ (0, 1], define
Vα,β,ℓ(A) := B(A)α−βfm(A) = B(A)α−βfg(A)(I + fg(A))−1,
so that Vα,β,ℓ(A) ∈ L(X) by the above.
The next statement shows that this operator cancels resolvent growth in
the case of a singularity at zero. It is a counterpart of Theorem 5.5 dealing
with singularities at infinity. In the proof we again use Karamata's Theo-
rem 2.15, Proposition 2.4 on domination properties of complete Bernstein
functions and complex analysis arguments shown in Lemma 5.3.
thm.bounded0 Theorem 7.3. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
and
0001
0001
0boundedd1
0boundedd1
0 ∈ σ(A) ⊂ C+ ∪ {0}
X, with generator −A such that
(7.3)
Let ℓ be a slowly varying function on R+ such that g : s 7→ s1−βℓ(s) is
increasing for some β ∈ (0, 1]. If there exists α > 1 such that
(7.4)
s → 0,
k(is + A)−1k = O(1),
s → ∞.
k(is + A)−1k = O(cid:18)
1
sαℓ(1/s)(cid:19) ,
then
sup
λ∈C+ k(λ + A)−1Vα,β,ℓ(A)k < ∞.
alpha1 Remark 7.4. We are interested only in the case when 0 ∈ σ(A), so k(is +
A)−1k ≥ s−1. Thus the assumption that α > 1 restricts generality only
slightly. Moreover, if α = 1, then the function ℓ can be increasing only if
it is bounded. In this case, (7.4) is equivalent to the same estimate with
ℓ ≡ 1. This situation is settled by Corollary 7.5 below. Hence if α = 1 we
are interested only in the case when ℓ is decreasing and limλ→∞ ℓ(λ) = 0,
and then necessarily β ∈ (0, 1). The following proof shows that Theorem
7.3 is also true in that case, provided thatPn ℓ(2n) converges.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
67
estimwithv2
estimwithv2
0resident3
0resident3
Proof. By the assumption (7.3) and Lemma 5.3 it suffices to prove that
sup(cid:8)k(is + A)−1Vα,β,ℓ(A)k : s ∈ R, 0 < s ≤ 1(cid:9) < ∞.
(7.5)
To this aim, observe first that if k ∈ N, then
(7.6)
(λ + A)−1B(A)k =
k−1Xi=0
(−λ)k−1−iAi(I + A)−k + (−λ)k(λ + A)−1(I + A)−k.
Let α− β = m + γ where m ∈ N∪{0}, and 0 ≤ γ < 1. By the assumption
(7.4),
kA(is + A)−1k = kI − is(is + A)−1k = O(cid:18)
1
sα−1ℓ(1/s)(cid:19) ,
s → 0.
By the moment inequality (4.2),
kAγ (is + A)−1k = O(cid:18)
1
sα−γℓ(1/s)(cid:19) ,
s → 0.
Using (7.6) for k = m, it follows that
bounded11
(7.7)
kB(A)α−β(is + A)−1k = kAγ(is + A)−1B(A)m(I + A)−γk
= O(cid:18)
= O(cid:18)
1
sα−γ−mℓ(1/s)(cid:19)
sβℓ(1/s)(cid:19) ,
1
s → 0.
Let fg and fm be the complete Bernstein functions defined by (7.1) and
(7.2). Then
lim
λ→0+
fg(λ) = lim
λ→0+
fm(λ) = 0,
and in addition fm is bounded. Hence fm has Stieltjes representation of the
form (0, 0, ν). Moreover, by (5.18),
fg(s) = O(cid:16)sβℓ(1/s)(cid:17) ,
s → 0.
fgfg
(7.8)
By Proposition 2.4 and (7.8),
bratio
bratio
(7.9)
fm(is) ≤ Cfm(s) =
Cfg(s)
1 + fg(s)
= O(cid:16)sβℓ(1/s)(cid:17) ,
s → 0,
Now fm(A) is a bounded operator, and, by (3.3), for every x ∈ dom (A),
fm(A)x =Z ∞
0+
A(λ + A)−1x dν(λ).
estimateondom
estimateondom
We shall show that there exists C > 0 such that
(7.10)
k(is + A)−1B(A)α−βfm(A)xk ≤ Ckxk,
x ∈ dom(A), 0 < s ≤ 1.
Since dom(A) is dense in X, this will imply that (7.10) holds for all x ∈ X,
thus giving us (7.5) and completing the proof.
68
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
If x ∈ dom(A) and s ∈ R \ {0}, then
(is + A)−1fm(A)x
(is + A)−1A(λ + A)−1x dν(λ)
0+
= Z ∞
= Z ∞
(λ + A)−1x −
= fm(−is)(is + A)−1x +Z ∞
0+ (cid:18) λ
λ − is
0+
(is + A)−1x(cid:19) dν(λ)
(λ + A)−1x dν(λ).
is
λ − is
λ
λ − is
Thus
3summand
(7.11)
(is + A)−1fm(A)B(A)α−βx
= fm(−is)(is + A)−1B(A)α−βx +Z 1
0+
λ
λ − is
(λ + A)−1B(A)α−βx dν(λ)
(λ + A)−1B(A)α−βx dν(λ).
+Z ∞
1+
λ
λ − is
we shall estimate each of the three summands above separately.
To bound the first summand, note that by (7.7) and (7.9), there exists
C > 0 such that
fm(−is)k(is + A)−1B(A)α−βxk ≤ Ckxk,
To estimate the second summand in (7.11), we take into account that
x ∈ X, 0 < s ≤ 1.
α > β and consider the following two cases.
If α − β ∈ (0, 1) then (4.1) yields
kB(A)α−β(λ + A)−1k = k(I + A)β−αAα−β(λ + A)−1k ≤ Cλα−β−1,
by Proposition 4.1. Then we obtain
λα−β−1 dν(λ)kxk.
We have to show that the integral here is finite. For this, let In be the
interval (2−(n+1), 2−n] for n ∈ N ∪ {0}. Then
0+
λ
0+
λ − is
2−(n−1)
2−n + λ
B(A)α−β(λ + A)−1x dν(λ)(cid:13)(cid:13)(cid:13)(cid:13) ≤ CZ 1
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
dν(λ) ≤ZIn
ZIn
λα−β−1 dν(λ) ≤ 2−n(α−β−1)+1ZIn
ZIn
Z 1
∞Xn=0ZIn
λα−β−1 dν(λ) =
0+
For α > 1, it follows from (2.6) with γ = (α − 1)/2 that
by (7.9). Hence
λα−β−1 dν(λ) < ∞.
dν(λ) ≤ 2fm(2−n) ≤ C2−nβℓ(2n),
dν(λ) ≤ C2−n(α−1)ℓ(2n).
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
69
and
If α − β ≥ 1, then
The same conclusion holds if α = 1 andP∞n=0 ℓ(2n) converges.
kB(A)α−β(λ + A)−1k ≤ kB(A)α−β−1kk(I + A)−1(cid:0)I − λ(λ + A)−1(cid:1)k ≤ C,
(cid:13)(cid:13)(cid:13)(cid:13)Z 1
B(A)α−β(λ + A)−1x dν(λ)(cid:13)(cid:13)(cid:13)(cid:13) ≤ CZ 1
B(A)α−β(λ + A)−1x dν(λ)(cid:13)(cid:13)(cid:13)(cid:13) ≤ KZ ∞
For the third summand in (7.11) we have for any s ∈ R,
dν(λ)
λ kB(A)α−βkkxk
The integral here is finite, by (2.3).
(cid:13)(cid:13)(cid:13)(cid:13)Z ∞
1+
= Ckxk,
dν(λ)kxk.
0+
λ
λ − is
λ
0+
λ − is
where K = supt≥0 kT (t)k, and C is finite by (2.3).
proof of the theorem is complete.
Finally, summarizing all the estimates above we get (7.10), and thus the
(cid:3)
1+
7.2. Resolvent growth equal to s−α. Here we consider the case when
we assume that the resolvent grows (at most) like s−α. We note first a
corollary from the arguments of the proof of Theorem 7.3 that will be useful
in the next subsection and also covers the case α = 1, ℓ ≡ 1 of Theorem 7.3
excluded in Remark 7.4.
cor.polybound0 Corollary 7.5. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
X with generator −A. Assume that σ(A) ∩ iR = {0} and there exists α ≥ 1
such that
k(is + A)−1k =(O (s−α) ,
O(1),
s → 0,
s → ∞.
Then
sup
λ∈C+ k(λ + A)−1B(A)αk < ∞.
Proof. The proof follows the same lines as the proof of Theorem 7.3, but
it is much simpler. It follows immediately from (7.8) (now for β = 0 and
ℓ = 1) and Lemma 5.3.
(cid:3)
The following result, which is analogous to [17, Theorem 2.4], gives the
optimal result in the case of exactly polynomial growth of the resolvent for
a singularity at zero.
polydec00 Theorem 7.6. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X with generator −A. Assume that σ(A) ∩ iR = {0}, and let α ≥ 1. The
following are equivalent:
resolvpolynom1
resolvpolynom2
(i) k(is + A)−1k =(O (s−α),
O (1),
(ii) kT (t)B(A)αk = O (t−1),
s → 0,
s → ∞.
t → ∞.
70
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
resolvpolynom3
(iii) kT (t)B(A)k = O (t−1/α),
Proof. By Corollary 7.5, the property (i) implies that
t → ∞.
sup
λ∈C+ k(λ + A)−1B(A)αk < ∞.
Then Theorem 4.7 yields (ii).
Properties (ii) and (iii) are equivalent by Lemma 4.2. Property (iii) im-
(cid:3)
plies (i) by Theorem 6.10.
7.3. Resolvent growth slower than s−α. Here we give a counterpart of
Theorem 5.6.
thm.main0 Theorem 7.7. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
0boundinf
0boundinf
1
X, with generator −A. Assume that σ(A) ∩ iR = {0} and
s → 0,
s → ∞,
sαℓ(1/s)(cid:19) ,
(7.12)
O(1),
where α > 1 and ℓ is increasing and slowly varying. Then
O(cid:18)
k(is + A)−1k =
kT (t)A(I + A)−1k = O(cid:18)
1
(tℓ(t1/α))1/α(cid:19) ,
t → ∞.
0BTest
0BTest
(7.13)
Proof. We can assume that T (t)A 6= 0 for each t > 0. Since (7.12) implies
property (i) of Theorem 7.6, we obtain from property (iii) of that theorem
that
We use the notation of Definitions 7.1 and 7.2, and apply Theorem 7.3,
with β = 1 and g = ℓ. We conclude that there exists C > 0 such that
kT (t)B(A)k = O(cid:0)t−1/α(cid:1),
k(λ + A)−1B(A)α−1fm(A)k ≤ C,
t → ∞.
λ ∈ C+.
Then, by Theorem 4.7, we have
kT (t)B(A)α−1fm(A)k ≤
C
t
,
t > 0,
Hence by Theorem 4.3(b), with γ = α − 1, and t1 = t2 = t,
fm(kT (t)B(A)k) ≤
CkT (t)B(A)k
tkT (2t)B(A)αk
.
Putting fα,m(s) = sα−1fm(s) we then obtain that
0fatest2
0fatest2
(7.14)
fα,m(kT (t)B(A)k) ≤
C
t
kT (t)B(A)kα
kT (2t)B(A)αk
.
By Theorem 2.15 with g = ℓ and ρ = σ = 1, Sℓ(t) ∼ t−1ℓ(t) as t → ∞.
The function fm(s) has the same decay at zero as fℓ(s), hence fα,m(s) ∼
sαℓ(1/s) as s → 0+. Let k(s) = 1/ℓ(s1/α). Then
sα
fα,m(s) ∼
k(s−α)
,
s → 0 + .
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
71
By Proposition 2.11(c),
f−1
α,m(s) ∼(cid:18)
s
k#(1/s)(cid:19)1/α
,
s → 0 + .
If the right-hand side of (7.14) is sufficiently small,
CkT (t)B(A)k
kT (t)B(A)k ≤
(tℓ(t)kT (2t)B(A)αk)1/α
,
where
L(t) = k#(cid:18)tkT (2t)B(A)k
kT (t)B(A)kα(cid:19) .
Let ψ(s) = (sk#(s))1/α. Since ψ is regularly varying with positive index,
we can choose k# so that ψ is strictly increasing and continuous (see the
remarks following Definition 2.9; in fact, we can choose k# to be increasing,
since k is decreasing). Then
0psiest
(7.15)
kT (t)B(A)k−1 ≥ ct1/α kT (2t)B(A)αk1/α
kT (t)B(A)k
kT (t)B(A)kα (cid:19) .
= c ψ(cid:18)tkT (2t)B(A)αk
L(t)1/α
If the right-hand side of (7.14) is bounded away from 0, then (7.15) also
holds for some c > 0, since kT (t)B(A)k is bounded and ψ is bounded on
bounded intervals. So (7.15) holds for all t > 0, for some c > 0.
Since ψ is strictly increasing,
tkT (2t)B(A)αk
kT (t)B(A)kα ≤ ψ−1(cid:0)CkT (t)B(A)k−1(cid:1) ,
ψ−1(s) ∼ sαk##(sα) ∼
for some constant C > 0. By Proposition 2.11(b),
s → ∞,
sα
ℓ(s)
,
so
Then (7.13) yields
kT (2t)B(A)αk ≤
C
tℓ (kT (t)B(A)k−1)
.
kT (2t)B(A)αk ≤
C
tℓ(t1/α)
,
for sufficiently large t. Since B(A) is sectorial, we can apply Lemma 4.2
with B = B(A), γ = α and δ = 1 and we obtain
kT (t)B(A)k ≤
C
(tℓ(t))1/α
,
for large t, and the proof is finished.
(cid:3)
Now we can formulate a counterpart of Corollary 5.7, showing that the
optimal estimate (6.16) holds when ℓ is increasing and dB-symmetric.
72
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Corollary 7.8. In addition to the assumptions of Theorem 7.7, assume that
ℓ is dB-symmetric. Then
where m−1 is any asymptotic inverse of sα/ℓ(s).
kT (t)B(A)k = O(cid:0)m−1(t)(cid:1) ,
t → ∞,
Proof. The proof is completely analogous to the proof of Corollary 5.7. (cid:3)
The discussion following the proof of Corollary 5.7 and addressing the
situation when ℓ may not be dB-symmetric applies to the current setting as
well. We omit the details.
Remark 7.9. Note the following duality between singularities at zero and at
infinity. If (T (t))t≥0 is a C0-semigroup of contractions on a Hilbert space
X with generator −A and the range of A is dense in X, then A is injective
and the operator −A−1, with dense domain ran(A), also generates a C0-
semigroup of contractions on X, by a direct application of the Lumer-Phillips
theorem. Thus the case of a singularity at infinity for A corresponds to the
case of a singularity at zero for A−1. Conversely, the case of a singularity at
zero for A corresponds, in general, to the case of singularities at both zero
and at infinity for A−1. This case will be studied in Section 8.
Remark 7.10. We are unable to give a result corresponding to Theorem
5.12 because of the lack of an initial estimate for the rate of decay that is
sufficiently close to (6.16) for our technique to work. When k(is + A)−1k
grows slightly faster than s−α, our best estimate is
kT (t)B(A)k = O(cid:0)t−1/β(cid:1),
t → ∞,
for each β > α (see Example 6.16(a) and Theorem 7.6). Our technique does
not improve this.
sectwosing
8. Singularities at both zero and infinity
In this section we study the rates of decay in the context of Corollary 6.2
and similar situations, so we are interested in the property
uniform2
uniform2
(8.1)
lim
t→∞kT (t)A(I + A)−2k = 0,
for a bounded C0-semigroup (T (t))t≥0 on a Banach space or Hilbert space.
Note that the range of A(I + A)−2 is ran(A)∩ dom(A) (Proposition 3.10(i)).
Thus we consider orbits starting in ran(A) ∩ dom(A) or similar spaces.
We shall assume that (8.1) holds, or equivalently that σ(A)∩iR ⊂ {0} (see
Corollary 6.2). We wish to examine the relation between the rate of decay
in (8.1) and the rate of growth of k(is + A)−1k as s → 0 or as s → ∞. Let
M : [1,∞) → R+ be a continuous increasing function, and m : (0, 1] → R+
be a continuous decreasing function, such that
k(is + A)−1k ≤(m(s),
M (s),
0 < s < 1,
s ≥ 1.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
73
We assume that 0 ∈ σ(A) and that lims→∞ M (s) = ∞ (otherwise we are in
the situation of Section 5, or Sections 6 and 7, respectively). By standard
theory, m(s) ≥ 1/s.
(6.8):
For example, M and m can be defined by modified versions of (5.5) and
defM2
defm2
(8.2)
(8.3)
M (s) = sup(cid:8)k(ir + A)−1k : 1 ≤ r ≤ s(cid:9) ,
m(s) = sup(cid:8)k(ir + A)−1k : s < r ≤ 1(cid:9) ,
Let N2 be a continuous, decreasing function such that
s ≥ 1,
0 < s < 1.
kT (t)A(I + A)−2k ≤ N2(t),
t ≥ 0.
For example, we can take
N2(t) = sup{kT (τ )A(I + A)−2k : τ ≥ t},
t ≥ 0.
By Corollary 6.2, our assumptions imply that limt→∞ N2(t) = 0. Let N−1
denote any right inverse for N2.
2
Under our assumptions, Mart´ınez [55, Corollary 3.3] has shown that
martinez
martinez
(8.4)
kT (t)A(I + A)−2k ≤ C max m−1
log(c′t),
log (C′t)! ,
1
M−1
where Mlog is defined by (5.3) and
mlog(s) = m(s) log(cid:18) 1 + m(s)
s
(cid:19) ,
0 < s ≤ 1.
We have omitted a term t−1 which appears in [55], because m−1(t) ≥ t when
0 ∈ σ(A).
The upper bound (8.4) is analogous to the upper bound for kT (t)A−1k
given in Theorem 1.1 and there is also a lower bound for kT (t)A(I + A)−2k.
Theorem 6.9 extends to our present situation, with an almost unchanged
proof. If lim supt→∞ tkT (t)A(I + A)−2k = 0, then X splits as in Theorem
6.9(ii).
The following is an analogue of Theorem 6.10, with a similar proof. The
term s−1 in the estimate (8.5) is relevant only when splitting occurs and
then k(is + A)−1k is comparable to s−1 when s is sufficiently small.
resbound0+ Theorem 8.1. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
X, with generator −A. Assume that σ(A)∩iR = {0}, and let N2 be as above.
Then, for any c ∈ (0, 1),
resest0+
resest0+
(8.5)
k(is + A)−1k =(O(cid:0)N−1
O(cid:0)N−1
2 (cs) + s−1(cid:1) ,
(cid:0)cs−1(cid:1)(cid:1)
2
s → 0,
s → ∞.
74
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
Proof. Let K = supt≥0 kT (t)k. Let s ∈ R \ {0}, t > 0 and x ∈ dom(A). We
use the formula
iseistx = iseistZ t
= iseistZ t
0
0
e−isτ T (τ )(is + A)x dτ + isT (t)x
e−isτ T (τ )(is + A)x dτ + T (t)(1 + A)−1(is + A)x
− (1 − is)T (t)A(I + A)−2((is + A) + (1 − is))x.
Since kT (t)(I + A)−1k ≤ K, this gives
skxk ≤ K(st + 1)k(is + A)xk + 1 − isN2(t) (k(is + A)xk + 1 − iskxk) .
Hence
(8.6) (cid:0)s − 1 − is2N2(t)(cid:1)kxk ≤ (K(st + 1) + 1 − isN2(t))k(is + A)xk.
Set t = N−1
2 (cs). For s sufficiently small,
Nest0
Nest0
For any K′ > K, (8.6) gives
s − 1 − is2N2(t) = s(cid:0)1 − c(1 + s2)(cid:1) > 0.
k(is + A)−1k ≤
2 (cs) + 1) + 1 − iscs
s(1 − c(1 + s2))
≤
2 (cs) + s−1(cid:1) ,
s − 1 − is2N2(t) = s(cid:0)1 − c(1 + s−2)(cid:1) > 0.
2 (c/s). For s sufficiently large,
K(sN−1
K′
1 − c(cid:0)N−1
for s sufficiently small.
Now set t = N−1
So we obtain
k(is + A)−1k ≤
K(cid:0)sN−1
= O(cid:0)N−1
2 (cs−1) + 1(cid:1) + 1 − iscs−1
2 (cs−1)(cid:1) ,
s (1 − c(1 + s−2))
s → ∞.
(cid:3)
As in Corollary 6.11 and (5.6), we obtain the following lower bound.
lowbound0+ Corollary 8.2. Let (T (t))t≥0 be a bounded C0-semigroup on a Banach space
X, with generator −A. Assume that σ(A) ∩ iR = {0} and that
Define M and m by (8.2) and (8.3) respectively. Then there exist c, c′, C′ > 0
such that
lim
s→0
max(cid:0)ks(is + A)−1k,ks(−is + A)−1k(cid:1) = ∞.
M−1(C′t)(cid:19)
kT (t)A(I + A)−2k ≥ c max(cid:18)m−1(c′t),
1
for all sufficiently large t.
Now we return to upper bounds, and we consider polynomially growing m
and M . In this case we can omit the logarithmic terms in (8.4) in a similar
way to Theorem 7.6.
poldecayzero Theorem 8.4. Let (T (t))t≥0 be a bounded C0-semigroup on a Hilbert space
X with generator −A. Assume that σ(A) ∩ iR = {0} and that there exist
α ≥ 1 and β > 0 such that
(8.7)
resolvpolynom+
resolvpolynom+
k(is + A)−1k =(O (s−α) ,
O(cid:0)sβ(cid:1) ,
s → 0,
s → ∞.
polynomdecayzero+
polynomdecayzero+
polynomdecay1+
polynomdecay1+
Then
(8.8)
and
(8.9)
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
75
seifert Remark 8.3. (a) For contraction semigroups, the assumption in Corollary
8.2 can be weakened in the same way as in Remark 6.12.
(b)
If
(T (t))t≥0
is bounded and eventually differentiable,
then
(I + A)2T (τ ) is a bounded operator for some τ > 0. Moreover,
AT (t + τ ) = (I + A)2T (τ )T (t)A(I + A)−2,
so limt→∞ kAT (t+τ )k = 0 if σ(A)∩iR = {0}. By Theorem 6.10, k(is+A)−1k
is bounded for s ≥ 1. Then (8.4) gives
kAT (t)k = O(cid:16)m−1
log(c′t)(cid:17) .
kT (t)Aα(I + A)−(α+β)k = O (t−1),
kT (t)A(I + A)−2k = O(cid:16)t−1/γ(cid:17) ,
t → ∞,
t → ∞,
where γ = max(α, β).
Conversely, if (8.9) holds for some γ > 0, then (8.7) holds for α =
max(γ, 1) and β = γ.
Proof. The argument which establishes (7.7) in the proof of Theorem 7.3
(where one puts β = 0 and ℓ ≡ 1) shows that
sup(cid:8)k(is + A)−1(A(I + A)−1)αk : 0 < s ≤ 1(cid:9) < ∞.
Similarly, the argument that (i) implies (5.17) in the proof of Theorem 5.5
shows that
Using product and composition rules similarly to (4.4),
sup(cid:8)k(is + A)−1(I + A)−βk : s ≥ 1(cid:9) < ∞.
sup(cid:8)k(is + A)−1Aα(I + A)−(α+β)k : s 6= 0(cid:9) < ∞.
Now we can apply Lemma 5.3 to the function
f (z) = (z + A)−1Aα(I + A)−(α+β),
and we deduce that
sup(cid:8)k(λ + A)−1Aα(I + A)−(α+β)k : λ ∈ C+(cid:9) < ∞.
76
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
By Theorem 4.7,
kT (t)Aα(I + A)−(α+β)k = O(t−1),
Since Aγ−α(I + A)−(2γ−α−β) is a bounded operator,
t → ∞.
kT (t)(A(I + A)−2)γk = O(t−1),
Since A(I + A)−2 is sectorial, by Lemma 4.2
kT (t)A(I + A)−2k = O(t−1/γ),
for γ = max(α, β).
t → ∞.
t → ∞,
The converse statement follows from Theorem 8.1.
(cid:3)
In Theorem 8.4, the case γ < 1 can occur only if the space X splits as
described before Theorem 8.1.
Remark 8.5. Proposition 3.10 (i) and (8.8) show that kT (t)xk = O(t−1)
for all x ∈ ran(Aα) ∩ dom(Aβ). Moreover, Proposition 3.10 (ii) shows that
Aa(I + A)−1 is sectorial whenever 0 < a < 1, in particular for a = α/(α+ β).
By Lemma 4.2 one obtains that
kT (t)Aαγ (I + A)−(αγ+βγ)k = O(cid:0)t−γ(cid:1) ,
t → ∞,
for every γ > 0. Hence kT (t)xk = O(t−γ) for all x ∈ ran(Aαγ) ∩ dom(Aβγ).
We do not consider regularly varying rates of resolvent growth in the
context of this section. It is not easy to find a single operator which cancels
resolvent growth at both zero and infinity simultaneously, without losing
information about the fine scale of the behaviour.
Finally we make some remarks about quasi-multiplication semigroups. In
many cases (including normal semigroups on Hilbert spaces, and multipli-
cation semigroups on Lp-spaces) the space X can be split into a direct sum
of two closed invariant subspaces X0 and X1 so that the generator of the
semigroup is bounded when restricted to X0 and invertible when restricted
to X1. Then the rate of decay on X is the maximum of the rates on X0 and
X1, so upper and lower bounds on X can be deduced from those on X0 and
X1. However knowing only the rate of decay on X is not sufficient to detect
whether it is controlled on X0 and X1. Consequently we do not think it is
possible to formulate a succinct result such as Propositions 5.1 and 6.13 in
terms of k(is + A)−1k in this case. If k(is + A)−1k dominates k(is−1 + A)−1k
whenever s > 1, then the rate of decay is determined by the behaviour
If k(is−1 + A)−1k dominates
on X1, and one can apply Proposition 5.1.
k(is + A)−1k whenever s > 1, then the behaviour on X0 dominates and
Proposition 6.13 is applicable.
References
[1] F. Alabau-Boussouira, P. Cannarsa, and R. Guglielmi, Indirect stabilization of
weakly coupled systems with hybrid boundary conditions, Math. Control Relat. Fields
1 (2011), 413 -- 436.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
77
[2] K. Ammari, E. Feireisl, S. Nicaise, Polynomial stabilization of some dissipative hy-
perbolic systems, Discret Contin. Dynam. Systems 34 (2014), 4371 -- 4388.
[3] K. Ammari, D. Mercier, V. R´egnier, and J. Valein, Spectral analysis and stabilization
of a chain of serially connected Euler-Bernoulli beams and strings, Commun. Pure
Appl. Anal. 11 (2012), 785 -- 807.
[4] N. Anantharaman and M. L´eautaud, Decay rates for the damped wave equation
on the torus, preprint, with appendix by S. Nonnenmacher, Anal. PDE, 7 (2014),
159 -- 214.
[5] W. Arendt and C. J. K. Batty, Tauberian theorems and stability of one-parameter
semigroups, Trans. Amer. Math. Soc. 306 (1988), 837 -- 852.
[6] W. Arendt, C. J. K. Batty, M. Hieber, and F. Neubrander, Vector-valued Laplace
Transforms and Cauchy Problems, 2nd ed., Monographs in Mathematics, vol. 96,
Birkhauser, Basel, 2011.
[7] W. Arendt and J. Pruss, Vector-valued Tauberian theorems and asymptotic behavior
of linear Volterra equations, SIAM J. Math. Anal. 23 (1992), 412 -- 448.
[8] M. Bassam, D. Mercier, S. Nicaise, and A. Wehbe, Stabilisation fronti`ere indirecte
du syst`eme de Timoshenko, C. R. Math. Acad. Sci. Paris 349 (2011), 379 -- 384.
[9] A. B´atkai, K. -- J. Engel, J. Pruss and R. Schnaubelt, Polynomial stability of operator
semigroups, Math. Nachr. 279 (2006), 1425 -- 1440.
[10] C. J. K. Batty, Tauberian theorems for the Laplace-Stieltjes transform, Trans. Amer.
Math. Soc. 322 (1990), 783 -- 804.
[11] C. J. K. Batty, Asymptotic behaviour of semigroups of operators, in: Functional
analysis and operator theory (Warsaw, 1992), Banach Center Publ. 30, Polish Acad.
Sci., Warsaw, 1994, pp. 35 -- 52.
[12] C. J. K. Batty, Differentiability of perturbed semigroups and delay semigroups, in:
Perspectives in operator theory, Banach Center Publ. 75, Polish Acad. Sci., Warsaw,
2007, 39 -- 53.
[13] C. J. K. Batty and T. Duyckaerts, Non-uniform stability for bounded semi-groups
on Banach spaces, J. Evol. Eq. 8 (2008), 765 -- 780.
[14] C. J. K. Batty, A. Gomilko and Yu. Tomilov, Product formulas in functional calculi,
Math. Z. 279 (2015), 479 -- 507.
[15] N. H. Bingham, C. M. Goldie, and J. L. Teugels, Regular variation, Encyclopedia
of Mathematics and its Applications 27, Cambridge University Press, Cambridge,
1987.
[16] J.-F. Bony and V. Petkov, Resolvent estimates and local energy decay for hyperbolic
equations, Ann. Univ. Ferrara Sez. VII Sci. Mat. 52 (2006), 233 -- 246.
[17] A. Borichev and Yu. Tomilov, Optimal polynomial decay of functions and operator
semigroups, Math. Ann. 347 (2010), 455 -- 478.
[18] N. Burq, D´ecroissance de l'´energie locale de l'´equation des ondes pour le probl`eme
ext´erieur et absence de r´esonance au voisinage du r´eel, Acta Math. 180 (1998),
1 -- 29.
[19] N. Burq and M. Hitrik, Energy decay for damped wave equations on partially rect-
angular domains, Math. Res. Letters 14 (2007), 35 -- 47.
[20] R. Chill and Yu. Tomilov, Stability of operator semigroups:
ideas and results, in:
Perspectives in operator theory, Banach Center Publ. Polish Acad. Sci. 75, Warsaw,
2007, 71 -- 109.
[21] H. Christianson, Applications of the cutoff resolvent estimates to the wave equation,
Math. Res. Letters 16 (2009), 577 -- 590.
[22] H. Christianson, E. Schenck, A. Vasy, and J. Wunsch, From resolvent estimates to
damped waves, J. Anal. Math., to appear.
[23] J.B. Conway, Functions of one complex variable, I. Second edition. Graduate Texts
in Mathematics, 11, Springer-Verlag, New York-Berlin, 1978.
78
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
[24] N. Dungey, On time regularity and related conditions for power-bounded operators,
Proc. London Math. Soc. 97 (2008), 97 -- 116.
[25] N. Dungey, Difference estimates for continuous and discrete operator semigroups,
Semigroup Forum 78 (2009), 226 -- 237.
[26] M. Eller and P. Toundykov, Carleman estimates for elliptic boundary value problems
with applications to the stabilization of hyperbolic systems, Evolution Equations and
Control Theory 1(2012), 271 -- 296.
[27] K.-J. Engel and R. Nagel, One-Parameter Semigroups for Linear Evolution Equa-
tions, Graduate Texts in Mathematics 194, Springer, New York, 2000.
[28] J. Esterle, E. Strouse, and F. Zouakia, Stabilit´e asymptotique de certains semi-
groupes d'op´erateurs et id´eaux primaires de L1(R+), J. Operator Theory 28 (1992),
203 -- 227.
[29] L. Fatori and R. Monteiro, The optimal decay rate for a weak dissipative Bresse
system, Appl. Math. Lett. 25 (2012), 600 -- 604.
[30] H. Fernandez Sare and J. Munoz Rivera, Optimal rates of decay in 2-d thermoelas-
ticity with second sound, J. Math. Phys. 53 (2012), 073509.
[31] A. Gomilko, M. Haase, and Yu. Tomilov, Bernstein functions and rates in mean
ergodic theorems for operator semigroups, J. Anal. Math. 118 (2012), 545 -- 576.
[32] M. Grobbelaar-van Dalsen, Strong stabilization of models incorporating the thermoe-
lastic Reissner-Mindlin plate equations with second sound, Appl. Anal. 90 (2011),
1419 -- 1449.
[33] M. Grobbelaar-van Dalsen, On the dissipative effect of a magnetic field in a Mindlin-
Timoshenko plate model, Z. Angew. Math. Phys. 63 (2012), 1047 -- 1065.
[34] M. Grobbelaar-van Dalsen, Stabilization of a thermoelastic Mindlin-Timoshenko
plate model revisited, Z. Angew. Math. Phys. 64 (2013), 1305 -- 1325.
[35] M. Haase, Spectral mapping theorems for holomorphic functional calculi, J. London
Math. Soc. (2) 71 (2005), 723 -- 739.
[36] M. Haase, The Functional Calculus for Sectorial Operators. Operator Theory: Ad-
vances and Applications 169, Birkhauser, Basel, 2006.
[37] J. Hao and Z. Liu, Stability of an abstract system of coupled hyperbolic and parabolic
equations, Z. Angew. Math. Phys. 64 (2013), 1145 -- 1159.
[38] Z. Harper, Laplace transform representations and Paley-Wiener theorems for func-
tions on vertical strips, Doc. Math. 15 (2010), 235 -- 254.
[39] E. Hille and R. S. Phillips, Functional Analysis and Semi-Groups, 3rd printing of
rev. ed. of 1957, Colloq. Publ. 31, Amer. Math. Soc., Providence, 1974.
[40] F. Hirsch, Transformation de Stieltjes et fonctions op´erant sur les potentiels ab-
straits, in: Th´eorie du potentiel et analyse harmonique (Journ´ees Soc. Math. France,
Inst. Recherche Math. Avanc´ee, Strasbourg, 1973), Lecture Notes in Math. 404,
Springer, Berlin, 1974, 149 -- 163.
[41] F. Hirsch, Int´egrales de r´esolvantes et calcul symbolique, Ann. Inst. Fourier (Greno-
ble) 22 (1972), 239 -- 264.
[42] F. Hirsch, Familles d'op´erateurs potentiels, Ann. Inst. Fourier (Grenoble) 25 (1975),
263 -- 288.
[43] F. Hirsch, Domaines d'op´erateurs repr´esent´es comme int´egrales de r´esolvantes, J.
Funct. Anal. 23 (1976), 199 -- 217.
[44] N. Kalton, S. Montgomery-Smith, K. Oleszkiewicz, and Yu. Tomilov, Power-
bounded operators and related norm estimates, J. London Math. Soc. 70 (2004),
463 -- 478.
[45] N. Kalton and P. Portal, Remarks on ℓ1 and ℓ∞-maximal regularity for power-
bounded operators, J. Aust. Math. Soc. 84 (2008), 345 -- 365.
[46] Y. Katznelson, An introduction to harmonic analysis, Wiley, New York, 1968.
[47] J. Korevaar, Tauberian Theory. A century of developments, Springer, Berlin, 2004.
FINE SCALES OF DECAY OF OPERATOR SEMIGROUPS
79
[48] P.C. Kunstmann and L. Weis, Maximal Lp-regularity for parabolic equations, Fourier
multiplier theorems and H ∞-functional calculus, in: Functional analytic methods for
evolution equations, Lecture Notes in Math. 1855, Springer, Berlin, 2004, 65 -- 311.
[49] M. Kwa´snicki, Spectral analysis of subordinate Brownian motions on the half-line,
Studia Math. 206 (2011), 211 -- 271.
[50] Yu. Latushkin and R. Shvydkoy, Hyperbolicity of semigroups and Fourier multi-
pliers, in: Systems, approximation, singular integral operators, and related topics
(Bordeaux, 2000), Oper. Theory Adv. Appl., 129, Birkhauser, Basel, 2001, 341 -- 363.
[51] G. Lebeau, ´Equation des ondes amorties, in: Algebraic and geometric methods in
mathematical physics (Kaciveli, 1993), Math. Phys. Stud. 19, Kluwer Acad. Publ.,
Dordrecht, 1996, 73 -- 109.
[52] G. Lebeau and L. Robbiano, Stabilisation de l'´equation des ondes par le bord, Duke
Math. J. 86 (1997), 465 -- 491.
[53] Z. Liu and B. Rao, Characterization of polynomial decay rate for the solution of
linear evolution equation, Z. Angew. Math. Phys. 56 (2005), 630 -- 644.
[54] C. Mart´ınez Carracedo and M. Sanz Alix, The theory of fractional powers of opera-
tors, North-Holland Mathematics Studies 187, North-Holland, Amsterdam, 2001.
[55] M. Mart´ınez, Decay estimates of functions through singular extensions of vector-
valued Laplace transforms, J. Math. Anal. Appl. 375 (2011), 196 -- 206.
[56] J. M. A. M. van Neerven, The asymptotic behaviour of semigroups of linear operators,
Operator Theory: Advances and Applications 88, Birkhauser, Basel, 1996.
[57] O. Nevanlinna, Convergence of iterations for linear equations, Lectures in Mathe-
matics, ETH Zurich, Birkhauser, Basel, 1993.
[58] S. Nicaise, Stabilization and asymptotic behavior of dispersive medium models, Sys-
tems Control Lett. 61 (2012), 638 -- 648.
[59] S. Nicaise and J. Valein, Stabilization of non-homogeneous elastic materials with
voids, J. Math. Anal. Appl. 387 (2012), 1061 -- 1087.
[60] N. Okazawa, Logarithms and imaginary powers of closed linear operators, Integral
Equations Operator Theory 38 (2000), 458 -- 500.
[61] E. I. Pustyl'nik, Some inequalities in the theory of self-adjoint operators, Izv.
Uchebn. Zaved. Mat. 1979, 52 -- 57 (in Russian).
[62] E. I. Pustyl'nik, Functions of a positive operator, Mat. Sb. (N.S.) 119 (1982), 32 -- 47
(in Russian), Math. USSR Sbornik 47(1984), 27 -- 42.
[63] M. L. Santosa, D. S. Almeida J´uniora, and J.E. Munoz Rivera, The stability number
of the Timoshenko system with second sound J. Diff. Eq. 253 (2012), 2715 -- 2733.
[64] R. Schilling, R. Song, and Z. Vondracek, Bernstein functions, de Gruyter Studies in
Mathematics 37, Walter de Gruyter, Berlin, 2010.
[65] E. Seneta, Regularly varying functions, Lecture Notes in Math. 508, Springer, Berlin,
1976.
[66] L. Tebou, Energy decay estimates for some weakly coupled Euler-Bernoulli and wave
equations with indirect damping mechanisms, Math. Control Relat. Fields 2 (2012),
45 -- 60.
[67] Vu Quoc Ph´ong, Theorems of Katznelson-Tzafriri type for semigroups of operators,
J. Funct. Anal. 103 (1992), 74 -- 84.
[68] D. V. Widder, The Laplace Transform, Princeton Mathematical Series 6, Princeton
University Press, Princeton, 1941.
80
CHARLES J.K. BATTY, RALPH CHILL, AND YURI TOMILOV
St. John's College, Oxford OX1 3JP, United Kingdom
E-mail address: [email protected]
Technische Universitat Dresden, Institut fur Analysis, 01062 Dresden, Ger-
many
E-mail address: [email protected]
Institute of Mathematics, Polish Academy of Sciences, ´Sniadeckich 8, 00-
956 Warszawa, Poland
E-mail address: [email protected]
|
1302.4054 | 3 | 1302 | 2013-05-20T08:12:18 | Universal conformal weights on Sobolev spaces | [
"math.FA"
] | The Riemann Mapping Theorem states existence of a conformal homeomorphism $\varphi$ of a simply connected plane domain $\Omega\subset\mathbb C$ with non-empty boundary onto the unit disc $\mathbb D\subset \mathbb C$. In the first part of the paper we study embeddings of Sobolev spaces $\overset{\circ}{W_{p}^{1}}(\Omega)$ into weighted Lebesgue spaces $L_{q}(\Omega,h)$ with an {}"universal" weight that is Jacobian of $\varphi$ i.e. $h(z):=J(z,\varphi)=| \varphi'(z)|^2$. Weighted Lebesgue spaces with such weights depend only on a conformal structure of $\Omega$. By this reason we call the weights $h(z)$ conformal weights. In the second part of the paper we prove compactness of embeddings of Sobolev spaces $\overset{\circ}{W_{2}^{1}}(\Omega)$ into $L_{q}(\Omega,h)$ for any $1\leq q<\infty$. With the help of Brennan's conjecture we extend these results to Sobolev spaces $\overset{\circ}{W_{p}^{1}}(\Omega)$. In this case $q$ is not arbitrary and depends on $p$ and the summability exponent for Brennan's conjecture. Applications to elliptic boundary value problems are demonstrated in the last part of the paper. | math.FA | math |
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
V.GOL'DSHTEIN AND A.UKHLOV
Abstract. The Riemann Mapping Theorem states the existence of a con-
formal homeomorphism ϕ of a simply connected plane domain Ω ⊂ C with
non-empty boundary onto the unit disc D ⊂ C. In the first part of the pa-
◦
p (Ω) into weighted Lebesgue
per, we study embeddings of Sobolev spaces
spaces Lq(Ω, h) with a "universal" weight that is the Jacobian of ϕ;
i. e.,
h(z) := J(z, ϕ) = ϕ′(z)2. Weighted Lebesgue spaces with such weights de-
pend only on the conformal structure of Ω. For this reason, we call the weights
h(z) conformal weights. In the second part of the paper, we prove compact-
W 1
◦
ness of embeddings of Sobolev spaces
2 (Ω) into Lq(Ω, h) for any 1 ≤ q < ∞.
With the help of Brennan's Conjecture, we extend these results to the Sobolev
W 1
◦
W 1
spaces
p (Ω). In this case, q depends on p and the integrability exponent for
Brennan's Conjecture. The last part of the paper is devoted to applications
to elliptic boundary value problems.
Key words and phrases: conformal mappings, Sobolev spaces, elliptic equa-
tions.
1. Introduction
Let Ω ⊂ C be an arbitrary simply connected plane domain with non-empty
boundary. By the Riemann Mapping Theorem, there exists a conformal homeo-
morphism ϕ of Ω onto the unit disc D ⊂ C.
This study is focused on the weighted Poincaré-Sobolev inequalities
(1.1)
1
r
(cid:18)
Ω
f (z)rh(z) dµ(cid:19)
≤ K(cid:18)
Ω
∇f (z)pdµ(cid:19)
1
p
◦
W 1
p (Ω) and a special weight h(z) := J(z, ϕ) =
for functions f of the Sobolev space
◦
ϕ′(z)2 induced by ϕ. Recall that
W 1
p (Ω) is closure of the set of all smooth functions
with compact support in Ω in the Sobolev space W 1
p (Ω) and J(z, ϕ) is Jacobian of
a conformal homeomorphisms ϕ : Ω → D at z ∈ Ω.
f (z) = f (x, y) is a real-valued function, ∇f (z) = ( ∂f
f , µ is the Lebesgue measure.
Here we have used the following notations: z = x + iy is a complex number,
∂y ) is the weak gradient of
∂x , ∂f
In the present paper, we give an essentially self-contained exposition of a method
for the study of such weighted inequalities and develop its applications to elliptic
boundary value problems.
The novel points include the use of a "transfer" scheme of Sobolev type embed-
ding theorems from regular domains to non-regular domains (proposed in [5]) in
combination with the Riemann Mapping Theorem and the Brennan's Conjecture.
1
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
2
The "transfer" scheme is based on systematic applications of the theory of com-
position operators on Sobolev spaces [20, 24].
In [9] this scheme was applied to
weighted Sobolev-type embedding operators in non-regular domains.
The Poincaré-Sobolev-type inequalities have essential applications in geometric
analysis (see, for example, [14]). The existence of the weighted inequalities (1.1) is
interesting even for bounded domains in the case of unbounded weights h(z). For
unbounded weights, the inequalities (1.1) contain additional information about the
boundary behavior of the functions f ∈
the Poisson problem −∆u = f , uP = g on a polyhedral domain P ⊂ R3.
The "transfer" scheme [5] is simplified here because of the well-known fact: For
any conformal homeomorphism w = ϕ(z) : Ω → Ω′ and any smooth function f
with square integrable derivatives, we have
Note, that such type inequalities have an application [1] in a regularity result for
◦
W 1
p (Ω).
Ω
∇(f ◦ ϕ)(z))2dµ =
∇f (w)2dµ.
Ω′
This equality means that ϕ induces an isometry of homogeneous Sobolev spaces
2(Ω′). It is one of the basic facts that makes it possible to "transfer"
2(Ω) and L1
L1
the Poincaré-Sobolev inequalities from the unit disc D ⊂ C to an arbitrary simply
connected plane domain with non-empty boundary Ω ⊂ C.
In the first part of the paper, we prove existence of bounded (compact) embed-
◦
W 1
2 (Ω) into weighted Lebesgue spaces Lq(Ω, h) with the
dings of the Sobolev space
conformal weight h(z) := J(z, ϕ) for any q ∈ [1,∞). Since two different conformal
homeomorphisms ϕ : Ω → D and ϕ : Ω → D can be connected by a conformal
automorphism η : D → D (i.e., ϕ = ϕ ◦ η), the conformal weights induced by ϕ and
ϕ are equivalent. It means that h(z) = J(z, ϕ) . J(z, ϕ) . J(z, ϕ) = h(z), and
therefore the weighted Lebesgue space Lq(Ω, h) does not depend on the choice of a
conformal homeomorphism and depends only on the conformal structure of Ω.
In the second part of the paper, we study the more complicated case of Sobolev
p (Ω), p 6= 2. Applying results of Brennan's Conjecture about the inte-
spaces
grability of the derivatives of conformal homeomorphisms ϕ : Ω → D, we prove
that such homeomorphisms induce bounded composition operators from the homo-
geneous Sobolev spaces L1
q(D) under some constraints on p and q that
are consequences of the results of Brennan's Conjecture. Another ingredient of this
study is necessary and sufficient conditions [20] for boundedness of the composition
operators from L1
q(D) induced by Sobolev homeomorphisms. This result
is rather general and is rearranged here for the conformal case.
p(Ω) into L1
◦
W 1
p(Ω) into L1
In the last part of the paper, we show a standard application of the main results
to boundary value problems for the Laplace operator.
We use a version of Brennan's Conjecture for composition operators on Sobolev
spaces [12] proposed recently by the authors. The original Brennan's Conjecture
concerns the integrability of derivatives of plane conformal homeomorphisms ϕ :
Ω → D that map a simply connected plane domain with non-empty boundary
Ω ⊂ C onto the unit disc D ⊂ C.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
3
The conjecture [3] is that
(1.2)
ϕ′(z)s dµ < +∞,
for all
4
3
< s < 4.
Ω
For 4/3 < s < 3, it is a comparatively easy consequence of the Koebe distortion
theorem (see, for example, [2]). J. Brennan [3] (1973) extended this range to 4/3 <
s < 3 + δ, where δ > 0, and conjectured it to hold for 4/3 < s < 4. The example
of Ω = C \ (−∞,−1/4] shows that this range of s cannot be extended. The upper
bound of those s for which (1.2) is known to hold has been increased to s ≤ 3.399
by Ch. Pommerenke, to s ≤ 3.421 by D. Bertilsson, and then to s ≤ 3.752 by
Hedenmalm and Shimorin (2005). These results and more information can be
found in [2, 17, 19].
For conformal homeomorphisms ψ : D → Ω, Brennan's Conjecture can be refor-
mulated as the Inverse Brennan's Conjecture
D
ψ′(w)α dµ < +∞,
for all − 2 < α < 2/3
where α = 2 − s.
The results of Inverse Brennan's Conjecture lead to the conjecture on the exis-
◦
W 1
q (D) for all 4/3 < p < 2
tence of bounded composition operators of
and all 1 ≤ q < 2p/(4 − p) (Theorem 4.5). As a corollary, we obtain a conjecture
about the existence of compact embeddings of
◦
W 1
p (Ω) into Lr(Ω, h) for all
◦
W 1
p (Ω) to
1 ≤ r <
p
2 − p
.
If α0 > −2 is the best known estimate in the Inverse Brennan's Conjecture, i. e.
the Inverse Brennan's Conjecture holds for any α ∈ [α0, 2
3 ), then
1 ≤ r ≤
2p
2 − p ·
α0
2 + α0
<
p
2 − p
is the best estimate for r for these embeddings [12].
A connection between Brennan's Conjecture and composition operators was es-
tablished in [12]:
Theorem 1.1. Equivalence Theorem. Brennan's Conjecture (1.2) holds for
a number s ∈ (4/3; 4) if and only if any conformal homeomorphism ϕ : Ω → D
induces a bounded composition operator
ϕ∗ : L1
p(D) → L1
q(p,s)(Ω)
for any p ∈ (2; +∞) and q(p, s) = ps/(p + s − 2).
Remark. Brennan's Conjecture is correct for some special classes of domains:
starlike domains, bounded domains which boundaries are locally graphs of contin-
uous functions etc.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
4
2. Notation and Preliminary Results about Composition Operators
We follow [13] for notation and basic facts about weighted Lebesgue spaces.
Let Ω ⊂ Rn be a domain and let v : Ω → R be a locally integrable almost
everywhere positive real valued function in Ω ( i.e v(z) > 0 almost everywhere).
Then a Radon measure ν is canonically associated with the weight function v:
ν(E) := E
v(z)dµ.
By the local integrability of v, the measure ν and the Lebesgue measure µ are
absolutely continuous with respect to one another:
dν = v(z)dµ.
In what follows, the weight v and the measure ν will be identified. The sets of
measure zero are the same for the Lebesgue measure µ and for ν. It means that we
do not need to specify what convergence almost everywhere is.
Denote by
V(Ω) := {v ∈ L1,loc(Ω) : v(z) > 0 a. e. on Ω}
the set of all such weights. Here L1,loc(Ω) is the space of locally integrable functions
in Ω.
For 1 ≤ p < ∞ and v ∈ V(Ω), consider the weighted Lebesgue space
Lp(Ω, v) :=(f : Ω → R : kf Lp(Ω, v)k :=(cid:18)Ω f (z)pdν(cid:19)1/p
< ∞) .
It is a Banach space for the norm kf Lp(Ω, v)k.
The space Lp(Ω, v) may fail to embed into L1,loc(Ω).
1
1−p ∈ L1,loc(Ω) and 1 < p < ∞ then the embedding
If v−1 ∈ L∞,loc(Ω) then the embedding operator i : L1(Ω, v) → L1,loc(Ω) is
Proposition 2.1. [13] If v
operator i : Lp(Ω, v) → L1,loc(Ω) is bounded.
bounded.
For 1 < p < ∞, we put
and for p = 1,
1
Vp(Ω ) :=nv ∈ V(Ω ) : v
1 −p ∈ L1 ,loc(Ω )o
V1 (Ω ) :=(cid:8)v ∈ V(Ω ) : v −1 ∈ L∞,loc(Ω )(cid:9).
Corollary 2.2. If a weight v is continuous and positive then i : Lp(Ω, v) →
L1,loc(Ω) is bounded.
This follows immediately from Proposition 2.1 because a continuous and positive
Define the Sobolev space W 1
weight belongs to Vp(Ω) and also to V1(Ω).
p (Ω), 1 ≤ p < ∞, as a normed space of locally
integrable weakly differentiable functions f : Ω → R equipped with the following
norm:
kf W 1
p (Ω)k =(cid:18)
Ω
f (z)p dµ(cid:19)1/p
+(cid:18)
Ω
∇f (z)p dµ(cid:19)1/p
.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
5
We will also need homogeneous seminormed Sobolev spaces L1
entiable functions f : Ω → R equipped with the following seminorms:
p(Ω) of weakly differ-
kf L1
p(Ω)k =(cid:18)
Ω
.
∇f (z)p dµ(cid:19)1/p
p(Ω) → L1,loc(Ω) is bounded.
Recall that the embedding operator i : L1
◦
W 1
The Sobolev space
smooth functions with compact supports C∞
p (Ω), 1 ≤ p < ∞, is defined as the closure of the space of
Let Ω and Ω′ be domains in C. We say that a conformal homeomorphism
0 (Ω) in the norm of W 1
p (Ω).
ϕ : Ω → Ω′ induces a bounded composition operator
ϕ∗ : L1
q(Ω), 1 ≤ q ≤ p ≤ ∞,
by the composition rule ϕ∗(f ) = f ◦ ϕ, if for any f ∈ L1
ϕ∗(f ) ∈ L1
q(Ω) and there exists a constant K < ∞ such that
p(Ω′) → L1
p(Ω′), the composition
kϕ∗(f ) L1
q(Ω)k ≤ Kkf L1
p(Ω′)k.
n(Ω) and L1
The theory of composition operators on Sobolev spaces goes back to Reshetnyak's
problem about the description of all isomorphisms of the homogeneous Sobolev
n(Ω′) which are induced by quasiconformal mappings of Eu-
spaces L1
clidean domains. In [22], it was proved that a homeomorphism ϕ : Ω → Ω′ between
Euclidean domains Ω ⊂ Rn and Ω′ ⊂ Rn induces an isomorphism of L1
n(Ω) and
n(Ω′) by the composition rule ϕ∗(f ) = f ◦ ϕ if and only if ϕ is quasiconformal.
L1
In the framework of this approach to geometric function theory, there appears
the problem about the description of homeomorphisms inducing isomorphisms of
p (Ω) and W 1
Sobolev spaces by the composition rule. The Sobolev spaces W 1
p (Ω′),
p > n, were considered in [23], the Sobolev spaces W 1
p (Ω′), n− 1 < p <
p (Ω) and W 1
n, were treated in [8], and Sobolev spaces W 1
p (Ω′), 1 ≤ p < n, were
considered in [16]. In [15], the theory of multipliers was applied to this composition
problem. Bounded composition operators on Sobolev spaces were studied in [21]
from another (but close) point of view. A geometric description of homeomorphisms
preserving the Sobolev spaces L1
p(Ω) was obtained in [6] for p > n − 1.
New problems arise when we study composition operators on Sobolev spaces
with decreasing integrability of the first weak derivatives. This problem was first
studied in [20]. In this case, a significant role in the description of the composition
operators is played by the so-called (quasi)additive set functions defined on open
sets.
p (Ω) and W 1
p(Ω′) and L1
The main result of [20] asserts that
Theorem 2.3. A homeomorphism ϕ : Ω → Ω′ between two domains Ω, Ω′ ⊂ Rn
induces a bounded composition operator
p(Ω′) → L1
q(Ω), 1 ≤ q < p < ∞,
ϕ∗ : L1
if and only if ϕ ∈ W 1
1,loc(Ω), has finite distortion, and
Kp,q(f ; Ω) =(cid:18)
Ω (cid:18)Dϕ(x)p
J(x, ϕ)(cid:19)
q
p−q
p−q
pq
dµ(cid:19)
< ∞.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
6
Here Dϕ(x) is the formal Jacobi matrix of ϕ at x ∈ Ω and J(x, ϕ) = det Dϕ(x)
is its Jacobian. The norm Dϕ(x) of the matrix is the norm of the linear operator
defined by this matrix in the Euclidean space Rn. In [5], this class of mappings was
studied in connection with the Sobolev-type embedding theorems.
The detailed study of composition operators on Sobolev spaces was carried out
in [24]. The composition operators on Sobolev spaces in the limit case p = ∞ were
studied in [10, 11].
Define the p-dilatation of a diffeomorphism ϕ : Ω → Ω′ as
Kp(x, ϕ) = Dϕ(x)p
J(x, ϕ)
and introduce the (p, q)-dilatation as
Kp,q(f ; Ω) =(cid:18)
Ω (cid:18)Kp(x, ϕ)(cid:19)
q
p−q
p−q
pq
.
dµ(cid:19)
For plane conformal homeomorphisms (n = 2), the p-dilatation is equal to
ϕ′(z)p−2 for any p ∈ [1,∞). Of course, for p = 2 the 2-dilatation is the clas-
sical conformal dilatation ϕ′(z))2/J(z, ϕ) and is equal to the unity for conformal
mappings.
3. Weighted Poincaré-Sobolev Inequalities for the Sobolev Space
◦
W 1
2 (Ω)
In this section, we study the weighted Poincaré-type inequalities for conformal
weights.
First we formulate a well known property of conformal homeomorphisms:
Lemma 3.1. Let Ω and Ω′ be two plane domains. Any conformal homeomorphism
w = ϕ(z) : Ω → Ω′ induces an isometry of spaces L1
Proof. Let f ∈ L1
belongs to L1
2(Ω′) be a smooth function. Then the smooth function g = f ◦ ϕ
2(Ω′) and L1
2(Ω).
2(Ω) because
1
2
Ω
k∇g L2(Ω)k =(cid:18)
∇(f ◦ ϕ(z))2 dµ(cid:19)
∇f2(ϕ(z))J(z, ϕ) dµ(cid:19)
∇f2(ϕ(z))ϕ′(z))2 dµ(cid:19)
∇f2(w) dµ(cid:19) = k∇f L2(Ω′)k.
We used the following conformal equality: ϕ′(z))2 = J(ϕ, z) for every z ∈ Ω.
2(Ω′) by smooth functions, we obtain
(cid:3)
Approximating an arbitrary function f ∈ L1
an isometry between L1
=(cid:18)
=(cid:18)
=(cid:18)
2(Ω′) and L1
2(Ω).
Ω
1
2
1
2
Ω
Ω′
Definition 3.2. Let Ω ⊂ C be a simply connected domain with non-empty bound-
ary and let ϕ be a conformal homeomorphism of Ω onto the unit disc D. We call
the smooth positive real-valued function h(z) = J(z, ϕ) = ϕ′(z)2 the universal
conformal weight in Ω (or simply the conformal weight).
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
7
Recall that the Lebesgue spaces Lp(Ω, h) does not depend on the choice of the
conformal homeomorphism ϕ; i.e. the dependence is only on the conformal struc-
ture of Ω. This is a reason to call the weight h(z) the (universal) conformal weight
on Ω.
Theorem 3.3. Let Ω ⊂ C be a simply connected domain with non-empty boundary.
Then the inequality
holds for every function f ∈ C∞
depending only on r.
0 (Ω) and for any 1 ≤ r < ∞. Here K is a constant
kf Lr(Ω, h)k ≤ Kk∇f L2(Ω)k
Remark. The constant K is equal to the exact constant for the corresponding
Poincaré-Sobolev inequality in the unit disc D ⊂ C, i. e., for
kf Lr(D)k ≤ Kk∇f L2(D)k, f ∈ C∞
0 (Ω). By the Riemann Mapping Theorem, there exists a confor-
0 (D)
Proof. Let f ∈ C∞
mal homeomorphism ϕ : Ω → D. Then the function g = f ◦ ϕ−1 belongs to C∞
and, by Lemma 3.1, k∇g L2(D)k = k∇f L2(Ω)k.
Using the Poincaré-Sobolev inequality for the function g in the disc D we infer
0 (D).
kf Lr(Ω, h)k =(cid:18)
Ω
f (z)rJ(z, ϕ) dµ(cid:19)
1
r
=(cid:18)
D
f ◦ ϕ−1(w)r dµ(cid:19)
1
r
= kg Lr(D)k ≤ Kk∇g L2(D)k = Kk∇f L2(Ω)k.
(cid:3)
Now we give some examples of conformal weights.
Example 3.4. Let Ωpl = C \ D = {z ∈ C : x2 + y2 > 1} be the plane without the
unit disc. The diffeomorphism
is conformal and maps Ωpl onto the unit disc D. The conformal weight is
w = ϕ(z) =
1
z
, z = x + iy,
Example 3.5. Let Ωh = C+ = {z ∈ C : y > 0} be the upper half-plane. The
diffeomorphism
is conformal and maps Ωh onto the unit disc D. Then the conformal weight is
h(z) =
1
z22 =
1
(x2 + y2)2 .
w = ϕ(z) =
, z = x + iy,
z − i
z + i
h(z) =
4
z + i4 =
4
(x2 + (y + 1)2)2 .
Example 3.6. Let Ωs = {z ∈ C : − π
4 < Re z < π
4} be a strip. The diffeomorphism
w = ϕ(z) =
= tan z, z = x + iy,
1
i
e2iz − 1
e2iz + 1
is conformal and maps Ω onto the unit disc D. Then the conformal weight is
h(z) =
1
z2 + 12 =
1
(x2 + y2)2 + x2 − y2 + 1
.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
8
Example 3.7. Let Ωc be the interior of the cardioid: r = 1
diffeomorphism
2 (1 + cos θ). The
w = ϕ(z) = √z − 1, z = x + iy,
is conformal and maps Ωc onto the unit disc D. Then the conformal weight
h(z) =
=
1
2pz
1
.
2 4px2 + y2
4. Embedding into Lebesgue Spaces with Conformal Weights
In this section, we prove existence of compact embeddings of the Sobolev spaces
◦
W 1
p (Ω) into the Lebesgue spaces Lr(Ω, h) with the (universal) conformal weights
h. We begin with a general fact about Sobolev spaces
Theorem 4.1. Let ϕ : D → Ω be a conformal homeomorphism. Then the compo-
sition operator
◦
W 1
2 (Ω).
is bounded.
ϕ∗ :
◦
W 1
2 (Ω) →
◦
W 1
2 (D)
◦
W 1
2 (Ω) be a smooth function. We first prove that ϕ∗f belongs to
Proof. Let f ∈
◦
W 1
2 (D).
By the definition of the composition operators ϕ∗f = f ◦ ϕ is a smooth function
defined everywhere in Ω and by Lemma 3.1, we have the equality
Since the support supp(f ) of f is compact, its inverse image ϕ−1(supp(f )) is
k∇(ϕ∗f )L2(D)k = k∇fL2(Ω)k.
also compact. Hence, we obtain the following estimate of kϕ∗fL2(D)k:
kϕ∗fL2(D)k =(cid:18)
f ◦ ϕ2 dµ(cid:19)
1
2
ϕ−1(supp(f ))
=(cid:18)
ϕ−1(supp(f ))
f ◦ ϕ2J(z, ϕ)
1
J(z, ϕ)
1
2
dµ(cid:19)
≤
z∈ϕ−1(supp(f ))(cid:18)
max
J(z, ϕ)
1
1
2(cid:19)(cid:18)
z∈ϕ−1(supp(f ))(cid:18)
max
≤
ϕ−1(supp(f ))
1
2
f ◦ ϕ2J(z, ϕ) dµ(cid:19)
2(cid:19)(cid:18)
supp(f )
1
1
2
f2(w) dµ(cid:19)
J(z, ϕ) 1
Using the notation
we conclude that
Q(ϕ, f ) :=
1
max
J(z, ϕ) 1
z∈ϕ−1(supp(f ))(cid:18)
2(cid:19).
kϕ∗fL2(D)k ≤ Q(ϕ, f )kfL2(Ω)k.
◦
W 1
2 (D).
So, the composition ϕ∗f belongs to the Sobolev space
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
9
Let us prove that the composition operator ϕ∗ is bounded.
By the Poincaré inequality
kϕ∗fL2(D)k ≤ Ck∇(ϕ∗f )L2(D)k
◦
W 1
2 (D), here the constant C does not depends on f .
for every function g = ϕ∗f ∈
Hence,
kϕ∗f
◦
W 1
2 (D)k = kϕ∗fL2(D)k + k∇(ϕ∗f )L2(D)k
≤ Ck∇(ϕ∗f )L2(D)k + k∇(ϕ∗f )L2(D)k
= (C + 1)k∇fL2(Ω)k ≤ (C + 1)kf
◦
W 1
2 (Ω)k
for every smooth function f ∈
◦
W 1
2 (Ω).
Using the density of smooth functions with compact supports in
◦
W 1
2 (Ω), we can
◦
W 1
2 (Ω) (see, for example,
extend the last inequality to an arbitrary function f ∈
[24, 7]). This means that ϕ∗f ∈
◦
W 1
2 (D) and the composition operator
◦
W 1
◦
W 1
2 (D)
ϕ∗ :
2 (Ω) →
is bounded.
(cid:3)
The previous theorem leads to compactness of the Sobolev type embeddings
◦
W 1
jr :
2 (Ω) ֒→ Lr(Ω, h) in the case of the universal conformal weight h:
Theorem 4.2. Let Ω ⊂ C be a simply connected domain with non-empty boundary.
Then the embedding operator
jr :
◦
W 1
2 (Ω) ֒→ Lr(Ω, h)
is compact for any 1 ≤ r < ∞. Here h is the universal conformal weight.
Proof. By the Riemann Mapping Theorem, there exists a conformal homeomor-
phism w = ϕ(z) : Ω → D. The inverse mapping is also conformal and by, the
previous theorem, there exists a constant 0 < K < ∞ such that
◦
W 1
2 (D)k ≤ Kkf
◦
W 1
2 (Ω)k
kf ◦ ϕ−1
◦
W 1
2 (Ω).
for any function f ∈
◦
W 1
By the classical Sobolev embedding theorem for the unit disc D, the embedding
operators ir :
2 (D) ֒→ Lr(D) are bounded and compact for any 1 ≤ r < ∞.
Using the change of variable formula and the boundedness of ir, we obtain
kf Lr(Ω, h)k =(cid:18)
Ω
f (z)rJ(z, ϕ) dµ(cid:19)
1
r
= kf ◦ ϕ−1 Lr(D)k ≤ kirkkf ◦ ϕ−1
=(cid:18)
D
◦
W 1
1
r
f ◦ ϕ−1(w)r dµ(cid:19)
2 (D)k ≤ kirk · K · kf
◦
W 1
2 (Ω)k.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
10
Therefore, the embedding operators jr :
◦
W 1
positions of the compact embedding operators ir :
composition operators.
2 (Ω) ֒→ Lr(Ω, h) are compact as com-
2 (D) ֒→ Lr(D) and bounded
(cid:3)
◦
W 1
The next theorem about composition operators was formulated and proved in
[20, 24] for Sobolev homeomorphisms. In the case of conformal homeomorphisms,
this statement is much simpler. Here we reproduce a comparatively simple proof
for the conformal case that enables us to avoid the main technicalities. As prelim-
inary information, this proof uses only the result about the existence of a bounded
monotone countably additive function [20, 24]:
Theorem 4.3. Suppose that a mapping ϕ : Ω → Ω′ induces a bounded composition
operator
Then
q(Ω),
ϕ∗ : L1
p(Ω′) → L1
p(A′)∩C0(A′) (cid:13)(cid:13)ϕ∗f L1
q(Ω)(cid:13)(cid:13)
(cid:13)(cid:13)f L1
p(A′)(cid:13)(cid:13)
where the number κ is defined by 1/κ = 1/q− 1/p, is a bounded monotone countably
additive function defined on open bounded subsets A′ ⊂ Ω′.
1 ≤ q < p ≤ ∞.
!κ
Φ(A′) =
sup
f ∈L1
,
Using this "localization principle" for composition operators, we prove
Theorem 4.4. Let Ω, Ω′ ⊂ C be plane domains. A conformal homeomorphism
w = ϕ(z) : Ω → Ω′ induces a bounded composition operator ϕ∗ : L1
q(Ω),
1 ≤ q < p < ∞, if and only if
p(Ω′) → L1
(cid:18)
Ω
ϕ′(z)
(p−2)q
p−q dµ(cid:19)
p−q
pq
= K < +∞.
Proof. Necessity. Suppose that the composition operator
q(Ω), 1 ≤ q < p < ∞,
p(Ω′) → L1
ϕ∗ : L1
is bounded. Then by Theorem 4.3 there exists a bounded monotone countably
additive function Φ defined on open bounded subsets A′ ⊂ Ω′ such that, for every
function f ∈ L1
(4.1)
Fix a cut function η ∈ C∞
and is equal to zero outside of the set {w ∈ C : w < 2}. Inserting the functions
0 (C) which is equal to one on the set {w ∈ C : w < 1}
p(Ω′) ∩ C0(A′),
kϕ∗(f ) L1
q(Ω)k ≤(cid:0)Φ(A′)(cid:1)
pq kf L1
p(Ω)k.
p−q
fR(w) = Re(w − w0)η(cid:18) w − w0
fI (w) = Im(w − w0)η(cid:18) w − w0
r
r
(cid:19), w0 ∈ Ω′,
(cid:19), w0 ∈ Ω′,
and
in this inequality, we obtain
(cid:18)
ϕ−1(B(w0,r))
1
q
ϕ′(z)q dµ(cid:19)
≤ CΦ(B(w0, 2r))
p−q
pq B(w0, r)
1
p
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
11
when B(w0, r) = {w ∈ C : w − w0 < r}, B(w0, 2r) ⊂ Ω′.
By the change of variable formula, using the equality ϕ′(z)2 = J(z, ϕ), we infer
ϕ′(z)q dµ =
ϕ′(z)q−2J(z, ϕ) dµ
ϕ−1(B(w0,r))
ϕ−1(B(w0,r))
=
B(w0,r)
ϕ′(ϕ−1(w))q−2 dµ.
Hence,
(cid:18)
B(w0,r)
ϕ′(ϕ−1(w))q−2 dµ(cid:19)
1
q
≤ CΦ(B(w0, 2r))
p−q
pq B(w0, r)
1
p
and we have
1
B(w0, r)
B(w0,r)
B(w0, r) (cid:19)
ϕ′(ϕ−1(w))q−2 dµ ≤ C(cid:18) Φ(B(w0, 2r)
p−q
p
.
Passing to the limit as r → 0, we get
Hence, for every open bounded subset V ⊂ Ω′
ϕ′(ϕ−1(w))q−2 ≤ C(cid:0)Φ′(w)(cid:1)
p−q dµ ≤ C
ϕ′(ϕ−1(w))
p(q−2)
p−q
p
for almost all w ∈ Ω′.
Φ′(w) dµ ≤ CΦ(V ) ≤ Ckϕ∗k
pq
p−q .
V
V
Since V is an arbitrary subset of Ω′ we have that
ϕ′(ϕ−1(w))
p(q−2)
p−q dµ ≤ Ckϕ∗k
pq
p−q .
Ω′
Therefore
Ω
ϕ′(z)
(p−2)q
p−q dµ =
ϕ′(z)
p(q−2)
p−q J(z, ϕ) dµ
Ω
=
Ω′
ϕ′(ϕ−1(w))
p(q−2)
p−q dµ ≤ Ckϕ∗k
pq
p−q .
Sufficiency. Let f ∈ L1
p(Ω′). Then, since ϕ is a smooth mapping, the compo-
sition ϕ∗(f ) = f ◦ ϕ is defined almost everywhere and weakly differentiable in Ω.
Hence, using the equality ϕ′(z)2 = J(z, ϕ), we obtain
=(cid:18)
∇(f ◦ ϕ(z))q dµ(cid:19)
q(Ω)k =(cid:18)
kϕ∗(f ) L1
1
q
1
q
Ω
Ω
∇fq(ϕ(z))ϕ′(z)q dµ(cid:19)
dµ(cid:19)
p ϕ′(z)
(p−2)q
p
q
1
q
.
=(cid:18)
Ω
∇fq(ϕ(z))J(z, ϕ)
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
12
Thus, using Hölder's inequality, we have
∇fq(ϕ(z))J(z, ϕ)
kϕ∗(f ) L1
≤(cid:18)
q(Ω)k =(cid:18)
∇fp(ϕ(z))J(z, ϕ) dµ(cid:19)
Ω
Ω
(p−2)q
p
q
p ϕ′(z)
1
q
dµ(cid:19)
1
p(cid:18)
Ω
p−q
pq
ϕ′(z)
(p−2)q
p−q dµ(cid:19)
= Kkf L1
p(Ω′)k.
(cid:3)
Theorem 4.5. Let Ω ⊂ C be a simply connected domain with non-empty bound-
ary and ϕ : Ω → D be a conformal homeomorphism. Suppose that the Inverse
Brennan's Conjecture holds for the interval [α0, 2/3) where α0 ∈ (cid:0) − 2, 0(cid:1) and
p ∈(cid:0)(α0 + 2)/(α0 + 1), 2(cid:1).
Then the inverse mapping ϕ−1 induces a bounded composition operator
for any q such that
(ϕ−1)∗ : L1
p(Ω) → L1
q(D)
1 ≤ q ≤ p α0/(2 + α0 − p) < 2p/(4 − p).
Proof. By the Inverse Brennan's Conjecture
D
(ϕ−1)′(w)α dµ < +∞,
for all − 2 < α < 2/3.
Choose α ∈ (−2, 0). By Theorem 4.4, the composition operator
p(Ω) → L1
q(D)
(ϕ−1)∗ : L1
is bounded for all 1 ≤ q < p < 2 such that
(p − 2)q
p − q
The inequality
(p − 2)q
p − q
holds for all 1 < p < 2 and 1 ≤ q < p < 2.
Because p < 2 the inequality
= α.
< 0
is equivalent to
−2 < α =
(p − 2)q
p − q
By Theorem 4.4 we have the restriction on q: q ≥ 1. Hence
q <
< p.
2p
4 − p
2p
4 − p
> 1, i. e. p >
4
3
.
It is the best possible estimate on p under the assumption that the Inverse
Brennan's conjecture is correct. Because the conjecture is not proved completely,
we can suppose in the statement of this Theorem that it holds for some −2 < α0 < 0.
In this case restrictions on p will depend on α0.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
13
The inequality
is equivalent to
α0 ≤
1 ≤ q ≤
Because the function u(α0) = α0p
p < 2 we obtain
(p − 2)q
p − q
α0p
.
2 + α0 − p
2+α0−p is an increasing function, α0 < 2 and
α0p
< p < 2.
<
q ≤
2 + α0 − p
2p
4 − p
Let us check for which numbers p the condition q ≥ 1 is correct, i. e.
numbers p the following inequality
for which
holds. By simple calculations we obtain that it holds for
1 <
α0p
2 + α0 − p
p > α0 + 2
α0 + 1
>
4
3
.
Two previous inequalities permit us to conclude that for any fixed p ∈ ((α0 + 2)/(α0 + 1), 2)
and for any q such that
1 ≤ q ≤
α0p
2 + α0 − p
<
2p
4 − p
< 2
the inequality
k∇(f ◦ ϕ−1) Lq(D)k ≤ Kk∇f Lp(Ω)k
p(Ω).
holds for every function f ∈ L1
(cid:3)
Proposition 4.6. Suppose that Ω ⊂ C is a simply connected domain with non
empty boundary, the Inverse Brennan's Conjecture holds for the interval [α0, 2/3)
where α0 ∈ (cid:0) − 2, 0(cid:1) and h(z) = J(z, ϕ) is the conformal weight defined by a
conformal homeomorphism ϕ : Ω → D. Then for every (α0 + 2)/(α0 + 1) < p <
2 and every function f ∈ C∞
0 (Ω), the inequality
holds for any r such that
kf Lr(Ω, h)k ≤ Mk∇f Lp(Ω)k
1 ≤ r ≤
2p
2 − p ·
α0
2 + α0
<
p
2 − p
.
The constant M does not depend on f .
Proof. By the Riemann Mapping Theorem, there exists a conformal homeomor-
phism w = ϕ(z) : Ω → D, and by the Inverse Brennan's Conjecture,
(ϕ−1)′(w)α dw < +∞,
for all − 2 < α0 < α < 2/3.
D
Hence, by Theorem 4.5, the inequality
holds for every function f ∈ L1
k∇(f ◦ ϕ−1) Lq(D)k ≤ Kk∇f Lp(Ω)k
p(Ω) and for any q such that
1 ≤ q ≤ p α0/(2 + α0 − p) < 2p/(4 − p).
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
14
Choose arbitrarily f ∈ C∞
Poincaré-Sobolev inequality,
0 (Ω). Then f ◦ ϕ−1 ∈ C∞
0 (D) and, by the classical
(4.2)
kf ◦ ϕ−1 Lr(D)k ≤ Ak∇(f ◦ ϕ−1) Lq(D)k
for any r such that
1 ≤ r ≤
2q
2 − q
Combining inequalities for q and r we conclude that the inequality (4.2) holds for
any r such that
2 − p
Using the change of variable formula, we finally infer
2 + α0
1 ≤ r ≤
2p
2 − p ·
α0
<
p
.
kf Lr(Ω, h)k =(cid:18)
Ω
f (z)rJ(z, ϕ) dµ(cid:19)
1
r
=(cid:18)
D
f (ϕ−1(w))r dµ(cid:19)
1
r
= kf ◦ ϕ−1 Lr(D)k ≤ Ak∇(f ◦ ϕ−1) Lq(D)k ≤ AKk∇f Lp(Ω)k.
(cid:3)
Theorem 4.7. Let Ω ⊂ C be a simply connected domain with non-empty bound-
ary and ϕ : Ω → D be a conformal homeomorphism. Suppose that the Inverse
Brennan's Conjecture holds for the interval [α0, 2/3) where α0 ∈ (cid:0) − 2, 0(cid:1) and
p ∈(cid:0)(α0 + 2)/(α0 + 1), 2(cid:1).
Then the inverse mapping ϕ−1 induces a bounded composition operator
(ϕ−1)∗ :
◦
W 1
p (Ω) →
◦
W 1
q (D)
for any q such that
1 ≤ q ≤ p α0/(2 + α0 − p) < 2p/(4 − p).
Remark. Of course, we can choose the best known estimate α0 > −2 in the
Inverse Brennan's Conjecture.
Proof. By the Inverse Brennan's Conjecture, we have
D
(ϕ−1)′(w)α dµ < +∞,
for all − 2 < α0 < α < 2/3.
◦
W 1
p (Ω) be a smooth function with a compact support. By Theorem 4.5,
Let f ∈
for any function f ∈ C∞
0 (Ω), the inequality
kf ◦ ϕ−1 L1
q(D)k ≤ Ckf L1
p(Ω)k
holds for all q : 1 ≤ q ≤ p α0/(2 + α0 − p) < 2p/(4 − p) < 2. Here the constant
0 < C < ∞ is independent of f .
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
15
0 (Ω), its supp(f ) is compact and its image ϕ(supp(f )) is compact
Since f ∈ C∞
too. The following estimate holds for kf ◦ ϕ−1Lq(D)k:
kf ◦ ϕ−1Lq(D)k =(cid:18)
f ◦ ϕ−1q dµ(cid:19)
1
q
ϕ(supp(f ))
=(cid:18)
ϕ(supp(f ))
f ◦ ϕ−1qJ(w, ϕ−1)
q
p
1
J(w, ϕ−1)
q
p
1
q
dµ(cid:19)
≤(cid:18)
ϕ(supp(f ))
J(w, ϕ−1)
q
q−p dµ(cid:19)
Putting
p−q
pq (cid:18)
ϕ(supp(f ))
f ◦ ϕ−1pJ(w, ϕ−1) dµ(cid:19)
1
p
.
Q(ϕ, f ) :=(cid:18)
ϕ(supp(f ))
J(w, ϕ−1)
p−q
pq
,
q
q−p dµ(cid:19)
we conclude that
kf ◦ ϕ−1Lq(D)k ≤ Q(ϕ, f )kfLp(Ω)k.
Thus, the function f ◦ ϕ−1 belongs to the Sobolev space
Poincaré-Sobolev inequality for the unit disc
◦
W 1
q (D). Using the
kf ◦ ϕ−1 Lq(D)k ≤ Ak∇(f ◦ ϕ−1) Lq(D)k
we obtain the inequality
kf ◦ ϕ−1 W 1
q (D)k = kf ◦ ϕ−1 Lq(D)k + k∇(f ◦ ϕ−1) Lq(D)k
≤ (A + 1)k∇(f ◦ ϕ−1) Lq(D)k ≤ (A + 1)Ckf L1
p(Ω)k ≤ Kkf W 1
p (Ω)k
that holds for every f ∈ C∞
0 (Ω).
◦
W 1
p (Ω), we
Using the density of smooth functions with compact supports in f ∈
can extend the last inequality to arbitrary f ∈
It allows us to conclude finally that the composition operators
◦
W 1
p (Ω) (see, for example, [24, 7]).
(ϕ−1)∗ :
◦
W 1
p (Ω) →
◦
W 1
q (D)
are bounded for all q satisfying the conditions 1 ≤ q ≤ p α0/(2 + α0 − p) <
2p/(4 − p).
(cid:3)
Theorem 4.8. Suppose that Ω ⊂ C is a simply connected domain with non empty
boundary, the Inverse Brennan's Conjecture holds for the interval [α0, 2/3) where
α0 ∈ (cid:0) − 2, 0(cid:1), p ∈ (cid:0)(α0 + 2)/(α0 + 1), 2(cid:1) and h(z) = J(z, ϕ) is the conformal
weight defined by a conformal mapping ϕ : Ω → D.
Then the embedding operator
is compact for every r such that
jr :
◦
W 1
p (Ω) ֒→ Lr(Ω, h)
1 ≤ r <
2p
2 − p ·
α0
2 + α0
<
p
2 − p
.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
16
Remark. Of course, we can choose the best known estimate α0 > −2 in the
Inverse Brennan's Conjecture.
◦
W 1
Proof. Choose an arbitrary function f ∈
a conformal homeomorphism. By the previous theorem, (ϕ−1)∗f ∈
1 ≤ q ≤ p α0/(2 + α0 − p) < 2p/(4 − p) < 2 and
q (D)k ≤ Kkf
p (Ω). Let w = ϕ(z) : Ω → D be
◦
W 1
q (D) for
k(ϕ−1)∗f
p (Ω)k
◦
W 1
◦
W 1
where K < ∞ does not depend on f . By the classical embedding theorem, the
function (ϕ−1)∗f ∈ Lr(D) for r : 1 ≤ r < 2p α0/(2 α0 + 4 − p(2 + α0)) <
p/(2 − p) and the inequality
k(ϕ−1)∗f Lr(D)k ≤ Ak(ϕ−1)∗f
◦
W 1
q (D). Using the change of variable formula, we obtain
q (D)k
◦
W 1
1
r
f (z)rJ(z, ϕ) dµ(cid:19)
=(cid:18)
D
f (ϕ−1(w))r dµ(cid:19)
1
r
holds for any f ∈
kf Lr(Ω, h)k =(cid:18)
Ω
= k(ϕ−1)∗f Lr(D)k ≤ Ak(ϕ−1)∗f
Consequently, the embedding operator
◦
W 1
q (D)k ≤ AKkf
◦
W 1
p (Ω)k.
is compact as the composition of the bounded composition operator
◦
W 1
jr :
p (Ω) ֒→ Lr(Ω, h)
compact embedding operator
(ϕ−1)∗ :
◦
W 1
p (Ω) →
◦
W 1
1 (D),
and the bounded composition operator
◦
W 1
ir :
q (D) ֒→ Lr(D)
ϕ∗ : Lr(D) → Lr(Ω, h).
(cid:3)
5. Applications to Elliptic Equations
As a standard application, we prove the solvability of the classical Dirichlet
problem for the degenerate Laplace operator on an arbitrary simply connected
plane domain Ω ⊂ C with non-empty boundary.
p (Ω, h, 1), 1 ≤ p < ∞, as the normed space
of locally integrable weakly differentiable functions f : Ω → R equipped with the
following norm:
Define the weighted Sobolev space W 1
kf W 1
p (Ω)k =(cid:18)
Ω
fp(z) h(z) dµ(cid:19)1/p
+(cid:18)
Ω
∇f (z)p dµ(cid:19)1/p
.
Here h is the universal conformal weight.
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
17
C∞
◦
W 1
Such type weighted spaces were introduced and investigated in [7] for Vp-weights.
The Sobolev space
p (Ω, h, 1), 1 ≤ p < ∞, is defined as the closure of the space
0 (Ω) of smooth functions with compact support in the norm of W 1
We shall use a short notation for the inner products: < u, v >:= ´Ω u(z)v(z)dµ
in L2(Ω) and < u, v >h:= ´Ω u(z)v(z)h(z)dµ in L2(Ω, h) and also the notation
[u, v] :=< ∇u,∇v >.
p (Ω, h, 1).
The problem is as follows:
(5.1)
(5.2)
∆u = f h in Ω,
u∂Ω = 0.
The weak statement of this Dirichlet problem is as follows:
A function u solves the previous problem iff u ∈
◦
W 1
2 (Ω, h, 1) and
[u, v] =< ∇u,∇v >= Ω
f (z)v(z)h(z)dµ
◦
W 1
2 (Ω, h, 1), under provided that this integral makes sense.
for all v ∈
We prove that, for any simply connected plane domain with non-empty boundary
and for any f ∈ Lp(Ω, h), 1 < p < ∞, there exists a unique (weak) solution to the
problem (5.1,5.2).
◦
W
Theorem 5.1. Let Ω be a simply connected plane domain with non-empty boundary
and let 1 < p < ∞. If f ∈ Lp(Ω, h) then there exists the unique weak solution
u ∈
Proof. The function f ∈ Lp(Ω, h) induces a linear functional F : C∞
the standard rule
1
2(Ω, h, 1) of the problem (5.1,5.2).
0 (Ω) → R by
F (v) =
Ω
f (z)v(z)h(z) dµ, v ∈ C∞
0 (Ω).
By Theorem 3.3:
The last inequality shows that the norm [u, u]
kvL2(Ω, h)k ≤ Kk∇vL2(Ω)k.
1
to the W 1
2 (Ω, h, 1)-norm on
◦
W
induces a Hilbert structure on
2 := k∇uL2(Ω)k is equivalent
1
2(Ω, h, 1) and the corresponding inner product [u, v]
◦
W 1
2 (Ω, h, 1).
◦
W
1
2(Ω, h, 1), we show that F is
Using Theorem 3.3 and the density of C∞
0 (Ω) in
◦
W 1
2 (Ω, h, 1).
0 (Ω) by Theorem 3.3 we have
a bounded linear functional in
For any v ∈ C∞
F (v) ≤ kf · v L1(Ω, h)k
≤ kf Lp(Ω, h)k · kv Lp′(Ω, h)k ≤ Ckf Lp(Ω, h)k · kv
◦
W
1
2(Ω, h, 1)k,
Therefore, the bounded functional F defined on the dense subset C∞
0 (Ω) can be
for any 1 ≤ p < ∞, p′ = p/(p − 1).
extended to
◦
W
1
2(Ω, h, 1).
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
18
Hence, by the Riesz theorem about linear functionals in a Hilbert space, [u, v] =
◦
W
1
2(Ω, h, 1), where B is a bounded linear operator from Lp(Ω, h)
[Bf, v] for all v ∈
into
◦
W
1
2(Ω, h, 1). Thus, u = Bf is a unique solution for (5.1,5.2).
(cid:3)
References
[1] B. Ammann, V. Nistor, Weighted Sobolev spaces and regularity for polyhedral do-
mains, Comput. Methods Appl. Mech. Engrg., 196, (2007), 3650 -- 3659.
[2] D. Bertilsson, On Brennan's Conjecture in conformal mapping, Doctoral Thesis,
Royal Institute of Technology, Stockholm, Sweden 1999.
[3] J. Brennan, The integrability of the derivative in conformal mapping, J. London
Math. Soc., 18, (1978), 261 -- 272.
[4] H. Federer, Geometric measure theory, Springer Verlag. 1969.
[5] V. Gol'dshtein, L. Gurov, Applications of change of variables operators for exact
embedding theorems, Integral Equations Operator Theory, 19, (1994), 1 -- 24.
[6] V. Gol'dshtein, L. Gurov, A. S. Romanov, Homeomorphisms that induce monomor-
phisms of Sobolev spaces, Israel J. Math., 91, (1995), 31 -- 60.
[7] V. Gol'dshtein, V. V. Motreanu, A. Ukhlov, Embeddings of weighted Sobolev spaces
and degenerate Dirichlet problems involving the weighted p-Laplacian, Complex Vari-
ables and Elliptic Equations, 56, (2011), 905 -- 930.
[8] V. Gol'dshtein, A. S. Romanov, On mappings preserving the Sobolev spaces. Siberian
Math. J., 25, (1984), 55 -- 61.
[9] V. Gol'dshtein, A. Ukhlov, Weighted Sobolev spaces and embedding theorems, Trans.
Amer. Math. Soc., 361, (2009), 3829-3850.
[10] V. Gol'dshtein, A. Ukhlov, Sobolev homeomorphisms and composition operators,
Sobolev spaces in mathematics. Int. Math. Ser (N. Y.), 11, (2010), 207 -- 220.
[11] V. Gol'dshtein, A. Ukhlov, About homeomorphisms that induce composition op-
erators on Sobolev spaces, Complex Variables and Elliptic Equations, 55, (2010),
833-945.
[12] V. Gol'dshtein, A. Ukhlov, Brennan's Conjecture for composition operators on
Sobolev spaces, Eurasian Math. J., 3, (2012), 35-43.
[13] J. Heinonen, T. Kilpelinen, O. Martio, Nonlinear Potential Theory of Degenerate
Elliptic Equations. Clarendon Press. Oxford, New York, Tokio. 1993.
[14] V. Maz'ya, Sobolev spaces, Berlin: Springer Verlag. 1985.
[15] V. G. Maz'ya, T. O. Shaposhnikova, Multipliers in spaces of differentiable functions,
Leningrad Univ. Press. Leningrad, 1986.
[16] I. G. Markina, A change of variable preserving the differential properties of functions.
Siberian Math. J., 31, (1990), 73 -- 84.
[17] Ch. Pommerenke, Boundary Behavior of Conformal Maps, Springer-Verlag, 1992.
[18] F. Riesz, Sur les operations fonctionelles lineaires, C. R. Acad. Sci. Paris, 149, (1909),
974 -- 977.
[19] H. Hedelman, S. Shimorin, Weighted Bergman Spaces and the integral spectrum of
conformal mappings. Duke Mathematical Journal, 127, 2, (2005), 341-393.
[20] A. Ukhlov, On mappings, which induced embeddings of Sobolev spaces, Siberian
Math. J., 34, (1993), 185 -- 192.
[21] S. K. Vodop'yanov, The Taylor formula and function spaces, Novosibirsk Univ. Press.
Novosibirsk, 1988.
[22] S. K. Vodop'yanov, V. M. Gol'dshtein, Structure isomorphisms of spaces W 1
n and
quasiconformal mappings. Sib. Mat. Zh. 16, (1975), 224 -- 246.
[23] S. K. Vodop'yanov, V. M. Gol'dshtein, Functional characteristics of quasiisometric
mappings. Sib. Mat. Zh. 17,(1976), 768 -- 773.
[24] S. K. Vodop'yanov, A. D. Ukhlov, Superposition operators in Sobolev spaces. Russian
Mathematics (Iz. VUZ). 46, No. 4, (2002), 11 -- 33.
Vladimir Gol'dshtein
Department of Mathematics
Ben-Gurion University of the Negev
Alexander Ukhlov
Department of Mathematics
Ben-Gurion University of the Negev
UNIVERSAL CONFORMAL WEIGHTS ON SOBOLEV SPACES
19
P.O.Box 653, Beer Sheva, 84105, Israel
E-mail: [email protected]
P.O.Box 653, Beer Sheva, 84105, Israel
E-mail: [email protected]
|
1502.03572 | 1 | 1502 | 2015-02-12T09:19:39 | Geometrical Characterization of RN-operators between Locally Convex Vector Spaces | [
"math.FA"
] | For locally convex vector spaces (l.c.v.s.) $E$ and $F$ and for linear and continuous operator $T: E \rightarrow F$ and for an absolutely convex neighborhood $V$ of zero in $F$, a bounded subset $B$ of $E$ is said to be $T$-V-dentable (respectively, $T$-V-s-dentable, respectively, $T$-V-f-dentable) if for any $\epsilon>0$ there exists an $x\in B$ so that $ x\notin \overline{co} (B\setminus T^{-1}(T(x)+\epsilon V))$ (respectively, so that $ x\notin s$-$co (B\setminus T^{-1}(T(x)+\epsilon V)),$ respectively, so that $ x\notin {co} (B\setminus T^{-1}(T(x)+\epsilon V)) ). $ Moreover, $B$ is called $T$-dentable (respectively, $T$-s-dentable, $T$-f-dentable) if it is $T$-V-dentable (respectively, $T$-V-s-dentable, $T$-V-f-dentable) for every absolutely convex neighborhood $V$ of zero in $F.$ RN-operators between locally convex vector spaces have been introduced in [5]. We present a theorem which says that, for a large class of l.c.v.s. $E, F,$ if $T: E \rightarrow F$ is a linear continuous map, then the following are equivalent: 1) $T \in RN(E,F);$ 2) Each bounded set in $E$ is $T$-dentable; 3) Each bounded set in $E$ is $T$-s-dentable; 4) Each bounded set in $E$ is $T$-$f$-dentable. Therefore, we have a generalization of Theorem 1 in [8], which gave a geometric characterization of RN-operators between Banach spaces. | math.FA | math |
Geometrical Characterization of RN-operators between Locally
Convex Vector Spaces
OLEG REINOV
ASFAND FAHAD
St. Petersburg State University
Dept. of Mathematics and Mechanics
Universitetskii pr. 28, 198504 St, Petersburg
Government College University
Abdus Salam School of Math. Sciences
68-B, New Muslim Town, 54600 Lahore
RUSSIA
[email protected]
PAKISTAN
[email protected]
Abstract: For locally convex vector spaces (l.c.v.s.) E and F and for linear and continuous operator T : E → F
and for an absolutely convex neighborhood V of zero in F , a bounded subset B of E is said to be T -V-dentable
(respectively, T -V-s-dentable, respectively, T -V-f-dentable ) if for any ǫ > 0 there exists an x ∈ B so that
x /∈ co (B \ T −1(T (x) + ǫV ))
(respectively, so that x /∈ s-co (B \ T −1(T (x) + ǫV )), respectively, so that x /∈ co (B \ T −1(T (x) + ǫV )) ).
Moreover, B is called T -dentable (respectively, T -s-dentable, T -f-dentable ) if it is T -V-dentable (respectively,
T -V-s-dentable, T -V-f-dentable ) for every absolutely convex neighborhood V of zero in F.RN-operators between
locally convex vector spaces have been introduced in [5]. We present a theorem which says that, for a large class
of l.c.v.s. E, F, if T : E → F is a linear continuous map, then the following are equivalent:
1) T ∈ RN (E, F ).
2) Each bounded set in E is T -dentable.
3) Each bounded set in E is T -s-dentable.
4) Each bounded set in E is T -f -dentable.
Therefore, we have a generalization of Theorem 1 in [8], which gave a geometric characterization of RN-operators
between Banach spaces.
Key -- Words: dentability, dentable set, locally convex space, Radon-Nikodym operator
1 Introduction
We use standard notations and terminology from the
theory of operators in Banach spaces and measure the-
ory (see, e.g., [1],[3]). Our main references on the the-
ory of the locally convex vector spaces is [12]. Banach
spaces are usually denoted by the letters X, Y, . . . ,
and the usual notations for the l.c.v.s. (locally convex
vector spaces) are E, F, . . . . All the spaces we con-
sider in the paper are over the field of the real num-
bers.
Some of the main notions and notations we will
use are ones of the convex, s-convex and closed con-
vex hulls of sets in Banach or l.c.v. spaces. Namely,
co(A)
co(A),
(respectively,
s-
co(A)) denotes the convex hull (respectively, the clo-
sure of the convex hull, the s-closure of the convex
hull) of the set A. The s-closure is defined as s-
aixi, where xi ∈ A, ai ≥ 0 and
respectively,
co(A) := {Pi≥1
Pi≥1
ai = 1}.
Let us recall some more notions:
Definition 1 Let X be a Banach space and (Ω, Σ) a
measureable space, consider m : Σ → X. m is called
1 of pair-
a vector measure if for every sequence {Ai}∞
i=1 Ai) =
wise disjoint sets from Σ one has m(S∞
P∞
i=1 m(Ai).
Definition 2 Let m : Σ → X be a vector measure.
Variation of a vector measure is a non-negative, ex-
tended real valued function with value on set A ∈ Σ
is
m(A) = supQ XAi∈Q
km(Ai)k
where Q denotes all finite partitions of A with pair-
wise disjoint sets in Σ; m is called vector measure of
bounded variation if m(Ω) is finite.
Let (Ω, Σ, µ) be a measure space.
Definition 3 A function
: Ω → X is µ-
measureable if there exists a sequence of simple func-
(see
tions {fn}∞
[1],page 41).
kfn − f k = 0 µ-a.e.
n=1 with lim
n
f
Definition 4 A µ-measureable function f : Ω → X is
called Bochner integrable if there exists a sequence of
simple functions fn such that lim
n RΩ kfn − f kdµ = 0
n RA fn
(see [1],page 44); in this case we setRA f = lim
for every A ∈ Σ.
The following notions have been introduced in [9]
and, independently, in [6].
Definition 5 A linear continuous operator T : X →
Y is said to be RN-operator (T ∈ RN (X, Y )), if for
every X-valued measure m of bounded variation, for
any measure space (Ω, Σ, µ) with m << µ there ex-
ists a function f : Ω → Y which is Bochner integrable
and such that
T (m)(A) = ZA
f dµ
for every A ∈ Σ.
Definition 6 Let T : X → Y be a linear continu-
ous operator A bounded subset B of Banach space X
is called T -dentable (respectively, T -f-dentable, T -s-
dentable), if for every ǫ > 0 there exists x ∈ B such
that
x /∈ co (B \ T −1(Dǫ(T (x)))).
(respectively, so that
x /∈ co (B \ T −1(Dǫ(T (x)) )),
resp.,
x /∈ s − co (B \ T −1(Dǫ(T (x)) ))),
The geometrical characterization of RN-operators
between Banach spaces has been given by following
theorem in [8].
Theorem 7 Let X, Y be Banach spaces and T :
X → Y , linear and continuous then the following
statements are equivalent:
1, T ∈ RN (X, Y )
2. Each bounded set in X is T -dentable.
3. Each bounded set in X is T -s-dentable.
4. Each bounded set in X is T -f -dentable.
Our aim is to generalize the above theorem for
operators between some l.c.v.
spaces. Firstly, we
shall define the notion of RN-operators between lo-
cally convex vector spaces. For this, we recall the def-
initions of Banach spaces of type EB and EV . Then
we define the notions of T -dentability, T -s-dentability
and T -f-dentability for operators in l.c.v.s. (in the next
section).
Let E be a l.c.v.s. and B ⊆ E be bounded and
n=1 nB. Define
absolutely convex. Let EB = S∞
Minkowski function on EB w.r.t. B :
kxkρB = ρB(x) = inf{λ ≥ 0 : x ∈ λB}.
It is a seminorm in general but we show that when
B is bounded and absolutely convex then it is a norm.
Suppose x ∈ EB such that ρB(x) = 0 so inf{λ ≥ 0 :
x ∈ λB} = 0 this implies for any ǫ > 0 ∃ λ ≥ 0 such
that x ∈ λB and λ ≤ ǫ, take ǫ = 1
n B ∀
1
n B = {0}. So x = 0
and (EB, ρB) is a normed space and its completion
EB is Banach space in E. Moreover, if B is complete
then EB is Banach space.
n ∈ N this implies x ∈ T∞
n , so x ∈ 1
n=1
Similarly
absolutely
an
zero V
for
neighborhood of
open
= V (0) we set
n=1 nV = E. Similarly we define
Minkowski function on EV w.r.t V , so that for x ∈ E
the semi norm of x is
EV = S∞
convex
kxkρV = ρV (x) = inf{λ ≥ 0 : x ∈ λV }
.
We identify two elements x, y ∈ E w.r.t. ρV ,
obtaining a quotient EV with corresponding elements
x, y, by x = y iff ρV (x − y) = 0. We get a normed
space EV and its completion gives us a Banach space
( EV , ρV ).
We will use the following important theorem (see
[10]) in the proof of our main theorem.
Theorem 8 (see [10]) Let E be a locally convex
vector space, V = V (0) be an absolute convex
neighborhood of 0, let B ⊆ E be a closed, bounded,
convex, sequentially complete and metrizable subset,
The following are equivalent:
(i) B is subset V -dentable.
(ii) B is subset V -f-dentable.
It has been shown in [10] that it follows from the
above theorem:
Let E be a locally convex vector space and let
B ⊆ E have the following properties:
(i) B is closed, bounded, convex and sequentially
complete,
(ii) for every bounded M ⊆ E and for x ∈ M there
exist a sequence xn ∈ M such that lim
n
(iii) each separable subset of B is metrizable.
xn = x,
Then the following are equivalent:
(i) B is subset V -dentable,
(ii)B is subset V -s-dentable,
(ii)B is subset V -f-dentable.
We will say that a locally convex vector space E
is an SBM-space if
(i) every closed bounded convex subset of E is
sequentially complete,
(ii) for every bounded M ⊆ E and for x ∈ M
xn = x
there exists a sequence xn ∈ M such that lim
n
and
(iii) each separable bounded subset of E is
metrizable.
Therefore,
if E is an SBM-space,
then every
bounded subset of E is V -dentable iff every bounded
subset of E is V -s-dentable iff every bounded subset
of E is V -f-dentable.
All quasi-complete (BM)-spaces [11], in particu-
lar, all Fr´echet spaces are SBM-spaces.
2 Main Results
Definition 9 Let E and F be locally convex vector
spaces (l.c.v.s.), let T : E → F be a linear and
continuous operator and let V be an absolutely con-
vex neighborhood of zero in F. A bounded subset B
of E is said to be T -V-dentable (respectively, T -V-
s-dentable, respectively, T -V-f-dentable ) if for any
ǫ > 0 there exists an x ∈ B so that
x /∈ co (B \ T −1(T (x) + ǫV ))
(respectively, so that
x /∈ s − co (B \ T −1(T (x) + ǫV )) ,
respectively, so that
x /∈ co (B \ T −1(T (x) + ǫV )) ).
Definition 10 Let T : E → F be a linear and
continuous operator. A bounded subset B of E is
said to be T -dentable (respectively, T -s-dentable, T -
f-dentable ) if for every absolutely convex neighbor-
hood V of zero in F there exists an x ∈ B such that
x /∈ co (B \ T −1(T (x) + V ))
(respectively, so that
x /∈ s − co (B \ T −1(T (x) + V ))
and, respectively, so that
x /∈ co (B \ T −1(T (x) + V )) ).
Remark 11 From the above two definitions it is clear
that B ⊆ E is T -dentable if and only if for every V it
is T -V-dentable. The same is true for corresponding
properties of s-dentability and f-dentability.
The following is the definition of an RN-operator
between locally convex vector spaces, and this is our
main definition.
Definition 12 (see [5]) Let T : E → F be linear and
continuous (in l.c.v.s.). T ∈ RN (E, F ) (a Radon-
Nikodym operator or RN-operator) if for every com-
plete, absolutely convex and bounded set B ⊆ E and
for any absolutely convex neighborhood V ⊆ F of
zero the natural operator ΨV ◦ T ◦ φB : EB → E →
F → FV is RN -operator between Banach spaces EB
and FV .
Here, ΨV and φB are the natural maps (cf. defi-
nitions in Section 1).
Remark 13 For the operators in Banach spaces our
definition of RN -operators coincides with the origi-
nal definitions from [9], [6]. This must be clear.
Remark 14 The usual definition of a weakly compact
operator between locally convex spaces is: T : E →
F is weakly compact if T maps a neighborhood of
zero in E to a relatively weakly compact subset of F.
The analogous definition can be given for the compact
operators. For the weakly compact case this means
that
(1) There exists an absolutely convex neighbor-
hood V = V (0) in E such that if D := T (V ),
then D is bounded in F and the natural injection
φD : FD → F is weakly compact.
Since every weakly compact operator in Banach
spaces is Radon-Nikodym, we get from the definition
that every weakly compact operator in l.c.v.s.
is an
RN-operator.
Following this way, we can defined also a class of
"bounded RN-operators". Namely, let T maps E to
F. T is a bounded RN-operator if T takes a neighbor-
hood V to a bounded subset of F and the natural map
φT (V ) : FT (V ) → F is "Radon-Nikodym". But how
to understand "Radon-Nikodym" in this case where
an operator maps a normed (or Banach) space to a lo-
cally convex space? We can go by a geometrical way
(saying that the image of φT (V ) is subset s-dentable.
Here the map φT (V ) is one-to-one, and we can follow
the assertion from [8] for operators in Banach spaces:
if U : X → Y is one-to-one then U is RN iff the U-
image of the unit ball is subset s-dentable). Another
way is just to apply the main definition from this paper.
Or, we can go by a traditional way: for T : X → F
with X Banach and F locally convex, say that T is of
type RN if for every operator U from an L1-space to
X the composition T U admits an integral represen-
tation with a (strongly) integrable F -valued function.
But in this case we are to give a good definition of
the "integrability" of an F -valued function. All these
are the topic for the further considerations in another
paper.
Let us mention that in every "right" definition
of RN-operator there must be an "ideal property":
AT B is RN for all linear continuous A, B if T is RN.
Thus, if we apply the definition 12 above as the main
definition (in this paper),
then every operator "of
type RN" considered above in this remark is "right"
Radon-Nikodym. This is one of the reason that here
we will deal only with Definition 12.
If E = X is a Banach space then T ∈ RN (X, F )
iff for every (a.c.) neighborhood V = V (0) ⊂ F the
composition ΨV T belongs to RN (X, bFV ). If F is a
Banach space too, then, evidently, T ∈ RN in the
sense of Definition 12 iff it is an RN-operator in the
usual sense of [9] and [6]. If E is a l.c.v.s. and F = Y
is Banach, then T ∈ RN (E, Y ) iff for every bounded
complete absolutely convex subset B ⊂ E the natu-
ral map T φB : EB → E → Y is Radon-Nikodym,
what means that T ∈ RN (E, Y ) iff for any finite
measure space (Ω, Σ, µ) and for every µ-continuous
E-valued measure ¯m : Σ → E with bounded µ-
average { ¯m(A)
: A ∈ Σ, µ(A) 6= 0} the measure
µ(A)
T ¯m : Σ → Y has a Bochner derivative with respect
to µ. From this it follows that an operator T between
locally convex spaces E, F is Radon-Nikodym (in the
sense of Definition 12) iff for any finite measure space
(Ω, Σ, µ) and for every µ-continuous E-valued mea-
sure ¯m : Σ → E with bounded µ-average, for ev-
ery a.c. neighborhood V = V (0) ⊂ F the measure
ΨV T ¯m : Σ → bFV has a Bochner derivative with re-
spect to µ.
Proposition 15 Let ΨV : E → EV , where V =
V (0) is an absolutely convex open neighborhood of
−1(D1(0)) = V , where D1(0) is open
zero then ΨV
unit ball in EV .
Proof: " ⊇ ": Let x ∈ V , then by definition of k.k EV
kΨV (x)k < 1 so V ⊆ ΨV
" ⊆ ": Let x ∈ ΨV
−1(D1(0))
−1(D1(0)); then ΨV (x) ∈
D1(0). This implies kΨV (x)k = 1 − c < 1 for some
c > 0, so inf{λ > 0 : x ∈ λV } = 1 − c < 1. This
implies x ∈ (1 − c
2 )V . Since V is absolutely convex
so λV ⊆ V for every λ < 1, therefore x ∈ V .
Hence ΨV
⊓⊔
−1(D1(0)) = V.
,
Theorem 16 Let E, F be locally convex vector
spaces and T : E → F be a linear continuous op-
erator. Consider the following conditions:
1) T ∈ RN (E, F ).
2) Each bounded set in E is T -dentable.
3) Each bounded set in E is T -s-dentable.
4) Each bounded set in E is T -f -dentable
We have 1) ⇔ 4), 2) ⇒ 3) ⇒ 4). If the space E is
an SBM-space then all the conditions are equivalent.
Proof: 1) ⇒ 4). Let B0 ⊆ E be bounded and V ⊆ F
be an absolutely convex neighborhood of 0; put B =
Γ(B0). By assumption in 1), the composition operator
ΨV ◦ T ◦ φB : EB → E → F → FV is an RN-
operator from EB to FV . By theorem 7 each bounded
subset in EB is ΨV T φB-f-dentable. In particular, B0
is ΨV T φB-f-dentable. Let ǫ > 0, there exists x ∈ B0
such that
(i) x /∈ co (B0 \ (ΨV T φB)−1(Dǫ(ΨV T φB(x)))).
−1(Dǫ(ΨV T φB(x))) = ΨV
15
implies
that
−1(ΨV (T (x))) + ǫV ,
Proposition
ΨV
so we get
(ii) x /∈ co (B0 \ T −1(T (x) + ǫV ),
0 ≤ λk ≤ 1 and Pn
Pn
otherwise there exist x1, x2 ..., xn in B0 where
/∈ T −1(T (x) + ǫV ) and λ1, λ2 ...,λn where
xk
k=1 λk = 1 such that
k=1 λkxk = x. Since xk /∈ T −1(T (x) + ǫV )
/∈
−1(ΨV (T (x))) for each k = 1, 2 ... n. So,
−1(Dǫ(ΨV T φB(x))) and
k=1 λkxk = x which is contradicts (i), therefore
(ii) holds and hence B0 is T -V-f-dentable, which
proofs 4).
so T (xk)
ΨV
we get xk /∈ φB
−1T −1ΨV
/∈
(T (x) + ǫV ) hence T (xk)
Pn
4) ⇒ 1). Let B ⊆ E be a bounded, absolutely
convex and complete subset and V ⊆ F be an ab-
solutely convex neighborhood of 0. We need to show
that the operator ΨV ◦ T ◦ φB : EB → E → F → FV
is RN. By theorem 7, it is enough to show that each
bounded B0 is ΨV T φB-f-dentable. Let ǫ > 0; by as-
sumption in 4) there exists an x ∈ B0 such that
x /∈ co (B0 \ T −1(T (x) + ǫV )) ),
which, by above proposition, gives that
x /∈ co (B0 \ (ΨV T φB)−1(Dǫ(ΨV T φB(x))) ).
This proofs 1).
2) ⇒ 3) ⇒ 4). Follows from definitions.
Suppose now that E is an SBM-space.
4) ⇒ 2). Let B ⊆ E be closed bounded and con-
vex and let V be an absolutely convex neighborhood
Acknowledgements: The research was supported by
the Higher Education Commission of Pakistan and by
the Grant Agency of RFBR (grant No. 15-01-05796).
References:
[1] J. Diestel and J. J. Uhl Jr., Vector measures,
American Mathematical Society, Providence, RI
1977.
[2] J. Diestel, Geometry of Banach Spaces, Se-
lected topics, Lecture notes in Mathematics,
485, Springer Verlag, Berlin 1975.
[3] N. Dunford, J. T. Schwartz, Linear operators.
Part I, Wiley Classics Library, John Wiley Sons,
New York 1988.
[4] L. Egghe, The Radon-Nikodym Property, s-
dentability and martingales in locally convex
spaces. Pacific J. of Math. 87, 1980, pp. 313-322.
[5] A. Fahad, O. I. Reinov, On dentability in locally
convex topological vector Space, Preprint 521,
May 2013, Abdus Salam School of Mathemati-
cal Sciences, GC University, Lahore, Pakistan.
[6] W. Linde, An operator ideal in connection with
the Radon-Nikodym property of Banach spaces,
Math. Nachr. 71, 1076, pp. 65-73.
[7] A. Fahad, O. I. Reinov , On dentability in
locally convex vector spaces, arXiv:1302.6019
[math.FA], 2013, pp. 1-5
[8] O.I. Reinov, Geometric characterization of RN-
operators, Matematicheskie Zametki 22, 1977,
pp. 189 -- 202.
[9] O. I. Reinov, Operators of type RN in Banach
spaces, Dokl. Akad. Nauk SSSR 220, 1975, pp.
528-531.
[10] O. I. Reinov, A. Fahad, On dentability in lo-
cally convex vector spaces, arXiv:1302.6019
[math.FA], 2013, pp. 1-5.
[11] Elias Saab, On the Radon-Nikodym property
in a class of locally convex spaces, Pacific J.
Math. 75, 1978, pp. 281-291.
[12] H. H. Schaefer, Topological vector spaces,
Graduate texts in Mathematics, 3, Springer Ver-
lag, Berlin 1971.
of zero in F . We show that B is subset T -V-dentable.
Clearly, B0 ⊂ B is T -V-f-dentable if and only if B0
is U = T −1(V )-f-dentable. From assumption, it fol-
lows that B is subset U-f-dentable. Applying the con-
sequence of theorem 8, we get that B is subset U-
dentable, or B is subset T -V-dentable, which proofs
2).
⊓⊔
Example 17 Consider any uncountable set Γ and the
classical Banach spaces c0(Γ) and l1(Γ); c0(Γ)∗ =
l1(Γ). The closed unit ball B of l1(Γ) is weak∗ com-
pact and an Eberlein compact in this weak∗ topol-
ogy (see, e.g., [2]). This implies that every separa-
ble subset of (B, w∗) is metrizable and that (B, w∗)
has the Frechet-Uryson property, i.e. every point in
the closure of a subset is a limit of a sequence of
this subset. Thus, the space (l1(Γ), w∗) is of type
SBM. The Banach space l1(Γ) has the RNP. There-
fore, every bounded subset of the space is dentable
(s-dentable, f-dentable), and thus every bounded sub-
set of (l1(Γ), w∗) is (w∗)-dentable too. The identity
map (l1(Γ), w∗) → (l1(Γ, w∗)) is Radon-Nikodym in
the sense of Definition 12. Moreover the space has
the RNP in the sense of the paper [11]. This directly
follows by Theorem 16, but not directly from the re-
sults of [11]. On the other case, for l1(Γ) this fact is
trivial. We can also say (by the same considerations)
that if X is any weakly generated Banach space (see
[2]) with X ∗ ∈ RN, then the space (X ∗, σ(X ∗, X))
has all just mentioned properties (clearly, or by appli-
cation of Theorem 16).
Finally, if X is WCG, then (X ∗, σ(X ∗, X)) is
SBM. So, the theorem can be applied for all such
spaces.
On the other hand,
there is an example (see
there exists
[4], Theorem 4), which shows that
a separable Banach space Y such that
the space
Yσ:=(Y, σ(Y, Y ∗)) does not have the RNP (in the
sense of [11]) and every bounded subset of which (of
Yσ) is s-dentable. Thus, the above theorem is not true
for general l.c.v.s.
Let us note that a l.c.v.s. E has the RNP in sense
of [11] iff the identity map E → E is an RN-operator.
Therefore, we obtain a generalization of a theorem
from [11]:
Corollary 18 For every SBM-space E (in particu-
lar, for every quasi complete l.c.v.s. with metrizable
bounded subsets, or for every Frechet space) the fol-
lowing are equivalent:
1. E has the RNP of [11].
2. Each bounded set in E is dentable.
3. Each bounded set in E is s-dentable.
4. Each bounded set in E is f -dentable.
|
1307.0924 | 1 | 1307 | 2013-07-03T06:48:59 | Beurling's Theorem And Invariant Subspaces For The Shift On Hardy Spaces | [
"math.FA"
] | Let $G$ be a bounded open subset in the complex plane and let $H^{2}(G)$ denote the Hardy space on $G$. We call a bounded simply connected domain $W$ perfectly connected if the boundary value function of the inverse of the Riemann map from $W$ onto the unit disk $D$ is almost 1-1 rwith respect to the Lebesgure on $\partial D$ and if the Riemann map belongs to the weak-star closure of the polynomials in $H^{\infty}(W)$. Our main theorem states: In order that for each $M\in Lat(M_{z})$, there exist $u\in H^{\infty}(G)$ such that $ M = \vee\{u H^{2}(G)\}$, it is necessary and sufficient that the following hold:
1) Each component of $G$ is a perfectly connected domain.
2) The harmonic measures of the components of $G$ are mutually singular.
3) % $P^{\infty}(\omega)
The set of polynomials is weak-star dense in $ H^{\infty}(G)$.
\noindent Moreover, if $G$ satisfies these conditions, then every $M\in Lat(M_{z})$ is of the form $u H^{2}(G)$, where %$u\in H^{\infty}(G)$ and the restriction of $u$ to each of the components of $G$ is either an inner function or zero. | math.FA | math |
Beurling's Theorem And Invariant Subspaces
For The Shift On Hardy Spaces)
Zhijian Qiu
Department of mathematics, Southwestern Univ of Finance and Economics
Chengdu 610072, China, E-mail:[email protected]
Abstract
Let G be a bounded open subset in the complex plane and let H 2(G)
denote the Hardy space on G. We call a bounded simply connected domain
W perfectly connected if the boundary value function of the inverse of the
Riemann map from W onto the unit disk D is almost 1-1 rwith respect to the
Lebesgure on ∂D and if the Riemann map belongs to the weak-star closure
of the polynomials in H ∞(W ). Our main theorem states: In order that for
each M ∈ Lat(Mz), there exist u ∈ H ∞(G) such that M = ∨{uH 2(G)}, it
is necessary and sufficient that the following hold:
1) Each component of G is a perfectly connected domain.
2) The harmonic measures of the components of G are mutually singular.
3) The set of polynomials is weak-star dense in H ∞(G).
Moreover, if G satisfies these conditions, then every M ∈ Lat(Mz) is of the
form uH 2(G), where the components of G is either an inner function or zero.
Keywords. Hardy space, invariant subspace, shift operator
AMS subject classification. 30H05 30E10, 46E15, 47B20
Introduction
In [4] A. Beurling completely described all the invariant subspaces for the operator
"multiplication by z" on the classical Hardy space H 2(D) on the unit disk D.
H 2(D) consists of analytic functions on D whose coefficients of Taylor expansion
are square summable. A closed linear subspace M of H 2(D) is said z-invariant
(or shift invariant) if zf ∈ M for every f ∈ M .
Beurling's theorem states that every z-invariant subspace of H 2(D) is either
zero subspace {0} or of the form uH 2(D), where u is an inner function in H 2(D),
that is, a bounded analytic function on D with nontangential boundary values of
modulus 1 almost everywhere respect to the Lebesgue measure on the unit circle
∂D.
Beurling's theorem is viewed as one of the most celebrated theorems in operator
theory and it has been extended to many directions. Unfortunately, the Beurling
type theorem does not hold for the shift (the operator of multiplication by z) on
the Hardy spaces on an arbitrary domain. The lattice of the invariant subspaces
of the shift is unknown in general.
In the case of Bergman space on the unit
disk, a theorem of Aleman, Richter and Sundberg says that if M is an invariant
1
subspace for the Bergman shift (which is the operator of the multiplication by
z) on the Bergman space L2
a(G), which consists of the analytic functions on D
that are square integrable with respect to the area measure, then M is generated
by M ⊖ zM , the orthogonal complement of zM in M . This theorem is regarded
as the Beurling's theorem for the Bergman shift. Comparing to the case of the
shift on H 2(D), this result does not tell what is M . Unlike the Hardy space
case, the dimension of M ⊖ zM could be any positive integer or ∞ ([2, 11]) and
characterization of the invariant subspaces for the Bergman shift is really difficult.
The lattice of invariant subspaces of the Bergman shift is unknown.
If we have our attention back to the Hardy spaces, then we may ask: for which
Hardy space H 2(G) it is possible to describe the lattice of invariant subspaces of
the shift operator?
For this question, let us begin with the definition of the Hardy spaces. For a
bounded domain G in the complex plane and a positive number q in [1, ∞), let
H q(G) be the set of analytic functions f for which there is a function u harmonic
on G such that f (z)q ≤ u(z) on G. Fix a point a ∈ G, then
kf k = inf{u(a)
1
q : f (z)q ≤ u(z), z ∈ G and u is harmonic on G}
defines a norm on H q(G), which is called the Hardy space on G (see [6]). By the
Harnack's inequality, the norms induced by different points are equivalent.
If G is a bounded open subsets with components G1, G2, ...Gn, ..., let H q(G)
denote the collection of the functions to G whose restriction to Gn belongs to
H q(Gn) for each n. If we define the addition on H q(G) by (f + g)(z) = f (z) + g(z)
pointwise on G and the scalar multiplication in the obvious way, then H q(G) is a
linear space over the complex field. Moreover, it is easy to verify that
kf kH q(G) = {Xn=1
1
2n kf Gnkq
H q(Gn)}
1
q
for each f ∈ H q(G)
defines a norm on H q(G) that makes it a complex Banach space of analytic func-
tions on G.
For q = ∞, H ∞(G) denotes the Banach algebra of the bounded analytic func-
tions on G with the supremum norm.
In this paper, we study the invariant space problem for the shift operator on
Hardy space H 2(G) over a general open subset set G. We investigate the maximum
extent of Beurling's theorem in its original form in Hardy spaces. In other words,
we try to find the most general open subsets G in the complex plane so that the
invariant subspace for the shift on H 2(G) have as simple forms as those for the
shift on H 2(D).
The main result of this paper, Theorem 5, gives a complete description of such
open subsets in the plane. Our characterization is in terms of harmonic measure
and perfectly connected domain as well as weak-star density of polynomials in
H ∞(G).
In Section 1 we give some preliminaries.
In Section 2, we study perfectly
connected domains and the related weak-star density problem. In Section 3, we
present our main theorem in this paper.
2
1 A structure theorem for the mean closure of
the rational functions
Let G be a bounded domain with no singleton boundary component. For a fixed
z ∈ G, there is a unique probability measure ωz supported on ∂G such that
u(z) =Z∂G
u(λ)dωz(λ)
for every function u(λ) that is continuous on G and harmonic on G. The measure
ωz is called the harmonic measure of G evaluated at z. By the Harnack's inequality,
any two harmonic measures for G are boundedly equivalent. The normalized
Lebesgue on ∂D is the harmonic measure evaluated at 0 for D.
For an arbitrary open subset G, let {Gi}∞
i=1
of G. Then a harmonic measure for G is defined to be ω = P∞
i=1 be the collection of the components
ωi
2i , where ωi is
2i , i = 1, 2, ..., are chosen so that ω
a harmonic measure for Gi. The constants 1
become a probability measure on G.
Let K be a compact subset in the plane and let µ be a positive finite Borel
measure supported on K. Let R2(K, µ) denote the closure of Rat(K), the set of
the rational functions with poles off K, in Lq(µ).
Let P denote the set of (analytic) polynomials and P q(µ) denote the closure
of P in L2(µ).
For q ∈ [1, ∞), a point w in C is called an analytic bounded point evaluation
(abpe) for Rq(K, µ) if there is a neighborhood G of w and a positive constant c
such that for λ ∈ G
f (λ) ≤ c kf kLq(µ),
f ∈ Rat(K).
So the map, f → f (λ), extends to a functional in Rq(K, µ)∗, the dual space of
Rq(K, µ). Thus, there is a (kernel) function kλ in Rq(K, µ)∗ such that
f (λ) =Z f kλ dµ, f ∈ Rat(K).
Clearly, the set of abpes is open. For each f ∈ Rq(K, µ), let f (λ) = R f kλ dµ.
Then f (λ) is analytic on the set of abpes.
We use ∇Rq(K, µ) to denote the set of abpes for Rq(K, µ).
The proof of our main theorem needs the following structure theorem for
Rq(K, µ), which can be found in [16]. Recall that the connectivity of a connected
domain is defined to be the number of the components of its complement in the
complex plane. A connected open subset is called circular domain if its boundary
consists of a finite number of disjoint circle.
Theorem 1 Let K be a compact subset whose complement in the plane has only
finitely many components and let µ be a positive finite Borel measure supported on
K. Then there is a Borel partition {∆n}∞
n=0 of the support of µ such that
Rq(K, µ) = Lq(µ∆0) ⊕ Rq(K, µ∆1) ⊕ ... ⊕ Rq(K, µ∆n) ⊕ ...
and for each n ≥ 1, if Un denotes ∇Rq(K, µ∆n), then
1)U n ⊃ ∆n and Rq(K, µ∆n) = Rq(U n, µ∆n);
2) the map e, defined by e(f ) = f , is an isometrical isomorphism and weak-star
homeomorphism from Rq(K, µ∆n)T L∞(µ∆n) onto H ∞(Un).
3
3) Un is conformally equivalent to a circular domain and the connectivity of
Un does not exceed the connectivity of the component of G that contains Un;
4) µ∂Un ≪ ωUn , the harmonic measure of Un; and if un is a conformal map
from a circular domain Wn onto Un, then every f ∈ H ∞(Un) has nontangential
limits on ∂Un a.e. [µ] and
e−1(f )(a) = lim
z→u−1
n (a)
f ◦un(z) for almost every on a ∈ ∂G with respect to µ∂Un;
5) for each f ∈ H ∞(Un), if let f ∗ be equal to its nontangential limit values on
∂Un and f ∗ = f on Un, then the map m, defined by m(f ) = f ∗∆n, is the inverse
of the map e.
An observation is that 3) in the theorem above implies that each Rq(U n, µ∆n)
contains no non-trivial characteristic function (because H ∞(Un) does not).
For a given µ, let K be a closed disk that contains the support of µ, then
R2(K, µ) = P 2(µ) and R2(K, µ∆i) = P 2(µ∆i). So Theorem 1 is also a structural
theorem for P 2(µ). In 1991 J. Thomson first proved Theorem 1 for P 2(µ) except
4) and 5) in [20]. Some applications of Theorem 1 can be found in [14, 17].
2 On Perfectly Connected Domains
Following [10] we call a simply connected domain G nicely connected if the bound-
ary value of a conformal map from the unit disk D onto G is univalent almost
everywhere on ∂D with respect the Lebesgue measure. So every bounded analytic
function in H ∞(G) has a well-defined boundary valuefunction on ∂G if G is a
nicely connected domain. A simply connected domain G is nicely connected do-
main if and only if every function in H ∞(G) can be approximated by a bounded
sequence of functions analytic on G and continuous on G (for example, see [5]).
For a simply connected domain G, recall that the weak-star topology for
H ∞(G) is defined as follows: let m denote the normalized Lebesgue measure and
let P ∞(m) denote the weak-star closure of P in L∞(m). The map f → f (where
f denotes the boundary value function of f on ∂D) from H ∞(D) to P ∞(m) is an
isometrical isomorphism. Since P ∞(m) is the dual space of L1(m)/P ∞(m)⊥, we
have a weak-star topology on H ∞(D) that is induced from P ∞(m) via the isomet-
rical isomorphism. For a bounded simply connected domain G, a conformal map
from D onto G induces an isometrical isomorphism from H ∞(D) onto H ∞(G) in
the obvious way. Thus the weak-star topology on H ∞(G) is defined to be the one
induced from H ∞(D) by that map.
Now we introduce the following:
Definition. A bounded simply connected domain G is called perfectly connected
if the Riemann map from G onto D belongs to the weak-star closure of the poly-
nomials in H ∞(G).
Theorem 2 Let G be a bounded simply connected domain and let ω be a harmonic
measure for G. Then the following are equivalent:
1) G is a perfectly connected domain.
2) G is a nicely connected domain and the boundary value function of the
Riemann map from G onto D belongs to P ∞(ω).
3) The set of polynomials is weak-star dense in H ∞(G).
4
Proof.
1) =⇒ 2). Suppose that G is a perfectly connected domain. Let ψ denote the
Riemann map from G onto D and let ϕ denote its inverse function. Then ψ can
be weak-star approximated by polynomials in H ∞(G). Without loss of generality,
we may assume that ω = m ◦ ψ. Choose a net of polynomials {qα} such that it
weak-star converges to ψ in H ∞(G). So the net {qα ◦ ϕ} weak-star converges to z
in H ∞(D). This is equivalent to that {qα ◦ ϕ} weak-star converges to z in P ∞(m),
where we still use ϕ to denote the boundary value of ϕ on ∂D. We claim that z
belongs to the closure of {p ◦ ϕ : p ∈ P} in P 2(m). In fact, recall that a convex
set is norm closed if and only if it is weakly closed. Since the weak topology and
weak-star topology coincide on a Hilbert space and since z belongs the weak-star
closure of {p ◦ ϕ : p ∈ P} in P ∞(m), the claim clearly follows.
To show G is nicely connected, choose a sequence of polynomials {pn ◦ ϕ} such
that it converges to z in P 2(m). By passing to a subsequence if necessary, we
have that pn ◦ ϕ → z almost everywhere on ∂D with respect to m. This clearly
implies that ϕ is univalent almost everywhere on ∂D. Thus, G is nicely connected.
Therefore, every f in H ∞(G) has a boundary value function, f , on ∂G.
Since {qα ◦ ϕ} weak-star converges to z in P ∞(m), we have that
α Z qα ◦ ϕf dm =Z zf dm,
lim
f ∈ L1(m).
So
α Z qα(f ◦ ψ)dm ◦ ψ =Z ψ(f ◦ ψ)dm ◦ ψ,
lim
f ∈ L1(m).
Since L1(ω) = {f ◦ ψ : f ∈ L1(m)}, it follows that limαR qαgdω = R ψgdω, f ∈
L1(ω). Therefore, ψ is in P ∞(ω).
2) =⇒ 3). With the same notations as above. We suppose that G is a nicely
connected domain and ψ belongs to P ∞(ω).
Let f ∈ H ∞(G), for convenience, we still use f to denote its boundary value
function. Then f ◦ ϕ ∈ H ∞(D). So there exists a net of polynomials {qτ } such
that it weak-star converges to f ◦ ϕ in H ∞(G). This, in turn, implies that {qτ ◦ ψ}
weak-star converges to f ∈ L∞(ω). Since qτ ◦ ψ belongs to P ∞(ω) for each n, it
follows that f ∈ P ∞(ω). Hence H ∞(G) ⊂ P ∞(ω). Obviously, P ∞(ω) ⊂ H ∞(G),
so we conclude that P ∞(ω) = H ∞(G). Therefore, P is weak-star dense in H ∞(G).
3) =⇒ 1) follows by the definition.
Remark 1 There is no simple topological descrition for perfectly connected do-
mains, however, a Jordan domain (enclosed by a Jordan curve) is perfectly con-
nected and so is a Caratheodory domain (a domains whose boundary is equal to
the boundary of their complement in the plane).
With Theorem 2, we know that a perfectly connected domain is the image
of an bounded analytic function on D known as a weak-star generator, which is
characterized by D. Sarason in [18]. Though a perfectly connect domain may look
so 'bad' (see [18, 19]), however, it represents the class of simply connected domains
most resemble to the unit disk from the view of the theory of Hardy spaces. Our
main theorem itself would be an example that reflects this view.
Let 1 ≤ q ≤ ∞ and let G be a nicely connected domain. Then each function
in H q(G) has a well-defined boundary value function on ∂G. For f ∈ H q(G), let
f denote the boundary value function on ∂G. The map, f → f is an isometrical
5
isomorphism from H q(G) onto eH q(G), which is defined to be the subspace { f :
f ∈ H q(G)} in Lq(ω) (where ω is chosen so that krkH q(G) = krkLq(ω) for each
∈ Rat(G)).
Now, for a bounded open subset, if each of the components of G is nicely
connected and the harmonic measures of G are mutually singular, then every
function f in H q(G) also has a well defined boundary value function f on ∂G. As
before, we use eH q(G) to denote the subspace { f : f ∈ H q(G)} in Lq(ω) (again, ω
is chosen so that krkH q (G) = krkLq(ω) for each r ∈ Rat(G)). Moreover, f → f is
an isometrical isomorphism.
In this paper, for seek of convinince, we often don't distinguish between the
two spaces. So, for any compact subset K that contains ∂G, when we say that
Rq(K, ω) = H q(G), we really means that Rq(K, ω) = eH q(G), provided that ω is
chosen so that the map f → f is an isometry (note, in the case that q = ∞, any
harmonic measure ω will do the job). Keep this observation in mind.
Lemma 1 Let G be a bounded domain and let ω be a harmonic measure for G.
If K is a compact subset that contains G, then ∇R2(K, ω) is a connected subset
that contains G.
Proof.
Without loss of generality, we may assume that ω is chosen so that krkL2(ω) =
krkH 2(G) for each r ∈ Rat(K). Since G ⊂ K, Rat(K) ⊂ H 2(G). So we can define
a linear operator J : R2(K, ω) → H 2(G) by J(r) = r for each r ∈ Rat(K).
Then kJ(r)k = krkH 2(G) = krkL2(ω). So J is an isometry from R2(K, ω) onto
J[R2(K, ω)] and the later is a closed subspace of H 2(G). Moreover, it is easy to
verify that J(f g) = J(f )J(g) for f, g ∈ R2(K, ω)∩L∞(ω). Now let u ∈ R2(K, ω)∩
L∞(ω) be a non-zero characteristic function. Then [J(u)]2 = J(u2) = J(u). So
we have that J(u − 1) = J(u) − 1 = 0, and hence u = 1. This means that R2(K, ω)
is irreducible and therefore ∇R2(K, ω) is connected.
Lastly, the definition of harmonic measure and the Harnack's inequality implies
that G ⊂ ∇R2(K, ω). So the proof is complete.
The following result extends Theorem 1 in [17].
Theorem 3 Let G = ∪Gi and let ω be a harmonic meausre for G. Suppose that
each Gi is nicely connected and the harmonic measures of the components of G
are mutually singular. Then P 2(ω) = H 2(G) if and only if ∇P 2(ω) = G.
Proof. Sufficiency.
Suppose ∇P 2(ω) = G. By Lemma 1, we have that Gi ⊂ ∇P 2(ωi). This
together with our hypothesis implies that Gi = ∇P 2(ωi). So it follows by Theorem
1 in [17] that P 2(ω∂Gi) = H 2(Gi). According to 5) of Theorem 1, we have that
χ∂Gi, the characteristic function of Gi, is in P 2(ω) for each i ≥ 1. Thus, we
conclude that
P 2(ω) = P 2(ω∂Gi) ⊕ ... ⊕ P 2(ω∂Gi) ⊕ ...
= H 2(G1) ⊕ ... ⊕ H 2(Gi) ⊕ ...
= H 2(G).
Necessity.
The proof is the same as that in the proof of Theorem 1 in [17].
6
Recall that a function g in H ∞(D) is called inner if its boundary values has
modulus 1 almost everywhere on ∂D. An analytic function on D is called outer if
it is of the form
1
2π Z π
−π
eit + z
eit − z
f (z) = α exp{
γ(t)dt},
where α is a constant with modulus 1 and γ is a real-valued function in L1(m).
The classical Hardy space theory guarantees that every function in H 2(D) can
be expressed as a product of an inner and outer functions. The expression is
unique up to a constant with modulus 1 [7].
Let ψ is the Riemann map from G onto D. We call g ∈ H 2(G) an outer function
if g ◦ ψ−1 is outer in H 2(D). The Beurling's theorem implies that φH 2(D) is dense
in H 2(D) if φ is outer in H 2(D).
The following theorem together with Theorem 2 also gives a characterization
for perfectly connected domain.
Theorem 4 Let G = ∪Gi and let ω be a harmonic meausre for G. Suppose that
each Gi is nicely connected and the harmonic measures of the components of G are
mutually singular. Then P ∞(ω) = H ∞(G) if and only if P} is dense in H 2(G)
for each g in H 2(G) for which gχGi is outer in H 2(Gi), i = 1, 2, ...
Proof. Sufficiency.
Suppose that P} is dense in H 2(G) for each g in H 2(G) for which gχGi is outer
in H 2(Gi). We claim that
∇P 2(ν) ⊂ G if ν is a finite measure such that [ν] = [ω].
Note, if we take g=1, then we see that P is dense in H 2(G). So it follows by
Proposition 3 that ∇P 2(ω) = G.
Set y = dν
dω + 1. Then both y and log y belong to L1(ω). It follows by Szego's
theorem that there exist an outer function xi in H 2(Gi) such that xi2 = yχGi
i=1 xi2 ≤ y.
i=1 xi in H 2(G) exists. Set
i=1 xi. Then x belongs to H 2(G) and x2 = y. Define a linear operator A
for each i ≥ 1. Clearly, for each positive integer k, Pk
It follows by the dominated theorem that limk→∞Pk
x =P∞
from P 2(yω) to P 2(ω) densely by A(p) = xp
i=1 xi2 =Pk
for each p ∈ P. Then
Z A(p)2dω =Z xp2dω =Z p2d(yω).
Thus, A is an isometry. By our hypothesis, A has a dense range. So it follows
that A is an unitary operator. For any f ∈ P 2(yω), there is {pn} ⊂ P such that
lim pn = f in P 2(yω). So
lim A(p) = lim pnx = f x = A(f )
in P 2(ω)
and hence cf x is analytic on G. Since x 6= 0 on G, we see that f extends to be
analyitc on G. This implies that G ⊂ ∇P 2(yω). Similarly, we can show that
∇P 2(yω) ⊂ G. Thus, we conclude that
∇P 2(ν) ⊂ ∇P 2(yω) = ∇P 2(ω) = G.
This proves the claim. It is a well-known result (see [20, 12]) that there is a positive
finite measure η such that [η] = [ω] and P ∞(ω) = P 2(η) ∩ L∞(η). So it follows
7
that
H 2(G) ⊃ P 2(ω)
⊃ P ∞(ω)
= P 2(η) ∩ L∞(η)
= H ∞(∇P 2(η)).
Consequently, we have that ∇P 2(η) = G. Hence, it follows by the inequality above
that P ∞(ω) = H ∞(G). So we are done.
Necessity.
Suppose P ∞(ω) = H ∞(G). Let g be a function in H 2(G) such that gχGi is
outer in H 2(Gi) for each i ≥ 1. Pick an arbitrary functin f ∈ H 2(G). For given
ǫ > 0, and each i ≥ 1, since gi is outer, there is a function hi in H ∞(Gi) such that
kf χ∂Gi − gχ∂Gihik <
1
2
(
ǫ
2i ).
Thus, when k is sufficiently large, we have
kf − g
kXi=1
hik ≤ k
(f χGi − gχGihi)k + k
kXi=1
kXi=1
1
2
≤
(
ǫ
2i ) +
ǫ
2
= ǫ.
∞Xi=k+1
f χGsk
Set h =Pk
H 2(G), Finally, since
i=1 hi, then h ∈ H ∞(G) and kf − ghk < ǫ. Hence, H ∞(G)g is dense in
H ∞(G) = P ∞(ω) ⊂ P 2(ω),
it follows clearly that P} is dense in H 2(G). '
3 The main result
For a bounded domain G, a point a in ∂G is said to be removable for H 2(G) if
every function in H 2(G) extends analyticly to a neighborhood of a. Every isolated
point in ∂G is removable for H 2(G) (see [3]).
For a bounded linear operator T , let Lat(T ) denote the lattice of invariant
subspaces of T .
Let X be a linear topological space and let F be a subset of X. We use ∨{F }
to denote the closure of the linear span of F in X.
Lemma 2 Let G be a bounded domain that has no removable boundary points for
H 2(G). Let Mz denote the shift on H 2(G). In order for that every M ∈ Lat(Mz)
there exist a function u ∈ H ∞(G) such that M = ∨{uH 2(G)} it is necessary and
sufficient that G is a perfect connected domain.
Proof. Necessity.
Let K be the union of G with the bounded components of the complement of
G in the plane and let ω denote a harmonic measure of G. Define I: R2(K, ω) →
H 2(G) by I(r) = r, for each r ∈ Rat(K). Then I is continuous and bounded
below. Then N = RI(2(K, ω)) is the closed subspace generated by Rat(K) in
H 2(G).
8
We first show that G must be simply connected. Suppose that G is not simply
connected, then the complement of G contains a component E that is a com-
pact subset. Let W denote ∇R2(K, ω). Since the components of K are simply
connected, it follows by Lemma 1 and Theorem 1 that W is a simply connected
domain that contains G. Moreover, W contains E. Since isolated points in ∂G
are removable for H 2(G), we see that E is a continum. So it follows that there
exists a function, say h, in H 2(G) that has essential singularity points contained
in E.
On the other hand, for each f ∈ N , there is a sequence {fn} in Rat(K) such
that lim fn = f in H 2(G). So this implies that fn → I −1(f ) and hence {fn}
converges to I −1(f ) uniformly on the compact subsets of W . Consequently, we
see that f extends to W analyticly. This contradicts to the fact that h has essential
singularities on E. Hence, G is simply connected.
By our hypothesis, there is a function u ∈ H ∞(G) such that N = ∨{uH 2(G)}.
So we see that u ∈ N . Now, let u = vh, where ψ is the Riemann map from G
onto D, v ◦ ψ−1 is inner and h ◦ ψ−1 is outer in H 2(D). Then we have
N = v ∨ {hf : f ∈ H 2(G)}.
Since h is outer, 1 = limn→∞vαn for some sequence {αn} ∈ H 2(G). Thus,
ψ = limn→∞vαnψ ∈ N . Therefore, there is a sequence of {rn} ∈ R(K) such that
lim rn = ψ in H 2(G). But this implies that there is a subsequence {rnj } such that
lim rnj ◦ ψ−1 = z pointwise almost everywhere on ∂D. Hence ψ−1 is 1-1 a.e. [m]]
on ∂D and thus it follows by the definition that G is a nicely connected domain.
Now, if G was not perfectly connected, then it would follows from Theorem 4
that there is an outer function g ∈ H 2(D) such that {p(g ◦ ψ) : p ∈ P} is not
dense in H 2(G). Set M = ∨{p(g ◦ ψ) : p ∈ P}. Then M ∈ Lat(Mz) and M
is nontrivial. Again, by our hypothesis, there is a function u ∈ H 2(G) such that
M = ∨{uH 2(G)}. Let u = vh, where v◦ψ−1 is inner and h◦ψ−1 is outer in H 2(D).
Then there exists a sequence {fn} in H 2(G) such that g ◦ψ = limn vhfn in H 2(G).
This implies that g ∈ v ◦ ψ−1H 2(D). Applying Beuring's theorem, we conclude
that v ◦ ψ−1 must be a constant. Consequently, we get that M = ∨{gH 2(G)} =
H 2(G), contradicting the fact that M is nontrivial. Therefore, G is a perfectly
connected domain.
Sufficiency.
Let M be a closed subspace of H 2(G) such that zM ⊂ M . Then ψ−1F ⊂
F, in where F = {f ◦ ψ−1 : f ∈ M }. We claim that zF ⊂ F . For this fix a
function h ∈ F . Let g ∈ F ⊥. Then R hgdm = 0 and thus
Z p ◦ ψ−1hgdm = 0
for each p ∈ P.
Since G is perfectly connected, ψ is in the weak-star closure of P. So it follows
that z belongs to the weak-star closure of {p ◦ ψ−1 : p ∈ P}. This implies that
R zhgdm = 0. By the Hahn-Banach theorem, we conclude that zh ∈ F for each
h ∈ F . Now applying the Beurling's theorem, we have that F = uH 2(D) for an
inner function u and therefore, we conclude that M = u ◦ ψH 2(G).
The following is our main theorem.
Theorem 5 Let G be a bounded open subset such that no point in the boundaries
of the components of G is removable for H 2(G).
In order that for each M ∈
9
Lat(Mz) there exist u ∈ H ∞(G) such that M = ∨{uH 2(G)} it is necessary and
sufficient that the following hold:
1) Each component of G is a perfectly connected domain.
2) The harmonic measures of the components of G are mutually singular.
3) The set of polynomials is weak-star dense in H ∞(G).
Moreover, if G satisfies these conditions, then every M ∈ Lat(Mz) is of the form
uH 2(G), where the restriction of u to each of the components of G is either an
inner function or zero.
Remark 2 1) and 2) together insures that H ∞(G) can be embeded in L∞(ω),
where ω is a harmonic measure of G. So H ∞(G) has a weak-star topology. With-
out 1) and 2), 3) may not make sense. However, 1) and 2) do not imply 3). For
example, Let a1 = (1, 1), a2 = (1, −1), a3 = (1, i) and a4 = (1, −i); and let Gi be
the open disks that has radius 1 and is centered at ai, i = 1, 2, 3, 4, respectively.
Set G = ∪4
i=1 ωi, where ωi is the harmonic measure evaluated
at ai. Then G satisfies conditions 1) and 2).
i=1Gi and set ω =P4
Let W be the bounded component of the interior of the complement of G and
let µ = ω∂W . Then µ is a multiple of the arclength on ∂W . So it follows that
W = ∇P 2(µ) ⊂ ∇P 2(ω).
Since W ∩ G = ∅, it follows by Proposition 3 that P 2(ω) 6= H 2(G). Consequently,
P ∞(ω) 6= H ∞(G) (applying Theorem 4 by taking g = 1).
Proof. Necessity. Suppose that every M ∈ Lat(Mz) is of the form ∨{uH ∞(G)}
for some u ∈ H 2(G). Let G1, G2, ...Gn, ... be all of the components of G and let
ωn be the harmonic measure for Gn so that for each r ∈ Rat(Gn), krkH 2(Gn) =
krkL2(ωn) for n = 1, 2, ... Fix such an integer n and let us consider H 2(Gn).
Suppose that N is a closed subspace of H 2(Gn) such that zH 2(Gn) ⊂ H 2(Gn).
Extend each f ∈ N to be a function in H 2(G) by letting its value to be zero off
Gn. Then χGn N would be a closed invariant subspace for the shift on H ∞(G).
Thus, by our hypothesis χGnN = ∨{unH 2(G)} for some un ∈ H 2(G). Therefore,
N = χGn ∨ {χGnunH 2(G)}
= ∨{(χGn un)H 2(Gn)}.
Since (χGn un) is in H 2(Gn), it follows by Lemma 2 that Gn must be a perfectly
connected domain.
Now set ω = P∞
1
2n ωn. Then ω is a harmonic measure for G. Choose a
compact subset K that contains K such that the complement of K has finitely
many components. Let W = ∇R2(K, ω). We want to show that W = G. For this
let I be the linear operator from R2(K, ω) to H 2(G) induced by I(f ) = f for each
f ∈ Rat(K). Then, for every f ∈ Rat(K)
n=1
Z f 2dω = Xn=1
Z f 2d(
1
2n ωn) = Xn=1
1
2n kf k2
H 2(Gn) = kf k2
H 2(G).
So I is an isometry. Clearly, I[R2(K, ω)] is a closed invariant subspace for the
shift on H 2(G). By our hypothesis, there exists a function u ∈ H ∞(G) such that
I[R2(K, ω)] = ∨{uH 2(G)}.
For each n, since χGn ∈ H 2(G), it follows that uχGn ∈ I[R2(K, ω)]. This implies
that uχGn ∈ H ∞(W ). Let U be the component of W that contains Gn. Then, we
10
I(∆i) = c∆i = χWi
for each i.
have either Gn = U or uχGn is identically zero on U . But the later implies that
I[R2(K, ω)] = ∨{uH 2(G − Gn)} and this is clearly impossible. Hence we have
that U = Gn.
Let {Wi}∞
i=1 be the collection of the components of W . By Theorem 1 , there
exists a Borel partition {∆i}∞
i=0 of the support of ω such that
R2(K, ω) = R2(W 1, ω∆1)... ⊕ R2(W i, ω∆i) ⊕ ...
Moreover, W i ⊃ ∆i and (ω∆i)∂Wi ≪ ωWi , where ωWi is a harmonic measure
for Wn; the map f → f is an isomorphism and weak-star homeomorphism from
R2(W i, ω∆i) ∩ L∞(ω∆i) onto H ∞(Wi).
Since χ∆i ∈ R2(K, ω) for i ≥ 1 and since {∆i}∞
i=1 are pairwise disjoint, it
Note that f = I(f )
follows from the bounded convergence theorem that 1 = limPi χ∆i in R2(K, ω).
so that fn → χ∆i in L2(ω). Then fn → dχ∆i uniformly compact subset of W .
for every f ∈ Rat(K). Choose a sequence {fn} in Rat(K)
Thus, it follows that
But I(∆i) ∈ H ∞(G) and thus we conclude that χWi belongs to H 2(G) for each i.
This means that χWi must be equal to χGj for some j. Consequently, Wi = Gj.
Therefore, it follows that W = G.
Now, since ∪Wi = W = ∪Gi and and since each Gi is a component of W ,
rearranging the indexes if necessary, we may assume that Wi = Gi for each i.
From the definition of ω, it is clear that
ωi = ωGi ≪ ω∂Gi.
On the other hand, according to Theorem 1, ω∂Gi ≪ ωGi. Hence, we have
that [ωi] = [ω∂Gi] for each i ≥ 1. Since ∆i ⊂ ∂Gi and since ω∆i and ω∆j
are mutually singular, it follows that ωi and ωj are mutually singular if i 6= j.
Therefore, we conclude that the harmonic measures of the components of G are
mutually singular.
Lastly, we want to show P ∞(G) = H ∞(G). Suppose the contrary. It follows
from Theorem 4 that there exist an outer function g in H 2(G) such that P} is
not desne in H 2(G). Set X = ∨{pg : p ∈ P}. Then X is a nontrivial invariant
subspace of Mz. So there is u ∈ H ∞(G) such that X = ∨{uH 2(G)}. Let u = vh
be the inner and outer factorization, then X = vH 2(G). Let {fn} ⊂ H 2(G) be a
sequence such that lim vfn = g in H 2(G). Thus, we have lim gχ∂Gi = vf χ∂Gi for
each i. Since gχ∂Gi is outer in H 2(Gi), it follows by the Buerling's theorem that
vχ∂Gi is a constant. Consequently, we get that M = H 2(G). This contradiction
shows that P must be weak-star dense in P ∞(G) = H ∞(G).
Sufficiency.
Suppose G is such that each of its components is perfectly connected, the
harmonic measures of its components are mutually singular and P ∞(ω) = H ∞(G).
By Proposition 3, we have P 2(ω) = H 2(G). In particular, χ∂Gi ∈ P 2(ω) for each
i ≥ 1. Thus, we have
P 2(ω) = P 2(ω∂G1) ⊕ ... ⊕ P 2(ω∂Gi) ⊕ ...
= H 2(G1) ⊕ ... ⊕ H 2(Gi) ⊕ ...
= H 2(G).
Suppose that M is a closed z invariant subspace of H 2(G). Pick a function a ∈ M
and set Ma = ∨{pa : pP}. Let a = uh be the inner and outer factorization. Then
11
it follows by Theorem 4 that Ma = uH 2(G). Since hχGi ∈ H 2(G), we see that
mχGi = uvχGi ∈ M for each i ≥. Set Mi = M χGi. Then we have
M ⊂ M1 + ... + Mi + ... ⊂ M1 + ... + Mi + ...
⊂ M1 ⊕ ... ⊕ Mi ⊕ ...
⊂ M.
Hence, Mi is a closed subspace for each i ≥ 1. Evidently, zMi = χGi zM ⊂ Mi.
Applying Lemma 2, we have that χGn M = unH 2(Gn), where un = 0 if M = {0}
and otherwise un ∈ H ∞(Gn) with un = 1 a.e.
[ωn] on ∂Gn. Extend u to be
a function in H 2(G) by defining its value to be zero off Gn. Then, for integers
n ≥ m ≥ 1, we have that
k
nXi=1
ui −
mXi=1
uikH 2(G) = (
nXi=m
nXi=m
≤ (
1
2i kuikH 2(Gi))
1
2
1
2i )
1
2 → 0
as m, n → ∞.
Thus, u = limnPn
i=1 ui exists. Moreover,
M = u1H 2(G1) ⊕ ... ⊕ unH 2(Gn) ⊕ ...
= uH 2(G1) ⊕ ... ⊕ vH 2(Gn) ⊕ ...
= uH 2(G).
Clearly, we have either u = 1 or u = 0 on ∂Gn for each n.
References
[1] A. Aleman, S. Richter, C. Sundberg, Beurling's theorem for the Bergman
space, Acta Math. 177 (1996), 275-310.
[2] C. Apostol, H. Bercovici, C. Foias, C. Pearcy, Invariant subspaces, dilation
theory, and the structure of the predual of a dual algebra, J. Functional
Analysis, 63 (1985), 369-404.
[3] S. Axler, Harmonic functions from a complex analysis viewpoint, American
Math. Monthly 93 (1986), 246-258.
[4] A. Beurling, On two problems concerning linear transformations in Hilbert
spaces, Acta Math., Vol 81 (1949), 239-255.
[5] A. Davie, Dirichlet algebras of analytic functions, J. Functional Analysis
6(1967), 348-356.
[6] P. Duren, Theory of H 2 spaces, Academic Press, New York, 1970.
[7] K. Hoffman, Banach spaces of analytic functions, Prentice-Hall, Englewood
Cliffs, N.J., 1962.
[8] T. Gamelin, Uniform Algebras, Prentice Hall, Englewood Cliffs, N.J., 1969.
12
[9] T. Gamelin, J. Garnett, Constructive techniques in rational approximation,
Trans. Amer. Math. Soc. 143 (1969), 187-200.
[10] I. Glicksburg, The abstract F. and M. Riese theorem, J. Functional Anal-
ysis 1 (1967), 109-122.
[11] H. Hedenhalm, S. Richter, K. Seip, Interpolating sequence and invariant
subspaces of given index in the Bergman spaces, J. Reine Angew. Math.
477 (1996) 13-30.
[12] J. McCarthy, Analytic structures for subnormal operators, Integral Equa-
tion and Operator Theorey 13(1990), 251-270.
[13] Z. Qiu, Density of polynomials, Houston Journal of Mathematics, Vol. 21,
Number 1 (1995), 109-118. 27 (1997).
[14] Z. Qiu, On quasisimilarity of subnormal operators, Science in China, Series
A Math. 2007 Vol. 50, No. 3, 305-312.
[15] Z. Qiu, Carleson measure and polynomial approximation, Chinese Annals
of Mathematics, Series A, 2007 Vol. 28, No. 2, 159-167.
[16] Z. Qiu, The structure of the closure of the rational functions in L2(µ), to
appear.
[17] Z. Qiu Density of rational functions in Hardy and Bergman spaces, Indian
Journal of Pure and Applied Mathematics
[18] D. Sarason, Weak-star generator of H ∞, Pacific J. Math., 17 (1966), 519-
528.
[19] D. Sarason, On order of a simply connected domain, Michigan Math. J., 15
(1968), 129-133.
[20] J. Thomson, Approximation in the mean by polynomials, Annals of Math-
ematics 133 (1991), 477-507.
13
|
1306.4093 | 1 | 1306 | 2013-06-18T08:02:49 | Factorizations of EP Banach space operators and EP Banach algebra elements | [
"math.FA"
] | EP Banach space operators and EP Banach algebra elements are characterized using different kinds of factorizations. The results obtained generalize well-known characterizations of EP matrices, EP Hilbert space operators and EP $C^*$-algebra elements. Furthermore, new results that hold in these contexts are presented. | math.FA | math |
Factorizations of EP Banach space operators
and EP Banach algebra elements
Enrico Boasso
Abstract
EP Banach space operators and EP Banach algebra elements are characterized using
different kinds of factorizations. The results obtained generalize well-known character-
izations of EP matrices, EP Hilbert space operators and EP C ∗-algebra elements. Fur-
thermore, new results that hold in these contexts are presented.
Keywords: Moore-Penrose inverse, EP Banach space operator, EP Banach algebra elem-
ent, factorization.
1.
Introduction
A complex matrix T is said to be EP, if it commutes with its Moore-Penrose inverse T †.
Moreover, the notion under consideration was extended to Hilbert space operators and C ∗-
algebra elements, and it consists in a generalization of normal matrices and operators, see the
introductory section of [6]. Furthermore, thanks to the concept of hermitian Banach algebra
element, in [14] V. Rakocevi´c extended the notion of Moore-Penrose inverse to elements of
a Banach algebra, which led to study EP Banach space operators and EP Banach algebra
elements, see [1].
The relationships among EP matrices and operators and the product operation have been
object of a particular attention. On the one hand, several articles studied when the product
of two EP matrices, Hilbert or Banach space operators, or elements of a C ∗-algebra or a
Banach algebra is again EP, see [1, 9] and the bibliography of these articles. On the other
hand, D. Drivarialis, S. Karamasios and D. Pappas in [6] and D. S. Djordjevi´c, J. J. Koliha
and I. Straskaba in [4] have recently characterized EP Hilbert space operators and EP C ∗-
algebra elements respectively through several different factorizations. Note that one of the
main lines of research concerning EP matrices and EP operators consists in characterizing
them through factorizations.
The objective of the present article is to characterize EP Banach space operators and
EP Banach algebra elements using factorizations to extend results of [6, 4] to the mentioned
contexts. Actually, three different kind of factorizatiosn will be considered.
It is worth
noticing that due to the lack of involution on a Banach algebra, and in particular on the
Banach algebra of bounded and linear maps defined on a Banach space, the proofs not only
are different from the ones known for matrices or Hilbert space operators, but also they give
a new insight into the cases where the involution does exist. Furthermore, thanks to the
approach developed in this work, new results in the frames both of Banach algebras and of
C ∗-algebras will be presented.
Enrico Boasso
2
2. Preliminary definitions and results
From now on X and Y will denote Banach spaces and L(X, Y ) will stand for the Banach
algebra of all bounded and linear maps defined on X with values in Y . As usual, when
X = Y , L(X, Y ) will be denoted by L(X). In addition, if T ∈ L(X, Y ), then N (T ) and R(T )
will stand for the null space and the range of T respectively.
On the other hand, A will denote a unital Banach algebra and e ∈ A will stand for the unit
element of A. If a ∈ A, then La, Ra : A → A will denote the maps defined by left and right
multiplication respectively, that is, La(x) = ax and Ra(x) = xa, where x ∈ A. Moreover, the
following notation will be used:
N (La) = a−1(0), R(La) = aA, N (Ra) = a−1(0), R(Ra) = Aa.
Recall that an element a ∈ A is said to be regular, if it has a generalized inverse, namely
if there exists b ∈ A such that a = aba. Furthermore, a generalized inverse b of a regular
element a ∈ A will be said to be normalized, if b is regular and a is a generalized inverse of
b, equivalently, a = aba and b = bab. Note that if b is a generalized inverse of a, then c = bab
is a normalized generalized inverse of a.
Next follows the key notion in the definition of Moore-Penrose invertible Banach algebra
elements, see [16].
Definition 2.1. Given a unital Banach algebra A, an element a ∈ A is said to be hermitian,
if k exp(ita) k= 1, for all t ∈ R.
Recall that if A is a C ∗-algebra, then a ∈ A is hermitian if and only if a is self-adjoint,
see [2, Proposition 20, Chapter I, Section 12]. Moreover, H = {a ∈ A : a is hermitian} ⊆ A
is a closed linear vector space over the real field, see [16, 5]. Since A is unital, e ∈ H, which
implies that a ∈ H if and only if e − a ∈ H. As regard equivalent definitions and the main
properties of hermitian Banach algebra elements and hermitian Banach space operators, see
[2, 5, 10, 12, 16].
In [14] V. Rakocevi´c introduced the notion of Moore-Penrose invertible Banach algebra
element. Next the definition of this object will be considered.
Definition 2.2. Let A be a unital Banach algebra and consider a ∈ A. If there exists a
normalized generalized inverse x ∈ A of a such that ax and xa are hermitian elements of A,
then x will be said to be the Moore-Penrose inverse of a and it will be denoted by a†.
In the conditions of Definition 2.2, recall that according to [14, Lemma 2.1], there exists at
most one Moore-Penrose inverse of a ∈ A. In addition, if a ∈ A has a Moore-Penrose inverse,
then (a†)† exists. In fact, (a†)† = a. Concerning the properties of the Moore-Penrose inverse
in the frames of Banach space operators and Banach algebras, see [14, 15, 1], in C ∗-algebras
see [7, 8, 11], for the original definition see [13].
In order to study the factorization that will be considered in the next section, the notion
of Moore-Penrose inverse for operators defined between different Banach spaces need to be
introduced. First of all, however, some preliminary results will be recalled.
Enrico Boasso
3
Remark 2.3. Let X and Y be two Banach spaces and consider T ∈ L(X, Y ) and S ∈ L(Y, X)
such that S is a normalized generalized inverse of T , i.e., T = T ST and S = ST S. Then, it
is not difficult to verify the following facts:
(i) ST ∈ L(X) is an idempotent, N (ST ) = N (T ), R(ST ) = R(S) and X = N (T ) ⊕ R(S).
(ii) T S ∈ L(Y ) is an idempotent, N (T S) = N (S), R(T S) = R(T ) and Y = N (S) ⊕ R(T ).
Definition 2.4. Let X and Y be two Banach spaces and consider T ∈ L(X, Y ). The operator
T will be said to be Moore-Penrose invertible, if there exists S ∈ L(Y, X) such that T = T ST ,
S = ST S, and ST ∈ L(X) and T S ∈ L(Y ) are hermitian operators.
Before going on some basic results concerning the objects of Definition 2.4 will be consid-
ered.
Lemma 2.5. Let X and Y be two Banach spaces and consider T ∈ L(X, Y ) and Si ∈ L(Y, X)
such that T and Si complies the four conditions of Definition 2.4, i = 1, 2. Then S1 = S2.
Proof. Adapt the proof of [14, Lemma 2.1] to the conditions of the Lemma.
Since according to Lemma 2.5 T ∈ L(X, Y ) has at most one Moore-Penrose inverse, when
the Moore-Penrose inverse of T exists, it will be denoted by T †. On the other hand, note
that in the next proposition, given T ∈ L(X, Y ), X and Y Banach spaces, T ∗ ∈ L(Y ∗, X ∗)
will denote the adjoint map of T and X ∗ and Y ∗ will stand for the dual space of X and Y
respectively.
Proposition 2.6. Let X and Y be two Banach spaces and consider T ∈ L(X, Y ).
(i) Necessary and suffcicient for T † to exist is the fact that there are two hermitian idempo-
tents P ∈ L(X) and Q ∈ L(Y ) such that N (P ) = N (T ) and R(Q) = R(T ).
(ii) If T † ∈ L(Y, X) exists, then (T ∗)† ∈ L(X ∗, Y ∗) also exists, moreover, (T ∗)† = (T †)∗.
(iii) Suppose that T † ∈ L(Y, X) exists, and let X ′ ⊆ X and Y ′ ⊆ Y be two closed vector
subspaces such that T (X ′) ⊆ Y ′ and T †(Y ′) ⊆ X ′. If T ′ = T Y ′
=
T † X ′
(iv) In the conditions of statement (iii), if T : X/X ′ → Y /Y ′ and T † : Y /Y ′ → X/X ′ denote
the operators induced by T and T † respectively, then the Moore-Penrose inverse of T exists
and ( T )† = T †.
Y ′ ∈ L(Y ′, X ′), then the Moore-Penrose inverse of T ′ exists and (T ′)† = T †′
X ′∈ L(X ′, Y ′) and T †′
.
Proof. Adapt the proofs of [1, Theorem 6], [1, Proposition 7] and [1, Theorem 10] to the
present situation.
The final point of this section concerns the definition of EP Banach algebra elements.
Definition 2.7. Let A be a unital Banach algebra. Given a ∈ A, a will be said to be EP, if
a† exists and aa† = a†a.
Properties, characterizations and other facts regarding EP Banach space operators and
EP Banach algebra elements were studied in [1]. In the following remark some of the most
relevant results on these objects will be recalled.
Enrico Boasso
4
Remark 2.8. Let A be a unital Banach algebra and consider a ∈ A.
(i) Note that a ∈ A is EP if and only if a† is EP.
(ii) According to [1, Remark 12], necessary and sufficient for a ∈ A to be EP is the fact that
La ∈ L(A) is EP.
(iii) Let A = L(X), X a Banach space, and considet T ∈ L(X). Then, according to [1,
Theorem 16], T is EP if and only if R(T ) = R(T †) or N (T ) = N (T †).
3. Factorization of the form a = bc
In this section, given a unital Banach algebra A, EP elements of the form a = bc will
be characterized, where a, b, c ∈ A, a is Moore-Penrose invertible, b−1(0) = 0 and cA = A.
Concerning this kind of factorization, see the introductory section of [6]. In addition, compare
the results of this section with [6, section 5] and [4, sections 1.3, 2.3]. However, to prove the
main results of this section, some preliminary facts must be considered. Note that in what
follows the identity map on the Banach space X (respectively Y ) will be denoted by I ∈ L(X)
(respectively I ′ ∈ L(Y )).
Proposition 3.1. Let X and Y be two Banach spaces and consider T ∈ L(X), C ∈ L(X, Y )
and B ∈ L(Y, X) such that C is surjective, B is injective and T = BC. Suppose, in addition,
that T † exists. Then, the following statements hold.
(i) There exists B† ∈ L(X, Y ) such that B†B = I ′,
(ii) there exists C † ∈ L(Y, X) such that CC † = I ′,
(iii) T † = C †B†, T T † = BB†, T †T = C †C, B† = CT † and C † = T †B.
Proof. According to [1, Theorem 6(ii)], there exist two hermitian idempotents P and Q ∈
L(X) such that N (P ) = N (T ) and R(Q) = R(T ). Since N (B) = 0 and R(B) = R(T ),
according to Proposition 2.6(i), B† ∈ L(X, Y ) exists. Actually, I ′ ∈ L(Y ) and Q ∈ L(X)
are two hermitian idempotents such that N (B) = N (I ′) and R(B) = R(Q). Furthermore,
according to Remark 2.3, R(I ′ − B†B) = N (B†B) = N (B) = 0, i. e., B†B = I ′.
Similarly, since R(C) = Y and N (C) = N (T ), according to Proposition 2.6(i), C † ∈
In fact, I ′ ∈ L(Y ) and P ∈ L(X) are two hermitian idempotents such
L(Y, X) exists.
that R(C) = R(I ′) and N (C) = N (P ). Moreover, note that according to Remark 2.3,
R(CC †) = R(C). However, since CC † and I ′ are hermitian idempotents of L(Y ) whose
ranges coincide, according to [12, Theorem 2.2], CC † = I ′.
Next consider S = C †B† ∈ L(X). A straightforward calculation proves that T = T ST ,
S = ST S, T S = BB† and ST = C †C. However, since BB† and C †C are two hermitian
idempotents, according to [14, Lemma 2.1], S = T †. The remaining two identities can be
derived from statements (i) and (ii).
In the following theorems Moore-Penrose invertible operators of the form T = BC will
be characterized.
Theorem 3.2. In the conditions of Proposition 3.1, the following statements are equivalent.
(i) T is an EP operator,
(ii) BB† = C †C,
(iii) N (B†) = N (C),
(iv) R(B) = R(C †).
Enrico Boasso
5
Proof. According to Proposition 3.1(iii), statements (i) and (ii) are equivalent. In addition,
since N (T ) = N (C), N (T †) = N (B†), R(T ) = R(B) and R(T †) = R(C †), according to [1,
Theorem 16], T is EP if and only if statements (iii)-(iv) holds.
Remark 3.3. Note that when H is a Hilbert space and T ∈ L(H), in the conditions of [6,
Theorem 5.1], necessary and sufficient for T to be EP is the fact that R(B) = R(C ∗). In
fact, apply Theorem 3.2 and use that R(C ∗) = R(C †), see section 2 of [6].
Theorem 3.4. In the conditions of Proposition 3.1, necessary and sufficient for T to be EP
is that one the following statements holds.
(i) (I − C †C)B = 0 and C(I − BB†) = 0,
(ii) B†(I − C †C) = 0 and C(I − BB†) = 0,
(iii) (I − C †C)B = 0 and (I − BB†)C † = 0,
(iv) B†(I − C †C) = 0 and (I − BB†)C † = 0,
(v) C(I − BB†) = 0 and B†(I − C †C)B = 0,
(vi) B†(I − C †C) = 0 and C(I − BB†)C † = 0.
Proof. Suppose that T is EP. Then, according to Theorem 3.2, R(B) = R(C †) and N (B†) =
N (C). However, since N (I −C †C) = R(C †C) = R(C †) and R(I −BB†) = N (BB†) = N (B†)
(Remark 2.3), statement (i) holds.
On the other hand, note that the conditions in statement (i) are equivalent to R(T ) =
R(B) ⊆ N (I − C †C) = R(C †) = R(T †) and N (T †) = N (B†) = R(I − BB†) ⊆ N (C) = N (T ).
However, since according to Remark 2.3 N (T †)⊕R(T ) = X = N (T )⊕R(T †), it is not difficult
to prove that N (T ) = N (T †) and R(T ) = R(T †). In particular, according to [1, Theorem
16], T is EP.
The equivalence among the condition of being EP and statements (ii)-(iv) can be proved
using similar arguments.
Next consider statement (v). According to what has been proved, if T is EP, then C(I −
BB†) = 0. In addition, if T is EP, then T †T = T †T †T T , equivalently, C †C = C †B†C †CBC.
Now well, according to Proposition 3.1(i)-(ii), B†B = I ′ = B†C †CB, which is equivalent to
B†(I − C †C)B = 0.
On the other hand, if statement (v) holds, then according to what has been proved,
N (B†) ⊆ N (C). However, since X = N (B†) ⊕ R(B) (Remark 2.3), if in addition B†(I −
C †C)B = 0, then a straightforward calculation proves that B†(I − C †C) = 0. Since R(I −
C †C) = N (C) (Remark 2.3), N (C) ⊆ N (B†). Therefore, according to Theorem 3.2(ii), T is
EP.
The equivalennce between the condition of being EP and statement (vi) can be proved in
a similar way.
Theorem 3.5. In the conditions of Proposition 3.1, the following statements are equivalent.
(i) T is an EP operator,
(ii) there exists an isomorphism U ∈ L(Y ) such that C = U B†,
(iii) there exists an injective map U1 ∈ L(Y ) such that C = U1B†,
(iv) there exist U2, U3 ∈ L(Y ) such that C = U2B† and B† = U3C,
(v) there exists an isomorphism W ∈ L(Y ) such that B = C †W ,
(vi) there exists a surjective map W1 ∈ L(Y ) such that B = C †W1,
Enrico Boasso
6
(vii) there exist W2, W3 ∈ L(Y ) such that B = C †W2 and C † = BW3,
(viii) there exist H1, H2 ∈ L(Y ) such that B = C †H2 and C = H1B†,
(ix) there exist K1, K2 ∈ L(Y ) such that C † = BK2 and B† = K1C,
(x) there exists an injective map S1 ∈ L(Y ) such that B† = S1C,
(xi) there exists a surjective map S2 ∈ L(Y ) such that C † = BS2.
Proof. Suppose that T is EP and define U = CT C † ∈ L(Y ). According to Proposition 3.1(ii)
and the second identity of Theorem 3.4(i), U = CB and U B† = C. To prove that U ∈ L(Y )
is an isomorphic map, consider Z = B†C † ∈ L(Y ). According to the second identity of
Theorem 3.4(i) and Proposition 3.1(ii), U Z = I ′. In addition, according to the first identity
of Theorem 3.4(i) and Proposition 3.1(i), ZU = I ′.
It is clear that statement (ii) implies statement (iii). On the other hand, if statement (iii)
holds, then N (C) = N (B†), which, according to Theorem 3.2(iii), is equivalent to statement
(i).
Clearly, statement (ii) implies statement (iv), which in turn implies that N (C) = N (B†).
Consequently, according to Theorem 3.2(iii), T is an EP operator.
On the other hand, if T is EP, then, according to the first identity of Theorem 3.4(i),
C †U = C †CB = B. As a result, statement (v) holds with W = U .
Statement (v) implies statement (vi), which in turn implies that R(B) = R(C †). However,
according to Theorem 3.2(iv), T is EP.
In addition, statement (v) implies statement (vii), which in turn implies that R(B) =
R(C †). In particular, according to Theorem 3.2(iv), T is EP.
It is clear that statements (iv) and (vii) implies statements (viii) and (ix). To prove that
both statement (viii) and statement (ix) implies that T is EP, consider the decomposition
of X defined by T T † = BB† and T †T = C †C, i.e., X = N (B†) ⊕ R(B) = N (C) ⊕ R(C †).
Now well, if statement (viii) (respectively statement (ix)) holds, then N (B†) ⊆ N (C) and
R(B) ⊆ R(C †) (respectively N (C) ⊆ N (B†) and R(C †) ⊆ R(B)). However, according the
decompositions of X, if statement (viii) or statement (ix) holds, then it is not difficult to
prove that N (B†) = N (C) and R(B) = R(C †). Therefore, according to Theorem 3.2, T is
EP.
Next, statement (ii) implies statement (x). On the other hand, if statement (x) holds,
since B† = S1C, Y = R(B†) ⊆ R(S1). However, since N (S1) = 0, statement (ii) holds.
Similarly, statement (v) implies statement (xi). To prove the converse, if C † = BS2, then,
according to Proposition 3.1, N (S2) = 0. Since S2 is surjective, statement (v) holds.
Next the Banach algebra case will be studied. Firstly some preliminary facts will be
considered.
Proposition 3.6. Let A be a unital Banach algebra and consider a, b, c ∈ A such that a†
exists, a = bc, b−1(0) = 0 and cA = A. Then, the following statements hold.
(i) There exists b† ∈ A such that b†b = e,
(ii) there exists c† ∈ A and cc† = e,
(iii) a† = c†b†, aa† = bb†, a†a = c†c, b† = ca† and c† = a†b.
Proof. Consider the maps La, Lb, Lc ∈ L(A). It is clear that La = LbLc, N (Lb) = 0 and
R(Lc) = A. What is more, according to [1, Remark 5(ii)], (La)† exists and (La)† = La† ∈
Enrico Boasso
7
L(A). Therefore, according to Proposition 3.1, (Lb)† exists and (Lb)†Lb = IA, (Lc)† exists
and Lc(Lc)† = IA, La† = (Lc)†(Lb)†, (Lb)† = Lca† and (Lc)† = La†b, where IA ∈ L(A) denotes
the identity map of A.
Next consider c′ = (Lc)†(e) = a†b. Since Lc(Lc)† = IA, cc′ = e. As a result, c = cc′c and
c′ = c′cc′. In addition, c′c = a†bc = a†a, which is a hermitian idempotent. Consequently,
according to [14, Lemma 2.1], c† exists, c† = c′ = a†b and cc† = e.
Let b′ = ca†. Since c† = a†b and cc† = e, b′b = e. In particular, b = bb′b and b′ = b′bb′.
Moreover, bb′ = bca† = aa†, which is a hermitian idempotent. Thus, according to [14, Lemma
2.1], b† exists, b† = b′ = ca† and b†b = e.
Finally, define a′ = c†b†. Then, since b†b = e and cc† = e, aa′a = a, a′aa′ = a′, aa′ = bb†
and a′a = c†c. Therefore, according to [14, Lemma 2.1], a† = a′ = c†b†.
In the following theorem, given A a unital Banach algebra, A−1 will stand for the set of
all invertible elements of A.
Theorem 3.7. In the conditions of Proposition 3.6, the following statements are equivalent.
(i) a ∈ A is EP,
(iii) (b†)−1(0) = c−1(0),
(v) b−1(0) = (c†)−1(0),
(vii) (e − c†c)b = 0 and c(e − bb†) = 0,
(ix) (e − c†c)b = 0 and (e − bb†)c† = 0,
(xi) c(e − bb†) = 0 and b†(e − c†c)b = 0,
(xiii) ∃ x ∈ A−1 : c = xb†,
(xv) ∃ z1, z2 ∈ A : c = z1b† and b† = z2c,
(xvii) ∃ v ∈ A : vA = A and b = c†v,
(xix) ∃ h1, h2 ∈ A : b = c†h2 and c = h1b†,
(xxi) ∃s1 ∈ A : s−1
(xxiii) bA−1 = c†A−1,
(xxiv) a ∈ c†A ∩ Ab†,
1 (0) = 0 and b† = s1c,
(ii) bb† = c†c,
(iv) bA = c†A,
(vi) Ac = Ab†,
(viii) b†(e − c†c) = 0 and c(e − bb†) = 0,
(x) b†(e − c†c) = 0 and (e − bb†)c† = 0,
(xii) b†(e − c†c) = 0 and c(e − bb†)c† = 0,
(xiv) ∃ y ∈ A : y−1(0) = 0 and c = yb†,
(xvi) ∃ u ∈ A−1 : b = c†u,
(xviii) ∃ w1, w2 ∈ A : b = c†w1 and c† = bw2,
(xx) ∃ k1, k2 ∈ A : c† = bk2 and b† = k1c,
(xxii) ∃s2 ∈ A : s2A = A and c† = bs2,
(xxiv) A−1c = A−1b†,
(xxvi) a† ∈ bA ∩ Ac.
Proof. Let T = La, B = Lb and C = Lc ∈ L(A). Then, according to Proposition 3.6 and [1,
Theorem 5(ii)], T , B and C are Moore-Penrose invertible operators, what is more, T † = La†,
B† = Lb† and C † = Lc†. Recall also that according to [1, Remark 12], a ∈ A is EP if and only
if La ∈ L(A) is EP. Therefore, since T = BC, N (B) = 0, R(C) = A, according to Theorem
3.2 and Theorem 3.4, it is easy to prove that statements (i)-(iv) and (vii)-(xii) are equivalent.
On the other hand, according to Proposition 3.6(i)-(ii), a−1(0) = b−1(0) and (a†)−1(0) =
(c†)−1(0). Consequently, according to [1, Theorem 18(viii)], statements (i) and (v) are equiva-
lent. In addition, a straightforward calculation proves that Ra = Ra†a = Rc†c = Rc. Similarly,
Ra† = Raa† = Rbb† = Rb†. Consequently, Aa = Ac and Aa† = Ab†. Therefore, according to
[1, Theorem 18(ix)], statements (i) and (vi) are equivalent.
Enrico Boasso
8
To prove that statement (i) and statements (xiii) -(xxii) are equivalent, use as before the
multiplication operators and adapt the proof of Theorem 3.5 to the present situation. Note
that x = u = cb.
Observe that statement (xxiii) (respectively (xxiv)) is equivalent to statement (xvi) (re-
spectively (xiii)).
Finally, according to [1, Theorem 18(xiii)-(xiv)], statement (i) and statements (xxv)-(xxvi)
are equivalent.
In the frame of unital C ∗-algebras, the results of Theorem 3.7 can be reformulated using
the adjoint instead of the Moore-Penrose inverse. However, to this end some preparation is
needed.
Remark 3.8. Recall that given a unital C ∗-algebra A and x ∈ A, then x†A = x∗A,
(x†)−1(0) = (x∗)−1(0), Ax† = Ax∗ and (x†)−1(0) = (x∗)−1(0), see [9, Lemma 1.5].
Compare the following lemma with [9, Lemma 1.5].
Lemma 3.9. Let A be a unital C ∗-algebra and consider a ∈ A such that a is Moore-Penrose
invertible.
(i) If p = aa†, then v = e − p + (a†)∗a† ∈ A−1, a† = a∗v and aa∗v = vaa∗ = p.
(ii) If q = a†a, then w = e − q + a†(a†)∗ ∈ A−1, a† = wa∗ and wa∗a = a∗aw = q.
Suppose, in addition, that a = bc, where b and c are such that b−1(0) = 0 and cA = A.
Then,
(iii) b†(b†)∗ ∈ A−1 and (b†(b†)∗)−1 = b∗b,
(iv) (c†)∗c† ∈ A−1 and ((c†)∗c†)−1 = cc∗,
(v) vb = (b†)∗(c†)∗c† and cw = b†(b†)∗(c†)∗.
Proof. Note that (a†)∗a† and aa∗ belong to the subalgebra pAp. What is more, (a†)∗a†aa∗ =
aa∗(a†)∗a† = p. Consequently, v ∈ A−1 and a straightforward calculation proves that a† =
a∗v and aa∗v = vaa∗ = p.
Interchanging a with a∗, statement (ii) can be derived from statement (i).
Statements (iii)-(iv) can be easily derived from Proposition 3.6(i)-(ii).
Note that v = e − bb† + (b†)∗(c†)∗c†b†. Consequently, according to Proposition 3.6(i), vb =
(b†)∗(c†)∗c†. A similar argument, using Proposition 3.6(ii), proves the remaining identity.
Theorem 3.10. Let A be a unital C ∗-algebra and consider a ∈ A such that a† exists. Suppose
that there exist b, c ∈ A such that a = bc, b−1(0) = 0 and cA = A. Then, the following
statements are equivalent.
(i) a ∈ A is EP,
(iii) (b∗)−1(0) = c−1(0),
(v) b−1(0) = (c∗)−1(0),
(vii) bA−1 = c∗A−1,
(ix) ∃ x ∈ A−1 : c = xb∗,
(xi) ∃ z1, z2 ∈ A : c = z1b∗ and b∗ = z2c,
(xiii) ∃s1 ∈ A : s−1
1 (0) = 0 and b∗ = s1c,
(ii) a ∈ c∗A ∩ Ab∗,
(iv) bA = c∗A,
(vi) Ac = Ab∗,
(viii) A−1c = A−1b∗,
(x) ∃ y ∈ A : y−1(0) = 0 and c = yb∗,
(xii) ∃ v ∈ A : vA = A and b = c∗v,
(xiv) ∃s2 ∈ A : s2A = A and c∗ = bs2.
Enrico Boasso
9
Proof. The equivalence among statements (i)-(vi) can be derived from Theorem 3.7 and the
relationships among the null spaces and the ranges of the operators Lx† and Lx∗ ∈ L(A), and
of the operators Rx† and Rx∗ ∈ L(A), for x = b and c (Remark 3.8).
Note that statement (ix) is equivalent to the fact that there exists z ∈ A−1 such that
b = c∗z, which in turn is equivalent to statements (vii) and (viii). Now well, if there exists
z ∈ A−1 such that b = c∗z, then bA = c∗A = c†A which, according to Theorem 3.7(iv),
implies that a is EP. On the other hand, if a is EP, according to Theoren 3.7(xvi), there
exists u ∈ A−1 such that b = c†u. Then, according to Proposition 3.6(i) and statements (i)
and (v) of Lemma 3.9,
b = c†b†bu = a†bu = a∗vbu = c∗b∗(b†)∗(c†)∗c†u = c∗z,
where z = (c†)∗c†u. However, according to Lemma 3.9 (iv), z ∈ A−1.
Clearly, statement (ix) implies statement (x), which in turn implies statement (iii).
Similarly, statement (ix) implies statement (xi), which in turn implies statement (iii).
Statement (ix) implies statement (xii), which in turn implies statement (iv).
Finally, statement (ix) implies statements (xiii) and (xiv). On the other hand, statement
(xiii) implies statement (iii) and statement (xiv) implies statement (iv).
Before the next theorem, recall that if A is a unital C ∗-algebra and a ∈ A, then (aa∗)−1(0) =
(a∗)−1(0) and (a∗a)−1(0) = a−1(0). Moreover, if a ∈ A is Moore-Penrose invertible, then
aa∗A = aA and a∗aA = a∗A, see [4, Lemma 1.1(ii)].
Theorem 3.11. In the conditions of Theorem 3.10, the following statements are equivalent.
(i) a ∈ A is EP,
(ii) a∗a = c∗b∗bcbb† and a∗a = c∗b∗c†cbc,
(iii) aa∗ = bcc∗b∗c∗(c∗)† and aa∗ = bcbb†c∗b∗,
(iv) a∗a = c∗b∗bcbb† and aa∗ = bcc∗b∗c∗(c∗)†,
(v) aa∗ = c†b†bcbcc∗b∗ and a∗a = c∗b∗bcbb†,
(vi) a∗a = bcc†b†b∗c∗bc and aa∗ = bcc∗b∗c∗(c∗)†,
(vii) aa∗ = c†b†bcbcc∗b∗ and a∗a = bcc†b†b∗c∗bc.
Proof. If a ∈ A is EP, then a = aaa† and a = a†aa. Using these identities, statement (i)
implies all the others.
On the other hand, suppose that statement (ii) holds. If a∗a = c∗b∗bcbb†, then (b†)−1(0) ⊆
(a∗a)−1(0) = a−1(0) = c−1(0). In addition, if c∗b∗bc = a∗a = c∗b∗c†cbc, according to Propos-
ition 3.6(ii), b∗b = b∗c†cb, equivalently, b∗(e− c†c)b = 0. Now well, since (b∗)−1(0) = (b†)−1(0)
([9, Lemma 1.5]), according to the decomposition of A defined by bb†, i.e., A = bA⊕(b†)−1(0),
b∗(e − c†c) = 0. Consequently, since c−1(0) = (c†c)−1(0) = (e − c†c)A ⊆ (b∗)−1(0), according
to Theorem 3.10(iii), a is EP.
To prove that statement (iii) implies that a is EP, note that if aa∗ = bcc∗b∗c∗(c∗)†, then
according to [9, Lemma 1.5] and Proposition 3.6(ii),
c−1(0) = ((c∗)∗)−1(0) = ((c∗)†)−1(0) ⊆ (aa∗)−1(0) = (a∗)−1(0) = (a†)−1(0) = (b†)−1(0).
Moreover, if bcc∗b∗ = aa∗ = bcbb†c∗b∗, then according to Proposition 3.6(i), cc∗ = cbb†c∗,
equivalently c(e − bb†)c∗ = 0. Now well, since c∗A = c†A ( [9, Lemma 1.5]), according
Enrico Boasso
10
to the decomposition of A defined by c†c, i.e., A = c†A ⊕ c−1(0), c(e − bb†) = 0, equiva-
lently (b†)−1(0) = (bb†)−1(0) = (e − bb†)A ⊆ c−1(0). Therefore, since (b∗)−1(0) = (b†)−1(0),
according to Theorem 3.10(iii), a is EP.
Next consider statement (iv). Note that according to what has been proved, (b∗)−1(0) =
c−1(0), which, according again to Theorem 3.10(iii), implies that a is EP.
If statement (v) holds, then according to [4, Lemma 1.1(ii)], [9, Lemma 1.5] and the first
identity of statement (v), bA = aA = aa∗A ⊆ c†A = c∗A. What is more, according to what
has been proved, the second identity of statement (v) implies that (b†)−1(0) = (b∗)−1(0) ⊆
c−1(0). Consequently, according to the decompositons of A defined by c†c and bb† considered
above, (b∗)−1(0) = c−1(0) and bA = c∗A. Therefore, according to Theorem 3.10(iii)-(iv), a is
EP.
Next suppose that statement (vi) holds. Then, according to the first identity of statement
(vi), c∗A = c†A = a†A = a∗A = a∗aA ⊆ bA. Moreover, according to what has been
proved, the second identity of statement (vi) implies that c−1(0) ⊆ (b∗)−1(0) = (b†)−1(0).
Therefore, according to the decompositons of A defined by c†c and bb† considered above,
c−1(0) = (b∗)−1(0) and c∗A = bA, which according to Theorem 3.10(iii)-(iv), is equivalent to
the fact that a is EP.
If statement (vii) holds, according to what has been proved, c∗A = bA, which, according
to Theorem 3.10(iv), implies that a is EP.
4. Factorization of the form a† = sa
In this section, given a unital Banach algebra A, EP elements of the form a† = sa will
be characterized, a, s ∈ A. Recall that given a Banach algebra A and a ∈ A, according to [1,
Theorem 18(xviii)], necessary and sufficient for a to be EP is the fact that there is z ∈ A−1
such that a† = za. In what follows this result will be refined. Compare this section with [6,
section 4] and [4, sections 1.1, 2.1]
Theorem 4.1. Let A be a unital Banach algebra and consider a ∈ A such that a† exists.
Then, the following statements are equivalent.
(i) a is EP,
(ii) there exists s ∈ A such that s−1(0) = 0 and a† = sa,
(iii) there exist s1 and s2 ∈ A such that a† = s1a and a = s2a†,
(iv) there exists u ∈ A such that uA = A and a† = au,
(v) there exist u1 and u2 ∈ A such that a† = au1 and a = a†u2,
(vi) there exists t ∈ A such that t−1(0) = 0 and a† = at,
(vii) there exists x ∈ A such that Ax = A and a† = xa,
(viii) there exists v ∈ A−1 such that a†a = vaa†,
(ix) there exists v1 ∈ A such that v−1
(x) there exist v2 and v3 ∈ A such that a†a = v2aa† and aa† = v3a†a,
(xi) there exists w ∈ A−1 such that a†a = aa†w,
(xii) there exists w1 ∈ A such that w1A = A and a†a = aa†w1,
(xiii) there exist w2 and w3 ∈ A such that a†a = aa†w2 and aa† = a†aw3,
(xiv) there exist z1 and z2 ∈ A such that a†a = az1a† and aa† = a†z2a.
1 (0) = 0 and a†a = v1aa†,
Enrico Boasso
11
Proof. According to [1, Theorem 18(xviii)], if a ∈ A is EP, then statement (ii) and (iii) hold.
On the other hand, if statement (ii) or (iii) holds, then a−1(0) = (a†)−1(0). Consequently,
according to [1, Theorem 18(iii)], a is EP.
Similarly, if a is EP, according to [1, Theorem 18(xvii)], statements (iv) and (v) holds. If,
on the other hand, statements (iv) or (v) holds, then aA = a†A. In particular, according to
[1, Theorem 18(iv)], a is EP.
To prove the equivalences among statement (i) and statements (vi)-(vii) apply arguments
similar to the ones in the proofs of the equivalences among statement (i), (ii) and (iv), and
use [1, Theorem 18(xvii)] and [1, Theorem 18(viii)-(ix)].
If a is EP, clearly statements (viii)-(x) hold. On the other hand, if one of the statements
(viii)-(x) holds, then it is not difficult to prove that a−1(0) = (a†)−1(0) (recall that according
to Remark 2.3, (aa†)−1(0) = (a†)−1(0) and (a†a)−1(0) = a−1(0)). However, according to [1,
Theorem 18(iii)] a is EP.
In a similar way, using in particular [1, Theorem 18(iv)], the equivalence among the
condition of being EP and statements (xi)-(xiii) can be proved.
It is clear that if a is EP, then statement (xiv) holds. On the other hand, statement (xiv)
implies that a−1(0) = (a†)−1(0). Therefore, according to [1, Theorem 18(iii)], a is EP.
In the following theorem the condition of being EP wil be considered in the context of
C ∗-algebras .
Theorem 4.2. Let A be a unital C ∗-algebra and consider a ∈ A such that a† exists. Then,
the following statements are equivalent.
(i) a is EP,
(ii) there exists s ∈ A such that s−1(0) = 0 and a∗ = sa,
(iii) there exist s1 and s2 ∈ A such that a∗ = s1a and a = s2a∗,
(iv) there exists u ∈ A such that uA = A and a∗ = au,
(v) there exist u1 and u2 ∈ A such that a∗ = au1 and a = a∗u2,
(vi) there exists t ∈ A such that t−1(0) = 0 and a∗ = at,
(vii) there exists x ∈ A such that Ax = A and a∗ = xa,
(viii) there exists v ∈ A−1 such that a∗a = vaa∗,
(ix) there exists v1 ∈ A such that v−1
(x) there exist v2 and v3 ∈ A such that a∗a = v2aa∗ and aa∗ = v3a∗a,
(xi) there exists w ∈ A−1 such that a∗a = aa∗w,
(xii) there exists w1 ∈ A such that w1A = A and a∗a = aa∗w1,
(xiii) there exist w2 and w3 ∈ A such that a∗a = aa∗w2 and aa∗ = a∗aw3,
(xiv) there exist z1 and z2 ∈ A such that a∗a = az1a∗ and aa∗ = a∗z2a.
(xv) there exists h1 ∈ A−1 such that a∗a = ah1h∗
(xvi) there exists h2 ∈ A such that (h2)−1(0) = 0 and a∗a = ah2h∗
3a∗.
(xvii) there exists h3 ∈ A such that h3A = A and a∗a = ah3h∗
1 (0) = 0 and a∗a = v1aa∗,
1a∗,
2a∗,
Proof. According to Theorem 4.1(ii), a is EP if and only if there exists s ∈ A such that
(s)−1(0) = 0 and a† = sa. Now well, since according to Lemma 3.9(ii) there exists w ∈ A−1
such that a† = wa∗, if s = w−1s, then s−1(0) = 0 and a∗ = sa. On the other hand, if
statement (ii) holds, then (a∗)−1(0) = a−1(0). Thus, according to [9, Theorem 3.1(iv)], a is
EP.
Enrico Boasso
12
To prove that statement (i) is equivalent to statements (iii)-(vii), apply an argument
similar to the one in the previous paragraph, using in particular the corresponding statements
of Theorem 4.1 and Lemma 3.9(i)-(ii).
Next, if a is EP, then, according to Lemma 3.9(i)-(ii), there exist v, w ∈ A−1 such that
a† = a∗v and a† = wa∗. Then, according to Lemma 3.9,
a∗a = w−1a†a = w−1aa† = w−1aa∗v = w−1vaa∗.
Consequently, since v, w ∈ A−1, statement (viii) holds. In addition, it is clear that statement
(viii) implies statement (ix) and (x). On the other hand, if one of these statements holds,
then it is not difficult to prove that a−1(0) = (a∗)−1(0) (recall that (aa∗)−1(0) = (a∗)−1(0)
and (a∗a)−1(0) = a−1(0)). Therefore, according to [9, Theorem 3.1(iv)], a is EP.
To prove that statements (xi)-(xiii) are equivalent to the fact that a is EP, use an argument
similar to the one in the previous paragraph and the identities aa∗A = aA and a∗aA = a∗A
([4, Lemma 1.1(ii)]).
If a is EP, according to of [9, Theorem 3.1(vi)-(vii)], there exist m1, m2, m3 and m4
such that a = a∗m1 = m2a∗ and a∗ = am3 = m4a. As a result, a∗a = a(m3m2)a∗ and
aa∗ = a∗(m1m4)a. Therefore, statement (xiv) holds. On the other hand, if this statement
holds, then it is not difficult to prove that a−1(0) = (a∗)−1(0). Consequentely, according to
[9, Theorem 3.1(iv)], a is EP.
Recall that if a is EP, then according to [9, Theorem 3.1(viii)], there exists h1 ∈ A−1
such that a∗ = ah1. Thus, statement (xv) holds. Clearly, statement (xv) implies statements
(xvi)-(xvii).
On the other hand, if statement (xvi) holds, then a∗a = (ah2)(ah2)∗. Consequently, a
straightforward calculation proves that a−1(0) = ((ah2)∗)−1(0). However, since (h2)−1(0) =
0, it is not difficult to prove that a−1(0) = (a∗)−1(0). Thus, according to [9, Theorem 3.1(iv)],
a is EP.
Finally, if statement (xvii) holds, then a∗a = (ah3)(ah3)∗. Therefore, since h3A = A,
a∗A = ah3A = aA ([4, Lemma 1.1(ii)]). However, according to [9, Theorem 3.1(vi)], a is
EP.
5. Factorization of the form a = ucv
In this section, given a unital Banach algebra A, EP elements of the form a = ucv will
be studied, a, u, c, v ∈ A; compare with [4, sections 1.2]. However, in first place EP Banach
space operators will be characterized as block operators. Note that unlike to the Hilbert
space context where the direct sum of two Hilbert spaces is again a Hilbert space, there is
no canonical way to give a norm to the direct sum of two Banach spaces. Moreover, given a
Hilbert space H and H1 and H2 two orthogonal and complementary subspaces of H, H1 ⊕ H2
with its canonical Hilbert norm is isometrically isomorphic to H. However, in the case of two
closed and complementary subspaces X1 and X2 of a fixed Banach space X, although the sum
norn is equivalent to the original one, the natural identification between X1 ⊕X2 and X is not
in general an isometry. Consequently, since the norm is a key concept involved in the notions
of hermitian and Moore-Penrose invertible Banach space operators ([1, Remark 4]), and since
Enrico Boasso
13
an isometry is in general necessary to preserve the property of being hermitian and Moore-
Penrose invertible ([1, Remark 9]), some results of [6, section 3] can not be reformulated in
terms of Banach space isomorphisms. Compare the results presented in this section with [6,
section 3].
In the following proposition, given two Banach spaces X1 and X2, X1 ⊕p X2 will denote
the Banach space X1 ⊕ X2 with the p-norm, 1 ≤ p ≤ ∞, see [3, page 74]. Note that the
identity maps of X1 and X2 will be denoted by I1 and I2 respectivelly.
Proposition 5.1. Let X1 and X2 be two Banach space and consider T1 ∈ L(X1) a Banach
space isomorphism. Then, T1 ⊕ 0 ∈ L(X1 ⊕p X2) has a Moore-Penrose inverse. Furthermore,
(T1 ⊕ 0)† = (T −1
1 ⊕ 0) and T1 ⊕ 0 is an EP operator.
Proof. It is clear that T −1
1 ⊕ 0 is a normalized generalized inverse of T1 ⊕ 0. Note that
(T1 ⊕ 0)(T −1
1 ⊕ 0) = (T −1
1 ⊕ 0)(T1 ⊕ 0) = I1 ⊕ 0 = P1, where P1 ∈ L(X1 ⊕p X2) is the
projection onto X1. Therefore, in order to conclude the proof, it is enough to prove that P1
is an hermitian idempotent. However, a straightforward calculation shows that exp(itP1) =
P2 + eitP1, where P2 ∈ L(X1 ⊕p X2) is the projection onto X2 and t ∈ R. As a result,
k exp(itP1) k= 1, for all t ∈ R, equivalently P1 is an hermitian map.
Next the case of complementary closed subspaces of a given Banach spaces will be studied.
Proposition 5.2. Let X1 and X2 be two Banach spaces and consider T1 ∈ L(X1) an iso-
morphic operator. Let X be a Banach space and consider T ∈ L(X) such that there exists
a linear and bounded isomorphism J : X1 ⊕1 X2 → X with the property T = J(T1 ⊕ 0)J −1.
Then, the following statements are equivalent.
(i) T has a Moore-Penrose inverse,
(ii) T is EP,
(iii) Q1 = J(I1 ⊕ 0)J −1 ∈ L(X) is a hermitian idempotent,
(iv) Q2 = J(0 ⊕ I2)J −1 ∈ L(X) is a hermitian idempotent.
In particular, if J is an isometry, the four statements hold.
Proof. It is not difficult to prove that T ′ = J(T −1
1 ⊕ 0)J −1 is a normalized generalized inverse
of T and that Q1 and Q2 are the projections onto the closed and complemented subspaces
J(X1 ⊕ 0) and J(0 ⊕ X2) respectively. Furthermore, since T T ′ = T ′T = Q1 = I − Q2,
statements (i)-(iv) are equivalent. Concerning the last statement, apply Proposition 5.1 and
[1, Remark 9].
As a result, the characterization of EP bounded and linear maps as block operators can
be stated.
Theorem 5.3. Let X be a Banach space and consider T ∈ L(X). Then, the following
statements are equivalent.
(i) T is EP,
(ii) There exist two Banach spaces X1 and X2, T1 ∈ L(X1) an isomorphic operator, and
J : X1 ⊕1 X2 → X a linear and bounded isomorphism such that T = J(T1 ⊕ 0)J −1 and
J(I1 ⊕ 0)J −1 ∈ L(X) is a hermitian idempotent.
Enrico Boasso
14
Proof. According to Proposition 5.2, statement (ii) implies that T is EP. On the other hand,
if T is EP, according to [1, Theorem 13], there exist P ∈ L(X) a hermitian idempotent
such that N (P ) = N (T ) and R(P ) = R(T ). Denote then X1 = R(P ), X2 = N (P ) and T1 =
T X1
X1 : X1 → X1. It is clear that T1 ∈ L(X1) is an isomorphism. Moreover, J : X1 ⊕1 X2 → X
is the map J(x1 ⊕ x2) = x1 + x2. Since J −1 : X → X1 ⊕1 X2 is such that J −1 = P ⊕ (I − P ),
I ∈ L(X) the identity map, T = J(T1 ⊕ 0)J −1 and J(I1 ⊕ 0)J −1 = P .
In the following theorem instead of isomorphic operators, injective and surjective bounded
and linear maps will be considered.
Theorem 5.4. Let X be a Banach space and consider T ∈ L(X). Then, the following
statements are equivalent.
(i) T is EP,
(ii) there exist Banach spaces X1 and X2, T1 ∈ L(X1) an isomorphism, S ∈ L(X1 ⊕1 X2, X)
injective, U ∈ L(X, X1 ⊕1 X2) surjective, and P ∈ L(X) a hermitian idempotent such that
T = S(T1 ⊕ 0)U , R(P ) = S(X1 ⊕ 0) and N (P ) = U −1(0 ⊕ X2).
Proof. If T is EP, then let X1, X2 and T1 ∈ L(X1) be as in Theorem 5.3 and define S = J ∈
L(X1 ⊕1 X2, X) and U = J −1 ∈ L(X, X1 ⊕1 X2), J as in Theorem 5.3. Moreover, define P =
J(I1 ⊕ 0)J −1 ∈ L(X). Since R(P ) = R(T ) = S(X1 ⊕ 0) and N (P ) = N (T ) = U −1(0 ⊕ X2),
statement (ii) holds. On the other hand, if statement (ii) holds, then P ∈ L(X) is a hermitian
idempotent such that R(P ) = R(T ) and N (P ) = N (T ). Therefore, according to [1, Theorem
13], T is EP.
Next the conditions of Theorem 5.3 will be weekend.
Theorem 5.5. Let X be a Banach space and consider T ∈ L(X) such that T † exists. Then,
the following statements are equivalent.
(i) T is EP.
(ii) (a) There exist Banach spaces X1 and X2, A1 ∈ L(X1) injective, B1 ∈ L(X2), V1 and
W1 ∈ L(X1 ⊕1 X2, X), V1 injective, and S1 ∈ L(X, X1 ⊕1 X2) such that T = V1(A1 ⊕ 0)S1
and T † = W1(B1 ⊕ 0)S1.
(b) There exist Banach spaces X3 and X4, A2 ∈ L(X3), B2 ∈ L(X2) injective, V2 and
W2 ∈ L(X1 ⊕1 X2, X), W2 injective, and S2 ∈ L(X, X1 ⊕1 X2) such that T = V2(A2 ⊕ 0)S2
and T † = W2(B2 ⊕ 0)S2.
(iii) (a) There exist Banach spaces Y1 and Y2, A3 ∈ L(Y1) surjective, B3 ∈ L(Y2), V3 ∈
L(Y1 ⊕1 Y2, X) and S3, S4 ∈ L(X, Y1 ⊕1 Y2), S3 surjective, such that T = V3(A2 ⊕ 0)S3 and
T † = V3(B3 ⊕ 0)S4,
(b) there exist Banach spaces Y3 and Y4, A4 ∈ L(Y3), B4 ∈ L(Y4) surjective, V4 ∈
L(Y3 ⊕1 Y4), X and S5, S6 ∈ L(X, Y3 ⊕1 Y4), S6 surjective, such that T = V4(A4 ⊕ 0)S5 and
T † = V4(B4 ⊕ 0)S6.
Proof. According to Theorem 5.3, if T is EP, then statement (ii) and (iii) hold. On the other
hand, if statement (ii) hold, then it no difficult to prove that N (T ) = N (T †). Therefore,
according to [1, Theorem 16(ii)], T is EP. Similarly, if statement (iii) holds, then R(T ) =
R(T †). Consequently, according to [1, Theorem 16(iii)], T is EP.
Next the Banach algebra frame will be considered.
Enrico Boasso
15
Theorem 5.6. Let A be a unital Banach algebra, and consider a ∈ A such that a† exists.
Then, the following statement are equivalent.
(i) The element a is EP,
(ii)there exist b1, c1, d1, f1 and g1 ∈ A such that a = b1c1g1, a† = f1d1g1, (c1)−1(0) =
(d1)−1(0), and (b1)−1(0) = (f1)−1(0) = 0,
(iii) there exist h1, k1, l1, m1 and n1 ∈ A such that a = h1k1l1, a† = h1m1n1, and l1A =
A = n1A, and k1A = m1A.
(iv)there exist b2, c2, d2, g2 and g3 ∈ A such that a = b2c2g2, a† = b2d2g3, (c2)−1(0) =
(d1)−1(0), and (g2)−1(0) = (g3)−1(0) = 0,
(v) there exist h2, h3, k2, l2 and m2 ∈ A such that a = h2k2l2, a† = h3m2l2, and Ak2 = Am2,
and Ah2 = Ah3 = A.
Proof. If a ∈ A is EP, then consider b1 = g1 = f1 = e, c1 = a and d1 = a† and use [1,
Theorem 18(iii)]. On the other hand, if statement (ii) holds, a straightforward calculation
shows that a−1(0) = (a†)−1(0) which, according again to [1, Theorem 18(iii)], implies that a
is EP.
A similar argument, using in particular that necessary and sufficient for a ∈ A to be
EP is the fact that aA = a†A ([1, Theorem 18(iv)]), proves that statements (i) and (iii) are
equivalent.
To prove that statement (i) and statements (iv)-(v) are equivalent, apply arguments
similars to the ones used to prove that the condition of being EP is equivalent to statements
(ii)-(iii), using in particular [1, Theorem 18(viii)-(ix)] instead of statements [1, Theorem
18(iii)-(iv)].
Note that under the hypothesis of Theorem 5.6, others statements equivalent to the
condition of being EP can be obtained if instead of a and a†, a†a and aa† are considered. In
fact, using the fact that (a†a)−1(0) = a−1(0) and (aa†)−1(0) = (a†)−1(0), arguments similar
to the ones in Theorem 5.6 prove the corresponding statements.
References
[1] E. Boasso, On the Moore-Penrose inverse, EP Banach space operators, and EP Banach
algebra elements, J. Math. Anal. Appl. 339 (2) (2008) 1003-1014.
[2] F. F. Bonsall, J. Duncan, Complete normed algebras, Springer Verlag, Berlin, Heidel-
berg, New York, 1973.
[3] J. B. Conway, A Course in Functional Analysis, Springer-Verlag, New York, Berlin,
Heidelberg, Tokyo, 1985.
[4] D. S. Djordjevi´c, J. J. Koliha, I. Straskaba, Factorizations of EP elements of C ∗-algebras,
Linear and Multilinear Algebra 57 (6) (2009) 587-594.
[5] H. R. Dowson, Spectral Theory of Linear operators, Academic Press, London, New York,
San Francisco, 1978.
Enrico Boasso
16
[6] D. Drivaliaris, S. Karamasios, D. Pappas, Factorizations of EP operators, Linear Algebra
Appl. 429 (7) (2008) 1555-1567.
[7] R. Harte, M. Mbekhta, On generalized inverses in C ∗-algebras, Studia Math. 103 (1992)
71-77.
[8] R. Harte, M. Mbekhta, On generalized inverses in C ∗-algebras II, Studia Math. 106
(1993) 129-138.
[9] J. J. Koliha, Elements of C ∗-algebras commuting with their Moore-Penrose inverse,
Studia Math. 139 (1) (2000) 81-90.
[10] G. Lumer, Semi-Inner-Product Spaces, Trans. Amer. Math. Soc. 100 (1) (1961) 29-43.
[11] M. Mbekhta, Conorme et inverse g´en´eralis´e dans les C ∗-alg`ebres, Canadian Math. Bull.
35 (4) (1992) 515-522.
[12] T. Palmer, Unbounded normal operators on Banach spaces, Trans. Amer. Math. Soc.
133 (2) (1968) 385-414.
[13] R. Penrose, A generalized inverse for matrices, Proc. Cambridge Philos. Soc. 51 (1955)
406-413.
[14] V. Rakocevi´c, Moore-Penrose inverse in Banach algebras, Proc. R. Ir. Acad. 88A (1)
(1988) 57-60.
[15] V. Rakocevi´c, On the continuity of the Moore-Penrose inverse in Banach algebras, Facta
Univ. Ser. Math. Inform. 6 (1991) 133-138.
[16] I. Vidav, Eine metrische Kennzeichnung der selbstadjungierten Operatoren, Math. Z. 66
(1956) 121-128.
Enrico Boasso
E-mail address: enrico [email protected]
|
1609.08909 | 2 | 1609 | 2016-12-07T11:44:37 | Log-majorization and Lie-Trotter formula for the Cartan barycenter on probability measure spaces | [
"math.FA"
] | We extend Ando-Hiai's log-majorization for the weighted geometric mean of positive definite matrices into that for the Cartan barycenter in the general setting of probability measures on the Riemannian manifold of positive definite matrices equipped with trace metric. The main key is the settlement of the monotonicity problem of the Cartan barycenteric map on the space of probability measures with finite first moment for the stochastic order induced by the cone. We also derive a version of Lie-Trotter formula and related unitarily invariant norm inequalities for the Cartan barycenter as the main application of log-majorization. | math.FA | math | LOG-MAJORIZATION AND LIE-TROTTER FORMULA FOR THE
CARTAN BARYCENTER ON PROBABILITY MEASURE SPACES
FUMIO HIAI AND YONGDO LIM
Abstract. We extend Ando-Hiai's log-majorization for the weighted geometric
mean of positive definite matrices into that for the Cartan barycenter in the gen-
eral setting of probability measures on the Riemannian manifold of positive defi-
nite matrices equipped with trace metric. The main key is the settlement of the
monotonicity problem of the Cartan barycenteric map on the space of probability
measures with finite first moment for the stochastic order induced by the cone. We
also derive a version of Lie-Trotter formula and related unitarily invariant norm
inequalities for the Cartan barycenter as the main application of log-majorization.
6
1
0
2
c
e
D
7
]
.
A
F
h
t
a
m
[
2
v
9
0
9
8
0
.
9
0
6
1
:
v
i
X
r
a
2010 Mathematics Subject Classification. 15A42, 47A64, 47B65, 47L07
Key words and phrases. Positive definite matrix, Cartan barycenter, Wasserstein
distance, log-majorization, Lie-Trotter formula, unitarily invariant norm
1. Introduction
Let A be an m × m positive definite matrix with eigenvalues λj(A), 1 ≤ j ≤ m,
arranged in decreasing order, i.e., λ1(A) ≥ · · · ≥ λm(A) with counting multiplicities.
B between positive definite matrices A and B is defined if
The log-majorization A ≺log
k
k
λi(A) ≤
λi(B)
for 1 ≤ k ≤ m − 1, and det A = det B.
Yi=1
Yi=1
The log-majorization gives rise to powerful devices in deriving various norm inequali-
ties and has many important applications in operator means, operator monotone func-
tions, statisticalmechanics, quantum information theory, eigenvalue analysis, etc., see,
e.g., [3, 6, 11]. For instance, A ≺log
B implies A ≤ B for all unitarily invariant
norms · .
Date: September 10, 2018.
1
2
HIAI AND LIM
As a complementary counterpart of the Golden-Thompson trace inequality, Ando
and Hiai [2] established the log-majorization on the matrix geometric mean of two
positive definite matrices: for positive definite matrices A, B and 0 ≤ α ≤ 1,
At#αBt ≺log
(A#αB)t,
t ≥ 1,
where A#αB := A1/2(A−1/2BA−1/2)αA1/2, the α-weighted geometric mean of A and
B. This provides various norm inequalities for unitarily invariant norms via the Lie-
Trotter formula limt→0(At#αBt)
t
increases to e(1−α) log A+α log B as r ց 0 for any unitarily invariant norm. Ando-
Hiai's log-majorization has many important applications in matrix analysis and in-
t = e(1−α) log A+α log B. For instance, (At#αBt)
1
1
equalities, together with Araki's log-majorization [4] extending the Lieb-Thirring and
the Golden-Thompson trace inequalities.
The matrix geometric mean A#αB, that plays the central role in Ando-Hiai's
log-majorization, appears as the unique (up to parametrization) geodesic curve α ∈
[0, 1] 7→ A#αB between A and B on the Riemannian manifold Pm of positive def-
inite matrices of size m, an important example of Cartan-Hadamard Riemannian
manifolds. Alternatively, the geometric mean A#αB is the Cartan barycenter of the
finitely supported measure (1 − α)δA + αδB on Pm, which is defined as the unique
minimizer of the least squares problem with respect to the Riemannian distance d
(see Section 2 for definition). Indeed, for a general probability measure µ on Pm with
finite first moment, the Cartan barycenter of µ is defined as the unique minimizer as
follows:
G(µ) := arg min
Z∈Pm ZPm(cid:2)d2(Z, X) − d2(Y, X)(cid:3)dµ(X)
(see Section 2 for more details). In particular, when µ = Pn
j=1 wjδAj is a discrete
probability measure supported on a finite number of A1, . . . , An ∈ Pm, the Cartan
barycenter G(µ) is the Karcher mean of A1, . . . , An, which has extensively been dis-
cussed in these years by many authors as a multivariable extension of the geometric
mean (see [7, 15, 20] and references therein).
The first aim of this paper is to establish the log-majorization (Theorem 4.4) for the
Cartan barycenter in the general setting of probability measures in the Wasserstein
space P 1(Pm), the probability measures on Pm with finite first moment. In this way,
we first establish the monotonicity of the Cartan barycenteric map on P 1(Pm) for
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
3
the stochastic order induced by the cone of positive semidefinite matrices, and then
generalize the log-majorization in [2] (as mentioned above) and in [10] (for the Karcher
mean of multivariables) to the setting of probability measures. Our second aim is to
derive the Lie-Trotter formula (Theorem 5.7) for the Cartan barycenter
G(µt)
lim
t→0
1
t = expZPm
log A dµ(A)
under a certain integrability assumption on µ, where µt is the tth power of the mea-
sure µ inherited from the matrix powers on Pm. Moreover, to demonstrate the useful-
ness of our log-majorization, we obtain several unitarily invariant norm inequalities
(Corollary 5.8) based on the above Lie-Trotter formula.
The main tools of the paper involve the theory of nonpositively curved metric
spaces and techniques from probability measures on metric spaces and the recent
combination of the two (see [18, 1, 19]). Not only are these tools crucial for our
developments, but also, we believe, significantly enhance the potential usefulness
of the Cartan barycenter of probability measures in matrix analysis and inequalities.
They overcome the limitation to the multivariable (finite number of matrices) setting,
and provide a new bridge between two different important fields of studies of matrix
analysis and probability measure theory on nonpositively curved metric spaces.
2. Cartan barycenters
Let Hm be the Euclidean space of m×m Hermitian matrices equipped with the inner
product hX, Y i := tr(XY ). The Frobenius norm k · k2 defined by kXk2 = (tr X 2)1/2
for X ∈ Hm gives rise to the Riemannian structure on the open convex cone Pm of
m × m positive definite matrices with the metric
hX, Y iA := tr(A−1XA−1Y ),
A ∈ Pm, X, Y ∈ Hm,
(2.1)
where the tangent space of Pm at any point A ∈ Pm is identified with Hm. The
Riemannian exponential at A ∈ Pm is given by
2 exp(A− 1
2 XA− 1
2 )A
expA(X) = A
1
1
2
and its inverse is
logA(X) = A
1
2 log(A− 1
2 XA− 1
2 )A
1
2 .
4
HIAI AND LIM
Then Pm is a Cartan-Hadamard Riemannian manifold, a simply connected complete
Riemannian manifold with nonpositive sectional curvature (the canonical 2-tensor is
nonnegative). The Riemannian trace metric (i.e., the geodesic distance with respect
to (2.1)) on Pm is given by
d(A, B) :=(cid:13)(cid:13) log A− 1
1
2 BA− 1
2(cid:13)(cid:13)2,
and the unique (up to parametrization) geodesic shortest curve joining A and B
is t ∈ [0, 1] 7→ A#tB = A
equivalently stated as
2 . The nonpositively curved property is
2 BA− 1
2 (A− 1
2 )tA
1
d2(A#tB, C) ≤ (1 − t)d2(A, C) + td2(B, C) − (1 − t)td2(A, B).
(2.2)
See [14, 5] for more about these Riemannian structures.
Let B = B(Pm) be the algebra of Borel sets, the smallest σ-algebra containing the
open sets of Pm. We note that the Euclidean topology on Pm coincides with the
metric topology of the trace metric d. Let P = P(Pm) be the set of all probability
measures on (Pm,B) and Pc = Pc(Pm) the set of all compactly supported µ ∈ P. Let
P0 = P0(Pm) be the set of all µ ∈ P of the form µ = (1/n)Pn
j=1 δAj , where δA is the
point measure of mass 1 at A ∈ P. For p ∈ [1,∞) let P p = P p(Pm) be the set of
probability measures with finite p-moment, i.e., for some (and hence all) Y ∈ Pm,
ZPm
dp(X, Y ) dµ(X) < ∞.
We say that ω ∈ P(Pm × Pm) is a coupling for µ, ν ∈ P if µ, ν are the marginals
of ω, i.e., if for all B ∈ B, ω(B × Pm) = µ(B) and ω(Pm × B) = ν(B). We note that
one such coupling is the product measure µ × ν. We denote the set of all couplings
for µ, ν ∈ P(Pm) by Π(µ, ν).
The p-Wasserstein distance dW
p on P p is defined by
π∈Π(µ,ν)ZPm×Pm
dW
p (µ, ν) :=(cid:20) inf
is a complete metric on P p and P0 is dense in P p [18]. Note that
It is known that dW
p
P0 ⊂ Pc ⊂ P q ⊂ P p ⊂ P 1 and dW
for 1 ≤ p ≤ q < ∞. We note that these
basic results on probability measure spaces hold in general setting of complete metric
dp(X, Y ) dπ(X, Y )(cid:21)
p ≤ dW
.
q
1
p
spaces in which cases separability assumption is necessary.
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
5
The following result on Lipschitz property of push-forward maps between met-
p on
ric spaces appears in [17], where X, Y are metric spaces and the distance dW
P p(X),P p(Y ) are defined as above.
Lemma 2.1. Let f : X → Y be a Lipschitz map with Lipschitz constant C. Then
the push-forward map f∗ : P p(X) → P p(Y ), f∗(µ) = µ◦ f −1, is Lipschitz with respect
to dW
p with Lipschitz constant C for 1 ≤ p < ∞.
Definition 2.2. The Cartan barycenter map G : P 1(Pm) → Pm is defined by
µ ∈ P 1(Pm)
Z∈Pm ZPm(cid:2)d2(Z, X) − d2(Y, X)(cid:3) dµ(X),
G(µ) := arg min
for a fixed Y . The uniqueness and existence of the minimizer is well-known and the
unique minimizer is independent of Y (see [18, Proposition 4.3]). On P 2(Pm), the
Cartan barycenter is determined by
G(µ) = arg min
d2(Z, X) dµ(X).
Z∈Pm ZPm
For a discrete measure µ =Pn
a weight (w1, . . . , wn), see, e.g., [15, 20].
j=1 wjδAj , G(µ) is the Karcher mean of A1, . . . , An with
The following contraction property appears in [18].
Theorem 2.3 (Fundamental Contraction Property). For every µ, ν ∈ P p(Pm), p ≥ 1,
d(G(µ), G(ν)) ≤ dW
1 (µ, ν) ≤ dW
p (µ, ν).
3. Karcher equations and monotonicity
A map g : Pm → R is called uniformly convex if there is a strictly increasing
function φ : [0,∞) → [0,∞) such that
g(A#B) ≤
(g(A) + g(B)) − φ(δ(A, B))
1
2
for all A, B ∈ Pm. For a continuous uniformly convex function g, it has a unique
minimizer of g (see [18]) and coincides with the unique point that vanishes the (either
Riemannian or Euclidean) gradient, whenever it is differentiable, see [16].
By (2.2), the map
Z 7→ZPm(cid:2)d2(Z, X) − d2(Y, X)(cid:3) dµ(X),
µ ∈ P 1(Pm)
6
HIAI AND LIM
is uniformly convex. The next theorem is a characterization of G(µ) in terms of the
unique solution to the Karcher equation.
Theorem 3.1. For every µ ∈ P 1(Pm), G(µ) is the unique solution Z ∈ Pm to the
Karcher equation
log Z −1/2XZ −1/2 dµ(X) = 0.
ZPm
Proof. Let µ ∈ P 1(Pm). We first show that the Euclidean gradient of the function
Z ∈ Pm 7−→ ϕ(Z) :=ZPm(cid:2)d2(Z, X) − d2(Y, X)(cid:3) dµ(X)
Z −1/2(cid:18)ZPm
log Z 1/2X −1Z 1/2 dµ(X)(cid:19)Z −1/2.
is
More precisely, with
we shall prove that
F (Z, X) := 2Z −1/2(cid:0)log Z 1/2X −1Z 1/2(cid:1)Z −1/2,
ϕ(Z + H) = ϕ(Z) +ZPm
Z, X ∈ Pm,
tr F (Z, X)H dµ(X) + o(kHk2)
(3.1)
as kHk2 → 0 for H ∈ Hm.
For each fixed X ∈ Pm, let
ψ(Z) := d2(Z, X) = tr(cid:0)log X −1/2ZX −1/2(cid:1)2
,
Z ∈ Pm.
It is not difficult to compute the gradient of ψ(Z) is F (Z, X), i.e.,
ψ(Z + H) = ψ(Z) + tr F (Z, X)H + o(kHk2)
as kHk2 → 0 for H ∈ Hm. Then, for every Z ∈ Pm and H ∈ Hm, by the Lebesgue
convergence theorem one can prove that
d
dt
ϕ(Z + tH)(cid:12)(cid:12)(cid:12)t=0
= lim
t→0
ϕ(Z + tH) − ϕ(Z)
t
ψ(Z + tH) − ψ(Z, X)
t
dµ(X)
tr F (Z, X)H dµ(X).
= lim
t→0ZPm
=ZPm
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
7
This formula for the directional derivative is enough to give (3.1) (due to the finite
dimensionality).
(cid:3)
For any U ⊂ Pm, we define U ↑ = {B ∈ Pm : A ≤ B for some A ∈ U}. A set
U is an upper set if U ↑ = U. For µ, ν ∈ P(Pm), we define µ ≤ ν if µ(U) ≤ ν(U)
for all open upper sets U. This partial order on P(Pm) is a natural extension of
the usual one; Aj ≤ Bσ(j) for some permutation σ and j = 1, . . . , n if and only if
(1/n)Pn
j=1 δAj ≤ (1/n)Pn
We recall the well-known Lowner-Heinz inequality:
j=1 δBj , as seen from the marriage theorem.
0 < A ≤ B implies At ≤ Bt,
t ∈ [0, 1].
The next theorem is the monotonicity property of the Cartan barycenter G on P 1(Pm)
and extends the recent works of Lawson-Lim [15] and Bhatia-Karandikar [8] on the
space of finitely (and uniformly) supported measures, which can be viewed as a mul-
tivariate Lowner-Heinz inequality.
Theorem 3.2. Let µ, ν ∈ P 1(Pm). If µ ≤ ν, then G(µ) ≤ G(ν).
Proof. Assume that µ, ν ∈ P 1(Pm) and µ ≤ ν. For each n ∈ N let Σn := {X ∈ Pm :
(1/n)I ≤ X ≤ nI} and
µn := µΣn + µ(Pm \ Σn)δ(1/n)I ,
νn := νΣn + ν(Pm \ Σn)δnI.
Then, as in the proof of [12, Section 6], we have µn ≤ νn. Since µn, νn ∈ Pc(Pm),
we have G(µn) ≤ G(νn) by [12, Theorem 5.5 (6)]. We now prove that dW
1 (µn, µ) → 0
as n → ∞. From a basic fact on the convergence in Wasserstein spaces (see [19,
Theorem 7.12]) we may prove that µn → µ weakly and
lim
n→∞ZPm k log Xk2 dµn(X) =ZPm k log Xk2 dµ(X).
Since µ(Pm \ Σn) → 0, it is obvious that µn → µ weakly. Note that
ZPm k log Xk2 dµn(X) =ZΣn k log Xk2 dµ(X) + k log((1/n)I)k2 µ(Pm \ Σn)
=ZΣn k log Xk2 dµ(X) + √m (log n) µ(Pm \ Σn).
8
HIAI AND LIM
Note also that if X ∈ Pm \ Σn, then either the largest eigenvalue of X satisfies
λ1(X) > n or the smallest one does λm(X) < 1/n, so we have k log Xk2 ≥ log n.
Therefore, since RPm k log Xk2 dµ(X) < ∞, we have
(log n) µ(Pm \ Σn) ≤ZPm\Σn k log Xk2 dµ(X) −→ 0 as n → ∞,
so that
lim
n→∞ZPm k log Xk2 dµ(X) = lim
n→∞ZΣn k log Xk2 dµ(X) =ZPm k log Xk2 dµ(X).
We thus have dW
1 (µn, µ) → 0, which implies δ(G(µn), G(µ)) → 0 by the fundamental
contraction property, so kG(µn)− G(µ)k2 → 0. Since kG(νn)− G(ν)k2 → 0 similarly,
G(µ) ≤ G(ν) follows by taking the limit of G(µn) ≤ G(νn).
(cid:3)
4. Log-majorization
For 1 ≤ k ≤ m and A ∈ Pm, let ΛkA be the kth antisymmetric tensor power of A.
See [2, 8, 11] for basic properties of Λk; for instance,
Λk(AB) = (ΛkA)(ΛkB),
Λk(At) = (ΛkA)t,
t > 0,
λ1(ΛkA) =
λj(A).
k
Yj=1
(4.1)
(4.2)
The kth antisymmetric tensor power map Λk maps Pm continuously into Pℓ where
ℓ :=(cid:0)m
k(cid:1). This induces the push-forward map
∗(µ)(O) = µ((Λk)−1(O)) for all Borel sets O ⊂ Pℓ.
Λk
∗ : P(Pm) → P(Pℓ),
Λk
∗(µ) := µ ◦ (Λk)−1,
that is, Λk
Proposition 4.1. The map Λk : Pm → Pℓ is Lipschitzian, that is,
A, B ∈ Pm,
∗ : P p(Pm) → P p(Pℓ) is Lipschitzian for
where αm,k := qk(cid:0)m−1
every p ≥ 1, that is,
dW
p (Λk
d(ΛkA, ΛkB) ≤ αm,k d(A, B),
k−1(cid:1). Furthermore, Λk
∗(ν)) ≤ αm,k dW
∗(µ), Λk
p (µ, ν),
µ, ν ∈ P p(Pm).
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
9
Proof. The eigenvalue list of Λk(A− 1
2 BA− 1
2 ) = (ΛkA)− 1
2 (ΛkB)(ΛkA)− 1
2 is
k
Yj=1
Hence
λij (A− 1
2 BA− 1
2 ),
1 ≤ i1 < · · · < ik ≤ m.
λij (A− 1
2 BA− 1
2 )!
2 )#2
log λij (A− 1
2 BA− 1
log2 λij (A− 1
2 BA− 1
2 )
log2 λi(A− 1
2 BA− 1
2 )
2
2
2 BA− 1
2 )(cid:13)(cid:13)
log2 k
Yj=1
d2(ΛkA, ΛkB) = (cid:13)(cid:13) log Λk(A− 1
= X1≤i1<···<ik≤m
= X1≤i1<···<ik≤m" k
Xj=1
Xj=1
≤ X1≤i1<···<ik≤m
k − 1(cid:19) m
= k(cid:18)m − 1
Xi=1
k − 1(cid:19)d2(A, B).
= k(cid:18)m − 1
k
k
The Lipschitz continuity of Λk
∗ follows by Lemma 2.1.
(cid:3)
The following is an extension of the result by Bhatia and Karandikar [8, Theorem
4.4] for finitely supported measures to general probability measures in P 1(Pm).
Theorem 4.2. For p ≥ 1, the following diagram commute:
Pm
Gx
P p(Pm)
that is,
Λk
−−−→ Pℓ
G
,
x
Λk
∗−−−→ P p(Pℓ)
G ◦ Λk
∗ = Λk ◦ G.
(4.3)
Proof. Let µ ∈ P 1(Pm). By Theorem 3.1, letting Z := G(µ), we may prove that
ZPℓ
log(cid:0)Λk(Z)(cid:1)−1/2X(cid:0)Λk(Z)(cid:1)−1/2 d(Λk
∗µ)(X) = 0,
10
HIAI AND LIM
log(cid:2)Λk(cid:0)Z −1/2XZ −1/2(cid:1)(cid:3) dµ(X) = 0. Note that
i.e., RPm
log(cid:2)Λk(cid:0)Z −1/2XZ −1/2(cid:1)(cid:3) = log(cid:0)Z −1/2XZ −1/2(cid:1)⊗k(cid:12)(cid:12)(cid:12)(Cm)Λk
= k
Xj=1
I ⊗(j−1) ⊗(cid:0)log Z −1/2XZ −1/2(cid:1) ⊗ I ⊗(k−j)!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(Cm)Λk
,
where (Cm)Λk is the k-fold antisymmetric tensor space of Cm.
log Z −1/2XZ −1/2 dµ(X) = 0, we have
Since RPm
ZPm
log(cid:2)Λk(cid:0)Z −1/2XZ −1/2(cid:1)(cid:3) dµ(X)
= k
Xj=1
I ⊗(j−1) ⊗(cid:18)ZPm
log Z −1/2XZ −1/2 µ(X)(cid:19) ⊗ I ⊗(k−j)!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(Cm)Λk
= 0.
(cid:3)
Next, we introduce powers of probability measures on Pm.
Definition 4.3. For t ∈ R \ {0} and O ∈ B(Pm), we let Ot := {At : A ∈ O} and
µt(O) := µ(O
1
t ).
In terms of push-forward measures, µt = g∗µ, where g(X) := X t. Note that µt ∈
P p(Pm) if µ ∈ P p(Pm).
By (4.1) and the definition of push-forward map, we have
Λk
∗(µt) = Λk
∗(µ)t,
µ ∈ P p(Pm), t 6= 0.
(4.4)
In [13], Kim-Lee-Lim established that kG(µt)k ≤ kG(µ)tk for µ ∈ P 2(Pm) and
It follows from the monotonicity
t ≥ 1, where k · k denotes the operator norm.
of Cartan barycenter and its the characterization via the Karcher equation. In the
present situation, the same method based on Theorems 3.1 and 3.2 proves that
kG(µt)k ≤ kG(µ)tk,
µ ∈ P 1(Pm),
t ≥ 1.
(4.5)
The main result of this section is the following:
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
11
Theorem 4.4. For every µ ∈ P 1(Pm) and t ≥ 1,
G(µ)t.
G(µt) ≺log
In particular, for any unitary invariant norm · ,
G(µt) ≤ G(µ)t,
t ≥ 1.
Proof. For 1 ≤ k ≤ m we have
k
Yj=1
λj(G(µt)) = λ1(ΛkG(µt)) = kΛkG(µt)k
∗(µt)(cid:1)(cid:13)(cid:13) =(cid:13)(cid:13)G(cid:0)(Λk
Yj=1
∗(µ))(cid:1)(cid:13)(cid:13)
∗ (µ))t(cid:1) is a positive scalar, the
det G(µ)t. When k = m, since Λm = det and G(cid:0)(Λm
∗ (µ))t(cid:1). In the one-dimensional
equalities shown above say that det G(µt) = G(cid:0)(Λm
case on P1 = (0,∞), we find by a direct computation that
log x dν(x)
where (4.2), (4.3), (4.4), and (4.5) have been used. It remains to show that det G(µt) =
= (cid:13)(cid:13)G(cid:0)Λk
≤ (cid:13)(cid:13)G(cid:0)(Λk
∗(µ))t(cid:1)(cid:13)(cid:13)
λj(G(µ)t),
t =
k
G(ν) = expZ(0,∞)
for every ν ∈ P 1((0,∞)). Therefore,
G(cid:0)(Λm
∗ (µ))t(cid:1) = expZ(0,∞)
= expZPm
log x d(Λm
∗ (µ)t)(x) = expZPm
t tr(log A) dµ(A) = dett(cid:18)expZPm
= det G(µ)t,
log(dettA) dµ(A)
log A dµ(A)(cid:19)
implying that det G(µt) = det G(µ)t.
(cid:3)
By a consequence of the preceding theorem, we have the following:
Corollary 4.5. For every µ ∈ P 1(Pm),
G(µq)
1
q ≺(log)
G(µp)
1
p ,
G(µp)
1
p ≺(log)
G(µ) ≺(log)
G(µ
0 < p ≤ q,
p )p,
1
p ≥ 1,
12
HIAI AND LIM
and therefore
G(µq)
1
1
q ≤ G(µp)
1
p,
G(µp)
p ≤ G(µ) ≤ G(µ
1
0 < p ≤ q,
p )p,
p ≥ 1
(4.6)
for all unitarily invariant norms · .
5. Lie-Trotter formula
The Lie-Trotter formula for the Cartan (or Karcher) mean of multivariable positive
definite matrices is
see [10, 9, 6].
G(At
1, . . . , At
n)
lim
t→0
1
t = exp 1
n
log Aj! ,
n
Xj=1
In this section we establish the Lie-Trotter formula and associated norm inequalities
for probability measures in a certain sub-class of P 1(Pm).
Lemma 5.1. For every X ∈ Pm,
k log Xk ≤ log(kXk + kX −1k).
Moreover, for every r > 0 there exists a constant cr > 0 such that
k log Xk2 ≤ cr(kXk + kX −1k)r,
X ∈ Pm.
Proof. Since kX −1k−1I ≤ X ≤ kXkI, we have (− log kX −1k)I ≤ log X ≤ (log kXk)I
so that
k log Xk = max(cid:8)log kXk, logkX −1k(cid:9) ≤ log(kXk + kX −1k).
k log Xk2 ≤ √mk log Xk ≤ √m log(cid:0)kXk + kX −1k(cid:1)
Next, for any r > 0, since limx→∞(log x)/xr = 0, br := supx≥1(log x)/xr < ∞. Noting
that kXk + kX −1k ≥ 2pkXk kX −1k ≥ 2, we have
≤ √m br(cid:0)kXk + kX −1k(cid:1)r,
X ∈ Pm.
(cid:3)
Now, for µ ∈ P(Pm) we consider the condition
ZPm(cid:0)kXk + kX −1k(cid:1) dµ(X) < ∞.
(5.1)
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
13
Lemma 5.2. If µ ∈ P(Pm) satisfies (5.1), then µ ∈ P p(Pm) for every p ∈ [1,∞).
Proof. By Lemma 5.1 with r = 1/p we have
k log Xk2 ≤ k1/p(cid:0)kXk + kX −1k(cid:1)1/p
2 dµ(X) ≤ kp
Therefore,
dp(X, I) dµ(X) =ZPm k log Xkp
ZPm
implying µ ∈ P p(Pm).
,
X ∈ Pm.
1/pZPm(cid:0)kXk + kX −1k(cid:1) dµ(X) < ∞,
(cid:3)
When µ satisfies (5.1), one can define the arithmetic and the harmonic means of µ
as
X dµ(X),
ZPm
(cid:18)ZPm
X −1 dµ(X)(cid:19)−1
,
respectively. By Lemma 5.2 one can also define the Cartan barycenter G(µ).
The next lemma will be useful in the proof of our main result of this section.
Lemma 5.3. Assume that µ ∈ P(Pm) satisfies (5.1). Then there exist a sequence
{µn}∞
n=1 in Pc(Pm) such that, as n → ∞,
dW
1 (µn, µ) −→ 0,
and
ZPm
X dµn(X) −→ZPm
X dµ(X),
ZPm
X −1 dµn(X) −→ZPm
X −1 dµ(X).
Proof. For each n ∈ N let Σn be as in the proof of Theorem 3.2 and define µn ∈ Pc(Pm)
as
µn := µΣn + µ(Pm \ Σn)δI.
We then have dW
weakly to µ and
1 (µn, µ) → 0 as in the proof of Theorem 3.2, since µn converges
ZPm k log Xk2 dµn(X) =ZΣn k log Xk2 dµ(X) −→ZPm k log Xk2 dµ(X)
as n → ∞. On the other hand, by assumption (5.1) we have
ZPm
and similarly RPm
X dµn(X) =ZΣn
X −1 dµn(X) →RPm
X dµ(X) + µ(Pm \ Σn)I −→ZPm
X −1 dµ(X).
X dµ(X),
(cid:3)
14
HIAI AND LIM
The following AGH (arithmetic-geometric-harmonic) mean inequalities were shown
for µ ∈ P0(Pm) in [21, Theorem 2] and extended in [12] to the case of µ ∈ Pc(Pm).
We further extend it to the case of µ satisfying (5.1).
Proposition 5.4 (AGH inequalities). If µ ∈ P(Pm) satisfies (5.1), then
(cid:18)ZPm
X −1 dµ(X)(cid:19)−1
≤ G(µ) ≤ZPm
X dµ(X).
(5.2)
Proof. By Lemma 5.3 choose a sequence {µn} in P1(Pm) such that dW
(hence G(µn) → G(µ) by Theorem 2.3) and
ZPm
X −1 dµn(X) −→ZPm
X dµn(X) −→ZPm
ZPm
1 (µn, µ) → 0
X dµ(X),
X −1 dµ(X).
Since inequalities (5.2) hold for µn, the result follows by taking the limit of (5.2) for
µn.
(cid:3)
Remark 5.5. Let Xij and (X −1)ij denote the (i, j)-entries of X, X −1, respectively.
Then it is clear that the functions X ∈ Pm 7→ Xij, (X −1)ij are integrable with respect
to µ for all i, j = 1, . . . , m if and only if condition (5.1) holds. Hence (5.1) is the best
possible assumption for the AGH mean inequalities in Proposition 5.4 to make sense.
Lemma 5.6. For every µ ∈ P(Pm) with (5.1),
X t dµ(X) −→ZPm
logZPm
1
t
or equivalently,
log X dµ(X),
(cid:18)ZPm
as t → 0 with t ≤ 1.
Proof. First, note that RPm
write
1
t
X tdµ(X)(cid:19)
= expZPm
log X dµ(X)
log X dµ(X) exists by Lemma 5.1. For any X ∈ Pm we
where
X t = et log X = I + t log X + R(t, X),
R(t; X) :=
(log X)n.
tn
n!
∞
Xn=2
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
15
Assuming t ≤ 1 we have
Xn=0
kR(t, X)k ≤ t2
∞
by Lemma 5.1. Therefore,
1
n! k log Xkn = t2ek log Xk ≤ t2(cid:0)kXk + kX −1k(cid:1)
ZPm kR(t, X)k dµ(X) ≤ t2ZPm(cid:0)kXk + kX −1k(cid:1) dµ(X) = O(t2) as t → 0,
so that we have
ZPm
X t dµ(X) = I + tZPm
log X dµ(X) + O(t2).
This implies that
1
t
and hence
logZPm
lim
t→0
1
t
X t dµ(X) =ZPm
logZPm
X t dµ(X) =ZPm
log X dµ(X) + O(t),
log X dµ(X).
Finally, for µ ∈ P(Pm) we consider the condition
(cid:3)
(5.3)
ZPm(cid:0)kXk + kX −1k(cid:1)r dµ(X) < ∞
for some r > 0. It is obvious that if (5.3) holds for r > 0, then it also holds for any
r′ ∈ (0, r]. Moreover, for any r > 0, condition (5.3) is equivalent to
ZPm(cid:0)kX rk + kX −rk(cid:1) dµ(X) < ∞,
so that both µr and µ−r satisfy (5.1).
Our main result of this section is the following:
Theorem 5.7 (Lie-Trotter formula). Let µ ∈ P(Pm) satisfying (5.3) for some r > 0.
Then we have
G(µt)
lim
t→0
1
t = expZPm
log X dµ(X).
(5.4)
16
HIAI AND LIM
Proof. First, assume that µ satisfies (5.1). For any t ∈ [−1, 1] \ {0}, by using Propo-
sition 5.4 to µt we have
(cid:18)ZPm
X −t dµ(X)(cid:19)−1
X −1 dµt(X)(cid:19)−1
=(cid:18)ZPm
≤ G(µt) ≤ZPm
X dµt(X) =ZPm
X t dµ(X).
Since log x is operator monotone on (0,∞), the above inequalities give
−
−
1
t
1
t
logZPm
logZPm
X −t dµ(X) ≤ log G(µt)
X −t dµ(X) ≥ log G(µt)
1
t ≤
1
t ≥
1
t
1
t
logZPm
logZPm
X t dµ(X)
if 0 < t ≤ 1,
X t dµ(X)
if −1 ≤ t < 0.
From Lemma 5.6 this implies that
log G(µt)
lim
t→0
1
t =ZPm
log X dµ(X).
(5.5)
Next, assume that µ satisfies (5.3) for some r > 0, that is, µr satisfies (5.1). The
above case yields
lim
t→0
Note that the left-hand side in the above is
log G(cid:0)(µr)t(cid:1)
1
t =ZPm
log X dµr(X).
log G(µrt)
lim
t→0
1
t = r lim
t→0
log G(µt)
1
t ,
while the right-hand side is
rZPm
log X dµ(X).
Hence we have (5.5) again, which implies (5.4).
(cid:3)
The next corollary extends [6, Corollary 2] to the case of probability measures
satisfying (5.3).
Corollary 5.8. Assume that µ ∈ P(Pm) satisfies (5.3) for an r > 0 and · is any
unitarily invariant norm. Then
(a) For every t > 0,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)G(µ−t)− 1
1
1
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
log X dµ(X)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)(cid:12)
expZPm
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)G(µt)
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) increases to (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) expRPm
log X dµ(X)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) as t ց 0.
,
and (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)G(µt)
(5.6)
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
17
(b) If 0 < t ≤ r, then
1
1
1
.
(5.7)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)ZPm
X −t dµ(X)(cid:19)− 1
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
t (cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)ZPm
≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)G(µt)
(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)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
log X dµ(X)(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)(cid:12)
X t dµ(X)(cid:19)
expZPm
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) decreases to(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) expRPm
Furthermore,(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:0)RPm
X t dµ(X)(cid:1)
X −t dµ(X)(cid:1)− 1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:0)RPm
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) increases to (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) expRPm
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
t (cid:12)(cid:12)(cid:12)(cid:12)
log X dµ(X)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) and
log X dµ(X)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) as r ≥ t ց 0.
Proof. When µ ∈ P 1(Pm) (without condition (5.3)), from the invariance G(µ−1) =
G(µ)−1 as immediately seen from Theorem 3.1, we find that G(µ−t)− 1
t = G(µt)
t ,
implying the equality in (5.6). It follows from (4.6) that (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)G(µt)
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) is increasing as
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)G(µt)
(b) Assume that 0 < t′ < t ≤ r and prove that
dµ(X) ≤(cid:18)ZPm
≥(cid:18)ZPm
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) being increasing noted above.
X t dµ(X)(cid:19)
X −t dµ(X)(cid:19)− t′
(a) The inequality in (5.6) is immediately seen from Theorem 5.7 together with
t ց 0. In the rest, assume (5.3) for an r > 0.
dµ(X)(cid:19)−1
(cid:18)ZPm
ZPm
X −t′
(5.8)
(5.9)
t′
t
,
t
.
1
1
1
X t′
j=1 Aℓ,j1Qℓ,j , ℓ ∈ N, with Aℓ,j ∈ Σn and Borel partitions {Qℓ,j}kℓ
For each n ∈ N let Σn be as in the proof of Lemma 5.3. Since X t and X t′ are
uniformly continuous on the compact set Σn, one can choose a sequence of simple
functions Pkℓ
j=1 of Σn
such that, as ℓ → ∞,
Xj=1
ℓ,jµ(Qℓ,j) −→ZΣn
ℓ,jµ(Ql,j) −→ZΣn
X t µ(X),
Xj=1
µ(X).
X t′
At′
At
kℓ
kℓ
Due to the operator concavity of xt′/t on (0,∞), we have
kℓ
Xj=1
µ(Qℓ,j)At′
ℓ,j + µ(Pm \ Σn)I ≤ kℓ
Xj=1
µ(Qℓ,j)At
ℓ,j + µ(Pm \ Σn)I!
t′
t
.
18
HIAI AND LIM
Letting l → ∞ gives
X t′
ZΣn
dµ(X) + µ(Pm \ Σn)I ≤(cid:18)ZΣn
X t dµ(X) + µ(Pm \ Σn)I(cid:19)
t′
t
.
Since kX tk and kX t′k are integrable with respect to µ, (5.8) follows by taking the
limit of the above inequality as n → ∞. Then, (5.9) also follows by replacing µ
with µ−1 in (5.8). Now, similarly to the proof of [6, Theorem 1] we see that for
1 ≤ j ≤ m, as r ≥ t ց 0, the jth eigenvalue of (cid:0)RPm
is decreasing and
that of (cid:0)RPm
(cid:18)ZPm
X −t dµ(X)(cid:1)− 1
X t dµr(X)(cid:19)
Furthermore, by applying Lemma 5.6 to µr we have
−→ expZPm
as t → 0 with t ≤ 1,
X t dµ(X)(cid:1)
log X dµr(X)
t is increasing.
1
t
1
t
which is rephrased as
1
t
(cid:18)ZPm
X t dµ(X)(cid:19)
log X dµ(X) as t → 0 with t ≤ r.
1
−→ expZPm
X t dµ(X)(cid:1)
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) increases to the same.
In view of (a) it remains to
log X dµ(X)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
t(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) decreases to (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) expRPm
Hence, as r ≥ t ց 0, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:0)RPm
X −t dµ(X)(cid:1)− 1
while (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:0)RPm
show the first inequality in (5.7). But this is immediately seen by applying (5.2) to
µt for 0 < t ≤ r.
Remark 5.9. The following example shows that condition (5.3) is not satisfied for
any r > 0 even if we have µ ∈ P p(Pm) for all p > 0. For instance, choose Xn ∈ Pm
such that Xn ≥ I and kXnk = nn, and define
1
Xn=1
2n δXn.
µ :=
(cid:3)
∞
Then, for any r > 0,
while
ZPm k log Xkp
2 dµ(X) ≤
(nr)n
2n = ∞,
∞
ZPm kXkr dµ(X) =
Xn=1
Xn=1 (cid:0)m log2 kXnk(cid:1)p/2
2n
∞
= mp/2
∞
Xn=1
(n log n)p
2n
< ∞
LOG-MAJORIZATION AND LIE-TROTTER FORMULA
19
for all p > 0.
Problem 5.10. Do Theorem 5.7 and part (a) of Corollary 5.8 hold for general µ ∈
P 1(Pm) without assumption (5.3)? In part (b) of Corollary 5.8, we cannot define
RPm
X ±t dµ(X) for general µ ∈ P 1, while part (a) makes sense for general µ ∈ P 1.
Acknowledgments. The authors thank Hiroyuki Osaka and Takeaki Yamazaki for
inviting the workshop on Quantum Information Theory and Related Topics 2016
in Ritsumeikan University where this work was initiated. The work of F. Hiai was
supported by Grant-in-Aid for Scientific Research (C)26400103. The work of Y. Lim
was supported by the National Research Foundation of Korea (NRF) grant funded
by the Korea government(MEST) No.2015R1A3A2031159 and 2016R1A5A1008055.
References
[1] L. Ambrosio, N. Gigli and G. Savar´e, Gradient Flows in Metric Spaces and in the Space of
Probability Measures, Second edition, Lectures in Mathematics ETH Zurich, Birkhauser Verlag,
Basel, 2008.
[2] T. Ando and F. Hiai, Log majorization and complementary Golden-Thompson type inequalities,
Linear Algebra Appl. 197 (1994), 113 -- 131.
[3] N. Bebiano, R. Lemos, J. Providencia, Inequalities for quantum relative entropy, Linear Algebra
Appl. 401 (2005), 159-172.
[4] H. Araki, On an inequality of Lieb and Thirring, Lett. Math. Phys. 19 (1990), 167 -- 170.
[5] R. Bhatia, Positive Definite Matrices, Princeton University Press, Princeton, NJ, 2007.
[6] R. Bhatia and P. Grover, Norm inequalities related to the matrix geometric mean, Linear
Algebra Appl. 437 (2012), 726 -- 733.
[7] R. Bhatia and J. Holbrook, Riemannian geometry and matrix geometric means, Linear Algebra
Appl. 413 (2006), 594 -- 618.
[8] R. Bhatia and R. Karandikar, Monotonicity of the matrix geometric mean, Math. Ann. 353
(2012), 1453 -- 1467.
[9] R. Bhatia, T. Jain and Y. Lim, On the Bures-Wasserstein distance between positive definite
matrices, submitted.
[10] F. Hiai and D. Petz, Riemannian metrics on positive definite matrices related to means II,
Linear Algebra Appl. 436 (2012), 2117 -- 2136.
[11] F. Hiai and D. Petz, Introduction to Matrix Analysis and Applications, Springer, Cham, 2014.
[12] S. Kim and H. Lee, The power mean and the least squares mean of probability measures on the
space of positive definite matrices, Linear Algebra Appl. 465 (2015), 325 -- 346.
20
HIAI AND LIM
[13] S. Kim, H. Lee and Y. Lim, An order inequality characterizing invariant barycenters on sym-
metric cones, J. Math. Anal. Appl. 442 (2016), 1 -- 16.
[14] J. Lawson and Y. Lim, The geometric mean, matrices, metrics, and more, Amer. Math. Monthly
108 (2001), 797 -- 812.
[15] J. Lawson and Y. Lim, Monotonic properties of the least squares mean, Math. Ann. 351 (2011),
267 -- 279.
[16] J. Lawson and Y. Lim, The least squares mean of positive Hilbert-Schmidt operators, J. Math.
Anal. Appl. 403 (2013), 365 -- 375.
[17] J. Lawson and Y. Lim, Contractive barycentric maps, to appear in J. Operator Theory.
[18] K.-T. Sturm, Probability measures on metric spaces of nonpositive curvature, Heat kernels
and analysis on manifolds, graphs, and metric spaces (Paris, 2002), 357 -- 390, Contemporary
Mathematics, 338, Amer. Math. Soc., Providence, RI, 2003.
[19] C. Villani, Topics in Optimal Transportation, Graduate Studies in Mathematics, Vol. 58, Amer.
Math. Soc., Providence, RI, 2003.
[20] T. Yamazaki, The Riemannian mean and matrix inequalities related to the Ando-Hiai inequality
and chaotic order, Operators and Matrices 6 (2012), 577 -- 588.
[21] T. Yamazaki, An elementary proof of arithmetic-geometric mean inequality of the weighted
Riemannian mean of positive definite matrices, Linear Algebra Appl. 438 (2013), 1564 -- 1569.
Tohoku University (Emeritus), Hakusan 3-8-16-303, Abiko 270-1154, Japan
E-mail address: [email protected]
Department of Mathematics, Sungkyunkwan University, Suwon 440-746, Korea
E-mail address: [email protected]
|
1009.4359 | 2 | 1009 | 2010-11-22T12:48:43 | Compactly Supported Shearlets | [
"math.FA"
] | Shearlet theory has become a central tool in analyzing and representing 2D data with anisotropic features. Shearlet systems are systems of functions generated by one single generator with parabolic scaling, shearing, and translation operators applied to it, in much the same way wavelet systems are dyadic scalings and translations of a single function, but including a precise control of directionality. Of the many directional representation systems proposed in the last decade, shearlets are among the most versatile and successful systems. The reason for this being an extensive list of desirable properties: shearlet systems can be generated by one function, they provide precise resolution of wavefront sets, they allow compactly supported analyzing elements, they are associated with fast decomposition algorithms, and they provide a unified treatment of the continuum and the digital realm.
The aim of this paper is to introduce some key concepts in directional representation systems and to shed some light on the success of shearlet systems as directional representation systems. In particular, we will give an overview of the different paths taken in shearlet theory with focus on separable and compactly supported shearlets in 2D and 3D. We will present constructions of compactly supported shearlet frames in those dimensions as well as discuss recent results on the ability of compactly supported shearlet frames satisfying weak decay, smoothness, and directional moment conditions to provide optimally sparse approximations of cartoon-like images in 2D as well as in 3D. Finally, we will show that these compactly supported shearlet systems provide optimally sparse approximations of an even generalized model of cartoon-like images comprising of $C^2$ functions that are smooth apart from piecewise $C^2$ discontinuity edges. | math.FA | math |
Compactly Supported Shearlets
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
Abstract Shearlet theory has become a central tool in analyzing and representing 2D
data with anisotropic features. Shearlet systems are systems of functions generated
by one single generator with parabolic scaling, shearing, and translation operators
applied to it, in much the same way wavelet systems are dyadic scalings and trans-
lations of a single function, but including a precise control of directionality. Of the
many directional representation systems proposed in the last decade, shearlets are
among the most versatile and successful systems. The reason for this being an exten-
sive list of desirable properties: shearlet systems can be generated by one function,
they provide precise resolution of wavefront sets, they allow compactly supported
analyzing elements, they are associated with fast decomposition algorithms, and
they provide a unified treatment of the continuum and the digital realm.
The aim of this paper is to introduce some key concepts in directional representa-
tion systems and to shed some light on the success of shearlet systems as directional
representation systems. In particular, we will give an overview of the different paths
taken in shearlet theory with focus on separable and compactly supported shearlets
in 2D and 3D. We will present constructions of compactly supported shearlet frames
in those dimensions as well as discuss recent results on the ability of compactly sup-
ported shearlet frames satisfying weak decay, smoothness, and directional moment
conditions to provide optimally sparse approximations of cartoon-like images in
2D as well as in 3D. Finally, we will show that these compactly supported shearlet
systems provide optimally sparse approximations of an even generalized model of
Gitta Kutyniok
Institute of Mathematics, University of Osnabrück, 49069 Osnabrück, Germany, e-mail:
[email protected]
Jakob Lemvig
Institute of Mathematics, University of Osnabrück, 49069 Osnabrück, Germany, e-mail:
[email protected]
Wang-Q Lim
Institute of Mathematics, University of Osnabrück, 49069 Osnabrück, Germany, e-mail:
[email protected]
1
2
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
cartoon-like images comprising of C2 functions that are smooth apart from piece-
wise C2 discontinuity edges.
1 Introduction
Recent advances in modern technology have created a brave new world of enor-
mous, multi-dimensional data structures. In medical imaging, seismic imaging, as-
tronomical imaging, computer vision, and video processing, the capabilities of mod-
ern computers and high-precision measuring devices have generated 2D, 3D, and
even higher dimensional data sets of sizes that were infeasible just a few years ago.
The need to efficiently handle such diverse types and huge amounts of data initiated
an intense study in developing efficient multivariate encoding methodologies in the
applied harmonic analysis research community.
In medical imaging, e.g., CT lung scans, the discontinuity curves of the image are
important specific features since one often wants to distinguish between the image
'objects' (e.g., the lungs) and the 'background'; that is, it is important to precisely
capture the edges. This observation holds for various other applications than medi-
cal imaging and illustrates that important classes of multivariate problems are gov-
erned by anisotropic features. Moreover, in high-dimensional data most information
is typically contained in lower-dimensional embedded manifolds, thereby also pre-
senting itself as anisotropic features. The anisotropic structures can be distinguished
by location and orientation/direction which indicates that our way of analyzing and
representing the data should capture not only location, but also directional informa-
tion.
In applied harmonic analysis, data is typically modeled in a continuum setting as
square-integrable functions or, more generally, as distributions. Recently, a novel di-
rectional representation system -- so-called shearlets -- has emerged which provides
a unified treatment of such continuum models as well as digital models, allowing,
for instance, a precise resolution of wavefront sets, optimally sparse representations
of cartoon-like images, and associated fast decomposition algorithms. Shearlet sys-
tems are systems generated by one single generator with parabolic scaling, shearing,
and translation operators applied to it, in the same way wavelet systems are dyadic
scalings and translations of a single function, but including a directionality charac-
teristic owing to the additional shearing operation (and the anisotropic scaling).
The aim of this survey paper is to introduce the key concepts in directional rep-
resentation systems and, in particular, to shed some light on the success of shearlet
systems. Moreover, we will give an overview of the different paths taken in shearlet
theory with focus on separable and compactly supported shearlets, since these sys-
tems are most well-suited for applications in, e.g., image processing and the theory
of partial differential equations.
Compactly Supported Shearlets
3
1.1 Directional Representation Systems
In recent years numerous approaches for efficiently representing directional features
of two-dimensional data have been proposed. A perfunctory list includes: steer-
able pyramid by Simoncelli et al. [40], directional filter banks by Bamberger and
Smith [2], 2D directional wavelets by Antoine et al. [1], curvelets by Candès and
Donoho [4], contourlets by Do and Vetterli [10], bandlets by LePennec and Mal-
lat [39], and shearlets by Labate, Weiss, and two of the authors [37]. Of these,
shearlets are among the most versatile and successful systems which owes to the
many desirable properties possessed by shearlet systems: they are generated by one
function, they provide optimally sparse approximation of so-called cartoon-like im-
ages, they allow compactly supported analyzing elements, they are associated with
fast decomposition algorithms, and they provide a unified treatment of continuum
and digital data.
Cartoon-like images are functions that are C2 apart from C2 singularity curves,
and the problem of sparsely representing such singularities using 2D representation
systems has been extensively studied; only curvelets [5], contourlets [10], and shear-
lets [21] are known to succeed in this task in an optimal way (see also Section 3). We
describe contourlets and curvelets in more details in Section 1.4 and will here just
mention some differences to shearlets. Contourlets are constructed from a discrete
filter bank and have therefore, unlike shearlets, no continuum theory. Curvelets, on
the other hand, are a continuum-domain system which, unlike shearlets, does not
transfer in a uniform way to the digital world. It is fair to say that shearlet theory is
a comprehensive theory with a mathematically rich structure as well as a superior
connection between the continuum and digital realm.
The missing link between the continuum and digital world for curvelets is caused
by the use of rotation as a means to parameterize directions. One of the distinctive
features of shearlets is the use of shearing in place of rotation; this is, in fact, de-
cisive for a clear link between the continuum and digital world which stems from
the fact that the shear matrix preserves the integer lattice. Traditionally, the shear
parameter ranges over a non-bounded interval. This has the effect that the direc-
tions are not treated uniformly, which is particularly important in applications. On
the other hand, rotations clearly do not suffer from this deficiency. To overcome
this shortcoming of shearing, Guo, Labate, and Weiss together with two of the au-
thors [37] (see also [20]) introduced the so-called cone-adapted shearlet systems,
where the frequency plane is partitioned into a horizontal and a vertical cone which
allows restriction of the shear parameter to bounded intervals (Section 2.1), thereby
guaranteeing uniform treatment of directions.
Shearlet systems therefore come in two ways: One class being generated by a
unitary representation of the shearlet group and equipped with a particularly 'nice'
mathematical structure, however causes a bias towards one direction, which makes
it unattractive for applications; the other class being generated by a quite similar
procedure, but restricted to cones in frequency domain, thereby ensuring an equal
treatment of all directions. To be precise this treatment of directions is only 'almost
equal' since there still is a slight, but controllable, bias towards directions of the
4
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
coordinate axes, see also Figure 4 in Section 2.2. For both classes, the continuous
shearlet systems are associated with a 4-dimensional parameter space consisting of
a scale parameter measuring the resolution, a shear parameter measuring the orien-
tation, and a translation parameter measuring the position of the shearlet (Section
1.3). A sampling of this parameter space leads to discrete shearlet systems, and it is
obvious that the possibilities for this are numerous. Using dyadic sampling leads to
so-called regular shearlet systems which are those discrete systems mainly consid-
ered in this paper. It should be mentioned that also irregular shearlet systems have
attracted some attention, and we refer to the papers [27 -- 29]. We end this section by
remarking that these discrete shearlet systems belong to a larger class of represen-
tation systems -- the so-called composite wavelets [23 -- 25].
1.2 Anisotropic Features, Discrete Shearlet Systems, and Quest for
Sparse Approximations
In many applications in 2D and 3D imaging the important information is often
located around edges separating 'image objects' from 'background'. These fea-
tures correspond precisely to the anisotropic structures in the data. Two-dimensional
shearlet systems are carefully designed to efficiently encode such anisotropic fea-
tures. In order to do this effectively, shearlets are scaled according to a parabolic
scaling law, thereby exhibiting a spatial footprint of size 2− j times 2− j/2, where 2 j
is the (discrete) scale parameter; this should be compared to the size of wavelet foot-
prints: 2− j times 2− j. These elongated, scaled needle-like shearlets then parametrize
directions by slope encoded in a shear matrix. As mentioned in the previous section,
such carefully designed shearlets do, in fact, perform optimally when representing
and analyzing anisotropic features in 2D data (Section 3).
In 3D the situation changes somewhat. While in 2D we 'only' have to handle one
type of anisotropic structures, namely curves, in 3D a much more complex situation
can occur, since we find two geometrically very different anisotropic structures:
Curves as one-dimensional features and surfaces as two-dimensional anisotropic
features. Our 3D shearlet elements in spatial domain will be of size 2− j times 2− j/2
times 2− j/2 which corresponds to 'plate-like' elements as j → ¥
. This indicates that
these 3D shearlet systems have been designed to efficiently capture two-dimensional
anisotropic structures, but neglecting one-dimensional structures. Nonetheless, sur-
prisingly, these 3D shearlet systems still perform optimally when representing and
analyzing 3D data that contain both curve and surface singularities (Section 4).
Of course, before we can talk of optimally sparse approximations, we need to
actually have these 2D and 3D shearlet systems at hand. Several constructions of
discrete band-limited 2D shearlet frames are already known, see [6, 20, 28, 29]. But
since spatial localization of the analyzing elements of the encoding system is im-
mensely important both for a precise detection of geometric features as well as
for a fast decomposition algorithm, we will mainly follow the sufficient conditions
for and construction of compactly supported cone-adapted 2D shearlet systems by
Compactly Supported Shearlets
5
Kittipoom and two of the authors [27] (Section 2.3). These results provide a large
class of separable, compactly supported shearlet systems with good frame bounds,
optimally sparse approximation properties, and associated numerically stable algo-
rithms.
1.3 Continuous Shearlet Systems
Discrete shearlet systems are, as mentioned, a sampled version of the so-called con-
tinuous shearlet systems. These continuous shearlets come, of course, also in two
different flavors, and we will briefly describe these in this section.
1.3.1 Cone-Adapted Shearlet Systems
Anisotropic features in multivariate data can be modeled in many different ways.
One possibility is the cartoon-like image class discussed above, but one can also
model such directional singularities through distributions. One would, for example,
model a one-dimensional anisotropic structure as the delta distribution of a curve.
The so-called cone-adapted continuous shearlet transform associated with cone-
adapted continuous shearlet systems was introduced by Labate and the first author
in [30] in the study of resolutions of the wavefront set for such distributions. It was
shown that the continuous shearlet transform is not only able to identify the singular
support of a distribution, but also the orientation of distributed singularities along
curves. More precisely, for a class of band-limited shearlet generators y ∈ L2(R2),
the first author and Labate [30] showed that the wavefront set of a (tempered) dis-
tribution f is precisely the closure of the set of points (t,s), where the shearlet
transform of f
(a,s,t) 7→D f ,a−3/4y (A−1
a S−1
s (· − t))E, where Aa =(cid:18)a 0
0 a1/2(cid:19) and Ss =(cid:18)1 s
0 1(cid:19) ,
is not of fast decay as the scale parameter a → 0. Later Grohs [18] extended this
result to Schwartz-class generators with infinitely many directional vanishing mo-
ments, in particular, not necessarily band-limited generators. In other words, these
results demonstrate that the wavefront set of a distribution can be precisely cap-
tured by continuous shearlets. For constructions of continuous shearlet frames with
compact support, we refer to [19].
1.3.2 Shearlets from Group Representations
Cone-adapted continuous shearlet systems and their associated cone-adapted con-
tinuous transforms described in the previous section have only very recently -- in
6
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
2009 -- attracted attention. Historically, the continuous shearlet transform was first
introduced in [20] without restriction to cones in frequency domain. Later, it was
shown in [7] that the associated continuous shearlet systems are generated by a
strongly continuous, irreducible, square-integrable representation of a locally com-
pact group, the so-called shearlet group. This implies that these shearlet systems
possess a rich mathematical structure, which in [7] was used to derive uncertainty
principles to tune the accuracy of the shearlet transform, and which in [6] allowed
the usage of coorbit theory to study smoothness spaces associated with the decay of
the shearlet coefficients.
Dahlke, Steidl, and Teschke generalized the shearlet group and the associated
continuous shearlet transform to higher dimensions Rn in the paper [8]. Further-
more, in [8] they showed that, for certain band-limited generators, the continuous
shearlet transform is able to identify hyperplane and tetrahedron singularities. Since
this transform originates from a unitary group representation, it is not able to capture
all directions, in particular, it will not capture the delta distribution on the x1-axis
(and more generally, any singularity with 'x1-directions'). We also remark that the
extension in [8] uses another scaling matrix as compared to the one used for the
three-dimensional shearlets considered in this paper; we refer to Section 4 for a
more detailed description of this issue.
1.4 Applications
Shearlet theory has applications in various areas. In this section we will present two
examples of such: Denoising of images and geometric separation of data. Before,
in order to show the reader the advantages of digital shearlets, we first give a short
overview of the numerical aspects of shearlets and two similar implementations of
directional representation systems, namely contourlets and curvelets, discussed in
Section 1.1.
Curvelets [3]. This approach builds on directional frequency partitioning and the
use of the Fast Fourier transform. The algorithm can be efficiently implemented
using (in frequency domain) multiplication with the frequency response of a filter
and frequency wrapping in place of convolution and down-sampling. However,
curvelets need to be band-limited and can only have very good spatial localiza-
tion if one allows high redundancy.
Contourlets [10]. This approach uses a directional filter bank, which produces direc-
tional frequency partitioning similar to those of curvelets. As the main advantage
of this approach, it allows a tree-structured filter bank implementation, in which
aliasing due to subsampling is allowed to exist. Consequently, one can achieve
great efficiency in terms of redundancy and good spatial localization. However,
the directional selectivity in this approach is artificially imposed by the special
sampling rule of a filter bank which introduces various artifacts. We remark that
also the recently introduced Hybrid Wavelets [17] suffer from this deficiency.
Compactly Supported Shearlets
7
(a) Original image
(b) Noisy image
(c) Denoised using curvelets
(d) Denoised using shearlets
(e) Denoised using curvelets (zoom)
(f) Denoised using shearlets (zoom)
Fig. 1 Denoising of the Goldhill image (512 × 512) using shearlets and curvelets. The noisy image
in (b) has a peak signal-to-noise ratio of 20.17 dB. The curvelet-denoised image in (c) and (e) has
a PSNR of 28.70 dB, while the shearlet-denoised image in (d) and (f) has a PSNR of only 29.20
dB.
8
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
Shearlets [38]. Using a shear matrix instead of rotation, directionality is natu-
rally adapted for the digital setting in the sense that the shear matrix preserves
the structure of the integer grid. Furthermore, excellent spatial localization is
achieved by using compactly supported shearlets. The only drawback is that
these compactly supported shearlets are not tight frames and, accordingly, the
synthesis process needs to be performed by iterative methods.
To illustrate how two of these implementations perform, we have included a
denoising example of the Goldhill image using both curvelets1 and shearlets, see
Figure 1. We omit a detailed analysis of the denoising results and leave the visual
comparison to the reader. For a detailed review of the shearlet transform and asso-
ciated aspects, we refer to [14, 16, 36, 38]. We also refer to [26, 35] for MRA based
algorithmic approaches to the shearlet transform.
The shearlet transform, in companion with the wavelet transform, has also been
applied to accomplish geometric separation of 'point-and-curve'-like data. An artifi-
cially made example of such data can be seen in Figure 2a. For a theoretical account
of these separation ideas we refer to the recent papers by Donoho and the first au-
thor [12, 13]. Here we simply display the result of the separation, see Figure 2. For
real-world applications of these separation techniques we refer to the paper [33] on
neurobiological imaging.
(a) Original point-curve data
of size 256 × 256.
(b) Separated point-like data
(captured by wavelets).
(c) Separated curve-like data
(captured by shearlets).
Fig. 2 Geometric separation of mixed 'point-and-curve' data. (a): Input data. (b) and (c): The
output of the separation algorithm.
In the spirit of reproducible research [15], we wish to mention that Figure 1d,
1f and 2 have been produced by the discrete shearlet transform implemented in the
Matlab toolbox Shearlab which has recently been released under a GNU license and
is freely available at http://www.shearlab.org.
1 Produced using Curvelab (Version 2.1.2), which is available from http://curvelet.org.
Compactly Supported Shearlets
1.5 Outline
9
In Section 2 we present a review of shearlet theory in L2(R2), where we focus
on discrete shearlet systems. We describe the classical band-limited construction
(Section 2.2) and a more recent construction of compactly supported shearlets (Sec-
tion 2.3). In Section 3 we present results on the ability of shearlets to optimally
sparsely approximate cartoon-like images. Section 4 is dedicated to a discussion on
similar properties of 3D shearlet systems.
2 2D Shearlets
In this section, we summarize what is known about constructions of discrete shear-
let systems in 2D. Although all results in this section can easily be extended to
(irregular) shearlet systems associated with a general irregular set of parameters for
scaling, shear, and translation, we will only focus on the discrete shearlet systems
associated with a regular set of parameters as described in the next section. For a
detailed analysis of irregular shearlet systems, we refer to [27]. We first start with
various notations and definitions for later use.
2.1 Preliminaries
For j ≥ 0,k ∈ Z, let
A2 j =(cid:18)2 j 0
0 2 j/2(cid:19) , Sk =(cid:18)1 k
0 1(cid:19) ,
and Mc =(cid:18)c1 0
0 c2(cid:19) ,
where c = (c1,c2) and c1,c2 are some positive constants. Similarly, we define
A2 j =(cid:18)2 j/2 0
0 2 j(cid:19) , Sk =(cid:18)1 0
k 1(cid:19) ,
and
Mc =(cid:18)c2 0
0 c1(cid:19) .
Next we define discrete shearlet systems in 2D.
Definition 1. Let c = (c1,c2) ∈ (R+)2. For f ,y , y ∈ L2(R2) the cone-adapted 2D
discrete shearlet system SH(f ,y , y ;c) is defined by
SH(f ,y , y ;c) = F (f ;c1) ∪Y (y ;c) ∪ Y ( y ;c),
where
F (f ;c1) = {f (· − m) : m ∈ c1Z2},
Y (y ;c) = {2
3
4 jy (SkA2 j · −m) : j ≥ 0, −⌈2 j/2⌉ ≤ k ≤ ⌈2 j/2⌉,m ∈ McZ2},
10
and
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
Y ( y ;c) = {2
3
4 j y ( Sk A2 j · −m) : j ≥ 0, −⌈2 j/2⌉ ≤ k ≤ ⌈2 j/2⌉,m ∈ McZ2}.
If SH(f ,y , y ;c) is a frame for L2(R2), we refer to f as a scaling function and y
and y as shearlets.
Our aim is to construct compactly supported functions f ,y
to obtain com-
pactly supported shearlets in 2D. For this, we will describe general sufficient con-
ditions on the shearlet generators y and y
, which lead to the construction of com-
pactly supported shearlets. To formulate our sufficient conditions on y and y
(Sec-
tion 2.3), we will first need to introduce the necessary notational concepts.
, and y
For functions f ,y , y ∈ L2(R2), we define Q
: R2 × R2 → R by
Q (x ,w ) = f (x ) f (x + w ) +Q 1(x ,w ) +Q 2(x ,w ),
where
and
Q 1(x ,w ) = (cid:229)
j≥0
Q 2(x ,w ) = (cid:229)
k≤⌈2 j/2⌉(cid:12)(cid:12)
k≤⌈2 j/2⌉(cid:12)(cid:12)
y (ST
k A2− jx )(cid:12)(cid:12)(cid:12)(cid:12)
y (Sk A2− jx )(cid:12)(cid:12)(cid:12)(cid:12)
y (Sk
T A2− jx + w )(cid:12)(cid:12)
y (Sk A2− jx + w )(cid:12)(cid:12) .
j≥0
Also, for c = (c1,c2) ∈ (R+)2, let
R(c) = (cid:229)
m∈Z2\{0}(cid:0)G 0(c−1
+ (G 2( M−1
1 m)G 0(−c−1
1 m)(cid:1)
c m)G 2(− M−1
c m))
1
2 ,
1
2 +(cid:0)G 1(M−1
c m)G 1(−M−1
c m)(cid:1)
(1)
1
2
where
G 0(w ) = esssup
x ∈R2
f (x ) f (x + w )
and G i(w ) = esssup
x ∈R2
Q
i(x ,w )
for i = 1,2.
2.2 Classical Construction
We now first describe the construction of band-limited shearlets which provides
tight frames for L2(R2). Constructions of this type were first introduced by Labate,
Weiss, and two of the authors in [37]. The classical example of a generating shearlet
is a function y ∈ L2(R2) satisfying
y (x ) = y (x 1,x 2) = y 1(x 1) y 2(
x 2x 1
),
(cid:229)
(cid:229)
Compactly Supported Shearlets
11
where y 1 ∈ L2(R) is a discrete wavelet, i.e., satisfies the discrete Calderón condition
given by
y 1(2− jx )2 = 1
for a.e. x ∈ R,
j∈Z
with y 1 ∈ C¥
tion, namely
(R) and supp y 1 ⊆ [− 5
4 , − 1
4 ] ∪ [ 1
4 , 5
4 ], and y 2 ∈ L2(R) is a bump func-
1(cid:229)
y 2(x + k)2 = 1 for a.e. x ∈ [−1,1],
k=−1
satisfying y 2 ∈ C¥
(R) and supp y 2 ⊆ [−1,1]. There are several choices of y 1 and y 2
satisfying those conditions, and we refer to [20] for further details. The tiling of the
C2
C3
C1
R
C4
Fig. 3 The cones C1 -- C4 and the
centered rectangle R in the fre-
quency domain.
Fig. 4 Tiling of the frequency
domain induced by band-limited
shearlets.
frequency domain given by these band-limited generators and choosing y (x1,x2) =
y (x2,x1) is illustrated in Figure 4. As described in Figure 3, a conic region C1 ∪ C3 is
covered by the frequency support of shearlets inY (y ;c) while C2 ∪ C4 is covered by
Y ( y ;c). For this particular choice, using an appropriate scaling function f
for the
centered rectangle R (see Figure 3), it was proved in [20, Thm. 3] that the associated
cone-adapted discrete shearlet system SH(f ,y , y ; (1,1)) forms a Parseval frame for
L2(R2).
2.3 Constructing Compactly Supported Shearlets
We are now ready to state general sufficient conditions for the construction of shear-
let frames.
Theorem 1 ([27]). Let f ,y ∈ L2(R2) be functions such that
f (x 1,x 2) ≤ C1 · min {1, x 1−g
} · min{1, x 2−g
}
and
(cid:229)
12
y (x 1,x 2) ≤ C2 · min{1, x 1
(2)
for some positive constants C1,C2 < ¥ and a > g > 3. Define y (x1,x2) = y (x2,x1),
and let Linf,Lsup be defined by
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
} · min{1, x 1−g
} · min {1, x 2−g
},
a
Linf = essinf
x ∈R2
Q (x ,0) and Lsup = esssup
x ∈R2
Q (x ,0).
Suppose that there is a constant Linf > 0 such that 0 < Linf ≤ Linf. Then there exist
a sampling parameter c = (c1,c2) with c1 = c2 and a constant Lsup < ¥
such that
R(c) < Lin f ≤ Lin f and Lsup ≤ Lsup,
and, further, SH(f ,y , y
satisfying
;c) forms a frame for L2(R2) with frame bounds A and B
1
detMc
[Lin f − R(c)] ≤ A ≤ B ≤
1
detMc
[Lsup + R(c)].
(3)
For a detailed proof, we refer to the paper [27] by Kittipoom and two of the authors.
Obviously, band-limited shearlets (from Section 2.2) satisfy condition (2). More
interestingly, also a large class of spatially compactly supported functions satisfies
this condition. In fact, in [27], various constructions of compactly supported shear-
lets are presented using Theorem 1 and generalized low-pass filters; an example of
such a construction procedure is given in Theorem 2 below. In Theorem 1 we as-
Mc), the only reason for this being
sumed c1 = c2 for the sampling matrix Mc (or
the simplification of the estimates for the frame bounds A,B in (3). In fact, the esti-
mate (3) generalizes easily to non-uniform sampling constants c1,c2 with c1 6= c2.
For explicit estimates of the form (3) in the case of non-uniform sampling, we refer
to [27].
The following result provides a specific family of functions satisfying the general
sufficiency condition from Theorem 1.
Theorem 2 ([27]). Let K,L > 0 be such that L ≥ 10 and 3L
a shearlet y ∈ L2(R2) by
2 ≤ K ≤ 3L−2, and define
y (x ) = m1(4x 1) f (x 1) f (2x 2),
x = (x 1,x 2) ∈ R2,
where m0 is the low pass filter satisfying
m0(x 1)2 = (cos(px
1))2K
L−1
n=0(cid:18)K − 1 + n
n
(cid:19)(sin(px
1))2n,
x 1 ∈ R,
m1 is the associated bandpass filter defined by
m1(x 1)2 = m0(x 1 + 1
2 )2,
x 1 ∈ R,
and f
is the scaling function given by
(cid:229)
Compactly Supported Shearlets
13
f (x 1) =
m0(2− jx 1),
x 1 ∈ R.
j=0
Then there exists a sampling constant c1 > 0 such that the shearlet system Y (y ;c)
forms a frame for L2(C1 ∪ C3) := (cid:8) f ∈ L2(R2) : supp f ⊂ C1 ∪ C3(cid:9) for any sam-
pling matrix Mc with c = (c1,c2) ∈ (R+)2 and c2 ≤ c1 ≤ c1.
For these shearlet systems, there is a bias towards the vertical axis, especially
at coarse scales, since they are defined for L2(C1 ∪ C3), and hence, the frequency
support of the shearlet elements overlaps more significantly along the vertical axis.
In order to control the upper frame bound, it is therefore desirable to apply a denser
sampling along the vertical axis than along the horizontal axis, i.e., c1 > c2.
Having compactly supported (separable) shearlet frames for L2(C1 ∪ C3) at hand
by Theorem 2, we can easily construct shearlet frames for the whole space L2(R2).
The exact procedure is described in the following theorem from [27].
Theorem 3 ([27]). Let y ∈ L2(R2) be the shearlet with associated scaling function
f 1 ∈ L2(R) both introduced in Theorem 2, and set f (x1,x2) = f 1(x1)f 1(x2) and
y (x1,x2) = y (x2,x1). Then the corresponding shearlet system SH(f ,y , y ;c) forms
a frame for L2(R2) for any sampling matrices Mc and Mc with c = (c1,c2) ∈ (R+)2
and c2 ≤ c1 ≤ c1.
For the horizontal cone C1 ∪ C3 we allow for a denser sampling by Mc along the
vertical axis, i.e., c2 ≤ c1, precisely as in Theorem 2. For the vertical cone C2 ∪ C4
we analogously allow for a denser sampling along the horizontal axis; since the
position of c1 and c2 is reversed in Mc compared to Mc, this still corresponds to
c2 ≤ c1.
We wish to mention that there is a trade-off between compact support of the
shearlet generators, tightness of the associated frame, and separability of the shear-
let generators. The known constructions of tight shearlet frames do not use sep-
arable generators (Section 2.2), and these constructions can be shown to not be
applicable to compactly supported generators. Tightness is difficult to obtain while
allowing for compactly supported generators, but we can gain separability as in
Theorem 3, hence fast algorithmic realizations. On the other hand, when allowing
non-compactly supported generators, tightness is possible, but separability seems to
be out of reach, which makes fast algorithmic realizations very difficult.
We end this section by remarking that the construction results above even gener-
alize to constructions of irregular shearlet systems [28, 29].
3 Sparse Approximations
After having introduced compactly supported shearlet systems in the previous sec-
tion, we now aim for optimally sparse approximations. To be precise, we will show
that these compactly supported shearlet systems provide optimally sparse approxi-
mations when representing and analyzing anisotropic features in 2D data.
¥
(cid:213)
14
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
3.1 Cartoon-Like Image Model
Following [11], we introduce STAR2(n ), a class of sets B with C2 boundaries ¶ B
and curvature bounded by n , as well as E 2n (R2), a class of cartoon-like images. For
this, in polar coordinates, we let r
: [0,2p ) → [0,1] be a radius function and define
the set B by
B = {x ∈ R2 : x ≤ r (q ), x = (x ,q ) in polar coordinates}.
In particular, we will require that the boundary ¶ B of B is given by the curve
b (q ) =(cid:18)r (q )cos(q )
r (q )sin(q )(cid:19) ,
and the class of boundaries of interest to us are defined by
r ≤ r 0 < 1,
(q ) ≤ n ,
supr ′′
(4)
(5)
where r 0 < 1 needs to be chosen so that y + B ⊂ [0,1]2 for some y ∈ R2.
The following definition now introduces a class of cartoon-like images.
Definition 2. For n > 0, the set STAR2(n ) is defined to be the set of all B ⊂ [0,1]2
such that B is a translate of a set obeying (4) and (5). Further, E 2n (R2) denotes the
set of functions f ∈ L2(R2) of the form
f = f0 + f1c B,
0(R2) with supp fi ⊂ [0,1]2 and k fikC2 =
where B ∈ STAR2(n ) and f0, f1 ∈ C2
a ≤2 kDa fik¥ ≤ 1 for i = 1,2.
One can also consider a more sophisticated class of cartoon-like images, where the
boundary of B is allowed to be piecewise C2, and we refer to the recent paper by
two of the authors [34] and to similar considerations for the 3D case in Section 4.2.
Donoho [11] proved that the optimal approximation rate for such cartoon-like
image models f ∈ E 2n (R2) which can be achieved for almost any representation sys-
tem under a so-called polynomial depth search selection procedure of the selected
system elements is
k f − fNk2
2 ≤ C · N−2
as N → ¥
,
where fN is the best N-term approximation of f . As discussed in the next section
shearlets in 2D do indeed deliver this optimal approximation rate.
3.2 Optimally Sparse Approximation of Cartoon-Like Images
Let SH(f ,y , y ;c) be a shearlet frame for L2(R2). Since this is a countable set
of functions, we can denote it by SH(f ,y , y ;c) = (s i)i∈I. We let ( s i)i∈I be a
(cid:229)
a
} · min{1, x 1−g
x 1(cid:17)−g
≤ h(x 1) ·(cid:16)1 + x 2
,
¶
¶x 2
(ii) (cid:12)(cid:12)(cid:12)
y (x )(cid:12)(cid:12)(cid:12)
Compactly Supported Shearlets
15
dual frame of (s i)i∈I. As our N-term approximation fN of a cartoon-like image
f ∈ E 2n (R2) by the frame SH(f ,y , y ;c), we then take
fN = (cid:229)
h f ,s ii s i,
i∈IN
where (h f ,s ii)i∈IN are the N largest coefficients h f ,s ii in magnitude. As in the tight
frame case, this procedure does not always yield the best N-term approximation,
but, surprisingly, even with this rather crude selection procedure, we can prove an
(almost) optimally sparse approximation rate. We speak of 'almost' optimality due
to the (negligible) log-factor in (6). The following result shows that our 'new' com-
pactly supported shearlets (see Section 2.3) deliver the same approximation rate as
band-limited curvelets [5], contourlets [10], and shearlets [21].
Theorem 4 ( [32]). Let c > 0, and let f ,y , y ∈ L2(R2) be compactly supported.
Suppose that, in addition, for all x = (x 1,x 2) ∈ R2, the shearlet y
(i) y (x ) ≤ C1 · min{1, x 1
}
} · min{1, x 2−g
satisfies
and
where a > 5, g ≥ 4, h ∈ L1(R), and C1 is a constant, and suppose that the shearlet
y
satisfies (i) and (ii) with the roles of x 1 and x 2 reversed. Further, suppose that
SH(f ,y , y ;c) forms a frame for L2(R2).
Then, for any n > 0, the shearlet frame SH(f ,y , y ;c) provides (almost) opti-
mally sparse approximations of functions f ∈ E 2n (R2) in the sense that there exists
some C > 0 such that
k f − fNk2
2 ≤ C · N−2 · (logN)3
as N → ¥
,
(6)
where fN is the nonlinear N-term approximation obtained by choosing the N largest
shearlet coefficients of f .
Condition (i) can be interpreted as both a condition ensuring (almost) separable
behavior as well as a moment condition along the horizontal axis, hence enforc-
ing directional selectivity. This condition ensures that the support of shearlets in
frequency domain is essentially of the form indicated in Figure 4. Condition (ii)
(together with (i)) is a weak version of a directional vanishing moment condition2,
which is crucial for having fast decay of the shearlet coefficients when the corre-
sponding shearlet intersects the discontinuity curve. Conditions (i) and (ii) are rather
mild conditions on the generators; in particular, shearlets constructed by Theorem 2
and 3, with extra assumptions on the parameters K and L, will indeed satisfy (i)
and (ii) in Theorem 4. To compare with the optimality result for band-limited gen-
erators we wish to point out that conditions (i) and (ii) are obviously satisfied for
band-limited generators.
2 For the precise definition of directional vanishing moments, we refer to [10].
16
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
We remark that this kind of approximation result is not available for shearlet
systems coming directly from the shearlet group. One reason for this being that
these systems, as mentioned several times, do not treat directions in a uniform way.
4 Shearlets in 3D and Beyond
Shearlet theory has traditionally only dealt with representation systems for two-
dimensional data. In the recent paper [8] (and the accompanying paper [9]) this
was changed when Dahlke, Steidl, and Teschke generalized the continuous shear-
let transform (see [7, 30]) to higher dimensions. The shearlet transform on L2(Rn)
by Dahlke, Steidl, and Teschke is associated with the so-called shearlet group in
R \ {0} × Rn−1 × Rn, with a dilation matrix of the form
Aa = diag (a,sgn(a) a1/n , . . . ,sgn(a) a1/n),
a ∈ R \ {0},
and with a shearing matrix with n − 1 shear parameters s = (s1, . . . ,sn−1) ∈ Rn−1 of
the form
Ss =(cid:20)1 s
0 In−1(cid:21) ,
where In denotes the n × n identity matrix. This type of shearing matrix gives rise to
shearlets consisting of wedges of size a−1 ×a−1/n ×· · ·×a−1/n in frequency domain,
where a−1 ≫ a−1/n for small a > 0. Hence, for small a > 0, the spatial appearance
is a surface-like element of co-dimension one.
In the following section we will consider shearlet systems in L2(R3) associ-
ated with a sightly different shearing matrix. More importantly, we will consider
pyramid-adapted 3D shearlet systems, since these systems treat directions in a uni-
form way as opposed to the shearlet systems coming from the shearlet group; this
design, of course, parallels the idea behind cone-adapted 2D shearlets. In [22], the
continuous version of the pyramid-adapted shearlet system was introduced, and it
was shown that the location and the local orientation of the boundary set of cer-
tain three-dimensional solid regions can be precisely identified by this continuous
shearlet transform. The pyramid-adapted shearlet system can easily be generalized
to higher dimensions, but for brevity we only consider the three-dimensional setup
and newly introduce it now in the discrete setting.
4.1 Pyramid-Adapted Shearlet Systems
We will scale according to paraboloidal scaling matrices A2 j, A2 j or A2 j, j ∈ Z, and
encode directionality by the shear matrices Sk, Sk, or Sk, k = (k1,k2) ∈ Z2, defined
by
Compactly Supported Shearlets
A2 j =
2 j 0
0
0 2 j/2 0
0 2 j/2
0
,
A2 j =
2 j/2 0
0
0 2 j 0
0 2 j/2
0
,
and
Sk =
1 k1 k2
0 1 0
0 0 1
,
Sk =
1 0 0
k1 1 k2
0 0 1
,
and
and
17
2 j/2 0
0
A2 j =
0 2 j/2 0
0 2 j
0
,
1 0 0
Sk =
0 1 0
k1 k2 1
,
respectively. The translation lattices will be defined through the following matrices:
Mc = diag(c1,c2,c2), Mc = diag(c2,c1,c2), and Mc = diag(c2,c2,c1), where c1 > 0
and c2 > 0.
Fig. 5 The partition of the fre-
quency domain: The centered
rectangle R. The arrangement
of the six pyramids is indi-
cated by the 'diagonal' lines.
See Figure 6 for a sketch of
the pyramids.
We next partition the frequency domain into the following six pyramids:
{(x 1,x 2,x 3) ∈ R3 : x 1 ≥ 1, x 2/x 1 ≤ 1, x 3/x 1 ≤ 1} : i = 1,
{(x 1,x 2,x 3) ∈ R3 : x 2 ≥ 1, x 1/x 2 ≤ 1, x 3/x 2 ≤ 1} : i = 2,
{(x 1,x 2,x 3) ∈ R3 : x 3 ≥ 1, x 1/x 3 ≤ 1, x 2/x 3 ≤ 1} : i = 3,
{(x 1,x 2,x 3) ∈ R3 : x 1 ≤ −1, x 2/x 1 ≤ 1, x 3/x 1 ≤ 1} : i = 4,
{(x 1,x 2,x 3) ∈ R3 : x 2 ≤ −1, x 1/x 2 ≤ 1, x 3/x 2 ≤ 1} : i = 5,
{(x 1,x 2,x 3) ∈ R3 : x 3 ≤ −1, x 1/x 3 ≤ 1, x 2/x 3 ≤ 1} : i = 6,
Pi =
and a centered rectangle
R = {(x 1,x 2,x 3) ∈ R3 : k(x 1,x 2,x 3)k¥ < 1}.
The partition is illustrated in Figures 5 and 6. This partition of the frequency
space allows us to restrict the range of the shear parameters. In the case of 'shear-
let group' systems one must allow arbitrarily large shear parameters, while the
18
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
'pyramid-adapted' systems restrict the shear parameters to (cid:2)−⌈2 j/2⌉, ⌈2 j/2⌉(cid:3). It is
exactly this fact that gives a more uniform treatment of the directionality properties
of the shearlet system.
(a) Pyramids P1
and P4 and the x 1
axis.
(b) Pyramids P2
and P5 and the x 2
axis.
(c) Pyramids P3
and P6 and the x 3
axis.
Fig. 6 The partition of the frequency domain: The 'top' of the six pyramids.
These considerations are now made precise in the following definition.
Definition 3. For c = (c1,c2) ∈ (R+)2, the pyramid-adapted 3D shearlet system
SH(f ,y , y , y ;c) generated by f ,y , y
, y ∈ L2(R3) is defined by
SH(f ,y , y , y ;c) = F (f ;c1) ∪Y (y ;c) ∪ Y ( y ;c) ∪ Y ( y ;c),
where
F (f ;c1) =(cid:8)f m = f (· − m) : m ∈ c1Z3(cid:9) ,
Y (y ;c) =ny
Y ( y ;c) = { y
j,k,m = 2 jy (SkA2 j · −m) : j ≥ 0, k ≤ ⌈2 j/2⌉,m ∈ McZ3o ,
j,k,m = 2 j y ( Sk A2 j · −m) : j ≥ 0, k ≤ ⌈2 j/2⌉,m ∈ McZ3},
and
Y ( y ;c) = { y
j,k,m = 2 j y ( Sk A2 j · −m) : j ≥ 0, k ≤ ⌈2 j/2⌉,m ∈ McZ3},
where j ∈ N0 and k ∈ Z2. Here we have used the vector notation k ≤ K for k =
(k1,k2) and K > 0 to denote k1 ≤ K and k2 ≤ K.
The construction of pyramid-adapted shearlet systems SH(f ,y , y , y ;c) runs
along the lines of the construction of cone-adapted shearlet systems in L2(R2) de-
scribed in Section 2.3. For a detailed description, we refer to [31].
We remark that the shearlets in spatial domain are of size 2− j/2 times 2− j/2 times
2− j which shows that the shearlet elements will become 'plate-like' as j → ¥
. One
could also use the scaling matrix A2 j = diag (2 j,2 j,2 j/2) with similar changes for
A2 j and A2 j . This would lead to 'needle-like' shearlet elements instead of the 'plate-
like' elements considered in this paper, but we will not pursue this further here, and
simply refer to [31]. More generally, it is possible to even consider non-paraboloidal
Compactly Supported Shearlets
19
scaling matrices of the form A j = diag (2 j,2a
j) for 0 < a ,b ≤ 1. One drawback
of allowing such general scaling matrices is the lack of fast algorithms for non-
dyadic multiscale systems. On the other hand, the parameters a and b allow us to
precisely shape the shearlet elements, ranging from very plate-like to very needle-
like, according to the application at hand, i.e., choosing the shearlet-shape that is the
best 'fit' for the geometric characteristics of the considered data.
j,2b
4.2 Sparse Approximations of 3D Data
We now consider approximations of three-dimensional cartoon-like images using
shearlets introduced in the previous section. The three-dimensional cartoon-like im-
ages E 2n (R3) will be piecewise C2 functions with discontinuities on a closed C2
surface whose principal curvatures are bounded by n . In [31] it was shown that the
optimal approximation rate for such 3D cartoon-like image models f ∈ E 2n (R3)
which can be achieved for almost any representation system (under polynomial
depth search selection procedure of the approximating coefficients) is
k f − fNk2
2 ≤ C · N−1
as N → ¥
,
where fN is the best N-term approximation of f . The following result shows that
compactly supported pyramid-adapted shearlets do (almost) deliver this approxima-
tion rate.
Theorem 5 ([31]). Let c ∈ (R+)2, and let f ,y , y
ported. Suppose that, for all x = (x 1,x 2,x 3) ∈ R3, the function y
} · min{1, x 1−g
} · min{1, x 2−g
(i) y (x ) ≤ C1 · min{1, x 1
x 1(cid:17)−g
x 1(cid:17)−g
(cid:16)1 + x 3
≤ h(x 1) ·(cid:16)1 + x 2
, y ∈ L2(R3) be compactly sup-
} · min{1, x 3−g
where a > 8, g ≥ 4, t 7→ th(t) ∈ L1(R), and C1 a constant, and suppose that y and
y
satisfy analogous conditions with the obvious change of coordinates. Further,
suppose that the shearlet system SH(f ,y , y , y ;c) forms a frame for L2(R3).
Then, for any n > 0, the shearlet frame SH(f ,y , y , y ;c) provides (almost) opti-
mally sparse approximations of functions f ∈ E 2n (R3) in the sense that there exists
some C > 0 such that
y (x )(cid:12)(cid:12)(cid:12)
(ii) (cid:12)(cid:12)(cid:12)
,
i = 2,3,
satisfies:
},
a
¶
¶x
i
k f − fNk2
2 ≤ C · N−1 · (logN)2
as N → ¥
.
(7)
In the following we will give a sketch of the proof of Theorem 5 and, in partic-
ular, give a heuristic argument (inspired by a similar one for 2D curvelets in [5]) to
explain the exponent N−1 in (7).
Proof (Theorem 5, Sketch). Let f ∈ E 2n (R3) be a 3D cartoon-like image. The main
j,k,m(cid:11).
concern is to derive appropriate estimates for the shearlet coefficients (cid:10) f ,y
We first observe that we can assume the scaling index j to be sufficiently large,
20
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
since f as well as all shearlet elements are compactly supported and since a finite
number does not contribute to the asymptotic estimate we are aiming for. In par-
ticular, this implies that we do not need to take frame elements from the 'scaling'
system F (f ;c1) into account. Also, we are allowed to restrict our analysis to shear-
lets y
j,k,m, since the frame elements y
j,k,m can be handled in a similar
way.
Letting q ( f )n denote the nth largest shearlet coefficient (cid:10) f ,y
j,k,m(cid:11) in absolute
value and using the frame property of SH(f ,y , y , y ;c), we conclude that
j,k,m and y
k f − fNk2
2 ≤
1
A
n>N
q ( f )2
n,
for any positive integer N, where A denotes the lower frame bound of the shear-
let frame SH(f ,y , y , y ;c). Thus, for completing the proof, it therefore suffices to
show that
as N → ¥
.
(8)
q ( f )2
n ≤ C · N−1 · (logN)2
n>N
For the following heuristic argument, we need to make some simplifications. We
will assume to have a shearlet of the form y (x) = h (x1)j (x2)j (x3), where h
is a
wavelet and j a bump (or a scaling) function. Note that the wavelet 'points' in the
short direction of the plate-like shearlet. We now consider three cases of coefficients
(cid:10) f ,y
(a) Shearlets y
(b) Shearlets y
(c) Shearlets y
j,k,m whose support does not overlap with the boundary ¶ B.
j,k,m whose support overlaps with ¶ B and is nearly tangent.
j,k,m whose support overlaps with ¶ B, but not tangentially.
j,k,m(cid:11):
(a) Sketch of shearlets whose
support does not overlap
with ¶ B.
Fig. 7 The three types of shearlet y
argument (explaining the approximation rate N−1). Note that only a section of ¶ B is shown.
(b) Sketch of shearlets whose
support overlaps with ¶ B and
is nearly tangent.
j,k,m and boundary ¶ B interactions considered in the heuristic
(c) Sketch of shearlets whose
support overlaps with ¶ B,
but not tangentially.
As we argue in the following, only coefficients from case (b) will be significant.
Case (b) is -- loosely speaking -- the situation in which the wavelet h breaches, in an
almost normal direction, through the discontinuity surface; as is well known from
wavelet theory, 1D wavelets efficiently handle such a 'jump' discontinuity.
(cid:229)
(cid:229)
Compactly Supported Shearlets
21
Case (a). Since f is C2 smooth away from ¶ B, the coefficients (cid:10) f ,y
be sufficiently small owing to the wavelet h
cients of smooth functions).
j,k,m(cid:11) will
Case (b). At scale j > 0, there are at most O(2 j) coefficients, since the plate-like
elements are of size 2− j/2 times 2− j/2 (and 'thickness' 2− j). By assumptions on f
and the support size of y
(and the fast decay of wavelet coeffi-
j,k,m, we obtain the estimate
(cid:10) f ,y
y
j,k,m(cid:11) ≤ k f k¥ (cid:13)(cid:13)
j,k,m(cid:13)(cid:13)1 ≤ C1 (2−2 j)1/2(cid:13)(cid:13)
y
j,k,m(cid:13)(cid:13)
1/2
2 ≤ C2 · 2− j
for some constants C1,C2 > 0. In other words, we have O(2 j) coefficients bounded
by C2 · 2− j. Assuming the case (a) and (c) coefficients are negligible, the nth largest
coefficient q ( f )n is then bounded by
q ( f )n ≤ C · n−1.
Therefore,
n>N
q ( f )2
n ≤ (cid:229)
n>N
C · n−2 ≤ C ·Z
N
x−2dx ≤ C · N−1
and we arrive at (8), but without the log-factor. This in turn shows (7), at least
heuristically, and still without the log-factor.
Case (c). Finally, when the shearlets are sheared away from the tangent position
in case (b), they will again be small. This is due to the vanishing moment conditions
in condition (i) and (ii). ⊓⊔
Clearly, Theorem 5 is an 'obvious' three-dimensional version of Theorem 4.
However, as opposed to the two-dimensional setting, anisotropic structures in three-
dimensional data comprise of two morphologically different types of structure,
namely surfaces and curves. It would therefore be desirable to allow our 3D im-
age class to also contain cartoon-like images with curve singularities. On the other
hand, the pyramid-adapted shearlets introduced in Section 4.1 are plate-like and
thus, a priori, not optimal for capturing such one-dimensional singularities. Surpris-
ingly, these plate-like shearlet systems still deliver the optimal rate N−1 for three-
dimensional cartoon-like images E 2n ,L(R3), where L indicates that we allow our dis-
continuity surface ¶ B to be piecewise C2 smooth; L ∈ N is the maximal number
of C2 pieces and n > 0 is an upper estimate for the principal curvatures on each
piece. In other words, for any n > 0 and L ∈ N, the shearlet frame SH(f ,y , y , y ;c)
provides (almost) optimally sparse approximations of functions f ∈ E 2n ,L(R3) in the
sense that there exists some C > 0 such that
k f − fNk2
2 ≤ C · N−1 · (logN)2
as N → ¥
.
(9)
The conditions on the shearlets y , y , y
are similar to these in Theorem 5, but
more technical, and we refer to [31] for the precise statements and definitions as well
as the proof of the optimal approximation error rate. Here we simply remark that
there exist numerous examples of shearlets y , y
satisfying these conditions,
which lead to (9); one large class of examples are separable generators y , y , y ∈
, and y
(cid:229)
¥
22
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
L2(R3), i.e.,
y (x) = h (x1)j (x2)j (x3),
where h ,j ∈ L2(R) are compactly supported functions satisfying:
(i) h (w ) ≤ C1 · min{1, w
y (x) = j (x1)h (x2)j (x3),
} · min{1, w −g
},
a
y (x) = j (x1)j (x2)h (x3),
¶
for ℓ = 0,1,
for w ∈ R, where a > 8, g ≥ 4, and C1,C2 are constants.
¶w (cid:1)ℓ j (w )(cid:12)(cid:12)(cid:12)
≤ C2 · min{1, w −g
(ii) (cid:12)(cid:12)(cid:12)(cid:0)
}
5 Conclusions
Designing a directional representation system that efficiently handles data with
anisotropic features is quite challenging since it needs to satisfy a long list of desired
properties: it should have a simple mathematical structure, it should provide opti-
mally sparse approximations of certain image classes, it should allow compactly
supported generators, it should be associated with fast decomposition algorithms,
and it should provide a unified treatment of the continuum and digital realm.
In this paper, we argue that shearlets meet all these challenges, and are, therefore,
one of the most satisfying directional systems. To be more precise, let us briefly
review our findings for 2D and 3D data:
• 2D Data. In Section 2, we constructed 2D shearlet systems that efficiently capture
anisotropic features and satisfy all the above requirements.
• 3D Data. In 3D, as opposed to 2D, we face the difficulty that there might exist
two geometrically different anisotropic features; 1D and 2D singularities. The
main difficulty in extending shearlet systems from the 2D to 3D setting lies,
therefore, in introducing a system that is able to represent both these geometri-
cally different structures efficiently. As shown in Section 4, a class of plate-like
shearlets is able to meet these requirements. In other words: the extension from
2D shearlets to 3D shearlets has been successful in terms of preserving the de-
sirable properties, e.g., optimally sparse approximations. It does therefore seem
that an extension to 4D or even higher dimensions is, if not straightforward then,
at the very least, feasible. In particular, the step to 4D now 'only' requires the
efficient handling of yet 'another' type of anisotropic feature.
Acknowledgements The first and third author acknowledge support from DFG Grant SPP-1324,
KU 1446/13. The first author also acknowledges support from DFG Grant KU 1446/14.
Compactly Supported Shearlets
References
23
1. J. P. Antoine, P. Carrette, R. Murenzi, and B. Piette, Image analysis with two-dimensional
continuous wavelet transform, Signal Process. 31 (1993), 241 -- 272.
2. R. H. Bamberger and M. J. T. Smith, A filter bank for the directional decomposition of images:
theory and design, IEEE Trans. Signal Process. 40 (1992), 882 -- 893.
3. E. J. Candés, L. Demanet, D. Donoho, L. Ying, Fast discrete curvelet transforms, Multiscale
Model. Simul. 5 (2006), 861 -- 899.
4. E. J. Candés and D. L. Donoho, Curvelets -- a suprisingly effective nonadaptive representation
for objects with edges, in Curve and Surface Fitting: Saint-Malo 1999, edited by A. Cohen,
C. Rabut, and L. L. Schumaker, Vanderbilt University Press, Nashville, TN, 2000.
5. E. J. Candés and D. L. Donoho, New tight frames of curvelets and optimal representations of
objects with piecewise C2 singularities, Comm. Pure and Appl. Math. 56 (2004), 216 -- 266.
6. S. Dahlke, G. Kutyniok, G. Steidl, and G. Teschke, Shearlet coorbit spaces and associated
Banach frames, Appl. Comput. Harmon. Anal. 27 (2009), 195 -- 214.
7. S. Dahlke, G. Kutyniok, P. Maass, C. Sagiv, H.-G. Stark, and G. Teschke, The uncertainty
principle associated with the continuous shearlet transform, Int. J. Wavelets Multiresolut.
Inf. Process. 6 (2008), 157 -- 181.
8. S. Dahlke, G. Steidl, and G. Teschke, The continuous shearlet transform in arbitrary space
dimensions, J. Fourier Anal. Appl. 16 (2010), 340 -- 364.
9. S. Dahlke and G. Teschke, The continuous shearlet transform in higher dimensions: varia-
tions of a theme, in Group Theory: Classes, Representation and Connections, and Applica-
tions, edited by C. W. Danellis, Math. Res. Develop., Nova Publishers, 2010, 167 -- 175.
10. M. N. Do and M. Vetterli, The contourlet transform: an efficient directional multiresolution
image representation, IEEE Trans. Image Process. 14 (2005), 2091 -- 2106.
11. D. L. Donoho, Sparse components of images and optimal atomic decomposition, Constr. Ap-
prox. 17 (2001), 353 -- 382.
12. D. L. Donoho and G. Kutyniok, Geometric separation using a wavelet-shearlet dictionary,
SampTA'09 (Marseille, France, 2009), Proc., 2009.
13. D. L. Donoho and G. Kutyniok, Microlocal analysis of the geometric separation problem,
preprint.
14. D. L. Donoho, G. Kutyniok, M. Shahram, and X. Zhuang, A rational design of a digital
shearlet transform, preprint.
15. D. L. Donoho, A. Maleki, M. Shahram, V. Stodden, and I. Ur-Rahman, Fifteen years of repro-
ducible research in computational harmonic analysis, Comput. Sci. Engrg. 11 (2009), 8 -- 18.
16. G. Easley, D. Labate, and W.-Q Lim, Sparse directional image representations using the
discrete shearlet transform, Appl. Comput. Harmon. Anal. 25 (2008), 25 -- 46.
17. R. Eslami and H. Radha, A new family of nonredundant transforms using hybrid wavelets
and directional filter banks, IEEE Trans. Image Process. 16 (2007), 1152 -- 1167.
18. P. Grohs, Continuous shearlet frames and resolution of the wavefront set, Monatsh. Math., to
appear.
19. P. Grohs, Continuous shearlet tight frames, J. Fourier Anal. Appl., to appear.
20. K. Guo, G. Kutyniok, and D. Labate, Sparse multidimensional representations using
anisotropic dilation and shear operators, in Wavelets and Splines (Athens, GA, 2005), Nash-
boro Press, Nashville, TN, 2006, 189 -- 201.
21. K. Guo and D. Labate, Optimally sparse multidimensional representation using shearlets,
SIAM J. Math Anal. 39 (2007), 298 -- 318.
22. K. Guo and D. Labate, Analysis and detection of surface discontinuities using the 3D contin-
uous shearlet transform, preprint.
23. K. Guo, D. Labate, W.-Q Lim, G. Weiss, and E. Wilson, Wavelets with composite dilations,
Electron. Res. Announc. Amer. Math. Soc. 10 (2004), 78 -- 87.
24. K. Guo, D. Labate, W.-Q Lim, G. Weiss, and E. Wilson, The theory of wavelets with compos-
ite dilations, Harmonic analysis and applications, Appl. Numer. Harmon. Anal., Birkhäuser
Boston, Boston, MA, 2006, 231 -- 250.
24
Gitta Kutyniok, Jakob Lemvig, and Wang-Q Lim
25. K. Guo, W.-Q Lim, D. Labate, G. Weiss, and E. Wilson, Wavelets with composite dilations
and their MRA properties, Appl. Comput. Harmon. Anal. 20 (2006), 220 -- 236.
26. B. Han, G. Kutyniok, and Z. Shen. A unitary extension principle for shearlet systems, preprint.
27. P. Kittipoom, G. Kutyniok, and W.-Q Lim, Construction of compactly supported shearlet
frames, preprint.
28. P. Kittipoom, G. Kutyniok, and W.-Q Lim, Irregular shearlet frames: Geometry and approx-
imation properties, preprint.
29. G. Kutyniok and D. Labate, Construction of regular and irregular shearlets, J. Wavelet The-
ory and Appl. 1 (2007), 1 -- 10.
30. G. Kutyniok and D. Labate, Resolution of the wavefront set using continuous shearlets, Trans.
Amer. Math. Soc. 361 (2009), 2719 -- 2754.
31. G. Kutyniok, J. Lemvig, and W.-Q Lim, Compactly supported shearlet frames and optimally
sparse approximations of functions in L2(R3) with piecewise C2 singularities, preprint.
32. G. Kutyniok and W.-Q Lim, Compactly supported shearlets are optimally sparse, preprint.
33. G. Kutyniok and W.-Q Lim, Image separation using shearlets, preprint.
34. G. Kutyniok and W.-Q Lim, Shearlets on bounded domains, in Approximation Theory XIII
(San Antonio, TX, 2010), Springer, to appear.
35. G. Kutyniok and T. Sauer, Adaptive directional subdivision schemes and shearlet multireso-
lution analysis, SIAM J. Math. Anal. 41 (2009), 1436 -- 1471.
36. G. Kutyniok, M. Shahram, and D. L. Donoho, Development of a digital shearlet transform
based on pseudo-polar FFT, in Wavelets XIII, edited by V. K. Goyal, M. Papadakis, D. Van
De Ville, SPIE Proc. 7446, SPIE, Bellingham, WA, 2009, 7446-12.
37. D. Labate, W.-Q Lim, G. Kutyniok, and G. Weiss. Sparse multidimensional representation
using shearlets, in Wavelets XI, edited by M. Papadakis, A. F. Laine, and M. A. Unser, SPIE
Proc. 5914, SPIE, Bellingham, WA, 2005, 254 -- 262,
38. W.-Q Lim, The discrete shearlet transform: A new directional transform and compactly sup-
ported shearlet frames, IEEE Trans. Image Process. 19 (2010), 1166 -- 1180.
39. E. L. Pennec and S. Mallat, Sparse geometric image representations with bandelets, IEEE
Trans. Image Process. 14 (2005), 423 -- 438.
40. E. P. Simoncelli, W. T. Freeman, E. H. Adelson, D. J. Heeger, Shiftable multiscale transforms,
IEEE Trans. Inform. Theory 38 (1992), 587 -- 607.
|
1704.04392 | 1 | 1704 | 2017-04-14T11:02:44 | A note on triangular operators on Smooth Sequence Spaces | [
"math.FA"
] | For a scalar sequence {(\theta_n)}_{n \in \mathbb{N}}, let C be the matrix defined by c_n^k = \theta_{n-k+1} if n > k, c_n^k = 0 if n < k. The map between K\"{o}the spaces \lambda(A) and \lambda(B) is called a Cauchy Product map if it is determined by the triangular matrix C. In this note we introduced some necessary and sufficient conditions for a Cauchy Product map on a nuclear K\"{o}the space \lambda(A) to nuclear G_1-space \lambda(B) to be linear and continuous. Its transpose is also considered. | math.FA | math |
A NOTE ON TRIANGULAR OPERATORS ON
SMOOTH SEQUENCE SPACES
ELIF UYANIK AND MURAT H. YURDAKUL
n = θn−k+1 if n ≥ k, ck
Abstract. For a scalar sequence (θn)n∈N, let C be the matrix
defined by ck
n = 0 if n < k. The map
between Kothe spaces λ(A) and λ(B) is called a Cauchy Product
map if it is determined by the triangular matrix C. In this note we
introduced some necessary and sufficient conditions for a Cauchy
Product map on a nuclear Kothe space λ(A) to nuclear G1-space
λ(B) to be linear and continuous. Its transpose is also considered.
Dedicated to the memory of Prof. Dr. Tosun Terzioglu
1. Introduction
We refer the reader to [3], [4] and [5] for the terminology used but
n)n,k∈N be a matrix of real numbers such
ak
n > 0. The ℓ1 - Kothe space
not defined here. Let A = (ak
that 0 ≤ ak
for all n, k and sup
n ≤ ak+1
n
k
λ(A) defined by the matrix A is the space of all sequences of scalars
x = (xn) such that
kxkk =Xn
xnak
n < ∞,
∀k ∈ N
With the topology generated by the system of seminorms {k.kk , k ∈ N},
it is a Fr´echet space.
The topological dual of λ(A) is isomorphic to the space of all se-
quences u for which un ≤ Cak
n for some k and C > 0.
It is well known that a Kothe space λ(A) associated with the matrix
A is nuclear if and only if for each k there exists m such that
ak
n
am
n
Xn
< +∞
2010 Mathematics Subject Classification. 47B37, 46A45.
Key words and phrases. Kothe spaces, Smooth Sequence Spaces, Cauchy
Product.
This research was partially supported by the Turkish Scientific and Technological
Research Council.
1
2
ELIF UYANIK AND MURAT H. YURDAKUL
and in this case the fundamental system of norms kxkk = Xn
can be replaced by the equivalent system of norms
xnak
n
kxkk = sup
n
xnak
n,
k ∈ N.
The infinite and finite type power series spaces are well known exam-
ples of Kothe spaces given by the matrices (ekαn) respectively (e−
k )
where (αn) is a monotonically increasing sequence going to infinity.
The space A(C) of all entire functions on C and the space A(D) of all
holomorphic functions on the unit disc can be represented as an infinite
respectively finite type power series spaces.
αn
Smooth sequence spaces were introduced in [6] as a generalization
of power series spaces. A Kothe set A = (cid:8)(ak
and the corresponding Kothe space λ(A) a G∞-space if A satisfies the
followings :
n)(cid:9) is called a G∞-set
n = 1, ak
n ≤ ak
(1) a1
(2) ∀k ∃j with (ak
n+1 for each k and n;
n)2 = O(aj
n)
A Kothe set B = (cid:8)(bk
n)(cid:9) is called a G1-set and the corresponding
Kothe space λ(B) a G1-space if B satisfies the followings :
(1) 0 < bk
n+1 ≤ bk
(2) ∀k ∃j with bk
We need the following result [1].
n < 1 for each k and n;
n = O((bj
n)2)
Lemma 1.1. Let λ(A) and λ(B) be Kothe spaces. A map T : λ(A) −→
λ(B) is continuous linear map if and only if for each k there exists m
such that
sup
n
kT enkk
kenkm
< +∞
If (an), (bn) are two sequences of scalars, then the Cauchy product
n
(cn) = (an) ∗ (bn) of (an) and (bn) is defined by cn =
an+1−kbk.
Now let θ = (θn) be a fixed sequence of scalars and let λ(A), λ(B) be
two nuclear ℓ1-Kothe spaces. We define the Cauchy Product mapping
Tθ from λ(A) into λ(B) by Tθx = θ ∗ x,
x = (xn) ∈ λ(A). So, Tθ :
λ(A) −→ λ(B) can be determined by the lower triangular matrix
Xk=1
0
0
θ1
0
θ2 θ1
θ3 θ2 θ1
...
C =
0
0
0
. . .
· · ·
· · ·
· · ·
CAUCHY PRODUCT OPERATOR
3
2. Cauchy Product Map on Kothe spaces
In this section we introduce some necessary and sufficient conditions
for the map Tθ to be linear and continuous.
Theorem 2.1. Let λ(A) be a nuclear Kothe space, λ(B) be a nuclear
G1-space. Then the Cauchy product map Tθ : λ(A) −→ λ(B) is linear
continuous operator if and only if the following hold:
i) θ ∈ λ(B)
ii) λ(A) ⊂ λ(B)
Proof. Let Tθ : λ(A) −→ λ(B) be a continuous linear operator.
θj−n+1bk
Note that kTθenkk = k(0, 0, ..., 0, θ1, θ2, · · · )kk = sup
j
j≥n
n ∈ N. Clearly kenkm = am
that
n . So, by Lemma 1.1 ∀k, ∃m, ∃ρ > 0 such
for
θj−n+1bk
j ≤ ρam
n ,
∀n ∈ N.
sup
j≥n
Choose j = n. Then ∀k, ∃m, ∃C > 0 such that
bk
n ≤ Cam
n
i.e, λ(A) ⊂ λ(B). Since Tθe1 ∈ λ(B), it follows that θ ∈ λ(B).
Conversely, since B is a G1-set and by ii and i we have for a given
k, there are m1(k) and m2(m1) such that
kTθenkk = sup
j≥n
θj−n+1bk
j
θj−n+1(bm1
j )2
(θj−n+1bm1
j )(bm1
n )
(θj−n+1bm1
j )(am2
n )
(θj−n+1bm1
j−n+1)(am2
n )
≤ C1 sup
j≥n
≤ C1 sup
j≥n
≤ C2 sup
j≥n
≤ C2 sup
j≥n
≤ Cam2
n .
Therefore, ∀k, ∃m2 such that
sup
n
kTθenkk
kenkm2
< ∞
that is, Tθ is continuous.
(cid:3)
(cid:3)
We consider the map Tθ
′ : λ(A) −→ λ(B) which is determined by the
matrix C t (the transpose of C) and try to find necessary and sufficient
conditions for the continuity of Tθ
′.
4
ELIF UYANIK AND MURAT H. YURDAKUL
Theorem 2.2. Let λ(A) be a nuclear G∞-space, λ(B) be a nuclear
′ : λ(A) −→ λ(B) which is given above is linear
Kothe space. Then, Tθ
continuous operator if and only if the following hold:
i) θ ∈ λ(A)′
ii) λ(A) ⊂ λ(B)
Proof. The matrix C t of the operator Tθ
lowing upper triangular matrix:
′ : λ(A) −→ λ(B) is the fol-
θ1 θ2 θ3
0
θ1 θ2
0
0
θ1
...
θ4
θ3
θ2
. . .
C t =
· · ·
· · ·
· · ·
Let Tθ
Note that kTθ
′ : λ(A) −→ λ(B) be a continuous linear operator.
′enkk = k(θn, θn−1, · · · , θ1, 0, 0, · · · )kk = sup
θn+1−ibk
i
1≤i≤n
for n ∈ N. So, by Lemma 1.1 ∀k, ∃m, ∃µ > 0 such that
sup
1≤i≤n
θn+1−ibk
i ≤ µam
n ,
∀n ∈ N.
Let i = 1. Hence ∃m, ∃C = µ
bk
1
> 0 such that
θn ≤ Cam
n ,
∀n
i.e, θ ∈ λ(A)′.
Let i = n. Then ∀k, ∃m such that
i.e,
bk
n ≤
µ
θ1
am
n
λ(A) ⊂ λ(B)
On the other hand, since A is a G∞-set and by (i) and (ii) for a given
k, there are m1 and m2(k) and m = max {m1, m2} such that
kTθ
′enkk = sup
1≤i≤n
θn−i+1bk
i
≤ C1 sup
1≤i≤n
≤ C1 sup
1≤i≤n
am1
n−i+1bk
i
am1
n−i+1am2
i
≤ C1am1
≤ C2(am
n am2
n )2.
n
CAUCHY PRODUCT OPERATOR
5
Since λ(A) is G∞ - space, for this m, ∃j such that
n )2
(am
aj
n
sup
n
< ∞
Therefore, ∀k, ∃j such that
sup
n
kTθ
′enkk
kenkj
< ∞
that is, Tθ
′ is continuous.
(cid:3)
(cid:3)
It is known that S is a normal sequence space if whenever xi < yi
and y = (yi) ∈ S, then x = (xi) ∈ S [2].
Remark 2.1. Now we write θ ∈ S when the Cauchy product map
Tθ : λ(A) −→ λ(B) above is continuous. If θ, η ∈ S, λ ∈ K, then
clearly Tθ+η and Tλθ will be continuous since Tθ and Tη are continuous.
Hence S is a vector space.
Now, let θi < ηi , ∀i, η ∈ S. Since Tη is continuous, for all k we
find m so that
sup
n (sup
j≥n
θj−n+1
bk
j
am
n ) ≤ sup
n (sup
j≥n
ηj−n+1
bk
j
am
n ) < ∞
i.e. Tθ is continuous.
Therefore θ ∈ S. Hence we obtain that S is a normal sequence space.
References
[1] L. Crone and W. Robinson, Diagonal maps and diameters in Kothe spaces,
Israel J. Math. 20, (1975), 13 -- 22.
[2] G. Kothe, Topological Vector Spaces 1, Springer-Verlag, Berlin, Heidelberg,
New York, 1969.
[3] R. Meise and D. Vogt, Introduction to Functional Analysis, Clarendon
Press, Oxford, 1997.
[4] A. Pietsch, Nuclear Locally Convex Spaces, Springer-Verlag, Berlin-New
York, 1972.
[5] M. S. Ramanujan and T. Terzioglu, Subspaces of smooth sequence spaces,
Studia Math. 65, (1979), 299 -- 312.
[6] T. Terzioglu, Die diametrale Dimension von lokalkonvexen Raumen, Col-
lect. Math. 20, (1969), 49 -- 99.
Elif Uyanık, Department of Mathematics, Middle East Technical
University, 06800 Ankara Turkey
E-mail address: [email protected]
Murat H. Yurdakul, Department of Mathematics, Middle East Tech-
nical University, 06800 Ankara Turkey
E-mail address: [email protected]
|
0909.1867 | 2 | 0909 | 2010-03-30T22:24:07 | Characterising derivations from the disc algebra to its dual | [
"math.FA",
"math.CV"
] | We show that the space of all bounded derivations from the disc algebra into its dual can be identified with the Hardy space $H^1$; using this, we infer that all such derivations are compact. Also, given a fixed derivation $D$, we construct a finite, positive Borel measure $\mu_D$ on the closed disc, such that $D$ factors through $L^2(\mu_D)$. Such a measure is known to exist, for any bounded linear map from the disc algebra to its dual, by results of Bourgain and Pietsch, but these results are highly non-constructive. | math.FA | math | CHARACTERISING DERIVATIONS FROM THE DISC ALGEBRA TO ITS
DUAL
Y. CHOI AND M. J. HEATH
Abstract. We show that the space of all bounded derivations from the disc algebra into its dual
can be identified with the Hardy space H 1; using this, we infer that all such derivations are compact.
Also, given a fixed derivation D, we construct a finite, positive Borel measure µD on the closed
disc, such that D factors through L2(µD). Such a measure is known to exist, for any bounded
linear map from the disc algebra to its dual, by results of Bourgain and Pietsch, but these results
are highly non-constructive.
0
1
0
2
r
a
M
0
3
]
.
A
F
h
t
a
m
[
2
v
7
6
8
1
.
9
0
9
0
:
v
i
X
r
a
1. Introduction
Although derivations from Banach algebras to particular coefficient modules have been much
studied, interest has usually focused on the existence or otherwise of non-trivial bounded deriva-
tions, rather than their characterisation. Even in the special case of derivations from a commutative
Banach algebra to its dual (as in [2] or [5] for example), there are comparatively few examples where
the space of such derivations is explicitly characterised or parametrised.
For uniform algebras, very little is known: the only examples with a complete characterisation
are the trivial uniform algebras C(X) -- and these have no nonzero bounded derivations into any
dual C(X)-bimodule. In this paper we provide the first example of a non-trivial uniform algebra --
namely, the disc algebra -- where the space of derivations to its dual can be completely characterised.
For the disc algebra, this is equivalent to characterising those complex Borel measures µ on the
unit circle, for which the bilinear functional (f, g) 7→RT f ′g dµ on the space of analytic polynomials
Indeed, our proof identifies this space, in a natural way, with a rather well-known space of
functions on the unit circle. Some of our techniques should be applicable to other uniform algebras,
but for sake of brevity we do not pursue this here.
is bounded in the supremum norm on the unit disc.
As a by-product of our characterisation, we obtain a short proof that every bounded derivation
from the disc algebra to its dual is automatically compact, resolving a question raised in [4]. Finally,
using some ideas from the proof of our characterisation, we show how one can construct a 'Pietsch
control measure' for a given derivation, which witnesses the (known) fact that each such derivation
is 2-summing. Our construction is explicit and does not rely on the previous results of Bourgain
on arbitrary linear maps from the disc algebra to its dual.
Notation and other preliminaries. Throughout, 'derivation' means 'bounded derivation'. Given
a Banach algebra A, we denote by D(A) the space of bounded derivations from A to is dual space A∗.
2010 Mathematics Subject Classification. Primary 46J15; Secondary 30H10, 47B47.
Key words and phrases. Derivation, disc algebra, Hardy space .
Second author supported by post-doctoral grant SFRH/BPD/40762/2007 from FCT (Portugal).
1
2
Y. CHOI AND M. J. HEATH
We briefly review the (mostly standard) notation used in describing the disc algebra and related
spaces.
• ∆ denotes the closed unit disc in the complex plane C; D denotes its interior, and T its
boundary. We write A(D) for the disc algebra, that is, the space of all functions which
are continuous on ∆ and analytic on D. Via the maximum modulus principle this can be
regarded as a closed subalgebra of C(T), and we shall do this without further comment.
• dz denotes arc length measure on T, normalised so that the length of T is 2π. For 1 ≤ p <
∞, the norm on Lp(T) is denoted by k · kp, and normalised so that k1kp = (2π)1/p.
• For 1 ≤ p ≤ ∞, we denote by H p the familiar Hardy space on D. We will identify (via radial
limits) elements of H p with those functions in Lp(T) whose negative Fourier coefficients are
zero. It is also convenient to write H p
0 for the subspace
H p
0 :=(cid:26)f ∈ H p : f (0) =ZT
f (z)dz = 0(cid:27) .
Finally, if X is a set of complex-valued functions, we denote by Conj X the set(cid:8)f : f ∈ X(cid:9).
2. The characterisation
Our characterisation is as follows:
Theorem 2.1. Let D ∈ D(A(D)). Then there is a unique hD ∈ Conj H 1
for all f ∈ A(D).
f (z)hD(z)dz
D(f )(1) =ZT
0 such that
Furthermore, E : D 7→ hD defines a bounded, linear isomorphism from D(A(D)) onto Conj H 1
0 .
As we shall see, the tools required to prove this result already exist in the literature. It is their
combination which appears to be new.
2.1. Everything except surjectivity of E. In this subsection we prove most of Theorem 2.1;
more precisely we prove the following.
Proposition 2.2. Let D ∈ D(A(D)). Then there is a unique hD ∈ Conj H 1
for all f ∈ A(D).
f (z)h(z)dz
0 such that
Dh(f )(1) =ZT
Furthermore E : D 7→ hD defines an injective, bounded, linear map from D(A(D)) to Conj H 1
0 .
To prove this we shall require some lemmas, and some auxiliary notation. Define B : D(A(D)) →
A(D)∗ by
B(D)(f ) := D(f )(1)
(f ∈ A(D)).
Lemma 2.3. B is a continuous, injective linear map.
Proof. It is clear that B is linear and contractive. Suppose D ∈ D(A(D)) and that B(D) = 0.
Given analytic polynomials f and g, let u be the polynomial with constant term zero and whose
derivative equals f ′g. Then
D(f )(g) = D(Z)(f ′g) = D(u)(1) = B(D)(u) = 0 ;
and so by continuity D = 0. Thus B is injective.
(cid:3)
CHARACTERISING DERIVATIONS FROM THE DISC ALGEBRA TO ITS DUAL
3
It follows from work of Morris in [4] that B factors through the natural inclusion of Conj H 1
0 into
A(D)∗. The main ingredient is the following proposition.
Proposition 2.4. Let A be a unital uniform algebra on a compact space X. Let M be a symmetric,
contractive Banach A-bimodule, and let D : A → M be a bounded derivation.
(i) If f ∈ A ∩ Conj A then D(f ) = 0.
(ii) If h ∈ A then kD(h)k ≤ 2ekDkk Re hk∞.
(iii) If f, g ∈ A then kD(f )k ≤ 2ekDkkf − gk∞.
As [4] is not easily available, and since the proof of Proposition 2.4 is short and instructive, we
give the argument below, paraphrased slightly from the original wording.
Proof. Let a ∈ A. Since M is symmetric, an easy induction using the Leibniz identity shows that
D(an) = nan−1 · D(a) for all n ≥ 1; on taking a = 1A and n = 2, 3 we have
D(1A) = 2 · 1A · D(1A) = 3 · 12
A · D(1A) = 3 · 1A · D(1A),
It follows, by considering the power series expansion for exp, that
which forces D(1A) = 0.
D(ea) = ea · D(a), and therefore
(2.1)
for all a ∈ A,
(Note that this does not use the fact that A is a uniform algebra.)
e−aD(ea) = D(a)
Now let f ∈ A ∩ Conj A. For every t > 0 we have keitfk∞ ≤ 1 and ke−itfk∞ ≤ 1; so by (2.1) we
have kD(tf )k ≤ kDk. Since t is arbitrary, D(f ) = 0, and (i) is proved.
Now suppose h ∈ A. If Re h = 0 then ih ∈ A∩ Conj A, and hence by part (i) iD(h) = D(ih) = 0.
We may thus assume without loss of generality that k Re hk∞ > 0. By (2.1), for every a ∈ A we
have
kD(a)k ≤ kDkkeak∞kDkke−ak∞ = kDkkeRe ak∞ke− Re ak∞ ≤ kDk exp(2k Re ak∞) .
Putting a = (2k Re hk∞)−1h we get (2k Re hk∞)−1D(h) ≤ e, and (ii) follows.
Finally, given f, g ∈ A, note that
Re(f − g) = Re(f − g)
,
Im(f − g) = Re(−i(f + g)) .
Applying (ii) with h = f − g and then with h = −i(f + g) gives
kD(f ) − D(g)k ≤ 2ekDkk Re(f − g)yk∞ ≤ 2ekDkkf − gk∞ ,
kD(f ) + D(g)k ≤ 2ekDkk Im(f − g)k∞ ≤ 2ekDkkf − gk∞ .
Combining these last two inequalities and using the triangle inequality, we obtain (iii).
Proof of Proposition 2.2. Take A = A(D) and M = A(D)∗ in Proposition 2.4. Then for each D ∈
D(A(D)), the functional B(D) ∈ A(D)∗ extends to a continuous linear functional on C(T)/ Conj A(D),
with norm at most 2ekDk. By the F. & M. Riesz Theorem, (C(T)/ Conj A(D))∗ consists precisely
of functionals of the form
(cid:3)
ψ : f + Conj A(D) 7→ZT
f hψ dz,
where hψ ∈ Conj H 1
E(D) ∈ Conj H 1
0 (necessarily unique) and kψk = khψkL1. Set E(D) := hB(D). We then have
0 , and kE(D)kL1 ≤ 2ekDk. That E is linear and injective follows from Lemma 2.3.
(cid:3)
4
Y. CHOI AND M. J. HEATH
2.2. Surjectivity. To go further, we shall use some standard (albeit non-trivial) results from the
theory of BMOA. Recall that a function f ∈ H 2 lies in the space BMOA if and only if it satisfies
one of the following three equivalent conditions:
(i) there exists a constant C1 such that
1
IZI f − fIdz ≤ C1
for every arc I ⊂ T,
where fI denotes the average of f over I.
(iii) there exists a constant C3 such that
(ii) there exists a constant C2 such that RT f hdz ≤ C2khk1 for all h ∈ H 1
for all k ∈ H 2,
ZD f ′(z)k(z)2(1 − z)2 dA(z) ≤ C 2
3ZT k2 dz
0 .
where dA is the usual area measure on D.
By taking the least possible constants C1, C2 and C3 which satisfy the above, one obtains three
mutually equivalent seminorms k · k(1), k · k(2) and k · k(3) on the space BMOA. (The kernel of each
of these seminorms is the space of constant functions.)
Remark 2.5. Strictly speaking, (i) is the true definition. The equivalence of (i) and (ii) is Feffer-
man's celebrated duality theorem, while the equivalence of (i) and (iii) is essentially the Carleson
embedding theorem in a special case. Both results are well-known to specialists, and proofs can
be found in [3], although assembling them from scattered parts takes some work on the reader's
part. (The duality theorem is covered in [3, Ch. VI, §4] -- see the discussion at the top of p. 240
for how one recovers the complex form of Fefferman's theorem from the real form -- while a proof
of the Carleson multiplier theorem can be extracted from a combination of [3, Theorem I.5.6] and
[3, Theorem VI.3.4].)
We can now give the proof of Theorem 2.1. The remainder of the argument is essentially
contained, phrased differently, in results from [1] (see in particular Theorem 3.4 of that article).
However, the required facts about BMOA predate [1], and it is easier to use them directly, rather
than try to deduce our theorem from that paper's results.
Proof of Theorem 2.1. Having proven Proposition 2.2, it only remains to show that E is surjective.
Let h ∈ H 1
0 . We claim that there is a constant K > 0 such that, for any f and g which are
holomorphic on a neighbourhood of ∆, we have
(2.2)
u(z)h(z) dz ≤ Kkfk∞kgk∞
where u is defined by u(0) = 0 and u′ = f ′g.
If this is the case, then we can extend the map
(f, g) 7→RT u(z)h(z) dz to obtain a bounded bilinear form on A(D), which we denote by D. It can
then be easily verified that D ∈ D(A(D)), and that E(D) = h as required.
It remains only to prove the inequality (2.2). Thus, let f and g be holomorphic on a neighbour-
hood of ∆. It is immediate from the definition of k · k(3) that kuk(3) ≤ kfk(3)kgk∞, and since the
norms k · k(3) and k · k(2) are equivalent, there exists a constant K1 -- independent of f , g and h --
such that
kuk(2) ≤ K1kfk(2)kgk∞ .
The desired inequality (2.2) now follows, since kfk(2) ≤ kfk∞.
(cid:3)
ZT
CHARACTERISING DERIVATIONS FROM THE DISC ALGEBRA TO ITS DUAL
5
Remark 2.6. It is possible to prove (2.2) more directly, using arguments from the proof that the
seminorms k · k(2) and k · k(3) are equivalent; see Remark 4.2 below.
3. Compactness
It is shown in [4] that every bounded derivation from the disc algebra to its dual is weakly
compact.
(In fact, by results of Bourgain, every bounded linear map from A(D) to its dual is
2-summing, and hence weakly compact, but the proof for general linear maps is significantly harder
than the special case of derivations. For relevant details, see [6, §III.I] for instance.)
We show that rather more is true: each derivation can be approximated in the operator norm
by finite rank derivations and so, in particular, is compact. This answers a question raised in [4].
The key observation is the following.
Lemma 3.1. Let h be an analytic polynomial with zero constant term. Then E−1(h) has finite
rank.
Proof. It suffices by linearity to consider the case where h(z) = zn for some n ≥ 1. Let D := E−1(h);
then
D(f )(g) =ZT
f ′(z)g(z)z−n dz ,
showing that D(f ) = 0 for every f in the finite-codimension subspace zn+2A(D).
(cid:3)
Since the polynomials are dense in H 1, Theorem 2.1 and Lemma 3.1 give the following corollary.
Corollary 3.2. Each D ∈ D(A(D)) is approximable and hence compact.
Remark 3.3. Given D ∈ D(A(D)), we can construct an explicit approximating sequence of finite-
rank derivations -- just use one's favourite method of approximating H 1 functions by analytic
polynomials, such as convolution with the F´ejer kernel for example. This may be useful for studying
other operator-theoretic properties of D; for instance, its singular values can be estimated by using
known approximation results for H 1-functions.
As remarked in the previous section, we know that every D ∈ D(A(D)) is 2-summing. Hence,
by Pietsch's factorization theorem, there exists a finite, positive Borel measure µD on ∆ such that
4. 2-summing norms
(⋆)
kD(f )k ≤ kfkL2(µ) =(cid:18)Z∆ f (z)2 dµD(z)(cid:19)1/2
for all f ∈ A(D).
However, these abstract arguments give us no way to find an 'explicit' measure µD. In this section
we shall show how this can be done, using ideas extracted from existing proofs that the seminorms
k · k(2) and k · k(3) are equivalent. See [3, Ch. 3 and 4].
0 we define a bilinear functional Dh on the space
of polynomials as follows: given polynomials f and g, let u be the unique polynomial satisfying
u(0) = 0 and f ′g = u′; then set
As in the proof of Theorem 2.1, for each h ∈ H 1
Dh(f )(g) :=ZT
u(z)h(z) dz .
6
Y. CHOI AND M. J. HEATH
By Proposition 2.2, every D ∈ D(A(D)) has the form Dh for some h ∈ H 1
0 . We may write
h(z) = αz + z2F (z) where F ∈ H 1 and α ∈ C. Then, since F + kFk1 and F − kFk1 are H 1-
functions vanishing nowhere on D, they have square roots in H 2, and we can therefore write
z2F (z) ="z(cid:18) F + kFk1
2
(cid:19)1/2#2
+"z(cid:18) F − kFk1
2
(cid:19)1/2#2
where k1, k2 ∈ H 2
(4.1)
0 . Thus
Dh = αDZ + Dk2
1
+ Dk2
2
,
and we shall estimate these three terms separately.
= k2
1 + k2
2 , say,
Estimating kDZk. Since
we find that
DZ (f )(g) =Z 2π
kDZ (f )k = 2πbf (1) ≤(cid:18)2πZT f (z)2 dz(cid:19)1/2
u(eiθ)e−iθ dθ = 2πf ′(0)g(0) = 2πbf (1)g(0),
= √2πkfkL2(λ) ,
(♦)
where λ denotes arc length measure on T, regarded as a Borel measure on ∆.
0
Estimating kDk2(f )k. Fix k ∈ H 2
on ∆, such that
0 . We shall construct finite, positive, Borel measures µ1 and µ2
(♠)
kDk2(f )k ≤ kfkL2(µ1) + kfkL2(µ2)
for all polynomials f .
The key is a well-known identity of Littlewood-Paley type. To state it, and for later convenience,
we introduce some notation. Let Λ denote the measure on D\{0} defined by dΛ(z) = 4log 1
z dA(z).
We denote the corresponding inner product on C(D) by h·,·iΛ and the induced Hilbert-space norm
by k · kΛ.
Lemma 4.1. Let u, v ∈ H 2. Then
In particular,
(4.2)
(u(ζ) − u(0))(v(ζ) − v(0)) dζ
(cid:10)u′, v′(cid:11)Λ =ZT
ku′kΛ =ZT (u(ζ) − u(0)2 dζ ≤ kuk2.
The lemma can be proved either by using Fourier series and Parseval's identity, or using Green's
formula; see [3, Lemma 3.1] for a proof using the latter approach.
Proof of the inequality (♠). Applying Lemma 4.1 yields
(4.3a)
Dk2(f )(g) =ZT
u(ζ)h(ζ) dζ =(cid:10)u′, (k2)′(cid:11)Λ =(cid:10)f ′g, 2k′k(cid:11)Λ .
CHARACTERISING DERIVATIONS FROM THE DISC ALGEBRA TO ITS DUAL
7
Then
(4.3b)
(cid:12)(cid:12)(cid:10)f ′g, 2kk′(cid:11)Λ(cid:12)(cid:12) ≤ 2ZD (f ′gk′k)(z) dΛ(z)
≤ 2kgk∞(cid:10)f ′k,k′(cid:11)Λ
≤ 2kgk∞kk′kΛkf ′kkΛ
= 2kgk∞kkk2kf ′kkΛ
(Cauchy-Schwarz)
(Equation (4.2))
We now consider kf ′kkΛ. Since f ′k = (f k)′ − f k′, we have
(4.3c)
kf ′kkΛ ≤ k(f k)′kΛ + kf k′kΛ = kf kk2 + kf k′kΛ = kfkL2(µB ) + kfkL2(µI )
where µB and µI are the positive Borel measures on D given by
µB(X) :=ZX∩T k(z)2 dz
; µI (X) :=ZX∩D k′(z)2 dΛ(z) .
Clearly µB is finite, since k ∈ H 2, and µI is finite since
µI (∆) =ZD k′(z)2 dΛ(z) = kk′k2
Dk2(f )(g) ≤ 4kgk∞kkk2(cid:0)kfkL2(µB ) + kfkL2(µI )(cid:1) ,
Putting the inequalities (4.3a), (4.3b) and (4.3c) together, we have
which on rescaling gives us the desired inequality.
(4.4)
Λ = kkk2
2 < ∞ .
(cid:3)
Finally, using (4.1), and combining the estimates (♦) and (♠), we obtain finite, positive Borel
measures µ1, . . . , µ5 such that
kDh(f )k ≤
kfkL2(µj )
for all polynomials f .
To get a single L2-upper bound, observe that there is a crude estimate
(4.5)
kfkL2(µj ) ≤
√5kfkL2(µ1+···+µ5) ,
5Xj=1
5Xj=1
which is a straightforward consequence of the elementary inequality
mXj=1
2
aj
=
mXj=1
mXk=1
ajak ≤
1
2
mXj=1
mXk=1
(a2
j + a2
k) ≤ m
a2
j
for all a1, . . . , am ≥ 0.
mXj=1
Thus in (⋆), we may take µD = √5(µ1 + ··· + µ5). Although we have not written out an explicit
formula for µD, it is clear from the calculations above that it could be easily done for a given
h ∈ H 1
0 .
Remark 4.2. The inequality (4.4) implies, in particular, that Dk2 ∈ D(A(D)) with kDk2k ≤ 8kkk2
2.
One can use this to obtain a direct proof of the earlier inequality (2.2), which does not rely on first
knowing the equivalence of the seminorms k · k(2) and k · k(3). Thus Theorem 2.1 could have been
proved without explicit mention of BMOA, although the argument would then have been longer
and more opaque.
8
Y. CHOI AND M. J. HEATH
References
6 (1996), 263 -- 275.
[1] A. B. Aleksandrov and V. V. Peller, Hankel operators and similarity to a contraction, Internat. Math. Res. Notices
[MR 97c:47006b]
[2] W. G. Bade, P. C. Curtis, Jr., and H. G. Dales, Amenability and weak amenability for Beurling and Lipschitz
[MR 88f:46098]
[3] J. Garnett, Bounded analytic functions, revised first ed. Graduate Texts in Mathematics, vol. 236, Springer, New
[MR 2007e:30049]
algebras, Proc. London Math. Soc. (3) 55 (1987), no. 2, 359 -- 377.
York, 2007.
[4] S. E. Morris, Bounded derivations from uniform algebras, Ph.D. thesis, University of Cambridge, 1993.
[5] V. Runde, A functorial approach to weak amenability for commutative Banach algebras, Glasgow Math. J. 34
[MR 93d:46084]
[6] P. Wojtaszczyk, Banach Spaces for Analysts, Cambridge Studies in Advanced Mathematics, vol. 25, Cambridge
[MR 93d:46001]
University Pess, Cambridge, 1991.
(1992), no. 2, 241 -- 251.
D´epartement de math´ematiques et de statistique, Pavillon Alexandre-Vachon, Universit´e Laval,
Qu´ebec, QC, Canada, G1V 0A6
E-mail address: [email protected]
Departamento de Matem´atica, Instituto Superior T´ecnico, Av. Rovisco Pais, 1049-001 Lisboa, Por-
tugal
E-mail address: [email protected]
|
1111.1510 | 1 | 1111 | 2011-11-07T09:08:28 | The Gyro-Kinetic Approximation. An attempt at explaining the method based on Darboux Algorithm and Lie Transform | [
"math.FA",
"math-ph",
"math-ph",
"math.SG",
"physics.plasm-ph"
] | This Proceeding presents the method that allows us to get the Gyro-Kinetic Approximation of the Dynamical System satisfied by the trajectory of a particle submitted to a Strong Magnetic Field. The goal of the method is to build a change of coordinates in order to make the dynamic of two components of the trajectory to disappear. This change of coordinates is based on a Darboux mathematical Algorithm and on a Lie Transform | math.FA | math |
The Gyro-Kinetic Approximation
An attempt at explaining the method based on Darboux Algorithm and
Lie Transform
Emmanuel Frénod∗
Mathieu Lutz†
Summer school Fusion - Paris - September 2011
Abstract
This Proceeding presents the method that allows us to get the Gyro-Kinetic Approximation of the
Dynamical System satisfied by the trajectory of a particle submitted to a Strong Magnetic Field. The
goal of the method is to build a change of coordinates in order to make the dynamic of two components of
the trajectory to disappear. This change of coordinates is based on a Darboux mathematical Algorithm
and on a Lie Transform
Introduction
1
Scientific framework - At the end of the 70', Littlejohn [11, 12, 13] shed new
light on what is called the Guiding Center Approximation. His approach was based
on mathematical theories - Hamiltonian Mechanics, Differential Geometry, Symplectic
Geometry - in order to clarify what has been done for years in the domain (see Kruskal
[10], Gardner [4], Northrop [14], Northrop & Rome [15]). His papers claim that it has
been for him an enormous effort to reach this goal, since he had to incorporate into a
physical affordable theory high level mathematics. Sure, this theory is a nice success.
It has been being widely used by physicists to deduce related models (Finite Larmor
Radius Approximation, Drift Kinetic Model, Quasi-Neutral Gyro-kinetic Model, etc.,
see for instance Koseleff [9], Brizard [1], Dubin et al. [2], Frieman & Chen [3], Hahm
[7], Hahm, Lee & Brizard [8], Parra & Catto [16, 17, 18]) making up the Gyro-Kinetic
Approximation Theory, which is the basis of all kinetic codes used to simulate Plasma
Turbulence emergence and evolution in Tokamak (see for instance Grandgirard et al.
[5, 6]).
The incorporation of mathematical concepts into a physical theory has been done so
nicely that the resulting Gyro-Kinetic Approximation Theory is now very difficult for
mathematicians.
∗Université Européenne de Bretagne, Lab-STICC (UMR CNRS 3192), Université de Bretagne-Sud, Cen-
tre Yves Coppens, Campus de Tohannic, F-56017, Vannes & Projet INRIA Calvi, Université de Strasbourg,
IRMA, 7 rue René Descartes, F-67084 Strasbourg Cedex, France.
†Université de Strasbourg, IRMA, 7 rue René Descartes, F-67084 Strasbourg Cedex, France & Projet
INRIA Calvi.
1
This talk (and proceeding) is the first step into the rewriting of the Gyro-Kinetic
Approximation Theory into a mathematically affordable theory. It is only a summa-
rize, to understand, with a mathematical slant, several aspects of what is done in the
references just cited.
Charge particles submitted to Strong Magnetic Field - The context of
the Gyro-kinetic Approximation is Tokamak Physics. An artist vision of Iter, which
is a Tokamak, is given in Figure 1. The vessel of a Tokamak is the interior of a torus
with a vertical axis of symmetry. Along the torus, electromagnets can generate a large
magnetic field.
Figure 1: Artist vision of Iter.
We begin by considering a charged particle within the vessel of a Tokamak in service. In
Usual Coordinates (x, v) = (x1, x2, x3, v1, v2, v3), where x stands for the position vari-
able and v for the velocity variable, the position and the velocity at time t (X(t; x, v, s),
V(t; x, v, s)) of the particle which is in x with velocity v at time s is the solution to
∂X
∂t
∂V
∂t
= V,
(E(X) + V × B(X)).
=
q
m
(1)
(2)
This dynamical system means nothing but that time derivative of the position is the
velocity and that time derivative of the velocity, which is the acceleration, in linked with
Lorentz force by Newton's law. In this dynamical system magnetic field B is composed
of a strong applied piece, of a strong self induced piece and of a self induced pertur-
bations. The strong applied piece is generated by the electromagnets and wind around
the axis of symmetry of the tokamak. Charged particles winding, with a relatively
large velocity, around the axis of symmetry within the vessel generate the indispensable
vertical component of the large magnetic field. This is what is called here the strong
self induced piece. The part of the self induced magnetic field which is not vertical is
called here the self induced perturbations and is, from now, forgotten. Electric field E
is induced by the particles of the vessel.
Without loss of generality, we can consider that E is the opposite of the gradient of
an electric potential, and B the curl of vector potential, i.e.:
E = −∇Φ, B = ∇ × A.
(3)
Helicoidal trajectories - Larmor Radius -
It is well known that the trajec-
tory of a charged particle submitted to a magnetic field is a helix. The helix axis is
the magnetic field direction and its radius, which is called in the context of Tokamak
2
Plasma the Larmor Radius, equals the norm of the projection of the velocity on the
plan orthogonal to the magnetic field divided by the particle's mass time the norm of
the magnetic field. As illustrated in left picture of Figure 2, because of the mass ratio
Figure 2: Left: Helicoidal trajectories of ions and electrons submitted to a magnetic field.
(Source: S. Jardin's Lectures at Cemracs'10). Right: Screen effect and Debye Length in a
plasma.
between ions and electrons, the order of magnitude of electron Larmor Radius is much
smaller than that of ions. In Tokamak, those order of magnitude are ∼ 5 · 10−4m for
electrons and ∼ 10−2m for ions. Those two lengths are two important scales of Tokamak
Plasma Physics.
To simplify the purpose, from now we will forget the scale of electron Larmor Radius
and only take ions under consideration.
Debye Length - We now explain an important phenomenon of Plasma Physics
which is the Screen Effect, with its related length which is the Debye Length.
It is well known that a charged particle generates around it an electric potential
which depends on the inverse of the distance to it, say ∼ 1/r. Since this function is
the Green Kernel of the opposite to the Laplace operator, when following the Mean
Field routine to deduce the model taking under consideration interactions of many
particles, it is gotten that the electric field that fills the space is the solution to −∆Φ =
Cst Charge Density. Because, of the long range of ∼ (1/r)-potential, the regularization
properties of minus Laplacian are such that space variations of the electric potential (and
consequently of the electric field) occure at a large scale.
Now, if the density of particles is relatively large, as it happens in Plasma, and
if we consider a test particle (which is drawn in blue in right picture of Figure 2),
the other particles (in the picture, orange particles are positively charged and green
particles negatively charged) which are beyond a given length λ, which is the Debye
Length (the black circle in the picture is censed being of radius λ), may be gathered
into subsets (the two grey ellipses in the picture are such subsets) which are such that
the resulting action of pairs of those subsets on the test particle is negligible. This has
the following consequence: it can be considered that the potential ∼ 1/r generated by
a particle may be approximated by a ∼ e−r/λ/r potential. Then, following the Mean
Field routine with this new assumption leads an electric field that fills the space which
is the solution to −DΦ = Cst Charge Density. Operator −D is the pseudo-differential
operator which Green Kernel is e−r/λ/r whose regularization properties are much less
than that of minus Laplacian. Then the regularity of the electric field is not as high as
we could think before this formal analysis. In particular the electric field can be prone
3
to variations over lengths ranging from several Debye Lengths to several hundreds of
Debye Lengths. Those variations are maybe not extremely large, but as no principle
fends them off, they are stable.
What Ions see -
In a very simplified slant, Electric field in a plasma is a bed with
variations which are not so large but not negligible, and, more than anything, stable
at scales ranging from several Debye Lengths to several hundreds of Debye Lengths. In
Figure 3: Symbolic picture of oscillating trajectory of ion in a electric field with variations
with typical size comparable with trajectory oscillation.
Figure 3 we symbolically represent this bed as a charcoal grey surface. (It is symbolic
because a surface is drawn while, in real world, Electric Field fills a 3D space.) This
surface shows variations at two scales which are cenced to be about 100 times the Debye
Lengths and 1.000 times the Debye Lengths. Those variations are not of large amplitude
but are stable. Now, ion motion, symbolically represented in red in the figure, shows
an oscillation of amplitude which is about 100 times the Debye Length (ion Larmor
Radius ∼ 10−2m, Debye Length : ∼ 10−4m). Hence, because of the stability of bed
variations, its resulting effect can be important on ion trajectory after several periods
of oscillation. To take into account this resulting effect, the concept of gyro-average was
introduced.
Dimensionless Dynamical System - Here, we do not present the scaling rou-
tine, we only mention that two small parameters may be introduced. The first one
is
ε ∼ Ion Larmor Radius
Tokamak size
∼ 10−2m
10m
∼ 10−3
(4)
The second one is η the Debye Length linked variations of the Electric Field and Po-
tential. It is defined as
η ∼ Characteristic scale of variations of the Electric Field
Tokamak size
.
(5)
4
00.511.522.5300.511.522.53012345678910 If we follow a scaling routine, we will get that the dimensionless Electric Potential and
Field write
Φ(x) = Φ0(x) + ηΦ1(
x
η
) and E(x) = E0(x) + E1(
x
η
),
with
E(x) = −(∇Φ)(x),
and that the dimensionless Magnetic Field write
= (∇ × A
ε
B(x)
ε
)(x).
Then, the dimensionless trajectory, is solution to the following dynamical system:
∂X
∂t
∂V
∂t
= V
= E0(X) + E1(
X
η
) + V × B(X)
ε
.
(6)
(7)
(8)
(9)
(10)
We guess that anybody agree with the fact that this system is a good model to describe
motion of charged particle within Tokamak.
Gyro-kinetic Model - Yet, there is the Gyro-Kinetic Model.
It claims that a
trajectory (r, ψ, ϕ, W(cid:107), J, Γ) in a coordinates system (r, ψ, ϕ, w(cid:107), j, γ) we will discuss
on later on, is solution to:
= (EDr + M CDr),
=
W(cid:107)
R
,
∂ψ
∂t
∂W(cid:107)
∂t
∂r
∂t
∂ϕ
∂t
where
=
+
EDψ + M CDψ
W(cid:107)
r
q(r)R
= (E0 + (cid:104)E1(cid:105))(cid:107) − J
∇(cid:107)B +
ε
,
(11)
W(cid:107)
B ED · (∇B),
Figure 4: Position coordinates in use in a Tokamak. (Source: V. Grandgirard's Lectures at
Cemracs'10).
R = R0 + r cos(ψ), ED = ε
B × (E0 + (cid:104)E1(cid:105))
B2
, M CD = ε
B × (∇B)
B3
(W 2(cid:107) + J
B
ε
),
(12)
5
Ψ and
(cid:16) ∂
∂ϕ
∇(cid:107) =
1
R
(cid:17)
.
+
1
q(r)
∂
∂ψ
(13)
We now explain how to read equations (11) - (13). A Tokamak is the interior of a
torus with a vertical symmetry axis. (A Tokamak and its associated position coordinate
system (r, ψ, ϕ) is given in Figure 4.) Hence, it is thought as a small disc carried by
a large circle of radius R0. Then, for a given point of the Tokamak, r is (almost) its
distance from the center of the small disk, ψ is (almost) the angle that the point makes
in the small disk and ϕ is (almost) the angle that it makes on the large circle.
In what concerns the velocity variable system (w(cid:107), j, γ), if a moving point is in (r, ψ, ϕ),
the w(cid:107)-component of its velocity stands (almost) for the projection of its velocity on the
direction the Magnetic Field has in (r, ψ, ϕ) (which is not far from being tangent to the
large circle but which has a small component tangent to the small circle), j is (almost)
half the square of the norm of its velocity on the plan orthogonal to the Magnetic Field
direction, and γ is (almost) the angle that this projection makes in this plan.
Above, precision "almost" is essential, and constitutes the core of the Gyro-Kinetic
Approximation which consists, as we will see, in building infinitesimal changes of coor-
dinates to set out a dynamical system shaped like (11).
We now comment system (11), forgetting that the variables are only almost what
they are supposed to be. The first equation in (11) says that the r-component of a
particle trajectory varies with a velocity which is the projection on the direction of the
vector joining the center of the small disk and the point of the Electric Drift velocity ED
and the Magnetic Curvature Drift velocity M CD, both defined in (12). The Electric
Drift velocity ED involves the cross product of the Magnetic Field and what is called
the gyro-average of the Electric Field, which is indicated by symbol (cid:104) (cid:105). The Magnetic
Curvature Drift velocity M CD involves the cross product of the Magnetic Field with
the gradient of the Magnetic Field norm.
In the second equation of (11), q(r) is called the quality factor. Essentially, it is the
number of revolutions in the small disk a particle makes while making one revolution
along the large circle. Then, term W(cid:107)/(q(r)R), where R is defined in (12), generates
the winding of the particle trajectory in the small disk (this is illustrated by the red
curves drawn on the left of the figure). The second term of the second equation of (11)
is linked with the action of the Electric Drift velocity and Magnetic Curvature Drift
velocity.
The third equation involves the contribution W(cid:107)/R of the w(cid:107)-component of the particle
velocity to the variation of the ϕ-component of the particle position.
The last equation describes the variation of w(cid:107)-component of the particle velocity. It
involves, among other terms, the projection of the gyro-average of the Electric Field on
the direction of the Magnetic Field and the variation of the Magnetic Field norm in the
direction of the Magnetic Field, materialized by operator ∇(cid:107) defined by (13).
In system (11), their is neither equation for J (the j-component of the particle
velocity) nor for Γ (the γ-component of the particle velocity). This may be considered
as a strange fact. The reason why this happens is that the dynamics of r, ψ, ϕ, W(cid:107),
the r-, ψ-, ϕ-, w(cid:107)-components of the particle trajectory and velocity do not depend on
the γ-component of the particle velocity and that the its j-component is invariant. This
is summarized as:
∂J
∂t
= 0,
∂Γ
∂t
= Does not matter.
(14)
6
What we explain and what we do not explain -
In this document we give
the key ideas that bring the deduction of the Gyro-Kinetic Dynamical System (12) from
the original one (9) - (10).
Among thing that we do not tackle, there is the question of the specific position
coordinates (r, ψ, ϕ). We will remain with a generic position coordinate system. We
will not talk about Gyro-average either. The question of the coupling of Gyro-Kinetic
Approximation with computation of Electric and Magnetic Fields, using Quasi-Neutral
Poisson Equation or Maxwell Equation is also not addressed here.
To end this paragraph we mention that the method only works in case when η is
larger than ε, or in other words, when η = ε1−κ for κ > 0. This is a important limitation
in the theory.
2 Methode summarize
Key result - The method allowing us to go from system (9) - (10) to a system of
the kind of (11), i.e. to make disappear two equations, has nothing to do with magic.
It is based on the following result:
Theorem 2.1. If, in a given coordinate system r = (r1, r2, r3, r4, r5, r6), a Hamiltonian
Dynamical System writes:
= P(R)∇rH(R),
P(R) =
∂R
∂t
M 0
:
0
0··· 0
0
0··· 0 −1
,
0
:
0
1
0
(15)
(16)
(17)
(for a trajectory R = (R1, R2, R3, R4, R5, R6)T )) with a Hamiltonian Function that does
not depend on the last variable, i.e.
Then, submatrix M does not depend on the two last variables, i.e.
∂H
∂r6
= 0,
∂M
∂r5
= 0 and ∂M
∂r6
= 0,
and consequently, time-evolution of the four first components R1, R2, R3, R4 is indepen-
dent of the last component R6; and, the penultimate component R5 of the trajectory in
not time-evolving, i.e.
∂R5
∂t
= 0.
(18)
The proof of property (17) is relatively straightforward by a calculation. Equation
(18) is a direct consequence of (16) that yields a sixth component of ∇rH(R) with worth
0 and consequently, because of the form of P(R) a fifth line of Dynmical System (15)
which is exactly (18).
Essentially, the Gyro-Kinetic version of Dynamical System (9) - (10) exhibit es-
sentially a property like (18) and the property that the four first components of the
trajectory are independent of the last component.
As a matter of fact, building the Gyro-Kinetic Approximation consists in building a
change of coordinates that bring a system of coordinates in which P(R) has the same
form as in (15) and in which (16) is true.
7
Usual Coordinates
(x, v)
= V
3
Cylindrical Coordinates
(x, v(cid:107), v⊥, θ)
= E0(X)+E1(
X
η
)+V× B(X)
ε
4: Darboux Method
∂X
∂t
∂V
∂t
1: Hamiltonian?
2
Canonical Coordinates
(q, p)
H ε
η = H ε
η (q, p):
∂Q
∂t
∂P
∂t
η
= ∇p H ε
= −∇q H ε
η
Darboux Almost
Canonical Coordinates
(y, u(cid:107), k, θ)
5: Lie Method
Lie Coordinates
(z, w(cid:107), j, γ)
Usual Coordinates
η(x, v), P ε
H ε
η (x, v) s.t:
(x, v)
= P ε
∂X
∂t
∂V
∂t
η ∇x,vH ε
η
1: Hamiltonian?
2
Canonical Coordinates
(q, p)
η(q, p), P ε
H ε
η (q, p)=S
= S∇q,p H ε
η
s.t: ∂Q
∂t
∂P
∂t
Cylindrical Coordinates
(x, v(cid:107), v⊥, θ)
(cid:101)H ε
η(x, v(cid:107), v⊥, θ), (cid:101)P ε
η (x, v(cid:107), v⊥, θ)
3
4: Darboux Method
Darboux Almost
Canonical Coordinates
(y, u(cid:107), k, θ)
ε
η(y, u(cid:107), k, θ), P ε
H
η (y, u(cid:107), k, θ)
5: Lie Method
Lie Coordinates
(z, w(cid:107), j, γ)
(cid:98)H ε
η(z, w(cid:107), j), (cid:98)P ε
η (z, w(cid:107), j, γ)
= P ε
η (z, w(cid:107), j, γ)
Figure 5: The method is made of 5 steps. 1: Check that the Dynamical System (9)- (10) is
well Hamiltonian. 2: Write the system using Hamiltonian Function and Poisson Matrix. 3:
Write the system in a cylindrical in velocity coordinate system using the formula giving how
the Poisson Matrix and the Hamiltonian Function is transformed by change of coordinates.
4: Make another change of coordinates in order to have the Poisson Matrix form allowing
the application of the Key Result (Theorem 2.1). 5: Make a last change of coordinates,
leaving the Poisson Matrix form unchanged and leading to a Hamiltonian independent of
the last variable.
8
η and the gradient ∇x,vH ε
Panorama - As illustrated in Figure 5, the method to build the desired change of
coordinates is made of 5 steps. The first one consists in checking that the Dynamical
System (9) - (10) is well Hamiltonian. This is symbolized by arrow 1 in the top picture
of the Figure. Once this is done, we can go back into the Usual Coordinate System but
knowing that the system writes as in the square which is in the top-left of the bottom
picture, i.e. involving a Poisson Matrix P ε
η of a Hamiltonian
Function. It may be written in that form in any Coordinate System, and formula give
how to transform the Hamiltonian Function and the Poisson Matrix while changing
of Coordinates. The goal of the third step is to introduce a Cylindrical in Velocity
Coordinate system which is known has being close to the Gyro-Kinetic Coordinate
System. Writing the system in this coordinate system uses the formula giving how the
Poisson Matrix and the Hamiltonian Function are transformed by change of coordinates.
(At this level, it would be also possible to make a position change of coordinates fitting
Tokamak geometry.) In the fourth step, we make another change of coordinates in order
to have the Poisson Matrix form allowing the application of the Key Result (Theorem
2.1). In the fifth step, we make a last change of coordinates, leaving the Poisson Matrix
form unchanged and leading to a Hamiltonian independent of the last variable.
3 Hamiltonian System
We make here the first step: we check that Dynamical System (9) - (10) is Hamiltonian.
Essentially, this means nothing but that there exists a coordinate system in which it
writes:
∂Q
∂t
∂P
∂t
= ∇p H ε
η ,
= −∇q H ε
η ,
∂Q
= S∇q,p H ε
η , with S =
or
∂t
∂P
∂t
η is a function.
where H ε
(19)
(20)
(21)
(cid:18) 0
−I3
(cid:19)
,
I3
0
The Canonical Coordinates -
the Usual Coordinates (x, v) = (x1, x2, x3, v1, v2, v3), trajectory : (X(t; x, v, s), V(t; x, v, s))
is solution to
If we use (7) and (8) in (9) - (10), we get that in
∂X
∂t
∂V
∂t
= V,
= −∇(cid:2)Φ0(X) + ηΦ1(
)(cid:3) + V × ∇ × A(X)
ε
.
X
η
(22)
(23)
We recall that (X, V) = (X1, X2, X3, V1, V2, V3).
We will show, in the next paragraph, that in coordinate system (q, p) = (q1, q2, q3, p1, p2, p3)
defined by
q = x, p = v +
A(x)
ε
,
9
(24)
with reverse transformation given by
x = q, v = p − A(q)
ε
,
(25)
the trajectory (Q(t; q, p, s), P(t; q, p, s)) ((Q, P) = (Q1, Q2, Q3, P1, P2, P3)) which is
given by
Q = X, P = V +
A(X)
ε
,
X = Q, V = P − A(Q)
ε
,
(26)
is solution of a Dynamical System of the form (19) - (20) or (21), with
H ε
η (q, p) =
1
2
+ Φ0(q) + ηΦ1(
q
η
).
(27)
Those (q, p) coordinates will be called Canonical Coordinates.
Check of Canonical nature of Canonical Coordinates - Making calcula-
tions suggested by (19) - (20) with H ε
η given by (27) gives
(cid:12)(cid:12)(cid:12)p − A(q)
ε
(cid:12)(cid:12)(cid:12)2
∂Q
∂t
∂P
∂t
= ∇p H ε
= −∇q H ε
η (Q, P) = P − A(Q)
,
(∇A(Q))T
ε
η (Q, P) =
ε
(cid:16)
(cid:17) − ∇(cid:2)Φ0(Q) + ηΦ1(
)(cid:3),
Q
η
P − A(Q)
ε
which applying the following formula
(∇A)T (p − A) = (∇A)(p − A) + (p − A) × (∇ × A),
yields
∂Q
∂t
∂P
∂t
= P − A(Q)
− (∇A(Q))
(cid:16)
ε
ε
,
P − A(Q)
ε
(cid:17)
(cid:16)
=
(cid:17) × ∇ × A(Q)
ε
− ∇(cid:2)Φ0(Q) + ηΦ1(
P − A(Q)
ε
Now,using (26) we get
∂X
∂t
= V,
∂
P − A(Q)
(cid:16) ∂Q
(cid:17)
− (∇A(Q))
(cid:17) × ∇ × A(Q)
− ∇(cid:2)Φ0(Q) + ηΦ1(
− ∇(cid:2)Φ0(X) + ηΦ1(
)(cid:3).
∂t
∂t
=
ε
ε
ε
Q
η
(cid:104)
X
η
(cid:105)
)(cid:3)
=
∂V
∂t
∂P
∂t
P − A(Q)
(cid:16)
= V × ∇ × A(X)
ε
=
ε
(28)
(29)
(30)
(31)
)(cid:3).
Q
η
(32)
(33)
(34)
which is (22) - (23), and then proving the equivalence between the two systems and
leads to the conclusion that the Dynamical System we take interest in is Hamiltonian.
10
As by-products : Poisson Matrix, Poisson Bracket, Change of Coordi-
nates Formula - To end the first step we now list some properties we can moreover
deduce from the fact that Dynamical System (9) - (10) is Hamiltonian.
In any coordinate system r = (r1, r2, r3, r4, r5, r6), the Dynamical System writes:
for a Poisson Matrix P, which is antisymmetric, and a Hamiltonian Function H.
= P(R)∇rH(R),
∂R
∂t
(35)
We can consider the Poisson Bracket which is defined for two regular functions f
and g : R6 → R by
{f, g}(r) = (∇rf (r)) · (P(r)(∇rg(r))).
(36)
If one of the two function is vector-valued, i.e. if for instance f : R6 → R6 and g : R6 →
R, then it is defined, for any component i = 1, . . . , 6, by
({f , g}(r))i = (∇rfi(r)) · (P(r)(∇rg(r))).
(37)
Then, if we introduce the Coordinate Function I, defined by I(r) = r, then (35) reads
also:
= {I, H}(R).
∂R
∂t
(38)
Remark 3.1. In papers of physicists, formula (38) is generally read ∂r
∂t
= {r, H}.
In another coordinate system r = (r1, r2, r3, r4, r5, r6) which is tied with the first one
by change-of-coordinates formula r = ρ(r) and r = ρ(r) = ρ−1(r), the system writes
∂ R
∂t
= P( R)∇r H( R),
where R = ρ(R), with Hamiltonian Function and Poisson Matrix given by
H(r) = H(ρ(r)) and ( P(r))ij =(cid:8)ρi, ρj
(cid:9)(ρ(r)).
(39)
(40)
Remark 3.2. From (27) and applying (40), in Usual Coordinates, Hamiltonian Func-
tion expression is
x
η
),
1
2
+ Φ0(x) + ηΦ1(
(cid:12)(cid:12)v(cid:12)(cid:12)2
(cid:18)∇(cid:2)Φ0(x) + ηΦ1( x
η )(cid:3)
(cid:19)
(cid:32)
v
,
(cid:33)
.
0
−I3
I3
(∇A(x))T −(∇A(x))
ε
11
(41)
(42)
(43)
leading to
H ε
η(x, v) =
∇x,vH ε
η =
and Poisson Matrix expression is
P ε
η (x, v) =
(cid:18) 0
−I3
(cid:19)
I3
0
, ∇υ6(q, p) =
(45)
.
(q)
(q)
(q)
1
ε
1
ε
1
ε
∂A3
∂q1
∂A3
∂q2
∂A3
∂q3
0
0
1
(cid:17)
This can be easily checked calculating
P ε
η (x, v)∇x,vH ε
η =
−∇(cid:2)Φ0(x) + ηΦ1(
v
)(cid:3) +
x
η
(∇A(x))T − (∇A(x))
ε
,
v
(44)
using (29),and comparing with (22) - (23).
Expression (41) is consistent with (40), let us check that it is the same for (43). To
do this, let us write the change of coordinates in the following way: (x, v) = υ(q, p)
and (q, p) = π(x, v). Then, for instance: υ5(q, p) = p2 − A2(q)
, υ6(q, p) = p3 − A3(q)
and
ε
ε
(q)
(q)
(q)
1
ε
1
ε
1
ε
∂A2
∂q1
∂A2
∂q2
∂A2
∂q3
0
1
0
, S =
∇υ5(q, p) =
(cid:16) ∂A3
∂q2
Hence, since {υ5, υ6} = (∇υ5) · (S∇υ6), expression of (P ε
is
η )56, obtained applying (40),
{υ5, υ6}(π(x, v)) =
1
ε
(x) − ∂A2
∂q3
(x)
,
(46)
which is also the expression of (P ε
led for the other entries of the matrix also.)
η )56 obtained applying (43). (This may of course be
4 Cylindrical Coordinates
Cylindrical Coordinates in velocity - We now turn to the second step which
consists in setting the expression of Dynamical System (22) - (23) (or (9) - (10) or (19),
(20), (27)) in a Cylindrical in velocity Coordinate System (x, v(cid:107), v⊥, θ) which is such
that
Figure 6: Helicoidal trajectory and Cylindrical Coordinates for the velocity variable.
12
B v (cid:757) v v θ ⎮
⎮
v(cid:107) = v · B
B , v⊥ =
(cid:12)(cid:12)(cid:12)v −(cid:16)
(cid:17) B
B
v · B
B
(cid:12)(cid:12)(cid:12), θ s.t. v −(cid:16)
v · B
B
(cid:17) B
B = v⊥(cos θ, sin θ).
(47)
This means, as shown in Figure 6 where a helicoidal trajectory induced by a Magnetic
Field pointing upward, that v(cid:107) is the projection of the velocity on the direction of the
Magnetic Field, v⊥ is the norm of the projection of the velocity on the plan orthogonal
to the Magnetic Field and θ is the angle that this projection makes in this plan.
System:
Hamiltonian Function and Poisson Matrix in Cylindrical Coordinates -
Applying formula (41), we get the expression of the Hamiltonian Function in this
(cid:101)H ε
η(x, v(cid:107), v⊥, θ) =
1
2
(cid:0)v2(cid:107) + v2⊥(cid:1) + Φ0(x) + ηΦ1(
x
η
).
(48)
We can also get the expression of the Poisson Matrix, which is heavy, applying formula
(41).
Here, we give its expression in the case when
, b > 0,
b(x)
0
0
B(x) =
0
0
0
− b(x)
ε
0
0
0
0
0
0
sin(θ)
cos(θ)
v⊥
only. It is:
(cid:101)P ε
η (x, v(cid:107), v⊥, θ) =
where
(49)
(50)
,
b(x)
0
0
0
0
ε
0
0
0
cos(θ) −$$
− sin(θ)
−$
v⊥
0
0
− sin(θ) − cos(θ)
v⊥
− cos(θ)
sin(θ)
v⊥
$
− b(x)
εv⊥
0
$$
0
b(x)
εv⊥
$ =
v(cid:107)
ε
(
∂b
∂x2
(x) +
∂b
∂x3
(x)) and $$ =
v(cid:107)
εv⊥
(sin(θ)
∂b
∂x2
(x) +
1 + sin2(θ)
cos(θ)
∂b
∂x3
(x)).
(51)
In more general cases, we do not give the Poisson Matrix expression. Yet we mention
a very important fact which is that
((cid:101)P ε
η (x, v(cid:107), v⊥, θ))56 = ((cid:101)P ε
η (x, v(cid:107), v⊥, θ))v⊥θ =
B(x)
εv⊥
> 0,
(52)
is always true. This is Important for the Darboux Method which will be reached in a
few lines. We will denote:
B(x)
εv⊥
= ω(x, v⊥).
13
(53)
5 Darboux Algorithm
Darboux Algorithm Target - The third step is the application of a mathematical
algorithm, so called the Darboux Algorithm, to build a Coordinate System (y, u(cid:107), k, θ)
in which the Poisson Matrix has the required form, given by (15), to apply the Key
Result. In fact, in order to manage the small parameter ε, we will build the Coordinate
System (y, u(cid:107), k, θ) in order to get P ε
η with the following form:
.
0
:
0
1
ε
0
M 0
:
0
0··· 0
0
0··· 0 − 1
ε
(54)
We introduce the following notations to manage change-of-coordinates mappings: (y, u(cid:107),
k, θ) = Υ(x, v(cid:107), v⊥, θ) and (x, v(cid:107), v⊥, θ) = ξ(y, u(cid:107), k, θ), (ξ = Υ−1).
An important and constitutive fact in the Darboux Algorithm is that the θ-variable is
left unchanged.
Now, Since
(P ε
η (y, v(cid:107), k, θ))ij = {Υi, Υj}(ξ(y, v(cid:107), k, θ)),{Υi, Υj} = (∇Υi) · ((cid:101)P ε(∇Υj)),
(55)
the bottom-right of form given in (54), results from:
{Υ6, Υ5} = − 1
ε
(56)
Remark 5.1. If we write Υ6 = Υθ and Υ5 = Υk, (56) may be also read {Υθ, Υk} = − 1
ε .
In articles of physicists, this last equation reads {θ, k} = − 1
ε .
.
In the same way, the fact that the two last lines (or columns) contain only zeros
results from:
{Υ1, Υ5} = 0, {Υ1, Υ6} = 0,
{Υ2, Υ5} = 0, {Υ2, Υ6} = 0,
{Υ3, Υ5} = 0, {Υ3, Υ6} = 0,
{Υ4, Υ5} = 0, {Υ4, Υ6} = 0.
Remark 5.2. Using the same conventions as in Remark 5.1, (57) may also read
{Υy1 , Υk} = 0, or {y1, k} = 0,
{Υy2 ,Υk} = 0, or {y2, k} = 0,
{Υy3 ,Υk} = 0, or {y3, k} = 0,
{Υy4 ,Υk} = 0, or(cid:8)v(cid:107), k(cid:9) = 0,
{Υy1 ,Υθ} = 0, or {y1, θ} = 0,
{Υy2 ,Υθ} = 0, or {y2, θ} = 0,
{Υy3 ,Υθ} = 0, or {y3, θ} = 0,
(cid:9) = 0, or(cid:8)v(cid:107), θ(cid:9) = 0.
(cid:8)Υv(cid:107) ,Υθ
(57)
(58)
Equations (56) and (57) are hyperbolic PDEs that need to be solve to get change-
of-coordinates mapping Υ.
First equation processing - Since (and this is a consequence of the fact that the
θ-variable is left unchanged in the sought change-of-coordinates)
∇Υ6(= ∇Υθ) = (0, 0, 0, 0, 0, 1)T ,
(59)
14
we deduce {Υ6, Υ5} = (∇Υ6) · ((cid:101)P ε(∇Υ5)) is the last component of ((cid:101)P ε(∇Υ5)). Hence,
equation (57) reads
F1
∂Υ5
∂x1
+ F2
∂Υ5
∂x2
+ F3
∂Υ5
∂x3
+ F(cid:107)
∂Υ5
∂v(cid:107)
+ ω
∂Υ5
∂v⊥
= − 1
ε
,
(60)
where Fn and ω are functions of (x, v(cid:107), v⊥, θ). Function ω is given by (53) and is positive.
Solving (60) will give Υ5(x, v(cid:107), v⊥, θ) which is the expression of component k of the
Darboux Coordinates in terms of the Cylindrical in Velocity Coordinates (x, v(cid:107), v⊥, θ).
To solve this equation we will use the Method of Characteristics.
Method of Characteristics - Dividing by ω, which is possible since ω is positive,
(60) gives
∂Υ5
∂v⊥
+ ε
v⊥F1
B
∂Υ5
∂x1
+ ε
v⊥F2
B
∂Υ5
∂x2
+ ε
v⊥F3
B
∂Υ5
∂x3
+ ε
v⊥F(cid:107)
B
∂Υ5
∂v(cid:107)
=
v⊥
B .
(61)
In order to get a solution to this equation, we need to add a boundary (or initial)
condition in a point where the involved vector field (εv⊥/B)(F1, F2, F3, F(cid:107)) does not
vanish. Hence we set
Υ5v⊥=ν = 0
for a small ν > 0.
(62)
We consider the following characteristics (X ,V(cid:107)) = (X1,X2,X3,V(cid:107)):
X1(v⊥; x, v(cid:107), u⊥) such that
X2(v⊥; x, v(cid:107), u⊥) such that
X3(v⊥; x, v(cid:107), u⊥) such that
V(cid:107)(v⊥; x, v(cid:107), u⊥) such that
∂X1
∂v⊥
∂X2
∂v⊥
∂X3
∂v⊥
∂V(cid:107)
∂v⊥
= ε
= ε
= ε
= ε
and the solution to (61) - (62) is given by
v⊥F1(X1,X2,X3,V(cid:107), v⊥, θ)
B(X1,X2,X3)
v⊥F2(X1,X2,X3,V(cid:107), v⊥, θ)
B(X1,X2,X3)
v⊥F3(X1,X2,X3,V(cid:107), v⊥, θ)
B(X1,X2,X3)
v⊥F(cid:107)(X1,X2,X3,V(cid:107), v⊥, θ)
B(X1,X2,X3)
, X1(u⊥) = x1,
, X2(u⊥) = x2,
, X3(u⊥) = x3,
, V(cid:107)(u⊥) = v(cid:107),
(63)
(cid:90) v⊥
ν
(cid:90) v⊥
Υ5(x, v(cid:107), v⊥, θ) = Υ5(X (ν; x, v(cid:107), v⊥),V(cid:107)(ν; x, v(cid:107), v⊥), ν, θ)+
s
B(X (s; x, v(cid:107), v⊥)) ds
s
(64)
=
ν
B(X (s; x, v(cid:107), v⊥)) ds.
(The last equality is gotten because of (62).)
Rewriting system (63) in the following compact form,
(cid:19)
(cid:18)X
V(cid:107)
∂v⊥
∂
= εF(X 1,V(cid:107), v⊥, θ),
(u⊥) =
(cid:19)
(cid:18)X
V(cid:107)
(cid:18) x
(cid:19)
v(cid:107)
,
(65)
with F = (v⊥/B)(F1, F2, F3, F(cid:107)) and applying the formula that gives the expansion of
the solution of a dynamical system in terms of its parameter, we get that
15
(cid:19)
(cid:18)X
V(cid:107)
(cid:19)
(cid:18)X
V(cid:107)
=
(v⊥; x, v(cid:107), u⊥),
(66)
(cid:19)
writes(cid:18)X
V(cid:107)
(cid:18) x
(cid:19)
v(cid:107)
=
=
+
v⊥
1!
(cid:18) x
(cid:19)
v(cid:107)
εF(x, v(cid:107), v⊥, θ) +
LεF(εF)(x1, v(cid:107), v⊥, θ)
v2⊥
2!
+
+
εv⊥
1!
F(x, v(cid:107), v⊥, θ) +
v3⊥
L2
εF(εF)(x1, v(cid:107), v⊥, θ) + . . .
3!
ε2v2⊥
LF(F)(x1, v(cid:107), v⊥, θ)
2!
ε3v3⊥
3!
L2
F(F)(x1, v(cid:107), v⊥, θ) + . . .
+
(67)
where LF(F) is the the Lie Derivative of vector field F in the direction of vector field
F, L2
F(F) = LF(L2
F(F)), and so on.
F(F) = LF(LF(F)) and L3
Rewriting this expansion, we get
X = x + εv⊥X 1 + ε2v2⊥X 2 + ε3v3⊥X 3 + . . .
V(cid:107) = v(cid:107) + εv⊥V 1(cid:107) + ε2v2⊥V 2(cid:107) + ε3v3⊥V 3(cid:107) + . . .
(cid:19)
(cid:19)
(cid:18)X 2
(cid:18)X 3
LF(F)(x, v(cid:107), v⊥, θ),
=
=
V 2(cid:107)
1
2!
V 3(cid:107)
= F(x, v(cid:107), v⊥, θ),
(cid:19)
with(cid:18)X 1
V 1(cid:107)
(68)
(69)
L2
F(F)(x, v(cid:107), v⊥, θ), . . .
1
3!
(70)
Now, injecting (69) in expression (64) on Υ5, we get:
Υ5(x, v(cid:107), v⊥, θ) =
(cid:90) v⊥
(cid:90) v⊥
ν
ν
ν
s
s
1
B(x) )X 1ds+
B(x) ds + ε
(cid:90) v⊥
(cid:90) v⊥
s3(cid:16)T 2(
(cid:90) v⊥
(cid:90) v⊥
s3(cid:16)T 2(
B(X (s; x, v(cid:107), v⊥)) ds =
s2 T 1(
B(x) )X 1 + T 1(
B(x) )X 1ds +
B(x) )X 1 + T 1(
s2 T 1(
ε2
ε2
B(x) )X 2(cid:17)
B(x) )X 2(cid:17)
1
1
ν
ν
1
1
1
ν
ds + . . .
ds + . . . ,
(71)
(v⊥ − ν)2
2B(x) + ε
=
where T i(1/B(x)) is the ith coefficient of the Taylor expansion of 1/B(x).
Formula (71) gives the expression of Υ5, i.e. the expression of new variable k in
terms of Cynlindrical Coordinates (x, v(cid:107), v⊥, θ), as an expansion in ε.
Remark 5.3. In (71), it is possible to choose ν as small as we want. Hence, the first
term of the expansion is essentially the Magnetic Moment v⊥/B(x).
16
On other equations and on Poisson Matrix in Darboux Coordinates
Equation (56) was processed and gave expression of k. Equations (57) can be processed
using similar Methods of Characteristics. A special attention needs to be given to the
fact that Υ1, . . . Υ4 are solutions of a PDE involving Υ5 and of another one involving
Υ6 that need to be tackled together. They will give y and u(cid:107) as expansions in ε. Once
this done, we will have
Υ = Υ0 + εΥ1 + ε2Υ2 + ε3Υ3 + . . . ,
(72)
or, in other words, the expression of (y, u(cid:107), k) in terms of (x, v(cid:107), v⊥, θ) and as an ex-
pansion in ε.
The other terms of the new Poisson matrix P ε
((cid:98)P ε
η )12 = {Υ1,Υ2}(={Υy1,Υy2}={y1, y2}), ((cid:98)P ε
η (y, u(cid:107), k, θ) are given computing:
η )13 = {Υ1,Υ3}(={Υy1,Υy3}={y1, y3}),
. . . .
Hamiltonian Function in Darboux Coordinates - We now have to com-
pute the Hamiltonian Function in the Darboux Coordinates. For this, we build from
expansion (72) of Υ, an asymptotic expansion of ξ = Υ−1:
ξ = ξ0 + εξ1 + ε2ξ2 + ε3ξ3 + . . . .
(73)
Then, we use the expression of the Hamiltonian Function in Cylindrical in Velocity
Coordinates:
(cid:0)v2(cid:107) + v2⊥(cid:1) + Φ0(x) + ηΦ1(
x
η
),
(74)
η(x, v(cid:107), v⊥, θ) =
(cid:101)H ε
η(ξ(y, u(cid:107), k, θ)) = (cid:101)H ε
η(y, u(cid:107), k, θ) = (cid:101)H ε
(cid:101)H ε
η(ξ0) + εT 1((cid:101)H ε
1
2
ε
H
and the following formula, which is gotten from (40),
η(ξ0 + εξ1 + ε2ξ2 + ε3ξ3 + . . . ) =
η)(ξ0) · ξ1 + ··· =
u2(cid:107)
2
+ B(y)k + Φ0(y) + ηΦ1(
y
η
) + εH1,η(y, u(cid:107), k, θ) + ε2H2,η(y, u(cid:107), k, θ) + . . . .
(75)
Notice that for this expansion to be valid, it is necessary to have the property we men-
tion on page 7: η = ε1−κ for κ > 0.
In expression (75), there is an important fact for the setting out of the to come Lie
Transform based Method: the first term is independent of θ.
6 Lie Transform based Method
Lie Transform based Method Target - As a result of the Darboux Algorithm,
we obtained a Poisson Matrix P ε
η (y, u(cid:107), k, θ) with the required form to apply the Key
Result (Theorem 2.1), but the resulting Hamiltonian Function
H
ε
η(y, u(cid:107), k, θ) = H0,η(y, u(cid:107), k) + εH1,η(y, u(cid:107), k, θ) + ε2H2,η(y, u(cid:107), k, θ) + . . . ,
(76)
depends on θ. And we need to make this dependency to vanish.
17
Let us notice two important facts, which are in fact linked: The first term in the
asymptotic expansion (75) does not depend on θ and the Hamiltonian Function ex-
pressed in the Cylindrical in Velocity Coordinates does not depend on θ. The conse-
quence of those facts is that it is certainly possible to make the last variable to vanish
using a mapping parametrized by ε and close to identity for small ε.
Moreover, we need to build this sought mapping in such a way that it does not change
the Poisson Matrix expression. This means that, as viewed as functions, P ε
must be the same, i.e.:
η and (cid:98)P ε
η
(cid:98)P ε
η (z, w(cid:107), j, γ) = P ε
η (z, w(cid:107), j, γ) or P ε
η (y, u(cid:107), k, θ) = (cid:98)P ε
η (y, u(cid:107), k, θ),
(77)
for any z, w(cid:107), j, γ or y, u(cid:107), k, θ. Yet, among changes of variables, the symplectic ones do
not change the Poisson Matrix expression. And among symplectic changes of variables,
are the flows of Hamiltonian Vector Fields, which are moreover close to identity for
small values of their parameter.
Hence, the Lie Transform based Method consists in building a change of variables,
parametrized by ε, reading
(y, u(cid:107), k, θ) (cid:55)→ (z, w(cid:107), j, γ) = ζ(ε; y, u(cid:107), k, θ),
with
ζ(ε; y, u(cid:107), k, θ) = (y, u(cid:107), k, θ) + εζ1(y, u(cid:107), k, θ), +ε2ζ2(y, u(cid:107), k, θ) + . . . ,
which is the flow of parameter ε of a Hamiltonian Vector Fields.
(78)
(79)
The way to do - Looking for ζ as the flow of a Hamiltonian Vector Field means
looking for a Hamiltonian Function:
G(ε; y, u(cid:107), k, θ) = G0(y, u(cid:107), k, θ) + εG1(y, u(cid:107), k, θ) + ε2G2(y, u(cid:107), k, θ) + . . . ,
(80)
such that ζ(ε; y, u(cid:107), k, θ) is solution to:
= P ε
η∇G, ζ(ε = 0; y, u(cid:107), k, θ) = (y, u(cid:107), k, θ).
∂ζ
∂ε
Hence the target becomes to find functions G0, G1, G2, . . . , such that:
η(z, w(cid:107), j) = H
η(λ(z, w(cid:107), j, γ)) where λ = ζ
ε
−1.
(cid:98)H ε
(81)
(82)
(85)
A result - We will use the following theorem
Theorem 6.1. For any two Hamiltonian Functions:
H
ε
η(y, u(cid:107), k, θ) = H0,η(y, u(cid:107), k) + εH1,η(y, u(cid:107), k, θ) + ε2H2,η(y, u(cid:107), k, θ) + . . . ,
(83)
and
(cid:98)H ε
η(z, w(cid:107), j) = (cid:98)H0,η(z, w(cid:107), j) + ε(cid:98)H1,η(z, w(cid:107), j) + ε2(cid:98)H2,η(z, w(cid:107), j) + . . . ,
(84)
expressed in two different variable systems (y, u(cid:107), k, θ and z, w(cid:107), j, γ) , with the property
that
(cid:98)H0,η(z, w(cid:107), j) = H0,η(z, w(cid:107), j),
18
There exists a Hamitonian Function writing
G(ε; y, u(cid:107), k, θ) = G0(y, u(cid:107), k, θ) + εG1(y, u(cid:107), k, θ) + ε2G2(y, u(cid:107), k, θ) + . . . ,
(86)
such that the solution ζ of the Hamiltonian Dynamical System associated with G and
with parameter ε, i.e. solution to:
is such that
= P ε
η∇G, ζ(ε = 0; z, w(cid:107), j, γ) = (y, u(cid:107), k, θ),
∂ζ
∂ε
(cid:98)H ε
η(z, w(cid:107), j) = H
η(λ(z, w(cid:107), j, γ)) where λ = ζ
ε
−1.
Moreover, the Hamiltonian Function g(ε; z, w(cid:107), j, γ) such that λ is solution to
= P ε
η∇g, λ(ε = 0; y, u(cid:107), k, θ) = (z, w(cid:107), j, γ),
∂λ
∂ε
writes
g(ε; z, w(cid:107), j, γ) = g0(z, w(cid:107), j, γ) + εg1(z, w(cid:107), j, γ) + ε2g2(z, w(cid:107), j, γ) + . . . ,
where g0, g1, g2, . . . are given by
(cid:8)g0, H0,η
(cid:9) = O0(H0,η), (cid:8)g1, H0,η
(cid:8)g2, H0,η
(cid:9) = O1(H0,η, H1,η, (cid:98)H1,η, g0),
(cid:9) = O2(H0,η, H1,η, (cid:98)H1,η, H2,η, (cid:98)H2,η, g0, g1), . . . .
(87)
(88)
(89)
(90)
(91)
for differential operators O0, O1, O2, . . . defined by recursive formula.
Remark 6.1. Despite we do not explicitly write the dependency of λ, g0, g1, g2, . . . ,
λ and G0, G1, G2, . . . with respect η to all those functions do depend on η. Here again,
to be able to write the expansions with respect to ε, it is necessary to have the property
mentioned on page 7: η = ε1−κ for κ > 0.
The Lie Transform based Method - As in the case of the Darboux Algorithm,
we do not do the computations that give ζ and λ. We only give the steps that allows
us to get them. The Lie Transform based Method may be summarized as:
1. Fix : (cid:98)H1,η(z, w(cid:107), j), (cid:98)H2,η(z, w(cid:107), j), . . . .
They can be fixed, a priory, with no restriction. Nevertheless those functions are
involved in the PDEs of (91). Hence, they need to be chosen in a way that leads
to PDEs which are as simple as possible to solve.
2. Solve PDE in (91) to get g0, g1, g2 recursively.
They are quite complex but we can solve them using the Method of Characteristics.
3. Get ζ solving (87).
It is gotten as an expansion in ε.
4. Compute expression of the Lie Variables: (z, w(cid:107), j, γ) = ζ(y, u(cid:107), k, θ)
5. Since (cid:98)H1,η, (cid:98)H2,η, . . . were fixed in the begining of the process, it is not necessary to
compute λ to get: (cid:98)H ε
η(λ(z, w(cid:107), j, γ)). Nonetheless, λ may be gotten
η(z, w(cid:107), j) = H
ε
as an expansion in ε by inverting expansion of ζ.
19
7 Brief conclusion
At the end of the day, in the Lie Coordinates System, the trajectory (Z, W(cid:107), J, Γ) of a
charged particle is solution to a Dynamical System which has the following form:
∂Γ
∂t
∂J
∂t
∂
= 0,
= Something complicated,
(cid:18) Z
(cid:19)
(92)
(93)
(94)
W(cid:107)
∂t
= Something independent of Γ.
This form is essentially the same as the one of (11).
As a conclusion, we just say that the method which is summarized here is the one
that allows us to get (11) or any other Gyro-Kinetic Approximation Model.
Insisting one more time, the important facts in system (92) - (94) is that the evolution
of the components Z and W(cid:107) of the trajectory and of its γ-component Γ are uncoupled
and that component J does not evolve and then becomes a parameter. Hence, even if
the γ-component Γ is not computed, it does not preclude the computation of J, Z and
W(cid:107) solving a collection, parametrized by J, of four-dimensional Dynamical Systems.
References
[1] J. A. Brizard. Nonlinear gyrokinetic Vlasov equation for toroidally rotating ax-
isymmetric tokamaks. Physics of Plasmas, 2(2):459 -- 471, 1995.
[2] D. H. E. Dubin, J. A. Krommes, C. Oberman, and W. W. Lee. Nonlinear gyroki-
netic equations. Physics of Fluids, XXVI(12):3524 -- 3535, 1983.
[3] E. A. Frieman and L. Chen. Nonlinear gyrokinetic equations for low-frequency
electromagnetic waves in general plasma equilibria. Physics of Fluids, 25(3):502 --
508, 1982.
[4] C. S. Gardner. Adiabatic invariants of periodic classical systems. Physical Rieview,
115, 1959.
[5] V. Grandgirard, M. Brunetti, P. Bertrand, N. Besse, X. Garbet, P. Ghendrih,
G. Manfredi, Y. Sarazin, O. Sauter, E. Sonnendrücker, J. Vaclavik, and L. Villard.
A drift-kinetic semi-lagrangian 4d code for ion turbulence simulation. Journal of
Computational Physics, 217(2):395 -- 423, 2006.
[6] V. Grandgirard, Y. Sarazin, P Angelino, A. Bottino, N. Crouseilles, G. Darmet,
G. Dif-Pradalier, X. Garbet, Ph. Ghendrih, S. Jolliet, G. Latu, E. Sonnendrücker,
and L. Villard. Global full-f gyrokinetic simulations of plasma turbulence. Plasma
Physics and Controlled Fusion, 49(12B):B173, 2007.
[7] T. S. Hahm. Nonlinear gyrokinetic equations for tokamak microturbulence. Physics
of Fluids, 31(9):2670 -- 2673, 1988.
[8] T. S. Hahm, W. W. Lee, and A. Brizard. Nonlinear gyrokinetic theory for finite-
beta plasmas. Physics of Fluids, 31(7):1940 -- 1948, 1988.
[9] P.-V. Koseleff. Comparison between deprit and dragt-finn perturbation methods,
1994.
20
[10] M. D. Kruskal. Plasma Physics, chapter Elementary Orbit and Drift Theory.
International Atomic Energy Agency, Vienna, 1965.
[11] R. G. Littlejohn. A guiding center Hamiltonian: A new approach. Journal of
Mathematical Physics, 20(12):2445 -- 2458, 1979.
[12] R. G. Littlejohn. Hamiltonian formulation of guiding center motion. Physics of
Fluids, 24(9):1730 -- 1749, 1981.
[13] R. G. Littlejohn. Hamiltonian perturbation theory in noncanonical coordinates.
Journal of Mathematical Physics, 23(5):742 -- 747, 1982.
[14] T. G. Northrop. The guiding center approximation to charged particle motion.
Annals of Physics, 15(1):79 -- 101, 1961.
[15] T. G. Northrop and J. A. Rome. Extensions of guiding center motion to higher
order. Physics of Fluids, 21(3):384 -- 389, 1978.
[16] F. I. Parra and P. J. Catto. Limitations of gyrokinetics on transport time scales.
Plasma Physics and Controlled Fusion, 50(6):065014, 2008.
[17] F. I. Parra and P. J. Catto. Gyrokinetic equivalence. Plasma Physics and Controlled
Fusion, 51(6):065002, 2009.
[18] F. I. Parra and P. J. Catto. Turbulent transport of toroidal angular momentum in
low flow gyrokinetics. Plasma Physics and Controlled Fusion, 52(4):045004, 2010.
21
|
1812.06766 | 2 | 1812 | 2019-06-21T18:38:03 | A pair of commuting hypergeometric operators on the complex plane and bispectrality | [
"math.FA",
"math.CA",
"math.RT",
"math.SP"
] | We consider the standard hypergeometric differential operator $D$ regarded as an operator on the complex plane $C$ and the complex conjugate operator $\overline D$. These operators formally commute and are formally adjoint one to another with respect to an appropriate weight. We find conditions when they commute in the Nelson sense and write explicitly their joint spectral decomposition. It is determined by a two-dimensional counterpart of the Jacobi transform (synonyms: generalized Mehler--Fock transform, Olevskii transform). We also show that the inverse transform is an operator of spectral decomposition for a pair of commuting difference operators defined in terms of shifts in imaginary direction. | math.FA | math |
A pair of commuting hypergeometric operators
on the complex plane and bispectrality
Vladimir F. Molchanov1, Yury A. Neretin2
We consider the standard hypergeometric differential operator D regarded as an op-
erator on the complex plane C and the complex conjugate operator D. These operators
formally commute and are formally adjoint one to another with respect to an appropriate
weight. We find conditions when they commute in the Nelson sense and write explicitly
their joint spectral decomposition. It is determined by a two-dimensional counterpart of
the Jacobi transform (synonyms: generalized Mehler -- Fock transform, Olevskii transform).
We also show that the inverse transform is an operator of spectral decomposition for a
pair of commuting difference operators defined in terms of shifts in imaginary direction.
1. Introduction
1.1. Spectral problem. Denote by C the complex plane without the points 0
and 1, by D( C) the space of smooth compactly supported functions on C. Denote
by d z the standard Lebesgue measure on C.
Fix real a and b. Consider the following measure on C
(1.1)
and the corresponding space L2(C, µa,b),
µa,b(z) d z := z2a+2b−21 − z2a−2b d z
hf, gi =ZC
f (z)g(z) µa,b(z) d z.
Consider the following pair of differential operators in the space L2(C, µa,b):
(1.2)
(1.3)
D :=z(1 − z)
D :=z(1 − z)
These operators formally commute, i.e.,
DDf = DDf,
∂2
∂2
∂z2 +(cid:0)a + b − (2a + 1)z(cid:1) ∂
∂z2 +(cid:0)a + b − (2a + 1)z(cid:1) ∂
where f ∈ D( C).
∂z − a2;
∂z − a2.
A straightforward calculation shows that they are formally adjoint,
hDf, gi = hf, Dgi,
2 (D + D), 1
where f , g ∈ D( C).
Therefore the operators 1
2i (D − D) are symmetric on the domain D( C).
The purpose of this paper is to construct an explicit spectral decomposition of
this pair, i.e., a unitary operator U , which diagonalizes both operators D, D.
As we know after the famous work of Edward Nelson [33], 1959, (see, also [42],
Sect. VIII.5) a question about commutativity of two unbounded self-adjoint opera-
tors can be highly nontrivial3. Recall that two self-adjoint operators A, B commute
1Supported by the State order of the Ministry of Education and Science of the Russian Feder-
ation 3.8515.2017.
2Supported by the grants FWF Der Wissenschaftsfonds, P28421, P31591.
3The topic of the Nelson paper was finite-dimensional Lie algebras of unbounded operators in
Hilbert spaces. However, his results remain to be non-trivial for pairs of commuting operators
and even for one operator.
1
2
b
1
1
a
Figure 1. To Theorem 1.1. The domain Π of commutativity, and
the domain Πcont ⊂ Π, where the spectrum is purely continuous.
if they can be simultaneously realized as operators of multiplication by functions
in some L2. Equivalently, the corresponding one-parametric groups commute:
eisAeitB = eitBeisA,
where s, t in R.
Equivalently, resolvents (A − λ)−1 and (B − µ)−1, commute. However these prop-
erties do not follow from the identity AB = BA and are difficult for a verification.
There are some useful sufficient conditions and necessary conditions for commuta-
tivity (for necessity we use the result of Kostyuchenko and Mityagin [23]-[24]), but
quite often a question remains to be heavy4.
Define two domains Π ⊃ Πcont of the parameters (a, b):
(1.4)
Π : 0 < a + b < 2,
−1 < a − b < 1.
(1.5)
Πcont : 0 6 a 6 1,
0 6 b 6 1,
and (a, b) 6= (±1,±1), (±1,∓1).
Theorem 1.1. The operators 1
commuting self-adjoint operators if and only if (a, b) ∈ Π.
2 (D + D), 1
2i (D − D) admit extensions to a pair of
Next, we define a natural domain for our operators. Consider the subspace
Ra,b( C) ⊂ L2( C, µa,b) consisting of smooth functions f on C satisfying the following
conditions5:
1◦. In a neighborhood of z = 0 a function f has an expansion of the form6
f (z) =(α(z) + β(z)z2−2a−2b,
α(z) + β(z) lnz,
if a + b 6= 1;
if a + b = 1,
(1.6)
(1.7)
where α(z), β(z) are smooth functions.
2◦. In a neighborhood of z = 1 a function f has an expansion of the form
f (z) =(γ(z) + δ(z)z − 12b−2a,
γ(z) + δ(z) lnz − 1,
if a − b 6= 0;
if a − b = 0,
where γ(z), δ(z) are smooth.
4A famous example is a problem, see [12], which was raised by Irving Segal in 1958 and which
was discussed during almost 30 years: Let Ω be an open connected domain in Rn. Assume that
the operators i∂/∂xk in D(Ω) admit commuting self-adjoint extensions. Is it correct that Ω is
essentially a fundamental domain of Rn with respect to a certain discrete group? The answer is
affirmative.
5If (a, b) /∈ Π, then Ra,b( C) is not contained in L2(C, µa,b).
6Boundary conditions in this spirit sometimes arise in spectral theory of ordinary differential
operators D for operators with deficiency indices (1, 1) or (2, 2), see, e.g., [36], Section 1.
3◦. For each p, q, N we have
3
(1.8)
∂p+qf
∂zp∂zq = O(cid:16)z−2a−p−q (ln z)−N(cid:17)
Theorem 1.2. a) For (a, b) ∈ Π the operators 1
2 (D + D), 1
self-adjoint on Ra,b( C) and commute in the Nelson sense.
as z → ∞.
2i (D− D) are essentially
b) If (a, b) ∈ Πcont, then the spectrum of the problem
(1.9)
Df = ζf,
Df = ζf
is multiplicity free and consists of ζ having the form
(1.10)
If (a, b) ∈ Π\ Πcont, then the spectrum consists on the same set plus one eigenvalue
ζ0 > 0.
where k ∈ Z, s ∈ R.
,
2 (cid:1)2
ζ =(cid:0) k+is
Let us explain the obstacle for commutativity. Consider a second order differen-
tial operator D on an interval. For each ζ ∈ R the differential equation Df = ζf
has two solutions, and we can select generalized eigenfunctions of D as solutions
that have L2- or almost L2-asymptotics at the ends of the interval. In our case the
system (1.9) locally has 4 solutions. Furthermore, C is not simply connected, solu-
tions are ramified at 0, 1, ∞. As a result there are few single valued solutions and
we have no freedom for selection of asymptotics. Such considerations (see Section
4) allow to establish necessity of the conditions of Theorem 1.1.
Unfortunately, we do not know an a priori proof of sufficiency and obtain it as
a byproduct of the explicit joint spectral decomposition of the operators D, D.
Such detour makes our work long and requires numerous explicit calculations and
estimates.
1.2. The index hypergeometric transform. Our work is a counterpart of
the following classical topic. Consider the hypergeometric differential operator
+ a2
1
d2
dx
D := x(x + 1)
(1.11) Ia,bf (s) :=
on the half-line R+, i.e., x > 0 . Consider the integral operator
dx2 +(cid:0)(a + b) + (2a + 1)x(cid:1) d
f (x) 2F1(cid:20)a + is, a − is
(1.12) L2(cid:16)R+, xa+b+1(1 + x)a−b dx(cid:17) → L2(cid:16)R+, π−1(cid:12)(cid:12)(cid:12)
0
Then Ia,b is a unitary operator
Γ(a + b)Z ∞
a + b
; x(cid:21) xa+b−1(1 + x)a−b dx.
ds(cid:17).
Γ(a + is)Γ(b + is)
Γ(2is)
2
(cid:12)(cid:12)(cid:12)
The operator Ia,b sends D to the multiplication by s2, see [45], [43], [40], [22], [21],
[34], [37]. This transform7 is known as 'the generalized Mehler-Fock transform', 'the
Olevskii transform', or 'the Jacobi transform'.
Such operators arise in a natural way in the analysis on rank one Riemannian
symmetric spaces, on the other hand they are special cases of multi-dimensional
7A special case a = 1/2, b = 1 of this transform was discovered by Gustav Mehler in 1881, the
general transform was obtained by Weyl [45] in 1910. The Ia,b is a representative of a large family
of index integral transforms, which involve indices of hypergeometric functions, see numerous
examples in [46], [16], [38].
4
Harish-Chandra spherical transforms and more general Heckman -- Opdam [18] inte-
gral transforms, which arise as spectral decompositions of certain families of com-
muting partial differential operators.
Next, consider the following difference operator in the space of even functions
depending on the variable s:
(1.13) Lg(s) :=
(a − is)(b − is)
(−2is)(−2is + 1)(cid:0)g(s + i) − g(s)(cid:1)+
+
(a + is)(b + is)
(2is)(2is + 1) (cid:0)g(s − i) − g(s)(cid:1),
where i2 = −1. A domain of this operator is a space of even functions holo-
morphic in the strip Im s < 1 + ε with some condition of decreasing at infin-
ity.
It turns out that L is essentially self-adjoint in the space of even functions
L2
a,b sends it to the operator of mul-
tiplication by x.
ds(cid:1) and the operator I −1
even(cid:0)R,(cid:12)(cid:12) Γ(a+is)Γ(b+is)
So we have a bispectrality in the spirit of Grunbaum [17], [8]. Notice that simpler
index integral transforms as the Kontorovich -- Lebedev transform and the 1F1-Wimp
transforms also are bispectral, see [38].
(cid:12)(cid:12)2
Γ(2is)
Cherednik showed [4] that inverse Heckman -- Opdam transforms provide spectral
decompositions of families of commuting difference operators, see also van Diejen,
Emsiz [6].
1.3. Radial parts of Laplace operators. Recall one more classical topic.
Consider the usual sphere S2
R:
x2 + y2 + z2 = 1,
the orthogonal group SO(3) acts in L2(S2
R) by rotations. Recall one of possible ways
to decompose this unitary representation into irreducible components. Consider the
Beltrami -- Laplace operator ∆ on the sphere and restrict it to the space of functions
depending on the height z. We get a differential operator
Lz := (1 − z2)
∂2
∂z2 − 2z
∂
∂z
in L2[−1, 1]. Eigenfunctions of Lz are the Legendre polynomials. Simple arguments
show that the spectral decomposition of ∆ is a priori equivalent to the spectral
decomposition of Lz (the reason of this equivalence is compactness of the group
SO(2) of rotations of S2
R about the vertical axis).
Now consider the complex manifold S2
C ⊂ C3 defined by the same equation
x2 + y2 + z2 = 1. The complex orthogonal group SO(3, C) (the Lorentz group) acts
on the quadric S2
C, the action admits an SO(3, C)-invariant measure, and again we
come to a problem8 of decomposition of the unitary representation of SO(3, C) in L2
on S2
C. Now we have two Beltrami-Laplace operators, a holomorphic operator ∆ and
an antiholomorphic operator ∆. They commute in the Nelson sense. Restricting
them to functions depending on the coordinate z ∈ C we get two operators9:
Lz := (1 − z2)
∂2
∂z2 − 2z
∂
∂z
,
Lz := (1 − z2)
∂2
∂z2 − 2z
∂
∂z
.
8This problem was solved by Naimark in [30] in a completely different way.
9This pair corresponds to a = b = 1/2 in our parameters.
5
However, now the stabilizer of the point (x, y, z) = (0, 0, 1) is a noncompact sub-
group SO(2, C), and this breaks the a priori argumentation. A joint spectral de-
composition of ∆, ∆ can be reformulated as a certain problem10 for Lz, Lz, but
this is not precisely a problem of a joint spectral decomposition of Lz, Lz.
Notice that a similar separation of variables can be done for L2 on an arbitrary
rank one complex symmetric space GC/HC (and, more generally, for spaces of L2-
sections of line bundles on GC/HC). In all the cases we get pairs of hypergeometric
operators of our type. We hope that our spectral decomposition allows to write
the explicit Plancherel formula for such spaces and to give another proof of old
Naimarks's results [30] -- [32] on tensor products of representations of the Lorentz
group. However, the present paper does not have such purposes.
1.4. Homographic transformations of the operators D, D. Our next
purpose is to present the explicit joint spectral decomposition of the pair D, D. We
need some preparations.
Consider the following 8 transformations of functions on C:
f (z) 7→ γj(z)f (1 − z),
f (z) 7→ γj(z)f (z),
where
γj(z) = 1,
1 − z2(b−a),
z2(1−a−b),
z2(1−a−b)1 − z2(b−a),
cf. Erd´elyi etc., [9], Subsect. 2.6.1. It can be readily checked that these transfor-
mations send the operators D, D to operators of the same type with other values
of the parameters (a, b), as
(a, b) 7→ (b, a),
(a, b) 7→ (1 − a, b),
etc.
Thus we get all isometries of the square Π. In particular, such transformations send
spectral problems to equivalent spectral problems.
1.5. Notation. Generalized powers. Denote by C× the multiplicative group
of C. We need a notation for characters of C×. Let z ∈ C× and a, a′ ∈ C satisfy
a − a′ ∈ Z. We define a generalized power of z by
z a = zaa′
:= zaza′
= ea ln z+a′ln z = z2a z a′−a,
Denote by ΛC the set of all pairs aa′ such that a − a′ ∈ Z. Denote by Λ ⊂ ΛC the
set of all pairs
a(cid:12)(cid:12)a′ = 1
2 (k + is)(cid:12)(cid:12)(cid:12) 1
2 (−k + is),
where k ∈ Z, s ∈ R.
Equivalently, aa′ ∈ Λ if a − a′ ∈ Z, a + a′ ∈ iR. We also will use the notation
(1.15)
2 Re(a + a′).
[a] = [aa′] := 1
(1.14)
We have
in particular, for a ∈ Λ we have(cid:12)(cid:12)zaa′(cid:12)(cid:12) = 1.
(cid:12)(cid:12)zaa′(cid:12)(cid:12) = z2[aa′],
10Such reductions for families of Laplace operators were widely used by Harish-Chandra (in his
famous works on the Plancherel formula for real semisimple Lie groups) and by his successors. The
problem for Lz, Lz is more similar to decompositions of L2 on real rank one pseudo-Riemannian
symmetric spaces, which was solved by one of the authors of the present paper [27] -- [29].
6
We fix the standard Lebesgue measure edλ on the set Λ:
2 (cid:1) ds.
ϕ(λ)edλ :=Xk ZR
ϕ(cid:0) k+is
ZΛ
1.6. Hypergeometric function of the complex field. Following Gelfand,
Graev, and Retakh [13], we define the gamma function ΓC, the beta function BC,
and the hypergeometric function 2F C
1 of the complex field (see, also, Gelfand, Graev,
Vilenkin, [14], Subsect. II.3.7, and Mimachi [26]). The gamma function ΓC is
(1.16) ΓC(a) = ΓC(aa′) :=
z a−1e2i Re z d z :=
:=
za−1a′−1e2i Re z d z =
1
πZC
πZC
1
= ia−a′ Γ(a)
Γ(1 − a′)
= ia′−a Γ(a′)
Γ(1 − a)
=
ia′−a
π
Γ(a)Γ(a′) sin πa′.
The beta function BC is11
(1.17)
BC(a, b) :=
1
πZC
(1.18)
2F C
1 [a, b; c; z] = 2F C
ta−1(1 − t)b−1 d t =
ΓC(a)ΓC(b)
ΓC(a + b)
=
Γ(a)Γ(b)Γ(1 − a′ − b′)
Γ(a + b)Γ(a′)Γ(b′)
.
The hypergeometric function of the complex field is defined by
c
1haa′, bb′
1ha; b
; zi = 2F C
πBC(b, c − b)ZC
; zi :=
cc′
tb−1(1 − t)c−b−1(1 − zt)−a d t.
:=
1
Recall that the Gauss hypergeometric functions are defined by
2F1[a, b; c; z] :=
1
B(b, c − b)Z 1
0
tb−1(1 − t)c−b−1(1 − zt)−adt =
(a)p(b)p
p! (c)p
zp,
∞Xp=0
where (c)p := c(c + 1) . . . (c + p − 1) is the Pochhammer symbol. The functions
2F C
1 [a, b; c; z] admit expressions in the terms of 2F1, see Theorem 3.9.
1.7. Spectral decomposition. For (a, b) ∈ Π we define the kernel Ka,b(z, λ)
on C × Λ by
(1.19)
Ka,b(z, λ) =
1
ΓC(a + ba + b) 2F C
1 (cid:20)a + λa − λ, a − λa + λ
a + ba + b
; z(cid:21) .
Theorem 1.3. Let (a, b) ∈ Πcont. Then the operator
Ja,bf (λ) :=ZC
Ka,b(z, λ) f (z) µa,b(z) d z
is a unitary operator from L2(C, µa,b) to L2
respect to the Plancherel measure
even(Λ, κa,b) of even functions on Λ with
(1.20)
dKa,b(λ) = κa,b(λ)edλ =
1
4π2(cid:12)(cid:12)(cid:12)λ ΓC(a − λa + λ) ΓC(b + λb − λ)(cid:12)(cid:12)(cid:12)
2
edλ.
11This integral has a multi-dimensional counterpart of the Selberg type, see [7].
Next, we modify the definition of the measure for (a, b) ∈ Π \ Πcont. Due to
the homographic transformations12 it is sufficient to consider the case a < 0. We
7
define the Plancherel measure dKa,b(λ) on ΛC that is the sum of κa,b edλ and two
δ-measures located at the points ±a ± a ∈ ΛC,
(1.21)
Define a constant function v(z) on C by
ΓC(a + ba + b) ΓC(b − ab − a) ΓC(2a2a) ·(cid:0)δaa + δ−a−a(cid:1).
v(z) = ΓC(a + ba + b)−1.
For f ∈ D( C) we define an even function Ja,b(λ) on the support of dKa,b(λ) given
by the same formula (1.20) on Λ, its value at (±a ± a) is
Ja,bf (±a ± a) := hf, viL2(C,µa,b).
Theorem 1.4. If (a, b) ∈ Π and a < 0, then the operator Ja,b is unitary as an
operator L2(C, µa,b) to L2
even(ΛC, dKa,b).
Our operator really determines the spectral decomposition:
Theorem 1.5. For each (a, b) ∈ Π for any f ∈ Deven( C) we have
Ja,bf (λ).
Ja,bDf (λ) = λ2Ja,bf (λ),
Ja,bDf (λ) = λ
2
This means that
DK(z, λ) = λ2K(z, λ),
DK(z, λ) = λ
2
K(z, λ).
Next, we consider the space Deven( Λ), which consists of even smooth compactly
supported functions on Λ that are zero on a neighborhood of the point 00. The
following statement explains the appearance of the space Ra,b and also is one of the
arguments for the proofs of our main statements.
Theorem 1.6. If F ∈ Deven( Λ), then J ∗
a,bF ∈ Ra,b.
The images of δ-functions also are contained in Ra,b.
1.8. The transformation Ja,b in the complex domain. Let us extend our
kernel K to the complex domain. For
2 (cid:12)(cid:12) −k+σ
2 (cid:9) ∈ ΛC
{λλ′} =(cid:8) k+σ
we set
(1.22) Ka,b(z; λλ′) = Ka,b(z; k, σ) :=
2 (cid:12)(cid:12)a + −k+σ
1 (cid:20)a + k+σ
ΓC(a + ba + b) 2F C
:=
1
2
, a + −k−σ
2
a + ba + b
(cid:12)(cid:12)a + k−σ
2
; z(cid:21) ,
where k ranges in Z, σ ranges in C. The previous expression (1.19) corresponds to
a pure imaginary σ.
For f ∈ D( C) we define a meromorphic function on ΛC by
f (z) K(z; k, σ) dµa,b d z.
Ja,bf (k, σ) :=Z C
Theorem 1.7. For f ∈ D( C) the function Ja,bf is contained in the space Wa,b
defined as follows.
12Changing of kernels Ka,b by the homographic transformation can be observed from Propo-
sition 3.5.
8
We define a space Wa,b as the space of all meromorphic functions13 F (k, σ) on
ΛC satisfying the conditions a) -- d):
a) F is even, i.e., F (−k,−σ) = F (k, σ).
b) Possible poles of F (k, σ) are located at the points
(1.23) σ = ±(−2a + k + 2j),
A maximal possible order of a pole at a point (l, c) is a multiplicity of (l, c) in the
collection14 (1.23)
±(−2b + k + 2j),
where j = 1, 2, 3, . . .
c) For each A > 0 for each N > 0 in the union of strips Re σ < A we have an
estimate
(1.24)
F (k, σ) = O(cid:0)k2 + (Im σ)2(cid:1)−N
d) For each p, q ∈ Z
as k2 + (Im σ)2 → ∞.
(1.25)
F (p, q) = F (q, p).
Next, we extend the spectral density κa,b to the complex domain.
(1.26) κa,b(λλ′) = κa,b(k, σ) :=
:=
1
4π2 (k + σ)(k − σ) ΓC(cid:0)a + k+σ
2
2 (cid:12)(cid:12)a + −k+σ
2 (cid:12)(cid:12)b + −k+σ
ΓC(cid:0)b + k+σ
(cid:1) ΓC(cid:0)a + −k−σ
(cid:1) ΓC(cid:0)b + −k−σ
(cid:12)(cid:12)a + −k−σ
(cid:1)×
(cid:12)(cid:12)b + −k−σ
2
2
2
2
2
(cid:1).
In the case a < 0 discussed above, κa,b has a pole at k = 0, σ = a and the inner
product in L2
even(Λ, dKa,b) can be written as
hF, Gi =
1
iXk Z i∞
−i∞
F (k, σ) G(k,−σ)κa,b(k, σ) dσ+
+ 2 ress=a(cid:16)F (k, σ) G(k,−σ)κa,b(0, σ)(cid:17).
If a > 1, then the spectral density has a zero at k = 0, σ = a− 1 but both functions
F (k, σ), G(k,−σ) admit simple poles at this point, and we have a similar formula.
1.9. Difference spectral problem. It turns out that our problem is bispectral,
and the bispectrality is a crucial argument of our proof. We define analogs of the
difference operator (1.13). Consider meromorphic functions Φ depending on
and the operators in the space of meromorphic functions defined by
λλ′ = 1
2 (−k + is) ∈ ΛC
2 (k + is)(cid:12)(cid:12)(cid:12) 1
T Φ(k, s) = Φ(k + 1, s − i),
or, equivalently,
(1.27)
T Φ(λλ′) = Φ(λ + 1λ′),
eT Φ(k, s) = Φ(k + 1, s + i),
eT Φ(λλ′) = Φ(λλ′ + 1).
13We say that a function F (k, σ) is meromorphic if it is meromorphic as a function in σ for
any fixed k.
14For (a, b) ∈ Π orders of poles 6 2. Poles of order 2 arise only if a = b, a = 1, b = 1.
We define the following difference operators
9
(1.28)
L :=
(1.29)
L :=
Formally,
(a + λ)(b + λ)
2λ(1 + 2λ)
(a + λ′)(b + λ′)
2λ′(1 + 2λ′)
(T − 1) +
(eT −1 − 1) +
(a − λ)(b − λ)
−2λ(1 − 2λ)
(a − λ′)(b − λ′)
−2λ(1 − 2λ′)
(T −1 − 1);
(eT − 1).
Theorem 1.8. a) The operators 1
are essentially self-adjoint and commute in the Nelson sense.
2i (L − L) defined on the space Wa,b
LL = LL.
2 (L + L), 1
b) For Φ ∈ Ja,bD( C) we have
L Φ(z) = z J −1
J −1
a,b
(1.30)
a,b Φ(z),
J −1
a,b
L Φ(z) = z J −1
a,b Φ(z).
Thus the operator J −1
a,b determines a joint spectral decomposition of 1
2 (L + L)
and 1
2i (L − L).
1.10. The structure of the proofs. We derive asymptotics of the kernel
K(z, λ) as z → 0, 1, ∞ for fixed λ (Theorem 3.9) and as λ → ∞ for fixed z
(Theorem 7.1). Next, we prove inclusions
a,b Deven( Λ) ⊂ Ra,b,
J ∗
Ja,b D( C) ⊂ Wa,b
(Proposition 5.2 and Corollary 8.2) and symmetries
hDf, giL2(C,µa,b) = hf, DgiL2(C,µa,b),
(1.31)
where f , g ∈ Ra,b;
hLF, GiL2( Λ,dKa,b) = hF, LGiL2(Λ,dKa,b), where F , G ∈ Wa,b,
(1.32)
see Proposition 5.5 and Theorem 8.4. This implies a generalized orthogonality, i.e.,
hJ ∗
a,bF, J ∗
a,bGiL2(C,µa,b) = 0
if supports of F , G ∈ Deven( Λ) are disjoint,
and a similar statement for Ja,b, see Lemmas 9.2, 6.4. Next, we show that for any
F , G ∈ Deven( Λ) the inner products of their preimages can be written as:
hJ ∗
a,bF, J ∗
a,bGiL2(C,µa,b) = hF, GiL2(Λ,dKa,b) +ZΛ ZΛ
H(λ1, λ2)F (λ1)G(λ2)edλ1 edλ2,
where H is a locally integrable function, see Lemma 6.4. We also prove a similar
statement for Ja,b, see Lemma 9.4. Then generalized orthogonality implies H(·,·) =
0. Thus we get
(1.33)
J ∗
a,bJa,b = 1,
Ja,bJ ∗
a,b = 1,
and this is our main statement.
Some steps of this double way are straightforward, some points require long
calculations and estimates, and we meet some points of good luck (the proofs of
Theorem 8.4 and Lemma 9.4). We also need a lot of information about functions
2F C
1 (in particular, to cover the cases a + b ∈ Z and a − b ∈ Z we need a tedious
examination of possible degenerations of functions 2F C
The bispectrality allows to avoid a direct proof of completeness of the system of
1 ).
generalized eigenfunctions of D, D.
To prove the necessary conditions of self-adjointness in Theorem 1.1 we analyze
common generalized eigenfunctions of D, D for (a, b) /∈ Π and after a natural
10
selection we reduce a set of possible candidates to a finite family. This is done in
Section 4.
This text is focused to a proof of unitarity of Ja,b. An introduction to functions
1 in Section 3 can be a point of an independent interest. Also, we get two
2F C
relatively pleasant statements about asymptotic behavior of integrals
and
tα−1α′−1(ε − t)β−1β ′−1ψ(t) d t
M (ε) =ZC
I(λ) =ZC f (t)2ei Re(cid:0)λϕ(t)(cid:1) d t
as λ → ∞,
where f , ϕ are holomorphic and λ ∈ C (Theorems 2.3 and 7.2).
as ε → 0
1.11. Final remarks. The index hypergeometric transform (1.11) can be
applied as a heavy tool of theory of special functions, see [22], [35], [37]. In [39] we
use our operators Ja,b to obtain a beta integral over Λ, which is a counterpart of
the Dougall 5H5-summation formula and of the de Branges -- Wilson integral.
Also, we notice that functions, which can be regarded as higher hypergeometric
functions 4F C
3 of the complex field, arise in a natural way in the work of Ismagilov
[20] as analogs of the Racah coefficients for unitary representations of the Lorentz
group SL(2, C) (see, also a continuation in [5]).
It seems that our problem can be a representative of some family of spectral
problems, but now it is too early to claim something certainly.
2. Preliminaries. Gamma function, the Mellin transform, weak
singularities
This section is a union of 3 disjoint topics:
-- some properties of the function ΓC, which are intensively used below:
-- some properties of the Mellin transform on C, they are used in a proof of
Proposition 3.1 and in Sections 7 -- 9:
-- a lemma from asymptotic analysis, which is used only in a proof of Theo-
rem 3.9 (the last statement can be independently established by a straightforward
tiresome way):
2.1. Some properties of the gamma function. The usual functional equa-
tions for the Γ-function can be easily rewritten for ΓC (recall that a − a′ ∈ Z!):
(2.1)
(2.2)
(2.3)
(2.4)
Also,
ΓC(aa′) = ΓC(a′a);
ΓC(a + 1a′) = i a ΓC(aa′);
ΓC(aa′) ΓC(1 − a1 − a′) = (−1)a−a′
ΓC(aa′) = (−1)a−a′
ΓC(aa′).
;
m−1Yp=0
ΓC(cid:16)a +
p − 1
m (cid:12)(cid:12)(cid:12)a′ +
The identity (2.4) implies
p − 1
m (cid:17) = m1−m(a+a′) ΓC(mama′).
(2.5)
BC[aa′, bb′] = BC[aa′, bb
′
].
Let k1, k2 ∈ Z. Then
(2.6)
ΓC(k1k2) =
11
∞,
0,
ik1−k2 (k1−1)!
(−k2)! ,
ik2−k1 (k2−1)!
(−k1)! ,
if k1, k2 ∈ Z−,
if k1, k2 ∈ N,
if k1 ∈ N, k2 ∈ Z−,
if k2 ∈ N, k1 ∈ Z−,
where Z− denotes the set of integers 6 0.
The following lemma gives us the asymptotics of the Plancherel density (1.20).
Lemma 2.1. We have the following asymptotics in λ ∈ Λ:
ΓC(a − λa + λ) ΓC(b + λb − λ) ∼ λa+b−1a+b−1
(2.7)
The asymptotics is uniform in a, b if they range in a bounded domain.
as λ → ∞.
Proof. Denote Re λ = k/2. Let arg λ < π − ε. Then we write our expression
as
and apply the standard asymptotic formula Γ(z + α)/Γ(z + β) ∼ zα−β in the sector
arg z < π− ε, see Erd´elyi etc., [9], formula (1.18.4). If arg(−λ) < π− ε, we write
ikΓ(a + λ)
Γ(1 − a + λ) ·
i−kΓ(b + λ)
Γ(1 − b + λ)
i−kΓ(a − λ)
Γ(1 − a − λ) ·
ikΓ(b − λ)
Γ(1 − b − λ)
and come to the same asymptotics.
(cid:3)
2.2. The Mellin transform. Denote by C× := C \ 0 the multiplicative group
of C. The Mellin transform (see, e.g., [13]) on C× is defined by
(2.8)
g(µ) = Mf (µ) =
f (z)zµ−1 d z,
where µ = {µµ′} =(cid:8) k+is
is the Fourier transform on the group C× ≃ (R/2πZ) × R, so it is reduced to the
Fourier transforms on (R/2πZ) and on R. Indeed, changing variables
2 (cid:9) ∈ ΛC (here we allow complex s). This operator
2 −k+is
1
2πZC
we come to
g(k, s) =
1
2πZ 2π
0
The inversion formula is given by
f (z) = M−1g(z) =
z = eρeiϕ
−∞
Z ∞
f(cid:0)eρeiϕ(cid:1)eikϕ+isρ dρ dϕ.
2πZΛ
g(cid:0)µ(cid:12)(cid:12)−µ(cid:1)z−µµedµ.
1
Equivalently, M is a unitary operator L2(C×,z−2) → L2(Λ).
2.3. The Mellin transform of even functions. We say that a function
f on C× is ×-even if f (z−1) = f (z). Denote by L2
+(C×,z−2) the subspace of
L2(C×,z−2) consisting of ×-even functions. Obviously, the Mellin transform sends
×-even functions in z to even functions in µ. Also, for a ×-even function f we have
(2.9)
f (z)(zµ−1 + z−µ−1) d z,
Mf (µ) =
where f is ×-even.
1
2ZC
2.4. The Mellin transform of smooth compactly supported functions.
12
Theorem 2.2. a) Let f be a compactly supported smooth function on C. Then
Mf (µµ′) extends to a meromorphic function in the variable µ with possible poles
located at the points µµ′ ∈ Z−× Z−. Moreover, for any p, p′ ∈ Z+ for Re(µ+ µ′) >
−p − p′ we have
(2.10)
zµ−p−1µ′−p′−1 ∂p+p′
I(µµ′) =
2π(−µ)p (−µ′)p′ ZC
(−1)p+p′
∂zp ∂z p′ f (z) d z.
The residues at the poles are
(2.11)
resµµ′=−p−p′I(µµ′) =
Mf(cid:0) k+is
2 −k+is
(cid:1) = O(k2 + s2)−N
2
1
(p − 1)! (p′ − 1)!
∂p+p′
f (0, 0)
∂zp ∂z p′
.
as k2 + s2 → ∞.
b) For each N for each A for all pairs (k, s) satisfying Im s < A we have
For a proof of statement a), see Gelfand, Shilov [15], Sect. B.1.3, or equivalently
Russian edition of Gelfand, Graev, Vilenkin [14], Addendum 1.3 (the term 'Mellin
transform' in that place is absent, but the statement is proved). Formula (2.10)
is obtained from (2.8) by integration by parts. The statements about location of
poles and about residues require more careful arguments.
Proof of statement b). We pass to polar coordinates, z = eiθ and get
Mf(cid:0) k+is
2 −k+is
2
(cid:1) =
1
2πZ ∞
0 Z 2π
0
H(θ, r) r−1+iseiθk dθ dr,
where H(θ, r) := Φ(reiθ) is a smooth function 2π-periodic in θ, the H(θ, 0) does
not depend on θ, also H(θ + π,−r) = H(θ, r). Integrating by parts in r, we get
(cid:3)
If s > k we take m = 0 and large l, if k > s, we take l > Im s and large m.
2.5. Weak singularities. Here we imitate one standard trick of asymptotic
analysis, see, e.g., [10], Sect. I.4. Fix R and a smooth function ψ(t) on C. Consider
the integrals of the following type
M (ε) = Mα,β(ε) :=Zt6R
tα−1α′−1(ε − t)β−1β ′−1ψ(t) d t.
Clearly, Mα,β(ε) is holomorphic in α, β in the domain of convergence and admits
a meromorphic continuation15 to (α, β) ∈ Λ2.
15For instance, see the proof of Proposition 3.1 below.
For l > A the integral absolutely converges. Integrating by parts in θ, we get
Mf(cid:0) k+is
0
2
(−1)l
(cid:1) =
(−1)k+l
2π (is)lZ 2π
2 −k+is
Z ∞
2π (ik)m(is)lZ 2π
Mf(cid:0) k+is
2 −k+is
0
0
and
Z ∞
0
∂l
∂rl H(θ, r) r−1+is+l dr eiθk dθ.
∂l+m
∂rl ∂θm H(θ, r)r−1+is+ldreiθk dθ,
(cid:1) 6
(2πik)m(is)l
const
.
2
Theorem 2.3. Let α, β satisfy the condition
(2.12)
α, β, α + β − 1 /∈ Z− × Z−.
Then M (ε) (defined in the sense of analytic continuation) admits the following
asymptotic expansion at 0:
13
(2.13)
M (ε) ∼ Xi,i′>0
BC(α + iα′ + i′, ββ′) ·
1
∂i+i′
ψ(0, 0)
i! i′!
∂ti∂ti′
· εα+β+i−1α′+β ′+i′−1+
rjj ′ εjj ′
.
+ Xj,j ′>0
The coefficients of the expansion are meromorphic in αα′, ββ′.
If 16 [αα′] > 0, [ββ′] > 0, [αα′] + [ββ′] > 1, then
(2.14)
tα+β−2α′+β ′−2ψ(t) d t.
r00 =Zt<R
First, we prove the following lemma
Lemma 2.4. Let α, α′, β, β′ satisfy the condition (2.12). Then the following
integral (defined in the sense of analytic continuation)
(2.15)
(2.16)
admits an asymptotic expansion of the form
R(ε) =Zt<R
tα−1α′−1(ε − t)β−1β ′−1 d t
R(ε) = BC(αα′, ββ′) · εα+β−1α′+β ′−1 + Xj>0,j ′ >0
pjj ′ εjj ′
.
Moreover, the series
(2.17)
Xj>0,j ′ >0
pjj ′ εjj ′
converges in the circle ε < 1/R, and the coefficients pjj ′ (a, b) are holomorphic in
the domain (2.12).
Proof. Set
Qα,β(ε1, ε2) := (−1)β ′−β Zt>R
tα+β−2α′+β ′−2(cid:16)1 −
ε1
t (cid:17)β−1(cid:16)1 −
ε2
t (cid:17)β ′−1
d t.
This function is meromorphic in α, β, and in ε1, ε2 in the bidisk ε1 < 1/R,
Under these conditions the integral R(ε) converges, and
[ββ′] > 0,
[αα′] + [ββ′] < 1.
= BC(αα′, ββ′) · εα+β−1α′+β ′−1 − Qα,β(ε, ε).
ε2 < 1/R. Let
(2.18)
[αα′] > 0,
R(ε) =ZC −Zt>R
16Recall notation (1.15).
14
Expanding the integrand in Qα,β in a series in ε1, ε2 and integrating termwise we
come to
(2.19) Qα,β(ε1, ε2) =
=
X
(−β + 1)j (−β′ + 1)j Rα+α′+β+β ′−j−j ′
1εj ′
εj
2 .
j>0,j ′ >0: α+β−j=α′+β ′−j ′
(j + j′ − α − α′ − β − β′) j! j′!
Now we can omit restrictions (2.18).
Indeed, under conditions (2.12) the series
(2.19) converges in the bidisk ε1 < 1, ε2 < 1 and therefore its sum coincides with
the meromorphic continuation.
(cid:3)
Proof of Theorem 2.3. We expand the function ψ as a sum
ψ(t, t) = Xj+j ′ 6N
1
j! j′!
ψ(0)
∂j+j ′
∂tj ∂tj ′
tjj ′
+ HN (t),
where HN (t) is a smooth function and
HN (t) = O(tN +1)
as t → 0.
Substituting this to the initial integral we get a sum of integrals of the form (2.15),
we apply Lemma 2.4 to each summand. Also we get a summand
I(ε) =Zt6R
tα−1α′−1(ε + t)β−1β ′−1HN (t) d t.
We wish to show that T (ε) has partial derivatives at 0 up to order N − k, where k
is constant depending only on α and β. Consider a partition of unity, 1 = ϕ1 + ϕ2
such that ϕ2 is zero at some smaller circle t < R′. According to this, we split
I = I1 + I2. Obviously, I2 has an expansion of the form
I2 ∼ Xj,j ′ >0
cjj ′ εjj ′
1
with coefficients meromorphic in α, β. Next, we integrate I1 by parts several times,
I1(ε) =
(β)m(β′)mZt6R
(ε + t)β−1+m(cid:12)(cid:12)β ′+m−1 ∂t2m
Choosing m we can make β + m − 1, β′ + m − 1 as large, as we want, say > q.
Next, we choose a large N , such that
can differentiate I1(ε) with respect to ε, ε
q times at 0 and consider its Taylor
expansion. This finishes a derivation of the asymptotic expansion for R(ε).
∂m∂tm(cid:16)tα−1α′−1HN (t)ϕ1(t)(cid:17) d t.
∂tm∂tm(cid:0). . .(cid:1) is continuous at 0. Now we
∂ 2m
If the integral R(0) converges, we substitute ε = 0 to the expansion and get the
(cid:3)
expression for r00.
3. The hypergeometric function of the complex field
Here we discuss basic properties of the functions 2F C
3.1. Domain of convergence and analytic continuation. The hyperge-
1 [a, b; c; z] of the complex field is defined by the Euler type
1 [·].
ometric function 2F C
integral (1.18):
(3.1)
2F C
1 [a, b; c; z] =
1
πBC(b, c − b)ZC
tb−1(1 − t)c−b−1(1 − zt)−a d t.
15
(3.2)
For z 6= 0, 1, the integral absolutely converges (see notation (1.15) if a, b, c is
contained in the following tube bΞ,
In other words, the integral absolutely converges if and only if the point(cid:0)[a], [b], [c](cid:1)
is contained in the simplex Ξ in R3 with vertices
[c] − [b] > 0,
[c] − [a] < 2.
bΞ :
[b] > 0,
[a] < 1,
(1, 2, 2).
(1, 0, 1),
(1, 0, 0),
(−1, 0, 0),
(3.3)
We have ΛC ≃ C× Z, therefore triples (a, b, c) depend on 3 integers and 3 complex
parameters. Clearly, each component of the set Z3 × C3 has an open intersection
with the domain of convergence17.
Proposition 3.1. For z ∈ C, the expression 2F C
1 [a, b; c; z] as a function of a, b,
c admits a meromorphic extension to arbitrary values of a, b, c with poles at a
countable union of surfaces
(3.4)
a ∈ N × N, b ∈ N × N, c − a ∈ N × N, c − b ∈ N × N,
(3.5)
and vanishes for all z ∈ C at
(3.6)
c ∈ Z− × Z−
c ∈ N × N.
Proof. Consider a partition of unity 1 = ϕ0(t) + ϕ1(t) + ϕ1/z(t) + ϕ∞(t) +
ϕ∅(z), where all summands are smooth and nonnegative, ϕ0, ϕ1, ϕ1/z, ϕ∞ are
zero outside small neighborhoods of of 0, 1, 1/z, ∞ respectively, and ϕ∅ = 0 in
small neighborhoods of these points. Denote P (t, t) the integrand in the integral
representation of 2F C
πBC(b, c − b) 2F C
=Z ϕ0P d t +Z ϕ1P d t +Z ϕ1/zP d t +Z ϕ∞P d t +Z ϕ∅P d t.
1 [a, b; c; z]. Then
1 [a, b; c; z] =
The last summand is an entire function in a, b, c. By Theorem 2.2 other summands
are meromorphic and can have poles at
b ∈ Z− × Z−, c − b ∈ Z− × Z−, a ∈ N × N, c − a ∈ N × N.
However, BC-factor in the front of the integral (3.1) kills the first and the second
families of poles and produces new poles and also zeros. This gives us (3.4) -- (3.6),
in particular the factor ΓC(c) produces poles (3.5) and zeros (3.6).
All these possible poles really are poles, the simplest way to observe this is to
look at formulas (3.26) -- (3.34) derived below. Formulas (3.26) -- (3.28) show that
(3.4) are poles. To check a presence of poles (3.5) we apply (3.32) -- (3.34).
(cid:3)
3.2. Kummer symmetries. This section contains a collection of elementary
formulas, they partially depend on Theorem 3.9 proved below. However, our proof
of this theorem is based on differential equations and asymptotic analysis and is
independent of our formulas.
17The map (a, b, c) → (cid:0)[a], [b], [c](cid:1) from Λ3
→ R3 is surjective on all components.
16
(3.7)
(3.8)
First we notice two trivial identities
2F C
cc′
1 (cid:20)aa′, bb′
1 (cid:20)a′a, b′b
; z(cid:21) = 2F C
; z(cid:21) = 2F C
1 (cid:20)a′a, b′b
1 (cid:20)a′a, b
c′c
c′c
c′c
b
; z(cid:21) ;
; z(cid:21) .
′
2F C
To verify (3.7) we substitute t 7→ t to the integral (3.1).
Proposition 3.2. a) (Gauss identity) Let [c] − [a] − [b] > 0. Then
ΓC(c) ΓC(c − a − b)
(3.9)
ΓC(c − a) ΓC(c − b)
; 1(cid:21) := lim
1 (cid:20)a, b
1 (cid:20)a, b
; z(cid:21) =
2F C
2F C
z→1
c
c
.
b) Let18 [c] < 1. Then
2F C
1 [a, b; c; 0] := lim
z→0
2F C
1 [a, b; c; z] = 1.
Proof. a) We substitute z = 1 to (3.1) and come to a beta function,
πBC[b, c − a]/πBC[b, c − c].
However, this argument is valid only if the beta integral BC[b, c − a] converges.
The general statement follows from Theorem 3.9.b proved below.
b) also is reduced to a beta-function with the same problem with the domain of
(cid:3)
convergence. The general statement follows from Theorem 3.9.a.
Proposition 3.3.
(3.10)
2F C
1 [a, b; c; z] = 2F C
1 [b, a; c; z].
This will become obvious after Theorem 3.9. We use this symmetry in the next
two proofs.
Proposition 3.4. (Euler and Pfaff transformations),
(3.11)
(3.12)
(3.13)
2F C
1 (cid:20)a, b
c
2F C
; z(cid:21) = (1 − z)−a
= (1 − z)−b
= (1 − z)c−a−b
2F C
;
c
z
z − 1(cid:21)
1 (cid:20)a, c − b
1 (cid:20)c − a, b
z − 1(cid:21)
1 (cid:20)c − a, c − b
; z(cid:21)
2F C
z
c
c
;
Proof. We substitute t = 1 − s to (3.1) and get (3.11). Applying (3.10) we get
(3.12). Applying (3.11) and (3.12), we get (3.13).
(cid:3)
18If [c] > 1, then limz→0(cid:12)
1 [a, b; c; z](cid:12)
(cid:12)2F C
(cid:12) = ∞.
17
Proposition 3.5. (Kummer symmetries) The following functions uC
j (z) are equal19:
(3.14)
uC
1 (z) = 2F C
1 (cid:20)a, b
c
; z(cid:21)
(compare with [9], (2.2.9.1));
(3.15) uC
5 (z) =
=
(−1)c−a−bΓC(c) ΓC(c − 1)
ΓC(a) ΓC(b) ΓC(c − a) ΓC(c − b)
z1−c
2F C
1 (cid:20)b − c + 1; a − c + 1
2 − c
; z(cid:21)
(see [9], (2.2.9.17)) and ratio of coefficients at u1, u5 in (2.2.10.35));
(see [9], (2.2.9.9) and B1 in (2.2.10.5));
(3.16)
uC
3 (z) =
(3.17)
uC
4 (z) =
(3.18)
uC
2 (z) =
(3.19)
uC
6 (z) =
2F C
2F C
(−z)−a
a − b + 1
ΓC(c) ΓC(a − b)
ΓC(a) ΓC(c − b)
ΓC(c) ΓC(b − a)
ΓC(b) ΓC(c − a)
1 (cid:20)a, a − c + 1
1 (cid:20)b, b − c + 1
(−z)−b
1 (cid:20)
ΓC(c) ΓC(c − a − b)
ΓC(c − a)ΓC(c − b) 2F C
ΓC(c) ΓC(a + b − c)
; z−1(cid:21)
; z−1(cid:21)
; 1 − z(cid:21)
1 (cid:20) c − a, c − b
(1 − z)c−a−b
a + b + 1 − c
b − a + 1
ΓC(a)ΓC(b)
c − a − b + 1
2F C
; 1 − z(cid:21)
(see [9], (2.2.9.10) and B2 in (2.2.10.5));
a, b
(see [9], (2.2.9.5) and A1 in (2.2.10.5));
(see [9], (2.2.9.5) and A1 in (2.2.10.5)).
Remark. For each expression (3.14) -- (3.19) we can apply one of the transfor-
⊠
mations (3.11)-(3.13). In this way we get 24 expressions of this type.
Proof. The formula for u3. Changing a variable t = 1/s in (3.1) we come to
(−1)c−a−b−1z−a
sa−c(1 − s)c−b−1(1 − s/z)−a d s =
πBC(b, c − b) ZC
(−1)c−b−1πBC(a − c + 1, c − b)
πBC(b, c − b)
(−z)−a
2F C
1 (cid:20)a, a − c + 1
a − b + 1
; z−1(cid:21) .
We cancel ΓC(c − b) and apply (2.3) two times.
The formula for u4. We transpose a and b in the formula for u3.
The formula for u5. We combine the transformations (3.16) and (3.17).
The formula for u2. We combine the transformations (3.16), (3.11), and again
(3.16).
We combine transformations (3.16), (3.12), and again (3.16).
(cid:3)
Remark. Proposition 3.5 is a self-closed collection of identities. However, they
are reflections of the Kummer table of solutions of the hypergeometric equation
(cid:16)z(1 − z)
∂2
∂z2 +(cid:0)c − (a + b + 1)z(cid:1) ∂
∂z − ab(cid:17) u(z) = 0,
19The meaning of subscripts j in uC
j , references, and comments are explained in a remark after
the proof.
18
see Erd´elyi, et al., [9], Section 2.2.9, formulas (1) -- (24). The Kummer table contains
6 solutions, each of them is defined in a neighborhood of one of the singular points
0, 1, ∞.
u1(z) = α1(z),
u3(z) = (−z)−aα3(z−1),
u2(z) = α2(1 − z),
u5(z) = z1−cα5(z),
u4(z) = (−z)−bα4(z−1),
u6(z) = (1 − z)c−a−bα6(1 − z),
where αj(x) are power series, αj(0) = 1. Generally, these solutions are ramified
at the points 0, 1, ∞. Each solution is represented in 4 forms, which can be
obtained one from another by the Pfaff transformations, see Erd´elyi, et al., [9],
Sect. 2.1, (22) -- (23). In the table above we present the corresponding expressions
for 2F C
1 [a, b; c; z], they correspond to Kummer's expressions with change (a, b, c) 7→
(a, b, c). The resulting functions uC
j are non-ramified (by definition) and differ by
factors independent of z, we normalize them to make them equal one to another.
Counterparts of these factors (except one formula) are present in the Kummer
formulas as coefficients of transfer-matrices (u1, u5) to (u3, u4) and (u2, u6), with
the same replacement (a, b, c) 7→ (a, b, c), see Erd´elyi, et al., [9], display (2.2.10.5)
and the coefficients A1, A2, B1, B2. So, in each line of Proposition 3.5 we give
a reference to the corresponding formula in Erd´elyi, et al., [9], (2.2.9.1) -- (2.2.9.24)
and to the corresponding coefficient in [9], display (2.2.10.5).
⊠
3.3. Differential equations.
Lemma 3.6.
∂
∂z 2F C
∂
∂z 2F C
1haa′; bb′
1haa′; bb′
; zi =
; zi =
cc′
cc′
1ha + 1a′; b + 1b′
; zi;
1haa′ + 1; bb′ + 1
; zi.
integral of the same form. The calculation is valid if Ξ∩(cid:0)Ξ + ( 1
2 )(cid:1) 6= ∅, where
Ξ is the simplex defined by (3.2) -- (3.3). This intersection is open and nonempty. It
remains to refer to the meromorphic continuation.
(cid:3)
Proof. We differentiate the integral with respect to the parameter z, and get an
ab
c 2F C
a′b′
c′
cc′ + 1
c + 1c′
2F C
2 , 1
2 , 1
Denote
(3.20)
(3.21)
D = D[a, b, c] := z(1 − z)
∂2
D′ = D′[a′, b′, c′] := z(1 − z)
∂2
∂z2 +(cid:0)c − (a + b + 1)z(cid:1) ∂
∂z2 +(cid:0)c′ − (a′ + b′ + 1)z(cid:1) ∂
∂z − ab;
∂z − a′b′.
1 [a, b; c; z] sat-
Proposition 3.7. The complex hypergeometric function F(z) = 2F C
isfies the following system of partial differential equations
(3.22)
D[a, b, c] F = 0;
D′[a′, b′, c′] F = 0.
We call these equation by complex hypergeometric system.
Proof. This follows from the identity
(3.23) D[a, b, c](cid:16)tb−1(1 − t)c−b−1(1 − tz)−a(cid:17) = −a
∂
∂t(cid:16)tb(1 − t)c−b(1 − tz)−a−1(cid:17)
[9], (2.1.3.11)). Consider sufficiently small positive ε, δ, κ and take a, b, c
(cf.
such that
19
[bb′] = ε,
[cc′] = ε + δ,
We multiply both parts of (3.23) by t
and integrate
In the left hand side for such values of the parameter we can permute
over C.
integration and differentiation in z.
In the right hand side the integrand is an
integrable derivative of an integrable function. Therefore the right hand side is
zero.
(cid:3)
[aa′] = − 1
2 + ε + δ + κ.
(1 − t)c′−b′−1(1 − tz)−a′
b′−1
Proposition 3.8. a) Any solution of system (3.20) -- (3.21) is real analytic in z.
b) Let z0 6= 0, 1, ∞. Denote by ϕ1(z), ϕ2(z) a pair of independent holomorphic
solutions of the ordinary differential equation D[a, b, c] f (z) = 0 at a neighborhood
of z0. Denote by ψ1(z), ψ2(z) a pair of antiholomorphic solutions of the ordinary
differential equation D′[a′, b′, c′] f (z) = 0. Then any solution of the system (3.20) --
(3.21) can be represented as
(3.24)
τij (a, b, c)ϕi(z)ψj(z).
Xi,j=1,2
c) If we choose ϕi, ψj meromorphic in the parameters a, b, c in some domain
in Λ3
C, then the coefficients τij also are meromorphic in the parameters a, b, c.
Proof. a) Indeed, D[a, b, c] is an elliptic differential operator, therefore solutions
of the equation DF = 0 are analytic functions, see, e.g., [19],Theorem 9.5.1.
b) Consider a solution
2F C
1 [a, b; c; z] = h00 + h10(z − z0) + h01(z − z0) + h11(z − z0)(z − z0) + . . .
of the system of partial differential equations (3.22). These equations determine
recurrence relations for the Taylor coefficients hij of 2F C
1 [. . . ] at z0. It can be easily
checked that all the coefficients hij admit linear expressions in terms of h00, h01,
h10, h11. On the other hand, for given h00, h01, h10, h11, we can find a local solution
of the complex hypergeometric system (3.22) in the formP Cij ϕi(z)ψj(z).
c) By Lemma 3.6, the coefficients h00, h10, h01, h11 depend on a, b, c meromor-
j(z0) are meromorphic in the parameters, then
(cid:3)
phically. If ϕi(z0), ϕ′
the coefficients Cij also are meromorphic.
i(z0), ψj(z0), ψ′
3.4. Expressions for 2F C
1 . Let us write expansions of 2F C
1 [. . . ; z] near the
singular points z = 0, 1, ∞. Explicit formulas for fundamental systems of solutions
of the hypergeometric differential equation are well-known, see Erd´elyi, et al., [9],
2.9 (the Kummer series). For definiteness, consider z0 = 0.
If c /∈ Z, then for
generic values of the parameters the hypergeometric equation D[a, b, c]f (z) = 0 has
two holomorphic (ramified) solutions on a punctured neighborhood of 0,
ϕ1(z) = 2F1[a, b; c; z],
ϕ2(z) = z1−cF [a + 1 − c, b + 1 − c; 2 − c; z].
The equation D′[a′, b′, c′]f (z) = 0 has two antiholomorphic solutions
ψ1(z) = 2F1[a′, b′; c′; z],
ψ2(z) = z1−c′
F [a′ + 1 − c′, b′ + 1 − c′; 2 − c′; z].
Therefore near z = 0 we have solutions of system (3.20) -- (3.21) of the same form
(3.24) with new ϕ, ψ. We get a family of functions depending of 4 parameters τij,
therefore for generic a, b, c this formula gives all multivalued solutions near z = 0.
20
Solutions (3.24) that are single valued in a neighborhood of 0 have the form
(3.25)
Theorem 3.9. a) In the disk z < 1 we have the following expansion:
A1ϕ1(z)ψ1(z) + A2ϕ2(z)ψ2(z).
c
2F C
1ha; b
+ A1 · z1−c1−c′
; zi = A0 · 2F1ha, b
; zi 2F1ha′, b′
2F1ha + 1 − c, b + 1 − c
; zi+
; zi 2F1ha′ + 1 − c′, b′ + 1 − c′
2 − c′
2 − c
c′
c
; zi,
A0 = 1,
A1 = (−1)c−a−b
ΓC(c) ΓC(c − 1)
.
ΓC(a) ΓC(b) ΓC(c − a) ΓC(c − b)
b) In the disk z − 1 < 1 the following expansion holds:
(3.29)
2F C
1ha; b
; zi = B0 · 2F1h
+ B1 · (1 − z)c−a−bc′−a′−b′
c
a, b
a + b + 1 − c
2F1h c − a, c − b
c + 1 − a − b
a′, b′
; 1 − zi2F1h
; 1 − zi2F1h c′ − a′, c′ − b′
a′ + b′ + 1 − c′; 1 − zi+
c′ + 1 − a′ − b′; 1 − zi,
(3.26)
where
(3.27)
(3.28)
where
(3.30)
(3.31)
B0 =
B1 =
ΓC(c) ΓC(c − a − b)
ΓC(c − a)ΓC(c − b)
ΓC(c)ΓC(a + b − c)
ΓC(a) ΓC(b)
,
.
c) In the disk z > 1 the following expansion holds:
(3.32)
2F C
1ha; b
c
where
(3.33)
(3.34)
; zi = C0 · (−z)−a−a′
+ C1 · (−z)−b−b′
a + 1 − b
2F1ha, a + 1 − c
2F1hb, b + 1 − c
b + 1 − a
; z−1i2F1ha′, a′ + 1 − c′
; z−1i2F1hb′, b′ + 1 − c′
b′ + 1 − a′
a′ + 1 − b′
; z −1i+
; z −1i,
C0 :=
C1 :=
ΓC(c) ΓC(b − a)
ΓC(c − a) ΓC(b)
ΓC(c) ΓC(a − b)
ΓC(c − b) ΓC(a)
,
.
3.5. Proof of Theorem 3.9. Forms (3.26), (3.29), (3.32) for the desired expres-
sions follow from the preceding considerations. Also we know that the coefficients
A0, A1, B0, B1, C0, C1 are meromorphic in a, b, c. Now we apply asymptotic
expansions from Theorem 2.3.
1. Asymptotic of 2F C
1 [a, b; c; z] as z → ∞. Assume that the defining integral
1 [a, b; c; z] converges, and also
for 2F C
(3.35)
[b] − [a] > 0
Then
21
1
πBC(b, c − b)ZC
tb−1(1 − t)c−b−1(1 − zt)−a d t =
πBC(b, c − b)ZC
z−a
=
∼
BC(b − a, c − b)
BC(b, c − b)
tb−1(1 − t)c−b−1(z−1 − t)−a d t ∼
pii′ z−i−i′(cid:17)+
· (−z)−a ·(cid:16)1 + X(i,i′)6=(0,0)
BC(b, c − b) · z−b ·(cid:16)1 + X(i,i′)6=(0,0)
BC(b, 1 − a)
+
qii′ z−i−i′(cid:17).
Precisely, denote z−1 by ε, and denote the integrand in the last integral by H(·).
Let ϕ(t) > 0, ψ(t) > 0 be smooth functions such that ϕ(t) + ψ(t) = 1, ϕ(t) = 1
near 0, and ψ(t) = 1 near ∞. A straightforward differentiation with respect to the
parameter ε shows that
is smooth near ε = 0. For
Z H(t; ε)ψ(t) d t
ZC
H(t; z)ϕ(t; ε) d t
we apply Theorem 2.13, due to the restriction (3.35) we can also apply (2.14). Thus
we get explicit coefficients C0, C1 in the expansion (3.32). To remove restrictions
for the parameters, we refer to the analytic continuation.
Finally, we transform BC(b, 1 − a) with formula (2.3),
BC(b, 1 − a) =
ΓC(b) ΓC(1 − a)
ΓC(1 + b − a)
= (−1)b ΓC(b)ΓC(a − b)
ΓC(a)
.
2. Asymptotic of 2F C
(1.18) of 2F C
1 , we get
1 [a, b; c; z] as z → 0. Substituting t = 1/s to the definition
2F C
1 [a, b; c; z] =
(−1)c−a−b
πBC(b, c − b)ZC
∼
(−1)c−a−bBC(a − c + 1, 1 − a)
BC(b, c − b)
· z 1−c ·(cid:16)1 + X(i,i′)6=(0,0)
s−c+a(z − s)−a(1 − s)c−b−1 d s ∼
pii′ zii′(cid:17)+
·(cid:16)1 + X(i,i′)6=(0,0)
BC(b, c − b)
qii′ zii′(cid:17).
(−1)c−bBC(1 − c, c − b)
+
22
get
3. Asymptotic of 2F C
2F C
1 [a, b; c; z] =
(−1)c−bBC(c − b, 1 − a)
∼
BC(b, c − b)
1 [a, b; c; z] as z → 1. We substitute t = 1
πBC(b, c − b)ZC
sc−b−1(1 − s)a−c(1 − z − s)−a d s ∼
(−1)c−b
1−s to (1.18) and
· (1 − z)c−b−a ·(cid:16)1 + X(i,i′)6=(0,0)
pii′ (1 − z)ii′(cid:17)+
qii′ zjj ′(cid:17).
·(cid:16)1 + X(j,j ′)6=(0,0)
(−1)c−b−aBC(c − b − a, 1 + a − c)
+
BC(b, c − b)
Remark. Another way of a proof of Theorem 3.9. Applying the Kum-
mer formulas, Erd´elyi, et al., [9], Section 2.9, we can write the analytic continuation
of (3.25) to a neighborhood of this point. The resulting expression for 2F C
1 must
be non-ramified at z = 1. This gives us the coefficients in (3.25) up to a common
factor. In fact this calculation is done below in the proof of Proposition 3.11. The
scalar factor can be evaluated using (3.9). It remains to apply the Kummer formu-
las ([9], Section 2.9) for the analytic continuation again and to get an expansion at
∞.
⊠
3.6. Additional symmetry.
Proposition 3.10. Let a − b ∈ Z. Then
(3.36)
2F C
1 (cid:20)aa′, bb′
cc′
; z(cid:21) = 2F C
1 (cid:20)ab′, ba′
cc′
; z(cid:21) .
Proof. The expansions (3.26) -- (3.28) at 0 for both functions are identical. We
only must verify the equality of the denominators in (3.28):
(3.37) ΓC(aa′) ΓC(bb′) ΓC(c − ac′ − a′) ΓC(c − bc′ − b′) =
ΓC(ab′) ΓC(ba′) ΓC(c − ac′ − b′) ΓC(c − bc′ − a′).
The both sides are equal to
(−1)c′−cπ4Γ(a)Γ(a′)Γ(b)Γ(b′)Γ(c − a)Γ(c − a′)Γ(c − b)Γ(c − b′)
sin πa′ sin πb′ sin π(c′ − a′) sin π(c′ − b′)
.
(cid:3)
3.7. Degenerations and logarithmic expressions.
a) Residues and zeros. Notice that poles and zeros of 2F C
1 [a, b; c; z] as func-
tion of a, b, c depend on a choice of a normalizing factor in the front of the integral
(3.1).
It is easy to see that residues at poles also are solutions of the complex hyperge-
ometric system (3.22). The expressions for the residues can be obtained from our
expansions.
For obtaining the residues at {aa′} ∈ N × N we can use the expansion of 2F C
1
at z = 0, see (3.26) -- (3.28). We get
z1−c1−c′
2F1ha + 1 − c, b + 1 − c
2 − c
; zi 2F1ha′ + 1 − c′, b′ + 1 − c′
2 − c′
; zi
with an obvious ΓC-factor. Applying the Pfaff transformations of 2F1, we observe
that these expressions are elementary functions. Formulas (3.26) -- (3.28) allow to
calculate residues at the poles of all the types (3.4).
Next, consider another normalization20 of the functions 2F C
1 :
23
(3.38)
2eF C
1 [a, b; c; z] :=
1
ΓC(c) 2F C
1 [a, b; c; z].
This operation cancels the factor ΓC(c) in expansion of 2F C
1 [z] at ∞, see (3.32) --
(3.34). So we get a finite expression at the poles (3.5) and non-zero function at the
zeros (3.6).
Thus, at all exceptional planes (3.4) -- (3.6) we get explicit nonzero expressions.
Such expressions also depend on normalization of 2F C
1 [. . . , z], but for a point
(a0, b0, c0) being in a general position on an exceptional plane such nonzero ex-
pression is canonically defined up to a constant factor.
b) Further degenerations. Classical hypergeometric differential equation
has a sophisticated list of degenerations, see [9], Sect. 2.2. In our case new diffi-
culties arise if at least two of the parameters a, b, c − a, c − b are contained in
Z × Z. We stop here further analysis and only notice that for exceptional values
of the parameters a solution of the complex hypergeometric system (3.22) can be
non-unique.
For instance, if a ∈ Z− × Z−, c − b ∈ N × N, then both summands in (3.25) are
single-valued (since all hypergeometric series are terminating).
c) Logarithmic expressions. For definiteness we discuss the case
c ∈ N × N
(which is interesting for our further purposes). Consider the function 2eF C
1 defined
by (3.38). It has a removable singularity at our c. Recall that for c = n ∈ N the
usual hypergeometric differential equation D[a, b, n]f = 0 has two solutions. The
first is 2F1[a, b; c; z] and the second has the form
(3.39)
Ψ[a, b; n; z] =
∞Xj=−n+1
pjzj + ln z · 2F1[a, b; n; z],
where pj are explicit coefficients, p−n+1 6= 0, and this form does not depend on
further degenerations, see [1], Section 2.3. Passing around 0 we get
Thus the system
Ψ[a, b; n; eiϕz](cid:12)(cid:12)(cid:12)ϕ=2π − Ψ[a, b; n; z] = 2F1[a, b; n; z].
D[a, b; n]F = 0,
D[a′, b′; n′]F = 0
has two solutions that are single-valued near zero, the first is obvious
2F1[a, b, n; z] 2F1[a′, b′; n′; z],
and the second is
(3.40)
2F1[a, b; n; z] Ψ[a′, b′; n′; z] + Ψ[a, b; n; z] 2F1[a′, b′; n′; z].
1 [a, b; n; z] is certain linear combination of these solutions.
d) On uniqueness of a solution of the hypergeometric system.
Our function 2eF C
20In fact, in the main part of our work we use this normalization of the kernel, see (1.19). Due
to this we do not lose the case of L2 on the complex quadric discussed in Subsect. 1.3.
24
Proposition 3.11. Let
a, b, c, c − a − b, c − a, c − b /∈ Z,
a′, b′, c′, c′ − a′ − b′, c′ − a′, c′ − b′ /∈ Z.
Let the system D[a, b, c] F = 0, D′[a′, b′, c′] F = 0 have a non-ramified non-zero
solution. Then c − c′ ∈ Z and
(3.41)
a − a′, b − b′ ∈ Z or a − b′, b − a′ ∈ Z
Such solution is unique up to a scalar factor and therefore is 2F C
or 2F C
1 [ab′, ba′; cc′; z]
1 [aa′, bb′; cc′; z]
Proof. First, we examine the behavior of a solution near z = 0. Let
ϕ(z) := 2F1(cid:20)a, b
c
; z(cid:21) ,
ψ(z) := 2F1(cid:20)a + 1 − c, b + 1 − c
2 − c
; z(cid:21) ,
i.e., ϕ, z1−cψ are the Kummer solutions of the hypergeometric equation D[a, b, c]f =
0 at 0, see [9], (2.9.1), (2.9.17). Denote by ´ϕ(z), ´ψ(z) the similar functions obtained
be the change a 7→ a′, b 7→ b′, c 7→ c′, z 7→ z. A solution of our system near 0 has
the form
G(z) = σ ϕ(z) ´ϕ(z) + µ z1−c′
ϕ(z) ´ψ(z) + ν z1−cψ(z) ´ϕ(z) + τ z1−c1−c ′
ψ(z) ´ψ(z).
Passing m times around 0 we come to
G(z) = σ ϕ(z) ´ϕ(z) + µe2πmc′i z1−c′
ϕ(z) ´ψ(z)+
+ νe−2πmci z1−cψ(z) ´ϕ(z) + τ e2πm(c′−c)i z1−c1−c ′
ψ(z) ´ψ(z).
Since c, c′ /∈ Z, we have e2πmc′i, e−2πmci 6= 1, on the other hand they are 6=
e2πm(c′−c)i.
If G(z) is single-valued, then µ = ν = 0. Also, we need τ = 0 or
c − c′ ∈ Z.
To examine the behavior of G near z = 1 we apply a formula for analytic
continuation, see [9], Subsect. 2.10. Near z = 1 we have
(3.42) F(cid:20)a, b
c
; z(cid:21) = A1(a, b, c) 2F1(cid:20)
a, b
a + b − c + 1
; 1 − z(cid:21) +
+ A2(a, b, c) (1 − z)c−a−b
2F1(cid:20) c − a, c − b
c − a − b + 1
; 1 − z(cid:21) ,
where
(3.43)
A1(a, b, c) :=
Γ(c)Γ(c − a − b)
Γ(c − a)Γ(c − b)
,
A2(a, b, c) :=
Γ(c)Γ(a + b − c)
Γ(a)Γ(b)
.
Since c − a − b, c′ − a′ − b′ /∈ Z, the expression ϕ(z) ´ϕ(z) is not single valued.
Thus τ 6= 0, c − c′ ∈ Z, and
G(z) = σ ϕ(z) ´ϕ(z) + τ z1−c1−c ′
ψ(z) ´ψ(z).
Applying for ϕ, ´ϕ, ψ, ´ψ formula (3.42) and the identity
2F1(α, β; γ; z) = (1 − z)γ−α−β
2F1(γ − α, γ − β; γ; z),
we come to
25
G(z) = σA(a, b, c)A(a′, b′, c′) 2F1(cid:20)a, b
+nσA1(a, b, c)A2(a′, b′, c′)+ τ A1(a+1 − c, b+1 − c, 2 − c)A2(a′ +1 − c′, b′ +1 − c′, 2 − c′)o×
; 1 − z(cid:21) 2F1(cid:20)a′, b′
; 1 − z(cid:21) +
c′
c
× (1 − z)c′−a′−b′
2F1(cid:20)a, b
c
; 1 − z(cid:21) 2F1(cid:20) c′
c′
− a′, c′
− a′
− b′
− b′ + 1
; 1 − z(cid:21) +
+nσA2(a, b, c)A1(a′, b′, c′)+ τ A2(a+1 − c, b+1 − c, 2 − c)A1(a′ +1 − c′, b′ +1 − c′, 2 − c′)o×
× (1 − z)c−a−b
2F1(cid:20) c − a, c − b
c − a − b + 1
; 1 − z(cid:21) 2F1(cid:20)a′, b′
c′
; 1 − z(cid:21) +
+ A2(a + 1 − c, b + 1 − c, 2 − c)A2(a′ + 1 − c′, b′ + 1 − c′, 2 − c′)×
× (1 − z)c−a−bc′−a′−b′
2F1(cid:20) c − a, c − b
; 1 − z(cid:21)
The coefficients A1(·), A2(·) have no zeros and no poles under our restrictions.
The expression is single-valued if and if two curly brackets are zero and (c − a −
b) − (c′ − a′ − b′) ∈ Z. This implies
; 1 − z(cid:21) 2F1(cid:20) c′ − a′, c′ − b′
c − a − b + 1
− b′ + 1
− a′
c′
(a + b) − (a′ + b′) ∈ Z.
Two curly brackets give a system of linear equations for σ, τ . It has a nonzero
solution if and only if its determinant ∆ is zero. Straightforward calculations give
∆ = π−4Γ(c)Γ(c′)Γ(2−c)Γ(2−c′)Γ(c−a−b)Γ(c′−a′−b′)Γ(a+b−c)Γ(a′+b′−c′)·Ξ,
where
Ξ = sin π(c − a) sin π(c − b) sin πa′ sin πb′ − sin π(c′ − a′) sin π(c′ − b′) sin πa sin πb.
Clearly, the set Ξ(a, b, c, a′, b′, c′) = 0 is invariant with respect to the shifts a 7→ a+1,
b 7→ b + 1, c 7→ c + 1. Therefore to examine the set of zeros we can assume c′ = c,
b′ = a + b − a′. Under these conditions Ξ can be reduced to the following form:
Ξ(a, b, c, a′, b′, c′) = sin π(a − a′) sin π(a′ − b) sin πc sin π(c − a − b)
(this non-obvious identity can be verified by decompositions of both sides into sums
if exponentials). This implies (3.41).
If ∆ = 0 then σ, τ are defined up to a common scalar factor, this proves the
(cid:3)
uniqueness (and gives an expression for σ/τ ).
e) Non-interesting solutions. However, we have seen that the complex
hypergeometric system for some values of the parameters has two single-valued
solutions. Also, there are solutions that do not seem reasonable. For instance, we
have
D[0, b1, c1] · 1 = 0,
D′[0, b2, c2] · 1 = 0
for arbitrary b1, c1, b2, c2 ∈ C.
3.8. Differential-difference equations for 2F C
1 . We can regard 2F1[a, b; c; z]
as a family of functions of a complex variable z depending on 3 parameters a, b, c.
But we also can regard 2F1[a, b; c; z] as one function of the four complex variables
a, b, c, z. Then 2F1[a, b; c; z] satisfy a non-obvious system of linear differential-
difference equations, some examples of such equations are in Erd´elyi, et al., [9],
(2.8.20-45). Below we show that such equations can be automatically transformed
26
to differential-difference equations for the function 2F C
variables.
1 [aa′, bb′; cc′; z] of 7 complex
Consider a space of functions in the variables a, b, c, z. Define operators
Taf (a, b, c, z) = f (a + 1, b, c, z),
Tcf (a, b, c, z) = f (a, b, c + 1, z).
Consider finite sums of the form
Tbf (a, b, c, z) = f (a, b + 1, c, z),
(3.44)
L =Xj>0 Xk,l,m∈Z
Uj,k,l,m(a, b, c, z) T k
a T l
b T m
c
∂j
∂zj ,
where Uj,k,l,m(a, b, c, z) are polynomial expressions in z with coefficients rationally
depending on a, b, c.
Assume that
L 2F1[a, b; c; z] = 0.
We can regard an operator (3.44) as an operator on functions f (aa′, bb′, cc′, z) on
Λ3 × C. We also define operators
For such an operator L we define the operator L′ by
Ta′ f (aa′, bb′, cc′, z) = f (aa′ + 1, bb′, cc′, z),
Tb′ f (aa′, bb′, cc′, z) = f (aa′, bb′ + 1, cc′, z),
Tc′f (aa′, bb′, cc′, z) = f (aa′, bb′, cc′ + 1, z).
L′ =Xj>0 Xk,l,m∈Z
Uj,k,l,m(a′, b′, c′, z) T k
b′ T m
c′
a′ T l
∂j
∂zj .
From the definition it follows that
LL′ = L′L.
Proposition 3.12. Let the function Q(a, b, c, z) = 2F1[a, b; c; z] satisfy an equation
L Q = 0. Then the function
R(aa′, bb′, cc′, z) := 2F C
1 [aa′, bb′, cc′, z]
satisfies the system of equations
(3.45)
L′ R(aa′, bb′, cc′, z) = 0.
Lemma 3.13. Let Q = 2F1[a, b; c; z] satisfy an equation LQ = 0. Then
L R(aa′, bb′, cc′, z) = 0,
(3.46)
eπi(c−a−b)Γ(c)Γ(c − 1)
Γ(a)Γ(b)Γ(c − a)Γ(c − b)
satisfies the same equation.
z1−c
2F1[a + 1 − c, b + 1 − c; 2 − c, z]
Remark. The same statement holds for the functions
(3.47)
(3.48)
u1 =
u2 =
u3 =
Γ(c − a)Γ(c − a − b)
Γ(c)Γ(c − b)
Γ(c)Γ(a + b − c)
Γ(a)Γ(b)
Γ(c)Γ(b − a)
Γ(b)Γ(c − b)
z−a
a, b
a + b − c + 1
2F1(cid:20)
(1 − z)c−a−b
2F1(cid:20)a, 1 − c + a
; 1 − z(cid:21) ;
2F1(cid:20) c − a, c − b
; z−1(cid:21)
c − a − b + 1
1 − b + a
; 1 − z(cid:21) ;
27
and also for other summands in the right-hand sides of formulas Erd´elyi, et al. [9],
(2.10.1) -- (2.10.4).
(cid:3)
Proof of Lemma 3.13. First, let a, b, c be in a general position. By Erd´elyi,
et al. [9], (2.10.1), (2.10.5),
(3.49)
F (a, b; c; z) = u1 + u2,
where u1, u2 are given by (3.47) -- (3.48). The function u2 is ramified at z = 1.
Passing around this point we get a function
eF := u1 + e2πi(c−a−b)u2.
By analytic continuation, LeF = 0. The factor e2πi(c−a−b) does not change under
the shifts Ta, Tb, Tc. Therefore the summands u1, u2 satisfy the same equation,
Lu1 = 0, Lu2 = 0. We apply the same transformation (3.49) to the summand u1
and repeat the same reasoning. We observe that
πΓ(c)Γ(c − 1)
Γ(a)Γ(b)Γ(c − a)Γ(c − b) sin π(a + b − c)
z1−c
2F1[a + 1 − c, b + 1 − c; 2 − c, z]
satisfies the same equation. This expression differs from (3.49) by the factor
eiπ(a+b−c) sin π(a + b − c), which is invariant under the shifts Ta, Tb, Tc.
Passing to a limit we omit restrictions to a, b, c.
Proof of Proposition 3.12. We use the expression (3.26) for 2F C
1 [a, b; c; z].
Obviously, the first summand satisfies the system (3.45). By Lemma 3.13, the
expression
(cid:3)
eπi(c−a−b)Γ(c)Γ(c − 1)
Γ(a)Γ(b)Γ(c − a)Γ(c − b)
z1−c
2F1(cid:20)a + 1 − c, b + 1 − c
; z(cid:21)×
×
eπi(c′−a′−b′)Γ(c′)Γ(c′ − 1)
Γ(a′)Γ(b′)Γ(c′ − a′)Γ(c′ − b′)
2 − c
z1−c′
2F1(cid:20)a′ + 1 − c′, b′ + 1′ − c
2 − c′
; z(cid:21) .
satisfies the system (3.45). It differs from the second summand in (3.26) by a factor
i0 sin πa′ sin πb′ sin π(c′ − a′) sin π(c′ − b′)
sin πc′ sin π(c′ − 1)
.
This expression is invariant with respect to shifts Ta, Ta′, . . . . Therefore the second
summand in (3.26) also satisfies the system.
(cid:3)
3.9. One difference operator. By [34], formula (2.3), the Gauss hypergeo-
metric function 2F1[p, q; r; z] satisfies the following difference equation
(3.50) − z · 2F1(p, q; r; z) =
q(r − p)
=
(q − p)(1 + q − p) 2F1(p − 1, q + 1; r; z)−
q(r − p)
p(r − q)
+
−h
(q − p)(1 + q − p)
(p − q)(1 + p − q)i2F1(p, q; r; z)+
(p − q)(1 + p − q) 2F1(p + 1, q − 1; r; z).
p(r − q)
+
28
Define the difference operators L, L′ acting on functions of the variables a, b, c,
z by
(3.51)
L =
(3.52) L′ =
b(c − a)
(b − a)(1 + b − a)(cid:0)T −1
(b′ − a′)(1 + b′ − a′)(cid:0)T −1
a Tb − 1(cid:1) −
a′ Tb′ −1(cid:1)−
b′(c′ − a′)
a(c − b)
(a − b)(1 + a − b)(cid:0)TaT −1
b − 1(cid:1);
b′ −1(cid:1).
(a′ − b′)(1 + a′ − b′)(cid:0)Ta′T −1
a′(c′ − b′)
Corollary 3.14. The complex hypergeometric function 2F C
following system of difference equations
1 [a, b; c; z] satisfies the
(3.53)
(3.54)
L 2F C
L′
2F C
1 [a, b; c; z] = z 2F C
1 [a, b; c; z] = z 2F C
1 [a, b; c; z];
1 [a, b; c; z].
3.10. Some properties of the kernel K. We have the following corollaries
from our previous considerations.
1) By (3.10) Ka,b is even,
(3.55)
2) By (3.8),
(3.56)
Ka,b(z;−k,−σ) = Ka,b(z; k, σ).
Ka,b(z; k,−σ) = Ka,b(z; k, σ).
In particular, Ka,b(z; k, σ) is real on Λ.
3) By Proposition 3.7, Ka,b(z; k, σ) satisfies the following differential equations:
(3.57)
(3.58)
D Ka,b(z; k, σ) = 1
D Ka,b(z; k, σ) = 1
4 (k + σ)2 Ka,b(z; k, σ);
4 (k − σ)2 Ka,b(z; k, σ).
4) By Corollary 3.14, Ka,b(z; k, σ) satisfies the following difference equations:
(3.59)
(3.60)
L Ka,b(z; k, σ) = z Ka,b(z; k, σ);
L Ka,b(z; k, σ) = z Ka,b(z; k, σ).
4. Nonexistence of commuting self-adjoint extensions
Here we prove that for (a, b) /∈ Π the operators 1
2i (D − D) defined
on D( C) do not admit commuting self-adjoint extensions. We analyze the set of
possible generalized eigenfunctions and show that this set is too small.
2 (D + D), 1
4.1. Generalized eigenfunctions. Denote by D′( C) the space of distributions
on C. We have a nuclear rigging (see [2], Section 14.2)
D( C) ⊂ L2(C, µa,b) ⊂ D′( C),
and apply the usual formalism of generalized eigenfunctions, see [2], Chapter 15.
Recall that we have formally symmetric and formally commuting operators
D+ := 1
2 (D + D),
D− := 1
2i (D − D)
in L2(C, µ) (defined on the domain D( C)) and the spectral problem
(4.1)
DΦ = ζΦ,
DΦ = ζΦ.
29
Suppose that the operators D+, D− admit commuting self-adjoint extensions.
Then the operator U of spectral decomposition can be written in terms of gener-
alized eigenfunctions. Precisely, there exist a space R equipped with a measure ρ
and an injective measurable map r 7→ ϕr from R to D′( C) such that
D+ϕr = a(r)ϕr,
D−ϕr = b(r)ϕr,
where a(r), b(r) are real-valued functions, and the pairing
U f (r) = {f, ϕr}
of f ∈ D( C) and ϕr determines a unitary operator L2(C, µ) → L2(R, ρ), see text-
book [2], Subsect. 15.2.321).
Since the operator D is elliptic, generalized eigenfunctions are smooth on C, see
e.g., [2], Theorem 16.2.1. Therefore in our case generalized eigenfunctions ϕr are
usual smooth solutions of the system of differential equations.
We also can identify the measure space R with its image, and so we can think
that the measure ρ is sitting on the space Ω of smooth solutions of the systems (4.1),
where ζ ranges in C. We intend to show that for any measure ρ on Ω the operator
J : L2(Ω, ρ) → L2(C, µa,b) defined by
U h(z) =ZΩ
h(r)ϕr(z) dρ(r)
is not unitary. Precisely:
Lemma 4.1. Let (a, b) /∈ Π. Let ρ be a measure on Ω, and let the corresponding
operator U be bounded. Then ρ is an atomic measure supported by a finite set.
The idea of a proof is simple, it is explained in the next subsection, a formal
proof is completed in Subsect. 4.3.
Lemma 4.1 implies that for (a, b) /∈ Π the operators D+, D− have no commuting
self-adjoint extensions.
4.2. Almost proof of Lemma 4.1. For ζ being in a general position, the
1 [·; z]. Denote by Ωhyp
1 [·; z]. We wish to prove the following
system (4.1) has a unique solution, and it has the form 2F C
the subset of Ω consisting of the functions 2F C
statement:
Lemma 4.2. Let (a, b) /∈ Π and a + b, a − b, a, b /∈ Z. Let ρ be a measure on Ω,
and let the corresponding operator U be bounded. Then ρ is atomic on Ωhyp.
Proof. Set ζ = λ2. Then a hypergeometric solution of the system (4.1) has
one of the two forms:
2F C
1 (cid:20)a + λa − λ, a − λa + λ
a + ba + b
, z(cid:21) ,
2F C
1 (cid:20)a + λa + λ, a − λa − λ
, z(cid:21) .
a + ba + b
In the first case we have (a + λ) − (a− λ) = 2 Re λ ∈ Z, hence λ = 1
k ∈ Z, s ∈ R. We come to the functions Ka,b(z; k, is).
, z(cid:21) .
1 (cid:20)a + τa + τ, a − τa − τ
K(z; 0, τ ) = 2F C
a + ba + b
In the second case we have λ− λ ∈ Z, i.e., λ = τ ∈ R. We come to the functions
2 (k + is), where
(4.2)
21Basically, this is a result of Kostyuchenko and Mityagin [23] -- [24] with weaker conditions for
a rigging.
30
Next, we will show that
(4.3)
the measure ρ is zero on the set of all λ = 1
2 (k + is) with s 6= 0.
(4.4)
(4.5)
Our kernel has the following asymptotics at z = 0 and z = 1:
K(z; k, is) =(cid:0)1 + O(z)(cid:1) + B(k, is)z2−2a−2b(cid:0)1 + O(z)(cid:1)
K(z; k, is) =C(k, is)(cid:0)1 + O(1 − z)(cid:1)+
+ D(k, is)1 − z2b−2a(cid:0)1 + O(1 − z)(cid:1)
where the coefficients B, C, D are continuous non-vanishing functions on Λ and all
O(·) are uniform on compact subsets of Λ (see formulas (3.26) -- (3.31)).
For definiteness, assume that a + b > 2. Consider a point (k0, is0) ∈ Λ, s0 6= 0
and a neighborhood N of (k0, is0). Assume that ρ(N) > 0. Denote by IN the
indicator function of the set N. The function U IN has the following asymptotics
at z = 0:
as z → 0,
as z → 1,
as z → 0.
Due to uniformity O(·), for a sufficiently small neighborhood N we have α 6= 0,
β 6= 0. Since a + b > 2, the actual asymptotics is
U IN(z) = α (1 + O(z)) + β z2−2a−2b(cid:0)1 + O(z)(cid:1)
U IN(z) = β z2−2a−2b(1 + O(z)).
U IN /∈ L2(cid:0)C,z2a+2b−21 − z2a−2b d z(cid:1).
Therefore
This contradicts to boundedness of U . Thus any point has a neighborhood of zero
measure, and this implies claim (4.3) in the case a + b > 1.
In domains a + b < 0, a − b < −1, a − b > 1 we get the same effect.
Next, examine the complementary series K(z; 0, τ ) of eigenfunctions, see (4.2).
We have the same asymptotics (4.4) -- (4.5), we only must write the coefficients of
the form A(0, τ ), B(0, τ ), C(0, τ ), D(0, τ ) in (4.4) -- (4.5). These functions have zeros
and poles on the axis τ ∈ R. The same argument as above shows that if τ0 is not a
zero and not a pole of all our coefficients, then the measure ρ is zero on a sufficiently
small neighborhood of τ0. The set of zeros and poles is countable. This completes
the proof of the lemma.
(cid:3)
4.3. Proof of non-self-adjointness. However, our system of differential equa-
tions (4.1) has solutions that have not the form 2F C
1 , and enumeration of all possible
degenerations is tedious. So we continue the proof of Lemma 4.1 without constrains
of Lemma 4.2. Due to the homographic transformations, without loss of generality
we can set
(4.6)
a + b > 2.
First, we examine asymptotics in a neighborhood of z = 0.
Asymptotics at z = 0. Non-logarithmic case. If a + b 6= 2, 3, . . . , then
the equation DΦ = λ2Φ has two holomorphic solutions,
a + b
2
The equation DΦ = λ
Ψ1(z) := 2F1(cid:20)a + λ, a − λ
; z(cid:21) , Ψ2(z) := z1−a−b
eΨ1(z) := 2F1(cid:20)a + λ, a − λ
; z(cid:21) ,
eΨ2(z) := z 1−a−b
a + b
2 − a − b
2F1(cid:20)1 − b + λ, 1 − b − λ
; z(cid:21) .
2F1(cid:20)1 − b + λ, 1 − b − λ
; z(cid:21) .
2 − a − b
Φ has two antiholomorphic solutions
Therefore a single-valued solution of the system must have the form
The first term has L2(C, µa,b)-asymptotics at z = 0, by (4.6) the second term
has non-L2-asymptotics. Thus the spectral measure ρ is supported by the set of
AΨ1(z)eΨ1(z) + BΨ2(z)eΨ2(z).
31
functions of the form Ψ1(z)eΨ1(z).
Asymptotics at z = 0. Logarithmic case. Now let a + b = n = 2, 3,. . . .
Then the equation DΦ = λ2Φ has two holomorphic solutions,
Ψ1(z) = 2F1[a + λ, a − λ; n, z],
Ψ2(z),
where Ψ2(z) is a logarithmic solution, which has the form (3.39). The equation
DΦ = λ
Φ has two antiholomorphic solutions,
2
A single valued solution must have the form
eΨ1(z) = 2F1[a + λ, a − λ; n, z],
eΨ2(z).
AΨ1(z)eΨ1(z) + B(cid:16)Ψ1(z)eΨ2(z) + Ψ2(z)eΨ1(z)(cid:17).
The asymptotics of the second summand is (z−n+1 + z −n+1) + O(z−n+2) if n > 3.
If n = 2 we have (z−1 + z −1) + O(z−ε). We get a non-L2 asymptotics.
Thus, for a + b > 2 the spectral measure is supported by set of functions of the
form Ψ1(z)eΨ1(z).
Single-valuedness near z = 1. Non-logarithmic case. Assume that
a − b /∈ Z. We apply formulas Erd´elyi, et al., [9], (2.10.1), (2.10.5) and write
explicit expansions of Ψ1, Ψ1 at z = 1.
Ψ1(z) = A1G1(1 − z) + A2(1 − z)b−aG2(1 − z);
eΨ1(z) = eA1eG1(1 − z) + eA2(1 − z)b−aeG2(1 − z),
where G1, G2 are certain series 2F1 and the coefficients A1, A2 are products of
gamma functions, see the explicit formulas (3.42) -- (3.43) above. Clearly, the prod-
uct Ψ1(z)eΨ1(z) can be single-valued only if A2 = eA2 = 0, or A1 = eA1 = 0. Looking
and only if both series G2(1 − z), eG2(1 − z) are terminating (i.e., b − λ = 0, −1,
to the explicit expressions for the gamma-coefficients, we observe that the first case
happens if both hypergeometric series G1(z), G2(z) are terminating (i.e., a− λ = 0,
−1, . . . or a + λ = 0, −1, . . . , in particular, λ is real). The second variant holds if
. . . or b + λ = 0, −1, . . . ).
Single-valuedness near z = 1. Logarithmic case. Now let b− a ∈ Z. The
transposition a ↔ b corresponds to a homographic transformation of differential
operators, it preserves the condition a+b > 2. Therefore we can assume m := b−a >
equations at the point z = 1,
0. Represent Ψ1(z), eΨ1(z) as combinations of basic solutions of the hypergeometric
Ψ1(z) = A 2F1[a + λ, a − λ; b − a + 1; z] + B Θ(1 − z);
eΨ1(z) = eA 2F1[a + λ, a − λ; b − a + 1; z] + eBeΘ(1 − z),
where Θ(1 − z) is a logarithmic series of the type (3.39), see Erd´elyi, et al., [9],
(2.10.12). A straightforward calculation shows that the product Ψ1(z)eΨ1(z) can
be single valued near z = 1 only if B = eB = 0. Therefore Ψ1(z) is single valued
near z = 1, and therefore it is a single valued solution of a hypergeometric equation
32
on the whole plane C. Hence (see Erd´elyi, et al., [9], Subset. 2.2.1) Ψ1(z) is a
polynomial.
Behavior at infinity. Thus the spectral measure ρ is supported by generalized
eigenfunctions of the following types
p1(z) p2(z),
(1 − z)b−ab−a q1(z) q2(z),
where pj, qj are polynomials. However, our density µa,b(z) has a behavior ∼ z4a−2
at infinity and therefore the space L2 can contain only a finite number orthogonal
functions of such a type.
(cid:3)
5. Symmetry of differential operators
Here we show that J ∗
a,b sends Deven( Λ) to Ra,b and verify that D and D are
adjoint one to another on Ra,b.
5.1. The map J ∗
In this section we denote by Dr(u) ⊂ C (resp. Dr(u)) the open (resp. closed)
disc in C of radius r with center at u. By Sr(u) we denote the circle z − u = r.
a,b on the space Deven( Λ). Introduce a natural topology in
the space Ra,b( C) defined in Subsect. 1.1. Consider the space R(0) of functions
in D1/3(0) having the form α(z) + β(z)z2a+2b−2, where α(z), β(z) are smooth in
D1/3(0) up to the boundary. Let C∞
flat(D1/3(0)) ⊂ C∞(D1/3(0)) be the subspace
consisting of all functions that are flat at 0. The space R(0) is a quotient space
R(0) ≃hC∞(D1/3(0)) ⊕ z2a+2b−2C∞(D1/3(0))i/C∞
flat(D1/3(0)).
We equip R(0) with the topology of a quotient space. In the same way we define a
topology in the space R(1) of smooth functions in D1/3(1) having the form γ(z) +
δ(z)1 − z2a−2b.
We define a topology in Ra,b as a weakest topology such that:
a) The restriction operators
Ra,b → R(0), Ra,b → R(1), Ra,b → C∞(cid:16)D2(0) \ (D1/3(0) ∩ D1/3(1))(cid:17)
are continuous.
b) For all α, β, N the following seminorms are continuous
where the expression B(k, s; u) for fixed k is smooth s except the point (k, s) = (0, 0).
Proof. We refer to expansion (3.32) -- (3.34). Notice that for k = 0, s = 0 we
have a singularity in this expansion (but the kernel itself is analytic at this point).
(cid:3)
(cid:12)(cid:12)−a+ −k+is
2 B(−k,−s; z−1),
(5.1)
pα,β,N (f ) = sup
Recall that Λ := Λ \ {(0, 0)}.
Lemma 5.1. For z > 2, (k, s) ∈ Λ we have the following expansion
(5.2) K(z; k, is) = z−a− k+is
2
∂α+βf (z)
∂zα∂zβ (cid:12)(cid:12)(cid:12)(cid:12).
C\D2(0)z2+α+β(lnz)N(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)−a− −k+is
2 B(k, s; z−1)+
+ z−a+ k+is
2
33
Proposition 5.2. a) Let Φ ∈ Deven( Λ). Then J ∗
a,bΦ ∈ Ra,b.
b) Moreover, the operator J ∗
Proof. Forms of asymptotics of J ∗
a,b is a continuous operator from Deven( Λ) to Ra,b.
a,bΦ at 0 and 1 follow from the expressions
(3.26), (3.29). Let us examine the asymptotics at z → ∞. Without loss of generality
we can assume that k is fixed. We write
2 − is
2 −a+ k
a,bΦ(z) = z−a− k
J ∗
2 B(k, s; z−1) Φ(k, s) ds +nsimilar
.o
term
2ZR
z− is
Differentiating the first summand by ∂α+β
3.6, we get an expression of the form
∂zα∂zβ and keeping in mind (3.32) and Lemma
z−a− k
2 −α−a+ k
z− is
2 − is
2 U α,β
2 −β X06p6α,06q6βZR
p,q (a, b, k, s)×
)p (a + k−is
×
ΓC(−k − isk + is) (a + k+is
ΓC(b − k+is
2 b − −k+is
2 + p, a + −k+is
2
2 + p
× 2F1(cid:20)a + k+is
2 )p (a + −k+is
)ΓC(a + k+is
2
2
2 a + −k+is
; z−1(cid:21) 2F1(cid:20)a + −k−is
a + b + p
2 )q (a + −k−is
) (a + b)p (a + b)q ×
)q
2
2 + q, a + k−is
2 + q
a + b + q
; z −1(cid:21) ×
× Φ(k, s) ds,
where U α,β
smooth compactly supported function on Λ. Next, we write
p,q (a, b, k, s) are polynomials. It is easy to verify that the integrand is a
z−is =
i
lnz
∂
∂sz−is,
integrate our expansion by parts N times and observe that pα,β,N (J ∗
The continuity follows from the same considerations.
As a corollary, we obtain the following lemma.
a,bΦ) < ∞.
(cid:3)
Lemma 5.3. The operator J ∗
space L2(C, µa,b).
a,b is continuous as an operator from Deven( Λ) to the
Proof.
Indeed, for (a, b) ∈ Π the identical embedding f 7→ f of Ra,b to
(cid:3)
L2(C, µa,b) is continuous.
Lemma 5.4. If f ∈ Ra,b, then Df ∈ Ra,b.
Proof. Let us check the behavior of Df at 0, For definiteness assume that
a + b 6= 1. Then near zero we have
Df = D(cid:0)α(z) + β(z)z1−a−b(cid:1) =
dz(cid:17)z1−a−bo · z1−a−bβ(z) +nthe resto.
Obviously, the rest has the form eα(z) + eβ(z)z2−a−b with smooth eα, eβ. The ex-
pression in the curly brackets22 is −(a + b)(a + b − 1)z1−a−b.
=n(cid:16)z(1 − z)
5.2. Symmetry of differential operators.
d2
dz2 + (a + b)
(cid:3)
d
22Cf. [11], Sect.I.2.
34
Proposition 5.5. For any f , g ∈ Ra,b( C)
hDf, gi = hf, Dgi.
Proof. Let f , g ∈ Ra,b. We wish to show that
By Lemma 5.3, Df , Dg ∈ L2(C, µa,b). Therefore our improper integral absolutely
converges, we write it as
lim
Z C(cid:0)Df (z) · g(z) − f (z) · Dg(z)(cid:1) µa,b d z = 0.
ε→0ZD1/ε(0)\(cid:0)Dε(0)∪Dε(1)(cid:1)(cid:0). . .(cid:1) dz ∧ dz
V (z) dz(cid:27),
V (z) dz −ZSε(1)
V (z)d z −ZSε(0)
∂z (cid:17)z(1 − z) µa,b(z).
∂z · g(z) − f (z) ·
∂g(z)
lim
ε→0(ZS1/ε(0)
V (z) =(cid:16) ∂f
.
2i
(5.3)
where
Next, we integrate two times by parts in z (with the Green formula) and after a
simple calculation come to
We claim that all summands in (5.3) tend to 0. For the first summand this is clear.
For the second summand we represent f , g as
f (z) = α(z) + β(z) z1−a−b1−a−b,
g(z) = γ(z) + δ(z) z1−a−b1−a−b.
Then V (z) transforms to an expression of the following type:
(cid:16)A(z) + B(z)z−a−b1−a−b + C(z)z2−2a−2b2−2a−2b(cid:17)×
× z(1 − z) · za+b−1a+b−1(1 − z)a−ba−b,
where A(z), B(z), C(z) are smooth near 0. We emphasize that the term with
z1−2a−2b2−2a−2b in the bracket appears with the coefficient
Thus we get summands with the following behavior at 0:
(2 − 2a − 2b)(cid:0)β(z)δ(z) − β(z)δ(z)(cid:1) = 0.
∼ A(0) za+ba+b−1,
∼ B(0) z0,
∼ C(0) z2−a−b1−a−b.
Since 0 < a + b < 2 all powers are > −1 and therefore23Rz=ε(. . . ) dz tends to 0.
(cid:3)
6. The operator J ∗
a,b is an isometry
Here we prove half of Theorem 1.3.
6.1. The statement. First, denote by Λ+ the subset of Λ consisting of (k+is)/2
such that k > 0 or k = 0 and s > 0. We have an obvious identification Deven( Λ) ≃
D(Λ+).
Lemma 6.1. Let u(λ), v(λ) be smooth compactly supported function on Λ+. Then
a,bu, J ∗
hJ ∗
a,bviL2( C,µa,b) = 2hu, viL2(Λ+,κa,b).
23See a discussion of a parallel situation for ordinary differential operators in [36], Section 1.
However, in the one-dimensional case we must impose boundary conditions in such points.
Our proof is based on heuristic arguments outlined in Berezin, Shubin [3], Section
2.6, for ordinary differential operators. However, this way is tiresome.
35
6.2. Preliminary remarks. Recall that
J ∗
a,bu(z) = 2ZΛ+
u(λ) K(z, λ) κa,b(λ)edλ.
By Lemma 5.3, this operator is continuous as an operator D( Λ+) → L2( C, µa,b).
Therefore the sesquilinear form
(6.1)
is continuous as a form D(Λ+) × D(Λ+) → C. By the kernel theorem (see, e.g.,
[19], Sect. 5.2) it is determined by a distribution. Formally, we transform (6.1) as
a,bviL2( C,µa,b)
T (u, v) := hJ ∗
a,bu, J ∗
v(ν)K(z, ν)κa,b(ν)edν(cid:17) µa,b(z) d z =
(6.2)
(6.3)
where
(6.4)
ZC (cid:16)ZΛ+
=ZΛ+ZΛ+
u(λ)K(z, λ)κa,b(λ)edλ(cid:17) ·(cid:16)ZΛ+
u(λ)v(ν)H(λ, ν)κa,b(λ)κa,b(ν)edλedν,
H(λ, ν) =Z C
K(z, λ) K(z, ν) µa,b(z) d z.
Notice that all integrals in line (6.2) converge absolutely. However, the triple
However, we regard H(λ, ν) as a distribution, then Lemma 6.1 can be reformu-
integralRΛ+RΛ+R C is not absolutely convergent. The integrand in (6.4) decreases
as z−2 and the integral diverges.
lated in the form:
Lemma 6.2. We have the following identity of distributions on D(Λ+) × D(Λ+):
(6.5)
H(λ, ν) = δ(λ − ν).
6.3. Orthogonality of packets.
Lemma 6.3. Let u, v ∈ D(Λ+) and supports supp(u), supp(v) have empty inter-
section. Then
a,bu, J ∗
hJ ∗
Proof. Denote D+ := 1
2 (D + D), D− = 1
a,bviL2( C,µa,b) = 0.
2i ((D − D). By Proposition 5.2,
J ∗
a,bu is contained in the space Ra,b. By Proposition 5.5, the operators D+, D− are
formally symmetric on Ra,b. Since they formally commute, for any real polynomial
p(D+, D−) we have
hp(D+, D−)J ∗
a,bu, J ∗
a,bvi = hJ ∗
a,bu, p(D+, D−)J ∗
a,bvi,
or
a,bvi = hJ ∗
a,b p(Re λ, Im λ) · u, J ∗
hJ ∗
(6.6)
where · denotes the operator of multiplication by a function. We choose a sequence
pN of polynomials such that pN uniformly converges to 1 on supp(u) with all
derivatives and converges to 0 on supp(v). By Lemma 5.3 the map J ∗
a,b is continuous
as a map D(Λ+) → L2(C, µa,b). Replacing p by pN in (6.6) and passing to a limit,
we come to the desired statement.
a,b p(Re λ, Im λ) · vi,
a,bu, J ∗
(cid:3)
36
6.4. Next reduction of our statement. Let S(u, v) be an Hermitian form
on D(Λ+). We say that S is Cω-smooth if it has the form
S(u, v) =ZΛ+ZΛ+
M (λ, ν) u(λ)v(ν)edλedν,
where M is a real analytic function on Λ+ × Λ+.
Lemma 6.4. We have
(6.7)
hJ ∗
a,bu, J ∗
a,bviL2(C,µa,b) = hu, viL2(Λ,κa,b) + S(u, v),
where S(u, v) is Cω-smooth.
This lemma together with Lemma 6.3 imply the desired statement, i.e., the
identity (6.5). Indeed, for any u, v with disjoint support, we have
ZΛ+ZΛ+
M (λ, ν) u(λ) v(ν) κa,b(λ)κa,b(ν)edλedν = 0,
therefore M (λ, ν) = 0.
The rest of this section is occupied by the proof of Lemma 6.4.
6.5. Beginning of the proof of Lemma 6.4. Cleaning of the problem.
Step 1. Represent
u =Xk
uk δ(Re λ − k/2),
vl δ(Re λ − l/2),
v =Xl
in fact the sums are finite and uk, vl depend on a real variable s. By Lemma 6.3,
we have
for k 6= l.
Therefore it is sufficient to examine only inner products
a,bvli = 0
a,buk, J ∗
hJ ∗
where
R(z) d z,
a,buk, J ∗
hJ ∗
a,bvki =Z C
2 (k + is)(cid:1)κa,b(cid:0) 1
×ZΛ+
vk(it) K(cid:0)z, 1
R(z) :=ZΛ+
uk(is) K(cid:0)z, 1
2 (k + is)(cid:1) ds×
2 (k + it)(cid:1)κa,b(cid:0) 1
Step 2. Represent the integral asRz62 R +Rz>2 R.
absolutely converges and can be written as
where
Zz62
L(s, t) =Zz62
R d z =ZRZR
K(cid:0)z, 1
2 (k + it)(cid:1) dt µa,b(z).
Let us show that the first summand is Cω-smooth. In this case the triple integral
uk(is)vk(it)L(s, t) ds dt,
2 (k + is)(cid:1) K(cid:0)z, 1
2 (k + it)(cid:1) µa,b(z) d z.
Integrand makes sense for complex s, t that are sufficiently close to R and the
integral absolutely converges (singularities at z = 0 and 1 have the forms (4.4),
(4.5)). Therefore L(s, t) is a holomorphic function in s, t near R × R.
37
Therefore our question is reduced to an examination the integral
Zz>2
R(z) d z
Step 3. A decomposition of the kernel. Applying Theorem 3.9.c, we write K(z, λ)
in the domain z > 2 as
(6.8) K(z, λ) = W1 + W2 + W3 :=
= A(λ)(−z)−a−λ−a+λ + A(−λ)(−z)−a+λ−a−λ + Ψ(z, λ),
A(λ) =
ΓC(2λ − 2λ)
ΓC(b − λb + λ)ΓC(a − λa + λ)
as z → ∞.
Ψ(z, λ) = O(z−2a−1)
where
and
Notice that
(6.9)
A(λ)2 = A(λ) A(−λ) = κ−1
a,b(λ).
α, β = 1, 2, or 3,
Therefore the integralRz>2 R(z) d z splits into a sum of 9 summands Vαβ , where
Vαβ :=Zz>2ZR
Wα(z; k, s)uk(is) κa,b(cid:0) 1
2 (k + is)(cid:1) ds×
Wβ(z; k, t)vk(it) κa,b(cid:0) 1
×ZR
2 (k + it)(cid:1) dt · µa,b(z) d z.
Step 4. For five summands V13, V23, V31, V32, V33 we immediately get absolute
convergence of triple integrals and Cω-smoothness. For instance,
V13 =ZRZR
2 (k − is)(cid:1)−1
uk(is)uk(it) A(cid:0) 1
×(cid:20)Zz>2(cid:16) z
z(cid:17)−k
×
z−2a+is Ψ(cid:0)z, 1
2 (k + is)(cid:1) µa,b(z) d z(cid:21) ds dt.
(we simplified the integrand using (6.9)). The expression in the square brackets is
real analytic (the integrand decreases as z−3).
Step 5. Non-obvious summands are V11, V12, V21, V22. We start with V11,
V11 :=Zz>2ZR
z(cid:17)−k/2
2 (k + is)(cid:1)(cid:16) z
uk(is)A(cid:0) 1
z−2a−isκa,b(cid:0) 1
×ZR
2 (k + it)) vk(it)(cid:16) z
z(cid:17)k/2
A(cid:0) 1
z−2a+itκa,b(cid:0) 1
2 (k + is)(cid:1) ds×
2 (k + it)(cid:1) dt µa,b(z) d z.
For k = 0 we must keep in mind that the integrationRR actually is taken over a
V11 := Zz>2 ZR ZR
z−4a−is+it ds dt×
ray [ε,∞) for some ε > 0. Applying (6.9), we come to
(6.10)
uk(is)vk(it)A(cid:0) 1
A(cid:0) 1
2 (k + it)(cid:1)−1
2 (k − is)(cid:1)−1
× µa,b(z) d z.
38
Next, we notice that
µa,b(z) = z2a+2b−21 − z2a−2b = z4a−2 + O(z4a−3)
as z → ∞.
We write
(6.11)
µa,b(z) = z4a−2 + (µa,b(z) − z4a−2),
substitute this to (6.10) and decompose (6.10) as a sum of two integrals. The second
summand immediately gives a Cω-smooth term. The first summand is the topic of
our interest. It equals the following expression:
(6.12)
I(u, v) := Zz>2 ZR ZR
uk(is)vk(it)A(cid:0) 1
2 (k − is)(cid:1)−1
×
A(cid:0) 1
2 (k + it)(cid:1)−1
× z−2−is+it ds dt d z.
6.6. Application of the Sokhotski formula and disappearance of a
singular term.
Step 6. Extension to the complex domain. Now consider a function I(u, v, ε)
obtained by replacing s 7→ s − iε in the boxed term, ε > 0. The new triple
integral absolutely converges, we can change the order of integrations and explicitly
integrate in z. We get
I(u, v, ε) =ZR ZR
uk(is)vk(it)A(cid:0) 1
2 (k − is)(cid:1)−1
2 (k + it)(cid:1)−1 2−is−ε+it
A(cid:0) 1
−is − ε + it
ds dt.
Next, we claim that
Indeed, we integrate I(u, v, ε) two times by parts in s and come to
I(u, v) = lim
ε→+0
I(u, v, ε).
I(u, v, ε) = Zz>2 ZR
vk(it)A(cid:0) 1
2 (k + it)(cid:1)−1ZR
∂2
∂s2(cid:20)uk(is)A(cid:0) 1
2 (k − is)(cid:1)−1(cid:21)×
× z−2−ε−is+it
i2 ln2 z
ds dt d z
The new triple integral absolutely converges and is continuous at ε = +0.
Thus we come to the so-called distribution 1
x−iε , see, e.g., [15]. Recall the Sokhot-
ski formula
(6.13)
f (y) dy
x − y
where p.v. denotes the principal value of an integral.
f (y) dy
x − y − iε
α
= p.v.Z β
α
ε→+0Z β
lim
+ πif (x),
Applying this formula and keeping in mind (6.9), we come to
(6.14)
I(u, v) = p.v.ZR ZR
uk(is)vk(it) A(cid:0) 1
2 (k − is)(cid:1)−1
+ πZR
ds dt+
−is + it
2 (k + it)(cid:1)−1 2−is+it
A(cid:0) 1
uk(is)vk(is) κa,b(cid:0) 1
2 (k + is)(cid:1) ds.
Step 8. We deal with V22 in the same way and come to
39
Next, we take the sum V11 + V22 modulo Cω-smooth terms. The expression
ds dt+
2 (k + is)(cid:1)−1 2−it+is
A(cid:0) 1
−it + is
uk(is)vk(it)A(cid:0) 1
uk(is)vk(is)κa,b(cid:0) 1
2−is+it − A(cid:0) 1
2 (k − it)(cid:1)−1
2 (k + is)(cid:1) ds +na Cω-smooth termo.
2 (k − it)(cid:1)−1
A(cid:0) 1
2 (k + is)(cid:1)−1
2−it+is
−i(s − t)
(6.15) V22 = p.v.ZR ZR
+ πZR
A(cid:0) 1
2 (k − is)(cid:1)−1
has the form
A(cid:0) 1
2 (k + it)(cid:1)−1
L(t, s) − L(s, t)
s − t
with analytic L(t, s). It has a removable singularity on the line t = s. Thus the first
summands in (6.14) and (6.15) give us a Cω-smooth term, the second summands
give us the first term in (6.7), i.e., the desired delta-function.
6.7. End of the proof of Lemma 6.4.
Step 9. Next, we examine the term V12. We write the integral and apply the
transformation (6.11). We get a sum of a Cω-smooth term and the integral
J(u, v) =Zz>2ZRZR
uk(is)vk(it)A(cid:0)− 1
A(cid:0)− 1
2 (k + is)(cid:1)−1
2 (k + it)(cid:1)−1
z(cid:17)k
×(cid:16) z
z−2−is−it d z dt ds.
×
As above, we change s 7→ s − iε in the box and get integrals J(u, v, ε) with ε > 0.
As above,
J(u, v; ε) =ZRZR
uk(is)vk(it)A(cid:0)− 1
2 (k + it)(cid:1)−1
A(cid:0)− 1
2 (k + is)(cid:1)−1
×(cid:20)Zz>2(cid:16) z
z−2−ε−is−it d z(cid:21) dt ds.
z(cid:17)−k
×
If k > 0, then the term in square brackets is zero (we pass to polar coordinates and
get 0 after the integration with respect to the angle coordinate). If k = 0, then we
get
2−ε−is−it
ε + i(s + t)
.
However, supp(u0), supp(v0) are contained in domains s > 0, t > 0, and actually
we have no singularity. Thus V12 is Cω-smooth.
The same examination shows Cω-smoothness of V21. This completes the proof
(cid:3)
of Lemma 6.4.
40
7. Asymptotics of the kernel in the parameters
7.1. The statement. Let us modify a notation for the kernel K. Set
2a − λ + σ
2 , a − λ − σ
2a + λ − σ
a + −k+σ+is
2
, a + −k−σ−is
2
a + ba + b
a + ba + b
2
; z(cid:21) =
; z(cid:21) ,
a + k−σ−is
2
K◦(z; λ; σ) :=
1
:=
ΓC(a + ba + b) 2F C
1
=
ΓC(a + ba + b) 2F C
where λ ∈ Λ, σ ∈ R. In fact,
2
1 (cid:20)a + λ + σ
1 (cid:20)a + k+σ+is
K◦(cid:0)z; k+is
Denote
2 ; σ(cid:1) = K(z; k, σ + is).
t±(z) = 1 ±p1 − 1/z.
However, in calculations of this section the variables σ and is have different mean-
ings.
Theorem 7.1. Then for a fixed z we have the following asymptotic expansion
(7.1) K◦(z; λ; σ) =
1
2b − λ + σ
2a + λ − σ
2(cid:1) ΓC(cid:0)b + λ + σ
ΓC(cid:0)a − λ − σ
× 1 − 1/z−1/2 · 1 − zb−a · z−a−b×
2 −λ Xk>0, l>0, k+l<N
2 −λ Xk>0, l>0, k+l<N
2(cid:1) · λ×
Ak(σ,√1 − z) Al(σ,√1 − z)+
Ak(σ,−√1 − z) Al(σ,−√1 − z)(cid:21)+
λ −l
λ −l
k! l!
k! l!
−k
−k
λ
λ
2 +λ σ
×(cid:20)(cid:16) t−(z)
t+(z)(cid:17) σ
+(cid:16) t+(z)
t−(z)(cid:17) σ
2 +λ σ
+ RN (z, σ, λ),
where Ak(ξ) are rational expressions in ξ (depending on the parameters a, b) having
poles at ξ = 0, ±1 and A0 = 1. The reminder RN (z) satisfies
as λ → ∞,
RN (z, σ, λ) = O(λ−N ),
moreover O(·) is uniform in z and σ on compact subsets in C × R.
The proof occupies the rest of this section.
Remark. This formula is a counterpart of Watson's [44] formula for asymptotics
of the Gauss hypergeometric functions 2F1[a− λ, b + λ; c; z] in the parameter λ (see
an exposition of Watson's results in [25], Sect. 7.2, see also a remark in [41], p.162,
on typos in [44]). We do not see a way to reduce our statement to Watson's work.
⊠
Remark. Lemma 2.1 gives us an asymptotics of the gamma-factor in (7.1). ⊠
7.2. Stationary phase approximation. We transform K◦(z, λ, σ) as
ΓC(cid:0)a−λ− σ
1
2(cid:1) ΓC(cid:0)b+λ+ σ
2 a+λ− σ
2 b−λ+ σ
(7.2)
where
(7.3)
R(t, z) := ta− σ
2 −1a− σ
2 −1(1 − t)b+ σ
2 −1b+ σ
2(cid:1)·λZC
R(t, z, σ) expnQ(t, z, λ, σ)o d t,
2 −1(1 − tz)−a− σ
2
and
41
(7.4) Q(t, z, λ) := λ ln(cid:16) t(1 − zt)
1 − t (cid:17) − λ ln(cid:16) t(1 − zt)
1 − t (cid:17) =
= ik Im ln(cid:16) t(1 − zt)
1 − t (cid:17) + is ln(cid:16) t(1 − zt)
1 − t (cid:17).
The function Im ln(. . . ) is ramified, however the exponent is well-defined and for-
mulas below contain only partial derivatives of ln(. . . ), which are independent of
the choice of a branch.
We apply the stationary phase approximation, see, e.g., Fedoryuk [10], Hormander
[19]. Singular points are 0, 1, ∞. Stationary points are
t± = 1 ±p1 − 1/z,
they are the same for both summands in (7.4). This could be a fatal obstacle for
an evaluation of a uniform asymptotics, however this does not happen. Also the
domain of convergence of the integral (7.2) is smaller than it is necessary for our
purposes.
Consider a partition of unity
1 = ρ0 + ρ1 + ρz−1 + ρ∞ + ρt+ + ρt− + τ,
where ρα is zero outside a small neighborhood of α, and τ is zero in neighborhoods
of 0, 1, z−1, ∞, t±. According to this partition we expand (7.2) into a sum of 7
integrals,
I = I0 + I1 + Iz−1 + I∞ + It+ + It− + J.
Obviously (see [10], Lemma III.2.1), for each N we have
as n + is → ∞.
J = O(k2 + s2)−N
7.3. Preparatory statement.
Theorem 7.2. Let Ω be a domain in C, f (t), ϕ(t) be holomorphic in Ω. Let t0
be a unique zero of ϕ′(t) in Ω and ϕ′′(t0) 6= 0. Let ρ(t) be a C∞-smooth function
compactly supported by Ω such that ρ = 1 in a neighborhood of t0. Consider the
integral
(7.5)
I(λ) =ZΩ
ρ(t) f (t) f (t) expni Re(λϕ(t)o d t,
where λ ∈ C is a parameter. Then
a) For λ > 1 we have the following expansion
(7.6)
I(λ) =
1
f ′′(t0)λ
expni Re(λϕ(t0)o×
×(cid:16) Xk>0, j>0, k+l<N
−l
λ−kλ
k! l!
ak(f, ϕ) al(f , ϕ) + RN (λ)(cid:17),
where ak are rational expressions
ak = ak(cid:0)ϕ(t0), ϕ′(t0), . . . ; f (t0), f ′(t0), . . . ; ϕ′′(t0)−1(cid:1)
and a0 = 1. The reminder RN satisfies
(7.7)
RN (λ) = O(λN )
as λ → ∞.
c) Let ϕ = ϕα, f = fα smoothly depend on a parameter α, where α ranges in
a compact domain K ⊂ C and the conditions of the preamble of the theorem are
satisfied for all α. Then O(·) in (7.7) is uniform in α ∈ K.
Proof. b) We use Fedoryuk [10], Proposition III.2.2 or Hormander [19], Theo-
rem 7.7.5. Let f be a smooth compactly supported function on Rn, let S be smooth.
Consider an n-dimensional integral
I(σ) :=ZRn
f (x) exp{iσS(x)} dx,
t > 1.
42
b) The asymptotic expansion
I(λ) ∼ λ−1 Xk>0, l>0
ckl
l
λkλ
as λ → ∞
can be written as
(7.8)
I(λ) ∼
× expn
× expn
2λ ϕ′′(t0)
i
2λ ϕ′′(t0)
1
f ′′(t0)λ
i
expni Re(λϕ(t0)o×
∂2
∂t2o(cid:16)f (t) expnλ(cid:0)ϕ(t) − ϕ′(t0) − 1
∂t2o(cid:16)f (t) expnλ(cid:0)ϕ(t) − ϕ′(t0) − 1
∂2
2 ϕ′′(t0)(t − t0)2(cid:1)o(cid:17)(cid:12)(cid:12)(cid:12)(cid:12)t=t0
2 ϕ′′(t0)(t − t0)2(cid:1)o(cid:17)(cid:12)(cid:12)(cid:12)(cid:12)t=t0
×
.
Let x0 be a unique critical point of S on the support of f , let it be nondegenerate.
Let H(x0) be the Hessian of S at x0 (i.e., the matrix composed of second partial
derivatives), let sgn H(x0) denote the signature of the Hessian (the number of pos-
itive eigenvalues minus the number of negative eigenvalues). Consider the second
order differential operator
L :=
i
2hH(x0)−1∇x,∇xi,
where ∇x denotes the vector column composed of
(7.9)
S(x, x0) := S(x) − S(x0) −
1
2hH(x0)(x − x0), (x − x0)i,
∂
∂x1
, . . . ,
∂
∂xn
. Denote
this expression is the part of the Taylor expansion of S(x) at x0 starting cubic
terms. Then the following expansion take holds:
(7.10)
σ (cid:17)n/2
I(σ) =(cid:16) 2π
×(cid:18)N −1Xk=0
4
det H(x0)−1/2 exph iπ
sgn H(x0)i×
Lk(cid:0)f (x) exp{iσS(x, x0)}(cid:1)(cid:12)(cid:12)(cid:12)x=x0
σ−k
k!
where V (σ) is bounded.
+ σ−N +[2N/3]V (σ)(cid:19),
Let us return to our integral (7.5). Without loss of generality, we can set t0 = 0,
ϕ′′(t0) = 1, i.e.,
ϕ(t) = 1
2 t2 + r(t),
where r(0) = r′(0) = r′′(0) = 0.
Set λ = seiθ, s > 0. Set z = x + iy, then
ϕ(x, y) = 1
2 (x2 − y2 + 2ixy) + r(x, y).
Thus we come to an oscillating integral in s with the parameter θ,
I(s, θ) =Z ρ(x, y)f (x, y)f (x, y) expnis(cos θ Re ϕ(x, y) + sin θ Im ϕ(x, y))o dx dy.
We wish to apply the general statement formulated above. The Hessian is given by
43
H = 2(cid:18)cos θ
sin θ − cos θ(cid:19) ,
sin θ
H −1 =
The signature is 0. The differential operator L is
1
2(cid:18)cos θ
sin θ − cos θ(cid:19) .
sin θ
L =
i
4(cid:16)cos θ(cid:16) ∂2
∂x2 −
∂2
∂y2(cid:17) + 2 sin θ
∂2
∂x∂y(cid:17) =
i
2(cid:16)eiθ ∂2
∂t2(cid:17).
∂t2 + e−iθ ∂2
Next, we rewrite our phase function S(·) as
e−iθϕ(t) + eiθϕ(t).
Therefore the expression (7.9) is
e−iθr(t) + eiθr(t).
Applying (7.10), we get
I(s, θ) :=
2π
s
∂t2 + e−iθ ∂2
∂t2(cid:17)o×
expn i
2s(cid:16)eiθ ∂2
×(cid:16)f (t)f (t) exp(cid:8)is(cid:0)e−iθr(t) + eiθr(t)(cid:1)(cid:9)(cid:17)(cid:12)(cid:12)(cid:12)(cid:12)t=0
∂t2o(cid:16)f (t) exp(cid:8)ise−iθr(t)(cid:9)(cid:17)(cid:12)(cid:12)(cid:12)(cid:12)t=0×
expn
× expn i
2se−iθ
2π
s
2seiθ
∂2
=
=
i
∂2
∂t2o(cid:16)f (t) exp(cid:8)iseiθr(t)(cid:9)(cid:17)(cid:12)(cid:12)(cid:12)(cid:12)t=0
.
We obtained asymptotics in s for fixed θ. However, θ ranges in a compact set, by
[10], Theorem III.2.2, we get that the term V (·) in (7.10) is bounded uniformly in
θ.
a) follows from b).
c) We again refer to the parametric version of the stationary phase approxima-
(cid:3)
tion, see [10], Theorem III.2.2.
7.4. Contribution of the stationary points. Let us apply Theorem 7.2 to
our integral (7.2). We have
t
1 − zt(cid:17)a
f (t) = R(t, z) =(cid:16)
ϕ(t) = 2 ln(cid:16) 1 − t
t(1 − zt)(cid:17).
(1 − t)b(cid:16) 1 − t
t(1 − zt)(cid:17) σ
2 (cid:0)t(t − 1)(cid:1)−1
;
ζ =p1 − 1/z.
Denote
We substitute t = t+ and transform the factors of R(t, z) = f (t)f (t):
=(cid:0)(z − 1)z(cid:1)−a/2−a/2
;
44
(7.11)
(7.12)
(7.13)
(7.14)
Next,
therefore
Finally,
and
(7.15)
;
.
t
2 σ
2
(cid:16)
1 − zt(cid:17)aa(cid:12)(cid:12)(cid:12)(cid:12)t=t+
−ζ (cid:17)aa
=(cid:16) 1 − ζ2
(1 − t)bb(cid:12)(cid:12)(cid:12)t=t+
z (cid:17)b/2b/2
=(cid:16) 1 − z
2(cid:12)(cid:12)(cid:12)t=t+
=(cid:16) 1 − ζ
1 + ζ(cid:17) σ
t(1 − zt)(cid:17) σ
(cid:16) 1 − t
(cid:0)t(t − 1)(cid:1)−1−1(cid:12)(cid:12)(cid:12)t=t+
=(cid:16) −1
ζ(1 + ζ)(cid:17)11
1 + ζ(cid:17),
ϕ(t+) = 2 ln(cid:16) 1 − ζ
1 + ζ(cid:17)λ−λ
2 (k + is)(cid:1)o =(cid:16) 1 − ζ
(1 − tz)2 ,
(1 − t)2 +
ϕ′′(t+) =
expni Re(cid:0)ϕ(t+) 1
ϕ′′(t) = −2
2
t2 +
2z2
−4
ζ(1 + ζ)2 .
t+(cid:17)λ−λ
=(cid:16) t−
.
Uniting these data we get that the leading term at the point t+ is
− ζ1 − zb−a z−a−b(cid:16) t−
t+(cid:17)λ−λ
1
(k2 + s2)1/2 .
·
Th general form of the asymptotic expansion at t = t+ follows from Theorem
7.2.
7.5. Contributions of the singular points.
Lemma 7.3. Contributions at the singular points 0, 1, ∞ are O(λ−N ) for any
N .
Proof. For definiteness examine the point 0. We have the integral
I0(λ) =ZC
ρ0(t) ta−1a−1(1 − t)c−a−1(1 − zt)−a(cid:16) t(1 − zt)
1 − t (cid:17)λ−λ
d t,
defined as an analytic continuation. Keeping in mind that a support of ρ0 can be
chosen sufficiently small, we pass to a new variable in a neighborhood of 0,
and come to an integral of the form
u =
t(1 − zt)
1 − t
,
I0(λ) =ZC
ua−λ−1a+λ−1Φ(u) d u,
where Φ is a smooth compactly supported function. It remains to apply Theorem
2.2.
Argumentation for other singular points is the same.
(cid:3)
45
8. Symmetry of difference operators
Here we prove Theorem 1.7, i.e., show that if f ∈ D( C), then Ja,bf is contained
in the space Wa,b of meromorphic functions on ΛC. Also we show that L and L are
formally adjoint one to another on Wa,b, see Theorem 8.4.
8.1. Beginning of the proof of Theorem 1.7. We follow the list of properties
in the definition of Wa,b, see Subsect. 1.8,
a) is a corollary of the symmetry Ka,b(z;−k,−σ) = Ka,b(z; k, σ).
b) We must examine poles of Ka,b(z; k, σ) as a function of the variable σ for a
fixed z ∈ C, k ∈ Z. Let a + b 6= 1. We look to the expansion (3.26) of 2F C
1 [·] at
z = 0. The only source of poles of K are zeros of the denominators in (3.28), i.e.,
zeros of the expression
(8.1) R(k, σ) := ΓC(cid:0)a + k+σ
2
2 (cid:12)(cid:12)a + −k+σ
ΓC(cid:0)b + k+σ
(cid:1) ΓC(cid:0)a + −k−σ
2 (cid:12)(cid:12)b + −k+σ
2
2
2
(cid:12)(cid:12)a + −k−σ
(cid:1)×
(cid:1) ΓC(cid:0)b + −k−σ
2
(cid:12)(cid:12)b + −k−σ
2
(cid:1).
This gives us the desired list of possible poles.
Let us examine the case a + b = 1. The decomposition of the hypergeometric
functions (3.26) at z = 0 produces an expression of the type
(8.2)
Ka,b(z; k, σ) =
ua,b(z, k, σ) − vz,a,b(z, k, σ)
a + b − 1
with ua,b, va,b having poles at zeros of R(k, σ). A decomposition at z = 1 gives
Ka,b(z; k, σ) =
Ua,b(z, k, σ) − Va,b(z, k, σ)
,
a − b
therefore the singularity in (8.2) at a + b = 1 is removable.
d) Indeed, we have Ka,b(p, q) = Ka,b(q, p), i.e.,
2F C
1 (cid:20)a + p+q
2 (cid:12)(cid:12)a + −p+q
2
, a + −p−q
2
a + b(cid:12)(cid:12)a + b
= 2F C
2
(cid:12)(cid:12)a + p−q
1 (cid:20)a + p+q
; z(cid:21) =
2 (cid:12)(cid:12)a + p−q
This is a special case of the symmetry (3.36).
We also mention the following similar identity for (8.1):
2 , a + −p−q
2
a + b(cid:12)(cid:12)a + b
(cid:12)(cid:12)a + −p+q
2
; z(cid:21) .
(8.3)
R(p, q) = R(q, p),
it is a special case of (3.37).
The statement c) about the behavior at infinity is a corollary of the expansion
(7.1) and the following lemma
Lemma 8.1. Let t±(z) be as in Theorem 7.1. Let Φ ∈ D( C) be a function with
a simply connected support. Then for any A > 0 for any N > 0 in the strip
Re σ < A we have
Z C
Φ(z)(cid:16) t−(z)
as (k2 + (Im σ)2) → ∞.
t+(z)(cid:17)(k+σ)/2(−k+σ)/2
d z = O(cid:0)k2 + (Im σ)2(cid:1)−N
46
We need a simply connected support since the integrand is ramified at the points
z = 0, z = 1. A proof of the lemma requires some preparations.
8.2. A change of variable. We define a new variable
(8.4)
The inverse map is done by
(8.5)
p :=
t+(z)
t−(z)
,
z = ζ(p) :=
(p + 1)2
4p
.
The map ζ(p) determines a two-sheet covering map from
C := C \ {0, 1,−1}
p1 − 1/z =
t− =
p − 1
p + 1
2
p + 1
,
,
ζ′(p) =
p2 − 1
4p2 ,
t+
t−
= p.
1 − z = −
(p − 1)2
2p
4p
,
t+ =
,
p + 1
(8.6)
to C. Notice that
(8.7)
(8.8)
Also,
(8.9)
ζ(p−1) = ζ(p),
ζ′(p−1) p−1 = ζ(p) p.
8.3. Proof of Theorem 1.7.c.
Proof of Lemma 8.1 We substitute z = ζ(p) to the integral and get
1
16ZC
p(k+σ)/2(−k+σ)/2(cid:16)Φ(ζ(p))p2 − 12 p−2(cid:17) d p.
This is a Mellin transform of a function compactly supported by C. In virtue of
Theorem 2.2 the integral rapidly decreases in the union of strips Re σ < A.
Proof of the statement c) of Theorem 1.7. We represent ϕ(z) as a sum
of functions in D( C) with simply connected supports. Next, we decompose the
kernel according to Theorem 7.1 and apply the lemma to each summand.
8.4. Continuity.
Corollary 8.2. The map Ja,b is a continuous map from D( C) to L2
even(Λ, κa,b).
Proof. Define the following seminorms on the space of smooth functions on Λ:
pα,N (F ) = sup
λ∈Λ(cid:12)(cid:12)(cid:12)
∂N F
∂σN (1 + λ)α(cid:12)(cid:12)(cid:12),
and the space Y defined by these seminorms. Clearly, our proof provides a continuity
of Ja,b as a map D( C) to Y . It remains to notice that the identical embedding f 7→ f
of Y to L2 is continuous.
If k = 0 and a = 1 (or b = 1), then elements of Wa,b have a pole of order two24
at k = 0, s = 0. In this case we write λ2F instead of F in the definition of the
seminorms.
(cid:3)
24At the same point the spectral density has a zero of order 4.
8.5. Invariance of Wa,b. Consider the difference operators L, L defined above
47
(1.28),
(8.10) LF (k, σ) =
(8.11) LF (k, σ) =
(a + k+σ
2 )(b + k+σ
2 )
(k + σ)(1 + k + σ) (cid:0)F (k + 1, σ + 1) − F (k, σ)(cid:1)+
+
2
(a + −k−σ
)(b + −k−σ
(−k − σ)(1 − k − σ) (cid:0)F (k − 1, σ − 1) − F (k, σ)(cid:1);
)
2
2
(a + −k+σ
)(b + −k+σ
(−k + σ)(1 − k + σ) (cid:0)F (k − 1, σ + 1) − F (k, σ)(cid:1)+
)
2
+
(a + k−σ
2 )(b + k−σ
2 )
(k − σ)(1 + k − σ) (cid:0)F (k + 1, σ − 1) − F (k, σ)(cid:1).
Lemma 8.3. The space Wa,b is invariant with respect to the operators L, L.
Proof. Since F (0,−1) = F (1, 0) = F (−1, 0) = F (0, 1), the expressions
F (k + 1, σ + 1) − F (k, σ)
,
F (k − 1, σ − 1) − F (k, σ)
1 − k − σ
have no poles at k = −1, σ = 0 and k = 1, σ = 0 respectively.
1 + k + σ
Since a function F (k, σ) is even, it can not have a pole of order 1 at k = 0, σ = 0.
New poles of F (k + 1, σ + 1) that are not poles of F (k, σ) are annihilated by the
rational factor in (8.10).
The condition LF (p, q) = LF (q, p) follows from a straightforward calculation. (cid:3)
8.6. Symmetry.
Theorem 8.4. For (a, b) ∈ Π, for F , G ∈ Wa,b we have
(8.12)
hLF, GiL2(ΛC,dKa,b) = hF, LGiL2(Λ,dKa,b).
Corollary 8.5. Operators 1
of D( C).
2 (L + L), 1
2i (L − L) are symmetric on the Ja,b-image
Remark. In fact, the proof uses only properties of F ∈ Wa,b in strips Re σ <
1 + ε. So we can define operators L, L on a space of meromorphic functions in the
strip satisfying an obvious list of conditions.
⊠
8.7. Proof of Theorem 8.4 for the case (a, b) ∈ Πcont. First, we notice that
for pure imaginary σ we have G(k, σ) = G(k,−σ), the last function is meromorphic
and also is contained in Wa,b. Let R(k, σ) be given by (8.1). Then
(8.13) 4π2i hLF, Gi =Xk ZiR(cid:26) (a + k+σ
2 )(b + k+σ
2 )
(k + σ)(1 + k + σ) (cid:0)F (k + 1, σ + 1) − F (k, σ)(cid:1)+
+
2
(a + −k−σ
)(b + −k−σ
(−k − σ)(1 − k − σ) (cid:0)F (k − 1, σ − 1) − F (k, σ)(cid:1)(cid:27)×
)
2
× G(k,−σ) (k − σ)(k + σ)R(k, σ) dσ.
Let us expand the expression in the curly brackets {. . .} as a sum of 4 summands
that include F (k + 1, σ + 1), F (k, σ), F (k− 1, σ− 1), F (k, σ). The whole expression
{. . .} is holomorphic near the contour of integration. The summands have simple
poles on the contour, and we pass to an integration in the sense of principal values.
Let us examine the summand corresponding F (k + 1, σ + 1). We get
48
(8.14)
where
(8.15)
1 + k + σ
k − σ
v.p.ZiR
Xk
eR(k, σ) :=(cid:0)a + 1 + k+σ
= ΓC(cid:0)a + 1 + k+σ
F (k + 1, σ + 1) G(k,−σ)eR(k, σ) dσ,
2 (cid:1)(cid:0)b + 1 + k+σ
2 (cid:12)(cid:12)a + −k+σ
ΓC(cid:0)b + 1 + k+σ
2 (cid:1)R(k, σ) =
(cid:1) ΓC(cid:0)a + −k−σ
2 (cid:12)(cid:12)b + −k+σ
2
2
2
2
(cid:12)(cid:12)a + −k−σ
(cid:1)
(cid:12)(cid:12)b + −k−σ
(cid:1) ΓC(cid:0)b + −k−σ
2
2
(cid:1).
Lemma 8.6. For 0 < a < 1, 0 < b < 1 the integrand in (8.14) has no poles in the
strip −1 < Im σ < 0.
Proof. We enumerate possible (simple) poles of the factors.
a) Factor G(k,−σ). In this case we can have poles if k = 0. Since a < 1, b < 1
the poles 2 − 2a, 2 − 2b are outside our strip. On other hand the pole 2a − 2 (resp.
2b − 2) is contained in the strip if 1/2 < a < 1 (resp. if 1/2 < b < 1).
b) Factor F (k + 1, σ + 1) has a pole in our strip for k = −1 at σ = 2a − 1 (resp.
σ = 2b − 1) if 0 < a < 1/2, (resp. 0 < b < 1/2).
c) Since a > 0, b > 0 the expression eR(k, σ) has no poles in our strip.
However, the poles of G(k,−σ) and of F (k + 1, σ + 1) are zeros of eR(k, σ).
Therefore the product is holomorphic.
(cid:3)
Lemma 8.7. In (8.14), we can change the integration contour to 1 + iR.
Proof. The integrand has no poles between contours iR and 1 + iR, but has
poles on contours, the integral is taken in the sense of principal values. We have
only two such poles, σ = 0 on the contour iR for k = −1 and σ = −1 for k = 0.
Thus the difference between the two integrals is 2π by half of the sum of residues,
i.e.,
2π
2 ((−1 − σ)F (0, σ + 1)G(−1,−σ)(cid:0)a + 1 + −1+σ
+ (0 − σ)F (1, σ + 1)G(0,−σ)(cid:0)a + 1 + σ
2
+
2
(cid:1)(cid:0)b + 1 + −1+σ
2(cid:1)(cid:0)b + 1 + σ
(cid:1)R(−1, σ)(cid:12)(cid:12)(cid:12)σ=0
2(cid:1)R(0, σ)(cid:12)(cid:12)(cid:12)σ=−1(cid:27).
Let us show that the sum is zero. Since F , G are even and satisfy (1.25), we have
F (0, 1) = F (1, 0),
G(−1, 0) = G(0,−1)
By (8.3), we have
The remaining factors give
R(−1, 0) = R(0,−1).
−(a + 1
2 )(b + 1
2 ) and (a + 1
2 )(b + 1
2 ),
i.e., the same expressions with different signs.
(cid:3)
End of the proof of Theorem 8.4 for the case (a, b) ∈ Πcont. Thus we
can replace the integration in (8.14) by the integration over the contour −1 + iR.
49
We change the variables l = k + 1, t = σ + 1 and get
Xl
Next,
and we come to
v.p.ZiR
Xl
F (l, t) G(l − 1,−t + 1)eR(l − 1, t − 1) dt.
l − t
−1 + l + t
v.p.ZiR
eR(l − 1, t − 1) = R(l, t)(cid:0)a + −l−t
F (l, t)(cid:20)(cid:0)a + −l−t
2 (cid:1)(cid:0)b + −l−t
2 (cid:1)
(−l − t)(1 − l − t)
2 (cid:1)(cid:0)b + −l−t
2 (cid:1),
G(l − 1,−t + 1)(cid:21)×
× (l − t)(l + t) R(l, t) dt.
We transform the expression in the big brackets to the form U (l,−t) , where
U (l, t) = (cid:0)a + −l+t
2 (cid:1)(cid:0)b + −l+t
2 (cid:1)
(−l + t)(1 − l + t)
G(l − 1, t + 1).
Thus we finished the transformation of the summand of the (8.13) corresponding to
F (k+1, σ+1). The transformation of the summand corresponding to F (k−1, σ−1)
is similar. The case of the summands F (k, σ) is obvious. We come to the desired
expression.
(cid:3)
8.8. End of the proof of Theorem 8.4. Due to the homographic transfor-
mations, it is sufficient to examine the case a < 0. Let Φ, Ψ ∈ Wa,b. Denote
(8.16)
For (a, b) ∈ Πcont we have
(8.17)
U (a, b; k, σ) := Φ(k, σ) Ψ(k,−σ) κa,b(k, σ).
U (a, b; k, σ) dσ.
4π2ihΦ, ΨiL2(Λ,κa,b) =Xk ZiR
We wish to write the analytic continuation of this expression to the domain (a, b) ∈
Π, a 6 0.
Possible singularities of U as a function in σ in the strip Re σ < 1 are the
following:
-- if b > 1/2, then both functions Φ, Ψ have poles at (k, σ) = (0,±(2 − 2b);
-- κa,b(k, σ) has poles at (k, σ) = (0,±2a).
Due to our restrictions 2b − 2 < 2a < −2a < 2 − 2b.
Thus all summands of (8.17) except 0-th are holomorphic in a < 1 − b.
Lemma 8.8. Fix b. Assume that Φ, Ψ be even rapidly decreasing meromorphic
functions in the strip Re σ < 1 satisfying the condition (1.25) and having poles
only at the points (0,±(2 − 2b). Then the following expression is holomorphic in
the domain a < 1 − b:
(8.18)
γb(a) :=(RiR U (a, b; 0, σ) dσ,
γiR(a) =ZiR
Proof of Lemma 8.8. Denote
RiR U (a, b; 0, σ) dσ + 4πi resσ=2aU (a, b; 0, σ),
U (a, b; 0, σ) dσ, Ξ±(a) := 2πi resσ=±2aU (a, b; 0, σ).
if a > 0,
if a 6 0.
Since U is even in σ, we have Ξ−(a) = −Ξ+(a). Due to the factor (k + σ)(k − σ)
in the Plancherel density, we have Ξ±(0) = 0. Therefore Ξ±(a) are holomorphic in
the disk a < 1 − b.
Consider a contour L on the plane σ ∈ C composed of the ray (−∞, b − 1 + ε],
the upper half of the circle σ = 1− b− ε and the ray [1− b− ε, +∞]. The function
γL(a) :=ZL
U (a, b; 0, σ) dσ
is holomorphic in a for a < 1−b. For Re a > 0 we have γL(a) = γiR(a)−Ξ+(a). For
Re a < 0 we have γL(a) = γiR(a) − Ξ−(a). This gives us the analytic continuation.
Proof of Theorem 8.4 for a < 0. Thus the analytic continuation of
4π2ihΦ, ΨiL2(Λ,κa,b) to the domain a < 0 is given by
γb(a) +Xk6=0ZiR
U (a, b; k, σ) dσ,
50
(cid:3)
i.e., for a < 0 we get 4π2ihΦ, ΨiL2(ΛC,dKa,b).
coincide for a > 0. Therefore they coincide for a < 0.
Now we see that both sides of (8.12) are real analytic in the parameter a and
(cid:3)
9. The operator Ja,b is an isometry
Here we prove the second part of Theorem 1.3.
9.1. Statement.
Lemma 9.1. Let f , g be smooth compactly supported functions on C. Then
hJa,bf, Ja,bgiL2(Λ,dKa,b) = hf, giL2( C,µa,b).
Here a way of a proof is simpler than in Section 6. We show that Ja,b is a
perturbation of a version of the Mellin transform.
9.2. Orthogonality of packets.
Lemma 9.2. Let f , g ∈ D( C). Let supp(f ) ∩ supp(g) = ∅. Then
hJa,bf, Ja,bgiL2
even(ΛC,dKa,b) = 0.
Proof. By Corollary 8.2 the operator Ja,b is continuous as an operator D( C) →
L2(ΛC, dKa,b), by Theorem 8.4 it is symmetric on the image of D( C, µa,b). We
consider the difference operators
and repeat the proof of Lemma 6.3.
1
2 (L + L),
1
2i (L − L)
(cid:3)
9.3. Decomposition of the kernel. Starting from this place we examine the
restriction of Ja,bf to Λ. Recall that the operator Ja,b is defined by the formula
(9.1)
f (z) K(z, λ)µa,b(z) d z.
Ja,bf (λ) =Z C
Decompose the kernel K(z, λ) according to (7.1) with N = 3. We consider λ ∈ Λ,
and therefore we set σ = 0. Denote by ω the factor depending on λ in the front of
the expansion. We have
(9.2)
ω(λ) ω(λ) = κ−1
a,b(λ).
Notice also that the expression in brackets in (7.1) has a singularity at λ = 0.
Denote by Θ(λ) a smooth function, which equals 0 for λ 6 1/3 and 1 for λ > 1/2.
Represent the kernel as
51
K(z, λ) = ω(λ)1 − zb−az−a−b×
t+(z)(cid:17)λ−λ(cid:21)+
+(cid:16) t−(z)
×(cid:26)(cid:20)(cid:16) t+(z)
t−(z)(cid:17)λ−λ
Xk>0, l>0, 16k+l62
−k
+ Θ(λ)(cid:20)(cid:16) t+(z)
t−(z)(cid:17)λ−λ
+(cid:16) t−(z)
t+(z)(cid:17)λ−λ Xk>0, l>0, 16k+l62
where R3(z, λ) is a smooth function in z ∈ C and λ,
λ −l
k! l!
−k
λ
λ
k! l!
λ −l
Ak(√1 − z) Al(√1 − z)+
Ak(−√1 − z) Al(−√1 − z)(cid:21) + R3(z, λ)(cid:27),
R3(z, λ) = O(λ−3)
as λ → ∞.
uniformly on compact subsets of C. The summands corresponding to k = 0, l = 0
are smooth at λ = 0, so we do not multiply them by the patch function Θ(λ).
Next, we change the variable as in (8.4) -- (8.9):
ζ(p) :=
(p + 1)2
4p
and represent the operator Ja,b in the form
Ja,bf (λ) = ω(λ)ZC
where the summands correspond to the summands of the previous formula. We
also denote
γ(p) := 1 − ζ(p)a−b−1/2ζ(p)−1/2ζ′(p)2.
9.4. The main term.
Lemma 9.3. The operator 1
to L2
even(Λ, κa,b).
2π (V +
0,0 + V −
0,0) is a unitary operator from L2(C, µa,b)
f(cid:0)ζ(p)(cid:1)1 − ζ(p)a−b−1/2ζ(p)−1/2ξ′(p)2×
×(cid:26)(cid:20)pλ−λ + (p−1)λ−λ(cid:21)+
p + 1(cid:17)Al(cid:16) p − 1
p + 1(cid:17)+
Ak(cid:16) p − 1
Xk>0, l>0, 16k+l62
p + 1(cid:17)(cid:21) + R(cid:0)ζ(p), λ(cid:1)(cid:27) d p,
Ak(cid:16)−
p + 1(cid:17)Al(cid:16)−
p − 1
p − 1
λ −l
k! l!
−k
λ
λ
−k
λ −l
+ Θ(λ)(cid:20)pλ−λ
+ (p−1)λ−λ Xk>0, l>0, 16k+l62
where C denotes C := C \ {0, 1, −1} as above.
0,0] +X V +
Ja,b = [V +
0,0 + V −
(9.3)
k! l!
k,l +X V −
k,l + Vrem,
It is convenient to split the operator Ja,b into a sum of operators,
52
Proof.
0,0 + V −
(cid:10)(V +
=ZΛ (cid:16)ZC
0,0 + V −
0,0)f, (V +
0,0)g(cid:11)L2(Λ,κ) =
f(cid:0)ζ(p)(cid:1) γ(p)(cid:0)pλ−λ+p−λλ(cid:1) d p(cid:17)(cid:16)ZC
g(cid:0)ζ(q)(cid:1)γ(q)(cid:0)q−λ+λ+qλ−λ(cid:1) d q(cid:17)edλ
(we also applied (9.2). Transform this expression as
(9.4)
ZΛ (cid:16)ZC
f(cid:0)ζ(p)(cid:1) γ(p)p2(cid:0)pλ−1−λ−1 + p−λ−1λ−1(cid:1) d p(cid:17)×
×(cid:16)ZC
g(cid:0)ζ(q)(cid:1) γ(q)q2(cid:0)q−λ−1+λ−1 + qλ−1−λ−1(cid:1) d q(cid:17)edλ.
Now we apply the remark about Mellin transforms of even functions from Subsect.
2.3. Keeping in mind (8.9), we observe that functions f (ζ(p))γ(p)p2 are ×-even.
Therefore both integrals over C in (9.4) are Mellin transforms of even functions,
and we can apply the Plancherel formula for the Mellin transform. We come to
ZC
f(cid:0)ζ(p)(cid:1) g(cid:0)ζ(p)(cid:1) γ(p)2 p4 d p
p2 =
=ZC
f(cid:0)ζ(p)(cid:1) g(cid:0)ζ(p)(cid:1)1 − ζ(p)2a−2b ζ(p)2a+2b−2ζ′(p)2×
×(cid:16)1 − ζ(p)−1ζ(p)−1p2ζ′(p)2(cid:17) d p.
By (8.7) -- (8.8) the expression in the big brackets is 1. Now we return to the variable
z = ζ(p) and get the desired expression
Z C
f (z)g(z) µa,b(z) d z.
9.5. Other terms.
Lemma 9.4. The Hermitian form
(9.5) {f, g} :=(cid:10)Ja,bf, Ja,bg(cid:11)L2(Λ,κa,b) −(cid:10)(V +
on D( C) can be written as
0,0 + V −
0,0)f, (V +
0,0 + V −
0,0)g(cid:11)L2(Λ,κa,b)
(9.6)
{f, g} =ZCZC
K(p, q) f (ζ(p)) g(ζ(q)) d p d q,
where K is a locally integrable function on C × C smooth outside the sets p = q,
p = q−1.
Proof. Expanding Ja,b according to (9.3), we get many summands in (9.5).
We wish to show that each summand can be written as (9.6) with its own K. Let
us start the discussion with the summand
53
f(cid:0)ζ(p)(cid:1) γ(p)p2 pλ−1−λ−1 d p(cid:17)×
q + 1(cid:17)(q−1)λ−1−λ−1 d q(cid:17)edλ.
q − 1
(9.7)
0,0f, V −
hV +
1 − Θ(λ)
λ
(cid:16)ZC
0,1giL2(Λ,κa,b) =ZΛ
×(cid:16)ZC
g(cid:0)ζ(q)(cid:1) γ(q)q2 A1(cid:16)−
F (p) := f(cid:0)ζ(p)(cid:1) γ(p)p2.
G(q) = g(cid:0)ζ(q−1)(cid:1) γ(q−1)q−2 A1(cid:16)−
0,1giL2(Λ,κa,b) =ZΛ
MF (λ) MG(λ)
Thus we get
0,0f, V −
hV +
The integral in the first big bracket is the Mellin transform of the function
The integral in the second big bracket is a complex conjugate to the Mellin trans-
form of
q−1 + 1(cid:17).
q−1 − 1
1 − Θ(λ)
λ
edλ.
Denote by L(p) the inverse Mellin transform of 1−Θ(λ)
. It is easy to see that L(p)
is an integrable function with a unique singularity of the type 1/(1 − p) at p = 1.
We rewrite our integral as
λ
L(pq) F (p) G(q) d p d q,
ZCZC
k′,l′ , where ε, ε′ = ±1, we have similar calculations. Instead
and it has the desired form.
For other pairs V ε
k,l, V ε′
of the boxed factor in (9.7), we get
(9.8)
1 − Θ(λ)
λl+k′
λ
k+l′
.
For k + l + k′ + l′ 6 2 we repeat the same considerations, in these cases inverse
Mellin transforms of the functions (9.8) have (integrable) singularities25 at p = 1
of types
1
,
1 − p
1 − p
1 − p
,
ln1 − p,
1 − p
1 − p
.
If k + l + k′ + l′ > 3, then this expression is integrable in λ, the triple integral
is convergent, we can change the order of integrations and we immediately get an
expression of the form (9.6) with real analytic K(p, q).
For the pairs including Vrem we get absolutely convergent triple integrals and
analytic kernels K(p, q).
9.6. Proof of Lemma 9.1. Now let f , g ∈ D( C) have disjoint supports. Then
both terms in (9.5) are zero (see Lemma 9.2). Therefore the kernel K(p, q) satisfy
the following property:
{f, g} =ZCZC
K(p, q) ϕ(p) ψ(q) dp dq = 0
25We can refer to corresponding formulas for the Fourier transform, see [14], Addendum, Sect.
1.7 (Russian edition) or [15], Sect. B.1.3 (English translation).
54
if ϕ, ψ are ×-even elements D( C) with disjoint supports.
we take a ×-even partition of unity τj with small supports, and decompose
We claim that {f, g} = 0 for any ×-even functions f , g ∈ D( C). To observe this,
{f, g} =Xk,l
{τkf, τlg}.
Clearly, we can make this sum as close to zero as we want by refinement of a
partition of unity. We omit trivial details.
10. Domains of self-adjointness
Thus Ja,b is unitary. Clearly the multiplication operators
f (z) 7→ 1
2 (z + z)f (z),
f (z) 7→ 1
2i (z − z)f (z)
defined on D( C) are essentially self-adjoint in L2(C, µa,b) and commute. Therefore
2 (L + L), 1
the operators 1
2i (L − L) are essentially self-adjoint and commute on
the subspace Ja,bD( C) ⊂ L2
even(ΛC, dKa,b). But Wa,b contains this image. This
establishes Theorem 1.2.a.
Theorem 1.8.a follows from the same argumentation.
References
[1] Andrews G. E., Askey R., Roy, R. Special functions. Cambridge University Press, Cambridge,
1999.
[2] Berezansky, Yu. M.; Sheftel, Z. G.; Us, G. F. Functional analysis. Vol. II. Birkhauser Verlag,
Basel, 1996.
[3] Berezin F. A., Shubin M. A. The Schrodinger equation. Kluwer, Dordrecht, 1991.
[4] Cherednik, I. Inverse Harish-Chandra transform and difference operators. Internat. Math.
Res. Notices 1997, no. 15, 733-750.
[5] Derkachov S. E., Spiridonov V. P. On the 6j-symbols for SL(2, C) group. Theoretical and
Mathematical Physics, 198, Issue 1, 2019, 29-47.
[6] van Diejen J. F., Emsiz E. Difference equation for the Heckman-Opdam hypergeometric func-
tion and its confluent Whittaker limit. Adv. Math. 285 (2015), 1225-1240.
[7] Dotsenko V. S., Fateev V. A., Four-point correlation functions and the operator algebra in
2D conformal invariant theories with central charge C1, Nucl. Phys. B 251 (1985), 691-734.
[8] Duistermaat, J. J., Grunbaum, F. A. Differential equations in the spectral parameter. Comm.
Math. Phys. 103 (1986), no. 2, 177240.
[9] Erd´elyi, A.; Magnus, W.; Oberhettinger, F.; Tricomi, F. G. Higher transcendental functions.
Vol. I, Based, in part, on notes left by Harry Bateman. McGraw-Hill Book Company, Inc.,
New York-Toronto-London, 1953.
[10] Fedoryuk, M. V. Asymptotics: integrals and series. (Russian), Nauka, Moscow, 1987.
[11] Fedoryuk, M. V. Asymptotic analysis. Linear ordinary differential equations. Springer-Verlag,
Berlin, 1993.
[12] Fuglede, B. Commuting self-adjoint partial differential operators and a group theoretic prob-
lem. J. Functional Analysis 16 (1974), 101-121.
[13] Gelfand, I. M., Graev, M. I.; Retakh, V. S. Hypergeometric functions over an arbitrary field.
Russian Math. Surveys 59 (2004), no. 5, 831-905.
[14] Gelfand, I. M., Graev, M. I., Vilenkin, N. Ya. Generalized functions. Vol. 5. Integral geometry
and representation theory. (Russian) Fizmatlit, Moscow, 1962; English transl. New York and
London: Academic Press. XVII, 449 p. (1966).
[15] Gelfand, I. M.; Shilov, G. E. Generalized functions. Vol. I: Properties and operations. Aca-
demic Press, New York-London 1964.
[16] Groenevelt, W. The Wilson function transform. Int. Math. Res. Not. 2003, no. 52, 2779-2817.
[17] Grunbaum F.A., Some nonlinear evolution equations and related topics arising in medical
imaging. In Inverse problems (Montecatini Terme, 1986), 113-141, Lecture Notes in Math.,
1225, Springer, Berlin, 1986.
55
[18] Heckman, G. J.; Opdam, E. M. Root systems and hypergeometric functions. I. Compositio
Math. 64 (1987), no. 3, 329 -- 352.
[19] Hormander, L. The analysis of linear partial differential operators. I. Distribution theory and
Fourier analysis. Springer-Verlag, Berlin, 1983.
[20] Ismagilov R. S., Racah operators for principal series of representations of the group SL(2, C).
Sb. Math., 198:3 (2007), 369-381.
[21] Koornwinder T. H. A new proof of a Paley-Wiener theorem for the Jacobi transform, Ark.
Mat. 13 (1975), 145-159
[22] Koornwinder T. H., Jacobi functions and analysis on noncompact symmetric spaces, in Spe-
cial functions: Group theoretical aspects and applications, Reidel, Dordrecht 1984, pp. 1-85.
[23] Kostyuchenko, A. G., Mityagin, B. S. Positive definite functionals on nuclear spaces. Soviet
Math. Dokl. 1, 1960, 177 -- 180.
[24] Kostyuchenko, A. G., Mityagin, B. S. Positive-definite functionals on nuclear spaces. (Rus-
sian) Trudy Moskov. Mat. Obsc., 9, 1960, 283-316.
[25] Luke, Y. L. The special functions and their approximations, Vol. I. Academic Press, New
York -- London, 1969.
[26] Mimachi, K. Complex hypergeometric integrals. In Konno, H. (ed.) et al., Representation
theory, special functions and Painvlev´e equations. Tokyo: Mathematical Society of Japan,
469-485 (2018).
[27] Molchanov, V. F. Spherical functions on hyperboloids, Mat. Sb. (N.S.), 99(141):2 (1976),
Math. USSR-Sb., 28:2 (1976), 119-139.
[28] Molchanov, V. F. Plancherel formula for hyperboloids, Proc. Steklov Inst. Math., 147 (1981),
63 -- 83.
[29] Molchanov, V. F. Harmonic analysis on homogeneous spaces. in Encyclopaedia Math. Sci., 59,
Representation theory and noncommutative harmonic analysis, II, 1 -- 135, Springer, Berlin,
1995.
[30] Naimark, M. A. Decomposition of a tensor product of irreducible representations of the proper
Lorentz group into irreducible representations. Part I. The case of a tensor product of repre-
sentations of the principal series. (Russian) Trudy Moskov. Mat. Obs. 8 (1959) 121 -- 153.
[31] Naimark, M. A. Decomposition of a tensor product of irreducible representations of the proper
Lorentz group into irreducible representations. Part II The case of a tensor product of repre-
sentations of the principal and complementary series. (Russian) Trudy Moskov. Mat. Obs. 9
(1960) 237-282.
[32] Naimark, M. A. Decomposition of a tensor product of irreducible representations of the proper
Lorentz group into irreducible representations. Part III. The case of a tensor product of
representations of the complementary series. (Russian) Tr. Moskov. Mat. Obs., 10 (1961),
181 -- 216.
[33] Nelson, E. Analytic vectors. Ann. of Math. (2) 70 (1959) 572-615.
[34] Neretin, Yu. A. The index hypergeometric transform and an imitation of the analysis of
Berezin kernels on hyperbolic spaces. Sb. Math. 192 (2001), no. 3-4, 403-432.
[35] Neretin Yu. A., Beta-integrals and finite orthogonal systems of Wilson polynomials, Sb.
Math., 193:7 (2002), 1071-1089.
[36] Neretin, Yu. A. Perturbation of Jacobi polynomials, and piecewise hypergeometric orthogonal
systems. Sb. Math. 197 (2006), no. 11-12, 1607-1633.
[37] Neretin, Yu. A. Index hypergeometric integral transform. (Russian), addendum to Russian
translation of [1] (2013) MCCME, Moscow, 607-624; English version is available as preprint
arXiv:1208.3342.
[38] Neretin, Yu. A. Difference Sturm-Liouville problems in the imaginary direction. J. Spectr.
Theory 3 (2013), no. 3, 237-269.
[39] Neretin, Yu. A. An analog of the Dougall formula and of the de Branges -- Wilson integral.
Preprint arXiv:1812.07341.
[40] Olevskii, M. N. On the representation of an arbitrary function in the form of an integral with
a kernel containing a hypergeometric function. (Russian) Dokl. Akad. Nauk SSSR (N.S.) 69,
(1949). 11-14.
[41] Olver, F. W. J. Asymptotics and special functions. Reprint of the 1974 original. A K Peters,
Ltd., Wellesley, MA, 1997.
[42] Reed M., Simon, B. Methods of modern mathematical physics. I. Functional analysis. Second
edition. Academic Press, Inc., New York, 1980.
56
[43] Titchmarsh, E. C. Eigenfunction expansions with second-order differential operators, Claren-
don Press, Oxford, 1946.
[44] Watson G.N., Asymptotic expansions of hypergeometric functions, Trans. Cambridge Philos.
Soc 22 (1918), 277-308.
[45] Weyl H. Uber gewonliche lineare Differentialgleichungen mit singularen Stellen und ihre
Eigenfunktionen (2 Note), Nachr. Konig. Gess. Wiss. Gottingen. Math.-Phys. (1910), 442 --
467; reprinted in Weyl H. Gesammelte Abhandlungen, vol. 1, Springer-Verlag, Berlin 1968,
pp. 222-247.
[46] Yakubovich S. B. Index transforms. London: World Scientific, 1996.
Vladimir Molchanov
Institute for Mathematics, Natural sciences,
and informational technologies,
Tambov State University, Tambov, Russia.
Yury Neretin
Pauli Institute/c.o. Math. Dept., University of Vienna
&Institute for Theoretical and Experimental Physics (Moscow);
&MechMath Dept., Moscow State University;
&Institute for Information Transmission Problems;
URL: http://mat.univie.ac.at/∼neretin/
|
1707.09307 | 1 | 1707 | 2017-07-28T16:21:11 | Extremal structure and Duality of Lipschitz free spaces | [
"math.FA"
] | We analyse the relationship between different extremal notions in Lipschitz free spaces (strongly exposed, exposed, preserved extreme and extreme points). We prove in particular that every preserved extreme point of the unit ball is also a denting point. We also show in some particular cases that every extreme point is a molecule, and that a molecule is extreme whenever the two points, say $x$ and $y$, which define it satisfy that the metric segment $[x, y]$ only contains $x$ and $y$. The most notable among them is the case when the free space admits an isometric predual with some additional properties. As an application, we get some new consequences about norm-attainment in spaces of vector valued Lipschitz functions. | math.FA | math |
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ
FREE SPACES
LUIS GARC´IA-LIROLA, COLIN PETITJEAN, ANTON´IN PROCH ´AZKA,
AND ABRAHAM RUEDA ZOCA
Abstract. We analyse the relationship between different extremal notions
in Lipschitz free spaces (strongly exposed, exposed, preserved extreme and
extreme points). We prove in particular that every preserved extreme point
of the unit ball is also a denting point. We also show in some particular
cases that every extreme point is a molecule, and that a molecule is extreme
whenever the two points, say x and y, which define it satisfy that the metric
segment [x, y] only contains x and y. The most notable among them is the
case when the free space admits an isometric predual with some additional
properties. As an application, we get some new consequences about norm-
attainment in spaces of vector valued Lipschitz functions.
1. Introduction
The Lipschitz free space F (M) of a metric space M (also known as Arens-
Eells space) is a Banach space such that every Lipschitz function on M admits
a canonical linear extension defined on F (M) (see below for details). This
fundamental linearisation property makes of Lipschitz free spaces a precious
magnifying glass to study Lipschitz maps between metric spaces, and for ex-
ample it relates some well known open problems in the Banach space theory to
some open problems about Lipschitz-free spaces (see [14]). A considerable effort
to study the linear structure and geometry of these spaces has been undergone
by many researchers in the last two or three decades.
Date: July, 2017.
2010 Mathematics Subject Classification. Primary 46B20; Secondary 54E50.
Key words and phrases. Extreme point; Dentability; Lipschitz free; Duality; Uniformly
discrete.
The research of L. Garc´ıa-Lirola was supported by the grants MINECO/FEDER
MTM2014-57838-C2-1-P and Fundaci´on S´eneca CARM 19368/PI/14.
The research of C. Petitjean and A. Proch´azka was supported by the French "Investisse-
ments d'Avenir" program, project ISITE-BFC".
The research of A. Rueda Zoca was supported by a research grant Contratos predoctorales
FPU del Plan Propio del Vicerrectorado de Investigaci´on y Transferencia de la Universidad
de Granada, by MINECO (Spain) Grant MTM2015-65020-P and by Junta de Andaluc´ıa
Grants FQM-0185.
1
2
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
In the present paper we want to focus on the extremal structure of F (M).
The study of extremal structure plays an important role in optimisation (indeed
we obtain some consequences in norm-attainment of Lipschitz maps). It has
probably started in [24]. It is proved for instance in Weaver's book that pre-
served extreme points of the unit ball BF (M ) are always molecules (i.e. measures
of the form mxy := (δ(x) − δ(y))/d(x, y)). Recently Aliaga and Guirao pushed
further this work (see [2]). In particular, answering a question of Weaver, they
showed in the compact case that the extreme points are in fact preserved, and
are exactly the molecules mxy for which there are no points except x and y
in the metric segment [x, y]. They also give a metric characterisation of pre-
served extreme points in full generality, which we prove also here by a different
argument. More results in the same line appeared in [11], where a metric
characterisation of the strongly exposed points is given.
However, the two main questions in this domain remain open:
a) If µ ∈ ext(BF (M )), is µ necessarily of the form µ = mxy for some x 6= y ∈ M?
b) If the metric segment [x, y] does not contain any other point of M than x
and y, is mxy an extreme point of BF (M ) ?
The goal of the present article is to continue the effort in exploring the
extremal structure of F (M) and provide affirmative answers to both previous
questions a) and b) in some particular cases. For instance, we prove that for
the following chain of implications
strongly exposed
(1)
=⇒ denting
(2)
=⇒ preserved extreme
(3)
=⇒ extreme,
the converse of (2) holds true in general (Theorem 2.4) but the converse of
(1) and (3) are both false (Examples 6.4 and 5.4 respectively). However, some
of the previous implications are equivalences in some special classes of metric
spaces. The most notable among them is the case when F (M) admits an
isometric predual with some additional properties. We are thus led to the
study of preduals of free spaces which seems interesting on its own (see Section
3).
The paper is organised as follows. In Section 2 we prove that every preserved
extreme point of BF (M ) is also a denting point in full generality (Theorem 2.4)
and we provide a different proof of the metric charactisation of preserved ex-
treme points given in [2]. We also show that the canonical image δ(M) of
M inside F (M) as well as the set V of molecules are weakly closed in F (M)
(Proposition 2.9 and Proposition 2.13 respectively). Next in Section 3, based
on ([6, 19, 24]), we study under what circumstances F (M) is isometric to a
dual space. We also pin down a distinguished class of preduals, called natu-
ral preduals (see Definition 3.1), which turns out to be of particular interest
in the later sections. In Section 4, we study the extremal structure of spaces
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
3
admitting such a natural predual. In particular, we show under an additional
assumption that the set of extreme points coincides with the set of strongly ex-
posed points (Corollary 4.2). Then in Section 5 we focus on the case when M
is uniformly discrete and bounded. Under this assumption, the question b) has
an affirmative answer (Proposition 5.1), the implication (1) admits a converse
(Proposition 5.3), and the question a) has also an affirmative answer if moreover
F (M) admits a natural predual (Proposition 5.5). In Section 6 we show that
the converse of (3) holds for certain compact spaces since the norm F (M) turns
out to be weak* asymptotically uniformly convex. To finish, in Section 7 we
apply our work to deduce results about norm attainment of Lipschitz functions.
Notation. Throughout the paper we will only consider real Banach spaces.
Given a Banach space X, we will denote by BX (respectively SX ) the closed
unit ball (respectively the unit sphere) of X. We will also denote by X ∗ the
topological dual of X. The notations ext(BX ), exp(BX ), and strexp(BX) stand
for the set of extreme, exposed, and strongly exposed points of BX , respectively
(we refer to [4] for formal definitions and background on this concepts). Given
a norming subspace Y of X, we denote by σ(X, Y ) the topology on X of
pointwise convergence on elements of Y . Given a topological space (T, τ ), we
denote Cτ (T ) the space of continuous functions on T .
Given a metric space M, B(x, r) denotes the closed ball in M centered at
x ∈ M with radius r. We will denote by Lip0(M, X) (or simply Lip0(M) if
X = R) the space of all X-valued Lipschitz functions on M which vanish at a
designated origin 0 ∈ M. We will consider the norm in Lip0(M) given by the
best Lipschitz constant, denoted k · kL. Of particular interest to us is the space
of little-Lipschitz functions,
= 0) .
lip0(M) :=(f ∈ Lip0(M) : lim
ε→0
sup
0<d(x,y)<ε
f (x) − f (y)
d(x, y)
We denote δ the canonical isometric embedding of M into F (M), which is given
by hf, δ(x)i = f (x) for x ∈ M and f ∈ Lip0(M). By a molecule we mean an
element of F (M) of the form
mxy :=
δ(x) − δ(y)
d(x, y)
for x, y ∈ M, x 6= y. The set of all molecules in M will be denoted by V . Note
by passing that V is a norming set for Lip0(M) and so BF (M ) = co(V ). We
will need for every x, y ∈ M, x 6= y, the function
fxy(t) :=
d(x, y)
2
d(t, y) − d(t, x)
d(t, y) + d(t, x)
.
4
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
The properties collected in the next lemma have been proved already in [17].
They make of fxy a useful tool for studying the geometry of BF (M ).
Lemma 1.1. Let x, y ∈ M with x 6= y. We have
d(u,v)
≤
(a) fxy(u)−fxy(v)
(b) fxy is Lipschitz and kfxykL ≤ 1.
(c) Let u 6= v ∈ M and ε > 0 be such that fxy(u)−fxy(v)
max{d(x,u)+d(u,y),d(x,v)+d(v,y)} for all u 6= v ∈ M.
d(x,y)
d(u,v)
> 1 − ε. Then
(1 − ε) max{d(x, v) + d(y, v), d(x, u) + d(y, u)} < d(x, y).
(d) If u 6= v ∈ M and fxy(u)−fxy(v)
d(u,v)
= 1, then u, v ∈ [x, y].
2. General results
Our first goal is to show that every preserved extreme point of BF (M ) is
also a denting point. In order to prove this proposition we need the following
characterisation of preserved extreme points which appears in [15] and that we
state for future reference.
Proposition 2.1 (Proposition 9.1 in [15]). Let X be a Banach space and let
x ∈ BX . The following are equivalent:
(i) x is an extreme point of BX ∗∗.
(ii) The slices of BX containing x are a neighbourhood basis of x for the
weak topology in BX .
(iii) For every sequences {yn} and {zn} in BX such that yn+zn
k·k
→ x we have
2
that yn
w→ x.
It is easy to check that conditions above are also equivalent to the following:
(iii') For every λ ∈ (0, 1) and sequences {yn} and {zn} in BX such that
λyn + (1 − λ)zn
k·k
→ x we have that yn, zn
w→ x.
Next lemma asserts that a net of molecules which converges to a molecule
in the weak topology in fact converges in the norm topology. This lemma will
be useful in the proof of Theorem 2.4 and also in order to show that the set of
molecules is not far from being weakly closed (Proposition 2.13)
Lemma 2.2. Assume {mxαyα} is a net in V which converges weakly to mxy.
Then limα d(xα, x) = 0 and limα d(yα, y) = 0.
Proof. Assume that ε < min{d(x, y), lim supα d(xα, x)}. Consider the map f
given by f (t) = (ε − d(x, t))+ and let g = f − f (0) ∈ Lip0(M). Note that
hg, mxyi = ε
d(x,y) > 0. However,
lim inf
α
hg, mxαyαi = lim inf
α
−f (yα)
d(xα, yα)
≤ 0,
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
5
a contradiction. Therefore, limα xα = x. Analogously we get that limα yα =
y.
(cid:3)
We need the following variation of Asplund–Bourgain–Namioka superlemma
[4, Theorem 3.4.1].
Lemma 2.3. Let A, B ⊂ X be bounded closed convex subsets and let ε > 0.
Assume that diam(A) < ε and that there is x0 ∈ A \ B which is a preserved
extreme point of co(A ∪ B). Then there is a slice of co(A ∪ B) containing x0
which is of diameter less than ε.
Proof. For each r ∈ [0, 1] let
Cr = {x ∈ X : x = (1 − λ)y + λz, y ∈ A, z ∈ B, λ ∈ [r, 1]}.
The proof of the Superlemma says that there is r so that diam(co(A ∪ B) \
Cr) < ε. We will show that x0 /∈ Cr. Thus, any slice separating x0 from Cr
will do the work. To this end, assume that there exist sequences {yn} ⊂ A,
{zn} ⊂ B and λn ⊂ [r, 1] such that x0 = limn(1 − λn)yn + λnz. By extracting
a subsequence, we may assume that {λn} converges to some λ ∈ [r, 1]. Note
that then x0 = limn(1 − λ)yn + λzn. Since x0 is a preserved extreme point, this
implies that {zn} converges weakly to x0 by Proposition 2.1. That is impossible
since x0 /∈ B and B is weakly closed as being convex and closed.
(cid:3)
Theorem 2.4. Every preserved extreme point of BF (M ) is a denting point.
Proof. Let µ be a preserved extreme point of BF (M ), which must be an element
of V . Denote by S the set of weak-open slices of BF (M ) containing µ. Consider
the order S1 ≤ S2 if S2 ⊂ S1 for S1, S2 ∈ S. Using (ii) of Proposition 2.1, every
finite intersection of elements of S contains an element of S and so (S, ≤) is
a directed set. Assume that µ is not a denting point. Then, there is ε > 0 so
that diam(S) > 2ε for every S ∈ S.
We distinguish two cases. Assume first that for every slice S of BF (M ) there
is µS ∈ (V ∩ S) \ B(µ, ε/4). Then {µS} is a net in V which converges weakly to
µ. By Lemma 2.2, it also converges in norm, which is impossible. Thus, there
is a slice S of BF (M) such that diam(V ∩ S) < ε/2. Note that
BF (M ) = co(V ) = co(co(V ∩ S) ∪ co(V \ S))
and so the hypothesis of Lemma 2.3 are satisfied for A = co(V ∩ S), B =
co(V \ S), and µ ∈ A \ B. Then there is a slice of BF (M ) containing µ of
diameter less than ε, a contradiction.
(cid:3)
Theorem 2.4 provides a new proof of the following result given in [11].
Corollary 2.5. Let M be a length space. Then BF (M ) does not have any
preserved extreme point.
6
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
Proof. The space F (M) has the Daugavet property whenever M is a length
space [16]. In particular, every slice of BF (M ) has diameter two. Thus, BF (M )
does not have any denting point.
(cid:3)
During the preparation of this preprint we have learnt that Aliaga and
Guirao [2] characterised metrically the preserved extreme points of free spaces.
In the following pages we provide an alternative proof of their result which
accidentally reproves our Theorem 2.4.
Theorem 2.6. Let M be a metric space and x, y ∈ M. The following are
equivalent:
(i) The molecule mxy is a denting point of BF (M ).
(ii) For every ε > 0 there exists δ > 0 such that every z ∈ M satisfies
(1 − δ)(d(x, z) + d(z, y)) < d(x, y) =⇒ min{d(x, z), d(y, z)} < ε.
Proof of (i)⇒(ii). In fact we are going to show that negation of (ii) implies
that mxy is not a preserved extreme point. Since denting points are trivially
preserved extreme points, this will show at once that mxy is not denting.
So let us fix ε > 0 such that for every n ∈ N there exists zn ∈ M such that
(cid:18)1 −
1
n(cid:19) (d(x, zn) + d(zn, y)) < d(x, y)
but min {d(x, zn), d(y, zn)} ≥ ε. Let µ be a w∗-cluster point of {zn} ({zn} is
clearly bounded). By lower semicontinuity of the norm we have
kδ(x) − µk + kµ − δ(y)k = d(x, y).
If µ ∈ {δ(x), δ(y)}, say µ = δ(x) then by Lemma 2.2 we get that zn → x in
(M, d) which is a contradiction.
Thus µ /∈ {δ(x), δ(y)}. Then
δ(x) − δ(y)
kδ(x) − δ(y)k
=
kδ(x) − µk
kδ(x) − δ(y)k
δ(x) − µ
kδ(x) − µk
+
kµ − δ(y)k
kδ(x) − δ(y)k
µ − δ(y)
kµ − δ(y)k
.
Thus µ is a non-trivial convex combination and so it is not preserved extreme
which concludes the proof of (i)⇒(ii).
(cid:3)
For the proof of the other implication we need a couple of lemmata. The first
of them shows that the diameter of the slices of the unit ball can be controlled
by the diameter of the slices a subset of the ball that is norming for the dual.
Lemma 2.7. Let X be a Banach space and let V ⊂ SX be such that BX =
co(V ). Let f ∈ BX ∗ and 0 < α, ε < 1. Then
diam(S(f, BX, εα)) ≤ 2 diam(S(f, BX, α) ∩ V ) + 4ε.
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
7
Proof. Fix a point x0 ∈ S(f, BX, α) ∩ V . It suffices to show that kx − x0k <
diam(S(f, BX, α) ∩ V ) + 2ε for every x ∈ S(f, BX, εα) ∩ co(V ). To this end,
i=1 λixi, with xi ∈ V ,
let x ∈ BX be such that f (x) > 1 − εα, and x = Pn
Pn
i=1 λi = 1 and λi > 0 for all 1 ≤ i ≤ n. Define
G = {i ∈ {1, . . . , n} : f (xi) > 1 − α}
and B = {1, . . . , n} \ G. We have
1 − εα < f (x) =Xi∈G
λif (xi) +Xi∈B
λif (xi)
≤Xi∈G
λi + (1 − α)Xi∈B
λi = 1 − αXi∈B
λi,
which yields thatPi∈B λi < ε. Now,
λikxi − x0k +Xi∈B
kx − x0k ≤Xi∈G
λikxi − x0k ≤ diam(S(f, BX, α) ∩ V ) + 2ε.
(cid:3)
Lemma 2.8. Let x, y ∈ M, x 6= y such that d(x, y) = 1. For every 0 < ε < 1/4
and 0 < τ < 1 there is a function f ∈ Lip0(M) such that kf kL = 1, hf, mxyi >
1 − 4ετ and satisfying that for every u, v ∈ M, u 6= v, such that if u, v ∈ B(x, ε)
or u, v ∈ B(y, ε), then hf, muvi ≤ 1 − τ .
Proof. Define f : B(x, ε) ∪ B(y, ε) → R by
f (t) =( 1
1
1+4ετ (τ + (1 − τ )d(y, t))
1+4ετ (1 − τ )d(y, t)
if t ∈ B(x, ε),
if t ∈ B(y, ε).
Note that
hf, mxyi = f (x) − f (y) =
1
1 + 4ετ
> 1 − 4ετ.
Moreover, note that if u, v ∈ B(x, ε) or u, v ∈ B(y, ε) then hf, muvi ≤ 1−τ
1+4ετ ≤
1 − τ , so the last condition in the statement is satisfied. Now we compute the
Lipschitz norm of f .
It remains to compute hf, muvi with u ∈ B(x, ε) and
v ∈ B(y, ε). In that case we have
hf, muvi =
≤
τ + (1 − τ )(d(u, y) − d(v, y))
(1 + 4ετ )d(u, v)
1
1 + 4ετ (cid:18) τ
1 − 2ε
+ 1 − τ(cid:19) ≤
≤
τ + (1 − τ )d(u, v)
(1 + 4ετ )d(u, v)
τ (1 + 4ε) + 1 − τ
1 + 4ετ
= 1
8
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
where we are using that (1 − 2ε)−1 ≤ 1 + 4ε since ε < 1/4. This shows that
kf kL ≤ 1. Next, find an extension of f with the same norm. Finally, replace
f with the function t 7→ f (t) − f (0).
(cid:3)
Proof of (ii)⇒ (i) of Theorem 2.6. Now, assume that (ii) holds. We can as-
sume that d(x, y) = 1. Fix 0 < ε < 1/4. We will find a slice of BF (M )
containing mxy of diameter smaller than 32ε. Let δ > 0 be given by property
(ii), clearly we may assume that δ < 1. Let f be the function given by Lemma
2.8 with τ = δ/2.
h(t) =
fxy(t) + f (t)
2
.
It is clear that khkL ≤ 1. Moreover, note that
hfxy, mxyi + hf, mxyi
hh, mxyi =
2
> 1 − 2ετ = 1 − εδ.
Take α = δ/4 and consider the slice S = S(h, BF (M ), α). Note that mxy ∈
S(h, BF (M ), 4εα). We will show that diam(S ∩ V ) ≤ 8ε and as a consequence
of Lemma 2.7 we will get that diam S(f, BF (M ), α) ≤ 32ε. So let u, v ∈ M be
such that muv ∈ S. First, note that hfxy, muvi > 1 − δ, since otherwise we
would have
hh, muvi =
1
2
(hfxy, muvi + hf, muvi) ≤
1
2
(1 − δ) +
1
2
= 1 −
δ
2
< 1 − α
Thus, from the property (c) of the function fxy and the hypothesis (ii) we have
that
min{d(x, u), d(u, y)} < ε
and min{d(x, v), d(y, v)} < ε.
On the other hand,
1 − α < hh, muvi ≤
1
2
+
1
2
hf, muvi
and so hf, muvi > 1 − 2α = 1 − δ
2 = 1 − τ . Thus, we have that u and v
If d(x, v) < ε
do not belong simultaneously to neither B(x, ε) nor B(y, ε).
and d(y, u) < ε, then it is easy to check that hfxy, muvi ≤ 0. So necessarily
d(x, u) < ε and d(y, v) < ε. Now, use the estimate
kmxy − muvk =
≤
kd(u, v)(δ(x) − δ(y)) − d(x, y)(δ(u) − δ(v))k
d(x, y)d(u, v)
k(δ(x) − δ(y)) − (δ(u) − δ(v))k
d(x, y)
+
d(u, v) − d(x, y)kδ(u) − δ(v)k
d(x, y)d(u, v)
≤ 2
d(x, u) + d(y, v)
d(x, y)
≤ 4ε.
Therefore, diam(S ∩ V ) ≤ 8ε.
(cid:3)
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
9
2.1. Weak topology in free spaces. The results which follow are indepen-
dent of the rest of the article. The reader interested only in the extremal
structure of the free spaces can skip until the end of this section.
A simple examples (Examples 5.6 and 5.7 ) show that δ(M) is not necessarily
weak∗ closed when F (M) is a dual space. The next proposition shows that the
situation is different for the weak topology.
Proposition 2.9. Let M be a complete metric space. Then δ(M) ⊂ F (M) is
weakly closed.
The proposition could be deduced more or less easily from Proposition 2.1.6
in [24] but we propose a self-contained proof. For the proof we will need the
next observation (essentially already present in [24]). The weak* closures of
subsets of F (M) below are taken in the bidual F (M)∗∗ = Lip0(M)∗.
Lemma 2.10. Let M be a complete metric space. Let µ ∈ δ(M)
Then there exists ε > 0 such that for all q1, . . . , qn ∈ M we have that
w∗
\ δ(M).
µ ∈ δ M \
B(qi, ε)!w∗
.
n[i=
w∗
Proof. Indeed, otherwise we could find a sequence (qn) ⊂ M such that µ ∈
for every n ≥ 1. It follows that kµ − δ(qn)k ≤ 2−n for every n
δ(B(qn, 2−n))
and thus {qn} is Cauchy. By completeness of M it follows that µ = limn δ(qn) ∈
δ(M). This contradiction proves the claim.
(cid:3)
Proof of Proposition 2.9. It is enough to show that if µ ∈ δ(M)
µ is not w∗-continuous. Indeed, this yields that µ /∈ F (M) and so δ(M)
δ(M)
∩ F (M) = δ(M).
\ δ(M), then
=
w∗
w
w∗
w∗
So let µ ∈ δ(M)
\ δ(M) and let ε > 0 be as in Lemma 2.10. Now
let U be an open neighborhood of 0 in (BLip0(M ), w∗). Since the w∗ topol-
ogy and the topology of pointwise convergence coincide on the ball BLip0(M ),
we may assume that there are x1, . . . , xn ∈ M and α > 0 such that U =
(cid:8)f ∈ BLip0(M ) : f (xi) < α for i = 1, . . . n(cid:9). We define f (x) := dist(x, {x1, . . . , xn}).
We clearly have f ∈ U. Moreover since
µ ∈ δ M \
B(xi, ε)!w∗
,
n[i=1
we have that µ(f ) ≥ ε. Thus µ is not weak*-continuous as V was arbitrary. (cid:3)
We observe the following curious corollary (which also admits an independent
proof by combinatorial methods).
10
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
Corollary 2.11. Let M be a complete metric space. If {xn} ⊂ M is a sequence
such that δ(xn) converges weakly to some µ ∈ F (M), then there exists x ∈ M
such that µ = δ(x) and d(xn, x) → 0.
Proof. The fact that µ = δ(x) follows from Proposition 2.9. For the rest it is
enough to pose f (·) := d(·, x) − d(0, x) and use that d(xn, x) = hδ(xn), f i −
hδ(x), f i → 0.
(cid:3)
Given a complete metric space M and µ ∈ F (M) \ δ(M) there is a weak
neighborhood that separates µ from δ(M). The next example shows that con-
trary to what one might expect, such a neighborhood is not necessarily of the
form {γ ∈ F (M) : hf, γ − µi < ε} for some f ∈ Lip0(M) and ε > 0.
Example 2.12. Let M = [0, 1] with the usual metric and let µ be the Lebesgue
measure on [0, 1]. It is well known and can be shown easily using the Riemann
sums that µ ∈ F (M).
0 f (t)dt.
Now the mean value theorem implies that for every f ∈ Lip0(M) there exists
x ∈ [0, 1] such that hδ(x), f i = hµ, f i.
It acts on Lip0([0, 1]) as follows hµ, f i = R 1
In light of Proposition 2.9, it is also natural to wonder if the set V of molecules
is weakly closed. It is known that 0 is in the weak-closure of V whenever M
is not bi-Lipschitz embeddable in RN (see Lemma 4.2 in [12]). The following
proposition shows that 0 is the only point that we can reach taking the weak-
closure of V .
Proposition 2.13. Let (M, d) be a complete metric space. Then V
w
⊂ V ∪{0}.
Proof. The proof is based on [24, Theorem 2.5.3]. Let us begin with an expla-
nation of this result. To this end, we need to introduce the same kind of map
as the one used in the proof of 3.3. Let M := {(x, y) ∈ M 2 : x 6= y} and
Φ : Lip0(M) → Cb( M )
f (x, y) 7→
f (x) − f (y)
d(x, y)
(here Cb( M ) stands for the continuous and bounded functions on M ).
It is
easy to see that Φ is an isometry. Now let us denote β M the Stone- Cech
compactification of M . As usual, we can canonically identify Cb( M ) with
C(β M) so that we now see Φ as a map from Lip0(M) to C(β M). Thus Φ∗
goes from C(β M)∗ = M(β M) to Lip0(M)∗. According to Weaver, we say
that µ ∈ Lip0(M)∗ is normal if hµ, fii converge to hµ, f i whenever {fi} is a
bounded and decreasing (meaning that fi ≥ fj for i ≤ j) net in Lip0(M) which
w∗-converges to f ∈ Lip0(M). Clearly normality is implied by w∗-continuity.
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
11
Finally, [24, Theorem 2.5.3] asserts that if x ∈ β M with Φ∗δ(x) 6= 0, then
Φ∗δ(x) is normal if and only if x ∈ M .
Let us now prove the assertion of the proposition. Since
w
V
= V
w∗
∩ F (M) = {µ ∈ V
w∗
: µ is w∗-continuous},
\(V ∪ {0}) then µ is not w∗-continuous.
it is enough to show that if µ ∈ V
So let us fix such a µ. We identify, as we may, M with δ( M ) ⊂ M(β M). We
claim that δ( M ) is homeomorphic to (V, w∗). Indeed, it is clear that
w∗
Φ∗↾δ( M ) : δ(x, y) ∈ δ( M ) 7→ mxy ∈ (V, w∗)
is continuous and bijective. The fact that the inverse mapping is also continu-
, w∗) is clearly a
ous follows from Lemma 2.2. So the claim is proved. Now (V
compactification of V . Thus the universal property of the Stone- Cech compacti-
fication provides a surjective extension of Φ∗↾δ( M ) that goes from δ(β M) to V
.
It is easy to check that the latter extension is in fact Φ∗↾δ(β M ) : δ(β M) → V
w∗
w∗
w∗
.
Now consider x ∈ β M such that Φ∗δ(x) = µ 6= 0. Since µ ∈ V
\(V ∪ {0}),
we deduce that x ∈ M . Thus, according to [24, Theorem 2.5.3], Φ∗δ(x) = µ is
not normal and therefore not w∗-continuous. This ends the proof.
(cid:3)
w∗
From the previous proposition, we deduce a result similar to Corollary 2.11.
Corollary 2.14. Let M be a complete metric space. If {mxnyn} ⊂ F (M) is a
sequence of molecules which converges weakly to some µ ∈ F (M), then there
exist x 6= y ∈ M such that µ = mxy and {mxnyn} actually converges in norm
to mxy.
Proof. The fact that µ = mxy follows from Proposition 2.13 and the fact that
{mxnyn} actually converges in norm follows from Lemma 2.2.
(cid:3)
This last corollary should be compared with [1, Theorem 5.2].
Now that we know that V
⊂ V ∪ {0} we get an easy proof of Weaver's
theorem [24] which claims that the preserved extreme points are molecules.
We include it for completeness as it is directly related to the main subject of
this paper.
w
Corollary 2.15. Let M be a complete metric space and let µ be a preserved
extreme point of BF (M ). Then µ = mxy for some x 6= y ∈ M.
Proof. Indeed, we have that co(V ) = BF (M ) and BF (M )
cow∗(V ) = BLip0(M )∗ and so by Milman's theorem ext(BLip0(M )∗) ⊂ V
we get that F (M)∩ext(BLip0(M )∗) ⊂ V
ext(BLip0(M )∗) ⊂ V .
= BLip0(M )∗. Thus
. Finally
and so Proposition 2.13 yields F (M)∩
(cid:3)
w∗
w
w∗
12
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
3. Duality of some Lipschitz free spaces
Many of our results in Sections 4 and 5 use the hypothesis that F (M) admits
an isometric predual which makes δ(M) w∗-closed. Even though for some of
these results we do not know whether this hypothesis is superfluous, we take
the opportunity to study the Lipschitz free spaces which admit such a predual.
Definition 3.1. Let M be a bounded metric space. We will say that a Banach
space X is a natural predual of F (M) if X ∗ = F (M) isometrically and δ(M)
is σ(F (M), X)-closed.
It is obvious that when M is a compact metric space then every isometric
predual of F (M) is natural. We will show in Examples 5.6 and 5.7 that there
are isometric preduals to F (M) which are not natural.
Let us state for the future reference an almost obvious characterisation of
natural preduals.
Proposition 3.2. Let M be a bounded metric space and let X be an isometric
predual of F (M). Then the following are equivalent:
(i) There is a compact Hausdorff topology τ on M such that X ⊂ Lip0(M)∩
Cτ (M).
(ii) δ(M) is σ(F (M), X)-closed.
Proof. We only need to show (i)⇒(ii). To this end, note that the w∗-topology
of F (M) and the τ -topology coincide on δ(M). Indeed, every w∗-open set in
δ(M) is also τ -open since X is made up of τ -continuous functions, so that
the w∗-topology is weaker than τ on δ(M). By compactness of the Hausdorff
topology τ , we have that they agree on δ(M).
(cid:3)
The natural preduals are quite common. In fact, the known constructions of
isometric preduals to F (M) when M is bounded all produce natural preduals.
Indeed, this is the case for Theorem 3.3.3 in [24] as well as Theorem 2.1 in [6]
because of the compactness. In the next theorem we will show that it is also
true for Theorem 6.2 in [19]. We will say that a subspace X of Lip0(M) 1-
separates points uniformly (shortened 1-S.P.U.) if for every x, y ∈ M and every
ε > 0 there is f ∈ X such that f (x) − f (y) = d(x, y) and kf kL < 1 + ε.
Proposition 3.3 (Theorem 6.2 in [19]). Let M be a separable bounded pointed
metric space and let τ be a topology on M so that (M, τ ) is compact. Assume
that X = lip0(M) ∩ Cτ (M) 1-S.P.U. Then X is a natural predual of F (M).
In what follows we provide a slightly different proof of Kalton's result, based
now on Petun¯ın-Pl¯ıcko theorem (see [13, 22]). We recall that this last theorem
asserts that a closed subspace S ⊂ X ∗ of the dual of a separable Banach space
X is an isometric predual of X (that is S∗ = X) if, and only if, S is composed
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
13
of norm-attaining functionals and S separates the points of X. The use of
this theorem to produce preduals to free spaces has become quite common
(see [5, 6, 7, 10] and also our Examples 5.6 and 5.7). The benefit of this proof
is that it avoids the metrizability assumption of the topology τ present in
Kalton's original exposition of this result.
In the proof we will also need the following lemma which restates in a general
framework the first step of Kalton's proof.
Lemma 3.4. Let (M, d) be a metric space such that there is a topology τ on
M and a subset X ⊂ Lip0(M) ∩ Cτ (M) which 1-S.P.U. Then d : (M, τ )2 → R
is l.s.c.
Proof. Let {xα}, {yα} be τ -convergent nets in M with limits x and y, respec-
tively. Given ε > 0, find f ∈ X such that f (y) − f (x) ≥ d(x, y) − ε and
f L = 1. Then
d(x, y) − ε ≤ lim
α
f (yα) − f (xα) ≤ lim inf
α
d(xα, yα)
and the arbitrariness of ε yields the desired conclusion.
(cid:3)
Proof of Proposition 3.3. First of all, according to Lemma 3.4, note that d is
τ -l.s.c.
Now, we need to verify the conditions of Petun¯ın and Pl¯ıcko's theorem. First,
since M is bounded, we see that S is a closed subspace of Lip0(M). Second, S
is separating since it is a lattice and separates the points of M uniformly (see
[19][Proposition 3.4]).
f (xn)−f (yn)
Finally it remains to show that X is made of norm-attaining functionals.
To this end,
let f ∈ SX and take sequences {xn}, {yn} in M such that
limn
d(xn,yn) = 1. Note that inf n d(xn, yn) =: θ > 0 since f ∈ lip0(M). By
the compactness of (M, τ ) and the boundedness of d, we can find subnets {xα}
τ→ y and d(xα, yα) → C > 0.
of {xn} and {yα} of {yn} such that xα
Then,
τ→ x, yα
1 = lim
α
f (xα) − f (xα)
d(xα, yα)
→
f (x) − f (y)
C
≤
f (x) − f (y)
d(x, y)
.
Thus X is made up of norm-attaining functionals.
To conclude, we get that S is a natural predual by just applying Proposition
(cid:3)
3.2.
The next proposition testifies that Kalton's theorem is the only way to build
a natural predual if the predual is moreover required to be a subspace of little
Lipschitz functions.
Proposition 3.5. Let M be a bounded metric space and let X ∗ = F (M) be
a natural predual such that X ⊆ lip0(M). Then there exists a topology τ on
14
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
M such that (M, τ ) is compact, the metric d : (M, τ )2 → R is l.s.c. and
X = lip0(M) ∩ Cτ (M).
Proof. We put τ := {δ−1(U) : U ∈ σ(F (M), X)}. Since δ(M) is σ(F (M), X)-
closed and bounded, (M, τ ) is compact. Recall that d(x, y) = kδ(x) − δ(y)k
and k·k is σ(F (M), X)-lsc, so the metric d is τ -lsc. Since
X = {x∗ ∈ F (M)∗ : x∗ is σ(F (M), X) − continuous}
and X ⊂ lip0(M), we get that X ⊆ lip0(M, d) ∩ Cτ (M) =: Y . This means
that σ(F (M), Y ) is stronger than σ(F (M), X). On the other hand, Propo-
sition 3.3 yields that Y ∗ = F (M). Therefore, by compactness, σ(F (M), X)
and σ(F (M), Y ) coincide on BF (M ). As a consequence of Banach-Dieudonn´e
theorem, they coincide on F (M). This means that
X = {x∗ ∈ F (M)∗ : x∗ is σ(F (M), X) − continuous}
= {x∗ ∈ F (M)∗ : x∗ is σ(F (M), Y ) − continuous} = Y.
(cid:3)
But one should be aware that not all natural preduals are contained in the
space of little Lipschitz functions.
Example 3.6. Let M =(cid:8) 1
n : n ∈ N(cid:9) ∪ {0} with the distance comming from
the reals. Then it is well known that F (M) is isometrically isomorphic to ℓ1.
Further we know (Theorem 2.1 in [6]) that lip0(M) is isometrically a predual.
Since M is compact, every predual is natural. So our Proposition 3.5 and the
fact that M is compact show that any isometric predual of ℓ1 which is not
isometric to lip0(M) intersects the complement of lip0(M).
Note that Lip0(M) = lip0(M) when M is uniformly discrete. This observa-
tion and the previous results yield the following corollary.
Corollary 3.7. Let (M, d) be a uniformly discrete bounded separable metric
space with 0 ∈ M. Let X be a Banach space. Then it is equivalent:
(i) X is a natural predual of F (M).
(ii) There is a Hausdorff topology τ on M such that (M, τ ) is compact, d is
τ -l.s.c. and X = Lip0(M, d) ∩ Cτ (M) equipped with the norm k·kL.
Proof. (ii) ⇒ (i) Given x, y ∈ M, x 6= y, define f : {x, y} → R by f (x) = 0
and f (y) = d(x, y). By Matouskova's extension theorem [20], there is f ∈
Lip0(M) ∩ Cτ (M) extending f such that f L = 1. Thus, the hypotheses of
Proposition 3.3 are satisfied.
The implication (i) ⇒ (ii) is contained in Proposition 3.5.
(cid:3)
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
15
In what follows we are going to develop yet another sufficient condition for
an isometric predual to be natural with the goal to show that certain preduals
constructed by Weaver in [23] are natural.
Proposition 3.8. Let M be a uniformly discrete, bounded, separable metric
space and let X ⊂ Lip0(M) be a Banach space such that X ∗ = F (M) isometri-
cally. If for every x ∈ M \ {0} the indicator function 1{x} belongs to X, then
X is a natural predual of F (M). Moreover 0 is the unique accumulation point
of (δ(M), w∗) and X is isomorphic to c0.
The proof will be based on the following general fact.
Lemma 3.9. Let X, Y be Banach spaces such that X ∗ = Y isometrically, Y
admits a bounded Schauder basis {un} and the biorthogonal functionals {u∗
n}
belong to X. Then un → 0 weakly*.
Proof. We will show that every subsequence of {un} admits a further subse-
quence that converges weakly* to 0. So let us consider such subsequence. By
the weak* compactness and separability, it admits a weak* convergent subse-
quence, let us call it {un} again. So we have un → u ∈ X weakly*. But this
means that for every m ∈ N we have
hu∗
m, uni = hu∗
m, ui .
0 = lim
n→∞
Thus u = 0.
(cid:3)
Proof of Proposition 3.8. Since M is bounded and uniformly discrete, the se-
quence {δ(x)}x∈M \{0} is a Schauder basis which is equivalent to the unit vector
basis of ℓ1. The biorthogonal functionals are exactly the indicator functions
1{x} for x 6= 0. Applying Lemma 3.9 we get that δ(M) is weak* closed and
that 0 is the unique w∗-accumulation point of δ(M). Let τ be the restriction of
the w∗-topology to M. Now Corollary 3.7 yields that X = Lip0(M) ∩ Cτ (M).
But, since M is bounded an uniformly discrete, we have that Lip0(M) are
just all bounded functions that vanish at 0. It follows immediately that X =
c0(M \ {0}).
(cid:3)
Remark 3.10. In [23], Weaver proved a duality result for rigidly locally compact
metric spaces. We recall that a locally compact metric space is said to be
rigidly locally compact (see the paragraph before Proposition 3.3 in [23]) if for
every r > 1 and every x ∈ M, the closed ball B(x, d(0,x)
) is compact. The
duality result of Weaver in particular implies that for a separable uniformly
discrete bounded metric space M which is rigidly compact, the space
r
X =(cid:26)f ∈ Lip0(M) :
f (·)
d(·, 0)
∈ C0(M)(cid:27)
16
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
is an isometric predual of F (M). Here C0(M) denotes the set of continuous
functions which are arbitrarily small out of compact sets. Since it is obvious
that the indicator functions 1{x} belong to X, Proposition 3.8 implies that X
is a natural predual of F (M) and that X is isomorphic to c0. This shows that
in the case of uniformly discrete bounded spaces, Corollary 3.7 covers the cases
in which Weaver's result ensures the existence of a predual.
Moreover, there is a metric space which satisfies the hypotheses of Corollary
3.7 and which is not rigidly locally compact.
Example 3.11. Let us consider the metric space M = {0, 1} × N equipped
with the following distance: d((0, n), (1, m)) = 2 for n, m ∈ N, and if n 6= m
we have d((0, n), (0, m)) = 1 and d((1, n), (1, m)) = 1. Then M satisfies the
assumptions of Corollary 3.7.
Indeed, declare (0, 1) to be the accumulation
point of the sequence {(0, n)}, (1, 1) to be the accumulation point of the se-
quence {(1, n)}, and then declare all the other points isolated. Now indepen-
dently of the choice of the distinguished point 0M , M is not rigidly locally
compact. For instance, say that 0M = (0, n). Then for every r > 1, the ball
B((1, 1), d(0M, (1, 1))/r) = B((1, 1), 2/r) contains all the elements of the form
(1, m) with m ∈ N. Consequently the considered ball is not compact, which
proves that M is not rigidly locally compact.
4. Extremal structure for spaces with natural preduals
We are going to focus now on the extreme points in the free spaces that
admit a natural predual. Assuming moreover that the predual is a subspace
of little Lipschitz functions we get an affirmative answer to one of our main
problems. Note that this is an extension of Corollary 3.3.6 in [24], where it is
obtained the same result under the assumption that M is compact.
Proposition 4.1. Let M be a bounded metric space. Assume that there is a
subspace X of lip0(M) which is a natural predual of F (M). Then
ext(BF (M )) ⊂ {mxy : x, y ∈ M, x 6= y}.
Proof. By the separation theorem we have that BF (M ) = cow∗(V ). Thus, ac-
w∗
cording to Milman theorem (see [8, Theorem 3.41]), we have ext(BF (M )) ⊂ V
.
So let us consider γ ∈ ext(BF (M )). Take a net {mxα,yα} in V which w∗-converges
to γ. By w∗-compactness of δ(M), we may assume (up to extracting subnets)
that {δ(xα)} and {δ(yα)} converge to some δ(x) and δ(y) respectively.
Next, we claim that we may also assume that {d(xα, yα)} converges to C > 0.
Indeed, since M is bounded, we may assume up to extract a further subnet
that {d(xα, yα)} converges to C ≥ 0. By assumption, there is f ∈ X such that
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
17
hf, γi > kγk/2 = 1/2. Since f ∈ lip0(M), there exists δ > 0 such that whenever
z1, z2 ∈ M satisfy d(z1, z2) ≤ δ then we have f (z1) − f (z2) ≤ 1
2 d(z1, z2). Since
lim
α
hf, mxα,yαi = hf, γi >
1
2
,
there is α0 such that hf, mxα,yαi > 1/2 for every α > α0. Thus d(xα, yα) > δ for
α > α0, which implies that C ≥ δ > 0. Summarizing, we have a net {mxα,yα}
. So, by uniqueness of the limit, γ = δ(x)−δ(y)
which w∗-converges to δ(x)−δ(y)
.
Since γ ∈ ext(BF (M )) ⊂ SF (M ), we get that C = d(x, y) and so γ = mxy.
(cid:3)
C
C
We have learned that a weaker version (for compact M) of the following
proposition appears in the preprint [2] for compact metric spaces. Our ap-
proach, which is independent of [2], also yields a characterisation of exposed
points of BF (M ).
Corollary 4.2. Let M be a bounded separable metric space. Assume that there
is a subspace X of lip0(M) which is a natural predual of F (M). Then given
µ ∈ BF (M ) the following are equivalent:
(i) µ ∈ ext(BF (M )).
(ii) µ ∈ exp(BF (M )).
(iii) There are x, y ∈ M, x 6= y, such that [x, y] = {x, y} and µ = mxy.
Proof. (i)⇒(iii) follows from Proposition 4.1. Moreover, (ii)⇒(i) is clear, so it
only remains to show (iii)⇒(ii). To this end, let x, y ∈ M, x 6= y, be so that
[x, y] = {x, y}. Consider
A = {µ ∈ BF (M ) : hfxy, µi = 1}.
We will show that A = {mxy} and so mxy is exposed by fxy in BF (M ). Let
µ ∈ ext(A). Since A is an extremal subset of BF (M ), µ is also an extreme point
of BF (M ) and so µ ∈ V ∩ A. Recall that if hfxy, mu,vi = 1 then u, v ∈ [x, y],
therefore V ∩ A = {mxy}. Thus ext(A) ⊂ {mxy}. Finally note that A is a
closed convex subset of BF (M ) and so A = co(ext(A)) = {mxy} since the space
F (M) has (RNP) as being a separable dual.
(cid:3)
It is proved in Aliaga and Guirao's paper [2] that if (M, d) is compact, then
a molecule mxy is extreme in BF (M ) if and only if it is preserved extreme if and
only if [x, y] = {x, y}. Thus, if lip0(M) 1-S.P.U. (and thus F (M) = lip0(M)∗),
Proposition 4.1 and Aliaga and Guirao's result provides a complete description
of the extreme points: they are the molecules mxy such that [x, y] = {x, y}. It
is possible to obtain the same kind of complete descriptions in some different
settings as it is proved in the following result (see also Section 5).
Proposition 4.3. Let (M, d) be a metric space for which there is a Hausdorff
topology τ such that (M, τ ) is compact and d : (M, τ )2 → R is l.s.c. Let
18
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
0 < p < 1 and let (M, dp) be the p-snowflake of M. Then given µ ∈ BF (M ) the
following are equivalent:
(i) µ ∈ ext(BF (M,dp)).
(ii) µ ∈ strexp(BF (M,dp)).
(iii) There are x, y ∈ M, x 6= y, such that µ = mxy.
Observe that under the hypotheses above it is not necessarily true that F (M)
is a dual space, but F (M, dp) already is.
Proof. (iii) =⇒ (ii). Let us fix x 6= y ∈ M. Since 0 < p < 1, it is readily
seen that [x, y] = {x, y}. Moreover it is proved in [24, Proposition 2.4.5] that
there is a peaking function at (x, y). Thus mxy is a strongly exposed point
([11, Theorem 4.4]). The implication (ii) =⇒ (i) is obvious. To finish, the
implication (i) =⇒ (iii) follows directly from Proposition 4.1 and the fact that
[x, y] = {x, y} for every x 6= y ∈ M.
(cid:3)
Next we will show that the extremal structure of a free space has impact
on its isometric preduals. If a metric space M is countable and satisfies the
assumptions of Proposition 4.1, then ext(BF (M )) is also countable. Therefore,
any isometric predual of F (M) is isomorphic to a polyhedral space by a theorem
of Fonf [9], and so it is saturated with subspaces isomorphic to c0. This applies
for instance in the following cases.
Corollary 4.4. Let M be a countable compact metric space. Then any iso-
metric predual of F (M) (in particular lip0(M)) is isomorphic to a polyhedral
space.
Corollary 4.5. Let (M, d) be a uniformly discrete bounded separable metric
space such that F (M) admits a natural predual. Then any isometric predual of
F (M) is isomorphic to a polyhedral space.
5. The uniformly discrete case
We have already witnessed that in the class of uniformly discrete and bounded
metric spaces, many results about F (M) become simpler. Yet another example
of this principle is the following main result of this section.
Proposition 5.1. Let (M, d) be a bounded uniformly discrete metric space.
Then a molecule mxy is an extreme point of BF (M ) if and only if [x, y] = {x, y}.
Also we will need the following observation, perhaps of independent inter-
est: Since a point x ∈ BX is extreme if and only if x ∈ ext(BY ) for every
2-dimensional subspace Y of X, the extreme points of BF (M ) are separably
determined. Let us be more precise.
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
19
Lemma 5.2. Assume that µ0 ∈ BF (M ) is not an extreme point of BF (M ). Then
there is a separable subset N ⊂ M such that µ0 ∈ F (N) and µ0 /∈ ext(BF (N )).
n} of finitely supported measures such that µi = limn→∞ νi
Proof. Write µ0 = 1
{νi
Let N = {0} ∪ supp{νi
an isometry and νi
subspace of F (M). Thus µ0, µ1, µ2 ∈ F (N) and so µ0 /∈ ext(BF (N )).
2 (µ1 + µ2), with µ1, µ2 ∈ BF (M ). We can find sequences
n for i = 0, 1, 2.
n}. Note that the canonical inclusion F (N) ֒→ F (M) is
n ∈ F (N) for each n, i. Since F (N) is complete, it is a closed
(cid:3)
Proof of Proposition 5.1. Let mxy be a molecule in M such that [x, y] = {x, y}
and assume that mxy /∈ ext(BF (M )). By Lemma 5.2, we may assume that M
is countable. Write M = {xn : n ≥ 0}. Let {en : n ≥ 1} be the unit vector
basis of ℓ1. It is well known that the map δ(xn) 7→ en for n ≥ 1 defines an
isomorphism from F (M) onto ℓ1. Thus {δ(xn) : n ≥ 1} is a Schauder basis for
F (M).
Now, let x, y ∈ M, x 6= y be such that [x, y] = {x, y}. Assume that mxy =
n=1 anδ(xn). Fix n ∈ N such that
1
2(µ + ν) for µ, ν ∈ BF (M ) and write µ =P∞
xn /∈ {x, y}. Then, there is εn > 0 such that
(1 − εn) (d(x, xn) + d(xn, y)) ≤ d(x, y).
Let gn = fxy + εn1{xn}, which is an element of Lip0(M) since M is uniformly
discrete. We will show that gnL ≤ 1. To this end, take u, v ∈ M, u 6= v.
Since kf kL ≤ 1, it is clear that hgn, muvi ≤ 1 if u, v 6= xn. Thus we may
assume v = xn. Therefore (c) in Lemma 1.1 yields that hfxy, muvi ≤ 1 − εn
and so hgn, muvi ≤ 1. Exchanging the roles of u and v, we get that gnL ≤ 1.
Moreover, note that
1 = hgn, mxyi =
1
2
(hgn, µi + hgn, νi) ≤ 1
and so hgn, µi = 1. Analogously we show that hfxy, µi = 1. Thus an =
h1{xn}, µi = 0. Therefore µ = aδ(x) + bδ(y) for some a, b ∈ R. Finally, let
f1(t) := d(t, x) − d(0, x) and f2(t) := d(t, y) − d(0, x). Then fiL = 1 and
hfi, mxyi = 1, so we also have hfi, µi = 1 for i = 1, 2. It follows from this that
a = −b = 1
d(x,y) , that is, µ = mxy. This implies that mxy is an extreme point
of BF (M ).
(cid:3)
Next we show that preserved extreme points are automatically strongly ex-
posed for uniformly discrete metric spaces. Notice that, contrary to other
results in this section, no boundedness assumption is needed.
Proposition 5.3. Let M be a uniformly discrete metric space. Then every
preserved extreme point of BF (M ) is also a strongly exposed point.
20
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
Proof. Let x, y ∈ M such that mxy is a preserved extreme point of BF (M ).
Assume that mxy is not strongly exposed. By Theorem 4.4 in [11], the pair
(x, y) enjoys property (Z). That is, for each n ∈ N we can find zn ∈ M \ {x, y}
such that
d(x, zn) + d(y, zn) ≤ d(x, y) +
1
n
min{d(x, zn), d(y, zn)}.
Thus,
(1 − 1/n)(d(x, zn) + d(y, zn)) ≤ d(x, y)
so it follows from condition (ii) in Theorem 2.6 that min{d(x, zn), d(y, zn)} → 0.
Since M is uniformly discrete, this means that {zn} is eventually equal to either
x or y, a contradiction.
(cid:3)
Aliaga and Guirao proved in [2] that, in the case of compact metric spaces,
every molecule which is an extreme point of BF (M ) is also a preserved extreme
point. However, that result is no longer true for general metric spaces, as the
following example shows.
Example 5.4. Consider the sequence in c0 given by x1 = 2e1, and xn = e1+(1+
1/n)en for n ≥ 2, where {en} is the canonical basis. Let M = {0}∪{xn : n ∈ N}.
This metric space is considered in [2, Example 4.2], where it is proved that the
molecule m0x1 is not a preserved extreme point of BF (M ). Let us note that this
fact also follows easily from Theorem 2.6. Moreover, by Proposition 5.1 we
have that m0x1 ∈ ext(BF (M )).
On the other hand, if we restrict our attention to uniformly discrete bounded
metric spaces satisfying the hypotheses of the duality result, then all the fami-
lies of distinguished points of BF (M ) that we have considered coincide.
Proposition 5.5. Let (M, d) be a uniformly discrete bounded metric space such
that F (M) admits a natural predual. Then for µ ∈ BF (M ) it is equivalent:
(i) µ ∈ ext(BF (M )).
(ii) µ ∈ strexp(BF (M )).
(iii) There are x, y ∈ M, x 6= y, such that µ = mxy and [x, y] = {x, y}.
Proof. (i) ⇒ (iii) follows from Proposition 4.1. Moreover, (ii)⇒(i) trivially.
Now, assume that µ = mxy with [x, y] = {x, y}. We will show that the pair
(x, y) fails property (Z) and thus µ is a strongly exposed point. Assume, by
contradiction, that there is a sequence {zn} in M such that
d(x, zn) + d(y, zn) ≤ d(x, y) +
1
n
min{d(x, zn), d(y, zn)}.
and so
(1 − 1/n)(d(x, zn) + d(y, zn)) ≤ d(x, y).
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
21
The compactness with respect to the w∗-topology ensures the existence of a
w∗-cluster point z of {zn} (M and δ(M) ⊂ F (M) being naturally identified).
Now, by the lower semicontinuity of the distance, we have
d(x, z) + d(y, z) ≤ lim inf
n→∞
(1 − 1/n)(d(x, zn) + d(y, zn)) ≤ d(x, y).
Therefore, z ∈ [x, y] = {x, y}. Suppose z = x. Denote θ = inf{d(u, v) : u 6=
v} > 0. The lower semicontinuity of d yields
θ + d(x, y) ≤ lim inf
n→∞
≤ lim inf
n→∞
(1 − 1/n)(θ + d(y, zn))
(1 − 1/n)(d(x, zn) + d(y, zn)) ≤ d(x, y),
which is impossible. The case z = y yields a similar contradiction. Thus the
pair (x, y) has property (Z).
(cid:3)
We now give some examples in which the preduals of F (M) have interesting
properties. The first one is a uniformly discrete and bounded metric space
M such that F (M) is isometric to a dual Banach space but cannot admit a
natural predual. This example comes from [2][Example 4.2] and has already
been introduced in Example 5.4.
Example 5.6. Consider the sequence in c0 given by x0 = 0, x1 = 2e1, and
xn = e1 + (1 + 1/n)en for n ≥ 2, where {en} is the canonical basis. Let
M = {0} ∪ {xn : n ∈ N}. Then
a) F (M) does not admit any natural predual.
b) the space X = {f ∈ Lip0(M) : lim f (xn) = f (x1)/2} satisfies X ∗ = F (M).
Our Corollary 3.7 guarantees that in order to prove a) it is enough to show
that there is no compact topology τ on M such that d is τ -l.s.c. Assume that
τ is such a topology. Then the sequence {xn} admits a τ -accumulation point
x ∈ M. Since d is τ -l.s.c. we get that x ∈ B(0, 1) ∩ B(x1, 1). But this is a
contradiction as the latter set is clearly empty.
For the proof of b) we will employ the theorem of Petun¯ın and Pl¯ıchko. The
space X is a clearly separable closed subspace of F (M)∗. Further, a simple
case check shows that for any x 6= y ∈ M, y 6= 0, the function f (x) = 0, f (y) =
d(x, y) can be extended as an element of X without increasing the Lipschitz
norm. Thus since X is clearly a lattice, Proposition 3.4 of [19] shows that X
is separating. Finally, if f ∈ X and
f (xnk ) − f (xmk )
d(xnk, xmk )
→ kf kL
then without loss of generality the sequence {mk} does not tend to infinity.
Passing to a subsequence, we may assume that it is constant, say mk = m for
22
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
all k ∈ N. If {nk} does not tend to infinity, then f (xi)−f (xm)
i 6= m. Otherwise, since f ∈ X, we have
d(xi,xm) = kf kL for some
f (xnk) − f (xm)
d(xnk, xm)
→
f (x1)
2 − f (m)
d(x1, xm)
.
So in this case the norm is attained at
d(x1,xm) (δ(x1)/2 − δ(xm)) ∈ BF (M ). It
follows that every f ∈ X attains its norm. Thus by the theorem of Petun¯ın
and Pl¯ıchko, X ∗ = F (M).
1
Next we show that F (M) can actually have both natural and non-natural
preduals.
Example 5.7. Let M = {0} ∪ {1, 2, 3, . . .} be a graph such that the edges
are couples of the form (0, n) with n ≥ 1. Let d be the shortest path dis-
tance on M. Then it is obvious and well known that F (M) is isometric to
ℓ1. Moreover F (M) admits both natural and non-natural preduals.
Indeed,
an example of a natural predual is X = {f ∈ Lip0(M) : lim f (n) = f (1)} (this
is immediate using Corollary 3.7). An example of a non-natural predual is
Y = {f ∈ Lip0(M) : lim f (n) = −f (1)}. We leave to the reader the verifica-
tion of the hypotheses of the theorem of Petunin and Plichko.
Our last example shows that there are uniformly discrete bounded metric
spaces such that their free space does not admit any isometric predual at all.
Such observation is relevant to the open problem whether F (M) has (MAP) for
every uniformly discrete and bounded metric space M (see also Problem 6.2
in [14]). Using a well known theorem of Grothendieck, in order to get an
affirmative answer it would be enough to show that F (M) is isometrically a
dual space. Our example shows that such a proof cannot work in general.
Nevertheless, for M in this example, F (M) enjoys the (MAP).
Example 5.8. Let M = {0}∪{1, 2, 3, . . .} ∪{a, b} with the following distances:
d(0, n) = d(a, n) = d(b, n) = 1 + 1/n,
d(a, b) = d(0, a) = d(0, b) = 2, and
d(n, m) = 1
for n, m ∈ {1, 2, 3, . . .}. Then F (M) is not isometrically a dual space.
Indeed, let us assume that F (M) = X ∗. Then some subsequence of {δ(n)} is
w∗-convergent to some µ ∈ F (M). We have kµ − δ(a)k ≤ 1 and kµk ≤ 1. But
Proposition 5.1 implies that δ(a)/2 is an extreme point of BF (M ). This means
that BF (M )(0, 1) ∩ BF (M )(δ(a), 1) = {δ(a)/2}. Similarly for δ(b)/2. Hence
δ(a)/2 = µ = δ(b)/2. Contradiction.
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
23
Let us now prove that F (M) has the (MAP). Let Mn := {0, a, b, 1, . . . , n}
and define fn: M → Mn by fn(x) = x if x ∈ Mn and f (x) = n otherwise.
The function fn is obviously a retraction from M to Mn. Moreover a simple
computation leads to kfnkL ≤ 1 + 1/n. Let us denote fn: F (M) → F (Mn) the
linearisation of fn which is in fact a projection of the same norm: k fnk ≤ 1+1/n.
Then define Pn := (1 + 1/n)−1 fn. Obviously, kPnk ≤ 1, Pn is of finite rank and
kPnγ − γk → 0 for every γ ∈ F (M). Thus F (M) has the (MAP).
6. Compact metric spaces
In this section we focus on the case in which M is a compact metric space
and F (M) is the dual of lip0(M). Recall that in this case all extreme points of
BF (M ) are molecules by Corollary 3.3.6 in [24]. We will show that indeed F (M)
satisfies a geometrical property, namely being weak* asymptotically uniformly
convex, which implies in particular that the norm and the weak* topologies
agree in SF (M ) and so every extreme point of the closed ball is also a denting
point.
If X is a separable Banach space then the modulus of weak*-asymptotic
uniform convexity of X ∗ can be computed as follows ([3]):
δ
∗
X ∗(t) = inf
x∗∈BX ∗
inf
w∗
→0
x∗
n
nk≥t
kx∗
lim inf
n→∞
kx∗ + x∗
nk − 1.
Recall that X ∗ is said to be weak*-asymptotically uniformly convex (weak*-
AUC for short) if δ
∗
X ∗(t) > 0 for each t > 0.
Proposition 6.1. Let M be a compact metric space. Assume that lip0(M) is
1-norming. Then F (M) is weak*-AUC.
For the proof we need the following easy lemma.
Lemma 6.2. Let {x∗
F ⊂ X ∗ be a finite dimensional subspace. Then lim inf n→∞ d(x∗
n} ⊂ X ∗ be a weak*-null sequence such that kx∗
nk ≥ 1 and
n, F ) ≥ 1
2.
Proof of Proposition 6.1. We will use the same arguments as in the proof of
Proposition 8 in [21]. Fix t > 0 and take γ ∈ SF (M ) and a weak*-null sequence
{γn} such that γn ≥ t for every n ∈ N. We will prove that
(6.1)
lim inf
n→∞
γ + γn ≥ 1 +
t
2
.
We may assume that γ is finitely supported. Pick f ∈ lip0(M) with f L =
1 and hf, γi > 1 − ε. Take θ > 0 such that supd(x,y)≤θ f (x) − f (y) ≤
εd(x, y). Pick δ < εθ
2(1+ε) . By compactness, there exists a finite subset E ⊂ M
containing the support of γ and such that supy∈M d(y, E) < δ. We have
24
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
lim inf n→∞ d(γn/t, F (E)) ≥ 1
2 by Lemma 6.2. Now, by Hahn-Banach theo-
rem, there exist a sequence {fn} ⊂ (1 + ε)BLip0(M ) such that fnE = 0 and
lim inf n→∞hfn, γni ≥ t
2 . Consider gn = f + fn. By distinguishing the cases
d(x, y) < θ and d(x, y) > θ, one can show that gnL ≤ 1 + ε. Now we have
lim inf
n→∞
γ + γn ≥ lim inf
n→∞
1
1
1 + ε
hgn, γ + γni
lim inf
n→∞
(hf, γi + hf, γni + hfn, γi + hfn, γni)
=
≥
1 + ε
1
1 + ε
(1 − ε +
t
2
− ε)
since γn
∗
F (M )(t) ≥ 1
δ
w∗
→ 0 and f ∈ lip0(M). Letting ε → 0 proves (6.1). It follows that
(cid:3)
2t and so F (M) is weak*-AUC.
It is well-known and easy to show that if X ∗ is weak*-AUC then then every
point of the unit sphere has weak*-neighbourhoods of arbitrarily small diameter.
This fact and the Choquet's lemma yield that if x∗ ∈ ext(BX ∗) then then there
are weak*-slices of BX ∗ containing x∗ of arbitrarily small diameter. That is,
every extreme point of BX ∗ is also a weak*-denting point.
Corollary 6.3. Let M be a compact metric space. Assume that lip0(M) 1-
S.P.U. Then every extreme point of BF (M ) is also a denting point.
At this point one could be inclined to believe that the denting points and the
strongly exposed points of BF (M ) coincide, at least when M is compact. We
are going to give an example of a compact metric space for which the inclusion
strexp(BF (M )) ⊂ ext(BF (M )∗∗) ∩ F (M) is strict.
Example 6.4. Let (T, d) be the following set with its real-tree distance
We will consider (Ω, d) as the set
[0, 1] × {0} ∪
∞[n=2(cid:26)1 −
{(0, 0), (1, 0)} ∪(cid:26)(cid:18)1 −
1
1
n(cid:27) ×(cid:20)0,
n2(cid:21) .
n2(cid:19) : n ≥ 2(cid:27)
1
1
n
,
together with the distance inherited from (T, d). Let us call for simplicity
0 := x1 := (0, 0), x∞ := (1, 0) and xn := (1 − 1
n2 ) if n ≥ 2.
n , 1
Since the couple (x∞, 0) is not a peak couple, the characterisation of the
points in strexp(BF (M )) given in [11] yields that δ(x∞) is not a strongly exposed
point of BF (Ω). Aliaga and Guirao [2] have proved that for a compact M, the
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
25
condition [x, y] = {x, y} implies that δ(x)−δ(y)
d(x,y)
BF (M ). In particular δ(x∞) is a preserved extreme point of BF (Ω).
is a preserved extreme point of
7. Application to norm attainment
Given a metric space M and a Banach space X, we have the following isomet-
ric identification Lip0(M, X) = L(F (M), X). Considering f ∈ Lip0(M, X), we
say that f strongly attains its norm if there are two different points x, y ∈ M
such that kf (x) − f (y)k = kf kd(x, y). In view of the results of [11, 14, 18], we
wonder when the classical notion of norm attainment coincides with the one
defined just above. In light of Bishop-Phelps theorem, we are also interested
in the denseness of the class of Lipschitz functions which strongly attain their
norm in Lip0(M, X).
We will mean by LipSN A(M, X) (respectively NA(F (M), X)) to the class of
all functions in Lip0(M, X) which strongly attain its norm (respectively which
attain its norm as a linear and continuous operator from F (M) to X). Let us
recall that a Banach space is said to have the Krein-Milman property (KMP)
if every non-empty closed convex bounded subset has an extreme point. It is
well known that (RNP) implies (KMP), although the converse is still an open
question (there are important classes of spaces for which the answer is yes).
We shall begin by stating the scalar case of previous result.
Proposition 7.1. Let (M, d) be a metric space such that F (M) has (KMP)
and such that ext(BF (M )) ⊆ V . Then every f ∈ Lip0(M) which attains its
norm on F (M) also strongly attains it. In other words, the following equality
holds:
NA(F (M), R) = LipSN A(M, R).
Therefore, LipSN A(M, R)
k · k
= Lip0(M).
Proof. Notice that the inclusion LipSN A(M, R) ⊆ NA(F (M), R) always holds.
Thus we just have to prove the reverse one. Let f be a function in Lip0(M)
which attains its norm on BF (M ). Since F (M) has (KMP), f also attains its
norm on an extreme point. Indeed, the set
C = {µ ∈ BF (M ) : hf, µi = 1}
is a non-empty closed convex bounded subset of F (M), so there is µ ∈ ext(C).
Then it is easy to check that µ is also an extreme point of BF (M ). Since
ext(BF (M )) ⊆ V , f attains its norm on a molecule δ(x)−δ(y)
d(x,y) with x 6= y.
The last part follows from Bishop-Phelps theorem.
(cid:3)
As a consequence of Proposition 4.1, we get the following.
26
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
Corollary 7.2. Let M be a separable bounded metric space such that F (M)
admits a natural predual X ⊂ lip0(M). Then
NA(F (M), R) = LipSN A(M, R) and LipSN A(M, R)
k · k
= Lip0(M).
We give some examples where the previous corollary applies.
Example 7.3.
(1) M compact metric space such that lip0(M) separates points uniformly
(note that this result was first proved by Godefroy using M-ideal theory,
see [14]). For instance M being compact and countable (see [5]), being
the middle third Cantor set (see [24]), or being any compact metric
space where the distance is composed with a nontrivial gauge (see [19]).
(2) M uniformly discrete metric space satisfying the assumptions of Propo-
sition 5.5.
(3) (BX ∗, k · kp) unit ball of a separable dual Banach space where the dis-
tance is the norm to the power p ∈ (0, 1) (see Proposition 6.3 in [19]).
Following the ideas of the last section in [10] we can extend this kind of result
in the vector valued case.
Proposition 7.4. Let X be a Banach space and M be a separable bounded
metric space such that F (M) admits a natural predual S ⊂ lip0(M). Assume
that X ∗ and F (M) has (RNP), and that X ∗ or F (M) has (AP ). Then
NA(F (M), X ∗∗)
k · k
= L(F (M), X ∗∗).
k · k
= Lip0(M, X ∗∗).
Equivalently, LipSN A(M, X ∗∗)
Proof. First of all, note that f ∈ Lip0(M, X ∗∗) strongly attains its norm if and
only if f : F (M) → X ∗∗ attains its operator norm. Indeed, assume that there
is γ ∈ F (M), kγk ≤ 1 and kf (γ)kX ∗∗ = kf kL. By the Hahn-Banach theorem
there is z ∈ X ∗∗∗ such that hz, f (γ)i = kf (γ)k. Since z ◦ f : M → R is a real
valued Lipschitz function which attains its operator norm on γ, Proposition 7.1
provides a molecule mxy ∈ F (M) such that f attains its operator norm on it.
This exactly means that f strongly attains its norm. The other implication is
trivial.
Next, the following isometric identification are well known (see [10] for exam-
ple): Lip0(M, X ∗∗) = L(F (M), X ∗∗) = (F (M)b⊗πX ∗)∗. Moreover, according
to [10, Lemma 4.3], if f ∈ L(F (M), X ∗∗) = (F (M)b⊗πX ∗)∗ attains its norm as
a linear form on F (M)b⊗πX ∗, then it also attains its norm as an operator in
To finish, apply Bishop-Phelps theorem to get the desired norm-denseness.
(cid:3)
L(F (M), X ∗∗).
EXTREMAL STRUCTURE AND DUALITY OF LIPSCHITZ FREE SPACES
27
Acknowledgment. The authors are grateful to Ram´on Aliaga and An-
tonio Guirao for sending them their preprint. The first and the last authors
are grateful to the Laboratoire de Math´ematiques de Besan¸con for the excel-
lent working conditions during their visit in June 2017. The third author is
grateful to Departamento de An´alisis Matem´atico of Universidad de Granada
for hospitality and excelent working conditions during his visit in June 2017.
The authors would like to thank Mat´ıas Raja for useful conversations.
References
[1] F. Albiac and N. J. Kalton, Lipschitz structure of quasi-Banach spaces, Israel J.
Math., 170 (2009), pp. 317–335.
[2] R. Aliaga and A. Guirao, On the preserved extremal structure of Lipschitz-free
spaces. arXiv:1705.09579, 2017.
[3] L. Borel-Mathurin, Isomorphismes non lin´eaires entre espaces de Banach, PhD the-
sis, Universit´e Paris 6, 2010.
[4] R. D. Bourgin, Geometric aspects of convex sets with the Radon-Nikod´ym property,
vol. 993 of Lecture Notes in Mathematics, Springer-Verlag, Berlin, 1983.
[5] A. Dalet, ´Etude des espaces Lipschitz-libres, PhD thesis, Universit´e de Franche-Comt´e,
2015.
[6] A. Dalet, Free spaces over countable compact metric spaces, Proc. Amer. Math. Soc.,
143 (2015), pp. 3537–3546.
[7]
, Free spaces over some proper metric spaces, Mediterr. J. Math., 12 (2015),
pp. 973–986.
[8] M. Fabian, P. Habala, P. H´ajek, V. Montesinos, and V. Zizler, Banach space
theory, CMS Books in Mathematics/Ouvrages de Math´ematiques de la SMC, Springer,
New York, 2011. The basis for linear and nonlinear analysis.
[9] V. P. Fonf, Massiveness of the set of extremal points of the unit sphere of some con-
jugate Banach spaces, Ukrain. Mat. Zh., 30 (1978), pp. 846–849, 863.
[10] L. Garc´ıa-Lirola, C. Petitjean, and A. R. Zoca, On the structure of spaces of
vector-valued Lipschitz functions, Studia Math., (2017). To appear.
[11] L. Garc´ıa-Lirola, A. Proch´azca, and A. R. Zoca, A characterisation of the
Daugavet property in spaces of Lipschitz functions. arXiv:1705.05145, 2017.
[12] L. Garc´ıa-Lirola and A. Rueda Zoca, Unconditional almost squareness and appli-
cations to spaces of Lipschitz functions, J. Math. Anal. Appl., 451 (2017), pp. 117–131.
[13] G. Godefroy, Boundaries of a convex set and interpolation sets, Math. Ann., 277
(1987), pp. 173–184.
[14]
, A survey on Lipschitz-free Banach spaces, Comment. Math., 55 (2015), pp. 89–
118.
[15] A. J. Guirao, V. Montesinos, and V. Zizler, On preserved and unpreserved ex-
treme points, in Descriptive topology and functional analysis, vol. 80 of Springer Proc.
Math. Stat., Springer, Cham, 2014, pp. 163–193.
[16] Y. Ivakhno, V. Kadets, and D. Werner, The Daugavet property for spaces of
Lipschitz functions, Math. Scand., 101 (2007), pp. 261–279.
[17]
, Corrigendum to: The Daugavet property for spaces of Lipschitz functions, Math.
Scand., 104 (2009), p. 319.
28
L. GARC´IA-LIROLA, C. PETITJEAN, A. PROCH ´AZKA, AND A. RUEDA ZOCA
[18] V. Kadets, M. Mart´ın, and M. Soloviova, Norm-attaining Lipschitz functionals,
Banach J. Math. Anal., 10 (2016), pp. 621–637.
[19] N. J. Kalton, Spaces of Lipschitz and Holder functions and their applications, Collect.
Math., 55 (2004), pp. 171–217.
[20] E. Matouskov´a, Extensions of continuous and Lipschitz functions, Canad. Math. Bull.,
43 (2000), pp. 208–217.
[21] C. Petitjean, Lipschitz-free spaces and Schur properties, J. Math. Anal. Appl., 453
(2017), pp. 894–907.
[22] J. I. Petun¯ın and A. N. Pl¯ıcko, Some properties of the set of functionals that attain
a supremum on the unit sphere, Ukrain. Mat. Z., 26 (1974), pp. 102–106, 143.
[23] N. Weaver, Duality for locally compact Lipschitz spaces, Rocky Mountain J. Math., 26
(1996), pp. 337–353.
[24]
, Lipschitz algebras, World Scientific Publishing Co., Inc., River Edge, NJ, 1999.
(L. Garc´ıa-Lirola) Universidad de Murcia, Facultad de Matem´aticas, Departa-
mento de Matem´aticas, 30100 Espinardo (Murcia), Spain
E-mail address: [email protected]
(C. Petitjean) Universit´e Bourgogne Franche-Comt´e, Laboratoire de Math´ematiques
UMR 6623, 16 route de Gray, 25030 Besanc¸on Cedex, France
E-mail address: [email protected]
(A. Proch´azka) Universit´e Bourgogne Franche-Comt´e, Laboratoire de Math´ematiques
UMR 6623, 16 route de Gray, 25030 Besanc¸on Cedex, France
E-mail address: [email protected]
(A. Rueda Zoca) Universidad de Granada, Facultad de Ciencias. Departamento
de An´alisis Matem´atico, 18071-Granada (Spain)
E-mail address:
URL: https://arzenglish.wordpress.com
[email protected]
|
1901.03196 | 1 | 1901 | 2019-01-09T13:46:25 | Analogs of certain quasi-analiticity results on Riemannian symmetric spaces of noncompact type | [
"math.FA"
] | An $L^2$ version of the celebrated Denjoy-Carleman theorem regarding quasi-analytic functions was proved by Chernoff \cite{CR} on $\mathbb R^d$ using iterates of the Laplacian. In $1934$ Ingham \cite{I} used the classical Denjoy-Carleman theorem to relate the decay of Fourier transform and quasi-analyticity of integrable functions on $\mathbb R$. In this paper we extend both these theorems to Riemannian symmetric spaces of noncompact type and show that the theorem of Ingham follows from that of Chernoff. | math.FA | math |
ANALOGS OF CERTAIN QUASI-ANALITICITY RESULTS ON
RIEMANNIAN SYMMETRIC SPACES OF NONCOMPACT TYPE
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
Abstract. An L2 version of the celebrated Denjoy-Carleman theorem regarding quasi-
analytic functions was proved by Chernoff [10] on Rd using iterates of the Laplacian.
In 1934 Ingham [23] used the classical Denjoy-Carleman theorem to relate the decay
of Fourier transform and quasi-analyticity of integrable functions on R. In this paper
we extend both these theorems to Riemannian symmetric spaces of noncompact type
and show that the theorem of Ingham follows from that of Chernoff.
1. Introduction
In this paper we will concern ourselves with the classical problem of determining
the relationship between the rate of decay of the Fourier transform of an integrable
function at infinity and the size of support of the function in the context of a Riemannian
symmetric space X = G/K of noncompact type, where G is a connected noncompact
semisimple Lie group with finite center and K is a maximal compact subgroup of G.
To understand the context better let us consider the following well known statement:
Suppose f ∈ L1(X) and its Fourier transform satisfies the estimate
f (λ, k) ≤ Cbht(λ), λ ∈ a∗,
for some C positive. If f vanishes on a nonempty open subset of X then f vanishes
identically (we refer the reader to Section 2 for meaning of relevant symbols). Unlike its
obvious Euclidean counterpart, the proof of this statement is not quite elementary as it
uses a nontrivial result such as the Kotake-Narasimhan theorem ([26]). However, sharp
conditions on the decay of the Fourier transform which prohibits a nonzero integrable
function to vanish on an open set is known from some classical results proved in [12, 22,
23, 25, 28, 31]. Recently, some of these results have appeared again in the context of
uniqueness of solution of Schrodinger equation [15]. This motivated us to have a fresh
look at these results and explore the possibility of extending them beyond Euclidean
spaces. Very recently we could extend the main result of [28] for Riemannian symmetric
spaces X (see [5]) and in this paper our objective is to do the same with the result
2010 Mathematics Subject Classification. Primary 43A85; Secondary 22E30, 33C67.
Key words and phrases. Riemannian symmetric space, Quasi-analyticity, Ingham's theorem.
The first author is supported by the Post Doctoral fellowship from IIT Bombay. Second author is
supported partially by SERB, MATRICS, MTR/2017/000235.
We are grateful to Professors Sayan Bagchi and Angela Pasquale for their helpful comments and
suggestions.
1
2
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
of Ingham [23]. A recently proved analogue of the result of Ingham on Rd states the
following.
Theorem 1.1 ([4], Theorem 2.2). Let θ : Rd → [0,∞) be a radially decreasing function
with limkξk→∞ θ(ξ) = 0 and we set
(1.1)
θ(ξ)
I =Zkξk≥1
kξkd dξ.
(a) Let f ∈ Lp(Rd), p ∈ [1, 2], be such that its Fourier transform F f satisfies the
estimate
(1.2)
F f (ξ) ≤ Ce−θ(ξ)kξk,
for almost every ξ ∈ Rd.
If f vanishes on a nonempty open set in Rd and I is infinite, then f is the zero
function on Rd.
(b) If I is finite then given any positive number L, there exists a nontrivial radial
function f ∈ C∞c (Rd) supported in B(0, L), satisfying the estimate (1.2).
Here,
F f (ξ) =ZRd
f (x)e−2πix·ξdx,
with x · ξ being the Euclidean inner product of the vectors x and ξ. It might be worth
noting that the function ξ → θ(ξ)kξk may not be radially increasing and hence the
condition (1.2) does not immediately imply a pointwise decay of the Fourier transform.
However, divergence of the integral I does imply that
θ(ξ)kξk = ∞.
lim sup
kξk→∞
For d = 1, using certain extra assumption on the function θ, Ingham showed that the
condition (1.2) together with the divergence of the integral I implies that
∞Xm=1(cid:13)(cid:13)(cid:13)(cid:13)
dmf
dxm(cid:13)(cid:13)(cid:13)(cid:13)
− 1
m
∞
= ∞,
(see [23, p. 30] for details). It then follows from the Denjoy-Carleman theorem ([34,
Theorem 19.11]) that f is quasi-analytic and hence vanishes identically under the as-
sumption that it vanishes on a nonempty open set.
It is fairly natural to anticipate
that an extension of Ingham's result on Rd or on a Riemannian symmetric space X
will involve a suitable extension of the Denjoy-Carleman theorem for these spaces. One
such result was obtained by Bochner in [7, Theorem 3] which provides an analogue of
the Denjoy-Carleman theorem for Rd involving iterates of the Laplace-Beltrami opera-
tor. An important variant of the result of Bochner was later proved by Chernoff in [10,
Theorem 6.1].
Theorem 1.2. Let f : Rd → C be a smooth function.
(a) (Bochner) If
Xm∈N k∆m
Rdfk− 1
∞ = ∞,
m
THEOREMS OF INGHAM AND CHERNOFF
3
and for all m ∈ N ∪ {0}, ∆m
nation then f vanishes identically.
Rdf (x) = 0 for all x in a set U of analytic determi-
(b) (Chernoff ) Suppose that for all m ∈ N ∪ {0}, ∆m
Rdf ∈ L2(Rd), and
Xm∈N k∆m
Rdfk− 1
2
2m
= ∞.
If f and all its partial derivatives vanish at the origin then f is identically zero.
In the above ∆Rd denotes the usual Laplacian on Rd. It was noted in [7] itself that part
(a) of the theorem above generalizes to Riemannian manifolds and hence is applicable on
Riemannian symmetric spaces. However, we could not use this result to prove Theorem
1.1 even for d = 1 due to the difficulty in estimating the relevant L∞ norms. On the
other hand, part (b) of the theorem above can be suitably used to prove Theorem 1.1.
It is thus natural to look for an analogue of Chernoff's result on Riemannian symmetric
spaces of noncompact type and then try to use it to prove an analogue of Theorem 1.1.
Our first result in this paper is the following weaker version of Theorem 1.2, (b) for a
Riemannian symmetric space X of noncompact type with rank d ≥ 1.
Theorem 1.3. Let f ∈ C∞(G/K) be such that ∆mf ∈ L2(G/K), for all m ∈ N ∪ {0}
and
(1.3)
∞Xm=1
k∆mfk− 1
2
2m
= ∞.
If f vanishes on any nonempty open set in G/K then f is identically zero.
Though the assumption that f vanishes on a nonempty open set drastically changes
the nature of the theorem but it is still sufficient for the application we have in mind. The
main idea here is to suitably use the connection between the Carleman type condition
(1.3) and approximation by polynomials proved in [13].
In fact, we will use (1.3) to
obtain completeness of Opdam hypergeometric functions in a suitable L2-space (see
Lemma 3.2). For a discussion on quasianaliticity and approximation by polynomials
on Lie groups, we refer the reader to [14] and references therein. We also construct an
example (see Example 3.8) to show the impossibility of proving an exact analogue of
Chernoff's result Theorem 1.2, (b) on symmetric spaces if we restrict ourselves only to
the class of G-invariant differential operators on X. It would be interesting to explore
the possibility of extending Theorem 1.3 (and more generally, Theorem 1.2, (b)) to more
general Riemannian manifolds. Our final result in this paper is the following analogue
of the theorem of Ingham (Theorem 1.1) on X.
Theorem 1.4. Let θ : [0,∞) → [0,∞) be a decreasing function with limr→∞ θ(r) = 0
and
(1.4)
where d = rank(X).
I =Z{λ∈a∗
+ kλkB≥1}
θ(kλkB)
kλkd
B
dλ,
4
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
(1.5)
(a) Suppose f ∈ L1(X) and its Fourier transform ef satisfies the estimate
Za∗×K ef (λ, k) eθ(kλkB )kλkB c(λ)−2dλ dk < ∞.
If f vanishes on a nonempty open set in X and I is infinite then f is the zero
function.
(b) If I is finite then given any positive number L, there exists a nontrivial f ∈
C∞c (G//K) supported in B(o, L) satisfying the estimate (1.5).
We refer the reader to Theorem 4.2 for an exact analogue of Theorem 1.1. Recently
we could prove a result analogous to Theorem 1.4 for Riemannian symmetric spaces X
(see [5, Theorem 1.2]) where the function kλkBθ(λ) was assumed to be increasing (as
in [28]). It turns out that with some effort, one can construct a function θ satisfying
the conditions of Theorem 1.4 such that kλkθ(λ) is decreasing on an unbounded set
of positive Lebesgue measure. This essentially proves that the results of Ingham and
Levinson are independent of each other.
This paper is organised as follows: In the next section, we will recall the required pre-
liminaries regarding harmonic analysis on Riemannian symmetric spaces of noncompact
type. We will prove Theorem 1.3 and Theorem 1.4 in section 3 and section 4 respectively.
2. Riemannian symmetric spaces of noncompact type
In this section, we describe the necessary preliminaries regarding semisimple Lie
groups and harmonic analysis on associated Riemannian symmetric spaces. These are
standard and can be found, for example, in [16, 18, 19, 20]. To make the article self-
contained, we shall gather only those results which will be used throughout this paper.
Let G be a connected, noncompact, real semisimple Lie group with finite centre and
g its Lie algebra. We fix a Cartan involution θ of g and write g = k ⊕ p where k and p
are +1 and −1 eigenspaces of θ respectively. Then k is a maximal compact subalgebra
of g and p is a linear subspace of g. The Cartan involution θ induces an automorphism
Θ of the group G and K = {g ∈ G Θ(g) = g} is a maximal compact subgroup of G.
Let a be a maximal subalgebra in p; then a is abelian. We assume that dim a = d, called
the real rank of G. Let B denote the Cartan Killing form of g. It is known that B p×p
is positive definite and hence induces an inner product and a norm k · kB on p. The
homogeneous space X = G/K is a smooth manifold with rank(X) = d. The tangent
space of X at the point o = eK can be naturally identified to p and the restriction of
B on p then induces a G-invariant Riemannian metric d on X. For a given g ∈ G and
a positive number L we define
B(gK, L) = {xK x ∈ G, d(gK, xK) < L},
to be the open ball with center gK and radius L. We can identify a with Rd endowed
with the inner product induced from p and let a∗ be the real dual of a. The set of
restricted roots of the pair (g, a) is denoted by Σ. It consists of all α ∈ a∗ such that
gα = {X ∈ g [Y, X] = α(Y )X,
for all Y ∈ a} ,
THEOREMS OF INGHAM AND CHERNOFF
5
is nonzero with mα = dim(gα). We choose a system of positive roots Σ+ and with
for all α ∈ Σ+}.
respect to Σ+, the positive Weyl chamber a+ = {X ∈ a α(X) > 0,
We denote by
n = ⊕α∈Σ+ gα.
Then n is a nilpotent subalgebra of g and we obtain the Iwasawa decomposition g =
k⊕ a⊕ n. If N = exp n and A = exp a then N is a Nilpotent Lie group and A normalizes
N. For the group G, we now have the Iwasawa decomposition G = KAN, that is, every
g ∈ G can be uniquely written as
g = κ(g) exp H(g)η(g), κ(g) ∈ K, H(g) ∈ a, η(g) ∈ N,
and the map
(k, a, n) 7→ kan
2Pα∈Σ+
is a global diffeomorphism of K × A × N onto G. Let ρ = 1
mαα be the half
sum of positive roots counted with multiplicity. Let M′ and M be the normalizer and
centralizer of a in K respectively. Then M is a normal subgroup of M′ and normalizes N.
The quotient group W = M′/M is a finite group, called the Weyl group of the pair (g, k).
W acts on a by the adjoint action. It is known that W acts as a group of orthogonal
transformation (preserving the Cartan-Killing form) on a. Each w ∈ W permutes the
Weyl chambers and the action of W on the Weyl chambers is simply transitive. Let
A+ = exp a+. Since exp : a → A is an isomorphism we can identify A with Rd. If A+
denotes the closure of A+ in G, then one has the polar decomposition G = KAK, that
is, each g ∈ G can be written as
g = k1(exp Y )k2, k1, k2 ∈ K, Y ∈ a.
In the above decomposition, the A component of g is uniquely determined modulo W .
In particular, it is well defined in A+. The map (k1, a, k2) 7→ k1ak2 of K × A× K into G
induces a diffeomorphism of K/M × A+ × K onto an open dense subset of G. It follows
that if gK = k1(exp Y )K ∈ X then
(2.1)
d(o, gK) = kY kB.
We extend the inner product on a induced by B to a∗ by duality, that is, we set
where Yλ is the unique element in a such that
hλ, µi = B(Yλ, Yµ), λ, µ ∈ a∗, Yλ, Yµ ∈ a,
This inner product induces a norm, again denoted by k · kB, on a∗,
λ(Y ) = B(Yλ, Y ),
for all Y ∈ a.
The elements of the Weyl group W acts on a∗ by the formula
kλkB = hλ, λi
1
2 , λ ∈ a∗.
sYλ = Ysλ,
s ∈ W, λ ∈ a∗.
6
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
Let a∗C denote the complexification of a∗, that is, the set of all complex-valued real
linear functionals on a. If λ : a → C is a real linear functional then ℜλ : a → R and
ℑλ : a → R, given by
ℜλ(Y ) = Real part of λ(Y ),
ℑλ(Y ) = Imaginary part of λ(Y ),
for all Y ∈ a,
for all Y ∈ a,
are real-valued linear functionals on a and λ = ℜλ + iℑλ. The usual extension of B to
a∗C, using conjugate linearity is also denoted by B. Hence a∗C can be naturally identified
with Cd such that
kλkB =(cid:0)kℜλk2
B(cid:1) 1
B + kℑλk2
2 ,
λ ∈ a∗C.
Through the identification of A with Rd, we use the Lebesgue measure on Rd as the
Haar measure da on A. As usual on the compact group K, we fix the normalized Haar
measure dk and dn denotes a Haar measure on N. The following integral formulae
describe the Haar measure of G corresponding to the Iwasawa and Polar decomposition
respectively.
f (k exp Y n) e2ρ(Y ) dn dY dk,
f ∈ Cc(G)
f (k1ak2) J(a) dk1 da dk2,
ZG
f (g)dg = ZKZaZN
= ZKZA+ZK
J(exp Y ) = c Yα∈Σ+
where dY is the Lebesgue measure on Rd and
(sinh α(Y ))mα ,
for Y ∈ a+,
c being a normalizing constant. If f is a function on X = G/K then f can be thought
of as a function on G which is right invariant under the action of K. It follows that on
X we have a G invariant measure dx such that
ZX
f (x) dx =ZK/MZa+
f (k exp Y ) J(exp Y ) dY dkM ,
where dkM is the K-invariant measure on K/M. For a sufficiently nice function f on
(2.2)
f (g)e(iλ−ρ)H(g−1k)dg,
λ ∈ a∗C, k ∈ K,
ef (λ, k) =ZG
X, its Fourier transform ef is defined on a∗C × K by the formula
whenever the integral exists ([19, P. 199]). As M normalizes N the function k 7→ ef (λ, k)
is right M-invariant. It is known that if f ∈ L1(X) then ef (λ, k) is a continuous function
If in addition ef ∈ L1(a∗ × K,c(λ)−2 dλ dk) then the following Fourier inversion holds,
of λ ∈ a∗, for almost every k ∈ K (in fact, holomorphic in λ on a domain containing a∗).
f (gK) = W−1Za∗×K ef (λ, k) e−(iλ+ρ)H(g−1k) c(λ)−2dλ dk,
(2.3)
THEOREMS OF INGHAM AND CHERNOFF
7
denotes Harish Chandra's c-function. Moreover, f 7→ ef extends to an isometry of
for almost every gK ∈ X ([19, Chapter III, Theorem 1.8, Theorem 1.9]). Here c(λ)
L2(X) onto L2(a∗+ × K,c(λ)−2 dλ dk) ([19, Chapter III, Theorem 1.5]).
Remark 2.1. It is known that ([1, P. 394], [11, P. 117]) for λ ∈ a∗+ there exists a positive
number C such that
(2.4)
c(λ)−2 ≤ C(1 + kλkB) dim n.
If rank(X) = 1, then a similar lower estimate holds ([2], P. 653); there exist two positive
numbers C1 and C2 such that for all λ ≥ 1
(2.5)
C1λdim n ≤ c(λ)−2 ≤ C2λdim n.
We now specialize to the case of K-biinvariant function f on G. Using the polar
decomposition of G we may view a K-biinvariant function f on G as a function on A+,
or by using the inverse exponential map we may also view f as a function on a solely
determined by its values on a+. Henceforth, we shall denote the set of K-biinvariant
functions in L1(G) by L1(G//K). If f ∈ L1(G//K) then the Fourier transform ef can
also be written as
f (g)φ−λ(g) dg,
ef (λ, k) = bf (λ) =ZG
φλ(g) =ZK
e−(iλ+ρ)(cid:0)H(g−1k)(cid:1) dk,
λ ∈ a∗C,
is Harish Chandra's elementary spherical function. We now list down some well known
properties of the elementary spherical functions which are important for us ([16], Prop
3.1.4 and Chapter 4, §4.6; [19], Lemma 1.18, P. 221).
Theorem 2.2.
(1) φλ(g) is K-biinvariant in g ∈ G and W -invariant in λ ∈ a∗C.
(2) φλ(g) is C∞ in g ∈ G and holomorphic in λ ∈ a∗C.
(3) For all λ ∈ a∗+ and g ∈ G we have
(2.6)
where
(2.7)
(2.8)
(2.9)
φλ(g) ≤ φ0(g) ≤ 1.
(4) For all Y ∈ a+ and λ ∈ a∗+
0 < φiλ(exp Y ) ≤ eλ(Y )φ0(exp Y ).
(5) If ∆ denotes the Laplace-Beltrami operator on X then
λ ∈ a∗C.
∆(φλ) = −(kλk2
B + kρk2
B)φλ,
For K-biinvariant Lp functions on G the following Fourier inversion formula is well
if f ∈ Lp(G//K), 1 ≤ p ≤ 2 with
known ([36], Therem 3.3 and [29], Theorem 5.4):
f ∈ L1(a∗+,c(λ)−2 dλ) then for almost every g ∈ G,
(2.10)
f (g) = W−1Za∗
f (λ)φ−λ(g)c(λ)−2 dλ dk.
8
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
The spherical Fourier transform and the Euclidean Fourier transform on a are related
by the so-called Abel transform. For f ∈ L1(G//K) its Abel transform Af is defined
by the integral
Af (exp Y ) = eρ(Y )ZN
f ((exp Y )n) dn, Y ∈ a,
([16, P. 107], [21, p.27]). We will need the following theorem regarding Abel transform
([16, Prop 3.3.1, Prop 3.3.2]), which we will refer as the slice projection theorem.
Theorem 2.3. The map A : C∞c (G//K) → C∞c (a)W is a bijection. If f ∈ C∞c (G//K)
then
F(cid:0)Af(cid:1)(λ) = bf (λ), λ ∈ a∗,
(2.11)
where F (Af ) denotes the Euclidean Fourier transform on a ∼= Rd.
Remark 2.4. It is easy to see that a special case of Theorem 1.4, namely when f ∈
C∞c (G//K), can be proved simply by using the slice projection theorem (see [6] for a
more general result in this direction). However, this approach cannot be used to prove
Theorem 1.4, because if an integrable K-biinvariant function f vanishes on an open set
then it is not necessarily true that Af also vanishes on an open subset of a.
We end this section with a short discussion on Opdam hypergeometric functions which
will play a crucial role in the next section. Let areg be the subset of regular elements in
a, that is,
areg = ∪α∈Σ(ker α)∁.
α(ξ)
For ξ ∈ a, let Tξ be the Dunkl-Cherednik operator [30, Definition 2.2], defined for
f ∈ C 1(a) and for x ∈ areg, by
∂
∂ξ
f (x) + Xα∈Σ+
mα
Tξf (x) =
1 − e−α(x) (f (x) − f (rαx)) − ρ(ξ)f (x),
where rα is the orthogonal reflection with respect to the hyperplane ker α and ∂
∂ξ f is the
directional derivative f in the direction of ξ. In particular rα is an isometry of (a,k·kB).
Let {ξ1, ξ2,· · · , ξd} be an orthonormal basis of a with respect to the inner product given
by B. Then the Heckman-Opdam Laplacian L defined by
L =
T 2
ξi,
dXi=1
is the K-invariant part of the Laplace-Beltrami operator ∆ on X ([17, Theorem 2.2,
Remark 2.3]).
Now, suppose that f is a K-biinvariant smooth, function on G which vanishes on the
ball B(o, L) for some positive number L. Using the polar decomposition of G we can
view f as a function on A and hence on the Lie algebra a. Using (2.1) it follows that this
latter function (again denoted by f ) vanishes on the set {X ∈ a kXkB < L} which we
will continue to denote by the same symbol B(o, L). Since rα is an isometry it follows
from the expression of the Dunkl-Cherednik operator that in this case Tξf also vanishes
on B(o, L) for all ξ ∈ a.
THEOREMS OF INGHAM AND CHERNOFF
9
The Opdam hypergeometric function Gλ, λ ∈ a∗C is defined to be the unique analytic
function on a such that
(2.12)
TξGλ = iλ(ξ) Gλ,
ξ ∈ a,
Gλ(0) = 1,
((see [30, p. 89]). The elementary spherical function φλ and the Opdam hypergeometric
function Gλ are related by [30, p. 89]
(2.13)
φλ(x) =
Gλ(wx), λ ∈ a∗C, x ∈ a.
1
W Xw∈W
However, there exists an alternative relation between φλ and Gλ which is important for
us [30, (4), p. 119]: For all x ∈ a,
(2.14)
g(wλ)Gwλ(x),
φλ(x) =
1
W Xw∈W
g(λ) = Yα∈Σ0
+
for almost every λ ∈ a∗,
! ,
λα = hλ, αi
hα, αi
,
1 −
mα
2 + mα/2
4
λα
where
(2.15)
and Σ0
Gλ which will be needed:
+ = {α ∈ Σ+ 2α 6∈ Σ}. We now list down a few results regarding the function
(1) For λ ∈ a∗
(2.16)
g(λ)c(λ)−1 ≤ C1 + C2kλkp,
for some C1, C2, p ≥ 0 (follows from [30, equation (8.1)]).
(2) For λ ∈ a∗, the function Gλ is known to be bounded on a ([30, Proposition 6.1]).
(3) More generally, for any polynomial p of degree N there exists a constant Cp such
that for all λ ∈ a∗, x ∈ a
(2.17)
where ∂
sition 3.2]).
(cid:12)(cid:12)(cid:12)(cid:12)p(cid:18) ∂
∂x(cid:19) Gλ(x)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Cp(1 + λ)N φ0(x),
∂x Gλ is the directional derivative of Gλ in the direction x (see [35, Propo-
3. Chernoff's theorem for symmetric spaces
In this section our aim is to prove the theorem of Chernoff (Theorem 1.3). We start
with a few results which will be needed to prove Theorem 1.3. The following lemma is
just a restatement of [13, Theorem 2.3] in view of the identification of a∗ with Rd.
Lemma 3.1. Let µ be a finite Borel measure on a∗ such that, for m ∈ N and 1 ≤ j ≤ d
the quantity Mj(m), defined by
Mj(m) =Za∗ λ(ξj)m dµ(λ),
10
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
is finite, where {ξ1, ξ2,· · · , ξd} is an orthonormal basis of a. If for each j ∈ {1,· · · , d},
the sequence {Mj(2m)}∞m=1 satisfies the Carleman's condition
(3.1)
Mj(2m)− 1
2m = ∞,
Xm∈N
then the polynomials in a∗ are dense in L2(a∗, dµ).
Given a positive number L we consider the following function space
GL(a∗) = span(cid:8)χx : a∗ → C x ∈ B(o, L), χx(λ) = Gλ(x), λ ∈ a∗(cid:9).
Lemma 3.2. Let µ be a finite Borel measure on a∗ such that for each j ∈ {1,· · · , d} and
each m ∈ N the quantity Mj(m) (as in Lemma 3.1) is finite. If for each j ∈ {1,· · · , d}
the sequence Mj(2m) satisfies the Carleman's condition (3.1) then for each L positive,
GL(a∗) is dense in L2(a∗, dµ).
Proof. We first note that because of finiteness of µ and boundedness of Gλ, for λ ∈ a∗,
the space GL(a∗) is a subset of L2(a∗, dµ). Let f ∈ L2(a∗, dµ) be such that
(3.2)
f (λ)Gλ(x) dµ(λ) = 0,
for all x ∈ B(o, L).
We define a function F on a by
Za∗
F (x) =Za∗
f (λ)Gλ(x) dµ(λ),
for x ∈ a.
It follows from (3.2) that F vanishes on the ball B(o, L). Estimate (2.17) together with
dominated convergence theorem implies that
P(cid:18) ∂
∂x(cid:19) F (x) =Za∗
f (λ) P(cid:18) ∂
∂x(cid:19) Gλ(x) dµ(λ),
for any polynomial P . Hence, for α = (α1,· · · , αd) ∈ (N ∪ {0})d we have from the
eigenvalue equation (2.12) that
ξ1 · · · T αd
T α1
ξd
(3.3)
F (x) = Za∗
= Za∗
f (λ) T α1
ξ1 · · · T αd
ξd
Gλ(x) dµ(λ)
f (λ) (iλ(ξ1))α1 · · · (iλ(ξd))αd Gλ(x) dµ(λ).
ξ1 · · · T αd
Since F vanishes on the ball B(o, L) so does T α1
taking x = 0, that for all α = (α1,· · · αd) ∈ (N ∪ {0})d
ξd
F . It now follows from (3.3) by
Za∗
(λ(ξ1))α1 · · · (λ(ξd))αd f (λ) dµ(λ) = 0.
This implies that f annihilates all polynomials and hence by Lemma 3.1, f is the zero
(cid:3)
function.
We will also need the following elementary lemma.
Proof. If lim sup an or lim inf an is nonzero then the result follows trivially. Hence, it
suffices to prove the result for the case limn→∞ an = 0. Without loss of generality we
can also assume that an ∈ (0, 1), for all n ∈ N. Let us define
n2(cid:9), B =(cid:8)n ∈ N : an >
As Pn∈N an diverges it follows that B is an infinite set. The result now follows by
A =(cid:8)n ∈ N : an ≤
observing that
1
n2(cid:9).
1
1+ m
a
n
n
an
lim
n→∞,n∈B
= lim
n→∞,n∈B
m
n
a
n = lim
n→∞,n∈B
e− m
n log 1
an = 1,
as for n ∈ B
1 ≤
1
an ≤ n2, and hence 0 ≤ log
1
an ≤ 2 log n.
(cid:3)
For f ∈ L1(X), we define the K-biinvariant component Sf of f by the integral
(3.4)
and for g ∈ G, we define the left translation operator lg on L1(X) by
Sf (x) =ZK
f (kx) dk, x ∈ X,
lgf (x) = f (gx), x ∈ X.
THEOREMS OF INGHAM AND CHERNOFF
11
Lemma 3.3. Let {an} be a sequence of positive numbers such that the series Pn∈N an
diverges. Then given any m ∈ N,
1+ m
a
n
n
= ∞.
Xn∈N
Remark 3.4. The operator lg is usually defined as left translation by g−1. The reason
we have defined lg differently because then it follows that S(lgf ) = S(lg1f ) if gK = g1K.
For a nonzero integrable function f , its K-biinvariant component S(f ) may not be
nonzero. However, the following lemma shows that there always exists g ∈ G such that
S(lgf ) is nonzero.
Lemma 3.5. ([5, Lemma 4.6]) If f ∈ L1(X) is nonzero then for every L positive, there
exists g ∈ G with gK ∈ B(o, L) such that S(lgf ) is nonzero.
We now present the proof of Theorem 1.3.
Proof of Theorem 1.3. The following steps will lead to the proof of the theorem.
Step 1: Using translation invariance of ∆, we can assume without loss of generality that
f ∈ C∞(G/K) and vanishes on the ball B(0, L) for some L positive. We first show that
it suffices to prove the result under the additional assumption that f is K-biinvariant.
To see this, suppose f ∈ C∞(G/K) vanishes on B(0, L) and satisfies the hypothesis
(1.3). Since ∆ is left-translation invariant operator and ∆mf ∈ L2(G/K), it is easy to
check that
∆m(Sf ) = S(∆mf ), for all m ∈ N.
12
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
Therefore
Hence
k∆m(Sf )k2
2 =ZG S(∆mf )(g)2 dg ≤ZGZK ∆mf (kg)2 dkdg = k∆mfk2
∞Xm=0
k∆m(Sf )k− 1
k∆mfk− 1
∞Xm=0
= ∞.
≥
2m
2m
2
2
2.
If f is not identically zero but vanishes on B(o, L), then by Lemma 3.5, there exists
g0K ∈ B(o, L/2) such that S(lg0f ) is non zero but vanishes on B(o, L/2). Hence, if the
theorem is true for K-biinvariant function, then S(lg0f ) must vanishes identically which
is a contradiction.
So, we assume that f ∈ C∞(G//K) is such that ∆mf ∈ L2(G//K) for all m ∈ N∪{0}
and satisfies (1.3). We will show that if f vanishes on B(o, L) then f is the zero function.
Step 2: Given such a function f we define a measure µ on a∗ by
µ(E) =ZE f (λ)c(λ)−2dλ,
for all Borel subsets E of a∗. Note that the measure µ is W -invariant. We claim that
that the space GL(a∗) is dense in L2(a∗, µ) for any given positive number L. Since
∆nf ∈ L2(a∗,c(λ)−2dλ) for all n ∈ N ∪ {0} and c(λ)−2 is of polynomial growth
(Remark 2.1) it follows by a simple application of Cauchy-Schwarz inequality that µ is
a finite measure and GL(a∗) is contained in L2(a∗, µ). The same argument also implies
that polynomials are contained in L2(a∗, µ). To prove the claim it suffices, in view of
Lemma 3.2, to show that
∞Xm=1
Mj(2m)− 1
2m = ∞,
where Mj(m) are as in Lemma 3.1. Now, for a large enough r ∈ N and for all m ∈ N
we have
Mj(2m) ≤ Za∗
≤ (cid:18)Za∗
(kλk2 + kρk2)2r dλ(cid:19) 1
2 (cid:18)Za∗
c(λ)−2
2
(kλk2 + kρk2)m bf (λ) c(λ)−2 dλ
(kλk2 + kρk2)(2m+2r) bf (λ)2 c(λ)−2 dλ(cid:19) 1
(kλk2 + kρk2)2r dλ(cid:19) 1
Ar =(cid:18)Za∗
c(λ)−2
.
2
= Ark∆m+rfk2,
where
Therefore,
Mj(2m)− 1
2m ≥ A− 1
r
2m
k∆(m+r)fk− 1
2
2m
= A− 1
r
2m
(cid:18)k∆(m+r)fk− 1
2
2(m+r)
m )
(cid:19)(1+ r
.
Since, limm→∞ A− 1
r
2m
THEOREMS OF INGHAM AND CHERNOFF
13
= 1, it follows from Lemma 3.3 and the hypothesis (1.3) that
It follows that for each positive number L the space GL(a∗) is dense in L2(a∗, µ).
∞Xm=1
Mj(2m)− 1
2m = ∞.
In view of the relation (2.14) between Gλ and φλ one would expect that
Step 3:
the previous step implies something regarding completeness of the elementary spherical
functions φλ. In this regard we consider the space
ΦL(a∗) = span(cid:8)λ 7→ φλ(x) x ∈ B(o, L), λ ∈ a∗(cid:9),
for any given positive number L and claim that ΦL(a∗) is dense in L1(a∗, µ)W . Here
L1(a∗, µ)W is the W -invariant functions in L1(a∗, µ). To prove this we consider a W -
invariant function h ∈ L∞(a∗, µ) such that
h(λ)φλ(x) dµ(λ) = 0,
for all x ∈ B(o, L). By (2.14) and the W -invariance of h it follows that
(3.5)
h(λ)g(λ)Gλ(x) dµ(λ) = 0,
Za∗
Za∗
for all x ∈ B(o, L). We will now repeatedly apply the operators Tξ on the integral in
(3.5) by viewing this as a function of the variable x ∈ a. To justify the differentiation
under the integral it is necessary to show that for each n ∈ N ∪ {0}
Za∗ h(λ) kλkn g(λ)Gλ(x) dµ(λ) < ∞.
By using the estimate (2.16) and the boundedness of Gλ, we get that
Za∗ h(λ) kλkn g(λ)Gλ(x) dµ(λ)
≤ Ckhk∞Za∗ kλkn g(λ) bf(λ)c(λ)−2 dλ
≤ Ckhk∞Za∗ kλkn (C1 + C2kλkp) bf (λ)c(λ)−1 dλ
≤ Ckhk∞Za∗
(kλk2 + kρk2)Mbf (λ)c(λ)−1 dλ,
(kλk2 + kρk2)2M +d+1bf (λ)2c(λ)−2 dλ(cid:19) 1
≤ Ckhk∞(cid:18)Za∗
2(cid:18)Za∗
M ∈ N, M ≥ n + p
1
(kλk2 + kρk2)d+1 dλ(cid:19) 1
2
< ∞,
where d = dim a. Note that in the last step we have used the assumption ∆mf ∈
L2(G/K) for all m ∈ N ∪ {0}. For each α ∈ Σ+
0 , we now choose ξα ∈ a in such a way
14
that
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
for all λ ∈ a∗.
Applying the composition of the operators Tξα, for all α ∈ Σ+
follows that
λ(ξα) = λα =
,
λ(α)
hα, αi
(3.6)
Za∗
h(λ)g(λ)Yα∈Σ+
0
iλα Gλ(x) dµ(λ) = 0,
0 on both sides of (3.5) it
for all x ∈ B(o, L). From the expression of the function g given in (2.15) it is easy to
see that the function g(λ)(cid:16)Qα∈Σ+
function it follows that the function h(λ)g(λ)(cid:16)Qα∈Σ+
iλα(cid:17) is of polynomial growth. Since h is a bounded
iλα(cid:17) is in L2(a∗, µ). As GL(a∗)
is dense in L2(a∗, dµ) it follows from (3.6) that h = 0 for almost every λ.
0
0
Step 4: By the Fourier inversion (2.10) we have that for all x ∈ B(o, L)
f (x) = W−1Za∗ bf (λ)φλ(x)c(λ)−2dλ = 0.
Therefore,
This implies that for all u ∈ ΦL(a∗)
(3.7)
Za∗ bf (λ)u(λ)c(λ)−2dλ = 0.
As bf ∈ L1(a∗, dµ)W by the completeness of ΦL(a∗) in L1(a∗, dµ)W we can approximate
bf by the elements of ΦL(a∗), that is, given ǫ > 0, there exists u0 ∈ ΦL(a∗) such that
kbf − u0kL1(a∗,dµ) < ǫ.
Za∗ bf (λ)2c(λ)−2dλ = Za∗ bf (λ)bf (λ)c(λ)−2dλ
= (cid:12)(cid:12)(cid:12)(cid:12)Za∗(cid:16)bf (λ) − u0(λ) + u0(λ)(cid:17) bf (λ)c(λ)−2dλ(cid:12)(cid:12)(cid:12)(cid:12)
≤ Za∗ bf (λ) − u0(λ) dµ(λ) +(cid:12)(cid:12)(cid:12)(cid:12)Za∗ bf (λ)u0(λ)c(λ)−2dλ(cid:12)(cid:12)(cid:12)(cid:12)
< ǫ.
(cid:3)
It follows that bf is zero and hence so is f .
Remark 3.6.
(1) We will like to point out that for rank one symmetric spaces it is
possible to prove Theorem 1.3 without appealing to Dunkl-Cherednik operator
Tξ and the Opdam hypergeometric functions Gλ. To see this, note that if µ is
an even, finite Borel measure on R and the sequence
m ∈ N,
λ2mdµ(λ),
M(2m) =ZR
THEOREMS OF INGHAM AND CHERNOFF
15
satisfies the Carleman condition (3.1) then by Lemma 3.1 the polynomials which
are even functions form a dense subspace of L2(R, µ)e where
L2(R, µ)e = {f ∈ L2(R, µ) f (λ) = f (−λ), for almost every λ ∈ R}.
We note that given any such polynomial P there exists a polynomial Q such that
P (λ) = Q(λ2 + ρ2),
for all λ ∈ R.
Obviously the same conclusion is not valid for W -invariant polynomials on Rd,
d > 1 and this is the main reason why we needed to use the Dunkl-Cherednik
operators and the Opdam hypergeometric functions in the proof of Theorem 1.3.
Now, if f ∈ L2(R, µ)e is such that for all n ∈ N ∪ {0}
ZR
f (λ)(λ2 + ρ2)ndµ(λ) = 0,
then it follows that f annihilates all polynomials which are even functions. Con-
sequently, f is the zero function. This can be used to prove that the space ΦL(R)
is dense in L2(R, µ)e (and hence in L1(R, µ)e). Precisely, if f ∈ L2(R, µ)e is such
that
f (λ)φλ(x)dµ(λ) = 0,
for all x ∈ B(o, L),
ZR
then by defining
h(x) =ZR
ZR
f (λ)φλ(x)dµ(λ),
x ∈ G
(as in Step 3 of the proof above) and applying the Laplacian ∆ repeatedly to h
and putting x = e we get that for all n ∈ N ∪ {0}
f (λ)(λ2 + ρ2)ndµ(λ) = 0.
which implies that f is the zero function. The rest of the proof then goes as it
is.
(2) It was noted in [10] that Theorem 1.2, (b) fails for d = 1 if dm
dxm f (0) vanishes only
for even natural numbers m. An analogous phenomena occurs for symmetric
spaces also and shows that an exact analogue of Theorem 1.2, (b) is not true for
X if we restrict ourselves only to the class of G-invariant differential operators on
X. In the following we will illustrate this for the n-dimensional real hyperbolic
space Hn by constructing a nonzero square integrable function f on Hn such that
∆mf (x0) = 0,
for all m ∈ N ∪ {0},
for some x0 ∈ Hn and satisfies (1.3).
We start with some preliminaries on Jacobi functions ([24]). A Jacobi function
φ(α,β)
λ
(3.8)
(α, β, λ ∈ C, α 6= −1,−2,· · · ) is the unique even C∞ function on R satisfying
(cid:18) d2
+ λ2 + (α + β + 1)2(cid:19) φ(α,β)
dt2 + ((2α + 1) coth t + (2β + 1) tanh t)
(0) = 1.
d
dt
φ(α,β)
λ
(t) = 0,
λ
16
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
In this paper we shall assume that α ≥ β ≥ − 1
2 . Let
∆(α,β) =
d2
dt2 + ((2α + 1) coth t + (2β + 1) tanh t)
d
dt
.
Then rewriting (3.8) we get that
(3.9)
(∆(α,β) + λ2 + (α + β + 1)2)φ(α,β)
λ
= 0,
φ(α,β)
λ
(0) = 1.
The Fourier-Jacobi transform of a suitable even function f on R is defined by
(3.10)
f (t)φ(α,β)
λ
(t)(2 sinh t)2α+1(2 cosh t)2β+1 dt,
F (α,β)f (λ) =Z ∞
0
for all complex numbers λ, for which the right hand side is well-defined. We point out
that this definition coincides with the group Fourier transform when (α, β) arises from
geometric cases. We also have the inversion and Plancherel formula for the Fourier-
Jacobi transformation (see [24, Theorem 2.2, Theorem 2.3] for the statement).
The Real hyperbolic space Hn is defined by
Hn = {x ∈ Rn+1 −x2
1 − x2
2 − · · · − x2
n + x2
n+1 = 1, xn+1 > 0}.
This is a rank one symmetric space of noncompact type and Hn = SO(n, 1)/SO(n). In
this particular case, we have (see [9, p.212]),
m1 = dim gα = n − 1, m2 = dim gα = 0.
It is well known ([24, (3.4)]) that the spherical function φλ on Hn is same as the Jacobi
function φ(α,β)
where
λ
(3.11)
α =
m1 + m2 − 1
2
=
n − 2
2
,
β =
m2 − 1
2
1
2
,
= −
and the half sum of positive roots for Hn is given by
ρ = α + β + 1 =
n − 1
2
.
Similarly for Hn+2l, l ∈ N the spherical function is equal to the Jacobi function φ(αl,βl)
where
λ
(3.12)
1
βl = −
2
and the half sum of positive roots for Hn+2l is given by
n + 2l − 1
ρl = αl + βl + 1 =
n + 2l − 2
αl =
2
,
,
= ρ + l.
2
If ξ ∈ Hn then using the Cartan decomposition of SO(n, 1) we can write ξ = kat · ξ0,
where ξ0 = (0,· · · , 1), k ∈ K/M = Sn−1 and
at =
cosh t
01×n−1
sinh t
01×n−1
sinh t
In−1×n−1 01×n−1
01×n−1
cosh t .
THEOREMS OF INGHAM AND CHERNOFF
17
We will need the following version of Hecke Bochner identity on Hn [8, Proposition 3.3.3]:
Lemma 3.7. If f (x) = f0(t)Yl(k) for x = kat.ξ0, where Yl is spherical harmonic of
degree l on K/M ∼= Sn−1 then
ef (λ, k) = dn,lQl(iλ − ρ)(cid:18)Z ∞
0
= dn,lQl(iλ − ρ)F (αl,βl)(cid:18) f0
(sinh t)l(cid:19) (λ)Yl(k),
f0(t)φH n+2l
λ
(t)(sinh t)2ρ+l dt(cid:19) Yl(k)
(3.13)
where φH n+2l
depending only on n and l, Ql(iλ − ρ) is a polynomial in λ given by
is the elementary spherical function on H n+2l, dn,l is some fixed constant
λ
(3.14)
Ql(iλ − ρ) =
l−1Ym=0
(iλ − ρ − m),
and F (αl,βl)f is the Jacobi transfrom of f defined in (3.10).
Example 3.8. Let ht be the heat kernel on Hn+2l ([3]). Since ht is a K-biinvariant
function on Hn+2l, using polar decomposition it can be viewed as an even function on
R and hence it can also be viewed as a K-biinvariant function on Hn. We now choose
a spherical harmonic Yl of degree l such that Yl(k0) = 0, for some k0 ∈ K/M. We now
define a function f on Hn by
f (ξ) = (sinh r)lh1(r)Yl(k),
(3.15)
It follows from the point wise estimate of the heat kernel ([3, (3.1)]) that f ∈ L2(G/K)
and
for ξ = kar · ξ0.
for all r in [0,∞). We now claim that
f (k0ar · ξ0) = 0,
(∆mf )(k0ar · ξ0) = 0,
for all r.
To prove this claim we will show that
(3.16)
for all ξ = kar · ξ0, where
∆mf (ξ) = (sinh r)l(cid:0)∆(αl,βl) + δ(cid:1)m
δ = (ρ + l)2 − ρ2 = ρ2
l − ρ2.
h1(r)Yl(k),
Taking Fourier transform the left hand sides of (3.16) and using Hecke Bochner identity
(3.13) we get
(3.17)
(∆mfe)(λ, k) = (cid:0)−(λ2 + ρ2)(cid:1)m ef (λ, k)
= (cid:0)−(λ2 + ρ2)(cid:1)m dn,lQl(iλ − ρ)F (αl,βl)(h1)(λ)Yl(k).
18
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
On the other hand, using (3.13) we get that the Fourier transform of the right hand side
of (3.16) is equal to
dn,lQl(iλ − ρ)F αl,βl(cid:0)(cid:0)∆(αl,βl) + δ(cid:1)m h1(cid:1) (λ)Yl(k)
= dn,lQl(iλ − ρ)(cid:0)−(λ2 + ρ2
l ) + δ(cid:1)m F αl,βlh1(λ)Yl(k)
= dn,lQl(iλ − ρ)(cid:0)−(λ2 + ρ2)(cid:1)m F αl,βlh1(λ)Yl(k),
which proves (3.16). Now, using (3.17) it follows that
k∆mfk2
L2(G/K) = k ^(∆mf )(λ, k)kL2(R×K,c(λ)−2dλdx)
= d2
n,lZR Ql(iλ − ρ)2(λ2 + ρ2)2m F (αl,βl)h1(λ)2 c(λ)−2 dλ.
Using (2.5), (3.14) we have
λ ≥ 1, for some n0,
c(λ)−2 ≤ Cλn0,
Ql(iλ − ρ)2 ≤ C(λ2 + ρ2)p0, for some p0 > 0 and
F (αl,βl)h1(λ) = e−λ2
.
Therefore
k∆mfk2
Consequently,
2
e−2λ2
L2(G/K) ≤ CZR(cid:0)λ2 + ρ2(cid:1)2(m+p0)
c(λ)−2 dλ
= C 2(m+p0) + C1Z ∞
+ ρ2(cid:17)2(m+p0) (cid:16)y
1 (cid:16) y
Z ∞
≤ C 2(m+p0) + C 2(m+p0)
Γ(cid:18)2m + 2p0 +
= C 2(m+p0) + C 2(m+p0)
Γ(cid:18)2m + 2p0 +
2 (cid:19) .
Γ(cid:18)2m + 2p0 +
≤ C 2(m+p0)
y2(m+p0) y
n0 + 1
n0 + 1
−2(m+p0)
n0−1
C
0
4m
0
1
2
2
Xm∈N k∆mfk− 1
2
2m
≥Xm∈N
4m
2 (cid:19)− 1
e−y dy
2
e−y dy
2
2(cid:17) n0−1
2 (cid:19)
n0 + 1
Now, using the fact that ([32, p. 30])
Γ(n + α)
Γ(n)nα = 1,
lim
n→∞
for α ∈ C,
it follows that
k∆mfk− 1
2
2m
−2(m+p0)
4m
−2(m+p0)
4m
≥ C
≥ C
(Γ(2m))− 1
(2m)− 1
2m (2m)− 4p0+n0+1
8m
2 (2m)− 4p0+n0+1
8m
.
Hence, for large m,
and
Therefore,
THEOREMS OF INGHAM AND CHERNOFF
19
k∆mfk− 1
2
2m
≥ C(2m)− 4m+4p0 +n0+1
8m
,
(2m)
4m+4p0 +n0+1
8m
≤ (2m)
8m
8m = 2m.
Xm≥m0
k∆mfk− 1
2
2m
≥ Xm≥m0
1
2m
= ∞.
This shows that the function f satisfies (1.3) and ∆mf (k0ar.ξ0) is zero for all m ∈ N∪{0}
and r in [0,∞].
4. Ingham's theorem for symmetric spaces
In this section we will prove Ingham's theorem (Theorem 1.4), using Theorem 1.3.
Proof of Theorem 1.4. We will first prove part (b) by reducing matters to Rd with the
help of Abel transform A. Since the integral in (1.4) is finite, we have
Z ∞
1
θ(r)
r
dr < ∞.
F h0(ξ) ≤ Ce−kξkθ(kξk),
Since θ is decreasing by part (b) of Theorem 1.1 for L positive, there exists a nontrivial
radial function h0 ∈ C∞c (Rd) with supp h0 ⊆ B(0, L/2) such that
for all ξ ∈ Rd.
(4.1)
Since h0 is a radial function on Rd, it can be thought of as a W -invariant function on
A ∼= Rd. Hence by Theorem 2.3, there exists h ∈ C∞c (G//K) such that A(h) = h0
with supp h ⊆ B(o, L/2). For a nontrivial φ ∈ C∞c (G//K) with support contained in
B(o, L/2), we consider the function f = h ∗ φ ∈ C∞c (G//K).
It follows from Paley-
Wiener theorem ([20, Theorem 7.1, Chapter IV]) that the support of f is contained
in B(o, L). Using the slice projection theorem (Theorem 2.3) and the estimate (4.1) it
follows that
Za∗ bf (λ) ekλkB θ(kλkB ) c(λ)−2 dλ
= Za∗ bh(λ) bφ(λ) ekλkB θ(kλkB ) c(λ)−2 dλ
= Za∗ F h0(λ) bφ(λ) ekλkB θ(kλkB ) c(λ)−2 dλ
≤ CZa∗ bφ(λ) c(λ)−2 dλ.
(4.2)
Since,bφ is a Schwartz function on a∗ it follows from the estimate (2.4) of c(λ)−2 that the
integral in (4.2) is finite and consequently, bf satisfies the condition (1.5). This completes
We will prove part (a) under the additional assumptions that f is continuous, K-
biinvariant and vanishes on an open ball centered at o. The general case then can be
the proof of part (b).
20
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
deduced from this case by mimicking the arguments given in the proof of [5, Theorem
1.2, steps 1-2].
So, we assume that f is K-biinvariant, continuous, integrable function which vanishes
on B(o, L) and satisfies the hypothesis (1.5)
(4.3)
We observe from (4.3) that bf ∈ L1(a∗,c(λ)−2dλ). As f is an integrable function, bf is
a bounded function and hence from (4.3) it follows that
(4.4)
Za∗ bf (λ) ekλkθ(kλkB ) c(λ)−2 dλ < ∞.
Za∗ bf (λ)2 eθ(kλkB )kλkB c(λ)−2dλ < ∞.
We now consider the following two cases as in [23].
(4.5)
Case I: Suppose θ satisfies the inequality
4
√r
θ(r) ≥
From (4.3) and (4.5), it follows that
,
for r ≥ 1.
(4.6)
and hence in particular
Za∗ bf (λ) e4√kλkB c(λ)−2dλ < ∞,
Za∗ bf (λ) kλkN
B c(λ)−2dλ < ∞,
for all N ∈ N. It follows that f ∈ C∞(G//K). To apply Theorem 1.3 we need to verify
condition (1.3). In this regard, an application of (4.4) implies that
k∆mfkL2(G//K) = (cid:18)Za∗(cid:0)kλk2
B(cid:1)2m bf (λ)2 c(λ)−2 dλ(cid:19) 1
B + kρk2
λ∈a∗(cid:0)kλk2
B(cid:1)m
B + kρk2
B + r2(cid:1)m e− rθ(r)
r∈(0,∞)(cid:0)kρk2
≤ sup
≤ C sup
kλkθ(kλkB
(4.7)
e−
2
2
2
,
(cid:18)Za∗ bf (λ)2 ekλkθ(kλkB ) c(λ)−2 dλ(cid:19) 1
2
and the latter quantity is finite by (4.5). In particular ∆mf ∈ L2(G//K) for all m ∈
N ∪ {0}. From now on we shall consider m ≥ max{2,kρkB}. To estimate the L∞ norm
(4.7) let us define
Then
gm(r) =(cid:0)kρk2
B + r2(cid:1)m
e− rθ(r)
2
,
r ∈ [0,∞).
If r ∈ (m4,∞) then by (4.5) we have
kgmkL∞[0,∞) ≤ kgmkL∞[0,1] + kgmkL∞[1,m4] + kgmkL∞(m4,∞).
B + r2)m e−2√r ≤ 2mr2me−2√r := γm(r) (say).
gm(r) ≤ (kρk2
THEOREMS OF INGHAM AND CHERNOFF
21
The function γm attains its maximum at r = 4m2. As m4 ≥ 4m2 and γm is decreasing
on (4m2,∞) we have
(4.8)
= (2m8e−2m)m.
:= ηm(r).
The function ηm attains its maximum at r = 4m/θ(m4). As m ≥ 2 we have
(4.10)
2
Also
(4.9)
For r ∈ [1, m4], as θ is a decreasing function,
.
e− θ(m4)r
r2me− θ(m4 )r
kgmkL∞(m4,∞) ≤ kγmkL∞(m4,∞) = 2mm8me−2m2
B(cid:1)m
kgmkL∞[0,1) ≤(cid:0)1 + kρk2
2 ≤(cid:0)1 + kρk2
B(cid:1)m
gm(r) ≤(cid:0)kρk2
B + r2(cid:1)m
B(cid:1)m(cid:18) 4m
θ(m4)(cid:19)2m
kηmkL∞[1,m4] ≤(cid:0)1 + kρk2
e−2m ≤(cid:0)1 + kρk2
θ(1)(cid:19)2m
θ(m4)(cid:19)2m
≥(cid:18) 4m
(cid:18) 4m
B(cid:1)m(cid:18) 4m
kgmkL∞[0,∞] ≤ 3(cid:0)1 + kρk2
θ(m4)(cid:19)2m
k∆mfkL2(G//K) ≤(cid:18) Cm
θ(m4)(cid:19)2m
> 1,
.
.
we have for all large m ∈ N
Since the right-hand side of (4.8) goes to zero as m goes to infinity and for all large
m ∈ N
B(cid:1)m(cid:18) 4m
θ(m4)(cid:19)2m
.
Therefore, using inequality (4.7) it follows from above that for all large m ∈ N and a
positive number C
(4.11)
Applying the change of variable kλkB = p4 in the integral (1.4) defining I, it follows
that
dp = ∞.
As θ is decreasing in [0,∞) this, in turn, implies that
= ∞.
θ(m4)
m
p
θ(p4)
1
Z ∞
Xm∈N
The inequality (4.11) then implies that
Xm∈N k∆mfk− 1
2m
L2(G//K) = ∞.
Since f vanishes on B(o, L), it follows from Theorem 1.3 that f vanishes identically.
This completes the proof under the assumption (4.5) on the function θ.
22
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
Case II. We now consider the general case, that is, θ is any nonnegative function de-
creasing to zero at infinity. Again, as in [23], we define
θ1(r) =
,
r ∈ [0,∞).
8
pr + 1
It is clear that the integral I in (1.4) is finite if θ is replaced by θ1. Hence, by case (b)
there exists a nontrivial f1 ∈ C∞c (G//K) such that suppf1 ⊆ B(o, L/2) and satisfies the
estimate
(4.12)
We now consider the function h = f ∗ f1 ∈ L1(G//K). Since f vanishes on B(o, L), the
function h vanishes on the open set B(o, L/2). Indeed, if g1K ∈ B(o, L/2) then for all
gK ∈ B(o, L/2) it follows by using G-invariance of the Riemannian metric d that
bf1(λ) ≤ Ce−kλkB θ1(kλkB ), λ ∈ a∗.
d(o, g1gK) ≤ d(o, g1K) + d(g1K, g1gK) < L.
that is, g1gK ∈ B(o, L). This implies that f (g1g) is zero for all gK ∈ B(o, L/2) = supp f1
and hence
f ∗ f1(g1) =ZG
f (g1g)f1(g−1) dg =Zsupp f1
f (g1g)f1(g−1) dg = 0.
We observe that
θ(r) + θ1(r) ≥
8
√r + 1
,
for r ≥ 1.
Using the estimate (4.12) and the hypothesis (4.3) it follows that
Za∗ bh(λ)ekλkB(cid:0)θ(kλkB )+θ1(kλkB )(cid:1)c(λ)−2 dλ < ∞.
Therefore h satisfies all the conditions used in case I and hence is the zero function.
This implies that
bh(λ) = bf (λ)bf1(λ) = 0,
Remark 4.1.
for almost every λ ∈ a∗. Since bf1 is a real analytic function, it follows that f vanishes
identically.
(cid:3)
(1) It is worth pointing out that Ingham's theorem (Theorem 1.4)
can also be proved directly using Lemma 3.2, without appealing to Theorem
1.3. For the sake of completeness we sketch the line of argument for part (a) of
Theorem 1.4. We assume that f is K-biinvariant, vanishes on B(o, L), satisfies
the hypothesis (1.5) and the function θ satisfies the estimate (4.5) with I = ∞.
In this case one can easily show that that the measure µ defined in Step 2 of the
proof of Theorem 1.3 is again a finite W -invariant measure on a∗. Using (4.3),
(4.5) and the arguments used in verifying (1.3), one can show that
(4.13)
Xk∈N
Mj(2k)− 1
2k = ∞,
THEOREMS OF INGHAM AND CHERNOFF
23
where Mj(k) is as given in Lemma 3.1. By the Fourier inversion (2.3) we get
that
This implies that for all u ∈ ΦL(a∗)
f (x) = W−1Za∗ bf (λ) φλ(x) c(λ)−2 dλ = 0,
Za∗ bf (λ) u(λ) c(λ)−2 dλ = 0.
(4.14)
if x ∈ B(o, L).
As in Step 3 of the proof of Theorem 1.3, it can be shown that ΦL(a∗) is dense
in L1(a∗, dµ)W . Therefore we can approximate bf by the elements of ΦL(a∗) and
(2) It is easy to see that part (a) of Theorem 1.4 remains true if the integral I in
hence by (4.14) we get that f is identically zero.
(1.2) is replaced by the integral
Z{λ∈a∗
θ(kλkB)
kλkη
B c(λ)−2dλ,
+ kλkB≥1}
where η = d + dim n, is the dimension of the symmetric space X. This follows
from the estimate (2.4) of c(λ)−2 as
Z ∞
1
θ(r)
r
dλ
dr = CZ{λ∈a∗
= CZ{λ∈a∗
≥ CZ{λ∈a∗
+ kλkB≥1}
+ kλkB≥1}
+ kλkB≥1}
B
θ(kλkB)
kλkd
θ(kλkB)
kλkη
θ(kλkB)
kλkη
B
dλ
B kλkdim n
B c(λ)−2dλ = ∞.
Moreover, because of the estimate (2.5), part (b) of Theorem 1.4 also remains
true in this case if the real rank of G is one.
An Lp version of Theorem 1.4 can actually be proved by using Theorem 1.4 itself. To
illustrate this we prove an L∞ version of the above theorem which can be thought of as
an exact analogue of Theorem 1.1.
Theorem 4.2. Let θ and I be as in Theorem 1.4.
ef (λ, k) ≤ Ce−kλkB θ(kλkB ),
(a) Suppose f ∈ L1(X) and the Fourier transform ef satisfies the estimate
for all λ ∈ a∗, k ∈ K.
(4.15)
If f vanishes on a nonempty open subset of X and I is infinite, then f = 0.
(b) If I is finite then given any L > 0, there exists a nontrivial f ∈ C∞c (G//K)
supported in B(o, L) satisfying the estimate (4.15).
Proof. As in Theorem 1.4, it suffices to prove the theorem for f ∈ L1(G//K) vanishing
on an open ball of the form B(o, L) such that bf satisfies the estimate
for all λ ∈ a∗+.
bf (λ) ≤ Ce−kλkB θ(kλkB ),
24
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
We choose a nonzero φ ∈ C∞c (G//K) with supp φ ⊆ B(o, L/2) and consider the function
f ∗ φ. Since f vanishes on B(o, L) and the support of the function φ is contained in
B(o, L/2) it follows as before that f ∗ φ vanishes on B(o, L/2). Now,
Za∗ [f ∗ φ(λ) ekλkB θ(kλkB ) c(λ)−2 dλ
= Za∗ bφ(λ) bf (λ) ekλkB θ(kλkB ) c(λ)−2 dλ
≤ CZa∗ bφ(λ) c(λ)−2 dλ < ∞.
It now follows from Theorem 1.4 that f ∗ φ is zero almost everywhere. Sincebφ is nonzero
almost everywhere we conclude that bf vanishes almost everywhere on a∗ and so does f .
To prove part (b) we observe that if I is finite then the function h constructed in the
(cid:3)
proof of Theorem 1.4, (b) satisfies the estimate (4.15).
Remark 4.3. Theorem 1.3 and Theorem 1.4 can be proved in the context of Damek-
Ricci spaces [2] by similar arguments.
It will be interesting to see whether both the
theorems hold true for hypergeometric transforms associated to root systems [29, 35].
References
[1] Anker, Jean-Philippe, A basic inequality for scattering theory on Riemannian symmetric spaces of
the noncompact type, Amer. J. Math. 113 (1991), no. 3, 391-398. MR1109344 (92k:43008)
[2] Anker, Jean-Philippe; Damek, Ewa; Yacoub, Chokri, Spherical analysis on harmonic AN groups,
Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 23 (1996), no. 4, 643-679 (1997). MR1469569 (99a:22014).
[3] Anker, Jean-Philippe; Ostellari, Patrick, The heat kernel on noncompact symmetric spaces. Lie
groups and symmetric spaces, 27-46, Amer. Math. Soc. Transl. Ser. 2, 210, Adv. Math. Sci., 54,
Amer. Math. Soc., Providence, RI, 2003. MR2018351 (2005b:58031)
[4] Bhowmik, M.; Ray, Swagato K.; Sen, S., Around theorems of Ingham-type regarding decay of
Fourier transform on Rn, Tn and two step nilpotent Lie Groups, to appear in Bull. Sci. Math.
[5] Bhowmik, M.; Ray, Swagato K., A theorem of Levinson for Riemannian symmetric spaces of
noncompact type, arXiv:1808.09710.
[6] Bhowmik, M.; Sen, S., Uncertainty principles of Ingham and Paley-Wiener on semisimple Lie
groups, Israel J. Math. 225 (2018), no. 1, 193221. MR3805646
[7] Bochner, S., Quasi-analytic functions, Laplace operator, positive kernels Ann. of Math. (2) 51
(1950), 68-91. MR0032708 (11,334g)
[8] Bray, William O., Aspects of harmonic analysis on real hyperbolic space. Fourier analysis, 77-102,
Lecture Notes in Pure and Appl. Math., 157, Dekker, New York, 1994. MR1277819
[9] Bray, William O. Generalized spectral projections on symmetric spaces of noncompact type: Paley-
Wiener theorems, J. Funct. Anal. 135 (1996), no. 1, 206-232. MR1367630
[10] Chernoff, Paul R., Quasi-analytic vectors and quasi-analytic functions, Bull. Amer. Math. Soc. 81
(1975), 637-646. MR0636276 (58#30432)
[11] Cowling, M.; Giulini, S.; Meda, S., Lp − Lq estimates for functions of the Laplace-Beltrami op-
erator on noncompact symmetric spaces I, Duke Math. J. 72 (1993), no. 1, 109-150. MR1242882
(95b:22031)
[12] de Branges, Louis, Hilbert spaces of entire functions. Prentice-Hall, Inc., Englewood Cliffs, N.J.
1968. MR0229011
[13] de Jeu, Marcel, Determinate multidimensional measures, the extended Carleman theorem and
quasi-analytic weights, Ann. Probab. 31 (2003), no. 3, 1205-1227. MR1988469 (2004f:44006)
THEOREMS OF INGHAM AND CHERNOFF
25
[14] de Jeu, Marcel, Subspaces with equal closure, Constr. Approx. 20 (2004), no. 1, 93157. MR2025416
(2005e:41083)
[15] Escauriaza, L.; Kenig, C. E.; Ponce, G.; Vega, L., Uniqueness properties of solutions to Schrdinger
equations, Bull. Amer. Math. Soc. (N.S.) 49 (2012), no. 3, 415-442. MR2917065
[16] Gangolli, R.; Varadarajan V. S., Harmonic Analysis of Spherical Functions on Real Reductive
Groups, Springer-Verlag, Berlin, 1988. MR954385 (89m:22015)
[17] Heckman, G. J. , Dunkl operators. S´eminaire Bourbaki, Vol. 1996/97. Ast´erisque No. 245 (1997),
Exp. No. 828, 4, 223-246. MR1627113
[18] Helgason, S., Differential geometry, Lie groups, and symmetric spaces, Graduate Studies in Math-
ematics, 34, American Mathematical Society, Providence, RI, 2001. MR1834454 (2002b:53081)
[19] Helgason, S.,Geometric Analysis on Symmetric Spaces, Mathematical Surveys and Monographs 39.
American Mathematical Society, Providence, RI, 1994. MR1280714 (96h:43009)
[20] Helgason, S., Groups and geometric analysis, Integral geometry, invariant differential operators,
and spherical functions, Mathematical Surveys and Monographs, 83. American Mathematical So-
ciety, Providence, RI, 2000. MR1790156 (2001h:22001)
[21] Helgason, S,. A duality for symmetric spaces with applications to group representations, Advances
in Math. 5, 1-154 (1970). MR0263988 (41 #8587)
[22] Hirschman, I. I., On the behaviour of Fourier transforms at infinity and on quasi-analytic classes
of functions, Amer. J. Math. 72 (1950), 200-213. MR0032816 (11,350f)
[23] Ingham, A. E., A Note on Fourier Transforms, J. London Math. Soc. S1-9 (1934), no. 1, 29-32.
MR1574706
[24] Koornwinder, Tom H., Jacobi functions and analysis on noncompact semisimple Lie groups. Special
functions: group theoretical aspects and applications, 1-85, Math. Appl., Reidel, Dordrecht, 1984.
MR0774055
[25] Koosis, P., The logarithmic integral I , Cambridge Studies in Advanced Mathematics, 12. Cam-
bridge University Press, Cambridge, 1998. xviii+606 pp. MR1670244 (99j:30001)
[26] Kotake, Takeshi; Narasimhan, Mudumbai S., Regularity theorems for fractional powers of a linear
elliptic operator, Bull. Soc. Math. France 90 1962 449-471. MR0149329
[27] Krotz, B; ´Olafsson, G; Stanton, R., The image of the heat kernel transform on Riemannian sym-
metric spaces of the noncompact type, Int. Math. Res. Not. 2005, no. 22, 1307-1329. MR2152539
[28] Levinson, N., On a Class of Non-Vanishing Functions, Proc. London Math. Soc. (2) 41 (1936), no.
5, 393-407. MR1576177
[29] Narayanan, E. K.; Pasquale, A.; Pusti, S., Asymptotics of Harish-Chandra expansions, bounded
hypergeometric functions associated with root systems, and applications, Adv. Math. 252 (2014),
227-259. MR3144230
[30] Opdam, E. M., Harmonic analysis for certain representations of graded Hecke algebras. Acta Math.
175 (1995), no. 1, 75-121. MR1353018
[31] Paley, R. E. A. C.; Wiener, N., Notes on the theory and application of Fourier transforms. I, II,
Trans. Amer. Math. Soc. 35 (1933), no. 2, 348-355. MR1501688
[32] Paris, R. B.; Kaminski, D., Asymptotics and Mellin-Barnes integrals. Encyclopedia of Mathematics
and its Applications, 85. Cambridge University Press, Cambridge, 2001. MR1854469
[33] Ray, Swagato K.; Sarkar, Rudra P., Chaotic behaviour of the Fourier multipliers on Riemannian
symmetric spaces of noncompact type, arXiv:1805.10048
[34] Rudin, W., Real and complex analysis Third edition. McGraw-Hill Book Co., New York, 1987.
MR0924157 (88k:00002)
[35] Schapira, Bruno, Contributions to the hypergeometric function theory of Heckman and Opdam:
sharp estimates, Schwartz space, heat kernel, Geom. Funct. Anal. 18 (2008), no. 1, 222250.
MR2399102 (2009f:33019)
[36] Stanton, Robert J.; Tomas, Peter A., Pointwise inversion of the spherical transform on Lp(G/K),
1 ≤ p < 2, Proc. Amer. Math. Soc. 73 (1979), no. 3, 398404. MR0518528.
26
MITHUN BHOWMIK, SANJOY PUSTI AND SWAGATO K. RAY
(Mithun Bhowmik, Sanjoy Pusti) Department of Mathematics, IIT Bombay, Powai,
Mumbai-400076, India
E-mail address: [email protected], [email protected]
(Swagato K Ray) Stat-Math Unit, Indian Statistical Institute, 203, B.T. Road, Kolkata-
700108, India
E-mail address: [email protected]
|
1607.00689 | 3 | 1607 | 2017-09-05T21:42:06 | Besov continuity for pseudo-differential operators on compact homogeneous manifolds | [
"math.FA"
] | In this paper we study the Besov continuity of pseudo-differential operators on compact homogeneous manifolds $M=G/K.$ We use the global quantization of these operators in terms of the representation theory of compact homogeneous manifolds. | math.FA | math | BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
ON COMPACT HOMOGENEOUS MANIFOLDS
DUV ´AN CARDONA
Abstract. In this paper we study the Besov continuity of pseudo-differential oper-
ators on compact homogeneous manifolds M = G/K. We use the global quantization
of these operators in terms of the representation theory of compact homogeneous
manifolds.
MSC 2010. Primary 19K56; Secondary 58J20, 43A65.
Contents
1.
Introduction
2. Main results
3. Pseudo-differential operators on compact Lie groups
3.1. Fourier analysis and Sobolev spaces on compact Lie groups
3.2. Differential and difference operators
3.3. Besov spaces
3.4. Global operators on compact homogeneous manifolds in Lebesgue spaces
4. Global pseudo-differential operators in Besov spaces
4.1. Examples
References
1
3
6
6
7
7
8
9
15
16
1. Introduction
In this work we study the mapping properties of pseudo-differential operators on
Besov spaces defined on compact Lie groups. The Besov spaces Br
p,q arose from attempts
to unify the various definitions of several fractional-order Sobolev spaces. In order to
illustrate the mathematical relevance of the Besov spaces, we recall that from the
context of the applications, a function belonging to some of these spaces admits a
decomposition of the form
7
1
0
2
p
e
S
5
]
.
A
F
h
t
a
m
[
3
v
9
8
6
0
0
.
7
0
6
1
:
v
i
X
r
a
f = f0 +Xj≥1
gj, gj = fj+1 − fj,
satisfying
kf kBr
p,q := k{2jrkgjkLp}∞
j=0klq(N) < ∞, g0 = f0, 0 < p, q ≤ ∞, r ∈ R,
(1.1)
(1.2)
Date: Received: xxxxxx; Revised: yyyyyy; Accepted: zzzzzz.
.
2010 Mathematics Subject Classification. 19K56; Secondary 58J20, 43A65.
Key words and phrases. Besov space, Compact homogeneous manifold, Pseudo-differential opera-
tors, Global analysis.
1
2
D. CARDONA
where the sequence (fj)j consists of approximations to the data (or the unknown) f of
a given problem at various levels of resolution indexed by j. In practice such approx-
imations can be defined by using the Fourier transform and this description is useful
in the numerical analysis of wavelet methods, and some areas of applied mathematics
as signal analysis and image processing. In the field of numerical analysis, multi-scale
and wavelet decompositions, Besov spaces have been used for three main task: precon-
ditioning large systems arising from the discretization of elliptic differential problems,
adaptive approximations of functions which are not smooth, and sparse representation
of initially full matrices arising in the discretization of integral equations (see A. Cohen
[9]).
In our work, we are interesting in the study of pseudo-differential problems associ-
ated to Besov spaces on compact Lie groups and more generally compact homogeneous
manifolds. More precisely, we want to study the action of pseudo-differential operators
on these spaces. We will use the formulation of Besov spaces on compact homogeneous
manifolds in terms of representation theory as in ([21]). Our main goal is to show
that, under certain conditions, the Lp boundedness of Fourier multipliers on compact
homogeneous manifolds gives to rise to results of continuity for pseudo-differential op-
erators on Besov spaces. In our analysis we use the theory of global pseudo-differential
operators on compact Lie groups and on compact homogeneous manifolds, which was
initiated in the PhD thesis of V. Turunen and was extensively developed by M. Ruzhan-
sky and V. Turunen in [24]. In this theory, every operator A mapping C ∞(G) itself,
where G is a compact Lie group, can be described in terms of representations of G as
irreducible unitary representations on G), the Ruzhansky-Turunen approach establish
follows. Let bG be the unitary dual of G (i.e, the set of equivalence classes of continuous
that A has associated a matrix-valued global (or full) symbol σA(x, ξ) ∈ Cdξ×dξ , [ξ] ∈ bG,
on the non-commutative phase space G × bG satisfying
(1.3)
σA(x, ξ) = ξ(x)∗(Aξ)(x) := ξ(x)∗[Aξij(x)]
dξ
i,j=1.
Then it can be shown that the operator A can be expressed in terms of such a symbol
as [24]
In the last five years, applications of this theory have been considered by many authors.
Advances in this framework includes the characterization of Hormander classes Sm
ρ,δ(G)
on compact Lie groups in terms of the representation theory of such groups (c.f [27] ),
the sharp Garding inequality on compact Lie groups, (c.f [25]), the behavior of Fourier
multipliers in Lp(G) spaces (c.f.
[29]), global functional calculus of operator on Lie
groups (c.f [28]), r-nuclearity of operators, Grothendieck-Lidskii formula and nuclear
[10, 12, 13]), the Gohberg lemma,
traces of operators on compact Lie groups (c.f.
characterization of compact operators, and the essential spectrum of operators on L2
(c.f [11]), Lp-boundedness of pseudo-differential operators in Hormander classes (c.f.
[14]), Besov continuity and nuclearity of Fourier multipliers on compact Lie groups (c.f
[5, 6, 7]), diffusive wavelets on groups and homogeneous spaces [15], and recently, a
reformulation of Ruzhansky and Turunen approach on the pseudo-differential calculus
in compact Lie groups (c.f. Fischer, V [17]), including a version of the Calder´on-
Vaillancourt Theorem in this framework.
Af (x) = X[ξ]∈ bG
dξTr[ξ(x)σA(x, ξ)bf (ξ)].
(1.4)
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
3
In the euclidean case of Rn, the Hormander's symbol class Sm
ρ,δ(Rn), m ∈ R and
0 ≤ δ, ρ ≤ 1, is defined by those functions a(x, ξ), x, ξ ∈ Rn satisfying
∂β
x ∂α
ξ a(x, ξ) ≤ Cα,βhξim−ρα+δβ, α, β ∈ Nn,
(1.5)
hξi = (1 + ξ)2. The corresponding pseudo-differential operator A with symbol a(·, ·)
is defined on the Schwartz space S (Rn) by
Af (x) =Z ei2πx·ξa(x, ξ)bf (ξ)dξ.
(1.6)
Consequently, on every differential manifold M, pseudo-differential operators A : C ∞(M ) →
C ∞(M ) associated to Hormander classes Sm
ρ,δ(M ), 0 ≤ δ < ρ ≤ 1, ρ ≥ 1 − δ can be
defined by the use of coordinate charts. When M = G is a compact Lie group and
1 − ρ ≤ δ, the exceptional results in [27] gives an equivalence of the Hormander classes
defined by charts and Hormander classes defined in terms of the representation theory
of the group G.
If K is a closed subgroup of a compact Lie group G, there is a canonical way to iden-
tify the quotient space M = G/K with a analytic manifold. Besov spaces on compact
Lie groups and general compact homogeneous manifolds where introduced in terms of
representations and analyzed in [21], they form scales Br
p,q(M ) carrying three indices
r ∈ R, 0 < p, q ≤ ∞. For 1 ≤ p < ∞, 1 ≤ q ≤ ∞ the Besov spaces Br
p,q(M ) coincide
p,q(Rn).
with the Besov spaces defined trough of localization with the euclidean space Br
1,δ(Rn), 0 ≤ δ < 1, then the corresponding operator
It is well known that if a ∈ Sm
A : Br+m
p,q(Rn) is bounded for 1 < p < ∞, 1 ≤ q < ∞ and r ∈ R, (c.f Bor-
daud [4], and Gibbons [18]). This implies the same result for compact Lie groups when
1 < p < ∞ and 1 ≤ q ≤ ∞. With this fact in mind, in order to obtain Besov continuity
for operators, we concentrate our attention to pseudo-differential operators A whose
symbols a = σA have limited regularity almost in one of the variables x, ξ. (Since, ξ in
the case of compact Lie groups has a discrete nature, the notion of differentiation is
related with difference operators). The results of this paper have been announced in
[6]. The Besov continuity of multipliers in the context of graded Lie groups has been
considered by the author and M. Ruzhansky in [8].
(Rn) → Br
p,q
This paper is organized as follows. In section 2 we present and briefly discuss our
main theorems. In Section 3, we summarizes basic properties on the harmonic analy-
sis in compact Lie groups including the Ruzhansky-Turunen theory of global pseudo-
differential operators on compact-Lie groups and the definition of Besov spaces on such
groups. Finally, in section 4 we proof our results on the boundedness of invariant
and non-invariant pseudo-differential operators on Besov spaces and some examples on
differential problems are given in Subsection 4.1.
2. Main results
In this section we present and briefly discuss our main theorems. The following is a
generalization of Theorem 1.2 of [5] to the case of homogeneous compact manifolds.
Theorem 2.1. Let M := G/K be a compact homogeneous manifold and let A = Op(σ)
be a Fourier multiplier on M . If A is bounded from Lp1(M ) into Lp2(M ), then A extends
to a bounded operator from Br
p2,q(M ), for all r ∈ R, 1 ≤ p1, p2 ≤ ∞, and
0 < q ≤ ∞.
p1,q(M ) into Br
4
D. CARDONA
As a consequence of this fact, we establish the following theorems. First, we present
a theorem on the boundedness of operators on compact homogeneous spaces.
Theorem 2.2. Let us consider A : C ∞(G/K) → C ∞(G/K) be a pseudo-differential
operator on the compact homogeneous manifold G/K. Let n = dim(G/K) and 1 < p1 ≤
2 ≤ p2 < ∞. Let us assume that the (global) matrix valued symbol a(x, π) of A satisfies
in terms of the Plancherel measure µ of bG0 the inequality,
x a(x, π)kop > s}]
1
p1
− 1
p2 < ∞,
sup
s>0
s[µ{π ∈ bG0 : k∂β
for all β ≤ [ n
p1
Br
p2,q(G/K), for all r ∈ R and 0 < q ≤ ∞.
] + 1. Then A extends to a bounded operator from Br
p1,q(G/K) into
(2.1)
It is important to mention that the Theorem above can be obtained as consequence
of Theorem 2.1 and the results in [1].
Remark 2.3. A classical result by Hormander (Theorem 1.11 of [20]) establish the
boundedness of a Fourier multiplier of the form (1.6) from Lp1(Rn) into Lp2(Rn) if its
symbol a(x, ξ) := a(ξ) satisfies the relation
s[µ{ξ ∈ Rn : a(ξ) > s}]
1
p1
− 1
p2 < ∞,
(2.2)
sup
s>0
where µ is the Lebesgue measure and 1 < p1 ≤ 2 ≤ p2 < ∞. For compact homogeneous
manifolds M := G/K, Ruzhansky, Akylzhanov and Nursultanov [1] have obtained the
boundedness from Lp1(G) into Lp2(G) of pseudo-differential operators A with symbols
satisfying the condition (2.1) and 1 < p1 ≤ 2 ≤ p2 < ∞. (See also the references [2]
and [3]).
Now for the case of compact Lie groups, which are important cases of homogeneous
manifolds, we have the following theorems on boundedness of operators associated to
symbols satisfying conditions of Hormander type.
Theorem 2.4. Let G be a compact Lie group, n = dim(G) and let 0 ≤ ρ, δ ≤ 1. Denote
by κ the smallest even integer larger that n
p ] + 1. Let A
from C ∞(G) into C ∞(G) be a pseudo-differential operator with global symbol a(x, ξ)
satisfying
2 . Let 1 < p < ∞ and l = [ n
kDα
with m ≥ κ(1 − ρ) 1
Br
ξ ∂β
p − 1
x a(x, ξ)kop ≤ Cα,βhξi−m−ρα+δβ, α ≤ κ, β ≤ l,
(2.3)
2 + δl. Then A extends to a bounded operator from Br
p,q(G) into
p,q(G) for all 0 < q ≤ ∞ and r ∈ R. Moreover, if we assume that
x a(x, ξ)kop ≤ Cα,βhξi−ρα, α ≤ κ, β ≤ l,
kDα
ξ ∂β
(2.4)
then A extends to a bounded operator from B
1 < p < ∞, 0 < q ≤ ∞ and r ∈ R.
r+κ(1−ρ) 1
p,q
p − 1
2
(G) into Br
p,q(G) for all
Theorem 2.5. Let G be a compact Lie group, n = dim(G), 0 ≤ ρ < 1 and 0 ≤ ν <
n
2 (1 − ρ). Let A from C ∞(G) into C ∞(G) be a pseudo-differential operator with global
symbol a(x, ξ) satisfying
ξ ∂β
x a(x, ξ)kop ≤ Cα,βhξi−ν−ρα, α ∈ Nn, β ≤ l,
kDα
(2.5)
with 1 < p < ∞ and l = [ n
into Br
p,q(G) for all p with 1
p ] + 1. Then A extends to a bounded operator from Br
p,q(G)
p − 1
2 ≤ ν
n (1 − ρ)−1, 0 < q ≤ ∞ and r ∈ R.
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
5
Now, we provide some remarks on our main results.
• Theorems 2.4 and (4.4) can be proved by using Theorem (2.1), and the Lp
boundedness theorems in [14]. For the definition of the difference operators Dα
ξ ,
α ∈ Nn (which were introduced in [29]) we refer to Definition 3.2.
• Recently in [5], Theorem 1.6, the boundedness of pseudo-differential operators
A on every Br
p,q(G)-space with symbols (of order zero) satisfying
ξ ∂β
x a(x, ξ)kop ≤ Cα,βhξi−α, α ≤ κ, β ≤ l,
kDα
(2.6)
has been shown. This result has been obtained as consequence of the Lp(G)-
boundedness of Fourier multipliers with symbols a(ξ) satisfying the analogous
condition
kDα
ξ a(ξ)kop ≤ Cαhξi−α, α ≤ κ.
(2.7)
• In Theorem 2.4, the condition (2.4), generalizes the Theorem 1.6 of [5] (in fact,
we only need to consider ρ = 1). In the hypothesis (2.3) we do not use the usual
conditions ρ > δ and ρ ≥ 1 − δ for the invariance under coordinates charts of
the Hormander classes Ψm
ρ,δ(G) (see [20]).
• For operators with symbols whose derivatives Dα
ξ a(ξ) are bounded by Cαhξi−m−ρα
in the operator norm, m → 0+, the Lp- boundedness is valid only for finite in-
tervals centered at p = 2, (c.f Delgado and Ruzhansky [14]). Since our Besov
estimates are obtained from these Lp-estimates, we obtain the boundedeness of
operators A on Br
p,q(G) around of p = 2.
• It was proved by Fischer that for ρ ≥ 1 − δ, ρ > δ, an operator A in the
Hormander class Ψm
ρ,δ(G) has matrix valued symbol a = σA satisfying
p,q
kDα
ξ ∂β
(Rn) into Br
x σA(x, ξ)kop ≤ Cα,βhξim−ρα+δβ, α, β ∈ Nn.
(2.8)
1,δ(Rn),
The Bordaud result which asserts that every operator A with symbol in Ψm
0 ≤ δ < 1, is a bounded operator from Br+m
p,q(Rn) implies the same
result for the classes Ψm
1,δ(G). The boundedness on Besov spaces for operators
associated to Hormander classes with 0 < ρ < 1 has a different behavior to
the case ρ = 1. In fact, as consequence of the results in Park [23], an operator
p,q(Rn)
A with symbol in Ψm
under the condition s = m + r + n(1 − ρ) 1
2 (notice that for ρ = 1 this
result is nothing else that the Bordaud theorem). In particular if m = 0, then
p,q(Rn) is bounded. Again, the Fischer result
A : B
mentioned above together with the Park result implies for ρ ≥ 1 − δ, ρ > δ,
σA ∈ Ψm
p,q(G), for all
r ∈ R. The novelty of our results is that we consider matrix valued symbols of
limited regularity, and we do not impose the condition ρ ≥ 1 − δ, ρ > δ.
ρ,δ(Rn), ρ ≥ δ, is bounded from Bs
ρ,δ(G) the boundedness of A : B
p,q(Rn) into Br
r+n(1−ρ) 1
p,q
r+n(1−ρ) 1
p,q
(Rn) → Br
(G) → Br
p − 1
p − 1
2
p − 1
2
• Although in our main results we consider symbols with order less than or equal
to zero, these results can be extended to symbols of arbitrary order by using
standard techniques. In fact, if A : C ∞(G) → C ∞(G) is a linear and bounded
operator, then under any one of the following conditions
-- kDα
-- kDα
ξ ∂β
ξ ∂β
x σA(x, ξ)kop ≤ Cαhξim−α, for all
x σA(x, ξ)kop ≤ Cα,βhξim−ν−ρα, α ∈ Nn, β ≤ l,
α ≤ κ,
β ≤ l,
p − 1
1
2 ≤ ν
n (1 −
ρ)−1, 0 ≤ ν < n
2 (1 − ρ),
-- kDα
ξ ∂β
x a(x, ξ)kop ≤ Cα,βhξim−κ(1−ρ) 1
p − 1
2 −δl−ρα+δβ, α ≤ κ, β ≤ l,
6
D. CARDONA
the corresponding pseudo-differential operator A extends to a bounded operator
from Br+m
p,q(G) for all r ∈ R, 1 < p < ∞ and 0 < q ≤ ∞.
Additionally, we observe that the condition
(G) into Br
p,q
-- kDα
ξ ∂β
x a(x, ξ)kop ≤ Cα,βhξim−ρα, α ≤ κ, β ≤ l,
(G) into
p,q(G) for all 1 < p < ∞, 0 < q ≤ ∞ and r ∈ R similar to the Park result for
assures that A extends to a bounded operator from B
Br
euclidean symbols.
m+r+κ(1−ρ) 1
p,q
p − 1
2
3. Pseudo-differential operators on compact Lie groups
3.1. Fourier analysis and Sobolev spaces on compact Lie groups. In this section
we will introduce some preliminaries on pseudo-differential operators on compact Lie
groups and some of its properties on Lp-spaces. There are two notions of pseudo-
differential operators on compact Lie groups. The first notion in the case of general
manifolds (based on the idea of local symbols) and, in a much more recent context,
the one of global pseudo-differential operators on compact Lie groups as defined by
Ruzhansky and Turunen [24] (see also [26]). We adopt this last notion for our work. We
will always equip a compact Lie group with the Haar measure µG. For simplicity, we will
writeRG f dx forRG f (x)dµG(x), Lp(G) for Lp(G, µG), etc. The following assumptions
are based on the group Fourier transform
The Peter-Weyl Theorem on G implies the Plancherel identity on L2(G),
ϕ(x)ξ(x)∗dx,
bϕ(ξ) =ZG
kϕkL2(G) =X[ξ]∈ bG
ϕ(x) = X[ξ]∈ bG
dξTr(bϕ(ξ)bϕ(ξ)∗)
1
2
kAkHS = Tr(AA∗),
dξTr(ξ(x)bϕ(ξ)).
= kbϕkL2( bG).
Here
denotes the Hilbert-Schmidt norm of matrices. Any linear operator A on G mapping
C ∞(G) into D′(G) gives rise to a matrix-valued global (or full) symbol σA(x, ξ) ∈ Cdξ×dξ
given by
(3.1)
which can be understood from the distributional viewpoint. Then it can be shown that
the operator A can be expressed in terms of such a symbol as [24]
σA(x, ξ) = ξ(x)∗(Aξ)(x),
Af (x) = X[ξ]∈ bG
dξTr[ξ(x)σA(x, ξ)bf (ξ)].
(3.2)
In this paper we use the notation Op(σA) = A. Lp(bG) spaces on the unitary dual can
be well defined. If p = 2, L2(bG) is defined by the norm
dξkΓ(ξ)k2
kΓk2
L2( bG)
HS .
= X[ξ]∈ bG
Now, we want to introduce Sobolev spaces and, for this, we give some basic tools.
Let ξ ∈ Rep(G) := ∪bG = {ξ : [ξ] ∈ bG}, if x ∈ G is fixed, ξ(x) : Hξ → Hξ is an
unitary operator and dξ := dim Hξ < ∞. There exists a non-negative real number λ[ξ]
depending only on the equivalence class [ξ] ∈ G, but not on the representation ξ, such
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
7
that −LGξ(x) = λ[ξ]ξ(x); here LG is the Laplacian on the group G (in this case, defined
as the Casimir element on G). Let hξi denote the function hξi = (1 + λ[ξ])
Definition 3.1. For every s ∈ R, the Sobolev space H s(G) on the Lie group G is defined
2 .
1
by the condition: f ∈ H s(G) if only if hξisbf ∈ L2(bG). The Sobolev space H s(G) is
a Hilbert space endowed with the inner product hf, gis = hΛsf, ΛsgiL2(G), where, for
every r ∈ R, Λs : H r → H r−s is the bounded pseudo-differential operator with symbol
hξisIξ. In Lp spaces, the p-Sobolev space of order s, H s,p(G), is defined by functions
satisfying
kf kH s,p(G) := kΛsf kLp(G) < ∞.
(3.3)
3.2. Differential and difference operators. In order to classify symbols by its reg-
ularity we present the usual definition of differential operators and the difference oper-
ators used introduced in [29].
j=1
Definition 3.2. Let (Yj)dim(G)
be a basis for the Lie algebra g of G, and let ∂j be
the left-invariant vector fields corresponding to Yj. We define the differential operator
associated to such a basis by DYj = ∂j and, for every α ∈ Nn, the differential operator
x is the one given by ∂α
∂α
n . Now, if ξ0 is a fixed irreducible representation,
dξ0
i,j=1 = ξ0(·) − Idξ0
the matrix-valued difference operator is the given by Dξ0 = (Dξ0,i,j)
.
If the representation is fixed we omit the index ξ0 so that, from a sequence D1 =
Dξ0,j1,i1, · · · , Dn = Dξ0,jn,in of operators of this type we define Dα
n , where
α ∈ Nn.
1 · · · Dαn
1 · · · ∂αn
ξ = Dα1
x = ∂α1
3.3. Besov spaces. We introduce the Besov spaces on compact Lie groups using the
Fourier transform on the group G as follow.
Definition 3.3. Let r ∈ R, 0 ≤ q < ∞ and 0 < p ≤ ∞. If f is a measurable function
on G, we say that f ∈ Br
p,q(G) if f satisfies
If q = ∞, Br
p,∞(G) consists of those functions f satisfying
kf kBr
p,q :=
∞Xm=0
kf kBr
p,∞ := sup
m∈N
2mrqk X2m≤hξi<2m+1
2mrk X2m≤hξi<2m+1
< ∞.
1
q
Lp(G)
dξTr[ξ(x)bf (ξ)]kq
dξTr[ξ(x)bf (ξ)]kLp(G) < ∞.
(3.4)
(3.5)
(3.6)
If we denote by Op(χm) the Fourier multiplier associated to the symbol
we also write,
χm(η) = 1{[ξ]:2m≤hξi<2m+1}(η),
kf kBr
p,q = k{2mrkOp(χm)f kLp(G)}∞
m=0klq(N), 0 < p, q ≤ ∞, r ∈ R.
Remark 3.4. For every s ∈ R, H s,2(G) = H s(G) = Br
2,2(G). Besov spaces according
to Definition (3.3) were introduced in [21] on compact homogeneous manifolds where,
in particular, the authors obtained its embedding properties. On compact Lie groups
such spaces where characterized, via representation theory, in [22].
Remark 3.5. In connection with our comments in the introduction, a detailed descrip-
tion on euclidean models and the role of Besov spaces in the context of applications
we refer the reader to the book of A. Cohen [9]. The reference Hairer, [19] explains
8
D. CARDONA
the importance of the Besov spaces in the setting of the theory of regularity structures
as well as a theorem of reconstruction and some interactions with stochastic partial
differential equations; on the other hand, as it was pointed out in in [16] (see also
references therein) several problems in signal analysis and information theory require
non-euclidean models. These models include: spheres, projective spaces and general
compact manifolds, hyperboloids and general non-compact symmetric spaces, and fi-
nally various Lie groups. In connection with these spaces it is important to study Besov
spaces on compact and non-compact manifolds.
3.4. Global operators on compact homogeneous manifolds in Lebesgue spaces.
Now we introduce the notion of homogeneous manifold. Let us consider a closed sub-
group K of G and identify M := G/K as a analytic manifold in a canonical way. (In
the case K = {e} where e is the identity element of the group G, we identify G/K
respect to the subgroup K. This means that π(h)(a) = a for all h ∈ K. Besov spaces
on homogeneous manifolds M = G/K can be defined, and the Besov norms are de-
with G). Let us denote by bG0 the subset of bG that are representations of type I with
fined as in (3.4) y (3.5), but the representations ξ in the sums are in bG0. The following
Lp −Lq-theorem will be useful in our analysis of Besov continuity for pseudo-differential
operators on homogeneous manifolds (c.f. [1]).
Theorem 3.6. Let us consider A : C ∞(G/K) → C ∞(G/K) be a pseudo-differential
operator operator on the compact homogeneous manifold G/K. Let n = dim(G/K) and
1 < p1 ≤ 2 ≤ p2 < ∞. Let us assume that the (global) symbol matrix valued a(x, π) of
A satisfies
sup
s>0
s[µ{π ∈ bG0 : k∂β
x a(x, π)kop > s}]
1
p1
− 1
p2 < ∞,
(3.7)
for all β ≤ [ n
p1
] + 1. Then A extends to a bounded operator from Lp1(G) into Lp2(G).
The following sharp Lp theorem on G allow us to investigate Besov continuity for
pseudo-differential operators on compact Lie groups. (c.f. Delgado and Ruzhansky
[14]).
Theorem 3.7. Let G be a Compact Lie group, n = dim(G) and let 0 ≤ ρ, ≤ δ ≤ 1.
Denote by κ the smallest even integer larger that n
p ] + 1.
Let A : C ∞(G) into C ∞(G) be a pseudo-differential operator with global symbol a(x, ξ)
satisfying
2 . Let 1 < p < ∞ and l = [ n
kDα
ξ ∂β
x a(x, ξ)kop ≤ Cα,βhξi−m−ρα+δβ, α ≤ κ, β ≤ l,
(3.8)
with m ≥ κ(1 − ρ) 1
Lp(G).
p − 1
2 + δl. Then A extends to a bounded operator from Lp(G) into
Theorem 3.8. Let G be a compact Lie group, n = dim(G), 0 ≤ ρ < 1 and 0 ≤
2 (1 − ρ). Let A : C ∞(G) into C ∞(G) be a pseudo-differential operator with global
ν < n
symbol a(x, ξ) satisfying
kDα
ξ ∂β
x σ(x, ξ)kop ≤ Cα,βhξi−ν−ρα+δβ, α ∈ Nn, β ≤ l,
(3.9)
with 1 < p < ∞ and l = [ n
into Lp(G) for all p with 1
p ] + 1. Then A extends to a bounded operator from Lp(G)
p − 1
n (1 − ρ)−1.
2 ≤ ν
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
9
4. Global pseudo-differential operators in Besov spaces
In this section we prove our main results. For the case of compact Lie groups,
Our starting point is the following theorem, which gives a relationship between Lp
boundedness and Besov continuity on homogeneous compact manifolds. A Fourier
multiplier on M := G/K is an operator A = Op(σ) with symbol σ(ξ) satisfying σ(ξ)ij =
0 for i > kξ or j > kξ, [ξ] ∈ bG0.
Theorem 4.1. Let M := G/K be a compact homogeneous manifold and let A = Op(σ)
be a Fourier multiplier on M . If A is bounded from Lp1(M ) into Lp2(M ), then A extends
to a bounded operator from Br
p2,q(M ), for all r ∈ R, 1 ≤ p1, p2 ≤ ∞, and
0 < q ≤ ∞.
p1,q(M ) into Br
Proof. First, let us consider a multiplier operator Op(σ) bounded from Lp1(M ) into
Lp2(M ), and f ∈ C ∞(M ). Then, we have kT f kLp2 (M ) ≤ Ckf kLp1 (M ), where C =
kT kB(Lp1 ,Lp2 ) is the usual operator norm. We denote by χm(ξ) the characteristic func-
tion of Dm := {ξ ∈ bG0 : 2m ≤ hξi < 2m+1} and Op(χm) the corresponding Fourier
multiplier of the symbol χm(ξ)Iξ. Here, Iξ := (aij) is the matrix in Cdξ×dξ , defined by
aii = 1 if 1 ≤ i ≤ kξ and aij = 0 in other case. By the definition of Besov norm, if
0 < q < ∞ we have
dξTr[ξ(x)F (Op(σ)f )(ξ)]kq
Lp2 (M )
dξ · χm(ξ)Tr[ξ(x)F (Op(σ)f )(ξ)]kq
Lp2 (M )
2mrqk X2m≤hξi<2m+1
2mrqk X[ξ]∈ bG0
2mrqk X[ξ]∈ bG0
2mrqk X[ξ]∈ bG0
kOp(σ)f kq
Br
=
p2 ,q
=
=
=
=
∞Xm=0
∞Xm=0
∞Xm=0
∞Xm=0
∞Xm=0
dξ · Tr[ξ(x)χm(ξ)σ(ξ)(F f )(ξ)]kq
Lp2 (M )
dξ · Tr[ξ(x)σ(ξ)F (Op(χm)f )(ξ)]kq
Lp2 (M )
2mrqkOp(σ)[(Op(χm)f )]kq
Lp2 (M ).
10
D. CARDONA
By the boundedness of Op(σ) from Lp1(M ) into Lp2(M ) we get,
2mrqC qkOp(χm)f kq
Lp1 (M )
kOp(σ)f kq
Br
≤
p2 ,q
=
=
=
∞Xm=0
∞Xm=0
∞Xm=0
∞Xm=0
2mrqC qk X[ξ]∈ bG0
2mrqC qk X2m≤hξi<2m+1
2mrqC qk X2m≤hξi<2m+1
= C qkf kq
Br
p1 ,q
dξ · Tr[ξ(x)χm(ξ)IξF (f )(ξ)]kq
Lp1 (M )
dξ · Tr[ξ(x)IξF (f )(ξ)]kq
Lp1 (M )
dξ · Tr[ξ(x)IξF (f )(ξ)]kq
Lp1 (M )
Hence,
kOp(σ)f kBr
p2 ,q ≤ Ckf kBr
p1 ,q .
If q = ∞ we have
kOp(σ)f kBr
p2,∞ = sup
m∈N
= sup
m∈N
= sup
m∈N
= sup
m∈N
= sup
m∈N
2mrk X2m≤hξi<2m+1
2mrk X[ξ]∈ bG
2mrk X[ξ]∈ bG
2mrk X[ξ]∈ bG
dξTr[ξ(x)F (Op(σ)f )(ξ)]kLp2 (M )
dξ · χm(ξ)Tr[ξ(x)F (Op(σ)f )(ξ)]kLp2 (M )
dξ · Tr[ξ(x)χm(ξ)σ(ξ)(F f )(ξ)]kLp2 (M )
dξ · Tr[ξ(x)σ(ξ)F (Op(χm)f )(ξ)]kLp2 (M )
2mrkOp(σ)[(Op(χm)f )]kLp2 (M ).
Newly, by using the fact that Op(σ) is a bounded operator from Lp1(M ) into Lp2(M )
we have,
2mrCkOp(χm)f kLp1 (M )
kOp(σ)f kBr
p2,∞ ≤ sup
m∈N
= sup
m∈N
= sup
m∈N
= sup
m∈N
2mrCk X[ξ]∈ bG
2mrCk X2m≤hξi<2m+1
2mrCk X2m≤hξi<2m+1
= Ckf kBr
p1 ,∞.
dξ · Tr[ξ(x)χm(ξ)IξF (f )(ξ)]kLp1 (M )
dξ · Tr[ξ(x)IξF (f )(ξ)]kLp1 (M )
dξ · Tr[ξ(x)IξF (f )(ξ)]kLp1 (M )
This implies that,
With the last inequality we end the proof.
(cid:3)
kOp(σ)f kBr
p2,∞ ≤ Ckf kBr
p1 ,∞.
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
11
Theorem 4.2. Let us consider A : C ∞(G/K) → C ∞(G/K) be a pseudo-differential
operator operator on the compact homogeneous manifold G/K. Let n = dim(G/K) and
1 < p1 ≤ 2 ≤ p2 < ∞. Let us assume that the (global) matrix valued symbol a(x, π) of
A satisfies in terms of the Plancherel measure µ on bG0 the inequality,
− 1
1
p1
p2 < ∞,
(4.1)
sup
s>0
s[µ{π ∈ bG0 : k∂β
x a(x, π)kop > s}]
for all β ≤ [ n
p1
Br
p2,q(G/K), for all r ∈ R and 0 < q ≤ ∞.
] + 1. Then A extends to a bounded operator from Br
p1,q(G/K) into
Proof. If we assume that A has symbol σ(x, π) = σ(π) independent of x ∈ M = G/K,
then by Theorem 3.6 we have that A is bounded from Lp1(M ) into Lp2(M ). By Theorem
2.1, A extends to a bounded operator from Br
p2,q(M ), for all r ∈ R and
0 < q ≤ ∞. Next, we consider the general case where a(x, π) depends on x. To do this
we write for f ∈ C ∞(M ) :
p1,q(M ) into Br
Af (x) = X[ξ]∈ bG0
dξTr[ξ(x)σ(x, ξ)bf (ξ)]
=ZM
[ X[ξ]∈ bG0
=ZM
[ X[ξ]∈ bG0
dξTr[ξ(y−1x)σ(x, ξ)]]f (y)dy
dξTr[ξ(y)σ(x, ξ)]]f (xy−1)dy.
Hence A = Op(σ)f (x) = (κ(x, ·) ∗ f )(x), where
κ(z, y) = X[ξ]∈ bG0
dξTr[ξ(y)σ(z, ξ)],
(4.2)
and ∗ is the right convolution operator. Moreover, if we define Azf (x) = (κ(z, ·) ∗ f )(x)
for every element z ∈ M, we have
Axf (x) = Af (x), x ∈ M.
For all 0 ≤ β ≤ [n/p]+ 1 we have ∂β
argument on Fourier multipliers, for every z ∈ M, ∂β
operator from Br
z σ(z, ·))f (x). So, by the precedent
z σ(z, ·))f is a bounded
p2,q for all r ∈ R and 0 < q ≤ ∞. Now, we want to
z Azf (x) = Op(∂β
p1,q(M ) into Br
z Azf = Op(∂β
12
D. CARDONA
estimate the Besov norm of Op(σ(·, ·)). First, we observe that
p
Lp2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
dξTr[ξ(x)F (Op(σ)f )(ξ)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
X2m≤hξi<2m+1
Op(σ)f (y)ξ(y)∗dy](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
:=ZM(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dξTr[ξ(x)ZM
X2m≤hξi<2m+1
Ayf (y)ξ(y)∗dy](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
=ZM(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dξTr[ξ(x)ZM
X2m≤hξi<2m+1
z∈M(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dξTr[ξ(x)F (Az f )(ξ)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ZM
X2m≤hξi<2m+1
sup
dx
p2
p2
dx
p2
dx
By the Sobolev embedding theorem, we have
From this, and the Sobolev embedding theorem we have
dξTr[ξ(x)F (Op(∂β
dξTr[ξ(x)F (Op(∂β
p2
sup
. sup
z∈M(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
X2m≤hξi<2m+1
. Xβ≤lZM(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
β≤lZM(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
z∈G(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ZM
X2m≤hξi<2m+1
. Xβ≤lZMZM(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
β≤lZMZM(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
β≤l,z∈MZM(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dξTr[ξ(x)F (Az f )(ξ)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
X2m≤hξi<2m+1
X2m≤hξi<2m+1
dξTr[ξ(x)Azf (y)ξ(y)∗](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
X2m≤hξi<2m+1
X2m≤hξi<2m+1
X2m≤hξi<2m+1
k X2m≤hξi<2m+1
= sup
≤ sup
. sup
β≤l,z∈M
sup
dξTr[ξ(x)F (Op(∂β
dξTr[ξ(x)F (Op(∂β
dξTr[ξ(x)F (Op(∂β
p2
p2
dz
dz
z σ(z, ·))f )(ξ)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
z σ(z, ·))f )(ξ)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
p
dy
p2
p2
z σ(z, ·))f )(ξ)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
z σ(z, ·))f )(ξ)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
z σ(z, ·))f )(ξ)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dzdx
dxdz
p2
dx
dξTr[ξ(x)F (Op(∂β
z σ(z, ·))f )(ξ)]kp2
Lp2
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
13
Hence,
dξTr[ξ(x)F (Op(σ)f )(ξ)]kLp2
dξTr[ξ(x)F (Op(∂β
z σ(z, ·))f )(ξ)]kLp2
We define for every z ∈ M the non-negative function z 7→ g(z) by
Thus, considering 0 < q < ∞ we obtain
kOp(σ)f kBr
k X2m≤hξi<2m+1
sup
2mrq
β≤l,z∈M
. sup
g(z) = sup
k X2m≤hξi<2m+1
2mrq(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
p2 ,q(M ) :=
∞Xm=0
X2m≤hξi<2m+1
β≤l,z∈M(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
.
∞Xm=0
β≤l(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
X2m≤hξi<2m+1
k→∞ kXm=0
g(z)! 1
k→∞ sup
∞Xm=0
β≤l,z∈M(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
p,q(M ) .
∞Xm=0
≤" sup
kOp(∂β
kOp(∂β
≤ sup
= lim
≤ lim
β≤l,z∈M
β≤l,z∈M
2mrq
sup
z∈M
q
We write,
∞Xm=0
2mrq sup
z∈M
Hence, we can write
kOp(σ)f kBr
dξTr[ξ(x)F (Op(σ)f )(ξ)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
X2m≤hξi<2m+1
dξTr[ξ(x)F (Op(∂β
q
1
q
Lp2
z σ(z, ·))f )(ξ)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
q
Lp2
1
q
dξTr[ξ(x)F (Op(∂β
2mrq sup
z∈M
q
g(z)! 1
2mrqg(z)! 1
q
q
q
.
Lp2
z∈M
= lim
2mrqg(z)! 1
z σ(z, ·))f )(ξ)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
k→∞ sup
kXm=0
z∈M ∞Xm=0
2mrqg(z)! 1
z σ(z, ·))f )(ξ)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
= sup
.
q
dξTr[ξ(x)F (Op(∂β
q
Lp2
1
q
X2m≤hξi<2m+1
z σ(z, ·))f kBr
p2 ,q
z σ(z, ·))kB(Br
p1 ,q,Br
p2 ,q)# kf kBr
p1 ,q .
So, we deduce the boundedness of A = Op(σ). Now, we treat of a similar way the
boundedness of Op(σ) if q = ∞. In fact, from the inequality
dξTr[ξ(x)F (Op(σ)f )(ξ)]kLp2
k X2m≤hξi<2m+1
. sup
β≤l,z∈M
k X2m≤hξi<2m+1
dξTr[ξ(x)F (Op(∂β
z σ(z, ·))f )(ξ)]kLp2
14
we have
2mrk X2m≤hξi<2m+1
D. CARDONA
dξTr[ξ(x)F (Op(σ)f )(ξ)]kLp2
dξTr[ξ(x)F (Op(∂β
z σ(z, ·))f )(ξ)]kLp2 .
. 2mr
So we get
β≤l,z∈M
sup
k X2m≤hξi<2m+1
p,∞(M ) ." sup
β≤l,z∈M
kOp(σ)f kBr
kOp(∂β
z σ(z, ·))kB(Br
p1 ,q ,Br
p2,q)# kf kBr
p1 ,∞.
With the last inequality we end the proof.
(cid:3)
Theorem 4.3. Let G be a Compact Lie group, n = dim(G) and let 0 ≤ ρ, δ ≤ 1.
Denote by κ the smallest even integer larger that n
p ] + 1.
Let A : C ∞(G) into C ∞(G) be a pseudo-differential operator with global symbol a(x, ξ)
satisfying
2 . Let 1 < p < ∞ and l = [ n
kDα
with m ≥ κ(1 − ρ) 1
Br
ξ ∂β
p − 1
x a(x, ξ)kop ≤ Cα,βhξi−m−ρα+δβ, α ≤ κ, β ≤ l,
(4.3)
2 + δl. Then A extends to a bounded operator from Br
p,q(G) into
p,q(G) for all 0 < q ≤ ∞ and r ∈ R. Moreover, if we assume that
x a(x, ξ)kop ≤ Cα,βhξi−ρα, α ≤ κ, β ≤ l,
ξ ∂β
kDα
(4.4)
then A extends to a bounded operator from B
1 < p < ∞, 0 < q ≤ ∞ and r ∈ R.
r+κ(1−ρ) 1
p,q
p − 1
2
(G) into Br
p,q(G) for all
Proof. If A = Op(a) is a Fourier multiplier, i.e, a(x, ξ) = a(ξ), by using Theorem 3.7
we have that A is bounded operator from Lp1 into Lp2 and consequently A extends to
a bounded operator from Br
p,q(G) for all 0 < q ≤ ∞ and r ∈ R. For the
general case where a(x, ξ) as in (4.3) depends on the spatial variable, we have, as in
the previous proof that
p,q(G) into Br
kOp(a)f kBr
p,q(G) ." sup
β≤l,z∈G
kOp(∂β
z a(z, ·))kB(Br
p,q ,Br
(4.5)
p,q)# kf kBr
p,q .
In fact, every multiplier Op(∂β
z a(z, ·)) is bounded on Br
p,q(G) because we only needs
kDα
ξ (∂β
z a(z, ξ))kop . hξi−m−ρα, α ≤ κ, β ≤ l.
(4.6)
For the proof of this necessary condition, we use the fact that m ≥ κ(1 − ρ) 1
In fact,
p − 1
2 + δl.
kDα
ξ (∂β
z a(z, ξ))kop . hξi−m−ρα+δβ . hξi−m−ρα+δl . hξi−κ(1−ρ) 1
p − 1
2 −ρα
which shows the boundedness of the multiplier (∂β
of operators (∂β
z a(z, ·))z∈G has norm uniformly bounded in z we have,
z a(z, ·)) on Br
p,q(G). Since the family
sup
β≤l,z∈M
kOp(∂β
z a(z, ·))kB(Br
p,q ,Br
p,q) < ∞.
So, we end the proof for this case. If the symbol a(ξ) satisfies
ξ a(x, ξ)kop ≤ Cαhξi−ρα, α ≤ κ,
kDα
(4.7)
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
15
then the corresponding operator Ta is bounded from H mp,p(G) into Lp(G), mp = κ(1 −
ρ) 1
2 , (Corollary 5.1 of [29]), so we have for 0 < q < ∞ the estimate
p − 1
kTaf kq
Br(G) =Xl≥0
.Xl≥0
=Xl≥0
2lrqkTa[(Op(χm)f )]kq
Lp(G)
2lrqk(Op(χm)f )kq
Hmp ,p(g)
2lrqkΛmp[(Op(χm)f )]kq
2lrqk(Op(χm)Λmpf )kq
Lp(M )
Lp(G) =Xl≥0
= kΛmpf kq
Br(G) . kf kq
Br+mp (G)
.
which proves the boundedness of Ta. Now, we extend the boundedness result for non-
invariant symbols a(x, ξ) as in (4.4) by using the inequality (4.5). The proof for q = ∞
is analogous.
(cid:3)
Theorem 4.4. Let G be a compact Lie group, n = dim(G), 0 ≤ ρ < 1 and 0 ≤
ν < n
2 (1 − ρ). Let A : C ∞(G) into C ∞(G) be a pseudo-differential operator with global
symbol a(x, ξ) satisfying
ξ ∂β
x a(x, ξ)kop ≤ Cα,βhξi−ν−ρα, α ∈ Nn, β ≤ l,
kDα
(4.8)
with 1 < p < ∞ and l = [ n
p − 1
into Br
p,q(G) for all 1
p ] + 1. Then A extends to a bounded operator from Br
2 ≤ ν
n (1 − ρ)−1, 0 < q ≤ ∞ and r ∈ R.
p,q(G)
Proof. Again, if a(·, ·) is independent of the spatial variable, the Fourier multiplier
A = Op(a) is bounded on Lp(G) as consequence of Theorem 3.8. Newly, by theorem
2.1 we obtain that the Fourier multiplier A is bounded on Br
p,q(G). We know that for
l = [ n
p ] + 1
kOp(a)f kBr
p,q(G) ." sup
β≤l,z∈M
kOp(∂β
z a(z, ·))kB(Br
p,q ,Br
p,q)# kf kBr
p,∞
provide that every multiplier Op(∂β
from the fact that
z a(z, ·)) is bounded on Br
p,q(G). But, this it follows
kDα
ξ ∂β
x a(x, ξ)kop ≤ Cα,βhξi−ν−ρα, α ∈ Nn, β ≤ l.
(4.9)
(cid:3)
4.1. Examples. Now, we consider examples of differential problems which could not
treated with the classical pseudo-differential calculus (based in the notion of local sym-
bols). We follow [29].
Example 4.5. Let us consider G = SU(2), and let {X, Y, Z} be a basis of its Lie
algebra g = su(2). Let us consider the differential operators
• Lsub = X 2 + Y 2 (sub-Laplacian) and
• H := X 2 + Y 2 − Z,
which are hypoelliptic operators by Hormander's sum of squares theorem. A parametrix
P of Lsub has matrix valued symbol σ ∈ S−1
(SU(2)). On every coordinate chart U ⊂ R3
2 ,0
of SU(2), Lsub has a local symbol in the class S−1
(U × R3). The classical Hormander
2 , 1
classes on compact manifolds M require the condition ρ ≥ 1 − δ and ρ > δ (which
1
1
2
16
D. CARDONA
implies that ρ > 1
2 ), hence such calculus cannot be used for the analysis of the sub-
Laplacian. The global description of the Hormander classes trough Ruzhansky-Turunen
calculus gives together with Theorem 2.4 that P is a bounded operator on Br
p,q(SU(2)),
r ∈ R, 0 < q ≤ ∞, and 1 < p < ∞. Hence, if we consider the problem
and we assume that the problem has almost one solution u ∈ Bs
p,q(SU(2)), we have
Lsubu = f, f ∈ Br
p,q(SU(2))
(4.10)
kukBr
p,q (G) . kf k
B
r−1+ κ
p,q
2 1
p − 1
2
+ kukBs
p,q (G).
(G)
(4.11)
On the other hand, since the operator H := X 2 + Y 2 − Z has parametrix with symbol
in S−1
2 ,0
(SU(2)), we have the estimate
1
kukBr
p,q(G) . kHuk
B
r−1+ κ
p,q
2 1
p − 1
2
+ kukBs
p,q(G),
(G)
as in the sub-Laplacian case for the following differential problem
Hu = f, f ∈ Br
p,q(SU(2)).
(4.12)
(4.13)
We end this section with the following example on vector fields on arbitrary compact
Lie groups.
Example 4.6. Let X be a real left invariant vector field on a compact Lie group G.
There exists an exceptional discrete set C ⊂ iR, such that X + c is globally hypoelliptic
for all c /∈ C. We recall that an differential operator A : D ′(G) → D ′(G) is globally
hypoelliptic, if u ∈ D ′(G), Au = f, and f ∈ C ∞(G) implies u ∈ C ∞(G). If G = SU(2),
X + c is globally hypoelliptic if and only if c /∈ C = 1
2 iZ. Moreover, on a compact Lie
group G, the inverse P = (X +c)−1 of X +c has global symbol in S0
0,0(G). As in the sub-
Laplacian case, the classical pseudo-differential calculus cannot be used for the analysis
of X + c. However, if we use Theorem 4.4, P is a bounded operator from B
into Br
(G)
2,q(G) for all r ∈ R and 0 < q ≤ ∞. Hence, we obtain the (sub-elliptic) estimate
(4.14)
.
kukBr
2,q (G) . k(X + c)uk
B
p − 1
2
r+κ 1
2,q
p − 1
2
r+κ 1
2,q
(G)
Acknowledgments: The author is indebted with Alexander Cardona for helpful
comments on an earlier draft of this paper. This project was supported by Faculty
of Sciences of Universidad de los Andes, Proyecto: Una clase de operadores pseudo-
diferenciales en espacios de Besov. 2016-1, Periodo intersemestral.
References
1. Akylzhanov, R. Nursultanov, E. Ruzhansky, M. Hardy-Littlewood-Paley type inequalities on com-
pact Lie groups, Math. Notes, 100 (2016), 287-290.
2. R. Akylzhanov, M. Ruzhansky, Fourier multipliers and group von Neumann algebras. C. R. Math.
Acad. Sci. Paris, 354 (2016), 766 -- 770.
3. R. Akylzhanov, E. Nursultanov, M. Ruzhansky, Hardy-Littlewood inequalities and Fourier multi-
pliers on SU(2), Studia Math., 234 (2016), 1-29.
4. Bourdaud, G.. Lp-estimates for certain non-regular pseudo-differential operators, Comm. Partial
Differential Equations, 7, 1023 -- 1033, (1982)
5. Cardona, D. Besov continuity for Multipliers defined on compact Lie groups. Palest. J. Math. Vol.
5(2) 35 -- 44 (2016)
6. Cardona, D. Besov continuity of pseudo-differential operators on compact Lie groups revisited. C.
R. Math. Acad. Sci. Paris Vol. 355, Issue 5, May 2017, Pag. 533 -- 537
7. Cardona, D. Nuclear pseudo-differential operators in Besov spaces on compact Lie groups, to appear
in J. Fourier Anal. Appl. 2017.
BESOV CONTINUITY FOR PSEUDO-DIFFERENTIAL OPERATORS
17
8. Cardona, D. Ruzhansky M. Multipliers for Besov spaces on graded Lie groups. C. R. Math. Acad.
Sci. Paris. Vol. 355(4), 400 -- 405, (2017)
9. Cohen, A. Numerical analysis of wavelet methods, Elsevier. 2003.
10. Delgado, J. Ruzhansky, M.: Lp-nuclearity, traces, and Grothendieck-Lidskii formula on compact
Lie groups., J. Math. Pures Appl. (9), 102(1), 153-172 (2014)
11. Dasgupta, A. Ruzhansky, M. The Gohberg Lemma, compactness, and essential spectrum of oper-
ators on compact Lie groups. J. Anal. Math. 128, 179-190, (2016)
12. Delgado, J. Ruzhansky, M.: Schatten classes on compact manifolds: Kernel conditions. J. Funct.
Anal., 267(3), 772-798 (2014)
13. Delgado, J. Ruzhansky, M.: Kernel and symbol criteria for Schatten classes and r-nuclearity on
compact manifolds., C. R. Acad. Sci. Paris. Ser. I. 352. 779-784 (2014)
14. Delgado, J. Ruzhansky, M. Lp-bounds for pseudo-differential operators on compact Lie groups. J.
Inst. Math. Jussieu, to appear, arXiv:1605.07027. DOI: /10.1017/S1474748017000123.
15. Ebert, S. Wirth, J. Diffusive wavelets on groups and homogeneous spaces, Proc. Roy. Soc. Edin-
burgh Sect. A 141 , no. 3, 497520. (2011)
16. Feichtinger, H. Fuhr, H. Pesenson,
I. Geometric Space-Frequency Analysis on Manifolds,
arXiv:1512.08668.
17. Fischer, V. Intrinsic pseudo-differential calculi on any compact Lie group. J. Funct. Anal. 268 , no.
11, 34043477. (2015)
18. Gibbons, G. Operateurs pseudo-differentiels et espaces de Besov, C. R. Acad. Sci. Paris, 286 (1978),
Serie A, 895 -- 897.
19. Hairer, M. Labb´e, C. The reconstruction theorem in Besov spaces. arXiv:1609.04543.
20. Hormander, L. Estimates for translation invariant operators in Lp spaces, Acta Math. 104, 93 -- 140
(1960).
21. E. Nursultanov, M. Ruzhansky, and S. Tikhonov, Nikolskii inequality and Besov, Triebel-Lizorkin,
Wiener and Beurling spaces on compact homogeneous manifolds. Ann. Sc. Norm. Super. Pisa Cl.
Sci., Vol. XVI, 981 -- 1017. (2016)
22. Nursultanov E., Ruzhansky M., Tikhonov S., Nikolskii inequality and functional classes on compact
Lie groups, Funct. Anal. Appl., 49, 226 -- 229. (2015)
23. Park, B. On the boundedness of Pseudo-differential operators on Triebel-Lizorkin and Besov spaces,
arXiv:1602.08811.
24. Ruzhansky, M., Turunen, V.:Pseudo-differential Operators and Symmetries: Background Analysis
and Advanced Topics Birkhauser-Verlag, Basel, (2010)
25. Ruzhansky, M., Turunen, V. Sharp Garding inequality on compact Lie groups. J. Funct. Anal., 260
2881-2901. (2011)
26. Ruzhansky M., Turunen V., Global quantization of pseudo-differential operators on compact Lie
groups, SU(2) and 3-sphere, Int. Math. Res. Not. IMRN 2013, no. 11, 2439-2496.
27. Ruzhansky M., Turunen V., Wirth J.: Hormander class of pseudo-differential operators on compact
Lie groups and global hypoellipticity, J. Fourier Anal. Appl., 20 476-499. (2014)
28. Ruzhansky, M. Wirth, J. Global functional calculus for operators on compact Lie groups. J. Funct.
Anal. 267 no. 1, 144 -- 172. (2014)
29. M. Ruzhansky, and J. Wirth, Lp Fourier multipliers on compact Lie groups Math. Z. 621 -- 642
(2015).
Duv´an Cardona:
Department of Mathematics
Pontificia Universidad Javeriana
Bogot´a
Colombia
E-mail address [email protected]; [email protected]
|
1609.07777 | 2 | 1609 | 2017-01-23T00:53:29 | Polydisc version of Arveson's conjecture | [
"math.FA",
"math.OA"
] | In the present paper, we solve the polydisc-version of Arveson Conjecture by giving a complete criteria for essential normality of homogeneous quotient modules of the Hardy module over the polydisc, and it turns out that our method applies to quotient modules of the weighted Bergman modules $A^2_s(\mathbb D^d)$. | math.FA | math |
Polydisc version of Arveson's conjecture
Penghui Wang∗ Chong Zhao †
School of Mathematics, Shandong University
Jinan, Shandong 250100, The People's Republic of China
March 23, 2018
Abstract
In the present paper, we solve the polydisc-version of Arveson Conjecture by giving a com-
plete criterion for essential normality of homogeneous quotient modules of the Hardy module
over the polydisc, and it turns out that our method applies to quotient modules of the weighted
Bergman modules A2
s(Dd).
2000 MSC. 47A13, 46H25
Key Words. Essential normality, Hardy space over the polydisc, quotient module
1
Introduction
Throughout this paper, D denotes the open unit disc of the complex plane. The Hardy space
H 2(Dd) over the polydisc is defined as the Hilbert space of analytic functions over Dd satisfying
f2 = sup
0<r<1ZTd f (rz)2dm(z) < ∞,
where dm is the normalized Haar measure on Td. H 2(Dd) can be viewed as a Hilbert module [5, 10]
in the sense that H 2(Dd) as a Hilbert space admits a natural module structure over the polynomial
ring C[z1, . . . , zd] with respect to the obvious multiplication action.
A closed subspace M of H 2(Dd) that is invariant under multiplication by polynomials is called
a submodule, and N = H 2(Dd) ⊖ M which is invariant under the adjoint module actions is called
a quotient module, whose module structure is given by
f · g = Sf g,
f ∈ C[z1, . . . , zd], g ∈ N ,
where Sf = PN Mf N . The closure of an ideal I ⊂ C[z1, . . . , zd] is a submodule, which is called
the submodule generated by I and denoted by [I]. For convenience we denote by [I]⊥ the quotient
∗Partially supported by NSFC(No. 11471189), E-mail: [email protected]
†Partially supported by NSFC(No. 11501329) and Shandong Province Natural Science Foundation ZR2014AQ009,
E-mail:
[email protected].
1
module associate to the ideal I. If all the commutators [S∗zi, Szj ], 1 ≤ i, j ≤ d belong to the compact
operator algebra K, then N is said to be essentially normal.
The present paper is a continuation of [25], aiming at giving a complete answer to the polydisc
version of Arveson's conjecture. In [2], Arveson conjectured that homogenous submodules of the
d-shift module over the unit ball are essentially normal, and many efforts have been made along
this line, such as [2, 3, 8, 9, 12, 13, 14, 15, 16, 17, 22] and references therein.
The essential normality of quotient modules of canonical analytic function Hilbert modules over
the polydiscs was initiated in [11], where some special modules such as [(z − w)2]⊥, [zi − wj]⊥ in
H 2(D2) were considered. Clark[6] proved that [B1(z1) − B2(z2),··· , Bd−1(zd−1) − Bd(zd)]⊥ can
be identified with a kind of Bergman space over the associated varieties, where Bi(zi) are finite
Blaschke products, and hence they are essentially normal. For recent development on essential
normality over polydiscs, refer to [18, 19, 20, 21, 23, 24] and references therein.
Compared to the existing results on Arveson's conjecture over the unit ball, the situations over
the polydisc are totally different. Briefly speaking, over the unit ball all the submodules generated
by polynomials and their associated quotient modules are believed to be essentially normal, while
over polydiscs no non-trivial submodule and few quotient modules are essentially normal.
For an ideal I ⊂ C[z1,··· , zd], denote by Z(I) the zero variety of I, and ZDd(I) = Z(I) ∩ Dd.
If dim CZ(I) = 0, the quotient module [I]⊥ is of finite dimension, on which the essential normality
is trivial. Therefore in what follows, we always assume dim CZ(I) ≥ 1. It is mentioned in [25] and
will be proved in Section 2 that, if the homogenous quotient module [I]⊥ is essentially normal, then
dim CZ(I) = 1. Therefore by homogeneity, ZDd(I) = Si Vi for several different discs Vi. Taking
ui ∈ ∂Vi and we can write Vi = {λui : λ ∈ D}. To continue, we set
Λi = {ij : ui,ij = 1} ⊂ {1, 2,··· , d}.
Obviously, Λi depends only on Vi. In what follows, for u ∈ ∂Dd and Vu = {λu : λ ∈ D}, we will
write Λ = {ij : uij = 1} if there is no ambiguity, and uΛ = (ui1,··· , uik ). Next, we introduce the
following condition on the variety of I.
n
Condition A. Let I be a homogenous ideal, ZDd(I) =
say that I(or the zero variety ZDd(I)) satisfies Condition A, if one of the following items holds
Vi, where {Vi} are different discs. We
Si=1
1) Λi 6= Λj for i 6= j;
2) For any pair (i, j) such that Λi = Λj, let Λi = {i1,··· , ik}, then (ui,i1,··· , ui,ik ) and
(uj,i1,··· , uj,ik ) are linearly independent.
To state our main result, we need the following
Definition 1.1. Let u ∈ ∂Dd, and I be a homogenous ideal such that ZDd(I) = Vu. Denote by Ju
the ideal of C[z1, . . . , zd] generated by {¯uizi − ¯ujzj : i, j ∈ Λ}, and I′ the ideal generated by I and
Ju.
1) If I′ = √I the prime ideal associated to Vu, then I is called quasi-prime.
2) If √I/I′ is of finite dimension, then I is called essentially quasi-prime.
The quotient module [I]⊥ is called (essentially) quasi-prime provided I is (essentially) quasi-prime.
2
We can state our main result, which is a complete criterion for essential normality of homoge-
neous quotient modules of H 2(Dd).
Theorem 1.2. Let I ⊂ C[z1, . . . , zd] be a homogeneous ideal for which [I]⊥ is of infinite dimension.
Then the quotient module [I]⊥ of H 2(Dd) is essentially normal if and only if the following items
hold,
1) ZDd(I) satisfies Condition A, hence ZDd(I) = Si
2) Let I = Tm
tially quasi-prime.
j=1 Iuj be the primary decomposition with ZDd(Iuj ) = Vuj , then each Iuj is essen-
Vui for several different discs Vui,
The criterion in Theorem 1.2 is purely algebraic, which does not depend on the structure of
Hardy module H 2(Dd). Notice that in the case d = 2, Theorem 1.2 reduces to [19, Theorem 1.1]. In
Section 4, we will give two surprising examples, which show that the essential normality in higher
dimensional case is more interesting than the 2-dimensional case.
Remark 1.3.
1) Theorem 1.2 suggests that the essential normality of quotient modules is closely
related to the distinguished variety[1].
2) The method in [19], where d=2 is assumed, relies heavily on the fact that Mzi is isometric,
and can not be applied to the weighted Bergman space.
In [25, Remark 2.15], we can only deal with the essential normality of quotient modules of the
s(D)
s(Dd) in some simplest cases, where A2
s(Dd) = A2
s(D) ⊗ ··· ⊗ A2
weighted Bergman module A2
and A2
s(D) is the space of analytic functions f on D such that
kfk2
s = ZD f (z)2(1 − z2)sdA(z) < ∞,
where s > 0 and dA is the normalized area measure on D. It can be verified that all the proofs in
the present paper are valid to the weighted Bergman module case. We have
Theorem 1.4. For a homogenous ideal I of C[z1, . . . , zd], the quotient module [I]⊥ of A2
essentially normal if and only if 1)-2) in Theorem 1.2 are satisfied.
s(Dd) is
It is worth to point out that Theorem 1.4 is the first nontrivial result on essential normality of
quotient modules of A2
s(Dd).
The present paper is organized as following. in section 2 we give a geometric characterization
of Z(I) for essential normality of [I]⊥. In Section 3 we construct a large class of essentially normal
quotient modules of H 2(Dd). In Section 4 the complete criterion for essentially normal homogeneous
quotient modules is obtained, and Section 5 contains some discussion on the K-homology.
3
2 The variety of an essentially normal quotient module
We begin the present section with some preliminaries. Recall that H 2(Dd) is the Hilbert space with
reproducing kernel
d
Kλ(z) =
(1 − ¯λizi)−1, ∀z ∈ Dd
Yi=1
at λ ∈ Dd. Obviously Kλ2 = Qd
For each A ∈ B(H 2(Dd)) the function
i=1(1 − λi2)−1 diverges to infinity as λ approaches ∂Dd.
A(z) = hAkz, kzi, z ∈ Dd
is called the Berezin transform of A, where kz = Kz−1Kz is the normalized reproducing kernel
at z. It is routine to compute for z, λ ∈ Dd that
hkz, Kλi = kz(λ) =
1
Kz
d
Yi=1
1
1 − ¯λizi
,
which converges to 0 as z approaches ∂Dd. Since linear combinations of {Kλ : λ ∈ Dd} are dense
in H 2(Dd), kz converges to 0 in the weak topology as z approaches ∂Dd. If A is compact then as
z approaches ∂Dd, Akz converges to 0 in norm and then A(z) converges to 0. Then we have the
following well-known lemma.
Lemma 2.1. It holds for each compact operator A ∈ B(H 2(Dd)) that limz→∂Dd A(z) = 0.
By an application of Berezin transform, we get the following result.
Lemma 2.2. Suppose I ⊂ C[z1, . . . , zd] is an ideal. Suppose u, v ∈ ∂Dd are two different accumu-
lation points of ZDd(I), and there is a subset Λ ⊂ {1, . . . , d} such that
ui = vi ∈ T,∀i ∈ Λ,
and ui, vi ∈ D whenever i ∈ Λc, then [I]⊥ is not essentially normal.
Proof. Since u 6= v, there is a polynomial h ∈ C[z1, . . . , zd] such that h(u) = 0 and h(v) 6= 0. Let
{u(n)},{v(n)} be sequences in ZDd(I) with limits u, v respectively. As n → ∞, both ku(n) and kv(n)
converge to 0 in the weak topology. On the other hand,
lim
n→∞h[S∗h, Sh]ku(n), ku(n)i
n→∞hShku(n), Shku(n)i − lim
n→∞hShku(n), kv(n)i2 − lim
n→∞h(v(n))hku(n) , kv(n)i2 − h(u)2
n→∞hS∗hku(n), S∗hku(n)i
n→∞h(u(n))2
n→∞hku(n), kv(n)i2
p(1 − ui2)(1 − vi2)
1 − ¯uivi
= lim
≥ lim
= lim
= h(v)2 lim
= h(v)2 Yi∈Λc
> 0.
Therefore {[S∗h, Sh]ku(n)} does not converge to 0 as n → ∞, and [S∗h, Sh] cannot be compact by
Lemma 2.1.
4
Example 2.3. In the case d = 3, let I ⊆ C[z1, z2] be a homogenous ideal such that Z(I)∩∂D2 ⊂ T2,
and p(z) = (α1z1 − z3)(α2z2 − z3) with α1 6= α2, αi < 1. Then the quotient module [I, p]⊥ of
H 2(D3) cannot be essentially normal according to Lemma 2.2.
The following lemma and its idea of proof originate from [25], and we perform the improved
proof here. It characterizes the zero varieties of homogeneous ideals I for witch [I]⊥ is essentially
normal.
Lemma 2.4. Let I ⊂ C[z1, . . . , zd] be a homogeneous ideal such that dim CZ(I) ≥ 2, then N = [I]⊥
is not essentially normal.
Proof. Suppose dim CZ(I) = k ≥ 2. Let V be any irreducible component of Z(I) of complex
dimension k. For i = 1, . . . , d, denote by
Ei = {z ∈ V : zi = max
1≤j≤dzj}.
Clearly each Ei is closed as a subset of V , and V = S1≤i≤d Ei. Then by the Baire Category
Theorem, some Ei is of the second category as a subset of V , and therefore we can find an open
subset Ω1 ⊂ Ei. Then since Sing(V ), the set of singular points of V , is of the first category, there
must be a nonsingular point u ∈ Ω1 ⊂ Ei. By homogeneity we may assume u ∈ ∂Dd. Let g1, . . . , gm
be a set of generators for the prime ideal of V , then the matrix (cid:16) ∂gj
is of rank d− k. By the
Implicit Function Theorem, there is an open ball B ⊂ Ck, an open neighbourhood Ω2 ⊂ Ω1 of u,
and an analytic bijection ϕ that maps B onto Ω2.
Define the maps ψj : Ω2 → Cd, z 7→ zj/zi, 1 ≤ j ≤ d, and set ψ = (ψ1, . . . , ψd). Since Ω2 ⊂ Ei,
ψj maps Ω2 into ¯D, and ψ maps Ω2 into Z(I) ∩ ∂Dd. Then ψj ◦ ϕ maps B analytically into ¯D.
Therefore ψj◦ϕ maps B either onto an open subset of D or to a constant cj in ¯D. If ψ is constant on
Ω2, then Ω2 ⊂ {z ∈ Cd : zj = cjzi, 1 ≤ j ≤ d}, which cannot be of dimension k ≥ 2. Therefore some
ψj is nonconstant on Ω2, and hence there exists v ∈ Ω2 such that ψj(v), ψj (u) are two different
points in D. Clearly for 1 ≤ k ≤ d, ψk(u) and ψk(v) should either both be in D, or be of the
same value in T. Then since ψ(u), ψ(v) ∈ Z(I)∩ ∂Dd, [I]⊥ cannot be essentially normal by Lemma
2.2.
(u)(cid:17)i,j
∂zi
Remark 2.5. Lemma 2.4 shows that, in the polydisc case, if the homogenous quotient module is
essentially normal, then the dimension of zero variety is less than or equal to 1. We speculate that
in Lemma 2.4 the requirement of homogeneity could be dropped.
Combine Lemma 2.2 and Lemma 2.4 and we immediately obtain the following result.
Proposition 2.6. Let I ⊂ C[z1, . . . , zd] be a homogeneous ideal for which [I]⊥ is essentially normal
and of infinite dimension, then the zero variety of I satisfies Condition A.
3 Construction of essentially normal quotient modules
In this section, we establish a strategy for constructing essentially normal quotient modules, be-
ginning with those associated to essentially quasi-prime ideals. Suppose u ∈ ∂Dd, and Iu is a
homogenous ideal with ZDd(Iu) = Vu. By [25, Theorem 2.17], we know that [Iu]⊥ is essentially
5
normal provided u ∈ Td. To study essential normality of non-distinguished homogenous quotient
modules, we have to investigate more on [Iu]⊥.
To continue, we need some terminologies. Let H1, H2 be Hilbert spaces. A ∈ B(H1, H2) is
called essentially bounded from below if there is a constant c > 0 and compact operator K such
that A∗A + K ≥ cIH1. Similarly A ∈ B(H1, H2) is called an essential isometry if IH1 − A∗A is
compact. A is said essentially unitary if both IH2 − AA∗ and IH1 − A∗A are compact. According
to [4], every essentially unitary operator A ∈ B(H) can be decomposed as A = S + K where K is
compact, and S is either a unitary operator, or a shift of multiplicity n, or the adjoint of a shift of
multiplicity n, according to its Fredholm index being 0, n,−n.
The following lemma can help us to prove the Fredholmness of some essential isometries.
Lemma 3.1. If A ∈ B(H1, H2) is an essential isometry, then the range of A is closed.
Proof. Since IH1 − A∗A is compact, A∗A is Fredholm and therefore has closed range. Then by
the Inverse Mapping Theorem A∗A is invertible on (ker A)⊥, and therefore bounded from below on
(ker A)⊥. Consequently A is bounded from below on (ker A)⊥ and has closed range.
We begin our construction of essentially normal quotient modules with a careful consideration
on the result of [25]. Recall that the radical of an ideal I ⊂ C[z1, . . . , zd] is the ideal
√I = {f ∈ C[z1, . . . , zd] : ∃n ∈ N, f n ∈ I}.
Lemma 3.2. Let u ∈ Td, and I ⊂ C[z1, . . . , zd] be an homogeneous ideal satisfying ZDd(I) = Vu.
Then N = [I]⊥ is essentially normal, and there is an essentially unitary operator S ∈ B(N )
and compact operators Ki such that Szi = uiS + Ki for 1 ≤ i ≤ d. Consequently for h(z) =
d−1Pd
i=1 ¯uizi, the operator S∗h is essentially unitary on N .
Proof. Recall that Ju ⊂ C[z1, . . . , zd] is the ideal generated by {ujzi − uizj : 1 ≤ i, j ≤ d}. Since
u ∈ Td, Ju = √I. Since C[z1, . . . , zd] is a Neotherian ring, J n
u ⊂ I for some positive integer n. By
u ]⊥ and therefore S∗z1 is essentially isometric on
[25, Remark 2.8], M∗z1 is essentially isometric on [J n
N . Then by [25, Theorem 2.14],
PN − S∗z1Sz1 = PN − Sz1S∗z1 + [Sz1, S∗z1]
is compact, and Sz1 is essentially unitary. [25, Remark 2.8] also shows the compactness of (uiM∗zi −
u ]⊥ for 1 ≤ i ≤ d, which implies the compactness of uiS∗zi − u1S∗z1. Set S = ¯u1Sz1 then
u1M∗z1) [J n
Ki = Szi − ui ¯u1Sz1 is compact, and Szi = uiS + Ki. Moreover,
S∗h = d−1
d
Xi=1
uiS∗zi = S∗ + d−1
d
Xi=1
K∗i ,
is essentially unitary.
The following lemma reveals the structure of quotient modules associated to primary ideals
with variety Vu. Recall that for u ∈ ∂Dd\Td, Λ = {i : ui ∈ T, 1 ≤ i ≤ d}. For abbreviation, denote
by uΛ = (ui1,··· , uik ) and DuΛ = {λuΛ, λ ∈ D}. Moreover, the polynomial ring on the variable zΛ
is denoted by C[zΛ], and the Hardy module on zΛ is denoted by H 2(DΛ).
6
Lemma 3.3. Let u ∈ ∂Dd\Td, and I ⊂ C[z1, . . . , zd] be a homogeneous ideal with variety Vu.
Suppose k = Λ, and set I0 = I ∩ C[zΛ]. Then I0 is a homogeneous ideal of C[zΛ] with variety
ZDk (I0) = DuΛ and
[I]⊥ ⊂ nf ∈ H 2(Dd) : f (z) = Xα∈Zd−k
+
zα
Λc fα(zΛ), fα ∈ N0o,
(3.1)
where N0 = [I0]⊥ is the quotient module of H 2(DΛ) associated to I0. Moreover if we denote by
h(z) = k−1Xi∈Λ
¯uizi,
then S∗h is essentially unitary and for i, j ∈ Λ, uiS∗zi − ujS∗zj is compact.
Proof. Clearly uΛ ∈ Z(I0), which implies DuΛ ⊂ ZDk (I0). Conversely suppose v ∈ ZDk (I0), then
choose v ∈ Dd such that vΛ = v and vj = ujh(v),∀j ∈ Λc. Denote by J ⊂ C[z1, . . . , zd] the ideal
generated by {zj − ujh : j ∈ Λc}, then by the Hilbert Nullstellensatz there is an integer N > 0 such
that J N ⊂ I. By a division algorithm, each g ∈ I can be decomposed as
g = Xj∈Λc
(zj − ujh)gj + g1, g1 ∈ C[zΛ], gj ∈ C[z1, . . . , zd].
Set g2 = − Pj∈Λc
(zj − ujh)gj ∈ J, then we have
N−1
Xk=0
This equality together with gN
2 ∈ I implies gN
1 ∈ I0. It follows from g1(v)N = 0 and g2(v) = 0
that g(v) = 0, which ensures v ∈ ZDd(I) = Du and consequently v ∈ DuΛ. Hence we have proved
ZDk (I0) = DuΛ.
1 − gN
gN
1 gN−1−k
gk
2 = g
∈ I.
2
Notice that each function f ∈ H 2(Dd) has an expansion
f (z) = Xα∈Zd−k
+
Λc fα(zΛ), fα ∈ H 2(DΛ).
zα
For f ∈ [I]⊥, since M∗g f = 0 for each g ∈ I0 we have fα ∈ N0,∀α ∈ Zd−k
j (z) = (zj − ujh(z))N ∈ I we have for n ≥ N that
For j ∈ Λc, since hN
+ . Hence (3.1) is proved.
zj = [¯ujS∗h + (Szj − ¯ujS∗h)]n
S∗n
=
=
Then for n ≥ 2N it holds
n
Xl=0
N−1
Xl=0
(cid:18)n
l(cid:19)¯un−l
(cid:18)n
l(cid:19)¯un−l
j S∗n−l
h
(Szj − ¯ujS∗h)l
j S∗n−l
h
(Szj − ¯ujS∗h)l
Sn
zj ≤
N−1
Xl=0
(cid:18)n
l(cid:19)¯ujn−l(1 + uj)l
≤ N(cid:18) n
N − 1(cid:19)¯ujn−N +1(1 + uj)N−1.
7
(3.2)
For positive integers n, denote by Qn the projection from [I]⊥ to the subspace of H 2(Dd)
spanned by {zα
Λc g : α ∈ Zd−k
+ , maxi αi < n, g ∈ N0}. Then by (3.2) it holds for f ∈ [I]⊥ that
f − Qnf2 = Xmaxj αj >n
Λcfα2
zα
d−k
Xj=1
Λc fα2
zα
Xαj >n
≤
= Xj∈Λc S∗n
zj f2
N − 1(cid:19)2
≤ f2N 2(cid:18) n
Xj∈Λc ¯uj2n−2N +2,
which converges uniformly for f ≤ 1 as n → ∞. Then we conclude limn→∞ Qn = P[I]⊥. By
Lemma 3.2 M∗h N0 is essentially isometric, and (uiM∗zi − ujM∗zj ) N0 is compact for i, j ∈ Λ. Then
for any integer n > 0 and sequence {gm} of unit vectors in [I]⊥ that converges weakly to 0, we
obtain by Lemma 3.2 that for i ∈ Λ
m→∞S∗hgm ≥ lim
m→∞M∗h Qngm = lim
m→∞QnM∗hgm ≥ lim
m→∞Qngm,
lim
which implies limm→∞ S∗hgm = 1, proving that S∗zi is essentially isometric. Similarly we have
lim
m→∞(uiM∗zi − ujM∗zj )gm = lim
n→∞
which induces the compactness of uiS∗zi − ujS∗zj .
dimension. For homogeneous f ∈ coker S∗h, it holds for each g ∈ [I]⊥ that
m→∞Qn(uiM∗zi − ujM∗zj )gm = 0,
Since S∗h is essentially isometric, PN − ShS∗h is compact and therefore ker S∗h must be of finite
lim
which induces hf ∈ I. Then since I is primary and h /∈ √I, f belongs to I. Therefore
hhf, gi = hf, S∗hgi = 0,
ker Sh = coker S∗h = {0},
and S∗h is Fredholm by Lemma 3.1. Then Sh is the inverse of S∗h in the Calkin algebra on N , which
implies the compactness of PN − S∗hSh, and ensures that Sh is essentially unitary.
The following proposition ensures the essential normality of essentially quasi-prime quotient
modules.
Proposition 3.4. Let u ∈ ∂Dd, and I ⊂ C[z1, . . . , zd] be an essentially quasi-prime homogeneous
ideal with variety Vu. Then N = [I]⊥ is essentially normal. Moreover there is an essentially unitary
operator S ∈ B(N ) and compact operators Ki such that
Szi = uiS + Ki,∀1 ≤ i ≤ d.
(3.3)
8
Proof. If u ∈ Td, then the conclusion is contained in Lemma 3.2. So we assume u ∈ ∂Dd\Td in the
remaining of the proof. By Lemma 3.3 S∗h is essentially unitary, and if i, j ∈ Λ then ¯uiSzi − ¯ujSzj
is compact. It follows that
[S∗h, Sh] = (S∗hSh − PN ) − (ShS∗h − PN ) ∈ K.
This equality together with the compactness of uiS∗zi − ujS∗zj gives
[S∗zj , Szi] ∈ K,∀i, j ∈ Λ.
Set S = Sh and Ki = Szi − uiSh for i ∈ Λ, then Ki is compact and Szi = uiS + Ki. For j ∈ Λc,
clearly hj ∈ √I and there hjhn ∈ √I for every natural number n. Since dim√I/I′ < +∞, there is
an integer N > 0 such that f ∈ I′ provided that f ∈ √I is homogeneous and deg f ≥ N . Therefore
hjhN ∈ I′, and there exists fj ∈ Ju such that hjhN −fj ∈ I. As a consequence S∗hj
is compact by Lemma 3.3. Then since S∗h is essentially unitary S∗hj
the proof of (3.3). As a consequence, [I]⊥ is essentially normal.
, which
must be compact, completing
S∗N
h = S∗fj
Remark 3.5. In either Lemma 3.2 or Proposition 3.4, Sh is essentially unitary, and therefore
Fredholm. Clearly 1 ∈ ker S∗h and dim ker S∗h > 0. Take an arbitrary homogeneous g ∈ ker Sh, then
hg ∈ I. But since h(u) 6= 0, h cannot belong to √I. Then g ∈ I and hence ker Sh = {0}. Therefore
the Fredholm index of Sh is nonzero.
Corollary 3.6. Let N be as in Proposition 3.4. Then P ⊥
Proof. By Proposition 3.4, Szi is essentially unitary for i ∈ Λ, and then PN − S∗ziSzi is compact.
Therefore since
N MziPN is compact for i ∈ Λ.
0 ≤ PN M∗ziP ⊥
N MziPN = PN M∗ziMziPN − S∗ziSzi ≤ PN − S∗ziSzi,
PN M∗ziP ⊥
N MziPN is compact.
To prove the essential normality of sums of essentially quasi-prime quotient modules, we need
the concept of asymptotical orthogonality. As in [19], if subspaces N1, N2 of the Hilbert space H
satisfy that PN1PN2 is compact, then N1 and N2 are said to be asymptotically orthogonal to each
other. As usual, let π : B(H) → B(H)/K(H) be the quotient map.
Lemma 3.7. Let u, v be two different points in ∂Dd, and 1 ≤ i, j ≤ d satisfy ui, vi, uj, vj ∈ T and
¯uivi 6= ¯ujvj. Let Iu, Iv ⊂ C[z1, . . . , zd] be essentially quasi-prime homogeneous ideals with variety
Vu, Vv respectively, then [Iu]⊥ is asymptotically orthogonal to [Iv]⊥.
Proof. By Corollary 3.6, both P[Iu]⊥M∗ziP[Iu] and P[Iv]MziP[Iv]⊥ are compact. Then by Proposition
3.4 we obtain
π(P[Iu]⊥P[Iv]⊥) = π(P[Iu]⊥P[Iv]⊥M∗ziP[Iv]⊥MziP[Iv]⊥)
= π(P[Iu]⊥M∗ziP[Iv]⊥MziP[Iv]⊥)
= π(P[Iu]⊥M∗ziP[Iu]⊥P[Iv]⊥MziP[Iv]⊥)
= π(¯uiviS∗uSv),
(3.4)
and similarly π(P[Iu]⊥P[Iv]⊥) = π(¯ujvjS∗uSv). Hence (¯uivi − ¯ujvj)π(S∗uSv) = 0, inducing that S∗uSv
is compact. Then the conclusion of the lemma follows from (3.4).
9
Lemma 3.8. Let u, v be two different points in ∂Dd, and 1 ≤ i ≤ d satisfy ui ∈ T and vi ∈ D. Let
Iu, Iv ⊂ C[z1, . . . , zd] be essentially quasi-prime homogeneous ideals with variety Vu, Vv respectively,
then [Iu]⊥ is asymptotically orthogonal to [Iv]⊥.
Proof. By Corollary 3.6 P[Iu]MziP[Iu]⊥ is compact. If P[Iu]M n
ziP[Iu]⊥ is compact, then
P[Iu]M n+1
zi P[Iu]⊥ = P[Iu]MziP[Iu]⊥M n
ziP[Iu]⊥ + P[Iu]MziP[Iu]M n
ziP[Iu]⊥ ∈ K.
Hence by inductive method P[Iu]M n
3.3 and Proposition 3.4 it holds for any integer n > 0 that
ziP[Iu]⊥ is compact for any positive integer n. Then by Lemma
π(P[Iv]⊥P[Iu]⊥) = π(P[Iv]⊥P[Iu]⊥M n
ziM∗n
v Sn∗u ).
= π(P[Iv]⊥M n
= ¯un
i π(Sn
i vn
zi P[Iu]⊥)
ziP[Iu]⊥M∗n
zi P[Iu]⊥)
(3.5)
Then π(P[Iv ]⊥P[Iu]⊥) ≤ vin for every positive integer n, inducing that π(P[Iv]⊥P[Iu]⊥) = 0.
Combining these two lemmas, we immediately get the following result.
Corollary 3.9. Let u, v ∈ ∂Dd, such that Vu 6= Vv and Vu ∪ Vv satisfies Condition A. If Iu, Iv ⊂
C[z1, . . . , zd] are essentially quasi-prime homogeneous ideals with variety Vu, Vv respectively, then
[Iu]⊥ is asymptotically orthogonal to [Iv]⊥.
Asymptotically orthogonal quotient modules posses the following additive property.
Lemma 3.10. If essentially normal quotient modules N1,N2,N3 of H 2(Dd) are asymptotically
orthogonal to each other, then the quotient module N = N1 + N2 is essentially normal and asymp-
totically orthogonal to N3.
Proof. By [19, Proposition 3.1], N is closed and therefore a quotient module. By [19, Theorem
3.3], if 1 ≤ j, k ≤ d then [S∗zj + ¯ilS∗zk , Szj + ilSzk ] is compact for l = 0, 1, 2, 3. Then by the identity
[S∗zj , Szk ] =
3
Xl=0
il[S∗zj + ¯ilS∗zk , Szj + ilSzk ]
we obtain the compactness of [S∗zj , Szk ], which leads to the essential normality of N . By the
definition of asymptotical orthogonality, PN3PN1 and PN3PN2 are compact. By [19, Proposition
3.2], PN − PN1 − PN2 is compact, and therefore
PN3PN = PN3(PN − PN1 − PN2) + PN3PN1 + PN3PN2 ∈ K,
indicating the asymptotical orthogonality between N and N3.
Theorem 3.11. Let I ⊂ C[z1, . . . , zd] be a homogeneous ideal with primary decomposition Tm
j=1 Iuj
with ZDd(Iuj ) = Vuj where uj ∈ ∂Dd, and assume that ZDd(I) satisfies Condition A. If each Iuj is
essentially quasi-prime, then [I]⊥ is essentially normal.
Proof. By Proposition 3.4 each [Iuj ]⊥ is essentially normal, and by Corollary 3.9 [Iui]⊥ is asymp-
totically orthogonal to [Iuj ]⊥ whenever i 6= j. Then the essential normality of [I]⊥ follows from
Lemma 3.10.
10
4 The ideal associated to an essentially normal quotient module
The present section is devoted to prove that, Theorem 3.11 produces all the essentially normal
homogeneous quotient modules of H 2(Dd).
Throughout this section we assume that the ideal I ⊂ C[z1, . . . , zd] is homogeneous, and denote
by N = [I]⊥. By proposition 2.6, if N is essentially normal then Z(I) satisfies Condition A. Write
its primary decomposition as
m
I =
Ik,
\k=1
with the zero variety ZDd(Ik) = Vuk for some uk ∈ ∂Dd. Denote by Nk = [Ik]⊥ for 1 ≤ k ≤ m, then
clearly Nk ⊂ N .
A basic fact is that, even if ZDd(I) satisfies Condition A, [I]⊥ might not be essentially normal.
Example 4.1. Consider the quotient module of H 2(D2) associated to the ideal I = z2
1
Clearly ZD2(I) = V(0,1), for which Condition A holds. Then
C[z1, z2].
[√I]⊥ = span {zn
2 : n = 0, 1, . . .},
being essentially normal.
[I]⊥ ⊖ [√I]⊥, which cannot be compact. Notice that [I]⊥ ⊖ [√I]⊥ is of infinite dimension.
It is routine to verify that [Sz1, S∗z1] is the orthogonal projection onto
We have the following necessary conditions to essential normality of quotient modules.
Lemma 4.2. If N is essentially normal, then Sf is compact whenever f ∈ √I.
Proof. By definition, there is a natural number n such that f n ∈ I, and therefore S∗n
since [S∗f , Sf ] is compact, (S∗f Sf )n = (S∗f Sf )n − S∗n
of Sf .
Lemma 4.3. Suppose N is essentially normal, and 1 ≤ k ≤ m. Then S∗f PNk is compact whenever
f ∈ √Ik.
Proof. Since f ∈ √Ik there is an integer N > 0 such that f 2N ∈ Ik. Therefore
f = 0. Then
f also be compact, leading to the compactness
f Sn
f Sn
PNk S2N
f S∗2N
f PNk = 0 ∈ K.
The essential normality of N assures that [S∗f , Sf ] is compact, and then we have
PNk ∈ K,
PNk (Sf S∗f )2N −1
PN (Sf S∗f )2N −1
PNk = PNk (Sf S∗f )2N
PNk is compact and so PNk (Sf S∗f )2N −1
and therefore PN (Sf S∗f )2N −1
2, . . . , 0. Then the lemma follows from PNk Sf S∗f PNk ∈ K.
By similar procedure we can recursively obtain the compactness of PNk (Sf S∗f )2n
PNk is.
PNk , n = N −
The following lemma is a consequence of Lemma 3.3, and we admit the notations from its proof.
11
Lemma 4.4. Let u ∈ ∂Dd\Td. Fix j0 ∈ Λc, and denote by Λ1 = {1, . . . , d}\{j0}. Suppose
I ⊂ C[z1, . . . , zd] is the prime ideal with variety Vu. Set I0 = I ∩ C[zΛ1], then I0 is the prime ideal
of C[zΛ1] with variety ZDd−1(I0) = DuΛ1 such that
[I]⊥ = (cid:8)
∞
Xn=0
zn
j0 ¯un
j0M∗n
h f0(zΛ1) : f0 ∈ N0(cid:9),
where N0 = [I0]⊥ is the quotient module of H 2(DΛ1) associated to I0. Moreover Sh is essentially
unitary on [I]⊥.
Proof. If f0 ∈ N0, set
f (z) =
∞
Xn=0
zn
j0 ¯un
j0M∗n
h f0(zΛ1).
Since ¯uj0 < 1 we have f ∈ H 2(Dd). If g ∈ I0, then since M∗g f0 = 0 we have M∗g f = 0. It is routine
f = 0 for hj = zj − ujh,∀j ∈ Λc. Therefore M∗g f = 0,∀g ∈ I since I is generated
to verify that M∗hj
by I0 and {hj : j ∈ Λc}. Hence f ∈ [I]⊥.
Conversely, each f ∈ [I]⊥ can be represented as
f (z) =
∞
Xn=0
j0fn(z), fn ∈ H 2(DΛ1).
zn
Since (M∗zj0 − ¯uj0M∗h)f = 0, we have
∞
Xn=1
zn−1
j0
fn(zΛ1 ) =
∞
Xn=0
j0 ¯uj0M∗h fn(zΛ1)
zn
and therefore fn+1 = ¯uj0M∗hfn,∀n ∈ Z+. Hence fn = ¯un
j0M∗n
h f0 and
f (z) =
∞
Xn=0
j0 ¯un
zn
j0M∗n
h f0(zΛ1).
(4.1)
For g ∈ I0, we have M∗g f = 0 and consequently M∗g f0 = 0 by (4.1), which ensures f0 ∈ N0.
Since I is prime, it follows directly that I0 is prime. By Lemma 3.3 ZDd−1(I0) = DuΛ1, and by
Proposition 3.4 Sh is essentially unitary. The proof is completed.
Lemma 4.5. Let u ∈ ∂Dd\Td. Fix j0 ∈ Λc, and denote by Λ1 = {1, . . . , d}\{j0}. Suppose
I ⊂ C[z1, . . . , zd] is an ideal with variety Vu such that Ju ⊂ I, and hj ∈ I,∀j ∈ Λc\{j0}, and
j0 ∈ I. Set I0 = √I ∩ C[zΛ1], then I0 is the prime ideal of C[zΛ1] with variety ZDd−1(I0) = DuΛ1
h2
such that
[I]⊥ ⊂ (cid:8)
∞
Xn=0
zn
j0 ¯un
j0M∗n
h f0(z) +
∞
Xn=1
nzn
j0 ¯un−1
j0
M∗(n−1)
h
g0(z) : f0, g0 ∈ N0(cid:9),
(4.2)
where N0 = [I0]⊥ is the quotient module of H 2(DΛ1) associated to I0. Moreover Sh is essentially
unitary on [I]⊥.
12
Proof. By Lemma 4.4, I0 is the prime ideal of C[zΛ1] associated to the variety DuΛ1, and
[√I]⊥ = nf ∈ H 2(Dd) : f =
∞
Xn=0
zn
j0 ¯un
j0M∗n
h f0, f0 ∈ N0o.
If f ∈ [I]⊥, then M∗2
hj0
f = 0, namely
M∗2
zj0
f = 2¯udM∗zj0
Applying (4.4) to the expansion
M∗hf − ¯u2
dM∗2
h f.
(4.3)
(4.4)
and comparing the terms involving zn
f (z) =
j0fn(z), fn ∈ H 2(DΛ1),
zn
∞
Xn=0
j0, we find
fn+2 = 2¯uj0M∗h fn+1 − ¯u2
j0M∗2
h fn,∀n ∈ Z+.
Thus f is determined by f0 and f1, and we can solve that
f =
∞
Xn=0
zn
j0 ¯un
j0M∗n
h f0 +
∞
Xn=1
nzn
j0 ¯un−1
j0
M∗(n−1)
h
g0,
(4.5)
where g0 := f1 − ¯uj0M∗hf0. Clearly I0 ⊂ I, and therefore for any g ∈ I0 we have M∗g f = 0. Then it
follows from (4.5) that M∗g f0 = 0, namely f0 ∈ N0. Then since
M∗hj0
f =
∞
Xn=0
zn
j0 ¯un
j0M∗n
h g0 ∈ [I]⊥,
we have g ∈ N0.
By Lemma 3.3 S∗h is essentially unitary. The proof of the lemma is completed.
The following lemma plays the key role in finding the necessary condition for the essential
normality.
Lemma 4.6. Suppose the ideal I ⊂ C[z1, . . . , zd] satisfies the condition of Lemma 4.5. Then M∗hj0
is essentially bounded from below on [I]⊥ ⊖ [√I]⊥.
Proof. If [I]⊥ ⊖ [√I]⊥ is of finite dimension then the conclusion is trivial, and therefore we assume
dim [I]⊥ ⊖ [√I]⊥ = ∞ in the rest of the proof. Since h /∈ I0, we have
Then since PN0Mh N0 is Fredholm, ran M∗h N0= N0. By Lemma 4.5 each f ∈ [I]⊥ ⊖ [√I]⊥ has
the expansion
ker PN0Mh N0= {0}.
f =
∞
Xn=0
zn
j0 ¯un
j0M∗n
h f0 +
∞
Xn=1
nzn
j0 ¯un−1
j0
M∗(n−1)
h
g0 ∈ [I]⊥, f0, g0 ∈ N0.
13
There is a unique f∗ ∈ N0 ⊖ ker(M∗h N0) such that M∗hf∗ = g0. Define
∞
Xn=0
uj0
Af := −
zn
j0 ¯un
j0M∗n
h f∗ +
nzn
j0 ¯un−1
j0
M∗(n−1)
h
g0,
1 − uj02
then by Lemma 4.4, f − Af ∈ [√I]⊥.
converge to 0 in the weak topology. We find successively that {f (n)
weakly. Since limm→∞ Qm = P[I]⊥ we have
Let {f (n)} and {h(n)} be sequences of unit vectors in [I]⊥ ⊖ [√I]⊥ and [√I]⊥ respectively, that
∗ } converge to 0
0 },{f (n)
0 },{g(n)
∞
Xn=1
lim
n→∞hAf (n), h(n)i
= lim
n→∞
= lim
m→∞
h−
lim
lim
uj0
j0M∗l
h f (n)
m→∞hAf (n), Qmh(n)i
n→∞Xl≤m
1 − uj02 ¯un
n→∞Xl≤m
1 − uj02h¯ul
uj0¯uj02l
1 − uj02 + luj0¯ul−1
n→∞Xl≤m
j0f (n)
(−
(−
, ¯ul
uj0
lim
lim
∗
= lim
m→∞
= lim
m→∞
= 0,
∗ + l¯ul−1
j0
M∗(l−1)
h
g(n)
0 , ¯ul
j0M∗l
h h(n)
0 i
j0g(n)
0 i + hl¯ul−1
j0
f (n)
∗
, ¯ul
j0h(n)
0 i)
j0 2)hf (n)
∗
, h(n)
0 i
where the last equality follows from
∞
Xl=0
uj0¯uj02l
1 − uj02 =
∞
Xl=1
luj0¯uj02l−2 =
uj0
1 − uj02
1
1 − uj02 .
As a consequence we have
lim
n→∞
(f (n) − Af (n)) = lim
n→∞
P[√I]⊥(f (n) − Af (n)) = − lim
n→∞
P[√I]⊥Af (n) = 0.
Therefore
= lim
= lim
lim
n→∞S∗hj0
n→∞S∗hj0
∞
Xl=0
n→∞
∞
Xl=0
¯uj02l
1
1 − uj02 ,
=
=
0
f (n)2 · g(n)
−2
Af (n)2 · g(n)
−2
2 · g(n)
h g(n)
j0 ¯ul
zl
j0M∗l
0
0
0
−2
14
j0M∗l
h f (n)
∗ +
lzl
j0 ¯ul−1
j0
M∗l
h f (n)
∗
2 · f (n)
∗
−2
∞
Xl=1
uj02
(1 − uj02)2
uj02
j0 2 +
and
= lim
= lim
0
lim
uj0
n→∞f (n)2 · g(n)
−2
n→∞Af (n)2 · f (n)
−2
∗
∞
Xl=0
zl
j0 ¯ul
n→∞ −
1 − uj02
∞
+ l¯ul−1
1 − uj02 ¯ul−1
Xl=1
∞
Xl=0
1 − uj02 ¯ul
∞
Xl=0
2 − (1 + uj02) + u4
− uj02
− uj02
(cid:20)(l + 1)(l + 2) − (l + 1)
j0
(1 − uj02)3
1
(1 − uj02)3 ,
=
=
=
=
=
inducing that
j0 + (l + 1)¯ul
j02 +
1 + uj02
1 − uj02 +
j0 + uj02(1 − uj02)
(1 − uj02)2
uj04
(1 − uj02)2(cid:21)uj02l +
uj02
(1 − uj02)2
lim
n→∞
S∗hj0
f (n)
f (n)
= 1 − uj02.
is essentially bounded from below on [I]⊥ ⊖ [I′]⊥.
is essentially bounded from below on [I]⊥ ⊖ [√I]⊥.
Thus S∗hj0
Corollary 4.7. Let u ∈ ∂Dd\Td. Define polynomials h(z) = Pi∈Λ ¯uizi and hj(z) = zj − ujh(z) for
j ∈ Λc. Let I ⊂ C[z1, . . . , zd] be an ideal of variety Vu such that Ju ⊂ I, and hihj ∈ I,∀i, j ∈ Λc.
Then Pj∈Λc Shj S∗hj
Proof. Without loss of generality we assume Λ = {1, . . . , k}. The conclusion is trivial if dim [I]⊥ ⊖
[√I]⊥ < ∞, so we assume dim [I]⊥ ⊖ [√I]⊥ = ∞ in the remaining of the proof. For j ∈ Λc denote
by Ij ⊂ C[z1, . . . , zd] the ideal generated by I and hk+1, . . . , hj, . . . , hd, where hj means hj being
omitted. Then I = Tj∈Λc Ij.
If homogeneous f ∈ [I]⊥ ⊖ [√I]⊥ is orthogonal to Pj∈Λc[Ij]⊥, then f belongs to each Ij and
therefore f ∈ I, inducing that f = 0. Then we can find homogeneous fj ∈ [Ij]⊥ ⊖ [√I]⊥ of the
same degree as f and such that f = Pj∈Λc fj. Let {f (n)} be any sequence of nonzero homogeneous
polynomials in [I]⊥ ⊖ [√I]⊥, such that limn→∞ deg(fn) = ∞. Then by the proof of the previous
lemma we have
Shj S∗hj f (n), f (n)i · f (n)−2
Shj S∗hj f (n)
i · f (n)−2
2 · f (n)−2
(1 − uj2)2f (n)
, f (n)
j
j
j
lim
n→∞hXj∈Λc
n→∞hXj∈Λc
n→∞ Xj∈Λc
= lim
≥ lim
≥ min
j∈Λc
(1 − uj2)2 > 0.
15
cannot be compact.
Corollary 4.8. Let u ∈ ∂Dd\Td, and Λ ⊂ {1, . . . , d} such that ui ∈ T if i ∈ Λ, and ui ∈ D if i ∈ Λc.
Define polynomials h(z) = Pi∈Λ ¯uizi and hj(z) = zj − ujh(z) for j ∈ Λc. Let I ⊂ C[z1, . . . , zd] be
an ideal of variety Vu which is not essentially quasi-prime, then Pj∈Λc Shj S∗hj
Proof. Let J ⊂ C[z1, . . . , zd] be the ideal generated by {hj : j ∈ Λc}. For positive integer n, denote
by In ⊂ C[z1, . . . , zd] the ideal generated by I′ and J n. Since J ⊂ √I′, there is a positive integer N
such that J N ⊂ I′, and therefore I′ = IN . Since dim [I′]⊥ ⊖ [√I]⊥ = √I/I′ = ∞, for each integer
m > 0 we can find a homogeneous f ∈ [IN ]⊥ ⊖ [√I]⊥ of degree greater than m + N . Suppose
f ∈ [In+1]⊥ but f /∈ [In]⊥ where 1 ≤ n < N . If n = 1 then f ∈ [I2]⊥ ⊖ [√I]⊥. If n > 1 then
there is some homogeneous g0 ∈ J (n−1) of degree n − 1 such that S∗g0f /∈ [√I]⊥, since otherwise
S∗g f = 0 for all g ∈ J n, which contradicts to f /∈ [In]⊥. Then since f ∈ [In+1]⊥, we have S∗g S∗g0f = 0
whenever g ∈ J 2, implying S∗g0f ∈ [I2]⊥. Therefore S∗g0f − P[I ′]⊥S∗g0f is a non-vanishing element
in [I2]⊥ ⊖ [√I]⊥, of degree greater than m + N − n + 1 > m. Hence in either cases there is a
sequence {fm} of unit vectors in [I2]⊥ ⊖ [√I]⊥ that converges to 0 weakly. Then by Corollary 4.7,
P[I ′]⊥ Pj∈Λc Mhj M∗hj [I ′]⊥ cannot be compact, proving the conclusion of the corollary.
Corollary 4.8 together with Lemma 4.3 indicates that, Theorem 3.11 actually gives the necessary
and sufficient condition for a homogeneous quotient module to be essentially normal. We summarize
these results in the following theorem, which is aforementioned as Theorem 1.2.
Theorem 4.9. Let I ⊂ C[z1, . . . , zd] be a homogeneous ideal for which [I]⊥ is of infinite dimension.
Then N = [I]⊥ is essentially normal if and only if the following items hold,
1) ZDd(I) satisfies Condition A, and hence ZDd(I) = Si
2) Let I = Tm
j=1 Iuj be the primary decomposition with ZDd(Iuj ) = Vuj , then each Iuj is essen-
Vui for several different discs Vui,
tially quasi-prime.
Finally we give two examples to illustrate how the algebraic structure of the ideal determines
the essential normality of its quotient module.
Example 4.10. Let I ⊂ C[z1, z2, z3] be the ideal generated by {(z1 − z2)2, z3(z1 + z2), z2
3}, then
3}, and √I is generated by
ZD3(I) = {(z, z, 0) : z ∈ D}. Clearly I′ is generated by {z1 − z2, z3z1, z2
{z1 − z2, z3}. It is not hard to verify √I = I′ + Cz3, and therefore dim√I/I′ = 1. Then [I]⊥ is
essentially normal by Theorem 3.11.
Example 4.11. Let I ⊂ C[z1, z2, z3] be the ideal generated by {(z1 − z2)2, z3(z1 − z2), z2
3}, then
3}, and √I is the ideal
ZD3(I) = {(z, z, 0) : z ∈ D}. Clearly I′ is the ideal generated by {z1 − z2, z2
generated by {z1 − z2, z3}. For natural number n the polynomial en(z) = z3Pn
i=0 zi
belongs to
√I, but does not lie in I′. Then √I/I′ cannot be of finite dimension, and by Theorem 1.2 [I]⊥ is
not essentially normal.
1zn−i
2
Although the forms of the ideals in these examples look similar, the essential normality of their
quotient modules are totally different.
16
5 K-homology for homogenous quotient module
Let I be a homogenous ideal such that [I]⊥ is essentially normal, and σe([I]⊥) be the joint essential
spectrum of the commuting tuple (Szd,··· , S∗zd
). Similar to [3, Proposition 2.5], it is routine to
verify that C∗([I]⊥) is irreducible. Therefore if [I]⊥ is essentially normal then K ⊂ C∗([I]⊥), and
we obtain the following short exact sequence
0 → K ֒→ C∗([I]⊥) → C(σe([I]⊥)) → 0,
(5.1)
The C∗−algebra extension theory as well as the K-homology is based on this short exact sequence[4].
To investigate the K-homology, we first calculate the essential spectrum σe([I]⊥); then examine
whether extension (5.1) yields a non-trivial element. Similar to [25], we have the following lemma.
Lemma 5.1. Suppose u ∈ ∂Dd and I be a essentially quasi-prime ideal with variety Vu, then
σe([I]⊥) = Z(I) ∩ ∂Dd.
Proof. By Theorem 1.2, λi − Szi(i = 1, . . . , d) are essentially normal. By [7, Corollary 3.9], the
(λi − Szi)(λi − Szi)∗ is Fredholm. First
tuple (λ1 − Sz1, . . . , λd − Szd) is Fredholm if and only if
we prove the assertion
d
Pi=1
Z(I) ∩ ∂Dd ⊂ σe([I]⊥).
(λi − Szi)(λi − Szi)∗
Otherwise, there is some λ = (λ1, . . . , λd) ∈ Z(I) ∩ ∂Dd making T =
Fredholm. Since T is positive, there is an invertible positive operator B and a compact operator
n} in Vu ∩ Dd that converges to λ as n → ∞. Since
K such that T = B + K. Take a sequence {µ
{kµ
Pi=1
n} converges to 0 weakly, there is a positive number c such that
n→∞hBkµ
n→∞h(B + K)kµ
n→∞hT kµ
, kµ
ni ≥ c.
ni = lim
, kµ
lim
, kµ
n
d
n
n→∞λ − µ
n2 = 0,
(5.2)
contradicting to the previous inequality. Hence the assertion is proved.
Conversely, f (Sz1, . . . , Szd) = 0 for each f ∈ I, and then the Spectral Mapping Theorem
ensures σe([I]⊥) ⊆ Z(f ). It follows that σe([I]⊥) ⊆ Z(I). Then since kSzik ≤ 1 for each i, we have
σe([I]⊥) ⊆ D
. By Lemma 3.3, there is some i ∈ Λu such that Szi is essentially unitary, which
implies σe(Szi) ⊂ T. Then for each (λ1,··· , λd) ∈ Dd, the tuple (λ1−Sz1,··· , λd−Szd) is Fredholm
and therefore σe([I]⊥) ⊂ ∂Dd. The proof of the lemma is completed.
d
Now let I be an arbitrary homogenous ideal such that [I]⊥ is essentially normal, then Z(I)
satisfies Condition A by Theorem 1.2. Let ZDd(I) = Si
Vi for different discs Vi, and I = Ti Ii be
the primary decomposition with Z(Ii) ∩ Dd = Vi. Then by Corollary 3.9, [Ii]⊥ is asymptotically
orthogonal to [Ij]⊥ whenever i 6= j. It follows that
σe([I]⊥) = [i
σe([Ii]⊥) = Z(I) ∩ ∂Dd.
We summarize this in the following proposition.
17
However, since µ
n ∈ Vu, kµ
n
ni = lim
n ∈ [I]⊥ and it holds that
n→∞hT kµ
ni = lim
lim
, kµ
n
Proposition 5.2. Let I be a homogenous ideal such that [I]⊥ is essentially normal, then
σe([I]⊥) = Z(I) ∩ ∂Dd.
At the end of this paper, we prove the non-triviality of extension (5.1).
Theorem 5.3. Let I be a homogenous ideal such that [I]⊥ is essentially normal, then the short
exact sequence
0 → K ֒→ C∗([I]⊥) → C(Z(I) ∩ ∂Dd) → 0
is not split.
Proof. Suppose ZDd(I) = Si Vui for different discs {Vui}. Let I = Ti Iui be the primary decompo-
sition such that Vui = Z(Iui) ∩ Dd. Since I′ui/Iui is of finite dimension, and Iui is asymptotically
orthogonal to Iuj whenever i 6= j, it suffices to show that the short exact sequence
0 → K ֒→ C∗([Iuj ]⊥) → C(∂Vuj ) → 0
(5.3)
is not split. By [17, Lemma 5.5], it is enough to find a Fredholm operator in C∗([Iuj ]⊥) with nonzero
Fredholm index. Fix a k ∈ Λuj , then S∗zk [Iuj ]⊥ is essentially unitary. Moreover by Remark 3.5,
Ind (S∗zk [Iuj ]⊥) 6= 0, completing the proof.
Acknowledgements. The authors would like to thank Yanyue Shi from Ocean University of
China, whose critical example of essentially normal quotient module stimulated us to improve the
concept of quasi-prime ideals. We thank Professor Kunyu Guo for his encouragements and interests.
We also thank Li Chen for his comments on the earlier version of this manuscript.
References
[1] J. Agler and J. E. McCarthy, Distinguished varieties, Acta Math. 194 (2005), no. 2, 133-153,
DOI: 10.1007/BF02393219.
[2] W. Arveson, The Dirac operator of a commuting d-tuple, J. Funct. Anal. 189 (2002), 53-79.
[3] W. Arveson, Quotients of standard Hilbert modules, Trans. Amer. Math. Soc. 359 (2007), no.
12, 6027-6055.
[4] L. Brown, R. Douglas and P. Fillmore, Extension of C∗-algebras and K-homology, Ann. of
Math. 105 (1977), 265-324.
[5] X. Chen and K. Guo, Analytic Hilbert Modules, CRC Research Notes, 413, 2003.
[6] D. N. Clark, Restrictions of H p functions in the polydisk, Amer. J. Math. 110 (1988), 1119-
1152.
[7] R. Curto, Fredholm and invertible n-tuples of operators. The deformation problem. Trans.
Amer. Math. Soc. 266 (1981), no. 1, 129-159.
18
[8] R. G. Douglas, Essentially reductive Hilbert modules, J. Oper. Theory. 55 (2006), 117-133.
[9] R. G. Douglas, Essentially reductive Hilbert modules. II. Hot topics in operator theory, 79-87,
Theta Ser. Adv. Math., 9, Theta, Bucharest, 2008.
[10] R. G. Douglas and V. I. Paulsen, Hilbert Modules over Function Algebras, Longman Research
Notes, 217, 1989.
[11] R. G. Douglas and G. Misra,Some calculations for Hilbert modules, J.Orissa Math. Soc. 12-15
(1993-1996), 75-85.
[12] R. G. Douglas, X. Tang and G. Yu, An Analytic Grothendieck Riemann Roch Theorem, Adv.
Math. 294 (2016), 307-331.
[13] R. G. Douglas and K. Wang, A harmonic analysis approach to essential normality of principal
submodules, J. Funct. Anal. 261 (2011), 3155-3180.
[14] M. Englis, Density of algebras generated by Topelitz operators on Bergman spaces, Ark. Mat.
30 (1992), 227-240.
[15] Q. Fang and J. Xia, Essential normality of polynomial-generated submodules: Hardy space and
beyond, J. Funct. Anal. 265 (2013), 2991-3008.
[16] K. Guo, Defect operators for submodules of H 2
d , J. Reine Angew. Math. 573 (2004), 181-209.
[17] K. Guo and K. Wang, Essentially normal Hilbert modules and K-homology, Math. Ann., 340
(2008), 907-934.
[18] K. Guo and K. Wang, Beurling type quotient modules over the bidisk and boundary represen-
tations, J. Funct. Anal. 257 (2009), 3218-3238.
[19] K. Guo and P. Wang, Essentially normal Hilbert modules and K-homology, III. Homogenous
quotient modules of Hardy modules on the bidisk, Sci. China. Ser. A 50 (2007), no. 3, 387-411.
[20] K. Guo and P. Wang, Essentially normal Hilbert modules and K-homology IV: Quasi-
homogenous quotient modules of Hardy module on the polydisks, Sci. China Math, 55 (2012),
no. 8, 1613-1626.
[21] K. Guo, K. Wang and G. Zhang, Trace formulas and p-essentially normal properties of quotient
modules on the bidisk, J. Oper. Theory, 67 (2012), no. 2, 511-535.
[22] M. Kennedy and O. Shalit, Essential normality, Essential norms and hyperrigidity, J. Funct.
Anal., 268 (2015), 2990-3016.
[23] B. Krishna Das, S. Gorai and J. Sarkar, On Quotient modules of H 2(Dn): Essential Normality
and Boundary Representations, arXiv: 1410.5633.
[24] P. Wang, The essential normality of Nη-type quotient module of Hardy module on the polydisc,
Proc. Amer. Math. Soc. 142 (2014), no. 1, 151-156.
[25] P. Wang and C. Zhao, Essential normality of homogenous quotient modules over the polydisc:
distinguished variety case, arXiv: 1508.03481.
19
|
1712.01047 | 2 | 1712 | 2019-02-21T12:14:02 | Approximation properties of hybrid shearlet-wavelet frames for Sobolev spaces | [
"math.FA"
] | In this paper, we study a newly developed shearlet system on bounded domains which yields frames for $H^s(\Omega)$ for some $s\in \mathbb{N}$, $\Omega \subset \mathbb{R}^2$. We will derive approximation rates with respect to $H^s(\Omega)$ norms for functions whose derivatives admit smooth jumps along curves and demonstrate superior rates to those provided by pure wavelet systems. These improved approximation rates demonstrate the potential of the novel shearlet system for the discretization of partial differential equations. Therefore, we implement an adaptive shearlet-based algorithm for the solution of an elliptic PDE and analyze its computational complexity and convergence properties. | math.FA | math |
Approximation properties of hybrid shearlet-wavelet frames for
Sobolev spaces
Philipp Petersen∗
Mones Raslan†
February 22, 2019
Abstract
In this paper, we study a newly developed hybrid shearlet-wavelet system on bounded domains which
yields frames for H s(Ω) for some s ∈ N, Ω ⊂ R2. We will derive approximation rates with respect to
H s(Ω) norms for functions whose derivatives admit smooth jumps along curves and demonstrate superior
rates to those provided by pure wavelet systems. These improved approximation rates demonstrate the
potential of the novel shearlet system for the discretization of partial differential equations. Therefore, we
implement an adaptive shearlet-wavelet-based algorithm for the solution of an elliptic PDE and analyze
its computational complexity and convergence properties.
Keywords: shearlets, wavelets, Sobolev spaces, Approximation properties
MSC classification: 42C40, 65M60, 41A25, 65T99, 94A12
1
Introduction
A well-established paradigm of applied harmonic analysis is the usage of representation systems to store,
analyze and manipulate data. An especially famous representation system is provided by wavelets (see e.g.
[21]), which can be considered as a standard tool in signal and image processing as well as in the numeri-
cal analysis of partial differential equations. While being perfectly suited to approximate one-dimensional
functions with point singularities, these systems admit one serious defect in 2D. In fact, two-dimensional
wavelet systems yield significantly suboptimal approximation rates when dealing with functions that admit
anisotropically structured singularities. Most importantly curvilinear singularities cannot be handled ade-
quately. This fact is even more severe since such structures appear very frequently in natural data as, for
instance, virtually every photograph admits at least one jump in color value along a line or a curve.
As a remedy anisotropic directional systems were invented, starting with curvelets [5, 4]. Later contourlets
[23] and shearlets [41] were developed. All of these systems yield optimal approximation rates of functions
that exhibit curvilinear singularities. From these systems, shearlets stand out due to their unique combination
of desirable features which include a unified treatment of digital and continuous realm, compactly supported
elements, fast implementations, and optimal approximation rates. Besides, the shearlet transform can be
interpreted as a unitary representation of the so-called shearlet group. This offers a natural definition of
associated smoothness spaces, customarily called shearlet coorbit spaces [14, 16]. In addition, 2D wavelet
systems, as well as shearlet systems, are both examples of the concept of wavelets with composite dilations
[34], making both systems very close conceptually such that we can think of shearlets as an extension of 1D
wavelets to 2D which retains the optimal approximation properties for functions with jump singularities.
It is because of these properties that a very effective line of research deals with the identification of
application areas of wavelets, where improvements can be made by switching to shearlet systems. Examples
∗Mathematical Institute, University of Oxford, Andrew Wiles Building, Woodstock Road, Oxford, OX2 6GG, UK, e-mail:
[email protected]
†Institut fur Mathematik, Technische Universitat Berlin, Strasse des 17. Juni 136, 10623 Berlin, Germany, e-mail:
[email protected]
1
where wavelet results were improved using shearlets include denoising and inpainting [40, 27], edge detection
[36], regularization of ill-posed problems [2, 35] and the reconstruction from Fourier samples [44].
One of the strongholds of wavelet methods is their remarkable ability to yield efficient discretizations of
PDEs. These discretizations lead to very effective numerical algorithms, which we will review below. These
results are only possible because wavelet systems can be constructed to yield frames or Banach frames for
Sobolev spaces. Additionally, wavelet systems can be built on bounded domains also incorporating boundary
conditions. Since PDEs are usually considered on bounded domains, this property is crucial.
Following the idea laid out earlier, we plan to use shearlets -- just like wavelets -- for the discretization of
PDEs. We build our results on a recently developed shearlet system on bounded domains [30], which yields
frames for H s(Ω), while allowing for boundary conditions.
1.1 Motivation: Adaptive frame methods
A first important step in utilizing systems from harmonic analysis for the solution of partial differential
equations was achieved by Cohen, Dahmen, and DeVore in [10] who developed a wavelet-based solver for
linear, elliptic PDEs guaranteeing optimal convergence and complexity provided that the solution of the PDE
contains only point singularities. The results from [10] were extended in [49, 13, 15] such that it becomes
possible to work with general frames instead of wavelet bases. Furthermore [12] broadened the results of [13]
to nonlinear elliptic PDEs.
Utilizing anisotropic frames for solving PDEs numerically is a relatively new topic, mainly due to the
difficulty of constructing them on bounded domains. Let us shortly review which conditions should be
fulfilled such that the methods introduced in the aforementioned papers work efficiently.
For linear problems, the aim is to obtain a well-conditioned discrete system of linear equations equivalent
to the PDE by using a frame, which then can be solved by standard numerical methods. By now there exist
two main approaches to end up with such a discrete system: the first one, introduced in [49], uses frames
for the Sobolev space H s(Ω), the solution space of the PDE. The second one uses the concept of Gelfand
frames, where a frame for L2(Ω) is used, which has the additional property that its synthesis operator is
bounded as a map from a weighted sequence space into H s(Ω) and the analysis operator with respect to its
dual frame is a bounded map from H s(Ω) into the sequence space.
Of course, in both cases, the frames should be constructed in such a way that boundary conditions on
the PDE can be imposed. Furthermore, the involved frame should yield optimally sparse approximations
to the solution in order to obtain optimal asymptotic convergence rates of the numerical algorithm to the
solution with low computational work and high accuracy.
So far, some first attempts have been made in utilizing anisotropic frames for solving elliptic PDEs.
In [28, 31] optimal ridgelet-based solvers were developed for linear advection equations, whereas in [17, 19]
shearlet-based solvers for general advection equations were developed. Although these works constitute major
successes in advancing anisotropic frame-based solvers, it was not possible to impose boundary conditions.
To overcome these problems, a hybrid shearlet-wavelet frame was constructed in [30] which will be further
reviewed in the following subsection.
1.2 Anisotropic multiscale systems on bounded domains
Shearlet systems on R2 admit two central properties. First of all, it was demonstrated in [33, 38], that
they admit optimal approximation rates for so-called cartoon-like functions. Here, cartoon-like functions are
functions which are piecewise C2-functions on R2 with a C2-discontinuity curve. In fact, using shearlets,
the L2-error of best N -term approximation for cartoon-like functions is decaying with a rate of N −1 (up to
logarithmic terms) -- a considerable improvement compared to the approximation rate N −1/2 provided by
wavelets. We also mention a recent extension of analyzing the approximation rates of functions whose higher-
order derivatives are cartoon-like functions [45]. Also in this scenario shearlets yield superior approximation
rates over wavelets.
The second cornerstone of shearlet theory is provided by the fact that shearlet systems yield stable
decompositions and reconstructions of functions in L2(R2), more specifically they form frames for L2(R2).
2
A construction of compactly supported shearlets admitting this property was provided in [37].
While both these properties appear very beneficial especially given our long-term goal to use shearlets
as a discretization tool for partial differential equations, we can observe that the standard construction is
not yet fully satisfactory for that task. In fact, a significant obstacle that needs to be tackled is that the
standard construction only constitutes a representation system for L2(R2), while a PDE is usually defined on
a bounded domain Ω ⊂ R2. Therefore, it is necessary to introduce multiscale systems on bounded domains,
which retain the frame and approximation properties of its R2 counterpart.
After extensive research, this task is quite well understood for wavelets (see for example [11, 9, 8, 6]) but
there are still open questions despite the isotropic character of wavelets. All the constructions of wavelets on
bounded domains hinge upon the multiresolution structure of the systems in order to construct boundary
adapted elements. Therefore it is clear that shearlet systems for bounded domains are even harder to
construct since such a structure is missing for these systems. Furthermore, the anisotropic shape of the
support can intersect the boundary to various degrees and at various angles requiring different boundary
adaptation for each element.
The first attempt in this direction has been made in [39], where an L2-frame with optimal approximation
properties for cartoon-like functions was constructed. Unfortunately, the resulting system is not boundary
adapted; hence it is not possible to characterize Sobolev spaces or impose boundary conditions on a PDE.
Another attempt has been made in [17, 19], but the resulting systems neither constitute frames for L2(Ω)
nor characterize Sobolev spaces.
A system that provides all the desired properties was finally developed in [30, 46]. It was demonstrated
how to construct a hybrid frame for H s(Ω), where s ∈ N0 = N ∪ {0} and Ω ⊂ R2 by combining a shearlet
frame with a wavelet frame on Ω. Roughly speaking, the resulting frame consists in one part of all elements
of some shearlet frame for H s(R2) with compact support (see for instance [37, 16] for a construction) the
support of which is fully contained in Ω. Additionally, boundary adapted wavelets in a small tubular region
around the boundary are included in the system, so that it becomes possible to impose boundary conditions.
If a frame for L2(Ω) is constructed in the just described way, then it can be shown that the system
characterizes Sobolev space norms by weighted ℓ2 norms of the associated analysis coefficients. Moreover, it
was numerically established in [46] that the system also yields a Gelfand frame so that it can be used for the
numerical solution of PDEs using the Gelfand frame approach introduced in [13]. Since only a few wavelets
are used in the construction, it is still possible to show that the resulting system approximates an extended
class of cartoon-like functions, defined on Ω optimally.
Additionally, by changing the construction slightly, the system also constitutes a frame for H s(Ω), which
is the set-up we will use in this work. In this case, the approximation properties of the system, at least with
respect to H s(Ω) norms have not yet been analyzed in [30].
1.3 Our contribution
In this paper, we provide further studies of a hybrid shearlet-wavelet frame for H s(Ω). Our analysis is
two-fold. First of all, we establish novel approximation rates for functions with curvilinear singularities
within their derivatives with respect to pure shearlet frames for Sobolev spaces on R2 as well as on bounded
domains Ω with respect to a hybrid shearlet-wavelet frame. We will establish improved rates over pure
wavelet systems. Secondly, we implement an adaptive numerical algorithm for the solution of a model PDE.
We will observe that the convergence of this algorithm matches our theoretical results and is of higher-order
than that provided by wavelet discretizations.
1.4 Outline
The paper is organized as follows. First of all, in Section 2 we present the main concepts needed throughout
subsequent sections. A short revision of shearlets on H s(R2) for s ∈ N0 is included. Most importantly, it
also contains a review of hybrid shearlet-wavelet frames on bounded domains for H s(Ω) for some bounded
domain Ω ⊂ R2 which were introduced in [30, 46]. In Section 3 we prove approximation results with respect
to the best N -term approximation in the H s(Ω)-norm. In particular, we will provide fast approximation
3
rates for functions which have cartoon-like first- or higher-order derivatives and will also explain why these
functions appear frequently as solutions of elliptic PDEs. In Section 4 we implement an adaptive algorithm
based on the constructed Sobolev frame, solving an exemplary PDE. We will observe very fast convergence
rates, which thereby highlight the potential of the new system.
2 Preliminaries
In this section, we first introduce shearlet and wavelet systems as well as basic notation.
2.1 Basic notation
1 · · · Dad
for a = (a1, ..., ad) ∈ Nd
0 let a = a1 + ... + ad and Da = Da1
For any Banach space (X, k·kX ) let X ′ be its topological dual. We furthermore use the usual multiindex
d . On Rd the Euclidean
notation:
scalar product shall be denoted by hx, yi and the induced norm by x for x, y ∈ Rd. We also denote by
x the absolute value for some x ∈ C. For a measurable subset Ω ⊆ Rd denote by Lp(Ω), p ∈ [1, ∞]
the usual Lebesgue spaces and for a countable index set Λ and p ∈ [1, ∞] by ℓp(Λ) the usual sequence
spaces. The cardinality of some set I shall be denoted by #I and the Lebesgue measure of some measurable
set Ω ⊂ R2 by Ω. The Fourier transform we use is given by (F f )(ξ) = f (ξ) = RRd f (x)e−2πihx,ξidx for
f ∈ L1(Rd) ∩ L2(Rd) which can be uniquely extended to L2(Rd). We note that for f ∈ L2(Rd) the Plancherel
identity kf k2 = kF f k2 holds and F −1 exists. For any subset Ω ⊆ Rd let ∂Ω be its boundary. For some
Hilbert space H and some closed linear subspace M of H, let PM denote the orthogonal projection onto
M. For some normed space X, some M ⊆ X and x ∈ X let d(x, M ) := inf{kx − ykX : y ∈ M } denote the
distance of x to M. Furthermore, for some r > 0 let Br(x) denote the open ball with radius r and center
x. If we have A ≤ C · B for two quantities A, B and some constant C > 0 we write A . B and A & B if
B . A. Furthermore A ∼ B shall be written if A . B and B . A hold. Furthermore, for a real number x
we denote by ⌊x⌋ the largest integer less than x.
2.2 Frames
In this subsection, we will introduce frames for Hilbert spaces.
Definition 2.1 ([26, 7]). A countable subset (ϕλ)λ∈Λ of some Hilbert space H is called a frame for H, if
there exist 0 < A ≤ B < ∞, such that
A kf k2
H ≤ Xλ∈Λ
hf, ϕλiH2 ≤ B kf k2
H ,
for all f ∈ H.
(1)
For every frame Φ = (ϕλ)λ∈Λ, we can define the linear and bounded analysis operator TΦ : H → ℓ2(Λ),
f 7→ (hf, ϕλiH)λ∈Λ as well as its adjoint, the synthesis operator, given by T ∗
(cλ)λ∈Λ 7→Pλ∈Λ cλϕλ. Then, due to the frame property (1), it is easy to see that the frame operator
ΦTΦ is a bounded and boundedly invertible operator from H onto itself. Therefore, we can define
Φ ϕλ)λ∈Λ, which is a frame itself. Afterwards, one can obtain the
Φ : ℓ2(Λ) → H,
)λ∈Λ := (S−1
SΦ = T ∗
the canonical dual frame (ϕdual
reconstruction formula
λ
f = Xλ∈Λ
hf, ϕdual
λ
iHϕλ = Xλ∈Λ
hf, ϕλiHϕdual
λ
,
which holds for every f ∈ H.
Let Φ = (ϕλ)λ∈Λ be a frame for a Hilbert space H. In view of real-world applications, it is reasonable to as-
sume that we cannot store all frame coefficients or all frame elements. Hence, for a given f =Pλ∈Λ cλϕλ ∈ H
we would like to approximate f in an optimal way by only using linear combinations of N frame elements.
To make this formal, following [22], we define the space of nonlinear approximation by
ΣN := [#M≤N
span{ϕλ : λ ∈ M }.
4
Then
σN,Φ,H(f ) := inf
fN ∈ΣN
kf − fN kH
is called the error of best N-term approximation of f with respect to Φ. There exists a very important
connection between the error of best N -term approximation and weak ℓp-spaces which are defined by
ℓp
w := {(cj)j∈N : sup
n∈N
n1/pc∗
n < ∞},
where (c∗
in [22] that for c ∈ ℓ2 we have that
n for all n ≥ 1. It was shown
n)n∈N is a rearrangement of c = (cλ)λ∈Λ ∈ ℓ2(Λ) such that c∗
n+1 ≥ c∗
σN,Φ,H(c) . N −s ⇐⇒ c ∈ ℓp
w
for p = ( 1
basis of ℓ2(N). If (ϕn)n∈N is a frame for some Hilbert space H one can still get
2 + s)−1, where H = ℓ2(N) and the N -term approximation is taken with respect to the canonical
(hf, ϕniH)n∈N ∈ ℓp
w =⇒ σN,Φ,H(f ) . N −s,
(2)
where the N -term approximation is taken with respect to the canonical dual frame. We add that ℓp ֒→
ℓp
w ֒→ ℓp+ǫ for 0 < p < 2 and ǫ > 0.
2.3 Function spaces
Since we aim to derive approximation properties of shearlet frames for Sobolev spaces on R2 as well as on
bounded domains, we briefly recall their definition.
Definition 2.2. Let Ω ⊆ R2 be an open domain and s ∈ N. Then the Sobolev space W s,2(Ω) of order s is
given by
W s,2(Ω) := {f ∈ L2(Ω) : Daf ∈ L2(Ω) for all a ≤ s}.
These spaces, equipped with the inner product hf, giW s,2(Ω) :=Pa≤shDaf, DagiL2(Ω), are Hilbert spaces.
If Ω = R2, then we can characterize W s,2(R2) by using Bessel potential spaces:
Definition 2.3. Let s ∈ N. Then the Bessel potential space of order s is given by
H s(R2) := {f ∈ L2(R2) : F −1[(1 + · 2)
s
2 F f ] ∈ L2(R2)}.
These spaces, equipped with the inner product hf, giHs(R2) := h(1 + · 2)
spaces.
s
2 f , (1 + · 2)
s
2 giL2(R2), are Hilbert
In fact, it holds that k·kW s,2(R2) ∼ k·kHs(R2), which is why in the coming sections we will sometimes
work on H s(R2) instead of W s,2(R2). If Ω ( R2, we will always work with W s,2(Ω), but in order to simplify
0 (Ω) := {u ∈ H 1(Ω) : u = 0 on ∂Ω} and
notation we will write H s(Ω) := W s,2(Ω). We furthermore define H 1
H −1(Ω) := (H 1
s,p we denote the usual Besov spaces.
0 (Ω))′. On top of that, by Bt
2.4 Wavelet systems
Let Ω ⊂ R2 be a bounded open domain.
In the sequel, we will work with wavelet systems on bounded
domains. Later on we will assume a number of properties that the wavelet systems should have. For now
we only stipulate that they should be indexed in a specific way.
Definition 2.4. A set W ⊂ L2(Ω) is called a boundary wavelet system for some J0 ∈ Z, if it can be written
as
W = {ωJ0,m,0 : m ∈ KJ0} ∪ {ωj,m,υ : j ≥ J0, m ∈ Kj, υ = 1, 2, 3},
where Kj ⊂ Ω, #Kj ∼ 22j. We denote the index set by
Θ := {(J0, m, 0) : m ∈ KJ0} ∪ {(j, m, υ) : j ≥ J0, m ∈ Kj, υ = 1, 2, 3}.
5
Explicit constructions of boundary wavelet systems that yield biorthogonal bases for L2(Ω) can be found,
for example, in [11, 20, 18, 6, 1, 47].
2.5 Shearlet systems on R2
Shearlet systems were introduced in [32, 33] with the intent to improve on suboptimal approximation rates
of wavelet systems for natural images.
The key towards faster approximation rates for functions displaying curvilinear singularities is to replace
the isotropic scaling of wavelets by an anisotropic variant paired with an operation to adjust the orientation
of the elements of the system. Towards such a construction, we define for j, k ∈ Z the parabolic scaling and
shearing matrices by
Aj =(cid:18) 2j
0
0
2
2 (cid:19) , and Sk =(cid:18) 1 k
0 1 (cid:19) .
j
Using the matrices above, we present the definition of a cone-adapted shearlet system:
Definition 2.5 ([38, 32]). Let φ, ψ ∈ L2(R2), c = [c1, c2]T ∈ R2 with c1, c2 > 0. Then the cone-adapted
shearlet system is defined by
where
SH(φ, ψ, ψ, c) := Φ(φ, c1) ∪ Ψ(ψ, c) ∪ Ψ( ψ, c),
Φ(φ, c1) :=(cid:8)ψ0,0,m,0 = φ(· − c1m) : m ∈ Z2(cid:9) ,
Ψ(ψ, c) :=nψj,k,m,1 = 2
Ψ( ψ, c) :=nψj,k,m,−1 = 2
4 ψ(ST
k
3j
4 ψ(SkAj · −Mcm) : j ∈ N0, k ≤ 2⌈ j
3j
Aj · −Mcm) : j ∈ N0, k ≤ 2⌈ j
2 ⌉, m ∈ Z2o ,
2 ⌉, m ∈ Z2o ,
with ψ(x1, x2) = ψ(x2, x1), Mc := diag(c1, c2), Mc = diag(c2, c1), Aj = diag(2j/2, 2j).
For cone-adapted shearlet systems we will employ the index set
Λ := {(j, k, m, ι) : ιj ≥ j ≥ 0, k ≤ ι2
j
2 , m ∈ Z2, ι ∈ {1, 0, −1}}.
Then we can write the cone-adapted shearlet system as (ψj,k,m,ι)(j,k,m,ι)∈Λ.
It was shown in [37, 16] that compactly supported shearlets can be constructed to yield frames for L2(R2).
A further sufficient condition was presented in [43]. An easily fulfillable condition, under which reweighted
shearlet systems (2−jsψj,k,m,ι)(j,k,m,ι)∈Λ constitute frames for H s(R2), s ≥ 0, was given in [30].
For ν > 0 let ST AR2(ν) denote the set of all star-shaped subsets of R2 with C2-boundary and curvature
bounded by ν. Then we can define the set of all cartoon-like functions in the following way:
Definition 2.6. For ν > 0, let E 2(ν) be the set of all functions f ∈ L2(R2), for which there exist some
B ∈ ST AR2(ν) and fi ∈ C2(R2) with compact support in (0, 1)2 as well as kfikC 2(R2) ≤ 1 for i = 1, 2 such
that
We call E 2(ν) the set of cartoon-like functions.
f = f1 + XBf2.
It was established in [25] that for an arbitrary dictionary Φ = (ϕi)i∈I for L2(R2) the optimal achievable
best N -term approximation rate for the class of cartoon-like functions is given by
σN,Φ,L2(R2)(f ) = O(N −1),
provided that only polynomial depth search is used to compute the approximation. Let us now describe
the approximation rate which shearlets achieve. First of all, for a shearlet system (ψj,k,m,ι)(j,k,m,ι)∈Λ and
6
f ∈ L2(R2) we denote by (c∗
Under mild assumptions on the generator functions φ and ψ it was shown in [33, 38] that
n(f ))n∈N the non-increasing rearrangement of (hf, ψj,k,m,ιiL2(R2)2)(j,k,m,ι)∈Λ.
Xn≥N
c∗
n(f ) . N −2 log(N )3.
(3)
If (ψj,k,m,ι)(j,k,m,ι)∈Λ constitutes a frame, then (3) yields that the best N -term approximation rate with
respect to any dual frame of the shearlet system Φdual obeys
σN,Φdual,L2(R2)(f ) = O(N −1 log(N )
3
2 ) for all f ∈ E 2(ν),
and, for the sequence (c∗
of these results to functions in H s(R2) whose higher-order derivatives are cartoon-like.
n(f ) . n−3 log(n)3. One aim of this paper is the generalization
n(f ))n∈N, we have c∗
2.6 Hybrid shearlet-wavelet systems on bounded domains
In [30], a hybrid shearlet-wavelet system on a bounded domain Ω ⊂ R2 was constructed, resulting from a
combination of parts of a wavelet frame on a bounded domain and parts of a shearlet system on R2. In order
to still obtain good approximation properties it is essential not to include all elements of the wavelet system
in the hybrid system. In particular, the elements which will be included should be located only in a thin
strip Γγ(j) close to the boundary of Ω, where for r > 0 and some fixed qsh > 0, which will be specified later,
Γr = {x ∈ Ω : d(x, ∂Ω) < qsh2−r},
γ(j) depends on the scale j of the wavelets, and Γγ(j) shrinks for increasing scales. Before we continue with
the defintion of a hybrid shearlet-wavelet system and the analysis of the frame property, we need to introduce
the following assumptions imposed on the underlying wavelet and shearlet systems.
Assumption 2.7. [30] Let s ∈ N0, W be a boundary wavelet system and SH(φ, ψ, ψ, c) be a shearlet system.
Then, we assume the following properties of the boundary wavelet system:
(W1) (2−jsωj,m,υ)(j,m,υ)∈Θ is a frame for H s(Ω) and there exists a dual frame
Hs = (2−jsωj,m,υ)dual
(j,m,υ)∈Θ and for all (j, m, υ) ∈ Θ with ∂Ω ∩ supp (2−jsωj,m,υ)dual = ∅ it holds
W dual
that
\(2−jsωj,m,υ)dual(ξ) . 2−js · 2−j
min{1, 2−jξiαw }
max{1, 2−jξ1βw } max{1, 2−jξ2βw }
,
(4)
for at least one i ∈ {1, 2}, some αw, βw > 0 and all ξ ∈ R2. Here the Fourier transform is to be
understood on L2(R2) after extension by 0. Furthermore, we assume that the elements of W dual
Hs have
compact support and let
q(0)
w := inf(cid:8)q > 0 : supp (2−jsωj,m,υ)dual ⊂ B2−j q(m),
for all (j, m, υ) ∈ Θ(cid:9) > 0.
(W2) The elements of W have compact support and
q(1)
w := inf(cid:8)q > 0 : supp ωj,m,υ ⊂ B2−j q(m),
Moreover, we have that
for all (j, m, υ) ∈ Θ(cid:9) > 0.
m − m′ ≥ 2−jq(2)
w ,
for all j ≥ J0 and m, m′ ∈ Kj, m 6= m′,
for some q(2)
w > 0.
Furthermore, we assume the following properties of SH(φ, ψ, ψ, c):
7
(S1) (2−jsψj,k,m,ι)(j,k,m,ι)∈Λ is a frame for H s(R2) with dual frame (2−jsψj,k,m,ι)dual
(j,k,m,ι)∈Λ.
(S2) For some C1, C2 > 0 the decay conditions
and
bψ(ξ1, ξ2) ≤ C1
bψ(ξ1, ξ2) ≤ C2
min{1, ξ1αsh}
max{1, ξ1βsh} max{1, ξ2βsh}
min{1, ξ2αsh}
max{1, ξ1βsh} max{1, ξ2βsh}
(5)
(6)
are obeyed for all (ξ1, ξ2) ∈ R2 and some αsh, βsh > 0.
(S3) For all (j, k, 0, ι) ∈ Λ and some qsh > 0 we have that
supp ψj,k,0,ι ⊂ B2−j/2qsh/2(0).
Essentially, Assumption 2.7 requires a wavelet frame and a shearlet frame for H s(Ω), H s(R2) respectively.
Moreover, the shearlet and wavelet frames are required to be built from compactly supported elements.
For the wavelet system it is additionally required that the associated dual frame is compactly supported
as well. For wavelet systems forming orthonormal bases this property is automatically satisfied. Finally,
three regularity assumptions are made in (4) (5), and (6).
In combination with the assumptions on the
supports, these regularity assumptions are necessary to control the correlation between wavelets elements
and shearlets elements in the following definition of a hybrid shearlet-wavelet system. The construction of
the hybrid shearlet-wavelet system functions by including wavelet elements supported close to the boundary
of the domain and shearlet elements that are supported inside the domain.
Definition 2.8. Let s ∈ N0, SH(φ, ψ, eψ, c) = (ψj,k,m,ι)(j,k,m,ι)∈Λ be a shearlet system fulfilling (S3), let
τ > 0 and t > 0. Further, let W be a boundary wavelet system fulfilling (W1) and (W2) and set
Wt,τ := {ωj,m,υ ∈ W : (j, m, υ) ∈ Θt,τ } ,
where
In addition, let
Θt,τ :=n(j, m, υ) ∈ Θ : B2−j (q(0)
w )(m) ∩ Γτ (j−t) 6= ∅o .
w +q(1)
Then, the hybrid shearlet-wavelet system with offsets t and τ is defined as
Λ0 := {(j, k, m, ι) ∈ Λ : supp ψj,k,m,ι ⊂ Ω} .
HSW s
t,τ (W; φ, ψ, ψ, c) := {ψj,k,m,ι : (j, k, m, ι) ∈ Λ0} ∪ Wt,τ .
The parameters τ and t in Definition 2.8 control the size of the boundary strip Γτ (j−t) and thereby
determine which wavelets from the underlying wavelet frame are included in the hybrid shearlet-wavelet
system. Intuitively a large τ and a small t imply an asymptotically quickly shrinking boundary strip for
j → ∞. On the other hand, decreasing τ or increasing t will increase the width of the strip.
The frame property for H s(Ω) of a hybrid shearlet-wavelet system on minimally smooth domains Ω ⊂ R2
(see [48]) has been analyzed in [30]:
Theorem 2.9 ([30]). Let SH(φ, ψ, ψ, c) and W satisfy (S1), (S2), (S3) and (W1), (W2), respectively, with
s ∈ N0, αw > 1, αsh > 0, βw ≥ s, βsh > 2 + 2s + 2αw, τ > 0 and ǫ > 0 such that ((1 − ǫ)/τ − 2)αw > 5
2 .
Then there exists some T > 0 such that, for any t ≥ T , then we have that the hybrid shearlet-wavelet system
(ϕn)n∈N = HSW s
t,τ (W; φ, ψ, ψ, c) satisfies
kf k2
Hs(Ω) ∼ Xn∈N(cid:12)(cid:12)(cid:12)(cid:10)f, 2−jnsϕn(cid:11)Hs(Ω)(cid:12)(cid:12)(cid:12)
2
,
8
for all f ∈ H s(Ω).
As already outlined in Subsection 2.5, the second core property of shearlets, besides the frame property,
is the optimal approximation rate of shearlets for cartoon-like functions, which are defined on Ω in the
following way:
Definition 2.10. Let ν > 0, and Ω ⊂ R2 be a domain, B ∈ ST AR2(ν), and fi ∈ C2(R2), kfikC 2(R2) < ∞
If #(∂B ∩ ∂Ω) ≤ M for some M ∈ N and ∂Ω and ∂B only intersect transversely, then
for i = 1, 2.
PΩ(f1 + χBf2) is a cartoon-like function on Ω. We denote the set of cartoon-like functions on Ω by E 2(ν, Ω).
It was shown in [30, 46] that for all functions f ∈ H l(Ω) such that for all a = l, Daf ∈ E 2(ν, Ω) one
obtains
σN,Φdual,L2(Ω)(f ) = O(N − l+2
2
log(N )
3
2 ),
(7)
with respect to any dual frame of the hybrid shearlet-wavelet system and the L2(Ω)-norm. For l = 0 and up
to the log factor, this is the optimal approximation rate.
In Section 3, we will prove a similar approximation rate to (7). There we measure the best N -term
approximation error with respect to the hybrid shearlet-wavelet system and the H l(Ω)-norm instead of the
L2(Ω)-norm.
3 Approximation rates
The two essential properties of shearlet systems on L2(R2) and L2(Ω), for a domain Ω ⊂ R2, are the
frame property as well as the property to obtain optimally sparse approximations of functions that exhibit
anisotropic structures, in our case cartoon-like functions. We recalled these properties in Subsection 2.5. As
we have seen already, reweighted shearlet systems will constitute frames for the Sobolev spaces H s(R2) or
H s(Ω), for s ∈ N0.
Certainly, cartoon-like functions with jump singularities do not belong to any Sobolev space H s(Ω) for
s ≥ 1
2 . Hence, we need to introduce a suitable and relevant generalization of the cartoon-model to Sobolev
spaces. Given that we are interested in capturing curve-like singularities we will analyze functions with
cartoon-like derivatives, i.e., functions f such that for all multi-indices b ≤ s we have that Dbf ∈ E 2(ν, Ω).
Of course this definition includes functions with smooth derivatives.
In the upcoming Subsection 3.1 we will demonstrate that this class of functions with cartoon-like deriva-
tives does appear in practice, particularly as solutions of elliptic PDEs.
Thereafter we will establish approximation rates admitted by reweighted shearlet systems for these func-
tions in various Sobolev norms. We will focus on the analysis of the case of bounded domains since this
is more involved. Nonetheless, all approximation results hold for H s(R2) with minor changes. Similarly to
the approximation rates on L2(R2) we will observe that for this function class our shearlet system admits
considerably faster approximation rates than wavelets in Subsection 3.3.
3.1 Regularity of solutions of elliptic PDEs
In this subsection we will recall some literature describing the regularity of the derivatives of the solution
of an elliptic PDE. We will ultimately see that the solution of an elliptic PDE with cartoon-like right-hand
side will have cartoon-like first- or higher-order derivatives.
We consider PDEs of the type:
Lu = − X1≤i,j≤2
Di(aijDju) + X1≤i≤2
biDiu + cu = f,
(8)
on a bounded domain Ω ⊂ R2, where aij, bi, c are bounded and measurable and L fulfills the uniform
ellipticity condition
λx2 ≤ X1≤i,j≤2
aij(x)xixj,
for all x ∈ Ω and some λ > 0.
(9)
9
In particular, the special case L = −∆ is included in the setting above. Furthermore, we are interested in
the case that f is a cartoon-like function on Ω or f = D1g1 + D2g2 ∈ H −1(Ω), where g1, g2 are cartoon-like
functions. To have any chance of good approximation rates by shearlets, we need the solution u to be smooth
away from curvilinear singularities. A first naive approach in this direction would be to examine standard
Sobolev estimates in order to analyze the global regularity of the solution u. A standard elliptic result, see
e.g. [24] yields that for the homogeneous Dirichlet problem (i.e., (8) on Ω and u = 0 on ∂Ω), provided that
f ∈ H m(Ω), ∂Ω ∈ Cm+2 and aij, bi, c are sufficiently smooth, the solution u ∈ H m+2(Ω). However, due
to the jump curve of the right-hand side f , in general we have that f /∈ H m(Ω) for any m ≥ 1. Hence,
we can only expect u ∈ H 2(Ω). To improve upon such estimates it is clear, that we need local estimates
to show that the solution u is piecewise smooth. As a first approach in this direction, in the case where f
is cartoon-like, we can employ classical domain decomposition and Schauder estimates, see [29], to observe
that (Dαu) ∈ C2(B) ∪ C2(Ω \ ¯B). It is, however, still possible that Dαu explodes near ∂B. This would
render an analysis of shearlet approximation rates more complicated. If in addition (Dαu)B ∈ C2( ¯B) and
(Dαu)Bc ∈ C2( ¯Bc), then it would immediately follow by the well-known Whitney extension theorem [50]
that Dαu is cartoon-like on Ω.
In this scenario the approximation rates that will be established in the
following subsection can be used.
Such an analysis of the behavior of Dαu close to the discontinuity of f can be found in [42]. We briefly
describe their setup. Let us now consider
Lu = − X1≤i,j≤2
Di(aijDju) = f with f =Xi
hi + Digi,
0
0
where a fulfills the uniform ellipticity condition (9). For the special case that h1 is cartoon like, h2, g1, g2 = 0
1(cid:19) we have a standard Poisson equation with cartoon-like right hand side.
and a =(cid:18)1
Under the assumptions that we can decompose Ω as Ω = SN
m=1 Ωm, where Ωm ⊂ Ω, ∂Ωm ∈ C∞ and
, hiΩm, giΩm ∈ C∞( ¯Ωm) for all m ≤ N, we have that u ∈ C∞( ¯Ωm ∩ Ω) [42,
we further assume that ai,j Ωm
Proposition 1.4].
This result implies, that if we stipulate some additional conditions on the boundary behavior of u, we
can conclude, that if f is cartoon-like and u is a solution of
then the second derivatives of u are cartoon-like. Furthermore, using the same argument, if
f = D1g1 + D2g2 ∈ H −1(Ω), with g1, g2 being cartoon-like, then the solution u of (10) has derivatives, which
are cartoon-like.
Lu = f,
(10)
3.2 Approximation rates for functions with cartoon-like derivatives
We provide approximation rates for two cases. First of all, we examine the approximation rates of reweighted
hybrid shearlet-wavelet systems in the H s(Ω)-norm of functions whose s-th order derivatives are cartoon-like.
Such functions appear as solutions of elliptic differential equations if the right hand side is a derivative of a
cartoon-like function, as described in Subsection 3.1. From now on, let Ω ⊂ R2 be a bounded domain, whose
boundary has finite length.
Theorem 3.1. Let s ∈ N, ν > 0, c ∈ R2 and φ, ψ, ψ ∈ L2(R2). For a ≤ s, let
{(Daψ)j,k,m,ι : (j, k, m, ι) ∈ Λ} := SH(cid:16)Daφ, Daψ, Da ψ, c(cid:17)
and let (θa
that for all f ∈ E 2(ν, Ω) we have that
n(f))n∈N be a non-increasing rearrangement of (hf, (Daψ)j,k,m,ιiL2(Ω))(j,k,m,ι)∈Λ. Further assume
for all a ≤ s. Furthermore, let (ωj,m,υ)(j,m,υ)∈Θ be a wavelet system on Ω such that for all (j, m, υ) ∈ Θ:
N (f) . N − 3
θa
2 log
3
2 N, for all N ∈ N,
(11)
10
• supp ωj,m,υ . 2−2j,
• kDaωj,m,υk∞ . 2(a+1)j,
• ωj,m,υ has two vanishing moments for all υ 6= 0.
On top of that, let t ∈ N, τ > 1/3 and (ϕn)n∈N := HSW s
Dau ∈ E 2(ν, Ω) for all a = s, we have
t,τ (W; φ, ψ, ψ, c). Then, for all u such that
ΞN (u) . N − 3
2 log
3
2 N for all N ∈ N,
where (Ξm(u))m∈N is a non-increasing rearrangement of (hu, 2−jnsϕniHs(Ω))n∈N.
Proof. First of all, we have that with (Ξm(u))m∈N being a non-increasing rearrangement of
(hu, 2−jnsϕniHs(Ω))n∈N that
ΞN (u) . θsh
⌊N/2⌋(u) + θw
⌊N/2⌋(u),
(12)
where (θsh
is a non-increasing rearrangement of (hu, 2−jsωj,m,υiHs(Ω))(j,m,υ)∈Θt,τ .
n (u))n∈N is a non-increasing rearrangement of (hu, 2−jsψj,k,m,ιiHs(Ω))(j,k,m,ι)∈Λ0 and (θw
n (u))n∈N
We start by estimating θsh
N (u) and we use that
hu, 2−jsψj,k,m,ιiHs(Ω) ≤ X0≤a≤s
hDau, 2−jsDaψj,k,m,ιiL2(Ω).
Let us now assume without loss of generality that ι = 1, then there holds for 0 ≤ a ≤ s that
hDau, 2−jsDaψj,k,m,1iL2(Ω) ≤ hDau, 2−jaDaψj,k,m,1iL2(Ω)
= hdDau, 2−ja(2πiξ)abψj,k,m,1iL2(R2)
≤ hdDau, (2πi)s2−ja2/2(ST
+ hdDau, (2πi)s(2−ja2/2(ST
= I + II.
−kA−j ξ)abψj,k,m,1iL2(R2)
−kA−jξ)a − 2−jaξa)bψj,k,m,1iL2(R2)
Part I can be estimated by
We continue with Part II and observe that
I ≤ hdDau, \(Daψ)j,k,m,1iL2(R2) = hDau, (Daψ)j,k,m,1iL2(Ω).
2−ja2/2(ST
−kA−jξ)a − 2−jaξa = Xb≤a,b2<a2
cb1,b2ξb,
where cb1,b2 . 2−ja. Hence, to estimate II it suffices to estimate
(13)
(14)
(15)
where b2 < a2. At this point we are again in the situation of (14) and continue the estimate until we finally
arrive at a point, where a2 = 0. In this situation, we have
hdDau, (2πi)s2−jaξbbψj,k,m,1iL2(R2),
hdDau, (2πi)s2−jaξa1
1 bψj,k,m,1iL2(R2) = (2πi)s−a1 hdDau, 2−ja1(2πiξ1)a1bψj,k,m,1iL2(R2)
= (2πi)s−a1 hdDau, (2πiST
= (2πi)s−a1 hdDau, \(D(a1,0)ψ)j,k,m,1iL2(R2)
= (2πi)s−a1 hDau, (D(a1,0)ψ)j,k,m,1iL2(Ω).
−kA−jξ)(a1,0)bψj,k,m,1iL2(R2)
11
From the considerations above we can estimate for some C > 0
θsh
N (u) . Xa≤s Xb≤s
θsh,b
⌊N/C⌋(Dau),
(16)
where (θsh,b
N (g))N ∈N denotes the non-increasing rearrangement of (hg, 2−js(Dbψ)j,k,m,ιiL2(Ω)). By (11)
we can estimate (16) by N − 3
N (u). We continue by
estimating θw
N (u). Assume first that supp ωj,m,υ intersects the singularity curve of Dau, which shall be
denoted by γ a for 0 ≤ a ≤ s. Recall that supp ωj,m,υ . 2−2j and kDaωj,m,υk∞ ≤ 2(a+1)j. Then, we
have
2 N for all N ∈ N. This yields the estimate of θsh
2 log
3
hDau, 2−jsDaωj,m,υiL2(Ω) ≤ 2−jskDauk∞kDaωj,m,υk∞ supp ωj,m,υ
. 2−j(s+2)2(a+1)j ≤ 2−j(s+2)2j(s+1) = 2−j.
From the width of Γτ (j−t) we know that we have, up to some multiplicative constant, only 2(1−τ )j wavelet
elements on scale j intersecting γ a. Consequently, since τ > 1/3, we get that
(hDau, 2−jsDaωj,m,υiL2(Ω))(j,m,υ)∈Θt,τ ,(supp ωj,m,υ∩γ a)6=∅ ∈ ℓ
2
3 .
Lastly, we want to address the wavelets associated with the smooth part of Dau. We assumed that Daωj,m,υ
has two vanishing moments for all j, m and υ 6= 0 and hence by a taking a Taylor approximation of Dau on
supp ωj,m,υ we obtain
hDau, 2−jsDaωj,m,υiL2(Ω) . 2−jskDaωj,m,υk∞2−2j2−2j . 2−3j.
Since there are 2(2−τ )j wavelets on every scale we obtain that
(hDau, 2−jsDaωj,m,υiL2(Ω))(j,m,υ)∈Θt,τ ,(supp ωj,m,υ∩γ a)=∅ ∈ ℓ
2
3 .
Combining both estimates for the wavelets, we obtain that
(hDau, 2−jsDaωj,m,υiL2(Ω))(j,m,υ)∈Θt,τ ∈ ℓ
2
3 .
Remark 3.2.
This yields that θw
2 . Invoking (12) concludes the proof.
N (u) . N − 3
• The assumptions of Theorem 3.1 are easily satisfied. As long as ((Daψ)j,k,m,ι)(j,k,m,ι)∈Λ
is a shearlet system meeting the criteria for optimally sparse approximations of cartoon-like functions
the assumptions are fulfilled. The necessary theoretical machinery has been established in [38].
• Theorem 3.1 also establishes the coefficient decay of a shearlet frame for H s(R2). To observe this one
can simply choose Ω ⊃ (0, 1)2 sufficiently large so that no wavelet element intersects (0, 1)2. Then all
scalar products of wavelets and u are zero, demonstrating the result.
Theorem 3.1 implies the following corollary, which describes the best N -term approximation rate of the
reweighted hybrid shearlet-wavelet system, if it constitutes a frame.
Corollary 3.3. Under the assumptions of Theorem 3.1 and the additional assumption that (2−jnsϕn)n∈N is
a frame for H s(Ω) we have that
where
ku − uN k2
Hs(Ω) . N −2 log3(N ),
uN := Xn∈EN
hu, 2−jnsϕniHs(Ω)(2−jnsϕn)dual
and EN contains the indices corresponding to the N largest coefficients in modulus of hu, 2−jnsϕniHs (Ω).
12
Proof. From the frame property we obtain that the synthesis operator of the dual frame is bounded and
hence
ku − uN k2
Hs(Ω) . k(hu, 2−jnsϕniHs(Ω))n6∈EN k2
ℓ2.
From Theorem 3.1 we conclude that k(hu, 2−jnsϕniHs(Ω))n6∈EN k2
ℓ2 . N −2 log3(N ).
We will observe in Subsection 3.3 that the approximation rate of the hybrid shearlet-wavelet system
is much faster than that provided by a pure wavelet system. Moreover, we can also discuss whether the
approximation rate of Corollary 3.3 is optimal. If s = 2, the discussion in Subsection 3.1 indicates that
every cartoon-like function g can be expressed as g = ∆u for a function u such that Dbu ∈ E 2(ν, Ω) for all
multi-indices b ≤ s. While the arguments of Subsection 3.1 are technically only proved for piecewise C∞
functions, they are expected to also hold for piecewise C2 functions.
Of course this is not a formal proof, since the argument of Subsection 3.1 only holds for piecewise smooth
functions g.
However, if we assume the above statement to be true, then Corollary 3.3 implies that the new hybrid
shearlet-wavelet system yields almost the optimal approximation rate, due to the following argument. As-
sume we have a dictionary Φ. Further assume that for all u such that Dbu ∈ E 2(ν, Ω) for all b ≤ 2 we have
that Φ provides an N -term approximation rate of
ku − uN k2
H2(Ω) . N −2−ǫ,
for some ǫ > 0, where uN is the best N -term approximation of u with respect to Φ. Clearly this also implies
(cid:13)(cid:13)D2
x1u − ux1
N(cid:13)(cid:13)2
L2(Ω) . N −2−ǫ, (cid:13)(cid:13)D2
N(cid:13)(cid:13)2
x2u − ux2
L2(Ω) . N −2−ǫ,
N is the best N -term approximation of D2
where uxι
xιφ : φ ∈ Φ}.
Consequently Φ1 ∪ Φ2 yields a best N -term approximation rate of D2
x2u = ∆u which is faster than
O(N −1). Such a result would be in conflict with the well-known optimality result of [25] which states that
the optimal achievable best N -term approximation rate for the class of cartoon-like functions is given by
σN,Φ(g) = O(N −1), for cartoon-like functions g, provided that only polynomial depth search is used to
compute the approximation. Hence the approximation rate of Corollary 3.3 is very likely to be optimal.
xιu with respect to the system Φι := {D2
x1u + D2
This argument can not be extended easily to s 6= 2 since it is not necessarily true that a given cartoon-like
function g is the s-th derivative of a function u such that Dbu ∈ E 2(ν, Ω) for all multi-indices b ≤ s.
Now we turn our attention to a further important case where the (s + l)-th derivatives of u are cartoon-
like. Again, such functions appear frequently as solutions of elliptic differential equations with cartoon-like
right-hand side (see Subsection 3.1).
Theorem 3.4. Let s ∈ N, 0 < l ∈ N, ν > 0 and φ, ψ, ψ ∈ L2(R2), and c ∈ R2. For a ≤ s, let
{(Daψ)j,k,m,ι : (j, k, m, ι) ∈ Λ} := SH(cid:16)Daφ, Daψ, Da ψ, c(cid:17) .
Further assume that for all f ∈ L2(Ω) such that Dbf ∈ E 2(ν, Ω) for all b ≤ l we have that
k(hf, (Daψ)j,k,m,ιiL2(Ω))(j,k,m,ι)∈Λk
< ∞,
2
l+3
ℓ
(17)
for all a ≤ s. Furthermore, let (ωj,m,υ)(j,m,υ)∈Θ be a wavelet system on Ω such that for all (j, m, υ) ∈ Θ:
• supp ωj,m,υ . 2−2j,
• kDaωj,m,υk∞ . 2(a+1)j,
• ωj,m,υ has l + 2 vanishing moments for all υ 6= 0.
13
If τ > 1/3, then
k(hu, 2−jnsϕniHs(Ω))n∈Nk
< ∞.
2
l+3
ℓ
for all u such that Dau ∈ E 2(ν, Ω) for all a = s + l.
Proof. First of all, for p = 2
l+3 we can certainly make the decomposition
k(hu, 2−jnsϕniHs(Ω))n∈Nkp
ℓp =k(hu, 2−jsψj,k,m,ιiHs(Ω))(j,k,m,ι)∈Λ0 kp
ℓp
+ k(hu, 2−jsωj,m,υiHs(Ω))(j,m,υ)∈Θt,τ kp
ℓp = I + II.
Thus, it sufficies to prove the finiteness of I and II. We start with I. Repeating the computations from (13)
to (15), we obtain that
I . Xa≤s Xb≤s
k(hDau, (Dbψ)j,k,m,ιiL2(Ω))(j,k,m,ι)∈Λ0 kp
ℓp < ∞,
where the finiteness follows by the assumption on the regularity of u and (17).
We continue by estimating II. Assume first that supp ωj,m,υ intersects the discontinuity of Dau. Recall
that supp ωj,m,υ . 2−2j and kDaωj,m,υk∞ . 2(a+1)j and that ω has more than l vanishing moments.
Thus, using a Taylor approximation of Dau on supp ωj,m,υ we can estimate
hDau, 2−jsDaωj,m,υiL2(Ω) . 2−js2−ljkDaωj,m,υk∞ supp ωj,m,υ . 2−(l+1)j.
From the width of Γτ (j−t) we know that, up to some constant, we have only 2(1−τ )j wavelet elements
intersecting the boundary curve of Dau, which we denote by γ a.
Consequently, we get that (hDau, 2−jsDaωj,m,υiL2(Ω))(j,m,υ)∈Θt,τ ,(supp ωj,m,υ∩γ a)6=∅ ∈ ℓp since τ > 1/3.
We proceed by estimating the norm associated to the wavelets that do not intersect γ a. We assumed that
Daωj,m,υ has l + 2 vanishing moments for all j, m and υ 6= 0 and hence
hDau, 2−jsDaωj,m,υiL2(Ω) . 2−jskDaωj,m,υk∞2−(l+2)j2−2j . 2−(l+3)j.
Since there are 2(2−τ )j wavelets on every scale we obtain that
(hDau, 2−jsDaωj,m,υiL2(Ω))(j,m,υ)∈Θt,τ ,(supp ωj,m,υ∩γ a)=∅ ∈ ℓp.
Combining both estimates for the wavelets we obtain that
(hu, 2−jsωj,m,υiHs(Ω))(j,m,υ)∈Θt,τ ∈ ℓp.
This finishes the estimate of II and thus the proof.
Remark 3.5. The assumptions of Theorem 3.4 can be easily checked by considering the results in [45], which
yields the required approximation rates for functions with cartoon-like derivatives by shearlets.
We can deduce the following corollary establishing best N -term approximation rates in the case that the
reweighted hybrid shearlet-wavelet system yields a frame.
Corollary 3.6. Under the assumptions of Theorem 3.4 and the additional assumption that (2−jnsϕn)n∈N is
a frame for H s(Ω) we have that
where
ku − uN k2
Hs(Ω) . N −(l+2),
uN := Xn∈EN
hu, 2−jnsϕniHs(Ω)(2−jnsϕn)dual,
and EN contains the indices corresponding to the N largest coefficients hu, 2−jnsϕniHs(Ω).
Proof. By Theorem 3.4 we obtain that (hu, 2−jnsϕniHs(Ω))n∈N ∈ ℓ
result.
2
l+3
2
֒→ ℓ
l+3
w . Invoking (2) yields the
14
3.3 Comparison to wavelets
To assess the quality of the approximation rates achieved by hybrid shearlet-wavelet systems in Corollaries
3.3 and 3.6, we should put them into perspective with approximation rates which are achievable by pure
wavelet systems.
We will observe with the help of the following theorem that the approximation rate of Corollary 3.3
cannot be achieved by pure wavelet systems.
Theorem 3.7 ([8]). Assume that for t ∈ [s, s′] with s′ > s the spaces Bt
wavelet characterization of the form
q,q(Ω), t − s = d/q − d/p admit a
kukBt
p,p(Ω) ∼ k(2tj22( 1
2 − 1
p )jk(cλ)λ∈Θj kℓp )j≥0kℓp,
for all u =Pλ∈Θ cλωλ. Then, for t ∈]s, s′[, t − s = 2/q − 2/p, we have the norm equivalence
kukBt
q,q(Ω) ∼ kukBs
p,p(Ω) + k(2j(t−s) inf
g∈Σj
ku − gkBs
p,p(Ω))j≥0kℓq ,
where Σj is the space of all u =Pn∈Θ cnωn with all but 22j coefficients cn equal to zero.
One can easily compute, that a function in H s(Ω) whose s-th order partial derivatives are cartoon-like
is not in Bs+1
q,q (Ω) for any q. Hence we obtain with t = s + 1 and p = 2 and an accordingly chosen q that
k(2j inf
g∈Σj
ku − gkHs(Ω))j≥0kℓq ,
is not finite. Therefore it cannot hold that
ku − gkHs(Ω) . 2−j−ǫ,
inf
g∈Σj
for some ǫ > 0. Consequently, since #Σj = 22j, the best N -term approximation rate in H s(Ω)-norm does
not decrease faster than N − 1
2 −ǫ for any ǫ > 0. Invoking for some l ∈ N the same argument for a function
u ∈ H s+l(Ω) whose (s + l)-th derivative is cartoon-like yields that the approximation rate with respect to
the H s(Ω)-norm will not decrease faster than N − l+1
If one compares these results to
the approximation rates by hybrid shearlet-wavelet systems of Corollaries 3.3 and 3.6 we observe that the
approximation rates by reweighted hybrid shearlet-wavelet systems are always superior.
2 −ǫ for any ǫ > 0.
4 Numerical examples
We aim to demonstrate the future potential of shearlet systems in view of possible applications in discretiza-
tion of PDEs. Towards this goal, we will provide some numerical examples for the solution of a Poisson
equation with cartoon-like right-hand side using an adaptive discretization by the new shearlet system.
4.1 Adaptive solution of PDEs
The adaptive algorithm is based on the works of [10, 49], which establish a method to adaptively solve
In particular, let L : H → H′ ∼= H be a linear, bounded and boundedly invertible
operator equations.
operator that induces a symmetric and elliptic bilinear form. Moreover, let f ∈ H′. We want to find the
uniquely determined u ∈ H such that
Using a frame for H, the problem can now be transformed into a discrete one by setting u = T ∗
Φu. We define
L : ℓ2(Λ) → ℓ2(Λ), L = TΦLT ∗
Φ, and f = TΦf ∈ ℓ2(Λ).
(19)
Lu = f.
(18)
15
It was shown in [49] that, if L fulfills the properties described above, solving (18) is equivalent to solving the
discretized system
Lu = f.
(20)
Additionally in [10] an algorithm named SOLVE was developed that provides a solution to (20) using
adaptive refinements. Under certain assumptions on the mappings L, PranL, a discretized solution uN with
N non-zero entries can be calculated with O(N ) operations. The convergence rate of the uN to the exact
solution u of (20) is given by the error of the best N -term approximation rate of u with respect to Φ. Hence
the algorithm's convergence rate is asymptotically optimal.
4.2 Setup
Our theoretical findings show that the hybrid shearlet-wavelet system yields a frame that is very well suited
to approximate functions with anisotropic characteristics such as jump singularities in their derivatives and,
as we have seen, such anisotropic structures appear in the solutions of PDEs.
Because of this, we examine an implementation based on SOLVE, and analyze its convergence rate.
The first step is to establish an implementation of the analysis and synthesis operators of a hybrid shearlet-
wavelet system. Contrary to the aforementioned results, which are valid for wavelet systems W defined on
general domains Ω, we choose Ω = (0, 1)2 for our numerical example. This is because, our implementation
is based on the standard wavelet library WaveLab [3] and the shearlet library ShearLab [40], which both
operate under the assumption Ω = (0, 1)2. For a given digitization of the domain Ω and a function h on Ω
modeled as an n × n pixel image, WaveLab provides an analysis operator TW, such that TW(h) is a vector
containing all boundary wavelet coefficients, computed with the L2-scalar product, of a wavelet basis up to
a certain scale. To turn this operator into an analysis operator based on the H 1-scalar product, we define
TW (h) := TW (h − ∆discrh),
(21)
where ∆discr is Matlab's built-in discrete Laplacian. This definition is justified because we choose a wavelet
system for which all involved wavelets vanish at the boundary of Ω.
Similarly, ShearLab provides a built-in analysis operator TS. Using the same construction as (21) we
construct an analysis operator TS with respect to the H 1-scalar product. Besides, the analysis operator
of ShearLab, called TS, computes the coefficients associated to all possible translates, i.e., it effectively
implements the analysis operator of the system
nψj,k,(Aj m),ι : 0 ≤ j ≤ J, ι ∈ {−1, 0, 1}, k ≤ ι2⌊j/2⌋, m ∈ cZo ,
(22)
for a suitable c > 0 and a maximal scale J. However, comparing with Definition 2.5, we require an analysis
operator for the system
nψj,k,m,ι : 0 ≤ j ≤ J, ι ∈ {−1, 0, 1}, k ≤ ι2⌊j/2⌋, m ∈ c1Z, supp ψj,k,m,ι ∩ ∂Ω = ∅o
(23)
instead. To turn the analysis operator supplied by ShearLab into one more appropriate for our purposes,
we introduce a mask MS which sets every coefficient associated with the system (22) that is not in (23) to
zero. Moreover, for τ, t > 0 we construct a mask MW which sets every component of TW not associated to
Wt,τ to zero. Finally, to compute the coefficients with respect to the reweighted systems, we introduce two
matrices DW and DS that multiply each wavelet or shearlet coefficient with 2−j where j is the associated
scale. Ultimately, we now implement the analysis operator of the hybrid shearlet wavelet system by
DSMS TS (cid:19).
THSW :=(cid:18)DWMW TW
In the case that L denotes the Laplacian, the construction (19) then leads to the following discrete operator
L := −THSW∆discrTHSW
∗,
16
completing our implementation of the operator equation (18).
Since the analysis operator of ShearLab computes the coefficients of the shearlet transform by using
convolutions, it always computes all coefficients on a given scale. As a result, this transformation is inherently
non-adaptive. To analyze the adaptive routine SOLVE we thus need to simulate the adaptivity by replacing
the involved adaptive steps by threshholding. This is extensively documented in [46, Section 4.3.1].
4.3 Experiment
We solve a Poisson problem on a bounded domain Ω = (0, 1)2 with right-hand side f
−∆u = f on Ω,
u = 0 on ∂Ω,
(24)
where f = D1g + D2g and g = XB 1
of g. We have depicted both g and f in Figure 1.
6
(0.5), i.e. g is a cartoon-like function and f is the sum of the derivatives
50
100
150
200
250
300
350
400
450
500
50
100
150
200
250
300
350
400
450
500
50
100
150
200
250
300
350
400
450
500
50
100
150
200
250
300
350
400
450
500
Figure 1: Left: The cartoon-like function g. Right: The right-hand side f = D1g + D2g.
Moreover, the exact solution u is depicted in Figure 2 alongside with its derivatives. We can clearly see,
that the derivatives of u are cartoon-like functions.
50
100
150
200
250
300
350
400
450
500
100
200
300
400
500
100
200
300
400
500
100
200
300
400
500
100
200
300
400
500
100
200
300
400
500
Figure 2: Left: Solution of the PDE (24). Middle: Cartoon-like derivative in x1 direction, Right: Cartoon-
like derivative in x2 direction.
In Figure 3 we can observe an approximation rate of O(N −1) after executing our version of SOLVE,
defined in Subsection 4.2. Moreover, since SOLVE yields an approximation rate of the order of the best
N -term approximation rate, we conclude with the results of Subsection 3.3 that the approximation rate of
O(N −1) is faster than that provided by wavelet frames, which can only achieve O(N −1/2).
17
0
10
-1
10
-2
10
-3
10
-4
10
-5
10
1
10
0
10
-1
10
-2
10
10
-6
2
10
3
10
4
10
5
10
6
10
-3
10
0
500
1000
1500
2000
Figure 3: Left: Squared error of approximation, blue actual approximation, red N −1. Right: Error in each
iteration within SOLVE against number of iterations.
On top of the study of the approximation rate, we can also analyze the approximation quality. Here we
provide in Figure 4 the reconstructed functions, the error committed and the active shearlet elements. One
can clearly observe, that the error vanishes uniformly, which demonstrates the good approximation quality
of shearlets at positions of curvilinear singularities.
Additionally, we can observe, that the algorithm finds the elements that are most strongly associated
with the jump singularity of the derivatives of u.
4.4 Outlook
The discretization of elliptic PDEs by the new system appears to be a very promising line of research.
Towards a proper implementation of an adaptive scheme the following issues need to be analyzed and will
be subject of future work.
• Mapping properties:
In order to obtain theoretical guarantees for the convergence rate of a hybrid
shearlet-wavelet-based adaptive frame method, we certainly need to analyze the assumptions concern-
ing the mapping properties of the discretized operator equation in advance.
• Approximation rates with respect to the primal frame: The convergence rate of SOLVE depends on
the N -term approximation rate provided by the underlying frame. Our theoretical results only provide
approximation rates with respect to the dual of the hybrid shearlet-wavelet frame, which is standard
in shearlet literature. Nonetheless, in order to guarantee the approximation rates of SOLVE a proper
analysis of the N -term approximation rates with respect to the primal frame has to be conducted.
• Implementation for a model problem: As we mentioned before, our current implementation is based on
thresholding procedures because it is based on the available shearlet code. Clearly, an implementation
of a hybrid shearlet-wavelet-based solver that carries out all the adaptive routines properly, is desirable
and will be developed in the future.
Acknowledgments
P. Petersen and M. Raslan thank P. Grohs and G. Kutyniok for valuable discussions. P. Petersen and
M. Raslan acknowledge support from the DFG Collaborative Research Center TRR 109 "Discretization in
Geometry and Dynamics" and the Berlin Mathematical School. P. Petersen is supported by a DFG Research
Fellowship "Shearlet-based energy functionals for anisotropic phase-field methods".
18
-3
× 10
6
4
2
0
-2
-4
-6
-3
× 10
2.5
2
1.5
1
0.5
0
∗v(i) for i = 1, . . . 5; Middle: Errors of the reconstructions, i.e.,
∗v(i) − u for i = 1, . . . 5; Bottom: Active shearlet elements, i.e., the non-zero coefficients of the
Figure 4: Top: Reconstructions THSW
THSW
shearlet part of v(i).
References
[1] K. Bittner. Biorthogonal spline wavelets on the interval. In Wavelets and Splines: Athens 2005, pages
93 -- 104. Nashboro Press, Brentwood, TN, 2006.
[2] T. A. Bubba, D. Labate, G. Zanghirati, S. Bonettini, and B. Goossens. Shearlet-based regularized
ROI reconstruction in fan beam computed tomography. In Wavelets and Sparsity XVI, pages 95970K --
95970K -- 11. Proceedings of the SPIE, San Diego, CA, 2015.
[3] J. Buckheit and D. Donoho. Wavelab and reproducible research.
In Wavelets and Statistics, pages
55 -- 81. Springer-Verlag, 1995.
[4] E. Cand`es and D. Donoho. New tight frames of curvelets and optimal representations of objects with
piecewise C2 singularities. Comm. Pure Appl. Math., 57(2):219 -- 266, 2004.
[5] E. J. Cand`es and D. L. Donoho. Curvelets: a surprisingly effective nonadaptive representation of objects
with edges. In Curve and surface fitting, pages 105 -- 120. Vanderbilt University Press, Nashville, TN,
2000.
[6] C. Canuto, A. Tabacco, and K. Urban. The wavelet element method. I. Construction and analysis.
Appl. Comput. Harmon. Anal., 6(1):1 -- 52, 1999.
19
[7] O. Christensen. An introduction to frames and Riesz bases. Birkhauser Boston, Inc., Boston, MA, 2003.
[8] A. Cohen. Wavelet methods in numerical analysis. North-Holland, Amsterdam, 2000.
[9] A. Cohen, W. Dahmen, and R. DeVore. Multiscale decompositions on bounded domains. Trans. Amer.
Math. Soc., 352(8):3651 -- 3685, 2000.
[10] A. Cohen, W. Dahmen, and R. DeVore. Adaptive wavelet methods for elliptic operator equations:
convergence rates. Math. Comp., 70(233):27 -- 75, 2001.
[11] A. Cohen, I. Daubechies, and P. Vial. Wavelets on the interval and fast wavelet transforms. Appl.
Comput. Harmon. Anal., 1(1):54 -- 81, 1993.
[12] S. Dahlke, M. Fornasier, M. Primbs, T. Raasch, and M. Werner. Nonlinear and adaptive frame approx-
imation schemes for elliptic PDEs: theory and numerical experiments. Numer. Methods Partial Differ.
Equ., 25(6):1366 -- 1401, 2009.
[13] S. Dahlke, M. Fornasier, and T. Raasch. Adaptive frame methods for elliptic operator equations. Adv.
Comput. Math., 27(1):27 -- 63, 2007.
[14] 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.
[15] S. Dahlke, T. Raasch, M. Werner, M. Fornasier, and R. Stevenson. Adaptive frame methods for elliptic
operator equations: the steepest descent approach. IMA J. Numer. Anal., 27(4):717 -- 740, 2007.
[16] S. Dahlke, G. Steidl, and G. Teschke. Shearlet coorbit spaces: compactly supported analyzing shearlets,
traces and embeddings. J. Fourier Anal. Appl., 17(6):1232 -- 1255, 2011.
[17] W. Dahmen, C. Huang, G. Kutyniok, W.-Q Lim, C. Schwab, and G. Welper. Efficient resolution of
anisotropic structures. In Extraction of quantifiable information from complex systems, volume 102 of
Lect. Notes Comput. Sci. Eng., pages 25 -- 51. Springer, Cham, 2014.
[18] W. Dahmen, A. Kunoth, and K. Urban. A wavelet Galerkin method for the Stokes equations. Computing,
56(3):259 -- 301, 1996.
[19] W. Dahmen, G. Kutyniok, W. Lim, C. Schwab, and G. Welper. Adaptive anisotropic Petrov-Galerkin
methods for first order transport equations. J. Comput. Appl. Math., 340:191 -- 220, 2018.
[20] W. Dahmen and R. Schneider. Wavelets with complementary boundary conditions -- function spaces on
the cube. Results Math., 34(3-4):255 -- 293, 1998.
[21] I. Daubechies. Ten lectures on wavelets. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia, PA, 1992.
[22] R. A. DeVore. Nonlinear approximation.
In Acta numerica, pages 51 -- 150. Cambridge Univ. Press,
Cambridge, 1998.
[23] M. N. Do and M. Vetterli. Contourlets. In Beyond wavelets, volume 10 of Stud. Comput. Math., pages
83 -- 105. Academic Press/Elsevier, San Diego, CA, 2003.
[24] M. Dobrowolski. Angewandte Funktionalanalysis. Funktionalanalysis, Sobolev-Raume und elliptische
Differentialgleichungen. Springer-Verlag Berlin Heidelberg, 2010.
[25] D. L. Donoho. Sparse components of images and optimal atomic decompositions. Constr. Approx.,
17(3):353 -- 382, 2001.
[26] R. Duffin and A. C. Schaeffer. A class of nonharmonic Fourier series. Trans. Amer. Math. Soc.,
72:341 -- 366, 1952.
20
[27] G. R. Easley, D. Labate, and F. Colonna. Shearlet-based total variation diffusion for denoising. Trans.
Img. Proc., 18(2):260 -- 268, February 2009.
[28] S. Etter, P. Grohs, and A. Obermeier. FFRT - a fast finite ridgelet transform for radiative transport.
Multiscale Model. Simul., 13(1):1 -- 42, 2014.
[29] D. Gilbarg and N. S. Trudinger. Elliptic partial differential equations of second order. Springer-Verlag,
Berlin, 2001.
[30] P. Grohs, G. Kutyniok, J. Ma, P. Petersen, and M. Raslan. Anisotropic multiscale systems on bounded
domains. arXiv preprint arXiv:1510.04538, 2015.
[31] P. Grohs and A. Obermeier. Optimal adaptive ridgelet schemes for linear advection equations. Appl.
Comput. Harmon. Anal., 41(3):768 -- 814, 2016.
[32] K. Guo, G. Kutyniok, and D. Labate. Sparse multidimensional representations using anisotropic dilation
and shear operators. In Wavelets and splines: Athens 2005, pages 189 -- 201. Nashboro Press, Brentwood,
TN, 2006.
[33] K. Guo and D. Labate. Optimally sparse multidimensional representation using shearlets. SIAM J.
Math. Anal., 39(1):298 -- 318, 2007.
[34] K. Guo, D. Labate, W.-Q Lim, G. Weiss, and E. Wilson. The theory of wavelets with composite dilations.
In Harmonic analysis and applications, pages 231 -- 250. Birkhauser Boston, Boston, MA, 2006.
[35] S. Hauser and G. Steidl. Convex multiclass segmentation with shearlet regularization. Int. J. Comput.
Math., 90(1):62 -- 81, 2013.
[36] E. J King, R. Reisenhofer, J. Kiefer, W.-Q Lim, Z. Li, and G. Heygster. Shearlet-based edge detec-
tion: flame fronts and tidal flats. In SPIE Optical Engineering+ Applications, pages 959905 -- 959905.
International Society for Optics and Photonics, 2015.
[37] P. Kittipoom, G. Kutyniok, and W.-Q Lim. Construction of compactly supported shearlet frames.
Constr. Approx., 35(1):21 -- 72, 2012.
[38] G. Kutyniok and W.-Q Lim. Compactly supported shearlets are optimally sparse. J. Approx. Theory,
163(11):1564 -- 1589, 2011.
[39] G. Kutyniok and W.-Q Lim. Shearlets on bounded domains.
In Approximation Theory XIII: San
Antonio 2010, pages 187 -- 206. Springer, 2012.
[40] G. Kutyniok, W.-Q Lim, and R. Reisenhofer. ShearLab 3D: Faithful digital shearlet transforms based
on compactly supported shearlets. ACM T. Math. Software, 42(1), 2015.
[41] D. Labate, W.-Q Lim, G. Kutyniok, and G. Weiss. Sparse multidimensional representation using
shearlets. In Wavelets XI,, pages 254 -- 262. Proceedings of the SPIE, San Diego, CA, 2005.
[42] Y. Li and L. Nirenberg. Estimates for elliptic systems from composite material. Commun. Pure Appl.
Math., 56(7):892 -- 925, 2003.
[43] W.-Q Lim. The discrete shearlet transform: A new directional transform and compactly supported
shearlet frames. IEEE Trans. Image Process., 19:1166 -- 1180, 2010.
[44] J. Ma. Generalized sampling reconstruction from Fourier measurements using compactly supported
shearlets. Appl. Comput. Harmon. Anal., 42(2):294 -- 318, 2017.
[45] P. Petersen. Shearlet approximation of functions with discontinuous derivatives. J. Approx. Theory,
207(C):127 -- 138, 2016.
21
[46] P. Petersen. Shearlets on Bounded Domains and Analysis of Singularities Using Compactly Supported
Shearlets. Dissertation. Technische Universitat Berlin, 2016.
[47] M. Primbs. Stabile biorthogonale Spline-Waveletbasen auf dem Intervall. Dissertation. Universitat
Duisburg, 2006.
[48] E. M. Stein. Singular integrals and differentiability properties of functions. Princeton University Press,
Princeton, N.J., 1970.
[49] R. Stevenson. Adaptive solution of operator equations using wavelet frames. SIAM J. Numer. Anal.,
41(3):1074 -- 1100, 2003.
[50] H. Whitney. Hassler Whitney Collected Papers, chapter Analytic Extensions of Differentiable Functions
Defined in Closed Sets, pages 228 -- 254. Birkhauser Boston, Boston, MA, 1992.
22
|
1905.10122 | 1 | 1905 | 2019-05-24T10:11:45 | A Subnormal Completion Problem for Weighted Shifts on Directed Trees, II | [
"math.FA"
] | The subnormal completion problem on a directed tree is to determine, given a collection of weights on a subtree, whether the weights may be completed to the weights of a subnormal weighted shift on the directed tree. We study this problem on a directed tree with a single branching point, $\eta$ branches and the trunk of length $1$ and its subtree which is the "truncation" of the full tree to vertices of generation not exceeding $2$. We provide necessary and sufficient conditions written in terms of two parameter sequences for the existence of a subnormal completion in which the resulting measures are $2$-atomic. As a consequence, we obtain a solution of the subnormal completion problem for this pair of directed trees when $\eta < \infty$. If $\eta=2$, we present a solution written explicitly in terms of initial data. | math.FA | math |
A Subnormal Completion Problem for
Weighted Shifts on Directed Trees, II
George R. Exner, Il Bong Jung, Jan Stochel, and Hye Yeong Yun
Abstract. The subnormal completion problem on a directed tree is to deter-
mine, given a collection of weights on a subtree, whether the weights may be
completed to the weights of a subnormal weighted shift on the directed tree.
We study this problem on a directed tree with a single branching point, η
branches and the trunk of length 1 and its subtree which is the "truncation"
of the full tree to vertices of generation not exceeding 2. We provide neces-
sary and sufficient conditions written in terms of two parameter sequences for
the existence of a subnormal completion in which the resulting measures are
2-atomic. As a consequence, we obtain a solution of the subnormal completion
problem for this pair of directed trees when η < ∞. If η = 2, we present a
solution written explicitly in terms of initial data.
1. Introduction
The class of unilateral weighted shifts on Hilbert space has been a standard and
important source of examples with which to study the properties of bounded linear
operators on Hilbert space, including especially the investigation of subnormality
(see [18] and [5]). A recently introduced class of weighted shifts on directed trees
provides a more extensive collection of objects for study (see e.g., [15, 14, 1, 2, 3]).
In [10], we initiated the study of a subnormal completion problem for weighted
shift operators on directed trees. For a classical weighted shift, the subnormal
completion problem is to be given an initial finite sequence of positive weights and
to determine whether or not they may be extended to the weights of an injective,
bounded, subnormal unilateral weighted shift; such a shift is called a subnormal
completion of the initial weight sequence (see [19, 16]; see also [6, 7, 8, 9]). We
consider the analogous task in the setting of weighted shifts on directed trees.
In the present paper we continue the study of the subnormal completion prob-
lem for weighted shifts on directed trees. As in [10], we restrict our attention to the
directed tree Tη,κ with a single branching point, η branches and the trunk of length
κ, and we consider the subnormal completion problem with respect to the subtree
Tη,κ,p, which is the "truncation" of Tη,κ to vertices of generation not exceeding
1991 Mathematics Subject Classification. Primary 47B20, 47B37; Secondary 05C20.
Key words and phrases. Subnormal operator, weighted shift on a directed tree, subnormal
completion problem, 2-atomic measures.
The research of the second author was supported by the National Research Foundation of
Korea (NRF) grant funded by the Korea government (MSIT) (2018R1A2B6003660).
1
2
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
2
p. If η is finite and the p-generation subnormal completion problem on Tη,κ has a
solution for given initial data λ, then we can always find a subnormal completion
λ
i,1, which are canonically associated with S λ at vertices
S λ such that the measures µ
of the first generation, are at most ⌊ κ+p+2
⌋-atomic (see Theorem 2.2). So far as we
know, there is no solution of the p-generation subnormal completion problem on
Tη,κ written in terms of initial data for p > 2. The only explicit solution is that for
the 1-generation subnormal completion problem on Tη,1 (see [10, Theorem 5.1]).
In view of the above discussion, if the 2-generation subnormal completion problem
on Tη,1 has a solution for given initial data λ, then we can always find a subnor-
λ
mal completion S λ of λ on Tη,1 with the property that each measure µ
i,1 is 1-
or 2-atomic. In this paper we provide necessary and sufficient conditions written
in terms of two parameter sequences for λ to admit a subnormal completion S λ
λ
on Tη,1 with 2-atomic measures µ
i,1. As a consequence, we solve the 2-generation
subnormal completion problem on Tη,1 for η < ∞. The results, taken as a whole,
suggest that a complete answer to the subnormal completion problem even for this
class of directed trees is at present out of reach.
The paper is organized as follows. In Section 2, we provide notation, termi-
nology and results that are needed in this paper.
In Section 3 we carry out an
in-depth analysis of the first two "negative" moments of the Berger measure as-
sociated with the Stampfli completion of three increasing weights to the weight
sequence for a subnormal unilateral weighted shift (see Lemma 3.6). In Section 4
we state and prove the solutions of the 2-generation subnormal completion problem
on Tη,1 with 2-atomic measures (the case η = ∞ is included, see Theorem 4.1) and
without any restrictions on supports of the associated measures (only for η < ∞; see
Theorem 4.2). In view of Proposition 4.4(iii), to solve the problem with 2-atomic
measures we have to compute the infimum of a quadratic form in η variables subject
to some constraints. This task is extremely complicated. In Section 5 we give a
solution of the 2-generation subnormal completion problem on T2,1 with 2-atomic
measures written entirely in terms of the initial data (see Theorem 5.2). This
confirms the scale of the complexity of this problem.
2. Preliminaries
In this section we sketch briefly the notation and results necessary for the
present discussion, but the reader is encouraged to consult [10] for a considerably
more complete presentation.
Given a complex Hilbert space H, we denote by B(H) the C∗-algebra of all
bounded linear operators on H. Recall that an operator T ∈ B(H) is said to be
subnormal if there exists a complex Hilbert space K containing H and a normal
operator N ∈ B(K) such that T h = N h for all h ∈ H. We refer the reader to
[4, 5, 12] for the foundations of the theory of subnormal operators.
Denote by Z+, N, R, R+ and C the sets of nonnegative integers, positive inte-
gers, real numbers, nonnegative real numbers and complex numbers, respectively.
Set N2 = {2, 3, 4, . . .} and ¯N2 = N2 ∪ {∞}. Define Jι by
Jι = {k ∈ N : k 6 ι} for ι ∈ Z+ ∪ {∞},
using the convention that J0 = ∅. We write card(X) for the cardinality of a set X.
In what follows, δx stands for the Dirac Borel measure on R+ at the point x ∈ R+.
The closed support of a Borel probability measure µ on R+ is denoted by supp µ.
A SUBNORMAL COMPLETION PROBLEM, II
3
A pair G = (V, E) is a directed graph if V is a nonempty set and E is a subset
of V × V . A member of V is a vertex of G , and an element of E is called an edge
of G . For u ∈ V , we set Chi(u) = {v ∈ V : (u, v) ∈ E}. A member of Chi(u) is
called a child of u. A vertex v of G is called a root of G , which we may also write
as v ∈ Root(G ), if there is no vertex u of G such that (u, v) is an edge of G . We
set V ◦ = V \ Root(G ). We say that T = (V, E) is a directed tree if T is a directed
graph, which is connected, has no circuits and has the property that for each vertex
u ∈ V there exists at most one vertex v ∈ V , called the parent of u and denoted
here by par(u), such that (v, u) ∈ E. The reader is referred to [15, 17] for more
information on directed trees needed in this paper.
−κ −(κ−1)
•
•
···
−1
•
(1,1)
0
(2,1)
•
•
☛☛☛☛
✠✠✠✠✠✠✠
•
✹✹✹✹✹✹✹✹
✶✶✶✶
......
•
...
(1,2)
•
(1,3)
•
(1,4)
•
(2,2)
•
(2,3)
•
(2,4)
•
(3,1) (3,2)
•
...
(3,3)
•
...
(3,4)
•
...
···
···
···
Figure 1. An illustration of the directed tree Tη,κ.
We shall consider a certain class of directed trees with a single branching point
obtained as follows: given η ∈ ¯N2 and κ ∈ N, we define the directed tree Tη,κ =
(Vη,κ, Eη,κ) by (see Figure 1):
Vη,κ = {−k : k ∈ Jκ} ⊔ {0} ⊔ {(i, j) : i ∈ Jη, j ∈ N},
Eη,κ = Eκ ⊔ {(0, (i, 1)) : i ∈ Jη} ⊔ {((i, j), (i, j + 1)) : i ∈ Jη, j ∈ N},
Eκ = {(−k,−k + 1) : k ∈ Jκ}.
Define as well a subtree of Tη,κ on which we may be given "some of" the weights
of a proposed shift as initial data: for η ∈ ¯N2, κ ∈ N and p ∈ N, let the directed
tree Tη,κ,p = (Vη,κ,p, Eη,κ,p) be defined by (see Figure 2)
Vη,κ,p = {−k : k ∈ Jκ} ⊔ {0} ⊔ {(i, j) : i ∈ Jη, j ∈ Jp},
Eη,κ,p = Eκ ⊔ {(0, (i, 1)) : i ∈ Jη} ⊔ {((i, j), (i, j + 1)) : i ∈ Jη, j ∈ Jp−1}.
Given a directed tree T = (V, E), let ℓ2(V ) be the Hilbert space of all square
summable complex functions on V equipped with the usual inner product. The
family {eu}u∈V defined by
eu(v) =(1
0
if v = u,
otherwise,
v ∈ V,
is clearly an orthonormal basis of ℓ2(V ). Given a system λ = {λv}v∈V ◦ ⊆ C, we
define the operator Sλ in ℓ2(V ) by Sλf = ΛT f for f ∈ ℓ2(V ) such that ΛT f ∈
/
/
/
/
/
/
/
/
D
D
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
4
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
−κ −(κ−1)
•
•
···
−1
•
(1,1)
0
(2,1)
•
•
☛☛☛☛
✠✠✠✠✠✠✠
•
✹✹✹✹✹✹✹✹
✶✶✶✶
......
•
...
(1,2)
•
(2,2)
•
•
...
···
···
···
(1,p)
•
(2,p)
•
(3,p)
•
...
(3,1) (3,2)
Figure 2. An illustration of the directed tree Tη,κ,p.
ℓ2(V ), where ΛT acts on functions f : V → C via
(ΛT f )(v) =(λv · f (par(v))
0
if v ∈ V ◦,
if v is a root of T .
We call Sλ the weighted shift on the directed tree T with weights {λv}v∈V ◦ .
Throughout this paper it is assumed that the resulting operator Sλ is bounded
(see [10] or more generally [15] for an approach suitable even for unbounded shifts,
and for discussions of when Sλ is indeed bounded). If Sλ ∈ B(ℓ2(V )), in view of
[15, (3.1.4)], one may more easily express Sλ by
Sλeu = Xv∈Chi(u)
λvev.
(We adopt the convention thatPv∈∅ xv = 0.) The weighted shifts desired are those
which are subnormal operators. According to [15, Theorem 6.1.3 and Notation
6.1.9], the following assertion holds.
Sλ ∈ B(ℓ2(V )) is subnormal if and only if for any u ∈ V , there
exists (a unique compactly supported ) Borel measure on R+, denoted
by µλ
u , such that
(2.1)
λeuk2 =ZR+
kSn
tndµλ
u (t), n ∈ Z+.
The characterizations of subnormality of Sλ on the directed tree Tη,κ can be found
in [15, Corollary 6.2.2].
Matters are in hand for the statement of the fundamental problem considered.
Let T = (V, E) be a subtree of a directed tree T = ( V , E) and let λ = {λv}v∈V ◦
be a system of positive real numbers. We say that a weighted shift S λ on T
with weights λ = {λv}v∈ V ◦ ⊆ (0,∞) is a subnormal completion of λ on T if
S λ ∈ B(ℓ2( V )), λ ⊆ λ, i.e., λv = λv for all v ∈ V ◦, and S λ is subnormal. If such
a completion exists, we may sometimes say λ admits a subnormal completion on
T . The subnormal completion problem for (T , T ) consists of seeking necessary
and sufficient conditions for a system λ = {λv}v∈V ◦ ⊆ (0,∞) to have a subnormal
completion on T .
Following [10], the subnormal completion problem for (Tη,κ,p, Tη,κ), where
p ∈ N, is called the p-generation subnormal completion problem on Tη,κ (sometimes
/
/
/
/
/
/
/
/
D
D
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
η
ZR+
Xi=1
Xi=1
Xi=1
η
η
dµi(s) = 1,
λ2
λ2
1
1
s
sk+1
i,1ZR+
i,1ZR+
i,1ZR+
sup supp µi < ∞.
sκ+1
1
λ2
dµi(s) =
dµi(s) 6
1
1
Qk−1
Qκ−1
j=0 λ2
−j
j=0 λ2
−j
k ∈ Jκ−1,
,
,
A SUBNORMAL COMPLETION PROBLEM, II
5
abbreviated to p-generation SCP on Tη,κ). In this particular case, our initial data
takes the form
λ = {λv}v∈V ◦
η,κ,p = {λ−κ+1, . . . , λ0} ∪ {λi,1}i∈Jη ∪ . . . ∪ {λi,p}i∈Jη .
The following result, which gives the measure-theoretic way of solving the
is a consequence of [10,
p-generation subnormal completion problem on Tη,κ,
Lemma 4.7 and Theorem 4.9].
Lemma 2.1. Suppose η ∈ ¯N2, κ, p ∈ N and λ = {λv}v∈V ◦
given. Then the following conditions are equivalent:
(i) λ admits a subnormal completion on Tη,κ,
(ii) there exist Borel probability measures {µi}η
following conditions:
i=1 on R+ which satisfy the
η,κ,p ⊆ (0,∞) are
sndµi(s) =
n+1
Yj=2
λ2
i,j , n ∈ Jp−1, i ∈ Jη,
(2.2)
(2.3)
(2.4)
sup
i∈Jη
Moreover, if {µi}η
i=1 are as in (ii), then there exists a subnormal completion S λ of
λ on Tη,κ such that µ
(2.5)
λ
i,1 = µi for all i ∈ Jη.
In [10, Theorem 5.1] we gave an explicit solution of the 1-generation sub-
normal completion problem on Tη,1 written in terms of initial data. As shown
in [10, Theorem 6.2], the 1-generation subnormal completion problem on Tη,κ,
where κ ∈ N, reduces to seeking necessary and sufficient conditions for a sys-
tem λ = {λv}v∈V ◦
η,κ,1 ⊆ (0,∞) to have a 2-generation flat subnormal completion
on Tη,κ.
It was also proved in [10, Theorem 8.3] that the problem of finding a
p-generation flat subnormal completion on Tη,κ can be solved by using the well-
known solutions of the subnormal completion problem for unilateral weighted shifts
given in [19, 6, 7, 8, 9, 16]. However, in most cases these solutions are not writ-
ten explicitly in terms of initial data. This means that we have no explicit (i.e.,
written in terms of initial data) solution of the p-generation subnormal completion
problem on Tη,κ for p > 2, even when κ = 1. It is the right moment to make the
following important observation which is implicitly contained in the proof of [10,
Theorem 7.1].
Theorem 2.2. Suppose η ∈ N2, κ, p ∈ N and λ = {λv}v∈V ◦
η,κ,p ⊆ (0,∞) are
given. If λ admits a subnormal completion on Tη,κ, then it admits a subnormal
completion S λ on Tη,κ such that
card(supp µ
λ
i,1) 6j κ + p + 2
2
k,
i ∈ Jη,
6
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
where ⌊x⌋ = min(cid:8)n ∈ Z+ : n 6 x < n + 1(cid:9) for x ∈ R+.
Proof. Applying [6, Theorems 5.1(iii) and 5.3(iii)] and arguing as in the proof
of the implication (iii)⇒(iv) of [10, Theorem 7.1], we may assume without loss of
generality that for every i ∈ Jη, the measure ρi appearing in the proof of the
implication (iv)⇒(v) of [10, Theorem 7.1] satisfies the following condition:
card(supp ρi) 6j κ + p + 2
2
k and supp ρi ⊆ (0,∞).
Next, by arguing as in the proof of the implication (iv)⇒(v) of [10, Theorem 7.1],
we get a subnormal completion with the desired property.
(cid:3)
It follows from Theorem 2.2 that if η ∈ N2 and the 2-generation subnormal
completion problem on Tη,1 has a solution for given data λ = {λv}v∈V ◦
η,1,2 ⊆ (0,∞),
then one can always find a subnormal completion S λ of λ on Tη,1 with the property
that each measure µ
λ
i,1 is 1- or 2-atomic.
3. Preparatory lemmas
One of the goals of this paper is to explore when initial data λ = {λv}v∈V ◦
λ
i,1 (see (2.1)) is 2-atomic (under present circumstances, the measures µ
on Tη,1,2 admits a subnormal completion S λ on Tη,1 such that each of the measures
λ
i,1 may be
µ
taken to be 1- or 2-atomic due to Theorem 2.2). We first require some background
on the Stampfli completion of three increasing weights to the weight sequence for
a subnormal unilateral weighted shift (cf. [19]).
η,1,2
In [19] the author gives an explicit construction of the completion of an ini-
tial finite sequence of three weights x, y, z satisfying 0 < x < y < z to the se-
quence of positive weights {αn}∞n=0 for a (bounded, injective) subnormal unilateral
weighted shift Wα. This includes a construction of the associated Berger measure
of Wα (i.e., a unique Borel probability measure ξ on R+ such that for any n ∈ N,
tndξ(t)), which turns out to be 2-atomic. The completion
γn := α2
sequence {αn}∞n=0, which is called the Stampfli completion of (x, y, z), is custom-
arily denoted by (x, y, z)∧; it is known that the moment sequence {γn}∞n=0 (with
γ0 = 1) for Wα satisfies a recursion.
n−1 = RR+
0 ··· α2
Definition 3.1. The Berger measure of Wα will be called the Berger measure
associated with the Stampfli completion (x, y, z)∧ of (x, y, z) and denoted by ξx,y,z.
One may see [6, 7, 8] for an alternative approach to these same results. We
note also that one may consider the cases 0 < x < y = z (yielding a 2-atomic
measure with an atom at zero) and 0 < x = y = z, and this last case yields a
completion whose Berger measure is 1-atomic.
Our technique will be to try to choose yi and zi, which will become respectively
λi,3 and λi,4 of the completion S λ, in such a way that the Berger measures associated
λ
to (xi, yi, zi)∧ will become the µ
i,1 and have good properties (and of course they
will automatically be 2-atomic). Figure 3 summarizes the task and our notation.
We begin by calculating the first two "negative" moments of the Berger measure
associated with the Stampfli completion (x, y, z)∧ of (x, y, z).
A SUBNORMAL COMPLETION PROBLEM, II
7
•
λ1,1
λ2,1
λ0
•
☛☛☛☛
✠✠✠✠✠✠✠
•
•
✹✹✹✹✹✹✹✹
·
·
✶✶✶✶
·
•
λn,1
λ1,2
•
λ2,2
•
·
·
·
•
λn,2
Figure 3. An illustration of a 2-generation SCP on Tη,1
with λ0, {λi,1}i∈Jη and {λi,2}i∈Jη as given data.
(3.1)
(3.2)
Lemma 3.2. Suppose that (x, y, z) ∈ R are such that 0 < x < y < z. Then
1
s
ZR+
dξx,y,z(s) =
x4 − 2x2y2 + y2z2
x2y2(z2 − y2)
=
1
x2 " 1 − 2 y2
y2
x2 ( z2
z2
x2
x2 + y2
x2 − y2
x2 ) #
x2
and
1
s2
ZR+
dξx,y,z(s) = −x6 + y2z4 + x4(y2 + 2z2) + x2(y4 − 4y2z2)
# .
x4 "−1 + y2
x4 − 4y2
x2 · z2
x2 + y4
=
x2
1
x2 · z4
x4(y3 − yz2)2
x4 + y2
x2 + 2z2
x2(cid:1)2
x2 − y2
x2(cid:0) z2
ξ = ρδs0 + (1 − ρ)δs1 ,
y2
Proof. To simplify notation, set ξ = ξx,y,z. By [7, Example 3.14], we have
where ρ = s1−x2
2 (ψ1 −pψ2
s1−s0
−y2
ψ0 = −x2y2 z2
y2−x2 and ψ1 = y2 z2
, s0 = 1
1 + 4ψ0) and s1 = 1
1 + 4ψ0) with
−x2
y2−x2 . Straightforward computations now yield
2 (ψ1 +pψ2
(3.1) and (3.2).
(cid:3)
Next we investigate the function f which comes from the expression appearing
on the right-hand side of the second equality in (3.1)
Lemma 3.3. Let f be a real function on Ω = {(u, v) ∈ R2 : v > u > 1} given by
(u, v) ∈ Ω.
Then f (Ω) = (1,∞) and for any (u, v) ∈ Ω and r ∈ (1,∞),
f (u, v) =
,
1 − 2u + uv
u(v − u)
f (u, v) = r
if and only if
v =
.
(3.3)
1 − 2u + ru2
(r − 1)u
Further, for any u ∈ (1,∞), the map ϕu : (1,∞) → (u,∞) defined by
ϕu(r) =
is a bijection.
1 − 2u + ru2
(r − 1)u
,
r ∈ (1,∞),
(3.4)
/
/
D
D
/
/
/
/
/
/
/
/
8
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
Proof. It is a routine matter to verify that for any u ∈ (1,∞), the function
ϕu is a well-defined bijection from (1,∞) to (u,∞). Clearly, the function f is well
defined. Since (u, ϕu(r)) ∈ Ω and f (u, ϕu(r)) = r for all r ∈ (1,∞) and u ∈ (1,∞),
we deduce that (1,∞) ⊆ f (Ω). Next, a simple argument shows that f (Ω) ⊆ (1,∞),
so f (Ω) = (1,∞). It is a computation to show that (3.3) holds.
(cid:3)
Corollary 3.4 below provides more information on the behaviour of the first
"negative" moment of the Berger measure appearing in Lemma 3.2.
Corollary 3.4. Let x, y ∈ R be such that 0 < x < y. Then for any z ∈ (y,∞),
there exists a unique r ∈ (1,∞) such that
1
s
ZR+
dξx,y,z(s) = r
1
x2
.
(3.5)
Conversely, for any r ∈ (1,∞), there exists a unique z ∈ (y,∞) such that (3.5)
holds; the number z is determined by the formula v = ϕu(r) with u = y2
x2 and
v = z2
x2 .
Proof. Making the substitutions u = y2
x2 and v = z2
x2 , we see that (u, v) ∈ Ω
and
1
x2 " 1 − 2 y2
x2 ( z2
x2 + y2
x2 − y2
y2
x2
z2
x2
x2 ) # =
1
x2
f (u, v).
Now applying Lemmata 3.2 and 3.3 completes the proof.
(cid:3)
The expression on the right-hand side of the second equality in (3.2) leads to the
function g which appears in a lemma below. The proof of this lemma follows from
straightforward computations via Lemma 3.3. The details are left to the reader.
Lemma 3.5. Let Ω be as in Lemma 3.3 and g be a real function on Ω given by
g(u, v) = −1 + u · v2 + u + 2v + u2 − 4u · v
,
(u, v) ∈ Ω.
For r ∈ (1,∞), let hr be a real function on (1,∞) defined by 1
u(v − u)2
hr(u) = g(cid:0)u, ϕu(r)(cid:1),
u ∈ (1,∞),
,
u ∈ (1,∞),
where ϕu is as in (3.4). Then the following statements hold for each r ∈ (1,∞):
u−1
(i) hr(u) = 1−2r+r2u
(ii) limu→1+ hr(u) = ∞,
(iii) limu→∞ hr(u) = r2,
(iv) hr(cid:0)(1,∞)(cid:1) = (r2,∞) and hr : (1,∞) → (r2,∞) is a bijection.
Putting together the last three lemmas, we obtain the following crucial lemma.
Lemma 3.6. For any (x, y, z) ∈ R3 such that 0 < x < y < z, there exist
r, ϑ ∈ (1,∞) such that
1
s
dξx,y,z(s) = r
1
x2
,
1
s2
dξx,y,z(s) = ϑ r2 1
x4
.
ZR+
ZR+
(3.6)
(3.7)
1 In view of Lemma 3.3, the definition of hr is correct.
A SUBNORMAL COMPLETION PROBLEM, II
9
Moreover, for any x ∈ (0,∞) and any r, ϑ ∈ (1,∞), there exists (y, z) ∈ R2 with
x < y < z satisfying (3.6) and (3.7).
Proof. Assume (x, y, z) ∈ R3 is such that 0 < x < y < z. Set u = y2
x2 and
x2 . Then (u, v) ∈ Ω. By Lemma 3.3, r := f (u, v) ∈ (1,∞) and v = ϕu(r), so
v = z2
1
s
ZR+
dξx,y,z(s)
(3.1)
=
f (u, v) = r
1
x2
1
x2
,
which gives (3.6). In turn, by Lemma 3.5(iv), hr(u) ∈ (r2,∞) and
1
s2
ZR+
dξx,y,z(s)
(3.2)
=
1
x4
g(u, v) =
hr(u),
1
x4
which implies that (3.7) holds for some ϑ ∈ (1,∞).
Suppose now that x ∈ (0,∞) and r, ϑ ∈ (1,∞). Since ϑ r2 ∈ (r2,∞), we
infer from Lemma 3.5(iv) that there exists u ∈ (1,∞) such that hr(u) = ϑ r2. Set
v = ϕu(r). Then (u, v) ∈ Ω and
(3.8)
Setting y = x√u and z = x√v, we see that 0 < x < y < z and
= ϑ r2 1
x4
g(cid:0)u, v(cid:1) = hr(u) = ϑ r2.
g(cid:0)u, v(cid:1) (3.8)
which gives (3.7). Since v = ϕu(r), we have
ZR+
dξx,y,z(s)
(3.2)
=
1
x4
1
s2
,
1
s
ZR+
dξx,y,z(s)
(3.1)
=
1
x2
f (u, v) =
1
x2
f (u, ϕu(r))
(3.3)
= r
1
x2
,
which implies (3.6). This completes the proof.
(cid:3)
4. The 2-generation SCP on Tη,1
We begin by solving the 2-generation subnormal completion problem on Tη,1
with 2-atomic measures (the case η = ∞ is included).
{λi,2}η
Theorem 4.1. Suppose η ∈ ¯N2, κ = 1, p = 2 and λ = {λ0} ∪ {λi,1}η
i=1 ⊆ (0,∞) are given. Then the following statements are equivalent:
(i) λ has a subnormal completion S λ on Tη,1 such that each measure µ
i=1 ∪
λ
i,1 is
2-atomic,
(ii) there exist sequences {ri}η
η
η
ri
= 1,
sup
i∈Jη
i=1,{ϑi}η
i=1 ⊆ (1,∞) such that
ϑiri − 1
λ2
i,2 < ∞,
(ϑi − 1)ri
λ2
i,1
Xi=1
λ2
i,2
Xi=1
i,1({0}) = 0 for all i ∈ Jη and
Xi=1
λ2
i,1 < sup
i∈Jη
Xi=1
λ2
i,1
λ2
i,2
λ2
i,1
λ4
i,2
1
λ2
0
1
λ2
0
ϑir2
i
< 1,
<
,
6
.
λ
η
η
(4.1)
(4.2)
(4.3)
λ2
i,2.
(4.4)
Moreover, if (i) holds, then µ
λ2
0 <
λ2
i,1,
η
Xi=1
λ2
i,1
λ4
i,2
η
Xi=1
10
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
simplifies. In particular, the statement (4.1) can be dropped.
Proof. We concentrate on proving the case when η = ∞. If η < ∞, the proof
∞,1 such that each measure
(i)⇒(ii) Let S λ be a subnormal completion of λ on T
λ
i,1 is 2-atomic. It follows from [15, Corollary 6.2.2(ii)] that
µi := µ
∞
Xi=1
Xi=1
∞
λ2
i,1ZR+
i,1ZR+
λ2
1
s
dµi(s) = 1,
1
s2
dµi(s) 6
1
λ2
0
.
(4.5)
(4.6)
(4.7)
(4.8)
Applying (2.1) to u = ei,1 and using the Berger-Gellar-Wallen theorem (see [11,
13]), we deduce that for every i ∈ N, the unilateral weighted shift Wα(i) with
weights α(i) = {λi,n+2}∞n=0 is subnormal and µi is the Berger measure of Wα(i) .
It follows from (4.5) that µi cannot have an atom at zero. Hence each µi has two
atoms in (0,∞). As a consequence, λi,2 < λi,3 < λi,4 and µi = ξλi,2,λi,3,λi,4
for every
i ∈ N (see [10, Lemma 2.3] and [7, Example 3.14, Theorem 3.9(iii)], respectively).
Applying Lemma 3.6 to x = xi := λi,2, y = yi := λi,3 and z = zi := λi,4, we deduce
that there exist sequences {ri}∞i=1 ⊆ (1,∞) and {ϑi}∞i=1 ⊆ (1,∞) such that
1
s
ZR+
ZR+
dµi(s) = ri
1
λ2
i,2
,
i ∈ N,
1
s2
dµi(s) = ϑi r2
i
1
λ4
i,2
,
i ∈ N.
According to the proof of Lemma 3.6, the sequences {ri}∞i=1 and {ϑi}∞i=1 are con-
structed via the following process:
yi = xi√ui, zi = xi√vi, vi = ϕui (ri) and hri(ui) = ϑir2
where 0 < xi < yi < zi and 1 < ui < vi.
i for i ∈ N,
(4.9)
(Recall that by Lemmata 3.3 and 3.5, the functions ϕui : (1,∞) → (ui,∞) and
hri : (1,∞) → (r2
i ,∞) are bijections.) Combining (4.5) with (4.7), we obtain (4.2).
In turn, (4.6) and (4.8) imply (4.3). To get (ii), it remains to prove (4.1).
For this, we show that under the circumstances of (4.9) the following assertion
holds:
sup supp ξxi,yi,zi < ∞ ⇐⇒ sup
i∈N
Indeed, according to [7, Example 3.14], we have
Ξ := sup
i∈N
x2
i
ϑiri − 1
(ϑi − 1)ri
< ∞.
(4.10)
sup supp ξxi,yi,zi =
1
2(cid:18)ψ1,i +qψ2
1,i + 4ψ0,i(cid:19),
i ∈ N,
(4.11)
where ψ0,i = −x2
i y2
i
i
z2
i −y2
y2
i −x2
i
for i ∈ N and
ψ1,i = y2
i
i
i − x2
z2
i − x2
y2
i
(4.9)
= x2
i ui
vi − 1
ui − 1
,
i ∈ N.
(4.12)
Since ψ0,i < 0 and ψ2
1,i + 4ψ0,i > 0 for every i ∈ N, we deduce from (4.11) that
Ξ < ∞ ⇐⇒ sup
i∈N
ψ1,i < ∞.
(4.13)
A SUBNORMAL COMPLETION PROBLEM, II
11
By (4.9), hri (ui) = ϑir2
i for all i ∈ N, so Lemma 3.5(i) leads to
ui =
1 − 2ri + ϑir2
(ϑi − 1)r2
i
i
,
i ∈ N.
Using the identity vi = ϕui (ri) and (3.4), we get
(ui − 1)(riui − 1)
vi − 1 =
(ri − 1)ui
Combining (4.12), (4.14) and (4.15), we see that
,
i ∈ N.
ψ1,i =
x2
i
ri − 1
(riui − 1) = x2
i
ϑiri − 1
(ϑi − 1)ri
,
i ∈ N.
This together with (4.13) yields (4.10). It follows from [15, Eg. (6.3.9)] that Ξ < ∞.
Hence, by (4.10), (4.1) is satisfied, which shows that (ii) holds.
(ii)⇒(i) Suppose now that (ii) holds. Applying the "moreover" part of Lemma 3.6
to x = xi := λi,2, r = ri and ϑ = ϑi, we deduce that for every i ∈ N, there exists
(yi, zi) ∈ R2 with xi < yi < zi such that (4.7) and (4.8) hold with µi = ξxi,yi,zi.
According to the proof of Lemma 3.6, the sequences {yi}∞i=1 and {zi}∞i=1 are con-
structed via the procedure (4.9). It follows from (4.1) and (4.10) that
(4.14)
(4.15)
(4.16)
(4.17)
(4.18)
(4.19)
sup
i∈N
sup supp µi < ∞.
sdξx,y,z(s) = x2 whenever 0 < x < y < z, we get
ZR+
sdµi(s) = λ2
i,2,
i ∈ N.
Since in general RR+
It follows that
∞
λ2
i,1ZR+
Xi=1
i,1ZR+
Xi=1
1
s2
λ2
∞
1
s
dµi(s)
(4.7)
=
dµi(s)
(4.8)
=
ϑir2
i
ri
λ2
i,1
λ2
i,2
λ2
i,1
λ4
i,2
(4.2)
= 1,
(4.3)
6
1
λ2
0
.
∞
Xi=1
Xi=1
∞
Putting together the conditions (4.16), (4.17), (4.18) and (4.19), and applying
∞,1 such
Lemma 2.1, we conclude that λ has a subnormal completion S λ on T
λ
that µ
i,1 is 2-atomic,
which gives (i).
i,1 = µi = ξxi,yi,zi for every i ∈ N. As a consequence, each µ
λ
λ
Now we prove the "moreover" part. Suppose (i) is satisfied. The fact that
i,1({0}) = 0 for every i ∈ N is a direct consequence of [10, Theorem 3.5(i)]. Notice
µ
that the second inequality in (4.4) follows from (4.3) while the third and the fourth
can be deduced from (4.2). Thus, it remains to prove the first inequality in (4.4).
It follows from [10, Theorem 4.9(i)] that λ2
i,1. Suppose, to the contrary,
that λ2
i,1. Then, by the "moreover" part of [10, Propositon 7.6], we see
λ
that µ
i,1, which contradicts our assumption
(cid:3)
0 =P∞i=1 λ2
i,1 = δc for every i ∈ N with c =P∞i=1 λ2
λ
i,1 is 2-atomic. This completes the proof.
0 6P∞i=1 λ2
that each µ
Now we are ready to solve the 2-generation subnormal completion problem on
Tη,1 for η ∈ N2 without any restrictions on the supports of the resulting mea-
sures µ
λ
i,1.
12
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
{λi,2}η
Theorem 4.2. Suppose η ∈ N2, κ = 1, p = 2 and λ = {λ0} ∪ {λi,1}η
i=1 ⊆ (0,∞) are given. Then the following statements are equivalent:
(i) λ has a subnormal completion S λ on Tη,1,
(ii) there exists a sequence {ri}η
i=1 ⊆ [1,∞) such that
Xi=1
λ2
i,1
λ2
i,2
= 1,
ri
η
r2
i
η
Xi=1
6 1
λ2
0
< 1
λ2
0
λ2
i,1
λ4
i,2
if ri = 1 for every i ∈ Jη,
if ri > 1 for some i ∈ Jη.
Moreover, if S λ is a subnormal completion of λ on Tη,1 such that each measure µ
is 1- or 2-atomic, then (4.20) and (4.21) hold for some {ri}η
ri = 1 (resp. ri > 1) if and only if µ
λ
i,1 is 1-atomic (resp. 2-atomic).
i=1 ⊆ [1,∞) such that
(4.22)
Conversely, if (ii) holds, then λ admits a subnormal completion S λ on Tη,1 which
satisfies (4.22).
Proof. In view of Theorem 2.2, the statement (i) is equivalent to the fact
λ
that λ admits a subnormal completion S λ on Tη,1 such that each measure µ
i,1 is
1- or 2-atomic. Hence, in the proof of the implication (i)⇒(ii), we can define the
partition (A, B) of Jη as follows
λ
λ
A =(cid:8)i ∈ Jη : µ
i,1 is 1-atomic(cid:9) and B =(cid:8)i ∈ Jη : µ
λ
i,1 is 2-atomic(cid:9).
i,1, i ∈ Jη, satisfy (2.2)-(2.4) with
By [15, Corollary 6.2.2(ii)] the measures µi = µ
κ = 1 and p = 2.
for every
i ∈ Jη and so (4.20) and (4.21) hold with ri = 1 for all i ∈ Jη. If B 6= ∅, then
by combining reasonings used above and in the proof of the implication (i)⇒(ii) of
Theorem 4.1, we get two sequences {ri}i∈B,{ϑi}i∈B ⊆ (1,∞) such that
If B = ∅, then it is easily seen that µ
λ
i,1 = δλ2
i,2
i=1 ∪
(4.20)
(4.21)
λ
i,1
λ2
i,1
λ2
i,2
Xi∈A
Xi∈A
λ2
i,1
λ4
i,2
+Xi∈B
+Xi∈B
ri
λ2
i,1
λ2
i,2
λ2
i,1
λ4
i,2
ϑir2
i
= 1,
6
1
λ2
0
.
(4.23)
(4.24)
Since ϑi > 1 for any i ∈ B, we see that (4.23) and (4.24) imply (4.20) and (4.21)
with ri = 1 for i ∈ A. Putting all of this together yields (ii) and (4.22).
To prove the converse implication (ii)⇒(i), we will define a family {µi}i∈Jη
of Borel probability measures on R+ satisfying (2.2)-(2.5) with κ = 1 and p = 2.
First, we define the partition (A′, B′) of Jη by
A′ = {i ∈ Jη : ri = 1} and B′ = {i ∈ Jη : ri > 1}.
If B′ = ∅, then the probability measures µi = δλ2
B′ 6= ∅, then we proceed as follows. By (4.21) there exists τ ∈ (1,∞) such that
, i ∈ Jη, satisfy (2.2)-(2.5). If
i,2
λ2
i,1
λ4
i,2
Xi∈A′
r2
i
+ τ Xi∈B ′
λ2
i,1
λ4
i,2
<
1
λ2
0
.
(4.25)
A SUBNORMAL COMPLETION PROBLEM, II
13
η
λ2
If i ∈ A′, then we set µi = δλ2
. If i ∈ B′, then we define µi as in the proof of the
implication (ii)⇒(i) of Theorem 4.1 with ϑi = τ . As in that proof, we verify that
i,2
sdµi(s) = λ2
ZR+
i,1ZR+
i,1ZR+
Xi=1
Xi=1
i ∈ Jη,
+ Xi∈B ′
+ Xi∈B ′
Hence (2.2)-(2.5) hold. Now, by applying Lemma 2.1, we conclude that λ has a
subnormal completion S λ on Tη,1 such that µ
i,1 = µi for every i ∈ Jη, which
i,1 : i ∈ A′} are 1-atomic while the measures
gives (i). Clearly, the measures {µ
i,1 : i ∈ B′} are 2-atomic. This completes the proof of the "moreover" part and
{µ
the proof of the theorem.
dµi(s) = Xi∈A′
dµi(s) = Xi∈A′
(4.20)
= 1,
λ2
i,1
λ2
i,2
λ2
i,1
λ4
i,2
λ2
i,1
λ4
i,2
1
λ2
0
.
ri
λ2
i,1
λ2
i,2
1
s
1
s2
η
λ2
ϑir2
i
(4.25)
<
(cid:3)
i,2,
λ
λ
λ
It follows from Theorem 4.2 that the 2-generation subnormal completion prob-
lem on Tη,1 may have a solution admitting simultaneously 1- and 2-atomic mea-
sures. Below we give an example showing that the problem on T2,1 may have a
solution of this kind with distinct atoms.
•
1
λ0
•
☛☛☛☛
✠✠✠✠✠✠✠
✹✹✹✹✹✹✹✹
•
✶✶✶✶
•
1
√2
•
•
√3
Figure 4. Mixed completions discussed in Example 4.3.
Example 4.3. Consider the 2-generation subnormal completion problem on
i=1 ⊆ (0,∞), where λ0 ∈
T2,1 with the initial data λ = {λ0} ∪ {λi,1}2
(0,∞) is arbitrary and the remaining weights are given by (see Figure 4)
i=1 ∪ {λi,2}2
λ1,1 = λ2,1 = 1, λ1,2 = √2 and λ2,2 = √3.
λ
1,1 is 1-atomic while µ
We are looking for λ0 for which λ admits a subnormal completion S λ on T2,1 such
λ
that µ
2,1 is 2-atomic, and all these atoms are different. In
view of [10, Theorem 4.9], any subnormal completion S λ of the above λ on T2,1
comes from compactly supported Borel probability measures µ1 and µ2 on R+
which satisfy the following conditions
sdµi(s) = λ2
i,2 =(2
3
ZR+
if i = 1,
if i = 2,
(4.26)
/
/
D
D
/
/
/
/
14
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
ZR+
ZR+
1
s2
1
s
dµ1(s) +ZR+
dµ1(s) +ZR+
1
s
1
s2
dµ2(s) = 1,
dµ2(s) 6
1
λ2
0
.
(4.27)
(4.28)
Let µ1 and µ2 be Borel probability measures on R+ given by
µ1 = δx
and µ2 =
δy +
δz.
1
2
1
2
It is a matter of routine to verify that µ1 and µ2 with x = 2, y = 3 − √3 and
z = 3 + √3 satisfy (4.26) and (4.27). Hence (4.28) holds if and only if λ2
0 6 12
7 ,
meaning that for such λ0's the corresponding λ admits a subnormal completion on
T2,1 with the desired properties.
If η ∈ N2, Theorem 4.1 takes a much simpler form which is a direct consequence
of Theorem 4.2.
{λi,2}η
Proposition 4.4. Suppose η ∈ N2, κ = 1, p = 2 and λ = {λ0} ∪ {λi,1}η
i=1 ⊆ (0,∞) are given. Then the following statements are equivalent:
(i) λ has a subnormal completion S λ on Tη,1 such that each measure µ
i=1 ∪
λ
i,1 is
2-atomic,
η
(ii) there exists a sequence {ri}η
i=1 ⊆ (1,∞) such that
Xi=1
Xi=1
λ2
i,1
λ4
i,2
λ2
i,1
λ2
i,2
1
λ2
0
= 1,
r2
i
ri
<
η
,
, where2
λ2
0
(iii) −∞ < β(η) < 1
β(η) := inf( η
Xi=1
r2
i
λ2
i,1
λ4
i,2
: {ri}η
i=1 ⊆ (1,∞) and
ri
λ2
i,1
λ2
i,2
= 1) .
η
Xi=1
(4.29)
(4.30)
(4.31)
λ
i,1 is 2-atomic, then there exists i ∈ Jη such that Pη
Below we discuss the 2-generation SCP on Tη,1 with 2-atomic measures under
some constraints. Let us suppose temporarily that η ∈ N2. According to the
"moreover" part of Theorem 4.1, if S λ is a subnormal completion of λ on Tη,1 such
that each measure µ
j,1 <
λ2
i,2 (use the fourth inequality in (4.4)). If the last inequality holds for all i ∈ Jη,
then the solution of the 2-generation subnormal completion problem on Tη,1 with
2-atomic measures takes a simple form. Namely, the first inequality in (4.4), which
is a necessary condition for solving the 2-generation subnormal completion problem
on Tη,1, becomes now sufficient (see Corollary 4.5(ii) below). It is worth pointing
out that in general the inequality λ2
i,1 holds whenever the p-generation
subnormal completion problem on Tη,κ has a solution (see [10, Theorem 4.9(i)]).
j=1 λ2
i=1 λ2
0 6Pη
2 We adhere to the convention that inf ∅ = −∞; in particular, −∞ < β(η) means that the
set in (4.31) is nonempty.
A SUBNORMAL COMPLETION PROBLEM, II
15
{λi,2}η
Corollary 4.5. Suppose η ∈ N2, κ = 1, p = 2 and λ = {λ0} ∪ {λi,1}η
i=1 ⊆ (0,∞) are given. Assume that
λ2
j,1 < λ2
i,2,
η
i=1 ∪
(4.32)
i ∈ Jη.
Xj=1
Then the following statements are equivalent:
(i) λ has a subnormal completion S λ on Tη,1 such that each measure µ
λ
i,1 is
2-atomic,
i,1.
i=1 λ2
0 <Pη
(ii) λ2
Proof. (i)⇒(ii) This is a direct consequence of the first inequality in (4.4).
(ii)⇒(i) Define {ri}η
i=1 ⊆ (1,∞) by
λ2
i,2
Pη
j=1 λ2
j,1
ri =
,
i ∈ Jη.
(4.33)
i=1 the equation (4.29) holds. Since
It is easily seen that with this choice of {ri}η
1
Pη
Xi=1
λ2
i,1
λ4
i,2
(4.33)
r2
i
=
η
(ii)
<
1
λ2
0
,
i=1 λ2
i,1
we obtain (4.30). Applying Proposition 4.4 completes the proof.
(cid:3)
Remark 4.6. In view of Proposition 4.4(iii), if η ∈ N2 one may create many
sufficient conditions for solving the 2-generation subnormal completion problem on
Tη,1 with 2-atomic measures. The procedure goes as follows. First, we look for a
sequence {ri}i∈Jη ⊆ (1,∞) satisfying the equation (4.29). Second, we substitute
this particular choice of {ri}i∈Jη into the expressionPη
appearing in (4.30).
Finally, we require the value so obtained to be strictly less than 1
λ2
0
i=1 r2
i
λ2
λ4
i,1
i,2
.
It is worth mentioning that the procedure described in Remark 4.6 has been
applied in the proof of Corollary 4.5 (see (4.33)). Below we give a few more examples
illustrating this procedure.
{λi,2}η
that each measure µ
Corollary 4.7. Suppose η ∈ N2, κ = 1, p = 2 and λ = {λ0} ∪ {λi,1}η
i=1 ∪
i=1 ⊆ (0,∞) are given. Then λ has a subnormal completion S λ on Tη,1 such
λ
i,1 is 2-atomic provided any of the following conditions holds:
(i) Pη
i=1
i,1 < λ2
λ2
λ2
1
λ2
(ii) ηλ2
< η2,
λ2
λ2
λ2
λ4
i=1
i=1
i,2
i,1
,
i,1
0 Pη
< 1 and λ2
i,2 for each i ∈ Jη and λ2
<(cid:16)Pη
0 Pη
< 1 for each i ∈ Jη and λ2
1
λ2
j,2
i=1
i,2
1
λ2
i,1λ4
i,2
(iii) λ2
j=1
i,1Pη
Proof. Apply Remark 4.6 to
<(cid:16)Pη
i=1
.
1
λ2
i,2(cid:17)2
i,1
i,2(cid:17)2
0Pη
i=1
i,1
1
,
ri =
1
η
λ2
i,2
λ2
i,1
and
ri =
λ2
λ2
j,1
j,2
ri =
j=1
Pη
in the cases (i), (ii) and (iii), respectively.
1
j=1
λ2
i,1Pη
1
λ2
j,2
(cid:3)
The above discussion can be applied to solve the 1-generation subnormal com-
pletion problem on Tη,1 with 2-atomic measures.
16
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
Corollary 4.8. Suppose η ∈ N2, κ = 1, p = 1 and λ = {λ0} ∪ {λi,1}η
(0,∞) are given. Then the following statements are equivalent:
(i) λ has a subnormal completion S λ on Tη,1 such that each measure µ
i=1 ⊆
λ
i,1 is
2-atomic,
i,1.
i=1 λ2
0 <Pη
(ii) λ2
Proof. (i)⇒(ii) Let S λ be a subnormal completion of λ on Tη,1 such that
λ
i,1 is 2-atomic. Then clearly this S λ is a subnormal completion of
each measure µ
λ ∪ {λi,2}i∈Jη on Tη,1. Hence, by the condition (4.4), (ii) is valid.
i=1 ⊆ (0,∞) that satisfies (4.32). By Corol-
lary 4.5, λ ∪ {λi,2}i∈Jη has a subnormal completion S λ on Tη,1 such that each
λ
i,1 is 2-atomic. As a consequence, this S λ is a subnormal completion of
measure µ
λ on Tη,1 as well. This completes the proof.
(cid:3)
(ii)⇒(i) Choose any sequence {λi,2}η
In concluding this section, we consider the 2-generation subnormal completion
problem on Tη,1 from the point of view of the condition (4.32).
Remark 4.9. Suppose that η ∈ N2. Let us consider the situation in which we
i=1 ⊆ (0,∞) on Tη,1,2 satisfying
are given initial data λ = {λ0} ∪ {λi,1}η
the equations
i=1 ∪ {λi,2}η
λi,2 = λ1,2,
i ∈ Jη.
Notice that under the above assumption, if λ admits a subnormal completion on
Tη,1, then it admits a 2-generation flat subnormal completion on Tη,1 (see [10,
Theorem 8.3]), and any 2-generation flat subnormal completion S λ of λ on Tη,1
has the property that µ
1,1 for all i ∈ Jη (apply (2.1) to u = ei,1). Below, we
discuss a few cases related to the condition (4.32).
λ
i,1 = µ
λ
1◦ If λ2
i=1 λ2
i,1, then there is no completion of the sequence
1,2 <Pη
η
λ0,vuut
α =
Xi=1
λ2
i,1, λ1,2
i=1 λ2
1,2 = Pη
to a subnormal unilateral weighted shift, because it is well known that a subnormal
unilateral weighted shift is hyponormal and so its weights must be monotonically
increasing. Therefore, by [10, Theorem 8.3(iii)] there is no subnormal completion
of λ on Tη,1.
2◦ If λ2
i,1, then the "moreover" part of Theorem 4.1 shows that
λ
there is no subnormal completion S λ of λ on Tη,1 with 2-atomic measures µ
i,1 since
we lack the third inequality in (4.4). Alternatively, we may use the "moreover" part
of [10, Proposition 7.6].
i=1 λ2
i,1 we may apply Corollary 4.5 to deduce
λ
i,1.
i,1 we may apply the condition (4.4) of
Theorem 4.1 to deduce that there is no subnormal completion S λ of λ on Tη,1 with
λ
i,1 2-atomic. Alternatively, again using the "moreover" part of [10, Proposi-
the µ
λ
i,1 for any subnormal completion S λ
that there is a subnormal completion S λ of λ on Tη,1 with 2-atomic measures µ
1,2 >Pη
1,2 >Pη
0 <Pη
0 =Pη
tion 7.6] we may deduce that the measures µ
of λ on Tη,1 (provided it exists) must be 1-atomic.
i,1 and λ2
i,1 and λ2
4◦ If λ2
3◦ If λ2
i=1 λ2
i=1 λ2
i=1 λ2
A SUBNORMAL COMPLETION PROBLEM, II
17
5. An explicit solution of the 2-generation SCP on T2,1
In this section we give necessary and sufficient conditions for solving the 2-
generation subnormal completion problem on T2,1 with 2-atomic measures explic-
itly in terms of the initial data. In view of Proposition 4.4(iii), the solution of the
2-generation subnormal completion problem on Tη,1 with 2-atomic measures re-
duces to computing the infimum β(η) of the quadratic formPη
the constraints that {ri}η
= 1. This task is extremely
complicated. The main difficulty comes from the requirement that the numbers
ri should belong to the open interval (1,∞). Even in the case of η = 2, there
are three different formulas for the infimum depending on weights in question (see
Theorem 5.2 below). However under some additional restrictive assumptions, the
infimum of the above quadratic form can be computed explicitly.
i=1 ⊆ (1,∞) and Pη
subject to
i=1 r2
i
i=1 ri
λ2
λ2
i,1
i,2
i,2
i,1
λ2
λ4
i=1 ⊆ (0,∞) and {bi}η
airi = 1(cid:27) =
η
i=1 ⊆ (0,∞). Then
.
a2
i
bi
(5.1)
1
i=1
Pη
η
bir2
i : {ri}η
i=1 ⊆ (0,∞) and
Proposition 5.1. Let η ∈ N2, {ai}η
min(cid:26)
Xi=1
Moreover, if Pη
min(cid:26)
Xi=1
Xi=1
for every i ∈ Jη, then
Xi=1
i=1 ⊆ (1,∞) and
i : {ri}η
< ai
bi
bir2
a2
j
bj
j=1
η
η
In both cases the minimum is attained for
airi = 1(cid:27) =
.
a2
i
bi
1
i=1
Pη
(5.2)
(5.3)
It follows from the
,
a2
j
bj
i ∈ Jη.
ri =
1
j=1
bi
ai Pη
i=1 ⊆ (0,∞) and Pη
ai√bi ·pbi ri 6(cid:18)
Xi=1
Xi=1
η
η
i=1 airi = 1.
a2
i
bi(cid:19)1/2(cid:18)
η
Xi=1
bir2
i(cid:19)1/2
.
Proof. Suppose {ri}η
Cauchy-Schwarz inequality that
1 =
η
Xi=1
airi =
Therefore, we have
bir2
i >
η
Xi=1
1
i=1
Pη
.
a2
i
bi
(5.4)
Substituting {ri}η
This proves (5.1). The equation (5.2) is a direct consequence of (5.1).
i=1 as in (5.3) shows that the inequality in (5.4) becomes equality.
(cid:3)
As shown below, a solution of the 2-generation subnormal completion problem
on T2,1 with 2-atomic measures can be written entirely in terms of the initial data.
This is done by computing the infimum β(2).
Theorem 5.2. Suppose η = 2, κ = 1, p = 2 and λ = {λ0} ∪ {λi,1}2
i=1 ∪
i=1 ⊆ (0,∞) are given. Then λ has a subnormal completion S λ on T2,1 with
{λi,2}2
18
G. R. EXNER, I. B. JUNG, J. STOCHEL, AND H. Y. YUN
2-atomic measures µ
λ
1,1 and µ
λ
2,1 if and only if β(2) < 1
λ2
0
and 1 < τ , where
β(2) =
with σ = a2
a1+b1
, τ = a2(b2−b1)
a1b2
(a1−a2)2+a1b1
a2
2b1
1
a1+b1
(b2−b1)2+a1b1
a1b2
2
if σ 6 1,
if 1 < σ < τ,
if τ 6 σ,
, aj = λ2
1,j and bj = λ2
2,j for j = 1, 2.
Proof. In view of Proposition 4.4, it is enough to compute β(2). If we replace
r1 and r2 by x and y, respectively, then (4.31) with η = 2 takes the form
β(2) = infnΘ(x, y) : x, y ∈ (1,∞) and φ(x, y) = 1o,
where
Θ(x, y) = x2 a1
a2
2
+ y2 b1
b2
2
and φ(x, y) = x
a1
a2
+ y
b1
b2
for x, y ∈ (1,∞).
(5.5)
If x, y ∈ (1,∞) are such that φ(x, y) = 1, then
a1
a2
y =
b2
b1(cid:16)1 −
x(cid:17) > 1,
so 1 < x < τ . As a consequence, we see that −∞ < β(2) if and only if 1 < τ .
Substituting y as in (5.5) into Θ(x, y), we obtain
Θ(x, y) = Ax2 − 2Bx + C,
1 < x < τ,
where
A =
a1
a2
2 (cid:18)1 +
a1
b1(cid:19), B =
a1
a2b1
and C =
1
b1
.
Note that A, B, C > 0. The quadratic polynomial Ω(x) := Ax2 − 2Bx + C regarded
as a function on R has a minimum at x = B
A . Observe that
B
A
= σ and Ω(σ) =
1
a1 + b1
> 0.
It is now a routine matter to compute β(2) by considering three possible disjoint
cases σ 6 1, 1 < σ < τ and τ 6 σ. What we get is β(2) = Ω(1) in the first case,
β(2) = Ω(σ) in the second and β(2) = Ω(τ ) in the third one, where
Ω(1) =
(a1 − a2)2 + a1b1
a2
2b1
and Ω(τ ) =
(b2 − b1)2 + a1b1
a1b2
2
.
This completes the proof.
(cid:3)
Acknowledgement. The authors take this opportunity to express their appre-
ciation both for the support of their universities Bucknell University, Jagiellonian
University and Kyungpook National University materially aiding this collaboration,
and to the Departments of Mathematics at which they have been guests in 2018
and 2019 for warm hospitality.
A SUBNORMAL COMPLETION PROBLEM, II
19
References
[1] P. Budzy´nski, Z.J. Jab lo´nski, I.B. Jung, J. Stochel, Unbounded subnormal weighted shifts on
directed trees, J. Math. Anal. Appl. 394 (2012), 819-834.
[2] P. Budzy´nski, Z.J. Jab lo´nski, I.B. Jung, J. Stochel, Unbounded subnormal weighted shifts on
directed trees. II, J. Math. Anal. Appl. 398 (2013), 600-608.
[3] P. Budzy´nski, Z.J. Jab lo´nski, I.B. Jung, J. Stochel, Unbounded weighted composition opera-
tors in L2-spaces, Lect. Notes Math., 2209, Springer, 2018.
[4] J.B. Conway, Subnormal operators, Research Notes in Mathematics, 51, Pitman Publ. Co.,
London, 1981.
[5] J.B. Conway, The theory of subnormal operators, Mathematical Surveys and Monographs,
36, American Mathematical Society, Providence, RI, 1991.
[6] R.E. Curto, L.A. Fialkow, Recursiveness, positivity, and truncated moment problems, Hous-
ton J. Math. 17 (1991), 603-635.
[7] R.E. Curto, L.A. Fialkow, Recursively generated weighted shifts and the subnormal comple-
tion problem, Integr. Equ. Oper. Theory 17 (1993), 202-246.
[8] R.E. Curto, L.A. Fialkow, Recursively generated weighted shifts and the subnormal comple-
tion problem II, Integr. Equ. Oper. Theory 18 (1994), 369-426.
[9] R. Curto, L. Fialkow, Solution of the truncated complex moment problem for flat data, Mem.
Amer. Math. Soc. 119 (1996), no. 568, x+52 pp.
[10] G.R. Exner,
I.B. Jung,
lem for weighted shifts on directed trees,
https://doi.org/10.1007/s00020-018-2496-9.
J. Stochel, H.Y. Yun, A subnormal
completion prob-
Integr. Equ. Oper. Theory, 90 (2018);
[11] R. Gellar, L.J. Wallen, Subnormal weighted shifts and the Halmos-Bram criterion, Proc.
Japan Acad. 46 (1970), 375-378.
[12] P.R. Halmos, Normal dilations and extensions of operators, Summa Brasil. Math. 2 (1950),
125-134.
[13] P.R. Halmos, Ten problems in Hilbert space, Bull. Amer. Math. Soc. 76 (1970), 887-933.
[14] Z.J. Jab lo´nski, I.B. Jung, J. Stochel, A non-hyponormal operator generating Stieltjes moment
sequences, J. Funct. Anal. 262 (2012), 3946-3980.
[15] Z.J. Jablonski, I.B. Jung, J. Stochel, Weighted shifts on diected trees, Mem. Amer. Math.
Soc. 216 (2012), viii+106 pp.
[16] Z.J. Jab lo´nski, I.B. Jung, J.A. Kwak, J. Stochel, Hyperexpansive completion problem via
alternating sequences; an application to subnormality, Linear Algebra Appl. 434 (2011),
2497-2526.
[17] O. Ore, Theory of graphs, American Mathematical Society Colloquium Publications, Vol.
XXXVIII, American Mathematical Society, Providence, R.I. 1962.
[18] A.L. Shields, Weighted shift operators and analytic function theory, Topics in operator theory,
pp. 49-128. Math. Surveys, No. 13, Amer. Math. Soc., Providence, R.I., 1974.
[19] J.G. Stampfli, Which weighted shifts are subnormal, Pacific J. Math. 17 (1966), 367-379.
Department of Mathematics, Bucknell University, Lewisburg, Pennsylvania 17837,
USA
E-mail address: [email protected]
Department of Mathematics, Kyungpook National University, Daegu 41566, Korea
E-mail address: [email protected]
Instytut Matematyki, Uniwersytet Jagiello´nski, ul. Lojasiewicza 6, PL-30348 Kra-
k´ow, Poland
E-mail address: [email protected]
Department of Mathematics, Kyungpook National University, Daegu 41566, Korea
E-mail address: [email protected]
|
1503.04968 | 1 | 1503 | 2015-03-17T09:55:41 | Comparison between some norm and order gauge integrals in Banach lattices | [
"math.FA"
] | Henstock-type integrals are considered, for functions defined in a Radon measure space and taking values in a Banach lattice X considering the norm and the order structure of the space.
A number of results are obtained, highlighting the differences between them, specially in the case of L-space-valued functions. | math.FA | math |
Comparison between some norm and order
gauge integrals in Banach lattices
Domenico Candeloro
University of Perugia (Italy)
Department of Mathematics and Computer Sciences
via Vanvitelli, 1 I-06123 Perugia, Italy
[email protected]
Anna Rita Sambucini
University of Perugia (Italy)
Department of Mathematics and Computer Sciences
via Vanvitelli, 1 I-06123 Perugia, Italy
[email protected]
Abstract
Henstock-type integrals are considered, for functions defined in a
Radon measure space and taking values in a Banach lattice X con-
sidering the norm and the order structure of the space. A number of
results are obtained, highlighting the differences between them, spe-
cially in the case of L-space-valued functions.
AMS Subject Classification: 28B05, 28B20, 46B42, 46G10, 18B15
Key Words: Banach lattices, L and M spaces, gauges, Henstock
integral, Mc Shane integral, variational Henstock integral
1
Introduction
In the last century many different notions of integral were introduced
for real-valued functions, in order to generalize the Riemann one. The
wide literature in this topic attests the interest for this field of research:
see for example [1,2,5,6,8,10,13,14,21 -- 27,29 -- 31,34,35,41 -- 43,45]. Af-
terwards the notions of order-type integral were also introduced and
studied, for functions taking their values in suitable ordered vector
spaces: see [4, 6, 9, 11, 12, 18, 19].
This paper aims to compare some norm- and order-type integrals, with
some interesting results when X is an L-space.
2
D. Candeloro and A. R. Sambucini
Throughout this paper (T, d) is a compact metric Hausdorff topolog-
ical space, Σ its Borel σ-algebra and µ : Σ → R+
0 a regular, pointwise
non atomic measure, and X a Banach (lattice) space.
In Sections 2 and 3 norm and order-type integrals are compared, show-
ing a first showy difference: order integrals in general do not respect
almost everywhere equality, except for order-bounded functions. An-
other noteworthy difference is that order integrals enjoy the Henstock
Lemma: this fact has interesting consequences in L-spaces, as shown
in Theorem 3.9.
In the subsection 3.1 integrability in [0, 1] is discussed and it is proven
that monotone mappings are McShane order-integrable. In the same
framework, two further notions of order-type integrals, inspired at the
paper [39], have been introduced: they are the corresponding notions
of the Henstock and McShane order-type integrals, but requiring mea-
surability of the involved gauges: in the normed case, these notions are
discussed in [39] where their strict connection with the Birkhoff inte-
grability has been shown and their equivalence is proven, for bounded
functions; here we prove that, also for the order-type integral, for
bounded functions the two notions are equivalent.
Some of these results have been exposed at the SISY 2014 - IEEE
12th International Symposium on Intelligent Systems and Informat-
ics (see [20]) and they will be applied to the multivalued case in a
subsequent paper [7].
2 Henstock (and McShane) integrals
Given a compact metric Hausdorff topological space (T, d) and its
Borel σ-algebra Σ, let µ : Σ → R+
0 be a σ-additive (bounded) regular
measure, so that (T, d, Σ, µ) is a Radon measure space. Let X be a
Banach space. The following concept of Henstock integrability was
presented in [29] for bounded measures in the Banach space context.
See also [9] for the following definitions and investigations.
A gauge is any map γ : T → R+. A decomposition Π of T is a finite
family Π = {(Ei, ti) : i = 1, . . . , k} of pairs such that ti ∈ Ei, Ei ∈ Σ
and µ(Ei ∩ Ej) = 0 for i 6= j. The points ti, i = 1, . . . , k, are called
tags.
i=1 Ei = T , Π is called a partition. A Perron
partition (or also Henstock partition) is a partition for which each tag
ti belongs to the corresponding set Ei; while, if this condition is not
necessarily fulfilled, the partition is said to be free.
If moreover Sk
Comparison between some norm and order gauge integrals...
3
Given a gauge γ, Π is said to be γ-fine (Π ≺ γ) if dist(w, ti) < γ(ti)
for every w ∈ Ei and i = 1, . . . , k.
Definition 2.1. A function f : T → X is H-integrable if there exists
I ∈ X such that, for every ε > 0 there is a gauge γ : T → R+ such
that for every γ-fine Perron partition of T , Π = {(Ei, ti), i = 1, . . . , q},
one has:
kσ(f, Π) − Ik ≤ ε,
i=1 f (ti)µ(Ei). In this case, the
where the symbol σ(f, Π) means Pq
following notation will be used: I = (H)RT f dµ.
It is not difficult to deduce, in case f is H-integrable in the set T ,
that also the restrictions f 1E are, for every measurable set E, thanks
to the Cousin Lemma (see [41, Proposition 1.7]). This is not true in
the classical theory, where T = [0, 1] and the partitions allowed are
only those consisting of sub-intervals.
Definition 2.2. A function f : T → X is MS-integrable if there exists
I ∈ X such that, for every ε > 0 there is a gauge γ : T → R+ such
that for every γ-fine partition of T , Π = {(Ei, ti), i = 1, . . . , q}, one
has:
kσ(f, Π) − Ik ≤ ε,
i=1 f (ti)µ(Ei). In this case, the
where the symbol σ(f, Π) means Pq
following notation will be used: I = (H)RT f dµ.
As it is well-known, in the classical theory H-integrability is dif-
ferent from McShane integrability and also from Pettis integrability.
Nevertheless, McShane integrability of a mapping f : [0, 1] → X (here
X is any Banach space) is equivalent to Henstock and Pettis simulta-
neous integrabilities to hold (see [29, Theorem 8]).
In the papers [24, 27, 43] one can find an interesting discussion of
the cases in which Henstock, Pettis and McShane integrability are
equivalent, and also counterexamples in some related problems.
Since the measure µ is assumed to be pointwise non atomic, a soon
as partitions consist of arbitrary measurable sets, all concepts in the
Henstock sense will turn out to be equivalent to the same concepts
4
D. Candeloro and A. R. Sambucini
in the McShane sense (i.e. without requiring that the tags are con-
tained in the corresponding sets of the involved partitions); however
the Henstock terminology and notations will remain unchanged. The
next proposition clarifies this fact.
Proposition 2.3. Assume, in the previous setting, that µ is pointwise
non atomic, i.e. µ({t}) = 0 for all t ∈ T . If f : T → X is H-integrable
in T , then it is also Mc Shane-integrable.
Proof. Of course, it will be sufficient to prove that, for every gauge γ
and any γ-fine free partition Π0, there exists a Henstock-type γ-fine
partition Π′ satisfying σ(f, Π0) = σ(f, Π′).
So, fix γ and Π0 := {(Bi, ti) : i = 1, ..., k} as above. Without loss
of generality, all the tags ti can be supposed to be distinct, otherwise
it will be sufficient for each tag to take the union of all the sets Bi
paired with it. Now, set A := {ti, i = 1, ..., k} and define, for each
j: B′
j is measurable and is
contained in γ(tj). Then the pairs (Bj, tj) form a γ-fine Henstock-
type partition Π′ and µ(B′
j) = µ(Bj) for all j, so σ(f, Π0) = σ(f, Π′).
This concludes the proof.
j := (Bj \ A) ∪ {tj}. Of course, each B′
So this Proposition shows that the use of free γ-fine partitions does
not modify the integral.
From now on suppose that X is a Banach lattice with order-
continuous norm, X + is its positive cone and X ++ is the subset of
strictly positive elements of X. The symbols , k k refer to modulus
and norm of X; for the relation between them and applications in
Banach lattices see for example [3, 9, 16, 17, 32, 36, 38].
An element e of X + is an order unit if for every u ∈ X there is an
n ∈ N such that u ≤ ne.
An L-space is a Banach lattice L such that ku + vk = kuk + kvk when-
ever u, v ∈ L+.
An M -space is a Banach lattice M in which the norm is an order-unit
norm, namely there is an order unit e and an equivalent Riesz norm
k · ke defined on M by the formula kuke := min{α : u < αe} for every
u ∈ M . In this case one has also ku + vk = kuk ∨ kvk. For further
properties about L- and M -spaces see also [32, Section 354].
In this setting the notion of order-type integral can be given.
Comparison between some norm and order gauge integrals...
5
Definition 2.4. A function f : T → X is order-integrable in the
Henstock sense ((oH)-integrability for short) if there exists J ∈ X
together with an (o)-sequence (bn)n in X and a corresponding sequence
(γn)n of gauges, such that for every n and every γn-fine Henstock
partition of T , Π = {(Ei, ti), i = 1, . . . , q}, it is σ(f, Π) − J ≤ bn and
the integral J will be denoted with (oH)R f .
Remark 2.5. Also in this case there is no difference in taking all
free γn-fine partitions. It is easy to see that any (oH)-integrable f is
also H-integrable and the two integrals coincide thanks to the order
continuity of the norm in X. The converse implication holds when X
is an M -space, but not in general: see for example Theorem 3.9 and
the following remarks. Moreover the (oH)-integrability of a function
f implies also its Pettis integrability, thanks to [29, Theorem 8].
The following Cauchy-type criterion holds:
Theorem 2.6. ( [20, Theorem 5]) Let f : T → X be any function.
Then f is (oH)-integrable if and only if there exist an (o)-sequence
(bn)n and a corresponding sequence (γn)n of gauges, such that for
every n, as soon as Π, Π′ are two γn-fine Henstock partitions, the
following holds:
σ(f, Π) − σ(f, Π′) ≤ bn.
(1)
3 Properties of the Norm and the Or-
der integrals
Some interesting properties of the order integral can be described com-
paring it with the norm one. A first singular fact is that, in general,
almost equal functions may have a different behavior with respect to
the (oH)-integral, as proven in [9, Example 2.8]. However, for order-
bounded functions, the following result holds.
Proposition 3.1. [20, Proposition 4] Let f, g : T → X be two bounded
maps, such that f = g µ-almost everywhere. Then, f is (oH)-integrable
if and only if g is, and the integrals coincide.
Another interesting difference concerns the Henstock Lemma, which
is valid in the present form for the order integral but not, in gen-
eral, for the norm one. The technique of the proof is similar to that
of [18, Theorem 1.4].
6
D. Candeloro and A. R. Sambucini
Proposition 3.2. [20, Proposition 6] Let f : T → X be any (oH)-
integrable function. Then, there exist an (o)-sequence (bn)n and a
corresponding sequence (γn)n of gauges, such that, for all n and all
γn-fine Henstock partitions Π one has
XE∈Π
sup
E,Π′′
Π′
E
{XF ′′∈Π′′
E
f (τF ′′)µ(F ′′) −XF ′∈Π′
E
f (τF ′)µ(F ′)} ≤ bn
E, Π′′
and Π′
E.
E belong to the family of all γn-fine Henstock partitions of
Proof. By theorem 2.6, an (o)-sequence (bn)n exists, together with a
corresponding sequence of gauges (γn)n, so that
XF ′∈Π′
f (τF ′)µ(F ′) − XF ′′∈Π′′
f (τF ′′)µ(F ′′) ≤ bn
(2)
E and Π′′
E.
holds, for all γn-fine partitions Π′, Π′′. Now, consider any γn-fine
partition Π, and, for every E in Π, take two arbitrary subpartitions
Π′
By varying E, this gives rise to two γn-fine partitions of T for which
(2) holds true. Denote by E1 the first element of Π, and, in the
summation at left-hand side, take the supremum with respect just to
the sets of the type F ′ that are contained in E1: this yields
sup
Π′
E1
σ(f, Π′
E1) + XF ′∈Π′,
F ′6⊂E1
f (τF ′)µ(F ′) − XF ′′∈Π′′
σ(f, Π) ≤ bn.
Now, by varying only the F ′′s that are contained in E1, one obtains:
sup
Π′
E1
σ(f, Π′
E1) − inf
Π′′
E1
σ(f, Π′′
E1) + XF ′∈Π′,
F ′6⊂E1
f (τF ′)µ(F ′) − XF ′′6⊂E1
σ(f, Π) ≤ bn,
namely
ωn(f, E1) +XF ′∈Π′,
F ′6⊂E1
f (τF ′)µ(F ′) − XF ′′∈Π′′,
F ′′6⊂E1
f (τF ′′)µ(F ′′) ≤ bn.
where
ωn(E) := sup
E,Π′′
Π′
E
{ XF ′′∈Π′′
E
f (τF ′′)µ(F ′′) − XF ′∈Π′
E
f (τF ′)µ(F ′)}.
Comparison between some norm and order gauge integrals...
7
Now, starting from the last inequality, make the same operation
in the second subset of Π (say E2), by keeping fixed all the F ′ and F ′′
that are not contained in it: so
ωn(f, E1) + ωn(f, E2) + XF ′∈Π′,
F ′6⊂E1∪E2
f (τF ′)µ(F ′) − XF ′′∈Π′′,
F ′′6⊂E1∪E2
f (τF ′′)µ(F ′′) ≤ bn.
Continuing this procedure the assertion follows.
A consequence of Proposition 3.2 is that (oH)-integrability is hered-
itary on every measurable subset A of T .
In fact taking the same
(o)-sequence (bn)n together with the corresponding sequence (γn)n for
each n any γn-fine partition of A can be extended to a γn-fine par-
tition of T thanks to the Cousin Lemma and so, for any two γn-fine
partitions Π, Π′ of A, it follows
σ(f, Π) − σ(f, Π′) ≤ ωn(f, A) ≤ bn.
Then, the Theorem 2.6 yields the conclusion. The additivity of the
integral can be obtained as well. Moreover by Proposition 3.2 it fol-
lows:
Corollary 3.3. ( [20, Theorem 9]) Let f : T → X be any (oH)-
integrable function. Then there exist an (o)-sequence (bn)n and a cor-
responding sequence (γn)n of gauges, such that:
3.3.1) for every n and every γn-fine partition Π one has
f (τE)µ(E) − (oH)ZE
XE∈Π
f dµ ≤ bn.
3.3.2) for every n and every γn-fine partition Π it holds
f (τE)µ(E) − f (τ ′
E)µ(E) ≤ bn,
XE∈Π
when all the tags satisfy the condition E ⊂ γn(τ ′
γn(τE) for all E.
E) and E ⊂
Remark 3.4. Observe that in Corollary 3.3 all partitions may also be
free, since, as already noticed, the restriction τE ∈ E does not affect
the results.
A consequence of this theorem is the following
8
D. Candeloro and A. R. Sambucini
Theorem 3.5. ( [20, Theorem 11]) If f : T → X is (oH)-integrable,
then also f is.
Now one can prove that the modulus of the indefinite oH-integral
of f is precisely the indefinite oH-integral of f ; a result of this type
was given in [28, Theorem 1] for the Bochner integral of functions
taking values in a Banach lattice X.
Definition 3.6. Let f : T → X be any (oH)-integrable mapping, and
be the indefinite integral of f . The modulus of µf , denoted by µf , is
set µf (A) = (oH)RA f dµ for all Borel sets A ∈ B. Then µf is said to
defined for each A ∈ B as follows: µf (A) = sup{PB∈π µf (B) : π ∈
Π(A)} where Π(A) is the family of all finite partitions of A.
Then
Theorem 3.7. ( [20, Theorem 13]) If f : T → X is (oH)-integrable
then one has µf = µf .
Proof. The (oH)-integrability of f follows from Theorem 3.5, to-
gether with µf ≤ µf and so µf is bounded. Now for the reverse
inequality it will be sufficient to prove that µf (T ) = µf (T ) thanks to
the additivity of µf and µf . Let (bn)n and (γn)n be an (o)-sequence
and its corresponding sequence of gauges such that, for every n and
every γn-fine partition π ≡ (Ei, ti)i it holds
) ≤ bn,
(3)
∨(Xi
So,
≤ bn.
f dµ(cid:12)(cid:12)(cid:12)(cid:12)
f dµ(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
f (ti)µ(Ei) − (oH)ZEi
f (ti)µ(Ei) − (oH)ZEi
,Xi
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
f dµ(cid:12)(cid:12)(cid:12)(cid:12)
f (ti)µ(Ei) −(cid:12)(cid:12)(cid:12)(cid:12)
(oH)ZEi
Xi
µf (T ) − µf (T ) ≤ Xi (cid:18)µf (Ei) −(cid:12)(cid:12)(cid:12)(cid:12)
f dµ(cid:12)(cid:12)(cid:12)(cid:12)
(cid:19) ≤
≤ Xi (cid:0)µf (Ei) − f (ti)µ(Ei)(cid:1) +
+ Xi (cid:18)f (ti)µ(Ei) −(cid:12)(cid:12)(cid:12)(cid:12)
f dµ(cid:12)(cid:12)(cid:12)(cid:12)
(cid:19) ≤ 2bn
(oH)ZEi
(oH)ZEi
thanks to (3). Since bn ↓ 0, then µf (T ) = µf (T ), from which the
assertion follows.
Comparison between some norm and order gauge integrals...
9
Another appealing consequence is obtained in L-spaces.
In the
sequel, when the Banach lattice is assumed to be an L-space, it will
be denoted by L rather than X. The following definition, related with
norm-integrability, is introduced.
Definition 3.8. [25, Definition 3] f : T → X is variationally H inte-
grable (in short vH-integrable) if for every ε > 0 there exists a gauge
γ such that, for every γ-fine partition Π ≡ (E, tE )E the following
inequality holds:
kf (tE)µ(E) − (H)ZE
XE∈Π
f dµk ≤ ε.
(4)
For results on variational integrability see also [15, 25, 37].
Theorem 3.9. ( [20, Theorem 15]) If f : T → L is (oH)-integrable
then f is Bochner integrable.
Proof. It will be enough to prove that f is vH-integrable thanks to [26,
Theorem 2]. By Corollary 3.3, there exist an (o)-sequence (bn)n and
a corresponding sequence (γn)n of gauges, such that, for every n and
every γn-fine partition Π one has
f (τE)µ(E) − f (τ ′
E)µ(E) ≤ bn.
XE∈Π
By order continuity of the norm it is limn kbnk = 0. So, fix ε > 0 and
let N ∈ N be such that kbN k ≤ ε. Then, if Π is any γN -fine partition,
one has
f (τE)µ(E) − f (τ ′
≤ kbN k ≤ ε,
XE∈Π
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
XE∈Π (cid:12)(cid:12)kf (τE)k − kf (τ ′
E)µ(E)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
E)k(cid:12)(cid:12)µ(E) ≤ kbN k ≤ ε,
in accordance with condition 3.3.2). Since k · k is an L-norm it follows
that
(5)
when Π is γN -fine, both for the tags τE and for the tags τ ′
E. Now, pro-
ceeding as in Theorem 3.5, it is not difficult to prove that kf k satisfies
the Cauchy criterion for the Henstock integrability, and therefore it is
integrable. From (5) and (3.3.1) the condition (4) follows.
10
D. Candeloro and A. R. Sambucini
This result is apparently in contrast with the common situation in
general Banach spaces (when only norm integrals are involved); more-
over, in some normed lattices, Bochner (norm) integrability does not
imply oH-integrability, as shown in the example ( [9, Example 2.8]).
Theorem 3.9 can be used to show that the (oH)-integrability is
stronger than (H)-integrability:
it is enough to consider for example
the function f defined in [45, Theorem 3] when X is any L-space of
infinite dimension. The function f is McShane integrable (i.e. H-
integrable), but not Bochner integrable. Since f takes values in an
L-space, it cannot be oH-integrable, in view of the previous theorem.
Another consequence of the previous results is that an M -space X
admits an equivalent L-norm only if it is of finite dimension. In fact
in such an M -space X the (H)-integrability and the (oH)-integrability
are the same: so, if X has an equivalent L-norm then the previous
argument immediately shows that X must be of finite dimension.
3.1
Integrability in [0, 1]
In this section the space T is [0, 1] endowed with the usual Lebesgue
measure λ. Rather than using partitions made up with arbitrary mea-
surable subsets, here they will be taken more traditionally as free par-
titions consisting of subintervals.
Indeed, in [31] it is proven that
there is equivalence between the two types: though the proof there is
related only to norm integrals, the technique is the same. The sym-
bol (oM)-integral will be used instead of (oH)-integral. Analogously
to [35, Lemma 5.35], it holds:
Lemma 3.10. ( [20, lemma 18]) Let f : [0, 1] → X be given and
suppose that there exists an (o)-sequence (bn)n such that, for every n
two (oM)-integrable functions g1 and g2 can be found, with the same
regulating (o)-sequence (βn)n, such that g1 ≤ f ≤ g2 and
(oM)Z g2dλ ≤ (oM)Z g1dλ + bn.
Then f is (oM)-integrable.
The (oM)-integrability of increasing functions can be obtained as
in [35, example 5.36].
Comparison between some norm and order gauge integrals...
11
Theorem 3.11. ( [20, Theorem 19]) Let f : [0, 1] → X be increasing.
Then f is (oM)-integrable.
In a similar way, one can prove the (oM)-integrability more gen-
erally for (order bounded) mappings that are Riemann-integrable in
the order sense.
Above and also in the previous Theorem, one can easily see that
the gauges involved in the proof are constant mappings. More gener-
ally, one can consider notions of McShane, or Henstock integrability,
in which the gauges involved are required to be measurable positive
mappings. These notions have been studied in [39] and related bib-
liography, in terms of the norm integral. The corresponding notions
for the order-type integral are now considered.
Definition 3.12. The mapping f : [0, 1] → X is said to be order-
Birkhoff integrable if there exist an element J ∈ X, an o-sequence
(pn)n in X and a corresponding sequence of measurable gauges γn on
[0, 1] such that, for every integer n, and every γn-fine free partition
P ≡ (ti, Ei)k
i=1 of [0, 1] it holds σ(f, P ) − J ≤ pn, where σ(f, P ) =
i=1 f (ti)Ei.
Pk
(oB)R 1
Observe that the partitions P consist of non-overlapping intervals,
and the tags ti do not belong necessarily to the corresponding subinter-
vals Ei. When f is order-Birkhoff integrable (shortly, (oB)-integrable),
the element J will be called the oB-integral of f and denoted by
0 f dx. The terminology used here, i.e. Birkhoff integrability,
is due to the fact that, in the case of norm-type integrals, this notion
is equivalent precisely to the Birkhoff (norm) integral: see [39, Remark
1].
For further reference, the notation (oBH)-integrability (i.e. (order-
Henstock-Birkhoff )-integrability) is used if, in the above definition,
only Perron-type partitions are allowed, i.e. partitions for which the
tags belong to the corresponding sub-intervals. Standard techniques
allow to prove usual properties of the (oB)-integral, according to the
next Proposition.
Proposition 3.13. Let f and g be oB-integrable mappings on [0, 1],
and fix arbitrarily α, β in R. Then αf + βg is oB-integrable and
(oB)Z 1
0
(αf + βg)dx = α (oB)Z 1
0
f dx + β(oB)Z 1
0
g dx.
12
D. Candeloro and A. R. Sambucini
Moreover, f 1[a,b] is oB-integrable, for every a, b ∈ [0, 1], a < b, and
0
(oB)Z 1
f dx = (oB)Z 1
Clearly, the notation (oB)R b
of f 1[a,b].
f 1[0,a] dx + (oB)Z 1
0
f 1[a,1] dx
0
a f dx will be used, to indicate the integral
In [44, Example 2.1], it is proven that, in general, H-integrable
mappings are not Birkhoff (neither, a fortiori, Bochner) integrable:
since the mapping given there takes values in an M -space (i.e. l∞([0, 1])),
this is also an example of an oM-integrable function that is not Birkhoff
integrable. This example will be used to discuss other implications.
First, denote by D the set of all dyadic rational points in [0, 1] and enu-
merate the elements of D as a sequence (dn)+∞
n=0. Next, set ϕ(dn) = 1
2n
for each n. Then define the following σ-additive measure µ in the space
([0, 1], P([0, 1])):
µ(E) := Xdn∈E
ϕ(dn),
for each E ⊂ [0, 1]. Clearly µ is finite and the space L := L1([0, 1], P([0, 1]), µ)
is an L-space, with the norm
kf k1 = Xd∈D
f (d)ϕ(d).
Observe that each function in l∞([0, 1] is also a member of L, since
the series Pn ϕ(dn) is convergent, and the two spaces L and l∞ have
the same natural ordering. Moreover, if (pn)n is a decreasing sequence
in l∞ with infimum 0, it is clear that limn kpnk1 = 0 (when (pn)n is
thought of as a sequence in L).
Now, recall the Phillips' example, i.e. the function f : [0, 1] → l∞([0, 1])
defined as follows:
f (t)(s) =(cid:26) 1,
0.
if t − s
is a dyadic rational,
otherwise.
As earlier proved by Phillips in [40, Example 10.2], f is not Birkhoff
integrable, but it is McShane integrable, as later shown by Rodr´ıguez
in [44]. Since l∞([0, 1]) is an M -space, f is also (oM)-integrable.
Now, since the range of f is contained in L (endowed with the
same ordering as l∞), by the previous remarks it follows that f is
Comparison between some norm and order gauge integrals...
13
oH-integrable in this space, and therefore Bochner integrable thanks
to Theorem 3.9. So, the same mapping f is Bochner integrable with
respect to the k·k1-norm but not even Birkhoff integrable with respect
to k · k∞.
The example in [9, Example 2.8] shows that also for this type of
integral generally a.e. equality is not sufficient for two functions to
have the same integral. However, the proof of Proposition 3.1 shows
that this is actually the case for bounded functions. For this reason
here only bounded integrable mappings f : [0, 1] → X are considered
(unless otherwise specified).
One interesting result in this topic is that, at least for bounded
functions, (oB)-integrability can also be defined by means of Perron-
type partitions, i.e. (oB)-integrability is equivalent to (oBH)-integrability.
This is not easy to prove: to give an idea, a similar equivalence for the
(oM)-integral as defined at the beginning of this section, i.e. without
requiring measurability of the involved gauges, is still an open prob-
lem.
In order to achieve the purpose, some more notions are necessary. The
following definition is inspired at [39, Definition 2(b)].
Definition 3.14. Let f : [0, 1] → X be any mapping, and fix any
measurable subset E ∈ [0, 1]. The function f will be said to be (o)-
Riemann measurable in E if there exist an (o)-sequence (pn)n in X
and a corresponding sequence (Fn)n of closed subsets of E, such that
λ(E \ Fn) ≤ n−1 and
k
f (ti) − f (t′
i)λ(Ei) ≤ pn
(6)
Xi=1
hold true, for every n ∈ N and every finite collection of pairs {(ti, Ei)k
where the sets Ei are pairwise non-overlapping subintervals, with the
tags ti belong to Ei ∩ Fn for all i and max{λ(Ei)} ≤ n−1.
i=1},
Remark that, in this definition, the formula (6) is (formally) quite
stronger than the corresponding one, introduced in [39, Definition
2(b)]: the norm in the latter is outside the summation, while here
the modulus is inside.
It is easy to see that, if f is (o)-Riemann measurable on a set E,
then also αf is, for every real constant α. Moreover, if f and g are
14
D. Candeloro and A. R. Sambucini
(o)-Riemann measurable on E, the same holds also for f + g. Other
useful facts are collected in the following Proposition, whose proof is
similar to that of [39, Theorem 1].
Proposition 3.15. Let f : [0, 1] → X be any mapping, and fix any
measurable set E in [0, 1]. The following properties hold:
(3.15.a) If f is (o)-Riemann measurable on E, then it is on every
measurable subset E1 ⊂ E.
(3.15.b) f is (o)-Riemann measurable on E if and only if f 1E is
(o)-Riemann measurable on [0, 1].
Moreover
Theorem 3.16. Let f : [0, 1] → X be any (oBH)-integrable mapping
on [0, 1]. Then f is (o)-Riemann measurable.
Proof. Thanks to the Henstock Lemma, in particular 3.3.2), it follows
that, also for this type of integral, there exist an (o)-sequence (pn)n
and a corresponding sequence (γn)n of measurable gauges such that
f (ti)λ(Ei) − f (t′
i)λ(Ei) ≤ pn,
k
Xi=1
as soon as all the pairs (ti, Ei)k
i=1 form a γn-fine Henstock-type par-
tition.
(The fact that here the sets Ei are subintervals is irrele-
vant). Now, fix n. Since γn is measurable, there exist a positive
δn > 0 and an open set Gn ⊂ [0, 1] with λ(Gn) < n−1 such that
{t ∈ [0, 1] : γn(t) < δn} ⊂ Gn. Setting Fn = [0, 1] \ Gn, then Fn is
closed and, by the previous inequality,
f (ti) − f (t′
i)λ(Ei) ≤ pn
k
Xi=1
holds true, as soon as the intervals Ei are non-overlapping and satisfy
maxi λ(Ei) ≤ δn, and the points ti, t′
i are in Ei ∩ Fn.
Now, it will be proven that, as soon as f is bounded and (o)-
Riemann measurable, then it is (oB)-integrable. Of course, thanks to
Theorem 3.16, this will imply that (oBH)-integrability for bounded
functions is equivalent to (oB)-integrability. Proceeding as in [39], a
technical Lemma concerning the oscillation ω of a bounded function
f will be stated. Its proof is omitted since it is completely similar to
that of [39, Lemma 1].
Comparison between some norm and order gauge integrals...
15
Lemma 3.17. Let f : [0, 1] → X be any bounded mapping, and fix
any measurable set F ⊂ [0, 1]. Assume that a ∈ X and δ ∈ R are two
positive elements such that
N
(βi − αi)ω(f ; [αi, βi] ∩ F ) ≤
Xi=1
a
5
as soon as ([αi, βi])N
maxi(βi − αi) ≤ 10δ. Then
i=1 are pairwise disjoint subintervals of [0, 1] with
(βk − αk)ω(f ; [αk − δ, βk + δ] ∩ F ) ≤ a
K
Xk=1
as soon as ([αk, βk])K
maxk(βk − αk) ≤ 2δ.
k=1 are pairwise disjoint subintervals of [0, 1] with
So it follows:
Theorem 3.18. Let f : [0, 1] → X be a bounded and (o)-Riemann
measurable map. Then f is (oB)-integrable.
Proof. First, denote by M the supremum of f (x), as x runs in [0, 1].
Next, let (pn)n and (Fn)n denote the (o)-sequence and the correspond-
ing sequence of measurable sets in [0, 1] related to the (o)-Riemann
measurability of f . Now, define a new sequence (γ′
n)n of measurable
gauges as follows:
γ′
n(t) =(cid:26) γn(t)/20,
dist(t, F ),
t ∈ Fn
t /∈ Fn
Next, fix n and choose two free partitions of [a, b], say P and P ′,
respectively with pairs (Ei, ti), i = 1...k, and (E′
j), j = 1, ...k′, sub-
ordinated to γ′
j, t′
n. It holds
(f (ti) − f (t′
j))λ(Ei ∩ E′
j) ≤
σ(f, P ) − σ(f, P ′) = Xi,j
≤ X(i,j)∈S
+ X(i,j) /∈S
(f (ti) − f (t′
j))λ(Ei ∩ E′
j) +
f (ti) − f (t′
j)λ(Ei ∩ E′
j),
16
D. Candeloro and A. R. Sambucini
where S denotes the set of all couples (i, j) such that ti and t′
to Fn. By definition of the sets Fn, one has λ(F c
thanks to the definition of γ′
n:
j belong
n) ≤ n−1, and so, also
f (ti) − f (t′
j)λ(Ei ∩ E′
j) ≤ 2M n−1;
X(i,j) /∈S
moreover, as to the first summand, one has clearly
X(i,j)∈S
(f (ti) − f (t′
j))λ(Ei ∩ E′
j) ≤ X(i,j)∈S
f (ti) − f (t′
j)λ(Ei ∩ E′
j).
Now, from the definition of γ′
n, it follows that
max
(i,j)∈S
λ(Ei ∩ E′
j) ≤ 5−1γn :
so, thanks to the Lemma 3.17 and (o)-Riemann measurability,
f (ti) − f (t′
j)λ(Ei ∩ E′
j) ≤ 5pn.
X(i,j)∈S
Now, from the previous inequalities, one gets
σ(f, P ) − σ(f, P ′) ≤ 5pn + 2M n−1.
Since (5pn +2M n−1)n is clearly an (o)-sequence, the Cauchy condition
for the (oB)-integrability is satisfied; and, since X is complete, f turns
out to be (oB)-integrable.
Conclusions
In this paper the notions of Henstock (Mc Shane) integrability for
functions defined in a metric compact regular space and taking values
in a Banach lattice with an order-continuous norm are investigated.
Both the norm-type and the order-type integrals have been examined.
In general the order-type integral is stronger than the norm-one, while
in M -spaces the two notions coincide and in L-spaces the order-type
Henstock integral implies the Bochner one.
Acknowledgement
The authors have been supported by University of Perugia -- Depart-
ment of Mathematics and Computer Sciences -- Grant Nr 2010.011.0403,
Prin "Metodi logici per il trattamento dell'informazione", Prin "Descartes"
and by the Grant prot. U2014/000237 of GNAMPA - INDAM (Italy).
Comparison between some norm and order gauge integrals...
17
References
[1] E. J. Balder and A. R. Sambucini, A note on strong convergence
for Pettis integrable functions, Vietnam J. Math. 31 (3) (2003),
341 -- 347.
[2] E. J. Balder and A. R. Sambucini, On weak compactness and
lower closure results for Pettis integrable (multi)functions, Bull.
Pol. Acad. Sci. Math. 52 (1) (2004), 53 -- 61.
[3] A. Boccuto, A. V. Bukhvalov and A. R. Sambucini, Inequalities
in Classical Spaces with Mixed Norms, Positivity 6 (2002) ,393 --
411.
[4] A. Boccuto and D. Candeloro, A survey of decomposition and
convergence theorems for l-group-valued measures, Atti Sem.
Mat. Fis. Univ. Modena e Reggio Emilia 53 (2005), 243 -- 260.
[5] A. Boccuto, D. Candeloro and A. R. Sambucini, A Fubini Theo-
rem in Riesz Spaces for the Kurzweil-Henstock Integral, Journal
of Function Spaces and Applications 9, No. 3 (2011), 283 -- 304.
doi:10.1155/2011/158412
[6] A. Boccuto, D. Candeloro and A.R. Sambucini, Vitali-type theo-
rems for filter convergence related to vector lattice-valued modu-
lars and applications to stochastic processes, J. Math. Anal. Appl.
vol 419 (2) (2014), 818 -- 838. DOI: 10.1016/j.jmaa.2014.05.014
[7] A. Boccuto, D. Candeloro and A. R. Sambucini, Henstock multi-
valued integrability in Banach lattices with respect to pointwise
non atomic measures, submitted for publication to Atti della Ac-
cademia Nazionale dei Lincei (2015).
[8] A. Boccuto and X. Dimitriou, Convergence theorems for Lattice
Group-Valued Measures, Bentham Science Publ. U.A.E. (2014).
[9] A. Boccuto, A.M. Minotti and A.R. Sambucini, Set-valued
Kurzweil-Henstock integral in Riesz space setting, PanAmerican
Mathematical Journal 23 (1) (2013), 57 -- 74.
[10] A. Boccuto, B. Riecan and M. Vrabelova, Kurzweil-Henstock in-
tegral in Riesz spaces, e-book, Bentham Science Publ., U. A. E.,
2009.
[11] A. Boccuto and A. R. Sambucini, A McShane Integral for Mul-
tifunctions, J. Concr. Appl. Math. 2 (4) (2004), 307-325.
18
D. Candeloro and A. R. Sambucini
[12] A. Boccuto and A. R. Sambucini, A note on comparison be-
tween Birkhoff and McShane-type integrals for multifunctions,
Real Anal. Exchange 37 (2) (2012), 315-324.
[13] A. Boccuto, A.R. Sambucini and V.A. Skvortsov Integration by
parts for Perron type integrals of order 1 and 2 in Riesz spaces,
Results in Mathematics 51 (2007), 5 -- 27.
[14] B. Bongiorno, The Henstock -- Kurzweil integral, in Handobook of
measure theory, vols. I, II, North -- Holland, Amsterdam, 2002,
587 -- 615.
[15] B. Bongiorno, L. Di Piazza and K. Musia l, A variational Henstock
integral characterization of the Radon-Nikod´ym property, Illinois
Journal of Mathematics 1 (2009), 87 -- 99.
[16] A. V. Bukhvalov, A. I. Veksler and V. A. Geiler, Normed lattices,
Journal of Soviet Mathematics 18 (4) (1982), 516 -- 551.
[17] A. V. Bukhvalov, A. I. Veksler and G. Ya Lozanovskii, Banach
Lattices - Some Banach Aspects of Their Theory, Russian Math-
ematical Surveys 34 (2) (1979),159 -- 212.
[18] D. Candeloro, Riemann-Stieltjes integration in Riesz Spaces,
Rend. Mat. Roma (Ser. VII), 16 (2) (1996), 563-585.
[19] D. Candeloro and A.R. Sambucini, Filter convergence and de-
compositions for vector lattice-valued measures, Mediterranean
J. Math. (2014) doi: 10.1007/s00009-014-0431-0, arXiv:1401.7818
[math.FA].
[20] D. Candeloro and A.R. Sambucini, Order-type Henstock and Mc
Shane integrals in Banach lattice setting, Proceedings of the SISY
2014 - IEEE 12th International Symposium on Intelligent Sys-
tems and Informatics, pages 55-59; ISBN 978-1-4799-5995-2.
[21] S. Cao, The Henstock integral for Banach-valued functions, SEA
Bull. Math. 16 (1992), 35-40.
[22] K. Cich´on and M. Cich´on Some Applications of Nonabsolute In-
tegrals in the Theory of Differential Inclusions in Banach Spaces,
in: G.P. Curbera, G. Mockenhaupt, W.J. Ricker (Eds.), Vector
Measures, Integration and Related Topics, in: Operator Theory:
Advances and Applications, vol. 201, BirHauser-Verlag, ISBN:
978-3-0346-0210-5 (2010), 115-124.
[23] A. Croitoru and A. Gavrilut¸, Comparison Between Birkhoff Inte-
gral and Gould Integral, Mediterr. J. Math. DOI 10.1007/s00009-
014-0410-5 (2014).
Comparison between some norm and order gauge integrals...
19
[24] R. Deville and J. Rodr´ıguez, Integration in Hilbert generated
Banach spaces, Israel J. Math. 177 (2010), 285-306.
[25] L. Di Piazza, and V. Marraffa, The McShane, PU and Henstock
integrals of Banach valued functions, Czech. Math. Journal 52
(2002), 609-633.
[26] L. Di Piazza and K. Musia l, A characterization of variationally
McShane integrable Banach-space valued functions, Illinois J.
Math. 45 (1) (2001), 279-289.
[27] L. Di Piazza and D. Preiss, When do McShane and Pettis in-
tegrals coincide?, Illinois Journal of Math. 47 (4) (2003), 1177 --
1187.
[28] L. Drewnowski and W. Wnuk, On the modulus of indefinite vec-
tor integrals with values in Banach lattices, Atti Sem. Mat. Fis.
Univ. Modena 47 (1999), 221-233.
[29] D. H. Fremlin, The Henstock and McShane integrals of vector-
valued functions, Illinois J. Math. 38 (3) (1994), 471 -- 479.
[30] D. H. Fremlin, J. Mendoza, On the integration of vector-valued
functions, Illinois J. Math. 38 (1) (1994), 127 -- 147.
[31] D. H. Fremlin, The generalized McShane integral, Illinois J.
Math. 39 (1) (1995), 39 -- 67.
[32] D. H. Fremlin, Measure theory. Vol. 3. Measure Algebras, Torres
Fremlin, Colchester, 2002.
[33] D. H. Fremlin, Measure theory. Vol. 4. Topological measure
spaces, Torres Fremlin, Colchester, 2006.
[34] A.C. Gavrilut¸, A.E. Iosif and A. Croitoru, The Gould integral in
Banach lattices, Positivity (2014), DOI 10.1007/s11117-014-0283-
7.
[35] D.S.Kurtz and C.W.Swarz, Theories of Integration. The integrals
of Riemann, Lebesgue,Henstock-Kurzweil, and McShane, World
Scientific Series in Real Analysis vol. 9 (2004).
[36] C. C. A. Labuschagne and B. A. Watson, Discrete stochastic
integration in Riesz spaces, Positivity 14 (4) (2010), 859-875.
[37] V. Marraffa, The Variational McShane Integral in Locally Con-
vex Spaces, Rocky Mountain Journal of Mathematics 39 (2009),
1993 -- 2013. DOI: 10.1216/RMJ-2009-39-6-1993
20
D. Candeloro and A. R. Sambucini
[38] P. Meyer-Nieberg, Banach lattices, Springer-Verlag, Berlin-
Heidelberg, 1991.
[39] K.M. Naralenkov, A Lusin type measurability property for
vector-valued functions, J. Math. Anal. Appl. 417 (1) (2014),
293 -- 307.
[40] R.S. Phillips, Integration in a convex linear topological space,
Trans. Amer. Math. Soc. 47 (1940), 114-145.
[41] B. Riecan, On the Kurzweil Integral in Compact Topological
Spaces, Radovi Mat. 2 (1986), 151-163.
[42] J. Rodr´ıguez, On the existence of Pettis integrable functions
which are not Birkhoff integrable, Proc. Amer. Math. Soc. 133
(4) (2005), 1157 -- 1163.
[43] J. Rodr´ıguez, On the equivalence of McShane and Pettis integra-
bility in non-separable Banach spaces, J. Math. Anal. Appl. 350
(2009), 514-524.
[44] J. Rodr´ıguez, Some examples in vector integration, Bull. of the
Australian Math. Soc. 80 (3) (2009), 384 -- 392.
[45] V.A. Skvortsov and A.P. Solodov, A variational
integral for
Banach-valued functions, Real Anal. Exchange 24 (1998-1999),
799-806.
|
1904.11880 | 1 | 1904 | 2019-04-26T15:00:55 | Subadditive inequalities for operators | [
"math.FA"
] | In this article, we present a new subadditivity behavior of convex and concave functions, when applied to Hilbert space operators. For example, under suitable assumptions on the spectrum of the positive operators $A$ and $B$, we prove that \[{{2}^{1-r}}{{\left( A+B \right)}^{r}}\le {{A}^{r}}+{{B}^{r}}\quad\text{ for }r>1\text{ and }r<0,\] and \[{{A}^{r}}+{{B}^{r}}\le {{2}^{1-r}}{{\left( A+B \right)}^{r}}\quad\text{ for }r\in \left[ 0,1 \right].\] These results provide considerable generalization of earlier results by Aujla and Silva.
Further, we present several extensions of the subadditivity idea initiated by Ando and Zhan, then extended by Bourin and Uchiyama. | math.FA | math |
SUBADDITIVE INEQUALITIES FOR OPERATORS
HAMID REZA MORADI, ZAHRA HEYDARBEYGI AND MOHAMMAD SABABHEH
Abstract. In this article, we present a new subadditivity behavior of convex and concave
functions, when applied to Hilbert space operators. For example, under suitable assumptions
on the spectrum of the positive operators A and B, we prove that
21´rpA ` Bqr ď Ar ` B r
for r ą 1 and r ă 0,
and
Ar ` B r ď 21´rpA ` Bqr
for r P r0, 1s .
These results provide considerable generalization of earlier results by Aujla and Silva.
Further, we present several extensions of the subadditivity idea initiated by Ando and Zhan
then extended by Bourin and Uchiyama.
1. Introduction
Let B pHq be the algebra of all (bounded linear) operators on a complex Hilbert space H.
An operator A on H is said to be positive (in symbol: A ě 0) if xAx, xy ě 0 for all x P H. We
write A ą 0 if A is positive and invertible; and we say that A is strictly positive in this case.
For self-adjoint operators A and B, we write A ě B if A´ B is positive, i.e., xAx, xy ě xBx, xy
for all x P H. We call it the usual order. In particular, for some scalars m and M, we write
m ď A ď M if mxx, xy ď xAx, xy ď M xx, xy for all x P H. In what follows we denote the
weighted arithmetic mean and the weighted geometric mean of strictly positive operators A
and B by A∇vB " p1 ´ vq A` vB and A7vB " A
2 , respectively. For the case
v " 1{2 we write A∇B and A7B. Notice that the definition of A∇vB is still valid for positive
(not necessarily strictly positive) operators.
2´A´ 1
2¯v
2 BA´ 1
1
1
A
A real-valued continuous function f on an interval J is said to be operator convex (resp.
operator concave) if f pA∇vBq ď presp. ěq f pAq ∇vf pBq for all v P r0, 1s and for all self-
adjoint operators A, B P BpHq whose spectra are contained in J. A continuous function f on
J is called operator monotone increasing (resp. decreasing), if
A ď B ñ f pAq ď presp. ěq f pBq .
2010 Mathematics Subject Classification. Primary 47A63; Secondary 47A64, 47B65, 15A60.
Key words and phrases. Operator inequality, operator convex, operator concave, positive operator.
1
2
H.R. Moradi, Z. Heydarbeygi & M. Sababheh
An important observation that relates operator convex and operator monotone functions is
given by the following result [1, Theorem 2.1, Theorem 3.1, Theorem 2.3 and Theorem 3.7].
Proposition 1.1. Let f : p0,8q Ñ r0,8q be continuous. Then
(1) f is operator monotone decreasing if and only if f is operator convex and fp8q ă 8.
(2) f is operator monotone increasing if and only if f is operator concave.
For a more recent reference for this proposition, we refer the reader to [13, Theorem 2.4].
We remark that in this reference, it is shown that a function f defined on pa,8q is operator
monotone if and only if f is operator concave and fp8q ą ´8. We notice that the condition
fp8q ą ´8 is implicitly assumed in Proposition 1.1 since f is non-negative.
Proposition 1.1 will be used frequently in the sequel; where treatment of operator monotone
functions and that of operator convexity are interchangeable.
It is well-known that a concave function f with fp0q ě 0 is subadditive in the sense that
(1.1)
fpa ` bq ď fpaq ` fpbq
for the non-negative numbers a, b. A similar inequality is not necessarily valid for operator
concave functions. That is, an operator concave function f does not necessarily satisfy
fpA ` Bq ď fpAq ` fpBq
for the positive operators A, B. In 1999 Ando and Zhan [2] proved a subadditivity inequality
for operator concave functions, which says that if A, B are two positive matrices, then
(1.2)
}f pA ` Bq}u ď }f pAq ` f pBq}u
for any unitarily invariant norm }}u and any non-negative operator monotone function f on
r0,8q. Bourin and Uchiyama [6] showed that the condition operator concavity in (1.2) can be
replaced by scalar concavity.
We refer the reader to [2, 6] for different variants of (1.2). Our main target in this paper is to
discuss subadditivity inequalities for convex and operator convex functions, without appealing
to unitarily invariant norms. However, this will be at the cost of additional conditions or weaker
estimates. The first result in this direction will be an operator convexity behavior for convex
functions. More precisely, we prove that a convex function satisfies the operator convexity
inequality
under some conditions on the spectra of the positive operators A, B.
f A ` B
2 ď
fpAq ` fpBq
2
,
Subadditive inequalities for operators
3
Moreover, we show that concave functions (not necessarily operator concave) satisfy the
operator subadditivity inequality
k fpA ` Bq ď fpAq ` fpBq,
for a positive scalar k ď 1. We consider this an interesting extension of (1.2).
Another result of this type is due to Aujla [3] which asserts that if the function f is completely
f pkq pxq ě 0 for all k " 0, 1, . . . and all x P r0,8q, where f p0q " 0
monotone (in the sense, p´1qk
and f pkq denotes the k-th derivative of f ) on r0,8q, then
(1.3)
2}f pA ` Bq}u ď }f p2Aq ` f p2Bq}u.
In Sec. 2, we extend the norm order in (1.3) to the operator order (see Corollary 2.1).
Naturally, this generalization imposes additional conditions.
Inspired by the result given in
Theorem 2.1, we present some extensions of the inequalities due to Aujla and Silva [4]. Our
inequalities refine earlier results in this direction obtained in [3] and [10].
2. Main Results
Our first main result proposes a mild condition under which operator convexity follows from
scalar convexity.
Theorem 2.1. Let A, B P B pHq be two positive operators such that n ď A ď N and m ď B ď
M for some positive real numbers n ă N and m ă M. Further, let f : p0,8q Ñ r0,8q be a
convex function. If, for some v P p0, 1q,
rn∇vm, N∇vMs X rn, Ns ,rn∇vm, N∇vMs X rm, Ms " ∅,
then
(2.1)
In particular,
(2.2)
f pA∇vBq ď f pAq ∇vf pBq .
f A ` B
2 ď
f pAq ` f pBq
2
provided that the above empty intersection condition is fulfilled when v " 1
2 .
The reverse inequalities hold when f is concave.
Proof. We use an idea from [11, Theorem 1]. For simplicity, let a " n∇vm and b " N∇vM.
Now since ra, bs X rm, Ms " ∅ and ra, bs X rn, Ns " ∅, we will consider the secant of f on
the interval ra, bs. So, let
yptq "
b ´ t
b ´ a
fpaq `
t ´ a
b ´ a
fpbq.
4
H.R. Moradi, Z. Heydarbeygi & M. Sababheh
Since f is convex, it satisfies
(2.3)
fptq ď yptq,
a ď t ď b.
Since n ď A ď N and m ď B ď M, it follows that n∇vm ď A∇vB ď N∇vM, and hence
(2.4)
fpA∇vBq ď ypA∇vBq
by applying a functional calculus argument to (2.3).
Further, by the empty intersection assumption, we have
fptq ě yptq,
t P rn, Ns Y rm, Ms,
because f is convex on p0,8q.
Since n ď A ď N and m ď B ď M, functional calculus implies
fpBq ě ypBq.
fpAq ě ypAq
and
Noting that ypAq∇vypBq " ypA∇vBq, the above inequalities together with (2.4) imply
fpAq∇vfpBq ě ypAq∇vypBq " ypA∇vBq ě fpA∇vBq,
which completes the proof.
(cid:3)
Related to Theorem 2.1, we present the following version for v ě 1. Proceeding similarly, one
can obtain similar results for v ď 0.
Theorem 2.2. Let A, B P B pHq be two positive operators such that n ď A ď N and m ď B ď
M for some positive real numbers n ă N and m ă M. Further, let f : p0,8q Ñ r0,8q be a
convex function. If, for some v ě 1,
rN∇vm, n∇vMs X rm, Ms " ∅ and
rn, Ns Ă rN∇vm, n∇vMs ,
then
The reverse inequalities hold when f is concave.
f pA∇vBq ď f pAq ∇vf pBq .
Proof. Notice first that when n ď A ď N, m ď B ď M and v ě 1, then
c :" N∇vm ď A∇vB ď n∇vM :" d.
Let
yptq "
Then noting the assumptions and applying a functional calculus argument, similar to Theorem
2.1, imply the desired inequality.
(cid:3)
d ´ t
d ´ c
fpcq `
t ´ c
d ´ c
fpdq.
Subadditive inequalities for operators
5
Remark 2.1. In Theorem 2.1, we assumed that f is convex on p0,8q. In fact, it is enough to as-
sume convexity on an interval J containing the three intervals rn, Ns,rm, Ms and rn∇vm, N∇vMs.
Also, in Theorem 2.1 we assumed that A and B are positive operators. It is clear that self
adjointeness is enough.
As an immediate consequence of Theorem 2.1, we have the following subadditivity result for
convex functions.
Corollary 2.1. Under the same assumptions of theorem 2.1, with v " 1
ditivity inequality holds for the convex function f .
2, the following subad-
2f pA ` Bq ď f p2Aq ` f p2Bq ô f A ` B
2 ď
fpAq ` fpBq
2
.
If in addition f p2tq ď 2f ptq, then
f pA ` Bq ď f pAq ` f pBq .
Proof. We have
2f pA ` Bq " 2f2A ` 2B
2
This completes the proof.
ď f p2Aq ` f p2Bq .
(cid:3)
Notice that the assumption f p2tq ď 2f ptq can be dropped for the non-negative decreasing
functions. This follows from the fact that each non-negative decreasing function f satisfies
f p2tq ď 2f ptq.
Remark 2.2. Although our assumptions in Corollary 2.1 are different from [4, Corollary 2.5],
we present an operator inequality which is much stronger than the eigenvalue inequality in [4].
The following corollary presents a stronger version of [3, Corollary 2.5].
In [3], unitarily
invariant versions for complex matrices have been obtained. In the following result, we present
an operator version under some conditions on the spectra. In this result, we use the notation
τpXq to denote the smallest closed interval containing the spectrum of the operator X P BpHq.
Corollary 2.2. Let A, B P B pHq be two strictly positive operators. Then
and
21´rpA ` Bqr ď Ar ` B r
for r ą 1 and r ă 0,
for r P r0, 1s
whenever τ pAq X τ` A`B
2
Ar ` B r ď 21´rpA ` Bqr
, τ pBq X τ` A`B
" ∅.
2
Remark 2.3. The inequalities in Corollary 2.2 are also stronger than the one given in [4,
Remark 2.11] because of the same reasoning mentioned in Remark 2.2.
6
H.R. Moradi, Z. Heydarbeygi & M. Sababheh
The following example shows how Corollary 2.2 works.
Example 2.1. Taking A " 3 1
1 5¸, B " 10 ´1
9 ¸. It is not hard to check that A and B
´1
satisfy in the conditions of Corollary 2.2.
(i) For r " 6,
Ar ` B r ´ 21´rpA ` Bqr «985931.21 ´476992
433279 ¸ ş 0.
´476992
(ii) For r " ´2,
Ar ` B r ´ 21´rpA ` Bqr « 0.0956 ´0.0384
0.0229 ¸ ş 0.
´0.0384
(iii) For r " 1{3 ,
21´rpA ` Bqr ´ Ar ´ B r « 0.1519 ´0.061
´0.061 0.0486¸ ş 0.
Although Corollary 2.2 does not present equivalent conditions; and only necessary conditions
are proposed, we present the following example which presents an example where the conditions
and the conclusions are not satisfied.
Example 2.2. Let
A " 1 1
1 1 ¸ , B " 3 1
1 1 ¸ .
This example was given in [5, Example V.1.4., P. 114] to show that the function fptq " t3 is
not operator convex. That is, it is shown there that the inequality
21´rpA ` Bqr ď Ar ` B r
is not true for r " 3 and the above matrices. Here we explain why this inequality of Corollary
2.2 does not apply. The reason lies in the computations of the spectra:
τpAq " r0, 2s, τpBq " r2 ´
?2, 2 `
Clearly,
?2s and τ A ` B
τpAq X τ A ` B
2 ‰ ∅ and τpBq X τ A ` B
That is, the conditions of Corollary 2.2 are not satisfied .
inequality 21´rpA ` Bqr ď Ar ` B r is not valid, for r " 3.
3 ` ?5
,
.
2
2
2 "„3 ´ ?5
2 ‰ ∅.
It is also readily seen that the
Our next result is a Hermite-Hadamard type inequality for operators satisfying certain con-
ditions on their spectra.
Subadditive inequalities for operators
7
Corollary 2.3. Let A, B P B pHq be two positive operators such that n ď A ď N and m ď
B ď M for some positive real numbers n ă N and m ă M. Further, let f : p0,8q Ñ r0,8q be
a convex function. If, for all v P p0, 1q,
rn∇vm, N∇vMs X rn, Ns ,rn∇vm, N∇vMs X rm, Ms " ∅,
then
(2.5)
f A ` B
2 ď
1
ż0
f pp1 ´ vq A ` vBq dv ď
f pAq ` f pBq
2
.
Proof. As we have shown, if the assumptions of Theorem 2.1 are satisfied then
(2.6)
f pp1 ´ vq A ` vBq ď p1 ´ vq f pAq ` vf pBq ,
@v P p0, 1q.
It follows from (2.6) that, for v P p0, 1q,
f A ` B
2 " fp1 ´ vq A ` vB ` p1 ´ vq B ` vA
2
(2.7)
ď
ď
1
2 rf pp1 ´ vq A ` vBq ` f pp1 ´ vq B ` vAqs
f pAq ` f pBq
.
2
Now, integrating over v P r0, 1s the inequalities in (2.7) and taking into account that
1
ż0
f pp1 ´ vq A ` vBq dv "
1
ż0
f pp1 ´ vq B ` vAq dv
we infer
(2.8)
f A ` B
2 ď
1
ż0
f pp1 ´ vq A ` vBq dv ď
f pAq ` f pBq
2
.
Notice that (2.8) nicely extend the main result of [7].
(cid:3)
Aujla showed that if f : r0,8q Ñ r0,8q is an operator monotone decreasing function f , then
[3, Theorem 2.6]
(2.9)
2f pA ` Bq ď f pAq ` f pBq ,
where A, B are two positive matrices. Indeed, (2.9) follows by adding the following observations
f pA ` Bq ď f pAq
and
f pA ` Bq ď f pBq .
On account of Proposition 1.1, we can improve (2.9) as follows.
8
H.R. Moradi, Z. Heydarbeygi & M. Sababheh
Proposition 2.1. Let A, B P B pHq be two positive operators. If f : r0,8q Ñ r0,8q is an
operator monotone decreasing function, then
(2.10)
2f pA ` Bq ď 2f pA∇Bq ď f pAq ` f pBq .
Proof. It follows from the assumption on f that
A∇B ď A ` B ñ f pA ` Bq ď f pA∇Bq .
It is known that such operator monotone decreasing function is also operator convex (see [13,
Theorem 2.4] and [1]), i.e.,
f pA ` Bq ď f pA∇Bq ď f pAq ∇f pBq
and the proof is complete.
(cid:3)
Remark 2.4. Notice that (2.10) is also stronger than [10, Proposition 3.14], due to the fact
that f pA∇Bq ď f pAq7f pBq, when f : p0,8q Ñ p0,8q is operator monotone decreasing.
In the following, we provide a reverse inequality for the subadditivity property of operator
monotone decreasing functions. To reach this end, we need the following lemma which can be
proved using the well-known Mond -- Pecari´c method. A comprehensive survey on this topic can
be found in [8, Chapter 2].
Lemma 2.1. Let A1, . . . , An P B pHq be positive operators with spectra contained in rm, Ms
and w1, . . . , wn be positive scalars with řn
i"1 wi " 1. If f is a positive operator convex function
on rm, Ms, then
n
where
ÿi"1
K pm, M, fq " max" 1
wiAi¸
wif pAiq ď K pm, M, fq f n
ÿi"1
t ´ m
M ´ m
f ptq M ´ t
M ´ m
f pmq `
f pMq : t P rm, Ms* .
The next result is stated in terms of operator monotone decreasing function. Notice that this
condition maybe replaced by operator convexity with the additional condition that fp8q ă 8,
as we have seen in Proposition 1.1.
Theorem 2.3. Let A, B P B pHq be two positive operators such that n ď A, B ď N.
If
f : p0,8q Ñ p0,8q is an operator monotone decreasing function and 0 ă m ď 2n ď 2N ď M,
then
(2.11)
f pAq ` f pBq ď 4K pm, M, fq f pA ` Bq
where K pm, M, fq is defined as in Lemma 2.1.
Subadditive inequalities for operators
9
Proof. By Proposition 1.1, f is operator convex and Lemma 2.1 applies. By taking n " 2,
w1 " w2 " 1{2 , A1 " 2A, and A2 " 2B in Lemma 2.1, we get
f p2Aq ` f p2Bq
2
ď K pm, M, fq f pA ` Bq .
Since f : p0,8q Ñ p0,8q is operator monotone decreasing, we have f pαtq ě 1
α ě 1 by [9, Lemma 2.2]. This implies the desired result.
α f ptq for each
(cid:3)
A better estimate than that in Theorem 2.3 may be obtained for operator concave functions
as follows. To this end, an argument similar to that in Lemma 2.1 implies
f n
ÿi"1
wiAi¸ ď
1
k pm, M, fq
n
ÿi"1
wif pAiq
where
(2.12)
k pm, M, fq " min" 1
f ptq M ´ t
M ´ m
f pmq `
t ´ m
M ´ m
f pMq : t P rm, Ms* ,
for the positive operator concave function f . Now we are ready to present a subadditive
property for operator concave functions.
First, we notice that any non-negative concave function f on p0,8q satisfies the property
fpαtq ď αfptq for α ě 1, t ą 0. This can be easily seen by considering the derivative of the
function gptq " fpαtq´ αfptq. Of course, if f is not differentiable, it is still can be approximated
by smooth concave functions and the differential argument holds.
Theorem 2.4. Let A, B P B pHq be two positive operators such that n ď A, B ď N.
f : p0,8q Ñ p0,8q is an operator concave function and 0 ă m ď 2n ď 2N ď M, then
(2.13)
k pm, M, fq f pA ` Bq ď f pAq ` f pBq
If
where k pm, M, fq is as in (2.12).
Proof. Proceeding like Theorem 2.3 and noting that fpαtq ď αfptq when f is operator concave
and α ě 1 , we obtain
2
1
fp2Aq ` fp2Bq
fpA ` Bq " f 2A ` 2B
kpm, M, fq
kpm, M, fq pfpAq ` fpBqq ,
1
ď
ď
2
completing the proof.
(cid:3)
10
H.R. Moradi, Z. Heydarbeygi & M. Sababheh
Remark 2.5. The inequalities (2.11) and (2.13) can be extended in the following way:
f ℓ
ÿi"1
Ai¸ ď
1
k pm, M, fq
ℓ
ÿi"1
f pAiq,
whenever f is operator concave on rm, Ms and n ď Ai ď N, where m ď ℓ n ď ℓ N ď M. In
addition,
f pAiq ď f ℓ
ÿi"1
whenever f is an operator monotone decreasing on rm, Ms.
ℓ2K pm, M, fq
1
ℓ
ÿi"1
Ai¸
We conclude our discussion of subadditive-type inequalities by the following versions for
convex and concave functions (not necessarily operator convex or operator concave).
For this, we remind the reader of the following simple observations. If f is convex (concave)
on an interval J containing the spectrum of a self adjoint operator A, then
(2.14)
f phAx, xiq ď pěq hfpAqx, xi ,
for any unit vector x P H. In [12, Theorem 6], reverses of these celebrated inequalities were
found as follows
(2.15)
Kpm, M, fqf phAx, xiq ě hfpAqx, xi ,
where f is convex on rm, Ms and m ď A ď M. On the other hand, if f is concave on rm, Ms
and m ď A ď M, we have
(2.16)
f phAx, xiq ď kpm, M, fq hfpAqx, xi .
Utilizing (2.16) and (2.15) implies the following superadditive and subadditive versions.
Theorem 2.5. Let m, M be be positive scalars and m ď A, B ď M. If f : r0, Ms Ñ r0,8q is
convex with fp0q " 0, then
fpAq ` fpBq ď Kpm, M, fqfpA ` Bq,
where K pm, M, fq is defined as in Lemma 2.1. On the other hand, if f is concave, then
fpAq ` fpBq ě kpm, M, fqfpA ` Bq,
where k pm, M, fq is defined as in (2.12).
Subadditive inequalities for operators
11
Proof. We prove the first inequality for convex function. The second inequality follows similarly.
Let x P H be a unit vector. Then, for the convex f with fp0q " 0,
xpf pAq ` f pBqq x, xy " xf pAq x, xy ` xf pBq x, xy
ď K pm, M, fqpf pxAx, xyq ` f pxBx, xyqq
ď K pm, M, fq f pxpA ` Bq x, xyq
ď K pm, M, fqxf pA ` Bq x, xy
pby (1.1) for convex functionsq
pby (2.14)q.
pby (2.15)q
Since this is true for an arbitrary unit vector x, the first inequality follows immediately. The
second follows similarly.
(cid:3)
Acknowledgement
The authors would like to express their gratitude to the anonymous referee for valuable
comments and corrections.
References
[1] T. Ando and F. Hiai, Operator log -- convex functions and operator means, Math. Ann., 350(3) (2011), 611 --
630.
[2] T. Ando and X. Zhan, Norm inequalities related to operator monotone functions, Math. Ann., 315 (1999),
771 -- 780.
[3] J.S. Aujla, Some norm inequalities for completely monotone functions, SIAM J. Matrix Anal. Appl., 22(2)
(2000), 569 -- 573.
[4] J.S. Aujla and F.C. Silva, Weak majorization inequalities and convex functions, Linear Algebra Appl., 369
(2003), 217 -- 233.
[5] R. Bhatia, Matrix analysis, Springer-Verlag, New York, 1997.
[6] J.C. Bourin and M. Uchiyama, A matrix subadditivity inequality for f pA ` Bq and f pAq ` f pBq, Linear
Algebra Appl., 423 (2007), 512 -- 518.
[7] S.S. Dragomir, Hermite-Hadamards type inequalities for operator convex functions, Appl. Math. Comput.,
218(3) (2011), 66 -- 772.
[8] T. Furuta, J. Mi´ci´c, J. Pecari´c and Y. Seo, Mond -- Pecari´c method in operator inequalities, Inequalities for
Bounded Selfadjoint Operators on a Hilbert Space. Element, Zagreb, 2005.
[9] I.H. Gumu¸s, H.R. Moradi and M. Sababheh, More accurate operator means inequalities, J. Math. Anal.
Appl., 465 (2018) 267 -- 280.
[10] M. Kian and S.S. Dragomir, f -Divergence functional of operator log-convex functions, Linear Multilinear
Algebra, 64(2) (2016), 123 -- 135.
[11] J. Mi´ci´c, Z. Pavi´c and J. Pecari´c, Jensen's inequality for operators without operator convexity, Linear
Algebra Appl., 434(5) (2011), 1228 -- 1237.
[12] J. Mi´ci´c, Y. Seo, S.E. Takahasi and M. Tominaga, Inequalities of Furuta and Mond -- Pecari´c, Math. Ineq.
Appl., 2 (1999), 83 -- 111.
[13] M. Uchiyama, Operator monotone functions, positive definite kernels and majorization, Proc. Amer. Math.
Soc., 138(11) (2010), 3985 -- 3996.
12
H.R. Moradi, Z. Heydarbeygi & M. Sababheh
(H.R. Moradi) Young Researchers and Elite Club, Mashhad Branch, Islamic Azad University, Mashhad, Iran.
E-mail address: [email protected]
(Z. Heydarbeygi) Department of Mathematics, Payame Noor Universtiy (PNU), P.O. BOX, 19395 -- 4697, Tehran, Iran.
E-mail address: [email protected]
(M. Sababheh) Department of Basic Sciences, Princess Sumaya University for Technology, Amman 11941, Jordan.
E-mail address: [email protected]
|
1003.0279 | 3 | 1003 | 2010-11-21T06:29:54 | Improved bounds in the metric cotype inequality for Banach spaces | [
"math.FA",
"math.MG"
] | It is shown that if (X, ||.||_X) is a Banach space with Rademacher cotype q then for every integer n there exists an even integer m< n^{1+1/q}$ such that for every f:Z_m^n --> X we have
$\sum_{j=1}^n \Avg_x [ ||f(x+ (m/2) e_j)-f(x) ||_X^q ] < C m^q \Avg_{\e,x} [ ||f(x+\e)-f(x) ||_X^q ]$,
where the expectations are with respect to uniformly chosen x\in Z_m^n and \e\in \{-1,0,1\}^n, and all the implied constants may depend only on q and the Rademacher cotype q constant of X. This improves the bound of m< n^{2+\frac{1}{q}} from [Mendel, Naor 2008]. The proof of the above inequality is based on a "smoothing and approximation" procedure which simplifies the proof of the metric characterization of Rademacher cotype of [Mendel, Naor 2008]. We also show that any such "smoothing and approximation" approach to metric cotype inequalities must require m> n^{(1/2)+(1/q)}. | math.FA | math | IMPROVED BOUNDS IN THE METRIC COTYPE INEQUALITY
FOR BANACH SPACES
OHAD GILADI, MANOR MENDEL, AND ASSAF NAOR
Abstract. It is shown that if (X,k·kX) is a Banach space with Rademacher cotype q then
for every integer n there exists an even integer m . n1+ 1
m → X
we have
q such that for every f : Zn
n
Xj=1
Ex"(cid:13)(cid:13)(cid:13)
m
2
f(cid:16)x +
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
q
X# . mq Eε,xhkf (x + ε) − f (x)kq
Xi,
(1)
where the expectations are with respect to uniformly chosen x ∈ Zn
m and ε ∈ {−1, 0, 1}n, and
all the implied constants may depend only on q and the Rademacher cotype q constant of X.
This improves the bound of m . n2+ 1
q from [13]. The proof of (1) is based on a "smoothing
and approximation" procedure which simplifies the proof of the metric characterization of
Rademacher cotype of [13]. We also show that any such "smoothing and approximation"
approach to metric cotype inequalities must require m & n
2 + 1
q .
1
0
1
0
2
v
o
N
1
2
]
.
A
F
h
t
a
m
[
3
v
9
7
2
0
.
3
0
0
1
:
v
i
X
r
a
Contents
Introduction
1.
1.1. The smoothing and approximation scheme
1.2. A lower bound on smoothing and approximation with general kernels
1.3. The relation to nonembeddability results and some open problems
2. Proof of Theorem 1.2
3. Proof of Lemma 2.1
3.1. Estimates for the bivariate Bernoulli numbers
3.2. Some combinatorial identities
3.3. Putting things together
4. Lower bounds
4.1. A lower bound for general convolution kernels: Proof of Proposition 4.1
4.2. A sharp lower bound for Ej averages: Proof of Proposition 4.2
4.3. Symmetrization
References
2
3
6
7
9
13
13
15
18
20
20
24
25
27
2010 Mathematics Subject Classification. 46B80,46B85,51F99,05C12.
1
1. Introduction
m
2
for every integer n there exists an even integer m such that for every f : Zn
A metric space (M , dM ) is said [13] to have metric cotype q > 0 with constant Γ > 0 if
m → X we have
(2)
n
Xj=1
ej(cid:17) , f (x)(cid:17)q# 6 ΓqmqEε,xhdM (f (x + ε), f (x))qi.
Ex"dM (cid:16)f(cid:16)x +
In (2) the expectations are taken with respect to x chosen uniformly at random from the
discrete torus Zn
m, and ε chosen uniformly at random from {−1, 0, 1}n (the ℓ∞ generators of
m). Also, in (2) and in what follows, {ej}n
Zn
A Banach space (X,k · kX) is said to have Rademacher cotype q > 0 if there exists a
constant C < ∞ such that for every n ∈ N and for every x1, x2, ...., xn ∈ X,
j=1 denotes the standard basis of Zn
m.
(3)
n
Xj=1
kxjkq
Eε"(cid:13)(cid:13)(cid:13)
q
X# .
εjxj(cid:13)(cid:13)(cid:13)
n
Xj=1
X 6 C qEε"(cid:13)(cid:13)(cid:13)
X# 6 T p
εjxj(cid:13)(cid:13)(cid:13)
Xj=1
n
p
n
Xj=1
X is said to have Rademacher type p > 0 if there exists a constant T < ∞ such that for
every n ∈ N and for every x1, x2, ...., xn ∈ X,
kxjkp
X.
(4)
The smallest possible constants C, T in (3), (4) are denoted Cp(X), Tp(X), respectively. We
refer to [16, 9] for more information on the notions of type and cotype, though the present
paper requires minimal background of this theory. We shall use throughout standard Banach
space notation and terminology, as appearing in, say, [21].
The following theorem was proved in [13]:
Theorem 1.1 ([13]). A Banach space (X,k · kX) has Rademacher cotype q if and only if it
has metric cotype q.
Thus, for Banach spaces the linear notion of Rademacher cotype q is equivalent to the
notion of metric cotype q, which ignores all the structure of the Banach space except for
its metric properties. Theorem 1.1 belongs to a comprehensive program, first formulated
by Bourgain in [2], which is known as the Ribe program, whose goal is to recast the local
theory of Banach spaces as a purely metric theory. A byproduct of this program is that
linear properties such as Rademacher cotype can be made to make sense in general metric
spaces, with applications to metric geometry in situations which lack any linear structure.
We refer to [13] and the references therein for more information on the Ribe program and
its applications.
Definition (2) and Theorem 1.1 suppress the value of m, since it is irrelevant for the
purpose of a metric characterization of Rademacher cotype. Nevertheless, good bounds on
m are important for applications of metric cotype to embedding theory, some of which will
be recalled in Section 1.3. It was observed in [13] that if the metric space M contains at
least two points then the value of m in (2) must satisfy m & n1/q (where the implied constant
depends only on Γ). If X is a Banach space with Rademacher type p > 1 and Rademacher
cotype q, then it was shown in [13] that X satisfies the metric cotype q inequality (2) for
every m > n1/q (in which case Γ depends only on p, q, Tp(X), Cq(X)). Such a sharp bound
2
on m is crucial for certain applications [13, 14] of metric cotype, and perhaps the most
important open problem in [13] is whether this sharp bound on m holds true even when the
condition that X has type p > 1 is dropped. The bound on m from [13] in Theorem 1.1 is
m & n2+ 1
q . Our main result improves this bound to m & n1+ 1
q :
Theorem 1.2. Let X be a Banach space with Rademacher cotype q > 2. Then for every
n ∈ N, every integer m > 6n1+ 1
n
Xj=1
Ex"(cid:13)(cid:13)(cid:13)
m
2
f(cid:16)x +
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
q which is divisible by 4, and every f : Zn
q
m → X, we have
X# .X mq Eε,xhkf (x + ε) − f (x)kq
Xi.
(5)
In (5), and in what follows, .X , &X indicate the corresponding inequalities up to constants
which may depend only on q and Cq(X). Similarly, we will use the notation .q, &q to indicate
the corresponding inequalities up to constants which may depend only on q.
Though a seemingly modest improvement over the result of [13], the strengthened metric
cotype inequality (5) does yield some new results in embedding theory, as well as a new proof
of a result of Bourgain [3]; these issues are discussed in Section 1.3. More importantly, our
proof of Theorem 1.2 is based on a better understanding and sharpening of the underlying
principles behind the proof of Theorem 1.1 in [13]. As a result, we isolate here the key
approach to the metric characterization of Rademacher cotype in [13], yielding a simpler and
clearer proof of Theorem 1.1, in addition to the improved bound on m. This is explained
in detail in Section 1.1. While the bound m & n1+ 1
is far from the conjectured optimal
bound m > n1/q, our second main result is that (a significant generalization of) the scheme
for proving Theorem 1.1 and Theorem 1.2 (which is implicit in [13] and formulated explicitly
here) cannot yield a bound better than m & n
q . Our method for proving this lower bound
is presented in Section 4, and might be of independent interest.
2 + 1
1
q
We remark in passing that in [13] a one parameter family of variants of the notion of
metric cotype is studied, corresponding to raising the distances to powers other than q, and
modifying the right-hand side of (2), (5) accordingly (we refer to [13] for more details).
The argument presented here can be modified to yield simplifications and improvements of
all the corresponding variants of Theorem 1.1. While these variants are crucial for certain
applications of metric cotype [13, 11], we chose to present Theorem 1.2 only for the simplest
"vanilla" version of metric cotype (2), for the sake of simplicity of exposition.
Notation for measures. Since our argument uses a variety of averaging procedures over
several spaces, it will be convenient to depart from the expectation notation that we used
thus far. In particular, throughout this paper µ will denote the uniform probability measure
on Zn
m (m, n will always be clear from the context), σ will denote the uniform probability
measure on {−1, 0, 1}n, and τ will denote the uniform probability measure on {−1, 1}n.
1.1. The smoothing and approximation scheme. We start with a description of an
abstraction of the approach of [13] to proving the metric characterization of Rademacher
cotype of Theorem 1.1.
For a Banach space X, a function f : Zn
use the standard notation for the convolution f ∗ ν : Zn
m → X and a probability measure ν on Zn
m, we
m → X:
f (x − y)dν(y).
f ∗ ν(x) =ZZn
m
3
Assume that we are given n probability measures ν1, . . . , νn on Zn
m, and two additional
probability measures β1, β2 on the pairs in Zn
m of ℓ∞ distance 1, i.e., on the set
E∞(Zn
m) def= n(x, y) ∈ Zn
m × Zn
m : x − y ∈ {−1, 0, 1}no.
m × Zn
For A, S, q > 1, we shall say that the measures ν1, . . . , νn, β1, β2 are a (q, A, S)-smoothing and
approximation scheme on Zn
m → X
we have the following two inequalities:
m if for every Banach space (X,k · kX) and every f : Zn
m kf ∗ νj(x) − f (x)kq
X dµ(x) 6 AqZE∞(Zn
m) kf (x) − f (y)kq
Xdβ1(x, y).
(6)
(7)
(8)
(A) Approximation property:
(S) Smoothing property:
n
1
n
Xj=1ZZn
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ZZn
Xj=1
n
q
εj(cid:16)f ∗ νj(x + ej) − f ∗ νj(x − ej)(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
6 SqZE∞(Zn
X
dτ (ε)dµ(x)
m) kf (x) − f (y)kq
Xdβ2(x, y).
Often, when the underlying space Zn
m is obvious from the context, we will not mention it
explicitly, and simply call ν1, . . . , νn, β1, β2 a (q, A, S)-smoothing and approximation scheme.
In some cases, however, it will be convenient to mention the underlying space Zn
m so as to
indicate certain restrictions on m.
We introduce these properties for the following simple reason. We wish to deduce the
metric cotype inequality (5) from the Rademacher cotype inequality (3).
In essence, the
Rademacher cotype condition (3) is the same as the metric cotype inequality (5) when
restricted to linear mappings. This statement is not quite accurate, but it suffices for the
purpose of understanding the intuition behind the ensuing argument; we refer to Section 5.1
in [13] for the precise argument. In any case, it stands to reason that in order to prove (5)
from (3), we should first smooth out f , so that it will be locally well approximated (on
average) by a linear function. As we shall see momentarily, it turns out that the appropriate
way to measure the quality of such a smoothing procedure is our smoothing property (8).
Of course, while the averaging operators corresponding to convolution with the measures
ν1, . . . , νn yield a better behaved function, we still need the resulting averaged function to
be close enough to the original function f , so as to deduce a meaningful inequality such
as (5) for f itself. Our approximation property (7) is what's needed for carrying out such
an approach.
The above general scheme is implicit in [13]. Once we have isolated the crucial approxima-
tion and smoothing properties, it is simple to see how they relate to metric cotype. For this
purpose, assume that the Banach space X has Rademacher cotype q, and for each x ∈ Zn
apply the Rademacher cotype q inequality to the vectors {f ∗ νj(x + ej) − f ∗ νj(x − ej)}n
(where the averaging in (3) is with respect to ε ∈ {−1, 1}n, rather than ε ∈ {−1, 0, 1}n; it is
an easy standard fact that these two variants of Rademacher cotype q coincide):
j=1
m
4
n
Z{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xj=1
εj(cid:16)f ∗ νj(x + ej) − f ∗ νj(x − ej)(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xj=1
&X
n
q
X
dτ (ε)
kf ∗ νj(x + ej) − f ∗ νj(x − ej)kq
X .
(9)
The triangle inequality, combined with the convexity of the function t 7→ tq, implies that for
every x ∈ Zn
m and j ∈ {1, . . . , n} we have
m
(cid:13)(cid:13)(cid:13)
f(cid:16)x +
2
+ 3q−1kf ∗ νj(x) − f (x)kq
X.
m
2
m
2
(10)
X
q
q
f ∗ νj(cid:16)x +
ej(cid:17) − f(cid:16)x +
At the same time (recalling that m is divisible by 4), a combination of the triangle inequality
and Holder's inequality bounds the first term in the right hand side of (10) as follows:
X
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
+ 3q−1(cid:13)(cid:13)(cid:13)
6 3q−1(cid:13)(cid:13)(cid:13)
f ∗ νj(cid:16)x +
m
2
q
ej(cid:17) − f ∗ νj(x)(cid:13)(cid:13)(cid:13)
ej(cid:17)(cid:13)(cid:13)(cid:13)
X
m
2
(cid:13)(cid:13)(cid:13)
f ∗ νj(cid:16)x +
q
X
m/4
6
ej(cid:17) − f ∗ νj(x)(cid:13)(cid:13)(cid:13)
Xt=1
6(cid:16)m
4(cid:17)q−1
Xt=1
kf ∗ νj(x + 2tej) − f ∗ νj(x + 2(t − 1)ej)kq
X .
kf ∗ νj(x + 2tej) − f ∗ νj(x + 2(t − 1)ej)kX
m/4
(11)
q
n
Substituting (11) into (10), summing up over j ∈ {1, . . . , n}, and integrating with respect
to x ∈ Zn
m while using the translation invariance of the measure µ, we deduce the inequality
Xj=1ZZn
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
Xj=1ZZn
m kf ∗ νj(x + ej) − f ∗ νj(x − ej)kq
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
m kf ∗ νj(x) − f (x)kq
dµ(x) . 3q
X dµ(x).
X dµ(x)
+ mq
m
2
(12)
X
n
n
q
Xj=1ZZn
We can now bound the first term in the right hand side of (12) using the approximation
property (7), and the second term in the right hand side of (12) using (9) and the smoothing
property (8). The inequality thus obtained is
n
Xj=1ZZn
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
m
2
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
q
X
dµ(x)
.X (nAq + mqSq)ZE∞(Zn
m) kf (x) − f (y)kq
Xdβ3(x, y),
(13)
where β3 = (β1 + β2)/2. Note in passing that when m . A, an inequality such as (13), with
perhaps a different measure β3 on E∞(Zn
m), is a consequence of the triangle inequality, and
therefore holds trivially on any Banach space X. Thus, for our purposes, we may assume
throughout that a (q, A, S)-smoothing and approximation scheme on Zn
m satisfies m & A.
5
Assuming that
inequality (13) becomes
m &
A
S · n1/q,
(14)
(15)
n
Xj=1ZZn
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
m
2
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
q
X
dµ(x) .X SqmqZE∞(Zn
m) kf (x) − f (y)kq
Xdβ3(x, y).
If we could come up with a smoothing and approximation scheme for which S . 1, and m
satisfied (14), then inequality (15) would not quite be the desired metric cotype inequality (5),
but it would be rather close to it. The difference is that the probability measure β3 is
not uniformly distributed on all ℓ∞ edges E∞(Zn
m), as required in (5). Nevertheless, for
many measures β3, elementary triangle inequality and symmetry arguments can be used to
"massage" inequality (15) into the desired inequality (5). This last point is a technical issue,
but it is not the heart of our argument: we wish to design a smoothing and approximation
scheme satisfying S . 1 with A as small as possible. In [13] such a scheme was designed
with A . n2. Here we carefully optimize the approach of [13] to yield a smoothing and
approximation scheme with A . n, in which case (14) becomes the desired bound m & n1+ 1
q .
The bounds that we need in order to establish this improved estimate on m are based
on the analysis of some quite delicate cancelations; indeed the bounds that we obtain are
sharp for our smoothing and approximation scheme, as discussed in Section 1.2. In proving
such sharp bounds, a certain bivariate extension of the Bernoulli numbers arises naturally;
these numbers, together with some basic asymptotic estimates for them, are presented in
Section 3.1. The cancelations in the Rademacher sums corresponding to our convolution
kernels are analyzed via certain combinatorial identities in Section 3.2.
1.2. A lower bound on smoothing and approximation with general kernels. One
might wonder whether our failure to prove the bound m & n1/q without the non-trivial
Rademacher type assumption is due to the fact we chose the wrong smoothing and ap-
proximation scheme. This is not the case. In Section 4 we show that any approach based
on smoothing and approximation is doomed to yield a sub-optimal dependence of m on
n (assuming that the conjectured n1/q bound is indeed true). Specifically, we show that
for any (q, A, S)-smoothing and approximation scheme on Zn
m, with m & A, we must have
AS &q √n. Thus the bound S . 1 forces the bound A &q √n, and correspondingly (14)
q . Additionally, we show in Section 4 that for the specific smoothing and
becomes m &q n
approximation scheme used here, the bound m & n1+ 1
q is sharp.
1
2 + 1
It remains open what is the best bound on m that is achievable via a smoothing and
approximation scheme. While this question is interesting from an analytic perspective, our
current lower bound shows that we need to use more than averaging with respect to positive
measures in order to prove the desired bound m & n1/q.
1
2 + 1
Note that the lower bound m &q n
q for smoothing and approximation schemes rules
out the applicability of this method to some of the most striking potential applications of
metric cotype to embedding theory in the coarse, uniform, or quasisymmetric categories,
as explained in Section 1.3; these applications rely crucially on the use of a metric cotype
inequality with m ≍ n1/q.
6
The cancelation that was exploited in [13] in order to prove the sharp bound on m in the
presence of non-trivial Rademacher type was also related to smoothing properties of convolu-
tion kernels, but with respect to signed measures: the smoothed Rademacher sums in the left
hand side of (8) are controlled in [13] via the Rademacher projection, and the corresponding
smoothing inequality (for signed measures) is proved via an appeal to Pisier's K-convexity
theorem [15]. It would be of great interest to understand combinatorially/geometrically the
cancelations that underly the estimate m > n1/q from [13], though there seems to be a lack of
methods to handle smoothing properties of signed convolution kernels in spaces with trivial
Rademacher type and finite Rademacher cotype.
dX (x,z)(cid:17).
dY (f (x),f (z) 6 η(cid:16) dX (x,y)
1.3. The relation to nonembeddability results and some open problems. We recall
some standard terminology. Let (X, dX) and (Y, dY ) be metric spaces. X is said to embed
with distortion D into Y if there exists a mapping f : X → Y and (scaling factor) λ > 0,
such that for all x, y ∈ X we have λdX(x, y) 6 dY (f (x), f (y)) 6 DλdX(x, y). X is said to
embed uniformly into Y if there exists an into homeomorphism f : X → Y such that both
f and f −1 are uniformly continuous. X is said to embed coarsely into Y if there exists a
mapping f : X → Y and two non-decreasing functions α, β : [0,∞) → [0,∞) such that
limt→∞ α(t) = ∞, and for all x, y ∈ X we have α(dX(x, y)) 6 dY (f (x), f (y)) 6 β(dX(x, y)).
X is said to admit a quasisymmetric embedding into Y if there exists a mapping f : X → Y
and an increasing (modulus) η : (0,∞) → (0,∞) such that for all distinct x, y, z ∈ X we
have dY (f (x),f (y))
For a Banach space X, let qX denote the infimum over those q > 2 such that X has
Rademacher (equiv. metric) cotype q. It was shown in [13, 14] that if X, Y are Banach spaces,
Y has Rademacher type p > 1, and X embeds uniformly, coarsely, or quasisymmetrically
into Y , then qX 6 qY . Thus, under the Rademacher type > 1 assumption on the target
space, Rademacher cotype q is an invariant that is stable under embeddings of Banach
spaces, provided that the embedding preserves distances in a variety of (seemingly quite
weak) senses. The role of the assumption that Y has non-trivial Rademacher type is via the
metric cotype inequality with optimal m: the proofs of these results only use that Y satisfies
the metric cotype q inequality (2) for some m ≍ n1/q (under this assumption, Y can be a
general metric space and not necessarily a Banach space). This fact motivates our conjecture
that for any Banach space Y with Rademacher cotype q, the metric cotype inequality (5)
holds for every m &Y n1/q. The same assertion for general metric spaces of metric cotype q
is too much to hope for; see [19].
Perhaps the simplest Banach spaces for which we do not know how to prove a sharp metric
cotype inequality are L1 and the Schatten-von Neumann trace class S1 (see, e.g., [21]). Both
of these spaces have Rademacher cotype 2 (for S1 see [18]), yet the currently best known
bound on m in the metric cotype inequality (5) (with q = 2) for both of these spaces is
the bound m & n3/2 obtained here. The above embedding results in the uniform, coarse
or quasisymmetric categories do hold true for embeddings into L1 (i.e., a Banach space X
that embeds in one of these senses into L1 satisfies qX = 2). This fact is due to an ad-hoc
argument, which fails for S1 (see Section 8 in [13] for an explanation). We can thus ask the
following natural questions (many of which were already raised in [13]):
Question 1. Can Lr admit a uniform, coarse, or quasisymmetric embedding into S1 when
r > 2? More ambitiously, can a Banach space X with qX > 2 embed in one of these senses
7
into S1? In greatest generality: can a Banach space X embed in one of these senses into a
Banach space Y with qY < qX ?
If a Banach space X admits a uniform or coarse embedding into S1, then X must have finite
cotype. This fact, which could be viewed as a (non-quantitative) step towards Question 1,
was communicated to us by Nigel Kalton. To prove it, note that it follows from [17, Lem. 3.2]
that for any ultrapower (S1)U of S1, the unit ball of (S1)U is uniformly homeomorphic to a
subset of Hilbert space. Thus (S1)U has Kalton's property Q (see [8] for a detailed discussion
of this property). If the unit ball of X is uniformly homeomorphic to a subset of S1 (resp. X
admits a coarse embedding into S1), then the unit ball in any ultrapower of X is uniformly
homeomorphic to a subset of (S1)U (resp. any ultrapower of X admits a coarse embedding
into (S1)U ). By the proof of [8, Thm. 4.2], it follows that any ultrapower of X has property
Q, and hence it cannot contain c0. Thus X cannot have infinite cotype by the Maurey-Pisier
theorem [10] and standard Banach space ultrapower theory (see [4, Thm. 8.12]).
Question 2. Does S1 admit a uniform, coarse, or quasisymmetric embedding into a Banach
space Y with Rademacher type p > 1? More ambitiously, does S1 embed in one of these
senses into Banach space Y with Rademacher type p > 1 and qY = 2? In greatest generality:
does every Banach space X embed in one of these senses into a Banach space Y with
Rademacher type p > 1? Perhaps we can even ensure in addition that qY = qX?
Question 2 relates to Question 1 since embeddings into spaces with type > 1 would allow
us to use the nonembeddability results of [13].
While the improved bound on m in Theorem 1.2 does not solve any of these fundamental
questions, it does yield new restrictions on the possible moduli of embeddings in the uniform,
coarse, or quasisymmetric categories.
Instead of stating our nonembedding corollaries in
greatest generality, let us illustrate our (modest) improved nonembeddability results for
snowflake embeddings of L4 into S1 (this is just an illustrative example; the method of [13]
yields similar results for embeddings of any Banach space X with qX > 2 into S1, and
S1 itself can be replaced by general Banach spaces of finite cotype). Take θ ∈ (0, 1) and
assume the metric space (L4,kx− ykθ
4) admits a bi-Lipschitz embedding into S1. Our strong
conjectures imply that this cannot happen, but at present the best we can do is give bounds
on θ. An application of Theorem 1.2 shows that θ 6 4/5, i.e., we have a definite quantitative
estimate asserting that a uniform embedding of L4 into S1 must be far from bi-Lipschitz.
The previous bound from [13] for S1 was m = n5/2, yielding θ 6 8/9. Our lower bound
shows that by using a smoothing and approximation scheme we cannot hope to get a bound
of θ < 2/3.
the ℓn
∞ metric. We denote this metric space by [m]n
Turning to bi-Lipschitz embeddings, consider the grid {0, 1, . . . , m}n ⊆ Rn, equipped with
∞. Bourgain [3] proved that if Y is a
qin
Banach space with Rademacher cotype q, then any embedding of hn1+ 1
into Y incurs
distortion &Y n1/q. The same result follows from Theorem 1.2, while the previous estimate
q(2q+1) for embeddings
on m from [13] only yields the weaker distortion lower bound of &Y n
into Y . The sharp bound on m from [13] when Y has Rademacher type > 1
∞ into Y incurs distortion & n1/q (where
the implied constant is now allowed to depend also on the Rademacher type parameters of
qin
of hn1+ 1
implies that in this case, any embedding of (cid:2)n1/q(cid:3)n
∞
∞
q+1
8
Y ). Our main conjecture implies the same improvement of Bourgain's result without the
assumption that Y has non-trivial Rademacher type.
Bourgain's theorem [3] is part of his more general investigation of embeddings of ε-nets
in unit balls of finite dimensional normed spaces. Bourgain's approach in [3] is based on
ideas similar to ours, that are carried out in the continuous domain. Specifically, given a
mapping f : [m]n
∞ → Y , he finds a mapping g : Rn → Y which is L-Lipschitz and close in
an appropriate sense (depending on L, m, n) to f on points of the grid [m]n
∞. Once this is
achieved, it is possible to differentiate g to obtain the desired distortion lower bound. Bour-
gain's approximate Lipschitz extension theorem (an alternative proof of which was found
in [1]) is a continuous version of a smoothing and approximation scheme; it seems plausi-
ble that our method in Section 4 for proving impossibility results for such schemes can be
used to prove similar restrictions on Bourgain's approach to approximate Lipschitz exten-
sion. When Y has non-trivial Rademacher type, the improvement in [13] over Bourgain's
nonembeddability result for grids is thus based on a more delicate cancelation than was used
in [3, 1].
Question 3. Is it true that for any Banach space Y of Rademacher cotype q, any embedding
of (cid:2)n1/q(cid:3)n
∞ into Y incurs distortion &Y n1/q (if true, this is a sharp bound). Specializing to
the Schatten-von Neumann trace class S1, we do not even know whether the distortion of
[√n]n
∞ in S1 is & √n. Theorem 1.2 implies a distortion lower bound of & n1/6, while the
bound on m from [13] only yields a distortion lower bound of & n1/10. Our results in Section 4
show that one cannot get a distortion lower bound asymptotically better than n1/4 by using
smoothing and approximation schemes.
We did not discuss here metric characterizations of Rademacher type. We refer to [12]
for more information on this topic. It turns out that our approach to Theorem 1.2 yields
improved bounds in [12] as well; see [5].
Acknowledgements. O. G. was partially supported by NSF grant CCF-0635078. M. M.
was partially supported by ISF grant no. 221/07, BSF grant no. 2006009, and a gift from
Cisco research center. A. N. was supported in part by NSF grants CCF-0635078 and CCF-
0832795, BSF grant 2006009, and the Packard Foundation.
2. Proof of Theorem 1.2
For n ∈ N denote [n] = {1, . . . , n}. When B ⊆ [n], and x ∈ ZB
m, we will sometimes slightly
m, with the understanding that for i ∈ [n]\ B
m, we denote by yB the restriction of y to the coordinates in B.
abuse notation by treating x as an element of Zn
we have xi = 0. For y ∈ Zn
As in [13], for j ∈ [n] and an odd integer k < m/2, we define S(j, k) ⊆ Zn
m by
S(j, k) def= ny ∈ [−k, k]n ⊆ Zn
m : yj is even ∧ ∀ℓ ∈ [n] \ {j} yℓ is oddo.
(16)
The parameter k will be fixed throughout the ensuing argument, and will be specified later.
For every j ∈ [n] let νj be the uniform probability measure on S(j, k). Following the notation
of [13], for a Banach space (X,k · kX ) and f : Zn
µ(S(j, k))ZS(j,k)
m → X, we write f ∗ νj = Ejf , that is,
Ejf (x) def=
f (x + y)dµ(y).
(17)
1
9
Recall that E∞(Zn
m), defined in (6), is the set of all ℓ∞ edges of Zn
m. Similarly, we denote
the ℓ1 edges of Zn
m by E1(Zn
m), i.e.,
E1(Zn
m) def= n(x, y) ∈ Zn
m × Zn
m : x − y ∈ {±e1, . . . ,±en}o.
(18)
Clearly E1(Zn
m) ⊆ E∞(Zn
m).
Let β◦
1 denote the uniform probability distribution on the pairs (x, y) ∈ E∞(Zn
1 denote the uniform probability distribution on E1(Zn
x − y ∈ {−1, 1}n, and let β◦◦
shall consider the probability measure on E∞(Zn
m) given by β1 = (β◦
1 + β◦◦
1 )/2.
m) with
m). We
Lemma 5.1 in [13] implies that for all q > 1 and f : Zn
m → X we have:
1
n
n
Xj=1ZZn
m kEjf − fkq
X dµ . (2k)qZE∞(Zn
m) kf (x) − f (y)kq
Xdβ1(x, y).
(19)
Inequality (19) corresponds to the approximation property (7), with A . k. The relevant
smoothing inequality is the main new ingredient in our proof of Theorem 1.2, and it requires
a more delicate choice of probability measure β2 on E∞(Zn
m) then
x − y ∈ {−1, 0, 1}n. Let S = {i ∈ [n] : xi = yi}, and define
(n/k)qS
If (x, y) ∈ E∞(Zn
m).
where Z is a normalization factor ensuring that β2 is a probability measure, i.e.,
β2(x, y) def=
1
Z ·
,
2n−Smn(cid:0) n
S(cid:1)
Xℓ=0(cid:16) n
k(cid:17)qℓ
≍ 1,
n
Z =
(20)
(21)
(22)
provided that, say,
k > 2n.
Our final choice of k will satisfy (22), so we may assume throughout that Z satisfies (21).
The key smoothing property of the averaging operators {Ej}n
j=1 is contained in the follow-
ing lemma:
Lemma 2.1. Let X be a Banach space, q > 1, n, m ∈ N, where m > 4n is divisible by 4,
and f : Zn
m → X. Suppose that k is an odd integer satisfying 2n 6 k < m
2 . Then,
n
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ZZn
Xj=1
where S .q 1.
εj [Ejf (x + ej) − Ejf (x − ej)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
q
X
dτ (ε)dµ(x)
6 SqZE∞(Zn
m) kf (x) − f (y)kq
Xdβ2(x, y),
(23)
We shall postpone the proof of Lemma 2.1 to Section 3, and proceed now to deduce
Theorem 1.2 assuming its validity. Before doing so, we recall for future use the following
simple lemma from [13]:
10
Lemma 2.2 (Lemma 2.6 from [13]). For every q > 1 and for every f : Zn
1
n
n
Xj=1ZZn
m kf (x + ej)− f (x)kq
Xdµ(x) . 2qZ{−1,0,1}nZZn
m kf (x + δ)− f (x)kq
m → X,
Xdµ(x)dσ(δ). (24)
Proof of Theorem 1.2. The argument in the introduction leading to (15), when specialized
to our smoothing and approximation scheme using (19) and (23), shows that if k > 2n and
m > 2kn1/q, then
n
Xj=1ZZn
m(cid:13)(cid:13)(cid:13)
m
2
f(cid:16)x +
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
where
q
X
dµ(x) .X mqZE∞(Zn
m) kf (x) − f (y)kq
Xdβ3(x, y),
(25)
β3 =
β1 + β2
2
6 β◦
1 + β◦◦
1 + β2.
Note that β◦
bound:
n
Xj=1ZZn
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
1 . β2 due to the contribution of S = ∅ in (20). Thus, (25) implies the following
m
2
q
X
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
+ mq XS⊆[n]
(cid:0) n
S(cid:1)
(n/k)qS
dµ(x) .X
n
mq
n
m kf (x + ej) − f (x)kq
Xj=1ZZn
Z{−1,1}[n]\SZZn
m kf (x + ε) − f (x)kq
X dµ(x)
Xdµ(x)dτ (ε),
(26)
where the first term in the right hand side of (26) corresponds to β◦◦
1 .
In order to deduce the desired metric cotype inequality (5) from (26), we shall apply (26)
m. Note that we are allowed to do so since our requirements
to lower dimensional sub-tori of Zn
on k, namely k > 2n and m > 2kn1/q, remain valid for smaller n.
m , we obtain
m . We can then consider the mapping g : ZB
Fix ∅ 6= B ⊆ [n] and x[n]\B ∈ Z[n]\B
m → X
given by g(xB) = f (x[n]\B, xB). Applying (26) to g, and averaging the resulting inequality
over x[n]\B ∈ Z[n]\B
Xj∈BZZn
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
m kf (x + ej) − f (x)kq
dµ(x) .X
Xdµ(x)
m
2
mq
X
q
m kf (x + ε) − f (x)kq
Xdµ(x)dτ (ε).
(27)
BXj∈BZZn
Z{−1,1}B\SZZn
def= 2B−1
3n−1 . Multiplying (27) by WB and summing over
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
(B/k)qS
+ mq XS⊆B
(cid:0)B
S(cid:1)
For B ⊆ [n] define the weight WB
∅ 6= B ⊆ [n], we obtain the bound
Xj=1ZZn
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
+ XB⊆[n]
ej(cid:17)− f (x)(cid:13)(cid:13)(cid:13)
(B/k)qS
WBXS⊆B
(cid:0)B
S(cid:1)
1
mq
m
2
B6=∅
X
n
q
dµ(x) .X XB⊆[n]
Z{−1,1}B\SZZn
B6=∅
WB
B Xj∈BZZn
m kf (x + ε) − f (x)kq
m kf (x + ej)− f (x)kq
X dµ(x)
Xdµ(x)dτ (ε),
(28)
11
where we used the identity
XB⊆[n]
m
2
f(cid:16)x +
WBXj∈BZZn
m(cid:13)(cid:13)(cid:13)
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
B Xj∈BZZn
m kf (x + ej) − f (x)kq
WB
XB⊆[n]
B6=∅
q
X
dµ(x) =
Xdµ(x) .
1
n
The first term in the right hand side of (28) is easy to bound, using Lemma 2.2, as follows:
q
X
dµ(x).
Xdµ(x)
n
n
m
2
Xj=1ZZn
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
Xj=1ZZn
m kf (x + ej) − f (x)kq
m kf (x + δ) − f (x)kq
ℓ−1(cid:1) 2ℓ−1
ℓ=1(cid:0)n−1
m kf (x + ε) − f (x)kq
aT ZZn
m kf (x + ε) − f (x)kq
3n−1 . 1
Xdµ(x)
Xdµ(x)dσ(δ),
(29)
n . To bound
Xdµ(x),
(30)
(24)
. 2qZ{−1,0,1}nZZn
(B/k)qS
S(cid:1)ZZn
2B−S(cid:0)B
3n XT ⊆[n] Xε∈{−1,1}T
1
.
the second term in the right hand side of (28), note that it equals
where in the first inequality of (29) we used the fact that Pn
C def= XS⊆[n] XS⊆B⊆[n]
B6=∅ Xε∈{−1,1}B\S
2B−1
3n−1 ·
where we used the change of variable T = B \ S, and for every T ⊆ [n] we write,
2ℓ−T (ℓ/k)q(ℓ−T )
.
2B−T (B/k)q(B−T )
=
n
aT
def= XB⊇T
(cid:0) B
B−T (cid:1)
ℓ − T(cid:19) ·
Xℓ=T (cid:18)n − T
v(cid:1)v
6 (cid:0)u
Fix T ⊆ [n]. Using the standard bounds (cid:0) u
v(cid:1) 6 (cid:0) eu
k(cid:19)q(ℓ−T )
ℓ (cid:19)ℓ−T (cid:18) ℓ
ℓ − T (cid:19)ℓ−T (cid:18)ℓ − T
0 6 v 6 u, we can bound aT as follows:
Xℓ=T (cid:18)e(n − T)
aT 6
2ℓ−T
n
(cid:0) ℓ
ℓ−T (cid:1)
v (cid:1)v, which hold for all integers
=
n
kq
Xℓ=T (cid:18)2e(n − T)ℓq−1
9 (cid:19)(ℓ−T )
. 1.
(cid:19)ℓ−T
.
(31)
Thus, assuming that k > 3n, and recalling that q > 2, we get the bound
aT 6
n
Xℓ=T (cid:18)2e(n − T)nq−1
(3n)q
(cid:19)ℓ−T
6
n
Xℓ=T (cid:18)2e
Combining (31) with (30), we see that the second term in the right hand side of (28) is
C .
1
3n XT ⊆[n] Xε∈{−1,1}TZZn
m kf (x + ε) − f (x)kq
X dµ(x)
=Z{−1,0,1}nZZn
m kf (x + δ) − f (x)kq
X dµ(x)dσ(δ).
12
In combination with (29), inequality (28) implies that
1
mq
n
Xj=1ZZn
m(cid:13)(cid:13)(cid:13)
f(cid:16)x +
m
2
q
X
ej(cid:17) − f (x)(cid:13)(cid:13)(cid:13)
dµ(x)
.X Z{−1,0,1}nZZn
m kf (x + δ) − f (x)kq
X dµ(x)dσ(δ),
which is precisely the desired inequality (5). Recall that in the above argument, our require-
ment on k was k > 3n, and our requirement on m was m > 2kn1/q (and that it is divisible
by 4). This implies the requirement m > 6n1+ 1
(cid:3)
q of Theorem 1.2.
3. Proof of Lemma 2.1
Lemma 2.1 is the main new ingredient of the proof of Theorem 1.2. Its proof is based on
combinatorial identities which relate the "smoothed out Rademacher sum"
n
Xj=1
εj [Ejf (x + ej) − Ejf (x − ej)]
(32)
to a certain bivariate extension of the Bernoulli numbers. We shall therefore first, in Sec-
tion 3.1, do some preparatory work which introduces these numbers and establishes estimates
that we will need in the ensuing argument. We shall then derive, in Section 3.2, certain com-
binatorial identities that relate (32) to the bivariate Bernoulli numbers. In Section 3.3 we
shall combine the results of Section 3.1 and Section 3.2 to complete the proof of Lemma 2.1.
3.1. Estimates for the bivariate Bernoulli numbers. There are two commonly used
definitions of the Bernoulli numbers {Br}∞
r=0. For more information on these two conventions,
we refer to http://en.wikipedia.org/wiki/Bernoulli_number. Here we shall refer to the
variant of the Bernoulli numbers that was originally defined by J. Bernoulli, for which B1 = 1
2,
and which is defined via the recursion
r−1
Xa=0
r =
a(cid:19).
Ba(cid:18)r
Observe that (33) contains the base case B0 = 1 when substituting r = 1. The recursion (33)
extends naturally to a bivariate sequence {Br,s}n
a(cid:19) −
Ba,s(cid:18)r
b(cid:19).
Br,b(cid:18)s
r,s=0, given by
r − s =
(33)
(34)
s−1
Xb=0
r−1
Xa=0
It is well-known (cf. [20, Sec. 2.5]) that the exponential generating function for {Br}∞
r=0 is
We shall require the following analogous computation of the bivariate exponential generating
function of {Br,s}n
r,s=0:
F (x) def=
xex
ex − 1
=
Br
xr
r!
.
∞
Xr=0
13
Lemma 3.1. For all x, y ∈ C with x,y < π we have
Xs=0
Xr=0
(x − y)ex+y
ex − ey =
F (x, y) def=
∞
∞
Br,s
xrys
r! · s!
,
(35)
where the series in (35) is absolutely convergent on {(x, y) ∈ C × C :
r < π.
x,y 6 r} for all
r=0P∞
s=0 zr,sxrys, for some {zr,s}∞
Proof. The function F (x, y) is analytic on Dπ = {(x, y) ∈ C × C :
x,y < π}, since its
only non-removable singularities are when x− y ∈ 2πi(Z \{0}). It follows that we can write
F (x, y) =P∞
r,s=0 ⊆ C, where the series converges absolutely
on any compact subset of Dπ (see, e.g., [6, Thm. 2.2.6]). Note that
(ex − ey)F (x, y) = ∞
Xn=1
n! ! ∞
xn − yn
Xr=0
∞
zr,sxrys!
Xs=0 r−1
Xa=0
∞
∞
Xs=0
Xr=0
=
za,s
(r − a)! −
s−1
Xb=0
zr,b
(s − b)!! xrys.
(36)
(37)
(cid:3)
(38)
At the same time,
(ex − ey)F (x, y) = (x − y)exey = (x − y)
∞
∞
Xr=0
Xs=0
xrys
r!s!
=
∞
∞
Xr=0
Xs=0
(r − s)
xrys
r!s!
.
By equating coefficients in (36) and (37), we see that for all r, s ∈ N ∪ {0},
r − s = r!s! r−1
Xa=0
za,s
(r − a)! −
s−1
Xb=0
zr,b
(s − b)!! =
r−1
Xa=0(cid:18)r
a(cid:19)a!s!za,s −
s−1
Xb=0(cid:18)s
b(cid:19)r!b!zr,b.
Since z0,0 = 1, the recursive definition (34) implies that zr,s = Br,s
r!s! , as required.
An immediate corollary of Lemma 3.1 is that since F (x, y) = F (y, x),
∀r, s ∈ N ∪ {0}, Br,s = Bs,r.
Another (crude) corollary of Lemma 3.1 is that since the power series in (35) converges
absolutely on {(x, y) ∈ C × C :
x,y 6 2}, for all but at most finitely many r, s ∈ N ∪ {0}
we have Br,s/(r!s!)1/(r+s) 6 1/2. Thus,
∀r, s ∈ N ∪ {0},
Remark 3.1. Since B2m = (−1)m−12ζ(2m)(2m)!
, where ζ(s) is the Riemann zeta function (and
B2m+1 = 0 for m > 1), one has the sharp asymptotics B2m ∼ 2(2m)!
(2π)2m for the classical
Bernoulli numbers. We did not investigate the question whether similar sharp asymptotics
can be obtained for the bivariate Bernoulli numbers.
Br,s .
(39)
(2π)2m
r!s!
2r+s .
14
3.2. Some combinatorial identities. We start by introducing some notation. For y ∈ Zn
write:
m
↑y↑ def= (cid:12)(cid:12){l : yl = k
↓y↓ def= (cid:12)(cid:12){l : yl = −k
lyl def= ↑y↑ + ↓y↓
def= ny ∈ [−k, k]n ⊆ Zn
m and ε ∈ {−1, 1}n, let x ⊙ ε ∈ Zn
(mod m)}(cid:12)(cid:12),
(mod m)}(cid:12)(cid:12),
↑y↓ def= ↑y↑ − ↓y↓ .
m : yt is odd ∀t ∈ [n]o.
and
S
We also define
m be the coordinate wise multiplication, i.e.,
where
Definition 3.2. For f : Zn
m → X, k < m
∆Bf (x) def=
For x ∈ Zn
(x ⊙ ε)j = xjεj. Also for ε, ε′ ∈ {−1, 1}n let hε, ε′i =Pn
j=1 εjε′
j.
We need to define additional auxiliary averaging operators.
2 odd, and B ⊆ [n], let
1
µ(LB)ZLB
f (x + y)dµ(y),
m : ∀i /∈ B, yi = 0 ∧ ∀i ∈ [n] yi is eveno.
m, ε ∈ {−1, 1}n, i ∈ [n], and 0 6 j 6 i,
(cid:16)1δk+ε[n]\S+L[n]\S (z) − 1δk−ε[n]\S+L[n]\S (z)(cid:17) ,
def= ny ∈ (−k, k)n ⊆ Zn
Definition 3.3. Define for z ∈ Zn
bi,j(z, ε) def= XS⊆[n]
S=i Xδ∈{−1,1}S
Xj=1
εj(cid:0)1ej+S(j,k)(z) − 1−ej+S(j,k)(z)(cid:1) ,
a(z, ε) def=
hδ,εSi=i−2j
LB
n
where we recall that S(j, k) was defined in (16).
(40)
(41)
The next lemma follows immediately from an inspection of our definitions.
Lemma 3.2. The following identities hold true:
Xy∈Zn
m
bi,j(y − x, ε)f (y)
= kn−i XS⊆[n]
S=i Xδ∈{−1,1}S
hδ,εSi=i−2j
(cid:16)∆[n]\Sf (x + δk + ε[n]\S) − ∆[n]\Sf (x + δk − ε[n]\S)(cid:17),
(42)
n
a(y − x, ε)f (y) = k(k + 1)n−1
Xy∈Zn
Claim 3.3. If there exits t ∈ [n] such that zt is even, then a(z, ε) = bi,j(z, ε) = 0 for all
ε ∈ {−1, 1}n.
εj(cid:0)Ejf (x + ej) − Ejf (x − ej)(cid:1).
Xj=1
(43)
m
15
Proof. This follows directly from the definitions of the sets S(j, k), and LB, since all values
of the coordinates are odd in all the points of the sets
δk + ε[n]\S + L[n]\S, δk − ε[n]\S + L[n]\S, ej + S(j, k), −ej + S(j, k),
for every S ⊆ [n], δ ∈ {−1, 1}S and ε ∈ {−1, 1}n.
(cid:3)
Claim 3.4. If zt is odd and either zt 6= ±k (mod m) for all t ∈ [n], or zt0 > k for some
t0 ∈ [n], then a(z, ε) = bi,j(z, ε) = 0.
Proof. We may assume that z ∈ [−m/2, m/2]n. If there is t0 ∈ [n] for which zt0 > k then
all the terms in the right hand side of (40) and (41) are 0. If zt < k for all t ∈ [n], then all
the terms in the right hand side of (40) and (41) cancel out.
(cid:3)
It follows that for z /∈ S we have a(z, ε) = bi,j(z, ε) = 0 for every ε ∈ {−1, 1}n and every
Xy∈x+S
0 6 j 6 i 6 n. Thus, in particular, identity (42) can be rewritten as:
(cid:16)∆[n]\Sf (x + δk + ε[n]\S) − ∆[n]\Sf (x + δk − ε[n]\S)(cid:17).
bi,j(y − x, ε)f (y)
= kn−i XS⊆[n]
S=i Xδ∈{−1,1}S
hδ,εSi=i−2j
Note that the definition (41) shows that for z ∈ S we have
a(z, ε) = Xt∈[n]
zt=k
εt − Xt∈[n]
zt=−k
εt = ↑z ⊙ ε↓ .
(44)
(45)
(46)
Using Claim 3.3 and Claim 3.4, in conjunction with (43) and (45), we conclude that:
Lemma 3.5. The following identity holds for all x ∈ Zn
m and ε ∈ {−1, 1}n:
n
Xj=1
εj(cid:0)Ejf (x + ej) − Ejf (x − ej)(cid:1) =
1
k(k + 1)n−1 Xy∈x+Sx(y − x) ⊙ εy f (y).
Lemma 3.6. If z ∈ S and i > lzl then ∀j ∈ {0, . . . , i} and ∀ε ∈ {−1, 1}n, we have
bi,j(z, ε) = 0.
Proof. If i > lzl then
z /∈(cid:16)δk + ε[n]\S + L[n]\S(cid:17)[(cid:16)δk − ε[n]\S + L[n]\S(cid:17),
for every S ⊆ [n] with S = i, and δ ∈ {−1, 1}S. If i = lzl then there exists exactly one
subset S ⊆ [n] in (40) where z can appear, namely S = {ℓ ∈ [n] : zℓ ∈ {−k, k}}. If
for some δ ∈ {−1, 1}S, then
z ∈(cid:16)δk + ε[n]\S + L[n]\S(cid:17)[(cid:16)δk − ε[n]\S + L[n]\S(cid:17),
z ∈(cid:16)δk + ε[n]\S + L[n]\S(cid:17)\(cid:16)δk − ε[n]\S + L[n]\S(cid:17),
since for all coordinates i ∈ [n] \ S we have zi < k. Hence in this case the terms in the sum
in the right hand side of (40) cancel out.
(cid:3)
16
Lemma 3.7. For every z ∈ S, ε ∈ {−1, 1}n, and 0 6 j 6 i < lzl,
bi,j(z, ε) =
Proof. By looking at the elements of
(cid:0)lzl−j
i−j (cid:1)
−(cid:0)lzl−(i−j)
(cid:1)
0
j
↓z ⊙ ε↓ = j,
↑z ⊙ ε↑ = i − j,
otherwise.
(cid:16)δk + ε[n]\S + L[n]\S(cid:17)[(cid:16)δk − ε[n]\S + L[n]\S(cid:17),
it is clear that we must have S ⊆ {h : zh ∈ {−k, k}} in order to get a nonzero contribution
to the right hand side of (40). For such an S there is at most one δ ∈ {−1, 1}S which
can contribute to the sum, namely δh = sgn(zh) for every h ∈ S. But since this δ should
also satisfy hδ, εSi = i − 2j, we conclude that a non-zero contribution can occur only when
↓zS ⊙ εS↓ = j. In those cases, there is an actual contribution only if either sgn(zhεh) = 1
for every h ∈ {ℓ : zℓ ∈ {−k, k}}\ S, or sgn(zhεh) = −1 for every h ∈ {ℓ : zℓ ∈ {−k, k}}\ S,
and those contributions have different signs. The claim now follows.
(cid:3)
The following lemma relates, via Lemma 3.7, what we have done so far to the bivariate
Bernoulli numbers.
Lemma 3.8. There exists a sequence {hα,β}06α6n
ε ∈ {−1, 1}n,
06β6α ⊆ R such that for all y ∈ Zn
m and all
n
α
↑y ⊙ ε↓ =
hα,βbα,β(y, ε),
, for all 0 6 β 6 α
Xβ=0
Xα=0
(α − β)!β!
hα,β .
2α
hα,β = hα,α−β.
(47)
(48)
(49)
Proof. Write r = ↑z ⊙ ε↑ and s = ↓z ⊙ ε↓. Thus r + s = lzl and r − s = ↑z ⊙ ε↓. With this
notation, if we substitute the values of bα,β(y, ε) from Lemma 3.7, the desired identity (47)
becomes:
r − s =
n
Xα=s
hα,s(cid:18) r
α − s(cid:19) −
n
Xβ=0
hβ+r,β(cid:18)s
β(cid:19) (♠)
=
n−s
Xa=0
n
ha+s,s(cid:18)r
Xa=0
b(cid:19)
hb+r,b(cid:18)s
a(cid:19) −
Xb=0
b(cid:19),
hb+r,b(cid:18)s
a(cid:19) −
ha+s,s(cid:18)r
Xb=0
(♣)
=
r−1
s−1
(50)
where in (♠) we used the change of variable β = b, α = a + s, and in (♣) we noted that
r + s = lzl 6 n and that the terms corresponding to a > r or b > s vanish, while the terms
corresponding to a = r and b = s cancel out. Thus, the desired identity (50) shows that we
must take ha+b,b = Ba,b, or hα,β = Bα−β,β. The bound (48) is now the same as (39), and the
identity (49) is the same as (38).
(cid:3)
17
3.3. Putting things together. We are now ready to complete the proof of Lemma 2.1
using the tools developed in the previous two sections.
be the sequence from Lemma 3.8. Then for all f : Zn
m → X
q
X dµ(x).
(51)
06β6α
Lemma 3.9. Let {hα,β}06α6n
and all ε ∈ {−1, 1}n we have,
m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ZZn
Xj=0
k(k + 1)n−1 n
Xi=0
1
i
hi,j Xy∈x+S
Xℓ=0
.q
n
q
X
S=ℓ
(n/k)ℓq
dµ(x)
bi,j(y − x, ε)f (y)!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ZZn
ℓ(cid:1) XS⊆[n]
m(cid:13)(cid:13)f (x + ε[n]\S) − f (x)(cid:13)(cid:13)
(cid:0)n
bi,j(y − x, ε)f (y)!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X
2i+1hi,jDi,j(x)!q
Xi=0
Xj=0
Xj=0
(∗∗)
6
n
.
i
i
Proof. For every x ∈ Zn
m and 0 6 j 6 i 6 n write
k(k + 1)n−1 Xy∈x+S
Note that,
1
Di,j(x) def= (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
hi,jDi,j(x)!q
= n
n
Xi=0
Xj=0
Xi=0
2i+1hi,jDi,j(x)!q
2−(i+1) i
Xj=0
Xi=0
2−(i+1)
(∗)
6
n
i
2(i+1)(q−1)(i + 1)q−1hi,jqDi,j(x)q,
(52)
where in (∗) we used the convexity of the function t 7→ tq and that P∞
i=0 2−(i+1) = 1, and
in (∗∗) we used Holder's inequality. It follows from (52), combined with the bound (48) on
hi,j, that,
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
bi,j(y − x, ε)f (y)
k(k + 1)n−1
Xi=0
hi,jDi,j(x)!q
6 n
Xi=0
hi,j Xy∈Zn
Xj=0
Xj=0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1
X
n
n
m
q
i
i
i
Di,j(x)q.
(53)
2(i+1)(q−1)(i + 1)q−1(cid:18)(i − j)!j!
2i
(cid:19)q
.
Xi=0
Xj=0
Now, Di,j(x) can be estimated using the identity (44) as follows:
kiDi,j(x) 6
(k + 1)n−1
kn−i−1 Di,j(x)
6 XS⊆[n]
S=i Xδ∈{−1,1}S
hδ,εSi=i−2j
(cid:13)(cid:13)(cid:13)
∆[n]\Sf (x + δk + ε[n]\S) − ∆[n]\Sf (x + δk − ε[n]\S)(cid:13)(cid:13)(cid:13)X
.
(54)
18
Di,j(x)q 6
Note that the number of terms in the sum in the right hand side of (54) is (cid:0)n
i(cid:1)(cid:0)i
j(cid:1). Thus
1
j(cid:19)q−1
i(cid:19)q−1(cid:18)i
kiq(cid:18)n
(cid:13)(cid:13)(cid:13)
∆[n]\Sf (x + δk + ε[n]\S) − ∆[n]\Sf (x + δk − ε[n]\S)(cid:13)(cid:13)(cid:13)
· XS⊆[n]
S=i Xδ∈{−1,1}S
hδ,εSi=i−2j
q
X
.
(55)
If we integrate inequality (55) with respect to x, use the translation invariance of µ to
eliminate the additive term δk in the argument of the integrands, and use the fact that ∆B
is an averaging operator for all B ⊆ [n], we obtain the bound
ZZn
Di,j(x)qdµ(x) 6
kiq(cid:18)n
dµ(x)
1
X
m
q
6
2q
S=i
i(cid:19)q−1(cid:18)i
j(cid:19)q
ZZn
m(cid:13)(cid:13)(cid:13)
f (x + ε[n]\S) − f (x − ε[n]\S)(cid:13)(cid:13)(cid:13)
XS⊆[n]
i(cid:19)q−1(cid:18)i
j(cid:19)q
kiq(cid:18)n
ZZn
f (x + ε[n]\S) − f (x)(cid:13)(cid:13)(cid:13)
m(cid:13)(cid:13)(cid:13)
XS⊆[n]
+ 2q−1(cid:13)(cid:13)(cid:13)
f (x + ε[n]\S)− f (x)(cid:13)(cid:13)(cid:13)
6 2q−1(cid:13)(cid:13)(cid:13)
S=i
X
X
q
q
q
X
f (x)− f (x− ε[n]\S)(cid:13)(cid:13)(cid:13)
q
X
,
dµ(x),
(56)
where in the last step of (56) we used the triangle inequality as follows:
f (x + ε[n]\S)− f (x− ε[n]\S)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)
while noticing that upon integration with respect to x, by translation invariance, both terms
become equal.
Integrating (53) with respect to x, and using (56), we see that that the left hand side
of (51) is at most
n
i
Xi=0
Xj=0
j(cid:1)(cid:17)q
2(i+1)(q−1)+q(i + 1)q−1(cid:16) (i−j)!j!
2iki (cid:0)n
i(cid:1)(cid:0)i
(cid:0)n
i(cid:1)
ki(n − i)!(cid:19)q
i(cid:1) (cid:18)
2i(cid:0)n
Xi=0
= 22q−1
(i + 1)q
n!
n
S=i
ZZn
f (x + ε[n]\S) − f (x)(cid:13)(cid:13)(cid:13)
m(cid:13)(cid:13)(cid:13)
XS⊆[n]
ZZn
f (x + ε[n]\S) − f (x)(cid:13)(cid:13)(cid:13)
m(cid:13)(cid:13)(cid:13)
XS⊆[n]
S=i
X
q
q
X
dµ(x)
dµ(x).
(57)
Inequality (57) implies the desired bound (51), since (i+1)q2−i .q 1 and n!/(n−i)! 6 ni. (cid:3)
Proof of Lemma 2.1. It follows from (43) and (46) that
m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Z{−1,1}nZZn
q
n
dµ(x)dτ (ε)
εj [Ejf (x + ej) − Ejf (x + ej)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xl=1
m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=Z{−1,1}nZZn
k(k + 1)n−1 Xy∈x+Sx(y − x) ⊙ εy f (y)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1
X
q
X
19
dµ(x)dτ (ε).
(58)
An application of identity (47) now shows that
1
m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
k(k + 1)n−1 Xy∈x+Sx(y − x) ⊙ εy f (y)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Z{−1,1}nZZn
m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
k(k + 1)n−1 n
=Z{−1,1}nZZn
Xi=0
hi,j Xy∈x+S
Lemma 2.1 now follows from Lemma 3.9.
Xj=0
1
i
q
X
dµ(x)dτ (ε)
bi,j(y − x, ε)f (y)!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
q
X
dµ(x)dτ (ε).
(59)
(cid:3)
4. Lower bounds
In this section we establish lower bounds for the best possible value of m in Theorem 1.2
that is achievable via a smoothing and approximation scheme. Our first result deals with
general convolution kernels:
Proposition 4.1. Assume that the probability measures ν1, . . . , νn, β1, β2 are a (q, A, S)-
smoothing and approximation scheme on Zn
m, i.e., conditions (7) and (8) are satisfied for
every Banach space X and every f : Zn
m → X. Assume also that m > cA for a large enough
universal constant c > 0. Then
S &q
.
(60)
√n
A
Recall, as explained in Section 1.1, that in order for a smoothing and approximation
scheme to yield the metric cotype inequality (5), we require S . 1, in which case the bound
on m becomes m & An1/q. Proposition 4.1 shows that S . 1 forces the bound A &q √n,
and correspondingly m &q n
2 + 1
q .
1
For the particular smoothing and approximation scheme used in our proof of Theorem 1.2,
the following proposition establishes asymptotically sharp bounds.
Proposition 4.2. Fix an odd integer k 6 m/2 and consider the averaging operators {Ej}n
j=1
used in our proof of Theorem 1.2, i.e., they are defined as in (17).
If there exist prob-
ability measures β1, β2 on E∞(Zn
m) for which the associated approximation and smoothing
inequalities (7) and (8) are satisfied for every Banach space X and every f : Zn
m → X, then
A & k
and S & min(cid:26)r n
k
,
n
k(cid:27) .
(61)
Proposition 4.2 shows that in order to have S . 1 we need to require k & n, in which case
A & n, and correspondingly m & n1+ 1
q , matching the bound obtained in Theorem 1.2.
4.1. A lower bound for general convolution kernels: Proof of Proposition 4.1.
Assume that the probability measures ν1, . . . , νn, β1, β2 are a (q, A, S)-smoothing and ap-
proximation scheme, i.e., they satisfy (7) and (8). It will be convenient to think of these
measures as functions defined on the appropriate (finite) spaces, i.e., ν1, . . . , νn : Zn
m → [0, 1]
and β1, β2 : E∞(Zn
m) → [0, 1].
20
For a probability measure ν on Zn
m, let Pj(ν) be the probability measure on Zm which is
the marginal of ν on the jth coordinate, i.e.,
Pj(ν)(r) def= Xx∈Zn
m
xj =r
ν(x).
Define the absolute value of x ∈ Zm to be x = min{x, m − x}.
Lemma 4.3. Assume that ν1, . . . , νn, β1 satisfy (7). Then for every s ∈ N we have:
n
1
n
A
s
.
νj(x) .
Xj=1 Xx∈Zn
∞. Let gs : R → R be the truncated jigsaw function
m
xj>s
(62)
Proof. We shall apply (7) with X = ℓn
with period 12s, depicted in Figure 1.
gs(x)
4s
s
5s
7s
12s
x
Figure 1. gs is truncated jigsaw function.
Define fs : Zn
m → X by
fs(x) def= (gs(x1), gs(x2), . . . , gs(xn)).
The Lipschitz constant of fs with respect to the ℓ∞ metric on Zn
from (7) that
m is 1, and therefore it follows
n
n
1
Xj=1ZZn
For every x ∈ Zn
(fs ∗ νj − fs)(x) = Xy∈Zn
m
m kfs ∗ νj − fskℓn
m and j ∈ [n],
∞
Assume that
dµ!q
6
1
n
n
Xj=1ZZn
m kfs ∗ νj − fskq
ℓn
∞
dµ . Aq.
(63)
νj(y)(fs(x − y) − fs(x))
= Xy∈Zn
(xj mod 12s) ∈ [0, s] ∪ [12s − s, 12s − 1].
m
νj(y)(cid:16)gs(x1 − y1) − gs(x1), . . . , gs(xn − yn) − gs(xn)(cid:17).
(64)
21
When 3s 6 yj 6 4s, we have gs(xj − yj) − gs(xj) > s, and for every yj ∈ Zm, we have
gs(xj − yj) − gs(xj) > 0. Hence,
∞ > Xy∈Zn
k(fs ∗ νj − fs)(x)kℓn
νj(y)(cid:0)gs(xj − yj) − gs(xj)(cid:1)
m
> sPj(νj)(cid:16){z ∈ Zm : 3s 6 z 6 4s}(cid:17).
m, and hence by integrating (65) over
(65)
Note that (64) holds for a constant fraction of x ∈ Zn
m we obtain:
Zn
νj(y) .
Xy∈Zn
m
3s6yj64s
1
sZZn
m k(fs ∗ νj − fs)(x)kℓn
∞dµ(x).
Averaging (66) over j ∈ {1, . . . , n} and using (63) we get,
n
1
n
Xj=1 Xy∈Zn
m
3s6yj64s
(66)∧(63)
νj(y)
.
A
s
.
Therefore
1
n
n
Xj=1 Xy∈Zn
m
yj>3s
νj(y) =
∞
Xℓ=0
1
n
n
Xj=1
3j( 4
3 )ℓ
X
sk6yj64j( 4
3)ℓ
(67)
.
νj(y)
∞
Xℓ=0
A
s · (4/3)ℓ
.
A
s
.
sk
(66)
(67)
(cid:3)
Corollary 4.4. Assume that m > cA for a large enough universal constant c ∈ N. Then:
(68)
n
.
Pj(νj)(z + 1) − Pj(νj)(z − 1) &
1
A
1
n
Xj=1 Xz∈Zm
Proof. We may assume that A is an integer. By Lemma 4.3, for c large enough we have
1
n
n
Xj=1 Xz6cA
Pj(νj)(z) >
3
4
and
Therefore,
n
3cA+2
Xj=1
Xz=cA+2
Pj(νj)(z) 6
1
4
.
1
n
n
Xj=1 Xz−2cA−26cA
Pj(νj)(z)
Pj(νj)(z) −
1
n
1
2
6
=
=
.
1
n
1
n
1
n
A
n
as required.
n
n
n
Xj=1 Xz6cA
Xj=1 Xz6cA
Xj=1 Xz6cA
Xj=1 Xz∈Zm
n
[Pj(νj)(z) − Pj(νj)(z + 2cA + 2)]
cA+1
[Pj(νj)(z + 2(t − 1)) − Pj(νj)(z + 2t)]
Xt=1
Pj(νj)(z + 1) − Pj(νj)(z − 1) ,
22
(cid:3)
Proof of Proposition 4.1. We shall apply the smoothing inequality (8) when X = L1(Zn
and f : Zn
is
m → X is defined as f (x) = mn · δ{x}, i.e., for x ∈ Zn
m the function f (x) : Zn
m, µ)
m → R
f (x)(y) def= (mn
0
x = y,
otherwise.
(69)
m we have:
n
For every ε ∈ {−1, 1}n and x ∈ Zn
Xj=1
εj (f ∗ νj(x + ej) − f ∗ νj(x − ej))
=
n
Xj=1
εj
Xy∈Zn
m(cid:0)νj(y + ej) − νj(y − ej)(cid:1)f (x − y)
(70)
By Kahane's inequality [7, 21] and the fact that L1(Zn
m, µ) has cotype 2 (see [21]),
n
Z{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xj=1
εj
Xy∈Zn
q
dτ (ε)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
m(cid:0)νj(y − ej) − νj(y + ej)(cid:1)f (x − y)
&q
Xj=1
m(cid:0)νj(y − ej) − νj(y + ej)(cid:1)f (x − y)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xy∈Zn
L1(Zn
m,µ)
n
2
(71)
q/2
L1(Zn
m,µ)
Note that by the definition of f , for every x ∈ Zn
m and j ∈ [n] we have,
Hence,
m
m,µ)
m(cid:0)νj(y − ej) − νj(y + ej)(cid:1)f (x − y)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L1(Zn
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xy∈Zn
(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)
= Xw∈Zm
Xz∈Zn
> Xw∈Zm
zj=w(cid:0)νj(z − ej) − νj(z + ej)(cid:1)
m(cid:0)νj(y − ej) − νj(y + ej)(cid:1)f (x − y)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xy∈Zn
Xj=1 Xw∈Zm
> 1
Xj=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
L1(Zn
(72)
n
n
n
2
1
n
23
= Xz∈Zn
m
νj(z − ej) − νj(z + ej)
Pj(νj)(w − 1) − Pj(νj)(w + 1).
(72)
m,µ)
Pj(νj)(w − 1) − Pj(νj)(w + 1)!2
(68)
&
1
A2 .
(73)
m,µ) 6 kf (x)kL1(Zn
Finally, since kf (x) − f (y)kL1(Zn
we can use the smoothing inequality (8) to deduce that
Sq & SqZE∞(Zn
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
> ZZn
εj (f ∗ νj(x + ej) − f ∗ νj(x − ej))(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
m) kf (x) − f (y)kq
m,µ)dβ2(x, y)
Xj=1
L1(Zn
(8)
n
q
where in the last step we used the identity (70), combined with the inequalities (71) and (73).
The proof of Proposition 4.1 is complete.
(cid:3)
m,µ) + kf (y)kL1(Zn
m,µ) 6 2 for all x, y ∈ Zn
m,
dτ (ε)dµ(x)&q
nq/2
Aq ,
L1(Zn
m,µ)
4.2. A sharp lower bound for Ej averages: Proof of Proposition 4.2. Recall that
S(j, k) is defined in (16), and in the setting of Proposition 4.2 we have:
νj(x) =
1S(j,k)(x)
k(k + 1)n−1 .
Let s ∈ {(k + 1)/2, (k + 3)/2} be an odd integer. By the definition of S(j, k) we have
νj(x) =
Xx∈Zn
m
xj>s
(k − s)(k + 1)n−1
k(k + 1)n−1
& 1.
Plugging this estimate into (62) we see that A/k & 1, proving the first assertion in (61).
To prove the second assertion of Proposition 4.2, we shall apply the smoothing inequal-
m, µ) and the function f from (69),
ity (8), as in Section 4.1, to the Banach space X = L1(Zn
i.e., f (x) = mnδ{x} ∈ L1(Zn
In our setting, the value of
m, µ). We shall use here notation from Section 3.2.
n
εj [Ejf (x + ej) − Ejf (x − ej)](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L1(Zn
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xj=1
m and ε ∈ {−1, 1}n. Thus the left hand side of (8) equals (by
does not depend on x ∈ Zn
Lemma 3.5),
(Ejf (ej) − Ejf (−ej))(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
!q
k(k + 1)n−1Xy∈S(cid:12)(cid:12)(cid:12)↑y↓(cid:12)(cid:12)(cid:12)
At the same time, as noted in Section 4.1, the right hand side of (8) is . Sq. It follows that
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xj
=
L1(Zn
m,µ)
m,µ)
1
.
q
S &
1
k(k + 1)n−1Xy∈S(cid:12)(cid:12)(cid:12)↑y↓(cid:12)(cid:12)(cid:12)
= E [Z] ,
(74)
k−1
i=1 ξi, and {ξi}n
where Z = Pn
i=1 are i.i.d. random variables taking the 0 with probability
1
k+1, and each of the values {−1, 1} with probability
k+1. The last equality in (74) is
an immediate consequence of the definitions of S and (cid:12)(cid:12)(cid:12)↑y↓(cid:12)(cid:12)(cid:12)
k+1, we have
E [Z 2] = np and E [Z 4] = np + n(n − 1)p2. By Holder's inequality it then follows that we
4 ≍ min(cid:8)√np, np(cid:9), completing the proof of Proposition 4.2. (cid:3)
have S & E[Z] > kZk3
. Writing p = 2
2/kZk2
24
4.3. Symmetrization. We do not know what is the smallest m for which the metric cotype
inequality (5) can be shown to hold true via a smoothing and approximation scheme: all we
know is that it is between n1+ 1
q . In this short section, we note that the special
symmetric structure of the smoothing and approximation scheme that we used in the proof
of Theorem 1.2 can be always assumed to hold true without loss of generality. This explains
why our choice of convolution kernels is natural. Additionally, this fact might be useful in
improving the lower bound on m of Proposition 4.1, though we do not know how to use it
in our current proof of Proposition 4.1.
q and n
2 + 1
1
For π ∈ Sn, i.e., a permutation of [n], and x ∈ Zn
m, write
For f : Zn
measure on Zn
m → X we define f π : Zn
m then for all x ∈ Zn
xπ def= (cid:0)xπ(1), xπ(2), . . . , xπ(n)(cid:1) .
m → X by f π(x) = f (xπ). Note that if ν is a probability
m we have
f ∗ νπ =(cid:16)f π−1 ∗ ν(cid:17)π
(75)
.
Indeed,
f ∗ νπ(x) =ZZn
m
It follows from (75) that
f (x − y)ν(yπ)dµ(y)=ZZn
m
f(cid:16)x − zπ−1(cid:17) ν(z)dµ(z)
=ZZn
m
f π−1
(xπ − z) ν(z)dµ(z) = f π−1
∗ ν(xπ) =(cid:16)f π−1
(x).
∗ ν(cid:17)π
kf ∗ νπ − fkLq(Zn
m,X) =(cid:13)(cid:13)(cid:13)
f π−1 ∗ ν − f π−1(cid:13)(cid:13)(cid:13)Lq(Zn
m,X)
.
(76)
Lemma 4.5. Assume that the probability measures ν1, . . . , νn, β1, β2 are a (q, A, S)-smoothing
and approximation scheme. Then there exist probability measures ¯ν1, . . . , ¯νn on Zn
m and two
probability measures ¯β1, ¯β2 on E∞(Zn
1. The sequence ¯ν1, . . . , ¯νn, ¯β1, ¯β2 is also a (q, A, S)-smoothing and approximation scheme,
2. For any j, h ∈ [n] we have ¯νj = ¯ν(j,h)
, where (j, h) ∈ Sn is the transposition of j and h.
3. For every j, h ∈ [n] \ {i} we have Pj(¯νi) = Ph(¯νi).
Proof. Define for j ∈ [n],
m), such that
h
νπ−1
π(j).
(77)
1
n! Xπ∈Sn
def=
¯νj
We also define for (x, y) ∈ E∞(Zn
m),
n! Xπ∈Sn
¯β1(x, y) def=
1
β1 (xπ, yπ)
and
¯β2(x, y) def=
1
n! Xπ∈Sn
25
β2 (xπ, yπ) .
(78)
Fix f : Zn
m → X and assume the validity of the approximation and smoothing inequali-
ties (7), (8). Then, by the convexity of k · kq
X ,
n! Xπ∈Sn
m kf ∗ ¯νj − fkq
Xj=1ZZn
Xdµ
(76)∧(77)
1
n
1
n
6
1
n
n
q
Lq(Zn
m,X)
Xj=1(cid:13)(cid:13)f π ∗ νπ(j) − f π(cid:13)(cid:13)
6 AqZE∞(Zn
(7)∧(78)
m) kf (x) − f (y)kq
Xd ¯β1(x, y).
(79)
This is precisely the approximation property for ¯ν1, . . . , ¯νn, ¯β1, ¯β2.
Similarly,
n
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
εj(cid:16)f ∗ ¯νj(x + ej) − f ∗ ¯νj(x − ej)(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ZZn
Xj=1
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n! Xπ∈SnZZn
Xj=1
εj(cid:16)f ∗ νπ−1
(78)
6
1
n
π(j)(x + ej) − f ∗ νπ−1
q
X
Note that
q
X
π(j)(x − ej)(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
dτ (ε)dµ(x)
dτ (ε)dµ(x).
(80)
n
Xj=1
εj(cid:16)f ∗ νπ−1
π(j)(x − ej)(cid:17)
π(j)(x + ej) − f ∗ νπ−1
επ−1(i)(cid:16)f π ∗ νi(cid:16)xπ−1
Xi=1
(75)
=
n
+ ei(cid:17) − f π ∗ νi(cid:16)xπ−1
− ei(cid:17)(cid:17) ,
(81)
q
n
(81)
π(j)(x + ej) − f ∗ νπ−1
where we made the change of variable j = π−1(i) and used the fact that eπ−1
r ∈ [n] and π ∈ Sn. Hence,
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ZZn
εj(cid:16)f ∗ νπ−1
Xj=1
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
= ZZn
Xr=1
εj(cid:16)f ∗ ¯νj(x + ej) − f ∗ ¯νj(x − ej)(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
εr(cid:16)f π ∗ νr(x + er) − f π ∗ νr(x − er)(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
mZ{−1,1}n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ZZn
Xj=1
The smoothing inequality for ¯ν1, . . . , ¯νn, ¯β1, ¯β2 now follows:
π(j)(x − ej)(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
dτ (ε)dµ(x)
dτ (ε)dµ(x)
(80)∧(82)∧(8)
6
X
X
X
n
n
q
q
r = eπ(r) for all
dτ (ε)dµ(x).
(82)
SqZE∞(Zn
m) kf (x) − f (y)kq
Xd ¯β2(x, y).
Assertions 2. and 3. of Lemma 4.5 follow directly from the definition (77).
(cid:3)
26
References
[1] B. Begun. A remark on almost extensions of Lipschitz functions. Israel J. Math., 109:151 -- 155, 1999.
[2] J. Bourgain. The metrical interpretation of superreflexivity in Banach spaces. Israel J. Math., 56(2):222 --
230, 1986.
[3] J. Bourgain. Remarks on the extension of Lipschitz maps defined on discrete sets and uniform homeo-
morphisms. In Geometrical aspects of functional analysis (1985/86), volume 1267 of Lecture Notes in
Math., pages 157 -- 167. Springer, Berlin, 1987.
[4] J. Diestel, H. Jarchow, and A. Tonge. Absolutely summing operators, volume 43 of Cambridge Studies
in Advanced Mathematics. Cambridge University Press, Cambridge, 1995.
[5] O. Giladi and A. Naor. Improved bounds in the scaled Enflo type inequality for Banach spaces. Available
at http://arxiv.org/abs/1004.4221, 2010.
[6] L. Hormander. An introduction to complex analysis in several variables, volume 7 of North-Holland
Mathematical Library. North-Holland Publishing Co., Amsterdam, third edition, 1990.
[7] J.-P. Kahane. Sur les sommes vectorielles P±un. C. R. Acad. Sci. Paris, 259:2577 -- 2580, 1964.
[8] N. J. Kalton. Coarse and uniform embeddings into reflexive spaces. Q. J. Math., 58(3):393 -- 414, 2007.
[9] B. Maurey. Type, cotype and K-convexity. In Handbook of the geometry of Banach spaces, Vol. 2, pages
1299 -- 1332. North-Holland, Amsterdam, 2003.
[10] B. Maurey and G. Pisier. S´eries de variables al´eatoires vectorielles ind´ependantes et propri´et´es
g´eom´etriques des espaces de Banach. Studia Math., 58(1):45 -- 90, 1976.
[11] M. Mendel. Metric dichotomies. In Limits of graphs in group theory and computer science, pages 59 -- 76.
EPFL Press, Lausanne, 2009.
[12] M. Mendel and A. Naor. Scaled Enflo type is equivalent to Rademacher type. Bull. Lond. Math. Soc.,
39(3):493 -- 498, 2007.
[13] M. Mendel and A. Naor. Metric cotype. Ann. of Math. (2), 168(1):247 -- 298, 2008.
[14] A. Naor. An application of metric cotype to quasisymmetric embeddings. Preprint, 2006. Available at
http://arxiv.org/abs/math/0607644.
[15] G. Pisier. Holomorphic semigroups and the geometry of Banach spaces. Ann. of Math. (2), 115(2):375 --
392, 1982.
[16] G. Pisier. Probabilistic methods in the geometry of Banach spaces. In Probability and analysis (Varenna,
1985), volume 1206 of Lecture Notes in Math., pages 167 -- 241. Springer, Berlin, 1986.
[17] Y. Raynaud. On ultrapowers of non commutative Lp spaces. J. Operator Theory, 48(1):41 -- 68, 2002.
[18] N. Tomczak-Jaegermann. The moduli of smoothness and convexity and the Rademacher averages of
[19] A. Veomett and K. Wildrick. Spaces of
trace classes Sp(1 6 p < ∞). Studia Math., 50:163 -- 182, 1974.
http://arxiv.org/abs/1001.3326.
small metric cotype. Preprint, 2010. Available at
[20] H. S. Wilf. generatingfunctionology. A K Peters Ltd., Wellesley, MA, third edition, 2006.
[21] P. Wojtaszczyk. Banach spaces for analysts, volume 25 of Cambridge Studies in Advanced Mathematics.
Cambridge University Press, Cambridge, 1991.
Courant Institute, New York University
E-mail address: [email protected]
Computer Science Division, The Open University of Israel
E-mail address: [email protected]
Courant Institute, New York University
E-mail address: [email protected]
27
|
1906.00394 | 3 | 1906 | 2019-10-09T17:54:24 | On slow decay of Peetre's K-functional | [
"math.FA"
] | We characterize when Peetre's K-functional decays to zero slowly and we use this characterization to demonstrate certain strict inclusions between real interpolation spaces. | math.FA | math |
ON SLOW DECAY OF PEETRE'S K-FUNCTIONAL
J. M. ALMIRA, P. FERN ´ANDEZ-MART´INEZ
Abstract. We characterize when Peetre's K-functional slowly decays to zero and we
use this characterization to demonstrate certain strict inclusions between real interpola-
tion spaces.
1. Introduction
In this paper we consider a couple pX, Y q, where pX, } }Xq is a quasi-Banach space
and Y Ă X is a quasi-semi-normed space, pY, } }Y q which is continuously embedded into
X. Under these conditions, we also consider Peetre's K-functional, which is defined by
Kpx, t, X, Y q " inf
yPY
}x ´ y}X ` t}y}Y
This functional is obviously connected to the approximation properties of the elements of
X by elements of Y . Concretely, if Y is dense in X, then
Kpx, 0`, X, Y q " lim
tÑ0`
Kpx, t, X, Y q " 0
and the claim
is equivalent to the following property:
Kpx, t, X, Y q ě c
If y P Y satisfies }x ´ y}X ă c, then }y}Y ě
In particular, if }x ´ y}X ă c
t
infinity when t tends to 0. Let A be a subset of X, we define
2 with y P Y , then }y}Y ě c´}x´y}X
KpA, t, X, Y q " sup
xPA
Kpx, t, X, Y q.
Thus, if SpXq denotes the unit sphere of X, the condition
(1)
KpSpXq, t, X, Y q ą c for all t ą 0
c ´ }x ´ y}X
t
ě c
2t, which diverges to
implies the existence of elements in X of small norm (in particular, of norm }x}X " 1)
which are badly approximable by elements of Y with small norm in Y , which is a property
that holds true in many cases. For example, if we set X " Cra, bs and Y " C pmqra, bs
Key words and phrases. Lethargy results, Rate of convergence, K-functional.
Corresponding author.
1
2
J. M. Almira, P. Fern´andez-Mart´ınez
with m ě 1, then there are oscillating functions of small uniform norm that cannot be
uniformly approximated by functions of small C pmq-semi-norm.
The connection between the K-functional and approximation theory is, in fact, a strong
one, and there are many results which connect the behavior of this functional to the prop-
erties of best approximation errors with respect to an approximation scheme pX, tAnuq. In
particular, the so called Central Theorems in Approximation Theory lead to a proof that,
when the approximation scheme pX, tAnuq satisfies Jackson's and Bernstein's inequalities
with respect to the space Y ãÑ X, which are given by
Epy, Anq ď Cn´r}y}Y for all y P Y and all n ě 0
and
}a}Y ď Cnr}a}X for all a P An,
respectively, then the approximation spaces Aα
are completely characterized as interpolation spaces by the formula (see [2])
q pX, tAnuq " tx P X : t2αkEpx, A2k qu P ℓqu
Aα
q pX, tAnuq " pX, Y qα{r,q for all 0 ă α ă r and all 0 ă q ď 8.
Here pX, Y qθ,q denotes the real interpolation space
pX, Y qθ,q " tx P X : ρθ,qpxq " }t´pθ` 1
q qKpx, t, X, Y q}Lqp0,8q ă 8u,
which, in case that Y ãÑ X, can be renormalized with the following equivalent quasi-norm:
}x}θ,q " }t2θkKpx,
1
2k
, X, Y qu8
k"0}ℓq.
It follows that, when pX, tAnuq satisfies Jackson's and Bernstein's inequalities with respect
to a subspace Y , the rates of convergence to zero of the sequence of best approximation
errors Epx, Anq and the sequence of evaluations of Peetre's K-functional Kpx, t, X, Y q at
points 1{n seem to have similar roles and, in particular, serve to describe the very same
subspaces of X. Thus, it is an interesting question to know under which conditions on
the couple pX, Y q, with Y ãÑ X, the K-functional Kp, , X, Y q approaches to zero slowly,
not only in presence of an approximation scheme but also for the general case. In this
paper we characterize couples whose associated K-functional decays to zero slowly. As an
application, we use this characterization to demonstrate that some natural embeddings
between real interpolation spaces are strict, a question that was posed by Lions in the
decade of 1960 and was solved by Janson, Nilsson, Peetre and Zafran in 1984 [4] with some
sophisticated tools [6]. A complete solution, for the complex method of interpolation, was
previously shown by Stafney in 1970 [7]. Bergh and Lofstrom presented a partial solution
for the real interpolation method in their famous book [1]. Their approach was inspired by
Stafney's work. In this paper we study the very same cases by means of less sophisticated
tools.
In section 2 of this paper we characterize K-functionals which slowly decay to zero as
those satisfying (1) and we also characterize them in terms of the approximation properties
Slow decay of Peetre's K-functional
3
of Y as a subspace of X. In section 3 we demonstrate that for any couple pX, Y q of quasi-
Banach spaces such that Y ãÑ X, Y ‰ X, the K-functional Kp, , X, Y q decays to
zero slowly and, as a consequence, we prove that, under the very same hypotheses, all
interpolation spaces pX, Y qθ,p are strictly embedded between X and Y and that, if θ1 ‰ θ2,
then pX, Y qθ1,p ‰ pX, Y qθ2,q for all 0 ă p, q ď 8. Moreover, if 0 ă θ ă 1 and 0 ă p, q ď 8
are such that pX, Y qθ,p ‰ pX, Y qθ,q, then we show that the spaces pX, Y qθ,r with r P rp, qs
are pairwise distinct. The proofs we introduce for these cases are new, elementary, and
very natural ones. Moreover, the results we get about K-functionals with slow decay are
also new and, in our opinion, interesting by themselves. In the last section of the paper,
we show that, if Y is densely embedded into X and Z ãÑ X is a subspace compactly
embedded into X, then there is a function ϕptq satisfying limtÑ0` ϕptq " 0 such that
Kpz, t, X, Y q ď ϕptq}z}Z for all z P Z. Thus, all elements of Z share a common decay to
zero in their K-functionals.
2. Characterizations of slowly decaying K-functionals
The importance of condition (1) comes from the following result:
Theorem 1. The following are equivalent claims:
paq KpSpXq, t, X, Y q ą c for all t ą 0 and a certain constant c ą 0.
pbq For every non-increasing sequences tεnu, ttnu P c0, there are elements x P X such
that
Kpx, tn, X, Y q ‰ Opεnq
Proof. paq ñ pbq. Let us assume, on the contrary, that Kpx, tn, X, Y q " Opεnq for all
x P X and certain sequences tεnu, ttnu P c0. This can be reformulated as
where
X "
Γm,
8
ďm"1
Γm " tx P X : Kpx, tn, X, Y q ď mεn for all n P Nu.
Now, Γm is a closed subset of X for all m and Baire category theorem implies that
Γm0 has nonempty interior for some m0 P N. On the other hand, Γm " ´Γm since
Kpx, t, X, Y q " Kp´x, t, X, Y q for all x P X and t ě 0. Furthermore, if C ą 1 is the
quasi-norm constant of X and Y , then
since, if x, y P Γm and λ P r0, 1s, then
convpΓmq Ď ΓCm,
Kpλx ` p1 ´ λqy, tn, X, Y q ď CpKpλx, tn, X, Y q ` Kpp1 ´ λqy, tn, X, Y qq
" CpλKpx, tn, X, Y q ` p1 ´ λqKpy, tn, X, Y qq
ď Cpλmεn ` p1 ´ λqmεnq
" Cmεn
4
J. M. Almira, P. Fern´andez-Mart´ınez
Thus, if BXpx0, rq " tx P X : }x0´x}X ă ru Ď Γm0, then convpBXpx0, rqYBXp´x0, rqq Ď
ΓCm0. In particular,
Now, it is clear that
1
2
pBXpx0, rq ` BXp´x0, rqq Ď ΓCm0.
BXp0, rq Ď
1
2
pBXpx0, rq ` BXp´x0, rqq,
since, if x P BXp0, rq, then } ´ x0 ´ px ` x0q} " }x0 ´ px ´ x0q} " }x}X "ď r, so that
x ´ x0 P BXpx0, rq, x ` x0 P BXpx0, rq, and x " 1
2ppx ´ x0q ` px ` x0qq. Thus,
This means that, if x P Xzt0u, then
BXp0, rq Ď ΓCm0.
Hence
Kp
rx
}x}
, tn, X, Y q ď Cm0εn for all n " 1, 2,
Kpx, tn, X, Y q ď
}x}
r
Cm0εn for all n " 1, 2,
On the other hand, paq implies that, for each n, there is xn P SpXq such that Kpxn, tn, X, Y q ą
c, so that
c ă Kpxn, tn, X, Y q ď
1
r
Cm0εn for all n " 1, 2, ,
which is impossible, since εn converges to 0 and c ą 0. This proves paq ñ pbq.
Let us demonstrate the other implication. Assume that paq does not hold. Then there
are non-increasing sequences ttnu, tcnu P c0 such that
KpSpXq, tn, X, Y q ď cn for all n P N
In particular, if x P X is not the null vector, then
Kp
x
}x}X
, tn, X, Y q ď KpSpXq, tn, X, Y q ď cn for all n P N,
so that
Kpx, tn, X, Y q " }x}XKp
x
}x}X
, tn, X, Y q ď }x}Xcn for all n P N,
and Kpx, tn, X, Y q " Opcnq for all x P X. This proves pbq ñ paq.
(cid:3)
Definition 2. We say that the K-functional Kp, , X, Y q slowly decays to zero if either
condition paq or pbq of Theorem 1 hold true.
Remark 3. Note that, if Y Ď X and the K-functional Kp, , X, Y q decays to zero slowly,
then pX, Y qθ,q is properly contained into X for all 0 ă θ ă 1 and all 0 ă q ď 8.
Next proposition, whose proof is omitted, states that if the elements of SpXq are not
easily approximated by elements of Y , they cannot be easily approximated by elements
of any other subspace Y1 ãÑ Y .
Slow decay of Peetre's K-functional
5
Proposition 4. Assume that Y is densely embedded into X and KpSpXq, t, X, Y q ą c
for all t ą 0 and a certain constant c ą 0. Then:
(a) If Y1 ãÑ Y is another quasi-semi-normed space continuously embedded into Y ,
then KpSpXq, t, X, Y1q ą c for all t ą 0.
(b) If Y1 ãÑ Y is another quasi-semi-normed space continuously embedded into Y ,
then for every every non-increasing sequences tεnu, ttnu P c0, there are elements
x P X such that
Kpx, tn, X1, Y1q ‰ Opεnq.
Remark 5. Assume that }},X is an equivalent quasi-norm of X and }},Y is an equivalent
quasi-semi-norm of Y , and let us denote by
K px, t, X, Y q " inf
yPY
}x ´ y},X ` t}y},Y
the K-functional associated to the couple pX, Y q when doted with the norms } },X and
} },Y , respectively. Then there are two constants M, N ą 0 such that
M K px, t, X, Y q ď Kpx, t, X, Y q ď N K px, t, X, Y q
In particular, if SpXq " tx P X : }x},X " 1u, the following are equivalent claims:
paq K pSpXq, t, X, Y q ą c for all t ą 0 and a certain constant c ą 0.
pbq KpSpXq, t, X, Y q ą c for all t ą 0 and a certain constant c ą 0.
Let us now characterize, in terms of the approximation properties of Y as a subspace
of X, the couples pX, Y q whose associated K-functional slowly decays to zero:
Theorem 6. The following are equivalent claims:
paq KpSpXq, t, X, Y q ą c for all t ą 0 and a certain constant c ą 0.
pbq For a certain constant δ ą 0 there exist sequences txnu Ă SpXq and tbnu Ăs0, 8r,
lim bn " 8, such that, for every n P N,
}xn ´ y}X ă δ and y P Y imply that }y}Y ě bn
Proof. Note that we can assume 0 ă c, δ ă 1 without loss of generality. Assume paq and
take, for each t ą 0, an element xt P SpXq such that Kpxt, t, X, Y q ą c then we have
2 with yt P Y , then }yt}Y ě c´}xt´yt}X
already shown that if }xt ´ yt}X ă c
2t. Thus,
pbq holds for δ " c{2 and bn " c
for every decreasing sequence ttnu P c0. This proves
2tn
paq ñ pbq.
ě c
t
Let us now assume that Kpxn, 1
bn
there is yn P Y such that }xn ´ yn}X ` 1
bn
}xn ´ yn}X ă δ and yn P Y imply }yn}Y ě bn, so that
, X, Y q ă δ with txnu, δ and tbnu satisfying pbq. Then
}yn}Y ă δ, which leads to a contradiction since
1 ą δ ą }xn ´ yn}X `
1
bn
}yn}Y ą 1.
It follows that KpSpXq, 1
bn
monotonicity of K and the fact that limnÑ8
, X, Y q ě δ for all n and paq follows with c " δ from the
(cid:3)
" 0.
1
bn
6
J. M. Almira, P. Fern´andez-Mart´ınez
Theorem 6 reveals an easy way to confirm that Peetre's K-functional decays to zero
slowly, since condition pbq of this theorem is, in many cases, easy to verify. We include
here just a few examples to show the way this theorem can be used.
Example 7. Set X " Cra, bs with the uniform norm }f }Cra,bs " suptPra,bs f ptq and Y "
C p1qra, bs with the seminorm }g}Cp1qra,bs " }g1}Cra,bs. Take αn " a`b
2 ` 1
and set
2n , βn " a`b
2 ´ 1
2n
´1
2
βn´αn
1
x ´ αn`βn
βn´αn
x P ra, αns
x P rαn, βns
x P rβn, bs
fnpxq "$&
%
Then }fn}Cra,bs " 1 for all n and, if }fn ´ g}Cra,bs ă 1
and gpβnq ą 1{2 so that, there is a point ρn P rαn, βns Ď ra, bs such that
2 with g P C p1qra, bs then gpαnq ă ´1{2
g1pρnq "
gpβnq ´ gpαnq
βn ´ αn
ě
1
1{n
" n
and condition pbq of Theorem 6 holds true with δ " 1
also works when we consider on Y " C p1qra, bs the norm }g}1 " }g}Cra,bs ` }g1}Cra,bs.
2 and bn " n. Note that this example
1
n"0 anq
}tanu}q " `ř8
Example 8. Given 0 ă p ă q ď 8 we set X " ℓq with the standard norm (or quasi-norm)
q , for q ă 8, and X " c0 with the supremum norm }tanu}8 "
supn an, if q " 8. We also set Y " ℓp with the standard norm (or quasi-norm). Take
k " 0 for k ě 2n. If }an´b}8 ă 1
an " tan
(analogously if }an ´b}q ă 1
2 for all k P t0, 1, , 2n´1u,
which implies that }b}p ě 2´1` 1
p . Thus, condition pbq of Theorem 6 holds true in all
these cases with δ " 1
2), with b " tbku P ℓp, then bk ą 1
k " 1 for k " 0, 1, , 2n´1 and an
k"0, where an
k u8
2 and bn " 2´1` 1
p n
1
p
1
p n
2
Example 9. If we do not impose Y ãÑ X but we maintain Y Ď X (i.e., we do not require
continuity of the inclusion) then it is easy to find a couple pX, Y q whose K-functional
does not decay to zero slowly. Concretely, we can take X " ℓ1 with its natural norm,
} }1, and Y " c00 " t finite sequences u with the norm of supremum , } }8. Obviously
Y Ă X. Take panq P Spℓ1q and take δ ą 0 arbitrary. Then there is a number N0 P N such
n"N0`1 an ă δ. Take pbnq P c00 given by bk " ak for k ď N0 and bk " 0 otherwise.
that ř8
Then
Kppanq, t, ℓ1, c00q ď }panq ´ pbnq}1 ` t}pbnq}8 ď δ ` t,
and this holds for every δ ą 0, so that
and this K-functional does not decay to zero slowly.
Kppanq, t, ℓ1, c00q ď t
Slow decay of Peetre's K-functional
7
3. The Quasi-Banach setting
The following result shows that, under very mild conditions on the couple pX, Y q, the
K-functional Kp, , X, Y q slowly decays to zero.
Theorem 10. Assume that pX, } }Xq and pY, } }Y q are quasi-Banach spaces. Then
paq If Y ãÑ X, Y ‰ X (i.e., Y is properly embedded into X). Then KpSpXq, t, X, Y q ą
c for all t ą 0 and a certain constant c ą 0.
pbq If X and Y are both p-normed spaces, Y ãÑ X, then either KpSpXq, t, X, Y q " 1
for all t ą 0, or X " Y .
Proof. It follows from Aoki-Rolewicz Theorem (see [2, Theorem 2.1.1] or [5, pages 7-
8]) that X and Y can be renormed with equivalent quasi-norms in such a way that they
become p-normed spaces for a certain p Ps0, 1s (indeed, Aoki-Rolewicz gives two equivalent
quasi-norms } },X and } },Y , respectively and two numbers pX, pY Ps0, 1s such that
}x1 ` x2}pX
,Y for all x1, x2 P X
and all y1, y2 P Y . Thus, if we set p " mintpX, pY u then }x1 ` x2}p
,X
and }y1 ` y2}p
,Y for all x1, x2 P X and all y1, y2 P Y and both spaces
are p-normed with the very same p. It follows from this and from Proposition 5 that we
only need to demonstrate part pbq of the Theorem.
,X and }y1 ` y2}pY
,Y ď }y1}pY
,X ď }x1}pX
,X ` }x2}pX
,Y ď }y1}p
,Y ` }y2}p
,Y ` }y2}pY
,X ď }x1}p
,X ` }x2}p
Thus, we assume that X, Y are quasi-Banach p-normed spaces and Y ãÑ X and we
want to demonstrate that, if KpSpXq, t, X, Y q ă 1 for a certain t ą 0, then X " Y .
Assume, on the contrary, that KpSpXq, t0, X, Y q " c ă 1 and X ‰ Y . Take ρ Ps0, 1r such
that c ă ρ1{p ă 1. Then Kp x
, t0, X, Y q ă ρ1{p for every x P X, x ‰ 0. Hence every
}x}X
element x of X which is different from zero satisfies
Kpx, t0, X, Y q ă ρ1{p}x}X,
which implies that }x ´ y0}X ` t0}y0}Y ă ρ1{p}x}X for certain y0 P Y . Thus, if we set
x0 " x ´ y0, we have that
(2)
x " x0 ` y0 with x0 P X, and y0 P Y
}x0}X ă ρ1{p}x}X
}y0}Y ă t´1
0 ρ1{p}x}X
Let us take x P XzY and apply (2) to this concrete element. We can repeat the argument
just applying it to x0 (if x0 " 0 then x " y0 P Y , which contradicts our assumption).
Hence, there are elements x1 P X and y1 P Y such that
$&
%
$&
%
x0 " x1 ` y1 with x1 P X, and y1 P Y
}x1}X ă ρ1{p}x0}X ă pρ1{pq2}x}X
}y1}Y ă t´1
0 pρ1{pq2}x}X
Moreover, x " x0 ` y0 " x1 ` y1 ` y0. Again x1 ‰ 0 since x R Y . We can repeat the
argument m times to get a decomposition x " xm ` ym ` ` y0 with xm P X, xm ‰ 0,
8
J. M. Almira, P. Fern´andez-Mart´ınez
yk P Y for all 0 ď k ď m and
" }xm}X ă pρ1{pqm`1}x}X
}yk}Y ă t´1
0 pρ1{pqk`1}x}X for all 0 ď k ď m.
Let us set zm " x ´ xm " y0 ` ` ym. Then
}x ´ zm}X " }xm}X ă pρ1{pqm`1}x}X Ñ 0 for m Ñ 8.
and x is the limit of zm in the norm of X. On the other hand, if n ą m, then
n
n
}zn ´ zm}p
Y " }ym`1 ` ` yn}p
Y ď
}yk}p
Y ď t´p
0 }x}p
X
ÿk"m`1
ρk`1,
ÿk"m`1
which converges to 0 for n, m Ñ 8. Hence tzmu is a Cauchy sequence in Y and its limit
belongs to Y since Y is topologically complete. This implies x P Y , which contradicts our
assumptions. Thus, we have demonstrated, for p-normed quasi-Banach spaces X and Y
satisfying Y ãÑ X, that if KpSpXq, t0, X, Y q ă 1 for a certain t0 ą 0, then X " Y . In
particular, if X ‰ Y then KpSpXq, t, X, Y q " 1 for all t ą 0.
(cid:3)
Corollary 11. Let pX, Y q be an ordered couple of quasi-Banach spaces, Y ãÑ X, Y ‰ X.
Then there exist a constant δ ą 0; a sequence tbnu Ă r0, 8q with pbnq Ñ 8 as n Ñ 8 and
a sequence txnu Ď SpXq such that for all n P N and y P Y
}xn ´ y}X ă δ ñ }y}Y ą bn.
Proof. The result is a direct application of Theorems 10 and 6.
(cid:3)
Theorem 10 has some interesting consequences on the theory of interpolation spaces.
Concretely, if X, Y are quasi-Banach spaces and Y ãÑ X, Y ‰ X, then all interpolation
spaces pX, Y qθ,q are strictly embedded into X and, if Y is not closed in X, they strictly
contain Y . Moreover, the inclusions we get by applying the reiteration theorem are also
strict. This is stated in the following theorems:
Theorem 12. Let pX, Y q be a couple of quasi-Banach spaces, Y ãÑ X, Y ‰ X and Y
not closed in X. Then
Y ãÑ pX, Y qθ,q ãÑ X, with strict inclusions, for 0 ă θ ă 1 and 0 ă q ď 8.
Proof. The fact that the inclusion pX, Y qθ,q ãÑ X is strict follows directly from Theo-
rems 1 and 10, since they allow us to claim that there are elements x P X such that
Kpx, 1
2k , X, Y q goes to zero as slow as we want, which implies that we can find x P X
such that }x}pX,Y qθ,q " 8. The inclusion Y ãÑ pX, Y qθ,q is also strict. This should be well
known, but we include here a proof for the sake of completeness. We assume q ă 8, since
the case q " 8 admits a similar proof. The fact that Y ãÑ pX, Y qθ,q as soon as θ ă 1 is a
direct computation, based on the fact that, for y P Y , Kpy, t, X, Y q ď t}y}Y for all t ą 0.
As Y ãÑ X with strict inclusion, we know that }y}X ď M}y}Y for a certain M ą 0 and all
y P Y and, moreover, there exists a sequence tynu8
n"0 Ď SpY q such that }yn}X converges
Slow decay of Peetre's K-functional
9
to 0 when n goes to infinity. In particular, for every N0 P N there exists yN0 P SpY q such
that }yN0}X ď 2´N0. This obviously implies that
KpyN0, t, X, Y q ď" 1
2N0
t
if t ě 1
2N0
if t ă 1
2N0
Hence
2kθq2´N0q `
2kθqKpyN0,
}yN0}θ,q " « 8
ÿk"0
ď « N0
ÿk"0
" „2θqpN0`1q ´ 1
ď „2pθ´1qqN0
2θq ´ 1
2θq
2θq ´ 1
`
1
q
1
2k
, X, Y qqff
ÿk"N0`1
8
1
q
2pθ´1qqpN0`1q
2pθ´1qkqff
1 ´ 2pθ´1qq
1 ´ 2pθ´1qq
2pθ´1qqpN0`1q
1
q
,
1
q
2´N0q `
which converges to 0 when N0 goes to infinity. This demonstrates that the norms } }Y
and } }θ,q are not equivalent on Y , which implies that Y ‰ pX, Y qθ,q.
(cid:3)
Theorem 13. Let pX, Y q be a couple of quasi-Banach spaces, Y ãÑ X, Y ‰ X. Then
paq Assume that 0 ă θ0, θ1 ă 1, θ0 ‰ θ1, and 0 ă p, q ď 8. Then
pX, Y qθ0,p ‰ pX, Y qθ1,q
pbq Let θ Ps0, 1r and assume that 0 ă p, q ď 8 are such that pX, Y qθ,p ‰ pX, Y qθ,q and
that r1, r2 P rp, qs, r1 ‰ r2. Then
pX, Y qθ,r1 ‰ pX, Y qθ,r2
pcq If 0 ă θ ă 1 and 0 ă p ď 8, then pX, Y qθ,p is an infinite-codimensional subspace
of X and Y is an infinite-codimensional subspace of pX, Y qθ,p.
Proof. Part paq follows from Theorem 10 and the fact that X is a space of class Cp0, pX, Y qq,
which implies that we can use the following reiteration formula:
(3)
pX, pX, Y qθ,pqα,q " pX, Y qαθ,q
Indeed, if 0 ă θ1 ă θ0 ă 1, then θ1 " αθ0 for a certain α Ps0, 1r, which implies that
(4)
pX, Y qθ1,q " pX, Y qαθ0,q " pX, pX, Y qθ0,pqα,q
Thus, Theorem 10 guarantees that pX, Y qθ0,p is (a quasi-Banach space) strictly embedded
into X, so that applying the very same result to the couple pX, pX, Y qθ0,pq, and formula
(4), we get that pX, Y qθ1,q is strictly embedded between pX, Y qθ0,p and X. In particular,
pX, Y qθ1,q ‰ pX, Y qθ0,p.
10
J. M. Almira, P. Fern´andez-Mart´ınez
Let us assume p ą q ą 0. To demonstrate part pbq we use the reiteration formula
(5)
ppX, Y qθ,p, pX, Y qθ,qqη,r " pX, Y qθ,r, where
1
r
"
1 ´ η
p
`
η
q
,
which, in conjunction with Theorem 10, implies that, if pX, Y qθ,p ‰ pX, Y qθ,q then all
interpolation spaces pX, Y qθ,r with r Psp, qr are strictly embedded between the spaces
pX, Y qθ,p and pX, Y qθ,q (note that, if p ą q, then pX, Y qθ,q ãÑ pX, Y qθ,p since ℓq ãÑ ℓp).
This is so because 1
p , 1
q s
which is equivalent to r P rq, ps. Thus, if r1, r2 Psq, pr, r1 ă r2, then pX, Y qθ,r1 is strictly
embedded in pX, Y qθ,p and henceforth pX, Y qθ,r2 is strictly embedded between the spaces
pX, Y qθ,p and pX, Y qθ,r1, since r1 ă r2 ă p. In particular, pX, Y qθ,r1 ‰ pX, Y qθ,r2.
q with 0 ď η ď 1 is another way to claim that 1
r P r 1
r " 1´η
p ` η
Part pcq follows directly from paq and Theorem 12.
(cid:3)
Remark 14. Observe that statement pbq of Theorem 13 also follows from part paq and
formula (5), since choosing two distinct r1, r2 in sp, qr is the same as choosing two distinct
η1, η2 in the lefthand side of formula (5).
In the case we consider a couple pX, Y q where pX, } }Xq is a quasi-Banach space and
Y Ă X is a quasi-semi-normed space pY, } }Y q which is continuously included into X,
we can no longer guarantee the slow decay of the associated K-functional. For example,
in [3] it was proved that, if 0 ă p ă 1 and r P N is a positive natural number, then
Kpf, t, Lp, W r
p with the
quasi-seminorm }g}p,r " }gprq}Lpr0,1s, so that
p q " 0 for all f P Lpr0, 1s, when we deal the Sobolev space W r
Kpf, t, Lp, W r
p q "
inf
gpr´1q is absolutely continuous
}f ´ g}Lpr0,1s ` t}gprq}Lpr0,1s.
In spite of this, we can use Theorem 10 to prove the following:
Theorem 15. Assume that pX, Y q is a couple, where pX, } }Xq is a quasi-Banach space
and Y Ă X is a quasi-semi-normed space pY, } }Y q which is continuously embedded into
X. If the interpolation space pX, Y qθ0,p0 is strictly embedded into X for some choice of
0 ă θ0 ă 1 and 0 ă p0 ď 8, then
KpSpXq, t, X, Y q ą c for all t ą 0 and a certain c ą 0.
In particular, the strict inclusion of one interpolation space pX, Y qθ0,p0 into X implies
that all interpolation spaces pX, Y qθ,p are strictly embedded into X, where 0 ă θ ă 1 and
0 ă p ď 8.
Proof. A direct application of Theorem 10 to the couple of quasi-Banach spaces pX, pX, Y qθ0,p0q
leads to
KpSpXq, t, X, pX, Y qθ0,p0q ą c for all t ą 0 and a certain c ą 0.
Thus, for each t, ε ą 0 there is xt P SpXq such that Kpxt, t, X, pX, Y qθ0,p0q ą p1 ´ εqc. On
the other hand, we have that Y ãÑ pX, Y qθ0,p0, which means that }y}θ0,p0 ď M}y}Y for all
Slow decay of Peetre's K-functional
11
y P Y and a certain constant M " Mpθ0, p0q. Hence
Kpxt, tM, X, Y q " inf
yPY
ě inf
yPY
}xt ´ y}X ` tM}y}Y
}xt ´ y}X ` t}y}θ0,p0
ě
inf
zPpX,Y qθ0 ,p0
}xt ´ z}X ` t}z}θ0,p0 (since Y Ď pX, Y qθ0,p0 )
" Kpxt, t, X, pX, Y qθ0,p0q ą p1 ´ εqc
Hence
Kpx t
M
, t, X, Y q " Kpx t
M
,
t
M
M, X, Y q ą p1 ´ εqc,
which implies that KpSpXq, t, X, Y q ě c ą 0 for all t ą 0, since ε, t ą 0 are arbitrary.
This ends the proof.
(cid:3)
4. Spaces whose elements share a common decay in their K-functionals
respect to a given couple
Another interesting question is to know under which conditions on a subspace Z of X
we have that, for each decreasing sequence ttnu all elements z P Z share a common rate
of convergence to zero for the sequence tKpz, tn, X, Y qu. We give a sufficient condition in
the following result:
Theorem 16. Assume that Y is a continuously embedded dense subspace of X. If Z is
a compactly embedded subspace of X, and ttnu P c0 is non-increasing, then there exists a
non increasing sequence tεnu P c0 such that
Kpz, tn, X, Y q ď }z}Zεn for all n P N and all z P Z
In fact, if we set ϕptq " KpSpZq, t, X, Y q, then limtÑ0` ϕptq " 0 and Kpz, t, X, Y q ď
}z}Zϕptq for all z P Z.
Proof. Take εn " KpSpZq, tn, X, Y q, where SpZq denotes the unit sphere in Z. Then
Kpz, tn, X, Y q " }z}ZKp
z
}z}Z
, tn, X, Y q ď }z}Zεn for all n P N,
so that the result is proved as soon as we demonstrate that tεnu P c0. Now, the sequence
tεnu is non-increasing because of the monotony of K. Thus, if this sequence does not
converge to 0, is uniformly bounded by a positive constant c ą 0. In particular, there is
a sequence tznu Ď SpZq such that Kpzn, tnX, Y q ě c for all n. The compactness of the
imbedding Z Ñ X implies that there is a subsequence tznk u8
k"1 and a point z P X such
that limkÑ8 }znk ´ z}X " 0. Hence
0 ă c ď Kpznk , tnk, X, Y q ď CrKpznk ´ z, tnk , X, Y q ` Kpz, tnk , X, Y qs Ñ 0,
which is impossible. Thus, we have demonstrated that, if ϕptq " KpSpZq, t, X, Y q, then
limtÑ0` ϕptq " 0 and Kpz, t, X, Y q ď }z}Zϕptq for all z P Z.
(cid:3)
12
J. M. Almira, P. Fern´andez-Mart´ınez
Example 17. If pX, Y q is a couple and Y ãÑ Z ãÑ X with Z compactly embedded into X,
then we can define the space pZ , } }q " pZ, } }Xq which results from considering on Z
the norm of X and, thanks to Theorem 16, the associated K-functional will satisfy:
Kpz, t, Z , Y q " inf
yPY
}z ´ y} ` t}y}Y " inf
yPY
}z ´ y}X ` t}y}Y ď }z}Zϕptq,
where ϕptq " KpSpZq, t, X, Y q. Thus, Kpz, t, Z , Y q does not decay to zero slowly.
References
[1] J. Bergh and J. Lofstrom, Interpolation spaces. An introduction, Springer, Berlin, 1976.
[2] R.A. DeVore, G.G. Lorentz, Constructive Approximation, Springer, 1993.
[3] Z. Ditzian, V.H. Hristov, K.G. Ivanov, Moduli of smoothness and K-functionals in Lp, 0 ă
p ă 1, Constructive Approx. 11 (1995) 67 -- 83.
[4] S. Janson, P. Nilsson, J. Peetre, M. Zafran, Notes on Wolff's note on interpolation spaces,
Proc. London Math. Soc. (3), 48 (1984) 283-299.
[5] N.J. Kalton, N.T. Peck, J.W. Roberts, An F-Space Sampler, in: London Math. Soc. Lecture
Note Series, vol. 89, Cambridge University Press, 1984.
[6] M. Levy, L'espace d'interpolation reel pA0, A1qθ,p contient ℓp, C. R. Acad. Sci. Paris, 289 (1979)
675 -- 677.
[7] J. Stafney, Analytic interpolation of certain multiplier spaces, Pacific J. Math., 32 (1970) 241-248.
J. M. ALMIRA
Departamento de Ingenier´ıa y Tecnolog´ıa de Computadores, Universidad de Murcia.
30100 Murcia, SPAIN
e-mail: [email protected]
P. FERN ´ANDEZ-MART´INEZ
Departamento de Matem´aticas, Universidad de Murcia.
30100 Murcia, SPAIN
e-mail: [email protected]
|
1708.00091 | 2 | 1708 | 2017-10-04T23:42:04 | Discrete probabilistic and algebraic dynamics: a stochastic commutative Gelfand-Naimark Theorem | [
"math.FA",
"math-ph",
"math.CT",
"math-ph",
"math.PR"
] | We introduce a category of stochastic maps (certain Markov kernels) on compact Hausdorff spaces, construct a stochastic analogue of the Gelfand spectrum functor, and prove a stochastic version of the commutative Gelfand-Naimark Theorem. This relates concepts from algebra and operator theory to concepts from topology and probability theory. For completeness, we review stochastic matrices, their relationship to positive maps on commutative $C^*$-algebras, and the Gelfand-Naimark Theorem. No knowledge of probability theory nor $C^*$-algebras is assumed and several examples are drawn from physics. | math.FA | math |
Discrete probabilistic and algebraic dynamics:
a stochastic commutative Gelfand-Naimark Theorem
Arthur J. Parzygnat
Abstract
We introduce a category of stochastic maps (certain Markov kernels) on compact Haus-
dorff spaces, construct a stochastic analogue of the Gelfand spectrum functor, and prove
a stochastic version of the commutative Gelfand-Naimark Theorem. This relates concepts
from algebra and operator theory to concepts from topology and probability theory. For
completeness, we review stochastic matrices, their relationship to positive maps on commu-
tative C ∗-algebras, and the Gelfand-Naimark Theorem. No knowledge of probability theory
nor C ∗-algebras is assumed and several examples are drawn from physics.
Contents
1 An algebraic perspective on probability theory
1.1 Brief background and motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Overview of results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
2
3
2 Positive maps and stochastic matrices
5
5
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
2.2 Some categories for finite probability theory . . . . . . . . . . . . . . . . . . . . . .
6
2.3 C ∗-algebras and states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Two types of morphisms of C ∗-algebras . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 From probability theory to algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6 From algebra to probability theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.7 Some quantum mechanics
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 An equivalence between spaces and algebraic structures
23
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1 From spaces to algebras
3.2 Some topological preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 The spectrum of a commutative C ∗-algebra . . . . . . . . . . . . . . . . . . . . . . . 28
3.4 The Gelfand transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5 The commutative Gelfand-Naimark equivalence . . . . . . . . . . . . . . . . . . . . 33
2010 Mathematics Subject Classification. 60B05 (Primary) 47B65, 18A40 (Secondary).
Key words and phrases. C ∗-algebra, positive map, regular measure, Markov kernel, categorical probability,
Radon monad, compact Hausdorff, Dynkin π-λ Theorem, states on C ∗-algebras, spectrum, Choquet theory.
1
4 Abstract probability theory
35
4.1 Compact Hausdorff spaces and complex measures . . . . . . . . . . . . . . . . . . . 36
4.2 Probability measures on compact Hausdorff spaces . . . . . . . . . . . . . . . . . . . 42
4.3 Continuous stochastic maps
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 From continuous probability theory to algebra . . . . . . . . . . . . . . . . . . . . . 51
4.5 The stochastic Gelfand spectrum functor . . . . . . . . . . . . . . . . . . . . . . . . 53
4.6 The stochastic commutative Gelfand-Naimark Theorem . . . . . . . . . . . . . . . . 62
5 Closing discussion
64
5.1 Categories of classical and quantum dynamics . . . . . . . . . . . . . . . . . . . . . 65
5.2 Categorical probability theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 The Baire approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Index of notation
70
1 An algebraic perspective on probability theory
That's another thing about categories, not that they give you the proofs, but they tell
you what to prove! -Mikhail Gromov1
1.1 Brief background and motivation
The fact that every compact Hausdorff space can be recovered from its C ∗-algebra of complex-
valued continuous functions was discovered by Gelfand and Naimark in 1943 (see Lemma 1 in
[18]). This involved constructing a space out of a commutative C ∗-algebra and a continuous
map out of a ∗-homomorphism of algebras. Extending earlier work of Riesz and Markov (and
many others), Kakutani proved that finite regular measures on such spaces correspond to positive
linear functionals on the algebras (see Theorem 9 in [23]). One can combine these ideas in the
following way. Continuous maps between spaces can be viewed as deterministic assignments in
the sense that their values on inputs are specified precisely.
Instead, we can imagine a non-
deterministic analogue as a "smeared out" function, or more precisely, an assignment sending
inputs to probability measures. Such "stochastic maps" correspond to (completely) positive linear
maps between the algebras. The latter are well-defined for non-commutative C ∗-algebras and
describe non-deterministic dynamics in quantum systems [27]. Although this correspondence is
well-understood to a large extent [9], [12] some important features are missed without considering
the category of such spaces and stochastic maps. For example, how are the compositions of
stochastic maps defined, are compositions respected under the correspondence, what is the explicit
procedure to go back from a positive map to a stochastic map, and is there a way to formulate all
of these dualities in the same context?
Our aim is to introduce these ideas in a suitable language and answer these questions.
In
the spirit of the Rosetta Stone [4], we will use category theory to do this, though we will only
1This quote is taken from Mikhail Gromov's Lecture 1 on "Probability, symmetry, linearity" given at the Institut
des Hautes ´Etudes Scientifiques on October 10, 2014.
2
use the bare minimum necessary.
In particular, we only assume the reader knows what cate-
gories, functors, and natural transformations are and how to compose them though the reader
may also understand the results without even knowing these definitions verbatim from memory.
Using category theory has the immediate benefit of formulating our questions precisely thereby
making the goals clear: we must identify the appropriate topological, algebraic, deterministic, and
non-deterministic categories, construct mappings between them, and prove the above mentioned
equivalences.
This work fits into the general context of categorical probability theory, which was pioneered by
Lawvere who first defined a category of stochastic maps in 1962 [26]. Further work in the area of
category-theoretic aspects of measure theory was developed by Linton in 1963 [28] and the Polish
school in the early 1970's, notably Semadeni [42] and ´Swirszcz [46]. In the early 1980's, Giry further
developed these ideas by describing the monadic structures available in probability theory on a
larger class of spaces [19]. More recently, this area was revisited in the area of computer science
where functional programming makes use of category-theoretic ideas, particularly monads. The
relationship to the category of ordinary relations, used in the theory of computing, was generalized
to the stochastic setting by Panangaden in 1998 [31] (an earlier version of this article was titled
"Probabilistic Relations") based on earlier work of Kozen in 1981 [24]. A monadic and algebraic
viewpoint on the questions raised above were worked out by Furber and Jacobs recently [15]. Our
goals are similar, but we approach the problem from an analytic perspective.
1.2 Overview of results
We prove a generalization of the commutative Gelfand-Naimark Theorem (henceforth referred to
as just the Gelfand-Naimark Theorem) that is valid for continuous stochastic maps on compact
Hausdorff spaces. This is done at three different levels of increasing complication, the first two
of which are well known. In Section 2, we review the equivalence between stochastic matrices on
finite sets and positive maps on finite-dimensional commutative (unital) C ∗-algebras, which are
all basically of the form Cn for some n ∈ N with pointwise algebraic structure. Specializing to the
case of functions instead of stochastic matrices corresponds to ∗-homomorphisms on the C ∗-algebra
side. In Section 3, we review the Gelfand-Naimark Theorem phrased in an appropriate categorical
context as an equivalence between the category of compact Hausdorff spaces with continuous maps
and the category of commutative unital C ∗-algebras and unit-preserving ∗-homomorphisms. We
do not prove the fundamental theorems here since they are well documented in standard references,
but we do introduce the necessary concepts needed to understand the statements in what follows.
Finally, in Section 4, we extend the Gelfand-Naimark Theorem to an equivalence between
compact Hausdorff spaces with continuous stochastic maps and commutative C ∗-algebras with
positive unital maps. To even state this theorem, an appropriate category of such spaces and
stochastic maps must be defined. Although versions of this category have been defined in the
literature (see [19] and [31] for example), the spaces are typically either measurable spaces or Polish
spaces. There is an important distinction between these two cases. Briefly, a Markov kernel from
a measurable space (X,MX) to another one (Y,MY ) is a function X × MY → [0, 1] such that
fixing the left variable gives a probability measure and fixing the right variable gives a measurable
function. A Markov kernel may equivalently be defined as a stochastic map, i.e. a measurable
3
function X → ProbMeas(Y ), where ProbMeas(Y ) is the set of all probability measures on Y
equipped with the smallest σ-algebra for which the evaluation function evE : ProbMeas(Y ) →
[0, 1], defined by ProbMeas(Y ) ∋ ν 7→ evE(ν) := ν(E), is measurable for all E ∈ MY . In other
words, in the case of measurable spaces, stochastic maps are in one-to-one correspondence with
Markov kernels. However, in the case of Polish spaces, many applications demand a more restrictive
class of Markov kernels that have additional continuity properties, but demanding that the Markov
kernel is continuous is too restrictive in several contexts (see Remarks 4.72 and 4.142). Fortunately,
one can use the second perspective and define stochastic maps as continuous functions X →
ProbMeas(Y ), where ProbMeas(Y ) is now equipped with the vague topology. As a result, it is
a-priori unclear whether the associated kernel function is Borel measurable after fixing the right
variable. Nevertheless, it follows from the fact that the spaces are Polish [19].
If Polish spaces are replaced by compact Hausdorff spaces, one can still define stochastic maps,
but the above proof of measurability of the associated kernels fails. To circumvent this difficulty,
we restrict our probability measures to be regular (this additional assumption is automatically
satisfied for Polish spaces). We first prove that the stochastic maps on such spaces have measurable
Markov kernels in Lemmas 4.64 and 4.76. Lemma 4.64 extends results due to Billingsley (see proof
of Theorem 2.1 in [5]). This leads to an explicit definition of composition of such stochastic maps
in Proposition 4.79 without referring to the Riesz-Markov-Kakutani Representation theorem as is
done in [15] and [12]. We then prove that stochastic maps on compact Hausdorff spaces form a
category, denoted by cHausStoch, in Theorem 4.102.
In addition, we provide an explicit and geometric construction of the Gelfand spectrum functor
from commutative C ∗-algebras and positive maps to the category of compact Hausdorff spaces and
continuous stochastic maps without using the using algebras of functions and without appealing to
the Riesz-Markov-Kakutani Representation Theorem. Instead, this "stochastic" Gelfand spectrum
functor is constructed using techniques from convex analysis and Choquet theory in Theorems 4.134
and 4.151. Finally, we prove that the stochastic Gelfand spectrum functor exhibits a categorical
inverse to the continuous functions functor in Theorem 4.162.
Our results can be succinctly summarized in that we complete the diagrammatic cube (the op
superscript refers to flipping the directionality of all morphisms)
cHausop
cC*-Alg
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧/
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧/
B✆✆✆✆✆✆✆✆✆✆✆✆
B✆✆✆✆✆✆✆✆✆✆✆✆
FinSetFunop
fdcC*-Alg
cHausStochop
cC*-AlgPos
(1.1)
FinSetStochop
fdcC*-AlgPos
by adding cHausStoch and the functors into and from it. The arrows between the left and right
4
/
/
o
o
/
/
/
/
/
/
o
o
o
o
o
o
/
/
o
o
/
/
o
o
/
?
/
?
?
?
1
B
1
B
faces describe the equivalences between topology and analysis on the left and algebra on the right.
The arrows from top to bottom describe the generalization of going from deterministic dynamics
to stochastic dynamics. The arrows from front to back describe the inclusion of finite systems to
possibly infinite systems. The arrows from the left to the right are particularly important in physics
for the following reason. Much of classical and quantum dynamics, whether deterministic or non-
deterministic, can be formulated in a single category: C ∗-algebras and completely positive maps.
Although, classical evolution is typically thought of as being quite distinct from quantum evolution,
they both take place in the same category allowing a precise comparison between dynamics. In
particular, it allows mixing classical and quantum systems. For example, measurement of quantum
systems by macroscopic (classical) beings can be thought of as a map of C ∗-algebras and pulling
back states from a non-commutative C ∗-algebra to a commutative one (see Example 2.99). The
importance of commutative subalgebras in the theory of measurement and their relationship to
quantum mechanics has been emphasized and summarized nicely in Heunen's work [21].
In summary, we first analyze the front and right face in Section 2, then the top face in Section
3, and finally the category cHausStoch and how it fits into this cube via a stochastic Gelfand-
Naimark Theorem in Section 4. Concluding remarks, including a more detailed discussion of the
relationship between our work and others, are in Section 5. For the reader who has all the necessary
background, our main results are contained in Sections 4.3 through 5.1. The reader may find the
Index of notation at the end helpful.
Probability theory evolved before the formulation of measure theory, and, in particular, before
the birth of category theory. As such, it has developed its own language and culture, which the
author is largely unfamiliar with resulting in a potentially contentious presentation. Caveat lector.
Acknowledgements.
This work is based on a series of lectures given at the Analysis Learning Seminar at the
University of Connecticut, Storrs and the Mathematical Physics, Fourier Analysis and Applications
Seminar at the CUNY Graduate Center in New York during the Spring 2017 semester. The author
is grateful for the invitations by Matthew Badger and Vyron Vellis to the former and by Azita
Mayeli and Max Yarmolinsky to the latter. The author has greatly benefited from discussions
with Stefan Andronache, Iddo Ben-Ari, Behrang Forghani, Marcelo Nomura, Benjamin Russo,
Ambar Sengupta, Scott O. Wilson, and Yun Yang. The author also thanks Markus Haase for
several insightful comments and suggestions on an earlier version of this work. Parts of Sections 2
and 3 were worked out when the author was partially supported by the CUNY Graduate Center
Capelloni Dissertation Fellowship and NSF grant PHY-1213380.
2 Positive maps and stochastic matrices
2.1
Introduction
We start off with basic probability theory on finite sets and linear algebra. We first learned about
this perspective from the work of Baez and Fritz on entropy [2]. We find their notation, particularly
distinguishing between deterministic and non-deterministic processes, incredibly lucid. The former
. We introduce the
are depicted with straight arrows −→ while the latter with curvy arrows
5
/
/
/
o
/
o
concepts of C ∗-algebras and states on C ∗-algebras. There are two main categories described here
both of whose objects consist of C ∗-algebras. The difference occurs at the level of morphisms.
In one category, the morphisms consist of algebra ∗-homomorphisms and in the other case they
consist of completely positive maps. We will show that there is a correspondence between these
categories when restricted to finite-dimensional commutative C ∗-algebras and ordinary concepts
in probability theory, namely stochastic matrices. Thus, the category of operations in quantum
mechanics can be viewed as a non-commutative extension of the category of stochastic (non-
deterministic) processes in classical mechanics.
In this section, we introduce
i. FinSetFun, the category of finite sets and functions (deterministic processes),
ii. FinSetStoch, the category of finite sets and stochastic matrices (non-deterministic processes),
iii. fdcC*-Alg, the category of finite-dimensional commutative C∗-algebras and ∗-homomorphisms
(the non-commutative analogue of deterministic processes), and
iv. fdcC*-AlgPos, the category of finite-dimensional commutative C∗-algebras and completely
positive maps (the non-commutative analogue of stochastic maps)
and show that there are functors (the op superscript will be explained in Theorem 2.72)
FinSetFunop
fdcC*-Alg
FinSetStochop
fdcC*-AlgPos
(2.1)
so that the horizontal functors are fully faithful and so that the diagram commutes.
These concepts will be rigorously defined and sufficient intuition will be provided to illustrate
that they are nothing more than simple ideas in probability theory. However, we do assume the
reader is (at least vaguely) familiar with the terms category, functor, and natural transformation.
2.2 Some categories for finite probability theory
We begin with probability theory as it might be more intuitive than algebra.
Definition 2.2. A finite probability space is a pair (X, p) consisting of a finite set X and a function
p : X → R satisfying
and
The elements of X are called events and p is called a probability measure on X. The notation
px := p(x) will often be used.
(2.3)
(2.4)
∀ x ∈ X
p(x) = 1.
p(x) ≥ 0
Xx∈X
6
/
/
/
/
Note that p associates a number to any subset E ⊆ X given by
px
(2.5)
p(E) :=Xx∈E
qy = Xx∈f −1(y)
and is the reason we refer to p as a measure (more on this will be discussed in Section 4). Tech-
nically, these subsets are called events.
Definition 2.6. Let (X, p) and (Y, q) be two finite probability spaces. A probability-preserving
function from (X, p) to (Y, q) is a function f : X → Y satisfying
px
(2.7)
for all y ∈ Y.
Composition is given by the usual composition of functions and the composition of two probability-
preserving functions is immediately seen to be probability-preserving. The meaning of a probability-
preserving function between probability spaces can be seen nicely in the following example.
Example 2.8. Consider the set
X :=(cid:26) • , •
• , ••• , •
•
• (cid:27)
•• ••
•••
• , •
• , •
•
with the probability measure given by 1
6 for each element. Consider the set
Y := {O, E}
(2.9)
(2.10)
consisting of just two elements (O stands for "odd" and E stands for "even") with probability
measure given by 1
2 for each element. Also consider the function f : X → Y defined by
O
E
o❫❫❫❫❫❫❫❫❫❫❫❫❫❫❫❫❫❫❫❫❫
s❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣
t✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐
j❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯
k❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲
p❵❵❵❵❵❵❵❵❵❵❵❵❵❵❵❵❵❵❵❵❵
•
•
•
•••
•
•
•
•
•••
•
•
•
•• ••
•
(2.11)
,
•••
•••
• to O and sending •
, and •
•• ••
• to E. Therefore, the probability-
sending •
, and •
preserving function f is associated to the process of rolling a die and considering the likelihood
of rolling an odd or an even roll as opposed to rolling a 1, 2, 3, 4, 5, or 6. This describes a
deterministic process since we know with certainty that if we see odd, the die must be 1, 3, or 5,
and analogously for even.
,
•
•
•
•
•
Notation 2.12. Let FinProb be the category whose objects are finite probability spaces (X, p)
and whose set of morphisms (cid:8)(X, p) → (Y, q)(cid:9) are probability-preserving functions from (X, p)
to (Y, q). Similarly, let FinSetFun be the category whose objects are finite sets and whose mor-
phisms are functions. Composition in both categories is defined to be the ordinary composition of
functions.
7
s
o
j
t
p
k
The category FinProb was recently used by Baez, Fritz, and Leinster to provide a new cat-
egorical characterization of entropy in the case of finite sets [3]. Extensions of this result were
obtained more recently for relative entropy on finite sets by Baez and Fritz [2] and relative entropy
on Polish spaces by Gagn´e and Panangaden [17]. In all cases, categories of stochastic maps were
crucial in these characterizations. While functions take in one input and spit out one output,
stochastic maps take in one input and have a "spread" of outputs described by some probability
distribution.
Definition 2.13. Let X and Y be two finite sets. Let Pr(Y ) denote the set of probability measures
on Y. A stochastic map from X to Y is a function
f
−→ Pr(Y )
X
x 7→ f (x)
whose evaluation on elements in Y is written as
Y ∋ y
f (x)
7−−→ fyx ∈ R≥0.
Note that by definition of f (x) being a probability measure on Y, this means that
fyx = 1
Xy∈Y
(2.14)
(2.15)
(2.16)
for all x ∈ X. The numbers {fyx} labelled by x ∈ X and y ∈ Y form what is often called
Y following the
a stochastic matrix. We denote stochastic maps using squiggly arrows X
convention of Baez and Fritz [2] to distinguish stochastic maps from ordinary functions. By abuse
Y and we
of notation, we often use the same notation f for both f : X → Pr(Y ) and f : X
may often refer to X → Pr(Y ) as the stochastic map (the use of straight and curly arrows avoids
potential confusion).
Example 2.17. Let X and Y be finite sets and let f : X → Y be a function. Associated to f is
a canonical stochastic map f : X
Y defined by
Here the function δ : Y × Y → R is the Kronecker-delta function defined by
X × Y ∋ (x, y) 7→ fyx := δyf (x).
δyy′ :=(1
0
if y = y′
otherwise
.
(2.18)
(2.19)
In other words, f associates to x the probability measure that is 1 on the element f (x) and 0
elsewhere, i.e. if we have a function, we know with certainty where a point will go.
Example 2.20. Recall Example 2.8 with rolling a die.
If instead of interpreting the set Y as
the set of odd or even die, suppose that the O and E just stand for boxes labelled by O or E.
Y can be interpreted as saying that for any given
A probability-preserving stochastic map X
roll of a die, that particular side has a probability of being placed in box O and a complementary
probability of being placed in box E.
8
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
Example 2.21. A random walk on a circle is described by the following stochastic map. Let
N ∈ N be the number of points on a circle, labeled x0, x1, . . . , xN −1 with xN ≡ x0 and so on
so that X := {x0, x1, . . . , xN −1}. The random walk on X is described by the stochastic map
f : X
o X defined by
fji :=( 1
2
0
if j ≡ i ± 1 mod N
otherwise
(2.22)
where the shorthand notation fji is used instead of fxj xi. The reason this is called the random walk
is because starting at any point, you have a 1
2 probability of moving one step in either direction.
A random walk on a line or plane can also be described though the set of points would be infinite,
which we are not considering at the moment.
Example 2.23. With such a definition in place, notice that the set of all stochastic maps
o X from a single element set {•} to a finite set X is isomorphic (bijective) to the set
{•}
Pr(X) of all probability measures on X.
Given finite sets X, Y, and Z together with stochastic maps f : X
Z,
one would like to iterate these maps. To do so, the probability of x ∈ X evolving to z ∈ Z is
given by summing over all possible intermediate elements y ∈ Y weighted by their corresponding
probabilities.
Y and g : Y
Z
•
•z
•
gzy
Y
•y
•
•
•
•
fyx
X
•
•x
•
This motivates the following definition.
Definition 2.24. Let X, Y, and Z be finite sets. The composition of the stochastic map f :
Z, is the function
X
g ◦ f : X → Pr(Z) defined by sending x ∈ X to the probability measure on Z defined by
Y followed by the stochastic map g : Y
Z, written as g ◦ f : X
Z ∋ z
(g◦f )(x)
7−−−−→ (g ◦ f )zx :=Xy∈Y
gzyfyx.
(2.25)
Note that g ◦ f is indeed a stochastic map because
gzyfyx =Xy∈Y Xz∈Z
{z }1
def= Xz∈ZXy∈Y
(g ◦ f )zx
Xz∈Z
gzy
9
fyx =Xy∈Y
fyx = 1.
(2.26)
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
Notation 2.27. Let FinSetStoch be the category whose objects are finite sets and whose mor-
phisms are stochastic maps. We leave checking the axioms of a category to the reader.
Notice that FinSetFun can be viewed as a subcategory of FinSetStoch by Example 2.17 and
we therefore denote the associated functor by δ : FinSetFun → FinSetStoch.
Remark 2.28. The definition of a stochastic map is closely related to the notion of a stochastic
o X from a finite set X to itself can be used to describe
process. A stochastic map f : X
the possible evolution of a state. In the special case of a known initial condition, i.e. an element
x0 ∈ X, one can consider the sequence of probability distributions (the first element of which is
just the Kronecker delta distribution at x0) given by (cid:0)x0, f (x0), f 2(x0), f 3(x0), f 4(x0), . . .(cid:1). Such
a sequence is called a Markov chain.
If each iteration of f is interpreted as a time-step, then
the probability distribution f n(x0) is to be interpreted in the following way. The probability of
finding the initial condition x0 to evolve to the position xn ∈ X after n time steps is (f n)xnx0.
More generally, the set of events can change under a given process (as we have described above).
An evolution of some initial states can therefore be described by a sequence of stochastic maps
X0
o X1
o X2
o X3
· · · ,
(2.29)
where each of the Xi is a set of possible events. A point in X0, described by a function {•} → X0,
evolves just as a probability measure {•}
o X0, namely by precomposing:
{•}
o X0
o X1
o X2
o X3
· · · .
(2.30)
Example 2.31. Let us model one-dimensional diffusion via a stochastic map on a finite set,
thought of as an approximation to diffusion on the interval [−1, 1] with periodic boundary condi-
tions. Namely, let X be the finite set given by
X :=(cid:26) k
10
: k ∈ {−10,−9, . . . , 9}(cid:27) ,
(2.32)
where the point 1 is to be interpreted as being identified with −1. According to the heat diffusion
equation, after some time step, a particle at a point has the following probability distribution.
Approximately, it has a 0.56 likelihood to stay in the same place, 0.21 likelihood to move one
unit over in either direction, and a 0.01 likelihood to move two units over in either direction. The
stochastic matrix associated to this process is given by
(2.33)
0
· · ·
0.56 0.21 0.01
0.21 0.56 0.21 0.01 · · ·
0.01 0.21 0.56 0.21 · · ·
0
0.01 0.21 0.56 · · ·
...
...
. . .
· · ·
0.01
0
0.21 0.01
· · ·
...
0
0
...
0
0
0.01 0.21
0.01
0
0
0
0
0
...
...
0.56 0.21
0.21 0.56
Under a few iterations of this process (matrix multiplication), the distribution (described by a unit
vector) becomes the uniform distribution (as one would expect-systems tend to equilibriate). This
is depicted in Figure 1.
10
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
o
t = 0
•
t = 1
t = 2
• • • • • • • • • • • • • • • • • • • •
t = 3
•
•
•
t = 10
• • • • • • • • •
• • • • • • • • •
• • • • • • • • •
• • • • • • • • •
•
•
•
t = 100
• • •
• • • • • • • • •
• • • • • • • • •
• • • • • • • • • • • • • • • • • • • • •
• • • • • • • • • • • • • • • • • • • • •
Figure 1: A sharp Gaussian representing a definite initial condition at the origin tends
towards a uniform distribution after several iterations of the heat flow described by the
matrix (2.33). t = n with n ∈ {0} ∪ N stands for the n-th time step.
Note that the formalism described above a-priori only describes discrete stochastic processes.
One can adapt these ideas to study continuous processes though we do not discuss this here except
for an example from quantum mechanics in Section 2.7. The reader is referred to Giry's work,
which is not only a foundational paper on categorical probability theory but also discusses this in
some detail [19].
2.3 C∗-algebras and states
There are several definitions that must be provided so that this work can be somewhat self-
contained. A more thorough account of many definitions and results can be found in Dixmier's
book [11] with additional references provided throughout.
Definition 2.34. An (associative and unital) algebra is a vector space A (over C) together with
(a) a binary multiplication operation A × A → A,
(b) an element 1A ∈ A \ {0}.
These data must satisfy the following conditions:
i. (ab)c = a(bc) (associativity of multiplication),
ii. a(b + c) = ab + ac and (a + b)c = ac + bc (distributivity of multiplication over vector addition),
iii. (λa)b = a(λb) = λ(ab), (distributivity of multiplication over scalar multiplication), and
iv. a1A = 1Aa = a (1A is a unit for the multiplication),
for all a, b, c ∈ A and λ ∈ C.
11
Definition 2.35. An algebra A together with a norm k · k : A → R≥0 for which A with this norm
is a normed vector space and
is a normed algebra. A is a Banach algebra iff all Cauchy sequences converge with respect to this
norm.
kabk ≤ kakkbk
∀ a, b ∈ A
(2.36)
Definition 2.37. An algebra A together with an anti-automorphism ∗ : A → A, i.e.
and
(λa + b)∗ = λ∗a∗ + b∗
∀ λ ∈ C, a, b ∈ A
(ab)∗ = b∗a∗
∀ a, b ∈ A,
(2.38)
(2.39)
is an involutive algebra. ∗ is known as an involution on A.
Definition 2.40. Let A be both an involutive algebra and a normed algebra. If, in addition, A
satisfies
ka∗k = kak
∀ a ∈ A
(2.41)
then A is called an normed involutive algebra. If, in addition, all Cauchy sequences in A converge,
then A is called a involutive Banach algebra.
Definition 2.42. A C ∗-algebra is an involutive Banach algebra A for which
ka∗ak = kak2
∀ a ∈ A.
(2.43)
A C ∗-algebra is commutative iff ab = ba for all a, b ∈ A.
Remark 2.44. If A is a normed algebra with an involution for which kak2 ≤ ka∗ak for all a ∈ A,
then A is an involutive algebra. In fact, if A is a Banach algebra with involution satisfying this
condition, then A is a C ∗-algebra [11].
All of the data of these algebraic structures are summarized in the following table (with no
reference to their conditions).
data/
vector
structure
space (VS)
algebra
0
+
scalar mult.
X
X
X
×
k · k
∗
1
X
X
X
X
X
normed
VS
X
X
X
X
Banach
algebra
involutive
algebra (IA)
normed IA/
C ∗-algebra
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
Definition 2.45. Let A and B be two C ∗-algebras. A ∗-homorphism of C ∗-algebras from B to A
is a bounded (i.e. continuous) linear map f : B → A such that
12
i. f (b∗) = f (b)∗,
ii. f (b1b2) = f (b1)f (b2), and
iii. f (1B) = 1A
for all b, b1, b2 ∈ B.
Remark 2.46. The word bounded in the definition of a ∗-homomorphism of C ∗-algebras is re-
dundant. It is a theorem that such a map satisfying conditions i., ii., and iii.
is automatically
bounded [11]. Thus, the reader may safely ignore that condition.
There are a few basic examples of C ∗-algebras that we will focus on in this section.
Example 2.47. C is a C ∗-algebra with its usual structure associated with complex numbers. The
involution ∗ is the complex conjugate. This is an example of a commutative C ∗-algebra.
This previous example automatically induces a huge class of examples.
Example 2.48. Let X be a finite set and let CX denote the set of all functions X → C. There is
a unique algebra structure and involution on CX so that the evaluation functions evx : CX → C,
defined by sending ϕ : X → C to ϕ(x), are ∗-homomorphisms for all x ∈ X (these are just the
pointwise algebraic structures).2 The norm on CX is given by
CX ∋ f 7→ kfk := sup
x∈X{f (x)}.
(2.49)
The previous examples were commutative C ∗-algebras. The following example is the quintessen-
tial C ∗-algebra of quantum mechanics of a finite system (such as a qubit).
Example 2.50. Fix n ∈ N. The set Mn(C) of n× n complex-valued matrices is a C ∗-algebra with
product given by matrix multiplication and ∗ is the transpose complex conjugate. The norm can
be taken to be either the operator norm or the standard Euclidean norm (the topology will be the
same so it does not matter).
It is a theorem that every finite-dimensional C ∗-algebra is (∗-isomorphic to) a direct sum of
these (Theorem III.1.1 in [10]).
Example 2.51. More generally, for any Hilbert space H, the set of bounded operators B(H) on
H is a C ∗-algebra via operator composition, ∗ given by the adjoint, and the norm given by the
operator norm.
Notation 2.52. Let C*-Alg, cC*-Alg, and fdcC*-Alg denote the categories whose objects are
C ∗-algebras, commutative C ∗-algebras, and finite-dimensional commutative C ∗-algebras, respec-
tively, and whose the morphisms are all ∗-homomorphisms of C ∗-algebras.
2This way of phrasing the natural algebraic structure on CX makes precise in what sense it is the "obvious"
algebraic structure. Note that X does not have to be finite for this statement. However, one must be careful about
the topology when X is infinite-demanding the smallest topology for which evx are all continuous would result in
the topology of pointwise convergence, which is not the one we will use later.
13
Definition 2.53. Given a C ∗-algebra A, a state on A is a bounded (i.e. continuous)3 linear
function ω : A → C such that ω(1A) = 1 and ω(a∗a) ≥ 0 for all a ∈ A.
The following theorem relates states to density matrices and motivates why density matrices
are the states of quantum mechanics as opposed to vectors (really, rays) in a Hilbert space [20].
Theorem 2.54. There is a one-to-one correspondence between states on Mn(C) and matrices
ρ ∈ Mn(C) satisfying tr(ρ) = 1 and ρ∗ = ρ (such matrices are known as density matrices). More
precisely, for any density matrix ρ, the function tr(ρ · ) : Mn(C) → C, sending A ∈ Mn(C) to
tr(ρA), is a state. Conversely, for any state ω : Mn(C) → C, there exists a unique density matrix
ρ such that ω = tr(ρ · ).
This theorem extends to states on B(H) with H a separable Hilbert space and states that
satisfy an additional continuity assumption.4 In the context of quantum mechanics, B(H) is the
set of physical observables including non self-adjoint operators, such as ladder operators, so that
it remains closed under multiplication by complex numbers. A state on B(H) assigns to every
observable a number, interpreted as the expectation value of that observable for that state. States
on C ∗-algebras and their relationship to the more standard approach of quantum mechanics in
terms of Hilbert spaces is described in a categorical framework in [32].
2.4 Two types of morphisms of C∗-algebras
There is another notion of a morphism that is important for C ∗-algebras and utilizes less of its
structure.
Definition 2.55. Let A be a C ∗-algebra. An element a ∈ A is positive iff there exists an element
b ∈ A such that a = b∗b. Positivity of elements induces a partial ordering on A and the notation
a ≥ 0 will be used to indicate that a is positive (b ≥ a iff b − a ≥ 0).
Definition 2.56. Let A and B be two C ∗-algebras. A positive (unital) map from B to A is a
bounded linear function ϕ : B → A such that ϕ(1B) = 1A and ϕ(b∗b) ≥ 0 for all b ∈ B.
A state is an example of a positive map. Every ∗-homomorphism of C ∗-algebras is a positive
map. The converse, however, is far from true. Examples are given in Section 2.7. Non-unital
positive maps seem to be the weakest structure needed to make sense of operations and processes
on quantum mechanical systems though one may also relax the unit-preserving condition [25].
Notation 2.57. Let A be an involutive algebra and let n ∈ N. Mn(A) is the set of of n × n
matrices with coefficients in A with product given by matrix multiplication and ∗ given by the
transpose and applying ∗ in each entry.
3The word bounded/continuous here is redundant. It is a theorem that such a map on a C ∗-algebra satisfying
these conditions is automatically bounded.
4Such states go by the name of normal states for those readers who would like to look up more information.
However, normal states require a certain notion of continuity not available on arbitrary C ∗-algebras, and one
requires the notion of a von Neumann algebra, a special type of C ∗-algebra that is closed with respect to a certain
topology [36]. B(H) is an example of a von Neumann algebra.
14
Theorem 2.58. Let A be a C ∗-algebra. Then there exists a unique norm k · k on Mn(A) such
that
and
kABk ≤ kAkkBk
∀ A, B ∈ Mn(A)
kA∗Ak = kAk2
∀ A ∈ Mn(A).
Furthermore, Mn(A) with this structure is a C ∗-algebra.
Proof. See Chapter 1 of Paulsen [33] or Chapters 4 and 6 of Warner [49].
Remark 2.61. There are ∗-isomorphisms
Mn(A) ∼= Mn(C) ⊗ A ∼= A ⊗ Mn(C).
(2.59)
(2.60)
(cid:4)
(2.62)
Therefore, all of these involutive algebras have a unique norm giving them the structure of C ∗-
algebras. The theory of tensor products of C ∗-algebras is notoriously complicated. Since we will
work with commutative C ∗-algebras for the most part, we will avoid such complications.
Definition 2.63. Let A and B be two C ∗-algebras. A completely positive (unital) map from B to
A is a function ϕ : B → A such that ϕ ⊗ idn : B ⊗ Mn(C) → A ⊗ Mn(C) is a positive map for all
n ∈ N. The abbreviation "CP map" will often be used for such maps.
Remark 2.64. One motivation for this definition comes from quantum mechanics [27], [25]. The
concept of tensoring with the algebra Mn(C) and the positive map with the identity means to
adjoin another finite system that does not interact with the original system on B during the
operation. The resulting map ϕ⊗ idn : B⊗ Mn(C) → A⊗ Mn(C) should therefore still be positive.
There are examples of positive maps that are not completely positive so requiring this condition
to be satisfied is an additional constraint [33], [25]. Unital maps of C ∗-algebras correspond to
the trace-preserving condition that is often referred to in texts on quantum information [30]. See
Example 2.93 for an illustration. All maps in this work are unital so we will not use the more
standard notation "CPU" [30].
The composition of CP maps is CP [33].
Theorem 2.65. Let A be a C ∗-algebra. Every positive map A → C is CP. In particular, a state
on A is CP. Furthermore, every positive map between commutative C ∗-algebras is CP.
Proof. See Theorems 3.9 and 3.11 in [33].
Notation 2.66. Henceforth, a ∗-homomorphism from a C ∗-algebra B to another C ∗-algebra A
will be denoted by a straight arrow B → A. A CP map, on the other hand, will be denoted with
a curvy arrow B
o A.
(cid:4)
We now collect all the categories of C ∗-algebras that will be needed in the prequel.
Notation 2.67. Let C*-AlgCP, cC*-AlgPos, and fdcC*-AlgPos denote the categories whose
objects are C ∗-algebras, commutative C ∗-algebras, and finite-dimensional commutative C ∗-algebras,
respectively, and whose the morphisms are all CP maps of C ∗-algebras (note that in the latter two
cases, CP maps are the same as positive maps).
15
/
/
/
o
/
On a given physical system corresponding to some C ∗-algebra, a state provides expectation
values for observables. One can define a category of C ∗-algebras equipped with a state together
with the various flavors of morphisms analogous to the case of finite probability. However, this is
tangential to our main goal and will follow as an immediate consequence anyway. Examples are
provided in Section 2.7, but first we connect algebra with probability theory.
2.5 From probability theory to algebra
Now, we will construct a functor C : FinSetStoch → fdcC*-AlgPos and show that it restricts
to a well-defined functor C : FinSetFun → fdcC*-Alg. The same notation for these functors will
be used as it will be clear from context which one is being referred to. The properties of these
functors will be discussed in Section 2.6. Associated to each finite set Y, one has the C ∗-algebra
CY discussed in Example 2.48. Note that any element v ∈ CY can be uniquely expressed as
where ey : Y → C is defined by
vyey,
v =Xy∈Y
Y ∋ y′ 7→ ey(y′) := δyy′
and vy ∈ C is given by v(y) =: vy. Hence, if X is another finite set, any CP map ϕ : CY
determined by its value on the set of elements {ey}y∈Y by linearity. Namely,
ϕ(v) =Xy∈Y
vyϕ(ey).
(2.68)
(2.69)
o CX is
(2.70)
Hence, to every stochastic map f : X
determined by
Y, associate the linear function C f : CY
o CX uniquely
C f (ey) :=Xx∈X
for all y ∈ Y. These assignments define the functor C.
Theorem 2.72. The assignment5
fyxex
FinSetStochop C−→ fdcC*-AlgPos
o CX(cid:17),
(cid:16)f : X
X 7→ CX
Y(cid:17) 7→(cid:16)C f : CY
where C f is defined in (2.71), is a well-defined functor.
(2.71)
(2.73)
Proof. The proof will be broken into three steps. It will be shown that C f is CP, the identity
stochastic map goes to the identity CP map, and the composition of stochastic maps gets sent to
the composition of CP maps.
5The op superscript is only meant to keep track that the directionality of the arrows get reversed. This is shown
in (2.71), where the domain of C f is CY instead of CX . The directionality change is also shown in (2.77).
16
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
i. By Theorem 2.65, it suffices to show that C f is positive. Every positive element in CY is of
the form
Xy∈Y
ayey,
with ay ≥ 0 ∀ y ∈ Y.
(2.74)
Hence, by linearity, it suffices to show that C f (ey), given by (2.71), is positive for every y ∈ Y.
By definition of f being a stochastic map, fyx ≥ 0 for all y ∈ Y and x ∈ X. Hence, C f (ey) is
a positive element in CX. Furthermore, C f is unital because
C f (1Y ) = C f Xy∈Y
ey! =Xy∈Y (cid:0)C f (ey)(cid:1) =Xy∈Y Xx∈X
fyxex =Xx∈X Xy∈Y
fyx!
}
{z
1
ex = 1X.
(2.75)
δx′xex = ex′
(2.76)
ii. The identity function id : X → X has the corresponding stochastic map given by the proba-
bility measure idx′x = δx′x for all x, x′ ∈ X. Therefore,
idx′xex′ =Xx∈X
C id(ex′) =Xx∈X
for all x′ ∈ X. Thus, C id is the identity CP map.
iii. Now suppose that f : X
Y and g : Y
o Z are two composable stochastic maps between
finite sets. This induces CP maps C f : CY
o CX and C g : CZ
o CY . Therefore,
C f(cid:0)C g(ez)(cid:1) = C f Xy∈Y
gzyey!
gzyC f (ey)
=Xy∈Y
gzyXx∈X
=Xy∈Y
=Xx∈X Xy∈Y
= C g◦f (ez)
ex
gzyfyx!
}
{z
(g◦f )zx
fyxex
(2.77)
for all z ∈ Z.
Theorem 2.78. The functor C : FinSetStochop → fdcC*-AlgPos from Theorem 2.72 restricts
to a well-defined functor C : FinSetFunop → fdcC*-Alg.
Proof. Using the notation of Theorem 2.72 and its proof, all that is left to show is that C f :
CX is a ∗-homomorphism when f is a function. By linearity, it suffices to prove that
CY
(cid:4)
17
/
/
/
o
/
o
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
o
C f (eyey′) = C f (ey)C f (ey′) for all y, y′ ∈ Y. This follows from6
C f (ey)C f (ey′)
C f (eyey′)
C f (δyy′ey)
δyy′C f (ey)
δyf (x)δy′f (x′)exex′
Xx,x′∈X
δyf (x)δy′f (x′)δxx′ex′
Xx,x′∈X
δyf (x)ex
δyy′Xx∈X
Xx∈X
δ2
yf (x)ex
δyy′Xx∈X
δyf (x)δy′f (x)ex
(2.79)
(cid:4)
2.6 From algebra to probability theory
We will show that the functors of Theorems 2.72 and 2.78 are fully faithful.
Definition 2.80. A functor F : D → E from a category D to a category E is fully faithful iff for
every pair of objects d and d′ in D, the function F : D(d, d′) → E(cid:0)F (d), F (d′)(cid:1) is a bijection (full
means this function is surjective and faithful means it is injective). Here, D(d, d′) denotes the set
of morphisms from d to d′ in the category D.
In other words, if there is a fully faithful functor D → E that is also injective on objects, this
allows one to study D in a broader context without altering the relations between its objects.
Being faithful means that no information is lost when viewing D as inside of E and being full
means that no extra information is added to D when viewing D inside of E. In particular, D when
viewed on its own has the same set (if it even is a set) of isomorphism classes of objects as when
viewing D inside of E.
Theorem 2.81. The functors C : FinSetStochop → fdcC*-AlgPos and C : FinSetFunop →
fdcC*-Alg from Theorems 2.72 and 2.78 are fully faithful and injective on objects.
Proof. Injectivity on objects is immediate from the definition of C. Fix any two sets X and Y and
a CP map ϕ : CY
Y
o CX. The goal is to show there exists a unique stochastic map f : X
6Diagrams like this are to be read from top to bottom going along either direction to recreate the thought-
process involved in proving the claim. The reason is because at some point near the bottom, a step which does not
seem obvious in one direction is perhaps more intuitive from the other direction (namely, δ2
yf (x) = δyf (x)).
18
u
u
g
g
/
/
/
o
/
/
/
/
o
/
o
such that C f = ϕ. Furthermore, when ϕ is a ∗-homomorphism, then the associated stochastic map
f is a function. These two claims are proven in two steps.
i. First note that for any y ∈ Y, ϕ(ey) is an element of CX and, because {ex}x∈X is a basis of
CX , there exist unique numbers {fyx ∈ C}x∈X such that
fyxex.
ϕ(ey) =Xx∈X
Positivity of ϕ demands that
fyx ≥ 0
for all x ∈ X (and y ∈ Y ). Since ϕ is linear and unital,
Xx∈X
ϕ(ey) =Xy∈Y Xx∈X
Since {ex}x∈X is a linearly independent set of vectors in CX , this implies
ex = 1X = ϕ(1Y ) = ϕ Xy∈Y
fyxex =Xx∈X Xy∈Y
ey! =Xy∈Y
fyx = 1.
Xy∈Y
This shows that for every x ∈ X, the assignment
Y ∋ y 7→ fyx
fyx! ex.
(2.82)
(2.83)
(2.84)
(2.85)
(2.86)
is a probability measure on Y. In other words, the numbers {fyx}y∈Y,x∈X define a stochastic
map f : X
Y. It is readily checked that C f = ϕ. This proves that C is full.
To see that C is faithful, suppose that f ′ : X
C f = C f ′
. In particular, this implies that for any y ∈ Y,
(ey)
C f (ey)
C f ′
Y is another stochastic map such that
Xx∈X
By linear independence of the set {ex}x∈X, this implies
Xx∈X
fyxex
f ′
yxex
fyx = f ′
yx
∀ x ∈ X.
(2.87)
(2.88)
Since this calculation was independent of y, this also proves the equality for all y ∈ Y. Hence,
f = f ′ proving that C is faithful. This concludes the proof that C is fully faithful.
ii. Using the same notation as in the previous step, it suffices to show that if ϕ : CY → CX is
Y comes from
actually a ∗-homomorphism, then the corresponding stochastic map f : X
19
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
a function f : X → Y. First notice that for any pair of distinct elements y, y′ ∈ Y, because ϕ
is a homomorphism,
ϕ(eyey′)
ϕ(ey)ϕ(ey′)
ϕ(0)
0
Xx∈X Xx′∈X
fyxfy′x′exex′
(2.89)
fyxfy′x′δxx′ex′
Xx∈X Xx′∈X
fyxfy′xex
Xx∈X
proving that
fyxfy′x = 0
∀ x ∈ X.
(2.90)
Putting this result aside for the time being, recall the result (2.85) which, together with the
fact that fyx ≥ 0 for all y ∈ Y and x ∈ X, implies that for each x ∈ X, there exists a y ∈ Y
such that fyx > 0 (for otherwise, the sum over all y would not be 1). For any other y′ ∈ Y \{y},
the result (2.90) then implies
fy′x = 0
∀ y′ ∈ Y \ {y}.
(2.91)
Using (2.85) again implies
(2.92)
In other words, for every x ∈ X, there exists a unique y ∈ Y such that fyx = 1 and fy′x = 0
for all y′ ∈ Y \{y}. Set f (x) to be this unique element y, i.e. f (x) := y. This assignment then
defines a function f : X → Y.
fyx = 1.
(cid:4)
It is a fact that every n-dimensional commutative C ∗-algebra is ∗-isomorphic to CX for some
set X with cardinality n (this will follow from results in Section 3). This, together with Theorem
2.81, says that FinSetStochop is equivalent to the category fdcC*-AlgPos consisting of finite-
dimensional commutative C ∗-algebras and positive maps. An explicit inverse will be constructed
in greater generality in Section 4.5.
2.7 Some quantum mechanics
Completely positive maps represent general processes that are allowed in quantum mechanics. We
will provide three different examples illustrating the versatility of the C ∗-algebraic approach. The
first two examples are actually examples of ∗-homomorphisms. We thank Stefan Andronache and
Marcelo Nomura for discussions leading to these examples.
20
Example 2.93. Let H be a finite-dimensional7 Hilbert space and H a self-adjoint operator,
thought of as a Hamiltonian. For each t ∈ R, the operator exp(cid:0)− itH
(cid:1) , where is the reduced
Planck constant, is unitary and describes unitary time evolution in quantum mechanics. Its action
on observables is given by the adjoint action (the Heisenberg picture)
B(H) → B(H)
A 7→ exp(cid:18)itH
(cid:19) A exp(cid:18)−
itH
(cid:19) .
(2.94)
Infinitesimally, this gives the differential equation that describes the time evolution of an observable
dA
dt
=
1
i
[A, H].
(2.95)
Now, given an initial state tr(ρ · ) : B(H)
notation) this pulls back to a new state tr(ρ · ) ◦ Adexp(− itH
) under this evolution
o C for some density matrix ρ (see Theorem 2.54 for
Ad
exp(− itH
)
B(H)
tr(ρ · )◦Ad
exp(− itH
)
C
B(H)
tr(ρ · )
and is given by the assignment
B(H) → C
A 7→ tr(cid:18)ρ exp(cid:18) itH
(cid:19) A exp(cid:18)−
itH
(cid:19)(cid:19) = tr(cid:18)exp(cid:18)−
itH
(cid:19) ρ exp(cid:18) itH
(cid:19) A(cid:19)
(2.96)
(2.97)
by cyclicity of the trace. By the uniqueness of density matrices representing states (Theorem 2.54),
this proves that the time evolution of ρ is given by the differential equation
dρ
dt
=
1
i
[H, ρ].
(2.98)
This is the quantum Liouville equation. Notice how it has the opposite sign of the evolution
for observables. Also notice that if ρ is a pure state, this reduces to Schrodinger's equation
and reproduces the Schrodinger picture. Thus, the C ∗-algebraic approach naturally incorporates
both perspectives. It is important to notice that the straight horizontal arrow in (2.96) indicates
deterministic evolution! The "non-determinacy" of this example only appears in the states and
their expectation values.
Example 2.99. Measurement is an example of a CP map, in fact a ∗-homomorphism. We think of
measuring a quantum mechanical system in terms of numbers. For example, in the Stern-Gerlach
experiment, we measure a particle (such as a silver atom) to be spin up or spin down depending
on its position on a screen after moving through a magnetic field [41]. A state of a quantum-
mechanical system is therefore reduced to a probability measure on the set of the eigenvalues of
the observable being measured.
7We use finite-dimensional Hilbert spaces to avoid technicalities involving domains of operators.
21
/
/
/
o
/
/
/
?
?
?
?
?
_
_
_
_
_
The general (finite-dimensional) situation is as follows.8 Let H be a finite-dimensional Hilbert
space, let A be a self-adjoint operator on H, and let ω be a state on B(H). Let σ(A) denote the
spectrum of A (in this case, σ(A) is just the set of eigenvalues of A). A is an observable and
σ(A) are the possible values that an observer will see when trying to measure the observable A.
It is a fact of life that an observer can only measure elements of σ(A) (this is the origin of the
"discreteness" of quantum mechanics). In this situation, what is the probability measure that an
observer expects to see on σ(A)? The resulting probability measure is obtained by pulling back ω
along the "measurement" map
m : Cσ(A) → B(H)
eλ 7→ Pλ,
(2.100)
where Pλ is the projection operator onto the eigenspace associated with the eigenvalue λ ∈ σ(A)
of A. This produces a state on Cσ(A) via pullback
Cσ(A)
m
B(H)
ω◦m
ω
C
(2.101)
By our analysis from Section 2.6, p := ω ◦ m precisely provides us with a probability measure
on σ(A). This probability measure is interpreted as the probability distribution of measuring the
corresponding eigenvalues of A with respect to the state ω. Note that this example does not have
an evolution of the form B(H) → B(K) between observables on Hilbert spaces H and K. Instead,
it is an "evolution" from the quantum phase space B(H) to the classical one σ(A). Hence, the
C ∗-algebraic formulation allows one to treat classical and quantum systems in the same category.
In both of the previous situations, the evolution was described by a ∗-homomorphism. In the
following example, we provide completely positive evolution that is not a ∗-homomorphism. Such
evolutions can be used to model noise and measurement among other things [30].
Example 2.102. Let(cid:8)↑i,↓i(cid:9) denote the standard Euclidean basis in C2. Let ↑ih↑ denote the
projection operator onto span(cid:0)↑i(cid:1) and similarly for ↓i. The states on B(C2) given by tr(cid:0)↑ih↑ · (cid:1)
and tr(cid:0) ↓ih↓
· (cid:1) can be interpreted as a qubit being either in the spin up or spin down state
with respect to an appropriately chosen coordinate frame. Consider an operation on the space of
all density matrices whose action on these two states is given by
↑ih↑ 7→ p↑ih↑ + (1 − p)↓ih↓
↓ih↓ 7→ (1 − p)↑ih↑ + p↓ih↓
(2.103)
for some fixed p ∈ [0, 1]. This is to be interpreted as an operation that has probability 1 − p of
flipping a spin up state to a spin down state and has probability 1 − p of flipping a spin down
B(C2) that
state to a spin up state. Does there exist a completely positive map f : B(C2)
8For infinite systems, one must use projection-valued measures [30].
22
/
/
\
\
\
\
\
B
B
B
B
B
/
/
/
o
/
o
&
E := √p(cid:20)1 0
0 1(cid:21)
B(C2) ∋ A 7→ E∗AE + F ∗AF.
F :=p1 − p(cid:20)0 1
1 0(cid:21)
(2.104)
(2.105)
pulls back the pure states above on the left in (2.103) to the states on the right in (2.103)? The
answer is yes. Let
and define f to be given by
One can check that f is completely positive and is not a ∗-homomorphism. The induced action
on the density matrices can be obtained by looking at the pullback of an initial density matrix ρ
B(C2) ∋ A 7→ tr(cid:0)ρf (A)(cid:1) = tr(ρE∗AE + ρF ∗AF ) = tr(cid:0)(EρE∗ + F ρF ∗)A(cid:1).
Hence, the resulting density matrix under such an operation is given by EρE∗ + F ρF ∗. One can
check that (2.103) holds under this transformation. So again, the C ∗-algebraic framework is able
to reproduce familiar operations such as this one, known as the bit flip channel [30], all in the
same category of physical processes.
(2.106)
A generalization of Example 2.93 to CP maps gives rise to Lindblad's equation describing
non-deterministic dynamics in (open) quantum systems [27].
3 An equivalence between spaces and algebraic structures
On their own, the study of topological spaces and continuous maps is quite different from the study
of algebra and homomorphisms. Topological spaces consist of sets equipped with a subset of the
power set (known as open sets) and continuous maps are functions whose pre-image function takes
open sets to open sets. Algebras are sets equipped with binary operations satisfying certain laws
such as associativity and maps between algebras are those that respect this algebraic structure.
In the plain context of set theory, it simply does not make sense to compare the two areas of
mathematics. However, viewing these two areas as a whole, which is achieved by viewing them
as categories, it makes sense to compare them. In fact, under additional suitable restrictions on
the respective sides, the two categories are equivalent. There are several other versions of such
equivalences between topological categories and algebraic ones [23], [22]. We will discuss one such
equivalence, the equivalence between the category of commutative (unital) C ∗-algebras and ∗-
homomorphisms with the category of compact Hausdorff topological spaces and continuous maps
following the expositions in Folland [13], Rudin [39], and Tao's online notes [47]. The references
already have well-written proofs of the main results. Our purpose here is to frame these results in
the appropriate categorical context and establish certain tools that will be used in Section 4 when
we relax the class of morphisms to include non-deterministic processes.
3.1 From spaces to algebras
In Section 2.5, we showed how given a finite set X, the set of functions CX can be made into a
C ∗-algebra in a natural manner. This can also be done if X is replaced with a compact Hausdorff
space and the function space is chosen to be the subspace of continuous functions from X to C.
23
Example 3.1. Let X be a compact Hausdorff space. Then the space of complex-valued continuous
functions, denoted by C(X), with the algebraic structure from CX and with norm given by the
uniform norm, i.e.
C(X) ∋ ϕ 7→ kϕk := sup
x∈X ϕ(x),
(3.2)
is a commutative C ∗-algebra. Cauchy completeness follows from the continuous (uniform) limit
theorem (see Theorem 7.12 in [37] or Theorem 21.6 in [29]). The constant function 1X is the
unit for C(X). Note that although compactness is needed for C(X) to be a C ∗-algebra with this
norm, it is not necessary that X be Hausdorff. Proving that C(X) is a C ∗-algebra under the only
assumption that X is compact is a good exercise.9 If X is not Hausdorff, there exist points that
cannot be separated by continuous complex valued functions so that continuous functions cannot
distinguish them (see Remark 3.65 and Theorem 4.102 for further details).
Proposition 3.3. Let X and Y be compact Hausdorff spaces and let f : X → Y be a continuous
function. The function C f : C(Y ) → C(X) defined by
C(Y ) ∋ β 7→ C f (ϕ) := ϕ ◦ f
(3.4)
is a ∗-homomorphism.
Proof. First note that C f (ϕ) is continuous since it is the composition of continuous functions. To
see that C f respects the product, let ψ, ϕ ∈ C(Y ) and let x ∈ X. Then
(3.5)
(cid:0)C f (ψϕ)(cid:1)(x) = (ψϕ)(cid:0)f (x)(cid:1)
by def'n of C f
=(cid:16)ψ(cid:0)f (x)(cid:1)(cid:17)(cid:16)ϕ(cid:0)f (x)(cid:1)(cid:17)
=(cid:0)(ψ ◦ f )(x)(cid:1)(cid:0)(ϕ ◦ f )(x)(cid:1)
=(cid:0)(ψ ◦ f )(ϕ ◦ f )(cid:1)(x)
=(cid:0)C f (ψ)C f (ϕ)(cid:1)(x)
by def'n of ψϕ
by def'n of ◦
by def'n of (ψ ◦ f )(ϕ ◦ f )
by def'n of C f
proving that C f (ψϕ) = C f (ψ)C f (ϕ). Linearity, ∗-preservation, and preservation of units follow
from analogous calculations.
(cid:4)
Notation 3.6. Let cHaus denote the category whose objects are compact Hausdorff spaces and
whose morphisms are continuous maps.
Proposition 3.7. The assignment
cHausop C−→ cC*-Alg
−→ Y(cid:17) 7→(cid:16)C(X) C f
(cid:16)X
X 7→ C(X)
f
←− C(Y )(cid:17)
from Example 3.1 and Proposition 3.3 defines a functor.
(3.8)
9Sketch of proof: start with a Cauchy sequence in C(X), use Cauchy completeness in C to obtain a candidate
function pointwise, show that the resulting function is bounded, then show that the sequence of functions converges
to it uniformly, and finally apply the uniform limit theorem to guarantee it is continuous.
24
Proof. From the definition of C, it is manifest that C is a functor: the identity idX : X → X gets
sent to C idX = idC(X) and the composition X
C(Z) ∋ ϕ 7→ C g◦f (ϕ) = ϕ ◦ (g ◦ f ) = (ϕ ◦ g) ◦ f = C f(cid:0)ϕ ◦ g(cid:1) = C f(cid:0)C g(ϕ)(cid:1) = (C f ◦ C g)(ϕ) (3.9)
−→ Z gets sent to the function
since composition of functions is associative.
(cid:4)
g
f
−→ Y
3.2 Some topological preliminaries
Given a Banach algebra A, a certain subspace, denoted by A∗, of the linear dual space A∨ has a
natural topology (the operator norm topology) coming from the norms on A and C. However, this
topology has too many open sets for some purposes. Another topology on A∗ that has fewer open
sets is the weak* topology. Fewer open sets allows more sets to be compact and more compact
sets means that sequences and nets have a better chance of converging (see Tao's notes for an
illuminating discussion [47]). There are several ways to describe this topology. The simplest for
our purposes is in terms of convergent nets, though it will also be useful to describe it in terms
of topological vector spaces whose topology is induced by a family of seminorms. This latter
perspective will also become more important in Section 4.
Definition 3.10. A seminorm on a vector space V is a function k · k : V → R satisfying
i. kvk ≥ 0 for all v ∈ V,
ii. kλvk = λkvk for all λ ∈ C, v ∈ V, and
iii. kv + wk ≤ kvk + kwk for all v, w ∈ V.
The only difference between a norm and a seminorm is that a norm satisfies the additional
condition that kvk = 0 implies v = 0.
Example 3.11. Let V be a vector space and (W,k · k) be a normed vector space. For any linear
map ϕ : V → W, the function k · kϕ : V → R defined by sending v ∈ V to (cid:13)(cid:13)ϕ(v)(cid:13)(cid:13) is a seminorm
on V. A particular case of interest is when W := C.
A single norm provides a natural topology via open balls of varying radii at different points.
While a single seminorm also provides a natural topology in a similar way, it is rarely Hausdorff
and does not enjoy many desirable properties (see Example 3.15). However, it does become a
topological vector space.
Definition 3.12. A topological vector space consists of a vector space V equipped with a topology
in which the addition of vectors + : V × V → V and scalar multiplication C × V → V are both
continuous.
Definition 3.13. A base on a set X is a collection B ⊆ P(X) of subsets of X satisfying
i. for each x ∈ X, there exists a B ∈ B such that x ∈ B,
ii. for every pair B1, B2 ∈ B and every x ∈ B1 ∩ B2, there exists a B ∈ B such that x ∈ B ⊆
B1 ∩ B2.
25
A base B as above can be used to construct a topology on X by setting τB to be the smallest
topology containing B. Explicitly, its elements are given by unions of all possible elements in B,
τB =([B∈C
B : C ⊆ B) .
This topology is called the topology generated by B.
Example 3.15. Let k · k be a seminorm on a vector space V. Then the set
where
nBr(v) ⊆ V : v ∈ V and r > 0o,
Br(v) :=(cid:8)w ∈ V : kv − wk < r(cid:9),
(3.14)
(3.16)
(3.17)
forms a base on V. With respect to the topology generated by this base, V is a topological vector
space. Furthermore, it is Hausdorff if and only if k · k is a norm (see Tao [47]). Br(v) is called the
open ball of radius r ≥ 0 centered at v ∈ V.
A sufficiently robust family of seminorms typically has a more manageable topology that nat-
urally arises frequently enough to merit study.
Notation 3.18. For n ∈ N, set n := {1, . . . , n}. Let X be a set and S ⊆ P(X) a subset of the
power set.10 The value of a function S : n → S at i ∈ n is denoted by Si instead of S(i).
Definition 3.19. A subbase on a set X is a collection S ⊆ P(X) of subsets of X such that the
set
B := [n∈{0}∪N( n
\i=1
Si
: n S−→ S)
(3.20)
Si is the empty intersection,
0
\i=1
of all finite intersections of elements in S is a base for X. Here,
which is taken to be X itself. The topology generated by S is the topology generated by B.
It is a fact that any collection of subsets of a set forms a subbase for some topology on that
set and that topology is the smallest topology containing those subsets (see Section 5 of [50]).
Proposition 3.21. Let V be a vector space with {k · kα}α∈A a family of seminorms on V indexed
by some set A. Let τα denote the topology associated to k · kα. Then S := [α∈A
τα is a subbase on V
and V with the topology generated by S is a topological vector space. Furthermore, V is Hausdorff
if and only if for each v ∈ V \ {0}, there exists an α ∈ A such that kvkα 6= 0. Such a family of
seminorms is said to separate points of V or is said to be a separating family.
Proof. See Theorem 5.14 and Proposition 5.16 in Folland [14].
(cid:4)
10S is a Fraktur "S."
26
Remark 3.22. The use of seminorms is not necessary in the conclusion that V is a topological
vector space. All that is required is that each of the topologies τα provide V with the structure of
a topological vector space [47].
Let V be a vector space and let V ∨ denote its linear dual (linear functions from V to C) In
many applications where V has a norm, V ∨ is too large to deal with. Instead, one uses the norm
to restrict to a certain subspace of linear functionals.
Proposition 3.23. Let V be a normed vector space with norm k · k. For χ ∈ V ∨, set (abusing
notation slightly)
kχk := infnM ≥ 0 : (cid:12)(cid:12)χ(v)(cid:12)(cid:12) ≤ Mkvk ∀ v ∈ Vo,
which can be infinite. Set V ∗ ⊆ V ∨ to be the subset given by
V ∗ :=(cid:8)χ ∈ V ∨ : kχk < ∞(cid:9).
(3.24)
(3.25)
Then k · k is a norm on V ∗ and V ∗ equipped with this norm is a Banach space, known as the
topological dual space of V. Elements of V ∗ are called bounded/continuous linear functionals on
V.
Proof. This is a general fact about operator norms (see Proposition 5.4 in [14]).
(cid:4)
Proposition 3.26. Let V be a vector space. For each v ∈ V, define k · kv : V ∨ → C by
V ∨ ∋ χ 7→ kχkv := χ(v).
(3.27)
Then k · kv is a seminorm on V ∨. Furthermore, for each χ ∈ V ∨ \ {0}, there exists a v ∈ V such
that χ(v) 6= 0, i.e. the family of seminorms {k · kv}v∈V is separating. The same statement holds
for V ∗ ⊆ V ∨ with the same family.
Proof. This immediately follows from the definitions.
(cid:4)
Definition 3.28. Let V be a vector space and let {k · kv}v∈V be the family of seminorms as in
Proposition 3.26. The Hausdorff topology generated by this family as described in Proposition
3.21 is known as the weak* topology on V ∗.
Proposition 3.29. Let V be a normed vector space. For each χ ∈ V ∗, ǫ > 0, n ∈ N, and αn : n →
V, set
Bαn
ǫ (χ) := \j∈nnξ ∈ V ∗ : kχ − ξkαn(j) < ǫo.
ǫ (χ) ⊆ V ∨ : ǫ > 0, n ∈ N, αn : n → V, χ ∈ V ∗o
B :=nBαn
(3.30)
(3.31)
Then
is a base for the weak* topology on V ∗.
Proof. This is a general fact for the topology generated by a separating family of seminorms-see
Section 2.4 in [34].
(cid:4)
27
In other words, a base for the weak* topology consists of finite intersections of balls coming
from the different seminorms.
Proposition 3.32. Let V be a vector space. A net11 χ : Θ → V ∗ in V ∗ converges to an element
lim χ ∈ V ∗ in the weak* topology if and only if
Proof. See Tao [47] or Section 5.4 of Folland [14].
lim
θ∈Θ(cid:0)χθ(v)(cid:1) = (lim χ)(v)
∀ v ∈ V.
(3.33)
(cid:4)
Nets are needed here because not all compact Hausdorff spaces are separable (see Section 4.2).
Proposition 3.34. Let V be a vector space and let V ∗ denote its topological dual space. The
weak* topology on V ∗ is the weakest topology (meaning that it has the fewest open sets) such that
the functions
V ∗ evv−−→ C
χ 7→ χ(v)
are continuous for all v ∈ V.
Proof. See Section 3.14 of [39].
(3.35)
(cid:4)
Theorem 3.36 (Banach-Alaoglu Theorem). Let V be a Banach space and let V ∗ denote its topo-
logical dual space. The closed unit ball (with respect to the norm on V ∗) in V ∗ is compact with
respect to the weak* topology.
Proof. See Theorem 5.18 in [14].
(cid:4)
This theorem is surprising because the closed (and bounded) unit ball is compact with respect
to the norm topology if and only if the normed vector space is finite-dimensional [47].
3.3 The spectrum of a commutative C∗-algebra
It is perhaps not so surprising that one can obtain a commutative C ∗-algebra from a space as
in Section 3.1. What is more surprising is that there is a way to go back from C ∗-algebras to
topological spaces. Although C ∗-algebras have a norm (and hence a topology), keep in mind that
a C ∗-algebra is a linear space so obtaining a non-trivial topology from such an object will require
some work.
Definition 3.37. Let A be a commutative Banach algebra. A character on A is a continuous
nonzero homomorphism χ : A → C of Banach algebras, i.e. χ is linear, satisfies χ(ab) = χ(a)χ(b)
for all a, b ∈ A, and satisfies the condition that there exists an a ∈ A for which χ(a) 6= 0.
Characters form a subset of A∗, the topological dual space of A. This dual space has a natural
topology coming from the norms on both A and C, but the topology of interest for us is the weak*
topology on A∗.
11A net is a function whose domain is a directed set (see Section 11 of [50]).
28
Definition 3.38. Let A be a commutative Banach algebra. The spectrum of A is the set
σ(A) :=nA
χ
−→ C : χ is a charactero .
(3.39)
equipped with the subspace topology coming from A∗ via the weak* topology.
Remark 3.40. The relationship between the spectrum of a commutative C ∗-algebra and the
spectrum of a certain operator on a Hilbert space is described in Proposition 1.15 in [13]. Briefly, if
A is bounded and normal, then the operator-theoretic spectrum of A is canonically homeomorphic
with the spectrum of the commutative unital ∗-subalgebra generated by A. This is the closure of
the set of polynomials in A and A∗. By definition of being normal, A and A∗ commute so that the
order does not matter in which the A's and A∗'s appear in such polynomials guaranteeing that
this C ∗-algebra is commutative.
Proposition 3.41. Let A be a Banach algebra and let χ ∈ σ(A) be a character. Then
i. χ(1A) = 1,
ii. χ(a) 6= 0 for all invertible elements a ∈ A, and
iii. χ(a) ≤ kak for all a ∈ A.
Proof. See Proposition 1.10 in [13].
(cid:4)
Proposition 3.42. Let A be a commutative Banach algebra. Then the spectrum σ(A) is a
nonempty compact Hausdorff space.
Proof. See Theorem 11.9 in [39]. σ(A) is nonempty because A is unital.
Proposition 3.43. Let B
function
−→ A be ∗-homomorphism of commutative Banach algebras. Then the
(cid:4)
f
σ(A) σf
−→ σ(B)
χ 7→ χ ◦ f
is continuous.
Proof. First, note that since
(3.44)
(3.45)
(3.46)
(χ ◦ f )(xy) = χ(cid:0)f (x)f (y)(cid:1) = χ(cid:0)f (x)(cid:1)χ(cid:0)f (y)(cid:1) = (χ ◦ f )(x)(χ ◦ f )(y)
for all x, y ∈ B, χ◦ f is indeed an element of σ(B). We will give two proofs of continuity, one using
open sets and the other using nets.
i. It suffices to show that the inverse image of a base element gets sent to an open set. By
Proposition 3.29 and by definition of the subspace topology, a base for the topology on σ(B)
is given by
σ(B) ∩ Bαn
ǫ (ξ) =(cid:8)χ ∈ σ(B) : kχ − ξkαn(j) < ǫ ∀ j ∈ n(cid:9)
29
over all ξ ∈ B∗, αn : n → B, ǫ > 0, n ∈ N. σf is continuous if12 for every χ ∈ σ(A) and every
basic set13 σ(B) ∩ Bαn
ǫ (χ ◦ f ) there exists an open set containing χ whose image is contained
in this set. In fact, the basic set Bf ◦αn
(χ) accomplishes this goal because
ǫ
which shows that
kχ − ξkf ◦αn(j) =(cid:12)(cid:12)(cid:12)χ(cid:16)f(cid:0)αn(j)(cid:1)(cid:17) − ξ(cid:16)f(cid:0)αn(j)(cid:1)(cid:17)(cid:12)(cid:12)(cid:12) = kχ ◦ f − ξ ◦ fkαn(j),
σf(cid:16)σ(A) ∩ Bf ◦αn
(χ)(cid:17) =nξ ◦ f ∈ σ(B) : kχ − ξkf ◦αn(j) < ǫ ∀ j ∈ no
ǫ
=nξ ◦ f ∈ σ(B) : kχ ◦ f − ξ ◦ fkαn(j) < ǫ ∀ j ∈ no
⊆ σ(B) ∩ Bαn
ǫ (χ ◦ f ).
(3.47)
(3.48)
Thus, σf is continuous.
ii. Let χ : Θ → σ(A) be a net converging to lim χ (in the weak* topology). The goal is to show
that the net σf ◦ χ : Θ → σ(B) converges to σf (lim χ) ≡ (lim χ) ◦ f. For every b ∈ B,
lim
θ∈Θ(cid:0)σf (χθ)(b)(cid:1) = lim
θ∈Θ(cid:16)χθ(cid:0)f (b)(cid:1)(cid:17) = (lim χ)(cid:0)f (b)(cid:1) =(cid:0)σf (lim χ)(cid:1)(b),
(3.49)
which establishes the required weak* convergence and thus shows that σf is continuous.
As the reader may have noticed, the proof was much shorter using nets. The technique of using
nets to prove continuity of various functions will be heavily used in Section 4.
Proposition 3.50. The assignment
(cid:4)
cC*-Algop σ−→ cHaus
A 7→ σ(A)
−→ A(cid:17) 7→(cid:16)σ(B)
(cid:16)B
f
σf
←−− σ(A)(cid:17)
from Definition 3.38 and Proposition 3.43 defines a functor.
(3.51)
Proof. The proof is completely analogous to the proof of Proposition 3.7.
(cid:4)
12Let f : X → X ′ be a function and let B and B′ be bases for topologies on X and X ′, respectively. f is
continuous if for every x ∈ X, and for any B ′ ∈ B′ with f (x) ∈ B ′, there exists a B ∈ B with x ∈ B such that
f (B) ⊆ B ′ (this is a simple exercise in the definitions).
13Technically, we should have said that for every χ ∈ σ(A) and every basic set B containing χ ◦ f, there exists
an open set U containing χ whose image is contained in B. However, due to the seminorms, it suffices to take the
basic set B containing χ ◦ f to be centered at χ ◦ f by choosing it to be sufficiently small.
30
3.4 The Gelfand transform
Up to this point, we have constructed functors
cC*-Algop
σ
C
/ cHaus
.
In general, the diagrams
cHaus
id
cHaus
cC*-Algop
_❄❄❄❄❄❄❄❄❄
σ
⑧⑧⑧⑧⑧⑧⑧⑧⑧
C
&
cC*-Alg
id
cC*-Alg
_❄❄❄❄❄❄❄❄❄
C
⑧⑧⑧⑧⑧⑧⑧⑧⑧
σ
cHausop
(3.52)
(3.53)
do not commute so that σ and C are not inverses of each other. However, they are close. In the
present section, we will construct natural isomorphisms
cHaus
cHaus
id
h
_❄❄❄❄❄❄❄❄❄
⑧⑧⑧⑧⑧⑧⑧⑧⑧
σ
C
cC*-Algop
&
cC*-Alg
id
cC*-Alg
_❄❄❄❄❄❄❄❄❄
C
Γ
⑧⑧⑧⑧⑧⑧⑧⑧⑧
σ
cHausop
(3.54)
indicating a precise sense in which σ and C are close to being inverses of each other. A natural
isomorphism between functors is analogous to a homotopy between continuous functions or a uni-
tary intertwiner between representations and what it amounts to more precisely will be described
presently. This will show that the categories cC*-Algop and cHaus are equivalent.
Notation 3.55. Let A be a commutative Banach algebra and set
A
ΓA−→ C(cid:0)σ(A)(cid:1)
a 7→(cid:16)χ 7→(cid:0)ΓA(a)(cid:1)(χ) := χ(a)(cid:17).
ΓA is called the Gelfand transform on A.
Proposition 3.57. The following facts are true regarding the Gelfand transform.
(3.56)
i. For every commutative Banach algebra A, the Gelfand transform ΓA : A → Cc(σ(A)) is a
homomorphism of Banach algebras (preserving units).
ii. If A is a commutative C ∗-algebra, then ΓA is a ∗-isomorphism (in particular, it is isometric).
Proof. See Theorem 1.13 parts (a) and (d) and Theorem 1.20 in [13] for the first and second claims,
respectively.
(cid:4)
Remark 3.58. If A is just an involutive Banach algebra, then ΓA(a∗) is not always equal to
(cid:0)ΓA(a)(cid:1)∗ for all a ∈ A. Hence, ΓA is not necessarily a ∗-homomorphism (see Proposition 1.14 in
[13]).
31
/
o
o
_
o
o
_
o
o
_
o
o
_
o
o
Proposition 3.59. The assignment
cC*-Algop
0
1
Γ−→ cC*-Algop
A 7→(cid:16)A
ΓA−→ C(cid:0)σ(A)(cid:1)(cid:17)
from Definition 3.55 defines a natural isomorphism Γ : idcC*-Alg ⇒ C ◦ σ.
Proof. Associated to any ∗-homomorphism B
f
ΓB
ΓA
B
f
A
−→ A of C ∗-algebras is the diagram
C(cid:0)σ(B)(cid:1)
C(cid:0)σ(A)(cid:1)
C σ(f )
.
(3.60)
(3.61)
To see that it commutes, let b ∈ B. Going along the top and right arrows gives C σ(f )(cid:0)ΓA(b)(cid:1) =
ΓA(b) ◦ σf by definition of C as a functor (see Proposition 3.7). Going along the left and bottom
arrows gives ΓA(cid:0)f (b)(cid:1). To see that these two elements of C(cid:0)σ(A)(cid:1) are equal, let χ ∈ σ(A). Then
(3.44)
ΓB(b)(cid:16)σf(cid:0)χ(cid:1)(cid:17)
ΓB(b)(cid:0)χ ◦ f(cid:1)
(3.56)
(3.56)
ΓA(cid:0)f (b)(cid:1)(cid:0)χ(cid:1)
χ(cid:0)f (b)(cid:1)
(χ ◦ f )(b)
,
(3.62)
which proves that Γ is a natural transformation. It is a natural isomorphism by ii. of Proposition
3.57.
(cid:4)
Proposition 3.63. Let X be a compact Hausdorff space. The assignment
hX : X → σ(cid:0)C(X)(cid:1)
x 7→(cid:16)C(X) ∋ f 7→(cid:0)hX (x)(cid:1)(cid:0)f(cid:1) := f (x)(cid:17)
is a well-defined homeomorphism.
(3.64)
Proof. See Theorem 1.16 of [13]. and Example 2 after Theorem 1.30 in [13].
(cid:4)
Remark 3.65. If we had only assumed that X was compact and not Hausdorff, then hX would
not be one-to-one. To see this, let x, x′ ∈ X be two distinct points that cannot be separated by
open sets. For any continuous function ϕ : X → C, the value of ϕ on these two points is forced to
be the same, ϕ(x) = ϕ(x′), i.e. hX(x) = hX(x′).
Proposition 3.66. The assignment
cHaus0
h−→ cHaus1
X 7→(cid:16)X
hX−→ σ(cid:0)C(X)(cid:1)(cid:17)
from Proposition 3.63 defines a natural isomorphism h : idcHaus ⇒ σ ◦ C.
32
(3.67)
/
/
/
/
Proof. Associated to any morphism X
f
−→ Y of compact Hausdorff spaces is the diagram
X
f
Y
hX
hY
σC(f )
σ(cid:0)C(X)(cid:1)
σ(cid:0)C(Y )(cid:1)
.
(3.68)
To see that this diagram commutes,
let x ∈ X. Going along the top and right arrows gives
σC(f )(cid:0)hX(x)(cid:1) = hX (x) ◦ C f by definition of σ as a functor (see Proposition 3.43). Going along the
left and bottom arrows gives hY(cid:0)f (x)(cid:1). To see that these two elements of σ(cid:0)C(Y )(cid:1), i.e. non-zero
characters C(Y ) → C, are equal, let ϕ : Y → C be a continuous function on Y. Then
(3.4)
(cid:0)hX (x)(cid:1)(cid:0)C f (ϕ)(cid:1)
(cid:0)hX (x)(cid:1)(cid:0)ϕ ◦ f(cid:1)
(3.64)
(3.64)
(cid:16)hY(cid:0)f (x)(cid:1)(cid:17)(cid:16)ϕ(cid:17)
ϕ(cid:0)f (x)(cid:1)
(ϕ ◦ f )(x)
,
(3.69)
which proves that h is a natural transformation.
3.63.
It is a natural isomorphism by Proposition
(cid:4)
3.5 The commutative Gelfand-Naimark equivalence
The following theorem is a categorical phrasing of a theorem due to Gelfand and Naimark (Lemma
1 in [18]). It describes in what sense the functors C and σ are inverses of each other. The natural
transformations Γ and h also satisfy a universal property that indicates in what sense they are
inverses of each other.
Theorem 3.70. Using the notation from above,
(cid:16)cC*-Algop σ−→ cHaus, cHaus
is an adjoint equivalence of categories.
C−→ cC*-Algop, id Γ=⇒ C ◦ σ, σ ◦ C h−1
==⇒ id(cid:17)
(3.71)
Proof. Because of Propositions 3.59 and 3.66, it suffices to check the zig-zag identities. The first
one is given by
✈
✁
cHaus
C
❥ ❴ ❚ ■
idC
❂
C /
cC*-Algop
σ
h−1
idcHaus
idcC*-Algop
Γ
cHaus
❂
■
C /
cC*-Algop
= cHaus
idC
cC*-Algop
(3.72)
C
✁
idC
❚ ❴ ❥ ✈
C
33
C
/
/
/
/
/
/
/
/
;
;
#
#
A
A
✤
✤
✤
✤
✤
✤
✤
✤
:
:
$
$
which translates to commutativity of the diagram14
C(cid:16)σ(cid:0)C(X)(cid:1)(cid:17)
?⑧⑧⑧⑧⑧⑧⑧
_❄❄❄❄❄❄❄
idC(X)
ΓC(X)
C(X)
C(h−1
X )
C(X)
(3.73)
for every compact Hausdorff space X. Because all morphisms here are invertible, it is equivalent
to show that the diagram
C(cid:16)σ(cid:0)C(X)(cid:1)(cid:17)
_❄❄❄❄❄❄❄
⑧⑧⑧⑧⑧⑧⑧
idC(X)
ΓC(X)
C(X)
C(hX )
C(X)
(3.74)
(3.75)
(3.78)
(cid:4)
commutes, which would provide a sense in which h is the left inverse of Γ. Thus, let ϕ ∈ C(X).
Applying the composition along the top two arrows to this element gives
C hX(cid:0)ΓC(X)(ϕ)(cid:1) = ΓC(X)(ϕ) ◦ hX
by Proposition 3.3. This is a map X → σ(cid:0)C(X)(cid:1) → C and is therefore determined pointwise so
let x ∈ X and apply this map to it. The result is
which proves the first zig-zag identity. The other zig-zag identity
(cid:0)ΓC(X)(ϕ)(cid:1)(cid:0)hX(x)(cid:1) (3.56)
== hX(x)(cid:0)ϕ(cid:1) (3.64)
== ϕ(x),
(3.76)
idcC*-Alg
Γ
σ
cHausop
C /
cC*-Alg
❂
✁
■
idσ
❚ ❴ ❥ ✈
σ
✁
cC*-Alg
h−1
idcHausop
σ
❥ ❴ ❚ ■
idσ
✈
❂
σ
σ
cHausop
= cC*-Alg
idσ
cHausop
(3.77)
σ
follows by a similar calculation. For a commutative C ∗-algebra A, this identity says that the
diagram
σ(cid:16)C(cid:0)σ(A)(cid:1)(cid:17)
_❄❄❄❄❄❄❄
⑧⑧⑧⑧⑧⑧⑧
idσ(A)
σ(ΓA)
σ(A)
hσ(A)
σ(A)
commutes, which provides a sense in which h is the right inverse of Γ.
The upshot of the Gelfand-Naimark Theorem that we emphasize here is that it provides one
with the perspective that the study of topology (compact Hausdorff spaces) and their continu-
ous functions is equivalent to the study of commutative C ∗-algebras and their ∗-homomorphisms.
14The arrows are in the direction required by contravariance of the functors.
34
o
o
_
?
o
o
_
/
/
/
/
/
;
;
#
#
B
B
✤
✤
✤
✤
✤
✤
✤
✤
;
;
#
#
o
o
_
Therefore, non-commutative C ∗-algebras can be interpreted as non-commutative topology though
a more satisfying relationship to topological concepts is still an area of active research [21]. Fur-
thermore, the ∗-homomorphisms between these non-commutative spaces can be interpreted as
deterministic processes. In the next section, we will describe how to fit in stochastic maps (non-
deterministic processes) and motivate non-commutative probability theory.
4 Abstract probability theory
The table below collects several of the categories that have been used along with a few new ones.
Category name
FinSetFun
FinSetStoch
cHaus
cHausStoch
fdcC*-Alg
Objects
finite sets
finite sets
compact Hausdorff spaces
compact Hausdorff spaces
finite-dimensional commutative C ∗-algebras
fdcC*-AlgPos finite-dimensional commutative C ∗-algebras
cC*-Alg
cC*-AlgPos
commutative C ∗-algebras
commutative C ∗-algebras
Morphisms
functions
stochastic maps
continuous functions
(continuous) stochastic maps
∗-homomorphisms
positive maps
∗-homomorphisms
positive maps
In this section, we define the category cHausStoch and generalize the (commutative) Gelfand-
Naimark Theorem to prove an adjoint equivalence of categories
cHausStochop
C
σ
/ cC*-AlgPos
.
(4.1)
During this procedure, we provide an explicit formula for the composition in cHausStoch in
Proposition 4.79 by showing that a stochastic map induces a canonical Markov kernel in Lemma
4.76. This formula was shown to be valid in the category of stochastic maps on Polish spaces in [19]
but seems to have not been verified in the case of compact Hausdorff spaces. One can nevertheless
define the composition by passing to the algebra of continuous functions and using the Riesz-
Markov-Kakutani Representation Theorem [12] or states [15] though these constructions are a
bit formal. In addition, we construct a stochastic version of the Gelfand spectrum functor that
does not use the Gelfand transform nor the Riesz-Markov-Kakutani Theorem in Theorem 4.145.
Instead, Choquet theory [35] is used to send states onto probability measures on the spectrum.
Several of the key ingredients in these constructions are standard [23], [9], [14], [16] though we
illustrate how these results fit into a broader context by showing how all of these categories fit into
the cube from (1.1) in the introduction.
Sections 4.1 and 4.2 review some measure-theoretic preliminaries. The main results begin in
Section 4.3 with the construction of the category cHausStoch and end in Section 4.6 with the
stochastic Gelfand-Naimark Theorem. The reader with the analytic background may skip ahead
directly to these sections. We discuss the relationship between our work and that of others in more
detail in Section 5.2.
35
/
o
o
4.1 Compact Hausdorff spaces and complex measures
Given a measurable space (Y,MY ), where Y is a set and MY is a set of measurable subsets,
known as a σ-algebra, one can define the notion of a positive measure and a complex measure.
Our measurable spaces will come from compact Hausdorff topological spaces equipped with the
Borel subsets for their σ-algebra. As a result, with the additional assumption of finite total
measure, the set of positive measures will be a subset of the set of complex measures. We review
the definitions, which will be important in specifying a particular topology on the set of all such
measures (Proposition 4.29), following Chapters 1 and 6 of [38]. We will obtain from this procedure,
a topology on the set of all probability measures on a compact Hausdorff space. For technical
reasons, we will then look at a particular subset of all measures satisfying a condition compatible
with respect to the topology of the underlying space-such measures are called regular measures
(also Radon measures). We will explain the technicalities and why they are needed in remarks.
Henceforth, all topological spaces will be assumed compact and Hausdorff and the set of measurable
subsets will always be taken to be the set of Borel subsets. A discussion of the pros and cons of
using Baire sets is deferred to Section 5.3. Most of this background material does not require the
measurable space to be compact Hausdorff though we phrase everything in this setting since all
our applications are for such spaces.
Definition 4.2. Let Y be a compact Hausdorff space with associated Borel measurable subsets
MY . A complex measure on (Y,MY ) is a function µ : MY → C that is countably additive in the
sense that for every measurable set E ∈ MY ,
µ(E) =Xi∈I
µ(Ei)
Ei ∩ Ej = ∅
[i∈I
Ei = E.
(4.3)
(4.4)
(4.5)
for every at most countable collection {Ei}i∈I⊆N of measurable sets satisfying15
whenever i 6= j and
Such a collection {Ei}i∈I is called a measurable partition of E. A positive measure is a complex
measure whose value on all measurable sets is non-negative. A probability measure is a positive
measure satisfying µ(Y ) = 1. The set of all complex measures on Y will be denoted by CMeas(Y ).
Occasionally, we may write Y instead of (Y,MY ) since our set of measurable sets will always
be Borel. Note that every positive measure can be used to define an integral using simple functions
(see Chapter 1 of [38]). Since the construction is used frequently for positive measurable functions,
we state it here while also setting some notation.
Definition 4.6. Let Y be a compact Hausdorff space and let E ∈ MY . The function χE : Y → R
defined by
Y ∋ y 7→ χE(y) :=(1
0
if y ∈ E
otherwise
(4.7)
15∅ denotes the empty set.
36
is called the characteristic function on E. A simple function is a function s : Y → R≥0 of the form
s =
siχEi
(4.8)
finite
Xi
for some finite set of non-negative numbers {si} and measurable subsets {Ei}. Let µ be a positive
measure on Y. For a measurable function f : Y → R≥0, the integral of f with respect to µ is the
number
ZY
f dµ := sup
0≤s≤f(Xi
siµ(Ei) : s =Xi
siχEi is simple) .
(4.9)
The integral of a complex measurable function can be defined by splitting the function up into
a sum of its real and imaginary parts each of which can be decomposed into a sum of positive
and negative parts. Similarly, every complex measure can be uniquely decomposed via Jordan
decomposition into an appropriate sum and difference of positive measures so as to define the
integral of a positive function with respect to a complex measure and then finally the integral of
any complex measurable function with respect to a complex measure (see Section 6.18 of [38]-we
will review a necessary fact to make sense of this definition in Theorem 4.23 below).
The following Lemma is useful for several of the facts that will follow. It is unfortunately long,
but is used so often that we felt it was important to state it in its full form.
Lemma 4.10. Let Z be a compact topological space and let ϕ : Z → R≥0 be a bounded measurable
function. Then there exists a sequence s : N → RZ of simple functions satisfying
sn(z) = ϕ(z) ∀ z ∈ Z.
Furthermore, the coefficients {sn,in} and measurable sets {En,in} in the expression
0 ≤ s1 ≤ s2 ≤ s3 ≤ · · · ≤ ϕ
lim
n→∞
(4.11)
&
sn =Xin
can be chosen so that in ∈ {0, 1, . . . , 2n − 1},
En,in = En+1,2in ∪ En+1,2in+1
sn,inχEn,in
&
En+1,2in ∩ En+1,2in+1 = ∅,
and
sn,in ≤ min(cid:8)sn+1,2in, sn+1,2in+1(cid:9)
for all in ∈ {0, 1, . . . , 2n − 1} and n ∈ N.
Proof. Since ϕ is bounded, there exists an M ≥ 0 such that ϕ(Z) ⊆ [0, M]. For each n ∈ N, set
h inM
h inM
for in ∈ [0, 2n − 1) ∩ Z
for in = 2n − 1,
where in ∈ {0, 1, . . . , 2n − 1}. Since ϕ is Borel measurable, the preimages En,in := ϕ−1(cid:0)Fn,in(cid:1) form
a measurable partition of Z for each n. Set sn : Z → R≥0 to be the simple function
Fn,in :=
2n , (in+1)M
2n , (in+1)M
2n (cid:17)
i
(4.15)
2n
(4.12)
(4.13)
(4.14)
sn :=
2n−1
Xin=0
inM
2n χEn,in .
37
(4.16)
R
F2,3n
F2,2n
F2,1n
F2,0n
R
F3,7{
F3,6{
F3,5{
F3,4{
F3,3{
F3,2{
F3,1{
F3,0{
E2,0
E2,1 E2,2
E2,3
E2,2 E2,1
E2,0
E3,0
E3,1
E3,3
E3,5
E3,7
E3,5
E3,3
E3,1
E3,2
E3,4
E3,6
E3,6
E3,4
E3,2
E3,0
Z
Z
Figure 2: A non-negative bounded measurable function is an increasing sequence of
simple functions satisfying several convenient properties as described in Lemma 4.10.
An example of such a simple function together with the next function in the sequence is depicted
in Figure 2. The sequence N ∋ n 7→ sn of simple functions satisfies the required properties because
for every n, the pointwise difference between sn and ϕ is at most M
2n . Furthermore, sn ≤ sn+1
for each n ∈ N because {Fn+1,in+1} is a refinement of {Fn,in} and hence likewise {En+1,in+1} is a
refinement of {En,in} (this was the reason for the choice of 2n as opposed to merely n in the above
formulas). The other properties are immediate from the construction.
(cid:4)
Theorem 4.17 (Monotone Convergence Theorem). Let Y be a compact Hausdorff space and let
ϕ : N → [0,∞]Y be an increasing sequence of Borel measurable functions, i.e. ϕn+1(y) ≥ ϕn(y)
for all y ∈ Y and n ∈ N. Then lim ϕ is Borel measurable and
lim ϕ dµ
(4.18)
lim
n→∞ZY
ϕn dµ =ZY
for every Borel measure µ on Y.
Proof. See Theorem 1.26 in [38].
Definition 4.19. Let Y be a compact Hausdorff space. Given a complex measure µ ∈ CMeas(Y ),
the total variation of µ is the function µ : MY → R defined by
(cid:4)
(4.20)
where the supremum is over all measurable partitions {Ei}i∈I of E.
MY ∋ E
µ
7−→ sup(Xi∈I (cid:12)(cid:12)µ(Ei)(cid:12)(cid:12)) ,
38
Proposition 4.21. Let Y be a compact Hausdorff space and µ ∈ CMeas(Y ). Then, the total
variation of µ is a positive measure on Y. Furthermore, the set of all complex measures on Y has
the structure of a normed complex vector space with norm given by
Proof. See Chapter 6 of [38].
CMeas(Y ) ∋ µ 7→ µ(Y ).
(4.22)
(cid:4)
Theorem 4.23 (Polar decomposition for complex measures). Let Y be a compact Hausdorff space
and µ ∈ CMeas(Y ). Then there exists a measurable function h : Y → C such that h(y) = 1 for
all y ∈ Y and
µ(E) =ZE
h dµ
∀ E ∈ MY .
This is often written in shorthand form as dµ = h dµ.
Proof. See Theorem 6.12 in [38].
(4.24)
(cid:4)
(4.25)
This is an important consequence of the Lebesgue-Radon-Nikodym Theorem.
It allows one
to define the integral of complex-valued measurable functions with respect to complex measures,
namely using the notation of Theorem 4.23, the integral of a complex-valued measurable function
ϕ : Y → C with respect to µ is
ZY
ϕ dµ :=ZY
ϕh dµ.
This definition satisfies the usual properties of the integral (see Section 6.18 in [38]). This theorem
can also be used to prove the following result, which we will need.
Lemma 4.26. Let Y be a compact Hausdorff space and µ ∈ CMeas(Y ). Then
(cid:12)(cid:12)(cid:12)(cid:12)
ZY
ϕ dµ(cid:12)(cid:12)(cid:12)(cid:12) ≤ZY ϕ dµ
for all ϕ ∈ C(Y ).
Proof. Using a polar decomposition dµ = h dµ,
(cid:12)(cid:12)(cid:12)(cid:12)
ZY
ϕ dµ(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)
ZY
ϕh dµ(cid:12)(cid:12)(cid:12)(cid:12) ≤ZY ϕh dµ =ZY ϕ dµ
for all ϕ ∈ C(Y ).
(4.27)
(4.28)
(cid:4)
Being a normed space, CMeas(Y ) has a natural topology on it. However, as for many infinite-
dimensional vector spaces, too few sequences have convergent subsequences with respect to this
topology. There are several ways to implement topologies on CMeas(Y ). Rather than using all
measurable functions, we will use continuous functions to define a collection of seminorms (recall
the discussion in Section 3.2).
39
Proposition 4.29. Let Y be a compact Hausdorff space. For every continuous function ϕ : Y →
C, set k · kϕ : CMeas(Y ) → R to be the function given by
CMeas(Y ) ∋ µ 7→ZY ϕ dµ.
(4.30)
Then k · kϕ is a seminorm on CMeas(Y ).
Proof. This follows immediately from the definition of the integral and the fact that the total
variation · defines a norm on CMeas(Y ). Note that the function k · kϕ makes sense even when
ϕ is just a measurable function.
i. kµkϕ =RY ϕ dµ ≥ 0 for all complex measures µ ∈ CMeas(Y ) follows immediately from the
ii. Fix λ ∈ C and µ ∈ CMeas(Y ). By Lemma 4.10, there exists a sequence s : N → RY of simple
definition of the integral of a positive function with respect to a positive measure.
functions on Y of the form
satisfying
and
sn :=
finite
Xin
sn,inχEn,in ,
0 ≤ s1 ≤ s2 ≤ s3 ≤ · · · ≤ ϕ
lim
n→∞
sn(y) = ϕ(y)
∀ y ∈ Y.
(4.31)
(4.32)
(4.33)
(4.34)
By the definition of the integral
lim
n→∞ZY
sn dλµ =ZY ϕ dλµ = kλµkϕ.
Meanwhile, by definition of simple functions, the left-hand-side of this expression equals
lim
n→∞ZY
sn dλµ = lim
n→∞
finite
n→∞
finite
finite
= lim
n→∞
λµ(En,in)sn,in
Xin=1
Xin=1
λµ(En,in)sn,in
Xin=1
= λ lim
n→∞ZY
sn dµ
= λ lim
= λZY ϕ dµ
= λkµkϕ.
µ(En,in)sn,in
(4.35)
Hence, kλµkϕ = λkµkϕ as needed.
40
iii. Let µ, ν ∈ CMeas(Y ) and let s be a sequence as in the previous part. A similar argument as
above shows that
kµ + νkϕ =ZY ϕ dµ + ν ≤ZY ϕ dµ +ZY ϕ dν = kµkϕ + kνkϕ.
All of the conditions of a seminorm have therefore been checked.
(4.36)
(cid:4)
Henceforth, we equip CMeas(Y ) with the topology generated by the family of seminorms
(cid:8)k · kϕ(cid:9)ϕ∈C(Y ). This topology is known as the vague topology on CMeas(Y ). It is helpful to know
when nets converge with respect to this topology so that we may use nets to prove whether or not
certain functions are continuous.
Proposition 4.37. Let Y be a compact Hausdorff space. A net µ : Θ → CMeas(Y ) converges to
lim µ with respect to the vague topology if and only if
lim
θ∈ΘZY
ϕ dµθ =ZY
ϕ d lim µ
(4.38)
for all ϕ ∈ C(Y ).
Proof. space
(⇒) Suppose that µ : Θ → CMeas(Y ) converges to lim µ. Fix ǫ > 0. By assumption, for any
continuous function ϕ ∈ C(Y ), there exists a θ0 ∈ Θ such that
∀ θ ≥ θ0.
klim µ − µθkϕ < ǫ
(4.39)
Then, by Lemma 4.26,
ϕ d lim µ −ZY
(cid:12)(cid:12)(cid:12)(cid:12)
ZY
ϕ d(lim µ − µθ)(cid:12)(cid:12)(cid:12)(cid:12)
ϕ dµθ(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)(cid:12)
ZY
≤ZY ϕ d lim µ − µθ
= klim µ − µθkϕ
∀ θ ≥ θ0.
< ǫ
(4.40)
θ∈ΘZY
ϕ d lim µ for all ϕ ∈ C(Y ). Fix ǫ > 0, n ∈ N, and α :
(⇐) Suppose that lim
n → C(Y ) whose value at i ∈ n will be denoted by αi (we will show that µ is eventually in a
neighborhood base of lim µ). By assumption, for each i ∈ n, there exist θi ∈ Θ such that
Set
Then, µθ ∈ Bα
topology.
ǫ (lim µ) for all θ ≥ θ0. Hence µ converges to lim µ with respect to the vague
ϕ dµθ = ZY
(cid:12)(cid:12)(cid:12)(cid:12)
ZY
ϕ d lim µ −ZY
41
< ǫ
∀ θ ≥ θi.
ϕ dµθ(cid:12)(cid:12)(cid:12)(cid:12)
θ0 := max
i∈n {θi}.
(4.41)
(4.42)
(cid:4)
4.2 Probability measures on compact Hausdorff spaces
Naively, one might use the set of all probability measures on Y and define suitably continuous
stochastic maps between compact Hausdorff spaces using this definition. However, in order to ob-
tain an equivalence of categories between such spaces and the category of commutative C ∗-algebras
and positive maps, an additional restriction needs to be made on the set of probability measures
to ensure that the equivalence holds. This restriction is solely due to the fact that not all compact
Hausdorff spaces are separable, or equivalently, metrizable (for instance, see counterexamples 24,
43, 107, and 111 from [45]). The measures we must restrict our attention to are regular (Radon)
measures. Since we will mainly be concerned with probability measures, we will immediately
assume throughout that all measures are finite.
Definition 4.43. Let Y be a compact Hausdorff space. A positive Borel measure µ : MY → R≥0
is regular iff
and
µ(E) = inf(cid:8)µ(U) : E ⊆ U with U open in Y(cid:9)
µ(E) = sup(cid:8)µ(K) : K ⊆ E with K compact in Y(cid:9)
(4.45)
for all E ∈ MY . A complex Borel measure µ : MY → C is regular iff the positive and negative
parts of the real and complex parts of µ are all regular.
(4.44)
Because being regular is a property of a measure, the set of regular measures on Y is a subset
CReg(Y ) ⊆ CMeas(Y ) of all complex measures. Since CMeas(Y ) has a normed vector space
structure, one could ask if CReg(Y ) inherits the properties of this structure as well.
Theorem 4.46. CReg(Y ) with the total variation norm is a Banach space for all compact Haus-
dorff spaces Y.
Proof. See Exercise 3 in Chapter 6 of [38].
(cid:4)
Notation 4.47. Because CReg(Y ) ⊆ CMeas(Y ), it also inherits the vague topology described
after Proposition 4.29. Let ProbMeas(Y ) denote the subset of CReg(Y ) of regular measures that
are also probability measures. Equip ProbMeas(Y ) with the subspace topology coming from the
vague topology on CMeas(Y ).
Notice that when Y is a finite set with the discrete topology, ProbMeas(Y ) = Pr(Y ).
Remark 4.48. When Y is just a locally compact Hausdorff space, one must use Radon probability
measures instead (see Theorem 7.2 in [14]). When Y is a compact Hausdorff space, it is a fact that
every Radon probability measure is regular (see Corollary 7.6 in [14]). When Y is, in addition,
second countable, every Borel measure on Y is regular (see Theorem 7.8 in [14]). Hence, many of
the complications that follow can be ignored provided that Y is second countable (and therefore
separable). In fact, many treatments of probability theory on topological spaces assume the spaces
are Polish. We do not need this requirement, which is one of the interesting features of our findings.
We mention several facts regarding regular measures on compact Hausdorff spaces that will be
needed for many of our proofs.
42
Lemma 4.49 (Urysohn's Lemma). Let Y be a compact Hausdorff space and let K ⊆ U ⊆ Y with
K closed (and hence compact) and U open. Then there exists a continuous function ϕ ∈ C(cid:0)Y, [0, 1](cid:1)
such that ϕ = 1 on K and ϕ = 0 outside a compact subset of U.
Proof. See Theorem 4.32 in [14].
(cid:4)
Definition 4.50. Let Y be a topological space. A function ϕ : Y → R is lower semi-continuous
iff f −1(cid:0)(a,∞)(cid:1) is open for all a ∈ R and is upper semi-continuous iff f −1(cid:0)(−∞, a)(cid:1) is open for all
a ∈ R.
It is immediate from the definition that semi-continuous functions are measurable. This is
because sets of the form (a,∞) also generate the Borel σ-algebra on R and because a function is
measurable if and only if the preimages of generating sets are measurable.
Proposition 4.51. Let Y be a topological space.
(a) If U ⊆ Y is open, then χU , the characteristic function on U, is lower semi-continuous.
(b) If K ⊆ Y is closed, then χK is upper semi-continuous.
(c) If Y is compact and Hausdorff and χ : Y → R≥0 is lower semi-continuous, then
χ(y) = sup
ϕ∈C(Y )(cid:8)ϕ(y) : 0 ≤ ϕ ≤ χ(cid:9)
∀ y ∈ Y.
(d) If Y is compact and Hausdorff and χ : Y → R≥0 is upper semi-continuous, then
χ(y) = inf
ϕ∈C(Y )(cid:8)ϕ(y) : 0 ≤ χ ≤ ϕ(cid:9)
∀ y ∈ Y.
(4.52)
(4.53)
Proof. See Proposition 7.11 in [14] (though the reader may also check that this follows from the
definitions and Urysohn's Lemma).
(cid:4)
Corollary 4.54. Let µ be a positive regular measure on a compact Hausdorff space Y and let
χ : Y → R≥0 be a lower semi-continuous function on Y. Then
ZY
χ dµ = sup
ϕ∈C(Y )(cid:26)ZY
ϕ dµ : 0 ≤ ϕ ≤ χ(cid:27) .
Proof. See Corollary 7.13 in [14].
(4.55)
(cid:4)
Remark 4.56. Corollary 4.54 follows from Lusin's theorem, which is a more quantitative version
of Urysohn's Lemma for compact Hausdorff spaces equipped with a (finite) regular measure [14].
All regular probability measures are regular positive measures, and therefore all of these results
apply to regular probability measures on a compact Hausdorff space.
43
4.3 Continuous stochastic maps
We are finally ready to define the category cHausStoch, the generalization of the category
FinSetStoch to compact Hausdorff spaces and (continuous) stochastic maps.
In the process
of doing so, we prove several facts of independent interest (see Lemma 4.64 and 4.76 for instance).
Definition 4.57. Let X and Y be two compact Hausdorff spaces. A (continuous) stochastic map
from X to Y is a continuous function µ : X → ProbMeas(Y ) with respect to the vague topology
Y. The value of µ at x ∈ X is denoted by µx and is a regular
and is denoted by µ : X
probability measure on Y.
Example 4.58. Let f : X → Y be a continuous function of compact Hausdorff spaces. Then
δf−→ ProbMeas(Y )
X
x 7→ δf (x),
where δf (x) is the measure defined by
MY ∋ E 7→ δf (x)(E) :=(1
0
if f (x) ∈ E
if f (x) /∈ E,
(4.59)
(4.60)
is a stochastic map from X to Y. δf (x) is called the Dirac delta measure on Y concentrated at
the point f (x). The fact that δf (x) is a regular probability measure follows immediately from the
definition. Before proving that this is a stochastic map, notice that for any continuous function
ϕ ∈ C(Y ) and for any x ∈ X,
ZY
To see that X ∋ x 7→ δf (x) is continuous, let x : Θ → X be a net converging to lim x ∈ X. Then,
for any continuous function ϕ ∈ C(Y ),
ϕ dδf (x) = ϕ(cid:0)f (x)(cid:1).
(4.61)
lim
θ∈ΘZY
ϕ dδf (xθ) = lim
θ∈Θ
ϕ(cid:0)f (xθ)(cid:1) = ϕ(cid:0)f (lim x)(cid:1) =ZY
ϕ dδf (lim x).
(4.62)
Since ϕ was arbitrary, this shows that lim
θ∈Θ
δf (xθ) = δf (lim x) in the vague topology.
In particular, the identity function from a space to itself induces a stochastic map known as
the identity stochastic map. Why it is called the identity will be justified after the definition of
composition of stochastic maps is provided.
The composition of stochastic maps is quite subtle for compact Hausdorff spaces due to the
possibility that the underlying spaces need not be metrizable. Before going into the details, we
isolate the issue that occurs by what we expect the composition to be heuristically based on our
understanding of the finite set case. In this regard, let X, Y, and Z be compact Hausdorff spaces
Z be two stochastic maps. The composition of µ followed by
and let µ : X
o Z which when applied to a Borel set E ⊆ Z and a point
ν should be some stochastic map X
x ∈ X should give the probability of x evolving through the intermediate possibilities in Y that
end up at the "window" E as sketched in Figure 3.
Y and ν : Y
44
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
Z
E
νy
νy ′
Y
y
•
•y′
X
•x
µx
Figure 3: x ∈ X evolves to the probability measure µx and all the points in Y that have
neighborhoods with non-vanishing measure themselves get sent to probability measures
on Z. To calculate the overlap with a measurable set E ⊆ Z, the measures are evaluated
on E and the results are integrated over all of Y. Only νy and νy′ for two points y, y′ ∈ Y
have been sketched on the left for ease of visualization.
In other words, we sum up νy(E) over all y using the measure µx on Y to obtain
ZY
νy(E) dµx(y).
(4.63)
However, this formula assumes that the function Y ∋ y 7→ νy(E) is measurable for all Borel sets
E ∈ MZ. The validity of this formula is well known if all the spaces in question are Polish [19]
but does not seem to be known for arbitrary compact Hausdorff spaces. We will prove this result
soon, but first we prove a crucial result that is interesting in its own right.
Lemma 4.64. Let Y be a compact Hausdorff space and let E ∈ MY be a Borel set. Then the
evaluation function
ProbMeas(Y )
evE−−→ R
µ 7→ µ(E)
(4.65)
is Borel measurable with respect to the vague topology on ProbMeas(Y ). In fact, evE is lower
semi-continuous when E is open and upper semi-continuous when E is closed.
To prove this lemma, we recall a useful theorem, which we learned through Sengupta's lucid
notes [43].
Theorem 4.66 (Dynkin's π-λ Theorem). Let X be a set and let P and L be collections of subsets
of X satisfying the following conditions.
i. P is closed under finite intersections (such a collection is called a π-system).
ii. L contains the empty set, is closed under taking complements (in X), and is closed under
countable unions of pairwise disjoint elements (such a collection is called a λ-system).
45
iii. P ⊆ L.
Then the σ-algebra generated by P is contained in L.
Proof. See Theorem 3.2 in [6] or the notes [43].
(cid:4)
We are indebted to Iddo Ben-Ari for the crucial insight that Dynkin's π-λ theorem can be used
to give a nice proof of Lemma 5.8, which lead to the proof of Lemma 4.64.
Proof of Lemma 4.64. The proof will be broken up into several steps. First, it will be shown that
the collection, P, of all open subsets of Y forms a π-system. Then it will be shown that the set
L :=nE ⊆ Y : ProbMeas(Y )
evE−−→ [0, 1] is Borel measurableo
(4.67)
is a λ-system. Finally, it will be shown that evU is lower semi-continuous (and hence Borel
measurable), for all open sets U ⊆ Y showing that P ⊆ L.
i. The fact that open sets are closed under finite intersections is part of the definition of a
topology.
ii. For each µ ∈ ProbMeas(Y ),
ev∅(µ) = µ(∅) = 0
so that ev∅ is constant and hence measurable. For E ∈ L,
evEc(µ) = µ(Ec) = µ(Y ) − µ(E) = (1 − evE)(µ)
(4.68)
(4.69)
is the difference of two measurable functions with respect to µ and is therefore measurable
with respect to µ. For {En} a countable collection of pairwise disjoint measurable sets in L
with E :=Sn En,
∞
evE(µ) = µ ∞
[n=1
En! =
µ(En) = ∞
Xn=1
Xn=1
evEn! (µ)
(4.70)
by additivity of the measure µ. Hence, since this is a bounded countable sum of measurable
functions, the sum is measurable.
iii. Let U ⊆ Y be open and non-empty (if U is empty, the result was already proved in part ii.).
By Proposition 4.51, χU is lower semi-continuous. Hence, for any a ∈ R, the preimage of
(a,∞) under evU is
by Corollary 4.54
(4.71)
ev−1
U (cid:0)(a,∞)(cid:1) =(cid:8)µ ∈ ProbMeas(Y ) : µ(U) > a(cid:9)
=(cid:8)µ ∈ ProbMeas(Y ) : ZY
=
= [ϕ∈C(Y )
χU dµ > a(cid:9)
ϕ dµ(cid:27) > a
(cid:26)ZY
ϕ dµ > a(cid:27) .
(cid:26)µ ∈ ProbMeas(Y ) : ZY
µ ∈ ProbMeas(Y ) :
ϕ∈C(Y )
0≤ϕ≤χU
sup
0≤ϕ≤χU
46
By definition of the vague topology, the set in curly brackets in the last equality is open.
Hence, this is a union of open sets and is therefore an open subset of ProbMeas(Y ). Therefore,
this shows that evU is lower semi-continuous for all open sets U in Y. An analogous argument
using infima shows that evK is upper semi-continuous for all closed sets K in Y.
Thus, since all the conditions of Dynkin's π-λ theorem are satisfied, the σ-algebra generated by P
which is the Borel σ-algebra, is contained in L.
(cid:4)
Remark 4.72. Although evU was shown to be lower semi-continuous for all open sets U, it is in
general not continuous. For an explicit example, enumerate the rationals in [0, 1] by some bijective
function q : N → Q ∩ [0, 1] and set
U := ∞
ǫ
2n+1 , qn +
(4.73)
[n=1(cid:16)qn −
ǫ
2n+1(cid:17)! ∩ [0, 1]
for some ǫ ∈ (0, 1). Then, U is open, contains all of the rationals in [0, 1], and has Lebesgue measure
ǫ < 1 so that it does not contain all of [0, 1]. Fix x∞ ∈ [0, 1] \ U. Since Q ∩ [0, 1] is dense in [0, 1],
there exists a sequence x : N → Q ∩ [0, 1] converging to x∞ in [0, 1]. Then, the corresponding
sequence of Dirac measures n 7→ δxn converges to δx∞ in the vague topology but
while
δxn(U) = 1
&
δxn(cid:0)[0, 1] \ U(cid:1) = 0
∀ n ∈ N
(4.74)
(4.75)
δx∞(U) = 0
&
δx∞(cid:0)[0, 1] \ U(cid:1) = 1
showing that evU is not continuous at δx∞. In fact, checking that evU is lower semi-continuous at
δx∞ directly follows from the definition since evU (µ) ≥ evU (δx∞) − ǫ reads µ(U) ≥ 0 − ǫ = −ǫ,
which is true for all ǫ > 0 and probability measures µ. However, one runs into trouble checking if
evU is upper semi-continuous for the reason illustrated above.
Lemma 4.76. Let µ : X
E ∈ MY , the function
Y be a stochastic map as in Definition 4.57. For every Borel set
X ∋ x 7→ µx(E)
(4.77)
is Borel measurable.
continuous.
In fact, if E is open (closed), then this function is lower (upper) semi-
Proof. This follows from Lemma 4.64 and the fact that this function is the composition
X
µ
−→ ProbMeas(Y )
evE−−→ R
(4.78)
of Borel measurable functions.
If E is open (closed), then this function is the composition of
a continuous function followed by a lower (upper) semi-continuous function, which is still lower
(upper) semi-continuous (this is immediate from the definitions).
(cid:4)
Note that this shows, in particular, that a stochastic map gives rise to a Markov kernel. The
converse is not true as an arbitrary Markov kernel takes no account of continuity. Hence, although
the composition of Markov kernels is a Markov kernel, the requirement of being a stochastic map
imposes additional restrictions.
47
/
/
/
o
/
o
Proposition 4.79. Let X, Y, and Z be compact Hausdorff spaces and let µ : X
ν : Y
o Z be two stochastic maps. For each x ∈ X, the assignment
MZ ∋ E 7→ (ν ◦ µ)x(E) :=ZY
νy(E) dµx(y),
is a regular probability measure on Z. Furthermore, the assignment
X → ProbMeas(Z)
x 7→ (ν ◦ µ)x,
Y and
(4.80)
(4.81)
is a stochastic map. This stochastic map is called the composition of µ followed by ν and is denoted
by ν ◦ µ.
Proof. The proof will be split into three parts to prove the following claims: i. (ν ◦ µ)x is a Borel
probability measure on Z, ii. it is regular, and iii. the assignment X ∋ x 7→ (ν ◦ µ)x is a stochastic
map X
o Z.
i. First,
(ν ◦ µ)x(Z) =ZY
dµx(y) =ZY
dµx(y) = µx(Y ) = 1.
(4.82)
Now, fix a Borel set E ∈ MZ. Since Y ∋ y 7→ νy(E) is Borel measurable (by Lemma 4.76)
and since νy(E) is non-negative for all Borel sets E in Z, its integral with respect to a positive
measure is non-negative. We now check that (ν ◦ µ)x is additive on (at most) countable
measurable partitions. Let {Ei}i∈I be a countable measurable partition of E. Then
νy(Z)
{z}1
(ν ◦ µ)x(E) =ZY
=ZY Xi∈I
=Xi∈I ZY
=Xi∈I
νy(E) dµx(y)
by definition of (ν ◦ µ)x
νy(Ei) dµx(y)
since νy is a measure
νy(Ei) dµx(y)
by the Monotone Convergence Theorem
(4.83)
(ν ◦ µ)x(Ei)
by definition of (ν ◦ µ)x,
In the third equality, νy(Ei) is non-negative and measurable for all y ∈ Y so the partial sums
form an increasing sequence of measurable functions. Hence, the Monotone Convergence
Theorem allows one to exchange the sum operation with the integral. Therefore, (ν ◦ µ)x is a
Borel probability measure.
ii. To prove regularity of the measure (ν ◦ µ)x, fix E ∈ MZ and ǫ > 0. The goal is to find an
open set U ⊆ Z and a compact set K ⊆ Z such that K ⊆ E ⊆ U and
(ν ◦ µ)x(U) − (ν ◦ µ)x(E) < ǫ
&
(ν ◦ µ)x(E) − (ν ◦ µ)x(K) < ǫ.
(4.84)
Note, that it suffices to show that (by a single application of the triangle inequality)
(ν ◦ µ)x(U) − (ν ◦ µ)x(K) < ǫ.
(4.85)
48
/
/
/
o
/
o
/
/
/
o
/
/
/
/
o
/
Fix y ∈ Y. Since νy is regular, there exist an open set Uy ⊆ Z and a closed set Ky ⊆ Z such
that Ky ⊆ E ⊆ Uy and
(4.86)
Since Y ∋ y′
7→ νy′(Ky) is upper semi-
continuous by Lemma 4.76, there exists an open neighborhood Vy ⊆ Y containing y such that
7→ νy′(Uy) is lower semi-continuous and Y ∋ y′
νy(Uy) − νy(Ky) <
ǫ
3
.
ǫ
3
νy(Uy) − νy′(Uy) ≤
(4.87)
The collection of such open and closed sets can be obtained for every y ∈ Y. Hence, {Vy}y∈Y
forms an open cover of Y. By compactness of Y, there exists a finite collection {y1, . . . , yn} of
points in Y for which {Vy1, . . . , Vyn} covers Y. Set
νy′(Ky) − νy(Ky) ≤
∀ y′ ∈ Vy.
&
ǫ
3
U :=
n
\i=1
n
[i=1
Uyi
&
K :=
Kyi.
(4.88)
Since these intersections and unions are finite, U is open and K is closed in Z. Furthermore,
they satisfy K ⊆ E ⊆ U. Now, fix y ∈ Y. There exists some i ∈ n such that y ∈ Vyi so that
(cid:12)(cid:12)νy(U) − νy(K)(cid:12)(cid:12) ≤(cid:12)(cid:12)νy(Uyi) − νy(Kyi)(cid:12)(cid:12) by (4.88) and sub-additivity of νy
=(cid:12)(cid:12)νy(Uyi) − νyi(Uyi) + νyi(Uyi) − νyi(Kyi) + νyi(Kyi) − νy(Kyi)(cid:12)(cid:12)
≤(cid:12)(cid:12)νy(Uyi) − νyi(Uyi)(cid:12)(cid:12) +(cid:12)(cid:12)νyi(Uyi) − νyi(Kyi)(cid:12)(cid:12) +(cid:12)(cid:12)νyi(Kyi) − νy(Kyi)(cid:12)(cid:12)
< ǫ by (4.86) and (4.87).
Since this inequality is true for every y ∈ Y,
(4.89)
(ν ◦ µ)x(U) − (ν ◦ µ)x(K) =ZY (cid:16)νy(U) − νy(K)(cid:17) dµx(y) <ZY
ǫ dµx = ǫ
(4.90)
iii. Now, we check that the function X ∋ x 7→ (ν ◦ µ)x is continuous so that ν ◦ µ : X
because µx is a probability measure. This proves that (ν ◦ µ)x is regular for arbitrary x ∈ X.
o Z is a
stochastic map. In this regard, fix an arbitrary ϕ ∈ C(Z) and a net x : Θ → X converging to
lim x. The goal is to show that
lim
θ∈ΘZZ
ϕ d(ν ◦ µ)xθ =ZZ
ϕ d(ν ◦ µ)lim x.
(4.91)
By the Jordan decomposition theorem, it suffices to assume that ϕ is positive. Since ϕ is
continuous, by Lemma 4.10 there exists a sequence of simple functions 0 ≤ s1 ≤ s2 ≤ · · · ≤ ϕ
of the form
sn,inχEn,in
(4.92)
defined on Borel sets {En,in} in Z satisfying all of the conclusions of that Lemma. Hence
θ∈ΘZZ
lim
ϕ d(ν ◦ µ)xθ = lim
lim
θ∈Θ
lim
n→∞Xin
sn,in(ν ◦ µ)xθ(En,in) by definition of the integral
sn,inZY
n→∞Xin
n→∞ZY Xin
νy(En,in) dµxθ by definition of ν ◦ µ
sn,inνy(En,in) dµxθ by linearity of the integral.
lim
= lim
θ∈Θ
= lim
θ∈Θ
(4.93)
sn =Xin
49
/
/
/
o
/
Now, to verify that the limit can be interchanged with the integral, note that by the properties
of the simple functions constructed above and since νy is a measure,
Xin
sn,inνy(En,in) =Xin
=Xin+1
≤Xin+1
=Xin+1
even
even
s
n,
in+1
sn,inνy(En+1,2in) +Xin
νy(cid:0)En+1,in+1(cid:1) +Xin+1
sn+1,in+1νy(cid:0)En+1,in+1(cid:1) +Xin+1
odd
odd
2
sn+1,in+1νy(En+1,in+1)
sn,inνy(En+1,2in+1) by (4.13)
s
n,
2
in+1−1
νy(cid:0)En+1,in+1(cid:1)
sn,in+1−1νy(cid:0)En+1,in+1(cid:1)
(4.94)
by (4.14)
for all y ∈ Y. Therefore, the sequence of functions
N × Y ∋ (n, y) 7→Xin
sn,inνy(En,in)
(4.95)
on Y is a monotonic increasing sequence of Borel measurable functions. Hence, continuing on
from equation (4.93),
lim
θ∈ΘZZ
lim
n→∞Xin
θ∈ΘZY
ϕ d(ν ◦ µ)xθ = lim
θ∈ΘZY (cid:18)ZZ
= lim
by the Monotone Convergence Theorem
(4.96)
ϕ dνy(cid:19) dµxθ by definition of the integral.
sn,inνy(En,in) dµxθ
To continue with this calculation, note that by assumption, the assignment y 7→ νy is continu-
ous with respect to the vague topology on ProbMeas(Z). This means that for any continuous
function such as ϕ ∈ C(Z) above, y 7→ RZ ϕ dνy is a continuous real-valued function on Y.
Hence, following from (4.96),
lim
θ∈ΘZZ
ϕ d(ν ◦ µ)xθ =ZY (cid:18)ZZ
ϕ dνy(cid:19) dµlim x
(4.97)
since µ is continuous with respect to the vague topology on ProbMeas(Y ). By applying an
analogous procedure (except without taking any limit in Θ), one similarly concludes that
ZZ
ϕ d(ν ◦ µ)lim x =ZY (cid:18)ZZ
ϕ dνy(cid:19) dµlim x.
Since these two results are equal, continuity of ν ◦ µ has been exhibited.
Therefore, ν ◦ µ : X
o Z is a stochastic map.
(4.98)
(cid:4)
From this fact, it immediately follows that the identity stochastic map is indeed an identity
for the composition of stochastic maps.
50
/
/
/
o
/
Proposition 4.99. The composition of stochastic maps is associative.
Proof. Let µ : X
Y, ν : Y
Z, and ζ : Z
o W be stochastic maps of compact Hausdorff
spaces. The goal is to show that (cid:0)ζ ◦ (ν ◦ µ)(cid:1)x(E) =(cid:0)(ζ ◦ ν) ◦ µ(cid:1)x(E) for all Borel sets E ∈ MW
and points x ∈ X. By expanding out the definitions, it must be shown that
ζz(E) dνy(z)(cid:19) dµx(y).
ζz(E) d(ν ◦ µ)x(z) =ZY (cid:18)ZZ
ZZ
(4.100)
Contrary to what one might naively think, Fubini's Theorem does not apply here. Instead, since
Z ∋ z 7→ ζz(E) is Borel measurable and bounded, by Lemma 4.10, there exists a sequence
s : N → RZ of simple functions satisfying
0 ≤ s1(z) ≤ s2(z) ≤ s3(z) ≤ · · · ≤ ζz(E)
&
lim
n→∞
sn(z) = ζz(E)
∀ z ∈ Z.
(4.101)
and all the other properties of that Lemma. Hence, by a completely analogous calculation as in
(4.93)–(4.96) but without a net, the equality in (4.100) is obtained. This proves that composition
of stochastic maps is associative.
(cid:4)
In conclusion, we have proved the first part of the following theorem.
Theorem 4.102. The collection of compact Hausdorff spaces together with morphisms as stochas-
tic maps forms a category, denoted by cHausStoch. In addition, the Dirac measure furnishes a
faithful (but not full) functor δ : cHaus → cHausStoch sending a space X to X and a continuous
function f : X → Y to δf : X
Proof. The faithfulness of δ follows immediately from the Hausdorff condition. More precisely,
seeking to prove the contrapositive, suppose that f, g : X → Y were two different functions. Then
there exists an x ∈ X such that f (x) 6= g(x). Since Y is Hausdorff, there exist open sets U and
V in Y such that x ∈ U, y ∈ V, and U ∩ V = ∅. Then δf (x)(U) = 1 while δg(x)(U) = 0. Hence
δf 6= δg.
Y as in Example 4.58.
(cid:4)
4.4 From continuous probability theory to algebra
We have already associated to every compact Hausdorff space a commutative C ∗-algebra by taking
continuous functions on that space. To every stochastic map, we can also associate a canonical
positive map of C ∗-algebras.
Proposition 4.103. Let µ : X
every continuous function ϕ : Y → C, the function C µ(ϕ) : X → C defined by
Y be a stochastic map between compact Hausdorff spaces. For
X ∋ x 7→ZY
ϕ dµx
(4.104)
is a continuous function on X. Furthermore, C µ : C(Y ) → C(X) is a (unital) positive map.
51
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
o
Proof. Linearity and positivity of C µ follow immediately from the definitions and properties of a
probability measure. To prove that C µ(ϕ) is a continuous function for any continuous function
ϕ ∈ C(Y ), let x : Θ → X be a net converging to lim x ∈ X. By definition of the vague topology
on ProbMeas(Y ) and because µ : X → ProbMeas(Y ) is continuous,
ϕ dµxθ =ZY
ϕ dµlim x = C µ(ϕ)(lim x)
(4.105)
(cid:4)
(4.107)
(4.108)
lim
θ∈ΘZY
θ∈Θ(cid:0)C µ(ϕ)(xθ)(cid:1) = lim
showing that C µ(ϕ) is continuous.
Theorem 4.106. The assignment
cHausStochop C−→ cC*-AlgPos
X 7→ C(X)
Y ) 7→(cid:0)C µ : C(Y )
(µ : X
o C(X)(cid:1)
is a functor. Moreover, the diagram
cHausop
δ
cHausStochop
C
C
cC*-Alg
cC*-AlgPos
commutes (the functor on the right is the inclusion since every ∗-homomorphism is positive).
Proof. The fact that the codomain of C is as claimed follows from Proposition 4.103.
If idX : X
o X is the identity stochastic map, then
(cid:0)C idX (ϕ)(cid:1)(x) =ZX
ϕ dδx = ϕ(x)
(4.109)
for all x ∈ X and for all ϕ ∈ C(X). Hence C idX (ϕ) = ϕ for all ϕ ∈ C(X). Thus C preserves
identities.
To see that C preserves the composition, let µ : X
Z be two stochastic
maps. The goal is to show that C ν◦µ = C µ ◦ C ν. The proof will be similar to the proof that
composition in cHausStoch is associative. To do this, fix ϕ ∈ C(Z). By the Jordan decomposition
theorem, it suffices to assume that ϕ(Z) ⊆ R≥0 since any measurable complex-valued function can
be decomposed into its real and imaginary parts and each of those into positive and negative parts.
Since ϕ is continuous and Z is compact, let s be a sequence of simple functions as in Lemma 4.10.
Y and ν : Y
52
/
/
/
o
/
o
/
/
/
o
/
/
/
/
/
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
o
= lim
= lim
sn,inνy(En,in) dµx(y) by (finite) linearity of the integral
νy(En,in) dµx(y) by definition of (ν ◦ µ)x
ϕ d(ν ◦ µ)x by definition of C ν◦µ
n→∞Xin
sn,in(ν ◦ µ)x(En,in) by definition of the Lebesgue integral
sn,inZY
n→∞Xin
n→∞ZY Xin
=ZY
n→∞Xin
=ZY (cid:18)ZZ
=ZY (cid:0)C ν(ϕ)(cid:1)(y) dµx(y) by definition of C ν
=(cid:16)C µ(cid:0)C ν(ϕ)(cid:1)(cid:17)(x) by definition of C µ
ϕ dνy(cid:19) dµx(y) by definition of the Lebesgue integral
lim
sn,inνy(En,in) dµx(y) by Monotone Convergence Theorem
(4.110)
Then,
(cid:0)C ν◦µ(ϕ)(cid:1)(x) =ZZ
= lim
for all x ∈ X. The Monotone Convergence Theorem can be applied in this calculation due to
Lemma 4.10 and the way the simple functions were constructed in that Lemma (recall step iii. in
the proof of Proposition 4.79). Since ϕ was arbitrary, this shows that C ν◦µ = C µ ◦ C ν.
Finally, if f : X → Y is a continuous function, let ϕ ∈ C(Y ) and x ∈ X. Then
(cid:0)C f (ϕ)(cid:1)(x) =ZY
ϕ dδf (x) = ϕ(cid:0)f (x)(cid:1)
showing that the diagram in the statement of the theorem commutes.
(4.111)
(cid:4)
4.5 The stochastic Gelfand spectrum functor
In this section, we prove that the functor C : cHausStochop → cC*-AlgPos constructed in
Section 4.4 is fully faithful. To do this, we review the Riesz-Markov-Kakutani Representation
Theorem. Then we use these results to construct an analogue of the spectrum functor for positive
maps on arbitrary commutative C ∗-algebras.
Theorem 4.112 (Riesz-Markov-Kakutani (RMK) Representation Theorem). Let Y be a compact
o C be a state on C(Y ). Then, there exists a unique regular
Hausdorff space and let I : C(Y )
probability measure µ ∈ ProbMeas(Y ) such that
I(ϕ) =ZY
ϕ dµ
∀ ϕ ∈ C(Y ).
Furthermore, this measure µ satisfies
µ(U) = sup(cid:8)I(ϕ) : ϕ ∈ C(Y ) and 0 ≤ ϕ ≤ χU(cid:9)
53
(4.113)
(4.114)
/
/
/
o
/
for all open sets U ⊆ Y and
for all closed (and hence compact) sets K ⊆ Y.
µ(K) = inf(cid:8)I(ϕ) : ϕ ∈ C(Y ) and ϕ ≥ χK(cid:9)
(4.115)
Note that by our definition, a state is a positive linear functional that is unit-preserving. This
is why the measure µ in the statement of the theorem is a probability measure. We will not prove
this theorem in its entirety, but an explicit construction of the measure is useful and provides a
close comparison to the finite-dimensional case. Namely, when X is finite, C(X) = CX and it
makes sense to assign µ(E) := I(χE) for any finite subset E ⊆ X as we have done in Section 2.
This is because the characteristic function χE is continuous on a finite set. Hence, heuristically,
given a measurable set E ∈ MY , one would expect µ(E) to be given by I(χE). However, χE is not
continuous and so this expression simply does not make sense because χE is not in the domain of
I. Nevertheless, we know by regularity that E can be approximated using open and closed sets.
One can use this fact to first extend I to be a positive linear functional on the larger space of lower
semi-continuous functions as is done in [48].
Sketch of proof. First, let U ⊆ Y be open and let K ⊆ Y be closed. Set
µ(U) := sup
0≤ϕ≤χU(cid:8)I(ϕ)(cid:9)
ϕ∈C(Y )
µ(K) := inf
0≤χK ≤ϕ(cid:8)I(ϕ)(cid:9).
ϕ∈C(Y )
For an arbitrary subset E of Y, set
µ+(E) = inf
U open
E⊆U (cid:8)µ(U)(cid:9)
It is a fact that for all Borel subsets E ∈ MY ,
µ+(E) = µ−(E)
µ−(E) = sup
K⊆E (cid:8)µ(K)(cid:9).
K closed
and we therefore denote this assignment by µ. It is also a fact that µ is the unique regular
probability measure on Y satisfying the condition in the statement of the theorem. See Theorem
7.2 in [14], Theorem 8 in [48], or Theorem 2.14 in [38] for proofs of these claims.
(cid:4)
Remark 4.119. In general, the assignments µ± on arbitrary subsets E of Y given by
(4.116)
(4.117)
(4.118)
(4.120)
(4.121)
and
µ+(E) := inf
U open
E⊆U
µ−(E) := sup
K closed
K⊆E
sup
ϕ∈C(Y )
0≤ϕ≤χU(cid:8)I(ϕ)(cid:9)
0≤χK ≤ϕ(cid:8)I(ϕ)(cid:9)
ϕ∈C(Y )
inf
are not the same. These are known as outer and inner measures associated to I. The collection of
subsets E for which µ+(E) = µ−(E) contains the Borel sets of Y. In general, however, the resulting
σ-algebra is larger than the Borel σ-algebra (see Theorem 2.14 in [38]). Nevertheless, µ restricted
to the Borel σ-algebra associated to Y provides Y with a regular probability measure. This is the
measure constructed in the proof of Theorem 4.112.
&
&
54
Proposition 4.122. cHausStochop C−→ cC*-AlgPos is fully faithful. More explicitly, let X and
o C(X) be a positive (unital) map. Then there
Y be compact Hausdorff spaces and let f : C(Y )
exists a unique stochastic map µ : X
Y such that
(cid:0)f (ϕ)(cid:1)(x) =ZY
ϕ dµx
∀ x ∈ X
∀ ϕ ∈ C(Y ).
Proof. First note that for every ϕ ∈ C(Y ) and x ∈ X,
(4.123)
(4.124)
(cid:0)f (ϕ)(cid:1)(x) = (evx ◦ f )(ϕ),
where evx : C(X) → C is the evaluation at x function (recall Example 2.48), which is a ∗-
o C(X) evx−−→ C is
homomorphism and therefore a positive map. Hence, the composition C(Y )
a state on C(Y ). By the RMK Representation Theorem, there exists a unique regular probability
measure µx ∈ ProbMeas(Y ) satisfying (4.123). The only thing left to check is that the assignment
X ∋ x 7→ µx ∈ ProbMeas(Y ) is continuous. However, this immediately follows from the fact that
ϕ(f ) is continuous since if x : Θ → X is a net converging to lim x, then
f
lim
θ∈ΘZY
Hence, C is fully faithful.
ϕ dµxθ = lim
θ∈Θ(cid:0)f (ϕ)(cid:1)(xθ) =(cid:0)f (ϕ)(cid:1)(lim x) =ZY
ϕ dµlim x.
(4.125)
(cid:4)
Thus, the RMK Representation Theorem precisely encodes the fact that the functor C :
cHausStochop → cC*-AlgPos is fully faithful while the Gelfand-Naimark Theorem shows that
this functor is essentially surjective. Together, these results hold if and only if C is part of adjoint
equivalence of categories. However, this result is somewhat formal. It is useful to have an ex-
plicit functor σ : cC*-AlgPos → cHausStochop that (1) requires no reference to any underlying
space structure and (2) agrees with the usual spectrum functor from the ordinary (commutative)
Gelfand-Naimark correspondence when restricted to ∗-homomorphisms of C ∗-algebras.
We have already constructed the functor σ at the level of objects, namely, given any com-
mutative C ∗-algebra A, take σ(A) to be the spectrum of A (see Definition 3.38). Given a ∗-
homomorphism of commutative C ∗-algebras f : B → A, the associated map σ(f ) assigns to each
character χ ∈ σ(A) the linear functional χ ◦ f : B → C. This linear functional is a character
precisely because f and χ are both (non-vanishing) ∗-homomorphisms. Hence, when f is simply
a positive map, χ ◦ f is no longer a character in general. Nevertheless, from these data, namely a
o A and a character χ ∈ σ(A), we must construct a canonical regular prob-
positive map f : B
χ on σ(B) in such a way so that this assignment σf : σ(A) → ProbMeas(cid:0)σ(B)(cid:1)
ability measure σf
is a stochastic map. The idea should be that we smear χ ◦ f onto σ(B) in some way that captures
the information of the state χ ◦ f as depicted in Figure 4.
This can be done formally using the Gelfand transform and the fact that C is fully faithful.
Namely, we can fill in the diagram
B
Γ−1
B
C(cid:0)σ(B)(cid:1)
f
C σ(f )
55
ΓA
A
C(cid:0)σ(A)(cid:1)
(4.126)
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
O
O
/
/
/
o
/
o
/
o
/
o
• χ ◦ f
σf
χ
Figure 4: A character provides a Dirac point measure on the spectrum. A state is
"smeared" onto the spectrum in terms of a probability measure representing the asso-
ciated "weight" of that state on the characters in the spectrum.
on the bottom by using the inverse of the Gelfand transform, the positive map f, and then the
o C(cid:0)σ(A)(cid:1) which we
Gelfand transform once more. This defines a positive map C σ(f ) : C(cid:0)σ(B)(cid:1)
know is uniquely characterized by a stochastic map σf : σ(A)
σ(B) by the RMK theorem. We
could simply define σ(f ) in this manner and move on. However, a more geometrically appealing
formula for σ(f ) is available in terms of the convex structure on the set of states with reference
to neither the Gelfand transform nor the RMK theorem. As the reader may have guessed from
the suggestive drawing above, this merits a review of convex sets, their extreme points, the Krein-
Milman theorem, and Choquet theory [44], [35].
Definition 4.127. Let V be a topological vector space. A subset C ⊆ V is said to be convex iff
(4.128)
λv + (1 − λ)w ∈ C
∀ λ ∈ [0, 1] whenever v, w ∈ C.
An extreme point of C is a point u ∈ C such that if u = λv + (1 − λ)w for some v, w ∈ C and
λ ∈ [0, 1], then v = u and/or w = u. The set of extreme points of C is denoted by ex(C).
Example 4.129. Let A be a commutative C ∗-algebra. Then, S(A), the set of all states, is a
compact convex subset of A∗ in the weak* topology. σ(A), the spectrum of A, is the set of
extreme points of S(A). This follows, for instance, from Theorem 11.33 in Rudin [39].
Definition 4.130. Let V be a topological vector space and let X ⊆ V. The convex hull of X is
given by
i.e. the set of all finite convex combinations of elements of X.
The following is an application of the Krein-Milman theorem suitable for our purposes [44].
Theorem 4.132. Let A be a commutative C ∗-algebra. Then the set of all states on A is the
closure of the convex hull of the set of characters on A, i.e.
S(A) = ch(cid:0)σ(A)(cid:1).
Proof. This follows from the usual Krein-Milman theorem, the conclusions of Example 4.129,
and the fact that A∗ is a locally convex topological vector space (it is locally convex due to the
separating family of seminorms generating its topology).
(4.133)
(cid:4)
56
ch(X) :=(finite
Xi
λivi
: vi ∈ X, λi ∈ [0, 1], Xi
λi = 1) ,
(4.131)
/
/
/
o
/
/
/
/
o
/
o
This important result implies the existence of a probability measure associated to any state.
Theorem 4.134. Let A be a commutative C ∗-algebra and let ω ∈ S(A) be a state on A. Then
there exists a unique regular probability measure µ on σ(A) such that
Zσ(A)
eva dµ = ω(a)
∀ a ∈ A.
(4.135)
Here eva : σ(A) → C denotes the evaluation at the point a sending a character χ ∈ σ(A) to χ(a).
The statement of the theorem can be obtained from the RMK Representation Theorem com-
bined with the Gelfand-Naimark Theorem. However, we prefer to prove this theorem directly
using Choquet theory, which provides a concrete and geometric construction of the measure. We
thank Benjamin Russo for discussions leading to this construction.
Proof. We first establish existence and then uniqueness.
By the Krein-Milman Theorem, ω ∈ ch(cid:0)σ(A)(cid:1). Hence, there exists a net ξ : Θ → ch(cid:0)σ(A)(cid:1)
such that lim ξ = ω in the weak* topology on S(A). Since each ξθ is in the convex hull of σ(A),
there exists an Nθ ∈ N and a convex decomposition of ξθ into characters, i.e.
ξθ =
NθXiθ=1
λiθχiθ ,
where χiθ ∈ σ(A) for all iθ ∈ {1, . . . , Nθ}. Set
µθ :=
NθXiθ=1
λiθδχiθ
(4.136)
(4.137)
to be the convex combination of Dirac measures on these characters. This defines a net of regular
exists a subnet (µθ′)θ′∈Θ′⊆Θ of these measures converging in the vague topology to some regular
probability measure, which we denote by µ. This probability measure satisfies (4.135) since
probability measures on σ(A). Since ProbMeas(cid:0)σ(A)(cid:1) is compact in the vague topology, there
Zσ(A)
by continuity of eva and vague convergence
eva dµ = lim
eva dµθ′
λiθ′ Zσ(A)
eva dδχi
θ′
by linearity of the measure and integral
λiθ′ χiθ′ (a)
by definition of the Dirac measure
(4.138)
ξθ′(a)
by definition of ξ
= lim
θ′∈Θ′
Nθ′
θ′∈Θ′ Zσ(A)
Xiθ′ =1
Xiθ′ =1
Xiθ′ =1
Nθ′
Nθ′
= lim
θ′∈Θ′
= lim
θ′∈Θ′
= ω(a)
by weak* convergence.
57
This proves existence of such a measure.
Uniqueness is more subtle and follows, for instance, from the Choquet-Meyer Theorem. More
specifically, it follows from Corollary 10.9 in [35], whose notation we follow. We do not introduce
all the necessary definitions to explain this in full detail, but we outline the idea using the notation
and terminology from [35]. A∗, the topological dual space of A with the weak* topology, is a
locally convex vector space (this follows, for instance, from Proposition 3.29 above and Theorem
1.37 in [39]). The set of all states S(A) on a commutative (unital) C ∗-algebra is a compact convex
subset of A∗ and is a Choquet simplex (see Theorem 3.1.18 in [40] and note that states on a
commutative C ∗-algebra are automatically tracial). Because ex(cid:0)S(A)(cid:1) = σ(A) is a closed set
(since it is compact), every maximal measure is supported on σ(A) (see the comments preceding
the Bishop-de Leeuw Theorem in Section 4 of [35]). To see that
ZS(A)
ϕ dµ = ZS(A)
ϕ dδω
(4.139)
(µ ∼ δω in the notation of [35]) for all continuous real-valued affine functions ϕ on S(A), notice
that all functions of the form eva : A∗ → C are continuous and linear and hence remain continuous
when restricted to S(A) ⊆ A∗. By construction of µ from the subnet (µθ′) and the definition of
the Dirac measure,
ZS(A)
eva dµ = Zσ(A)
eva dµ = ω(a) = ZS(A)
eva dδω
∀ a ∈ A.
(4.140)
Now, suppose that ϕ is an arbitrary continuous real-valued affine function on S(A). Then, ϕ
restricted to σ(A) is a continuous function and therefore equals eva for some a ∈ A by Proposition
3.57. If we show that ϕ = eva on all of S(A), then we will be done by using the previous calculation.
Note that because S(A) is a compact convex subset of a locally convex space, the supremum and
infimum of ϕ are achieved on a face of S(A). This is a consequence, for instance, of Proposition
8.26 in [44] by noting that the image of a non-empty compact convex set under such a continuous
map is an interval in R and the supremum and infimum of an interval are at most two points.
These endpoints are faces of the interval and so their pre-images are faces of the simplex S(A).
Furthermore, the extreme points of a face are extreme points of the original convex subset (a
simple calculation from the definitions shows this). Hence, ϕ achieves its supremum on σ(A)
(and possibly on a larger domain as well). Hence, by the Hahn-Banach Theorem (see Section 5.22
of [38]), there exists a unique extension of ϕσ(A) = eva to S(A). But we already know that eva is
such an extension. Hence ϕ = eva. Since the result was already proved for such functions, µ ∼ δω
so uniqueness of the regular probability measure follows from the Choquet-Meyer theorem.
(cid:4)
Remark 4.141. Uniqueness also follows from a theorem due to Bauer since S(A) is a Bauer
simplex for A a commutative C ∗-algebra (see Theorem II.4.1 in [1]).
Remark 4.142. Note that because the evaluation map evE : ProbMeas(cid:0)σ(A)(cid:1) → R is not
necessarily continuous for every Borel subset E ⊆ σ(A), in general
λiθ(E),
(4.143)
µ(E) 6= lim
θ∈ΘXiθ∈Iθ
58
where Iθ is the subset of {1, . . . , Nθ} for which χiθ ∈ E. A counterexample is given by the sequence
N ∋ n 7→ µn :=
1
n
n
Xk=1
δ k
n
(4.144)
of a uniform sum of Dirac measures on [0, 1]. This sequence converges to the standard Lebesgue
measure µ on [0, 1] in the vague topology. However, µn(cid:0)Q ∩ [0, 1](cid:1) = 1 while µ(cid:0)Q ∩ [0, 1](cid:1) = 0.
In the finite-dimensional setting, this does not occur, nets are not required, and ω is uniquely
expressed as a convex decomposition of extremal states. The measure of E is the sum of the
weights corresponding to restricting this sum to the contributions coming only from E. This is
because the definition of a Choquet simplex is one that agrees with the definition of a simplex in
finite dimensional vector spaces (see Proposition 10.10 in [35]).
Theorem 4.145. Let B and A be commutative C ∗-algebras and let f : B
o A be a positive
(unital) map. For each character χ ∈ σ(A), there exists a unique regular probability measure
χ ∈ ProbMeas(cid:0)σ(B)(cid:1) such that
σf
Zσ(B)
∀ b ∈ B.
evb dσf
(4.146)
χ = χ(cid:0)f (b)(cid:1)
Furthermore, the assignment
σ(A)
σ(f )≡σf
−−−−−→ ProbMeas(cid:0)σ(B)(cid:1)
χ 7−−−−−→ σf
χ,
(4.147)
χ is the unique regular probability measure on σ(B) associated to the state χ ◦ f ∈ S(B)
where σf
via Theorem 4.134, defines a stochastic map σ(A)
Proof. Since χ ◦ f is a state on A, Theorem 4.134 guarantees there exists a unique regular proba-
bility measure σf
χ satisfying (4.146). The only thing left to check is that this defines a stochastic
map, i.e. that the function σf : σ(A) → ProbMeas(cid:0)σ(B)(cid:1) is continuous with respect to the vague
topology. Let χ : Θ → σ(A) be a net of characters converging to lim χ. The goal is to show that
σ(B).
lim
θ∈Θ Zσ(B)
ϕ dσf
χθ = Zσ(B)
ϕ dσf
lim χ
∀ ϕ ∈ C(cid:0)σ(B)(cid:1).
(4.148)
The Gelfand transform ΓB : B → C(cid:0)σ(B)(cid:1) sending b ∈ B to the evaluation function evb is a
C ∗-algebra isomorphism by Proposition 3.57. Hence, every continuous function on the spectrum
of B is of this form. This fact simplifies the above goal to simply showing that
evb dσf
lim χ
∀ b ∈ B.
(4.149)
lim
θ∈Θ Zσ(B)
evb dσf
χθ = Zσ(B)
59
/
/
/
o
/
/
/
/
o
/
o
This follows from the previous results since
lim
θ∈Θ Zσ(B)
evb dσf
χθ = lim
θ∈Θ
χθ(cid:0)f (b)(cid:1) by (4.146)
= (lim χ)(cid:0)f (b)(cid:1)
= Zσ(B)
evb dσf
lim χ by (4.146)
by weak* convergence of the net χ
(4.150)
for all b ∈ B. Therefore σf is a stochastic map.
Theorem 4.151. The assignment
(cid:4)
(4.152)
cC*-AlgPosop σ−→ cHausStoch
f
A 7→ σ(A)
o A(cid:17) 7→(cid:16)σ(A)
(cid:16)B
σ(f )≡σf
σ(B)(cid:17)
from Definition 3.38 and Theorem 4.145 defines a functor, henceforth referred to as the stochastic
spectrum functor.
Proof. The identity positive map gets sent to the identity stochastic map by uniqueness of the
measure and the fact that the Dirac measure is a regular measure satisfying the required conditions.
Let C,B, and A be commutative C ∗-algebras and let f : B
B be positive
maps. Let σf : σ(A) → ProbMeas(cid:0)σ(B)(cid:1), σg : σ(B) → ProbMeas(cid:0)σ(C)(cid:1), and σf ◦g : σ(A) →
ProbMeas(cid:0)σ(C)(cid:1) denote the images of f, g, and f ◦ g, under σ, respectively. Let σg ◦ σf : σ(A) →
ProbMeas(cid:0)σ(C)(cid:1) denote the composition σ(A)
σ(C) of the stochastic map σf followed
σ(B)
by σg. The goal is to show σf ◦g = σg◦σf . Since σf ◦g
and (σg◦σf )χ are regular probability measures
for each χ ∈ σ(A), by uniqueness in Theorem 4.145, it suffices to show
o A and g : C
χ
Zσ(C)
evc d(σg ◦ σf )χ = χ(cid:16)f(cid:0)g(c)(cid:1)(cid:17)
(4.153)
for all c ∈ C and for all χ ∈ σ(A). Therefore, first let c ∈ C be positive so that evc ≥ 0, and fix
χ ∈ σ(A), and let
(4.154)
be a sequence s : N → Rσ(C) of simple functions converging to evc and satisfying all the conditions
0 ≤ s1 ≤ s2 ≤ · · · ≤ evc
60
/
/
/
o
/
/
/
/
o
/
o
/
o
/
o
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
o
as in Lemma 4.10. Then,
χ
evg(c)(ψ) dσf
evc dσg
dσf
ψ
χ(ψ) by definition of evg(c)
χ(ψ) by definition of σf
χ(ψ) by definition of σg
χ(cid:16)f(cid:0)g(c)(cid:1)(cid:17) = Zσ(B)
= Zσ(B)
ψ(cid:0)g(c)(cid:1) dσf
= Zσ(B)
Zσ(C)
= Zσ(B)
n→∞Xin
sn,in Zσ(B)
n→∞Xin
n→∞Xin
sn,in(σg ◦ σf )χ(En,in) by definition of (σg ◦ σf )χ
= Zσ(C)
n→∞Xin
= Zσ(C)
evc d(σg ◦ σf )χ by definition of s.
σg
ψ(En,in) dσf
sn,inσg
ψ(En,in) dσf
lim
= lim
= lim
ψ
lim
sn,inχEn,in d(σg ◦ σf )χ by definition of Lebesgue integral
χ(ψ) by definition of Lebesgue integral
(4.155)
χ(ψ) by Monotone Convergence Theorem
Again, the Monotone Convergence Theorem applies by similar arguments as in step iii.
in the
proof of Proposition 4.79. The case of more general c ∈ C for which evc need not be positive is
handled by splitting up evc using a Jordan decomposition and applying a similar procedure. (cid:4)
Example 4.156. Let ρ : [−1, 1] → R be the restriction of the Gaussian used to describe the
distribution of heat after some time with initial condition a source of heat at the origin, namely
[−1, 1] ∋ x 7→ ρ(x) :=
,
(4.157)
exp (−100x2)
R 1
−1 exp (−100y2) dy
which has been properly normalized so that the integral of ρ with respect to the Lebesgue measure
is 1. Consider the sequence of probability measures given by
N ∋ n 7→ µn :=
n
Xk=1
−1+ 2k
nZ−1+ 2(k−1)
n
ρ(x) dx
δ−1+ 2k−1
n
.
(4.158)
The distribution and resulting convex sum approximation is depicted in Figure 5 for several values
of n ∈ N. Although the sequence of measures dµn does not converge to ρ dµLebesgue in the total
variation norm topology, the convergence holds in the vague topology, i.e.
lim
n→∞Z 1
−1
ϕ dµn =Z 1
−1
ϕ(x)ρ(x) dx
61
∀ ϕ ∈ C(cid:0)[−1, 1](cid:1).
(4.159)
•
n = 1
n = 10
• •
• • • •
• • • •
n = 31
•••
••••••••••••••
••••••••••••••
n = 50
n = 71
n = 100
•
••••••••••••••••••••••
•
••••••••••••••••••••••
••
•
•
•••••
•
•
••••••••••••••••••••••••••••••••
••••••••••••••••••••••••••••••••
••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••
Figure 5: The curve drawn in red is the graph of ρ in (4.157) (cut off from above to
fit on the page) while the (larger) blue bullets represent the heights of the Dirac point
distributions. The heights for the point measures decrease because the area over which
they are used to approximate functions decreases as n increases.
This gives an example of a probability measure on [−1, 1] that can be approximated by a sequence of
convex sums of Dirac point measures with respect to the vague topology. Again, it is not true that
ρ(x) dx for all Borel measurable sets E. For example, take E := [−1, 1]∩ (R\ Q),
lim
n→∞
all irrationals in [−1, 1]. The term on the left is 0 while the term on the right is 1.
µn(E) =ZE
4.6 The stochastic commutative Gelfand-Naimark Theorem
Notice that when A = C(X) with X a compact Hausdorff space and ω ∈ S(cid:0)C(X)(cid:1), Theorem
4.134 says that there exists a unique regular probability measure µ on σ(cid:0)C(X)(cid:1) such that
evϕ dµ = ω(ϕ)
∀ ϕ ∈ C(X).
(4.160)
Meanwhile, by the RMK Representation Theorem, there exists a unique regular probability mea-
sure ν on X such that
ϕ dν
∀ ϕ ∈ C(X).
(4.161)
By the Gelfand-Naimark Theorem, we know that the spaces X and σ(cid:0)C(X)(cid:1) are homeomorphic
and one expects these two measures to be related. In fact, the function hX : X → σ(cid:0)C(X)(cid:1) pushes
X : σ(cid:0)C(X)(cid:1) → X pushes forward the measure µ to
forward the measure ν to µ and the inverse h−1
ν. This is explained in more detail in the following theorem.
Z
σ(cid:0)C(X)(cid:1)
ω(ϕ) =ZX
62
Theorem 4.162. The stochastic spectrum functor together with the natural isomorphisms
cHausStoch
cHausStoch
id
h
a❉❉❉❉❉❉❉❉❉
σ
}③③③③③③③③③
C
cC*-AlgPosop
&
cC*-AlgPos
cC*-AlgPos
id
Γ
a❉❉❉❉❉❉❉❉❉
}③③③③③③③③③
σ
C
cHausStochop
(4.163)
from the commutative Gelfand-Naimark Theorem form an adjoint equivalence of categories.
Proof. The only thing that has not been shown is that the transformations Γ and h still satisfy
naturality. Note that each is an isomorphism in its respective category on objects due to the usual
Gelfand-Naimark Theorem.
i. Proof that h is natural. Let µ : X
denote the stochastic map obtained from applying the functor C and then σ to µ. The goal is
to prove that the diagram
Y be a stochastic map and let σC(µ) : σ(cid:0)C(X)(cid:1)
σ(cid:0)C(Y )(cid:1)
X
µ
Y
hX
σ(C(X))
σC(µ)
σ(C(Y ))
hY
(4.164)
commutes. Note that for any Borel set E ∈ Mσ(C(Y )) and for any x ∈ X, the lower left
composition δhY ◦ µ applied to x evaluated at E gives
χh−1
(4.165)
(δhY ◦ µ)x(E) =ZY
δhY (y)(E) dµx(y) =ZY
Y (E)(y) dµx(y) = µx(cid:0)h−1
Y (E)(cid:1).
In other words, (δhY ◦ µ)x is the pushforward of the measure µx along the function hY . Now
we check that this equals (cid:0)σC(µ) ◦ hX(cid:1)x for every x ∈ X by integrating evE with respect to
this pushforward measure and ϕ ∈ C(Y ). By the change of variables formula for integrating
measurable functions with respect to pushforward measures (see Theorem 3.6.1 in [7]),
Zσ(C(Y ))
evϕ d(δhY ◦ µ)x = Zσ(C(Y ))
evϕ d(cid:0)µx ◦ h−1
Y (cid:1) =ZY
since evϕ is continuous (and hence measurable). Therefore,
(evϕ ◦ hY ) dµx
(4.166)
Zσ(C(Y ))
evϕ(cid:0)hY (y)(cid:1) dµx(y)
hY (y)(cid:0)ϕ(cid:1) dµx(y)
evϕ d(δhY ◦ µ)x =ZY
=ZY
=ZY
=(cid:0)C µ(ϕ)(cid:1)(x)
= hX (x)(cid:0)C µ(ϕ)(cid:1)
= Zσ(C(Y ))
evϕ dσC(µ)
hX (x)
ϕ(y) dµx(y)
63
by (4.166)
by definition of evϕ
by definition of hY (3.64)
(4.167)
by definition of C µ (4.104)
by definition of hX (3.64)
by definition of σC(µ)
hX (x).
}
a
o
o
}
a
o
o
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
/
O
O
O
O
O
O
O
Since this is true for all ϕ ∈ C(Y ), the uniqueness of measures on σ(C(Y )) from Theorem
4.134 implies σC(µ)
hX (x) = (δhY ◦ µ)x for all x ∈ X. Therefore, the diagram above commutes and
h : idcHausStoch ⇒ σ ◦ C is a natural isomorphism.
ii. Proof that Γ is natural. Let f : B
diagram
o A be a positive map. The goal is to show that the
B
f
A
ΓB
C(σ(B))
C σ(f )
ΓA
C(σ(A))
(4.168)
commutes, i.e. for arbitrary b ∈ B, the two functions C σ(f )(cid:0)ΓB(b)(cid:1) and ΓA(cid:0)f (a)(cid:1) are equal on
σ(A). To check equality of these functions, evaluating them on an arbitrary χ ∈ σ(A) gives
(4.169)
evb dσf
(cid:16)C σ(f )(cid:0)ΓB(b)(cid:1)(cid:17)(χ) = Zσ(B)
χ = χ(cid:0)f (b)(cid:1) =(cid:16)ΓA(cid:0)f (a)(cid:1)(cid:17)(χ),
which is the desired result.
(cid:4)
We should also check that this equivalence agrees with the usual Gelfand-Naimark equiva-
lence for deterministic processes, i.e. for ∗-homomorphisms on the C ∗-algebra side and measure-
preserving functions on the analytic side.
Theorem 4.170. Let B and A be commutative C ∗-algebras, let f : B
and let χ ∈ σ(A) be a character on A. Then the corresponding regular probability measure σf
Dirac measure at χ◦ f, i.e. σf
σ(A) → σ(B) agreeing with the usual spectrum functor on cC*-Alg, i.e. the diagram
o A be a ∗-homomorphism,
χ is the
χ = δχ◦f . Therefore, σ(f ) uniquely determines a continuous function
cHaus
δ
cHausStoch
σ
σ
cC*-Algop
cC*-AlgPosop
(4.171)
commutes.
Proof. This follows from the fact that χ◦ f is a character on σ(B) and that it immediately satisfies
the conditions of Theorem 4.145 so that the measure is uniquely specified.
(cid:4)
Remark 4.172. The full power of Theorem 4.145 is not needed to prove this last result. It follows
from a simpler result due to Bauer (see Proposition 1.4 in [35]).
5 Closing discussion
We briefly discuss the broad physical scope of our results in the context of quantum mechanics
and then the relationship to similar conclusions by other authors.
64
/
/
/
o
/
/
/
/
/
O
O
O
O
O
O
O
/
/
/
o
/
o
o
o
o
5.1 Categories of classical and quantum dynamics
In this paper, we have collected functors
cHausop
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
δ
FinSetFunop
cC*-Alg
B✆✆✆✆✆✆✆✆✆✆✆✆
fdcC*-Alg
C
σ
C
σ
cHausStochop
cC*-AlgPos
(5.1)
δ
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
B✆✆✆✆✆✆✆✆✆✆✆✆
FinSetStochop
fdcC*-AlgPos
between a variety of categories. The left face of the cube consists of dynamics on classical systems
(phase space). The right face describes the algebraic counterparts in terms of observables of
those classical systems (functions on phase space) and how those evolve under the dynamics. The
top face describes deterministic dynamics in both space and algebraic settings. The bottom face
describes stochastic dynamics in both space and algebraic settings. The front face describes finite
classical systems (systems whose phase space is finite). The back face describes classical systems
whose phase space is possibly infinite. The functors between the left face and the right face are
all equivalences. The equivalences on the top face are due to Gelfand and Naimark while the back
edge of the bottom face has been described in this work.
This cube has a natural extension to the right for arbitrary C ∗-algebras. ∗-homomorphisms
are therefore non-commutative analogues of deterministic processes while completely positive maps
are non-commutative analogues of non-deterministic processes.
cHausop
cC*-Alg
C*-Alg
FinSetStochop
fdcC*-AlgPos
fdC*-AlgCP
The theorems presented in this work offer more motivation for thinking of completely positive
maps and the dynamics associated with them in quantum theory as a natural extension of stochas-
tic dynamics in classical theory. This can be seen more concretely if one adds more data to these
65
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
?⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧
C
σ
C
σ
B✆✆✆✆✆✆✆✆✆✆✆✆
B✆✆✆✆✆✆✆✆✆✆✆✆
B✆✆✆✆✆✆✆✆✆✆✆✆
B✆✆✆✆✆✆✆✆✆✆✆✆
FinSetFunop
fdcC*-Alg
fdC*-Alg
cHausStochop
cC*-AlgPos
C*-AlgCP
(5.2)
/
/
o
o
/
/
o
o
/
/
o
o
/
/
o
o
/
?
/
?
1
B
1
B
/
/
o
o
/
/
o
o
/
/
o
o
/
/
o
o
/
?
/
?
1
B
1
B
/
/
/
/
/
/
/
/
1
B
1
B
categories. The categories above only describe dynamics of the observables of a system, but not
of individual states. In other words, we have so far described the Heisenberg picture of quantum
mechanics. However, this is a red herring. Evolution of states is also described in this setting
o A describes
because a state is a completely positive map. If a completely positive map f : B
o C is a
the evolution of observables described by A to observables described by B, and ω : A
o C under the evolution described by f. Therefore, one
state on A, then ω evolves to ω ◦ f : B
can include both evolution of observables and states by taking the slice category of C*-AlgCP
over C, denoted by C*-AlgCP/C. The objects of this category are C ∗-algebras equipped with
o C consists of a completely positive
o C. A morphism from ξ : B
states ω : A
o A such that ξ = ω ◦ f. This is a non-commutative generalization of spaces equipped
map f : B
with probability measures. This situation was described in Example 2.93 though more general
dynamics is encompassed in Lindblad's formalism [27].
o C to ω : A
5.2 Categorical probability theory
As mentioned earlier, Sections 2 and 3 were entirely a review though phrased in a categorical
setting that was used to state and prove various results in Section 4. Many of the results in
Section 4 relied on well known facts in analysis though our categorical framework brought some
questions to surface which seem to have not been addressed earlier. Similar results were obtained
by Furber and Jacobs [15], who proved an equivalence of categories between cHausStochop and
cC*-AlgPos by realizing the category cHausStoch as the Kleisli category of a particular monad,
a slight variant of the Giry monad, on the category cHaus. There are three main differences
between our findings.
i. Rather than constructing the Giry monad, we worked directly with the objects and mor-
phisms of cHausStoch bypassing the monad entirely. This saves us quite a bit of space from
explaining background material on monads, Kleisli categories, and how these are related to
adjunctions, which is all beautifully explained in [42]. Furthermore, it makes our presentation
more accessible to those with a minimal categorical background. However, as a result, some
proofs become a bit more involved compared to [15]. For instance, given two stochastic maps
Z, proving that the composition ν ◦ µ is well-defined, continu-
µ : X
ous, regular, and a Markov kernel was quite involved (see Proposition 4.79 and the preceding
lemmas). From the Giry monad perspective, ProbMeas can be viewed as a functor from the
category cHaus to itself (one must prove this). In addition, there are two natural transforma-
tions that provide ProbMeas with the structure of a monad. The inclusion of a space X into
ProbMeas(X) via the Dirac measure defines a natural transformation δ : idcHaus ⇒ ProbMeas.
In addition, there is a natural transformation b : ProbMeas◦ ProbMeas ⇒ ProbMeas given on
a space X by the assignment
Y and ν : Y
ProbMeas(cid:0)ProbMeas(X)(cid:1) bX−→ ProbMeas(X)
MX ∋ E
Ω 7→
bX (Ω)
7−−−→ Z
evE dΩ
ProbMeas(X)
(5.3)
.
66
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
/
/
/
o
/
o
/
/
/
o
/
o
Notice that one must still show that evE is measurable to define this map. bX is surprisingly
easy to interpret and has significant physical meaning. ProbMeas(X) is a convex subset of
the vector space of all finite measures. A probability measure on it therefore has a barycenter,
its center of mass. The map bX is precisely the assignment of that probability measure to its
barycenter. This follows from the fact that for any Borel set E ∈ MX,
evE dΩ =
evE(µ) dΩ(µ) =
µ(E) dΩ(µ).
(5.4)
ProbMeas(X)
ProbMeas(X)
ProbMeas(X)
Z
Z
Z
µ
X
A physically relevant instance where the barycenter naturally occurs will be discussed after
point iii. Going back to the Giry monad and cHausStoch, using the barycenter map, the
definition of the composition ν ◦ µ can then be taken to be the composition of the functions
(5.5)
ProbMeas(ν)
−−−−−−−→ ProbMeas(cid:0)ProbMeas(Z)(cid:1) bZ−→ ProbMeas(Z)
from which it immediately follows that the composition is continuous, since each function in
this composition is continuous. The relationship between our formula for the composition and
this composition of functions follows from the change of variables formula since applying the
above composition to a point gives
−→ ProbMeas(Y )
X ∋ x 7→ µx 7→ ν∗µx 7→ bZ(ν∗µx),
(5.6)
where ν∗(µx) denotes the pushforward measure of µx along the (measurable) function ν : Y →
ProbMeas(Z). This evaluates on measurable sets E ∈ MZ to
evE d(ν∗µx) =ZY
(evE ◦ ν) dµx,
(5.7)
Z
(cid:0)bZ(ν∗µx)(cid:1)(E) =
ProbMeas(Z)
which agrees with our definition of composition (4.80). By using an intermediate amount of
category theory, our work offers a bridge towards more categorical approaches to probability
theory [19].
ii. In [15], Furber and Jacobs use the RMK Representation Theorem immediately so that a
Y is defined to be a continuous function X → S(cid:0)C(Y )(cid:1) into states
stochastic map X
on the C ∗-algebra of continuous functions on Y instead of a continuous function X →
ProbMeas(Y ) into regular probability measures on Y (this is how the Kleisli category as-
sociated to their version of the Giry monad was defined). There is no significant difference
between these perspectives. However, we prefer to think of the left side of our cube in Section
5.1 as purely analytical, topological, and measure-theoretic.
In particular, we provided an
explicit definition of the composition using Markov kernels and without any reference to the
RMK theorem. The right face of this cube is more algebraic and utilizes algebras of function
spaces, states, positive maps, etc. Of course, the point of the equivalence between these two
faces is that this distinction is just psychological. Nevertheless, it is helpful to distinguish
the different tools used in both subjects. As a result, many of the proofs in our paper are
completely different in nature from those of Furber and Jacobs.
67
/
/
/
o
/
o
iii. Finally, Furber and Jacobs showed that C : cHausStochop → cC*-AlgPos is fully faithful
and essentially surjective [15]. However, any abstract nonsense construction of an explicit
inverse will necessarily relate any commutative C ∗-algebra to one of continuous functions on
some space and will therefore use the RMK theorem in its construction. We have therefore
provided an explicit construction of an inverse without using the Gelfand transform nor the
RMK theorem. Furthermore, we showed that the usual Gelfand transform is part of the adjoint
equivalence for the stochastic Gelfand-Naimark Theorem. Our construction of the stochastic
spectrum functor applies to positive maps and take characters to states. Choquet theory was
used to "spread" this state out onto the spectrum in terms of a probability measure which was
obtained as a limit of successive approximations via decompositions of the state into convex
combinations of Dirac measures supported on the space of characters. This perspective may
be useful for explorations in the category of non-commutative C ∗-algebras.
One instance where the barycenter map is important in physics occurs in preparing systems
In an experiment, one has a protocol for preparing a state,
on which to perform experiments.
which will then be probed in some way, say, by measuring some observable. Because the probe
may alter the state, one needs the protocol to be as accurate as possible so that the same state
can be constructed so that it can be probed in subsequent experiments.
Indeed, this is how a
state is defined mathematically-via its expectation values. However, in practice, protocols, as
well as instruments, are rarely perfect. As a result, the prepared states may differ slightly from
one another so that they are more accurately represented by a probability measure on the set of
states. Hence, an experimentalist is not always obtaining expectation values for the same state.
Nevertheless, after viewing the set of states as probability measures via the RMK Representation
Theorem, the barycenter map provides one with a canonical state that represents the average
state prepared by the protocol. This is the state very often referred to in most textbooks, but one
should keep in mind that the barycenter represents only one part of all the data.
5.3 The Baire approach
One may have circumvented many of the difficulties we encountered by using Baire sets instead of
Borel sets. This perspective is emphasized in works such as [12]. In this case, one has the following
analogue of Lemma 4.64.
Lemma 5.8. Let Y be a compact Hausdorff space, let E ⊆ Y be a Baire set, and let Pr(Y ) denote
all probability measures on Y. Then the evaluation function
Pr(Y )
evE−−→ R
µ 7→ µ(E)
is Borel measurable.
Proof. Let P be the collection of all compact Gδ subsets of Y and let
L :=nE ⊆ Y : Pr(Y )
evE−−→ [0, 1] is Borel measurableo
68
(5.9)
(5.10)
i. The intersection of compact sets is compact and the finite intersection of a countable collection
of open sets is still a countable intersection of open sets. Hence P is a π-system.
ii. This is the same as in the proof of Lemma 4.64 so L is a λ-system.
iii. Let K = U1∩U2∩U3∩· · · be a compact Gδ set in Y with {Un} an at most countable collection
of open sets in Y. It suffices to assume that U1 ⊇ U2 ⊇ U3 ⊇ · · · since if this were not the case,
the sequence U1, U1 ∩ U2, U1 ∩ U2 ∩ U3, . . . would also have intersection equal to K. For each n,
by Urysohn's Lemma, there exists a continuous function fn : Y → [0, 1] such that fn(y) = 1
for y ∈ K and fn(y) = 0 for y ∈ U c
n. Similarly, it suffices to assume that f1 ≥ f2 ≥ f3 ≥ · · ·
since if this were not the case, the sequence N × Y ∋ (n, y) 7→ min(cid:8)f1(y), f2(y), . . . , fn(y)(cid:9)
would also have the same pointwise limit. From these assumptions, it follows that
lim
n→∞
fn(y) = χK(y)
∀ y ∈ Y.
Hence, for any a ∈ R, the preimage of (a,∞) under evK is
ev−1
K (cid:0)(a,∞)(cid:1) =(cid:8)µ ∈ Pr(Y ) : µ(K) > a(cid:9)
=(cid:26)µ ∈ Pr(Y ) :
n→∞ZY
\n=1(cid:26)µ ∈ Pr(Y ) : ZY
lim
by the Monotone Convergence Theorem
∞
=
fn dµ > a(cid:27)
fn dµ > a(cid:27) since · · · ≥ fn ≥ fn+1 ≥ · · · .
(5.11)
(5.12)
By definition of the vague topology, the set in curly brackets in the last equality is open.
Hence, this is a countable intersection of open sets and is therefore a Borel subset of Pr(Y ).
This shows that evK is Borel measurable for all compact Gδ sets K.
Since all the conditions of Dynkin's π-λ theorem are satisfied, the sigma algebra generated by P,
which is the Baire σ-algebra on Y, is contained L.
(cid:4)
This allows one to define the composition of stochastic maps on Baire sets. One can then
use the theorem that every Baire measure extends uniquely to a regular measure under these
assumptions (see Corollary 7.3.4. in [8]). However, this would not guarantee that formula (4.80)
for the composition of stochastic maps is still a valid mathematical expression. The construction
of the regular Borel extension of this measure is somewhat formal and does not provide one with
the explicit formula we have produced.
69
Index of notation
Notation
FinSetFun
FinProb
Pr(Y )
Y
X
fyx
δyy′
FinSetStoch
A,B,C
1A
k · k
iff
CX
Mn(C)
C*-Alg
cC*-Alg
fdcC*-Alg
CP map
B → A
o A
B
C*-AlgCP
cC*-AlgPos
fdcC*-AlgPos
C
C(X)
C f ≡ C(f )
cHaus
C
B
Name and/or description
Location
Page
category of finite sets and functions
Notation 2.12
category of probability measures on finite sets
and measure-preserving functions
probability measures on a finite set Y
stochastic map from X to Y
yx-component of a stochastic map
f : X
Y
the value of the Kronecker delta function
on (y, y′) ∈ Y × Y
category of finite sets and stochastic maps
an algebra (often a unital C ∗-algebra)
multiplicative unit in an algebra A
a norm or seminorm
"if and only if" (used only in definitions)
set of all functions from X to C
set of all complex n × n matrices
category of C ∗-algebras and ∗-homomorphisms
the subcategory of C*-Alg of commutative
the subcategory of cC*-Alg of finite-dimensional
C ∗-algebras and ∗-homomorphisms
C ∗-algebras and ∗-homomorphisms
completely positive (unital) map
∗-homomorphism of C ∗-algebras
completely positive map of C ∗-algebras
category of C ∗-algebras
and completely positive (CP) maps
the subcategory of C*-AlgCP of
commutative C ∗-algebras and CP maps
the subcategory of cC*-AlgPos of
finite-dimensional C ∗-algebras and CP maps
functor C : FinSetStochop → fdcC*-AlgPos
and its restriction to FinSetFun
continuous functions on X
pullback C(Y ) → C(X) associated to f : X → Y
category of compact Hausdorff spaces
and continuous maps
functor cHausop → cC*-Alg
base on a set (for a topology)
70
Notation 2.12
Definition 2.13
Definition 2.13
Definition 2.13
Equation 2.19
Notation 2.27
Definition 2.34
Definition 2.34
Definition 2.35
Definition 2.35
Example 2.48
Example 2.50
Definition 2.52
Definition 2.52
Definition 2.52
Definition 2.63
Notation 2.66
Notation 2.66
Notation 2.67
Notation 2.67
Notation 2.67
Theorem 2.72
Theorem 2.78
Example 3.1
Proposition 3.3
Notation 3.6
Proposition 3.7
Definition 3.13
7
7
8
8
8
8
10
11
11
12
12
13
13
13
13
13
15
15
15
15
15
15
16
17
24
24
24
24
25
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
n
k · kα
V ∨
V ∗
Θ
σ(A)
σ
Γ
h
σ(f ) ≡ σf
(Y,MY )
∅
µ, ν
CMeas(Y )
χE
s
R · dµ
ProbMeas(Y )
µ : X
Y
δf
ν ◦ µ
cHausStoch
δ
C
RMK
ex(C)
S(A)
ch(X)
σ
Notation 3.18
Proposition 3.26
Proposition 3.23
Proposition 3.23
Proposition 3.32
Definition 3.38
Proposition 3.43
Proposition 3.50
Proposition 3.59
Proposition 3.66
Section 4.1
Definition 4.2
Definition 4.2
Definition 4.2
Definition 4.6
Definition 4.6
Definition 4.6
Notation 4.47
Definition 4.57
Example 4.58
26
27
27
27
28
29
29
30
32
32
36
36
36
36
36
36
36
42
44
44
Proposition 4.79
48
Theorem 4.102
Theorem 4.102
Theorem 4.106
Theorem 4.112
Definition 4.127
Example 4.129
Definition 4.130
Theorem 4.151
51
51
52
53
56
56
56
60
the set {1, 2, . . . , n} with n ∈ N
seminorm indexed by α in some set
algebraic (linear) dual of vector space V
topological dual of normed space V
ordered index set for nets
spectrum of commutative Banach algebra A
spectrum of a homomorphism f
spectrum functor cC*-Algop → cHaus
Gelfand transform Γ : idcC*-Alg ⇒ C ◦ σ
natural transformation h : idcHaus ⇒ σ ◦ C
compact Hausdorff space Y
with its Borel σ-algebra MY
the empty set
measure (positive, real, or complex)
set of complex measures on Y
characteristic/indicator function on E
simple function
integral with respect to a measure µ
set of regular probability measures on Y
(equipped with the vague topology)
stochastic map described by a
continuous function µ : X → ProbMeas(Y )
Dirac delta stochastic map
associated to a continuous function f
composition of µ : X
Y
followed by ν : Y
o Z
category of compact Hausdorff spaces
and stochastic maps
Dirac measure functor δ : cHaus → cHausStoch
functor cHausStochop → cC*-AlgPos
short for "Riesz-Markov-Kakutani"
extreme points of a convex set C
states on commutative C ∗-algebra A
convex hull of X
stochastic spectrum functor
cC*-AlgPosop → cHausStoch
71
/
/
/
o
/
o
/
/
/
o
/
o
/
/
/
o
/
References
[1] Erik M. Alfsen, Compact convex sets and boundary integrals, Ergebnisse der Mathematik und ihrer Grenzge-
biete. 2. Folge, Springer-Verlag Berlin Heidelberg, 1971. ↑58
[2] John C. Baez and Tobias Fritz, A bayesian characterization of relative entropy, Theory and Applications of
Categories 29 (2014), no. 16, 422–456, available at 1402.3067. ↑5, 8
[3] John C. Baez, Tobias Fritz, and Tom Leinster, A characterization of entropy in terms of information loss,
Entropy 13 (2011), 1945–1957, available at 1106.1791. ↑8
[4] John C. Baez and Mike Stay, Physics, topology, logic and computation: a Rosetta Stone, New structures for
physics, 2011, pp. 95–172. ↑2
[5] Patrick Billingsley, The invariance principle for dependent random variables, Trans. Amer. Math. Soc. 83
(1956), 250–268. ↑4
[6]
, Probability and measure, 3rd ed., New York: John Wiley & Sons, Inc., 1995. ↑46
[7] Vladimir I. Bogachev, Measure theory, Vol. 1, Springer-Verlag Berlin Heidelberg, 2007. ↑63
[8]
, Measure theory, Vol. 2, Springer-Verlag Berlin Heidelberg, 2007. ↑69
[9] Fabio Cipriani, Positive maps on c∗-algebras, Encyclopedia of Mathematical Physics 3 (2006), 88–94. ↑2, 35
[10] Kenneth R. Davidson, C ∗-algebras by example, Fields Institute Monographs, vol. 6, American Mathematical
Society, Providence, RI, 1996. ↑13
[11] Jacques Dixmier, C ∗-algebras, North-Holland Publishing Co., Amsterdam-New York-Oxford, 1977. Translated
from the French by Francis Jellett, North-Holland Mathematical Library, Vol. 15. ↑11, 12, 13
[12] Tanja Eisner, B´alint Farkas, Markus Haase, and Rainer Nagel, Operator theoretic aspects of ergodic theory,
Graduate Texts in Mathematics, vol. 272, Springer, Cham, 2015. ↑2, 4, 35, 68
[13] Gerald B. Folland, A course in abstract harmonic analysis, Textbooks in Mathematics, CRC Press, 1994. ↑23,
29, 31, 32
[14]
54
, Real analysis: Modern techniques and their applications, 2nd ed., Wiley, 2007. ↑26, 27, 28, 35, 42, 43,
[15] Robert Furber and Bart Jacobs, From kleisli categories to commutative C ∗-algebras: probabilistic gelfand du-
ality, Logical Methods in Computer Science 11 (2015), 1–28. ↑3, 4, 35, 66, 67, 68
[16] Alex Furman, What is . . . a stationary measure?, Notices Amer. Math. Soc. 58 (2011), no. 9, 1276–1277. ↑35
[17] Nicolas Gagn´e and Prakash Panangaden, A categorical characterization of relative entropy on polish spaces
(2017), available at 1703.08853. ↑8
[18] Israel Gelfand and Mark Neumark, On the imbedding of normed rings into the ring of operators in Hilbert
space, Rec. Math. [Mat. Sbornik] N.S. 12(54) (1943), 197–213. ↑2, 33
[19] Mich`ele Giry, A categorical approach to probability theory, Categorical aspects of topology and analysis (Ottawa,
Ont., 1980), 1982, pp. 68–85. ↑3, 4, 11, 35, 45, 67
[20] Brian C. Hall, Quantum theory for mathematicians, Graduate Texts in Mathematics, vol. 267, Springer, New
York, 2013. ↑14
[21] Chris Heunen, The many classical faces of quantum structures, Entropy 19 (2017), no. 4, 144, available at
1412.2177. ↑5, 35
[22] Richard V. Kadison, A representation theory for commutative topological algebra, Mem. Amer. Math. Soc.,
No. 7 (1951), 39. ↑23
72
[23] Shizuo Kakutani, Concrete representation of abstract (M )-spaces. (A characterization of the space of continuous
functions.), Ann. of Math. (2) 42 (1941), 994–1024. ↑2, 23, 35
[24] Dexter Kozen, Semantics of probabilistic programs, J. Comput. System Sci. 22 (1981), no. 3, 328–350. Special
issue dedicated to Michael Machtey. ↑3
[25] Karl Kraus, States, effects, and operations: Fundamental notions of quantum theory, Lecture Notes in Physics
Vol. 190, Springer Berlin Heidelberg, 1983. ↑14, 15
[26] William F. Lawvere, The category of probabilistic mappings, 1962. preprint. ↑3
[27] Goran Lindblad, On the generators of quantum dynamical semigroups, Comm. Math. Phys. 48 (1976), no. 2,
119–130. ↑2, 15, 23, 66
[28] Fred E. J. Linton, The functorial foundations of measure theory, Columbia University, 1963. dissertation thesis.
↑3
[29] James R. Munkres, Topology: a first course, 2nd ed., Prentice-Hall, Inc., 2000. ↑24
[30] Michael A. Nielsen and Isaac L. Chuang, Quantum computation and quantum information, 10th Anniversary,
Cambridge University Press, Cambridge, 2010. ↑15, 22, 23
[31] Prakash Panangaden, The category of Markov kernels, PROBMIV'98: First International Workshop on Prob-
abilistic Methods in Verification (Indianapolis, IN), 1999, pp. 17. ↑3
[32] Arthur J. Parzygnat, From observables and states to hilbert space and back: a 2-categorical adjunction (2016),
available at 1609.08975. ↑14
[33] Vern Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Advanced Mathematics,
Cambridge University Press, 2003. ↑15
[34] Gert K. Pedersen, Analysis now, Graduate Texts in Mathematics, vol. 118, Springer, 1989. ↑27
[35] Robert R. Phelps, Lectures on choquet's theorem, 2nd ed., Lecture Notes in Mathematics, vol. 1757, Springer-
Verlag Berlin Heidelberg, 2001. ↑35, 56, 58, 59, 64
[36] Mikl´os R´edei and Stephen J. Summers, Quantum probability theory, Studies in History and Philosophy of
Modern Physics 38 (2007), 390–417, available at quant-ph/0601158. ↑14
[37] Walter Rudin, Principles of mathematical analysis, Third, McGraw-Hill Book Co., New York-Auckland-
Dusseldorf, 1976. International Series in Pure and Applied Mathematics. ↑24
[38]
[39]
, Real and complex analysis, 3rd ed., McGraw-Hill Inc., 1987. ↑36, 37, 38, 39, 42, 54, 58
, Functional analysis, 2nd ed., International Series in Pure and Applied Mathematics, McGraw-Hill
Inc., 1991. ↑23, 28, 29, 56, 58
[40] Shoichiro Sakai, c∗-algebras and w∗-algebras, 1st ed., Ergebnisse der Mathematik und ihrer Grenzgebiete,
Springer-Verlag Berlin Heidelberg, 1971. ↑58
[41] Jun John Sakurai, Modern Quantum Mechanics; revised ed., Addison-Wesley, 1993. ↑21
[42] Zbigniew Semadeni, Monads and their eilenberg-moore algebras in functional analysis, Queen's Papers in Pure
and Applied Mathematics, Issue 33, Queen's University, 1973. ↑3, 66
[43] Ambar Sengupta, Lecture 6. the dynkin π-λ theorem, Louisiana State University, 2009. Available as:
https://www.math.lsu.edu/~sengupta/7360f09/DynkinPiLambda.pdf. ↑45, 46
[44] Barry Simon, Convexity: An analytic viewpoint, 1st ed., Cambridge Tracts in Mathematics (Book 187), Cam-
bridge University Press, 2011. ↑56, 58
[45] Lynn Arthur Steen and J. Arthur Seebach Jr., Counterexamples in topology, Dover Publications, Inc., Mineola,
NY, 1995. Reprint of the second (1978) edition. ↑42
73
[46] Tadeusz ´Swirszcz, Monadic functors and convexity, Bull. Acad. Polon. Sci. S´er. Sci. Math. Astronom. Phys.
22 (1974), 39–42. ↑3
[47] Terence Tao,
245b,
notes
11: The
strong
and
weak
topologies,
2009. Available
as:
https://terrytao.wordpress.com/2009/02/21/245b-notes-11-the-strong-and-weak-topologies/.
↑23, 25, 26, 27, 28
[48]
, 245b, notes 12: Continuous functions on locally compact hausdorff spaces, 2009. Available as:
https://terrytao.wordpress.com/2009/03/02/245b-notes-12-continuous-functions-on-locally-compact-hausdorff-spaces/.
↑54
[49] Garth Warner, C*-algebras, 2010. ↑15
[50] Stephen Willard, General topology, Dover Books on Mathematics, Dover Publications, 2004. ↑26, 28
Mathematics Department, University of Connecticut, 341 Mansfield Road U1009,
Storrs, CT 06269, USA
Email: [email protected]
74
|
1105.0877 | 1 | 1105 | 2011-05-04T17:38:08 | Fundamental solutions of evolutionary PDOs and rapidly decreasing distributions | [
"math.FA",
"math.AP"
] | Let $P(\partial_0,\partial_1,...,\partial_n)$ be a PDO on $\symR^{1+n}$ with constant coefficients. It is proved that (i) the real parts of the $\lambda$-roots of the polynomial $P(\lambda,i\xi_1,...,i\xi_n)$ are bounded from above when $(\xi_1,...,\xi_n)$ ranges over $\symR^n$ if and only if (ii) $P$ has a fundamental solution with support in $H_+=\{(x_0,x_1,\allowbreak..., x_n)\in \symR^{1+n}:x_0\ge0\}$ having some special properties expressed in terms of the L. Schwartz space $\calO^{\prime}_C$ of rapidly decreasing distributions. Moreover, it is proved that the fundamental solution with support in $H_+$ having these special properties is unique. | math.FA | math | Fundamental solutions of evolutionary PDOs
and rapidly decreasing distributions
Jan Kisy´nski
Institute of Mathematics, Polish Academy of Sciences
´Sniadeckich 8, 00-956 Warszawa, Poland
E-mail: [email protected]
1
1
0
2
y
a
M
4
]
.
A
F
h
t
a
m
[
1
v
7
7
8
0
.
5
0
1
1
:
v
i
X
r
a
Abstract
Let P (∂0, ∂1, . . . , ∂n) be a PDO on R1+n with constant coeffi-
cients. It is proved that
(i) the real parts of the λ-roots of the polynomial P (λ, iξ1, . . . , iξn)
are bounded from above when (ξ1, . . . , ξn) ranges over Rn
if and only if
(ii) P has a fundamental solution with support in H+ = {(x0, x1,
. . . , xn) ∈ R1+n : x0 ≥ 0} having some special properties ex-
pressed in terms of the L. Schwartz space O′
C of rapidly de-
creasing distributions.
Moreover, it is proved that the fundamental solution with support in
H+ having these special properties is unique.
1
Introduction and the main result
1.1 Rapidly decreasing distributions
By Theorem IX in Sec. VII.5 of L. Schwartz's book [S2], for every distribu-
tion T ∈ D′(Rn) the following two conditions are equivalent:
(1.1) T ∗ ϕ ∈ S(Rn) for every ϕ ∈ D(Rn),
2010 Mathematics Subject Classification: Primary 35E05, 46F99.
Key words and phrases: PDO with constant coefficients, Petrovskiı condition, rapidly
decreasing distributions, slowly increasing functions.
1
2
J. Kisy´nski
(1.2) for every k ∈ N0 there is mk ∈ N0 such that T = Pα≤mk
∂αFk,α
where, for every multiindex α = (α1, . . . , αn) ∈ N0 of length α =
α1 + · · · + αn ≤ mk, Fk,α is a continuous function on Rn such that
supx∈Rn(1 + x)kFk,α(x) < ∞.
In the above, and everywhere in the following, ∂α = ∂α1
n where
1
∂1, . . . , ∂n are partial derivatives of the first order not multiplied by any
factor. Each of the conditions (1.1), (1.2) is satisfied if and only if the dis-
tribution T is rapidly decreasing, where the definition of rapid decrease, due
to L. Schwartz, refers to the notion of boundedness of a distribution. The
C(Rn). From
space of rapidly decreasing distributions on Rn is denoted by O′
(1.2) it follows that
. . . ∂αn
(1.3) whenever T ∈ O′
C(Rn) and ϕ ∈ C ∞
b (Rn),
then ϕT ∈ O′
C(Rn).
It is clear from (1.2) that O′
FT makes sense for every T ∈ O′
[S2],
C(Rn) ⊂ S ′(Rn), so that the Fourier transform
C(Rn). By Theorem XV in Sec. VII.8 of
(1.4)
FO′
C(Rn) = OM (Rn),
where OM (Rn) denotes the space of infinitely differentiable slowly increasing
functions on Rn. Recall that φ ∈ OM (Rn) if and only if φ ∈ C ∞(Rn) and
for every α ∈ Nn
0 there is mα ∈ N0 such that
sup
ξ∈Rn
(1 + ξ)−mα∂αφ(ξ) < ∞.
Complete proofs of theorems about O′
C(Rn) and OM (Rn) needed in the
present paper may be found in [K].
1.2 The main result
Our object of interest will be the differential operator P (∂0, ∂1, . . . , ∂n) on
R1+n = {(x0, x1, . . . , xn) : xν ∈ R for ν = 0, . . . , n} with constant coeffi-
cients, and the associated polynomial P (λ, iξ1, . . . , iξn) defined on C × Rn.
A distribution N on R1+n such that
P N ≡ δ
is called a fundamental solution for the operator P . Let
H+ = {(x0, x1, . . . , xn) ∈ R1+n : x0 ≥ 0}.
Evolutionary PDOs and rapidly decreasing distributions
3
If there exists a fundamental solution N for P such that supp N ⊂ H+,
then the operator P is said to be evolutionary with respect to H+. For every
fixed λ ∈ C let e−λ be the function on R1+n given by e−λ(x0, x1, . . . , xn) =
exp(−λx0) for (x0, x1, . . . , xn) ∈ R1+n. For ϑ ∈ D(R), denote by ϑ0 the
function on R1+n such that ϑ0(x0, x1, . . . , xn) = ϑ(x0). Let
O′
LOC(H+) = {T ∈ D′(R1+n) : supp T ⊂ H+,
ϑ0T ∈ O′
C(R1+n) for every ϑ ∈ D(R)}.
Theorem. Let P (∂0, ∂1, . . . , ∂n) be the differential operator on R1+n with
constant coefficients. Let
ω0 = sup{Re λ : λ ∈ C and there is (ξ1, . . . , ξn) ∈ Rn
such that P (λ, iξ1, . . . , iξn) = 0}.
Then the following two conditions are equivalent:
(i) ω0 < ∞,
(ii) the differential operator P (∂0, ∂1, . . . , ∂n) has a fundamental solution
N belonging to O′
LOC(H+).
Furthermore, if (i) and (ii) are satisfied, then the fundamental solution N
as in (ii) is unique and satisfies
(iii) ω0 = inf{Re λ : λ ∈ C, e−λN ∈ O′
C(R1+n)}, and e−λN ∈ O′
whenever Re λ > ω0.
C(R1+n)
1.3 Remarks
Condition (i) can be called the Petrovskiı condition because it first ap-
peared in I. G. Petrovskiı's paper [P]. Namely, in [P], in the footnote on
p. 24, it was conjectured that, if the polynomial P (λ, iξ1, . . . , iξn) is unital
with respect to λ, then this condition is equivalent to a certain formally
weaker condition also concerning the λ-roots of P (λ, iξ1, . . . , iξn). The va-
lidity of this conjecture was proved by L. Garding in [G]. I. G. Petrovskiı
noticed the significance of smooth slowly increasing functions for the theory
of evolutionary PDEs with constant coefficients. L. Schwartz explained in
[S1] how the results of Petrovskiı may be elucidated by placing them in the
framework of rapidly decreasing distributions and smooth slowly increasing
functions. (Condition (i) was not mentioned in [S1]; notice that [S1] was
earlier than [G].)
4
J. Kisy´nski
L. Hormander proved in [H1] that if P (ζ0, ζ1, . . . , ζn) is a polynomial of
1 + n complex variables, then the following two conditions are equivalent:
(i)∗ there are constants A ∈ ]−∞,∞[ and r ∈ ]0,∞[ such that
inf{Re F (ζ1, . . . , ζn) : (ζ1, . . . , ζn) ∈ Biξ1,...,iξn;r} ≤ A
for every (ξ1, . . . , ξn) ∈ Rn and every function F holomorphic in the
ball
Biξ1,...,iξn;r =n(ζ1, . . . , ζn) ∈ Cn :
nXν=1
ζν − iξν2 < r2o
such that P (F (ζ1, . . . , ζn), ζ1, . . . , ζn) = 0 in Biξ1,...,iξn;r,
(ii)∗ the differential operator P (∂0, ∂1, . . . , ∂n) has a fundamental solution
with support in H+.
The equivalence (i)∗⇔(ii)∗ was reproved in Sec. 12.8 of [H2]. The funda-
mental solution occurring in (ii)∗ need not be unique. It is non-unique if
(i)∗ holds and the boundary of H+ is characteristic for P (∂0, ∂1, . . . , ∂n).
Obviously (i) implies (i)∗. Furthermore, as indicated in [H1], the operator
∂0 − i(∂1 + 1)2 satisfies (i)∗ but does not satisfy (i). Therefore condition (i)∗
is essentially weaker than (i).
Let us stress that in [H1], and in the present paper, the largest power
of λ in P (λ, iξ1, . . . , iξn) is multiplied by a polynomial of ξ1, . . . , ξn which,
in contrast to the assumption (5) in Sec. 3.10 of [R], may vanish for some
(ξ1, . . . , ξn) ∈ Rn.
2 Existence of a fundamental solution satis-
fying (ii) and (iii)
2.1 Application of the Tarski -- Seidenberg theorem
We are going to prove that if (i) holds, then the differential operator
P (∂0, ∂1, . . . , ∂n) has a fundamental solution N satisfying the conditions
(ii) and (iii). So, suppose that (i) holds and let
N = {(σ, ξ0, . . . , ξn) ∈ R2+n : P (σ + iξ0, iξ1, . . . , iξn) = 0}.
Then N ⊂ {(σ, ξ0, . . . , ξn) ∈ R2+n : σ ≤ ω0}, and hence, by Theorem A.3
from the Appendix to [T] or by Theorem 3.2 of [Go]∗), there are c, µ, µ′ ∈
∗)Following the idea of L. Hormander, these theorems are deduced from the Tarski --
Seidenberg theorem about projections of semi-algebraic sets.
Evolutionary PDOs and rapidly decreasing distributions
5
]0,∞[ such that whenever σ ∈ ]ω0,∞[, and (ξ0, . . . , ξn) ∈ R1+n, then
(2.1)
P (σ + iξ0, iξ1, . . . , iξn) ≥ c(dist((σ, ξ0, . . . , ξn);N ))µ
· (1 + (σ2 + ξ2
≥ c(σ − ω0)µ(1 + σ + iξ0 + (ξ2
0 + · · · + ξ2
1 + · · · + ξ2
n)1/2)−µ′
n)1/2)−µ′
.
2.2 The slowly increasing functions cNσ and the rapidly
decreasing distributions Nσ
For every σ ∈ ]ω0,∞[ let
(2.2)
cNσ(ξ0, . . . , ξn) = (P (σ + iξ0, iξ1, . . . , iξn))−1
for (ξ0, . . . , ξn) ∈ R1+n. Then, for every multiindex α ∈ N1+n,
∂αcNσ(ξ0, . . . , ξn) = (P (σ + iξ0, iξ1, . . . , iξn))−1−αQα(σ, ξ0, . . . , ξn)
where Qα is a polynomial. Consequently, (2.1) implies that
cNσ ∈ OM (R1+n)
for every σ ∈ ]ω0,∞[.
Nσ = F−1cNσ
(2.3)
Let
(2.4)
where F denotes the Fourier transformation on R1+n such that
(Fϕ)(ξ0, . . . , ξn) = bϕ(ξ0, . . . , ξn)
=Z · · ·Z
e−i Pn
ν=0 xν ξν ϕ(x0, . . . , xn) dx0 . . . dxn
R1+n
for ϕ ∈ S(R1+n), and F is extended onto S ′(R1+n) by duality. From (1.4)
and (2.3) it follows that
(2.5)
Nσ ∈ O′
C(R1+n)
for every σ ∈ ]ω0,∞[.
Furthermore, from (2.2) it follows that
(2.6) if σ ∈ ]ω0,∞[ then Nσ is a fundamental solution for the differential
operator P (σ + ∂0, ∂1, . . . , ∂n).
6
J. Kisy´nski
Take σ ∈ ]ω0,∞[, and consider the distribution eσNσ ∈ D′(R1+n). By the
Parseval equality, for every ϕ ∈ D(R1+n) one has
heσNσ, ϕi = hNσ, eσϕi = (2π)−1−nhcNσ,deσϕ∨i
= (2π)−1−nZ · · ·Z
(deσϕ(−ξ0, . . . ,−ξn)(P (σ + iξ0, iξ1, . . . , iξn))−1 dξ0 . . . dξn.
R1+n
For every ϕ ∈ D(R1+n) the Fourier integral
bϕ(ζ0, . . . , ζn) =Z · · ·Z
R1+n
e−i Pn
ν=0 xν ζν ϕ(x0, . . . , xn) dx0 . . . dxn
makes sense for (ζ0, . . . , ζn) ∈ C1+n and defines the holomorphic extension
of bϕ from R1+n onto C1+n. This holomorphic extension satisfies
deσϕ(ζ0, . . . , ζn) = bϕ(ζ0 + iσ, ζ1, . . . , ζn).
Consequently, whenever ϕ ∈ D(R1+n) and σ ∈ ]ω0,∞[, then
R1+n bϕ(−ξ0 + iσ,−ξ1, . . . ,−ξn)
heσNσ, ϕi = (2π)−1−nZ · · ·Z
(2.7)
· P (σ + iξ0, ξ1, . . . , ξn)−1 dξ0 . . . dξn.
Integration by parts shows that whenever ϕ ∈ D(R1+n) and l ∈ N, then
(1 + ξ0 − iσl + ξ1l + · · · + ξnl)bϕ(−ξ0 + iσ,−ξ1, . . . ,−ξn)
(2.8)
νϕkL1(R1+n)(cid:17) exp(Hϕ(σ))
k∂l
≤(cid:16)kϕkL1(R1+n) +
nXν=0
for every σ, ξ0, . . . , ξn ∈ R where
(2.9)
Hϕ(σ) = sup{σx0 : (x0, . . . , xn) ∈ supp ϕ}.
From (2.1), (2.7) -- (2.9) and the Cauchy integral theorem it follows that
(2.10) the distribution eσNσ ∈ D′(R1+n) does not depend on σ provided
that σ ∈ ]ω0,∞[,
(2.11) limσ→∞heσNσ, ϕi = 0 whenever ϕ ∈ D(R1+n) and supp ϕ⊂ R1+n\H+.
Evolutionary PDOs and rapidly decreasing distributions
7
2.3 The fundamental solution N
Thanks to (2.10) we may define the distribution N ∈ D′(R1+n) by the
equality
(2.12)
N = eσNσ
for every σ ∈ ]ω0,∞[.
From (2.11) it follows that
(2.13)
For every σ ∈ R let
supp N ⊂ H+.
(2.14)
Sσ = P (σ + ∂0, ∂1, . . . , ∂n)δ.
Since (−∂0)k(e−σϕ) = e−σ(σ − ∂0)kϕ, it follows that
(2.15)
P (−∂0,−∂1, . . . ,−∂n)(e−σϕ) = e−σP (σ − ∂0,−∂1, . . . ,−∂n)ϕ
for every σ ∈ R and ϕ ∈ D(R1+n). From (2.15) one infers that
hS0, e−σϕi = [P (−∂0,−∂1, . . . ,−∂n)(e−σϕ)](0)
= [eσP (−∂0,−∂1, . . . ,−∂n)(e−σϕ)](0)
= [P (σ − ∂0,−∂1, . . . ,−∂n)ϕ](0) = hSσ, ϕi,
proving that
(2.16)
Sσ = e−σS0
for every σ ∈ R.
From (2.6), (2.12) and (2.15) it follows that whenever σ ∈ ]ω0,∞[, then
P N = S0 ∗ N = (eσSσ) ∗ (eσNσ) = eσ(Sσ ∗ Nσ) = eσδ = δ,
so that
(2.17)
N is a fundamental solution for the operator P .
Above we have used the fact that whenever T, U ∈ D′(R1+n), σ ∈ R, and
one of T, U has compact support, then eσ(T ∗ U) = (eσT ) ∗ (eσU). This is
true under the additional assumption that T, U ∈ L1
loc(R1+n), and this case
implies the general assertion by regularization.
8
J. Kisy´nski
2.4 Properties of N
If ϑ ∈ D(R) and σ ∈ ]ω0,∞[, then ϑ0eσ is bounded on R1+n together with
all its partial derivatives, so that, by (1.3), ϑ0N = (ϑ0eσ)Nσ ∈ O′
C(R1+n)
because Nσ ∈ O′
(2.18)
C(R1+n). Hence, by (2.13),
N ∈ O′
LOC(H+).
The relations (2.17) and (2.18) show that (i) implies (ii). We are going
to prove that N defined by (2.12) satisfies (iii). To this end, take λ ∈ C
such that Re λ ∈ ]ω0,∞[. Let σ = 1
2(ω0 + Re λ). Then e−λN = eσ−λNσ ∈
O′
C(R1+n) because Nσ ∈ O′
C(R1+n), supp Nσ ⊂ H+, and eσ−λ is bounded
together with all its partial derivatives on the set {(x0, . . . , xn) ∈ R1+n :
x0 > −1}. It remains to prove that
if λ ∈ C and e−λN ∈ O′
(2.19)
C(R1+n), then Re λ ≥ ω0.
C(R1+n). Take any σ ∈ ]Re λ,∞[.
So, suppose that λ ∈ C and e−λN ∈ O′
Since eλ−σ is bounded on {(x0, . . . , xn) ∈ R1+n : x0 > −1} together with
all its partial derivatives, it follows by (1.3) that e−σN = eλ−σ(e−λN) ∈
C(R1+n). Furthermore
O′
Sσ ∗ (e−σN) = (e−σS0) ∗ (e−σN) = e−σ(S0 ∗ N) = e−σδ = δ.
Let φ = F(e−σN). Then φ ∈ OM (R1+n) and
P (σ + iξ0, iξ1, . . . , iξn) · φ(ξ0, . . . , ξn) = [F(Sσ ∗ (e−σN))](ξ0, . . . , ξn) = 1
for every (ξ0, . . . , ξn) ∈ R1+n. This implies that P (σ + iξ0, iξ1, . . . , iξn) 6= 0
for every (ξ0, . . . , ξn) ∈ R1+n. Since this is true for every σ ∈ ]Re λ,∞[, it
follows that Re λ ≥ ω0, proving (2.19).
3 Uniqueness of the fundamental solution be-
longing to O′LOC(H+)
3.1 An associativity relation for convolution
Lemma 3.1. Suppose that (i) holds. Fix σ ∈ ]ω0,∞[ and define Nσ and
Sσ by (2.4) and (2.14). Suppose moreover that U ∈ O′
LOC(H+) and that
P (∂0, ∂1, . . . , ∂n)U has compact support. Then
(3.1)
(Nσ ∗ Sσ) ∗ (e−σU) = Nσ ∗ (Sσ ∗ (e−σU)).
Evolutionary PDOs and rapidly decreasing distributions
9
Proof. Notice that both sides of (3.1) are well defined because every sign ∗
in (3.1) denotes a convolution of two distributions on R1+n one of which has
compact support. To see this it is sufficient to observe that supp Sσ = {0},
Nσ ∗ Sσ = Sσ ∗ Nσ = δ, and, by (2.15),
Sσ ∗ (e−σU) = P (σ + ∂0, ∂1, . . . , ∂n)(e−σU) = e−σ(P (∂0, ∂1, . . . , ∂n)U)
has compact support. However, from the three factors Nσ, Sσ and e−σU
occurring in (3.1) only one has compact support, so that (3.1) does not
follow from any of the simple criterions of the associativity of convolution.
In order to prove that both sides of (3.1) are equal we will apply an argument
going back to C. Chevalley ([Che, pp. 120 -- 121], proof of Theorem 2.2) which
reduces the problem to the Fubini -- Tonelli theorem.
Since the set {ϕ1 ∗ ϕ2 ∗ ϕ3 : ϕi ∈ D(R1+n) for i = 1, 2, 3} is dense in
D(R1+n), (3.4) will follow once it is proved that
(3.2) [(Nσ∗Sσ)∗(e−σN)]∗[ϕ1∗ϕ2∗ϕ3] = [Nσ∗(Sσ∗(e−σN))]∗[ϕ1∗ϕ2∗ϕ3]
for every ϕ1, ϕ2, ϕ3 ∈ D(R1+n). In order to prove (3.2), fix ϕ1, ϕ2, ϕ3 and
let
f = Nσ ∗ ϕ1,
g = Sσ ∗ ϕ2,
h = (e−σU) ∗ ϕ3.
Then f, g, h ∈ C ∞(R1+n) and using commutativity and associativity of con-
volution of distributions when all factors except at most one have compact
support, one can prove that
(3.3)
and
(3.4)
[(Nσ ∗ Sσ) ∗ (e−σU)] ∗ [ϕ1 ∗ ϕ2 ∗ ϕ3] = (f ∗ g) ∗ h
[Nσ ∗ (Sσ ∗ (e−σU))] ∗ [ϕ1 ∗ ϕ2 ∗ ϕ3] = f ∗ (g ∗ h).
Let us stress that in the proof of (3.3) and (3.4) (and in particular in the
proof that the right sides of (3.3) and (3.4) make sense) we have to make
use of the facts that Nσ ∗ Sσ = δ and Sσ ∗ (e−σU) has compact support. The
equalities (3.3) and (3.4) reduce the problem of proving (3.2) to proving the
equality
(3.5)
(f ∗ g) ∗ h = f ∗ (g ∗ h).
To do this, we need some more detailed information about f , g, h. Since
Nσ ∈ O′
(3.6)
C(R1+n), by (1.1) one has
f ∈ S(R1+n) ⊂ L1(R1+n).
10
J. Kisy´nski
Since supp Sσ = {0}, one has
(3.7)
g ∈ D(R1+n) ⊂ L1(R1+n).
Furthermore
(3.8)
h ∈ C ∞(R;S(Rn)) ⊂ C(R; L1(Rn)).
Indeed, for the proof of (3.8) it is sufficient to show that [(e−σU) ∗
ϕ3][−a,a]×Rn ∈ C ∞([−a, a];S(Rn)) for every a ∈ ]0,∞[. So, take a ∈ ]0,∞[
and b ∈ ]0,∞[ such that supp ϕ3 ⊂ [−b, b] × Rn. Take ϑ ∈ D(R) such that
ϑ = 1 on [−a − b, a + b]. Then
(3.9)
[(e−σU) ∗ ϕ3][−a,a]×Rn = [(ϑ0e−σU) ∗ ϕ3][−a,a]×Rn.
Since ϑ0e−σU ∈ O′
that (3.9) implies (3.8).
C(R1+n), by (1.1) one has (ϑe−σU) ∗ ϕ3 ∈ S(R1+n), so
Since supp Nσ, supp e−σU ⊂ H+ there is c ∈ ]0,∞[ (depending on ϕ1,
ϕ2, ϕ3, which however are fixed) such that
(3.10)
supp f, supp g, supp h ⊂ {(x0, x1, . . . , xn) ∈ R1+n : x0 ≥ −c}.
From (3.6) -- (3.8) and (3.10) it follows that (f ∗ g) ∗ h ∈ C([−3c,∞[;
L1(Rn)). Hence
ZR1+n(cid:18)ZR1+n f (v0, . . . , vn) g(u0 − v0, . . . , un − vn) dv0 . . . dvn(cid:19)
· h(x0 − u0, . . . , xn − un) du0 . . . dun < ∞
for every (x0, . . . , xn) ∈ R1+n, so that, by the Fubini -- Tonelli theorem, the
two iterated integrals corresponding to the integral
ZR1+n×R1+n
f (v0, . . . , vn)g(u0 − v0, . . . , un − vn)h(x0 − u0, . . . , xn − vn)
dv0 . . . dvn du0 . . . dun
are equal for every (x0, . . . , xn) ∈ R1+n. This means that (3.5) holds.
3.2 Uniqueness as a consequence of the associativity
relation (3.1)
The uniqueness of the fundamental solution belonging to O′
LOC(H+) for the
operator P (∂0, ∂1, . . . , ∂n) satysfying (i) is a consequence of the following
lemma.
Evolutionary PDOs and rapidly decreasing distributions
11
Lemma 3.2. Suppose that (i) holds and that F ∈ E ′(R1+n) has support
contained in H+. Then the equation
(3.11)
P (∂0, ∂1, . . . , ∂n)U = F
has exactly one solution U belonging to O′
tion and every σ ∈ ]ω0,∞[ one has
(3.12)
U = (eσNσ) ∗ F.
LOC(H+). Moreover, for this solu-
Proof. Suppose that (i) holds. Take σ ∈ ]ω0,∞[. Then, in view of (2.18) and
(2.17), N = eσNσ belongs to O′
LOC(H+) and is a fundamental solution for
P (∂0, ∂1, . . . , ∂n). It follows that U defined by (3.12) belongs to O′
LOC(H+)
and satisfies (3.11). It remains to prove that in O′
LOC(H+) there are no other
solutions of (3.11). To this end suppose that U ∈ O′
LOC(H+) and U satisfies
(3.11). Take σ ∈ ]ω0,∞[ and define Sσ by (2.14). Then, by (2.15),
Sσ ∗ (e−σU) = P (σ + ∂0, ∂1, . . . , ∂n)(e−σU) = e−σ(P (∂0, ∂1, . . . , ∂n)U) = e−σF,
whence, by (2.14), (2.6) and (3.1),
e−σU = δ ∗ (e−σU) = (Nσ ∗ Sσ) ∗ (e−σU) = Nσ ∗ (Sσ ∗ (e−σU))
= Nσ ∗ (e−σF ) = e−σ((eσNσ) ∗ F )
so that U = (eσNσ) ∗ F .
4 Proof of (ii)⇒(i)
4.1 The distributions ϑ0N (ϕ ⊗ ·)
Let N ∈ O′
LOC(H+) be a fundamental solution for P (∂0, ∂1, . . . , ∂n). Fix a, b
such that 0 < a < b < ∞, and ϑ ∈ D(R) such that ϑ = 1 on [−b, b]. For
every ϕ ∈ D(R) consider the mapping
T (ϕ) : D(Rn) ∋ φ 7→ hϑ0N, ϕ ⊗ φi ∈ C.
Then T (ϕ) ∈ D′(Rn). Since ϑ0N ∈ O′
every k ∈ N0 there is mk ∈ N0 such that
ϑ0N = Xp+α≤mk
∂p
0 ∂α1
(4.1)
1
· · · ∂αn
n Fk;p,α
C(R1+n), from (1.2) it follows that for
where every Fk;p,α is a continuous function on R1+n = {(t, x) ∈ R × Rn} for
which
sup
(t,x)∈R1+n
(1 + t + x)kFk;p,α(t, x) < ∞.
12
J. Kisy´nski
Consequently, whenever ϕ ∈ D(R), then
∂α1
(4.2)
1
· · · ∂αn
n fk;α;ϕ
where
T (ϕ) = Xα≤mk
ZR
fk;α;ϕ(x) = Xp≤mk−α
((−∂0)pϕ(t))Fk;p;α(t, x) dt.
It follows that, whenever α ≤ mk, ϕ ∈ D(R), and x ∈ Rn, then
0 ϕ(t)(1 + t + x)−k dt
fk;α;ϕ(x) ≤ Ck Xp≤mk−α
(4.3)
Zsupp ϑ ∂p
≤ Dk(1 + x)−k sup{∂p
0 ϕ(t) : p = 0, . . . , mk, t ∈ R},
where Ck, Dk ∈ ]0,∞[ depend only on k. In particular this shows that
(4.4)
T (ϕ) ∈ O′
C(Rn)
for every ϕ ∈ D(R).∗)
Since N is the fundamental solution for P (∂0, ∂1, . . . , ∂n) with support in
H+, and ϑ = 1 on [−b, b], it follows that
(4.5) T (ϕ) = 0 whenever supp ϕ ⊂ ]−∞, 0[,
(4.6) Pm
[−b,b](R) where δ
is the Dirac distribution on Rn and Qk(∂1, . . . , ∂n), k = 0, . . . , m, are
k=0 Qk(∂1, . . . , ∂n)T ((−∂0)kϕ) = ϕ(0)δ for all ϕ ∈ C ∞
PDOs on Rn such that P (∂0, ∂1, . . . , ∂n) =Pm
0 Qk(∂1, . . . , ∂n).
k=0 ∂k
In the subsequent lemmas it will be tacitly assumed that (ii) holds and
N, a, b, ϑ, T are fixed. Recall that 0 < a < b < ∞, ϑ ∈ D(R), ϑ = 1 on
[−b, b], N ∈ O′
LOC(H+) is a fundamental solution for P (∂0, ∂1, . . . , ∂n) and
C(Rn) for every ϕ ∈ D(R). For every ϕ ∈ D(R)
T (ϕ) = ϑ0N(ϕ ⊗ ·) ∈ O′
denote by bT (ϕ) the image of T (ϕ) under the Fourier transformation on Rn.
Then bT (ϕ) ∈ OM (Rn), by (4.4) and (1.4).
Lemma 4.1. There are p0, m0 ∈ N0 and C ∈ ]0,∞[ such that
bT (ϕ)(ξ) ≤ C(1 + ξ)m0 sup{∂p
for every ξ ∈ Rn and ϕ ∈ C ∞
[a,b](R).
0 ϕ(t) : p = 0, . . . , p0, a ≤ t ≤ b}
∗) After introducing the topology in O′
C(Rn), it is possible to prove that the mapping
C (Rn) is a vector-valued distribution. However this is insignificant
D(R) ∋ ϕ 7→ T (ϕ) ∈ O′
for the present proof.
Evolutionary PDOs and rapidly decreasing distributions
13
Proof. If in (4.1) we take k > n, then, by (4.2) and (4.3),
T (ϕ) = Xα≤mk
∂α1
1
· · · ∂αn
n fk;α;ϕ
for every ϕ ∈ C ∞
[a,b](R)
where
kfk;α;ϕkL1(Rn) ≤ D sup{∂p
0 ϕ(t) : p = 0, . . . , mk, a ≤ t ≤ b}
for every α with α ≤ mk and every ϕ ∈ C ∞
ing only on k. Consequently, whenever ϕ ∈ C ∞
bT (ϕ)(ξ) ≤ (1 + ξ)mkgϕ(ξ)Mm×m
where gϕ ∈ Cb(Rn) and
[a,b](R), with D ∈ ]0,∞[ depend-
[a,b](R), then
for every ξ ∈ Rn
ξ∈Rn gϕ(ξ) ≤ C sup{∂p
sup
0 ϕ(t) : p = 0, . . . , mk, a ≤ t ≤ b}
for some C ∈ ]0,∞[ depending only on k.
4.2 An inequality of Chazarain type
Lemma 4.2. Suppose that P (∂0, ∂1, . . . , ∂n) is a PDO on R1+n with con-
stant coefficients for which there is a fundamental solution belonging to
O′
LOC(H+). Then there are a, b ∈ ]0,∞[ such that whenever (λ, ξ) ∈ C× Rn
and
Re λ > a + b log(1 + λ + ξ) ∗),
then P (λ, iξ) 6= 0.
Proof. From (4.6) it follows that
(4.7)
mXk=0
Qk(iξ)bT ((−∂0)kϕ) = ϕ(0)
[−b,b](R) and ξ ∈ Rn. Take ϕ0 ∈ C ∞
for every ϕ ∈ C ∞
[−b,b](R) such that ϕ0 = 1
on [−a, a]. Following J. Chazarain [Cha], pp. 394 -- 395, consider functions
∗) This inequality and its proof are similar to the inequality (1.2) on p. 394 of [Cha]
and the argument presented on p. 395 of [Cha]. There is however an important difference.
In [Cha] the inequality (1.2) does not involve ξ and determines the "logarithmic region"
Λ ⊂ C such that for every λ ∈ Λ an abstract operator Q(λ) = λmAm + ··· + λA1 + A0 is
invertible. In our case the inequality involves ξ but the operator Q(λ) is replaced by the
polynomial P (λ, iξ), and Lemma 4.2 is not the final step of the argument.
14
J. Kisy´nski
of the form ϕ = e−λϕ0 where λ ranges over C. Since T (ϕ) = 0 whenever
supp ϕ ⊂ ]−∞, 0[, by (4.7) and the Leibniz formula one has
(4.8)
P (λ, iξ)[bT (e−λϕ0)](ξ) =(cid:16) mXk=0
λkQk(iξ)(cid:17)[bT (e−λϕ0)](ξ)
mXk=0
Qk(iξ)[bT (ψk,λ)](ξ)
= 1 −
where ψk,λ ∈ C ∞
[a,b](R) is determined by the equality
ψk,λ(t) = (−1)k
j(cid:19)∂j
(cid:18)k
0ϕ0(t)(−λ)k−je−λt
kXj=1
for t ∈ [a, b].
By Lemma 4.1 there are C, K ∈ ]0,∞[ such that, if Re λ ≥ 0, then
(4.9)
[bT (ψk,λ)](ξ)
≤ C(1 + ξ)m0 sup{∂p
≤ C(1 + ξ)m0K(1 + λ)m−1+p0e−a Re λ
0 ψk,λ(t) : p = 0, . . . , p0, a ≤ t ≤ b}
for every k = 1, . . . , m and ξ ∈ Rn. Furthermore, there are l ∈ N and
L ∈ ]0,∞[ such that
mXk=0
Qk(iξ) ≤ L(1 + ξ)l
for every ξ ∈ Rn.
(4.10)
CKL(1 + λ + ξ)µe−a Re λ < 1,
Let µ = m0+l+m−1+p0. From (4.8) -- (4.10) it follows that if (λ, ξ) ∈ C×Rn,
Re λ ≥ 0, and
then Pm
(λ, ξ) ∈ C × Rn and
k=0 Qk(iξ)[bT (ψk,λ)](ξ) < 1, and hence P (λ, iξ) 6= 0. Therefore, if
Re λ > a−1 log(CKL + 1) + a−1µ log(1 + λ + ξ),
then P (λ, iξ) 6= 0.
4.3 The Chazarain type inequality implies (i)
The implication (ii)⇒(i) is an immediate consequence of Lemma 4.2 and
the following
Evolutionary PDOs and rapidly decreasing distributions
15
Lemma 4.3. Let Q be a polynomial of 1 + n variables with complex coeffi-
cients. Suppose that there are a ∈ R and b ∈ ]0,∞[ such that
Re λ ≤ a + b log(1 + λ + ξ)
whenever (λ, ξ) ∈ C × Rn and Q(λ, ξ) = 0.
(4.11)
Then
sup{Re λ : (λ, ξ) ∈ C × Rn, Q(λ, ξ) = 0} < ∞.
The proof follows the scheme due to L. Garding and L. Hormander. Let
σ(r) = sup{Re λ : λ ∈ C and there is ξ ∈ Rn such that
λ2 + ξ2 ≤ 1
2 r2 and Q(λ, ξ) = 0}.
Then, by (4.11),
(4.12)
σ(r) ≤ a + b log(1 + r)
for every r ∈ [0,∞[.
Following an idea of L. Hormander (presented in the Appendix to [H2]),
the Tarski -- Seidenberg theorem is used to show that there is a polynomial
V (z, w) (not vanishing identically) of two variables such that V (r, σ(r)) = 0
for every r ∈ [0,∞[. Then, as in L. Garding's proof of the Lemma on p. 11
of [G], the Puiseux expansions of the w-roots of V (z, w) for large z show
that (4.12) is possible only if sup{σ(r) : r ∈ [0,∞[} < ∞.
References
[Cha] J. Chazarain, Probl`emes de Cauchy abstraits et applications `a
quelques probl`emes mixtes, J. Funct. Anal. 7 (1971), 386 -- 446.
[Che] C. Chevalley, Theory of Distributions, lectures given at Columbia Uni-
versity, 1950 -- 1951. Notes prepared by K. Nomizu (mimeographed).
[G]
L. Garding, Linear hyperbolic partial differential equations with con-
stant coefficients, Acta Math. 85 (1951), 1 -- 62.
[Go] E. A. Gorin, Asymptotic properties of polynomials and algebraic func-
tions of several variables, Uspekhi Mat. Nauk 16 (1961), no. 1, 91 -- 118
(in Russian).
[H1] L. Hormander, On the characteristic Cauchy problem, Ann. of Math.
88 (1968), 341 -- 370.
16
J. Kisy´nski
[H2] L. Hormander, The Analysis of Linear Partial Differential Operators
II. Differential Operators with Constant Coefficients, Springer, 1983.
[K]
[P]
J. Kisy´nski, Equicontinuity and convergent sequences in the spaces
O′
C and OM , Bull. Polish Acad. Sci. Math., to appear.
I. G. Petrovskiı, Uber das Cauchysche Problem fur ein System lin-
earer partieller Differentialgleichungen im Gebiete der nichtanalyti-
schen Funktionen, Bulletin de l'Universit´e d'Etat de Moscou 1 (1938),
no. 7, 1 -- 74.
[R]
J. Rauch, Partial Differential Equations, Springer, 1991.
[S1] L. Schwartz, Les ´equations d'´evolution li´ees au produit de composi-
tion, Ann. Inst. Fourier (Grenoble) 2 (1950), 19 -- 49.
[S2] L. Schwartz, Th´eorie des Distributions, nouvelle ´ed., Hermann, Paris,
1966.
[T]
F. Treves, Lectures on Linear Partial Differential Equations with Con-
stant Coefficients, Inst. Mat. Pura Apl., Rio de Janeiro, 1961.
Evolutionary PDOs and rapidly decreasing distributions
17
Appendix
Proof of Lemma 4.3
Let R(σ, τ, ξ) and S(σ, τ, ξ) be real polynomials on R2+n such that
R(σ, τ, ξ) + iS(σ, τ, ξ) = Q(σ + iτ, ξ).
Then
E = {(r, σ, τ, ξ) ∈ R3+n : r ≥ 0, σ2 + τ 2 + ξ2 ≤ 1
2 r2,
R(σ, τ, ξ) = 0, S(σ, τ, ξ) = 0}
is a semi-algebraic subset of R3+n and, by the Tarski -- Seidenberg theorem
(see Appendix to [H2]) its projection on R2 defined by
F = {(r, σ) ∈ R2 : ∃τ,ξ (r, σ, τ, ξ) ∈ E}
is a semi-algebraic subset of R2. If σ(r) is defined as in Sec. 4.3, then for
every r ∈ [0,∞[ one has
(A.1)
σ(r) = sup{σ : (r, σ) ∈ F}.
Since F is semi-algebraic, it may be represented in the form
(A.2)
where
(A.3)
F =
k[i=1
Fi ∩ Gi,1 ∩ · · · ∩ Gi,j(i)
Fi = {x ∈ R2 : Pi(x) = 0}, Gi,j = {x ∈ R2 : Qi,j(x) > 0},
Pi and Qi,j being real polynomials on R2. It is not excluded that some Pi are
identically zero and some Qi,j are strictly positive on the whole R2. From
(A.1) it follows that whenever r ∈ [0,∞[ is fixed, there is i(r) ∈ {1, . . . , k}
such that
(A.4)
σ(r) = sup{σ : (r, σ) ∈ Fi(r) ∩ Gi(r),1 ∩ · · · ∩ Gi(r),j(i(r))}.
By (4.12), for every r ∈ [0,∞[ one has σ(r) < ∞, so that there is a bounded
sequence (σν(r))∞
ν=1 such that
(A.5) (r, σν(r)) ∈ Fi(r) ∩ Gi(r),1 ∩ · · · ∩ Gi(r),j(i(r))
for every ν = 1, 2, . . .
18
and
(A.6)
lim
ν→∞
σν(r) = σ(r).
J. Kisy´nski
If Pi(r) 6≡ 0, then (A.5) and (A.6) imply that Pi(r)(r, σ(r)) = 0. If Pi(r) ≡ 0,
then, again by (A.5) and (A.6), for some j0 ∈ {1, . . . , j(i(r))} one has
Qi(r),j0 6≡ 0 and Qi(r),j0(r, σ(r)) = 0, because otherwise Fi(r) = R2 and there
would be ε > 0 such that Qi(r),j(r, σ(r)+ε) > 0 for every j ∈ {1, . . . , j(i(r))}
contrary to (A.4). Consequently, whenever r ∈ [0,∞[, then either Wr ≡ Pi(r)
or Wr ≡ Qi(r),j0 is a real polynomial on R2 such that
Wr 6≡ 0 and Wr(r, σ(r)) = 0.
Therefore if V is equal to the product of all those polynomials Pi and Qi,j,
that occur in (A.3) and do not vanish identically on R2, then
(A.7)
V 6≡ 0 and V (r, σ(r)) = 0
for every r ∈ [0,∞[.
Now we are going to show that (4.12) and (A.7) imply sup{σ(r) : r ∈
[0,∞[} < ∞. To this end we consider V as a polynomial V (z, w) of two com-
plex veriables, and, following L. Garding [G, proof of the Lemma on p. 11],
we use the Puiseux expansions of the w-roots of V (z, w). Concerning these
expansions we will give exact references to [S-Z]. Consider the factorization
V (z, w) = V1(z, w) · V2(z, w) · . . . · Vl(z, w),
z ∈ C \
l[k=1
Sk,
x ∈ C,
where
(i) every Vk, k = 1, . . . , l, belongs to the ring K(z)[w] of polynomials of
w over the field K(z) of rational functions of z, so that
Vk(z, w) =
dkXj=0
Ak,j(z)wj
for every z ∈ C \ Sk and x ∈ C
where Ak,j ∈ K(z) for j = 0, . . . , dk, Ak,dk 6≡ 0, and the finite set Sk
consists of those points of C at which some Ak,j, j = 0, . . . , dk, has a
pole,
(ii) every Vk, k = 1, . . . , l, is an irreducible element of K(z)[w].
The assumption that Ak,dk 6≡ 0 implies that all the sets
Nk = {z ∈ C \ Sk : Ak,dk(z) = 0},
k = 1, . . . , l,
Evolutionary PDOs and rapidly decreasing distributions
19
are finite. Define
Mk = {z ∈ C \ (Sk ∪ Nk) : not all the w-roots of Vk(z, w) are simple},
Nk = {(z, w) ∈ (C \ (Sk ∪ Nk ∪ Mk)) × C : Vk(z, w) = 0}.
From Theorems VI.13.7, VI.14.2 and VI.14.3 of [S-Z] it follows that
(a) for every k = 1, . . . , l the set Mk is finite and
Nk ∩ [(C \ (Sk ∪ Nk ∪ Mk)) × C]
is equal to the graph of a dk-variate function Rk analytic on the set
C \ (Sk ∪ Nk ∪ Mk),
(b) there is R ∈ ]0,∞[ such that for every k = 1, . . . , l one has {z ∈ C :
R < z < ∞} ⊂ C \ (Sk ∪ Nk ∪ Mk), and if z ∈ C and R < z < ∞,
then
Rk(z) = {φk(ζ) : ζ ∈ C, 0 < ζ < R−1/dk , ζ dk = z−1}
where φk is a function of one complex variable holomorphic in the
annulus {ζ ∈ C : 0 < ζ < R−1/dk},
(c) every φk, k = 1, . . . , l, has at zero either a removable singularity or a
pole.
Consequently, for every k = 1, . . . , l one has
(A.8) Nk ∩ ({z ∈ C : z > R} × C)
p=pk
=n(cid:16)z,
∞Xp=pk
ak,pζ p(cid:17) : (z, ζ) ∈ C2, z > R, ζ dk = z−1o
where P∞
ak,pζ p is the Laurent expansion of φk in the annulus {ζ ∈
C : 0 < ζ < R−1/dk}. We assume that either ak,pk 6= 0 or 0 = ak,pk =
ak,pk+1 = · · · . The equality (A.8) is nothing but the exact form of the
Puiseux series expansion of Rk(z) for z → ∞. It follows that if r ∈ ]R,∞[,
then (r, σ(r)) ⊂Sl
k=1 Nk and σ(r) is equal to one of the numbers
ak,p(cid:18) ei2πd/dk
k√r (cid:19)p
∞Xp=pk
k = 1, . . . , l, d = 1, . . . , dk,
σk,d(r) =
d
,
where dk√rk is the positive dk-th root of r and the series is absolutely con-
vergent, so that σk,d(r) = ck,dr−pk/dk(1 + o(1)) as r → ∞ where ck,d =
ak,pkei2πdpk/dk . If for some k = 1, . . . , l and d = 1, . . . , dk the set
{r ∈ ]R,∞[ : σ(r) = σk,d(r)}
20
J. Kisy´nski
is unbounded, then ck,d must be real, and, by the estimation (4.12) of σ(r),
either ck,d ≤ 0, or ck,d > 0 and pk ≥ 0. In both cases sup{σk,d(r) : r ∈
]R,∞[} < ∞. This implies that sup{σ(r) : r ∈ [0,∞[} < ∞, completing
the proof.
References
[S-Z] S. Saks and A. Zygmund, Analytic Functions, 3rd ed., PWN,
Warszawa, 1959 (in Polish); English transl.: PWN, 1965; French
transl.: Masson, 1970.
|
1301.3447 | 1 | 1301 | 2013-01-15T18:41:21 | Hermite-Hadamard type inequalities via preinvexity and prequasiinvexity | [
"math.FA"
] | In this paper, we obtain some Hermite-Hadamard type inequalities for functions whose third derivatives in absolute value are preinvex and prequasiinvex. | math.FA | math |
HERMITE-HADAMARD TYPE INEQUALITIES VIA
PREINVEXITY AND PREQUASIINVEXITY
M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC∗,♦
Abstract. In this paper, we obtain some Hermite-Hadamard type inequali-
ties for functions whose third derivatives in absolute value are preinvex and
prequasiinvex.
1. INTRODUCTION
Several researchers have been studied on convexity and a lot of papers have been
written on this topic which give new generalizations, extensions and applications.
A huge amount of these studies on refinements of celebrated Hermite-Hadamard
Inequality for convex functions. Invex functions introduced by Hanson as a gen-
eralization of convex functions in [2]. Some properties of preinvex functions have
been discussed in the papers [3]-[9]. It is well-known that there are many applica-
tions of invexity in nonlinear optimization, variational inequalities and in the other
branches of pure applied sciences.
Now it is time to give the following definitions and results which will be used in
this paper (see [5], [6] and [7]):
Let K be a nonempty closed set in Rn. We denote by h., .i and k.k the inner
product and norm respectively. Let f : K → R and η : K × K → R be continuous
functions.
Definition 1. (See [6]) Let u ∈ K. Then the set K is said to be invex at u with
respect to η (., .) , if
u + tη(v, u) ∈ K,
∀u, v ∈ K,
t ∈ [0, 1].
K is said to be invex set with respect to η, if K is invex at each u ∈ K. The invex
set K is also called a η−connected set.
Remark 1. (See [5]) We would like to mention that the Definition 1 of an invex
set has a clear geometric interpretation. This definition essentially says that there
is a path starting from a point u which is contained in K. We don't require that
the point v should be one of the end points of the path. This observation plays
an important role in our analysis . Note that, if we demand that v should be an
end point of the path for every pair of points, u, v ∈ K, then η(v, u) = v − u and
consequently invexity reduces to convexity. Thus, it is true that every convex set is
also an invex set with respect to η(v, u) = v − u, but the converse is not necessarily
true.
Key words and phrases. Hermite-Hadamard inequality, preinvex functions, prequasiinvex func-
tions, power-mean inequality, Holder inequality.
♦Corresponding Author.
1
2
M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC∗,♦
Definition 2. (See [6]) The function f on the invex set K is said to be preinvex
with respect to η, if
f (u + tη(v, u)) ≤ (1 − t)f (u) + tf (v),
∀u, v ∈ K,
t ∈ [0, 1].
The function f is said to be preconcave if and only if −f is preinvex. Note that
every convex function is a preinvex function, but the converse is not true. For
example, the function f (u) = − u is not a convex function, but it is a preinvex
function with respect to η, where
η(v, u) =(cid:26)
v − u,
u − v,
if v ≤ 0, u ≤ 0
and
v ≥ 0, u ≥ 0
otherwise
.
Definition 3. (See [4]) The function f on the invex set K is said to be prequasiinvex
with respect to η, if
f (u + tη(v, u)) ≤ max{f (u), f (v)} ,
∀u, v ∈ K,
t ∈ [0, 1].
The following inequality is well known in the literature as the Hermite-Hadamard
integral inequality:
f(cid:18) a + b
2 (cid:19) ≤
1
b − aZ b
a
f (x)dx ≤
f (a) + f (b)
2
where f : I ⊆ R → R is a convex function on the interval I of real numbers and
a, b ∈ I with a < b.
Lemma 1. (See [1]) Let f : I ⊆ R → R be a three times differentiable function on
I ◦ with a, b ∈ I, a < b. If f ′′′ ∈ L[a, b], then
b − aZ b
t (1 − t) (2t − 1) f ′′′ (ta + (1 − t)b) dt.
2
(b − a)3
[f ′(b) − f ′(a)]
f (x)dx −
b − a
12
f (a) + f (b)
−
12
=
1
a
Z 1
0
The main purpose of this paper is to prove some new inequalities of Hermite-
Hadamard type for preinvex and prequasiinvex functions by using a new version of
Lemma 1.
2. Inequalities for preinvex functions
To prove our main results we need the following equality which is a generalization
of Lemma 1 to invex sets:
Lemma 2. Let A ⊆ R be an open invex subset with respect to η : A × A → R and
a, b ∈ A with η (b, a) 6= 0. Suppose that f : A → R is a three times differentiable
function. If f ′′′ is integrable on the η−path Pbc, c = b + η (a, b) , then the following
equality holds:
(2.1)
Z b+η(a,b)
(η (a, b))4
b
12
=
f (x)dx − η (a, b)
Z 1
0
t (1 − t) (2t − 1) f ′′′(b + tη (a, b))dt.
f (b) + f (b + η (a, b))
2
(η (a, b))2
12
−
[f ′(b) − f ′(b + η (a, b))]
Proof. The proof of Lemma 2 is left to the reader.
(cid:3)
HERMITE-HADAMARD TYPE INEQUALITIES VIA PREINVEXITY AND PREQUASIINVEXITY3
Theorem 1. Let A ⊆ R be an open invex subset with respect to η : A × A → R.
Suppose that f : A → R is a three times differentiable function. If f ′′′q is preinvex
on A , then the following inequality holds:
Z b+η(a,b)
b
(η (a, b))4
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
192
≤
f (b) + f (b + η (a, b))
f (x)dx − η (a, b)
(cid:20)f ′′′(a)q + f ′′′(b)q
2
2
1
q
(cid:21)
(η (a, b))2
12
−
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b
2
=
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(η (a, b))2
(η (a, b))4
f (b) + f (b + η (a, b))
Z b+η(a,b)
for q ≥ 1 and every a, b ∈ A with η (b, a) 6= 0.
Proof. Since a, b ∈ A and A is an invex set with respect to η, it is obvious that
b + tη (a, b) ∈ A for t ∈ [0, 1] . By using Lemma 2, the power-mean inequality and
preinvexity of f ′′′q, we can write
f (x)dx − η (a, b)
Z 1
t (1 − t)2t − 1f ′′′(b + tη (a, b))
q (cid:18)Z 1
(cid:18)Z 1
t (1 − t)2t − 1 dt(cid:19)
q (cid:18)Z 1
(cid:18)Z 1
t (1 − t)2t − 1 dt(cid:19)
(cid:20)f ′′′(a)q + f ′′′(b)q
(cid:21)
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dt(cid:19)
t (1 − t)2t − 1f ′′′(b + tη (a, b))q
t (1 − t)2t − 1(cid:2)tf ′′′(a)q + (1 − t)f ′′′(b)q(cid:3) dt(cid:19)
(η (a, b))4
(η (a, b))4
(η (a, b))4
192
≤
≤
−
0
0
0
0
1
1−
1
1−
12
12
12
12
=
2
1
q
1
q
0
where we use the fact that
Z 1
0
t (1 − t)2t − 1 dt =
1
16
and
Z 1
0
t2 (1 − t)2t − 1 dt =Z 1
0
t (1 − t)2 2t − 1 dt =
1
32
.
(cid:3)
The proof is completed.
1
q
Corollary 1. In Theorem 1, if we choose q = 1, we obtain
Z b+η(a,b)
b
(η (a, b))4
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
384
f (x)dx − η (a, b)
f (b) + f (b + η (a, b))
2
(η (a, b))2
12
−
[f ′′′(a) + f ′′′(b)] .
≤
Corollary 2. In Theorem 1, if we take f ′(b) = f ′(b + η (a, b)), then we have
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z b+η(a,b)
(η (a, b))4
384
≤
f (b) + f (b + η (a, b))
f (x)dx − η (a, b)
(cid:2)f ′′′(a)q + f ′′′(b)q(cid:3)
2
1
q .
4
M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC∗,♦
Theorem 2. Let A ⊆ R be an open invex subset with respect to η : A × A → R.
Suppose that f : A → R is a differentiable function. If f ′′q is preinvex on A, then
the following inequality holds:
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b
Z b+η(a,b)
f (x)dx − η (a, b)
(p + 1) (p + 3)(cid:19)
q (cid:18)
(η (b, a))4
24 × 6
1
1
1
p
f (b) + f (b + η (a, b))
(η (a, b))2
2
−
(cid:2)f ′′′(a)q + f ′′′(b)q(cid:3)
12
1
q
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
for every a, b ∈ A with η (b, a) 6= 0 where q > 1, p−1 + q−1 = 1.
Proof. By using Lemma 2, Holder inequality and preinvexity of f ′′′q, we can write
f (b) + f (b + η (a, b))
Z b+η(a,b)
b
(η (b, a))4
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
12
(η (b, a))4
12
f (x)dx − η (a, b)
(cid:18)Z 1
(cid:18)Z 1
t (1 − t)2t − 1p dt(cid:19)
t (1 − t)2t − 1p dt(cid:19)
0
0
2
1
1
0
p (cid:18)Z 1
p (cid:18)Z 1
0
Computing the above integrals, we deduce
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b
Z b+η(a,b)
f (x)dx − η (a, b)
q (cid:18)
(p + 1) (p + 3)(cid:19)
(η (b, a))4
24 × 6
1
1
1
p
f (b) + f (b + η (a, b))
2
−
(cid:2)f ′′′(a)q + f ′′′(b)q(cid:3)
which completes the proof.
Theorem 3. Under the assumptions of Theorem 2, we have
q
1
12
−
dt(cid:19)
(η (a, b))2
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
t (1 − t)f ′′(b + tη (a, b))q
t (1 − t)(cid:2)tf ′′′(a)q + (1 − t)f ′′′(b)q(cid:3) dt(cid:19)
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(η (a, b))2
12
q ,
1
1
q
.
≤
≤
≤
≤
≤
≤
≤
=
(cid:3)
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dt(cid:19)
1
q
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z b+η(a,b)
b
(η (b, a))4
96
Z b+η(a,b)
b
(η (b, a))4
12
(η (b, a))4
12
(η (b, a))4
12
f (b) + f (b + η (a, b))
(η (a, b))2
12
−
f (x)dx − η (a, b)
p Γ (1 + p)
Γ(cid:0) 3
2 + p(cid:1)
(cid:0)√π(cid:1)
1
2
1
q
1
!
p (cid:18) 1
q + 1(cid:19)
(cid:2)f ′′′(a)q + f ′′′(b)q(cid:3)
q .
Proof. From Lemma 2, using Holder inequality and preinvexity of f ′′q , we have
1
f (x)dx − η (a, b)
(cid:18)Z 1
dt(cid:19)
0 (cid:0)t − t2(cid:1)p
(cid:18)Z 1
dt(cid:19)
0 (cid:0)t − t2(cid:1)p
(√π)
p × 4 Γ (1 + p)
Γ(cid:0) 3
2 + p(cid:1)
2
p
1
1
f (b) + f (b + η (a, b))
2
(η (a, b))2
−
12
1
p (cid:18)Z 1
0 2t − 1q f ′′′(b + tη (a, b))q
p (cid:18)Z 1
0 2t − 1q(cid:2)tf ′′(a)q + (1 − t)f ′′(b)q(cid:3) dt(cid:19)
!
p (cid:20)f ′′′(a)q + f ′′′(b)q
2 (q + 1)
(cid:21)
1
1
q
1
q
HERMITE-HADAMARD TYPE INEQUALITIES VIA PREINVEXITY AND PREQUASIINVEXITY5
where we used the fact that
dt =
Z 1
0 (cid:0)t − t2(cid:1)p
t2t − 1q dt =Z 1
0
2−1−2p√πΓ (1 + p)
Γ(cid:0) 3
2 + p(cid:1)
(1 − t)2t − 1q dt =
1
2 (q + 1)
.
and
Z 1
0
The proof is completed.
(cid:3)
3. Inequalities for prequasiinvex functions
In this section, we obtain Hermite-Hadamard type inequalities for prequasiinvex
functions via Lemma 2.
Theorem 4. Let A ⊆ R be an open invex subset with respect to η : A × A →
R. Suppose that f : A → R is a three times differentiable function. If f ′′′q is
prequasiinvex on A , then the following inequality holds:
f (b) + f (b + η (a, b))
2
(η (a, b))2
12
−
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z b+η(a,b)
b
(η (a, b))4
f (x)dx − η (a, b)
(cid:2)max(cid:8)f ′′′(a)q
1
≤
192
,f ′′′(b)q(cid:9)(cid:3)
for q ≥ 1 and every a, b ∈ A with η (b, a) 6= 0.
Proof. By using Lemma 2, the power-mean inequality and prequasiinvexity of f ′′′q,
we can write
q
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z b+η(a,b)
b
(η (a, b))4
12
(η (a, b))4
12
(η (a, b))4
12
f (b) + f (b + η (a, b))
1
0
2
1−
f (x)dx − η (a, b)
q (cid:18)Z 1
(cid:18)Z 1
t (1 − t)2t − 1 dt(cid:19)
q (cid:18)Z 1
(cid:18)Z 1
t (1 − t)2t − 1 dt(cid:19)
q " max(cid:8)f ′′′(a)q ,f ′′′(b)q(cid:9)
16(cid:19)1−
(cid:18) 1
16
1−
0
0
0
1
1
1
(η (a, b))2
12
−
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dt(cid:19)
,f ′′′(b)q(cid:9)(cid:3) dt(cid:19)
t (1 − t)2t − 1f ′′′(b + tη (a, b))q
t (1 − t)2t − 1(cid:2)max(cid:8)f ′′′(a)q
#
.
1
q
q
1
q
≤
≤
=
The proof is completed.
(cid:3)
Corollary 3. In Theorem 4, if we choose q = 1, we obtain
Z b+η(a,b)
b
(η (a, b))4
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
192
f (x)dx − η (a, b)
f (b) + f (b + η (a, b))
2
(η (a, b))2
12
−
[max{f ′′′(a) ,f ′′′(b)}] .
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤
6
M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC∗,♦
Corollary 4. In Theorem 4, if we take f ′(b) = f ′(b + η (a, b)), then we have
b
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z b+η(a,b)
(η (a, b))4
192
≤
f (b) + f (b + η (a, b))
2
f (x)dx − η (a, b)
(cid:2)max(cid:8)f ′′′(a)q
1
q .
,f ′′′(b)q(cid:9)(cid:3)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Theorem 5. Let A ⊆ R be an open invex subset with respect to η : A × A → R.
Suppose that f : A → R is a differentiable function. If f ′′q is preinvex on A, then
the following inequality holds:
b
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(η (a, b))2
f (b) + f (b + η (a, b))
Z b+η(a,b)
f (x)dx − η (a, b)
q (cid:18)
(p + 1) (p + 3)(cid:19)
−
,f ′′′(b)q(cid:9)(cid:3)
for every a, b ∈ A with η (b, a) 6= 0 where q > 1, p−1 + q−1 = 1.
Proof. By using Lemma 2, Holder inequality and preinvexity of f ′′′q, we can write
(cid:2)max(cid:8)f ′′′(a)q
(η (b, a))4
24 × 3
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤
12
2
1
p
1
1
1
q
f (b) + f (b + η (a, b))
2
(η (a, b))2
12
−
Z b+η(a,b)
b
(η (b, a))4
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
12
(η (b, a))4
12
≤
≤
f (x)dx − η (a, b)
(cid:18)Z 1
(cid:18)Z 1
t (1 − t)2t − 1p dt(cid:19)
t (1 − t)2t − 1p dt(cid:19)
0
0
1
1
0
p (cid:18)Z 1
p (cid:18)Z 1
0
q
1
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dt(cid:19)
,f ′′′(b)q(cid:9)(cid:3) dt(cid:19)
1
q
.
(cid:3)
t (1 − t)f ′′(b + tη (a, b))q
t (1 − t)(cid:2)max(cid:8)f ′′′(a)q
Computing the above integrals, we obtain the desired result.
Theorem 6. Under the assumptions of Theorem 5, we have
f (b) + f (b + η (a, b))
(η (a, b))2
−
12
f (x)dx − η (a, b)
p Γ (1 + p)
Γ(cid:0) 3
2 + p(cid:1)
(cid:0)√π(cid:1)
1
2
1
q
1
!
p (cid:18) 1
q + 1(cid:19)
(cid:2)max(cid:8)f ′′′(a)q
Proof. From Lemma 2, using Holder inequality and preinvexity of f ′′q , we have
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z b+η(a,b)
b
(η (b, a))4
48
Z b+η(a,b)
b
(η (b, a))4
12
(η (b, a))4
12
≤
≤
≤
f (x)dx − η (a, b)
(cid:18)Z 1
dt(cid:19)
0 (cid:0)t − t2(cid:1)p
(cid:18)Z 1
dt(cid:19)
0 (cid:0)t − t2(cid:1)p
f (b) + f (b + η (a, b))
2
(η (a, b))2
−
12
1
1
p (cid:18)Z 1
p (cid:18)Z 1
0 2t − 1q f ′′′(b + tη (a, b))q
0 2t − 1q(cid:2)max(cid:8)f ′′′(a)q
Computing the above integrals, we obtain the desired result.
1
q .
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
,f ′′′(b)q(cid:9)(cid:3)
[f ′(b) − f ′(b + η (a, b))](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
dt(cid:19)
,f ′′′(b)q(cid:9)(cid:3) dt(cid:19)
.
1
1
q
q
(cid:3)
HERMITE-HADAMARD TYPE INEQUALITIES VIA PREINVEXITY AND PREQUASIINVEXITY7
References
[1] L. Chun and F. Qi, Integral inequalities for Hermite-Hadamard type for functions whose 3rd
derivatives are s−convex, Applied Mathematics, 3 (2012), 1680-1885.
[2] M.A. Hanson, On sufficiency of the Kuhn-Tucker conditions, J. Math. Anal. Appl. , 80 (1981)
545-550.
[3] R. Pini, Invexity and generalized Convexity, Optimization 22 (1991) 513-525.
[4] S. R. Mohan and S. K. Neogy, On invex sets and preinvex function, J. Math. Anal. Appl. 189
(1995) 901-908.
[5] T. Antczak, Mean value in invexity analysis, Nonlinear Analysis 60 (2005) 1471-1484.
[6] T. Weir, and B. Mond, Preinvex Functions in multiple Objective Optimization, Journal of
Mathematical Analysis and Applications, 136 (1998) 29-38.
[7] X. M. Yang and D. Li, On properties of preinvex functions, J. Math. Anal. Appl. 256 (2001)
229-241.
[8] X. M. Yang, X. Q. Yang, K.L. Teo, Generalized invexity and generalized invariant monotonoc-
ity, J. Optim. Theo. Appl. 117 (2003) 607-625.
[9] M.A. Noor, Some new classes of nonconvex functions, Nonl. Funct. Anal. Appl. 12(2006),
165 -- 171.
(cid:7)ATATURK UNIVERSITY, K.K. EDUCATION FACULTY, DEPARTMENT OF MATH-
EMATICS, 25240 CAMPUS, ERZURUM, TURKEY
E-mail address: [email protected]
∗ADIYAMAN UNIVERSITY, FACULTY OF SCIENCE AND ART, DEPARTMENT OF
MATHEMATICS, 02040, ADIYAMAN, TURKEY
E-mail address: [email protected]
|
1110.6766 | 2 | 1110 | 2011-12-12T11:05:05 | Duality and Distance Formulas in Spaces Defined by Means of Oscillation | [
"math.FA",
"math.CV"
] | For the classical space of functions with bounded mean oscillation, it is well known that VMO** = BMO and there are many characterizations of the distance from a function f in BMO to VMO. When considering the Bloch space, results in the same vein are available with respect to the little Bloch space. In this paper such duality results and distance formulas are obtained by pure functional analysis. Applications include general M\"obius invariant spaces such as Q_K-spaces, weighted spaces, Lipschitz-H\"older spaces and rectangular BMO of several variables. | math.FA | math |
DUALITY AND DISTANCE FORMULAS IN SPACES
DEFINED BY MEANS OF OSCILLATION
KARL-MIKAEL PERFEKT
Abstract. For the classical space of functions with bounded mean
oscillation, it is well known that V M O∗∗ = BM O and there are
many characterizations of the distance from a function f in BM O
to V M O. When considering the Bloch space, results in the same
vein are available with respect to the little Bloch space. In this pa-
per such duality results and distance formulas are obtained by pure
functional analysis. Applications include general Mobius invari-
ant spaces such as QK-spaces, weighted spaces, Lipschitz-Holder
spaces and rectangular BM O of several variables.
1. Introduction
It is well known that the bidual of V MO is BMO, that is, the second
dual of the space of functions on the unit circle T (or the line R) with
vanishing mean oscillation can be naturally represented as the space
of functions with bounded mean oscillation. The same holds true for
the respective subspaces V MOA and BMOA of those functions in
V MO or BMO whose harmonic extensions are analytic, and there has
been considerable interest in estimating the distance from a function
f ∈ BMOA to V MOA, starting with Axler and Shapiro [5], continuing
with Carmona and Cuf´ı [8] and Stegenga and Stephenson [26]. For
the Bloch space B and little Bloch space B0 the situation is similar.
B∗∗
0 = B and the distance from f ∈ B to B0 has been characterized by
Attele [4] and Tjani [27].
Numerous people have explored the validity of the biduality Hv0(Ω)∗∗ =
Hv(Ω), where Hv(Ω) is a weighted space, consisting of analytic func-
tions bounded under the weighted supremum norm given by a weight
function v on Ω ⊂ C, and Hv0(Ω) denotes the corresponding little
space (see Example 4.4 for details). For example, Rubel and Shields
[24] considered the problem when Ω = D is the unit disc and v is radial.
Later, Anderson and Duncan [2] addressed the problem for Ω = C, and
Date: July 1, 2018.
1
2
KARL-MIKAEL PERFEKT
Bierstedt and Summers [6] went on to give a characterization of the
weights v for which the biduality holds for general Ω.
The purpose of this paper is to obtain such duality results and distance
formulas in a very general setting. Working with a Banach space M
defined by a big-O condition and a corresponding "little space" M0,
we will under mild assumptions show that M ∗∗
0 = M and prove an
isometric formula for the distance from f ∈ M to M0 in terms of
the defining condition for M0. Using a theorem of Godefroy [18], we
will additionally obtain as a corollary that M ∗
0 is the unique isometric
predual of M. When M = B this specializes to a result of Nara
[23]. The methods involved are purely operator-theoretic, appealing
to embeddings into spaces of continuous vector-valued functions rather
than analyticity, invariance properties or geometry.
The power of these general results is illustrated in the final section.
Many examples will be given there, where the main theorems are ap-
plied to general Mobius invariant spaces of analytic functions including
a large class of so-called QK-spaces, weighted spaces, rectangular BMO
of several variables and Lipschitz-Holder spaces.
A brief outline of our approach is given below. Fixing the notation for
this, and for the rest of this paper, X and Y will be two Banach spaces,
with X separable and reflexive. L will be a given collection of bounded
operators L : X → Y that is accompanied by a σ-compact locally
compact Hausdorff topology τ such that for every x ∈ X, the map
Tx : L → Y given by TxL = Lx is continuous. Here Y is considered
with its norm topology. Note that we impose no particular algebraic
structure on L. Z ∗ will denote the dual of a Banach space Z and we
shall without mention identify Z as a subset of Z ∗∗ in the usual way.
Our main objects of study are the two spaces
M(X, L) = (cid:26)x ∈ X : sup
L
kLxkY < ∞(cid:27)
and
M0(X, L) = (cid:26)x ∈ M(X, L) :
lim
L∋L→∞
kLxkY = 0(cid:27) ,
where the limit L → ∞ is taken in the sense of the one-point com-
pactification of (L, τ ). By replacing X with the closure of M(X, L) in
X and making appropriate modifications to the setting just described,
we may as well assume that M(X, L) is dense in X. Furthermore, we
assume that L is such that
kxkM (X,L) = sup
L
kLxkY
DUALITY AND DISTANCE FORMULAS
3
defines a norm which makes M(X, L) into a Banach space continuously
contained in X. Note that M0(X, L) is then automatically a closed
subspace of M(X, L). These assumptions are mostly for convenience
and will hold trivially in all examples to come.
Example 1.1. Let X = L2(T)/C, Y = L1(T) and
L = (cid:26)LI : LIf = χI
1
I
(f − fI),
∅ 6= I ⊂ T is an arc(cid:27) ,
where χI is the characteristic function of I, I is its length and fI =
1
I RI f ds is the average of f on I. Each arc I is given by its midpoint
a ∈ T and length b, 0 < b ≤ 2π. We give L the quotient topology τ of
T × (0, 2π] obtained when identifying all pairs (a1, b1) and (a2, b2) with
b1 = b2 = 2π. Then M(X, L) = BMO(T) is the space of functions of
bounded mean oscillation on the circle. LI → ∞ in τ means exactly
that I → 0, so it follows that M0(X, L) = V MO(T) are the functions
of vanishing mean oscillation (see Garnett [17], Chapter VI).
M0(X, L) may be trivial even when M(X, L) is not. This happens for
example when M(X, L) is the space of Lipschitz-continuous functions
f on [0, 1] with f (0) = 0 (see Example 4.6). In the general context
considered here we shall not say anything about this, but instead make
one of the following two assumptions. They say that M0(X, L) is dense
in X (under the X-norm) with additional norm control when approxi-
mating elements of M(X, L). This is a natural hypothesis that is easy
to verify in the examples we have in mind. In fact, the assumptions
are necessary for the respective conclusions of the main theorems.
Assumption A. For every x ∈ M(X, L) there is a sequence (xn) in
M0(X, L) such that xn → x in X and supn kxnkM (X,L) < ∞.
Assumption B. For every x ∈ M(X, L) there is a sequence (xn) in
M0(X, L) such that xn → x in X and supn kxnkM (X,L) ≤ kxkM (X,L).
Note that the assumptions could have equivalently been stated with
the sequence (xn) tending to x only weakly in X. The main theorems
are as follows.
Theorem 1.2. Suppose that assumption A holds. Then X ∗ is contin-
uously contained and dense in M0(X, L)∗. Denoting by
I : M0(X, L)∗∗ → X
the adjoint of the inclusion map J : X ∗ → M0(X, L)∗, the operator I is
a continuous isomorphism of M0(X, L)∗∗ onto M(X, L) which acts as
4
KARL-MIKAEL PERFEKT
the identity on M0(X, L). Furthermore, I is an isometry if assumption
B holds.
Theorem 1.3. Let assumption A hold. Then, for any x ∈ M(X, L),
it holds that
(1)
dist(x, M0(X, L))M (X,L) = lim
L∋L→∞
kLxkY .
Example 1.4. Let X, Y , L and τ be as in Example 1.1. Then As-
sumption B holds by letting fn = f ∗ P1−1/n for f ∈ M(X, L), where Pr
is the Poisson kernel for the unit disc, Pr(θ) = 1−r2
eiθ−r2 . The theorems
say that V MO(T)∗∗ ≃ BMO(T) isometrically via the L2(T)-pairing,
and that
(2)
dist(f, V MO)BM O = lim
I→0
1
I ZI
f − fI ds.
This improves upon a result in [26]. Note that if we repeat the con-
struction with Y = Lp(T) for some 1 < p < ∞, we still obtain
that M(X, L) = BMO(T) with an equivalent norm, due to the John-
Nirenberg theorem (see [17]). This gives us a distance formula, corre-
sponding to (2), involving the p-norm on the right-hand side.
We say that Z is a unique (isometric) predual if for any Banach space
W , W ∗ isometric to Z ∗ implies W isometric to Z. Note that the
canonical decomposition
(3)
Z ∗∗∗ = Z ∗ ⊕ Z ⊥
induces a projection π : Z ∗∗∗ → Z ∗ with kernel Z ⊥,
(πz∗∗∗)(z) = z∗∗∗(z),
∀z ∈ Z.
We say that Z is a strongly unique predual if this is the only projec-
tion π from Z ∗∗∗ to Z ∗ of norm one with Ker π weak-star closed. An
excellent survey of these matters can be found in Godefroy [18].
Corollary 1.5. Suppose that Assumption B holds. Then M0(X, L)∗ is
the strongly unique predual of M(X, L).
2. Preliminaries
One of our main tools will be the isometric embedding V : M(X, L) →
Cb(L, Y ) of M(X, L) into the space Cb(L, Y ) of bounded continuous
Y -valued functions on L, given by
V x = Tx,
DUALITY AND DISTANCE FORMULAS
5
where TxL = Lx as before. Cb(L, Y ) is normed by the usual supremum-
norm, so that V indeed is an isometry. Note that V embeds M0(X, L)
into the space C0(L, Y ), consisting of those functions T ∈ Cb(L, Y )
vanishing at infinity.
In order to study duality via this embedding, we will make use of vector-
valued integration theory. Of central importance will be the Riesz-
Zinger theorem [28] for C0(L, Y ), representing the dual of C0(L, Y ) as
a space of measures. Let B0 be the σ-algebra of all Baire sets of L,
generated by the compact Gδ-sets. By cabv(L, Y ∗) we shall denote the
Banach space of countably additive vector Baire measures µ : B0 → Y ∗
with bounded variation
kµk = supX kµ(Ei)kY ∗ < ∞,
where the supremum is taken over all finite partitions of L = ∪Ei into
disjoint Baire sets Ei. Excellent references for these matters are found
for example in Dobrakov [11], [12] and [13].
Theorem 2.1 ([13], Theorem 2). For every bounded linear functional
ℓ ∈ C0(L, Y )∗ there is a unique measure µ ∈ cabv(L, Y ∗) such that
(4)
ℓ(T ) = ZL
T (L) dµ(L),
T ∈ C0(L, Y ).
Furthermore, every µ ∈ cabv(L, Y ∗) defines a continuous functional on
C0(L, Y ) via (4), and kℓk = kµk.
Remark 2.2. In our situation, (L, τ ) being σ-compact, every continuous
function T : L → Y is Baire measurable ([19], pp. 220–221).
In
particular, every T ∈ Cb(L, Y ) induces a bounded functional m ∈
cabv(L, Y ∗)∗ via
m(µ) = ZL
T (L) dµ(L).
It is clear that kmk = kT kCb(L,Y ). Hence, Cb(L, Y ) isometrically em-
beds into cabv(L, Y ∗)∗ in a way that extends the canonical embedding
of C0(L, Y ) into C0(L, Y )∗∗ = cabv(L, Y ∗)∗.
In the proof of Theorem 1.3 we shall require the following simple lemma.
Lemma 2.3. Suppose that m ∈ Cb(L, Y )∗ annihilates C0(L, Y ). Then
(5)
m(T ) ≤ kmk lim
L→∞
kT (L)kY ,
T ∈ Cb(L, Y ).
Proof. Let K1 ⊂ K2 ⊂ · · · be an increasing sequence of compact subsets
of (L, τ ) such that L = ∪Kn. Denote by αL = L ∪ {∞} the one point
6
KARL-MIKAEL PERFEKT
compactification of L. For each n, let sn : αL → [0, 1] be a continuous
function such that s−1
n (0) ⊃ Kn and sn(∞) = 1. Then
m(T ) = m(snT ) ≤ kmk sup
L∈L\Kn
kT (L)kY .
In the limit we obtain (5).
(cid:3)
Corollary 1.5 will follow as an application of a result in [18].
Theorem 2.4 ([18], Theorem V.1). Let Z be a Banach space. Suppose
that for every z∗∗ ∈ Z ∗∗ it holds that: z∗∗ ∈ Z if and only if
z∗∗(z∗) = lim
n
for every weak Cauchy sequence (z∗
n) in Z ∗ with weak-star limit z∗.
Then Z is the strongly unique predual of Z ∗.
z∗∗(z∗
n)
3. Main Results
We shall now proceed to prove the main theorems and Corollary 1.5.
Proof of Theorem 1.3. Seeing as M0(X, L) is continuously contained
in X, every x∗ ∈ X ∗ is clearly continuous also on M0(X, L). Assump-
tion A implies that M0(X, L) is dense in X, so that each element of
X ∗ induces a unique functional on M0(X, L). This proves that X ∗ is
continuously contained in M0(X, L)∗.
We shall now demonstrate that X ∗ is dense in M0(X, L)∗. Thus, let
ℓ ∈ M0(X, L)∗. By Theorem 2.1 and the Hahn-Banach theorem there
is a measure µ ∈ cabv(L, Y ∗) such that
ℓ(x) = (ℓ ◦ V −1)(Tx) = ZL
Lx dµ(L),
x ∈ M0(X, L).
Let K1 ⊂ K2 ⊂ · · · be an increasing sequence of compact subsets of
(L, τ ) such that L = ∪Kn. By ([19], 50.D) we can choose the Kn
to be Gδ and hence Baire sets. Put µn = µKn and let ℓn be the
corresponding functionals
ℓn(x) = ZL
Lx dµn(L).
The operators L ∈ Kn are uniformly bounded by the Banach-Steinhaus
theorem, from which it is clear that ℓn ∈ X ∗. Finally, note that
limn kµn − µk = 0, which implies that the functionals ℓn converge to ℓ
in M0(X, L)∗.
X ∗ being dense ensures the injectivity of I = J ∗. Moreover, it is clear
that I acts as the identity on M0(X, L). Let m ∈ M0(X, L)∗∗ and
DUALITY AND DISTANCE FORMULAS
7
x = Im ∈ X. Note that the unit ball of M0(X, L) is weak-star dense
in the unit ball of M0(X, L)∗∗ ([9], Proposition 4.1). Furthermore,
the weak-star topology of M0(X, L)∗∗ is metrizable on the unit ball,
since M0(X, L)∗ was just proven to be separable. Accordingly, choose
a sequence xn ∈ M0(X, L) with kxnk ≤ kmk such that xn → m weak-
star. Then, for y∗ ∈ Y ∗ and L ∈ L,
(JL∗y∗)(xn)
y∗(Lx) = (L∗y∗)(x) = m(JL∗y∗) = lim
n
y∗(Lxn).
(L∗y∗)(xn) = lim
n
= lim
n
It follows that x ∈ M(X, L) and
(6)
since
kxkM (X,L) = kImkM (X,L) ≤ kmkM0(X,L)∗∗ ,
kLxkY = sup
ky∗k=1
y∗(Lx) ≤ sup
n
kxnkM (X,L) ≤ kmk ,
∀L ∈ L.
We have thus proved that I maps M0(X, L)∗∗ into M(X, L) contrac-
tively.
Given x ∈ M(X, L) choose xn ∈ M0(X, L) such that xn → x in X and
supn kxnkM (X,L) < ∞ (≤ kxkM (X,L) if assumption B holds). Define
x ∈ M0(X, L)∗∗ by x(Jx∗) = x∗(x) = limn(Jx∗)(xn) for x∗ ∈ X ∗. It
is clear from the last equality that this defines x as a bounded linear
functional on M0(X, L)∗ and if assumption B holds, then
(7)
kxkM0(X,L)∗∗ ≤ kxkM (X,L) .
Obviously, I x = x. This proves that I is onto. If assumption B holds,
then we obtain from (6) and (7) that I is an isometry.
(cid:3)
Proof of Theorem 1.3. Let m ∈ M(X, L)∗. Then m ◦ V −1 acts on
V M(X, L). As in Remark 2.2 we naturally view Cb(L, Y ) as a sub-
space of cabv(L, Y ∗)∗. With this identification, m ◦ V −1 extends by
Hahn-Banach to a functional ¯m ∈ cabv(L, Y ∗)∗∗ with k ¯mk = kmk.
Applying the decomposition (3) with Z = C0(L, Y ) we obtain
cabv(L, Y ∗)∗∗ = cabv(L, Y ∗) ⊕ C0(L, Y )⊥,
and we decompose ¯m = ¯mω∗ + ¯ms accordingly. Let µ ∈ cabv(L, Y ∗) be
the measure corresponding to ¯mω∗, so that, in particular,
¯mω∗(T ) = ZL
T (L) dµ(L), T ∈ Cb(L, Y ).
8
KARL-MIKAEL PERFEKT
Let I : M0(X, L)∗∗ → M(X, L) be the isomorphism given by Theorem
1.2. With Z = M0(X, L), (3) gives
(8)
M(X, L)∗ ≃ M0(X, L)∗∗∗ = M0(X, L)∗ ⊕ M0(X, L)⊥,
and we obtain a second decomposition m ◦ I = (m ◦ I)ω∗ + (m ◦ I)s.
Our first goal is to show that the former decomposition is an extension
of the latter. More precisely, we have
Claim. (m ◦ I)ω∗ ≡ 0 if and only if ¯mω∗ annihilates V M(X, L).
To prove this, let x ∈ M(X, L) and let x = I −1x ∈ M0(X, L)∗∗. As in
the proof of Theorem 1.2, choose xn ∈ M0(X, L) with kxnkM0(X,L) ≤
kxk such that xn → x weak-star. Note that (xn) in particular converges
to x weakly in X. Hence Lxn → Lx weakly in Z for every L ∈ L. Since
also supn,L kLxnkY < ∞, it follows from ([13], Theorem 9) that
(9)
ZL
Lx dµ(L) = lim
n ZL
Lxn dµ(L).
To be more precise, Theorem 9 in [13] allows us to move the limit
inside the integral when integrating over a compact Gδ-set K ⊂ L.
However, as in the proof of Theorem 1.2, we obtain (9) by an obvious
approximation argument. We thus have
¯mω∗(V x) = lim
n ZL
Lxn dµ(L) = lim
n
¯mω∗(V xn) = lim
n
¯m(V xn) =
lim
n
m(xn) = lim
n
(m ◦ I)(xn) = lim
n
(m ◦ I)ω∗(xn) = (m ◦ I)ω∗(x),
so that the claim is proven.
We can now calculate the distance from x ∈ M(X, L) to M0(X, L)
using duality.
dist(x, M0(X, L))M (X,L) = sup
kmk=1
m(x) =
sup
kmk=1
¯ms(V x).
(m◦I)ω∗ ≡0
¯mω∗ ⊥V M (X,L)
Since k ¯msk ≤ kmk we obtain by Lemma 2.3 that
dist(x, M0(X, L))M (X,L) ≤ lim
L→∞
kLxkY .
The converse inequality is trivial; for any x0 ∈ M0(X, L) we have
kx − x0kM (X,L) ≥ lim
L→∞
≥ lim
L→∞
kLx − Lx0kY
(kLxkY − kLx0kY ) = lim
L→∞
kLxkY .
(cid:3)
DUALITY AND DISTANCE FORMULAS
9
Proof of Corollary 1.5. As in the preceding proof, for m ∈ M0(X, L)∗∗∗,
write
m = mω∗ + ms,
in accordance with (8). Suppose that m /∈ M0(X, L)∗, or equiva-
lently, ms 6= 0. Pick x ∈ M0(X, L)∗∗ such that ms(x) 6= 0 and let
xn ∈ M0(X, L) converge to x weak-star. Then (xn), as a sequence in
M0(X, L)∗∗, is weakly Cauchy, since
lim
n
m′(xn) = m′
ω∗(x),
∀m′ ∈ M0(X, L)∗∗∗.
On the other hand,
m(x) = mω∗(x) + ms(x) 6= mω∗(x),
so that
m(x) 6= lim
n
m(xn).
We have thus verified the condition of Theorem 2.4 for Z = M0(X, L)∗,
proving the corollary.
(cid:3)
4. Examples
Example 4.1. Denoting by L2
space on the unit disc D, let X = L2
with f (0) = 0. Let Y = C,
a = L2(D) ∩ Hol(D) the usual Bergman
a/C be the space of functions f ∈ L2
a
L = {Lw : Lwf = (1 − w2)f ′(w), w ∈ D},
and let τ be the topology of D. Then M(X, L) = B/C is the Bloch
space modulo constants and M0(X, L) = B0/C is the little Bloch
space (up to constants). For f ∈ B it is clear that the dilations fr,
fr(z) = f (rz), converge to f in L2
a as r → 1− and that kfrkB ≤ kf kB,
verifying the hypothesis of Assumption B. From the theorems we ob-
tain that (B0/C)∗∗ ≃ B/C isometrically via the L2
a-pairing, as well as
the distance formula
dist(f, B0/C)B/C = lim
w→1
(1 − w2)f ′(w).
This improves a result previously obtained in [4] and [27]. Corollary
1.5 says furthermore that the the Bloch space has a unique predual,
reproducing a result found in [23].
Example 4.2. Much more generally, the main theorems can be applied
to Mobius invariant spaces of analytic functions through the following
10
KARL-MIKAEL PERFEKT
construction. Denote by G the Mobius group, consisting of the con-
formal disc automorphisms φ : D → D. Each function in G is of the
form
φa,λ(z) = λ
,
a ∈ D, λ ∈ T.
a − z
1 − ¯az
G is a topological group with the topology of D × T.
φa,λ → ∞ equivalently means that a → 1.
Let X be a Banach space whose members f ∈ X are functions analytic
in D with f (0) = 0. We assume that X is continuously contained in
Hol(D)/C, the latter space equipped with the compact-open topology,
and that it satisfies the properties:
In particular,
[I] X is reflexive.
[II] The holomorphic polynomials p with p(0) = 0 are contained
and dense in X.
[III] For each fixed f ∈ X, the map Tf , Tf φ = f ◦ φ − f (φ(0)) is a
continuous map from G to X.
[IV] limG∋φ→∞ kφ − φ(0)kX = 0.
We now let Y = X and let L be the collection of composition operators
induced by G,
L = {Lφ : Lφf = f ◦ φ − f (φ(0)), φ ∈ G},
equipping it with the topology of G.
M(X, L) and M0(X, L) are then Mobius invariant Banach spaces in
the sense that if f is in either space then so is f ◦ φ − f (φ(0)), φ ∈
G, and furthermore kf ◦ φ − f (φ(0))kM (X,L) = kf kM (X,L). Property
[IV] implies that B1/C is continuously contained in M0(X, L), where
B1 denotes the analytic Besov 1-space, the minimal Mobius invariant
space (see [3]). In particular M0(X, L) contains the polynomials. The
construction of the space M(X, L) has been considered by Aleman and
Simbotin in [1].
General Mobius invariant spaces are studied by Arazy, Fisher and Pee-
tre in [3]. The next proposition, saying that Assumption B holds, is
essentially contained there. Arazy et al. have a stricter definition of
what a Mobius invariant space is, however, so its proof is included here
for completeness.
Proposition 4.3. Given f ∈ M(X, L), f (z) = P∞
n + 1(cid:19) akzk.
(cid:18)1 −
fn(z) =
k
n
Xk=1
1 akzk, let
Then kfnkM (X,L) ≤ kf kM (X,L) and fn → f weakly in X.
DUALITY AND DISTANCE FORMULAS
11
That kfnkM (X,L) ≤ kf kM (X,L) is immediate from the formula
k=−n(cid:16)1 − k
n+1(cid:17) e−ikθ the Fej´er kernels.
Proof. Denote by Φn(θ) = Pn
2π Z 2π
fn ◦ φ − fn(φ(0)) =
1
0
(cid:0)f (eiθφ( · )) − f (eiθφ(0))(cid:1) Φn(θ) dθ,
where the the integral is to be understood as an X-valued function of θ
integrated against the measure Φn(θ) dθ. That fn → f weakly follows
from the same formula with φ(z) = z, because if ℓ ∈ X ∗, then
ℓ(fn) =
1
2π Z 2π
0
ℓ(f (eiθ · ))Φn(θ) dθ → ℓ(f )
as n → ∞, by a standard argument about the Fej´er kernels.
(cid:3)
Applying the theorems, we obtain that M0(X, L)∗∗ ≃ M(X, L) isomet-
rically, that M0(X, L)∗ is the unique isometric predual of M(X, L) and
the formula
dist(f, M0(X, L))M (X,L) = lim
a→1
kf ◦ φa,λ − f (φa,λ(0))kX .
There are many examples of Mobius invariant spaces. Letting X =
L2
a/C we once again obtain the Bloch space, M(X, L) = B/C and
M0(X, L) = B0/C, but with a different norm than in Example 4.1.
When X = H 2/C is the Hardy space modulo constants we get the
space of analytic BMO functions with its conformally invariant norm,
M(X, L) = BMOA/C and M0(X, L) = V MOA/C (see [17]).
The QK-spaces provide a wide class of Mobius invariant spaces that
includes both B and BMOA. For a non-zero, right-continuous, non-
decreasing function K : [0, ∞) → [0, ∞), denote by QK the space of all
f ∈ Hol(D) with f (0) = 0 such that
kf k2
QK = sup
φ∈GZD
f ′(z)2K (cid:18)log
1
φ(cid:19) dA(z) < ∞.
See Ess´en and Wulan [14] for a survey of QK-spaces. See also [1].
Clearly, QK = M(XK, L), if XK is the space of all f ∈ Hol(D) with
f (0) = 0 such that
kf k2
XK = ZD
f ′(z)2K (cid:18)log
1
z(cid:19) dA(z) < ∞.
to verify that XK is a Hilbert space for which properties [II] and [IV]
If K (cid:16)log 1
hold. Furthermore, if K (cid:16)log 1
z(cid:17) is integrable on D and K(ρ) → 0 as ρ → 0+, it is easy
z(cid:17) is a normal weight in the sense of
12
KARL-MIKAEL PERFEKT
Shields and Williams [25], standard arguments show that if φa,λ → φ
in G, then
kf ◦ φa,λ − f (φa,λ(0))kXK
→ kf ◦ φ − f (φ(0))kXK ,
f ∈ XK.
Since f ◦ φa,λ − f (φa,λ(0)) also tends weakly to f ◦ φ − f (φ(0)), we in
fact have norm convergence, verifying [III] under these assumptions.
Hence, if we denote by QK,0 = M0(XK, L) the space of those functions
f ∈ QK such that
lim
a→1ZD
f ′(z)2K (cid:18)log
1
φa,λ(cid:19) dA(z) = 0,
we have proven that Q∗∗
K,0 = QK and that
dist(f, QK,0)QK = lim
a→1ZD
f ′(z)2K (cid:18)log
1
φa,λ(cid:19) dA(z).
Example 4.4. For an open subset Ω of C, let v be a strictly posi-
tive continuous function on Ω. In this example we shall consider the
weighted space Hv(Ω) of analytic functions on Ω bounded under the
weighted supremum norm given by the weight v.
For the purpose of applying our construction, choose an auxiliary strictly
positive continuous weight function w : Ω → R+ such that w is inte-
grable on Ω. Define X = L2
a(v2w) to be the weighted Bergman space
on Ω with weight v2w, consisting of those f ∈ Hol(Ω) such that
kf kL2
a(v2w) = ZΩ
f (z)2v(z)2w(z) dA(z) < ∞.
One easily verifies that X is a Hilbert space continuously contained in
Hol(D). Furthermore, let Y = C,
L = {Lz : Lzf = v(z)f (z),
z ∈ Ω},
and let τ be given by the usual topology of Ω.
It is then clear that M(X, L) = Hv(Ω) is the Banach space of all
f ∈ Hol(Ω) such that vf is bounded, and that M0(X, L) = Hv0(Ω)
is the corresponding little space, consisting of those f such that vf
vanishes at infinity on Ω.
Assumption A (Assumption B) holds if and only if there for each f ∈
Hv(Ω) exists a sequence (fn) ⊂ Hv0(Ω) such that fn → f pointwise in
Ω and sup kfnkHv(Ω) < ∞ (sup kfnkHv(Ω) ≤ kf kHv(Ω)). We have hence
recovered a result of Bierstedt and Summers [6]; Hv0(Ω)∗∗ ≃ Hv(Ω)
via the natural isomorphism if and only if this pointwise weighted ap-
proximation condition holds. The isomorphism is isometric precisely
when Assumption B holds.
DUALITY AND DISTANCE FORMULAS
13
When either assumption holds we furthermore obtain the distance for-
mula
dist(f, Hv0(Ω))Hv(Ω) = lim
Ω∋z→∞
w.r.t. Ω
v(z)f (z).
Bierstedt and Summers give sufficient conditions for radial weights v
which ensure that Assumption B holds. For example, it holds when
v is a radial weight vanishing at ∂Ω, where Ω is a balanced domain
such that Ω is a compact subset of {z ∈ C : rz ∈ Ω} for every 0 <
r < 1. When Ω = C, we may take any radial weight v on C decreasing
rapidly at infinity to obtain a space Hv(C) of entire functions satisfying
Assumption B. See [6] for details.
For simplicity the above considerations have not been carried out in
their full generality. We could, for example, have considered weighted
spaces M(X, L) of harmonic functions, or of functions defined on Cn,
n > 1. However, problems arise when Ω is an open subset of an infinite-
dimensional Banach space, since local compactness of Ω is lost. This
case has been considered by Garc´ıa, Maestre and Rueda in [16]. In a
different direction, the biduality problem has been studied for weighted
inductive limits of spaces of analytic functions. See Bierstedt, Bonet
and Galbis [7] for results in this context.
Example 4.5. We now turn to rectangular bounded mean oscillation
on the 2-torus. The space BMORect(T2) consists of those f ∈ L2(T2)/C
such that
1
f (ζ, λ) − fJ (ζ) − fI(λ) + fI×J2 ds(ζ)ds(λ) < ∞,
sup
IJ ZI ZJ
J RJ f (ζ, λ) ds(λ) and φI(λ) = 1
where the supremum is taken over all subarcs I, J ⊂ T, ds is arc length
measure, φJ (ζ) = 1
the averages of f (ζ, ·) and f (·, λ) on J and I, respectively, and fI×J
is the average of f on I × J. BMORect(T2) is one of several possible
generalizations of BMO(T) to the two-variable case. We focus on this
particular one because it fits naturally into our scheme. A treatment
of rectangular BMO can be found in Ferguson and Sadosky [15].
To obtain BMORect(T2) via our construction, let X = L2(T2)/C, Y =
L2(T2) and
I RI f (ζ, λ) ds(ζ) are
L = (cid:26)LI,J : LI,Jf = χI×J
1
IJ
(f − fJ − fI + fI×J )(cid:27) ,
where I and J range over all non-empty arcs. Denoting by τ the
quotient topology considered in Example 1.1, we equip L with the cor-
responding product topology τ × τ , so that LI,J → ∞ means that
14
KARL-MIKAEL PERFEKT
min(I, J) → 0. By construction M(X, L) = BMORect(T2). Accord-
ingly, M0(X, L) will be named V MORect(T2). Assumption B is verified
exactly as in Example 1.4, letting fn = P1−1/n(ζ) ∗ P1−1/n(λ) ∗ f be a
double Poisson integral.
From Theorem 1.2 we hence obtain that V MORect(T2)∗∗ is isometrically
isomorphic to BMORect(T2) via the L2(T2)-pairing, and Theorem 1.3
gives
dist(f, V MORect)BM ORect =
lim
min(I,J)→0
kLI,J f kL2(T2) .
Another possible generalization of BMO(T) to several variables is
known as product BMO.
In [21], Lacey, Terwilleger and Wick ex-
plore the corresponding product V MO space. It would be interesting
to apply our techniques also to this case, but one meets the difficulty
of defining a reasonable topology on the collection of all open subsets
of Rn. On the other hand, the predual of BMORect(T2) is given as a
space spanned by certain "rectangular atoms" (see [15]) and is as such
more difficult to understand than the predual of product BMO, which
is the Hardy space H 1(Tn) of the n-torus.
Example 4.6. Let 0 < α ≤ 1 and let Ω be a compact subset of Rn.
In this example we shall treat the Lipschitz-Holder space Lipα(Ω). By
definition, a real-valued function f on Ω is in Lipα(Ω) if and only if
kf kLipα(Ω) = sup
x,y∈Ω
x6=y
f (x) − f (y)
x − yα < ∞.
As usual we identify f and f + C, C constant, in order to obtain a
norm.
X will be chosen as a quotient space of an appropriate fractional
Sobolev space (Besov space) W l,p(Rn). For 0 < l < 1 and 1 < p < ∞,
W l,p consists of those f ∈ Lp(Rn) such that
ZRn ZRn
f (x) − f (y)p
x − ypl+n dx dy < ∞.
Choose l and p such that 0 < l < α and pl > n. By a Sobolev type
embedding theorem ([22], Proposition 4.2.5) it then holds that W l,p
continuously embeds into the space of continuous bounded functions
Cb(Rn). Let
AΩ = {f ∈ W l,p : f (x) = 0, x ∈ Ω}.
We set X = W l,p/AΩ.
DUALITY AND DISTANCE FORMULAS
15
To obtain Lipα(Ω) through our construction, let Y = C and let every
operator Lx,y ∈ L, x, y ∈ Ω, x 6= y, be of the form
Lx,yf =
f (x) − f (y)
x − yα
.
We give L the topology of {(x, y) ∈ Ω × Ω : x 6= y}. It then holds that
M(X, L) = Lipα(Ω). One inclusion is obvious and to see the other one
let f ∈ Lipα(Ω). As in [10], f can be extended to f ∈ Lipα(Rn), f = f
on Ω. Letting χ ∈ C ∞
c (Rn) be a cut-off function such that χ(x) = 1
for x ∈ Ω, it is straightforward to check that χ f ∈ M(X, L), verifying
that Lipα(Ω) ⊂ M(X, L).
Note that M0(X, L) = lipα(Ω) is the corresponding little Holder space,
consisting of those f ∈ Lipα(Ω) such that
lim
x−y→0
f (x) − f (y)
x − yα = 0.
When α = 1, the space lipα(Ω) is empty in many cases, so that As-
sumption A may fail. However, for α < 1 we can verify Assumption B
in general. Let Pt, t > 0, be the n-dimensional Poisson kernel,
Pt(x) = c
t
(t2 + x2)(n+1)/2 ,
x ∈ Rn,
for the appropriate normalization constant c, and let χ denote the same
cut-off as before. For f ∈ M(X, L) it is straightforward to verify that
χ · (Pt ∗ f ) ∈ M0(X, L),
kχ · (Pt ∗ f )kLipα(Ω) ≤ kf kLipα(Ω) ,
and that χ · (Pt ∗ f ) → f = χf weakly in X as t → 0+, the final
statement following from the reflexivity of X and the fact that Pt ∗ f
tends to f pointwise almost everywhere.
We conclude that, for 0 < α < 1,
dist(f, lipα(Ω))Lipα(Ω) = lim
x−y→0
f (x) − f (y)
x − yα
.
In addition, Theorem 1.2 says that lipα(Ω)∗∗ ≃ Lipα(Ω) isometrically.
In Kalton [20] it is proven that lipα(M)∗∗ ≃ Lipα(M) under very gen-
eral conditions on M, for example whenever M is a compact metric
space. It would be interesting to see if the theorems of this paper can
be applied in this situation and, if this is the case, which space X to
embed Lipα(M) into.
16
KARL-MIKAEL PERFEKT
Acknowledgements. The author is grateful to Alexandru Aleman
and Sandra Pott for helpful discussions, ideas and support regarding
this paper, and to Jos´e Bonet for pointing out the application consid-
ered in Example 4.4.
References
[1] Aleman, A., Simbotin, A-M, Estimates in Mobius invariant spaces of analytic
functions, Complex Var. Theory Appl. 49 (2004), no. 7-9, 487–510.
[2] Anderson, J. M., Duncan, J., Duals of Banach spaces of entire functions, Glas-
gow Math. J. 32 (1990), no. 2, 215–220.
[3] Arazy J., Fisher S. D., Peetre J, Mobius invariant function spaces, J. Reine
Angew. Math. 363 (1985), 110–145.
[4] Attele, K. R. M., Interpolating sequences for the derivatives of Bloch functions,
Glasgow Math. J. 34 (1992), no. 1, 35–41.
[5] Axler, S., Shapiro, J. H., Putnam's theorem, Alexander's spectral area esti-
mate, and VMO, Math. Ann. 271 (1985), no. 2, 161–183.
[6] Bierstedt, K. D., Summers, W. H., Biduals of weighted Banach spaces of ana-
lytic functions, J. Austral. Math. Soc. Ser. A 54 (1993), no. 1, 70–79.
[7] Bierstedt, K. D., Bonet, J., Galbis, A., Weighted spaces of holomorphic func-
tions on balanced domains, Michigan Math. J. 40 (1993), no. 2, 271–297.
[8] Carmona, J., Cuf´ı, J.,, On the distance of an analytic function to VMO, J.
London Math. Soc. (2) 34 (1986), no. 1, 52–66.
[9] Conway, J. B, "A course in functional analysis", Second edition, Graduate
Texts in Mathematics, 96. Springer-Verlag, New York, 1990.
[10] Czipszer, J., Geh´er, L, Extension of functions satisfying a Lipschitz condition,
Acta Math. Acad. Sci. Hungar. 6 (1955), 213–220.
[11] Dobrakov, I., On integration in Banach spaces. I, Czechoslovak Math. J.
20(95) (1970), 511–536.
[12] Dobrakov, I., On integration in Banach spaces. II, Czechoslovak Math. J.
20(95) (1970), 680–695.
[13] Dobrakov, I., On representation of linear operators on C0 (T, X), Czechoslovak
Math. J. 21 (96) 1971 13–30.
[14] Ess´en, M., Wulan, H., On analytic and meromorphic functions and spaces of
QK-type, Illinois J. Math. 46 (2002), no. 4, 1233–1258.
[15] Ferguson, S. H., Sadosky, C., Characterizations of bounded mean oscillation
on the polydisk in terms of Hankel operators and Carleson measures, J. Anal.
Math. 81 (2000), 239–267.
[16] Garc´ıa, D., Maestre, M., Rueda, P, Weighted spaces of holomorphic functions
on Banach spaces, Studia Math. 138 (2000), no. 1, 1–24.
[17] Garnett, J. B., Bounded analytic functions. Revised first edition, Graduate
Texts in Mathematics, 236. Springer, New York, 2007
[18] Godefroy, G., Existence and uniqueness of isometric preduals: a survey, "Ba-
nach space theory" (Iowa City, IA, 1987), 131–193, Contemp. Math., 85, Amer.
Math. Soc., Providence, RI, 1989.
[19] Halmos, P. R., "Measure Theory", D. Van Nostrand Company, Inc., New York,
N. Y., 1950.
DUALITY AND DISTANCE FORMULAS
17
[20] Kalton, N. J., Spaces of Lipschitz and Hlder functions and their applications,
Collect. Math. 55 (2004), no. 2, 171–217.
[21] Lacey, M. T., Terwilleger, E.; Wick, B. D., Remarks on product VMO, Proc.
Amer. Math. Soc. 134 (2006), no. 2, 465–474 (electronic).
[22] Maz'ya, V. G., Shaposhnikova, T. O., "Theory of Sobolev multipliers. With
applications to differential and integral operators", Grundlehren der Mathe-
matischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
337. Springer-Verlag, Berlin, 2009.
[23] Nara, C., Uniqueness of the predual of the Bloch space and its strongly exposed
points, Illinois J. Math. 34 (1990), no. 1, 98–107.
[24] Rubel, L. A., Shields, A. L., The second duals of certain spaces of analytic
functions, J. Austral. Math. Soc. 11 1970 276–280.
[25] Shields, A. L., Williams, D. L., Bounded projections, duality, and multipliers
in spaces of analytic functions, Trans. Amer. Math. Soc. 162 1971 287–302.
[26] Stegenga, D. A., Stephenson, K., Sharp geometric estimates of the distance to
VMOA, The Madison Symposium on Complex Analysis (Madison, WI, 1991),
421–432, Contemp. Math., 137, Amer. Math. Soc., Providence, RI, 1992.
[27] Tjani, M., Distance of a Bloch function to the little Bloch space, Bull. Austral.
Math. Soc. 74 (2006), no. 1, 101–119.
[28] Zinger, I., Linear functionals on the space of continuous mappings of a compact
Hausdorff space into a Banach space, (Russian) Rev. Math. Pures Appl. 2 1957
301–315
|
1504.04087 | 2 | 1504 | 2015-05-27T02:53:12 | On $L^p-$boundedness of pseudo-differential operators of Sj\"ostrand's class | [
"math.FA"
] | We extended the known result that symbols from modulation spaces $M^{\infty,1}(\mathbb{R}^{2n})$, also known as the Sj\"{o}strand's class, produce bounded operators in $L^2(\mathbb{R}^n)$, to general $L^p$ boundedness at the cost of lost of derivatives. Indeed, we showed that pseudo-differential operators acting from $L^p$-Sobolev spaces $L^p_s(\mathbb{R}^n)$ to $L^p(\mathbb{R}^n)$ spaces with symbols from the modulation space $M^{\infty,1}(\mathbb{R}^{2n})$ are bounded, whenever $s\geq n|1/p-1/2|.$ This estimate is sharp for all $1\leq p\leq\infty$. | math.FA | math |
(1)
e2πix·ξσ(x, ξ) f (ξ) dξ,
f ∈ S(Rn),
σ(x, D)f (x) =ZRn
ON Lp−BOUNDEDNESS OF PSEUDO-DIFFERENTIAL OPERATORS
OF SJ OSTRAND'S CLASS
JAYSON CUNANAN
Abstract. We extended the known result that symbols from modulation spaces M ∞,1(R2n),
also known as the Sjostrand's class, produce bounded operators in L2(Rn), to general Lp
boundedness at the cost of lost of derivatives. Indeed, we showed that pseudo-differential
operators acting from Lp-Sobolev spaces Lp
s(Rn) to Lp(Rn) spaces with symbols from the
modulation space M ∞,1(R2n) are bounded, whenever s ≥ n1/p − 1/2. This estimate is
sharp for all 1 ≤ p ≤ ∞.
1. Introduction
Given a symbol σ ∈ S ′(R2n), we define a pseudo-differential operator in Rn by
where f (ξ) = Ff (ξ) = RRn e−i2πx·ξf (x) dx is the Fourier transform of f. It is clear that
σ(x, D)f is well-defined as a temperate distribution. These operators play essential roles in
many fields including, partial differential equations (PDEs), quantum mechanics, and signal
processing. Boundedness results of pseudo-differential operators on Lp−Sobolev spaces are
of special interest since they imply regularity results of the corresponding PDEs.
A thorough analysis of such operators have been carried on for Hormander's classes,
Sm
ρ,δ, m ∈ R, 0 ≤ δ ≤ ρ ≤ 1, of smooth functions σ(x, ξ) satisfying the estimates
∂α
x ∂β
ξ σ(x, ξ) ≤ Cα,β(1 + ξ)m−ρβ+δα,
for all multi-indexes α and β. In particular, the classical Calderon-Vaillancourt theorem
[15] states that symbols belonging to S0
0,0 produce bounded operators on L2. In [16, 17],
Sjostrand extended this result on L2 boundedness by introducing a new class of sym-
bols, containing S0
0,0, which do not require boundedness of the derivatives of the symbols.
This new class is later identified as the modulation space M ∞,1, first introduced in time-
frequency analysis by Feichtinger [7, 8, 9]. Roughly speaking, elements of M p,q are distribu-
tions with the same local regularity as a function whose Fourier transform is in Lq and has
decay properties of an Lp function (see section 2 for details). In [11, 12], Grochenig and
Heil then significantly extended Sjostrand's result to boundedness of pseudo-differential
operators on all modulation spaces.
2010 Mathematics Subject Classification. 35S05, 42B35.
Key words and phrases. Pseudo-differential operators, modulation spaces, Wiener amalgam spaces.
1
2
JAYSON CUNANAN
For general Lp boundedness, Hormander showed in [13] that mp = n(1 − ρ)1/p − 1/2
is the critical decreasing order for the Lp boundedness of pseudo-differential operators in
ρ,δ. Several sufficient conditions are known for the Lp boundedness of pseudo-differential
Sm
operators of Homander classes, we refer the reader to [1] and the references therein for a
detailed account. For symbols in M ∞,1, it remains to know the minimal loss of derivatives
required to achieve Lp boundedness for all 1 ≤ p ≤ ∞.
Our goal in this note is to give sufficient and necessary conditions for the boundedness
s(Rn) to
of pseudo-differential operators whose symbols come from M ∞,1(R2n) acting on Lp
Lp(Rn). The following theorem is our main result.
Theorem 1.1. Let 1 ≤ p ≤ ∞. Then the pseudo-differential operator σ : S(Rn) → S ′(Rn)
having symbols in M ∞,1(R2n) extends to a bounded operator from Lp
s(Rn) to Lp(Rn) if and
only if s ≥ n1/p − 1/2.
In classical literatures for Lp boundedness of operators in Hormander's classes, results
often require much stronger assumption, than simply belonging to a particular Sm
ρ,δ, when
dealing with the endpoints p = 1, ∞. As an illustration, it is known that the converse of
Theorem 4.1 in Section 4 does not hold for p = 1 and p = ∞. Although, Fefferman proved
in [6] the converse of Theorem 4.1 for the case 1 < p < ∞. Here, Theorem 1.1 shows the
advantage of taking symbols in M ∞,1(R2n) since we are able to show boundedness for all
1 ≤ p ≤ ∞.
We organize this note as follows. Section 2 will consist of notations, definitions and
In Section 3 and Section 4, we prove the
properties of the involved functions spaces.
sufficient and necessary conditions, respectively, for the boundedness of pseudo-differential
operators Lp
s(Rn) → Lp(Rn).
2. Preliminaries
Notations. The Schwartz class of test functions on Rn shall be denoted by S(Rn) and
its dual, the space of tempered distributions, by S ′(Rn). The Lp(Rn) norm is given by
f Lp = (RRn f (x)p dx)1/p whenever 1 ≤ p < ∞, and f L∞ = ess.supx∈Rnf (x). The
Fourier transform of a function f ∈ S(Rn) is given by
Ff (ξ) = f (ξ) =ZRn
e−i2πx·ξf (x) dx
which is an isomorphism of the Schwartz space S(Rn) onto itself that extends to the
tempered distributions S ′(Rn) by duality. The inverse Fourier transform is given by
F −1f (x) = f (x) =RRn ei2πξ·xf (ξ) dξ. Given 1 ≤ p ≤ ∞, we denote by p′ the conjugate ex-
ponent of p (i.e. 1/p + 1/p′ = 1). The translation operator is defined by Txf (t) = f (t − x).
We now recall the definitions of the function spaces to be used in this article.
Lp-Sobolev: We adapted Stein's notation [18] and define the Lp-Sobolev norm by
We remark that this notation should not be interchanged with the set of all f such that
(1 + x2)s/2f belongs to Lp.
f Lp
s
= ((1 + · 2)sbf (·))∨Lp.
ON Lp−BOUNDEDNESS OF PSEUDO-DIFFERENTIAL OPERATORS OF SJ OSTRAND'S CLASS
3
Modulation spaces. Let s ∈ R, we denote hξis = (1 + ξ2)s/2, ξ ∈ Rn. The class of
weight functions to be used in this note will be denoted by νs1,s2(x, ξ) = hxis1hξis2, si ∈
R, i = 1, 2, on R2n or m(x, ξ, ζ1, ζ2) = hxis1hξis2 = (νs1,s2 ⊗ 1)(x, ξ, ζ1, ζ2), on R4n. For
1 ≤ p, q ≤ ∞, the modulation space M p,q
m consists of all tempered distributions f ∈ S ′(Rn)
such that the norm
f M p,q
m
=
Z
Rn
Z
Rn
Vgf (x, ξ)pm(x, ξ)p dx
q/p
1/q
,
dξ
is finite, with usual modifications if p or q = ∞. Here Vgf denotes the short-time Fourier
transform (STFT) of f ∈ S ′(Rn) with respect to the window 0 6= g ∈ S(Rn) defined by
Vgf (x, ξ) =ZRn
f (y)g(y − x)e−2πiy·ξ dy.
The modulation space M p,q
m (Rn) is a Banach space whose norm is independent of the choice
of the window g. If m ≡ 1, we write M p,q. A weighted modulation space in R2n will be
denoted by M p,q
νs1 ,s2 ⊗1(R2n).
Wiener amalgam spaces: For 1 ≤ p, q ≤ ∞, s ∈ R and 0 6= g ∈ S, the Wiener
(Rn) consists of all tempered distributions f ∈ S ′(Rn) such that the
amalgam space W p,q
norm
s
f W p,q
s
:= f W (F Lq
s,Lp) =
Z
Rn
Z
Rn
F(f · Tyg)(ω)qhωisq dω
p/q
1/p
,
dy
0
is finite, with usual modifications if p or q = ∞. If s = 0 we simply write W p,q instead of
W p,q
. We note that this definition is independent of the choice of window g. We denote
the closure of the Schwartz class S(Rn) in the W p,q
s (Rn). If 1 ≤ p, q < ∞
then W p,q
-norm by W p,q
s = W p,q
s
s
.
We collect properties of Wiener amalgam spaces in the following lemma. We note also
that analogous properties hold for modulation spaces. In fact, FW p,q = M p,q.
Lemma 2.1. Let p, q, pi.qi ∈ [1, ∞] for i = 1, 2 and sj ∈ R for j = 1, 2. Then
(1) S(Rn) ֒→ W p,q(Rn) ֒→ S ′(Rn);
(2) S is dense in W p,q if p and q < ∞;
(3) If q1 ≤ q2 and p1 ≤ p2, then W p1,q1 ֒→ W p2,q2;
(4) If s1 ≥ s2, then W p,q
s1
(5) hDi−s : W p,q → W p,q
s
(6) (Convolution) If FLq1 ∗ FLq2 ֒→ FLq and Lp1 ∗ Lp2 ֒→ Lp, then
, f 7→ (bf (·)h·i−s)∨ , is an isomorphism;
֒→ W p,q
s2 ;
W (FLq1, Lp1) ∗ W (FLq2, Lp2) ֒→ W (FLq, Lp).
4
JAYSON CUNANAN
(7) (Complex interpolation) For 0 < θ < 1. Let
s = θs1 + (1 − θ)s2. Then
1
p
=
θ
p1
+
1 − θ
p2
,
1
q
=
θ
q1
+
1 − θ
q2
and
[W p1,q1
s1
, W p2,q2
s2
][θ] = W p,q
s
;
(8) (Duality) (W p,q
s )′ = W p′,q′
−s , where 1/p + 1/p′ = 1 = 1/q + 1/q′,
p, q 6= ∞.
The proofs of these statements can be found in [7, 10, 9, 19].
Pseudodifferential operators in the form (1) are referred as Kohn-Nirenberg correspon-
dence. Now we give another form which is called the Weyl quantization, and can be defined
as follows:
Lσf (x) =Z Z e2π(x−y)ξσ(
, ξ)f (y) dydξ.
x + y
2
Our interest in using this form is due to its adjoint, namely, L∗
σ = L¯σ. Moreover, for
symbols in modulation spaces, boundedness results can be obtained using either the Kohn-
Nirenberg or Weyl form. This statement will be made clear by Remark 2.1.
We quote the following result in [?, Theorem 14.3.5].
Theorem 2.1. Let T be a continuous linear operator mapping S(Rn) into S ′(Rn). Then
there exist tempered distributions K, σ, a ∈ S ′(R2n), such that T has the following repre-
sentations:
(i) as an integral operator hT f, gi = hK, g ⊗ ¯f i, for f, g ∈ S(Rn);
(ii) as a pseudodifferential operator T = Lσ, with Weyl symbol σ and T = a(x, D) with
Kohn-Nirenberg symbol a.
The relations between K, σ, a are given by
(2)
σ = F2τsK,
a = U σ
where F2 is the partial Fourier transform in the second variable, τs is the symmetric co-
ordinate transformation τsK(x, y) = K(x + y/2, x − y/2) and the operator U is defined by
cU σ(ω, u) = eπiωu σ(ω, u).
Remark 2.1. Since, σ(x, D) = LU −1σ, and the fact that M ∞,1 is invariant under U −1
implies that boundedness results for Kohn-Nirenberg and Weyl form are interchangeable
for symbols in M ∞,1 (see [?, Corollary 14.5.5]).
3. Sufficient condition
The proof for the sufficient conditions of Theorem 1.1 utilize the inclusion relations
between Lp−Sobolev and Wiener amalgam spaces W p,q described in [5]. For (1/p, 1/q) ∈
[0, 1] × [0, 1], the indices τ1(p, q) and τ2(p, q) are defined as follows:
τ1(p, q) =
if (1/p, 1/q) ∈ I ∗
0
1/p + 1/q − 1 if (1/p, 1/q) ∈ I ∗
if (1/p, 1/q) ∈ I ∗
1/q − 1/2
1 : min(1/p′, 1/2) ≥ 1/q
2 : min(1/q, 1/2) ≥ 1/p′
3 : min(1/p′, 1/q) ≥ 1/2
ON Lp−BOUNDEDNESS OF PSEUDO-DIFFERENTIAL OPERATORS OF SJ OSTRAND'S CLASS
5
1/q
1
1/2
I ∗
3
I ∗
1
1/q
1
I ∗
2
1/2
I2
I1
I3
0
1/2
1
1/p
0
1/2
1
1/p
Figure 1. The index sets for Lp
s − W p,q inclusion
τ2(p, q) =
if (1/p, 1/q) ∈ I1 : max(1/p′, 1/2) ≤ 1/q
0
1/p + 1/q − 1 if (1/p, 1/q) ∈ I2 : max(1/q, 1/2) ≤ 1/p′
if (1/p, 1/q) ∈ I3 : max(1/p′, 1/q) ≤ 1/2
1/q − 1/2
where 1/p + 1/p′ = 1 = 1/q + 1/q′. Refer to Figure 1 for a visualization.
Theorem 3.1 ([5]). Let 1 ≤ p, q ≤ ∞ and s ∈ R. Then Lp
conditions is satisfied.
s ֒→ W p,q if one of the following
(1) p > q, q < 2 and s > nτ1(p, q);
(2) p 6= 1, max(1/p, 1/2) ≥ 1/q and s ≥ nτ1(p, q);
(3) p = 1, q = ∞ and s ≥ nτ1(1, ∞);
(4) p = 1, q 6= ∞ and s > nτ1(1, q).
Theorem 3.2 ([5]). Let 1 ≤ p, q ≤ ∞ and s ∈ R. Then W p,q ֒→ Lp
conditions is satisfied.
s if one of the following
(1) p < q, q > 2 and s < nτ2(p, q);
(2) p 6= ∞, min(1/p, 1/2) ≤ 1/q and s ≤ nτ2(p, q);
(3) p = ∞, q = 1 and s ≤ nτ2(∞, 1);
(4) p = ∞, q 6= 1 and s < nτ2(1, q).
In the work of Cordero, Tabacco and Wahlberg [4, Corollary 3.11], they showed how
symbols in modulation spaces M p,q(R2n) produces bounded Fourier integral operators in
Wiener amalgam spaces W p,q(Rn). Choosing the phase Φ(x, ξ) = x · ξ, this result reduces
to boundedness of pseudo-differential operators of Kohn-Nirenberg correspondence. For
the ease of reference, we write a particular case of their result as follows.
Corollary 3.1. Given σ ∈ M ∞,1(R2n). Then the operator σ(x, D) extends to a bounded
operator on Wiener amalgam spaces W p,q(Rn).
6
JAYSON CUNANAN
Proof of IF part on Theorem 1.1. First, we let
q = 2 if 1 ≤ p ≤ 2.
Then using the embedding results Theorem 3.1 and Theorem 3.2, we can say that W p,q ֒→
Lp and Lp
s ֒→ W p,q whenever s ≥ n(1/p + 1/q − 1) for 1 ≤ p ≤ 2. The desired boundedness
now follows from the following commutative diagram.
W p,q
Inclusionx
Lp
s
σ
−−−−→ W p,q
yInclusion
σ
−−−−→ Lp
Substituting the appropriate value of q we have s ≥ n(1/p − 1/2) for 1 ≤ p ≤ 2. By
Remark 2.1 the result for p > 2 follows by duality on the corresponding Weyl quantization.
(cid:3)
Remark 3.1. The inclusion relations between Lp−Sobolev spaces and modulation
spaces M p,q are known due to [14]. Using a similar argument as above, one can conclude
boundedness results but with worse estimates for s. Namely, s ≥ 2n1/p − 1/2.
4. Necessary condition
Boundedness results of pseudo-differential operators having symbols in weighted mod-
ulation spaces acting on Lp spaces is equivalent to boundedness results of operators with
symbols in unweighted modulation spaces acting on Lp−Sobolev spaces. More specifically,
we have the following proposition.
Proposition 4.1. Let σ(x, D) and σ(x, D) be pseudodifferential operators such that
σ(x, ξ) = σ(x, ξ)hξis,
x, ξ ∈ Rn, s ∈ R.
Then
(1) the operator σ(x, D) is bounded from Lp(Rn) to itself if and only if σ(x, D) is
bounded from Lp
s(Rn) to Lp(Rn).
(2)
σ ∈ M ∞,1
ν0,s⊗1(R2n) ⇐⇒ σ ∈ M ∞,1(R2n).
Proof. The proof of (1) is a direct consequence of the following commutative diagram.
Indeed, the vertical arrows define isomorphisms:
Lp
hDisx
Lp
s
σ
−−−−→ Lp
yIdentity
σ−−−−→ Lp
Also, by [19, Corollary 2.3], multiplication by a weight function η(x, ξ, ζ1, ζ2) = hξis,
x, ξ, ζ1, ζ2 ∈ Rn is an isomorphism from M ∞,1
(cid:3)
ν0,s⊗1(R2n) to M ∞,1(R2n), thus (2).
ON Lp−BOUNDEDNESS OF PSEUDO-DIFFERENTIAL OPERATORS OF SJ OSTRAND'S CLASS
7
Remark. For the general case of Fourier integral operators of Proposition 4.1, we refer
the reader to [3, Proposition 2.8], where the authors rephrased boundedness of FIO with
symbols in weighted modulation spaces acting on unweighted modulation spaces by FIO
with symbols in unweighted modulation spaces acting on weighted modulation spaces.
We recall the following theorem by Hormander which gives the critical decreasing order
for the Lp boundedness of pseudo-differential operators σ(x, D) in Op(Ss
ρ,δ).
Theorem 4.1 (Hormander). Let 0 ≤ δ ≤ ρ ≤ 1 and δ < 1. Then
Op(Ss
ρ,δ) ⊂ L(Lp(Rn)) =⇒ s ≤ −n(1 − ρ)1/p − 1/2
Proof of the converse of Theorem 1.1. By Proposition 4.1 it suffice to show that
(3)
Op(M ∞,1
ν0,s⊗1(R2n)) ⊂ L(Lp(Rn)) =⇒ s ≥ n1/p − 1/2.
In [12], Grochenig and Heil proved the following embedding S0
fact that,hDi−sM ∞,1(R2n) = M ∞,1
0,0 ֒→ M ∞,1
we conclude that S−s
4.1 gives the desired estimate.
0,0 ֒→ M ∞,1. Using the
0,0) whenever σ ∈ S0
0,0,
ν0,s⊗1(R2n). Using this inclusion on (3) together with Theorem
ν0,s⊗1(R2n), σ(x, D)hDi−s ∈ Op(S−s
(cid:3)
References
[1] R. Ashino, M. Nagase, and R. Vaillancourt. Pseudodifferential operators in Lp(Rn) spaces,. Technical
report, Centre de recherches mathematiques, 2003.
[2] A. B´enyi, K. Grochenig, K. Okoudjou, and L. Rogers. Unimodular Fourier multipliers for modulation
spaces. J. Funct. Anal., 246(2):366 -- 384, 2007.
[3] E. Cordero and F. Nicola. Boundedness of Schrodinger type propagators on modulation spaces. J.
Fourier Anal. Appl., 16(3):311 -- 339, 2010.
[4] E. Cordero, A. Tabacco, and P. Wahlberg. Schrodinger-type propagators, pseudodifferential operators
and modulation spaces. J. London Math. Soc., 88(2):375 -- 395, 2013.
[5] J. Cunanan, M. Sugimoto, and M. Kobayashi. Inclusion relations between Lp-Sobolev and Wiener
amalgam spaces. J. Funct. Anal., 419(2):738 -- 747, 2014.
[6] C. Fefferman. Lp-bounds for pseudo-differential operators. Israel Journal of Mathematics, 14(4):413 --
417, 1973.
[7] H. G. Feichtinger. Banach convolution algebras of wiener's type. In Proc. Conf. "Function, Series,
Operators", Budapest, Colloq. Math. Soc. J´anos Bolyai, pages 509 -- 524. North-Holland, Amsterdam,
1980.
[8] H. G. Feichtinger. Banach spaces of distributions of wiener type and interpolation. In Functional
Analysis and Approximation, volume 60, pages 153 -- 165. Birkhauser Basel, 1981.
[9] H. G. Feichtinger. Modulation spaces on locally compact abelian groups. Technical report, University
of Vienna, 1983.
[10] H. G. Feichtinger. Generalized amalgams, with applications to Fourier transform. Canad. J. Math.,
42(3):395 -- 409, 1990.
[11] K. Grochenig. Time-Frequency Analysis of Sjostrand's Class. Rev. Mat. Iberoamericana, 22(2):703 -- 724,
09 2006.
[12] K. Grochenig and C. Heil. Modulation spaces and pseudodifferential operators. Integral Equations
Operator Theory, 34(4):439 -- 457, 1999.
[13] L. Hormander. Pseudo-differential operators and hypoelliptic equations. In Singular integrals (Proc.
Sympos. Pure Math., Vol. X, Chicago, Ill., 1966), pages 138 -- 183. Amer. Math. Soc., Providence, R.I.,
1967.
8
JAYSON CUNANAN
[14] M. Kobayashi and M. Sugimoto. The inclusion relation between Sobolev and modulation spaces. J.
Funct. Anal., 260(11):3189 -- 3208, 2009.
[15] A. P. Calderon and R. Vaillancourt. On the boundedness of pseudo-differential operators. J. Math.
Soc. Japan, 23(2):374 -- 378, 04 1971.
[16] J. Sjostrand. An algebra of pseudodifferential operators. Math. Res. Lett., 1(2):185 -- 192, 1994.
[17] J. Sjostrand. Wiener type algebras of pseudodifferential operators. In S´eminaire sur les ´Equations aux
D´eriv´ees Partielles, 1994 -- 1995, pages Exp. No. IV, 21. ´Ecole Polytech., Palaiseau, 1995.
[18] E. M. Stein. Harmonic Analysis. Princeton University Press, Princeton, 1993.
[19] J. Toft. Continuity properties for modulation spaces, with applications to pseudo-differential calculus-
II. Ann. Global Anal. Geom., 26:73 -- 106, 2004.
Graduate School of Mathematics, Nagoya University, Furocho, Chikusa-ku, Nagoya 464-
8602, Japan
E-mail address: [email protected]
|
1708.07345 | 1 | 1708 | 2017-08-24T10:25:28 | Reducibility of WCE operators on $L^2(\mathcal{F})$ | [
"math.FA"
] | In this paper we characterize the closed subspaces of $L^2(\mathcal{F})$ that reduce the operators of the form $E^{\mathcal{A}}M_u$, in which $\mathcal{A}$ is a $\sigma$- subalgebra of $\mathcal{F}$. We show that $L^2(A)$ reduces $E^{\mathcal{A}}M_u$ if and only if $E^{\mathcal{A}}(\chi_A)=\chi_A$ on the support of $E^{\mathcal{A}}(|u|^2)$, where $A\in \mathcal{F}$. Also, some necessary and sufficient conditions are provided for $L^2(\mathcal{B})$ to reduces $E^{\mathcal{A}}M_u$, for the $\sigma$- subalgebra $\mathcal{B}$ of $\mathcal{F}$. | math.FA | math |
REDUCIBILITY OF WCE OPERATORS ON L2(F )
Y. ESTAREMI
Abstract. In this paper we characterize the closed subspaces of
L2(F ) that reduce the operators of the form EAMu, in which A is a
σ- subalgebra of F . We show that L2(A) reduces EAMu if and only
if EA(χA) = χA on the support of EA(u2), where A ∈ F . Also,
some necessary and sufficient conditions are provided for L2(B) to
reduces EAMu, for the σ- subalgebra B of F .
1. Introduction
Let (X, F , µ) be a complete σ-finite measure space. For any sub-
σ-finite algebra A ⊆ F , the L2-space L2(X, A, µA) is abbreviated by
L2(A), and its norm is denoted by k.k2. All comparisons between two
functions or two sets are to be interpreted as holding up to a µ-null
set. The support of a measurable function f is defined as S(f ) =
{x ∈ X; f (x) 6= 0}. We denote the vector space of all equivalence
classes of almost everywhere finite valued measurable functions on X by
L0(F ). For a sub-σ-finite algebra A ⊆ F , the conditional expectation
operator associated with A is the mapping f → EAf , defined for all
non-negative measurable function f as well as for all f ∈ L2(F ), where
EAf , by the Radon-Nikodym theorem, is the unique A-measurable
function satisfying
ZA
f dµ = ZA
EAf dµ,
∀A ∈ A.
As an operator on L2(F ), EA is idempotent and EA(L2(F )) = L2(A).
This operator will play a major role in our work.
Composition of conditional expectation operators and multiplication
operators appear often in the study of other operators such as multi-
plication operators and weighted composition operators. Specifically,
in [9], S.-T. C. Moy characterized all operators on Lp of the form
f → EA(f g) for g in Lq with EA(g) bounded. And R. G. Douglas, [1],
analyzed positive projections on L1 and many of his characterizations
are in terms of combinations of multiplication operators and conditional
expectations. Also, P.G. Dodds, C.B. Huijsmans and B. De Pagter,
[2], extended these characterizations to the setting of function ideals
2010 Mathematics Subject Classification. 47A15.
Key words and phrases. conditional expectation operator, invariant subspace,
reducing subspace, sigma subalgebra.
1
2
Y. ESTAREMI
and vector lattices. Some other people studied weighted conditional
expectation operators on measurable function spaces in [6, 7, 8] and
references therein. Moreover, we investigated some classic properties
of weighted conditional expectation operators of the form MwEAMu
on Lp spaces in [3, 4, 5]. In this paper we characterize reducible closed
subspaces of L2(F ) for the weighted conditional expectation operators
of the form EAMu.
2. Reducibility of WCE operators
Throughout this section we assume that u ∈ D(EA) := {f ∈ L0(F ) :
EA(f ) ∈ L0(A)}. Now we recall the definition of weighted conditional
expectation operators on L2(F ).
Definition 2.1. Let (X, F , µ) be a σ-finite measure space and let A
be a σ-subalgebra of F such that (X, A, µA) is also σ-finite. Let EA be
the conditional expectation operator relative to A. If u ∈ L0(F ) such
that uf is conditionable and EA(uf ) ∈ L2(F ) for all f ∈ D ⊆ L2(F ),
where D is a linear subspace, then the corresponding weighted condi-
tional expectation (or briefly WCE) operator is the linear transforma-
tion EAMu : D → L2(F ) defined by f → EA(uf ).
For a bounded WCE operator T = EAMu on the Hilbert space L2(F )
we have
T ∗ = M¯uEA, T ∗T = M¯uEAMu,
T T ∗ = EAMEA(u2).
In this section we will be concerned with the reducibility of the WCE
operators of the form EAMu. Recall that an operator T on a Hilbert
space H is reducible if there is a proper closed subspace M of H, other
that {0} and H, for which T M ⊆ M and T ∗M ⊆ M. In this case we
say that M reduces T . Equivalently, M reduces T if and only if the
orthogonal projection P onto M commutes with T . We shall make use
of the fact that if a projection commutes with T , then it commutes
with any polynomial in T and T ∗. Now we have the next proposition.
Proposition 2.2. Let T = EAMu be a bounded operator on L2(F ),
M ⊆ L2(F ) be a subspace such that reduces EAMu and P be the or-
thogonal projection onto M. Then we have the followings:
(a) EA(uP f ) = P EA(uf ) for every f ∈ L2(F ).
(b) ¯uEA(P f ) = P (¯uEA(f )) for every f ∈ L2(F ).
(c) EA(u2)EA(P f ) = P EA(EA(u2)f ) for every f ∈ L2(F ).
(d) ¯uEA(uP f ) = P (¯uEA(uf )) for every f ∈ L2(F ).
(e) P (EA(u2)EA(uf )) = EA(u2)P (EA(uf )) for every f ∈ L2(F ).
3
Proof. The items (a), (b), (c) and (d) follow from the fact that P
commute with T , T ∗, T T ∗ and T ∗T , respectively. By applying T to
the equation in item (c) we get the item (e).
(cid:3)
Now by using Proposition 2.2 we give a necessary and sufficient con-
ditions under which the subspace L2(A) reduces EAMu for A ∈ F .
Theorem 2.3. Let EAMu be bounded on L2(F ) and A ∈ F . Then
L2(A) reduces EAMu if and only if EA(χA) = χA on S(EA(u2)).
Proof. If L2(A) reduces EAMu, then by the Proposition 2.2 part (c)
we have EA(u2)EA(χAf ) = EA(u2)χAEA(f ) for all f ∈ L2(F ). Let
{Fn} be a sequence in F increasing to X with µ(Fn) < ∞, and put
fn = χFn. Then fn ∈ L2(F ) and fn ↑ 1, hence EA(fn) ↑ 1 and
EA(χAfn) ↑ EA(χA). Thus we get that EA(u2)EA(χA) = EA(u2)χA
and so EA(χA) = χA on S(EA(u2)). Now we prove the converse,
let EA(χA) = χA on S(EA(u2)). By conditional type Holder in-
equality we have that S(EA(uf )) ⊆ S(EA(u2)) and S(EA(χA)uf ) ⊆
S(EA(u2)) for all f ∈ L2(F ). Therefore we have
χAEA(uf ) = EA(EA(χA)uf ) = EA(χAuf ),
this completes the proof.
(cid:3)
Let B ⊆ F be a σ- subalgebra.
In the next theorem we give a
sufficient condition under which the subspace L2(B) reduces EAMu.
Theorem 2.4. Let EAMu be bounded on L2(F ). Then for every σ-
subalgebra B ⊆ F such that B ⊆ A or A ⊆ B and u is B-measurable,
the subspace L2(B) reduces EAMu.
Proof. If we decompose L2(F ) as a direct sum L2(B) ⊕ K(EB), then
the corresponding block matrix of EAMu is as follow:
EAMu = (cid:18)
EBEAMuEB
(I − EB)EAMuEB (I − EB)EAMu(I − EB) (cid:19).
EBEAMu(I − EB)
If B ⊆ A or A ⊆ B, then in both cases we have EBEA = EAEB. Hence
we get that
EBEAMu(I − EB) = EA(EBMu − EBMuEB).
Since u is B-measurable, then we have EBEAMu(I −EB) = 0. Similarly
we have (I − EB)EAMuEB = 0. These imply that L2(B) reduces
EAMu.
(cid:3)
Let A, B ⊆ F be σ-sub algebras. In the next theorem we determine
the relation between A and B if L2(B) reduces EAMu and the converse.
Theorem 2.5. Let EAMu be bounded on L2(F ). Then we have the
followings:
4
Y. ESTAREMI
(a) If B ⊆ F is a sigma subalgebra such that EBEA = EB∩A and u
is B-measurable, then the subspace L2(B) reduces EAMu.
(b) If L2(B) reduces EAMu, then EBEAf = EAEBf for every f ∈
L2(S(EA(u2))) and u is B-measurable.
(c) If S(EA(u2)) = X, then L2(B) reduces EAMu if and only if
EBEA = EB∩A and u is B-measurable.
Proof. (a) The condition EBEA = EB∩A implies that EBEA = EAEB.
Hence by the same method of Theorem 2.4 we get the result.
(b) Suppose that L2(B) reduces EAMu. Then by Proposition 2.2
part (b) we have EBM¯uEA = M¯uEAEB. Let {Fn} be a sequence in F
increasing to X with µ(Fn) < ∞, and put fn = χFn. Then fn ∈ L2(F )
and fn ↑ 1, hence EA(fn) ↑ 1. These observations shows that ¯u should
be B-measurable. Similarly by Proposition 2.2 part (c) we get that
EA(u2) is B- measurable. Again by part (c) of Proposition 2.2 we get
that EBEAf = EAEBf for every f ∈ L2(S(EA(u2))).
(c) It is a direct consequence of parts (a) and (b).
(cid:3)
Here we find a σ-subalgebra B of A such that L2(B) reduces EAMu.
Theorem 2.6. Let A ⊆ F be a σ-algebra and EAMu be a bounded
operator on L2(F ). Then there exists a σ-subalgebra B ⊆ F such that
L2(B) reduces EAMu.
Proof. Let B be the σ-algebra generated by the sets in u−1(C) ∪ A,
in which C is the σ-algebra of C such that u : (X, F ) → (C, C) is
If u is A-measurable, then B ⊆ A and if u is not A-
measurable.
measurable, then A ⊆ B. In both cases we have EAEB = EBEA = EA
and u is B-measurable. Hence by Theorem 2.4 we get that L2(B)
reduces EAMu.
(cid:3)
Now we recall a fundamental lemma from general operator theory.
Lemma 2.7. Let (X, F , µ) be a finite measure space, T ∈ B(L2(F ))
and S be a closed operator on L2(F ). If T = S on a dense subset of
L2(F ), then S is bounded and T = S.
A ∗-subalgebra of B(L2(F )) is called a Von Neumann algebra on the
Hilbert space L2(F ) if it is closed in strong operator topology (SOT).
Now we get a Von Neumann algebra of WCE operators in the next
theorem.
Theorem 2.8. If (X, F , µ) is a finite measure space and A ⊂ F is a
σ-subalgebra, then W = {EAMg : g ∈ L∞(A)} is a unital commutative
Von Neumann algebra with unit EA.
5
Proof. It is easy to see that W is a self-adjoint operator subalgebra
of B(H). Then we only need to prove W is strongly closed. Let
{EAMuα}α ⊆ W and T ∈ B(L2(F )) such that kEA(uαf ) − T (f )kL2 →
0 for all f ∈ L2(F ). Hence for the constant function 1 we have
kuα − T (1)kL2 → 0 and so T (1) is A-measurable. Also, for every
f ∈ L∞(F ) and α we have
kT (1)EA(f ) − T (f )kL2 ≤ kT (1)EA(f ) − EA(uαf )kL2 + kEA(uαf ) − T (f )kL2
≤ kT (1) − uαkL2kf k∞ + kEA(uαf ) − T (f )kL2.
This implies that T = EAMT (1) on L∞(F ). Since L∞(F ) is dense in
L2(F ) and EAMT (1) is closed, then by Lemma 2.7 we get that EAMT (1)
is bounded and T = EAMT (1). Therefore W is strongly closed and
consequently is a unital commutative Von Neumann algebra with unit
EA.
(cid:3)
Let M be a closed subspace of L2(F ) that reduces EAMu. Let B0
be the σ- subalgebra generated by {f −1(C) : f ∈ M} ∪ u−1(C), B1 be
the σ-algebra generated by {f −1(C) : f ∈ M} ∪ u−1(C) ∪ A and B2 be
the σ-algebra generated by {f −1(C) : f ∈ M} ∪ A. Then B0 ⊆ B1,
M ⊆ L2(B0) ⊆ L2(B1) and the subspaces L2(B0), L2(B1) are reduc-
ing subspaces for EAMu. Also, we have B0 ⊆ B2 ⊆ B1, if u is A-
measurable, moreover L2(B2) reduces EAMu.
In the sequel we provide some necessary and sufficient conditions for
a reducing subspace to be of the form L2(A) or L2(B).
Theorem 2.9. Let (X, F , µ) be a finite measure space, M ⊆ L2(F ) be
a reducing subspace for EAMu, P be the orthogonal projection onto M
and p = P (1). Then there exists a σ- algebra M ⊆ A such that L2(M)
reduces EAMu and the followings hold:
(a) If S(EA(u2)) = X, then P f = EA(p)f , for all f ∈ L2(M).
(b) If M = L2(A) for some A ∈ F , then M is equal to the σ- algebra
generated by AA, in which AA = {B ∈ A : A ∩ B ∈ A}.
(c) If M = A, then EA(M) = f.L2(A) ⊆ L1(S(f )), for some
f ∈ L2(A).
(d) L2(M) ⊆ M if and only if 1 ∈ M.
(e) L2(M) ⊆ M ⊥ if and only if 1 ∈ M ⊥.
Proof. Put
W = {EAMg : g ∈ L∞(A),
EAMgP = P EAMg}.
6
Y. ESTAREMI
It is easy to see that W is a Von Neumann subalgebra of {EAMg :
g ∈ L∞(A)}. Let
M0 = {A ∈ A : EAMχA ∈ W }.
Easily we have that ∅ ∈ M0 and A ∩ B ∈ M0, for all A, B ∈ M0. Also
we have W = {EAMg : g ∈ L∞(M0)}. If we put M = M0 ∪ {X},
then M is a σ- subalgebra of A. Also, M0 is a σ- algebra when EAP =
P EA, in this case we set M = M0. In general we suppose that M is
the σ- algebra generated by M0. Now we show that L2(M) reduces
EAMu, for this, it suffices to show that L∞(M0) reduces EAMu. Let
g ∈ L∞(M0) and f ∈ L2(F ). Then
Also we have
EAMEA(u)gP f = EAMuEAMgP f
= EAMuP EAMgf
= P EAMuP EAMgf
= P EAMEA(u)gf.
P EAMugf = P EAMgEAMuf
= EAMgP EAMuf
= EAMgEAMuP f
= EAMugP f.
By these observations we get that L∞(M0) is invariant under EAMu
and MuEA. Hence we get that L∞(M) reduces EAMu. Consequently,
L2(M) reduces EAMu
Also for f ∈ L2(M0) we have
P (f ) = P EA(f.1)
= P EAMf (1)
= EAMf P (1)
= f.EA(P (1))
= MEA(P (1))(f ).
Hence we have (a).
(b) Suppose that M = L2(A) for some A ∈ F . Then P = MχA and
M0 = {B ∈ A : A ∩ B ∈ A} = AA. Thus M is equal to the σ- algebra
generated by AA.
(c) If M = A, then by part (a) for every f ∈ L2(A) we have P (f ) =
EA(p)f . Also, for every B ∈ A \ {X} we have B ∈ M0 and so for all
7
f ∈ L2(F ),
EA(χBP (f )) = P (EA(χBf )) = EA(p)χBEA(f ).
This implies that EAP (f ) = EA(p)EA(f ) for all f ∈ L2(F ). Therefore
EA(M) = EA(p).L2(A).
(d) If L2(M) ⊆ M, then 1 ∈ L2(M) ⊆ M. Conversely, if 1 ∈ M,
then p = P (1) = 1 and so EA(p) = EA(1) = 1. Hence by (a) we get
that P (f ) = f for all f ∈ L2(M). This means that L2(M) ⊆ M.
(e) If L2(M) ⊆ M ⊥, then 1 ∈ L2(M) ⊆ M ⊥. Conversely, if 1 ∈ M ⊥,
then p = P (1) = 0 and so P (f ) = 0.f , for all f ∈ L2(M). This implies
that L2(M) ⊆ M ⊥.
(cid:3)
References
[1] R. G. Douglas, Contractive projections on an L1 space, Pacific J. Math. 15
(1965), 443-462.
[2] P. G. Dodds, C.B. Huijsmans and B. De Pagter, Characterizations of condi-
tional expectation-type operators, Pacific J. Math. 141 (1990), 55-77.
[3] Y. Estaremi, Some classes of weighted conditional type operators and their
spectra, Positivity 19 (2015), 83-93.
[4] Y. Estaremi, On properties of Multiplication conditional type operators be-
tween Lp-space, Filomat. Preprint.
[5] Y. Estaremi and M. R. Jabbarzadeh, Weighted lambert type operators on
Lp-spaces, Oper. Matrices 1 (2013), 101-116.
[6] J. J. Grobler and B. de Pagter, Operators representable as multiplication-
conditional expectation operators, J. Operator Theory 48 (2002), 15-40.
[7] Herron. J. Weighted conditional expectation operators, Oper. Matrices 1
(2011), 107-118.
[8] A. Lambert, Lp multipliers and nested sigma-algebras, Oper. Theory Adv.
Appl. 104 (1998), 147-153.
[9] Shu-Teh Chen, Moy, Characterizations of conditional expectation as a trans-
formation on function spaces, Pacific J. Math. 4 (1954), 47-63.
[10] M. M. Rao, Conditional measure and applications, Marcel Dekker, New York,
1993.
Y. Estaremi
E-mail address: [email protected]
Department of Mathematics, Payame Noor University , P. O. Box:
19395-3697, Tehran, Iran
|
1703.02260 | 1 | 1703 | 2017-03-07T08:00:35 | Strong factorizations of operators with applications to Fourier and Ces\'aro transforms | [
"math.FA"
] | Consider two continuous linear operators $T\colon X_1(\mu)\to Y_1(\nu)$ and $S\colon X_2(\mu)\to Y_2(\nu)$ between Banach function spaces related to different $\sigma$-finite measures $\mu$ and $\nu$. We characterize by means of weighted norm inequalities when $T$ can be strongly factored through $S$, that is, when there exist functions $g$ and $h$ such that $T(f)=gS(hf)$ for all $f\in X_1(\mu)$. For the case of spaces with Schauder basis our characterization can be improved, as we show when $S$ is for instance the Fourier operator, or the Ces\`aro operator. Our aim is to study the case when the map $T$ is besides injective. Then we say that it is a~representing operator ---in the sense that it allows to represent each elements of the Banach function space $X(\mu)$ by a~sequence of generalized Fourier coefficients---, providing a complete characterization of these maps in terms of weighted norm inequalities. Some examples and applications involving recent results on the Hausdorff-Young and the Hardy-Littlewood inequalities for operators on weighted Banach function spaces are also provided. | math.FA | math |
STRONG FACTORIZATIONS OF OPERATORS WITH APPLICATIONS
TO FOURIER AND CES ´ARO TRANSFORMS
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
Abstract. Consider two continuous linear operators T : X1(µ) → Y1(ν) and S : X2(µ) →
Y2(ν) between Banach function spaces related to different σ-finite measures µ and ν. We
characterize by means of weighted norm inequalities when T can be strongly factored through
S, that is, when there exist functions g and h such that T (f ) = gS(hf ) for all f ∈ X1(µ).
For the case of spaces with Schauder basis our characterization can be improved, as we show
when S is for instance the Fourier operator, or the Ces`aro operator. Our aim is to study
the case when the map T is besides injective. Then we say that it is a representing operator
-- in the sense that it allows to represent each elements of the Banach function space X(µ)
by a sequence of generalized Fourier coefficients -- , providing a complete characterization
of these maps in terms of weighted norm inequalities. Some examples and applications
involving recent results on the Hausdorff-Young and the Hardy-Littlewood inequalities for
operators on weighted Banach function spaces are also provided.
1. Introduction
Let X1(µ), X2(µ), Y1(ν), Y2(ν) be Banach function spaces related to different σ-finite
measures µ, ν and consider two continuous linear operators T : X1(µ) → Y1(ν), S : X2(µ) →
Y2(ν). In this paper we provide a characterization in terms of weighted norm inequalities
of when T can be factored through S via multiplication operators, that is, when there are
functions g and h satisfying that T (f ) = gS(hf ) for all f ∈ X1(µ).
This problem was studied in [6] for the case when µ and ν are the same finite measure.
However, the results developed there do not allow to face the problem we study here, in
which different σ-finite measures µ and ν appear in order to consider the relevant case of
the classical sequence spaces ℓp. The reason is that we are interested in considering standard
cases as the Fourier and the Ces`aro operators, that will be in fact our main examples.
In this direction, we will show that in the case when the Kothe dual Y1(ν)′ of Y (ν) and
X1(µ) have Schauder basis, the norm inequality which characterizes the factorization of T
through S can be weakened. After showing this, we will develop with some detail some the
examples regarding Fourier operators, operators factoring though infinite matrices and the
2010 Mathematics Subject Classification. Primary 46E30, 47B38; Secondary 46B15, 43A25.
Key words and phrases. Banach function space, Strong factorization, Schauder basis, Fourier operator,
representing operator, Ces`aro operator.
The first author gratefully acknowledge the support of the Ministerio de Econom´ıa y Competitividad
(project #MTM2015-65888-C4-1-P) and the Junta de Andaluc´ıa (project FQM-7276), Spain. The second au-
thor was supported by National Science Centre, Poland, project no. 2015/17/B/ST1/00064. The third author
acknowledges with thanks the support of the Ministerio de Econom´ıa y Competitividad (project MTM2016-
77054-C2-1-P), Spain.
1
2
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
Ces`aro operator. This will allow to introduce the notion of representing operator and to
study it in the second part of the paper.
Let us explain briefly this notion. With the notation introduced above, assume that Y1(ν)
and Y2(ν) have unconditional basis U1 := {vi : i ∈ N} and U2 := {ei : i ∈ N}, respectively.
Suppose that there exists a Schauder basis B : {fi :∈ N} for the space X2(µ) and write
αi(f ) for the i-th basic coefficient of f ∈ X2(µ), i ∈ N. We will say that an operator
T : X1(µ) → Y1(ν) is a representing operator for X1(µ) on Y1(ν) (associated to the basis B of
X2(µ)) if each element x ∈ X1(µ) can be represented univocally by a sequence of coefficients
i=1 βi(x)vi ∈ Y1(ν), where the coefficients βi(x) can be computed by means
of the associated values of αi by a simple transformation provided by multiplication operators.
Thus, the last part of the paper is devoted to find a characterization of such operators
in terms of vector norm inequalities that they must satisfy. We provide also classical and
recently published examples of such kind of maps, using for instance an improvement of
the Hausdorff-Young inequality given in [8], or the continuity of the Fourier operator Hp :
Lp[−π, π] → ℓp(W ) -- where ℓp(W ) is a weighted ℓp-space -- that can be found in [1].
(βi(x)) such thatP∞
2. Preliminaries
Let (Ω, Σ, µ) be a σ-finite measure space and denote by L0(µ) the space of all measurable
real functions defined on Ω, where functions which are equal µ-a.e. are identified. By a Banach
function space we mean a Banach space X(µ) ⊂ L0(µ) with norm k · kX satisfying that if
f ∈ L0(µ), g ∈ X(µ) and f ≤ g µ-a.e. then f ∈ X(µ) and kfkX ≤ kgkX . In particular
X(µ) is a Banach lattice for the µ-a.e. pointwise order, in which the convergence in norm
of a sequence implies the convergence µ-a.e. for some subsequence. Note that every positive
linear operator between Banach lattices is continuous, (see [10, p. 2]). So, all inclusions
between Banach function spaces are continuous. General information about Banach function
spaces can be found for instance in [17, Ch. 15] considering the function norm ρ defined there
as ρ(f ) = kfkX if f ∈ X(µ) and ρ(f ) = ∞ in other case.
A Banach function space X(µ) is said to be saturated if there is no A ∈ Σ with µ(A) > 0
such that f χA = 0 µ-a.e. for all f ∈ X(µ). This is equivalent to the existence of a function
g ∈ X(µ) such that g > 0 µ-a.e.
Given two Banach function spaces X(µ) and Y (µ), the Y (µ)-dual space of X(µ) is defined
by
X Y =(cid:8)h ∈ L0(µ) : f h ∈ Y (µ) for all f ∈ X(µ)(cid:9).
Every h ∈ X Y defines a continuous multiplication operator Mh : X(µ) → Y (µ) via Mh(f ) =
f h for all f ∈ X(µ). The space X Y is a Banach function space with norm
khkX Y = sup
f ∈BX khfkY ,
h ∈ X Y ,
if and only if X(µ) is saturated. As usual BX denotes the closed unit ball of X(µ). Note that
X L1
is just the classical Kothe dual space X(µ)′ of X(µ). If X(µ) is saturated then X(µ)′ is
also saturated. This does not hold in general for X Y . For issues related to generalized dual
spaces see [3] and the references therein.
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
3
A saturated Banach function space X(µ) is contained in its Kothe bidual X(µ)′′ with
kfkX ′′ ≤ kfkX for all f ∈ X(µ). It is known that kfkX ′′ = kfkX for all f ∈ X(µ) if and
only if X(µ) is order semi-continuous, that is, if for every f, fn ∈ X(µ) such that 0 ≤ fn ↑ f
µ-a.e. it follows that kfnkX ↑ kfkX . Even more, X(µ) = X(µ)′′ with equal norms if and only
if X(µ) has the Fatou property, that is, if for every fn ∈ X(µ) such that 0 ≤ fn ↑ f µ-a.e.
and supn kfnkX < ∞, we have that f ∈ X(µ) and kfnkX ↑ kfkX .
Denote by X(µ)∗ the topological dual of a saturated Banach function space X(µ). Every
function h ∈ X(µ)′ defines an element η(h) ∈ X(µ)∗ via hη(h), fi =R hf dν for all f ∈ X(µ).
The map η : X(µ)′ → X(µ)∗ is a continuous linear injection, since the norm of every h ∈ X(µ)′
can be computed as
khkX ′ = sup
f ∈BX(cid:12)(cid:12)(cid:12)Z hf dµ(cid:12)(cid:12)(cid:12)
and so η is an isometry.
It is known that η is surjective if and only if X(µ) is σ-order
continuous, that is, if for every (fn) ⊂ X(µ) with fn ↓ 0 µ-a.e. it follows that kfnkX ↓ 0.
Note that σ-order continuity implies order semi-continuity.
The σ-order continuous part Xa(µ) of a saturated Banach function space X(µ) is the
largest σ-order continuous closed solid subspace of X(µ), which can be described as
Xa(µ) =(cid:8)f ∈ X(µ) : f ≥ fn ↓ 0µ-a.e. implies kfnkX ↓ 0(cid:9).
Also, a function f ∈ Xa(µ) if and only if f ∈ X(µ) satisfies that kf χAnkX ↓ 0 whenever
(An) ⊂ Σ is such that An ↓ with µ(∩An) = 0. Note that Xa(µ) could be the trivial space
as in the case of X(µ) = L∞(µ) when µ is nonatomic. In the case when Xa(µ) is saturated,
Xa(µ) is order dense in L0(µ) and so by the Monotone Convergence Theorem, it follows easily
that Xa(µ)′ = X(µ)′ with equal norms.
The π-product space XπY of two Banach function spaces X(µ) and Y (µ) is defined as the
space of functions h ∈ L0(µ) such that h ≤Pn fngn µ-a.e. for some sequences (fn) ⊂ X(µ)
and (gn) ⊂ Y (µ) satisfyingPn kfnkXkgnkY < ∞. For h ∈ XπY , consider the norm
π(h) = infnXn
kfnkXkgnkYo,
where the infimum is taken over all sequences (fn) ⊂ X(µ) and (gn) ⊂ Y (µ) such that
h ≤ Pn fngn µ-a.e. and Pn kfnkXkgnkY < ∞. The space XπY is a saturated Banach
function space with norm π if and only if X(µ), Y (µ) and X Y ′
are saturated and, in this case,
(XπY )′ = X Y ′
with equal norms (see [5, Proposition2.2]). The calculus of product spaces is
nowadays well-known (see [3, 5, 9, 16]); the reader can find all the information that is needed
on this construction in these papers.
Banach function spaces on the measure space (N,P(N), λ) with the counting measure λ
are are called usually Banach sequence spaces. The classical Banach sequence spaces ℓp for
1 ≤ p ≤ ∞ is saturated which is σ-order continuous if and only if p < ∞. As usual for each
n ∈ N, we denote by (en) the standard unit vector basis in c0.
4
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
We recall the well known easily verified formula (ℓp)ℓq
= ℓspq with equal norms, where
1 ≤ spq =
= ℓp′
pq
if 1 ≤ q < p < ∞
p−q
if 1 ≤ q < p = ∞
q
∞ if 1 ≤ p ≤ q ≤ ∞
.
(2.1)
In particular, (ℓp)′ = (ℓp)ℓ1
p′ = 1).
Note that ℓp has the Fatou property as (ℓp)′′ = ℓp. Also note that spq = 1 if and only if q = 1
and p = ∞.
where p′ denote the conjugate exponent of p ( 1
p + 1
3. Strong factorization of operators on Banach function spaces
Let (Ω, Σ, µ), (∆, Γ, ν) be σ-finite measure spaces, X1(µ), X2(µ), Y1(ν), Y2(ν) saturated
Banach function spaces and T : X1(µ) → Y1(ν), S : X2(µ) → Y2(ν) nontrivial continuous
linear operators. For h ∈ X X2
1 , we will say that T factors strongly through S and Mh if there
exists g ∈ Y
such that the diagram
Y ′′
1
2
X1(µ)
Mh
X2(µ)
T
S
Y1(ν)
i
/ Y1(ν)′′
Mg
/ Y2(ν)
(3.1)
commutes. Here i denotes the inclusion map. Note that if Y1(ν) has the Fatou property the
diagram above looks as
X1(µ)
Mh
X2(µ)
T
S
Y1(ν)
Mg
/ Y2(ν)
In the case when µ and ν are the same finite measure and under certain extra conditions, [6,
Theorem 4.1] characterizes when T factors strongly through S and Mh. In this section we
extend this theorem to our more general setting and improve it by relaxing the conditions.
The extension will be obtained from the following broader factorization result.
is saturated and consider a function h ∈ X X2
1 . The follow-
Theorem 3.1. Assume that Y
ing statements are equivalent:
Y ′′
1
2
(a) There exists a constant C > 0 such that the inequality
nXi=1Z T (xi)y′
i dν ≤ C(cid:13)(cid:13)(cid:13)
nXi=1
i(cid:13)(cid:13)(cid:13)Y2πY ′
1
S(hxi)y′
,
n ∈ N,
holds for every n ∈ N, x1, ..., xn ∈ X1(µ) and y′
1, ..., y′
n ∈ Y1(ν)′.
1)∗ satisfying the following factorization between the operators T
(b) There exists ξ∗ ∈ (Y2πY ′
and S:
X1(µ)
Mh
X2(µ)
T
S
Y1(ν)
i
/ Y1(ν)′′
η
/ Y1(ν)′∗
/ Y2(ν)
Rξ∗
/
/
/
/
9
9
/
/
/
O
O
/
/
/
/
/
4
4
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
5
1i for y2 ∈ Y2(ν) and y′
1i = hξ∗, y2y′
being saturated assures that Y2πY ′
where η is the continuous linear injection of Y1(ν)′′ into Y1(ν)′∗ and Rξ∗ is the continuous
linear operator defined by hRξ∗(y2), y′
1 ∈ Y1(ν)′.
Proof. Note that the condition of Y Y ′′
1 is a saturated
Banach function space. Also note that the map Rξ∗ : Y2(ν) → Y1(ν)′∗ defined in (b) is a well
defined continuous linear operator as
1i ≤ kξ∗k(Y2πY ′
1 ∈ Y1(ν)′.
hRξ∗(y2), y′
for all y2 ∈ Y2(ν) and y′
1 ≤ kξ∗k(Y2πY ′
1 )∗ky2kY2ky′
1)∗ky2y′
1kY2πY ′
1kY ′
2
1
1
(a) ⇒ (b) For every n ∈ N, x1, ..., xn ∈ X1(µ) and y′
1, ..., y′
n ∈ Y1(ν)′ we take the convex
function φ : B(Y2πY ′
1 )∗ → R given by
φ(ξ∗) =
nXi=1
hξ∗, S(hxi)y′
ii
1)∗ we have that φ is
1 )∗. Considering the weak* topology on (Y2πY ′
for all ξ∗ ∈ B(Y2πY ′
a continuous map on a compact convex set. Moreover, from the Hahn-Banach Theorem
there exists ξ∗
nXi=1Z T (xi)y′
i dν − C
φ ∈ B(Y2πY ′
1 )∗ such that
and so, by (a), it follows that φ(ξ∗
nXi=1
(cid:13)(cid:13)(cid:13)
S(hxi)y′
i(cid:13)(cid:13)(cid:13)Y2πY ′
1
φ) ≤ 0.
=Dξ∗
φ,
nXi=1
S(hxi)y′
iE
Since the family F of functions φ defined in this way is concave, Ky Fan's lemma (see for
1 )∗ such that φ(ξ∗) ≤ 0
instance [14, E.4]) guarantees the existence of an element ξ∗ ∈ B(Y2πY ′
for all φ ∈ F. In particular, for every x ∈ X1(µ) and y′ ∈ Y1(ν)′, we have that
By taking −y′ instead of y′, we obtain that
Z T (x)y′ dν ≤ Chξ∗, S(hx)y′i.
−Z T (x)y′ dν ≤ −Chξ∗, S(hx)y′i
hη(cid:0)T (x)(cid:1), y′i = hRCξ∗(cid:0)S(hx)(cid:1), y′i.
Therefore, η(cid:0)T (x)(cid:1) = RCξ∗(cid:0)S(hx)(cid:1) for all x ∈ X1(µ) and the factorization in (b) holds for
Cξ∗ ∈ (Y2πY ′
n ∈ Y1(ν)′ we have that
(b) ⇒ (a) For each n ∈ N and every x1, ..., xn ∈ X1(µ), y′
1, ..., y′
and so
1)∗.
nXi=1Z T (xi)y′
i dν =
ii
=
ii =
nXi=1
nXi=1
hη(cid:0)T (xi)(cid:1), y′
ii =Dξ∗,
nXi=1
hξ∗, S(hxi)y′
nXi=1
≤ kξ∗k(Y2πY ′
S(hxi)y′
S(hxi)y′
hRξ∗(cid:0)S(hxi)(cid:1), y′
iE
nXi=1
i(cid:13)(cid:13)(cid:13)Y2πY ′
.
1
1 )∗(cid:13)(cid:13)(cid:13)
Note that kξ∗k(Y2πY ′
1 )∗ > 0 as T is nontrivial.
(cid:3)
6
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
Note that the condition of Y
Y ′′
1
2
which is equivalent to L∞(ν) ⊂ Y
equivalent to
being saturated is obtained for instance if Y2(ν) ⊂ Y1(ν)′′
. Also note that the condition (a) of Theorem 3.1 is
Y ′′
1
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
nXi=1Z T (xi)y′
i dν(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ C(cid:13)(cid:13)(cid:13)
nXi=1
S(hxi)y′
,
n ∈ N
i(cid:13)(cid:13)(cid:13)Y2πY ′
1
n ∈ Y1(ν)′. Indeed, we only have to take −y′
1, ...,−y′
n
for every x1, ..., xn ∈ X1(µ) and y′
instead of y′
1, ..., y′
n in Theorem 3.1.(a).
1, ..., y′
As a consequence of Theorem 3.1, we obtain the following generalization and improvement
of [6, Theorem 4.1].
Corollary 3.2. Assume that Y
1 ∈ Y1(ν)′. Given h ∈ X X2
and y′
Y ′′
1
2
is saturated and that y2y′
1 ∈ (Y2πY ′
1 , the following statements are equivalent:
1)a for all y2 ∈ Y2(ν)
(a) T factors strongly through S and Mh.
(b) There exists a constant C > 0 such that the inequality
nXi=1Z T (xi)y′
i dν ≤ C(cid:13)(cid:13)(cid:13)
nXi=1
S(hxi)y′
,
n ∈ N
i(cid:13)(cid:13)(cid:13)Y2πY ′
1
holds for all every x1, ..., xn ∈ X1(µ) and y′
1, ..., y′
n ∈ Y1(ν)′.
Proof. First note that (Y2πY ′
Y1(ν)′ we have that 0 < y2y′
1)a is saturated. Indeed, by taking 0 < y2 ∈ Y2(ν) and 0 < y′
1 ∈ (Y2πY ′
1)a. Then,
(Y2πY ′
1)′
a = (Y2πY ′
1)′ = Y
Y ′′
1
2
.
1 ∈
(b) ⇒ (a) From Theorem 3.1 there exists ξ∗ ∈ (Y2πY ′
1)∗ such that
hη(cid:0)T (x)(cid:1), y′i = hRξ∗(cid:0)S(hx)(cid:1), y′i = hξ∗, S(hx)y′i
Y ′′
1
2
a = Y
for all x ∈ X1(µ) and y′ ∈ Y1(ν)′. Denote by eξ∗ the restriction of ξ∗ to (Y2πY ′
1)a is σ-order continuous and eξ∗ ∈ (Y2πY ′
(Y2πY ′
1)′
(Y2πY ′
and y′ ∈ Y1(ν)′, we have that
hη(cid:0)T (x)(cid:1), y′i = hξ∗, S(hx)y′i = heξ∗, S(hx)y′i =Z gS(hx)y′ dν = hη(cid:0)gS(hx)(cid:1), y′i
, that is, heξ∗, zi =R gz dν for all z ∈ (Y2πY ′
a, we can identify eξ∗ with a function g ∈
1)a. Then, for every x ∈ X1(µ)
1)a. Since
1)∗
and so T (x) = gS(hx).
(a) ⇒ (b) Let g ∈ Y
Y ′′
2 = (Y2πY ′
1
1)′ be such that T (x) = gS(hx) for all x ∈ X1(µ). Consider
1)∗ satisfies
1)′ → (Y2πY ′
the continuous linear injection eη : (Y2πY ′
1)∗. Then eη(g) ∈ (Y2πY ′
hReη(g)(cid:0)S(hx)(cid:1), y′i = heη(g), S(hx)y′i =Z gS(hx)y′ dµ
= Z T (x)y′ dµ = hη(cid:0)T (x)(cid:1), y′i
for all x ∈ X1(µ) and y′ ∈ Y1(ν)′ and so Theorem 3.1.(b) holds for ξ∗ =eη(g).
1 ∈ Y1(ν)′
Remark 3.3. Of course, the condition y2y′
holds when Y2πY ′
1 is σ-order continuous. But also this condition is obtained for instance if
any of Y2(ν) or Y1(ν)′ is σ-order continuous. Indeed, suppose that Y2(ν) is σ-order continuous
1)a for all y2 ∈ Y2(ν) and y′
1 ∈ (Y2πY ′
(cid:3)
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
7
and take y2 ∈ Y2(ν) = (Y2)a(ν) and y′
ν(∩An) = 0, we have that
1 ∈ Y1(ν)′. For every (An) ⊂ Σ such that An ↓ with
ky2y′
1χAnkY2πY ′
1 ≤ ky2χAnkY2 · ky′
1kY ′
1 → 0
and so y2y′
1 ∈ (Y2πY ′
1)a. We get the case when Y1(ν)′ is σ-order continuous in a similar way.
Remark 3.4. Note that if the σ-order continuous part Xa(µ) of a saturated Banach function
space X(µ) is also saturated then kxkX = kxkX ′′ for all x ∈ Xa(µ). Indeed, for every x ∈
Xa(µ) we have that kxkX = kxkX ′′
a = kxkX ′′
since Xa(µ)′ = X(µ)′ with equal norms. Then, the norm in Corollary 3.2.(b) can be computed
as
a since Xa(µ) is order semi-continuous and kxkX ′′
nXi=1
(cid:13)(cid:13)(cid:13)
S(hxi)y′
i(cid:13)(cid:13)(cid:13)Y2πY ′
1
= (cid:13)(cid:13)(cid:13)
=
nXi=1
S(hxi)y′
sup
f ∈B
Y ′′
1
2
Y
Z (cid:12)(cid:12)(cid:12)f
S(hxi)y′
i(cid:13)(cid:13)(cid:13)(cid:0)Y
Y ′′
1
2 (cid:1)′
1)′′
i(cid:13)(cid:13)(cid:13)(Y2πY ′
nXi=1
S(hxi)y′
=(cid:13)(cid:13)(cid:13)
nXi=1
i(cid:12)(cid:12)(cid:12) dν.
4. Strong factorization involving Schauder basis
Let (Ω, Σ, µ), (∆, Γ, ν) be σ-finite measure spaces, X1(µ), X2(µ), Y1(ν), Y2(ν) saturated
Banach function spaces and T : X1(µ) → Y1(ν), S : X2(µ) → Y2(ν) nontrivial continuous
linear operators. In this section we assume the existence of a Schauder basis (γn) for Y1(ν)′
and denote by (γ∗
n) the sequence of coefficient functionals with respect to this basis.
Theorem 4.1. Assume that Y
continuous. Given h ∈ X X2
(a) T factors strongly through S and Mh.
Y ′′
1
2
is saturated and that any of Y2(ν) or Y1(ν)′ is σ-order
1 , the following statements are equivalent:
(b) There exists a constant C > 0 such that the inequality
nXi=1Z T (xi)γi dν ≤ C(cid:13)(cid:13)(cid:13)
nXi=1
S(hxi)γi(cid:13)(cid:13)(cid:13)Y2πY ′
1
,
n ∈ N
holds for every x1, ..., xn ∈ X1(µ).
Moreover, if Y2(ν) ⊂ Y1(ν)′′ and the functions (γn) have pairwise disjoint support, then the
condition
(c) There exists a constant C > 0 such that the inequality
Z T (x)γn dν ≤ CZ S(hx)γn dν,
n ∈ N
holds for every x ∈ X1(µ) and n ≥ 1.
implies (a)-(b). In the case when Y
Y ′′
2 = L∞(ν), we have that (c) is equivalent to (a)-(b).
1
Proof. (a) ⇔ (b) From Remark 3.3, we only have to prove that the condition (b) of the
present theorem implies the condition (b) of Corollary 3.2. The converse implication follows
8
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
=
=
Y ′′
1
1
1
i)mkY ′
1
(4.1)
1, ..., y′
S(hxi)(y′
.
1
Since (y′
i)m → y′
by taking y′
(y′
i)m =Pm
k, y′
ii γk. It follows that
i)m dν =
nXi=1Z T (xi)(y′
i in Y1(ν)′ as m → ∞ and
i dν −Z z(y′
i = γi. Let x1, ..., xn ∈ X1(µ) and y′
k=1hγ∗
n ∈ Y1(ν)′. Fix m ∈ N and denote
iiZ T (xi)γk dν
nXi=1
mXk=1
k, y′
hγ∗
mXk=1Z (cid:16) nXi=1
ii T (xi)(cid:17)γk dν
k, y′
hγ∗
mXk=1Z T(cid:16) nXi=1
ii xi(cid:17)γk dν
k, y′
hγ∗
ii xi(cid:17)γk(cid:13)(cid:13)(cid:13)Y2πY ′
≤ C(cid:13)(cid:13)(cid:13)
S(cid:16)h
mXk=1
nXi=1
k, y′
hγ∗
ii S(hxi)γk(cid:13)(cid:13)(cid:13)Y2πY ′
= C(cid:13)(cid:13)(cid:13)
mXk=1
nXi=1
k, y′
hγ∗
= C(cid:13)(cid:13)(cid:13)
i)m(cid:13)(cid:13)(cid:13)Y2πY ′
nXi=1
i)m dν(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)Z z(cid:0)y′
i)m(cid:1) dν(cid:12)(cid:12)(cid:12) ≤ kzkY1ky′
i − (y′
i − (y′
i=1R T (xi)y′
i=1R T (xi)(y′
i)m dν →Pn
for every z ∈ Y1(ν), we have thatPn
= kz(cid:0)y′
i)m(cid:1)kY2πY ′
i − (y′
i − (y′
i)m →Pn
for every z ∈ Y2(ν), we have that Pn
nXi=1Z T (xi)y′
i dν ≤ C(cid:13)(cid:13)(cid:13)
nXi=1
Assume that Y2(ν) ⊂ Y1(ν)′′ and that the functions (γn) have pairwise disjoint support.
Let us see that (c) implies (b). The condition Y2(ν) ⊂ Y1(ν)′′ is equivalent to L∞(ν) ⊂
1(cid:1)′′ ⊂ L∞(ν)′ = L1(ν). Denote by K the continuity
1 ⊂(cid:0)Y2πY ′
1(cid:1)′
2 =(cid:0)Y2πY ′
and so Y2πY ′
Y
constant of the inclusion Y2πY ′
1 ⊂ L1(ν). For every n ∈ N and x1, ..., xn ∈ X1(µ), noting that
i=1 S(hxi)γi =(cid:12)(cid:12)Pn
i=1 S(hxi)γi(cid:12)(cid:12) pointwise (as (γk) have disjoint support), we have that
Pn
nXi=1Z S(hxi)γi dν = CZ (cid:12)(cid:12)(cid:12)
nXi=1Z T (xi)γi dν ≤ C
≤ CK(cid:13)(cid:13)(cid:13)
nXi=1
i=1 S(hxi)(y′
Then, taking limit as m → ∞ in (4.1), we obtain
S(hxi)γi(cid:13)(cid:13)(cid:13)Y2πY ′
other hand, since
kzy′
S(hxi)γi(cid:12)(cid:12)(cid:12) dν
i)mkY ′
i in Y2πY ′
1
i dν as m → ∞. On
(cid:12)(cid:12)(cid:12)Z zy′
i(cid:13)(cid:13)(cid:13)Y2πY ′
1
i − z(y′
i)mkY2πY ′
1
1 ≤ kzkY2ky′
1 as m → ∞.
i=1 S(hxi)y′
S(hxi)y′
.
then (a) implies (c), as if g ∈ Y
is such that T (x) = gS(hx) for
nXi=1
.
1
Y ′′
1
2
If moreover L∞(ν) = Y
all x ∈ X1(µ), it follows that
Y ′′
1
2
Z T (x)γn dν =Z gS(hx)γn dν ≤Z gS(hx)γn dν ≤ kgk∞Z S(hx)γn dν.
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
9
(cid:3)
Now suppose that there is also a Schauder basis (βn) for X1(µ) and denote by (β∗
n) the
sequence of its coefficient functionals. Then, the equivalent inequalities for the strong factor-
ization can be relaxed.
Y ′′
1
2
is saturated and that any of Y2(ν) or Y1(ν)′ is σ-order
Theorem 4.2. Assume that Y
continuous. Given h ∈ X X2
(a) T factors strongly through S and Mh.
(b) There exists g ∈ Y
(c) There exists a constant C > 0 such that the inequality
Y ′′
1
2
such that T (βn) = gS(hβn) for each n ∈ N.
1 , the following statements are equivalent:
nXi=1
mXj=1
rijZ T (βj)γi dν ≤ C(cid:13)(cid:13)(cid:13)
nXi=1
mXj=1
rijS(hβj )γi(cid:13)(cid:13)(cid:13)Y2πY ′
1
,
n, m ∈ N
holds for every (rij) ⊂ Bℓ∞.
Moreover, if Y2(ν) ⊂ Y1(ν)′′ and the functions (γn) have pairwise disjoint support, then the
condition
(d) There exists a constant C > 0 such that the inequality
mXj=1
rjZ T (βj)γn dν ≤ CZ (cid:12)(cid:12)(cid:12)
mXj=1
rjS(hβj )γn(cid:12)(cid:12)(cid:12) dν
holds for every n, m ∈ N and (rj) ⊂ Bℓ∞.
implies (a)-(c). In the case when L∞(ν) = Y
Y ′′
1
2
, we have that (d) is equivalent to (a)-(c).
Proof. (a) ⇒ (b) Let g ∈ Y
for x = βn we obtain (b).
(b) ⇒ (c) Since g ∈ Y
Y ′′
1
2
be such that T (x) = gS(hx) for all x ∈ X1(µ). In particular,
Y ′′
2 = (Y2πY ′
1
1)′, for every n, m ∈ N and (rij) ⊂ Bℓ∞, it follows that
nXi=1
mXj=1
rijZ T (βj)γi dν =
rijS(hβj )γi dν
rijZ gS(hβj )γi dν
mXj=1
nXi=1
= Z g
mXj=1
nXi=1
≤ Z (cid:12)(cid:12)(cid:12)g
mXj=1
nXi=1
1)′(cid:13)(cid:13)(cid:13)
rijS(hβj)γi(cid:12)(cid:12)(cid:12) dν
mXj=1
nXi=1
≤ kgk(Y2πY ′
rijS(hβj)γi(cid:13)(cid:13)(cid:13)Y2πY ′
1
.
(c) ⇒ (a) Let us show that the condition (b) of Theorem 4.1 holds. Let x1, ..., xn ∈
X1(µ) which can be assumed to be non-null. Fix m ∈ N large enough such that (xi)m =
Pm
j=1hβ∗
it follows
j , xii βj 6= 0 and denote α = max i=1,..,n
j , xii. By taking rij =
j=1,...,m hβ∗
hβ∗
j ,xii
α
10
that
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
nXi=1Z T(cid:0)(xi)m(cid:1)γi dν =
= α
j , xiiZ T (βj)γi dν
mXj=1
nXi=1
hβ∗
rijZ T (βj)γi dν
nXi=1
mXj=1
rijS(hβj )γi(cid:13)(cid:13)(cid:13)Y2πY ′
≤ α C(cid:13)(cid:13)(cid:13)
nXi=1
mXj=1
j , xiiS(hβj )γi(cid:13)(cid:13)(cid:13)Y2πY ′
= C(cid:13)(cid:13)(cid:13)
nXi=1
mXj=1
hβ∗
S(cid:0)h(xi)m(cid:1)γi(cid:13)(cid:13)(cid:13)Y2πY ′
= C(cid:13)(cid:13)(cid:13)
nXi=1
.
1
1
1
(4.2)
Denoting by kTk the operator norm of T , since (xi)m → xi in X1(µ) as m → ∞ and
(cid:12)(cid:12)(cid:12)Z T (xi)z dν −Z T(cid:0)(xi)m(cid:1)z dν(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)Z T(cid:0)xi − (xi)m(cid:1)z dν(cid:12)(cid:12)(cid:12)
1 kT(cid:0)xi − (xi)m(cid:1)kY1
1 kTkkxi − (xi)mkX1
≤ kzkY ′
≤ kzkY ′
On other hand, denoting by kSk the operator norm of S, since
for every z ∈ Y1(ν)′, we have that Pn
kS(hxi)z − S(cid:0)h(xi)m(cid:1)zkY2πY ′
1
i=1R T(cid:0)(xi)m(cid:1)γi dν → Pn
i=1R T (xi)γi dν as m → ∞.
1
= kS(cid:0)h(xi − (xi)m)(cid:1)zkY2πY ′
1 kS(cid:0)h(xi − (xi)m)(cid:1)kY2
1 kSkkh(xi − (xi)m)kX2
1 kSkkhkX
≤ kzkY ′
≤ kzkY ′
≤ kzkY ′
X2
1 kxi − (xi)mkX1
i=1 S(cid:0)hxi(cid:1)γi in Y2πY ′
1 as m → ∞.
for every z ∈ Y1(ν)′, we have thatPn
Then, Taking limit as m → ∞ in (4.2), we obtain
i=1 S(cid:0)h(xi)m(cid:1)γi →Pn
1
.
nXi=1
We have already noted that in this case Y2πY ′
S(hxi)γi(cid:13)(cid:13)(cid:13)Y2πY ′
(d) implies (c). For every n, m ∈ N and (rij) ⊂ Bℓ∞ we have that
Assume that Y2(ν) ⊂ Y1(ν)′′ and that the functions (γn) have pairwise disjoint support.
1 ⊂ L1(ν) (denote by K its continuity constant)
and Pn
nXi=1Z T (xi)γi dν ≤ C(cid:13)(cid:13)(cid:13)
i=1 fiγi =(cid:12)(cid:12)Pn
i=1 fiγi(cid:12)(cid:12) pointwise for every n ∈ N and (fi) ⊂ L0(ν). Let us see that
nXi=1Z (cid:12)(cid:12)(cid:12)
rijZ T (βj)γi dν ≤ C
rijS(hβj )γi(cid:12)(cid:12)(cid:12) dν
mXj=1
mXj=1
nXi=1
= CZ (cid:12)(cid:12)(cid:12)
rijS(hβj )γi(cid:12)(cid:12)(cid:12) dν
mXj=1
nXi=1
rijS(hβj )γi(cid:13)(cid:13)(cid:13)Y2πY ′
≤ CK(cid:13)(cid:13)(cid:13)
mXj=1
nXi=1
.
1
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
11
If moreover L∞(ν) = Y
all x ∈ X1(µ), it follows that
Y ′′
1
2
rjZ T (βj)γn dν =
mXj=1
then (a) implies (d), as if g ∈ Y
Y ′′
1
2
is such that T (x) = gS(hx) for
mXj=1
≤ Z (cid:12)(cid:12)(cid:12)g
rjZ gS(hβj )γn dν =Z g
mXj=1
rjS(hβj )γn(cid:12)(cid:12)(cid:12) dν ≤ kgk∞Z (cid:12)(cid:12)(cid:12)
mXj=1
rjS(hβj )γn dν
mXj=1
rjS(hβj)γn(cid:12)(cid:12)(cid:12) dν.
(cid:3)
5. Examples: the Fourier and Ces`aro operators
In this section we show how the results obtained in the previous one can be applied in
concrete contexts. In particular, we will deal with the Fourier operator acting in different
weighted Lp-spaces, we will show factorization through infinite matrices and, as a special
case, we will analyze the case provided by the Ces`aro operator.
5.1. Strong factorization through the Fourier operator. Consider the measure space
given by the interval T = [−π, π], its Borel σ-algebra and the Lebesgue measure m and denote
by (φn) the real trigonometric system on T, that is,
φn(x) =
1
√2π
cos(kx)
√π
sin(kx)
√π
if n = 1
if n = 2k
.
if n = 2k + 1
−π φi(x)φj(x) dx = 0 if i 6= j andR π
Note thatR π
−π φi(x)φi(x) dx = 1. Each function f ∈ L1(m)
is associated to its Fourier series S(f ) =Pn≥1 anφn where an =RT f φn dm. If f ∈ Lr(m)
for 1 < r < ∞ then S(f ) converges to f in Lr(m) and so (φn) is a Schauder basis on Lr(m).
Let F be the Fourier operator defined by
F(f ) =(cid:16)ZT
f φn dm(cid:17),
f ∈ L1(m).
The Hausdorff-Young inequality (see for instance [7, (8.5.7)]) guarantees that
F : Lr(m) → ℓr′
is a well defined continuous operator for every 1 < r ≤ 2.
Fix 1 < r ≤ 2, r ≤ p < ∞, 1 < q ≤ ∞ and let T : Lp(m) → ℓq be a non-trivial continuous
linear operator. We have that (φn) is a Schauder basis for Lp(m) (as 1 < p < ∞) and (en) is
a Schauder basis for (ℓq)′ (as q > 1). Also, Lp(m) ⊂ Lr(m) (as r ≤ p) and so χT ∈ (Lp)Lr
.
Proposition 5.1. The following statements are equivalent:
12
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
(a) T factors strongly through F, that is, there exists g ∈ ℓsr′q such that
Lp(m)
T
/ ℓq
i
Mg
Lr(m)
F
/ ℓr′
(see (2.1) in the preliminaries for the definition of sr′q).
(c) There exists a constant C > 0 such that the inequality
(b) T (φn)i = 0 for all i 6= n and (cid:0)T (φi)i(cid:1) ∈ ℓsr′q .
rijT (φj)i ≤ C(cid:16) min{n,m}Xi=1
mXj=1
nXi=1
riis′
s′
r′q ,
r′q(cid:17) 1
n, m ∈ N
holds for every (rij) ⊂ Bℓ∞.
Moreover, in the case when r′ ≤ q, the conditions (a)-(c) are equivalent to
(d) There exists a constant C > 0 such that the inequality
rjT (φj)n ≤ C( rn
0
if n ≤ m
if n > m
mXj=1
holds for each n, m ∈ N and all (rj) ⊂ Bℓ∞.
)ℓq
and (ℓq)′ are σ-order continuous (as r, q > 1) and that (ℓr′
Proof. Note that both ℓr′
=
(ℓr′
= ℓsr′q where sr′q is defined as in (2.1). For the equivalence among (a), (b) and (c), let
us see that conditions (b) and (c) are just respectively conditions (b) and (c) of Theorem 4.2
rewritten for X1(µ) = Lp(m), X2(µ) = Lr(m), Y1(ν) = ℓq (ν being the counting measure λ
on N), Y2(ν) = ℓr′
(b) ⇒ Theorem 4.2.(b). Take g =(cid:0)T (φi)i(cid:1) ∈ ℓsr′q . Then, for every n, i ∈ N we have that
Theorem 4.2.(b) ⇒ (b). Let g ∈ ℓsr′q be such that T (φn) = gF(φn) for all n ∈ N. Then
T (φn)i = T (φi)iF(φn)i = giF(φn)i and so T (φn) = gF(φn).
, S = F, h = χT, (βn) = (φn) and (γn) = (en).
)(ℓq)′′
T (φn)i = giF(φn)i =( gi
0
if i = n
if i 6= n
and so(cid:0)T (φi)i(cid:1) = g ∈ ℓsr′q .
(c) ⇔ Theorem 4.2.(c). From Remark 3.4 and noting that (cid:0)ℓr′
r′q with equals norms and s′
(ℓsr′q )′ = ℓs′
=
r′q < ∞ (as sr′q > 1), for each n, m ∈ N and all
π(ℓq)′(cid:1)′′
= (cid:0)(ℓr′
)(ℓq)′′(cid:1)′
/
/
O
O
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
13
)(ℓq)′′
= ℓ∞. Then (d) is equivalent
(rij) ⊂ Bℓ∞ it follows that
π(ℓq)′
π(ℓq)′)′′
(cid:13)(cid:13)(cid:13)
s′
r′q
s′
r′q
mXj=1
nXi=1
rijF(φj)ei(cid:13)(cid:13)(cid:13)ℓr′
rijF(φj)ei(cid:13)(cid:13)(cid:13)(ℓr′
= (cid:13)(cid:13)(cid:13)
mXj=1
nXi=1
rijF(φj)ei(cid:13)(cid:13)(cid:13)ℓ
= (cid:13)(cid:13)(cid:13)
mXj=1
nXi=1
rijF(φj)iei(cid:13)(cid:13)(cid:13)ℓ
= (cid:13)(cid:13)(cid:13)
mXj=1
nXi=1
rijF(φj)i(cid:12)(cid:12)(cid:12)
= (cid:16) nXi=1(cid:12)(cid:12)(cid:12)
r′q(cid:17) 1
mXj=1
= (cid:16) min{n,m}Xi=1
r′q(cid:17) 1
nXi=1
mXj=1
In the case when r′ ≤ q we have that sr′q = ∞ and so (ℓr′
mXj=1
rijZ T (φj)ei dλ =
rjZ T (φj)en dλ =
nXi=1
mXj=1
mXj=1
riis′
rijT (φj)i.
rjT (φj)n
to (a)-(c) as (d) is to rewrite condition (d) of Theorem 4.2. Indeed,
s′
s′
r′q
s′
r′q
and
and
Z (cid:12)(cid:12)(cid:12)
mXj=1
rjF(φj )en(cid:12)(cid:12)(cid:12) dλ =(cid:12)(cid:12)(cid:12)
mXj=1
rjF(φj)n(cid:12)(cid:12)(cid:12) =( rn
0
if n ≤ m
if n > m
.
(cid:3)
5.2. Strong factorization for infinite matrices and the Ces`aro operator. Consider
the measure space (N,P(N), λ) with λ being the counting measure on N. Let X1(λ), X2(λ),
Y1(λ), Y2(λ) be saturated Banach function spaces in which (en) is a Schauder basis and
T : X1(λ) → Y1(λ), S : X2(λ) → Y2(λ) be nontrivial continuous linear operators. Then, the
operators T and S can be described by infinite matrices (aij) and (bij) respectively, namely
aij = T (ej)i and bij = S(ej )i. We also require that (en) is a Schauder basis for Y1(λ)′.
is saturated and that any of Y2(λ) or Y1(λ)′ is σ-order
1 , the following statements are equivalent:
Y ′′
1
2
Proposition 5.2. Assume that Y
continuous. Given h ∈ X X2
(a) T factors strongly through S and Mh.
such that aij
(b) There exists g ∈ Y
bij
Y ′′
1
2
bij = 0.
= gihj whenever bij 6= 0 and aij = 0 whenever
(c) There exists a constant C > 0 such that the inequality
nXi=1
mXj=1
rijaij ≤ C(cid:13)(cid:13)(cid:13)
nXi=1(cid:16) mXj=1
hjrijbij(cid:17)ei(cid:13)(cid:13)(cid:13)Y2πY ′
1
,
n, m ∈ N
holds for every (rij) ⊂ Bℓ∞.
14
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
Moreover, if Y2(λ) ⊂ Y1(λ)′′, then the condition
(d) There exists a constant C > 0 such that the inequality
mXj=1
rjanj ≤ C(cid:12)(cid:12)(cid:12)
holds for every (rj) ⊂ Bℓ∞,
implies (a)-(c). In the case when ℓ∞ = Y
mXj=1
hjrjbnj(cid:12)(cid:12)(cid:12),
n, m ∈ N
Y ′′
1
2
, we have that (d) is equivalent to (a)-(c).
Proof. We only have to see that conditions (b), (c) and (d) are just respectively conditions
(b), (c) and (d) of Theorem 4.2 rewritten for µ = ν being the counting measure λ and (βn) =
(γn) = (en). Note that for every i, j ∈ N we have that aij = T (ej)i and gihjbij = gihjS(ej)i =
giS(hjej)i = giS(hej)i. So (b) ⇔ Theorem 4.2.(b). Since R T (ej)ei dλ = T (ej)i = aij and
S(hej )ei = S(hjej)ei = hjS(ej)ei = hjS(ej)iei = hjbijei we have that (c) ⇔ Theorem 4.2.(c).
Moreover as
mXj=1
Z (cid:12)(cid:12)(cid:12)
mXj=1
mXj=1
rjhjbnjen(cid:12)(cid:12)(cid:12) dλ =(cid:12)(cid:12)(cid:12)
rjS(hej )en(cid:12)(cid:12)(cid:12) dλ =Z (cid:12)(cid:12)(cid:12)
i=1 xi(cid:1). It is well known that C : ℓr → ℓr continuously for every
nPn
rjhjbnj(cid:12)(cid:12)(cid:12),
(cid:3)
Let C be the Ces`aro operator which maps a real sequence x = (xn) into the sequence of its
1 < r < ∞ (see [7, Theorem 326]) and it can be described by the infinite matrix (bij) where
bij = 1
it follows that (d) ⇔ Theorem 4.2.(d).
Ces`aro means C(x) =(cid:0) 1
i if j ≤ i and bij = 0 if j > i, that is,
0
1
2
1
3
1
4
(bij) =
1
1
2
1
3
1
4
...
...
0
0
1
3
1
4
...
0 0 ···
0 0 ···
0 0 ···
0 ···
...
...
. . .
1
4
.
Fix 1 ≤ p < ∞, 1 < q, r < ∞ and let T : ℓp → ℓq be a nontrivial continuous operator
described by the infinite matrix (aij) with aij = T (ej)i. Note that (en) is a Schauder Basis
on ℓp, ℓq, ℓr and (ℓq)′.
Proposition 5.3. Let h ∈ ℓspr (see (2.1) for the definition of spr). The following statements
are equivalent:
(a) T factors strongly through C and Mh, that is, there exists g ∈ ℓsrq such that
ℓp
Mh
ℓr
T
C
ℓq
Mg
/ ℓr
(b) There exists g ∈ ℓsrq such that
aij =( gihj
0
i
if j ≤ i
if j > i
/
/
/
O
O
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
15
(c) There exists a constant C > 0 such that the inequality
nXi=1
mXj=1
rijaij ≤ C(cid:16) nXi=1
min{i,m}Xj=1
1
is′
rq(cid:12)(cid:12)(cid:12)
hjrij(cid:12)(cid:12)(cid:12)
s′
rq ,
s′
rq(cid:17) 1
n, m ∈ N
holds for every (rij) ∈ Bℓ∞.
Moreover, in the case when r ≤ q, the conditions (a)-(c) are equivalent to
(d) There exists a constant C > 0 such that the inequality
mXj=1
rjanj ≤ C
holds for every (rj) ⊂ Bℓ∞.
min{n,m}Xj=1
1
n(cid:12)(cid:12)(cid:12)
hjrj(cid:12)(cid:12)(cid:12),
n, m ∈ N
and s′
= (ℓsrq )′ = ℓs′
rq with equals norms
= ℓsrq and (ℓp)ℓr
nXi=1(cid:16) mXj=1
so (b) ⇔ Proposition 5.2.(b).
i if j ≤ i and bij = 0 if j > i,
Proof. Note that both ℓr and (ℓq)′ are σ-order continuous (as r < ∞ and q > 1), (ℓr)(ℓq)′′
=
= ℓspr . Also note that if r ≤ q then srq = ∞ and so (ℓr)(ℓq)′′
(ℓr)ℓq
= ℓ∞.
Then, we only have to see that (b), (c), (d) is just to rewrite respectively conditions (b), (c),
(d) of Proposition 5.2 for X1(λ) = ℓp, X2(λ) = ℓr, Y1(λ) = ℓq, Y2(λ) = ℓr and S = C.
As noted above, the elements of the matrix of C are bij = 1
By Remark 3.4 and noting that(cid:0)ℓrπ(ℓq)′(cid:1)′′
=(cid:0)(ℓr)(ℓq)′′(cid:1)′
rq < ∞ (as srq > 1), for every n, m ∈ N and (rj) ⊂ Bℓ∞ it follows that
= (cid:13)(cid:13)(cid:13)
hjrijbij(cid:17)ei(cid:13)(cid:13)(cid:13)ℓrπ(ℓq)′
nXi=1(cid:16) mXj=1
= (cid:13)(cid:13)(cid:13)
nXi=1(cid:16) mXj=1
= (cid:16) nXi=1(cid:12)(cid:12)(cid:12)
mXj=1
rq(cid:12)(cid:12)(cid:12)
= (cid:16) nXi=1
min{n,m}Xj=1
hjrijbij(cid:17)ei(cid:13)(cid:13)(cid:13)(ℓrπ(ℓq)′)′′
hjrijbjj(cid:17)ei(cid:13)(cid:13)(cid:13)ℓs′
hjrijbij(cid:12)(cid:12)(cid:12)
rq(cid:17) 1
min{i,m}Xj=1
hjrij(cid:12)(cid:12)(cid:12)
Hence, (c) ⇔ Proposition 5.2.(c).
(d) ⇔ Proposition 5.2.(d) holds as
s′
rq .
s′
rq(cid:17) 1
hjrjbnj =
hjrj.
rq
s′
s′
rq
mXj=1
1
n
(cid:13)(cid:13)(cid:13)
1
is′
Finally we show how the matrix of T must looks for T can be strongly factored through
the Ces`aro operator.
Proposition 5.4. Let h ∈ ℓspr and suppose that h1 6= 0. The following statements are
equivalent:
(a) T factors strongly through C and Mh.
(cid:3)
16
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
(b) aij = 0 for j > i, aij = hjai1
h1
(c) The matrix of T looks as
for j ≤ i and (iai1) ∈ ℓsrq .
0
0
0
0
0
0
h1α1
h1α2 h2α2
h1α3 h2α3 h3α3
h1α4 h2α4 h3α4 h4α4
h1α5 h2α5 h3α5 h4α5 h5α5
...
0
0
0
0
...
...
...
...
···
···
···
···
···
. . .
for some (αn) ∈ RN such that (nαn) ∈ ℓsrq (note that the element hj0α1 is positioned at the
j0-th row and the j0-th column of the matrix).
6. Domination by basis operators and representing operators
As a result of the active research in several branches of the Harmonic Analysis, a lot of
information is known about weighted norm inequalities for classical operators on weighted
Banach function spaces, mainly regarding weighted Lp and Lorentz spaces. The bibliography
on the subject is extremely broad; we refer the reader to [7] for the classical inequalities, and
Proof. (a) ⇒ (b) From Proposition 5.3 there exists g ∈ ℓsrq such that
i
.
where (αn) ∈ RN is such that (nαn) ∈ ℓsrq .
aij =( gihj
(b) ⇒ (c) Taking (αn) = (cid:0) an1
h1(cid:1) we have that hjαi = hjai1
(c) ⇒ (a) Taking g = (iαi) ∈ ℓsrq it follows that
aij =( hjαi = gihj
for all i and so aij = hjai1
h1
(nan1) ∈ ℓsrq.
if j ≤ i
if j > i
if j ≤ i
if j > i
0
0
.
i
Then ai1 = gih1
i
(nαn) = 1
h1
for every j ≤ i. Also note that (iai1) = h1g ∈ ℓsrq.
= aij for every j ≤ i and
h1
Then, from Proposition 5.3, (a) holds.
(cid:3)
If T factors strongly through C and Mh then there exists hj 6= 0 as T is non trivial. So,
given 0 6= h ∈ ℓspr and denoting j0 = min(cid:8)j ∈ N : hj 6= 0(cid:9), similarly to Proposition 5.4, we
have that T factors strongly through C and Mh if and only if its matrix looks as
0 ···
...
0 ···
0 ···
0 ···
0 ···
0 ···
...
0
...
0
0
...
0
0
0
...
0
0
0
0
...
0
0 hj0α1
0 hj0α2 hj0+1α2
0 hj0α3 hj0+1α3 hj0+2α3
0 hj0α4 hj0+1α4 hj0+2α4 hj0+3α4
...
0
...
0
0
0
0
...
...
...
...
···
···
···
···
···
···
. . .
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
17
to [2, 4] and the references therein for an updated review of the state of the art. We will
use also some concrete results and ideas concerning weighted norm inequalities that can be
found in the papers [1, 8, 12, 13].
We will show in what follows the characterization in terms of vector norm inequalities of
what we call a representing operator for a Banach function space X(µ). This is essentially
a modification of a basis operator F : L(µ) → Λ that allows to identify each function in
X(µ) with some easy transformation of the basic coefficients of certain univocally associated
function. Our motivation is given by the fact that, although the coefficients are not associ-
ated to a basis of the space X(µ), such kind of operator -- that we will call a representing
operator -- allows to find an easy representation of the functions of the space by means of
some basic-type coefficients. If F is a basis operator, we write (αi(f ))∞
i=1 ∈ Λ for the basic
coefficients of a function f , that is, F(f ) = (αi(f ))∞
i=1.
Definition 6.1. Let X(µ) be a Banach function space over µ and ℓ a sequence space over
the counting measure c on N. Let B be a Schauder basis of a Banach function space L(µ) and
suppose that the basic coefficients of the functions of L(µ) are in a sequence space Λ defined
as the Banach lattice given generated by an unconditional basis of a Banach space. Consider
an operator T : X(µ) → ℓ. We will say that T is a representing operator for X(µ) (with
respect to F) if it is an injective two-sides-diagonal transformation of the basis operator F.
Thus, technically a representing operator is an injective map such that there are a sequence
g = (gj) ∈ (Λ)ℓ with gj 6= 0 for all j ∈ N and a function h ∈ X L, h 6= 0 µ-a.e., such that for
every x ∈ X1(µ1), the sequence T (x) = (β(x)j) ∈ ℓ can be written as
β(x)j = Pj ◦ T (x) = gjF(hx) = gjαj(hx).
That is, for the elements y ∈ h · X(µ) ⊆ L(µ) we have that
j β(h−1y)j.
αj(y) = F(y) = g−1
Equivalently, for each x ∈ X(µ), there is a sequence (βj ) ∈ ℓ such that
x = T −1((βj)) = h−1F −1((g−1
j βj)).
Example 6.2.
(i) An easy example of the above introduced notion is the so called generalized Fourier
series. Consider p = 2, an interval I of the real line, the space L2(I) endowed
with Lebesgue measure dx and a weight function w : I → R+, w > 0. Note that
the multiplication operator Mw1/2 : L2(wdx) → L2(I) defines an isometry. Take
a sequence of functions (φn)n belonging to L2(wdx) and such that the associated
sequence (bn)n, where bn = w1/2φn for all n defines an orthonormal basis B in L2(I),
that is, it is orthogonal, norm one and complete. Note that this is equivalent to
say that it defines an orthonormal basis in the weighted space L2(wdx). Consider
the Fourier operator FB associated to the basis B of L2(I). Then the operator T :
L2(wdx) → ℓ2 given by T := idℓ2 ◦FB ◦ Mw1/2 is a representing operator for L2(wdx).
Concrete examples of this situation are given by classical orthogonal basis of poly-
nomials in weighted L2-spaces. For example, for the trivial case of the weight equal to
18
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
1 and the space L2[−1, 1], we can define the functions φn to be the Legendre polyno-
mials, that are solutions to the Sturm-Liouville problem and define the corresponding
Fourier-Legendre series. Other non trivial cases also for I = (−1, 1) are given by the
weight functions w(x) = (1 − x2)−1/2(x) and w(x) = (1 − x2)1/2(x) and the Cheby-
shev polynomial of the first and second kinds, respectively. Laguerre polynomials give
other example for I = (0,∞) and weight function w(x) = e−x.
(ii) Take a function 0 < h ∈ X(µ)Lp(µ) and consider a sequence 0 < λ = (λi) ∈ ℓ2. Let
us write c for the counting measure in N. Consider the space ℓ1(λc) = {(τi) : (λiτi) ∈
ℓ1} with the corresponding norm k(τi)kℓ1(λc) = Pλiτi. Then we have that ℓ2 ֒→
ℓ1(λc), and so the space of multiplication operators (ℓ2)ℓ1(λc) is not trivial. A direct
computation shows also that (ℓ2)ℓ1(λc) = ℓ2(λ2c). Then, for every τ = (τi) ∈ ℓ2(λ2c)
with τi 6= 0 for all i ∈ N we have that the operator T : X(µ) → ℓ1(λc) given by
T (·) = τF(h·) is a representing operator for the space X(µ).
Let J be a finite subset of N, and write PJ : ℓ → ℓ for the standard projection on the
subspace generated by the elements of B with subindexes in J.
If T : X(µ) → ℓ is an
operator, consider the net {PJ ◦ T =: TJ : N ⊃ J finite}, where the order is given by the
inclusion of the set of subindexes, that is PJ ◦ T ≤ PJ ′ ◦ T if and only if J ⊆ J ′. By definition,
T = lim
B
PB ◦ T
as a pointwise limit. In what follow we will characterize representing operators in terms of
inequalities using this approximation procedure and a compactness argument. Thus, consid-
ering the basic (biorthognal) functionals b′
i ∈ L(µ)′, i ∈ J, associated to the basis of L(µ)
that defines the Fourier operator that we are considering, we have
PJ (x) :=Xj∈J
hf, b′
jiej,
f ∈ Lp(µ).
Fix a function h ∈ (X(µ))L(µ) and suppose that Λℓ is non-trivial. Assume that the condi-
= (Λ′πℓ)∗. The domination inequality
tions are given in order to obtain that (Λ)ℓ = (ℓ′)Λ′
that must be considered in this case is given by the following expression.
nXi=1Z PJ ◦ T (xi)y′
i dc ≤(cid:13)(cid:13)(cid:13)(cid:16) nXi=1Xj∈J
g∈B(Λ)ℓ(cid:16) nXi=1Xj∈J
that is, we are considering the sequence(cid:0)Pn
hhxi, b′
the dual of (Λ)ℓ given by
iihej,· y′
= sup
(cid:16) nXi=1Xj∈J
ii(cid:17) : Λℓ → R.
hhxi, b′
iihej ,· y′
ii(cid:17)(cid:13)(cid:13)(cid:13)Λ′πℓ
hhxi, b′
i=1hhxi, b′
ii(cid:17),
iihej, gy′
i)j(cid:1)j∈J ∈ Λ′πℓ as the functional of
ii(y′
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
19
After taking into account the particular descriptions of the elements of the spaces involved,
we get the equivalent expression for the inequality
nXi=1Xj∈J
jihej, (gj (y′
hhxi, b′
i)j)i
hT (xi), ejihej , ((y′
i)j)i(cid:17) = sup
g∈B(Λ)ℓ(cid:16) nXi=1Xj∈J
T (xi)j (y′
i)j ≤ sup
g∈BΛ)ℓ(cid:16) nXi=1Xj∈J
hhxi, b′
i)j(cid:17)
jigj(y′
αj(hxi) gj (y′
i)j(cid:17),
jdµ, j ∈ J, are the j-th Fourier coefficients of the function hxi
≤ sup
g∈BΛℓ(cid:16) nXi=1Xj∈J
nXi=1Xj∈J
where αj(hxi) = R hxib′
associated to the basis B.
and so the initial inequality is equivalent to the following one,
Thus, the assumptions on the properties of (X(µ))L(µ) and Λℓ provides the following
Theorem 6.3. Suppose that Λ′ and ℓ satisfies that Λ′πℓ′ is saturated and (Λ′πℓ′)∗ = (Λ′)ℓ,
and let h be a measurable function such that 0 < h ∈ (X(µ))L(µ). The following statements
are equivalent for an operator T : X(µ) → ℓ.
(i) For every finite set J ⊆ N the inequality
i)j ≤ C sup
holds for every x1, ..., xn ∈ X1 and y′
nXi=1Xj∈J
T (xi)j (y′
g∈BΛℓ(cid:16) nXi=1Xj∈J
αj(hxi) gj (y′
i)j(cid:17),
1, ..., y′
n ∈ ℓ′.
(ii) T is a representing operator with respect to F, that is, there is a sequence g ∈ Λℓ such
that (T (x))j = gj · αj(hx) for all x ∈ X(µ) and j ∈ N. In other words, T factors
through F as
T
X(µ)
Mh
L(µ)
F
/ Λ.
ℓ
Mg
(6.1)
Proof. Let us see that (i) implies (ii). We can assume without loss of generality that C = 1.
Note that as a consequence of Remark 3.3 the requirements on Λ and ℓ provides the conditions
on these spaces for applying Corollary 3.2. By the computations above, we obtain that for
each finite set J we have a norm one sequence gJ ∈ Λℓ satisfying that
PJ ◦ T (x) = gJ · PJ ◦ F(hx).
Consider the net N := {gJ : J ⊆ N finite}, where the order is given by the inclusion of the
finite sets used for the subindexes. We can assume without loss of generality that the support
of each function gJ is in J, that is, the coefficients (gJ )k of the sequence gJ are 0 for k /∈ J.
Since all the functions of the net are in the unit ball and due to the product compatibility of
the pair defined by Λ′ and ℓ, we have that the net is included in the weak* compact set BΛℓ.
Therefore, it has a convergent subnet N0, that is, there is a sequence g0 ∈ BΛℓ such that
lim
η∈N0
gη = g0
/
/
/
O
O
nXi=1Xj∈J
T (xi)j (y′
g∈Bℓ2(λ2c)(cid:16) nXi=1Xj∈J
i)j ≤ C sup
1, ..., y′
αj(hxi) gj (y′
i)j(cid:17),
20
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
in the weak* topology given by the dual pair(cid:10)Λ′πℓ, Λℓ(cid:11).
Note now that for a fixed x ∈ X(µ), due to the fact that we are assuming that ℓ has a
unconditional basis with associated projections PJ , we have
PJ ◦ T (x) = T (x).
J∈N finite
lim
Then,
T (x) = lim
J∈N finite
PJ ◦ T (x) = lim
η∈N0
Pη ◦ T (x) = lim
η∈N0
gη · F(hx) = g0 · F(hx).
This gives (ii) and finishes the proof, since the converse holds by a direct computation.
(cid:3)
Let us provide an example. Consider again Example 6.2(ii), and recall that (ℓ2)ℓ1(λc) =
ℓ2(λ2c). Theorem 6.3 gives that an injective operator T : X(µ) → ℓ1(λc) is a representing
operator by means of the Fourier operator if and only if for every finite set J ⊆ N the
inequality
holds for every x1, ..., xn ∈ X(µ) and y′
Remark 6.4. Let us gives some sufficient conditions for the product sequence space appearing
in Theorem 6.3 to satisfy what is needed. The product compatibility of the pair ℓp′
and ℓ
means that
n ∈ ℓ1(λc)′.
)ℓ = (ℓ′)ℓp
For example, if ℓ′ is p-convex we have that ℓ′πℓp′
is saturated and so, a Banach function space
(see Proposition 2.2 in [15]). Moreover, the quoted result provides also the equality (under
the assumption of saturation of the product)
(ℓp′
.
=(cid:0)ℓ′πℓp′(cid:1)∗
= (ℓp′
)ℓ.
(cid:0)ℓ′πℓp′(cid:1)′
Consequently, if the product is order continuous, we get the desired result. Conditions under
which this space is order continuous are given in Proposition 5.3 in [3]: for example, if the
norm of the product is equivalent to
kλkπ ∼ inf(cid:8)kηkℓp′ · kγkℓ′ : λ = η · γ, η ∈ ℓp′
, γ ∈ ℓ′(cid:9),
the space is order continuous if ℓ′ is assumed to be order continuous (recall that p > 1 and
so ℓp′
is order continuous too). The formula above for the product space works for example
if ℓ is p′-concave, since this implies that ℓ′ is p-convex that together with the p′-convexity of
ℓp′
provides the result. Concrete examples for ℓp spaces has been given in Example 6.2.
7. Operators associated to trigonometric series
Relevant historical examples are the ones associated to the Fourier series and the corre-
sponding Fourier coefficients. We finish the paper by explicitly writing the results presented
previously in this setting. We will write x(·) for the i-th Fourier (real) coefficients of the
function x with indexes in the set Z, writing the coefficients an asociated to cos functions as
x(i) with positive i and the coefficients bn for the functions sin as x(i) with negative i.
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
21
• Due to the Hausdorff-Young inequality, we know that for 1 < p ≤ 2, the Fourier
transform Fp -- sending Lp[−π, π] → ℓp′
that assigns to each function the sequence
of its Fourier coefficients is well-defined and continuous. The Fourier transform is
defined as F2 : L2 → ℓ2. Suppose that we want to check if a particular operator
G2 : L2[−π, π] → ℓ2 can be extended to Lp[−π, π] through Fp. That is, is there a
factorization for G2 as
L2[−π, π]
i
Lp[−π, π]
G2
Fp
ℓ2
Mλ
/ ℓp′
for the operator G2 for some multiplication operator given by a sequence λ.
We have shown that this is equivalent to the following inequalities to hold for the
operator G2. For each x1, ..., xn ∈ L2[−π, π] and λi ∈ ℓ2,
∞Xk=1
nXi=1
(G2(xi))k(λi)k
≤ C(cid:13)(cid:13)(cid:0) nXi=1
xi(k)(λi)k(cid:1)(cid:13)(cid:13)ℓr = C(cid:16) ∞Xk=1(cid:12)(cid:12)
nXi=1
xi(k)(λi)k(cid:12)(cid:12)r(cid:17)1/r
.
• For 1 < p ≤ 2 again, Kellogg proved an improvement of the Hausdorff-Young inequal-
ity, that assures that the corresponding Fourier coefficients of the functions in Lp can
be found in the smaller mixed norm space Lp′,2 ⊆ ℓp′
. Fix 1 ≤ p, q ≤ ∞. The mixed
norm sequence space Lp,q was defined in [8] as the space of sequences λ = (λk)∞
−∞
such that
kλk =(cid:16) ∞Xm=∞(cid:16) Xk∈I(m)
λkp(cid:17)q/p(cid:17)1/q
< ∞,
where I(m) = {k ∈ Z : 2m−1 ≤ k ≤ 2m} if m > 0, I(0) = {0} and I(m) = {k ∈ Z :
−2−m ≤ k ≤ −2−m−1} if m < 0. It is easy to see that Lp′,2 ⊆ ℓp′
, and so we have a
factorization for the Fourier map as
Lp[−π, π]
i
Lp[−π, π]
Fp
Kp
ℓp′
i
/ Lp′,2.
In Theorem 1 of [8], it is proved that the space of multiplication operators (multi-
can in fact be identified with ℓ∞. Consequently, our results
pliers) from Lp′,2 to ℓp′
imply that for every finite set J ⊆ Z the inequality
nXi=1Xj∈J
(Fpxi)j (λ′
i)j ≤ C sup
g∈Bℓ∞(cid:16) nXi=1Xj∈J
xi(j) gj (λ′
i)j(cid:17),
holds for every x1, ..., xn ∈ X1 and λ′
n ∈ ℓp, what is obvious. However, note that
this is essentially a characterization, since any other operator Gp from Lp and having
1, ..., λ′
/
/
/
O
O
/
/
/
O
O
22
O. DELGADO, M. MASTY LO, AND E. A. S ´ANCHEZ P´EREZ
values in a sequence space ℓ such that (Lp′,2)ℓ = ℓ∞ satisfying these inequalities has
to be of the form g · Kp for a certain sequence g ∈ ℓ∞.
• The Hardy-Littlewood inequality, also for 1 < p ≤ 2, provides an example of an
operator Hp sending the Fourier coefficients of the functions in Lp to a weighted ℓp
space. For 1 < p < 2, consider the weighted sequence space ℓp(W ), where the weight
W is given by W = (Wn) = (1/(n + 1)2−p). The Hardy-Littlewood inequality can be
understood as the fact that the Fourier operator can be defined as Hp : Lp[−π, π] →
ℓp(W ) (see [1, S.2], in particular Theorem B). Note that the multiplication operator
Mγ : ℓp(W ) → ℓp given by the sequence γ = (cid:16)(1/(n + 1)
p (cid:17) defines an isometry.
Therefore, the factorization scheme
2−p
Lp[−π, π]
i
Lp[−π, π]
γ·Hp
Hp
ℓp
Mγ
/ ℓp(W )
provides other example of the situation we are describing. Indeed, for every multi-
plication operator τ for τ ∈ (ℓp(W ))ℓp
we can give an operator τ · Hp satisfying this
factorization. Our results implies the class of all these operators is characterized in
the following way: if T : Lp[−π, π] → ℓp satisfies the inequalities
i)j ≤ C
sup
xi(j) gj (λ′
(T (xi))j (λ′
nXi=1Xj∈J
g∈B(ℓp(W ))ℓp(cid:16) nXi=1Xj∈J
i)j(cid:17),
n ∈ ℓp′
1, ..., λ′
for each finite subset J ⊂ Z, for every x1, ..., xn ∈ Lp[−2π, 2π] and λ′
then it has a factorization as the one above for a certain τ ∈ (ℓp(W ))ℓp
.
• Let us recall Example 6.2(i). A representing operator T : L2(wdx) → ℓ2 associated
to a weight function w and an orthogonal basis B with respect to the corresponding
weight function was considered. It allowed a factorization as
,
L2(wdx)
M
w1/2
L2[I]
T
FB
ℓ2
id
/ ℓ2.
The corresponding vector norm inequality characterizing this factorization is
nXi=1Xj∈J
(T (xi))j (λ′
(T (xi))j (λ′
i)j ≤ C(cid:13)(cid:13)(cid:13)
nXi=1Xj∈J
nXi=1
(FB(xi))j (λ′
i)j(cid:12)(cid:12)(cid:12)
= CXj∈J(cid:12)(cid:12)(cid:12)
i)j(cid:13)(cid:13)(cid:13)ℓ1
for a given constant C > 0 and for each finite subset J ⊂ N, for every x1, ..., xn ∈
L2(wdx) and λ′
1, ..., λ′
n ∈ ℓ2.
/
/
/
O
O
/
/
/
O
O
STRONG FACTORIZATIONS OF FOURIER TYPE REPRESENTING OPERATORS
23
References
[1] J. M. Ash, S. Tikhonov and J. Tung, Wiener's positive Fourier coefficients theorem in variants of Lp
spaces, Michigan Math. J. 59 (2010), no. 1, 143 -- 151.
[2] J. J. Benedetto and H. P. Heinig, Weighted Fourier inequalities: new proofs and generalizations, J. Fourier
Anal. Appl. 9 (2003), no. 1, 1 -- 37.
[3] J. M. Calabuig, O. Delgado and E. A. S´anchez P´erez Generalized perfect spaces, Indag. Math. 19 (2008),
359 -- 378.
[4] D. Cruz-Uribe, J. M. Martell and C. P´erez, Sharp weighted estimates for classical operators, Adv. Math.
229 (2012), no. 1, 408-441.
[5] O. Delgado and E. A. S´anchez P´erez, Summability properties for multiplication operators on Banach
function spaces, Integr. Equ. Oper. Theory 66 (2010), 197 -- 214.
[6] O. Delgado and E. A. S´anchez P´erez, Strong factorizations between couples of operators on Banach func-
tion spaces, J. Convex Anal. 20 (2013), 599 -- 616.
[7] G. H. Hardy, J. E. Littlewood and G. P´olya, Inequalities, Cambridge University Press, 1934.
[8] C. N. Kellogg, An extension of the Hausdorff-Young theorem, Michigan Math. J. 18 (1971), 121 -- 127.
[9] P. Kolwicz, K. Le´snik and L. Maligranda, Pointwise products of some Banach function spaces and fac-
torization, J. Funct. Anal. 266 (2014), no. 2, 616 -- 659.
[10] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces, vol. II, Springer-Verlag, Berlin, 1979.
[11] L. Maligranda and L. E. Persson, Generalized duality of some Banach function spaces, Indag. Math.
51 (1989), 323 -- 338.
[12] H. Mhaskar and S. Tikhonov, Wiener type theorems for Jacobi series with nonnegative coefficients, Proc.
Amer. Math. Soc. 140 (2012), no. 3, 977 -- 986.
[13] C. P´erez, Sharp Lp-weighted Sobolev inequalities, Ann. Inst. Fourier 45 (1995), no. 3, 809-824.
[14] A. Pietsch, Operator Ideals, North Holland, Amsterdam, 1980.
[15] E. A. S´anchez P´erez, Factorization theorems for multiplication operators on Banach function spaces,
Integr. Equ. Oper. Theory 80,1 117-135 (2014).
[16] A. R. Schep, Products and factors of Banach function spaces, Positivity 14 (2010), no. 2, 301-319.
[17] A. C. Zaanen, Integration, 2nd rev. ed. North Holland, Amsterdam, 1967.
Departamento de Matem´atica Aplicada I, E. T. S. de Ingenier´ıa de Edificaci´on, Universidad
de Sevilla, Avenida de Reina Mercedes, 4 A, Sevilla 41012, Spain.
E-mail address: [email protected]
Faculty of Mathematics and Computer Science, Adam Mickiewicz University in Pozna´n, Umul-
towska 87, 61-614 Pozna´n, Poland.
E-mail address: [email protected]
Instituto Universitario de Matem´atica Pura y Aplicada, Universitat Polit`ecnica de Val`encia,
Camino de Vera s/n, 46022 Valencia, Spain.
E-mail address: [email protected]
|
1102.0336 | 4 | 1102 | 2011-06-22T07:23:01 | The Bargmann transform on a broad family of Banach spaces, with applications to Toeplitz and pseudo-differential operators | [
"math.FA",
"math.AP",
"math.CV"
] | We investigate mapping properties for the Bargmann transform on modulation spaces whose weights and their reciprocals are allowed to grow faster than exponentials. We prove that this transform is isometric and bijective from modulation spaces to convenient Lebesgue spaces of analytic functions. We use this to prove that such modulation spaces fulfill most of the continuity properties which are well-known when the weights are moderated. Finally we use the results to establish continuity properties of Toeplitz and pseudo-differential operators in the context of these modulation spaces. | math.FA | math | THE BARGMANN TRANSFORM ON A BROAD FAMILY OF
BANACH SPACES, WITH APPLICATIONS TO TOEPLITZ AND
PSEUDO-DIFFERENTIAL OPERATORS
JOACHIM TOFT
Abstract. We investigate mapping properties for the Bargmann transform on
modulation spaces whose weights and their reciprocals are allowed to grow faster
than exponentials. We prove that this transform is isometric and bijective from
modulation spaces to convenient Lebesgue spaces of analytic functions. We use
this to prove that such modulation spaces fulfill most of the continuity properties
which are well-known when the weights are moderated. Finally we use the results
to establish continuity properties of Toeplitz and pseudo-differential operators in
the context of these modulation spaces.
0. Introduction
In this paper we introduce and establish basic continuity properties for a broad
family of (quasi-)Banach spaces of functions and distributions of Gelfand-Shilov
types, in the framework of harmonic analysis. We establish close links between
these spaces and (weighted) Lebesgue spaces Ap,q
(ω) of analytic functions related to
Bargmann-Fock spaces. The family of spaces consists of modulation spaces, where
each modulation space is obtained by imposing a weighted mixed norm estimate on
the short-time Fourier transforms of the involved distributions. Important cases
of such spaces are given by M p,q
(ω), where the weighted mixed norm estimate is
constituted by the Lp,q
(ω) norm.
Among the involved parameters p, q and the weight ω, it follows that ω is
most important concerning imposing regularity, or relaxing growth, oscillations
and singularity conditions on the involved distributions. In comparison to already
established theories of such spaces (cf. [22, 27, 47] and the references therein) the
conditions for the involved weight functions are significantly relaxed in the present
paper. This leads to that our family of modulation spaces are significantly larger
compared to the "usual" families of such spaces. For example, for each fixed
s > 1/2, the modulation space M p,q
(ω) can be made "arbitrary close" to the Gelfand-
Shilov space Ss or to S′s, by choosing the weight ω in appropriate ways.
An essential part of our investigations concerns the establishment of the links
between the modulation spaces and the Ap,q
(ω) spaces, by proving that the Bargmann
transform is isometric and bijective between these spaces. One of the benefits is
that any property valid for the Ap,q
(ω) spaces, carry over to the modulation spaces, and
vise versa. For example, we prove that any modulation space is a Banach or quasi-
Banach space, and that convenient density, duality and interpolation properties
1
1
1
0
2
n
u
J
2
2
]
.
A
F
h
t
a
m
[
4
v
6
3
3
0
.
2
0
1
1
:
v
i
X
r
a
hold for such spaces, because similar properties are valid for corresponding spaces
of analytic functions.
Finally we use our results to extend the theory of pseudo-differential operators
to involve more extreme symbols and target distributions comparing to earlier
investigations.
We recall that the (classical) modulation space M p,q
(ω), p, q ∈ [1,∞], as introduced
and carefully investigated by Feichtinger and Grochenig in [18 -- 21, 27], consists of
all tempered distributions whose short-time Fourier transforms (STFT) have finite
mixed Lp,q
(ω) norm. Here the weight ω quantifies the degree of asymptotic decay and
singularity of the distribution in M p,q
(ω). In general it is assumed that ω should be
moderate, which imposes several properties on ω and thereby on the modulation
space M p,q
(ω). (See Sections 1 and 2 for strict definitions.) For example, the moderate
property implies that ω is not allowed to grow or decay faster than exponentials,
that M p,q
(ω) are invariant (but not norm invariant) under pullbacks of translations,
and that several properties valid for weighted Lebesgue spaces (e. g. density, duality
and interpolation properties) carry over to classical modulation spaces.
A major idea behind the design of these spaces was to find useful Banach spaces,
which are defined in a way similar to Besov spaces, in the sense of replacing the
dyadic decomposition on the Fourier transform side, characteristic to Besov spaces,
with a uniform decomposition. From the construction of these spaces, it turns out
that modulation spaces and Besov spaces in some sense are rather similar, and sharp
embeddings between these spaces can be found in [54,55], which are improvements
of certain embeddings in [26]. (See also [48, 60] for verification of the sharpness.)
During the last 15 years, several results have been proved which confirm the
usefulness of the modulation spaces in time-frequency analysis, where they occur
naturally. For example, in [19, 28, 30], it is shown that all modulation spaces admit
reconstructible sequence space representations using Gabor frames.
Parallel to this development, modulation spaces have been incorporated into the
calculus of pseudo-differential operators, which also involve Toeplitz operators. (See
e. g. [28,31,32,35,36,48,54 -- 58] and the references therein concerning symbol classes
embedded in S ′, and [15, 28, 42 -- 44, 49, 50] for results involving ultra-distributions.
Here and in what follows we use the usual notations for the usual function and
distribution spaces, see e. g. [37].)
The Bargmann transform can easily be reformulated in terms of the short-time
Fourier transform, with a particular Gauss function as window function. By refor-
mulating the Bargmann transform in such way, and using the fundamental role of
the short-time Fourier transform in the definition of modulation spaces, it easily
follows that the Bargmann transform is continuous and injective from M p,q
(ω) to Ap,q
(ω).
Furthermore, by choosing the window function as a particular Gaussian function
in the M p,q
(ω) norm, it follows that V : M p,q
These facts and several other mapping properties for the Bargmann transform
on (classical) modulation spaces were established in [20, 22, 27, 33, 47]. In fact, here
(ω) → Ap,q
(ω) is isometric.
2
it is proved that the Bargmann transform from M p,q
but in fact bijective.
(ω) to Ap,q
(ω) is not only injective,
For the modulation space M p,q
(ω), the weight function ω is important for imposing
or relaxing conditions on the distributions f in M p,q
(ω). More precisely, the weight ω =
ω(x, ξ) depends on both the space (or time) variables x as well as the momentum
(or frequency) variables ξ. Roughly speaking, the weight function posses (cf. [16,
18, 28, 31, 32]):
• ω tending rapidly to infinity as x tends to infinity, imposes that f tends
rapidly to zero at infinity;
• ω tending rapidly to zero as x tends to infinity, relaxes the growth conditions
on f at infinity;
• ω tending rapidly to infinity as ξ tends to infinity, imposes high regularity
for f ;
• ω tending rapidly to zero as ξ tends to infinity, relaxes the conditions on
singularities of f ;
• ω tending rapidly to infinity as both x and ξ tends to infinity, imposes
stronger restrictions on oscillations for f at infinity;
• ω tending rapidly to zero as both x and ξ tends to infinity, relax the re-
strictions on oscillations for f at infinity.
The condition that ω should be moderate implies that
ω + 1/ω ≤ v
(0.1)
for some v = Cec · , where c, C > 0 are constants.
In this case, ω is called a
weight of exponential type. We remark that corresponding modulation spaces M p,q
(ω)
are subsets of appropriate spaces of Gelfand-Shilov distributions, and for certain
choices of ω we may have that M p,q
(ω) is contained in S , or that S ′ is contained in
M p,q
(ω). A more restrictive case appears when (0.1) is true for some v = Cec · s, with
0 ≤ s < 1. In this case, ω is called a weight of subexponential type. If instead v in
(0.1) can be chosen as polynomial, then ω is said to be of polynomial type. In this
case, M p,q
(ω) contains S , and is contained in S ′.
Several properties for the modulation spaces might be violated when passing
from the subexponential type weights into exponential type weights. For example,
if ω is of exponential type, then M p,q
(ω) might be contained in the set of real analytic
functions, which in particular implies that there are no non-trivial compactly sup-
ported elements in M p,q
(ω). Consequently, there are no compactly supported Gabor
atoms, implying the time-frequency machinery breaks in those parts were com-
pactly supported Gabor atoms are needed.
In the present paper we go beyond these situations and relax the assumptions
on v even more. For example, we permit v in (0.1) to be superexponential, i. e.
v = Cec · γ , where 1 < γ < 2. In this situation, almost no arguments in classical
modulation space theory can be used, because the main results in that theory are
3
based on the fact that ω should be moderate. This condition is violated when v in
(0.1) has to be superexponential.
In Sections 1 and 2 we give the explicit conditions on the weight functions, and
in Sections 3 and 4 we prove:
(1) any extended weight class contains all weights in classical modulation space
theory, including weights which are moderated by exponential type weights.
Furthermore, any superexponential weight with γ above less than 2 are
included, as well as weights of the form ω = h · ih· i and ω = Γ(h · i + 1).
Here hxi = (1 + x2)1/2 and Γ is the Gamma function.
(ω) and M p,q
(ω) are Banach spaces and fulfill convenient density, duality and
(2) Ap,q
interpolation properties.
(3) the Bargmann transform is isometric and bijective from M p,q
(ω) to Ap,q
(ω).
In the last section we establish new forms of pseudo-differential calculi in the
framework of these modulation spaces. This means that the spans of the spaces
for operator symbols, target functions and image functions, are significantly larger
comparing to earlier theories. Therefore, these spaces may be smaller as well as
larger comparing the usual situations. The approach here is similar to [56, 57, 59],
where similar results were obtained in background of classical modulation space
theory. The results here are, to some extent, also related to the results in [11, 42 --
44, 46, 49, 50], when v in (0.1) is bounded by a subexponential function.
We remark that in contrast to classical theory of pseudo-differential operators,
(cf. e. g. [37]), there are no explicit regularity assumptions on the symbols. On the
other hand, if 1 < γ1 < γ < γ2 < 2 with c > 0, and the weight ω is given by
ω(x, ξ) = ec(xγ+ξγ),
(0.2)
then the corresponding modulation spaces are contained in the Gelfand-Shilov
space S1/γ1, and contain S1/γ2. In particular, this means that the involved func-
tions and their derivatives are extendable to entire analytic functions and fulfill
estimates of the form
f (x) ≤ Ce−cxγ1 ,
and bf (ξ) ≤ Ce−cξγ1 ,
for some positive constants c and C. It is therefore obvious that in this situation,
the elements in these modulation spaces posses strong regularity properties.
contain the dual S′1/γ1
the space of tempered distributions.
On the other hand, if c < 0 in (0.2), then the corresponding modulation spaces
of S1/γ1, which in turn is significantly larger than e. g. S ′,
Finally, in Section 5 we apply the continuity results for modulation spaces to
establish continuity properties for Toeplitz operators with symbols in weighted
mixed norm space of Lebesgue types. (Cf. [5, 7, 10, 13, 14, 58] and the references
therein for similar and related investigations.)
4
Acknowledgement
First I would like to express my gratitude to M. Signahl for interesting and
valuable discussions on questions related to certain parts of the present paper
during 2009 and 2010. I also thank K. Grochenig for valuable discussions. In fact,
important ideas to the paper appeared when I communicated with him during the
spring 2010. I am also grateful to A. Galbis, C. Fernandez, P. Wahlberg and S.
Pilipovi´c for supports and careful reading of the original paper, leading to several
improvements of the style and content.
1. Preliminaries
In this section we give some definitions and recall some basic facts. The proofs
are in general omitted. In the first part we consider appropriate cnditions on the in-
volvd weight functions. Thereafter we review some facts for Gelfand-Shilov spaces.
Then we discuss basic properties of the short-time Fourier transform, which is there-
after used in the definition of modulation spaces, and obtaining basic properties
for such spaces. The last part of the section is devoted to the Bargmann transform
and appropriate Banach spaces of entire functions, which are appropriate in the
background of the Bargmann transform.
1.1. Weight functions. We start by discussing general properties on the involved
weight functions. A weight on Rd is a positive function ω on Rd such that ω ∈
L∞loc(Rd), and for each compact set K ⊆ Rd, there is a constant c > 0 such that
ω(x) ≥ c
when
x ∈ K.
A usual condition on ω is that it should be v-moderate for some positive function
v ∈ L∞loc(Rd). This means that
(1.1)
for some constant C which is independent of x, y ∈ Rd. We note that (1.1) implies
that ω fulfills the estimates
ω(x + y) ≤ Cω(x)v(y),
x, y ∈ Rd,
C−1v(−x)−1 ≤ ω(x) ≤ Cv(x).
We say that v is submultiplicative when (1.1) holds with ω = v. In the sequel,
v and vj for j ≥ 0, always stand for submultiplicative weights if nothing else is
stated.
The weight ω is called a weight of exponential type, if v in (1.1) can be chosen
as v(x) = Cecx for some c, C > 0.
If, more restrictive, v can be chosen as a
polynomial, then ω is called a weight of polynomial type. We let P(Rd) and
PE(Rd) be the sets of all weights on Rd of polynomial type and exponential type,
respectively. Obviously, P(Rd) ⊆ PE(Rd).
A broader class of moderate weights comparing to P(Rd) is obtained by re-
placing the polynomial assumption on v by the so called GRS condition (Gelfand-
Raikov-Shilov condition). That is, v ∈ L∞loc(Rd) is positive and satisfies
log v(nx)
= 0.
lim
n→∞
n
5
An important class of submultiplicative weights which fulfills the GRS condition is
the so called subexponential weights, i. e. weights of the form
ω(x) = Cecxs
,
(1.2)
when ω = v, and c, C and s are positive constants such that s < 1. On the other
hand, if v is a weight of exponential type, then the GRS condition is violated.
Furthermore, if ω is v-moderate for some v, then it is moderated by an exponential
type weight. Consequently, the PE(Rd) contains all weights on Rd which are
moderated by some functions, including those weights moderated by v which fulfills
the GRS-conditions. We refer to [29] and the references therein for these facts.
In this paper we permit weights where the moderate condition (1.1) on ω has
been relaxed by appropriate local and global conditions. In most of the situations,
the local condition is
C−1ω(x) ≤ ω(x + y) ≤ Cω(x) when Rc ≤ x ≤ c/y, R ≥ 2,
(1.3)
for some positive constants c and C. However,
condition (1.3) is relaxed into
C−1ω(x)2 ≤ ω(x+y)ω(x−y) ≤ Cω(x)2 when Rc ≤ x ≤ c/y, R ≥ 2, (1.3)′
for some positive constants c and C.
in most of the situations, the
Important examples of weights satisfying (1.3) are those which satisfy (1.2), when
C and s being positive such that s ≤ 2, and c ∈ R. Especially we note that if ω
is given by (1.2) with 1 < s ≤ 2, then ω is not moderated by any weight v, but
satisfies (1.3) for some choices of c > 0 and C > 0. On the other hand, if ω > 0
and satisfies (1.3), then Proposition 2.6 in Section 2 shows that
holds for some positive constants c and C.
C−1e−cx2
≤ ω(x) ≤ Cecx2
,
(1.4)
Definition 1.1. Let ω be a weight on Rd.
(1) ω is called a weight of Gaussian type (weakly Gaussian type) on Rd, if (1.3)
holds (if (1.3)′ holds) for some positive c and C, and (1.4) holds for some
positive c and C. The set of Gaussian type and weakly Gaussian type
weights on Rd are denoted by PG(Rd) and PQ(Rd), respectively;
(2) ω is called a weight of subgaussian type (weakly subgaussian type) on Rd, if
(1.3) holds (if (1.3)′ holds) for some positive c and C, and for every c > 0,
there is a constant C > 0 such that (1.4) holds. The set of subgaussian
type and weakly subgaussian type weights on Rd are denoted by P 0
G(Rd)
and P 0
Q(Rd), respectively.
We note that each one of the families of weight functions in Definition 1.1 are
groups under the ordinary multiplications.
Remark 1.2. The family P 0
G, but its definition is somewhat
more complicated. An important reason for introducing this family is that we may
prove that the general modulation spaces, introduced later on, can be made close
Q is larger than P 0
6
to Gelfand-Shilov spaces in the sense of Proposition 3.9 in Section 3. So far we are
unable to prove any similar result when the family P 0
Q is replaced by P 0
G.
On the other hand, for any weight in P 0
G one may find an equivalent smooth
weight (cf. Proposition 2.6 in Section 2). So far we are unable to extend this
property to all weights in P 0
Q.
We note that if ω ∈ P 0
(1) there are invertible d× d-matrices T1, . . . , TN whose norms are at most one,
Q(Rd), then ω satisfies the following conditions:
i. e. kTjk ≤ 1, j = 1, . . . , N, and such that
C−1ω(x)N ≤
NYj=1
ω(x + Tjy) ≤ Cω(x)N
for some positive constants c and C;
when Rc ≤ x ≤ c/y, R ≥ 2,
(1.3)′′
(2) for every c > 0, there is a constant C > 0 such that (1.4) holds.
Hence if we modify the definition of P 0
Q(Rd) in such way that it should contain
all weights ω satisfying (1) and (2), then we obtain a larger family of weights,
comparing to Definition 1.1. By straight-forward computations it follows that all
results in the paper are true after the definition of P 0
Q in Definition 1.1 has been
modified in this way.
A special situation appears for Proposition 3.9 in Section 3, where the symmetry
condition in y in (1.3)′ is essential for its proof. However, it follows that Proposition
3.9 is true, after (1.3)′ in the definition of P 0
Q has been replaced by
C−1ω(x)2N ≤
ω(x + Tjy)ω(x − Tjy) ≤ Cω(x)2N ,
NYj=1
when Rc ≤ x ≤ c/y and R ≥ 2, where Tj are invertible matrices with norm at
most one.
In Section 2 we introduce other convenient subfamilies of PQ(Rd).
Example 1.3. Let c, s ∈ R, C > 0 and t > 1/2. Then
σs(x) ≡ hxis = (1 + x2)s/2, ω1(x) = ecx1/t
and ω2(x) = ecx2
,
(1.5)
are weights of polynomial type, subgaussian type and Gaussian type, respectively.
Definition 1.4. Let Ω ⊆ PQ(Rd). Then Ω is called an admissible family of
weights, if there is a rotation invariant function 0 < ω0(x) ∈ L∞loc(Rd) ∩ L1(Rd)
which decreases with x and such that
ω · ω0 ∈ Ω and ω/ω0 ∈ Ω when ω ∈ Ω.
We note that for some choice of ω0 in Definition 1.4 we have
for some constant C > 0.
ω0(x) ≤ Chxi−d
7
(1.6)
Example 1.5. Every family in Definition 1.1 are admissible. Moreover, if ω0 ∈
PQ(Rd) and Ω is a family of admissible weights, then
(1) { σN ; N ∈ Z} is admissible;
(2) ω0 · Ω ≡ { ω0ω ; ω ∈ Ω} is admissible.
1.2. Gelfand-Shilov spaces. Next we recall the definition of Gelfand-Shilov spaces.
Let 0 < h, s ∈ R be fixed. Then we let Ss,h(Rd) be the set of all f ∈ C∞(Rd)
such that
kfkSs,h ≡ sup xβ∂αf (x)
hα+β(α! β!)s
is finite. Here the supremum should be taken over all α, β ∈ Nd and x ∈ Rd.
Obviously Ss,h ⊆ S is a Banach space which increases with h and s. Furthermore,
if s > 1/2 or s = 1/2 and h ≥ 1, then Ss,h contains all finite linear combinations of
Hermite functions. Since such linear combinations are dense in S , it follows that
the dual S′s,h(Rd) of Ss,h(Rd) is a Banach space which contains S ′(Rd).
limit respectively of Ss,h(Rd). This implies that
The Gelfand-Shilov spaces Ss(Rd) and Σs(Rd) are the inductive and projective
Ss(Rd) = [h>0
Ss,h(Rd)
and Σs(Rd) = \h>0
Ss,h(Rd),
(1.7)
and that the topology for Ss(Rd) is the strongest possible one such that each
inclusion map from Ss,h(Rd) to Ss(Rd) is continuous. The space Σs(Rd) is a Fr´echet
space with semi norms k · kSs,h, h > 0.
We remark that the space Ss(Rd) is zero when s < 1/2, and that Σs(Rd) is zero
when s ≤ 1/2. Furthermore, for each ε > 0 and s ≥ 1/2 we have
(F f )(ξ) = bf (ξ) ≡ (2π)−d/2ZRd
8
f (x)e−ihx,ξi dx
Σs(Rd) ⊆ Ss(Rd) ⊆ Σs+ε(Rd).
On the other hand, in [41] there is an alternative elegant definition of Σs1(Rd) and
Ss2(Rd) such that these spaces agrees with the definitions above when s1 > 1/2
and s2 ≥ 1/2, but Σ1/2(Rd) is non-trivial and contained in S1/2(Rd).
From now on we assume that s > 1/2 when considering Σs(Rd).
The Gelfand-Shilov distribution spaces S′s(Rd) and Σ′s(Rd) are the projective and
inductive limit respectively of S′s,h(Rd). This means that
S′s(Rd) = \h>0
S′s,h(Rd)
and Σ′s(Rd) = [h>0
S′s,h(Rd).
(1.7)′
We remark that already in [24] it is proved that S′s(Rd) is the dual of Ss(Rd), and
if s > 1/2, then Σ′s(Rd) is the dual of Σs(Rd) (also in topological sense).
The Gelfand-Shilov spaces are invariant under several basic transformations. For
example they are invariant under translations, dilations, tensor products and under
any Fourier transformation.
From now on we let F be the Fourier transform which takes the form
when f ∈ L1(Rd). Here h · , · i denotes the usual scalar product on Rd. The map F
extends uniquely to homeomorphisms on S ′(Rd), S′s(Rd) and Σ′s(Rd), and restricts
to homeomorphisms on S (Rd), Ss(Rd) and Σs(Rd), and to a unitary operator on
L2(Rd).
The following lemma shows that elements in Gelfand-Shilov spaces can be char-
acterized by estimates of the form
f (x) ≤ Ce−εx1/s
The proof is omitted, since the result can be found in e. g. [12, 24].
Lemma 1.6. Let f ∈ S′1/2(Rd). Then the following is true:
and bf (ξ) ≤ Ce−εξ1/s
.
(1.8)
(1) if s ≥ 1/2, then f ∈ Ss(Rd), if and only if there are constants ε > 0 and
C > 0 such that (1.8) holds;
(2) if s > 1/2, then f ∈ Σs(Rd), if and only if for each ε > 0, there is a
constant C such that (1.8) holds.
Gelfand-Shilov spaces posses several other convenient properties. For example,
they can easily be characterized by Hermite functions. We recall that the Hermite
function hα with respect to the multi-index α ∈ Nd is defined by
hα(x) = π−d/4(−1)α(2αα!)−1/2ex2/2(∂αe−x2
).
The set (hα)α∈Nd is an orthonormal basis for L2(Rd). In particular,
cαhα,
cα = (f, hα)L2(Rd),
(1.9)
f =Xα
and
kfkL2 = k{cα}αkl2 < ∞,
when f ∈ L2(Rd). Here and in what follows, ( · , · )L2(Rd) denotes any continuous
extension of the L2 form on S1/2(Rd).
It is well-known that f here belongs to S (Rd), if and only if
k{cαhαit}αkl2 < ∞
(1.10)
for every t ≥ 0. Furthermore, for every f ∈ S ′(Rd), the expansion (1.9) still
holds, where the sum converges in S ′, and (1.10) holds for some choice of t ∈ R,
depending on f . The same conclusion holds after the l2 norm has been replaced
by any lp norm with 1 ≤ p ≤ ∞.
The following proposition, which can be found in e. g. [25], shows that similar
conclusion for Gelfand-Shilov spaces hold, after the estimate (1.10) is replaced by
k{cαetα1/2s
}αklp < ∞.
(1.11)
(Cf. formula (2.12) in [25].)
Proposition 1.7. Let p ∈ [1,∞], f ∈ S′1/2Rd), s ≥ 1/2 and let cα be as in (1.9).
Then the following is true:
(1) f ∈ S′s(Rd), if and only if (1.11) holds for every t < 0. Furthermore, (1.9)
holds where the sum converges in S′s;
9
(2) f ∈ Σ′s(Rd), if and only if (1.11) holds for some t < 0. Furthermore, (1.9)
holds where the sum converges in Σ′s;
(3) f ∈ Ss(Rd), if and only if (1.11) holds for some t > 0. Furthermore, (1.9)
holds where the sum converges in Ss;
(4) f ∈ Σs(Rd), if and only if (1.11) holds for every t > 0. Furthermore, (1.9)
holds where the sum converges in Σs.
1.3. The short-time Fourier transform. Let φ ∈ S (Rd) \ 0 be fixed. For
every f ∈ S ′(Rd), the short-time Fourier transform Vφf is the distribution on
R2d defined by the formula
(1.12)
We note that the right-hand side defines an element in S ′(R2d) ∩ C∞(R2d), and
that if f ∈ Lq
(ω) for some ω ∈ P(Rd), then Vφf takes the form
(Vφf )(x, ξ) = F (f φ( · − x))(ξ).
Vφf (x, ξ) = (2π)−d/2ZRd
f (y)φ(y − x)e−ihy,ξi dy.
(1.12)′
In order to extend the definition of the short-time Fourier transform we re-
formulate (1.12) in terms of partial Fourier transforms and tensor products. More
presisely, we let F2F be the partial Fourier transform of F (x, y) ∈ S ′(R2d) with re-
spect to the y-variable, and we let U be the map which takes F (x, y) into F (y, y−x).
Then it follows that
(1.13)
Vφf = (F2 ◦ U)(f ⊗ φ)
when f ∈ S ′(Rd) and φ ∈ S (Rd).
We remark that tensor products of elements in Gelfand-Shilov spaces are defined
in similar ways as for tensor products for distributions (cf. Chapter V in [37]). Let
f, g ∈ S′1/2(Rd) and let s ≥ 1/2. Then it follows that f ⊗ g ∈ S′s(R2d), if and only
if f, g ∈ S′s(Rd). Similar fact holds for any other choice of Gelfand-Shilov spaces
of functions or distributions.
The following result is essentially a restatement of ?? in [15] and concerns the
map
(f, φ) 7→ Vφf,
(1.14)
and follows immediately from (1.13), and the facts that tensor products, F2 and
U are continuous on Gelfand-Shilov spaces. (See also [15, 34] for general properties
of the short-time Fourier transform in background of Gelfand-Shilov spaces.)
Proposition 1.8. Let s ≥ 1/2 and let f, φ ∈ S′1/2(Rd) \ 0. Then the map (1.14)
from S (Rd)× S (Rd) to S ′(R2d) is uniquely extendable to a continuous map from
S′1/2(Rd) × S′1/2(Rd) to S′1/2(R2d). Furthermore, the following is true:
(1) the map (1.14) restricts to a continuous map from Ss(Rd) × Ss(Rd) to
Ss(R2d). Moreover, Vφf ∈ Ss(R2d), if and only if f, φ ∈ Ss(Rd);
(2) the map (1.14) restricts to a continuous map from S′s(Rd) × S′s(Rd) to
S′s(R2d). Moreover, Vφf ∈ S′s(R2d), if and only if f, φ ∈ S′s(Rd).
10
Similar facts hold after the spaces Ss and S′s have been replaced by Σs and Σ′s
respectively.
We also have the following proposition.
Proposition 1.9. Let s ≥ 1/2, φ ∈ Ss(Rd) \ 0 be even, and let f ∈ S′1/2(Rd).
Then the following is true:
(1) f ∈ Ss(Rd), if and only if for some ε > 0 and some constant Cε it holds
Vφf (x, ξ) ≤ Cεe−ε(x1/s+ξ1/s);
(1.15)
(1.16)
(2) if f ∈ S′s(Rd), then there are constants ε > 0 and Cε > 0 such that
Vφf (x, ξ) ≤ Cεeε(x1/s+ξ1/s);
(3) if for every ε > 0, there is a constant Cε such that (1.16) holds, then
f ∈ S′s(Rd).
Proposition 1.9 can be found in [15] and to some extend also in [34]. Since the
arguments in the proof are important later on, we present here an explicit proof,
based on reformulation of the statements in terms of Wigner distributions.
First let f, g ∈ L2(Rd). Then the Wigner distribution of f and g is defined by
the formula
Wf,g(x, ξ) = (2π)−d/2Z f (x − y/2)g(x + y/2)eihy,ξi dy.
We note that the Wigner distribution is closely connected to the short-time Fourier
transform, since
Vφf (x, ξ) = 2−deihx,ξi/2Wf, φ(−x/2, ξ/2),
which follows by straight-forward computations. Here f (x) = f (−x). From this
relation it follows that most of the properties which involve short-time Fourier
transform also hold after replacing the short-time Fourier transforms by Wigner
distributions. For example, Propositions 1.8 and 1.9 remain the same after such
replacements.
Proof. (1) If f ∈ Ss(Rd), then it follows from Lemma 1.6 and Proposition 1.8 that
(1.15) holds for some constants ε > 0 and Cε > 0.
Now assume instead that (1.15) holds for some constants ε > 0 and Cε > 0.
Then (1.15) still holds after Vφf has been replaced by Wf, φ = Wf,φ, provided the
constants ε and Cε have been replaced by other suitable ones, if necessary. Since
F (Wf,φ)(ξ, x) = Vφf (−x, ξ),
by Parseval's formula, it follows that (1.15) holds for both Wf,φ and F (Wf,φ).
Hence, f ∈ Ss(Rd) by Lemma 1.6. This proves (1).
The assertion (2) follows by straight-forward computations, using the fact that
in combination with Lemma 1.6.
Vφf (x, ξ) = hf, φ( · − x)e−ih · ,ξii
11
On the other hand, if for every ε > 0 there is a constant Cε > 0 such that (1.16)
holds and ϕ is a finite sum of Hermite functions, then Vφϕ ∈ Ss(R2d), and
(f, ϕ)L2(Rd) = c(Vφf, Vφϕ)L2(R2d)
(1.17)
is well-defined. Here c = kφk−2
L2 > 0. Now, by (1.16), (1) and the fact that finite
sums of Hermite functions are dense in Ss(Rd), it follows that the right-hand side
of (1.17) defines a continuous linear form on Ss(Rd) with respect to ϕ. Hence,
f ∈ S′s(Rd), which gives (3), and the proof is complete.
Remark 1.10. There is obviously a gap between the necessary and sufficiency con-
ditions in (2) and (3) of Proposition 1.9. In general it seems to be difficult to find
convenient equivalent conditions for the short-time Fourier transform of f ∈ S′1/2
in order for f should belong to S′s, for some s > 1/2.
On the other hand, for each s ≥ 1/2 and φ ∈ Ss(Rd) \ 0, let Υs,φ(Rd) be the
space which consists of all f ∈ S′s(Rd) such that for every ε > 0 there is a constant
Cε > 0 such that (1.16) holds. Then Υs,φ(Rd) is still a "large space" in the sense
that it contains every S′t(Rd) for t > s.
For future references we set Υ = Υs,φ when s = 1/2 and φ(x) = π−d/4e−x2/2.
(cid:3)
1.4. Modulation spaces. We shall now discuss modulation spaces and recall
some basic properties. In what follows we let B be a mixed norm space on Rd.
This means that for some p1, . . . , pn ∈ [1,∞] and vector spaces
V1, . . . , Vn ⊆ Rd
such that V1 ⊕ · · · ⊕ Vn = Rd,
(1.18)
then B = Bn, where Bj, j = 1, . . . , n is inductively defined by
Bj =(Lp1(V1),
Lpj (Vj; Bj−1),
j = 1
j = 2, . . . , n.
(1.19)
The minimal and maximal exponents min(p1, . . . , pn) and max(p1, . . . , pn) are de-
noted by ν1(B) and ν2(B) respectively, and the norm of B is given by kfkB ≡
kFn−1kLpn (Vn), where F0 = f and
Fj(xj+1, . . . , xn) = kFj−1( · , xj+1, . . . , xn)kLpj (Vj ),
j = 1, . . . , n − 1.
In several situations the notation Lp(V ) is used instead of B, where
V = (V1, . . . , Vn)
and p = (p1, . . . , pn).
(1.20)
We set B′ = Lp′
(V ), where p′ = (p′1, . . . , p′2) and p′j ∈ [1,∞] is the conjugate
exponent of pj, j = 1, . . . , n. That is, pj and p′j should satisfy 1/pj + 1/p′j = 1. If
ν2(B) < ∞, then the dual of B with respect to ( · , · )L2 is given by B′.
should belong to (0,∞] instead of [1,∞]. Still we set
In some situations we relax the conditions on p1, . . . , pn in such way that they
kfkLpj (Vj ) ≡
(cid:16)RVj f (xj)pj dxj(cid:17)1/pj
xj∈Vj (cid:0)f (xj)(cid:1),
ess sup
12
, when 0 < pj < ∞
pj = ∞,
when
and
kFkLp,q ≡(cid:16)ZRd(cid:16)ZRd F (x, ξ)p dx(cid:17)q/p
tw ≡(cid:16)ZRd(cid:16)ZRd F (x, ξ)q dξ(cid:17)p/q
kFkLp,q
dξ(cid:17)1/q
dx(cid:17)1/p
< ∞ .
< ∞ ,
where f is measurable on Vj. (Cf. [8].) We note that k · kLpj (Vj ) is a quasi-norm,
but not a norm, when pj < 1. Furthermore, in this case, Lpj (Vj) is a quasi-Banach
space, with topology defined by this quasi-norm.
Now, if p1, . . . , pn ∈ (0,∞], and V = (V1, . . . Vn) is the same as above, then Lp(V )
is called mixed quasi-norm space on Rd, and is defined as Bn in (1.19).
Example 1.11. Let p, q ∈ [1,∞], and Lp,q(R2d) and its twisted space Lp,q
be the Banach spaces, which consist of all F ∈ L1
loc(R2d) such that
tw (R2d)
respectively (with obvious modifications when p = ∞ or q = ∞). Then it follows
that both Lp,q(R2d) and Lp,q
tw are defined in analogous ways, where
loc(R2d) with r = min(p, q).
If instead p, q ∈ (0,∞], then Lp,q and Lp,q
the condition F ∈ L1
In this case, one obtains mixed quasi-norm spaces.
loc(R2d) has to be replaced by F ∈ Lr
tw (R2d) are mixed norm spaces.
The definition of modulation spaces is given in the following.
Q(R2d),
Definition 1.12. Let B be a mixed quasi-norm space on R2d, ω ∈ P 0
and let φ = π−d/4e−x2/2. Then the modulation space M(ω, B) consists of all
f ∈ S′1/2(Rd) such that
kfkM (ω,B) ≡ kVφf ωkB < ∞.
(1.21)
If ω = 1, then the notation M(B) is used instead of M(ω, B).
s = M p
(ω) = M p,p
(ω) = W p,p
Since the cases B = Lp,q(R2d) and B = Lp,q
(ω)(Rd) = M(ω, Lp,q(R2d)) and W p,q
tw (R2d) are especially important to us
we set M p,q
(ω)(Rd) = M(ω, Lp,q
tw (R2d)). We recall
that if ω ∈ P(R2d), then the former space is a "classical modulation space", and
the latter space is related to certain types of classical Wiener amalgam spaces. For
convenience we set M p
(σs) and
M p
(σs), where σs is given by (1.5), and if ω = 1, then we use the notations
M(B), M p,q, W p,q and M p, instead of M(ω, B), M p,q
(ω), respectively.
Here we note that
and σs(x, ξ) = hx, ξis = (1 + x2 + ξ2)s/2.
σs(x) = hxis = (1 + x2 + ξ2)s/2
For exponential type weights we have the following proposition. We omit the
proof, since the result can be found in [18, 20, 21, 28, 56]. Here and in what follows
we write p1 ≤ p2, when
(ω). Furthermore, we set M p,q
(ω) and M p
s = M p,q
(ω), W p,q
p1 = (p1,1, . . . , p1,n) ∈ (0,∞]n
and p2 = (p2,1, . . . , p2,n) ∈ (0,∞]n
(1.22)
satisfy p1,j ≤ p2,j for every j = 1, . . . , n.
13
Proposition 1.13. Let p, q, pj, qj ∈ [1,∞], ω, ωj, v ∈ PE(R2d) for j = 1, 2 be such
that v is submultiplicative and even, ω is v-moderate, and let B be a mixed normed
space on R2d. Then the following is true:
(1) if φ ∈ M 1
(v)(Rd) \ 0, then f ∈ M(ω, B) if and only if (1.21) holds, i. e.
M(ω, B) is independent of the choice of φ. Moreover, M(ω, B) is a Banach
space under the norm in (1.21), and different choices of φ give rise to
equivalent norms;
(2) if (1.18), (1.20) and (1.22) hold with p1 ≤ p2, and ω2 ≤ Cω1 for some
constant C, then
Σ1(Rd) ⊆ M(ω1, Lp1(V )) ⊆ M(ω2, Lp2(V )) ⊆ Σ′1(Rd);
(3) the sesqui-linear form ( · , · )L2 on Σ1(Rd) extends to a continuous map from
(ω)(Rd)×M p′,q′
M p,q
(1/ω)(Rd) to C. This extension is unique, except when p = q′ ∈
{1,∞}. On the other hand, if kak = sup(a, b)L2, where the supremum is
taken over all b ∈ M p′,q′
(1/ω) ≤ 1, then k·k and k·kM p,q
(1/ω)(Rd) such that kbkM p′,q′
(ω)
are equivalent norms;
(4) if p, q < ∞, then Σ1(Rd) is dense in M p,q
(ω)(Rd) can be identified with M p′,q′
M p,q
over, Σ1(Rd) is weakly dense in M∞(ω)(Rd).
(ω)(Rd), and the dual space of
(1/ω)(Rd), through the form ( · , · )L2. More-
1.5. The Bargmann transform. We shall now consider the Bargmann transform
which is defined by the formula
(Vf )(z) = π−d/4ZRd
exp(cid:16) −
1
2
(hz, zi + y2) + 21/2hz, yi(cid:17)f (y) dy,
when f ∈ L2(Rd). We note that if f ∈ L2(Rd), then the Bargmann transform Vf
of f is the entire function on Cd, given by
(Vf )(z) =Z Ad(z, y)f (y) dy,
or
where the Bargmann kernel Ad is given by
(Vf )(z) = hf, Ad(z, · )i,
(1.23)
1
2
Ad(z, y) = π−d/4 exp(cid:16) −
zjwj, when z = (z1, . . . , zd) ∈ Cd
(hz, zi + y2) + 21/2hz, yi(cid:17).
and w = (w1, . . . , wd) ∈ Cd,
Here
hz, wi =
dXj=1
and otherwise h · , · i denotes the duality between test function spaces and their
corresponding duals. We note that the right-hand side in (1.23) makes sense when
14
f ∈ S′1/2(Rd) and defines an element in A(Cd), since y 7→ Ad(z, y) can be inter-
preted as an element in S1/2(Rd) with values in A(Cd). Here and in what follows,
A(Cd) denotes the set of all entire functions on Cd.
It was proved by Bargmann that f 7→ Vf is a bijective and isometric map from
L2(Rd) to the Hilbert space A2(Cd), the set of entire functions F on Cd which
fulfills
kFkA2 ≡(cid:16)ZCd F (z)2dµ(z)(cid:17)1/2
< ∞.
(1.24)
Here dµ(z) = π−de−z2 dλ(z), where dλ(z) is the Lebesgue measure on Cd, and the
scalar product on A2(Cd) is given by
F (z)G(z) dµ(z), F, G ∈ A2(Cd).
(1.25)
(F, G)A2 ≡ZCd
F (z) =ZCd
Furthermore, Bargmann proved that there is a convenient reproducing kernel on
A2(Cd), given by the formula
e(z,w)F (w) dµ(w), F ∈ A2(Cd),
(1.26)
where (z, w) is the scalar product of z ∈ Cd and w ∈ Cd (cf. [3, 4]). Note that this
reproducing kernel is unique in view of [38].
From now on we assume that φ in (1.12), (1.12)′ and (1.21) is given by
φ(x) = π−d/4e−x2/2,
(1.27)
if nothing else is stated. Then it follows that the Bargmann transform can be
expressed in terms of the short-time Fourier transform f 7→ Vφf . More precisely,
let S be the dilation operator given by
(SF )(x, ξ) = F (2−1/2x,−2−1/2ξ),
loc(R2d). Then it follows by straight-forward computations that
when F ∈ L1
(Vf )(z) = (Vf )(x + iξ) = (2π)d/2e(x2+ξ2)/2e−ihx,ξiVφf (21/2x,−21/2ξ)
(1.28)
= (2π)d/2e(x2+ξ2)/2e−ihx,ξi(S−1(Vφf ))(x, ξ),
(1.29)
or equivalently,
Vφf (x, ξ) = (2π)−d/2e−(x2+ξ2)/4e−ihx,ξi/2(Vf )(2−1/2x,−2−1/2ξ).
For future references we observe that (1.29) and (1.30) can be formulated into
= (2π)−d/2e−ihx,ξi/2S(e− · 2/2(Vf ))(x, ξ).
(1.30)
V = UV ◦ Vφ,
and U−1
V ◦ V = Vφ,
where UV is the linear, continuous and bijective operator on D′(R2d) ≃ D′(Cd),
given by
(1.31)
(UVF )(x, ξ) = (2π)d/2e(x2+ξ2)/2e−ihx,ξiF (21/2x,−21/2ξ).
15
Definition 1.14. Let ω1 ∈ P 0
space on R2d = Cd, and let r > 0 be such that r ≤ ν1(B).
Q(R2d), ω2 ∈ PQ(R2d), B be a mixed quasi-norm
(1) The space B(ω2, B) is the modified weighted B-space which consists of all
F ∈ Lr
loc(R2d) = Lr
loc(Cd) such that
kFkB(ω2,B) ≡ k(U−1
V F )ω2kB < ∞.
Here UV is given by (1.31);
(2) The space A(ω2, B) consists of all F ∈ A(Cd) ∩ B(ω2, B) with topology
inherited from B(ω2, B);
(3) The space A0(ω1, B) is given by
A0(ω1, B) ≡ { (Vf ) ; f ∈ M(ω1, B) },
and is equipped with the quasi-norm kFkA0(ω1,B) ≡ kfkM (ω1,B), when F =
Vf .
We note that the spaces in Definition 1.14 are normed spaces when ν1(B) ≥ 1.
For conveneincy we set kFkB(ω,B) = ∞, when F /∈ B(ω, B) is measurable, and
kFkA(ω,B) = ∞, when F ∈ A(Cd) \ B(ω, B). We also set
(ω)(Cd) = B(ω, B), Ap,q
(ω) = Ap,q
(ω)(Cd) = A(ω, B)
Bp,q
(ω) = Bp,q
when B = Lp,q(Cd), Ap
Ap,q, Bp and Ap instead of Bp,q
(ω) = Ap,p
(ω), and if ω = 1, then we use the notations Bp,q,
(ω), Ap,q
(ω), Bp
(ω) and Ap
(ω), respectively.
For future references we note that the Bp
(ω) quasi-norm is given by
kFkBp
(ω)
= 2d/p(2π)−d/2(cid:18)ZCd e−z2/2F (z)ω(21/2z)p dλ(z)(cid:19)1/p
= 2d/p(2π)−d/2(cid:18)ZZR2d e−(x2+ξ2)/2F (x + iξ)ω(21/2x,−21/2ξ)p dxdξ(cid:19)1/p
(1.32)
(with obvious modifications when p = ∞). Especially it follows that the norm and
scalar product in B2
(ω)(Cd) take the forms
kFkB2
(ω)
(F, G)B2
(ω)
=(cid:18)ZCd F (z)ω(21/2z)2 dµ(z)(cid:19)1/2
=ZCd
F (z)G(z)ω(21/2z)2 dµ(z),
(cf. (1.24) and (1.25)).
,
F ∈ B2
(ω)(Cd)
F, G ∈ B2
(ω)(Cd)
The following result shows that the norm in A0(ω,B) is well-defined.
Q(R2d), B be a mixed norm space on R2d and let
Proposition 1.15. Let ω ∈ P 0
φ be as in (1.27). Then A0(ω, B) ⊆ A(ω, B), and the map V is an isometric
injection from M(ω, B) to A(ω, B).
16
Proof. The result is an immediate consequence of (1.29), (1.30) and Definition
1.14.
(cid:3)
In the case ω = 1 and B = L2, it follows from [3] that Proposition 1.15 holds, and
the inclusion is replaced by equality. That is, we have A2
0 = A2 which is called the
Bargmann-Fock space, or just the Fock space. In Section 4 we improve the latter
property and show that for any choice of ω ∈ P 0
Q and every mixed quasi-norm
space B, we have A0(ω, B) = A(ω, B).
2. Weight functions
In this section we establish results on weight functions which are needed. In the
first part we investigate weights belonging to PE. Here we are especially focused
on finding properties which are needed to show that PE is contained in convenient
subfamilies of PQ, which are introduced in the second part of the section.
2.1. Moderate weights. For a moderate weight ω there are convenient ways to
find smooth weigths ω0 which are equivalent in the sense that for some constant
C > 0 we have
C−1ω0 ≤ ω ≤ Cω0.
(2.1)
In fact, we have the following result, which extends Lemma 1.2 (4) in [55]. Here
the weight ω ∈ PE(Rd) is called elliptic if ω ∈ C∞(Rd), and for each multi-index
α, we have
(2.2)
Lemma 2.1. Let ω ∈ PE(Rd). Then it exists an elliptic weight ω0 ∈ PE(Rd)
such that (2.1) holds.
(∂αω0)/ω0 ∈ L∞(Rd).
Lemma 2.1 follows by similar arguments as in the proof of Lemma 1.2 (4) in [55].
In order to be self-contained we here present a proof.
Proof. Let ψ ∈ C∞(Rd) be a positive function which is bounded by Gauss functions
together with all its derivatives, and let ω0 = ψ∗ ω. Also let v ∈ PE(Rd) be chosen
such that ω is v-moderate. Then ω0 is smooth, and
ω0(x) =Z ω(x − y)ψ(y) dy ≤ C1ω(x),
C1 =Z v(−y)ψ(y) dy < ∞.
C2 =Z v(y)−1ψ(y) dy < ∞,
with
Furthermore, if
then
C2ω(x) =Z ω(x − y + y)
ψ(y)
v(y)
dy ≤ CZ ω(x − y)ψ(y) dy = Cω0(x),
for some constant C. This proves that ω0 ∈ PE ∩ C∞, and that (2.1) is fulfilled.
17
By differentiating ω0, the first part of the proof gives that for some Gaussian ψ1,
and some constants C1 and C2 we have
∂αω0 = (∂αψ) ∗ ω ≤ ∂αψ ∗ ω ≤ ψ1 ∗ ω ≤ C1ω ≤ C2ω0.
Hence, (2.2) is fulfilled for ω = ω0. The proof is complete.
(cid:3)
We also need some properties concerning the minimal weight which moderates
a specific weight ω ∈ PE(Rd). For any such weight, the fact that ω is v-moderate
for some function v, implies that
is well-defined. Furthermore, by straight-forward computations it follows that
v0(x) ≡ sup
y0∈Rd
ω(x + y0)
ω(y0)
(2.3)
ω(x + y) ≤ ω(x)v(y) and v(x + y) ≤ v(x)v(y),
x, y ∈ Rd,
(2.4)
holds for v = v0. The following result shows that v0 is minimal among elements
which moderates ω, and satisfies (2.4). Furthermore, here we establish differential
properties of v0 in terms of the functionals
(J1ω)(x, y) ≡ inf
(J2ω)(x, y) ≡ sup
(Jkω)(x, y) ≤ sup
ω(y0)
y0∈Rd(cid:16)(∂yω)(x + y0)
y0∈Rd(cid:16)(∂yω)(x + y0)
y0∈Rd(cid:16)(∇ω)(x + y0)
(cid:17) = inf
y0∈Rd(cid:16)h(∇ω)(x + y0), yi
(cid:17) = sup
y0∈Rd(cid:16)h(∇ω)(x + y0), yi
(cid:17)y ≤ Cv0(x)y,
ω(y0)
ω(y0)
ω(y0)
ω(y0)
(cid:17),
(cid:17),
k = 1, 2,
when in additional ω is elliptic. We note that J1ω and J2ω satisfy
for such ω and some constant C, depending on ω only.
Lemma 2.2. Let ω ∈ PE(Rd), 0 < v ∈ L∞loc(Rd) be such that ω(x + y) ≤ ω(x)v(y)
when x, y ∈ Rd, and let v0 be as in (2.3). Then v0 ≤ v, and (2.4) holds.
Moreover, if in addition ω is elliptic, then the following is true:
(1) (J1ω)(x, y) and (J2ω)(x, y) are continuous functions which are positively
homogeneous in the y-variable of order one;
(2) if x, y ∈ Rd, then
inf
(J1ω)(x + ty, y) ≤ v0(x + y) − v0(x) ≤ sup
0≤t≤1
0≤t≤1
In particular, v0 is locally Lipschitz continuous;
(J2ω)(x + ty, y).
(3) if x, y ∈ Rd, then
(J1ω)(x, y) ≤ lim inf
h→0
v0(x + hy) − v0(x)
h
v0(x + hy) − v0(x)
h
≤ (J2ω)(x, y).
≤ lim sup
h→0
18
Proof. The first part and (1) are simple consequences of the definitions. The details
are left for the reader. It is also obvious that (3) follows from (1) and (2) if we
replace y in (2) by hy, and let h turns to zero.
It remains to prove (2), and then we may assume that y 6= 0 and x are fixed. By
(2.3) it follows that for every ε > 0, there is an element y0 ∈ Rd such that
ω(x + y + y0)
+ ε.
v0(x + y) ≤
ω(y0)
By Taylor's formula we get for some θ ∈ [0, 1],
v0(x + y) ≤
+ h(∇ω)(x + θy + y0), yi
ω(x + y0)
ω(y0)
ω(y0)
+ ε
≤ v0(x) + (J2ω)(x + θy, y) + ε ≤ v0(x) + sup
0≤t≤1
(J2ω)(x + ty, y) + ε.
Since ε was arbitrary chosen, the last inequality in (2) follows.
In the same way, let ε > 0 be arbitrary and choose y1 ∈ Rd such that
ω(x + y1)
+ ε.
v0(x) ≤
ω(y1)
By Taylor's formula we get for some θ ∈ [0, 1],
v0(x + y) ≥
ω(x + y + y1)
ω(x + y1)
ω(y1)
=
ω(y1)
ω(y1)
≥ v0(x) + (J1ω)(x + θy, y) − ε ≥ v0(x) + inf
0≤t≤1
and the first inequality in (2) follows. The proof is complete.
+ h(∇ω)(x + θy + y1), yi
(J1ω)(x + ty, y) − ε,
(cid:3)
2.2. Subfamilies of Gaussian type weights. Next we discuss further appropri-
ate conditions for subfamilies to P 0
Q(Rd) and show that these subfamilies contain
both PE(Rd) as well as all weights of the form ω(x) = Cecxγ , when C > 0, c ∈ R
and 0 ≤ γ < 2.
In the following definition we list most of the needed properties. The definitions
involve global conditions of the form
ω(21/2x)e−ε(1−λ2)x2/2
ω(21/2λx)
≤ Cε,
1 − θε < λ < 1, x ∈ Rd,
(2.5)
and
lim
λ→1−(cid:16) sup
x2∈V ⊥
ω0(21/2x)e−ε(1−λ2)x2/2
ω0(21/2λx)
(cid:17) ≤ 1,
x1 ∈ V, x2 ∈ V ⊥, x = x1 + x2 ∈ Rd,
(2.6)
for some vector space V . Here the dilation factor 21/2, is needed because of the
relation between the Bargmann transform and the short-time Fourier transform in
(1.29), and the quasi-norms in Definition 1.14.
19
Definition 2.3. Let V ⊆ Rd be a vector space.
(1) The weight ω ∈ P 0
Q(Rd) is called dilated suitable with respect to ε ∈ (0, 1],
if there are constants Cε > 0 and θε ∈ (0, 1) such that (2.5) holds. If both
ω and 1/ω are dilated suitable with respect to every ε ∈ (0, 1], then ω is
called strongly dilated suitable;
(2) The weight ω ∈ P 0
Q(Rd) is called narrowly dilated suitable with respect to
ε ∈ (0, 1] and V , if it is dilated suitable with respect to ε, and (2.6) holds
for every x1 ∈ V and some equivalent continuous weight ω0 to ω. If ω and
1/ω are narrowly dilated suitable with respect to V and every ε ∈ (0, 1],
then ω is called strongly narrowly dilated suitable.
The set of strongly dilated and strongly narrowly dilated suitable weights with
D,V (Rd) respectively. Furthermore we
D,V (Rd) is increasing
respect to V are denoted by PD(Rd) and P 0
set P 0
with respect to V , and that P 0
D,V (Rd), when V = {0}. We note that P 0
D(Rd) = P 0
If ω is v-moderate for some v, then ω is moderated by some function which grow
exponentially. It follows easily that ω satisfies (2.5) in this case. Hence, any weight
in PE is dilated suitable.
D,V (Rd) ⊆ PD(Rd).
In the following proposition we stress the latter property and prove that we in
fact have that PE ⊆ P 0
D.
Proposition 2.4. Let V ⊆ Rd be a vector space. Then the following inclusions
are true:
P(Rd) ⊆ PE(Rd) ⊆ P 0
D,V (Rd) ⊆ PD(Rd),
G(Rd), P 0
D(Rd)\ P 0
Q(Rd)\ PG(Rd) ⊆ P 0
D(Rd) ⊆ P 0
Q(Rd)[ PG(Rd) ⊆ PQ(Rd)
P 0
G(Rd) ⊆ P 0
Furthermore, if C > 0, 0 ≤ γ < 2, t ∈ R, and ω(x) = Cetxγ
, x ∈ Rd, then
G(Rd).
D(Rd) ∩ P 0
ω ∈ P 0
Proof. All the inclusions, except PE ⊆ P 0
definitions.
D are immediate consequences of the
we may assume that ω is elliptic.
Therefore, let ω ∈ PE(Rd), and let v0 be the same as in (2.3). By Lemma 2.1
We shall prove that (2.5) and (2.6) holds for ω and 1/ω, for every choice of
ε ∈ (0, 1], and since PE is a group under multiplications, it suffices to prove these
relations for ω. Since the left-hand side of (2.5) is equal to 1 as x = 0, the result
follows if we prove
ω(x)e−ε(1−λ2)x2/2
≤ 1 + C0√1 − λ,
for some constant C0 which depends on ε ∈ (0, 1).
R > 0 such that
ω(λx)
Let ε ∈ (0, 1) be fixed. We have v0(x) ≤ CeCx for some C > 1, and we choose
0 < λ < 1,
(2.7)
CeRC−εR2/2 < 1.
20
(1 − λ)xe−ε(1−λ)x2/2 ≤ R√1 − λ
as x ≤ R/√1 − λ, we obtain
v0((1 − λ)x)e−ε(1−λ)x2/2 ≤ 1 + C1R√1 − λ,
x ≤ R/√1 − λ,
Next we consider the case R/√1 − λ ≤ x ≤ R/(ε(1 − λ)). Then we have
and (2.7) follows in this case.
Since 0 < λ < 1 we have λ2 < λ. Hence Lemma 2.2 gives
ω(x)e−ε(1−λ2)x2/2
ω(λx)
≤ v0((1 − λ)x)e−ε(1−λ)x2/2.
(2.8)
We shall estimate the right-hand side, and start by considering the case when
x ≤ R/√1 − λ. Since v0(0) = 1 and (1 − λ)x stays bounded, the right-hand side
of (2.8) can be estimated by
v0((1 − λ)x)e−ε(1−λ)x2/2 ≤ v0((1 − λ)x) − v0(0)e−ε(1−λ)x2/2 + e−ε(1−λ)x2/2.
≤ C1(1 − λ)xe−ε(1−λ)x2/2 + 1,
where the last inequality follows from the fact that v0 is locally Lipschitz continuous,
in view of Lemma 2.2. Since
v0((1 − λ)x) ≤ CeC(1−λ)x ≤ CeRC/ε
and
which gives
e−(1−λ)x2/2 ≤ e−R2/2,
v0((1 − λ)x)e−ε(1−λ)x2/2 ≤ CeRC/ε−R2/2 = C 1−1/ε(cid:16)CeRC−εR2/2(cid:17)1/ε
< 1,
(2.9)
where the last inequality follows from our choice of R, together with the fact that
C > 1. This proves (2.7) in this case.
Finally, we consider the case R/(ε(1 − λ)) ≤ x. From the assumptions on R, it
follows that R > 2C/ε, which implies that
C − x/2 < C − R/(2ε) < C − εR/2 < 0.
(2.10)
Then
v0((1 − λ)x)e−(1−λ2)x2/2 ≤ CeC(1−λ)xe−(1−λ)x2/2
≤ Ce(1−λ)x(C−x/2) ≤ CeR(C−εR/2)/ε < 1,
where the last inequalities follows from (2.9) and (2.10). This gives (2.7) also for
R/(ε(1 − λ)) ≤ x, and the proof is complete.
(cid:3)
In most of our investigations, the pairs of weights and mixed norm spaces fulfill
the conditions in the following definition.
21
Definition 2.5. Let B = Lp(V ) be a mixed norm space on Rd such that (1.18)
Q(Rd). Then the pair (B, ω)
and (1.20) hold for some p ∈ [1,∞]n, and let ω ∈ P 0
is called feasible (strongly feasible) on Rd if one of the following conditions hold:
(1) ν1(B) > 1 and ω is dilated suitable with respect to ε = 1 (ω is strongly
dilated suitable);
(2) ν2(B) < ∞, and ω is dilated suitable with respect to ε = 1 (ω is strongly
dilated suitable);
(3) p1 = ∞, 1 < p2, . . . , pn−1 < ∞, pn = 1, and ω is narrowly dilated suitable
with respect to ε = 1 and V = {0} (ω is strongly narrowly dilated suitable
with respect to V = {0}).
In some situations it is convenient to separate the case (3) in Definition 2.5 from
the other ones. Therefore we say that the pair (B, ω) is narrowly feasible (strongly
narrowly feasible) if it is feasible and satisfies (3) in Definition 2.5.
We note that if B fulfills (1) or (2) in Definition 2.5 and ω ∈ PE(Rd), then the
pair (B, ω) is feasible. If instead B fulfills (3), then Proposition 2.4 shows that
the pair (B, ω) is narrowly feasible.
The following result related to Lemma 2.1 shows that for any weight in PG(Rd),
it is always possible to find a smooth equivalent weight.
Proposition 2.6. Let ω be a weight on Rd such that (1.3) holds. Then ω ∈
PG(Rd). Furthermore, there is a weight ω0 ∈ PG(Rd) ∩ C∞(Rd) such that the
following is true:
(1) for every multi-index α, there is a constant Cα such that
∂αω0(x) ≤ Cαω0(x)hxiα;
(2) (2.1) holds for some constant C;
(3) if in addition ω is rotation invariant, then ω0 is rotation invariant.
In the proof and later on we let Br(a) denote an open ball in Rd or Cd with
radius r ≥ 0 and center at a in Rd or Cd, respectively.
Proof. Let B = Bc(0), where c be the same as in (1.3), and let 0 ≤ ψ ∈ C∞0 (B)
invariant such that 0 ≤ ψ0 ≤ 1 and ψ0(x) = 1 when x ∈ 2B. Then it follows by
straight-forward computations that (1) -- (3) are fulfilled when
be rotation invariant such that R ψ dx = 1. Also let ψ0 ∈ C∞0 (3B) be rotation
ω0(x) ≡ ψ0(x) + (1 − ψ0(x))xdZ ψ(x(x − y))ω(y) dy.
It remains to prove that ω0, ω ∈ PG(Rd), and then it suffices to prove that ω0
satisfies (1.4). Let 1 ≤ t ∈ R, x0 ∈ Rd be fixed such that x0 = 1, and let
C1 = inf
x=1
ω0(x)
and C2 = sup
x=1
ω0(x).
Then (1) implies that
ψ(t) = ψx0(t) ≡ ω0(tx0),
22
t ≥ 1,
satisfies
ψ′(t) − Ctψ(t) ≤ 0 and ψ′(t) + Ctψ(t) ≥ 0,
0 < C1 ≤ ψ(1) ≤ C2,
for some constant C > 0 which is independent of x0. This implies that
C1e−C(t2−1)/2 ≤ ψ(t) ≤ C2eC(t2−1)/2,
t ≥ 1.
Since C1 and C2 are independent of the choice of x0 on the unit sphere, (1.4) follows
from the latter inequalities. Hence ω0 ∈ PG, and the proof is complete.
(cid:3)
We finish this section by proving the following result on existence rotation in-
variant weights in our families of weights.
Proposition 2.7. Let P be equal to
P(Rd), PE(Rd), P 0
D(Rd), P 0
G(Rd),
PD(Rd), P 0
Q(Rd), PG(Rd) or PQ(Rd).
If ω ∈ P, then there are rotation invariant weights ω1, ω2 ∈ P such that
C−1ω1 ≤ ω ≤ Cω2.
For the case P = P 0
Q we need the following lemma.
Lemma 2.8. Let f ∈ L∞loc(Rd) be such that for each ε > 0, there is a constant
Cε > 0 such that
Then there is a rotation invariant ω ∈ P 0
Proof. Let
Q(Rd) such that f ≤ ω.
f (x) ≤ Cεeεx2
.
Then
g ≡ f + e
and h0(x) ≡
log g(x)
x2
.
h(x) = 0,
lim
x→∞
(2.11)
(2.12)
for h = h0 due to the assumptions.
Now set
h1(x) =Z ϕ(x − y)" sup
y0≥y−1
h0(y0)# dy,
x ≥ 2,
where 0 ≤ ϕ ∈ C∞0 (Rd) is rotation invariant, supported in the unit ball, and
satisfies kϕkL1 = 1. Then it follows that h1 is rotation invariant, smooth and
larger than h0 in Ω2, where Ωr ≡ Rd \ Br(0). Furthermore, (2.12) holds for h = h1,
and since h0 is bounded in Ω1, it follows that h(α)
is bounded in Ω2 for every
1
multi-index α.
Hence, if y ≤ 1/x with x ∈ Ω3, Taylor's formula gives
h1(x + y) + h1(x − y) − 2h1(x) ≤ Cy2 ≤ C/x2.
23
By again using the fact that h1 is bounded in Ω2, it follows that
ω(x) ≡(supy≤3 g(y) when x ≤ 3,
eh1(x)x2
when x ≥ 3,
Q(Rd), and the result follows.
is rotation invariant, larger than g and fulfills (1.3)′. This together with (2.12)
shows that ω ∈ P 0
(cid:3)
Proof of Proposition 2.7. If P = P 0
Q, then the result is an immediate consequence
of Lemma 2.8 and the fact that P 0
Q is a group under multiplications. By straight-
forward computations it also follows that ω in Lemma 2.8 is dilated suitable or
strongly dilated suitable when this is true for f . This proves the statement for
P = PD(Rd) and for P = P 0
For P = P we may choose ωj = Cjh · iNj for appropriate constants Cj and Nj,
and if P = PE we may choose ωj = Cjecj · sj for appropriate constants 0 < sj ≤ 1,
cj and Cj. If instead P = PG or P = PQ, then similar is true after the condition
on sj is replaced by sj = 2.
G, and
It remains to consider the case P = P 0
G. Therefore, assume that ω ∈ P 0
D(Rd).
set
ω2(x) = sup
z=x
ω(z)
and ω1(x) = inf
z=x
ω(z).
Now choose c and C such that (1.3) holds, and let x, y be such that x ≥ 2c and
y ≤ c/x. Then for each ε > 0, there exists z ∈ Rd such that z = x and
By (1.3) we get
ω2(x) ≤ ω(z) + ε.
ω2(x) ≤ ω(z) + ε ≤ C inf
y0≤c/x
ω(z + y0) + ε
y0≤c/x
≤ C inf
ω2(x + y0) + ε ≤ Cω2(x + y) + ε.
This proves that ω2(x) ≤ Cω2(x + y). In the same way it follows that ω2(x + y) ≤
Cω2(x). Since ω2 fulfills similar types of estimates as ω, it follows that ω2 is
subgaussian.
In the same way it follows that ω1 is subgaussian, and the result
follows for P = P 0
G.
(cid:3)
2.3. Examples. Next we give some examples on weights in P 0
D(Rd). First we
note that any weight of the form σs and ω1 in Example 1.3 belongs to P(Rd)
in view of Proposition 2.4.
In order to give other examples it is convenient to
consider corresponding logarithmic conditions on those weights. We note that if ω
is a weight on Rd and ϕ(x) = log ω(x), then (1.4) is equivalent to
ϕ(x) ≤ C + εx2,
(2.13)
for some positive constants C and ε. The conditions (1.3) and (1.3)′ are the same
as
ϕ(x + y) − ϕ(x) ≤ C,
24
x ≥ 2c,
y ≤ c/x,
(2.14)
and
ϕ(x + y) + ϕ(x − y) − 2ϕ(x) ≤ C,
x ≥ 2c,
y ≤ c/x,
(2.14)′
respectively, for some positive constants c and C. Finally, ω and 1/ω are dilated
suitable with respect to ε ∈ (0, 1], if and only if there are constants Cε > 0 and
θε ∈ (0, 1) such that
ϕ(λx) − ϕ(x) ≤ Cε + ε(1 − λ2)x2,
(2.15)
and ω and 1/ω are narrowly dilated suitable with respect to ε ∈ (0, 1] and V = {0}
when
(2.16)
λ→1− (cid:16)ϕ(λx) − ϕ(x) − ε(1 − λ2)x2(cid:17) = 0.
In particular, the following lemma is an immediate consequence of the definitions.
1 − θε < λ < 1,
lim sup
Lemma 2.9. Let ω be a weight on Rd, and let ϕ(x) = log ω(x). Then the following
is true:
(1) ω ∈ P 0
(2) ω ∈ P 0
G(Rd), if and only if (2.14) holds for some positive constants c and
C, and for every ε > 0, there is a positive constant C such that (2.13) holds:
Q(Rd), if and only if (2.14)′ holds for some positive constants c and
C, and for every ε > 0, there is a positive constant C such that (2.13) holds:
D(Rd), if and only if (2.14)′ holds for some positive constants c and
C, and for every ε > 0, there are positive constants C, Cε and θε ∈ (0, 1)
such that (2.13), (2.15) and (2.16) hold.
(3) ω ∈ P 0
Example 2.10. Let ω(x) = hxithxi, x ∈ Rd, for some choice of t ∈ R. Then it
follows by straight-forward computations that ϕ(x) = log ϕ(x) = hxi loghxi satisfies
(2.13) -- (2.16). Hence ω ∈ P 0
Example 2.11. Let ω(x) = Γ(hxi + 1 + r) and ϕ(x) = log ω(x), where Γ is the
gamma function, and r > −2 is real. Then we have
D(Rd) ∩ P 0
G(Rd).
ω(x) = (2π(hxi + r))1/2(cid:18)hxi + r
e (cid:19)hxi+r
(1 + o(hxi−1)),
by Stirling's formula. This is gives
1 + 2r
ϕ(x) =
log(2π) +
1
2
loghxi + hxi(loghxi − 1) + ψ(x),
2
where ψ(x) is continuous and satisfies
lim
x→∞hxiψ(x) = 0.
(2.17)
(2.18)
By straight-forward computations it follows that the first three terms in (2.17)
satisfy the conditions (2.13) -- (2.16). Furthermore, the condition (2.18) together
with the proof of Proposition 2.4 show that also ψ(x) satisfies (2.13) -- (2.16). Con-
sequently, ω ∈ P 0
D(Rd) ∩ P 0
G(Rd).
The following result shows that there are weights in (P 0
D∩P 0
G)\P which fulfills
(0.1) for some polynomial v.
25
Proposition 2.12. Let ω(x, ξ) = (1+xrξr)s for some r > 0 and s ∈ R\0. Then
ω belongs to (P 0
G(R2d))\ P(R2d) and fulfills (0.1) for some polynomial
v.
D(R2d)∩P 0
D and P are groups under multiplications, and invariant under
G, P 0
Proof. Since P 0
mappings ω 7→ ωt for t 6= 0, and that ω is equivalent to
ω0(x, ξ) = (1 + x1/2ξ1/2)2rs,
we may assume that r = 1/2 and s = 1.
It is obvious that ω fulfills (0.1) for some polynomial v, and by straight-forward
computations it also follows that (1.3) is fulfilled. Furthermore, by choosing x, y, ξ, η ∈
Rd in such way that η = 1/y and x = ξ = 1, it follows that
1 + x + y1/2ξ + η1/2
ω(x + y, ξ + η)
2
= ∞.
sup
y,η
ω(y, η)
≥ sup
η=1/y
It remains to prove that ω ∈ P 0
This implies that ω /∈ P.
D, which follows if we prove that for every ε >,
then (2.6) holds for V = {0}, after Rd and x have been replaced by R2d and (x, ξ),
respectively.
Here it is obvious that (2.6) is true with ω = ω0, since
ω(λx, λξ)e−ε(1−λ2)(x2+ξ2)/2
ω(x, ξ)
≤ 1
with equality when x = ξ = 0. Let
h(t1, t2) =
Since
(1 + √t1t2)e−ε(1−λ)(t2
1+t2
2)/2
1 + λ√t1t2
,
t1, t2 ≥ 0.
h(x,ξ) =
ω(x, ξ)e−ε(1−λ)(x2+ξ2)/2
ω(λx, λξ)
≥
ω(x, ξ)e−ε(1−λ2)(x2+ξ2)/2
ω(λx, λξ)
,
and h(0, 0) = 1, the result follows if we prove
lim
λ→1−(cid:0) sup
t1,t2≥0
h(t1, t2)(cid:1) = 1.
Now, by straight-forward computations it follows that
h(t1, t2) = h(t0, t0), where
sup
t1,t2≥0
t0 = −
1 + λ
2λ
+s(cid:18)1 + λ
2λ (cid:19)2
+
1 − ε
ελ
,
and it is straight-forward to control that h(t0, t0) → 1 as λ → 1−. The proof is
complete.
(cid:3)
26
3. Harmonic estimates and mapping properties for the Bargmann
transform on modulation spaces
In the first part of the section we establish certain invariance properties of spaces
of harmonic or analytic functions. Thereafter we apply these properties to prove
that A0(ω, B) = A(ω, B), for appropriate weights ω. In the end of the section we
use these results to prove general properties of A(ω, B) and M(ω, B), for example
that they are Banach spaces when B is a mixed norm space.
3.1. Analytic and harmonic estimates. Let Ω be an admissible family of weights
on Rd, and let B be a mixed quasi-norm space on Rd. (Cf. Definition 1.4.) Then
we prove that subsets of
E1(Ω, B) ≡ { f ∈ Lr
loc(Rd) ; f ω ∈ B for every ω ∈ Ω}
and
E2(Ω, B) ≡ { f ∈ Lr
loc(Rd) ; f ω ∈ B for some ω ∈ Ω},
r = ν1(B),
of analytic or harmonic functions are independent of the choice of B. Recall
here that ν1(B) is the smallest involved Lebesgue exponent in B, and belongs to
(0,∞]. Also recall that B is a mixed norm space, if and only if ν1(B) ≥ 1. Some
restrictions are needed when considering subsets of harmonic functions.
It is easy to prove one direction. In fact, by the definitions and Holder's inequality
we get
Ej(Ω, L∞(Rd)) ⊆ Ej(Ω, B) ⊆ Ej(Ω, Lr(Rd)),
where j = 1, 2 and r = min(1, ν1(B))
(3.1)
(with continuous inclusions).
In order to establish opposite inclusions to (3.1), for corresponding subsets of
analytic or harmonic functions, we will use techniques based on harmonic estimates
for such functions.
We start with the following result. Here and in what follows we let H(Rd) be
the set of harmonic functions on Rd.
Proposition 3.1. Let Ω ⊆ PG(Rd) be an admissible family of weights on Rd, and
let B1 and B2 be mixed norm space on Rd. Then
Ej(Ω, B1)\H(Rd) = Ej(Ω, B2)\H(Rd),
j = 1, 2.
(3.2)
Proof. It suffices to prove Ej(Ω, L1) ∩ H ⊆ Ej(Ω, L∞) ∩ H, in view of (3.1).
Assume that f ∈ E2(Ω, L1) ∩ H. By (1.6) and the assumptions we have
kfkL1
(ω)
< ∞ and ω1 ≤ Ch · i−dω,
(3.3)
for some ω, ω1 ∈ Ω and some constant C > 0. Let c and C be the same as in (1.3).
Then the result follows if we prove
kf χkL∞
(ω1)
27
< ∞,
(3.4)
when χ is the characteristic function of { x ∈ Rd ; x ≥ 2c }.
Since f ∈ H(Rd), the mean-value property for harmonic functions gives
f (x) = c−1
f (x + y) dy,
where cd is the volume of the d-dimensional unit ball. If x ≥ 2c, then hxi ≤ C1x
for some constant C1 > 0, and (1.3) and (3.3) give
f (x)ω1(x) =(cid:12)(cid:12)(cid:12)(cid:12)c−1
d (x/c)dZy≤c/x
f (x + y) dy(cid:12)(cid:12)(cid:12)(cid:12)
f (x + y)ω(x) dy
d (x/c)dω1(x)Zy≤c/x
≤ C1Zy≤c/x
≤ C2Zy≤c/x
< ∞,
for some constants C1 and C2. This gives (3.2) for j = 2. By similar arguments,
(3.2) also follows for j = 1. The details are left for the reader. The proof is
complete.
(cid:3)
f (x + y)ω(x + y) dy ≤ C2kfkL1
(ω)
Next we discuss similar questions for spaces of analytic functions, i. e. we present
sufficient conditions in order for the identity
Ej(Ω, B1) ∩ A(Cd) = Ej(Ω, B2) ∩ A(Cd),
j = 1, 2.
(3.5)
should hold.
Theorem 3.2. Let Ω be an admissible family of weights on Cd ≃ R2d, and let B1
and B2 be mixed quasi-norm spaces on Cd. Then (3.5) holds.
Furthermore, for every fixed ω ∈ Ω, there are ω1, ω2 ∈ Ω and constant C > 0
such that
C−1kFkA(ω1,B2) ≤ kFkA(ω,B1) ≤ CkFkA(ω2,B2),
F ∈ A(Cd).
(3.6)
For the proof we need the following lemma, concerning mean-value properties
for analytic functions.
Lemma 3.3. Let ν be a positive Borel measure on Cd which is rotation invariant
under each coordinate z1, . . . , zd ∈ C, T1, . . . , Tn be (complex) d × d-matrices, and
let F1, . . . , Fn ∈ A(Cd). Also let r > 0, and let Ω ⊆ Cd be compact and convex,
which is rotation invariant under each coordinate z1, . . . , zd ∈ C. Then
nYj=1
nYj=1
Fj(z) =
Fj(z)r ≤
1
ν(Ω)ZΩ
ν(Ω)ZΩ
1
nYj=1
nYj=1
28
Fj(z + Tjw) dν(w)
(3.7)
Fj(z + Tjw)r dν(w).
(3.8)
and
mean-value property for harmonic functions we get
j=1 Fj(z + Tjw). Then w 7→ G(z, w) is analytic. By the
Proof. Let G(z, w) = Qn
G(z, 0) =
which is the same as (3.7).
1
ν(Ω)ZΩ
G(z, w) dν(w),
From the same analyticity property it follows that the map w 7→ G(z, w)r is
subharmonic, in view of [38, Corollary 2.1.15]. Hence, by [38, Theorem 2.1.4] we
get
G(z, 0)r ≤
1
ν(Ω)ZΩ G(z, w)r dν(w),
which is the same as (3.8), and the proof is complete.
(cid:3)
Proof of Theorem 3.2. We only prove (3.5) for j = 2, leaving the small modifica-
tions of the case j = 1 for the reader. By (3.1) it suffices to prove
E2(Ω, Lr) ∩ A(Cd) ⊆ E2(Ω, L∞) ∩ A(Cd),
r = ν1(B).
Therefore, assume that F ∈ E(Ω, Lr) ∩ A(Cd). We shall mainly follow the ideas
in Proposition 3.1. By (1.6) and the assumptions we have
< ∞ and ω1 ≤ Ch · i−2dω,
kFkLr
(ω)
for some ω, ω1 ∈ Ω and some constant C > 0. Let c and C be the same as in (1.3)′.
Then the result follows if we prove that (3.4) holds when χ is the characteristic
function of { z ∈ Cd ; z ≥ 2c }.
By Lemma 3.3 and Cauchy-Schwartz inequality we get
F (z)ω1(z)r ≤ c−1
F (z + w)F (z − w)r/2 dλ(w)
2d (z/c)dω1(z)rZw≤c/z
≤ C1Zw≤c/z
F (z + w)F (z − w)ω(z)2r/2 dλ(w)
≤ C2Zw≤c/z
F (z + w)ω(z + w)r/2F (z − w)ω(z − w)r/2 dλ(w)
≤ C2ZCd F (z + w)ω(z + w)r/2F (z − w)ω(z − w)r/2 dλ(w)
≤ C2(cid:18)ZCd F (z + w)ω(z + w)r λ(w)(cid:19)1/2
·(cid:18)ZCd F (z − w)ω(z − w)r λ(w)(cid:19)1/2
< ∞,
for some constants C1 and C2. Here recall that dλ(z) is the Lebesgue measure on
Cd. This proves (3.5) for j = 2, and (3.6). By similar arguments, (3.5) follows for
j = 1. The details are left for the reader, and the proof is complete.
(cid:3)
= C2kFkr
Lr
(ω)
29
3.2. Mapping properties for the Bargmann transform on modulation
spaces. Next we prove that A0(ω, B) is equal to A(ω, B) for every choice of ω in
P 0
Q and mixed quasi-norm space B.
Theorem 3.4. Let B be a mixed quasi-norm space on R2d ≃ Cd and let ω ∈
P 0
Q(Cd). Then A0(ω, B) = A(ω, B), and the map f 7→ Vf from M(ω, B) to
A(ω, B) is isometric and bijective.
We need some preparations for the proof, and start to show that M(ω, B) is
a Banach space when ν1(B) ≥ 1, which is a consequence of the following result.
Here, for each ψ ∈ S1/2(Rd) \ 0, 0 < ω ∈ L∞loc(R2d) and mixed norm space B on
R2d, we let Mψ(ω, B) be the set of all f ∈ (S1/2)′(Rd) such that
kfkMψ(ω,B) ≡ kVψf · ωkB < ∞.
Proposition 3.5. Let ψ ∈ S1/2(Rd) \ 0 and ω ∈ L∞loc(R2d) be such that for every
ε > 0 there is a constant Cε > 0 such that
ω ≥ C−1
ε e−ε · 2
Also let B be a mixed norm space on R2d. Then Mψ(ω, B) is a Banach space.
Proof. We may assume that kψkL2 = 1, and start by proving that Mψ(ω, B) is
continuously embedded in (S1/2,ε)′(Rd), for every choice of ε > 0. We have ψ ∈
S1/2,ε0, for some ε0 > 0.
Therefore let ε > 0 be arbitrary. By a straight-forward combination of Lemma
1.6 and Proposition 1.8 and their proofs it follows that for some δ > 0 we have
.
(3.9)
kVψϕkS1/2,δ ≤ CεkϕkS1/2,εkψkS1/2,ε0
,
(3.10)
for some constant Cε > 0, which only depends on ε. (Cf. [15].)
Now we recall that
(f, ϕ)L2(Rd) ≡ (Vψf, Vψϕ)L2(R2d), ϕ ∈ S1/2,ε(Rd),
for any f ∈ Mψ(ω, B). Hence, (3.9), (3.10) and Holder's inequality give
(f, ϕ)L2 = (Vψf, Vψϕ)L2 ≤ kfkMψ(ω,B)kVψϕ/ωkB′
≤ C1kfkMψ(ω,B)kVψϕ · eδ · 2
· h · i−2d−1kB′
≤ C2kfkMψ(ω,B)kVψϕ · eδ · 2
kL∞kh · i−2d−1kB′ ≤ C3kfkMψ(ω,B)kVψϕkS1/2,δ,
≤ C4kfkMψ(ω,B)kϕkS1/2,ε,
for some constants C1, . . . , C4 > 0. Consequently, Mψ(ω, B) is continuously em-
bedded in (S1/2,ε)′(Rd) and in (S1/2)′(Rd) for every ε > 0.
Now let {fj}∞j=1 be a Cauchy sequence in Mψ(ω, B). Since (S1/2,ε)′(Rd) decreases
when ε decreases, (1.7)′ in combination with the previous embedding properties
show that there is a element f ∈ (S1/2)′(Rd) (which is independent of ε) such that
fj → f in (S1/2,ε)′ as j → ∞.
30
Since Vψfj → Vψf pointwise, it now follows that f ∈ Mψ(ω, B) by Fatou's
lemma. This shows that Mψ(ω, B) is a Banach space, and the proof is complete.
(cid:3)
Next we consider the case when B = L2 and ω is rotation invariant in each
coordinate. In this case we have the following.
Lemma 3.6. Let ω ∈ P 0
Q(R2d) be rotation invariant. If
then {cαzα}α∈Nd
cα = kzαk−1
(ω)(Cd).
A2
(ω)
,
is an orthonormal basis for A2
In the proof of Lemma 3.6 and in several other situations later on, we encounter
the integral
Z∆d
eihn,θi dθ =((2π)d, n = 0,
0,
n 6= 0,
when ∆d = [0, 2π]d
and n ∈ Zd.
(3.11)
Proof. First we prove that the scalar product (zα, zβ)A2
polar coordinates we have
(ω)
is zero when α 6= β. By
where
z = (r1eiθ1, . . . , rdeiθd),
r = (r1, . . . , rd) ∈ [0,∞)d,
and θ = (θ1, . . . , θd) ∈ ∆d.
Furthermore it follows from the assumptions that ω(z) = ω0(2−1/2r), for some
positive function ω0 on [0,∞)d. Hence (1.32) and (3.11) give
(zα, zβ)A2
(ω)
= π−dI(ω0)(α + β) ·Z∆d
eihα−β,θi dθ = 2dI(ω0)(2α)δα,β,
where
I(ω0)(α) =Z[0,∞)d
ω0(r)rαe−r2
r1 · · · rd dr > 0,
α ∈ Nd,
and δα,β is the Kronecker's delta function. In particular, (zα, zβ)A2
only if α 6= β, and we have proved that {cαzα} is an orthonormal system for A2
Hahn-Banach's theorem, it suffices to prove that if F ∈ A2
= 0 for all α ∈ Nd,
It remains to prove that the set of linear combinations of cαzα spans A2
= 0, if and
(ω).
(ω). By
(F, zα)A2
(ω), and
(3.12)
(ω)
(ω)
implies that F = 0.
Therefore assume that (3.12) holds. Since F is entire, it follows that its Taylor
series expansion
F (z) =Xβ
aβzβ,
aβ =
F (β)(0)
β!
31
(3.13)
is locally uniformly convergent, and that
Xβ
aβrβ < ∞
(3.14)
holds. Hence (3.11) gives
0 = (F, zα)A2
(ω)
=ZCd Xβ
= π−dZ[0,∞)d Z∆d Xβ
= π−dZ[0,∞)d Xβ
aβzβ! zαω(21/2z) dµ(z)
aβrα+βeihβ−α,θi! dθ! ω0(r)e−r2
eihβ−α,θi dθ(cid:19)! ω0(r)e−r2
aβrα+β(cid:18)Z∆d
r1 · · · rd dr
r1 · · · rd dr
= 2dI(ω0)(2α)aα.
Since I(ω0)(2α) > 0, we get aα = 0 for every α. Consequently, F is identically zero,
and the proof is complete.
(cid:3)
We may now prove Theorem 3.4 in the important special case that B = L2 and
ω is rotation invariant in every coordinate.
Proposition 3.7. If ω ∈ P 0
A0(ω, L2) = A(ω, L2).
Q(Cd) is rotation invariant in each coordinate, then
Proof. We use the same notations as in the proof of Lemma 3.6. The image under
the Bargmann transform V of the hermite function hα is zα/(α!)1/2. Since M 2
(ω) is
a Banach space in view of Proposition 3.5, and V is isometric and injective from
M 2
(ω), it follows from Lemma 3.6 that
(ω) to A2
is an orthonormal basis of M 2
(ω), and that V is bijective from M 2
(ω) to A2
(ω).
{(α!)1/2cαhα}α∈Nd
(cid:3)
Remark 3.8. Lemma 3.6 and Proposition 3.7 give equalities between weighted lp-
norms of the Taylor coefficients and weighted Lp norm of corresponding entire
functions, when the involved weights are rotation invariant and p = 2. In general
it is difficult to find such equalities between coefficients and functions in other
situations when the weights are not rotation invariant, or p is not equal to 2. We
refer to [9] for positive results in this directions.
Proof of Theorem 3.4. By Proposition 1.15 it follows that the map f 7→ Vf is an
isometric injective map from M(ω, B) to A(ω, B). We have to show that this map
is surjective.
Therefore assume that F ∈ A(ω, B). Since M(ω1, B) ⊆ M(ω2, B), as ω2 ≤
Cω1, it follows from Theorem 3.2 and Proposition 3.7 that there is an element
32
f ∈ M 2
(ω1)(Rd) ⊆ S′1/2(Rd) such that F = Vf , for some ω1. We have
kfkM (ω,B) = kVfkA(ω,B) = kFkA(ω,B) < ∞.
Hence, f ∈ M(ω, B), and the result follows. The proof is complete.
(cid:3)
As a consequence of Theorems 3.2 (2) and 3.4, we may identify the space Υ(Rd)
in Remark 1.10 with unions of modulation spaces. More precisely we have the
following.
Proposition 3.9. Let B be a mixed quasi-norm space on R2d. Then
[ M(ω, B) = Υ(Rd),
where the union is taken over all rotation invariant ω ∈ P 0
Furthermore, if 0 < γ < 2, 0 < ε < min(γ, 2 − γ) and ω(x, ξ) = Cec(xγ+ξγ),
for some constants C > 0 and c ∈ R which are independent of x, ξ ∈ Rd, then the
following is true:
Q(R2d).
(1) if c > 0, then S1/(γ+ε)(Rd) ⊆ M(ω, B) ⊆ S1/(γ−ε)(Rd);
(2) if c < 0, then S′1/(γ−ε)(Rd) ⊆ M(ω, B) ⊆ S′1/(γ+ε)(Rd).
Proof. By Theorems 3.2 (2) and 3.4, we may assume that B = L∞. By the
assumptions on ω we get M∞(ω) ⊆ Υ.
that P 0
Let f ∈ Υ(Rd). Then it follows from Proposition 1.9, Lemma 2.8 and the fact
Q is a group under multiplications that
Vφf ≤ 1/ω
for some ω ∈ P 0
and the first part follows.
Q(R2d), where φ(x) = π−d/4e−x2/2. This is the same as f ∈ M∞(ω),
The assertions (1) and (2) follow by similar arguments in combination of Propo-
(cid:3)
sition 1.9 and are left for the reader. The proof is complete.
4. Basic properties for spaces of analytic functions and
modulation spaces
In this section we establish basic properties for the spaces A(ω, B) and M(ω, B)
when ω is an appropriate weight and B is a mixed quasi-norm space. In view of
Theorem 3.4 any property of A(ω, B) carry over to M(ω, B), and vice versa. We
start by proving that these spaces are quasi-Banach spaces. Then we prove that if
ν2(B) < ∞, then P (Cd) is dense in A(ω, B), and that the dual of A(ω, B) can be
identified with A(1/ω, B′) through a unique extension of the A2 form on P (Cd). A
straight-forward consequence of the latter results is that P (Cd) is dense in A(ω, B)
with respect to the weak∗-topology, when ν1(B) > 1. (Recall Subsection 1.4 for
the definitions of ν(B) and ν2(B2).) Thereafter we introduce the concept of narrow
convergence to get convenient density properties for certain B with ν1(B) = 1 and
ν2(B) = ∞. Finally we formulate corresponding results for modulation spaces.
33
A cornerstone of these investigations concerns the projection operator
(ΠAF )(z) =Z F (w)e(z,w) dµ(w),
(4.1)
(4.2)
loc(Cd)
related to the reproducing formula (1.26). Here recall that dµ(z) = π−de−z2 dλ(z),
where dλ(z) is the Lebesgue measure on Cd. The minimal assumption on F is that
it should be locally integrable on Cd and satisfy
where p ∈ (0,∞] is fixed. We note that (4.2) is fulfilled for p = 1 if F ∈ L1
and satisfies
kLp < ∞ for every N ≥ 0,
dλ(z) < ∞ for some γ < 1.
(4.3)
kF · eN · − · 2
Z F (z)e−γz2
Lemma 4.1. Let γ, δ > 0 and let F ∈ L1
loc(Cd). Then the following is true:
(1) if p ∈ [1,∞] and F satisfies (4.2), then ΠAF is an entire function on Cd;
(2) if p ∈ (0,∞] and F satisfies (4.2) and in addition is entire, then ΠAF = F ;
(3) if 4δ(1 − γ) ≥ 1, then for some constant C > 0 it holds
k(ΠAF ) · e−δ · 2
kL1 ≤ CkF e−γ · 2
kL1
when F ∈ L1
loc(Cd) and satisfies (4.3);
(4) if γ < 3/4 and (4.3) is fulfilled, then
(F, G)B2 = (F, ΠAG)B2 = (ΠAF, G)B2,
(4.4)
for every polynomial G (which is analytic) on Cd.
Proof. By Hoder's inequality we may assume that p = 1 when proving (1). Let
E(z, w) = F (w)e(z,w)−w2. The condition (4.2) implies that if z ∈ K, where K ⊆
Cd is compact, then for each multi-index α we have that ∂α
z E(z, w) is uniformly
bounded in L1 with respect to w. The assertion (1) is now a consequence of the
fact that z 7→ E(z, w) is entire.
Since { etz−z2 ∈ L∞loc(Cd) ; t ∈ R} is an admissible family of weights, we may
assume that p = 1 in view of Theorem 3.2, when proving (2). The assertion then
follows from the same arguments as for the proof of Lemma A.2 in [47]. In order
to be self-contained we give here a proof.
For every mutli-index α we have
(∂α
z E)(z, w) = e(z,w)−w2
wαF (w)
and
(∂α
z ∂zE)(z, w) = 0.
Hence, by the assumptions it follows that the map w 7→ (∂α
z E)(z, w) belongs to
L1(Cd) for every z ∈ Cd, and that F0 ≡ ΠAF in (4.1) is analytic with derivatives
(4.5)
e(z,w)wαF (w) dµ(w).
(∂αF0)(z) ≡ZCd
34
In particular we have
(∂αF0)(0) ≡ZCd
wαF (w) dµ(w).
(4.5)′
We have to prove that F0 = F . Since both F and F0 are entire functions it
suffices to prove
for every multi-index α.
∂αF0(0) = ∂αF (0),
If w = (r1eiθ1, . . . , rdeiθd), where r = (r1, . . . , rd) ∈ [0,∞)d and θ = (θ1, . . . , θd) ∈
∆d = [0, 2π]d, then (3.13) and (4.5)′ give
∂αF0(0) =ZCd
wαF (w) dµ(w)
= π−dZ[0,∞)d(cid:16)Z∆d
rαe−ihθ,αiF (r1eiθ1, . . . , rdeiθd)r1 · · · rde−r2
dθ(cid:17) dr
= π−dZ[0,∞)d
rαr1 · · · rde−r2
Jα(r) dr,
(4.6)
where
Jα(r) =Z∆d
e−ihθ,αiF (r1eiθ1, . . . , rdeiθd) dθ =Z∆d(cid:16)Xβ
aβrβeihθ,β−αi(cid:17) dθ
By (3.14) it follows that we may interchange the order of summation and integra-
tion. This gives
Jα(r) =Xβ
aβrβZ∆d
eihθ,β−αi dθ = (2π)daαrα,
(4.7)
in view of (3.11)
By inserting (4.7) into (4.6) and taking uj = r2
j as new variables of integration,
(3.13) gives
∂αF0(0) = 2daαZ[0,∞)d
r2αr1 · · · rde−r2
= aαZ[0,∞)d
dr
uαe−(u1+···+ud) du = aαα! = ∂αF (0).
(4.8)
This proves (2).
The assertion (3) follows from the inequality
F (w)e(z,w)−w2
e−δz2
≤ F (w)e−γw2
e−(1−γ)w−(1−γ)−1z/22
,
and (4) is obtained by choosing δ > 1 in the latter estimate, giving that
(z, w) 7→ F (w)G(z)e(z,w)−w2
35
e−z2
belongs to L1(Cd × Cd) when G is a polynomial. The relation (4.4) is now an im-
mediate consequence of the reproducing formula (1.26) applied on G, and Fubbini's
theorem.
(cid:3)
Remark 4.2. We note that if F ∈ A(Cd) and satisfies (4.2), then
zαF (z) =Z wαF (w)e(z,w) dµ(w)
∂αF (z) =Z wαF (w)e(z,w) dµ(w),
∂αF (0) = (F, zα)A2
(4.9)
and
giving that
(see also [3]).
In fact, the first formula follows by replacing F by zαF in the reproducing formula
and using (2) in Lemma 4.1. For the second formula we note that the condition
(4.2) and reproducing formula give
∂αF (z) = ∂α(cid:16)Z F (w)e(z,w) dµ(w)(cid:17)
=Z ∂α
and the result follows.
z(cid:0)F (w)e(z,w)(cid:1) dµ(w) =Z wαF (w)e(z,w) dµ(w),
Remark 4.3. It follows from the the proof, and especially (4.8), of the previous
lemma that if F1, F2 ∈ A(Cd), (4.2) holds for F = Fj and p = 1, j = 1, 2, and that
(F1, G)A2 = (F2, G)A2,
then F1 = F2.
G ∈ P (Cd),
Next we prove that A(ω, B) and M(ω, B) are Banach spaces when ν1(B) ≥ 1.
Q(Cd), ω2 ∈ PQ(Cd) and B be a mixed quasi-norm
Theorem 4.4. Let ω1 ∈ P 0
space on Cd. Then the following is true:
(1) M(ω1, B) and A(ω2, B) are quasi-Banach spaces;
(2) if in addition ν1(B) ≥ 1, then M(ω1, B) and A(ω2, B) are Banach spaces.
Proof. By Theorem 3.4 it suffices to prove the result for A(ω2, B). Since the
statement is invariant under dilations, we may assume that (1.4) holds for ω = ω2
and c = 1/8. Furthermore, since it is obvious that k · kB(ω2,B) is a norm when
ν1(B) ≥ 1, it suffices to prove (1).
By Theorem 3.2 it follows that A(ω2, B) is continuously embedded in A(ω0, L1),
for some choice of ω0 ∈ PQ. Furthermore, it follows from the proof of Theorem
3.2 that we may choose ω0 such that it satisfies (1.4) with c = 1/6. Consequently,
any F in A(ω0, L1) fulfills (4.2) with p = 1.
36
Now let (Fj)∞j=1 be a Cauchy sequence in A(ω2, B). Since both B(ω2, B) and
B(ω0, L1) are quasi-Banach spaces, it exists an element F ∈ B(ω2, B) ∩ B(ω0, L1)
such that Fj → F in B(ω2, B) and B(ω0, L1) as j → ∞.
We have to prove that F ∈ A(Cd). By the assumptions and Lemma 4.1 it follows
that F0 = ΠAF in (4.1) defines an analytic function, and that
Fj(z) =Z Fj(w)e(z,w) dµ(w)
for every j. Furthermore, for each compact set K ⊆ Cd there is a constant C > 0
such that
sup
K Fj(z) − F0(z) ≤ π−dZ Fj(w) − F (w)e−w2+Cw dλ(w)
= π−dZ (cid:12)(cid:12)(cid:12)(Fj(w) − F (w))e−w2/2ω0(21/2w)(cid:12)(cid:12)(cid:12) · (e−w2/2+Cwω0(21/2w)−1) dλ(w)
≤ Cω0Z Fj(w) − F (w)e−w2/2ω0(21/2w) dλ(w) = CkFj − FkB(ω0,L1),
(4.10)
where
Cω0 = ess sup
w∈Cd (cid:0)e−w2/2+Cw/ω0(21/2w)(cid:1) < ∞,
and C is a constant. Since the right-hand side of (4.10) turns to zero as j → ∞, it
follows that Fj → F0 locally uniformly as j → ∞. This proves that F = F0, which
is analytic, and the result follows.
(cid:3)
4.1. Density and duality properties. Next we prove that if B is a mixed norm
space with ν2(B) < ∞, and that the weight ω ∈ P 0
Q(Cd) in addition should be
dilated suitable, then the set P (Cd) of polynomials on Cd is dense in A(ω, B). Fur-
thermore, in this situation we also prove that the dual of A(ω, B) can be identified
with A(1/ω, B′), through a unique extention of the A2 form on P (Cd).
An important part of these considerations concerns possibilities to approximate
elements F in A(ω, B) with their dilations F (λ · ) for 0 < λ < 1. We note that the
latter functions belong to
AP (Cd) ≡ { F ∈ A(Cd) ; F · e−(1−ε)z2/2 ∈ B for some ε > 0 },
and that
(4.11)
P (Cd) ⊆ AP (Cd) ⊆ A(ω, B), when ω ∈ P 0
Q(Cd).
This is a straight-forward consequence of Theorem 3.2 and the definitions.
P 0
Proposition 4.5. The set AP (Cd) in (4.11) is independent of the mixed quasi-
norm space B on Cd. Furthermore, if B is a mixed norm space on Cd and ω ∈
Q(Cd), then P (Cd) is dense in AP (Cd) with respect to the topology in A(ω, B).
Proof. The first part follows immediately from Theorem 3.2, and the observation
that
Ω = { e−(1−ε)z2/2 ; 0 < ε < 1, z ∈ Cd }
37
is an admissible family of weights on Cd.
The first part then shows that we may assume that B = L1 and ω = eε0 · 2
for some small ε0 > 0 which depends on λ > 0, when proving the second part.
The result is then an immediate consequence of [47, Prop. 3.2]. The proof is
complete.
(cid:3)
Remark 4.6. By similar arguments, using Proposition 3.1 instead of Theorem 3.2,
it follows that
{ f ∈ H(Rd) ; f · e−(1−ε)x2/2 ∈ B for some ε > 0 }
is independent of the mixed norm space B on Rd. Here recall that H(Rd) is the
set of harmonic functions on Rd.
Our result on duality is the following.
Theorem 4.7. Let ω ∈ P 0
on Cd such that ν2(B) < ∞. Then the following is true:
Q(Cd) be dilated suitable and B be a mixed norm space
(1) the A2 form on P (Cd) extends uniquely to a continuous sesqui-linear form
on A(ω, B) × A(1/ω, B′);
(2) the dual of A(ω, B) can be identified by A(1/ω, B′) through the extension
of the A2 form on P (Cd).
given by (1.20).
Here we recall that B′ = Lp′
(V ), when B = Lp(V ), and p ∈ [1,∞]n and V are
We also have the following result on density, which is strongly connected to the
proof of Theorem 4.7.
Theorem 4.8. Let ω ∈ P 0
Q(Cd) be dilated suitable, B be a mixed quasi-norm
space on Cd such that ν2(B) < ∞, and let F ∈ A(ω, B) and G ∈ A(ω, B′). Then
the following is true:
(1) P (Cd) is dense in A(ω, B). If in addition ν1(B) ≥ 1, then P (Cd) is dense
in A(ω, B′) with respect to the weak∗-topology;
(2) if 0 < λ < 1, then F (λ · ) → F in A(ω, B). If in addition ν1(B) ≥ 1, then
G(λ · ) → G with respect to the weak∗-topology in A(ω, B′) .
We start by proving the first parts of (1) and (2) in Theorem 4.8. Thereafter we
prove Theorem 4.7, and finally we prove the last parts of (1) and (2) in Theorem
4.8.
Some preparations for the proofs are needed. We start by recalling the following
generalization of Lebesgue's theorem.
Lemma 4.9. Let p ∈ (0,∞), dµ be a positive measure, and let fj, f, gj, g ∈ Lp(dµ)
be such that fj → f and gj → g pointwise a. e. as j → ∞, fj ≤ gj, f ≤ g and
kgjkLp(dµ) → kgkLp(dµ) as j → ∞. Then fj → f in Lp(dµ) as j → ∞.
Proof. Let q = max(1, p). Then the result follows by applying Fatou's lemma on
2q(gp + gp
(cid:3)
j ) − f − fjp. The details are left for the reader.
38
The first part of Theorem 4.8 (2) is an immediate consequence of the following
lemma.
Lemma 4.10. Let 0 < ω0 ∈ L∞loc(Rd) be such that
ω0(x) ≤ Cω0(λx),
x ∈ Rd, 1 − θ < λ < 1,
for some constants θ ∈ (0, 1) and C > 0, and let and B be a mixed quasi-norm
space on Rd with ν2(B) < ∞. Also let f ∈ Lr
loc(Rd) be such that f · ω0 ∈ B, where
r = ν1(B). Then f (λ · )ω0 ∈ B when 1 − θ < λ < 1, and
λ→1−k(f − f (λ · ))ω0kB = 0.
(4.12)
lim
Proof. We only consider the case when B = Lp(Rd), p ∈ (0,∞). The straight-
forward modifications to the general case are left for the reader.
By the assumptions we have
f (λx)ω0(x) ≤ Cf (λx)ω0(λx),
and it follows that kf (λ · )ω0kLp ≤ Cλ−d/pkf ω0kLp < ∞, by a simple change of
variables in the integral. Hence {f (λ · )ω0}1−θ<λ<1 is a bounded set in Lp.
By straight-forward approximations, it follows that we may assume that f ∈
C0(Rd) and ω0 ∈ C(Rd) when proving (4.12). Then the result follows if we choose
gλ(x) = f (λx), g(x) = f (x) and dµ(x) = ω0(x) dx in Lemma 4.9. The proof is
complete.
(cid:3)
Proof of the first parts of Theorem 4.8. By Lemma 4.10 it suffices to prove that if
F ∈ A(ω, B) and λ < 1 with 1 − λ small enough, then kF (λ · ) − GkA(ω,B) can be
made arbitrarily small as G ∈ P (Cd).
Let s ∈ R be chosen such that λ < s < 1. Then it follows from the assumptions
that
for some constant C. In particular, we have
C−1ω(z)e−z2/2 ≤ e−sz2/2 ≤ Cω(λz)e−λ2z2/2,
kF (λ · ) − GkA(ω,B) ≤ Ck(F (λ · ) − G)e−s · 2/2kB
and
kF (λ · )e−s· 2/2kB < ∞,
for some constant C > 0.
(4.13)
(4.14)
By (4.14) it follows that F (λ · ) belongs to the set
Ap,t(Cd) ≡ { F ∈ A(Cd) ; kF · e−t · 2/2kB < ∞ },
when t = s. Since the same is true for any choice of t ∈ (λ, s), Ap,t increases with t,
and that { e−t · 2/2 ; λ < t < s } is an admissible family of weights, it follows from
Theorem 3.2 that (4.13) and (4.14) hold after s has been replaced by an appropriate
t ∈ (λ, s), and p = (1, . . . , 1). Since P (Cd) is dense in Ap,t(Cd) with p = (1, . . . , 1),
in view of [47, Proposition 3.2], it follows that the right-hand side of (4.13) can be
made arbitrarily small, and the assertion follows. This completes the proof of the
first parts of Theorem 4.8 (1) and (2).
(cid:3)
39
Next we turn to the proof of Theorem 4.7. First we note that if A(ω, B) is the
same as in Theorem 4.7 and l ∈ (A(ω, B))′, then
l(F ) =Z F (z)G(z) dµ(z),
F ∈ A(ω, B),
(4.15)
for some G ∈ B(1/ω, B′). (Note that the mixed norm space B′ is the dual to B
when ν2(B) < ∞ and the L2 form is used.) In fact, it follows from the definitions
that there is a constant C ≥ 0 such that
l(F ) ≤ CkFkB(ω,B)
(4.16)
when F ∈ A(ω, B). By Hahn-Banach's theorem it follows that l is extendable to
a linear continuous form on B(ω, B) and that (4.16) still holds for F ∈ B(ω, B).
The formula (4.15) now follows from well-known results in measure theory.
From now on we let lG be the continuous linear form on A(ω, B), defined by
(4.15) when G ∈ B(1/ω, B′). Then it follows from the previous investigations that
the map G 7→ lG is surjective from B(1/ω, B′) to (A(ω, B))′.
In what follows we link the kernel of the latter map with the kernel
N(ω, B) ≡ { F ∈ B(ω, B) ; ΠAF = 0 }
of the projection operator. Here we note that every F in B(ω, B) for ω ∈ P 0
and mixed norm space B, fulfill the required properties in Lemma 4.1.
Lemma 4.11. Let B be a mixed norm space on Cd and assume that ω ∈ P 0
is dilated suitable. Then the following is true:
Q(Cd)
Q(Cd)
(1) N(ω, B) is a closed subspace of B(ω, B);
(2) if ν2(B) < ∞, then the kernel of the map G 7→ lG, from B(1/ω, B′) to
(A(ω, B))′ is equal to N(1/ω, B′).
Proof. Assume that Fj ∈ N(ω, B), j ≥ 1, converges to F ∈ B(ω, B) in B(ω, B),
as j → ∞. If K ⊆ Cd is compact, then for some constant CK, depending on K,
Holder's inequality gives
sup
z∈K ΠA(F − Fj) ≤ CKkF − FjkB(ω,B) → 0
z∈K ΠAF = sup
as j → ∞. This proves (1).
(2) By the first part of Theorem 4.8 (1) it suffices to prove that lG(F ) = 0 for
every polynomial F on Cd, if and only if G ∈ N(1/ω, B′). By Lemma 4.1 (4) it
follows that
(4.17)
lG(F ) = lΠAG(F ) = (F, G)A2,
and G ∈ B(1/ω, B′). This proves that lG = 0 when G ∈ N(1/ω, B′).
show that the entire function ΠAG satisfies
On the other hand, if lG(F ) = 0 for every polynomial F , then (4.9) and (4.17)
F ∈ P (Cd),
(∂αΠAG)(0) = 0,
for every α. This implies that ΠAG = 0, i. e. G ∈ N(1/ω, B′), and the proof is
complete.
(cid:3)
40
Remark 4.12. Let ω ∈ P 0
Q(Cd), B be a mixed norm space on Cd, and let A∗(ω, B)
be the completion of P (Cd) under the norm k · kA(ω,B). Then the same arguments
as in the proof of Lemma 4.11 give that the kernel of the map G 7→ lG, from
B(1/ω, B′) to (A∗(ω, B))′ is equal to N(1/ω, B′).
We note that this generalizes Lemma 4.11 (2), since if in addition ν2(B) < ∞,
then A∗(ω, B) = A(ω, B) in view of the first part of Theorem 4.8 (1).
As a consequence of Lemma 4.11 it follows that if ν2(B) < ∞, then the surjective
and continuous map G 7→ lG from B(1/ω, B′) to (A(ω, B))′ induces a homeomor-
phism from the quotent space
C(1/ω, B′) ≡ B(1/ω, B′)/N(1/ω, B′)
to (A(ω, B))′. Here note that Lemma 4.11 implies that C(1/ω, B′) is a Banach
space under the usual quotent topology.
Proof of Theorem 4.7. Let Cw(1/ω, B′) be equal to C(1/ω, B′) equipped by the
induced weak∗-topology on (A(ω, B))′. For each G ∈ B(1/ω, B′) we write G∗ = G
mod N(1/ω, B′) for its image in C(1/ω, B′) under the quotent map. Then it
follows that Cw(1/ω, B′) is a local convex topological vector space, and that the
separating vector space of linear functionals on Cw(1/ω, B′) is equal to
{ Λ ; Λ(G∗) = lG(F ) for some F ∈ A(ω, B) }.
Note here that the equality Λ(G∗) = lG(F ) makes sense since Lemma 4.11 gives
lG1(F ) = lG2(F ) when G1, G2 ∈ B(1/ω, B) are two different representatives of G
mod N(1/ω, B′) and F ∈ A(ω, B).
Since ΠA(F ) = F when F ∈ A(1/ω, B′), it follows that the map G 7→ G
mod N(1/ω, B′) from A(1/ω, B′) to C(1/ω, B′) is continuous and injective. Let
C0(1/ω, B′) be the image of this map. The result follows if we prove that C0(1/ω, B′) =
C(1/ω, B′).
By Hahn-Banach's theorem it suffices to prove that if Λ is a linear and continuous
functional on Cw(1/ω, B′) which is zero on C0(1/ω, B′), then Λ is identically zero.
Since the dual of Cw(1/ω, B′) is equal to A(ω, B) when using the A2 form, we have
Λ(G∗) = ΛF (G∗) ≡ (G, F )A2,
for some F ∈ A(ω, B).
Now ΛF (G∗) = 0 when G ∈ A(1/ω, B′). In particular ΛF (zα) = (zα, F )A2 = 0
for every multi-index α. Since ∂αF (0) = (F, zα)A2 in view of Remark 4.2, it follows
that the entire function F is zero together with all its derivatives at origin. This
implies that F is identically zero, and hence, Λ is zero. This proves the result. (cid:3)
Q(Cd) is dilated
Remark 4.13. By Theorem 4.7 and its proof it follows that if ω ∈ P 0
suitable and B is a mixed norm space such that ν2(B) < ∞, then the mappings
G 7→ lG
and
G 7→ G∗
from A(1/ω, B′) to (A(ω, B))′ and from A(1/ω, B′) to C(1/ω, B′) respectively,
are bijective and continuous. Hence these mappings are homeomorphisms by the
41
open mapping theorem. In particular, the norms in respective space are equivalent,
giving that for some C > 0 it holds
C−1klGk(A(ω,B))′ ≤ kG∗kC(1/ω,B′) ≤ kGkA(1/ω,B′) ≤ CklGk(A(ω,B))′,
when G ∈ A(1/ω, B′).
The end of the proof of Theorem 4.8. We start to prove the second part of (2). Let
F ∈ A(1/ω, B) be fixed, and choose the vector spaces Vj ⊆ Cd and p ∈ [1,∞)
such that B = Lp(V ) when V = (V1, . . . Vn). Then it follows by straight-forward
computations that
(G − G(λ · ), F )A2
=(cid:12)(cid:12)(cid:12)(cid:12)π−dZ (cid:0)G(z)e−z2/2(cid:1)(cid:0)F (z)e−z2/2 − λ−2dF (z/λ)e−(2−λ2)z2/(2λ2)(cid:1)dλ(z)(cid:12)(cid:12)(cid:12)(cid:12)
≤ CkS(F · e− · 2/2 − λ−2dF ( · /λ)e−(2−λ2) · 2/(2λ2))/ωkLp(V ),
where C = C0kGkA(ω,B′) < ∞, for some constant C0 > 0, and S is the operator in
(1.28). By taking 2−1/2z/λ as new variables of integration we get
(G − G(λ · ), F )A2
≤ C1λck(F (λ · )e−λ2 · 2/2 − λ−2dF · e−(1−λ2/2) · 2
for some constants c and C1.
)/ω(S−1(λ · ))kLp(V ),
(4.18)
Now we set
Φλ(z) = (F (λz)e−λ2z2/2 − λ−2dF (z)e−(1−λ2/2)z2
for the last integrand. By (2.5), we get Φλ ≤ Ψλ, where
)/ω(21/2λz)
Ψλ(z) = F (λz)e−λ2z2/2/ω(21/2λz) + Cλ−2dF (z)e−z2/2)/ω(21/2z),
provided the constant C is chosen sufficiently large. Since Φλ → 0 pointwise, and
Ψλ(z) → (C + 1)F (z)e−z2/2/ω(21/2z)
in Lp(V ) as λ → 1− in view of Lemma 4.10, it follows from Lemma 4.9 that Φλ → 0
in Lp(V ) as λ → 1−. This implies that the right-hand side of (4.18) turns to zero
as λ → 1−, and the second part of (2) follows.
It remains to prove the second part of (1). Let G ∈ A(ω, B′) and F ∈ A(1/ω, B).
By the second part of (2) it suffices to prove that for some 0 < λ < 1 and some
polynomials Gj we have
(4.19)
Therefore, let Gj be a sequence of polynomials on Cd. By Proposition 4.5 we
(G(λ2 · ) − Gj, F )A2 → 0
as
j → ∞.
get G(λ · )e−· 2/2 ∈ L1(Cd), and
(G(λ2 · ) − Gj, F )A2 ≤ kG(λ2 · ) − GjkA(ω,B′)kFkA(1/ω,B)
≤ Ck(G(λ · ) − Gj( · /λ))e−· 2/2kL1kFkA(1/ω,B),
42
(4.20)
In the last inequality we have applied Theorem 3.2 with
for some constant C.
B1 = L1 and B2 = B on the estimates
G(λw) − Gj(w/λ)e−w2/(2λ2)ω(21/2w/λ)
≤ CG(λw) − Gj(w/λ)e−w2/(2λ) ≤ CG(λw) − Gj(w/λ)e−w2/2,
for some constant C.
Consequently, if ω0 = 1 and Gj are chosen such that Gj( · /λ) → G(λ · ) in
A(ω0, L1) as j → ∞, then (4.19) follows from (4.20). The proof is complete.
4.2. Narrow convergence. Next we introduce the narow convergence for A(ω, B).
We note that Theorem 4.8 give no explicit possibilities to approximate elements
in A(ω, B) with polynomials when ν1(B) = 1 and ν2(B) = ∞. In this context,
the narrow convergence makes such approximations possible in some of these situ-
ations. The assumptions on the involved weight functions and B is that the pair
(B, ω) should be narrowly feasible (cf. Definition 2.5).
(cid:3)
In order to define the narrow convergence we introduce the functional
JF,ω(ζ2) ≡ sup
ζ1∈V1(cid:16)S(F (z)e−z2/2)ω(z)(cid:17),
z = (ζ1, ζ2) ∈ V1 ⊕ V ⊥1 ≃ Cd,
when F ∈ A(ω, B). Here we recall that S is the dilation operator, given by (1.28).
Definition 4.14. Let (B, ω) be a narrowly feasible space weight pair on Cd, let p
and V be the same as in Definition 2.5, and let q = (p2, . . . , pn) and U = (V2, . . . Vn).
Also let F, Fj ∈ A(ω, B), j ≥ 1. Then Fj is said to converge to F narrowly as
j → ∞, if the following conditions are fulfilled:
(1) S(Fje− · 2/2)ω → S(F e− · 2/2)ω in S′1/2(Cd) as j → ∞;
(2) JFj,ω,p → JF,ω,p in Lq(U) as j → ∞.
The following result gives motivations for introducing the narrow convergence.
Theorem 4.15. Let (B, ω) be a narrowly feasible pair on Cd. Then the following
is true:
(1) P (Cd) is dense in A(ω, B) with respect to the narrow convergence;
(2) if F ∈ A(ω, B) and 0 < λ < 1, then F (λ · ) → F narrowly as λ → 1−.
Proof. We start by proving (2). We may assume that ω = ω0, where ω0 is the same
as in Definition 4.14. Then
S(F (λ · )e−· 2/2)ω → S(F e− · 2/2)ω
pointwise and in S′1/2(Rd) as λ → 1−.
Furthermore, if z = (ζ1, ζ2) ∈ V1 ⊕ V ⊥1 = Cd, then
where
JF (λ · ),ω,p(ζ2) ≤ CλJF,ω,p(λζ2),
Cλ = sup
z∈Cd(cid:16)ω(z)e−(1−λ2)z2/4
ω(λz)
43
(cid:17),
in view of the definition of S. Since Cλ → 1 as λ → 1−, and JF,ω,p(λ · ) → JF,ω,p
in Lq(U) as λ → 1−, in view of Lemma 4.10, it follows from Lemma 4.9 that
JF (λ · ),ω,p → JF,ω,p in Lq(U) as λ → 1−. This proves (2).
It remains to prove (1). Let F ∈ A(ω, B) and 0 < λ < 1. By Cantor's diagonal
principle it suffices to prove that there is a sequence Fj of polynomials which
converges to F (λ · ) narrowly as j → ∞. Since
kJF (λ· ),ω,p − JFj,ω,pkLq(U ) ≤ kF (λ · ) − FjkA(ω,B),
it suffices to prove that Fj → F (λ · ) in A(ω, B). However, this fact is an immediate
consequence of (2.5), Proposition 4.5 and (2). The proof is complete.
Proposition 4.16. Let (B, ω) be a narrowly feasible pair on Cd, Fj, F ∈ A(ω, B),
j = 1, 2, . . . , be such that Fj → F narrowly as j → ∞, and let G ∈ B(1/ω, B′).
Then
(cid:3)
(Fj, G)B2 → (F, G)B2
as
j → ∞.
Proof. We may assume that ω = ω0, where ω0 is the same as in Definition 4.14,
and we let p, q, U and V be the same as in Definition 4.14. It follows from the
assumptions that
Fj(z)G(z)e−z2
= F (z)G(z)e−z2
,
and that
lim
j→∞
S(Fj Ge− · 2
S(F Ge− · 2
(4.21)
(4.22)
)(z) ≤ JFj,ω0,p(ζ2)S(Ge− · 2/2)(z)/ω(z),
)(z) ≤ JF,ω0,p(ζ2)S(Ge− · 2/2)(z)/ω(z).
Furthermore, by Holder's inequality we get
JFj ,ω0,p(ζ2)S(Ge− · 2/2)(z)/ω(z) → JF,ω0,p(ζ2)S(Ge−· 2/2)(z)/ω(z)
in L1(Cd) as j → ∞. A combination of (4.21), (4.22) and Lemma 4.9 now implies
that S(FjGe− · 2) → S(F Ge− · 2) in L1(Cd) as j → ∞, which implies
(Fj, G)B2 =Z Fj(z)G(z) dµ(z) →Z F (z)G(z) dµ(z) = (F, G)B2
as
j → ∞.
The proof is complete.
(cid:3)
4.3. General properties for modulation spaces. We finish the section by using
Theorem 3.4 in order to carry over basic results on A(ω, B) spaces into correspond-
ing result for modulation spaces.
The following result is an immediate consequences of Theorems 3.4, 4.7 and 4.8.
Here and in what follows we let S0(Rd) be the vector space which consists of all
finite linear combinations of Hermite functions.
Theorem 4.17. Let ω ∈ P 0
on R2d such that ν2(B) < ∞. Then the following is true:
Q(Cd) be dilated suitable and B be a mixed norm space
(1) the L2 form on S1/2(Rd) extends uniquely to a continuous sesqui-linear form
on M(ω, B) × M(1/ω, B′);
44
(2) the dual of M(ω, B) can be identified by M(1/ω, B′) through the extension
of the L2 form on S1/2(Rd);
weak∗-topology.
(3) S0(Rd) is dense in M(ω, B), and dense in M(ω, B′) with respect to the
The definition of narrow convergence for elements in certain modulation spaces
is the following. Here the functional which corresponds to JF,ω is
Hf,ω(ζ2) ≡ sup
ζ1∈V1(cid:0)Vφf (x, ξ)ω(x, ξ)(cid:1),
z = (ζ1, ζ2) ∈ V1 ⊕ V ⊥1 = R2d,
Definition 4.18. Let (B, ω) be a narrowly feasible pair on R2d, let p and V be
the same as in Definition 2.5, and let q = (p2, . . . , pn) and U = (V2, . . . Vn). Also
let f, fj ∈ M(ω, B), j ≥ 1. Then fj is said to converge to f narrowly as j → ∞,
if the following conditions are fulfilled:
(1) fj → f in S′1/2(Rd) as j → ∞;
(2) Hfj,ω,p → Hf,ω,p in Lq(U) as j → ∞.
By (1.30) it follows that Hf,ω = (2π)−d/2JVf,ω. Consequently, fj → f narrowly
in M(ω, B) as j → ∞, if and only if Vfj → Vf narrowly in A(ω, B) as j → ∞.
The following result is therefore an immediate consequence of Theorems 3.4, 4.15
and Proposition 4.16.
Theorem 4.19. Let (B, ω) be a narrowly feasible pair on R2d. Also let fj, f ∈
M(ω, B), j = 1, 2, . . . , be such that fj → f narrowly as j → ∞, and let g ∈
M(1/ω, B′). Then the following is true:
(1) S0(Rd) is dense in M(ω, B) with respect to the narrow convergence;
(2) (fj, g)L2 → (f, g)L2 as j → ∞.
In Section 6 we shall use these results to form a pseudo-differential calculus
involving symbols and target distributions in background of the modulation space
theory presented here.
5. Some consequences
In this section we show some consequences of the results in the previous sections.
We start by establishing continuity properties of ΠA on appropriate B(ω, B) spaces.
From these results it follows that A(ω, B) is a retract of B(ω, B) under ΠA. There-
after we use this property for establishing interpolation properties for A(ω, B), and
continuity properties of Toeplitz operators.
5.1. Continuity properties of ΠA on B(ω, B). We start with the following
result related to Lemma 4.1 (3).
Q(R2d)
Theorem 5.1. Let B be a mixed norm space on R2d and assume that ω ∈ P 0
is dilated suitable. Then the map ΠA is continuous from B(ω, B) to A(ω, B). In
particular, A(ω, B) is a retract of B(ω, B) under ΠA.
45
Proof. By Remark 4.12 it follows that the mappings
from A(ω, B) and B(ω, B)/N(ω, B) respectively to
F 7→ LF
and F mod N(ω, B)
{ lF ; F ∈ B(ω, B) } ⊆ (A∗(1/ω, B′))′
are well-defined, continuous and bijective. By the open mapping theorem it follows
that the inverse of the map F 7→ F mod N(ω, B) is continuous and bijective, and
that
kF0kA(ω,B) ≤ C inf
G∈N (ω,B)kF + GkB(ω,B),
F − F0 ∈ N(ω, B), F0 ∈ A(ω, B),
(5.1)
for some constant C.
Now let F ∈ B(ω, B), set F1 = ΠAF , and choose F0 ∈ A(ω, B) such that
lF = lF0. Then lF0 = lF1 on P (Cd), and Lemma 4.1 and Remark 4.2 shows that
∂αF0(0) = ∂αF1(0) for every multi-index α. Consequently, F0(z) = F1(z), since
both F0 and F1 are entire. Furthermore, F − F0 ∈ N(ω, B), since the restriction
of ΠA on A(ω, B) is the identity map.
By (5.1) we now get
kΠAFkA(ω,B) = kF0kA(ω,B) ≤ C inf
G∈N (ω,B) kF + GkB(ω,B) ≤ CkFkB(ω,B),
and the assertion follows.
(cid:3)
We recall that (real and complex) interpolation properties carry over under re-
tracts. These properties also include interpolation techniques with more than two
spaces involved. (Cf. [1, 2, 8].) In particular, the following result is an immedi-
ate consequence of Theorem 5.1. Here recall that (B1, B2)[θ] denotes the complex
interpolation space with respect to θ ∈ [0, 1], when (B1, B2) is a compatible pair.
Proposition 5.2. Any interpolation property valid for the B(ω, B) spaces also
holds for the A(ω, B) spaces, when the weights ω are dilated suitable and belong to
Q(Cd), and B are mixed norm spaces on Cd. In particular, if 0 ≤ θ ≤ 1, (ω, B1)
P 0
and (ω, B2) are feasible pairs on Cd, and that B = (B1, B2)[θ], for θ ∈ [0, 1], then
(B(ω, B1), B(ω, B2))[θ] = B(ω, B) and (A(ω, B1), A(ω, B2))[θ] = A(ω, B).
Next we discuss further density properties of A(ω, B). We recall that A∗(ω, B)
denotes the completion of P (Cd) under the norm k · kB(ω,B), when ω ∈ PQ(Cd)
and B is a mixed norm space on Cd. We also let B∗(ω, B) be the completion of
C∞0 (Cd) under the norm k · kB(ω,B). The following result links the A∗(ω, B) with
B∗(ω, B).
Proposition 5.3. Let ω ∈ P 0
space on Cd. Then
Q(Cd) be dilated suitable, and let B be a mixed norm
A(Cd)\ B∗(ω, B) = A(ω, B)\ B∗(ω, B) = A∗(ω, B).
46
Proof. Since A(ω, B) ⊆ A(Cd), the result follows if we prove
A∗(ω, B) ⊆ A(ω, B)\ B∗(ω, B) and A(Cd)\ B∗(ω, B) ⊆ A∗(ω, B).
(5.2)
First we prove the first inclusion in (5.2). Since A∗(ω, B) ⊆ A(ω, B), it suffices to
prove that A∗(ω, B) ⊆ B∗(ω, B).
Let F ∈ A∗(ω, B). Then there is a sequence Fj ∈ P (Cd) such that kF −
FjkA(ω,B) → 0 as j → ∞. Hence (5.2) follows if we prove that for each G ∈ P (Cd),
there are elements Gj ∈ C∞0 (Cd) such that kG − GjkB(ω,B) → 0 as j → ∞.
Let ϕj ∈ C∞0 (Cd) be chosen such that 0 ≤ ϕj ≤ 1 and ϕj(z) = 1 when z ≤ j,
and let Gj = ϕjG. By Holder's inequality, there is a constant C such that
kG − GjkB(ω,B) ≤ C sup
z∈Cd(cid:0)G(z) − Gj(z)e−z2/4(cid:1) ≤ C(cid:0) sup
z≥j G(z)e−z2/4(cid:1) → 0
In order to prove the second inclusion we instead assume that F ∈ A(Cd) ∩
as j → ∞. This gives the first inclusion in (5.2).
B∗(ω, B), and we let Fj ∈ C∞0 (Cd) be a sequence such that
j → ∞.
kF − FjkB(ω,B) → 0
(5.3)
By straight-forward computations it follows that ΠAFj(z) ≤ CjeCjz, for some
constants Cj > 0, which implies that ΠAFj ∈ AP (Cd) ⊆ A∗(ω, B). Hence Theorem
5.1 and (5.3) give
kF − ΠAFjkB(ω,B) = kΠA(F − Fj)kB(ω,B) ≤ CkF − FjkB(ω,B) → 0 as
j → ∞,
for some constant C. Since any element in AP (Cd) can be approximated by ele-
ments in P (Cd) with respect to the norm A(ω, B), in view of Proposition 4.5, the
result follows. The proof is complete.
(cid:3)
as
We may now extend (4.4). In fact, we have the following.
Proposition 5.4. Let (B, ω) be a feasible pair on Cd, and let F ∈ A(ω, B) and
G ∈ B(1/ω, B′). Then (4.4) holds.
Proof. First assume that ν2(B) < ∞. Then A(ω, B) = A∗(ω, B). Let Fj ∈ P (Cd)
be such that Fj → F in A(ω, B) as j → ∞. Since ΠAF = F and ΠAFj = Fj, it
follows from Lemma 4.1 and Theorem 5.1 that
(ΠAF, G)B2 = (F, G)B2 = lim
j→∞
(Fj, G)B2 = lim
j→∞
(Fj, ΠAG)B2 = (F, ΠAG)B2,
(5.4)
and the result follows in this case.
Next we consider the case when ν1(B) > 1. Then B(1/ω, B′) = B∗(1/ω, B′),
and the result follows by similar arguments after approximating G with elements
in C∞0 and using the fact that
(ΠAG)(z) ≤ CeCz
for some C > 0,
when G ∈ C∞0 which is needed when applying Lemma 4.1 (cf. (4.3)).
Finally, if (B, ω) is narrowly feasible, then we choose a sequence Fj ∈ P (Cd)
which converges to F narrowly as j → ∞. By Lemma 4.1, Proposition 4.16 and
Theorem 5.1, it follows that (5.4) holds also in this case. The proof is complete. (cid:3)
47
The next result concerns convenient equivalent norms on A(ω, B).
Proposition 5.5. Let (B, ω) be a feasible pair on Cd, and let
F ∈ A(ω, B),
kFk ≡ sup (F, G)A2,
where the supremum is taken over all G ∈ A(1/ω, B′) such that kGkA(1/ω,B′) ≤ 1.
Then k · k is a norm on A(ω, B) which is equivalent to k · kA(ω,B).
Proof. By Holder's inequality we get kFk ≤ kFkA(ω,B). We have to prove that
kFkA(ω,B) ≤ CkFk,
F ∈ A(ω, B),
for some constant C which is independent of F ∈ A(ω, B).
Let ε > 0 be fixed, and let Ω be the set of all G ∈ B(1/ω, B′) such that
kGkB(1/ω,B′) ≤ 1. Then the converse of Holder's inequality and Proposition 5.4
give
(5.5)
kFkA(ω,B) = sup
G∈Ω (F, G)B2 ≤ (F, G0)B2 + ε = (F, ΠAG0)B2 + ε,
for some choice of G0 ∈ Ω. Since ΠAG0 ∈ A(1/ω, B′) and kΠAG0kA(1/ω,B′) ≤
CkG0kB(1/ω,B′) for some constant C, by Theorem 5.1, we obtain
kFkA(ω,B) ≤ (F, G0)B2 + ε = (F, ΠAG0)B2 + ε ≤ C sup (F, G)B2 + ε,
where the supremum is taken over all G ∈ Ω ∩ A(1/ω, B′). Since ε > 0 was
arbitrary chosen, (5.5) follows. The proof is complete.
(cid:3)
5.2. Consequences for modulation spaces and Toeplitz operators. Next
we use Theorem 3.4 to carry over the previous results for A(ω, B) spaces into
modulation spaces. First we need to investigate how ΠA is linked to the composition
Vφ ◦ V ∗φ , where V ∗φ is the (Hilbert-) adjoint of Vφ. Here we let V ∗φ F be the unique
element in S′1/2(Rd) which satisfies
(V ∗φ F, g)L2(Rd) = (F, Vφg)L2(R2d),
g ∈ S0(Rd),
when F ∈ S′1/2(R2d). We also let ΠM be the continuous operator on S′1/2(R2d)
given by ΠM = Vφ ◦ V ∗φ , and note that ΠM is a projection from S′1/2(R2d) onto
Vφ(S′1/2(Rd)).
Lemma 5.6. Let ω ∈ P 0
Q(Cd), and let B be a mixed norm space on R2d. Then
ΠA = UV ◦ ΠM ◦ U−1
V ,
on B(ω, B), where UV is given by (1.31).
Proof. Let F ∈ B(ω, B), F0 = U−1
Then ΠM G0 = G0, and
V F , g ∈ S0(Rd), G0 = Vφg and G = UVG0.
(F, G)B2 = (UVF0, UVG0)B2 = (F0, G0)L2 = (F0, Vφg)L2
= (V ∗φ F0, g)L2 = (ΠM F0, Vφg)L2 = (ΠM F0, G0)L2
= (UV(ΠM F0), G)B2 = ((UV ◦ ΠM ◦ U−1
V )F, G)B2,
48
where UV(ΠM F0) = V(V ∗φ F0) is analytic and satisfies (4.2). Furthermore, (F, G)B2 =
(ΠAF, G)B2 in view of Lemma 4.1 (4). A combination of these equalities now gives
Since also ΠAF is analytic it follows from Remark 4.3 that
((UV ◦ ΠM ◦ U−1
V )F, G)B2 = (ΠAF, G)B2.
and the result follows.
(UV ◦ ΠM ◦ U−1
V )F = ΠAF,
(cid:3)
The following result is now an immediate consequence of Theorems 3.4 and 5.1,
and Lemma 5.6. Here and in what follows we let B(ω) be the set of all f ∈ L1(Rd)
such that kfkB(ω) ≡ kf ωkB is finite, when B is a mixed norm space on Rd and
ω ∈ PQ(Rd). In this situation we also set B′(ω) = (B′)(ω).
Theorem 5.7. Let B be a mixed norm space on Cd and assume that ω ∈ P 0
is dilated suitable. Then the following is true:
Q(Cd)
(1) ΠM is continuous from B(ω) to Vφ(M(ω, B));
(2) V ∗φ is continuous from B(ω) to M(ω, B).
In particular, Vφ(M(ω, B)) is a retract of BB(ω) under ΠM .
The next results are immediate consequences of Theorem 3.4, Propositions 5.2,
5.4 and 5.5, and Lemma 5.6.
Proposition 5.8. Any interpolation property valid for the B(ω, B) spaces also
hold for the M(ω, B) spaces, when the weights ω are dilated suitable and belong to
Q(R2d), and B are mixed norm space on R2d. In particular, if 0 ≤ θ ≤ 1, (ω, B1)
P 0
and (ω, B2) are feasible pairs on R2d, and that B = (B1, B2)[θ], for θ ∈ [0, 1], then
(M(ω, B1), M(ω, B2))[θ] = M(ω, B).
Proposition 5.9. Let (B, ω) be a feasible pair on R2d, and let f ∈ M(ω, B) and
G ∈ B′(1/ω). Then
(Vφf, G)L2(R2d) = (f, V ∗φ G)L2(Rd).
Proposition 5.10. Let (B, ω) be a feasible pair on R2d, and let
kfk ≡ sup (f, g)L2,
f ∈ M(ω, B),
where the supremum is taken over all g ∈ M(1/ω, B′) such that kgkM (1/ω,B′) ≤ 1.
Then k · k is a norm on M(ω, B) which is equivalent to k · kM (ω,B).
Next we consider Toeplitz operators and Berezin-Toeplitz operators. Let a ∈
S1/2(R2d) be fixed. Then the Toeplitz operator Tp(a) is the linear and continuous
operator on S1/2(Rd), defined by the formula
(Tp(a)f, g)L2(Rd) = (a Vφf, Vφg)L2(R2d).
(5.6)
There are several characterizations and several ways to extend the definition of
Toeplitz and Berezin-Toeplitz operators (see e. g. [5, 7, 13, 14, 17, 23, 31, 32, 47, 54,
58] and the references therein). For example, the definition of Tp(a) is uniquely
extendable to every a ∈ S′1/2(R2d), and then Tp(a) is continuous on S1/2(Rd).
49
Furthermore, it follows from (5.6) that Tp(a) is uniquely extendable to a continuous
operator from Υ(Rd) to S′1/2(Rd).
Toeplitz operators arise in pseudo-differential calculus in [23,39], in the theory of
quantization (Berezin quantization) in [7]), and in signal processing in [17] (under
the name of time-frequency localization operators or STFT multipliers). There
are also strong connection between such operators and Berezin-Toeplitz operators
which we shall consider now.
Let B1, B2 be mixed norm spaces on R2d ≃ Cd, ω1, ω2 ∈ P 0
Q(Cd),
a ∈ B1(ω1), and S be as in (1.28). Then the Berezin-Toeplitz operator TV(a) is
the operator from A(ω2, B2) to A(Cd), given by the formula
Q(R2d) ≃ P 0
TV(a)F = ΠA((S−1a)F ),
F ∈ A(ω2, B2).
It follows from (5.6) that if a ∈ S′1/2(R2d) and f ∈ S′1/2(Rd), then
(Vφ ◦ Tpφ(a))f = ΠM (a · F0), where F0 = Vφf.
Hence
(5.7)
(5.8)
(5.9)
TV(a) ◦ V = V ◦ Tp(a),
for appropriate a ∈ S′1/2(R2d), by Lemma 5.6.
We have now the following result. Here we recall that if pj = (pj,1, . . . , pj,n) ∈
[1,∞]n, j = 0, 1, 2, then 1/p1 + 1/p2 = 1/p0 means that 1/p1,k + 1/p2,k = 1/p0,k for
every k = 1, . . . , n.
Proposition 5.11. Let Bj = Lpj (V ), j = 0, 1, 2, be mixed norm space such that
1/p1 + 1/p2 = 1/p0, and let ωj ∈ P 0
D(Cd) for j = 0, 1, 2. Also let
a ∈ B1(ω1). Then the following is true:
D(R2d) ≃ P 0
(1) the Toeplitz operator Tp(a) is continuous from M(ω2, B2) to M(ω0, B0);
(2) the Berezin-Toeplitz operator TV(a) is continuous from A(ω2, B2) to A(ω0, B0).
Proof. The assetion (2) follows immediately from the definitions and Holder's in-
equality, and (1) is then an immediate consequence of (2) and (5.9). The proof is
complete.
(cid:3)
6. Pseudo-differential operators
In this section we state results on pseudo-differential operators in background
of the modulation space theory from the previous sections. The proofs are in
general omitted, since the results follows by the same arguments as in [57 -- 59] in
combination with the results in previous sections.
6.1. General properties of pseudo-differential operators. We start with the
definition of pseudo-differential operators. Let t ∈ R be fixed and let a ∈ S1/2(R2d).
Then the pseudo-differential operator Opt(a) with symbol a is the linear and con-
tinuous operator on S1/2(Rd), defined by the formula
(Opt(a)f )(x) = (2π)−dZZ a((1 − t)x + ty, ξ)f (y)eihx−y,ξi dydξ.
50
(6.1)
The definition of Opt(a) extends to each a ∈ S′1/2(R2d), and then Opt(a) is contin-
uous from S1/2(Rd) to S′1/2(Rd). (Cf. e. g. [15], and to some extent in [37].) More
precisely, for any a ∈ S′1/2(R2d), the operator Opt(a) is defined as the linear and
continuous operator from S1/2(Rd) to S′1/2(Rd) with distribution kernel given by
(6.2)
Here F2F is the partial Fourier transform of F (x, y) ∈ S′1/2(R2d) with respect to
the y variable. This definition makes sense, since the mappings F2 and F (x, y) 7→
F ((1 − t)x + ty, y − x) are homeomorphisms on S′1/2(R2d).
On the other hand, let T be an arbitrary linear and continuous operator from
S(Rd) to S′(Rd). Then it follows from Theorem 2.2 in [40] that for some K =
KT ∈ S′1/2(R2d) we have
2 a)((1 − t)x + ty, x − y).
Ka,t(x, y) = (F −1
(T f, g)L2(Rd) = (K, g ⊗ f )L2(R2d),
for every f, g ∈ S(Rd). Now by letting a be defined by (6.2) after replacing Ka,t
with K it follows that T = Opt(a). Consequently, the map a 7→ Opt(a) is bijective
from S′1/2(R2d) to L (S(Rd),S′(Rd)).
If t = 1/2, then Opt(a) is equal to the Weyl quantization Opw(a) of a. If instead
t = 0, then the standard (Kohn-Nirenberg) representation a(x, D) is obtained.
In particular, if a ∈ S′1/2(R2d) and s, t ∈ R, then there is a unique b ∈ S′1/2(R2d)
such that Ops(a) = Opt(b). By straight-forward applications of Fourier's inversion
formula, it follows that
Ops(a) = Opt(b) ⇐⇒ b(x, ξ) = ei(t−s)hDx,Dξia(x, ξ).
(6.3)
(Cf. Section 18.5 in [37].) Note here that the right-hand side makes sense, be-
cause ei(t−s)hDx,Dξi on the Fourier transform side is a multiplication by the function
ei(t−s)hx,ξi, which is a continuous operation on S′1/2(R2d), in view of the definitions.
Let t ∈ R and a ∈ S′1/2(R2d) be fixed. Then a is called a rank-one element with
respect to t, if the corresponding pseudo-differential operator is of rank-one, i. e.
(6.4)
for some f1, f2 ∈ S′1/2(Rd). Here f ∈ S1/2(Rd). By straight-forward computations
it follows that (6.4) is fulfilled, if and only if a = (2π)d/2W t
t-Wigner distribution, defined by the formula
f1,f2, where the W t
Opt(a)f = (f, f2)L2f1,
f1,f2
W t
f1,f2(x, ξ) ≡ F (f1(x + t· )f2(x − (1 − t) · ))(ξ),
which takes the form
W t
f1,f2(x, ξ) = (2π)−d/2Z f1(x + ty)f2(x − (1 − t)y)e−ihy,ξi dy,
when f1, f2 ∈ S1/2(Rd). By combining these facts with (6.3), it follows that
W t
f1,f2 = ei(t−s)hDx,DξiW s
f1,f2,
51
for each f1, f2 ∈ S′1/2(Rd) and s, t ∈ R. Since the Weyl case is important to us,
we set W t
f1,f2 = Wf1,f2 when t = 1/2. Then Wf1,f2 is the usual (cross-) Wigner
distribution of f1 and f2.
Next we discuss the Weyl product, twisted convolution and related objects. Let
a, b ∈ S′1/2(R2d) be appropriate. Then the Weyl product a#b between a and b is
the function or distribution which fulfills Opw(a#b) = Opw(a) ◦ Opw(b), provided
the right-hand side makes sense. More generally, if t ∈ R, then the product #t is
defined by the formula
Opt(a#tb) = Opt(a) ◦ Opt(b),
(6.5)
provided the right-hand side makes sense as a continuous operator from S1/2(Rd)
to S′1/2(Rd).
The Weyl product can also, in a convenient way, be expressed in terms of the
symplectic Fourier transform and twisted convolution. More precisely, the sym-
plectic Fourier transform for a ∈ S1/2(R2d) is defined by the formula
(Fσa)(X) = π−dZ a(Y )e2iσ(X,Y ) dY.
Here σ is the symplectic form, which is defined by
σ(X, Y ) = hy, ξi − hx, ηi,
X = (x, ξ) ∈ R2d, Y = (y, η) ∈ R2d,
where h·,·i denotes the usual scalar product on Rd, as before.
omorphism on S′1/2(R2d), and to a unitary map on L2(R2d). Furthermore, F 2
the identity operator.
It follows that Fσ is continuous on S1/2(R2d), and extends as usual to a home-
σ is
Let a, b ∈ S1/2(R2d). Then the twisted convolution of a and b is defined by the
formula
(a ∗σ b)(X) = (2/π)d/2Z a(X − Y )b(Y )e2iσ(X,Y ) dY.
The definition of ∗σ extends in different ways. For example, it extends to a con-
tinuous multiplication on Lp
(v)(R2d) when p ∈ [1, 2] when v ∈ PE(R2d) is sub-
multiplicative (cf. [59]), and to a continuous map from S′1/2(R2d) × S1/2(R2d) to
S′1/2(R2d)∩ C∞(R2d). If a, b ∈ S′1/2(R2d), then a#b makes sense if and only if a∗σbb
makes sense, and then
(6.6)
a#b = (2π)−d/2a ∗σ (Fσb).
We also remark that for the twisted convolution we have
Fσ(a ∗σ b) = (Fσa) ∗σ b = a ∗σ (Fσb),
(6.7)
where a(X) = a(−X) (cf. [51, 53]). A combination of (6.6) and (6.7) gives
Fσ(a#b) = (2π)−d/2(Fσa) ∗σ (Fσb).
52
For admissible a, b, c ∈ S′1/2(R2d), it follows by straight-forward computations
that
(a1 ∗σ a2, b) =(a1, b ∗σea2) = (a2,ea1 ∗σ b),
(a1#a2, b) = (a1, b#a2) = (a2, a1#b),
(a1 ∗σ a2) ∗σ b = a1 ∗σ (a2 ∗σ b)
(a1#a2)#b = a1#(a2#b).
6.2. Pseudo-differential operators and modulation spaces. Next we con-
sider questions on Weyl quantizations of pseudo-differential operators in the con-
text of modulation space theory. It is then convenient to use the symplectic Fourier
transform and the symplectic short-time Fourier transform, instead of correspond-
ing "ordinary" transformations. Here the symplectic short-time Fourier transform
of a ∈ S′1/2(R2d) with respect to the window function Ψ ∈ S′1/2(R2d) is defined by
VΨa(X, Y ) = Fσ(cid:0)a Ψ( · − X)(cid:1)(Y ), X, Y ∈ R2d.
Let (B, ω) be an admissible pair on R4d. Then M(ω, B) denotes the modulation
spaces of Gelfand-Shilov distributions on R2d, where the symplectic short-time
Fourier transform is used instead of the usual short-time Fourier transform and the
window function Ψ(X) here above is equal to
Φ(X) = (2/π)d/2e−X2
,
in the definitions of the norms. In a way similar as for the usual modulation spaces,
we set
Mp,q
(ω)(R2d) = M(ω, Lp,q(R4d)),
when p, q ∈ [1,∞]. It follows that any property valid for the modulation spaces in
previous sections carry over to spaces of the form M(ω, B).
The choice of window function is motivated by the simple form that the fol-
lowing results attain. For the proof we refer to the results in previous sections in
combination with the proof of Proposition 4.1 in [57].
Proposition 6.1. Let pj, qj, p, q ∈ [1,∞] such that p ≤ pj, qj ≤ q, for j = 1, 2,
and that
Also let ω1, ω2 ∈ P 0
and (Lp,q(R4d), ω) are admissible pairs and satisfy
1/p1 + 1/p2 = 1/q1 + 1/q2 = 1/p + 1/q.
Q(R2d) and ω ∈ P 0
Q(R4d) be such that (Lpj,qj (R2d), ωj), j01, 2,
ω(X, Y ) ≤ Cω1(X − Y )ω2(X + Y ).
Then the map (f1, f2) 7→ Wf1,f2 from S′1/2(Rd) × S′1/2(Rd) to S′1/2(R2d) restricts to
a continuous mapping from M p1,q1
kWf1,f2kMp,q
(ω1) (Rd) × M p2,q2
(ω) ≤ Ckf1kM
(ω2) (Rd) to Mp,q
(ω)(R2d), and
(ω1) kf2kM
,
(ωj ) (Rd), j = 1, 2.
where the constant C is independent of fj ∈ M pj,qj
p2,q2
(ω2)
p1,q1
We now arrive on the following continuity result for pseudo-differential opera-
tors. Again we omit the proof, since the result is a straight-forward consequence
of Proposition 4.12 in [57] and its proof in combinations with the results on mod-
ulation spaces in previous sections.
53
Theorem 6.2. Let p, q, pj, qj ∈ [1,∞] for j = 1, 2, be such that
=
1
q1
1
p1
1
q2 −
1
1
p2 −
p
Q(R4d) and ω1, ω2 ∈ P 0
=
+
Also let ω ∈ P 0
(Lp,q(R4d), ω) are admissible pairs and satisfy
1
q − 1,
q ≤ p2, q2 ≤ p.
(6.8)
Q(R2d) be such that (Lpj ,qj (R2d), ωj) and
ω2(X − Y )
ω1(X + Y ) ≤ Cω(X, Y ),
If a ∈ Mp,q
(6.9)
for some constant C > 0.
S′1/2(Rd) extends uniquely to a continuous map from M p1,q1
k · kMp,q
, then
(ω)
(ω)(R2d), then Opw(a) from S1/2(Rd) to
(ω1) (Rd) to M p2,q2
(ω2) (Rd).
Moreover, if in addition a belongs to the closure of S0(R2d) under the norm
Opw(a) : M p1,q1
(ω1) (Rd) → M p2,q2
(ω2) (Rd)
is compact.
Next we consider algebraic properties of modulation spaces under twisted con-
volution and Weyl product. These investigations are based on the following lemma
together with the observations
Φ#Φ = π−dΦ and Φ ∗σ Φ = 2dΦ
(cf. [53]). We refer to [52, 59] for its proof.
Lemma 6.3. Assume that a1 ∈ Υ(R2d), a2 ∈ S1/2(R2d), Φ(X) = (2/π)d/2e−X2
and X, Y ∈ R2d. Then the following is true:
(1) the map
Z 7→ e2iσ(Z,Y )(VΦa1)(X − Y + Z, Z) (VΦa2)(X + Z, Y − Z)
belongs to L1(R2d), and
VΦ(a1#a2)(X, Y )
=Z e2iσ(Z,Y )(VΦa1)(X − Y + Z, Z) (VΦa2)(X + Z, Y − Z) dZ;
(2) the map
Z 7→ e2iσ(X,Z−Y )(VΦa1)(X − Y + Z, Z) (VΦa2)(Y − Z, X + Z)
belongs to L1(R2d), and
VΦ(a1 ∗σ a2)(X, Y )
=Z e2iσ(X,Z−Y )(VΦa1)(X − Y + Z, Z) (VΦa2)(Y − Z, X + Z) dZ.
54
The following two theorems now follows by combining the results in previous
sections and using similar arguments as in the proofs of Theorem 0.3′ in [36] and
Theorem 2.2 in [59]. The details are left for the reader. Here and in what follows,
the involved Lebesgue exponents should satisfy conditions of the form
1
p1
1
p1
+
+
1
p2 −
1
p2 −
0 ≤
and
1
p0
= 1 −(cid:16) 1
q1
1
p0 ≤
1
pj
,
1
qj ≤
1
q1
+
1
q0(cid:17),
1
q2 −
1
q2 −
+
1
q0
(6.10)
,
j = 0, 1, 2,
(6.11)
1
q1
+
1
q2 −
1
q0 ≤
1
pj
,
1
qj ≤
1
p1
+
1
p2 −
1
p0
,
0 ≤
j = 0, 1, 2.
(6.12)
Furthermore, the involved weight functions should satisfy
ω0(X, Y ) ≤ Cω1(X − Y + Z, Z)ω2(X + Z, Y − Z), X, Y, Z ∈ R2d,
(6.13)
or
ω0(X, Y ) ≤ Cω1(X − Y + Z, Z)ω2(Y − Z, X + Z), X, Y, Z ∈ R2d.
(6.14)
Q(R4d) and pj, qj ∈ [1,∞] be such that (Lpj,qj (R4d), ωj)
Theorem 6.4. Let ωj ∈ P 0
are admissible pairs for j = 0, 1, 2, and (6.10), (6.11) and (6.13) hold. Then the
map (a1, a2) 7→ a1#a2 on S0(R2d) extends uniquely to a continuous map from
Mp1,q1
(ω2) (R2d) to Mp0,q0
(ω1) (R2d) × Mp2,q2
(ω0) (R2d), and
(ω0) ≤ Cka1kM
p0,q0
ka1#a2kM
,
p1,q1
p2,q2
(ω2)
(ω1) ka2kM
(ωj) (R2d), j = 1, 2.
where the constant C is independent of aj ∈ Mpj,qj
Q(R4d) and pj, qj ∈ [1,∞] be such that (Lpj,qj (R4d), ωj)
Theorem 6.5. Let ωj ∈ P 0
are admissible pairs for j = 0, 1, 2, and (6.10), (6.12) and (6.14) hold. Then the
map (a1, a2) 7→ a1 ∗σ a2 on S0(R2d) extends uniquely to a continuous map from
W p1,q1
(ω2) (R2d) to W p0,q0
(ω1) (R2d) × W p2,q2
(ω0) (R2d), and
(ω0) ≤ Cka1kW
p0,q0
ka1 ∗σ a2kW
where the constant C is independent of aj ∈ W pj,qj
Remark 6.6. We note that ωj, j = 0, 1, 2, fulfills all the required properties in
Theorem 6.4, if
ω0(X, Y ) =
ω0(X, Y ) =
ν3(X − Y )
ν1(X + Y )
ν3(X − Y )
ν2(X + Y )
,
.
ω1(X, Y ) =
ν2(X − Y )
ν1(X + Y )
,
(6.15)
for some appropriate ν1, ν2, ν3 ∈ P 0
appears for ω in Theorem 6.2.
Q(R2d). Note here that such types of conditions
55
,
p1,q1
p2,q2
(ω2)
(ω1) ka2kW
(ωj ) (R2d), j = 1, 2.
6.3. Examples on calculi of pseudo-differential operators. Next we give
some examples on symbol classes and continuity properties for corresponding pseudo-
differential operators. We are especially focused on symbol classes of the form
M∞,1
(ω) (R2d), because they are to some extend linked to certain classical symbol
classes in the pseudo-differential calculus. We have for example that
S(ω)(R2d) = \N≥0
M∞,1
(1/ωN )(R2d),
(6.16)
where ω ∈ P(R2d) and ωN (X, Y ) = ω(X)hY i−N . Here S(ω)(R2d) is the symbol
class which consists of all a ∈ C∞(R2d) such that (∂αa)/ω ∈ L∞(R2d) for every
multi-index α. (Cf. [36, Rem. 2.18].)
We also remark that (6.16) can be used to carry over properties valid for modula-
tion spaces into S(ω) spaces. For example, in [36, Rem. 2.18] it is proved that The-
orem 6.4 and (6.16) imply S(ω1)#S(ω2) ⊆ S(ω1ω1) when ω1, ω2 ∈ P. (See [37, Sec.
18.5] for an alternative proof of the latter fact.)
As a consequence of Theorems 6.2 and 6.4, and Remark 6.6 we have the following.
Proposition 6.7. Let p, q ∈ [1,∞], ν1, ν2, ν3 ∈ P 0
be such that (6.15) is fulfilled. Then the following is true:
D(R2d) and ω0, ω1, ω2 ∈ P 0
D(R4d)
(1) if aj ∈ M∞,1
Opw(a1) : M p,q
(ωj)(R2d), then the mappings
(ν1)(Rd) → M p,q
(ν2)(Rd), Opw(a2) : M p,q
(ν2)(Rd) → M p,q
(ν3)(Rd)
are continuous;
(2) the map (a1, a2) 7→ a1#a2 is continuous from M∞,1
(ω1)(R2d) × M∞,1
(ω2)(R2d) to
M∞,1
(ω0)(R2d).
Corollary 6.8. Let p, q ∈ [1,∞], ν ∈ P 0
ω(X, Y ) =
D(R2d) and
ν(X − Y )
ν(X + Y )
.
Then the following is true:
(6.17)
(ω) (R2d), then Opw(a) is continuous on M p,q
(ν) (Rd);
(1) if a ∈ M∞,1
(2) (M∞,1
(ω) (R2d), #) is an algebra.
Proof. The result follows by letting ν1 = ν2 = ν3 = ν and ω0 = ω1 = ω2 = ω in
Proposition 6.7.
(cid:3)
Example 6.9. Let ν(X) = ecXγ , for 0 ≤ γ < 2 and some constant c ∈ R. In this
case, ω in (6.17) is given by
ω(X, Y ) = ec(X−Y γ−X+Y γ ).
In the case γ ≤ 1 one may use the inequality ω(X, Y ) ≤ e2c·Y γ to conclude
(ω0) and ω0(X, Y ) = e2c·Y γ .
that Opw(a) is continuous on M p,q
(ν) (Rd), when a ∈ M∞,1
56
More generally, let νj(X) = ecjXγ , for 0 ≤ γ < 2 and some constants cj ∈ R,
j = 1, 2. In this case, ω1 in (6.15) is given by
ω1(X, Y ) = ec2X−Y γ−c1X+Y γ
.
In the case γ ≥ 1, c1 = 2γ−1 and c2 = 1 we have
X − Y γ − 2γ−1X + Y ≤ (X + Y + 2Y )γ − 2γ−1X + Y ≤ 22γ−1Y γ,
and ω(X, Y ) ≤ e22γ−1Y γ . Hence, if
,
ω0(X, Y ) = e22γ−1Y γ
ν1(X) = e2γ−1Xγ
,
(ω0), then Opw(a) is continuous from M p,q
and a ∈ M∞,1
Example 6.10. Let νj(X) = hXicjhXi or νj(X) = ΓhXicj for some constant cj ∈
R, j = 1, 2. In this case, ω1 in (6.15) is given by
(ν1) to M p,q
(ν2);
ν2(X) = eXγ
,
or
Hence, if a ∈ M∞,1
ω1(X, Y ) = hX − Y ic2hX−Y ihX + Y i−c1hX+Y i
ω1(X, Y ) =(cid:0)Γ(hX − Y i)(cid:1)c2(cid:0)Γ(hX + Y i)(cid:1)−c1.
(ω1), then Opw(a) is continuous from M p,q
(ν1) to M p,q
(ν2).
We note that different situations appear depending on the sign on c1 and c2:
(1) if c1 > 0 and c2 > 0, then the weights νj(X), j = 1, 2, turn rapidly to
infinity at infinity. This implies that the target space M p,q
(ν1) as well as the
image space M p,q
(ν2) are small, in the sense that their elements turns rapidly
to zero at infinity, fulfill hard restrictions on oscillations at infinity, and are
extendable to entire functions on Cd.
The corresponding weight ω1 turns rapidly to zero as X = Y and X →
∞, while ω1 turns rapidly to infinity as X = −Y and X → ∞.
(2) if c1 < 0 and c2 < 0 (i. e. the adjoint situation comparing to (1)), then the
target space M p,q
(ν2) are large, in the sense
that their elements are allowed to turns rapidly to infinity at infinity, with
small restrictions on oscillations and singularities at infinity.
(ν1) as well as the image space M p,q
The corresponding weight ω1 turns rapidly to zero as X = −Y and
X → ∞, while ω1 turns rapidly to infinity as X = Y and X → ∞;
(3) if c1 < 0 and c2 > 0, then the target space M p,q
(ν1) is large and the image
space M p,q
(ν2) is small.
the symbols in M∞,1
on oscillations at infinity, and are extendable to entire functions on C2d;
The corresponding weight ω1 turns rapidly to infinity at infinity. Hence,
(ω1) turn rapidly to zero at infinity, fulfill hard restrictions
(4) if c1 > 0 and c2 < 0, then the target space M p,q
(ν1) is small and the image
space M p,q
(ν2) is large.
57
The corresponding weight ω1 turns rapidly to zero at infinity. Hence, the
(ω1) are allowed to turn rapidly to infinity at infinity, with
symbols in M∞,1
small restrictions on oscillations and singularities at infinity.
6.4. The case of moderate weights. It follows from the general results in pre-
vious sections that almost all results on pseudo-differential operators in [56,57] can
be extended to include weights in the class PE. In what follows we state these
extensions, and leave most of the verifications for the reader.
We start with the following general form of Feichtinger-Grochenig's kernel theo-
rem. The proof is the same as in [57, Prop. 4.7], where Theorem 2.2 in [40] should
be applied instead of the classical Schwartz kernel theorem.
Proposition 6.11. Let d = d1 + d2, ωj ∈ PE(R2dj ) for j = 1, 2 and let ω ∈
PE(Rd ⊕ Rd) be such that
ω(x, y, ξ, η) = ω2(x, ξ)/ω1(y,−η).
Also let T be a linear and continuous map from S1/2(Rd1) to S′1/2(Rd2). Then T
(ω1)(Rd1) to M∞(ω2)(Rd2), if and only if it
extends to a continuous mapping from M 1
exists an element K ∈ M∞(ω)(Rd) such that
(T f )(x) = hK(x,·), fi.
For the proof of the following result we refer to [57, Prop. 4.8] and its proof.
Proposition 6.12. Let t ∈ R, a ∈ S′1/2(R2d), and let K ∈ S′1/2(R2d) be the
distribution kernel for the pseudo-differential operator Opt(a). Also let p ∈ [1,∞],
and ω, ω0 ∈ PE(R2d ⊕ R2d) be such that
ω(x, ξ, η, y) = ω0(x − ty, x + (1 − t)y,−ξ + (1 − t)η, ξ + tη).
(ω)(R2d) if and only if K ∈ M p
(ω0)(R2d). Moreover, if φ ∈ S1/2(R2d)
Then a ∈ M p
and
ψ(x, y) =Z φ((1 − t)x + ty, ξ)eihx−y,ξi dξ,
then kakM p,φ
(ω)
.
= kKkM p,ψ
(ω0)
The next result shows that pseudo-differential operators with symbols in modula-
tion are to some extent invariant under the choice of t in (6.1). We refer to [57, Prop.
1.7] for the proof. Here we let SΦ be the linear and continuous map on S1/2(Rd)
and on S′1/2(Rd), defined by the formula
f 7→ SΦf ≡ (eiΦ ⊗ δV2) ∗ f,
(6.18)
where δV2 is the delta function on the vector space V2 ⊆ Rd and Φ is a real-valued
and non-degenerate quadratic form on V1 = V ⊥2 .
Proposition 6.13. Let φ ∈ S1/2(Rd), ω ∈ PE(R2d), p, q ∈ [1,∞], V1, V2 ⊆ Rd be
vector spaces such that V2 = V ⊥1 . Also let Φ be a real-valued and non-degenerate
58
quadratic form on V1, and let AΦ/2 be the corresponding matrix. If ξ = (ξ1, ξ2)
where ξj ∈ Vj for j = 1, 2, then
(ωΦ)
kSΦfkM p,q,φ
= kfkM p,q,ψ
ωΦ(x, ξ) = ω(x − A−1
(ω)
, where
f ∈ S′1/2(Rd),
Φ ξ1, ξ) and ψ = SΦφ.
In particular, the following are true:
(1) the map (6.18) on S′1/2(Rd) restricts to a homeomorphism from M p,q
(ω)(Rd)
to M p,q
(ωΦ)(Rd);
(2) if t ∈ R, ω0 ∈ PE(R2d ⊕ R2d), and
ωt(x, ξ, η, y) = ω0(x − ty, ξ − tη, y, η),
then the map eithDx,Dξi on S′1/2(R2d) restricts to a homeomorphism from
M p,q
(ω0)(R2d) to M p,q
(ωt)(R2d).
By combining Propositions 6.11 -- 6.12 we get the following result (cf. [57, Thm.
4.6]).
Theorem 6.14. Let t ∈ R and p, q, pj, qj ∈ [1,∞] for j = 1, 2, be such that (6.8)
holds. Also let ω ∈ PE(R4d) and ω1, ω2 ∈ PE(R2d) be such that
ω2(x − ty, ξ + (1 − t)η)
ω1(x + (1 − t)y, ξ − tη) ≤ Cω(x, ξ, η, y),
(6.19)
for some constant C > 0. If a ∈ M p,q
extends uniquely to a continuous map from M p1,q1
(ω)(R2d), then Opt(a) from S1/2(Rd) to S′1/2(Rd)
(ω1) (Rd) to M p2,q2
(ω2) (Rd).
Moreover, if in addition a belongs to the closure of S0(R2d) under the norm
k · kM p,q
(ω)
, then
Opt(a) : M p1,q1
(ω1) (Rd) → M p2,q2
(ω2) (Rd)
is compact.
Theorem 6.15. Let t ∈ R, a ∈ S′1/2(R2d), ω ∈ PE(R2d ⊕ R2d), and ω1, ω2 ∈
PE(R2d) such that (6.9) holds. Then the following is true:
(1) the operator Opt(a) from S1/2(Rd) to S′1/2(Rd) extends to a continuous map-
ping from M 1
(ω1)(Rd) to M∞(ω2)(Rd), if and only if a ∈ M∞(ω)(R2d);
(2) the map a 7→ Opt(a) from M∞(ω)(R2d) to the set of linear and continuous
operators from M 1
(ω1)(Rd) to M∞(ω2)(Rd).
Finally we consider Schatten-von Neumann properties. Let ω1, ω2 ∈ PE(R2d).
Then the set st,p(ω1, ω2) consists of all a ∈ S′1/2(R2d) such that Opt(a) belongs to
Ip(ω1, ω2), the set of Schatten-von Neumann operator of order p ∈ [1,∞] from
M 2
(ω2)(Rd). Note that
(ω1)(Rd) to M 2
I1(ω1, ω2), I2(ω1, ω2)
59
and I∞(ω1, ω2),
are the sets of trace-class, Hilbert-Schmidt and continuous operators respectively,
from M 2
(ω2)(Rd). The space st,p(ω1, ω2) is equipped by the norm
(ω1)(Rd) to M 2
kakst,p(ω1,ω2) ≡ k Opt(a)kIp(ω1,ω2).
By Theorem 6.15 it follows that the map a 7→ Opt(a) from st,p(ω1, ω2) to Ip(ω1, ω2)
is continuous and bijective.
It is easy to obtain a complete characterization of symbols to Hilbert-Schmidt
operators. In fact, we have the following result. We refer to [57, Prop4.11] for the
proof.
Proposition 6.16. Let a ∈ S′1/2(R2d), ω1, ω2 ∈ PE(R2d) and that ω ∈ PE(R2d ⊕
R2d) be such that equality is attained in (6.9) for t = 1/2 and some constant C.
Then Opw(a) ∈ I2(ω1, ω2), if and only if a ∈ M 2
(ω)(R2d). Moreover, for some
constant C > 0 it holds
C−1kakM 2
(ω) ≤ kaksw
2 (ω1,ω2) ≤ CkakM 2
,
(ω)
for every a ∈ S′1/2(R2d).
We have now the following result.
Theorem 6.17. Let t ∈ R and p, q, pj, qj ∈ [1,∞] for j = 1, 2, satisfy
p1 ≤ p ≤ p2,
q1 ≤ min(p, p′) and q2 ≥ max(p, p′).
Also let ω ∈ PE(R2d⊕R2d) and ω1, ω2 ∈ PE(R2d) be such that equality is attained
in (6.19), for some constant C. Then
M p1,q1
(ω) (R2d) ⊆ st,p(ω1, ω2) ⊆ M p2,q2
(ω) (R2d)
(6.20)
Moreover, for some constant C > 0 it holds
C−1kakM
p2,q2
(ω) ≤ kakt,p(ω1,ω2) ≤ CkakM
p1,q1
(ω)
for every a ∈ S′1/2(R2d).
Proof. By Proposition 1.13 and Theorem 6.15 it follows that st,∞ ⊆ M∞(ω), and
by Theorem 6.14 we get M∞,1
(ω) ⊆ st,1 and
st,1 ⊆ M 1,∞(ω) Furthermore, If p1 = p2 = q1 = q2 = 2, then (6.20) follows from and
Propositions 6.13 and 6.16. The result now follows for general p by interpolating
these cases. The proof is complete.
(cid:3)
(ω) ⊆ st,∞. By duality we obtain M 1
6.5. A pseudo-differential calculus in the Bargmann-Fock setting. In this
section we show some possibilities to establish a pseudo-differential calculus on Ba-
nach spaces of analytic functions, in the frame-work of the theory of the Bargmann
transform. The definition of the calculus is in some sense similar to the usual
pseudo-differential calculus, defined in Section 1 (cf. (6.1)). We show that usual
partial differential operators have convenient forms, and remark that the usual
calculus in Section 1, to some extent, can be considered as a part of this pseudo-
differential calculus on analytic functions.
60
Before the definition of the calculus on analytic functions, we consider properties
of compositions of the Bargmann transform with the Fourier transform, translations
or modulations. It is then convenient to introduce some notations.
The Fourier-Bargmann transform FV,t of any function or distribution F on Cd
of order t ∈ R is given by
(FV,tF )(z) = F (e−itπ/2z).
V,t = FV,−t, and that (F −1
We also set FV = FV,1, and call this map the Fourier-Bargmann transform. We
note that F −1
V F )(z) = F (iz). The following lemma shows
that the latter formula is strongly related to Fourier's inversion formula.
Lemma 6.18. Let ω ∈ P 0
ω(e−itπ/2z) = (FV,tω)(z). Then the following is true:
Q(Cd), B be a mixed norm space on Cd, and set ωt(z) =
(1) FV,t restricts to continuous bijective mappings from B(ω, B) to B(ωt, B),
and A(ω, B) to A(ωt, B);
(2) V ◦ F is equal to FV ◦ V as mappings from M(ω, B) to A(ω1, B);
(3) if f ∈ M(ω, B), then
(V(f ( · − x/√2)))(z) = ehz,xi−x2/4(Vf )(z − x),
(V(ei√2h · ,ξif ))(z) = (Vf )(z + iξ).
We note that (2) and (3) in Lemma 6.18 in some special cases were proved
already in [3, 22, 27, 30].
Proof. The assertions (1) and (3) follows immediately from the definitions, and (2)
follows by a straight-forward application of Fourier's inversion formula. The details
are left for the reader.
(cid:3)
By Lemma 6.18 and the investigations in Section 1, it follows that eihx,ξi, F and
dx in (6.1) concerning the usual pseudo-differential calculus correspond to ei(z,w),
FV and dµ(z) respectively. The following definition of our complex version of
pseudo-differential operators, is based on these observations.
Definition 6.19. Let t ∈ R and let a ∈ (S1/2)′(Cd ⊕ Cd) be such that
(1) a(z, w)e−z2−w2+N (hzi+hwi) ∈ S ′(Cd ⊕ Cd) for every N ≥ 0;
(2) if p ∈ P (Cd), then z 7→ ha(z, i· ), e− · 2+(z, · )pi is entire.
Then the (complex) pseudo-differential operator OpV,t(a) with respect to the sym-
bol a is the linear operator from P (Cd) to A(Cd), given by
(OpV,t(a)F )(z)
=ZZ a((1 − t)z + tw1, w2)F (w1)ei((z,w2)−(w2,w1)) dµ(w1)dµ(w2)
=ZZ a((1 − t)z + tw1, iw2)F (w1)e((z,w2)+(w2,w1)) dµ(w1)dµ(w2),
61
(6.21)
when F ∈ P (Cd).
We note that the reproducing kernel in combination with the fact that w1 7→
a((1−t)z +tw1, w2)F (w1) in (6.21) is analytic and satisfying appropriate conditions
give
If
(OpV,t(a)F )(z) =Z a((1 − t)z + tw, w)F (w)ei(z,w) dµ(w).
(w1, w2) 7→ a((1 − t)z + tw1, w2)F (w1)ei((z,w2)−(w2,w1))
in (6.21) is not an integrable function, then OpV,t(a) is defined as the operator with
kernel
(z, w) 7→ π−dΠA(a((1 − t)z + tw, i· )e((z, · )−( · ,w)))e−w2
.
For conveniency we also set OpV = OpV,0.
The following proposition gives motivations for considering operators of the form
OpV,t(a).
Proposition 6.20. Let N ≥ 0 be an integer, aβ ∈ A(Cd) for every β ∈ Nd such
that β ≤ N, and let
a(z, w) = Xβ≤N
aβ(z)wβ.
Then
(OpV(a)F )(z) = Xβ≤N
aβ(z)(DβF )(z),
F ∈ P (Cd).
Proof. The result follows by straight-forward computations, using Remark 4.2 (cid:3)
Remark 6.21. We may use Lemma 6.18 and the mapping properties for the Bargmann
transform to reformulate certain pseudo-differential operators of the form Opt(a)
into pseudo-differential operators given by Definition 6.19. The details are left for
the reader.
Remark 6.22. If a(z, w) = (S−1b)(w/i), then it follows by the definitions that
OpV,t(a) = TV(b). Hence the set of Berezin-Toeplitz operators can be considered
as a subclass of the Bargmann pseudo-differential operators.
Remark 6.23. Let a fulfills the conditions in Definition 6.19, and assume in addition
that w 7→ a(z, w) is analytic. Then it follows by the reproducing formula that
(OpV,t(a)F )(z) = a(z, z)F (z),
when F ∈ P (Cd).
62
References
[1] I. Asekritova, N. Krugliac Real interpolation of vector-valued spaces in non-diagonal case,
Proc. Amer. Math. Soc. 133 (2005), 1665 -- 1675.
[2] I. Asekritova, N. Krugliac Interpolation of Besov spaces in the nondiagonal case, St. Pe-
tersburg Math. J. 18 (2007), 511 -- 516.
[3] V. Bargmann, On a Hilbert space of analytic functions and an associated integral transform,
Comm. Pure Appl. Math., 14 (1961), 187 -- 214.
[4]
On a Hilbert space of analytic functions and an associated integral transform. Part
II. A family of related function spaces. Application to distribution theory., Comm. Pure
Appl. Math., 20 (1967), 1 -- 101.
[5] W. Bauer Berezin-Toeplitz quantization and composition formulas, J. Funct. Anal., 256
(2007), 3107 -- 3142.
[6] R. Beals Characterization of pseudodifferential operators and applications, Duke Math. J.
44 (1977), 45 -- 57.
[7] F. A. Berezin Wick and anti-Wick symbols of operators, Mat. Sb. (N.S.), 86 (1971), 578 -- 610.
[8] J. Bergh, J. Lofstrom Interpolation Spaces, An Introduction, Springer-Verlag, Berlin Hei-
delberg NewYork, 1976.
[9] O. Bilasco, A. Galbis On Taylor coefficients of entire functions integrable against exponential
weights, Math. Nach. 223 (2001), 5 -- 21.
[10] P. Boggiatto Localization operators with Lp symbols on modulation spaces in: P. Boggiatto,
R. Ashino, M. W. Wong (Eds) Advances in Pseudo-Differential Operators, Operator The-
ory: Advances and Applications 155, Birkhauser Verlag, Basel 2004, pp. 149 -- 163.
[11] M. Cappiello Fourier integral operators of infinite order and SG-hyperbolic problems, Thesis,
Department of Mathematics, University of Turin, Turin, 2004.
[12] J. Chung, S.-Y. Chung, D. Kim Characterizations of the Gelfand-Shilov spaces via Fourier
transforms, Proc. Amer. Math. Soc. 124 (1996), 21012108.
[13] L.A. Coburn The Bargmann isometry and Gabor-Daubechies wavelet localization opera-
tors In: Systems, approximation, singular integral operators, and related topics, (Bordeaux,
2000), Oper. Theory Advances and Applications 129, Birkhauser, Basel, 2001, pp. 169 -- 178.
[14] E. Cordero, K. H. Grochenig Time-frequency analysis of localization operators, J. Funct.
Anal. (1) 205, (2003), 107 -- 131.
[15] E. Cordero, S. Pilipovi´c, L. Rodino, N. Teofanov Quasianalytic Gelfand- Shilov spaces with
applications to localization operators, Rocky Mt. J. Math. 40 (2010), 1123-1147.
[16] S. Coriasco, K. Johansson, J. Toft Global Wave Front Set of Modulation Space types,
(preprint, available at arXiv:0912.3366).
[17] I. Daubechies Time-frequency localization operators: a geometric phase space approach,
IEEE Trans. Inform. Th. (4) 34, (1988), 605 -- 612.
[18] H. G. Feichtinger Modulation spaces on locally compact abelian groups. Technical report,
University of Vienna, Vienna, 1983; also in: M. Krishna, R. Radha, S. Thangavelu (Eds)
Wavelets and their applications, Allied Publishers Private Limited, NewDehli Mumbai
Kolkata Chennai Hagpur Ahmedabad Bangalore Hyderbad Lucknow, 2003, pp. 99 -- 140.
[19]
Gabor frames and time-frequency analysis of distributions, J. Funct. Anal. (2) 146
(1997), 464 -- 495.
[20] H. G. Feichtinger, K. H. Grochenig Banach spaces related to integrable group representations
and their atomic decompositions, I, J. Funct. Anal., 86 (1989), 307 -- 340.
[21]
Banach spaces related to integrable group representations and their atomic decom-
positions, II, Monatsh. Math., 108 (1989), 129 -- 148.
[22] H. G. Feichtinger, K. H. Grochenig, D. Walnut Wilson bases and modulation spaces, Math.
Nach. 155 (1992), 7 -- 17.
[23] G. B. Folland Harmonic analysis in phase space, Princeton U. P., Princeton, 1989.
63
[24] I. M. Gelfand, G. E. Shilov Generalized functions, I -- III, Academic Press, NewYork London,
1968.
[25] T. Gramchev, S. Pilipovi´c, L. Rodino Classes of degenerate elliptic operators in Gelfand-
Shilov spaces in: L. Rodino, M. W. Wong (Eds) New developments in pseudo-differential
operators, Operator Theory: Advances and Applications 189, Birkhauser Verlag, Basel
2009, pp. 15 -- 31.
[26] P. Grobner Banachraume Glatter Funktionen und Zerlegungsmethoden, Thesis, University
of Vienna, Vienna, 1992.
[27] K. H. Grochenig Describing functions: atomic decompositions versus frames, Monatsh.
Math.,112 (1991), 1 -- 42.
Foundations of Time-Frequency Analysis, Birkhauser, Boston, 2001.
[28]
[29] K. Grochenig Weight functions in time-frequency analysis in: L. Rodino, M. W. Wong (Eds)
Pseudodifferential Operators: Partial Differential Equations and Time-Frequency Analysis,
Fields Institute Comm., 52 2007, pp. 343 -- 366.
[30] K. Grochenig, M. Leinert Wiener's lemma for twisted convolution and Gabor frames, J.
Amer. Math. Soc., 17 (2004), 1 -- 18.
[31] K. Grochenig, J. Toft Isomorphism properties of Toeplitz operators and pseudo-differential
operators between modulation spaces, J. Math. An. (to appear), also available at arXiv,
arXiv:0905.4954.
[32]
The Range of
localization operators and lifting theorems for modulation and
Bargmann-Fock Spaces, (submitted).
[33] K. Grochenig, D. Walnut A Riesz basis for Bargmann-Fock space related to sampling and
interpolation, Ark. Mat. 30 (1992), 283 -- 295.
[34] K. Grochenig,G. Zimmermann Spaces of test functions via the STFT J. Funct. Spaces Appl.
2 (2004), 25 -- 53.
[35] F. H´erau Melin -- Hormander inequality in a Wiener type pseudo-differential algebra, Ark.
Mat., 39 (2001), 311 -- 38.
[36] A. Holst, J. Toft, P. Wahlberg Weyl product algebras and modulation spaces, J. Funct.
Anal., 251 (2007), 463 -- 491.
[37] L. Hormander The Analysis of Linear Partial Differential Operators, vol I -- III, Springer-
Verlag, Berlin Heidelberg NewYork Tokyo, 1983, 1985.
[38] S. Krantz Function theory of several complex variables, Wiley, NewYork Chichester Brisbane
Toronto Singapore, 1982.
[39] N. Lerner The Wick calculus of pseudo-differential operators and some of its applications,
CUBO, 5 (2003), 213 -- 236.
[40] Z. Lozanov-Crvenkovi´c, D. Perisi´c, M. Taskovi´c Gelfand-Shilov spaces structural and kernel
theorems, (preprint), arXiv:0706.2268v2.
[41] S. Pilipovic Generalization of Zemanian spaces of generalized functions which have or-
thonormal series expansions, SIAM J. Math. Anal. 17 (1986), 477484.
[42] S. Pilipovi´c, N. Teofanov Wilson Bases and Ultramodulation Spaces, Math. Nachr. 242
(2002), 179 -- 196.
[43]
[44]
On a symbol class of Elliptic Pseudodifferential Operators, Bull. Acad. Serbe Sci.
Arts 27 (2002), 57 -- 68.
Pseudodifferential operators on ultra-modulation spaces, J. Funct. Anal.208 (2004),
194 -- 228.
[45] M. Reed, B. Simon Methods of modern mathematical physics, Academic Press, London New
York, 1979.
[46] L. Rodino Linear partial differential operators in Gevrey spaces, World scientific publishing
Co., Singapore, 1993.
[47] M. Signahl, J. Toft Mapping properties for the Bargmann transform on modulation spaces,
arXiv:1002.3061.
64
[48] M. Sugimoto, N. Tomita The dilation property of modulation spaces and their inclusion
relation with Besov Spaces, J. Funct. Anal. (1), 248 (2007), 79 -- 106.
[49] N. Teofanov Ultramodulation spaces and pseudodifferential operators, Endowment An-
drejevi´c, Beograd, 2003.
[50]
Modulation spaces, Gelfand-Shilov spaces and pseudodifferential operators, Sampl.
Theory Signal Image Process, 5 (2006), 225 -- 242.
[51] J. Toft Continuity and Positivity Problems in Pseudo-Differential Calculus, Thesis, Depart-
ment of Mathematics, University of Lund, Lund, 1996.
[52]
[53]
[54]
[55]
[56]
[57]
Subalgebras to a Wiener type Algebra of Pseudo-Differential operators, Ann. Inst.
Fourier (5) 51 (2001), 1347 -- 1383.
Continuity properties for non-commutative convolution algebras with applications in
pseudo-differential calculus, Bull. Sci. Math. (2) 126 (2002), 115 -- 142.
Continuity properties for modulation spaces with applications to pseudo-differential
calculus, I, J. Funct. Anal. (2), 207 (2004), 399 -- 429.
Convolution and embeddings for weighted modulation spaces in: P. Boggiatto, R.
Ashino, M. W. Wong (Eds) Advances in Pseudo-Differential Operators, Operator Theory:
Advances and Applications 155, Birkhauser Verlag, Basel 2004, pp. 165 -- 186.
Continuity properties for modulation spaces with applications to pseudo-differential
calculus, II, Ann. Global Anal. Geom., 26 (2004), 73 -- 106.
Continuity and Schatten properties for pseudo-differential operators on modulation
spaces in: J. Toft, M. W. Wong, H. Zhu (eds) Modern Trends in Pseudo-Differential Op-
erators, Operator Theory: Advances and Applications, Birkhauser Verlag, Basel, 2007,
173 -- 206.
[58]
Continuity and Schatten properties for Toeplitz operators on modulation spaces in: J.
Toft, M. W. Wong, H. Zhu (eds) Modern Trends in Pseudo-Differential Operators, Operator
Theory: Advances and Applications, Birkhauser Verlag, Basel, 2007, 313 -- 328.
[59]
Multiplication properties in pseudo-differential calculus with small regularity on the
symbols, J. Pseudo-Differ. Oper. Appl., 1 (2010), 101 -- 138.
[60] B. Wang, C. Huang Frequency-uniform decomposition method for the generalized BO, KdV
and NLS equations, J. Differential Equations, 239, 2007, 213 -- 250.
Department of Computer science, Physics and Mathematics, Linnaeus University,
Vaxjo, Sweden
E-mail address: [email protected]
65
|
1806.03728 | 2 | 1806 | 2019-12-22T13:26:26 | Hardy inequalities on metric measure spaces | [
"math.FA",
"math.AP",
"math.SP"
] | In this note we give several characterisations of weights for two-weight Hardy inequalities to hold on general metric measure spaces possessing polar decompositions. Since there may be no differentiable structure on such spaces, the inequalities are given in the integral form in the spirit of Hardy's original inequality. We give examples obtaining new weighted Hardy inequalities on $\mathbb R^n$, on homogeneous groups, on hyperbolic spaces, and on Cartan-Hadamard manifolds. | math.FA | math |
HARDY INEQUALITIES ON METRIC MEASURE SPACES
MICHAEL RUZHANSKY AND DAULTI VERMA
Dedicated to G. H. Hardy on the occasion of 100 years of his famous inequality
Abstract. In this note we give several characterisations of weights for two-weight
Hardy inequalities to hold on general metric measure spaces possessing polar de-
compositions. Since there may be no differentiable structure on such spaces, the
inequalities are given in the integral form in the spirit of Hardy's original inequality.
We give examples obtaining new weighted Hardy inequalities on Rn, on homoge-
neous groups, on hyperbolic spaces, and on Cartan-Hadamard manifolds. We note
that doubling conditions are not required for our analysis.
Contents
1.
Introduction
2. Main results
3. Applications and examples
3.1. Homogeneous groups
3.2. Hyperbolic spaces
3.3. Cartan-Hadamard manifolds
4. Equivalence of weight conditions
5. Equivalent conditions for the Hardy inequality
References
1. Introduction
In [Har20], Hardy showed his famous inequality
1
3
5
5
6
7
8
13
16
f (x)pdx,
(1.1)
b (cid:18)R x
Z ∞
b f (t)dt
x
(cid:19)p
dx ≤(cid:18) p
p − 1(cid:19)pZ ∞
b
where p > 1, b > 0, and f ≥ 0 is a nonnegative function. The original discrete version
of this inequality goes back to [Har18] making this year the 100th anniversary of this
topic, see [RS19] for a historical discussion. Consequently, we do not try to provide a
comprehensive historical account here but refer to [RS19] for an extensive historical
overview of the subject.
In particular, since then a lot of work has been done on Hardy inequalities in
different forms and in different settings. It is clearly impossible to give a complete
Date: December 24, 2019.
1991 Mathematics Subject Classification. 26D10, 22E30.
Key words and phrases. Hardy inequalities, metric measure spaces, homogeneous group, hyper-
bolic space, Riemannian manifolds with negative curvature.
1
2
MICHAEL RUZHANSKY AND DAULTI VERMA
overview of the literature, so let us only refer to books and surveys by Opic and Kufner
[OK90], Davies [Dav99], Kufner, Persson and Samko [KP03, KPS17], Edmunds and
Evans [EE04], Mazya [Maz85, Maz11], Ghoussoub and Moradifam [GM13], Balinsky,
Evans and Lewis [BEL15], and references therein. Hardy inequalities in 1D with
weights have been also studied in [AM90, Muc72] in a similar spirit to our approach.
Another list of conditions was given in [GKPW04], however, both in more limited
settings, as well as, more recently, in [GKP10]. More precisely, in our Theorem
2.1 in the special case X = R, the conditions Di correspond to conditions Ai(s),
s = 1, 2, 3, 4, in [GKP10, Theorem 2].
In this paper we show that the inequality (1.1) actually holds in a much more
general setting, also with rather general pairs of weights. However, the weights have
to satisfy certain compatibility conditions for such inequalities to hold true, and these
conditions are necessary and sufficient.
We note that the only known cases of our results are essentially only the Euclidean
ones. The examples we give on homogeneous groups and hyperbolic spaces are new,
but the main result itself is of course more general, completely characterising the
weights for the integral Hardy inequality on metric measure spaces. The importance
of such results is, in particular, in that they lead to a variety of hypoelliptic Hardy-
Sobolev and other inequalities, once we apply it with the weights associated to Riesz
kernels (for hypoelliptic operators, see [RY18b]).
More specifically, we consider metric spaces (X, d) with a Borel measure dx allowing
for the following polar decomposition at a ∈ X: we assume that there is a locally
integrable function λ ∈ L1
(1.2)
loc such that for all f ∈ L1(X) we have
f (r, ω)λ(r, ω)dωrdr,
f (x)dx =Z ∞
0 ZΣr
ZX
for the sets Σr = {x ∈ X : d(x, a) = r} with a measure on it denoted by dω = dωr.
The condition (1.2) is rather general since we allow the function λ to depend on the
whole variable x = (r, ω).
The reason to assume (1.2) is that since X does not have to have a differentiable
structure, the function λ(r, ω) can not be in general obtained as the Jacobian of the
polar change of coordinates. However, if such a differentiable structure exists on X,
the condition (1.2) can be obtained as the standard polar decomposition formula. In
particular, let us give several examples of X for which the condition (1.2) is satisfied
with different expressions for λ(r, ω):
(I) Euclidean space Rn: λ(r, ω) = rn−1.
(II) Homogeneous groups: λ(r, ω) = rQ−1, where Q is the homogeneous dimension
of the group. Such groups have been consistently developed by Folland and
Stein [FS82], see also an up-to-date exposition in [FR16].
(III) Hyperbolic spaces Hn: λ(r, ω) = (sinh r)n−1.
(IV) Cartan-Hadamard manifolds: Let KM be the sectional curvature on (M, g).
A Riemannian manifold (M, g) is called a Cartan-Hadamard manifold if it is
complete, simply connected and has non-positive sectional curvature, i.e., the
sectional curvature KM ≤ 0 along each plane section at each point of M. Let
us fix a point a ∈ M and denote by ρ(x) = d(x, a) the geodesic distance from
x to a on M. The exponential map expa : TaM → M is a diffeomorphism,
HARDY INEQUALITIES ON METRIC MEASURE SPACES
3
see e.g. Helgason [Hel01]. Let J(ρ, ω) be the density function on M, see e.g.
[GHL04]. Then we have the following polar decomposition:
ZM
f (x)dx =Z ∞
0 ZSn−1
f (expa(ρω))J(ρ, ω)ρn−1dρdω,
so that we have (1.2) with λ(ρ, ω) = J(ρ, ω)ρn−1.
(V) Complete manifolds: Let M be a complete manifold. Let p ∈ M and let C(p)
denote the cut locus of p. Let Dp := M\C(p) and S(p; r) := {x ∈ Mp : x =
r}, where · is the Riemannian length, where Mp stands for the tangent space
at p. Then for any p ∈ M and any integrable function f on M we have (e.g.
see [Cha06, Formula III.3.5, P.123]) the polar decompsition
ZM
f dV =Z +∞
0
drZr−1S(p;r)∩Dp
f (exp rξ)√g(r; ξ)dµp(ξ)
(1.3)
for some function √g on Dp, where r−1S(p, r)∩Dp is the subset of Sp obtained
by dividing each of the elements of S(p, r) ∩ Dp by r, and Sp := S(p; 1). Here
dµp(ξ) is the Riemannian measure on Sp induced by the Euclidean Lebesgue
measure on Mp. We refer to [Cha06], [Li12, Chapter 4] and [CLN06, Chapter
1, Paragraph 12] for more details on this decompsition.
Throughout this paper, by A ≈ B we will always mean that the expressions A and
B are equivalent.
2. Main results
We denote by B(a, r) the ball in X with centre a and radius r, i.e
where d is the metric on X. Once and for all we will fix some point a ∈ X, and we
will write
B(a, r) := {x ∈ X : d(x, a) < r},
xa := d(a, x).
Our first main result is the following characterisation of weights u and v for the
corresponding Hardy inequality to hold on X, with the characterisation for the conju-
gate Hardy inequality given in Theorem 2.2. The first condition is the Muckenhoupt
condition while the other ones are equivalent to it.
Theorem 2.1. Let 1 < p ≤ q < ∞ and let s > 0. Let X be a metric measure space
with a polar decomposition (1.2) at a. Let u, v > 0 be measurable functions positive
a.e in X such that u ∈ L1(X\{a}) and v1−p′ ∈ L1
loc(X). Denote
and
Then the inequality
u(y)dy
v1−p′
(y)dy.
U(x) :=ZX\B(a,xa)
V (x) :=ZB(a,xa)
u(x)dx(cid:19) 1
q
(cid:18)ZX(cid:18)ZB(a,xa) f (y)dy(cid:19)q
≤ C(cid:26)ZX f (x)pv(x)dx(cid:27) 1
p
(2.1)
4
MICHAEL RUZHANSKY AND DAULTI VERMA
holds for all measurable functions f : X → C if and only if any of the following
equivalent conditions hold:
q
1
1
q (x)V
p′ (x)(cid:27) < ∞.
(1) D1 := supx6=a(cid:26)U
(2) D2 := supx6=a(cid:26)RX\B(a,xa) u(y)V q( 1
(3) D3 := supx6=a(cid:26)RB(a,xa) u(y)V q( 1
(4) D4 := supx6=a(cid:26)RB(a,xa) v1−p′(y)U p′( 1
(5) D5 := supx6=a(cid:26)RX\B(a,xa) v1−p′(y)U p′( 1
p′ −s)(y)dy(cid:27) 1
p′ +s)(y)dy(cid:27) 1
q −s)(y)dy(cid:27) 1
q +s)(y)dy(cid:27) 1
L1(X).
p′
q
p′
u, v1−p′ ∈ L1(X).
V s(x) < ∞.
V −s(x) < ∞, provided that u, v1−p′ ∈
U s(x) < ∞.
U−s(x) < ∞, provided that
by
and
Moreover, the constant C for which (2.1) holds and quantities D1 − D5 are related
D1 ≤ C ≤ D1(p′)
1
p′ p
1
q ,
(2.2)
1
1
D1 ≤ (max(1, p′s))
q D2, D2 ≤ (max(1,
qD3 ≤ D1 ≤ (1 + sp′)
1
1
(
sp′
1 + p′s
)
D1 ≤ (max(1, qs))
p′ D4, D4 ≤ (max(1,
p′ D5 ≤ D1 ≤ (1 + sq)
1
(
sq
1 + qs
)
1
p′s
))
1
qD1,
qD3,
1
qs
p′ D5.
1
))
1
p′ D1,
In particular, Theorem 2.1 is an extension of (1.1) to the setting of metric measures
spaces X with the polar decomposition (1.2): in particular, for p = q and real-valued
nonnegative measurable f ≥ 0, inequality (2.1) becomes
ZX(cid:18)ZB(a,xa)
f (y)dy(cid:19)p
u(x)dx ≤ CZX
f (x)pv(x)dx,
Indeed, in this case we can take u(x) = 1
as an extension of (1.1).
X = [b,∞), a = b, so that Theorem 2.1 implies (1.1).
For the results in the case of X = R we can refer to [KP03, PS01], and also to
[PSW07] for inequalities for q < p. For X = Rn, the result has been proved in [Ver08],
with related inequalities obtained in one dimension in [GKPW04, OK90]. For related
works on hyperbolic spaces we can refer to [LY17, RY18a], and to [Ngu17, RY18a] for
inequalities on Cartan-Hadamard manifolds, with the background analysis available
in [GHL04, Hel01]. For the analysis of Hardy inequalities on homogeneous groups we
can refer to [RS17, RSY18].
xp , v(x) = 1,
Let us also briefly discuss the conjugate Hardy inequality to that in Theorem 2.1:
holds for all measurable functions f if and only if any of the following equivalent
conditions holds:
≤ C(cid:26)ZX f (x)pv(x)dx(cid:27) 1
p
(2.3)
u(y)dy
v1−p′
(y)dy.
and
Then the inequality
q
(cid:18)ZX(cid:18)ZX\B(a,xa) f (y)dy(cid:19)q
U(x) =ZB(a,xa)
V (x) =ZX\B(a,xa)
u(x)dx(cid:19) 1
(1) D∗1 := supx6=a(cid:26)U
p′ (x)(cid:27) < ∞.
(2) D∗2 := supx6=a(cid:26)RB(a,xa) u(y)V q( 1
(3) D∗3 := supx6=a(cid:26)RX\B(a,xa) u(y)V q( 1
(4) D∗4 := supx6=a(cid:26)RX\B(a,xa) v1−p′(y)U p′( 1
(5) D∗5 := supx6=a(cid:26)RB(a,xa) v1−p′(t)U p′( 1
u, v1−p′ ∈ L1(X).
q (x)V
1
1
u, v1−p′ ∈ L1(X).
HARDY INEQUALITIES ON METRIC MEASURE SPACES
5
Theorem 2.2. Let 1 < p ≤ q < ∞ and s > 0. Let X be a metric measure space with
a polar decomposition as in (1.2). Let u, v > 0 be measurable functions positive a.e.
such that u ∈ L1
loc(X) and v1−p′ ∈ L1(X\{a}). Let
V s(x) < ∞.
q
q
p′ −s)(y)dy(cid:27) 1
p′ +s)(y)dy(cid:27) 1
q −s)(y)dy(cid:27) 1
q +s)(t)dt(cid:27) 1
p′
p′
V −s(x) < ∞, provided that
U s(x) < ∞.
U−s(x) < ∞, provided that
3. Applications and examples
In this section we will give examples of the application of Theorem 2.1 in the
settings of homogeneous groups, hyperbolic spaces, and Cartan-Hadamard manifolds.
3.1. Homogeneous groups. Let G be a homogeneous group of homogeneous di-
mension Q, equipped with a quasi-norm · . For the general description of the setup
of homogeneous groups we refer to [FS82] or [FR16]. Particular example of homo-
geneous groups are the Euclidean space Rn (in which case Q = n), the Heisenberg
group, as well as general stratified groups (homogeneous Carnot groups) and graded
groups.
In relation to the notation of this paper, let us take a = 0 to be the identity of the
group G. We can also simplify the notation denoting xa by x, which is consistent
with the notation for the quasi-norm · .
If we take power weights
u(x) = xα and v(x) = xβ,
6
MICHAEL RUZHANSKY AND DAULTI VERMA
then the inequality (2.1) holds for 1 < p ≤ q < ∞ if and only if
D1 = sup
r>0(cid:18)σZ ∞
r
ραρQ−1dρ(cid:19) 1
q(cid:18)σZ r
0
ρβ(1−p′)ρQ−1dρ(cid:19) 1
p′
< ∞,
where σ is the area of the unit sphere in G with respect to the quasi-norm · . For
this supremum to be well-defined we need to have α + Q < 0 and β(1 − p′) + Q > 0.
Consequently, we have
D1 = σ( 1
q + 1
p′ ) sup
r>0(cid:18)Z ∞
r
ρα+Q−1dρ(cid:19) 1
q(cid:18)Z r
0
ρβ(1−p′)+Q−1dρ(cid:19) 1
p′
= σ( 1
q + 1
p′ ) sup
r>0
α+Q
q
r
α + Q
1
q
β(1−p′)+Q
p′
r
(β(1 − p′) + Q)
,
1
p′
which is finite if and only if the power of r is zero. Summarising, we obtain
Corollary 3.1. Let G be a homogeneous group of homogeneous dimension Q, equipped
with a quasi-norm · . Let 1 < p ≤ q < ∞ and let α, β ∈ R. Then the inequality
(cid:18)ZG(cid:18)ZB(0,x) f (y)dy(cid:19)q
q
xαdx(cid:19) 1
≤ C(cid:26)ZG f (x)pxβdx(cid:27) 1
p
holds for all measurable functions f : G → C if and only if α+Q < 0, β(1−p′)+Q > 0
and α+Q
= 0. Moreover, the constant C for (3.1) satisfies
(3.1)
q + β(1−p′)+Q
q + 1
p′
p′
1
σ
1
α + Q
q (β(1 − p′) + Q)
p′ ≤ C ≤ (p′)
1
1
p′ p
1
q
1
q + 1
p′
σ
1
α + Q
q (β(1 − p′) + Q)
,
1
p′
(3.2)
where σ is the area of the unit sphere in G with respect to the quasi-norm · .
3.2. Hyperbolic spaces. Let Hn be the hyperbolic space of dimension n and let
a ∈ Hn. Let us take the weights
u(x) = (sinhxa)α and v(x) = (sinhxa)β.
Then, passing to polar coordinates, D1 is equivalent to
D1 ≃ sup
xa>0(cid:18)Z ∞
(sinh ρ)α+n−1dρ(cid:19) 1
q(cid:18)Z xa
(sinh ρ)β(1−p′)+n−1dρ(cid:19) 1
p′
xa
For the integrability of the first and the second terms we need, respectively, α+n−1 <
0 and β(1 − p′) + n > 0.
Let us now analyse conditions for this supremum to be finite. For xa ≫ 1, it can
be written as
0
.
xa≫1(cid:18)Z ∞
sup
xa
(exp ρ)α+n−1dρ(cid:19) 1
q(cid:18)Z xa
0
(exp ρ)β(1−p′)+n−1dρ(cid:19) 1
p′
(exp xa)(cid:18) α+n−1
≃ sup
xa≫1
q + β(1−p′)+n−1
p′
(cid:19),
p′
ρβ(1−p′)+n−1dρ(cid:19) 1
(sinh ρ)α+n−1dρ(cid:19) 1
q(cid:18)Z xa
0
ρβ(1−p′)+n−1dρ(cid:19) 1
p′
.
For some small R we have sinh ρxa≤ρ<R ≈ ρ, so that the above supremum is
xa
supxa≪1(cid:18)Z ∞
≃ supxa≪1(cid:18)Z R
≈ supxa≪1(cid:18)xα+n
(sinh ρ)α+n−1dρ(cid:19) 1
q(cid:18)Z xa
(sinh ρ)α+n−1dρ+Z ∞
a + CR(cid:19) 1
xa
β(1−p′)+n
xa
R
.
p′
0
q
Now, for α + n ≥ 0, this is
≈ supxa≪1 xa
β(1−p′)+n
time, for for α + n < 0 it is
p′
HARDY INEQUALITIES ON METRIC MEASURE SPACES
7
which is finite if and only if α+n−1
as
q + β(1−p′)+n−1
p′
≤ 0. For xa ≪ 1, it can be written
, which is finite if and only if β(1−p′)+n
p′
≥ 0. At the same
α+n
q + β(1−p′)+n
≈ supxa≪1 xa
Summarising, we obtain the following
p′
, which is finite if and only if α+n
q + β(1−p′)+n
p′
≥ 0.
Corollary 3.2. Let Hn be the hyperbolic space, a ∈ Hn, and let xa denote the
hyperbolic distance from x to a. Let 1 < p ≤ q < ∞ and let α, β ∈ R. Then the
inequality
(cid:18)ZHn(cid:18)ZB(a,xa) f (y)dy(cid:19)q
(sinh xa)αdx(cid:19) 1
q
≤ C(cid:26)ZHn f (x)p(sinhxa)βdx(cid:27) 1
p
holds for all measurable functions f : Hn → C if and only if the parameters satisfy
either of the following conditions
(A) for α + n ≥ 0, if and only if α + n < 1, β(1− p′) + n > 0 and α+n
q + β(1−p′)+n
p′
≤
1
q + 1
p′ ;
(B) for α + n < 0, if and only if β(1−p′) + n > 0 and 0 ≤ α+n
q + β(1−p′)+n
p′
q + 1
p′ .
≤ 1
3.3. Cartan-Hadamard manifolds. Let (M, g) be a Cartan-Hadamard manifold
and assume that the sectional curvature KM is constant. In this case it is known that
J(t, ω) is a function of t only. More precisely, if KM = −b for b ≥ 0, then J(t, ω) = 1
if b = 0, and J(t, ω) = ( sinh√bt√bt
)n−1 for b > 0, see e.g. [BC01] or [GHL04, p. 166-167].
When b = 0, then let us take u(x) = xα
a , then the inequality
(2.1) holds for 1 < p ≤ q < ∞ if and only if
a and v(x) = xβ
< ∞.
After changing to the polar coordinates, this is equivalent to
dy(cid:19) 1
q(cid:18)ZB(a,xa) yβ(1−p′)
a dy(cid:19) 1
supxa>0(cid:18)ZM\B(a,xa) yα
q(cid:18)R xa
ρβ(1−p′)+n−1dρ(cid:19) 1
supxa>0(cid:18)R ∞
ρα+n−1dρ(cid:19) 1
xa
,
p′
p′
a
0
which is finite if and only if conditions of Corollary 3.1 hold with Q = n (which is
natural since the curvature is zero).
8
MICHAEL RUZHANSKY AND DAULTI VERMA
q(cid:18)ZB(a,xa)
(sinh√bya)αdy(cid:19) 1
After changing to the polar coordinates, this supremum is equivalent to
When b > 0, let us take u(x) = (sinh√bxa)α and v(x) = (sinh √bxa)β. Then he
inequality (2.1) holds for 1 < p ≤ q < ∞ if and only if
supxa>0(cid:18)ZM\B(a,xa)
supxa>0(Z ∞
(sinh√bt)α(
×(Z xa
≃ supxa>0(cid:18)Z ∞
(sinh√bt)β(1−p′)+n−1dt(cid:19) 1
(sinh √bya)β(1−p′)dy(cid:19) 1
(sinh √bt)α+n−1dt(cid:19) 1
which has the same conditions for finiteness as the case of the hyperbolic space in
Corollary 3.2 (which is also natural since it is the negative constant curvature case).
(sinh√bt)β(1−p′)(
q(cid:18)Z xa
0
√bt
sinh √bt
)n−1tn−1dt)
1
p′
p′
< ∞.
sinh √bt
)n−1tn−1dt)
1
q
√bt
p′
,
xa
xa
0
4. Equivalence of weight conditions
In this section we prove that the quantities D1 -- D5 involving the weights in Theorem
2.1 are equivalent. However, it will be convenient to formulate it in the following
slightly more general form:
Theorem 4.1. Let α, β, s > 0 and let f ∈ L1(X\{a}), g ∈ L1
f, g > 0 are positive a.e in X. Let us denote
loc(X), be such that
Then the following quantities are equivalent:
provided that G(β−s)/α(y) makes sense.
(1) A1 := supx6=a A1(x; α, β) := supx6=a F α(x)Gβ(x).
(2) A2 := supx6=a A2(x; α, β, s) := supx6=a(cid:18)RX\B(a,xa) f (y)G(β−s)/α(y)dy(cid:19)α
(3) A3 := supx6=a A3(x; α, β, s) := supx6=a(cid:18)RB(a,xa) g(y)F (α−s)/β(y)dy(cid:19)β
(4) A4 := supx6=a A4(x; α, β, s) := supx6=a(cid:18)RB(a,xa) f (y)G(β+s)/α(y)dy(cid:19)α
(5) A5 := supx6=a A5(x; α, β, s) := supx6=a(cid:18)RX\B(a,xa) g(y)F (α+s)/β(y)dy(cid:19)β
provided that f, g ∈ L1(X) and that G−s(x) makes sense.
provided f, g ∈ L1(X) and that that F −s(x) makes sense.
provided that F (α−s)/β(y) makes sense.
Gs(x),
F s(x),
G−s(x),
F −s(x),
Moreover, we have the following relations between the above quantities:
and
F (x) :=ZX\B(a,xa)
G(x) :=ZB(a,xa)
f (y)dy,
g(y)dy.
HARDY INEQUALITIES ON METRIC MEASURE SPACES
9
• A1 ≤ (max(1, s
• A1 ≤ (max(1, s
• ( s
β+s )αA4 ≤ A1 ≤ (1 + s
β ))αA2 and A2 ≤ (max(1, β
α))βA3 and A3 ≤ (max(1, α
s ))αA1;
s ))βA1;
β )αA4 and ( s
α+s )βA5 ≤ A1 ≤ (1 + s
α )βA5.
Proof of Theorem 4.1. A1 ≈ A2
G(β−s)/α(x). Consequently, we can estimate
We will first consider the case s ≤ β. Then for ya ≥ xa we have G(β−s)/α(y) ≥
A2(x; α, β, s) =(cid:18)ZX\B(a,xa)
≥(cid:18)ZX\B(a,xa)
f (y)G(β−s)/α(y)dy(cid:19)α
f (y)dy(cid:19)α
Gβ(x)
= F α(x)Gβ(x),
Gs(x)
which implies A2 ≥ A1. For s > β, let us first introduce some notation, using the
polar decomposition (1.2). First, we denote
W (x) :=ZX\B(a,xa)
xaZΣr
=Z ∞
=Z ∞
xafW (r)dr
=: fW1(xa),
fG1(r) :=Z r
0 ZΣs
fW (r) :=ZΣr
fF1(r) :=Z ∞
r ZΣs
g(s, σ)λ(s, σ)dσs, and
where
with eG(s) :=RΣs
Moreover, we denote
f (y)G(β−s)/α(y)dy
f (r, ω)λ(r, ω)fG1(r)(β−s)/αdωrdr
g(s, σ)λ(s, σ)dσsds =Z r
0 eG(s)ds,
f (r, ω)λ(r, ω)fG1(r)(β−s)/αdωr.
λ(s, σ)f (s, σ)dσsds =Z ∞
r
eF (s)ds.
10
MICHAEL RUZHANSKY AND DAULTI VERMA
Using the function W defined above, we can estimate
(s−β)/s
(β−s)/α
(s−β)/α
f (y)G(β−s)/α(y)G(s−β)/α(y)W (s−β)/s(y)W (β−s)/s(y)dy(cid:19)α
(r)fW1
(r)fG1
λ(r, ω)f (r, ω)fG1
(r)(cid:19)Gβ(x)(cid:18)Z ∞
xafW (r)(Z ∞
r fW (s)ds)
(r)fW1
Gs(y)W α(y)(cid:19)(s−β)/s(cid:18) s
β(cid:19)α
Gs(x)W α(x)(cid:19)β/s
Gs(y)W α(y)(cid:19)(1−β/s)(cid:18) sup
(r)fW1
Gβ(x)W (βα)/s(x)
(s−β)α/s
(β−s)/s
xa>0
dr(cid:19)α
F α(x)Gβ(x)
(s−β)
= Gβ(x)(cid:18)ZX\B(a,xa)
= Gβ(x)(cid:18)Z ∞
xaZΣr
≤(cid:18) sup
r>xafG1
=(cid:18) sup
β(cid:19)α(cid:18) sup
≤(cid:18) s
≤(cid:18) s
β(cid:19)α
sup
xa>0
ya>xa
ya>xa
(β−s)/s
(r)dωrdr(cid:19)α
A2(x; α, β, s).
Therefore, we obtain
β(cid:19)α
A1 ≤(cid:18) s
A2.
Hence, we have for every s > 0 the inequality
A1 ≤(cid:18)max(1,
s
β
)(cid:19)α
A2.
Conversely, we have for s < β,
Gs(x)W α(x)
= Gs(x)(cid:18)ZX\B(a,xa)
= Gs(x)(cid:18)Z ∞
xaZΣr
(r)fF1
×fF1
= Gs(x)(cid:18)Z ∞
xaZΣr
(β−s)/β
λ(s, σ)g(s, σ)dsdσs(cid:19)(β−s)/α
f (y)G(β−s)/α(y)F (β−s)/β(y)F (s−β)/β(y)dy(cid:19)α
λ(r, ω)f (r, ω)(cid:18)Z r
0 ZΣs
(r)drdωr(cid:19)α
λ(r, ω)f (r, ω)fG1
(r)fF1
(r)fF1
(β−s)/α
(s−β)/β
(β−s)/β
(s−β)/β
(r)dωrdr(cid:19)α
.
HARDY INEQUALITIES ON METRIC MEASURE SPACES
11
Consequently, we can estimate
Gs(x)W α(x)
eF (s)ds(cid:19)(s−β)/β
dr(cid:19)α
(β−s)/β
(β−s)/α
(r)fF1
(r)(cid:19)(β−s)/β
(r)fF1
Gβ(y)F α(y)(cid:19)(β−s)/β(cid:18) β
(r)(cid:19)α
Gs(x)(cid:18)Z ∞
xa eF (r)(cid:18)Z ∞
s(cid:19)α
Gs(x)(cid:18)β
fF1
Gβ(x)F α(x)(cid:19)s/β
s(cid:19)α(cid:18) sup
xa>0
(αs)/β
(r)
r
β
α
≤(cid:18) sup
r>xafG1
=(cid:18) sup
r>xafG1
≤(cid:18) sup
s(cid:19)α
≤(cid:18)β
which gives A2 ≤(cid:0) β
s(cid:1)α
sup
xa>0
ya>xa
A1(x; α, β),
A1.
On the other hand, for s ≥ β, when ya > xa we have G(β−s)/α(y) ≤ G(β−s)/α(x).
Therefore, we can estimate
A2(x; α, β, s) = Gs(x)(cid:18)ZX\B(a,xa)
≤ Gs(x)(cid:18)ZX\B(a,xa)
f (y)G(β−s)/α(y)dy(cid:19)α
f (y)dy(cid:19)α
G(β−s)(x)
= F α(x)Gβ(x)
i.e. A2 ≤ A1. Therefore, we have for s > 0, the overall estimate
A2 ≤(cid:18)max(1,
β
s
)(cid:19)α
A1.
Hence we have also shown that A1 ≈ A2.
of A1 ≈ A2, where we just need to interchange the roles of F and G.
Next we observe that the proof of A1 ≈ A3 follows along the same lines as that
A1 ≈ A4
Let us denote
f (y)G(β+s)/α(y)dy
λ(r, ω)f (r, ω)G(β+s)/α(r, ω)dωrdr
W0(x) :=ZB(a,xa)
ZΣr
=Z xa
=:Z xa
0 fW0(r)dr,
0
so that we can write
A4(x; α, β, s) = G−s(x)W α
0 (x).
12
MICHAEL RUZHANSKY AND DAULTI VERMA
We rewrite A1 as
A1(x; α, β) = Gβ(x)(cid:18)ZX\B(a,xa)
= Gβ(x)(cid:18)Z ∞
xaZΣr
= Gβ(x)(cid:18)Z ∞
xafG1
We can estimate this by
≤ Gβ(x)(cid:18)fG1
≤ Gβ(x)(cid:18) sup
×(cid:18)fG1
≤ Gβ(x) sup
ya>0
−(β+s)/α
(∞)W0(∞) +
α
−β/α
(β + s)
G−s(y)W α
ya>xa
(∞) +
0 (y)(cid:19)×
Z ∞
xafG1
A4(y; α, β, s)(cid:18)fG1
A4(y; α, β, s)(cid:20)(β + s)
+(cid:18)1 −
β(cid:19)α
A4(y; α, β, s),
sup
ya>0
−β/α
β
s
= sup
ya>0
≤(cid:18)1 +
−(β+s)/α
.
d
(r)
(β + s)
f (y)G(β+s)/α(y)G−(β+s)/α(y)dy(cid:19)α
λ(r, ω)f (r, ω)G(β+s)/α(r, ω)G−(β+s)/α(r, ω)dωrdr(cid:19)α
dr(cid:18)Z r
0 fW0(s)ds(cid:19)dr(cid:19)α
Z ∞
xa eG(r)(fG1(r))
dr(cid:18)Z r
0 eG(s)ds(cid:19)dr(cid:19)α
β (cid:18)fG1
(xa) −fG1
β (cid:19)(cid:18) G(x)
G(∞)(cid:19)β/α(cid:21)α
α −1W0(r)dr(cid:19)α
(∞) +
−(β/α)−1
(β + s)
(β + s)
−(β/α)
−β/α
(r)
−(β+s)
α
d
(∞)(cid:19)(cid:19)α
where the expressions like G(∞) make sense since g ∈ L1(X). Therefore, we obtain
A1 ≤ (1 + s/β)αA4.
To prove the opposite inequality, we assume that
sup
xa>0
A1(x; α, β) < ∞.
Then we have
d
(r)
(β+s)/α
A4(x; α, β, s) = G−s(x)(cid:18)ZB(a,xa)
= G−s(x)(cid:18)Z xa
0 fG1
(r)fF1(r)(cid:12)(cid:12)(cid:12)(cid:12)
(r)fF1
0<r<xafG1
G(β+s)/α(y)f (y)dy(cid:19)α
dr(cid:18) −Z ∞
eF (s)ds(cid:19)dr(cid:19)α
dr(cid:18)Z r
α Z xa
0 fF1(r)fG1
(r)(cid:19)(cid:18) β + s
0 eG(s)ds)dr(cid:19)α
α Z xa
(Z r
0 fG1
(β+s)/α−1
s/α−1
β + s
d
dr
xa
(r)
(r)
+
(β+s)/α
d
α
β
0
r
0 eG(s)ds(cid:19)dr(cid:19)α
= G−s(x)(cid:18)fG1
≤ G−s(x)(cid:18) sup
HARDY INEQUALITIES ON METRIC MEASURE SPACES
13
α (cid:19)α
≤(cid:18) β + s
Gs/α(x)(cid:19)α
s
sup
ya>0
Gβ(y)F α(y)G−s(x)(cid:18)α
=(cid:18)β + s
s (cid:19)α
sup
xa>0
A1(x; α, β),
where we have used that f ∈ L1(X). Hence we have proved that A1 ≈ A4.
interchange the roles of F and G.
follows the same lines as that of the case A1 ≈ A4 if we
The proof of A1 ≈ A5
(cid:3)
5. Equivalent conditions for the Hardy inequality
In this section we prove Theorem 2.1 and also give some comments concerning
Theorem 2.2. Without loss of generality we can assume that f ≥ 0. Then we observe
that if in Theorem 4.1 we take
f (x) = u(x), g(x) = v1−p′
(x), α =
1
q
, β =
1
p′
,
then it follows that we have the equivalence of the quantities
D1 ≈ D2 ≈ D3 ≈ D4 ≈ D5.
So, we first assume that any one of these equivalent conditions holds true. In partic-
ular, D1 < ∞, and using polar coordinates (1.2), we have for every a > 0 that
λ(r, ω)v1−p′
(r, ω)dωrdr(cid:27) 1
p′
≤ D1.
(5.1)
(cid:26)Z ∞
a ZΣr
We denote
and
λ(r, ω)u(r, ω)dωrdr(cid:27) 1
q(cid:26)Z a
0 ZΣr
h(t) :=(cid:18)Z t
0 ZΣs
eU1(t) :=ZΣt
F1(s) :=ZΣs
H1(t) :=Z t
0 ZΣs
λ(s, σ)v1−p′
(s, σ)dsdσs(cid:19) 1
pp′
,
λ(t, ω)u(t, ω)dωt,
λ(s, σ)[f (s, σ)v
1
p (s, σ)h(s)]pdσs,
λ(s, σ)[v
1
p (s, σ)h(s)]−p′
dσsds.
14
MICHAEL RUZHANSKY AND DAULTI VERMA
Then using polar coordinates (1.2), Holder's inequality, and Minkowski's inequality,
the left side of (2.1) can be estimated as
1
p (s, σ)h(s)]pdsdσs(cid:19) q
p
dx
f (y)dy(cid:19)q
u(x)(cid:18)ZB(a,xa)
ZX
λ(r, ω)u(r, ω)(cid:18)Z r
≤Z ∞
0 ZΣs
0 ZΣr
×(cid:18)Z r
0 ZΣs
eU1(r)(cid:18)Z r
=Z ∞
≤(cid:18)Z ∞
F1(s)(cid:18)Z ∞
F1(s)ds(cid:19) q
eU1(r)H
λ(s, σ)[v
H
q
p′
q
p′
0
0
0
s
1
p
λ(s, σ)[f (s, σ)v
p′
dsdσs(cid:19) q
p (s, σ)h(s)]−p′
dωrdr
1 (r)dr
q
1 (r)dr(cid:19) p
p
ds(cid:19) q
.
(5.2)
Denoting
V1(s) :=ZΣs
λ(s, σ)v1−p′
(s, σ)dσs,
and using inequality (5.1) we can estimate
p
1
0
0
dr
λ(r, σ)[v
λ(r, σ)v1−p′
p (r, σ)h(r)]−p′
(r, σ)(cid:18)Z r
0 ZΣρ
V1(ρ)dρ(cid:19)− 1
H1(t) =Z t
0 ZΣr
=Z t
0 ZΣr
V1(r)(cid:18)Z r
=Z t
= p′(cid:18)Z t
V1(r)dr(cid:19) 1
(r, σ)drdσr(cid:19) 1
= p′(cid:18)Z t
0 ZΣr
λ(r, σ)u(r, σ)drdσr(cid:19)− 1
×(cid:18)Z ∞
t ZΣr
≤ p′D1(cid:18)Z ∞
eU1(s)ds(cid:19)− 1
λ(r, σ)v1−p′
.
p′
0
t
q
q
dσrdr
λ(ρ, ω)v1−p′
(ρ, ω)dρdωρ(cid:19)− 1
p
drdσr
p′(cid:18)Z ∞
t ZΣr
λ(r, σ)u(r, σ)drdσr(cid:19) 1
q
(5.3)
HARDY INEQUALITIES ON METRIC MEASURE SPACES
15
At the same time we can also estimate
Z ∞
s
eU1(t)(cid:18)Z ∞
t
p′
eU1(τ )dτ(cid:19)− 1
p
p
p
t
s
s
d
dt
dt(cid:18)Z ∞
eU1(τ )dτ(cid:19) 1
dt = −pZ ∞
= p(cid:18)Z ∞
eU1(t)dt(cid:19) 1
λ(t, ω)u(t, ω)dtdωt(cid:19) 1
= p(cid:18)Z ∞
s ZΣt
= p((cid:18)Z ∞
λ(t, ω)u(t, ω)dtdωt(cid:19) 1
s ZΣt
p′) q
(t, ω)dtdωt(cid:19) 1
×(cid:18)Z s
0 ZΣt
(t, ω)dtdωt(cid:19)− q
×(cid:18)Z s
0 ZΣt
λ(t, ω)v1−p′
λ(t, ω)v1−p′
p′ p
q
p
q
p
1 h−q(s),
≤ pD
(5.4)
in view of (5.1). Therefore using (5.3) and (5.4) in (5.2), we have
ZX
u(x)(cid:18)ZB(a,xa)
f (y)dy(cid:19)q
dx ≤ Dq
1p′
q
p′ p(cid:18)ZX
v(x)f (x)pdx(cid:19) q
p
.
Hence, it follows that (2.1) holds with C ≤ D1(p′)
in (2.2).
1
p′ p
1
q proving one of the relations
Conversely, let us assume that inequality (2.1) holds, and consider the function
f (x) = v1−p′
(x)χ(0,t)(xa)
for some t > 0, and where χ is the cut-off function. With this function, the right
hand side of of (2.1) takes the form
At the same time, the left hand side of of (2.1) takes the form
(cid:18)ZX(cid:18)ZB(a,xa) f (y)dy(cid:19)q
q
u(x)dx(cid:19) 1
(cid:18)ZX
v(x)f (x)pdx(cid:19) 1
p
p
.
v1−p′
(x)dx(cid:19) 1
=(cid:18)Zxa≤t
≥(cid:18)Zxa≥t(cid:18)ZB(a,xa) f (y)dy(cid:19)q
(y)dy(cid:19)q
=(cid:18)Zxa≥t(cid:18)Zya≤t
u(x)dx(cid:19) 1
=(cid:18)Zxa≥t
q(cid:18)Zya≤t
v1−p′
v1−p′
q
q
u(x)dx(cid:19) 1
u(x)dx(cid:19) 1
(y)dy(cid:19).
16
MICHAEL RUZHANSKY AND DAULTI VERMA
Altogether the inequality (2.1) takes the form
(cid:18)Zxa≥t
u(x)dx(cid:19) 1
q(cid:18)Zya≤t
v1−p′
(y)dy(cid:19) ≤ C(cid:18)Zxa≤t
v1−p′
p
(x)dx(cid:19) 1
,
which gives D1 ≤ C. Hence, we have the equivalence and the second relation in (2.2).
q in
As for Theorem 2.2, if we take f (x) = v1−p′ , g(x) = u(x), α = 1
p′ and β = 1
Theorem (4.1), we find that
D∗1 ≈ D∗2 ≈ D∗3 ≈ D∗4 ≈ D∗5.
Consequently, we can show Theorem 2.2 by the argument similar to that in Section
5 where Theorem 2.1 was proved. We also note that in the case of homogeneous
groups, we can actually also derive it from Theorem 2.1 by the involutive change of
variables x 7→ x−1.
Data accessibility. No new data was collected or generated during the course of
research.
Competing interests. We have no competing interests.
Authors' contributions. The authors contributed equally to this paper.
Acknowledgements. The authors would like to thank Nurgissa Yessirkegenov for
discussions and for checking our calculations. We would also like to thank all 5
referees of this paper for useful comments.
Funding statement. The first author was supported in parts by the FWO Odysseus
Project, EPSRC grant EP/R003025/1 and by the Leverhulme Grant RPG-2017-151.
Ethics statement. The work did not involve any collection of human data.
References
[AM90]
[BEL15]
[BC01]
[Cha06]
[CLN06]
[Dav99]
[EE04]
[FR16]
[FS82]
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.
A. A. Balinsky, W. D. Evans, and R. T. Lewis. The analysis and geometry of Hardy's
inequality. Universitext. Springer, Cham, 2015.
R. Bishop and R. Crittenden. Geometry of manifolds. Reprint of the 1964 original. AMS
Chelsea Publishing, Providence, RI, 2001.
I. Chavel. Riemannian geometry: a modern introduction. New York Cambridge Univer-
sity Press, Cambridge studies in advanced mathematics, second edition, 2006.
B. Chow, P. Lu, L. Ni. Hamilton's Ricci Flow. AMS Graduate Studies in Mathematics,
77, Providence, RI, 2006.
E. B. Davies. A review of Hardy inequalities. In The Maz'ya anniversary collection, Vol.
2 (Rostock, 1998), volume 110 of Oper. Theory Adv. Appl., pages 55 -- 67. Birkhauser,
Basel, 1999.
D. E. Edmunds and W. D. Evans. Hardy operators, function spaces and embeddings.
Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2004.
V. Fischer and M. Ruzhansky. Quantization on nilpotent Lie groups, volume 314 of
Progress in Mathematics. Birkhauser. (open access book), 2016.
G. B. Folland and E. M. Stein. Hardy spaces on homogeneous groups, volume 28 of
Mathematical Notes. Princeton University Press, Princeton, N.J.; University of Tokyo
Press, Tokyo, 1982.
HARDY INEQUALITIES ON METRIC MEASURE SPACES
17
[KP03]
[KPS17]
[GHL04]
[GKP10]
[Li12]
[LY17]
[Har20]
[Hel01]
S. Gallot, D. Hulin, and J. Lafontaine. Riemannian geometry. Universitext. Springer-
Verlag, Berlin, third edition, 2004.
A. Gogatishvili, A. Kufner, and L.-E. Persson. Some new scales of characterization of
Hardys inequality. Proceedings of the Estonian Academy of Sciences, 59(1):7 -- 18, 2010.
[GKPW04] A. Gogatishvili, A. Kufner, L.-E. Persson, and A. Wedestig. An equivalence theorem for
integral conditions related to Hardy's inequality. Real Anal. Exchange, 29(2):867 -- 880,
2003/04.
N. Ghoussoub and A. Moradifam. Functional inequalities: new perspectives and new
applications, volume 187 of Mathematical Surveys and Monographs. American Mathe-
matical Society, Providence, RI, 2013.
G. H. Hardy. Notes on some points in the integral calculus. Messenger Math, 48:107 -- 112,
1918.
G. H. Hardy. Note on a theorem of Hilbert. Math. Z., 6(3-4):314 -- 317, 1920.
S. Helgason. Differential geometry, Lie groups, and symmetric spaces, volume 34 of
Graduate Studies in Mathematics. American Mathematical Society, Providence, RI,
2001. Corrected reprint of the 1978 original.
A. Kufner and L.-E. Persson. Weighted inequalities of Hardy type. World Scientific Pub-
lishing Co., Inc., River Edge, NJ, 2003.
A. Kufner, L.-E. Persson, and N. Samko. Weighted inequalities of Hardy type. World
Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, second edition, 2017.
P. Li. Geometric analysis. Cambridge Studies in Advanced Mathematics, 134, 2012.
G. Lu and Q. Yang. Sharp Hardy-Adams inequalities for bi-Laplacian on hyperbolic
space of dimension four. Adv. Math., 319:567 -- 598, 2017.
V. G. Maz'ja. Sobolev spaces. Springer Series in Soviet Mathematics. Springer-Verlag,
Berlin, 1985. Translated from the Russian by T. O. Shaposhnikova.
V. Maz'ya. Sobolev spaces with applications to elliptic partial differential equations, vol-
ume 342 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles
of Mathematical Sciences]. Springer, Heidelberg, augmented edition, 2011.
B. Muckenhoupt. Hardy's inequality with weights. Studia Mathematica , T.XLJV, 1972,
31 -- 38.
V. H. Nguyen. Sharp Hardy and Rellich type inequalities on Cartan-Hadamard mani-
folds and their improvements. arXiv:1708.09306, 2017.
B. Opic and A. Kufner. Hardy-type inequalities, volume 219 of Pitman Research Notes
in Mathematics Series. Longman Scientific & Technical, Harlow, 1990.
L. E. Persson and V. D. Stepanov. Weighted integral inequalities with a geometric mean.
Dokl. Akad. Nauk, 377(4):439 -- 440, 2001.
L.-E. Persson, V. Stepanov, and P. Wall. Some scales of equivalent weight character-
izations of Hardy's inequality: the case q < p. Math. Inequal. Appl., 10(2):267 -- 279,
2007.
M. Ruzhansky and D. Suragan. Hardy and Rellich inequalities, identities, and sharp
remainders on homogeneous groups. Adv. Math., 317:799 -- 822, 2017.
M. Ruzhansky and D. Suragan. Hardy inequalities on homogeneous groups. Progress in
Math., Vol. 327, Birkhauser, 2019. (open access book)
[RS17]
[RS19]
[OK90]
[PS01]
[Maz85]
[Maz11]
[Muc72]
[Ngu17]
[GM13]
[Har18]
[PSW07]
[RSY18] M. Ruzhansky, D. Suragan, and N. Yessirkegenov. Extended Caffarelli-Kohn-Nirenberg
inequalities, and remainders, stability, and superweights for Lp-weighted Hardy inequal-
ities. Trans. Amer. Math. Soc. Ser. B, 5:32 -- 62, 2018.
M. Ruzhansky and N. Yessirkegenov. Hardy, weighted Trudinger-Moser and Caffarelli-
Kohn-Nirenberg type inequalities on Riemannian manifolds with negative curvature.
arXiv:1802.09072, 2018.
[RY18a]
[RY18b] M. Ruzhansky
and N. Yessirkegenov. Hypoelliptic
functional
inequalities.
[Ver08]
https://arxiv.org/abs/1805.01064, 2018.
D. Verma. Equivalent conditions for Hardy inequalities. Proc. A. Razmadze Math. Inst.,
148:107 -- 116, 2008.
18
MICHAEL RUZHANSKY AND DAULTI VERMA
Michael Ruzhansky:
Department of Mathematics
Imperial College London
180 Queen's Gate, London, SW7 2AZ
United Kingdom
and
Department of Mathematics: Analysis, Logic and Discrete Mathematics
Ghent University, Belgium
and
School of Mathematical Sciences
Queen Mary University of London
United Kingdom
E-mail address [email protected]
Daulti Verma:
Miranda House College
University of Delhi
Delhi 110007
India
and
Department of Mathematics
Imperial College London
180 Queen's Gate, London, SW7 2AZ
United Kingdom
E-mail address [email protected]
|
1812.00582 | 3 | 1812 | 2019-12-11T09:29:41 | Eigenvalues of the Neumann-Poincare operator in dimension 3: Weyl's law and geometry | [
"math.FA",
"math.SP"
] | We consider the asymptotic properties of the eigenvalues of the Neumann-Poincare (NP) operator in three dimensions. The region $\Omega\subset \mathbb{R}^3$ is bounded by a compact surface $\Gamma=\partial \Omega$, with certain smoothness conditions imposed. The NP operator is defined by $ \mathcal{K}_{\Gamma}[\psi](\mathbf{x}):=\frac{1}{4\pi}\int_\Gamma\frac{\langle \mathbf{y}-\mathbf{x},\mathbf{n}(\mathbf{y})\rangle}{|\mathbf{x}-\mathbf{y}|^3}\psi(\mathbf{y})dS_{\mathbf{y}},$ where $dS_{\mathbf{y}}$ is the surface element and $\mathbf{n}(\mathbf{y})$ is the outer unit normal vector on $\Gamma$. The first-named author established earlier that the singular numbers $s_j(\mathcal{K}_{\Gamma})$ of $\mathcal{K}_{\Gamma}$ and the ordered moduli of its eigenvalues $\lambda_j(\mathcal{K}_{\Gamma})$ satisfy the Weyl law with coefficient expressed in geometric terms. Our main purpose here is to investigate the asymptotic behavior of positive and negative eigenvalues separately under the condition of infinite smoothness of the boundary $\Gamma$. These formulas are used, in particular, to obtain certain answers to the long-standing problem of the existence or finiteness of negative eigenvalues of $\mathcal{K}_{\Gamma}$. | math.FA | math |
EIGENVALUES OF THE NEUMANN-POINCARE OPERATOR IN
DIMENSION 3: WEYL'S LAW AND GEOMETRY
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
To Volodya Maz'ya, an outstanding mathematician
Abstract. We consider the asymptotic properties of the eigenvalues of the Neumann-
Poincare (NP) operator in three dimensions. The region Ω ⊂ R3 is bounded by a
compact surface Γ = ∂Ω, with certain smoothness conditions imposed. The NP
operator KΓ, called often 'the direct value of the double layer potential', acting in
L2(Γ), is defined by
1
4π ZΓ
KΓ[ψ](x) :=
hy − x, n(y)i
x − y3 ψ(y)dSy,
where dSy is the surface element and n(y) is the outer unit normal vector on Γ.
The first-named author proved in [26] that the singular numbers sj(KΓ) of KΓ and
the ordered moduli of its eigenvalues λj(KΓ) satisfy the Weyl law
sj(K(Γ)) ∼ λj(KΓ) ∼(cid:26) 3W (Γ) − 2πχ(Γ)
128π
2
(cid:27) 1
j−
1
2 ,
under the condition that Γ belongs to the class C 2,α with α > 0, where W (Γ) and
χ(Γ) denote, respectively, the Willmore energy and the Euler characteristic of the
boundary surface Γ. Although the NP operator is not self-adjoint (and therefore
no general relations between eigenvalues and singular number exist), the ordered
moduli of the eigenvalues of KΓ satisfy the same asymptotic relation.
Our main purpose here is to investigate the asymptotic behavior of positive and
negative eigenvalues separately under the condition of infinite smoothness of the
boundary Γ. These formulas are used, in particular, to obtain certain answers to
the long-standing problem of the existence or finiteness of negative eigenvalues of
KΓ. A more sophisticated estimation allows us to give a natural extension of the
Weyl's law for the case of a smooth boundary.
1. Introduction and Results
1.1. Introduction. The Neumann -- Poincar´e (abbreviated by NP) operator is the
boundary integral operator which appears naturally when solving classical boundary
value problems using layer potentials. Its study (for the Laplace operator) goes back
to C. Neumann [27] and H. Poincar´e [29] as the name of the operator suggests (this
names combination was first used by T. Carleman in his Thesis [13] and became
conventional afterwards). If the boundary of the domain on which the NP operator
is defined, is C 1,α smooth, then the NP operator is compact. Thus the second
2010 Mathematics Subject Classification. 47A75 (primary), 58J50 (secondary).
Key words and phrases. Neumann-Poincare operator, Eigenvalues, Weyl's law, Pseudo-
differential operators, Willmore energy.
The second-named author is supported by grant RFBR No 17-01-00668.
1
2
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
kind Fredholm integral equation, which appears when solving Dirichlet or Neumann
problems, is subject to the Fredholm index theory.
The study of spectral properties of the NP operator was initiated by S. Zaremba,
[38]. Later on, it was proved in [18] that the NP operator, not self-adjoint, generally,
in L2, can be however realized as a self-adjoint operator in the H −1/2- Sobolev space,
provided a new inner product is introduced there, and therefore the NP spectrum
is real, may consist of a continuous spectrum and a discrete spectrum (and possibly
limit points of the discrete spectrum). If the domain is only Lipschitz, in particular,
has corners, the corresponding NP operator does, in fact, possess a continuous spec-
trum (as well as eigenvalues). If the domain has a smoother boundary, say, C 1,α,
then the spectrum consists of eigenvalues (converging to 0 if there are infinitely
many of them) and the point zero.
It turns out that the properties of eigenvalues in the two-dimensional and higher
dimensional cases are quite different. In the two-dimensional case, the rate of decay
of these eigenvalues depends in a crucial way on the smoothness of the boundary, and
their exact asymptotic behavior is presently known only for a few examples when
these eigenvalues can be calculated explicitly. We refer to [7, 25] for the progress on
the convergence rate of NP eigenvalues in two dimensions.
As for the three-dimensional case, a general reasoning implies that the NP op-
erator, being in the smooth case a pseudodifferential operator of order −1, should
have eigenvalues λj having the order j− 1
2 . In fact, the first-named author [25] has
proved the asymptotic formula for the nonincreasingly ordered moduli of eigenval-
ues (see [26] and see also Theorem 1.4). However, the exact asymptotic formula
for the positive and negative eigenvalues ordered separately was previously never
known. One of complications stems from the circumstance that the NP operator is
not self-adjoint in L2.
With all this in mind, the purpose of this paper is to prove the Weyl law for the
asymptotic behavior of positive and negative NP eigenvalues, separately, in three
dimensions. The reasoning is based upon the pseudodifferential representation of
the NP operator and the formulas for the eigenvalue asymptotics for pseudodifferen-
tial operators. These formulas are well known for self-adjoint operators on smooth
manifolds, see [9], and this consideration takes care of the smooth case, since it
turns out that the spectral problem for the non-self-adjoint NP operator can be
reduced to the one for a self-adjoint pseudodifferential operator. As it concerns the
case of a boundary Γ of finite smoothness, exactly, of the class C 2,α, the eigenvalue
formulas of [9], established there for smooth manifolds, cannot be applied directly
and require certain additional perturbation reasoning. We present such reasoning,
thus supporting the calculations in [26].
The coefficients in the asymptotic formulas are expressed in geometrical terms,
involving the principal curvatures and the Gauss curvature of the surface. The most
esthetic is the result for a convex domain, where the coefficient in the Weyl formula
is expressed via the Euler characteristic and the Willmore energy of the surface.
Especially for the case of a non-convex domain, some interesting problems arise
as well. One of important by-products of our eigenvalue calculations is a certain
progress in the long-standing question on the existence of negative eigenvalues of
NEUMANN-POINCARE OPERATOR
3
the NP operator in three dimensions. (There is a kind of inconsistency in the lit-
erature in fixing the sign in the expression for the fundamental solution for the
Laplacian and, consequently, in the integral kernel of the NP operator, therefore the
question on positive and negative eigenvalues arises alternatively. The agreement
we adopt follows, say, [3] or [18], so the fundamental solution is −(4πx)−1. Be-
sides it, we employ the second fundamental form as the inner products of the outer
normal derivative and the second partial derivatives of a regular parametrization
of a surface. So the principal curvatures of a sphere, for instance, are negative).
The existence of at least one negative eigenvalue was established in [3], by means of
explicit computations, for the case of an oblate spheroid. Recently, the existence,
again, of a negative eigenvalue was established in [15] for at least one in a pair of
two surfaces related by inversion, provided there exists at least one point of non-
convexity. Besides it, the NP operator on the standard torus has infinitely many
negative eigenvalues as well as infinitely many positive ones, which was established,
again, by means of explicit eigenvalue calculations, see [16].
Our results on this latter problem are the following. Suppose that the surface
Γ = ∂Ω is infinitely smooth. First, there always exist infinitely many positive eigen-
values. Further on, if there exists a point where at least one of principal curvatures
is positive (so, the body Ω is not convex near this point), there exist infinitely many
negative eigenvalues of the NP operator. On the other hand, if the surface is uni-
formly convex in the sense that the principal curvatures are everywhere negative
(due to smoothness, this implies that they are separated from zero) then there may
exist only finitely many negative eigenvalues. We note again that the results on
the asymptotics of eigenvalues and on the negative eigenvalues are obtained un-
der the condition of infinite smoothness of the surface while the singular numbers
asymptotics is proved for surfaces of the class C 2,α.
1.2. Main results. To state the results in a more precise manner, let Ω be a C 1,α
bounded region in R3 (it is allowed that the boundary Γ = ∂Ω consists of several
connected components.) The NP operator KΓ : L2(Γ) → L2(Γ) is defined by
KΓ[ψ](x) :=
1
4πZΓ
hy − x, n(y)i
x − y3 ψ(y) dSy
(1.1)
where dSy is the surface element and n(y) is the outer normal unit vector to Γ at
the boundary point y. This operator is known to be non-selfadjoint in L2(Γ), unless
each component of Γ is a sphere. However, KΓ is symmetrizable, in other words,
there exists a self-adjoint operator in the Sobolev space H − 1
2 (Γ) equipped with the
norm defined as
kuk− 1
2
= h−SΓu, ui
1
2 ,
(1.2)
where SΓ is the single layer operator on Γ,
SΓ[φ](x) = −
x − y−1φ(y)dSy, x ∈ Γ.
1
4πZΓ
As explained above, we know already that KΓ is a compact operator on L2(Γ) and
the set of its eigenvalues consists of at most countable set of real numbers, with 0 the
only possible limit point. It is also known that the eigenvalues of the NP operator lie
4
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
in the interval (−1/2, 1/2] and 1/2 is the eigenvalue corresponding to the constant
eigenfunction. We denote the set of NP eigenvalues counting multiplicities by
0 (Γ) ≥ ±λ±
σp(KΓ) = { ±λ±
1 (Γ) ≥ . . . },
(1.3)
where ±λ±
j (∂Ω) are positive, resp., moduli of negative, eigenvalues of KΓ; it is a
priori possible that the set of negative eigenvalues is finite or even void. Note that,
by the symmetrization, there are no associated eigenfunctions and therefore the geo-
metric multiplicity of each eigenvalue coincides with its algebraic multiplicity. The
union of the sets of positive and of moduli of negative eigenvalues, again, numbered
in the non-increasing order, form the sequence
sp(KΓ) = {
1
2
= λ0(Γ) ≥ λ1(Γ) ≥ . . . }.
(1.4)
This sequence coincides with the non-increasingly ordered set of the singular num-
bers (s-numbers) of the operator KΓ considered as an operator in the Sobolev space
H −1/2(Γ) with the above norm (we note here that when changing the norm in a
Hilbert space to an equivalent one, the s-numbers of an operator may change, un-
like the eigenvalues). The set of s-numbers of the operator KΓ in L2(Γ) is, generally,
different from sp(KΓ) and it is denoted by
µµµ(KΓ) = { µ0(Γ) ≥ µ1(Γ) ≥ . . . }.
(1.5)
As long as this does not cause confusion, we will omit the specification Γ in the
notations of operators, spaces, eigenvalues and singular values.
The counting functions for the sequences (1.3), (1.4), (1.5) are denoted by n±(λ) =
#{j : ±λ±
j > λ}, resp., n(λ), m(λ). Of course, n(λ) = n+(λ) + n−(λ).
Theorem 1.1. Let Ω be a bounded domain with C ∞ boundary. Then
λ±
j ∼ A±(Γ)
1
2 j− 1
2 , j → ∞,
(1.6)
where
while
A±(Γ) =
1
128π2Z
Γ
dSxZ 2π
0
[(k1(x) cos2 θ + k2(x) sin2 θ)∓]2dθ,
(1.7)
A+(Γ) + A−(Γ) = A(Γ),
(1.8)
the latter given in (1.11). Here k1(x), k2(x) are the principal curvatures of the
surface Γ at the point x with the direction of the normal vector being chosen to be
the exterior one; dSx is the surface element of Γ at x. If the coefficient A+(Γ) or
A−(Γ) turns out to be zero, formula (1.6) should be understood as λ±
2 ).
j = o(j− 1
We say that the surface is almost convex at a point x if k1(x), k2(x) are non-
positive (of course, if the surface is convex near a point, it is almost convex there).
By (1.7), for a surface which is almost convex at each point, A+(Γ) = A(Γ) and
A−(Γ) = 0. Further on, the surface is called strictly convex at x if, moreover,
the principal curvatures of the surface at x are negative. For a body with smooth
boundary in R3 there necessarily must exist a region in Γ where the surface is strictly
convex. Therefore, the coefficient A+(Γ) may never vanish.
Corollary 1.2. Under the conditions of Theorem 1.1, there exist infinitely many
positive eigenvalues of the NP operator.
NEUMANN-POINCARE OPERATOR
5
On the other hand, there may exist an open subset in Γ where the surface is not
almost convex, in other words, where at least one of principal curvatures is positive.
In this case, the coefficient A−(Γ) is nonzero. By continuity, this happens even if
there is just one point on the boundary where the latter is not convex.
Corollary 1.3. If the surface Γ is not almost convex at one point, at least, there
exist infinitely many negative eigenvalues of the NP operator.
Further on, we can express the asymptotic formulas for the NP eigenvalues (and
their counting functions) using some notions of surface geometry. To do this, we
recall the definition of the Willmore energy W (Γ):
W (Γ) :=ZΓ
H 2(x) dSx,
(1.9)
where H(x) is the mean curvature of the surface at x. If Γ consists of several con-
nected components, W (Γ) is the sum of the Willmore energies of these components.
Theorem 1.4. Let Ω be a C ∞ bounded region. Then
µj(KΓ) ∼ λj(KΓ) ∼ A(Γ)
1
2 j−1/2
as j → ∞,
where
A(Γ) =
3W (Γ) − 2πχ(Γ)
128π
.
(1.10)
(1.11)
Here W (Γ) and χ(Γ) denote, respectively, the Willmore energy and the Euler char-
acteristic of the surface Γ.
Thus, the NP operator has always the infinite rank (such question was discussed in
[18]) and the decay rate of NP eigenvalues is j−1/2 for smooth regions. Furthermore,
the integral (1.9) is especially interesting because it has the remarkable property of
being invariant under Mobius transformations of R3, see [11]. Thus we find that
the asymptotic behavior of moduli of NP eigenvalues and NP singular numbers is
also Mobius invariant since the Euler characteristic is topologically invariant. We
will present some further facts and applications later on (see section 5). Theorem
1.4 holds true even for C 2,α surfaces [26]. This means that a kind of spectral cut-off
happens: the eigenvalues of the NP operator may never have a faster decay rate
than j− 1
2 and thus the operator may not ever belong to a smaller Schatten class
than Σ2.
Remark 1.5. Formula (1.10) can be written in the equivalent form, using the count-
ing functions:
n(λ) ∼ A(Γ)λ−2, n±(λ) ∼ A±(Γ)λ−2, m(λ) = n+(λ) + n−(λ) ∼ A(Γ)λ−2.
(1.12)
The above theorems are proved by means of finding the pseudodifferential oper-
ator representation for the NP operator or for a certain approximation of KΓ and
further application of classical results on the asymptotics of eigenvalues or s-numbers
of negative order pseudodifferential operators. Some further properties of these op-
erators enable us to obtain sufficient conditions for the finiteness and for infiniteness
of the set of negative eigenvalues.
6
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
Theorem 1.6. Let the body Ω with smooth boundary ∂Ω be strictly convex at all
points. Then there may exist only finitely many negative eigenvalues of the NP
operator.
To illustrate the meaning of the general results, let us consider the case ∂Ω = S2.
It has been proved already by Poincar´e [29] that the NP eigenvalues on the two-
dimensional sphere are
2(2k+1) for k = 0, 1, 2 . . . and their multiplicities are 2k + 1.
1
It easily follows that the j = k2-th eigenvalue satisfies
λj(KS 2) =
1
2(2k + 1)
∼
1
4
j−1/2.
On the other hand, by Theorem 1.4, λj(KS 2) ∼ 1
4 j−1/2 since W (S2) = 4π and
χ(S2) = 2. This calculation, is, of course, consistent with the above asymptotics of
the explicit eigenvalues. Moreover, due to the convexity of the sphere, by Theorem
1.6, there may exist only a finite number of negative eigenvalues - in this example
there are no such eigenvalues at all, so λ+
4j−1/2. In the example of an
oblate spheroid, it was found in [3] that there exists at least one negative eigenvalue.
Our result complements it by stating that there may exist only a finite set of such
eigenvalues.
j (KS 2) ∼ 1
Another illustration concerns the surface ∂Ω being diffeomorphic to a torus in
R3. Such surface, obviously, contains points where it is not convex. Therefore,
by Corollary 1.3, the NP operator possesses infinitely many negative eigenvalues.
For the standard metric torus, this property was recently established in [16] by
means of an explicit calculation. More generally, for any surface which is not simply
connected, there are infinitely many negative eigenvalues.
Finally, suppose that the body Ω has some holes inside, so that the surface Γ
In this case, the interior part of the
consists of several connected components.
boundary possesses necessarily a fragment where the surface is not almost convex.
Therefore, for the NP operator the set of negative eigenvalues is infinite.
It is worth comparing our decay rate results for the three dimensional NP eigen-
values, obtained here, with those for the two-dimensional NP eigenvalues. There,
in the latter case, it is well known that the eigenvalues of the integral operator KΓ
are symmetric with respect to the origin. The only exception is the eigenvalue 1/2
corresponding to the constant eigenfunction. NP eigenvalues are invariant under
Mobius transformations [32]. One of the main distinguished features here is that
the decay rate depends essentially on the smoothness of the boundary. Indeed, it is
proved in [7, 25] that if the boundary is C k smooth then for any τ > −k + 3/2,
λ±
j (KΓ) = o(jτ ) as j → ∞.
Moreover for an analytic boundary, the eigenvalues have at least the exponential
decay rate:
λ±
j (KΓ) ≤ Ce−jǫ
as j → ∞,
Here ǫ is the modified Grauert radius of ∂Ω (see [7] for the precise statement). For a
piecewise analytic smooth Γ the second-named author has established recently the
estimate λ±
2 ). All these results contain upper estimates for eigenvalues. It
is a challenge: with exception for Mobius transformed ellipses, there exists presently
not a single example of curves where the asymptotics of NP eigenvalues is known. In
j = O(e−cj
1
NEUMANN-POINCARE OPERATOR
7
particular, it is unknown whether there exist curves for which the eigenvalue decay
is super-exponential, in other words, the question on the spectral cut-off property is
still open here.
The paper is organized as follows. In the next section we establish some important
relationships between singular- and eigenvalues using Ky-Fan's theorem and the
Plemelj's symmetrization principle for NP operators. In section 3, we introduce the
approximate pseudo-differential operators for NP operators. Further on, in section
4 we show that 4 the relations established in section 3 yield the Weyl law for NP
eigenvalues. Some applications and a discussion are provided in section 5.
V. Maz'ya made an essential contribution to the field of boundary integral equa-
tions. His paper [23] had a great influence in the topic. We are happy to be able to
contribute to the special volume dedicated to his jubilee and wish him many more
years of productive activity.
2. Preliminaries on Schatten classes and perturbations
As preliminaries, we shall recall some results on Schatten classes, used in this
paper.
Let K be a compact operator in a separable Hilbert space H. The singular values
(s-numbers) {sj(K)} are the eigenvalues of (K ∗K)1/2 numbered in the nonincreasing
order, counting multiplicities.
σsing(K) = { sj(K) s1(K) ≥ s2(K) ≥ s3(K) ≥ · · · }.
(2.1)
The algebra of operators satisfying kKkp
j=1 sj(K)p < ∞ is
called the Schatten class Sp. It is known that for K ∈ Sp, the singular values satisfy
Sp = tr(K ∗K)p/2 = P∞
sj(K) = o(j−1/p).
(2.2)
The class of operators K which satisfy (2.2) is called small weak Schatten class
Σ0
p, and the operators satisfying (2.2) with o(j−1/p) replaced by O(j−1/p), form the
weak Schatten class Σp (see, e.g., [33]). Thus, Sp ⊂ Σ0
p ⊂ Σp. In particular, the
Schatten class S2 carries the individual name Hilbert-Schmidt class. An integral
operator K with kernel K(x, y) acting in L2(M) for some measure space M, belongs
2, Σ2 there exist no
exact conditions for this kind of inclusion. An important subset in Σ2 consists of
operators for which the asymptotic relation
to S2 if and only if R RM ×M K(x, y)2dxdy < ∞; for classes Σ0
sj(K) ∼ Cj− 1
2
(2.3)
holds. The following statement (by Ky Fan, [19]) shows that the property (2.3) is
stable under perturbations of K, belonging to the class Σ0
2.
Proposition 2.1. If for a compact operator K the asymptotics (2.3) holds and
R ∈ Σ0
2 is valid
with the same constant C.
2 then for the operator K ′ = K + R, the asymptotics sj(K ′) ∼ Cj− 1
As well known, starting from H. Weyl, for self-adjoint compact operators, a similar
stability of the asymptotic law for separately positive and negative eigenvalues under
weaker perturbations takes place as well.
8
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
2 and R ∈ Σ0
2 is valid with the same constants C±.
Proposition 2.2. If for compact self-adjoint operators K, R there is the asymptotics
j (K) ∼ C±j− 1
λ±
2 then for the operator K ′ = K + R, the asymptotics
j (K ′) ∼ C±j− 1
λ±
For non-selfadjoint operators, generally, (2.3) does not imply that the eigenvalues
or their absolute values follow a similar asymptotic law. The statement to follow
enables us, however, to pass from the asymptotics of s-numbers to the asymptotics
of (moduli of) eigenvalues under some additional conditions.
Proposition 2.3. Let K be a symmetrizable compact operator, namely, there exists
an invertible operator S such that S −1KS is self-adjoint. Assume
(1) S −1KS − K ∈ Σ0
2;
(2) sj(K) ∼ Cj−1/2
as j → ∞.
Then λj(K) ∼ sj(K) ∼ Cj−1/2
as j → ∞, where, recall, λj(K) are the moduli of
eigenvalues of K numbered in the non-increasing order counting algebraic multiplic-
ities.
Proof. We notice that the operator S −1KS has the same eigenvalues of K. Since
the singular values of self-adjoint operator are the absolute values of eigenvalues,
the result follows from Proposition 2.1.
(cid:3)
3. Layer potentials as pseudo-differential operators
The single layer potential operator −SΓ is positive, invertible, self-adjoint, and it
satisfies Plemelj's symmetrization principle (also known as the Calder´on's identity):
SΓK∗
Γ = KΓSΓ.
Thus we can symmetrize the NP operator via single layer potentials, namely,
KKKΓ = (−SΓ)−1/2KΓ(−SΓ)1/2
(3.1)
is self-adjoint in L2(Γ). Our aim in this section is to find a pseudo-differential repre-
sentation of the symmetrized NP operator KKKΓ (3.1). To this purpose, we represent
the single layer potential SΓ and the double layer potential KΓ as PsDO. Thanks to
the smoothness condition, the asymptotic expansion of the integral kernel in terms,
homogeneous in x − y, yields a PsDO for any local chart on the boundary manifold.
The calculations in [26] show that the principal symbol of the NP operator, of order
−1, equals in local co-ordinates x′, ξ
pΓ(x′, ξ) = −
(3.2)
L(x′)ξ2
2 − 2M(x′)ξ1ξ2 + N(x′)ξ2
1
4 det(gjk)(cid:16)Pj,k gjk(x′)ξjξk(cid:17)3/2
(note especially here the minus sign in front of the fraction.) Here gjk(x′) denotes the
metric tensor and L(x′), M(x′), N(x′) are the coefficients of the second fundamental
form on the boundary Γ in the local co-ordinates x′ in some domain in R2 with,
recall (this is highly important), the normal vector directed to the exterior of Ω.
The single layer potential SΓ is also PsDO of order −1 with principal symbol
(−4πξx′)−1. Here ξx′ =qPj,k gjk(x′)ξjξk on each local chart [1]. It is known that
the operator −SΓ is positive and invertible, it is elliptic, and therefore the complex
powers of −SΓ are pseudodifferential operators. The operator (−SΓ)z has principal
NEUMANN-POINCARE OPERATOR
9
symbol (4πξx)z. Therefore the principal symbol of KKK equals the product of principal
symbols of K, S and S −1, and thus, the principal symbol of the symmetrized NP
operator coincides with the one of the original NP operator. In this way, we obtained:
Theorem 3.1. Let Ω ⊂ R3 be a bounded smooth region. Then the symmetrized NP
operator KKKΓ is an order −1 pseudodifferential operator on Γ with principal symbol
pΓ(x′, ξ) given by (3.2).
For the boundary of finite smoothness C 2,α, the above reasoning is not valid since
a convenient symbolic calculus for operators on such manifolds is not developed. We
repeat, for further reference, the local approximation result from [26] in a convenient
form:
Proposition 3.2. Let Ω ⊂ R3 be a bounded domain with C 2,α boundary. Let V ⊂ Γ
be a coordinate patch so that π : V → R2 is a co-ordinate mapping generating the
isometry π of L2(V ) to L2(U), U = πV ⊂ R2. If ψ1, ψ2 are bounded functions with
support in U then the projected operator πψ1KΓψ2π−1 is the sum of a pseudodiffer-
ential operator in U with symbol
h(x′, ξ) = hψ1,ψ2(x′, ξ) = πψ1(x′)πψ2(x′)pΓ(x′, ξ)
(3.3)
and a Hilbert-Schmidt class operator, with pΓ given in (3.2). This symbol is smooth
in ξ variable and belongs to C α in x′.
4. Local considerations and delocalization
Our aim now is to obtain the eigenvalue asymptotics for the operator KKKΓ. For a
smooth surface Γ = ∂Ω, this would follow from Theorem 3.1 and the basic result
by Birman-Solomyak, [9], on the asymptotics of singular numbers and eigenvalues
for negative order pseudodifferential operators. For C 2,α surfaces, a perturbation
approach was used in [26], with a representation in local charts of the NP operator as
a sum of a pseudodifferential operator and a Hilbert-Schmidt one, with again further
using Birman-Solomyak's result to obtain the asymptotics of moduli of eigenvalues.
However, the analysis of the paper [9] shows that some more explanations are needed
when applying the results of this paper to our setting.
In the original paper [9], only the case of a pseudodifferential operator in a domain
of the Euclidean space, with moderately regular symbol, was considered in detail.
Just a brief remark was included that the results, for a homogeneous symbol, carry
over 'easily' to operators on smooth manifolds. Somehow, this statement migrated to
later publications by various authors, even applied to manifolds with finite smooth-
ness. It turns out, however, that certain complications arise in this generalization,
even for the smooth case. The authors of [9] were quite aware of these complications.
In particular, when applying in [10] the results of [9] to the study of the asymptotics
of scattering phases, they presented the corresponding reasoning, which turned out
to be rather involved, requiring some additional technicalities. The problem consists
just in the passage from the eigenvalue asymptotics of operators in local charts to
the whole manifold. We call this stage 'delocalization'. In this section, we describe
such delocalization, including the case of a finite smoothness, considered in [26].
In our setting, even for a finite smoothness, the 'delocalization' is somewhat easier
than in [10], since in our very special particular case of the NP operator on a two-
dimensional manifold, the most simple Hilbert-Schmidt estimates are sufficient. In
10
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
other dimensions our reasoning is also possible, however it becomes somewhat more
technical, due to the absence of sharp analytical criteria for an integral operator to
belong to Schatten classes other than S2. We consider the case of a surface of the
class C 2,α; the case of an infinitely smooth surface is more simple.
We reproduce here Theorem 2 from [9], adapted to our particular case.
Theorem 4.1. Let T be a pseudodifferential operator of order −1 with homogeneous
symbol a(x′, ξ) in a bounded domain U ⊂ R2, smooth in the variable ξ and belonging
to C α, α > 0, in x′ variable. Let also b(x′), c(x′) be bounded weight functions sup-
ported in U . Then for the operator L = bT c, the asymptotic formula for singular
numbers holds:
sj(L)j
1
2 = C
1
2 (1 + o(1)), C =
(b(x′)c(x′)a(x′, ω))2dx′dω,
(4.1)
and, for a self-adjoint operator T , with b(x′) = ¯c(x′), the formula for eigenvalues
holds:
λ±
j j
1
2 = C
1
2
±(1 + o(1)), C± =
(b(x′)2a(x′, ω)±)2dx′dω.
(4.2)
1
8π2ZUZS 1
1
8π2ZUZS 1
S Γm = Γ. We denote by Γs
On the surface Γ, we consider a finite system of disjoint open subsets Γm so that
m the 'star' of Γm, namely, the union of Γm and those
sets Γm′ for which Γm ∩ Γm′ 6= ∅. We can suppose that this decomposition is so fine
that, for some choice of points xm ∈ Γm, the orthogonal projection πm of Γm to the
tangent plane Tm at xm is a homeomorphism, moreover, it is a homeomorphism of
m. We denote by Um the range of Γs
Γs
m). The smoothness
conditions imposed on Γ imply that π−1
m , considered as a mapping from Um to
R3, is of the class C 2,α. Further on, the mapping πm generates in a usual way
m), which are also isometries
m under πm, Um = πm(Γs
isometries of Hilbert spaces fπm : L2(Um) → L2(Γs
fπm : L2(πm(Γm)) → L2(Γm).
represented as
Let χm be the characteristic function of the set Γm. The NP operator K can be
Km,m′, Km,m′ = χmKχm′.
(4.3)
K = Xm,m′
Our first statement about the singular numbers asymptotics is the following.
Proposition 4.2. For any m, as j → ∞,
sj(Km,m)j
1
2 = C
1
2
m(1 + o(1)), Cm =
h(x′, θ)2dx′dθ,
(4.4)
1
8π2ZπmΓmZ 2π
0
where h(x′, ξ) is given by (3.3) with ψ1 = ψ2 = χm.
Proof. By means of the isometry fπm of Hilbert spaces L2(Γm) and L2(Um), the
operator Km,m turns out to be unitary equivalent to the pseudodifferential operator
Pm with symbol determined by (3.2) and weight functions b = c = πmχm(x′) plus a
Hilbert-Schmidt operator. Further on, the symbol of the pseudodifferential operator
Pm, is smooth in the ξ variable and belongs to C α in x′ variable. Therefore, to
this operator we can apply Theorem 4.1 establishing the asymptotics of singular
numbers of a pseudodifferential operator in a domain in the Euclidean space, which
is given by formulas (4.1). The addition of a Hilbert-Schmidt operator, by the Weyl
NEUMANN-POINCARE OPERATOR
11
inequality and Proposition 2.3 does not change the leading term in these asymptotic
formulas.
(cid:3)
The asymptotics, just found, is responsible for the diagonal terms m = m′ in the
representation (4.3). The proposition to follow proves that the non-diagonal terms
in (4.3) do not affect these asymptotics.
Proposition 4.3. Let m 6= m′. Then for the operator Km,m′ = χmKχm′ ,
sj(Km,m′) = o(j− 1
2 ).
(4.5)
Proof. We consider two cases: Γm ∩ Γm′
6= ∅ and Γm ∩ Γm′ = ∅.
In the first case, again, the operator Km,m′ is unitary equivalent to the sum of a
Hilbert-Schmidt operator and a pseudodifferential operator in L2(ΓmS Γm′) having
the form
PΓm,Γm
′ u(x′) = (2π)−2χm(x′)ZR2
Z
P (x′, ξ)e(x′−y′)ξχm′(y′)dy′dξ.
(4.6)
Γm S Γm
′
By Theorem 4.1, the singular numbers of the operator (4.6) have asymptotics
Cm,m′j− 1
2 with coefficient Cm,m′ determined by the integration of the symbol and the
weight functions χm, χ′
the product χm(x′)χ′
Therefore, the coefficient in front of j− 1
numbers of the operator Km,m′ equals zero. This means that sj(Km,m′) = o(j− 1
just what we were aiming to.
m(x′) equals zero almost everywhere and the integral annules.
2 in the asymptotic formula for the singular
2 ),
m over πm(ΓmS Γm′) × S1. However, in this particular case,
In the second case, the distance between the compact sets Γm and Γm′ is positive,
therefore the integral kernel of the NP operator is bounded on Γm×Γm′. This implies,
in particular, that this kernel is square integrable over Γm × Γm′, and therefore the
operator Km,m′ belongs to the Hilbert-Schmidt class. As we explained before, this
means that sj(Km,m′) = o(j− 1
(cid:3)
2 ).
Now we collect the local parts of K to obtain the spectral asymptotics.
Theorem 4.4. Let Γ be a C 2,α. Then the following asymptotic formula is valid:
σj(K) ∼ A(Γ)
1
2 j− 1
2 , A(Γ) =
pΓ(x, ω)2 dSxdω
(4.7)
1
8π2ZS ∗Γ
Proof. We use the block-matrix representation of the operator K with respect to the
system of orthogonal subspaces Hm = L2(Γm) :
K = Xm,m′
χmKχm′ =X Km,m′.
(4.8)
The terms on the diagonal, m = m′ in (4.8) act in orthogonal subspaces, therefore
the set of s-numbers of Pm Km,m is the union of such sets of Km,m, and thus the
distribution function of its s-numbers equals the sum of distribution functions for
Km,m. Therefore, the asymptotic coefficients in these formulas, found in the previous
lemmas, should be added up, which produces the coefficient in (4.7). On the other
hand, off-diagonal terms in (4.8) have singular numbers decaying faster than j− 1
and therefore, by Propositions 2.3, 2.1 they do not contribute to the leading term
2
12
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
of the asymptotics. This reasoning takes care of the asymptotics for the singular
numbers.
(cid:3)
This theorem justifies the reasoning in [26] concerning the reference to the result
by Birman-Solomyak.
For the case of smooth boundary, a similar reasoning should be repeated as applied
to the self-adjoint operator KKK. The passage from the local asymptotic formula (4.2)
to the global one goes as above, just by using Proposition 2.2 and the pseudolocality
property of classical pseudodifferential operators.
Theorem 4.5. Let Γ be a smooth boundary of Ω. Then the following asymptotic
formula is valid
λ±
j (K) ∼ A±(Γ)
1
2 j− 1
2 , A±(Γ) =
pΓ(x′, ω)2
± dx′dω.
(4.9)
1
8π2ZS ∗Γ
To finish the proof of Theorem 1.1 and Theorem 1.4, we calculate the positive
constants A(Γ), A±(Γ). To achieve this, we may use the isothermal charts, so that
2). Then the surface
1 + dx2
element is dS = E(x′)dx′ and, after summing the local contributions, we obtain
1
A±(Γ) =
2 − 2M(x)ξ1ξ2 + N(x)ξ2
1
locally the metric has the formPi,j gijdxidxj = E(x′)(dx2
4 det(gij){qPj,k gjk(x)ξjξk}3i2
8π2Z∂ΩZS 1h L(x)ξ2
8π2Z∂ΩZS 1h L(x) cos2 θ − 2M(x) cos θ sin θ + N(x) sin2 θ
128π2Z∂ΩZS 1
4E2(x)E−3/2(x)
=
=
E2(x)
dξdx′
1
1
∓
[(L(x) cos2 θ − 2M(x) cos θ sin θ + N(x) sin2 θ)∓]2
i2
∓
dθdx′
dθdSx.
For fixed x one can diagonalize the last equation by an orthogonal matrix and so
A±(Γ) =
1
128π2Z∂ΩZS 1
[(k1(x) cos2 θ + k2(x) sin2 θ)∓]2 dθdSx.
This gives us the required formula for A±(Γ) as a coordinate free representation, For
the sum A(Γ) = A+(Γ) + A−(G), we can add up the above expressions for A±(Γ) to
obtain
4 N 2(x) + πM 2(x) + π
2 L(x)N(x))
dSx
E2(x)
4 N 2(x) + π(L(x)N(x) − E2(x)K(x)) + π
2 L(x)N(x))
dSx
A(G) =
=
=
=
=
1
1
( 3π
( 3π
4 L2(x) + 3π
4 L2(x) + 3π
128π2Z∂Ω
128π2Z∂Ω
512πZ∂Ωh(cid:18) L(x) + N(x)
512πZ∂Ω
4H 2(x) dSx −
E(x)
3
3
1
64
3W (∂Ω) − 2πχ(∂Ω)
128π
.
(cid:19)2
−
E2(x)
4
3
K(x)i dSx
χ(∂Ω)
Here we used the Gauss-Bonnet theorem in the integration of Gaussian curvature.
NEUMANN-POINCARE OPERATOR
13
5. Applications and remarks
5.1. Plasmonic eigenvalues. The interest in the NP operator, especially in its
spectral properties, has been growing rapidly recently, due to its connection to
the plasmon resonance and the anomalous localized resonance in meta materials
possessing negative material characteristics, for example, dielectric constants. These
resonances occur at eigenvalues and at the accumulation points of eigenvalues of the
NP operator, respectively (see [4, 20] and references therein). The spectral nature
of the NP operator is also related to stress concentration between hard inclusions
[12].
As an application of our results, let us consider plasmonic eigenvalues (see e.g.
[14] and references therein). A real number ǫ is called a plasmonic eigenvalue if the
following problem admits a solution u in the space H 1(R3):
∆u = 0
u− = u+
ǫ∂nu− = −∂nu+
in R3\∂Ω,
on ∂Ω,
on ∂Ω.
(5.1)
where the subscript ± on the left-hand side respectively denotes the limit (to ∂Ω)
from the outside and inside of Ω. The well-known relation [8] between the plasmonic
eigenvalue ǫ and the NP eigenvalue λ gives
−2λj
ǫj − 1 =
λj − 1/2
∼ 4A±j−1/2.
(5.2)
Hence the plasmonic eigenvalues constitute the sequence with 1 as the limit, and
the (R.H.S.) of (5.2) gives its converging rate. Positive and negative NP eigenvalues
correspond to left limits and right limits respectively.
5.2. (In)finiteness of the set of negative eigenvalues. We consider now the
question on negative eigenvalues of the NP operator, mentioned in the Introduction.
We give here the proof of Theorem 1.6.
So, we suppose here that the surface Γ is infinitely smooth and convex. The latter
condition means that the principal symbol k(x, ξ) of the NP operator K, considered
as a pseudodifferential operator calculated in Section 3, is positive for all x ∈ ∂Ω
and ξ : ξ = 1, therefore, by the compactness of the cospheric bundle,
k(x, ξ) ≥ Cξ−1
(5.3)
Therefore, the principal symbol of the self-adjoint pseudodifferential operator KKK =
S − 1
2 is the same, k(x, ξ). It follows that the symbol
2 KS
1
l(x, ξ) = (k(x, ξ))− 1
2
(5.4)
T ∗(∂Ω) as
is well defined as a smooth positive function on the cotangent bundle of
a positive function, degree − 1
2 positively homogeneous.
Now we consider some pseudodifferential operator L on ∂Ω with principal symbol
l. This operator can be constructed in the usual way, by means of gluing together
local operators with this symbol, defined by means of the Fourier transform. We
denote by R the operator L∗L, where the adjoint operator is considered in the sense
of the space H − 1
2 (Γ) with norm as in (1.2). The operator R, thus constructed,
is a nonnegative operator in H − 1
2 , moreover, it is an elliptic operator of order 1.
14
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
Therefore, the zero subspace of R has finite dimension. If this subspace is nontrivial,
we add to R the orthogonal, finite rank, projection onto this subspace, thus changing
R by a smoothing operator. After this operation, the operator, which we still
denote by R, becomes a first order positive elliptic operator, still, with principal
symbol l(x, ξ)2 = k(x, ξ)−1. By construction, the operator R is invertible and the
principal symbol of the positive operator K = R−1 equals, again, k(x, ξ). So, since
the operators KKK and K have the same principal symbol, their difference, M =
K − K is a self-adjoint pseudodifferential operator of lower order, no greater than
−2. Therefore, M has the form
M = KZK,
with an operator Z = K−1MK−1, a zero order pseudodifferential operator, bounded
in all Sobolev spaces H s(∂Ω).
Now, consider, for some t > 0 the subspace Lt spanned by all eigenfunctions of
K with eigenvalues smaller than t. This subspace has finite codimension, and for
u ∈ Lt, we have kKuk2 ≤ t(Ku, u) (by the Spectral Theorem.) Therefore, (all
scalar products and norms are in the sense of H − 1
2 (Γ)) (Mu, u) = (KZKu, u) =
(ZKu, Ku) ≤ kZkkKuk2, u ∈ Lt.
So we have
(KKKu, u) = (Ku, u) + (Mu, u) ≤ (Ku, u) − tkZk(Ku, u) = (1 − tkZk)(Ku, u).
We choose t so small that 1 − tkZk > 1
2. This choice gives us
1
2
(Ku, u) ≥ 0.
(KKKu, u) ≥
(5.5)
In this way, we have found a subspace with finite codimension on which the quadratic
form of the operator KKK is nonnegative. Again, by the Spectral Theorem, this means
that the operator KKK, and together with it, the operator K, may have only a finite
number of positive eigenvalues.
On the other hand, we consider a surface which is not almost convex. This means
that somewhere at the surface, the integrand in A−(Γ) in (1.7) is positive, thus
A−(Γ) > 0. This means that for the negative eigenvalues there exist the power
asymptotics with a nonvanishing coefficients, which implies their infiniteness. As
explained in the Introduction, the set of positive eigenvalues is always infinite.
5.3. Mobius (non)invariance. As described in the previous section, the asymp-
totic behavior of absolute values of NP eigenvalues is related closely with the Will-
more energy and the Euler characteristics. Some applications in this direction can
be found in [26]. The separate behavior of positive and negative eigenvalues involves
more detailed structures. For instance, the asymptotic of the absolute value of NP
eigenvalues is invariant under Mobius transforms as is the case with two-dimensional
NP operators. However, Mobius transforms of ellipsoids are gourd-shaped surfaces
(they may have negative curvature points). So we have infinite many negative eigen-
values on the gourd-shaped surface and its asymptotic of negative eigenvalues varies
from an ellipsoid. As a result, the asymptotics of NP eigenvalues may vary under
Mobius transforms, while Weyl's asymptotics of moduli of eigenvalues is invariant.
NEUMANN-POINCARE OPERATOR
15
5.4. Remarks. We should note that the results on the asymptotics of eigenvalues
and on the negative eigenvalues are obtained under the condition of infinite smooth-
ness of the surface while the singular numbers asymptotics is proved for surfaces of
the class C 2,α. We are convinced that these results can be extended to less smooth
surfaces, the ones where the asymptotic coefficients still make sense, thus to surfaces
of the class C 1,1, i.e., those which are described locally by once differentiable func-
tions with derivatives satisfying the Lipschitz condition. The natural way to pursue
this aim is to study the approximation of non-smooth surfaces by smooth ones, in
the flavor of [31].
General results on the eigenvalue asymptotics for a smooth surface can be easily
extended to the multi-dimensional case, without considerable complications. What
may cause certain trouble is finding an expression for the asymptotic coefficients
in geometrical terms. However, for the case of finite smoothness, the possibility
of such extension is presently unclear, in particular, since the reasoning involving
Hilbert-Schmidt operators should be replaced by some other Schatten classes and
presently known conditions for an integral operator to belong to such classes might
be not sufficient for our aims.
By our opinion, the results on the eigenvalue asymptotics can also be extended
to the NP operator defined in a proper way with the fundamental solution for the
Laplacian being replaced by the fundamental solution for some general second order
elliptic operator with smooth coefficients.
The authors mean to pursue these topics in the future.
References
[1] M.S. Agranovich, B.Z. Katsenelenbaum, A.N. Sivov and N.N. Voitovich, Generalized Method
of Eigenoscillations in Diffraction Theory, Wiley-VCH, Berlin, 1999.
[2] L. V. Ahlfors, Remarks on the Neumann-Poincar´e integral equation, Pacific. J. Math., 2
(1952), 271 -- 280.
[3] J. F. Ahner, On the eigenvalues of the electrostatic integral operator II, J. Math. Anal. Appl.,
181 (1994), 328 -- 334.
[4] H. Ammari, G. Ciaolo, H. Kang, H. Lee and G. W. Milton. Spectral theory of a Neumann-
Poincar´e-type operator and analysis of cloaking due to anomalous localized resonance, Arch.
Ration. Mech. An. 208 (2013), 667 -- 692.
[5] H. Ammari, H. Kang, and H. Lee, Layer potential techniques in spectral analysis, Mathemat-
ical Surveys and Monographs, 153 American Math. Soc., Providence RI, (2009).
[6] K. Ando and H. Kang, Analysis of plasmon resonance on smooth domains using spectral
properties of the Neumann -- Poincar´e operator, J. Math. Anal. Appl. 435(1) (2016), 162 -- 178.
[7] K. Ando, H. Kang, Y. Miyanishi, Exponential decay estimates of the eigenvalues for the
Neumann -- Poincar´e operator on analytic boundaries in two dimensions, To appear in J. Inte-
gral Equations, arXiv:1606.01483.
[8] K. Ando, H. Kang, Y. Miyanishi and E. Ushikoshi, The first Hadamard variation of Neumann --
Poincar´e eigenvalues, To appear in Proc. Amer. Math. Soc., arXiv:1805.02414.
[9] M. Birman and M. Solomyak, Asymptotic behavior of the spectrum of pseudodifferential
operators with anisotropically homogeneous symbols, Vestnik Leningrad Univ. 13 (1977),
13 -- 21; English translation in Vestin. Leningr. Univ. Math. 10, 237 -- 247.
[10] M. Birman, D. Yafaev, Asymptotic behavior of the spectrum of the scattering matrix. (Rus-
sian) Boundary value problems of mathematical physics and related questions in the theory of
functions, 13. Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 110 (1981),
3-29. English translation in: J. Sov. Math. 25 (1984), 793-814.
[11] W. Blaschke, Voresungen Uber Differentialgeometrie III, Berlin: Springer (1929)
16
YOSHIHISA MIYANISHI AND GRIGORI ROZENBLUM
[12] E. Bonnetier and F. Triki, On the spectrum of Poincar´e variational problem for two close-to-
touching inclusions in 2D, Arch. Ration. Mech. An. 209 (2013), 541 -- 567.
[13] T. Carleman, Uber das Neumann-Poincar´esche problem fur ein gebiet mit ecken, Almquist
and Wiksells, Uppsala, 1916.
[14] D. Grieser, The plasmonic eigenvalue problem, Rev. Math. Phys. 26 (2014), 1450005.
[15] Y. Ji, H.Kang, A concavity condition for existence of a negative Neumann-Poincar´e eigenvalue
in three dimensions. arXiv:1808.10621
[16] K. Ando, Y. Ji, H. Kang, D. Kawagoe, and Y. Miyanishi Spectral structure of the Neumann --
Poincar´e operator on tori, arXiv:1810.09693
[17] H. Kang, M. Lim and S. Yu, Spectral resolution of the Neumann-Poincar´e operator on inter-
secting disks and analysis of plasmon resonance, arXiv:1501.02952, 2015.
[18] D. Khavinson, M. Putinar, and H.S. Shapiro, Poincar´e's variational problem in potential
theory. Arch. Ration. Mech. Anal. 185(1) (2007), 143 -- 184.
[19] Ky Fan, Maximum properties and inequalities for the eigenvalues of completely continuous
operators. Proc. Nat. Acad. Sci., U. S. A. 37, (1951), 760 -- 766.
[20] I. D. Mayergoyz, D. R. Fredkin and Z. Zhang, Electrostatic (plasmon) resonances in nanopar-
ticles, Phys. Rev. B, 72 (2005), 155412.
[21] F. C. Marques and A. Neves, Min-Max theory and the Willmore conjecture, Anal. Math. 179
(2014), 683 -- 782.
[22] E. Martensen, A spectral property of the electrostatic integral operator, J. Math. Anal. Appl.,
238 (1999), 551 -- 557.
[23] V. Maz'ya. Boundary Integral Equations, Itogi Nauki i Techniki, Fundamental Directions, vol.
27, Analysis-4, Viniti, 1988, pp. 131 -- 228.
[24] G.W. Milton and N.-A.P. Nicorovici, On the cloaking effects associated with anomalous lo-
calized resonance, Proc. R. Soc. A 462 (2006), 3027 -- 3059.
[25] Y. Miyanishi and T. Suzuki, Eigenvalues and eigenfunctions of double layer potentials, Trans.
Amer. Math. 369 (2017), 8037 -- 8059
[26] Y. Miyanishi, Weyl's law for the eigenvalues of the Neumann -- Poincar´e operators in three
dimensions: Willmore energy and surface geometry, arXiv:1806.03657
[27] C. Neumann, Uber die Methode des arithmetischen Mittels, Erste and zweite Abhandlung,
Leipzig 1887/88, in Abh. d. Kgl. Sachs Ges. d. Wiss., IX and XIII.
[28] K. Perfekt and M. Putinar, The essential spectrum of the Neumann -- Poincar´e operator on a
domain with corners, Arch. Ration. Mech. Anal. 223 (2017), no. 2, 1019 -- 1033.
[29] H. Poincar´e, La m´ethode de Neumann et le probl`eme de Dirichlet, Acta Math. 20 (1897),
59 -- 152.
[30] S. Ritter, The spectrum of the electrostatic integral operator for an ellipsoid, in "Inverse
Scattering and Potential Problems in Mathematical Physics," (R.F.Kleinman, R.Kress, and
E.Marstensen, Eds.), Lang, Frankfurt/Bern, (1995), 157 -- 167.
[31] G. Rozenblum and G. Tashchiyan, Eigenvalue asymptotics for potential type operators on
Lipschitz surfaces, Russ. J. Math. Phys. 13 (3), (2006), 326 -- 339.
[32] M. Schiffer, The Fredholm eigenvalues of plane domains, Pacific J. Math. 7 (1957), 1187 -- 1225.
[33] B. Simon, Trace ideals and their applications, 2nd ed., Amer. Math. Soc. (2005).
[34] O. Steinbach and W. L. Wendland, On C.Neumann's method for second-order elliptic systems
in domains with non-smooth boundaries, J. Math. Anal. Appl.262 (2001), 733 -- 748.
[35] M. E. Taylor, Tools for PDE: Pseudodifferential Operators, Paradifferential Operators, and
Layer Potentials, Mathematical Surveys and Monographs, 81 American Math. Soc., Provi-
dence RI, (2000).
[36] G.C. Verchota, Layer potentials and boundary value problems for Laplace's equation in Lip-
schitz domains, J. Funct. Anal. 59 (1984), 572 -- 611.
[37] J. H. White, A global invariant of conformal mappings in space, Proc. Amer. Math. Soc. 38
(1973), 162 -- 164.
[38] S. Zaremba, Les fonctions fondamentales de M. Poincar´e et la m´ethode de Neumann pour une
fronti´ere compos´e de polygones curvilignes, Journal de Math´ematiques Pures et Appliqu´ees
10 (1904), 395 -- 444.
NEUMANN-POINCARE OPERATOR
17
Center for Mathematical Modeling and Data Science, Osaka University, Japan
E-mail address: [email protected]
Chalmers University of Technology and The University of Gothenburg (Swe-
den); St.Petersburg State University, Dept. Math. Physics (St.Petersburg, Rus-
sia)
E-mail address: [email protected]
|
1806.06816 | 1 | 1806 | 2018-06-18T16:45:47 | $C_0$-semigroups of 2-isometries and Dirichlet spaces | [
"math.FA",
"math.CV"
] | In the context of a theorem of Richter, we establish a similarity between $C_0$-semigroups of analytic 2-isometries $\{T(t)\}_{t\geq0}$ acting on a Hilbert space $\mathcal H$ and the multiplication operator semigroup $\{M_{\phi_t}\}_{t\geq 0}$ induced by $\phi_t(s)=\exp (-st)$ for $s$ in the right-half plane $\mathbb{C}_+$ acting boundedly on weighted Dirichlet spaces on $\mathbb{C}_+$. As a consequence, we derive a connection with the right shift semigroup $\{S_t\}_{t\geq 0}$ $$ S_tf(x)=\left \{ \begin{array}{ll} 0 & \mbox { if }0\leq t\leq x, \\ f(x-t)& \mbox { if } x>t, \end{array} \right . $$ acting on a weighted Lebesgue space on the half line $\mathbb{R}_+$ and address some applications regarding the study of the invariant subspaces of $C_0$-semigroups of analytic 2-isometries. | math.FA | math |
C0-SEMIGROUPS OF 2-ISOMETRIES AND DIRICHLET SPACES
EVA A. GALLARDO-GUTI´ERREZ AND JONATHAN R. PARTINGTON
Abstract. In the context of a theorem of Richter, we establish a similarity between
C0-semigroups of analytic 2-isometries {T (t)}t≥0 acting on a Hilbert space H and
the multiplication operator semigroup {Mφt }t≥0 induced by φt(s) = exp(−st) for s
in the right-half plane C+ acting boundedly on weighted Dirichlet spaces on C+. As
a consequence, we derive a connection with the right shift semigroup {St}t≥0
Stf (x) =( 0
f (x − t)
if 0 ≤ x ≤ t,
if x > t,
acting on a weighted Lebesgue space on the half line R+ and address some appli-
cations regarding the study of the invariant subspaces of C0-semigroups of analytic
2-isometries.
1. Introduction
The concept of a 2-isometry was introduced by Agler in the early eighties (cf.
[1]);
this is related to notions due to J. W. Helton (see [8] and [9]) and characterized in
terms of their extension properties (see [2]). Recall that a bounded linear operator T
on a separable, infinite dimensional complex Hilbert space H is called a 2-isometry if it
satisfies
T ∗2T 2 − 2T ∗T + I = 0,
where I denotes the identity operator. In addition, such operators are called analytic if
no nonzero vector is in the range of every power of T . It turns out that Mz, i.e. the
multiplication operator by z, acting on the classical Dirichlet space, is a cyclic analytic
2-isometry. But, moreover, in [15] (see also [14]) Richter proved that any cyclic analytic
2-isometry is unitarily equivalent to Mz acting on a generalized Dirichlet space D(µ).
More precisely, let µ be a finite non-negative Borel measure on the unit circle T and
D(µ) the generalized Dirichlet space associated to µ, that is, the Hilbert space consisting
of analytic functions on the unit disc D such that the integral
ZD f ′(z)2 Zξ=1
1 − z2
ξ − z2 dµ(ξ)! dm(z)
π
Date: June 2016, Revised February 2017.
1991 Mathematics Subject Classification. Primary 47B38.
Key words and phrases. 2-isometries, right-shift semigroups, Dirichlet space.
The authors are partially supported by Plan Nacional I+D grant no. MTM2013-42105-P and
MTM2016-77710-P.
1
2
EVA A. GALLARDO-GUTI ´ERREZ AND JONATHAN R. PARTINGTON
is finite (here dm(z) denotes the Lebesgue area measure in D). Note that if µ = 0, the
space D(µ) is defined to be the classical Hardy space H 2 and for non-zero, finite, non-
negative Borel measures µ on T, the space D(µ) is contained in the Hardy space (see [7,
Chapter 7]). Then Richter's Theorem reads as follows:
Theorem (Richter). Let T be a bounded linear operator on an infinite dimensional
complex Hilbert space H. Then the following condition are equivalent:
(i) T is an analytic 2-isometry with dim Ker T ∗ = 1,
(ii) T is unitarily equivalent to (Mz, D(µ)) for some finite non-negative Borel mea-
sure on T, where
kfk2
D(µ) = kfk2
H 2 +ZD f ′(z)2 Zξ=1
1 − z2
ξ − z2 dµ(ξ)! dm(z)
π
.
One of the main applications of Richter's Theorem concerns the study of the invariant
subspaces for the multiplication operator Mz in the spaces D(µ) and its relationship
with the classical Beurling Theorem for the Hardy space H 2 (see [3]). For instance,
regarding the Dirichlet space D = D(cid:16) dξ
closed, invariant subspace M under Mz satisfies that dimM ⊖ zM = 1. Moreover, if
ϕ ∈ M ⊖ zM with kϕkD = 1, then ϕ(z) ≤ 1 for z ≤ 1 and M = ϕD(mϕ), where
dmϕ is the measure on T given by dmϕ(ξ) = ϕ(ξ)2 dξ
2π . For general D(µ) spaces, an
analogous result holds. We refer the reader to Chapters 7 and 8 in the recent monograph
2π(cid:17), Richter and Sundberg [16] proved that any
"A primer on the Dirichlet space" [7] for more on the subject.
Motivated by the Beurling-Lax Theorem and the work carried out by Richter, the
aim of this work is taking further the study of the 2-isometries and considering C0-
semigroups of 2-isometric operators. In particular, we will establish a similarity between
C0-semigroups of analytic 2-isometries {T (t)}t≥0 acting on a Hilbert space H and the
multiplication operator semigroup {Mφt}t≥0 induced by φt(s) = exp(−st) for s in the
right-half plane C+ acting boundedly on weighted Dirichlet spaces eDC+ (ν) on C+ (see
Definition 2.3). As a consequence, by means of the Laplace transform, we derive a
connection with the right shift semigroup {St}t≥0
Stf (x) =( 0
f (x − t)
if 0 ≤ t ≤ x,
if x > t,
acting on a weighted Lebesgue space on the half line R+. Finally, some applications
regarding the study of the invariant subspaces of C0-semigroups of analytic 2-isometries
are also discussed in Section 3.
2. C0-semigroups of analytic 2-isometries
First, we introduce some basic concepts and terminology regarding C0-semigroups of
bounded linear operators. For more on this topic, we refer the reader to the Engel–Nagel
monograph [6].
C0-SEMIGROUP OF 2-ISOMETRIES
3
A C0-semigroup {T (t)}t≥0 of operators on a Hilbert space H is a family of bounded
linear operators on H satisfying the functional equation
T (t + s) = T (t)T (s)
for all t, s ≥ 0,
T (0) = I,
and such that T (t) → I in the strong operator topology as t → 0+. Given a C0-semigroup
{T (t)}t≥0, there exists a closed and densely defined linear operator A that determines
the semigroup uniquely, called the generator of {T (t)}t≥0, defined by means of
Ax := lim
t→0+
T (t)x − x
t
,
where the domain D(A) of A consists of all x ∈ H for which this limit exists (see [6,
Chapter II], for instance). Although the generator is, in general, an unbounded operator,
it plays an important role in the study of a C0-semigroup, reflecting many of its properties.
However, if 1 is in the resolvent of A, that is, in the set
ρ(A) = {λ ∈ C : (A − λI) : D(A) ⊂ H → H is bijective},
then (A − I)−1 is a bounded operator on H by the Closed Graph Theorem, and the
Cayley transform of A defined by
V := (A + I)(A − I)−1
is a bounded operator on H, since V − I = 2(A − I)−1. Therefore V determines the
semigroup uniquely, since A does. This operator is called the cogenerator of the C0-
semigroup {T (t)}t≥0. Observe that 1 is not an eigenvalue of V .
Recall that if A is a closed operator, then the spectral bound s(A) of A is defined by
s(A) := sup{Re λ : λ ∈ σ(A)},
where σ(A) = C \ ρ(A) is the spectrum of A, and in case that A is the generator of a
C0-semigroup, then s(A) is always dominated by the growth bound of the semigroup, that
is,
−∞ ≤ s(A) ≤ w0 = inf(w ∈ R :
there exists Mw ≥ 1 such that
kT (t)k ≤ Mw ewt for all t ≥ 0 ) .
Indeed, if r(T (t)) denotes the spectral radius of T (t), it follows that w0 = 1
t log r(T (t))
for each t > 0 (see [6, Section 2, Chapter IV], for instance). The following lemma will be
useful in the context of our main result later.
Lemma 2.1. Let {T (t)}t≥0 be a C0-semigroup on a separable, infinite dimensional com-
plex Hilbert space H consisting of 2-isometries and A its generator. Then 1 ∈ ρ(A) and
therefore, the cogenerator V of {T (t)}t≥0 is well-defined.
4
EVA A. GALLARDO-GUTI ´ERREZ AND JONATHAN R. PARTINGTON
Proof. By induction it follows that, for any n ≥ 1 and t ≥ 0, T (t) satisfies
T (t)∗nT (t)n − nT (t)∗T (t) + (n − 1)I = 0,
and so
kT (t)nxk2 = nkT (t)xk2 − (n − 1)kxk2
for x ∈ H. From here, it follows that kT (t)nk ≤ C √n, where C is a constant independent
of n, and therefore the spectral radius r(T (t)) ≤ 1 for any t. Therefore, s(A) ≤ 0; and
therefore 1 ∈ ρ(A).
(cid:3)
The next result consists of a particular instance of [11, Theorem 1], where C0-semigroups
of hypercontractions are considered. We state it for C0-semigroups of 2-isometries and
include its proof for the sake of completeness.
Proposition 2.2. Let {T (t)}t≥0 be a C0-semigroup on a separable, infinite dimensional
complex Hilbert space H. Then the following conditions are equivalent:
(i) T (t) is a 2-isometry for every t ≥ 0.
(ii) The mapping t ∈ R+ 7→ kT (t)xk2 is affine for each x ∈ H.
(iii) RehA2y, yi + kAyk2 = 0
(iv) The cogenerator V of {T (t)}t≥0 exists and is a 2-isometry.
Proof. (i) ⇐⇒ (ii): If each T (t) is a 2-isometry, then for t ≥ 0 and τ > 0 we have
(y ∈ D(A2)).
hT (t + 2τ )x, T (t + 2τ )xi − 2hT (t + τ )x, T (t + τ )xi + hT (t)x, T (t)xi = 0,
so that
(1)
kT (t + τ )xk2 =
1
2
(kT (t)xk2 + kT (t + 2τ )xk2).
Since t ∈ R+ → kT (t)xk2 is continuous, the mapping is affine.
Conversely, taking t = 0 we see that (1) implies that T (τ ) is a 2-isometry.
(ii) ⇐⇒ (iii): For t > 0 we calculate the second derivative of the function g : t 7→
kT (t)yk2 for y ∈ D(A2). We have
g′′(t) =
d2
dt2hT (t)y, T (t)yi
= hA2T (t)y, T (t)yi + 2hAT (t)y, AT (t)yi + hT (t)y, A2T (t)yi.
For g affine, g′′ is zero, and Condition (iii) follows on letting t → 0. Conversely, Condi-
tion (iii) implies Condition (ii) for y ∈ D(A2), and hence for all y by density.
(iii) ⇐⇒ (iv): We calculate
h(I − 2V ∗V + V ∗2V 2)x, xi
C0-SEMIGROUP OF 2-ISOMETRIES
5
for x = (A − I)2y (note that (A − I)−2 : H → H is defined everywhere and has dense
range). We obtain
h(A − I)2y, (A − I)2yi − 2h(A2 − I)y, (A2 − I)yi + h(A + I)2y, (A + I)2yi
= 4hA2y, yi + 8hAy, Ayi + 4hy, A2yi.
Thus V is a 2-isometry if and only if Condition (iii) holds.
(cid:3)
Before stating the main result of the section, let us introduce the following definition.
Definition 2.3. Let ν be a finite positive Borel measure supported on the imaginary
axis. The Dirichlet space eDC+ (ν) is defined as the space of analytic functions F on right
half-plane C+ such that
1
πZC+ F ′(s)2(cid:18)x +
1
πZ ∞
−∞
x
x2 + (y − τ )2 dν(τ )(cid:19) dx dy < ∞
kFk2 = F (1)2 +
where s = x + iy.
The spaces eDC+ (ν) arise, in a natural way, when we analyze C0-semigroups of analytic
2-isometries in Hilbert spaces, as it is stated in our main result:
Theorem 2.4. Let {T (t)}t≥0 be a C0-semigroup on a separable, infinite dimensional
complex Hilbert space H consisting of analytic 2-isometries for every t > 0 such that
(2)
dim\t>0
ker(cid:0)T ∗(t) − e−t I(cid:1) = 1.
Then there exists a finite positive Borel measure ν supported on the imaginary axis such
that {T (t)}t≥0 is similar to the semigroup of multiplication operators induced by exp(−ts)
acting on the space eDC+ (ν). Moreover, if the multiplication operators induced by exp(−ts)
act continuously for every t > 0 on a Dirichlet space eDC+ (ν) where ν is a finite positive
Borel measure supported on the imaginary axis, then the corresponding semigroup consists
of analytic 2-isometries and satisfies (2).
Before proceeding further, let us remark that our main result yields similarity for the
semigroup {T (t)}t≥0 because of the definition of the norm in eDC+ (ν). In addition, as
we shall see later, condition (2) is a way of expressing the property that dim ker V ∗ = 1,
where V is the cogenerator of the semigroup {T (t)}t≥0.
In order to prove Theorem 2.4, we need the following auxiliary results.
Proposition 2.5. Let {T (t)}t≥0 be a C0-semigroup on a separable, infinite dimensional
complex Hilbert space H consisting of analytic 2-isometries. Then the cogenerator V is
an analytic 2-isometry.
Proof. First, we observe that V is well-defined by Lemma 2.1 and, it is a 2-isometry by
Proposition 2.2. So, we are required to show that V is analytic.
6
EVA A. GALLARDO-GUTI ´ERREZ AND JONATHAN R. PARTINGTON
The Wold Decomposition Theorem for 2-isometries (see [12], for instance), yields that
V can be decomposed as V = S⊕ U with respect to H = H1⊕H2, where U is the unitary
part on H2 =Tn V nH and S is an analytic 2-isometry. We will show that U = 0.
Let us assume, on the contrary, that U 6= 0.
First, we observe that since 1 is not an eigenvalue of V , the generator A of the
semigroup {T (t)}t≥0 may be expressed as the (possibly) unbounded operator
(V + I)(V − I)−1.
Moreover, since T (t) commutes with (A − I)−1 and hence with V , it holds that H2 is
invariant under T (t) for every t ≥ 0.
In addition, the generator B of the restricted
semigroup {T (t)H2}t≥0 is the restriction of A to the D(A) ∩ H2 (see [6, Ch. 2, Sec. 2],
for instance); and the cogenerator is U .
Now, taking into account the fact that U is unitary, one deduces that B is skew-adjoint
(i. e., B⋆ = −B). Then the restriction of T (t) to H2 is unitary for every t ≥ 0 and,
therefore, every vector in H2 is in the range of (powers of) T (t). Since T (t) is analytic,
it follows that H2 = {0}, a contradiction. Hence, U = 0 and the proof is completed. (cid:3)
Lemma 2.6. Let {T (t)}t≥0 be a C0-semigroup on a separable, infinite dimensional com-
plex Hilbert space H and A its generator. The following conditions are equivalent:
(1) Ax0 = −x0 for some x0 ∈ D(A).
(2) T (t)x0 = e−tx0 for all t ≥ 0 and x0 ∈ D(A).
In addition, if 1 ∈ ρ(A) and V is the cogenerator, any of the previous conditions is
equivalent to
(3) V x0 = 0 for some x0 ∈ D(A).
Note that the equivalence between (1) and (2) in Lemma 2.6 just follows from the
relationship between the eigenspaces of A and the semigroup {T (t)}t≥0, that is,
Ker(λI − A) = \t≥0
Ker(eλt − T (t)),
with λ ∈ C (see [6, Corollary 3.8, Section IV], for instance). The last statement follows
from the definition of V .
We are now in position to prove Theorem 2.4.
Proof of Theorem 2.4 Assume that {T (t)}t>0 consists of analytic 2-isometries. Let V
denote its cogenerator; this is well-defined by Lemma 2.1, and it is an analytic 2-isometry
by Proposition 2.5.
In addition, the hypotheses dimTt>0 ker (T ∗(t) − e−t I) = 1 along with Lemma 2.6
applied to the adjoint semigroup {T ∗(t)}t≥0, yields that dim Ker V ∗ = 1.
By means of Richter's Theorem, it follows that V is similar to Mz acting on the space
D(µ) for some finite non-negative Borel measure µ on T considered with the equivalent
C0-SEMIGROUP OF 2-ISOMETRIES
7
norm
(3)
kfk2
D(µ) ≈ f (0)2 +ZD f ′(z)2 Zξ=1
= f (0)2 +ZD f ′(z)2Pµ(z)
dm(z)
.
π
1 − z2
ξ − z2 dµ(ξ)! dm(z)
π
Observe that the similarity is the price paid when we consider the equivalent norm.
Hence, for any t ≥ 0, it follows that T (t) is unitarily equivalent to the multiplication
operator induced by exp(−t(1 + z)/(1 − z)) on D(µ). Now, we migrate to the right
half-plane C+ = {Re s > 0} applying the change of variables s = (1 + z)/(1 − z), or
z = (s − 1)/(s + 1).
First, we observe that
Pµ(cid:18) s − 1
s + 1(cid:19) =Zξ=1
1 − s−1
s+12
ξ − s−1
s+12
dµ(ξ)
(s ∈ C+)
is a positive harmonic function in C+; so there exists a non-negative constant ρ and a
finite positive Borel measure ν supported on the imaginary axis such that
(4)
Pµ(cid:18) s − 1
s + 1(cid:19) = ρ x +
1
πZ ∞
−∞
x
x2 + (y − τ )2 dν(τ ),
(s = x + iy)
(see [10, Exercise 6, p. 134], for instance).
We can express ν in terms of µ, since with ξ = (u − 1)/(u + 1) for u = iτ ∈ iR, we
have
Pµ(cid:18) s − 1
s + 1(cid:19) = µ(1) s + 12 − s − 12
(s + 1) − (s − 1)2 +Zξ∈T\{1}
xu + 12
u − s2 dµ(ξ),
x2 + (y − τ )2 dµ(ξ),
= µ(1)x +Zξ∈T\{1}
= µ(1)x +Zξ∈T\{1}
x(1 + τ 2)
(s + 12 − s − 12)u + 12
(u − 1)(s + 1) − (u + 1)(s − 1)2 dµ(ξ)
where s = x + iy ∈ C+. So in (4) we have
and
(5)
ρ = µ(1)
dν(τ )
π(1 + τ 2)
= dµ(ξ).
Then, upon applying the change of variables s = (1 + z)/(1− z) in (3), we deduce that
T (t) is similar to the multiplication operator induced by exp(−ts) acting on the space
eDC+ (ν) consisting of analytic functions F on C+ such that
(6)
1
πZC+ F ′(s)2(cid:18)x +
1
πZ ∞
−∞
x
x2 + (y − τ )2 dν(τ )(cid:19) dx dy < ∞,
where s = x + iy and F (s) = f (z). This proves the first half of Theorem 2.4.
In order to conclude the proof, let us assume that the multiplication operators induced
by exp(−ts) act continuously for every t > 0 on a Dirichlet space eDC+ (ν) where ν is a
finite positive Borel measure supported on the imaginary axis. Reversing the steps above
8
EVA A. GALLARDO-GUTI ´ERREZ AND JONATHAN R. PARTINGTON
and taking into account the fact that (5) defines a measure µ on T, where µ(1) = ν(0) = ρ,
we deduce that the given semigroup is similar to the semigroup of multiplication operators
induced by φt(z) = exp(−t(1 + z)/(1 − z)) on D(µ). Since the cogenerator of such a
semigroup is Mz, which is a 2-isometry, it follows by Proposition 2.2 that {Mφt}t≥0
consists of 2-isometries.
It remains to show that Mφt is analytic for every t > 0. If not, then there are a t0 > 0
and a F ∈ eDC+ (ν) such that the function s 7→ ent0sF (s) lies in eDC+ (ν) for n = 1, 2, 3, . . ..
In particular,
Transferring to the disc by letting s = (1 + z)/(1 − z) and F (s) = f (z), we have
ZC+ (F (s)ent0s)′2x dx dy < ∞.
ZD [f (z) exp(nt0(1 + z)/(1 − z))]′2 1 − z2
1 − z2 dA(z) < ∞,
so that the function z 7→ f (z) exp(nt0(1 + z)/(1− z)) lies in the weighted Dirichlet space
D(δ1) corresponding to a Dirac measure at 1, and hence in H 2(D), by [7, Thm. 7.1.2].
We conclude that f is identically zero, since no nontrivial H 2 function can be divisible
by an arbitrarily large power of a nonconstant inner function. Hence the analyticity is
also established.
A connection with the right-shift semigroup in weighted L2(R+). Now, by means
of the Laplace transform, we will establish a connection of C0-semigroups of analytic
2-isometries {T (t)}t≥0 acting on a Hilbert space H and the the right shift semigroup
{St}t≥0
Stf (x) =( 0
f (x − t)
if 0 ≤ x ≤ t,
if x > t,
acting on a weighted Lebesgue space on the half line R+.
First, let us begin by recalling a result asserting that for each α > −1, a function
α(C+), that is, the space
G analytic in C+ belongs to the weighted Bergman space A2
consisting of analytic functions on C+ for which
α(Π+) =ZC+ G(x + iy)2 xα dx dy < ∞ ,
kGk2
A2
if and only if it has the form
G(s) := Lg(s) =Z ∞
0
e−st g(t) dt ,
s ∈ C+ ,
where g is a measurable function on R+ with
Z ∞
0
g(t)2 t−1−α dt < ∞ .
C0-SEMIGROUP OF 2-ISOMETRIES
9
Moreover,
kGk2
α(C+) =
A2
π Γ(1 + α)
2α
Z ∞
0
(see [4] or [5, Theorem 1], for instance). In other words, the Laplace transform is an
isometric isomorphism between A2
means of a density argument, and taking α = 1, it follows that for any F ∈ eDC+ (ν),
there exists f ∈ L2(R+) (unique in the usual sense of equivalence classes), such that
α(C+) and L2(R+,(cid:16) π Γ(1+α)
t−1−α dt). Hence, by
g(t)2 t−1−α dt ,
(cid:17)1/2
2α
(i) L(tf (t)) = F ′(s),
(ii)
1
πZC+ F ′(s)2 x dx dy =
f (t)2dt,
which corresponds to the first sum in (6); and
2 Z ∞
1
0
(s = x + iy),
(iii)
1
1
x
2π Z ∞
x2 + (y − τ )2 dx dy =
π2ZC+ F ′(s)2
cation operator induced by exp(−ts) acting on the space eDC+ (ν) yields, by means of the
Laplace transform, that {T (t)}t≥0 is transformed to the right-shift semigroup {St}t≥0
acting on the Hilbert space H which consists of functions f defined on R+ such that
These three items along with the fact that for any t ≥ 0, T (t) is similar to the multipli-
u f (u)e−iτ u du(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)Z t
(s = x + iy).
dt
t2 ,
0
2
0
Z ∞
0
f (t)2 dt +Z ∞
−∞(cid:12)(cid:12)(cid:12)(cid:12)Z t
0 Z ∞
0
2
f (u)e−iτ uu du(cid:12)(cid:12)(cid:12)(cid:12)
dν(τ )
dt
t2 < ∞.
3. A final remark on invariant subspaces of C0-semigroups of analytic
2-isometries
As an application of our main result, we deal with the study of the lattice of the closed
invariant subspaces of a C0-semigroup {T (t)}t≥0 of analytic 2-isometries.
Here we shall use the following result from [15, Thm. 7.1] and [16, Thm. 3.2].
Theorem 3.1. Let M be a non-zero invariant subspace of (Mz, D(µ)). Then M = φDµφ
where φ ∈ M ⊖ zM is a multiplier of D(µ) and dµφ = φ2dµ.
In the continuous case we have the following result:
Theorem 3.2. Let {T (t)}t≥0 denote the semigroup of multiplication operators induced
subspace invariant under all the operators T (t). Then there is a function ψ ∈ M such
by exp(−ts) on the space eDC+ (ν), as in Theorem 2.4, and let M be a non-zero closed
that M = ψeDC+ (νψ).
10
EVA A. GALLARDO-GUTI ´ERREZ AND JONATHAN R. PARTINGTON
Proof. If M is invariant under the semigroup, then it is also invariant under the cogener-
ator V , and after transforming to the disc as in the proof of Theorem 2.4, we may apply
Theorem 3.1.
Note that under the equivalence between D(µ) and eDC+ (ν), as detailed in (4) and
(5), the subspace φDµφ maps to a space ψeDC+ (νψ), where ψ(s) = φ((s − 1)/(s + 1)) and
dνψ = ψ2dν.
(cid:3)
Acknowledgements
The authors thank the referees for a careful reading of the manuscript as well as for
their suggestions, which improved the final version of the submitted manuscript.
References
[1] J. Agler, Hypercontractions and subnormality, J. Operator Theory 13 (1985), no. 2, 203–217.
[2] J. Agler, A disconjugacy theorem for Toeplitz operators, Amer. J. Math. 112 (1990), no. 1, 1–14.
[3] A. Beurling, On two problems concerning linear transformations in Hilbert space Acta Math. 81
(1948), 239–255.
[4] N. Das and J. R. Partington, Little Hankel operators on the half-plane, Integral Equations Operator
Theory 20 (1994), no. 3, 306–324.
[5] P. Duren, E. A. Gallardo-Guti´errez and A. Montes-Rodr´ıguez, A Paley–Wiener theorem for
Bergman spaces with application to invariant subspaces, Bull. London Math. Soc. 39 (2007), no. 3,
459–466.
[6] K. J. Engel and R. Nagel, One-parameter Semigroups for Linear Evolution Equations, Graduate
Text in Mathematics, vol. 194, Springer–Verlag, New York, 2000.
[7] O. El-Fallah, K. Kellay, J. Mashreghi, and T. Ransford, A primer on the Dirichlet space, volume
203 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 2014.
[8] J. W. Helton, Operators with a representation as multiplication by × on a Sobolev space, Hilbert
space operators and operator algebras (Proc. Internat. Conf., Tihany, 1970), pp. 279–287, North-
Holland, Amsterdam, 1972.
[9] J. W. Helton, Infinite dimensional Jordan operators and Sturm-Liouville conjugate point theory,
Trans. Amer. Math. Soc. 170 (1972), 305–331.
[10] K. Hoffman, Banach spaces of analytic functions, Dover Publication, Inc., 1988.
[11] B. Jacob, J. R. Partington, S. Pott, and A. Wynn, β-admissibility of observation and control
operators for hypercontractive semigroups, J. Evol. Equ. 18 (2018), no. 1, 153–170.
[12] A. Olofsson, A von Neumann-Wold decomposition of two-isometries, Acta Sci. Math. (Szeged) 79
(2004), 715–726.
[13] B. Sz-Nagy and C. Foias, Harmonic Analysis of Operators on Hilbert Space, North-Holland Pub-
lishing Co., 1970.
[14] S. Richter, Invariant subspaces of the Dirichlet shift, J. Reine Angew. Math. 386 (1988), 205–220.
[15] S. Richter, A representation theorem for cyclic analytic two-isometries, Trans. Amer. Math. Soc.
328 (1991), no. 1, 325–349.
[16] S. Richter and C. Sundberg, Multipliers and invariant subspaces in the Dirichlet space, J. Operator
Theory, 28 (1992), no. 1, 167–186.
C0-SEMIGROUP OF 2-ISOMETRIES
11
Universidad Complutense de Madrid e ICMAT
Departamento de An´alisis Matem´atico,
Facultad de Ciencias Matem´aticas,
Plaza de Ciencias 3
28040, Madrid (SPAIN)
E-mail address: [email protected]
School of Mathematics,
University of Leeds,
Leeds LS2 9JT, U.K.
E-mail address: [email protected]
|
1811.04268 | 1 | 1811 | 2018-11-10T15:15:45 | Lebesgue inequalities for Chebyshev Thresholding Greedy Algorithms | [
"math.FA"
] | We establish estimates for the Lebesgue parameters of the Chebyshev Weak Thresholding Greedy Algorithm in the case of general bases in Banach spaces. These generalize and slightly improve earlier results in [9], and are complemented with examples showing the optimality of the bounds. Our results also correct certain bounds recently announced in [18], and answer some questions left open in that paper. | math.FA | math |
LEBESGUE INEQUALITIES FOR CHEBYSHEV THRESHOLDING GREEDY
ALGORITHMS
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
ABSTRACT. We establish estimates for the Lebesgue parameters of the Chebyshev Weak
Thresholding Greedy Algorithm in the case of general bases in Banach spaces. These gen-
eralize and slightly improve earlier results in [9], and are complemented with examples
showing the optimality of the bounds. Our results also correct certain bounds recently an-
nounced in [18], and answer some questions left open in that paper.
1. INTRODUCTION
Let X be a Banach space over K = R or C, let X∗ be its dual space, and consider a system
{en, e∗n}∞
n=1 ⊂ X× X∗ with the following properties:
a) 0 < infn{kenk,ke∗nk} ≤ supn{kenk,ke∗nk} < ∞
b) e∗n(em) = δn,m, for all n, m ≥ 1
c) X = span{en : n ∈ N}
d) X∗ = span{e∗n : n ∈ N}
Under these conditions B = {en}∞
for X (or M-basis for short), with dual system {e∗n}∞
following special cases
w∗.
n=1 is called a seminormalized Markushevich basis
n=1. Sometimes we shall consider the
N-th partial sum operator
operator. In this case we use the constant
e) B is a Schauder basis if Kb := supN kSNk < ∞, where SNx := ∑N
f) B is a Cesàro basis if supN kFNk < ∞, where FN := 1
N ∑N
N kI − FNk(cid:9).
With every x ∈ X, we shall associate the formal series x ∼ ∑∞
that limn e∗n(x) = 0. As usual, we denote supp x = {n ∈ N : e∗n(x) 6= 0}.
β = max(cid:8)sup
N kFNk, sup
(1.1)
n=1 e∗n(x)en is the
n=1 Sn is the N-th Cesàro
n=1 e∗n(x)en, where a)-c) imply
We recall standard notions about (weak) greedy algorithms; see e.g. the texts [22, 24]
for details and historical background. Fix t ∈ (0, 1]. We say that A is a t-greedy set for x of
1991 Mathematics Subject Classification. 41A65, 41A46, 46B15.
Key words and phrases. thresholding Chebyshev greedy algorithm, thresholding greedy algorithm, quasi-
greedy basis, semi-greedy bases.
The research of the first, third and fourth authors are partially supported by the grants MTM-2016-76566-P
(MINECO, Spain) and 19368/PI/14 (Fundación Séneca, Región de Murcia, Spain). Also, the first author is
supported by a PhD Fellowship from the program "Ayudas para contratos predoctorales para la formación de
doctores 2017" (MINECO, Spain). The second author is supported by Grant MTM-2014-53009-P (MINECO,
Spain). Third author partially supported by grant MTM2017-83262-C2-2-P (Spain). The fourth author has
received funding from the European Union's Horizon 2020 research and innovation programme under the
Marie Skłodowska-Curie grant agreement No 777822.
1
2
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
order m, denoted A ∈ G(x, m, t), if A = m and
(1.2)
min
A t-greedy operator of order m is any mapping G t
form
n∈A e∗n(x) ≥ t · max
n6∈A e∗n(x).
m : X → X which at each x ∈ X takes the
G t
m(x) = ∑
n∈A
e∗n(x)en,
for some set A = A(x, G t
m) ∈ G(x, m, t).
m for the set of all t-greedy operators of order m. The approximation scheme
We write Gt
m=1 to each vector x ∈ X is called a Weak Thresholding
which assigns a sequence {G t
Greedy Algorithm (WTGA), see [15, 23]. When t = 1 one just says Thresholding Greedy
Algorithm (TGA), and drops the super-index t, that is G 1
m(x)}∞
m = Gm, etc.
It is standard to quantify the efficiency of these algorithms, among all possible m-term
approximations, in terms of Lebesgue-type inequalities. That is, for each m = 1, 2, ..., we
look for the smallest constant Lt
(1.3)
where
mσm(x), ∀ x ∈ X, ∀ G t
m(x)k ≤ Lt
m such that
kx− G t
m ∈ Gt
m,
B ≤ mo.
σm(x) := infn(cid:13)(cid:13)(cid:13)x− ∑
n∈B
bnen(cid:13)(cid:13)(cid:13)
: bn ∈ K,
We call the number Lt
m the Lebesgue parameter associated with the WTGA, and we just
write Lm when t = 1. We refer to [24, Chapter 3] for a survey on such inequalities, and to
[11, 9, 1, 4, 5] for recent results. It is known that Lt
m = O(1) holds for a fixed t if and only
if it holds for all t ∈ (0, 1], and if and only if B is unconditional and democratic; see [14]
and [22, Thm 1.39]. In this special case B is called a greedy basis.
In this paper we shall be interested in Chebyshev thresholding greedy algorithms. These
were introduced by Dilworth, Kalton and Kutzarova, see [7, §3], as an enhancement of the
TGA. Here, we use the weak version considered in [9]. Namely, for fixed t ∈ (0, 1] we say
that CGt
m : X → X is a Chebyshev t-greedy operator of order m if for every x ∈ X the set
A = supp CGt
m(x) ∈ G(x, m, t) and moreover
kx− CGt
m(x)k = minn(cid:13)(cid:13)x− ∑
n∈A
anen(cid:13)(cid:13) : an ∈ Ko.
Finally, we define the weak Chebyshevian Lebesgue parameter L
such that
ch,t
m as the smallest constant
kx− CGt
m(x)k ≤ Lch,t
m σm(x), ∀ x ∈ X, ∀ CGm ∈ Gch,t
m ,
where Gch,t
when t = 1 we shall omit the index t, that is Lch
m := L
ch,1
m .
m is the collection of all Chebyshev t-greedy operators of order m. As before,
When Lch
m = O(1) the system B is called semi-greedy; see [7]. We remark that the first
author recently established that a Schauder basis B is semi-greedy if and only if is quasi-
greedy and democratic; see [2].
In this paper we shall be interested in quantitative bounds of L
ch,t
m in terms of the quasi-
greedy and democracy parameters of a general M-basis B. Earlier bounds were obtained
by Dilworth, Kutzarova and Oikhberg in [9] when B is a quasi-greedy basis, and very
recently, some improvements were also announced by C. Shao and P. Ye in [18, Theorem
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
3
3.5]. Unfortunately, various arguments in the last paper seem not to be correct, so one of
our goals here is to give precise statements and proofs for the results in [18], and also settle
some of the questions which are left open there.
To state our results, we recall the definitions of the involved parameters. Given a finite
set A ⊂ N, we shall use the following standard notation for the indicator sums:
en
1A = ∑
n∈A
and 1εA = ∑
n∈A
εnen, ε ∈ ϒ
where ϒ is the set of all ε = {εn}n ⊂ K with εn = 1. Similarly, we write
The relevant parameters for this paper are the following:
PA(x) = ∑
n∈A
e∗n(x)en.
• Conditionality parameters:
km := sup
• Quasi-greedy parameters:
A≤mkPAk and
kc
m = sup
A≤mkI − PAk.
gm := sup
Gk∈Gk, k≤m kGkk
and gc
m := sup
Gk∈Gk, k≤m kI − Gkk.
Below we shall also use the variant
gm := sup
G′<G
G∈G
k, k≤m
kG − G ′k,
where G ′ < G means that A(x, G ′) ⊂ A(x, G ) for all x; see [4].
• Super-democracy parameters:
k1εAk
k1ηBk
k1εAk
k1ηBk
• Quasi-greedy parameters for constant coefficients (see [4, (3.11)])
A=B≤m, A∩B=/0
A=B≤m
ε=η=1
µm = sup
µd
m =
ε=η=1
and
sup
.
γm = sup
ε=1
B⊂A, A≤m
k1εBk
k1εAk
.
Note that γm ≤ gm ≤ gm ≤ 2gm, but in general γm may be much smaller than gm; see e.g.
[4, §5.5]. Likewise, in §5 below we show that µd
m may be much smaller than µm, except for
Schauder bases in which both quantities turn out to be equivalent; see Theorem 5.2.
Our first result is a general upper bound, which improves and extends [18, Theorem 2.4].
Theorem 1.1. Let B be an M-basis in X, and let K = supn, j ke∗nkke jk. Then,
(1.4)
t ) K m , ∀ m ∈ N, t ∈ (0, 1].
Lch,t
m ≤ 1 + (1 + 1
Moreover, there exists a pair (X, B) where the equality is attained for all m and t.
The second result is a slight generalization of [9, Theorem 4.1], and gives a correct ver-
sion of [18, Theorem 3.5].
4
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
Theorem 1.2. Let B be an M-basis in X. Then, for all m ≥ 1 and t ∈ (0, 1],
(1.5)
Lch,t
m ≤ gc
2m +
2
t
min(cid:8) gm µm , γ2m g2m µd
m(cid:9).
Our next result concerns lower bounds for L
ch,t
m , for which we need to introduce weaker
versions of the democracy parameters with an additional separation condition. For two
finite sets A, B ⊂ N and c ≥ 1, the notation A > cB will stand for min A > cmax B.
(1.6) ϑm,c := sup(cid:26)k1εAk
k1ηBk
Theorem 1.3. If B is a Cesàro basis in X with constant β, then for every c ≥ 2
ε = η = 1, A = B ≤ m with A > cB or B > cA(cid:27) .
• Given an integer c ≥ 2, we define
:
Lch,t
m ≥
1
tβ2
c− 1
c + 1
ϑm,c, ∀ m ∈ N, t ∈ (0, 1].
We shall also establish, in Theorem 3.8 below, a similar lower bound valid for more
general M-bases (not necessarily of Cesàro type), in terms of a new parameter θm which is
invariant under rearrangements of B.
Remark 1.4. One may compare the bounds for Lch
m above with those for Lm given in [4]
(1) Lm ≤ 1 + 3Km,
(2) Lm ≤ kc
2m + gm µm,
and
(3) Lm ≥ µd
m,
which illustrate a slightly better behavior of the Chebishev TGA. Observe that one also has
the trivial inequalities
m ≤ kc
m is direct by definition, while Lt
Lch,t
m ≤ Lt
m Lch,t
m .
m ≤ kc
mL
m(x). Pick a Chebyshev greedy operator CGt
ch,t
m can be proved as follows: take
m(x) =
m such that supp CGt
ch,t
m ≤ Lt
Indeed, L
x∈ X and A = supp G t
A. Then
m(x)k = k(I − PA)xk = k(I − PA)(x− CGt
ch,t
m . Hence, when B is unconditional then Lt
so Lt
tional quasi-greedy and democratic bases we have Lch
kx− G t
m ≤ kc
mL
m(x))k ≤ kc
m ≈ L
mkx− CGt
m(x)k,
ch,t
m . However for all condi-
m = O(1), but Lm → ∞.
The paper is organized as follows. Section 2 is devoted to preliminary lemmas. In Sec-
tion 3 we prove Theorems 1.1, 1.2 and 1.3, and also establish the more general lower bound
in Theorem 3.8, giving various situations in which it applies. Section 4 is devoted to ex-
amples illustrating the optimality of the results; in particular, an optimal bound of Lch
m for
the trigonometric system in L1(T), settling a question left open in [18]. In Section 5 we
investigate the equivalence between µd
m and µm and show Theorem 5.2. Finally, in Section
6 we study the convergence of CGm(x) and Gm(x) to x under the strong M-basis assumption,
settling a gap in [18, 27].
2. PRELIMINARY RESULTS
We recall some basic concepts and results that will be used later in the paper; see [7, 4].
For each α > 0 we define the α-truncation of a scalar y ∈ K as
Tα(y) = αsign y if y ≥ α,
and Tα(y) = y if y ≤ α.
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
5
We extend Tα to an operator in X by formally assigning Tα(x) ∼ ∑∞
n=1 Tα(e∗n(x))en, that is
Tα(x) := α1εΛα(x) + (I − PΛα(x))(x),
where Λα(x) = {n : e∗n(x) > α} and ε = {sign (e∗n(x))}. Of course, this operator is well
defined since Λα(x) is a finite set. In [4] we can find the following result:
Lemma 2.1. [4, Lemma 2.5] For all α > 0 and x ∈ X, we have
kTα(x)k ≤ gc
Λα(x)kxk.
We also need a well known property from [7, 8], formulated as follows.
Lemma 2.2. [4, Lemma 2.3] If x ∈ X and ε = {sign (e∗n(x))}, then
(2.1)
min
n∈G e∗n(x)k1εGk ≤ gGkxk,
∀G ∈ G(x, m, 1).
The following version of (2.1), valid even if G is not greedy, improves [9, Lemma 2.2].
Lemma 2.3. Let x∈ X and ε ={sign (e∗n(x))}. For every set finite A ⊂ N, if α = minn∈Ae∗n(x),
then
(2.2)
where Λα(x) = {n : e∗n(x) > α}.
Proof. Call G = A∪ Λα(x), and notice that it is a greedy set for x. Then,
αk1εAk ≤ γA∪Λα(x) gA∪Λα(x)kxk,
αk1εAk ≤ αγGk1εGk ≤ γG gGkxk,
using (2.1) in the last step.
Remark 2.4. The following is a variant of (2.2) with a different constant
(2.3)
min
n∈A e∗n(x) k1εAk ≤ kAkxk.
A similar proof as the one in Lemma 2.3 can be seen in [3, Proposition 2.5].
(cid:3)
Finally, using convexity as in [4, Lemma 2.7], one has the elementary lemma.
(cid:13)(cid:13)(cid:13) ∑
Lemma 2.5. For all finite sets A ⊂ N and scalars an ∈ K it holds
ε=1(cid:13)(cid:13)1εA(cid:13)(cid:13).
m ∈ Gch,t
3.1. Proof of Theorem 1.1. Let x ∈ X and CGt
operator, and denote by A = supp CGt
B = m. By definition of the Chebyshev operators,
anen(cid:13)(cid:13)(cid:13) ≤ max
3. PROOF OF THE MAIN RESULTS
n∈A an sup
n∈A
kx− CGt
m(x)k ≤ kx− PA∩B(x)k ≤ kPB\A(x)k +kx− PB(x)k.
On the one hand, using (1.2),
m be a fixed Chebyshev t-greedy
mx ∈ G(x, m, t). Pick any z = ∑n∈B bnen such that
kPB\A(x)k ≤ sup
n kenk ∑
j∈B\Ae∗j (x) ≤
sup
n kenk ∑
j∈A\Be∗j(x− z) ≤
1
t
1
t
Kmkx− zk.
On the other hand, using the inequality (3.9) of [4],
kx− PB(x)k = k(I − PB)(x− z)k ≤ kc
mkx− zk ≤ (1 + Km)kx− zk.
ch,t
Hence, L
witnessed by Examples 4.1 and 4.2 below.
t(cid:1) Km. Finally, the fact that the equality in (1.4) can be attained is
m ≤ 1 +(cid:0)1 + 1
6
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
3.2. Proof of Theorem 1.2. The scheme of the proof follows the lines in [7, Theorem 3.2]
and [9, Theorem 4.1], with some additional simplifications introduced in [4].
Given x ∈ X and CGt
m ∈ Gch,t
m , we denote by A = supp CGt
mx ∈ G(x, m, t). Pick any
z = ∑n∈B bnen such that B = m. By definition of the Chebyshev operators,
(3.1)
We make the selection of p suggested in [7]. Namely, if α = maxn /∈Ae∗n(x), we let
for any p = ∑n∈A anen.
mxk ≤ kx− pk,
kx− CGt
It is easily verified that
p = PA(x)− PA(cid:0)Tα(x− z)(cid:1).
(3.2)
Since Λα(x− z) = {n :
(3.3)
Next we treat the first term in (3.2). Observe that maxn∈B\Ae∗n(x− Tα(x− z)) ≤ 2α, so
Lemma 2.5 gives
x− p = (I − PA)(cid:0)x− Tα(x− z)(cid:1) + Tα(x− z)
= PB\A(cid:0)x− Tα(x− z)(cid:1) + Tα(x− z).
e∗n(x− z) > α} ⊂ A∪ B, then Lemma 2.1 gives
(cid:13)(cid:13)Tα(x− z)(cid:13)(cid:13) ≤ gc
(cid:13)(cid:13)PB\A(cid:0)x− Tα(x− z)(cid:1)(cid:13)(cid:13) ≤ 2α sup
2mkx− zk.
ε=1(cid:13)(cid:13)1ε(B\A)(cid:13)(cid:13)
n∈A\Be∗n(x− z) sup
min
2
t
≤
ε=1(cid:13)(cid:13)1ε(B\A)(cid:13)(cid:13) = (∗).
At this point we have two possible approaches. Let ηn = sign [e∗n(x− z)]. In the first ap-
proach we pick a greedy set Γ ∈ G(x− z,A\ B, 1), and control (3.4) by
µm gmkx− zk,
(3.5)
(∗) ≤
min
2
t
2
t
using Lemma 2.2 in the last step. In the second approach, we argue as follows
(3.4)
(3.6)
(∗) ≤
2
t
2
t
γ2m g2m µd
mkx− zk,
n∈Γ e∗n(x− z) µm(cid:13)(cid:13)1ηΓ(cid:13)(cid:13) ≤
m(cid:13)(cid:13)1η(A\B)(cid:13)(cid:13) ≤
n∈A\Be∗n(x− z) µd
min
using in the last step Lemma 2.3 and the fact that, if δ = minA\Be∗n(x− z), then the set
(A\ B)∪{n :
e∗n(x− z) > δ} ⊂ A∪ B and hence has cardinality ≤ 2m.
We can now combine the estimates displayed in (3.1)-(3.6) and obtain
kx− CGt
mxk ≤ (cid:2)gc
2m +
min(cid:8) gm µm , γ2m g2m µd
m(cid:9)(cid:3)kx− zk,
2
t
which after taking the infimum over all z establishes Theorem 1.2.
(cid:3)
Remark 3.1. In [18, Theorem 3.5] a stronger inequality is stated (for t = 1), namely
(3.7)
Lch
m ≤ gc
2m + 2 gm µd
m.
The proof, however, seems to contain a gap, and a missing factor kc
m should also appear in
the last summand. Nevertheless, it is still fair to ask whether the inequality (3.7) asserted in
[18] may be true with a different proof.
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
7
Remark 3.2. Using Remark 2.4 in place of Lemma 2.3 in (3.6) above leads to an alternative
and slightly simpler estimate
(3.8)
Lch,t
m ≤ gc
2m +
2
t
km µd
m .
However, this would not be as efficient as (1.5) when B is quasi-greedy and conditional.
Remark 3.3. When B is quasi-greedy with constant q = supm gm < ∞, then Theorem 1.2
implies the following
Lch,t
m ≤ q + 4t−1 q2 µd
m.
3.3. Proof of Theorem 1.3. Recall that SN = ∑N
This is a slight improvement with respect to [9, Theorem 4.1].
n=1 e∗n(·)en and
n=1(cid:0)1−
∑
FN(x) =
n− 1
N (cid:1)e∗n(x)en.
Sn(x) =
∑
n=1
1
N
N
N
For M > N we define the operators (of de la Vallée-Poussin type)
VN,M(x) =
(3.9)
=
M
N
M − N
∑
n=1
e∗n(x)en +
FM(x)−
N
FN(x)
M
M − N
n=N+1(cid:0)1−
∑
n− N − 1
M − N (cid:1) e∗n(x)en.
In particular, observe that, for β as in (1.1) we have
(3.10)
max(cid:8)kVN,Mk,kI −VN,Mk(cid:9) ≤
M + N
M− N
β.
We next prove that, if c ≥ 2, then for all A, B ⊂ N such that B > cA with A = B ≤ m it
holds
(3.11)
k1εAk
k1ηBk
Pick any set C > B such that B∪C = m, and let
Lch,t
m ≥
c− 1
c + 1
1
tβ
,
∀ ε = η = 1.
x = 1εA + t1ηB + t1C.
Then B∪C ∈ G(x, m, t), and hence there is a Chebyshev t-greedy operator so that
x− CGt
m(x) = 1εA + ∑
n∈B∪C
anen,
for some scalars an ∈ K. Clearly,
kx− CGt
m(x)k ≤ Lch,t
m σm(x) ≤ Lch,t
m kt1ηBk,
using z = 1εA + t1C an m-term approximant. On the other hand, let N = max A. Since
min B∪C > cN, then (3.9) yields
Therefore, (3.10) implies that
VN,cN(x− CGt
mx) = 1εA.
kx− CGt
m(x)k ≥ kVN,cN(x− CGt
kVN,cNk
mx)k
We have therefore proved (3.11).
c− 1
(c + 1)β k1εAk.
≥
8
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
We next show that when A = B ≤ m satisfy A > cB then
(3.12)
Lch,t
m ≥
1
tβ2
c− 1
c + 1
k1εAk
k1ηBk
, ∀ ε = η = 1.
This together with (3.11) is enough to establish Theorem 1.3. We shall actually show a
slightly stronger result:
Lemma 3.4. Let A = B ≤ m and let y ∈ X be such that y∞ := supne∗n(y) ≤ 1 and
A > c(B ·∪ supp y). Then
Lch,t
(3.13)
m ≥
, ∀ ε = η = 1.
c− 1
c + 1
1
tβ2
k1εAk
k1ηB + yk
Observe that the case y = 0 in (3.13) yields (3.12). We now show (3.13). Pick a large
integer λ > 1 and a set C > λA such that B∪C = m. Let
x = 1εA + ty + t1ηB + t1C.
As before, B∪C ∈ G(x, m, t), and hence for some Chebyshev t-greedy operator we have
x− CGt
m(x) = 1εA + ty + ∑
n∈B∪C
anen,
for suitable scalars an ∈ K. Choosing 1εA + t1C as m-term approximant of x we see that
which letting λ → ∞ yields (3.13). This completes the proof of Lemma 3.4, and hence of
Theorem 1.3.
Remark 3.5. When B is a Schauder basis, a similar proof gives the following lower bound,
which is also obtained in [18, Theorem 2.2]
Lch,t
m ≥
1
(Kb + 1)t
supnk1εAk
k1ηBk
:
A = B = m, A > B or B > A , ε = η = 1o.
The statement for Cesàro bases, however, will be needed for the applications in §4.3.
3.4. Lower bounds for general M-bases. Observe that
ϑm,c = sup
A≤m
ϑc(A), where ϑc(A) = sup
B :
B=A
B>cA
ε,η∈ϒ
We consider a new parameter
(3.14)
ϑm = sup
A≤m
inf
c≥1
ϑc(A).
maxnk1εAk
k1ηBk
k1εAko.
, k1ηBk
On the other hand, calling N = max(B ·∪ supp y) and L = max A we have
Thus,
kx− CGt
Therefore we obtain
m(x)k ≥
kx− CGt
m σm(x) ≤ Lch,t
m(x)k ≤ Lch,t
(I −VN,cN)◦ VL,λL(cid:0)x− CGt
kI −VN,cNkkVL,λLk ≥
λ− 1
Lch,t
m ≥
λ + 1
c− 1
c + 1
k1εAk
1
tβ2
m t k1ηB + yk.
mx(cid:1) = 1εA
c− 1
(c + 1)β
k1εAk
k1ηB + yk
λ− 1
(λ + 1)β k1εAk.
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
9
We remark that, unlike ϑm,c, the parameter ϑm depends on {en}∞
n=1 but not on the reorder-
ch,t
m in terms of ϑm in a less restrictive
ings of the system. We shall give a lower bound for L
situation than the Cesàro basis assumption on {en}∞
n=1.
Given ρ ≥ 1, we say that {en}∞
n=1 is ρ-admissible if the following holds: for each finite
set A ⊂ N, there exists n0 = n0(A) such that, for all sets B with min B ≥ n0 and B ≤ A,
(3.15)
Observe that (3.15) implies that
(3.16)
n∈A
n∈A∪B
(cid:13)(cid:13) ∑
αnen(cid:13)(cid:13) ≤ ρ(cid:13)(cid:13) ∑
αnen(cid:13)(cid:13) ≤ (ρ+ 1)(cid:13)(cid:13) ∑
(cid:13)(cid:13) ∑
αnen(cid:13)(cid:13), ∀ αn ∈ K.
αnen(cid:13)(cid:13),
∀ αn ∈ K.
n∈A∪B
n∈B
This condition is clearly satisfied by all Schauder and Cesàro bases (with ρ = Kb or ρ > β),
but we shall see below that it also holds in more general situations.
Proposition 3.6. Let {en, e∗n}∞
Lch,t
(3.17)
m ≥
n=1 be an M-basis such that {en}∞
t ∈ (0, 1].
n=1 is ρ-admissible. Then
, ∀ m ∈ N,
(ρ+ 1)t
ϑm
Proof. Fix A ⊂ N such that A ≤ m. Choose C disjoint with A such that A∪ C = m. Let
n0 = n0(A∪C) as in the above definition, which we may assume larger than max A∪C. Pick
any B with min B ≥ n0 and B = A, and any ε,η ∈ ϒ. Let x = t1εA + t1C + 1ηB. Then
A∪ C ∈ G(x, m, t), and there is a Chebyshev t-greedy operator with CGt
m(x) supported in
A∪C. Thus,
m kx− (1ηB + t1C)k = Lch,t
m σm(x) ≤ Lch,t
m(x)k ≤ Lch,t
m t k1εAk.
kx− CGt
On the other hand, using the property in (3.16) one obtains
Thus,
Lch,t
m ≥
1
(ρ+ 1)t
kx− CGt
m(x)k ≥ k1ηBk
ρ+ 1
.
.
k1ηBk
k1εAk
We now assume additionally that min B ≥ n0 + m, and pick D ⊂ [n0, n0 + m− 1] such that
B +D = m. Let y = 1εA + t1ηB + t1D. Then B∪ D ∈ G(y, m, t) and a similar reasoning
gives
k1εAk
ρ ≤ ky− CGt
m(y)k ≤ Lch,t
m t k1ηBk.
Thus,
Lch,t
m ≥
1
(ρ+ 1)t
maxnk1ηBk
k1εAk
m σm(y) ≤ Lch,t
k1ηBko,
, k1εAk
infc≥1 ϑc(A)
(ρ+ 1)t
.
Lch,t
m ≥
ϑn0+m(A)
(ρ+ 1)t ≥
and taking the supremum over all B = A with B ≥ (n0 + m)A and all ε,η∈ ϒ, we see that
Finally, a supremum over all A ≤ m leads to (3.17).
(cid:3)
10
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
We now give some general conditions in {en, e∗n}∞
n=1 and X under which ρ-admissibility
holds. We recall a few standard definitions; see e.g. [12]. We use the notation [en]n∈A =
span{en}n∈A, for A ⊂ N. A sequence {en}∞
n=1 is weakly null if
Given a subset Y ⊂ X∗, we shall say that {en}∞
lim
n→∞
x∗(en) = 0, ∀ x∗ ∈ X∗.
n=1 is Y -null if
lim
n→∞
y(en) = 0, ∀ y ∈ Y.
sup
Given κ ∈ (0, 1], we say that a set Y ⊂ X∗ is κ-norming whenever
x∗∈Y,kx∗k≤1x∗(x) ≥ κkxk, ∀ x ∈ X.
n=1 be a biorthogonal system in X × X∗. Suppose that the
n=1 ⊂ X is Y -null, for some subset Y ⊂ X∗ which is κ-norming.
Proposition 3.7. Let {en, e∗n}∞
sequence {en := ke∗nk en}∞
Then {en}∞
Proof. Consider a finite set A ⊂ N with say A = m and denote
n=1 is ρ-admissible for every ρ > 1/κ.
E := [en]n∈A.
Given ε > 0, one can find a finite set S ⊂ Y ∩{x∗ ∈ X∗ : kx∗k = 1} so that
(3.18)
max
x∗∈S x∗(e) ≥ (1− ε)κkek, ∀ e ∈ E.
Indeed, it suffices to verify the above inequality for e of norm 1. Pick an εκ/2-net (zk)N
in the unit sphere of E. For any k find a norm one z∗k ∈ Y so that z∗k(zk) > (1− ε/2)κ. We
claim that S = {z∗k : 1 ≤ k ≤ N} has the desired properties. To see this, pick a norm one
e ∈ E, and find k with ke− zkk ≤ εκ/2. Then
k=1
max
x∗∈S x∗(e) ≥ z∗k(e) ≥ z∗k(zk)−ke− zkk ≥ (1− ε/2)κ− εκ/2 = (1− ε)κ.
Next, since the sequence {ke∗nk en} is Y -null, for each δ > 0 we can find an integer n0 >
max A so that
max
x∗∈S x∗(en)ke∗nk ≤
δκ
m
, ∀ n ≥ n0.
Pick any B of cardinality m with min B ≥ n0, and let
G := [en]n∈B.
n∈Bx∗(en)ke∗nkk fk ≤ δκk fk.
For f = ∑n∈B e∗n( f )en ∈ G, we have
(3.19)
x∗∈S x∗( f ) ≤ max
x∗∈S ∑
max
We claim that
(3.20)
To show this, we fix γ > 0 (to be chosen later), and assume first that k fk ≥ (1 + γ)kek.
Then,
for any e ∈ E, f ∈ G.
1 + δκ kek,
ke + fk ≥
(1− ε− δ)κ
ke + fk ≥ k fk−kek ≥ γkek.
Next assume that k fk < (1 + γ)kek, then using (3.18) and (3.19) we obtain that
ke + fk ≥ max
x∗∈S x∗(e + f ) ≥ (1− ε)κkek− δκk fk > (1− ε− δ(1 + γ))κkek.
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
11
We now choose γ so that γ = (1− ε− δ(1 + γ))κ, that is,
γ =
(1− ε− δ)κ
1 + δκ
,
which shows the claim in (3.20). Now, given ρ > 1/κ, we may pick δ = ε sufficiently
small so that the above number γ > 1/ρ. Then, (3.20) becomes
ke + fk ≥
1
ρkek,
for any e ∈ [en]n∈A, f ∈ [en]n∈B,
for all B with min B ≥ n0 and B = A = m. Thus, {en}∞
n=1 is ρ-admissible.
(cid:3)
We mention a few cases where the hypotheses in the above proposition can be applied:
n=1 is weakly null, since Y = X∗ is always 1-norming.
n=1 is Y -null. In particular, when {en}∞
(1) When the sequence {en}∞
(2) When supn≥1kenkke∗nk < ∞ and Y = [e∗n]n∈N is κ-norming, since the first condition
implies that {en}∞
n=1 is a Schauder basis in X, in
which case the above conditions hold with κ = 1/Kb; see [19, Theorems I.3.1 and I.12.2].
(3) In every separable Banach space X, if one picks {en, e∗n}∞
n=1 to be an M-basis with the
properties in (2) and κ = 1; see e.g. [20, Theorem III.8.5] for the existence of such bases.
(4) Let X = C(K) where K is a compact Hausdorff set and let µ be a Radon probability
measure in K with suppµ = K. Then, the natural embedding of C(K) into L∞(µ) is iso-
metric, and therefore Y = L1(µ) is 1-norming in X. Let {en}∞
n=1 be a complete system in X
which is orthonormal with respect to µ and uniformly bounded, that is, RK enem dµ = δn,m
and supnkenk∞ < ∞. Then the sequence {en}∞
n=1 is L1(µ)-null in X. Indeed, this follows
from case (2), and the fact that C(K) is dense in L1(µ).
Examples of such systems in C(K) include the trigonometric system in C[0, 1] (in the real
or complex case), as well as certain polygonal versions of the Walsh system [6, 16, 26], or
any reorderings of them (which may cease to be Cesàro bases).
(5) As a dual of the previous, if X = L1(µ) then every system {en}∞
null, and hence case (1) applies.
(6) Recall the definition of the right fundamental function: ϕr(m) = sup{k1Ak : A ≤ m}.
If {en}∞
n=1 is such that ϕr(m) = o(m), then this system is weakly null. Indeed, first note that
also ϕr(m) = sup{k1ηAk : A ≤ m, η = 1} = o(m). Assume that the system is not weakly
null. Then there exist a norm one x∗ ∈ X∗ and ε0 > 0 so that the set A = {n ∈ N : x∗(en) ≥
ε0} is infinite. Pick any F ⊂ A with F = m and let ηn = sign[x∗(en)]; then
n∈F x∗(en) ≥ mε0,
ϕr(m) ≥ k1ηFk ≥ x∗( ∑
n∈F
n=1 as in (4) is weakly
ηnen) = ∑
contradicting our assumption.
Finally, as a consequence of Propositions 3.6 and 3.7 one obtains
Theorem 3.8. Let {en, e∗n}∞
n=1 be a seminormalized M-basis such that the sequence {en}∞
is Y -null for some subset Y ⊂ X∗ which is κ-norming. Then, if ϑm is as in (3.14), we have
(3.21)
κϑm
n=1
Lch,t
m ≥
(κ+ 1)t
, ∀ m ∈ N,
t ∈ (0, 1].
12
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
The first two examples are variants of those in [4, §5.1] and [5, §8.1].
4. EXAMPLES
4.1. Example 4.1: The summing basis. Let X be the closure of the set of all finite se-
quences a = (an)n ∈ c00 with the norm
The canonical system B = {en}∞
kak = sup
n=1 is a Schauder basis in X with Kb = 1 and kenk = 1
for all n. Also, ke∗1k = 1, ke∗nk = 2 if n ≥ 2, so K = 2 in Theorem 1.1; see [4, §5.1]. We now
show that, for this example of (X, B), the bound of Theorem 1.1 is sharp. As in [4, §5.1],
we consider the element:
∑
n=1
m (cid:12)(cid:12)(cid:12)
m
an(cid:12)(cid:12)(cid:12).
,
2
,
,
1
2
,
1
t
,
, ...,
1
t
,
1
2
1
2
;
1
2
min
kx− CGt
m(x)k =
, 0, ...(cid:17),
;−1, 1
{z}
x =(cid:16) 1
, ...,−1, 1
{z}
{z }
{z }
2(cid:1) and m blocks of (−1, 1). Picking A = {n : xn = −1}
where we have m blocks of (cid:0) 1
t , 1
2 , 1
as a t-greedy set of x, we see that
ai,i=1,...,m(cid:13)(cid:13)(cid:13)(cid:16)1
; a1, 1, a2, 1, ..., am, 1, 0, ...(cid:17)(cid:13)(cid:13)(cid:13)
≥ (cid:13)(cid:13)(cid:13)(cid:16)1
σm(x) ≤ (cid:13)(cid:13)(cid:13)x−
= (cid:13)(cid:13)(cid:13)(cid:16)1
m ≥ 1 + 2(1 + 1
;−1, 1, ...,−1, 1, 0...(cid:17)(cid:13)(cid:13)(cid:13) =
(0, 1, 0, ..., 0, 1, 0; 0, ...)(cid:13)(cid:13)(cid:13)
2
t )m and we conclude that L
Hence, L
As a consequence, observe that in this case CGt
; 0, ...(cid:17)(cid:13)(cid:13)(cid:13) = m +
t )m by Theorem 1.1.
ch,t
m = 1 + 2(1 + 1
On the other hand,
1
2
1
2
1
2
1
t
,
1
t
1
2
,
,−1,
,−1,
t + 1
1
2
;
, ...,
1
2
;
1
2
, ...,
1
2
;
1
2
1
2
.
, ...,
1
t
,
1
2
1
t
,
1
2
.
,
2
m
t
+
t
1
2
,
2
ch,t
1
2
m(x) = 0.
Remark 4.1. The above example strengthens [18, Theorem 2.4], where the authors are only
able to show that 1 + 4m ≤ Lch
4.2. Example 4.2: the difference basis. Let {en}∞
define the elements
n=1 be the canonical basis in ℓ1(N) and
m ≤ 1 + 6m.
y1 = e1, yn = en − en−1, n = 2, 3, ...
n=1 is called the difference basis of ℓ1. We recall some basic
The new system B = {yn}∞
properties used in [5, §8.1]. If (bn)n ∈ c00 then
∞
∑
n=1bn− bn+1.
bnynk =
∞
∑
n=1
k
ch,t
m ≤ 1 + 2(1 + 1
Also, B is a monotone basis with ky1k = 1, kynk = 2 if n ≥ 2, and ky∗nk = 1 for all n ≥ 1
(in fact, the dual system corresponds to the summing basis). So, K = 2 and Theorem 1.1
t )m for all t ∈ (0, 1]. To show the equality we consider the vector
gives L
x = ∑n bnyn with coefficients (bn) given by
, 0, ...(cid:17),
(cid:16)1, 1, 1,− 1
, ..., 1, 1,− 1
t , 1
t , 1
}
}
{z
{z
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
13
where the block (cid:16)1, 1, −1
t , 1(cid:17) is repeated m times. If we take Γ = {2, 6, ..., 4m− 2} as a
t-greedy set for x of cardinality m, then
kx− CGt
m(x)k = inf
(a j)m
j=1kx−
m
∑
j=1
a jy4 j−2k
= inf
(a j)m
= inf
(a j)m
j=1
j=1(cid:13)(cid:13)(cid:13)(cid:16)1, 1− a1, 1, −1
j=1a j + 2m(cid:16)1 +
∑
, 1, ..., 1− am, 1, −1
t(cid:17) + 1 = 2m(cid:16)1 +
1
2
m
t
t
, 1, 0, ...(cid:17)(cid:13)(cid:13)(cid:13)
t(cid:17) + 1.
1
Hence, in this case we also have CGt
m(x) = 0. On the other hand
m
σm(x) ≤(cid:13)(cid:13)x + (1 + 1
∑
j=1
ch,t
m = 1 + 2(1 + 1
t )m.
t )
y4 j(cid:13)(cid:13) = k(1, 1, 1, 1, 1, ..., 1, 1, 1, 1, 0, ...)k = 1.
This shows that L
4.3. Example 4.3: the trigonometric system in Lp(T). Consider B = {einx}n∈Z in Lp(T)
for 1 ≤ p < ∞, and in C(T) if p = ∞. In [21], Temlyakov showed that
cpm 1
p− 1
p− 1
2 ≤ Lm ≤ 1 + 3m 1
2,
for some cp > 0 and all 1 ≤ p ≤ ∞. Adapting his argument, Shao and Ye have recently
established, in [18, Theorem 2.1], that for 1 < p ≤ ∞ it also holds
(4.1)
The case p = 1 is left as an open question, and only the estimate √m
m . √m is given;
see [18, (2.24)]. Moreover, the proof of the case p = ∞ seems to contain some gaps and
may not be complete.
m ≈ m 1
p− 1
Lch
2.
ln(m) . Lch
Here, we shall give a short proof ensuring the validity of (4.1) in the full range 1 ≤ p ≤ ∞,
with a reasoning similar to [4, §5.4]. More precisely, we shall prove the following.
Proposition 4.2. Let 1 ≤ p ≤ ∞. Then there exists cp > 0 such that
(4.2)
t ∈ (0, 1].
m ≥ cp t−1 m 1
p− 1
Lch,t
2, ∀ m ∈ N,
We remark that in the cases p = 1 and p = ∞ the trigonometric system is not a Schauder
basis, but it is a Cesàro basis1. So we may use the lower bounds in Theorem 1.3, namely
Lch,t
m ≥ c′p t−1
sup
sup
A=B≤m
A>2B or B>2A
ε=η=1
k1εAk
k1ηBk
.
(4.3)
(4.4)
• Case 1 < p ≤ 2. Assume that m = 2ℓ + 1 or 2ℓ + 2 (that is, ℓ = ⌊ m−1
B = {−ℓ, ..., ℓ}, so that 1B = Dℓ is the ℓ-th Dirichlet kernel, and hence
2 ⌋). We choose
Next we take a lacunary set A = {2 j : j0 ≤ j ≤ j0 + 2ℓ}, so that
1− 1
p .
k1Bkp = kDℓkLp(T) ≈ m
k1Akp ≈ √m,
1We equip B with its natural ordering {1, eix, e−ix, e2ix, e−2ix, . . .}.
k1Ak∞ = kD2L−1 − 1kL∞(T) ≈ m.
So, (4.3) implies the desired bound.
• Case p = 1. We use the lower bound in Lemma 3.4, namely
(4.5)
Lch,t
m ≥ c′1 t−1
k1Ak
k1B + yk
,
14
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
and where j0 is chosen such that 2 j0 ≥ m, and hence A > 2B. Then, (4.3) implies
Lch,t
m ≥ cp t−1 m1/2
1− 1
m
p
= cp t−1 m 1
p− 1
2.
• Case 2 ≤ p < ∞. The same proof works in this case, just reversing the roles of A
and B.
• Case p = ∞. We replace the lacunary set by a Rudin-Shapiro polynomial of the
form
R(x) = eiNx
2L−1
∑
n=0
εneinx, with εn ∈ {±1},
where L is such that 2L ≤ m < 2L+1; see e.g. [13, p. 33]. Then, R = 1εB with
B = N +{0, 1, . . ., 2L − 1} and
k1εBk∞ = kRkL∞(T) ≈ √m.
If we pick N ≥ 2· 2L, then B > 2A with A = {±1, . . . ,±(2L − 1)}. Finally,
for all A = B ≤ m and all y such that A > 2(B ·∪ supp y) and supne∗n(y) ≤ 1. As
before, let m = 2ℓ + 1 or 2ℓ + 2, and choose the same sets A and B as in the case
1 < p ≤ 2. Next choose y so that the vector
Vℓ = 1B + y
is a de la Vallée-Poussin kernel as in [13, p. 15]. Then, the Fourier coeffients e∗n(y)
have modulus ≤ 1 and are supported in {n : ℓ < n ≤ 2ℓ + 1}, so the condition
A > 2(B ·∪ supp y) holds if 2 j0 ≥ 2m + 1. Finally,
so the bound L
ch,t
m & t−1√m follows from (4.5).
k1B + yk1 = kVℓkL1(T) ≤ 3,
Remark 4.3. Using the trivial upper bound L
L
ch,t
m ≈ t−1m 1
p− 1
2 for all 1 ≤ p ≤ ∞.
ch,t
m ≤ Lt
m . t−1m 1
p− 1
2, we conclude that
5. COMPARISON BETWEEN µm AND µd
m
In this section we compare the democracy constants µm and µd
m defined in §1 above. Let
us first note that
(5.1)
and
(5.2)
m)2
µd
m ≤ µm ≤ ( µd
m ≤ µm ≤ (1 + 2κ)γm µd
µd
m,
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
15
where κ = 1 or 2 depending if K = R or C. Indeed, the left inequality in (5.1) is immediate
by definition, and the right one follows from
= k1ηBk
k1Ck
k1Ck
m)2,
k1εAk ≤ ( µd
for any A = B ≤ m and any C disjoint with A∪ B with C = A = B. Concerning the
right inequality in (5.2), we use that if A = B ≤ m then
k1ηBk
k1εAk
k1ηBk ≤ k1ε(A\B)k +k1ε(A∩B)k
k1εAk
k1ηBk
≤ γm k1ε(A\B)k
k1η(B\A)k
+ k1ε(A∩B)k
k1ηBk ≤ γm µd
m + 2κγm,
using in the last step [4, Lemma 3.3]. From (5.2) we see that µm ≈ µd
greedy for constant coefficients.
m when B is quasi-
In the next subsection we shall show that µm ≈ µd
seems new in the literature.
m for all Schauder bases, a result which
5.1. Equivalence for Schauder bases. We begin with a simple observation.
Lemma 5.1.
(5.3)
µd
m = supnk1ηBk
k1εAk
:
B ≤ A ≤ m, A∩ B = /0, ε = η = 1o.
Proof. Let ε =η = 1 and B ≤A ≤ m with A∩B = /0. We must show that k1ηBk/k1εAk ≤
µd
m. Pick any set C disjoint with A∪ B such that B +C = A. We now use the elementary
inequality
2
+
x + y
x− y
kxk =(cid:13)(cid:13)(cid:13)
2 (cid:13)(cid:13)(cid:13) ≤ max{kx + yk,kx− yk},
(5.4)
with x = 1ηB and y = 1C. Let η′ ∈ ϒ be such that η′B = ηB and η′C = ±1, according to
the sign that reaches the maximum in (5.4). Then k1ηBk ≤ k1η′(B∪C)k ≤ µd
mk1εAk, and the
result follows.
(cid:3)
Theorem 5.2. If Kb is the basis constant and κ = supnke∗nkkenk, then
µm ≤ 2(Kb + 1) µd
(5.5)
Proof. Let A = B ≤ m, and ε = η = 1. Then
m + κ Kb.
k1εAk ≤ k1η(B\A)k
k1ηBk
k1εAk
m. We now bound II. Pick an integer n0 such that A1 = {n ∈
+ k1η(B∩A)k
k1εAk
= I + II.
Lemma 5.1 implies I ≤ µd
A : n ≤ n0} and A2 = A\ A1 satisfy
A1 = A2
(if A is even), or
Then
A1 = A− 1
2
= A2− 1 (if A is odd).
II ≤ k1η(B∩A1)k
k1εAk
+ k1η(B∩A2)k
k1εAk
≤ (Kb + 1)k1η(B∩A1)k
k1εA2k
+ Kbk1η(B∩A2)k
k1εA1k
= II1 + II2,
16
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
using in the second line the basis constant bound for the denominator. Since B∩ A1 ≤
A1 ≤ A2, we see that
II1 ≤ (Kb + 1) µd
m.
On the other hand, picking any number n1 ∈ B∩ A2, and using ke∗n1kk1εAk ≥ e∗n1(1εA) = 1,
we see that
II2 ≤ Kbk1η(B∩A2\{n1})k
+ Kbken1kke∗n1k ≤ Kb µd
m + κKb,
k1εA1k
the last bound due to B∩ A2\{n1} ≤ A2− 1 ≤ A1 and Lemma 5.1. Putting together the
previous bounds easily leads to (5.5).
(cid:3)
Remark 5.3. A similar argument shows the equivalence of the standard (unsigned) democ-
racy parameters
(5.6)
m = sup
A=B≤m
A∩B=/0
Indeed, in this case, the analog of (5.3) takes the weaker form
µm = sup
A=B≤m
and µd
k1Bk
k1Ak
.
k1Bk
k1Ak
(5.7)
µd
m ≤ sup
B≤A≤m
A∩B=/0
k1Bk
k1Ak ≤ Kbµd
m.
Then, (5.7) and the same proof we gave for Theorem 5.2 (with η = ε ≡ 1) leads to
(5.8)
m + κ Kb.
µm ≤ 2(Kb + 1)Kb µd
5.2. An example where µm grows faster than µd
to be new in the literature. As in (5.6), we denote by µm, µd
corresponding to constant signs.
m. The following example also seems
m the democracy parameters
Theorem 5.4. There exists a Banach space X with an M-basis B such that
lim sup
m→∞
µm
m]2−ε = lim sup
[ µd
m→∞
µm
m]2−ε = ∞, ∀ ε > 0.
[µd
Proof. Let N0 = 1, and define recursively Nk = 22Nk−1 , and N′k = N1 + . . . + Nk−1. Consider
the blocks of integers
Sk =(cid:8)N′k + 1, . . . , N′k + Nk(cid:9),
and denote the tail blocks by Tk = ∪ j≥k+1S j. Finally, let
and ∑
j∈Sk
Nk =(cid:8)(σj) j∈Sk : σj ∈ {±1}
σj = 0(cid:9).
We define a real Banach space X as the closure of c00 with the norm
kxk = maxn kxk∞, sup
k≥1
αk sup
σ∈Nk(cid:12)(cid:12)h1σSk, xi(cid:12)(cid:12), sup
k≥1
βk sup
S⊂Tk
S=Nk
∑
j∈Sx jo,
where the weights αk and βk are chosen as follows:
αk = 2−Nk−1 =
1
log2 Nk
and βk =
1
√Nk
.
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
17
Observe that
N′k = N1 + . . . + Nk−1 ≤ 2Nk−1 = 2 log2 log2 Nk
and
αk
βk
=
√Nk
log2 Nk
.
Claim 1:
µNk ≥ µNk ≥
Nk/2
(log2 Nk)plog2 log2 Nk
, for all k ≥ 1.
Proof. Pick any A ⊂ Sk ∪ Sk+1 such that A = Nk and A∩ Sk = A∩ Sk+1 = Nk/2. Then
k1Ak ≥ αk Nk/2 =
Nk/2
log2 Nk
.
Next, pick B = Sk, so that B = A = Nk and
k1Bk = maxn 1, αk · 0, sup
n≤k−1
Nk/2
Then µNk ≥ k1Ak/k1Bk ≥
(log2 Nk)√log2 log2 Nk
βnNno = βk−1Nk−1 =pNk−1 =plog2 log2 Nk.
.
(cid:3)
Nk ≤ µd
Nk ≤pNk, for all k ≥ 2.
Claim 2: µd
Proof. Let A, B be any pair of disjoint sets with A = B ≤ Nk, and let ε = η = 1. If
A = B ≤ √Nk, then the trivial bounds k1εAk ≤ A and k1ηBk ≥ 1 give
So, it remains to consider the cases √Nk < A = B ≤ Nk. We split A into three parts
k1εAk
k1ηBk ≤pNk.
A0 = A∩ Sk, A+ = A∩ Tk, A− = A∩ [S1 ∪ . . .∪ Sk−1].
Then, we have the following upper bound
k1εAk ≤ maxn1, sup
n<k
αnA−, αkA0, sup
n>k
αnNk, sup
n<k
βnNn, sup
n≥k
βnAo
≤ maxn N′k, αkA0, βkAo,
due to the elementary inequalities
• supn<k αnA− ≤ A− ≤ N′k
• supn>k αnNk = αk+1Nk = Nk2−Nk ≤ 1
• supn<k βnNn = √Nk−1 ≤ Nk−1 ≤ N′k
• supn≥k βnA = βkA.
Moreover, since βkA ≤ min{βkNk = √Nk, αkA }, we derive
k1εAk ≤ max{pNk,αkA0} and k1εAk ≤ max{N′k,αkA}.
(5.9)
We now give a lower bound for k1ηBk. The key estimate will rely on the following
Lemma 5.5. Let B0 = B∩ Sk and Bc
(5.10)
0 = Sk \ B0. Then
sup
σ∈Nk(cid:12)(cid:12)h1σSk, 1ηB0i(cid:12)(cid:12) ≥ min{B0,Bc
0}.
18
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
Proof. If B0 ≤ Nk/2, then we may select any σ ∈ Nk such that σB0 = η (which is possible
since Bc
0 ≥ B0), which gives
h1σSk, 1ηB0i = B0 = min{B0,Bc
0}.
0 = Nk − B0. Choose
Assume now that B0 > Nk/2. Pick any S ⊂ B0 with S = Bc
ν∈ {−1, 1}Bc
νi = 0. Choose τ∈ {−1, 1}B0\S so that ∑i∈B0\S τi = 0.
Replacing τ by −τ, if necessary, we may assume that ∑i∈B0\S τiηi ≥ 0. Finally, define
σ ∈ Nk by setting
0 so that ∑i∈S ηi + ∑i∈Bc
0
Then,
σS = ηS, σBc
0 = νBc
0, σB0\S = τB0\S.
h1σSk, 1ηB0i = ∑
i∈S
η2
i + ∑
i∈B0\S
τiηi ≥ S = Bc
0 = min{B0,Bc
0} . (cid:3)
From the lemma and the definition of the norm we see that
(5.11)
k1ηBk ≥ maxn1, αk min{B0,Bc
0}, βkB+o.
We shall finally combine the estimates in (5.9) and (5.11) to establish Claim 2. We distin-
guish two cases
Case 1: min{B0,Bc
0. Then, since A0 ⊂ Bc
0, we see that
0} = Bc
and therefore the first estimate in (5.9) gives
αkA0 ≤ αkBc
0 ≤ k1ηBk,
k1εAk
k1ηBk ≤
max{√Nk,k1ηBk}
k1ηBk
≤pNk.
Case 2: min{B0,Bc
0} = B0. Then, (5.11) reduces to
k1ηBk ≥ max(cid:8)αkB0, βkB+(cid:9) ≥ βkB0 +B+
≥ βkB/4,
since B− ≤ N′k ≤ √Nk/2 ≤ B/2, if k ≥ 2. Also, the second bound in (5.9) reads
since N′k ≤ √Nk/ log2 Nk = αk√Nk ≤ αkA, if k ≥ 2. Thus
4√Nk
log2 Nk ≤pNk.
= βkB−B−
k1εAk ≤ αkA,
k1εAk
k1ηBk ≤
αkA
βkB/4
4αk
βk
=
=
2
2
This establishes Claim 2.
From Claims 1 and 2 we now deduce that
and therefore
µNk
]2−ε ≥
[ µd
Nk
ε/2
k /2
N
(log2 Nk)plog2 log2 Nk → ∞,
lim sup
N→∞
µN
N]2−ε
[µd
= lim sup
N→∞
µN
N]2−ε
[ µd
= ∞.
(cid:3)
(cid:3)
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
19
6. NORM CONVERGENCE OF CGt
In this section we search for conditions in B = {en}∞
mx
mx AND G t
n=1 under which it holds
(6.1)
In [18, Theorem 1.1] this convergence is asserted for all "bases" {en, e∗n}∞
n=1 satisfying (a)-
(b)-(c). The proof however, does not seem complete, so we investigate here whether (6.1)
may be true in that generality.
kx− CGm(x)k → 0, ∀ x ∈ X.
The solution to this question requires the notion of strong M-basis; see [20, Def 8.4]. We
say that B is a strong M-basis if additionally to the conditions (a)-(d) in §1 it also holds
(6.2)
span{en}n∈A = (cid:8)x ∈ X :
supp x ⊂ A(cid:9), ∀ A ⊂ N.
Clearly, all Schauder or Cesàro bases (in some ordering) are strong M-bases; see e.g. [17]
for further examples. However, there exist M-bases which are not strong M-bases, see e.g.
[20, p. 244], or [10]1 for seminormalized examples in Hilbert spaces.
Lemma 6.1. If B is an M-basis which is not a strong M-basis, then there exists an x0 ∈ X
such that, for all Chebyshev greedy operators CGm,
(6.3)
lim inf
m→∞ kx0 − CGmx0k > 0.
Proof. If B is not a strong M-basis there exists some set A ⊂ N (necessarily infinite) and
some x0 ∈ X with supp x0 ⊂ A such that
δ = dist(x0, [en]A) > 0.
Since supp CGmx0 is always a subset of A, this implies (6.3).
(cid:3)
Remark 6.2. The above reasoning also implies that lim infmkx0− Gmx0k > 0, for all greedy
operators Gm. In particular, for M-bases which are not strong, the quasi-greedy condition
(6.4)
Cq := sup
Gm∈Gm
m∈N
kGmk < ∞
does not imply that Gmx converges to x for all x ∈ X. So the standard characterization in
[27, Theorem 1] needs the extra assumption that B is a strong M-basis.
A corrected version of [18, Theorem 1.1] (and also of "3 ⇒ 1" in [27, Theorem 1]) is the
following.
Proposition 6.3. If B is a strong M-basis then, for all Chebyshev t-greedy operators CGt
m,
(6.5)
lim
m→∞kx− CGt
mxk = 0, ∀ x ∈ X.
If additionally Cq < ∞, then for all t-greedy operators G t
m,
(6.6)
lim
m→∞kx− G t
mxk = 0, ∀ x ∈ X.
Proof. Given x ∈ X and ε > 0, by (6.2) there exists z = ∑n∈B bnen such that kx− zk < ε, for
some finite set B ⊂ supp x. Let α = minn∈Be∗n(x) and
1We thank V. Kadets for kindly providing this reference.
¯Λα = {n :
e∗n(x) ≥ α}.
20
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
Since α > 0, this is a finite greedy set for x which contains B. Moreover, we claim that
(6.7)
Indeed, if this was not the case there would exist n0 ∈ ¯Λα \ A, and since A is a t-greedy
set for x, then minn∈Ae∗n(x) ≥ te∗n0(x) ≥ tα. So, A ⊂ ¯Λtα, which is a contradiction since
m = A > ¯Λtα. Therefore, (6.7) holds and hence
mx =: A, ∀ m > ¯Λtα.
¯Λα ⊂ supp CGt
kx− CGt
mxk ≤ kx− ∑
n∈B
bnenk < ε,
∀ m > ¯Λtα.
This establishes (6.5).
We now prove (6.6). As above, let z = ∑n∈B bnen with B ⊂ supp x and kx − zk < ε.
Performing if necessary a small perturbation in the bn's, we may assume that bn 6= e∗n(x) for
all n ∈ B. Let now
α1 = min
and α = min{α1,α2} > 0.
n∈B e∗n(x), α2 = min
n∈B e∗n(x− z),
Consider the sets
¯Λtα = {n :
e∗n(x) ≥ tα} = {n :
e∗n(x− z) ≥ tα},
if m > ¯Λtα and A := supp G t
mx,
which for all t ∈ (0, 1] are greedy sets for both x and x− z, and contain B. We claim that,
(6.8)
The assertion ¯Λα ⊂ A is proved exactly as in (6.7). Next, we must show that
k /∈A e∗k(x).
if n ∈ A then e∗n(x− z) ≥ t max
¯Λα ⊂ A and A ∈ G(x− z, m, t).
k /∈A e∗k(x− z) = t max
then
This is clear if n ∈ A\ B since e∗n(x− z) = e∗n(x), and A ∈ G(x, m, t). On the other hand, if
n ∈ B, then e∗n(x− z) ≥ α2 ≥ α≥ maxk∈Ac e∗k(x), the last inequality due to ¯Λα ⊂ A. Thus
(6.8) holds true, and therefore
G t
m(x)− z = ∑
n∈A
e∗n(x− z)en = ¯G t
m(x− z),
for some ¯G t
m ∈ Gt
m. Thus,
kG t
m(x)− xk = k(I − ¯G t
m)(x− z)k ≤ (1 +k ¯G t
mk)ε,
and the result follows from supmk ¯G t
mk ≤ (1 + 4Cq/t)Cq, by [9, Lemma 2.1].
(cid:3)
REFERENCES
[1] F. ALBIAC AND J.L. ANSORENA, Characterization of 1-almost greedy bases. Rev. Matem. Compl. 30
(1) (2017), 13 -- 24.
[2] P. M. BERNÁ, Equivalence between almost-greedy and semi-greedy bases. J. Math. Anal. Appl. 417
(2019), 218 -- 225.
[3] P. M. BERNÁ, Ó. BLASCO, Characterization of greedy bases in Banach spaces. J. Approx. Theory 215
(2017), 28 -- 39.
[4] P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, Lebesgue inequalities for the greedy algorithm in general
bases. Rev. Mat. Complut. 30 (2017), 369 -- 392.
[5] P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, T. OIKHBERG, Embeddings and
Lebesgue-Type Inequalities for the Greedy Algorithm in Banach Spaces. Constr. Approx. 48 (3) (2018),
415 -- 451.
[6] Z. CIESIELSKI, A bounded orthonormal system of polygonals. Studia Math. 31 (1968) 339 -- 346.
LEBESGUE INEQUALITIES FOR CHEBYSHEV GREEDY ALGORITHMS
21
[7] S.J. DILWORTH, N.J. KALTON, D. KUTZAROVA, On the existence of almost greedy bases in Banach
spaces, Studia Math. 159 (1) (2003), 67 -- 101.
[8] S.J. DILWORTH, N.J. KALTON, D. KUTZAROVA, AND V.N. TEMLYAKOV, The Thresholding Greedy
Algorithm, Greedy Bases, and Duality, Constr. Approx. 19, (2003), 575 -- 597.
[9] S.J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, Lebesgue constants for the weak greedy algorithm,
Rev. Matem. Compl. 28 (2) (2015), 393 -- 409.
[10] L. N. DOVBYSH, N. K. NIKOLSKII, V. N. SUDAKOV, How good can a nonhereditary family be? J.
Sov. Math. 34 (6) (1986), 2050 -- 2060.
[11] G. GARRIGÓS, E. HERNÁNDEZ AND T. OIKHBERG, Lebesgue-type inequalities for quasi-greedy
bases, Constr. Approx. 38 (3) (2013), 447 -- 470.
[12] P. HAJEK, V. MONTESINOS SANTALUCÍA, J. VANDERWERFF, V. ZIZLER, Biorthogonal systems in
Banach spaces. Springer-Verlag 2008.
[13] Y. KATZNELSON, An introduction to Harmonic Analysis, 2nd ed. Dover Publ Inc, New York, 1976.
[14] S.V. KONYAGIN AND V.N. TEMLYAKOV, A remark on greedy approximation in Banach spaces, East.
J. Approx. 5 (1999), 365 -- 379.
[15] S.V. KONYAGIN AND V.N. TEMLYAKOV, Greedy approximatin with regard to bases and general mini-
mal systems, Serdica Math. J. 28 (2002), 305 -- 328.
[16] S. ROPELA, Properties of bounded orthogonal spline bases. In Approximation theory (Papers, VIth Se-
mester, Stefan Banach Internat. Math. Center, Warsaw, 1975), pp. 197 -- 205, Banach Center Publ., 4,
Warsaw, 1979.
[17] W.H. RUCKLE, On the classification of biorthogonal sequences. Canadian J. Math 26 (1974), 721 -- 733.
[18] C. SHAO, P. YE, Lebesgue constants for Chebyshev thresholding greedy algorithms, Journal of Inequal-
ities and Applications (2018), Paper No. 102, 23 pp.
[19] I. SINGER. Bases in Banach spaces I. Springer-Verlag, 1970.
[20] I. SINGER. Bases in Banach spaces II. Springer-Verlag, 1981.
[21] V. N. TEMLYAKOV, Greedy algorithm and n-term trigonometric approximation, Const.Approx. 14
(1998), 569 -- 587.
[22] V.N. TEMLYAKOV, Greedy approximation. Cambridge University Press, 2011.
[23] V. N. TEMLYAKOV, The best m-term approximation and greedy algorithms, Adv. Comput. 8 (1998),
249 -- 265.
[24] V. N. TEMLYAKOV, Sparse approximation with bases. Ed. by S. Tikhonov. Advanced Courses in Math-
ematics. CRM Barcelona. Birkhäuser-Springer, 2015.
[25] V. N. TEMLYAKOV, M. YANG, P. YE, Lebesgue-type inequalities for greedy approximation with respect
to quasi-greedy bases, East J. Approx 17 (2011), 127 -- 138.
[26] F. WEISZ, On the Fejér means of bounded Ciesielski systems. Studia Math. 146 (3) (2001), 227 -- 243.
[27] P. WOJTASZCZYK, Greedy Algorithm for General Biorthogonal Systems, Journal of Approximation
Theory, 107, (2000), 293 -- 314.
22
P. M. BERNÁ, Ó. BLASCO, G. GARRIGÓS, E. HERNÁNDEZ, AND T. OIKHBERG
PABLO M. BERNÁ, DEPARTAMENTO DE MATEMÁTICAS, UNIVERSIDAD AUTÓNOMA DE MADRID,
28049 MADRID, SPAIN
E-mail address: [email protected]
ÓSCAR BLASCO, DEPARTMENT OF ANALYSIS MATHEMATICS, UNIVERSIDAD DE VALENCIA, CAM-
PUS DE BURJASSOT, VALENCIA, 46100, SPAIN
E-mail address: [email protected]
GUSTAVO GARRIGÓS, DEPARTAMENTO DE MATEMÁTICAS, UNIVERSIDAD DE MURCIA, 30100 MUR-
CIA, SPAIN
E-mail address: [email protected]
EUGENIO HERNÁNDEZ, DEPARTAMENTO DE MATEMÁTICAS, UNIVERSIDAD AUTÓNOMA DE MADRID,
28049 MADRID, SPAIN
E-mail address: [email protected]
TIMUR OIKHBERG, DEPARTMENT OF MATHEMATICS, UNIVERSITY OF ILLINOIS URBANA-CHAMPAIGN,
URBANA, IL 61807, USA
E-mail address: [email protected]
|
0902.4869 | 2 | 0902 | 2010-11-02T11:04:41 | Higher rank numerical ranges of normal matrices | [
"math.FA",
"math-ph",
"math-ph",
"quant-ph"
] | The higher rank numerical range is closely connected to the construction of quantum error correction code for a noisy quantum channel. It is known that if a normal matrix $A \in M_n$ has eigenvalues $a_1, \..., a_n$, then its higher rank numerical range $\Lambda_k(A)$ is the intersection of convex polygons with vertices $a_{j_1}, \..., a_{j_{n-k+1}}$, where $1 \le j_1 < \... < j_{n-k+1} \le n$. In this paper, it is shown that the higher rank numerical range of a normal matrix with $m$ distinct eigenvalues can be written as the intersection of no more than $\max\{m,4\}$ closed half planes. In addition, given a convex polygon ${\mathcal P}$ a construction is given for a normal matrix $A \in M_n$ with minimum $n$ such that $\Lambda_k(A) = {\mathcal P}$. In particular, if ${\mathcal P}$ has $p$ vertices, with $p \ge 3$, there is a normal matrix $A \in M_n$ with $n \le \max\left\{p+k-1, 2k+2 \right\}$ such that $\Lambda_k(A) = {\mathcal P}$. | math.FA | math |
HIGHER RANK NUMERICAL RANGES OF NORMAL MATRICES
HWA-LONG GAU∗, CHI-KWONG LI† , YIU-TUNG POON‡ , AND NUNG-SING SZE§
Abstract. The higher rank numerical range is closely connected to the construction of quantum
error correction code for a noisy quantum channel. It is known that if a normal matrix A ∈ Mn
has eigenvalues a1, . . . , an, then its higher rank numerical range Λk(A) is the intersection of convex
polygons with vertices aj1 , . . . , ajn−k+1 , where 1 ≤ j1 < · · · < jn−k+1 ≤ n. In this paper, it is
shown that the higher rank numerical range of a normal matrix with m distinct eigenvalues can
be written as the intersection of no more than max{m, 4} closed half planes. In addition, given a
convex polygon P a construction is given for a normal matrix A ∈ Mn with minimum n such that
Λk(A) = P. In particular, if P has p vertices, with p ≥ 3, there is a normal matrix A ∈ Mn with
n ≤ max {p + k − 1, 2k + 2} such that Λk(A) = P.
AMS subject classifications. 15A60, 15A90, 47N50, 81P68
Key words. Quantum error correction, higher rank numerical range, normal matrices, convex
polygon
1. Introduction. Let Mn be the algebra of n× n complex matrices regarded as
linear operators acting on the n-dimensional Hilbert space Cn. The classical numerical
range of A ∈ Mn is defined and denoted by
W (A) = {x∗Ax ∈ C : x ∈ Cn with x∗x = 1},
which is a useful concept in studying matrices and operators; see [6].
In the context of quantum information theory, if the quantum states are repre-
sented as matrices in Mn, then a quantum channel is a trace preserving completely
positive map L : Mn → Mn with the following operator sum representation
r(cid:88)
j=1
(1.1)
where E1, . . . , Er ∈ Mn satisfy(cid:80)r
j=1 EjE∗
L(A) =
E∗
j AEj,
j = In. The matrices E1, . . . , Er are known
as the error operators of the quantum channel L. A subspace V of Cn is a quantum
error correction code for the channel L if and only if the orthogonal projection P ∈ Mn
i EjP = γijP for all i, j ∈ {1, . . . , r}; for example, see
with range space V satisfies P E∗
[7, 8, 9]. In this connection, for 1 ≤ k < n researchers define the rank-k numerical
range of A ∈ Mn by
Λk(A) = {λ ∈ C : P AP = λP for some rank-k orthogonal projection P},
∗ Department of Mathematics, National Central University, Chung-Li 320, Taiwan (Email: hl-
[email protected]). Research of Gau was supported by National Science Council of the Republic
of China.
† Department of Mathematics, College of William & Mary, Williamsburg, VA 23185 (Email:
[email protected]). Research of Li was supported by USA NSF. Li was also supported by the
William and Mary Plumeri Award. Li is an honorary professor of the University of Hong Kong and
an honorary professor of the Taiyuan University of Technology.
[email protected]). Research of Poon was supported by USA NSF.
‡ Department of Mathematics,
yt-
§ Department of Applied Mathematics, The Hong Kong Polytechnic University, Hung Hom, Hong
Kong (Email: [email protected]). Research of Sze was supported by the Hong Kong
Polytechnic University startup grant.
Iowa State University, Ames,
IA 50051 (Email:
1
and the joint rank-k numerical range of A1, . . . , Am ∈ Mn by Λk(A1, . . . , Am) to be
the collection of complex vectors (a1, . . . , am) ∈ C1×m such that P AjP = ajP for a
rank-k orthogonal projection P ∈ Mn. Evidently, there is a quantum error correction
code V of dimension k for the quantum channel L described in (1.1) if and only if
Λk(A1, . . . , Am) is non-empty for (A1, . . . , Am) = (E∗
r Er). Also, it is
easy to see that if (a1, . . . , am) ∈ Λk(A1, . . . , Am) then aj ∈ Λk(Aj) for j = 1, . . . , m.
When k = 1, Λk(A) reduces to the classical numerical range W (A).
1 E2, . . . , E∗
1 E1, E∗
Recently, interesting results have been obtained for the rank-k numerical range
and the joint rank-k numerical range; see [1, 2, 3, 4, 5, 11, 12, 13, 14, 16]. In particular,
an explicit description of the rank-k numerical range of A ∈ Mn is given in [14],
namely,
(1.2)
Λk(A) =
{µ ∈ C : e−iξµ + eiξµ ≤ λk(e−iξA + eiξA∗)},
(cid:92)
ξ∈[0,2π)
where λk(X) is the kth largest eigenvalue of a Hermitian matrix X.
In the study of quantum error correction, there are channels such as the ran-
domized unitary channels and Pauli channels whose error operators are commuting
normal matrices. Thus, it is of interest to study the rank-k numerical ranges of nor-
mal matrices. Although the error operators of a generic quantum channel may not
commute, a good understanding of the special case would lead to deeper insights and
more proof techniques for the general case.
Given S ⊆ C, let conv S denote the smallest convex subset of C containing S.
For a normal matrix A ∈ Mn with eigenvalues a1, . . . , an, it was conjectured in [3, 4]
that
(cid:92)
(1.3)
Λk(A) =
1≤j1<···<jn−k+1≤n
conv {aj1 , . . . , ajn−k+1},
which is a convex polygon including its interior (if it is non-empty). This conjecture
was confirmed in [14] using the description of Λk(A) in (1.2). In our discussion, a
polygon would always mean a convex polygon with its interior.
In this paper, we improve the description (1.3) of the rank-k numerical range
of a normal matrix. In particular, in Section 2 we show that for a normal matrix
A with m distinct eigenvalues, Λk(A) can be written as the intersection of no more
than max{m, 4} closed half planes in C. Moreover, if Λk(A) (cid:54)= ∅, then it is a polygon
with no more than m vertices. We then consider the "inverse" problem, namely, for
a given polygon P, construct a normal matrix A ∈ Mn with Λk(A) = P. In other
words, we study the necessary condition for the existence of quantum channels whose
error operators have prescribed rank-k numerical ranges.
It is easy to check that
Λk( A) = P if A = A ⊗ Ik with W (A) = P. Our goal is to find a normal matrix
A with smallest size so that Λk( A) = P. To achieve this, we give a necessary and
sufficient condition for the existence of a normal matrix A ∈ Mn so that Λk(A) = P
in terms of k-regular sets in C (see Definition 3.3). Furthermore, we show that the
problem of finding a desired normal matrix A is equivalent to a combinatorial problem
of extending a given p element set of unimodular complex numbers to a k-regular set.
We then give the solution of the problem in Section 4. As a consequence of our results,
if P is a polygon with p vertices, then there is a normal matrix A ∈ Mn with
n ≤ max{p + k − 1, 2k + 2}
2
such that Λk(A) = P. Moreover, this upper bound is best possible in the sense that
there exists P so that there is no matrix of smaller dimension with rank-k numerical
range equal to P.
Suppose A is normal. By (1.3), one can write Λk(A) as the intersection of (cid:0) n
2. Construction of higher rank numerical ranges. By (1.2), Λk(A) can be
obtained as the intersection of infinitely many closed half planes for a given A ∈ Mn.
k−1
In particular, it is well known that
(cid:1)
convex polygons so that Λk(A) is a polygon.
Λ1(A) = conv {a1, . . . , am}, where a1, . . . , am are the distinct eigenvalues of A.
There are nice interplay between the algebraic properties of A ∈ Mn and the
geometric properties of Λ1(A) = W (A). For instance, Λ1(A) is always non-empty;
Λ1(A) is a singleton if and only if A is a scalar matrix; Λ1(A) is a non-degenerate line
segment if and only if A is a non-scalar normal matrix and its eigenvalues lie on a
straight line. Unfortunately, these results have no analogs for Λk(A) if k > 1. First,
the set Λk(A) may be empty, see [13]; there are non-scalar matrices A such that Λk(A)
is a singleton; and there are non-normal matrices A such that Λk(A) is a line segment.
Even for a normal matrix A, it is not easy to determine whether Λk(A) is empty, a
point or a line segment without actually constructing the set Λk(A). Moreover, there
is no easy way to express the vertices of the polygon Λk(A) (if it is non-empty) in
terms of the eigenvalues of the normal matrix A as in the case of Λ1(A). Of course,
one can use (1.3) to construct Λk(A) for the normal matrix A, but the number of
polygons needed in the construction will grow exponentially for large n and k. In the
following, we will study efficient ways to generate Λk(A) for a normal matrix A ∈ Mn.
While it is difficult to use the eigenvalues of A to determine the set Λk(A), it turns
out that we can use half planes determined by the eigenvalues to generate Λk(A)
efficiently. In the following, we will focus on the following problem.
Problem 2.1. Determine the minimum number of half planes needed to con-
struct Λk(A) using the eigenvalues of the normal matrix A ∈ Mn.
As by-products, we will show that for a normal matrix A with m distinct eigen-
values, Λk(A) is either empty or is a polygon with at most m vertices. In fact, by
examining the location of the eigenvalues of A on the complex plane, one may further
reduce the number of half planes needed to construct Λk(A).
Suppose the eigenvalues of A ∈ Mn are collinear. Then by a translation, followed
by a rotation, we may assume that A is Hermitian with eigenvalues a1 ≥ . . . ≥ an.
Then we have Λk(A) = [an−k+1, ak]. So we focus on those normal matrices whose
eigenvalues are not collinear.
Let us motivate our result with the following examples, which can be verified by
using (1.3).
Example 2.2. Let A = diag (1, w, w2, . . . , wn−1) with w = e2πi/n. Then for
k ≤ n/2, we have Λk(A) = ∩n−1
(cid:26)
j=0 Hj, where
(cid:16)
z ∈ C : Re
Hj =
e− (2j+k)πi
n
(cid:17) ≤ cos
z
(cid:27)
,
kπ
n
and only a small part of conv {wj−1, wj−1+k} lies in Λk(A).
3
Λ2(A) with n = 9 in Example 2.2 Λ3(A) with n = 9 in Example 2.2
More generally, we have the following.
Example 2.3. Let a1, . . . , an be the eigenvalues of A ∈ Mn, with n ≥ 3. Suppose
conv {a1, . . . , an} = P is an n-sided convex polygon containing the origin in the inte-
rior. We may assume that a1, . . . , an are arranged in the counterclockwise direction
on the boundary of P. For j ∈ {1, . . . , n}, let Lj be the line passing through aj and
aj+k, where aj+k = aj+k−n if j + k > n, and Hj be the closed half plane determined
by Lj which does not contain a(cid:96) for j < (cid:96) < j + k. Then
n(cid:92)
j=1
Λk(A) =
Hj.
Note that each Hj in Example 2.3 contains exactly n − k + 1 eigenvalues of A.
The situation is more complicated if Λ1(A) is not an n-sided convex polygon for
the normal matrix A ∈ Mn.
Example 2.4. Suppose B = diag (1, i,−1,−i, 2, 2i,−2,−2i, 3, 3i,−3,−3i). One
can see from the figures that the eigenvalues 1, i,−1,−i are interior points of Λ2(B)
while these eigenvalues are the vertices of Λ3(B).
Λ2(B)
Λ3(B)
To deal with normal matrices A ∈ Mn as in Example 2.4 that Λ1(A) is not an
n-side convex polygon, we need to construct some half spaces using the eigenvalues of
the normal matrix A. To do this, we introduce the following. Given any two distinct
complex numbers a and b, let L(a, b) be the (directed) line passing through a and b.
4
The closed half plane
H(a, b) = {z ∈ C : Im(cid:0)(¯b − ¯a)(z − a)(cid:1) ≥ 0}
is called the left closed half plane determined by L(a, b). For example, H(0, i) = {z ∈
C : Re (z) ≤ 0} and H(i, 0) = {z ∈ C : Re (z) ≥ 0}. Remark that in Example 2.2, the
set Hj is indeed the closed half plane H(wj, wj+k). Note that L(a, b) (cid:54)= L(b, a). In
our discussion, it is sometimes convenient to write
H(a, b) = {z ∈ C : Re (e−iξz) ≤ Re (e−iξa)}
with ξ = arg(b − a) − π/2. Also, we use H0(a, b) to denote the left open half plane
determined by L(a, b), i.e., H0(a, b) = H(a, b) \ L(a, b).
We have the following result showing that for a normal matrix A ∈ Mn with m
distinct eigenvalues, Λk(A) can be written as the intersection of at most max{m, 4}
half spaces. Even without any knowledge about the final shape of the set Λk(A), one
can use m(m − 1) half spaces to generate Λk(A). Evidently, the construction is more
efficient than the construction using (1.2) or (1.3). Furthermore, we can conclude
that Λk(A) is either an empty set, a singleton, a line segment, or a non-degenerate
polygon with at most m vertices.
Theorem 2.5. Let A ∈ Mn be normal with distinct eigenvalues a1, . . . , am that
are not collinear. Let S be the set of index pairs (r, s) such that H(ar, as) contains at
least n − k + 1 eigenvalues (counting multiplicities) of A, and
S0 = {(r, s) ∈ S : H0(ar, as) contains at most
n − k − 1 eigenvalues (counting multiplicities) }.
Then
(2.1)
(cid:92)
(r,s)∈S
Λk(A) =
H(ar, as) =
(cid:92)
(r,s)∈S0
H(ar, as).
Moreover, Λk(A) can be written as intersection of at most max{m, 4} half planes
H(ar, as), with (r, s) ∈ S0.
Proof. In the first part of the proof, we assume that A ∈ Mn has n eigenvalues
a1, . . . , an. For notational simplicity, we write H(ar, as) = H(r, s), H0(ar, as) =
H0(r, s), and L(ar, as) = L(r, s) for any two distinct eigenvalues ar and as of A. For
each (r, s) ∈ S, since H(r, s) is convex and contains at least n − k + 1 eigenvalues of
A, by (1.3), we have
conv {aj1, . . . , ajn−k+1} ⊆ H(r, s).
1≤j1<···<jn−k+1≤n
(cid:92)
Λk(A) ⊆ (cid:92)
Λk(A) =
It follows that
(2.2)
H(r, s).
(r,s)∈S
To prove the reverse inclusion of (2.2), note that if z is a point not in Λk(A),
then z will lie outside a convex polygon which equals the convex hull of n − k + 1
eigenvalues of A. So, it suffices to show that the convex hull W of any n − k + 1
5
eigenvalues of A can be written as an intersection of half planes, W = ∩(cid:96)
j=1H(rj, sj)
for some (r1, s1), . . . , (r(cid:96), s(cid:96)) ∈ S. We consider the following three cases.
Case 1 Suppose W is a singleton. Then W = {ar} for some eigenvalue ar with
multiplicity at least n − k + 1. Since the eigenvalues of A are non-collinear, there are
eigenvalues as and at such that ar, as, and at are not collinear. Then
W = H(r, s) ∩ H(s, r) ∩ H(r, t) ∩ H(t, r).
Case 2 Suppose W is a non-degenerate line segment. In this case, W = conv {ar, as}
for some eigenvalues ar and as with ar (cid:54)= as. Since the eigenvalues of A are non-
collinear, there is another eigenvalue at such that ar, as, and at are not collinear.
Without loss of generality, we assume that at ∈ H(r, s). Otherwise, we interchange
ar and as. Then
W = H(r, s) ∩ H(s, r) ∩ H(s, t) ∩ H(t, r).
Case 3 Suppose W is a non-degenerate polygonal disk. We may relabel the eigenval-
ues of A and assume that W has vertices a1, . . . , aq arranged in the counterclockwise
direction, where q ≥ 3. For convenience of notation, we will let aq+1 = a1 and
H(q, q + 1) = H(q, 1). Then
(cid:92)
1≤t≤q
W =
H(t, t + 1).
needed in the intersection (cid:84)
Thus, the first equality in (2.1) is proved. To prove the second equality in (2.1), we
claim the following.
Claim For each (r, s) ∈ S \ S0, there exist two ordered pairs (r1, s1) and (r2, s2) in
S0 such that H(r1, s1) ∩ H(r2, s2) ⊆ H0(r, s).
Once the claim is proved, all the half planes H(r, s) with (r, s) ∈ S \ S0 are not
(r,s)∈S H(ar, as) and hence the second equality in (2.1)
holds.
To prove the claim, suppose (r, s) ∈ S \ S0. Then H0(r, s) contains at least
n − k eigenvalues of A. By a translation followed by a rotation, we may assume that
H(r, s) = {z ∈ C : Im z ≥ 0} and we can relabel the index of eigenvalues so that for
1 ≤ j ≤ n − 1, either Im aj > Im aj+1 or Im aj = Im aj+1 with Re aj ≥ Re aj+1. Let
U = conv {a1, . . . , an−k}
and V = conv {an−k+1, . . . , an}.
Then U and V are disjoint if an−k (cid:54)= an−k+1 or U ∩ V = {an−k} if an−k = an−k+1.
By the assumption, U ⊆ H0(r, s) and {ar, as} ⊆ V. Define the set
W = conv {ai − aj : 1 ≤ i ≤ n − k < j ≤ n} = {u − v : u ∈ U and v ∈ V},
which is a convex polygon. Note that W ⊆ {z ∈ C : Im (z) ≥ 0} since Im (ai− aj) ≥ 0
for all 1 ≤ i ≤ n − k < j ≤ n. By the facts that U and V can intersect at at most
one point, and the union U ∪ V cannot be contained in any line, the set W does not
lie in any line that passes through the origin, and the point 0 can only be either an
extreme point of W or is not in W. Under these conditions, one can find two extreme
points w1 and w2 in W with Im ( ¯w1w2) (cid:54)= 0 such that
(2.3)
Im ( ¯w1w) ≥ 0 ≥ Im ( ¯w2w)
for all w ∈ W.
6
Since w1 is an extreme point in W, there are eigenvalues as1 ∈ U and ar1 ∈ V such
that w1 = as1 − ar1. Then (2.3) gives
Im (¯as1 − ¯ar1)(u − ar1) ≥ 0
and Im (¯ar1 − ¯as1 )(v − as1) ≥ 0
for all u ∈ U and v ∈ V, and thus, U ⊆ H(r1, s1) and V ⊆ H(s1, r1). With the fact
that ar1 and as1 lie in the line L(r1, s1), the closed half plane H(r1, s1) contains at
least n − k + 1 eigenvalues of A while the open half plane H0(r1, s1) contains at most
n − k − 1 eigenvalues only. Therefore, (r1, s1) ∈ S0. By a similar argument, one can
show that there are eigenvalues ar2 ∈ U and as2 ∈ V such that w2 = ar2 − as2. Then
(2.3) yields
Im (¯ar2 − ¯as2)(u − as2) ≤ 0
and Im (¯as2 − ¯ar2)(v − ar2 ) ≤ 0
for all u ∈ U and v ∈ V, and thus, U ⊆ H(r2, s2) and V ⊆ H(s2, r2), and one can
conclude that (r2, s2) ∈ S0. Observe that the two lines L(r1, s1) and L(r2, s2) are
not parallel as Im (¯ar1 − ¯as1)(as2 − ar2) = Im ( ¯w1w2) (cid:54)= 0. Using the fact that the
two distinct eigenvalues ar and as are in V, which is contained in the intersection
H(s1, r1) ∩ H(s2, r2), one can conclude that the intersection H(r1, s1) ∩ H(r2, s2)
must lie in H0(r, s), the interior of H(r, s). Therefore, the claim holds.
Next, we turn to the last part of the Theorem. It is trivial that if Λk(A) is either
an empty set, a singleton, or a non-degenerate line segment, then only at most 4 half
planes are needed in the construction of Λk(A).
Suppose A has m distinct eigenvalues a1, . . . , am and Λk(A) is a non-degenerate
polygon. Let T be a minimal subset of S0 such that Λk(A) = ∩(r,s)∈T H(r, s). Since
T is minimal, the half planes H(r, s), (r, s) ∈ T , are all distinct. We may further
assume that for all (r, s) ∈ T , {a1, . . . , am}∩ L(ar, as) ⊆ conv {ar, as}. Since Λk(A) is
a non-degenerate polygon, for each 1 ≤ t ≤ m, there exist at most two pairs (r, s) ∈ T
such that t ∈ {r, s}. Therefore, T contains at most m ordered pairs.
Example 2.6. Let A = diag (0, 0, 1, 1, i). Then
Λ2(A) = [0, 1] = H(0, 1) ∩ H(1, 0) ∩ H(1, i) ∩ H(i, 0)
Λ3(A) = ∅ = H(1, 0) ∩ H(1, i) ∩ H(i, 0)
and the intersection of any 2 half planes H(ar, as) is non-empty. This example also
shows that one cannot replace max{m, 4} by m in the conclusion in Theorem 2.5.
Example 2.7. Let A = diag (1,−1, i,−i). Then
Λ2(A) = {0} = H(1,−1) ∩ H(−1, 1) ∩ H(i,−i) ∩ H(−i, i)
and Λ2(A) cannot be written as an intersection of less than 4 half planes H(ar, as).
Corollary 2.8. Suppose A ∈ Mn is normal such that W (A) is an n-sided
polygon containing the origin as its interior point. Let v1, . . . , vn be the vertices of
W (A) having arguments 0 ≤ ξ1 < ··· < ξn < 2π. If k < n/2, then Λk(A) is an n-sided
convex polygon obtained by joining vj and vj+k, where vj+k = vj+k−n if j + k > n.
By Theorem 2.5, it is easy to see that the boundary of Λk(A) are subsets of
the union of line segments of the form conv {ar, as} such that ar and as satisfy the
H(ar, as) condition. However, it is not easy to determine which part of the line
segment actually belong to Λk(A) as shown in Examples 2.2, 2.3, and 2.4. By Theorem
7
2.5, if the normal matrix A ∈ Mn has m distinct eigenvalues, we need no more than
max{m, 4} half planes H(ar, as) to generate Λk(A). Can one determine these half
planes effectively? We will answer this question by presenting an algorithm in Section
5 based on the discussion in this section.
3. Matrices with prescribed higher rank numerical ranges. We study the
following problem in this section.
Problem 3.1. Let k > 1 be a positive integer, and let P be a p-sided polygon in
C. Construct a normal matrix A with smallest size (dimension) such that Λk(A) = P.
If P degenerates to a line segment joining two points a1 and a2. Then the smallest
n to get a normal matrix with Λk(A) = P is n = 2k, if a1 and a2 are distinct and
n = k if a1 = a2. So we focus on the case when the polygon P is non-degenerate.
A natural approach to Problem 3.1 is to reverse the construction of Λk(A) in
Example 2.3. Suppose we have a non-degenerate p-sided polygon P, with vertices,
v1, . . . , vp.
Without loss of generality, we may assume that 0 lies in the interior of P and
the arguments of vj in [0, 2π) are arranged in ascending order. Our goal is to use
the support line Lj which passes through vj, vj+1 for j = 1, . . . , p, where vp+1 = v1,
to construct A = diag (a1, . . . , ap) such that Λk(A) = P. Note that if the desired
values a1, . . . , ap exist and are arranged in counter-clockwise direction, then (by proper
numbering) the line Lj will coincide with the line passing through aj and aj+k, where
aj+k = aj+k−p if j+k > p. Consequently, aj will lie at the intersection of Lj and Lj−k,
where Lj−k = Lj−k+p if j − k < 0. Consequently, there exists A = diag (a1, . . . , ap)
satisfying Λk(A) = P if the following hold.
(1) k < p/2.
(2) There exist a1, . . . , ap ∈ C such that
(2.a) Lj ∩ Lj−k = {aj} for j = 1, . . . , p,
(2.b) a1, . . . , ap have arguments ξ1 < . . . < ξp in the interval [ξ1, ξ1 + 2π) and
0 lie in the interior of their convex hull.
Note that by Theorem 2.5, A has the smallest dimension among all normal ma-
trices B such that Λk(B) = P.
Clearly, conditions (1) and (2a) are necessary in the above construction. From the
following example, one can see that the above construction also fails when condition
(2b) is not satisfied.
Example 3.2. Let P be the 5-sided polygon with vertices {v1, . . . , v5} = {2 +
i, 1 + 2i,−1 + 3i,−1 − i, 3 − i}, see the below figure.
The polygon P
8
Then, with k = 2, we have
{a1, . . . , a5} = {−1 + 4i, 7 − i,−1 + 7i, 4 − i, (5 + 5i)/3},
which does not satisfy the condition (2b). Clearly, for A = diag (a1, . . . , a5), Λ2(A)
lies in the convex hull of {a1, . . . , a5}, which does not contain P.
Conditions (2a) and (2b) motivate the following definition.
Definition 3.3. Let Ω = {z ∈ C : z = 1}. A subset Π = {α1, . . . , αm}, with
distinct α1, . . . , αm ∈ Ω, is k-regular if every semi-circular arc of Ω without endpoints
contains at least k elements in Π.
Given distinct α1, α2 ∈ Ω, α2/α1 = eiθ for a unique 0 < θ < 2π. Then [α1, α2] =
{eitα1 : 0 ≤ t ≤ θ} is the closed arc on Ω from α1 to α2 in the counterclockwise
direction. Also define the open arc
(α1, α2) = [α1, α2] \ {α1, α2}.
The value θ is called the length of these intervals. Suppose 1 ≤ k ≤ n and Π ⊆ Ω.
Then Π is k-regular if for each α ∈ Π, (α,−α) ∩ Π contains at least k elements.
Note that if Π = {eiξj : 1 ≤ j ≤ n} with distinct ξ1, . . . , ξn ∈ [0, 2π), then Π is
k-regular if and only if for each r = 1, . . . , n, there are 1 ≤ j1 < . . . < jk ≤ n such
that
eiξj1 , . . . , eiξjk ∈(cid:16)
eiξr , ei(ξr+π)(cid:17)
(3.1)
For this reason, a set {ξ1, . . . , ξn} ⊆ [0, 2π) of n distinct numbers is also called k-
regular if {eiξj : 1 ≤ j ≤ n} is k-regular as defined in Definition 3.3. For ξ, ξ(cid:48) ∈ [0, 2π),
[ξ, ξ(cid:48)] will denote the subset {t ∈ [0, 2π) : eit ∈ [eiξ, eiξ(cid:48)
]}; the intervals [ξ, ξ(cid:48)), (ξ, ξ(cid:48)]
and (ξ, ξ(cid:48)) will also be defined similarly.
In Example 2.2, a direct computation shows that for 1 ≤ r, k ≤ n,
.
(cid:26) 2kπ/n
2kπ/n − 2π
ξr+k − ξr =
if r + k ≤ n,
if r + k > n.
Therefore, the set {ξ1, . . . , ξn} is k-regular and Λk(A) is nonempty for 1 ≤ k <
n/2. Otherwise, the set {ξ1, . . . , ξn} is not k-regular and Λk(A) is either empty or a
singleton.
In the following, we need an alternate formulation of (1.2). For any d, ξ ∈ R,
consider the closed half plane
(3.2)
H(d, ξ) = {µ ∈ C : Re (e−iξµ) ≤ d},
and its boundary, which is the straight line
L(d, ξ) = ∂H(d, ξ) = {µ ∈ C : Re (e−iξµ) = d}.
(3.3)
For A ∈ Mn, let ReA = (A + A∗)/2. Then (1.2) is equivalent to
(cid:92)
Λk(A) =
ξ∈[0,2π)
H(λk(Re (e−iξA)), ξ).
The following result is easy to verify.
Proposition 3.4. Let A ∈ Mn and Λk(A) = ∩m
j=1H(dj, ξj) (cid:54)= ∅, where H(dj, ξj)
is defined as in (3.2) for some d1, . . . , dm ∈ R and distinct ξ1, . . . , ξm ∈ [0, 2π).
9
(a) We have 0 ∈ Λk(A) if and only if d1, . . . , dm ≥ 0, and 0 is an interior point
j=1H(rdj, ξj + ξ) and
(b) If µ = reiξ with r > 0 and ξ ∈ R, then Λk(µA) = ∩m
of Λk(A) if and only if d1, . . . , dm > 0.
Λk(A + µI) = ∩m
j=1H(dj + r cos(ξ − ξj), ξj).
Theorem 3.5. Suppose P =(cid:84)p
In connection to Problem 3.1, we have the following.
j=1 H(dj, ξj) is a non-degenerate p-sided polygon,
where H(dj, ξj) is defined as in (3.2) with d1, . . . , dp ∈ R and distinct ξ1, . . . , ξp ∈
[0, 2π). Let q be a nonnegative integer. The following two statements are equivalent.
(I) There is a (p + q) × (p + q) normal matrix A such that Λk(A) = P.
(II) There are distinct ξp+1, . . . , ξp+q ∈ [0, 2π) such that {ξ1, . . . , ξp+q} is k-regular.
j=1 H(dj, ξj) to be a non-degenerate
Notice that a necessary condition for the set(cid:84)p
polygon is that
{eiξ1 , . . . , eiξp} is 1-regular.
(3.4)
By Proposition 3.4, one may assume that 0 lies in the interior of P in our proofs.
However, it is equally convenient for us not to impose this assumption so that we
need not verify dj > 0 in H(dj, ξj) in our proofs.
To prove Theorem 3.5, we need some lemmas.
Lemma 3.6. Given A = diag (a1, . . . , an) and 1 ≤ m < n. Suppose the eigenval-
ues am+1, . . . , an are in Λk(A) but not extreme points of Λk(A). Then
Λk(diag (a1, . . . , am)) = Λk(A).
Proof.
It suffices to show that if an is in Λk(A) but not an extreme point of
Λk(A), then Λk(diag (a1, . . . , an−1)) = Λk(A).
Suppose an satisfy the above assumption. Clearly, Λk(diag (a1, . . . , an−1)) is a
subset of Λk(A). On the other hand, for any 1 ≤ j1 < . . . < jn−k ≤ n − 1, Λk(A) ⊆
conv {aj1 , . . . , ajn−k , an}. Since an is not an extreme point of Λk(A), it follows that
an lies in conv {aj1, . . . , ajn−k , an} but is not its extreme point. Therefore,
conv {aj1, . . . , ajn−k} = conv {aj1, . . . , ajn−k , an}.
Thus,
Λk(A) =
(cid:92){conv {aj1 , . . . , ajn−k+1} : 1 ≤ j1 < . . . < jn−k < jn−k+1 ≤ n}
⊆(cid:92){conv {aj1 , . . . , ajn−k , an} : 1 ≤ j1 < . . . < jn−k ≤ n − 1}
(cid:92){conv {aj1 , . . . , ajn−k} : 1 ≤ j1 < . . . < jn−k ≤ n − 1}
=
= Λk(diag (a1, . . . , an−1)).
The next lemma shows that if a convex polygon P is the intersection of half planes
H(dj, ζj) for j = 1, . . . , m, such that the set {ζ1, . . . , ζm} is "almost" k-regular (in
the sense that {ζ1, . . . , ζm} is k-regular if we count the multiplicity of each element
in the set), one may replace these half planes by n other half planes H( dj, ζj) for
j = 1, . . . , n, n ≥ m, with ζi (cid:54)= ζj for all i (cid:54)= j, such that { ξ1, . . . , ξn} is k-regular and
the boundary L( dj, ζj) of H( dj, ζj) touches the polygon P for each j = 1, . . . , n.
j=1H(dj, ζj) such that 0 ≤ ζ1 ≤ . . . ≤ ζm < 2π and
for each r = 1, . . . , m, there are 1 ≤ j1 < ··· < jk ≤ m such that ζj1 , . . . , ζjk ∈ (ζr, ζr+
Lemma 3.7. Suppose P = ∩m
10
π). For every n ≥ m, there exist d1, . . . , dn ∈ R and distinct ζ1, . . . , ζn ∈ [0, 2π) with
{ζ1, . . . , ζn} being k-regular such that P = ∩n
j=1H( dj, ζj) and P ∩ L( dj, ζj) (cid:54)= ∅ for
each j = 1, . . . , n.
Proof. Set ζ1 = ζ1, and for s ∈ {2, . . . , m}, let ζs = ζs if ζs−1 < ζs. For the
remaining values, we have ζs−1 = ζs and we can set
ζs−t1 = ζs−t1+1 = ··· = ζs = ··· = ζs+t2 < ζs+t2+1
for some t1 ≥ 1 and t2 ≥ 0. Let (cid:96) = min{j : ζj > ζs−t1 + π}, then we can replace ζs+j
by ζs+j = ζs+j + j for sufficient small j > 0 for j = −t1 + 1,−t1 + 2, . . . , 0, . . . , t2
such that
ζs−t1 < ζs−t1+1 < ··· < ζs < ··· < ζs+t2 < min{ζ(cid:96) − π, ζs+t2+1}.
After this modification, ζ1, . . . , ζm are distinct and {ζ1, . . . , ζm} is k-regular. If n > m,
pick distinct ζm+1, . . . , ζn ∈ [0, 2π) \ {ζ1, . . . , ζm}. Then {ζ1, . . . , ζn} also forms a k-
regular set. Finally, let dj = maxµ∈P Re
for j = 1, . . . , n. Clearly, we
have P ∩ L( dj, ζj) (cid:54)= ∅ and P ⊆ H( dj, ζj) for all j. By construction, {ζ1, . . . , ζm} ⊆
{ζ1, . . . , ζn} and P = ∩m
e−iζj µ
(cid:16)
(cid:17)
j=1H( dj, ζj).
We can now present the proof of Theorem 3.5.
Proof of Theorem 3.5. Let P =(cid:84)p
j=1H(dj, ζj) = ∩n
polygon, where d1, . . . , dp ∈ R and ξ1, . . . , ξp ∈ [0, 2π).
j=1 H(dj, ξj) be a non-degenerate p-sided
Suppose (I) holds. We may assume that A = diag (a1, . . . , ap+q) and Λk(A) = P.
By Lemma 3.6, one can remove the eigenvalues of A in Λk(A) that are not extreme
points of Λk(A) to get A ∈ Mn for some positive integer n ≤ p + q so that Λk(A) =
Λk( A). We have the following.
Claim There are f1, . . . , fn ∈ R and ζ1, . . . , ζn ∈ [0, 2π) such that Λk( A) =
∩n
j=1H(fj, ζj). Furthermore, for each r = 1, . . . , n, there exist 1 ≤ j1 < . . . < jk ≤ n
such that ζj1, . . . , ζjk ∈ (ζr, ζr + π).
j=1H( dj, ξj) for
Once the claim holds, Lemma 3.7 will ensure that Λk( A) = ∩p+q
some d1, . . . , dp+q ∈ R and a k-regular set { ξ1, . . . , ξp+q}, with
H(dj, ξj) = P = Λk(A) = Λk( A) = ∩p+q
j=1H( dj, ξj) .
p(cid:92)
j=1
Then {ξ1, . . . , ξp} ⊆ { ξ1, . . . , ξp+q}. Thus, we can take ξp+1, . . . , ξp+q ∈ [0, 2π) so that
{ξ1, . . . , ξp+q} = { ξ1, . . . , ξp+q}. Therefore, (II) holds.
For notational convenience, we assume that A = A in the claim so that every
eigenvalue of A is either an extreme point of Λk(A) or does not lie in Λk(A).
We first construct ζ1, . . . , ζn ∈ [0, 2π) and f1, . . . , fn ∈ R. For r = 1, . . . , n, let Γr
be the set containing all ξ ∈ [0, 2π) such that the closed half plane H(Re (e−iξar), ξ)
contains at least n− k + 1 eigenvalues of A. As ar is either an extreme point of Λk(A)
or not in Λk(A), there is a ζ ∈ [0, 2π) such that Re (e−iζar) ≥ λk(Re (e−iζA)) and
hence the half plane H(Re (e−iζar), ζ) contains at least n − k + 1 eigenvalues. Then
Γr is always nonempty. Furthermore, by the definition of Γr, the set Γr is an union
of closed arcs of Ω. Clearly,
P = Λk(A) ⊆ (cid:92)
ξ∈Γr
11
H(Re (e−iξar), ξ).
Also the above intersection, which containing P, is a non-degenerate conical region.
Then Γr is contained in some open semi-circular arc of Ω; otherwise, the above inter-
section of half planes is equal to the singleton {ar}. As Γr is a union of closed arcs in
some open semi-circular arc of Ω, there exists a unique ζr ∈ Γr such that
(3.5)
Γr ⊆ (ζr − π, ζr].
Let fr = Re (e−iζr ar) for 1 ≤ r ≤ n. We show that Λk(A) = (cid:84)n
Suppose T is a minimal subset of S0 such that Λk(A) =(cid:84)
j=1 H(fj, ζj).
(r,s)∈T H(ar, as). We may
further assume that for all (r, s) ∈ T , {a1, . . . , am} ∩ L(ar, as) ⊆ conv {ar, as}. For
each (r, s) ∈ T , write H(ar, as) = H(Re (e−iζar), ζ) with ζ = arg(as − ar) − π/2.
Then ζ ∈ Γr. We claim that ζ = ζr. Suppose not. By the above assumption on ar,
one can see that for a sufficiently small > 0, the half plane H(Re (e−iζar), ζ) with
ζ = ζ − will contain all eigenvalues of A that are in H(ar, as), i.e., ζ ∈ Γr. With
(3.5), we have
Λk(A) ⊆ H(Re (e−iζr ar), ζr) ∩ H(Re (e−iζar), ζ) ⊆ H0(ar, as) ∪ {ar}.
So L(ar, as) ∩ Λk(A) contains at most one point. But this contradicts the fact that
(r, s) is an element in the minimal subset T . Therefore, ζ = ζr. Then for each
(r, s) ∈ T , H(ar, as) = H(fr, ζr) and so
H(fj, ζj) ⊆ (cid:92)
H(ar, as) = Λk(A) ⊆ n(cid:92)
(r,s)∈T
H(fj, ζj).
j=1
n(cid:92)
Thus, Λk(A) =(cid:84)n
j=1
j=1 H(fj, ζj) and the first part of the claim holds.
To prove the second part of the claim, without loss of generality, we may assume
that ar = 0 and ζr = 0. Then fr = 0 and
H(fr, ζr) = H = {z ∈ C : Re (z) ≤ 0}.
Thus, the closed left half plane contains at least n − k + 1 eigenvalues of A. Suppose
that the closed right half plane −H contains eigenvalues aj1, . . . , ajh of A with ζjt (cid:54)= 0
for t = 1, . . . , g, and ζjt = 0 for t = g + 1, . . . , h for some g ≤ h. Fix a sufficiently
small > 0. We choose g + 1 ≤ (cid:96) ≤ h so that
Re (e−iaj(cid:96)) = max
g+1≤t≤h
Re (e−iajt).
Then {ajg+1, . . . , ajh} ⊆ H(Re (e−iaj(cid:96)), ). On the other hand, this closed half plane
H(Re (e−iaj(cid:96) ), ) also contains all eigenvalues of A that are in the left open half plane.
Thus, this closed half plane H(Re (e−iaj(cid:96) ), ) has at least n− g eigenvalues of A. On
the other hand by (3.5), /∈ Γj(cid:96) and so H(Re (e−iaj(cid:96) ), ) can have at most n − k
eigenvalues. Thus, we have g ≥ k.
Now for each t = 1, . . . , k, let dt = Re (ajt), then dt ≥ 0 and H ⊆ H( dt, 0).
Thus, the closed half plane H( dt, 0) contains at least n − k + 1 eigenvalues of A, i.e.,
ζr = 0 ∈ Γjt. Recall that ζjt (cid:54)= 0. By (3.5), one see that
ζr ∈ (ζjt − π, ζjt)
for t = 1, . . . , k.
Equivalently, ζj1 , . . . , ζjk ∈ (ζr, ζr + π). Thus, our claim is proved, and (II) holds.
12
Suppose now (II) holds, namely, there are distinct ξp+1, . . . , ξp+q such that
{ξ1, . . . , ξp+q} is k-regular. For j = p + 1, . . . , p + q, define
Then P ⊆ H(dj, ξj) and so
P =
dj = max
µ∈P Re (e−iξj µ).
p(cid:92)
j=1
H(dj, ξj) =
n(cid:92)
j=1
H(dj, ξj)
with n = p + q. By Lemma 3.7, we may assume that P ∩ L(dj, ξj) (cid:54)= ∅ for all
j = 1, . . . , n, and 0 ≤ ξ1 < ··· < ξn < 2π such that condition (3.1) holds. For each
r = 1, . . . , n, let
(cid:0)eiξr dr+k − eiξr+k dr
(cid:1)
ar =
i
sin(ξr+k − ξr)
and A = diag (a1, . . . , an). Then
Re (e−iξr ar) = dr
and Re (e−iξr+k ar) = dr+k.
Note that ar ∈ L(dr, ξr)∩L(dr+k, ξr+k) is the vertex of the conical region H(dr, ξr)∩
H(dr+k, ξr+k), which contains P. Therefore,
Re (e−iξr (ar − µ)) ≥ 0
and Re (e−iξr+k (ar − µ)) ≥ 0,
for all µ ∈ P. Since ξr+k ∈ (ξr, ξr + π), we have
µ∈P Re (e−iξµ)
(3.6)
Let µj ∈ L(dj, ξj)∩P for j = r, r + k. As ξr+k ∈ (ξr, ξr + π), we have µr = ar − ieiξr br
and µr+k = ar + ieiξr+k cr for some br, cr ≥ 0. Note that
for all ξ ∈ [ξr, ξr+k].
Re (e−iξar) ≥ max
Re (e−iξ(µr − ar)) = br sin(ξr − ξ) ≥ 0
for all ξ ∈ [ξr − π, ξr],
and
Re (e−iξ(µr+k − ar)) = cr sin(ξ − ξr+k) ≥ 0
Since {ξ1, . . . , ξn} is k-regular, it is easily seen that
for all ξ ∈ [ξr+k, ξr+k + π].
[0, 2π) \ [ξr, ξr+k] = [ξr − π, ξr) ∪ (ξr+k, ξr+k + π].
Therefore, for ξ ∈ [0, 2π) \ [ξr, ξr+k], we have
max{Re (e−iξ(µr − ar)), Re (e−iξ(µr+k − ar))} ≥ 0.
Moreover, we have
(3.7)
max{Re (e−iξµr), Re (e−iξµr+k)} ≥ Re (e−iξar).
Let ξ ∈ [0, 2π). Then ξ ∈ [ξs, ξs+1) for some s ∈ {1, . . . , n}.
It follows that
ξ ∈ [ξr, ξr+k] for r = s − k + 1, . . . , s, and ξ ∈ [0, 2π) \ [ξr, ξr+k] for other r. By (3.6)
and (3.7),
r∈{s−k+1,...,s} Re (e−iξar) ≥ max
µ∈P Re (e−iξµ) ≥
min
r /∈{s−k+1,...,s} Re (e−iξar).
max
13
Thus, λk(Re (e−iξA)) = minr∈{s−k+1,...,s} Re (e−iξar) and so
P ⊆ H(cid:0)λk(Re (e−iξA)), ξ(cid:1) .
Hence, P ⊆ Λk(A). Furthermore, if ξ = ξs, then Re (e−iξs as) = ds. Thus
λk(Re (e−iξs A)) =
r∈{s−k+1,...,s} Re (e−iξsar) ≤ ds.
min
It follows that
Λk(A) =
(cid:92)
⊆ (cid:92)
ξ∈[0,2π)
H(cid:0)λk(Re (e−iξA)), ξ(cid:1)
H(cid:0)λk(Re (e−iξsA)), ξs
1≤s≤n
(cid:1) ⊆ (cid:92)
1≤s≤n
H (ds, ξs) = P.
Thus, P = Λk(A).
By Theorem 3.5, Problem 3.1 is equivalent to the following combinatorial problem,
whose solution will be given in the next section.
Problem 3.8. Suppose {ξ1, . . . , ξp} ⊆ [0, 2π) is 1-regular. For k > 1, determine
the smallest nonnegative integer q so that {ξ1, . . . , ξp+q} is k-regular for some distinct
ξp+1, . . . , ξp+q ∈ [0, 2π).
4. Solutions for Problems 3.1 and 3.8. In this section, we give the solutions
for Problems 3.1 and 3.8. Given a non-empty set Π = {ξ1, . . . , ξp} ⊆ Ω, Problem 3.8
is equivalent to the study of smallest nonnegative integer q so that {ξ1, . . . , ξp+q} is
k-regular for some distinct ξp+1, . . . , ξp+q ∈ Ω. We have the following.
Theorem 4.1. Let k > 1 be a positive integer and Π be a p element subset of
Ω, including s pairs of antipodal points: {β1,−β1}, . . . ,{βs,−βs}, where p ≥ 3 and
s ≥ 0. Suppose Π is 1-regular but not k-regular and q is the minimum number of
points in Ω one can add to Π to form a k-regular set.
(a) If k ≥ p − s, then
(4.1)
q =
(b) If k < p− s, then q is the smallest nonnegative integer t such that one can remove
t non-antipodal points from Π to get a (k − t)-regular set. More precisely,
q =min{t ∈ N : Π \ {β1,−β1, . . . , βs,−βs} has a t-element
subset T such that Π \ T is (k − t)-regular}.
(4.2)
Consequently,
(4.3)
The inequality in (4.3) becomes equality if Π = {1, i,−1, α4, . . . , αp} where α4, . . . , αp
lie in the open lower half plane.
q ≤ min{2k + 2 − p, k − 1}.
Several remarks concerning Theorem 4.1 are in order.
If condition (a) in the
theorem holds, then the value q can be determined immediately. However, it is im-
portant to consider two cases depending on whether Π has pairs of antipodal points
as illustrated by the following.
14
(cid:26) 2k + 1 − p
2k + 2 − p
if s = 0,
if s > 0.
Example 4.2. Suppose S1 = {1, w, w2, w3} with w = e2iπ/5. Then α (cid:54)= −β for
any two elements α, β ∈ S1 and adding w4 to S1 results in a 2-regular set. Suppose
S2 = {1,−1, i,−i}. Then we need to add at least two points, say, z,−z ∈ Ω \ S2, to
get a 2-regular set.
Suppose condition (b) in the theorem holds. We can determine the value q by
taking t non-antipodal elements away from Π at a time and check whether the re-
sulting set is (k − t)-regular. The value q can then be determined in no more than
(cid:1) = 2p−2s steps. The success of reducing Problem 3.8 to a problem
(cid:0) p − 2s
i
(cid:80)p−2s
i=0
which is solvable in finite steps depends on Lemma 4.7 and Proposition 4.8.
It would be nice to have a simple formula for q in terms of p, k, s in case (b) of
the theorem. However, the following example show that the value q depends not only
on the values p,k,s, but also on the relative positions of the points in Π.
Example 4.3. Let S1 = {1, w, w2, w3, w4, w5} with w = e2πi/7 and S2 =
{z2, z3, z7, z8, z12, z13} with z = e2πi/15. Notice that both S1 and S2 contain 6 el-
ements and have no antipodal pairs. Furthermore, both of them are 2-regular but
not 3-regular. Clearly, adding w6 to S1 results a 3-regular set. However, as each of
the open arcs (z3,−z3), (z8,−z8) and (z13,−z13) contains only two elements of S2
while the intersection of this three open arcs is empty, at least two elements has to
be added to S2 to form a 3-regular set.
Note that our proofs are constructive; see Lemma 4.7 and Propositions 4.6 and
4.8. One can actually construct a subset Π(cid:48) ⊆ Ω with q elements so that Π ∪ Π(cid:48) is
k-regular.
By Theorem 4.1, we can answer Problems 3.1 and 3.8, and obtain some additional
information on their solutions. We will continue to use the notation H(d, ξ) defined
in (3.2) in the following.
Theorem 4.4. For Problem 3.1, if a p-sided polygon P is expressed as P =
j=1H(dj, ξj) for some d1, . . . , dp ∈ R and ξ1, . . . , ξp ∈ [0, 2π), then the minimum
∩p
dimension n for the existence of a normal matrix A ∈ Mn such that Λk(A) = P is
equal to p + q, where q is determined in Theorem 4.1. Moreover,
(4.4)
n ≤ max{2k + 2, p + k − 1}.
The inequality in (4.4) becomes equality if (ξ1, ξ2, ξ3) = (0, π/2, π) and ξ4, . . . , ξp lie
in (π, 2π).
We break down the proofs of Theorems 4.1 and 4.4 in several propositions. We
first give a lower bound for the number of elements in a k-regular set.
Furthermore, if S contains a pair of antipodal points {α,−α}, then n ≥ 2k + 2.
Proposition 4.5. Suppose S = {α1, . . . , αn} ⊆ Ω is k-regular. Then n ≥ 2k + 1.
Proof. For any r ∈ {1, . . . , n}, each of the open arcs (αr,−αr) and (−αr, αr)
contains k elements of S. Thus, n ≥ 2k + 1. For the last statement, if we take αr = α,
then together with α and −α, we see that n ≥ 2k + 2. The proof of the assertion is
complete.
As shown in Proposition 4.5, the existence of a pair of antipodal points {α,−α}
has implication on the size of a k-regular set Π. The next result together with Propo-
sition 4.5 show that the lower bound in (4.1) is best possible.
Proposition 4.6. Let k > 1 and Π is a p element subset of Ω containing s pairs
of antipodal points, where p ≥ 3 and s ≥ 0. If Π is 1-regular but not k-regular and
15
k ≥ p− s, then one can extend Π to a k-regular set by adding 2k + 1− p or 2k + 2− p
elements, depending whether s is zero.
Proof. Assume k ≥ p − s. Suppose first that s > 0. Let Π(cid:48)(cid:48) be a set containing
(k − p + s + 1) pairs of antipodal points such that Π(cid:48)(cid:48) ∩ Π is empty. Take
Π(cid:48) = Π(cid:48)(cid:48) ∪ −(Π \ {β1,−β1, . . . , βs,−βs}).
Then Π(cid:48) contains (2k + 2− p) elements. Furthermore, the set Π∪ Π(cid:48) contains exactly
k + 1 pairs of antipodal points hence it is k-regular. Thus, the result follows if s > 0.
Next, suppose s = 0. Without loss of generality, we may assume that 1 ∈ Π.
Hence, −1 ∈ Π(cid:48). We now modify Π(cid:48). We first delete the point −1 in Π(cid:48). Then for
all other points α ∈ Π(cid:48), we replace α by eiξα if α lies in the upper open half plane
P = {z ∈ C : Im (z) > 0}, and by e−iξα if α lies in the lower open half plane −P ,
with sufficiently small ξ > 0. Then we see that for every α ∈ Π∪ Π(cid:48), αP still contains
exactly k elements. Thus, Π ∪ Π(cid:48) is k-regular. Furthermore, the modified set Π(cid:48) has
one fewer point, i.e., Π(cid:48) has only 2k + 1 − p elements. The proof of is complete.
A referee pointed out that each (k − 1)-regular set can be enlarged to a k-regular
set by adding in not more than 2 extra elements. The following result shows that
sometimes 2 may not be the minimum number needed.
antipodal point. The following are equivalent.
Lemma 4.7. Let k > 1 and Π be a subset of Ω containing at least one non-
(a) One can add a point β /∈ Π so that Π ∪ {β} is k-regular.
(b) One can delete a non-antipodal point γ ∈ Π so that Π\{γ} is (k − 1)-regular.
Here, an element α ∈ Π is called a non-antipodal point of Π if −α /∈ Π.
Proof. Suppose first that (b) holds. Let P = {z ∈ C : Im (z) > 0}. Without
loss of generality, we may assume that γ = 1 is a non-antipodal point in Π. Suppose
Π \ {γ} = {eiθ1 , . . . , eiθp−1} such that
0 < θ1 < ··· < θm < π < θm+1 < ··· < θp−1 < 2π.
As Π \ {γ} is (k − 1)-regular, by Proposition 4.5, Π \ {γ} has p − 1 ≥ 2(k − 1) + 1
elements. Therefore, for every α ∈ Ω, the open half plane αP contains at least k − 1
elements in Π\{γ} and either P or −P contains at least k elements in Π\{γ}. Hence,
we have either m = k − 1 or k ≤ m ≤ p − k.
Choose β = eiθ where
(cid:26) max{π + θm, θp−1}/2
min{2π + θ1, π + θm+1}/2
θ =
if m = k − 1,
if k ≤ m ≤ p − k.
Now for every α (cid:54)= ±1, the open half plane αP contains at least k − 1 elements of
Π\{γ} and either γ or β. Hence, αP contains at least k elements of Π∪{β}. On the
other hand, when α = ±1, the open half plane αP contains either k elements of Π or
k − 1 elements of Π and β. Again, αP contains at least k elements of Π ∪ {β}. Thus,
(a) holds.
Conversely, suppose (a) holds.
If −β ∈ Π, then it is easy to see that the set
Π \ {−β} is (k − 1)-regular. From now, we assume that −β /∈ Π. Without loss of
generality, we may assume that β = −1. Furthermore, by replacing Π with the set
{ ¯ξ : ξ ∈ Π}, if necessary, we can assume that the number of elements in Π ∩ P is
greater or equal to the number of elements in Π ∩ (−P ). Under this assumption, the
upper open half plane must contain at least one non-antipodal point of Π.
16
Let γ be the non-antipodal point in Π such that 0 < arg(γ) ≤ arg(α) for all
non-antipodal points α ∈ Π. Then γ ∈ P . We show that Π \ {γ} is (k − 1)-regular.
Take any α ∈ Π \ {γ}. Suppose α ∈ βP ∪ γP . Then the open half plane αP can
contain at most one of points β and γ. As the open half plane αP contains at least k
elements of Π ∪ {β}, αP contains at least k − 1 elements of Π \ {γ}. Thus, Π \ {γ} is
(k − 1)-regular if Π\{γ} ⊆ βP ∪ γP . Now suppose (Π\{γ})\ (βP ∪ γP ) is nonempty
and let ω1, . . . , ωt be the points in this set. Notice that all of them lie in the upper
open half plane P . Therefore, we may assume that
0 < arg(ω1) < ··· < arg(ωt) < arg(γ) < π.
Also by the choice of γ, ω1, . . . , ωt cannot be non-antipodal points and hence the
points −ω1, . . . ,−ωt are in Π. Clearly, each open half plane ωjP contains at least k
elements of Π ∪ {β}. Notice that (wjP ) \ P contains exactly j elements of Π ∪ {β},
namely, −ω1, . . . ,−ωj−1 and β. Also the set P \ (wjP ) contains exactly j elements
of Π ∪ {β}, namely, ω1, . . . , ωj. It follows that the half plane wjP contains the same
number of elements of P ∪{β} as the upper half plane P . By Proposition 4.5, Π∪{β}
contains at least 2k + 2 elements. Then by assumption, the upper open half plane P
contains at least k + 1 elements of Π∪{β}. Thus, every open half plane wjP contains
at least k + 1 elements of Π ∪ {β} and it contains at least k − 1 elements of Π \ {γ}.
Therefore, Π \ {γ} is a (k − 1)-regular set and the assertion follows.
Applying the above lemma inductively (repeatedly), we have the following.
Proposition 4.8. Let k > 1 and Π is a p element subset of Ω containing s
pairs of antipodal points, where p ≥ 3 and s ≥ 0. Suppose p > 2s. For any positive
t ≤ min{k, p − 2s, p − 1}, the following are equivalent.
(a) One can add t points β1, . . . , βt /∈ Π so that Π ∪ {β1, . . . , βt} is k-regular.
(b) One can delete t non-antipodal points γ1, . . . , γt ∈ Π so that Π \ {γ1, . . . , γt}
is (k − t)-regular.
Proof. Clearly, the result holds for t = 1 by Proposition 4.7. Assume the
statement holds for all (cid:96) < t. Suppose Π ∪ {β1, . . . , βt} is k-regular. Let Π1 =
Π ∪ {β1}. Then Π1 ∪ {β2, . . . , βt} is k-regular and it follows from the assumption
that one can find t − 1 non-antipodal points γ1, . . . , γt−1 ∈ Π ∪ {β1} such that
Π1 \ {γ1, . . . , γt−1} is (k − t + 1)-regular. If β1 /∈ {γ1, . . . , γt−1}, by applying Lemma
4.7 to the set (Π \ {γ1, . . . , γt−1}) ∪ {β1}, one can find another non-antipodal point
γt ∈ Π1 \{γ1, . . . , γt−1} so that Π\{γ1, . . . , γt} is (k − t)-regular. On the other hand,
if β1 is one of the γj, say β1 = γ1, then Π\{γ2, . . . , γt−1} is (k− t + 1)-regular. In this
case, take an arbitrary element γt ∈ Π \ {γ2, . . . , γt−1} and apply Lemma 4.7 to the
set (Π\{γ2, . . . , γt})∪{γt}, one can find another non-antipodal point γt+1 so that the
set Π \ {γ2, . . . , γt+1} is (k − t)-regular. Then (b) follows. The proof of (b) implying
(a) can also be done by induction in a similar way.
Suppose k < p−s. Given a p element subset Π of Ω containing s pairs of antipodal
points, β1,−β1, . . . , βs,−βs with s > 0, which is not k-regular, the set obtained from
Π by deleting all p − 2s non-antipodal points is a (s − 1)-regular set. On the other
hand, if Π does not have any pair of antipodal points, then k ≤ p − 1 and one can
always delete k elements to form a 0-regular set.
In both cases, one see that the
following minimum always exist.
q =min{t ∈ N : Π \ {β1,−β1, . . . , βs,−βs} has a t-element
subset T such that Π \ T is (k − t)-regular}.
17
By Proposition 4.8, one can always add this minimum number q of points to Π to
form a k-regular set. Furthermore, this number q is optimal in the sense that one
cannot add fewer than q elements to do so.
By definition, q is a positive integer bounded above by min{k, p−2s}. The follow-
ing proposition gives more information about the minimum value (4.2) in Theorem
4.1.
Proposition 4.9. Using the notation in Theorem 4.1. If k < p − s, then the
value q in (4.2) exists and satisfies
(cid:26) k
q ≤
min{k − 1, p − 2s}
if (p,s) = (k+1,0) or (k + 2, 1),
otherwise.
Also q is bounded below by 2k + 1 − p or 2k + 2 − p, depending whether s is zero.
Furthermore, q = 2k + 1 − p if p ≤ k + 2 with s = 0 and q = 2k + 2 − p if p ≤ k + 3
with s > 0.
Proof. The lower bound can be seen easily from Proposition 4.5. Also the case
when (p, s) = (k + 1, 0) or (k + 2, 1) has already discussed. Now we assume that
(p, s) /∈ {(k+1, 0), (k+2, 1)}. Consider the case when s ≥ 2. Take t = min{k−1, p−2s}
and delete t non-antipodal elements in Π. Then the resulting set is (s − 1)-regular
and hence (k − t)-regular as k − t = max{1, k − p + 2s} ≤ s − 1. Thus, q ≤ t.
Next we consider the case when s = 1 and p ≥ k + 3. Let {α, −α} be the
pair of antipodal points in Π. Since Π is 1-regular, there are α1 ∈ (α, −α) ∩ Π and
α2 ∈ (−α, α) ∩ Π. Pick another k − 1 non-antipodal points α3, . . . , αk+1 in Π. The
set Π \ {α3, . . . , αk+1} containing {α1, α2, α,−α} is 1-regular. Then q ≤ k − 1.
Finally consider the case when s = 0 and p ≥ k + 2. We may assume that
Π = {eiξj : 1 ≤ j ≤ p} with 0 = ξ1 < ··· < ξp < 2π. Since Π is 1-regular, we can
choose (cid:96) such that ξ(cid:96) = max{ξj : 0 < ξj < π}. Then S = {ξ1, ξ(cid:96), ξ(cid:96)+1} is 1-regular.
Then any p − k + 1 subset of Π containing S is 1-regular. Thus, q ≤ k − 1.
Now we are ready to present the following:
Proof of Theorems 4.1 and 4.4. The assertions on q and n follows by Propo-
sitions 4.6 and 4.8. For the last assertion in Theorem 4.1, we see that in order to get
a k-regular set by adding q points to Π, we need to add at least k − 1 points eiξ, with
0 < ξ < π. If 2k + 2− p > k− 1, then p− 3 < k and we need to add an extra k− (p− 3)
points eiξ, with π < ξ < 2π, giving a total of k − 1 + k − (p − 3) = 2k + 2 − p points.
This proves the equality in (4.3). The equality in (4.4) now follows readily.
To close this section, let us illustrate our results by the following example.
Example 4.10. Let the polygon P = conv {1, w, w2, w3, w4, w5, w6, w9} with
w = e2πi/12, see the following.
The polygon P
18
j=1 H(dj, ξj) with d1 = ··· = d6 = cos π
(ξ1, . . . , ξ8) =
,
3π
12
,
5π
12
,
7π
12
,
9π
12
,
4 , and
12 , d7 = d8 = cos π
11π
12
15π
12
21π
12
,
,
.
Π = {α1, . . . , α8} =
πi
e
12 , e
3πi
12 , e
5πi
12 , e
7πi
12 , e
9πi
12 , e
11πi
12 , e
15πi
12 , e
21πi
12
(cid:19)
(cid:110)
(cid:111)
.
(cid:111)
Then P =(cid:84)8
Thus,
(cid:110)
(cid:111)
(cid:18) π
(cid:110)
12
e 3πi
12 , e 15πi
12
12
In particular, Π has two pairs of antipodal points, namely,
and
, i.e., p = 8 and s = 2. By Theorem 4.4 and Proposition 4.9, for k ≥ 5,
12 , e 21πi
e 9πi
a (2k + 2) × (2k + 2) normal matrix A can be constructed so that Λk(A) = P.
It remains to consider the cases for k ≤ 4. Clearly, Π is 2-regular. Thus, a
8 × 8 normal matrix A2 can be constructed so that Λ2(A2) = P. However, Π is not
k-regular for k ≥ 3.
12 } is 2-regular. Then Theorem
4.4 shows that there is a 9 × 9 normal matrix A3 such that Λ3(A3) = P. Indeed,
following the proof of Lemma 4.7, we see that if Π(cid:48) = {e 18πi
12 } is 2-
regular. Thus, Theorem 4.4 shows that there is a 10× 10 normal matrix A4 such that
Λ4(A4) = P.
Finally, we turn to the case when k = 4. Notice that Π \ {e 5πi
Now we consider the case k = 3. Clearly, Π \ {e 5πi
12 }, Π ∪ Π(cid:48) is 3-regular.
12 , e 7πi
In the following, we display the higher rank numerical ranges of A2, A3, and A4.
In the figures, the points "o" correspond to the vertices of the polygon while the
points "∗" correspond to the eigenvalues of the normal matrices.
Λ2(A2) = P
Λ3(A3) = P
Λ4(A4) = P
19
5. An algorithm. In this section, we further present a detail procedure for
constructing the rank-k numerical ranges of normal matrices based on the discussion
in Section 2.
Given a normal matrix A with m distinct eigenvalues a1, . . . , am, one can easily
construct Λk(A) through the following algorithms.
Basic Algorithm First construct the set S0. For each ordered pair (r, s) with r < s,
count the number of eigenvalues of A (counting multiplicities) in the open planes
H0(ar, as) and H0(as, ar).
1. If H0(ar, as) has at most n − k − 1 eigenvalues while H0(as, ar) has at most
2. If H0(as, ar) has at most n − k − 1 eigenvalues while H0(ar, as) has at most
k − 1 eigenvalues, then collect the index pair (r, s) in S0.
k − 1 eigenvalues, then collect the index pair (s, r) in S0.
Notice that one can already construct Λk(A) by determine the intersection of all
the half planes H(ar, as) with (r, s) ∈ S0. Nevertheless, one can perform the following
additional steps to simplify the set S0 before constructing Λk(A).
Modified Algorithm 1 Suppose in basic algorithm, there is an index pair (p, q)
satisfying both (1) and (2), i.e., both pairs (p, q) and (q, p) are in S0. Then Λk(A) is
a subset of a line segment. In this case, Λk(A) can be constructed as follows.
Set aj = (aj − ap)/(aq − ap) and define S1 = {(r, s) ∈ S0 : Im (ar) (cid:54)= Im (as)}. If
S1 = ∅, then Λk(A) = ∅. Suppose S1 (cid:54)= ∅. For each (r, s) ∈ S1, compute
brs =
Im (ar) Re (as) − Im (as) Re (ar)
Im (ar) − Im (as)
.
Take
b1 = max{brs : (r, s) ∈ S1, Im (ar) ≥ 0 and Im (as) ≤ 0},
b2 = min{brs : (r, s) ∈ S1, Im (ar) ≤ 0 and Im (as) ≥ 0}.
Then Λk(A) is the line segment in C joining the points (aq − ap)b1 + ap and (aq −
ap)b2 + ap if b1 ≤ b2; otherwise, Λk(A) = ∅.
Modified Algorithm 2 Assume the situation mentioned in modified algorithm 1
does not hold. Check if the set S0 satisfy the following.
There are (r1, s1), . . . , (r(cid:96), s(cid:96)) ∈ S0 with (cid:96) ≥ 3 such that
{r1, s1} ∩ {r2, s2} ∩ ··· ∩ {r(cid:96), s(cid:96)} = {t}
for some 1 ≤ t ≤ m.
(5.1)
If yes, define
(cid:26) arg(asj − t)
arg(t − arj )
θj =
if rj = t,
if sj = t.
Relabel the indices so that 0 ≤ θ1 ≤ ··· ≤ θ(cid:96) < 2π. Consider the following three
cases.
1. If θ(cid:96) − θ1 < π, remove the all pairs (rj, sj) in S0 for j (cid:54)= 1, (cid:96). Then check
2. If θk+1 − θk > π for some k, remove the all pairs (rj, sj) in S0 for j (cid:54)= k, k + 1.
again whether the modified set still satisfies (5.1).
Then check again whether the modified set still satisfies (5.1)
3. If the above two items are not satisfied, then Λk(A) is either the empty set or
the singleton set {at}. In this case, check whether at lies in H(ar, as) for all
(r, s) ∈ S0. If yes, Λk(A) is the singleton set; otherwise it is the empty set.
20
Finally, if the modified set S0 does not satisfy (5.1), then one can construct Λk(A) by
determine the intersection of all the half planes H(ar, as) with (r, s) in the modified
set S0.
Acknowledgment
This research began at the 2008 IMA PI Summer Program for Graduate Students,
where the second author is a lecturer and the third author is a co-organizer. The
support of IMA and NSF for the program is graciously acknowledged. The hospitality
of the colleagues at Iowa State University is deeply appreciated. The authors would
also like to thank the referees for some helpful comments.
REFERENCES
[1] M.D. Choi, M. Giesinger, J. A. Holbrook, and D.W. Kribs, Geometry of higher-rank
numerical ranges, Linear and Multilinear Algebra 56 (2008), 53-64.
[2] M.D. Choi, J.A. Holbrook, D. W. Kribs, and K. Zyczkowski, Higher-rank numerical ranges
of unitary and normal matrices, Operators and Matrices 1 (2007), 409-426.
[3] M.D. Choi, D. W. Kribs, and K. Zyczkowski, Higher-rank numerical ranges and compres-
sion problems, Linear Algebra Appl. 418 (2006), 828-839.
[4] M.D. Choi, D. W. Kribs, and K. Zyczkowski, Quantum error correcting codes from the
compression formalism, Rep. Math. Phys., 58 (2006), 77 -- 91.
[5] H.L. Gau, C.K. Li, and P.Y. Wu, Higher-Rank Numerical Ranges and Dilations, J. Op-
erator Theory, 63 (2010), 181 -- 189.
[6] R.A. Horn and C.R. Johnson, Topics in Matrix Analysis, Cambridge University Press,
Cambridge, 1991.
[7] E. Knill and R. Laflamme, Theory of quantum error-correcting codes, Phys. Rev. A 55
(1997), 900-911.
[8] E. Knill, R. Laflamme, and L. Viola, Theory of quantum error correction for general noise,
Phys. Rev. Lett. 84 (2000), 2525.
[9] D.W. Kribs, R. Laflamme, D. Poulin, and M. Lesosky, Operator quantum error correction,
Quant. Inf. & Comp., 6 (2006), 383-399.
[10] S.R. Lay, Convex Sets and Their Applications, Pure and Applied Mathematics, John
Wiley & Sons, Inc., New York, 1982.
[11] C.K. Li and Y.T. Poon, Generalized numerical ranges and quantum error correction, J.
Operator Theory, to appear.
e-preprint http://arxiv.org/abs/0812.4772.
[12] C.K. Li, Y.T. Poon, and N.S. Sze, Higher rank numerical ranges and low rank perturbation
of quantum channels, J. Math. Anal. Appl. 348 (2008), 843-855.
[13] C.K. Li, Y.T. Poon, and N.S. Sze, Condition for the higher rank numerical range to be
non-empty, Linear and Multilinear Algebra, 57(2009), 365-368.
[14] C.K. Li and N.S. Sze, Canonical forms, higher rank numerical ranges, totally isotropic
subspaces, and matrix equations, Proc. Amer. Math. Soc., 136 (2008), 3013-3023.
[15] M.A. Nielsen and I.L. Chuang, Quantum computation and quantum information, Cam-
bridge, New York, 2000.
[16] H. Woerdeman, The higher rank numerical range is convex, Linear and Multilinear Alge-
bra 56 (2008), 65-67.
21
|
1611.06560 | 1 | 1611 | 2016-11-20T18:14:03 | Some topics in the perturbation theory for a functional calculus of closed operators on Banach spaces | [
"math.FA"
] | The wark is a contribution to the functional calculus constructed by of the author and A. A. Atvinovskii of closed operators on Banach spaces. The calculus is based on Markov and related functions as symbols. Estimates of bounded perturbations of operator functions with respect to general operator ideal norms, operator Lipschitzness, moment inequality, Freshet operator differentiability, and Livshitz-Krein trace formulae have been considered. | math.FA | math |
Некоторые вопросы теории возмущений в функциональном исчислении
замкнутых операторов в банаховом пространстве
А. Р. Миротин
[email protected]
The work is a continuation of research of A.A. Atvinovskii and the author of
functional calculus of closed operators on Banach spaces based on Markov and
related functions as symbols. The following topics in the perturbation theory are
considered: estimates of bounded perturbations of operator functions with respect
to general operator ideal norms, operator Lipschitzness, moment inequality, Freshet
operator differentiability, analyticity of operator functions under consideration with
respect to the perturbation parameter, spectral shift function and Livshits -- Krein
trace formulae.
1 Введение и предварительные сведения
Различным проблемам теории возмущений линейных операторов, относящимся
к классическому функциональному исчислению фон Неймана-Мурье-Данфорда
самосопряженных, унитарных и нормальных операторов в гильбертовом про-
странстве, посвящена обширная литература (см., например, [1] (cid:22) [5]). Данная
работа посвящена аналогичным вопросам теории возмущений линейных опера-
торов, возникающим в функциональном исчислении замкнутых операторов в
банаховом пространстве, построенном в [6], [7] (см. также [8], [9]). Вместе с тем,
ее можно читать независимо от указанных работ.
В статье рассматриваются следующие вопросы теории возмущений: оценки
ограниченных возмущений значений рассматриваемых в работе операторных
функций в нормах общих операторных идеалов, операторная и коммутаторная
липшицевость таких функций, неравенство моментов, операторная дифферен-
цируемость по Фреше, аналитическая зависимость значений рассматриваемых
операторных функций от параметра возмущения, функция спектрального сдви-
га и аналог формулы следа Лифшица-Крейна для ядерных возмущений.
Единственной известной автору работой по операторно и коммутаторно лип-
шицевым функциям в контексте банаховых пространств является недавняя ста-
тья [10]1. В этой работе рассмотрено функциональное исчисление для опера-
торов скалярного типа в банаховых пространствах, использующее в качестве
символов довольно широкий класс функций, тесно связанный с пространством
Соболева W 1,2(R) и однородным пространством Бесова B1
∞,1(R). С помощью
развитой теории двойных операторных интегралов для операторов скалярно-
го типа в банаховых пространствах в указанной статье получены, в частности,
оценки коммутаторов весьма общего вида в операторной норме и нормах опе-
раторных идеалов, удовлетворяющих некоторым естественным ограничениям.
Напомним необходимые понятия и результаты из [6] и [7].
Определение 1 Пусть a < b. Говорят, что функция g относится к классу
1Автор благодарит рецензента, указавшего ему на эту работу.
1
Маркова R[a, b], если она голоморфна в верхней полуплоскости, отображает
ее в свое замыкание, а также голоморфна и положительна на (−∞, a) и го-
ломорфна и отрицательна на (b, ∞).
Скажем, что функция g принадлежит классу R0, если она принадлежит
R[0, b] и непрерывна в нуле.
Определение 2 Обозначим через ZR(0, b] (ZR[a, b]) множество функций ви-
да zg(z), где g ∈ R0 (соответственно, g ∈ R[a, b]).
В силу интегрального представления функций Маркова [11, теорема П.6],
функции из ZR(0, b] имеют в точности вид
f (z) =
b
Z
0
z
t − z
dτ (t),
(1)
где τ (cid:22) единственная ограниченная положительная мера на (0, b], такая, что
0 dτ (t)/t < ∞ ("представляющая мера"). Аналогичное интегральное представ-
ление имеют и функции из ZR[a, b].
R b
Следующая лемма дает внутреннюю характеристику функций из ZR(0, b].
Лемма 1 Функция f принадлежит ZR(0, b] тогда и только тогда, когда она
голоморфна и отрицательна на промежутках (−∞, 0) и (b, +∞), причем функ-
ция f (z)/z непрерывна в нуле и отображает открытую верхнюю полуплос-
кость {z ∈ C : Imz > 0} в свое замыкание.
Доказательство. Достаточно заметить, что указанные свойства функции f
равносильны тому, что функция g(z) := f (z)/z принадлежит классу R[0, b]
функций Маркова и непрерывна в нуле, т. е. принадлежит классу R0 в смысле
[7].
Аналогичным образом можно дать внутреннюю характеристику функций
из ZR[0, b].
Примеры 1. Функция
z
b − z (cid:18) z
z − b(cid:19)α−1
=
sin πα
π
b
Z
0
z
t − z
tα−1(b − t)−αdt
(0 < α < 1)
принадлежит ZR[0, b], но не принадлежит ZR(0, b];
функция
z(cid:18)1 −(cid:18) z
z − b(cid:19)α(cid:19) =
sin πα
π
b
Z
0
z
t − z (cid:18) t
b − t(cid:19)α
dt
(0 < α < 1)
принадлежит ZR(0, b] (см. [12, 2.2.6, формулы № 7, 5]).
2
Всюду ниже A (cid:22) замкнутый плотно определенный оператор в комплексном
банаховом пространстве X, σ(A) и ρ(A) (cid:22) спектр и резольвентное множество
оператора A соответственно, LB(X) (cid:22) алгебра линейных ограниченных опера-
торов в X.
Определение 3 Будем говорить, что (замкнутый плотно определенный) опе-
ратор A в пространстве X принадлежит классу V(0,b](X), если (0, b] ⊂ ρ(A)
и для некоторой постоянной MA > 0 выполняется неравенство
kR(t, A)k ≤
MA
t
, t ∈ (0, b].
Класс V(0,b](X) весьма широк. Например, ему принадлежат все операто-
ры вида −A, где A есть слабо позитивный оператор [13]. Укажем еще один
класс операторов из V(0,b](X). Напомним, что ограниченный оператор T в про-
странстве X называется оператором Ритта, если σ(T ) ⊂ D и kR(t, T )k ≤
C/t − 1 (t > 1) (см., например, [15], [16]).
Пример 2. Если оператор Ритта T в пространстве X имеет левый обратный
T −1, то T −1 + I ∈ V(0,1](X). В самом деле, если 1/t ∈ ρ(T ) (в частности, если
t < 1), то легко проверить, что оператор (−1/t)T R(1/t, T ) = (−1/t)R(1/t, T )T
равен R(t, T −1). Следовательно, (0, 1] ⊂ ρ(T −1 + I) и при t ∈ (0, 1] имеем
kR(t, T −1 + I)k = kR(t − 1, T −1)k ≤ k1/(t − 1)T R(1/(t − 1), T )k ≤ CkT k/t.
Определение 4 Для любой функции f ∈ ZR(0, b] с представляющей мерой τ
и любого A ∈ V(0,b](X) положим
b
AR(t, A)dτ (t).
f (A) =
Z
0
Из хорошо известного тождества
(2)
(3)
AR(t, A) = −I + tR(t, A)
(t ∈ ρ(A))
следует, что при A ∈ V(0,b](X) интеграл в (2) существует в смысле Бохнера
(относительно операторной нормы), оператор f (A) определен на X и ограничен,
причем kf (A)k ≤ (1 + MA)τ ([0, b]).
Ясно, что f (A) = Ag(A), где g(A) понимается в смысле [7], т. е. g(A) =
0 R(t, A)dτ (t). Поэтому данное исследование можно рассматривать как про-
должение работ [6] (cid:22) [7].
R b
2 Оценка возмущений ограниченными операто-
рами
Теорема 1 Пусть f ∈ ZR(0, b]. Для любых операторов A, B ∈ V(0,b](X), та-
ких, что D(A) ⊇ D(B) и оператор A − B ограничен, справедливо неравенство
kf (A) − f (B)k ≤ −(MA + MB + MAMB)f (−kA − Bk).
3
Доказательство. Ясно, что
f (A) − f (B) =
b
Z
0
(AR(t, A) − BR(t, B))dτ (t).
(4)
Рассмотрим подынтегральную функцию ϕ(t) := AR(t, A) − BR(t, B). Посколь-
ку D(A) ⊇ D(B), то в силу формулы (3) и второго резольвентного тождества
имеем ϕ(t) = t(R(t, A) − R(t, B)) = tR(t, A)(A − B)R(t, B), откуда kϕ(t)k ≤
MAMBkA − Bk/t. Но так как kϕ(t)k ≤ MA + MB, то
kϕ(t)k ≤ MAMBkA − Bk min(cid:26) MA + MB
MAMBkA − Bk
,
1
t(cid:27) .
Полагая в неравенстве
min(cid:26)a,
1
t(cid:27) ≤
1 + ad
t + d
,
справедливом для всех a, d, t > 0, a = (MA+MB)/(MAMBkA−Bk), d = kA−Bk,
получим
kϕ(t)k ≤
(MA + MB + MAMB)kA − Bk
t + kA − Bk
.
Теперь из (4) следует, что
kf (A) − f (B)k ≤ (MA + MB + MAMB)kA − Bk
b
Z
0
dτ (t)
t + kA − Bk
=
−(MA + MB + MAMB)f (−kA − Bk),
что и требовалось доказать.
Теорема 2 Если в условиях теоремы 1 операторы A − B и R(t, B) коммути-
руют, то при всех x ∈ D(A) справедливо неравенство
k(f (A) − f (B))xk ≤ −(MA + MB + MAMB)f (−k(A − B)xk).
Эта теорема доказывается так же, как и теорема 1.
Следующее неравенство по своей формулировке аналогично неравенствам
моментов, установленным в [13], [17], и есть частный случай теоремы 2 (при
B = O).
Следствие 1 [14] Пусть A ∈ V(0,b](X), f ∈ ZR(0, b]. Тогда для любого x ∈
D(A), kxk = 1 справедливо неравенство
kf (A)xk ≤ −(2MA + 1)f (−kAxk).
4
Следствие 2 Если A ∈ V(0,b](X), f ∈ ZR(0, b], то при всех x ∈ D(A), kxk = 1
kf (A)xk ≤ (2MA + 1)f ′(−0)kAxk.
Доказательство. Заметим, что функция f (−s)/(−s) убывает на множестве
s > 0. Поэтому
f ′(−0) = lim
s→+0
f (−s)
−s
≥
f (−s)
−s
(s > 0).
В частности, −f (−kAxk) ≤ f ′(−0)kAxk.
Далее (I, k · kI) обозначает операторный идеал в X, т. е. двусторонний иде-
ал алгебры LB(X) ограниченных операторов в пространстве X, полный отно-
сительно нормы k · kI, удовлетворяющей условиям kASBkI ≤ kAkkSkIkBk,
kSk ≤ kSkI при всех A, B ∈ LB(X) и S ∈ I (случай I = LB(X) не исключа-
ется и представляет интерес).
Следующая теорема показывает, что функции класса ZR(0, b] сохраняют
возмущения операторами из I и, в частности, являются I-операторно липши-
цевыми (см., например, [1]) в классе операторов V c
(0,b](X) := {A ∈ V(0,b](X) :
MA ≤ c} (c = const).
Теорема 3 Пусть f ∈ ZR(0, b]. Для любых операторов A, B ∈ V(0,b](X) та-
ких, что D(A) ⊇ D(B) и A − B принадлежит I, оператор f (A) − f (B) тоже
принадлежит I, и выполняется неравенство
kf (A) − f (B)kI ≤ MAMBf ′(−0)kA − BkI.
Доказательство. Пусть A, B ∈ V(0,b](X) таковы что D(A) ⊇ D(B) и A − B ∈
I. Как показано в доказательстве теоремы 1,
AR(t, A) − BR(t, B) = tR(t, A)(A − B)R(t, B), t ∈ (0, b],
(5)
Следовательно, этот оператор принадлежит I, и из (4) и (5) вытекает, что
kf (A) − f (B)kI ≤ MAMBkA − BkI
b
Z
0
dτ (t)
t
= MAMBf ′(−0)kA − BkI,
поскольку R b
0 dτ (t)/t = limz→−0 f (z)/z = f ′(−0).
Рассматриваемое функциональное исчисление обладает следующим свой-
ством устойчивости.
Следствие 3 Если An, Bn ∈ V(0,b](X), D(An) ⊇ D(Bn), причем An − Bn ∈ I и
kAn − BnkI → 0, MAn, MBn ≤ const, то kf (An) − f (Bn)kI → 0 (n → ∞).
Замечание 1 Данное функциональное исчисление обладает также следую-
щим свойством непрерывности. Если fn ∈ ZR(0, b], A ∈ V(0,b](X) и f ′
n(−0) →
0, то kfn(A)k → 0 (n → ∞).
5
В самом деле, kfn(A)k ≤ (1 + MA)τn([0, b]) → 0 (τn (cid:22) представляющая мера
для fn), поскольку
1
b
τn([0, b]) ≤
b
Z
0
dτn(t)
t
= lim
z→−0
fn(z)
z
= f ′
n(−0).
Теорема 3 может быть применена также для оценки норм коммутаторов.
Ниже через [A, B] обозначается коммутатор AB − BA операторов A и B. Ес-
ли этот оператор допускает продолжение на все X, то это продолжение тоже
обозначается [A, B].
Следствие 4 Пусть U есть автоморфизм пространства X, f ∈ ZR(0, b],
A ∈ V(0,b](X). Если [A, U ] ∈ I, то [f (A), U ] ∈ I и
k[f (A), U ]kI ≤ M 2
Af ′(−0)k[A, U ]kI.
Доказательство. Заметим, что R(t, U AU −1) = U R(t, A)U −1 (t ∈ (0, b]). От-
сюда следует, прежде всего, что U AU −1 ∈ V(0,b](X) и что MU AU −1 = MA. Более
того, f (U AU −1) = U f (A)U −1, а потому [f (A), U ] = (f (A) − f (U AU −1))U . При
этом A − U AU −1 = [A, U ]U −1 ∈ I. Тогда по теореме 3
k[f (A), U ]kI ≤ kf (A)−f (U AU −1)kI ≤ M 2
Af ′(−0)kA−U AU −1kI ≤ M 2
Af ′(−0)k[A, U ]kI.
Замечание 2 Как уже отмечалось во введении, коммутаторная липшице-
вость в контексте банаховых пространств являлась основным предметом
пионерской работы [10], где использовалось функциональное исчисление, осно-
ванное на развитой там технике двойных операторных интегралов для опе-
раторов скалярного типа. Поскольку операторы класса V(0,b](X) не являются,
вообще говоря, операторами скалярного типа, теорема 3 и следствие 4 суще-
ственно отличаются от результатов работы [10] как по рассматриваемым
классам операторов, так и по используемым классам функций.
3 Операторная дифференцируемость
В соответствии с [6], обозначим через V[a,b](X) множество всех замкнутых плот-
но определенных операторов в X, для которых [a, b] ⊂ ρ(A). Если A ∈ V[a,b](X), f (z) =
zg(z), где функция g ∈ R[a, b] имеет представляющую меру τ , то мы полагаем
f (A) := Z b
a
AR(t, A)dτ (t).
Функция f (z) = zg(z) голоморфна в окрестности спектра оператора A ∈ V[a,b](X)
и в бесконечно удаленной точке, а потому определено ее значение на операторе
A в смысле голоморфного функционального исчисления. С помощью теоремы
6
Фубини легко показать, что функциональное исчисление, определенное выше,
является частью голоморфного.
В связи с нижеследующим отметим, что если A ∈ V[a,b](X), то и A + ∆A ∈
V[a,b](X), если ∆A ∈ I, k∆AkI < δA, где
δA := min
ζ∈[a,b]
kR(ζ, A)k−1.
Это следует из [18, замечание IV.3.2].
Определение 5 Пусть f ∈ ZR[a, b], A ∈ V[a,b](X) и пусть I (cid:22) оператор-
ный идеал. Ограниченный оператор f ∇
A на I (трансформатор) называется
I-производной Фреше операторной функции f в точке A, если для ∆A ∈ I
справедливо асимптотическое равенство
kf (A + ∆A) − f (A) − f ∇
A (∆A)kI = o(k∆AkI) при k∆AkI → 0.
(7)
Теорема 4 Функция f ∈ ZR[a, b] является I-дифференцируемой по Фреше,
причем ее I-производная Фреше в точке A ∈ V[a,b](X) имеет вид
f ∇
A (B) =
b
Z
a
R(t, A)BR(t, A)tdτ (t) (B ∈ I).
(8)
Доказательство. Прежде всего, заметим, что трансформатор F (B), опре-
деляемый правой частью формулы (8), действительно ограничен, поскольку
kF (B)kI ≤ m2
a tdτ (t)kBkI, где
AR b
mA := max
t∈[a,b]
kR(t, A)k = 1/δA.
Далее, рассуждая, как в доказательстве теоремы 1, получаем для ∆A ∈ I,
таких, что A + ∆A ∈ V[a,b](X), равенство
f (A + ∆A) − f (A) − F (∆A) =
b
Z
a
(R(t, A + ∆A) − R(t, A))∆AR(t, A)tdτ (t). (9)
Известно (см., например, [18, теорема IV.1.16]), что при k∆AkkR(t, A)k < 1
kR(t, A + ∆A) − R(t, A)k ≤
k∆AkkR(t, A)k2
1 − k∆AkkR(t, A)k
.
(10)
Выберем ∆A таким, что k∆AkI ≤ 1/2mA. Тогда в силу (10)
kR(t, A + ∆A) − R(t, A)k ≤
k∆AkIkR(t, A)k2
1 − k∆AkIkR(t, A)k
≤ 2m2
Ak∆AkI.
7
Отсюда и из равенства (9) следует, что
kf (A + ∆A) − f (A) − F (∆A)kI = o(k∆AkI) при k∆AkI → 0,
что завершает доказательство.
Для формулировки следствия теоремы 4 введем на V[a,b](X) отношение эк-
вивалентности, считая операторы A и A′ из V[a,b](X) I-эквивалентными, если
A′ − A ∈ I. Очевидно, что формула dist(A, A′) = kA′ − AkI задает метрику в
каждом классе эквивалентности.
Следствие 5 Отображение A 7→ f ∇
операторов из V[a,b](X).
A непрерывно в каждом классе I-эквивалентности
Доказательство. Пусть операторы A и A′ из V[a,b](X) эквивалентны. По тео-
реме 4 для любого B ∈ I имеем
b
Z
a
R(t, A′)B(R(t, A′)−R(t, A))tdτ (t)+
b
Z
a
(R(t, A′)−R(t, A))BR(t, A)tdτ (t).
(f ∇
A′−f ∇
A )B =
Поэтому
k(f ∇
A′ − f ∇
A )BkI ≤ 2mA′mAkBkI
b
Z
a
kR(t, A′) − R(t, A)ktdτ (t).
Пусть ε > 0. Если оператор ∆A := A′ − A таков, что k∆AkI < min{1/2mA, ε},
то из формулы (10) следует, что kR(t, A′)−R(t, A)k ≤ mA, а потому mA′ ≤ 2mA.
Из формулы (10) следует также, что kR(t, A′) − R(t, A)k ≤ 2m2
Aε. Значит, если
dist(A, A′) < min{1/2mA, ε}, то
k(f ∇
A′ − f ∇
kf ∇
A′ − f ∇
b
Z
a
A
A )BkI ≤
8m4
A kLB(I) ≤
b
Z
a
A
8m4
tdτ (t)
εkBkI,
tdτ (t)
ε,
т. е.
что и требовалось доказать.
Для формулировки еще одного следствия теоремы 4 заметим, что для произ-
водной функции f ∈ ZR[a, b] с представляющей мерой τ справедливо равенство
f ′(z) =
tdτ (t)
(t − z)2 .
b
Z
a
8
Для функции f ∈ ZR[a, b] и оператора A ∈ V[a,b](X) положим
f ′(A) :=
b
Z
a
R(t, A)2tdτ (t).
Ясно, что оператор f ′(A) ограничен.
Следствие 6 Если операторы A и B коммутируют, то f ∇
A (B) = f ′(A)B.
Теорема 5 Если f ∈ ZR[a, b], A ∈ V[a,b](X), B ∈ I, то I-значная функция
z 7→ f (A + zB) − f (A) аналитична в окрестности нуля OA,B := {z ∈ C : z <
δA/kBk}, и в этой окрестности имеет место разложение
f (A + zB) − f (A) =
∞
Xn=1
znCn,
(11)
где
Cn =
1
n!
dn
dzn f (A + zB)(cid:12)(cid:12)(cid:12)(cid:12)z=0
=
b
Z
a
(R(t, A)B)nR(t, A)tdτ (t)
(12)
(производные понимаются в смысле нормы k · kI).
Доказательство. Если z ∈ OA,B, то, как было отмечено перед определением
5, A + zB ∈ V[a,b](X). Рассуждая, как при доказательстве теоремы 3, получим,
что f (A + zB) − f (A) ∈ I. Пусть h настолько мало, что z + h ∈ OA,B. Тогда,
заменяя в формуле (7) A на A + zB, а ∆A (cid:22) на hB и разделив на h, получим,
что существует
d
dz
f (A + zB) = f ∇
A+zB(B) =
b
Z
a
R(t, A + zB)BR(t, A + zB)tdτ (t)
(13)
(мы воспользовались теоремой 4). Следовательно, функция f (A + zB) − f (A)
аналитична в OA,B, и имеет место разложение (11), в котором Cn определяется
первым из равенств в (12). Второе равенство в (12) будем доказывать по ин-
дукции. При n = 1 оно верно в силу (13). Если предположить, что оно верно
для некоторого n, то
dn+1
dzn+1 f (A + zB)(cid:12)(cid:12)(cid:12)(cid:12)z=0
b
= lim
h→0
1
h (cid:18) dn
dzn f (A + zB)(cid:12)(cid:12)(cid:12)(cid:12)z=h
−
dn
dzn f (A + zB)(cid:12)(cid:12)(cid:12)(cid:12)z=0(cid:19) =
n! lim
h→0
1
h
Z
a
((R(t, A + hB)B)nR(t, A + hB) − (R(t, A)B)nR(t, A))tdτ (t) =
9
n! lim
h→0
b
1
h
Z
a
(R(t, A + hB)B)n(R(t, A + hB) − R(t, A))tdτ (t)+
b
((R(t, A + hB)B)n − (R(t, A)B)n)R(t, A))tdτ (t) = L1 + L2.
(14)
n! lim
h→0
1
h
Z
a
Заметим, что
k(R(t, A + hB)B)n(R(t, A + hB) − R(t, A)kI ≤
kR(t, A + hB)knkBkn
IkR(t, A + hB) − R(t, A)k.
При этом из (10) следует, что при h < 1/(2kBkImA) выполняется неравенство
kR(t, A+hB)−R(t, A)k < 1, а потому kR(t, A+hB)k < 1+mA. Следовательно,
k(R(t, A + hB)B)n(R(t, A + hB) − R(t, A)kI ≤ (1 + mA)nkBkn
I.
Из (10) следует также, что k(R(t, A + hB) − R(t, A)kI → 0 при h → 0. Отсюда
в силу второго резольвентного тождества выводим, что
1
h
lim
h→0(cid:13)(cid:13)(cid:13)(cid:13)
(R(t, A + hB) − R(t, A)) − R(t, A)BR(t, A)(cid:13)(cid:13)(cid:13)(cid:13)I
= 0.
(15)
Воспользовавшись теоремой Лебега для интеграла Бохнера, получаем теперь,
что
L1 = n! lim
h→0
b
1
h
Z
a
(R(t, A + hB)B)n(R(t, A + hB) − R(t, A))tdτ (t) =
n!
b
Z
a
(R(t, A)B)n+1R(t, A)tdτ (t).
(16)
При вычисления предела L2 будем исходить из тождества
Sn − T n =
n
Xk=1
Sn−k(S − T )T k−1,
справедливого для любых ограниченных операторов S и T в X. Полагая в
этом тождестве S = R(t, A + hB)B, T = R(t, A)B и разделив обе его части на
h, заключаем, что
lim
h→0(cid:13)(cid:13)(cid:13)(cid:13)
1
h
((R(t, A + hB)B)n − (R(t, A)B)n) − n(R(t, A)B)n+1(cid:13)(cid:13)(cid:13)(cid:13)I
= 0.
10
Отсюда следует, что
L2 = n! lim
h→0
b
1
h
Z
a
((R(t, A + hB)B)n − (R(t, A)B)n)R(t, A))tdτ (t) =
n!n
b
Z
a
(R(t, A)B)n+1R(t, A)tdτ (t).
(17)
Подставляя (16) и (17) в (14), получаем требуемое равенство. Теорема доказана.
4 Формула следа Лифшица-Крейна
В связи со следующей теоремой отметим, что формула следа для ядерных воз-
мущений самосопряженных операторов была доказана в случае конечномерных
возмущений в [19], а в общем случае (cid:22) в [20], [21]. Обзор более поздних иссле-
дований (в контексте гильбертовых пространств) и библиографию см. в [22]. Из
последних работ на эту тему см. [23]. Ниже мы вводим функцию спектрального
сдвига ξ для пары операторов класса V[a,b](X) и доказываем для рассматри-
ваемого нами функционального исчисления аналог формулы следа Лифшица-
Крейна. Напомним, что для банахова пространства X со свойством аппрокси-
мации существует непрерывный линейный функционал tr нормы 1 (след ) на
операторном идеале (S1, k · kS1) ядерных операторов, определенных на X (см.,
например, [24, с. 64]).
Теорема 6 Пусть пространство X обладает свойством аппроксимации, A, B ∈
V[a,b](X) D(A) ⊇ D(B), A − B ∈ S1. Тогда существует функция ξ, ана-
литическая в окрестности отрезка [a, b] и такая, что для любой функции
f ∈ ZR[a, b] справедливо равенство
tr(f (A) − f (B)) =
1
2πi Z
Γ
ξ(z)f ′(z)dz,
где контур Γ охватывает [a, b] и лежит в связной компоненте множества
ρ(A) ∩ ρ(B), содержащей [a, b].
Доказательство. Тот факт, что f (A) − f (B) ∈ S1, доказывается так же, как
и в теореме 3. При этом в силу аналога формулы (4) (в которой нижний предел
интегрирования заменен на a) и формулы (5) справедливо равенство
f (A) − f (B) =
b
Z
a
R(t, A)(A − B)R(t, B)tdτ (t),
11
причем интеграл в правой части существует в смысле Бохнера относительно
ядерной нормы. Следовательно,
tr(f (A) − f (B)) =
b
Z
a
ϕ(t)tdτ (t),
где ϕ(t) = tr(R(t, A)(A − B)R(t, B)). Последняя функция аналитична в одно-
связной окрестности отрезка [a, b], содержащейся в ρ(A) ∩ ρ(B). В самом деле,
поскольку отрезок [a, b] лежит в ρ(A) ∩ ρ(B), для некоторой окрестности лю-
бой точки z0 ∈ [a, b] справедливы абсолютно сходящиеся по норме оператора
m=0(z − z0)mBm с ограни-
разложения R(z, A) = P∞
n=0(z − z0)nAn, R(z, B) = P∞
ченными операторными коэффициентами. Поэтому
R(z, A)(A − B)R(z, B) =
∞
Xn=0
∞
(z − z0)n+mAn(A − B)Bm,
Xm=0
причем ряд, стоящий в правой части, абсолютно сходится в ядерной норме,
поскольку
k(z − z0)n+mAn(A − B)BmkS1 ≤ z − z0n+mkAnkkBmkk(A − B)kS1.
Если ξ есть первообразная функции ϕ в достаточно малой односвязной окрест-
ности отрезка [a, b] (первообразная существует, так как ϕ аналитична в такой
окрестности), то, как известно, по формуле Коши
ϕ(t) =
1
2πi Z
Γ
ξ(z)
(z − t)2 dz,
где контур Γ охватывает [a, b] и лежит в выбранной окрестности.
Тогда, в силу теоремы Фубини,
tr(f (A) − f (B)) =
1
2πi Z
Γ
ξ(z)
b
Z
a
tdτ (t)
(z − t)2 dz =
1
2πi Z
Γ
ξ(z)f ′(z)dz
(теорема Фубини применима, поскольку подынтегральная функция в соответ-
ствующем двойном интеграле непрерывна на компакте Γ × [a, b]), что и требо-
валось доказать.
Список литературы
[1] Александров А.Б., Пеллер В.В., Операторно липшицевы функции Preprint,
http://arxiv.org/abs/1602.07994v1 (2016).
[2] Kissin E., Shulman V.S., Classes of operator-smooth functions. I. Operator
Lipschitz functions, Proc. Edinburgh Math. Soc. 48 151(cid:22)173 (2005).
12
[3] Kissin E., Potapov D., Sukochev F., and Shulman V.S., Lipschitz functions,
Schatten ideals and unbounded derivations, Functional Analysis and its
Applications 45(2) 93 (cid:22) 96 (2011).
[4] P. J. Ayre, M. G. Cowling and F. A. Sukochev, Operator Lipschitz estimates in
the unitary setting, Proc. Amer. Math. Soc. 144 (3), 1053 -- 1057 (2016).
[5] E. Kissin, D. Potapov, V. Shulman and F. Sukochev, Operator smoothness
in Schatten norms for functions of severalvariables: Lipschitz conditions,
differentiabilityand unbounded derivations, Proc. London Math. Soc. 105 (4)
661 -- 702 (2012).
[6] Атвиновский А.А., Миротин А.Р., Об одном функциональном исчислении
замкнутых операторов в банаховом пространстве, Изв. вузов. Матем.
№ 10, 3 -- 15 (2013).
[7] Атвиновский А.А., Миротин А.Р., Об одном функциональном исчислении
замкнутых операторов в банаховом пространстве. II, Изв. вузов. Матем.,
№ 5, 3 -- 16 (2015).
[8] Атвиновский А.А., Миротин А.Р., Обращение одного класса операторов в
банаховом пространстве и некоторые его применения, Проблемы физики,
математики и техники, № 3(16), 55 -- 60 (2013).
[9] Атвиновский А.А., Миротин А.Р., Обращение линейной комбинации значе-
ний резольвенты замкнутого оператора, Проблемы физики, математики
и техники, № 3(20), 77 -- 79 (2014).
[10] J. Rozendaal, F. Sukochev, and A. Tomskova, Operator Lipschitz functions on
Banach spaces, Studia Mathematica 232 (1) 57 -- 92 (2016).
[11] Крейн М.Г., Нудельман А.А. Проблема моментов Маркова и экстремаль-
ные задачи (Наука, М., 1973)
[12] Прудников А.П., Брычков Ю.А., Маричев О.И. Интегралы и ряды: в 3 т.
- Т. 1: Элементарные функции (Наука, М., 1981).
[13] Пустыльник Е.И., О функциях позитивного оператора, Матем. сб.,
119(161), № 1(9), 32 -- 47 (1982).
[14] Миротин А.Р., Атвиновский А.А., О некоторых свойствах функциональ-
ного исчисления замкнутых операторов в банаховом пространстве, Про-
блемы физики, математики и техники, № 4, (2016).
[15] Lancien F., Le Merdy C., On functional calculus properties of Ritt operators,
preprint, arXiv:1301.4875v1.
[16] Lyubich Yu., Spectral localization, power boundedness and invariant subspaces
under Ritt's type condition, Studia Math. 134, 153 -- 167 (1999).
13
[17] Миротин А.Р., О некоторых свойствах многомерного функционального ис-
числения Бохнера-Филлипса, Сиб. матем. журн., 52 (6), 1300 -- 1312 (2011).
[18] Като Т. Теория возмущений линейных операторов (Мир, M., 1972).
[19] Лифшиц И. М., Об одной задаче теории возмущений, Успехи мат.наук, 7
(1), 171 -- 180 (1952).
[20] Крейн М. Г., О формуле следов в теории возмущений, Матем. сб., 33 (3),
597 -- 626 (1953).
[21] Крейн М. Г., Об определителях возмущения и формуле следов для унитар-
ных и самосопряженных операторов, ДАН СССР, 144 (2), 268 -- 271 (1962).
[22] Бирман М.Ш., Яфаев Д.Р., Функция спектрального сдвига. Работы М. Г.
Крейна и их дальнейшее развитие, Алгебра и анализ, 4, 1 -- 44 (1992).
[23] Peller V.V., The Lifshitz-Krein trace formula and operator Lipschitz functions,
Proc. Amer. Math. Soc. Published electronically: August 1, 2016. DOI:
http://dx.doi.org/10.1090/proc/13140.
[24] Defant A., Floret K. Tensor norms and operator ideals (North-Holland,
Amsterdam, 1993).
14
|
1005.1914 | 1 | 1005 | 2010-05-11T19:04:01 | Dimensions of l^p-cohomology groups | [
"math.FA",
"math.GR"
] | Let G be an infinite discrete group of type FP-infinity and let p>1 be a real number. We prove that the l^p-homology and cohomology groups of G are either 0 or infinite dimensional. We also show that the cardinality of the p-harmonic boundary of a finitely generated group is either 0, 1, or infinity. | math.FA | math |
DIMENSIONS OF ℓp-COHOMOLOGY GROUPS
MARK S. GRINSHPON, PETER A. LINNELL, AND MICHAEL J. PULS
Abstract. Let G be an infinite discrete group of type FP∞ and let 1 < p ∈ R.
We prove that the ℓp-homology and cohomology groups of G are either 0 or
infinite dimensional. We also show that the cardinality of the p-harmonic
boundary of a finitely generated group is either 0, 1, or ∞.
Let G be a discrete group and for 1 ≤ p ∈ R, let
1. Introduction
ℓp(G) = {Xx∈G
axx ax ∈ C and Xx∈G
axp < ∞}.
This is a complex Banach space with respect to the norm kf kp := (cid:0)Px∈G axp(cid:1)1/p
Also G acts on the left of ℓp(G) according to the rule gPx∈G axx = Px∈G axgx,
and similarly on the right according to the rule (cid:0)Px∈G axx(cid:1)g = Px∈G axxg. These
actions make ℓp(G) into a CG-bimodule. Suppose we are given a free resolution of
the trivial CG-module C with free right CG-modules:
.
(1.1)
Let
(1.2)
(1.3)
· · · −→ CGen+1 dn−→ CGen
dn−1−→ · · · d1−→ CGe1 d0−→ CG −→ C −→ 0.
n : HomCG(CGen , ℓp(G)) −→ HomCG(CGen+1 , ℓp(G)),
d∗
∗ : CGen+1 ⊗CG ℓp(G) −→ CGen ⊗CG ℓp(G)
dn
be the maps induced by dn; for convenience we write d∗
the usual (unreduced) cohomology and homology groups
−1 = d−1
∗ = 0. Then one has
H n(G, ℓp(G)) = ker d∗
Hn(G, ℓp(G)) = ker dn−1
∗
n/ im d∗
n−1,
/ im dn
∗ .
We will be interested in the case en < ∞ for all n, whence HomCG(CGen , ℓp(G)) ∼=
ℓp(G)en as left CG-modules and CGen ⊗CG ℓp(G) ∼= ℓp(G)en as right CG-modules.
(Recall that HomCG(CGen , ℓp(G)) consists of right CG-maps θ : CGen → ℓp(G) with
left G-action defined by (gθ)(α) = g(θα) for α ∈ CGen . Also CGen+1 ⊗CG ℓp(G) is
the tensor product of the right CG-module CGen+1 with the left CG-module ℓp(G)
with right G-action defined by (α ⊗ u)g = α ⊗ ug.)
Bekka and Valette [2, Corollary 8] proved that if G is a finitely generated group,
then H 1(G, ℓ2(G)) is either zero or infinite dimensional. The motivation behind
Date: Tue May 11 07:51:40 EDT 2010.
2000 Mathematics Subject Classification. Primary: 43A15; Secondary: 20J06.
Key words and phrases. projective resolution, group of type FPn, ℓp-cohomology, translation
invariant functional.
The research of the third author was partially supported by PSC-CUNY grant 62598-00 40.
1
2
M. S. GRINSHPON, P. A. LINNELL, AND M. J. PULS
this paper is to see if this result holds for arbitrary 1 < p ∈ R. It turns out that
this result remains true not only for 1 < p ∈ R, but also for the other homology and
cohomology groups. Recall that G is of type FPn over C if there exists a resolution
(1.1) which has ed finite for all d ≤ n, and G is of type FP∞ if it is of type FPn
for all n ∈ N. Furthermore G is of type FP1 over C if and only if G is finitely
generated. We shall prove
Theorem 1.4. Let d, n be non-negative integers and let G be an infinite group of
type FPn over C. Let 1 < p ∈ R. Then
(i) H d(G, ℓp(G)) is either 0 or has infinite C-dimension for all d ≤ n.
(ii) Hd(G, ℓp(G)) is either 0 or has infinite C-dimension for all d ≤ n.
Theorem 1.4 immediately yields
Corollary. Let G be an infinite group of type FP∞ over C and let d be a non-
negative integer. Let 1 < p ∈ R. Then
(i) H d(G, ℓp(G)) is either 0 or has infinite C-dimension.
(ii) Hd(G, ℓp(G)) is either 0 or has infinite C-dimension.
We deduce Theorem 1.4 from our main theorem:
Theorem 1.5. Let G be an infinite group, let m be a non-negative integer, and let
A ⊆ B be closed left G-invariant subspaces of ℓp(G)m. Then either A = B or B/A
has infinite dimension over C.
Of course, Theorem 1.5 remains true if we replace "left" with "right".
The layout of this paper is as follows. In Section 2 we give some definitions and
recall some well-known results. In Section 3 we prove Theorems 1.4 and 1.5. In
Section 4, we shall use Theorem 1.5 to obtain a result concerning the cardinality
of the p-harmonic boundary of a finitely generated group and to prove a result
about translation invariant functionals on a certain function space of functions on
a finitely generated group.
2. Preliminaries
We denote the positive integers by N. Let 1 ≤ p ∈ R. Then for α = Pg∈G αgg ∈
ℓ1(G) and β = Pg∈G βgg ∈ ℓp(G), we define convolution by
αβ = Xg,h∈G
αgβhgh = Xg∈G(cid:0)Xx∈G
αgx−1 βx(cid:1)g ∈ ℓp(G).
Young's inequality [5, 32D] tells us that
(2.1)
kαβkp ≤ kαk1kβkp.
Thus in particular ℓ1(G) is a ring with multiplication being convolution. Let m be
a non-negative integer. While different norms can be defined on a finite direct sum
of normed spaces, they are all equivalent [6, §1.8]. The most natural and consistent
choice for ℓp(G)m is using the p-norm:
m
k(u1, . . . , um)kp = (cid:0)
Xk=1
kukkp
p(cid:1)1/p
for all (u1, . . . , um) ∈ ℓp(G)m.
DIMENSIONS OF ℓp-COHOMOLOGY GROUPS
3
The inequality (2.1) still holds for u ∈ ℓ1(G) and v = (v1, . . . , vm) ∈ ℓp(G)m with
convolution defined componentwise, because
m
m
kuvkp
p = ku(v1, . . . , vm)kp
p =
(2.2)
kuvkkp
p ≤
Xk=1
Similarly for u ∈ ℓp(G) and v = (v1, . . . , vm) ∈ ℓ1(G)m, we have
1kvkkp
kukp
=
1
Xk=1
Xk=1
m
Xk=1(cid:0)kuk1kvkkp)p
p = kukp
kvkkp
m
p = kukp
1kvkp
p.
(2.3)
kuvkp
p ≤ kukp
pkvkp
1.
Note that (2.2) tells us that ℓp(G)m is a left ℓ1(G)-module, and that closed left
G-invariant subspaces of ℓp(G)m are left ℓ1(G)-submodules.
We shall also need the following two well-known results.
Lemma 2.4. Suppose n is a positive integer and for 1 ≤ k ≤ n we have bounded
linear operators Tk : B → B on a normed space B such that the range Tk(B) is
dense in B for each Tk. Then the range of T1 · · · Tn is also dense in B.
Proof. First we prove the claim for n = 2. Let ǫ > 0 and b ∈ B. There exists
b1 ∈ B such that kb − T1b1k < ǫ/2, and then there exists b2 ∈ B such that
kb1 − T2b2k ≤ ǫ
2kT1k . Thus
kb − T1T2b2k ≤ kb − T1b1k + kT1b1 − T1T2b2k
≤ kb − T1b1k + kT1kkb1 − T2b2k
<
ǫ
2
+ kT1k ·
ǫ
2kT1k
= ǫ.
The lemma now follows by induction on n.
(cid:3)
Lemma 2.5. Let T : A → B be a bounded linear operator between the Banach
spaces A and B. If T (A) has finite codimension in B, then T (A) is closed in B.
Proof. See [1, p. 95, Exercise (1), §3.4]
(cid:3)
3. Proof of the main theorems
The critical case in the proof of Theorem 1.5 is when G is infinite cyclic, and
the reader will understand most of the proof by studying this special situation. To
prove the result in general, we have had to repeat some arguments almost verbatim
several times. However we have chosen to give full details over brevity and clarity.
Proof of Theorem 1.5. We will assume that B/A is finite dimensional and will prove
that A = B. First suppose G has an element g of infinite order. Write H = hgi.
Note that for α ∈ ℓ1(G) and β ∈ ℓp(G)m, we have kαβkp ≤ kαk1kkβkp by (2.2),
thus in particular ℓp(G)m is a left ℓ1(H)-module, and A and B are left ℓ1(H)-
submodules. The action of g on the finite dimensional vector space B/A has a
minimal polynomial, i.e. there exists F (x) ∈ C[x] such that F (g) = 0 on B/A, and
therefore F (g)b ∈ A for all b ∈ B. Factor F (x) into linear factors and notice that if
ω 6= 1, then (g −ω) is invertible in ℓ1(H). Thus since A and B are ℓ1(H)-invariant,
we may assume that F (g) consists of factors (g − ω) with ω = 1 only. If we prove
that F (g)B is dense in B, that will imply that A = B.
4
M. S. GRINSHPON, P. A. LINNELL, AND M. J. PULS
Fix ω with ω = 1 and for n ∈ N, let xn = 1
k=1 ω−kgk. Note that
n Pn
(3.1)
kxnkp = (cid:18)n ·
1
np(cid:19)1/p
= n(1−p)/p,
consequently limn→∞ kxnkp = 0.
Now pick arbitrary b ∈ B and ǫ > 0. Since CGm is dense in ℓp(G)m, there
exists c ∈ CGm such that kb − ckp < ǫ/2. Then we may choose n ∈ N such that
kxnkp < ǫ
and we have
2kck1
kxnbkp = kxn(b − c) + xnckp ≤ kxn(b − c)kp + kxnckp
≤ kxnk1kb − ckp + kck1kxnkp by (2.2) and (2.3)
< 1 ·
ǫ
2
+ kck1 ·
ǫ
2kck1
= ǫ.
Thus kb − (1 − xn)bkp = kxnbkp < ǫ. Now note that the homomorphism CH →
C induced by the identity on C and sending g to ω has (1 − xn) in its kernel,
consequently we may write 1 − xn = (g − ω)d, where d ∈ CG, and we deduce that
(1 − xn)b ∈ (g − ω)B. Thus (g − ω)B is dense in B. Since the product of operators
with dense ranges has dense range by Lemma 2.4, we conclude that F (g)B is dense
in B.
sequence {g1, g2, . . . } of distinct elements of N and let xn = Pn
Therefore we may assume that every element of G has finite order. Let N denote
the kernel of the action of G on B/A. Suppose N is infinite. Choose an infinite
1
n gk ∈ CG. Let
b ∈ B, let ǫ > 0 and follow the argument above. Since kxnkp = n(1−p)/p, we see
that limn→∞ kxnkp = 0. Also CGm is dense ℓp(G)m, hence there exists c ∈ CGm
such that kb − ckp < ǫ/2. Then we may choose n ∈ N such that kxnkp < ǫ
and
we have
2kck1
k=1
kxnbkp = kxn(b − c) + xnckp ≤ kxn(b − c)kp + kxnckp
≤ kxnk1kb − ckp + kck1kxnkp by (2.2) and (2.3)
< 1 ·
ǫ
2
+ kck1 ·
ǫ
2kck1
= ǫ.
Thus kb − (1 − xn)bkp = kxnbkp < ǫ. Since (1 − xn)b ∈ A for all n ∈ N, we see that
A is dense in B and we conclude that A = B.
Therefore we may assume that N is finite, so G/N is an infinite torsion group,
and its action on B/A tells us that it is also a linear group over C. By a theorem
of Schur [4, cf. 1.L.4], there is a normal abelian subgroup K/N of finite index
in G/N . Since a simple C[K/N ]-module has dimension one over C, there is a
one-dimensional K-invariant subspace U/A of B/A. Again follow the proof above.
Choose an infinite sequence {g1, g2, . . . } of distinct elements of K/N . Then there
exist ω1, ω2, . . . ∈ C with ωi = 1 such that (gi − ωi)U ⊆ A. As before for n ∈ N,
set xn = Pn
k=1
1
n ω−1
k gk. Again (cf. (3.1)), kxnkp = n(1−p)/p → 0 as n → ∞.
Now pick arbitrary u ∈ U and ǫ > 0. Since CGm is dense in ℓp(G)m, there
exists c ∈ CGm such that ku − ckp < ǫ/2. Then we may choose n ∈ N such that
DIMENSIONS OF ℓp-COHOMOLOGY GROUPS
5
kxnkp < ǫ
2kck1
and we have
kxnukp = kxn(u − c) + xnckp ≤ kxn(u − c)kp + kxnckp
≤ kxnk1ku − ckp + kck1kxnkp by (2.2) and (2.3)
< 1 ·
ǫ
2
+ kck1 ·
ǫ
2kck1
= ǫ.
Since (1 − xn)u ∈ A for all n, we deduce that A is dense in U , a contradiction. This
completes the proof of Theorem 1.5.
(cid:3)
Deduction of Theorem 1.4 from Theorem 1.5. For Theorem 1.4(i), note that the
n : ℓp(G)en → ℓp(G)en+1 are continuous, because they are given by right
maps d∗
multiplication by an en × en+1 matrix with entries in CG. Thus if im d∗
n−1 has
finite codimension in ker d∗
n, it will be closed by Lemma 2.5. Theorem 1.4(i) now
follows from Theorem 1.5. The proof of Theorem 1.4(ii) is almost exactly the same,
except we need to deal with right G-invariant subspaces of ℓp(G)en .
(cid:3)
We can also prove results for the corresponding real Banach spaces. For a group
G and 1 ≤ p ∈ R, let
lp(G) = {Xx∈G
axx ax ∈ R and Xx∈G
axp < ∞}.
This is a real Banach space with respect to the norm kf kp := (cid:0)Px∈G axp(cid:1)1/p
Also G acts on the left of lp(G) according to the rule gPx∈G axx = Px∈G axgx.
Then we have
.
Corollary 3.2. Let G be an infinite group, let m be a non-negative integer, and
let A ⊆ B be closed left G-invariant subspaces of lp(G)m. Then either A = B or
B/A has infinite dimension over R.
Proof. We can regard A and B as closed G-invariant real subspaces of ℓp(G)m. Set
X = A + iA and Y = B + iB. Then X and Y are closed left G-invariant complex
subspaces of ℓp(G)m. Since either Y = X or Y /X has infinite C-dimension by
Theorem 1.5, we see that either B = A or B/A has infinite R-dimension and the
result follows.
(cid:3)
4. Applications to finitely generated groups
In this section we will use Corollary 3.2 to obtain some new results concern-
ing finitely generated infinite groups. Let F (G) denote the set of all real valued
functions on G. This has a left and right G-action given by (gf )(x) = f (g−1x)
and (f g)(x) = f (xg−1) for f ∈ F (G) and g, x ∈ G, respectively. We will view
lp(G) as {f ∈ F (G) Px∈G f (x)p < ∞}. To make this identification, we send
f to Px∈G f (x)x. Also l∞(G) = {f ∈ F (G) supx∈G f (x) < ∞} with norm
kf k∞ = supx∈G f (x). Finally RG will denote the functions in F (G) with finite
support.
Throughout this section, p will always denote a real number greater than one
and G will be a group with a finite symmetric generating set S (so S = S−1).
For a real-valued function f on G, we define the p-th power of the gradient, the
6
M. S. GRINSHPON, P. A. LINNELL, AND M. J. PULS
p-Dirichlet sum, and the p-Laplacian of g ∈ G by
Df (g)p = Xs∈S
Ip(f ) = Xg∈G
∆pf (g) = Xs∈S
f (g) − f (gs)p,
Df (g)p, and
f (gs) − f (g)p−2(f (gs) − f (g)),
In the case 1 < p < 2, we make the convention that f (gs) −
respectively.
f (g)p−2(f (gs) − f (g)) = 0 if f (gs) = f (g). We shall say that f is p-Dirichlet
finite if Ip(f ) < ∞. The set of all p-Dirichlet finite functions on G will be denoted
by Dp(G). A function f is said to be p-harmonic if ∆pf (g) = 0 for all g ∈ G. The
set HDp(G) will consist of the p-harmonic functions contained in Dp(G). We iden-
tify the constant functions on G with R. Observe that R is contained in HDp(G).
Endowed with the norm
kf kDp := (Ip(f ) + f (e)p)1/p,
Dp(G) is a reflexive Banach space, where e is the identity element of G and f ∈
Dp(G). For X ⊆ Dp(G), let X Dp indicate its closure in the Dp-norm and let B(X)
denote the bounded functions in X, that is X ∩ l∞(G); sometimes we will write
BX for B(X). The set BDp(G) is closed under the usual operations of scalar
multiplication and addition. Also BDp(G) is a reflexive Banach space under the
norm
kf kBDp := (Ip(f ))1/p + kf k∞,
where f ∈ BDp(G). Furthermore, kf hkBDp ≤ kf kBDpkhkBDp for f, h ∈ BDp(G).
Thus BDp(G) is an abelian Banach algebra. For Y ⊆ BDp(G), let Y BDp denote
its closure in the BDp-norm. Note that if f ∈ BDp(G), then kf kDp ≤ kf kBDp and
that B(RGDp ) = B(lp(G)Dp
) is a closed ideal in BDp(G).
Our first application of Corollary 3.2 will be concerned with the cardinality of
the p-harmonic boundary of G, which we now define. For a more detailed discussion
of this boundary, see [9]. Let Sp(BDp(G)) denote the set of complex-valued charac-
ters on BDp(G), that is nonzero ∗-homomorphisms from BDp(G) to C. Then with
respect to the weak ∗-topology, Sp(BDp(G)) is a compact Hausdorff space. Given
a topological space X, let C(X) denote the ring of continuous complex-valued func-
tions on X. The Gelfand transform defined by f (χ) = χ(f ) yields a monomorphism
of Banach algebras from BDp(G) into C(Sp(BDp(G))) with dense image. Further-
more the map ι : G → Sp(BDp(G)) given by (ι(g))(f ) = f (g) is an injection, and
ι(G) is an open dense subset of Sp(BDp(G)). The p-Royden boundary of G, which
we shall denote by Rp(G), is the compact set Sp(BDp(G)) \ ι(G). The p-harmonic
boundary of G is the following subset of Rp(G):
∂p(G) := {χ ∈ Rp(G) f (χ) = 0 for all f ∈ B(RGDp )}.
We can now state
Theorem 4.1. Let 1 < p ∈ R and let G be a finitely generated infinite group. Then
the cardinality of ∂p(G) is either 0, 1 or ∞.
Proof. Let S := {s1, . . . , sd} be a symmetric generating set for G. We will use the
results of [9] with G being the Cayley graph of G with respect to the generating set
DIMENSIONS OF ℓp-COHOMOLOGY GROUPS
7
S; thus the vertices of this graph are the elements of G, and g1, g2 ∈ G are joined
by an edge if and only if g1 = g2s for some s ∈ S. If 1 ∈ B(RGDp ), then ∂p(G) = ∅
by [9, Theorem 2.1 and Proposition 4.2]. Thus we will assume that 1 /∈ B(RGDp ).
If BHDp(G) = R, then [9, Theorem 4.11] says that ∂p(G) = 1. Now suppose
∂p(G) > 1. We will complete the proof of the theorem by showing that if ∂p(G)
is finite, then there exist two closed left G-invariant subspaces A and B of lp(G)d
that violate Corollary 3.2. We start by showing how Dp(G) is related to lp(G)d.
Define a continuous linear map Dp(G) → lp(G)d by θ(f ) = (f (s1−1), . . . , f (sd−1)).
Then ker θ = R and so θ induces an embedding θ′ : Dp(G)/R ֒→ lp(G)d. Clearly
Dp(G)/R is a Banach space under the norm induced by the norm on Dp(G), and
θ′ preserves this norm and also the left G-action.
We now construct the subspaces of lp(G)d that will give us our contradiction.
The Gelfand transform yields a homomorphism of BDp(G)/B(RGDp ) onto a dense
subspace of C(∂p(G)). Now [9, Theorem 4.9] shows that if f ∈ BDp(G), then
f ∈ B(RGDp ) if and only if f = 0 on ∂p(G). This tells us that if ∂p(G) < ∞, then
dimR(cid:0)BDp(G)/B(RGDp )(cid:1)= ∂p(G). Suppose 1 < ∂p(G) < ∞. Then
1 ≤ dimR(cid:0)BDp(G)/(B(RGDp ) ⊕ R)(cid:1) < ∞
and we deduce that
1 ≤ dimR(cid:0)(BDp(G) + RGDp )/R ⊕ RGDp(cid:1) < ∞.
It follows that θ′(cid:0)(BDp(G))+ RGDp /R(cid:1) is a subspace of lp(G)d properly containing
the finite codimensional closed subspace θ′(R ⊕ RGDp /R). Hence θ′(cid:0)(BDp(G) +
RGDp )/R(cid:1) is closed in lp(G)d. Thus with B = θ′(cid:0)(BDp(G) + RGDp )/R) and
A = θ′(R ⊕ RGDp /R), we obtain a contradiction from Corollary 3.2. So it is
impossible for 1 < ∂p(G) < ∞. Therefore ∂p(G) is either 0, 1 or ∞.
(cid:3)
We will now use Corollary 3.2 to obtain some results about translation invariant
linear functionals (which we define below) on Dp(G)/R. Let X be a normed space
of functions on G. For f ∈ X and h ∈ G, the left translation of f by h, denoted
by fh, is the function fh(g) := f (hg). Assume that if f ∈ X, then fh ∈ X for all
h ∈ G; that is, X is left translation invariant. We shall say that T is a translation
invariant left functional (TILF) on X if T (fh) = T (f ) for f ∈ X and all h ∈ G. For
the rest of this section translation invariant will mean left translation invariant. A
common question to ask is that if T is a TILF on X, then is T continuous? For
background about the problem of automatic continuity, see [7, 10, 12, 13]. Define
Diff(X) := linear span{fh − f f ∈ X, h ∈ G}.
It is clear that Diff(X) is contained in the kernel of any TILF on X. Observe that
f ∈ Dp(G) if and only if fh − f ∈ lp(G) for all h ∈ G. By definition we have the
following inclusions:
Diff(lp(G)) ⊆ Diff(Dp(G)/R) ⊆ lp(G) ⊆ Dp(G)/R.
The set Dp(G)/R is a Banach space under the norm induced from the norm on
Dp(G). The norm of f ∈ Dp(G)/R will be indicated by kf kD(p) and the closure of
a set Y ⊆ Dp(G)/R will be denoted by Y D(p). We can now state
Lemma 4.2. Diff(Dp(G)/R)D(p) = lp(G)D(p).
8
M. S. GRINSHPON, P. A. LINNELL, AND M. J. PULS
Proof. Let f ∈ lp(G). By [13, Lemma 1] there is a sequence (fn) in Diff(lp(G)) that
converges to f in the lp-norm. It now follows from Minkowski's inequality that for
s ∈ S,
k(f − fn)s − (f − fn)kp
p = Xg∈G
f (sg) − fn(sg) − (f (g) − fn(g))p → 0
as n → ∞. Hence f ∈ Diff(lp(G))D(p) which implies lp(G) ⊆ Diff(lp(G))D(p), and
the result follows.
(cid:3)
Combining the Hahn-Banach theorem with Lemma 4.2 and the fact lp(G)D(p) =
Dp(G)/R if and only if Dp(G)/lp(G) ⊕ R
Dp
= 0, we obtain the following
Corollary 4.3. Let 1 < p ∈ R. Then Dp(G)/lp(G) ⊕ R
exists a nonzero continuous TILF on Dp(G)/R.
Dp
6= 0 if and only if there
It was shown in [11] that if G is nonamenable, then the only TILF on lp(G) is
the zero functional. (Consequently every TILF is automatically continuous!) We
now show that this is not true for Dp(G)/R.
Proposition 4.4. Let 1 < p ∈ R and suppose that G is a finitely generated non-
amenable group. If Dp(G)/lp(G) ⊕ R
(Dp) 6= 0, then there is a discontinuous TILF
on Dp(G)/R.
Proof. It is known that lp(G) is closed in Dp(G)/R if and only if G is nonamenable
[3, Corollary 1]. Also Dp(G)/(lp(G) ⊕ R) is infinite dimensional by Corollary 3.2.
Let B be a Hamel basis for lp(G) and extend it to a Hamel basis H for Dp(G)/R.
Now H \ B corresponds to a Hamel basis of Dp(G)/(lp(G) ⊕ R). Select a countable
subset C := {fn n ∈ N} from H \ B. Define a linear functional T on Dp(G)/R
by T (f ) = 0 for f ∈ H \ C and T (fn) = nkfnkD(p) for n ∈ N. By [13, Lemma 1]
Diff(lp(G)) = lp(G), which implies Diff(Dp(G)/R) = lp(G). Thus T is a TILF on
nkfnkD(p)(cid:17) = 1 for all n.
Dp(G)/R. However (cid:16)
nkfnkD(p)(cid:17) → 0 in Dp(G)/R and T (cid:16)
Thus T is discontinuous on Dp(G)/R.
fn
fn
(cid:3)
The free group on two generators provides an example of a group that satisfies
Proposition 4.4, see [8, Corollary 4.3] for details.
If G is an infinite amenable
group, then by using an argument similar to the proof of Proposition 4.4, we see
that there always exists a discontinuous TILF on Dp(G)/R (the key point is that
Dp(G)/(lp(G) ⊕ R) will still be infinite dimensional).
References
[1] William Arveson. A short course on spectral theory, volume 209 of Graduate Texts in Math-
ematics. Springer-Verlag, New York, 2002.
[2] Mohammed E. B. Bekka and Alain Valette. Group cohomology, harmonic functions and the
first L2-Betti number. Potential Anal., 6(4):313 -- 326, 1997.
[3] A. Guichardet. ´Etude de la l-cohomologie et de la topologie du dual pour les groupes de Lie
`a radical ab´elien. Math. Ann., 228(3):215 -- 232, 1977.
[4] Otto H. Kegel and Bertram A. F. Wehrfritz. Locally finite groups. North-Holland Publishing
Co., Amsterdam, 1973. North-Holland Mathematical Library, Vol. 3.
[5] Lynn H. Loomis. An introduction to abstract harmonic analysis. D. Van Nostrand Company,
Inc., Toronto-New York-London, 1953.
[6] Robert E. Megginson. An introduction to Banach space theory, volume 183 of Graduate Texts
in Mathematics. Springer-Verlag, New York, 1998.
DIMENSIONS OF ℓp-COHOMOLOGY GROUPS
9
[7] Gary Hosler Meisters. Some problems and results on translation-invariant linear forms. In
Radical Banach algebras and automatic continuity (Long Beach, Calif., 1981), volume 975
of Lecture Notes in Math., pages 423 -- 444. Springer, Berlin, 1983.
[8] Michael J. Puls. The first Lp-cohomology of some finitely generated groups and p-harmonic
functions. J. Funct. Anal., 237(2):391 -- 401, 2006.
[9] Michael J. Puls. Graphs of bounded degree and the p-harmonic boundary. available via
http://arxiv.org/abs/0806.3073, 2008.
[10] Sadahiro Saeki. Discontinuous translation invariant functionals. Trans. Amer. Math. Soc.,
282(1):403 -- 414, 1984.
[11] G. A. Willis. Translation invariant functionals on Lp(G) when G is not amenable. J. Austral.
Math. Soc. Ser. A, 41(2):237 -- 250, 1986.
[12] G. A. Willis. Continuity of translation invariant linear functionals on C0(G) for certain locally
compact groups G. Monatsh. Math., 105(2):161 -- 164, 1988.
[13] Gordon S. Woodward. Translation-invariant linear forms on C0(G), C(G), Lp(G) for non-
compact groups. J. Functional Analysis, 16:205 -- 220, 1974.
Department of Mathematics and Statistics, Georgia State University, Atlanta, GA
30303, USA
E-mail address: [email protected]
URL: http://www2.gsu.edu/~wwwmat/faculty_staff/instructors/grinshpon.html
Department of Mathematics, Virginia Tech, Blacksburg, VA 24061-0123, USA
E-mail address: [email protected]
URL: http://www.math.vt.edu/people/plinnell/
Department of Mathematics, John Jay College -- CUNY, 445 West 59th Street, New
York, NY 10019, USA
E-mail address: [email protected]
|
1804.10350 | 1 | 1804 | 2018-04-27T06:02:18 | Weak$^*$-sequential properties of Johnson-Lindenstrauss spaces | [
"math.FA"
] | A Banach space $X$ is said to have Efremov's property ($\mathcal{E}$) if every element of the weak$^*$-closure of a convex bounded set $C \subseteq X^*$ is the weak$^*$-limit of a sequence in $C$. By assuming the Continuum Hypothesis, we prove that there exist maximal almost disjoint families of infinite subsets of $\mathbb{N}$ for which the corresponding Johnson-Lindenstrauss spaces enjoy (resp. fail) property ($\mathcal{E}$). This is related to a gap in [A. Plichko, Three sequential properties of dual Banach spaces in the weak$^*$ topology, Topology Appl. 190 (2015), 93--98] and allows to answer (consistently) questions of Plichko and Yost. | math.FA | math |
WEAK∗-SEQUENTIAL PROPERTIES OF
JOHNSON-LINDENSTRAUSS SPACES
ANTONIO AVIL´ES, GONZALO MART´INEZ-CERVANTES, AND JOS´E RODR´IGUEZ
Abstract. A Banach space X is said to have Efremov's property (E) if every
element of the weak∗-closure of a convex bounded set C ⊆ X ∗ is the weak∗-
limit of a sequence in C. By assuming the Continuum Hypothesis, we prove
that there exist maximal almost disjoint families of infinite subsets of N for
which the corresponding Johnson-Lindenstrauss spaces enjoy (resp. fail) prop-
erty (E). This is related to a gap in [A. Plichko, Three sequential properties of
dual Banach spaces in the weak∗ topology, Topology Appl. 190 (2015), 93 -- 98]
and allows to answer (consistently) questions of Plichko and Yost.
1. Introduction
A Banach space X is said to have
(i) weak∗-angelic dual if every element of the weak∗-closure of a bounded set
B ⊆ X ∗ is the weak∗-limit of a sequence in B;
(ii) Efremov's property (E) if every element of the weak∗-closure of a convex
bounded set C ⊆ X ∗ is the weak∗-limit of a sequence in C;
(iii) Corson's property (C) if every element of the weak∗-closure of a convex
bounded set C ⊆ X ∗ belongs to the weak∗-closure of a countable subset
of C.
Clearly, (i)⇒(ii)⇒(iii). Property (E) was first considered in [4] and was studied
further by Plichko and Yost in [12, 13]. To clarify whether property (E) is actually
different to the weak∗-angelicity of the dual or property (C), they asked in [13,
p. 352] if Johnson-Lindenstrauss spaces enjoy property (E). It is well known that
the Johnson-Lindenstrauss space J L2(F ) associated to any almost disjoint family F
of subsets of N has property (C), but fails to have weak∗-angelic dual whenever F is
maximal (shortly, a MAD family). Plichko [12] claimed that Johnson-Lindenstrauss
spaces have property (E). However, his proof contains a gap. Under the Continuum
Hypothesis (CH), we will prove the existence of two MAD families F+ and F− such
that J L2(F+) has property (E), while J L2(F−) fails it. We do not know whether
such MAD families can be constructed in ZFC without any extra set-theoretic
assumption.
Date: August 29, 2018.
2010 Mathematics Subject Classification. 46A50, 46B26.
Key words and phrases. Johnson-Lindenstrauss space; weak∗-sequential closure; almost dis-
joint family.
Research supported by projects MTM2014-54182-P, MTM2017-86182-P (AEI/FEDER, UE)
and 19275/PI/14 (Fundaci´on S´eneca).
1
2
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J. RODR´IGUEZ
In particular, under CH, property (E) lies strictly between having weak∗-angelic
dual and property (C). We stress that other consistent examples of Banach spaces
with property (C) but not property (E) were already constructed by J.T. Moore
under ♦ (unpublished) and Brech, see [2, Section 3.3].
A Banach space X is said to have Gulisashvili's property (D) if σ(X ∗) = σ(Γ)
for any total set Γ ⊆ X ∗, where σ(Γ) denotes the σ-algebra on X generated by Γ.
Gulisashvili [5] proved that this property is enjoyed by any Banach space having
weak∗-angelic dual and asked whether the converse holds. One of the aims in [12]
was to show that Johnson-Lindenstrauss spaces separate both properties. This is
indeed the case but the argument cannot rely on property (E). To explain this we
need a couple of definitions. A Banach space X is said to have
• property (E ′) if every weak∗-sequentially closed convex bounded subset
of X ∗ is weak∗-closed (see [9]);
• property (D′) if every weak∗-sequentially closed linear subspace of X ∗ is
weak∗-closed (see [12]).
The following diagram summarizes the relations between all these properties:
weak∗-angelic dual
(E)
(E ′)
(D′)
(D)
(C)
The second-named author proved in [9] that any Johnson-Lindenstrauss space has
weak∗-sequential dual ball and so it has property (E ′). In particular, this implies
that the Johnson-Lindenstrauss space associated to any MAD family works as a
counterexample to Gulisashvili's question above. Note also that, under CH, the
space J L2(F−) based on our MAD family F− answers in the negative Plichko's
question [12] of whether properties (E) and (D′) are equivalent.
This paper is organized as follows.
In Section 2 we fix the terminology and
collect some preliminary facts on Johnson-Lindenstrauss spaces.
In Sections 3
and 4, by assuming CH, we construct MAD families such that the correspond-
ing Johnson-Lindentrauss spaces have/fail property (E). Finally, in Section 5 we
analyze Plichko's attempt to prove that all Johnson-Lindentrauss spaces have prop-
erty (E). For instance, we show that if X is a Banach space which is weak∗-
sequentially dense in X ∗∗, then X has property (E ′) (see Theorem 5.3).
2. Preliminaries
All our Banach spaces are real. The (topological) dual of a Banach space X is
denoted by X ∗ and the weak∗-topology on X ∗ is denoted by w∗. The linear span
of a set W ⊆ X is denoted by span(W ), while span(W ) stands for its closure; we
write co(W ) for the convex hull of W . The closed unit ball of X is denoted by BX .
Two sets are said to be almost disjoint if they have finite intersection. By an
almost disjoint family we mean a family of pairwise almost disjoint infinite subsets
WEAK∗-SEQUENTIAL PROPERTIES OF JOHNSON-LINDENSTRAUSS SPACES
3
of N. An almost disjoint family is said to be a maximal almost disjoint (MAD)
family if it is maximal with respect to inclusion.
Let F be an almost disjoint family. The Johnson-Lindenstrauss space J L2(F )
is defined as the completion of span(c0 ∪ {χN : N ∈ F }) ⊆ ℓ∞ with respect to the
norm
k
Xr=1
x +
(cid:13)(cid:13)(cid:13)
arχNr(cid:13)(cid:13)(cid:13)JL2(F )
:= max(cid:26)(cid:13)(cid:13)(cid:13)
x +
k
Xr=1
arχNr(cid:13)(cid:13)(cid:13)∞
, (cid:18) k
Xr=1
a2
r(cid:19)
1
2(cid:27),
where x ∈ c0, {N1, . . . , Nk} ⊆ F and a1, . . . , ak ∈ R. Here χN denotes the charac-
teristic function of a set N ⊆ N and k · k∞ is the supremum norm on ℓ∞. Johnson-
Lindenstrauss spaces first appeared in [6] and, in general, they refer to spaces of the
form J L2(F ) with F being a MAD family. However, we will avoid the maximality
assumption on F unless otherwise mentioned.
The dual J L2(F )∗ is isomorphic to ℓ1 ⊕ ℓ2(F ). More precisely, for each n ∈ N,
n(χN ) = χN (n) for all N ∈ F and
n ∈ J L2(F )∗ be the functional satisfying e∗
let e∗
e∗
n(ej) =(1 if n = j
0 otherwise
for all j ∈ N, where (ej)j∈N denotes the usual basis of c0. For each N ∈ F , let
e∗
N ∈ J L2(F )∗ be the functional satisfying e∗
N (x) = 0 for every x ∈ c0 and
e∗
N (χN ′ ) =(1 if N ′ = N
0 otherwise
n)n∈N is equivalent to the usual basis of ℓ1, (e∗
for all N ′ ∈ F . Then (e∗
N )N ∈F is
equivalent to the usual basis of ℓ2(F ) and J L2(F )∗ equals to the direct sum of
N : N ∈ F }). That is, every x∗ ∈ J L2(F )∗ has a
span({e∗
unique expression of the form
n : n ∈ N}) and span({e∗
x∗ = Xn∈N
ane∗
n + XN ∈F
aN e∗
N
with (an)n∈N ∈ ℓ1 and (aN )N ∈F ∈ ℓ2(F ), and we write
suppN x∗ := {n ∈ N : an 6= 0} and
suppF x∗ := {N ∈ F : aN 6= 0}.
We say that x∗ is finitely supported if suppN x∗ and suppF x∗ are both finite.
So, if F ⊆ F ′ are two almost disjoint families, then there is an isomorphic
embedding i : J L2(F )∗ → J L2(F ′)∗ such that i(e∗
n for all n ∈ N and
i(e∗
N for all N ∈ F (note that i is not weak∗-weak∗ continuous). This
allows us to see every element of J L2(F )∗ as an element of J L2(F ′)∗ through the
operator i (which will be omitted) and we write J L2(F )∗ ⊆ J L2(F ′)∗.
N ) = e∗
n) = e∗
For more information on Johnson-Lindentrauss spaces we refer the reader to [7],
[8], [17], [18] and [19].
4
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J. RODR´IGUEZ
3. A Johnson-Lindenstrauss space with property (E)
In this section we will prove that, under CH, there exists a MAD family F+
for which J L2(F+) has property (E). In order to do this, we construct by transfi-
nite induction an increasing family of countable almost disjoint families (Fα)α<ω1
Fα is a MAD family and such that every bounded sequence in
Fα)∗ containing 0 in its weak∗-closure is dealt with at some step to
guarantee that it admits a subsequence whose arithmetic means are weak∗-null.
such that Sα<ω1
J L2(Sα<ω1
Definition 3.1. Given an almost disjoint family F , we say that a sequence (x∗
n)n∈N
in J L2(F )∗ is semi-summable if supj∈NPn∈N x∗
Lemma 3.2. Let F be an almost disjoint family and (x∗
in J L2(F )∗ for which 0 is a weak∗-cluster point. Then (x∗
summable subsequence.
n(ej) < ∞.
n)n∈N a bounded sequence
n)n∈N admits a semi-
Proof. Observe first that for every x∗ ∈ J L2(F )∗ and c > 0 the cardinality of the
set {j ∈ N : x∗(ej) ≥ c} is less than or equal to kx∗kc−1.
Assume without loss of generality that kx∗
nk ≤ 1 for all n ∈ N. We next construct
ni (ej) < 2 for every k ∈ N
and j ∈ N. Of course, such subsequence is semi-summable. For the first step, just
take x∗
1. Now, suppose n1 < n2 < · · · < nk have been already chosen in
nk(cid:1)k∈N such that Pk
i=1 x∗
n1 := x∗
by induction a subsequence (cid:0)x∗
such a way that Pk
S :=nj ∈ N :
i=1 x∗
Xi=1
x∗
k
ni (ej) < 2 for every j ∈ N. Note that
k
ni (ej) ≥ 1o ⊆
[i=1nj ∈ N : x∗
ni (ej) ≥
1
ko,
so S is finite. Since 0 is a weak∗-cluster point of (x∗
such that
n)n∈N, there exists nk+1 > nk
On the other hand, given any j ∈ N \ S, we have
x∗
ni (ej) < 2 for every j ∈ S.
x∗
ni (ej) < 1 + x∗
nk+1 (ej) ≤ 2,
k+1
Xi=1
k+1
Xi=1
because kx∗
nk+1k ≤ 1. Therefore, Pk+1
i=1 x∗
The following result will be our key lemma in the inductive construction.
ni (ej) < 2 for every j ∈ N.
(cid:3)
Lemma 3.3. Let F be a countable almost disjoint family with N = S F and let
S be a countable family of semi-summable sequences of finitely supported elements
of J L2(F )∗. Let (x∗
k)k∈N be a semi-summable sequence of finitely supported ele-
ments of J L2(F )∗ such that
k is not contained in a finite union of elements of F .
(⋆) Sk∈N suppN x∗
Then there exists an infinite set N ⊆Sk∈N suppN x∗
(i) F ∪ {N } is an almost disjoint family;
(ii) limk→∞
2 + . . . + x∗
k)(χN ) = 0;
1 + x∗
1
k (x∗
k such that:
WEAK∗-SEQUENTIAL PROPERTIES OF JOHNSON-LINDENSTRAUSS SPACES
5
(iii) limk→∞
1
k (y∗
1 + y∗
2 + . . . + y∗
k)(χN ) = 0 for every sequence (y∗
k)k∈N ∈ S.
(We consider χN ∈ J L2(F ∪ {N }) and the embedding J L2(F )∗ ⊆ J L2(F ∪ {N })∗.)
Proof. Enumerate F = {Ni : i ∈ N} and S = {(cid:0)y∗
A :=Sk∈N suppN x∗
k. For each k ∈ N, set
suppN y∗
suppN x∗
Ak := [r,m≤2k
r,m ∪ [m≤2k
r,m(cid:1)m∈N : r ∈ N}. Define
m ∪ [i≤k
Ni,
so that A \ Ak is infinite (bear in mind (⋆) and the fact that N =S F ). Therefore,
we can take a sequence (tk)k∈N with tk ∈ A \ Ak for all k ∈ N and tk 6= tk′ whenever
k 6= k′. Define N := {tk : k ∈ N} ⊆ A. Let us check that N satisfies the required
properties. Given any i ∈ N, we have Ni ⊆ Ak for every k ≥ i, hence tk 6∈ Ni for
every k ≥ i, and so N ∩ Ni is finite. Therefore, F ∪ {N } is almost disjoint. Note
that for any
x∗ = Xn∈N
ane∗
n + XM∈F
aM e∗
M ∈ J L2(F )∗ ⊆ J L2(F ∪ {N })∗
(with (an)n∈N ∈ ℓ1 and (aM )M∈F ∈ ℓ2(F )) we have
(3.1)
x∗(χN ) = Xn∈N
an.
In order to prove (ii), fix j ∈ N with j ≥ 2 and let s(j) ∈ N so that
s(j) ≤ log2(j) < s(j) + 1.
suppN x∗
m ⊆ {t1, t2, . . . , ts(j)}
N ∩ [m≤j
Then
and so
1
j
(cid:12)(cid:12)(cid:12)
(x∗
1 + x∗
2 + . . . + x∗
j )(χN )(cid:12)(cid:12)(cid:12)
(3.1)
=
1
j(cid:12)(cid:12)(cid:12)
(x∗
1 + x∗
2 + . . . + x∗
etk(cid:19)(cid:12)(cid:12)(cid:12)
s(j)
j )(cid:18)
Xk=1
Xm=1
Xk=1
s(j)
j
≤
1
j
x∗
m(etk ) ≤
s(j)
j
C,
where we write C := supk∈NP∞
semi-summable). It follows that limj→∞
m=1 x∗
m(etk ) < ∞ (bear in mind that (x∗
m)m∈N is
1
j (x∗
1 + x∗
2 + . . . + x∗
j )(χN ) = 0.
For the proof of (iii), take any r ∈ N. For every j ∈ N with j ≥ max{r, 2} we
have
suppN y∗
r,m ⊆ {t1, t2, . . . , ts(j)}
N ∩ [m≤j
and, similarly as before, we conclude that limj→∞
The proof is complete.
1
j (y∗
r,1 + y∗
r,2 + . . .+ y∗
r,j)(χN ) = 0.
(cid:3)
6
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J. RODR´IGUEZ
Theorem 3.4. Under CH, there exists a MAD family F+ such that every bounded
n)n∈N in J L2(F+)∗ for which 0 is a weak∗-cluster point has a subse-
sequence (x∗
quence (cid:0)x∗
nk(cid:1)k∈N such that
w∗ − lim
k→∞
1
k
(x∗
n1 + x∗
n2 + . . . + x∗
nk ) = 0.
γ , η2
γ) : 0 < γ < ω1} be an enumeration of ω1 × ω1 with η1
Proof. Let {ηγ = (η1
γ < γ
for every 0 < γ < ω1. By transfinite induction on α < ω1, we will construct an
of countable almost disjoint families and an increasing
increasing chain (Fα)α<ω1
where each Sα is a countable family of semi-summable sequences of
chain (Sα)α<ω1
finitely supported elements of J L2(Fα)∗. In each step, the set Rα of all bounded
sequences of finitely supported elements of J L2(Fα)∗ has cardinality c = ℵ1 and is
enumerated as
Rα = {rξ : ξ ∈ {α} × ω1}.
Let F0 be any countable almost disjoint family with N = S F0 and let S0 be
any countable family of semi-summable weak∗-null sequences of finitely supported
elements of J L2(F0)∗.
Suppose now that (Fγ)γ<α and (Sγ)γ<α are already defined for some 0 < α < ω1.
Take Fα := Sγ<α Fγ and Sα := Sγ<α Sγ. Note that Fα is a countable almost
disjoint family and that Sα is a countable family of semi-summable sequences of
finitely supported elements of J L2( Fα).
Let (x∗
n)n∈N := rηα ∈ Rη1
α < α. That
n)n∈N is a bounded sequence of finitely supported elements of the space
α , which is already defined since η1
α )∗ ⊆ J L2( Fα). We now distinguish several cases:
Case 1. If 0 is not a weak∗-cluster point of (x∗
n)n∈N in J L2( Fα)∗, then we set
is, (x∗
J L2(Fη1
Fα := Fα and Sα := Sα.
Case 2. If 0 is a weak∗-cluster point of (x∗
n)n∈N in J L2( Fα)∗, then we can take
that Fα is countable) and so bounded subsets of J L2( Fα)∗ are weak∗-metrizable.
a weak∗-null subsequence (cid:0)x∗
By passing to a further subsequence, not relabeled, we can assume that (cid:0)x∗
nk(cid:1)k∈N, because J L2( Fα) is separable (bear in mind
nk(cid:1)k∈N
is semi-summable (apply Lemma 3.2).
nk is contained in a finite union of elements of Fα, then we
set Fα := Fα and Sα := Sα.
• If Sk∈N suppN x∗
• If not, then we apply Lemma 3.3 to Fα, Sα and(cid:0)x∗
an infinite set N ⊆Sk∈N suppN x∗
(i) Fα ∪ {N } is an almost disjoint family;
(ii) limk→∞
(iii) limk→∞
n2 + . . . + x∗
2 + . . . + y∗
nk )(χN ) = 0;
1
k (x∗
1
k (y∗
n1 + x∗
1 + y∗
k)(χN ) = 0 for every (y∗
nk satisfying the following conditions:
nk(cid:1)n∈N in order to obtain
In this case, we define Fα := Fα ∪ {N } and Sα := Sα ∪ {(cid:0)x∗
This finishes the inductive construction. We claim that
k)k∈N ∈ Sα.
nk(cid:1)k∈N}.
is a MAD family satisfying the required property. Clearly, F+ is almost disjoint.
F+ := [α<ω1
Fα
WEAK∗-SEQUENTIAL PROPERTIES OF JOHNSON-LINDENSTRAUSS SPACES
7
For the maximality, take any infinite set N ′ = {nk : k ∈ N} ⊆ N. Note that
(cid:0)e∗
nk(cid:1)k∈N ∈ R0, hence there is 0 < α < ω1 such that (cid:0)e∗
nk(cid:1)k∈N = rηα , that is,
(cid:0)e∗
nk(cid:1)k∈N is the sequence considered at step α in the inductive construction. It is
nk(cid:1)k∈N in J L2( Fα)∗ if and
easy to check that 0 is not a weak∗-cluster point of (cid:0)e∗
only if N ′ is contained in a finite union of elements of Fα; in this case, there is
N ′′ ∈ Fα ⊆ F+ such that N ′ ∩ N ′′ is infinite. On the other hand, if 0 is a weak∗-
nk )k∈N in J L2( Fα)∗, then we find in step α an infinite set N ⊆ N ′
cluster point of (e∗
such that Fα ∪ {N } = Fα ⊆ F+. This shows that F is a MAD family.
Let (x∗
Rβ and so there is 0 < α < ω1 such that (x∗
point. We can assume without loss of generality that each x∗
(because finitely supported elements are norm-dense in J L2(F+)∗). Then (x∗
n)n∈N be a bounded sequence in J L2(F+)∗ for which 0 is a weak∗-cluster
n is finitely supported
n)n∈N
n)n∈N = rηα . Note
n)n∈N in J L2( Fα)∗, so we are in Case 2 of the
that 0 is a weak∗-cluster point of (x∗
inductive construction at step α. Therefore, one of the following conditions holds:
nk is contained in a finite union of elements of Fα.
Indeed, since
nk (χN ′ ) = 0
belongs to Sβ<ω1
Condition 1: Sk∈N suppN x∗
In this case, we claim that (cid:0)x∗
nk(cid:1)k∈N is weak∗-null in J L2( Fα)∗, it suffices to show that limk→∞ x∗
(cid:0)x∗
nk(cid:1)k∈N is weak∗-null in J L2(F+)∗.
for every N ′ ∈ F+ \ Fα. To this end, for each k ∈ N, we write
x∗
nk = Xm∈N
am,k e∗
m + XN ∈ Fα
aN,k e∗
N
with (am,k)m∈N ∈ ℓ1 and (aN,k)N ∈ Fα
∈ ℓ2( Fα). Choose N1, N2, . . . , Np ∈ Fα such
i=1 Ni. Given any N ′ ∈ F+ \ Fα, we have
am,k = Xm∈Sp
i=1 N ′∩Ni
am,k
for all k ∈ N.
i=1 N ′ ∩ Ni is finite and limk→∞ am,k = limk→∞ x∗
nk (em) = 0 for every
nk (χN ′ ) = 0, as desired.
nk is not contained in a finite union of elements of Fα.
that Sk∈N suppN x∗
(3.2)
x∗
nk ⊆Sp
nk (χN ′ ) = Xm∈N ′
m ∈ N, from (3.2) we get limk→∞ x∗
Since Sp
Condition 2: Sk∈N suppN x∗
Then
(3.3)
lim
k→∞
1
k
(x∗
n1 + x∗
n2 + . . . + x∗
nk )(χN ′ ) = 0
for every N ′ ∈ F+ \ Fα. Indeed, for N ′ = N this follows from the very construction
at step α. If N ′ ∈ F+ \ Fα, then N ′ is added to F+ at step β for some α < β < ω1
nk(cid:1)k∈N ∈ Sα ⊆ Sβ.
and (3.3) holds because (cid:0)x∗
Similarly as before, (3.3) implies that (cid:0) 1
null in J L2(F+)∗. The proof is complete.
k (x∗
n1 + x∗
n2 + . . . + x∗
nk )(cid:1)k∈N is weak∗-
(cid:3)
Corollary 3.5. Under CH, there exists a MAD family F+ such that J L2(F+) has
property (E).
Proof. Let F+ be the MAD family given by Theorem 3.4. By linearity, it is enough
to prove that if C ⊆ BJL2(F+)∗ is a convex set with 0 ∈ C
, then there exists a
weak∗-null sequence contained in C. This is obvious if 0 ∈ C, so we assume that
w∗
8
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J. RODR´IGUEZ
0 6∈ C. Since (BJL2(F+)∗ , w∗) is sequential (see [9, Theorem 3.1]), it has countable
tightness and so there exists a sequence (x∗
n)n∈N in C with
(3.4)
0 ∈ {x∗
n : n ∈ N}
w∗
.
Note that the existence of such a sequence can also be deduced from Corson's
property (C) of J L2(F+). Since 0 is a weak∗-cluster point of (x∗
n)n∈N (by (3.4)
and the fact that x∗
n 6= 0 for all n ∈ N), the conclusion follows from Theorem 3.4,
which ensures the existence of a subsequence(cid:0)x∗
arithmetic means (cid:0) 1
nk(cid:1)k∈N such that the sequence of its
nk )(cid:1)k∈N (which is contained in the convex
set C) is weak∗-null.
n2 + . . . + x∗
n1 + x∗
k (x∗
(cid:3)
4. A Johnson-Lindenstrauss space without Property (E)
Given any infinite almost disjoint family F , we have 0 ∈ {e∗
in
J L2(F )∗, because w∗ − limk→∞ e∗
= 0 for every sequence (Nk)k∈N of pairwise
distinct elements of F and w∗ − limn∈N e∗
N for all N ∈ F . In this section we
will prove that, under CH, there exists a MAD family F− for which co({e∗
n : n ∈ N})
does not contain weak∗-null sequences and consequently J L2(F−) does not have
property (E). In order to construct F−, we will focus on the matrices (λi,j)i,j∈N
n : n ∈ N}
n = e∗
Nk
w∗
determined by sequences (cid:16)Pj∈N λi,j e∗
j(cid:17)i∈N
Definition 4.1. We say that a matrix (λi,j )i,j∈N ∈ [0, 1]N×N is
in co({e∗
n : n ∈ N}).
(i) convex if Pj∈N λi,j = 1 for every i ∈ N;
(ii) null if limi→∞ λi,j = 0 for every j ∈ N.
Lemma 4.2. Let (Nr)r∈N be a sequence of subsets of N and let (λi,j)i,j∈N be a
λi,j = 0 for every r ∈ N, then there exists an
convex null matrix. If limi→∞Pj∈Nr
infinite set N ′ ⊆ N such that N ′ ∩ Nr is finite for every r ∈ N and
lim sup
i→∞ Xj∈N ′
λi,j ≥
1
2
.
λi,j = 0 for every r ∈ N, we can find a strictly increas-
ing sequence (nr)r∈N in N such that
Proof. Since limi→∞Pj∈Nr
Xj∈N1∪N2∪...∪Nr
λnr ,j ≤
r
Xk=1 Xj∈Nk
λnr ,j <
1
2
for all r ∈ N.
For each r ∈ N we have Pj∈N λnr ,j = 1 and, therefore, we can find a finite set
Fr ⊆ N \ (N1 ∪ N2 ∪ . . . ∪ Nr) in such a way that Pj∈Fr
Set N ′ :=Sr∈N Fr. Clearly, N ′ ∩ Nr ⊆Ss<r Fs is finite for every r ∈ N. Notice
that Pj∈N ′ λnr ,j ≥Pj∈Fr
2 for every r ∈ N, so
λnr ,j ≥ 1
2 .
λnr ,j ≥ 1
lim sup
i→∞ Xj∈N ′
λi,j ≥
1
2
.
Bearing in mind that for each finite set F ⊆ N we have limi→∞Pj∈F λi,j = 0, we
conclude that N ′ is infinite.
(cid:3)
WEAK∗-SEQUENTIAL PROPERTIES OF JOHNSON-LINDENSTRAUSS SPACES
9
Lemma 4.3. Let {(cid:0)λα
cardinality ℵ1. Then there exists an almost disjoint family F such that for every
α < ω1 there is N ′
i,j(cid:1)i,j∈N : α < ω1} be a family of convex null matrices of
λα
i,j > 0.
α ∈ F with lim supi→∞Pj∈N ′
α
Proof. Let F0 be a countable almost disjoint family including a set N ′
0 ⊆ N for
λ0
i,j > 0. Indeed, the existence of such F0 follows from
Lemma 4.2 applied to an arbitrary countable infinite almost disjoint family and the
0
which lim supi→∞Pj∈N ′
convex null matrix (cid:0)λ0
i,j(cid:1)i,j∈N.
We now construct an increasing chain (Fα)α<ω1
of countable infinite almost
disjoint families by transfinite induction on α. Suppose that 0 < α < ω1 and that
Fβ is already constructed for every β < α. If
(4.1)
lim
i→∞Xj∈N
λα
i,j = 0 for every N ∈ [β<α
Fβ,
order to get an infinite set N ′
disjoint and
then we can apply Lemma 4.2 to Sβ<α Fβ (which is countable) and (cid:0)λα
α ⊆ N such that Fα := (Sβ<α Fβ) ∪ {N ′
i→∞ Xj∈N ′
i,j > 0.
lim sup
λα
α
i,j(cid:1)i,j∈N in
α} is almost
If (4.1) fails, then we just take Fα :=Sβ<α Fβ.
It is clear that F :=Sβ<ω1
Fβ is the desired almost disjoint family.
(cid:3)
The previous lemma combined with the fact that under CH there are only ℵ1-
many convex null matrices in [0, 1]N×N provide a MAD family for which the corre-
sponding Johnson-Lindenstrauss space does not have property (E):
Theorem 4.4. Under CH, there exists a MAD family F− such that J L2(F−) does
not have property (E).
Proof. Let F− be the almost disjoint family given by Lemma 4.3 applied to the
family of all convex null matrices (which has cardinality ℵ1 under CH). We claim
that F− is maximal. Indeed, if N = {nk : k ∈ N} ⊆ N is an infinite set, then we can
define a convex null matrix (λi,j)i,j∈N by the formula λi,j := 1 if ni = j and λi,j := 0
2 , which
otherwise. Then there exists N ′ ∈ F− such that lim supi→∞Pj∈N ′ λi,j ≥ 1
clearly implies that N ∩ N ′ is infinite. Therefore, F− is a MAD family.
We now prove that J L2(F−) does not have property (E). Let
C := co({e∗
n : n ∈ N}) ⊆ J L2(F−)∗.
As we explained at the beginning of this section, 0 ∈ C
to prove that no sequence in C is weak∗-convergent to zero. Let (x∗
arbitrary sequence in C and, for each i ∈ N, write x∗
. Thus, it is enough
i )i∈N be an
j , where λi,j ≥ 0
w∗
and Pj∈N λi,j = 1 (the sum being finitely supported). Hence
M := (λi,j )i,j∈N = (x∗
i (ej))i,j∈N ∈ [0, 1]N×N
i =Pj∈N λi,je∗
10
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J. RODR´IGUEZ
is a convex matrix. Clearly, if M is not null, then (x∗
the other hand, if M is null, then there exists N ∈ F− such that
i )i∈N is not weak∗-null. On
lim sup
i→∞
x∗
i (χN ) = lim sup
λi,j > 0,
i→∞ Xj∈N
so the sequence (x∗
i )i∈N cannot be weak∗-null either.
(cid:3)
5. Further remarks on weak∗-sequential properties
Let X be a Banach space. Given any set C ⊆ X ∗, we denote by S1(C) ⊆ X ∗
the set of all limits of weak∗-convergent sequences contained in C. Clearly, X has
property (E) if and only if S1(C) is weak∗-closed (equivalently, S1(C) = C
) for
every convex bounded set C ⊆ X ∗. The failed argument of [12] that all Johnson-
Lindenstrauss spaces have property (E) is based on the claim (see the proof of [12,
Proposition 8]) that S1(C) is norm-closed for every convex bounded set C ⊆ X ∗.
However, this is not always the case, as we show below.
w∗
Definition 5.1. We say that a Banach space X has property (P) if S1(C) is norm-
closed for every convex bounded set C ⊆ X ∗.
It is clear that property (E) implies property (P) and that every Grothendieck
space has property (P).
We next give an example of a Banach space failing property (P). Recall that
the cardinal d is defined as the least cardinality of a subset of NN which is cofinal
for the relation "f ≤∗ g if and only if f (i) ≤ g(i) for all but finitely many i's".
One has ℵ1 ≤ d ≤ c, but whether any of these are strict inequalities or equalities is
independent of ZFC, see e.g. [16].
Example 5.2. ℓ1(d) fails property (P).
Proof. For any function f : N → N we define Mf : N × N → R by
Mf (i, j) :=(1
1
i
if j < f (i),
if j ≥ f (i).
Let Γ ⊆ NN be a family of functions with cardinality d which is cofinal for ≤∗.
Write X := ℓ1(Γ) and identify X ∗ = ℓ∞(Γ). For each (i, j) ∈ N × N, define
i,j ∈ BX ∗ = [−1, 1]Γ by declaring x∗
x∗
i,j(f ) := Mf (i, j) for all f ∈ Γ. Let C ⊆ BX ∗
be the convex hull of the x∗
i,j 's.
Observe that for each i ∈ N we have 1
i χΓ ∈ S1(C), because
w∗ − lim
j→∞
x∗
i,j =
1
i
χΓ.
Therefore, 0 ∈ S1(C)
k·k
.
We claim that 0 6∈ S1(C). Indeed, let (y∗
n)n∈N be any sequence in C and, for
each n ∈ N, write
y∗
n = X(i,j)∈N×N
λ(n)
i,j x∗
i,j ,
WEAK∗-SEQUENTIAL PROPERTIES OF JOHNSON-LINDENSTRAUSS SPACES
11
where λ(n)
contradiction, suppose that (y∗
i,j ≥ 0 and P(i,j)∈N×N λ(n)
Step 1. Fix i ∈ N. Then
n)n∈N is weak∗-null.
i,j = 1 (the sum being finitely supported). By
(5.1)
λ(n)
i,j = 0,
lim
n→∞Xj∈N
because for an arbitrary f ∈ Γ and for every n ∈ N we have
y∗
n(f ) = X(k,j)∈N×N
λ(n)
k,j Mf (k, j) ≥Xj∈N
λ(n)
i,j Mf (i, j) ≥
λ(n)
i,j .
1
i Xj∈N
Choose n(i) ∈ N such that
(5.2)
λ(n)
i,j ≤
1
2i+1
Xj∈N
for every n ≥ n(i).
Now we choose f (i) ∈ N large enough such that
(5.3)
λ(n)
i,j = 0 for every j ≥ f (i) and every n < n(i).
Step 2. Pick f ∈ Γ such that f ≤∗ f and fix i0 ∈ N such that f (i) ≤ f (i) for
every i ≥ i0. By (5.1), there is n0 ∈ N such that
(5.4)
λ(n)
i,j ≤
1
4
Xi<i0Xj∈N
for every n ≥ n0.
Observe that for each n ∈ N we have
Xi: n≥n(i)Xj∈N
λ(n)
i,j
(5.2)
≤ Xi∈N
1
2i+1 =
1
2
and so
(5.5)
Xi: n<n(i) Xj< f (i)
λ(n)
i,j
(5.3)
= Xi: n<n(i)Xj∈N
λ(n)
i,j = 1 − Xi: n≥n(i)Xj∈N
λ(n)
i,j ≥
1
2
.
Therefore, for every n ≥ n0 we have
y∗
n(f ) = X(i,j)∈N×N
λ(n)
i,j
i≥i0 Xj<f (i)
λ(n)
i,j Mf (i, j) ≥ Xi: n<n(i)
i≥i0 Xj< f (i)
≥ Xi: n<n(i)
λ(n)
i,j
(5.4)
≥ Xi: n<n(i) Xj< f (i)
λ(n)
i,j −
1
4
(5.5)
≥
1
4
,
which contradicts the fact that (y∗
n)n∈N is weak∗-null.
(cid:3)
The remainder of this section is devoted to showing how the techniques of [12]
yield Theorem 5.3 below. Recall first that a Banach space X is said to have prop-
for every convex bounded set C ⊆ X ∗. Here,
erty (E ′) (see [9]) if Sω1 (C) = C
for any ordinal α ≤ ω1, the α-th w∗-sequential closure of a set D ⊆ X ∗ is defined
by transfinite induction as follows: S0(D) := D, Sα(D) := S1(Sβ(D)) if α = β + 1
w∗
and Sα(D) :=Sβ<α Sβ(D) if α is a limit ordinal.
12
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J. RODR´IGUEZ
Theorem 5.3. Let X be a Banach space which is weak∗-sequentially dense in X ∗∗.
Then:
w∗
k·k
= S1(C)
(i) C
(ii) X has property (E ′);
(iii) X has property (E) if and only if it has property (P).
= S2(C) for every convex bounded set C ⊆ X ∗;
In order to prove Theorem 5.3 we need some previous work. The following fact
will play a key role in our argument (see [10, Theorem 4], cf. [14, Proposition 3.9]):
Fact 5.4. If X is a Banach space which is weak∗-sequentially dense in X ∗∗, then
X contains no isomorphic copy of ℓ1.
The next result goes back to [4, Theorem 3.3.8], cf.
[12, Proposition 7]. We
provide another proof for the reader's convenience. Recall that, given a Banach
space X and a weak∗-compact set K ⊆ X ∗, a set B ⊆ K is said to be a James
0(x) = supx∗∈K x∗(x).
boundary of K if for every x ∈ X there is x∗
Proposition 5.5. Let X be a Banach space which is weak∗-sequentially dense
in X ∗∗. Then for every weak∗-compact set K ⊆ X ∗ and every James boundary B
of K we have
0 ∈ B such that x∗
co(K)
w∗
= co(B)
k·k
.
Proof. The inclusion co(K)
phic copy of ℓ1 (Fact 5.4), we can apply [3, Theorem 5.4] to get
⊇ co(B)
is obvious. Since X contains no isomor-
w∗
k·k
w∗
co(K)
= co(B)
γ
,
where γ denotes the topology on X ∗ of uniform convergence on bounded countable
subsets of X. Since X is sequentially weak∗-dense in X ∗∗, it is easy to check that
γ is stronger than the weak topology of X ∗, hence
weak
k·k
w∗
γ
co(K)
= co(B)
⊆ co(B)
= co(B)
,
which finishes the proof.
(cid:3)
The dual ball BX ∗ of a Banach space X is said to be convex block weak∗-compact
if every sequence in BX ∗ admits a weak∗-convergent convex block subsequence. By a
convex block subsequence of a sequence (fn)n∈N in a linear space we mean a sequence
anfn, where (Ik)k∈N is a sequence of finite
subsets of N with max(Ik) < min(Ik+1) and (an)n∈N is a sequence of non-negative
(gk)k∈N of vectors of the form gk =Pn∈Ik
real numbers such that Pn∈Ik
Lemma 5.6. Let X be a Banach space such that BX ∗ is convex block weak∗-
compact. Let D ⊆ X ∗ be a bounded set and x ∈ X. Then there is x∗
0 ∈ S1(co(D))
such that
an = 1 for all k ∈ N.
(5.6)
x∗
0(x) = sup(cid:8)x∗(x) : x∗ ∈ D
w∗
(cid:9).
Proof. Let α be the right hand side of (5.6). Take a sequence (x∗
that limn→∞ x∗
convex block subsequence (y∗
(y∗
n)n∈N in D such
n(x) = α. By the assumption on X, there is a weak∗-convergent
0 ∈ S1(co(D)). Since
(cid:3)
k(x))k∈N is a convex block subsequence of (x∗
n(x))n∈N, we have x∗
n)n∈N, with limit x∗
k)k∈N of (x∗
0(x) = α.
WEAK∗-SEQUENTIAL PROPERTIES OF JOHNSON-LINDENSTRAUSS SPACES
13
Proof of Theorem 5.3. Since X contains no isomorphic copy of ℓ1 (Fact 5.4), BX ∗
is convex block weak∗-compact, see [1, Proposition 3.11] (cf.
[11, Proposition 11]
and [15]). To prove (i), let C ⊆ X ∗ be a convex bounded set. By Lemma 5.6,
w∗
the convex set S1(C) is a James boundary of the convex weak∗-compact set C
.
k·k
Now, an appeal to Proposition 5.5 ensures that C
. Since S1(C)
k·k
is
w∗
= S1(C)
w∗
(obviously) contained in S2(C) ⊆ C
Statements (ii) and (iii) follow at once from (i).
, we conclude that C
w∗
k·k
= S1(C)
= S2(C).
(cid:3)
It is known that J L2(F ) is sequentially weak∗-dense in its bidual for every almost
disjoint family F . Therefore, under CH, the property of being sequentially weak∗-
dense in the bidual does not imply property (P).
Acknowledgements. This research was supported by projects MTM2014-54182-
P, MTM2017-86182-P (AEI/FEDER, UE) and 19275/PI/14 (Fundaci´on S´eneca).
References
[1] J. Bourgain, La propri´et´e de Radon-Nikod´ym, Publ. Math. Univ. Pierre et Marie Curie 36
(1979).
[2] C. Brech, Constru¸coes gen´ericas de espa¸cos de Asplund C(K), Universidade de Sao Paulo
and Universit´e Paris 7 (2008).
[3] B. Cascales, M. Munoz, and J. Orihuela, James boundaries and σ-fragmented selectors, Stu-
dia Math. 188 (2008), no. 2, 97 -- 122. MR 2430997
[4] N. M. Efremov, Duality conditions of a Banach space in terms of norm attaining linear
functionals, Ph.D. Thesis, Kharkiv State University, 1985 (in Russian).
[5] A. B. Gulisashvili, Estimates for the Pettis integral in interpolation spaces, and a general-
ization of some imbedding theorems, Soviet Math., Dokl. 25 (1982), 428 -- 432. MR 0651235
[6] W. B. Johnson and J. Lindenstrauss, Some remarks on weakly compactly generated Banach
spaces, Israel J. Math. 17 (1974), 219 -- 230. MR 0417760
[7] W. Marciszewski and R. Pol, On Banach spaces whose norm-open sets are Fσ-sets in the
weak topology, J. Math. Anal. Appl. 350 (2009), no. 2, 708 -- 722. MR 2474806 (2010a:46060)
[8] W. Marciszewski and R. Pol, On Borel almost disjoint families, Monatsh. Math. 168 (2012),
no. 3-4, 545 -- 562. MR 2993963
[9] G. Mart´ınez-Cervantes, Banach spaces with weak*-sequential dual ball, Proc. Amer. Math.
Soc. 146 (2018), no. 4, 1825 -- 1832. MR 3754364
[10] K. Musia l, The weak Radon-Nikod´ym property in Banach spaces, Studia Math. 64 (1979),
no. 2, 151 -- 173. MR 537118
[11] H. Pfitzner, Boundaries for Banach spaces determine weak compactness, Invent. Math. 182
(2010), no. 3, 585 -- 604. MR 2737706
[12] A. Plichko, Three sequential properties of dual Banach spaces in the weak∗ topology, Topology
Appl. 190 (2015), 93 -- 98. MR 3349508
[13] A. M. Plichko and D. Yost, Complemented and uncomplemented subspaces of Banach spaces,
Extracta Math. 15 (2000), no. 2, 335 -- 371. MR 1823896
[14] J. Rodr´ıguez and G. Vera, Uniqueness of measure extensions in Banach spaces, Studia Math.
175 (2006), no. 2, 139 -- 155. MR 2261730
[15] Th. Schlumprecht, On dual spaces with bounded sequences without weak∗-convergent convex
blocks, Proc. Amer. Math. Soc. 107 (1989), no. 2, 395 -- 408. MR 979052
[16] E. K. van Douwen, The integers and topology, Handbook of set-theoretic topology, North-
Holland, Amsterdam, 1984, pp. 111 -- 167. MR 776622
[17] D. Yost, The Johnson-Lindenstrauss space, Extracta Math. 12 (1997), no. 2, 185 -- 192.
MR 1607165
14
A. AVIL ´ES, G. MART´INEZ-CERVANTES, AND J. RODR´IGUEZ
[18] D. Yost, A different Johnson-Lindenstrauss space, New Zealand J. Math. 37 (2008), 47 -- 49.
MR 2406564
[19] V. Zizler, Nonseparable Banach spaces, Handbook of the geometry of Banach spaces, Vol. 2,
North-Holland, Amsterdam, 2003, pp. 1743 -- 1816. MR 1999608
Dpto. de Matem´aticas, Facultad de Matem´aticas, Universidad de Murcia, 30100 Es-
pinardo (Murcia), Spain
E-mail address: [email protected]
Dpto. de Matem´aticas, Facultad de Matem´aticas, Universidad de Murcia, 30100 Es-
pinardo (Murcia), Spain
E-mail address: [email protected]
Dpto. de Ingenier´ıa y Tecnolog´ıa de Computadores, Facultad de Inform´atica, Uni-
versidad de Murcia, 30100 Espinardo (Murcia), Spain
E-mail address: [email protected]
|
1503.08690 | 1 | 1503 | 2015-03-30T14:44:43 | Correlation Minimizing Frames in Small Dimensions | [
"math.FA"
] | A uniform tight frame of N vectors for a d dimensional space is correlation minimizing if among all such frames it is as "nearly" orthogonal as possible, i.e., it minimizes the maximal inner product of unequal vectors. In this paper we begin to catalog these frames for small dimensions, in particular, d=3. | math.FA | math |
CORRELATION MINIMIZING FRAMES IN SMALL
DIMENSIONS
GRANT GETZELMAN, NICOLE L. LEONHARD, AND VERN I. PAULSEN
Abstract. A uniform tight frame of N vectors for a d dimensional
space is correlation minimizing if among all such frames it is as "nearly"
orthogonal as possible, i.e., it minimizes the maximal inner product of
unequal vectors. In this paper we begin to catalog these frames for small
dimensions, in particular, d=3.
1. Introduction
In their study of the erasures problem, Holmes and the third author [8]
proved that a certain family of frames were optimal for 2 erasures and called
these the 2-optimal frames. Briefly, these are uniform Parseval frames for
which the minimal angle between any pair of vectors is as large as possi-
ble. Thus, the 2-optimal frames are exactly the frames that are correlation
minimizing among all frames in their family of frames.
They proved that when equiangular frames exist, then these are always
2-optimal and conversely. But very few examples are known of 2-optimal
frames in the cases when equiangular frames do not exist.
In particular,
even in dimension 3, very few 2-optimal frames are known and even less is
known about "uniqueness" of such frames, i.e., when up to some natural
notion of "equivalence" of frames a 2-optimal frame is unique. We provide
additional examples of 2-optimal frames and in some cases we are able to
prove uniqueness up to equivalence.
Optimal packings of N lines in R3 were studied by Conway, Hardin and
Sloan [6] for all values of N ≤ 55. For some values of N they were able to give
closed form descriptions of these sets of lines, along with proofs that they
were indeed optimal packings, while for many values of N they were only
able to give numerical approximations to these optimal packings. Holmes
and Paulsen [8] also did numerical experiments that attempted to compute
the minimum angle between vectors for 2-optimal frames of N vectors in R3.
Their computations showed that for some values of N the minimum angle
between vectors for 2-optimal frames appeared to be identical to the angle
determined by [6] for optimal line packings. This lead them to conjecture
that for these values on N , one could obtain uniform tight frames by choosing
a unit vector from each line in the optimal line packing.
2000 Mathematics Subject Classification. Primary 46L15; Secondary 47L25.
1
2
G. GETZELMAN, N. L. LEONHARD, AND V. I. PAULSEN
In this paper we find some of these 2-optimal frames in R3 for some
values of N and prove that for some of the values of N where these two
angles were shown to numerically agree, that one does indeed obtain tight
frames, which are necessarily 2-optimal, by choosing unit vectors from the
optimal line packing. Also, in some cases earlier work only provided the
Grammian matrices of these 2-optimal frames, but for the convenience of
the reader who might be interested in experimenting with these frames, we
also produce the actual frames and some geometric descriptions of the sets
of vectors.
In the next section we review the necessary background information needed
and then in the final section we prove the results claimed above and produce
the frames.
2. Background
We briefly review the key concepts from frame theory and the paper [8]
that we shall need.
2.1. Frames. A family F = {fi}i∈I of elements in a (real or complex)
Hilbert space H is called a frame for H
if there are constants 0 < A ≤
¯
B < ∞, called the lower and upper frame bounds, respectively so that for
all f ∈ H
(1)
hf, fii2 ≤ Bkfk2.
Akfk2 ≤Xi∈I
In general, a frame can have more vectors than the dimension of the
Hilbert space and, in the case that the space is finite dimensional, we call
the redundancy of the frame.
card(I)
dim(H)
If A = B, then F is called a tight frame and when A = B = 1, F is called
a Parseval frame. Thus, if F = {fi}i∈I is a tight frame with constant A
then { 1√A
While many authors prefer to work with unit norm tight frames, we will
mainly consider uniform Parseval frames. As we see above this is just a
matter of scaling.
fi}i∈I is a Parseval frame.
If F = {fi}i∈I is a frame for H, the analysis operator is the bounded
linear operator V : H → ℓ2(I) given by V (x)i = hx, fii for all i ∈ I. The
adjoint of the analysis operator, V ∗ : ℓ2(I) → H is defined by the formula,
V ∗(ei) = fi, where ei is the vector that is 1 in the i-th entry and 0 elsewhere.
Hence,
V ∗V x =Xi∈I
hx, fiifi.
In particular, F is a Parseval frame if and only if V is an isometry and
this is if and only if V ∗V = IH.
Thus, F = {fi}i∈I is a Parseval frame if and only if we have that
h =Xi∈I
hh, fiifi, ∀h ∈ H.
More generally, if F = {fi}i∈I is a tight frame for a Hilbert space H with
constant A then
3
h =
1
√AXi∈I
hh, fiifi, ∀h ∈ H.
This is known as the sampling and reconstruction formula.
On the other hand the analysis operator V is an isometry if and only if
the Grammian matrix V V ∗ = (hfj, fii) is a projection with rank equal to
dim(H), which gives another characterization of Parseval frames.
If F = {f1, ..., fN} is a Parseval frame for a d-dimensional Hilbert space,
then to shorten terminology we shall simply call F a (N, d)-frame. If F is
a uniform (N, d)-frame with analysis operator V , then V ∗V = Id the d × d
identity matrix and
d = rank(V V ∗) = T r(V V ∗) =
N
Xi=1
kfik2 = Nkfkk2,
for any k. Thus, for a (N, d)-frame,
is the redundancy.
1
kfkk2 =
N
d
,
infinite dimensional Hilbert space, we still call
For this reason, when F = {fi}i∈I is a uniform Parseval frame for an
kfkk2 the frame redundancy.
Following the notation of [8], we let E1(N, d) denote the set of all uniform
(N, d)-frames for Rd(respectively, Cd). This notation comes from a result
of Casazza and Kovalecevic [5] that the uniform Parseval frames are in a
certain sense optimal for the 1-erasure problem.
1
Given an indexed set S = {vi}i∈I of non-zero vectors, their maximum
correlation, denoted M∞(S), is defined as
vl
(2)
,
M∞(S) = sup{h
vk
kvkk
kvlki : k, l ∈ I, k 6= l}
so that
Θ(S) := arccos(M∞(S))
is the infimum of the angles between pairs of vectors. Thus, 0 ≤ M∞(S) ≤ 1
and M∞(S) = 0 if and only if Θ(S) = π
2 if and only if S is an orthogonal set.
Thus, smaller maximum correlation means a set is more nearly orthogonal.
If M∞(S) = 1 and the supremum is attained, then any two vectors where the
supremum is attained are parallel. So larger maximum correlation indicates
that the set contains vectors that are more nearly parallel.
4
G. GETZELMAN, N. L. LEONHARD, AND V. I. PAULSEN
Given F = {f1, ...fN} ∈ E1(N, d) we have
M∞(F) =
N
d · max{hfk, fli : k 6= l}.
Note that the factor N
vector in F has norm pd/N .
We set
d is included because F ∈ E1(N, d) implies that each
(3)
and
C(N, d) = inf{M∞(F) : F ∈ E1(N, d)}.
Note that since arccos is a decreasing function,
Θ(N, d) = arccos(C(N, d)).
Θ(N, d) = sup{Θ(F) : F ∈ E1(N, d)}.
We shall call C(N, d) the correlation constant for frames in E1(N, d) and so
Θ(N, d) is the maximum angle between vectors for frames in E1(N, d).
These constants were introduced in [8] where it was proven that C(N, d)
was always attained, i.e., the infimum is actually a minimum, and any uni-
form Parseval frame where it is attained was called 2-optimal. This termi-
nology arose from the result from [8] that such frames were optimal for the
2-erasures problem.
Perhaps a more descriptive terminology would have been to call a uniform
Parseval frame where this minimum correlation is attained a correlation
minimizing uniform Parseval frame or angle maximizing uniform Parseval
frame. In which case the result from [8] is that a uniform Parseval frame is
correlation minimizing if and only if it is optimal for the 2-erasure problem.
In any case, since we will not be considering the erasures problem in this
paper, we prefer to use the term correlation minimizing. Thus, a uniform
Parseval frame is correlation minimizing if and only if it is in the set
E2(N, d) = {F ∈ E1(N, d) : M∞(F) = C(N, d)}
and a uniform Parseval frame is correlation minimizing if and only if it is
2-optimal. In [3] the correlation minimizing frames are called Grassmannian
Parseval frames.
A frame is equiangular if for some α,hfi, fji = α for all i6= j.
Theorem 2.1. [8] Let N ≥ d, and let F ∈ E1(N, d). Then
(4)
M∞(F) ≥s N − d
d(N − 1)
,
and equality holds iff F is equiangular.
If there exists an equiangular frame F ∈ E1(N, d), then it is 2-optimal, i.e.,
correlation minimizing and in this case every frame in E2(N, d) is equian-
gular.
Futhermore, if N > d(d+1)
in the real case and N > d2 in the complex case,
2
then there is no equiangular frame in E1(N, d) and equality cannot hold in
the above equation.
Thus, we see that
5
C(N, d) ≥s N − d
d(N − 1)
,
with equality if and only if there exists an equiangular frame. The quantity
appearing on the right hand side of the above equation is known as the
Welch bound.
Since C(N, d) is the actual lower bound for the maximum correlation of
uniform Parseval frames and, consequently, for unit norm tight frames, it
should be an important constant to compute. But surprisingly very little is
known about it other than the numerical estimates in [8].
i=1 and G = {gi}n
The goal of this paper is to find more examples of correlation minimizing
uniform Parseval frames and to determine if they are unique, modulo an
equivalence relation, which we now explain.
Definition 2.2. Frames F = {fi}n
i=1, are type I equivalent
if there exists a unitary (orthogonal matrix, in the real case) U such that
gi = U fi for all i.
Theorem 2.3. [8] If V and W are the analysis operators for F and G,
respectively, then the following are equivalent
1. F and G are type I equivalent
2. there exists a unitary(respectively, orthogonal matrix) U such that V =
W U
3. V V ∗ = W W ∗.
i=1 and G = {gi}n
In [8] it is shown that there is a one-to-one correspondence between N ×N
rank d projections and type I equivalence classes of uniform (N, d)-frames.
Definition 2.4. Frames F = {fi}n
i=1 are type II equivalent
if they are a permutation of the same set of vectors and they are type III
equivalent if there exist numbers {λi}n
i=1 of modulus such that fi = λigi.
Thus, in the real case if they differ by multiplication by ±1. Two frames are
equivalent if they belong to the same equivalence class in the equivalence
relation generated by these three equivalence relations.
Theorem 2.5. [8] If F and G are (N, d)-frames with analysis operators
V and W , respectively, then they are equivalent if and only if U V V ∗U∗ =
W W ∗ for some N ×N unitary U that is the product of a permutation matrix
and a diagonal matrix with entries of modulus 1(±1, in the real case).
3. Grassmannian and Correlation Minimizing Frames in R3
The line packing problem is the problem of packing N lines in Rd so that
the minimal angle between any two of them is as large as possible. Any
solution to this problem is called a Grassmannian line packing. Given a
6
G. GETZELMAN, N. L. LEONHARD, AND V. I. PAULSEN
Grassmannian line packing with N ≥ d if we choose one unit vector from
each line, then this set of vectors always yields a frame for Rd. Any frame
obtained this way is called a Grassmannian frame by [11].
If a Grassmannian frame happens to be a tight frame, then after scaling
the vectors bypd/N we would obtain a uniform (N, d)-frame that is neces-
sarily correlation minimizing. Thus, Grassmannian frames yield 2-optimal
frames by scaling if and only if they are tight frames.
A geometric approach to solving the line packing problem and list of best-
known packings is posted on [10]. Conway, Hardin, and Sloane [6] find the
Grassmannina line packings of N lines through the origin in R3, describe
the packings and compute this minimal angle for 2 ≤ N ≤ 55. So a natu-
ral question that we shall study below is whether or not the Grassmannian
frames arising from these Grassmannian line packings are tight. The numer-
ical experiments of [8] indicates that the answer should be "yes" for some
values of N and "no" for other values.
In [8] the uniform (N, 2)-frames that are correlation minimizing were con-
structed, it was shown that these frames form a single equivalence class, and
that these are Grassmannian.
First note that whenever N = d then any Parseval frame must be an
orthonormal basis, since the frame operator V will be an isometry from Rd
to Rd and hence will be an orthogonal matrix. Hence, the rows of V will be
an orthonormal set. Moreover, every orthonormal basis is type I equivalent.
Thus, there is a unique equivalence class of (d, d)-frames and an orthonormal
basis is clearly correlation minimizing and Grassmannian.
So the first interesting case is the (4, 3)-frames. For this a theorem is
useful.
Theorem 3.1. Let F = {fi}N
i=1 be a correlation minimizing (N, d)-frame
with Grammian matrix G. Then IN − G is the Grammian matrix of a
correlation minimizing (N, N − d)-frame. Moreover, there is a one-to-one
correspondence between equivalence classes of correlation minimizing (N, d)-
frames and equivalence classes of correlation minimizing (N, N − d)-frames.
Proof. Let F = {fi}N
i=1 be any uniform (N, d)-frame with Grammian matrix
G. Then G is a rank d projection all of whose diagonal entries are equal to
d
N . Hence, IN − G is a rank N − d projection all of whose diagonal entries
are N−d
N . Hence, if we let W W ∗ = IN − G be any factorization, then by the
results of the last section the rows of W (or their complex conjugates in the
complex case) form a uniform (N, N − d)-frame whose Grammian is IN − G.
This frame is not uniquely determined by IN − G, since many factorizations
are possible, but it is unique up to type I equivalence by Theorem 2.3.
So choose one such uniform (N, N − d)-frame and denote it by F⊥. Now
if Fi i = 1, 2 are any uniform (N, d)-frames with Grammians Gi i = 1, 2 and
we let F⊥i
i = 1, 2 be (N, N −d)-frames with Grammians IN −Gi obtained as
above, then, by applying Theorem 2.5, we see that F1 and F2 are equivalent
if and only if F⊥1 and F⊥2 are equivalent.
Finally, since the maximum correlation of a frame is really just the maxi-
mum off-diagonal entry of its Grammian(appropriately scaled), we see that
7
M∞(F⊥) =
d
N − d
M∞(F),
and so whenever F is a correlation minimizing uniform (N, d)-frame that
F⊥ is a correlation minimizing uniform (N, N − d)-frame.
Remark 3.2. The above proof also shows that
(cid:3)
C(N, N − d) =
N − d
Let JN denote the N × N matrix of all 1's.
d
C(N, d).
Corollary 3.3. Up to equivalence there is a unique correlation minimizing
(N, 1)-frame and a unique correlation minimizing (N, N − 1)-frame. These
equivalence classes are represented by the uniform frames with Grammians
N JN and IN − 1
1
N JN , respectively. Moreover, both these frames are equian-
gular and so these frames are also Grassmannian.
Proof. To obtain a uniform (N, 1)-frame one must choose N numbers of
modulus 1/√N . But these are all equivalent to choosing the number 1/√N
N -times. Thus, up to equivalence there is only one uniform (N, 1)-frame
and it has Grammian 1
Hence, by the above theorem, up to equivalence there is only one (N, N −
All these frames are equiangular since all the off-diagonal entries in their
(cid:3)
1)-frame and it has Grammian given by IN − 1
Grammians are of constant modulus 1
N .
N JN .
N JN .
We can now give a concrete description of one representative of this equiv-
alence class of frames in the case N = 4, d = N − 1 = 3.
Theorem 3.4. The lines generated by opposite vertices of the inscribed cube
in the sphere centered at the origin is the optimal packing of 4 lines in 3-
√3
2 centered at the origin and consider
space. If we take the sphere of radius
the 8 vectors determined by the vertices of this cube, then any set of 4 of these
vectors that are not collinear yields a correlation minimizing, equiangular
(4, 3)-frame. In particular, one correlation minimizing, equiangular (4, 3)-
frame is given by
1
2
), (−
), (−
,−
,−
,−
,−
), (+
, +
, +
(+
, +
, +
1
2
)
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
and every other correlation minimizing (4,3)-frame is equivalent to this
frame.
Proof. Let
f1∗
f2∗
f3∗
f4∗
V =
=
2
+ 1
2 + 1
− 1
2 − 1
− 1
2 + 1
2 − 1
+ 1
2 + 1
2 + 1
2 − 1
2 − 1
2
2
2
.
8
G. GETZELMAN, N. L. LEONHARD, AND V. I. PAULSEN
Computing the Grammian yields G = V V ∗ = I4 − 1
representative of the unique correlation minimizing (4, 3)-frame.
4 J4 so that this is one
The remaining claims are now straightforward to verify.
(cid:3)
Theorem 3.5. A correlation minimizing (5, 3)-frame is given by the vectors:
c(1, 0, 0), c(−1 − √5
, −5 − 3√5
1 − √5
15 − √5
150 − 30√5
, 0)
c(
a
6
,
,
6
a
, −60√5
150 − 30√5
where a = q18(cid:0)15 − √5(cid:1), b = p150 − 30√5, and c = p3/5. Every other
), c(−1 + √5
,
6
1 + √5
correlation minimizing (5, 3)-frame is equivalent to this frame.
5 − 3√5
a
4√5
a
ab
ab
ab
c(
),
6
)
,
,
Proof. From [8] we have that the correlation minimizing (5, 2)-frame is unique
up to equivalence and one representative is given by the vectors
{(cid:0)cos(cid:18) πk
5 (cid:19) , sin(cid:18) πk
5 (cid:19)(cid:1) : k = 1, 2, 3, 4, 5}.
Thus, by Theorem 3.1 the correlation minimizing (5, 3)-frame will be unique
up to equivalence and a representative Grammian will be given by G =
I5 − G(5,2), where G(5,2) is the Grammian of the above vectors.
Computing this Grammian yields,
3/5
2
5 cos(π/5) 2
3/5
2
2
5 cos(π/5)
5 cos(2π/5) 2
5 cos(3π/5) 2
5 cos(4π/5) 2
2
2
5 cos(π/5)
5 cos(2π/5) 2
5 cos(3π/5) 2
5 cos(2π/5) 2
5 cos(π/5) 2
2
3/5
2
5 cos(4π/5)
5 cos(3π/5)
5 cos(2π/5) f rac25cos(3π/5)
5 cos(π/5)
2
2
5 cos(2π/5)
5 cos(π/5)
2
5 cos(π/5)
5 cos(2π/5) 2
3/5
5 cos(π/5)
G =
This can be factored as G = 3
U =
with a and b as above.
6
−1−√5
1−√5
−1+√5
1+√5
6
6
6
5 U U∗ where
1
0
a
15−√5
−5−3√5
5−3√5
a
4√5
a
a
0
0
ab
150−30√5
−60√5
150−30√5
ab
ab
.
(cid:3)
3/5
,
Corollary 3.6. The Grassmannian frame of 5 vectors in R3 is not a tight
frame.
Proof. By inspection the largest off diagonal entry of the above G is 2
and the smallest angle produced by the vectors of this Grammian is equal
to arccos( 2
3 cos(π/5)) which is approximately 57.361 degrees. This is not
equal to the angle of the optimal packing of 5 lines found in [6]. Thus, if we
5 cos(π/5),
9
take one unit vector from each of the 5 lines corresponding to the optimal
packing of 5 lines through the origin in R3, then this set of vectors can not
be a tight frame since its correlation is smaller.
(cid:3)
Thus, the correlation minimizing (5, 3)-frame is an example that is not
obtained via the optimal line packing. In the language of [11] the correlation
minimizing (5, 3)-frame is not a Grassmannian frame. In the language of [3]
the Grassmannian Parseval frame of 5 vectors in R3 is not a Grassmannian
frame.
In [2] it was shown that the correlation minimizing (6, 3)-frame is equian-
gular, that it is unique up to equivalence and its Grammian was given. Below
we give a geometric description of the set of vectors for one representative
of this equivalence class and give the vectors explicitly.
Theorem 3.7. The 6 vertices, that lie in the upper half plane of an icosa-
hedron centered at the origin and symmetric about the xy-plane, form a
correlation minimizing (6, 3)-frame. Set α = 1√5
, then these are the vectors
given by:
f1 = 1√2
(0, 0, 1),
, α(cid:19),
1+α ,q (1+2α)(1−α)
, α(cid:19),
1+α ,−q (1+2α)(1−α)
, α(cid:19),
1+α ,q (1−2α)(1+α)
, α(cid:19).
1−α ,−q (1−2α)(1+α)
f2 = 1√2(cid:16)√1 − α2, 0, α(cid:17),
f3 = 1√2(cid:18)αq 1−α
f4 = 1√2(cid:18)αq 1−α
f5 = 1√2(cid:18)−αq 1−α
f6 = 1√2(cid:18)−αq 1+α
Proof. From [2] we have the vectors defined above. For k 6= l, we compute
hfk, fli = 1√5
. Thus, this set of vectors is equiangular and and each vector
√3√6
has norm
Every other correlation minimizing (6, 3)-frame is equivalent to this frame.
so these must be a correlation minimizing (6, 3)-frame.
1+α
1+α
1−α
1−α
In [6] it is observed that the 6 lines obtained by taking antipodal pairs of
(cid:3)
points on an icosahedron are equiangular.
Before constructing a correlation minimizing (7, 3)-frame a little proposi-
tion will be useful.
Proposition 3.8. If {f1, . . . , fN} is a uniform (N, d)-frame and {g1, . . . , gM}
is a uniform (M, d)-frame, then {af1, . . . , afN , bg1, . . . , bgM} is a uniform
(M + N, d)-frame, where a =pN/(N + M ) and b =pM/(N + M ).
Proof. Since kfik =pd/N and kgjk =pd/M we have that kafik = kbgjk =
pd/(N + M ), so this set of vectors is uniform in norm. Finally, for any
10
G. GETZELMAN, N. L. LEONHARD, AND V. I. PAULSEN
vector x ∈ Rd, we have that
Xj=1
hx, afii2 +
Xi=1
M
N
hx, bgji2 = a2kxk2 + b2kxk2 = kxk2,
so the Parseval condition is met.
Theorem 3.9. Let {f1, f2, f3, f4} be the correlation minimizing (4, 3)-frame
of Theorem 3.4 and let {e1, e2, e3} be the standard orthonormal basis for R3,
then {p4/7 f1,p4/7 f2,p4/7 f3,p4/7 f4,p3/7 e1,p3/7 e2,p3/7 e3} is a
correlation minimizing (7, 3)-frame.
(cid:3)
Proof. By the above proposition, this set of vectors is a uniform (7, 3)-
frame. The inner products of pairs of these unequal vectors take on the
√3
3 . Since
values, {0, ±
cos−1(cid:16)√3
3 (cid:17) corresponds to the minimum angle for the Rhombic Dodecahe-
dron [7], which is an optimal line packing angle for 7 lines in 3 space found
by [6], this uniform Parseval frame must be correlation minimizing.
(cid:3)
7}, so that for this frame, M∞(F) = 7
√3
7 , ± 1
√3
7 =
3
Since the correlation minimizing (7,3)-frame corresponds to an optimal
line packing, every correlation minimizing (7,3)-frame would yield an op-
timal line packing. But we do not know if every correlation minimizing
(7,3)-frame is equivalent to this frame. In [6], they remark that the optimal
packing of 7 lines in 3 space appears to be unique, but do not supply a proof.
A related, and possibly easier, problem would be to decide if every optimal
line packing of 7 lines in 3 space yields a tight frame.
The optimal line packing for 10 lines in R3 is given numerically on Sloane's
web site [10]. In [6], it was determined that there are infinitely many solu-
tions to this optimal line packing problem. This occurs because the axial
line can rattle, that is, the vectors can move freely over a small range of
angles without affecting the minimum angle.
Theorem 3.10. The optimal line packing for 10 lines in R3 comprised of
2 axis vectors and the set of 8 vectors that are not collinear from the scaled
hexakis bi-antiprism, given by
, 0! ,(cid:18)β, 0,±βq√3 − 1(cid:19)
,
1
2
√3
2
(1, 0, 0) , (0,−1, 0) ,
±
√3
, βq√3 − 1! , −
2
,−β
β
√3
2
, β
2
β
2
,−β
,−βq√3 − 1! , −
β
2
√3
2
, βq√3 − 1! ,
√3
2
β
2
, β
,−βq√3 − 1! .
where β = 3− 1
of the axis that will create it a tight frame.
4 , is not a tight frame. Moreover, there does not exist a rattle
11
Proof. First, we construct the hexakis bi-antiprim by taking two hexagonal
antiprisms and joining them at the base. To create the first half, shift the
coordinates for the hexigonal antiprism from [9]. For the second half, we use
a shift and rotation of the same coordinates from [9]. Now, we join them
at the base to complete the construction. From the set of 18 unique scaled
vectors in the construction we consider 2 axis together with the set of 8
vectors that are not collinear. Set β = 3− 1
4 and define
V =
0
0
0
0
0
−1
1
0
√3
−
2
√3
2
β
β
β
2
1
2
1
2
0
0
√3
β
2
√3
− β
2 −β
2
√3
−β
√3
− β
β
2
β
2
2
βp√3 − 1
−βp√3 − 1
βp√3 − 1
βp√3 − 1
2 −βp√3 − 1
−βp√3 − 1
Recall that a set of vectors forms a uniform tight frame if and only if they
are of equal norm and when they are entered as the rows of a matrix, then
that matrix is a multiple of an isometry. Moreover, to be a multiple of an
isometry, the columns of the matrix must be orthogonal and of equal norm.
The rows of V are unit norm. By inspection we see the columns are
orthogonal. However, the columns of V do not have equal norm. Hence, no
multiple of V is an isometry and so the rows are not a tight frame.
Now, we will consider the case where the axial lines "rattle" to try to
gain equality in the norm of the columns. Consider the first row as v1 =
(a1, b1, c1) and the second as v2 = (a2, b2, c2). Since the vectors comprising
the last eight entries of V are orthogonal, to keep the columns orthogonal
we will need the vectors (a1, a2), (b1, b2) and (c1, c2) to be orthogonal. Since
this is three vectors in R2, one of them must be zero. The norm of the
first column is the largest so a1 = a2 = 0. The rows must be unit norm so
1 + c2
b2
2 = 1. We still need the norms of the three columns
to be equal. Thus, we get the system of equations.
1 = 1 and b2
2 + c2
1 = 1
2 = 1
2 = 1
1 + c2
b2
b2
2 + c2
1 + b2
b2
2 = 3√3 − 9
1 + c2
c2
2
.
By subtracting the third equation from the first we see that c2
ging into equation 2 we get c2
2. Plug-
2 = 1, which contradicts the fourth equation.
1 = b2
1+c2
12
G. GETZELMAN, N. L. LEONHARD, AND V. I. PAULSEN
Therefore, there is no choice of vectors that can make V a multiple of an
isometry.
(cid:3)
Theorem 3.11. A correlation minimizing (12, 3)-frame is given by scaling
the set vertices of the rhombicuboctahedron through the origin to be vectors
of length 1
2 . This frame is given by all non-collinear permutations of the
vectors 1
2(cid:18)± 1√2√2+5
,± 1√2√2+5
,± (1+√2)
√2√2+5(cid:19) .
Proof. In [6], the optimal line packing of 12 lines in R3 is a rhombicubocta-
hedron. Define V such that the rows are the vectors of the rhombicubocta-
hedron in [7] . So, V = 1
2
. Each row
V is of norm 1
2 . Additionally, we see the columns are orthogonal and of
equal norm. Therefore, V is an isometry and we can conclude that the rows
form a uniform Parseval frame that is a correlation minimizing frame. (cid:3)
Theorem 3.12. A correlation minimizing (16, 3)-frame is given by scaling
the set of unit norm opposite vertices of the Biscribed Pentakis Dodecahe-
dron with radius one centered at the origin. This frame is given by scaling
the vectors
(0, c0,±c4) , (c4, 0,±c0) , (c0,±c4, 0) , (c1, 0,±c3) , (c3,±c1, 0) , (0, c3,±c1) ,
(0,−c3,−c1) (c2, c2, c2) , (c2,−c2,−c2) , (−c2, c2,−c2) , (−c2,−c2, c2) .
√15+√3
√15−√3
c0 =
by the factor
q10(5−√5)
√3
3 ,c3 =
q10(5+√5)
, and c4 =
10
, c2 =
6
, c1 =
√3
4 .
10
6
(1+√2)
1√2√2+5
1√2√2+5
√2√2+5
−1−√2
1√2√2+5
1√2√2+5
√2√2+5
(1+√2)
1√2√2+5 − 1√2√2+5
√2√2+5
1√2√2+5 − 1√2√2+5 − (1+√2)
√2√2+5
(1+√2)
1√2√2+5
1√2√2+5
√2√2+5
(1+√2)
√2√2+5 − 1√2√2+5
1√2√2+5
−1− 2√2
1√2√2+5
1√2√2+5
√2√2+5
1√2√2+5 − (1+√2)
√2√2+5 − 1√2√2+5
(1+√2)
1√2√2+5
1√2√2+5
√2√2+5
(1+√2)
1√2√2+5 − 1√2√2+5
√2√2+5
(1+√2)
1√2√2+5
√2√2+5 − 1√2√2+5
(1+√2)
√2√2+5 − 1√2√2+5 − 1√2√2+5
13
Proof. Let the columns of W ∗ be the opposite vertices of the Biscribed
Pentakis Dodecahedron with radius one centered at the origin in [9]. Set
√15+√3
q10(5+√5)
q10(5−√5)
10
, and c4 =
.
6
√15−√3
c0 =
It follows that,
6
, c1 =
10
, c2 =
√3
3 ,c3 =
c0
c0
c4 −c4
0
0
W ∗ =
0
0
c0
c0
c4 −c4
c4
c4
0
0
c0 −c0
c2
c2 −c2
c2 −c2 −c2
By inspection we see that the columns of W ∗ are unit norm, the rows are
√3
equal norm and the rows are orthogonal. Hence, V =
4 W is an isometry
and so its rows are a uniform (16, 3)-frame. Since these vectors are the ver-
tices of the the optimal line packing in [6] they form a (16, 3)-frame, which
is correlation minimizing.
(cid:3)
c1
c1
0
0
c3 −c3
0
0
c3 −c3
c1 −c1
c3
c3
c1 −c1
0
0
c2 −c2 −c2
c2 −c2
c2
.
Remark 3.13. We do not know if the correlation minimizing (12, 3)-frame
and (16, 3)-frame are unique up to frame equivalence. This is largely because
it is unknown if the corresponding arrangements for the optimal packings of
16 lines and 12 lines, respectively, in R3 are unique.
References
[1] J. J. Benedetto and M. Fickus, Finite normalized tight frames, Adv. Comput. Math.,
vol. 18, no. 24, pp. 357-385, 2003.
[2] J. J. Benedetto and J. D. Kolesar, Geometric properties of Grassmannian frames for
R2 and R3, EURASIP Journal on Applied Signal Processing 4985 (2006).
[3] Bernhard G. Bodmann and John Hass, Frame potentials and the geometry of frames,
arXiv:1407.1663v1 [math.FA] 7 Jul 2014.
[4] Bernhard G. Bodmann and Vern I. Paulsen, Frames, graphs and erasures, Linear
Algebra Appl. 404 , 118-146, 2005.
[5] P. Casazza and J. Kovalecevic, Uniform tight frames with erasures, Adv. Com-
put.Math. 18 (2003), 387430.
[6] J. H. Conway, R H. Hardin, and N. J. A. Sloane, Packing lines, planes, etc.: packings
in Grassmannian spaces, Experiment. Math., vol. 5, no.2, pp. 139159, 1996.
[7] R.W. Gray, http://www.rwgrayprojects.com/rbfnotes/polyhed/PolyhedraData/
[8] R. Holmes and V. Paulsen, Optimal frames for erasures, Linear Algebra and its
Applications, vol. 377, pp. 3151, 2004.
[9] David I. McCooey, http://dmccooey.com/polyhedra
[10] N.J.A.Sloane, http://neilsloane.com/grass/grassTab.html
[11] T. Strohmer and R. W. Heath, Jr., Grassmannian frames with applications to coding
and communication, Appl. Comput. Harmon. Anal., vol. 14, no. 3, pp. 257275, 2003.
Department of Mathematics, University of Houston, Houston, Texas 77204-
3476, U.S.A.
Department of Mathematics, University of Houston, Houston, Texas 77204-
3476, U.S.A.
E-mail address: [email protected]
Department of Mathematics, University of Houston, Houston, Texas 77204-
3476, U.S.A.
E-mail address: [email protected]
|
1202.1628 | 1 | 1202 | 2012-02-08T08:54:34 | Strong convergence theorems for strongly relatively nonexpansive sequences and applications | [
"math.FA",
"math.OC"
] | The aim of this paper is to establish strong convergence theorems for a strongly relatively nonexpansive sequence in a smooth and uniformly convex Banach space. Then we employ our results to approximate solutions of the zero point problem for a maximal monotone operator and the fixed point problem for a relatively nonexpansive mapping. | math.FA | math |
STRONG CONVERGENCE THEOREMS FOR STRONGLY
RELATIVELY NONEXPANSIVE SEQUENCES AND
APPLICATIONS
KOJI AOYAMA, YASUNORI KIMURA, AND FUMIAKI KOHSAKA
Abstract. The aim of this paper is to establish strong convergence theorems
for a strongly relatively nonexpansive sequence in a smooth and uniformly
convex Banach space. Then we employ our results to approximate solutions of
the zero point problem for a maximal monotone operator and the fixed point
problem for a relatively nonexpansive mapping.
1. Introduction
xn+1 = J −1(cid:18) 1
n
Jx +(cid:16)1 −
1
n(cid:17)J(J + rnA)−1Jxn(cid:19)
Let E be a smooth and uniformly convex Banach space, E ∗ the dual of E,
A ⊂ E × E ∗ a maximal monotone operator with a zero point, and {rn} a sequence
of positive real numbers. Assume that {xn} is a sequence defined as follows: x1 ∈ E
and
for n ∈ N, where J and J −1 are the duality mappings of E and E ∗, respectively. It
is known [9] that if rn → ∞, then {xn} converges strongly to some zero point of A.
However, we have not known whether {xn} converges strongly or not without the
assumption that rn → ∞. In §5 we present an affirmative answer to this problem;
see Theorem 5.2 and Remark 5.3.
Furthermore, more general results are proved; see Theorem 4.1. This is a strong
convergence theorem for a strongly relatively nonexpansive sequence introduced
in [3].
In the proofs of Theorem 4.1, we use modifications of ideas developed
in [10, 16]. In particular, Lemma 3.2 due to Maing´e [10] is a fundamental tool; see
also Example 3.3 and Lemma 3.4.
In §5, using Theorem 4.1, we also show Theorem 5.5 which is a strong convergence
theorem for a relatively nonexpansive mapping in the sense of Matsushita and
Takahashi [11].
2. Preliminaries
Throughout the present paper, E denotes a real Banach space with norm k · k,
E ∗ the dual of E, hx, x∗i the value of x∗ ∈ E ∗ at x ∈ E, and N the set of positive
integers. The norm of E ∗ is also denoted by k · k. Strong convergence of a sequence
{xn} in E to x ∈ E is denoted by xn → x and weak convergence by xn ⇀ x.
Date: March 29, 2018.
2010 Mathematics Subject Classification. 47H09, 47H10, 41A65.
Key words and phrases. Strongly relatively nonexpansive sequence, common fixed point, strong
convergence theorem.
1
2
KOJI AOYAMA, YASUNORI KIMURA, AND FUMIAKI KOHSAKA
The (normalized) duality mapping of E is denoted by J, that is, it is a set-valued
mapping of E into E ∗ defined by
Jx = {x∗ ∈ E ∗ : hx, x∗i = kxk2 = kx∗k2}
for x ∈ E.
Let SE denote the unit sphere of E, that is, SE = {x ∈ E : kxk = 1}. The norm
k · k of E is said to be Gateaux differentiable if the limit
(2.1)
lim
t→0
kx + tyk − kxk
t
exists for all x, y ∈ SE. In this case E is said to be smooth and it is known that
the duality mapping J of E is single-valued. The norm of E is said to be uniformly
Gateaux differentiable if for each y ∈ SE the limit (2.1) is attained uniformly for
x ∈ SE. A Banach space E is said to be uniformly smooth if the limit (2.1) is
attained uniformly for x, y ∈ SE.
In this case it is known that J is uniformly
norm-to-norm continuous on each bounded subset of E; see [17] for more details.
A Banach space E is said to be strictly convex if x, y ∈ SE and x 6= y imply
kx + yk < 2. A Banach space E is said to be uniformly convex if for any ǫ > 0
there exists δ > 0 such that x, y ∈ SE and kx − yk ≥ ǫ imply kx + yk /2 ≤ 1 − δ.
It is known that E is reflexive and strictly convex if E is uniformly convex; E is
uniformly smooth if and only if E ∗ is uniformly convex; see [17] for more details.
In the rest of this section, unless otherwise stated, we assume that E is a smooth,
strictly convex, and reflexive Banach space. In this case it is known that the duality
mapping J of E is single-valued and bijective, and J −1 is the duality mapping of
E ∗.
We deal with a real-valued function φ on E × E defined by
φ(x, y) = kxk2 − 2 hx, Jyi + kyk2
for x, y ∈ E; see [1, 8]. From the definition of φ, it is clear that
(2.2)
for all x, y ∈ E. Since k · k2 is convex,
(kxk − kyk)2 ≤ φ(x, y)
(2.3)
(2.4)
holds for all x, y, w ∈ E and λ ∈ [0, 1]. It is known that
φ(cid:16)w, J −1(cid:0)λJx + (1 − λ)Jy(cid:1)(cid:17) ≤ λφ(w, x) + (1 − λ)φ(w, y)
φ(x, J −1x∗) ≤ φ(x, J −1(x∗ − y∗)) + 2(cid:10)J −1x∗ − x, y∗(cid:11)
holds for all x ∈ E and x∗, y∗ ∈ E ∗; see [9, Lemma 3.2].
Lemma 2.1 ([8, Proposition 2]). Let E be a smooth and uniformly convex Banach
space. Let {xn} and {yn} be bounded sequences in E.
If φ(xn, yn) → 0, then
xn − yn → 0.
Let {xn} and {yn} be bounded sequences in E. Then it is obvious from the
definition of φ that φ(xn, yn) → 0 if xn − yn → 0. From this fact and Lemma 2.1,
we deduce the following: If E is a uniformly convex and uniformly smooth Banach
space E, then
(2.5)
xn − yn → 0 ⇔ Jxn − Jyn → 0 ⇔ φ(xn, yn) → 0.
In the rest of this section, we assume that C is a nonempty closed convex subset
of E.
3
Let T : C → E be a mapping. The set of fixed points of T is denoted by F (T ).
A point p ∈ C is said to be an asymptotic fixed point of T [6, 14] if there exists a
sequence {xn} in C such that xn ⇀ p and xn − T xn → 0. The set of asymptotic
fixed points of T is denoted by F (T ). A mapping T is said to be of type (r) if
F (T ) 6= ∅ and φ(p, T x) ≤ φ(p, x) for all x ∈ C and p ∈ F (T ); T is said to be
relatively nonexpansive [11, 12] if T is of type (r) and F (T ) = F (T ). We know that
if T : C → E is of type (r), then F (T ) is closed and convex; see [12, Proposition 2.4].
It is known that, for each x ∈ E, there exists a unique point x0 ∈ C such that
φ(x0, x) = min{φ(y, x) : y ∈ C}.
Such a point x0 is denoted by QC(x) and QC is called the generalized projection
of E onto C; see [1, 8]. It is known that
(2.6)
or equivalently
hz − QC (x), Jx − JQC(x)i ≤ 0
(2.7)
φ(z, QC (x)) + φ(QC (x), x) ≤ φ(z, x)
holds for all x ∈ E and z ∈ C.
projection QC is of type (r).
It is obvious from (2.7) that the generalized
Let A be a set-valued mapping of E into E ∗, which is denoted by A ⊂ E × E ∗.
The effective domain of A is denoted by dom(A) and the range of A by R(A),
that is, dom(A) = {x ∈ E : Ax 6= ∅} and R(A) = Sx∈dom(A) Ax. A set-valued
mapping A ⊂ E × E ∗ is said to be a monotone operator if hx − y, x∗ − y∗i ≥ 0 for
all (x, x∗), (y, y∗) ∈ A. A monotone operator A ⊂ E × E ∗ is said to be maximal
if A = A′ whenever A′ ⊂ E × E ∗ is a monotone operator such that A ⊂ A′. It is
known that if A is a maximal monotone operator, then A−10 is closed and convex,
where A−10 = {x ∈ E : Ax ∋ 0}.
Let A ⊂ E × E ∗ be a maximal monotone operator and r > 0. Then it is known
that R(J + rA) = E ∗; see [15]. Thus a single-valued mapping Lr = (J + rA)−1J
of E onto dom(A) is well defined and is called the resolvent of A. It is also known
that F (Lr) = A−10 and
(2.8)
φ(u, Lrx) + φ(Lrx, x) ≤ φ(u, x)
for all x ∈ E and u ∈ F (Lr); see [7, 9]. It is obvious from (2.8) that the resolvent
Lr of A is of type (r) for all r > 0 whenever A−10 is nonempty.
The following lemma is well known; see [2, 18].
Lemma 2.2. Let {ξn} be a sequence of nonnegative real numbers, {γn} a sequence
of real numbers, and {αn} a sequence in [0, 1]. Suppose that ξn+1 ≤ (1 − αn)ξn +
αnγn for every n ∈ N, lim supn→∞ γn ≤ 0, and P∞
3. Eventually increasing functions and strongly relatively
n=1 αn = ∞. Then ξn → 0.
nonexpansive sequences
In this section, we provide some needed lemmas about an eventually increasing
function and a strongly relatively nonexpansive sequence.
A function τ : N → N is said to be eventually increasing if limn→∞ τ (n) = ∞
and τ (n) ≤ τ (n + 1) for all n ∈ N. By definition, we easily obtain the following:
Lemma 3.1. Let τ : N → N be an eventually increasing function and {αn} a
sequence of real numbers such that αn → 0. Then ατ (n) → 0.
4
KOJI AOYAMA, YASUNORI KIMURA, AND FUMIAKI KOHSAKA
We need the following lemma:
Lemma 3.2 (Maing´e [10, Lemma 3.1]). Let {ξn} be a sequence of real numbers.
Suppose that there exists a subsequence {ξni } of {ξn} such that ξni < ξni+1 for all
i ∈ N. Then there exist N ∈ N and a function τ : N → N such that τ (n) ≤ τ (n + 1),
ξτ (n) ≤ ξτ (n)+1, and ξn ≤ ξτ (n)+1 for all n ≥ N and limn→∞ τ (n) = ∞.
Under the assumptions of Lemma 3.2, we can not choose a strictly increasing
function τ ; see the following example:
Example 3.3. Let {ξn} be a sequence of real numbers define by
ξn =(0
if n is odd;
1/n if n is even.
Then the following hold:
(1) There exists a subsequence {ξni } of {ξn} such that ξni < ξni+1 for all i ∈ N;
(2) there does not exist a subsequence {ξmk } of {ξn} such that ξmk ≤ ξmk+1
and ξk ≤ ξmk+1 for all k ∈ N.
Proof. Define ni = 2i − 1 for each i ∈ N. Then it is clear that
ξni = ξ2i−1 = 0 <
1
2i
= ξ2i = ξni+1
for every i ∈ N. Thus (1) holds.
Let {ξmk } be a subsequence of {ξn}. Suppose that ξmk ≤ ξmk+1 for all k ∈ N.
Then it is easy to check that mk is odd and mk + 1 is even for every k ∈ N. We
now assume that ξk ≤ ξmk+1 for k ∈ N. Then it follows that
1
k
= ξk ≤ ξmk+1 =
1
mk + 1
if k is even. This implies that k ≥ mk + 1 ≥ k + 1, which is a contradiction.
(cid:3)
Using Lemma 3.2, we obtain the following:
Lemma 3.4. Let {ξn} be a sequence of nonnegative real numbers which is not
convergent. Then there exist N ∈ N and an eventually increasing function τ : N →
N such that ξτ (n) ≤ ξτ (n)+1 for all n ∈ N and ξn ≤ ξτ (n)+1 for all n ≥ N .
Proof. Since {ξn} is not convergent, for any n ∈ N there exists m ∈ N such that
m ≥ n and ξm < ξm+1, and hence there exists a subsequence {ξni} of {ξn} such
that ξni < ξni+1 for every i ∈ N. Lemma 3.2 implies that there exist N ∈ N and a
function σ : N → N such that σ(n) ≤ σ(n + 1), ξσ(n) ≤ ξσ(n)+1, and ξn ≤ ξσ(n)+1
for every n ≥ N and limn→∞ σ(n) = ∞. Let us define τ : N → N by τ (n) = σ(N )
for n ∈ {1, 2, . . . , N } and τ (n) = σ(n) for n > N , which completes the proof.
(cid:3)
In the rest of this section, unless otherwise stated, we assume that E is a smooth,
strictly convex, and reflexive Banach space and C is a nonempty closed convex
subset of E.
n=1 F (Tn) is
Let {Tn} be a sequence of mappings of C into E such that F = T∞
nonempty. Then
• {Tn} is said to be a strongly relatively nonexpansive sequence [3] if each Tn
is of type (r) and φ(Tnxn, xn) → 0 whenever {xn} is a bounded sequence
in E and φ(p, xn) − φ(p, Tnxn) → 0 for some point p ∈ F ;
5
• {Tn} satisfies the condition (Z) if every weak cluster point of {xn} belongs
to F whenever {xn} is a bounded sequence in C such that Tnxn − xn → 0.
Let A ⊂ E × E ∗ be a maximal monotone operator with a zero point and {rn}
a sequence of positive real numbers. Then (2.8) shows that the sequence {Lrn}
of resolvents of A is a strongly relatively nonexpansive sequence; see [3] for more
details.
In order to prove our main result in §4, we need the following lemmas:
Lemma 3.5. Let {Tn} be a sequence of mappings of C into E such that F =
n=1 F (Tn) is nonempty, τ : N → N an eventually increasing function, and {zn} a
bounded sequence in C such that φ(p, zn) − φ(p, Tτ (n)zn) → 0 for some p ∈ F . If
{Tn} is a strongly relatively nonexpansive sequence, then φ(Tτ (n)zn, zn) → 0.
T∞
Proof. Suppose that φ(Tτ (n)zn, zn) 9 0. Then there exist ǫ > 0 and a strictly
increasing function σ : N → N such that τ ◦ σ is also strictly increasing and
(3.1)
φ(Tτ ◦σ(n)zσ(n), zσ(n)) ≥ ǫ
for all n ∈ N. Set µ = τ ◦ σ and R(µ) = {µ(n) : n ∈ N}. Define a sequence {yn} in
C as follows: For each n ∈ N,
yn =(zσ◦µ−1(n)
p
if n ∈ R(µ);
if n /∈ R(µ).
It is clear that {yn} is bounded,
φ(p, yn) − φ(p, Tnyn) = φ(p, zσ◦µ−1(n)) − φ(p, Tτ (σ◦µ−1(n))zσ◦µ−1(n))
for n ∈ R(µ), and φ(p, yn) − φ(p, Tnyn) = 0 for n /∈ R(µ). Since σ ◦ µ−1 is
strictly increasing, it follows that φ(p, yn) − φ(p, Tnyn) → 0, so φ(Tnyn, yn) → 0
because {Tn} is a strongly relatively nonexpansive sequence. Therefore, noting that
yµ(n) = zσ(µ−1(µ(n))) = zσ(n) and µ is strictly increasing, we have
φ(Tτ ◦σ(n)zσ(n), zσ(n)) = φ(Tµ(n)yµ(n), yµ(n)) → 0,
which contradicts to (3.1).
(cid:3)
Lemma 3.6. Let {Tn} be a sequence of mappings of C into E such that F =
n=1 F (Tn) is nonempty, τ : N → N an eventually increasing function, and {zn} a
bounded sequence in C such that Tτ (n)zn − zn → 0. Suppose that {Tn} satisfies the
condition (Z). Then every weak cluster point of {zn} belongs to F .
T∞
Proof. Let z be a weak cluster point of {zn}. Then there exists a strictly increasing
function σ : N → N such that zσ(n) ⇀ z as n → ∞ and τ ◦ σ is strictly increasing.
Set µ = τ ◦ σ and R(µ) = {µ(n) : n ∈ N}. Define a sequence {yn} in C as follows:
For each n ∈ N,
yn =(zσ◦µ−1(n)
p
if n ∈ R(µ);
if n /∈ R(µ),
where p is a point in F . Then it is clear that {yn} is bounded,
yn − Tnyn = zσ◦µ−1(n) − Tτ (σ◦µ−1(n))zσ◦µ−1(n)
for n ∈ R(µ), and yn − Tnyn = 0 for n /∈ R(µ). Since zn − Tτ (n)zn → 0 and σ ◦ µ−1
is strictly increasing, it follows that yn − Tnyn → 0. Noting that µ is strictly
increasing and yµ(n) = zσ◦µ−1(µ(n)) = zσ(n) for every n ∈ N, we know that {zσ(n)}
6
KOJI AOYAMA, YASUNORI KIMURA, AND FUMIAKI KOHSAKA
is a subsequence of {yn}, and hence z is a weak cluster point of {yn}. Since {Tn}
satisfies the condition (Z), we conclude that z ∈ F .
(cid:3)
Lemma 3.7. Let {Tn} be a sequence of mappings of C into E, F be a nonempty
closed convex subset of E, {zn} a bounded sequence in C such that zn − Tnzn → 0,
and u ∈ E. Suppose that every weak cluster point of {zn} belongs to F . Then
lim sup
n→∞
hTnzn − w, Ju − Jwi ≤ 0,
where w = QF (u).
Proof. Since zn − Tnzn → 0 and {zn} is bounded, there exists a weakly convergent
subsequence {zni} of {zn} such that
lim sup
n→∞
hTnzn − w, Ju − Jwi = lim sup
n→∞
hzn − w, Ju − Jwi
= lim
i→∞
hzni − w, Ju − Jwi .
Let z be the weak limit of {zni}. By assumption, we see that z ∈ F . Thus (2.6)
shows that
lim
i→∞
hzni − w, Ju − Jwi = hz − w, Ju − Jwi ≤ 0,
which is the desired result.
(cid:3)
4. Strong convergence theorems for strongly relatively
nonexpansive sequences
In this section, we prove the following strong convergence theorem:
Theorem 4.1. Let E be a smooth and uniformly convex Banach space, C a
nonempty closed convex subset of E, {Sn} a sequence of mappings of C into E
n=1 F (Sn) is nonempty, and {αn} a sequence in [0, 1] such that
αn → 0. Let u be a point in E and {xn} a sequence defined by x1 ∈ C and
such that F = T∞
(4.1)
for n ∈ N. Suppose that
xn+1 = QC J −1(cid:0)αnJu + (1 − αn)JSnxn(cid:1)
• {Sn} is a strongly relatively nonexpansive sequence;
• {Sn} satisfies the condition (Z);
• αn > 0 for every n ∈ N and P∞
Then {xn} converges strongly to w = QF (u).
n=1 αn = ∞.
First, we show some lemmas; then we prove Theorem 4.1. In the rest of this
section, we set
for n ∈ N, so (4.1) is reduced to xn+1 = QC(yn).
yn = J −1(cid:0)αnJu + (1 − αn)JSnxn(cid:1)
Lemma 4.2. Both {xn} and {Snxn} are bounded, and moreover, the following
hold:
(1) yn − Snxn → 0;
(2) φ(w, xn+1) ≤ αnφ(w, u) + φ(w, Snxn) for every n ∈ N;
(3) φ(w, xn+1) ≤ (1 − αn)φ(w, xn) + 2αn hyn − w, Ju − Jwi for every n ∈ N.
7
Proof. Since QC and Sn are of type (r) and w ∈ F (Sn) ⊂ C, it follows from (2.3)
that
φ(w, xn+1) ≤ φ(w, yn)
(4.2)
≤ αnφ(w, u) + (1 − αn)φ(w, Snxn)
≤ αnφ(w, u) + (1 − αn)φ(w, xn)
for every n ∈ N. Thus, by induction on n, we have
φ(w, Snxn) ≤ φ(w, xn) ≤ max{φ(w, x1), φ(w, u)}.
Therefore, by virtue of (2.2), it turns out that {xn} and {Snxn} are bounded.
By αn → 0, it is clear that Jyn − JSnxn = αn(Ju − JSnxn) → 0. This shows
that
yn − Snxn = J −1Jyn − J −1JSnxn → 0
because E ∗ is uniformly smooth and J −1 is uniformly continuous on every bounded
set. Thus (1) holds.
(2) follows from (4.2).
Since Sn is of type (r), it follows from (4.2), (2.4), and (2.3) that
φ(w, xn+1) ≤ φ(w, yn)
(4.3)
≤ φ(cid:16)w, J −1(cid:0)αnJu + (1 − αn)JSnxn − αn(Ju − Jw)(cid:1)(cid:17)
+ 2 hyn − w, αn(Ju − Jw)i
≤ (1 − αn)φ(w, Snxn) + αnφ(w, w) + 2αn hyn − w, Ju − Jwi
≤ (1 − αn)φ(w, xn) + 2αn hyn − w, Ju − Jwi
for every n ∈ N. Therefore, (3) holds.
Lemma 4.3. Suppose that
(cid:3)
(4.4)
Then {xn} converges strongly to w.
lim sup
n→∞ (cid:0)φ(w, xn) − φ(w, xn+1)(cid:1) ≤ 0.
Proof. We first show that Snxn −xn → 0. Since Sn is of type (r), it follows from (2)
in Lemma 4.2 that
0 ≤ φ(w, xn) − φ(w, Snxn) ≤ φ(w, xn) − φ(w, xn+1) + αnφ(w, u)
for every n ∈ N, so φ(w, xn) − φ(w, Snxn) → 0 by (4.4) and αn → 0. Since {Sn}
is a strongly relatively nonexpansive sequence and {xn} is bounded by Lemma 4.2,
φ(Snxn, xn) → 0. Using Lemma 2.1, we conclude that Snxn − xn → 0.
We know that yn − Snxn → 0 by (1) in Lemma 4.2 and {Sn} satisfies the
condition (Z) by assumption, so Lemma 3.7 implies that
lim sup
n→∞
hyn − w, Ju − Jwi = lim sup
n→∞
hSnxn − w, Ju − Jwi ≤ 0.
It follows from (3) in Lemma 4.2 that
φ(w, xn+1) ≤ (1 − αn)φ(w, xn) + 2αn hyn − w, Ju − Jwi
for every n ∈ N. Therefore, noting that P∞
conclude that φ(w, xn) → 0, and hence xn → w by Lemma 2.1.
n=1 αn = ∞ and using Lemma 2.2, we
(cid:3)
Lemma 4.4. The real number sequence {φ(w, xn)} is convergent.
8
KOJI AOYAMA, YASUNORI KIMURA, AND FUMIAKI KOHSAKA
Proof. We assume, to obtain a contraction, that {φ(w, xn)} is not convergent. Then
Lemma 3.4 implies that there exist N ∈ N and an eventually increasing function
τ : N → N such that
(4.5)
φ(w, xτ (n)) ≤ φ(w, xτ (n)+1)
for every n ∈ N and
(4.6)
φ(w, xn) ≤ φ(w, xτ (n)+1)
for every n ≥ N .
We show that Sτ (n)xτ (n) − xτ (n) → 0. Since Sτ (n) is of type (r), it follows
from (4.5), (2) in Lemma 4.2, and Lemma 3.1 that
0 ≤ φ(w, xτ (n)) − φ(w, Sτ (n)xτ (n))
≤ φ(w, xτ (n)+1) − φ(w, Sτ (n)xτ (n))
≤ ατ (n)φ(w, u) → 0
as n → ∞. Since {xτ (n)} is bounded and {Sn} is a strongly relatively nonexpansive
sequence, it follows from Lemma 3.5 that φ(Sτ (n)xτ (n), xτ (n)) → 0, so we conclude
that Sτ (n)xτ (n) − xτ (n) → 0 by Lemma 2.1.
Finally, we obtain a contradiction that φ(w, xn) → 0. From (4.5) and (3) in
Lemma 4.2, we know that
φ(w, xτ (n)) ≤ φ(w, xτ (n)+1)
(4.7)
≤ (1 − ατ (n))φ(w, xτ (n)) + 2ατ (n)(cid:10)yτ (n) − w, Ju − Jw(cid:11)
for every n ∈ N, where yτ (n) = J −1(cid:0)ατ (n)Ju + (1 − ατ (n))JSτ (n)xτ (n)(cid:1) for n ∈ N.
Noting that ατ (n) > 0, (4.7) is reduced to
so that
(4.8)
φ(w, xτ (n)) ≤ 2(cid:10)yτ (n) − w, Ju − Jw(cid:11) ,
φ(w, xτ (n)+1) ≤ 2(cid:10)yτ (n) − w, Ju − Jw(cid:11)
for every n ∈ N. Since {Sn} satisfies the condition (Z), it follows from Lemma 3.6
that every weak cluster point of {xτ (n)} belongs to F . Using (4.8), (1) in Lemma 4.2,
and Lemma 3.7, we have
lim sup
n→∞
φ(w, xτ (n)+1) ≤ 2 lim sup
= 2 lim sup
n→∞ (cid:10)yτ (n) − w, Ju − Jw(cid:11)
n→∞ (cid:10)Sτ (n)xτ (n) − w, Ju − Jw(cid:11) ≤ 0.
Therefore, by virtue of (4.6), we conclude that
lim sup
n→∞
φ(w, xn) ≤ lim sup
n→∞
φ(w, xτ (n)+1) ≤ 0,
and hence φ(w, xn) → 0, which is a contradiction.
Proof of Theorem 4.1. Using Lemmas 4.3 and 4.4, we get the conclusion.
(cid:3)
(cid:3)
9
5. Applications
In this section, we study the zero point problem for a maximal monotone operator
and the fixed point problem for a relatively nonexpansive mapping. We employ
Theorem 4.1 to approximate solutions of these problems.
To prove the first theorem, we need the following lemma:
Lemma 5.1 ([4, Lemma 3.5]). Let E be a strictly convex and reflexive Banach
space whose norm is uniformly Gateaux differentiable, {rn} a sequence of positive
real numbers, and Lrn the resolvent of a maximal monotone operator A ⊂ E ×
E ∗. Suppose that inf n rn > 0 and A−10 is nonempty. Then {Lrn} satisfies the
condition (Z).
We adopt a modified proximal point algorithm introduced by Kohsaka and Taka-
hashi [9] in the following theorem:
Theorem 5.2. Let E be a uniformly convex Banach space whose norm is uni-
formly Gateaux differentiable, A ⊂ E × E ∗ a maximal monotone operator, {αn} a
sequence in (0, 1], and {rn} a sequence of positive real numbers. Suppose that A−10
n=1 αn = ∞, and inf n rn > 0. Let u be a point in E and
is nonempty, αn → 0, P∞
{xn} a sequence defined by x1 ∈ C and
xn+1 = J −1(cid:0)αnJu + (1 − αn)JLrn xn(cid:1)
(5.1)
for n ∈ N, where Lrn = (J + rnA)−1J. Then {xn} converges strongly to QA−10(u).
Proof. Set Sn = Lrn for n ∈ N. It is known that F (Sn) = A−10 and Lrn is a type (r)
n=1 F (Sn) = A−10 is nonempty. It is
also known that {Sn} is a strongly relatively nonexpansive sequence by [3, Example
3.2] and {Sn} satisfies the condition (Z) by Lemma 5.1. It is clear that QE is the
identity mapping on E. Therefore, Theorem 4.1 implies the conclusion.
(cid:3)
self-mapping of E for each n ∈ N. Hence T∞
Remark 5.3. Theorem 5.2 is similar to [9, Theorem 3.3]. In [9, Theorem 3.3], E
is assumed to be smooth and uniformly convex and {αn} in [0, 1] while {rn} is
assumed to diverge to infinity.
To prove the next theorem, we need the following lemma:
Lemma 5.4 ([3, Lemma 2.1]). Let {xn} and {yn} be two bounded sequences
in a uniformly convex Banach space E and {λn} a sequence in [0, 1] such that
lim inf n→∞ λn > 0. Suppose that
λn kxnk2 + (1 − λn) kynk2 − kλnxn + (1 − λn)ynk2 → 0.
Then (1 − λn)(xn − yn) → 0.
The following is a strong convergence theorem for a relatively nonexpansive
mapping; see [11, 12] for other convergence theorems and see also [3].
Theorem 5.5 ([13, Theorem 3.4]). Let E be a uniformly convex and uniformly
smooth Banach space, C a nonempty closed convex subset of E, T : C → E a
relatively nonexpansive mapping, {αn} a sequence in (0, 1], and {βn} a sequence
n=1 αn = ∞, and 0 < lim inf n→∞ βn ≤
lim supn→∞ βn < 1. Let u be a point in E and {xn} a sequence defined by x1 ∈ C
and
in [0, 1]. Suppose that αn → 0, P∞
(5.2)
xn+1 = QCJ −1(cid:16)αnJu + (1 − αn)(cid:0)βnJxn + (1 − βn)JT xn(cid:1)(cid:17)
10
KOJI AOYAMA, YASUNORI KIMURA, AND FUMIAKI KOHSAKA
for n ∈ N. Then {xn} converges strongly to QF (T )(u).
Proof. Set Sn = J −1(cid:0)βnJ + (1 − βn)JT(cid:1) for n ∈ N. Then it is easy to check that
each Sn is a mapping of type (r) and T∞
n=1 F (Sn) = F (T ); see [5, Corollary 3.8].
Moreover, it is clear that (5.2) coincides with (4.1). To finish the proof, it is enough
to show that {Sn} is a strongly relatively nonexpansive sequence and {Sn} satisfies
the condition (Z).
Let {yn} be a bounded sequence in C such that φ(p, yn) − φ(p, Snyn) → 0 for
n=1 F (Sn). Since T is of type (r), we have
some p ∈T∞
βn kJynk2 + (1 − βn) kJT ynk2 − kJSnynk2
= βnφ(p, yn) + (1 − βn)φ(p, T yn) − φ(p, Snyn)
≤ βnφ(p, yn) + (1 − βn)φ(p, yn) − φ(p, Snyn)
= φ(p, yn) − φ(p, Snyn) → 0.
Using Lemma 5.4 and (2.5), it turns out that
Jyn − JSnyn = (1 − βn)(Jyn − JT yn) → 0,
and hence φ(Snyn, yn) → 0. Thus {Sn} is a strongly relatively nonexpansive se-
quence.
Let {zn} be a bounded sequence in C such that zn − Snzn → 0. Then it follows
from (2.5) that
(1 − βn)(Jzn − JT zn) = Jzn − JSnzn → 0,
so we conclude that zn − T zn → 0 by lim supn→∞ βn < 1 and (2.5). Since T is
relatively nonexpansive, every weak cluster point of {zn} belongs to F (T ). This
means that {Sn} satisfies the condition (Z). Consequently, Theorem 4.1 implies the
conclusion.
(cid:3)
Remark 5.6. In [13, Theorem 3.4], {αn} and {βn} are assumed to be sequences in
(0, 1).
References
[1] Y. I. Alber, Metric and generalized projection operators in Banach spaces: properties and
applications, Theory and applications of nonlinear operators of accretive and monotone type,
Lecture Notes in Pure and Appl. Math., vol. 178, Dekker, New York, 1996, pp. 15–50.
[2] K. Aoyama, Y. Kimura, W. Takahashi, and M. Toyoda, Approximation of common fixed
points of a countable family of nonexpansive mappings in a Banach space, Nonlinear Anal.
67 (2007), 2350–2360.
[3] K. Aoyama, F. Kohsaka, and W. Takahashi, Strongly relatively nonexpansive sequences in
Banach spaces and applications, J. Fixed Point Theory Appl. 5 (2009), 201–224.
[4]
[5]
, Proximal point methods for monotone operators in Banach spaces, Taiwanese Jour-
nal of Mathematics 15 (2011), 259–281.
, Strong convergence theorems by shrinking and hybrid projection methods for rel-
atively nonexpansive mappings in Banach spaces, Nonlinear analysis and convex analysis,
Yokohama Publ., Yokohama, 2009, pp. 7–26.
[6] Y. Censor and S. Reich, Iterations of paracontractions and firmly nonexpansive operators
with applications to feasibility and optimization, Optimization 37 (1996), 323–339.
[7] S. Kamimura, F. Kohsaka, and W. Takahashi, Weak and strong convergence theorems for
maximal monotone operators in a Banach space, Set-Valued Anal. 12 (2004), 417–429.
[8] S. Kamimura and W. Takahashi, Strong convergence of a proximal-type algorithm in a Ba-
nach space, SIAM J. Optim. 13 (2002), 938–945 (electronic) (2003).
11
[9] F. Kohsaka and W. Takahashi, Strong convergence of an iterative sequence for maximal
monotone operators in a Banach space, Abstr. Appl. Anal. (2004), 239–249.
[10] P.-E. Maing´e, Strong convergence of projected subgradient methods for nonsmooth and non-
strictly convex minimization, Set-Valued Anal. 16 (2008), 899–912.
[11] S. Matsushita and W. Takahashi, Weak and strong convergence theorems for relatively non-
expansive mappings in Banach spaces, Fixed Point Theory Appl. (2004), 37–47.
[12]
, A strong convergence theorem for relatively nonexpansive mappings in a Banach
space, J. Approx. Theory 134 (2005), 257–266.
[13] W. Nilsrakoo and S. Saejung, Strong convergence theorems by Halpern-Mann iterations for
relatively nonexpansive mappings in Banach spaces, Applied Mathematics and Computation
217 (2011), 6577–6586.
[14] S. Reich, A weak convergence theorem for the alternating method with Bregman distances,
Theory and applications of nonlinear operators of accretive and monotone type, Lecture Notes
in Pure and Appl. Math., vol. 178, Dekker, New York, 1996, pp. 313–318.
[15] R. T. Rockafellar, On the maximality of sums of nonlinear monotone operators, Trans. Amer.
Math. Soc. 149 (1970), 75–88.
[16] S. Saejung, Halpern's iteration in Banach spaces, Nonlinear Anal. 73 (2010), 3431–3439.
[17] W. Takahashi, Nonlinear functional analysis, Yokohama Publishers, Yokohama, 2000.
[18] H.-K. Xu, Iterative algorithms for nonlinear operators, J. London Math. Soc. (2) 66 (2002),
240–256.
(Koji Aoyama) Department of Economics, Chiba University, Yayoi-cho, Inage-ku, Chi-
ba-shi, Chiba 263-8522, Japan
E-mail address: [email protected]
(Yasunori Kimura) Department of Information Science, Toho University, Miyama,
Funabashi-shi, Chiba 274-8510, Japan
E-mail address: [email protected]
(Fumiaki Kohsaka) Department of Computer Science and Intelligent Systems, Oita
University, Dannoharu, Oita-shi, Oita 870-1192, Japan
E-mail address: [email protected]
|
1812.04248 | 2 | 1812 | 2019-01-11T10:24:07 | A profile decomposition for the limiting Sobolev embedding | [
"math.FA"
] | For many known non-compact embeddings of two Banach spaces $E\hookrightarrow F$, every bounded sequence in $E$ has a subsequence that takes form of a profile decomposition - a sum of clearly structured terms with asymptotically disjoint supports plus a remainder that vanishes in the norm of $F$. In this note we construct a profile decomposition for arbitrary sequences in the Sobolev space $H^{1,2}(M)$ of a compact Riemannian manifold, relative to the embedding of $H^{1,2}(M)$ into $L^{2^*}(M)$, generalizing the well-known profile decomposition of Struwe ([Proposition 2.1]{Struwe}) to the case of arbitrary bounded sequences. | math.FA | math |
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV
EMBEDDING
GIUSEPPE DEVILLANOVA AND CYRIL TINTAREV
Abstract. For many known non-compact embeddings of two Banach spaces E ֒→ F ,
every bounded sequence in E has a subsequence that takes form of a profile decom-
position - a sum of clearly structured terms with asymptotically disjoint supports
plus a remainder that vanishes in the norm of F . In this note we construct a profile
2(M ) of a compact
decomposition for arbitrary sequences in the Sobolev space H 1
Riemannian manifold, relative to the embedding of H 1
(M ), generaliz-
ing the well-known profile decomposition of Struwe [12, Proposition 2.1] to the case
of arbitrary bounded sequences.
2(M ) into L2∗
,
,
1. Introduction
When the embedding of two Banach spaces E ֒→ F is continuous and not compact,
the lack of compactness can be manifested by the (behavior in F of the) difference
uk − u between the elements of a weakly convergent sequence (uk)k∈N ⊂ E and its
weak limit u. Therefore one may call defect of compactness of (uk)k∈N the (sequences
of) differences uk − u taken up to a suitable remainder that vanishes in the norm of F .
(Note that, if the embedding is compact and E is reflexive, the defect of compactness
is itself infinitesimal and so it can be identified with zero). For many embeddings there
exist well-structured representations of the defect of compactness, known as profile de-
compositions. Best studied are profile decompositions relative to Sobolev embeddings,
which are sums of terms with asymptotically disjoint supports, called elementary con-
centrations or bubbles. Profile decompositions were originally motivated by studies of
concentration phenomena in PDE in the early 1980's by Uhlenbeck, Brezis, Coron,
Nirenberg, Aubin and Lions, and they play a significant role in the verification process
of the convergence of sequences of functions in applied analysis, particularly when the
information available via the classical concentration-compactness method is not enough
detailed.
Profile decompositions are known to exist when the embedding E ֒→ F is cocompact
relative to some group G of isometries on E, see [11]. We recall that an embedding
E ֒→ F is called cocompact relative to a group G of isometries (G-cocompact for short)
if any sequence (uk)k∈N ⊂ E such that gk(uk) ⇀ 0 for any sequence of operators
2010 Mathematics Subject Classification. Primary 46E35, 46B50, Secondary 58J99, 35B44, 35A25.
Key words and phrases. concentration compactness, profile decompositions, multiscale analysis.
1
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
2
(gk)k∈N ⊂ G turns out to be infinitesimal in the norm of F . (An elementary example
due to Jaffard [7], which is easy to verify, is cocompactness of embedding of ℓ∞(Z)
into itself relative to the group of shifts G := {gm := (an)n∈N 7→ (an+m)n∈N m ∈ Z}.)
Up to the authors knowledge the first cocompactness result for functional spaces is
[8, Lemma 6] by E. Lieb which expresses (using different terminology than the present
note) that the nonhomogeneous Sobolev space H 1,p(RN ) is cocompactly embedded into
Lq(RN ), when N > p and q ∈ (p, p∗) (where p∗ = N p
N −p), relative to the group of shifts
u 7→ u(· − y), y ∈ RN . A profile decomposition relative to a group G of bijective
isometries on a Banach space E represents defect of compactness uk − u as a sum of
k ∈ G
and w(n) ∈ E. Note that in the above sum the index n = 0 is not allowed since, in
the existing literature, usually w(0) represents the weak-limit u of the sequence and
(g(0)
k )k∈N is the constant sequence of constant value the identity map of the space. So,
by using this convention, we can use defect of compactness to represent the sequence
k w(n) and a remainder vanishing in F . In the above sums
each of the elements w(n) (for n ≥ 1), called concentration profiles, is obtained as the
weak-limit (as k → ∞) of the "deflated" sequence ((g(n)
elementary concentrations, or bubbles, namely Pn∈N\{0} g(n)
(uk)k∈N as a sum of Pn∈N g(n)
k w(n) with some g(n)
k )−1(uk))k∈N .
p
Typical examples of isometries groups G, involved in profile decompositions, are the
above mentioned group of shifts u 7→ u(·−y) and the rescaling group, which is a product
group of shifts and dilations u 7→ tru(t·), t > 0, where, for instance, when u belongs to
the homogeneous Sobolev space H s,p(RN ) (N/s > p ≥ 1, s > 0), r = r(p, s) = N −ps
.
Existence of profile decompositions for general bounded sequences in H 1,p(RN ) (rel-
ative to the rescaling group) was proved by Solimini, see [10, Theorem 2], and inde-
pendently, but with a weaker form of remainder, by G´erard in [6], with an extension
to the case of fractional Sobolev spaces by Jaffard in [7]. Only in [9], for the first time,
the authors observed that profile decomposition (and thus concentration phenomena in
general) can be understood in functional-analytic terms, rather than in specific func-
tion spaces. Actually the results in [9] where extended in [11] to uniformly convex
Banach spaces with the Opial condition (without the Opial condition profile decom-
position still exists but weak convergence must be replaced by (a less-known) Delta
convergence, see [4]). Finally the result has been extended up to a suitable class of
metric spaces, see [5] and [3]. Despite the character of the statement in [11] is rather
general, profile decompositions are still true, for instance, when the space E is not
reflexive (e.g. [2]), or when one only has a semigroup of isometries (e.g. [1]), or when
the profile decomposition can be expressed without the explicit use of a group (e.g.
Struwe [12]) and so when [11, Theorem 2.10] does not apply.
The present paper generalizes, in the spirit of [10, Theorem 2], Struwe's result [12,
Proposition 2.1] (which provides a profile decomposition for Palais-Smale sequences of
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
3
particular functionals) to the case of general bounded sequences in H 1,2(M), where M
is a smooth compact manifold in dimension N ≥ 3.
The paper is organized as follows. In Section 2 we introduce some notation and state
the main theorem of the paper and the result on which the related proof is based. In
Section 3 we prove that the embedding H 1,2(M) ֒→ L2∗(M) is cocompact with respect
to a group of suitable transformations which are depending on the Atlas associated to
the manifold. Section 4 is devoted to the proof of (the main) Theorem 2.1.
2. Statement of the main result
Let N ≥ 3 and let (M, g) be a compact smooth Riemannian N-dimensional mani-
fold. We consider the Sobolev space H 1,2(M) equipped with the norm defined by the
quadratic form of the Laplace-Beltrami operator,
kuk2 = M
(du2 + u2)dvg,
(2.1)
(vg denotes the Riemannian measure of the manifold). For every y ∈ M we shall
denote by Ty(M) the tangent space in y to M, and by expy the exponential (local)
map at the point y (defined on a suitable set Uy ⊂ Ty(M) by setting, for all v ∈ Uy,
expy(v) := γv(1) where γv is the unique geodesic, contained in M, such that γv(0) = y
and γ′
v(0) = v and extended to the case v = 0 by setting expy(0) = y). Since we
will not use here any property of tangent bundles we will identify tangent spaces of
M at different points with RN and, for any ρ > 0, we shall denote by Bρ(0) the
Euclidean N-dimensional ball centered at the origin with radius ρ. On the other hand,
we shall denote by Bρ(y) the open coordinate ball (i.e. the subset in M such that
exp−1
y (Bρ(y)) = Bρ(0)) with center y and radius ρ > 0. For the reader's convenience
we recall that the injectivity radius ρy of a point y ∈ M is the radius of the largest ball
about the origin in Ty(M) that can be mapped diffeomorfically via the map expy, and
that, the injectivity radius of the mainfold M, ρM := inf y∈M ρy. Since M is compact,
ρM is strictly positive, so we can fix 0 < ρ < ρM
3 , moreover, there exists a finite set of
points (zi)i∈I ⊂ M such that (Bρ(zi), exp−1
zi )i∈I is a finite smooth atlas of M.
In what follows we shall fix χ ∈ C ∞
0 (Bρ(0)), χ = 1 on B ρ
2
(0), so that, set for i ∈ I
χi := χzi = χ ◦ exp−1
zi
and
χi :=
,
(2.2)
(χi)i∈I is a smooth partition of unity on M subordinated to the covering (Bρ(zi))i∈I.
Then, since ku ◦ expzi kL2∗ (Bρ(0)) is bounded by the H 1,2(Bρ(0))-norm of u ◦ expzi,
the Sobolev embedding H 1,2(M) ֒→ L2∗(M) can be deduced from the corresponding
one on the Euclidean space (by the use of the fixed partition of unity (χi)i∈I).
In
fact, Theorem 2.1 below will provide a profile decomposition for bounded sequences in
H 1,2(M).
χi
Pj∈I χj
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
4
Finally we recall that the scalar product associated with (2.1) can be written with
help of the partition of unity (χs)s∈I in the following coordinate form:
N
hΦ, Ψi := Xs∈I
Xs∈I
Bρ(0)
Bρ(0)
gzs
i,j∂i((χsΦ)(expzs(ξ)))∂j(Ψ(expzs(ξ)))qdet(gzs
Xi,j=1
(χsΦ)(expzs(ξ))Ψ(expzs(ξ))qdet(gzs
i,j)dξ.
i,j)dξ+
(2.3)
Before stating the theorem, we warn the reader that, given a bounded sequence
(vk)k∈N ⊂ H 1,2(Bρ(0)) and a vanishing sequence of positive numbers (tk)k∈N, and setting
r = r(2) = N
2 , we will say (with a slight abuse on the definition of weak
kvk(tk·))k∈N weakly converges to v ∈ H 1,2(RN ) if for
convergence) that the sequence (tr
any ϕ ∈ C ∞
0 (RN ) such that supp ϕ ⊂ Bρ(0)
2∗ = N −2
ϕ(x) tr
kvk(tkx) dx −→ ϕ(x) v(x) dx as k → ∞.
Theorem 2.1. Let M be a compact smooth Riemannian N-dimensional manifold (N ≥
3). Let ρ ∈ (0, ρM
(0), and let (χi)i∈I, defined by
(2.2), be a smooth partition of unity on M subordinated to the covering (Bρ(zi))i∈I.
Then, given a bounded sequence (uk)k∈N in H 1,2(M) and, with r = N
2 , there
exist:
0 (Bρ(0)), χ = 1 on B ρ
3 ), let χ ∈ C ∞
2∗ = N −2
2
• a sequence (Y (n))n∈N\{0} of sequences Y (n) := (y(n)
• a sequence (J (n))n∈N\{0} of sequences J (n) := (j(n)
• a sequence (w(n))n∈N\{0} of functions (profiles) w(n) ∈ H 1,2(RN ),
k )k∈N ⊂ M, y(n)
k )k∈N ⊂ R+,
k → ¯y(n) ∈ M,
such that, modulo subsequences,
j(n)
k −→ ∞ as k → ∞ ∀n ∈ N \ {0},
k − j(m)
j(n)
k
+ 2j
(n)
k d(y(n)
k , y(m)
k
) → ∞ whenever m 6= n,
(n)
k ruk ◦ expy
Moreover, setting for all k ∈ N
2−j
(n)
k
(2−j
(n)
k ·)⇀w(n)
in H 1,2(RN ) as k → ∞.
(2.4)
(2.5)
(2.6)
Sk(x) := Xn∈N\{0}
(n)
k r
2j
χ ◦ exp−1
y
(n)
k
(x) w(n)(2j
(n)
k exp−1
y
(n)
k
(x)), x ∈ M,
(2.7)
the series Sk ∈ H 1,2(M) are unconditionally convergent (with respect to n) and the
sequence (Sk)k∈N is uniformly convergent (with respect to k) in H 1,2(M), in addition
uk − u − Sk → 0 in L2∗
(M) .
(2.8)
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
5
Finally the following energy bound holds
Xn∈N\{0}
k∇w(n)k2
L2(RN ) + kuk2
H 1,2(M ) ≤ lim inf
k→∞
kukk2
H 1,2(M ).
(2.9)
We want to emphasize that (2.8) states that, modulo subsequence, the defect of
compactness uk − u of the bounded sequence (uk)k∈N (which, modulo subsequence,
weakly converges to u) has a representation given (up to a remainder which vanishes
in the norm of L2∗(M)) by the clearly structured terms in Sk.
The proof of this theorem is based on the following easy corollary to Solimini's profile
decomposition [10, Theorem 2].
Theorem 2.2. Given m ∈ N \ {0} and 1 < p < N
. Let (vk)k∈N
be a bounded sequence in the homogeneous Sobolev space H m,p(RN ) supported on a
compact set K ⊂ RN . Then, there exists a (renamed) subsequence (s.t. vk ⇀ v) whose
defect of compactness vk − v has the form
p∗(m) = N −mp
m let r = N
p
Sk = Xn∈N\{0}
(n)
k rw(n)(2j
2j
(n)
k (· − ξ(n)
k )),
where, for any n ∈ N \ {0}, Ξ(n) := (ξ(n)
that j(n)
+ξ(n)
k → + ∞ as k → ∞ and w(n) is the weak limit of the sequence (2−j
k ))k∈N. Moreover the addenda are asymptotically mutually orthogonal, i.e.
k )k∈N ⊂ K, and J (n) := (j(n)
k )k∈N ⊂ R are such
·
k rvk(2−j
(n)
k
(n)
j(n)
k − j(m)
k
+ 2j
(n)
k ξ(n)
k − ξ(m)
k
→ ∞ whenever m 6= n.
(2.10)
Proof We shall assume, without restrictions, that uk ⇀ 0. According to the profile
decomposition result [10, Theorem 2], modulo the extraction of a subsequence, each
term vk has concentration terms (depending on n) of the following shape
cn
k := 2j
(n)
k rw(n)(2j
(n)
k (· − ξ(n)
k ))
(n)
k ∈ RN , j(n)
· +ξ(n)
for some ξ(n)
(2−j
k rvk(2−j
Indeed, on the contrary, the assumption j(n)
has a bounded support, that
k ∈ R where w(n) is obtained as the weak limit of the sequence
k ))k∈N. We claim that the sequence J (n) is bounded from below.
k → − ∞ as k → ∞ would imply, since vk
(n)
k
k2−j
(n)
k rvk(2−j
(n)
k
· +ξ(n)
k )kp→0 as k → ∞,
and so that w(n) = 0.
As a consequence ξ(n)
k ∈ K for k large enough. Note also that J (n) cannot have any
bounded subsequence, since otherwise (vk)k∈N should have a nonzero weak limit, in
contradiction to our assumptions.
Finally, condition (2.10) is the condition of asymptotic orthogonality (decoupling)
of bubbles from [10].
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
6
3. Cocompactness in Sobolev spaces of compact manifolds
The Sobolev embedding H 1,2(M) ֒→ L2∗(M) has the following property of cocom-
pactness type.
3 . Let (Bρ(zi), exp−1
Theorem 3.1. Let M be a compact smooth Riemannian N-dimensional manifold (N ≥
3), and 0 < ρ < ρM
zi )i∈I be a finite smooth atlas of M and let
χ ∈ C ∞
0 (Bρ(0)) so that (χi)i∈I, defined by (2.2), is a smooth partition of unity on
M subordinated to the covering (Bρ(zi))i∈I. Set r = r(2) = N
2 . If (uk)k∈N is
any bounded sequence in H 1,2(M) such that for every i ∈ I, (yk)k∈N ⊂ Bρ(zi), and
(jk)k∈N ⊂ N such that jk → +∞
2∗ = N −2
2−jkr(χiuk) ◦ expyk(2−jk·)⇀0 as k → ∞,
(3.11)
then uk→0 in L2∗(M).
Proof We claim that for all sequences (ξk)k∈N ⊂ RN and (jk)k∈N ⊂ N such that
jk → +∞ and for every i ∈ I we have
2−jkr(χiuk) ◦ expzi(2−jk · +ξk)⇀0 as k → ∞.
(3.12)
zi ) and (Bρ(yk), exp−1
yk
Since (3.12) is obviously true when ξk ≥ ρ, (indeed the terms in (3.12) are identically
zero for k large enough), we shall assume ξk ∈ Bρ(0) for all k ∈ N. Given i ∈ I,
we set yk := expzi(ξk) ∈ M and denote by ψk the transition map between the charts
(Bρ(zi), exp−1
◦ expzi (so that expzi =
expyk ◦ψk and ψk(ξk) = 0). Therefore, for k large enough, by using Taylor expansion
of the first order at ξk (where, for a lighter notation, we denote by ψ′
k(ξk) the Jacobi
k(ξk))−1 the corresponding Jacobian,
matrix of ψk at ξk (ψ′
and drop the dot symbol for the rows-by-columns product) we get, since jk → +∞,
that
k(ξk))−1 its inverse and by (ψ′
) i.e. we set ψk := exp−1
yk
2−jkr(χiuk)(expzi(2−jkξ + ξk)) = 2−jkr(χiuk)(expyk ◦ψk)(2−jkξ + ξk) =
2−jkr(χiuk)(expyk(2−jk(ψ′
k(ξk) + o(1))ξ)).
(3.13)
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
7
(we are using the Landau symbol o(1) to denote any (matrix valued) function uniformly
convergent to zero). In correspondence to any test function ϕ ∈ C ∞
0 (RN ),
ϕ(ξ)2−jkr[(χiuk) ◦ expzi(2−jkξ + ξk) − (χiuk) ◦ expyk(2−jkψ′
k(ξk)ξ)]dξ =
ϕ(ξ)2−jkr[(χiuk) ◦ expyk ◦ψk(2−jkξ + ξk) − (χiuk) ◦ expyk (2−jkψ′
k(ξk)ξ)]dξ =
B2ρ(0)
B2ρ(0)
2 η<C2−jk
k(ξk))−12jk
[(χiuk) ◦ expyk (cid:0)ψk((ψ′
k(ξk))−12jk
2 1
N+2
N+2
(ψ′
(ψ′
0
ϕ(2jk(ψ′
k(ξk))−1η)×
k(ξk))−1η + ξk(cid:1) − (χiuk) ◦ expy(η)]dη =
ϕ(2jk(ψ′
k(ξk))−1η)×
dsη<C2−jk
∇((χiuk) ◦ expyk(sψk((ψ′
k(ξk))−1η + ξk) + (1 − s)η)) · (ψk((ψ′
k(ξk))−1η + ξk) − η)dη,
(the second equality holds by integrating with respect to the variable η = 2−jkψ′
k(ξk)ξ).
Set, for each s ∈ [0, 1], ζ := sψk((ψ′
k(ξk))−1η + ξk) + (1 − s)η, since for η → 0,
ζ = η + O(η2) and since the Jacobian of the transformation is close to 1 in the
domain of integration, the modulus of the last expression is bounded by the following
one, which, in turn, can be estimated by Cauchy inequality. So, we have
C2jk
N+2
2 ζ<C2−jk
ϕ(2jk(ψ′
k(ξk))−1η(ζ))∇(χiuk) ◦ expyk(ζ)ζ2dζ ≤
N+2
C2jk
2 k∇(χiuk) ◦ expyk k2(cid:18)ζ<C2−jk
2 kukkH 1,2(M )(cid:18)ξ<C
N+2
C2jk
ϕ(2jk(ψ′
k(ξk))−1η(ζ))2ζ4dζ(cid:19)
1
2
≤
ϕ(ξ)22−4jkξ42−jkN dξ(cid:19)
1
2
≤ C2−jk−→0.
Therefore, by taking into account (3.13), we deduce that both sequences (2−jkr(χiuk)(expyk(2−jk·)))k∈N
and (2−jkr(χiuk)(expzi(2−jk ·+ξk)))k∈N have the same weak limit and, since (3.11) holds
true, (3.12) holds too.
Consequently, from the cocompactness of the embedding H 1,2(RN ) ֒→ L2∗(RN ) ([10,
Theorem 1]), it follows that for every i ∈ I,
(χiuk) ◦ expzi →0
in L2∗
(RN ) as k → ∞,
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
8
and therefore, since (χi)i∈I is a partition of unity subordinated to the atlas (Bρ(zi), exp−1
we deduce that
zi )i∈I,
M
uk2∗
dvg = M
Xi∈I
χiuk2∗
dvg ≤ CXi∈I
Bρ(0)
C Xi∈I
Bρ(zi)
χiuk2∗
dvg ≤
uk ◦ expzi(ξ)2∗
dξ → 0,
which proves the statement of the theorem.
4. Proof of Theorem 2.1 (profile decomposition)
1. Without loss of generality we may assume (by replacing uk with uk − u) that
uk ⇀ 0.
Then, setting for all i ∈ I
vk,i := (χiuk) ◦ expzi
we get that the sequence (vk,i)k∈N is bounded in H 1,2
0 (Bρ(0)) (and weakly converges to
zero), and so we can consider a profile decomposition of (vk,i)k∈N given by Theorem 2.2
when m = 1 and r = N −2
2 . An iterated extraction allows to find a subsequence which
has a profile decomposition for every i ∈ I i.e. such that for all i ∈ I the defect of
compactness of vk,i has the following form
(4.14)
Sk,i = Xn∈N\{0}
2j
(n)
k,i rw(n)
i
(2j
(n)
k,i (· − ξ(n)
k,i )) =: Xn∈N\{0}
c(n)
k,i .
By taking into account (4.14) we will be able to get concentration terms of χiuk by
zi . More in detail we consider
k,i of vk,i with exp−1
composing each concentration term c(n)
for all i ∈ I the term, defined on Bρ(zi),
C(n)
k,i := c(n)
k,i ◦ exp−1
zi = 2j
(n)
k,i rw(n)
i
(2j
(n)
k,i (exp−1
zi (·) − ξ(n)
k,i )).
Setting
we have that
k,i := expzi(ξ(n)
y(n)
k,i )
k,i rw(n)
i
Since for all i ∈ I and n ∈ N \ {0}
C(n)
k,i = 2j
(n)
(2j
(n)
k,i (exp−1
zi (·) − exp−1
zi (y(n)
k,i ))).
(4.15)
(4.16)
w(n)
i
:= w − lim
k→∞
2−j
(n)
k,i r(χiuk) ◦ expzi(2−j
(n)
k,i
· +ξ(n)
k,i ),
we can see that w(n)
by exp−1
zi
i
"evaluates" χiuk on points belonging to Bρ(zi) which are mapped
in subsets of Bρ(0) which are (for large k) concentrated around the points
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
9
ξ(n)
k,i . So, due to (4.16), it is sufficient to evaluate w(n)
Bρ(y(n)
k,i ). So, setting
i
on points which belong also to
Bi,k,n := exp−1
y
(n)
k,i
(Bρ(y(n)
k,i ) ∩ Bρ(zi)) ⊂ Bρ(0),
(4.17)
we shall consider the transition map between the charts (Bρ(y(n)
i.e. the map
k,i ), exp−1
) and (Bρ(zi), exp−1
zi ),
y
(n)
k,i
defined on Bi,k,n. Note that ψi,k,n(0) = ξ(n)
k,i , moreover, by setting for any x ∈ Bi,k,n
ψi,k,n := exp−1
zi ◦ expy
(n)
k,i
(4.18)
η := 2j
(n)
k,i exp−1
y
(n)
k,i
(x),
(4.19)
(n)
zi (x) = ψi,k,n(2−j
we have exp−1
k,i η) for all x ∈ Bi,k,n. Therefore (by using Taylor
expansion of the first order of the transition map ψi,k,n at 0, where, to use a lighter
notation we denote by ψ′
i,k,n(0))−1 its
inverse and omit the dot symbol for the rows-by-columns product) we deduce
i,k,n(0) the Jacobi matrix of ψi,k,n at zero, (ψ′
(n)
k,i (ψi,k,n(2−j
(n)
k,i η) − ξ(n)
k,i ) = 2j
(n)
k,i (ψi,k,n(2−j
(n)
k,i η) − ψi,k,n(0)) =
(n)
2j
k,i (exp−1
zi (x) − ξ(n)
k,i ) = 2j
i,k,n(0)η + O(2−j
= ψ′
(n)
(n)
k,i η2) = 2j
k,i ψ′
i,k,n(0) exp−1
y
(n)
k,i
(x) + O(2j
(n)
k,i (exp−1
y
(n)
k,i
(x))2).
(4.20)
Without loss of generality, applying Arzel`a-Ascoli theorem and passing to a suitable
subsequence, we can assume that (ψi,k,n)k∈N converges in the norm of C 1(RN ) as k → ∞
to some function ψi,n. We claim that, under a suitable renaming of the profile w(n)
,
namely by renaming w(n)
k,i (of χiuk) in (4.15)
i
take the following form:
, concentration terms C(n)
i,n(0) ·) as w(n)
(ψ′
i
i
C(n)
k,i := 2j
(n)
k,i rw(n)
i
(2j
For this purpose we show that, as k → ∞,
(n)
k,i exp−1
y
(n)
k,i
(·)).
(n)
2j
k,i rd(w(n)
Bρ(y
Indeed, the previous relation written under the coordinate map expy
zi (x)−ξ(n)
k,i ))−w(n)
k,i (exp−1
(n)
k,i )∩Bρ(zi)
i,n(0) exp−1
k,i ψ′
(2j
(2j
(n)
(n)
i
i
y
(n)
k,i
(x)))2dvg→0.
, i.e. by setting
(n)
k,i
ξ = exp−1
y
(n)
k,i
(x) becomes (by taking into account (4.18) and (4.17))
Bi,k,n
2j
(n)
k,i r∇(w(n)
i
(2j
(n)
k,i (ψi,k,n(ξ) − ξ(n)
k,i )) − w(n)
i
(2j
(n)
k,i ψ′
i,n(0)ξ))2dξ→0 as k → ∞,
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
10
and, by taking into account (4.19) (and by a null extension to whole of RN of the
involved functions), the claim will follow if, as k → ∞,
(n)
k,i
2−j
N+2
2 RN
ψ′
i,k,n(2−j
(n)
k,i η)∇w(n)
i
(2j
(n)
k,i (ψi,k,n(2−j
(n)
k,i η) − ξ(n)
k,i )) − ψ′
i,n(0)∇w(n)
i
(ψ′
i,n(0)η)2dη→0.
This last convergence easily follows by Lebesgue dominated convergence theorem, in-
deed (for all n and for all i) ∇w(n)
k,i → + ∞,
and (by taking into account that convergence of (ψi,k,n)k∈N and (ψ′
i,k,n)k∈N to ψi,n
i,k,n(2−j
and ψ′
i,n(0),
2j
i,n respectively is uniform) the pointwise convergence of ψ′
i ∈ L2(RN ), and when k → ∞, we have j(n)
i,n(0)η (as easily follows by (4.20) and (4.19)).
k,i (ψi,k,n(2−j
It is easy to see now that the renamed profiles w(n)
are obtained as pointwise limits
k,i η) − ξ(n)
k,i η)→ψ′
)→ψ′
(n)
(n)
(n)
i
i
(and thus also as weak limits)
w(n)
i
(ξ) = lim
k→∞
2−j
(n)
k,i r(χiuk) ◦ expy
(2−j
(n)
k,i
(n)
k,i ξ), for a.e. ξ ∈ RN .
(4.21)
2. Since each Bρ(zi) ⊂ B2ρ(zi) ⊂ M and M is compact, we may assume that for all
n ∈ N \ {0} and for all i ∈ I, there exist, up to subsequences, points of concentration
y(n)
i
:= lim
k→∞
y(n)
k,i .
(4.22)
In order to achieve the orthogonality relation (2.5) we shall introduce the following
equivalence relation on the set of sequences in M × R. Namely given (yk, jk)k∈N and
(y′
k)k∈N in M × Z we shall write
k, j′
k)k∈N when (jk − j′
k + 2jkd(yk, y′
k))k∈N is a bounded sequence.
k, j′
(yk, jk)k∈N ≃ (y′
(R)
Since the set I is a finite set, the number of sequences (y(n)
k,i )k∈N which can be
equivalent to a fixed sequence (y(¯n)
k,¯ı )k∈N is finite. Therefore we can exploit the
unconditional convergence with respect to the indexes (n) of the series Sk,i and syn-
chronize them by replacing ¯n and all the indexes m in the finite set
k,¯ı , j(¯n)
k,i , j(n)
N¯n := {m ∈ N \ {0} ∃i ∈ I s.t. (y(n)
k,i , j(n)
k,i )k∈N ≃ (y(¯n)
k,¯ı , j(¯n)
k,¯ı )k∈N}
with, say, the smallest integer in N¯n.
Thanks to this synchronization procedure the following property
k,i2 , j(m)
holds true for all i1, i2 ∈ I and m, n ∈ N \ {0}.
k,i1)k∈N ≃ (y(m)
k,i1, j(n)
(y(n)
k,i2 )k∈N ⇐⇒ m = n,
Note also that when (y(n)
k,i1, j(n)
we can set, modulo subsequences
k,i1)k∈N ≃ (y(n)
k,i2, j(n)
k,i2), since (j(n)
k,i2 − j(n)
k,i1)k∈N is bounded,
j(i1, i2, n) := lim
k→+∞
j(n)
k,i2 − j(n)
k,i1 ∈ R,
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
11
so that, by redefining w(n)
assume that (j(n)
bounded, we get (by (2.4)) that (see (4.22))
k,i2)k∈N = (j(n)
i2 (2−j(i1,i2,n)·) as (the corresponding profile) w(n)
k,i1, y(n)
k,i1)k∈N. Moreover, since also (2j
k,i1 d(y(n)
(n)
i2 , we can
k,i2))k∈N is
¯y(n)
i1 = ¯y(n)
i2
for all (y(n)
k,i1, j(n)
k,i1)k∈N ≃ (y(n)
k,i2, j(n)
k,i2)k∈N.
(4.23)
Finally, we show that the elementary concentrations terms C (n)
k,i do not change (up
k,i )k∈N in the same equivalence class. Namely
to a vanishing term) by varying (y(n)
the following property holds true
k,i , j(n)
(y(n)
k,i1, j(n)
k,i1)k∈N ≃ (y(n)
k,i2, j(n)
k,i2)k∈N ⇒ kC (n)
k,i1 − C (n)
k,i2k → 0,
for all i1, i2 ∈ I. Since, as shown above, we can assume, without restrictions, that
(j(n)
k,i2)k∈N (and we shall denote, to shorten notation, their common value as
(j(n)
y(n)
k,i2, we
have
k,i1)k∈N = (j(n)
k )k∈N) it will suffice to prove that, set ¯ξ(n)
y(n)
k,i1 and ¯ξ(n)
k,i1 = exp−1
zi1
k,i2 = exp−1
zi1
Bρ(zi1 )
2j
(n)
k rd(w(n)
i1 (2j
(n)
k (exp−1
zi1
(x)−¯ξ(n)
k,i2))−w(n)
i1 (2j
(n)
k (exp−1
zi1
(x)−¯ξ(n)
k,i1)))2dvg→0 as k → ∞.
(4.24)
Indeed, we get, modulo subsequences, that
2j
(n)
k ¯ξ(n)
k,i2−¯ξ(n)
k,i1 = 2j
(n)
k exp−1
zi1
y(n)
k,i2−exp−1
zi1
y(n)
k,i1 = 2j
Then, (2.5) follows directly from (4.23).
(n)
k d(y(n)
k,i2, zi1)−(y(n)
k,i1, zi1) ≤ 2j
(n)
k d(y(n)
k,i2, y(n)
k,i1) → 0.
3. Consider now the sum Pn∈N\{0}Pi∈I
k,i )k∈N, which are synchronized at the Step 2 as (y(n)
(j(n)
¯y(n) and (4.21) takes the form
C(n)
k,i , with the sequences (y(n)
k )k∈N and (j(n)
k )k∈N, while y(n)
k,i )k∈N and
k →
w(n)
i
(ξ) = lim
k→∞
2−j
(n)
k r(χiuk) ◦ expy
(2−j
(n)
k
(n)
k ξ), for a.e. ξ ∈ RN .
(4.25)
Since j(n)
k → ∞ implies expy
(2−j
(n)
k
(n)
k ξ) → ¯y(n) in M, we have from (4.25)
w(n)
i
(ξ) = χi(¯y(n)) lim
k→∞
2−j
(n)
k ruk ◦ expy
(2−j
(n)
k
(n)
k ξ), for a.e. ξ ∈ RN ,
(4.26)
taking into account that for each ξ ∈ RN the limit is evaluated with k ≥ k(ξ) with
some k(ξ) sufficiently large. Set
w(n) := Xi∈I
w(n)
i
.
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
12
Then relation (2.6) immediately follows from (4.26), w(n)
i = χi(¯y(n))w(n), and since, by
Step 1, defect of compactness of χiuk is a unconditionally convergent series, we have
Xi∈I Xn∈N\{0}
Xn∈N\{0}Xi∈I
C(n)
k,i (x) = Xn∈N\{0}Xi∈I
C(n)
k,i (x) =
2j
(n)
k rw(n)
i
(2j
(n)
k exp−1
y
(n)
k
(x)) =
w(n)(2j
Xn∈N\{0}
(n)
k exp−1
y
(n)
k
(x)), x ∈ Bρ(y(n)
k ) ,
which gives (2.7).
4. In order to prove the "energy" estimate (2.9), assume, without loss of generality,
that the sum in (2.7) is finite and that all w(n) have compact support, and expand by
bilinearity the trivial inequality ku − uk + Skk2
H 1,2(M ) ≥ 0. Then, by using the norm
(2.1) and the representation (2.3) of the scalar product in H 1,2(M), we have
(n)
k r
k2j
0 ≤ kukk2 + kuk2 − 2huk, ui + 2hu − uk, Ski+
Xn
Xm6=n
χ ◦ exp−1
y
χ ◦ exp−1
y
k exp−1
y
w(m)(2j
exp−1
y
(m)
k
w(n)(2j
(m)
k
r
h2j
(n)
k
(m)
k
(·))k2−
(n)
k
(n)
(m)
k
(·)), 2j
(n)
k r
χ ◦ exp−1
y
(n)
k
w(n)(2j
(n)
k exp−1
y
(n)
k
(·))i.
(4.27)
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
13
The first line of (4.27) can be evaluated taking into account that uk ⇀ u, Sk ⇀ 0, that
the definition of profiles w(n) given by (2.6) and that r = N −2
2 .
kukk2 + kuk2 − 2huk, ui + 2hu − uk, Ski =
ku2
kk + ku2k − 2kuk2 + o(1) − 2Xn
huk, 2j
(n)
k r
χ ◦ exp−1
y
(n)
k
w(n)(2j
(n)
k exp−1
y
(n)
k
(·))i =
2j
2Xn
(n)
k r ξ<ρ
N
Xi,j=1
2Xn
(n)
k
gy
ij ∂i(uk(expy
(n)
k
(ξ)))∂j(χ(ξ) w(n)(2j
2j
(n)
k rξ<ρ
uk(expy
(n)
k
(ξ))χ(ξ) w(n)(2j
(n)
(n)
k
i,j (ξ)dξ−
kukk2 − kuk2 + o(1)−
k ξ))rdet gy
k ξ)rdet gy
kukk2 − kuk2 + o(1)−
i,j (ξ)dξ =
(n)
k
(n)
2Xn
η<ρ2j
(n)
k
N
Xi,j=1
(n)
k
gy
ij ∂i(2−j
(n)
k ruk ◦ expy
(n)
k
(2−j
(n)
k η))∂j(χ(2−j
(n)
k η) w(n)(η))·
2−2j
2Xn
(n)
k η<ρ2j
(n)
k
2−j
(n)
k ruk ◦ expy
(n)
k
(2−j
(n)
k η)χ(2−j
rdet gy
k η)w(n)(η)rdet gy
(n)
(n)
k
i,j (2−j
(n)
k η) dη−
(n)
k
i,j (2−j
(n)
k η)dη =
kukk2 − kuk2 + o(1) − 2Xn
RN
N
Xi
(n)
∂iw(n)(η)2dη − 2Xn
2−2j
k RN
kukk2 − kuk2 − 2Xn
w(n)(η)2dη =
k∇w(n)k2
2 + o(1).
(n)
(n)
(In the third equality we have set η = 2j
k ξ, while in the fourth we have used the fact,
due to (2.6) that 2−j
k ·)⇀χ(0)w(n) = w(n) as k → ∞ (in
our slightly modified sense of weak convergence). Note also we have still denoted by
∂i (resp. ∂j) the derivative with respect to the ith (resp jth) component of η = 2j
k ξ.
Finally in the last equality we have used (2.1)).
k ·)(uk ◦ expy
k χ(2−j
)(2−j
(n)
k
(n)
(n)
(n)
In order to estimate the second line of (4.27) we shall split (according to (2.1)) the
H 1,2(M)-norm into the L2-norm of the gradient (gradient part) and the L2-norm of
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
14
the function (L2 part) and consider first the latter. Since
Xn
k (N −2)Bρ(yn)
(n)
2j
Xn
(n)
k
k2j
N −2
2 χ ◦ exp−1
y
(n)
k
w(n)(2j
(n)
k exp−1
y
(n)
k
(·))k2
2 =
χ ◦ exp−1
y
(n)
k
(x) w(n)(2j
(n)
k exp−1
y
(n)
k
(x))2dvg =
2j
Xn
χ(2−j
χ(ξ)(w(n)(2j
(n)
k (N −2)ξ<ρ
k η)w(n)(η)2rdet gy
(n)
(n)
k ξ)2rdet gy
(n)
k
i,j (ξ)dξ =
(n)
k
i,j (2−j
(n)
k η)dη→0 as k → ∞,
2−2j
Xn
(n)
k η<ρ2j
(n)
k
(since j(n)
sum of the gradient terms as follows.
k →∞) as k → ∞, the second line of (4.27) is evaluated in the limit by the
N
2j
Xn
gy
Xi,j=1
gy
ij ∂i(χ(2−j
(n)
k
ij
(n)
k
Xn
Xn
(n)
2j
k (N −2)ξ<ρ
η<ρ2j
Xi,j=1
(n)
k
N
(n)
k (N −2)Bρ(y
d(χ ◦ exp−1
y
(n)
k
(x) w(n)(2j
(n)
k )
(n)
k exp−1
y
(n)
k
(x)))2dvg =
(ξ)∂i(χ(ξ) w(n)(2j
(n)
k ξ))∂j(χ(ξ) w(n)(2j
(n)
k ξ))rdet gy
(n)
k
i,j (ξ)dξ =
(n)
k η) w(n)(η))∂j(χ(2−j
(n)
k η) w(n)(η))rdet gy
(n)
k
i,j (2−j
(n)
k η) dη
→Xn
RN
∇w(n)(η)2 dη = Xn
k∇w(n)k2 as k → ∞.
Consider now the terms in the sum in third line of (4.27). Note that the L2-part of
the scalar product converges to zero by Cauchy inequality and by the calculations for
the first line of (4.27). At the light of the orthogonality condition (2.5) we have to face
two cases.
Case 1: The sequence (j(n)
k − j(m)
k
k → + ∞ as k → ∞. Then, using changes of variables ξ = exp−1
)k∈N is unbounded. Assume without loss of gener-
(x)
ality that j(n)
k − j(m)
y
(n)
k
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
15
and η = 2j
(n)
k ξ,
(m)
k
r
h2j
χ ◦ exp−1
y
(m)
k
(x) w(m)(2j
(m)
k
exp−1
y
(m)
k
(·)), 2j
(n)
k r
χ ◦ exp−1
y
(n)
k
(x) w(n)(2j
(n)
k exp−1
y
(n)
k
(·))i =
2j
(n)
k r2j
(m)
k
r Bρ(y
(m)
k
)∩Bρ(y
(n)
k )
d(χ ◦ exp−1
y
(m)
k
(x) w(m)(2j
(m)
k
exp−1
y
(m)
k
(x)))·
d(χ ◦ exp−1
y
(n)
k
(x) w(n)(2j
(n)
k exp−1
y
(n)
k
(x)))dvg + o(1) =
2j
(n)
k r2j
(m)
k
r ξ<ρ
N
Xi,j=1
gy
(n)
k
ij
(ξ)∂i(χ(ξ) w(n)(2j
(n)
k ξ))·
∂j(χ(exp−1
y
(m)
k
(expy
(n)
k
(ξ))) w(m)(2j
(m)
k
exp−1
y
(m)
k
(expy
(n)
k
(ξ))))rdet gy
(n)
k
i,j (ξ)dξ =
2−j
(n)
k r2j
(m)
k
rη<ρ2j
gy
(n)
k
ij
(2−j
(n)
k η)∂i((1 + o(1)) w(n)(η))·
N
(n)
k
Xi,j=1
(expy
(n)
k
∂j((1 + o(1)) w(m)(2j
(m)
k
exp−1
y
(m)
k
since, by (2.6),
(2−j
(n)
k η))))(1 + o(1))dη + o(1) → 0,
w − lim
k→∞
2−j
(n)
k r2j
(m)
k
rw(m)(2j
(m)
k
(exp−1
y
(m)
k
◦ expy
(n)
k
)(2−j
(n)
k ·)) = w − lim
2−j
(n)
k ruk(·) = 0.
k→∞
(n)
k d(y(n)
k , y(m)
Case 2: 2j
) → ∞ as k → ∞. Since case 1 has been ruled out, we can
assume without restrictions that the sequence j(m)
k = j ∈ R for all large k. Then,
by arguing as above (and in particular by taking into account that the L2-part of the
scalar product is negligible), we get that, as k → ∞,
k − j(n)
k
(m)
k
r
h2j
χ ◦ exp−1
y
(m)
k
w(m)(2j
(m)
k
exp−1
y
(m)
k
(·)), 2j
(n)
k r
χ ◦ exp−1
y
(n)
k
w(n)(2j
(n)
k exp−1
y
(n)
k
(·))i → 0,
since the values of w(m) and of w(n) are set to concentrate at sufficiently separated
points, indeed d(2j
k d(y(n)
k d(y(n)
k y(m)
) → ∞.
) ≥ 2j
) = 2j
(m)
(n)
(n)
k , 2jy(m)
k
(n)
k y(n)
k , 2j
k , y(m)
k
k
Then, by applying the estimates obtained for the three lines of inequality (4.27) we
finally deduce (2.9) concluding the proof of Theorem 2.1.
Acknowledgment. The first author is supported by GNAMPA of the "Istituto Nazionale
di Alta Matematica (INdAM)" and by MIUR - FFABR - 2017 research grant.
http://dx.doi.org/10.13039/501100003407
The second author had no academic affiliation when working on this paper.
A PROFILE DECOMPOSITION FOR THE LIMITING SOBOLEV EMBEDDING
16
References
[1] Adimurthi, A., and Tintarev, C. On compactness in the Trudinger-Moser inequality. Ann.
Sc. Norm. Sup. Pisa Cl. Sci. 5 (2014), 1 -- 18.
[2] Adimurthi, A., and Tintarev, C. Defect of compactness in spaces of bounded variation. J.
Func. Anal., 271 (2016), 37 -- 48.
[3] Devillanova, G. Multiscale weak compactness in metric spaces. J. Elliptic Parabol. Eq. 2, 131
(2016). http://dx.doi.org/10.1007/BF03377397.
[4] Devillanova, G., Solimini, S., and Tintarev, C. On weak convergence in metric spaces.
Nonl. Anal. Opt., Contemp. Math. 659 (2016), 43 -- 63.
[5] Devillanova, G., Solimini, S., and Tintarev, C. Profile decomposition in metric spaces.
Pure Appl. Funct. Anal. 2, 4 (2017), 599 -- 628.
[6] G´erard, P. Description de compacit´e de l'injection de Sobolev. ESAIM Control Optim. Calc.
Var., 3 (1988), 213 -- 233.
[7] Jaffard, S. Analysis of the lack of compactness in the critical Sobolev embeddings. J. Funct.
Analysis, 161 (1999), 384 -- 396.
[8] Lieb, E. On the lowest eigenvalue of the laplacian for the intersection of two domains. Invent.
Math., 74 (1983), 441 -- 448.
[9] Schindler, I., and Tintarev, K. An abstract version of the concentration compactness prin-
ciple. Revista Mat. Complutense, 15 (2002), 1 -- 20.
[10] Solimini, S. A note on compactness-type properties with respect to Lorentz norms of bounded
subsets of a Sobolev space. Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, 12 (1995), 319 -- 337.
[11] Solimini, S., and Tintarev, C. Analysis of concentration in the Banach space. Comm. Con-
temp. Math., 18 (2016). http://dx.doi.org/10.1142/S0219199715500388.
[12] Struwe, M. A global compactness result for elliptic boundary value problems involving limiting
nonlinearities. Math. Z., 187 (1984), 511 -- 517.
Dipartimento di Meccanica, Matematica e Management Politecnico di Bari, via E.
Orabona n. 4, 70125 Bari, Italy
E-mail address: [email protected]
Sankt Olofsgatan 66B, 75330 Uppsala, Sweden
E-mail address: [email protected]
|
1511.05702 | 1 | 1511 | 2015-11-18T09:26:02 | Algebras of singular integral operators with kernels controlled by multiple norms | [
"math.FA"
] | The purpose of this paper is to study algebras of singular integral operators on $\mathbb{R}^{n}$ and nilpotent Lie groups that arise when one considers the composition of Calder\'on-Zygmund operators with different homogeneities, such as operators that occur in sub-elliptic problems and those arising in elliptic problems. For example, one would like to describe the algebras containing the operators related to the Kohn-Laplacian for appropriate domains, or those related to inverses of H\"ormander sub-Laplacians, when these are composed with the more standard class of pseudo-differential operators. The algebras we study can be characterized in a number of different but equivalent ways, and consist of operators that are pseudo-local and bounded on $L^{p}$ for $1<p<\infty$. While the usual class of Calder\'on-Zygmund operators is invariant under a one-parameter family of dilations, the operators we study fall outside this class, and reflect a multi-parameter structure. | math.FA | math |
Algebras of singular integral operators
with kernels controlled by multiple norms
Alexander Nagel∗
Fulvio Ricci†
Elias M. Stein ‡
Stephen Wainger §
Contents
1 Introduction
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Some motivating examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Plan of the paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
3
4
5
2 The Classes P(E) and M(E)
11
2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Global norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Classes of distributions and multipliers . . . . . . . . . . . . . . . . . . . . . . . 14
3 Marked partitions and decompositions of RN
15
. . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Characterizing the sets EA
3.1 Dominant terms in Nj(t) and bNj(t)
3.2 Marked partitions and the sets ES and bES . . . . . . . . . . . . . . . . . . . . . 16
S and bEA
3.4 Estimates of kernels and multipliers on ES and bES . . . . . . . . . . . . . . . . 20
3.5 A coarser decomposition of RN associated to S ∈ S(n) . . . . . . . . . . . . . . 21
S . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 Fourier transform duality of kernels and multipliers
4.1 Fourier transforms of multipliers
4.2 Fourier transforms of kernels
22
. . . . . . . . . . . . . . . . . . . . . . . . . . 22
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5 Dyadic sums of Schwartz functions
28
5.1 Cones associated to a matrix E . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Inclusions among classes of kernels associated to different matrices . . . . . . . 29
5.2
5.3 Coarser decompositions and lower dimensional matrices
. . . . . . . . . . . . . 30
5.4 Size estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.5 Cancellation properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.6 Dyadic sums with weak cancellation . . . . . . . . . . . . . . . . . . . . . . . . 36
6 Decomposition of multipliers and kernels
38
6.1 New matrices ES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.2 Road map for the dyadic decomposition . . . . . . . . . . . . . . . . . . . . . . 39
6.3 Partitions of unity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.4 Dyadic decomposition of a multiplier . . . . . . . . . . . . . . . . . . . . . . . . 43
∗University of Wisconsin-Madison, Madison WI 53706, [email protected]
†Scuola Normale Superiore, Piazza dei Cavalieri 7, 56126 Pisa, [email protected]
‡Princeton University, Princeton NJ 08544. [email protected]
§University of Wisconsin-Madison, Madison WI 53706, [email protected]
2
6.5 Dyadic decomposition of a kernel
. . . . . . . . . . . . . . . . . . . . . . . . . . 45
7 The rank of E and integrability at infinity
47
7.1 The rank 1 case: Calder´on-Zygmund kernels . . . . . . . . . . . . . . . . . . . . 48
7.2 Higher rank and integrability at infinity . . . . . . . . . . . . . . . . . . . . . . 50
7.3 Higher rank and weak-type estimates near zero . . . . . . . . . . . . . . . . . . 51
8 Convolution operators on homogeneous nilpotent Lie groups
53
8.1 Convolution of scaled bump functions: compatibility of dilations and convolution 53
8.2 Automorphic flag kernels and Lp-boundedness of convolution operators . . . . 56
9 Composition of operators
57
9.1 A preliminary decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
9.2 Reduction to the case of finite sets . . . . . . . . . . . . . . . . . . . . . . . . . 60
9.3 Properties of the convolution [ϕI
T ]J . . . . . . . . . . . . . . . . . . . . . 61
9.4 A further decomposition of ΓZ(ES) × ΓZ(ET ) . . . . . . . . . . . . . . . . . . . 64
9.5 Fixed and free indices
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
9.6 A finer decomposition of RN . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
9.7 The matrix ES,T,W . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
9.8 The class P0(ES,T,W )
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
9.9 Estimates for [ϕ]I ∗ [ψ]J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
9.10 Proof of Theorem 9.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
S]I ∗ [ψJ
10 Convolution of Calder´on-Zygmund kernels
76
10.1 Calder´on-Zygmund kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
10.2 A general convolution theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
10.3 Convolution of two Calder´on-Zygmund kernels
. . . . . . . . . . . . . . . . . . 77
11 Two-flag kernels and multipliers
82
11.1 Flag kernels and multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
11.2 Pairs of opposite flags
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
11.3 Two-step flags
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
11.4 General two-flag kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
11.5 Proof of Lemma 11.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
12 Extended kernels and operators
91
12.1 The Lp boundedness of the operators . . . . . . . . . . . . . . . . . . . . . . . . 91
12.2 The algebra of operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
13 The role of pseudo-differential operators
95
13.1 The isotropic extended kernels
. . . . . . . . . . . . . . . . . . . . . . . . . . . 97
13.2 Proof of Theorem 13.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
13.3 The space Lp
1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
14 Appendix I: Properties of cones Γ(A)
102
14.1 Optimal inequalities and the basic hypothesis . . . . . . . . . . . . . . . . . . . 102
14.2 Partial matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
14.3 Projections
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
14.4 The dimension of Γ(E) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
14.5 The reduced matrix E♭ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
15 Appendix II: Estimates for homogeneous norms
108
16 Appendix III: Estimates for geometric sums
Index of symbols
1
Introduction
3
110
116
The purpose of this paper is to study algebras of singular integral operators on Rn and
nilpotent Lie groups that arise when one considers the composition of Calder´on-Zygmund
operators with different homogeneities, such as operators that occur in sub-elliptic problems
and those arising in elliptic problems. For example, one would like to describe the algebras
containing the operators related to the Kohn-Laplacian for appropriate domains, or those
related to inverses of Hormander sub-Laplacians, when these are composed with the more
standard class of pseudo-differential operators. The algebras we study can be characterized
in a number of different but equivalent ways, and consist of operators that are pseudo-local
and bounded on Lp for 1 < p < ∞. While the usual class of Calder´on-Zygmund operators is
invariant under a one-parameter family of dilations, the operators we study fall outside this
class, and reflect a multi-parameter structure.
This paper is the second in the series begun with [NRSW12].
1.1 Background
An initial impetus for the study of composition of singular integral operators of different
types came from the study of the ∂-Neumann problem in complex analysis.
In the case
of domains in Cn where matters are sufficiently well understood, the relevant "Calder´on
operator" for this boundary-value problem can be viewed as a composition of a sub-elliptic
type operator with a standard pseudo-differential operator of order 0. Early studies of such
compositions in the context of the Heisenberg group can be found [GS77] and in [PS82].
In [MRS95] more general such operators appear as singular integrals with "flag kernels".
The corresponding algebra was broad enough to contain, for example, all operators of the
form m(L, iT ), where L is the sub-Laplacian and T is the central invariant vector field, with
m a multiplier of Marcinkiewicz-type. The notion of flag kernels (having singularities on
appropriate flag varieties) and the properties of the corresponding singular integrals were
then extended to the higher step case in [NRS01], largely in the Euclidean setting, and then
in [NRSW12] in the context of automorphic flags on a general homogeneous group. In this
connection we should mention the general theory of Brian Street [Str14] which treats aspects
of both singular integrals and singular Radon transforms in the context of multi-parameter
analysis.
However, the general nature of these results did not provide an entirely satisfactory answer
to a main question that arises when composing singular integrals of different homogeneities:
that of characterizing the resulting class of kernels. It turns out that the class in question
is indeed narrower than the class of flag kernels: the kernels satisfy stricter differential in-
equalities, and in particular they are all smooth away from the origin, and the corresponding
operators are thus pseudo-local.
A different approach to the study of flag kernels can be found in the work of G lowacki; see
[G lo10a], [G lo10b], [G lo13]. There has been recent interest in the study of Hardy spaces asso-
ciated to flags and flag kernels (see for example [HL10], [HLL10], [Rua11], [HLL13], [HLS14],
[Wu14a]) and in the study of weighted norm inequalities (see for example [Wu14b]). Fur-
ther recent references that deal with flag kernels include [Yan09], [DLM10] [Rua10], [WL12],
[LZ13], [SY13], [Wu14b], and [MPR15].
4
1.2 Some motivating examples
We begin by describing a particularly simple situation occurring in studying the heat
equation on Rn × R, or in studying convolution of operators arising on the non-Abelian
Heisenberg group Hn = {(z, t) : z = (z1, . . . , zn) ∈ Cn, t ∈ R}. In the latter case, Calder´on-
Zygmund kernels adapted to the automorphic dilations δ · (z, t) = (δz, δ2t) are distributions
K on Hn which, away from the origin, are given by integration against a smooth function K
satisfying
and which satisfy appropriate cancellation conditions. Similarly the isotropic variants of these
kernels (the building blocks of the standard pseudodifferential operators) satisfy instead the
inequalities
1
t ∂α
t ∂α
(cid:12)(cid:12)(cid:12)∂β
z,¯zK(z, t)(cid:12)(cid:12)(cid:12) .(cid:0)z + t
2(cid:1)−2n−2−α−2β
(cid:12)(cid:12)(cid:12)∂β
z,¯zK(z, t)(cid:12)(cid:12)(cid:12) . (z + t)−2n−1−α−β .
z,¯zK(z, t)(cid:12)(cid:12)(cid:12) . (z + t)−2n−α (z2 + t)−1−β
t ∂α
(cid:12)(cid:12)(cid:12)∂β
(1.1)
(1.2)
(1.3)
If we consider kernels with compact support, it turns out that the relevant algebra of
kernels containing both kinds of distributions are those given by the mixed differential in-
equalities
for (z, t) in the unit ball, in addition to cancellation conditions.
There are several ways of thinking about the estimates in equation (1.3). One notices that
the inequalities in (1.3) are the best that can be satisfied on the unit ball for kernels K that
are assumed to be either of type (1.1) or of type (1.2), but this rather simple situation is not
repeated in the higher step case1. However, one can make two more productive observations.
a) Operators satisfying (1.3) can be understood in terms of the theory of flag kernels studied
in [NRS01] and [NRSW12]. Consider the following two sets of differential inequalities:
t ∂α
(cid:12)(cid:12)(cid:12)∂β
(cid:12)(cid:12)(cid:12)∂β
z,¯zK(z, t)(cid:12)(cid:12)(cid:12) . z−2n−α(cid:0)z2 + t(cid:1)−1−β
z,¯zK(z, t)(cid:12)(cid:12)(cid:12) .(cid:0)z + t(cid:1)−2n−α
t−1−β.
t ∂α
,
(1.4)
(1.5)
Equation (1.4) gives the differential inequalities satisfied by flag kernels for the flag (0) ⊂
Cn ⊂ Cn ⊕ R, while equation (1.5) give the differential inequalities for the opposite flag
(0) ⊂ R ⊂ Cn ⊕ R. Operators of type (1.1) satisfy (1.4), operators of type (1.2) satisfy
(1.5), and locally (for (z, t) in compact sets), operators of type (1.1) or (1.2) satisfy both
(1.4) and (1.5). Moreover, we will see that a kernel satisfying both (1.4) and (1.5) auto-
matically satisfies (1.3). Thus one is led to the study of two-flag kernels which locally are
simultaneously flag kernels for two opposite flags.
b) A second related perspective is to view kernels K satisfying the conditions of (1.3) as
satisfying differential inequalities appropriate to two different families of dilations on RN .
Derivatives in z or ¯z are controlled by the dilations (z, t) → (δz, δt), while derivatives with
respect to t are controlled by the dilations (z, t) → (δz, δ2t). This suggests that, more
generally, we should study operators whose kernels satisfy differential inequalities in which
different derivatives are controlled by different families of dilations.
1See Subsection 1.3.8 below for an interesting example, and Section 7 for a more complete discussion
5
In this paper we begin by adopting the second point of view.
It will then turn out that
the resulting class of distributions includes the class which belong to two flags. One of the
main goals of this paper is the study the composition of convolution operators with kernels
of this form. In the Euclidean context this reduces to the study of the product of the Fourier
multipliers, but in the non-Abelian situation the composition becomes a more serious issue.
A second goal of this paper is to establish regularity of two-flag kernels. The inequalities in
equations (1.4) and (1.5) are the differential inequalities for a kernel belonging simultaneously
to two opposite 2-step flags, and it is easy to see that these imply the inequalities in equation
(1.3). However in the case of higher step, the results are not so simple, and in fact are rather
surprising. Consider, for example, the case of two opposite 3-step flags in R3:
F :
F ⊥ :
(0) ⊂ {(x, y, z) : y = z = 0} ⊂ {(x, y, z) : z = 0} ⊂ R3,
(0) ⊂ {(x, y, z) : x = y = 0} ⊂ {(x, y, z) : x = 0} ⊂ R3.
A flag kernel K for the flag F with homogeneity λ · (x, y, z) = (λx, λy, λz) satisfies the
differential inequalities
x∂b
y∂c
x∂b
y∂c
A flag kernel K for the flag F ⊥ with homogeneity λ · (x, y, z) = (λx, λ
differential inequalities
zK(x, y, z)(cid:12)(cid:12) ≤ Ca,b,c x−1−a(x + y)−1−b(x + y + z)−1−c.
zK(x, y, z)(cid:12)(cid:12) ≤ Ca,b,c (x + y2 + z3)−1−a(y + z
1
2 y, λ
2 )−1−bz−1−c.
3
(cid:12)(cid:12)∂a
(cid:12)(cid:12)∂a
Suppose that K is a distribution with compact support which is a flag kernel for both flags.
Looking only at the inequalities in equations (1.6) and (1.7), these estimates provide no
information about K(x, y, z) if x = z = 0 and y 6= 0; in terms of the assumed size estimates
alone, the distribution K could be singular away from the origin. Unexpectedly, it turns out
that
1
3 z) satisfies the
(1.6)
(1.7)
x∂b
y∂c
(cid:12)(cid:12)∂a
zK(x, y, z)(cid:12)(cid:12) ≤ Ca,b,c (x + y2 + z3)−1−a(x + y + z
and the distribution K is indeed pseudo-local. It is important to note that this is a consequence
of both the differential inequalities and the cancellation conditions.
3
2 )−1−b(x + y + z)−1−c,
1.3 Plan of the paper
Our results can be divided roughly into two kinds: those concerning properties of the
kernels and those concerning the resulting convolution operators.
1. We begin by studying a class P(E) of distributions on RN given away from the origin by
integration against a smooth function. These functions are required to satisfy differential
inequalities adapted to a set of n ≤ N families of dilations on RN prescribed by an n × n
matrix E. Distributions K ∈ P(E) must also satisfy appropriate cancellation conditions.
2. We are primarily interested in the local behavior of these distributions2, and it will be
convenient to modify the class of kernels outside the unit ball so that they and all their
derivatives are rapidly decreasing at infinity. Under these hypotheses, we shall denote the
class of modified kernels by P0(E).
2We will see below in Section 7 that if the rank of the matrix E is greater than 1, then the kernels K ∈ P(E)
are integrable at infinity.
6
3. We then characterize these distributions in terms of their Fourier transform, and also in
terms of decompositions as infinite dyadic sums of dilates of normalized bump functions.
This material is discussed in Sections 2 - 6. In Section 7 we show that these distributions
fall outside the class of standard Calder´on-Zygmund kernels. In Section 10 we find the
smallest class P0(E) containing the convolution of Calder´on-Zygmund kernels with different
homogeneities. In Section 11 we show that two-flag kernels provide examples of the classes
P(E).
After establishing the basic properties of the distributions K ∈ P0(E), we fix a homogeneous
nilpotent Lie group G whose underlying space is RN and thus has a given automorphic
family of dilations. (Of course this includes the special case in which the group G is the
abelian Rn.)
4. If the n families of dilations on RN given by the matrix E are compatible3 with the
automorphic dilations, we show that convolution on G with a distribution K ∈ P(E)
defines a bounded operator on Lp(G) for 1 < p < ∞, and that the collection of these
convolution operators form an algebra under composition. This material is discussed in
Sections 8 - 9.
5. We also characterize the smallest algebra of convolution operators which arise when one
composes Calder´on-Zygmund operators of different homogeneities. This material appears
in Section 10.
6. Finally we describe a generalization of the class P(E) to allow variable coefficients, and
we investigate the commutation properties of these operators. In particular, we study the
role of the classical pseudo-differential operators. This material is considered in Sections
12 and 13.
In the following subsections we provide a more detailed summary of the results in this paper.
1.3.1 Section 2: The class of kernels
In Section 2 we provide a precise definition of a class of distributions and their associated
multipliers. We start with a fixed decomposition RN = RC1 ⊕ · · · ⊕ RCn. Write x ∈ RN
such as xj, to indicate a tuple of coordinates, and standard type, such as xk, to indicate a
single coordinate.) Each component RCj is equipped with a family of dilations, denoted by
λ · xj for λ > 0. For simplicity of exposition in this Introduction, we assume that these are
as (x1, . . . , xn) with xj = (cid:0)xk : k ∈ Cj(cid:1) ∈ RCj . (Throughout this paper we use boldface,
the standard isotropic dilations λ · xj =(cid:0)λxk : k ∈ Cj(cid:1) although we will allow more general
non-isotropic dilations in Section 2 below. If Qj is the homogeneous dimension of RCj , then
for isotropic dilations, Qj is the dimension of RCj .4 Let nj be a smooth homogeneous norm
on RCj so that nj(λ · xj ) = λ nj(xj ). Thus for isotropic dilations, we can take nj to be the
standard Euclidean norm on RCj .
The homogeneities on each component RCj can be weighted in various ways to give a
global family of dilations on RN . Let E = {e(j, k)} be an n × n matrix with strictly positive
entries, and for 1 ≤ j ≤ n, define a family of dilations on RN by setting
δj(λ)[x] =(cid:0)λ1/e(j,1) · x1, . . . , λ1/e(j,j) · xj , . . . , λ1/e(j,n) · xn(cid:1).
3A precise definition is given in Subsection 8.1 below.
4If λ · xj = exp[Aj log λ](xj) the homogeneous dimension of RCj is the trace of Aj.
(1.8)
Put
Nj(x) = n1(x1)e(j,1) + · · · + nj(xj )e(j,j) + · · · nn(xn)e(j,n)
so that Nj(cid:0)δj(λ)[x](cid:1) = λNj (x). Thus N1, . . . , Nn are homogeneous norms5 on RN for the
different dilation structures. We require that the components of the matrix E satisfy
7
(1.9)
e(j, j) = 1
e(j, k) ≤ e(j, l)e(l, k) 1 ≤ j, k, l ≤ n.
1 ≤ j ≤ n,
(1.10)
We study a family P = P(E) of distributions K on RN which are given away from the
origin 0 ∈ RN by integration against a smooth function K satisfying the following differential
inequalities6:
Nj(x)−Qj −αj.
(1.11)
x1 · · · ∂αn
(cid:12)(cid:12)∂α1
xn K(x)(cid:12)(cid:12) .
nYj=1
(Note that derivatives of K with respect to any of the variables in xj are controlled by the
family of dilations δj(λ) and the homogeneous norm Nj.) In addition, we impose certain can-
cellation conditions on the distributions K. These are analogous to cancellation conditions for
product kernels or flag kernels. Roughly speaking, they require that if x = (x′, x′′) is a decom-
position of the coordinates x1, . . . , xn into two subsets, then K(x′) = R K(x′, x′′)ψ(x′′) dx′′
is a distribution in the x′′ variables of the same type. The precise conditions are specified in
part (b) of Definition 2.2 below.
Note that in the example of the Heisenberg group given above, we have the natural de-
composition Hn = Cn ⊕ R ∼= R2n ⊕ R. The isotropic dilations on each component are given by
λ · (z1, . . . , zn) = (λz1, . . . , λzn) and λ · t = λt, so that the homogeneous dimension of Cn is 2n
2 and n2(t) = t,
and the homogeneous dimension of R is 1. Set n1(z) = z =(cid:2)Pn
so that n1(λ · z) = λn1(z) and n2(λ · t) = λn2(t). If we take E =(cid:20)1 1
j=1 zj2(cid:3) 1
2 1(cid:21) then
N1(z, t) = n1(z) + n2(t) ≈ z + t,
N2(z, t) = n1(z)2 + n2(t) ≈ z2 + t,
(1.12)
and the corresponding dilations are given by
δ1(λ)(z, t) = (λz, λt),
δ2(λ)(z, t) = (λ
1
2 z, λt).
The differential inequalities required by (1.11) are then
z,¯z∂β
(cid:12)(cid:12)∂α
t K(z, t)(cid:12)(cid:12) . N1(z, t)−2n−αN2(z, t)−1−β ≈ (z + t)−2n−α(z2 + t)−1−β
which is in agreement with (1.3).
1.3.2 Section 3: Marked partitions
Our analysis of the distributions K ∈ P0(E) is based on the decomposition of the unit ball
into the regions where one summand in each norm Nj(x) = n1(x1)e(j,1) + · · · + nn(xn)e(j,n)
is strictly larger than all the others. (There is an analogous decomposition of the Fourier
5This is standard terminology, although we only have Nj(x + y) ≤ Aj [Nj(x) + Nj(y)] for some constant
Aj instead of the usual triangle inequality.
6We use standard multi-index notation; see Section 2 for more details.
8
transform ξ-space for ξ large, which is needed to understand the differential inequalities of
the multiplier m = bK.) Matters are simplest in the case when n = 2 exemplified by the
Heisenberg group, where the two norms are given in equation (1.12). In this case there are
three regions near the origin.
1. The set z2 & t & t2, where the isotropic dilations are controlling. Here N1(z, t) ≈ z ≈
z + t and N2(z, t) ≈ z2 ≈ (z + t)2.
2. The set z . t . t 1
2 , where the automorphic dilations are controlling. Here N1(z, t) ≈
t ≈ z2 + t and N2(z, t) ≈ t ≈ z2 + t.
3. The intermediate region where t . z . t 1
2 where kernels of the type (1.3) behave like
"product kernels". Here N1(z, t) ≈ z and N2(z, t) ≈ t.
The case n ≥ 3 is more intricate. The required systematic decomposition of the x-space
depends on the notion of a 'marked partition' of the set {1, . . . , n}, which is a collection of
r=1 Ir = {1, . . . , n}, together with a
disjoint non-empty subsets I1, . . . , Is of {1, . . . , n} withSs
'marked' element kr in each subset Ir. This concept arises as follows.
Suppose the matrix E satisfies the conditions given in equation (1.10), and let x be a
point in the unit ball such that no two terms in any norm Nj(x) are equal. We will see in
Proposition 3.3 below that if nk(xk) is the dominant term in Nj(x) for some j 6= k then nk(x)
is also the dominant term in Nk(x). Thus given such a point x, there is a well-defined set
{k1, . . . , ks} of indices such that nkr (xkr ) is the dominant term in Nkr (x). For 1 ≤ r ≤ s, we
let Ir be the set of indices j such that nkr (xkr ) is dominant in Nj(x). Then kr ∈ Ir and this
produces a marked partition. The details are given below in Section 3.
version of a power of the function F (x) =Qn
One may note that the decomposition given via marked partitions is somewhat akin to
what happens in the theory of "resolutions of singularities", here applied to (say) a smooth
j=1 Nj(x). Using that theory (see e.g. [CGP13])
one would be led to decompose a neighborhood of the origin into a number of parts, and in
each part make an appropriate change of variables so that F becomes essentially a monomial
in the new variables. However, in our situation the decomposition by marked partitions is
much more explicit and tractable, and no auxiliary change of variables is required.
1.3.3 Section 4: Characterization by Fourier transforms
For ξ ∈ RN = RC1 ⊕ · · · ⊕ RCn write ξ = (ξ1, . . . , ξn) with ξk ∈ RCk and set
(1.13)
.
(1.14)
This is the 'dual norm' to Nj(x). Denote by M∞(E) the class of multipliers m ∈ C∞(RN )
satisfying
bNj(ξ) = ξ11/e(1,j) + · · · + ξj + · · · + ξn1/e(j,n).
(cid:12)(cid:12)(cid:12)∂α1
nYj=1h1 + bNj(ξ)i−αj
that K ∈ P0(E) if and only if its Fourier transform m = bK ∈ M∞(E).
m(ξ)(cid:12)(cid:12)(cid:12) .
· · · ∂αn
ξn
ξ1
1.3.4 Sections 5 and 6: Characterization by dyadic decompositions
On page 6 we indtroduced the subclass P0(E) of proper distributions. In Section 4 we show
In Section 5 we show that kernels K ∈ P0(E) can be characterized by their dyadic decompo-
sition into dilates of bump functions. Let {ϕI } be a uniformly bounded family of C∞-functions
supported in the unit ball, where I runs over the set of all n-tuples I = (i1, . . . , in) ∈ Zn. Let
9
[ϕI ]I (x) = 2− Pn
j=1 Qj ij ϕ(2−i1 x1, . . . , 2−in xn)
be the L1-invariant dilation of ϕI , and let
ΓZ(E) =nI = (i1, . . . , in) ∈ Zn : e(j, k)ik ≤ ij < 0o.
If
(1.15)
(1.16)
(1.17)
ZRCj
ϕI (x1, . . . , xj, . . . , xn) dxj = 0
for each I ∈ ΓZ(E) and for all 1 ≤ j ≤ n, then the infinite series PI∈ΓZ(E)[ϕI ]I converges
in the sense of distributions to a distribution K ∈ P0(E). More generally, we show that
the 'strong' cancellation condition given in (1.17) can be replaced by a variant of the weak
cancellation condition that was introduced in [NRSW12]. Roughly speaking7, we require that
if ϕI does not have cancellation in a variable xj , then there is a compensating gain in the size
of ϕI .
This more general result is then used in our study of convolutions of distributions in Section
9. In this study we also need a converse statement: if K ∈ P0(E), it can be written, modulo
distributions corresponding to coarser decompositions of RN , as a sumPI∈ΓZ(E)[ϕI ]I . This
is the subject of Section 6.
1.3.5 Section 7: Significance of the rank of E
In Section 7 we show that the class of classical Calder´on-Zygmund kernels corresponds
exactly to the case in which the rank of the matrix E is equal to 1. If the rank is greater than
1, distributions K ∈ P(E) differ from Calder´on-Zygmund kernels in two ways:
1) The distribution K is integrable at infinity, and so we can write K = K0 + K∞ with
K0, K∞ ∈ P(E), where K0 has support in the unit ball, and where K∞ is given by integra-
tion against a function K∞ ∈ C∞(RN ) ∩ L1(RN ). Thus for rank(E) ≥ 2, we focus on the
local behavior of distributions K ∈ P(E). We generally restrict our attention to the class
of proper distributions P0(E) ⊂ P(E) such that all derivatives are rapidly decreasing at
infinity.
and the rank of E is greater than 1, such weak type estimates may no longer be true.
We show, for example, that if rank(E) = n there exist distributions K ∈ P(E) such that
2) If K is a Calder´on-Zygmund kernel, then (cid:12)(cid:12)(cid:8)x ∈ RN : K(x) > λ(cid:9)(cid:12)(cid:12) . λ−1. If K ∈ P(E)
(cid:12)(cid:12)(cid:8)x ∈ RN : K(x) > λ(cid:9)(cid:12)(cid:12) ≈ λ−1(cid:2) log λ(cid:3)n−1
3) In general, there is a notion of 'reduced rank' which determines optimal bounds for esti-
.
mates of the sort considered in 2). This is discussed in Section 7.3.
7The precise statement is given in Definition 5.9.
10
1.3.6 Sections 8 and 9: Convolution operators on a Lie group
In Sections 8 and 9 we turn to the second part of our analysis dealing with convolution
operators TKf = f ∗ K on a homogeneous nilpotent Lie group G ∼= RN . Our two main results
are the following:
(A) If K ∈ P0(E) then each operator TK extends to a bounded operator on Lp(G) for 1 <
p < ∞.
(B) If K, L ∈ P0(E) then there exists M ∈ P0(E) such that TL ◦ TK = TM. Formally
M = K ∗ L, although the convolution of tempered distributions is not always defined.
Thus the space P0(E) is an algebra under composition (or convolution).
The first result follows rather easily from the observation that every distribution K ∈ P0(E)
is a flag kernel, since it was established in [NRSW12] that convolution with a flag kernel
extends to a bounded operator on Lp. The second result is analogous to the result in the
same paper that the space of kernels is closed under convolution. However this case is more
difficult and requires additional ideas.
We remark that both results are quite simple for Euclidean convolution, since dTKf (ξ) =
m(ξ)bf (ξ) where m is the Fourier transform of K and hence m ∈ M∞(E). It is not hard to
check that a multiplier m ∈ M∞(E) satisfies the conditions of the Marcinkiewicz multiplier
theorem, and the Lp-boundedness of the operator TK follows immediately. Since the product
of two such multipliers is again a multiplier of the same class, the result on composition of
operators also follows easily in this case.
To prove the results on a more general group, we rely very heavily on the decomposition
of distributions K ∈ P(E) as dyadic sums of dilates of bump functions. The heart of the
argument is then a careful analysis of sums of the formPI,J [ϕI ]I ∗ [ψJ ]J .
1.3.7 Section 10: Convolution of Calder´on-Zygmund kernels
In Section 10 we study the convolution of Calder´on-Zygmund kernels K1, . . . , Km with
compact support having different homogeneities. If the homogeneities are suitably adapted
to the decomposition on RN we show that there is an s×s matrix E so that the algebra P0(E)
contains each Kj . We also establish a kind of converse result: If E is an n×n matrix satisfying
the basic hypothesis given in (1.10), then there are finitely many different homogeneities on
RN so that P0(E) is the minimal algebra of the above type containing compactly supported
Calder´on-Zygmund kernels of each homogeneity.
1.3.8 Section 11: Regularity of two-flag kernels
In Section 11 we return to the study of kernels belonging to two opposite flags, and show
that these belong to P0(E) for an appropriate matrix E. In the example studied in section
1.2, where the kernel satisfied the two estimates
x∂b
y∂c
(cid:12)(cid:12)∂a
zK(x, y, z)(cid:12)(cid:12) ≤ Ca,b,c
x−1−a(x2 + y)−1−b(x3 + y3/2 + z)−1−c
(x + y2 + z3)−1−a(y + z 3
2 )−1−bz−1−c
,
it turns out that K ∈ P0(E) where
E =
1
2
3
1
1
3
2
1
1
1 and
11
N1(x, y, z) = x + y + z,
N2(x, y, z) = x2 + y + z,
N3(x, y, z) = x3 + y3/2 + z
.
1.3.9 Sections 12 and 13: Operators with variable coefficients
In Sections 12 and 13 we describe a wider class of operators with "variable coefficients".
An operator T belongs to this extended class if it is of the form
T [f ](x) =ZG
K(x, z)f (x · z−1) dz
(1.18)
where for each x, the distribution K(x, ·) is a kernel of the type we have been considering
that depends in an appropriately smooth way on x. One can then assert:
a) Operators of that type form an algebra, and each is bounded on Lp(G) for 1 < p < ∞.
b) The standard pseudo-differential operators of order 0 belong to this algebra, and the sub-
algebra of these operators is "central" in the sense that an operator of the extended algebra
commutes with a standard pseudo-differential operator of order 0, up to an error which is
an appropriate "smoothing operator."
2 The Classes P(E) and M(E)
In this section we define a class of distributions P(E) and the corresponding class of Fourier
multipliers M(E) on the space RN and its dual. We begin by introducing notation that will
be used throughout the paper.
2.1 Notation
2.1.1 Decompositions of RN
We begin with an initial decomposition of RN into direct sums of subspaces where only some
of the coordinates x1, . . . , xN are non-zero. Thus if C ⊂ {1, . . . , N } is non-empty, set
RC =nx = (x1, . . . , xN ) ∈ RN : xj = 0 for all j /∈ Co ⊂ RN .
(2.1)
We will abuse of this notation by allowing RC to also indicate the set of c-tuples (c being
the cardinality of C) with entries indexed by elements of C. We assume that an initial
decomposition of RN is given,
RN = RC1 ⊕ · · · ⊕ RCn,
(2.2)
where C1, . . . , Cn ⊂ {1, . . . , N } are disjoint non-empty subsets with Sn
j=1 Cj = {1, . . . , N }.
Given the decomposition (2.2), if x = (x1, . . . , xN ) ∈ RN we write x = (x1, . . . , xn) where
xj = {xl : l ∈ Cj}.
12
Unfortunately there are further notational complications. We shall sometimes need to
consider the direct sum of only a subset of the subspaces RC1, . . . , RCn , and we will also need to
partition the set {C1, . . . , Cn} into a union of disjoint subsets and consider the corresponding
coarser decomposition of RN . We shall then use the following notation.
(a) If L = {l1, . . . , lr} ⊂ {1, . . . , n}, write
An element of RL can then be written xL = {xk : k ∈ L}.
RL = RCl1 ⊕ · · · ⊕ RClr =nx ∈ RN : xj = 0 for all j /∈
r[k=1
(b) Let I1, . . . , Is ⊂ {1, . . . , n} be disjoint, non-empty subsets with Ss
RCk =nx ∈ RN : xj = 0 for all j /∈ [k∈Ir
RIr = Mk∈Ir
Write
We then have the coarser decomposition RN = RI1 ⊕ · · · ⊕ RIs.
x = (xI1 , . . . , xIs ) where xIr ∈ RIr .
Ck}.
Clko.
(2.3)
r=1 Ir = {1, . . . , n}.
(2.4)
If x ∈ RN we write
Note that we have slightly abused notation since the symbols RL and RIr in equations (2.3)
and (2.4) have a different meaning than the symbol RC in eqwuation (2.1). However C is a
subset of {1, . . . , N } while L and Ir are subsets of {1, . . . , n}.
2.1.2 Dilations on RN
The underlying space RN is equipped with a family of dilations given by
λ · x = λ · (x1, . . . , xN ) =(cid:0)λd1x1, . . . , λdN xN(cid:1)
where each dj > 0. We can restrict these dilations in equation (2.5) to each subspace RCj
and write λ · xj =(cid:8)λdlxl : l ∈ Cj(cid:9). The homogeneous dimension of the subspace RCj is then
Qj = Pl∈Cj
dl. Let nj : RCj → [0, ∞) be a smooth homogeneous norm for this family of
dilations so that
(2.5)
(2.6)
nj(xj ) ≈ Xl∈Cj
xl1/dl.
Thus nj(λ · xj) = λnj(xj ) for λ > 0. Put
B(ρ) =nx ∈ RN : nj(xj) < ρ, 1 ≤ j ≤ no.
We use standard multi-index notation for derivatives.
∂γ
(γ 1, . . . , γ n) ∈ NC1 ×· · ·×NCn. When estimating derivatives ∂γ
norms on each subspace RCj we need to use an appropriately weighted length. Thus put
If γ = (γ1, . . . , γN ) ∈ NN then
l=1 γl. Using the decomposition (2.2), we can also write γ =
x f (x) in terms of homogeneous
x = QN
l=1 ∂γl
xl and γ = PN
[[γj]] = Xl∈Cj
γldl
and
[[γ]] =
[[γj]] =
nXj=1
γldl.
NXl=1
(2.7)
The space of infinitely differentiable real-valued functions on RN with compact support
0 (RN ) and the space of Schwartz functions is denoted by S(RN ). The basic
is denoted by C∞
semi-norms on these spaces are given by
ϕ(m) = sup
ψM = sup
x ϕ(x)(cid:12)(cid:12) : γ ≤ mo
x∈RNn(cid:12)(cid:12)∂γ
x∈RN(cid:8)(cid:12)(cid:12)(1 + x)α∂β
x ψ(x)(cid:12)(cid:12) : α + β ≤ M(cid:9) for ψ ∈ S(RN ).
for ϕ ∈ C∞
0 (RN ),
13
Let A be an index set. We use the following terminology from Section 4 in [NRSW12].
1) If(cid:8)ϕα : α ∈ A(cid:9) ⊂ C∞
0 (RN ), then the family {ϕα} is normalized in terms of a function Φ ∈
C∞
0 (RN ) if there are constants C, Cm > 0 and integers pm with the following properties.
a) If the support of Φ is contained in the ball B(ρ) then the support of each ϕα is contained
in the ball B(Cρ).
b) For every m ∈ N and every α ∈ A, ϕα(m) ≤ CmΦ(m+pm).
2) A family {ϕα : α ∈ A} ⊂ C∞
0 (RN ) is uniformly bounded if all the members are supported
in a fixed ball B(ρ) and if there are constants Cm > 0 so that ϕα(m) ≤ Cm for all α ∈ A.
3) If {ψα : α ∈ A} ⊂ S(RN ), then the family {ψα} is normalized in terms of a function
Ψ ∈ S(RN ) if there are constants CM > 0 and integers pM ∈ N so that ψα[M] ≤
CM Ψ[M+pM ] for every M ∈ N.
4) A family {ψα : α ∈ A} ⊂ S(RN ) is uniformly bounded if there are constants CM > 0 so
that ψαM ≤ CM for all α ∈ A.
Given the decomposition RN = RC1 ⊕ · · · ⊕ RCn, define an n-parameter family of dilations
If (λ1, . . . , λn) = (2−t1, . . . , 2−tn) with
by setting (λ1, . . . , λn) · x = (λ1 · x1, . . . , λn · xn).
t = (t1, . . . , tn) ∈ Rn then for f ∈ L1(RN ), we define
(f )t(x) = f (2−t1 · x1, . . . , 2−tn · xn),
[f ]t(x) = 2− Pn
j=1 tj Qj f (2−t1 · x1, . . . , 2−tn · xn).
(2.8)
Note that (f )tL∞(RN ) = f L∞(RN ) while [f ]tL1(RN ) = f L1(RN ).
2.2 Global norms
The classes P(E) and M(E) of distributions and multipliers are defined using families of
norms {N1, . . . , Nn} and dual norms {bN1, . . . , bNn}. These in turn are defined in terms of an
n × n matrix E = {e(j, k)}, where each entry e(j, k) ∈ (0, ∞). For 1 ≤ j ≤ n set
Nj(t1, . . . , tn) = n1(t1)e(j,1) + · · · + nn(tn)e(j,n) ≈
nXk=1Xl∈Ck
nXk=1Xl∈Ck
tle(j,k)/dl ,
tl1/e(k,j)dl .
(2.9)
bNj(t1, . . . , tn) = n1(t1)
1
e(1,j) + · · · + nn(tn)
1
e(n,j) ≈
The entries of E are usually subject to certain natural constraints.8
Definition 2.1. The matrix E satisfies the basic hypothesis if :
e(j, j) = 1,
e(j, l) ≤ e(j, k)e(k, l)
for 1 ≤ j ≤ n,
for 1 ≤ j, k, l ≤ n.
(2.10)
In particular, 1 ≤ e(j, k)e(k, j) for 1 ≤ j, k ≤ n.
8The motivation for these constraints is discussed in Section 14 below.
14
We will need to restrict the norms {Nj} to certain subspaces of RN . Thus if L =
{l1, . . . , ls} ⊂ {1, . . . , n}, we use the notation introduced on page 12 and set RL =Ll∈L RCl.
For tL = (tl1 , . . . , tls) ∈ RL, and for each l ∈ L, put
Nl(tL) = Xm∈L
nm(tm)e(l,m).
If R = {R1, . . . , Rs} are positive real numbers and if ψ ∈ C∞
setting
0 (RL), define ψR ∈ C∞
0 (RL) by
ψR(tL) = ψ(R1 · tl1, . . . , Rs · tls).
2.3 Classes of distributions and multipliers
Let RN = RC1 ⊕ · · · ⊕ RCn , let E = {e(j, k)} be an n × n matrix satisfying (2.10), and let
{N1, . . . , Nn} and {bN1, . . . , bNn} be the norms and dual defined in (2.9).
Definition 2.2. A distribution K on RN belongs to the class P(E) if it satisfies the following
differential inequalities and cancellation conditions.
(a) [Differential Inequalities] Away from the origin the distribution K is given by integration
against a C∞-function K, and for every γ = (γ 1, . . . , γ n) ∈ NC1 × · · · × NCn , there is a
constant Cγ so that
∂γK(x) ≤ Cγ
Nj(x)−(Qj +[[γ j ]]).
nYj=1
(b) [Cancellation Conditions] Let L = {l1, . . . , lr} and M = {m1, . . . , ms} be any pair of
complementary subsets of {1, . . . , n}, let R = {R1, . . . , Rs} be any positive real numbers,
and let ψ ∈ C∞
0 (RM ) have support in the unit ball. Define a distribution Kψ,R on RL by
setting
(cid:10)Kψ,R, ϕ(cid:11) =(cid:10)K, ϕ ⊗ ψR(cid:11)
for every ϕ ∈ C∞
0 (RL). Then Kψ,R uniformly satisfies the analogue of the estimates
in (a) on the space RL: precisely, this means that away from the origin of RL the dis-
tribution Kψ,R is given by integration against a smooth function Kψ,R and for every
γ = (γ l1, . . . , γlr ) ∈ NCl1 × · · · × NClr there is a constant C′
γ depending only on the
constants {Cγ} in (a) and the norms {ψ(m)} (but independent of R) so that
(cid:12)(cid:12)∂γKψ,R(x)(cid:12)(cid:12) ≤ C′
γ
rYt=1
Nlt(xL)−(Qlt +[[γ lt
]]).
In particular, if L = ∅ then (cid:10)K, ψR(cid:11) is bounded independently of R.
product kernel on RN = Ln
Remark 2.3. Since Nj(x)−[[γ j ]] ≤ xj −[[γ j ]], it follows that a distribution K ∈ P(E) is a
j=1 RCj in the sense of [NRS01]. (The cancellation conditions
follow from the Fourier transform characterization. See [NRS01], Remark 2.16.) However,
the distributions in P(E) are in general more regular; they are singular only at the origin.
Definition 2.4. If m ∈ C∞(RN \ {0}) then m belongs to the class of multipliers M(E) if for
every γ = (γ1, . . . , γn) ∈ NC1 × · · · × NCn there is a constant Cγ so that
∂γm(ξ) ≤ Cγ
nYj=1 bNj(ξ)−[[γ j ]].
15
We shall see in Lemma 7.3 that unless e(j, k) = e(j, l)e(l, k) for all j, k, l ∈ {1, . . . , n},
the distribution K ∈ P(E) is integrable at infinity. We are primarily interested in the local
behavior of these kernels and the following restricted classes of distributions and multipliers.
Definition 2.5. Let x denote the usual Euclidean norm of a vector x ∈ RN .
(a) A distribution K on RN belongs to the class P0(E) if it belongs to the class P(E) and away
from the origin it is given by a smooth function K satisfying the differential inequalities
(cid:12)(cid:12)∂γ
x K(x)(cid:12)(cid:12) ≤ CM,γ
nYj=1
Nj(x)(Qj +[[γ j ]])(cid:0)1 + x(cid:1)−M
(2.11)
for every multi-index γ = (γ 1, . . . , γ n) ∈ NC1 × · · · × NCn and every M ≥ 0 . Moreover
the quotient distributions Kψ,R defined in part (b) of Definition 2.2 satisfy
(cid:12)(cid:12)∂γKψ,R(x)(cid:12)(cid:12) ≤ DM,γ(cid:0)1 + x(cid:1)−M
N L
lt (xL)−(Qlt +[[γ lt
]]).
(2.12)
rYt=1
(b) A function m ∈ C∞(RN ) belongs to M∞(E) if there is a constant Cγ so that
(cid:12)(cid:12)(cid:12)∂γ
ξ m(ξ)(cid:12)(cid:12)(cid:12) ≤ Cγ
nYj=1(cid:0)1 + bNj(ξ)(cid:1)−[[γ j ]]
for every multi-index γ = (γ 1, . . . , γn) ∈ NC1 × · · · × NCn.
We will see in Section 4 below that the multipliers m ∈ M∞(E) are precisely the Fourier
transforms of the distributions K ∈ P0(E), and conversely, that distributions K ∈ P0(E) are
the inverse Fourier transforms of multipliers m ∈ M∞(E). The family of norms
KM = Xγ,γ ′≤M
infnCM,γ + DM,γ ′ : (2.11) and (2.12) holdo
(2.13)
with M ∈ N induces a Fr´echet space topology on the space P0(E).
3 Marked partitions and decompositions of RN
3.1 Dominant terms in Nj(t) and bNj(t)
RN into regions where the norms Nj(t) =Pn
To study the kernels and multipliers introduced in Section 2, we need to partition the space
k=1 nk(tk)1/e(k,j)
are comparable to a single term nk(tk)e(j,k) or nk(tk)1/e(k,j). We begin by introducing the
notion of strict dominance and A-dominance.
k=1 nk(tk)e(j,k) or bNj(t) =Pn
Definition 3.1. Let A ≥ 1.
every l 6= k. The term tk is strictly dominant in Nj(t) if we can take A = 1.
(a) Let t ∈ B(1). The term tk is A-dominant in Nj(t) if nl(tl)e(j,l) < (cid:2)A nk(tk)(cid:3)e(j,k)
(b) Let t ∈ B(1)c. The term tk is A-dominant in bNj(t) if nl(tl)1/e(l,j) <(cid:2)A nk(tk)(cid:3)1/e(k,j)
for every l 6= k. The term tk is strictly dominant in bNj(t) if we can take A = 1.
for
16
Remarks 3.2.
1) If t ∈ B(1) and if tk is A-dominant in Nj(t) then Nj(t) ≈ nk(tk)e(j,k) where the implied
constant depends only on A, E, and n.
constant depends only on A, E, and n.
2) If t ∈ B(1)c and if tk is A-dominant in bNj(t) then bNj(t) ≈ nk(t)1/e(k,j) where the implied
3) If t ∈ B(1)c then each bNj(t) ≥ 1.
4) If t ∈ B(1)c and if tk is A-dominant in any bNj(t) then Ank(tk) > 1.
The following is a simple but key fact about dominance.
Proposition 3.3. Suppose the matrix E satisfies the basic hypothesis (2.10).
(a) Let t ∈ B(1). If tk is A-dominant in Nj(t) then tk is also A-dominant in Nk(t). In
particular, if tk is strictly dominant in Nj(t), then tk is strictly dominant in Nk(t).
(b) Let t ∈ B(1)c. If tk is A-dominant in bNj(t) then tk is also A-dominant in bNk(t). In
particular if tk is strictly dominant in bNj(t) then tk is also strictly dominant in bNk(t).
l 6= k, then nl(tl)e(j,l) <(cid:2)A nk(tk)(cid:3)e(j,k)
Proof. Let t ∈ B(1) and A ≥ 1. By the basic hypothesis, e(k, l) ≥ e(j, l)/e(j, k), and since
nl(tl) ≤ 1 it follows that nl(tl)e(k,l) ≤ nl(tl)e(j,l)/e(j,k). If tk is A-dominant in Nj(t) and if
, and so
nl(tl)e(k,l) ≤ nl(tl)e(j,l)/e(j,k) < A nk(tk).
Thus tk is A-dominant in Nk(t). On the other hand, if t ∈ B(1)c and tk is A-dominant in
bNj(t), then every l 6= k we have nl(tl)1/e(l,j) < [Ank(tk)]1/e(k,j). Since e(l, j) ≤ e(l, k)e(k, j)
and Ank(tk) > 1 we have
nl(tl)1/e(l,k) =hnl(tl)1/e(l,j)ie(l,j)/e(l,k)
<hAnk(tk)ie(l,j)/e(l,k)e(k,j)
≤ A nk(tk).
depends on the notion of a marked partition.
The notion of strict dominance allows us to decompose (up to a closed set of measure
Thus tk is A-dominant in bNk(t).
3.2 Marked partitions and the sets ES and bES
zero) the ball B(1) =(cid:8)t ∈ RN : nj(tj) ≤ 1(cid:9) into sets where one term in each Nj is strictly
dominant. The same is true for the complement B(1)c and the norms bNj. The decomposition
(a) A marked partition S = (cid:8)(I1, k1); . . . ; (Is; ks)(cid:9) of the set {1, . . . , n} is a collection
{I1, . . . , Is} of disjoint non-empty subsets of {1, . . . , n} such that Ss
(b) The partition S0 =(cid:8)({1}, 1); . . . ; ({n}, n)(cid:9) is called the principal partition of {1, . . . , n}.
(c) Any marked partition S =(cid:8)(I1, k1); . . . ; (Is; ks)(cid:9) with s < n is called a non-principal
together with a choice of a particular element kr ∈ Ir in each subset.
r=1 Ir = {1, . . . , n},
Definition 3.4.
partition.
(d) The set of all marked partitions of {1, . . . , n} is denoted by S(n).
17
We want to associate a unique marked partition to every point t ∈ RN , but to do this we
can be equal. Let
The set Ξ is closed and has measure zero, and RN \ Ξ is dense in RN . Note that if t /∈ Ξ, all
need to exclude the closed set of measure zero where two different terms in Nj(t) or bNj(t)
Ξ =nt ∈ RN :(cid:0)∃j(cid:1)(cid:0)∃l 6= m(cid:1) nl(tl)e(j,l) = nm(tm)e(j,m) or nl(tl)1/e(l,j) = nm(tm)1/e(m,j)o.
the summands in Nj(t) and bNj(t) are distinct.
cases, depending on whether the point t belongs to B(1) or to B(1)c.
Case 1: If t ∈ B(1) \ Ξ then for every j the terms n1(t1)e(j,1), . . . , nn(tn)e(j,n) in Nj(t) are
distinct. Thus for each j ∈ {1, . . . , n} there exists a unique integer k(j) ∈ {1, . . . , n} such
that tk(j) is strictly dominant in Nj(t). Let {k1, . . . , ks} be the set of distinct integers k(j)
which arise in this way. For 1 ≤ r ≤ s set
We associate to every point t ∈ RN \ Ξ a marked partition S(t) ∈ S(n). There are two
Ir =nj ∈ {1, . . . , n} : tkr is strictly dominant in Nj(t)o.
Every j belongs to a unique Ir since there is only one term in Nj(t) which is dominant, and
We have kr ∈ Ir by Proposition 3.3, and thus S = S(t) = {(I1, k1); . . . ; (Is, ks)} ∈ S(n).
and again for each j there exists a unique integer k(j) ∈ {1, . . . , n} such that tk(j) is strictly
according to Proposition 3.3, we have kr ∈ Ir. Thus S(t) =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S(n).
Case 2: If t ∈ B(1)c \ Ξ, the terms n1(t1)1/(1,j), . . . , nn(tn)1/e(n,j) in bNj(t) are distinct,
dominant in bNj(t). Let {k1, . . . , ks} be the set of distinct integers that arise, and again set
Ir =nj ∈ {1, . . . , n} : tkr is strictly dominant in bNj(t)o.
Definition 3.5. Let S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S(n), and let A ≥ 1. Then
tkr is A-dominant in Nj(t)(cid:1)o
S =nt = (t1, . . . , tn) ∈ B(1) : ∀r ∈ {1, . . . , s}, ∀j ∈ Ir,
=nt ∈ B(1) : ∀r ∈ {1, . . . , s}, ∀j ∈ Ir, ∀l 6= kr, nl(tl)e(j,l) <(cid:2)A nkr (tkr )(cid:3)e(j,kr ) o.
S =nt = (t1, . . . , tn) ∈ B(1)c : ∀r ∈ {1, . . . , s}, ∀j ∈ Ir, tkr is A-dominant in bNj(t)(cid:1)o
bEA
=nt ∈ B(1)c : ∀r ∈ {1, . . . , s}, ∀j ∈ Ir, ∀l 6= kr, nl(tl)1/e(l,j) <(cid:2)A nkr (tkr )(cid:3)1/e(kr ,j)o.
EA
When A = 1 we write E1
If A > 1 then
Proposition 3.6. The set EA
ES,
EA
S ,
S are open. Moreover,
S = bES.
S = ES and bE1
S and bEA
B(1) \ Ξ = [S∈S(n)
B(1) ⊂ [S∈S(n)
B(1)c \ Ξ = [S∈S(n) bES.
B(1)c ⊂ [S∈S(n)bEA
S is an open neighborhood of the closure of bES.
S .
since the closure of RN \ Ξ is all of RN , and if A > 1 then EA
Proof. The first line follows from the construction of ES and bES. The second line follows
closure of ES, and bEA
S is an open neighborhood of the
18
3.3 Characterizing the sets EA
S
S and bEA
S involve
inequalities between the norms nj(tj ) and nkr (tkr ) for all r ∈ {1, . . . , s} and all j 6= kr.
In this section we show that the sets are characterized by a smaller number of inequalities
relating, for given r, nkr (tkr ) to nj(tj ) only for j ∈ Ir ∪ {k1, . . . , ks}.
Let S = (cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S(n). The definitions of the sets EA
S and bEA
Definition 3.7. Let(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S(n). For 1 ≤ r ≤ s and 1 ≤ l ≤ n set
σS(kr, l) = min
j∈Ir
e(j, l)
e(j, kr)
,
τS(l, kr) = min
j∈Ir
e(l, j)
e(kr, j)
.
Proposition 3.8. Suppose that the matrix E = {e(j, k)} satisfies the basic hypothesis (2.10)
and that S =(cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) ∈ S(n).
(a) If 1 ≤ r ≤ s and if l ∈ Ir then
σS(kr, l) = e(l, kr)−1,
τS(l, kr) = e(kr, l)−1.
(b) Let 1 ≤ r, p ≤ s. If m ∈ Ip then
σS(kr, m) ≥ σS(kr, kp)σS(kp, m),
τS(m, kr) ≥ τS(m, kp)τS(kp, kr).
Proof. It suffices to prove the statements about σS; the statements about τS follow in the same
way. If l ∈ Ir we can take j = l in the definition of σS(kr, l) to see that σS(kr, l) ≤ e(l, kr)−1.
On the other hand it follows from the basic assumption (2.10) that e(j, l)/e(j, kr) ≥ e(l, kr)−1
for any l, and this gives the opposite inequality and establishes the first equality in (a).
Next let m ∈ Ip. Then using (a) and the basic hypothesis, if j0 ∈ Ir
σS(kp, m)σS(kr, kp) = e(m, kp)−1σS(kr, kp) = e(m, kp)−1 min
j∈Ir
e(j, kp)
e(j, kr)
≤ e(m, kp)−1 e(j0, kp)
e(j0, kr)
≤
e(j0, m)
e(j0, kr)
.
Taking the minimum over j0 ∈ Ir it follows that σS(kr, m) ≥ σS(kr, kp)σS(kp, m), which
establishes the first inequality in (b) and completes the proof.
Lemma 3.9. Let S =(cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) ∈ S(n).
(a) Let t ∈ B(1). If t ∈ EA
S then for every r ∈ {1, . . . , s}
nkr (tkr ) > A−1nj(tj)1/e(j,kr ) = A−1nj(tj)σS (kr,j) ∀j ∈ Ir, j 6= kr,
nkr (tkr ) > A−1nkp (tkp )σS (kr ,kp)
∀p ∈ {1, . . . , s}, p 6= r.
and
(3.1)
Conversely, if t ∈ B(1) satisfies (3.1) then t ∈ EAη
only on the matrix E.
S where η is a constant that depends
(b) Let t ∈ B(1)c. If t ∈ bEA
S then for every r ∈ {1, . . . , s}
nkr (tkr ) > A−1nj(tj)e(kr ,j) = A−1nj(tj)1/τS (j,kr ) ∀j ∈ Ir, j 6= kr,
nkr (tkr ) > A−1nkp (tkp )1/τS(kp,kr)
∀p ∈ {1, . . . , s}, p 6= r.
and
(3.2)
Conversely, if t ∈ B(1)c satisfies (3.2) then t ∈ EAη
only on the matrix E.
S where η is a constant that depends
19
Letting A = Aη = 1, we obtain the following characterization of the sets ES and bES.
Corollary 3.10. Let S =(cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) ∈ S(n).
(a) If t ∈ B(1) then t ∈ ES if and only if
nkr (tkr ) > nj(tj)1/e(j,kr ) = nj(tj )σS (kr ,j)
nkr (tkr ) > nkp (tkp )σS (kr ,kp)
for every j ∈ Ir with j 6= kr, and
for every p ∈ {1, . . . , s} with p 6= r.
(b) If t ∈ B(1)c then t ∈ bES if and only if
nkr (tkr ) > nj(tj )e(kr ,j) = nj(tj)1/τS (j,kr )
nkr (tkr ) > nkp (tkp )1/τS (kp,kr )
Proof of Lemma 3.9.
for every j ∈ Ir with j 6= kr, and
for every p ∈ {1, . . . , s} with p 6= r.
If t ∈ B(1) then t ∈ EA
S if and only if nkr (tkr ) > A−1nl(tl)e(j,l)/e(j,kr ) for 1 ≤ r ≤ s, for
l 6= kr, and for all j ∈ Ir. But since nl(tl) ≤ 1 this is equivalent to saying that
nkr (tkr ) > A−1 max
j∈Ir
nl(tl)e(j,l)/e(j,kr ) = A−1nl(tl)minj∈Ir e(j,l)/e(j,kr ) = A−1nl(tl)σS (kr ,l),
and this establishes both inequalities in (3.1). To prove the converse, let l ∈ {1, . . . , n} with
l 6= kr. Then there exists an index p ∈ {1, . . . , s} so that l ∈ Ip, and by the first inequality
in (3.1) it follows that nl(tl)σS (kp,l) < A nkp (tkp ). On the other hand, the second inequality
in (3.1) says that nkp (tkp )σS (kr,kp) ≤ A nkr (tkr ). Also, Proposition 3.8, part (b) says that
σS(kr, l) ≥ σS(kr, kp)σS(kp, l). Since nl(tl) < 1, it follows that
< A1+σS (kr,kp)nkr (tkr ).
where we take η = 1 + supr,p σS(kr, kp) + supr,p τS(kp, kr)−1. This shows that tkr is Aη-
dominant in Nj(t), and completes the proof of (a).
S if and only if
nkr (tkr ) > A−1 nl(tl)e(kr ,j)/e(l,j) for 1 ≤ r ≤ s, for l 6= kr, and for all j ∈ Ir. Suppose first
that nl(tl) ≥ 1. In this case it follows that
The proof of (b) is very similar. Suppose that t ∈ B(1)c. Then t ∈ bEA
nkr (tkr ) > A−1 max
j∈Ir
nl(tl)e(kr ,j)/e(l,j)
= A−1nl(tl)maxj∈Ir e(kr ,j)/e(l,j) = A−1nl(tl)1/τS (l,kr ).
On the other hand, if nl(tl) < 1 then we still have nkr (tkr ) > A−1nl(tl)1/τS (l,kr) since
nkr (tkr ) > A−1 by Remarks 3.2. In either case, this establishes both inequalities of (3.2). To
prove the converse, let l ∈ {1, . . . , n} with l 6= kr. If nl(tl) < 1, then since nkr (tkr ) > A−1 it
. If nl(tl) ≥ 1. There exists p ∈ {1, . . . , s}
so that l ∈ Ip, and by the first inequality in (3.2) it follows that nl(tl)1/τS (l,kp) < A nkp (tkp ).
On the other hand, the second inequality in (3.2) says that nkp (tkp ) ≤ AτS (kp,kr ) nkr (tkr ).
Also, Proposition 3.8, part (b) says that τS(l, kr) ≥ τS(l, kp)τS(kp, kr). It follows that
follows that nl(tl)1/e(l,j) < 1 <(cid:2)A nkr (tkr )(cid:3)1/e(kr ,j)
nl(tl)1/τS (l,kr ) ≤ nl(tl)1/τS (l,kp)τS(kp,kr ) <(cid:2)A nkp (tkp )(cid:3)1/τS (kp,kr)
< A1+1/τS (kp,kr)nkr (tkr ).
But then if j ∈ Ir and l 6= kr,
nl(tl)σS (kr ,l) ≤ nl(tl)σS (kp,l)σS (kr ,kp) <(cid:2)A nkp (tkt )(cid:3)σS (kr,kp)
nl(tl)e(j,l) =hnl(tl)e(j,l)/e(j,kr )ie(j,kr )
≤ nl(tl)σS (kr ,l)e(j,kr ) <hA(1+σS (kr ,kp))nkr (tkr )ie(j,kr )
≤hAηnkr (tkr )ie(j,kr )
20
If j ∈ Ir and l 6= kr,
nl(tl)1/e(l,j) ≤(cid:2)nl(tl)1/τS (l,kr)(cid:3)1/e(kr ,j)
<(cid:2)A1+1/τS (kp,kr )nkr (tkr )(cid:3)1/e(kr ,j)
.
, and this says that tkr is
Thus in all cases we have nl(tl)1/e(kr ,l) < (cid:2)Aη nkr (tkr )(cid:3)1/e(kr ,j)
Aη-dominant in bNj(t), which completes the proof of (b) and of Lemma 3.9.
3.4 Estimates of kernels and multipliers on ES and bES
ES and bES. First consider the principal marked partition S0 = (cid:8)({1}, 1); . . . ; ({n}, n)(cid:9). If
x ∈ B(1) ∩ ES0 then Nk(x) ≈ nk(xk), and if ξ ∈ bES0 then bNk(ξ) ≈ nk(ξk). Thus the
and bES0 :
differential inequalities for K ∈ P(E) and m ∈ M(E) take the following form on the sets ES0
The differential inequalities for kernels and multipliers take simpler forms on the sets
nk(xk)−(Qk+[[γ k]]),
x ∈ ES0 =⇒ (cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ Cγ
ξ ∈ bES0 =⇒ (cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ Cγ
nYk=1
nYk=1
nk(ξk)−[[γ k]].
j=1 RCj .
Note that these are precisely the differential inequalities satisfied by product kernels or product
(See [NRS01], page 34 and page 37.) Thus although a
distribution K ∈ P(E) is pseudo-local, if the set ES0 6= ∅ the function K satisfies nothing
better than the estimates for a product kernel on this set.
multipliers on RN = Ln
We have the following extension for non-empty sets ES and bES.
Proposition 3.11. Let S =(cid:8)(I1, k1); . . . ; (Is; ks)(cid:9) ∈ S(n).
(a) If ES ∩B(1) 6= ∅, if K ∈ P(E), and if K ∈ C∞(RN \{0}) is the associated smooth function,
then for every γ = (γ 1, . . . , γn), there is a constant Cγ so that for x ∈ ES ∩ B(1) the
inequality in Definition 2.2, part (a) is equivalent to
∂γK(x) ≤ Cγ
sYr=1
nkr (xkr )− Pj∈Ir
e(j,kr )(Qj +[[γ j ]]).
(b) If bES ∩ B(1)c 6= ∅ and if m ∈ M(E) then for every γ = (γ1, . . . , γn) there is a constant
Cγ so that for ξ ∈ bES the inequality in Definition 2.4 is equivalent to
∂γm(ξ) ≤ Cγ
nkr (ξkr )− Pj∈Ir
e(j,kr )[[γ j ]].
sYr=1
Proof. If x ∈ ES, then Nj(x) ≈ nkr (xkr )e(j,kr ) for every j ∈ Ir. Thus if x ∈ ES
∂γK(x) ≤ Cγ
nYj=1
Nj(x)−(Qj +[[γ j ]]) ≈ Cγ
sYr=1Yj∈Ir
nkr (xkr )−e(j,kr )(Qj +[[γ j ]]).
This establishes part (a), and again part (b) is done the same way.
3.5 A coarser decomposition of RN associated to S ∈ S(n)
21
Proposition 3.11 suggests that, on each non-empty set ES, a kernel K ∈ P0(E) satisfies
the same differential inequalities of a product kernel adapted to a coarser decomposition of
RN depending on S and, in a similar way, that on each non-empty set bES, a multiplier
For each 1 ≤ r ≤ s we gather the coordinates {tj : j ∈ Ir} into a single subspace as
m ∈ M∞(E) behaves like a product multiplier.
discussed on page 12. Thus for 1 ≤ r ≤ s set
RIr =Mj∈Ir
RCj
(3.3)
so that RN = RI1 ⊕ · · · ⊕ RIs. Note that if s < n, this decomposition is strictly coarser than
the original decomposition RN = RC1 ⊕ · · · ⊕ RCn. If t ∈ RN write t = (tI1 , . . . , tIS ) where
Next we introduce two new families of dilations xIr → λ ·S xIr and ξIr → λ·S ξIr on each
component RIr . These should have the property that if x ∈ ES then the associated norm of
tIr =(cid:8)tj : j ∈ Ir(cid:9).
xIr should be comparable to nkr (xkr ), and if ξ ∈ bES, then the associated norm of ξIr should
be comparable to nkr (ξkr ). Recall from Corollary 3.10 that
x ∈ ES =⇒ nj(xj )1/e(j,kr ) < nkr (xr)
nj(ξj) e(kr ,j) < nkr (ξkr ),
ξ ∈ bES =⇒
and from Proposition 3.8 that if j ∈ Ir then e(j, kr) = σS(kr, j)−1 and e(kr, j) = τ (j, kr)−1.
(3.4)
(3.5)
(3.6)
Definition 3.12. Let S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S(n).
(a) If x = (xI1 , . . . , xIs ) ∈ RN with xIr =(cid:8)xj : j ∈ Ir} ∈ RIr set
λ ·S xIr =(cid:16)λe(j,kr ) · xj : j ∈ Ir(cid:17) =(cid:16)λ1/σS (kr,j) · xj : j ∈ Ir(cid:17).
Let nS,r be a smooth homogeneous norm on RIr for this dilation.
(b) If ξ = (ξI1 , . . . , ξIs ) ∈ RN with ξIr =(cid:8)ξj : j ∈ Ir(cid:9) let
λ·S ξIr =(cid:16)λ1/e(kr ,j) · ξj : j ∈ Ir(cid:17) =(cid:16)λτS (j,kr ) · ξj : j ∈ Ir(cid:17).
Let bnS,r be a smooth homogeneous norm on RIr for this dilation.
Observe that
nS,r(xIr ) ≈ Xj∈Ir
bnS,r(ξIr ) ≈ Xj∈Ir
QS,r = Xj∈Ir
bQS,r = Xj∈Ir
nj(xj )1/e(j,kr ) = Xj∈Ir
nj(ξj)e(kr ,j) = Xj∈Ir
Qje(j, kr) = Xj∈Ir
Qje(kr, j)−1 = Xj∈Ir
nj(xj )σS (kr ,j),
nj(ξj)1/τS (j,kr),
QjσS(kr, j)−1,
QjτS(j, kr).
and that the homogeneous dimensions of RIr under the dilations λ·S and λ·S are given by
22
Moreover, it follows directly from and (2.9) and (3.5) that, for 1 ≤ r ≤ s,
for all t ∈ RN .
nS,r(tIr ) . bNkr (t),
bnS,r(tIr ) . Nkr (t),
(3.7)
4 Fourier transform duality of kernels and multipliers
In this section we show that the Fourier transforms of kernels K ∈ P0(E) are multipliers
m ∈ M∞(E), and conversely, that the inverse Fourier transform of multipliers m ∈ M∞(E)
are kernels K ∈ P0(E). Such results are well known for Calder´on-Zygmund kernels and
multipliers, and more generally, for product kernels and product multipliers or for flag kernels
and flag multipliers. (See [NRS01] for details.)
4.1 Fourier transforms of multipliers
Let m ∈ L∞(RN ) and let K = bm be its Fourier transform in the sense of tempered
distributions. Thus if ϕ ∈ S(RN ) we have (cid:10)K, ϕ(cid:11) = (cid:10)m,bϕ(cid:11) = RRN m(ξ)bϕ(ξ) dξ. Let χ ∈
C∞
0 (RN ) be identically equal to one in a neighborhood of the origin, and let χǫ(ξ) = χ(ǫξ).
Then
(cid:10)K, ϕ(cid:11) =ZRN
ǫ→0ZRN
m(ξ)bϕ(ξ) dξ = lim
χǫ(ξ)m(ξ)bϕ(ξ) dξ = lim
ǫ→0(cid:10)dχǫm, ϕ(cid:11)
to the class P0(E).
Proof. Assume as discussed above that m has compact support. The Fourier transform is
by parts) provided that we do not use any information about the size of this support in our
estimates.
so K = limǫ→0 dχǫm in the sense of distributions. Thus in making estimates of the Fourier
transform bm, we can assume that m has compact support (and hence can freely integrate
Theorem 4.1. Let m ∈ M∞(E). Then the Fourier transform bm is a distribution K belonging
given by K(x) = bm(x) =RRN e2πihx,ξim(ξ) dξ, and for every γ ∈ NN
To show that K = bm ∈ P0(E), we must verify the differential inequalities of Definition 2.2
and Definition 2.5 as well as the the cancellation conditions of Definition 2.2. We begin with
the differential inequalities.
∂γK(x) = (2πi)γZRN
e2πihx,ξi ξγ m(ξ) dξ.
If γ ∈ NC1 × · · · × NCn and M ≥ 0, we must show that there are constants Cγ , Cγ,M > 0
(4.1)
(independent of the support of m) so that
nYj=1
Nj(x)−(Qj +[[γ j ]])
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ Cγ
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ Cγ,M (1 + x)−M if x /∈ B(1),
if x ∈ B(1),
where x is the ordinary Euclidean norm of x ∈ RN .
(4.2)
The second inequality in (4.2) follows from a standard integration by parts. Let ∆ denote
the ordinary Laplace operator. Then
23
ZRN
e2πihx,ξi ξγ m(ξ) dξ = (4π2x)−2MZRN
= (4π2x)−2MZRN
(−∆ξ)Mhe2πihx,ξii ξγ m(ξ) dξ
e2πihx,ξi (−∆ξ)M [ξγ m(ξ)] dξ.
, so the last integral converges
absolutely independently of the support of m.
We turn to the heart of the matter which is the first estimate in (4.2). Suppose that
ξ [ξγ m(ξ)](cid:12)(cid:12) ≤ CM(cid:0)1 + ξ(cid:1)−N −1
If M is large enough, (cid:12)(cid:12)∆M
x ∈ B(1) ∩ ES where S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S(n). By (3.4), for 1 ≤ r ≤ s, if j ∈ Ir we
where nS,r(xIr ) ≈Pj∈Ir
nj(xj)1/e(j,kr ) is the norm defined in (3.5). Hence, for x ∈ B(1)∩ES,
Nj(x) ≈ nkr (xkr )e(j,kr ) ≈ nS,r(xIr ),
the estimate (4.2) is equivalent to the estimate
(4.3)
have
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) .
nkr (xkr )− Pj∈Ir
e(j,kr )(Qj +[[γ j ]]),
(4.4)
sYr=1
which, by (4.3), may also be interpreted as the differential inequality of a product kernel for
the decomposition RN = RI1 ⊕· · ·⊕RIs with the dilations ·S such that λ·S xIr =(cid:8)λe(j,kr ) ·xj :
j ∈ Ir(cid:9).
To obtain (4.4) we adapt the standard method used for proving that product kernels and
multipliers are related via the Fourier transform, cf. Theorem 2.1.11 in [NRS01]. This consists
in splitting the Fourier integral (4.1) into 2s regions depending on whether nS,r(ξIr ) is larger
or smaller9 than nkr (xkr )−1 for 1 ≤ r ≤ s. For any r, we take advantage of the smallness of
the domain of integration if we are considering nS,r(ξIr ) < nkr (xkr )−1, while we integrate by
parts in ξkr in the other case10.
Let ϕ0 and ϕ1 be smooth functions on the positive half-line with ϕ0(t) + ϕ1(t) ≡ 1,
ϕ0(t) ≡ 1 for t ≤ 1
2 and ϕ0(t) ≡ 0 for t ≥ 2. For ǫ = (ǫ1, . . . , ǫs) ∈ {0, 1}s set
sYr=1
ϕǫr(cid:0)nS,r(ξIr ) nkr (xkr )(cid:1),
e2πihx,ξimǫ(ξ) dξ.
mǫ(ξ) = m(ξ)
Kǫ(x) =ZRN
Note that m(ξ) = Pǫ∈{0,1}s mǫ(ξ) and K(x) = Pǫ∈{0,1}s Kǫ(x), so it will suffice to show
that
nkr (xkr )− Pj∈Ir
e(j,kr )(Qj +[[γ j ]]).
(4.5)
(cid:12)(cid:12)∂γKǫ(x)(cid:12)(cid:12) .
sYr=1
9It may seem strange that we measure the size of ξIr with nS,r(ξIr ) instead of bns,r(ξIr ). The point is that
we have estimates of derivatives of m in terms of bNkr (ξ) and we want a bound from below on this quantity
that involves a norm on RIr . By (3.7), we need to use nS,r(ξIr ).
10 We are not claiming that m is a product multiplier for the coarser decomposition.
If it were so, the
estimate
(cid:12)(cid:12)∂ γ K(x)(cid:12)(cid:12) .
sY
r=1
nS,r(xIr )− Pj∈Ir
e(j,kr )(Qj +[[γj ]]),
would hold for every x ∈ RN , which is not true. Also notice that formula (4.6) below, which gives a differential
inequality typical of a product multiplier, only holds for derivatives in the ξkr .
24
Since nS,r(ξIr ) is homogeneous of degree one under the dilations λ ·S ξIr , and since e(kr, kr) =
Since m ∈ M(E), and since bNkr (ξ) & nS,r(ξIr ) by (3.7), we have
δk1
ξk1
(cid:12)(cid:12)∂
· · · ∂δks
ξks
m(ξ)(cid:12)(cid:12) .
sYr=1
sYr=1 bNkr (ξ)−[[δkr ]] .
1, it follows from Proposition 15.1 in Appendix II that(cid:12)(cid:12)∂γ
all multi-indices γ,. Using the chain rule, we see that ∂δkr
of terms of the form
nS,r(ξIr )−[[δkr ]].
(4.6)
ξkr
nS,r(ξIr )(cid:12)(cid:12) ≤ CγnS,r(ξIr )1−[[γ]] for
ξkrhϕǫr(cid:0)nS,r(ξIr ) nkr (xkr )(cid:1)i is a sum
nS,r(ξIr ) · · · ∂αm
ξkr
nS,r(ξIr )
ϕ(m)
ǫr (cid:0)nS,r(ξIr )nkr (xkr )(cid:1) nkr (xkr )m ∂α1
ξkr
where 1 ≤ m ≤ δkr and α1 + · · · + αm = δkr . Since nkr (xkr ) ≈ nS,r(ξIr )−1 on the support
of ϕ(m)
, m ≥ 1, it follows that
ǫr
(cid:12)(cid:12)(cid:12)ϕ(m)
ǫr (cid:0)nS,r(ξIr )nkr (xkr )(cid:1) nkr (xkr )m ∂α1
Combining these estimates gives(cid:12)(cid:12)∂
is supported where
δk1
ξk1
ξkr
· · · ∂δks
ξks
nS,r(ξIr ) · · · ∂αm
ξkr
mǫ(ξ)(cid:12)(cid:12) ≤ Cδ Qs
nS,r(ξIr )(cid:12)(cid:12)(cid:12) . nS,r(ξIr )−[[δkr ]].
r=1 nS,r(ξIr )−[[δkr ]].
Now for each ǫ = (ǫ1, . . . , ǫs) ∈ {0, 1}s let A(ǫ) = {r ∈ {1, . . . , s} : ǫr = 1}. Note that mǫ
nS,r(ξIr ) ≤ 2nkr (xkr )−1
nkr (xkr )−1
nS,r(ξIr ) ≥
1
2
if r /∈ A(ǫ),
if r ∈ A(ǫ).
Choose δ = (δ1, . . . , δn) ∈ NC1 × · · · × NCn with δj = 0 unless j = kr for some r ∈ A(ǫ). If
r ∈ A(ǫ), we will choose the entries of each δkr = (δkr ,1, . . . , δkr ,ckr ) to be much larger than
the corresponding entries of γkr = (γkr ,1, . . . , γkr ,ckr ). Using (4.1) we integrate by parts to
obtain
(cid:0) Yr∈A(ǫ)
x
δkr
kr (cid:1)∂γ
x
x Kǫ(x) = (2πi)γZRN(cid:0) Yr∈A(ǫ)
= (2πi)γ−δZRN(cid:2) Yr∈A(ǫ)
= (−1)δ(2πi)γ−δZRN
δkr
∂δkr
kr (cid:1)e2πihx,ξiξγ mǫ(ξ) dξ
ξkr(cid:3)(cid:0)e2πihx,ξi(cid:1) ξγmǫ(ξ) dξ
e2πihx,ξi(cid:2) Yr∈A(ǫ)
∂δkr
ξkr(cid:3)(cid:0)ξγmǫ(cid:1)(ξ) dξ.
(4.7)
The product rule shows that(cid:2)Qr∈A(ǫ) ∂δkr
of the form
ξkr(cid:3)(cid:0)ξγmǫ(cid:1)(ξ) can be written as a finite sum of terms
(cid:16) Yr∈A(ǫ)
∂
µkr
ξkr (cid:17)(cid:2)ξγ(cid:3)(cid:16) Yr∈A(ǫ)
∂
δkr −µkr
ξkr
(cid:17)(cid:2)mǫ(cid:3)(ξ)
(4.8)
where µ = (µ1, . . . , µn) ≤ (δ1, . . . , δn). The first factor is zero if the entries of µkr are
larger than the corresponding terms γkr . Thus our choice of δkr guarantees that the entries
of δkr − µkr are large enough to make the integrals below converge independently of the
compact support of m. Note that µj = 0 unless j = kr for some r ∈ A(ǫ). Then
(cid:16) Yr∈A(ǫ)
∂
µkr
ξkr(cid:17)(cid:2)ξγ(cid:3) = Cµ,γ Yr∈A(ǫ)(cid:16)Yj∈Ir
γ j −µj
j
ξ
(cid:17) Yr /∈A(ǫ)(cid:0)Yj∈Ir
ξ
γ j
j (cid:1)
According to (4.6) we also have
∂
δkr −µkr
ξkr
(cid:12)(cid:12)(cid:12)(cid:16) Yr∈A(ǫ)
ξkr(cid:17)(cid:2)ξγ(cid:3)(cid:16) Yr∈A(ǫ)
nS,r(ξIr )−[[δkr
(cid:17)(cid:2)mǫ(cid:3)(ξ)(cid:12)(cid:12)(cid:12) ≤ Cδ,µ Yr∈A(ǫ)
(cid:17)(cid:2)mǫ(cid:3)(ξ)(cid:12)(cid:12)(cid:12)
]] Yr∈A(ǫ)(cid:16)Yj∈Ir
δkr −µkr
ξkr
]]+[[µkr
ξ
∂
Thus
∂
µkr
(cid:12)(cid:12)(cid:12)(cid:16) Yr∈A(ǫ)
≤ Cδ,γ Yr∈A(ǫ)
terms of the form
Now using (4.7), (4.8), and (4.9), we can estimate(cid:12)(cid:12)(cid:0) Yr∈A(ǫ)
Yr∈A(ǫ)"
nS,r (ξIr )≥nkr (xkr )−1 (cid:16)Yj∈Ir
γ j −µj
j
Z
25
nS,r(ξIr )−[[δkr ]]+[[µkr ]]
γ j −µj
j
γ j
ξ
(4.9)
j (cid:1).
(cid:17) Yr /∈A(ǫ)(cid:0)Yj∈Ir
x Kǫ(x)(cid:12)(cid:12) by a finite sum of
kr (cid:1)∂γ
δkr
x
ξ
(cid:17) nS,r(ξIr )−[[δkr ]]+[[µkr ]] dξIr#
× Yr /∈A(ǫ)"
nS,r (ξIr )<nkr (xkr )−1 (cid:16)Yj∈Ir
e(kr ,j)(Qj +[[γ j ]]) Yr∈A(ǫ)
nkr (xkr )− Pj∈Ir
Z
ξ
γ j
j (cid:17) dξIr#.
e(kr ,j)(Qj +[[γ j ]]−[[δr]])
nkr (xkr )[[δr]]
nkr (xkr )− Pj∈Ir
e(kr ,j)(Qj +[[γ j ]]).
nkr (xkr )− Pj∈Ir
Yr /∈A(ǫ)
= Yr∈A(ǫ)
sYr=1
According to Proposition 15.3 in the appendix, this last expression is bounded by
Comparing with (4.7), this shows that ∂γ
correct differential inequalities.
x K(x) satisfies (4.4), so the kernel K satisfies the
We also need to verify the cancellation conditions given in Definition 2.2, part (b). Let
x = (x′, x′′) ∈ RN1 × RN2 be a decomposition of the variables into two subsets, with x′ =
(xp1 , . . . , xpr ) and x′′ = (xq1 , . . . , xqt ). Let ψ be a normalized bump function in the x′′
variables. Then
Kψ,R(x′) =ZRN2
K(x′, x′′)ψ(R · x′′) dx′′
e2πi<x,ξ>m(ξ)ψ(R · x′′) dx′′ dξ
e2πi<x′′,ξ′′>ψ(R · x′′) dx′′(cid:21) m(ξ′, ξ′′) dξ′′# e2πi<x′,ξ′> dξ′
=ZRNZRN2
=ZRN1"ZRN2(cid:20)ZRN2
=ZRN1(cid:20)ZRN2hR−Q′′bψ(R−1 · ξ′′)i m(ξ′, ξ′′) dξ′′(cid:21) e2πi<x′,ξ′> dξ′
=ZRN1
e2πi<x′,ξ′>mψ,R(ξ′) dξ′
where
mψ,R(ξ′) =ZRN2hR−Q′′bψ(R−1 · ξ′′)i m(ξ′, ξ′′) dξ′′ =ZRN2 bψR(ξ′′)m(ξ′, ξ′′) dξ′′.
26
Since bψRL1 is finite and independent of R, it is easy to see that mψ,R belongs to the class
M∞(bNp1 , . . . , bNpr ) on the space RN1. Thus the estimates for Kψ,R follow from the same
arguments. This completes the proof of Theorem 4.1.
4.2 Fourier transforms of kernels
Theorem 4.2. Let K ∈ P0(E). Then the Fourier transform of K is a smooth function
belonging to the class M∞(E).
Proof. Suppose that K ∈ P0(E). According to Definition 2.5 we can write K = K0 + ψ where
support in the ball B(1). The distribution K then acts on C∞(RN ), and in particular, we set
K0 ∈ P0(E) has compact support in B(1) and ψ ∈ S(RN ). Since bψ ∈ M∞(E), it suffices to
show that cK0 ∈ M∞(E). Thus without loss of generality we can assume that K has compact
m(ξ) =(cid:10)K, exp[2πi < · , ξ >](cid:11). From general principles, this is a smooth function and
ξ m(ξ) = (2πi)γ(cid:10)K, xγ exp[2πi < · , ξ >](cid:11) = (2πi)γ(cid:10)xγ K, exp[2πi < · , ξ >](cid:11)
(4.10)
∂γ
e2πi<x,ξ>xγ K(x) dx.
= (2πi)γZRN
Note that if γ ≥ 1, the distribution xγK coincides with integration against the locally
integrable function xγK(x), (see Remark 2.1.7 on page 36 of [NRS01]), and this justifies the
last equality.
We want to show
.
(4.11)
(cid:12)(cid:12)∂γ
ξ m(ξ)(cid:12)(cid:12) ≤ Cγ
Since K has compact support, this inequality follows if ξ ∈ B(1). Thus we only need to
nYj=1(cid:0)1 + bNj(ξ)(cid:1)−[[γ j ]]
consider ξ ∈ B(1)c. Suppose ξ belongs to bES where S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S(n). Then
for j ∈ Ir, 1 + bNj(ξ) ≈ nkr (ξkr )1/e(kr ,j) and we want to prove
(cid:12)(cid:12)∂γ
ξ m(ξ)(cid:12)(cid:12) ≤ Cγ
we split the integral in (4.10) into 2s regions depending on whether bnS,r(xIr ) is larger or
sYr=1Yj∈Ir(cid:0)1 + bNj(ξ)(cid:1)−[[γ j ]]
Let ϕ0 and ϕ1 be smooth functions on the positive half-line with ϕ0(t) + ϕ1(t) ≡ 1,
The proof now follows the same pattern as the proof of Theorem 4.1. To establish (4.12)
smaller than nkr (ξkr )−1.
nkr (ξkr )− Pj∈Ir
[[γ j ]]/e(kr ,j).
sYr=1
ϕ0(t) ≡ 1 for t near zero, and ϕ0(t) ≡ 0 for t large. For ǫ = (ǫ1, . . . , ǫs) ∈ {0, 1}s set
(4.12)
≈ Cγ
sYr=1
ϕǫr(cid:0)nkr (ξkr )bnS,r(xIr )(cid:1),
e2πihx,ξiKǫ(x) dx.
Kǫ(x) = K(x)
mǫ(ξ) =ZRN
Note that m(ξ) = Pǫ∈{0,1}s mǫ(ξ) and K(x) = Pǫ∈{0,1}s Kǫ(x). Recall that bnS,r(ξr) =
Pj∈Ir
nj(ξj)e(kr ,j) so bnS,r is homogeneous relative to the family of dilations {ξj; j ∈ Ir} →
{λ1/e(kr ,j)ξj : j ∈ Ir}. It follows from Proposition 15.1 that for δkr ∈ NCkr and ǫ ∈ {0, 1}s
we have
(cid:12)(cid:12)∂
δk1
xk1
· · · ∂δks
xks Kǫ(x)(cid:12)(cid:12) ≤ Cδ
nYj=1
Nj(x)−Qj
sYr=1bnS,r(xkr )−[[δkr ]]
(4.13)
(4.14)
since e(kr, kr) = 1. Now fix ǫ ∈ {0, 1}s and let A(ǫ) = {r ∈ {1, . . . , s} : ǫr = 1}. Let
δkr ∈ Nckr . We will choose the entries of each δkr to be sufficiently large integers. Then
27
(cid:0) Yr∈A(ǫ)
ξδr
ξδr
ξ mǫ(ξ) = (−2πi)γZRN(cid:0) Yr∈A(ǫ)
kr(cid:1)∂γ
= (−2πi)γ−δZRN(cid:2) Yr∈A(ǫ)
= (−1)δ(−2πi)γ−δZRN
∂δr
∂δr
kr(cid:1)e−2πihx,ξixγ Kǫ(x) dx
xkr(cid:3)(cid:0)e−2πihx,ξi(cid:1) xγ Kǫ(x) dx
e−2πihx,ξi(cid:2) Yr∈A(ǫ)
xkr(cid:3)(cid:0)xγKǫ(cid:1)(x) dx.
xkr(cid:3)(cid:0)xγ Kǫ(cid:1)(x) can be written as a
(cid:0)Kǫ(x)(cid:1)(cid:1). In the first factor,
Yr /∈A(ǫ)(cid:0)Yj∈Ir
j (cid:1)x
γ kr −µkr
kr
j (cid:1)
(4.15)
γ j
γ j
x
The product rule shows that the derivative (cid:2)Qr∈A(ǫ) ∂δr
sum of terms of the form(cid:0)Qr∈A(ǫ) ∂µr
xkr(cid:1)(cid:0)xγ(cid:1)(cid:0)Qr∈A(ǫ) ∂δr−µr
the entries of µr can only be as large as the corresponding terms γr. Thus we will choose δr
to have entries much larger than those of γr so that the entries of δr − µr are large enough
to make the integrals below converge independently of the compact support of m. Then
xkr
(cid:16) Yr∈A(ǫ)
∂µr
xkr(cid:17)(cid:0)xγ(cid:1) = Cµ,γ Yr∈A(ǫ)(cid:0) Yj∈Ir
j6=kr
x
Note that e(j, l) ≤ e(j, kr)e(kr, l). Thus
According to (4.14)
(cid:12)(cid:12)(cid:12)(cid:16) Yr∈A(ǫ)
Nj(x) =
xle(kr ,l)e(J,kr )
nXl=1
xle(j,l) ≥Xl∈Ir
xle(j,l) ≥Xl∈Ir
&hXL∈Ir
xle(kr ,l)ie(j,kr )
=bnS,r(xIr )e(j,kr ).
(cid:0)Kǫ(cid:1)(x)(cid:17)(cid:12)(cid:12)(cid:12) ≤ Cδ,µ
nYj=1
Nj(x)−Qj Yr∈A(ǫ)bnS,r(xIr )−[[δr]]+[[µr]]
≤ Cδ,µYr=sbnS,r(xIr )− Pj∈Ir
Qj /e(kr ,j) Yr∈A(ǫ)bnS,r(xIr )−[[δr ]]+[[µr]].
∂δr −µr
xkr
(4.16)
x
γ j
terms of the form
ZRIr
ZRIr
bnS,r (xIr )<nkr (ξkr )−1
Thus according to (4.7) and (4.16), we can estimate (cid:12)(cid:12)(cid:0)Qr∈A(ǫ) xδr
x Kǫ(x)(cid:12)(cid:12) by a sum of
kr(cid:1)∂γ
h Yr /∈A(ǫ)
Qj e(j,kr ) dxIri×
(cid:12)(cid:12)(cid:12)bnS,r(xIr )− Pj∈Ir
h Yr∈A(ǫ)
≈h Yr /∈A(ǫ)
h Yr∈A(ǫ)
(cid:12)(cid:12)(cid:12)Yj∈Ir
j (cid:12)(cid:12)(cid:12)bnS,r(xIr )− Pj∈Ir
(cid:12)(cid:12)(cid:12)(cid:0) Yj∈Ir
j (cid:1)x
j (cid:12)(cid:12)(cid:12)bnS,r(xIr )− Pj∈Ir
(cid:12)(cid:12)(cid:12)Yj∈Ir
(cid:12)(cid:12)(cid:12)(cid:0)Yj∈Ir
j (cid:1)(cid:12)(cid:12)(cid:12)bnS,r(xIr )− Pj∈Ir
Qj e(j,kr ) dxIri×
Qj e(j,kr )−[[δr ]] dxIri.
Qj e(j,kr )−[[δr]]+[[µr ]] dxIri
bnS,r (xIr )≥nkr (ξkr )−1
bnS,r(xIr )<nkr (ξkr )−1
bnS,r (xIr )≥nkr (ξkr )−1
γ kr −µkr
kr
ZRIr
ZRIr
j6=kr
γ j
γ j
γ j
x
x
x
28
the appropriate γ j are equal to zero. However, for these terms, we use the cancellation prop-
It may happen that the integrals over the regions wherebnS,r(xIr ) < nkr (ξkr )−1 diverge if all
erties of the kernels since ϕǫr(cid:0)nkr (ξkr )bnS,r(xIr )(cid:1) are dilates of normalized bump functions.
Thus according to Proposition 15.3, this last expression is bounded by
(−([[γ j ]]−[[δr]])e(j,kr ))
Pj∈Ir
nkr (ξkr )
Yr /∈A(ǫ)
nkr (ξkr )− Pj∈Ir
= Yr∈A(ǫ)
[[γ j ]]e(j,kr ) Yr∈A(ǫ)
sYr=1
nkr (ξkr )[[δr]]e(j,kr )
nkr (ξkr )− Pj∈Ir
[[γj ]]e(j,kr ).
Comparing with (4.7), this shows that ∂γ
the correct differential inequalities, and this completes the proof of Theorem 4.2.
ξ m(ξ) satisfies (4.12), so the multiplier m satisfies
5 Dyadic sums of Schwartz functions
In Section 4 we showed that distributions K ∈ P0(E) can be characterized in terms of their
Fourier transforms. We now begin the study of a different kind of characterization in terms
of dyadic sums of normalized bump functions. In this section we show that appropriate sums
of dilates do converge to distributions K ∈ P0(E). In Theorem 6.14 below, we will see that,
modulo 'coarser' kernels, every distribution K ∈ P0(E) can be decomposed in this way.
If ϕ ∈ S(RN ), the dilates [ϕ]t and (ϕ)t were defined in equation (2.8). In this section
we show that appropriate dyadic sums of dilates [ϕ]t of Schwartz functions having suitable
cancellation properties converge to distributions K ∈ P0(E). Similarly we show that dyadic
sums of dilates (ϕ)t of Schwartz functions with appropriate vanishing conditions converge to
multipliers M∞(E). The main results appear in subsections 5.4-5.6. However, before turning
to the main results of this section, we begin with some further remarks about cones associated
to matrices, and then discuss inclusions among classes of kernels and multipliers defined by
different matrices.
5.1 Cones associated to a matrix E
Let E = {e(j, k)} be an n × n matrix satisfying the basic hypotheses in (2.10) and let
µ ≥ 0. We introduce the following notation.
Γ(E) =nt = (t1, . . . , tn) ∈ Rn : ∀j 6= k, e(j, k)tk ≤ tj < 0o,
Γµ(E) =nt = (t1, . . . , tn) ∈ Rn : ∀j 6= k, e(j, k)tk − µ ≤ tj < 0o,
Γo(E) = int(cid:0)Γ(E)(cid:1) =nt = (t1, . . . , tn) ∈ Rn : ∀j 6= k, e(j, k)tk < tj < 0o,
ΓZ,µ(E) =(cid:8)I = (i1, . . . , in) ∈ Zn : ∀j 6= k, e(j, k)ik − µ ≤ ij < 0o,
ΓZ(E) = ΓZ,0(E).
(5.1)
We will refer to properties of the cone Γ(E) which are proved in Appendix I (Section 14).
In particular, Lemma 5.1 summarizes the content of Lemma 14.7 and Lemma 14.9.
Since E satisfies the basic hypothesis, Γ(E) is not empty.11 However its interior, which is
an open convex cone, can be empty. For example, if E =(cid:20)1
1
2
2
1(cid:21), Γ(E) is the one-dimensional
11See Proposition 14.5 in Appendix I.
29
ray given by x1 = 2x2, x2 < 0, which has no interior. Note that in this case 1 = e(1, 2)e(2, 1).
The following Lemma shows that, more generally, any such equality is the obstruction to the
open cone being non-empty.
Lemma 5.1. Let E be a matrix satisfying the basic hypotheses (2.10).
(a) The interior of Γ(E) is non-empty if and only if E is reduced, i.e., 1 < e(j, k)e(k, j) for
every pair 1 ≤ j 6= k ≤ n.
(b) Γ(E) is contained in the intersection of the hyperplanes(cid:8)t ∈ Rn : e(j, k)tk = tj(cid:9) for all
pairs (j, k) such that e(j, k)e(k, j) = 1 and has non-empty interior in this subspace of
Rn.
5.2
Inclusions among classes of kernels associated to different ma-
trices
The cone Γ(E) contains the relevant information on the structure of kernels and multipliers
associated with the matrix E.
Proposition 5.2. Let E, E′ be two n × n matrices, both satisfying the basic hypotheses. The
following are equivalent:
(i) P0(E) ⊆ P0(E′);
(ii) M∞(E) ⊆ M∞(E′);
(iii) for every j = 1, . . . , n and ξ ∈ Bc, bN ′
(iv) for every j, k = 1, . . . , n, e(j, k) ≤ e′(j, k);
j(ξ) ≤ bNj(ξ);
(v) Γ(E) ⊆ Γ(E′).
Proof. The equivalence between (i) and (ii) is given by Theorems 4.1 and 4.2. The implications
(v)⇒(iv)⇒(iii)⇒(ii) are obvious. In order to prove that (ii) implies (v), assume by contradic-
tion that Γ(E) in not contained in Γ(E′). Then there exists a half-line R−α ∈ Γ(E) \ Γ(E′),
where α = (α1, . . . , αn) has strictly positive components.
j=1 nj(ξ)1/αj and consider multipliers m which satisfy the inequalities
Let Nα =Pn
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) .(cid:0)1 + Nα(ξ)(cid:1)− Pj αj [[γj ]]
(5.2)
for all γ ∈ NN . We show that every such m is in M∞(E) but there exist some which are not
in M∞(E′).
To prove the first statement it suffices to show that, for every j,
This condition is equivalent to the inequalities αk/αj ≤ e(k, j) for every j, k, and in turn
these are equivalent to the condition −α ∈ Γ(E).
(cid:0)1 + Nα(ξ)(cid:1)αj & 1 + bNj(ξ).
Similarly, since −α 6∈ Γ(E′), there exist j, k such that αk/αj > e′(k, j). Suppose in (5.2)
we take γ = (0, . . . , 0, γj, . . . , 0) and ξ = (0, . . . , 0, ξk, . . . , 0). The inequality becomes
(cid:12)(cid:12)∂γj m(0, . . . , ξk, . . . , 0)(cid:12)(cid:12) .(cid:0)1 + nk(ξk)1/αk(cid:1)−αj [[γj ]]
30
where the right-hand side is not dominated by
It is then sufficient to construct m satisfying (5.2) and such that, for some l ∈ Cj it satisfies
the opposite inequality
e.g.,
.
[[γj]]
e′(k,j)
≈(cid:0)1 + bN ′
j(ξ)(cid:1)−[[γj ]]
(cid:0)1 + nk(ξk)(cid:1)− 1
(cid:12)(cid:12)∂ξlm(0, . . . , ξk, . . . , 0)(cid:12)(cid:12) &(cid:0)1 + nk(ξk)1/αk(cid:1)−αj dl,
m(ξ) =
ξl
,
eNα(ξ)αj dl
where eNα is a smooth homogeneous norm equivalent to Nα and even in ξl. This m is not in
M∞(E′).
5.3 Coarser decompositions and lower dimensional matrices
We will often need to consider classes P0(E) where E is a lower-dimensional (say s × s)
matrix and the variables x1, x2, . . . , xn have correspondingly been grouped together into s
blocks with given dilation exponents within each block.
Precisely, consider a coarser decompositions of RN ,
RN = RA1 ⊕ RA2 ⊕ · · · ⊕ RAs,
where A = {A1, A2, . . . , As} is a partition of {1, 2, . . . , n} and RAr is defined according to
(2.3). We also assign an n-tuple α = (α1, α2, . . . , αn) of positive exponents to define dilations
on each RAr , with homogeneous norm
nAr (tAr ) = Xj∈Ar
nj(tj)αj ,
r = 1, . . . , n.
(5.3)
Given an s × s matrix E =(cid:0)e(t, r)(cid:1)1≤t,r≤s satisfying the basic hypotheses, with s < n, let
P0(E, A, α) be the class of associated kernels and M∞(E, A, α) the class of multipliers. The
associated norms are, for 1 ≤ t ≤ s,
sXr=1 Xk∈Ar
sXr=1 Xk∈Ar
e(cid:0)t, r).
sXr=1
sXr=1
Nt(x) =
bNt(ξ) =
nAr (xAr )e(t,r) ≈
nAr (ξAr )
1
e(r,t) ≈
nk(xk)αke(t,r)
nk(ξk)
αk
e(r,t) .
Define the n × n matrix E♯ =(cid:0)e♯(j, k)(cid:1) where, for j ∈ At and k ∈ Ar,
e♯(j, k) =
αk
αj
Then E♯ satisfies the basic assumptions and the associated dual norms are
sXr=1 Xk∈Ar
sXr=1 Xk∈Ar
N ♯
j (x) =
j (ξ) =
bN ♯
αk
αj
e(t,r)
nk(xk)
≈ Nt(x)1/αj
αk
αj
nk(ξk)
1
e(r,t) ≈ bNt(ξ)1/αj .
(5.4)
(5.5)
Lemma 5.3. We have the equalities M∞(E, A, α) = M∞(E♯), P0(E, A, α) = P0(E♯).
Proof. Let m ∈ M∞(E, A, α). Then, for every γ = (γ 1, . . . , γ s) ∈ NA1 × · · · × NAs,
31
where
,
ξ m(ξ)(cid:12)(cid:12) .
(cid:12)(cid:12)∂γ
[[γt]] = Xj∈At
sYt=1(cid:0)1 + bNt(ξ)(cid:1)−[[γ t]]
αj Xl∈Cj
γldl,
1
by (2.7), taking into account the dilations on RAt which make (5.3) a homogeneous norm.
1, . . . , γ ′
j)j∈At , γ can be written as an element of NC1 × · · · × NCn as γ = (γ ′
If γt = (γ ′
n).
Then
and therefore
γldl = [[γ′
j]]
Xl∈Cj
sYt=1(cid:0)1 + bNt(ξ)(cid:1)−[[γt]]
=
≈
αj
sYt=1 Yj∈At(cid:0)1 + bNt(ξ)(cid:1)− 1
sYt=1 Yj∈At(cid:0)1 + bN ♯
j (ξ)(cid:1)−[[γ ′
[[γ ′
j ]]
j ]]
=
nYj=1(cid:0)1 + bN ♯
j (ξ)(cid:1)−[[γ ′
j ]]
.
Then m ∈ M∞(E). The opposite implication is proved in the same way. The equality
P0(E) = P0(E♯) follows from Theorems 4.1 and 4.2.
Notice that Lemma 5.3 implies that kernels belonging to P0(E, A, α) satisfy stronger
cancellation properties than those required by the definition.
Combining together Proposition 5.2 and Lemma 5.3 we obtain the following statement.
Corollary 5.4. Let A and α be as above, and let E an s × s matrix satisfying the basic
hypotheses. E′ an n × n matrix also satisfying the basic hypotheses. We have the inclusion
P0(E, A, α) ⊆ P0(E′) if and only if the matrices E♯ and E′ satisfy the condition
e♯(j, k) ≤ e′(j, k),
∀ j, k, 1 ≤ j, k ≤ n,
together with the other equivalent conditions in Proposition 5.2.
This result can be applied in particular to the case where an n × n matrix E satisfying the
basic hypotheses is not reduced, i.e., there exist j 6= k such that e(j, k)e(k, j) = 1. As shown
in Appendix I (Section 14.5), in this case the j-th row of E equals e(j, k) times the k-th row,
and the same holds for the k-th and j-th column.
Following the notation of Lemma 14.9, we denote by k1, . . . , ks representatives of the
equivalence classes A1, . . . , As under the relation
j ∼ k ⇐⇒ e(j, k)e(k, j) = 1.
Then, for all j ∈ At,
Nj(x) ≈ Nkt(x)e(j,kt )
and
Moreover, for every r ∈ {1, . . . , s},
Nkt(x) ≈
sXr=1h Xk∈Ar
nk(xk)e(kr ,k)ie(kt,kr )
and
bNj(ξ) ≈ bNkt (ξ)e(j,kt).
sXr=1h Xk∈Ar
bNkt(ξ) ≈
(5.6)
nk(ξk)e(kr ,k)i1/e(kr ,kt)
.
32
Therefore we are in the previous situation, with αk = e(kr, k) if k ∈ Ar and with the reduced
matrix
E♭ =(cid:0)e(kt, kr)(cid:1)1≤t,r≤s
defined in Section 14.5. Notice that (E♭)♯ = E. By Lemma 5.3 we have the following
Proposition 5.5. P0(E) = P0(E♭) and M∞(E) = M∞(E♭).
Proposition 5.5 makes it possible to replace, whenever it is appropriate, E by E♭ and
reduce matters to the case of a reduced matrix.
5.4 Size estimates
We now turn to the construction of the dyadic decompositions of kernels K belonging
to classes P0(E). In the rest of this section, we shall assume that Γo(E) is non-empty, and
consequently that ΓZ(E) is non-empty, where Γo(E) and ΓZ(E) were defined in equation (5.1).
0 (RN ) or S(RN ). We will show
Let (cid:8)ϕI : I ∈ ΓZ(E)(cid:9) be a uniformly bounded family in C∞
that if all the functions ϕI satisfy a suitable cancellation condition, then the sum
[ϕI ]I
(5.7)
K = XI∈ΓZ(E)
converges in the sense of distributions to an element K ∈ P0(E). The precise statement is
given in Theorem 5.13 below.
We first show that, even without any cancellation hypotheses on the functions ϕI , the sum
in equation (5.7) satisfies the differential inequalities of Definition 2.2 for the class P0(E).
Lemma 5.6. Suppose that Γo(E) is non-empty, and let (cid:8)ϕI : I ∈ ΓZ(E)(cid:9) ⊂ S(RN ) be a
uniformly bounded family of Schwartz functions. Then the series K(x) =PI∈ΓZ(E)[ϕI ]I (x)
converges absolutely for all x 6= 0, K ∈ C∞(RN \ {0}), and
∂γK(x) = XI∈ΓZ(E)
2− Pn
j=1 ij (Qj +[[γ j ]])(∂γϕI )(cid:0)2−i1 · x1, . . . , 2−in · xn(cid:1).
(5.8)
Moreover, for every γ = (γ1, . . . , γ n) ∈ NN and M ≥ 0 there exists a constant Cγ,M > 0 so
that
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ Cγ,M
nYj=1
Nj(x)−(Qj +[[γ j ]])(1 + x)−M .
(5.9)
Proof. Since the family {ϕI } is uniformly bounded in S(Rn), for each M > 0 there is a
constant Cγ,M so that
j=1 ij (Qj +[[γ j ]])(∂γϕI )(cid:0)2−i1 · x1, . . . , 2−in · xn(cid:1)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)2− Pn
≤ Cγ,M
2−ij (Qj +[[γ j ]])(1 +
nYj=1
2−ij nj(xj ))−M .
nXj=1
It follows from Proposition 16.3 in Appendix III that the series on the right hand side of
j=1(1 +
. This shows that the formal differentiation of equation (5.7) yielding (5.8) is
equation (5.8) converges absolutely and is bounded by C Qn
j=1 Nj(x)−(Qj +[[γ j ]])Qn
nj(x))−M ′
justified, and also establishes the inequality in (5.9).
As an application of Lemma 5.6 we check that kernels satisfying the differential inequalities
in part (a) of Definition 2.2 just fail to be integrable near the origin. See also Lemma 7.6.
Corollary 5.7. Let E satisfy (2.10) and let Nj(x) be defined in equation (2.9). Suppose that
Γo(E) is non-empty. Then
33
nYj=1
Nj(x)−Qj dx = +∞;
(a) ZB(1)
(b) If α > 0 and 1 ≤ k ≤ n thenZB(1)
Nk(x)α
Nj(x)−Qj dx < +∞ .
nYj=1
F ⊂ Γ(E) is a finite set with M elements , thenPI∈F [ϕ]I (x) ≤ CQN
Proof. Choose ϕ ∈ C∞
ϕI = ϕ for every I, the family {ϕI } is normalized in S(RN ). According to equation (5.9), if
j=1 Nj(x)−Qj , and hence
0 (RN ) supported in B(1) withRRN ϕ(x) dx = 1 and ϕ(x) ≥ 0. Then if
ZB(1)
NYj=1
[ϕ]I (x) dx = C−1M.
Letting M → ∞ gives statement (a). To prove statement (b), observe that there is a constant
ck,l > 0 so that if x ∈ B(1), then Nk(x) ≤ Nl(x)ck,l . Thus given α > 0 and 1 ≤ k ≤ n there
Nj(x)−Qj dx ≥ C−1XI∈FZRN
exist positive numbers ǫ1, . . . , ǫn so that Nk(x)α ≤Qn
Nj(x)−Qj dx ≤ ZB(1)
nYj=1
Znj (xj )≤1
nYj=1
ZB(1)
nYj=1
Nk(x)α
≤
l=1 Nl(x)ǫl . Then
Nj(x)−Qj +ǫj dx
nj(xj)−Qj +ǫj dxj < +∞.
This completes the proof.
5.5 Cancellation properties
Definition 5.8. Let ϕ ∈ S(RN ).
E is an n × n matrix satisfying (2.10).
We next turn to a discussion of cancellation properties. Let RN =Ln
(a) ϕ has cancellation in the variable xj ifZRCj
ϕ(x1, . . . , xj , . . . , xn) dxj = 0.
j=1 RCj and suppose
(b) ϕ has strong cancellations if it has cancellation in xj for 1 ≤ j ≤ n.
This terminology was used in Definition 5.3 of [NRSW12]. We will also need a variant of the
notion of weak-cancellation that was introduced in Definition 5.5 of [NRSW12].
Definition 5.9. A function ϕ ∈ C∞
to the n-tuple I = (i1, . . . , in) ∈ ΓZ(E) provided that one can write
0 (RN ) has weak cancellation with parameter ǫ > 0 relative
ϕ = XA⊂{1,...,n−1}hYj∈A
2−ǫΛ(I,j)iϕA
where the sum is taken over all subsets A ⊂ {1, . . . , n − 1}, and for each subset A,
34
(1) ϕA ∈ C∞
0 (RN ), and is normalized relative to ϕ;
(2) ϕA has cancellation in the variable xj for each j /∈ A, and in particular for j = n;
(3) τA : A → {1, . . . , n} is a mapping with τA(j) 6= j for all j ∈ A;
(4) Λ(I, j) = ij − e(cid:0)j, τA(j)(cid:1) iτA(j).
Note that if I ∈ ΓZ(E), then ij − e(cid:0)j, τA(j)(cid:1) iτA(j) ≥ 0. Thus if ϕ has weak cancellation,
it can be written as a finite sum of functions, each of which has the property that for each
index j, either there is cancellation in xj or the lack of cancellation is compensated by a gain
2−ǫ[ij−e(j,τA(j)) iτA(j)].
The following is Lemma 5.1 in [NRSW12]:
Proposition 5.10. If ϕ ∈ S(RN ) and {J1, . . . , Jr} are disjoint subsets of {1, . . . , N }, then
the following are equivalent:
(a) R ϕ(x) dxJk = 0 for 1 ≤ k ≤ r.
with respect to ϕ, so that ϕ(x) =Pj1∈J1
Moreover, if ϕ ∈ C∞
relative to ϕ.
We will also need the following result.
(b) For every (j1, . . . , jr) ∈ J1 × · · · × Jr there is a function ϕj1,...,jr ∈ S(Rn), normalized
0 (RN ), then we can choose the functions ϕj1,...,jr ∈ C∞
0 (RN ) normalized
· · ·Pjr ∈Jr
∂j1 · · · ∂jr ϕj1,...,jr (x).
Proposition 5.11. Let L = {l1, . . . , la} and M = {m1, . . . , mb} be complementary subsets
of {1, . . . , n}, let A be a subset of M , and let R = {Rm1, . . . , Rmb} be positive real numbers.
Write any x ∈ RN as x = (x′, x′′) where x′ = (xl1 , . . . , xla) ∈ RL and x′′ = (xm1 , . . . , xmb) ∈
RM . Let ϕ ∈ S(RN ) and let ψ ∈ C∞
0 (RM ) have support in B(1). Put
Φ(x′) =ZRM
ϕ(x′, x′′)ψ(Rm1 · xm1 , . . . , Rmb · xmb ) dxm1 · · · dxmb .
(a) Φ ∈ S(cid:0)RL(cid:1) and is normalized relative to ϕ and ψ with constants independent of R. If
0 (cid:0)RL(cid:1) and is normalized relative to ϕ and ψ with constants
ϕ ∈ C∞
independent of R.
0 (RN ), then Φ ∈ C∞
(b) If some of the parameters {Rmj } are large, there is an improvement over statement (a):
there exists ǫ > 0 independent of ϕ, ψ and R so that
Φ(x′) =h Ymj ∈M
mj(cid:9)ieΦ(x′)
where eΦ ∈ S(RN ) is normalized relative to ϕ and ψ, with constants independent of R. If
0 (cid:0)RL(cid:1) and is normalized relative
min(cid:8)1, R−ǫ
0 (RN ) is supported in the unit ball, then eΦ ∈ C∞
ϕ ∈ C∞
to ϕ and ψ with constants independent of R.
(c) If ϕ has cancellation in some of the variables in M there is an improvement over statement
(b) even if Rmj is small. There exists ǫ > 0 independent of ϕ, ψ and R so that if ϕ has
cancellation in xj for each j ∈ A then
Φ(x′) =h Ymj ∈M\A
mj(cid:9)ieΦ(x′),
where eΦ is normalized relative to ϕ and ψ, with constants independent of of R.
mj(cid:9)ih Ymj ∈A
min(cid:8)1, R−ǫ
min(cid:8)Rǫ
mj , R−ǫ
35
Proof. Statement (a) is clear. Next, since ψ(x′′) is supported where nmj (xmj ) ≤ 1, we have
nmj (xmj ) ≤ R−1
mj if ψ(Rm1 · xm1 , . . . , Rmb · xmb) 6= 0. The integration in the definition of
Φ(x′) takes place over the set(cid:8)(xm1 , . . . , xmb) : nmj (xmj ) ≤ R−1
integral by the (small) size of the region of integration in the variables {xmj } for which Rmj
is large. This gives the estimate in part (b). To establish (c) we use Proposition 5.10: Φ(x′)
is a sum of terms of the form
mj(cid:9), and we can estimate the
ZRM
∂j1 · · · ∂ja ϕj1,...,ja (x′, x′′)ψ(Rm1 · xm1 , . . . , Rmb · xmb ) dxm1 · · · dxmb
where ∂jk is a derivative with respect to a variable in xmk for each mk ∈ A. Integrating by
parts in the variables in A for which Rmk < 1 then gives the estimate in part (c), and this
completes the proof.
We will want to show that sums of dilates of normalized bump functions with weak can-
cellation satisfy the cancellation properties of the distributions K ∈ P0(E), and we will use
the following.
Lemma 5.12. Suppose that ϕI has weak cancellation with parameter ǫ > 0. Let L =
{l1, . . . , la} and M = {m1, . . . , mb} be complementary subsets of {k1, . . . , ks}, and let R =
{Rm1, . . . , Rmb} be positive real numbers. Write x ∈ RN as x = (x′, x′′) where x′ =
(xl1 , . . . , xla ) and x′′ = (xm1 , . . . , xmb ), and similarly write I = (I ′, I ′′). Let ψ ∈ C∞
0 (RM ) be
a normalized bump function. Then the function
can be written as a finite sum of terms, indexed by subsets A ⊂ {k1, . . . , ks}, of the form
ϕI (x′, x′′)ψ(cid:0)(2im1 Rm1) · xm1 , . . . , (2imb Rmb) · xmb(cid:1) dxm1 · · · dxmb .
2−ǫ[ij−e(j,τ (j))iτ (j) ]ih Yj∈M\A
minn(2jRj)+ǫ, (2jRj )−ǫoieϕI
A(x′).
A} ⊂ C∞
0 (RL) is normalized relative to ϕI .
ΦI
ψ,R(x′) =ZRM
h Yj∈A∩M
The family {eϕI
Proof. To make the exposition and the notation simpler, assume without loss of generality that
x′ = (xk1 , . . . , xka ) and x′′ = (xka+1 , . . . , xks ) and write I = (I ′, I ′′) with I ′ = (i1, . . . , ika)
and I ′′ = (ika+1, . . . , iks). According to Definition 5.9, since ϕI has weak cancellation,
ϕI (x′, x′′) = XA⊂{1,...,n}hYj∈A
2−ǫ[ij −e(j,τ (j))iτ (j) ]iϕI
A(x′, x′′).
Then ΦI
R,ψ can be written as a finite sum of terms, indexed by A ⊂ {1, . . . , n}, of the form
hYj∈A
2−ǫ[ij −e(j,τ (j))iτ (j) ]iZ ϕI
A(x′, x′′)ψ(cid:0)(2im1 Rm1) · xm1 , . . . , (2imb Rmb ) · xms(cid:1) dx′′.
Using Proposition 5.11, and incorporatingQj∈M\A 2−ǫ[ij−e(j,τ (j))iτ (j) ] into the bump function,
it follows this function can be written
h Yj∈A∩M
2−ǫ[ij−e(j,τ (j))iτ (j) ]ih Yj∈M\A
minn(2imj Rmj )+ǫ, (2imj Rmj )−ǫoieϕI
A(x′),
which completes the proof.
36
5.6 Dyadic sums with weak cancellation
Fix the decomposition RN = RC1 ⊕ · · · ⊕ RCn and let E be an n × n matrix satisfying
the basic hypothesis given in equation (2.10). Suppose that Γo(E) 6= ∅. Let ΓZ(E) be as in
equation (5.1).
Theorem 5.13. Let (cid:8)ϕI : I ∈ ΓZ(E)(cid:9) ⊂ S(RN ) be a uniformly bounded family.
function ϕI has weak cancellation with parameter ǫ > 0 relative to I, then K =PI∈ΓZ(E)[ϕI ]I
converges in the sense of distributions to a kernel K ∈ P0(E).
If each
The arguments are similar to those in the proofs of Theorem 6.8 and Propositions 6.9,
6.10, and 6.11 of [NRSW12]. The proof consists of two steps:
(1) If F ⊂ ΓZ(E) is a finite set, the distribution KF = PI∈F [ϕI ]I belongs to the class
P0(E) with constants independent of the set F .
(Lemma 5.6 shows that KF satisfies
the appropriate differential inequalities, so we only need to verify that KF satisfies the
cancellation conditions.)
Proof of Step 1.
limm→∞ KFm exists in the sense of distributions.
(2) If {Fm} is any increasing sequence of finite subsets of ΓZ(E) with S Fm = ΓZ(E), then
Let F ⊂ ΓZ(E) be a finite set and let KF = PI∈F [ϕI ]I . To verify the cancella-
tion conditions, let L = {l1, . . . , la} and M = {m1, . . . , mb} be complementary subsets of
{1, . . . , n}. If x = (x1, . . . , xn) ∈ RN write x = (x′, x′′) where x′ = (xl1 , . . . , xla ) ∈ RL and
x′′ = (xm1 , . . . , xmb) ∈ RM . Let ψ ∈ C∞
0 (RL) be a normalized bump function with support
in the unit ball, and let R = {Rl1, . . . , Rla} be positive real numbers. We must show that the
function
KF,ψ,R(x′′) =XI∈FZRM
[ϕI ]I (x1, . . . , xn)ψ(Rl1 · xl1 , . . . , Rla · xla ) dxl1 · · · dxla
satisfies the correct differential inequalities in the x′′ variables.
To simplify the notation, and without any real loss of generality, we will only consider
the case when L = {1, . . . , a} and M = {a + 1, . . . , n}, so that x′ = (x1, . . . , xa) and x′′ =
(xa+1, . . . , xn). Let E′′ be the sub-matrix of E consisting of those entries e(j, k) for which
a + 1 ≤ j, k ≤ n, and let
ΓZ(E′′) =n(ia+1, . . . , in) ∈ Zn−a : e(j, k)ik ≤ ij < 0, a + 1 ≤ j, k ≤ no.
For each I ′′ = (ia+1, . . . , in) ∈ ΓZ(E′′) there are only finitely many elements I ∈ ΓZ(E) which
have the same last n − a entries. Let F ′′ ⊂ ΓZ(E′′) be a finite set and suppose (without loss
of generality) that
For each I ′′ ∈ ΓZ(E′′) let
F =nI = (i1, . . . , in) ∈ ΓZ(E) : I ′′ = (ia+1, . . . , in) ∈ F ′′o.
Λ(I ′′) =nI ′ = (i1, . . . , ia) ∈ Za : (I ′, I ′′) ∈ ΓZ(E)o.
Then for each I ∈ F we can write I = (I ′, I ′′) with I ′′ ∈ F ′′ and I ′ ∈ Λ(I ′′).
Let ψ ∈ C∞
0 (RM ) be a normalized bump function with support in the unit ball, and let
R = {R1, . . . , Ra} be positive real numbers. To complete step 1 we need to estimate
KF,ψ,R(x′′) = XI ′′∈F ′′ XI ′∈Λ(I ′′)ZRM
[ϕI ]I (x1, . . . , xn)ψ(R1 · x1, . . . , Ra · xa) dx1 · · · dxa
37
and its derivatives. Let
ΦI
ψ,R(x′′) =ZRM
ZRM
[ϕI ]I (x1, . . . , xn)ψ(R1 · x1, . . . , Ra · xa) dx1 · · · dxa =(cid:2)Φ(I ′,I ′′)
ϕI (x1, . . . , xn)ψ(cid:0)(2i1 R1) · x1, . . . , (2iaRa) · xa) dx1 · · · dxa.
ψ,R (cid:3)I ′′(x′′)
Then
and so
KF,ψ,R(x′′) = XI ′′∈F ′h XI ′∈Λ(I ′′)
Φ(I ′,I ′′)
ψ,R iI ′′
(x′′).
ψ,R ∈ C∞
If we can show that for each I ′′ ∈ F ′′ the inner sumPI ′∈Λ(I ′′) Φ(I ′,I ′′)
eΦI ′′
converges to a function
0 (RL), and this family is uniformly bounded independent of the choice of F , then
the required differential inequality estimates for KF,ψ,R will follow from Lemma 5.6. But
according to Lemma 5.12, Φ(I ′,I ′′)
is a sum of normalized bump functions multiplied by
exponential gains. Thus the proof of step 1 will be complete if we can show that for each
fixed I ′′ ∈ F ′′ ⊂ ΓZ(E′′) and each subset A ⊂ {1, . . . , n},
ψ,R
ψ,R
X(i1,...,ia)∈Λ(I ′′)" aYj=1
j∈A
2−ǫ[ij −e(j,τA(j))iτA (j)]#" aYj=1
j /∈A
minn(2ij Rj)+ǫ, (2ij Rj)−ǫo#
converges independently of I ′′ and the finite set F ′′. Without loss of generality, we can assume
that {1, . . . , a} ∩ A = {1, . . . , b} and {1, . . . , a} \ A = {b + 1, . . . , a}. Thus we need to estimate
X(i1,...,ia)∈Λ(I ′′)" bYj=1
2−ǫ[ij −e(j,τA(j))iτA (j)]#" aYj=b+1
minn(2ij Rj)+ǫ, (2ij Rj)−ǫo#.
The required estimate follows from Proposition 16.2 in Appendix III.
Proof of Step 2.
We need to check that if {Fm ⊂ ΓZ(E)} is a sequence of finite sets withS Fm = ΓZ(E),
then the sequence(cid:8)KFm =PI∈Fm
[ϕI ]I(cid:9) converges in the sense of distributions to an element
has cancellation in xn, it follows from Proposition 5.11 that we can write ϕI =Pk∈Cn
of P0(E). The argument is similar to that on pages 667-668 of [NRSW12]. Since each ϕI
∂xk ϕI
k
0 (RN ) is normalized relative to ϕI . It follows that
where each ϕI
k ∈ C∞
[ϕI ]I = Xk∈Cn
[∂xk ϕI
k]I = Xk∈Cn
2iks dk ∂xk [ϕI
k]I .
Thus if ψ ∈ S∞
0 (RN ) is a test function, we can integrate by parts to get
(cid:10)KFm , ψ(cid:11) = Xk∈Cn(cid:10) XI∈Fm
= − Xk∈CnZRNh XI∈Fm
2ikdk ∂xk [ϕI
k]I , ψ(cid:11) = − Xk∈Cn(cid:10) XI∈Fm
2ikdk 2− Pn
k=1 ikQk ϕI
2ikdk [ϕI
k]I , ∂xk ψ(cid:11)
k(2−i1 · x1, . . . , 2−in · xn)i∂xk ψ(x) dx.
38
Let d = min(cid:8)dk : 1 ≤ k ≤ N(cid:9) > 0. Since {ϕI
from Proposition 16.3 that
k} is a normalized family in S(RN ), it follows
(cid:12)(cid:12)(cid:12) XI∈Fm
2ikdk 2− Pn
k=1 ikQk ϕI
≤ CM XI∈Γ(E)h nYk=1
nYk=1
≤ CM Nn(x)d
k(2−i1 · x1, . . . , 2−in · xn)(cid:12)(cid:12)(cid:12)
2−ik nk(xk)(cid:17)−M
2−ikQki2ikd(cid:16)1 +
nXj=1
nYj=1
Nk(x)−Qk ≤ CM Nn(x)d
Nj(x)−Qj .
It follows from Corollary 5.7 that Nn(x)d Qn
j=1 Nj(x)−Qj is locally integrable. Thus by the
Lebesgue dominated convergence theorem it follows that as m tends to infinity, hKFm , ψi
tends to
− Xk∈CnZRN h XI∈Fm
This completes the proof.
2ikdk 2− Pn
k=1 ikQk ϕI
k(2−i1 · x1, . . . , 2−in · xn)i∂xk ψ(x)dx.
6 Decomposition of multipliers and kernels
Let K ∈ P0(E). The object of this section is to decompose K into a finite sum of distributions,
each of which can be written as a dyadic sum of dilates of normalized Schwartz functions (or
normalized bump functions) with appropriate cancellation. It is difficult to directly decompose
K and maintain the cancellation, so we work on the Fourier transform side. The effect of this
procedure is that the following steps will be governed by the decomposition of B(1)c in the
ξ-space according to the sets bES, while the decomposition of B(1) in the x-space will no longer
play any role.
6.1 New matrices ES
Corollary 3.10 we obtain the following equivalent condition.
Let S′(E) denote the set of marked partitions S such that bES ∩ B(1)c is non-empty. From
Corollary 6.1. Let S =(cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) ∈ S(n). Then bES ∩ B(1)c 6= ∅ if and only if
bΓS =na = (a1, . . . , as) ∈ Rs : τS(kr, kp)ap ≤ ar < 0 for all 1 ≤ p, r ≤ so,
has non-empty interior.
the cone
(6.1)
With the decomposition introduced in Section 3.5 and in the notation of Section 5.3,
for each S ∈ S′(E) we want to define new norms on RN which describe the behaviour of
multipliers on the set bES ∩ B(1)c in the same way that the original norms bN1(t), . . . , bNn(t)
describe it on the principal region bES0 ∩ B(1)c.
Let S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E). On each RIr we put the norm
nj(ξj)1/τS(j,kr ),
bnS,r(ξIr ) = Xj∈Ir
nj(ξj)e(kr ,j) = Xj∈Ir
39
defined in (3.5) and homogeneous with respect to the dilations ·S.
We now want an s×s matrix ES satisfying the basic hypotheses and such that Γ(ES) =bΓS.
We cannot simply use the coefficients τS(kr, kp) because in general they do not satisfy the
inequalities τS(kr, kp) ≤ τS(kr, ku) τS(ku, kp).
Luckily we can replace the coefficients τS(kp, kr) by new ones which do satisfy the basic
hypothesis. According to Lemma 14.2 in Appendix I, if the cone is non-empty, the coefficients
in the inequalities defining the cone can be replaced by new coefficients which satisfy the basic
hypothesis. Thus we have the following result.
Lemma 6.2. Let S = (cid:8)(I1, k1); . . . , (Is, ks)(cid:9) ∈ S′(E). There is a unique s × s matrix
ES = {eS(r, p)} such that eS(r, p) ≤ τS(kr, kp), and
Γ(ES) =na ∈ Rs : eS(r, p)ap ≤ ar < 0o =bΓS.
According to (5.4), we introduce global norms NS,r and bNS,r.
Definition 6.3. Let S =(cid:8)(I1, k1); . . . , (Is, ks)(cid:9) ∈ S′(E). Set
sXp=1Xk∈Ip
sXp=1Xk∈Ip
sXp=1bnS,p(xIp )eS (r,p) ≈
sXp=1bnS,p(ξIp )1/eS (p,r) ≈
bNS,r(ξ) =
NS,r(x) =
nk(xk)e(kp,k)eS (r,p),
nk(ξk)e(kp,k)/eS (p,r).
(6.2)
These norms coincide with the norms defined in (5.4), relative to the partition IS =
{I1, . . . , Is} of {1, . . . , n}, to the matrix ES and to the vector of exponents αS = (α1, . . . , αn)
with αj = e(kr, j) if j ∈ Ir.
We will use the shortened notation M∞(ES) instead of M∞(ES, IS, αS).
Proposition 6.4. We have the inclusion M∞(ES) ⊂ M∞(E).
Proof. By Corollary 5.4, it is sufficient to check that the entries of the matrix E♯
greater than the corresponding entries of the matrix E. By (5.5), if j ∈ Ir and k ∈ Ip,
S are not
e♯
S(j, k) =
e(kp, k)
e(kr, j)
eS(r, p) ≤
e(kp, k)
e(kr, j)
τS(kr, kp) ≤
e(kp, k)
e(kr, j)
e(kr, k)
e(kp, k)
≤ e(j, k).
6.2 Road map for the dyadic decomposition
(1) In Section 6.3 we will construct a partition of unity on RN consisting of functions Ψ0 and
0 (RN ) has support in B(2), and where each ΨS ∈ M∞(E)
Let m = bK be the Fourier transform so that m ∈ M∞(E). We proceed as follows.
and has support in bEA
S ∩ B(1)c. Here A > 1 depends only on the matrix E.
{ΨS}S∈S ′(E), where Ψ0 ∈ C∞
We will write
Then m0 ∈ C∞
m = Ψ0m + XS∈S ′(E)
ΨSm = m0 + XS∈S ′(E)
0 (B(2)) and each mS ∈ M∞(E) is supported in bEA
mS.
S ∩ B(1)c.
40
0 (RN ) is a uniformly bounded family with supports in the
(2) For each S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E), we show that
mS(ξI1 , . . . , ξIs ) = XJ∈Γ(ES)
where(cid:8)mJ
set where 4−1 <bnS,r(ξIr ) < 4 for 1 ≤ r ≤ s.
S : J ∈ Γ(ES)(cid:9) ⊂ C∞
S]J (ξI1 , . . . , ξIs ) = mJ
(3) Writing [mJ
mJ
S(2−j1 ·SξI1 , . . . , 2−js ·SξIs)
S(2−j1 ·SξI1 , . . . , 2−js ·SξIs ), it then follows that
m(ξ) = m0(ξ) + XS∈S ′([n]) XJ∈Γ(ES)
[mJ
S]J (ξ).
This is done in Section 6.4.
(4) Taking inverse Fourier transform, denoted by ∨, we obtain
K = m∨
0 + XS∈S ′([n]) XJ∈Γ(ES )(cid:0)[mJ
S]J(cid:1)∨
.
This gives a decomposition of K into sums of dilates of Schwartz functions, which in turn
can be written as sums of dilates of normalized bump functions. (See Lemma 6.15 below.)
6.3 Partitions of unity
In this section, we construct a partition of unity on the 'Fourier transform side'. We
j=1 RCj and E = {e(j, k)} is an
RCj .
l=1 nl(ξl)1/e(l,j) which might not be differentiable, we choose a
begin by slightly modifying our notation. As usual, RN =Ln
n × n matrix satisfying (2.10). Let S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E) and RIr =Lj∈Ir
Instead of taking bNj(ξ) =Pn
nXl=1
smooth version of this norm. Then
nj(ξj)e(kr ,j),
bnS,r(ξIr ) ≈ Xj∈Ir
bQS,r = Xj∈Ir
Qje(kr, j)−1,
nl(ξl)1/e(l,j),
bNj(ξ) ≈
λ·S ξIr =(cid:8)λ1/e(kr ,j) · ξj : j ∈ Ir(cid:9),
j∈Ir(cid:8)e(kp, j)/e(kr, j)(cid:9).
τS(kp, kr) = min
S be the set introduced in Definition 3.5 for A ≥ 1. Recall from Lemma 3.9 that there
S ∩ B(1)c =⇒(nj(ξj)e(kr ,j) < Ankr (ξkr )
∀j 6= kp in Ir
nkp (ξkp ) < Ankr (ξkr )τS(kp,kr) ∀1 ≤ p 6= r ≤ s
Let bEA
is a constant η > 1 so that
ξ ∈ bEA
S ∩ B(1)c.
=⇒ ξ ∈ bEAη
Now pick ψ ∈ C∞(R) and χ ∈ C∞(RN ) such that
ψ(t) =(1
0
if t ≤ 2n
if t ≥ 4n
,
and χ(ξ) =(0
1
if ξ ≤ 1
if ξ ≥ 3
2
.
41
For S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E), set
sYr=1Yj∈Ir
S (ξ) = χ(ξ)
Ψ#
Proposition 6.5.
(a) Let bES be the closure of bES. Then Ψ♯
(b) There is a constant A > 1 so that Ψ♯
(c) Ψ♯
S ∈ C∞(RN ).
(6.3)
ψ(cid:0)bNj(ξ)nkr (ξkr )−1/e(kr ,j)(cid:1).
S(ξ) ≡ 1 in an open neighborhood of the set bES \B(2).
S is supported in bEA
S \ B(1) .
Proof. Suppose that ξ ∈ bES \ B(2). Then χ(η) = 1 for η in a neighborhood of ξ. Moreover,
nl(ξl)1/e(l,j) ≤ nkr (ξkr )1/e(kr ,j) for every j ∈ Ir, and so
nl(ξl)1/e(l,j).
the same is true for Ψ♯
bNj(ξ)nkr (ξkr )−1/e(kr ,j) ≈ nkr (ξkr )−1/e(kr ,j)
It follows that ψ(cid:0)bNj(η)nkr (ηkr )−1/e(kr ,j)(cid:1) = 1 for η in an open neighborhood of ξ, and so
S (ξ) 6= 0. Then χ(ξ) 6= 0, and ψ(cid:0)bNj(ξ)nkr (ξkr )−1/e(kr ,j)(cid:1) 6= 0 for 1 ≤
, and since nj(ξ) . bNj(ξj) and nkp (ξkp ) .
bNj(ξ)e(kp,j), it follows from the assumption on the support of ψ that
r ≤ s and every j ∈ Ir. This shows that ξ ∈ B(1)
nj(ξj) . nkr (ξkr )1/e(kr ,j)
Next suppose Ψ#
for j ∈ Ir,
S(η).
c
nXl=1
nkp (ξkp ) . nkr (ξkr )e(kp,j)/e(kr ,j)
for 1 ≤ p, r ≤ s.
S is contained in the set bEAη
It follows that there exists a constant A > 0 (depending on n and the matrix E) so that
nj(ξj) ≤ A nkr (ξkr )1/e(kr ,j) and nkp (ξkp ) ≤ A nkr (ξkr )τS (kp,kr). But this means that the
support of Ψ♯
S where η is the constant from Lemma 3.9.
Finally, if ξ /∈ B(1) and if Ψ♯
the set where ξkr 6= 0, and it follows that Ψ#
S(ξ) 6= 0, then bNj(ξ) ≥ 1 for 1 ≤ j ≤ n and it follows that
nkr (ξkr ) is bounded away from 0. Since the function bNj(ξ)nkr (ξkr )−1/e(kr ,j) is smooth on
S ∈ M∞(bE); i.e. for every γ ∈ NC1 × · · · × NCn there is a constant Cγ > 0
Lemma 6.6. Ψ#
so that
S is smooth.
(cid:12)(cid:12)∂γ
ξ Ψ#
S (ξ)(cid:12)(cid:12) ≤ Cγ
nYj=1h1 + bNj(ξ)i−[[γ j ]]
.
(ξ1, . . . , ξn) → (λe(1,j) · ξ1, . . . , λe(n,j) · ξn).
Proof. The function bNj(ξ) is homogeneous of degree one relative to the dilations
and the function ρj,kr (ξ) = bNj(ξ)nkr (ξkr )−1/e(kr ,j) is homogeneous of degree zero. The
adapted length of γ isPn
homogeneous dimension of RCl is e(l, j)Ql, and if γ = (γ 1, . . . , γn) ∈ NC1 × · · · × NCn , the
l=1 e(l, j)[[γj]]. It follows from Proposition 15.1 in Appendix II that
for ξ ∈ B(1)c
∂γρj,kr (ξ) .h1 + bNj(ξ)i− Pn
l=1 e(l,j)[[γ l]]
.
42
But on the support of Ψ#
S which is contained in bEA
l=1 e(l,j)[[γ l]]
S ,
h1+bNj(ξ)i− Pn
=
sYp=1Yl∈Iph1 + bNj(ξ)i−[[γl]]e(l,j)
sYp=1Yl∈Iph1 + nkp (ξkp )i−[[γ l]]/e(kp,l)
≤
≤
sYp=1Yl∈Iph1 + nkp (ξkp )i−[[γ l]]e(l,j)/e(kp ,j)
.
sYp=1Yl∈Iph1 + Nl(ξ)i−[[γ l]]
since e(kp, j) ≤ e(kp, l)e(l, j), nkp (ξkp ) > 1, and Nl(ξ) ≈ nkp (ξkp )1/e(kp,l) on the support
of Ψ#
S if l ∈ Ip. The estimates of the Lemma now follow from the chain rule and product
rule.
Let Ψ#
0 be a compactly supported function which is identically equal to 1 on B(2). We
set
Ψ0 = Ψ#
0hΨ#
0 + XS∈S ′(E)
Ψ#
Si−1
,
ΨS = Ψ#
ShΨ#
0 + XS∈S ′(E)
Ψ#
Si−1
.
(6.4)
Using the fact that, if m ∈ M∞(E) is bounded away from zero, then also 1/m ∈ M∞(E),
we then have
Corollary 6.7.
(b) ΨS ∈ M∞(E) for S = 0 and for every S ∈ S′(E).
(a) Ψ0(ξ) +PS∈S ′(E) ΨS(ξ) ≡ 1 for every ξ ∈ RN .
(c) For S ∈ S′(E) the function ΨS is supported in bEA
S .
We now use Corollary 6.7 to obtain a decomposition of a multiplier m ∈ M∞(E).
Corollary 6.8. Let m ∈ M∞(E). Put m0 = Ψ0m, and mS = ΨSm for each S ∈ S′(E).
Then
S where A depends only on the matrix E;
(d) m0 ∈ C∞
0 (RN ) has support in B(2).
(a) m(ξ) = m0(ξ) +PS∈S ′(E) mS(ξ) for every ξ ∈ RN ;
(b) mS ∈ M∞(bE) for every S ∈ S′(E);
(c) mS is supported in bEA
Remark 6.9. Since mS is supported on bEA
sYr=1Yj∈Ir bNj(ξ)−[[γ j ]] ≈
sYr=1bnS,r(ξIr )
(cid:12)(cid:12)∂γ
ξ (mS)(ξ)(cid:12)(cid:12) .
S , we have
− P
j∈Ir
≈
sYr=1Yj∈Ir
[[γ j ]]/e(kr ,j)
.
nkr (ξkr )−[[γ j ]]e(kr ,j)
This says that the differential estimates for mS are equivalent to product multiplier estimates
relative to the norms {bnS,r} on the decomposition RN = RI1 ⊕ · · · ⊕ RIs .
All the functions mS belongs to the class M∞(E). If S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E)
is not the principal marked partition, one can say more: the function mS in fact belongs to
the class of multipliers M∞(ES) introduced in Section 6.1.
Proposition 6.10. For S ∈ S′(E), the function mS belongs to the class M∞(ES).
43
Proof. We must show that for any γ = (γI1 , . . . , γIs ) ∈ NI1 × · · · × NIS there is a constant
Cγ so that
[[γj]]/e(kr, j) is the length of the multi-index γIr adapted to the dila-
tions ·S. But this is a simple consequence of the estimate in Remark 6.9 and the fact that on
sYr=1(cid:0)1 + bNS,r(ξ)(cid:1)−[[γ Ir ]]S ,
(cid:12)(cid:12)∂γmS(ξ)(cid:12)(cid:12) ≤ Cγ
S we havebnS,r(ξIr ) ≈ 1 + bNS,r(ξ).
where [[γ Ir ]]S = Pj∈Ir
the set bEA
6.4 Dyadic decomposition of a multiplier
We now turn to the task of providing a dyadic decomposition for each multiplier mS. Fix
S = (cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E), and write RN = RI1 ⊕ · · · ⊕ RIs . On each factor RIr
choose a function θ ∈ C∞
0 (RIr ) satisfying12
Θj(ξIr ) = θj(ξIr ) − θj−1(ξIr ) = θ(2−j ·S ξIr ) − θ(2j−1 ·S ξIr ) = Θ0(2−j ·S ξIr ).
. Put θj(ξIr ) = θ(2−j ·S ξIr ), and
0
θ(ξIr ) =(1
Recall that 2−j ·S ξIr =(cid:8)2−j/e(kr ,j) · ξj(cid:9)j∈Ir
Note that Θj is supported wherebnS,r(ξIr ) ≈ 2j. Precisely:
if bnS,r(ξIr ) ≤ 1,
if bnS,r(ξIr ) ≥ 2.
For 0 6= ξIr ∈ RIr we have
Θj(ξIr ) 6= 0 =⇒ 2j−1 ≤bnS,r(ξIr ) ≤ 2j+2.
∞Xj=−∞
Θj(ξIr ) = θ0(ξIr ) +
Θj(ξIr ) = 1,
∞Xj=1
and the second equality holds for ξIr = 0 as well. For J = (j1, . . . , js) ∈ Zs define
ΘJ (ξ) = ΘJ (ξI1 , . . . , ξIs) = Θj1 (ξI1 ) · · · Θjs(ξIs).
Then ΘJ is supported where eachbnS,r(ξIr ) ≈ 2jr , andPJ∈Zs ΘJ (ξI1 , . . . , ξIs) ≡ 1 provided
each ξIr 6= 0. Recall that mS is supported on bEA
S ∩ B(1)c, and ξIr 6= 0 on this support. Thus
ΘJ (ξ)ΨS(ξ)m(ξ).
we can write
mS(ξ) = XJ∈Zs
ΘJ (ξ)mS(ξ) = XJ∈Zs
Proposition 6.11. There is a constant µ > 0 depending only on E so that if J ∈ Zs and
ΘJ (ξ)ΨS(ξ)m(ξ) 6= 0, then J belongs to
−ΓZ,µ(ES) =nJ = (j1, . . . , js) ∈ Zs : 0 < jp ≤ τS(kp, kr)jr + µo.
12To keep notation simpler, we do not indicate the dependence of θ on the index r.
44
Proof. Suppose ΘJ (ξ)ΨS(ξ)m(ξ) 6= 0. Then ΘJ (ξ) 6= 0, and it follows that 2jr−1 ≤
and so
S and so
for all r, p ∈ {1, . . . , s}. It follows that
bnS,r(ξIr ) ≤ 2jr+2. Also ΨS(ξ) 6= 0, and it follows that ξ ∈ bEA
1 ≤bnS,p(ξIp )1/τS (kp,kr) < AbnS,r(ξIr )
(cid:0)2jp−1(cid:1)1/τS (kp,kr)
≤bnS,p(ξIp )1/τS (kp,kr) < AbnS,r(ξIr ) ≤ A 2jr+1,
jp ≤ τS(kp, kr)jr +(cid:2)τS(kp, kr)(1 + log2(A)) + 1(cid:3) ≤ τS(kp, kr)jr + µ
mS(ξ) = XJ∈−ΓZ,µ(ES )
= XJ∈−ΓZ,µ(ES )
Θ0(2−j1 ·S ξI1 , . . . , 2−js ·S ξIs ) mS(ξI1 , . . . , ξIs ).
for an appropriate constant µ.
It now follows that
ΘJ (ξ)mS(ξ)
Put
mJ
S(ξ) = Θ0(ξ)mS(2j1 ·S ξI1 , . . . , 2js ·S ξIs)
so that we can write
family in C∞
mS(ξI1, . . . , ξIs) = XJ∈−ΓZ,µ(ES )
Proposition 6.12. The set of functions nmJ
Proof. Since mS is supported on bEA
∂γ(cid:0)mS(2j1 ·S ξI1 , . . . , 2js ·S ξIs )(cid:1)(ξ) = 2Ps
0 (RN ), each supported where 1
S ,
bounded away from zero (depending on γ).
But on the support of θ0 the product Qs
We have written mS(ξ) =PJ∈−ΓZ,µ(ES ) mJ
−ΓZ,µ(ES) by a sum over −ΓZ(ES).
mJ
S(2−j1·SξI1 , . . . , 2−js·SξIS ).
r=1 [[γ r ]]/e(kr ,j)
S : J ∈ −ΓZ,µ(ES)o is a uniformly bounded
4 <bnS,r(ξIr ) < 4 for 1 ≤ r ≤ s.
ξ (mS)(2j1 ·S ξI1 , . . . , 2js ·S ξIs)(cid:12)(cid:12)(cid:12)
r=1 [[γ r ]]/e(kr ,j) (cid:12)(cid:12)(cid:12)∂γ
sYr=1bnS,r(2jr ·S ξIr )−[[γ l]]/e(kr ,j)
sYr=1bnS,r(ξIr )−[[γ l]]/e(kr ,j)
r=1bnS,r(ξIr )−[[γ l]]/e(kr ,j) is uniformly bounded and
S(2−J ·S ξ). We want to replace the sum over
. 2Ps
= Cγ
For each J ∈ −ΓZ,µ(ES), choose J ′ = p(J) ∈ −ΓZ(ES) at minimal distance from J. It is
quite clear that there is a finite upper bound M for the number of J having the same p(J)
and that, if p(J) = J ′, then jr − j′
r ≤ c(µ) for each r = 1, . . . , s. It follows that we can
rewrite
S(2−J ·S ξ)
mJ
mS(ξ) = XJ∈−ΓZ,µ(ES )
= XJ ′∈−ΓZ(ES ) XJ:p(J)=J ′
= XJ ′∈−ΓZ(ES )emJ ′
S (2−J ′
mJ
S(2−J ·S ξ)
(6.5)
·S ξ) ,
45
mJ
S(2J ′−J ·S ξ)
where the functions
S (ξ) = XJ:p(J)=J ′
emJ ′
0 (RN ).
form a a uniformly bounded family in C∞
Applying this decomposition for each marked partition S ∈ S′(E) we have the following
decomposition for multipliers.
Theorem 6.13. Let m ∈ M∞(E). There is a uniformly bounded collection of functions
{mJ
0 (RN ) indexed by
S} ⊂ C∞
S ∈ S′(E) =(cid:8)T ∈ S(n) : bET ∩ B(1)c 6= ∅(cid:9)
and J ∈ ΓZ(ES) with the following properties.
(a) If S = (cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E) and J ∈ ΓZ(ES), the function mJ
where c−1 <bnS,r(ξIr ) < c for 1 ≤ r ≤ s and an appropriate constant c.
(b) The multiplier m can be written
S is supported
m(ξ) = m0(ξ) + XS∈S ′(E) XJ∈−ΓZ(ES )
mJ
S(2−j1 ·S ξi1 , . . . , 2−js ·S ξIs )
(6.6)
where m0 ∈ C∞
0 (RN ) is supported in B(2).
6.5 Dyadic decomposition of a kernel
We can now decompose a distribution K ∈ P0(E). We know from Theorem 4.2 that the
Fourier transform m = bK ∈ M∞(E), and by Theorem 6.13 we can write m as in equation
(6.6). Write
[ψ]J (ξI1 , . . . , ξIs ) = ψ(cid:0)2−j1 ·S ξI1 , . . . , 2−js ·S ξIs(cid:1)
so that
Then if ∨ denotes the inverse Fourier transform, we have
bK = m0 + XS∈S ′(E) XJ∈−ΓZ(ES )
+ XS∈S ′(E) XJ∈−ΓZ(ES )
∨
K = (m0)
[mJ
S]J .
([mJ
S]J )∨.
Computing the inverse Fourier transform of [mJ
S]J we have
([mJ
S]J )∨(xI1 , . . . , xIS ) = 2Ps
r=1 jr bQIr [mJ
S]∨(2j1 ·S xI1 , . . . , 2js ·S xIs ).
Since the family {mJ
{([mJ
S} is uniformly bounded in C∞
S]J )∨} is also a uniformly bounded family in S(RN ). Since each mJ
c−1 <bnS,r(ξIr ) < c for all r, and sincebnS,r(ξIr ) ≈ nkr (ξkr ), it follows that
S)∨(x1, . . . , xkr , . . . , xn) dxkr = 0
(mJ
0 (RN ) ⊂ S(RN ), it follows that the family
S is supported where
(6.7)
ZRCkr
for 1 ≤ r ≤ s; i.e. the function (mJ
Let us write
S)∨ has cancellation in each of the variables {xk1 , . . . , xks}.
ψ0 = (m0)
∨
,
ψJ
S = ([mJ
S]J )∨ = [(m−J
S )∨]−J .
(6.8)
46
Theorem 6.14. Let K ∈ P0(E). There is a function ψ0 ∈ S(RN ) and for each S ∈ S′(E)
there is a uniformly bounded collection of functions(cid:8)ψJ
(1) Let S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9) ∈ S′(E).
S(cid:9)J∈ΓZ(bES ) ⊂ S(RN ) so that:
(a) Each function ψJ
S has cancellation in each of the variables {xk1 , . . . , xks }, and so has
strong cancellation relative to the decomposition RN = RI1 ⊕ · · · ⊕ RIs.
(b) The series
KS(x) = XJ∈ΓZ(ES )(cid:2)ψJ
S(cid:3)J (x) = XJ∈ΓZ(ES )
2− Ps
r=1 ir bQIr ψJ
S(2−J·Sx)
converges to an element of P0(ES) ⊂ P0(E) in the sense of distributions .
(2) The distribution K decomposes as K(x) = ψ0(x) + XS∈S ′(E)
Proof. Let m = bK ∈ M∞(E). Using Theorem 6.13, write
m = m0 + XS∈S ′(E) XJ∈−ΓZ(ES )
[mJ
S]J .
KS(x).
Then for each S ∈ S′(E) and each J ∈ ΓZ(ES), let ψJ
observed in equation (6.7) that ψJ
gives assertion (1a).
S )∨(x). We have already
S has cancellation in each variable xkr , 1 ≤ r ≤ s, which
S (x) = (m−J
Next
KS = XJ∈ΓZ(ES)
[ψJ
S ]J = XJ∈−ΓZ(ES)
[ψ−J
S ]−J = XJ∈−ΓZ(ES )
([mJ
S]J )∨ = m∨
S.
According to Proposition 6.10, mS ∈ M∞(ES), and using Theorem 4.1 it follows that KS ∈
P0(ES). But according to Lemma 6.4, M∞(ES) ⊂ M∞(E), and every distribution in P0(ES)
belongs to P0(E). This establishes assertion (1b).
Finally assertion (2) is then a consequence of the identity m = m0 +PS∈S ′(E) mS, and
this completes the proof.
Theorem 6.14 shows that each kernel K can be written as sums of dilates of uniformly
bounded families of Schwartz functions. As in [NRSW12], we can improve this result and show
that we can replace Schwartz functions by uniformly bounded families of functions compactly
supported in the unit ball. We use the following result, which is Lemma 6.5 in [NRSW12].
Lemma 6.15. Let ψ ∈ S(RN ). Then there are functions {ϕk} ⊂ C∞
0 (RN ), k ∈ N such that
ψ(x) =
∞Xk=0
2−kQϕk(2−k · x)
where Q is the homogeneous dimension of RN relative to the given dilations. The functions
{ϕk} have the following properties.
(a) Each ϕk is supported in B(1).
(b) For any δ > 0 and any γ ∈ NN there exists a positive integer M so that
sup
x∈RN
∂γϕk(x) ≤ 2−kδ sup
x∈RN Xα≤M
∂αψ(x).
(c) If ψ has strong cancellation, then each ϕk also has strong cancellation.
47
S ]J (x), and each ψJ
S has cancellation in the variables
{xk1 , . . . , xks }. We apply Lemma 6.15 to each function ψJ
S:
Now let K ∈ P0(E) and let K = ψ0 +PS∈S ′(E) KS be the decomposition given in Theo-
rem 6.14 where KS(x) =PJ∈ΓZ(ES )[ψJ
∞Xk=0
S (2−k · x)
2−kQS ϕJ,k
S(x) =
ψJ
where each {ϕJ,k
each having cancellation in the variables {xk1 , . . . , xks }. Then
S } ⊂ C∞
0 (RN ) is a uniformly bounded family of functions supported in B(1),
KS(x) = XJ∈ΓZ(ES)" ∞Xk=0
KS(x) = XJ∈ΓZ(ES)
[ϕJ,k
S ]k#J
(x).
S]J (x)
[eϕJ
The argument on pages 661-663 of [NRSW12] shows that this last sum can be rewritten to
yield
S} is a uniformly bounded family in C∞
0 (RN ), all supported in B(1), and all having
cancellation in the variables {xk1 , . . . , xks }. Thus we have
Corollary 6.16. Let K ∈ P0(E). For each S ∈ S′(E) there is a uniformly bounded collection
where {eϕJ
of functions(cid:8)ϕJ
S(cid:9)J∈ΓZ(ES ) ⊂ C∞
0 (RN ) and a function ψ0 ∈ S(RN ) so that
(a) Each function ϕJ
S has cancellation in the variables xk1 , . . . , xks.
(b) Let
where
element KS ∈ P0(ES) ⊂ P0(E).
Then the series KS = PJ∈ΓZ(ES )(cid:2)ϕJ
(c) K(x) = ψ0(x) +PS∈S ′(E) KS(x).
[ϕJ
S]J (x) = 2−ϕJ
S(2−j1 ·S xI1 , . . . , 2−js ·S xIs ),
2−jr ·S xIr =n2−jr /e(kr ,j) · xjoj∈Ir
.
S(cid:3)J converges in the sense of distributions to an
7 The rank of E and integrability at infinity
So far we have analyzed kernel in the restricted class P0(E), which are the sum of a kernel in
P(E) with compact support and a Schwartz function.
In this Section we discuss the nature of a distribution K ∈ P(E) and show that it depends
in the first place on the rank of the matrix E = {e(j, k)}.
We will see that if the rank equals 1 then the distribution K is a standard (non-isotropic)
Calder´on-Zygmund kernel. However, if the rank is greater than 1, we will see that the kernel
48
K is integrable at infinity, but has worse behavior near the origin than a Calder´on-Zygmund
kernel.
We will also show that we have integrability at infinity in a more general situation. Sup-
pose that the Fourier transform m of a distribution K satisfies the differential inequalities in
Definition 2.4 for pure derivatives; i.e. for every 1 ≤ j ≤ n and for every γj ∈ NCj there is a
m(ξ) ≤ Cγ j Nj(ξ)−[[γ j ]]. Then if the rank of E is greater than
constant Cγ j > 0 so that ∂
1, the function K is integrable at infinity. This result will be used in Section 11 below when
we consider distributions which are simultaneously flag kernels for two different flags.
γj
ξj
7.1 The rank 1 case: Calder´on-Zygmund kernels
We begin by recalling the definition of a Calder´on-Zygmund kernel13. Start with the usual
decomposition
and use the standard notation of Section 2. For a = (a1, . . . , an) an n-tuple of positive
numbers, define a one-parameter family of dilations by setting
RN = RC1 ⊕ · · · ⊕ RCn ,
(7.1)
λ ·a x = (λa1 · x1, . . . , λan · xn).
With this homogeneity, RN has homogeneous dimension Qa =Pn
Na(x) = n1(x1)1/a1 + · · · + nn(xn)1/an
j=1 ajQj, and
(7.2)
is a homogeneous norm. We can then define the class of Calder´on-Zygmund kernels with this
homogeneity.
Definition 7.1. CZa is the space of tempered distributions K on RN such that
(a) away from the origin, K is given by integration against a smooth function K satisfying
the differential inequalities
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ CγNa(x)−Qa−[[γ]]a ,
j=1 aj[[γj]];
[NRS01], that, for distributions K ∈ CZa, the corresponding multi-
0 (RN ) with
Put
It is well known, cf.
(b) there is a constant C > 0 so that for any normalized bump function ψ ∈ C∞
for every multi-index γ, with [[γ]]a =Pn
support in the unit ball and for any R > 0,(cid:12)(cid:12)(cid:10)K, ψR(cid:11)(cid:12)(cid:12) ≤ C, where ψR(x) = ψ(R ·a x).
pliers m = bK are characterized by the differential inequalities
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ CγNa(ξ)−[[γ]]a .
Ea =
Then these differential inequalities can be written as(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ CγQn
the multipliers are the elements of the class M(Ea) where
aj
a1 + · · · + nn(ξ)
aj
an ≈ Na(ξ)aj .
Nj(ξ) = n1(ξ)
· · ·
· · ·
. . .
· · ·
a1
a2
1
...
an
a2
a1
a3
a2
a3
...
an
a3
1
a2
a1
...
an
a1
a1
an
a2
an
...
1
.
j=1 Nj(ξ)−[[γ j ]], and so
(7.3)
13For reasons of compatibility with the framework of this paper, we restrict ourselves to kernels and multi-
pliers which are C∞ away from the origin and to dilations which are compatible with the decomposition 7.1
and with the dilations in (2.5).
49
Proposition 7.2. Let E = {e(j, k)} satisfy the basic hypotheses (2.10). The following are
equivalent:
(1) rank(E) = 1;
(2) e(j, k) = e(j, l)e(l, k) for all 1 ≤ j, k, l ≤ n;
(3) there is a dilation structure on RN , compatible with the decomposition (7.1), so that if
m ∈ M(E), then m is a Mihlin-Hormander multiplier relative to that structure.
(4) there is a dilation structure on RN , compatible with the decomposition (7.1), so that if
K ∈ P(E), then K is a Calder´on-Zygmund kernel relative to that structure;
Proof. The rank of E equals 1 if and only if all the rows are proportional. Due to the 1's along
the diagonal this means that, for every j 6= k the j-th row equals the k-th row multiplied by
e(j, k). Thus (1) and (2) are equivalent.
all j. Moreover, in analogy with (5.6),
If (2) holds, we have in particular that e(j, k)e(k, j) = 1 for all j, k. Hence bNj = Nj for
Nj(x) ≈ N1(x)e(j,1).
Thus the differential inequalities in Definition 2.4 can be written
N1(ξ)−e(j,1)[[γ j ]] = N1(ξ)− Pn
j=1 e(j,1)[[γ j ]]
nYj=1 bNj(ξ)−[[γ j ]]
≈
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ Cγ
nYj=1
j=1 e(j, 1)[[γj]] =Pn
= N1(ξ)−[[γ]],
where [[γ]] =Pn
j=1 e(1, j)−1[[γj]] is the length of γ in the dilation structure
that makes N1 a homogeneous norm. Thus the multiplier m(ξ) satisfies the Mihlin-Hormander
differential inequalities relative to the family of dilations, and so (3) holds.
Conversely, assume that (3) holds. Then there are exponents β1 = 1, β2, . . . , βn such that,
setting
every m ∈ M(E) satisfies the inequalities
nj(ξ)βj ,
N (ξ) =
nXj=1
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) . N (ξ)− Pn
j=1 β−1
j
[[γ j ]].
This is equivalent to saying that m ∈ M(E′) with
e′(j, k) =
βk
βj
.
Hence in our hypothesis we have the inclusion M(E) ⊆ M(E′), which implies that M∞(E) ⊆
M∞(E′). By Proposition 5.2, this implies that Γ(E) ⊆ Γ(E′). By Lemma 14.9, the dimension
of Γ(E′) equals the reduced rank of E′, which is 1. Hence also Γ(E) has dimension 1, so it
must coincide with E′. This proves that (3) implies (1).
Finally, the equivalence of (3) and (4) is proved in [NRS01].
50
7.2 Higher rank and integrability at infinity
In this section let E = {e(j, k)} be an n × n matrix satisfying the basic hypothesis (2.10),
k=1 nk(ξk)1/e(k,j). Let Qj be
the homogeneous dimension of RCj . We begin by studying the integrability at infinity of a
measurable function K on RN which satisfies the inequality
k=1 nk(xk)e(j,k) and bNj(ξ) = Pn
and as usual let Nj(x) = Pn
K(x) ≤ C
Nj(x)−Qj .
nYj=1
(7.4)
Note that nothing is said about the size of the derivatives of K.
Lemma 7.3. Suppose that the rank of E is greater than or equal to 2. If K satisfies the
inequality in equation (7.4), then K is integrable over the complement of the unit ball B(1).
Proof. Since the rank of E is strictly larger than 1, it follows from Proposition 7.2 that
there are integers j, k, l ∈ {1, . . . , n} so that e(j, k) < e(j, l)e(l, k). Moreover, for every
k ∈ {1, . . . , n}, we have
(7.5)
K(x) ≤ C
Nj(x)−Qj ≤ C
nYj=1
j=1 e(j,k)Qj = Cxk− eQk
xk−e(j,k)Qj = Cxk− Pn
where ω ranges over the collection of non-empty subsets of {1, . . . , n}, and
j=1 e(j, k)Qj. Decompose the complement of B(1) as the union of sets Eω
nYj=1
where eQk = Pn
Fix ω, and let {θl}l∈ω be strictly positive numbers such that Pl∈ω θl = 1. Then it follows
from (7.5) that K(x) .Ql∈ω xl−θl eQl
θleQl > Ql for every l, then
ZEω
K(x) dx .Yl∈ω
Eω = {x ∈ RN : xk ≥ 1 for k ∈ ω, xk < 1 for k /∈ ω}.
, and if we can choose the constants {θl} such that
xl−θl eQl dxl < +∞.
Zxl∈RCl
l
xl≥1
e(j, k)/e(l, k) and so
We can choose {θl} with the required property if Pl∈ω QleQ−1
for every subset ω if Pn
e(l, k) Qlh nXj=1
nXl=1
l=1 QleQ−1
Qlh nXj=1
e(j, l)Qji−1
nXl=1
nXl=1
l =
≤
QleQ−1
l < 1, and we can do this
l < 1. But for any fixed k ∈ {1, . . . , n}, we have e(j, l) ≥
e(j, k)Qji−1
= 1.
(7.6)
Moreover, we can make the inequality in (7.6) strict if there exists a single triple j, k, l ∈
{1, . . . , n} such that e(j, k) < e(j, l)e(l, k). This completes the proof.
Corollary 7.4. Suppose that rank (E) ≥ 2. If K ∈ P(E) we can write K = K0 + K∞ where
K0, K∞ ∈ P(E), K0 is a distribution with compact support in the unit ball, and K∞ is a
distribution given by integration against a function K∞ ∈ C∞(RN ) ∩ L1(RN ) which satisfies
the differential inequalities of Definition 2.2.
We also will need the following variant of Lemma 7.3.
51
Lemma 7.5. Suppose that K is a tempered distribution on Rn and suppose that its Fourier
transform m = bK is a smooth function away from zero which satisfies the following differential
inequalities for pure derivatives: for 1 ≤ j ≤ n and for every γ j ∈ NCj there exists a constant
Cγ j > 0 so that
γ j
ξj
(cid:12)(cid:12)∂
m(ξ)(cid:12)(cid:12) ≤ Cγ j bNj(ξ)−[[γ j ]].
RN and bNj(t) =Pn
Pn
j=1 e(j, k)Qj. Choose χ ∈ C∞
Then for all γ j ∈ NCj , the product and chain rules show that we have
Then K is given by integration against a smooth function K away from the origin, and if the
rank of the matrix E is greater than 1, then K is integrable at infinity.14
Proof. If we can show that K is given by integration agains a function K with K(x) .
nk(xk)− Pn
j=1 e(j,k)Qj for 1 ≤ k ≤ n, then the argument proceeds as in the proof of Lemma 7.3.
Thus fix k, and define λ·k x = (λe(1,k) ·x1, . . . , λe(n,k) ·xn). This is a family of dilations on all of
k=1 nk(tk)1/e(k,j) is a homogeneous norm. The homogeneous dimension is
0 (RN ) with χ(ξ) ≡ 1 if bNj(ξ) ≤ 1. Let mR(ξ) = χ(R·j ξ)m(ξ).
(cid:12)(cid:12)∂
mR(ξ)(cid:12)(cid:12) ≤ Cγ j bNj(ξ)−[[γ j ]].
γ j
ξj
Now mR → m in the sense of distributions as R → 0, and hence K is the limit, in the sense
of distributions, of the inverse Fourier transform of mR. But it now follows from Proposition
15.4 in Appendix II that K(x) . nk(xk)− Pn
j=1 e(j,k)Qj , and this completes the proof.
7.3 Higher rank and weak-type estimates near zero
If K is the smooth function on RN \ {0} corresponding to a Calder´on-Zygmund kernel K,
then K satisfies the estimate(cid:12)(cid:12)(cid:8)x ∈ RN : K(x) > λ(cid:9)(cid:12)(cid:12) . λ−1. Moreover, there are Calder´on-
Zygmund kernels K such that(cid:12)(cid:12)(cid:8)x ∈ RN : K(x) > λ(cid:9)(cid:12)(cid:12) & λ−1. In this section we show that
the corresponding estimates for kernels K ∈ P(E) associated with the decomposition RN =
RC1 ⊕ · · · ⊕ RCn depend on the reduced rank of E. Recall from the discussion in Section 5.3
and Proposition 5.5 that if the reduced rank of E is m < n then there is an m × m matrix
E♭ so that P0(E) coincides with the space of distributions P0(E♭) associated with a coarser
decomposition decomposition RN = RA1 ⊕ · · · ⊕ RAm
Lemma 7.6. Suppose that the reduced rank of E is m.
(a) If K ∈ P(E), then for λ ≥ 1,(cid:12)(cid:12)(cid:12)nx ∈ RN : K(x) > λo(cid:12)(cid:12)(cid:12) . λ−1 log(λ)m−1.
(b) There exists K ∈ P(E) so that(cid:12)(cid:12)(cid:12)nx ∈ RN : K(x) > λo(cid:12)(cid:12)(cid:12) & λ−1 log(λ)m−1 for λ ≥ 1.
If m = 1, we are dealing with a Calder´on-Zygmund kernel and these facts are well-known, so
we may assume m > 1.
Proof of Lemma 7.6, part (a). As indicated above, we can assume that m = n by replacing
E by its reduced matrix and applying Proposition 5.5. Suppose that λ ≥ 1. For 1 ≤ k ≤ n
l=1 nl(xl)Ql . Since
l=1 e(l,k)Ql , and alsoQn
j=1 Nj(x)Qj ≥Qn
j=1 Nj(x)Qj ≥ nk(xk)Pn
j=1 Nj(x)−Qj , it follows that
we haveQn
K(x) ≤Qn
nx ∈ RN : K(x) > λo ⊂(x ∈ RN :(Qn
l=1 nl(xl)Ql < λ−1 and
nl(xl)Ql < λ−Ql/ Pn
k=1 e(k,l)Qk
1 ≤ l ≤ n).
14Note that we only make an hypothesis about "pure" derivatives of m. We are not assuming that m ∈
M(E).
52
greater than 1. Part (a) of the Lemma then follows from the following calculation.
k=1 e(k, l)Qki−1
Note that, as in equation (7.6),Pn
Proposition 7.7. Let A1, . . . , An, δ ∈ (0, ∞) and put A =Qn
l=1 QlhPn
j=1 Aj. Put
< 1 since the rank of E is strictly
E(A1, . . . , An, δ) =nx ∈ RN :
nYj=1
nj(xj )Qj < δ and nj(xj )Qj < Aj for 1 ≤ j ≤ no.
There is a constant cn depending only on n so that
(cid:12)(cid:12)E(A1, . . . , An, δ)(cid:12)(cid:12) & cn(A
δh1 +(cid:0)log(cid:0) A
δ(cid:1)(cid:1)n−1i
if A < δ,
if A ≥ δ
Proof. We argue by induction on n, with the case n = 1 following from Proposition 15.2 in
Appendix II. Thus suppose the result is true for n − 1. We divide the region E(A1, . . . , An, δ)
into two parts; the part where nn(xn)Qn < δ[A1 · · · An−1]−1 is a product of n balls, and its
volume is δ. For the region where nn(xn)Qn ≥ δ[A1 · · · An−1]−1 we can use the induction
hypothesis in the variables {x1, . . . , xn−1}.
k=1 θk(tk).
Proof of Lemma 7.6, part (b). As before we assume that E is reduced, i.e., m = n. Recall that
λ · x = (λd1 x1, . . . , λdN xN ). Choose a closed interval Jk ⊂ (1, 2dk) ⊂ R, and let θk ∈ C∞
0 (R)
be an odd function such that θk(t) ≥ 0 if t ≥ 0, θk(t) ≡ 0 if t /∈ (1, 2dk ), and θk(t) ≡ 1 if
The function Θ has strong cancellation in the sense of Definition 5.8.
It follows from
K ∈ P0(E). Note that the supports of the functions [Θ]I are disjoint. Let Σ denote the
set where Θ = 1, so that if 2−I · x ∈ Σ and 2− Pn
j=1 ij Qj > λ we have K(x) > λ. Let
t ∈ Jk. ThenRR θk(t) dt = 0 and if we put θk,m(t) = θk(2mdkt), the functions θk,m1 and θk,m2
have disjoint supports if m1 6= m2. Put Θ(t) =QN
Theorem 5.13 that K =PI∈ΓZ(E)[Θ]I converges in the sense of distributions to an element
j=1 ij Qj(cid:12)(cid:12)Σ(cid:12)(cid:12). We have
ΣI = {x ∈ RN : 2−I · x ∈ Σ}. Then(cid:12)(cid:12)ΣI(cid:12)(cid:12) = 2+ Pn
nx ∈ RN :(cid:16)∃I ∈ ΓZ(E)(cid:17)(cid:16)2−I·x ∈ Σ and 2− Pn
j=1 ij Qj > λ(cid:17)o ⊂ {x ∈ RN : K(x) > λ},
(cid:12)(cid:12)(cid:12)nx ∈ RN : K(x) > λo(cid:12)(cid:12)(cid:12) ≥(cid:12)(cid:12)(cid:12)nx ∈ RN :(cid:16)∃I ∈ ΓZ(E)(cid:17)(cid:16)2−I·x ∈ Σ and 2− Pn
j=1 ij Qj > λ(cid:17)o(cid:12)(cid:12)(cid:12)
with C a large constant. The cone Γ(E) =nt = (t1, . . . , tn) : e(j, k)tk < tj < 0o is open since
where ΓZ(E, λ) is the set of n-tuples I = (i1, . . . , in) ∈ ΓZ(E) such that λ < 2− Pn
ΣI & λ−1#(cid:16)ΓZ(E, λ)(cid:17)
E has rank n, and the number of lattice points (i1, . . . , in) in this cone satisfying
& XI∈ΓZ(E,λ)
j=1 ij Qj ≤ Cλ
so
log(λ) <
Qjij < log(λ) + 1
nXj=1
is of the order of(cid:0) log(λ)(cid:1)n−1
the proof of part (b).
. This shows that #(cid:16)ΓZ(E, λ)(cid:17) &(cid:0) log(λ)(cid:1)n−1
which completes
53
8 Convolution operators on homogeneous nilpotent Lie
groups
We now turn to the study of convolution operators f → f ∗ K with K ∈ P0(E) and the
convolution on a homogeneous nilpotent Lie group G. The underlying manifold of G is RN
for some integer N ≥ 1, and there is a one-parameter group of automorphic dilations, i.e., of
automorphisms δr : G → G which, in appropriate coordinates on G, take the form
δr(x1, . . . , xN ) = (rd1 x1, . . . , rdN xN ),
(8.1)
with di > 0 for every i. The ordering of the coordinates and the parametrization by r can be
chosen so that
1 ≤ d1 ≤ d2 ≤ · · · ≤ dN .
(8.2)
In these coordinates, the product on G has the following form. For x = (x1, . . . , xN ), y =
(y1, . . . , yN ),
xy =(cid:0)x1 + y1, x2 + y2 + M2(x, y), . . . , xN + yN + MN (x, y)(cid:1),
where M2, . . . , MN are polynomials which vanish for x or y equal to 0 and are such that
(8.3)
Ml(δrx, δry) = δrMl(x, y).
(8.4)
In particular, each Ml only depends on the variables xm, ym for which dm < dl. Moreover,
the Haar measure for G is Lebesgue measure on RN .
If f, g ∈ L1(G), the convolution f ∗ g is defined by
f ∗ g(x) =ZG∼= RN
f (xy−1)g(y) dy =ZG∼=RN
f (y)g(y−1x) dy.
8.1 Convolution of scaled bump functions: compatibility of dila-
tions and convolution
If f ∈ L1(G) and λ = (λ1, . . . , λN ) with each λj > 0, consider the N -parameter family of
dilations of f by λ given by
fλ(x1, . . . , xN ) =h NYj=1
λ−dj
j
i f (λ−d1
1 x1, . . . , λ−dN
N xN ),
(8.5)
If ϕ, ψ are normalized bump functions supported in the unit ball and if λ = (λ1, . . . , λN )
and µ = (µ1, . . . , µN ), it is easy to check that for Euclidean convolution ϕλ ∗ ψµ = θν where
θ ∈ C∞
0 (RN ) is normalized relative to ϕ and ψ, and νj = max{λj, µj} = λj ∨ µj. This fact is
very useful in studying the convolution of two sums of dilates of bump functions. However,
on a more general nilpotent Lie group, this need not be true unless we suitably restrict the
N -tuples λ = (λ1, . . . , λN ) and µ = (µ1, . . . , µN ). Let
EN =nλ = (λ1, . . . , λN ) ∈ (0, ∞)N : λ1 ≤ λ2 ≤ · · · ≤ λNo.
(8.6)
We say that a mapping (x1, . . . , xN ) → (λ−d1
N xN ) is compatible with the group
structure of G if λ ∈ EN . The significance of this notion comes from the following result,
which is Lemma 6.17 in [NRSW12].
1 x1, . . . , λ−dN
54
0 (G) be supported in B(1). If λ, µ ∈ EN there exists θ ∈ C∞
Lemma 8.1. Let ϕ, ψ ∈ C∞
0 (G) so
that ϕλ ∗ ψµ = θν where νj = λj ∨ µj = max{λj, µj}. Moreover, θ is normalized relative to ϕ
and ψ: there are constants ρ0 ≥ 1 and Cm > 0 independent of ϕ and ψ so that θ is supported
in B(ρ0) and sup
γ≤m
∂γθ∞ ≤ Cm sup
γ≤m
∂γϕ∞ sup
γ≤m
∂γψ∞.
It will be useful for our purposes to have a formulation of Lemma 8.1 for dilations with
fewer parameters and with the N variables split into n blocks according to a decomposition
G ∼= RN = RC1 ⊕ · · · ⊕ RCn. For reasons of compatibility with the group structure, we now
require that C1, . . . , Cn ⊂ {1, . . . , N } be sets of consecutive integers, indexed according to the
natural ordering:
C1 = {1, . . . , l1}, C2 = {l1 + 1, . . . , l2},
. . . Cn = {ln−1 + 1, . . . , N }.
(8.7)
This implies that each Gj ∼= RCj ⊕ · · · ⊕ RCn is a subgroup of G. Formula (8.3) remains true
if the individual variables xl, yl ∈ R are replaced by the blocks of variables xj, yj ∈ RCj and
the scalar-valued functions Ml by the RCj -valued functions Mj = (Mlj−1+1, . . . , Mlj ). Notice
that the functions Mj still satisfy (8.4).15
Let E = {e(j, k)} be an n × n matrix satisfying (2.10). Associated to E is the cone of
lattice points
ΓZ(E) =nI = (i1, . . . , in) ∈ Zn : e(j, k)ik ≤ ij < 0, 1 ≤ j, k ≤ no
and we consider dilations
[f ]I (x) = 2− Pn
j=1 ij Qj f (2−i1 · x1, . . . , 2−in · xn).
Note that (2i1, . . . , 2in ) is an n-tuple while elements of EN defined in equation (8.6) are N -
tuples. However [f ]I = fλ(I) where λ(I) is obtained from I by setting λl = 2ik if l ∈ Ck; that
is
(8.8)
λ(I) =(cid:0)
z
C1
}
{
z
2i1, . . . , 2i1,
2i2 , . . . , 2i2 , . . . ,
C2
}
{
Cn
{
z
2in , . . . , 2in (cid:1).
}
We want to impose conditions on the matrix E, in addition to the basic hypothesis (2.10),
which guarantees that if I ∈ ΓZ(E), then λ(I) ∈ EN .
Definition 8.2. An n × n matrix E = {e(j, k)} is doubly monotone if each row is weakly
increasing from left to right, and each column is weakly decreasing from top to bottom; i.e.
e(j, k) ≤ e(j, k + 1)
e(j, k) ≥ e(j + 1, k)
for 1 ≤ j ≤ n and 1 ≤ k < n,
for 1 ≤ j < n and 1 ≤ k ≤ n.
Proposition 8.3. Let E = {e(j, k)} be an n × n matrix satisfying the basic hypothesis (2.10).
(a) The matrix E is doubly monotone if and only if e(j + 1, j) ≤ 1 for 1 ≤ j ≤ n − 1.
(b) If the matrix E is doubly monotone and if I ∈ Γ(E) then λ(I) ∈ EN .
15In what follows below, after choosing marked partitions, we may have situations where we no longer
have sequences of consecutive integers. However this is all right if we can refine to a situation of consecutive
integers.
55
Proof. If E is doubly monotone and each row is weakly increasing, then since e(j +1, j +1) = 1
we certainly have e(j + 1, j) ≤ 1. Conversely, suppose that e(k + 1, k) ≤ 1 for 1 ≤ k ≤ n − 1.
Then by (2.10) we have
e(j, k) ≤ e(j, k + 1)e(k + 1, k) ≤ e(j, k + 1),
e(j + 1, k) ≤ e(j + 1, j)e(j, k) ≤ e(j, k),
so E is doubly monotone.
Next if E is doubly monotone and I = (i1, . . . , in) ∈ Γ(E), it follows from the definition of
λ(I) in equation (8.8) that we only need to show that ij ≤ ij+1 for 1 ≤ j ≤ n − 1. However,
since e(j + 1, j) ≤ 1 and ij < 0, we have ij ≤ e(j + 1, j)ij ≤ ij+1 since I ∈ ΓZ(E).
If ϕ, ψ ∈ C∞
Corollary 8.4. Suppose E is doubly monotone, and let I = (i1, . . . , in), J = (j1, . . . , jn) ∈
0 (Rn) are supported in the unit ball, then [ϕ]I ∗ [ψ]J = [θ]K where
ΓZ(E).
θ ∈ C∞
0 (RN ) is normalized relative to ϕ and ψ, and K = (k1, . . . , kn) ∈ ΓZ(E) with km =
max{im, jm} for 1 ≤ m ≤ n.
the s × s matrix ES = {eS(r, p)} was defined in Lemma 6.2. We observe that if we order the
sets {I1, . . . , Is} so that k1 ≤ k2 ≤ · · · ≤ ks and if E is doubly monotone, then the same is
true of the matrix ES.
Recall that S ∈ S′(E) means that bES ∩ B(1)c 6= ∅. If S =(cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) ∈ S′(E),
Lemma 8.5. Let S = (cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) ∈ S′(E) with k1 < k2 < · · · < ks, and let
ES = {eS(r, p)} be the s × s matrix whose existence is established in Lemma 6.2.
matrix E is doubly monotone, then the matrix ES is doubly monotone.
If the
Proof. From the construction of ES we have
eS(r + 1, r) ≤ τS(r + 1, r) = min
j∈Ir
e(kr+1, j)
e(kr, j)
≤ e(kr+1, kr).
If {e(j, k)} is doubly monotone, e(kr+1, kr) ≤ 1 because of our ordering of the sets {I1, . . . , Is},
and it follows that eS(r + 1, r) ≤ 1. Since ES satisfies the basic hypothesis (2.10), it then
follows from part (a) of Proposition 8.3 that it is also doubly monotone.
Continue to fix S =(cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) ∈ S′(E) with k1 < k2 < · · · < ks. In Definition
3.12 we introduced the family of dilations
where
2−J·S t =(cid:0)2−j1·S tI1 , . . . , 2−jr·S tIr , . . . , 2−js·S tIs(cid:1)
2−jr ·S tIr =n2−jr/e(kr ,j) · tj : j ∈ Iro =n2−jrτS(j,kr ) · tj : j ∈ Iro.
Recall that ΓZ(ES) =(cid:8)(j1, . . . , js) ∈ Zs : eS(p, r)jr ≤ jp < 0(cid:9).
Lemma 8.6. Suppose that E is doubly monotone. If J = (j1, . . . , js) ∈ ΓZ(ES) then the
mapping t → 2−J ·S t is compatible with the group structure.
Proof. Let l, m ∈ {1, . . . , n} with l ≤ m and l ∈ Ir, m ∈ Ip. According to Proposition
8.3, it suffices to show that jre(kr, l)−1 ≤ jpe(kp, m)−1. If J = (j1, . . . , js) ∈ ΓZ(ES) then
eS(p, r)jr ≤ jp. But then since jr < 0,
jp
jr
≤ eS(p, r) ≤ τS(kp, kr) = min
j∈Ir
e(kp, j)
e(kr, j)
≤
e(kp, l)
e(kr, l)
≤
e(kp, m)e(m, l)
e(kr, l)
≤
e(kp, m)
e(kr, l)
where the last inequality follows since l ≤ m and the matrix E is (weakly) increasing across
each row. This means that jre(kr, l)−1 ≤ jpe(kp, m)−1, completing the proof.
56
8.2 Automorphic flag kernels and Lp-boundedness of convolution
operators
In this section we discuss Lp-boundedness of operators
TKf = f ∗ K,
given by convolution (relative to some homogeneous group structure on RN ) with a kernel
K ∈ P(E). The matrix E is assumed to satisfy the basic hypotheses (2.10).
The reader should be aware that there are two possible interpretations of the symbol f ∗K,
depending on whether the underlying group is the standard abelian Rn or some non-abelian
group G. Whenever clarification is required, we will write T G
K to specify the group. For
ordinary convolution on RN , it can be easily verified that a kernel in P(E) is a product kernel,
as defined in Section 2 of [NRS01]. In fact, the differential inequalities of a product kernel
follow from the inequalities Nj(x) ≥ nj(xj ), j = 1, . . . , n, and the cancellation conditions are
the same as in Definition 2.2 (b). By [NRS01], we have the following theorem.
Theorem 8.7. Let δλ be any family of dilations on RN and RN = RC1 ⊕ · · · ⊕ RCn be any
decomposition of RN into homogeneous subspace. Suppose that the matrix E satisfies the basic
hypotheses. Then, for every K ∈ P(E), the operator TK is bounded on Lp for 1 < p < ∞.
Considering now a general homogeneous group law as in (8.3), we must be more specific
on all the ingredients needed to define a class P(E).
We fix a decomposition RN = RC1 ⊕· · ·⊕RCn where the Cj are as in (8.7). In the notation
of [NRS01], the subgroups Gj = RCj ⊕ · · · ⊕ RCn form the standard flag
F :
(0) ⊂ Gn ⊂ Gn−1 ⊂ · · · ⊂ G2 ⊂ G1 = G.
On each RCj we also fix a homogeneous norm nj for the automorphic dilations δλ in
(8.1). A distribution K is an automorphic flag kernel on the standard flag if it satisfies the
differential inequalities
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ Cγ
nYj=1hn1(x1) + · · · + nj(xj )i−Qj −[[γ j ]]
and the cancellation conditions in Definition 2.2 (b).
The differential inequalities for the automorphic flag multiplier m = bK are
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ Cγ
nYj=1(cid:2)nj(ξj) + · · · + nn(ξn)(cid:3)−[[γ j ]]
.
(See Definition 2.3.2 in [NRS01] for details.)
We show that, if E is doubly monotone, a distribution K ∈ P0(E) is always an automorphic
flag kernel for a standard flag.
Proposition 8.8. If K ∈ P0(E) where the class of distributions is defined by a doubly mono-
tone matrix, then K is an automorphic flag kernel on the standard flag.
Proof. Let K ∈ P0(E) where the class of distributions is defined by a doubly monotone matrix
nXk=1
bNj(ξ) =
E = {e(j, k)} and let m = bK. Note that for ξ ∈ B(1)c we have
nk(ξk)1/e(k,j) ≥
nk(ξk)1/e(k,j) ≥
nXk=j
nk(ξk)
nXk=j
57
since for k ≥ j we have e(k, j) ≤ 1. Thus if γ ∈ NC1 × · · · × NCn we have
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ Cγ
nYj=1(cid:0)1 + bNj(ξ)(cid:1)−[[γj ]]
≤ Cγ
nYj=1(cid:2)nj(ξj) + · · · + nn(ξn)(cid:3)−[[γ j ]]
.
Theorem 8.9. Let G be a homogeneous nilpotent Lie group, endowed with automorphic
dilations δλ, G ∼= RN = RC1 ⊕ · · · ⊕ RCn, the Cj being as in (8.7). If E is a doubly monotone
matrix satisfying the basic hypothesis, and if K ∈ P0(E), then the operator TK[ϕ] = ϕ ∗ K,
defined initially for ϕ ∈ S(RN ), extends uniquely to a bounded operator on Lp(G) for 1 < p <
∞. If the rank of E is greater than one, the same holds for K ∈ P(E).
Proof. If K ∈ P0(E), it follows from Proposition 8.8 that K is an automorphic flag kernel on
the standard flag. The result now follows from Theorem 8.14 in [NRSW12]. Assume now that
the rank of E is greater than 1 and K ∈ P(E). By Corollary 7.4, we can write K = K0 + K∞
where K0 ∈ P0(E) and K∞ ∈ L1(RN ). The operator f → f ∗ K∞ is bounded on Lp(G) for
1 ≤ p ≤ ∞.
9 Composition of operators
If K ∈ P0(E) let TK[f ] = f ∗ K be the corresponding left-invariant convolution operator on
the homogeneous nilpotent Lie group G. TK is initially defined on the Schwartz space S(RN ).
We are interested in the composition of two operators TK, TL with K, L ∈ P0(E). Formally
TL ◦ TK[f ] = TK[f ] ∗ L = (f ∗ K) ∗ L = f ∗ (K ∗ L) = TK∗L.
To give a concrete meaning to these equalities is easy if convolution is relative to the
(abelian) additive structure of RN , because we can refer to the Fourier multipliers m = bK,
m′ = bL. For f ∈ S(RN ), TL ◦ TK[f ] = F −1(mm′bf ), where F denotes the Fourier transform,
and the convolution kernel of TL ◦ TK is F −1(mm′) = K ∗ L. In fact, since M∞(E) is closed
under pointwise product, we can conclude as follows.
Theorem 9.1. Let E be a matrix satisfying the basic hypotheses and let K, L ∈ P0(E). With
respect to ordinary convolution on RN , then K ∗ L ∈ P0(E).
Suppose now that RN ∼= G is a noncommutative homogeneous group. It is still easy to
verify that K ∗ L is a well-defined distribution. In fact,
K = K0 + ϕ0,
L = L0 + ψ0,
where K0, L0 have compact support and ϕ0, ψ0 are Schwartz functions. Hence
K ∗ L = K0 ∗ L0 + K0 ∗ ψ0 + ϕ0 ∗ L0 + ϕ0 ∗ ψ0,
where each summand is well defined. However, it is not guaranteed that the convolution K ∗ L
is in the same class.
The main result of this section is that, under the assumptions introduced in Section 8,
K ∗ L ∈ P0(E), and so P0(E) is an algebra under convolution on the group G. We recall the
assumptions:
(i) the underlying dilations on RN are automorphisms of G;
(ii) the subspaces RCj are as in (8.7);
(iii) the matrix E is doubly monotone.
58
Theorem 9.2. Suppose that conditions (i)-(iii) above are satisfied. If K, L ∈ P0(E) then
K ∗ L ∈ P0(E).
The following outline provides an informal road map for the long proof which begins in
Section 9.1. Let K, L ∈ P0(E).
(A) According to Theorem 6.14,
K = ϕ0 + XS∈S ′(E)
KS
and L = ψ0 + XT ∈S ′(E)
LT
where ϕ0, ψ0 ∈ S(RN ) and
Here(cid:8)ϕI
S]I
[ϕI
[ψJ
T ]J .
and LT = XJ∈ΓZ(ET )
KS = XI∈ΓZ(ES )
T(cid:9) are uniformly bounded families of functions in C∞
S(cid:9) and(cid:8)ψJ
ΓZ(ES) =nI = (i1, . . . , is) ∈ Zs : τS(kp, kr)ir ≤ ip < 0, 1 ≤ r, p ≤ so,
ΓZ(ET ) =nJ = (j1, . . . , jq) ∈ Zq : τT (ℓb, ℓa)ja ≤ jb < 0, 1 ≤ a, b ≤ qo.
0 (RN ), and
discussed in Section 9.1.
Then the problem of studying the composition TL ◦ TK reduces essentially to the study
of the finite number of compositions (cid:8)TLT ◦ TKS(cid:9) for S, T ∈ S′(E). This reduction is
(B) We fix marked partitions S = (cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) and T = (cid:0)(J1, ℓ1); . . . ; (Jq, ℓq)(cid:1) in
S′(E), and we simplify notation by writing ϕI
study TLT ◦ TKS where
T = ψJ . Thus we need to
S = ϕI and ψJ
KS = XI∈ΓZ(ES )
LT = XJ∈ΓZ(ET )
F1րΓZ(ES ) XI∈F1
F2րΓZ(ET ) XJ∈F2
lim
[ϕI ]I =
lim
[ϕI ]I =
lim
F րΓZ(ES )
KF1,S,
[ψJ ]J =
[ψJ ]J =
lim
F2րΓZ(ES)
LF2,T .
Here F1, F2 are finite sets of indices, and the limits exist in the sense of distributions.
(C) Recall from Corollary 6.1 that, since S, T ∈ S′(E), ΓZ(ES) and ΓZ(ET ) have nonempty
interiors. By Lemma 6.2 in Section 6.1, there are new matrices ES = {eS(p, r)} and ET =
{eT (b, a)} which satisfy the basic hypothesis such that eS(p, r) ≤ τS(kp, kr), eT (b, a) ≤
τT (ℓb, ℓa).
If F1 ⊂ ΓZ(ES) and F2 ⊂ ΓZ(ET ) are any two finite sets of indices then TLF2,T ◦ TKF1,S =
THF1,F2 ,S,T where
HF1,F2,S,T = KF1,S ∗ LF2,T = XI∈F1 XJ∈F2
[ϕI ]I ∗ [ψJ ]J .
In Section 9.2 we see that it suffices to show that limF1րΓZ(ES) limF2րΓZ(ET ) HF1,F2,S,T
exists in the sense of distributions and belongs to P0(E). More generally, for arbitrary
finite subsets F ⊂ ΓZ(ES) × ΓZ(ET ), we show that
(a) HF,S,T = X(I,J)∈F
[ϕI ]I ∗ [ψJ ]J ∈ P0(E) with estimates independent of the set F ;
59
(b) In the sense of distributions,
lim
F րΓZ(ES )×ΓZ(ET )
HF,S,T exists and belongs to P0(E).
(D) In Section 9.3 we summarize the support, decay, and cancellation properties of the con-
volutions [ϕI ]I ∗ [ψJ ]J that are needed in our analysis.
(E) In order to study HF,S,T we write
ΓZ(ES) × ΓZ(ET ) = [W ∈W(S,T )
ΓW (S, T )
disjoint. This partition of ΓZ(ES) × ΓZ(ET ) is somewhat complicated, but it has the
where the index set W(S, T ) is finite, and the sets(cid:8)ΓW (S, T ) ⊂ ΓZ(ES) × ΓZ(ET )(cid:9) are
property that if (I, J) =(cid:0)(i1, . . . , is), (j1, . . . , jq)(cid:1) ∈ ΓW (S, T ), then for 1 ≤ r ≤ s and
1 ≤ a ≤ q, if Ir ∩ Ja 6= ∅ the ratio ir/ja is suitably restricted. Having made the partition,
it then suffices to show that for every W ∈ W(S, T ), limF րΓW (S,T ) HF,S,T converges in
the sense of distributions and belongs to P0(E). This partition is described in Section
9.4.
(F) Now fix W ∈ W(S, T ), and let F ⊂ ΓW (S, T ) be a finite set. We study of HF,S,T =
P(I,J)∈F [ϕI ]I ∗ [ψJ ]J .
(a) For each convolution [ϕI ]I ∗ [ψJ ]J there is a normalized bump function θI,J and a
set of dilations M (I, J) so that [ϕI ]I ∗ [ψJ ]J = [θI,J ]M(I,J).
(b) There are s parameters (i1, . . . , is) in I and q parameters (j1, . . . , jq) in J, and the
function [θI,J ]M(I,J) is dilated by some collection of these dilations. However not all
of these need appear in M (I, J). We will say that a parameter ir or ja from I ∪ J is
'fixed' if it appears among the parameters in M (I, J), and 'free' if it does not appear.
These notions are defined precisely in Section 9.5.
(c) Let I ∗
r be the indices in Ir which are fixed, and let J ∗
a be the indices in Ja which are
fixed. Some of these sets may be empty (if Ir or Ja consists entirely of free indices),
and we suppose there are α ≤ s non-empty sets I ∗
rα and β ≤ q non-empty
set J ∗
r1 , . . . , I ∗
a1, . . . J ∗
aβ .
(d) The sets I ∗
rα , J ∗
to a new decomposition
r1 , . . . , I ∗
a1, . . . J ∗
aβ are disjoint and their union is {1, . . . , n}. This leads
RN =
I ∗
rj ⊕
R
αMj=1
RJ ∗
ak .
βMk=1
This decomposition is discussed in Section 9.6.
(G) Associated to this last decomposition, we show in Section 9.7 that there is an (α + β) ×
(α + β) matrix E(S, T, W ) with associated family of norms N #
α+β on RN . In
Section 9.8 we show that the corresponding family of distributions P0(E(S, T, W )) is
contained in the original family P(E). Theorem 9.2 then follows if we can show that
1 , . . . , N #
X(I,J)∈ΓW (S,T )
[ϕI ]I ∗ [ψJ ]J ∈ P0(E(S, T, W )).
This is the heart of the matter, and we argue as follows.
60
a) Each convolution [ϕI ]I ∗ [ψJ ]J = [θI,J ]M(I,J) where θI,J ∈ C∞
0 (RN ) is normalized
relative to ϕI and ψJ , and where M (I, J) is an α + β-tuple of integers coming from
α + β of the integers in I = (i1, . . . , is) and J = (j1, . . . , jq). Since each function ϕI
and ψJ has strong cancellation, the function θI,J will have appropriate decay in the
free variables and weak cancellation in the fixed variables. (See Lemma 9.4 below.)
b) We let M(S, T, W ) denote the (infinite) set of α + β-tuples M = M (I, J) that arise in
this way as the dilation for some convolution [ϕI ]I ∗ [ψJ ]J . For each M ∈ M(S, T, W ),
we let F (M ) be the set of pairs (I, J) ∈ ΓW (S, T ) such that M (I, J) = M . It follows
that
X(I,J)∈ΓW (S,T )
[ϕI ]I ∗ [ψJ ]J = XM∈M(S,T,W )h X(I,J)∈F (M)
θI,JiM
.
c) Because of the decay properties of θI,J , for each M ∈ M(S, T, W ), the inner infinite
0 , and the family of functions
{ΘM } is uniformly bounded and has weak cancellation relative to M . Thus
sumP(I,J)∈F (M) θI,J converges to a function ΘM ∈ C∞
[ϕI ]I ∗ [ψJ ]J = XM∈M(S,T,W )
X(I,J)∈ΓW (S,T )
[ΘM ]M .
Theorem 5.13 then shows that P(I,J)∈ΓW (S,T )[ϕI ]I ∗ [ψJ ]J ∈ P0(E(S, T, W )). The
details are provided in Sections 9.9 and 9.10.
9.1 A preliminary decomposition
Let K, L ∈ P0(E). According to Corollary 6.16 we can write
K = ϕ0 + XS∈S ′(E)
L = ψ0 + XT ∈S ′(E)
KS = ϕ0 + XS∈S ′(E) XI∈ΓZ(ES )
LT = ψ0 + XT ∈S ′(E) XJ∈ΓZ(ET )
[ϕI
S]I ,
[ψJ
T ]J ,
where ϕ0, ψ0 ∈ S(RN ) and {ϕI
C∞
0 (RN ) supported in B(1). Each ϕI
ψJ
T has cancellation in the variables {xl1, . . . , xlq }. We have
T } are uniformly bounded families of functions in
S has cancellation in the variables {xk1 , . . . , xks} and each
S} and {ψJ
TK ◦ TL = ϕ0 ∗ L + K ∗ ψ0 + XS,T ∈S ′(E)
TKS ◦ TLT .
Since the first two terms on the right-hand side are Schwartz functions, to prove Theorem 9.2
it suffices to show that each composition TKS ◦ TLT is a convolution operator THS,T where
HS,T ∈ P0(E).
9.2 Reduction to the case of finite sets
It follows from Corollary 6.16 that
KS(x) =
lim
F րΓZ(ES )XI∈F
[ϕI
S]I (x),
LT (x) =
lim
GրΓZ(ET )XJ∈G
[ψJ
T ]J (x)
where the limits are taken in the sense of distributions, F ranges over finite subsets of ΓZ(ES),
and G ranges over finite subsets of ΓZ(ET ). It follows that
TKS ◦ TLT [f ] =
lim
F րΓZ(ES )
lim
GրΓZ(ET )hf ∗XI∈F XJ∈G
[ϕI
S]I ∗ [ψJ
T ]Ji.
Thus it will suffice to show that limF րΓZ(ES ) limGրΓZ(ET )PI∈FPJ∈G[ϕI
T ]J ∈ P0(E)
where the limits are take in the sense of distributions. To do this, we will show that for any
finite sets F1 ⊂ ΓZ(ES) and F2 ⊂ ΓZ(ET ),
S]I ∗ [ψJ
61
HF1,F2 = XI∈F1 XJ∈F2
[ϕI
S]I ∗ [ψJ
T ]J ∈ P0(E)
with estimates that are independent of the choice of the finite subsets F and G, and
HS,T =
lim
F1րΓZ(ES )
lim
F2րΓZ(ET )
HF,G ∈ P0(E)
Thus to prove Theroem 9.2, it will suffice establish the following result.
Lemma 9.3. For each S, T ∈ S′(E) there exists HS,T ∈ P0(E) so that
KS ∗ LT =
X
(I,J)∈ΓZ(ES )×ΓZ(ET )
[ϕI
S]I ∗ [ψJ
T ]J = HS,T
(9.1)
where the series converges in the sense of distributions as the limit of sums over finite subsets
of ΓZ(ES) × ΓZ(ET ).
9.3 Properties of the convolution [ϕI
S]I ∗ [ψJ
T ]J
In order to proceed further, we will need information about the support, decay, and cancel-
lation properties of each of the convolutions [ϕI
T ]J that appear in equation (9.1). The
required estimates are given in Lemma 9.4 below. The statement essentially incorporates the
results of Lemmas 6.17 and 6.18 and Corollary 6.19 in [NRSW12]. Since the setting now is
somewhat more complicated, we include a sketch of the proof.
S]I ∗ [ψJ
Suppose that G is a homogeneous nilpotent Lie group with underlying manifold RN , and
that the one-parameter family of automorphic dilations on G is given by
δr(x) = (rd1 x1, . . . , rdN xN ).
Let ∗ denote the convolution on G.
Lemma 9.4. Let ϕ, ψ ∈ C∞
0 (RN ) have support in the unit ball and put
[ϕ]λ(x) =h NYm=1
[ψ]µ(x) =h NYm=1
λ−dm
m iϕ(λ−d1
m iψ(µ−d1
µ−dm
1 x1, . . . , λ−dN
N xN ),
1 x1, . . . , µ−dN
N xN ).
Assume that 0 < λ1 ≤ λ2 ≤ · · · ≤ λN and 0 < µ1 ≤ µ2 ≤ · · · ≤ µN .
(a) [Support properties] There exists θ ∈ C∞
0 (RN ), normalized with respect to ϕ and ψ, so
that,
[ϕ]λ ∗ [ψ]µ(x) =h nYm=1
ν−dm
m iθ(ν−d1
where νj = λj ∨ µj = max{λj, µj}.
1 x1, . . . , ν−dN
N xN ) = [θ]ν (x)
(b) [Cancellation and decay properties] The estimates come in three parts, depending on
whether there is cancellation in any of the first N − 1 variables {x1, . . . , xN −1}, or in the
last variable xN .
62
(i) Let C, D ⊂ {1, . . . , N − 1}, and suppose that ϕ has cancellation in the variables xj
for j ∈ C, and that ψ has cancellation in the variables xk for k ∈ D. Put
C1 =nj ∈ {1, . . . , N − 1} : ϕ has cancellation in xj and λj < µj(cid:9),
C2 =nj ∈ {1, . . . , N − 1} : ϕ has cancellation in xj and λj ≥ µj(cid:9),
D1 =nj ∈ {1, . . . , N − 1} : ψ has cancellation in xj and µj ≤ λj(cid:9),
D2 =nj ∈ {1, . . . , N − 1} : ψ has cancellation in xj and µj > λj(cid:9),
C1 ∪ C2 = C,
D1 ∪ D2 = D,
C1 ∩ D1 = ∅,
C1 ∩ C2 = ∅,
D1 ∩ D2 = ∅,
C2 ∩ D2 = ∅.
so that
There exists ǫ > 0 depending only on the group structure so that [ϕ]λ ∗ [ψ]µ can be
written as a finite sum of terms of the form
µj(cid:19)ǫ
Yj∈C1(cid:20)(cid:18) λj
λj+1(cid:19)ǫ(cid:21) × Yk∈D1(cid:20)(cid:18) µk
λk(cid:19)ǫ
+(cid:18) λj
2(cid:18) λj
λj+1(cid:19)ǫ
× Yj∈C ′
µk+1(cid:19)ǫ(cid:21) ×
+(cid:18) µk
2(cid:18) µk
µk+1(cid:19)ǫ(cid:2)ΘC ′′
× Yk∈D′
2 ,D′′
2(cid:3)ν
(9.2)
C2 = C′
D2 = D′
2 ∪ C′′
2 ,
2 ∪ D′′
2 ,
C′
D′
2 ∩ C′′
2 ∩ D′′
2 = ∅,
2 = ∅.
cancellation in the variables xj for j ∈ C′′
2 ∪ D′′
2 .
0 (RN ) are normalized relative to ϕ and ψ and have
where
2 ,D′′
The functions (cid:8)ΘC ′′
2(cid:9) ⊂ C∞
µN(cid:17)ǫ
additional factor (cid:16) λN
(ii) Suppose that ϕ also has cancellation in the variable xN . If λN < µN , there is an
in equation (9.2). If λN ≥ µN then the function ΘC ′′
2 ,D′′
2
in equation (9.2) also has cancellation in the variable xN .
(iii) Suppose that ψ also has cancellation in the variable xN .
If λN < µN , then the
in equation (9.2) also has cancellation in the variable xN . If λN ≥
function ΘC ′′
2 ,D′′
2
µN there is an additional factor (cid:16) λN
µN(cid:17)ǫ
in equation (9.2).
Sketch of the proof.
Because of the monotonicity of the λ's and the µ's, assertion (a) follows immediately from
Lemma 8.1, so the real content of the lemma is assertion (b). Let us first consider the case of
Abelian convolution. Then there is no need to distinguish between the first N −1 variable and
the last variable. If (say) ϕ has cancellation in the variable xj then we can write ϕ = ∂xj φ
where φ is normalized in terms of ϕ. Suppose that λj < µj . Then moving the derivative ∂xj
across the convolution, we have
[ϕ]λ ∗ [ψ]µ = [∂xj φ]λ ∗ [ψ]µ = λdj
j ∂xj(cid:0)[φ]λ) ∗ [ψ]µ
= λdj
µj(cid:17)dj
j [φ]λ ∗ ∂xj(cid:0)[ψ]µ(cid:1) =(cid:16) λj
µj(cid:17)dj
[φ]λ ∗ [∂xj ψ]µ =(cid:16) λj
[θ]ν
where we have use part (a) to write [φ]λ ∗ [∂xj ψ]µ = [θ]ν where θ is normalized. On the other
hand, if λj ≥ µj, we can put the derivative in front of the whole convolution, and we have
63
[ϕ]λ ∗ [φ]µ = [∂xj φ]λ ∗ [φ]µ = λdj
j ∂xj(cid:0)[φ]λ) ∗ [φ]µ
= λdj
j ∂xj(cid:16)[φ]λ ∗ [ψ]µ(cid:17) = λdj
j ∂xj(cid:0)[θ′]ν(cid:1) = [∂xj θ′]ν
where [φ]λ ∗ [ψ]µ = [θ′]ν by part (a), and now ∂xj θ′ has cancellation in the variable xj . Thus
the lemma follows in the case of Abelian convolution on Rn.
For convolution ∗ on a homogeneous nilpotent Lie group, in general it is not the case that
∂xj f ∗ g = f ∗ ∂xj g or that ∂xj f ∗ g = ∂xj [f ∗ g]. However, we have the following facts.16
(a) If X is a left-invariant vector field and Y is a right-invariant vector field then X[f ∗ g] =
f ∗ X[g] and Y [f ∗ g] = Y [f ] ∗ g;
(c) For any 1 ≤ j ≤ N − 1 we can write
agreeing with X at the origin, then X[f ] ∗ g = f ∗ Y [g];
(b) If X is a left-invariant vector field and Y = eX is the unique right-invariant vector field
∂xk(cid:2)Pj,k(x)f (x)(cid:3)
NXk=j+1
NXk=j+1
(x) = Zj[f ](x) +
(x) = Zj[f ](x) +
∂f
∂xk
∂f
∂xj
Pj,k(x)
∂
where Zj is either a left- or a right-invariant vector field, and Pj,k is a polynomial of the
form
Pj,k(x1, . . . , xk−1) = Xα∈A(j,k)
cj,k(α)xα1
1 · · · xαk−1
k−1
where A(j, k) =(cid:8)(α1, . . . , αk−1) ∈ Zk−1 : α1d1 + · · · + αk−1dk−1 = dk − dj(cid:9) .
We can establish part (b) of the lemma by repeating the Abelian argument, but replacing the
derivative ∂xj by the appropriate invariant vector field Zj. When j < N this introduces an
error which is a sum of terms with cancellation in variables xk with k > j. Explicitly,
[∂xj φ]λ = λdj
j ∂xj(cid:0)[φ]λ(cid:1) = λdj
NXk=j+1
∂xk(cid:0)Pj,k(x)[φ]λ(x)(cid:1)
Now it follows as in Proposition 4.5 of [NRSW12] that
j
j Zj(cid:0)[φ]λ(cid:1) + λdj
NXk=j+1(cid:16) λj
λk(cid:17)dj
λk(cid:17)dj
NXk=j+1(cid:16) λj
j Zj(cid:0)[φ]λ(cid:1) +
j Zj(cid:0)[φ]λ(cid:1) +
[∂xj φ]λ = λdj
= λdj
λdk
k ∂xj(cid:0)[Θj,k]λ(cid:1)
[∂xj Θj,k]λ
where the functions Θj,k are normalized relative to φ. The term λdj
the Abelian case, since now the vector field Zj can either be moved across a convolution or
j Zj(cid:0)[φ]λ(cid:1) is handled as in
µj(cid:17)dj
put in front of the convolution. This gives a gain of(cid:16) λj
cancellation in variables xk with k > 1, and there is a gain (cid:16) λj
≤(cid:16) λj
λj+1(cid:17)dj
λk(cid:17)dj
. Of course
when j = N , there is no correction term. This completes the sketch. Complete details can
be found in Section 6 of [NRSW12].
. The remaining terms now have
16See Proposition 6.16 in [NRSW12] or page 22 in [FS82].
64
9.4 A further decomposition of ΓZ(ES) × ΓZ(ET )
Fix two marked partitions S = (cid:8)(I1, k1); . . . ; (Is, ks)(cid:9), T = {(J1, ℓ1); . . . ; (Jq, ℓq)} ∈ S′(E).
We return to the study of the infinite sum P(I,J)∈ΓZ(ES)×ΓZ(ET )[ϕI ]I ∗ [ψJ ]J .17 To analyze
this, we need a further decomposition of the index set ΓZ(ES)×ΓZ(ET ). Recall from equation
(1.15) that
[ϕI ]I (x) = 2− Ps
[ψJ ]J (x) = 2− Pq
r=1 ir bQS,r ϕI (2−i1 ·S xI1 , . . . , 2−is ·S xIs ),
a=1 ja bQT ,a ψJ (2−j1 ·T xJ1 , . . . , 2−jq·T xJq ),
where, using the identities e(kr, j) = τS(j, kr)−1 and e(ℓa, l) = τT (l, ℓa)−1 from Proposition
3.8, together with (3.6) and Definition 3.12 (b), we have
QjτS(j, kr),
Qj e(kr, j)−1
bQS,r = Xj∈Ir
bQT,a = Xl∈Ja
= Xj∈Ir
=n2−ir τS(j,kr ) · xv : j ∈ Iro,
2−ir ·S xIr =n2−ir /e(kr ,j) · xv : j ∈ Iro
= Xl∈Ja
2−ja ·T xJa =n2−ja/e(ℓa,v) · xv : v ∈ Jao =n2−jaτT (v,ℓa) · xv : v ∈ Jao.
Qae(ℓa, l)−1
QaτT (l, ℓa),
(9.3)
(9.4)
According to Lemma 9.4, for each index j, 1 ≤ j ≤ n, the scale of the support of the
convolution [ϕI ]I ∗ [ψJ ]J in the direction of each coordinate xj is the maximum of the scales
of the supports of each factor. We let 1 ≤ r ≤ s and 1 ≤ a ≤ q and focus on indices j ∈ Ir ∩Ja
provided that the intersection is non-empty. The size of the support of [ϕI ]I in the direction
of xj is on the order of 2irτS(j,kr ), while the size of the support of [ψJ ]J is on the order of
2jτT (vj ,ℓa). Since the indices ir and ja are negative, we have
2ir τS(j,kr ) ≥ 2jaτT (j,ℓa) ⇐⇒ irτS(j, kr) ≥ jaτT (j, ℓa) ⇐⇒
Write18 Ir ∩ Ja =(cid:8)v1, . . . , vm(cid:9), and order these indices so that
0 <
τT (v1, ℓa)
τS(v1, kr)
≤
τT (v2, ℓa)
τS(v2, kr)
≤ · · · ≤
τT (vm−1, ℓa)
τS(vm−1, kr)
≤
ir
ja
≤
τT (j, ℓa)
τS(j, kr)
=
e(kr, j)
e(ℓa, l)
.
τT (vm, ℓa)
τS(vm, kr)
< ∞.
(9.5)
The ratios in equation (9.5) divide the positive real axis into at most m + 1 subintervals. For
any I = (i1, . . . , is) ∈ ΓZ(ES) and J = (j1, . . . , jq) ∈ ΓZ(ET ), the ratio ir/ja must lie in one
of these intervals. We partition ΓZ(ES) × ΓZ(ET ) into disjoint sets so that for (I, J) in each
set, the ratios ir/ja always lie in the same interval.
To describe and index this decomposition, we proceed as follows. Let
ΣS,T = {(Ir, Ja) : 1 ≤ r ≤ s, 1 ≤ a ≤ q, Ir ∩ Ja 6= ∅},
For each (Ir, Ja) ∈ ΣS,T consider the set of mr,a + 1 intervals
τT (v2, ℓa)
τS(v2, kr)(cid:19) = Er,a
1 , . . .
((cid:20)0,
,
τS(v1, kr)
0 ,(cid:20) τT (v1, ℓa)
τS(vj+1, kr)(cid:19) = Er,a
τT (vj+1, ℓa)
j
τT (v1, ℓa)
τS(v1, kr)(cid:19) = Er,a
. . . ,(cid:20) τT (vj , ℓa)
. . . ,(cid:20) τT (vmr,a−1, ℓa)
τS(vmr,a−1, kr)
τS(vj , kr)
,
, . . .
(9.6)
,
τT (vmr,a, ℓa)
τS(vmr,a, kr)(cid:19) = Er,a
mr,a−1,(cid:20) τT (vmr,a , ℓa)
τS(vmr,a, kr)
mr,a),
, ∞(cid:19) = Er,a
17Since S and T are fixed, we now write ϕI and ψJ instead of ϕI
18We should write Ir ∩ Ja = {vr,a
1 , . . . , vr,a
S and ψJ
T .
m(r,a)}, but as long as the context is clear we omit the dependence
on (r, a) for simplicity of notation.
and let
fW(S, T ) = Y(Ir ,Ja)∈ΣS,T(cid:8)Er,a
0 , Er,a
1 , . . . , Er,a
mr,a(cid:9),
65
i.e., the set of functions which assign to each (Ir, Ja) ∈ ΣS,T one of the intervals in (9.6).
w(Ir ,Ja)
τS(vj ,kr )
τT (vj ,ℓa)
if j > 0,
∞
if j = 0,
τT (vj+1,ℓa)
τS (vj+1,kr)
W (Ir, Ja) = Er,a
w(Ir, Ja) is the number of that interval.) Setting j = w(Ir, Ja), let
Hence, if W ∈ fW(S, T ) and (Ir, Ja) ∈ ΣS,T , W (Ir, Ja) is one of these intervals, and we write
so that w(Ir, Ja) ∈(cid:8)0, 1, . . . , mr,a(cid:9). (Thus W (Ir, Ja) is one of the mr,a + 1 intervals, while
ΛW (Ir, Ja) =
ΛW (Ja, Ir) =
We now construct the desired partition. For W ∈ fW(S, T ), put
ΓW (S, T ) =n(cid:0)(i1, . . . , is),(j1, . . . , jq)(cid:1) ∈ ΓZ(ES) × ΓZ(ET ) :
< ΛW (Ir, Ja)o.
It may happen that for some W ∈ fW(S, T ) the set ΓW (S, T ) is empty,19 and so we put
so that ΛW (Ja, Ir)−1 and ΛW (Ir, Ja) are the left and right endpoints of Ew(Ir ,Ja).
(Ir, Ja) ∈ ΣS,T =⇒
∞
if j = mr,a,
if j < mr,a,
1
ΛW (Ja, Ir)
≤
ir
ja
(9.7)
W(S, T ) =nW ∈ fW(S, T ) : ΓW (S, T ) 6= ∅o.
The following Proposition is an immediate consequence of the definition of ΓW (S, T ) and the
ordering in equation (9.5).
Proposition 9.5. Let S =(cid:8)(I1, k1); . . . ; (Is, ks)(cid:9), T = {(J1, ℓ1); . . . ; (Jq, ℓq)} ∈ S′(E).
(a) The sets {ΓW (S, T ) : W ∈ W(S, T )} partition the set ΓZ(ES) × ΓZ(ET ):
ΓZ(ES) × ΓZ(ET ) = [W ∈W(S,T )
ΓW (S, T ),
W1, W2 ∈ W(S, T ), W1 6= W2 =⇒ ΓW1(S, T ) ∩ ΓW2 (S, T ) = ∅.
(b) Let W ∈ W(S, T ) and suppose Ir ∩ Ja = {v1, . . . , vm} 6= ∅, ordered as in equation (9.5).
Suppose that W (Ir, Ja) = Ew(Ir ,Ja). Let (cid:0)(i1, . . . , is), (j1, . . . , jq)(cid:1) ∈ ΓW (S, T ). Then for
1 ≤ k ≤ m,
k ≤ w(Ir, Ja) ⇐⇒ jaτT (vk, ℓa) ≥ irτS(vk, kr)
k > w(Ir, Ja) ⇐⇒ jaτT (vk, ℓa) < irτS(vk, kr)
In particular,
(1) if w(Ir, Ja) = 0, then irτS(vk, kr) > jaτT (vk, ℓa) for all 1 ≤ k ≤ m;
(2) if w(Ir, Ja) = m then irτS(vk, kr) ≤ jaτT (vk, ℓa) for all 1 ≤ k ≤ m;
19For instance, if there are equalities among the weak inequalities in (9.5), some of the intervals are empty.
66
(3) if kr ∈ Ir ∩ Ja then
(4) if ℓa ∈ Ir ∩ Ja then
kr ≤ w(Ir, Ja) =⇒ ir ≤ τT (kr, ℓa) ja < 0,
kr > w(Ir, Ja) =⇒ τT (kr, ℓa) ja < ir < 0;
ℓa > w(Ir, Ja) =⇒ ja < τS(ℓa, kr) ir < 0,
ℓa ≤ w(Ir, Ja) =⇒ τS(ℓa, kr) ir ≤ ja < 0.
To summarize, we have partitioned ΓZ(ES)×ΓZ(ET ) into a disjoint union of sets ΓW (S, T ).
Let I = (i1, . . . , is) and J = (j1, . . . , jq). Then (I, J) ∈ ΓW (S, T ) if and only if
eS(p, r)ir ≤ ip < 0
eT (b, a)ja ≤ jj < 0
ΛW (Ir, Ja)ja < ir
ΛW (Ja, Ir)ir ≤ ja
for r, p ∈ {1, . . . , s},
for a, b ∈ {1, . . . , q},
if Ir ∩ Ja 6= ∅,
if Ir ∩ Ja 6= ∅.
(9.8)
It can happen that ΛW (Ir, Ja) or ΛW (Ja, Ir) is +∞, in which case there is simply no corre-
sponding inequality between ir and ja. However, if there is no bound of an index ir by any
ja, then ΛW (Ir, Ja) = ∞ for every a such that Ir ∩ Ja 6= ∅, and this means that the scale
of the dilation by ir is always smaller than the scale of the dilation by the corresponding ja.
Similarly, if there is no bound of ja by any ir, then the scale of the dilation by ja is always
less than or equal to the scale of of the dilation by the corresponding ir.20
9.5 Fixed and free indices
Continue to fix S = (cid:8)(I1, k1); . . . ; (Is, ks)(cid:9), T = {(J1, ℓ1); . . . , (Jq, ℓq)} ∈ S′(E). The index
set W(S, T ) is finite, so formally, using the partition from Proposition 9.5, we have
KS ∗ LT (x) = XW ∈W(S,T )" X(I,J)∈ΓW (S,T )
[ϕI ]I ∗ [ψJ ]J (x)# = XW ∈W(S,T )
RS,T,W (x).
Thus to prove Lemma 9.3 (and hence prove Theorem 9.2) it suffices to show that for each
W ∈ W(S, T ) the inner infinite sum
RS,T,W = X(I,J)∈ΓW (S,T )
[ϕI ]I ∗ [ψJ ]J
converges in the sense of distributions, and that the sum belongs to the class P0(E). We will do
this by showing that there is a matrix E(S, T, W ) so that RS,T,W ∈ P0(E(S, T, W )) ⊂ P0(E).
Definition 9.6. Let S = (cid:8)(I1, k1); . . . ; (Is, ks)(cid:9), T = {(J1, ℓ1); . . . , (Jq, ℓq)} ∈ S′(E) and
W ∈ W(S, T ). Recall that if (Ir, Ja) ∈ ΣS,T then Ir ∩ Ja 6= ∅ and W (Ir, Ja) = Ew(Ir ,Ja) where
{E0, . . . , Em} are the intervals listed in equation (9.6).
20In the notation of the next section, this means that in the first case, ir is a free index, and in the second
case ja is a free index. See Definition 9.6.
67
(9.9)
(a) Set
I ∗
r =
=
J ∗
a =
=
Ir ∩Ja6=∅nj ∈ Ir ∩ Ja : ir τS(j, kr) > ja τT (j, ℓa)
q[a=1
Ir ∩Ja6=∅nvk ∈ Ir ∩ Ja : k > w(Ir, Ja)o;
q[a=1
Ir ∩Ja6=∅nj ∈ Ir ∩ Ja : ir τS(j, kr) ≤ ja τT (j, ℓa)
w[r=1
∀(cid:0)(i1, . . . , is), (j1, . . . , jq)(cid:1) ∈ ΓW (S, T )o
∀(cid:0)(i1, . . . , is), (j1, . . . , jq)(cid:1) ∈ ΓW (S, T )o
Ir ∩Ja6=∅nvk ∈ Ir ∩ Ja : k ≤ w(Ir, Ja)o.
q[a=1
(b) An index r, 1 ≤ r ≤ s is fixed if I ∗
r 6= ∅, and is free if I ∗
r = ∅.
(c) An index a, 1 ≤ a ≤ q is fixed if J ∗
a 6= ∅, and is free if J ∗
a = ∅.
r and J ∗
Remark 9.7. The notions of fixed and free indices and the definitions of the sets I ∗
a
depend on the choice of S, T ∈ S′(E) and on the choice of W ∈ W(S, T ). In particular, they
are the same for all choices of (I, J) ∈ ΓW (S, T ).
Clearly I ∗
r ⊂ Ir, J ∗
a ⊂ Ja and I ∗
r ∩ J ∗
pair (Ir, Ja) ∈ ΣS,T so that j ∈ Ir ∩ Ja, and then either j ∈ I ∗
disjoint decomposition
a = ∅. Also, for every j ∈ {1, . . . , n} there is a unique
a . Thus we have a
r or j ∈ J ∗
{1, . . . , n} =
I ∗
r ∪
J ∗
a .
(9.10)
s[r=1
q[a=1
The following Proposition provides an explanation of the significance of fixed and free
indices. The proof just involves unwinding the definitions.
Proposition 9.8.
(a) An index r ∈ {1, . . . , s} is free if and only if for every j ∈ Ir and every (I, J) ∈ ΓW (S, T ),
the scale of the support of [ϕI ]I is smaller in the direction of xj than the scale of the
support of [ψJ ]J .
(b) An index r ∈ {1, . . . , s} is fixed if and only if there exists at least one j ∈ Ir so that for
every (I, J) ∈ ΓW (S, T ), the scale of the support of [ϕI ]I in the direction of xj is larger
than the scale of the support of [ψJ ]J .
(c) The indices j ∈ {1, . . . , n} for which the scale of the support of [ϕI ]I is larger than the
scale of the support of [ψJ ]J in the direction of xj is preciselySs
r=1 I ∗
r .
(d) An index a ∈ {1, . . . , q} is free if and only if for every j ∈ Ja and for every (I, J) ∈
ΓW (S, T ), the scale of the support of [ψJ ]J in the direction of xj is smaller than the scale
of the support of [ϕI ]I .
68
(e) An index a ∈ {1, . . . , q} is fixed if and only if there exists at least one j ∈ Ja so that for
every (I, J) ∈ ΓW (S, T ) the scale of the support of [ψJ ]J is larger than the scale of the
support of [ϕI ]I in the direction of xj .
(f) The indices j ∈ {1, . . . , n} for which the scale of the support of [ψJ ]J is larger than the
scale of the support of [ϕI ]I in the direction of xj is preciselySq
Let S∗ = {r1, . . . , rα} denote the subset of {1, . . . , s} consisting of fixed indices, and let
T ∗ = {a1, . . . , aβ} denote the subset of {1, . . . , q} consisting of fixed indices. It may happen
that α = 0 or that β = 0, but we always have α + β ≥ 1. Then we can write
a=1 J ∗
a .
{1, . . . , n} = [r∈S ∗
I ∗
r ∪ [a∈T ∗
J ∗
a =
I ∗
rj ∪
α[j=1
J ∗
ak .
β[k=1
(9.11)
9.6 A finer decomposition of RN
Using the decomposition in equation (9.11) we study the new decomposition of RN indexed
by the fixed indices in {1, . . . , s} and {1, . . . , q}:
where
R
R
J ∗
aj
a =
RI ∗
RJ ∗
I ∗
rj ⊕
αMj=1
r ⊕ Ma∈T ∗
βMj=1
RN = Mr∈S ∗
r =(cid:8)x = (x1, . . . , xn) ∈ RN : xj 6= 0 =⇒ j ∈ I ∗
r(cid:9),
a(cid:9).
a =(cid:8)x = (x1, . . . , xn) ∈ RN : xj 6= 0 =⇒ j ∈ J ∗
r = {xj : j ∈ I ∗
r } and xJ ∗
a as xI ∗
RI ∗
RJ ∗
Write elements of RI ∗
a set of coordinates on RN .
r and RJ ∗
a = {xj : j ∈ J ∗
a }. This gives us
Order these coordinates so that the corresponding set of indices {kr1, . . . , krα, ℓa1, . . . , ℓaβ }
together are increasing. With this ordering, rename the subspaces as {RV1 , . . . , RVα+β }. Then
every element x ∈ RN is an (α + β)-tuple x = (x1, . . . , xm, . . . , xα+β) where xm ∈ RVk and Vk
is either I ∗
a for some a ∈ T ∗. By the choice of the ordering, if m1 < m2
then
r for some r ∈ S∗ or J ∗
Vm1 = I ∗
r ,
Vm1 = I ∗
r ,
Vm1 = J ∗
a ,
Vm1 = J ∗
a ,
Vm2 = I ∗
p
Vm2 = J ∗
a
Vm2 = I ∗
r
Vm2 = J ∗
b
=⇒ kr < kp,
=⇒ kr < ℓa,
=⇒
ℓa < kr,
=⇒
ℓa < ℓb.
On each component RVm we use the dilation structure inherited from either Ir or Ja. Thus
from (9.4) we define a dilation which we write ◦ by setting
λ ◦ xm =
nλτS (j,kr ) · xj : j ∈ I ∗
nλτT (j,ℓa) · xj : j ∈ J ∗
ro if xm = xI ∗
ao if xm = xJ ∗
a
r
.
(9.12)
(To keep notation simple we write ◦ instead of ◦S,T,W .) The corresponding homogeneous
norms are given by
nS,r(xI ∗
r ) = Xj∈I ∗
r
nj(xj )1/τS (j,kr),
nT,a(xJ ∗
a ) = Xj∈J ∗
a
nj(xj )1/τT (j,ℓa),
If(cid:0)(i1, . . . , is), (j1, . . . , jq)(cid:1) ∈ ΓW (S, T ) ⊂ ΓZ(ES) × ΓZ(ET ), we obtain an (α + β)-tuple
{ir1, . . . , irα, ja1, . . . , jaβ } by keeping only the ir and js corresponding to fixed indices S∗ and
T ∗. Let π : Zs+q → Zα+β be the mapping which takes (i1, . . . , is), (j1, . . . , jq) ∈ Zs+q to
(ir1, . . . , irα, ja1 , . . . , jaβ ) ∈ Zα+β. Our next objective is to describe the inequalities satisfied
by the entries of these (α + β)-tuples.
It follows from equation (9.8) that
eS(rj , rk)irk ≤ irj < 0
eT (al, am)jam ≤ jal < 0
ΛW (Irj , Jam)jam < irj
ΛW (Jam , Irj )irj ≤ jam
for j, k ∈ {1, . . . , α},
for l, m ∈ {1, . . . , β},
if Irj ∩ Jam 6= ∅,
if Irj ∩ Jam 6= ∅.
69
(9.13)
(9.14)
and the homogeneous dimension of RVm is then
Qm =
QI ∗
QJ ∗
r =Pj∈I ∗
a =Pj∈J ∗
a
r
Qj τS(j, kr)
if xm = xI ∗
r ;
Qj τT (j, ℓa)
if xm = xJ ∗
a .
9.7 The matrix ES,T,W
Let
Γ(S, T, W ) =n(ir1, . . . , irα, ja1 , . . . , jaβ ) ∈ Zα+β which satisfy (9.13)o.
Thus π : ΓW (S, T ) → Γ(S, T, W ).
Proposition 9.9. The mapping π : ΓW (S, T ) → Γ(S, T, W ) is onto.
Proof. Thus suppose that the (α + β)-tuple (ir1 , . . . , irα, ja1, . . . , jaβ ) satisfies the inequalties
in equation (9.13). First observe that the first two lines of inequalities describe the projections
of Γ(ES) and Γ(ET ) onto the smaller sets of coordinates. Since ES and ET satisfy the basic
hypotheses (2.10), it follows from Lemma 14.4 that there exists (i1, . . . , is) ∈ Γ(ES) and
(t1, . . . , tq) ∈ Γ(ET ) such that
π(i1, . . . , is, j1, . . . , jq) = (ir1, . . . , irα, ja1, . . . , jaβ ).
We claim that (i1, . . . , is, j1, . . . , jq) ∈ ΓW (S, T ), so we need to check all the inequalities in
equation (9.8). The first two lines are clearly satisfied, since they only depend on the fact
that (i1, . . . , is) ∈ Γ(ES) and (j1, . . . , jq) ∈ Γ(ET ). Also the second two lines are satisfied
provided that r = rj and a = am for some j ∈ {1, . . . , α} and m ∈ {1, . . . , β}. It remains
to check what happens if ir is a free index or ja is a free index. But as we observed in the
summary on page 66, if ir is a free index and Ir ∩ Ja 6= ∅ then ΛW (Ir, Ja) = ∞ so there is no
third line in the inequalities. Also if ir is a free index, then the scale of the dilation τS(j, kr)ir
must be smaller than the scale of the dilation τT (j, ℓa)ja for every j ∈ Ir ∩ Ja. This says that
the ratio ir/ja must lie in the largest (i.e. the unbounded) interval, and this is the statement
of the fourth inequality. A similar argument works if ja is a free index.
The cone Γ(S, T, W ) ⊂ Rα+β is defined by the inequalities in (9.13), which, using the
notation in Section 14.2, is defined by a partial matrix BW , with indices varying in the
disjoint union S∗ ⊔ T ∗ and given by
b(µ, ν) =
eS(µ, ν)
eT (µ, ν)
ΛW (Iµ, Jν)
ΛW (Jµ, Iν)
if µ, ν ∈ S∗
if µ, ν ∈ T ∗
if µ ∈ S∗ , ν ∈ T ∗
if µ ∈ T ∗ , ν ∈ S∗.
(9.15)
70
Lemma 9.10. The matrix BW is connected.
Proof. We must prove that, given µ 6= ν ∈ S∗ ⊔ T ∗, there are λ1, . . . , λa ∈ S∗ ⊔ T ∗ such that
bW (µ, λ1)bW (λ1, λ2) · · · bW (λa, ν) < ∞.
There is no doubt that this holds if both µ and ν belong either to S∗ or to T ∗ because all
the eS and eT are finite. For the other two cases it is necessary and sufficient to prove that
there exist r ∈ S∗, a ∈ T ∗ with ΛW (Ir, Ja) < ∞ and r′ ∈ S∗, a′ ∈ T ∗ with ΛW (Ja′ , Ir′ ) < ∞
(equivalently with w(Ir, Ja) 6= 0 and w(Ir′ , Ja′ ) 6= mr′,a′).
Assume that one of these conditions is violated, e.g., that w(Ir, Ja) = 0 for all r ∈ S∗, a ∈
r for every r, a. Hence T ∗ = ∅ and the proof is
T ∗. By (9.9), this implies that Ir ∩ Ja ⊂ I ∗
completed.
By Lemma 14.3, there is matrix ES,T,W which satisfies the basic hypotheses (2.10) so that
we have Γ(S, T, W ) = Γ(ES,T,W ).
Proposition 9.11. If (m1, . . . , mα+β) ∈ ΓZ(ES,T,W ), the corresponding dilation of RN given
by
(x1, . . . , xα+β) −→ (2−m1 ◦ x1, . . . , 2−mα+β ◦ xα+β)
is still compatible (in the sense of Section 8.1) with the nilpotent Lie group structure.
Proof. Let 1 ≤ l1 < l2 ≤ n. Then xl1 and xl2 are coordinates that belong to one of the
sets {I ∗
ak , 1 ≤ k ≤ β}. Write the (α + β)-tuple (m1, . . . , mα+β) =
(ir1, . . . , irα, ja1 , . . . , jaq ). There are thus four cases to consider.
rj , 1 ≤ j ≤ α} or {J ∗
If xl1 ∈ I ∗
r11
and xl2 ∈ I ∗
rl2
base 2 the scales of the corresponding dilations are irl1
eS(rl2 , rl1 )irl1
≤ irl2
< 0 so
, we proceed as in Lemma 8.6. After taking logarithm to
). Now
) and irl2
τS(l2, krl2
τS(l1, kr11
irl2
irl1
≤ eS(rl2 , rl1 ) ≤ τS(krl2
, krl1
) = min
j∈Irl1
e(krl2
e(krl1
, j)
, j)
≤
e(krl2
e(krl1
, l1)
, l1)
≤
e(krl2
, l2)e(l2, l1)
e(krl1
, l1)
≤
e(krl2
e(krl1
, l2)
, l1)
since e(l2, l1) ≤ 1. Thus
irl1
τS(l1, kr11
) =
irl1
e(krl1
, l1)
≤
irl2
e(krl2
, l2)
= irl2
τS(l2, kr12
),
which is the desired inequality. A similar argument works if xl1 ∈ J ∗
a11
and xl2 ∈ J ∗
al2
.
Now suppose xl1 ∈ I ∗
r11
∩ J ∗
al2
base 2 the scales of the corresponding dilations are irl1
as before, it follows that
and xl2 ∈ Irl2
∩ Jal1
. This time, after taking logarithm to
). Just
τS(l1, kr11
) and jal2
τT (l2, jal2
irl1
τS(l1, kr11
) ≤ irl2
τS(l2, kr12
).
But now since l2 ∈ J ∗
al2
j-scale at this variable. Thus
, we know that the i-scale at this variable is less than or equal to the
irl2
τS(l2, kr12
) ≤ jal2
τT (l2, lal2
).
Putting the last two inequalities together gives the required estimate. A similar argument
works if xl1 ∈ Ir11
, and this completes the proof.
and xl2 ∈ I ∗
rl2
∩ J ∗
al1
∩ Jal2
9.8 The class P0(ES,T,W )
71
Given the decomposition RN =Lα
j∗
aj and the matrix ES,T,W , we have the
space of distributions P0(ES,T,W ) and the corresponding space of multipliers M∞(ES,T,W ).
For notational convenience we use the symbol eW to denote the entries of ES,T,W . Recall that
their indices vary in the disjoint union S∗ ⊔ T ∗ of the sets S∗ ⊂ {1, . . . , s} and T ∗ ⊂ {1, . . . , q}
of fixed indices relative to W .
rj ⊕Lβ
j=1 R
j=1 R
I ∗
Lemma 9.12. M∞(ES,T,W ) ⊂ M∞(E) and consequently P0(ES,T,W ) ⊂ P0(E).
Proof. The dual norms for the space M∞(ES,T,W ) are given by
p
bNr,W (ξ) = Xp∈S ∗bnS,p(ξI ∗
bNa,W (ξ) = Xp∈S ∗bnS,p(ξI ∗
p
)1/eW (p,r) + Xb∈T ∗bnT,b(ξJ ∗
)1/eW (Ip,Ja) + Xb∈T ∗bnT,b(ξJ ∗
b
)1/eW (b,r),
b )1/eW (b,a),
r ∈ S∗;
a∈T ∗.
By Corollary 5.4, it suffices to show, for each (j, k), the entries e♯
E♯
S,T,W satisfy the inequality
W (j, k) of the n × n matrix
By (5.5), e♯
W (j, k) = αk
αj
eW (µ, ν), where
e♯
W (j, k) ≤ e(j, k).
(i) µ, ν ∈ S∗ ⊔ T ∗ are such that j ∈ I ∗
µ (if µ ∈ S∗) or j ∈ J ∗
µ (if µ ∈ T ∗) and k ∈ I ∗
ν or J ∗
ν
accordingly;
(ii) αk is the exponent of the norm of ξk in
) = Xk∈I ∗
p
p
bnS,p(ξI ∗
nk(ξk)e(kp,k) (if k ∈ I ∗
p ),
) = Xk∈J ∗
b
b
bnT,b(ξJ ∗
We then have the inequalities
nk(ξk)e(ℓb,k) (if k ∈ J ∗
b ).
e♯
W (j, k) =
e(kν, k)
e(kµ, j)
e(ℓν, k)
e(ℓµ, j)
e(ℓν, k)
e(kµ, j)
e(kν, k)
e(ℓµ, j)
eW (µ, ν)
if µ, ν ∈ S∗ , j ∈ I ∗
µ , k ∈ I ∗
ν
eW (µ, ν)
if µ, ν ∈ T ∗ , j ∈ J ∗
µ , k ∈ J ∗
ν
eW (µ, ν)
if µ ∈ S∗ , ν ∈ T ∗ , j ∈ I ∗
µ , k ∈ J ∗
ν
eW (µ, ν)
if µ ∈ T ∗ , ν ∈ S∗ , j ∈ J ∗
µ , k ∈ I ∗
ν .
If µ, ν ∈ S∗ and j ∈ I ∗
µ , k ∈ I ∗
ν , using the inequality eW (µ, ν) ≤ bW (µ, ν), we have
e♯
W (j, k) ≤
e(kν, k)
e(kµ, j)
bW (µ, ν) =
e(kν, k)
e(kµ, j)
min
l∈Iν
e(kµ, l)
e(kν, l)
≤
e(kµ, k)
e(kµ, j)
≤ e(j, k),
and similarly if µ, ν ∈ T ∗.
Assume now that µ ∈ S∗, ν ∈ T ∗, j ∈ I ∗
ν . In this case the inequality eW (µ, ν) ≤
bW (µ, ν) may not be sufficient (e.g., we may have bW (µ, ν) = ∞). However, we can use the
basic inequality
µ , k ∈ J ∗
eW (µ, ν) ≤ eW (µ, ν′)eW (ν′, ν) ≤ bW (µ, ν′)bW (ν′, ν) ≤ ΛW (Iµ, Jν ′ )τT (ℓν, ℓν ′),
72
which holds for all ν′ ∈ T ∗. Choosing ν′ so that j ∈ Jν ′ , we have, with m = w(µ, ν′),
e♯
W (j, k) ≤
=
≤
=
e(ℓν, k)
e(kµ, j)
e(ℓν, k)
e(kµ, j)
e(ℓν, k)
e(kµ, j)
e(kµ, vm)
e(ℓν ′, vm)
ΛW (Iµ, Jν ′ )τT (ℓν, ℓν ′)
e(kµ, vm)
e(ℓν ′, vm)
e(kµ, vm)
e(ℓν ′, vm)
e(ℓν ′, j)
e(kµ, j)
e(ℓν ′, i)
e(ℓν, i)
min
i∈Jν
e(ℓν ′ , k)
e(ℓν, k)
e(ℓν ′, k)
e(ℓν ′ , j)
.
(9.16)
Since j ∈ I ∗
Iµ ∩ J ν ′
, i.e.,
µ ∩ Jν ′ , it follows from (9.9) that k = vm′ with m′ > m in the ordering (9.5) on
e(kµ, vm)
e(ℓν., vm)
=
τT (vm, ℓν ′)
τS(vm, kµ)
≤
τT (j, ℓν ′ )
τS(j, kµ)
=
e(kµ, j)
e(ℓν ′, j)
.
Hence
e♯
W (j, k) ≤
≤ e(j, k).
e(ℓν ′, k)
e(ℓν ′ , j)
µ , k ∈ I ∗
The last case, with µ ∈ T ∗, ν ∈ S∗, j ∈ J ∗
ν , is treated similarly.
9.9 Estimates for [ϕ]I ∗ [ψ]J
Then
ir
ir
>
≤
I ∗
r =
J ∗
a =
ja
ja
e(kr, j)
e(ℓa, j)
e(kr, j)
e(ℓa, j)
W ∈ W(S, T ). Recall that
Ir∩Ja6=∅nj ∈ Ir ∩ Ja :
q[a=1
Ir∩Ja6=∅nj ∈ Ir ∩ Ja :
s[r=1
Fix marked partitions S = (cid:0)(I1, k1); . . . ; (Is, ks)(cid:1) and T = (cid:0)(J1, ℓ1); . . . ; (Jq, ℓq)(cid:1), and let
∀(cid:0)(i1, . . . , is), (j1, . . . , jq)(cid:1) ∈ ΓES,T ,Wo,
∀(cid:0)(i1, . . . , is), (j1, . . . , jq)(cid:1) ∈ ΓES,T ,Wo.
r 6= ∅(cid:9) = {r1, . . . , rα}
a 6= ∅(cid:9) = {a1, . . . , aβ}
r =(cid:8)xj : j ∈ I ∗
ro if xm = xI ∗
ao if xm = xJ ∗
S∗ =(cid:8)r ∈ {1, . . . , s} : I ∗
T ∗ =(cid:8)a ∈ {1, . . . , q} : J ∗
r(cid:9) for some r ∈ S∗ or is
a(cid:9) for some a ∈ T ∗. In equation (9.12) we introduced a family of
(x1, . . . , xα+β) where each xm is either equal to xI ∗
equal to xJ ∗
dilations on the space Vm by setting
are the fixed indices. Write RN = Lα+β
nλ1/e(kr ,j) · xj : j ∈ I ∗
nλ1/e(ℓa,j) · xj : j ∈ J ∗
j=1 Vj as in subsection 9.6. Thus if x ∈ RN , x =
a =(cid:8)xj : j ∈ J ∗
Let I = (i1, . . . , is) and J = (j1, . . . , jq) with (I, J) ∈ ΓW (S, T ). If ϕ, ψ ∈ C∞
0 (RN ) have
support in the unit ball, it follows from from part (a) of Lemma 9.4 that there is a function
θ ∈ C∞
0 (RN ), normalized with respect to the functions ϕ and ψ, so that
λ ◦ xm =
r ,
a .
[ϕ]I ∗ [ψ]J (x) = 2− Pα+β
j=1 mj Q∗
j θ(2−m1 · x1, . . . , 2−mα+β · xα+β),
73
where Q∗
j is the homogeneous dimension of Vj, and
mj = ir
mj = ja
if xm = xI ∗
if xm = xJ ∗
a .
r
Thus the indices {m1, . . . , mα+β} are some permutation of the indices {r1, . . . , rα, a1, . . . , aβ}.
Note that if S∗∗ = {rα+1, . . . , rs} ⊂ {1, . . . , s} and T ∗∗ = {aβ+1, . . . , aq} ⊂ {1, . . . , q} are the
free indices, they do not appear among the indices {m1, . . . , mα+β}.
Now suppose that ϕ has cancellation in the variables {xk1 , . . . , xks } and that ψ has cancel-
lation in the variables {xℓ1, . . . , xℓq }. In order to obtain additional estimates of the function
θ, we will use part (b) of Lemma 9.4. Using the notation of Lemma 9.4, we have
C = {k1, . . . , ks} ⊂ {1, . . . , n},
D = {ℓ1, . . . , ℓq} ⊂ {1, . . . , n}.
The dilations are given by
λj = 2ir /e(kr ,j)
µk = 2ja/e(ℓa,j)
if j ∈ Ir,
if j ∈ Ja.
The set C1 =(cid:8)kr : λkr < µkr(cid:9) is now the union of two disjoint subsets:
r = ∅o
a = ∅o
C′
C′′
1 =nkr : I ∗
1 =nkr : I ∗
1 =nℓa : J ∗
1 =nℓa : J ∗
D′
D′′
=nkr : r is freeo,
ro =nkr : r is fixed and kr /∈ I ∗
ro.
=nℓa : a is freeo,
ao =nℓa : a is fixed and ℓa /∈ J ∗
ao.
r 6= ∅ and kr /∈ I ∗
The same is true of the set D1 =(cid:8)ℓa : µℓa ≤ λℓa(cid:9):
a 6= ∅ and ℓa /∈ J ∗
The sets C2 and D2 are given by
Define maps
and note that
ro,
C2 =nkr : r is fixed and kr ∈ I ∗
D2 =nℓa : a is fixed and ℓa ∈ J ∗
ao.
σ1 :(cid:8)1, . . . , s(cid:9) →(cid:8)1, . . . , q(cid:9) such that kr ∈ Jσ1(r),
σ2 :(cid:8)1, . . . , q(cid:9) →(cid:8)1, . . . , s(cid:9) such that ℓa ∈ Iσ2(a),
1 ≤ r ≤ s and r is free =⇒ σ1(r) is fixed,
1 ≤ a ≤ q and a is free =⇒ σ2(a) is fixed.
Applying part (b) of Lemma 9.4 we have the following result.
Lemma 9.13. Let (I, J) ∈ Γ(ES,T,W ), let ϕ, ψ be normalized bump functions, and suppose
that ϕ has cancellation in the variables {xk1, . . . , xks } and that ψ has cancellation in the
variables {xℓ1, . . . xℓq }. Write [ϕ]I ∗ [ψ]J = [θ]M as in Lemma 9.3. There exists ǫ > 0
depending only on the group structure so that θ can be written as a finite sum of terms of the
form
Ykr∈C ′
1
2−ǫG1(ir )Ykr ∈C ′′
1
2−ǫG1(ir )Yℓa∈D′
1
2−ǫH1(ja) Yℓa∈D′′
1
2−ǫH1(a)Ykr ∈C ′
2
2−ǫG2(r)Yℓa∈D′
2
2−ǫH2(ja) θC ′′
2 ,D′′
2
where the factors are defined as follows.
74
a) C2and D2 are each partitioned into two disjoint subsets:
2 ∩ C′′
2 ∩ D′′
C2 = C′
D2 = D′
2 , C′
2 , D′
2 ∪ C′′
2 ∪ D′′
2 = ∅,
2 = ∅.
b) The functions(cid:8)θC ′′
c) The function θC ′′
2 ,D′′
2
2 ,D′′
2(cid:9) ⊂ C∞
0 (Rn) are normalized relative to ϕ and ψ.
has cancellation in each variable xm = xI ∗
r if r ∈ C′′
2 .
d) The function θC ′′
2 ,D′′
2
has cancellation in each variable xm = xJ ∗
a if a ∈ D′′
2 .
e) The expressions Gj (r) and Hj(a) are given by:
G1(ir) = minn
H1(ja) = minn
jσ1(r)
e(ℓσ1(r), kr)
kσ2(a)
e(kσ2(a), ℓa)
G2(ir) = ir+1 − ir;
H2(ja) = ja+1 − ja.
− ir, ir+1 − iro;
− ja, ja+1 − jao;
The meaning of this result is as follows. The original functions ϕI or ψJ have cancellation
in variables whose indices belong to C0 ∪ C1 ∪ C2 and to D0 ∪ D1 ∪ D2. Then with ϕI ∗ ψJ =
[θI,J ]M , the function θI,J can be written as a sum of terms in which a normalized bump
function ΘI,J
C ′
1,D1/ are multiplied by certain small factors.
A) If r ∈ {1, . . . , s} or a ∈ {1, . . . , q} is a free variable then r ∈ C′
1. If r is free, the
scale of [ϕI ]I in the coordinate xkr must be smaller than the scale of [ψJ ]J in xkr , which
is jσ1(r)/e(ℓσ1(r), kr). Thus there is a gain, either of order −ǫ[jσ1(r)/e(ℓσ1(r), kr) − ir] or of
order −ǫ[ir+1 − ir], due either to integration by parts or to replacing the derivative ∂xkr
with the corresponding invariant differential operator. This leads to the term 2−ǫG1(r),
and there is a similar explanation for the term 2−ǫH1(a). In particular, there is an gain
associated with all free indices.
1 or a ∈ D′
B) If r ∈ {1, . . . , s} or a ∈ {1, . . . , q} is a fixed variable, then r ∈ C′′
1 ∪ C2 or a ∈ D′′
1 ∪ D2.
a) For indices r ∈ C′′
1 , kr /∈ I ∗
r and so the scale of [ϕI ]I in the coordinate xkr must be
smaller than the scale of [ψJ ]J in xkr , which is jσ1(r)/e(ℓσ1(r), kr). Thus there is a
gain, either of order −ǫ[jσ1(r)/e(ℓσ1(r), kr) − ir] or of order −ǫ[ir+1 − ir], due either to
integration by parts or to replacing the derivative ∂xkr with the corresponding invariant
differential operator. This leads to the term 2−ǫG1(r) and there is a similar explanation
for the term 2−ǫH1(a).
b) For an index r ∈ C2, it may happen that the cancelation of [ϕI ]I in the variable xkr
2 be the
leads to cancellation in this variable of the convolution [ϕI ]I ∗ [ψJ ]J . We let C′′
elements of C2 where this happens and C′
does have cancellation in the variables xI ∗
have cancellation in the variables x∗
2 be the complementary set. Then θI,J
C ′′
r for r ∈ C′′
2 . In the same way,θI,J
C ′′
2 ,D′′
2
does
2 ,D′′
2
c) For indices r ∈ C′
2 = C2 \ C′′
2 , since cancellation does not persist
in the convolution, there is an exponential gain of order −ǫ[ir+1 − ir or −ǫ[ja+1 − ja
due replacing an ordinary derivative by a left- or right-invariant operator. This leads
to the terms 2−ǫG2(r) and 2−ǫH2(a).
2 or a ∈ D′
Ja for a ∈ D′′
2 .
2 = D2 \ D′′
Remarks 9.14. Let (I, J) ∈ Γ(ES,T,W ).
75
1) If kr ∈ C′
1 then ir ≤ jσ1(r)e(ℓσ1(r), kr)−1 and σ1(r) is a fixed index. Similarly, if ℓa ∈ D′
1
then ja ≤ kσ2(a)e(kσ2(a), ℓa)−1 and σ2(a) is a fixed index.
2) If kr ∈ C′′
1 then σ1(r) is also fixed, and
ir
jσ1(r)
≥ τW (Jσ1(r), Ir)−1 ≥ ΛW (Jσ1(r), Ir)−1 =
e(kr, vmr,σ1(r) )
e(ℓσ1(r), vmr,σ1(r) )
≥ e(ℓσ1(r), kr)−1
where the second inequality follows from the definition of τW since J ∗
σ(r) ∩ Ir 6= ∅, and
the last inequality follows from the basic hypothesis. Thus ir − e(ℓσ1(r), kr)−1jσ1(r) ≤ 0.
Similarly, if ja ∈ D′′
1 , it follows that ja − e(kσ2(a), ℓa)−1iσ2(a) ≤ 0.
9.10 Proof of Theorem 9.2
We have seen on page 61 that Theorem 9.2 follows from Lemma 9.3 which deals with sums
of convolutions [ϕI ]I ∗ [ψJ ]J where (I, J) ∈ ΓZ(ES) × ΓZ(ET ). In section 9.4 we partitioned
ΓZ(ES) × ΓZ(ET ) into disjoint sets ΓW (S, T ) for W ∈ W(S, T ). Thus the proof of Lemma
9.3, and hence Theorem 9.2, follows from the following result.
Lemma 9.15. If F ⊂ ΓW (S, T ) is any finite set, then
HF = X(S,T )∈F
[ϕI ]I ∗ [ψJ ]J ∈ P0(ES,T,W )
with constants independent of the finite set F . Moreover, limF րΓW (S,T ) HF = HS,T,W ∈
P0(ES,T,W ) with convergence in the sense of distributions.
In particular, it follows from
Lemma 9.12 that HS,T,W ∈ P0(E).
Proof. Recall from page 60 that
X(I,J)∈ΓW (S,T )
[ϕI ]I ∗ [ψJ ]J = XM∈M(S,T,W )h X(I,J)∈F (M)
θI,JiM
= XM∈M(S,T,W )
[ΘM ]M .
We first show that for each M ∈ M(S, T, W ), the sum P(I,J)∈F (M) θI,J converges to a
normalized bump function. Given M = (m1, . . . , mα+β), this determines the indices ir and
ja for which r and a are fixed; i.e. r ∈ {r1, . . . , rα} and a ∈ {a1, . . . , aβ}. Thus ir and ja
belong to the sets C′′
1 , C2, D2. We must sum over the remaining indices corresponding
to free variables r and a for which kr ∈ C′
1. We write the free variables as
1 and ℓa ∈ D′
1 , D′′
finite sum of terms (indexed by subsets C′′
{rα+1, . . . , rs} and {aβ+1, . . . , aq}. Using Lemma 9.13 it follows that P(I,J)∈F (M) θI,J is a
Ykr ∈C ′′
2−ǫG1(r) Yℓa∈D′′
2−ǫH1(a) Ykr∈C ′
2 ⊂ D2) of the form
2 ⊂ C2 and D′′
2−ǫH2(a)
1
1
2−ǫH1(av ) θC ′′
2 ,D′′
2i
2
2−ǫG2(r) Yℓa∈D′
h X(irα+1 ,...,irs )∈C ′
sYv=α+1
1
2
(jaβ+1 ,...,jaq )∈D′
1
2−ǫG1(rv )
qYv=β+1
where each function θC ′′
if a ∈ D′′
2 .
2 ,D′′
2
has cancellation in the variables xm = xI ∗
r if r ∈ C′′
2 and xm = xJ ∗
a
76
We have
X(irα+1 ,...,irs )∈C ′
(jaβ+1 ,...,jaq )∈D′
1
1
2−ǫG1(rv)
sYv=α+1
qYv=β+1
2−ǫH1(av ) θC ′′
2 ,D′′
2
irv ≤jσ1 (rv )e(ℓσ1 (rv ),krv )
≤
sYv=α+1h
qYv=β+1"
X
X
jav ≤iσ2 (av )e(kσ2 (av ),ℓav )
2−ǫ[jσ1(rv )e(ℓσ1(rv ),krv )−irv ]i
2−ǫ[iσ2(av )e(kσ2 (av ),ℓav )−jav ]# < ∞.
It follows thatP(I,J)∈F (M) θI,J is a finite sum of terms of the form
2−ǫG1(r) Yℓa∈D′′
1
2−ǫH1(a) Ykr ∈C ′
2
2−ǫG2(r) Yℓa∈D′
2
Ykr ∈C ′′
1
2−ǫH2(a) θI,J
M,C ′′
2 ,D′′
2
where θI,J
M,C ′′
2 ,D′′
2
is a normalized bump function. It remains to show that
XM∈M(S,T,W ) Ykr ∈C ′′
1
2−ǫG1(r) Yℓa∈D′′
1
2−ǫH1(a) Ykr∈C ′
2
2−ǫG2(r) Yℓa∈D′
2
2−ǫH2(a) θI,J
M,C ′′
2 ,D′′
2
converges in the sense of distributions to an element of P0(ES,T,W ). However, this follows from
Theorem 5.13 once we observe that each term in the sum is a function with weak cancellation
with parameter ǫ > 0 relative to the multi-index M . This completes the proof.
10 Convolution of Calder´on-Zygmund kernels
In this section we study the convolution of two or more compactly supported Calder´on-
Zygmund kernels with different homogeneities, and show that the result belongs to an appro-
priate class P0(E).
10.1 Calder´on-Zygmund kernels
Recalling the notion of Calder´on-Zygmund kernel given in Definition 7.1, we define the cor-
responding class of kernels with rapid decay at infinity.
Definition 10.1. CZa,0 is the space of tempered distributions K on RN such that
(a) away from the origin, K is given by integration against a smooth function K satisfying
the differential inequalities
for every γ and M ;
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ Cγ,M Na(x)−Qa−[[γ]]a(cid:0)1 + Na(x)(cid:1)−M
,
(b) there is a constant C > 0 so that for any normalized bump function ψ ∈ C∞
It follows from the results of Section 4 that, for distributions K ∈ CZa,0, the corresponding
support in the unit ball and for any R > 0,(cid:12)(cid:12)(cid:10)K, ψR(cid:11)(cid:12)(cid:12) ≤ C, where ψR(x) = ψ(R ·a x).
multipliers m = bK are characterized by the differential inequalities
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ Cγ(cid:2)1 + Na(ξ)(cid:3)−[[γ]]a .
0 (RN ) with
These differential inequalities characterize the multipliers belonging to the class M∞(Ea)
where Ea is the matrix in (7.3). It follows that, if K ∈ CZa,0, then K ∈ P0(Ea). Moreover,
the matrix Ea is doubly monotone if and only if a1 ≥ a2 ≥ · · · ≥ an. Hence this is the
compatibility condition with an underlying nilpotent Lie group structure in the sense of
Section 8.1.
77
Finally, it follows from our decomposition result that if K ∈ P0(Ea), there exists Φ ∈
S(RN ) and a uniformly bounded family {ϕj} ⊂ C∞
j, and, in the sense of distributions,
0 (RN ) such thatRRN ϕj(x) dx = 0 for all
K(x) =
0Xj=−∞
[ϕj]j(x) + Φ(x) =
0Xj=−∞
2−jQa ϕj(2−ja1 · x1, . . . , 2−jan · xn) + Φ(x).
(10.1)
10.2 A general convolution theorem
Let K1, . . . , Kp be Calder´on-Zygmund kernels on a homogeneous nilpotent Lie group G ∼=
RN = RC1 ⊕ · · · ⊕ RCn. Suppose that the kernel Kℓ is in the class CZaℓ,0 associated with the
dilations λ ·aℓ x =(cid:0)λaℓ
well as the compatibility condition aℓ
1 · x1, . . . , λaℓ
n · xn(cid:1). If G is not abelian, we assume that (8.7) holds, as
j+1 for all j, ℓ.
j ≥ aℓ
Theorem 10.2.
(a) The convolution K1 ∗ · · · ∗ Kp belongs to the class P0(E) where e(j, k) = max1≤ℓ≤p
aℓ
j
aℓ
k
.
(b) Conversely, every E arises in this way. Precisely, let E = {e(j, k)} be any n × n matrix
satisfying (2.10). There are n sequences ak such that e(j, k) = max1≤ℓ≤n
doubly monotone, each ak can be taken non-increasing.
aℓ
j
aℓ
k
. If E is
Proof. We have the identity CZaℓ,0 = P0(Eℓ), where eℓ(j, k) = (cid:0)aℓ
tion 5.2, each Kℓ ∈ P0(E). Moreover, if aℓ
part (a) follows from Theorems 9.1 and 9.2.
. By Proposi-
j+1 for all j, ℓ, E is doubly monotone. Then
j(cid:1)(cid:0)aℓ
k(cid:1)−1
j ≥ aℓ
To establish (b), put aℓ
2 ≥ · · · ≥ aℓ
1 ≥ aℓ
j+1, so aℓ
aℓ
j = e(j, ℓ). Since E is doubly monotone, aℓ
n. For any 1 ≤ j, ℓ ≤ n we have
j = e(j, ℓ) ≥ e(j + 1, ℓ) =
aℓ
j
aℓ
k
=
e(j, ℓ)
e(k, ℓ)
≤
e(j, k)e(k, ℓ)
e(k, ℓ)
= e(j, k)
≤ e(j, k). On the other hand
e(j, k)
e(k, k)
≤ max
1≤ℓ≤n
e(j, ℓ)
e(k, ℓ)
= max
1≤ℓ≤n
aℓ
j
aℓ
k
.
so max1≤ℓ≤m(cid:0)aℓ
j(cid:1)(cid:0)aℓ
k(cid:1)−1
e(j, k) =
This completes the proof.
10.3 Convolution of two Calder´on-Zygmund kernels
Part (a) of Theorem 10.2 determines the minimal class P0(E) containing a finite family of
Calder´on-Zygmund classes CZaℓ,0 and part (b) shows that every class P0(E) arises in this
way. It is possible, however, that a convolution
with Kj ∈ CZaj,0, is contained in a proper subalgebra of the class P0(E) which appears in
Theorem 10.2 (a).
K1 ∗ K2 ∗ · · · ∗ Kp,
78
We show that such a situation occurs with two Calder´on-Zygmund classes having different
homogeneities. Let a = (a1, . . . , an) and b = (b1, . . . , bn) with aj ≥ aj+1 and bj ≥ bj+1 for
all j, and set
λ ·a x = (λa1 x1, . . . , λan xn) ,
λ ·b x =(cid:0)λb1 x1, . . . , λbn xn(cid:1) .
Let K ∈ P0(Ea) and L ∈ P0(Eb). If we apply part (a) of Theorem 10.2, we obtain that
K ∗ L ∈ P0(E), where the entries of E are
Notice that
bj
,
ak
bko.
e(j, k) = maxn aj
e(j, k)e(k, j) = maxn bj/aj
bj/ajo,
bk/ak
bk/ak
,
(10.2)
so that the reduced rank of E can be as large as n. We will show that K ∗ L belongs to the
sum of n − 1 classes P0(Em) where each Em has reduced rank equal to 2.
As in (10.1), write
K(x) =
L(x) =
0Xj=−∞
0Xk=−∞
[ϕj]j(x) + Φ(x) =
2−jQa ϕj (2−ja1 · x1, . . . , 2−jan · xn) + Φ(x),
[ψk]k(x) + Ψ(x) =
2−kQb ψk(2−kb1 · x1, . . . , 2−kbn · xn) + Ψ(x).
0Xj=−∞
0Xk=−∞
Since K ∗ Ψ, Φ ∗ L, Φ ∗ Ψ ∈ S(RN ), we only need to analyze the double sum
0Xj=−∞
[ϕj ]j ∗ [ψk]k.
0Xk=−∞
(10.3)
Let σ be the permutation of {1, . . . , n} so that bσ(1)
aσ(1)
will suppose that all the inequalities are strict21.
≤ bσ(2)
aσ(2)
≤ · · · ≤ bσ(n)
aσ(n)
. For simplicity, we
We split the double sum in (10.3) into n + 1 two-parameter sums. Set
Nm = X(j,k)∈Γm
[ϕj ]j ∗ [ψk]k, 0 ≤ m ≤ n,
(10.4)
where
Γ0 =n(j, k) ∈ Z × Z : j < 0,
Γm =n(j, k) ∈ Z × Z : j < 0,
Γn =n(j, k) ∈ Z × Z : j < 0,
Then clearlyP0
j=−∞P0
k < 0,
k < 0,
k < 0,
≤
j
k
bσ(m)
aσ(m)
bσ(n)
aσ(n)
bσ(1)
aσ(1)o,
<
<
j
k
≤
j
ko.
k=−∞[ϕj ]j ∗ [ψk]k =Pn
m=0 Nm.
21 The proof below shows that this restriction does not affect the validity of the final result.
bσ(m+1)
aσ(m+1)o, 1 ≤ m ≤ n − 1,
(10.5)
79
Lemma 10.3. For 0 ≤ m ≤ n let Am = {σ(1), . . . , σ(m)} and Bm = {σ(m + 1), . . . , σ(n)}.
Then the distribution Nm ∈ P0(Em) where Em = {em(j, k)} is the n × n matrix whose entries
are given by
em(j, k) =
em(j, k) =
em(j, k) =
em(j, k) =
bj
bk
bjaσ(m+1)
akbσ(m+1)
ajbσ(m)
bkaσ(m)
aj
ak
if j ∈ Am and k ∈ Am,
if j ∈ Am and k ∈ Bm,
if j ∈ Bm and k ∈ Am,
if j ∈ Bm and k ∈ Bm.
Proof. We use the fact that [ϕj ]j ∗ [ψk]k(x) = 2− Pn
lp = jap ∨ kbp. Now if (j, k) ∈ Γm then
p=1 lpQp θj,k(2−l1 · x1, . . . , 2−ln · xn) where
kbp
if p ∈ {σ(1), . . . , σ(m)} = Am,
jap
if p ∈ {σ(m + 1), . . . , σ(n)} = Bm.
(10.6)
lp =
If L(j, k) is the multiindex with components lp as above, one easily verifies that
(j, k) ∈ Γm =⇒ L(j, k) ∈ Γ(Em).
Consider first the case 1 ≤ m ≤ n − 1, where Am and Bm are both nonempty. Then the map
(j, k) ∈ Γm 7−→ L(j, k) is injective. Since each term [ϕj]j ∗ [ψk]k has weak cancellation, it
follows22 from Lemma 5.13 that the series (10.4) convergens to a kernel in the class P0(Em).
Suppose now that m = 0 (the case m = n can be treated in the same way). In this case
the matrix E0 has rank 1 and the statement is that N0 ∈ CZa,0. Notice that L(j, k) = ja
only depends on j. Hence for each j we must verify that
Xk≤
aσ(1)
bσ(1)
j
[ϕj]j ∗ [ψk]k = [eϕj ]j
with uniformly bounded functionseϕj with mean value zero (the scaling by j being as in (10.1)).
It follows from Lemma 9.4 that each term in the sum is a normalized bump function times a
factor that decreases exponentially with k. The mean value zero property is obvious.
Remarks 10.4.
1. Recall that
< · · · <
< · · · <
, and that
z
bσ(1)
aσ(1)
Am
}
<
bσ(2)
aσ(2)
{
bσ(m)
aσ(m)
z
<
bσ(m+1)
aσ(m+1)
Bm
}
bσ(n−1)
aσ(n−1)
<
{
bσ(n)
aσ(n)
a1 ≥ a2 ≥ · · · ≥ an, b1 ≥ b2 ≥ · · · ≥ bn. We claim that e(j + 1, j) ≤ 1 for 1 ≤ j ≤ n − 1.
• If j, j + 1 ∈ Am, then since bj+1 ≤ bj,
e(j + 1, j) =
bj+1
bj
≤
bj
bj
= 1.;
• If j ∈ Am and j + 1 ∈ Bm then aj+1
bj+1
≤ aσ(m+1)
bσ(m+1)
≤ aσ(m)
bσ(m)
and so
e(j + 1, j) =
aj+1bσ(m)
bjaσ(m)
≤
aj+1bσ(m)
bj+1aσ(m)
≤ 1.
22 The fact that the components lp are not integers is irrelevant as long as they belong to a fixed lattice.
80
• If j ∈ Bm and j + 1 ∈ Am then bj+1
aj+1
≤ bσ(m)
aσ(m)
and so
e(j + 1, j) =
bj+1aσ(m+1)
ajbσ(m+1)
≤
bj+1aσ(m+1)
aj+1bσ(m+1)
≤
bσ(m)aσ(m+1)
aσ(m)bσ(m+1)
≤ 1.
• If j, j + 1 ∈ Bm, then since aj+1 ≤ aj,
e(j + 1, j) =
aj+1
aj
≤ 1.
It now follows from Proposition 8.3 that the matrix Em is doubly monotone.
2. For m = 0, n, Em has rank one and N0 ∈ CZa,0, Nn ∈ CZb,0. For 1 ≤ m ≤ n − 1, we can
also think of the distribution Nm on the coarser decomposition
with exponents, in the notation of Section 5.3, αm = (αm,1, . . . , αm,n) given by
RN =(cid:0) ⊕j∈Am RCj(cid:1) +(cid:0) ⊕j∈Bm RCj(cid:1).
bj
1
aj
αm,j =( 1
m =" 1
E♭
bσ(m)
aσ(m)
if j ∈ Am
if j ∈ Bm.
aσ(m+1)
bσ(m+1)
1
# .
(Note that the product of the off-diagonal elements is at least 1.)
3. The proof of Lemma 10.3 shows that, if G is the additive group RN , the inclusion (10.7)
holds without any monotonicity assumption on the aj and the bj and with the same defi-
nition of σ(m).23
Proposition 10.5. We have the inclusion
CZa,0 ∗ CZb,0 ⊆
P0(Em),
n−1Xm=1
(10.7)
m=1 P0(Em) is contained in P0(E), where E is the matrix with entries
Moreover, the classPn−1
(10.2), and is closed under convolution.
Proof. The inclusion follows since N0 ∈ P0(E1), Nn ∈ P0(En−1). By Proposition 5.2, the
first part of the statement follows from the inequalities
which can be easily verified. To prove the second part of the statement, observe that each
K ∈Pn−1
m=1 P0(Em) decomposes as
em(j, k) ≤ maxn aj
ak
,
bj
bko,
K =
n−1Xm=1
Nm +
Ml,
nXl=1
23 This can also be proved more directly by decomposing the multiplier µ = bK bL. Assuming, as we may, that
the variables have been ordered so that σ(m) = m for every m, µ can be decomposed as a sum µ = Pn−1
m=1 µm,
where µm ∈ M∞(Em) and Bc ∩ supp µm is contained in the set where
Nb(ξ1, . . . , ξm)bm ≤ ANa(ξm, . . . , ξn)am ,
Na(ξm+1, . . . , ξn)am+1 ≤ ANb(ξ1, . . . , ξm+1)bm+1 .
where Nm ∈ P0(Em) has the form
Nm = X(j,k)∈Γm
[ϕj,k]L(j,k),
81
where the ϕj,k are uniformly bounded and have cancellation in the variables xA(m) and xB(m),
and Ml ∈ CZγ l,0, where γl = (γl
1, . . . , γl
n) and
m =(aσ(l)bm if bm
am
bσ(l)am if bm
am
γl
≤ bσ(l)
aσ(l)
≥ bσ(l)
aσ(l)
,
.
Given a second kernel K′ =Pn−1
Nm ∗ N ′
m′ ,
m=1 N ′
m +Pn
Nm ∗ M′
l′ , Ml ∗ N ′
m′, Ml ∗ M′
l′.
l=1 M′
l, we consider the convolutions
We prove that, assuming m ≤ m′, Nm ∗ N ′
p=m P0(Ep). The proofs of the following
statements are left to the reader:
m′ ∈Pm′
Nm ∗ M′
p=m P0(Ep)
p=l′ P0(Ep)
if m < l′
if m ≥ l′.
(similarly for Ml ∗ N ′
m′)
Ml ∗ M′
P0(Ep)
(l < l′).
l′ ∈
l′ ∈(Pl′−1
Pm
l′−1Xp=l
Nm = X(j,k)∈Γm
Let
[ϕj,k]L(j,k),
N ′
m′ = X(j′,k′)∈Γm′
[ψj′,k′
]L′(j′,k′),
where the components of L′(j′, k′) are defined by (10.6) with m′ in place of m. We isolate a
single term [ϕj,k]L(j,k) ∗ [ψj′,k′
]L′(j′,k′) in the convolution, which is scaled by the multi-index
L(j, k) ∨ L′(j′, k′) = (lp)1≤p≤n, where
L(j, k)p =(kbp
jap
if p ∈ Am
if p ∈ Bm
L′(j′, k′)p =(k′bp
j′ap
if p ∈ Am′
if p ∈ Bm′ .
Consider first the case k ≤ k′.
If p ∈ Am ⊂ Am′ , then lp = k′bp.
Assume now that r, s ∈ {m + 1, . . . , m′}, r < s. Then p = σ(r), q = σ(s) ∈ Bm ∩ Am′. If
jap ≥ k′bp, then
If
j ≥ k′ bp
ap
≥ k′ bq
aq
.
m′′ = max(cid:8)r : k′bσ(r) > jaσ(r)(cid:9),
if p = σ(r) with r ≤ m′′
if p = σ(r) with m′′ < r ≤ m′.
lp =(k′bp
jap
(10.8)
(10.9)
Assume that R′ is nonempty, i.e., that, for q = σ(m′), k′bq ≤ jaq. Since (j′, k′) ∈ Γm′ , we
have
j′ < k′ bq
aq
≤ j.
82
Therefore lp = jap for all p ∈ Bm′. To conclude,
lp =(k′bp
jap
if p ∈ Am′′
if p ∈ Bm′′ .
Notice that (j, k′) ∈ Γm′′ by (10.9). We then have
X(j,k)∈Γm
(j′,k′)∈Γm′
k≤k′
[ϕj,k]L(j,k) ∗ [ψj′,k′
]L′(j′,k′) =
m′Xm′′=m X(j,k′)∈Γm′′ X(j,k)∈Γm
(j′,k′)∈Γm′
k≤k′ , j′≤j
[ϕj,k]L(j,k) ∗ [ψj′,k′
]L′(j′,k′).
(10.10)
By Lemma 9.4, the terms in the innermost sum are uniformly bounded up to a factor which
decays exponentially in j′ and k and have weak cancellation in the variables xAm′′ , xBm′′ . So
the sum in (10.10) belongs toPm′
m′′=m P0(Em′′).
Consider now k > k′. Then lp = kap for p ∈ Am. Since (j, k) ∈ Γm, for p = σ(m + 1)
we have jap ≥ kbp > k′bp. Hence lp = jap. By (10.8), we obtain that lp = jap for all
p ∈ Bm ∩ Am′.
In particular, since (j′, k′) ∈ Γm′, for p = σ(m′) we have jap ≥ k′bp > j′ap so that j > j′.
Then lp = jap for all p ∈ Bm′ . To conclude, L(j, k) ∨ L′(j′, k′) = L(j, k) ∈ Γm and
X(j,k)∈Γm
(j′,k′)∈Γm′
k>k′
[ϕj,k]L(j,k) ∗ [ψj′,k′
]L′(j′,k′) =
X(j,k)∈Γm X(j′,k′)∈Γm′
k′<k , j′<j
[ϕj,k]L(j,k) ∗ [ψj′,k′
]L′(j′,k′).
(10.11)
Repeating the same arguments used in the other case, we conclude that the sum in (10.11)
belongs to P0(Em).
11 Two-flag kernels and multipliers
In this Section we show the surprising fact, already mentioned in Section 1.2, that distributions
which are simultaneously flag kernels for two opposite flags are in fact the elements of an
appropriate class P(E). We will recall the definition of flag-kernel in Section 11.1, and define
two-flag kernels in Section 11.2. We consider the special case of step-two flags (which is
considerably simpler than the general case) in Section 11.3, and the general situation in
Section 11.4. The key point is that if m is the Fourier transform of a compactly supported
distribution satisfying the differential inequalities of flag kernels, then m actually satisfies
improved estimates. This is established in Section 11.5.
11.1 Flag kernels and multipliers
We recall the definition of flag kernel introduced in [NRS01]. Let RN = RC1 ⊕ · · · ⊕ RCn,
and use the notation of Section 2, so that nj is a homogeneous norm on RCj , which has
homogeneous dimension Qj. An n-step flag is a strictly increasing sequence of n subspaces
83
of RN , each of which is a direct sum of the subspaces {RC1, . . . , RCn}. In particular, denote
by F the flag
F :
(0) ( RCn ( RCn−1 ⊕ RCn ( · · · ( RC2 ⊕ · · · ⊕ RCn ( RN .
(11.1)
Given an n-tuple a = (a1, . . . , an) of positive real numbers, we consider the dilations
defined in (7.2) and introduce on RN the partial norms
λ ·a x = (λa1 · x1, . . . , λan · xn)
N ′
j(x) = n1(x1)1/a1 + · · · + nj(xj )1/aj ,
(11.2)
(11.3)
so that N ′
j does not involve the variables xj+1, . . . , xn. Recall that if L = {l1, . . . , lr} ⊂
{1, . . . , n} with l1 < · · · < lr, we denote by RL the space RCl1 ⊕ · · · ⊕ RClr . The "quotient
flag" of F on RL, denoted by FL, is then
(0) ( RClr ( RClr−1 ⊕ RClr ( · · · ( RCl2 ⊕ · · · RClr ( RL
and the corresponding partial norms are given for 1 ≤ j ≤ r by
N ′
L,j(xl1 , . . . , xlj ) = nl1(xl1 )1/al1 + · · · + nlj (xlj )1/alj .
(11.4)
(11.5)
Notice that intrinsic to the flag structure is an ordering of the subspaces RCj , which must
be taken into account in the labeling of variables, norms etc.
Definition 11.1. A flag kernel associated to the flag F and the dilations λ·αx is a distribution
K on RN with the following size estimates and cancellation conditions:
(A) [Differential Inequalities] Away from the subspace where x1 = 0 the distribution K is
given by integration against a smooth function K which satisfies
(cid:12)(cid:12)∂γ 1
x1 · · · ∂γn
xn K(x1, . . . , xn)(cid:12)(cid:12) ≤ Cγ
N ′
j(x1, . . . , xj)−aj (Qj +[[γj ]])
nYj=1
on RCl1 ⊕ · · · ⊕ RClr by setting(cid:10)Kψ,R, ϕ(cid:11) =(cid:10)K, ϕ ⊗ ψR(cid:11) for every ϕ ∈ C∞
(B) [Cancellation Conditions] Let L = {l1, . . . , lr} and M = {m1, . . . , ms} be complementary
subsets of {1, . . . , n}, let R = {R1, . . . , Rs} be positive real numbers, and let ψ ∈ C∞
0 (RM )
be a normalized bump function with support in the unit ball. Define a distribution Kψ,R
0 (RL). Then
Kψ,R uniformly satisfies the analogue of the estimates in (a) on the space RL; i.e. away
from the subspacee of RL where xcr = 0, the distribution Kψ,R is given by integration
against a smooth function Kψ,R and for every γ = (γl1 , . . . , γ lr ) ∈ NCl1 ×· · ·×NClr there
is a constant Cγ depending only on the constants in (a) and in particular independent of
ψ and R so that
γl1
xl1 · · · ∂
(cid:12)(cid:12)∂
γlr
xlr Kψ,R(xl1 , . . . , xlr )(cid:12)(cid:12) ≤ Cγ
rYj=1
As usual, if s = n, we require that (cid:10)K, ψR(cid:11) is bounded independently of R.
The following is Theorem 2.3.9 in [NRS01].
N ′
L,j(xl1 , . . . , xlj )−alj (Qlj +[[γ lj
]]).
Theorem 11.2. If K is a flag kernel adapted to the flag (11.1) and the family of dilations
(11.2), the Fourier transform m = bK is function smooth away from the subspace where ξn = 0
which satisfies the differential inequalities
ξ1
(cid:12)(cid:12)∂γ 1
· · · ∂γn
ξm
m(ξ1, . . . , ξn)(cid:12)(cid:12) ≤ Cγ
nYj=1(cid:16)nj(ξj)1/aj + · · · + nn(ξn)1/an(cid:17)−αj [[γ j ]]
.
(11.6)
Conversely, any function satisfying (11.6) is the Fourier transform of a flag kernel adapted
to the flag (11.1) and dilations (11.2).
84
11.2 Pairs of opposite flags
In Section 11.1, we chose a particular ordering of the subspaces {RC1, . . . , RCn} to define
the flag F in (11.1) and a particular family of dilations on RN in (11.2). We can consider
the opposite flag obtained by reversing the indices {1, . . . , n} and choosing a different one-
parameter family of dilations. Our objective is then to study two-flag distributions which are
simultaneously flag kernels for the two different flags and two different homogeneities. Thus
set
(0) ( RCn ( RCn−1 ⊕ RCn ( · · · ( RC2 ⊕ · · · ⊕ RCn ( RN ,
F :
F ⊥ (0) ( RC1 ( RC1 ⊕ RC2 ( · · · ( RC1 ⊕ · · · ⊕ RCn−1 ( RN .
(11.7)
Let a = (a1, . . . , an) and b = (b1, . . . , bn) be two n-tuples of positive real numbers. Define
two one-parameter families of dilations
λ ·a x = (λa1 · x1, . . . , λan · xn),
λ ·b x = (λb1 · x1, . . . , λbn · xn),
(11.8)
and associate the first dilation with the flag F and the second with the flag F ⊥. Along with
satisfying the appropriate cancellation conditions, a two-flag distribution K then satisfies the
following differential inequalities. Away from the set where x1 = 0 or where xn = 0, K is
given by integration against a smooth function K and
j=1(cid:16)n1(x1)aj /a1 + · · · + nj−1(xj−1)aj /aj−1 + nj(xj )(cid:17)−(Qj +[[γ j ]])
Qn
j=1(cid:16)nj(xj) + nj+1(xj+1)bj /bj+1 + · · · + nn(xn)bj /bn(cid:17)−(Qj +[[γ j ]])
Qn
.
.
(cid:12)(cid:12)∂γK(x)(cid:12)(cid:12) ≤ Cγ
(cid:12)(cid:12)∂γm(ξ)(cid:12)(cid:12) ≤ Cγ
11.3 Two-step flags
The corresponding Fourier transform m = bK satisfies
j=1(cid:16)nj(ξj) + nj+1(ξj+1)aj /aj+1 + · · · + nn(ξn)aj /an(cid:17)−[[γj ]]
Qn
j=1(cid:16)n1(ξ1)bj /b1 + · · · + nj−1(ξj−1)bj /bj−1 + nj(ξj)(cid:17)−[[γ j ]]
Qn
We first consider the case of two-step flags, where RN is written as a direct sum of two
subspaces: RN = Rn ⊕ Rm. As we will see, this situation is much simpler than the general
case considered later in Section 11.4. Write elements of RN as (x, y) with x ∈ Rn and y ∈ Rm.
For notational convenience we assume isotropic dilations on each component. There are two
flags F and F ⊥ adapted to this decomposition:
F :
(0) ( Rn ( Rn ⊕ Rm = RN and
F ⊥ :
(0) ( Rm ( Rn ⊕ Rm = RN .
Let a = (a1, a2) and b = (b1, b2) be two pairs of positive real numbers and put
λ ·a (x, y) = (λa1 x, λa2 y),
λ ·b (x, y) = (λb1 x, λb2 y).
(11.9)
Associate the family of dilations λ ·a (x, y) with the flag F and the family λ ·b (x, y) with
the flag F ⊥. A distribution K on Rn × Rm is then a two-flag kernel relative to this data if
it satisfies appropriate cancellation conditions and is given by integration against a smooth
function K which satisfies the differential inequalities
85
x ∂β
y K(x, y)(cid:12)(cid:12) ≤ Cα,β
x−n−α(cid:0)xa2/a1 + y(cid:1)−m−β
(cid:12)(cid:12)∂α
(cid:0)x + yb1/b2(cid:1)−n−α
If m(ξ, η) = bK(ξ, η) is the Fourier transform, then
η m(ξ, η)(cid:12)(cid:12) ≤ Cα,β
(cid:0)ξ + ηa1/a2(cid:1)−α
(cid:12)(cid:12)∂α
ξ−α(cid:0)ξb2/b1 + η(cid:1)−β
y−m−β
η−β
ξ ∂β
It follows that both K and m are smooth away from the origin.
for η 6= 0 (flag F )
for ξ 6= 0 (flag F ⊥)
for x 6= 0 (flag F )
for y 6= 0 (flag F ⊥)
(11.10)
.
(11.11)
Proposition 11.3. Suppose that a1
b1
< a2
b2
, and let
N1(x, y) = x + yb1/b2
N2(x, y) = xa2/a1 + y
be the norms associated with the matrix
E =(cid:20) 1
a2
a1
bN1(ξ, η) = ξ + ηa1/a2
bN2(ξ, η) = ξb2/b1 + η
1(cid:21)
b1
b2
(11.12)
satisfying the basic hypotheses.
If K is a two-flag kernel on Rn × Rm adapted to the dilations given in (11.9), then
(a) the corresponding function K is integrable at infinity;
supported in B(1).
(c) we can write K = K0 + K∞ where K∞ ∈ L1(RN ) ∩ C∞(RN ), and K0 is a two-flag kernel
(b) K ∈ P(E) and bK = m ∈ M(E);
Conversely, if K is in P(E), then K is a two-flag kernel (and bK = m a two-flag multiplier)
on Rn × Rm adapted to the dilations given in (11.9).
Proof. We have already observed that m is smooth away from the origin, and it follows from
equation (11.11) that we have the following estimates for pure derivatives of m:
(cid:12)(cid:12)∂α
ξ m(ξ, η)(cid:12)(cid:12) .(cid:0)ξ + ηa1/a2(cid:1)−α
(cid:12)(cid:12)∂β
η m(ξ, η)(cid:12)(cid:12) .(cid:0)ξb2/b1 + η(cid:1)−β
,
Since the rank of E is greater than 1, it follows from Lemma 7.5 that K is smooth away from
the origin and is integrable at infinity. This establishes (a).
Next let χ ∈ C∞
0 (Rn × Rn) have support in the unit ball, with χ(x, y) ≡ 1 if x + y ≤ 1
2 .
Obviously K0 = χK satisfies the differential inequalities (11.10) and it easy to verify that it
also satisfies the cancellations in Definition 11.1. By (a), K∞ = (1 − χ)K ∈ L1(Rn) ∩ C∞(Rn).
This establishes (c).
Being a two-flag kernel we know that K0 satisfies (11.10). A case by case analysis shows
that these estimates can be combined and improved.
86
(1) Suppose x ≥ yb1/b2 . Then x−1 . (cid:0)x + yb1/b2(cid:1)−1
shows that
and hence the first inequality
∂α
x ∂β
y K0(x, y) .(cid:0)x + yb1/b2(cid:1)−n−α(cid:0)xa2/a1 + y(cid:1)−m−β
.
(11.13)
, and the second inequality again
(2) Suppose y ≥ xa2/a1 . Then y−1 .(cid:0)xa2/a1 + y(cid:1)−1
gives the estimate in (11.13).
(3) Finally suppose x < yb1/b2 and y < xa2/a1 . Here we must have
x < x
a2 b1
a1 b2
and y < y
a2 b1
a1b2
Since a2b1 > a1b2, this means that x > 1 and y > 1, and this is outside the support of
K0. It follows that for all (x, y) ∈ Rn ⊕ Rm we have the improved estimates
∂α
x ∂β
y K(x, y) .(cid:0)x + y
b1
b2(cid:1)−n−α(cid:0)x
a2
a1 + y(cid:1)−m−β
.
This shows that K0 ∈ P0(E) since the cancellations conditions for the flag kernels are the
same as the cancellation conditions for P0(E). This establishes (b) and completes the proof.
The last part of the statement is obvious.
Remarks 11.4.
(1) For a matrix E of the form (11.12), the condition a1
b1
is equivalent to the basic
hypotheses. If the inequality is an equality, E is reducible and the class P(E) consists
of Calder´on-Zygmund kernels.
≤ a2
b2
The arguments used in the proof of Proposition 11.3 can adapted to prove that, if the
opposite inequality a1
holds, then a kernel K ∈ P(E) (defined by the same dif-
b1
ferential inequalities and cancellations as in Definition 2.2) is integrable near 0 instead
than at infinity. This suggests that the essentially local nature of the kernels strongly
depends on the basic hypotheses.
> a2
b2
(2) One may observe that a 2 × 2 matrix satisfying the basic hypotheses can always be put
in the form (11.12) for appropriate aj, bj with a1
. This means that, for every two-
b1
fold decomposition RN = RC1 ⊕ RC2 and every E satisfying the basic hypotheses, the
class P(E) conicides either with a class of two-flag kernels or with a class of Calder´on-
Zygmund kernels, depending on its rank.
≤ a2
b2
As we will see next, this is specific of two-fold decomposition. However, it follows from
Lemma 5.3 that the same also holds for general decompositions, as long as the reduced
rank of E is not greater than 2.
11.4 General two-flag kernels
In the case of the two-step flags, the improved estimates for the kernel followed easily from
the two-flag estimates simply by considering possible cases. However, if the number of norms
is greater than two, this argument no longer works. Observe that the differential inequalities
for the kernel K give no information when x1 = xn = 0 while other xj 6= 0. In particular,
it does not follow from the inequalities alone that the kernel is non-singular away from the
origin. Nevertheless, we have the following result.
Theorem 11.5. Suppose that a1
b1
K be a two-flag kernel for the flags F and F ⊥ with homogeneities a and b respectively.
with at least one strict inequality, and let
≤ · · · ≤ an
bn
≤ a2
b2
(a) The function K is integrable at infinity, and we can write K = K0 + K∞ where K∞ ∈
L1(RN ) ∩ C∞(RN ), and K0 is a two-flag kernel supported in B(1).
(b) The kernel K0 belongs to the class P0(E) associated to the matrix
87
E =
1
a2/a1
a3/a1
...
b1/b2
1
a3/a2
...
b1/b3
b2/b3
1
...
an−1/a1 an−1/a2 an−1/a3
an/a1
an/a3
an/a2
· · ·
· · ·
· · ·
. . .
· · ·
· · ·
b1/bn−1
b2/bn−1
b3/bn−1
...
1
b1/bn
b2/bn
b3/bn
...
bn−1/bn
an/an−1
1
.
(11.14)
(c) Conversely, every kernel in P0(E) is a two-flag kernel for the flags F and F ⊥ with
homogeneities a and b respectively.
Proof. If two ratios aj/bj, ak/bk are equal, the product e(j, k)e(k, j) is 1. By Proposition 5.5,
the matrix E can be reduced to E♭ of the same type. We may then assume that all inequalities
are strict.
The first part of the proof proceeds as the proof of Proposition 11.3. Combining the two
kinds of estimates for pure derivatives of the Fourier transform m, we see that m satisfies
γj
ξj
(cid:12)(cid:12)∂
m(ξ)(cid:12)(cid:12) . bNj(ξ)−[[γ j ]] where
bNj(ξ) = n1(ξ1)bj /b1 + · · · + nj−1(ξj−1)bj /bj−1 + nj(ξj)+
We want to show that e(j, k) ≤ e(j, l)e(l, k) where e(j, k) =(bj/bk
aj/ak
We want to apply Lemma 7.5, and so we need to check that the matrix E satisfies the basic
hypothesis of (2.10) and has rank greater than 1. We proceed to check this.
+ nj+1(ξj+1)aj /aj+1 + · · · + nn(ξn)aj /an .
if j ≤ k
if j ≥ k
. Thus we
need to show
e(j, k) =
e(j, k) =
bj
bk
aj
ak
≤
≤
aj
al
bj
bl
bj
bl
bl
bk
bl
bk
al
ak
bj
bl
aj
al
aj
al
al
ak
al
ak
bl
bk
= e(j, l)e(l, k)
= e(j, l)e(l, k)
= e(j, l)e(l, k)
if l ≤ j ≤ k,
if j ≤ l ≤ k,
if j ≤ k ≤ l,
if j ≤ k,
= e(j, l)e(l, k)
= e(j, l)e(l, k)
= e(j, l)e(l, k)
if l ≥ j ≥ k,
if j ≥ l ≥ k,
if j ≥ k ≥ l,
if j ≥ k.
However it is easy to check that all of these inequalities follow from the assumption in the
Theorem. Thus E satisfies the basic hypotheses. Also one can check that
det(E) =
nYj=2(cid:18)1 −
aj
bj
bj−1
aj−1(cid:19) 6= 0
so in fact the matrix E has rank n. Thus it follows from Lemma 7.5 that K is integrable at
infinity.
The proof of part (b) will use the following result, which shows that estimates for pure
derivatives leads to estimates for mixed derivatives. In order to keep our focus on the proof
of Theorem 11.5, we defer the proof to Section 11.5.
88
Lemma 11.6. Let m ∈ C∞(RN \ {0}). Suppose that for every γj ∈ NCj there is a constant
Cγ j > 0 so that
Then for every γ = (γ 1, . . . , γn) ∈ NC1 × · · · × NCn there is a constant Cγ > 0 so that if
ξ ∈ B(1)c,
γ j
ξj
(cid:12)(cid:12)∂
m(ξ)(cid:12)(cid:12) ≤ Cγ j bNj(ξ)−[[γ j ]].
m(ξ)(cid:12)(cid:12) ≤ Cγ
nYj=1 bNj(ξ)−[[γ j ]].
· · · ∂γ n
ξn
ξ1
(cid:12)(cid:12)∂γ 1
We now turn to the proof of Theorem 11.5 part (b), which is more involved than in the
situation considered in Section 11.3. We start by decomposing the Fourier transform m0 of
the kernel K0 by means of a smooth cutoff function supported in B(2) and equal to 1 on
B(1). We obtain m0 = m′ + m′′, where m′ is in C0(RN ), while m′′ is supported in B(1)c and
satisfies the inequalities (11.11). Denoting by K′, K′′ the inverse Fourier transforms of m′, m′′
respectively, we then have K0 = K′ + K′′.
We consider m′′ first. Since it satisfies the inequalities (11.11), the two types of estimates
can be combined together to give
γj
ξj
(cid:12)(cid:12)∂
m′′(ξ)(cid:12)(cid:12) . bNj(ξ)−[[γ j ]],
(11.15)
whenever we differentiate only in the ξj-variables (pure derivatives). Applying Lemma 11.6
and recalling that m′′ is supported in B(1)c, we obtain inequalities for all derivatives of m′′,
namely
ξ1
(cid:12)(cid:12)∂γ1
· · · ∂γ n
ξn
m′′(ξ)(cid:12)(cid:12) .
nYj=1(cid:0)1 + bNj(ξ)(cid:1)−[[γ j ]]
,
(11.16)
which tells us that m′′ ∈ M∞(E).
By Theorem 4.1 it follows that K′′ ∈ P0(E). Using a cutoff function in the x-variables,
we could then split K′′ as K0 + K′′′, with K0 ∈ P0(E) and K′′′ ∈ S(RN ).
We next observe that K′ ∈ C∞(Rn) because m′ has compact support. It follows that K
coincides with a smooth function away from the origin. We can then apply Theorem 7.5 and
conclude that K is integrable at infinity.
Hence K′ ∈ C∞(RN ) ∩ L1(RN ) and the same holds for K′ + K′′′ = K∞.
Finally, part (c) is an obvious consequence of (b).
Remark 11.7.
As mentioned in Remark 11.4, for a general matrix E with reduced rank larger than 2,
it is not true that the class P(E) coincides with a class of two-flag kernels. In fact, such a
coincidence occurs if and only if the equality e(j, ℓ) = e(j, k)e(k, ℓ) holds whenever the triple
j, k, ℓ is ordered (j < k < l or vice-versa).
The following statement is a direct consequence of Theorem 9.2.
Corollary 11.8. Let K, L be two compactly supported two-flag kernels as in Theorem 11.5,
for the same pair of flags and the same homogeneities. Then the convolution K ∗ L is also a
two-flag kernel of the same type.
11.5 Proof of Lemma 11.6
In this section we complete the proof of Theorem 11.5 by proving Lemma 11.6; i.e. we
show that if m ∈ C∞(RN ) satisfies the estimates for the class of multipliers M(E) for pure
derivatives, then the estimates for mixed derivatives follow automatically. We will use the
following result, which is easily proved by estimating the decay of the Fourier transform of f .
89
Proposition 11.9. Let f ∈ C∞
0 (RN ) be supported in the unit ball. Suppose that
for each j and each γj ∈ NCj . Then for every γ = (γ1, . . . , γn) ∈ NC1 × · · · × NCn we have
γ j
(cid:12)(cid:12)∂
ηj f (η)(cid:12)(cid:12) ≤ Cγ j
f (η)(cid:12)(cid:12) ≤ C′
(cid:12)(cid:12)∂γ 1
· · · ∂γ n
ηn
η1
γ
where the constants {C′
γ} depend only on the constants {Cγ j }.
we have
nj(ξj)e(kr ,j) on RIr . Let C ≥ 1 be a constant such that, for all 1 ≤ r ≤ s,
Proposition 11.10. Let S = {(I1, k1); . . . ; (Is, ks)} ∈ S(n) and let C be the constant from
bnS,r(ξIr + ηIr ) ≤ ChbnS,r(ξIr ) +bnS,r(ηIr )i.
S , we give a quan-
titative estimate of the size of a ball about ξ that we want contained in some larger re-
S . Recall from Definition 3.12 and equation (3.5) that we defined a dilation struc-
Next let S = {(I1, k1); . . . ; (Is, ks)} ∈ S(n). For ξ ∈ B(1)c ∩ bEA
gion bEA′
ture λb·S ξIr = (cid:8)λ1/e(kr ,j) · ξj : j ∈ Ir(cid:9) and a corresponding smooth homogeneous norm
bnS,r(ξIr ) ≈Pj∈Ir
equation (11.17). Let ξ = (ξ1, . . . , ξn) ∈ B(1)c ∩ bEA
nη ∈ RN :bnS,r(ξIr − ηIr ) = Xj∈Ir
Proof. Let ξ ∈ bEA
nj(ηj − ξj)e(kr ,j) < ǫ nkr (ξkr ), 1 ≤ r ≤ so ⊂ bEA′
S . Recall from part (b) of Lemma 3.9 that
nkr (ξkr ) > A−1nj(ξj)e(kr ,j)
and
nkr (ξkr ) > A−1nkt(ξkt )1/τS (kr ,kt) ∀t ∈ {1, . . . , s}, t 6= r.
∀j ∈ Ir, j 6= kr,
where A′ is a constant that depends only on A and C.
S . If ǫ < (2C)−1, then
(11.17)
S
Conversely, if η ∈ B(1)c satisfies
nkr (ηkr ) > A−1nj(ηj)e(kr ,j)
and
nkr (ηkr ) > A−1nkt(ηkt )1/τS(kr ,kt) ∀t ∈ {1, . . . , s}, t 6= r.
∀j ∈ Ir, j 6= kr,
then η ∈ EAτ
S where τ is a constant that depends only on the coefficients {e(j, k)}.
If nj(ηj − ξj)e(kr ,j) < ǫnkr (ξkr ) for 1 ≤ r ≤ s and all j ∈ IR, and if Cǫ < 1
nkr (ξkr ) ≤ C(cid:2)nkr (ξkr −ηkr )+nkr (ηkr )(cid:3) ≤ C(cid:2)ǫnkr (ξkr )+nkr (ηkr )(cid:3) < 1
so that for 1 ≤ r ≤ s,
2 , then
2 nkr (ξkr )+Cnkr (ηkr )
nkr (ξkr ) < 2Cnkr (ηkr ).
(11.18)
Next, if j ∈ Ir,
nj(ηj) ≤ C(cid:2)nj(ηj − ξj) + nj(ξj)(cid:3) < C(cid:2)(cid:0)ǫnk(ξkr )(cid:1)1/e(kr ,j)
and it follows from (11.18) that
+(cid:0)Ankr (ξkr )(cid:1)1/e(kr ,j)(cid:3),
nj(ηj) ≤ C′ nkr (ηkr )1/e(kr ,j)
(11.19)
where C′ depends on C, ǫ, and A. Finally, if 1 ≤ t ≤ s with t 6= r, it follows from (11.19)
that nkt (ηkt ) ≤ C′nkt (ξkt )1/e(kt,kt), and so
where C′′ depends on C, ǫ, and A. It now follows from the inequalities in (11.19) and (11.20)
nkt(ηkt ) ≤ C′(cid:0)Ankr (ξkr )(cid:1)τS (kr ,kt)
≤ C′′ nkr (ηkr )τS (kr ,kt)
(11.20)
S , and this completes the proof.
that η ∈ bEA′
90
We now prove Lemma 11.6.
Let ξ ∈ B(1)c. Then if A > 1 there exists S = (cid:0)(I1, k1); . . . , ; (IS, ks)(cid:1) ∈ S(n), so that
ξ ∈ bEA
S . We have bNj(ξ) ≈ nkr (ξkr )1/e(kr ,j) for every r ∈ {1, . . . , s} and every j ∈ Ir. (Here
the implied constants depend on A.) Thus to establish the Lemma we need to show that for
every γ = (γ 1, . . . , γn) ∈ NC1 × · · · × NCn there is a constant Cγ > 0 so that
ξ1
(cid:12)(cid:12)∂γ 1
· · · ∂γn
ξn
m(ξ)(cid:12)(cid:12) ≤ Cγ
nkr (ξkr )− Pj∈Ir
[[γ j ]]/e(kr ,j).
sYr=1
According to Proposition 11.10, there exists ǫ > 0 and a constant A′ depending only on A so
that
nη ∈ RN : nj(ηj − ξj)e(kr ,j) < ǫ nkr (ξkr ) for 1 ≤ r ≤ s and j ∈ Iro ⊂ bEA′
0 (RIr ) be supported where nj(ηj) ≤ 1 for all j ∈ Ir, with
S .
For 1 ≤ r ≤ s let ϕr ∈ C∞
2 for all j ∈ Ir. Put
bnS,r(ηIr ) ≡ 1 if nj(ηj) ≤ 1
f (η) =(cid:16) sYr=1
ϕr(ηIr )(cid:17) m(cid:16)ξI1 + [ǫ nk1 (ξk1)]b·S ηI1 , . . . , ξIS + [ǫ nks(ξks )]b·S ηIs(cid:17).
Here, the function m is evaluated at the point whose jth-coordinate is
vj = ξj + [ǫnkr (ξkr )]1/e(kr ,j) · ηj
when j ∈ Ir. Clearly f ∈ C∞(RN ) and is supported where nj(ηj) ≤ 1 for all 1 ≤ j ≤ n.
Moreover
if nj(ηj) ≤ 1
2 for all j. If f (η) 6= 0, so that nj(ηj) ≤ 1, then
f (η) = m(cid:16)ξI1 + [ǫ nk1 (ξk1)]b·S ηI1, . . . , ξIS + [ǫ nks(ξks )]b·S ηIs(cid:17)
nj(vj − ξj) = nj(cid:16)nkr ([ǫ ξkr )]1/e(kr ,j) · ηj(cid:17)
= [ǫ nkr (ξkr )]1/e(kr ,j) nj(ηj) ≤ [ǫ nkr (ξkr )]1/e(kr ,j),
γ j
S .
S , we get
Consider 'pure' derivatives ∂
r=1 ϕr(ηIr )(cid:1)
r=1 ϕr(ηIr )(cid:1) are bounded by
ηj f . Some of the derivatives will fall on (cid:0)Qs
constants independent of m. On the other hand, if γ j ∈ N(Ir ), using the chain rule, the
hypothesis about pure derivatives of m, and the fact that we are evaluating m at points in
and it follows that m is evaluated at a point of bEA′
and the rest fall on the term involving m. Derivatives of (cid:0)Qs
bEA′
ηj"m(cid:16)ξI1 + [ǫ nk1(ξk1 )]b·S ηI1 , . . . , ξIS + [ǫ nks(ξks)]b·S ηIs(cid:17)#
ηj mi(cid:16)ξI1 + [ǫ nk1 (ξk1)]b·S ηI1 , . . . , ξIS + [ǫ nks(ξks )]b·S ηIs(cid:17)Yj∈Ir(cid:2)ǫ nkr (ξkr )(cid:3)[[γ j ]]/e(kr ,j)
=h∂
≤ Cγ j Yj∈Ir
nkr (ξkr )−[[γ j ]]/e(kr ,j) Yj∈Ir(cid:2)ǫ nkr (ξkr )(cid:3)[[γ j ]]/e(kr ,j)
≤ Cγ j .
γj
∂
γj
91
Thus f ∈ C∞(RN ) is supported in the unit ball, and pure derivatives are bounded. It follows
from Proposition 11.9 that all derivatives are bounded by constants depending only on the
{Cγ j }. But for nj(ηj) ≤ 1
2 , we have
∂γ 1
η1
· · · ∂γn
ηn
f (η) = [∂γ 1
η1
· · · ∂γn
ηn
m](v1, . . . , vn)
nkr (ξkr )
Pj∈Ir
[[γ j ]]/e(kr ,j).
sYr=1
where vj = ξj + [ǫnkr (ξkr )]1/e(kr ,j) · ηj. When η = 0, the uniform bounds on f give the
desired estimates for the derivatives of m, and this completes the proof.
12 Extended kernels and operators
Operators of the form TKf = f ∗ K with K ∈ P0(E) are "constant coefficient"operators. In
this section we consider a more general class of operators which have "variable coefficients".
Identify the space RN with a nilpotent Lie group G ∼= RN as in Section 8, and consider
kernels of the form K(x, z), with (x, z) ∈ G × G, so that for each x ∈ G, K(x, ·) ∈ P(E),
with smooth dependence on x. More precisely, assume that for each norm · M defined in
(2.13) and each β,
X βK(x, ·)M < ∞ ,
sup
x
(12.1)
where X β is a monomial in the left-invariant vector fields of G acting on the x-variable. We
also assume (unless G = RN as an additive group) the hypotheses of Theorems 8.9 and 9.2
to guarantee that proper kernels define bounded convolution on Lp, 1 < p < ∞ and form a
class closed under convolution.
Kernels of this kind will be said to belong to the "extended class", to distinguish them
from the "proper" kernels in P0(E) which are independent of x. Operators corresponding to
these extended kernels are given formally by
T f (x) =ZG
K(x, y−1x)f (y)dy;
(12.2)
that is, if f ∈ S(RN ), T f (x) = hK(x, ·), fxi, with fx(y) = f (xy−1).
Note that the extended class of operators is left-invariant in the following sense. If a ∈ G
set Laf (x) = f (a−1x). Then L−1
a T La is an operator again of the form (12.2), with K(x, z)
replaced by K(ax, z). Observe that the semi-norms (12.1) of K(x, z) and K(ax, z) are the
same.
Our main results concerning this extended class of kernels and operators are as follows.
Theorem 12.1. Each operator T given by (12.2) is bounded on Lp(RN ),
for 1 < p < ∞.
Theorem 12.2. The class of these operators form an algebra under addition and composition
of operators.
12.1 The Lp boundedness of the operators
Recall that if δr : G → G is the one-parameter group of automorphic dilations of the group
G, the coordinates on RN are chosen so that δr(x1, . . . , xN ) = (rd1 x1, . . . , rdN xN ) with d1 ≤
d2 ≤ · · · ≤ dN . Let · G be a smooth homogeneous norm, so that δrxG = r xG. The
j=1 dj, and d(x, y) = y−1xG is a left-invariant quasi-
homogeneous dimension of G is Q =PN
distance. Let Br(x) =(cid:8)y ∈ G : d(x, y) < r(cid:9). Then(cid:12)(cid:12)Br(x)(cid:12)(cid:12) =(cid:12)(cid:12)Br(0)(cid:12)(cid:12) = rQ(cid:12)(cid:12)B1(0)(cid:12)(cid:12). We will
write Br(0) = Br.
92
with
In proving Theorem 12.1, we make a preliminary reduction to kernels K(x, z) that are
supported when z ∈ B1. In fact let η be a C∞ function supported in B1 which equals 1 when
hence, by Minkowski's inequality, is bounded on Lp, for all 1 ≤ p ≤ ∞.
see that K∞(x, z) is bounded and rapidly decreasing as z → ∞. Now the operator T∞, that
Thus suppose K(x, z) is supported when z ∈ B1. We make a further assumption that
K(x, z) is supported when x ∈ Bλ, where λ is so small so that B2λ is strictly contained in
z ∈ B1/2. Then writing K(x, z) = K0(x, z) + K∞(x, z) = K(x, z)η(z) + K(x, z)(cid:0)1 − η(z)(cid:1), we
is (12.2) with K∞ in place at K, can be written as T∞(f )(x) =R K∞(x, y)f (xy−1)dy, and
the torus T =(cid:8)x : xj < 1
j = 1, . . . , N(cid:9). (Later we will lift this hypothesis.) In this case
K(x, z) = Xk∈ZN
(2π)N ZT
we can extend K outside of T by periodicity, and expand K(x, z) as Fourier series in the
x-variable,
Hk(z) =
1
e−2πik·xK(x, z)dx .
(12.3)
2 ,
e2πik·xHk(z) ,
Each kernel Hk(z) can be written as αkKk(z) in such a way that the αk are rapidly
decreasing coefficients and the Kk are proper kernels, supported for z ∈ B1, whose semi-
norms KkM are uniformly bounded in k for every M . This can be done via the following
diagonal argument. Integrating by parts in (12.3), we obtain that, for every integer q,
HkM ≤
Cq
kq Kq,M ,
where Kq,M denotes the supremum of the semi-norms implied by (12.1) over β ≤ q. Then,
given q, we inductively choose increasing integers mq such that
For mq ≤ k < mq+1, we set
Cq
mq
q
Kq,q <
1
mq/2
q
.
αk =
Cq
kq Kq,q .
Then αk < k−q/2, showing that the αk are rapidly decreasing, and the kernels Kk = α−1
k Hk
are uniformly bounded in every M -norm (we remark that the M -norms are increasing in M ).
Upon relabeling the series and with a slight abuse of notation, this gives
K(x, z) =
∞Xn=1
an(x)Kn(z) .
(12.4)
Here an(x) = η(x)αke2πik·x, Kn(z) = Kk(z) (for a k that depends on n); also η is a suitable
C∞ cut-off function supported in x ∈ Bλ. With this choice, the an(x) are a rapidly decreasing
sequence of C∞ functions, each supported in Bλ. Clearly now the operator T whose kernel
is given by (12.4) is bounded on Lp, in view of Theorem 8.9.
Now, with the above λ fixed, we will drop the restriction that K(x, z) is supported for
x ∈ Bλ, by a further decomposition which writes RN as a union of balls Bλ(xm) = xm · Bλ,
with these balls chosen so they are "almost"disjoint. For this we need the following lemma.
Lemma 12.3. Given λ > 0, there is a collection {Bλ(xm)} so that
(1) (cid:8)Bλ/2(xm)(cid:9) covers RN ;
93
(2) for every µ > 0, the collection {Bµ(xm)} has the bounded intersection property: there
exists an integer M = Mµ so that no point x ∈ RM is contained in more than M of the
{Bµ(xm)}.
Proof. Fix a constant c1 (which is small with λ, and which will be chosen momentarily) and
choose the centers xm so that the balls {Bc1(xm)} are disjoint, and so that this is a maximal
Bλ/2(xm) does not cover RN , there exists an
family with respect to that property. Now if ∪
m
x0, so that d(x0, xm) ≥ λ/2 for all m. We then claim that Bc1(x0) must be disjoint from
all the Bc1(xm), violating the maximality. In fact if this disjointness fails, then there is an
x′ ∈ Bc1(x0) ∩ Bc1(xm) for some m. This means d(x′, x0) < c1 and d(x′, xm) < c1, which by
the quasi-triangle inequality d(x0, xm) ≤ c(d(x′, x0) + d(x′, xm)) gives a contradiction with
d(x0, xm) ≥ λ/2, when λ/2 ≥ 2cc1. Thus if we choose c1 = λ/4c we get conclusion (1).
To prove the second conclusion we may choose µ ≥ c1. Suppose x′ and x′′ belong to
Bµ(xm). Then d(x′, x′′) ≤ 2cµ, so x′′ ∈ B2cµ(x′), and hence Bµ(xm) ⊂ B2cµ(x′). So if
x′ belongs to M of the Bµ(xm), then Bµ(xm) ⊂ B2cµ(x′) for each of these m. Since then
Bc1(xm) ⊂ B2cµ(x′) and the Bc1(xm) are disjoint, we have
X m (Bc1(xm)) ≤ m (B2cµ(x′)) ,
where the sum is taken over those m. By homogeneity we get M cQ
(2) is established with M = (2c/c1)Q.
1 ≤ (2cµ)Q, and conclusion
We now return to the proof of Theorem 12.1. Let us fix a positive C∞ function ϕ that
equals 1 in Bλ/2 and vanishes outside Bλ. Let ϕm(x) = ϕ(cid:0)x−1
m · x(cid:1), with the xm as given in
Lemma (12.3). Then because of conclusions (1) and (2) of the lemma, 1 ≤P ϕm(x) ≤ M .
Set ψm(x) = ϕm(x)/P ϕm(x). Since {X αϕm} are bounded uniformly in m, the same is true
We let Km(x, z) = ψm(x)K(x, z) and Tm(f )(x) = R Km(x, y−1x)f (y)dy. Then T =
P Tm. Observe that Tm is of the form LxmT ′L−1
is an operator for which we already know its Lp boundedness by Theorem 8.9. Hence
of {ψm}. Notice that the {ψm} give a partition of unity, with ψm supported in Bλ(xm).
xm, where Lx is left-translation by x and T ′
Tm(f )Lp ≤ Apf Lp ,
(12.5)
uniformly in m,
for 1 < p < ∞. Now because T =P Tm, the lemma (with µ = λ) gives us
However by (12.5),
Z T pdx ≤ MλXm Z Tm(f )pdx
Z Tm(f )p dx ≤ Ap
ZBµ(xm)
f pdx.
p
This is because x ∈ Bλ(xm) and y−1x ∈ B1 imply y ∈ Bµ(xm) here with µ = c(λ + 1), in
view of the fact that d(y, xm) ≤ c(d(x, xm)+d(y, x)), and the support properties of Km(x, z).
Again by conclusion (2) of Lemma 12.3,
ZBµ(xm)
Xm
f pdx ≤ MµZ f pdx,
giving the conclusion T (f )Lp ≤ A′
pf Lp and so proving Theorem 12.1.
94
12.2 The algebra of operators
Proceeding to the proof of Theorem 12.2, we first summarize and slightly rephrase the essential
idea of the previous subsection. Whenever K(x, z) is a kernel of the extended class that is
supported for z ≤ 1, then we can write it as
K(x, z) =Xn,m
an,m(x)Kn,m(z) ,
(12.6)
where
(a) the Kn,m are proper kernels, each supported for z in the unit ball, with M -norms that
are uniform in n and m;
(b) the an,m(x) are C∞ functions, each supported in x ∈ Bλ(xm);
(c) for every α, sup
x
X αan,m(x)
is rapidly decreasing as n → ∞, uniformly in m.
To prove Theorem 12.2 we study the composition of two operators T and S given by
T (f )(x) =Z K(x, y−1x)f (y)dy, S(f )(x) =Z L(x, y−1x)f (y)dy
where K and L are both kernels of the extended class. We consider first their essential parts,
that is, we restrict ourselves to K(x, z) and L(x, z) that are supported when z is in the unit
ball B1. Now by (12.6), T =Pn,m Tn,m and S =Pn′,m′ Sn′,m′ where
Tn,m(f )(x) = an,m(x)(f ∗ Kn,m)(x),
Sn′,m′ (f )(x) = bn′,m′(x)(f ∗ Ln′,m′)(x).
and where {Kn,m} and {Ln′,m′} are uniform families of proper kernels, each supported in
z ≤ 1.
We next observe that, for each m, Tn,m ◦ Sn′,m′ = 0 unless m′ belongs to a subset of
boundedly many m′ (depending on m). In fact, by the support properties of an,m, bn′,m′,
and the kernels Kn,m, the product Tn,m · Sn′,m′ is non-vanishing only if the quasi-distance
between Bλ(xm) and Bλ(xm′ ) does not exceed 1. Thus by conclusion (2) of Lemma 12.3, this
only happens for at most Mµ of the m′, where µ = c2(2λ + 1).
From this, and the rapid decrease of an,m and bn′,m′ as n → ∞ and n′ → ∞, we see that
the analysis of the composition can be reduced to the case when both T and S are each one
summand. So we write T = a ◦ U, S = b ◦ V , where U (f ) = f ∗ K and V (f ) = f ∗ L, with
both K and L proper kernels supported for z ≤ 1. Now
T ◦ S = a ◦ U ◦ b ◦ V = a ◦ b ◦ U ◦ V + a ◦ [b, U ] ◦ V.
The term a◦b◦U ◦V is of the right form because it corresponds to the kernel a(x)b(x)(L∗K)(z),
and L ∗ K is a proper kernel by Theorem 9.2. On the other hand,
[b, U ](f )(x) =Z (cid:0)b(x) − b(y)(cid:1)K(y−1x)f (y) dy .
(12.7)
Write b(x) − b(y) = c(x, z), with z = y−1x; thus c(x, z) = b(x) − b(xz−1). Recall that
c(x, z) can be restricted to z ∈ B1, (the support of K), and so can be taken to have compact
z -- support. Also, c(x, z) is jointly C∞, and supx(cid:12)(cid:12)X α
At this stage we invoke the following lemma.
Lemma 12.4. Let K ∈ P0(E) and c ∈ C∞
x ∂β
z c(x, z)(cid:12)(cid:12) < ∞ , uniformly in z.
0 (RN ). Then c K ∈ P0(E).
Proof. We use the Fourier transform characterization of Theorems 4.1 and 4.2:
then m ∈ M∞(E) and
95
if m = bK,
.
cc K(ξ) = m ∗bc (ξ) =ZRN
m(ξ − η)bc(η)dη.
∂γ
∂γ
0 < δ < ab−1 and write
We have bNj(ξ + η) ≤ C(cid:2)bNj(ξ) + bNj(η)(cid:3), and if ξ is the Euclidean norm of ξ, there are
positive real numbers a and b so that ξa ≤ bNj(ξ) ≤ ξb for 1 ≤ j ≤ n and all ξ ≥ 1. Let
ξ m(ξ − η)bc (η)dη +Zη≥ξδ
∂γ(m ∗ bc ) =Zη≤ξδ
If η ≤ ξδ then bNj(ξ − η) ≈ bNj(ξ) for all 1 ≤ j ≤ n and all sufficiently large ξ, and hence
(cid:12)(cid:12)I(cid:12)(cid:12) ≤Zη<ξδ
nYj=1
(1 + bNj(ξ − η))−[[γ j ]](cid:12)(cid:12)bc(η)(cid:12)(cid:12) dη .
(cid:12)(cid:12)II(cid:12)(cid:12) ≤ cM ξ−M ≤ cM
ξ m(ξ − η)bc(η)dη = I + II.
nYj=1(cid:0)1 + bNj(ξ)(cid:1)−[[γ j ]]
nYj=1(cid:0)1 + bNj(ξ)(cid:1)−[[αj ]]
j=1[[αj ]]. As a result, c K ∈ P0(E).
On the other hand, because of the rapid decrease ofbc,
if M ≥Pn
supx,z(cid:12)(cid:12)∂α
z c(x, z)(cid:12)(cid:12) of c and the M -norms of K.
Coming to the operator [b, U ] ◦ V , we see that
x ∂β
We now apply Lemma 12.4 to the kernel c(x, z)K(z) for each fixed x. This shows that
c(x, z)K(z) is an extended kernel. An examination of the argument guarantees that ev-
ery semi-norm (12.1) of this kernel is controlled by only finitely many of the semi-norms
([b, U ] ◦ V )(f )(x) =Z M(x, y−1x)f (y)dy,
where M(x, · ) = L ∗ eK(x, ·), where V (f ) = f ∗ L, and where eK(x, z) = c(x, z)K(z).
Then M(x, z) is an extended kernel, in view of Theorem 9.2; moreover any semi-norm
(12.1) of M is controlled by finitely many semi-norms of b, K and L. Finally, the rapid
decrease of the an,m, as n → ∞ (and the corresponding decrease of the bn′,m′ as n′ → ∞)
then assures that T ◦ S is an operator of the extended class under the assumption to both
K(x, z) and L(x, z) are supported when z ≤ 1.
The case when either one or both K(x, z) and L(x, z) are supported for z away from the
origin is a simpler version of the above argument. In effect, when K is supported when z ≤ 1,
and L is supported when z ≥ c > 0, then we need only use the facts that a convolution of a
Schwartz function with a distribution of compact support is a Schwartz function, or that the
convolution of two Schwartz functions is again a Schwartz function. This allows us to finish
the proof of Theorem 12.2.
13 The role of pseudo-differential operators
In this section we study the interplay of pseudo-differential operators with the operators of
the extended class treated in Section 12. While more general cases can also be treated, for
simplicity we restrict ourselves to the situation where the homogeneous group underlying our
96
Euclidean space RN is a stratified group, and where the matrix E arises from the two-flag
example in Theorem 11.5 where a1 = · · · = an = 1 and bj = 1
j . Thus we take
E =
1
1
1
...
1
1
3
· · ·
2
1 3/2 · · ·
· · ·
1
...
. . .
· · ·
1
· · ·
1
1
...
1
1
n − 1
(n − 1)/2
(n − 1)/3
...
1
1
n
n/2
n/3
...
1
n/(n − 1)
.
In this case of a stratified group we have the double identification of RN = RC1 ⊕· · · ⊕ RCn
with the underlying group and also (via exponential coordinates) with its Lie algebra g. We
also assume that the subspace RCk = hk is the subspace homogeneous of degree k under
the automorphic dilations δr. The fact that g is stratified guarantees that hk+1 = [hk, h1].
Suppose the cardinality of Ck is ck. If x = (x1, . . . , xn) ∈ RN with xk ∈ RCk , write xk =
(xk,1, . . . , xk,ck ). The automorphic dilations on hk are given by δλxk = (λkxk,1, . . . , λkxk,ck ),
the homogeneous dimension of hk is k ck, and for αk ∈ Nck , we have [[αk]] = k αk, where
αk is the standard length of the multi-index.
We take the homogeneous norm on hk to be nk(tk) = tk 1
· denotes the Euclidean norm. Then
k =hPCk
j=1 tk,j 2i1/2k
, where
(13.1)
Nj(x) = x1 + x2
1
2 + · · · + xj−1
1
j−1 + xj
1
j + xj+1
1
j + · · · + xn
1
j
1
j + ξ2
1
j + · · · ξj−1
1
j + ξj
1
j + ξj+1
1
j+1 + · · · + ξn
1
n
bNj(ξ) = ξ1
Notice that Nn = bN0 is a homogeneous norm for the automorphic dilations δλ, and N1 is
We will consider standard pseudo-differential operators of order 0. By this we mean
equal to the Euclidean norm.
(13.2)
(13.3)
operators of the form
f 7→ P (f )(x) =ZRN
a(x, ξ) bf (ξ)e2πix·ξdξ,
defined for Schwartz functions f , where the symbol a(x, ξ) is assumed to be of compact
support in x and satisfy the differential inequalities
(cid:12)(cid:12)∂β
x ∂α
ξ a(x, ξ)(cid:12)(cid:12) ≤ Aα,β(1 + ξ)−α .
In what follows the errors introduced by commutators will be smoothing operators in the
following sense. For 1 < p < ∞, the space Lp
1 consists of all f in Lp so that Xf (taken
in the weak sense) belongs to Lp(G), for each left-invariant vector field X of the generating
sub-space h1 of g. We define a norm on Lp
1 by setting
f Lp
1
= f Lp +
gXj=1
Xj(f )Lp ,
(13.4)
where {X1, . . . , Xg} a basis for the subspace h1. We shall also see below that an f ∈ Lp
belongs to Lp
1 if and only if
f (xh) − f (x)Lp = O(cid:0)Nn(h)(cid:1), as h → 0.
An operator will be said to be smoothing if it maps Lp to Lp
1,
The following theorem will be proved in the next sections.
for all 1 < p < ∞.
(13.5)
97
Theorem 13.1. Suppose P is a standard pseudo-differential operator of order 0. Then P
belongs to the extended class treated in Section 12. Moreover if T is any operator of this class,
then the commutator [T, P ] is a smoothing operator.
13.1 The isotropic extended kernels
For the proof of Theorem 13.1 it is necessary to consider a special sub-class CZ0 ⊂ P0(E), (the
isotropic Calder´on-Zygmund kernels on RN which agree with a Schwartz function at infinity)
and a corresponding sub-class of the "extended" kernels. A distribution K ∈ CZ0 is assumed
to satisfy the stronger inequalities
∂γK(z) ≤ Aαz−N −γ(cid:0)1 + z(cid:1)−M
,
(13.6)
for every γ and M . Moreover these kernels are required to satisfy only the one type of
cancellation property which arises when n = 1. It follows from Theorems 4.1 and 4.2 (for the
case n = 1) that this is equivalent to the condition that the Fourier transform m = bK satisfies
∂γm(ξ) ≤ A′
hence CZ0 ⊂ P0(E). We call elements of CZ0 isotropic proper kernels.
α(1 + ξ)−γ. But then it follows from equation (13.1) that m ∈ M∞(E), and
Together with CZ0 we consider the extended kernels K(x, z) such that for x fixed, K(x, · ) ∈
CZ0. More precisely, we shall assume that these have compact support in x and all x-
derivatives ∂α
x K(x, · ) are isotropic proper kernels uniformly in x. Notice that, having assumed
compact support in x, this condition can be equivalently formulated with the x-derivatives
replaced by left-invariant, or right-invariant, vector fields. We will refer to these kernels as
isotropic extended kernels.
The key fact we need is the following.
Proposition 13.2. Suppose P is a pseudo-differential operator (13.2) of order 0, with sym-
bol a(x, ξ) having compact support in x. Then P can be represented in the following three
alternative forms: for any Schwartz function f :
(a) P (f )(x) = RRN
(b) P (f )(x) = RRN
(c) P (f )(x) = RRN
K(x, x − y)f (y)dy;
KL(x, y−1x)f (y)dy;
KR(x, xy−1)f (y)dy.
Here K, KL, and KR are each isotropic extended kernels.
The realization (a) is well-known; see for instance [Ste70], Sections 4 and 7.4 of Chapter 6.
In this case bK(x, ·) = a(x, ·) (where the Fourier transform acts on the second variable), and
reduces to the fact that a multiplier m(ξ) that satisfies the differential inequalities(cid:12)(cid:12)(cid:12)∂α
ξ m(ξ)(cid:12)(cid:12)(cid:12) ≤
Aα(1 + ξ)−α arises as the Fourier transform of an isotropic proper kernel, and so this is
covered by Theorems 4.1 and 4.2. The other two representations need further analysis.
First we assert that
y−1x = Lx,y(x − y),
(13.7)
where for each (x, y) ∈ G × G, Lx,y is a linear transformation that depends polynomially on
x and y. In fact, by the homogeneity of our group and using the Campbell-Hausdorff formula
we can write for each 1 ≤ k ≤ n and each 1 ≤ j ≤ ck,
(cid:0)y−1x(cid:1)k,j = xk,j − yk,j +Xℓ<k
cℓXj=1
pℓ,j
k (x, y) · (xℓ,j − yℓ,j)
98
with pℓ,j
k polynomials in x and y jointly homogeneous at degree k − ℓ (see for example Section
6.6 in [NRSW12], or Chapter 1 of [FS82]). This gives (13.7) with Lx,y represented by a block-
triangular matrix which in particular implies that det(Lx,y) = 1,
for all x and y. Now let
Mx,y denote the inverse of Lx,y. Then x − y = Mx,y(y−1x) and Mx,y is again polynomially
dependent on x and y.
Define a (non-linear) mapping Φ = Φx : RN → RN by setting Φx(z) = Mx,xz−1(z). Given
x and w,
Φx(cid:0)(x − w)−1x(cid:1) = Mx,xx−1(x−w)(cid:0)(x − w)−1x(cid:1) = x − (x − w) = w
x (w) = (x−w)−1x = Lx,x−w(w).
so Φx maps RN onto RN and the inverse of Φx is given by Φ−1
Thus both Φx and Φ−1
x are polynomial mappings of RN to itself (and hence diffeomorphisms)
that each depend polynomially on x. In our applications of these mappings below, the range
of x is restricted to a compact set, in view of the support hypothesis on the symbol a(x, ξ)
and the resulting fact that K(x, z) has compact-support in x.
We set down the following two simple properties of the mappings Φx:
1) Φx(z) ≈ z, for z ≤ 1;
2) z 1
n . Φx(z) . zn,
for
z ≥ 1.
The first assertion is a consequence of the fact that Φ is a diffeomorphism that fixes the
origin. For the second, we see first that Φx(z) ≤ czn, z ≥ 1 because of the polynomial
dependence (of degree n) of the linear mappings Mx,y. Since the same inequality holds for
the inverse Φ−1
x we get the reverse inequality z 1
n ≤ cΦx(z), again for large z.
Our proof of the second representation (b) in Proposition 13.2 starts with knowing the
first representation and defining KL(x, z) = K(x, Φx(z)). Then, at least formally, we have
Z KL(x, y−1x)f (y)dy =Z K(x, x − y)f (y)dy = P (f )(x).
(13.8)
In order to justify this it will suffice to see that KL(x, z) is, for each x, an isotropic proper
kernel, and (in both variables) an isotropic extended kernel. For this we make the following
observation.
Lemma 13.3. Let K(z) ∈ CZ0. Suppose Φ is a C∞ diffeomorphism of RN fixing the origin,
and assume that Φ(z) & zδ, for some δ > 0 and all z ≥ 1. Then K ◦ Φ ∈ CZ0 where
the distribution K ◦ Φ is defined by(cid:10)K ◦ Φ, ϕi = hK, ϕ(Φ−1)JΦ−1(cid:11) for every test function ϕ,
where JΦ is the Jacobian determinant of Φ.
Proof. Away from the origin K ◦ Φ is the C∞ function K(Φ(z)). Since Φ(0) = 0 it follows
that Φ(z) & cz for small z, and hence K ◦ Φ satisfies (13.6) on the unit ball because K
does. Moreover the assumption Φ(z) ≥ czδ for z ≥ 1, implies the rapid decrease at infinity
of K(Φ(z)), since the same holds for K(z). It remains to verify the cancellation conditions
for K ◦ Φ, i.e. (cid:12)(cid:12)(cid:10)K ◦ Φ, ϕ(R · )(cid:11)(cid:12)(cid:12) ≤ C for all R ≥ 1 wherever ϕ is a normalized C∞ function
supported in the unit ball. However(cid:12)(cid:12)(cid:10)K◦Φ, ϕ(R · )(cid:11)(cid:12)(cid:12) =(cid:10)K, ΨR(cid:11), where it is easy to verify that
for each R, ΨR(x) is supported in the ball x ≤ cR−1, and satisfies ∂α
x ΨR(x) ≤ cαRα, with
bounds c and cα independent of R. Thus we may think of ΨR(x) as of the form c1Ψ(c2Rx),
for some c1 and c2 independent of R, verifying that K ◦ Φ satisfies the requisite cancellation
condition. Thus the proof of Lemma 13.3 is complete.
Hence we have proved the representation (b) of Proposition 13.2. The representation (c)
is proved in the same way.
13.2 Proof of Theorem 13.1
99
Turning to the proof of Theorem 13.1, we see by the identity (b) in Proposition 13.2 that
the pseudo-differential operator P belongs to the extended class of operators, establishing the
first assertion of Theorem 13.1. Now let T be any operator in the extended class,
T (f )(x) =Z L(x, y−1x)f (y)dy,
with L(x, z) a kernel of the extended class. We examine the commutator [T, P ]. Let B
be a ball containing the x -- support of the symbol a(x, ξ) of P , and hence containing the
x-supports of K(x, z), KL(x, z) and KR(x, z). We now first decompose T as T0 + T∞ via
L(x, z) = L0(x, z) + L∞(x, z) where L0(x, z) = η(x)L(x, z), L∞(x, z) =(cid:0)1 − η(x)(cid:1)L∞(x, z)
with η ∈ C∞ of compact support which = 1 on the double B∗ at the ball B. Thus K0(x, z)
has compact x -- support, while K∞(x, z) is supported on the complement of B∗. (Note that
this is different from the decomposition T = T0 + T∞ in Section 12.1).
Now by reasoning as in Section 12.1, we can write
T0(f )(x) =Xn
Tn(f )(x) =Xn
an(x)(f ∗ Hn)(x)
(13.9)
where an(x) are C∞ functions, supported in a common compact set, for which supx ∂α
x an(x)
is rapidly decreasing in n, for each fixed α. The kernels Hn are proper kernels and their semi-
norms are each uniformly bounded in n.
For the pseudo-differential operator P we shall use the "right -- invariant" representation
(c) in Proposition 13.2. This gives by the reasoning establishing (13.9), that
P (f )(x) =Xn′
Pn′ (f )(x) =Xn′
bn′(x)(Kn′ ∗ f )(x) .
(13.10)
Here the bn′ (x) are C∞ function, supported in a compact set (in view of the support property
x bn′(x) is rapidly decreasing in n′ for each fixed α, and the
of KR(x, z)), for which supx ∂α
Kn′ are isotropic proper kernels and their semi-norms are each uniformly bounded in n′. The
rapid decrease in n and n′ pointed out above reduces the study of the commutator [T0, P ] to
that where only a single term of each of the sums, (13.9) and (13.10) appears.
Thus we look at the commutator [a ◦ U, b ◦ V ], where U (f ) = f ∗ H and V (f ) = K ∗ f .
This commutator is handled by the following
Lemma 13.4.
(1) U and V commute,
(2) [b, U ] is a smoothing operator,
(3) [a, V ] is a smoothing operator,
and the bounds of the mappings [b, U ] and [a, V ] depend only on finitely many of the semi-
norms controlling a, b, H, and K.
Suppose first that the lemma is proved. There aU ◦ bV = abU ◦ V + a[b, U ]V . However
U ◦ V = V ◦ U , by conclusion (1). Thus
[a ◦ U, b ◦ V ] = a ◦ [U, b ◦ V ] + [a, b ◦ V ] ◦ U
= a ◦ [U, b] ◦ V + ab ◦ [U, V ] + b ◦ [a, V ] ◦ U + [a, b] ◦ V ◦ U
= a ◦ [U, b] ◦ V + b ◦ [a, V ] ◦ U .
As well as multiplication by a and b, both U and V map Lp to itself, since they are
convolution operators with proper kernels. Thus by conclusions (2) and (3) of the lemma,
[a ◦ U, b ◦ V ] maps Lp to Lp
1, and we have achieved our desired result for [T0, P ].
100
Proof of Lemma 13.4 . The first conclusion (which is key) follows immediately because U
is left-invariant and V is right-invariant. Let us now consider [b, U ], and begin with the
representation given for [b, U ] in (12.7) (here H plays the role of K). We have that [b, U ] is
an operator of the extended class, represented by the kernel eH(x, z) = c(x, z)H(x, z), where
c(x, z) is the C∞ function given by b(x) − b(xz−1). Notice however that c(x, 0) = 0, and thus
where ck,j are again C∞ functions. Hence
c(x, z) =
ckXj=1
nXk=1
eH(x, z) =Xk,j
zk,j · ck,j (x, z),
zk,j · eHk,j(x, z)
(13.11)
where by Lemma 12.4 we know that each eHk,j (x, z) is an extended kernel. Now if X is a
left-invariant vector field, then
X[b, U ](f )(x) = X(cid:18)Z eH(x, y−1x)f (y)dy(cid:19) =Z H′(x, y−1x)f (y)dy ,
where by (13.11) we have that
H′(x, z) =
nXj=1(cid:16)Xx(cid:0)zj · eHj(x, z)(cid:1) + Xz(cid:0)zj · eHj(x, z)(cid:1)(cid:17) .
At this stage we need the following observation.
Lemma 13.5. Suppose H is a proper kernel and X is a left-invariant vector field of degree 1
(that is, in the subspace h1 of the Lie algebra). Then for each (k, j), 1 ≤ k ≤ n, 1 ≤ j ≤ ck,
This can be proved by using the Fourier transform characterization of proper kernels. First
X(cid:0)zk,jH(z)(cid:1) is a proper kernel.
nXℓ=2
write X as
∂z1,r
X =
∂
+
qℓ,s(z)
∂
∂zℓ,s
=
∂
∂z1,r
+
cℓXs=1
Qℓ(z) · ∇zℓ
nXℓ=2
· · · , zℓ−1) are homogeneous polynomials of degree ℓ − 1 (in the
where Qℓ(z) = Qℓ(z1,
homogeneity of the automorphic dilations). Consider first the case k = 1. The Fourier
transform of X(cid:0)z1,jH(z)(cid:1) is then a linear combination of m = bH and ξℓ · Qℓ(2πi∂ξ) ∂ξ1,j m(ξ).
However, it is clear from (13.1) that bNj(ξ)k ≥ ξk if j ≤ k. The the effect of the factor ξℓ,
which has size ξℓ, is counterbalanced by the effect of Qℓ(2πi∂ξ)∂ξk,j on m(ξ), in view of the
estimates for m. The situation when k > 1 (and hence ∂ξ1,j is replaced by ∂ξk,j ) leads to an
even better estimate, which then establishes Lemma 13.5.
Using the lemma we see therefore that Xx[b, U ] is an operator with an extended kernel,
1, establishing conclusion (2)
and hence by Theorem 12.1 we have that [b, U ] maps Lp to Lp
of Lemma 13.4.
Finally, for conclusion (3) we may appeal to the standard theory of pseudo-differential
operators that guarantees that the commutator [a, V ] is an operator of "order 1", and in
particular maps Lp to Lp
1; or we may argue in the same spirit as in the case of conclusion (2),
but here the details are simpler. In any case, Lemma 7.4 is now established and this proves
that [T0, P ] is bounded from Lp to Lp
1.
101
To complete the proof of Theorem 13.1 it remains to deal with [T∞, P ] = T∞P − P T∞.
We will see that the terms T∞P and P T∞ are separately smoothing operators (in fact,
infinitely smoothing) because of disjointness of relevant supports. Consider first T∞P . Then
T∞(P f ) = T∞(F ), where F is supported in the ball B (where the symbol a(x, ξ) is supported).
However
T∞(F )(x) =Z K∞(x, y−1x)F (y)dy
∞(x, y−1x)F (y)dy with K ′
and K∞(x, z) is supported for x outside B∗. Hence in the integral above we have that
y−1x ≥ c > 0, for an appropriate c. Moreover if X is any left -- invariant operator then
∞(x, z) = XxK∞(x, z) + XzK∞(x, z).
∞(x, z) is bounded and has rapid decay in z, (we can restrict ourselves here to
1. In fact the same argument shows that
X(T∞F )(x)) is given byR K′
Essentially the same argument works for P T∞. Here (P T∞)(f ) =R K L(x, y−1x)f (y)dy
Since now K′
z ≥ c > 0), this ensures that T∞P maps Lp to Lp
X α(T∞P ) maps Lp to itself for any monomial X α of left -- invariant differential operators.
by Proposition 13.2, with F supported in the complement of B∗ and K(x, z) supported where
x ∈ B. Then again y−1x ≥ c > 0, and we may proceed as before. Altogether then, the
proof of Theorem 13.1 is completed.
13.3 The space Lp
1
We conclude by proving the equivalence of the condition that f ∈ Lp
condition (13.5).
1 with the Lipschitz -- type
Proposition 13.6. Let 1 < p < ∞. Then
f ∈ Lp
1
⇐⇒
f (xh) − f (x)Lp = O(cid:0)Nn(h)(cid:1), as h → 0.
Proof. Let X be a left-invariant vector field in the subspace h1 of the Lie algebra, and set
ht = exp(tX). Then ht = ht ≈ t, t ∈ R. So
f (xh) − f (x)Lp = O(cid:0)Nn(h)(cid:1), as h → 0 =⇒
tn(cid:0)f (xhtn ) −
f (x)(cid:1) converges weakly in Lp and hence in the sense of distributions. Thus in that sense,
t(cid:0)f (xht) − f (x)(cid:1)Lp = O(1)
Thus by the weak compactness of Lp, there is a sequence {tn}, tn → 0 so that 1
X(f ) ∈ Lp. Since this holds for all X ∈ h1, we have that f ∈ Lp
1.
as t → 0.
1, and h = exp(sX) where X ∈ h1. Then Nn(h) = h . s
Conversely, suppose f ∈ Lp
1
and
f (xh) − f (x) =Z s
0
d
dt
f(cid:0)x exp(tX)(cid:1) dt =Z s
0
Xf(cid:0)x exp(tX)(cid:1) dt .
Hence, f (xh) − f (x)Lp ≤ s X(f )Lp = O(cid:0)Nn(h)(cid:1). To extend this to any element h we
use the following assertion.
Lemma 13.7. Suppose G is a stratified group. Then there is an integer M , so that every
h ∈ G, can be written as h = h1h2 · · · hM , where each hj ∈ exp h1 and hj ≤ cNn(h) for
all 1 ≤ j ≤ M .
Assuming for a moment this to be true, the desired estimate for f (xh) − f (x) then follows
for the special case when h ∈ exp h1 by writing
f (xh) − f (x) =
MXj=1
f (xh1 · · · hj) − f (xh1 · · · hj−1).
102
Proof of Lemma 13.7. Let V = exp(B1), where B1 is the unit ball in h1, Then V is an analytic
submanifold of G which generates G. By Proposition 1.1 of [RS88], there exists an integer
m such that V m contains an open set A. Since V = V −1, V 2m contains a neighborhood
of the identity. This proves the statement for h < δ for some δ > 0 with M = 2m. By
homogeneity, the same holds for every h ∈ G.
14 Appendix I: Properties of cones Γ(A)
In this appendix we study properties of the cone Γ(E) defined in (5.1). Our results provide
motivation for the basic hypothesis (2.10) imposed on the matrix E.
14.1 Optimal inequalities and the basic hypothesis
Definition 14.1. Let A =(cid:8)a(j, k) : 1 ≤ j, k ≤ n(cid:9) be an n × n matrix of strictly positive real
numbers. Then
Γo(A) =nt = (t1, . . . , tn) ∈ Rn : a(j, k)tk < tj < 0o
Γ(A) =nt = (t1, . . . , tn) ∈ Rn : a(j, k)tk ≤ tj < 0o
are (possibly empty) convex polyhedral cones contained in the negative orthant of Rn. The
dimension of either cone is the maximal number of linearly independent vectors it contains.
Equivalently, this is the dimension of the smallest subspace of Rn containing the cone.
Γ(A) is defined by a collection of linear inequalities involving pairs of coordinates. It may
happen that different matrices (i.e. different sets of such inequalities) give rise to the same
cone. The following Lemma shows that there is a unique optimal set of inequalities, and that
the corresponding matrix satisfies the basic hypothesis (2.10).
Lemma 14.2. Let A = {a(j, k)} be an n × n matrix of strictly positive real numbers, and
suppose that the corresponding cone Γ(A) is not empty. Then there exists a unique n × n
matrix eA =(cid:8)a(j, k)(cid:9) with the following properties:
(a) The coefficients of eA satisfy the inequalities (2.10) of the basic hypothesis:
1 = a(j, j)
0 < a(j, k) ≤ a(j, l)a(l, k)
for 1 ≤ j ≤ n,
for 1 ≤ j, k, l ≤ n.
(b) Γ(A) = Γ(eA).
Moreover, if eA satisfies (a) and (b), then a(j, k) ≤ a(j, k) for 1 ≤ j, k ≤ n.
Proof. For any t = (t1, . . . , tn) ∈ Γ(A), tj < 0 and 0 < tjt−1
Γ(A) 6= ∅ define
tj
tk
Then a(j, j) = 1 and for any j, k, l ∈ {1, . . . , n}
0 < a(j, k) = sup
t∈Γ(A)
a(j, k) = sup
t∈Γ(A)
tj
tk
This establishes (a).
= sup
t∈Γ(A)h tj
tl
tl
tki ≤ sup
t∈Γ(A)
k ≤ a(j, k) for 1 ≤ j, k ≤ n. If
≤ a(j, k).
(14.1)
tj
tl
sup
t∈Γ(A)
tl
tk
= a(j, l) a(l, k).
a(j, k)vk ≤ vj , and so vj v−1
Let v = (v1, . . . , vn) ∈ Γ(eA). Since vk < 0 and a(j, k) ≤ a(j, k), it follows that a(j, k)vk ≤
k ≤ a(j, k). Thus v ∈ Γ(A) and it follows that Γ(eA) ⊂ Γ(A).
103
Conversely, if v = (v1, . . . , vn) ∈ Γ(A), then vjv−1
k ≤ supt∈Γ(A) tjt−1
k = a(j, k) and so
a(j, k)vk ≤ vj . It follows that Γ(A) ⊂ Γ(eA), and this establishes (b).
Next we show that the matrix eA is unique. Let C =(cid:8)c(j, k) : 1 ≤ j, k ≤ n(cid:9) be any matrix
satisfying (a) and (b). Since Γ(C) = Γ(A), if t ∈ Γ(C) then c(j, k)tk ≤ tj, and so
a(j, k) = sup
t∈Γ(A)
tjt−1
k = sup
t∈Γ(C)
tjt−1
k ≤ c(j, k).
all 1 ≤ j, k, l ≤ n. Letting l = k, we have a(j, k)(−c(k, k)) ≤ −c(j, k), and since c(k, k) = 1
To establish the reverse inequality, consider the vectors wl =(cid:0)− c(1, l), . . . , −c(n, l)(cid:1). Since C
satisfies (a), we have c(j, l) ≤ c(j, k)c(k, l), or equivalently c(j, k)(cid:0) − c(k, l)(cid:1) ≤(cid:0)− c(j, l)(cid:1). This
shows that wl ∈ Γ(C) = Γ(eA) for 1 ≤ l ≤ n. It follows that a(j, k)(−c(k, l)) ≤ −c(j, l) for
we have c(j, k) ≤ a(j, k). Thus c(j, k) = a(j, k), which shows that eA is unique. We have
already observed that a(j, k) ≤ a(j, k), and since the matrix eA is unique, this completes the
a(j, k) =(min(cid:8)a(j, i1)a(i1, i2) · · · a(ir, k) : r ≥ 0, {i1, . . . , ir} ⊂ {1, . . . , n}(cid:9) if j 6= k
It can be shown that, in the hypotheses of Lemma 14.2,
if j = k
proof.
1
.
14.2 Partial matrices
In Lemma 14.2 it is assumed that the cone Γ(A) is defined by the full number n(n − 1)
inequalities a(j, k)tk ≤ tj where 1 ≤ j 6= k ≤ n. We also need to consider cases in which we
begin with fewer inequalities. We formulate this as follows. Let B = {b(j, k)} be an n × n
matrix, where this time each b(j, k) is either a positive real number or the symbol ∞. As
before we put
Γ(B) =nt = (t1, . . . , tn) ∈ Rn : b(j, k)tk ≤ tj < 0o,
but now if b(j, k) = ∞, the inequality b(j, k)tk ≤ tj < 0 puts no constraint on the relation
between the two negative real numbers tk and tj. It is still the case that Γ(B) is a polyhedral
cone in the negative orthant of Rn. We call B a partial matrix. We say that a partial
matrix B = {b(j, k)} is connected if for any two indices j, k ∈ {1, . . . , n} there are indices
i1, . . . , ia ∈ {1, . . . , n} (with a possibly dependent on j and k) so that
b(j, i1)b(i1, i2) · · · b(il, il+1) · · · b(ia−1, ia)b(ia, k) < +∞.
It then follows that
b(j, i1)b(i1, i2) · · · b(ij, ij+1) · · · b(ia−1, ia)b(ia, k)tk ≤ tj < 0
so if B is connected and t ∈ Γ(B) it follows that
sup
t∈Γ(B)
tj
tk
≤ b(j, i1)b(i1, i2) · · · b(ij, ij+1) · · · b(ia−1, ia)b(ia, k) < +∞.
We then have the following analogue of Lemma 14.2.
Lemma 14.3. Let B = {b(j, k)} be an n × n connected, partial matrix, and suppose that the
corresponding cone Γ(B) is not empty. Then there exists a unique n × n matrix eB =(cid:8)b(j, k)(cid:9)
with the following properties:
104
(a) The coefficients of eB are all finite positive real numbers, and
for 1 ≤ j ≤ n,
1 = b(j, j)
0 < b(j, k) ≤ b(j, l)b(l, k)
for 1 ≤ j, k, l ≤ n.
(b) Γ(B) = Γ(eB).
Moreover, if eB satisfies (a) and (b), then b(j, k) ≤ b(j, k) for 1 ≤ j, k ≤ n.
Proof. There is only a minor change from the proof of Lemma 14.2. Since b(j, k) might
equal ∞ we cannot directly conclude that supt∈Γ(B) tj t−1
is finite by observing that it is
k
bounded by b(j, k) as in equation (14.1). However, since B is connected, it is still the case
that supt∈Γ(B) tj t−1
k < ∞, and we can defineeb(j, k) to the this supremum. The proof of (a)
then proceeds as before, and we set Γ(eB) = nt = (t1, . . . , tn) ∈ Rn :eb(j, k)tk ≤ tj < 0o.
Note that if b(j, k) < ∞ then eb(j, k) = supt∈Γ(B)
Γ(eB) = Γ(B) and the uniqueness of B also follow as before.
≤ b(j, k) as before. The proofs that
tj
tk
14.3 Projections
Let 1 < p < n and let πn,p : Rn → Rp be the projection onto the first p coordinates:
πn,p(x1, . . . , xn) = (x1, . . . , xp). Let An = {a(j, k) : 1 ≤ j, k ≤ n} be an n × n matrix with
strictly positive entries which satisfies the basic hypotheses
a(j, j) = 1
a(j, l) ≤ a(j, k)a(k, l)
1 ≤ j ≤ n,
1 ≤ j, k, l ≤ n.
Let Ap = {a(j, k) : 1 ≤ j, k ≤ p} be the corresponding p × p matrix (which of course also
satisfies the basic hypotheses). We than have two cones:
Γ(An) =nt = (t1, . . . , tn) ∈ Rn : a(j, k)tk ≤ tj < 0, 1 ≤ j, k ≤ no,
Γ(Ap) =nu = (u1, . . . , up) ∈ Rp : a(j, k)uk ≤ uj < 0, ≤ j, k ≤ po.
Lemma 14.4. The projection πn,p maps the cone Γ(An) onto the cone Γ(Ap).
Proof. If a(j, k)tk ≤ tj < 0 for 1 ≤ j, k ≤ n, then a(j, k)tk ≤ tj < 0 for 1 ≤ j, k ≤ p, and so it
is clear that π maps Γ(An) into Γ(Ap). The main point of the Lemma is that the mapping
is onto. But since
πn,p = πp+1,p ◦ πp+2,p+1 ◦ · · · ◦ πn,n−1
it clearly suffices to show that the mapping πn,n−1 : Γ(An) → Γ(An−1) is onto.
Let (u1, . . . , un−1) ∈ Γ(An−1). We must show that there exists un < 0 so that
1 ≤ j ≤ n − 1 =⇒ a(j, n)un ≤ uj
and a(n, j)uj ≤ un
If we let Ej be the closed intervalha(n, j)uj, uj
For any j, k ∈ {1, . . . , n − 1} we have
a(j,n)i, then we must show thatTn−1
j=1 Ej 6= ∅.
a(n, j)uj ≤
a(k, j)
a(k, n)
uj
≤
uk
a(k, n)
since a(k, j) ≤ a(k, n)a(n, j) and uj < 0,
since (u1, . . . , un−1) ∈ Γ(An−1) and hence a(k, j)uj ≤ uk.
It follows that for any index 1 ≤ ℓ ≤ n − 1 we have
a(n, ℓ)uℓ ≤ supna(n, j)uj : 1 ≤ j ≤ n − 1o ≤ infn uk
and hence we have produce a closed, non-empty interval inTn−1
a(k, n)
(cid:20)supna(n, j)uj : 1 ≤ j ≤ n − 1o, infn uk
a(k, n)
105
uℓ
a(ℓ, n)
: 1 ≤ k ≤ n − 1o ≤
j=1 Ej :
: 1 ≤ k ≤ n − 1o(cid:21) ⊂
Ej .
n−1\j=1
14.4 The dimension of Γ(E)
Proposition 14.5. Let E be an n × n matrix satisfying the basic hypothesis (2.10). Then
the cone Γ(E) is not empty. In particular, the 2n vectors
vl =(cid:0) − e(1, l), . . . , −e(n, l)(cid:1),
wl =(cid:0) − e(l, 1)−1, . . . , −e(l, n)−1(cid:1), 1 ≤ l ≤ n
1 ≤ l ≤ n
all belong to Γ(E).
Proof. Fix l ∈ {1, . . . , n} and let vj = −e(j, l). Then it follows from (2.10) that e(j, l) ≤
e(j, k)e(k, l) and so e(j, k)vk = −e(j, k)e(k, l) ≤ −e(j, l) = vj . This shows that vl ∈ Γ(E).
Similarly, let wj = −e(l, j)−1. Then since e(l, k) ≤ e(l, j)e(j, k), e(j, k)wk = − e(j,k)
e(l,j) ,
and so wl ∈ Γ(E). This completes the proof.
e(l,k) ≤ − 1
Since the vectors vl in Proposition 14.5 are the negatives of the columns of E, we have
the following corollary.
Corollary 14.6. Let E be an n × n matrix satisfying the basic hypothesis (2.10). Then the
dimension of Γ(E) is greater than or equal to the rank of E.
Lemma 14.7. Let E be an n × n matrix satisfying the basic hypotheses (2.10). The following
are equivalent:
(i) the interior of Γ(E) is non-empty;
(ii) the dimension of Γ(E) is n;
(iii) 1 < e(j, k)e(k, j) for every pair 1 ≤ j 6= k ≤ n.
Proof. Obviously (i) implies (ii). Conversely, since Γ(E) is convex, if it contains a basis of Rn
it also containes the convex hull of this basis together with 0, which is open.
In order to prove that (iii) implies (i), we argue by induction on n. For n = 1, Γ(E) =
(−∞, 0) which has dimension 1. For n = 2, det(E) = 1 − e(1, 2)e(2, 1) < 0, so the rank of E
is 2, and hence by Corollary 14.6 the dimension of Γ(E) is 2 and its interior is a nontrivial
angle.
Thus assume that the Lemma is true for all (n−1)×(n−1) matrices E′ satisfying equation
(2.10). Let E be an n × n matrix satisfying equation (2.10) such that 1 < e(j, k)e(k, j) for
every pair 1 ≤ j 6= k ≤ n. Let E′ be the sub-matrix consisting of the first n − 1 rows and
columns. Then clearly E′ also satisfies equation (2.10) and the hypothesis of the Lemma, so
by the induction hypothesis, the dimension of Γ(E′) is n − 1, and Γ(E′) contains a non-empty
open set U ⊂ Rn−1.
106
Let t′ = (t1, . . . , tn−1) ∈ U . Then the components of t′ satisfy the strict inequalities
it follows that
e(k, j)tj < tk for all j 6= k. Since tj < 0 and e(k, j) ≤ e(k, n)e(n, j),
e(k, n)e(n, j)tj ≤ e(k, j)tj < tk and so e(n, j)tj < e(k, n)−1tk for all j, k. Therefore
maxne(n, j)tj : 1 ≤ j ≤ n − 1o < minne(k, n)−1tk : 1 ≤ k ≤ n − 1o.
(14.2)
But if t′ ∈ Γ(E′) then t = (t′, tn) = (t1, . . . , tn−1, tn) in Γ(E) if and only if for all 1 ≤ j, k ≤
n − 1,
maxne(n, j)tj : 1 ≤ j ≤ n − 1o ≤ tn ≤ minne(k, n)−1tk : 1 ≤ k ≤ n − 1o.
It follows from (14.2) that the set of tn for which this is true is non empty and contains
the open interval(cid:0)m(t′), M (t′)(cid:1) where
Then
m(t′) = maxne(n, j)tj : 1 ≤ j ≤ n − 1o,
M (t′) = minne(k, n)−1tk : 1 ≤ k ≤ n − 1o.
n(t′, tn) : t′ ∈ U and m(t′) < tn < M (t′)o
is an open subset of Γ(E).
Conversely, if 1 < e(j, k)e(k, j) for every pair 1 ≤ j 6= k ≤ n, it follows from Lemma
14.7 in Appendix I that Γ(E) has dimension n and in particular has non-empty interior. On
the other hand, if Γo(E) is non-empty, then the projection of Γ(E) onto the two dimensional
subspace spanned by xj and xk is open and non-empty. Since this projection is contained in
{(xj, xk) ∈ R2 : e(j, k)xk < xj < e(k, j)−1xk}, it follows that 1 < e(j, k)e(k, j).
Proposition 14.8. Let E be an n × n matrix satisfying the basic hypothesis (2.10), and
suppose that 1 < e(j, k)e(k, j) for all j 6= k. Then for any indices k1, . . . , ks it follows that
1 < e(k1, k2)e(k2, k3) · · · e(ks−1, ks)e(ks, k1)
provided that the indices {k1, . . . , ks} are not all the same.
Proof. The basic hypothesis implies that we always have
1 ≤ e(k1, k2)e(k2, k3) · · · e(ks−1, ks)e(ks, k1).
We argue by induction on s. If s = 2, the hypothesis of the Proposition implies that we cannot
have 1 = e(k1, k2)e(k2, k1) unless k1 = k2. Now suppose the statement is true for any set of
indices of length s − 1, and let k1, . . . , ks be indices which are not all equal. Consider the
product e(k1, k2)e(k2, k3) · · · e(ks−1, ks)e(ks, k1). Without loss of generality we can assume
k1 6= k2 for otherwise e(k1, k2) = 1 and the result follows by induction. But then
1 < e(k1, k2)e(k2, k1) ≤ e(k1, k2)e(k2, k3)e(k3, k1) ≤ · · ·
· · · ≤ e(k1, k2)e(k2, k3) · · · e(ks−1, ks)e(ks, k1)
which completes the proof.
107
14.5 The reduced matrix E♭
In this section we show that if the dimension of Γ(E) is s < n, there is an s × s sub-matrix
E♭ of E satisfying the hypothesis of Lemma 14.7, so that Γ(E♭) ⊂ Rs has dimension s, and
that Γ(E) is the graph of a linear mapping from Rs to Rn.
Let E = {e(j, k)} be an n × n matrix satisfying the basic hypothesis (2.10).
If j, k ∈
{1, . . . , n} write j ∼ k if and only if e(j, k)e(k, j) = 1.
It is immediate that j ∼ j and
j ∼ k ⇐⇒ k ∼ j. Suppose that j ∼ k and k ∼ l. Then since e(j, k)e(k, j) = 1 and
e(k, l)e(l, k) = 1,
1 ≤ e(j, l)e(l, j) ≤(cid:0)e(j, k)e(k, l)(cid:1)(cid:0)e(l, k)e(k, j)(cid:1) =(cid:0)e(j, k)e(k, j)(cid:1)(cid:0)e(k, l)e(l, k)(cid:1) = 1
so that j ∼ l. Thus '∼' is an equivalence relation.
There are other equivalent formulations of this relation, based on the following remarks.
(i) If e(j, k)e(k, j) = 1, the basic hypotheses imply that, for every ℓ,
1
e(k, j)
≤
e(j, ℓ)
e(k, ℓ)
≤ e(j, k),
1
e(j, k)
≤
e(ℓ, j)
e(ℓ, k)
≤ e(k, j).
Hence the above inequalities are all equalities, i.e., the j-th and k-th row of E are
proportional. Conversely, if the j-th and the k-th rows Ej, Ek are proportional, then
Ej = e(j, k)Ek and Ek = e(k, j)Ej, hence e(j, k)e(k, j) = 1.
In the same way, one proves that j ∼ k if and only if the j-th and k-th column are
proportional.
(ii) If t = (t1, . . . , tn) ∈ Γ(E) and if j ∼ k, then e(j, k)tk ≤ tj = e(j, k)e(k, j)tj ≤ e(j, k)tk
so for all t ∈ Γ(E),
t = (t1, . . . , tn) ∈ Γ(E), j ∼ k
=⇒
tj = e(j, k)tk.
(14.3)
Conversely, if Γ(E) is contained in the hyperplane tj = λtk, this condition is satisfied in
particular by the vectors vk, wk in (14.5). Hence e(j, k) = λ = e(k, j)−1 and therefore
j ∼ k.
Let {k1, . . . , ks} be representatives of the distinct equivalence classes in {1, . . . , n} and let
E♭ be the s × s matrix whose entries are {e(ka, kb) : 1 ≤ a, b ≤ s}. Then E♭ satisfies the
same basic hypothesis, and if a 6= b we have e(ka, kb)e(kb, ka) > 1. The matrix E♭ is called a
reduced matrix for E. It is not uniquely determined by E since it depends on the choice of
representatives of the s distinct equivalence classes. The number s is called the reduced rank
of E.
Let Π : Rn → Rs be the projection given by Π(t) = Π(t1, . . . , tn) = (tk1 , . . . , tks ).
If
t ∈ Γ(E) then clearly Π(t) ∈ Γ(E♭), so Π : Γ(E) → Γ(E♭). On the other hand, if t♭ =
(tk1 , . . . , tks) ∈ Γ(E♭), set σ(t♭) = (t1, . . . , tn) where tj = e(j, ka)tka if j ∼ ka. We claim
σ(t♭) ∈ Γ(E). To see this, let 1 ≤ j, k ≤ n and suppose j ∼ ka and k ∼ kb. Then since
e(ka, kb)tkb ≤ tka,
e(j, k)tk = e(j, k)e(k, kb)tkb = e(j, k)e(k, kb)e(ka, kb)−1e(ka, kb)tkb
≤ e(j, k)e(k, kb)e(ka, kb)−1tka = e(j, k)e(k, kb)e(ka, kb)−1e(j, ka)−1e(j, ka)tka
= e(j, k)e(k, kb)e(ka, kb)−1e(j, ka)−1tj ≤ tj
since e(j, ka)e(ka, kb) = e(ka,kb)
≤ e(j, k)e(k, kb). It follows that e(j, k)tk ≤ tj
as required, so σ(t♭) ∈ Γ(E). Thus if t = (t1, . . . , tn) ∈ Γ(E), the coordinates tj for j /∈
{k1, . . . , ks} are uniquely determined by Π(t). It follows that the mapping Π : Γ(E) → Γ(E♭) is
one-to-one and onto, with the inverse mapping given by σ. We have established the following.
e(ka ,j) ≤ e(ka,j)e(j,kb)
e(ka ,j)
108
Lemma 14.9. Let E be an n × n matrix satisfying the basic hypothesis (2.10). Let j ∼ k if
and only if 1 = e(j, k)e(k, j). This is an equivalence relation, and suppose there are s distinct
equivalence classes. Let {k1, . . . , ks} be representatives of the distinct equivalence classes, and
let E♭ be the s × s reduced matrix whose entries are {e(ka, kb) : 1 ≤ a, b ≤ s}.
a) The reduced matrix E♭ satisfies the basic hypothesis (2.10).
b) 1 < e(ka, kb)e(kb, ka) for all 1 ≤ a 6= b ≤ s.
c) The dimension of Γ(E) is s.
d) Γ(E♭) ⊂ Rs has dimension s, and thus has non-empty interior.
e) Let Π : Rn → Rs be given by Π(t) = Π(t1, . . . , tn) = (tk1 , . . . , tks ). Then Π : Γ(E) → Γ(E♭)
is one-to-one and onto.
f) Let V =nt = (t1, . . . , tn) ∈ Rn : tj = e(j, ka)tka if j ∼ kao. Then V is an s-dimensional
subspace of Rn, and Γ(E) ⊂ V .
15 Appendix II: Estimates for homogeneous norms
Consider a family of dilations on Rm given by λ · x = (λd1x1, . . . , λdmxm) with homogeneous
dimension Q = d1 + · · · + dm. Let N : Rm → [0, ∞) be a smooth homogeneous norm relative
to this family so that N (x) ≈ x11/d1 + · · · + xm1/dm. For any γ = (γ1, . . . , γm) ∈ Nm let
[[γ]] = d1γ1 + · · · + dmγm.
Proposition 15.1. Let ρ(x) be homogeneous of degree p. For γ ∈ Nm there is a constant
Cγ > 0 so that
x ρ(x) ≤ Cγ N (x)p−[[γ]].
∂γ
Proof. Let γ = (γ1, . . . , γm) ∈ Nm. Differentiating the equation λp ρ(x) = ρ(λ · x) we obtain
λp ∂γρ(x) = λ
j=1 djγj ∂γρ(λ · x) = λ[[γ]]∂γρ(λ · x).
Pm
It follows that ∂γ
positive on the compact set {x : N (x) = 1}, it follows that ∂γ
x ρ(x)1/(p−[[γ]]) is homogeneous of degree 1. Since N is continuous and strictly
x ρ(x) ≤ Cγ N (x)p−[[γ]].
Proposition 15.2. There is a constant c so that(cid:12)(cid:12){x ∈ Rm : N (x) ≤ A}(cid:12)(cid:12) = c AQ.
Proof. Let Bm(1) = {x ∈ Rm : N (x) < 1}. Then N (x) ≤ A if and only if N (A−1 · x) =
A−1N (x) ≤ 1, and this is true if and only if A−1 · x ∈ BN (1). If TA(x) = A · x, this means
that x ∈ TA(B(1)). Since det(TA) = AQ, the Proposition follows.
Proposition 15.3. If a = (a1, . . . , am) ∈ Nm, write xa =Qm
(a) For a ∈ Nm and M < [[a]] + Q there exists C > 0 so that
j=1 xaj
j .
ZN (x)≤A
xaN (x)−M dx ≤ C A[[a]]+Q−M .
109
(b) For a ∈ Nm, M > [[a]] + Q there exists C > 0 so that
ZN (x)≥A
xaN (x)−M dx ≤ C A[[a]]+Q−M .
(c) For a ∈ Nm, M = [[a]] + Q, and B > 4A, there exists C > 0 so that
Z
A<N (x)<B
xaN (x)−M dx ≤ C log(cid:20) B
A(cid:21) .
Proof. We have xa =Qm
ZN (x)≤A
xaN (x)−M dx .
j=1 xjaj .Qm
∞Xj=1
∞Xj=1
∞Xj=1
≈
≈
j=1 N (x)aj dj = N (x)[[a]]. Thus if [[a]] + Q − M > 0,
Z
2−j A<N (x)≤2−j+1A
N (x)[[a]]−M dx
(2−jA)[[a]]−M(cid:12)(cid:12){x : N (x) ≤ 2−j+1A}(cid:12)(cid:12)
(2−jA)[[a]]−M (2−j+1A)Q . A[[a]]+Q−M .
Similarly, if [[a]] + Q − M < 0,
ZN (x)≥A
xaN (x)−M dx .
≈
.
Z
2j A≤N (x)<2j+1A
N (x)[[a]]−M dx
(2jA)[[a]]−M(cid:12)(cid:12){x : N (x) ≤ 2−j+1A}(cid:12)(cid:12)
(2jA)[[a]]−M (2−jA)Q
∞Xj=0
∞Xj=0
∞Xj=0
. A[[a]]+Q−M .
Finally, if [[a]] + Q − M = 0,
Z
A≤N (x)≤B
xaN (x)−M dx .
Z
A≤N (x)≤B
N (x)[[a]]−M dx.
If B > 4A, choose j < k so that 2j−1 < A ≤ 2j ≤ 2k ≤ B < 2k+1. Then
Z
A≤N (x)≤B
N (x)[[a]]−M dx ≤
.
.
This completes the proof.
2l≤N (x)<2l+1
N (x)[[a]]−M dx
Z
2l([[a]]−M)(cid:12)(cid:12){x : N (x) ≤ 2l+1}(cid:12)(cid:12)
2l([[a]]−M+Q) = k − j + 2 ≈ log(cid:20) B
A(cid:21) .
k+1Xl=j−1
k+1Xl=j−1
k+1Xl=j−1
110
Proposition 15.4. Let m ∈ C∞(Rm \ {0}) have compact support, and let L = {l1, . . . , lr} ⊂
{1, . . . , m}. Suppose for every γ ∈ N there is a constant C > 0 so that if j ∈ L then
Put K(x) =RRm e2πihx,ξim(ξ) dξ. There is a constant C > 0 independent of the support of
m so that
Proof. Let ϕ ∈ C∞
We have
K(x) =ZRm
Since m is bounded, we have
ξj
m(ξ1, . . . , ξm)(cid:12)(cid:12) ≤ CN (ξ)−djγ.
(cid:12)(cid:12)∂γ
(cid:12)(cid:12)K(x)(cid:12)(cid:12) ≤ C(cid:0)Xj∈L
xj 1/dj(cid:1)−Q
e2πihx,ξiϕ(cid:0)λN (ξ)(cid:1)m(ξ) dξ +ZRm
.
(cid:12)(cid:12)I(cid:12)(cid:12) ≤ m∞(cid:12)(cid:12)(cid:8)ξ : N (ξ) ≤ 2λ−1(cid:9)(cid:12)(cid:12) . λ−Q.
(cid:12)(cid:12)∂γ
ξj(cid:2)1 − ϕ(cid:0)λN (ξ)(cid:1)(cid:3)m(ξ)(cid:12)(cid:12) ≤ CN (ξ)−djγ
0 (R) with ϕ(t) ≡ 1 if t ≤ 1, ϕ(t) ≡ 0 if t ≥ 2 and 0 ≤ ϕ(t) ≤ 1 for all t.
e2πihx,ξi(cid:2)1 − ϕ(cid:0)λN (ξ)(cid:1)(cid:3)m(ξ) dξ = I + II.
On the other hand, the chain, the product rule, and Proposition 15.1 show that
since any derivative of ϕ(λN (ξ)) is supported where λN (ξ) ≤ 2. For j ∈ L we can integrate
by parts M times in the variable ξj and obtain
(cid:12)(cid:12)II(cid:12)(cid:12) . xj−M ZN (ξ)≥λ−1(cid:12)(cid:12)(cid:12)∂M
. xj−M ZN (ξ)≥λ−1
. xj−M λ−Q+dj M .
ξjh(cid:2)1 − ϕ(cid:0)λN (ξ)(cid:1)(cid:3)m(ξ)i(cid:12)(cid:12)(cid:12) dξ
N (ξ)−Mdj dξ
Setting λ = xj 1/dj , it follows that K(x) . xj −Q/dj for every j ∈ L, and hence that
K(x) . min
j∈L
xj −Q/dj ≈(cid:16)Xj∈L
xj1/dj(cid:17)−Q
.
This completes the proof.
Remark 15.5. Note that in this Proposition, we only assume estimates on "pure" ξj deriva-
tives of m for j ∈ L. The conclusion is that the inverse Fourier transform K = m has the
expected decay relative to the partial norm NL(x) = Pj∈L xj1/dj . If L = {1, . . . , n}, the
conclusion is that K(x) . N (x)−Q.
16 Appendix III: Estimates for geometric sums
Proposition 16.1. Let ǫ > 0, a1, . . . , am > 0 with m ≥ 2 andQm
Γ =nI = (i1, . . . , im) ∈ Zm : ajij ≤ ij+1 < 0 for 1 ≤ j ≤ m − 1, amim ≤ i1 < 0o.
j=1 aj 6= 1. Set
For I = (i1, . . . , im) ∈ Γ set Λ(I) = [i2 − a1i1] + · · · + [im − am−1im−1] + [i1 − amim]. Then
there is a constant C > 0 depending on {ǫ, a1, . . . , am} so thatPI∈Γ 2−ǫΛ(I) ≤ C.
111
j=1 aj 6= 1, it follows that Λ(t) > 0 for t ∈ Γ.
j=1 tj = 1} is compact. Then Λ(t) ≥ 0 on Γ,
and if Λ(t) = 0 with t ∈ Γ, then t1 = amtm and tj+1 = ajtj for 1 ≤ j ≤ m − 1, so that
In particular,
by compactness there exists η > 0 so that if t ∈ Γ and t = 1 then Λ(t) > η. Since Λ is
j=1 tj = 1} on which Λ(t) > η. By
Proof. Γ is a convex cone and {t ∈ Γ : Pm
j=1 aj(cid:1)t1. Since Qm
t1 = (cid:0)Qm
continuous, there is an open neighborhood of {t ∈ Γ :Pm
homogeneity it follows that Λ(t) ≥ ηPm
If I ∈ Γ, then Λ(I) ≥ ηPn
j=1 ij ≤h ∞Xi=0
XI∈Γ
2−ǫΛ(I) ≤XI∈Γ
j=1 tj for allt ∈ Γ.
j=1 ij. Thus we have
2−ǫη Pm
2−ǫηiim
which is a constant depending only on ǫ and a1, . . . , am. This completes the proof.
Now let E = {e(j, k)} be an n × n matrix satisfying (2.10) such that 1 < e(j, k)e(k, j)
for all 1 ≤ j, k ≤ n. Let ǫ > 0. Let τ : {1, . . . , n} → {1, . . . , n} with τ (j) 6= j for all j. Let
0 ≤ b < a ≤ n. Put
ΓZ(E) =nI = (i1, . . . , in) ∈ Zn : e(j, k)ik ≤ ij,
1 ≤ j, k ≤ no,
ΓZ(E′) =nI ′ = (i1, . . . , ia) ∈ Za : e(j, k)ik ≤ ij,
1 ≤ j, k ≤ ao,
ΓZ(E′′) =nI ′′ = (ia+1, . . . , in) ∈ Zn−a : e(j, k)ik ≤ ij, a + 1 ≤ j, k ≤ no.
Note that ΓZ(E) ⊂ ΓZ(E′) × ΓZ(E′′). Also, if I ′′ ∈ ΓZ(E′′) then
Λ(I ′′) =nI ′ = (i1, . . . , ia) ∈ Za : (I ′, I ′′) ∈ ΓZ(E)o ⊂ ΓZ(E′).
Proposition 16.2. Let Rb+1, . . . , Ra be positive real numbers. Then
F = XI ′∈ΓZ(E′)" bYj=1
2−ǫ[ij −e(j,τ (j))iτ (j)]#" aYj=b+1
minn(2ij Rj)+ǫ, (2ij Rj )−ǫo#
converges and is bounded by a constant C, depending on ǫ, E but independent of Rb+1, . . . , Ra.
Here it is understood that if b = 0 the first product is empty, and if 0 < b = a the second
product is empty.
Proof. We argue by induction on n ≥ 1. When n = 1 we have
F ≤Xj∈Z
minn(2jR)+ǫ, (2jR)−ǫo.
which converges and is independent of R. Thus assume that the result is true for some
n − 1 ≥ 1. There are two possibilities. First suppose that b = 0. Then
minn(2ij Rj)+ǫ, (2ij Rj)−ǫo
aYj=1
F = XI ′∈ΓZ(E′)
minn(2ik Rk)+ǫ, (2ik Rk)−ǫo
aYk=1
≤ X(i1,...,ia)∈Za
aYj=1"Xij ∈Z
minn(2ij Rj)+ǫ, (2ij Rj)−ǫo#
=
112
which is independent of R1, . . . , Ra.
Next suppose that b ≥ 1 so that the productQb
are now two possibilities.
Case 1:
j=1 2−ǫ[ij −e(j,τ (j))iτ (j)] is non-empty. There
Suppose there exists j0 ∈ {1, . . . , b} such that j0 6= τ (k) for any k ∈ {1, . . . , b}. Without
loss of generality, assume j0 = 1. Set
eΓZ(E′) =neI ′ = (i2, . . . , ia) ∈ Za−1 : e(j, k)ik ≤ ij, 2 ≤ j, k ≤ ao.
Then ΓZ(E′) ⊂n(i1,eI ′) : e(1, τ (1))iτ (1) ≤ i1 and eI ′ ∈eΓZ(E′)o. Since the index i1 does not
appear in the productQb
XeI ′∈eΓZ(E′)" bYj=2
2−ǫ[ij −e(j,τ (j))iτ (j)]#" aYj=b+1
j=2 2−ǫ[ij −e(j,τ (j))iτ (j)], the sum F is dominated by
minn(2ij Rj)+ǫ, (2ij Rj)−ǫo#
2−ǫ[i1−e(1,τ (1))iτ (1)]#
×" Xi1∈Z
e(1,τ (1))iτ (1)≤i1
But
Xi1∈Z
e(1,τ (1))iτ (1)≤i1
and so
F ≤
1
1 − 2−ǫe(1,τ (1)) XeI ′∈eΓZ(E′)" bYj=2
,
2−ǫ[i1−e(1,τ (1))iτ (1)] ≤h1 − 2−ǫe(1,τ (1))i−1
2−ǫ[ij −e(j,τ (j))iτ (j)]#" aYj=b+1
minn(2ij Rj)+ǫ, (2ij Rj)−ǫo#.
The result now follows by induction since in the remaining sum we have replaced {1, . . . , n}
with {2, . . . , n}.
Case 2:
Now assume that each j ∈ {1, . . . , b} is equal to τ (k) for at least one k ∈ {1, . . . , b}. Let
σ(j) ∈ {1, . . . , b} be a choice of some pre-image of j. Then σ : {1, . . . , b} → {1, . . . , b}, and
since τ (σ(j)) = j, it follows that σ is one-to-one. Thus σ (and hence also τ ) is a permutation
of {1, . . . , b}. It follows that we can decompose {1, . . . , b} into a disjoint union of non-empty
subsets C1, . . . , Cp which are the minimal closed orbits of the mapping τ . We can write
Cl = {jl,1, . . . , jl,ml} with τ (jl,k) = jl,k+1 for 1 ≤ k ≤ ml − 1 and τ (jl,kl ) = jl,1. Then for
1 ≤ l ≤ p let
Γ′
l =n(il,1, . . . , il,ml) ∈ Zml : e(jl,k, τ (jl,k))iτ (jl,k) ≤ ijl,k 1 ≤ k ≤ mlo.
Then ΓZ(E′) ⊂ Γ′
1 × · · · × Γp, and so
l
F ≤
2−ǫ[ijl,k −e(jl,k,τ (jl,k))iτ (jl,k )]i
pYl=1h XIl∈Γ′
minn(2ik Rk)+ǫ, (2ik Rk)−ǫoi
mlYk=1
However PIl∈Γ′
lQml
and Proposition 16.1. Also,Pik∈Z minn(2ik Rk)+ǫ, (2ik Rk)−ǫo is bounded independently of
k=1 2−ǫ[ijl,k −e(jl,k,τ (jl,k))iτ (jl,k )] is uniformly bounded by Proposition 14.8
Rk. Thus F is bounded independently of is+1 and Ra+1, . . . , Rb. This completes the proof.
bYk=a+1hXik∈Z
that for all (A1, . . . , An) with Aj > 0,
j=1(e(αj + 1) + Mj) there is a constant C = C(M1, . . . , Mn, n, e) so
Proposition 16.3. Let E =(cid:8)e(j, k)(cid:9) be an n × n matrix satisfying (2.10), and suppose that
1 < e(j, k)e(k, j) for all j 6= k. Let e = max(cid:8)e(j, k) : 1 ≤ j, k ≤ n(cid:9). Let αj > 0, Mj ≥ 0 for
1 ≤ j ≤ n. If M ≥Pn
2−ij αjih1 +
XI∈ΓZ(E)h nYj=1
Proof. Let Nj(A) =Pn
nXk=1
. If I = (i1, . . . , in) ∈ ΓZ(E) then for any 1 ≤ j ≤ n we have
2−ik Aki−M
nYj=1hAe(j,1)
2−ik Akie
1 + 2−ij Nj(A) ≤ 1 +
−ij < −e(j, k)ik and so
2−ike(j,k)Ae(j,k)
+ · · · + Ae(j,n)
(1 + Aj)−Mj .
i−αj
nXk=1
k=1 Ae(j,k)
k
nYj=1
≤ C
1
n
113
.
≤
.
,
k
≤ 1 +
nXk=1
nXk=1
.h1 +
nXk=1(cid:2)2−ik Ak(cid:3)e(j,k)
2−ik Aki−e(αj +1) nYj=1h1 + Aji−Mj
nYj=1h1 +
nYj=1h1 + 2−ij Nj(A)i−(αj +1) nYj=1h1 + Aji−Mj
nXk=1
2−ij αjh1 + 2−ij Nj(A)i−(αj +1) nYj=1h1 + Aji−Mj
2−ij αjh1 + 2−ij Nj(A)i−αj −1i nYj=1h1 + Aji−Mj
nYj=1h1 + Aji−Mj
2−ik Aki−M
.
h1 +
nXk=1
2−ik Aki−M
XI∈ΓZ(E)h nYj=1
2−ij αjih1 +
nYj=1
. XI∈ΓZ(E)
nYj=1hXij ≤0
nYj=1
Nj(A)−αj
. C
.
Therefore
and so
This completes the proof.
114
References
[CGP13] Tristan C. Collins, Allan Greenleaf, and Malabika Pramanik. A multi-dimensional
resolution of singularities with applications to analysis. Amer. J. Math.,
135(5):1179 -- 1252, 2013.
[DLM10] Yong Ding, Guo Zhen Lu, and Bo Lin Ma. Multi-parameter Triebel-Lizorkin and
Besov spaces associated with flag singular integrals. Acta Math. Sin. (Engl. Ser.),
26(4):603 -- 620, 2010.
[FS82]
[G lo10a]
[G lo10b]
[G lo13]
[GS77]
[HL10]
[HLL10]
[HLL13]
[HLS14]
[LZ13]
G. B. Folland and Elias M. Stein. Hardy spaces on homogeneous groups, volume 28
of Mathematical Notes. Princeton University Press, Princeton, N.J.; University of
Tokyo Press, Tokyo, 1982.
Pawe l G lowacki. Composition and L2-boundedness of flag kernels. Colloq. Math.,
118(2):581 -- 585, 2010.
Pawe l G lowacki. Correction to "Composition and L2-boundedness of flag kernels"
[mr2602167]. Colloq. Math., 120(2):331, 2010.
Pawe l G lowacki. Lp-boundedness of flag kernels on homogeneous groups via sym-
bolic calculus. J. Lie Theory, 23(4):953 -- 977, 2013.
P. C. Greiner and E. M. Stein. Estimates for the ∂-Neumann problem. Princeton
University Press, Princeton, N.J., 1977. Mathematical Notes, No. 19.
Yongsheng Han and Guozhen Lu. Some recent works on multiparameter Hardy
space theory and discrete Littlewood-Paley analysis. In Trends in partial differ-
ential equations, volume 10 of Adv. Lect. Math. (ALM), pages 99 -- 191. Int. Press,
Somerville, MA, 2010.
Yongsheng Han, Ji Li, and Guozhen Lu. Duality of multiparameter Hardy spaces
H p on spaces of homogeneous type. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5),
9(4):645 -- 685, 2010.
Yongsheng Han, Ji Li, and Guozhen Lu. Multiparameter Hardy space theory on
Carnot-Carath´eodory spaces and product spaces of homogeneous type. Trans.
Amer. Math. Soc., 365(1):319 -- 360, 2013.
Yongsheng Han, Guozhen Lu, and Eric Sawyer.
Flag Hardy spaces and
Marcinkiewicz multipliers on the Heisenberg group. Anal. PDE, 7(7):1465 -- 1534,
2014.
Guo Zhen Lu and Yue Ping Zhu. Singular integrals and weighted Triebel-Lizorkin
and Besov spaces of arbitrary number of parameters. Acta Math. Sin. (Engl. Ser.),
29(1):39 -- 52, 2013.
[MPR15] Detlef Muller, Marco M. Peloso, and Fulvio Ricci. Analysis of the Hodge Laplacian
on the Heisenberg group. Mem. Amer. Math. Soc., 233(1095):vi+91, 2015.
[MRS95] Detlef Muller, Fulvio Ricci, and Elias M. Stein. Marcinkiewicz multipliers
and multi-parameter structure on Heisenberg (-type) groups. I. Invent. Math.,
119(2):199 -- 233, 1995.
[NRS01]
Alexander Nagel, Fulvio Ricci, and Elias M. Stein. Singular integrals with flag
kernels and analysis on quadratic CR manifolds. J. Funct. Anal., 181(1):29 -- 118,
2001.
115
[NRSW12] Alexander Nagel, Fulvio Ricci, Elias Stein, and Stephen Wainger. Singular inte-
grals with flag kernels on homogeneous groups, I. Rev. Mat. Iberoam., 28(3):631 --
722, 2012.
[PS82]
[RS88]
[Rua10]
[Rua11]
[Ste70]
[SY13]
[Str14]
[WL12]
D. H. Phong and E. M. Stein. Some further classes of pseudodifferential and
singular-integral operators arising in boundary value problems. I. Composition of
operators. Amer. J. Math., 104(1):141 -- 172, 1982.
Fulvio Ricci and Elias M. Stein. Harmonic analysis on nilpotent groups and
singular integrals. II. Singular kernels supported on submanifolds. J. Funct. Anal.,
78(1):56 -- 84, 1988.
Zhuoping Ruan. Weighted Hardy spaces in the three-parameter case. J. Math.
Anal. Appl., 367(2):625 -- 639, 2010.
Zhuoping Ruan. Non-isotropic flag singular integrals on multi-parameter Hardy
spaces. Taiwanese J. Math., 15(2):473 -- 499, 2011.
Elias M. Stein. Singular integrals and differentiability properties of functions.
Princeton Mathematical Series, No. 30. Princeton University Press, Princeton,
N.J., 1970.
Elias M. Stein and Po-Lam Yung. Pseudodifferential operators of mixed type
adapted to distributions of k-planes. Math. Res. Lett., 20(6):1183 -- 1208, 2013.
Brian Street. Multi-parameter singular integrals. Annals of Mathematics Studies,
Volume 189, Princeton University Press, Princeton, N.J., 2014.
Xinfeng Wu and Zongguang Liu. Characterizations of multiparameter Besov and
Triebel-Lizorkin spaces associated with flag singular integrals. J. Funct. Spaces
Appl., pages Art. ID 275791, 18, 2012.
[Wu14a]
Xinfeng Wu. An atomic decomposition characterization of flag Hardy spaces
H p
F (Rn × Rm) with applications. J. Geom. Anal., 24(2):613 -- 626, 2014.
[Wu14b]
Xinfeng Wu. Weighted norm inequalities for flag singular integrals on homoge-
neous groups. Taiwanese J. Math., 18(2):357 -- 369, 2014.
[Yan09]
Dachun Yang. Besov and Triebel-Lizorkin spaces related to singular integrals with
flag kernels. Rev. Mat. Complut., 22(1):253 -- 302, 2009.
Index of symbols
B(ρ), 12
CZa, 48
CZa,0, 76
CZ0, 97
S , 17
S , bEA
ES, bES, EA
E♯, 30
ES, 39
EN , 53
ES,T,W , 70
m0, mS, 42
nj, 12
Nj, bNj, 13
nS,r,bnS,r, 21
NS,rbNS,r, 39
bNr,W , 71
j, 83
L,j, 83
Na, 48
N ′
N ′
(f )t, [f ]t, 13
F , 83
FL, 83
F ⊥, 84
[[γ]], 12
Γ(E), Γµ(E), Γo(E), ΓZ,µ(E), ΓZ(E), 28
ΓW (S, T ), 65
Γ(S, T, W ), 69
g, 96
hk, 96
I ∗
r , J ∗
a , 67
λ · x, 12
λ ·S xIr , λ·S ξIr , 21
λ ·a x, 48
Λ(S, T, W ), 60
M(E), 14
M∞(E), 15
P(E), 14
P0(E), 15
Ψ0, ΨS, 42
Qj, 12
QS,r,bQS,r, 21
RCj , 11
RL, 12
S(n), 17
σS(kr, l), 18
S′(E), 38
ΣS,T , 64
τS(l, kr), 18
TK, 57
fW(S, T ), W(S, T ), 65
xCj , 11
xL, 12
116
|
1809.03882 | 1 | 1809 | 2018-09-09T20:10:22 | Generalized multipliers for left-invertible operators and applications | [
"math.FA"
] | We introduce generalized multipliers for left-invertible operators which formal Laurent series $U_x(z)=\sum_{n=1}^\infty(P_ET^{n}x) \frac{1}{z^n}+\sum_{n=0}^\infty(P_E{T^{\prime*}}^{n}x)z^n$ actually represent analytic functions on an annulus or a disc. | math.FA | math |
Generalized multipliers for left-invertible operators
and applications
Pawe l Pietrzycki
Abstract. We introduce generalized multipliers for left-invertible operators
which formal Laurent series Ux(z) = P∞
zn + P∞
x)zn
actually represent analytic functions on an annulus or a disc.
n=1(PE T nx) 1
n=0(PE T ′∗ n
1. Introduction
In [20] S. Shimorin obtain a weak analog of the Wold decomposition theorem,
representing operator close to isometry in some sense as a direct sum of a unitary
operator and a shift operator acting in some reproducing kernel Hilbert space of
vector-valued holomorphic functions defined on a disc. The construction of the
Shimorin's model for a left-invertible analytic operator T ∈ B(H) is as follows. Let
E := N (T ∗) and define a vector-valued holomorphic functions Ux as
Ux(z) =
∞Xn=0
(PET ′∗n
x)zn,
z ∈ D(r(T ′)−1),
where T ′ is the Cauchy dual of T . Then we equip the obtained space of analytic
functions H := {Ux : x ∈ H} with the norm induced by H. The operator U : H ∋
x → Ux ∈ H becomes a unitary operator. Moreover, Shimorin proved that H
is a reproducing kernel Hilbert space and the operator T is unitary equivalent to
the operator Mz of multiplication by z on H and T ′∗ is unitary equivalent to the
operator L given by the
(L f )(z) =
f (z) − f (0)
z
,
f ∈ H .
Following [20], the reproducing kernel for H is an B(E)-valued function of two
variables κH : Ω × Ω → B(E) that
(i) for any e ∈ E and λ ∈ Ω
κH (·, λ)e ∈ H
2010 Mathematics Subject Classification. Primary 47B20, 47B33; Secondary 47B37.
Key words and phrases. left-invertible operators, composition operator, weighted shift,
weighted shift on directed three.
1
2
P. PIETRZYCKI
(ii) for any e ∈ E, f ∈ H and λ ∈ Ω
hf (λ), eiE = hf, κH (·, λ)eiH
The class of weighted shifts on a directed tree was introduced in [9] and in-
tensively studied since then [7, 2, 4]. The class is a source of interesting exam-
ples (see e.g., [8, 13]). In [7] S. Chavan and S. Trivedi showed that a weighted
shift Sλ on a rooted directed tree with finite branching index is analytic therefore
can be modelled as a multiplication operator Mz on a reproducing kernel Hilbert
space H of E-valued holomorphic functions on a disc centered at the origin, where
E := N (S∗
λ). Moreover, they proved that the reproducing kernel associated with
H is multi-diagonal.
In [4] P. Budzy´nski, P. Dymek and M. Ptak introduced the notion of multiplier
In [8] P. Dymek, A. P laneta and M. Ptak
algebra induced by a weighted shift.
extended this notion to the case of left-invertible analytic operators.
2. Preliminaries
In this paper, we use the following notation. The fields of rational, real and
complex numbers are denoted by Q, R and C, respectively. The symbols Z, Z+,
N and R+ stand for the sets of integers, positive integers, nonnegative integers,
and nonnegative real numbers, respectively. Set D(r) = {z ∈ C : z 6 r} and
A(r−, r+) = {z ∈ C : r− 6 z 6 r+} for r, r−, r+ ∈ R+. The expression "a count-
able set" means a finite set or a countably infinite set.
All Hilbert spaces considered in this paper are assumed to be complex. Let
T be a linear operator in a complex Hilbert space H. Denote by T ∗ the adjoint
of T . We write B(H) for the C∗-algebra of all bounded operators and the cone
of all positive operators in H, respectively. The spectrum and spectral radius of
T ∈ B(H) is denoted by σ(T ) and r(T ) respectively. Let T ∈ B(H). We say that
T is left-invertible if there exists S ∈ B(H) such that ST = I. The Cauchy dual
operator T ′ of a left-invertible operator T ∈ B(H) is defined by
T ′ = T (T ∗T )−1
The notion of the Cauchy dual operator has been introduced and studied by Shi-
morin in the context of the wandering subspace problem for Bergman-type operators
[20]. We call T analytic if H∞ :=T∞
i=1 T i = {0}.
Let X be a set and φ : X → X. If n ∈ N then the n-th iterate of φ is given by
φ(n) = φ ◦ φ ◦ · · · ◦ φ, φ composed with itself n-times. For x ∈ X the set
[x]φ = {y ∈ X : there exist i, j ∈ N such that φ(i)(x) = φ(j)(y)}
is called the orbit of f containing x. If x ∈ X and φ(i)(x) = x for some i ∈ Z+
then the cycle of φ containing x is the set
Define the function [φ] : X → N by
Cφ = {φ(i)(x) : i ∈ N}
(i) [φ](x) = 0 if x is in the cycle of φ
(ii) [φ](x∗) = 0, where x∗ is a fixed element of X not containing a cycle,
(iii) [φ](φ(x)) = [φ](x) + 1 if x is not in a cycle of φ.
GENERALIZED MULTIPLIERS FOR LEFT-INVERTIBLE OPERATORS
3
We set
for m, n ∈ N.
Genφ (m, n) := {x ∈ X : m 6 [φ](x) 6 n}
Let (X, A , µ) be a µ-finite measure space, φ : X → X and w : X → C be
measurable transformations. By a weighted composition operator Cϕ,w in L2(µ) we
mean a mapping
(2.1)
D(Cϕ,w) = {f ∈ L2(µ) : w(f ◦ φ) ∈ L2(µ)},
Cϕ,wf = w(f ◦ φ),
f ∈ D(Cϕ,w).
Let us recall some useful properties of composition operator we need in this
paper:
Lemma 2.1. Let X be a countable set, φ : X → X and w : X → C be measurable
transformations. If Cϕ,w ∈ B(ℓ2(X)) then for any x ∈ X and n ∈ N
(i) C∗
(ii) C∗n
(iii) C n
(iv) C∗
ϕ,wex = w(x)eϕ(x)
ϕ,wex = w(x)w(φ(x)) · · · w(φ(n−1)(x))eφ(n)(x),
ϕ,wex =Py∈ϕ−n(x) w(y)w(φ(y)) · · · w(φ(n−1)(y))ey,
ϕ,wCϕ,wex =(cid:16)Py∈ϕ−1(x) w(y)2(cid:17)ex.
We now describe Cauchy dual of weighted composition operator
Lemma 2.2. Let X be a countable set, φ : X → X and w : X → C be measurable
transformations. If Cϕ,w ∈ B(ℓ2(X)) is left-invertible operator then the Cauchy
dual C′
ϕ,w of Cϕ,w is also a weighted composition operator and is given by:
C′
ϕ,wex = Xy∈ϕ−1(x)
w(y)
(cid:16)Pz∈ϕ−1(y) w(z)2(cid:17) ey.
Let T = (V ; E) be a directed tree (V and E are the sets of vertices and edges
of T , respectively). For any vertex u ∈ V we put Chi(u) = {v ∈ V : (u, v) ∈ E}.
Denote by par the partial function from V to V which assigns to a vertex u a
unique v ∈ V such that (v, u) ∈ E. A vertex u ∈ V is called a root of T if u
If T has a root, we denote it by root. Put V ◦ = V \ {root}
has no parent.
if T has a root and V ◦ = V otherwise. The Hilbert space of square summable
complex functions on V equipped with the standard inner product is denoted by
ℓ2(V ). For u ∈ V , we define eu ∈ ℓ2(V ) to be the characteristic function of the set
{u}. It turns out that the set {ev}v∈V is an orthonormal basis of ℓ2(V ). We put
V≺ := {v ∈ V : card(Chi(V )) ≥ 2} and call the a member of this set a branching
vertex of T
Given a system λ = {λv}v∈V ◦ of complex numbers, we define the operator Sλ
in ℓ2(V ), which is called a weighted shift on T with weights λ, as follows
D(Sλ) = {f ∈ ℓ2(V ) : ΛT f ∈ ℓ2(V )}
and Sλf = ΛT f
for
f ∈ D(Sλ),
where
(ΛT f )(v) =(cid:26) λvf (par(v))
0
if v ∈ V ◦,
otherwise.
4
P. PIETRZYCKI
Lemma 2.3. If Sλ is a densely defined weighted shift on a directed tree T with
weights λ = {λv}v∈V ◦ , then
(2.2)
N (S∗
λ) =(cid:26) herooti ⊕Lu∈V≺
Lu∈V≺
(ℓ2(Chi(u)) ⊖ hλui)
(ℓ2(Chi(u)) ⊖ hλui)
if T has a root,
otherwise,
where λu ∈ ℓ2(Chi(u)) is given by λu : ℓ2(Chi(u)) ∋ v → λv ∈ C
A subgraph of a directed tree T which itself is a directed tree will be called
a subtree of T . We refer the reader to [9] for more details on weighted shifts on
directed trees.
3. Generalized multipliers
In the recent paper [15] we introduced a new analytic model for left-invertible
operators. Now, we recall this model. Let T ∈ B(H) be a left-invertible operator
and E be a subspace of H denote by [E]T ∗,T ′ the direct sum of the smallest T ′-
invariant subspace containing E and the smallest T ∗-invariant subspace containing
E:
[E]T ∗,T ′ :=_{T ∗nx : x ∈ E, n ∈ N} ⊕_{T ′n
where T ′ is the Cauchy dual of T .
x : x ∈ E, n ∈ N},
To avoid the repetition, we state the following assumption which will be used
frequently in this section.
(LI)
The operator T ∈ B(H) is left-invertible and E is a subspace of H
such that [E]T ∗,T ′ = H.
Suppose (LI) holds.
In this case we may construct a Hilbert H associated
with T , of formal Laurent series with vector coefficients. We proceed as follows.
For each x ∈ H, define a formal Laurent series Ux with vector coefficients as
(3.1)
Ux(z) =
∞Xn=1
(PE T nx)
1
zn +
∞Xn=0
(PET ′∗n
x)zn.
Let H denote the vector space of formal Laurent series with vector coefficients
of the form Ux, x ∈ H. Consider the map U : H → H defined by U x = Ux. As
shown in [15] U is injective. In particular, we may equip the space H with the
norm induced from H, so that U is unitary.
By [15] the operator T is unitary equivalent to the operator Mz : H → H of
multiplication by z on H given by
(3.2)
(Mzf )(z) = zf (z),
f ∈ H
and operator T ′∗ is unitary equivalent to the operator L : H → H given by
(3.3)
(L f )(z) =
f (z) − (PN (M ∗
z )f )(z)
z
,
f ∈ H .
For left-invertible operator T ∈ B(H), among all subspaces satisfying condition
(LI) we will distinguish those subspaces E which satisfy the following condition
(3.4)
E ⊥ T nE
and
E ⊥ T ′n
E,
n ∈ Z+
GENERALIZED MULTIPLIERS FOR LEFT-INVERTIBLE OPERATORS
5
Observe that every f ∈ H can be represented as follows
where
(3.5)
f (n)zn,
f =
∞Xn=−∞
f (n) =(cid:26) PE T ′∗nU ∗f
PE T −nU ∗f
if n ∈ N
if n ∈ Z \ N.
Lemma 3.1. Let {fn}∞
n=0 ⊂ H and f ∈ H be such that limn→∞ fn = f . Then
lim
n→∞
fn(k) = f (k)
for k ∈ Z.
Proof. It follows directly from (3.5).
(cid:3)
In [4] P. Budzy´nski, P. Dymek and M. Ptak introduced the notion of multiplier
algebra induced by a weighted shift. In [8] P. Dymek, A. P laneta, M. Ptak extended
this notion to the case of left-invertible analytic operators.
We introduce generalized multipliers for left-invertible operators which formal
Laurent series (3.1) actually represent analytic functions on an annulus or a disc.
Define the Cauchy-type multiplication ∗ : B(E)Z × E Z → E Z given by
(3.6)
( ϕ ∗ f )(n) =
∞Xk=−∞
ϕ(k) f (n − k)
ϕ ∈ B(E)Z, f ∈ E Z.
We define the operator M ϕ : H ⊇ D(M ϕ) → H by
D(M ϕ) = {f ∈ H : there is g ∈ H such that ϕ ∗ f = g},
M ϕf = g if ϕ ∗ f = g.
Lemma 3.2. Let ϕ : Z → B(E). Then following assertions are satisfied:
(i) for every e ∈ E and n ∈ Z, ( ϕ ∗ d(U e))(n) = ϕ(n)e
(ii)
[M ϕf (n) =
+
∞Xk=−∞
∞Xk=0
ϕ(n − k) f (k) =
ϕ(n + k)PET kU ∗f
∞Xk=1
ϕ(n − k)PE T ′∗k
U ∗f,
f ∈ D(M ϕ), n ∈ Z
(iii)
(M ϕf )(z) =
∞Xn=−∞(cid:16) ∞Xk=1
ϕ(n + k)PE T kU ∗f +
∞Xk=0
ϕ(n − k)PET ′∗k
U ∗f(cid:17)zn,
f ∈ D(M ϕ).
Proof.
(i) Fix e ∈ E. By (3.4) and (3.6), we have
( ϕ ∗ [(U e))(n) =
∞Xk=1
ϕ(n + k)PE T ke +
∞Xk=0
ϕ(n − k)PE T ′∗k
e = ϕ(n)e, n ∈ Z.
(cid:3)
6
P. PIETRZYCKI
We call ϕ a generalized multiplier of T and M ϕ a generalized multiplication
operator if M ϕ ∈ B(H ). The set of all generalized multipliers of the operator T we
denote by GM(T ). One can easily verify that the set GM(T ) is a linear subspace
of B(E)Z. Consider the map V : GM(T ) ∋ ϕ → M ϕ ∈ B(H ). By Lemma 3.2, the
kernel of V is trivial. In particular, we may equip the space GM(T ) with the norm
k·k : GM(T ) → [0, ∞) induced from B(H ), so that V is isometry:
k ϕk := kM ϕk ,
ϕ ∈ GM(T ).
For operator A ∈ B(H) let ϕA : Z → B(E) be a function defined by
ϕA(m) =(cid:26) PE T ′∗mAE if m ∈ N
PE T −mAE if m ∈ Z \ N
Theorem 3.3. Let T be left-invertible. The following assertions are satisfied:
(i) For n ∈ N the sequence χ{n}IE is generalized multiplier and Mχ{n}IE =
M n
z . If n ∈ Z \ N then D(Mχ{n}IE ) = R(Mz) and Mχ{n}IE = L−n.
(ii) Mz commutes with M ϕ, for ϕ ∈ GM(T ),
(iii) ϕ ∗ ψ ∈ GM(T ) for every ϕ, ψ ∈ GM(T ) and
(iv) The space GM(T ) endowed with the Cauchy-type multiplication
M ϕM ψ = M ϕ∗ ψ,
(3.7)
( ϕ ∗ ψ)(n) =
∞Xk=−∞
is a Banach algebra.
Proof.
(i)
ϕ(k) ψ(n − k)
ϕ, ψ ∈ B(E)Z,
Consider first the case when n ∈ N. Fix f ∈ H and set g = M n
z f and
ϕ = χ{n}IE. Then ϕ ∗ f = g. If n ∈ Z \ N and f ∈ R(Mz) then by (3.3) we have
z f . Define g = L −nf . As in the previous case we obtain ϕ ∗ f = g. If
L f = 1
f ∈ N (M ∗
z ) \ {0} then
f (n) =(cid:26) 0
PET nU ∗f
if Z+
if n ∈ Z \ Z+.
Hence, ϕ ∗ f (0) = 0 and there exist some k ∈ Z such that ϕ ∗ f (k) 6= 0 which
contradicts (3.5).
(ii)
(M ϕMzf )(z) =
∞Xn=−∞(cid:16) ∞Xk=1
ϕ(n + k)PET kU ∗Mzf +
= Mz
∞Xn=−∞(cid:16) ∞Xk=1
= (MzM ϕf )(z),
for f ∈ H
(iii)
ϕ(n + k)PET kU ∗f +
ϕ(n − k)PET ′∗k
ϕ(n − k)PE T ′∗k
∞Xk=0
∞Xk=0
U ∗Mzf(cid:17)zn
U ∗f(cid:17)zn
V
M ϕM ψf
(n) =
=
=
∞Xk=−∞
∞Xk=−∞
∞Xl=−∞
ϕ(k)
∞Xl=−∞
∞Xk=−∞
∞Xj=−∞
∞Xk=−∞
∞Xl=−∞
∞Xl=−∞
GENERALIZED MULTIPLIERS FOR LEFT-INVERTIBLE OPERATORS
7
ϕ(k)[M ψf (n − k) =
∞Xk=−∞
ϕ(k)
ψ(j) f (n − k − j)
ψ(l − k) f (n − l) =
ϕ(k) ψ(l − k) f (n − l)
ϕ(k) ψ(l − k) f (n − l) =
ϕ ∗ ψ(l) f (n − l)
= ( ϕ ∗ ψ) ∗ f (n)
f ∈ H , n ∈ Z.
This implies that ( ϕ∗ ψ)∗ f = M ϕM ψf
for every f ∈ H . Hence D(M ϕ∗ ψ) = H
and M ϕM ψ = M ϕ∗ ψ. This in turn implies that ϕ ∗ ψ ∈ GM(T ), because M ϕ, M ψ ∈
H .
V
(iv)
Suppose that a sequence { ϕn}∞
n=0 ⊂ GM(T ) is a Cauchy sequence. Since the map
V : GM(T ) ∋ ϕ → M ϕ ∈ B(H ) is isometry, we see that the sequence {M ϕn}∞
n=0 ⊂
B(H ) is also a Cauchy sequence and there exists an operator A ∈ B(H ) such that
limn→∞ ϕn = A.
(cid:3)
Theorem 3.4. Let T ∈ B(H) be left-invertible and E ⊂ H be such that,
(i) [E]T ∗,T ′ = H and [E]T ′∗ ,T = H,
(ii)
(3.8)
T nT ′∗n
E ⊂ E,
n ∈ N,
(iii) formal Laurent series (3.1) converges absolutely in E on Ω such int Ω 6= ∅.
(iv) (3.4) holds
If A ∈ B(H) commutes with T , then ϕA ∈ A and A = U ∗M ϕAU .
Proof. Let A = U AU ∗. All we need to prove is the following equality
(3.9)
f ∈ H , n ∈ Z.
(cϕA ∗ f )(n) = cAf (n),
Fix n ∈ Z. Consider first the case when f = U T me, e ∈ E, m ∈ N. By (3.4)
ϕA(n + k)PET kU ∗f = ϕA(n + k)PE T k+me = χset0(k + m) ϕA(n)e,
and
ϕA(n − k)PE T ′∗k
U ∗f =
ϕA(n − k)PE T ′∗k−me = 0
ϕA(n − k)PE T m−ke = 0
ϕA(n − m)e
if k > m
if k < m
if k = m.
ϕA(n − m)e =
PE T ′∗n−mAe = PET ′∗nAT me
PE T m−nAe = PE T ′∗nAT me
PE T m−nAe = PE T −nAT me
if n > m
if n < m and n ≥ 0
if n < m and n < 0
8
P. PIETRZYCKI
This altogether implies that
(cϕA ∗ f )(n) =
= χNPE T ′∗n
∞Xk=1 cϕA(n + k)PET kU ∗f +
= cAf (n),
where f = U T me for e ∈ E, m ∈ N.
∞Xk=0 cϕA(n − k)PET ′∗k
U ∗f
AT me + χZ\NPET ′∗n
AT me
In turn, if f = U T ′∗me m ∈ Z+. It is plain that
ϕA(n − k)PE T ′∗k
U ∗f = ϕA(n − k)PE T ′∗k+m
e = 0.
It follows from (3.4) and inclusion (3.8) that
ϕA(n + k)PET k−m(T mT ′∗me) = 0
if k > m
ϕA(n + k)PET ′∗m−k(T mT ′∗me) = 0 if k < m
ϕA(n + m)(T mT ′∗me)
if k = m.
PE T ′∗n+mAge = PE T ′∗nAT ′∗me
PE T ′∗n+mAge = PE T −nAT ′∗me
PE T −(m+n)Age = PET −nAT ′∗me
if n + m ≥ 0, n ≥ 0
if n + m ≥ 0, n < 0
if n + m < 0
Let ge = T mT ′∗me then
ϕA(n + k)PET kU ∗f =
ϕA(n + m)ge =
(cϕA ∗ f )(n) =
As a consequence, we have
∞Xk=0 cϕA(n − k)PET ′∗k
AT ′∗m
e
U ∗f
e + χZ\NPET ′∗n
= χNPE T ′∗n
∞Xk=1 cϕA(n + k)PET kU ∗f +
= cAf (n),
AT ′∗m
where f = U T ′∗me, for e ∈ E, m ∈ N. We extend the previous equality by linearity
to the following space
lin{U T nx : x ∈ E, n ∈ N} ⊕ lin{U T ′∗n
x : x ∈ E, n ∈ N}.
An application of Lemma 3.1 gives (3.9) which completes the proof.
(cid:3)
It is interesting to observe that the class of left-invertible and analytic operators
and the class of weighted shift on leafless directed trees satisfy the assumptions of
the previous theorem.
Example 3.5. Let T ∈ B(H) be left-invertible and analytic and E := N (T ∗).
By [20, Proposition 2.7], H⊥
∞ = [E]T ′ . Since T is analytic H∞ = {0}, we see
that [E]T ′∗,T ⊃ [E]T = H and [E]T ∗,T ′ ⊃ [E]T ′ = H, which yields [E]T ∗,T ′ = H
and [E]T ′∗ ,T = H. Using equality E = N (T ∗) one can show that (3.4) holds and
T nT ′∗nE = {0} ⊂ E
Example 3.6. Let T be a rootless and leafless directed tree and λ = {λv}v∈V
be a system of weights. Let Sλ be the weighted shift on T and E := hωi ⊕ N (S∗
λ).
Assume that Sλ ∈ B(H) and formal Laurent series (3.1) converges absolutely in
E on Ω such int Ω 6= ∅. By [15, Lemma 4.2] and [10, Lemma 4.3.1] we have
[E]T ∗,T ′ = H and [E]T ′∗,T = H, where T = Sλ. It is a matter of routine to verify
that T nT ′∗nE ⊂ {ω} ⊂ E.
GENERALIZED MULTIPLIERS FOR LEFT-INVERTIBLE OPERATORS
9
4. Weighted shifts on directed trees
In [4] P. Budzy´nski, P. Dymek, M. Ptak.
introduced a notion of a multiplier
algebra induced by a weighted shift, which is defined via related multiplication
operators. Assume that
(4.1)
T = (V, E) is a countably infinite rooted and leafless directed tree,
and λ = {λv}v∈V ◦ ⊂ (0, ∞)
For u ∈ V and v ∈ Des(u) we set
λuv =(cid:26) 1
n=0 λparn(v)
if u = v
if park(v) = u.
Let ϕ : N → C. Define the mapping Γλ
ϕ : CV → CV by
(Γλ
ϕ)(v) =
λpark(v)v ϕ(k)f (park(v)),
v ∈ V.
Qk−1
vXk=0
The multiplication operator M λ
ϕ : ℓ2(V ) ⊇ D(M λ
ϕ ) = {f ∈ ℓ2(V ) : Γλ
ϕ f = Γλ
ϕf,
f ∈ D(M λ
ϕ )
D(M λ
M λ
ϕ ) → ℓ2(V ) is given by
ϕf ∈ ℓ2(V )},
We can write the above definition of M λ
ϕ in the following form
(4.2)
M λ
ϕ f =
∞Xn=0
ϕ(k)Sk
λf
Theorem 4.1. The following equality holds
(M ψf )(z) =
∞Xk=1
ψ(−k)S
′∗k
λ f +
∞Xk=0
ψ(k)Sk
λf
for f ∈ R(Sλ)
Proof. First, we note that
∞Xn=−∞
f (n − k)zn =(cid:26) M k
z f
L −kf
if k ∈ N and f ∈ H
if k ∈ Z \ N and f ∈ R(Mz).
Let ϕ be a sequence with finite support. By Theorem 3.3 the sequence ϕ induces
bounded operator M ϕ on subspace R(Mz). Changing order of summation we obtain
(M ψf )(z) =
=
=
=
∞Xn=−∞
∞Xk=−∞
∞Xk=−∞
∞Xk=0
∞Xk=−∞
ψ(k) f (n − k)zn =
∞Xk=−∞
∞Xn=−∞
ψ(k) f (n − k)zn
ψ(k)
∞Xn=−∞
f (n − k)zn
ψ(k)(χN(k)M k
z f + χZ\N(k)L −kf )
ψ(k)M k
z f +
∞Xk=1
ψ(−k)L kf.
10
P. PIETRZYCKI
Let, for n ∈ N , denote by cpn : Z → C the coefficients of the n-th Fejer kernel, i.e.,
cpn(m) =(cid:26) 1 − m
0
n+1
if m ≤ n
if m > n
As in the proof of [8, Proposition 15] one can show that M pn ϕ
SOT→ M ϕ in R(Mz).
(cid:3)
References
[1] A. Anand, S. Chavan, Z. J. Jab lo´nski, J. Stochel, Complete systems of unitary invariants for
some classes of 2-isometries, https://arxiv.org/abs/1806.03229
[2] P. Budzy´nski, P. Dymek, A. P laneta, M. Ptak. Weighted shifts on directed trees. Their
multiplier algebras, reflexivity and decompositions. arXiv preprint arXiv:1702.00765 (2017).
[3] P. Budzy´nski, P. Dymek, Z. J. Jab lo´nski, J. Stochel, Subnormal weighted shifts on directed
trees and composition operators in L2-spaces with non-densely defined powers, Abs. Appl.
Anal. (2014), 791-817,
[4] P. Budzy´nski, P. Dymek, M. Ptak. Analytic structure of weighted shifts on directed trees.
Mathematische Nachrichten (2016).
[5] P. Budzy´nski, Z. J. Jab lo´nski, I. B. Jung, J. Stochel, On unbounded composition operators
in L2-spaces, Ann. Mat. Pur. Appl. 193 (2014), 663-688.
[6] J. Carlson, The spectra and commutants of some weighted composition operators, Trans.
Amer. Math. Soc. 317 (1990), 631-654.
[7] S. Chavan, S. Trivedi, An analytic model for left-invertible weighted shifts on directed trees,
J. London Math. Soc. 94 (2016), 253-279.
[8] P. Dymek, A. Paneta, M. Ptak. Generalized multipliers for left-invertible analytic operators
and their applications to commutant and reflexivity. Journal of Functional Analysis (2018).
[9] Z. J. Jab lo´nski, I. B. Jung, J. Stochel, Weighted shifts on directed trees, Mem. Amer. Math.
Soc. 216 (2012), no. 1017.
[10] Z. J. Jab lo´nski, I. B. Jung, J. Stochel, A non-hyponormal operator generating Stieltjes mo-
ment sequences, J. Funct. Anal., 262 (2012), pp. 3946-3980.
domain, Proc. Amer. Math. Soc. 142 (2014), 3109-3116.
J. Math. Anal. Appl. 446 (2017), 823-842.
[11] E. A. Nordgren, Composition operators, Canad. J. Math. 20 (1968), 442-449. 21.
[12] E. A. Nordgren, Composition operators on Hilbert spaces, Hubert Space Operators, Lecture
Notes in Math., vol. 693, Springer-VerlagB, erlin, 1978, pp. 37-63.
[13] P. Pietrzycki, The single equality A∗nAn = (A∗A)n does not imply the quasinormality of
weighted shifts on rootless directed trees. Journal of Mathematical Analysis and Applications
(2016): 338-348.
[14] P. Pietrzycki, Reduced commutativity
of moduli
of
operators.
arXiv preprint
arXiv:1802.01007.
[15] P. Pietrzycki, A Shimorin-type analytic model for left-invertible operators on an annulus and
applications
[16] W. C. Ridge, Spectrum of a composition operator, Proc. Amer. Math. Soc. 37 (1973), 121-
127.
[17] W. Rudin, Functional Analysis, McGraw-Hill, International Editions 1991
[18] R. K. Singh, J. S. Manhas, Composition operators on function spaces, Elsevier Science Pub-
lishers B.V., North-Holland-Amsterdam, 1993.
[19] A. Shields, Weighted shift operators, analytic function theory, Math. Surveys, Amer. Math.
Soc., Providence, 1974, 49-128.
[20] S. Shimorin, Wold-type decompositions and wandering subspaces for operators close to isome-
tries. Journal fr die Reine und Angewandte Mathematik 531 (2001): 147-189.
[21] D. Xia, On the analytic model of a class of hyponormal operator, Integral Eq. Oper. Theory
6 (1983), 134-157.
GENERALIZED MULTIPLIERS FOR LEFT-INVERTIBLE OPERATORS
11
Wydzia l Matematyki i Informatyki, Uniwersytet Jagiello´nski, ul.
Lojasiewicza 6,
PL-30348 Krak´ow
E-mail address: [email protected]
|
1802.02722 | 3 | 1802 | 2019-03-25T05:53:49 | B-metric spaces, fixed points and Lipschitz functions | [
"math.FA"
] | The paper is concerned with b-metric and generalized b-metric spaces. One proves the existence of the completion of a generalized b-metric space and some fixed point results. The behavior of Lipschitz functions on b-metric spaces of homogeneous type, as well as of Lipschitz functions defined on, or with values in quasi-Banach spaces, is studied. | math.FA | math | B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
S. COBZAS¸
Abstract. The paper is concerned with b-metric and generalized b-metric spaces. One proves
the existence of the completion of a generalized b-metric space and some fixed point results. The
behavior of Lipschitz functions on b-metric spaces of homogeneous type, as well as of Lipschitz
functions defined on, or with values in quasi-Banach spaces, is studied.
MSC2010: 54E25 54E35 47H09 47H10 46A16 26A16
Keywords: metric space, generalized metric space, b-metric space, completion, metrizability,
fixed point, quasi-Banach space, Lipschitz mapping
9
1
0
2
r
a
M
5
2
]
.
A
F
h
t
a
m
[
3
v
2
2
7
2
0
.
2
0
8
1
:
v
i
X
r
a
Contents
Introduction
1. b-metric spaces
1.1. Topological properties and metrizability
1.2. An axiomatic definition of balls in b-metric spaces
1.3. Strong b-metric spaces and completion
1.4. Spaces of homogeneous type
1.5. Topological properties of f -quasimetric spaces
1.6. Historical remarks and further results
2. Generalized b-metric spaces
2.1. The completion of generalized b-metric spaces
3. Fixed points in b-metric spaces
3.1. Fixed points in b-metric spaces
3.2. Fixed points in generalized b-metric spaces
4. Lipschitz functions
4.1. Quasi-normed spaces
4.2. Lipschitz functions and quasi-normed spaces
4.3. Lipschitz functions on spaces of homogeneous type
References
1
1
2
6
8
9
10
14
15
16
17
17
22
23
23
26
29
30
Introduction
There are a lot of extensions of the notions of metric and metric space -- see for instance the
books [32], [53], [71], or the survey papers [22], [51]. In this paper we concentrate on b-metric and
generalized b-metric spaces, with emphasis on their topological properties, some fixed point results
and Lipschitz functions on such spaces.
A part of the results from this paper are included in [24].
In this section we present some results on b-metric spaces.
1. b-metric spaces
Date: March 26, 2019.
1
2
S. COBZAS¸
1.1. Topological properties and metrizability. A b-metric on a nonempty set X is a function
d : X × X → [0, ∞) satisfying the conditions
(1.1)
(i) d(x, y) = 0 ⇐⇒ x = y;
(ii) d(x, y) = d(y, x);
(iii) d(x, y) ≤ s[d(x, z) + d(z, y)],
for all x, y, z ∈ X, and for some fixed number s ≥ 1. The pair (X, d) is called a b-metric space.
Obviously, for s = 1 one obtains a metric on X.
Example 1.1. If (X, d) is a metric space and β > 1, then dβ(x, y) is a b-metric.
Indeed,
dβ(x, y) ≤ [d(x, z) + d(z, y)]β ≤ 2β (max{d(x, z), d(z, y)})β
≤ 2β[dβ(x, y) + dβ(x, y)] .
The s-relaxed triangle inequality implies
d(x0, xn) ≤ sd(x0, x1) + s2d(x1, x2) + · · · + sn−1d(xn−2, xn−1) + sn−1d(xn−1, xn) ,
(1.2)
for all n ∈ N and all x0, x1, . . . , xn ∈ X.
Indeed, we obtain successively
d(x0, xn) ≤ sd(x0, x1) + sd(x1, xn) ≤ sd(x0, x1) + s2d(x1, x2) + s2d(x2, xn) ≤ . . .
≤ sd(x0, x1) + s2d(x1, x2) + · · · + sn−1d(xn−2, xn−1) + sn−1d(xn−1, xn)
Along with the inequality (iii), called the s-relaxed triangle inequality, one considers also the
s-relaxed polygonal inequality
(iv)
for all x0, x1, . . . , xn ∈ X and all n ∈ N.
d(x0, xn) ≤ s[d(x0, x1) + d(x1, x2) + · · · + d(xn−1, xn)],
For n = 2 one obtains the inequality (iii). The following example shows that the converse is not
true -- there exist b-metrics that do not satisfy the relaxed polygonal inequality.
Example 1.2 ([53], Theorem 12.10). Let X = [0, 1] and d(x, y) = (x − y)2, x, y ∈ [0, 1]. Then d is
a 2-relaxed metric on X which is not polygonally s-relaxed for any s ≥ 1.
Indeed, it is easy to check that d satisfies the 2-relaxed triangle inequality. Suppose that, for
n , 1 ≤ i ≤ n − 1, we obtain
some s ≥ 1, d satisfies the s-relaxed polygonal inequality. Taking xi = i
1
s
=
1
s
n(cid:19)2
· d(0, 1) ≤ d(0, x1) + d(x1, x2) + · · · + d(xn−1, 1) = n ·(cid:18) 1
=
1
n
,
for all n ∈ N, which is impossible.
One can consider also an ultrametric version of (iii):
(iii′)
d(x, y) ≤ λ max{d(x, z), d(y, z)} ,
for all x, y, z ∈ X. It is obvious that
(iii′) =⇒ (iii) with s = λ;
(iii) =⇒ (iii′) with λ = 2s.
The condition
(iii′′)
for all ε > 0 and x, y, z ∈ X, is equivalent to (iii′) with λ = 2.
max{d(x, z), d(y, z)} ≤ ε =⇒ d(x, y) ≤ 2ε ,
Let now (X, d) be again a b-metric space. One introduces a topology on a b-metric space (X, d)
in the usual way. The "open" ball B(x, r) of center x ∈ X and radius r > 0 is given by
B(x, r) = {y ∈ X : d(x, y) < r} .
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
3
A subset Y of X is called open if for every x ∈ Y there exists a number rx > 0 such that
B(x, rx) ⊆ Y. Denoting by τd (or τ (d)) the family of all open subsets of X it follows that τd
satisfies the axioms of a topology. This topology is derived from a uniformity Ud on X having as
basis the sets
Uε = {(x, y) ∈ X × X : d(x, y) < ε},
ε > 0 .
The uniformity Ud has a countable basis {U1/n : n ∈ N} so that, by Frink's metrization theorem
([39]), the uniformity Ud is derived from a metric ρ, hence the topology τd as well. This was
remarked in the paper [57]. In [37] it is shown that the topology τd satisfies the hypotheses of the
Nagata-Smirnov metrizability theorem.
Concerning the metrizability of uniform and topological spaces, see the treatise [36].
There exist also direct proofs of the metrizability of the topology of a b-metric space.
Let (X, d) be a b-metric space. Put
ρ(x, y) = infn nXk=0
d(xi−1, xi)o ,
where the infimum is taken over all n ∈ N and all chains x = x0, x1, . . . , xn = y of elements in X
connecting x and y.
As remarked Frink [39], if a b-metric d satisfies (iii′) for λ = 2, then formula (1.3) defines a
metric equivalent to d. We present the result in the form given by Schroeder [72].
Theorem 1.3 (A. H. Frink [39] and V. Schroeder [72]). If d : X × X → [0, ∞) satisfies the
conditions (i), (ii) from (1.1) and (iii′) for some 1 ≤ λ ≤ 2, then the function ρ defined by (1.3) is
a metric on X satisfying the inequalities 1
2λ d ≤ ρ ≤ d.
V. Schroeder [72] also showed that for every ε > 0 there exists a b-metric d satisfying (1.1).(iii)
with s = 1 + ε such that the mapping ρ defined by (1.3) is not a metric. Other example showing
the limits of Frink's metrization method was given in An and Dung [10].
General results of metrizability were obtained in [2] and [64] by a slight modification of Frink's
technique.
Let (X, d) be a b-metric space. For 0 < p ≤ 1 define
(1.3)
(1.4)
ρp(x, y) = infn nXk=0
d(xi−1, xi)po ,
where the infimum is taken over all n ∈ N and all chains x = x0, x1, . . . , xn = y of elements in X.
The function ρp defined by (1.4) satisfies the conditions
(1) ρp(x, y) = ρp(y, x),
(2) ρp(x, y) ≤ ρp(x, z) + ρp(z, y),
(3) dp(x, y) ≤ ρp(x, y) ,
for all x, y, z ∈ X, i.e., ρ is a pseudometric on X and dp is dominated by ρ.
Theorem 1.4 ([64]). Let d be a b-metric on a nonempty set X satisfying the s-relaxed triangle
inequality (1.1).(iii), for some s ≥ 1. If the number p ∈ (0, 1] is given by the equation (2s)p = 2,
then the mapping ρp : X × X → [0, ∞) defined by (1.4) is a metric on X satisfying the inequalities
(1.5)
for all x, y ∈ X.
ρp(x, y) ≤ dp(x, y) ≤ 2ρp(x, y) ,
The same conclusions hold if d satisfies the conditions (i), (ii) from (1.1) and (iii′) for some
λ ≥ 2. In this case 0 < p ≤ 1 is given by λp = 2 and the metric ρp satisfies the inequalities
(1.6)
for all x, y ∈ X.
ρp(x, y) ≤ dp(x, y) ≤ 4ρp(x, y) ,
The inequalities (1.5) have the following consequences.
4
S. COBZAS¸
Corollary 1.5. Under the hypotheses of Theorem 1.4, τd = τρ, that is, the topology of any b-
metric space is metrizable, and the convergence of sequences with respect to τd is characterized in
the following way:
xn
τd−→ x ⇐⇒ d(x, xn) −→ 0 ,
for any sequence (xn) in X and x ∈ X.
Proof. The equality of topologies follows from the inclusions
valid for all x ∈ X and r > 0.
Bd(x, r1/p) ⊆ Bρ(x, r) and Bρ(cid:0)x, 4−1rp(cid:1) ⊆ Bd(x, r) ,
The statement concerning sequences is a consequence of the equality τd = τρ and of the inequal-
(cid:3)
ities (1.5).
Remark 1.6. In [2] the proof is given for a p satisfying the inequalities
(1.7)
1 ≥ p ≥ (log2(3s))−2 ,
while in Theorem 1.4 the result holds for
(1.8)
Putting
p = (log2(2s))−1 .
ρ(p) = sup{p ∈ (0, 1] : ρp is a metric, Lipschitz equivalent to dp} ,
the estimation 1.7 yields ρ(p) ≥ (log2(3s))−2, while from 1.8 one obtains the better evaluation
ρ(p) ≥ (log2(2s))−1, which cannot be improved, as it is shown by the example of the spaces ℓp
with 0 < p < 1.
A proof of Theorem 1.4 is also given in the book by Heinonen [42, Prop. 14.5], with the
evaluation p ≥ (log2 λ)−2, where λ is the constant from (iii′).
Remark 1.7. It follows that ρ(x, y) = ρp(x, y)1/p, x, y ∈ X, is a b-metric on X, Lipschitz equiv-
alent to d and satisfying the inequality
ρ(x, y)p ≤ ρ(x, z)p + ρ(z, y)p ,
for all x, y, z ∈ X. This is a well known fact in the theory of quasi-normed spaces, where a
quasi-norm k · k satisfying the inequality
kx + ykp ≤ kxkp + kykp,
for some 0 < p ≤ 1 is called a p-norm (see Subsection 4.1).
Let (X, d) be a b-metric space. The b-metric d is called
• continuous if
(1.9)
d(xn, x) → 0 and d(yn, y) → 0 =⇒ d(xn, yn) → d(x, y) ,
• separately continuous if the function d(x, ·) is continuous on X for every x ∈ X, i.e.,
(1.10)
d(yn, y) → 0 =⇒ d(x, yn) → d(x, y) ,
for all sequences (xn), (yn) in X and all x, y ∈ X.
The topology τd generated by a b-metric d has some peculiarities -- a ball B(x, r) need not be
τd-open and the b-metric d could not be continuous on X × X.
Remark 1.8. Let (X, d) be a b-metric space and x ∈ X. Then
B(x, r) is τd-open for every r > 0 ⇐⇒ d(x, ·) is upper semicontinuous on X.
Consequently, if the b-metric is separately continuous on X, then the balls B(x, r) are τd-open.
The equivalence follows from the equality
B(x, r) = d(x, ·)−1(cid:0)(−∞, r)(cid:1) .
We present now an example of a b-metric space where the balls are not necessarily open.
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
5
Example 1.9 ([64]). Consider a fixed number ε > 0. For X = N0 = {0, 1, . . . } let d : X × X →
[0, ∞) be defined by
d(0, 1) = 1,
d(1, m) = 1
m ,
d(0, m) = 1 + ε
d(n, m) = 1
n + 1
m
for m ≥ 2
for n ≥ 2
and extended to X × X by d(n, n) = 0 and symmetry.
Then
for all m, n, k ∈ X, B(cid:0)0, 1 + ε
r > 0, that is, for any 1 ∈ B(cid:0)0, 1 + ε
2(cid:1) is not τd-open.
ball B(cid:0)0, 1 + ε
Other examples are given in [11].
d(m, n) ≤ (1 + ε)[d(n, k) + d(k, m)] ,
2(cid:1) = {0, 1} and the ball B(1, r) contains an infinity of terms for every
2(cid:1) for every r > 0, showing that the
2(cid:1), B(1, r) * B(cid:0)0, 1 + ε
Remark 1.10. If, for some 0 < p < 1, a b-metric d on a set X satisfies the inequality
(1.11)
d(x, y)p ≤ d(x, z)p + d(z, y)p,
for all x, y, z ∈ X,
then the balls corresponding to d are τd-open. Moreover, the b-metric d is continuous.
By Remark 1.7, the b-metric ρ = ρp
p corresponding to the metric ρp constructed in Theorem 1.4
satisfies the inequality (1.11).
Indeed, let B(x, r) be a ball in (X, d) and y ∈ B(x, r). We have to show that there exists r′ > 0
such that B(y, r′) ⊆ B(x, r). Taking r′ := (rp − d(x, y)p)1/p > 0, then d(y, z) < r′ implies
d(x, z)p ≤ d(x, y)p + d(y, z)p
< d(x, y)p + r′p = rp,
that is, d(x, z) < r. The continuity of the b-metric d follows from the inequality
d(xn, yn)p − d(x, y)p ≤ d(xn, x)p + d(yn, y)p ,
which can be proved as in the metric case (using (1.11)).
Equivalence notions for b-metrics.
In connection to the metrizability of b-metric spaces, we mention the following notions of equiv-
alence for b-metrics.
Let d1, d2 be two b-metrics on the same set X. Then d1, d2 are called
• topologically equivalent if τd1 = τd2;
• uniformly equivalent if the identity mapping IX on X is uniformly continuous both from
(X, d1) to (X, d2) as well as from (X, d2) to (X, d1), i.e.
∀ε > 0, ∃δ(ε) > 0 such that ∀x, y ∈ X (d1(x, y) ≤ δ(ε) =⇒ d2(x, y) ≤ ε),
∀ε > 0, ∃δ(ε) > 0 such that ∀x, y ∈ X (d2(x, y) ≤ δ(ε) =⇒ d1(x, y) ≤ ε) .
• Lipschitz equivalent if there exist c1, c2 > 0 such that
c1d2(x, y) ≤ d1(x, y) ≤ c2d2(x, y) ,
for all x, y ∈ X
Of course, the above definitions applies to metrics as well, as particular cases of b-metrics.
Remark 1.11. It is obvious that, in general,
Lipschitz equivalence =⇒ uniform equivalence =⇒ topological equivalence.
For quasi-norms, topological equivalence is equivalent to Lipschitz equivalence.
So the expression "the topology τd generated by a b-metric d on a set X is metrizable" means
that there exists a metric ρ on X topologically equivalent to d.
The problem of the existence of a metric that is Lipschitz equivalent to a b-metric was solved
in [37], where this property was called metric boundedness.
6
S. COBZAS¸
Theorem 1.12 ([37], see also [53], Theorem 12.9). Let (X, d) be a b-metric space. Then d is
Lipschitz equivalent to a metric if and only if d satisfies the s-relaxed polygonal inequality (iv) for
some s ≥ 1.
1.2. An axiomatic definition of balls in b-metric spaces. H. Aimar [2] found a set of prop-
erties characterizing balls in b-metric spaces.
For a nonempty set X consider a mapping U : X × (0, ∞) → P(X) satisfying the following
properties:
(1.12)
(i) \r>0
(ii) [r>0
U (x, r) = {x};
U (x, r) = X;
(iii)
(iv)
0 < r1 ≤ r2 =⇒ U (x, r1) ⊆ U (x, r2);
there exists c ≥ 1 such that
y ∈ U (x, r) =⇒ U (x, r) ⊆ U (y, cr) and U (y, r) ⊆ U (x, cr),
for all x ∈ X and r > 0.
We call the sets U (x, r) formal balls.
Remark 1.13. By (i), x ∈ U (x, r) for all r > 0 and x ∈ X.
It is easy to check that if (X, d) is a b-metric space, where the b-metric d satisfies the relaxed
triangle inequality for some s ≥ 1, then the sets U (x, r) = Bd(x, r), x ∈ X, r > 0 satisfy the
properties from (1.12) with c = 2s.
Conditions (i) -- (iii) are easy to check. For (iv), if y ∈ Bd(x, r) and z ∈ Bd(x, r), then
d(y, z) ≤ s(d(y, x) + d(x, z)) < 2sr ,
i.e. z ∈ Bd(y, 2sr).
Similarly, y ∈ Bd(x, r) and z ∈ Bd(y, r) imply
d(x, z) ≤ s(d(x, y) + d(y, z)) < 2sr ,
i.e. z ∈ Bd(x, 2sr).
It can be shown that, conversely, a family of subsets of X satisfying the properties from (1.12)
generates a b-metric on X, the balls corresponding to d being tightly connected with the sets
U (x, r).
Theorem 1.14 ([2]). For a nonempty set X and a mapping U : X × (0, ∞) → P(X) satisfying
the properties from (1.12), define d : X × X → [0, ∞) by
(1.13)
Then:
d(x, y) = inf{r > 0 : y ∈ U (x, r) and x ∈ U (y, r)},
x, y ∈ X.
(1) d is a b-metric on X satisfying the relaxed triangle inequality for s = c;
(2) the open balls Bd(x, r) corresponding to d and the sets U (x, r) are related by the inclusions
(1.14)
for all x ∈ X and r > 0, where γ > 1.
U(cid:0)x, (γc)−1r(cid:1) ⊆ Bd(x, r) ⊆ U (x, r),
Proof. We shall verify only the relaxed triangle inequality. Let x, y, z ∈ X and ε > 0. By the
definition (1.13) of d there exists r1, r2 > 0 such that
(1.15)
0 < r1 < d(x, z) + ε and z ∈ U (x, r1), x ∈ U (z, r1);
0 < r2 < d(z, y) + ε and z ∈ U (y, r2), y ∈ U (z, r2).
Taking into account (iii) it follows
x, y ∈ U (z, r1 + r2) and z ∈ U (x, r1 + r2) ∩ U (y, r1 + r2) .
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
7
Applying (iv) to x ∈ U (z, r1 + r2) one obtains
y ∈ U (z, r1 + r2) ⊆ U (x, c(r1 + r2)).
Similarly, y ∈ U (z, r1 + r2) implies
x ∈ U (z, r1 + r2) ⊆ U (y, c(r1 + r2)),
so that, by the definition of d, d(x, y) ≤ c(r1 + r2). But then, by adding the inequalities (1.15),
one obtains
d(x, y) ≤ c(r1 + r2) < c(d(x, z) + d(z, y)) + 2cε .
Since these hold for all ε > 0, it follows
d(x, y) ≤ c(d(x, z) + d(z, y)) .
Let us prove now the inclusions (1.14).
If d(x, y) < r, then there exists 0 < r′ < r such that y ∈ U (x, r′) and x ∈ U (y, r′). By (iii),
U (x, r′) ⊆ U (x, r), so that Bd(x, r) ⊆ U (x, r).
Let now y ∈ U (x, (γc)−1r). By Remark 1.13 and (iv),
x ∈ U(cid:0)x, (γc)−1r(cid:1) ⊆ U (y, r/γ) , and
y ∈ U(cid:0)y, (γc)−1r(cid:1) ⊆ U (x, r/γ) .
By the definition of d, d(x, y) ≤ r/γ < r , i.e. y ∈ Bd(x, r).
(cid:3)
Remark 1.15. In [2] it is shown that d satisfies the relaxed triangle inequality with s = 2c. Also,
the first inclusion in (1.14) is proved for γ = 2, i.e. it is shown that U(cid:0)y, (2c)−1r(cid:1) ⊆ Bd(x, r).
In [1] a similar characterization is given in terms of some subsets of X × X. Denoting by ∆ the
diagonal of X × X,
∆ = {(x, x) : x ∈ X} ,
one considers a mapping V : (0, ∞) → P(X × X) satisfying the properties:
V (r) = ∆;
V (r) = X × X;
(j) \r>0
(jj) [r>0
(jjj)
(1.16)
(jv)
for all r > 0.
0 < r1 ≤ r2 =⇒ V (r1) ⊆ V (r2);
there exists c ≥ 1 such that V (r) ◦ V (r) ⊆ V (cr).
By analogy with the case of uniform spaces we call the sets V (r) antourages.
If (X, d) is a
b-metric space with d satisfying the relaxed triangle inequality for some s ≥ 1, then the sets
Wd(r) = {(x, y) ∈ X × X : d(x, y) < r},
, r > 0,
satisfy the conditions from (1.16) with c = 2s. Indeed, the conditions (j) -- (jjj) are easily verified.
To check (jv), suppose that (x, y) ∈ Wd(r)◦ V (r). Then there exists z ∈ X such that (x, z) ∈ Wd(r)
and (z, y) ∈ Wd(r), implying
d(x, y) ≤ s(d(x, z) + d(z, y)) < 2sr ,
showing that (jv) holds with c = 2s.
A converse result holds in this case too.
Theorem 1.16 ([1]). For a nonempty set X and a mapping V : (0, ∞) → P(X × X) satisfying
the properties from (1.16), define d : X × X → [0, ∞) by
(1.17)
Then:
d(x, y) = inf{r > 0 : (x, y) ∈ V (r)},
x, y ∈ X.
(1) d is a b-metric on X satisfying the relaxed triangle inequality for s = c;
8
S. COBZAS¸
(2) for any 0 < γ < 1 the following inclusions hold
(1.18)
V (γr) ⊆ Wd(r) ⊆ V (r),
for all r > 0.
Proof. Again, we prove only the validity of the relaxed triangle inequality. Let x, y, z ∈ X. By the
definition of d there exist r1, r2 > 0 such that
(1.19)
0 < r1 < d(x, z) + ε and (x, z) ∈ V (r1);
0 < r2 < d(z, y) + ε and (z, y) ∈ V (r2).
By (jjj) V (r1) ∪ V (r2) ⊆ V (r1 + r2) so that, by (jv), (x, y) ∈ V (c(r1 + r2)), implying d(x, y) ≤
c(r1 + r2).
Consequently, the inequalities (1.15) yield by addition
d(x, y) ≤ c(r1 + r2) < c(d(x, z) + d(z, y)) + 2cε .
Since these hold for all ε > 0, it follows
d(x, y) ≤ c(d(x, z) + d(z, y)) .
The proof of the inclusions (1.18) is simpler than in the case considered in Theorem 1.14.
Indeed, (x, y) ∈ W(r) is equivalent to d(x, y) < r. By the definition of d, there exists 0 < r′ < r
such that (x, y) ∈ V (r′). By (jjj), V (r′) ⊆ V (r), showing that (x, y) ∈ V (r), i.e. Wd(r) ⊆ V (r).
If (x, y) ∈ V (γr), then
d(x, y) ≤ γr < r ,
showing that (x, y) ∈ Wd(r), i.e. V (γr) ⊆ Wd(r).
(cid:3)
1.3. Strong b-metric spaces and completion. Let (X, d) be a b-metric space. As we have seen,
the topology τd generated by the b-metric d has some drawbacks in what concerns the continuity
property of d and the topological openness of the "open" balls. To remedy these shortcomings Kirk
and Shahzad [53, §12.4] introduced a special class of b-metrics. A mapping d : X × X → [0, ∞) is
called a strong b-metric if it satisfies the conditions (i) and (ii) from (1.1) and
(v)
d(x, y) ≤ d(x, z) + sd(y, z) ,
for some s ≥ 1 and all x, y, z ∈ X. It is obvious that (v) is equivalent to
(v′)
d(x, y) ≤ min{sd(x, z) + d(y, z), d(x, z) + sd(y, z)} ,
for all x, y, z ∈ X, and that (v) implies the s-relaxed triangle inequality (iii) from (1.1).
The topology generated by a strong b-metric has good properties as, for instance, the openness
of the balls B(x, r). Indeed, if y ∈ B(x, r) then
d(y, z) ≤ d(x, y + sd(y, z) < ε ,
provided sd(y, z) < ε − d(x, y), that is B(y, r′) ⊆ B(x, r), where r′ = (ε − d(x, y))/s.
Also the following inequality
(1.20)
holds for all x, y, x′, y′ ∈ X, implying the continuity of the b-metric: if d(xn, x) → 0 and d(yn, y) →
0, then the relations
d(x, y) − d(x′, y′) ≤ s[d(x, x′) + d(y, y′)] ,
d(xn, yn) − d(x, y) ≤ s[d(xn, x) + d(yn, y)] −→ 0 as n → ∞ ,
show that d(xn, yn) −→ d(x, y) as n → ∞.
A strong b-metric satisfies the s-polygonal inequality. Indeed,
d(x0, xn) ≤ sd(x0, x1) + d(x1, xn) ≤ sd(x0, x1) + sd(x1, x2) + d(x2, xn) ≤ . . .
≤ s[sd(x0, x1) + sd(x1, x2) + · · · + d(xn−1, xn)] .
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
9
Completeness and completion.
A Cauchy sequence in a b-metric space (X, d) is a sequence (xn) in X such that limm,n→∞ d(xn, xm) =
0. The inequality d(xn, xm) ≤ s [d(xn, x) + d(x, xm)] shows that every convergent sequence is
Cauchy. The b-metric space (X, d) is called complete if every Cauchy sequence converges to some
x ∈ X. By a completion of a b-metric space (X, d) one understands a complete b-metric space
(Y, ρ) such that there exists an isometric embedding j : X → Y with j(X) dense in Y .
By an isometric embedding of a b-metric space (X1, d1) into a b-metric space (X2, d2) one
understands a mapping f : X1 → X2 such that
d2(f (x), f (y)) = d1(x, y) ,
for all x, y ∈ X1. Two b-metric spaces (X1, d1), (X2, d2) are called isometric if there exists a
surjective isometric embedding f : X1 → X2.
The completeness is preserved by the uniform equivalence of b-metrics, but not by the topological
equivalence.
A question raised in [53, p. 128] is:
Does every strong b-metric space admit a completion?
This question was answered in the affirmative in [12].
Theorem 1.17. Let (X, d) be a strong b-metric space.
1. There exists a complete strong b-metric space ( X, d ) which is a completion of (X, d).
2. The completion is unique up to an isometry, in the sense that if (X1, d1), (X2, d2) are
two strong b-metric spaces which are completions of (X, d), then (X1, d1) and (X2, d2) are
isometric.
Proof. The proof follows the ideas from the metric case. On the family C(X) of all Cauchy
sequences in X one considers the equivalence relation
(xn) ∼ (yn) ⇐⇒ lim
n
d(xn, yn) = 0 .
On the quotient space X = C(X)/∼ one defines d by d(ξ, η) = limn d(xn, yn), where (xn) ∈ ξ and
(yn) ∈ η, and one shows that ( X, d) is a complete strong b-metric space containing X isometrically
as a dense subset.
(cid:3)
Remark 1.18. As it is mentioned in [12], the existence of a completion of an arbitrary b-metric
space is still an important open problem.
1.4. Spaces of homogeneous type. Completing some earlier results of Coifman and de Guz-
man [25], Mac´ıas and Segovia [57, 58] considered b-metrics (under the name quasi-distances) in
connection with some problems in harmonic analysis.
The framework in [57] is the following. Let (X, d) be a b-metric space. One considers a positive
measure µ defined on a σ-algebra of subsets of X containing the open sets and the balls B(x, r)
such that
(1.21)
0 < µ (B(x, ar)) ≤ βµ (B(x, r)) ,
for all x ∈ X and r > 0, where a > 1 and A > 0 are fixed numbers. A b-metric space equipped with
a measure µ satisfying (1.21) is called a space of homogeneous type and is denoted by (X, d, µ). If
further, there exist c1, c2 > 0 such that
(1.22)
0 < µ({x}) < r < µ(X) =⇒ c1r ≤ µ (B(x, r)) ≤ c2r ,
for all x ∈ X, then the space (X, d, µ) of homogeneous type is called normal.
Remark 1.19. One can show that if (X, d) is a b-metric space with a positive measure µ satisfying
(1.22), then the space (X, d, µ) is of homogeneous type.
Concerning the openness of balls in b-metric spaces we mention the following result.
10
S. COBZAS¸
Theorem 1.20 ([57]). Let (X, d) be a b-metric space. Then there exist a b-metric d′ on X,
Lipschitz equivalent to d, and the constants C > 0 and 0 < α < 1 such that
(1.23)
whenever max{d′(x, z), d′(y, z)} < r.
d′(x, z) − d′(y, z) ≤ Cr1−αd′(x, y)α,
Remark 1.21. The inequality (1.23) can be written in the equivalent form
(1.24)
and it is easy to check that the balls corresponding to a b-metric d′ satisfying (1.24) are τd′-open.
d′(x, z) − d′(y, z) ≤ Cd′(x, y)α (max{d′(x, z), d′(y, z)})1−α ,
Indeed, let Bd′(x, r) be a ball. We have to show that for every y ∈ Bd′(x, r) there exists r′ > 0
such that Bd′(y, r′) ⊆ Bd′(x, r), that is,
d′(y, z) < r′ =⇒ d′(x, z) < r ,
for any z ∈ X. Supposing that d′ satisfies the s′-relaxed triangle inequality for some s′ ≥ 1, choose
first 0 < r′ < r. Then
d′(x, z) ≤ s′d′(x, y) + s′d′(y, z) < 2s′r .
By (1.24),
d′(x, z) ≤ d′(x, y) + d′(x, z) − d′(x, y)
< d′(x, y) + Cd′(y, z)α (max{d′(x, z), d′(x, y)})1−α
< d′(x, y) + C(2s′r)1−α(r′)α .
Choosing 0 < r′ < r such that C(2s′r)1−α(r′)α < r − d′(x, y), it follows d′(x, z) < r.
Concerning the set of points x ∈ X with µ({x}) > 0 we mention.
Proposition 1.22 ([57]). Let (X, d, µ) be a space of homogeneous type and
M = {x ∈ X : µ({x}) > 0} .
Then the set M is at most countable and for every x ∈ M there exists r > 0 such that M ∩
B(x, r) = {x}.
We mention also the following result.
Theorem 1.23 ([57]). Let (X, d, µ) be a space of homogeneous type such that the balls are τd-open.
Let δ : X × X → [0, ∞) be given by
δ(x, y) = inf{µ(B) : B is a ball containing x, y} ,
if x 6= y and δ(x, x) = 0.
Then δ is a b-metric on X, (X, δ, µ) is a normal space and τδ = τd.
1.5. Topological properties of f -quasimetric spaces. We present now, following [14] a very
general class of metric type spaces. On a nonempty set X consider a mapping d : X × X → R+
satisfying only the condition
(1.25)
d(x, y) = 0 ⇐⇒ x = y ,
for all x, y ∈ X, and call such a function distance. One can define open balls with respect to d as
usual
for x ∈ X and r > 0, and a topology τd by
B(x, r) = {y ∈ X : d(x, y) < r} ,
G ∈ τd
def⇐⇒ ∀x ∈ G, ∃r > 0, B(x, r) ⊆ G ,
where G ⊆ X.
The topology τd is T1 because X r {x} is open, and so {x} closed.
Indeed, if y 6= x, then
r := d(y, x) > 0 and x /∈ B(y, r).
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
11
Along with the distance d one can consider the conjugate distance ¯d(x, y) = d(y, x), x, y ∈ X.
The ¯d-balls are given by
B ¯d(x, r) = {y ∈ X : ¯d(x, y) < r} = {y ∈ X : d(y, x) < r} ,
and the corresponding topology is denoted by τ ¯d.
Consider now a function f : R+ × R+ → R+ such that
(1.26)
(t1, t2) → (0, 0) =⇒ f (t1, t2) → (0, 0) ,
where (t1, t2) ∈ R+ × R+. we say that a distance d on a set X is an f -quasimetric if it satisfies the
inequality
(1.27)
for all x, y, z ∈ X.
d(x, y) ≤ f (d(x, z), d(y, z)) ,
Example 1.24. We present first some important particular cases of function f .
• f (t1, t2) = t1 + t2. In this case d is a quasimetric (see [23]) and a metric if the distance d
is symmetric.
• f (t1, t2) = s1t1 + s2t2, for some s1, s2 ≥ 1. In this case d is called an (s1, s2)-quasimetric
(see [15]), a b-quasimetric if s1 = s2 = s, respectively an (s1, s2)-metric, and b-metric if d
is symmetric.
From 1.27 one obtains the following result, called the asymptotic triangle inequality:
(1.28)
d(xn, yn) → 0 and d(yn, zn) → 0 =⇒ d(xn, zn) → 0 .
Conversely, if a distance functions satisfies (1.28), then there exists a function f , satisfying
(1.26), such that d is an f -quasimetric.
Indeed, define h : R+ → R+ by
h(t) = sup{d(u, v) : u, v ∈ X, ∃w ∈ X, d(u, w) + d(w, v) ≤ t} .
The function h is obviously nondecreasing and
h(t) = 0 .
lim
tց0
Indeed, if tn → 0, where tn ∈ R+, n ∈ N, then there exist un, vn, wn ∈ X such that
d(un, wn) + d(wn, vn) ≤ tn and d(un, vn) > f (tn) −
1
n
,
for all n ∈ N. The first inequality implies d(un, wn), d(wn, vn) → 0, so that, by (1.28), d(un, vn) → 0,
which, by the second inequality from above, yields f (tn) → 0.
By the definition of the function h,
for all x, y, z ∈ X, so we can take f (t1, t2) = h(t1 + t2).
Define the convergence of a sequence (xn) in a distance space (X, d) to x ∈ X by
h(d(x, z) + d(z, y)) ≥ d(x, y) ,
and, for Z ⊆ X and x ∈ X put
xn
d−→ x def⇐⇒ d(x, xn) → 0 ,
d(x, Z) = inf{d(x, z) : z ∈ Z} .
We have the following useful characterizations of the interior and closure.
Proposition 1.25. Let (X, d) be an f -quasimetric space and Z ⊆ X. Then
(1.29)
int(Z) = {x ∈ X : d(x, X r Z) > 0} and cl(Z) = {x ∈ X : d(x, A) = 0} .
12
Proof. Let
We show that
S. COBZAS¸
Z := {x ∈ X : d(x, X r Z) > 0} .
(i)
(ii)
(iii)
Z is open;
Z ⊆ Z;
int(Z) ⊆ Z ,
which will imply that int(Z) = Z.
Suppose that Z is not open. Then there exists x ∈ Z such that B(x, n−1) * Z for all n ∈ N.
Hence, for each n ∈ N, there exists yn ∈ X such that
d(x, yn) <
1
n
and d(yn, X r Z) = 0 .
Then, for every n ∈ N, there exists wn ∈ X rZ such that d(yn, wn) < 1/n . By (1.28), d(x, wn) → 0,
implying d(x, X r Z) = 0, in contradiction to the hypothesis that x ∈ Z.
The proof of (ii) is simple. If x /∈ Z, then x ∈ X r Z, so that d(x, X r Z) = 0, that is, x /∈ Z.
To prove (iii), suppose that x ∈ int(Z). Then there exists r > 0 such that B(x, r) ⊆ int(Z) ⊆ Z.
But then, for any y ∈ X r Z, d(x, y) ≥ r > 0, that is, x ∈ Z.
The proof of the formula for closure is based on the equality
cl(X r Y ) = X r int(Y ) ,
valid for any subset Y of X. Then
cl(Z) = X r int(X r Z) = X r {x ∈ X : d(x, Z) > 0}
= {x ∈ X : d(x, Z) = 0} .
(cid:3)
Proposition 1.25 has some important consequences.
Corollary 1.26. Let (X, d) be an f -quasimetric space.
of x.
1. For every x ∈ X and r > 0, x ∈ int(cid:0)B(x, r)(cid:1), or, equivalently, B(x, r) is a neighborhood
2. The topology τd satisfies the first axiom of countability, i.e. every point has a countable
base of neighborhood. Consequently, usual sequences suffices to characterize the topological
properties of X.
3. The convergence of a sequence (xn) in X to x ∈ X with respect to τd is characterized in
the following way:
(1.30)
xn
τd−→ x ⇐⇒ d(x, xn) −→ 0 .
Proof. 1. This follows from the following relations
y ∈ X r B(x, r) ⇐⇒ d(x, y) ≥ r
2. A countable base of neighborhoods of a point x ∈ X is B(x, n−1), n ∈ N. If V is a neighbor-
=⇒ d(x, X r B(x, r)) ≥ r > 0
⇐⇒ x ∈ int(cid:0)B(x, r)(cid:1) .
hood of x, then there exists G ∈ τd such that
x ∈ G ⊆ V .
By the definition of the topology τd, there exists r > 0 such that B(x, r) ⊆ G, implying
B(x, n−1) ⊆ B(x, r) ⊆ G ⊆ V ,
for some sufficiently large n ∈ N.
and
Here, by definition,
(1.32)
lim sup
t2ց0
t2ց0
θ(r) =(sup Λ(r)
r
if Λ(r) 6= ∅
if Λ(r) = ∅ .
f (t1, t2) = inf
δ>0(cid:0) sup{f (t1, t2) : 0 ≤ t2 < δ}(cid:1) .
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
13
3. Suppose that xn
τd−→ x and let r > 0. Then, by 1, B(x, r) ∈ V(x) so there exists nr ∈ N such
that xn ∈ B(x, r), or, equivalently, d(x, xn) < r for all n ≥ nr. Consequently, d(x, xn) −→ 0.
Suppose now that d(x, xn) −→ 0 and let V ∈ V(x). Then x ∈ int(V ) ∈ τd, so there exists r > 0
such that B(x, r) ⊆ int(V ) ⊆ V. By hypothesis, there exists n0 ∈ N, such that d(x, xn) < r for all
n ≥ n0, implying xn ∈ B(x, r) ⊆ V for all n ≥ n0.
(cid:3)
Remark 1.27. The convergence with respect to the topology τ ¯d generated by the conjugate f -
quasimetric ¯d(x, y) = d(y, x), x, y ∈ X, is characterized by
(1.31)
xn
τ ¯d−→ x ⇐⇒ d(xn, x) −→ 0 .
Since x ∈ int(cid:0)B(x, r)(cid:1), there exists r′ > 0 such that B(x, r′) ⊆ int(cid:0)B(x, r)(cid:1) ⊆ B(x, r). A
natural question is to find the biggest r′ such that B(x, r′) int(cid:0)B(x, r)(cid:1).
For r > 0 let
Λ(r) = {t1 ≥ 0 : lim sup
f (t1, t2) ≥ r} ,
Remark 1.28. Observe that Λ(r) = ∅ implies B(x, r) = X.
Also θ(r) > 0 in both cases.
Indeed, if Λ(r) = ∅, then putting t1 = d(x, y) for some arbitrary y ∈ X, it follows lim supt2ց0 f (t1, t2) <
r, so there exists δ > 0 such that sup{f (t1, t2) : 0 ≤ t2 < δ} < r, implying
that is, y ∈ B(x, r).
d(x, y) ≤ f (d(x, y), d(y, y)) = f (t1, 0) < r ,
If θ(r) = 0, then Λ(r) 6= ∅, so there exists a sequence (tk
(1.32) and the definition of Λ(r), for every k ∈ N there exists 0 ≤ tk
Taking into account (1.26), one obtains the contradiction.
1) in Λ(r) such that tk
2 < 1/k such that f (tk
1 → 0 as k → ∞. By
2) > r/2.
1, tk
0 = lim
k→∞
f (tk
1, tk
2) ≥ r/2 .
Proposition 1.29. Let (X, d) be an f -quasimetric space. Then
for every x ∈ X and r > 0.
B(x, θ(r)) ⊆ int(cid:0)B(x, r)(cid:1) ,
Proof. By Remark 1.28 we can suppose Λ(r) 6= ∅.
Let y ∈ B(x, θ(r)). Then t1 := d(x, y) < θ(r), implying t1 /∈ Λ(r) so that lim supt2ց0 f (t1, t2) <
r. By (1.32) there exists δ > 0 such that
(1.33)
We show that
(1.34)
sup{f (t1, t2) : 0 ≤ t2 < δ} < r .
B(y, δ) ⊆ B(x, r) .
If z ∈ B(y, δ), then, by (1.33)
d(x, z) ≤ f (d(x, y), d(y, z)) = f (t1, d(y, z)) < r ,
because d(y, z) < δ.
Since B(y, δ) is a neighborhood of y, the inclusion (1.34) shows that B(x, r) is also a neighbor-
(cid:3)
hood of y, that is, y ∈ int(cid:0)B(x, r)(cid:1).
14
S. COBZAS¸
Remark 1.30. If (X, d) is an (s1, s2)-quasimetric space, i.e. an f -quasimetric space for f (t1, t2) =
s1t1 + s2t2, then
θ(r) = r/s1 .
Indeed,
θ(r) = inf{t1 ≥ 0 : lim
t2ց0
(s1t1 + s2t2) ≥ r}
= inf{t1 ≥ 0 : s1t1 ≥ r} = r/s1 .
The authors define in [14] the notion of Cauchy sequence and completeness. A sequence in a
distance space (X, d) is called Cauchy if for every ε > 0 there exists n0 = n0(ε) such that
(1.35)
d(xn, xn+k) < ε ,
for all n ≥ n0 and all k ∈ N. The distance space (X, d) is called complete if every Cauchy sequence
is convergent to some x ∈ X.
The authors prove in [14] the validity of Baire category theorem in complete f -quasimetric spaces
which satisfy the separation axiom T3 (i.e. are regular). As in the metric case, the proof is based on
the nonemptiness of descending sequences of closed balls with radii tending to 0. They extend the
metrization Theorems 1.3 and 1.4 to this setting proving the quasimetrizability of f -quasimetric
spaces. In this case, the equivalent quasimetrics are also given by the formulae (1.3) and (1.4) and
are denoted by Inf d. They introduce the notion of weak symmetry of the f -quasimetric d by the
condition
(1.36)
d(x, xn) → 0 =⇒ d(xn, x) → 0 ,
for all sequences (xn) in X and x ∈ X. The topology generated by a weakly symmetric f -
quasimetric is normal and metrizable.
Remark 1.31. By (1.30) and (1.31), the condition (1.36) means that the identity mapping I :
(X, τd) → (X, τ ¯d) is continuous, or equivalently, τ ¯d ⊆ τd, i.e. the topology τd is finer than τ ¯d.
Remark 1.32. In the theory of quasimetric spaces (see Example 1.24) a sequence satisfying (1.35)
is called "left K-Cauchy" and the corresponding notion of completeness, "left K-completeness".
If the sequence (xn) satisfies the condition
(1.37)
d(xn+k, xn) < ε ,
for all n ≥ n0 and all k ∈ N, then it is called "right K-Cauchy", and the corresponding notion of
completeness, "right K-completeness".
Some authors call a sequence (xn) satisfying (1.35) "forward Cauchy" and "backward Cauchy"
if it satisfies (1.37). Also the convergence given by d(x, xn) → 0 is called "backward convergence",
while that given by d(xn, x) → 0 is called "forward convergence" (see, e.g.
[61]). Combining
these notions of Cauchy sequence and convergence one obtains various notions of completeness:
"forward-forward complete" meaning that every forward Cauchy sequence is forward convergent,
with similar definitions for forward-backward, backward-forward, etc -- completeness.
Due to the asymmetry of the quasimetric, there are several notions of Cauchy sequence (actually
7, see [68]), each of them agreeing with the usual notion of Cauchy (fundamental) sequence in
the metric case. Considering d-convergence and ¯d-convergence, from these 7 notions of Cauchy
sequence one obtains 14 notions of completeness (see the book [23]).
1.6. Historical remarks and further results. The relaxed triangle inequality and the corre-
sponding spaces were rediscovered several times under various names -- quasi-metric, near metric
(in [32]), metric type, etc.
• (1970) Coifman and de Guzman [25] in connection with some problems in harmonic analysis
(a b-metric is called "distance" function);
• (1979) the results of Coifman and de Guzman were completed by Mac´ıas and Segovia
[57, 58];
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
15
• (1989) Bakhtin [17] called them "quasi-metric spaces" and proved a contraction principle
for such spaces;
• (1993) Czerwik introduced them under the name "b-metric space", first for s = 2 in [26],
and then for an arbitrary s ≥ 1 in [27], with applications to fixed points;
• (1998,2003) Fagin et al. [37, 38] considered distances satisfying the s-relaxed triangle and
polygonal inequalities with applications to some problems in theoretical computer science;
• (2010) Khamsi [50] introduced them under the name "metric type spaces" and remarked
that if D is a cone metric on a set X with values in a Banach space ordered by a normal
cone C with normality constant K, then d(x, y) = kD(x, y)k, x, y ∈ X, is a b-metric on X
satisfying the K-relaxed polygonal inequality.
Some topological properties of b-metric spaces (e.g. compactness) were studied in [52]. Xia
[76] studied the properties of the space C(T, X) of continuous functions from a compact metric
space T to a b-metric space X, and geodesics and intrinsic metrics in b-metric spaces. The results
were applied to show that the optimal transport paths between atomic probability measures are
geodesics in the intrinsic metric. An, Tuyen and Dung [11] extended to b-metric spaces Stone's
paracompactness theorem.
2. Generalized b-metric spaces
The notions of generalized metric, meaning a mapping d : X × X → [0, ∞] satisfying the axioms
of a metric, and generalized metric space (X, d) were introduced by W. A. J. Luxemburg in [55] -- [56]
in connection with the method of successive approximation and fixed points. These results were
completed by A. F. Monna [62] and M. Edelstein [35]. Further results were obtained by J. B. Diaz
and B. Margolis [34, 59] and C. F. K. Jung [43]. G. Dezso [33] considered generalized vector metrics,
i.e. metrics with values in Rm
+ ∪{(+∞)m}, and extended to this setting Perov's fixed point theorem
(see [65] -- [66]) as well as other fixed point results (Luxemburg, Jung, Diaz-Margolis, Kannan).
For some recent results on generalized metric spaces see [18] and [28]. Recently, G. Beer and J.
Vanderwerf [19] -- [20] considered vector spaces equipped with norms that can take infinite values,
called "extended norms" (see also [30]).
Following these ideas, we consider here the notion of generalized b-metric on a nonempty set X
as a mapping d : X × X → [0, ∞] satisfying the conditions (i) -- (iii) from (1.1). If d satisfies further
the condition (v), then d is called a generalized strong b-metric and the pair (X, d) a generalized
strong b-metric space.
Let (X, d) be a generalized b-metric space. As in Jung [43], it follows that
(2.1)
x ∼ y
d⇐⇒ d(x, y) < +∞,
x, y ∈ X,
is an equivalence relation on X. Denoting by Xi, i ∈ I, the equivalence classes corresponding to ∼
and putting di = dXi×Xi , i ∈ I, then (Xi, di) is a b-metric space (a strong b-metric space if (X, d)
is a generalized strong b-metric space) for every i ∈ I. Therefore, X can be uniquely decomposed
into equivalence classes Xi, i ∈ I, called the canonical decomposition of X.
By analogy to [43] we have.
Theorem 2.1. Let (X, d) be a generalized b-metric space and Xi, i ∈ I, its canonical decomposi-
tion. Then the following hold.
1. The space (X, d) is complete if and only if (Xi, di) is complete for every i ∈ I.
2.
If (Yi, di), i ∈ I, are b-metric spaces (with the same s) and Yi ∩ Yj = ∅ for all i 6= j in I,
then
(2.2)
d(x, y) :=
di(x, y)
+∞
if x, y ∈ Yi, for some i ∈ I,
if x ∈ Yi and y ∈ Yj
for some i, j ∈ I with i 6= j,
is a generalized b-metric on Y =Si∈I Yi, with {Yi : i ∈ I} the family of equivalence classes
corresponding to the equivalence relation (2.1).
16
S. COBZAS¸
The same results are true for generalized strong b-metric spaces.
2.1. The completion of generalized b-metric spaces. In this subsection we shall prove the
existence of the completion of strong b-metric spaces. The existence of the completion of a gener-
alized metric space was proved in [29].
We start with the following lemma.
Lemma 2.2. Let (X, d) be a generalized b-metric space, (Z, D) a complete generalized b-metric
space, with continuous generalized b-metrics d, D and Y a dense subset of X. Then for every
isometric embedding f : Y → Z there exists a unique isometric embedding F : X → Z such that
F Y = f . If, in addition, X is complete and f (Y ) is dense in Z, then F is bijective (i.e. F is an
isometry of X onto Z).
Proof. For the sake of completeness we include the simple proof of this result. For x ∈ X let
(yn) be a sequence in Y such that d(yn, x) → 0. Then (yn) is a Cauchy sequence in (X, d) and
the equalities D(f (yn), f (ym)) = d(yn, ym), m, n ∈ N, show that (f (yn)) is a Cauchy sequence in
(Z, D). Since (Z, D) is complete, there exists z ∈ Z such that D(f (yn), z) → 0. If (y′
n) is another
sequence in Y converging to x, then (f (y′
n)) will converge to an element z′ ∈ Z. By the continuity
of the generalized b-metrics d and D,
D(z, z′) = D(lim
n
f (yn), lim
n
f (y′
n)) = lim
n
D(f (yn), f (y′
n)) = lim
n
d(yn, y′
n) = 0 ,
showing that z = z′. So we can unambiguously define a mapping F : X → Z by F (x) = limn f (yn),
where (yn) is a sequence in Y converging to x ∈ X. For y ∈ Y taking yn = y, n ∈ N, it follows
F (y) = y.
For x, x′ ∈ X, let (yn), (y′
n) be sequences in Y converging to x and x′, respectively. Then
D(F (x), F (x′)) = lim
n
D(f (yn), f (y′
n)) = lim
n
d(yn, y′
n) = d(x, x′) ,
i.e. F is an isometric embedding.
If f (Y ) is dense in Z, then, for any z ∈ Z, there exists a sequence (yn) in Y such that
D(f (yn), z) → 0. It follows that (f (yn)) is a Cauchy sequence in Z and so, as f is an isome-
try, (yn) will be a Cauchy sequence in X. As the space X is complete, (yn) is convergent to some
x ∈ X. But then
D(F (x), z) = lim
n
D(F (x), f (yn)) = lim
n
d(x, yn) = 0 ,
showing that F (x) = z.
(cid:3)
Remark 2.3. The proof can be adapted to show that, under the hypotheses of Lemma 2.2, every
uniformly continuous mapping f : Y → Z has a unique uniformly continuous extension to X. The
notion of uniform continuity for mappings between generalized b-metric spaces is defined as in the
metric case.
Let (X, d) be a generalized strong b-metric space with Xi, i ∈ I, the family of equivalence classes
corresponding to (2.1). For every i ∈ I, let (Yi, Di) be a completion of the strong b-metric space
(Xi, di). Denote by Ti : (Xi, di) → (Yi, Di) the isometric embedding with Ti(Xi) Di-dense in Yi
corresponding to this completion.
Replacing, if necessary, Yi with Yi = Yi × {i}, Di with Di((x, i), (y, i)) = Di(x, y), for x, y ∈ Yi,
and putting Ti(x, i) = (Ti(x), i), x ∈ Yi, we may suppose, without restricting the generality, that
Yi ∩ Yj = ∅ for all i, j ∈ I with i 6= j .
Put Y :=Si∈I Yi, and define
according to (2.2) and T : X → Y by
D : Y × Y → [0, ∞]
T (x) := Ti(x),
where i is the unique element of I such that x ∈ Xi.
We have the following result.
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
17
Theorem 2.4. Let (X, d) be a generalized strong b-metric space and (Y, D) the generalized strong
b-metric space defined above. Then
(Y, D) is a complete generalized strong b-metric space;
(i)
(ii) T : (X, d) → (Y, D) is an isometric embedding with T (X) D-dense in Y ;
(iii) any other complete generalized strong b-metric space (Z, ) that contains a ρ-dense isomet-
ric copy of (X, d), is isometric to (Y, D).
Proof. Since each strong b-metric space (Yi, Di) is complete, Theorem 2.1 implies that the gener-
alized strong b-metric space (Y, D) is complete.
Let x, y ∈ X. If x, y ∈ Xi, for some i ∈ I, then
D(T (x), T (y)) = Di(Ti(x), Ti(y)) = di(x, y) = d(x, y).
If x ∈ Xi, y ∈ Xj with i 6= j, then
so that
T (x) = Ti(x) ∈ Yi and T (y) = Tj(x) ∈ Yj ,
D(T (x), T (y)) = D(Ti(x), Tj(y)) = +∞ = d(x, y).
Now for ξ ∈ Y there exists a unique i ∈ I such that ξ ∈ Yi. Since Ti(Xi) is dense in (Yi, Di), there
exists a sequence (xn) in Xi such that
0 = lim
n→∞
Di(Ti(xn), ξ) = lim
n→∞
D(T (xn), ξ),
which means that T (X) is D-dense in (Y, D).
Finally, to verify (iii), let S : (X, d) → (Z, ρ) be an isometric embedding with S(X) dense in Z.
Define R : T (X) → X by R(T (x)) = x, x ∈ X. Then R is an isometry of T (X) onto X and S ◦ R
is an isometric embedding of T (X) into Z. Since T (X) is dense in Y and S(R(T (X))) = S(X) is
dense in Z, Lemma 2.2 yields the existence of an isometry U of Y onto Z, which ends the proof. (cid:3)
We shall prove some fixed point results in b-metric and in generalized b-metric spaces.
3. Fixed points in b-metric spaces
3.1. Fixed points in b-metric spaces. We start with the case of b-metric spaces. The second
result is an extension to b-metric spaces of Theorem 4.1 from [41].
Let (X, d) be a b-metric space with d satisfying the s-relaxed triangle inequality. We consider
functions ϕ : R+ → R+ satisfying the conditions
(3.1)
(a) ϕ is nondecreasing and
(b)
lim
n→∞
ϕn(t) = 0 for all t > 0 .
Remark 3.1. If ϕ : R+ → R+ satisfies the conditions (a) and (b) from above, then
(c) ϕ(t) < t for all t > 0;
(d)
lim
n→∞
ϕn(0) = 0 and ϕ(0) = 0 = lim
tց0
ϕ(t) .
Indeed, if ϕ(t) ≥ t for some t > 0, then, by (a), ϕ2(t) ≥ ϕ(t) ≥ t and, in general ϕn(t) ≥ t > 0
for all n, in contradiction to (b).
Also, 0 ≤ ϕ(0) ≤ ϕ(1) implies 0 ≤ ϕ2(0) ≤ ϕ2(1) and in general 0 ≤ ϕn(0) ≤ ϕn(1). Since
limn→∞ ϕn(1) = 0, this yields (d).
Similarly, 0 ≤ ϕ(0) ≤ ϕ(t) < t for any t > 0, implies ϕ(0) = 0 = limtց0 ϕ(t).
18
S. COBZAS¸
Theorem 3.2. Let (X, d) be a complete b-metric space, where d satisfies the s-relaxed triangle
inequality and let ϕ : R+ → R+ be a function satisfying the conditions (a), (b) from (3.1). Then
every mapping f : X → X satisfying the inequality
(3.2)
d(f (x), f (y)) ≤ ϕ(d(x, y)) ,
for all x, y ∈ X, has a unique fixed point z and, for every x ∈ X, the sequence (cid:0)f n(x)(cid:1)n∈N0
converges to z as n → ∞.
Proof. We present the proof given in [44]. Let x ∈ X. Put xn = T nx, n ∈ N0 , and let us show
that (xn)n∈N0 is a Cauchy sequence.
Observe first that (3.2) implies
(3.3)
d(T nu, T nv) ≤ ϕn(d(u, v)) ,
for all u, v ∈ X and n ∈ N. This implies
(3.4)
d(T nxkn, xkn) ≤ ϕkn(d(xn, x0)) ,
for all k, n ∈ N. Indeed, by (3.3),
d(T nxkn, xkn) = d(T knxn, T knx0) ≤ ϕkn(d(xnx0)) .
From (3.4) and (3.1).(b) one obtains
(3.5)
for every n ∈ N.
lim
k→∞
d(T nxkn, xkn) = 0 ,
Let ε > 0 be given.
Observe first that there exist ¯k, ¯n ∈ N s.t.
(3.6)
(i) T ¯n(B(x¯k ¯n, ε) ⊆ B(x¯k ¯n, ε);
(ii) xk¯n ∈ B(x¯k ¯n, ε) for all k ≥ ¯k;
(iii) d(xk1 ¯n, x¯k2 ¯n) < 2sε for all k1, k2 ≥ ¯k.
Indeed, by (3.1).(b), there exists ¯n ∈ N s.t.
(3.7)
and, by (3.5), there exists ¯k ∈ N s.t.
ϕn(ε) < ε/(2s) for all n ≥ ¯n,
(3.8)
d(T ¯nxk¯n, xk¯n) < ε/(2s) for all k ≥ ¯k .
But then, d(u, x¯k¯n) < ε implies
d(T ¯nu, T ¯nx¯k¯n) ≤ ϕ¯n(d(u, x¯k ¯n)) ≤ ϕ¯n(ε) < ε/(2s) ,
so that
d(T ¯nu, x¯k¯n) ≤ s(cid:2)d(T ¯nu, T ¯nx¯k¯n) + d(T ¯nx¯k¯n, x¯k¯n)(cid:3)
ε
+
< sh ε
2s
2si = ε ,
showing that (3.6).(i) holds.
Since x¯k¯n ∈ B(x¯k¯n, ε) it follows that x(¯k+1)¯n = T ¯nx¯k¯n ∈ B(x¯k¯n, ε) and, in general, by induction,
x(¯k+j)¯n ∈ B(x¯k ¯n, ε) for any j ∈ N0.
Now, by (ii), xk1 ¯n, xk2 ¯n ∈ B(x¯k ¯n, ε) for all k1, k2 ≥ ¯k, so that
d(xk1 ¯n, xk2 ¯n) ≤ s(d(xk1 ¯n, x¯k¯n) + d(x¯k¯n, xk2 ¯n) < 2sε ,
showing that (iii) holds too.
By (3.5) for n = 1 one obtains
lim
k→∞
d(xk+1, xk) = 0 .
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
19
It is easy to check that this implies
lim
k→∞
d(xk¯n+p, xk¯n) = 0 for p = 0, 1, . . . , ¯n − 1 ,
so there exists k0 ∈ N s.t.
(3.9)
d(xk¯n+p, xk¯n) < ε for all k ≥ k0 and p = 0, 1, . . . , ¯n − 1 .
Let now k := max{¯k, k0} and let m1 = k1 ¯n + p1, m2 = k2¯n + p2 with p1, p2 ∈ {0, 1, . . . , ¯n − 1}
and k1, k2 ≥ k.
Combining (3.9) and (3.6).(iii) one obtains
d(xm1 , xm2) ≤ sd(xk1 ¯n+p1 , xk1 ¯n) + s2d(xk1 ¯n, xk2 ¯n) + s3d(xk2 ¯n, xk2 ¯n+p2 )
< (s + 2s3 + s2)ε ≤ 4s3ε ,
which shows that (xn) is a Cauchy sequence.
The completeness of X implies the existence of a point z ∈ X s.t. limn→∞ d(xn, z) = 0.
We have
d(xn+1, T z) = d(T xn, T z) ≤ ϕ(d(xn, z)) ≤ d(xn, z)
for all n ∈ N, so that
d(z, T z) ≤ s [d(z, xn+1) + d(xn+1, T z)]
≤ s [d(z, xn+1) + d(xn, z)] .
Letting n → ∞ one obtains d(z, T z) = 0, that is, z = T z.
The uniqueness follows in the following way. Suppose zi = T zi, i = 1, 2. Then
d(z1, z2) = d(T z1, T z2) ≤ ϕ(d(z1, z2)) .
By Remark 3.1.(c) this can hold only for d(z1, z2) = 0, that is, for z1 = z2.
(cid:3)
Let (X, d) be a b-metric space with d satisfying the s-relaxed triangle inequality for some s ≥ 1.
An important particular case of a function ϕ satisfying the conditions (a),(b) from (3.1) is
If 0 < α < 1, then
ϕ(t) = αt, t ≥ 0 .
Since ϕ is also strictly increasing, it satisfies the conditions (a),(b) from (3.1).
The inequality (3.2) becomes in this case
ϕn(t) = αnt → 0 as n → ∞ .
d(f (x), f (y)) ≤ αd(x, y) ,
for all x, y ∈ X.
So, Theorem 3.2 has as consequence the analog of Banach contraction principle in b-metric
spaces. The following proposition also illustrates how various types of relaxed inequalities for the
b-metric influence the form this principle takes.
Proposition 3.3. Let (X, d) be a complete b-metric space, where d satisfies the s-relaxed triangle
inequality and f : X → X a mapping such that, for some 0 < α < 1,
(3.10)
d(f (x), f (y)) ≤ αd(x, y),
for all x, y ∈ X. Then f has a unique fixed point z and, for every x ∈ X, the sequence(cid:0)f n(x)(cid:1)n∈N
converges to z as n → ∞.
1. ([17]) If further 0 < α < 1/s, then the following evaluation of the order of convergence holds
(3.11)
for all n ∈ N.
d(xn, z) ≤
s2d(x0, x1)
1 − αs
· αn ,
20
S. COBZAS¸
2. ([53]) If d satisfies the s-relaxed polygonal inequality, then the following evaluation of the order
of convergence
(3.12)
d(xn, z) ≤
s2d(x0, x1)
1 − α
· αn, n ∈ N ,
holds for any 0 < α < 1.
Proof. Although, as we have remarked, the first statement of the proposition follows from Theorem
3.2, we show a proof based on Theorem 1.4. Our presentation follows [10].
Suppose that d satisfies the s-relaxed triangle inequality, for some s ≥ 1. If 0 < p ≤ 1 is given
by the equation (2s)p = 1, then, by Theorem 1.4, the functional ρp given by (1.4) is a metric on
X satisfying the inequalities
(3.13)
ρp ≤ dp ≤ 2ρp .
For x, y ∈ X let x = x0, x1, . . . , xn = y be an arbitrary chain in X connecting x and y. Then
yi = f (xi), i = 0, 1, . . . , n, is a chain in X connecting f (x) and f (y). Consequently, by (1.4) and
(3.10),
(3.14)
ρp(f (x), f (y)) ≤
d(yi, yi+1)p ≤ αp
n−1Xi=0
d(xi, xi+1)p.
n−1Xi=0
Since the inequality between the extreme terms in (3.14) holds for all chains x = x0, x1, . . . , xn =
y, n ∈ N, connecting x and y, it follows
ρp(f (x), f (y)) ≤ αpρp(x, y) ,
for all x, y ∈ X, where 0 < αp < 1. Consequently, f is a contraction with respect to ρp. The
inequalities (3.13) and the completeness of (X, d) imply the completeness of (X, ρp) and so, by
Banach's contraction principle, f has a unique fixed point z ∈ X and the sequence of iterates
(f n(x))n∈N is ρp-convergent to z, for every x ∈ X. Appealing again to the inequalities (3.13), it
follows that (f n(x))n∈N is also d-convergent to z for every x ∈ X.
1. The proof is similar to that of Banach's contraction principle in the metric case. Observe
first that, (3.10) implies
(3.15)
d(f n(x), f n(y)) ≤ αnd(x, y) ,
for all n ∈ N and x, y ∈ X.
For x0 ∈ X consider the sequence of iterates
xn = f (xn−1) = f n(x0), n ∈ N .
Let us prove that (xn) is a Cauchy sequence.
By (1.2) and (3.15),
d(xn, xn+k+1) ≤ sd(xn, xn+1) + s2d(xn+1, xn+2) + . . .
(3.16)
+ skd(xn+k−1, xn+k) + skd(xn+k, xn+k+1)
≤(cid:0)αns + αn+1s2 + · · · + αn+k−1sk(cid:1)d(x0, x1) + αn+kskd(x0, x1)
= αns(cid:18) 1 − (αs)k
+ αksk−1(cid:19) d(x0, x1)
1 − αs
1 − (αs)k−1α(s + 1 − αs)
= αns
d(x0, x1)
1 − αs
< αn sd(x0, x1)
1 − αs
,
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
21
for all n, k ∈ N. Since limn→∞ αn = 0, this shows that (xn) is a Cauchy sequence. By the
completeness of (X, d) there exists z ∈ X such that limn→∞ d(xn, z) = 0. We have
d(z, f (z)) ≤ sd(z, xn+1) + sd(xn+1, f (z))
≤ sd(z, xn+1) + sαd(xn, z) −→ 0 as n → ∞ .
Hence d(z, f (z)) = 0 and so z = f (z).
Taking into account (3.16),
d(xn, z) ≤ sd(xn, xn+k+1) + sd(xn+k+1, z)
< αn s2d(x0, x1)
1 − αs
+ sd(xn+k+1, z) .
Letting k → ∞, one obtains (3.11).
Suppose now that there exists two points z, z′ ∈ X such that f (z) = z and f (z′) = z′. Then
the relations
d(z, z′) = d(f (z), f (z′)) ≤ αd(z, z′)
show that d(z, z′) = 0, i.e. z = z′.
2. Let x0 ∈ X and xn = f (xn−1), n ∈ N. Taking into account the relaxed polygonal inequality
and (3.15), we obtain
d(xn, xn+k) ≤ s
k−1Xi=0
d(xn+i, xn+i+1) ≤ s(αn + αn+1 + · · · + αn+k)d(x0, x1)
= sαn 1 − αk+1
1 − α
· d(x0, x1) <
sd(x0, x1)
1 − α
· αn.
Based on these relations the proof goes as in case 1.
(cid:3)
Remark 3.4. The proof given here to statement 2 from Proposition 3.3 is simpler than that of
Theorem 12.4 in [53].
Remark 3.5. The proofs given in [26] and [53] to Theorem 3.2 go in the following way.
Let x be a fixed element of X and ε > 0. By (3.1).(b) there exists m = mε ∈ N such that
(3.17)
ϕm(ε) <
ε
2s
.
One considers the sequence xk = f km(x), k ∈ N, and one shows that there exists k0 ∈ N such
that
(3.18)
d(xk, xk′ ) < 2sε ,
for all k, k′ ≥ k0. One affirms that the inequality (3.18) shows that (xk) is a Cauchy sequence,
which is not surely true, because the inequality is true only for this specific ε.
Taking another ε, say 0 < ε′ < ε, we find another number m′ = mε′ (possibly different from
m), such that
(3.19)
ϕm′
(ε′) <
ε′
2s
.
The above procedure yields a sequence x′
k = f km′
(x), k ∈ N, satisfying, for some k1 ∈ N,
(3.20)
d(xk, xk′ ) < 2sε′ ,
for all k, k′ ≥ k1.
(xk) is Cauchy.
But the sequences (xk) and (x′
k) can be totally different, so we cannot infer that the sequence
As we have shown this flaw was fixed in the paper [44].
22
S. COBZAS¸
0 < α < 1/s, namelyP∞
Remark 3.6. Berinde [21] considers comparison functions satisfying a condition stronger than
k=1 ϕk(t) < ∞, allowing estimations of the order of convergence similar to
(3.11). He also shows that the sequence xn = f n(x0), n ∈ N0, is convergent to a fixed point of f if
and only if it is bounded. For various kinds of comparison functions, the relations between them
an applications to fixed points, see [71, §3.0.3].
3.2. Fixed points in generalized b-metric spaces. Theorem 3.2 admits the following extension
to generalized b-metric spaces.
Theorem 3.7. Let (X, d) be a complete generalized b-metric space and suppose that the mapping
f : X → X is such that
(3.21)
d (f (x), f (y)) ≤ ϕ (d(x, y)) ,
for all x, y ∈ X with d(x, y) < ∞, where the function ϕ : R+ → R+ satisfies the conditions (a),(b)
from (3.1).
Consider, for some x ∈ X, the sequence of successive approximations (f n(x))n∈N0. Then either
(A)
d(f k(x), f k+1(x)) = +∞ for all k ∈ N0,
or
(B)
the sequence (f n(x))n∈N is convergent to a fixed point of f .
Proof. Let X = Si∈I Xi be the canonical decomposition of X corresponding to the equivalence
relation (2.1). Assume that (A) does not hold. Then
d(f m(x), f m+1(x)) < +∞ ,
for some m ∈ N0 . If i ∈ I is such that f m(x), f m+1(x) ∈ Xi, then
implies f m+2(x) ∈ Xi , and so, by mathematical induction, f m+k(x) ∈ Xi for all k ∈ N0 .
d(cid:0)f m+1(x), f m+2(x)(cid:1) ≤ ϕ(cid:0)d(cid:0)f m(x), f m+1(x)(cid:1)(cid:1) < ∞ ,
Since
the inequality
z ∈ Xi ⇐⇒ d(z, f m(x)) < ∞ ,
d(f (z), f m+1(x)) ≤ ϕ(d(z, f m(x)) < ∞ ,
shows that the restriction fi = f Xi of f to Xi is a mapping from Xi to Xi satisfying
d(fi(y), fi(z)) ≤ ϕ(d(y, z)) ,
for all y, z ∈ Xi.
By Theorem 2.1, Xi is complete, so that, by Theorem 3.2, (cid:0)f m+k(x)(cid:1)k∈N0
fixed point zi ∈ Xi of fi, which is a fixed point for f .
is convergent to a
(cid:3)
Remark 3.8. For s = 1 and ϕ(t) = αt, t ≥ 0, where 0 ≤ α < 1, we get the Diaz and Margolis [34]
fixed point theorem of the alternative. At the same time this extends Theorem 2 from [31] and
give simpler proofs to Theorems 2.1 and 3.1 from [16].
Proposition 3.3 also admits extensions to this setting as results of the alternative. We formulate
only one of these results.
Corollary 3.9. Let (X, d) be a complete b-metric space, where d satisfies the s-relaxed triangle
inequality and let f : X → X be a mapping satisfying, for some 0 < α < 1, the inequality
d(f (x), f (y)) ≤ αd(x, y) ,
for all x, y ∈ X with d(x, y) < ∞.
Then, for every x ∈ X, either
(A′)
d(f k(x), f k+1(x)) = +∞ for all k ∈ N0,
or
(B′)
the sequence (f n(x))n∈N is convergent to a fixed point of f .
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
23
4. Lipschitz functions
In this section we shall discuss the behavior of Lipschitz functions defined on or taking values
in quasi-normed spaces and of Lipschitz functions on spaces of homogeneous type.
4.1. Quasi-normed spaces. We start by a brief presentation of an important class of b-metric
spaces -- quasi-normed spaces. Good references are [47], [49], [54, pp. 156-166], [70].
A quasi-norm on a vector space X (over K equal to R or C) is a functional k · k : X → R+ for
which there exists a real number k ≥ 1 so that
(QN1) kxk = 0 ⇐⇒ x = 0;
(QN2) kαxk = α kxk;
(QN3) kx + yk ≤ k(kxk + kyk),
for all x, y ∈ X and α ∈ K. The pair (X, k · k) is called a quasi-normed space. A complete
quasi-normed space is called a quasi-Banach space.
If k = 1, then k · k is a norm. The smallest constant k for which the inequality (QN3) is satisfied
for all x, y ∈ X is called the modulus of concavity of the quasi-normed space X.
For a linear operator T from a quasi-normed space (X, k · kX) to a normed space (Y, k · kY ) put
kT k = sup{kT xkY : x ∈ X, kxkX ≤ 1} .
In particular,
(4.1)
is a norm on the dual space X ∗ = (X, k · kX)∗.
kx∗k = sup{x∗(x) : x ∈ X, kxkX ≤ 1}, x∗ ∈ X ∗,
It follows that T is continuous if and only if kT k < ∞ and, in this case,
kT xkY ≤ kT kkxkX , x ∈ X,
kT k being the smallest number L ≥ 0 for which the inequality kT xkY ≤ LkxkX holds for all
x ∈ X. If Y is also a quasi-normed space, then k · k is only a quasi-norm on the space L(X, Y ) of
all continuous linear operators from X to Y .
Two quasi-norms k · k1, k · k2 on a vector space X are called equivalent if they generate the same
topology, or equivalently, if
for all sequences (xn) in X and x ∈ X. As in the case of norms, the equivalence of two quasi-norms
k · k1, k · k2 on a vector space X is equivalent to the existence of two numbers α, β > 0 such that
kxn − xk1 → 0 ⇐⇒ kxn − xk2 → 0 ,
αkxk1 ≤ kxk2 ≤ βkxk1 ,
for all x ∈ X.
A subset A of a topological vector space (TVS) (X, τ ) is called bounded if it is absorbed by any
0-neighborhood, i.e for every V ∈ Vτ (0) there exists t > 0 such that A ⊆ tV. A TVS is called locally
bounded if it has a bounded 0-neighborhood. A quasi-normed space (X, k · k) is locally bounded,
as the closed unit ball BX = {x ∈ X : kxk ≤ 1} is a bounded neighborhood of 0. One shows that,
conversely, the topology of every locally bounded TVS is generated by a quasi-norm.
A quasi-normed space (X, k · k) is normable (i.e. there exists a norm k · k1 on X equivalent
to the quasi-norm k · k) if and only if 0 has a bounded convex neighborhood (implying that X is
locally convex).
Definition 4.1. An F -norm on a vector space X is a mapping k · k : X → R+ satisfying the
conditions
(F1) kxk = 0 ⇐⇒ x = 0;
(F2) kλxk ≤ kxk for all λ ∈ K with λ ≤ 1;
(F3) kx + yk ≤ kxk + kyk;
(F4) kxnk → 0 =⇒ kλxnk → 0;
(F5) λn → 0 =⇒ kλnxk → 0,
24
S. COBZAS¸
for all x, y, xn ∈ X λ, λn ∈ K. An F -space is a vector space equipped with a complete F -norm.
It follows that d(x, y) = ky − xk, x, y ∈ X, is a translation-invariant metric on X defining a
vector topology. It is known that the metrizability of a TVS (X, τ ) is equivalent to the existence
of a countable basis of 0-neighborhoods, and in this case there exists a translation-invariant metric
d on X generating the topology τ .
One shows, see [54, p. 163], that the topology of a metrizable TVS can be always given by an
F -norm. If (X, τ ) is a TVS, then the topology τ generates a uniformity Wτ on X, a basis of it
being given by the sets
WU = {(x, y) ∈ X 2 : y − x ∈ U } ,
where U runs over a 0-neighborhood basis in X. Any translation-invariant metric generating the
topology τ generates the same uniformity Wτ , so that if X is complete with respect to Wτ , then
it is complete with respect to any translation-invariant metric generating the topology τ.
Typical examples of quasi-normed spaces are the spaces Lp[0, 1] and ℓp with 0 < p < 1 equipped
with the quasi-norms
(4.2)
kf kp =(cid:18)Z 1
0
f (t)dt(cid:19)1/p
for f ∈ Lp[0, 1] and x = (xk)k∈N ∈ ℓp , respectively.
and kxkp = ∞Xk=1
xkp!1/p
,
The quasi-norms k · kp satisfy the inequalities
(4.3)
kf + gkp ≤ 2(1−p)/p(kf kp + kgkp) and
kx + ykp ≤ 2(1−p)/p(kxkp + kykp) ,
for all f, g ∈ Lp[0, 1] and x, y ∈ ℓp.
The constant 2(1−p)/p > 1 is sharp, i.e. the moduli of concavity of the spaces Lp[0, 1] and ℓp are
both equal to 2(1−p)/p.
To show this, we start with the elementary inequalities
(4.4)
valid for all a, b > 0.
(a + b)p ≤ ap + bp ≤ 21−p(a + b)p ,
Let f, g ∈ Lp[0, 1]. The first inequality from above implies
f (t) + g(t)p ≤ (f (t) + g(t))p ≤ f (t)p + g(t)p ,
for almost all t ∈ [0, 1], so that
(4.5)
kf + gkp
p =Z 1
0
f (t) + g(t)pdt ≤Z 1
0
f (t)pdt +Z 1
0
g(t)pdt
= kf kp
p + kgkp
p .
This inequality and the second inequality from (4.4) yield
p(cid:1)1/p
kf + gkp =(cid:0)kf + gkp
≤ 2(1−p)/p(kf kp + kgkp) .
≤(cid:0)kf kp
p + kgkp
p(cid:1)1/p
Similar calculations can be done to show that
kx + ykp ≤ 2(1−p)/p(kxkp + kykp) ,
for all x, y ∈ ℓp.
To show that the constant 2(1−p)/p is sharp take x = (1, 0, 0 . . . ) and y = (0, 1, 0 . . . ) in the case
of the space ℓp. Then
kx + ykp = 21/p and 2(1−p)/p(kxkp + kykp) = 2(1−p)/p · 2 = 21/p ,
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
25
2 ) and g = χ[ 1
2 ,1] to obtain equality in the first inequality from (4.3).
that is, we have equality in the second inequality from (4.3). In the case of the space Lp[0, 1] take
f = χ[0, 1
Remark 4.2. Apparently similar, the quasi-normed spaces ℓp and Lp[0, 1] drastically differ. For
instance, the space Lp[0, 1] has trivial dual, (Lp[0, 1])∗ = {0}, while (ℓp)∗ = ℓ∞, see [54, pp. 156-
158]. D. Pallaschke [63] and Ph. Turpin [75]) have shown that every compact endomorphism of
Lp, 0 < p < 1, is null. N. Kalton and J. H. Shapiro [48] showed that there exists a quasi-Banach
space with trivial dual admitting non-trivial compact endomorphisms. The example is a quotient
space of H p, 0 < p < 1. Here, H p, 0 < p < 1, denotes the classical Hardy quasi-Banach spaces of
analytic functions in the unit disk of C.
A p-norm, where 0 < p ≤ 1, is a mapping k · k : X → R+ satisfying (QN1) and (QN2) and
(QN3′) kx + ykp ≤ kxkp + kykp , for all x, y ∈ X.
The quasi-norms of the spaces Lp[0, 1] and ℓp, 0 < p < 1, a p-norms, i.e.
kf + gkp
kx + ykp
p ≤ kf kp
p ≤ kxkp
p + kgkp
p and
p + kykp
p,
for all f, g ∈ Lp[0, 1] and x, y ∈ ℓp.
A famous result of T. Aoki [13] and S. Rolewicz [69] says that on any quasi-normed space
(X, k · k) there exists a p-norm equivalent to k · k, where p is determined from the equality 21/p = k,
k being the constant from (QN3).
Let 0 < p ≤ 1. A subset A of a vector space X is called p-convex if αx + βy ∈ A for all x, y ∈ A
and all α, β ≥ 0 with αp + βp = 1, and p-absolutely convex if αx + βy ∈ A for all x, y ∈ A and all
α, β ∈ K with αp + βp ≤ 1. For p = 1 one obtains the usual convex and absolutely convex sets,
respectively.
A TVS X is p-normable if and only if it has a bounded p-convex 0-neighborhood, see [54,
p. 161]. One shows first that under this hypothesis there exists a bounded p-absolutely convex
neighborhood V of 0 and one defines the p-norm as the Minkowski functional corresponding to V ,
i.e. kxk = inf{t : t > 0, x ∈ tV }.
Remark 4.3. In Kothe [54] by a p-norm on a vector space X one understands a mapping k · k′ :
X → R+ such that
kxk′ = 0 ⇐⇒ x = 0,
kαxk′ = αpkxk′ and kx + yk′ ≤ kxk′ + kyk′ ,
for all x, y ∈ X and α ∈ K. In this case the "p-norm" corresponding to a bounded absolutely
p-convex 0-neighborhood is given by kxk′ = inf{tp : t > 0, x ∈ tV }.
It follows that k · k is a p-norm in the sense given here if and only if k · kp is a p-norm in the
sense given in [54].
The Banach envelope.
Let (X, k · k) be a quasi-Banach space and BX = {x ∈ X : kxk ≤ 1} its closed unit ball.
Denote by k · kC the Minkowski functional of the set C = co(BX ). It is obvious that k · kC is
a seminorm on X and a norm on the quotient space X/N , where N = {x ∈ X : kxkC = 0}.
the completion of X/N with respect to the quotient-norm k · k bX corresponding to k · kC, whose
Since, for x 6= 0, x′ := x/kxk ∈ BX ⊆ C, it follows kx′kC ≤ 1, that is kxkC ≤ kxk. Denote by bX
(unique) extension to bX is denoted also by k · k bX. It follows kbxk bX ≤ kxk for all x ∈ X, hence the
embedding j : X → bX is continuous and one shows that j(X) is dense in bX. The space bX is called
We distinguish two situations.
I. X has trivial dual : X ∗ = {0}.
In this case C = co(BX ) = X (see [49, Proposition 2.1, p. 16]) and so k · kC ≡ 0, N = X and
the Banach envelope of the quasi-Banach space X.
X/X = {b0}. It follows bX = {0} and bX ∗ = {0} = X ∗. In particularcLp = {0}, where Lp = Lp[0, 1].
II. X has a separating dual.
26
S. COBZAS¸
This means that for every x 6= 0 there exists x∗ ∈ X ∗ with x∗(x) 6= 0 (e.g. X = ℓp with
0 < p < 1). In this case k · kC is a norm on X which can be calculated by the formula
(4.6)
kxkC = sup{x∗(x) : x∗ ∈ X ∗, kx∗k ≤ 1} .
where the norm of x∗ ∈ X ∗ is given by (4.1).
Consequently N = {0}, X/N = X and we can consider X as a dense subspace of bX (in fact,
continuously and densely embedded in bX).
It follows that:
(i) every continuous linear functional on (X, k · k) has a unique norm preserving extension to
(ii) every continuous linear operator T from (X, k · k) to a Banach space Y has a unique norm
(bX, k · k bX);
preserving extension bT : (bX, k · k bX) → Y .
Consequently (X, k · k)∗ can be identified with (bX, k · k bX)∗ and the norm k · k bX can also be
calculated by the formula (4.6) for all x ∈ bX.
One shows that the Banach envelope of ℓp is ℓ1, for every 0 < p < 1.
Another way to define the Banach envelope in the case of a quasi-Banach space with separating
dual is via the embedding jX of X into its bidual X ∗∗ (see [49]). Since X ∗ separates the points of
X, it follows that jX is injective of norm kjX k ≤ 1 (in this case one can not prove that kjX k = 1
because the Hahn-Banach extension theorem may fail for non-locally convex spaces).
By (4.6)
kxkC = sup{x∗(x) : kx∗k ≤ 1} = kjX (x)kX ∗∗ ,
so we can identify bX with the closure of jX (X) in (X ∗∗, k · kX ∗∗)
4.2. Lipschitz functions and quasi-normed spaces. It turns out that some results concerning
Banach space-valued Lipschitz functions fail in the quasi-Banach case and, in some cases, the
validity of some of them forces the quasi-Banach space to be locally convex, i.e. a Banach space.
In this subsection we consider only spaces over R.
Let (Z, d) be a b-metric space and (Y, k · k) a quasi-normed space. A function f : Z → Y is
called Lipschitz if there exists L ≥ 0 (called a Lipschitz constant for f ) such that
for all z, z′ ∈ Z. One denotes by Lip(Z, Y ) the space of all Lipschitz functions from Z to Y .
kf (z) − f (z′)k ≤ Ld(z, z′) ,
The Lipschitz norm kf kL of f is defined by
kf kL = sup(cid:26) kf (z) − f (z′)k
d(z, z′)
: z, z′ ∈ Z, z 6= z′(cid:27) .
It follows that kf kL is the smallest Lipschitz constant for f .
Since kf k = 0 if and only if f = const, k · k is actually only a seminorm on Lip(Z, Y ). To obtain
a norm, one considers a fixed element z0 ∈ Z and the space
Lip0(Z, Y ) = {f ∈ Lip(Z, Y ) : f (z0) = 0} .
If Z = X, where X is a quasi-normed space, then one take 0 for the fixed point z0, and, in this
case, Lip0(X, Y ) is a quasi-normed space, that is,
kf + gkL ≤ k(kf kL + kgkL) ,
for f, g ∈ Lip0(X, Y ), where k ≥ 1 is the constant from (QN3). It is complete, provided Y is a
quasi-Banach space.
If Y = R, then one uses the notation Lip(X), Lip0(X) and Lip0(X) is called the Lipschitz dual
of the quasi-normed space X.
We noted that the space Lp = Lp[0, 1] has trivial dual. F. Albiac [3] proved that it has also a
trivial Lipschitz dual, i.e. Lip0(Lp) = {0}. Later he showed that this is a more general phenomenon.
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
27
Proposition 4.4 (F. Albiac [4]). Let (X, k · k) be a quasi-Banach space and
x := sup{f (x) : f ∈ Lip0(X), kf kL ≤ 1} , x ∈ X .
Then
(i) · is a seminorm on X;
(ii) if Lip0(X) is nontrivial, then X has a nontrivial dual, i.e. X ∗ 6= {0};
(iii) if X has a separating Lipschitz dual, then X has a separating (linear) dual and · is a
norm on X.
One says that a family F of real valued functions on a quasi-normed space X is separating if
for every x 6= y in X there exists f ∈ F with f (x) 6= f (y).
It is known that every Lipschitz function f from a subset of a metric space (X, d) to R admits an
extension to X which is Lipschitz with the same Lipschitz constant (McShane's extension theorem).
The following result shows that the validity of this result for every subset of a quasi-Banach space
X forces this space to be Banach.
Proposition 4.5 ([4]). Let (X, k · k) be a quasi-Banach space. If for every subset Z of X, every
L-Lipschitz function f : Z → R admits an L′-Lipschitz extension, for some L′ ≥ L, then the space
X is locally convex, i.e. it is a Banach space.
It is known that every continuous linear operator from a quasi-Banach space X to a Banach
[4] has shown that this is true for Lipschitz mappings too: every Lipschitz mapping f : X → Y
space Y admits a norm preserving linear extension to the Banach envelope bX of X to Y . F. Albiac
admits a unique Lipschitz extension with the same Lipschitz constant bf : bX → Y .
Moreover, if X, Y are normed spaces and f : X → Y is Gateaux differentiable on the interval
[x, y] := {x + t(y − x) : t ∈ [0, 1]}, then
(4.7)
kf (x) − f (y)k ≤ kx − yk sup{kf ′(ξ)k : ξ ∈ [x, y]} .
Proposition 4.6 ([4]). Let (X, k · k) be a quasi-Banach space.
If every nonconstant Gateaux
differentiable Lipschitz function f : [0, 1] → X satisfies the mean value inequality (4.7) for all
x, y ∈ [0, 1], then the space X is locally convex, i.e. it is a Banach space.
Let α > 0. A function f : (X1, d1) → (X2, d2) between two b-metric spaces X1, X2 is called
Holder of order α if there exists L ≥ 0 such that
(4.8)
d2(f (x), f (y)) ≤ Ld1(x, y)α ,
for all x, y ∈ X1. As a consequence of the mean value theorem, every function f from [0, 1] to a
Banach space X which is Holder of order α > 1 is constant, a fact that is no longer true if X is a
quasi-Banach space.
Example 4.7. Let Lp = Lp[0, 1] for 0 < p < 1. The function f : [0, 1] → Lp given by f (t) = χ[0,t]
satisfies the equality
for all s, t ∈ [0, 1], where k · kp is the Lp-norm (see (4.2)).
kf (s) − f (t)kp = s − t1/p ,
Indeed, for 0 ≤ t < s ≤ 1,
kf (s) − f (t)kp =(cid:18)Z s
t
χp
(t,s](u)du(cid:19)1/p
= s − t1/p .
The Riemann integral of a function f : [a, b] → X, where [a, b] is an interval in R and X is a
Banach space, can be defined as in the real case, by simply replacing the absolute value · with the
norm sign k · k, and has properties similar to those from the real case. For instance, the following
result is true.
Proposition 4.8 ([4]). Let X be a Banach space. If f : [a, b] → X is continuous, then
28
S. COBZAS¸
(i) f is Riemann integrable, and
(ii)
the function
F (t) =Z t
a
(4.9)
f (s)ds, t ∈ [a, b] ,
is differentiable with F ′(t) = f (t) for all t ∈ [a, b].
Remark 4.9. However, there is a point where this analogy is broken, namely the Lebesgue criterion
of Riemann integrability: a function f : [a, b] → R is Riemann integrable if and only if it is
continuous almost everywhere on [a, b] (i.e. excepting a set of Lebesgue measure zero). In the
infinite dimensional case this criterion does not hold in general, leading to the study of those
Banach spaces for which it, or some weaker forms, are true , see, for instance, [40], [73], [74] and
the references quoted therein.
In the case of quasi-Banach spaces the situation is different. By a result attributed to S. Mazur
and W. Orlicz [60] (see also [70, p. 122]) an F -space X is locally convex if and only if every
continuous function f : [0, 1] → X is Riemann integrable.
M. M. Popov [67] investigated the Riemann integrability of functions defined on intervals in R
with values in an F -space. Among other results, he proved that a Riemann integrable function
f : [a, b] → X is bounded and that the function F defined by (4.9) is uniformly continuous, but
there exists a continuous function f : [0, 1] → ℓp, where 0 < 1 < p, such that the function F does
not have a right derivative at t = 0. He asked whether any continuous function f from [0, 1] to
Lp[0, 1], 0 < p < 1, (or more general, to a quasi-Banach space X with X ∗ = {0}) admits a primitive.
This problem was solved by N. Kalton [46] who proved that if X is a quasi-Banach space with
X ∗ = {0}, then every continuous function f : [0, 1] → X has a primitive. Kalton considered the
space C1
Kal(I, X), where I = [0, 1] and X is a quasi-Banach space, of all continuously differentiable
functions f : I → X such that the function f : I 2 → X given for s, t ∈ I by f (t, t) = f ′(t) and
f (s, t) = (f (s) − f (t))/(s − t) if s 6= t, is continuous. It follows that C1
Kal(I, X) is a quasi-Banach
space with respect to the quasi-norm
kf k = kf (0)k + kf kL .
Kal(I, X) was introduced in [5]; Kalton used the notation C1(I; X).
The notation C1
Denote by C(I, X) the Banach space (with respect to the sup-norm) of all continuous functions
from I to X. The core of a quasi-Banach space X is the maximal subspace Z of X (denoted by
core(X)) with Z ∗ = {0}. One shows that such a subspace always exists, is unique and closed. No-
tice that core(X) = {0} implies only that X has a nontrivial dual, but not necessarily a separating
one.
In [6] it is shown that if X is a quasi-Banach space with core(X) = {0}, then there exists a
continuous function f : [0, 1] → X failing to have a primitive.
Kalton, op. cit., called a quasi-Banach X a D-space if the mapping
Kal(I, X) → C(I, X) ,
given by Df = f ′, is surjective and proved the following result.
D : C1
Theorem 4.10 ([46]). Let X be a quasi-Banach with core(X) = {0}. Then X is a D-space if and
only if X is locally convex (or, equivalently, a Banach space).
It is known that every continuously differentiable function from an interval [a, b] ⊆ R to a
Banach space X is Lipschitz with kf kL = sup{kf ′(t)k : t ∈ [a, b]} (a consequence of the Mean
Value Theorem, see (4.7)). As it was shown in [5] this in no longer true in quasi-Banach spaces.
Theorem 4.11. Let X be a non-locally convex quasi-Banach space X. Then there exists a function
F : I → X such that:
(i) F is continuously differentiable on I;
(ii) F ′ is Riemann integrable on I and F (t) =R t
0 F ′(s)ds, t ∈ I;
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
29
(iii) F is not Lipschitz on I.
In [7] it is proved that the usual rule of the calculation of the integral (called Barrow's rule by
the authors, known also as Leibniz rule) holds in the quasi-Banach case in the following form.
Proposition 4.12. Let X be a quasi-Banach with separating dual. If F : [a, b] → X is differen-
tiable with Riemann integrable derivative, then
Z b
a
F ′(t)dt = F (b) − F (a) .
Another pathological result concerning differentiability of quasi-Banach valued Lipschitz func-
tions was obtained by N. Kalton [45, Theorem 3.3].
Theorem 4.13. Let X be an F -space with trivial dual. Then for every pair of distinct points
x0, x1 ∈ X there exists a function f : [0, 1] → X such that f (0) = x0, f (1) = x1 and
lim
s−t→0
f (s) − f (t)
s − t
= 0 uniformly for s, t ∈ [0, 1] .
In particular f ′(t) = 0 for all t ∈ [0, 1].
Remark 4.14. N. Kalton [45, Corollary 3.4] also remarked that if X is an F -space and x ∈ X \{0},
then in order to exist a function f : [0, 1] → X such that f (0) = 0, f (1) = x and f ′(t) = 0 for all
t ∈ [0, 1], it is necessary and sufficient that x ∈ core(X).
If X is a Banach space and f : [0, 1] → X is continuous then it is Riemann integrable and the
average function Ave[f ] : [a, b] × [a, b] → X, given by
Ave[f ](s, t) =
f (c)
t−sR t
s−tR s
1
1
s f (u)du
t f (u)du
if a ≤ s < t ≤ b,
if s = t = c ∈ [a, b],
if a ≤ t < s ≤ b,
is jointly continuous on [a, b]×[a, b], and so, separately continuous and bounded. Some pathological
properties of the average function in the quasi-Banach case are examined in [5], [8] and [67].
The analog of the Radon-Nikod´ym Property for quasi-Banach spaces and its connections with
the differentiability of Lipschitz mappings and martingales are discussed in [9].
4.3. Lipschitz functions on spaces of homogeneous type. Let (X, d, µ) be a space of homo-
geneous type (see Subsection 1.4). By B we shall denote balls of the form B(x, r). If ϕ is a function
integrable on bounded sets, then the mean value of ϕ on the ball B is defined by
(4.10)
mB(ϕ) = µ(B)−1ZB
ϕ(x)dµ(x) .
For 1 ≤ q < ∞ and 0 < β < ∞ one denotes by Lip(q, β) the set of all functions ϕ, integrable on
bounded sets, for which there exists a constant C ≥ 0 such that
(4.11)
(cid:18) 1
µ(B)ZB
ϕ(x) − mB(ϕ)qdµ(x)(cid:19)1/q
≤ Cµ(B)β ,
for all balls B. The least constant C for which (4.11) holds will be denoted by kϕkβ,q .
We shall denote by Lip(β) the set of all functions ϕ on X such that there exists a constant
C ≥ 0 satisfying
(4.12)
ϕ(x) − ϕ(y) ≤ Cd(x, y)β ,
for all x, y ∈ X, i.e. it is Holder of order β (see (4.8)). The least C ≥ 0 for which (4.12) holds is
denoted by kϕkβ .
The following results concerning these classes of Lipschitz functions were proved in [57].
30
S. COBZAS¸
Theorem 4.15. Let (X, d, µ) be a space of homogeneous type. Then there exists a constant C ≥ 0
(depending on β and q only) such that for every ϕ ∈ Lip(β, q) there exists a function ψ satisfying
(i)
(ii)
ϕ(x) = ψ(x) a.e. on X, and
ψ(x) − ψ(y) ≤ Ckϕkβ,qµ(B)β ,
for any ball B containing the x, y.
Theorem 4.16. Let (X, d, µ) be a space of homogeneous type. Then, given 0 < β < ∞, there
exists a b-metric δ on X such that (X, δ, µ) is a normal space of homogeneous type and for every
1 ≤ q < ∞ we have
ϕ ∈ Lip(β, q) of (X, d, µ) ⇐⇒ ∃ψ ∈ Lip(β) of (X, δ, µ) with ϕ a.e.= ψ.
Moreover, the norms kϕkβ,q and kψkβ are equivalent.
References
[1] H. Aimar, Balls as subspaces of homogeneous type: On a construction due to R. Mac´ıas and C. Segovia,
Recent developments in real and harmonic analysis. In honor of Carlos Segovia, Boston, MA: Birkhauser, 2010,
pp. 25 -- 36. 7
[2] H. Aimar, B. Iaffei, and L. Nitti, On the Mac´ıas-Segovia metrization of quasi-metric spaces, Rev. Uni´on Mat.
Argent. 41 (1998), no. 2, 67 -- 75. 3, 4, 6, 7
[3] F. Albiac, Nonlinear structure of some classical quasi-Banach spaces and F -spaces, J. Math. Anal. Appl. 340
(2008), no. 2, 1312 -- 1325. 26
[4] F. Albiac, The role of local convexity in Lipschitz maps, J. Convex Anal. 18 (2011), no. 4, 983 -- 997. 27
[5] F. Albiac and J. L. Ansorena, Lipschitz maps and primitives for continuous functions in quasi-Banach spaces,
Nonlinear Anal. 75 (2012), no. 16, 6108 -- 6119. 28, 29
[6]
[7]
[8]
[9]
, On a problem posed by M. M. Popov, Studia Math. 211 (2012), no. 3, 247 -- 258. 28
, Integration in quasi-Banach spaces and the fundamental theorem of calculus, J. Funct. Anal. 264
(2013), no. 9, 2059 -- 2076. 29
, Optimal average approximations for functions mapping in quasi-Banach spaces, J. Funct. Anal. 266
(2014), no. 6, 3894 -- 3905. 29
, On Lipschitz maps, martingales, and the Radon-Nikod´ym property for F-spaces, Mediterr. J. Math.
13 (2016), no. 4, 1963 -- 1980. 29
[10] T. V. An and N. V. Dung, Remarks on Frink's metrization technique and applications, arXiv preprint
arXiv:1507.01724 (2015), 15 p. 3, 20
[11] T. V. An, L. Q. Tuyen, and N. V. Dung, Stone-type theorem on b-metric spaces and applications, Topology
Appl. 185-186 (2015), 50 -- 64. 5, 15
[12]
, Answers to Kirk-Shahzad's questions on strong b-metric spaces, Taiwanese J. Math. 20 (2016), no. 5,
1175 -- 1184. 9
[13] T. Aoki, Locally bounded linear topological spaces, Proc. Imper. Acad. Tokyo 18 (1942), 588 -- 594. 25
[14] A. V. Arutyunov, A. V. Greshnov, L. V. Lokutsievskii, and K. V. Storozhuk, Topological and geometrical
properties of spaces with symmetric and nonsymmetric f -quasimetrics, Topology Appl. 221 (2017), 178 -- 194.
10, 14
[15] A. V. Arutyunov and A. V. Greshnov, Coincidence points of multivalued mappings in (q1, q2)-quasimetric
spaces, Dokl. Math. 96 (2017), no. 2, 438 -- 441. 11
[16] H. Aydi and S. Czerwik, Fixed point theorems in generalized b-metric, spaces, in Daras, Nicholas J. (ed.) et
al., Modern discrete mathematics and analysis. With applications in cryptography, information systems and
modeling. Springer Optimization and Its Applications 131, pp. 1 -- 9, Cham: Springer 2018. 22
[17] I. A. Bakhtin, The contraction mapping principle in quasimetric spaces, Funktionalnyi Analyz, Ulianovskii
Gos. Ped. Inst. 30 (1989), 26 -- 37 (Russian). 15, 19
[18] G. Beer, The structure of extended real-valued metric spaces, Set-Valued Var. Anal. 21 (2013), no. 4, 591 -- 602.
15
[19]
[20] G. Beer and J. Vanderwerff, Structural properties of extended normed spaces, Set-Valued Var. Anal. 23 (2015),
, Norms with infinite values, J. Convex Anal. 22 (2015), no. 1, 37 -- 60. 15
no. 4, 613 -- 630. 15
[21] V. Berinde, Generalized contractions in quasimetric spaces, Semin. Fixed Point Theory Cluj-Napoca 1993
(1993), 3 -- 9. 22
[22] V. Berinde and M. Choban, Generalized distances and their associate metrics. Impact on fixed point theory,
Creat. Math. Inform. 22 (2013), no. 1, 23 -- 32. 1
[23] S. Cobza¸s, Functional analysis in asymmetric normed spaces, Frontiers in Mathematics, Birkhauser/Springer
Basel AG, Basel, 2013. 11, 14
B-METRIC SPACES, FIXED POINTS AND LIPSCHITZ FUNCTIONS
31
[24] S. Cobza¸s and S. Czerwik, The completion of generalized b-metric spaces and fixed points, Fixed Point Theory
(2018), (to appear). 1
[25] R. R. Coifman and M. de Guzman, Singular integrals and multipliers on homogeneous spaces, Rev. Uni´on Mat.
Argent. 25 (1970), 137 -- 143. 9, 14
[26] S. Czerwik, Contraction mappings in b-metric spaces, Acta Math. Inform. Univ. Ostrav. 1 (1993), 5 -- 11. 15, 21
, Nonlinear set-valued contraction mappings in B-metric spaces, Atti Semin. Mat. Fis. Univ. Modena
[27]
46 (1998), no. 2, 263 -- 276. 15
[28] S. Czerwik and K. Kr´ol, Cantor, Banach and Baire theorems in generalized metric spaces, Mathematical
analysis, approximation theory and their applications, Cham: Springer, 2016, pp. 139 -- 144. 15
[29]
[30]
, Completion of generalized metric spaces, Indian J. Math. 58 (2016), no. 2, 231 -- 237. 16
, Generalized Minkowski functionals, Contributions in mathematics and engineering. In honor of Con-
stantin Carath´eodory. With a foreword by R. Tyrrell Rockafellar, Cham: Springer, 2016, pp. 69 -- 79. 15
, Fixed point theorems in generalized metric spaces, Asian-Eur. J. Math. 10 (2017), no. 2, 8 p. 22
[31]
[32] M. M. Deza and E. Deza, Encyclopedia of distances, 3rd ed., Springer, Berlin, 2014. 1, 14
[33] G. Dezso, Fixed point theorems in generalized metric spaces, PU.M.A., Pure Math. Appl. 11 (2000), no. 2,
183 -- 186. 15
[34] J. B. Diaz and B. Margolis, A fixed point theorem of the alternative, for contractions on a generalized complete
metric space, Bull. Amer. Math. Soc. 74 (1968), 305 -- 309. 15, 22
[35] M. Edelstein, A remark on a theorem of A.F. Monna., Nederl. Akad. Wet., Proc., Ser. A 67 (1964), 88 -- 89. 15
[36] R. Engelking, General topology, second ed., Sigma Series in Pure Mathematics, vol. 6, Heldermann Verlag,
Berlin, 1989, (translated from the Polish by the author). 3
[37] R. Fagin, R. Kumar, and D. Sivakumar, Comparing top k lists, SIAM J. Discrete Math. 17 (2003), no. 1,
134 -- 160. 3, 5, 6, 15
[38] R. Fagin and L. Stockmeyer, Relaxing the triangle inequality in pattern matching, International J. Computer
Vision 28 (1998), no. 3, 134 -- 160. 15
[39] A. H. Frink, Distance functions and the metrization problem, Bull. Amer. Math. Soc. 43 (1937), 133 -- 142. 3
[40] R. Gordon, Riemann integration in Banach spaces, Rocky Mountain J. Math. 21 (1991), no. 3, 923 -- 949. 28
[41] A. Granas and J. Dugundji, Fixed point theory, Springer, New York, NY, 2003. 17
[42] J. Heinonen, Lectures on analysis on metric spaces, Springer, New York, NY, 2001. 4
[43] C. F. K. Jung, On generalized complete metric spaces, Bull. Am. Math. Soc. 75 (1969), 113 -- 116. 15
[44] S. Kaj´ant´o, A. Luk´acs, A note on the paper "Contraction mappings in b-metric spaces" by Czerwik, Acta Univ.
Sapientiae Math. 10 (2018), no. 1, 8589 18, 21
[45] N. J. Kalton, Curves with zero derivative in F-spaces, Glasg. Math. J. 22 (1981), 19 -- 29. 29
[46]
, The existence of primitives for continuous functions in a quasi-Banach space, Atti Semin. Mat. Fis.
Univ. Modena 44 (1996), no. 1, 113 -- 117. 28
[47]
, Quasi-Banach spaces, Handbook of the geometry of Banach spaces. Volume 2, Amsterdam: North-
Holland, 2003, pp. 1099 -- 1130. 23
[48] N. J. Kalton and J. H. Shapiro, An F-space with trivial dual and non-trivial compact endomorphisms, Israel
J. Math. 20 (1975), 282 -- 291. 25
[49] N.J. Kalton, N.T. Peck, and J. W. Roberts, An F-space sampler, London Mathematical Society Lecture Note
Series, vol. 89, Cambridge University Press, Cambridge, 1984. 23, 25, 26
[50] M. A. Khamsi, Remarks on cone metric spaces and fixed point theorems of contractive mappings, Fixed Point
Theory Appl. 2010 (2010), 7 p. 15
[51]
[52] M. A. Khamsi and N. Hussain, KKM mappings in metric type spaces, Nonlinear Anal. 73 (2010), no. 9,
, Generalized metric spaces: a survey, J. Fixed Point Theory Appl. 17 (2015), no. 3, 455 -- 475. 1
3123 -- 3129. 15
[53] W. Kirk and N. Shahzad, Fixed point theory in distance spaces, Cham: Springer, 2014. 1, 2, 6, 8, 9, 20, 21
[54] G. Kothe, Topological vector spaces. I, Springer-Verlag, Berlin-Heidelberg-New York, 1969, (Translated by D.
J. H. Garling). 23, 24, 25
[55] W.A.J. Luxemburg, On the convergence of successive approximations in the theory of ordinary differential
equations, Can. Math. Bull. 1 (1958), 9 -- 20. 15
[56]
, On the convergence of successive approximations in the theory of ordinary differential equations. III,
Nieuw Arch. Wiskd., III. Ser. 6 (1958), 93 -- 98. 15
[57] R. A. Mac´ıas and C. Segovia, Lipschitz functions on spaces of homogeneous type, Adv. Math. 33 (1979),
257 -- 270. 3, 9, 10, 14, 29
[58]
[59] B. Margolis, On some fixed points theorems in generalized complete metric spaces, Bull. Amer. Math. Soc. 74
, Singular integrals on generalized Lipschitz and Hardy spaces, Stud. Math. 65 (1979), 55 -- 75. 9, 14
(1968), 275 -- 282. 15
[60] S. Mazur and W. Orlicz, Sur les espaces m´etriques lin´eaires. I, Studia Math. 10 (1948), 184 -- 208 (French). 28
[61] A. C. G. Mennucci, On asymmetric distances, Anal. Geom. Metr. Spaces 1 (2013), 200 -- 231. 14
[62] A.F. Monna, Sur un th´eor`eme de M. Luxemburg concernant les points fixes d'une classe d'application d'un
espace m´etrique dans lui-meme, Nederl. Akad. Wet., Proc., Ser. A 64 (1961), 89 -- 96. 15
32
S. COBZAS¸
[63] D. Pallaschke, The compact endomorphisms of the metric linear spaces Lϕ, Studia Math. 47 (1973), 123 -- 133.
25
[64] M. Paluszy´nski and K. Stempak, On quasi-metric and metric spaces, Proc. Amer. Math. Soc. 137 (2009),
no. 12, 4307 -- 4312. 3, 5
[65] A. I. Perov, On the Cauchy problem for a system of ordinary differential equations, Priblizhen. Metody Reshen.
Differ. Uravn. Vyp. 2, (1964), 115-134 (Russian). 15
[66] A. I. Perov and A. V. Kibenko, On a certain general method for investigation of boundary value problems, Izv.
Akad. Nauk SSSR, Ser. Mat. 30 (1966), 249 -- 264 (Russian). 15
[67] M. M. Popov, On integrability in F -spaces, Studia Math. 110 (1994), no. 3, 205 -- 220. 28, 29
[68] I. L. Reilly , P. V. Subrahmanyam, M. K. Vamanamurthy, Cauchy sequences in quasi-pseudo-metric spaces,
Monatsh. Math. 93 (1982), 127 -- 140. 14
, Metric linear spaces, (2nd ed.), D. Reidel Publishing Company, Boston, 1985. 23, 28
[69] S. Rolewicz, On a certain class of linear metric spaces, Bull. Acad. Pol. Sci., Cl. III 5 (1957), 471 -- 473. 25
[70]
[71] I. A. Rus, A. Petru¸sel, and G. Petru¸sel, Fixed point theory, Cluj University Press, Cluj-Napoca, 2008. 1, 22
[72] V. Schroeder, Quasi-metric and metric spaces, Conform. Geom. Dyn. 10 (2006), 355 -- 360. 3
[73] M. A. Sofi, Weaker forms of continuity and vector-valued Riemann integration, Colloq. Math. 129 (2012),
no. 1, 1 -- 6. 28
[74]
, Riemann integrability under weaker forms of continuity in infinite dimensional space, arXiv:1612.00642
[math.FA] (2016), 18 pp. 28
[75] Ph. Turpin, Op´erateurs lin´eaires entre espaces d'Orlicz non localement convexes, Studia Math. 46 (1973),
153 -- 165 (French). 25
[76] Q. Xia, The geodesic problem in quasimetric spaces, J. Geom. Anal. 19 (2009), no. 2, 452 -- 479. 15
S. Cobzas¸, Babes¸-Bolyai University, Department of Mathematics, Cluj-Napoca, Romania
E-mail address: [email protected]
|
1605.03769 | 1 | 1605 | 2016-05-12T11:46:49 | Operator Space Structures on $\ell^1(n).$ | [
"math.FA"
] | We show that the complex normed linear space $\ell^1(n),$ $n>1,$ has no isometric embedding into $k\times k$ complex matrices for any $k\in \mathbb{N}$ and discuss a class of infinite dimensional operator space structures on it. | math.FA | math |
OPERATOR SPACE STRUCTURES ON ℓ1(n)
RAJEEV GUPTA AND MD. RAMIZ REZA
Abstract. We show that the complex normed linear space ℓ1(n), n > 1, has no isometric
embedding into k × k complex matrices for any k ∈ N and discuss a class of infinite dimensional
operator space structures on it.
1. Introduction
In this paper, all the normed linear spaces considered are over the field of complex numbers
unless specified. It is well known that there are isometric embeddings of real ℓ1(n) into real
ℓ∞(k) for some k and hence into the space of k × k real matrices Mk(R). However, we prove
that ℓ1(n), n > 1, has no isometric embedding into Mk for any k ∈ N. This shows that there
is no operator space structure on ℓ1(n), n > 1, which can be induced by any k × k matrices
A1, . . . , An. Furthermore, we study the operators space structures on ℓ1(n). We recall some
definitions first.
Definition 1.1. An abstract operator space is a normed linear space V together with a sequence
of norms k · kk defined on the linear space
with the understanding that k · k1 is the norm of V and the family of norms k · kk satisfies the
compatibility conditions:
Mk(V ) :=(cid:8)(cid:0)(cid:0)vij(cid:1)(cid:1)vij ∈ V, 1 ≤ i, j ≤ k(cid:9), ∀k ∈ N,
1. kT ⊕ Skp+q = max(cid:8)kT kp, kSkq(cid:9) and
2. kASBkq ≤ kAkopkSkpkBkop
for all S ∈ Mq(V ), T ∈ Mp(V ), A ∈ Mq×p(C) and B ∈ Mp×q(C).
Let (V, k·kk) and (W, k·kk) be two operator spaces. A linear bijection T : V → W is said to be
a complete isometry if T ⊗ Ik : (Mk(V ), k · kk) → (Mk(W ), k · kk) is an isometry for every k ∈ N.
Operator spaces (V, k · kk) and (W, k · kk) are said to be completely isometric if there is a linear
complete isometry T : V → W . A well known theorem of Ruan says that any operator space
(V, k · kk) can be embedded, completely isometrically, into C∗-algebra B(H) for some Hilbert
space H. There are two natural operator space structures on any normed linear space V, which
may coincide. These are the MIN and the MAX operator space structures defined below.
Definition 1.2 (MIN). The MIN operator space structure denoted by MIN(V ) on a normed
linear space V is obtained by the isometric embedding of V into the C∗-algebra C((V ∗)1), the
Mk(V ), we set
space of continuous functions on the unit ball (V ∗)1 of the dual space V ∗. Thus for (cid:0)(cid:0)vij(cid:1)(cid:1) in
(cid:13)(cid:13)(cid:0)(cid:0)vij(cid:1)(cid:1)(cid:13)(cid:13)M IN = sup(cid:8)(cid:13)(cid:13)(cid:0)(cid:0)f (vij)(cid:1)(cid:1)(cid:13)(cid:13) : f ∈ (V ∗)1(cid:9) ,
where the norm of a scalar matrix (cid:0)(cid:0)f (vij)(cid:1)(cid:1) is the operator norm in Mk.
The results in this paper are from the first author's thesis "The Carath´eodory-Fej´er Interpolation Problems
and the von-Neumann Inequality" submitted to the Indian Institute of Science, Bangalore-560012.
Key words and phrases. Operator Spaces, Finite dimensional embedding, MIN structure.
The first named author was supported by the NBHM, Government of India.
The second named author was supported by CSIR, Government of India.
1
2
RAJEEV GUPTA AND MD. RAMIZ REZA
Definition 1.3 (Max). Let V be a normed linear space and (cid:0)(cid:0)vij(cid:1)(cid:1) ∈ Mk(V ). Define
where the supremum is taken over all isometry maps T and all Hilbert spaces H. This operator
space structure is denoted by MAX(V ).
(cid:13)(cid:13)(cid:0)(cid:0)vij(cid:1)(cid:1)(cid:13)(cid:13)M AX = sup(cid:8)(cid:13)(cid:13)(cid:0)(cid:0)T vij(cid:1)(cid:1)(cid:13)(cid:13) : T : V → B(H)(cid:9) ,
These two operator space structures are extremal in the sense that for any normed linear
space V , MIN(V ) and MAX(V ) are the smallest and the largest operator space structures on
V respectively. For any normed linear space V, Paulsen [Pau92] associates a constant, namely,
α(V ), which is defined as following.
α(V ) := sup(cid:8)kIV ⊗ Ikk(Mk(V ),k·kMIN)→(Mk(V ),k·kMAX) : k ∈ N(cid:9) .
The constant α(V ) is equal to 1 if and only if V has only one operator space structure on it.
There are only a few examples of normed linear spaces for which α(V ) is known to be 1. These
include α(ℓ∞(2)) = α(ℓ1(2)) = 1. In fact, it is known (cf.
[Pis03, Page 77]) that α(V ) > 1 if
dim(V ) ≥ 3.
The map φ : ℓ∞(n) → B(Cn) defined by φ(z1, . . . , zn) = diag(z1, . . . , zn), is an isometric
embedding of the normed linear space ℓ∞(n) into the finite dimensional C∗−algebra B(Cn).
Clearly, this is the MIN structure of the normed linear space ℓ∞(n). We shall, however prove that
there is no such finite dimensional isometric embedding for the dual space ℓ1(n). Nevertheless,
we shall construct, explicitly, a class of isometric infinite dimensional embeddings of ℓ1(n).
Unfortunately, all of these embeddings are completely isometric to the MIN structure. In the
end of this paper, using these embeddings and Parrott's example in [Mis94], we construct an
operator space structure on ℓ1(3), which is distinct from the MIN structure.
2. ℓ1(n) has no isometric embedding into any Mk
In this section, we will show that there does not exist an isometric embedding of ℓ1(n), n > 1,
into any finite dimensional matrix algebra Mk, k ∈ N. Without loss of generality, we prove this
for the case of n = 2. For the proof of the main theorem of this section, we shall need the
following lemma.
Lemma 2.1. For k ∈ N and θ1, . . . , θk ∈ [0, 2π), there exists a1, a2 ∈ C such that
Proof. For any two non-zero complex numbers a1, a2, we have
max
j=1,...,k(cid:12)(cid:12)a1 + eiθj a2(cid:12)(cid:12) <(cid:12)(cid:12)a1(cid:12)(cid:12) +(cid:12)(cid:12)a2(cid:12)(cid:12).
where φ1 and φ2 are the arguments of a1 and a2 respectively. Setting αj = θj + φ2 − φ1, we have
max
2
max
j=1,...,k(cid:12)(cid:12)a1 + eiθj a2(cid:12)(cid:12) = max
j=1,...,k(cid:12)(cid:12)a1 + eiθj a2(cid:12)(cid:12)
2 = max
j=1,...,k(cid:12)(cid:12)a1 + ei(θj +φ2−φ1)a2(cid:12)(cid:12),
j=1,...,k(cid:12)(cid:12)a1 + eiαj a2(cid:12)(cid:12)
j=1,...,k(cid:12)(cid:12)a12 + a22 + 2a1a2cosαj(cid:12)(cid:12).
= max
Therefore
if and only if cos αj = 1 for some j, that is, if and only if αj = 0 for some j. Choose a1 and a2
such that φ1 − φ2 6= θj for all j = 1, . . . , k. The existence of such a pair a1 and a2 proves the
lemma.
(cid:3)
max
j=1,...,k(cid:12)(cid:12)a1 + eiθj a2(cid:12)(cid:12) =(cid:12)(cid:12)a1(cid:12)(cid:12) +(cid:12)(cid:12)a2(cid:12)(cid:12)
OPERATOR SPACE STRUCTURES ON ℓ1(n)
3
The referee points out that the lemma is equivalent to the statement "There is no isometric
embedding of ℓ1(n), n > 1, into ℓ∞(k) for any k ∈ N." The argument below validates this
equivalence.
Suppose S : ℓ1(2) → ℓ∞(k) defined by S(z1, z2) := (a1z1 + b1z2, . . . , akz1 + bkz2) is an isometry
with smallest possible k ∈ N. Then, due to the minimality of k, it follows that aj = bj = 1
for j = 1, . . . , k. Without loss of generality, we can assume that aj = 1 for j = 1, . . . , n. Using
Lemma 2.2, we conclude that S can not be an isometry. For the converse part, we note that
Lemma 2.2 is equivalent to the statement that the linear map S : ℓ1(2) → ℓ∞(k) defined by
S(z1, z2) := (z1 + eiθ1 z2, . . . , z1 + eiθk z2) can not be an isometry.
Now, we prove the main theorem of this section.
Theorem 2.2. There is no isometric embedding of ℓ1(2) into Mk for any k ∈ N..
Proof. Suppose there is a k − dimensional isometric embedding φ of ℓ1(2). Then φ is induced
by a pair of operators T1, T2 ∈ Mk of norm 1, defined by the rule, φ(a1, a2) = a1T1 + a2T2. Let
U1 and U2 in M2k be the pair of unitary maps:
where DTi is the positive square root of the (positive) operator I − T ∗i Ti. Now, we have
Ui :=(cid:18) Ti DT ∗
DTi −T ∗i ,(cid:19) i = 1, 2,
i
PCk (a1U1 + a2U2)Ck = a1T1 + a2T2.
(This dilating pair of unitary maps is not necessarily commuting nor is it a power dilation!)
Thus ψ : ℓ1(2) → M2k(C) defined by ψ(a1, a2) = a1U1 + a2U2 is also an isometry. Since norms
are preserved under unitary operations, without loss of generality, we assume U1 = I and U2 to
be a diagonal unitary, say, D. Let D = diag(cid:0)eiθ1 , . . . , eiθ2k(cid:1). Applying Lemma 2.1, we obtain
complex numbers a1 and a2 such that
max
j=1,...,2k(cid:12)(cid:12)a1 + eiθj a2(cid:12)(cid:12) <(cid:12)(cid:12)a1(cid:12)(cid:12) +(cid:12)(cid:12)a2(cid:12)(cid:12).
Hence ψ cannot be an isometry, which contradicts the hypothesis that φ is an isometry.
(cid:3)
Remark 2.3. Let X be a finite dimensional normed linear space. Suppose X is embedded
isometrically in Mk for some k ∈ N, then the standard dual operator space structure on X∗
need not admit an embedding in Mn for any n ∈ N.
Remark 2.4. An amusing corollary to this theorem is that the two spaces ℓ∞(n) and ℓ1(n) cannot
be isometrically isomorphic for n > 1.
Remark 2.5. Prof. G. Pisier points out that Theorem 2.2 may be true if one replaces Mk by
K(H), the set of all compact operators on an infinite dimensional separable Hilbert space H.
3. Infinite dimensional embeddings of ℓ1(n)
In this section we construct operator space structure on ℓ1(n) (n ≥ 3) which is not completely
isometric to MIN structure of ℓ1(n).
Let Hi be a Hilbert space and Ti be a contraction on Hi for i = 1, . . . , n. Assume that the
unit circle T is contained in σ(Ti), the spectrum of Ti, for i = 1, . . . , n. Denote
T1 = T1 ⊗ I⊗(n−1), T2 = I ⊗ T2 ⊗ I⊗(n−2), . . . , Tn = I⊗(n−1) ⊗ Tn
and T = ( T1, . . . , Tn).
Theorem 3.1. Suppose the operators T1, . . . , Tn are defined as above. Then, the function
fT : ℓ1(n) → B(H1 ⊗ · · · ⊗ Hn)
defined by fT (a1, . . . , an) := a1 T1 + · · · + an Tn is an isometry.
4
RAJEEV GUPTA AND MD. RAMIZ REZA
Proof. Since T ⊂ σ(Ti) and Ti is a contraction for i = 1, . . . , n, it follows that T ⊂ ∂σ(Ti) for
i = 1, . . . , n. From (cf. [Con90, Proposition 6.7, Page 210]), we have T ⊂ σa(Ti) for i = 1, . . . , n,
where σa(Ti) is the approximate point spectrum of Ti. Thus for any i ∈ {1, . . . , n} and λ ∈ T,
there exists a sequence of unit vectors (xi
m)m∈N in Hi such that
m)k −→ 0 as m −→ ∞.
k(Ti − λ)(xi
Now, applying the Cauchy-Schwarz's inequality, we have
h(Ti − λ)(xi
m), (xi
m)i ≤ k(Ti − λ)(xi
= k(Ti − λ)(xi
m)kk(xi
m)k
m)k −→ 0
m)i −→ λ as m −→ ∞. Let (a1, . . . , an) be any vector in ℓ1(n)
as m −→ ∞. Hence hTi(xi
such that none of its co-ordinates is zero. Let λ1 = e−i arg(a1), λ2 = e−i arg(a2), . . . , λn = e−i arg(an).
Now for each i ∈ {1, . . . , n}, we have (xi
m), (xi
m)m∈N, a sequence of unit vectors from Hi, such that
m)i −→ λi as m −→ ∞.
hTi(xi
m), (xi
As m goes to ∞, we have
(cid:12)(cid:12)h(a1T1 ⊗ I⊗(n−1) + · · · + anI⊗(n−1) ⊗ Tn)(x1
=(cid:12)(cid:12)a1hT1(x1
= a1 + · · · + an = k(a1, . . . , an)k1.
m))i + · · · + anhTn(xn
m), (x1
m), (xn
m), (x1
m ⊗ · · · ⊗ xn
m ⊗ · · · ⊗ xn
m)i(cid:12)(cid:12)
m))i(cid:12)(cid:12) −→(cid:12)(cid:12)a1λ1 + · · · + anλn(cid:12)(cid:12)
Hence ka1 T1 + · · · + an Tnk ≥ k(a1, . . . , an)k1. Also
ka1 T1 + · · · + an Tnk ≤ a1kT1k + · · · + ankTnk.
Hence ka1 T1 + · · · + an Tnk = k(a1, . . . , an)k1, proving that fT is an isometry.
If some of the co-ordinates in the vector (a1, . . . , an) are zero, the same argument, as above,
(cid:3)
remains valid after dropping those co-ordinates.
An adaptation of the technique involved in the proof of Theorem 3.1, also proves the following
theorem.
Theorem 3.2. For i = 1, . . . , n, let Ti be a contraction on a Hilbert space Hi and T ⊆ σ(Ti).
Denote Ti = T1 ⊗ · · · ⊗ Ti ⊗ IHi+1⊗···⊗Hn and T = ( T1, . . . , Tn). Then, the function
gT : ℓ1(n) → B(H1 ⊗ · · · ⊗ Hn)
defined by gT (a1, . . . , an) := a1 T1 + · · · + an Tn is an isometry.
Remark 3.3. We show that all the operator spaces induced by the isometries defined in Theorem
3.1 are completely isometric to the MIN structure.
Suppose T1, . . . , Tn are contractions on Hilbert spaces H1, . . . , Hn respectively with the prop-
erty that T ⊆ σ(Ti) for i = 1, . . . , n. Denote T1 = T1 ⊗ IH2 ⊗ · · · ⊗ IHn, . . . , Tn = IH1 ⊗ · · · ⊗
IHn−1 ⊗Tn. Then the map fT defined as in the Theorem 3.1 is an isometry. The dilation theorem
due to Sz.-Nagy (cf.
[Pau02, Theorem 1.1, Page 7]), gives unitary maps Uj : Kj → Kj, dilat-
ing the contraction Tj, for j = 1, . . . , n. The operator space structure defined by the isometry
g : ℓ1(n) → B(K1 ⊗ · · ·⊗ Kn), where g(a1, . . . , an) = a1U1 ⊗ IK2⊗···⊗Kn + · · ·+ anIK1⊗···⊗Kn−1 ⊗ Un,
is no lesser than that of fT . Since U1, . . . , Un are unitary maps, C∗−algebra generated by
U1 ⊗ IK2⊗···⊗Kn, . . . , IK1⊗···⊗Kn−1 ⊗ Un is commutative. From (cf. [Pis03, Proposition 1.10, Page
24]), we conclude that g is a complete isometry.
It can similarly be shown that all the operator spaces induced by the isometries defined in
Theorem 3.2 are completely isometric to the MIN structure.
OPERATOR SPACE STRUCTURES ON ℓ1(n)
5
3.1. Operator space structures on ℓ1(n) different from the MIN structure. Parrott
[Par70] provides an example of a contractive homomorphism on A(D3) which is not completely
(Here A(D3) is the closure, with respect to the supremum norm on D3, of the
contractive.
polynomial in 3 complex variables.) Using a triple (I, U, V )(defined below) of 2 × 2 unitaries,
it was shown in [Mis94] that examples due to Parrott may be easily thought of as examples of
linear contractive maps on ℓ1(3) which are not completely contractive. Indeed this realization
shows that the operator space structure on ℓ1(3) can not be unique. In this section, using the
example from [Mis94], we give an explicit operator space structure k · kos on ℓ1(3), which is not
completely isometric to the MIN structure
Consider the following 2 × 2 unitary operators:
√3
2
2
√3
2 − 1
0 1 (cid:19) , U := 1
I2 =(cid:18) 1 0
2 ! and V := 1
2 −
√3
2
√3
2
2 ! .
1
It is clear that the map h : ℓ1(3) → M2, defined by h(z1, z2, z3) = z1I + z2U + z3V, is of norm
at most 1. The computations done in [Mis94] includes the following:
(3.1)
and
(3.2)
kI ⊗ I + U ⊗ U + V ⊗ V k = 3
sup
kz1I + z2U + z3V k < 3.
z1,z2,z3∈D
Choose a diagonal operator D on ℓ2(Z) such that kDk ≤ 1 and T ⊂ σ(D). Define
T1 :=(cid:20) I 0
0 D (cid:21) , T2 :=(cid:20) U 0
0 D (cid:21) , T3 :=(cid:20) V 0
0 D (cid:21)
and T1 = T1⊗I ⊗I, T2 = I ⊗ T2⊗I, T3 = I ⊗I ⊗ Tn. Let S1 := T1⊕I, S2 := T2⊕U, S3 := T3⊕V be
operators on a Hilbert space K. Define S : ℓ1(3) −→ B(K) by S(e1) = S1, S(e2) = S2, S(e3) =
S3 and extend it linearly.
From Theorem 3.1, we know that the function (z1, z2, z3) 7→ z1 T1 + z2 T2 + z3 T3 is an isometry
and since h is of norm at most 1, it follows that the map (z1, z2, z3) 7→ z1S1 + z2S2 + z3S3 is
also an isometry. Consequently, there is an operator space structure os on ℓ1(3) for which S is
a complete isometry. Also from (3.1), we have
kS1 ⊗ I + S2 ⊗ U + S3 ⊗ V k ≥ kI ⊗ I + U ⊗ U + V ⊗ V k = 3.
Thus k(I, U, V )kos = 3, as norm of (I, U, V ) is at most 3 under any operator space structure on
ℓ1(3). On the other hand, from (3.2), we have
k(I, U, V )kMIN = sup
kz1I + z2U + z3V k < 3.
z1,z2,z3∈D
It follows from [Pau02, Theorem 14.1] that if there is a map φ : (ℓ1(3), MIN) → (ℓ1(3), k · kos)
which is a complete isometry, then the identity I : (ℓ1(3), MIN) → (ℓ1(3), k · kos) must be also
a complete isometry. Therefore the operator space structure k · kos is different from the MIN
structure. However, although k(I, U, V )kMAX = 3, we are unable to decide whether the operator
space structure k · kos is completely isometric to the MAX operator space structure or not.
Acknowledgement: We are very grateful to G. Misra for several fruitful discussions and
suggestions. We are also thankful to the referee for many valuable suggestions.
References
[Con90] John B. Conway, A course in functional analysis, second ed., Graduate Texts in Mathematics, vol. 96,
Springer-Verlag, New York, 1990. MR 1070713 (91e:46001)
[Mis94] G. Misra, Completely contractive Hilbert modules and Parrott's example, Acta Math. Hungar. 63 (1994),
no. 3, 291 -- 303. MR 1261472 (96a:46097)
6
RAJEEV GUPTA AND MD. RAMIZ REZA
[Par70] Stephen Parrott, Unitary dilations for commuting contractions, Pacific J. Math. 34 (1970), 481 -- 490.
MR 0268710 (42 #3607)
[Pau92] Vern I. Paulsen, Representations of function algebras, abstract operator spaces, and Banach space geom-
etry, J. Funct. Anal. 109 (1992), no. 1, 113 -- 129. MR 1183607 (93h:46001)
[Pau02] Vern Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Advanced Mathe-
matics, vol. 78, Cambridge University Press, Cambridge, 2002. MR 1976867 (2004c:46118)
[Pis03] Gilles Pisier, Introduction to operator space theory, London Mathematical Society Lecture Note Series,
vol. 294, Cambridge University Press, Cambridge, 2003. MR 2006539 (2004k:46097)
Rajeev Gupta: Department of Mathematics, Indian Institute of Science, Bangalore
E-mail address: [email protected]
Md. Ramiz Reza: Department of Mathematics, Indian Institute of Science, Bangalore
E-mail address: [email protected]
|
1307.4480 | 1 | 1307 | 2013-07-17T01:44:55 | s-Numbers of compact embeddings of function spaces on quasi-bounded domains | [
"math.FA",
"math.NA"
] | We prove asymptotic formulas for the behavior of approximation, Gelfand, Kolmogorov and Weyl numbers of Sobolev embeddings between Besov and Triebel-Lizorkin spaces defined on quasi-bounded domains. | math.FA | math |
s-Numbers of compact embeddings of function spaces on
quasi-bounded domains
Shun Zhang, Alicja Gąsiorowska∗
Abstract
We prove asymptotic formulas for the behavior of approximation, Gelfand, Kol-
mogorov and Weyl numbers of Sobolev embeddings between Besov and Triebel-
Lizorkin spaces defined on quasi-bounded domains.
Keywords: approximation numbers, Gelfand numbers, Kolmogorov numbers,
Weyl numbers, Sobolev embedding, quasi-bounded domain
1. Introduction
Today we have a good knowledge about asymptotic behavior of entropy
and approximation numbers of compact embeddings between spaces of Sobolev-
Besov-Hardy type. The results regarding asymptotic behavior of the entropy
numbers and the approximation numbers of the embeddings of spaces defined
on bounded domains are the oldest part of this theory. We owe M.S. Birman,
M.Z. Solomjak [2], D. Edmunds, H. Triebel [9], A. Caetano [3] and others the
estimates in this respect. The Gelfand, the Kolmogorov and the Weyl numbers
on bounded domains were studied by C. Lubitz [14], R. Linde [13], A. Caetano
[5, 6] and J. Vybíral [19]. But, there is a wider class of domains for which the
embeddings of spaces of Besov-Sobolev type can be compact. These are so called
quasi-bounded domains. The quasi-bounded domain may be unbounded, and
can be even a domain of infinite Lebesgue measure. Recently H.-G. Leopold
and L. Skrzypczak [12] gave necessary and sufficient conditions for compactness
of these embeddings and studied the asymptotic behavior of entropy numbers
for function spaces defined on quasi-bounded domains. In this article we inves-
tigate asymptotic behavior of s-numbers of these embeddings. Our approach is
based on wavelet decomposition of function spaces on domains introduced by
H. Triebel in 2008; cf.
In the case of the Weyl numbers our estimates
complement the earlier results even for bounded domains.
[18].
The concept of quasi-bounded domains is not new. Function spaces on these
domains were studied in the 60s of the last century by Clark and Hewgill; see [7,
8]. Overview of some of these results can be found in [1]. All these results assume
∗Corresponding author
Preprint submitted to Elsevier
July 21, 2018
a certain regularity of the boundary of the domain. The assumptions that we
make here on the regularity of the boundary (uniformly E-porous domain) are
weaker than the previously assumed.
2. Function spaces on quasi-bounded domain
We assume that the reader is acquainted with the definition and basic prop-
p,q(Rd), 0 < p, q ≤ ∞, s ∈ R and Triebel-Lizorkin
erties of Besov spaces Bs
p,q(Rd), 0 < p < ∞, 0 < q ≤ ∞, s ∈ R. We denote the continuous
spaces F s
embeddings between quasi-Banach spaces by ֒→ and also we use the common
notations As
p,q(Ω) with A = B or A = F , if it makes no difference.
p,q(Rd), As
Let Ω be an open and nonempty set in Rd. Such set is called a domain. The
domain Ω in Rd is called quasi-bounded if
lim
x∈Ω,x→∞
d(x, ∂Ω) = 0.
(1)
In particular, each bounded domain is quasi-bounded.
Definition 1. Let Ω be the domain in Rd, such that Ω 6= Rd and let 0 < p, q ≤
∞, s ∈ R with p < ∞ for F -spaces.
(i) Let
As
p,q(Ω) = {f ∈ D′(Ω) : f = gΩ for some distribution g ∈ As
p,q(Rd)},
f As
p,q(Ω) = inf gAs
p,q(Rd),
where the infimum is taken over all g ∈ As
p,q(Rd) with f = gΩ.
(ii) Let
p,q( ¯Ω) = {f ∈ As
p,q(Rd) : supp f ⊂ ¯Ω},
eAs
(iii) We define
p,q( ¯Ω)},
p,q(Rd)k,
p,q(Ω)k = inf kgAs
kf eAs
eBs
B0
Bs
p,q(Ω),
p,q(Ω),
p,q(Ω),
p,q( ¯Ω) with f = gΩ.
eAs
p,q(Ω) = {f ∈ D′(Ω) : f = gΩ for some distribution g ∈ eAs
where the infimum is taken over all g ∈ eAs
p,q(Ω) =
p,q(Ω) =
eF s
− 1(cid:19)+
σp = d(cid:18) 1
0 < p ≤ ∞, 0 < q ≤ ∞, s > σp,
1 < p < ∞, 1 ≤ q ≤ ∞, s = 0,
0 < p < ∞, 0 < q ≤ ∞, s < 0,
0 < p < ∞, 0 < q ≤ ∞, s > σp,q,
1 < p < ∞, 1 ≤ q ≤ ∞, s = 0,
0 < p < ∞, 0 < q ≤ ∞, s < 0,
σp,q = d(cid:18)
p,q(Ω),
p,q(Ω),
p,q(Ω),
− 1(cid:19)+
1
min(p, q)
,
p
if
if
if
if
if
if
¯Bs
¯F s
and
where
F 0
F s
, 0 < p, q ≤ ∞.
2
In this paper we consider only the so called uniformly E-porous domain
introduced by H. Triebel; cf. [18]. This assumption of the uniformly E-porous
property allows us to use the wavelet characterization of these function spaces.
Definition 2.
(i) A close set Γ ⊂ Rd is said to be porous if there exists a
number 0 < η < 1 such that one finds for any ball B(x, r) ⊂ Rd centered
at x ∈ Γ and of radius r with 0 < r < 1, a ball B(y, ηr) with
B(y, ηr) ⊂ B(x, r)
and B(y, ηr) ∩ Γ = ∅.
(ii) A close set Γ ⊂ Rd is said to be uniformly porous if it is porous and there
is a locally finite positive Radon measure µ on Rd such that Γ = supp µ
and for some constants C, c > 0
C h(r) ≤ µ(B(x, r)) ≤ c h(r),
for x ∈ Γ, 0 < r < 1,
where h : [0, 1] → R is a continuous strictly increasing function with h(0) =
0 and h(1) = 1 (the constants C, c are independent of x and r).
Definition 3. Let Ω be an open set in Rd such that Ω 6= Rd and Γ = ∂Ω.
(i) The domain Ω is said to be E-porous if there is a number 0 < η < 1 such
that one finds for any ball B(x, r) ⊂ Rd centered at x ∈ Γ and of radius r
with 0 < r < 1, a ball B(y, ηr) with
B(y, ηr) ⊂ B(x, r)
and B(y, ηr) ∩ ¯Ω = ∅.
(ii) The domain Ω is called uniformly E-porous if it is E-porous and Γ = ∂Ω
is uniformly porous.
In 2008 H. Triebel presented wavelet characterization of function spaces on
[18, Chapter 2 and 3]. By Theorem 3.23 in [18] there
E-porous domains; cf.
exists an isomorphism between spaces ¯Bs
last space is defined below.
p,q(Ω) and ℓq(cid:0)2j(s−d/p)ℓMj
p (cid:1), where the
Definition 4. Let 0 < p ≤ ∞, 0 < q ≤ ∞, {βj}∞
numbers and {Mj}∞
j=0 ∈ ¯N, j ∈ N0, ¯N = N ∪ {∞}. Then
j=0 be a sequence of positive
ℓq(βj ℓMj
xℓq(βjℓMj
p ) =(cid:26)x : x = {xj,l}j∈N0,l=1,...,Mj with
< ∞(cid:27)
p ) =(cid:18) ∞Xj=0
xj,lp(cid:19)q/p(cid:19)1/q
j(cid:18) MjXl=1
βq
with the obvious modifications if p = ∞ or q = ∞.
3
In [12] H.-G. Leopold and L. Skrzypczak introduced also the so called box
packing constant of a domain in Rd.
Let Ω ⊂ Rd be the domain such that Ω 6= Rd. We put
bj(Ω) = sup(cid:26)k :
Qj,mℓ ⊂ Ω,
k[ℓ=1
Qj,mℓ being pairwise disjoint dyadic cubes of size 2−j(cid:27),
for j = 0, 1, . . . and we put
b(Ω) = sup(cid:26)t ∈ R+ : lim sup
j→∞
bj(Ω)2−jt = ∞(cid:27).
Moreover it is proved in [12], that for any nonempty open set Ω ⊂ Rd we have
d ≤ b(Ω) ≤ ∞. If Ω is unbounded and is not quasi-bounded, then b(Ω) = ∞.
If Lebesgue measure Ω is finite, then b(Ω) = d. In particular if the domain
is bounded, then b(Ω) = d. If the domain is quasi-bounded and its Lebesgue
measure is infinite then b(Ω) can be finite and bigger than d. However there are
quasi-bounded domains such that b(Ω) = ∞.
Furthermore in [12] it is shown that
Mj ∼ bj(Ω).
(2)
The box packing constant helps to prove the conditions for the compactness
of Sobolev embeddings.
Theorem 5. Let Ω be an uniformly E-porous quasi-bounded domain in Rd and
let 1
, b(Ω) < ∞. Then
− 1
p∗ =(cid:18) 1
p2
p1(cid:19)+
¯Bs1
p1,q1 (Ω) ֒→ ¯Bs2
p2,q2(Ω)
(3)
is compact if
s1 − s2 − d(cid:18) 1
p1
−
1
p2(cid:19) >
b(Ω)
p∗ .
If the embedding (3) is compact and 1
If the embedding (3) is compact and 1
p∗ = 0 then s1 − s2 − d(cid:18) 1
p∗ > 0 then s1 − s2 − d(cid:18) 1
p1
p1
− 1
p2(cid:19) > 0.
p2(cid:19) ≥ b(Ω)
p∗ .
− 1
Now we will formulate the key definitions, where we denote by L(X, Y ) the
class of all linear continuous operators from X into Y .
Definition 6. Let X and Y be quasi-Banach spaces and T ∈ L(X, Y ).
4
(i) For k ∈ N, we define the k-th approximation number by
ak(T ) := inf{kT − Ak : A ∈ L(X, Y ), rank(A) < k},
where rank(A) denotes the dimension of the range A(X) = {A(x), x ∈ X}.
(ii) For k ∈ N, we define the k-th Gelfand number by
ck(T ) := inf{kT J X
M k : M ⊂ X, codim(M ) < k},
where J X
of the quasi-Banach space X.
M is the natural injection of M into X and M is a closed subspace
(iii) For k ∈ N, we define the k-th Kolmogorov number by
dk(T ) := inf{kQY
N T k : N ⊂ Y, dim(N ) < k},
where QY
N is the natural surjection of Y onto the quotient space
Y /N = {y + N : ky + N k = inf
z∈N
ky + zk}
and N is a closed subspace of the quasi-Banach space Y .
(iv) For k ∈ N, we define the k-th Weyl number by
xk(T ) := sup{ak(T S) : S ∈ L(ℓ2, X) with kSk ≤ 1}.
The approximation numbers, the Gelfand, the Kolmogorov and the Weyl
numbers in the context of Banach spaces are examples of s-numbers, that were
introduced by A. Pietsch; cf.
[15]. In [6] we can find definition of s-numbers
extended to the context of quasi-Banach spaces.
Now we recall the well known properties of the above s-numbers.
Proposition 7. Let sk ∈ {ak, ck, dk, xk}, W, X, Y be quasi-Banach spaces and
let Z be a t-Banach space, t ∈ (0, 1]. Then
1. kT k = s1(T ) ≥ s2(T ) ≥ s3(T ) ≥ . . . ≥ 0
2. (additivity)
for all T ∈ L(X, Y ),
st
n+k−1(T1 + T2) ≤ st
for all T1, T2 ∈ L(X, Z) and n, k ∈ N,
k(T1) + st
n(T2)
3. (multiplicativity)
sn+k−1(T1T2) ≤ sk(T1)sn(T2)
for all T1 ∈ L(X, Y ), T2 ∈ L(W, X) and n, k ∈ N,
4. xk(T ) ≤ ck(T ) ≤ ak(T ),
dk(T ) ≤ ak(T )
for all T ∈ L(X, Y ) and k ∈ N.
Estimates of the Weyl numbers of embeddings of finite dimensional ℓN
p , N ∈
N, will be very useful for us therefore we recall the already known estimates.
Lemma 8 and Lemma 10 can be found in Caetano [4]. Lemma 9 is taken from
Konig [10, p. 186], see also in [4].
5
Lemma 8. Let 0 < p1 ≤ max(2, p2) ≤ ∞ and 1 ≤ k ≤ N/2, N ∈ N. Then
xk(id : ℓN
p1 → ℓN
p2 ) ∼
1
p2
− 1
p1 ,
1
2 − 1
1
p2
p1 ,
− 1
p1 ,
k
1,
k
N
if
if
if
if
0 < p1 ≤ p2 ≤ 2,
2 ≤ p1 ≤ p2 ≤ ∞,
0 < p1 ≤ 2 ≤ p2 ≤ ∞,
0 < p2 ≤ p1 ≤ 2.
(4)
(5)
(7)
Lemma 9. Let 2 ≤ p2 < p1 ≤ ∞. Then there is a positive constant C indepen-
dent of N and k such that for all k, N ∈ N
xk(cid:0)id : ℓN
p1 → ℓN
p2(cid:1) ≤ C (N/k)(1/p2−1/p1)/(1−2/p1) .
Lemma 10. Let 0 < p2 ≤ 2 < p1 ≤ ∞. Then there is a positive constant C
independent of N and k such that for all k, N ∈ N
xk(id : ℓN
p1 → ℓN
p2) ≥ C( N
N
1
p2
1
p2
− 1
2 ,
− 1
p1 ,
if
if
k ≤ N/2,
2
k ≤ N
p1 ,
(6)
Lemma 11. Let 2 < p2 < p1 ≤ ∞. Then there is a positive constant C
independent of N and k such that for all k, N ∈ N
xk(cid:0)id : ℓN
p1 → ℓN
p2(cid:1) ≥ C if k ≤ N/4.
Proof. Using the multiplicativity of the Weyl numbers, together with Lemma 8
and the hypothesis k ≤ N
4 , we can write
C1 ≤ x2k(cid:0)id : ℓN
2 → ℓN
p1 → ℓN
p2(cid:1) ≤ xk(cid:0)id : ℓN
≤ C2xk(cid:0)id : ℓN
p2(cid:1) xk(cid:0)id : ℓN
p2(cid:1) ,
p1 → ℓN
2 → ℓN
p1(cid:1)
which completes the proof.
The following result was already proved in the case 1 ≤ p2 ≤ 2 < p1 ≤ ∞;
see [10, p. 186].
Lemma 12. Let 0 < p2 ≤ 2 < p1 ≤ ∞. Then there is a positive constant C
independent of N and k such that for all k, N ∈ N
xk(cid:0)id : ℓN
p1 → ℓN
p2(cid:1) ≤ CN
1
p2 k− 1
2 .
(8)
Proof. Using the multiplicativity of the Weyl numbers, together with Lemma 8,
Lemma 9 and the hypothesis k ≤ N
2 , we can write
x2k(cid:0)id : ℓN
p1 → ℓN
p2(cid:1) ≤ xk(cid:0)id : ℓN
p2 k− 1
≤ CN
2 → ℓN
2 (N/k)(1/2−1/p1)/(1−2/p1) ≤ CN
p2(cid:1) xk(cid:0)id : ℓN
p1 → ℓN
2(cid:1)
1
1
p2 k− 1
2 ,
which finishes the proof, by the monotonicity of the Weyl numbers.
6
The following lemma for Gelfand numbers will be useful for the upper esti-
mates of Weyl numbers.
Lemma 13. If 1 ≤ k ≤ N < ∞ and 0 < p2 ≤ p1 ≤ ∞, then
ck(cid:0)id : ℓN
p1 → ℓN
p2(cid:1) = (N − k + 1)
1
p2
− 1
p1 .
The proof of this lemma follows literally [16, Section 11.11.4], see also [17].
Indeed the original proof is used only to deal with the Banach setting. However,
the same proof works also in the quasi-Banach setting 0 < p2 ≤ p1 ≤ ∞.
Based on the definitions of Weyl and Gelfand numbers, it is known that for
any linear continuous operator T between two complex quasi-Banach spaces,
xk(T ) ≤ ck(T ). Then in terms of Lemma 9, Lemma 12 and Lemma 13, we have
the following two propositions.
Proposition 14. Let 0 < p2 ≤ 2 < p1 ≤ ∞. Then there is a positive constant
C independent of N and k such that for all k, N ∈ N
xk(cid:0)id : ℓN
p1 → ℓN
p2(cid:1) ≤ C(N
N
− 1
1
p1 ,
p2
p2 k− 1
2 ,
1
if 1 ≤ k ≤ N
2
p1 ,
p1 ≤ k ≤ N.
2
if N
(9)
Proposition 15. Let 2 ≤ p2 < p1 ≤ ∞. Then there is a positive constant C
independent of N and k such that for all k, N ∈ N
xk(cid:0)id : ℓN
p1 → ℓN
1
p2
− 1
p1 ,
p2(cid:1) ≤ C(N
(cid:0) N
k(cid:1)(1/p2−1/p1)/(1−2/p1)
Of course when k > N then
if 1 ≤ k ≤ N
2
p1 ,
,
if N
2
p1 ≤ k ≤ N.
(10)
xk(id : ℓN
p1 → ℓN
p2 ) = 0.
3. The embeddings of sequence spaces
In this part we consider the embeddings of sequence spaces from Definition 4,
In [11] we find the following
that will play an important role in this paper.
theorem.
Theorem 16. Let 0 < p1, p2 ≤ ∞, 0 < q1, q2 ≤ ∞. {Mj}∞
sequence of natural numbers and {βj}∞
embedding
j=0 be an arbitrary
j=0 be an arbitrary weight sequence. The
id : ℓq1 (βjℓMj
p1 ) → ℓq2(ℓMj
p2 )
is compact if and only if
or
j M (1/p2−1/p1)+
j
(cid:8)β−1
j=0 ∈ ℓq∗
(cid:9)∞
if
q∗ < ∞
lim
j→∞
(β−1
j M (1/p2−1/p1)+
j
) = 0
if
q∗ = ∞,
7
We need also the following lemma which will be used in the proof of the next
where
1
q∗ =(cid:18) 1
q2
−
1
q1(cid:19)+
i.e.
q∗ =
theorem.
j
j M (1/p2−1/p1)+
Lemma 17. Let(cid:8)β−1
(cid:9)∞
idj x = (δj,kxk,l)k∈N0,l=1,...,Mk =(cid:26) 0,
p2 )(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:8)β−1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(id −
idj)xℓq2 (ℓMj
NXj=0
Then
∞,
(q1q2)/(q1 − q2),
q2,
if
if
if
0 < q1 ≤ q2 ≤ ∞,
0 < q2 < q1 < ∞,
q1 = ∞.
j=0 ∈ ℓq∗ and
if k 6= j,
if k = j and l = 1, . . . , Mj.
xj,l,
j M (1/p2−1/p1)+
j
j=N +1ℓq∗(cid:12)(cid:12)(cid:12)(cid:12)·xℓq1 (βjℓMj
(cid:9)∞
p1 ).
The proof can be found in [11].
Now we consider the asymptotic behavior of the Weyl, the Gelfand num-
bers and the approximation numbers for these embeddings. We start with the
estimates of the Weyl numbers.
Theorem 18. Let 0 < p1, p2 ≤ ∞, 0 < q1, q2 ≤ ∞ and 0 < b < ∞, δ >
b( 1
), Mj ∼ 2jb and that the operator
p2
id : ℓq1 (2jδℓMj
)+. Suppose λ = ( 1
p2
p1 ) → ℓq2 (ℓMj
p2 ) is compact.
)/(1 − 2
p1
− 1
p1
− 1
p1
(i) If 0 < p1 ≤ max(2, p2) ≤ ∞ or 0 < p2 ≤ 2 < p1 ≤ ∞, then
xk(id : ℓq1 (2jδℓMj
p1 ) → ℓq2 (ℓMj
p2 )) ∼ k−β,
where
β =
δ
,
− 1
p2
b + 1
p1
δ
b ,
δ
b + 1
− 1
2 ,
p1
p1
b + 1
2 ( δ
− 1
p2
p1
b + 1
2 − 1
,
p2
δ
if
if
if
if
if
0 < p1, p2 ≤ 2,
2 ≤ p1 ≤ p2 ≤ ∞,
0 < p1 ≤ 2 ≤ p2 ≤ ∞,
0 < p2 ≤ 2 < p1 ≤ ∞ and δ < b
p2
0 < p2 ≤ 2 < p1 ≤ ∞ and δ > b
p2
,
.
),
(ii) If 2 < p2 < p1 ≤ ∞ and δ > bλ, then
xk(id : ℓq1 (2jδℓMj
p1 ) → ℓq2 (ℓMj
p2 )) ∼ k− δ
b .
(iii) If 2 < p2 < p1 ≤ ∞ and δ < bλ, then
ck− δ
b ≤ xk(id : ℓq1 (2jδℓMj
p1 ) → ℓq2 (ℓMj
p2 )) ≤ Ck− p1
2 ( δ
b + 1
p1
− 1
p2
).
8
Proof. Step 1. Estimation from below. We consider the commutative diagram
ℓMj
p1
Sj
ℓq1(2jδℓMj
p1 )
id(j)
id
ℓMj
p2
Tj
ℓq2(ℓMj
p2 ),
where id(j) denotes the embedding from ℓMj
p1
projection, such that
in ℓMj
p2 . The operator Tj is a
Tjx = {δj,kxk,l}Mj
l=1,
and Sj is the natural embedding which maps {xl}Mj
ℓq1(2jδℓMj
p1 ),
l=1 to the j-th block in
Sj({xl}Mj
xv,l =(cid:26) 0,
xl,
l=1) = (xv,l), where
if v 6= j,
if v = j and 1 ≤ l ≤ Mj.
Then
and
Sj : ℓMj
p1 → ℓq1 (2jδℓMj
p1 ) = 2jδ,
Tj : ℓq2 (ℓMj
p2 ) → ℓMj
p2 = 1
id(j) = Tj ◦ id ◦Sj.
Consequently
xk(id(j) : ℓMj
p1 → ℓMj
p2 ) ≤ 2jδxk(id : ℓq1(2jδℓMj
p1 ) → ℓq2 (ℓMj
p2 )).
Let 0 < p1 ≤ max(2, p2) ≤ ∞. By (4) with k = [Mj/2], we obtain
xk(id) ≥ C
k−(δ/b+1/p1−1/p2),
k−δ/b,
k−(δ/b+1/p1−1/2),
if 0 < p1, p2 ≤ 2,
if 2 ≤ p1 ≤ p2 ≤ ∞,
if 0 < p1 ≤ 2 ≤ p2 ≤ ∞.
Let 0 < p2 ≤ 2 < p1 ≤ ∞.
If δ < b
p2
then by (6) with k = [M 2/p1
], we obtain
j
Ck−(p1/2)(δ/b+1/p1−1/p2) ≤ C2−jδ2jb(1/p2−1/p1) ≤ C2−jδM 1/p2−1/p1
j
≤ xk(id).
If δ > b
p2
then the formula (6) with k = [Mj/2], yields
Ck−(δ/b+1/2−1/p2) ≤ C2−jδM 1/p2−1/2
j
≤ xk(id).
Let 2 < p2 < p1 ≤ ∞. By (7) with k = [Mj/4], we obtain
Ck−δ/b ≤ C2−jδ ≤ xk(id).
9
/
/
o
o
Step 2. Estimation from above. Let idj be defined as in Lemma 17. Then
we have
where
idj)xℓq2 (ℓMj
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(id −
NXj=0
EN =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12){2−jδM (1/p2−1/p1)+
LXj=0
p2 )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ EN(cid:12)(cid:12)(cid:12)(cid:12)xℓq1 (2jδℓMj
p1 )(cid:12)(cid:12)(cid:12)(cid:12) ,
j=N +1ℓq∗(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12).
NXj=L+1
p1 ) → ℓq2(ℓMj
p2 )) ≤ Eρ
(idj) +
N +
xρ
kj
j
}∞
Let ρ = min(1, p2, q2), then lq2 (lMj
properties of the Weyl numbers, we get
p2 ) is a ρ-Banach space. Therefore from the
xρ
k(id : ℓq1(2jδℓMj
xρ
kj
(idj),
(11)
where
k =
NXj=0
kj − (N + 1).
(12)
We will choose N later.
Substep 2.1. We consider the commutative diagram
ℓq1 (2jδℓMj
p1 )
idj
ℓq2(ℓMj
p2 )
Tj
Sj
ℓMj
p1
id(j)
ℓMj
p2 ,
where Tj and Sj are defined similarly to Tj and Sj in Step 1. Now
Tj : ℓq1 (2jδℓMj
p1 ) → ℓMj
pj = 2−jδ
Therefore idj = Sj ◦ id(j) ◦Tj and
and Sj : ℓMj
p2 → ℓq2(ℓMj
p2 ) = 1.
xkj (idj) ≤ 2−jδxkj (id(j) : ℓMj
p1 → ℓMj
p2 ).
(13)
Consequently we get from (11)
xρ
k(id : ℓq1(2jδℓMj
p1 ) → ℓq2(ℓMj
p2 ))
≤ Eρ
N +
LXj=0
2−jδρxρ
kj
(id(j)) +
NXj=L+1
2−jδρxρ
kj
(id(j)),
with arbitrary L, N and kj satisfying (12).
Substep 2.2. Now we estimate the sumPL
j=0 xρ
kj
(idj).
10
/
/
o
o
Let 0 < p1 ≤ max(2, p2) ≤ ∞. First we show that there exists a constant
C > 0 independent of N and k, such that
xk(id : ℓN
p1 → ℓN
p2 ) ≤ C
k1/p2−1/p1,
1,
k1/2−1/p1 ,
N 1/p2−1/p1,
if
if
if
if
0 ≤ p1 ≤ p2 ≤ 2,
2 ≤ p1 ≤ p2 ≤ ∞,
0 < p1 ≤ 2 ≤ p2 ≤ ∞,
0 < p2 ≤ p1 ≤ 2,
(14)
for k ≤ N . We consider the commutative diagram
ℓN
p1
S /
ℓ2N
p1
id
Id
ℓN
p2
T
ℓ2N
p2 ,
where
S(λ1, . . . , λN ) = (λ1, . . . , λN , 0, . . . , 0) and
T (λ1, . . . , λ2N ) = (λ1, . . . , λN ).
Both norms S and T are equal to 1. Therefore xk(id) ≤ xk(Id), and then
by (4) we get (14).
Let
kj = [Mj2(L−j)ǫ]
for
j = 0, 1, . . . , L,
with 0 < ǫ < b. Then
LXj=0
We put
kj ≤
LXj=0
t =
[c2jb2(L−j)ǫ] ≤ [c2Lb]
LXj=0
2(j−L)(b−ǫ) ≤ [c2Lb].
(15)
1/p2 − 1/p1,
0,
1/2 − 1/p1,
if
if
if
0 < p1, p2 ≤ 2,
2 ≤ p1 ≤ p2 ≤ ∞,
0 < p1 ≤ 2 ≤ p2 ≤ ∞.
(16)
(17)
So by (13) and (14) we get
xkj (idj) ≤ C2−jδM t
j 2(L−j)ǫt,
if kj ≤ Mj and
xkj (idj) = 0,
if kj > Mj. Moreover if j < L − 1
xkj (idj) = 0. Therefore by (17) we have:
ǫ , then kj = [Mj2(L−j)ǫ] > Mj and then
11
/
o
o
for t < 0
xρ
kj
(idj) =
LXj=0
LXj=⌈L−1/ǫ⌉
xρ
kj
(idj)
LXj=⌈L−1/ǫ⌉
LXj=⌈L−1/ǫ⌉
≤ c(2−Lδ2Lbt)ρ
(2(L−j)δ2−(L−j)bt2(L−j)ǫt)ρ
≤ c2−Lbρ(δ/b−t)
(2δ/ǫ2−bt/ǫ2(L−j)ǫt)ρ ≤ C2−Lbρ(δ/b−t);
(18)
for t = 0
xρ
kj
(idj) =
LXj=0
LXj=⌈L−1/ǫ⌉
xρ
kj
(idj) ≤ c2−Lρδ
LXj=⌈L−1/ǫ⌉
2(L−j)ρδ
≤ c2−Lρδ
LXj=⌈L−1/ǫ⌉
2ρδ/ǫ ≤ c2−Lρδ 1
ǫ
2ρδ/ǫ ≤ C2−Lρδ;
(19)
for t > 0
xρ
kj
(idj) =
LXj=0
LXj=⌈L−1/ǫ⌉
xρ
kj
(idj) ≤ c(2−Lδ2Lbt)ρ
LXj=⌈L−1/ǫ⌉
(2(L−j)δ2−(L−j)bt)ρ
≤ c2−Lbρ(δ/b−t)
(2δ/ǫ2−(L−j)bt)ρ ≤ C2−Lbρ(δ/b−t);
(20)
LXj=⌈L−1/ǫ⌉
where ⌈x⌉ = inf{k ∈ N : k ≥ x} and constant C is dependent on ǫ, δ, t but not
on L.
Let 0 < p2 ≤ 2 < p1 ≤ ∞.
b < 1
p2
then we take
If δ
kj =(cid:2)M 2/p1
b < 1
p2
j
2(L−j)ǫ(cid:3)
for
j = 0, 1, . . . , L,
with ǫ > 0. Since δ
So we have
we can choose ǫ > 0, such that δ
b + 1
p1
− 1
p2
< ǫ
2b < 1
p1
.
kj ≤
LXj=0
LXj=0
[c2jb(2/p1)2(L−j)ǫ] ≤ [c2Lb(2/p1)]
LXj=0
2b(j−L)(2/p1−ǫ/b) ≤ [c2Lb(2/p1)].
(21)
Now by (13) and (9) we get
xkj (idj) ≤ C2−jδM 1/p2−1/p1
j
2−(L−j)(ǫ/2)
≤ C2−Lδ2Lb(1/p2−1/p1)2(L−j)δ2(L−j)b(1/p1−1/p2)2−(L−j)(ǫ/2),
(22)
12
since M 2/p1
j
≤ kj. Hence
xρ
kj
LXj=0
(idj) ≤ c2−Lbρ(δ/b+1/p1−1/p2)
LXj=0
2bρ(L−j)(δ/b+1/p1−1/p2−ǫ/(2b))
≤ C2−Lbρ(δ/b+1/p1−1/p2),
(23)
with constant C independent of L.
If δ
then we take
b > 1
p2
kj = [Mj2(L−j)ǫ]
for
j = 0, 1, . . . , L,
(24)
with a fixed ǫ, 0 < ǫ < b. Hence
kj ≤
LXj=0
LXj=0
[c2jb2(L−j)ǫ] ≤ [c2Lb]
LXj=0
2(j−L)(b−ǫ) ≤ [c2Lb].
(25)
If j < L − 1
ǫ then kj > Mj. Therefore from (13) and (9) we get
xρ
kj
(idj) =
LXj=0
LXj=⌈L−1/ǫ⌉
≤ c(2−Lδ2Lb(1/p2−1/2))ρ
Here
xρ
kj
(idj)
LXj=⌈L−1/ǫ⌉
(2(L−j)δ2−(L−j)b(1/p2−1/2)2−(L−j)(ǫ/2))ρ.
(2(L−j)δ2−(L−j)b(1/p2−1/2)2−(L−j)(ǫ/2))ρ
LXj=⌈L−1/ǫ⌉
So
LXj=0
Let 2 < p2 < p1 ≤ ∞.
If δ < bλ then we take
≤ 2ρδ/ǫ
LXj=⌈L−1/ǫ⌉
(2−(L−j)b(1/p2−1/2)2−(L−j)(ǫ/2))ρ ≤ C.
xρ
kj
(idj) ≤ C2−Lbρ(δ/b+1/2−1/p2).
(26)
kj = [M 2/p1
j
2(L−j)ε]
for j = 0, 1, . . . , L,
with ε > 0. Since δ < bλ we can choose ε > 0, such that 1
) = 2
We should mention that the equality 1
p1
if δ = bλ. Then
b + 1
p1
− 1
p2
λ ( δ
λ ( δ
− 1
p2
) < ε
b + 1
b < 2
.
p1
p1
holds true if and only
LXj=0
kj ≤ [c2Lb(2/p1)].
(27)
13
By means of (10) and (13), we have
xkj (idj) ≤ C2−jδM λ(1−2/p1)
j
2−(L−j)ελ ≤ C2−Lb( δ
b + 1
p1
− 1
p2
)2b(L−j)( δ
b + 1
p1
− 1
p2
− ελ
b ).
Therefore,
xρ
kj
(idj) ≤ C2−Lbρ( δ
b + 1
p1
− 1
p2
),
(28)
LXj=0
with constant C independent of L.
If δ > bλ then we take
kj = [Mj2(L−j)ε]
for j = 0, 1, . . . , L,
with fixed ε, 0 < ε < b. Then δ > λε and
kj ≤ [c2Lb].
LXj=0
If j < L − 1
ε , then kj > Mj. In view of (5) and (13), we have
(29)
(30)
xρ
kj
LXj=0
(idj) =
xρ
kj
(idj) ≤ C1
LXj=⌈L−1/ε⌉
(2−jδ2−(L−j)λε)ρ
≤ C12−Lδρ
(2(L−j)(δ−λε))ρ ≤ C22−Lδρ2(δ−λε)ρ/ε ≤ C32−Lδρ.
LXj=⌈L−1/ε⌉
LXj=⌈L−1/ε⌉
Substep 2.3. Now we estimate the sumPN
that EN is small enough.
Consider the first case 0 < p1 ≤ max(2, p2) ≤ ∞. We show that
j=L+1 xρ
kj
(idj) and choose N such
NXj=L+1
where t is defined in (16). Let
xρ
kj
(idj) ≤ C2−Lbρ(δ/b−t),
(31)
kj = max(cid:8)(cid:2)ML(j − L)−2(cid:3), 1(cid:9) for
Then
j = L + 1, . . . , N.
kj ≤ cML + (N − L).
(32)
NXj=L+1
If t ≤ 0 then kt
j ≤ (ML(j − L)−2)t. Therefore from (13) and (14) we get
NXj=L+1
xρ
kj
(idj) ≤ c(2−Lδ2Lbt)ρ
NXj=L+1
(2(L−j)δ(j − L)−2t)ρ ≤ C2−Lbρ(δ/b−t).
(33)
14
If t > 0 and since δ > b( 1
p2
− 1
p1
)+ = b
t then from (13) and (14) we obtain
NXj=L+1
xρ
kj
(idj) ≤ c(2−Lδ2Lbt)ρ
NXj=L+1
(2(L−j)δ2−(L−j)bt)ρ ≤ C2−Lbρ(δ/b−t),
(34)
where C is independent of L and N .
Now we can choose N such that EN ≤ c2−LδM t
L. This is possible because
{EN }∞
N =1 is a decreasing sequence such that limN→∞ EN = 0. The formulas
(12), (15) and (32) imply that k ≤ [cML]. So by (11), (18), (19), (20), (33) and
(34) we have x[cML](id) ≤ C2−Lb(δ/b−t). Finally
xk(id) ≤ Ck−(δ/b−t),
for all k ∈ N.
Consider the next case 0 < p2 ≤ 2 < p1 ≤ ∞.
b < 1
p2
then we take
If δ
Hence
L
kj = max(cid:8)(cid:2)M 2/p1
NXj=L+1
(j − L)−2(cid:3), 1(cid:9) for
j = L + 1, . . . , N.
kj ≤ cM 2/p1
L + (N − L).
(35)
Now by (9), (13) and since δ > b( 1
p2
− 1
p1
)+ we have
(idj) ≤ C(2−Lδ2Lb(1/p2−1/p1))ρ
NXj=L+1
xρ
kj
≤ c2−Lbρ(δ/b+1/p1−1/p2)
(2(L−j)δ2(L−j)b(1/p1−1/p2))ρ
NXj=L+1
NXj=L+1
2bρ(L−j)(δ/b+1/p1−1/p2) ≤ c2−Lbρ(δ/b+1/p1−1/p2).
(36)
As in the previous case we can choose N such that EN ≤ c2−LδML
The formulas (12), (21) and (35) imply k ≤ [cM 2/p1
we have x[cM 2/p1
](id) ≤ C2−Lb(δ/b+1/p1−1/p2). Finally
1/p2−1/p1 .
]. So by (11), (23) and (36)
L
L
xk(id) ≤ Ck−(p1/2)(δ/b+1/p1−1/p2),
for all k ∈ N.
But if we assume δ
b > 1
p2
, then taking
kj = max(cid:8)[ML(j − L)−2], 1(cid:9) for
we get
j = L + 1, . . . , N,
NXj=L+1
kj ≤ cML + (N − L).
(37)
15
We notice that if 1 ≤ kj ≤ M 2/p1
and (13) we have
j
then M 1/p2−1/p1
j
≤ M 1/p2
j
k−1/2
j
. So by (9)
NXj=L+1
xρ
kj
(idj) ≤
NXj=L+1
c(2−jδ2(bj)/p2 2−(bL)/2(j − L))ρ
≤ c2−Lbρ(δ/b−1/p2+1/2)
(2b(L−j)(δ/b−1/p2)(j − L))ρ
NXj=L+1
≤ C2−Lbρ(δ/b−1/p2+1/2).
(38)
1/p2−1/2. From (12), (25), (37) we
Now we choose N such that EN ≤ c2−LδML
have k ≤ [cML]. So by (11), (26), (38) we get x[cML](id) ≤ C2−Lb(δ/b+1/2−1/p2).
Then
xk(id) ≤ Ck−(δ/b+1/2−1/p2),
for all k ∈ N.
Consider the last case 2 < p2 < p1 ≤ ∞.
If δ < bλ then we take
kj = max{[M 2/p1
L
(j − L)−2], 1} for j = L + 1, . . . , N.
Hence, kj ≤ M 2/p1
L ≤ M 2/p1
j
and
NXj=L+1
kj ≤ cM 2/p1
L + (N − L).
(39)
Based on (10), (13) and the previous assumption δ > b( 1
p2
− 1
p1
)+, we have
xρ
kj
(idj) ≤ C
(2−jδ2jb( 1
p2
NXj=L+1
− 1
p1
))ρ ≤ C2−Lbρ( δ
b + 1
p1
− 1
p2
).
(40)
We choose N such that EN ≤ c2−LδM
j=0 kj − (N + 1) ≤ [cM 2/p1
L
− 1
p1
1
p2
. The formulas (27) and (39) imply
L
]. Finally, by (11), (28) and (40), we have
NXj=L+1
that k =PN
xk(id) ≤ Ck− p1
2 ( δ
b + 1
p1
− 1
p2
),
for all k ∈ N.
If δ > bλ then we take
kj = max{[ML(j − L)−2], 1} for j = L + 1, . . . , N,
and we get
NXj=L+1
kj ≤ cML + (N + L).
(41)
16
By (5) and (13), we have
NXj=L+1
xρ
kj
(idj) ≤ C
NXj=L+1
≤ C2−Lρδ
NXj=L+1
(2−jδ2jbλ2−Lbλ(j − L)2λ)ρ
2ρ(L−j)(δ−bλ)(j − L)2ρλ ≤ C2−Lρδ.
(42)
We choose N such that EN ≤ c2−Lδ. The formulas (29) and (41) yield that
j=0 kj − (N + 1) ≤ [cML]. Then, by (11), (30) and (42), we have
k =PN
xk(id) ≤ Ck− δ
b ,
for all k ∈ N.
b = p1
Remark 1. Still, there are minor gaps left open in the estimates for the case
δ < bλ. It should be mentioned that, for the bounds of the asymptotic order,
the inequality δ
) holds if and only if δ = bλ. Furthermore,
the problem becomes much more complicated if δ = bλ. How do the Weyl
numbers behave in such limiting situation? Different from those previous cases,
the corresponding behaviour herein maybe depend on the parameters q1 and q2
to a certain extent.
b + 1
p1
− 1
p2
2 ( δ
The asymptotic behavior of the approximation, the Gelfand and the Kol-
mogorov numbers can be found in [19, Theorem 3.5, 4.12 and 4.6]. Since a bit
different notation is used there we recall here the results using our symbols.
Some partial results were proved earlier in [3].
Theorem 19. Let
− 1
δ > b( 1
p1
p2
p1 ) → ℓq2 (ℓMj
id : ℓq1(2jδℓMj
p =
p2 ) is compact. Then
)+. Suppose Mj ∼ 2jb, 1
0 < p1, p2 ≤ ∞, 0 < q1, q2 ≤ ∞ and 0 < b < ∞,
1,p2) and that the operator
1
min(p′
ak(id : ℓq1 (2jδℓMj
p1 ) → ℓq2 (ℓMj
p2 )) ∼ k−β,
δ
b ,
p
δ
b ,
2
δ
δ
b + 1
2 − 1
p ,
b + 1
− 1
p1
p2
b + 1
2 − 1
,
p2
δ
if
if
if
if
if
0 < p1 ≤ p2 ≤ 2 or 2 ≤ p1 ≤ p2 ≤ ∞,
0 < p1 < 2 < p2 < ∞ and δ < b
p or
1 < p1 < 2 < p2 = ∞ and δ < b
p ,
0 < p1 < 2 < p2 ≤ ∞ and δ > b
p ,
0 < p2 ≤ p1 ≤ ∞,
0 < p1 ≤ 1 < p2 = ∞.
,
(43)
where
β =
Theorem 20. Let 0 < p1, p2 ≤ ∞, 0 < q1, q2 ≤ ∞ and 0 < b < ∞, δ >
b( 1
p2
id : ℓq1 (2jδℓMj
)+. Suppose Mj ∼ 2jb, θ =(cid:0) 1
2(cid:1) and that the operator
p2(cid:1)/(cid:0) 1
p2 ) is compact. Then
p1 ) → ℓq2 (ℓMj
− 1
p1
− 1
− 1
p1
p1
17
where
β =
ck(id : ℓq1 (2jδℓMj
p1 ) → ℓq2 (ℓMj
p2 )) ∼ k−β,
δ
b ,
b + 1
δ
p1
− 1
p2
,
p′
1
2
δ
b ,
δ
b + 1
p1
− 1
2 ,
if
if
if
if
2 ≤ p1 < p2 ≤ ∞,
0 < p2 ≤ p1 ≤ ∞ or
0 < p1 < p2 ≤ 2 and δ
b > θ
b < θ
1 < p1 < p2 ≤ 2 and δ
p′
1
1 < p1 < 2 < p2 ≤ ∞ and δ
0 < p1 < 2 < p2 ≤ ∞ and δ
p′
1
,
or
b < 1
b > 1
p′
1
p′
1
,
.
Theorem 21. Let 0 < p1, p2 ≤ ∞, 0 < q1, q2 ≤ ∞ and 0 < b < ∞, δ >
b( 1
p2
id : ℓq1 (2jδℓMj
)+. Suppose Mj ∼ 2jb, θ′ =(cid:0) 1
p2(cid:1) and that the operator
p2(cid:1)/(cid:0) 1
p2 ) is compact. Then
p1 ) → ℓq2 (ℓMj
2 − 1
− 1
p1
− 1
p1
where
β =
dk(id : ℓq1(2jδℓMj
p1 ) → ℓq2(ℓMj
p2 )) ∼ k−β,
δ
b ,
b + 1
δ
p2
δ
b ,
2
2 − 1
p2
,
δ
b + 1
p1
− 1
p2
,
if
if
if
if
0 < p1 ≤ p2 ≤ 2,
0 < p1 < 2 < p2 ≤ ∞ and δ
0 < p1 < 2 < p2 < ∞ and δ
b < θ ′
2 < p1 ≤ p2 ≤ ∞ and δ
p2
b > θ ′
2 < p1 ≤ p2 ≤ ∞ and δ
0 < p2 ≤ p1 ≤ ∞.
p2
b > 1
p2
b < 1
p2
,
or
,
or
4. The embeddings of function spaces on quasi-bounded domains
In this part, we use the estimates from the previous section to obtain the
asymptotic behavior of the s-numbers of the following embeddings
¯As1
p1,q1 (Ω) ֒→ ¯As2
p2,q2 (Ω),
defined on the quasi-bounded domains. Theorem 3.23 in [18] guarantees that
sk( ¯As1
p1,q1 (Ω) ֒→ ¯As2
p2,q2 (Ω)) ∼ sk(ℓq1 (2jδℓMj
p1 ) ֒→ ℓq2(ℓMj
p2 )),
for any s-numbers, if Ω is the uniformly E-porous domain described in Definition
3. If domain Ω is quasi-bounded and the assumptions of Theorem 5 are satisfied
then the embedding is compact. Moreover if
0 < lim inf
j→∞
bj(Ω)2−jb(Ω) ≤ lim sup
j→∞
bj(Ω)2−jb(Ω) < ∞,
(44)
then using (2) we get
Mj ∼ 2jb(Ω).
18
Using all the above remarks and the well known elementary embeddings
Bs
p,q1 (Rd) ֒→ F s
p,q(Rd) ֒→ Bs
p,q2 (Rd),
if q1 ≤ min(p, q), q2 ≥ max(p, q),
we get the following theorem.
Theorem 22. Let Ω be the quasi-bounded domain satisfying (44) and uniformly
E-porous in Rd with Ω 6= Rd. Let 0 < p1, p2 ≤ ∞, 0 < q1, q2 ≤ ∞, b = b(Ω) <
∞, δ = s1 − s2 − d( 1
),
p1
− 1
)/(1 − 2
p1
. Then
p∗ , 1
− 1
) > b(Ω)
− 1
p2
− 1
′
p∗ =(cid:0) 1
p2(cid:1)/(cid:0) 1
p2
2 − 1
− 1
p1(cid:1)+, λ = ( 1
p2(cid:1) and 1
p =
p2
− 1
p1
1
1,p2)
min(p
2(cid:1), θ′ =(cid:0) 1
p1
p2(cid:1)/(cid:0) 1
p1
θ =(cid:0) 1
(i)
p1
xk( ¯As1
p1,q1(Ω) ֒→ ¯As2
p2,q2 (Ω)) ∼ k−γ,
where
s1−s2
b + b−d
b ( 1
p1
b ( 1
b − d
− 1
p2
p1
− 1
p2
),
s1−s2
),
b( 1
p1
− 1
p1
s1−s2
b +
2(cid:0) s1−s2
b +
s1−s2
b + b−d
2 − 1
p2
b( 1
b
− 1
p2
2 )−d( 1
p1
b ( 1
− 1
p2
p1
− 1
p2
)−d( 1
p1
)
,
)(cid:1),
)
,
b
if
if
if
if
if
0 < p1, p2 ≤ 2,
2 ≤ p1 ≤ p2 ≤ ∞ or
2 < p2 < p1 ≤ ∞ and δ > bλ,
0 < p1 ≤ 2 ≤ p2 ≤ ∞,
0 < p2 ≤ 2 < p1 ≤ ∞ and δ
0 < p2 ≤ 2 < p1 ≤ ∞ and δ
b < 1
p2
b > 1
p2
,
.
γ =
(iii)
where
(ii) If 2 < p2 < p1 ≤ ∞ and δ < bλ, then
ck− s1−s2
b + d
b ( 1
p1
− 1
p2
) ≤ xk( ¯As1
p1,q1 (Ω) ֒→ ¯As2
p2,q2 (Ω))
≤ Ck− p1
2(cid:0) s1−s2
b + b−d
b ( 1
p1
− 1
p2
)(cid:1).
ak( ¯As1
p1,q1 (Ω) ֒→ ¯As2
p2,q2(Ω)) ∼ k−γ,
γ =
(iv)
)(cid:1),
b − d
b ( 1
p1
b ( 1
b − d
p1
− 1
),
p2
− 1
p2
s1−s2
p
2(cid:0) s1−s2
s1−s2
b( 1
2 − 1
p )−d( 1
p1
− 1
p2
)
,
b +
b + b−d
b ( 1
p1
2 − 1
b( 1
p2
b +
s1−s2
s1−s2
b
− 1
p2
)−d( 1
p1
),
− 1
p2
b
)
,
if
if
if
if
if
0 < p1 ≤ p2 ≤ 2 or 2 ≤ p1 ≤ p2 ≤ ∞,
0 < p1 < 2 < p2 < ∞ and δ
1 < p1 < 2 < p2 = ∞ and δ
0 < p1 < 2 < p2 ≤ ∞ and δ
0 < p2 ≤ p1 ≤ ∞,
b < 1
p or
b < 1
p ,
b > 1
p ,
0 < p1 ≤ 1 < p2 = ∞.
ck( ¯As1
p1,q1 (Ω) ֒→ ¯As2
p2,q2(Ω)) ∼ k−γ,
19
s1−s2
b ( 1
b − d
p1
b + b−d
b ( 1
p1
− 1
),
p2
− 1
p2
s1−s2
),
γ =
b − d
b ( 1
p1
− 1
p2
p′
1
2(cid:0) s1−s2
s1−s2
b +
b( 1
p1
− 1
2 )−d( 1
p1
b
if
if
if
if
)
,
2 ≤ p1 < p2 ≤ ∞,
0 < p2 ≤ p1 ≤ ∞ or
0 < p1 < p2 ≤ 2 and δ
b > θ
b < θ
1 < p1 < p2 ≤ 2 and δ
p′
1
1 < p1 < 2 < p2 ≤ ∞ and δ
0 < p1 < 2 < p2 ≤ ∞ and δ
p′
1
,
or
b < 1
b > 1
p′
1
p′
1
,
.
)(cid:1),
− 1
p2
dk( ¯As1
p1,q1(Ω) ֒→ ¯As2
p2,q2 (Ω)) ∼ k−γ,
where
(v)
where
b ( 1
p1
b( 1
2 − 1
p2
− 1
p2
),
)−d( 1
p1
s1−s2
b − d
s1−s2
b +
p2
2 ( s1−s2
b − d
b ( 1
p1
− 1
p2
)
,
)),
b
− 1
p2
γ =
s1−s2
b + b−d
b ( 1
p1
− 1
p2
),
if
if
if
if
0 < p1 ≤ p2 ≤ 2,
0 < p1 < 2 < p2 ≤ ∞ and δ
0 < p1 < 2 < p2 < ∞ and δ
b < θ ′
2 < p1 ≤ p2 ≤ ∞ and δ
p2
b > θ ′
2 < p1 ≤ p2 ≤ ∞ and δ
p2
0 < p2 ≤ p1 ≤ ∞.
b > 1
p2
b < 1
p2
,
or
,
or
The above formulas simplify if Ω is a domain of finite measure.
Corollary 23. Let Ω be uniformly E-porous in Rd of finite Lebesgue measure
and δ = s1 − s2 − d( 1
p denote the same as in
p1
Theorem 22. Then
)+ > 0. Let λ, θ, θ′ and 1
− 1
p2
(i)
where
γ =
xk( ¯As1
p1,q1(Ω) ֒→ ¯As2
p2,q2 (Ω)) ∼ k−γ,
s1−s2
,
d
s1−s2
d − 1
p1
+ 1
p2
,
s1−s2
d − 1
s1−s2
,
d + 1
p1
2
d
s1−s2
2 + 1
p2
2 − 1
p1
,
,
if
if
if
if
if
0 < p1, p2 ≤ 2,
2 ≤ p1 ≤ p2 ≤ ∞ or
2 < p2 < p1 ≤ ∞ and δ > dλ,
0 < p1 ≤ 2 ≤ p2 ≤ ∞,
0 < p2 ≤ 2 < p1 ≤ ∞ and δ
0 < p2 ≤ 2 < p1 ≤ ∞ and δ
d < 1
p2
d > 1
p2
,
.
(ii) If 2 < p2 < p1 ≤ ∞ and δ < dλ, then
ck− s1−s2
d + 1
p1
(iii)
− 1
p2 ≤ xk( ¯As1
p1,q1(Ω) ֒→ ¯As2
p2,q2 (Ω)) ≤ Ck− p1
2
s1 −s2
d
.
ak( ¯As1
p1,q1 (Ω) ֒→ ¯As2
p2,q2(Ω)) ∼ k−γ,
20
d − 1
p1
s1−s2
2 ( s1−s2
d − 1
p1
+ 1
,
p2
+ 1
p2
p
),
s1−s2
d + 1
2 − 1
p − 1
p1
+ 1
p2
,
s1−s2
,
d
s1−s2
d + 1
2 − 1
p1
,
if
if
if
if
if
0 < p1 ≤ p2 ≤ 2 or 2 ≤ p1 ≤ p2 ≤ ∞,
0 < p1 < 2 < p2 < ∞ and δ
1 < p1 < 2 < p2 = ∞ and δ
0 < p1 < 2 < p2 ≤ ∞ and δ
0 < p2 ≤ p1 ≤ ∞,
0 < p1 ≤ 1 < p2 = ∞.
d < 1
p or
d < 1
p ,
d > 1
p ,
ck( ¯As1
p1,q1 (Ω) ֒→ ¯As2
p2,q2(Ω)) ∼ k−γ,
s1−s2
d − 1
p1
+ 1
p2
,
s1−s2
,
d
p′
2 ( s1−s2
1
d − 1
p1
+ 1
p2
),
s1−s2
d − 1
2 + 1
p2
,
if
if
if
if
2 ≤ p1 < p2 ≤ ∞,
0 < p2 ≤ p1 ≤ ∞ or
d > θ
0 < p1 < p2 ≤ 2 and δ
d < θ
1 < p1 < p2 ≤ 2 and δ
p′
1
1 < p1 < 2 < p2 ≤ ∞ and δ
0 < p1 < 2 < p2 ≤ ∞ and δ
p′
1
,
or
d < 1
d > 1
p′
1
p′
1
,
.
dk( ¯As1
p1,q1(Ω) ֒→ ¯As2
p2,q2 (Ω)) ∼ k−γ,
,
s1−s2
d − 1
+ 1
p2
p1
s1−s2
2 − 1
d + 1
,
p1
p2
2 ( s1−s2
d − 1
+ 1
p2
p1
s1−s2
d
,
),
if
if
if
if
0 < p1 ≤ p2 ≤ 2,
0 < p1 < 2 < p2 ≤ ∞ and δ
0 < p1 < 2 < p2 < ∞ and δ
d < θ ′
2 < p1 ≤ p2 ≤ ∞ and δ
p2
d > θ ′
2 < p1 ≤ p2 ≤ ∞ and δ
p2
0 < p2 ≤ p1 ≤ ∞.
d > 1
p2
d < 1
p2
,
or
,
or
where
γ =
(iv)
where
(v)
γ =
where
γ =
Remark 2. Estimates of the approximation, the Gelfand, the Kolmogorov and
the Weyl numbers given in the last corollary coincide with the previous results
for these numbers in the case of bounded Lipschitz domains; cf.
[9, Section
3.3.4], [3], [19, Theorem 3.5, 4.12 and 4.6]. To the best of our knowledge the
estimates of the Weyl numbers were not complete to the very end; cf. A. Caetano
[5, 6]. The above corollary complements the previous estimates.
Acknowledgments
The first author was partially supported by the Natural Science Foundation
of China (Grant Nos 11171137, 61173187 and 61173188), the 211 Project of
Anhui University (Grant No. 33050069) and Doctoral Research Start-up Fund
Project of Anhui University (Grant No. 33190215).
21
References
[1] R.A. Adams and J.J.F. Fournier, Sobolev spaces, Elsevier, New York, 2003.
[2] M.S. Birman and M.Z. Solomjak, Quantitative analysis in Sobolev imbed-
ding theorems and applications to spectral theory, Amer. Math. Soc.
Transl., Series 2, 114 (1980).
[3] A.M. Caetano, About approximation numbers in function spaces, J. Ap-
prox. Theory 94 (1998), 383-395.
[4] A.M. Caetano, Asymptotic Distribution of Weyl Numbers and Eigenvalues,
D. Phil. thesis, University of Sussex, 1991.
[5] A.M. Caetano, Weyl numbers in function spaces, Forum Math. 2 (1990),
249-263.
[6] A.M. Caetano, Weyl numbers in function spaces II, Forum Math. 3 (1990),
613-621.
[7] C. Clark, An embedding theorem for function spaces, Pac. J. Math. 19
(1965), 243-251.
[8] C. Clark and D. Hewgill, One can hear whether a drum has finite area, P.
Am. Math. Soc. 18 (1967), 236-237.
[9] D.E. Edmunds and H. Triebel, Function spaces, entropy numbers, differ-
ential operators, Cambridge Tracts in Math. 120, Cambridge Univ. Press,
Cambridge, 1996.
[10] H. Konig, Eigenvalue Distribution of Compact Operators, Birkhauser,
Basel, 1986.
[11] H.-G. Leopold, Embeddings and entropy numbers for general weighted se-
quence spaces: the non-limiting case, Georgian Math. Journal, Volume 7
(2000), Number 4, 731-743.
[12] H.-G. Leopold, L. Skrzypczak, Compactness of embeddings of Besov and
Triebel-Lizorkin spaces defined on quasi-bounded domains (preprint).
[13] R. Linde, s-Numbers of diagonal operators and Besov embeddings, in: Pro-
ceedings of the 13th Winter School on Abstract Analysis (Srni, 1985), Rend.
Circ. Mat. Palermo 2 Suppl. No. 10 (1986), 83-110.
[14] C. Lubitz, Weylzahlen von Diagonaloperatoren und Sobolev-Einbettungen,
Bonner Math. Schriften 144 (1982).
[15] A. Pietsch, Eigenvalues and s-Numbers, Cambridge Stud. Adv. Math. 13,
Cambridge Univ. Press, Cambridge, 1987.
[16] A. Pietsch, Operator Ideals, Deutsch. Verlag Wiss., Berlin, 1978; North-
Holland, Amsterdam, London, New York, Tokyo, 1980.
22
[17] A. Pinkus, n-Widths in Approximation Theory, Springer-Verlag, Berlin,
1985.
[18] H. Triebel, Function spaces and wavelet on domains, European Mathemat-
ical Society Publishing House, Zurich 2008.
[19] J. Vybíral, Widths of embeddings in function spaces, J. Complexity 24
(2008), 545-570.
Shun Zhang
Key Laboratory of Intelligent Computing & Signal Processing of Ministry of
Education, and School of Computer Science and Technology,
Anhui University,
Hefei, 230601 Anhui,
P.R. CHINA
[email protected]
Alicja Gąsiorowska
Institute of Mathematics,
A. Mickiewicz University,
Umultowska 87, 61-614 Poznań,
POLAND
[email protected]
23
|
1301.5116 | 1 | 1301 | 2013-01-22T09:11:35 | Rational matrix solutions to the Leech equation: The Ball-Trent approach revisited | [
"math.FA"
] | Using spectral factorization techniques, a method is given by which rational matrix solutions to the Leech equation with rational matrix data can be computed explicitly. This method is based on an approach by J.A. Ball and T.T. Trent, and generalizes techniques from recent work of T.T. Trent for the case of polynomial matrix data. | math.FA | math |
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION:
THE BALL-TRENT APPROACH REVISITED
SANNE TER HORST
Abstract. Using spectral factorization techniques, a method is given by which
rational matrix solutions to the Leech equation with rational matrix data can
be computed explicitly. This method is based on an approach by J.A. Ball
and T.T. Trent, and generalizes techniques from recent work of T.T. Trent for
the case of polynomial matrix data.
0. Introduction
Consider H ∞-matrix functions G ∈ H ∞
+(Cp) →
ℓ2
+(Cm) and TK : ℓ2
+(Cm) be the corresponding (block) Toeplitz oper-
ators. See Section 1 below for the definitions of these spaces and operators. A
beautiful unpublished result of R.B. Leech (cf., [11]) tells us that there exists an
X ∈ H ∞
+(Cq) → ℓ2
m×p and K ∈ H ∞
m×q, and let TG : ℓ2
p×q such that
(0.1)
G(z)X(z) = K(z)
(z ∈ D),
and kXk∞ ≤ 1,
with D the open unit disc in C, if and only if
(0.2)
TGT ∗
G − TKT ∗
K is positive.
Note that (0.1) is equivalent to TGTX = TK and kTXk ≤ 1. Hence Leech's theorem
can be viewed as the analogue of the Douglas factorization lemma [7] within the
class of analytic Toeplitz operators. The necessity of (0.2) follows directly from
Douglas' factorization lemma and the reformulation of (0.1) in terms of Toeplitz
operators. The other implication is more involved. The solution criterion (0.2) can
also be formulated directly in terms of the functions G and K, it is equivalent to
the map
(0.3)
L(z, w) =
G(z)G(w)∗ − K(z)K(w)∗
1 − zw
(z, w ∈ D)
being a positive kernel in the sense of Aronszajn [1], that is, for any finite sequence
z1, . . . , zn ∈ D the block operator matrix [L(zi, zj)]i,j=1,...,n defines a positive op-
erator on the Hilbert space direct sum of n copies of Cm. We note that the actual
result by Leech is stated in the general context of Hilbert space operators inter-
twining shift operators, and in particular holds for operator-valued H ∞-functions
as well. Our interest is primarily in the case where G and K are rational matrix
functions.
There exists various proofs of Leech's theorem, see [10] and the references therein.
In [3] Ball and Trent prove a generalization of Leech's theorem to the polydisc in
Cd, adapting a technique coined the 'lurking isometry' approach in [2], and give a
Key words and phrases. Leech equation, Toeplitz operators, stable rational matrix functions,
outer spectral factor.
1
2
SANNE TER HORST
description of all X ∈ H ∞
here, specified to the single variable case.
p×q satisfying (0.1). We briefly outline the construction
G − TKT ∗
G − TKT ∗
K as TGT ∗
K = Λ◦Λ∗
The positivity of TGT ∗
K implies we can factor TGT ∗
◦, for some operator Λ◦ : H◦ → ℓ2
G −
TKT ∗
+(Cm) such that Ker Λ◦ = {0}. The
latter implies that dim H◦ = rank (TGT ∗
K). Such a factorization is often
referred to as a Kolmogorov decomposition in the literature, cf., [6]. Let Λ◦ be
the analytic operator-valued function on D, with values Λ◦(z) : H◦ → Cm, z ∈ D,
defined by Λ◦(z)h = (FmΛ◦h)(z), z ∈ D, h ∈ H◦. Here Fm is the Fourier transform
mapping ℓ2
m. Next one verifies that G,
K and Λ◦ satisfy the following identity:
+(Cm) isometrically onto the Hardy space H 2
G − TKT ∗
zw Λ◦(z)Λ◦(w)∗ + G(z)G(w)∗ =
(0.4)
= Λ◦(z)Λ◦(w)∗ + K(z)K(w)∗
(z, w ∈ D).
From this identity one derives the existence of a partial isometry
C◦ D◦(cid:21) :(cid:20)H◦
M◦ =(cid:20)A◦ B◦
Cq(cid:21) →(cid:20)H◦
Cp(cid:21)
(cid:2) z Λ◦(z) G(z) (cid:3) M◦ =(cid:2) Λ◦(z) K(z) (cid:3)
(z ∈ D).
(0.5)
such that
(0.6)
This in turn implies that the function X defined on D by
X(z) = D◦ + zC◦(I − zA◦)−1B◦
(0.7)
is in H ∞
p×q and satisfies (0.1). If one considers all contractions M◦ of the form (0.5)
such that (0.6) holds, possibly enlarging H◦, all solutions X to (0.1) are obtained
via (0.7).
(z ∈ D)
From the point of view of rational matrix functions the above construction has
one disadvantage. In general, the Hilbert space H◦ appearing in (0.5) is infinite
dimensional, and in that case it is hard to see when the solution X in (0.7) is
rational. In fact, even if both G and K are rational matrix functions, TGT ∗
G −TKT ∗
K
may very well be of infinite rank. More precisely, see Theorem 3.2 below, in the
rational matrix case TGT ∗
K has finite rank if and only if G(eit)G(eit)∗ =
K(eit)K(eit)∗ for all 0 ≤ t ≤ 2π. Overcoming this difficulty is the main theme of
the present paper.
G − TKT ∗
In the context of the Toeplitz-corona problem, which can be reduced to the
special case of (0.1) with q = m and K(z) = Im, z ∈ D, Trent [12] deduced a
modification of the above procedure for the special case that G is a row vector
(m = 1) polynomial, leading to a rational column vector solution of McMillan
degree at most the highest degree of the polynomials occurring in G. Throughout
this paper, the McMillan degree of a rational matrix function V will be denoted
by δ(V ); see Section 1 for the precise definition of δ(V ). The procedure of [12] was
recently extended in [13] to the general case of the Leech equation (0.1), with G
and K rational matrix functions, by reducing it to the case where G and K have
polynomial entries, and solving the latter problem via techniques similar to those
in [12].
In the present paper we also consider the Leech equation (0.1) with G and K
rational matrix functions. However, instead of reducing to the case of polynomial
data, we associate our problem with another Leech equation, with data functions G
and eK, i.e., with the same G. The advantage of our approach is that we keep better
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION
3
track of the McMillan degrees in our computations, leading to sharper bounds on
the McMillan degrees of the solutions. The construction of eK even works in the
case where G and K are not rational, provided that the function R ∈ L∞
by
m×m defined
(0.8)
R(eit) = G(eit)G(eit)∗ − K(eit)K(eit)∗
(a.e. t ∈ [0, 2π])
ΦTΦ and ker T ∗
admits an outer spectral factor, that is, a function Φ ∈ H ∞
r×m, for some r ≤ m,
with TR = T ∗
Φ = {0}. Note that outer spectral factors are unique
up to multiplication with a unitary constant matrix on the left, hence, with some
abuse of terminology, we will refer to the outer spectral factor, provided it exist. If
G and K are rational, then so is R, and this implies an outer spectral factor of R
exists.
Our method requires the following procedure:
1. Define R ∈ L∞
m×m by (0.8). Then TR is positive. Assume R admits an
outer spectral factor Φ ∈ H ∞
r×m, for some r ≤ m.
2. The subspace
(0.9)
MΦ := {f ∈ ℓ2
+(Cr) T ∗
Φf ∈ Im HG + Im HK }.
is invariant under the backward shift on ℓ2
Lax theorem, there exists an inner function Θ ∈ H ∞
that the range of TΘ is the orthogonal complement of MΦ.
+(Cr), and hence, by the Beurling-
r×k, for some k ≤ r, such
3. Define F ∈ L∞
m×k by F (eit) = Φ(eit)∗Θ(eit), for a.e. t ∈ [0, 2π]. Then
F ∈ H ∞
m×k.
The claims in the above steps will be proved Section 2. The function F defined in
Step 3 can be taken as a particular choice for the function F appearing in the next
theorem. This theorem provides the basis for our method and is the main result of
the present paper; a proof will be given in Section 2.
Theorem 0.1. Assume G ∈ H ∞
K is
positive and the function R defined in (0.8) admits an outer spectral factor. Then
there exists a function F ∈ H ∞
K − TF T ∗
m×k, for some k ≤ m, such that:
F is positive;
m×q such that TGT ∗
m×p and K ∈ H ∞
G − TKT ∗
G − TKT ∗
(i) TGT ∗
(ii) rank(cid:0)TGT ∗
G − TKT ∗
K − TF T ∗
F(cid:1) ≤ dim(cid:16)Im HG + Im HK(cid:17).
Here HG and HK denote the Hankel operators of G and K, respectively.
Given F as in Theorem 0.1, we apply the Ball-Trent approach with K replaced
(0.10)
by eK = [ K F ]. This yields H ∞-solutions eX = [ X Y ] of
G(z)(cid:2)X(z) Y (z)(cid:3) =(cid:2)K(z) F (z)(cid:3)
(z < 1), and k(cid:2)X Y(cid:3) k∞ ≤ 1.
Note that (0.10) implies that X satisfies (0.1). Whether or not all solutions of (0.1)
can be obtained via this procedure is still an open problem.
This procedure is specifically of interest in case G and K are rational matrix
functions. In that case the upper bound in (ii) is finite, and serves as an upper
bound on the least possible McMillan degree of solutions eX to (0.10), hence the
same upper bound applies to X. The following theorem provides some additional
results for the case of rational data functions; a proof will be given in Section 3.
Theorem 0.2. Let G ∈ H ∞
that TGT ∗
G−TKT ∗
m×p and K ∈ H ∞
K is positive. Then the function R ∈ L∞
m×q be rational matrix functions such
m×m defined by (0.8) admits
4
SANNE TER HORST
an outer spectral factor Φ. Moreover, in this case the functions R, Φ, Θ and F
defined in the above procedure are all rational matrix functions whose McMillan
degrees satisfy
(0.11)
1
2
δ(R) = δ(Φ) ≤ δ(Θ) = δ(F ) = dim MΦ < ∞,
and Θ is two-sided inner, i.e., k = r and Θ(eit)∗Θ(eit) = Ir = Θ(eit)Θ(eit)∗ for
each t ∈ [0, 2π]. Finally, we have
TGT ∗
G − TKT ∗
K − TF T ∗
F = HKH ∗
(0.12)
In particular, the left hand side in inequality (ii) in Theorem 0.1 is equal to rank (HK H ∗
HF H ∗
F − HGH ∗
G).
K + HF H ∗
F − HGH ∗
G.
K+
Thus, in case G and K are rational matrix functions, the problem reduces to
computing a Kolmogorov decomposition of the right hand side of (0.12). Note
that there are effective ways to computing Kolmogorov decompositions, cf., [6].
Moreover, the functions R, Φ, Θ and F can be computed explicitly using state space
techniques from mathematical systems theory (cf., [4, 8]), starting from a state
space representation of the function [ G K ]. This will be the topic of a forthcoming
paper of the present author together with A.E. Frazho and M.A. Kaashoek.
The paper consists of 4 sections, not counting the present introduction. Section
1 contains some of the notations and terminology as well as some operator theory
preliminaries used in the sequel. The main result, Theorem 0.1, is proved in Section
2.
In Section 3 the focus lays on the case that G and K are rational matrix
functions; a proof of Theorem 0.2 will be given as well as a criterion for the case
that TGT ∗
K has finite rank. The final section contains some general operator
theoretical results, and their proofs, that are used in the preceding sections.
G−TKT ∗
1. Preliminaries
In this section we introduce notations and terminology used throughout the
paper and we present some operator theory preliminaries.
With operator we mean a continuous linear map acting between two Hilbert
spaces. In particular, all operators in this paper are by definition bounded.
In-
vertibility of an operator means the operator has a bounded inverse. Let H be a
Hilbert space. A subspace of H is a closed linear manifold within H. The identity
operator on H will denoted by IH and the k × k identity matrix by Ik. Often these
subscripts H and k will be omitted. We say that an operator T on H is positive
whenever the inner product hT u, ui ≥ 0 for each u ∈ H, and T is said to be positive
definite whenever T is both positive and invertible. The notations T ≥ 0 and T > 0
will be used to indicate the positivity, respectively positive definiteness, of T . In
case T1 and T2 are selfadjoint operators on H, we will write T1 ≥ T2, resp. T1 > T2,
to indicate T1 − T2 ≥ 0, resp. T1 − T2 > 0.
The symbol H ∞
m×p will indicate the Hardy space of all uniformly bounded ana-
lytic m × p matrix-valued functions in the open unit disc. For any V ∈ H ∞
m×p the
supremum norm of V is defined by kV k∞ = supz<1 kV (z)k, making H ∞
m×p into
a Banach space. Here we follow the convention that the norm kM k of an m × p
matrix M is equal to the norm of the operator from Cp into Cm induced by M in
the canonical way. We write L∞
m×p for the Banach space consisting of all Lebesgue
measurable, essentially bounded m × p-matrix functions on the unit circle T to-
gether with the essential supremum norm, also denoted by k k∞. The space H ∞
m×p
Eku =(cid:2)u 0
0 · · ·(cid:3)⊤
. Note that I − SkS∗
k = EkE∗
k.
Let Z be a function in L∞
m×p and denote the Fourier coefficients of Z by . . . , Z−1, Z0, Z1, Z2, . . ..
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION
5
will be viewed both as a sub-Banach space of L∞
own right.
m×p and as a Banach space in its
With a function Z ∈ L∞
m×p we associate the functions Z ∗ ∈ L∞
p×m and Z t ∈ L∞
m×p
defined by
(1.1)
Z ∗(eit) = Z(eit)∗
and Z t(eit) = Z(e−it)
(a.e. t ∈ [0, 2π])
For V ∈ H ∞
m×p, the functions V ∗ and V t can be uniquely extended to bounded
analytic functions on the open exterior disc C\D, infinity included, via the formulas
V ∗(z) = V (1/¯z)∗ and V t(z) = V (1/z), z > 1.
By ℓ2(Ck) and ℓ2
+(Ck) we denote the Hilbert spaces consisting of bilateral,
respectively unilateral, square summable sequences with values in Ck. Viewing
ℓ2
+(Ck) as a sub-Hilbert space of ℓ2(Ck), we write ℓ2
−(Ck) for the orthogonal com-
+(Ck) in ℓ2(Ck). The symbol Sk stands for the (block) forward shift
plement of ℓ2
on ℓ2
+(Ck) defined by
+(Ck), and Ek denotes the canonical embedding of Ck into ℓ2
Then we define the (block) Toeplitz operator TZ and (block) Hankel operators HZ,+
and HZ,− associated with Z by the operators mapping ℓ2
+(Cm) given
by their infinite block matrix representations
+(Cp) into ℓ2
TZ =
Z0 Z−1 Z−2 ···
Z1 Z0 Z−1 ···
···
Z2 Z1 Z0
...
...
...
...
, HZ,+ =
Z1 Z2 Z3 ···
Z2 Z3 Z4 ···
Z3 Z4 Z5 ···
...
...
...
...
, HZ,− =
Z−1 Z−2 Z−3 ···
Z−2 Z−3 Z−4 ···
Z−3 Z−4 Z−5 ···
...
...
...
...
.
We shall refer to HZ,+ and HZ,− as the analytic, respectively anti-analytic, Hankel
operator associated with Z. Note that TZ ∗ = T ∗
Z,+ = HZ ∗,−. For V ∈
H ∞
m×p we have HV,− = 0, and we will simply write HV for HV,+.
Now consider U ∈ H ∞
m×q. Then the following useful
m×p and W ∈ H ∞
n×p, V ∈ H ∞
Z and H ∗
identities apply (cf., [5, Proposition 2.14]):
(1.2)
TV ∗W = T ∗
HV ∗W,+ = T ∗
V TW ,
TU V ∗ = TU T ∗
V + HU H ∗
V ,
V HW , HU V ∗,+ = HU T ∗
V t.
p×m- and H ∞
The sets of rational matrix L∞
m×p and RH ∞
p×m-functions will be denoted by
RL∞
m×p, respectively. For a m × p rational matrix function Z the
McMillan degree is denoted by δ(Z) and equals the sum of the local degrees,
δ(Z) = Pw∈C δ(Z, w). Here the local degree δ(Z, w) of Z at w is defined to be
the rank of the Hankel operator defined by the negative Fourier coefficients of the
Fourier expansion if Z in a deleted neighborhood of w. See Section 8.4 in [4] for more
details. It is well known that for V ∈ RH ∞
m×p the MacMillan degree δ(V ) equals
the rank of the Hankel operator HV . Moreover, for Z ∈ RL∞
m×p the MacMillan
degree δ(Z) is equal to rank (HZ,+) + rank (HZ,−).
2. Proof of Theorem 0.1
Let G ∈ H ∞
m×p and K ∈ H ∞
m×q, and define R ∈ L∞
m×m by (0.8). Throughout
this section we shall assume that TGT ∗
K. This implies that R is positive
on T. Indeed, note that the positivity of the kernel L in (0.3) implies that (1 −
z2)L(z, z) = G(z)G(z)∗ − K(z)K(z)∗ is positive for each z ∈ D. Hence the same is
true for the non-tangential limits of (1 − z2)L(z, z) to the unit circle, which exist
G ≥ TKT ∗
6
SANNE TER HORST
for almost all points on the unit circle, where the values coincide with the values of
R.
Since the function R is positive on T, it follows that TR is a positive operator
on ℓ2
+(Cm). Under some additional constraints on TR, the positivity of TR implies
that R admits an outer spectral factor (see [9, Proposition V.4.2]), that is, there
exists a function Φ ∈ H ∞
R = Φ∗Φ,
r×m, for some integer r ≤ m, such that
and Ker T ∗
i.e., TR = T ∗
(2.1)
Φ = {0}.
ΦTΦ,
The latter condition says that TΦ has dense range, i.e., Φ is outer. The function
Φ is unique up to a unitary constant matrix on the left, that is, if Ψ is another
outer function satisfying R = Ψ∗Ψ, then Φ and Ψ are matrix functions of the same
size, and Φ(·) = U Ψ(·) where U is a constant unitary matrix. With some abuse of
terminology, we shall refer to Φ as the outer spectral factor of R. See [9, 11] for
further details.
We start with a few preliminary results.
Lemma 2.1. Let Φ be the r × m outer spectral factor of the function R given by
(0.8). Set NΦ = Im HG + Im HK, and let MΦ be the inverse image of NΦ under
the map T ∗
Φ, i.e.,
MΦ = (T ∗
(2.2)
Then MΦ is a subspace of ℓ2
the backward shift S∗
r . Moreover,
Φ)−1 [NΦ] = {f ∈ ℓ2
+(Cr) T ∗
Φf ∈ Im HG + Im HK }.
+(Cr), dim MΦ ≤ dim NΦ, and MΦ is invariant under
(2.3)
Im HΦEm = Im S∗
r TΦEm ⊂ MΦ.
Proof. Since T ∗
manifold NΦ under T ∗
bound on dim MΦ follows from the injectivity of T ∗
and S∗
Φ is a continuous linear map, the inverse image of the closed linear
Φ is again linear and closed. Thus MΦ is a subspace. The
mHG = HGSp
Φ. The fact that S∗
mHK = HKSq implies that
S∗
m(cid:16)Im HG + Im HK(cid:17) ⊂ S∗
mIm HG + S∗
mIm HK
= Im HGSp + Im HKSq ⊂ Im HG + Im HK.
m. Take f ∈ MΦ, i.e., T ∗
Φf ∈ NΦ. Using SrTΦ =
Thus NΦ is invariant under S∗
TΦSm we have
T ∗
ΦS∗
r f = S∗
mT ∗
Φf ∈ S∗
mNΦ ⊂ NΦ.
Thus S∗
r f ∈ MΦ. Hence MΦ is invariant under the backward shift S∗
r .
Next we prove (2.3). Inspecting the first columns in HΦ and TΦ yields HΦEm =
S∗
r TΦEm. Hence the identity in (2.3) holds. Take u ∈ Cm, and put x = S∗TΦEmu.
Then
Φx = T ∗
T ∗
= S∗
ΦS∗
m(TGT ∗
r TΦEmu = S∗
G + HGH ∗
ΦTΦEmu = S∗
mT ∗
G)Emu − S∗
mTREmu
m(TKT ∗
mTGEp = HGEp and S∗
K)Emu.
K + HKH ∗
mHG = HGSp. Hence
Note that T ∗
GEm = EpG(0)∗, S∗
m(TGT ∗
S∗
Similarly, S∗
m(TKT ∗
Im HG + Im HK ⊂ NΦ, and thus x ∈ MΦ. Hence Im S∗
G + HGH ∗
K + HKH ∗
G)Emu = HG(EpG(0)∗ + SpH ∗
K)Emu ∈ Im HK. This shows that T ∗
r TΦEm ⊂ MΦ.
GE)u ∈ Im HG.
Φx belongs to
(cid:3)
Corollary 2.2. Let Φ be the r × m outer spectral factor of the function R given by
(0.8). Define MΦ by (2.2). Then Im HΦ ⊂ MΦ.
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION
7
Proof. By (2.3), we see that the range of the first block column of HΦ is in MΦ.
m = S∗l
Since HΦSm = S∗
r HΦ, it follows that HΦSl
r HΦ holds for any positive integer
l. The fact that MΦ is invariant under S∗
r then shows that for any positive integer
l
Im HΦSl
mEm = Im S∗l
r HΦEm = S∗l
r Im HΦEm ⊂ S∗l
r MΦ ⊂ MΦ.
This shows that the range of each column of HΦ is in MΦ, and thus the range of
HΦ is included in MΦ.
(cid:3)
By the Beurling-Lax-Halmos theorem, the fact that the space MΦ is invariant
under the backward shift implies MΦ = Ker T ∗
r×k,
with k some nonnegative integer, k ≤ r. This Θ is unique up to a constant unitary
matrix from the right. Despite this mild form of non-uniqueness, we shall refer to
Θ as the inner function associated with the space MΦ.
Θ for some inner function Θ ∈ H ∞
Proposition 2.3. Let Φ be the r × m outer spectral factor of the function R given
by (0.8), and let Θ be the r × k inner function associated with the space MΦ in
(2.2). Then F = Φ∗Θ belongs to H ∞
F = T ∗
K − TF T ∗
m×k. Moreover, we have
ΦPMΦ TΦ − HGH ∗
G + HKH ∗
F ) ≤ dim(Im HG + Im HK).
(i) TGT ∗
(ii) rank (TGT ∗
G − TKT ∗
G − TKT ∗
K − TF T ∗
K ≥ 0;
Here PMΦ is the orthogonal projection on ℓ2
HGH ∗
K ≥ 0, then rank (TGT ∗
G − HKH ∗
G − TKT ∗
+(Ck) with range MΦ. If in addition
K − TF T ∗
F ) ≤ dim MΦ.
Proof. Since Φ∗ and Θ are matrix-valued L∞-functions, we have F ∈ L∞
see that F ∈ H ∞
showing that HF ∗,+ = 0. Note that ker T ∗
Hence HΦTΘ = 0, by Corollary 2.2. Thus the third identity in (1.2) yields
m×k. To
m×k it suffices to show that HF,− = 0. However, this is the same as
+(Cr) ⊖ MG,K, by definition of Θ.
Θ = ℓ2
HF ∗,+ = HΘ∗Φ,+ = T ∗
ΦHΦ = 0,
and it follows that F ∈ H ∞
m×k, as claimed.
Next we deal with item (i). Since MΦ = Ker T ∗
and TΘT ∗
Applying the second identity in (1.2) yields
Θ is the orthogonal projection onto M⊥
Θ and Θ is inner, M⊥
Φ = Im TΘ
Φ. In particular, I − PMΦ = TΘT ∗
Θ.
(2.4)
TR = (TGT ∗
G − TKT ∗
K) + (HGH ∗
G − HKH ∗
K).
With (2.4) and I − PMΦ = TΘT ∗
(TGT ∗
G − TKT ∗
K) + (HGH ∗
ΦPMΦ TΦ + T ∗
ΦPMΦ TΦ + TF T ∗
F .
= T ∗
= T ∗
Θ we obtain
G − HKH ∗
Φ(I − PMΦ )TΦ = T ∗
K) = TR = T ∗
ΦTΦ =
ΦPMΦTΦ + T ∗
ΦTΘT ∗
ΘTΦ =
Here we used that T ∗
ΦTΘ = TΦ∗ TΘ = TF . This proves the identity in (i).
To show T ∗
ΦPMΦTΦ−HGH ∗
K is positive and to prove the rank constraint
on this operator, we apply Lemma 4.1 with the following choices of spaces and
operators:
G+HKH ∗
V = ℓ2
W = ℓ2
G − HKH ∗
+(Cm), V1 = NΦ, V2 = V ⊖ NΦ X = HGH ∗
K,
Y = T ∗
+(Cr), W1 = MΦ, W2 = W ⊖ MΦ,
Φ.
8
SANNE TER HORST
Here NΦ and MΦ are the spaces defined in Lemma 2.1. In particular,
XV = Im (HGH ∗
G − HKH ∗
Y −1[V1] = (T ∗
K) ⊂(cid:16)Im HG + Im HK(cid:17) = NΦ = V1,
Φ)−1[NΦ] = MΦ = W1.
Furthermore, we have
Y Y ∗ − X = T ∗
ΦTΦ − HGH ∗
G + HKH ∗
K = TR − HGH ∗
G + HKH ∗
K
= TGT ∗
G − TKTK ≥ 0.
Hence TΦPMΦ T ∗
rank constraint (ii) follows from Lemma 4.1 as well.
G + HKH ∗
Φ − HGH ∗
K = Y PW1 Y ∗ − X is positive by (4.1), and the
Moreover, note that HGH ∗
G − HKH ∗
last statement of Lemma 4.1 we find that
K ≥ 0 translates to X ≥ 0. Thus, by the
rank (T ∗
ΦPMΦTΦ − HGH ∗
G + HKH ∗
K) = rank (Y PW1Y ∗ − X) ≤ dim W1,
which, together with dim W1 = dim MΦ, proves the last claim.
(cid:3)
We will now prove the main result of the present paper.
Proof of Theorem 0.1. Let MΦ and NΦ be as in Lemma 2.1, and define F as
in Proposition 2.3. Thus F = Φ∗Θ, where Φ is the r × m outer spectral factor of
the function R, and Θ is the r × k inner function associated with the space MΦ
in (2.2). We know that F ∈ H ∞
m×k. With this choice of F , Proposition 2.3 tells us
directly that items (i) and (ii) in Theorem 0.1 are fulfilled.
(cid:3)
3. The case where G and K are rational matrix functions.
Let G ∈ RH ∞
G − TKTK ≥ 0. The aim of
this section is to prove Theorem 0.2. In addition we will derive a criterion for the
case that rank (TGT ∗
m×q such that TGT ∗
m×p and K ∈ RH ∞
G − TKT ∗
K) < ∞.
In the previous section we observed that TR ≥ 0, where R ∈ L∞
m×m is given by
(0.8). Since G and K are rational, so is R, and this, together with TR ≥ 0, implies
R admits an outer spectral factor Φ ∈ RH ∞
r×m, sor some r ≤ m, see [11, Section
6.6]. Also note that δ(G) = rank HG < ∞ and δ(K) = rank HK < ∞ imply that
the subspace NΦ of Lemma 2.1 is finite dimensional, and hence the subspace MΦ
in (2.2) is finite dimensional, since ker T ∗
Φ = {0}. Then Theorem 4.3.2 in [8] yields
that the inner function Θ associated with MΦ is a two-sided inner rational matrix
function, that is, Θ ∈ RH ∞
r×r and ΘΘ∗ = Θ∗Θ is identically equal to Ir.
The next proposition provides the relations between the McMillan degrees given
in Theorem 0.2.
Proposition 3.1. Let G ∈ RH ∞
K ≥ 0.
Define MΦ as in (2.2). Then the functions R, Φ, Θ and F defined in Section 2 are
all rational matrix functions and the following bounds on their McMillan degrees
apply:
m×p and K ∈ RH ∞
m×q with TGT ∗
G − TKT ∗
(3.1)
1
2
δ(R) = δ(Φ) ≤ δ(F ) = δ(Θ) = dim MΦ.
Moreover, we have HΘH ∗
Θ = PMΦ and HF H ∗
F = T ∗
ΦPMΦTΦ.
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION
9
Proof. Corollary 2.2 implies Im HΦ ⊂ MΦ. Hence
δ(Φ) = rank (HΦ) = dim Im HΦ ≤ dim MΦ.
Moreover, we have HR,+ = HΦ∗Φ,+ = T ∗
to V ∗W = Φ∗Φ. Since Φ is outer, Ker T ∗
rank HΦ = δ(Φ). By TR ≥ 0, we have HR,− = H ∗
rank H ∗
ΦHΦ, by the third identity in (1.2) applied
Φ = {0}, and therefore rank HR,+ =
R,+. In particular, rank HR,− =
R,+ = rank HR,+, and thus δ(R) = 2rank HR,+ = 2δ(Φ).
Θ implies TΘT ∗
The fact that Θ is inner with MΦ = Ker T ∗
Θ = I − PMΦ . Since
Θ is two-sided inner, we have ΘΘ∗ = Θ∗Θ = Ir, hence TΘΘ∗ = I. Now apply the
second identity of (1.2). This yields
HΘH ∗
Θ = TΘΘ∗ − TΘT ∗
Θ = I − TΘT ∗
Θ = PMΦ.
Hence
δ(Θ) = rank HΘ = rank (HΘH ∗
Θ) = rank PMΦ = dim MΦ.
Recall that F = Φ∗Θ. Hence, by the third identity of (1.2), we obtain that
Φ = {0}, which implies
ΦHΘ together
(cid:3)
HF = HΦ∗Θ = T ∗
δ(F ) = rank HF = rank (T ∗
with HΘH ∗
ΦHΘ) = rank HΘ = δ(Θ). Finally, HF = T ∗
ΦHΘ. Since Φ is outer, we have Ker T ∗
Θ = PMΦ implies HF H ∗
ΦPMΦTΦ.
F = T ∗
Note that dim(Im HG + Im HK) ≤ δ(G) + δ(K) < ∞, since G ∈ RH ∞
m×p and
K ∈ RH ∞
equation (0.1) to one where
m×q. Hence, replacing K by eK = [ K F ], reduces the original Leech
(3.2)
rank (TGT ∗
G − TKT ∗
K) < ∞.
We will next focus on the case of the Leech equation where the rank constraint
(3.2) holds. The following theorem provides necessary and sufficient conditions for
(3.2) to hold.
Theorem 3.2. Let G ∈ RH ∞
R ∈ RL∞
G − TKT ∗
m×m by (0.8). Then the following statements are equivalent:
m×p and K ∈ RH ∞
m×q with TGT ∗
K ≥ 0. Define
G − TKT ∗
(i) rank (TGT ∗
(ii) TR = 0;
(iii) G(eit)G(eit)∗ = K(eit)K(eit)∗
K) < ∞;
(t ∈ [0, 2π]).
Moreover, in this case
(3.3)
and
(3.4)
δ(G) ≤ δ(K),
TGT ∗
G − TKT ∗
K = HKH ∗
K − HGH ∗
G
δ(K) − δ(G) ≤ rank (TGT ∗
G − TKT ∗
K) ≤ δ(K).
Here δ(G) and δ(K) denote the McMillan degrees of G and K, respectively.
Proof. Note that (iii) is equivalent to R(eit) = 0 for each t ∈ [0, 2π], hence to
TR = 0, since R ∈ RL∞
m×m. Thus (ii) ⇔ (iii).
The fact that G and K are rational matrix H ∞-functions implies that HG and
HK have finite rank, and thus rank (HGH ∗
K) < ∞. From formula (2.4)
it then follows that (i) holds if and only if rank TR < ∞. However, R is a rational
matrix function with no poles of the circle, and thus continuous on the circle. This
implies that rank TR < ∞ holds if and only if R(eit) = 0 for all t ∈ [0, 2π], and thus
TR = 0. Hence (i) ⇔ (ii).
G − HKH ∗
The combination of TR = 0 and formula (2.4) gives (3.3).
10
SANNE TER HORST
Note that for any positive Hilbert space operators Z and Y on V, the inequality
Z ≥ Y implies rank Z ≥ rank Y . Indeed, by Douglas' Factorization Lemma there
exists a contraction Q on V such that Y
2 . Hence
2 = QZ
1
1
rank Y = rank Y
1
2 = rank (QZ
1
2 ) ≤ rank (Z
1
2 ) = rank (Z).
Applying this inequality with Z = HKH ∗
HKH ∗
K − HGH ∗
G = TGT ∗
δ(K) = rank (HKH ∗
K ≥ 0, we obtain
G − TKT ∗
K) = rank (Z) ≥ rank (Y ) = rank (HGHG) = δ(G).
K and Y = HGH ∗
G and noting that Z −Y =
If we take Z = HKH ∗
Z ≥ Y , and thus
K and Y = HKH ∗
K − HGH ∗
G = TGT ∗
G − TKT ∗
K, then clearly
δ(K) = rank (HKH ∗
In addition to Z and Y , set V = HGH ∗
K) = rank (Z) ≥ rank (Y ) = rank (TGT ∗
G. Then Z = Y + V implies
G − TKT ∗
K).
rank (Z) = rank (Y + V ) ≤ rank (Y ) + rank (V ).
Since δ(G) = rank (V ) and δ(K) = rank (Z), the last part of (3.4) holds.
(cid:3)
Remark 3.3. If the matrix H ∞-functions G and K are continuous, then the first
part of Theorem 3.2 goes through in a slightly altered form. One only has to replace
(i) by: TGT ∗
K is compact. The argumentation is similar to the one given in
the proof of Theorem 3.2, where we now use that HG and HK are compact, since
G and K are continuous, and that R being continuous together with TR compact
implies TR = 0, and hence R = 0.
G − TKT ∗
How restrictive condition (3.2) can be becomes evident when considering the
Toeplitz corona problem.
Corollary 3.4. Let G ∈ RH ∞
Im for each z ∈ D. Then rank (TGT ∗
matrix function whose value is a co-isometry.
m×p such that TGT ∗
G ≥ I, i.e., (0.2) holds with K(z) =
G − I) < ∞ holds if and only if G is a constant
Proof. Clearly if G is a constant matrix function whose value is a co-isometry,
then TGT ∗
G = I, and hence TGT ∗
G − I has finite rank.
Conversely, assume TGT ∗
G − I has finite rank. By Theorem 3.2, R = GG∗ −
Im = 0. Thus GG∗ = Im. In particular, the values of G are co-isometries. By
G. Thus
the second identity in (1.2) we have Iℓ2
−HGH ∗
G = TGT ∗
G − I ≥ 0. This can only occur if HG = 0, i.e., if G is constant
matrix function.
(cid:3)
+(Cm) = TGG∗ = TGT ∗
G + HGH ∗
Corollary 3.5. Let G ∈ RH ∞
R, Φ, Θ, and F as in Section 2. Then
m×p and K ∈ RH ∞
m×q with TGT ∗
G−TKT ∗
K ≥ 0. Define
(3.5)
Φ∗Φ = R = F F ∗
and Φ = ΘF ∗.
Moreover, TR > 0 if and only Φ is invertible outer, that is, r = m and Φ has
an inverse in H ∞
m×m with an anti-analytic
inverse.
m×m. In this case F is invertible in L∞
The first two identities in (3.5) say that Φ is a right and F a left spectral factors of
R. The last identity, together with Θ two-sided inner, provides a Douglas-Shapiro-
Shields factorization of Φ, cf., [8, Chapter 4].
GG∗ = eK eK ∗ = KK ∗ + F F ∗,
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION
11
Proof of Corollary 3.5. The identity Φ∗Φ = R holds by definition of Φ. Apply-
ing Theorem 3.2 with K replaced by eK = [ K F ], where we note that condition (i)
is satisfied by Theorem 0.1, yields
i.e. F F ∗ = GG∗ − KK ∗ = R.
Recall that F is defined as F = Φ∗Θ. Hence F ∗ = Θ∗Φ. Since Θ is two-sided inner,
ΘΘ∗ is identically equal to Ir. Hence Φ = ΘF ∗.
It is well known that TR > 0 holds if and only if its outer spectral factor is
invertible outer, c.f., [8, Proposition 10.2.1]. Assume TR > 0. Then Φ and Θ
are invertible with Φ−1 ∈ H ∞
m×m and Θ−1 = Θ∗. This shows that F = Φ∗Θ is
invertible in L∞
m×m, with inverse (Φ∗Θ)−1 = Θ∗(Φ∗)−1 = Θ∗(Φ−1)∗. Since Θ∗ and
(Φ−1)∗ are both anti-analytic, so is F −1.
(cid:3)
Proof of Theorem 0.2. We observed at the beginning of the present section that
R admits an outer spectral factor and that Θ is two-sided inner. The relations
between the McMillan degrees of R, Φ, Θ and F in (0.11) follow from Proposition
3.1. The identity (0.12) follows by replacing K in (3.3) by eK = [ K F ], noting that
F ) < ∞ by Theorem 0.1, and
(cid:3)
K − TF T ∗
) = rank (TGT ∗
= TKT ∗
K + TF T ∗
G − T eKT ∗
rank (TGT ∗
eK
the identities T eKT ∗
eK
G − TKT ∗
F and H eKH ∗
eK
K + HF T ∗
F .
= HKT ∗
In case (3.2) holds, the following proposition shows how the partial isometry M◦
in (0.6) can be computed.
Proposition 3.6. Let G ∈ RH ∞
m×q such that (3.2) holds. Let
ν = rank (TGT ∗
K) < ∞. Then ν ≤ δ(K), the space H0 in (0.5) can be
taken to be Cν, and in that case the partial isometry M◦ in (0.5) and (0.6) can be
computed via M◦ = M +
m×p and K ∈ RH ∞
G − TKT ∗
1 M∗ with
1
2π Z 2π
0
M1 =
V (eiω)∗V (eiω)dω,
M∗ =
V (eiω)∗W (eiω)dω,
V (eit) =(cid:2) eit Λ◦(eit) G(eit) (cid:3) , W (eit) =(cid:2) Λ◦(eit) K(eit) (cid:3) ,
1 the Moore-Penrose pseudo-inverse of M1.
and M +
a.e.
1
2π Z 2π
0
Proof. Recall from the introduction that dim H◦ = rank (TGT ∗
K) = ν.
Since ν < ∞, we can apply a linear transformation identifying H◦ with Cν, and
since H◦ comes from the factorization of TGT ∗
K, we can just as well apply
this transformation and take H0 to be Cν. The bound on ν is a direct consequence
of (3.4).
G − TKT ∗
G − TKT ∗
The formula for M◦ follows by applying Lemma 4.2 with the given choice of V
and W . Note that the identity (4.5) follows from (0.4). The square summability of
the Taylor coefficients of V and W follows from the boundedness of TGEp, TKEq
and Λ◦ (as defined in the introduction), as operators mapping into ℓ2
+(Cm). Hence
all conditions are satisfied, and Lemma 4.2 applies.
(cid:3)
We conclude this section with two examples.
Example 3.7. According to Proposition 2.3, if HGH ∗
K ≥ 0, then the
upper bound on the rank of TGT ∗
F in item (i) can be improved
to dim M, with M as defined in Lemma 2.1. This improvement can be arbitrarily
large. Let l be a positive integer, take for G any rational function of McMillan
degree l and take K = G. Clearly R = GG∗ − KK ∗ = 0, thus Φ = 0, which implies
G − HKH ∗
K − TF T ∗
G − TKT ∗
12
SANNE TER HORST
M = {0}. Hence dim M = 0; a solution with McMillan degree 0 is obviously
X(z) = 1, z ∈ C. On the other hand dim(Im HG + Im HK) = dim(Im HG) =
δ(G) = l. Hence we have an improvement of l.
Example 3.8. Let G and K are matrix polynomials whose values are matrices
of size m × p, respectively m × q, say with degrees d1, respectively d2. Assume
that the last coefficients of G and K, i.e, corresponding to zd1 and zd2, have full
rank and that p, q ≥ m. This implies that the last coefficients of G and K admit a
right inverse. Note that HG and HK only have entries on the first d1, respectively
d2, anti-diagonals, starting in the left upper corner. Since the last coefficients
of G and K admit a left inverse, it follows that δ(G) = rank HG = m · d1 and
δ(K) = rank HK = m · d2. Now also assume that TGT ∗
K ≥ 0. Applying
Theorem 0.1, and following the subsequent procedure we obtain that there exists
a rational matrix solution X to (0.1). The McMillan degree of X is bounded by
δ(G) + δ(K) = m(d1 + d2). However, in this case the rank constraint in item
(ii) of Theorem 0.1 gives a much sharper bound, namely dim(Im HG + Im HK) ≤
m max{d1, d2}, due to the specific structure of HG and HK. Note that this bound
is in line with [13] (where the factor m does not appear, but should be there).
G − TKT ∗
4. Appendix
In this appendix we prove two results of a general operator theoretical nature
that are used in the paper.
Lemma 4.1. Let V = V1 ⊕ V2 and W = W1 ⊕ W2 be Hilbert space direct sums,
and let X : V → V and Y : W → V be operators. Assume that X is selfadjoint and
XV ⊂ V1, and that W1 = Y −1[V1], i.e., W1 is the inverse image of V1 under Y .
Finally, let PW1 be the orthogonal projection of W onto W1. Then
(4.1)
Y Y ∗ − X ≥ 0 ⇐⇒ Y PW1 Y ∗ − X ≥ 0.
Moreover, rank (Y PW1Y ∗ − X) ≤ dim V1. Assume Y Y ∗ − X ≥ 0 and in addition
that Y is injective and X ≥ 0. Then rank (Y PW1Y ∗) = dim W1, rank X ≤ dim W1
and rank (Y PW1Y ∗ − X) ≤ dim W1.
Proof. Using the decompositions V = V1 ⊕ V2 and W = W1 ⊕ W2 we represent X
and Y as 2 × 2 operator matrices, as follows:
(4.2)
X =(cid:20)X1
0
0
0(cid:21) :(cid:20)V1
V2(cid:21) →(cid:20)V1
V2(cid:21) , Y =(cid:20)Y1 Y2
Y3(cid:21) :(cid:20)W1
W2(cid:21) →(cid:20)V1
V2(cid:21) .
0
Note that the zeros in the operator matrix for X follow from the fact that X is
selfadjoint and XV ⊂ V1. The zero in the left lower corner of the operator matrix
for Y is a consequence of W1 = Y −1[V1]. Indeed, the latter equality implies that Y
maps W1 into V1. The identity W1 = Y −1[V1] also implies that Y3 is one-to-one.
To see this, assume Y3u = 0 for some u ∈ W2. Then Y u ∈ V1. But the latter
can only happen when u ∈ Y −1[V1] = W1. Thus u ∈ W1 ∩ W2, and hence u = 0.
Therefore, Y3 is one-to-one.
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION
13
Next, observe that the partitionings in (4.2) imply that
(4.3)
(4.4)
Y Y ∗ − X =(cid:20)IV1 Y2
Y3(cid:21)(cid:20)Y1Y ∗
Y PW1 Y ∗ − X =(cid:20)Y1Y ∗
1 − X1
0
0
1 − X1
0
0
Y ∗
2
0
Y ∗
IW2(cid:21)(cid:20)IV1
3(cid:21) on (cid:20) V1
V2 (cid:21) ,
0(cid:21) on (cid:20) V1
V2 (cid:21) .
0
Now assume that the inequality in the right hand side of (4.1) holds. This
implies that the operator matrix in the right hand side of (4.4) is positive. But
then the same holds true for the operator defined by the second operator matrix
in the right hand side of (4.3). The equality (4.3) then shows that Y Y ∗ − X is a
positive operator, and the implication ⇐= in (4.1) is proved.
To prove the reverse implication assume that Y Y ∗ − X is a positive operator.
Since Y3 is one-to-one, the operator U from V1 ⊕ V2 to V1 ⊕ W2 defined by the third
operator matrix in the right hand side of (4.3) has a dense range. Using (4.4) and
the positivity of Y Y ∗ − X, we see that
But the range of U is dense. Hence, by continuity, we get
1 − X1
0
(cid:10)(cid:20)Y1Y ∗
(cid:10)(cid:20)Y1Y ∗
1 − X1
0
for all v ∈ V = V1 ⊕ V2.
0
IW2(cid:21) U v, U v(cid:11) ≥ 0
IM⊥(cid:21) y, y(cid:11) ≥ 0 for all y ∈ V1 ⊕ W2.
0
It follows that Y1Y ∗
operator Y PW1 Y ∗ − X. This proves the implication =⇒ in (4.1).
1 − X1 is positive, and by (4.4) the same holds true for the
The decomposition (4.4) shows clearly that rank (Y PW1 Y ∗ − X) ≤ dim V1.
Note that if Y is injective, we have rank (Y PW1Y ∗) = rank (PW1 ) = dim W1.
Assuming Y Y ∗ − X ≥ 0, we have Y PW1 Y ∗ ≥ X. By Douglas' Factorization
Lemma, X
2 = KPW1 Y ∗ for some contraction K, and hence
1
rank X = rank X
1
2 = rank (KPW1 Y ∗) ≤ rank (PW1 Y ∗) = rank (Y PW1 Y ∗).
Thus rank X ≤ dim W1. A similar argument applied to Y PW1 Y ∗ ≥ Y PW1 Y ∗ − X
shows rank (Y PW1 Y ∗ − X) ≤ rank (Y PW1 Y ∗) = dim W1.
(cid:3)
Lemma 4.2. Consider two matrix functions V and W , analytic on D, with valued
V (z) : Ck → Cp and W (z) : Cν → Cp, z ∈ D, and Taylor expansions V (z) =
j Wj <
j=0 W ∗
j=0 V ∗
P∞
j=0 zjVj and W (z) =P∞
∞. If
j=0 zjWj. AssumeP∞
j Vj < ∞ and P∞
(4.5)
V (z)V (w)∗ = W (z)W (w)∗
for all z, w ∈ D,
then there exists a partial isometry M : Cν → Ck such that V (z)M = W (z) for all
z in D. Moreover, this partial isometry M is given by M = M +
1 M∗ with
M1 =
M∗ =
(4.6)
0
1
2π Z 2π
2π Z 2π
1
0
V (eiω)∗V (eiω)dω =
V (eiω)∗W (eiω)dω =
∞Xj=0
∞Xj=0
V ∗
j Vj
V ∗
j Wj.
Here M +
1 denotes the Moore-Penrose pseudo inverse of M1.
14
SANNE TER HORST
Proof. The assumption yields we can define operators Ω1 and Ω2 by
: Ck → ℓ2
+(Cp) and Ω2 =
: Cν → ℓ2
+(Cp).
Ω1 =
V0
V1
V2
...
W0
W1
W2
...
For each z ∈ D we write Fz for the point evaluation operator
Fz = E∗
p (I − zS∗
p)−1 : ℓ2
+(Cp) → Cp,
i.e., Fz(x0, x1, x2, . . .) =
∞Xj=0
zjxj.
Note that V (z) = Fp,zΩ1 and W (z) = Fp,zΩ2, z ∈ D. Hence
Fp,z(Ω1Ω∗
1 − Ω2Ω∗
2)F∗
p,w = V (z)V (w)∗ − W (z)W (w)∗ = 0
(z, w ∈ D).
1 = Ω2Ω∗
2 as initial space and Im Ω∗
Since ∩z∈DKer Fp,z = {0}, it follows that Ω1Ω∗
2. By Douglas' factorization
lemma there exists a unique partial isometry M : Cν → Ck that satisfies Ω1M = Ω2
and has Im Ω∗
1 as final space. Multiplying both sides with
Fp,z yields V (z)M = W (z), z ∈ D. Note that the Moore-Penrose pseudo inverse
of Ω1 is given by Ω+
1 Ω1 is the orthogonal projection on
1. Thus M = Ω+
Im Ω∗
1Ω1 and
M∗ = Ω∗
(cid:3)
1 = (Ω∗
1 Ω1M = Ω+
1 M∗.
1Ω2. Note that M1 = Ω∗
1Ω2. Hence M = M +
1 Ω2 = (Ω∗
1Ω1)+Ω∗
1Ω1)+Ω∗
1. Then Ω+
Acknowledgement. The author thanks Art Frazho and Rien Kaashoek for the
useful discussions and their constructive suggestions during the preparation of this
paper.
References
[1] N. Aronszajn, Theory of reproducing kernels, Trans. Amer. Math. Soc. 68 (1950), 337 -- 404.
[2] J.A. Ball, Linear systems, operator model theory and scattering: multivariable generaliza-
tions, in: Operator Theory and Its Applications (Winnipeg, MB, 1998), pp. 151-178, Fields
Inst. Commun., Vol. 25, Amer. Math. Soc., Providence, 2000.
[3] J.A. Ball and T.T. Trent, Unitary colligations, reproducing kernel Hilbert spaces, and
Nevanlinna-Pick interpolation in several variables, J. Funct. Anal. 157 (1998), 161.
[4] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran, Factorization of matrix and operator
functions: the state space method, Oper. Theory Adv. Appl. 178, Birkhauser Verlag, Basel,
2008.
[5] A. Bottcher and B. Silbermann, Analysis of Toeplitz operators, Springer-Verlag, Berlin, 1990.
[6] T. Constantinescu, Schur parameters, factorization and dilation problems, Oper. Theory
Adv. Appl. 82, Birkhauser Verlag, Basel, 1996.
[7] R.G. Douglas, On majorization, factorization, and range inclusion of operators on Hilbert
space, Proc. Amer. Math. Soc. 17 (1966), 413-415.
[8] A.E. Frazho and W. Bosri, An operator perspective on signals and systems, Oper. Theory
Adv. Appl. 204, Birkhauser Verlag, Basel, 2010.
[9] B. Sz.-Nagy, C. Foias, H. Bercovici and L. K´erchy, Harmonic analysis of operators on Hilbert
space, Springer, New York, 2009.
[10] N.K. Nikol'skii, Treatise on the shift operator, Grundlehren 273, Springer Verlag, Berlin
1986.
[11] M. Rosenblum and J. Rovnyak, Hardy classes and operator theory, Oxford Mathematical
Monographs, Oxford Science Publications, The Clarendon Press, Oxford University Press,
New York, 1985.
[12] T.T. Trent, An algorithm for the corona solutions on H∞(D), Integr. Equ. Oper. Theory 59
(2007), 421 -- 435.
[13] T.T. Trent, A Constructive Proof of the Leech Theorem for Rational Matrix Functions,
Integr. Equ. Oper. Theory 75 (2013), 39 -- 48.
RATIONAL MATRIX SOLUTIONS TO THE LEECH EQUATION
15
School of Computer, Statistical and Mathematical Sciences, North-West Univer-
sity, Potchefstroom 2520, South Africa
E-mail address: [email protected]
|
1609.08231 | 1 | 1609 | 2016-09-27T00:41:18 | New properties for certain positive semidefinite matrices | [
"math.FA"
] | We bring in some new notions associated with $2\times 2$ block positive semidefinite matrices. These notions concern the inequalities between the singular values of the off diagonal blocks and the eigenvalues of the arithmetic mean or geometric mean of the diagonal blocks. We investigate some relations between them. Many examples are included to illustrate these relations. | math.FA | math | NEW PROPERTIES FOR CERTAIN POSITIVE
SEMIDEFINITE MATRICES
MINGHUA LIN
Abstract. We bring in some new notions associated with 2 × 2 block
positive semidefinite matrices. These notions concern the inequalities
between the singular values of the off diagonal blocks and the eigenvalues
of the arithmetic mean or geometric mean of the diagonal blocks. We
investigate some relations between them. Many examples are included
to illustrate these relations.
1. Introduction
6
1
0
2
p
e
S
7
2
]
.
A
F
h
t
a
m
[
1
v
1
3
2
8
0
.
9
0
6
1
:
v
i
X
r
a
Matrices considered here have entries from the complex number field.
We are interested in positive semidefinite matrices partitioned into 2 × 2
blocks
(cid:20)M11 M12
(cid:21)
M∗
.
M =
12 M22
(1.1)
In particular, we assume that the off-diagonal blocks are square, say n × n.
The study of eigenvalues or singular values is of central importance in
matrix analysis. It could date back to Cauchy, who established an interlacing
theorem for a bordered Hermitian matrix [10, p. 242]. Notable results also
include that of Schur [17, p. 300] and Fan [17, p. 308], who revealed a
majorization relation between the eigenvalues and diagonal entries (or the
eigenvalues of diagonal blocks) of a Hermitian matrix. For more information,
we refer to [4, Chapter III, VII, IX], [11, Chapter 3].
To proceed, let us fix some notation. For any n×n matrix A, the singular
values sj(A) are nonincreasingly arranged, s1(A) ≥ s2(A) ≥ ··· ≥ sn(A). If
A is Hermitian, we also arrange its eigenvalues λj(A) in nonincreasing order
λ1(A) ≥ λ2(A) ≥ ··· ≥ λn(A). The geometric mean of two n × n positive
definite matrices A and B, denoted by A(cid:93)B, is the positive definite solution
of the Riccati equation XB−1X = A and it has the explicit expression
A(cid:93)B = B1/2(B−1/2AB−1/2)1/2B1/2. The notion of geometric mean can be
2010 Mathematics Subject Classification. 15A45, 15A42, 47A30.
Key words and phrases. singular value, eigenvalue, positive partial transpose, inequality.
1
2
M. LIN
uniquely extended to all positive semidefinite matrices by a limit from above:
A(cid:93)B := lim
→0+
(A + In)(cid:93)(B + In),
where In is the n × n identity matrix [5, Chapter 4]. For two Hermitian
matrices A and B of the same size, we write A ≥ B (A > B) to mean that
A − B is positive semidefinite (positive definite).
Now we introduce new notions to be investigated in this paper.
Definition 1.1. Consider the matrix given in (1.1),
(i) M is said to have la-property if
λj(M11 + M22),
k = 1, . . . , n.
k(cid:89)
j=1
2sj(M12) ≤ k(cid:89)
sj(M12) ≤ k(cid:89)
k(cid:89)
j=1
(ii) M is said to have lg-property if
λj(M11(cid:93)M22),
k = 1, . . . , n.
j=1
j=1
(iii) M is said to have a-property if
2sj(M12) ≤ λj(M11 + M22),
j = 1, . . . , n.
(iv) M is said to have g-property if
sj(M12) ≤ λj(M11(cid:93)M22),
j = 1, . . . , n.
It is clear that g-property (resp. lg-property) is stronger than a-property
(resp. la-property) and g-property (resp. a-property) is stronger than lg-
property (resp. la-property). So one may conlude that la-property is the
weakest one and g-property is the strongest one among these four properties.
But not all positive semidefinite matrices have la-property. For example,
the positive semidefinite matrix
(cid:20)M11 M12
(cid:21)
M∗
12 M22
=
1 0 0 1
0 0 0 0
0 0 0 0
1 0 0 1
does not have la-property, as
2s1(M12) = 2 > λ1(M11 + M22) = 1.
NEW PROPERTIES
3
2. Basic observations
Our first observation is the following
Proposition 2.1. Consider the positive semidefinite matrix M given in
(1.1) with each block 2 × 2. If M has la-property, then M has either a-
property or lg-property.
Proof. As we have assumed that M has la-property, in particular,
(2.1)
s1(M12) ≤ 1
2
λ1(M11 + M22).
First of all, let us observe the following relation
λ1(M11(cid:93)M22)λ2(M11(cid:93)M22) = det M11(cid:93)M22
(cid:112)
det M11M22
=
≥ det M12 = s1(M12)s2(M12),
in which the inequality is due to the fact that determinant function is
Liebian [4, p. 269]. That is,
(2.2)
λ1(M11(cid:93)M22)λ2(M11(cid:93)M22) ≥ s1(M12)s2(M12),
Thus, from (2.2), if M does not have lg-property, then we must have
(2.3)
λ1(M11(cid:93)M22) < s1(M12).
Inequalities (2.2) and (2.3) together yield
λ2(M11(cid:93)M22) ≥ s2(M12).
Taking into account that λ2(M11(cid:93)M22) ≤ 1
2λ2(M11 + M22) gives
(2.4)
s2(M12) ≤ 1
2
λ2(M11 + M22).
Thus (2.1) and (2.4) together indicate that M has a-property.
Assume now that M does not have a-property, then in view of (2.1),
s2(M12) >
1
2
λ2(M11 + M22) ≥ λ2(M11(cid:93)M22).
Therefore, from (2.2) we conclude that
(2.5)
s1(M12) ≤ λ1(M11(cid:93)M22).
Thus (2.2) and (2.5) together indicates that M has lg-property
(cid:3)
The following numerical example shows that Proposition 2.1 needs not
be true if each block of M is 3 × 3. In other words, there are matrices that
have only la-property but no other three.
4
M. LIN
2
1.7353 −0.2433 1.7146
1.6438
0.7227
Example 2.2. Take M11 =
2.7266 −1.3731 −0.0930
2
−0.2433
1.7146
−1.3731
−0.0930
2.3151
0.0859
0.0859
0.7646
M22 =
and X =
,
0.7227
6.6795
−0.0445 −0.9170 −0.3964
.
0.6927 −0.3142
−0.7198 −0.2457
22 so that M is positive
0.6492
0.6492
As X is a contraction, we may take M12 = M 1/2
semidefinite (see [11, p. 207]). A calculation using Matlab shows
11 XM 1/2
λ(M11(cid:93)M22) = {7.2176, 5.5156, 1.0415};
s(M12) = {8.7154, 3.2243, 1.4755};
λ((M11 + M22)/2) = {26.9680, 9.2207, 1.0879}.
So M has la-property. But it neither has lg-property nor a-property.
Remark 2.3. In the computation of matrix geometric mean, we used the
programme developed by Bini and Iannazzo [8].
The next example shows that g-property could be more strict than the
other three.
Example 2.4. Take M11 =
(cid:20)1 0
(cid:21)2
0 2
(cid:20) 2 −1
(cid:21)2
−1
(cid:20)−1 1
(cid:21)
.
√
2
2
As X is a unitary, we may take M12 = M 1/2
semidefinite. In this case, det M12 =
culation using Matlab shows
√
, M22 =
and X =
1
11 XM 1/2
1
so that M is positive
det M11M22 = det M11(cid:93)M22. A cal-
1
22
λ(M11(cid:93)M22) = {3.0760, 0.6502};
s(M12) = {2.8284, 0.7071};
λ((M11 + M22)/2) = {4.5000, 1.5000}.
So M has lg-property (one only needs to check that λ1(M11(cid:93)M22) > s1(M12))
and a-property. But it does not have g-property.
There are examples that M has lg-property but no a-property; see Sec-
tion 3. Examples that M has a-property but no lg-property are considered
in Section 4. We may use a Venn diagram to illustrate relations between
these four notions.
NEW PROPERTIES
5
3. PPT matrices
(cid:20)M11 M12
(cid:21)
M∗
12 M22
is said to have positive
(cid:21)
(cid:20)M11 M∗
12
M12 M22
In this section, we present initial incentive for the investigation in this
paper.
A positive semidefinite matrix M =
partial transpose (PPT) if its partial transpose Mτ =
is also
positive semidefinite. The partial transpose is an intriguing operation, it
is different from the conventional transpose in many aspects, for example,
(Mτ )2 (cid:54)= (M2)τ in general.
The following theorem was proved in [13].
Theorem 3.1. Let M =
be PPT. Then
(cid:20)M11 M12
(cid:21)
M∗
12 M22
k(cid:89)
sj(M12) ≤ k(cid:89)
λj(M11(cid:93)M22),
j = 1, . . . , n.
j=1
j=1
In other words, PPT matrices have lg-property. It is noteworthy that
not all PPT matrices have a-property. We present two different examples.
Example 3.2. The matrix(cid:20) A2 + B2 AB + BA
(cid:21)
AB + BA A2 + B2
is PPT whenever A, B are n× n Hermitian matrices. The a-property in this
example is equivalent to
(3.1)
sj(AB + BA) ≤ λj(A2 + B2),
j = 1, . . . , n.
However, (3.1) fails in general; see [7, p. 2182]. This example has been given
in [13].
6
M. LIN
Example 3.3. It is easy to see that if A, B are n × n positive semidefinite
matrices, then
(cid:20)A + B A − B
A − B A + B
(cid:21)
is PPT. The a-property in this example is equivalent to
(3.2)
sj(A − B) ≤ λj(A + B),
j = 1, . . . , n.
Again, the concrete example in [7, p. 2179] shows (3.2) fails in general.
In the sequel, the norm (cid:107) · (cid:107) stands for the usual spectral norm, i.e.,
(cid:107) · (cid:107) = s1(·). Ando proved the following norm inequality.
Theorem 3.4. [2, Theorem 3.3] Let M =
be PPT. Then
(cid:20)M11 M12
(cid:21)
12 M22
M∗
(cid:107)M12(cid:107) ≤ (cid:107)M11(cid:93)M22(cid:107).
Though Theorem 3.4 looks weaker in form than Theorem 3.1, we use a
that
standard approach to show they are essentially equivalent.
If ∧k(X), 1 ≤ k ≤ n, denotes the k-th antisymmetric tensor power [4, p.
j=1 sj(X). Note
16] of an n × n matrix X, then (cid:107) ∧k (X)(cid:107) = s1(∧k(X)) =(cid:81)k
(cid:17)
(cid:16) ∧k (M11)
(cid:17)
(cid:16) ∧k (M22)
(cid:21)
(cid:20)∧k(M11) ∧k(M12)
Thus, to show that Theorem 3.4 implies Theorem 3.1, it suffices to show
(cid:20)M11 M12
∧k(M11(cid:93)M22) =
(cid:21)
(cid:93)
.
∧k(M∗
12) ∧k(M22)
. Without loss
that if
M∗
12 M22
is PPT, then so is
of generality, we assume M22 is positive definite (the general case follows by
a standard continuity argument). Consider the Schur complement
∧k(M11) − ∧k(M12)(∧k(M22))−1 ∧k (M∗
12)
= ∧k(M11) − ∧k(M12) ∧k (M−1
22 ) ∧k (M∗
12)
= ∧k(M11) − ∧k(M12M−1
22 M∗
12)
12) ≥ 0,
≥ ∧k(M11 − M12M−1
22 M∗
in which the first inequality is by [5, (4.20), p. 114]. Similarly,
12)(∧k(M22))−1 ∧k (M12)
22 M12) ≥ 0,
∧k(M11) − ∧k(M∗
12M−1
≥ ∧k(M11 − M∗
as desired.
We remark that a simple proof of Theorem 3.1 has appeared in [12].
NEW PROPERTIES
7
A typical example of PPT matrix is the Hua matrix which has the form
(cid:20)(In − A∗A)−1
(In − A∗B)−1
(cid:21)
,
(In − B∗A)−1
(In − B∗B)−1
where A, B are m × n strictly contractive matrices. So the Hua matrix has
lg-property. In [13, Theorem 3.3], we proved that it has a-property. Later,
we used a simpler argument to show that the Hua matrix has g-property;
see [16, Theorem 3.2].
(cid:20) A X
(cid:21)
In the next two examples, we assume that
, where A, X, B are
n × n, is positive semidefinite. The trace of a square matrix X is denoted
by trX.
X∗ B
Example 3.5. It is known that(cid:20) Φ(A) Φ(X)
(cid:21)
Φ(X∗) Φ(B)
,
where Φ : X (cid:55)→ X+(trX)In, is PPT (see [14]). So the matrix
has lg-property.
(cid:20) Φ(A) Φ(X)
(cid:21)
Φ(X∗) Φ(B)
It is recently proved [15] that the matrix
(cid:16)
namely,
2sj
(cid:17) ≤ sj
(cid:16)
Φ(X)
Φ(A) + Φ(B)
,
(cid:17)
(cid:20) Φ(A) Φ(X)
(cid:21)
Φ(X∗) Φ(B)
has a-property,
j = 1, . . . , n.
(cid:20) Φ(A) Φ(X)
(cid:21)
Φ(X∗) Φ(B)
Numerical experiments suggest that the matrix
has g-
property, which we haven't been able to prove yet.
Example 3.6. If we consider the map Ψ : X (cid:55)→ 2(trX)In − X, then using
the approach in [14] we can show that the matrix
(cid:20) Ψ(A) Ψ(X)
(cid:21)
Ψ(X∗) Ψ(B)
is PPT.
Though there are strong numerical evidence suggesting that this block
matrix also has g-property, yet we have not even been able to show that it
has a-property.
4. Non-PPT matrices
Consider the positive semidefinite matrix M =
(cid:20)M11 M12
(cid:21)
M∗
12 M22
. Assume
further that the off diagonal block M12 is unitary. It is easy to see that
8
M. LIN
the block matrix M is not PPT in general. The next proposition says that
under the extra unitary assumption on M12, the matrix M has lg-property.
Proposition 4.1. If M12 in the positive semidefinite matrix M =
is unitary, then
k(cid:89)
sj(M12) ≤ k(cid:89)
1 =
λj(M11(cid:93)M22),
k = 1, . . . , n.
j=1
j=1
Proof. As ∧k(M12), 1 ≤ k ≤ n, is again unitary, similar to the argument in
Section 3, the required inequality is equivalent to
(4.1)
(cid:107)M11(cid:93)M22(cid:107) ≥ 1.
Inequality (4.1) is due to Ando [2, Theorem 3.5]. We include a proof for
completeness. First of all, we notice that M11 and M22 are nonsingular, as
we may write M12 = M 1/2
for some contraction C (see [11, p. 207]).
We need to prove that
11 CM 1/2
22
M11 ≥ M12M−1
22 M∗
12 =⇒ (cid:107)M11(cid:93)M22(cid:107) ≥ 1.
Assume otherwise that (cid:107)M11(cid:93)M22(cid:107) < 1, i.e., M11(cid:93)M22 < In, then due to the
monotoncity of geometric mean
X := (M12M−1
22 M∗
12)(cid:93)M22 < In
and so (cid:107)X(cid:107) < 1. Moreoever,
M12M−1
22 M∗
12 = XM−1
22 X.
Taking norms on both sides gives
(cid:107)M−1
22 (cid:107) = (cid:107)M12M−1
22 M∗
12(cid:107) = (cid:107)XM−1
22 X(cid:107) ≤ (cid:107)X(cid:107)2(cid:107)M−1
(cid:20)M11 M12
(cid:21)
M∗
12 M22
22 (cid:107) < (cid:107)M−1
22 (cid:107).
(cid:21)
(cid:20)M11 M12
M∗
12 M22
(cid:3)
with
A contradiction.
To see that the positive semidefinite matrix M =
M12 unitary does not have a-property in general, consider a special case
M22 = M∗
11 M12. Then a-property in this case is equivalent to
λj(M11 + M∗
11 M12) ≥ 2,
j = 1, . . . , n.
12M−1
12M−1
(4.2)
Take M11 =
(cid:20)1 0
(cid:21)
0 2
(cid:20)0 1
(cid:21)
. Then
, M12 =
1 0
12M−1
M11 + M∗
11 M12 =
(cid:20)3/2 0
(cid:21)
0
3
.
And so,
NEW PROPERTIES
9
λ2(M11 + M∗
12M−1
11 M12) = 3/2 < 2s2(M12) = 2,
violating (4.2).
Now we present two examples about a positive semidefinite matrix that
has a-property but no lg-property.
Example 4.2. Bhatia and Kittaneh [6] proved that if A, B are n×n positive
semidefinite matrices, then
2sj(AB) ≤ λj(A2 + B2),
j = 1, . . . , n.
This in particular says that the matrix
(cid:20) A2 AB
(cid:21)
BA B2
has a-property. Now we explain that the matrix does not have lg-property.
It sufficies to show that
(4.3)
fails in general.
(cid:107)AB(cid:107) ≤ (cid:107)A2(cid:93)B2(cid:107).
Indeed, the correct result is that the inequality sign in (4.3) should be
reversed. In [3], Ando and Hiai proved
(cid:107)A2(cid:93)B2(cid:107) ≤ (cid:107)A(cid:93)B(cid:107)2.
Combinging with
(cid:107)A(cid:93)B(cid:107)2 ≤ (cid:107)A1/2B1/2(cid:107)2 = λ1(AB) ≤ (cid:107)AB(cid:107)
gives
(4.4)
(cid:107)AB(cid:107) ≥ (cid:107)A2(cid:93)B2(cid:107).
In particular, if A, B do not commute, then the inequalities in (4.4) are
strict.
Example 4.3. Let A, B be n×n positive semidefinite matrices. We consider
(cid:20)(cid:107)B(cid:107)A AB
BA (cid:107)A(cid:107)B
(cid:21)
.
10
M. LIN
This matrix is positive semidefinite, for (cid:107)B(cid:107)A is positive semidefinite
and the Schur complement
(cid:107)B(cid:107)A − AB((cid:107)A(cid:107)B)−1BA = (cid:107)B(cid:107)A − 1
is positive semidefinite.
≥ (cid:107)B(cid:107)
(cid:18)
(cid:19)
(cid:107)A(cid:107)ABA
A − 1
(cid:107)A(cid:107)A2
The matrix has a-property, which is proved in Proposition 4.4. However,
the matrix does not have lg-property as we have a simple numerical example.
Take
A =
(cid:20)1.7 1.3
(cid:21)
(cid:20) 2.2 −1.5
(cid:21)
(cid:112)(cid:107)A(cid:107)(cid:107)B(cid:107)(cid:107)A(cid:93)B(cid:107) ≈ 1.2055 < (cid:107)AB(cid:107) ≈ 2.6515.
−1.5
B =
1.3
1
,
.
1.1
Then a calculation gives
Proposition 4.4. Let A, B be n × n positive semidefinite matrices. Then
2sj(AB) ≤ λj((cid:107)B(cid:107)A + (cid:107)A(cid:107)B),
j = 1, . . . , n.
Proof. The Bhatia-Kittaneh-Drury inequality [9] says that if X, Y are n× n
positive semidefinite matrices, then
(cid:113)
sj(XY ) ≤ λj(X + Y ),
2
This implies
λj((cid:107)B(cid:107)A + (cid:107)A(cid:107)B) ≥ 2
sj((cid:107)A(cid:107)(cid:107)B(cid:107)AB) = 2
(cid:113)
j = 1, . . . , n.
(cid:113)(cid:107)A(cid:107)(cid:107)B(cid:107)sj(AB).
But it is clear that sj(AB) ≤ (cid:107)A(cid:107)(cid:107)B(cid:107) for j = 1, . . . , n. So the required
(cid:3)
inequality is confirmed.
Finally, we present a simple non-PPT matrix that has g-property.
Example 4.5. Let A be any n×n positive semidefinite matrix. The positive
semidefinite matrix
(cid:21)
(cid:20) I
A
A∗ A∗A
is not PPT in general, but it has g-property. This is because
sj(A) = λj(A) = λj(I(cid:93)A∗A)
for j = 1, . . . , n.
NEW PROPERTIES
11
5. Concluding remarks
We point out some closely related questions for future considerations.
†. Besides the challenging problems described in Example 3.5 and Example
3.6, other maps could be considered/constructed to meet these four prop-
erties.
‡. A generic criterion for lg-property is the PPT condition (Theorem 3.1).
It would be of great interest to know similar conditions for other three
properties.
§. One may add two new relations to Definition 1.1. More precisely,
(v) M is said to have ma-property if
λj(M11 + M22),
k = 1, . . . , n.
k(cid:88)
j=1
2sj(M12) ≤ k(cid:88)
sj(M12) ≤ k(cid:88)
k(cid:88)
j=1
(vi) M is said to have mg-property if
λj(M11(cid:93)M22),
k = 1, . . . , n.
j=1
j=1
This of course deserves further investigation.
Acknowledgments. The work is supported in part by a grant from NNSFC.
References
[1] T. Ando, Positivity of operator-matrices of Hua-type, Banach J. Math. Anal.
2 (2008) 1-8.
[2] T. Ando, Geometric mean and norm Schwarz inequality, Ann. Funct. Anal.
7 (2016) 1-8.
[3] T. Ando, F. Hiai, Log majorization and complementary Golden-Thompson
type inequalities, Linear Algebra Appl. 197/198 (1994) 113-131.
[4] R. Bhatia, Matrix Analysis, GTM 169, Springer-Verlag, New York, 1997.
[5] R. Bhatia, Positive Definite Matrices, Princeton University Press, Princeton,
2007.
[6] R. Bhatia, F. Kittaneh, On the singular values of a product of operators,
SIAM J. Matrix Anal. Appl. 11 (1990) 272-277.
[7] R. Bhatia, F. Kittaneh, The matrix arithmetic-geometric mean inequality
revisited, Linear Algebra Appl. 428 (2008) 2177-2191.
[8] D. Bini, B Iannazzo, The Matrix Means Toolbox, available at http:
//bezout.dm.unipi.it/software/mmtoolbox/
[9] S. W. Drury, On a question of Bhatia and Kittaneh, Linear Algebra Appl.
437 (2012) 1955-1960.
[10] R. A. Horn, C. R. Johnson, Matrix Analysis, Cambridge University Press,
2nd ed., 2013.
12
M. LIN
[11] R. A. Horn, C. R. Johnson, Topics in Matrix Analysis, Cambridge University
Press, 1991.
[12] E.-Y. Lee, The off-diagonal block of a PPT matrix, Linear Algebra Appl.
[13] M. Lin, Inequalities related to 2 × 2 block PPT matrices, Oper. Matrices, 9
486 (2015) 449-453.
(2015) 917-924.
[14] M. Lin, A completely PPT map, Linear Algebra Appl. 459 (2014) 404-410.
[15] M. Lin, A singular value inequality related to a linear map, Electron. J.
Linear Algebra, 31 (2016) 120-124.
[16] M. Lin, The Hua matrix and inequalities related to contractive matrices,
Linear Algebra Appl. 511 (2016) 22-30.
[17] A. W. Marshall, I. Olkin, B. Arnold, Inequalities: Theory of Majorization
and Its Applications, Springer, New York, 2nd ed., 2011.
Department of Mathematics, Shanghai University, Shanghai, 200444, China,
E-mail address: m [email protected]
|
1906.09014 | 1 | 1906 | 2019-06-21T08:58:50 | Cauchy-Riemann equations for free noncommutative functions | [
"math.FA"
] | In classical complex analysis analyticity of a complex function $f$ is equivalent to differentiability of its real and imaginary parts $u$ and $v$, respectively, together with the Cauchy-Riemann equations for the partial derivatives of $u$ and $v$. We extend this result to the context of free noncommutative functions on tuples of matrices of arbitrary size. In this context, the real and imaginary parts become so called real noncommutative functions, as appeared recently in the context of L\"owner's theorem in several noncommutative variables. Additionally, as part of our investigation of real noncommutative functions, we show that real noncommutative functions are in fact noncommutative functions. | math.FA | math |
CAUCHY-RIEMANN EQUATIONS FOR FREE
NONCOMMUTATIVE FUNCTIONS
S. TER HORST AND E.M. KLEM
Abstract. In classical complex analysis analyticity of a complex function f
is equivalent to differentiability of its real and imaginary parts u and v, respec-
tively, together with the Cauchy-Riemann equations for the partial derivatives
of u and v. We extend this result to the context of free noncommutative
functions on tuples of matrices of arbitrary size.
In this context, the real
and imaginary parts become so called real noncommutative functions, as ap-
peared recently in the context of Lowner's theorem in several noncommutative
variables. Additionally, as part of our investigation of real noncommutative
functions, we show that real noncommutative functions are in fact noncom-
mutative functions.
1. Introduction
Over the last decade a theory of free noncommutative (nc) functions that are
evaluated in tuples of matrices of arbitrary size was developed. The theory becomes
particularly rich when the functions have a domain that is assumed to be right (or
left) admissible, in which case the functions admit a Taylor expansion and, under
mild boundedness assumptions, are analytic. We refer to [7] for the first book that
presents a comprehensive account of the theory, as well the seminal paper [14] by
J.L. Taylor. Precise definitions will be given a little further in this introduction.
More recently, in connection with Lowner's theorem [10, 9, 8], the notion of
real nc functions appeared. These functions have domains that consist of tuples
on Hermitian matrices, precluding the right (or left) admissibility property, and
satisfy slightly different conditions. Another instance where real nc function come
up in a natural way is as the real and imaginary part of a nc function.
In the
present paper we derive the noncommutative Cauchy-Riemann equations for the
real and imaginary part of a nc function and consider the question when two real
nc functions satisfying the noncommutative Cauchy-Riemann equations appear as
the real and imaginary part of a nc function.
We will now provide more precise definitions and state our main result. Through-
out Cn×n denotes the complex vector space of n × n complex matrices and Hn the
real vector space of n× n Hermitian matrices. For a positive integer d, we consider
functions with domains in
Cd
nc :=
(Cd)n×n =
∞an=1
(Cn×n)d
∞an=1
or Hd
nc :=
∞an=1
(Hn)d.
2010 Mathematics Subject Classification. Primary 32A10; Secondary 46L52, 26B05.
Key words and phrases. Cauchy-Riemann equations, free noncommutative functions, real non-
commutative functions.
This work is based on the research supported in part by the National Research Foundation of
South Africa (Grant Numbers 90670, 118583, and 94069).
1
In case d = 1 we omit it as a superscript and simply write Cnc and Hnc. A subset
D of Cd
nc is said to be a nc set in case it respects direct sums:
nc or Hd
X, Y ∈ D =⇒ X ⊕ Y =(cid:20) X 0
Y (cid:21) ∈ D.
0
In some papers the converse implication as well as additional features are also
assumed, cf., [10, 9]. See Lemma 2.6 below as well as the paragraph preceding this
lemma. For a nc set D and a positive integer n we define Dn := D ∩ Cn×n. A nc
set D ⊂ Cd
X ∈ Dn, Y ∈ Dm, Z ∈ Cn×m =⇒(cid:20) X rZ
Y (cid:21) ∈ Dn+m for some 0 6= r ∈ C. (1.1)
nc is called right admissible in case
In case the nc set D is right admissible and closed under similarity, then the "for
some" part in the right-hand side of (1.1) can be replaced by "for all." There is
a dual notion of left admissibility, see page 18 and onwards in [7], but we will not
need this notion in the present paper.
0
nc is called a nc function
A function w : D → Cnc whose domain D is a nc set in Cd
in case it has the following properties:
(NC-i) w is graded, i.e., w(Dn) ⊂ Cn×n for n = 1, 2, . . .;
(NC-ii) w respects direct sums, i.e., for all X, Y ∈ D we have
w(X ⊕ Y ) = w(X) ⊕ w(Y );
(NC-iii) w respects similarities, i.e., for all X ∈ Dn, S ∈ Cn×n invertible so that
SXS−1 ∈ Dn, we have
w(SXS−1) = Sw(X)S−1.
Much of the theory of nc functions developed in [7] is for nc functions whose domains
are right (or left) admissible, in which case for each X, Y , Z and r 6= 0 as in (1.1)
one can define the right difference-differential operator ∆w(X, Y ) at the point Z
via
w(cid:18)(cid:20) X rZ
Y (cid:21)(cid:19) =(cid:20) w(X)
0
r∆w(X, Y )(Z)
0
w(Y )
(cid:21) ,
(1.2)
with the zero and two block diagonal entries following from (NCi) -- (NCiii). This
right difference-differential operator is linear in Z and provides a difference formula
for w leading to the so-called Taylor-Taylor expansion of w, and, under certain
boundedness assumptions on w, provides the Gateaux-derivative of w; see [7] for
an elaborate treatment. Recall that the Gateaux- or G-derivative of a function
g : Dg → Y with domain Dg ⊂ X , with X and Y Banach spaces over the field
K = C or K = R, at a point X ∈ X in Dg in the direction Z ∈ X is given by
Dg(X)(Z) := lim
K∋t→0
g(X + tZ) − g(X)
t
,
(1.3)
provided the limit exist. Then g is said to be Gateaux- or G-differentiable in case
Dg is open and Dg(X)(Z) exists for all X ∈ Dg and all Z ∈ X . In the case of nc
functions, G-differentiability means that for each positive integer n the restriction of
the domain to (Cn×n)d should be G-differentiable; see Section 3 for further details
and references on G-differentiability as well as Fr´echet- or F-differentiability.
A function w : Dw → Hnc is called a real nc function in case its domain D is a
nc which is graded and respects direct sums, i.e., (NC-i) and
nc set contained in Hd
(NC-ii) above hold, and
2
(RNC-iii) w respects unitary equivalence, i.e., for all X ∈ Dn, U ∈ Cn×n unitary so
that U XU ∗ ∈ Dn, we have
w(U XU ∗) = U w(X)U ∗.
Despite the seeming limitation of unitary equivalence over similarity, one of the
contributions of the present paper is the observation that real nc functions are
also nc functions, see Theorem 2.1 below. Hence (NC-i), (NC-ii) and (RNC-iii)
imply (NC-iii). This result relies heavily on the fact that the domains of real nc
functions consist of tuples of Hermitian matrices only. The latter also implies that
the domains of real nc functions are 'nowhere right admissible,' and hence much of
the theory developed in [7] does not apply to real nc functions.
Now, given an nc function f on a right admissible domain Df ⊂ Cd
nc, we write
f (A + iB) = u(A, B) + iv(A, B), A + iB ∈ Df ,
(1.4)
for A, B ∈ Hd
nc of the same size and with
u(A, B) := Re f (A + iB) and v(A, B) := Im f (A + iB).
This defines real nc functions u and v on domain D = {(A, B) ∈ H2d
nc : A+iB ∈ Df},
which is open in H2d
nc; in both cases open means
n and (Cn×n)d,
that the restriction of the domain to n × n matrices is open in H2d
respectively. Furthermore, in case f is G-differentiable, then so are u and v and their
G-derivatives satisfy the following noncommutative Cauchy-Riemann equations
nc precisely when Df is open in Cd
Du(A, B)(Z1, Z2) = Dv(A, B)(−Z2, Z1), (A, B) ∈ Dn, Z1, Z2 ∈ Hn, n ∈ N. (1.5)
See Theorem 4.1 for these claims as well as additional results.
Conversely, one may wonder whether G-differentiable real nc functions u and v
with open domains Du and Dv, respectively, in H2d
nc that satisfy (1.5) on D = Du∩Dv
define a nc function f via (1.4). For this purpose, G-differentiability does not seem
to be the appropriate notion of differentiability, and we will rather assume the
stronger notion of F-differentiability, in which case the derivative is still obtained
via (1.3); see Section 3 for further details. Even in classical complex analysis this
phenomenon occurs, see [2, 4] as well as Remark 5.6 below. Our main result is the
following theorem.
Theorem 1.1. Let u and v be real nc functions with open domains Du and Dv,
respectively, in H2d
nc that are F-differentiable and satisfy the nc Cauchy-Riemann
equations (1.5) on D = Du ∩ Dv. Define f on Df = {A + iB ∈ Cd
nc : (A, B) ∈ D}
via (1.4). Then f is a F-differentiable nc function.
Apart from the present introduction, this paper consists of four sections.
In
Section 2 we prove that real nc functions are nc function, consider some exam-
ples and look at domain extensions. Next, in Section 3 we review the notions of
Gateaux- and Fr´echet differentiability for nc functions. The domains of real nc
functions are not right-admissible so that the G-derivative cannot be determined
algebraically through the difference-differential operator. In the following section
we derive properties of the real and imaginary parts of an nc function, including
the nc Cauchy-Riemann equations. Finally, in Section 5 we consider the converse
direction and prove Theorem 1.1.
3
2. Real nc functions are nc functions
In this section we focus on real nc functions only, without assuming any form of
differentiability. Our main result is the following theorem.
Theorem 2.1. Real nc functions are nc functions.
In order to prove this result we first show that real nc functions also respect
intertwining.
Proposition 2.2. A graded function w : D → Hnc on a nc set D ⊂ Hnc respects
direct sums and unitary equivalence if and only if it respects intertwining: if X ∈
Dn, Y ∈ Dm, and T ∈ Cn×m so that XT = T Y , then w(X)T = T w(Y ).
Proof. The necessity follows from Proposition 2.1 in [7]. Assume w respects direct
sums and unitary equivalence, i.e., w is a real nc function. Let X ∈ Dn, Y ∈
Dm, and T0 ∈ Cn×m so that XT0 = T0Y .
If T0 = 0, then it is trivial that
w(X)T0 = T0w(Y ), so assume T0 6= 0. Set T = kT0k−1T0 so that kTk = 1.
Let DT := (I − T ∗T )1/2 and DT ∗ := (I − T T ∗)1/2 be the defect matrices of the
contractions T and T ∗, respectively. Since X and Y are Hermitian we have
T ∗X = Y T ∗.
T ∗ = X(I − T T ∗) = X − T Y T ∗ = X − T T ∗X = (I − T T ∗)X = D2
T ∗ X,
and similarly Y D2
Y DT = DT Y . Let UT be the unitary rotation matrix associated with T :
T Y . By the spectral theorem we have XDT ∗ = DT ∗ X and
T = D2
UT =(cid:20) T DT ∗
DT −T ∗ (cid:21) .
Therefore
XD2
Then
(cid:20) X 0
0
Hence
Since w respects direct sums and unitary similarities, we have that
0
0
U ∗
Y (cid:21) UT =(cid:20) XT XDT ∗
DT Y −T ∗X (cid:21) = UT(cid:20) Y
Y DT −Y T ∗ (cid:21) =(cid:20) T Y DT ∗ X
T(cid:20) X 0
Y (cid:21) UT =(cid:20) Y
0 X (cid:21) ∈ Dn+m.
w(X) (cid:21) = w(cid:18)(cid:20)Y
0 X(cid:21)(cid:19) = w(cid:18)U ∗
T(cid:20)X 0
T(cid:20)w(X)
T w(cid:18)(cid:20)X 0
Y(cid:21)(cid:19) UT = U ∗
(cid:20) w(X)T
w(Y )(cid:21) UT = UT(cid:20)w(Y )
w(Y )DT −w(Y )T ∗ (cid:21) =(cid:20)w(X)
Y(cid:21) UT(cid:19)
(cid:20) w(Y )
w(X)DT ∗
= U ∗
0
0
0
0
0
0
0
0
0
0
0 X (cid:21) .
0
w(Y )(cid:21) UT .
w(X)(cid:21) =
0
This shows that
=(cid:20) T w(Y ) DT ∗w(X)
DT w(Y ) −T ∗w(X) (cid:21) .
Comparing the left-upper corners in the above identity yields w(X)T = T w(Y ),
and thus
as desired.
w(X)T0 = kT0kw(X)T = kT0kT w(Y ) = T0w(Y )
(cid:3)
4
Proof of Theorem 2.1. This is now straightforward. By assumption w is graded
and respects direct sums. Let X ∈ Dn and T ∈ Cn×n invertible so that Y :=
T −1XT ∈ Dn. Then XT = T Y , and thus w(X)T = T w(Y ) holds by Proposition
2.2. Therefore, we have w(T −1XT ) = T −1T w(Y ) = T −1w(X)T .
(cid:3)
Remark 2.3. Theorem 2.1 shows that assumptions (NC-i),(NC-ii) and (RNC-iii)
imply (NC-iii), that is: For X ∈ Dn, S ∈ Cn×n invertible so that SXS−1 ∈ Dn, we
have
w(SXS−1) = Sw(X)S−1.
An important feature here is that Y := SXS−1 ∈ Dn implies, in particular, that Y
is Hermitian. In this case, by [6, Problem 4.1.P3], X and Y are not only similar,
but also unitarily equivalent. In fact, we have Y = U XU ∗, where U is the unitary
matrix from the polar decomposition of S. Consequently, we have w(SXS−1) =
w(U XU ∗) = U w(X)U ∗. However, to arrive at w(SXS−1) = Sw(X)S−1 it still
seems necessary to have a result like Proposition 2.2, at least for the case of positive
definite similarities.
Example 2.4. It also follows from Theorem 2.1 that real nc functions are only
distinguishable from other nc functions by the fact that their domains are contained
in Hd
nc for some positive integer d. Simple examples show that the assumption
D ⊂ Hd
nc cannot be removed without Theorem 2.1 losing its validity. Any one of
the functions
w1(X) = X ∗, w2(X) = (X ∗X)
1
2
can be defined on Cnc, where they satisfy (NC-i),(NC-ii) and (RNC-iii) but not
(NC-iii), hence they are not nc functions on Cnc, but their restrictions to Hnc are,
by Theorem 2.1.
Example 2.5. For d = 1 more intricate examples can easily be constructed. Via
the continuous functional calculus, any continuous function w with domain in R
can be extended to a real nc function on the nc set of Hermitian matrices whose
spectrum is contained in the domain of w, even when it is not differentiable. Clearly
the resulting real nc function is also not differentiable in case w is not.
It is not directly clear how a continuous function of several real variables can
be extended to a real nc function, except when the domain is restricted to tuples
of commuting matrices. In passing, we note that a (unintentional) non-example
is given in [3], where an extension of a function in several real variables to a non-
commutative domain is considered, which, after some minor modifications, can be
restricted to a nc domain in Hnc, leading to a non-graded function (it maps Hd
n to
Hnd) which does satisfy conditions (NC-ii) and (NC-iii).
Domain extensions. Since a real nc function w with domain D is a nc function,
it follows from Proposition A.3 in [7] that w can be uniquely extended to a nc
function, also denoted by w, on the similarity invariant envelop of D:
D(si) := {SXS−1 : X ∈ Dn, S ∈ Cn×n invertible}
via
w(Y ) = w(SXS−1) := Sw(X)S−1
(Y = SXS−1 ∈ D(si)).
However, in general, D(si) will not be contained in Hnc, although all matrices in
D(si) have real spectrum only and the only nilpotent matrix in D(si) is the zero
5
In the context of real nc functions it may be more
matrix 0, assuming 0 ∈ D.
natural to consider the extension of w to the unitary equivalence invariant envelop
D(ue) := {U XU ∗ : X ∈ Dn, U ∈ Cn×n unitary} = D(si) ∩ Hnc,
with w extended as before. The fact that D(ue) = D(si) ∩ Hnc again follows by [6,
Problem 4.1.P3]. As this is just the restriction to D(ue) of the extension of w to
D(si), clearly we end up with a real nc function extension of w to D(ue) which is
uniquely determined by w.
In [10, 9] real free sets (restricted to the case where tensoring is done with the
real topological vector space R = Rd) are nc sets D ⊂ Hnc that are closed under
unitary equivalence and have the following property:
(a) For X, Y ∈ Hnc we have X, Y ∈ D if and only of X ⊕ Y ∈ D.
n
One implication is true by the assumption that D is a nc set, but the other direction
need not be true for the unitary equivalence envelop of a nc set contained in Hnc.
Lemma 2.6. Let D ⊂ Hnc be a nc set. Then the unitary equivalence envelop D(ue)
of D is a real free set if and only if it is closed under injective intertwining:
If X ∈ D(ue)
, Y ∈ Hm and S ∈ Cn×m injective so that XS = SY , then Y ∈ D(ue).
Proof. Assume D(ue) is closed under injective intertwining. Since D is a nc set, so
is D(ue), by [7, Proposition A.1]. Hence it remains to show that for X, Y ∈ Hnc with
0 ] and S = S2 = [ 0
X ⊕ Y ∈ D also X, Y ∈ D. This follows by taking S = S1 = [ I
I ],
respectively, with sizes compatible with the decomposition of X⊕Y . Indeed, clearly
S1 and S2 are injective and we have (X ⊕ Y )S1 = S1X and (X ⊕ Y )S2 = S2Y .
Thus D(ue) is a nc set in Hnc which is closed under unitary equivalence and satisfies
(a), hence it is a real free set.
For the converse direction, assume D(ue) is a real free set. Take X ∈ D(ue)
,
Y ∈ Hm and S ∈ Cn×m injective so that XS = SY . Since D(ue) is closed under
0 (cid:3) with
unitary equivalence and S is injective, without loss of generality S = (cid:2) S1
S1 invertible. Then XS = SY implies Ran(S) is invariant for X. However, X is
Hermitian, so that Ran(S) is in fact a reducing subspace for X. Hence X = X1⊕X2
with respect to the same decomposition as for S. Then property (a) implies X1 ∈
Hm is in D(ue), and XS = SY yields X1S1 = S1Y , i.e., X1 = S1Y S−1
1 . Hence
X1 and Y are similar. Since X1 and Y are Hermitian, X1 and Y are also unitarily
equivalent, by [6, Problem 4.1.P3]. Hence Y is in D(ue).
(cid:3)
n
3. Differentiability of nc functions
For differentiation of vector-valued functions several notions exist, and these may
differ for real and complex vector spaces. We refer to Section III.3 in [5], Section
5.3 in [1] and Sections 2.3 and 2.4 in [11] for elaborate treatments, often at a much
higher level of generality than required here. In this paper we will only encounter
Gateaux (G-)differentiability and Fr´echet (F-)differentiability. In the context of nc
functions over complex Banach spaces these notions are discussed in Chapter 7 of
[7], with a few remarks dedicated to the case of real Banach spaces. Here we will
restrict to the case of nc functions on finite dimensional spaces, i.e., with domains
in Cd
nc and Hd
We start with the definitions of G-differentiability and F-differentiability, not
distinguishing whether the field K we work over is K = R or K = C, where in
nc with d finite, as we do throughout the paper.
6
the case of K = R we consider nc functions with domains contained in Hd
for K = C the nc functions are assumed to have a domain in Cd
a nc function defined on an open domain D in Hd
nc (for K = R) or in Cd
K = C). Then for each X ∈ Dn and n × n matrix Z (in Hd
is G-differentiable at X in direction Z in case the limit
nc and
nc. Now let w be
nc (for
n for K = R) we say w
Dw(X)(Z) := lim
K∋t→0
w(X + tZ) − w(X)
t
=
d
dt
w(X + tZ)(cid:12)(cid:12)(cid:12)(cid:12)t=0
(3.1)
exists. In that case Dw(X)(Z) is the G-derivative of w at X in direction Z. We
say that w is G-differentiable in X if w is G-differentiable at X in each direction
Z, and w is called G-differentiable if it is G-differentiable in any X ∈ Dw. If w is
G-differentiable at X ∈ D, then the map Z 7→ Dw(X)(Z) is linear in Z. We shall
usually refer to Z as the directional variable.
Following [5], we say that the nc function w is F-differentiable in X ∈ D in case
w is G-differentiable in X and the G-derivative Dw(X) at X satisfies
kw(X + Z) − w(X) − Dw(X)(Z)k
= 0.
(3.2)
lim
kZk→0
kZk
nc or Cd
Here, and in the sequel, the norm kZk for Z = (Z (1), . . . , Z (d)) in Hd
nc is
given by kZk = maxk kZ (k)k. Note that if for X ∈ D there exists a homogeneous
map Z 7→ Dw(X)(Z) that satisfies (3.2), then it must satisfy (3.1), so that w is
G-differentiable, and hence Dw(X)(Z) is in fact linear in the directional variable
Z. Hence, existence of a homogeneous map Z 7→ Dw(X)(Z) satisfying (3.2) can be
used as another definition of F-differentiability. Even in case w is F-differentiable,
we will refer to (3.1) as the G-derivative of w.
The case D ⊂ Cd
nc (K = C). This case is discussed in detail in Chapter 7 of
[7]. We just mention a few specific results relevant to the present paper and to
illustrate the contrast with the case of real nc functions. Since the domain D of w
is assumed to be open in Cd
nc it must be right-admissible and hence the difference-
differential operator ∆w(X, Y )(Z) defined via (1.2) exists for all X ∈ Dn, Y ∈ Dm
and Z ∈ (Cn×m)d.
By Theorem 7.2 in [7], w is G-differentiable in case w is locally bounded on
slices, that is, if for any n, X ∈ Dn and any Z ∈ Cn×n there exists a ε > 0
so that t 7→ w(X + tZ) is bounded for t < ε. Moreover, in that case we have
Dw(X)(Z) = ∆w(X, X)(Z), and hence the G-derivative can be determined alge-
braically by evaluating w in [ X rZ
0 X ] for small r. Furthermore, by Theorem 7.4 in
[7], w is F-differentiable in case w is locally bounded, that is, if for any n, X ∈ Dn
there exists a δ > 0 so that w is bounded on the set of Y ∈ Dn with kX − Y k < δ.
However, since we only consider the case of finite dimensional vector spaces, for
X ∈ Dn the linear map Z 7→ Dw(X)(Z) from (Cn×n)d to Cn×n is continuous,
hence G-differentiability and F-differentiability coincide, by a result of Zorn [15].
The case D ⊂ Hd
nc (K = R). The domain of real nc functions are 'nowhere right
admissible', hence one cannot in general define the difference-differential operator
∆w of a real nc function w in the way it is done for nc functions defined on a right
admissible nc set. Nonetheless, Proposition 2.5 in [10] provides a difference formula
for real nc functions, provided they are F-differentiable.
As pointed out in Example 2.5, for d = 1 any continuous function with domain
in R can be extended to a real nc function. Clearly G- or F-differentiability will not
7
follow under local boundedness properties; consider, for instance, the function w2
in Example 2.4. The theory of G- and F-differentiability for functions between real
Banach spaces is treated in Section 5.3 in [1] and Sections 2.3 and 2.4 in [11]. It is
not the case here that G- and F-differentiability coincide. By Proposition 5.3.4 in
[1] or Proposition 2.51 in [11], a sufficient condition under which G-differentiability
at a point X ∈ Dn implies F-differentiability at X is that the map Y 7→ Dw(Y )
from Dn into the space of linear operators from Hd
n to Hn is continuous at X. Even
if w is F-differentiable, there does not appear to be a general way to determine Dw
algebraically, since there is no difference-differential operator.
The formula presented in the next proposition can be seen as complementary to
the difference formula in [10, Proposition 2.5].
Proposition 3.1. Let w : D → Hnc be a G-differentiable real nc function on an
open domain D ⊂ Hd
Dw(cid:18)(cid:20)X 0
n, with n arbitrary, we have
Dw(X)(Z)
Dw(X)(Z)
0
0
nc. For X ∈ Dn and Z ∈ Hd
0 X(cid:21)(cid:19)(cid:18)(cid:20) 0 Z
Z 0(cid:21)(cid:19) =(cid:20)
X − tZ(cid:21) V =(cid:20) X tZ
tZ X(cid:21) , where
0
(cid:21) .
√2(cid:20)I
I −I(cid:21) .
1
I
Proof. Note that
V ∗(cid:20)X + tZ
0
t ∈ R, V =
Since D is open and X ∈ D, for small t both 2 × 2 block matrices are in D, and we
have
Using this formula we obtain
Dw(cid:18)(cid:20)X 0
0
0
t
t
1
=
0
0
w(cid:18)(cid:20) X tZ
w(cid:18)(cid:20) X tZ
X − tZ(cid:21) V(cid:19)
w(X − tZ)(cid:21) V
tZ X(cid:21)(cid:19) = w(cid:18)V ∗(cid:20)X + tZ
= V ∗(cid:20)w(X + tZ)
2(cid:20)w(X + tZ) + w(X − tZ) w(X + tZ) − w(X − tZ)
w(X + tZ) − w(X − tZ) w(X + tZ) + w(X − tZ)(cid:21) .
tZ X(cid:21)(cid:19) − w(cid:18)(cid:20)X 0
0 X(cid:21)(cid:19)
#
0 X(cid:21)(cid:19)(cid:18)(cid:20) 0 Z
Z 0(cid:21)(cid:19) = lim
t→0" w(X+tZ)+w(X−tZ)−2w(X)
t→0
2(cid:20)Dw(X)(Z) − Dw(X)(Z) Dw(X)(Z) + Dw(X)(Z)
Dw(X)(Z) + Dw(X)(Z) Dw(X)(Z) − Dw(X)(Z)(cid:21)
=(cid:20)
− w(X−tZ)−w(X)
+ w(X−tZ)−w(X)
w(X+tZ)+w(X−tZ)−2w(X)
w(X+tZ)−w(X−tZ)
w(X+tZ)−w(X−tZ)
Dw(X)(Z)
Dw(X)(Z)
w(X+tZ)−w(X)
w(X+tZ)−w(X)
w(X+tZ)−w(X)
w(X+tZ)−w(X)
(cid:21) .
1
2
1
−t
−t
=
lim
=
lim
t
t
t
1
2
=
0
t→0
0
8
t
t
t
t
−t
+ w(X−tZ)−w(X)
− w(X−tZ)−w(X)
−t
(cid:3)
4. Real and complex part of a nc function
the introduction, we define the real and imaginary parts of f as
Throughout this section, let f be a nc function with domain Df ⊂ Cd
u : Du → Hnc,
with Du = Dv = D :=
n, A + iB ∈ Df} ⊂ H2d
nc,
u : Dv → Hnc,
∞an=1
{(A, B) : A, B ∈ Hd
nc. As in
with u and v defined for (A, B) ∈ D by
u(A, B) := Re f (A + iB) =
(f (A + iB) + f (A + iB)∗),
(4.1)
(4.2)
1
2
1
2i
v(A, B) := Im f (A + iB) =
(f (A + iB) − f (A + iB)∗).
In particular, u, v and f satisfy (1.4). The following theorem is the main result of
this section.
Theorem 4.1. Let f be a G-differentiable nc function defined on an open nc
set Df ⊂ Cd
nc and define u and v as in (4.1) and (4.2). Then u and v are G-
differentiable real nc functions, whose G-derivatives at (A, B) ∈ Dn in direction
Z = (Z1, Z2) ∈ H2d
n , for any n, are given by
Du(A, B)(Z1, Z2) = Re Df (A + iB)(Z1 + iZ2),
Dv(A, B)(Z1, Z2) = Im Df (A + iB)(Z1 + iZ2),
(4.3)
and Du and Dv satisfy the nc Cauchy-Riemann equations:
Du(A, B)(Z1, Z2) = Dv(A, B)(−Z2, Z1), (A, B) ∈ Dn, Z1, Z2 ∈ Hn, n ∈ N. (4.4)
Finally, if f is F-differentiable, then u and v are F-differentiable as well.
In order to prove this result we first prove a lemma that will also be useful in the
sequel. The result may be well-known, but we could not find it in the literature,
hence we add a proof for completeness.
Lemma 4.2. For Z = Z1 + iZ2 ∈ Cd
nc with Z1, Z2 ∈ Hd
n we have
n ≤ kZ1 + iZ2k(Cn×n)d ≤ 2k(Z1, Z2)kH2d
n
k(Z1, Z2)kH2d
(4.5)
.
Proof. Set δ = kZ1 + iZ2k = kZk, ρ = k(Z1, Z2)k = max{kZ1k,kZ2k}. Then
δ2In ≥ Z ∗Z = Z 2
1 + Z 2
2 + [iZ1, Z2] and δ2In ≥ ZZ ∗ = Z 2
1 + Z 2
2 − [iZ1, Z2].
Here [T1, T2] denotes the commutator of the square matrices T1, T2, i.e., [T1, T2] =
T1T2 − T2T1, which is applied entrywise in case T1 and T2 are tuples of matrices of
the same size. Taking the average of the above two inequalities gives
δ2In ≥ Z 2
1 + Z 2
2 .
Hence Z 2
k(Z1, Z2)k ≤ δ = kZ1 + iZ2k. For the second inequality, note that Z 2
j = 1, 2. Also, we have
j ≤ δ2In, or equivalently, kZjk ≤ δ for both j = 1, 2. Therefore, we have
j ≤ ρ2In for
k[iZ1, Z2]k = kZ1Z2 − Z2Z1k ≤ 2kZ1kkZ2k ≤ 2ρ2.
This implies −2ρ2In ≤ [iZ1, Z2] ≤ 2ρ2In, since [iZ1, Z2] ∈ Hn. We then obtain
0 ≤ Z ∗Z = Z 2
1 + Z 2
2 + [iZ1, Z2] ≤ 4ρ2In,
(cid:3)
so that kZk ≤ 2ρ = 2k(Z1, Z2)k.
9
Since the inequalities in (4.6) provide a comparison between the norms in Cd
nc
nc, the following corollary is immediate.
and H2d
Corollary 4.3. The nc set Df is open if and only if D is open.
By applying the inequalities of Lemma 4.2 to both the denominator and numer-
1
2
m. Then
ator, we obtain the following corollary.
Corollary 4.4. Let Z1, Z2 ∈ Hd
n and T1, T2 ∈ Hk
k(T1, T2)k
k(Z1, Z2)k ≤ kT1 + iT2k
kZ1 + iZ2k ≤ 2 k(T1, T2)k
k(Z1, Z2)k
Proof of Theorem 4.1. The proof is divided into four parts.
Part 1: u and v are real nc functions. It is straightforward to check that u
and v are graded and respect direct sums, since f has these properties. Clearly D
is contained in H2d
nc. It remains to verify that u and v respect unitary equivalence.
Let (A, B) ∈ Dn and U ∈ Cn×n unitary so that (U AU ∗, U BU ∗) ∈ Dn. Set X =
A + iB ∈ Df . By definition of D we have U XU ∗ ∈ Df , and since f respects
similarities, and hence unitary equivalence, we have
f (U XU ∗) = U f (X)U ∗.
(4.6)
.
The left hand side specifies to
f (U XU ∗) = f (U AU ∗ + iU BU ∗) = u(U AU ∗, U BU ∗) + iv(U AU ∗, U BU ∗),
while on the right hand side we get
U f (X)U ∗ = U f (A + iB)U ∗ = U u(A, B)U ∗ + iU v(A, B)U ∗.
Since the values of u and v are Hermitian and Hn is closed under unitary equiva-
lence, it follows that
u(U AU ∗, U BU ∗) = U u(A, B)U ∗
and v(U AU ∗, U BU ∗) = U v(A, B)U ∗.
Hence, u and v respect unitary equivalence.
Part 2: Proof of (4.3). Let X = A + iB ∈ Df,n, Z = Z1 + iZ2 ∈ (Cn×n)d with
A, B, Z1, Z2 ∈ Hnc. Assume f is G-differentiable at X in direction Z. In this part
we show that u and v are G-differentiable at (A, B) in the direction (Z1, Z2) and
that their G-derivatives satisfy
Df (A + iB)(Z1 + iZ2) = Du(A, B)(Z1, Z2) + iDv(A, B)(Z1, Z2).
(4.7)
This proves (4.3) and shows that u and v are G-differentiable in case f is G-
differentiable.
To see that our claim holds, note that for 0 6= t ∈ R we have
f (X + tZ) − f (X)
f (A + iB + t(Z1 + iZ2)) − f (A + iB)
=
t
u(A + tZ1, B + tZ2) + iv(A + tZ1, B + tZ2) − u(A, B) − iv(A, B)
v(A + tZ1, B + tZ2) − v(A, B)
u(A + tZ1, B + tZ2) − u(A, B)
+ i
t
t
=
=
t
t
.
The result follows by letting t go to 0, and noting that in the right most side of the
above identities the limits of the real and imaginary parts are independent.
10
Part 3: Cauchy-Riemann equations. The proof follows along the same lines as
the classical complex analysis proof. For X = A + iB, Z = Z1 + iZ2 and h ∈ R we
have
f (X + ihZ) − f (X) = f (A + iB + ih(Z1 + iZ2) − f (A + iB)
= f (A − hZ2 + i(B + hZ1)) − f (A + iB)
= u(A − hZ2, B + hZ1) + iv(A − hZ2, B + hZ1) − u(A, B) − iv(A, B)
= u(A − hZ2, B + hZ1) − u(A, B) + i(v(A − hZ2, B + hZ1) − v(A, B)).
Dividing by ih and taking h → 0 we obtain
Df (X)(Z) = lim
h→0
f (X + ihZ) − f (X)
ih
= lim
h→0
v(A − hZ2, B + hZ1) − v(A, B)
h
+
− i lim
h→0
u(A − hZ2, B + hZ1) − u(A, B)
h
= Dv(A, B)(−Z2, Z1) − iDu(A, B)(−Z2, Z1).
Comparing with (4.7) provides the desired equations.
Part 4: F-differentiability. Assume f is F-differentiable. This implies that f
is G-differentiable and hence u and v are G-differentiable, by Part 2. Since Df
is C-linear in the directional variable, it is clear from (4.3) that Du and Dv are
R-linear in the directional variable. Now let X = A + iB with (A, B) ∈ Dn and
Z = Z1 + iZ2 with Z1, Z2 ∈ Hd
n. Then
f (X + Z) − f (X) − Df (X)(Z) =
= u(A + tZ1, B + tZ2) − u(A, B) − Du(A, B)(Z1, Z2)+
+ i(v(A + tZ1, B + tZ2) − v(A, B) − Dv(A, B)(Z1, Z2)).
Now apply Corollary 4.4 with Z1 and Z2 as above and
T1 = u(A + tZ1, B + tZ2) − u(A, B) − Du(A, B)(Z1, Z2),
T2 = v(A + tZ1, B + tZ2) − v(A, B) − Dv(A, B)(Z1, Z2),
(4.8)
and note that Z → 0 if and only if (Z1, Z2) → 0, by Lemma 4.2. It then follows
that
lim
kZk→0
holds if and only if
kf (X + Z) − F (X) − Df (X)(Z)k
kZk
= 0
(4.9)
and
lim
k(Z1,Z2)k→0
lim
k(Z1,Z2)k→0
ku(A + tZ1, B + tZ2) − u(A, B) − Du(A, B)(Z1, Z2)k
k(Z1, Z2)k
kv(A + tZ1, B + tZ2) − v(A, B) − Dv(A, B)(Z1, Z2)k
k(Z1, Z2)k
= 0
= 0.
In particular, since (4.9) holds, and (A, B) ∈ Dn and Z1, Z2 ∈ Hd
arbitrarily, it follows that u and v are F-differentiable.
n were chosen
(cid:3)
11
The fact that the G-derivative of a G-differentiable nc function on a complex-
open domain (and hence right-admissible) can be computed algebraically, via block
upper triangular matrices, provides additional structure for its real and imaginary
parts, which enables us to compute their G-derivatives algebraically as well.
Proposition 4.5. Let f be a nc function defined on an open nc set Df ⊂ Cd
nc and
define u and v as in (4.1) -- (4.2). Let X = A + iB ∈ Df,n and Y = C + iD ∈ Df,m
and Z ∈ (Cn×m)d such that [ X Z
1
2 Z
0 Y ] ∈ Df,n+m. Then
D (cid:21)(cid:19) ∈ D
− i
2 Z
2 Z ∗ C (cid:21) ,(cid:20) B
(4.10)
2 Z ∗
1
i
(cid:18)(cid:20) A
2 Z ∗ C (cid:21) ,(cid:20) B
2 Z ∗ C (cid:21) ,(cid:20) B
1
2 Z
i
1
1
2 Z
and there exist TX,Y,1, TX,Y,2 ∈ (Cn×m)2d so that
u(cid:18)(cid:20) A
v(cid:18)(cid:20) A
D (cid:21)(cid:19) =(cid:20) u(A, B)
− i
2 Z
D (cid:21)(cid:19) =(cid:20) v(A, B)
− i
2 Z
T ∗
Moreover, if X = Y , Z = Z1 + iZ2 with Z1, Z2 ∈ Hd
2 Z ∗
2 Z ∗
X,Y,2
X,Y,1
T ∗
1
i
slices, then
TX,Y,1
u(C, D) (cid:21) ,
v(C, D) (cid:21) .
TX,Y,2
(4.11)
n and f is locally bounded on
Du(A, B)(Z1, Z2) = TX,Y,1 + T ∗
X,Y,1, Dv(A, B)(Z1, Z2) = TX,Y,2 + T ∗
X,Y,2.
together with [ X Z
0 Y ] ∈ Df,n+m yields (4.10). Since f is a nc function, we have
D (cid:21) ,
− i
2 Z
(4.12)
Proof. The decomposition
(cid:20) X Z
Y (cid:21) =(cid:20) A + iB Z1 + iZ2
0
0
1
1
2 Z
2 Z ∗ C (cid:21) + i(cid:20) B
C + iD (cid:21) =(cid:20) A
Y(cid:21)(cid:19) =(cid:20)f (X) ∆f (X, Y )(Z)
(cid:21) ,
f (Y )
0
i
2 Z ∗
f(cid:18)(cid:20)X Z
0
with ∆f (X, Y )(Z) the right nc difference-differential operator applied to f , at the
point (X, Y ) and direction Z. Note that
(cid:20)f (X) ∆f (X, Y )(Z)
f (Y )
0
(cid:21) =(cid:20) 1
2 (f (X) + f (X)∗)
1
2 ∆f (X, Y )(Z)∗
1
2 ∆f (X, Y )(Z)
2 (f (Y ) + f (Y )∗)(cid:21) +
1
=(cid:20)
+ i(cid:20)− i
i
u(A, B)
2 ∆f (X, Y )(Z)
2 (f (Y ) − f (Y )∗)(cid:21)
2 (f (X) − f (X)∗) − i
− i
2 ∆f (X, Y )(Z)∗
(cid:21) +
1
2 ∆f (X, Y )(Z)
u(C, D)
− i
2 ∆f (X, Y )(Z)
v(A, B)
v(C, D)
i
2 ∆f (X, Y )(Z)∗
(cid:21) .
+ i(cid:20)
1
2 ∆f (X, Y )(Z)∗
0 Y ]) together with (4.12) proves (4.11), where we take TX,Y,1 =
This formula for f ([ X Z
2 ∆f (X, Y )(Z) and TX,Y,2 = − 1
1
and ∆f (X, Y )(Z) = Df (X)(Z). It now follows by Theorem 4.1 that
2 ∆f (X, Y )(Z).
Now assume X = Y and f is locally bounded on slices. Then f is G-differentiable
TX,Y,1 + T ∗
X,Y,1 = Re Df (X)(Z) = Du(A, B)(Z1, Z2),
and, similarly, Dv(A, B)(Z1, Z2) = TX,Y,2 + T ∗
X,Y,2.
(cid:3)
12
Not all real nc functions "respect diagonals" as in (4.11). Also, one may wonder
whether (4.11) in some form extends beyond points of the form (4.10) in case u
and v are the real and imaginary parts of a nc function. This is also not the case
in general. We illustrate this in the following example.
Example 4.6. Consider the following three real nc functions
v(A, B) = AB + BA, w(A, B) = A2
u(A, B) = A2 − B2,
((A, B) ∈ H2
nc).
Then u and v are the real and imaginary part of the nc function f (X) = X 2. For
an arbitrary 2 × 2 block point
AZ1 − BZ2 + Z1C − Z2D
2 C2 − D2 + Z ∗
1 BZ1 + AZ2 + Z2C + Z1D
1 CD + DC + Z ∗
1 Z1 − Z ∗
1 Z2 + Z ∗
2 Z2 (cid:21) ,
2 Z1 (cid:21) ,
2 D (cid:21)(cid:19) ∈ H2
nc
2
Z ∗
Z ∗
Z ∗
we obtain:
2 B + CZ ∗
Z ∗
1 A − Z ∗
1 C (cid:21) ,(cid:20) B Z2
1 − Z2Z ∗
1 − DZ ∗
2 + Z2Z ∗
2 + DZ ∗
2 A + CZ ∗
1 AZ1 + Z1C
1 C2 + Z1Z ∗
(E, F ) :=(cid:18)(cid:20) A Z1
u(E, F ) =(cid:20) A2 − B2 + Z1Z ∗
v(E, F ) =(cid:20) AB + BA + Z1Z ∗
w(E, F ) =(cid:20) A2 + Z1Z ∗
It follows that u(E, F ) =h u(A,B)
while v(E, F ) =h v(A,B)
v(C,D)i holds if and only if
2 = −Z2Z ∗
1 = Z2Z ∗
2
∗
1 (cid:21) .
∗
Z1Z ∗
1 A + CZ ∗
1 B + Z ∗
and Z ∗
and Z ∗
Z1Z ∗
Z ∗
∗
∗
u(C,D)i holds if and only if
1 Z1 = Z ∗
2 Z2,
(4.13)
1
2 Z1.
1 Z2 = −Z ∗
(4.14)
Both conditions are true in case Z2 = ±iZ1. Conversely, these conditions on Z1
and Z2 together imply Z2 = −iZ1, but, in general, neither implies Z2 = ±iZ1 by
itself. Indeed, the identities in (4.13) imply that the kernels and co-kernels of Z1
and Z2 coincide, so that we can reduce to the case where Z1 and Z2 are invertible.
In that case, by Douglas' Lemma, (4.13) is equivalent to the existence of unitary
matrices U and V so that Z1 = U Z2 = Z2V . Assume U and V are like this, and
Z1, Z2 invertible. Then (4.14) implies
2 Z1Z ∗
However, Z2 is invertible, hence Z2Z ∗
which implies U = ±iI. Hence Z1 = ±iZ2.
2 = −Z2Z ∗
2 = −Z2Z ∗
2 is invertible. Thus we find that U = −U ∗,
On the other hand, we have w(E, F ) =h w(A,B)
w(C,D)i precisely when Z1 = 0.
Hence (4.11) holds with u or v replaced by w if and only if Z = 0, which is true for
any real nc function.
2 U ∗Z2Z ∗
2 .
2 = Z2Z ∗
2 U Z2Z ∗
1 Z2Z ∗
Z2Z ∗
∗
∗
5. Cauchy-Riemann equations: Sufficiency
In this section we prove Theorem 1.1. Throughout, let
and v : Dv → H2d
u : Du → H2d
nc
nc
be real nc functions. For notational convenience we introduce the nc set
(5.1)
D := Du ∩ Dv.
13
Now we define f on Df := {A + iB : (A, B) ∈ D} by
f (A + iB) = u(A, B) + iv(A, B)
(A + iB ∈ Df ).
(5.2)
It is easy to see that f is graded, respects direct sums as well as unitary equivalence,
since u and v have these properties. However, it is not necessarily the case that f
respects similarities, despite the fact that u and v do. The following proposition
sums up the properties that f has without further assumptions on u and v (except
G-differentiability in the last part). The claims follow directly from (5.2), hence we
omit the proof.
Proposition 5.1. Let u and v be real nc functions as in (5.1) and define f as
in (5.2). Then f is graded, respects direct sums and respects unitary equivalence.
Moreover, in case X = A + iB with (A, B) ∈ Dn, Z = Z1 + iZ2 with Z1, Z2 ∈ Hn,
for any n ∈ N, and u and v are G-differentiable at (A, B) in direction (Z1, Z2),
then
lim
R∋t→0
f (X + tZ) − f (X)
t
= Du(A, B)(Z1, Z2) + iDv(A, B)(Z1, Z2).
(5.3)
Remark 5.2. Without additional assumptions on u and v it is possible to prove
something slightly stronger than the fact that f respects unitary similarity.
If
X = A + iB ∈ Df,n and S ∈ Cn×n is invertible are such that C := SAS−1 and
D := SBS−1 are in Hd
n, then it still follows easily that f (SXS−1) = Sf (X)S−1,
using the fact that u and v respect similarity. Note that in this case (A, B) and
(C, D) are not only similar via S, but also unitarily equivalent via the unitary
matrix in the polar decomposition of S, cf., Remark 2.3. In general, of course, it
will not be the case that C and D are Hermitian.
To prove, under the conditions of Theorem 1.1, that f respects similarity, and
hence is a nc function, we will use Lemma 2.3 of [10]. To apply this lemma, we
need to prove that f has the following two properties:
(i) f is F-differentiable;
(ii) the following identity holds
Df (X)([T, X]) = [T, f (X)], X ∈ Df,n, T ∈ Cn×n, n = 1, 2, . . . .
(5.4)
As before, [S, Q] denotes the commutator of square matrices S, Q of the same size,
applied entrywise in case S and Q are tuples of matrices. In case only one of S and
Q is a tuple, then the other one is identified with a tuple of the same length and
the given matrix in each entry. Note that if S and Q are Hermitian, then [S, Q] is
skew-Hermitian, and hence [iS, Q] = i[S, Q] is Hermitian.
To achieve more than in Proposition 5.1 we require the nc Cauchy-Riemann
equations (1.5) which, for convenience, we recall here: For n = 1, 2, . . .
Du(A, B)(Z1, Z2) = Dv(A, B)(−Z2, Z1), (A, B) ∈ Dn, Z1, Z2 ∈ Hn.
From Proposition 5.1 it is clear what the G-derivative of f should be in case f
(5.5)
is F-differentiable. For X = A + iB ∈ Df,n and Z1 + iZ2 ∈ (Cn×n)d we define
eDf (A + iB)(Z1 + iZ2) := Du(A, B)(Z1, Z2) + iDv(A, B)(Z1, Z2),
provided the G-derivatives of u and v exist in (A, B). As a first step we show that
(5.6)
eDf (X)(Z) is linear in Z.
14
Lemma 5.3. Let u and v be G-differentiable, real nc functions that satisfy the
linear in the directional variable Z.
nc Cauchy-Riemann equations (5.5). Then the map eDf (X)(Z) defined in (5.6) is
Proof. The maps Du and Dv are R-linear in the directional variable. Hence eDf
is additive and R-homogeneous in the directional variable. Write z ∈ C as z = reiθ
with r ≥ 0 and θ ∈ [0, 2π]. Note that
eiθZ = (cos θ + i sin θ)(Z1 + iZ2) = (Z1 cos θ − Z2 sin θ) + i(Z1 sin θ + Z2 cos θ).
Set Z1,θ := Z1 cos θ − Z2 sin θ and Z2,θ := Z1 sin θ + Z2 cos θ. It follows that
eDf (X)(zZ) = Du(A, B)(rZ1,θ, rZ2,θ) + iDv(A, B)(rZ1,θ, rZ2,θ)
= r(Du(A, B)(Z1,θ, Z2,θ) + iDv(A, B)(Z1,θ, Z2,θ)).
(5.7)
Using that G-derivatives Du and Dv are R-linear in the the directional variables
together with the Cauchy-Riemann equations (5.5) yields
Du(A, B)(Z1,θ, Z2,θ) = cos θDu(A, B)(Z1, Z2) + sin θDu(A, B)(−Z2, Z1)
= cos θDu(A, B)(Z1, Z2) − sin θDv(A, B)(Z1, Z2).
Similarly, we have
Dv(A, B)(Z1,θ, Z2,θ) = cos θDv(A, B)(Z1, Z2) + sin θDv(A, B)(−Z2, Z1)
= cos θDv(A, B)(Z1, Z2) + sin θDu(A, B)(Z1, Z2).
Combining these formulas shows
Du(A, B)(Z1,θ, Z2,θ) + iDv(A, B)(Z1,θ, Z2,θ) =
= (cos θ + i sin θ)Du(A, B)(Z1, Z2) + ((cos θ + i sin θ))iDv(A, B)(Z1, Z2)
= eiθ(Du(A, B)(Z1, Z2) + iDv(A, B)(Z1, Z2)).
(5.8)
Together with (5.7) this yields
eDf (X)(zZ) = z(Du(A, B)(Z1, Z2) + iDv(A, B)(Z1, Z2)),
so that eD is C-homogeneous in the directional variable, and hence C-linear.
With linearity out of the way, it is straightforward to prove f is F-differentiable
(cid:3)
in case u and v are F-differentiable.
Lemma 5.4. Let u and v be F-differentiable, real nc functions that satisfy the nc
Cauchy-Riemann equations (5.5). Then f defined by (5.2) is F-differentiable with
Proof. The proof is similar to the last part of the proof of Theorem 4.1. Since u
G-derivative given by Df (X)(Z) = eDf (X)(Z) as in (5.6).
and v are F-differentiable, they are G-differentiable, and thus eDf is C-linear in the
directional variable. To see that f is F-differentiable, note that for X = A + iB ∈
Df,n and Z = Z1 + iZ2, Z1, Z2 ∈ Hd
n, we have
f (X + Z) − f (X) − eDf (X)(Z) =
= (u(A + Z1, B + Z2) − u(A, B) − Du(A, B)(Z1, Z2))+
+ i(v(A + Z1, B + Z2) − v(A, B) − Dv(A, B)(Z1, Z2)).
Using T1 and T2 as in (4.8) the same argument applies, in the opposite direction,
to conclude that F-differentiability of u and v implies F-differentiability of f . (cid:3)
15
Lemma 5.5. Let u and v be F-differentiable, real nc functions that satisfy the nc
Cauchy-Riemann equations (5.5). Define f as in (5.2). Then (5.4) holds.
Proof. Let X = A + iB and T = T1 + iT2. Then
[T, X] = ([iT1, B] + [iT2, A]) + i([iT1,−A] + [iT2, B]).
Set Z1 = [iT1, B] + [iT2, A] and Z2 = [iT1,−A] + [iT2, B]. By Lemma 5.4 we obtain
Df (X)([T, X]) = Du(A, B)(Z1, Z2) + iDv(A, B)(Z1, Z2).
Note that
Du(A, B)(Z1, Z2) = Du(A, B)([iT1, B] + [iT2, A], [iT1,−A] + [iT2, B])
= Du(A, B)([iT2, A], [iT2, B]) + Du(A, B)([iT1, B], [iT1,−A])
= Du(A, B)([iT2, (A, B)]) + Du(A, B)([iT1, (B,−A)]).
Applying the Cauchy-Riemann equations (5.5) to the second summand gives
Du(A, B)(Z1, Z2) = Du(A, B)([iT2, (A, B)]) + Dv(A, B)([iT1, (A, B)]).
Now use that Part (a) Lemma 2.3 of [10] applies to u and v. This yields
Du(A, B)(Z1, Z2) = [iT2, u(A, B)] + [iT1, v(A, B)].
Similarly, for Dv(A, B)(Z1, Z2) we get
Dv(A, B)(Z1, Z2) = Dv(A, B)([iT2, (A, B)]) + Dv(A, B)([iT1, (B,−A)])
= Dv(A, B)([iT2, (A, B)]) + Dv(A, B)([iT1, (−A,−B)])
= Dv(A, B)([iT2, (A, B)]) − Du(A, B)([iT1, (A, B)])
= [iT2, v(A, B)] − [iT1, u(A, B)].
Therefore, we have
Df (X)([T, X]) =
= [iT2, u(A, B)] + [iT1, v(A, B)] + i([iT2, v(A, B)] − [iT1, u(A, B)])
= [iT2, u(A, B)] − i[iT1, u(A, B)] + [iT1, v(A, B)] + i[iT2, v(A, B)]
= [T1 + iT2, u(A, B)] + [T1 + iT2, iv(A, B)]
= [T, u(A, B) + iv(A, B)] = [T, f (X)].
(cid:3)
Proof of Theorem 1.1. The proof of this theorem is now straightforward. The
fact that f is graded and respects direct sums follows from Proposition 5.1. Lemma
5.4 yields the F-differentiability of f. Finally, from Lemma 5.5 we have that (5.4)
holds and combining this with the fact that f is F-differentiable we can apply
Lemma 2.3 of [10] to conclude that f respects similarities. Therefore, f is a F-
differentiable nc function.
(cid:3)
Remark 5.6. As pointed out in [10], even in classical complex analysis, G-differen-
tiability of u and v, i.e., existence of partial derivatives, together with the Cauchy-
Riemann equations is not strong enough to prove analyticity of f . Continuity
of the partial derivatives provides F-differentiability, which is strong enough; this
corresponds to the approach taken in the present paper. The Looman-Menchoff
theorem, cf., [13, Page 199], states that continuity of f , and hence of u and v, is
also sufficient. This in turn implies that u and v were F-differentiable from the
start. As the proof of the Looman-Menchoff theorem requires the Baire category
16
theorem and Lebesgue integration, it is not clear whether a similar relaxation of
Theorem 1.1 can be achieved in the context considered here.
In particular, the
theory of integration of nc functions does not appear to be well developed so far.
We are just aware of the paper [12] on the nc Hardy space over the unitary matrices.
Acknowledgments. This work is based on research supported in part by the Na-
tional Research Foundation of South Africa (NRF) and the DST-NRF Centre of
Excellence in Mathematical and Statistical Sciences (CoE-MaSS). Any opinion,
finding and conclusion or recommendation expressed in this material is that of the
authors and the NRF and CoE-MaSS do not accept any liability in this regard.
References
[1] K. Atkinson and W. Han, Theoretical numerical analysis. A functional analysis framework,
Texts in Applied Mathematics 39, Springer-Verlag, New York, 2001.
[2] S.A.R. Disney, J.D. Gray, and S.A. Morris, Is a function that satisfies the Cauchy-Riemann
equations necessarily analytic? Austral. Math. Soc. Gaz. 2 (1975), 6781.
[3] T. Jiang and H. Sendov, On differentiability of a class of orthogonally invariant functions on
several operator variables, Oper. Matrices 12 (2018), 711-721.
[4] J.D. Gray and S.A. Morris, When is a function that satisfies the Cauchy-Riemann equations
analytic? Amer. Math. Monthly 85 (1978), 246256.
[5] E. Hille and R.S. Phillips, Functional analysis and semi-groups, American Mathematical
Society Colloquium Publications 31, American Mathematical Society, Providence, R.I., 1957.
[6] R.A. Horn and C.R. Johnson, Matrix analysis. Second edition, Cambridge University Press,
Cambridge, 2013.
[7] D.S. Kaliuzhnyi-Verbovetskyi and V. Vinnikov, Foundations of free noncommutative func-
tion theory, Mathematical Surveys and Monographs 199, American Mathematical Society,
Providence, RI, 2014.
[8] M. P´alfia, Lowner's Theorem in several variables, preprint, arXiv:1405.5076.
[9] J.E. Pascoe, The noncommutative Lwner theorem for matrix monotone functions over oper-
ator systems, Linear Algebra Appl. 541 (2018), 54-59.
[10] J.E. Pascoe and R. Tully-Doyle, Free Pick functions: representations, asymptotic behavior
and matrix monotonicity in several noncommuting variables, J. Funct. Anal. 273 (2017),
283-328.
[11] J.-P. Penot, Calculus without derivatives, Graduate Texts in Mathematics 266, Springer,
New York, 2013.
[12] M. Popa and V. Vinnikov, H 2 spaces of non-commutative functions, Complex Anal. Oper.
Theory 12 (2018), 945-967.
[13] S. Saks, Theory of the integral. Second revised edition, Dover Publications, Inc., New York,
1964.
[14] J.L. Taylor, Functions of several noncommuting variables, Bull. Amer. Math. Soc. 79 (1973),
1-34.
[15] M.A. Zorn, Derivatives and Frchet differentials, Bull. Amer. Math. Soc. 52 (1946), 133137.
S. ter Horst, Department of Mathematics, Unit for BMI, North-West University,
Potchefstroom, 2531 South Africa, and DST-NRF Centre of Excellence in Mathemat-
ical and Statistical Sciences (CoE-MaSS)
E-mail address: [email protected]
E.M. Klem, Department of Mathematics, Unit for BMI, North-West University,
Potchefstroom, 2531 South Africa
E-mail address: [email protected]
17
|
1306.0893 | 1 | 1306 | 2013-06-03T20:01:08 | Powers of distances to lower dimensional sets as Muckenhoupt weights | [
"math.FA"
] | Let $(X,d,\mu)$ be an Ahlfors metric measure space. We give sufficient conditions on a closed set $F\subseteq X$ and on a real number $\beta$ in such a way that $d(x,F)^\beta$ becomes a Muckenhoupt weight. We give also some illustrations to regularity of solutions of partial differential equations and regarding some classical fractals. | math.FA | math |
POWERS OF DISTANCES TO LOWER DIMENSIONAL
SETS AS MUCKENHOUPT WEIGHTS
HUGO AIMAR, MARILINA CARENA, RICARDO DUR ´AN, AND MARISA TOSCHI
Abstract. Let (X, d, µ) be an Ahlfors metric measure space. We give
sufficient conditions on a closed set F ⊆ X and on a real number β
in such a way that d(x, F )β becomes a Muckenhoupt weight. We give
also some illustrations to regularity of solutions of partial differential
equations and regarding some classical fractals.
1. Introduction
Under different conditions for a domain Ω, weighted norm estimates for
solutions of linear and nonlinear equations have been studied by several
authors (see for example [DS04], [DST08], [DST10], [Sou04]), where the
weight of interest is a power of the distance to the boundary ∂Ω.
The class of Muckenhoupt weights Ap(Rn) is a fundamental tool in real
and harmonic analysis. The first non-trivial examples of weights in Ap(Rn)
are the weights xβ for −n < β < n(p − 1). The results in [DST08] show
that in a domain whose boundary has dimension n − 1, if the domain is
smooth enough then dβ(x, ∂Ω) ∈ Ap(Rn) for −1 < β < (p − 1). In [DLG10]
this result has been generalized to some s-dimensional compact sets F in
Rn with 0 ≤ s < n. They proved that dβ(x, F ) ∈ Ap(Rn) for −(n − s) <
β < (n − s)(p − 1).
Therefore thinking the domains on Rn as complements of closed sets, it
seems natural to try to consider the somehow heterogeneous situation in
which the boundary shows different dimensions at different points. With
that generality the problem looks hard. But with some extra hypotheses on
the structure of ∂Ω some extensions are possible. The techniques used here
extend naturally to general metric measure space (X, d, µ) satisfying the so
called Ahlfors condition, which is a particular case of space of homogeneous
type.
The paper is organized as follows. In Section 2 we give some definitions
and notation on metric spaces and the Hardy-Littlewood maximal function
of measures obtaining the finiteness of such function for certain measures.
The main results of this note are contained in Section 3. There we prove
that powers of the Hardy-Littlewood maximal function of measures belong
to the Muckenhoupt class and in some cases we can describe the behavior
of such function giving a family of weights on metric measure spaces. In
Section 4 we give some applications. First we obtain weighted bounded
2010 Mathematics Subject Classification. Primary 28A25; Secondary 28A78.
Key words and phrases. Ahlfors spaces, Hardy-Littlewood maximal operator, Mucken-
houpt weights, Hausdorff measure.
1
2
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
estimates for gradients of solutions of polyharmonic equations and then we
produce families of weights on some classical fractals.
2. Hardy-Littlewood maximal functions of measures
Let X be a set. A quasi-distance on X is a non-negative symmetric
function d defined on X × X such that d(x, y) = 0 if and only if x = y, and
there exists a constant K ≥ 1 such that the inequality
d(x, y) ≤ K(d(x, z) + d(z, y))
holds for every x, y, z ∈ X. We will refer to K as the triangle constant for
d. A quasi-distance d on X induces a topology through the neighborhood
system given by the family of all subsets of X containing a d-ball B(x, r) =
{y ∈ X : d(x, y) < r}, r > 0 (see [CW71]). In a quasi-metric space (X, d)
the diameter of a subset E is defined as
diam(E) = sup{d(x, y) : x, y ∈ E},
and the distance between a point x ∈ X and a set E is defined by d(x, E) =
inf{d(x, y) : y ∈ E}.
Throughout this paper (X, d) shall be a quasi-metric space such that the
d-balls are open sets.
We shall say that (X, d, µ) is a space of homogeneous type if µ is a
non-negative Borel measure µ satisfying the doubling condition
0 < µ(B(x, 2r)) ≤ Aµ(B(x, r)) < ∞
for some constant A, for every x ∈ X and every r > 0.
If we also have
µ({x}) = 0 for every x ∈ X then we say that (X, d, µ) is a non-atomic
space of homogeneous type.
We shall now recall a basic property of spaces of homogeneous type that
we shall need. This property is actually contained in [CW71], and reflects
the fact that spaces of homogeneous type have finite metric (or Assouad)
dimension (see [Ass79]). The expression finite metric dimension means
that there exists a constant N ∈ N such that no ball of radius 2r contains
more than N points of any r-disperse subset of X. A set U is said to be
r-disperse if d(x, y) ≥ r for every x, y ∈ U , x 6= y. An r-net is a maximal
r-disperse set. It is easy to check that U is an r-net in X if and only if U
is an r-disperse and r-dense set in X, where r-dense means that for every
x ∈ X there exists u ∈ U with d(x, u) < r. It is well known that if a quasi-
metric space (X, d) has finite metric dimension, then every bounded subset
F of X is totally bounded, so that for every r > 0 there exists a finite r-net
on F , whose cardinal depends on diam(F ) and on r.
On the other hand, every compact quasi-metric space with finite metric
dimension carries a nontrivial doubling measure (see [Wu98] or [VK87]).
Let (X, d, µ) be a space of homogeneous type. For a given locally in-
tegrable function f , the Hardy-Littlewood maximal operator is given
by
Mµf (x) = sup
1
µ(B)ZB
f dµ,
POWERS OF DISTANCES AS MUCKENHOUPT WEIGHTS
3
where the supremum is taken over the family of the d-balls B containing x.
Since µ is doubling then kMµf kLp(dµ) ≤ Cpkf kLp(dµ) for 1 < p < ∞ and
µ ({Mµf > λ}) ≤ C
λ kf kL1(dµ) (see [CW71]).
The definition of the Hardy-Littlewood maximal operator can be extended
to a non-negative Borel measure ν such that every ball has finite ν-measure
by
Mµν(x) = sup
ν(B)
µ(B)
,
where the supremum is taken over the family of the d-balls B containing x.
Since µ is doubling, Mµν(x) is equivalent to its centered version, i.e.
Mµν(x) = sup
r>0
ν(B(x, r))
µ(B(x, r))
.
When ν is a finite measure the analogous of the weak type inequality
given above shows that Mµν is finite µ-almost everywhere. We aim to give
sufficient conditions other than the finiteness of ν in order to have the µ-
almost everywhere finiteness of Mµν.
We shall deal with spaces of finite Hausdorff dimension. Some comments
regarding the terminology are in order.
In the bibliography belonging to
geometric measure theory, such as [Fal86], an s-set E is one for which 0 <
H s(E) < ∞ where H s is the Hausdorff measure of dimension s. In some
references related to problems of harmonic analysis and partial differential
equations, see for example [Sjo97], the expression s-set is used to name
a set that supports a measure ν for which ν(B(x, r)) behaves as rs for r
small. This condition implies the above one. On the other hand when
dealing with operators such as the Hardy-Littlewood maximal, the global
behavior of the given set, aside its local behavior, becomes relevant. In this
direction, again, two different terminologies appear in the literature. When
there exists a measure ν such that ν(B(x, r)) behaves as rs for r up to the
diameter of the support of ν, which could be unbounded, the space is said
to be s-Ahlfors. See for example [Sjo97] and [Gro07]. In [MS79] for a given
space of homogeneous type (X, d, µ), a quasi-distance δ can be constructed
on X satisfying that the measure of a δ-ball centered at x with radius r,
behaves as r when µ({x}) = 0 if r is less than µ(X). This property is named
normality.
For the sake of simplicity we shall adopt along this note the following
definitions.
Assume that (X, d) is a fixed quasi-metric space. A subspace (Y, d) of
(X, d) is said to be α-Ahlfors with measure µ if µ is a Borel measure
supported in Y such that there exists a constant c ≥ 1 satisfying the in-
equalities
(2.1)
c−1rα ≤ µ(B(x, r)) ≤ crα,
for every x ∈ Y and every 0 < r < diam(Y ).
It easy to see that each
α-Ahlfors space with measure µ is a non-atomic space of homogeneous type
with doubling constant for µ which only depends on c and α. We will refer to
the triangle constant K and the constants c and α in (2.1) as the geometric
constants of the space. When (2.1) holds only for 0 < r < r0 for some
positive r0, we say that (Y, d) is locally α-Ahlfors with measure µ. The
4
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
constant c in (2.1) is said to be a constant for the Ahlfors condition
of µ.
The next proposition shows that both concepts coincide when the space
is bounded.
Proposition 1. Let F be a closed and bounded subset of a quasi-metric
space (X, d). If (F, d) is locally s-Ahlfors with measure ν, then we have that
(F, d) is s-Ahlfors with measure ν.
Proof. If diam(F ) ≤ r0 the result is trivial, so that we shall assume that
diam(F ) > r0. Take x ∈ F . If 0 < r < r0 we use that (F, d) is locally s-
Ahlfors with measure ν, so that only we need to consider the case r0 ≤ r <
diam(F ). Being F a bounded set, since (X, d) has finite metric dimension,
there exists a finite r0
2 -net U in F , let us say U = {x1, . . . , xI }, where I
depends only on diam(F ) and r0. Then
ν(B(x, r)) ≤
I
Xi=1
ν(cid:16)B(cid:16)xi,
r0
2 (cid:17)(cid:17) ≤ Ic2−srs
0 < Icrs.
(cid:3)
The main result of this section, which we shall use in order to apply
Theorem 3 to build Muckenhoupt weights in the next section, is contained
in the following proposition.
Proposition 2. Let (X, d) be α-Ahlfors with measure µ and F closed in X.
If (F, d) is s-Ahlfors with measure ν, where 0 ≤ s < α, then Mµν(x) < ∞
for µ-almost every x ∈ X.
Proof. If x /∈ F and r ≤ d(x, F ), then B(x, r) ∩ F = ∅ and hence we have
ν(B(x, r)) = 0. So that, in this case, we can assume r > d(x, F ) > 0. Fix
y ∈ F such that d(x, y) < 3
2 d(x, F ). Then B(x, r) ⊆ B(y, 3Kr) and we
obtain
ν(B(x, r))
µ(B(x, r))
≤
ν(B(y, 3Kr))
µ(B(x, r))
≤ cc(3K)srs−α < Cd(x, F )s−α,
with C = cc(3K)s, where K denotes the triangle constant for d, and c and
c are constants for the Ahlfors condition of µ and ν respectively. In other
words, C depends only on the geometric constants of (X, d) and (F, d).
Hence Mµν(x) < ∞ for x /∈ F .
ber n let Fn = F ∩ B(x0, n). Then µ(F ) ≤P∞
So we have to prove that µ(F ) = 0. Fix x0 ∈ F and for each natural num-
n=1 µ(Fn), so that it is enough
to prove that µ(Fn) = 0 for every n. To see this, fix n ∈ N and 0 < ρ < n.
Since (X, d) has finite metric dimension and Fn is bounded, there exists a
finite ρ-net U in Fn, let us say U = {x1, x2, . . . , xIρ}. Hence {B(xi, ρ) : i =
i=1 µ(B(xi, ρ)) ≤ cραIρ. To
1, . . . , Iρ} is a cover of Fn, so that µ(Fn) ≤ PIρ
2K(cid:1) ∩ B(cid:0)xj, ρ
estimate Iρ, notice that B(cid:0)xi, ρ
2K(cid:1) = ∅ for i 6= j. Then
2K(cid:17)(cid:17) = ν
B(cid:16)xi,
[i=1
c(2K)s ≤
2K(cid:17)
ν(cid:16)B(cid:16)xi,
≤ ν(B(x0, 2Kn)).
Iρ
Xi=1
ρs
Iρ
ρ
Iρ
ρ
POWERS OF DISTANCES AS MUCKENHOUPT WEIGHTS
5
So that Iρ ≤ ν(B(x0, 2Kn))c(2K)sρ−s and we have that
µ(Fn) ≤ ccν(B(x0, 2Kn))(2K)α+sρα−s.
Taking ρ → 0 we obtain µ(Fn) = 0.
(cid:3)
Notice that if x ∈ F , then we have that
ν(B(x, r))
µ(B(x, r))
≥
1
cc
rs−α,
and taking supremum over r > 0 we obtain Mµν(x) = ∞.
3. Main results
The theory developed by Muckenhoupt in [Muc72] provides necessary and
sufficient conditions on a weight w defined on (Rn, · , λ) in order to obtain
weighted estimates for the maximal operator. These functions w are known
as Ap-Muckenhoupt weights.
Let (X, d, µ) be a quasi-metric measure space such that every d-ball has
positive and finite µ-measure. A weight w on X is a locally integrable non-
negative function defined on X. By locally we mean integrable over balls,
i.e. RB w dµ < ∞ for every d-ball B in X.
For 1 < p < ∞ the Muckenhoupt class Ap(X, d, µ) is defined as the
set of all weights w defined on X for which there exists a constant C such
that the inequality
(cid:18) 1
µ(B)ZB
w dµ(cid:19)(cid:18) 1
µ(B)ZB
w− 1
p−1 dµ(cid:19)p−1
≤ C
holds for every d-ball B in X. For p = 1, we say that w ∈ A1(X, d, µ) if
there exists a constant C such that
1
µ(B)ZB
w dµ ≤ C w(x)
holds for every d-ball B in X and µ-almost every x ∈ B. Set A∞(X, d, µ) =
Sp≥1 Ap(X, d, µ).
It is a well known result in the theory of Muckenhoupt weights that if
w ∈ Ap(X, d, µ), then wdµ is doubling on X provided that (X, d, µ) is a
space of homogeneous type. The classical reference for the basic theory of
Muckenhoupt weights is Chapter IV in the book [GCRdF85]. A celebrated
result of the theory proved in Rn by P. Jones in [Jon80], which extends to
space of homogeneous type, is the factorization theorem: every Ap weight w
can be written as w = w0 w1−p
1 with w0 and w1 in A1. This is an important
result of the Muckenhoupt weights and it is known as factorization property.
The basic general result which shall be useful for the construction of
Muckenhoupt weights as powers of the distance to some particular sets in
X is contained in the following statement.
Theorem 3. Let (X, d, µ) be a space of homogeneous type. Let ν be a
Borel measure such that Mµν(x) < ∞ for µ-almost every x ∈ X. Then
(Mµν)γ ∈ A1(X, d, µ) for every 0 ≤ γ < 1.
6
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
The proof of Theorem 3 is based in Kolmogorov inequality. We will follows
the lines that in Rn for ν absolutely continuous are given in [Duo01] and for
general ν in Rn in [GCRdF85]. However, non trivial technical modification
are needed to extend the result to a general space of homogeneous type.
Proof of Theorem 3. We have to prove that there exists a constant C such
that the inequality
(3.1)
1
µ(B)ZB
(Mµν)γ dµ ≤ C Mµν(x)γ
holds for every d-ball B in X and µ-almost every x ∈ B. Let us fix a d-
ball B = B(x0, r0) and write ν = ν1 + ν2, where ν1 is the restriction to
B(x0, 2Kr0) of ν and ν2 = ν − ν1. Then for 0 ≤ γ < 1 we have
(Mµν)γ ≤ (Mµν1)γ + (Mµν2)γ,
and it is enough to prove (3.1) with ν1 and ν2 on the left hand side.
Since Mµ is of weak type (1, 1) and ν1 is finite, we can apply the Kol-
mogorov's inequality on (X, d, µ), namely
(Mµν1)γ dµ ≤ µ(E)1−γ ν1(E)γ = µ(E)1−γν(E ∩ B(x0, 2Kr0))γ
ZE
for every measurable subset E of X with finite measure, see for example
[dG81], and obtain a constant C depending on γ such that
1
µ(B)ZB
µ(B)(cid:19)γ
(Mµν1)γ dµ ≤ C(cid:18) ν(B)
≤ CMµν(x)γ.
In order to analyze Mµν2, note that if y ∈ B and B1 = B(x1, r1) is
In fact, for
any d-ball containing y with ν2(B1) > 0, then r1 > 1
z ∈ B1 ∩ (X \ B(x0, 2Kr0)) we have
2K r0.
2Kr0 ≤ d(z, x0) ≤ K(d(z, y) + d(y, x0)) < K(2Kr1 + r0).
This implies that B ⊆ B(x1, 5K 3r1). To see this, notice that for x ∈ B we
have
d(x, x1) ≤ K 2(d(x, x0) + d(x0, y) + d(y, x1)) < K 2(2r0 + r1),
so that x ∈ B(x1, 5K 3r1). Therefore
ν2(B1)
µ(B1)
≤
µ(B(x1, 5K 3r1))
µ(B1)
Mµν(x).
Since B1 is an arbitrary d-ball containing y,
Mµν2(y) ≤
µ(B(x1, 5K 3r1))
µ(B1)
Mµν(x).
Integrating with respect to y we obtain
1
µ(B)ZB
(Mµν2)γ dµ ≤ CMµν(x)γ
and the theorem is proved.
(cid:3)
POWERS OF DISTANCES AS MUCKENHOUPT WEIGHTS
7
The first elementary use of Theorem 3 to produce weights in the Ap(X, d, µ)
class is obtained taking ν = δ0 the Dirac delta measure in x0. In this case
we have
Mµδ0(x) = sup
r>0
δ0(B(x, r))
µ(B(x, r))
=
1
µ(B(x, d(x, x0))
,
where B(x, s) = {y ∈ X : d(x, y) ≤ s}. Since for the doubling condition
we have µ(B(x, s)) ≤ µ(B(x, s)) ≤ Aµ(B(x, s)), applying Theorem 3 we
obtain that µ(B(x, d(x, x0)))−γ ∈ A1(X, d, µ) for 0 ≤ γ < 1. Consequently
µ(B(x, d(x, x0)))β ∈ Ap(X, d, µ) for −1 < β < p − 1.
As an immediate consequence of Proposition 2 and Theorem 3, we have
the following result.
Corollary 4. Let (X, d) be α-Ahlfors with measure µ and F closed in X.
If (F, d) is s-Ahlfors with measure ν, where 0 ≤ s < α, then (Mµν)γ ∈
i=1 Fi and
i=1 νi, where {F1, . . . , FH } is a family of pairwise disjoint closed and
bounded subsets of X such that (Fi, d) is si-Ahlfors with measure νi, where
0 ≤ si < α for i = 1, 2, . . . , H. Moreover the weight max{M γ
µ ν(x), C} ∈
A1(X, d, µ) for every C > 0.
A1(X, d, µ) for every 0 ≤ γ < 1. The same is true when F = SH
ν =PH
In the following result we explore the behavior of Mµν to obtain an explicit
family of Muckenhoupt weights. Here we use the notation f (x) ≃ g(x) to
indicate that there exist positive and finite constants k1 and k2 such that
k1g(x) ≤ f (x) ≤ k2g(x) for every x.
Theorem 5. Let (X, d) be α-Ahlfors with measure µ. Let {F1, . . . , FH } be a
family of pairwise disjoint closed and bounded subsets of X such that (Fi, d)
is locally si-Ahlfors with measure νi, where 0 ≤ si < α for i = 1, 2, . . . , H.
i=1 νi. Then there exist open sets U1, . . . , UH
Set F = SH
i=1 Fi and ν = PH
pairwise disjoint with Ui containing Fi such that
(i) if d(x, F ) < 2Kdiam(F ) then we have
(a) Mµν(x) ≃ d(x, Fi)si−α, for every x ∈ Ui \ Fi, for every i =
(ii) if d(x, F ) ≥ 2Kdiam(F ) then Mµν(x) ≃ d(x, F )−α ≃ d(x, x0)−α
i=1 Ui then Mµν(x) ≃ 1;
1, 2, . . . , H;
(b) if x /∈SH
for every x0 ∈ F .
Proof. Let us start by proving (ii). Let us fix x such that d(x, F ) ≥
2Kdiam(F ). Since F is bounded, we only need to prove the first equi-
valence in (ii). Since for r ≤ d(x, F ) we have B(x, r) ∩ F = ∅, in order to
estimate Mµν(x) we only have to consider balls centered at x with radius
r > d(x, F ). Hence
ν(B(x, r))
µ(B(x, r))
≤
cν(F )
d(x, F )α ,
where c denotes a constant for the Ahlfors condition of µ.
On the other hand, taking r = 3Kd(x, F ) we obtain
Mµν(x) ≥
ν(B(x, 3Kd(x, F )))
c(3Kd(x, F ))α
.
8
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
Now fix y ∈ F such that d(x, y1) < 3
2 d(x, F ). Then
B(y, 2Kdiam(F )) ⊆ B(y, d(x, F )) ⊆ B(x, 3Kd(x, F )),
and therefore
Mµν(x) ≥
ν(B(y, 2Kdiam(F )))
c(3Kd(x, F ))α
=
ν(F )
c(3Kd(x, F ))α .
To prove (i), let ci be the constant for the Ahlfors condition of νi.
In
other words, ci is a constant satisfying
c−1
i rsi ≤ νi(B(y, r)) ≤ cirsi,
for every 0 < r ≤ ri and every y ∈ Fi, for some ri > 0. Let us define
∆ = min{d(Fi, Fj ) : i 6= j} and
Ui = {x ∈ X : d(x, Fi) < κi} ,
where κi ≤ min{2ri, ∆/2} is a constant that we shall define later. Notice
that Ui ∩ Uj = ∅ for i 6= j.
In order to show (ia), let us fix i ∈ {1, . . . , H} and x ∈ Ui \ Fi. To obtain
the lower bound, let y ∈ Fi such that d(x, y) < 3
2 d(x, Fi). Then
ν (B (x, 2Kd(x, Fi)))
µ (B (x, 2Kd(x, Fi)))
νi (B (x, 2Kd(x, Fi)))
µ (B (x, 2Kd(x, Fi)))
νi(cid:0)B(cid:0)y, 1
2 d(x, Fi)(cid:1)(cid:1)
c(2Kd(x, Fi))α
Mµν(x) ≥
≥
≥
≥ Cid(x, Fi)si−α,
where Ci =
1
cci2si (2K)α .
To obtain the upper bound, let us keep consider the case x ∈ Ui \ Fi.
Hence if r ≤ d(x, Fi) we have that r < ∆/2, so that B(x, r) ∩ F = ∅
and ν(B(x, r)) = 0. Assume then r > d(x, Fi) and fix y ∈ Fi such that
d(x, y) < r. Then
ν(B(x, r))
µ(B(x, r))
≤
ν(B(y, 2Kr))
µ(B(x, r))
≤
cν(B(y, 2Kr))
rα
.
We shall consider two cases taking into account the size of r.
Case 1:
B(y, 2Kr) ∩ Fj = ∅ for every j 6= i. Then
d(x, Fi) < r < ∆
2K .
In this case we have 2Kr < ∆, so that
ν(B(y, 2Kr))
rα
=
νi(B(y, 2Kr))
rα
Case 2:
r ≥ ∆
2K . In this case
≤ ci(2Kr)sir−α < ci(2K)sid(x, Fi)si−α.
ν(B(x, r))
µ(B(x, r))
≤
cν(F )
∆ (cid:19)α
rα ≤ c(cid:18) 2K
ν(F ).
POWERS OF DISTANCES AS MUCKENHOUPT WEIGHTS
9
To obtain the result we only need to take x such that ci(2K)sid(x, Fi)si−α ≥
∆ (cid:1)α
c(cid:0) 2K
ν(F ). In other words, it is enough to take
κi = min(2ri,
∆
2
,(cid:18) c2αK αν(F )
ci∆α2siK si(cid:19)
1
si −α)
in the definition of Ui.
Finally, to prove (ib) take x /∈ SH
i=1 Ui such that d(x, F ) < 2Kdiam(F ).
Then d(x, Fi) ≥ κi for each i, so that d(x, F ) ≥ min κi =: κ. As before, in
order to estimate Mµν(x) we only need to consider r > d(x, F ) ≥ κ. In this
case
ν(B(x, r))
µ(B(x, r))
≤
cν(F )
κα .
On the other hand, since d(x, F ) < 2Kdiam(F ) we have F ⊆ B(x, 3K 2diam(F ))
and therefore
Mµν(x) ≥
ν(B(x, 3K 2diam(F )))
µ(B(x, 3K 2diam(F )))
≥
ν(F )
c(3K 2diam(F ))α .
(cid:3)
As a consequence of Corollary 4 and Theorem 5, we have the following
result.
Theorem 6. Let (X, d) be α-Ahlfors with measure µ. Let {F1, . . . , FH }
be a family of pairwise disjoint closed and bounded subsets of X such that
(Fi, d) is a locally si-Ahlfors space with measure νi, where 0 ≤ si < α for
i = 1, 2, . . . , H. Then
(i)
(ii)
w(x) =( d(x, Fi)(si−α)γ,
1,
for x ∈ Ui;
for x ∈(cid:16)SH
i=1 Ui(cid:17)c
belong to A1(X, d, µ) for every 0 ≤ γ < 1, where Ui is the open set
containing Fi given by Theorem 5;
v(x) =( d(x, Fi)βi,
1,
for x ∈ Ui;
for x ∈(cid:16)SH
i=1 Ui(cid:17)c
belongs to Ap(X, d, µ) for every −(α − si) < βi < (α − si)(p − 1).
Proof. Since the equivalence w1 ≃ w2 preserves A1(X, d, µ), in order to prove
(i), from Corollary 4 we only need to check that w(x) ≃ max{M γ
µ ν(x), C} =:
u(x) for some positive constant C.
As we saw in item (ia) in Theorem 5, for each i = 1, ..., H, there exists a
constant C such that if x ∈ Ui \ Fi then Mµν(x) ≥ Cd(x, Fi)si−α and, since
d(x, Fi) ≤ 2ri, taking s∗ = max{s1, . . . , sH} and r∗ = max{r1, . . . , rH }, we
have that M γ
µ ν(x) ≥ C γ(2r∗)γ(s∗−α).
Then, if we will consider C := C γ(2r∗)γ(s∗−α) in the definition of u we
obtain that for x ∈SH
i=1 Ui \ Fi we have
u(x) = M γ
µ ν(x) ≃ d(x, Fi)(si−α)γ = w(x).
10
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
For the case of x /∈ SH
consider the following possibilities for x:
i=1 Ui, w(x) = 1, and to see that u(x) ≃ 1 we shall
• d(x, F ) < 2Kdiam(F ). By (ii) of Theorem 5, M γ
µ ν(x) ≃ 1 and we
have that u(x) ≃ w(x), not depending on the value of u(x).
• 2Kdiam(F ) < d(x, F ) ≤ Λ, with
Λ = max(2Kdiam(F ),(cid:18) cν(F )
C 1/γ (cid:19)1/α)
and c a constant for the Ahlfors condition of µ. Item (ii) of Theo-
rem 5 together with the bounds for d(x, F ) implies that M γ
µ ν(x) ≃
d(x, F )−αγ ≃ 1. Then, as in the previous case, u(x) ≃ 1.
• d(x, F ) > Λ.
In the proof of (ii) in Theorem 5 we can see that
Mµν(x) ≤ cν(F )
d(x,F )α . Then
M γ
Λα (cid:19)γ
µ ν(x) ≤(cid:18) cν(F )
Hence u(x) = C ≃ 1.
≤ C.
Item (ii) is an immediate consequence of item (i) and the definition of
(cid:3)
Muckenhoupt class.
In the particular case of an s-Ahlfors space F , we can take out the re-
striction of boundedness and we obtain the following result.
Theorem 7. Let (X, d) be α-Ahlfors with measure µ and F closed in X. If
(F, d) is s-Ahlfors with measure ν, where 0 ≤ s < α, then d(x, F )γ(s−α) ∈
A1(X, d, µ) for every 0 ≤ γ < 1. Consequently d(x, F )β ∈ Ap(X, d, µ) for
−(α − s) < β < (α − s)(p − 1).
Proof. The results follows if we prove that Mµν(x) ≃ d(x, F )s−α for µ-
almost every x ∈ X.
Take x /∈ F . The upper bound was proved in Proposition 2. To obtain
the lower bound, let y ∈ F be such that d(x, y) < 3
2 d(x, F ). Then
Mµν(x) ≥
ν (B (x, 2Kd(x, F )))
µ (B (x, 2Kd(x, F )))
≥
ν(cid:0)B(cid:0)y, 1
c(2Kd(x, F ))α ≥ Cd(x, F )s−α,
2 d(x, F )(cid:1)(cid:1)
c(2K)α2s , with c and c constants for the Ahlfors condition of µ
where C =
and ν respectively.
c
For x ∈ F , as we proved in Proposition 2, Mµν(x) = ∞ but µ(F ) = 0. (cid:3)
4. Applications
This section contain three topics. The first one deal with weighted Sobolev
regularity for solutions of elliptic partial differential equations of large order.
Second we give a criteria to decide wether or not a given set is locally
s-Ahlfors in terms of the s-dimensional Hausdorff measure.
Third, we provide some non trivial Ap weights on the Sierpinski gasket
built from Theorem 7.
Weighted Sobolev regularity. Let (X, d) be α-Ahlfors with measure µ
and Ω subset of X with diam(Ω) < ∞.
POWERS OF DISTANCES AS MUCKENHOUPT WEIGHTS
11
Given an operator R : Ω × Ω −→ R and f : Ω −→ R we define
u(x) :=ZΩ
R(x, y) f (y) dµ(y).
If there exists a positive constant C such that
(4.1)
R(x, y) ≤ Cd(x, y)−α+1,
we can see that u satisfies
u(x) ≤ C Mµf (x),
where Mµf (x) is the Hardy-Littlewood maximal operator in (X, d, µ). In
fact, if we denote δ the diameter of Ω and using (4.1) we have
u(x) ≤ CZd(x,y)≤δ
f (y)
d(x, y)α−1 dµ(y)
= C
∞
Xk=0Z{2−(k+1)δ≤d(x,y)≤2−k δ}
f (y)
d(x, y)α−1 dµ(y)
∞
Xk=0
2−k
µ(B(x, 2−kδ))Z{d(x,y)≤2−k δ}
≤ C 2α−1
≤ C Mµf (x).
f (y)dµ(y)
Since Mµf (x) is a bounded operator in Lp(Ω, wdµ) if w ∈ Ap(X, d, µ),
follows immediately that
(4.2)
kukLp(Ω,wdµ) ≤ C kf kLp(Ω,wdµ)
i.e. there exists a positive constant C such that
(cid:18)ZΩ
up w dµ(cid:19)1/p
≤ C (cid:18)ZΩ
f p w dµ(cid:19)1/p
,
provided that w ∈ Ap(X, d, µ).
As in general we are interested in sources f as singular as possible the
above inequality allows unbounded growth of f close to some subset of Ω if
the weight w vanishes there. This is precisely the case for weights produced
by Theorem 6.
In the particular case of the n-dimensional Euclidean space (Rn, · , λ),
taking R as the Green function associated with the polyharmonic Dirichlet
problem in Ω
(cid:26) (−∆)mu = f
∂ν(cid:1)j
(cid:0) ∂
u = 0
in Ω
on ∂Ω 0 ≤ j ≤ m − 1,
∂ν is the normal derivative, it is known that the solution of (4.3) is
(4.3)
where ∂
given by
(4.4)
u(x) =ZΩ
R(x, y) f (y) dy.
Then, to prove some weighted Sobolev a priori estimates for the solution
of this problem we need estimates for the Green function R(x, y) and its
derivatives.
12
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
For Ω a regular domain in Rn, some of this estimates was given in [DST10],
where the authors proved that
(4.5)
kukW 2m,p(Ω,wdx) ≤ C kf kLp(Ω,wdx)
Let us remember that for η a multi-index, η = (η1, η2, . . . , ηn) ∈ ZZn
j=1 ηj, Dη = ∂η1
x1...∂ηn
+ we
xn . The Sobolev spaces are
for w ∈ Ap.
denote as usual η = Pn
defined by
W k,p
w (Ω) = {v ∈ Lp(Ω, wdx) : Dηv ∈ Lp(Ω, wdx) ∀ η ≤ k}
and the norm of v ∈ W k,p
w (Ω) is given by
kvkW k,p
w (Ω) = Xη≤k
kDηvkLp(Ω,wdx) .
For Ω a general bounded open set in Rn there exists a constant C such
that the Green function satisfies the estimate
DηR(x, y) ≤ Cx − y2m−n−η,
(4.6)
for n ∈ [3, 2m + 1] ∩ N odd and 0 ≤ η ≤ m − n
even and 0 ≤ η ≤ m − n
2 (see [MM09]).
2 + 1
Then, since η ≤ 2m − 1, it follows that
2 and for n ∈ [2, 2m] ∩ N
(4.7)
kukW k,p
w (Ω) ≤ C kf kLp(Ω,wdx)
for values of n, m and k = η given above, provided that w ∈ Ap(Rn).
Hausdorff measure based criteria for the local s-Ahlfors condition.
In the hypothesis of the results obtained in the previous section we require
that the spaces are locally s-Ahlfors. In order to check that a given (F, d)
is locally s-Ahlfors, we should be able to find a Borel measure ν supported
on F and a real number r0 > 0 such that (2.1) holds for every x ∈ F and
every 0 < r < r0. This does not seem to be an easy task. However, if (F, d)
is locally s-Ahlfors, then there exists essentially only one Borel measure ν
satisfying the condition required in the definition. This fact is known in the
Euclidean setting (see for instance [Tri11]), and for the sake of completeness
let us extend it to general metric measure spaces in Lemma 8. First we shall
need recall some definitions. The basic aspects of Hausdorff measure and
dimension can be found in [Fal86].
For ρ > 0, we say that a sequence {Ei} of subsets of X is a ρ-cover of
a set E if E ⊆ S Ei and diam(Ei) ≤ ρ for every i. Let E ⊆ X and s ≥ 0
fixed. We define
H s
ρ (E) = inf( ∞
Xi=1
diams(Ei) : {Ei} is a ρ-cover of E) .
Clearly H s
to 0 exists (although it may be infinite). Then we define
ρ (E) increases when ρ decreces, so that the limit when ρ tends
H s(E) = lim
ρ→0
H s
ρ (E) = sup
ρ>0
H s
ρ (E).
We shall refer to H s(E) as the s-dimensional Hausdorff measure of E.
POWERS OF DISTANCES AS MUCKENHOUPT WEIGHTS
13
If in the above definition we replace the arbitrary ρ-cover by coverings by
d-balls centering in the set E and with diameter less than ρ, we obtain the
s-dimensional spherical Hausdorff measure of E, which will denote S s.
Lemma 8. Let (X, d) be a quasi metric space.
(i) For every s > 0 and every E ⊆ X we have that
H s(E) ≤ S s(E) ≤ K s2sH s(E),
where K denotes de triangle constant for d. In other words, the
measures H s and S s are equivalents.
(ii) If F is locally s-Ahlfors with measure ν, then there exists a constant
c such that
c−1(diam(B(x, r)))s ≤ ν(B(x, r)) ≤ c(diam(B(x, r)))s
for every x ∈ F and every 0 < r < r0.
(iii) If (X, d) has finite metric dimension and (F, d) is locally s-Ahlfors
with measure ν, then (F, d) is locally s-Ahlfors with the restriction
of H s to F .
Proof. The first inequality in (i) is straightforward. For the second one, let
us fix ρ > 0 and let {Ei} a ρ-cover of E. We can assume that Ei ∩ E 6= ∅
for every i. Let ε > 0 given. For each i, let us fix xi ∈ Ei ∩ E and set
Bi = B(xi, diam(Ei) + ε). Then Ei ⊆ Bi, so that {Bi} is a covering of E by
d-balls with diameter less than 2K(ρ + ε). Then
S s
2K(ρ+ε)(E) ≤Xi
diams(Bi) < 2sK sXi
diams(Ei) + 2sK sεs.
Hence S s
ρ → 0 we obtain the result.
2K(ρ+ε)(E) ≤ 2sK sH s
ρ (E) + 2sK sεs. Taking ε = ρ and by making
To prove (ii), let c such that c−1rs ≤ ν(B(x, r)) ≤ crs for every x ∈ F
and every 0 < r < r0. Fix x ∈ F and 0 < r < r0 and set B = B(x, r). Since
diam(B) ≤ 2Kr, the first inequality follows immediately with c = c2sK s.
To obtain the second one, let us consider two cases according to the size
of diam(B).
If diam(B) < r0 there exists 0 < ε < diam(B) such that
diam(B)+ε < r0. Also for each x ∈ F we have B(x, r)∩F ⊂ B(x, diam(B)+
ε) ∩ F . Then
(4.8)
ν(B(x, r)) ≤ ν(B(x, diam(B) + ε)) ≤ c2s(diam(B))s.
Otherwise, if diam(B) ≥ r0, since r < r0 we have r < diam(B). So that
c = c2sK s works.
Finally, in order to prove (iii), let us fix x ∈ F and 0 < r < r0, and set
F (x, r) to denote the set F ∩ B(x, r). By hypothesis there exists a constant
c such that
c−1rs ≤ ν(F (x, r)) ≤ crs.
Let {Bj} be a covering of F (x, r) by d-balls centering in F (x, r), and let rj
the radio of Bj. Since we can assume 0 < rj < r0 for every j, from (4.8) we
have that
ν(Bj ∩ F ) ≤ c2s(diam(Bj))s.
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
14
Then
c−1rs ≤ ν(F (x, r)) ≤Xj
ν(Bj ∩F ) ≤ c2sS s(F (x, r)) ≤ c4sK sH s(F (x, r)).
On the other hand, we claim that there exists a constant Λ, which does
not depend on x or r, such that H s(F (x, r)) ≤ Λrs. To prove it we shall
consider two possibilities. First let us assume that 0 < r < r0/K, and
let 0 < ρ < r0/K − r. Set U a finite ρ-net in F (x, r), let us say U =
{x1, x2, . . . , xIρ}. Then {B(xi, ρ) : i = 1, . . . , Iρ} is a 2Kρ-cover of F (x, r)
and each y ∈ F (x, r) belongs to at most N of such balls, where N is the
constant from the finite metric dimension of X, which does not depend on
ρ, r or x. In fact, for a fixed y ∈ F (x, r), we have that y ∈ B(xi, ρ) if and
only if xi ∈ B(y, ρ), so that the number of balls B(xi, ρ) to which y belongs
is equal to the cardinal of U ∩ B(y, ρ). Then we have
H s
2Kρ(F (x, r)) ≤
Iρ
Iρ
ρs
≤ c
Xi=1
Xi=1
≤ cN ν
ν (B(xi, ρ))
Iρ
[i=1
B(xi, ρ)
≤ cN ν (B(x, K(r + ρ)))
≤ c2N K s(r + ρ)s.
Taking ρ → 0 we obtain the desired result for this case. Then only remains
to consider the case r0
K ≤ r < r0. In this case, being B(x, r) a bounded set,
there exists a finite r0(2K)−1-net in B(x, r), let us say U = {x1, . . . , xIr0
}.
Then F (x, r) ⊆SIr0
i=1 B(cid:0)xi, r0
H s(cid:16)B(cid:16)xi,
Xi=1
H s(F (x, r)) ≤
Ir0
2K(cid:1). Applying the previous case we obtain
r0
2K(cid:17) ∩ F(cid:17) ≤ Ir0Λ(cid:16) r0
2K(cid:17)s
≤ Ir0Λ2−srs.
Moreover, we have that Ir0 ≤ N 1+log2 K, since every s-disperse subset of X
has at most N m points in each ball of radius 2ms, for all m ∈ N and every
s > 0 (see [CW71] and [Ass79]).
(cid:3)
Muckenhoupt weights on the Sierpinski gasket. We shall apply
Theorem 7 to a classical fractal set. Set T any equilateral triangle in R2.
Let S be the Sierpinski's gasket constructed in T equipped with the usual
distance d inherited from R2 and with the α-dimensional Hausdorff measure
H α, where α = log 3
log 2 . It is well known that (S, d) is α-Ahlfors with measure
H α (see [Mos97]). Set F the boundary of the triangle T . Then F is 1-
Ahlfors with measure ν, taking ν as the length. So that Theorem 7 says
that
d(x, F )(1−α)γ ∈ A1(S, d, H α),
for every 0 ≤ γ < 1.
POWERS OF DISTANCES AS MUCKENHOUPT WEIGHTS
15
References
Patrice Assouad. ´Etude d'une dimension m´etrique li´ee `a la possibilit´e de
plongements dans Rn. C. R. Acad. Sci. Paris S´er. A-B, 288(15):A731 -- A734,
1979.
Ronald R. Coifman and Guido Weiss. Analyse harmonique non-commutative
sur certains espaces homog`enes. Springer-Verlag, Berlin, 1971. ´Etude de cer-
taines int´egrales singuli`eres, Lecture Notes in Mathematics, Vol. 242.
Miguel de Guzm´an. Real variable methods in Fourier analysis, volume 46 of
North-Holland Mathematics Studies. North-Holland Publishing Co., Amster-
dam, 1981.
Ricardo G. Dur´an and Fernando L´opez Garc´ıa. Solutions of the divergence
and analysis of the Stokes equations in planar Holder-α domains. Math. Mod-
els Methods Appl. Sci., 20(1):95 -- 120, 2010.
A. Dall'Acqua and G. Sweers. Estimates for Green function and Poisson ker-
nels of higher-order Dirichlet boundary value problems. J. Differential Equa-
tions, 205(2):466 -- 487, 2004.
R. G. Dur´an, M. Sanmartino, and M. Toschi. Weighted a priori estimates for
the Poisson equation. Indiana Univ. Math. J., 57(7):3463 -- 3478, 2008.
Ricardo G. Dur´an, Marcela Sanmartino, and Marisa Toschi. Weighted a priori
estimates for solution of (−∆)mu = f with homogeneous Dirichlet conditions.
Anal. Theory Appl., 26(4):339 -- 349, 2010.
Javier Duoandikoetxea. Fourier analysis, volume 29 of Graduate Studies in
Mathematics. American Mathematical Society, Providence, RI, 2001. Trans-
lated and revised from the 1995 Spanish original by David Cruz-Uribe.
K. J. Falconer. The geometry of fractal sets, volume 85 of Cambridge Tracts
in Mathematics. Cambridge University Press, Cambridge, 1986.
[Ass79]
[CW71]
[dG81]
[DLG10]
[DS04]
[DST08]
[DST10]
[Duo01]
[Fal86]
[MM09]
[GCRdF85] Jos´e Garc´ıa-Cuerva and Jos´e L. Rubio de Francia. Weighted norm inequal-
ities and related topics, volume 116 of North-Holland Mathematics Studies.
North-Holland Publishing Co., Amsterdam, 1985. , Notas de Matem´atica
[Mathematical Notes], 104.
Misha Gromov. Metric structures for Riemannian and non-Riemannian
spaces. Modern Birkhauser Classics. Birkhauser Boston Inc., Boston, MA,
english edition, 2007. Based on the 1981 French original, With appendices
by M. Katz, P. Pansu and S. Semmes, Translated from the French by Sean
Michael Bates.
Peter W. Jones. Factorization of Ap weights. Ann. of Math. (2), 111(3):511 --
530, 1980.
Svitlana Mayboroda and Vladimir Maz'ya. Pointwise estimates for the poly-
harmonic Green function in general domains. In Analysis, partial differential
equations and applications, volume 193 of Oper. Theory Adv. Appl., pages
143 -- 158. Birkhauser Verlag, Basel, 2009.
Umberto Mosco. Variational fractals. Ann. Scuola Norm. Sup. Pisa Cl. Sci.
(4), 25(3-4):683 -- 712 (1998), 1997. Dedicated to Ennio De Giorgi.
Roberto A. Mac´ıas and Carlos Segovia. Lipschitz functions on spaces of ho-
mogeneous type. Adv. in Math., 33(3):257 -- 270, 1979.
Benjamin Muckenhoupt. Weighted norm inequalities for the Hardy maximal
function. Trans. Amer. Math. Soc., 165:207 -- 226, 1972.
Tord Sjodin. On s-sets and mutual absolute continuity of measures on homo-
geneous spaces. Manuscripta Math., 94(2):169 -- 186, 1997.
Ph. Souplet. A survey on Lp
δ spaces and their applications to nonlinear elliptic
and parabolic problems. In Nonlinear partial differential equations and their
applications, volume 20 of GAKUTO Internat. Ser. Math. Sci. Appl., pages
464 -- 479. Gakk¯otosho, Tokyo, 2004.
Hans Triebel. Fractals and spectra. Modern Birkhauser Classics. Birkhauser
Verlag, Basel, 2011. Related to Fourier analysis and function spaces.
[Muc72]
[Sjo97]
[Sou04]
[Mos97]
[MS79]
[Gro07]
[Jon80]
[Tri11]
16
[VK87]
[Wu98]
H. AIMAR, M. CARENA, R. DUR ´AN, AND M. TOSCHI
A. L. Vol′berg and S. V. Konyagin. On measures with the doubling condition.
Izv. Akad. Nauk SSSR Ser. Mat., 51(3):666 -- 675, 1987.
Jang-Mei Wu. Hausdorff dimension and doubling measures on metric spaces.
Proc. Amer. Math. Soc., 126(5):1453 -- 1459, 1998.
Instituto de Matem´atica Aplicada del Litoral (CONICET-UNL), Departa-
mento de Matem´atica (FIQ-UNL), Santa Fe, Argentina.
E-mail address: [email protected]
Instituto de Matem´atica Aplicada del Litoral (CONICET-UNL), Departa-
mento de Matem´atica (FHUC-UNL), Santa Fe, Argentina.
E-mail address: [email protected]
Instituto de Investigaciones Matem´aticas "Luis A. Santal´o" (CONICET-
UBA), Departamento de Matem´atica (UBA), Buenos Aires, Argentina.
E-mail address: [email protected]
Instituto de Matem´atica Aplicada del Litoral (CONICET-UNL), Departa-
mento de Matem´atica (FIQ-UNL), Santa Fe, Argentina.
E-mail address: [email protected]
|
1707.09004 | 5 | 1707 | 2018-07-03T12:58:20 | Hypergroup Deformations of Semigroups | [
"math.FA",
"math.CO"
] | We view the well-known example of the dual of a countable compact hypergroup, motivated by the orbit space of p-adic integers by Dunkl and Ramirez (1975), as hypergroup deformation of the max semigroup structure on the linearly ordered set $\mathbb{Z}_+$ of the non-negative integers along the diagonal. This works as motivation for us to study hypergroups or semi convolution spaces arising from "max" semigroups or general commutative semigroups via hypergroup deformation on idempotents. | math.FA | math | HYPERGROUP DEFORMATIONS OF SEMIGROUPS
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
Abstract. We view the well-known example of the dual of a countable compact hypergroup,
motivated by the orbit space of p-adic integers by Dunkl and Ramirez (1975), as hypergroup
deformation of the max semigroup structure on the linearly ordered set Z+ of the non-negative
integers along the diagonal. This works as motivation for us to study hypergroups or semi convo-
lution spaces arising from "max" semigroups or general commutative semigroups via hypergroup
deformation on idempotents.
8
1
0
2
l
u
J
3
]
.
A
F
h
t
a
m
[
5
v
4
0
0
9
0
.
7
0
7
1
:
v
i
X
r
a
1. Introduction
We introduce and study hypergroups, same as convolution spaces, in short, convos, or, semi
convolution spaces, in short, semiconvos [12] arising from general commutative semigroups via
deformation of the part of diagonal consisting of the idempotents. The genesis was the well-
known example related to the orbit space of p-adic integers by Dunkl and Ramirez [7] viewed
with this perspective on one hand and a good account of the structure of measure algebras
of certain linearly ordered semigroups with order topology in [9] and [16] on the other hand
(see also [11]). There is a substantial development of hypergroup deformations of groups and
their applications. To get an idea one can see [17] and [13]. The present work can be seen as a
complement of that because a group has no idempotents other than the identity.
The next section gives basics of semigroups and hypergroups in the form that we need with
a little touch of novelty at a few places. We begin Section 3 with an attempt to make a "max"
semigroup (S, <, ·) with the discrete topology into a hermitian discrete hypergroup by deforming
the product on the diagonal. Amongst other things, we arrive at the result that this can be
done if and only if either S is finite or S is isomorphic to (Z+, <, max). We determine its
dual and show that it becomes a countable compact hermitian hypergroup with respect to
pointwise multiplication. In Section 4, we prove our main theorem on semiconvo or hypergroup
deformations of idempotents in commutative semigroups.
2010 Mathematics Subject Classification. Primary 43A62, 20M14.
Key words and phrases. Semigroups, "Max" semigroups, Discrete hypergroups, Discrete semiconvos, Dunkl-
Ramirez example, Hypergroup deformation of idempotents, Dual hypergroups, Semiconvo deformation.
1
2
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
Let Z+ = N ∪ {0} and ×, the usual multiplication. For a subset T of S, χ
denotes the
T
characteristic function of T defined on S. For notational convenience, we take empty sums to be
zero.
2. Basics of semigroups and hypergroups
For basics of semigroups and hypergroups one can refer to standard books, monographs and
research papers. For instance, one can see ([10], [8], [3], [19] [9], [16], [11]) for semigroups and
([6], [12], [20], [7], [2], [5], [18], [14]) for hypergroups. However, we give below some of them in
the form we need.
2.1. Basics of semigroups.
(i) As in [16], we consider a non-empty set S linearly ordered by the relation '<'.
(a) For m, n ∈ S, we define m · n = max{m, n}. This makes it into a commutative
semigroup. In this paper, we call such a semigroup (S, <, ·) a "max" semigroup. At
times, we will write it as (S, <, max). Further, for m, n ∈ S, we write Ln = {k ∈ S :
k < n}, Um = {k ∈ S : m < k}.
(b) The linear order '<' and the "max" semigroup operation as in (a) above are ap-
propriately related in the sense that m, n, k ∈ S and m ≤ n together imply that
mk ≤ nk as specified in Definition 0.1 of linearly ordered semigroup in [11].
(c) We may define another product '·' by min{m, n} but we stick to max unless needed
for some explicit purpose.
(ii) We call a commutative semigroup (S, ·) max-min type if for m, n ∈ S, m · n is m or n.
Clearly, a commutative semigroup is max-min type if and only if it becomes a "max"
semigroup via : for m, n ∈ S, m < n if and only if m · n = n 6= m.
(iii) Let (S, ·) be a semigroup with identity e. For m, n ∈ S we usually write m · n = mn.
(a) A non-empty subset T of S is called an ideal in S if T S ⊂ T and ST ⊂ T, where
T S := {ts : t ∈ T, s ∈ S} and similarly for ST. As in [15], an ideal T (= S) in S will
be called a prime ideal if the complement S\T of T in S is a semigroup.
(b) Let E(S) denote the set of idempotent elements in S, i.e., the set of elements n ∈ S
such that n2 = n. We write E0(S) = E(S)\{e} and eS = S\E(S).
(c) Let G(S) denote the set {g ∈ S : ∃ h ∈ S with gh = hg = e}. Then G(S) is a group
contained in S called the maximal group. Note that G(S)∩E(S) = {e}. Set G1(S) =
{g ∈ G(S) : gm = m for all m ∈ E0(S)}. Clearly, G1(S) is a subgroup of G(S). Note
that members of G1(S) act on E0(S) as the identity via left multiplication of (S, ·).
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
3
Definition 2.1. Let (S, ·) be a semigroup with identity e.
(i) We call (S, ·) inverse-free in case for m, n ∈ S, mn = e holds if and only if m = n = e.
This condition is equivalent to saying that m or n is equal to e.
(ii) (S, ·) is called action-free if G1(S) = {e}.
Example 2.1.
(i) (Z+, +) , ((0, 1], ×) and ([1, ∞), ×) are inverse-free semigroups which are
not max-min type.
(ii) (Z+, max) and (Z+ ∪ {∞}, min) are max-min type inverse-free semigroups.
(iii) A commuting set of orthogonal projections on a Hilbert space H containing the identity
operator IH on H, with respect to composition of two operators, form a semigroup with
identity which is inverse-free.
We know that if (S, ·) is a commutative semigroup then (E(S), ·) is a semigroup. Interrelations
are collected in the following proposition whose proof is straight-forward.
Proposition 2.2. Let (S, ·) be a semigroup with identity e.
(i) If (S, ·) is inverse-free then G(S) = {e}. Converse part is also true if S is commutative.
(ii) If (S, ·) is max-min type then (S, ·) is inverse-free.
(iii) Suppose S\{e} is non-empty.
(a) (S\{e}, ·) is a semigroup if and only if (S, ·) is inverse-free.
(b) If (S, ·) is inverse-free then (S\{e}, ·) is an ideal in (S, ·).
(iv) Suppose E0(S) is non-empty. If (S, ·) is inverse-free and commutative then (E0(S), ·) is
a semigroup.
(v) For a commutative semigroup (S, ·), eS is an ideal if and only if eS is a prime ideal.
Proposition 2.3. (Dichotomy). Let S be a semigroup. For m ∈ S either mj, j = 1, 2, . . . are
all distinct or mj is an idempotent for some j ∈ N.
We will say m is of infinite order in the first case and of finite order in the second case.
Example 2.2. Here we provide an example of a commutative semigroup (S, ·), in which G(S) =
G1(S) 6= {e} and thus, S is not action-free.
We take S = (Z+ × {0}) ∪ {(0, 1)} and define '·' as follows:
(i, 0) · (j, 0) = (max{i, j}, 0) for i, j ∈ Z+,
(j, 0) · (0, 1) = (0, 1) · (j, 0) = (j, 0) for j 6= 0,
(0, 0) · (0, 1) = (0, 1) · (0, 0) = (0, 1) and
4
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
(0, 1) · (0, 1) = (0, 0).
Then e = (0, 0) and E0(S) = N × {0}. Also G(S) = {(0, α) : α = 0, 1} = G1(S).
The semigroup in our next example is action-free but not inverse-free.
Example 2.3. Let S = (Z+, max) × ({0, 1}, addition mod 2), a commutative semigroup with
identity e = (0, 0). It is easy to see that E0(S) = N × {0}, G(S) = {(0, α) : α = 0, 1}, and
G1(S) = {e}.
In this paper, we shall mainly be concerned with semigroups equipped with the discrete
topology, in short, discrete semigroups.
Remark 1. Let (S, <) be a linearly ordered set as in Item 2.1 (i).
(i) Let (S, <, max) be the "max" semigroup as in Item 2.1 (i)(a) above. We equip S with
the order topology τ0 and assume that S has an identity e. Then τ0 is discrete if and
only if
(a) e has an immediate successor, and,
(b) each m 6= e has an immediate predecessor as well as an immediate successor.
(ii) Two different types of examples for (i) above can be provided by
(a) (Z+, <, max) or its non-empty finite subsets, and,
(b) {0} ∪ {m ± 1
n+2 : m, n ∈ N} considered as a subset of the real line with the usual
order and topology.
2.2. Basics of hypergroups. Here we come to the basics of hypergroups. Dunkl [6], Jewett
[12] and Spector [20] independently created locally compact hypergroups (same as 'convos' in
[12]) under different names with the purpose of doing standard harmonic analysis. In this paper
we are mostly concerned with commutative discrete semiconvos or hypergroups. It is convenient
to write the definition in terms of a minimal number of axioms. For instance, see ([14, Chapter
1], [1]).
Let K be a discrete space. Let M (K) be the space of complex-valued regular Borel measures
on K. Let MF (K) and Mp(K) denote the subset of M (K) consisting of measures with finite
support and probability measures respectively. Let MF,p(K) = MF (K) ∩ Mp(K). At times, we
do not distinguish between m and δm for any m ∈ K because m 7→ δm is an embedding from K
into Mp(K). Here δm is the unit point mass at m, i.e., the Dirac-delta measure at m.
We begin with a map ∗ : K × K → MF,p(K). Simple computations enable us to extend '∗'
to a bilinear map called convolution, denoted by '∗' again, from M (K) × M (K) to M (K). At
times, for certain n ∈ K we will write qn for δn ∗ δn and Qn for its support.
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
5
A bijective map ∨ : m 7→ m from K to K is called an involution if m = m. We can extend it
to M (K) in a natural way.
Definition 2.4. A pair (K, ∗) is called a discrete semiconvo if the following conditions hold.
• The map ∗ : K × K → MF,p(K) satisfies the associativity condition
(δm ∗ δn) ∗ δk = δm ∗ (δn ∗ δk) for all m, n, k ∈ K.
• There exists (necessarily unique) element e ∈ K such that
δm ∗ δe = δe ∗ δm = δm for all m ∈ K.
A discrete semiconvo (K, ∗) is called commutative if δm ∗ δn = δn ∗ δm for all m, n ∈ K.
Definition 2.5. A triplet (K, ∗, ∨) is called a discrete hypergroup if
• (K, ∗) is a discrete semiconvo,
• ∨ is an involution on K that satisfies
(i) (δm ∗ δn
) = δn ∗ δ m for all m, n ∈ K and
(ii) e ∈ spt(δm ∗ δn) if and only if m = n.
A discrete hypergroup (K, ∗, ∨) is called hermitian if the involution on K is the identity map,
i.e., m = m for all m ∈ K.
Note that a hermitian discrete hypergroup is commutative.
We write (K, ∗) or (K, ∗, ∨) as K only if no confusion can arise.
Let K be a commutative discrete hypergroup. For a complex-valued function χ defined on K,
we write χ(m) := χ( m) and χ(m ∗ n) = RK χ d(δm ∗ δn) for m, n ∈ K. Now, define two dual
objects of K :
Xb(K) = {χ ∈ ℓ∞(K) : χ 6= 0, χ(m ∗ n) = χ(m)χ(n) for all m, n ∈ K} ,
bK =nχ ∈ Xb(K) : χ = χ, i.e., χ( m) = χ(m) for all m ∈ Ko .
Each χ ∈ Xb(K) is called a character and each χ ∈ bK is called a symmetric character. With
the topology of pointwise convergence, Xb(K) and bK become compact Hausdorff spaces. In
contrast to the group case, these two dual objects need not be the same and also need not have
a hypergroup structure.
Now we give some examples of hypergroups.
Example 2.4. Polynomial hypergroups: This is a wide and important class of hermitian
discrete hypergroups in which hypergroup structures are defined on Z+. This class contains
6
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
Chebyshev polynomial hypergroups of first kind, Chebyshev polynomial hypergroups of second
kind, Jacobi hypergroups, Laguerre hypergroups etc. For more details see [2] and [14].
For Illustration, we describe CP, the Chebyshev polynomial hypergroup of first kind which
arises from the Chebyshev polynomials of first kind. In fact, they define the following convolution
'∗' on Z+ :
δm ∗ δn =
1
2
δn−m +
1
2
δn+m for m, n ∈ Z+.
For any k ∈ N, K = kZ+ with ∗K×K makes K a discrete hypergroup in its own right.
The Chebyshev polynomial hypergroup of second kind (Z+, ∗) arises from the Chebyshev
polynomials of second kind and the convolution '∗' on Z+ is given by
δm ∗ δn =
min{m,n}Xk=0
m − n + 2k + 1
(m + 1)(n + 1)
δm−n+2k.
Example 2.5. Let Ha = {0, 1, 2, . . . , ∞}, 0 < a ≤ 1
2 , be the one-point compactification of Z+.
Dunkl and Ramirez [7] defined a convolution structure on Ha to make it a (hermitian) countable
compact hypergroup. For a prime p, let ∆p be the ring of p-adic integers and W be its group of
units, that is , {x = x0 + x1p + . . . + xnpn + . . . ∈ ∆p : xj = 0, 1, . . . , p − 1 for j ≥ 0 and x0 6= 0}.
For a = 1
p , Ha derives its structure from W-orbits of action of W on ∆p by multiplication in ∆p.
Next, Dunkl and Ramirez make the symmetric dual space cHa of Ha into a hermitian discrete
hypergroup. The members of cHa are given by {χn : n ∈ Z+}, where, for k ∈ Ha,
0
if k < n − 1,
Then the convolution '∗' on K = Z+ identified with cHa = {χn : n ∈ Z+} is dictated by pointwise
product of functions in cHa, that is:
χmχn = χmax{m,n} for m 6= n,
0 = χ0, χ2
χ2
1 =
a
1 − a
χ0 +
1 − 2a
1 − a
χ1,
χ2
n =
an
1 − a
χ0 +
n−1Xk=1
an−kχk +
1 − 2a
1 − a
χn for n ≥ 2.
We call (K, ∗) a (discrete) Dunkl-Ramirez hypergroup.
Ha has a good spectral synthesis in the sense that every closed subset of Ha is a set of spectral
synthesis for the Fourier algebra A(Ha) [7, Theorem 10.6]. Chilana (now, Ajit Iqbal Singh) and
Ajay Kumar [4] have strengthened this further.
χn(k) =
a
a−1
if k = n − 1,
1
if k ≥ n (or k = ∞).
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
7
Remark 2. We note a few more standard facts about a Dunkl-Ramirez hypergroup (K, ∗). We
follow the notation as in 2.1 and 2.2 above.
(i) For n ∈ Z+, (Ln ∪ {n}, ∗) is a subhypergroup of (K, ∗).
(ii) Let a 6= 1
2 . For n ∈ Z+, Qn = Ln ∪ {n} and as a consequence, for n ∈ N, #Qn ≥ 2.
(iii) Let a = 1
2 .
(a) For n ∈ N, Qn = Ln. For n ≥ 2, #Qn ≥ 2 whereas for n = 1, #Qn = 1.
(b) In fact, (L1 ∪ {1}, ∗) is a group isomorphic to ({0, 1}, addition mod 2)
3. Hypergroups arising from hypergroup deformations of idempotent elements
of "max" semigroups
To begin with, we assume that (S, ·) is a commutative discrete semigroup with identity e
such that E0(S) 6= φ. For q ∈ Mp(S), let Q be its support, in short, spt(q) and q(j) = q({j})
for j ∈ S. Then Q is countable. Observe that q = Pj∈Q q(j)δj with q(j) > 0 for each j ∈ Q
and Pj∈Q q(j) = 1. We prefer this to the usual form Pj∈S q(j)δj with q(j) ≥ 0 for j ∈ S and
Pj∈S q(j) = 1, unless otherwise stated.
Definition 3.1. A probability measure q on S with #Q ≥ 2 will be called non-Dirac.
3.1. Motivation. We note that a Dunkl-Ramirez hypergroup (as in Example 2.5 above) is a
hermitian (hence commutative) discrete hypergroup K = cHa (0 < a ≤ 1
2 ) and its convolution
'∗' arises as a hypergroup deformation of the semigroup (Z+, ·), where m · n = max{m, n} in the
sense that δm ∗ δn = δmn for m 6= n or, m = n = 0 and for m = n 6= 0, we have
δ1 ∗ δ1 =
δn ∗ δn =
a
1 − a
an
1 − a
δ0 +
1 − 2a
1 − a
δ1,
δ0 +
n−1Xk=1
an−kδk +
1 − 2a
1 − a
δn for n ≥ 2.
Also, each element of Z+ is an idempotent in (Z+, ·). Further, for n ∈ Z+, (Ln ∪ {n}, ∗) arises
as a hypergroup deformation of the finite subsemigroup (Ln∪{n}, max) of (Z+, ·). This motivates
the construction of a commutative discrete semiconvo or discrete hypergroup structure on S for
a commutative discrete semigroup (S, ·) by deforming the product. We elaborate as follows.
3.2. Hypergroups deformations of "max" semigroups. We now confine our attention to
a "max" semigroup (S, <, ·) as in Item 2.1(i)(a). We assume that S has an identity e and it
is equipped with the discrete topology. We try to deform this discrete semigroup (S, ·) into a
hermitian (hence commutative) discrete hypergroup by deforming the product on the diagonal
of S\{e}.
8
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
3.2.1. Preparatory material. (i) For a discrete space K and a complex-valued regular Borel
or a non-negative measure µ on K, we usually write µ(j) for µ({j}).
(ii)(a) Let (K, ∗) be a discrete hermitian hypergroup. Then by [12, Theorem 7.1A] the Haar
measure λ on K is given by: λ(e) = 1 and for e 6= n ∈ K, λ(n) =
1
(δn∗δn)(e) .
(b) Let K be a Dunkl-Ramirez hypergroup cHa (0 < a ≤ 1
Further, for n ∈ N, we have
2 ). Then λ(n) = 1−a
an for all n ∈ N.
δn ∗ δn(0) =
an
1 − a
=
1
λ(n)
,
δn ∗ δn(k) = an−k =
λ(k)
λ(n)
for 1 ≤ k < n and
δn ∗ δn(n) =
1 − 2a
1 − a
= 1 − X0≤k<n
(δn ∗ δn)(k) = 1 − (δn ∗ δn)(Ln) =
λ(n) − λ(Ln)
λ(n)
.
(c) (δn ∗ δn)(n) = 0 for some n ∈ N if and only if a = 1
n ∈ N. In this case, λ(n) = 2n−1 for n ∈ N.
2 if and only if (δn ∗ δn)(n) = 0 for all
We try to replace (Z+, <, ·) in the Dunkl-Ramirez hypergroups by a discrete "max" semigroup
(S, <, max) with identity in our next theorem.
We give a set of necessary and sufficient conditions for (S, ∗) to become a hermitian discrete
hypergroup with convolution product '∗' defined as follows:
δm ∗ δn = δn ∗ δm = δm·n(= δmax{m,n}) for m, n ∈ S with m 6= n, or, m = n = e,
δn ∗ δn = qn
for n ∈ S\{e}.
Here, qn is a probability measure on S with finite support Qn containing e and has the form
Pj∈Qn
qn(j)δj with qn(j) > 0 for j ∈ Qn and Pj∈Qn
qn(j) = 1.
This is equivalent to looking for conditions on S, Qn's, an S-tuple {vn}n∈S in [1, ∞) with
ve = 1 and vn = 1
qn(e) for n ∈ S\{e} and qn(j) for j ∈ Qn\{e}, n ∈ S\{e}.
Theorem 3.2. Let (S, <, ·) be a discrete (commutative) "max" semigroup with identity e and '
∗' and other related symbols as above. Then (S, ∗) is a hermitian discrete hypergroup if and only
if the following conditions hold.
(i) Either S is finite or (S, <, ·) is isomorphic to (Z+, <, max).
(ii) For n ∈ S\{e}, we have Ln ⊂ Qn ⊂ Ln ∪ {n}.
(iii) If #S > 2, then for e 6= m < n in S, we have
(a) qn(e) = qn(m)qm(e) and
(b) qn(e)(cid:16)1 +Pe6=k∈Ln
1
qk(e)(cid:17) ≤ 1;
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
9
or, equivalently, with vn = 1
qn(e) for n ∈ S,
(iii)' If #S > 2, then for e 6= m < n in S, we have
(a) qn(m) = vm
vn
and
(b) Pk∈Ln
vk ≤ vn.
Proof. Suppose (S, ∗) is a hermitian discrete hypergroup.
We prove (i) in two steps.
Step (α). Let n ∈ S\{e} and m < n . Associativity of (S, ∗) demands that (δn ∗ δn) ∗ δm =
δn ∗ (δn ∗ δm). This gives that qn ∗ δm = qn. But e ∈ Qn; therefore, m ∈ Qn. Hence Ln ⊂ Qn.
Since #Qn < ∞ we have #Ln < ∞. Clearly, Le = φ is finite.
Step (β). Assume that S is not finite. Consider any n ∈ S. By Step (α) Ln ∪ {n} is finite
and, therefore, Un = S\(Ln ∪ {n}) is non-empty. Choose any m ∈ Un. Suppose that m is not
immediate successor of n. Then, there exists an element m′ ∈ S such that n < m′ < m. So
Un ∩ Lm is finite and non-empty because m′ ∈ Un ∩ Lm and Lm is finite by Step (α). Since
Un ∩ Lm is finite we can enumerate its elements in order, say, n1 < n2 < · · · < ns. Choose
the immediate successor n1 of n. Thus each element n of S has an immediate successor, say,
s(n). Now, start with x0 = e, take x1 = s(e), x2 = s(x1), · · ·, xj = s(xj−1) and so on. Set
W = {xj : j ∈ Z+}. We claim that W = S. Suppose not; then there exists t ∈ S\W. Therefore,
t 6= e, and hence t > e. But x1 is immediate successor of e so t > x1. Again, x2 is the immediate
successor of x1, so t > x2. Repeating this process, we get t > xj for every j ∈ Z+. Hence
W ⊂ Lt. This shows that W is finite because Lt is finite by Step (α). This gives a contradiction.
So, S = W.
Therefore, either S is finite or (S, <, ·) is isomorphic to (Z+, <, max).
(ii) Consider any n ∈ S\{e}. Let, if possible, Qn 6⊂ Ln∪{n}. Then there exists m > n with m ∈
Qn. The associativity of '∗' demands that (δn ∗ δn) ∗ δm = δn ∗ (δn ∗ δm), which in turn gives that
qn∗δm = δm. Now, L.H.S. = qn(m)δm ∗δm +Pm6=j∈Qn
qn(j)δjm = qn(m)qm +Pm6=j∈Qn
By Step (i) (α), Lm ⊂ Qm. Now, Qm ⊂ spt(L.H.S.) = spt(R.H.S.) = {m}. But n ∈ Lm. Thus,
qn(j)δjm.
we have n = m, which is a contradiction. Therefore, Qn ⊂ Ln ∪ {n}. Hence, using Step (i)(α),
we get Ln ⊂ Qn ⊂ Ln ∪ {n} for n ∈ S\{e}.
(iii) Suppose #S > 2 and e 6= m < n in S.
(a) As in the proof of Step(i)(α), qn ∗ δm = qn. Now, as by (ii), m ∈ Qn, we get
L.H.S. = qn(e)δm + Xj∈Qn
e6=j6=m
qn(j)δjm + qn(m)δm ∗ δm
10
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
= qn(e)δm + Xj∈Qn
qn(j)δjm + qn(m) Xj′∈Qm
e6=j6=m
qm(j′)δj′ and
R.H.S. = qn(e)δe + Xe6=k6=m, k∈Qn
qn(k)δk + qn(m)δm.
Since jm 6= e for j, m ∈ S\{e} and R.H.S. = L.H.S., we get qn(e) = qn(m) qm(e).
(b) Because n ∈ S\{e}, we get 0 < Pj∈Ln
qn(e)
qj(e) for all j ∈ Ln\{e}. Therefore, we get
qn(j) ≤ Pj∈Qn
qn(j) = 1. By (iii)(a), qn(j) =
qn(e)
1 + Xe6=j∈Ln
1
qj(e)
= qn(e) + Xe6=j∈Ln
qn(e)
qj(e)
= Xj∈Ln
qn(j) ≤ 1.
It can be seen easily that (iii)' is only a rewording of (iii).
For the converse part, assume that (i)-(iii) all hold. We note right at the outset that, because
condition (ii) is satisfied, for n ∈ S\{e}, we may write qn = δn ∗ δn = Pj∈Ln∪{n} qn(j)δj =
Pj≤n qn(j)δj with qn(j) > 0 for j ∈ Ln and qn(n) ≥ 0.
To see that (S, ∗) is a hermitian discrete hypergroup, first note that the identity element e of
semigroup (S, ·) works as the identity of (S, ∗), i.e., δn ∗ δe = δe ∗ δn = δn for every n ∈ S.
Next, we show that e ∈ spt(δm ∗ δn) if and only if m = n. It is trivial in case m = n = e.
If m = n 6= e then e ∈ spt(δn ∗ δn) = Qn. Now, assume that e ∈ spt(δm ∗ δn). Let, if possible,
m 6= n, so δm ∗ δn = δmn. Without loss of generality we can assume that m < n, so δmn = δn
and n > e. Therefore, e ∈ spt(δm ∗ δn) shows that n = e, which is a contradiction. Hence m = n.
Now, we prove the associativity condition of (S, ∗), i.e., for m, n, k ∈ S
(1*)
(δm ∗ δn) ∗ δk = δm ∗ (δn ∗ δk).
If any one or more of {m, n, k} are e then (1*) is trivial. We now consider the cases when none
of m, n and k are e.
Case (i) If m 6= n 6= k 6= m, we first consider the subcase m < n < k. So, L.H.S. of (1*)
= δn ∗ δk = δk and R.H.S. of (1*) = δm ∗ δk = δk. Therefore, L.H.S. =R.H.S. Other
subcases can be dealt with in a similar way and hence (1*) holds.
Case (ii) Out of m, n, k exactly two are distinct. We get #S > 2.
(α) We consider first the following form of (1*): (δn ∗ δn) ∗ δm = δn ∗ (δn ∗ δm) with
e 6= n 6= m 6= e, which is the same as
(2*)
qn ∗ δm = δn ∗ δnm
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
11
When n < m, R.H.S. of (2*) is equal to δn ∗ δm = δm. The condition that
Qn ⊂ Ln ∪ {n} implies that Qn · m = {m} and using this, we get L.H.S. of
qn(j) = 1. Therefore, L.H.S.= R.H.S. and hence (2*) holds.
qn(j)δjm = (cid:16)Pj∈Qn
qn(j)(δj ∗ δm) = Pj∈Qn
qn(j)(cid:17) δm = δm as
(2*) = Pj∈Qn
Pj∈Qn
We now come to the case when m < n. Then R.H.S. of (2*) is qn = Pj≤n qn(j)δj
= Xj≤n
with qn(j) > 0 for j < n and qn(n) ≥ 0. So L.H.S. of (2*)
qm(k)δk
qn(m)qm(k)δk + qn(m)qm(m)δm.
qn(j)(δj ∗ δm) =Xj≤n
qn(j)δj +
Xj<m
qn(j)δjm + qn(m)Xk≤m
qn(j)
δm + Xk<m
j6=m
= Xm<j≤n
By (iii)(a) qn(k) qk(e) = qn(e) for k < n. So we get qn(m) qm(j) = qn(j) for all
j < m. Therefore, L.H.S. of (2*)
qn(m)qm(j) + qn(m)qm(m)
qn(j)δj +
Xj<m
qn(j)δj + qn(m)
Xj≤m
= Xm<j≤n
= Xm<j≤n
Further, we have Pj≤m qm(j) = 1. So, L.H.S. of (2*) =Pk≤n qn(k)δk = qn.
δm + Xk<m
δm + Xk<m
Therefore, we get L.H.S. of (2*) = R. H.S. of (2*) and hence (2*) holds.
qm(j)
qn(k)δk.
qn(k)δk
(β) Now, consider the following form of (1*) with m < n, (δn ∗ δm) ∗ δn = δn ∗ (δm ∗ δn).
R.H.S. = δn ∗ δn = (δn ∗ δm) ∗ δn = L.H.S.
(γ) The case m > n of the form (1*) as in (β) above follows simply because L.H.S.
= δm ∗ δn = δm = R.H.S.
(iii) If m = n = k then (1*) follows from the fact δs ∗ δn = δn ∗ δs for s ∈ Qn.
Hence (1*) holds and consequently (S, ∗) is a hermitian discrete hypergroup.
(cid:3)
Remark 3. In view of condition (i) in Theorem 3.2 and Remark 1(i) we note that for "max"
semigroups (S, <, max) of interest to us the discrete topology and the order topology coincide
(cf. [9], [16]).
We now interpret conditions (i)-(iii) in Theorem 3.2 above. Because a finite "max" semigroup
is isomorphic to (Zk = {j : 0 ≤ j < k}, <, max) for some k ∈ N, we confine our attention to Z+
to begin with.
12
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
Let V be the set of sequences v = (vj)j∈Z+ in [1, ∞) which satisfy (i) v0 = 1 and (ii) vn ≥
Pj∈Ln vj for n ∈ N.
For n ∈ N, let un = vn −Pj∈Ln
vj. Then simple calculations give the following:
v0 = 1,
v1 = 1 + u1,
v2 = (2 + u1) + u2,
v3 = (22 + 2u1 + u2) + u3,
... =
...,
to elaborate,
vn = (2n−1 + 2n−2u1 + . . . + un−1) + un for n ≥ 3.
Alternatively, we may start with a sequence u = (un)n∈N in [0, ∞) and define (vn)n∈Z+ as
above. Let U be the set of such sequences (un)n∈N's. The sets U and V have cardinality c.
The same is true if we confine our attention to domain Zk = {j : 0 ≤ j < k} for k ∈ N,
instead of Z+.
Corollary 3.3. For each v ∈ V (or the corresponding u ∈ U ), there is one and only one
(hermitian) hypergroup deformation (Z+, ∗) of (Z+, <, max) which satisfies (δn ∗ δn)(0) = 1
vn
n ∈ Z+. Further, for this deformation, the convolution '∗' and the Haar measure λ satisfy the
,
following conditions.
(i) λ(n) = vn for n ∈ Z+ and λ(n) − λ(Ln) = un for n ∈ N.
(ii) λ(Ln) ≤ λ(n) for n ∈ N.
(iii) For n ∈ N,
(a) δn ∗ δn(m) = λ(m)
(b) δn ∗ δn(n) = λ(n)−λ(Ln)
λ(n) for m < n,
λ(n)
,
(c) δn ∗ δn(m) = 0 for m > n.
(iv) For n ∈ N, spt(δn ∗ δn) =
Ln
if λ(n) = λ(Ln),
Ln ∪ {n}
if λ(n) > λ(Ln).
Proof. (i) This is immediate from 3.2.1(ii)(a) and expressions for (vn)n∈Z+ and (un)n∈N given
above.
(ii) This follows from Theorem 3.2(iii)(b).
(iii) We have only to appeal to Theorem 3.2(iii)(a).
(iv) This follows from (iii).
(cid:3)
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
13
Remark 4. We compare Theorem 3.2 and Corollary 3.3 above with the Dunkl-Ramirez hyper-
groups (Example 2.5 above). We freely use Remark 2, Section 3.1 and Item 3.2.1 above.
Qm = Lm for some m ∈ N if and only if a = 1
(a) For the Dunkl-Ramirez hypergroup cHa (0 < a ≤ 1
qm(e)(cid:16)1 +Pe6=j∈Lm
latter condition can be shortened to um = 0.
1
(b) For our theorem, for any arbitrarily fixed m ∈ S\{e}, Qm = Lm if and only if
qj(e)(cid:17) = 1. Under the correspondence set up after Theorem 3.2, the
2 ), qn(0) = an
1−a , n ∈ N. Further,
2 if and only if Qn = Ln for all n ∈ N.
(c) Suppose #S > 2. For any strictly increasing sequence (nj) in S\{e} contained in S,
considered as a subsemigroup of (Z+, <, max) in view of Theorem 3.2 (i) above, we can
find (qn(e))n∈S leading to a deformation satisfying Qnj = Lnj ∀j and Qn = Ln ∪ {n} for
n ∈ S\{e} but not equal to any nj.
3.3. The dual objects of (S, ∗). We now come to the dual objects Xb(S) and bS of hypergroup
(S, ∗), as in Theorem 3.2 and Corollary 3.3 above. For this purpose, to begin with, we fix any
v ∈ V (or the corresponding u ∈ U ) and consider the corresponding deformation (Z+, ∗) as
obtained in Theorem 3.2 and Corollary 3.3. We freely follow the concepts, results and notation
set up in this process.
Theorem 3.4. Let (S, <, ·) ∼= (Z+, <, max) and other symbols satisfy the conditions (ii)-(iii)
of Theorem 3.2 and let (S, ∗) be the corresponding deformed hypergroup with the Haar measure
λ. Then the dual objects Xb(S) and bS of (S, ∗) are equal. Equipped with the topology of uniform
convergence on compact subsets of S, bS can be identified with the one point compactification
+(= Z+ ∪ {∞}) of Z+. More precisely, the identification is given by k 7→ χ
k
, where χ
(n) = 1
Z∗
∞
for all n ∈ Z+, and, for k ∈ Z+, χ
is given by
k
(n) =
χ
k
1
if n ≤ k,
βk
0
if n = k + 1,
if n > k + 1,
where, βk = −λ(Lk+1)
λ(k+1) =
−Pj∈Lk+1
vj
vk+1
= uk+1
vk+1
− 1 = δk+1 ∗ δk+1(k + 1) − 1 = qk+1(k + 1) − 1.
Proof. Before we start the proof, it is helpful to give a diagrammatic interpretation (Table
1) of the functions χ
's, k ∈ Z+. As in the proof of Theorem 3.2, for n ∈ N, we may write
k
qn =Pn
j=0 qn(j)δj with qn(j) > 0 for 0 ≤ j < n, but qn(n) ≥ 0 and Pn
We note that, for k ∈ Z+, βk = qk+1(k + 1) − 1 and −1 ≤ βk < 0.
j=0 qn(j) = 1.
14
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
k+1
···
0
1
1 β0
1
1
1
...
1
1
...
1
1
1
...
1
1
...
2
0
β1
1
1
...
1
1
...
3
0
0
β2
1
...
1
1
...
···
· · ·
· · ·
· · ·
· · ·
...
· · ·
· · ·
...
χ
χ
χ
0
1
2
χ
...
χ
3
k
χ
k+1
...
k
0
0
0
0
0
0
0
0
...
...
1 βk
1
...
1
...
· · ·
· · ·
· · ·
· · ·
· · ·
· · ·
· · ·
. . .
Table 1. Part of character table of (S, ∗)
First, we check that for k ∈ Z∗
+ = Z+ ∪ {∞}, χ
k
Fix k ∈ Z+. It is easy to see that χ
(m ∗ n) = χ
(m)χ
k
k
It remains to check that for m ∈ Z+, (χ
(m))2 = χ
k
k
is a character. Clearly, χ
∞
(n) for all m 6= n ∈ Z+.
k
(m ∗ m), i.e.,
is a character.
(3)
Let m ∈ Z+.
Case (i) m ≤ k : We have χ
(m))2 =
(χ
k
mXj=0
qm(j)χ
(j).
k
(m) = 1 and on the other hand, χ
k
k
j=1 qm(j) = 1. Therefore, (3) holds.
Case (ii) m = k + 1 : We have
Pm
(m ∗ m) = Pm
j=0 qm(j)χ
(j) =
k
(m ∗ m) =
χ
k
k+1Xj=0
qk+1(j)χ
(j) =
k
kXj=0
qk+1(j) + qk+1(k + 1)(qk+1(k + 1) − 1)
= (1 − qk+1(k + 1)) + (qk+1(k + 1))2 − qk+1(k + 1)
= (qk+1(k + 1) − 1)2 = (χ
(m))2.
k
Therefore, (3) holds.
Case (iii) m > k + 1 : We have χ
k
(n) = 0 for n > k + 1. So, χ
j=0 qm(j) + qm(k +
1)(qk+1(k +1)−1). Since by Theorem 3.2 and by proof of part (iii)(b) of Theorem 3.2, we
know that for 1 ≤ j < m, qm(j) = qm(0)
qj(0) and 1 − qk+1(k + 1) = qk+1(0)(cid:16)1 +Pk
j=1
1
qj(0)(cid:17) ,
(m ∗ m) = Pk
k
we get
(m ∗ m) = qm(0)
1 +
χ
k
kXj=1
1
qj(0)
+
1
qk+1(0)
(qk+1(k + 1) − 1)
= 0 = (χ
(m))2.
k
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
15
Therefore, χ
k
is a character. Also, χ
is real-valued and S is hermitian. So, χ
k
k
is a symmetric
character.
Now, note that for k1 6= k2 ∈ Z∗
+, we have χ
6= χ
k2
k1
. We will show that injective map k 7→ χ
k
from Z∗
+ to Xb(S) is onto.
To prove this take any χ ∈ Xb(S) such that χ 6≡ 1. Since χ(0) = 1, there exists a least s ∈ N
such that χ(s) 6= 1. Let m, n ∈ Z+ with m < n. We have
χ(m)χ(n) = χ(m ∗ n) = χ(n).
This shows that either χ(m) = 1 or χ(n) = 0. As a consequence, χ(n) = 0 for all n > s.
Now consider the set
A = {j ∈ N : χ(t) = 0 for all t ≥ j}.
Then s + 1 ∈ A and so A 6= φ. Hence, there exists a least element, say, w in A.
Let, if possible, w = 1. Then χ(0) = 1 and χ(t) = 0 for all t ≥ 1. So, for any m ≥ 1, we get
χ(m ∗ m) = χ(m)χ(m) = 0. On the other hand, we have χ(m ∗ m) =Pm
0, which is a contradiction. Therefore, w > 1.
j=0 qm(j)χ(j) = qm(0) >
Note that χ(n) = 1 for all n < w − 1. In fact, let, if possible, χ(n1) 6= 1 for some n1 < w − 1.
Therefore we get χ(t) = 0 for t ≥ w − 1, which shows that w − 1 ∈ A. This is a contradiction to
the fact w is the least element of A.
Now, we come to χ(w − 1) = β (say). Then, we have
β2 = (χ(w − 1))2 = χ((w − 1) ∗ (w − 1)) =
w−1Xj=0
qw−1(j)χ(j)
=
w−2Xj=0
qw−1(j) + qw−1(w − 1) β.
This implies that β2 − qw−1(w − 1) β − (1 − qw−1(w − 1)) = 0 and therefore, β = 1 or qw−1(w −
1) − 1 = βw−2 where w − 2 ≥ 0.
Let, if possible, β = 1. As w ∈ A, χ(w ∗ w) = χ(w)χ(w) = 0; on the other hand, we have
χ(w ∗ w) =Pw
j=0 qw(j)χ(j) =Pw−1
j=0 qw(j) ≥ qw(0) > 0. Hence, β = βw−2.
Therefore, the character χ is given by
χ(n) =
1
if n ≤ w − 2,
βw−2
if n = w − 1,
0
if n > w − 1,
i.e., χ = χ
.
w−2
16
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
Therefore, the map k 7→ χ
k
Further, it is clear from Table 1 that bS equipped with the topology of uniform convergence on
compact subsets of S (i.e., the topology of pointwise convergence in our case) is the one point
is onto. As an immediate consequence, Xb(S) = bS.
compactification Z∗
+ of Z+.
(cid:3)
Remark 5. (i) Consider any countable infinite compact Hausdorff space X. Then any proba-
bility measure µ on X has the form µ = Px∈X µ(x)δx = Px∈X γxδx (say), where γx ≥ 0 and
Px∈X γx = 1. Trivially, µ has compact support. In particular, if X = bS as in Theorem 3.4
above, µ =Pj∈Z+ γjδj + γ∞δ∞, where γj ≥ 0 ∀j ∈ Z∗
Let C =(cid:8)j ∈ Z∗
+ : γj > 0(cid:9) . Then the support of µ is given by
+ and Pj∈Z+ γj + γ∞ = 1.
C =
C
if C is finite or ∞ ∈ C,
C ∪ {∞} otherwise.
(ii) For a non-constant complex-valued function f on bS ∼= Z∗
sense that f (j) = 1 for j ≥ j0 for some j0 ∈ N, we have, for µ on bS as in (i) above,
= 1 + Xj∈Z+
1 − Xj∈Z+
f (j)γj +
f dµ = Xj∈Z+
(f (j) − 1)γj.
ZbS
j<j0
j<j0
j<j0
γj
+, which is eventually 1 in the
(iii) For concepts, notation and details related to our next theorem, we refer to [2, Chapter
2, particularly, Section 2.4], [12, Section 6] and [22]. For the sake of convenience, we state
Proposition 2.4.2 of [2] in the form we need as follows.
" bS can be made into a hypergroup (with respect to pointwise multiplication) if for χ, χ′ ∈
bS, there exists a (regular) probability measure µχ,χ′ = Pj∈Z+ µχ,χ′(j)δχj + µχ,χ′(∞)δχ∞ =
Pj∈Z+ γχ,χ′
∞ δχ∞ (say) which satisfies
δχj + γχ,χ′
j
(a) for v ∈ Z+,
χ(v)χ′(v) = Xj∈Z+
γχ,χ′
j χ
j
(v) + γχ,χ′
∞ ,
(b) the map (χ, χ′) 7→ spt(µχ,χ′) is continuous on Z∗
+ × Z∗
+ → the space C(Z∗
+) of compact
subsets of Z∗
+ with Michael topology,
(c) χ
∞
∈ spt(µχ,χ′) if and only if χ = χ′."
In this case, we may set δχ ∗ δχ′ = µχ,χ′, take involution as identity and make bS a hermitian
hypergroup, in fact.
Theorem 3.5. Let (S, <, ·) ∼= (Z+, <, max) and other symbols satisfy the conditions (ii)-(iii)
of Theorem 3.2 and let (S, ∗) be the corresponding deformed hypergroup with the Haar measure
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
17
λ. The dual space bS of (S, ∗) becomes a countable compact hermitian hypergroup with respect to
pointwise multiplication. More precisely, the convolution '∗' on bS is given by
for m, n ∈ Z+ with m 6= n or m = n = ∞,
δχ
m
∗ δχ
n
=
δχ
min{m,n}
Pj∈Z+ γm
j δχ
otherwise,
j
where, γm
j = 0 for j < m, γm
m = 1 + βm ≥ 0, and for p ≥ 1, γm
βj's are as in Theorem 3.4. Further, we also have Xb(bS) = bbS ∼= (S, ∗).
m+p =Qm+p−1
j=m
−βj
1−βj+1
> 0, and,
Proof. We will freely use Remark 5.
It is clear from Table 1 that δχm ∗ δχn := δχ
min{m,n}
for m, n ∈ Z∗
+ with m 6= n or m = n = ∞.
We now look for a suitable probability measure µm,m, in short, µm, for m ∈ Z+. Fix m ∈ Z+.
By Remark (5)(i) above, we may write µm =Pj∈Z+ γm
Pj∈Z+ γm
j + γm
+. We note from Table 1 and Remark 5(ii) that the condition in Remark 5 (iii)(a)
∞ = 1. Now, for any t ∈ N, consider the function et on bS defined by et(χ
j ≥ 0 for j ∈ Z∗
) = χ
j δj +γm
∞δ∞, where γm
for n ∈ Z∗
+ and
n
(t)
n
takes the form
We may express it as:
χ
m
(t)2 − 1 = Xn∈Z+
n<t
(χ
n
(t) − 1)γm
n
for all t ∈ N.
(4)
1
1
...
1 − β0
0
1 − β1
0
0
1
...
1 − β2
...
· · ·
· · ·
· · ·
. . .
γm
0
γm
1
γm
2
...
=
(1)2
(2)2
(3)2
1 − χ
1 − χ
m
m
m
1 − χ
...
.
We consider any principal truncated submatrix of the infinite matrix on the L.H.S. of (4) above.
It is a lower triangular matrix and its diagonal entries 1 − βj's are all non-zero, and, therefore,
it is invertible. Now, for j ∈ N with j ≤ m, 1 − χ
(j)2 = 0. So we get γm
j = 0 for j < m . So
m
the system (4) can be replaced by the system:
(5)
1
1
...
1 − βm
0
1 − βm+1
0
0
1
...
1 − βm+2
...
· · ·
· · ·
· · ·
. . .
γm
m
γm
m+1
m+2
γm
...
=
1 − β2
m
1
1
...
and γm
j = 0 for j < m, if any.
The first two rows of (5) readily give,
18
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
m = 1 + βm ≥ 0, and γm
γm
m+1 =
−βm
1 − βm+1
.
We note that by Theorem 3.4, γm
m > 0 if and only if δm+1 ∗ δm+1(m + 1) > 0.
Consecutive pairs of rows then give that, for p ≥ 2,
γm
m+p =
−βm+p−1
1 − βm+p
γm
m+p−1.
γm
m+p =
m+p−1Yj=m
−βj
1 − βj+1
> 0.
So, for p ≥ 1,
Let sp =Pp
j=0 γm
m+j for p ≥ 1. First note that
s1 = γm
m + γm
m+1 = 1 + βm −
βm
1 − βm+1
= 1 +(cid:18) −βm
1 − βm+1(cid:19) βm+1 = 1 + γm
m+1βm+1.
Assume that sp = 1 + γm
m+pβm+p. Then
sp+1 = sp + γm
m+p+1 = 1 + γm
m+p βm+p + γm
m+p+1
= 1 + γm
m+pβm+p − γm
m+p
βm+p
1 − βm+p+1
= 1 + γm
m+p(cid:18) −βm+p
1 − βm+p+1(cid:19) βm+p+1 = 1 + γm
m+p+1βm+p+1.
Therefore, by mathematical induction we get sp = 1 + γm
m+pβm+p for p ≥ 1.
Let p ≥ 1. Then
γm
m+p βm+p = βm
pYj=1
−βm+j
1 − βm+j
,
which, by using the fact that −1 ≤ βk < 0 for all k ∈ Z+, gives
0 < −γm
m+p βm+p ≤(cid:18) 1
2(cid:19)p
.
j = limp→∞ sp = 1. So, γm
∞ = 1 −Pj∈Z+ γm
j = 0.
Therefore, Pj∈Z+ γm
Hence, we may set µm = δχ
m
The continuity of the map (χ
or m = n = ∞, spt(δχ
nχ
k
∈ bS : k ≥ mo ∪ {χ
or δm+1 ∗ δm+1(m + 1) = 0.
∗ δχ
n
} or nχ
m
∞
j δχ
.
j
m
n
∗ δχ
, χ
∗ δχ
m
) is singleton {χ
) 7→ spt(δχ
=Pj∈Z+ γm
∈ bS : k > mo ∪ {χ
m
k
) is easy to see by noting that for m 6= n
n
min{m,n}
} and for m ∈ Z+, spt(δχ
∗ δχ
m
) =
m
} according as δm+1 ∗ δm+1(m + 1) > 0
∞
It can be readily checked that χ
∞
is the identity element. We see that χ
∈ spt(δχ
∗ δχ
)
n
m
∞
if and only if m = n. In view of Remark (5)(iii), bS is a compact hermitian hypergroup.
For the rest of the proof, it is convenient to index the elements of bS by Z∗
fication k 7→ χ
.
+ under the identi-
k
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
19
Now, we calculate the dual space Xb(bS), the space of continuous characters on bS. We know
that S ⊂ bbS via n 7→ ξ
(k) = 1 for k ≥ 1, and, for n ≥ 2,
out from Table 1; to elaborate, ξ
+. We can figure them
(n) for all k ∈ Z∗
, defined as ξ
n
k
≡ 1, ξ
(χ
) = ξ
(k) := χ
(0) = β0, ξ
n
n
k
0
1
1
(k) =
ξ
n
βn−1
if k < n − 1,
if k = n − 1,
if k ≥ n.
0
1
Next, let ξ ∈ Xb(bS) such that ξ 6≡ 1. Consider the set
B = {s ∈ Z+ : ξ(s) 6= 1} .
Then B 6= φ. Take any s ∈ B, so ξ(s) 6= 1. Note that for m ∈ Z+, n ∈ Z∗
+ with m < n, we
have ξ(m) = ξ(m)ξ(n), and therefore, either ξ(m) = 0 or ξ(n) = 1. In particular, we get that
for all t ∈ Z+ with t < s, ξ(t) = 0 and hence t ∈ B. This implies that Ls ∪ {s} ⊂ B. So either
B = Z+ or B has a maximum element.
Let, if possible, B = Z+. Then, for each n ∈ N, ξ = 0 on Ln. Since this holds for every n ∈ N
we get ξ ≡ 0 on Z+. Since ξ(∞) = 1, we get that ξ is not continuous, a contradiction. Hence, B
has a maximum element, say w. Therefore, it follows from the definition of w that ξ(k) = 1 for
all k > w. Also, since ξ(w) 6= 1 we get ξ(k) = 0 for k < w.
Now, we calculate the value of ξ(w) = β (say). We have
β2 = ξ(w)2 = ξ(w ∗ w) = γw
wβ + Xj∈Z+
γw
j .
j>w
This implies that β2 − γw
wβ − (1 − γw
w) = 0 and therefore, β = 1 or γw
w − 1. But we know that
β = ξ(w) 6= 1 and hence ξ(w) = γw
w − 1 = βw. Now, set n = w + 1 ≥ 1, then ξ = ξn.
(cid:3)
Hence (S, ∗) ∼= Xb(bS).
Remark 6. (i)(a) The last part of Theorem 3.5 can be proved directly with the help of [2,
Proposition 2.4.18]. Our proof is easy and self contained.
(b) γ's in Theorem 3.5 are related via γm
j =
j
γ 0
γ 0
m
for j > m > 0.
(ii)(a) Any hypergroup (bS, ∗) as in Theorem 3.5 above may be thought of as hypergroup
deformation of the compact linearly ordered semigroup (Z∗
+, <, min) with the order topology
(cf. [9], [16], [11]) on the diagonal of Z+ × Z+.
(b) The class of hypergroups bS as in Theorem 3.5 above lies between the class of countable
compact hypergroups Ha (see Section 2 above) of [7] and the class of compact almost discrete
hypergroups of [21].
20
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
(iii) Let (S, <, max) be a finite "max" semigroup of cardinality n. Then it can be identified
with (Zn, <, max). Suitably truncated table of Table 1 and truncated submatrices in (4) and (5)
can be used to formulate and prove suitable analogues of Theorem 3.4 and Theorem 3.5.
4. semiconvos or hypergroups arising from deformations of commutative
discrete semigroups
We consider a commutative discrete semigroup (S, ·) with identity e such that E0(S) is not
empty. We try to make (S, ·) into a commutative discrete semiconvo or discrete hypergroup
(S, ∗) by deforming the product on DE0(S) := {(m, m) : m ∈ E0(S)}, the diagonal of E0(S), or
the idempotent diagonal of S, say.
For n ∈ E(S), let qn be a probability measure on S with finite support Qn containing e.
We express qn = Pj∈Qn
qn(j)δj with qn(j) > 0 for j ∈ Qn and Pj∈Qn
qn(j) = 1. We look for
necessary and sufficient conditions on S and {qn : n ∈ E0(S)} such that (S, ∗) with '∗' defined
below, is a commutative discrete semiconvo:
δm ∗ δn = δn ∗ δm = δmn for (m, n) ∈ S × S\DE0(S),
δn ∗ δn = qn for n ∈ E0(S).
It is convenient to consider various relationships among the properties of (S, ·) and of (S, ∗)
presented in the following theorems before coming to our main Theorem 4.3.
Theorem 4.1. Let (S, ·) be a commutative discrete semigroup with identity e. Suppose (S, ∗)
is a commutative discrete semiconvo, where '∗' and other related notation and concepts are as
above. Then the following are equivalent.
(i) For n ∈ E0(S), Qn ⊂ E(S).
(ii) (E(S), ·) is a max-min type semigroup.
(iii) S is inverse-free.
(iv) G1(S) = {e}, i.e., S is action-free.
(v) For n ∈ E0(S) and m ∈ eS (= S\E(S)) we have n 6= nm.
Proof. We prove the following implications: (i) ⇒ (iv) ⇒ (iii) ⇒ (v) ⇒ (i) and (iii) ⇔ (ii).
(i) ⇒(iv) . Assume Qn ⊂ E(S) for n ∈ E0(S). Consider any such n. Let, if possible, G1(S) 6= {e}
and choose any m(6= e) ∈ G1(S) so that mn = n 6= m. As (S, ∗) is a semiconvo, the
associativity of '∗' demands that (δm ∗ δn) ∗ δn = δm ∗ (δn ∗ δn). L.H.S. is equal to
δmn ∗ δn = δn ∗ δn = qn and R.H.S. is equal to δm ∗ qn. Since both sides are equal, we
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
21
get qn = qn ∗ δm. Now, m ∈ spt(qn ∗ δm). So m ∈ Qn, which is a contradiction. Hence,
G1(S) = {e}, i.e., S is action-free.
(iv) ⇒ (iii) . Assume that S is not inverse-free. Then G(S) does not act as the identity on E0(S). So
there exists k ∈ E0(S) such that km 6= k for some m(6= e) ∈ G(S). Note that k 6= e and
also δk 6= qk since e ∈ spt(qk). Now, there exists n(6= e) ∈ G(S) such that mn = e. Note
that kn 6= n because if kn = n then k = e by multiplying by m on both sides, which is
not true. Similarly, km 6= m. Consider the three different cases:
(A) km 6= kn, (B) km = kn /∈ E0(S), and (C) km = kn ∈ E0(S).
Now, because (S, ∗) is a semiconvo, the associativity of '∗' demands that
(6*)
(δkm ∗ δkn) ∗ δk = δkm ∗ (δkn ∗ δk).
Case (A). When km 6= kn. First, note that kn 6= k because if kn = k then by
multiplying both sides by m, we get k = km so that kn = k = km, which can not
happen. The L.H.S. of (6*) is equal to δk2mn ∗ δk = qk because k ∈ E0(S) and R.H.S.
of (6*) = δkm ∗ δk2n = δkm ∗ δkn = δk2mn = δk because km 6= kn and k ∈ E0(S). Since
δk 6= qk, (6*) does not hold.
Case (B). When km = kn /∈ E0(S). Then L.H.S. of (6*) is equal to δk2mn ∗ δk =
δk ∗ δk = qk because k ∈ E0(S). And R.H.S. of (6*) is equal to δkm ∗ δkn = δk2mn = δk
because kn /∈ E0(S) and k ∈ E0(S). Therefore, (6*) does not hold.
Case (C). When km = kn ∈ E0(S). As (S, ∗) is a semiconvo, the associativity of '∗'
demands that
(7*)
(δk ∗ δkn) ∗ δn = δk ∗ (δkm ∗ δn).
Now, L.H.S.= δk2n ∗ δn = δkn ∗ δn = δkn2 and R.H.S.= δk ∗ δkmn = δk ∗ δk = qk. Since
e ∈ spt(qk) = spt(R.H.S.) = spt(L.H.S.) = {kn2}, we get kn2 = e, so that k ∈ G(S) and
k ∈ G(S) ∩ E(S) = {e}, which is a contradiction. Therefore, (7*) does not hold.
Therefore, in all three cases the associativity condition in Definition 2.4 does not hold
so that (S, ∗) is not a semiconvo, which is a contradiction to the assumption. Hence, S
is inverse-free.
(iii) ⇒ (v) . Let, if possible, n = nm. The associativity of '∗' demands that for m, n ∈ S, (δn ∗ δn) ∗
δm = δn ∗ (δn ∗ δm). Now, as n ∈ E0(S) and m ∈ eS, we get qn ∗ δm = δn ∗ δnm = qn.
Therefore, Pj∈Qn
qn(j)(δj ∗ δm) =Pk∈Qn
qn(k)δk. But as m ∈ eS, we get
qn(e)δm + Xe6=j∈Qn
qn(j)δjm = qn(e)δe + Xe6=k∈Qn
qn(k)δk.
22
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
This shows that there exists a je(6= e) ∈ Qn such that jem = e which implies that
je = m = e because S is inverse-free. This is a contradiction. Therefore, n 6= nm.
(v)⇒ (i) . Let, if possible, there exist m /∈ E(S), which is in Qn. Since n 6= m, by the associativity
of '∗', we have qn ∗ δm = δn ∗ δnm. But n 6= nm, so we have that δn ∗ δnm = δn2m = δnm.
Therefore, Pj∈Qn
qn(j)(δj ∗ δm) = δnm. By the definition of '∗', we get
qn(e)δm + Xe6=j6=m, j∈Qn
qn(j)δjm + qn(m)δm2 = δnm.
So m = nm and m2 = nm. This shows that m = m2 and hence m ∈ E(S), which is a
contradiction. Therefore Qn ⊂ E(S).
(iii) ⇔ (ii) . Suppose (iii) holds. We know that (E(S), ·) is a subsemigroup of (S, ·). Let, if possible,
there exist m, n ∈ E(S) with m 6= mn 6= n. Then e 6= m 6= n 6= e. So, by Proposition
2.2(iv), mn ∈ E0(S). Now, the associativity of '∗' demands
(δm ∗ δn) ∗ δmn = δm ∗ (δn ∗ δmn).
We see that L.H.S. = δmn ∗ δmn = qmn. Further, R.H.S. = δm ∗ δnmn = δm ∗ δmn using
the fact that n ∈ E(S). As m 6= mn we get R.H.S.= δm2n, which in turn = δmn because
m ∈ E(S). Thus qmn = δmn. Now, e ∈ Qmn so mn = e; this is not true as it shows that
m, n ∈ G(S), which is a contradiction. Hence E(S) is a max-min type semigroup, i.e.,
(ii) holds.
The reverse implication follows from Proposition 2.2(ii).
(cid:3)
Theorem 4.2. Let (S, ·) be a commutative discrete semigroup with identity e. Suppose (S, ∗)
is a commutative discrete semiconvo, where '∗' and other related notation and concepts are as
above. Then we have the following:
(i) For m, n ∈ E0(S) with m 6= n 6= nm, we have m /∈ Qn and Qn · m = {nm}.
(ii) Under any (hence all) conditions (i)-(v) of Theorem 4.1, we have the following.
(a) (E(S), ∗) is a hermitian (hence commutative) discrete hypergroup.
(b) E(S) is finite or E(S) is isomorphic to (Z+, <, max), where the order on E(S) is
defined by m < n if mn = n 6= m.
(c) If #E(S) > 2, then for e 6= m < n in E(S), we have the following.
(α) qn(e) = qn(m)qm(e) and
(β) qn(e)(cid:16)1 +Pe6=k∈Ln
1
qk(e)(cid:17) ≤ 1,
where for n ∈ E(S), Ln := {j ∈ E(S) : j < n}.
= Pj∈Qn
qn(j)(δj ∗ δm) = qn(e)δm + qn(m)qm +P j∈Qn
e6=j6=m
qn(j)δjm. Since e ∈ Qm and e 6= m,
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
23
(d) eS is an ideal of (S, ·).
(e) For n ∈ E0(S), m ∈ eS we get Qn · m = {nm}.
Proof. (i). Let m, n ∈ E0(S) with m 6= n 6= nm. The associativity of '∗' demands that (δn ∗ δn) ∗
δm = δn ∗ (δn ∗ δm). Now L.H.S. = qn ∗ δm. Further, R.H.S. = δn ∗ δnm which in turn = δnm using
the fact that n 6= nm and n ∈ E0(S). Thus qn ∗ δm = δnm. Let, if possible m ∈ Qn. Then L.H.S.
the measure on L.H.S. is non-Dirac but R.H.S. is a Dirac measure, which is a contradiction.
Therefore, m /∈ Qn and hence, by qn ∗ δm = δnm, we get Qn · m = {nm}.
(ii)(a). We know that (E(S), ·) is a subsemigroup of (S, ·). Under the assumption, (i) in
Theorem 4.1 is satisfied. Therefore, '∗' induces a product on E(S), again denoted by '∗', i.e.,
∗ = ∗E(S) : E(S) × E(S) → MF,p(E(S)). Since we assume that e ∈ Qn, using (iii) of Theorem
4.1 it is easy to see that for all m, n ∈ E(S), e ∈ spt(δm ∗ δn) if and only if m = n. The
associativity follows as (S, ∗) is a discrete semiconvo. Hence, '∗' induces a hermitian discrete
hypergroup structure on E(S).
(b). Under the assumption that E(S) is a max-min type semigroup, by Item 2.1(ii), it can be
given a linear order by m < n if mn = n 6= m and made a "max" semigroup with identity e.
Therefore, it follows from Theorem 3.2 that E(S) is finite or E(S) is isomorphic to (Z+, <, max).
(c). This follows by (b) above as (E(S), ·) is a "max" semigroup and (E(S), ∗) is a hermitian
discrete hypergroup.
(d). Let k, n ∈ eS; then by the associativity of '∗' we get (δnk ∗ δn) ∗ δk = δnk ∗ (δn ∗ δk) which
is equivalent to δ(nk)2 = δnk ∗ δnk as n, k ∈ eS. Under the assumption S is inverse-free, it follows
that nk 6= e. This forces nk /∈ E0(S). Therefore, eS is a subsemigroup of (S, ·). Now, to show that
eS is an ideal it is enough to show that for k ∈ E0(S) and n ∈ eS, nk ∈ eS. This can be proved in
a manner similar to the proof that S is a subsemigroup by using k 6= nk and the associativity
condition (δnk ∗ δk) ∗ δn = δnk ∗ (δk ∗ δn).
(e). Let n ∈ E0(S), m ∈ eS. The associativity of '∗' demands that (δn ∗ δn) ∗ δm = δn ∗ (δn ∗ δm).
By the definition of '∗', we get qn ∗ δm = δn ∗ δnm. Since, by (d) above, n 6= nm, we get
qn ∗ δm = δnm, which, using (i), shows that Qn · m = {nm}.
(cid:3)
Now, we state our main theorem.
Theorem 4.3. Let (S, ·) be a commutative discrete semigroup with identity e such that S is
action-free. Let '∗' and other related notation and concepts be as above. Then (S, ∗) is a com-
mutative discrete semiconvo if and only if the following conditions hold.
24
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
(i) E(S) is finite or E(S) is isomorphic to (Z+, <, max), where the order on E(S) is defined
by m < n if mn = n 6= m.
(iii) Qn ⊂ E(S) for n ∈ E0(S).
(ii) (eS, ·) is an ideal of (S, ·).
(iv) If n ∈ E0(S) and m ∈ eS then Qn · m = {nm}.
(v) For n ∈ E0(S), we have Ln ⊂ Qn ⊂ Ln ∪ {n}, where for n ∈ E(S), Ln := {j ∈ E(S) :
j < n}.
(vi) If #E(S) > 2, then for e 6= m < n in E(S), we have the following:
(α) qn(e) = qn(m)qm(e) and
(β) qn(e)(cid:16)1 +Pe6=k∈Ln
1
qk(e)(cid:17) ≤ 1.
Further, under these conditions, E(S) is a hermitian discrete hypergroup. Moreover, S is a
hermitian discrete hypergroup if and only if S = E(S).
Proof. We assume all the conditions and prove that (S, ∗) is a commutative discrete semiconvo.
Since S has an identity, it is enough to check the associativity of '∗' to prove that (S, ∗) is a
semiconvo. We will prove that for m, n, k ∈ S
(8*)
(δm ∗ δn) ∗ δk = δm ∗ (δn ∗ δk).
We divide the proof of (8*) into various steps.
(i) If m, n, k ∈ eS, then using the fact the eS is an ideal (hence subsemigroup) of S we have
that the product of any two elements is in eS. Therefore, both sides of (8*) become δmnk.
(ii) If m, n ∈ eS and k ∈ E(S) then both sides of (8*) become δmnk using the fact eS is an
(iii) Suppose m ∈ eS and n, k ∈ E(S). If n or k is e then (8*) is trivial. Therefore, we can
assume that n, k ∈ E0(S).
ideal of S.
(a) When n = k ∈ E0(S). The condition (8*) becomes (δm ∗ δn) ∗ δn = δm ∗ (δn ∗ δn).
Now, δm ∗ (δn ∗ δn) = δm ∗ qn = Pj∈Qn
Qn ⊂ E(S) and m · Qn = {mn}, we get R.H.S. of (8*)
qn(j) (δm ∗ δj). Because for n ∈ E0(S),
= δm ∗ (δn ∗ δn) = Xj∈Qn
qn(j)δmj = Xj∈Qn
qn(j)δmn =
Xj∈Qn
qn(j)
δmn = δmn.
Using the fact eS is an ideal and n2 = n, we get L.H.S. of (8*) = (δm ∗ δn) ∗ δn = δmn.
Therefore, the associativity (8*) holds.
(b) For n 6= k ∈ E0(S), R.H.S. and L.H.S. of (8*) both become δmnk using the fact the
eS is an ideal in (S, ·). Therefore, (8*) holds.
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
25
(iv) We now come to cases when n ∈ eS and m, k ∈ E(S), or k ∈ eS and m, n ∈ E(S). These
can be dealt with by arguments similar to those in (ii) and (iii) above.
(v) When m, n, k ∈ E(S), associativity of '∗' can be proved in a manner similar to that for
proof of Theorem 3.2.
Hence, (S, ∗) is a semiconvo.
Converse part of the first part of the theorem follows from Theorem 3.2 and Theorem 4.2.
Since (S, ∗) has a semiconvo structure and (S, ·) is action-free, it follows from Theorem 4.2
that (E(S), ∗) has a hermitian hypergroup structure.
Moreover, if S is hermitian hypergroup then for any e 6= m ∈ S, e ∈ spt(δm ∗ δm); so we get
m ∈ E0(S) or m ∈ eS with m2 = e. But eS is an ideal and e in not in eS. So the second condition
is not possible, which shows that m ∈ E0(S). Therefore, S ⊂ E(S). Hence S = E(S).
(cid:3)
Remark 7.
(i) We can formulate a suitable analogue of Corollary 3.3 based on Theorem 4.3
above.
(ii) In view of Theorem 3.2 and Theorem 4.3 above, identifying E(S) as a finite subset of Z+
or Z+ itself, for n > 1, #Qn ≥ 2. But #Q1 can very well be 1. As already remarked, this
is the case for Dunkl-Ramirez hypergroup for a = 1
2 . We note that some proofs become
simpler, shorter or different when we take #Qn > 1 for all n.
(iii) R. C. Vrem [23] defined the join of two hypergroups. Because S = E(S) ⊔ eS, at the
first glance, our semiconvo (S, ∗) looks like the join of a hermitian discrete hypergroup
(E(S), ∗) and a discrete semigroup (eS, ·), but it is something different from join.
(iv) We can utilize the examples in Section 2 to illustrate various parts of theorems above. In
particular, it is a consequence of Theorem 4.1 that for Example 2.3, '∗' does not induce
a semiconvo structure on S.
The following semigroup (S, ·) is an example of a semigroup (S, ·) satisfying the conditions of
Theorem 4.3 and for which eS is non-empty.
Example 4.1. For T = {1 − 1
r+1 : r ∈ Z+}, let S = T ∪ N with '·' defined as follows.
m · n =
max{m, n} if m or n ∈ T,
m + n
if m, n ∈ N.
Then S is an inverse-free commutative discrete semigroup with identity 0. Further, E(S) = T ,
which is isomorphic to (Z+, <, ·). Also, eS = N is a prime ideal in S.
26
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
Acknowledgment
Vishvesh Kumar thanks the Council of Scientific and Industrial Research, India, for its senior
research fellowship. He thanks his supervisors Ritumoni Sarma and N. Shravan Kumar for their
support and encouragement.
A preliminary version of a part of this paper was included in the invited talk by Ajit Iqbal
Singh at the conference "The Stone- Cech compactification : Theory and Applications, at Centre
for Mathematical Sciences, University of Cambridge, July 6-8 2016" in honour of Neil Hindman
and Dona Strauss. She is grateful to the organizers H.G. Dales and Imre Leader for the kind
invitation, hospitality and travel support. She thanks them, Dona Strauss and Neil Hindman
and other participants for useful discussion. She expresses her thanks to the Indian National
Science Academy for the position of INSA Emeritus Scientist and travel support.
References
[1] M. Alaghmandan and E. Samei, Weighted discrete hypergroups, Indiana Univ. Math. J.,
65(2), (2016) 423-451.
[2] W. R. Bloom and Herbert Heyer, Harmonic analysis on probability measures on hyper-
groups, De Gruyter, Berlin (1995) (Reprint: 2011).
[3] J. F. Berglund, H. D. Junghenn and P. Milnes, Analysis on semigroups: Function spaces,
compactifications, representations, Canadian Mathematical Society Series of Monographs
and Advanced Texts (John Wiley & Sons, Inc.), New York (1989).
[4] Ajit Kaur Chilana and Ajay Kumar, Ultra-Strong Ditkin Sets in Hypergroups, Proc. Amer.
Math. Soc., 77(3) (1979) 353-358.
[5] W. C. Connett, O. Gebuhrer and A. L. Schwartz, editors. Application of hypergroups and
related measure algebras, Providence, R. I., American Mathematical Society, Contemporary
Mathematics, 183 ( 1995).
[6] C. F. Dunkl, The measure algebra of a locally compact hypergroup, Trans. Amer. Math.
Soc., 179 (1973) 331-348.
[7] C. F. Dunkl and D. E. Ramirez, A family of countable compact P⋆-hypergroups, Trans.
Amer. Math. Soc., 202 (1975) 339-356.
[8] C. F. Dunkl and D. E. Ramirez, Representations of commutative semitopological semi-
groups, Lecture Notes in Mathematics, 435, Springer-Verlag, Berlin-New York, (1975).
[9] E. Hewitt and H. S. Zuckerman, Structure theory for a class of convolution algebras, Pacific
J. Math., 7(1) (1957) 913-941.
[10] E. Hewitt and K. A. Ross, Abstract harmonic analysis Vol. I: Structure of topological groups,
integration theory, group representations, 2nd edition, Grundlehren der Mathematischen
HYPERGROUP DEFORMATIONS OF SEMIGROUPS
27
Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 115, SpringerVer-
lag, Berlin-New York, (1979).
[11] K. H. Hofmann and J. D. Lawson, Linearly ordered semigroups: historical origins and A.
H. Cliffords influence, Semigroup theory and its applications (New Orleans, LA, 1994).
London Mathematical Society. Lecture Note Series vol. 231, Cambridge University Press,
Cambridge, (1996) 15-39.
[12] R. I. Jewett, Spaces with an abstract convolution of measures, Adv. Math., 18 (1975) 1-101.
[13] S. Kawakami, T. Tsurii and S. Yamanaka, Deformations of finite hypergroups, Sci. Math.
Jpn., 79(2) (2016) 213-223.
[14] R. Lasser, Discrete commutative hypergroups, Inzell Lectures on Orthogonal Polynomials
(Inzell, 2001), Advances in the Theory of Special Functions and Orthogonal Polynomials
2, Nova Sci. Publ. Hauppauge, New York, (2005) 55-102.
[15] M. Petrich, Semicharacters of the cartesian product of two semigroups, Pacific J. Math.,
12(2) (1962) 679-683.
[16] K. A. Ross, The structure of certain measure algebras, Pacific J. Math., 11(2) (1961)
723-737.
[17] K. A. Ross and Daming Xu, Hypergroup deformations and Markov chains, J. Theoret.
Probab., 7(4), (1994) 813-830.
[18] K. A. Ross, J. M. Anderson, G. L. Litvinov, Ajit Iqbal Singh, V. S. Sunder and N. J.
Wildberger, editors. Harmonic analysis and hypergroups, Trends Math., Birkhauser Boston,
Boston, MA, (1998).
[19] W. Ruppert, Compact semitopological semigroups: an intrinsic theory, Lecture Notes in
Mathematics, vol. 1079. Springer-Verlag, Berlin (1984)
[20] R. Spector, Apercu de la th´eorie des hypergroupes, Analyse Harmonique sur les Groupes
de Lie (1975) 643-673 (Sem. Nancy-Strasburg 1973-1975, Lecture Notes in Mathematics
497, Springer-Verlag, Berlin, Heidelberg, New York 1975).
[21] M. Voit, Compact almost discrete hypergroups, Canad. J. Math., 48 (1996) 210-224.
[22] R. C. Vrem, Harmonic analysis on compact hypergroups, Pacific J. Math., 85(1) (1979)
239-251.
[23] R. C. Vrem, Hypergroup joins and their dual objects, Pacific J. Math., 111(2) (1984)
483-495.
Vishvesh Kumar
Department of Mathematics
Indian Institute of Technology Delhi
New Delhi - 110 016, India.
E-mail address: [email protected]
Kenneth A. Ross,
28
VISHVESH KUMAR, KENNETH A. ROSS, AND AJIT IQBAL SINGH
Prof Emeritus
Department of Mathematics
University of Oregon
Eugene, OR 97403, USA.
E-mail address: [email protected]
Ajit Iqbal Singh
INSA Emeritus Scientist
The Indian National Science Academy
New Delhi - 110002, India.
E-mail address: [email protected]
|
1704.00122 | 1 | 1704 | 2017-04-01T05:33:56 | Perturbation bounds for the Mostow and the bipolar decompositions | [
"math.FA"
] | Perturbation bounds for Mostow's decomposition and the bipolar decomposition of matrices have been computed. To do so, expressions for the derivative of the geometric mean of two positive definite matrices have been derived. | math.FA | math |
PERTURBATION BOUNDS FOR THE MOSTOW AND THE
BIPOLAR DECOMPOSITIONS
PRIYANKA GROVER AND PRADIP MISHRA
Abstract. Perturbation bounds for Mostow's decomposition and the bipolar de-
composition of matrices have been computed. To do so, expressions for the deriv-
ative of the geometric mean of two positive definite matrices have been derived.
1. Introduction
Matrix factorizations have been used in numerical analysis to implement efficient
matrix algorithms.
In machine learning, matrix factorizations play an important
role to explain latent features underlying the interactions between different kinds
of entities. Many matrix factorizations namely, the polar decomposition, the QR
decomposition, the LR decomposition etc., have been of considerable interest for
many decades. Perturbation bounds for such factorizations have been of interest for
a long time (see [2, 21, 22] and the references therein). Some generalizations and
improvements on them have been obtained in the subsequent works, for example,
see [11, 12, 13, 16, 17, 18, 19, 23].
An interesting matrix factorization follows from the work of Mostow [20]. It states
that every non singular complex matrix Z can be uniquely factorized as
Z = W eiKeS,
(1.1)
where W is a unitary matrix, S is a real symmetric matrix and K is a real skew
symmetric matrix. Recently, Bhatia [6] showed that every complex unitary matrix
W can be factorized as
(1.2)
where L is a real skew symmetric matrix and T is a real symmetric matrix. Using
(1.1) and (1.2), it has been obtained in [6] that
W = eLeiT ,
Z = eLeiT eiKeS.
(1.3)
Our goal is to find the perturbation bounds for the factors arising in (1.1), (1.2) and
(1.3). In [1], Barbaresco has used Berger Fibration in Unit Siegel Disk for Radar
2010 Mathematics Subject Classification. 15A23, 47A55, 65F60, 47A64, 47A30, 15A45.
Key words and phrases. Perturbation bounds, Derivative, Matrix factorizations, The bipolar
decomposition, Mostow's decomposition theorem, Polar decomposition, Sylvester's equation, Geo-
metric mean.
1
2
GROVER AND MISHRA
Space-Time Adaptive Processing and Toeplitz-Block-Toeplitz covariance matrices
based on Mostow's decomposition.
Let M(n, C) be the space of n × n complex matrices, and U(n, C) be the set of
n × n complex unitary matrices. Let · be any unitarily invariant norm on
M(n, C), that is, for any U, V ∈ U(n, C) and A ∈ M(n, C), we have
UAV = A.
Two special examples of such norms are the operator norm k · k (also known as the
spectral norm) and Frobenius norm k · k2 (also known as Hilbert-Schmidt norm or
Schatten 2-norm). Various properties of unitarily invariant norms are known [3,
Chapter IV]. We would require the following important properties: for A, B, C ∈
M(n, C)
(1.4)
ABC ≤ kAk B kCk,
and
(1.5)
Let W be a subspace of M(n, C) and let T : W → M(n, C) be a linear map. As in
[2], we take
A = A∗ = At = ¯A.
T = sup{T (X) : X = 1}.
(1.6)
It has been shown in [1, 6] that the factors in the decomposition (1.1) are related
to the geometric mean. So to obtain the perturbation bounds for (1.1), we obtain
expressions for the derivative of the geometric mean and bounds on its norms in
Section 2. In Section 3 and Section 4, we exploit the idea in [2] to obtain bounds on
the derivative of the decomposition maps for (1.1) and (1.2), respectively. In Section
5, we discuss the first order perturbation bounds for maps on Lie groups and obtain
the perturbation bounds for the factorizations (1.1), (1.2) and (1.3).
2. Derivative of the geometric mean
Let H(n, C) be the space of n × n complex Hermitian matrices and let P(n, C)
be the set of n × n complex positive definite matrices. For A, B ∈ P(n, C) their
geometric mean is defined as
[5, Chapter 4]. It is the unique positive solution of the Riccati equation
A#B = A1/2(cid:0)A−1/2BA−1/2(cid:1)1/2
A1/2,
XA−1X = B.
The geometric mean of A and B is also given by
A#B = A(A−1B)1/2 = (AB−1)1/2B,
(2.1)
(2.2)
(2.3)
where (A−1B)1/2 and (AB−1)1/2 are the unique square roots of A−1B and AB−1,
respectively, with positive eigenvalues.
PERTURBATION BOUNDS FOR THE MOSTOW AND THE BIPOLAR DECOMPOSITIONS 3
Let G : P(n, C) × P(n, C) → P(n, C) be the map defined as
G(A, B) = A#B.
Since A 7→ A1/2 is a differentiable function on P(n, C), we get from (2.1) that G is
a differentiable map. The derivative is given by
DG(A, B)(X, Y ) =
G(A + tX, B + tY ) for all X, Y ∈ H(n, C).
d
dt(cid:12)(cid:12)(cid:12)(cid:12)t=0
The following proposition gives an expression for DG(A, B).
Proposition 2.1. For X, Y ∈ H(n, C)
DG(A, B)(X, Y ) =Z ∞
e−t(BA−1)1/2
0
(Y + (BA−1)1/2X(A−1B)1/2)e−t(A−1B)1/2
dt.
(2.4)
Proof. For sufficiently small t, by (2.2), we have
G(A + tX, B + tY )(A + tX)−1G(A + tX, B + tY ) = B + tY.
(2.5)
Differentiating with respect to t at 0, we get
(DG(A, B)(X, Y )) A−1G(A, B) − G(A, B)(A−1XA−1)G(A, B)
+G(A, B)A−1 (DG(A, B)(X, Y )) = Y.
Put D = DG(A, B)(X, Y ) and C = A−1G(A, B) = (A−1B)1/2. Then the above
equation can be rewritten as
C ∗D + DC = Y + C ∗XC.
(2.6)
This is a well studied Sylvester's equation (see [3, 9]). By [3, Theorem VII.2.3], we
obtain
DG(A, B)(X, Y ) =Z ∞
0
e−tC ∗
(Y + C ∗XC)e−tCdt.
Substituting C = (A−1B)1/2 in (2.7), we obtain (2.4).
(2.7)
(cid:3)
Some other expressions for the solution of the Sylvester's equation [3, 10] are
known. From these, one can obtain other expressions for DG(A, B)(X, Y ).
Suppose A and B commute. Then C = (A−1B)1/2 is Hermitian. Let λ1(C) ≥
· · · ≥ λn(C) denote the eigenvalues of C. Using [3, Theorem VII.2.15] for (2.6), we
obtain
DG(A, B)(X, Y ) ≤
π
4λn(C)Y + C ∗XC.
DG(A, B) = sup{DG(A, B)(X, Y )
By (1.6),
(X, Y ) = max{X,Y }. So we get
π
:
(X, Y ) = 1}, where
(2.8)
DG(A, B) ≤
4λn(C)(cid:0)1 + kCk2(cid:1) .
4
GROVER AND MISHRA
By Proposition 2.1, we obtain a better bound for DG(A, B). We mention this
in the following corollary for general A and B.
Corollary 2.2. For A, B ∈ P(n, C)
DG(A, B) ≤(cid:18)Z ∞
ke−t(A−1B)1/2
In the case when A and B commute, R ∞
obtain
0
DG(A, B) ≤
In some other cases, a bound on R ∞
λn (Re C) is nonnegative (where Re C = C+C ∗
2
k2dt(cid:19)(cid:0)1 + k(A−1B)1/2k2(cid:1) .
0 ke−tCk2dt = 1
1
2λn(C)(cid:0)1 + kCk2(cid:1) .
(2.9)
(2.10)
2λn(C) . So from (2.9), we
0 ke−tCk2dt is easy to calculate. For example, if
), then we have
Z ∞
0
ke−tCk2dt ≤
1
2λn (Re C)
.
This has been observed in [2, 7].
Remark 2.3. We observe that for A, B ∈ P(n, C), DG(A, B) is a positive linear map
from H(n, C) × H(n, C) to M(n, C). So by [5, Theorem 2.6.3], we obtain
kDG(A, B)k = kDG(A, B)(I, I)k.
3. Mostow's decomposition
The Mostow decomposition theorem (1.1) gives that any non singular matrix
Z can be uniquely factorized as Z = W eiKeS. Let P1 = eiK and P2 = eS. Then
P1 ∈ P(n, C) and P2 ∈ P(n, R), where P(n, R) stands for the set of n×n real positive
definite matrices. We also have P1P1 = I. Such matrices X which satisfy XX = I
are called circular (or coninvolutary) [15]. Let Pcirc be the set of circular positive
definite matrices. Then P1 ∈ Pcirc. Let : GL(n, C) → U(n, C) × Pcirc × P(n, R) be
the map
(3.1)
where 0(Z) = W , 1(Z) = P1, and 2(Z) = P2 . Since the factorizations in (1.1)
are unique, these maps are well defined. The product map τ (W, P1, P2) = W P1P2 is
the inverse of . For any matrix A, let cond(A) denotes the condition number of A.
(Z) = (0(Z), 1(Z), 2(Z)),
0 ke−t((Z ∗Z)−1Z ∗Z)1/2
k2dt. Then
Theorem 3.1. For Z ∈ GL(n, C) let β(Z) =R ∞
1 k kP −1
2 k
D0(Z) ≤ kP −1
D1(Z) ≤
2
(cid:0)1 + kP1kβ(Z)cond(Z)(cid:0)1 + cond(Z)4(cid:1)(cid:1) ,
cond(P1) kP −1
2 k
(cid:0)1 + kP1kβ(Z)cond(Z)(cid:0)1 + cond(Z)4(cid:1)(cid:1) ,
2
(3.2)
(3.3)
PERTURBATION BOUNDS FOR THE MOSTOW AND THE BIPOLAR DECOMPOSITIONS 5
and
D2(Z) ≤ β(Z)cond(Z)(cid:0)1 + cond(Z)4(cid:1) .
Proof. We know 2(Z) = eS = (Z ∗Z#Z ∗Z)1/2. Let f : (0,∞) → R be defined as
f (t) = t1/2, g : P(n, C) → P(n, C) × P(n, C) as g(A) = (A, A), and h : GL(n, C) →
P(n, C) as h(Z) = Z ∗Z. Then 2 = f◦G◦g◦h, where G is the geometric mean map as
defined in Section 2. By the chain rule, D2(Z) = Df (Z ∗Z#Z ∗Z)◦DG(Z ∗Z, Z ∗Z)◦
Dg(Z ∗Z) ◦ Dh(Z).
Now by [3, Theorem X.3.1], we obtain that if A ∈ P(n, C), then
(3.4)
So
Df (A) ≤
1
2kA−1k1/2.
(3.5)
(3.6)
(3.7)
1
2k(Z ∗Z#Z ∗Z)−1k1/2 DG(Z ∗Z, Z ∗Z)(Z ∗A + AZ, Z ∗A + AZ).
D2(Z)(A) ≤
We know that (A#B)−1 = A−1#B−1 and kA#Bk ≤ kAk1/2kBk1/2. Therefore
1
2kZ −1k DG(Z ∗Z, Z ∗Z)(Z ∗A + AZ, Z ∗A + AZ).
D2(Z)(A) ≤
Let C = (Z ∗Z)−1(cid:0)Z ∗Z#Z ∗Z(cid:1) = (cid:0)(Z ∗Z)−1Z ∗Z(cid:1)1/2
(2.9), we obtain
. Then kCk ≤ cond(Z)2. By
and so
1
D2(Z)(A) ≤
2kZ −1kβ(Z)(cid:0)1 + kCk2(cid:1)Z ∗A + AZ,
D2(Z)(A) ≤ β(Z) cond(Z)(cid:0)1 + cond(Z)4(cid:1) A.
Equation (3.4) follows from (3.7).
space at any point P1 is given by iP 1/2
Let SH(n, R) be the space of n × n real skew symmetric matrices. The tangent
. This follows from [2, p. 258].
Let D(Z) : M(n, C) → W SH(n, C) ⊕ iP 1/2
1 ⊕ H(n, R) be defined
, Y2), where X ∈ SH(n, C), Y1 ∈ SH(n, R) and
SH(n, R)P 1/2
SH(n, R)P 1/2
1 Y1P 1/2
1
1
as D(Z)(A) = (W X, iP 1/2
Y2 ∈ H(n, R). So we have
1
1
X ∗ = −X, Y1 = Y1, Y t
1 = −Y1, Y2 = Y2, Y t
2 = Y2.
(3.8)
The map D(Z) is the inverse of Dτ (W, P1, P2), and so
D0(Z)(A) = W X, D1(Z)(A) = iP 1/2
1 Y1P 1/2
1
, D2(Z)(A) = Y2,
(3.9)
and
Dτ (W, P1, P2)(W X, iP 1/2Y1P 1/2, Y2) = A.
(3.10)
6
Also,
GROVER AND MISHRA
Dτ (W, P1, P2)(W X, iP 1/2Y1P 1/2, Y2) =
=
τ (W etX , P 1/2
1 eitY1P 1/2
1
, P2 + tY2)
W etXP 1/2
1 eitY1P 1/2
1
(P2 + tY2)
d
d
dt(cid:12)(cid:12)(cid:12)(cid:12)t=0
dt(cid:12)(cid:12)(cid:12)(cid:12)t=0
= W XP1P2 + W P 1/2
1
(iY1)P 1/2
1 P2 + W P1Y2. (3.11)
By (3.10) and (3.11), we obtain
W XP1P2 + W P 1/2
1
(iY1)P 1/2
1 P2 + W P1Y2 = A,
that is,
X + P 1/2
1
(iY1)P −1/2
1
= (W ∗A − P1Y2)(P1P2)−1.
Taking conjugate transpose on both the sides and using (3.8), we get
− X + P −1/2
Adding (3.13) and (3.14) gives
1
(iY1)P 1/2
1 = (P1P2)−1(A∗W − Y2P1).
(3.12)
(3.13)
(3.14)
(P 1/2
(iY1)P 1/2
1 (P 1/2
By [3, Theorem VII.2.3], we obtain
1 + P −1
)P −1
1
1
1
(iY1)P 1/2
1
) = Re(cid:0)(W ∗A − P1Y2)(P1P2)−1(cid:1) .
1 dt. (3.15)
= Z ∞
0
e−tP −1
1 Re(cid:0)(W ∗A − P1Y2)(P1P2)−1(cid:1) e−tP −1
1 k2dt(cid:19) Re(cid:0)(W ∗A − P1Y2)(P1P2)−1(cid:1)
1
(iY1)P 1/2
ke−tP −1
P 1/2
1
(iY1)P 1/2
1
1
So
D1(Z)(A) = P 1/2
≤ (cid:18)Z ∞
≤ kP1k
≤
0
2 (W ∗A − P1Y2)(P1P2)−1
cond(P1) kP −1
2 k
2
The last inequality follows from (3.7) and (3.9). Hence we obtain (3.3).
(cid:0)1 + kP1kβ(Z) cond(Z)(cid:0)1 + cond(Z)4(cid:1)(cid:1) A.
By (3.12), we also have
XP1 + P 1/2
1
(iY1)P 1/2
1 = (W ∗A − P1Y2)P −1
2
.
(3.16)
Again taking conjugate transpose on both the sides and using (3.8), we obtain
Now, subtracting (3.17) from (3.16), we get
1
(iY1)P 1/2
P1X + P 1/2
1 = P −1
2 (A∗W − Y2P1).
XP1 + P1X = 2i Im(cid:0)(W ∗A − P1Y2)P −1
2 (cid:1) .
(3.17)
PERTURBATION BOUNDS FOR THE MOSTOW AND THE BIPOLAR DECOMPOSITIONS 7
Again by [3, Theorem VII.2.3], we get
X = Z ∞
0
e−tP1 Im(cid:0)(W ∗A − P1Y2)P −1
2 (cid:1) e−tP1dt.
Therefore
D0(Z)(A) = X
≤ (cid:18)Z ∞
≤ kP −1
0
2
From this, (3.2) follows.
ke−tP1k2dt(cid:19)(W ∗A − P1Y2)P −1
1 k kP −1
2 k
2
(cid:0)1 + kP1kβ(Z) cond(Z)(cid:0)1 + cond(Z)4(cid:1)(cid:1) A.
(cid:3)
Remark 3.2. We have used in (3.7) that kA#Bk ≤ kAk1/2kBk1/2. Better bounds
on D2(Z)(X) can be found using [8, Theorem 2]. For example, we also have
D2(Z) ≤ kZk β(Z)k(Z ∗Z)−1/4(Z ∗Z)−1/2(Z ∗Z)−1/4k
(cid:0)1 + kZ −1k4k(Z ∗Z)1/4(Z ∗Z)1/2(Z ∗Z)1/4k2(cid:1) .
Remark 3.3. One can find another bound for D1(Z) in Theorem 3.1 by using
the expression eiK = e−SZ ∗Ze−S given in [6]. This can be expressed as 1(Z) =
(2(Z)−1(Z ∗Z)2(Z)−1)1/2 . Using this approach, the factor cond(P1)
in (3.3) gets
replaced by kP1k2. By the chain rule, we get
2 )(2 Re(P −1
D1(Z)(A) = Df (P −1
2 (D(2(Z)(A))P −1
2 Z ∗)(AP −1
2 Z ∗ZP −1
2 )),
2
2 − ZP −1
where f is the square root function. By [3, Theorem X.3.1] and using ZP −1
we obtain
2 = W P1,
D1(Z) ≤ kP1k2kP −1
2 k(cid:0)1 + kP1kβ(Z) cond(Z) (1 + (cond(Z))4)(cid:1) .(3.18)
4. Decomposition of unitary matrices
Every complex unitary matrix W can be factorized as W = W1W2, by the second
or third polar decomposition of W . This decomposition is unique if W ′W doesn't
have −1 as an eigenvalue. Let U = {W ∈ U(n, C) − 1 /∈ σ(W ′W )}, where σ(A)
denotes the spectrum of A and Usym+ be the set of U ∈ U(n, C) such that U ′ = U
and U has all the eigenvalues in the open right half plane. Let O(n, R) be the
set of real orthogonal matrices. We define Φ : U → O(n, R) × Usym+ as Φ(W ) =
(Φ1(W ), Φ2(W )), where Φ1(W ) = W1 and Φ2(W ) = W2. The product map Ψ :
O(n, R) × Usym+ → U is the inverse of Φ.
8
GROVER AND MISHRA
∞
Theorem 4.1. Let σ(W2) = {eiθ1, . . . , eiθn}. Let {an} be any ℓ1-sequence such that
for all θ = θi − θj (1 ≤ i, j ≤ n)
Xn=−∞
DΦk(W ) ≤ 2 ∞
Xn=−∞
(−1)naneinθ =
an! .
Then for k = 1, 2
1 + eiθ .
(4.1)
(4.2)
1
Proof. The map
DΦ(W ) : W SH(n, C) → W1 SH(n, R) ⊕ W
1
2
1
2 iH(n, R) W
2
2
is an isomorphism and its inverse is DΨ(W1, W2). For X ∈ SH(n, R) and Y ∈
iH(n, R)
DΨ(W1, W2)(W1X, W
2 Y W
2 ) =
1
2
1
2
d
dt(cid:12)(cid:12)(cid:12)(cid:12)t=0
Ψ(cid:16)W1etX , W
2 etY W
1
2
2 (cid:17)
1
2
1
2
1
2
1
2
1
2
= W1(XW2 + W
2 Y W
2 ).
Let S ∈ SH(n, C) be such that DΦ(W )(W S) = (W1X, W
2 Y W
2 ). Then we have
W S = W1(XW2 + W
2 Y W
2 ),
1
2
1
2
that is,
(4.3)
Taking transpose on both the sides of the above equation (4.3) and adding the new
equation to (4.3), we get
2 = X + W 1/2
2 Y W −1/2
W2SW −1
.
2
W2Y + Y W2 = W
2 SW
2 + W
2 S ′W
2 .
3
2
−
1
2
−
1
2
3
2
(4.4)
By [3, Theorem VII.2.7], we obtain
∞
−n−
+ W
2
3
2
n+ 3
S ′W
2
2
(cid:17) .
Y =
This gives
Therefore
1
2
−n+ 1
2
2
n−
SW
2
Xn=−∞
an(−1)n(cid:16)W
Y ≤ 2 ∞
Xn=−∞
DΦ2(W )(W S) = Y
≤ 2 ∞
Xn=−∞
an!S.
an!W S.
(4.5)
PERTURBATION BOUNDS FOR THE MOSTOW AND THE BIPOLAR DECOMPOSITIONS 9
Equation (4.3) can also be written as
W 1/2
2 SW −1/2
2
= W −1/2
2 XW 1/2
2 + Y.
(4.6)
Taking complex conjugate on both sides of the above equation (4.6) and adding the
new equation to (4.6), we get
By similar calculations as done above, we get
XW2 + W2X = W2S − S ′W2.
W1X ≤ 2 ∞
Xn=−∞
an! S,
and so
DΦ1(W )(W S) = X ≤ 2 ∞
Xn=−∞
an! W S.
Equations (4.5) and (4.8) give the required result.
(4.7)
(4.8)
(cid:3)
5. Perturbation bounds
In this section, we discuss first order perturbation bounds for a map from a Lie
group to a manifold and use it to obtain perturbation bounds for the decomposition
maps. Let M ⊆ GL(n, C) be a differentiable manifold. For A ∈ M let A denote a
perturbation of A in a small neighbourhood of A in M. Suppose A = A1A2. Then
Ai denote the corresponding factors for A. Let f be a smooth function on M. If
M is a convex set, then by Taylor's theorem, we have
f ( A) − f (A) ≤ Df (A) A − A + O( A − A2).
(5.1)
We denote this as
f ( A) − f (A) . Df (A) A − A.
We also note here that if there is a M > 0 such that Df (A) < M, then in a
small neighborhood of A, we have
f ( A) − f (A) < M A − A.
(5.2)
5.1. First order perturbation bounds. The function log is well defined for all
non singular matrices A if we choose a branch of logarithm. In this case, exp is its
inverse. The map D log(A) : M(n, C) → M(n, C) is given by
D log(A)(X) =Z 1
0
(t(A − I) + I)−1 X (t(A − I) + I)−1 dt.
(5.3)
For ǫ > 0 define Uǫ = {X ∈ Mn(C) : kXk2 < ǫ} and Vǫ = exp(Uǫ). Let
G ⊆ GL(n, C) be a matrix Lie group with Lie algebra G and let A0 ∈ G. Then by
10
GROVER AND MISHRA
[14, Theorem 2.27], there exists an ǫ > 0 such that the map H : Uǫ ∩ G → A0Vǫ ∩ G
defined by H(X) = A0 exp(X) is a bijective map. For X ∈ G
DH(O)(X) = A0D exp(O)(X)
= A0Z 1
0
= A0X.
e(1−t)OXetOdt
This gives
DH(O) ≤ kA0k.
DH1(A)(X) = D log (A−1
(5.4)
Let H1 : A0Vǫ → M(n, C) be the map defined as H1(W ) = log(A−1
0 W ). Note that
the restriction of H1 to A0Vǫ∩ G is H −1, and so (H, Uǫ∩G, A0Vǫ ∩ G) is a local chart
around A0 ∈ G. For A ∈ A0Vǫ
= Z 1
= Z 1
0 A − I) + I(cid:1)−1 dt
0 (cid:0)t(A−1
(t(A − A0) + A0)−1 X (t(A − A0) + A0)−1 A0 dt.
0 A − I) + I(cid:1)−1 A−1
0 X(cid:0)t(A−1
0 A)(A−1
0 X)
0
By Taylor's theorem, we have
H1( A) − H1(A) . DH1(A) A − A
≤ kA0k(cid:18)Z 1
0 k (t(A − A0) + A0)−1 k2dt(cid:19) A − A.
In particular, when A = A0, we have
H1( A0) − H1(A0) . kA0k kA−1
(5.5)
Let G1 ⊆ GL(n, C) be a differential manifold (G1 may not be a group) and let
F : G → G1 be a smooth map. Then F ◦ H : Uǫ ∩ G → G1. Note that H1(A0) = O.
Let S ∈ Uǫ ∩ G be in a small neighbourhood of O. Then we have
0 k2 A0 − A0.
(F ◦ H)( S) − (F ◦ H)(O) . D(F ◦ H)(O) S.
Therefore
(F ◦ H)( S) − (F ◦ H)(O) . DF (A0) DH(O) S.
Let A0 = H( S). Using equations (5.4) and (5.7), we get
F ( A0) − F (A0) . DF (A0) kA0k S
= DF (A0) kA0k H1( A0) − H1(A0).
By (5.5), we obtain
(5.6)
(5.7)
F ( A0) − F (A0) . DF (A0) cond(A0)2 A0 − A0.
(5.8)
PERTURBATION BOUNDS FOR THE MOSTOW AND THE BIPOLAR DECOMPOSITIONS 11
In particular, when A0 is unitary matrix, we get
F ( A0) − F (A0) . DF (A0) A0 − A0.
(5.9)
5.2. Perturbation bounds for the bipolar decomposition. Equation (5.9) and
Theorem 4.1 together give the perturbation bounds for the decomposition (1.2). We
state this as a proposition below. The notations are as in Section 4.
Proposition 5.1. For W ∈ U and k = 1, 2
Wk − Wk . 2 ∞
Xn=−∞
an! W − W.
(5.10)
As observed in [6], the expression (1.2) gives both the second and third polar
decompositions for a unitary matrix W . Therefore Theorem 4.6 and Theorem 4.7
in [2] give that for each k
Wk − Wk .(cid:18)Z ∞
0 ke−tW2k2dt(cid:19) W − W.
(5.11)
We see that the bounds obtained in (5.10) are sometimes better than the ones
given by (5.11). For example, let W = diag(eiθ, eiθ), where π/3 < θ < π/2. Then
0 ke−tW2k2dt =
W1 = I, W2 = W andZ ∞
n 6= 0. Then {an} ∈ ℓ1 satisfies (4.1) and 2P∞
If the eigenvalues of W2 are close to i or −i, then the bounds in (5.11) are too
large. But the bounds we get in (5.10) depend upon how far the eigenvalues of W2
lie on the unit circle. We explain this below.
> 1. Let a0 = 1/2 and an = 0 for all
n=−∞ an = 1.
2 cos θ
1
Let Θ = {θi − θj : eiθj ∈ σ(W2)} ⊆ (−δ, δ), where 0 < δ < π. We define the
function f : [−π, π] → C as
Then f is periodic and absolutely continuous. Also,
2
1
1
1
2 + i tan δ
2 − i tan θ
2 + i
2 tan δ
f (θ) =
Z π
−π f ′(θ)2dθ =
2
2(π−δ) (θ + π)
2 (−1 + θ−δ
(π−δ) )
−π ≤ θ ≤ −δ,
−δ ≤ θ ≤ δ,
δ ≤ θ ≤ π.
tan2 δ
2
2(π − δ)
+
tan2 δ
2
12
+
tan δ
2
4
.
So f ′ ∈ L2[−π, π]. Let the Fourier coefficients of f be bn. For the sequence an =
(−1)nbn we have an ∈ ℓ1 and for θ ∈ (−δ, δ)
1
∞
(−1)naneinθ =
1 + eiθ =
1
2 −
i
2
tan
θ
2
.
(5.12)
Xn=−∞
an ≤ a0 +
π
√3kf ′kL2.
2
+
tan2 δ
2
3
+ tan
δ
2
.
(5.13)
12
GROVER AND MISHRA
From [4, p. 117], we know that
∞
∞
π
∞
2
Therefore
bn =
Xn=−∞
Xn=−∞
√3s 2 tan2 δ
(π − δ)
by (5.13), we see that 2P an is very close to 1.
decomposition (1.3). For Z ∈ GL(n, C) let
ke−t((Z ∗Z)−1Z ∗Z)1/2
an ≤ 1 +
Xn=−∞
k(Z) =(cid:18)Z ∞
0
and for W unitary let
If δ is very small, that is, if the eigenvalues of W2 are very close to each other, then
We now obtain the perturbation bounds for the factors S, K, L, T in the bipolar
k2dt(cid:19) cond(Z)(cid:0)1 + cond(Z)4(cid:1) ,
C(W ) =Z 1
0 k(t(W − I) + I)−1k2dt.
Before stating the theorem, we observe a few things about the decomposition
(1.3). For any Z ∈ GL(n, C) the matrices S and K are unique but L and T are not
unique. If eL and eiT do not have −1 as an eigenvalue, then we can use principal
logarithm to define L and T uniquely. But if eL or eiT have −1 as an eigenvalue,
then we choose α ∈ [−π, 0) such that eiα /∈ σ(W1) ∪ σ(W2). A branch of logarithm
for which arg z ∈ [α, α + 2π) gives unique S, K, L and T .
Theorem 5.2. Let Z ∈ GL(n, C) be such that −1 /∈ σ(Z ′Z((Z ∗Z)−1#(Z ∗Z)−1)).
Let Z = eLeiT eiKeS, where σ(eiT ) = {eiθ1, . . . , eiθn}. Let {an} be any ℓ1-sequence
such that for all θ = θi − θj (1 ≤ i, j ≤ n)
∞
Then
(−1)naneinθ =
Xn=−∞
1
1 + eiθ .
2
L − L . 2 C(eL) ∞
Xn=−∞
T − T . 2 C(eiT ) ∞
Xn=−∞
cond(eiK) ke−Sk
an! ke−iKk ke−Sk
an! ke−iKk ke−Sk
K − K . e−iK
2
2
(cid:0)1 + keiKk k(Z)(cid:1) Z − Z,
(cid:0)1 + keiKk k(Z)(cid:1) Z − Z,
(5.14)
(5.15)
(cid:0)1 + keiKk k(Z)(cid:1) Z − Z,
(5.16)
PERTURBATION BOUNDS FOR THE MOSTOW AND THE BIPOLAR DECOMPOSITIONS 13
Proof. Let Z = W eiKeS. Using notations as in section 3, we get
S − S . e−S k(Z) Z − Z.
W − W . D0(Z) Z − Z,
ei K − eiK . D1(Z) Z − Z,
e S − eS . D2(Z) Z − Z.
(5.20)
The matrices eS and eiK are both positive. So log(eS) = S and log(eiK) = iK.
We have
Similarly,
S − S . D log(eS) e S − eS
. D log(eS) D2(Z) Z − Z.
K − K . D log(eiK) D1(Z) Z − Z.
(5.22)
We know that if η is an operator monotone function on (0,∞), then for A ∈ P(n, C)
Dη(A) ≤ kη ′(A)k [3, Theorem X.3.4]. Now since log is an operator monotone
function on (0,∞), we obtain
and
and
and
(5.17)
(5.18)
(5.19)
(5.21)
(5.23)
(5.24)
D log(eS) ≤ e−S
D log(eiK) ≤ e−iK.
Equations (5.23) and (3.4) give (5.17). And, (5.24) and (3.3) give (5.16).
Since −1 /∈ σ(Z ′Z((Z ∗Z)−1#(Z ∗Z)−1)), −1 /∈ σ(W ′W ). Then by the second or
third polar decomposition, W can be uniquely factorized as W = W1W2, where W1
and W2 are also unitary matrices. If L = log W1 and iT = log W2, then we have
W = eLeiT . Now
L − L = log(e L) − log(eL)
. D log(eL) e L − eL.
By (5.3), we obtain
Equation (5.10) gives
L − L . C(eL) e L − eL.
Therefore we have
an! W − W.
L − L . 2C(eL) ∞
Xn=−∞
an! D0(Z) Z − Z.
L − L . 2C(eL) ∞
Xn=−∞
14
GROVER AND MISHRA
By (3.2), we obtain (5.14). Similar calculations by putting T = 1
(5.15).
i log W2 yield
(cid:3)
We illustrate the behaviour of the above bounds with the help of an example.
For a natural number n, consider a one-parameter family of matrices Zn(t) =
diag(esin t, esin(t+ π
n )), t ∈ R. For Zn(t), the factors in the bipolar decomposition
are given by Sn(t) = diag(sin t, sin(t + π
n )), Kn(t) = Tn(t) = Ln(t) = O. We consider
the operator norm in Theorem 5.2. Let fn(t) be the first order perturbation bounds
as given in (5.17), that is, fn(t) = ke−Sn(t)k k(Zn(t)). Then
fn(t) =
1
n ))(cid:1)2
2(cid:0)max(e− sin t, e− sin(t+ π
(cid:0)1 + max(e−4 sin t, e−4 sin(t+ π
max(esin t, esin(t+ π
n )) max(e4 sin t, e4 sin(t+ π
n ))
n ))(cid:1) .
The behavior of fn(t) can be seen in the following graph. We observe that for n = 2,
the perturbation bound for some of these matrices can be more than 1200.
1200
1000
800
600
400
200
-3
-2
-1
1
2
3
When n increases, the maximum value of fn(t) decreases. In particular, we observe
this for n = 500 in the below graph.
PERTURBATION BOUNDS FOR THE MOSTOW AND THE BIPOLAR DECOMPOSITIONS 15
2.5
2.0
1.5
1.0
0.5
-3
-2
-1
1
2
3
Bounds for other factors Kn(t), Ln(t) and Tn(t) are given by gn(t) := 1
k(Zn(t))) which also vary in a similar way.
2ke−Sn(t)k(1 +
Remark 5.3. Other perturbation bounds for L and T in Theorem 5.2 can also be
found using direct formulas, which we get from the principal logarithm. Let V be
2 log(W ′W ) do not
the set of complex unitary matrices W such that W ′W and W e−
2i log(W ′W ). Using the
have eigenvalue −1. Then L = log(W e−
chain rule and Taylor's theorem, we get
2 log(W ′W )) and T = 1
1
1
and
L − L ≤ (1 + C(W ′W )) W − W
T − T ≤ C(W ′W ) W − W.
Acknowledgement. The work of the first author is supported by the research
grant of INSPIRE Faculty Award [DST/INSPIRE/04/2014/002705] of Department
of Science and Technology, India.
References
[1] F. Barbaresco, Information geometry of covariance matrix: Cartan-Siegal homogeneous
bounded domains, Mostow/Berger fibration and Fréchet median, Matrix Information Ge-
ometry, eds. F. Nielsen, R. Bhatia, Springer, 2013, pp. 199-255.
[2] R. Bhatia, Matrix factorizations and their perturbations, Linear Algebra Appl. 197/198 (1994)
245 -- 276.
[3] R. Bhatia, Matrix Analysis, Springer, 1997.
[4] R. Bhatia, Fourier Series, Hindustan Book Agency, 2003.
[5] R. Bhatia, Positive Definite Matrices, Princeton University Press, 2007.
[6] R. Bhatia, The bipolar decomposition, Linear Algebra Appl. 439 (2013) 3031 -- 3037.
[7] R. Bhatia, L. Elsner, Higher order logarithmic derivatives of matrices in the spectral norm,
SIAM J. Matrix Anal. Appl. 25 (2003) 662 -- 668.
16
GROVER AND MISHRA
[8] R. Bhatia, P. Grover, Norm inequalities related to the matrix geometric mean, Linear Algebra
Appl. 437 (2012) 726 -- 733.
London Math. Soc. 29 (1997) 1 -- 21.
[9] R. Bhatia, P. Rosenthal, How and why to solve the operator equation AX − X B = Y , Bull.
[10] R. Bhatia, M. Uchiyama, The operator equationPn
= Y , Expo. Math. 27 (2009)
[11] X.-W. Chang, On the sensitivity of the SR decomposition, Linear Algebra Appl. 282 (1998)
251 -- 255.
An−iX B i
i=0
297 -- 310.
[12] X.-W. Chang, D. Stehle, Rigorous perturbation bounds of some matrix factorizations, SIAM
J. Matrix Anal. Appl. 31 (2010) 2841 -- 2859.
[13] A. Galántai, Perturbations of triangular matrix factorizations, Linear Multilinear Algebra 51
(2003) 175 -- 198.
[14] B.C. Hall, Lie Groups, Lie Algebras, and Representations : An Elementary Introduction,
Springer, 2003.
[15] R. Horn, C.R. Johnson, Topics in Matrix Analysis, Cambridge University Press, 1990.
[16] A. Largillier, Bounds for relative errors of complex matrix factorizations, Appl. Math. Lett. 9
(1996) 79 -- 84.
[17] W. Li, W. Sun, Perturbation bounds of unitary and subunitary polar factors, SIAM J. Matrix
Anal. Appl. 23 (2002) 1183 -- 1193.
[18] W. Li, W. Sun, New perturbation bounds for unitary polar factors, SIAM J. Matrix Anal.
Appl. 25 (2003) 362 -- 372.
[19] H. Li, H. Yang, H. Shao, Perturbation analysis for the hyperbolic QR factorization, Comput.
Math. Appl. 63 (2012) 1607 -- 1620.
[20] G.D. Mostow, Some new decomposition theorems for semi-simple groups, Mem. Amer. Math.
Soc. 14 (1955) 31 -- 54.
[21] G.W. Stewart, Perturbation bounds for the QR factorization of a matrix, SIAM J. Numer.
Anal. 14 (1977) 509 -- 518.
[22] G.W. Stewart, J.-G. Sun, Matrix Perturbation Theory, Academic Press, 1990.
[23] Z. Xie, W. Li, Sensitivity analysis for the SR decomposition, Linear Multilinear Algebra 63
(2015) 222 -- 234.
E-mail address: [email protected], [email protected]
Department of Mathematics, Shiv Nadar University, Dadri, U.P. 201314, India.
|
1207.3861 | 1 | 1207 | 2012-07-17T02:07:53 | A refinement of the companion of Ostrowski inequality for functions of bounded variation and applications | [
"math.FA",
"math.CA"
] | In this paper we establish a refinement of the companion of Ostrowski inequality for functions of bounded variation. Applications for the trapezoid inequality, the mid-point inequality, and to probability density functions are also given. | math.FA | math |
A REFINEMENT OF THE COMPANION OF OSTROWSKI INEQUALITY FOR
FUNCTIONS OF BOUNDED VARIATION AND APPLICATIONS
WENJUN LIU AND YUN SUN
Abstract. In this paper we establish a refinement of the companion of Ostrowski inequality for
functions of bounded variation. Applications for the trapezoid inequality, the mid-point inequality,
and to probability density functions are also given.
1. Introduction
In 1938, Ostrowski [21] established the following interesting integral inequality for differentiable
mappings with bounded derivatives:
Theorem 1.1. Let f : [a, b] → R be a differentiable mapping on (a, b) whose derivative is bounded
on (a, b) and denote kf ′k∞ = sup
t∈(a,b)
f ′(t) < ∞. Then, for all x ∈ [a, b], one has the inequality
(1.1)
f (x) −
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
b − aZ b
a
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
The constant 1
4 is sharp in the sense that it can not be replaced by a smaller one.
≤" 1
4
+
(x − a+b
2 )2
(b − a)2 # (b − a)kf ′k∞.
This inequality has attracted considerable interest over the years, and many authors proved
generalizations, modifications and applications of it. In [10], Dragomir extended this result to the
larger class of functions of bounded variation, as follows:
Theorem 1.2. Let f : [a, b] → R be a function of bounded variation on [a, b]. Denote by Wb
total variation on [a, b]. Then, for all x ∈ [a, b], one has the inequality
a(f ) its
(1.2)
The constant 1
2 is sharp in the sense that it can not be replaced by a smaller one.
f (x) −
1
b − aZ b
a
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤" 1
2
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
x − a+b
2
b − a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(f ).
# b
_a
The best inequality one can obtain from (1.1) and (1.2) is the midpoint inequality. The corre-
sponding version for the generalised trapezoid inequality was obtained by Cerone, Dragomir, and
Pearce in [7], from which one can derive the trapezoid inequality. Recently, by using a critical
Lemma, Dragomir [12] proved refinement of the generalised trapezoid and Ostrowski inequalities for
functions of bounded variation, the particular cases of which provide refinements of the trapezoid
and mid-point inequalities.
In [14], Guessab and Schmeisser, in the effort of incorporating together the mid-point and trape-
zoid inequalities, proved a companion of Ostrowski's inequality. Motivated by [14], Dragomir [11]
proved some companions of Ostrowski's inequality for functions of bounded variation, as follows:
2010 Mathematics Subject Classification. 26D15, 41A55, 41A80, 65C50.
Key words and phrases. companion of Ostrowski inequality; functions of bounded variation; mid-point inequality,
trapezoid inequality; probability density function.
1
2
W. J. LIU AND Y. SUN
Theorem 1.3. Assume that the function f : [a, b] → R is of bounded variation on [a, b]. Then the
following inequalities
x
4
1
2
4
1
2
≤
−
x − 3a+b
f (x) + f (a + b − x)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − aZ b
b − a"(x − a)
(f ) +(cid:18) a + b
_a
+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
# b
≤" 1
_a
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:0) 3a+b
4 (cid:1) + f(cid:0) a+3b
4 (cid:1)
4 is also the best possible.
2 ]. The constant 1
(f )
−
2
(1.3)
(1.4)
Here the constant 1
a
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
− x(cid:19) a+b−x
_x
(f ) + (x − a)
b
_a+b−x
(f )#
1
b − aZ b
a
≤
1
4
b
(f ).
_a
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
hold for all x ∈ [a, a+b
4 is best possible.
The best inequality one can obtain from (1.3) is the trapezoid type inequality, namely,
For other related results, the reader may be refer to [1, 2, 3, 4, 5, 6, 13, 15, 16, 17, 18, 19, 20, 22,
23, 24, 25, 26, 27, 28] and the references therein.
The main aim of this paper is to establish a refinement of the companion of Ostrowski inequality
(1.3) for functions of bounded variation. Applications for the trapezoid inequality, the mid-point
inequality, the trapezoid type inequality (1.4), and to probability density functions are also given.
2. Main results
To prove our main results, we need the following lemma, which is a slight improvement of [12,
Lemma 2.1].
Lemma 2.1. Let u, f : [a, b] → R. If u is continuous on [a, b] and f is of bounded variation on
[c, b] ⊇ [a, b], then
Z b
a
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
u(t)df (t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(2.1)
1
a
u(t)d t
(f )!
≤Z b
_c
q "Z b
u(t)pd t
≤" b
(f )#
_c
_a
_a
≤ max
t∈[a,b]
u(t)
(f ).
a
b
1
p
(f )!#
if p > 1,
1
p
+
1
q
= 1
Proof. See [12, Lemma 2.1].
The following result may be stated.
(cid:3)
Theorem 2.1. Assume that the function f : [a, b] → R is of bounded variation on [a, b]. Then
(2.2)
f (x) + f (a + b − x)
2
−
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
b − aZ b
a
≤ Q(x)
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
A REFINEMENT OF THE COMPANION OF OSTROWSKI INEQUALITY
3
for all x ∈ [a, a+b
2 ], where
a+b
2
a
(f ) +Z
− x(cid:19) a+b−x
_x
sgn(x − t)
(f ) + (x − a)
a+b−t
_t
(f )dt#
(f )#
_a+b−x
b
1
p
(f )
a+b−t
− (x − a)p(cid:21) a+b−x
− x(cid:19)p
_x
(f )dt)
_t
(f ) +(cid:18) a + b
_a
− x(cid:19)p a+b−x
_x
2
x
(f ) + (x − a)p
1
p
b
_a+b−x
(f ))
(2.3)
We also have
Q(x) ≤
1
x
4
2
1
4
4
1
≤
(f ).
Q(x) :=
x − 3a+b
− x(cid:19) a+b−x
b − a"2(cid:18) 3a + b
_x
b − a"(x − a)
(f ) +(cid:18) a + b
_a
+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤" 1
# b
_a
q ((cid:20)(cid:18) a + b
b − a" b
(f )#
_a
+pZ
b − a" b
(f )#
q ((x − a)p
_a
+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
# b
≤" 1
_a
rp(x, t)sgn(x − t)
x − 3a+b
q = 1 and rp : [a, b]2 → R with
p + 1
(f ),
≤
a+b
1
1
1
4
2
a
2
(2.4)
4
where p > 1, 1
(2.5)
rp(x, t) :=
(t − a)p−1,
t ∈ [a, x],
(cid:18) a + b
2
− x(cid:19)p−1
,
t ∈(cid:18)x,
a + b
2 (cid:21) .
Proof. Define the kernel K(x, t) by
(2.6)
t − a,
t ∈ [a, x],
t −
a + b
2
t − b,
,
t ∈ (x, a + b − x],
t ∈ (a + b − x, b],
K(x, t) :=
for all x ∈ [a, a+b
(see [11])
2 ]. Using the integration by parts formula for Riemann-Stieltjes integrals, we obtain
(2.7)
1
b − aZ b
a
K(x, t)df (t) =
f (x) + f (a + b − x)
2
−
1
b − aZ b
a
f (t)dt.
4
W. J. LIU AND Y. SUN
Now, if f is of bounded variation on [a, b], then on taking the modulus, applying the first inequality
in (2.1) and making a substitution of the form t = a + b − s, we deduce
2
x
x
x
a
a
a
a
a+b
2
a+b
t −
a+b−x
a + b
a + b
a + b
a + b
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
K(x, t)df (t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Z b
2 (cid:19) df (t)(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)
Z a+b−x
Z b
Z x
≤(cid:12)(cid:12)(cid:12)(cid:12)
(t − a)df (t)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)t −
d t
(f )! +Z b
t − ad t
(f )! +Z a+b−x
≤Z x
2 (cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
_a
_a
2 (cid:19) d t
(f )!
(t − a)d t
(f )! −Z
=Z x
(cid:18)t −
_a
_a
2 (cid:19) d t
(b − t)d t
(f )! +Z b
(f )!
+Z a+b−x
(cid:18)t −
_a
_a
a t
(f )! dt +(cid:18)x −
2 (cid:19) x
(f ) −Z x
(f ) +Z
_a
_a
_a
− x(cid:19) a+b−x
t
(f )! dt − (x − a)
(f ) −Z a+b−x
+(cid:18) a + b
_a
_a
− x(cid:19) a+b−x
(f )dt −Z b
=2(cid:18) 3a + b
(f ) −Z
_a
_x
− x(cid:19) a+b−x
=2(cid:18) 3a + b
(f ) −Z
(f )dt +Z
_x
_a
− x(cid:19) a+b−x
=2(cid:18) 3a + b
(f ) +Z
_t
_x
sgn(x − t)
sgn(x − t)
sgn(x − t)
=(x − a)
(f )dt
a + b
a+b−x
a+b−t
a+b
a+b
a+b
a+b
2
a+b
2
a+b
2
a+b
2
t
t
2
2
a
a
2
4
4
4
x
x
a
a
(t − b)df (t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
t − bd t
(f )!
_a
a+b−x
2
t
(f )! dt
_a
(f ) +Z b
_a
a+b−x
sgn(a + b − x − t)
a+b−x t
(f )! dt
_a
_a
(f )dt
t
a+b−s
sgn(x − s)
(f )ds
_a
=(b − a)Q(x),
for all x ∈ [a, a+b
Now, since W·
2 ], and the inequality (2.2) is proved.
a is monotonic nondecreasing on [a, b], then
a+b
2
a t
(f )! dt ≥ 0,
Z x
_a
t
(f )! dt ≥(cid:18) a + b
_a
Z
− x(cid:19)
2
x
a+b
2
_a
2
2
(f ),
t
(f )! dt ≤(cid:18) a + b
− x(cid:19)
_a
a+b−x t
(f )! dt ≤ (x − a)
Z b
_a
_a
(f ) and
a+b
(f ),
b
_a
Z a+b−x
a+b
2
for all x ∈ [a, a+b
2 ], which gives
− x(cid:19) a+b−x
(b − a)Q(x) ≤2(cid:18) 3a + b
_x
(f ) +(cid:18) a + b
_a
=(x − a)
4
2
x
(f ) + (x − a)
− x(cid:19) a+b−x
_x
b
(f )
_a
(f ) + (x − a)
b
_a+b−x
(f )
and the inequality (2.3) is proved.
A REFINEMENT OF THE COMPANION OF OSTROWSKI INEQUALITY
5
Utilising the second part of the second inequality in (2.1) and Holder's inequality, we deduce that
(b − a)Q(x)
1
p
1
1
a
1
p
≤" x
(f )#
_a
+" b
_a+b−x
≤" b
(f )#
_a
+Z b
+"a+b−x
_x
(f )!#
(f )! +Z a+b−x
q "Z x
t − apd t
(f )!#
_a
q "Z b
t − bpd t
(f )#
_a
q "Z x
t − apd t
_a
t − bpd t
(f )!#
_a
a+b−x
a+b−x
(cid:12)(cid:12)(cid:12)(cid:12)
1
p
x
a
.
1
(2.8)
Now, observe that
(f )#
1
q "Z a+b−x
x
1
p
p
d t
(f )!#
_a
t −
(cid:12)(cid:12)(cid:12)(cid:12)
a + b
2 (cid:12)(cid:12)(cid:12)(cid:12)
t −
d t
_a
(f )!
p
a + b
2 (cid:12)(cid:12)(cid:12)(cid:12)
t − bpd t
_a
(f )!
a+b−x
(b − t)p−1
(f )dt
t
_a
a+b−x
p
d t
(f )! +Z b
_a
d t
(f )!
_a
(b − t)pd t
(f )!
_a
− x(cid:19)p x
_a
− x(cid:19)p a+b−x
_a
(f ) + pZ b
a+b−x
(f )
(f )
2
_a
2
t
a
2
2
2
x
2
x
x
x
a
a
a+b
a+b
a+b
t −
a+b−x
a + b
a + b
=(x − a)p
(t − a)p−1
t − apd t
(f )! +Z a+b−x
R(x) :=Z x
(cid:12)(cid:12)(cid:12)(cid:12)
2 (cid:12)(cid:12)(cid:12)(cid:12)
_a
(t − a)pd t
(f )! +Z
− t(cid:19)p
=Z x
(cid:18) a + b
_a
d t
(f )! +Z b
2 (cid:19)p
+Z a+b−x
(cid:18)t −
_a
(f ) − pZ x
(f )dt −(cid:18) a + b
_a
_a
− t(cid:19)p−1 t
(cid:18) a + b
(f )dt +(cid:18) a + b
+ pZ
_a
2 (cid:19)p−1 t
− pZ a+b−x
(cid:18)t −
_a
− (x − a)p(cid:21) a+b−x
− x(cid:19)p
=(cid:20)(cid:18) a + b
_x
− pZ x
(f )dt + pZ
_a
− s(cid:19)p−1 a+b−s
− pZ
_a
− (x − a)p(cid:21) a+b−x
=(cid:20)(cid:18) a + b
_x
+ pZ
_a
(cid:18) a + b
(f )ds + pZ x
(f ) − pZ
(cid:18) a + b
− x(cid:19)p
(f )dt − (x − a)p
rp(x, s)sgn(x − s)
(t − a)p−1
(f )ds
a + b
a+b−s
(f )
a+b
a+b
a+b
2
a+b
2
a+b
2
x
2
2
2
a
a
2
2
x
2
2
a
a
t
− t(cid:19)p−1 t
_a
(f )dt
(s − a)p−1
(f )ds
a+b−s
_a
rp(x, t)sgn(x − t)
t
(f )dt
_a
6
W. J. LIU AND Y. SUN
=(cid:20)(cid:18) a + b
2
− x(cid:19)p
− (x − a)p(cid:21) a+b−x
_x
(f ) + pZ
a
a+b
2
rp(x, t)sgn(x − t)
a+b−t
(f )dt,
_t
where rp is given in (2.5). Utilising (2.8), we deduce the first part of (2.4).
a is monotonic nondecreasing on [a, b], we have
Since W·
2
R(x) ≤(cid:20)(cid:18) a + b
− p"Z
=(cid:20)(cid:18) a + b
2
x
=(x − a)p
which proves (2.4).
a+b
2
a+b
2
2
x
2
a+b
− x(cid:19)p
(cid:18) a + b
− x(cid:19)p
(f ) +(cid:18) a + b
_a
− (x − a)p(cid:21) a+b−x
(f ) + p"Z
_x
ds#
− s(cid:19)p−1
(f ) + p(cid:20)Z x
_a
− (x − a)p(cid:21) a+b−x
_x
− x(cid:19)p a+b−x
_x
(f ) + (x − a)p
2
x
a
(f ) + (x − a)p
(f )
a+b
2
_a
dt#
(f )
2
− t(cid:19)p−1
(cid:18) a + b
(s − a)p−1 ds(cid:21) b
_a
_a
(f )
b
b
(f ),
_a+b−x
(cid:3)
Corollary 2.1. Under the assumptions of Theorem 2.1 with x = 3a+b
type inequalities
4
, we have the refined trapezoid
≤
≤
≤
(2.9)
2
a
4
1
2
a+b
−
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f(cid:0) 3a+b
4 (cid:1) + f(cid:0) a+3b
4 (cid:1)
sgn(cid:18) 3a + b
b − aZ
b − a" b
(f )#
q "pZ
_a
_a
(f ),
1
4
1
a
b
2
1
a+b
a
1
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − aZ b
− t(cid:19) a+b−t
_t
rp(x, t)sgn(cid:18) 3a + b
(f )dt
4
1
p
− t(cid:19) a+b−t
_t
(f )dt#
where rp is given in (2.5).
Corollary 2.2. Under the assumptions of Theorem 2.1 with x = a, we have the refined trapezoid
inequalities
a
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
a+b−t
(f )! dt
_t
2 (cid:19)p b
(f ) − pZ
_a
2
b
a
1
2
1
a+b
−
1
2
(f ) −
b − aZ b
f (a) + f (b)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − aZ
_a
b − a" b
(f )#
q "(cid:18) b − a
_a
_a
(f ).
1
2
1
b
1
a
≤
≤
≤
(2.10)
a+b
2
(cid:18) a + b
2
− t(cid:19)p−1 a+b−t
_t
(f )dt#
1
p
A REFINEMENT OF THE COMPANION OF OSTROWSKI INEQUALITY
7
Corollary 2.3. Under the assumptions of Theorem 2.1 with x = a+b
inequalities
2 , we have the refined midpoint
≤
≤
≤
(2.11)
2
a
a
1
1
a+b
2 (cid:19) −
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − aZ b
f(cid:18) a + b
a+b−t
(f )! dt
b − aZ
_t
b − a" b
(f )#
q "pZ
_a
_a
(f ).
1
2
a+b
1
a
b
1
2
(t − a)p−1
1
p
(f )dt#
a+b−t
_t
Remark 1. The inequalities (2.2)-(2.4) provide a refinement of the companion of Ostrowski inequal-
ity (1.3), and the inequalities (2.10), (2.11) and (2.9) are refinements of the trapezoid inequality, the
mid-point inequality and the trapezoid type inequality (1.4), respectively, which were obtained in [8],
[9] and [11], respectively.
A new inequality of Ostrowski's type may be stated as follows:
Corollary 2.4. Let f be as in Theorem 2.1. Additionally, if f is symmetric about the line x = a+b
2 ,
i.e., f (a + b − x) = f (x), then for all x ∈ [a, a+b
2 ] we have
(2.12)
where Q(x) satisfies (2.3) and (2.4).
≤ Q(x),
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Remark 2. Under the assumptions of Corollary 2.4 with x = a, we have
a
1
1
f (a) −
f (x) −
b − aZ b
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
f (t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − aZ b
a+b−t
(f )! dt
b − aZ
_a
_t
q "(cid:18) b − a
b − a" b
(f )#
2 (cid:19)p b
(f ) − pZ
_a
_a
_a
(f ) −
(f ).
1
2
1
2
a+b
1
1
a
a
b
b
2
1
≤
≤
≤
(2.13)
a+b
2
(cid:18) a + b
2
− t(cid:19)p−1 a+b−t
_t
(f )dt#
1
p
a
3. Application to probability density functions
Now, let X be a random variable taking values in the finite interval [a, b], with the probability
density function f : [a, b] → [0, 1] and with the cumulative distribution function
The following results hold:
F (x) = P r(X ≤ x) =Z x
a
f (t)dt.
8
W. J. LIU AND Y. SUN
Theorem 3.1. With the assumptions of Theorem 2.1, we have
for all x ∈ [a, a+b
2 ], where E(X) is the expectation of X and
1
2
b − a
b − E(X)
[F (x) + F (a + b − x)] −
(cid:12)(cid:12)(cid:12)(cid:12)
− x(cid:19) [F (a + b − x) − F (x)] +Z
− x(cid:19) [F (a + b − x) − F (x)] + (x − a)(cid:21)
a+b
a
2
≤ T (x)
(cid:12)(cid:12)(cid:12)(cid:12)
sgn(x − t)[F (a + b − t) − F (t)]dt#
(3.1)
T (x) :=
≤
(3.2)
≤
We also have
.
4
1
1
1
4
1
4
T (x) ≤
4
x − 3a+b
b − a"2(cid:18) 3a + b
b − a(cid:20)2(cid:18) 3a + b
+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − a(cid:26)(cid:20)(cid:18) a + b
+pZ
b − a(cid:26)(cid:20)(cid:18) a + b
+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
b − a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2
x − 3a+b
p + 1
1
4
≤
≤
a+b
1
2
4
a
2
,
(3.3)
where p > 1, 1
q = 1 and rp is given in (2.5).
− x(cid:19)p
− (x − a)p(cid:21) [F (a + b − x) − F (x)]
rp(x, t)sgn(x − t)[F (a + b − t) − F (t)]dt)
1
p
− x(cid:19)p
− (x − a)p(cid:21) [F (a + b − x) − F (x)] + (x − a)p(cid:27)
1
p
Proof. By (2.2)-(2.4) on choosing f = F and taking into account
we obtain (3.1)-(3.3).
E(X) =Z b
a
tdF (t) = b −Z b
a
F (t)dt,
Corollary 3.1. Under the assumptions of Theorem 3.1 with x = 3a+b
4
, we have
(cid:3)
2
1
1
a+b
(cid:12)(cid:12)(cid:12)(cid:12)
2(cid:20)F (cid:18) 3a + b
b − aZ
b − a"pZ
4 (cid:19) + F (cid:18) a + 3b
sgn(cid:18) 3a + b
rp(x, t)sgn(cid:18) 3a + b
a+b
4
1
4
a
a
2
b − a
b − E(X)
4 (cid:19)(cid:21) −
− t(cid:19) [F (a + b − t) − F (t)]dt
(cid:12)(cid:12)(cid:12)(cid:12)
− t(cid:19) [F (a + b − t) − F (t)]dt#
1
4
,
1
p
≤
≤
≤
(3.4)
where rp is given in (2.5).
Remark 3. The inequalities (3.1)-(3.4) provide refinements of the inequalities given in [11, Theorem
5] and [11, Corollary 4].
Acknowledgments. This work was partly supported by the National Natural Science Foundation
of China (Grant No. 40975002) and the Natural Science Foundation of the Jiangsu Higher Education
Institutions (Grant No. 09KJB110005).
A REFINEMENT OF THE COMPANION OF OSTROWSKI INEQUALITY
9
References
[1] M. W. Alomari, A companion of Ostrowski's inequality with applications, Transylv. J. Math. Mech. 3 (2011),
no. 1, 9 -- 14.
[2] M. W. Alomari, A companion of Ostrowski's inequality for mappings whose first derivatives are bounded and
applications in numerical integration, RGMIA Res. Rep. Coll., 14 (2011) article 57.
[3] M. W. Alomari, A generalization of companion inequality of Ostrowski's type for mappings whose first derivatives
are bounded and applications in numerical integration, RGMIA Res. Rep. Coll., 14 (2011) article 58.
[4] G. A. Anastassiou, Ostrowski type inequalities, Proc. Amer. Math. Soc. 123 (1995), no. 12, 3775 -- 3781.
[5] G. A. Anastassiou and J. A. Goldstein, Higher order Ostrowski type inequalities over Euclidean domains, J.
Math. Anal. Appl. 337 (2008), no. 2, 962 -- 968.
[6] N. S. Barnett, S. S. Dragomir and I. Gomm, A companion for the Ostrowski and the generalised trapezoid
inequalities, Math. Comput. Modelling 50 (2009), no. 1-2, 179 -- 187.
[7] P. Cerone, S. S. Dragomir and C. E. M. Pearce, A generalized trapezoid inequality for functions of bounded
variation, Turkish J. Math. 24 (2000), no. 2, 147 -- 163.
[8] S. S. Dragomir, The Ostrowski integral inequality for mappings of bounded variation, Bull. Austral. Math. Soc.
60 (1999), no. 3, 495 -- 508.
[9] S. S. Dragomir, On the midpoint quadrature formula for mappings with bounded variation and applications,
Kragujevac J. Math. 22 (2000), 13 -- 19.
[10] S. S. Dragomir, On the Ostrowski's integral inequality for mappings with bounded variation and applications,
Math. Inequal. Appl. 4 (2001), no. 1, 59 -- 66.
[11] S. S. Dragomir, A companions of Ostrowski's inequality for functions of bounded variation and applications,
RGMIA Res. Rep. Coll., 5 (2002), Supp., article 28.
[12] S. S. Dragomir, Refinements of the generalised trapezoid and Ostrowski inequalities for functions of bounded
variation, Arch. Math. (Basel) 91 (2008), no. 5, 450 -- 460.
[13] J. Duoandikoetxea, A unified approach to several inequalities involving functions and derivatives, Czechoslovak
Math. J. 51(126) (2001), no. 2, 363 -- 376.
[14] A. Guessab and G. Schmeisser, Sharp integral inequalities of the Hermite-Hadamard type, J. Approx. Theory
115 (2002), no. 2, 260 -- 288.
[15] V. N. Huy and Q. -A. Ngo, New bounds for the Ostrowski-like type inequalities, Bull. Korean Math. Soc. 48
(2011), no. 1, 95 -- 104.
[16] W. J. Liu, Several error inequalities for a quadrature formula with a parameter and applications, Comput. Math.
Appl. 56 (2008), no. 7, 1766 -- 1772.
[17] W. J. Liu, Some weighted integral inequalities with a parameter and applications, Acta Appl. Math. 109 (2010),
no. 2, 389 -- 400.
[18] W. J. Liu, Q. L. Xue and S. F. Wang, New generalization of perturbed Ostrowski type inequalities and applica-
tions, J. Appl. Math. Comput. 32 (2010), no. 1, 157 -- 169.
[19] Z. Liu, Note on a paper by N. Ujevi´c, Appl. Math. Lett. 20 (2007), no. 6, 659 -- 663.
[20] Z. Liu, Some companions of an Ostrowski type inequality and applications, JIPAM. J. Inequal. Pure Appl. Math.
10 (2009), no. 2, Article 52, 12 pp.
[21] A. Ostrowski, Uber die Absolutabweichung einer differentiierbaren Funktion von ihrem Integralmittelwert, Com-
ment. Math. Helv. 10 (1937), no. 1, 226 -- 227.
[22] M. Z. Sarikaya, On the Ostrowski type integral inequality, Acta Math. Univ. Comenian. (N.S.) 79 (2010), no. 1,
129 -- 134.
[23] M. Z. Sarikaya, New weighted Ostrowski and Cebysev type inequalities on time scales, Comput. Math. Appl. 60
(2010), no. 5, 1510 -- 1514.
[24] E. Set and M. Z. Sarıkaya, On the generalization of Ostrowski and Gruss type discrete inequalities, Comput.
Math. Appl. 62 (2011), no. 1, 455 -- 461.
[25] K.-L. Tseng, S.-R. Hwang and S. S. Dragomir, Generalizations of weighted Ostrowski type inequalities for map-
pings of bounded variation and their applications, Comput. Math. Appl. 55 (2008), no. 8, 1785 -- 1793.
[26] K.-L. Tseng, S.-R. Hwang, G.-S. Yang and Y.-M. Chou, Weighted Ostrowski integral inequality for mappings of
bounded variation, Taiwanese J. Math. 15 (2011), no. 2, 573 -- 585.
[27] N. Ujevi´c, New bounds for the first inequality of Ostrowski-Gruss type and applications, Comput. Math. Appl.
46 (2003), no. 2-3, 421 -- 427.
[28] S. W. Vong, A note on some Ostrowski-like type inequalities, Comput. Math. Appl. 62 (2011), no. 1, 532 -- 535.
10
W. J. LIU AND Y. SUN
(W. J. Liu) College of Mathematics and Statistics, Nanjing University of Information Science and
Technology, Nanjing 210044, China
E-mail address: [email protected]
(Y. Sun) College of Mathematics and Statistics, Nanjing University of Information Science and Tech-
nology, Nanjing 210044, China
|
1812.09757 | 1 | 1812 | 2018-12-23T19:14:28 | Partial Classification of Polynomials and an Orthonormal Basis Construction on the Associated Basin of Attraction | [
"math.FA",
"math.CV"
] | In the paper "Infinite product representations for kernels and iterations of functions", the authors associate certain Fatou subsets with reproducing kernel Hilbert spaces. They also present a method for constructing an orthonormal basis for said Hilbert space, but the method depends on the polynomial of the given Fatou set. We provide a partial classification of those polynomials the method applies to. | math.FA | math |
Partial Classification of Polynomials and an Orthonormal Basis
Construction on the Associated Basin of Attraction
James Tipton
March 17, 2021
Abstract
In the paper Infinite product representations for kernels and iterations of functions, the authors
associate certain Fatou subsets with reproducing kernel Hilbert spaces. They also present a method for
constructing an orthonormal basis for said Hilbert space, but the method depends on the polynomial
of the given Fatou set. We provide a partial classification of those polynomials the method applies to.
1
Introduction
Complex Dynamics
(c) < 1. The point z0 is called
Recall that R : C → C has an attracting fixed point at z0 ∈ C if R
an attracting fixed point because all points within a certain neighborhood of z0 are "attracted" to z0
under repeated iteration of R. The nth iterate of R is denoted by
R◦n(z) = R ◦ R ◦ ··· ◦ R(z)
(cid:48)
(cid:123)(cid:122)
n times
(cid:124)
(cid:125)
The basin of attraction of R at the attracting fixed point z0 is the following subset of C:
BR,z0 = {z ∈ C :
n→∞ R◦n(z) = z0}
lim
For many polynomials with an attracting fixed point, the basin of attraction is a fractal.
Reproducing Kernel Hilbert Spaces
A reproducing kernel Hilbert space (RKHS) on C is a Hilbert space of functions on C in which
every linear evaluation functional is bounded. Uniquely associated to each RKHS is a kernel function
K : C × C → C with the reproducing property:
(cid:104)f (z), K(z, w)(cid:105)H = f (w)
Since a RKHS is, in particular, a Hilbert space, it must have an orthonormal basis (ONB). Although
ONBs are guaranteed to exist, explicitly constructing an ONB is a much harder task.
1
Kernel Functions on Basins of Attraction
1 INTRODUCTION
If R satisfies sufficient conditions, then one may construct a kernel function, represented as an infinite
product, on a subset of BR,z0. See [1] for the general result. In particular, if R is a polynomial and
z0 = 0, then the map K : C × C → C defined by
∞(cid:89)
(cid:16)
(cid:17)
K(z, w) =
1 + R◦n(z)R◦n(w)
(1)
is a kernel function on all of BR,0 [3]. The infinite product involves iterates of the map R and the
map 1 + zw, which is a kernel function on C. The kernel function 1 + zw can be used to construct an
ONB under certain circumstances.
n=0
The ONB Construction
First we take a moment to recall multi-index notation. Suppose J is an index set, then
J∞ = {v : v ∈ J N for some N = 1, 2, . . .}
Denote the RKHS associated to the previous kernel function (1) by H. The constant function 1(z) = 1
plays a crucial role in the construction, and in fact belongs to H. Consider a family of operators on
H, {Si : H → H}. For each v = (v1, v2, . . . , vN ) ∈ J N , define bv : C → C
bv(z) = (Sv1Sv2 ··· SvN 1)(z)
The next theorem, due to the authors of [1], gives sufficient conditions for the functions bv to form an
ONB.
Theorem 1. If a family of operators {Si : H → H}N
i=1 satisfies the Cuntz relations:
N(cid:88)
S∗
i Sj = δijI,
SiS∗
i = I
then B = {bv : v ∈ J∞} is an ONB for H.
i=1
In our particular set-up, the family we wish to consider is F = {S1, S2} where S1f (z) = f (R(z))
and S2f (z) = zf (R(z)). This family of operators can be shown to satisfy the Cuntz relations when
certain conditions are met, which we discuss now.
The Dagger Conditions
The family F of interest depends on the map R that is chosen. It can be shown that if R satisfies for
all z ∈ BR,0,
M (z) < ∞
where M (z) is the number of solutions to R(ζ) = z, counting multiplicity, and either
(cid:88)
(cid:88)
R(ζ)=z
1
M (z)
1
M (z)
R(ζ)=z
2
ei(ζ)ej(ζ) = δij,
∀i, j ∈ J
or
ei(ζ)ej(ζ) = δij,
∀i, j ∈ J,
(†)
(‡)
(cid:88)
(cid:88)
(cid:88)
R(ζ)=z
R(ζ)=z
R(ζ)=z
1 = M (z)
ζ = 0 =
(cid:88)
ζ
R(ζ)=z
ζ2 = M (z),
(cid:88)
(cid:88)
(cid:88)
R(ζ)=z
R(ζ)=z
1 = M (z)
ζ = 0
ζ 2 = M (z)
(†1)
(†2)
(†3)
(‡1)
(‡2)
(‡3)
2 PARTIAL CLASSIFICATION OF THE DAGGER CONDITIONS
then F satisfies the Cuntz relations [1]. The functions, ei(z), are taken from any ONB for the RKHS
associated to the underlying kernel function of the infinite product kernel function. For the family F
that we are interested in, we have that e1(z) = 1 and e2(z) = z; this comes from the underlying kernel
function 1 + zw mentioned earlier. For ease of exposition we will refer to the above conditions as the
dagger conditions. A natural question is when does the map R satisfy either of the above conditions?
In the context of the underlying kernel function 1 + zw, the † condition becomes
and the ‡ condition becomes
We examine which polynomials R satisfy the dagger conditions and offer a classification for R to
satisfy the ‡ condition.
R(ζ)=z
2 Partial Classification of the Dagger Conditions
The purpose of the dagger conditions is to construct an ONB for the RKHS corresponding to the
kernel function on BR,0. Thus our interest lies only with those polynomials with an attracting fixed
point at 0, even though the dagger conditions do not require R to have such a property. The first two
cases of either dagger condition is quite easily characterized.
Proposition 2. If P (z) is a degree n polynomial with an attracting fixed point at 0 then the following
hold:
a) P (z) satisfies †1 and ‡1.
b) P (z) satisfies †2 and ‡2 if and only if an−1 = 0.
Proof.
a) Since BP,0 is completely invariant with respect to P , we know that if P (ζ) = z for some z ∈ BP,0,
then we must have that ζ ∈ Ω. By the Fundamental Theorem of Algebra there are n solutions
to P (ζ) = z, counting multiplicity. Thus (cid:80)
b) We have that P (z) satisfies †2 and ‡2 if and only if (cid:80)
ζ = 0 = (cid:80)
ζ which by Vieta's
1 = n = M (z).
P (ζ)=z
formulas is equivalent to an−1 = 0.
P (ζ)=z
P (ζ)=z
3
2 PARTIAL CLASSIFICATION OF THE DAGGER CONDITIONS
Note: The conditions †2 and ‡2 are equivalent to each other since (cid:80)
ζ = (cid:80)
ζ
The next proposition will complete our characterization of ‡.
P (ζ)=z
P (ζ)=z
Proposition 3. A polynomial P of degree n ≥ 3, with an attracting fixed point at 0, satisfies ‡3 if
and only if an−2 =
−nan
2
Proof. The polynomial P satisfies ‡3 if and only if(cid:88)
which is equivalent to an−2 = −nan
2
, since Vieta's formulas and the Newton-Girard formulas give
ζ 2 = n
R(ζ)=z
(cid:88)
n =
R(ζ)=z
ζ 2 = −2e2 =
−2an−2
an
Combining the last two propositions, we may characterize those polynomials satisfying the ‡ con-
dition:
Theorem 4. Suppose P is a polynomial of degree n ≥ 3 with an attracting fixed point at 0. The
polynomial P satisfies ‡ if and only if P (z) =
Note: Since P has an attracting fixed point at 0, we have also that a1 < 1.
akzk where an−1 = 0, and an−2 = −nan
n(cid:80)
.
2
k=1
Characterizing the †3 condition seems to be more challenging. It is easy to find polynomials which
satisfy †3 at a particular point. But the dagger conditions are required to hold for all z ∈ BP,0. As such
we introduce "partial" conditions, †c and ‡c, each meaning that the corresponding set of equations
hold precisely at the point c ∈ C. The next proposition will help us determine which polynomials
might satisfy †.
Proposition 5. Suppose c ∈ BP,0.
a) If P (z) satisfies ‡c, then P (z) satisfies ‡.
b) If P (z) satisfies both †c and ‡c then the equation P (z) = c has only real solutions.
Proof.
a) If P (z) satisfies ‡c, then we must have that
(cid:88)
1 = n,
(cid:88)
(cid:88)
ζ 2 = n
ζ = 0,
By Vieta's formulas, we have that an−1 = 0. Thus we have that ‡1 and ‡2 hold by an application
of Proposition 1. By an application of the Newton-Girard formulas we have that
P (ζ)=c
P (ζ)=c
P (ζ)=c
(cid:88)
n =
ζ 2 = −2e2 =
−2an−2
an
Applying the same formula again we find that(cid:88)
P (ζ)=c
So P actually satisfies ‡.
P (ζ)=z
4
ζ 2 =
−2an−2
an
= n
2 PARTIAL CLASSIFICATION OF THE DAGGER CONDITIONS
b) Suppose that P (z) satisfies both †c and ‡c, so in particular, we have that
which requires that (cid:80)
R(ζ)=c
ζ 2 =
ζ 2
R(ζ)=c
R(ζ)=c
Im ζ 2 = 0. Thus we have that
(cid:88)
(cid:88)
(cid:88)
(cid:88)
Re ζ 2 =
ζ 2
If ζ 2 is not real, then ζ 2 > Re ζ 2, so the above equality holds only if all solutions to P (ζ) = c
are real.
R(ζ)=c
R(ζ)=c
So if P (z) has a non-real zero, then P (z) cannot satisfy both † and ‡. The only interesting polyno-
mials which might satisfy † are those P (z) which have the property that ‡c is not satisfied for any c ∈ Ω.
Such a polynomial requires the property that P (z) + c has at least one non-real zero for all c ∈ Ω.
An example of a polynomial with this property is any cubic of the form P (z) = az3 + bz + c with b a
scalar multiple of a. We show now that no polynomial can satisfy both † and ‡. This fact can be used
in turn to show that the aforementioned property is a necessary condition for a polynomial to satisfy †.
Proposition 6. If P is a polynomial with an attracting fixed point at 0, then P cannot satisfy both †
and ‡.
Proof. Suppose to the contrary that P satisfies both † and ‡. By Proposition 5, P (z) = c has only real
solutions, for any c ∈ BP,0. Thus the inverse image of BP,0 under P must be a subset of R. However,
BP,0 is an open set, and P is a continuous map, so that the inverse image of BP,0 under P must be
an open set. But no subset of R is open as a subset of C. Thus it cannot be that P satisfies both †
and ‡.
We can now state a necessary condition for a polynomial to satisfy †.
Proposition 7. If P is a polynomial with an attracting fixed point at 0 that satisfies †, then for all
c ∈ BR,0, the equation P (z) = c has at least one non-real solution.
Proof. Suppose that P satisfies †, and to the contrary, that P (z) = c has only real solutions. If ζ is
such a solution, then ζ2 = ζ 2. Since P satisfies †3, we have also that, by the previous observation,
that P satisfies ‡3. By Proposition 2, P also satisfies ‡1 and ‡2, so that P must satisfy ‡. But P
cannot satisfy both † and ‡, so that P (z) = c must have at least one non-real solution.
Another approach to showing that a polynomial satisfies one of the dagger conditions is to write it
as a product of two polynomials, each of which satisfy the same dagger condition. This is equivalent
to determining whether the product of two polynomials, both satisfying the same dagger condition,
will satisfy a dagger condition. This works quite well for the ‡ condition.
Proposition 8. If R(z) and Q(z) satisfy ‡, then R(z)Q(z) satisfies ‡.
Proof. Suppose that R and Q both satisfy ‡ and let S(z) = R(z)Q(z). Let ai denote the coefficients of
R, bi denote the coefficients of Q, and suppose deg R = r, deg Q = q so that deg S = r + q. Theorem 4
, and ar−1 = 0 = bq−1. If ci denotes the coefficients of S, then
tells us that ar−2 =
, bq−2 =
−rar
2
−rbq
2
−(r + q)cr+q
2
. By Theorem 4, the polynomial S
we have that cr+q = arbq, cr+q−1 = 0, and cr+q−2 =
satisfies the ‡ condition.
5
3 EXAMPLES
Figure 1: The domain of K(z,w): BR,0
3 Examples
Here we present an example of a polynomial satisfying ‡, and an example of a polynomial satisfying
†0.
Example 9. Consider the polynomial R(z) = iz4 − 2iz2 − 1+i
point of R since R(0) = 0 and R
2 z. We see that 0 is an attracting fixed
< 1. So we have that the map
(0) =
√
(cid:48)
2
2
∞(cid:89)
(cid:16)
K(z, w) =
1 + R◦n(z)R◦n(w)
(cid:17)
(cid:17)
is a kernel function on BR,0. The polynomial R has coefficients: a4 = i, a2 = −2i, a1 = − 1+i
a3 = a0 = 0. Since a3 = 0 and a2 = −4a4
2
that the operators S1 and S2, defined by:
2 , and
, Theorem 4 tells us that R satisfies ‡. This in turn shows
n=0
S1f (z) = f (R(z))
and
S2f (z) = zf (R(z))
satisfy the Cuntz relations. So we may apply Theorem 1 to conclude that the functions bv(z) form an
ONB for the Hilbert space associated to K. Recall that
where v ∈ J∞. The first few basis elements are:
bv(z) = (Sv1Sv2 ··· SvN 1)(z)
1, z, R(z), zR(z), R◦2(z), zR◦2(z), R(z)R◦2(z), zR(z)R◦2(z), . . .
So the basis elements may be calculated recursively, but obtaining a general formula appears to require
a general formula for R◦n.
Example 10. Consider the polynomial Q(z) = 1
0. So the map
4 z which also has an attracting fixed point at
2 z3 + 3
K(z, w) =
1 + Q◦n(z)Q◦n(w)
∞(cid:89)
(cid:16)
n=0
is a kernel function on BQ,0. There is a RKHS associated to K, however, we can't use the dagger
conditions to construct an ONB. By Theorem 4, Q doesn't satisfy ‡, in particular, the condition ‡3.
It turns out that Q does satisfy †0; this follows from Proposition 2 and the following observation:
(cid:88)
ζ2 =
(cid:12)(cid:12)(cid:12)0
(cid:12)(cid:12)(cid:12)2
(cid:12)(cid:12)(cid:12)i(cid:112)3/2
(cid:12)(cid:12)(cid:12)2
+
(cid:12)(cid:12)(cid:12) − i(cid:112)3/2
(cid:12)(cid:12)(cid:12)2
+
= 3
Q(ζ)=0
6
5 APPENDIX: NEWTON-GIRARD IDENTITIES AND THE VIETA FORMULA
Figure 2: The domain of K(z,w): BQ,0
It can be shown in a similar fashion that Q satisfies †c for c =
all c ∈ BQ,0: for c = i ∈ BQ,0 and with the aid of WolframAlpha, we have that
√
i
2
2
. However Q doesn't satisfy †c for
(cid:88)
ζ2 > 3
Q(ζ)=i
4 Open Questions
Here we discuss some open questions pertaining to this paper.
1. Classify the † condition. The main issue lies with †3. Since the sum involves the modulus of
the roots, Vieta's formula may not be applied.
2. Find a polynomial that satisfies †3 or show that no polynomial satifies †3. Just having
one example would be a nice starting point; but so would knowing that no examples exist.
3. What if we don't count multiplicity? Much of the theory presented in [1] should still work
if we don't count the multiplicity of the solutions to P (z) = c. The biggest issue with this change
would be in the application of Vieta's formula, since it does use multiplicity.
4. Classify polynomials satisfying †0. This could be another starting point for classifying the
† condition. Understanding when †0 is satisfied could help to understand when †c is satisfied.
5. Generalize the results presented here to other "underlying" kernel functions. There
are other underlying kernel functions one could use other than 1 + zw. However, changing the
underlying kernel function will change the dagger conditions. There are some kernel functions
for which the approach presented here might still work, in particular, kernel functions of the
form 1 + (zw)n, where n is a positive integer.
5 Appendix: Newton-Girard Identities and the Vieta
Formula
Here we take a brief look at the Newton-Girard identities and Vieta's formula; both quintessential
tools in this paper. See [2] for a more in depth historical introduction.
n(cid:80)
ajzj satisfies an (cid:54)= 0. If z1, . . . , zn are the roots
Theorem 11 (Vieta's formula). Suppose P (z) =
(counting multiplicity) of P , then (cid:88)
1≤j1<···<jk≤n
j=0
zi1 ··· zik =
(−1)kan−k
an
7
5 APPENDIX: NEWTON-GIRARD IDENTITIES AND THE VIETA FORMULA
Proof. By assumption we have that
n(cid:88)
n(cid:89)
ajzj = an
(z − zj) = anzn + an
j=0
j=1
equating coefficients yields Vieta's formula.
n(cid:88)
(cid:16)
k=1
(−1)k (cid:88)
1≤j1<···<jk≤n
(cid:17)
zn−k
zi1 ··· zik
The Newton-Girard identities involve symmetric polynomials, so we start with a few definitions
and notational conventions. The kth power sum in n variables is the polynomial pk,n =
zk
i . The
elementary symmetric polynomials in n variables are defined by
e0 = 1,
e1 =
zi,
e2 =
zizj,
. . . ,
en = z1z2 ··· zn,
ek = 0 ∀k > n
n(cid:80)
i=1
Theorem 12 (Newton-Girard identities). Let em denote the mth elementary symmetric poliynomial
in n variables and Pm denote the mth power sum in n variables. We have that
n(cid:88)
i=1
k−1(cid:88)
(cid:16)
l=0
(cid:88)
1≤i<j≤n
(cid:17)
(−1)lelPk−l
+ (−1)kkek = 0 ∀k, n ∈ N
(cid:17)
zn−l =
n(cid:88)
(−1)lelzn−l
l=0
zi1 ··· zil
(cid:17)
(cid:16)
k−1(cid:88)
h
l=0
(cid:33)
Proof. From the proof of Vieta's formula we have
n(cid:89)
(z − zj) = zn +
j=1
l=1
1≤j1<···<jl≤n
n(cid:88)
(cid:16)
k(cid:88)
l=0
(−1)l (cid:88)
(cid:16)
k−1(cid:88)
l=0
where we now think of the zj as free variables. Suppose k = n, set z = zh where h ∈ {1, . . . , k} to
obtain
0 =
elzk−l
h =
(−1)lelzk−l
+ (−1)kek
Now sum the right hand side over h to obtain:
(cid:32)k−1(cid:88)
(cid:16)
(cid:88)
h
l=0
0 =
(cid:17)
(−1)lelzk−l
h
+ (−1)kek
=
(−1)lelPk−l
(cid:17)
+ (−1)kkek
Demonstrating the identity for k = n. The case n < k follows from setting k − n of the zh to 0 and
the case k < n follows from setting n − k of the zh to 0.
The first two Newton-Girard identities are:
1. For k = 1: P1 − e1 = 0
2. For k = 2: P2 − e1P1 + 2e2 = 0
Solving the second formula for P2, we obtain P2 = e1P1 − 2e2. Now let n be the degree of some
polynomial R that satisfies †2 or ‡2, and plug the n roots (counting multiplicity) of R into P2. The
identity becomes P2 = −2e2 since †2 (or ‡2) implies that e1 = P1 = 0. By Vieta's formula, we have
that e2 = an−2
an
. So we may conclude by stating that
(cid:88)
P2 =
ζ 2 =
R(ζ)=0
−2an−2
an
Although we only made use of the first two identities here, the other identities will likely be of use
when the underlying kernel function is changed.
8
REFERENCES
References
REFERENCES
[1] Alpay, D., Jorgensen, P., Lewkowicz, I., & Martziano, I. (2015). Infinite product representations
for kernels and iterations of functions. In Recent advances in inverse scattering, Schur analysis
and stochastic processes (pp. 67-87). Birkhuser, Cham.
[2] Funkhouser, H. G. (1930). A short account of the history of symmetric functions of roots of
equations. The American mathematical monthly, 37(7), 357-365.
[3] Tipton, J. E. (2016). Reproducing kernel Hilbert spaces and complex dynamics. The University of
Iowa.
9
|
0809.1176 | 3 | 0809 | 2018-08-18T17:50:58 | On the connected components of the conjugacy class of projectors on $ \ell_p\oplus\ell_q $ | [
"math.FA",
"math.AT",
"math.OA"
] | We characterize the projectors $ P $ on a Banach space $ E $ having the property of being connected to all the others projectors obtained as a conjugation of $ P $. Using this characterization we show an example of Banach space where the conjugacy class of a projector splits into several path-connected components, and describe the conjugacy classes of projectors onto subspaces of $ \ell_p\oplus\ell_q $ with $ p\neq q $. | math.FA | math |
ON THE CONNECTED COMPONENTS OF THE CONJUGACY
CLASS OF PROJECTORS ON ℓp ⊕ ℓq
DANIELE GARRISI
Abstract. We characterize the projectors P on a Banach space E having the
property of being connected to all the others projectors obtained as a conju-
gation of P . Using this characterization we show an example of Banach space
where the conjugacy class of a projector splits into several path-connected
components, and describe the conjugacy classes of projectors onto subspaces
of ℓp ⊕ ℓq with p 6= q.
1. Introduction
Let A be a Banach algebra and denote by P(A) the space of projectors, that is the
set of the elements p ∈ A such that p2 = p, endowed with the subspace topology.
We denote by G(A) the group of invertible elements of A. Given a topological
space X, we use the notation π0(X) for set of path-connected components of X,
and [x] for the component which contains x.
Given two projectors p and q it is possible to define another relation
pRq : ∃g ∈ G(A) such that gp = qg, Rp := {q pRq}.
If pRq, we also say that p is conjugated to q, and Rp is the conjugacy class of p. In
general, for every Banach algebra, the inclusion [p] ⊆ Rp holds. A proof of this fact
can be found in [7, Proposition 4.2], or [4, Ch. I.4 §6]. Actually, the following fact
holds true: if [p] = [q], there exists a continuous path u : [0, 1] → G(A) such that
(1.1)
u(1)p = qu(1) and u(0) = 1.
In [7, 7.13.1], G. Porta and L. Recht provided an example of Banach algebra and a
projector p such that [p] ( Rp. The algebra A is C(S 3, M (2, C)) and the projector
is defined as follows: if x := (x1, x2, x3, x4) is a point of S 3, then
(1.2)
p(x) := (cid:18)z0z0
z0z1
z1z0
z1z1(cid:19) ,
q(x) := 1 − p(x)
where z0 := x1 + ix2 and z1 := x3 + ix4. Therefore, we wonder whether a similar
example may occur in algebras of bounded operators on a Banach space E, for a
suitable choice of E. That is, whether there are two projectors P and Q in P(L(E))
which are conjugated to each other, but [P ] 6= [Q], for a suitable choice of E. This
happens in the space E := ℓp ⊕ ℓq with 1 ≤ p, q and p 6= q.
In §2, we provide necessary and sufficient conditions to a projector P ∈ P(L(E))
ensuring the equality RP = [P ]. This conditions is expressed in terms of the
connected components of the general linear group of ker(P ), ran(P ) and E. In §3, we
1991 Mathematics Subject Classification. 46H05; 47L05.
Key words and phrases. Projectors; components; conjugation.
This work was supported by INHA UNIVERSITY Research Grant.
1
2
DANIELE GARRISI
address projectors P such that ran(P ) can be obtained as direct sum of a subspace of
ℓp ⊕{0} and a subspace of {0}⊕ℓq. We show that there are two possible behaviours:
either RP is path-connected or RP has infinitely many connected components.
2. A criterion to establish whether RP = [P ]
Let P be a projector of L(E). We set X := ran(P ) and Y := ker(P ).
Proposition 2.1. Given a Banach space E, RP = [P ] if and only if
ϕ∗ : π0(GL(X) × GL(Y )) → π0(GL(E)), ϕ(A, B) := A ⊕ B
is surjective.
Proof. Suppose that RP ⊆ [P ]. Given T ∈ GL(E), we set
(2.1)
Q := T P T −1.
Since Q is a projector, it is path-connected to P . Then, from (1.1), there exists a
path of invertible operators U such that
SP = QS, S := U (1).
(2.2)
From (2.1) and (2.2), it follows that S−1T commutes with P or, equivalently, S−1T
is in the image of ϕ. Hence S−1T ∈ Im(ϕ) while, from the definition of U (t), we have
[S−1T ] = [T ]. Then, [T ] ∈ Im(ϕ∗). Conversely, suppose that ϕ∗ is surjective and
consider T P T −1, a conjugation of P . Let (A, B) be a pair such that [A⊕B] = [T −1].
Since A ⊕ B commutes with P we have
(T (A ⊕ B))P (T (A ⊕ B))−1 = T P T −1,
while [T (A ⊕ B)] = [idE]. Then [T P T −1] = [P ].
(cid:3)
Remark 2.1. Actually, we proved that T P T −1 belongs to the same connected
component of P if and only if [T ] belongs to the image of ϕ∗.
Example 2.1. When GL(E) is path-connected, RP = [P ].
In fact, given U ∈
GL(E) and a path U (t) which connects U to idE, the path U (t)P U (t)−1 connects
P to U P U −1.
Example 2.2. Considering cases where the linear group is not connected does not
automatically imply the existence of a projector where [P ] ( RP . For instance,
if E = Rn, the linear group is GL(n, R) which consists of two connected com-
ponents characterized by the sign of the determinant; given T ∈ GL(n, R) such
that det(T ) < 0 it is possible to choose A ∈ GL(X) and B ∈ GL(Y ) such that
det(A) · det(B) < 0. Therefore ϕ∗([A ⊕ B]) = [T ], and ϕ∗ is surjective.
3. The conjugacy class in ℓp ⊕ ℓq
We will use a construction of A. Douady, [2], devised with the purpose of showing
the existence of infinite-dimensional Banach spaces whose linear group is not path-
connected.
Let E = F ⊕ G be Banach spaces such that F and G are isomorphic to their closed
subspaces of co-dimension one and such that L(F, G) = LSS(F, G) (the subspace
of strictly singular operators). We set
PF (x, y) := (x, 0), PG(x, y) = (0, y).
Given an operator A ∈ L(F ), we use the notation ind(A) for the Fredholm index,
whenever appropriate.
ON THE CONNECTED COMPONENTS OF THE CONJUGACY CLASS OF PROJECTORS 3
Lemma 3.1 (From [2, Proposition 1]). There exists a continuous, surjective group
homomorphism j : GL(E) → Z such that j(T ) = ind(PF T PF ). Moreover,
(a) ind(PF T PF ) + ind(PGT PG) = 0
(b) if GL(F ) and GL(G) are path-connected, j is injective.
For the proof of the fact that j is a group homomorphism and that is surjective, we
refer to Lemma 1 and Lemma 2 of [5]. An explicit invertible operator V such that
j(V ) = 1 is the example provided in [5, p. 6]. Let F1 and G1 be two closed subspaces
of co-dimension one in F and G, respectively and σ : F1 → F and τ : G1 → G be
two isomorphisms. Let also v and w be such that
We define
(3.1)
F = F1 ⊕ hvi, G = G1 ⊕ hwi.
V (x + sv, y) = (σ(x), sw + τ −1(y)).
where ind(PF V PF ) = 1. For the injectivity of the map j when GL(F ) and GL(G)
are path connected, we refer to [2] and [3].
Hereafter, we set F := ℓp and G := ℓq. Let P be such that ran(P ) = X1 ⊕ X2,
where X1 ⊆ ℓp and X2 ⊆ ℓq are two closed and complemented subspaces. Let Y1
and Y2 be closed subspaces such that
X1 ⊕ Y1 = ℓp, X2 ⊕ Y2 = ℓq.
Proposition 3.1. If both Y1 and Y2, or both X1 and X2 have infinite dimension,
then [P ] = RP . That is, any projector conjugated to P is also path-connected to
it. Otherwise, π0(RP ) ≈ Z, that is in the conjugacy class of P there are infinitely
many connected components.
Proof. From [1, Theorem 2.24, p. 35], closed infinite-dimensional and comple-
mented subspaces of ℓp and ℓq are isomorphic to ℓp and ℓq, respectively. Moreover,
from [6], the linear groups of ℓp and ℓq are contractible, and LSS(ℓp, ℓq) = L(ℓp, ℓq),
from [1, Pitt's Theorem, p. 32] and [1, Theorem 2.1.9, p. 33]. Therefore, the
decomposition
(X1 ⊕ Y1)M(X2 ⊕ Y2) = E, F := X1 ⊕ Y1, G := X2 ⊕ Y2
satisfies all the assumptions of Lemma 3.1, making j a group isomorphism on
π0(GL(E)). From Proposition 2.1, we need to check whether the inclusion
ϕ∗ : π0(GL(X1 ⊕ X2) × GL(Y1 ⊕ Y2)) → π0(GL(E))
is surjective. In particular, we want to know whether we can obtain an invertible
operator in T on E such that j(T ) = 1. Let
(A, B) ∈ GL(X1 ⊕ X2) × GL(Y1 ⊕ Y2)
be an arbitrary element and T := A ⊕ B. We denote by P1 and P2 the projectors
on X1 and X2, and by Q1 and Q2, the projectors on Y1 and Y2. We have
(3.2)
(3.3)
(3.4)
1 = ind(P1AP1) + ind(Q1BQ1)
0 = ind(Q1BQ1) + ind(Q2BQ2)
0 = ind(P1AP1) + ind(P2AP2).
The first equality comes from the requirement j(T ) = 1. The second and the third
equalities follows from (a) of Lemma 3.1 when it is applied to the spaces Y1, Y2 and
X1, X2.
4
DANIELE GARRISI
In finite-dimensional spaces, all the operators are Fredholm with index zero. Sup-
pose that the assumption of the proposition is not fulfilled, that is, for instance, X1
and Y1 have finite dimension. Therefore, in (3.2), we obtain 1 = 0 + 0; if X1 and
Y2 have finite dimension, therefore, from (3.3), we obtain
which yields to
ind(Q2BQ2) = 0 = ind(Q1BQ1)
1 = ind(P1AP1) = 0
combining (3.2) and the fact that X1 has finite dimension. A similar arguments
gives a contradiction when X2 and Y1 both have finite dimension.
Now, we suppose that both X1 and X2 have infinite dimension. Then, we can
define on X1 ⊕ X2 an invertible operator V similar to (3.1). Then
And, if both Y1 and Y2 have infinite dimension,
j(ϕ(V, idY1⊕Y2)) = 1.
j(ϕ(idX1⊕X2 , V )) = 1.
(cid:3)
Therefore, in ℓp ⊕ ℓq, the projector PF (or PG) is an example of projector such
that [P ] 6= RP . We notice that PF is not conjugated to idE − PF = PG (ℓp is
not isomorphic to ℓq). This might suggest that the features of the example in
(1.2) were not entirely transposed to the setting of linear operators. The following
remark explains why this is not possible
Remark 3.1. If P is in P(L(E)), and idE − P ∈ RP , then [idE − P ] = [P ].
We conclude this paper by showing that our example is substantially different from
the one obtained in [7]. We mean that C(S 3, M (2, C)) does not have the algebraic
structure of the space of bounded operators on a Banach space E. This would
already follow from Remark 3.1, but we prefer to give a proof which is independent
from the content of the work in [7].
Proposition 3.2. Given a non-empty compact space K, a natural number n and
a Banach space E, C(K, M (n, C)) is not isomorphic to L(E), unless E ≃ Cn and
K is a singleton.
Proof. We look at the centers of the two algebras. We have
Z(L(E)) = hidEi, Z(C(K, M (n, C))) = C(K, hidni).
If C(K, M (n, C)) ≃ L(E) there is an algebra isomorphism between the two centers.
Therefore, Z(L(E)) ≃ Z(C(K, M (n, C))). The first space has dimension one, while
the second space has infinite dimension unless K is finite.
If K is a singleton,
then the dimension is one. In that case C(K, M (n, C)) ≃ M (n, C). In conclusion,
M (n, C) is isomorphic to L(E) only in one case: when E ≃ Cn.
(cid:3)
References
[1] Fernando Albiac and Nigel J. Kalton, Topics in Banach space theory, Graduate Texts in
Mathematics, vol. 233, Springer, New York, 2006. MR MR2192298 (2006h:46005)
[2] Adrien Douady, Un espace de Banach dont le groupe lin´eaire n'est pas connexe, Nederl. Akad.
Wetensch. Proc. Ser. A 68 = Indag. Math. 27 (1965), 787 -- 789. MR MR0187056 (32 #4511)
[3] K. Janich, Vektorraumbundel und der raum der fredholm-operatoren, Math. Ann. 161 (1965),
no. 2, 129 -- 142.
ON THE CONNECTED COMPONENTS OF THE CONJUGACY CLASS OF PROJECTORS 5
[4] Tosio Kato, Perturbation theory for linear operators, Classics in Mathematics, Springer-Verlag,
Berlin, 1995, Reprint of the 1980 edition. MR MR1335452 (96a:47025)
[5] B. S. Mitjagin, The homotopy structure of a linear group of a Banach space, Uspehi Mat.
Nauk 25 (1970), no. 5(155), 63 -- 106. MR MR0341523 (49 #6274a)
[6] Gerhard Neubauer, Der Homotopietyp der Automorphismengruppe in den Raumen lp und c0,
Math. Ann. 174 (1967), 33 -- 40. MR MR0219088 (36 #2171)
[7] Horacio Porta and L´azaro Recht, Spaces of projections in a Banach algebra, Acta Cient.
Venezolana 38 (1987), no. 4, 408 -- 426 (1988). MR MR953576 (90a:46189)
E-mail address: [email protected]
West Building, Office No. 5W443, Department of Mathematics Education, Inha Uni-
versity, 253 Yonghyun-Dong, Nam-Gu, Incheon, South Korea 402-751
|
1901.06186 | 1 | 1901 | 2019-01-18T11:32:53 | Orlicz-Besov extension and Ahlfors $n$-regular domains | [
"math.FA"
] | Let $n\ge2$ and $\phi : [0,\fz) \to [0,\infty)$ be a
Young's function satisfying $\sup_{x>0} \int_0^1\frac{\phi( t x)}{ \phi(x)}\frac{dt}{t^{n+1} }<\infty. $
We show that Ahlfors $n$-regular domains are Besov-Orlicz ${\dot {\bf B}}^{\phi}$ extension domains,
which is necessary to guarantee the nontrivially of ${\dot {\bf B}}^{\phi}$. On the other hand, assume that $\phi$ grows sub-exponentially at $\fz$ additionally. If $\Omega$ is a Besov-Orlicz ${\dot {\bf B}}^{\phi}$ extension domain, then it must be Ahlfors $n$-regular. | math.FA | math | ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
TIAN LIANG AND YUAN ZHOU
0
dt
φ(tx)
φ(x)
Abstract Let n ≥ 2 and φ : [0, ∞) → [0, ∞) be a Young's function satisfying
supx>0R 1
tn+1 < ∞. We show that Ahlfors n-regular domains are Besov-Orlicz
φ
B
extension domains, which is necessary to guarantee the nontrivially of Bφ. On
the other hand, assume that φ grows sub-exponentially at ∞ additionally. If Ω is a
Besov-Orlicz Bφ extension domain, then it must be Ahlfors n-regular.
1. Introduction
Let φ : [0, ∞) → [0, ∞) be a Young function, that is, φ is a convex, φ(0) = 0, φ(t) > 0 for t > 0 and
limt→∞ φ(t) = +∞. Given any domain Ω ⊂ Rn, the Orlicz-Besov space Bφ(Ω) consists of all measurable
functions u in Ω whose (semi-)norms
kuk Bφ(Ω) := inf(α > 0 :ZΩZΩ
φ u(x) − u(y)
α
! dxdy
x − y2n ≤ 1)
9
1
0
2
n
a
J
8
1
]
s1
.
A
F
h
t
a
m
[
eq1.t1
eq1.t1
1
v
6
8
1
6
0
.
delta0
delta0
1
0
9
1
:
v
i
X
r
a
is finite. Modulo constant functions, Bφ(Ω) is a Banach space. We refer to [14] for the applications of Orlicz-
Besov spaces in qausi-conformal geometry. Note that, in the case of φ(t) = t p with p ≥ 1, the Bφ(Ω)-norms
are written as
(1.1)
= ZΩZΩ
kuk Bφ(Ω)
u(x) − u(y)p
x − y2n
dxdy!1/p
.
By this, when φ(t) = t p with p > n, Bφ(Ω) is exactly the Besov spaces Bn/p
p,p (Ω) (or fractional Sobolev spaces
W n/p,p(Ω)). However, when φ(t) = t p with p ≤ n, thanks to (1.1) and [4], the space Bφ(Ω) is trivial, that is,
only contains constant functions.
In general, to guarantee the nontrivially of Bφ(Ω), we always assume
(1.2)
Cφ := sup
x>0
tn
φ(t)Z t
0
φ(s)
sn
ds
s
< ∞.
Indeed, (1.2) does imply that Bφ(Ω) contains smooth functions with compact supports, and hence nontrivial;
see Lemma 2.2 below. If φ(t) = t p, observe that φ satisfies (1.2) if and only if p > n, and Bφ(Ω) is nontrivial if
and only if p > n. In this sense, we see that (1.2) is optimal to guarantee the nontrivially of Bφ(Ω). There are
some other interesting Young functions satisfying (1.2), for example, tn[ln(1 + t)]α with α > 1, t p[ln(1 + t)]α
with p > n and α ≥ 1, t pectα
j=0 (ctα) j/ j! where c > 0 with α > 0,
with p > n, c > 0 and α > 0, and ectα
where [n/α] is the maximum of integers no bigger than n/α.
−P[n/α]
In this paper, we obtain the following results for the Orlicz-Besov extension in Ahlfors n-regular domains.
Recall that a domain Ω is Ahlfors n-regular if there exists a constant CA(Ω) > 0 such that
B(x, r) ∩ Ω ≥ CA(Ω)rn ∀x ∈ Ω, 0 < r < 2 diam Ω.
A domain Ω is called Bφ-extension domain if any function u ∈ Bφ(Ω) can be extended to as a functioneu ∈
Bφ(Rn) continuously and linearly; in other words, there exists a bounded linear operator E : Bφ(Ω) → Bφ(Rn)
with EuΩ = u for any u ∈ Bφ(Ω).
t1.1 Theorem 1.1. Let φ be a Young function satisfying (1.2).
Date: January 21, 2019.
1
2
TIAN LIANG AND YUAN ZHOU
(i) If Ω ⊂ Rn is Ahlfors n-regular domain, then Ω is a Bφ-extension domain.
(ii) Assume that φ additionally satisfies
subexp
subexp
(1.3)
φ(x)e−cx = 0 ∀c > 0.
If Ω ⊂ Rn is a Bφ-extension domain, then Ω is Ahlfors n-regular.
lim sup
x→∞
with p > n and α ≥ 1, and ectα
Theorem 1.1 extends the known results for fractional Sobolev spaces Wn/p,p(Ω) or Besov space Bn/p
with p > n and α ∈ (0, 1), and ectα
−P[n/α]
Note that the condition (1.3) in Theorem 1.1 (ii) allows a large class of Young functions, including tn[ln(1 +
j=0 (ctα) j/ j!
j=0 (ctα) j/ j! where
t)]α with α > 1, t p[ln(1 + t)]α with p > n and α ≥ 1, t pectα
where c > 0 with α ∈ (0, 1). But (1.3) rules out t pectα
c > 0 with α ≥ 1,
p,p (Ω).
Recall that the extension problem for function spaces (including Sobolev, fracntional Sobolev, Hajlasz-Sobolev,
Besov and Triebel-Lizorkin spaces) have been widely studied in the literature, see [7, 8, 20, 9, 10, 2, 11, 16,
21, 22, 5, 17, 19] and the references therein. Given function spaces X(U) defined in any domain U ⊂ Rn in the
same manner, define X-extension domains similarly to Bφ-extension domains. It turns out that the extendabil-
ity of functions in X(Ω) not only relies on the geometry of the domain but also on the analytic properties of X.
In particular, it was essentially known that Ahlfors n-regular domains are fractional Sobolev Ws,p-extension
domains for any s ∈ (0, 1) and p ≥ 1; see Jonsson-Wallin [9] (also Shvartsman [17]). Here Ws,p(Ω) is the set
−P[n/α]
of all functions with
(1.4)
kuk Ws,p (Ω) := ZΩZΩ
u(x) − u(y)p
x − yn+sp
dxdy!1/p
< ∞.
Moreover, by Shvartsman [18] and Hajłasz et al [5, 6], Ahlfors n-regular domains also are Hajlasz-Sobolev
M1,p-extension domain with p ≥ 1. Recall that for a given function u in Ω, we say g is a Hajłasz gradient of u
(for short g ∈ D(u, Ω)) if
u(x) − u(y) ≤ x − y[g(x) + g(y)]
for almost all (x, y) ∈ Ω × Ω.
The Hajłasz Sobolev space M1,p(Ω) is the set of all functions u in Ω with
g∈D(u,Ω) kgkLp(Ω) < ∞
kuk M1,p(Ω) := inf
Conversely, Hajłasz [5, 6] essentially proved that Hajlasz-Sobolev M1,p-extension domains must be Ahlfors
n-regular; and by [24], similar results hold true for fractional Sobolev Ws,p-extension for any s ∈ (0, 1) and
p ≥ 1. Note that Wn/p,p(Ω) = Bn/p
p,p (Ω) = Bφ(Ω) for any p > n and φ(t) = t p.
To prove Theorem 1.1 (i), it suffices to define a suitable linear extension operator and prove its boundedness.
Following Jones [8], to define the extension operator we have to find suitable reflecting cubes of Whitney
cubes for Rn \ Ω. If we use the reflecting cubes the same as in [11, 5, 23, 24] which may have unbounded
overlaps, we cannot prove the boundedness of the extension operator in general since the Young function may
grows exponentially at ∞. See Remark 4.1 for details. Instead, we use the reflecting cubes of Shvartsman
[17, 18], which have bounded overlaps (see Lemma 2.2), to define extension operator. The bounded overlaps
of reflecting cubes allow us to use the convexity of φ, and also avoid using maximal functions. With some
careful analysis, we finally obtain the boundedness of extension operator.
Theorem 1.1 (ii) is proved in section 5 by borrowing some ideas from [5, 24]. Precisely, we first prove Ω
supports the following imbedding: there exists positive constants CI(Ω) and C(n) such that
im
(1.5)
c∈RZB∩Ω
inf
exp u − c
α ! dx ≤ C(n)B
for any ball B ⊂ Rn.
whenever u ∈ Bφ(Ω) and α > CI(Ω)kuk Bφ(Ω) > 0. Then we calculate the precise kuk Bφ(Ω)-norm of some cut-off
functions. Using this and the sub-exponential growth of φ following the idea from [5] (see also [6, 24]), we
are able to prove Ω is Ahlfors n-regular.
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
3
As a byproduct, we have the following result.
c1.2 Corollary 1.2. Suppose that φ is a Young function satisfying (1.2) and (1.3). Let Ω ⊂ Rn be any domain. The
following are equivalent:
(i) Ω is Ahlfors n-regular;
(ii) Ω is a Bφ-extension domain;
(iii) Ω supports the imbedding (1.5).
Remark 1.3. We conjecture that Theorem 1.1 (ii) holds without the additional assumption (1.3). The difficult
to remove (1.3) is to find a suitable imbedding properties of Bφ(Rn) better than (1.5) when φ does not satisfies
(1.3).
Note that (1.5) is always true when Ω is a Bφ-extension domain, but it is not enough to prove that Ω is
j=0 tγ j/ j! for t ≥ 0, by Lemma 2.5, Ω supports the imbedding
Ahlfors n-regular in general. If φ(t) = etα
?B
−P[n/α]
exp u(x) − uB
α
!γ
dx ≤ C(n)
whenever α > C(γ, n)kuk Bφγ (Rn). However, when Bφγ-extension domain, such a imbedding is also not enough
to prove Ω is Ahlfors n-regular.
Notation used in the following is standard. The constant C(n, α, φ) would vary from line to line and is
independent of parameters depending only on n, α, φ. Constants with subscripts would not change in different
occurrences , like Cφ. Given a domain, set BΩ(x, r) = B(x, r)∩Ω for convenience. We denote by uX the average
of u on X, namely, uX = >X
u dx. For a domain Ω and x ∈ Rn, we use d(x, Ω) to describe the distance
XRX
u ≡ 1
from x to Ω.
We list several basic properties of Orlicz-Besov spaces.
2. Some basic properties
12.4 Lemma 2.1. Suppose that φ is a Young function. Let Ω ⊂ Rn be any domain. Then Bφ(Ω) ⊂ L1(B ∩ Ω) for
any ball B ⊂ Rn, in particular, Bφ(Ω) ⊂ L1(Ω) when Ω is bounded.
Proof. For any α > kuk Bφ(Ω), we have
By Fubini's theorem, for almost all x ∈ Ω we have
Fix such a point x. For any B = B(z, r) ⊂ Rn with B ∩ Ω , ∅, we have x − y ≤ x + z + r for all y ∈ B ∩ Ω,
and hence
!
α
ZΩZΩ
ZΩ
φ u(x) − u(y)
φ u(x) − u(y)
!
α
φ u(x) − u(y)
α
u(x) − u(y)
α
?B∩Ω
φ ?B∩Ω
dydx
x − y2n ≤ 1.
< ∞.
dy
x − y2n
! dy < ∞.
dy! < ∞,
By Jessen's inequality, we have
which implies that
?B∩Ω u(y) dy ≤ u(x) + ?B∩Ω u(x) − u(y) dy < ∞,
that is, u ∈ L1(B ∩ Ω) as desired.
(cid:3)
4
TIAN LIANG AND YUAN ZHOU
l3.3 Lemma 2.2. Suppose that φ is a Young function satisfying (1.2). Let Ω ⊂ Rn be any domain. Then C1
c (Ω) ⊂
Bφ(Ω).
Proof. Assume that L = kukL∞(Ω) + kDukL∞(Ω) > 0. Let V = supp u ⋐ W ⋐ Ω. Then
H :=ZΩZΩ
By (1.2), we have
α
φ u(z) − u(w)
φ z − w
ZWZW
φ(cid:18) L
α(cid:19) dzdw
z − w2n
.
dz
α/L ! dzdw
+ 2ZΩ\WZV
z − w2n
φ z − w
α/L !
φ
α/L! dt
φ z − w
! dzdw
z − w2n ≤ZWZW
α/L ! dzdw
z − w2n ≤ZVZB(w,2 diam W)
= nωnZWZ 2 diam W
= nωnW(
= nωnW2−n diam W−nφ 2L diam W
z − w2n ≤ 2ωnφ(cid:18) L
α(cid:19)ZVZΩ\B(z, dist (V,W∁))
)nZ 2L diam W/α
z − w2n
dw
ds
sn+1
dwdz
φ (s)
tn+1
L
α
α
t
0
0
dw
! .
Moreover,
2ZΩ\WZV
α/L! dzdw
φ 1
z − w2n ≤ 2φ(cid:18) L
α(cid:19) V dist (V, W ∁)−n.
Obviously, letting α sufficiently enough and using the convexity of φ, we have H ≤ 1. That is, u ∈ Bφ(Ω) as
(cid:3)
desired.
The following Poincar´e type inequality is needed in Section 4. Below denote by ωn the area of the unit
sphere S n−1.
l2.5 Lemma 2.3. Suppose that φ is a Young function. For any ball B ⊂ Rn and u ∈ Bφ(B), we have
φ u(x) − uB
α
! dx ≤ ω2
n
?B
when α > kuk Bφ(B), and
?B u(x) − uB dx ≤ φ−1(ω2
n)kuk Bφ(B).
Proof. Let u ∈ Bφ(B). For any α > kuk Bφ(B), by Jensen's inequality, we have
>B u(x) − uB dx
α
φ
α
≤ ?B
≤ ?B?B
nZBZB
≤ ω2
φ u(x) − uB
! dx
φ u(x) − u(y)
φ u(x) − u(y)
?B u(x) − uB dx ≤ αφ−1(ω2
n).
α
α
! dydx
!
dydx
x − y2n ≤ ω2
n,
that is,
Letting α → kuk Bφ(B), we obtain
as desired.
?B u(x) − uB dx ≤ φ−1(ω2
n)kuk Bφ(B)
(cid:3)
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
5
As a consequence of Lemma 2.3, we have the following imbedding. Denote by BMO(Ω) the space of
functions with bounded mean oscillations, that is, the collection of u ∈ L1
B⊂Ω?B u(x) − uB dx < ∞.
kukBMO(Ω) = sup
loc (Ω) such that
c2.x Corollary 2.4. Suppose that φ is a Young function. Let Ω ⊂ Rn be any domain. We have Bφ(Ω) ⊂ BMO(Ω)
and kukBMO(Ω) ≤ φ−1(ω2
n)kuk Bφ (Ω) for all u ∈ Bφ(Ω).
Note that if
φγ(t) = etγ
[n/γ]Xj=0
tγ j
j!
−
with γ ≥ 1, then φγ a Young's function satisfying (1.2). Denote by Bφγ(Rn) the associated Orlicz-Besov space.
l2.xx5 Lemma 2.5. For γ ≥ 1, there exists constant C(γ, n) ≥ 1 such that for any u ∈ Bφγ(Rn) and ball B ⊂ Rn, we
have
exp u(x) − uB
α
!γ
?B
dx ≤ C(n)
whenever α > C(γ, n)kuk Bφγ (Rn).
Proof. By Corollary 2.4, we have u ∈ BMO(Rn) and kukBMO(Rn) ≤ φ−1(ω2
Nirenberg inequality, we have
n)kuk Bφ(Rn). Thus by the John-
Thus for all 1 ≤ j ≤ [n/γ], we have
?B u(x) − uB[n/γ] dx ≤ C(γ, n)kuk[n/γ]
BMO(Rn ) ≤ C(γ, n)kuk[n/γ]
!γ j
?B u(x) − uB
dx ≤ 1/n
α
.
Bφ (Rn)
when α ≥ C(γ, n)kukq
Bφ (Rn)
for some constant C(γ, n). Note that by Lemma 2.3, one has
φγ u(x) − uB
α
! dx ≤ 1/n
?B
when α > nω2
nkuk Bφ(Rn). Since
we obtain
when α > [nω2
n
+ C(γ, n)]kuk Bφ (Rn).
etγ
= φγ(t) + 1 +
exp u(x) − uB
α
?B
tγ j
j!
[n/γ]Xj=1
!γ
dx ≤ 3.
(cid:3)
s2
3. Whitney's decomposition and the reflected quasi-cubes
In this section, we always let Ω be an Ahlfors n-regular domain. Observe that ∂Ω = 0; see [17, Lemma
2.1] and also [24, 5]. Moreover, diam Ω = ∞ if and only if Ω = ∞. Write U := Rn \ Ω. Without loss of
generality, we assume U , ∅. It's well know that U admits a Whitney decomposition.
Lemma 3.1. There exists a collection W = {Qi}i∈N of (closed) cubes satisfying
(i) U = ∪i∈NQi, and Q◦k ∩ Q◦i
(ii) l(Qk) ≤ dist (Qk, ∂Ω) ≤ 4√nl(Qk);
(iii) 1
4 l(Qk) ≤ l(Qi) ≤ 4l(Qk) whenever Qk ∩ Qi , ∅.
= ∅ for all i, k ∈ N with i , k;
6
TIAN LIANG AND YUAN ZHOU
The following basic properties of Whitney's decomposition are used quite often in Section 4. For any
Q ∈ W , denote by N(Q) the neighbor cubes of Q in W , that is,
N(Q) := {P ∈ W , P ∩ Q , ∅}.
Then, by (iii) there exists an integer γ0 depending only on n such that
for all Q ∈ W .
♯N(Q) ≤ γ0
(3.1)
By (iii) again, for any P, Q ∈ W we know that
(3.2)
P ∈ N(Q) if and only if Q ∈ N(P), if and only if 9
e2.w1
e2.xx1
8 P ∩ 9
8 Q , ∅.
It then follows that
e2.w2
(3.3)
Indeed, by (3.2) we write
By lQ ≤ 4lP given in (iii), and (3.1), we arrive at
for all Q ∈ W .
χ 9
8 Q(x) dx.
1
1
QZU
QZU
QZU
χ 9
1
1
8 Q(x) dx ≤ 4nγ0
χ 9
QZP
8 Q(x) dx = XP∈N(Q)
8 Q(x) dx ≤ XP∈N(Q)
χ 9
P
Q ≤ 4nγ0
Wǫ := {Q ∈ W : lQ <
diam Ω}.
1
ǫ
eQ∗ := Q(x∗Q, lQ) ⊂ 10√nQ.
eQ∗ǫ := (ǫQ∗ ∩ Ω) \(cid:16)[{ǫP∗ : P ∈ Aǫ
Q}(cid:17) ,
Q :=nP ∈ Wǫ : ǫeP∗ ∩ ǫeQ∗ , ∅, lP ≤ ǫlQo .
Aǫ
as desired.
Below we recall the reflected quasi-cubes of Whitney's cubes as given by Shvartsman [17, Theorem 2.4].
For any ǫ > 0, set
Obviously, W = Wǫ for all ǫ > 0 if diam Ω = ∞, and Wǫ ( W for any ǫ > 0 if diam Ω < ∞.
For any Q = Q(xQ, lQ) ∈ Wǫ , fix any x∗Q ∈ Ω so that dist (Q, Ω) = dist (x, Q). By Lemma 3.1 (ii), one has
Set
where
properties; see [17, Theorem 2.4] for the proof, here we omit the details.
Below, when ǫ is small enough, we define eQ∗ǫ as reflected quasi-cubes of Q ∈ Wǫ so that they enjoy some nice
l2.1 Lemma 3.2. Let ǫ0 = [CA(Ω)/2γ0]1/n/(30√n). Denote by Q∗ = eQ∗ǫ0 as quasi-cubes of any cube Q ∈ Wǫ0 .
Then the following hold:
(ii)
Q ≤ γ1Q∗ whenever Q ∈ Wǫ0;
(i) Q∗ ⊂ (10√nQ) ∩ Ω for any Q ∈ Wǫ0;
(iii) PQ∈Wǫ0
If Ω is bounded, we let Q∗ = Ω as the reflected quasi-cube of any cube Q ∈ W \ Wǫ0
Above γ1 and γ2 are positive constants depending only on n and CA(Ω).
χQ∗ ≤ γ2.
, ∅. Write
where W
(0)
ǫ0
= Wǫ0. That is, W (k)
ǫ0
vk.1
vk.1
(3.4)
ǫ0
ǫ0
W (k)
is the kth-neighbors of Wǫ0.
= {Q ∈ N(P) : P ∈ W (k−1)
V (k) :=[{x ∈ Q; Q ∈ W (k)
ǫ0 } ∀k ≥ 0.
} ∀k ≥ 1,
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
7
Since Q∗ = Ω for Q < Wǫ0, by Lemma 3.3 (iii) we have
XQ∈W
(k)
ǫ0
χQ∗ ≤ XQ∈Wǫ0
χQ∗ + ♯(W
(k)
ǫ0 \ Wǫ0 )χΩ ≤ [γ2 + ♯(W (k)
ǫ0 \ Wǫ0)]χΩ ∀k ≥ 1.
(k)
For Q ∈ W
Lemma 3.1 (ii), we have
ǫ0 \ Wǫ0, observe that lQ ≥ 1
ǫ0
diam Ω and lQ ≤ 4klP ≤ 4k
ǫ0
diam Ω for some P ∈ Wǫ0 . Thus, by
Q ⊂ Q( ¯x, diam Ω + 8√n
4k
ǫ0
diam Ω)
for any fixed ¯x ∈ Ω, and hence
This yields that
eq2.xx2
eq2.xx2
(3.5)
4k
ǫ0
)nǫn
0 ≤ (ǫ0 + 4k+2 √n)n.
♯(W (k)
ǫ0 \ Wǫ0 ) ≤ (1 + 8√n
XQ∈W (k)
ǫ0
χQ∗ ≤ γ2 + (ǫ0 + 4k+2 √n)n ∀k ≥ 1.
Finally, associated to W , one has the following partition of unit of U.
Lemma 3.3. There exists a family {ϕQ : Q ∈ W } of functions such that
(i) for each Q ∈ W , 0 ≤ ϕQ ∈ C∞0 ( 17
(ii) for each Q ∈ W , ∇ϕQ ≤ L/lQ;
(iii) PQ∈W ϕ = χU .
It suffices to prove the existence of a bounded linear operator E : Bφ(Ω) → Bφ(Rn) such that EuΩ = u for
4. Proof of Theorem 1.1(i)
16 Q);
all u ∈ Bφ(Ω). Define the operator E by
Eu(x) ≡
u(x),
0
PQ∈W
x ∈ Ω,
x ∈ ∂Ω,
x ∈ U
ϕQ(x)uQ∗ ,
s3
eq3.y1
for any u ∈ Bφ(Ω). Recall that W is the Whitney cubes of U as in Lemma 3.1 and {ϕQ}Q∈W as in Lemma 3.3;
that Q∗ is the reflected quasi-cube of Q ∈ Wǫ0 as given in Lemma 3.2, and Q∗ = Ω if Q ∈ W \ Wǫ0 (when Ω is
bounded). By Lemma 2.1, uQ∗ = 1
u dx is always finite.
Obviously, E is linear, EuΩ = u in Ω, and moreover, if kuk Bφ(Ω)
constant function essentially. Thus, to prove the boundedness of E : Bφ(Ω) → Bφ(Rn), by the definition of the
norm k · k Bφ(Rn), we only need to find a constant M > 0 depending only on n, CA(Ω) and φ such that
(4.1)
dydx
= 0, then u and hence Eu must be a
Q∗RQ∗
α
!
φ Eu(x) − Eu(y)
= 1 and α > M. Below we assume that kuk Bφ(Ω)
!
H(α) =ZΩZΩ
+ 2ZUZΩ
+ZUZU
H(α) :=ZRnZRn
φ u(x) − u(y)
φ Eu(x) − Eu(y)
dydx
x − y2n
!
α
α
dydx
x − y2n
=: H1(α) + 2H2(α) + H3(α).
x − y2n ≤ 1.
= 1 Since ∂Ω = 0, one writes
φ Eu(x) − u(y)
!
dydx
x − y2n
α
whenever kuk Bφ(Ω)
To get (4.1), it suffices to find constants Mi ≥ 1 depending only on n, CA(Ω) and φ such that Hi(α) ≤ 1/4
whenever α ≥ Mi for i = 1, 2, 3. Indeed, by taking M = M1 + M2 + M3, we have H(α) ≤ 1 whenever α ≥ M.
8
TIAN LIANG AND YUAN ZHOU
Firstly, we may let M1 = 4. Indeed, if α > 4 that is, α/4 > 1, by the convexity of φ and kuk Bφ(Ω)
= 1, we
have
φ u(x) − u(y)
!
1
x − y2n ≤
4
To find M2 and M3, we consider two cases: diam Ω = ∞ and diam Ω < ∞.
Case diam Ω = ∞. To find M2, for any x ∈ U and y ∈ Ω, since PQ∈W
4ZΩZΩ
H1(α) ≤
dydx
α/4
1
.
ϕQ(x) = 1 by Lemma 3.3, one has
and hence, by the convexity of φ and Jensen's inequality,
φ Eu(x) − u(y)
α
Eu(x) − u(y) = XQ∈W
! ≤ φXQ∈W
≤ XQ∈W
≤ XQ∈W
ϕQ(x)[uQ∗ − u(y)],
α
ϕQ(x)uQ∗ − u(y)
ϕQ(x)φ ?Q∗
ϕQ(x)?Q∗
u(z) − u(y)
φ u(z) − u(y)
α
α
dz!
! dz.
If ϕQ(x) , 0, then x ∈ 17
l(Q), we have x − z ≤ 20nx − y, that is ,
16 Q. For z ∈ Q∗, by Q∗ ⊂ 10√nQ, we have x − z ≤ 20nl(Q). Since x − y ≥ d(x, Ω) ≥
So we have
ZΩ
Thus, by Lemma 3.2 (ii) we write
α
!
φ Eu(x) − u(y)
x − y2n ≤ (21n)2n XQ∈W
H2(α) ≤ 2(21n)2nZU XQ∈W
ϕQ(x)?Q∗ZΩ
≤ 2γ1(21n)2n XQ∈W 1
QZU
Since ϕQ ≤ χ 9
8 Q as given in Lemma 3.3, by (3.3) we have
y − z ≤ x − y + x − z ≤ 21nx − y
ϕQ(x)?Q∗ZΩ
dy
φ u(z) − u(y)
α
!
dzdy
z − y2n
.
φ u(z) − u(y)
ϕQ(x) dx!ZQ∗ZΩ
! dzdy
y − z2n
φ u(z) − u(y)
α
α
dx
! dzdy
y − z2n
.
which implies that
eq4.w2
(4.2)
1
1
ϕQ(x) dx ≤
QZU
QZU
H2(α) ≤ 2γ14nγ0(21n)2n XQ∈W ZQ∗ZΩ
H2(α) ≤ 2γ14nγ0γ2(21n)2nZΩZΩ
8 Q(x) dx ≤ 4nγ0,
χ 9
φ u(z) − u(y)
α
φ u(z) − u(y)
α
.
! dzdy
y − z2n
! dzdy
y − z2n
.
ByPQ∈W χQ∗ ≤ γ2 as in Lemma 3.2 (iii), we obtain
Take M2 = 8γ14nγ0γ2(21n)2n. By the convexity of φ again, if α > M2, we have H2(α) ≤ 1/4.
we have
XQ∈W
Eu(x) − Eu(y) = XP∈W XQ∈W
= XP∈W XQ∈W
! ≤ XQ∈W XP∈W
≤ XQ∈W XP∈W
α
φ Eu(x) − Eu(y
ϕQ(x) = XP∈W
ϕP(y) = 1,
ϕQ(x)ϕP(y)[uQ∗ − uP∗ ]
ϕQ(x)ϕP(y)?Q∗ ?P∗
ϕQ(x)ϕP(y)φ ?Q∗ ?P∗
ϕQ(x)ϕP(y)?Q∗ ?P∗
[u(z) − u(w)] dzdw.
α
u(z) − u(w)
φ u(z) − u(w)
α
dzdw!
! dwdz.
Applying the convexity of φ and Jensen's inequality, one obtains
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
9
To find M3, for each x ∈ U set
X1(x) :=(y ∈ U : x − y ≥
1
132n
max{d(x, Ω), d(y, Ω)})
and
X2(x) := U \ X1(x) =(y ∈ U : x − y <
1
132n
max{d(x, Ω), d(y, Ω)}) .
!
φ Eu(x) − Eu(y)
α
+ZUZX2(x)
dydx
x − y2n
Write
H3(α) =ZUZX1(x)
φ Eu(x) − Eu(y)
α
!
= H31(α) + H32(α)
dydx
x − y2n
then letting M3 := max{M31, M32}, for α > M3 we have H3 ≤ 1
Below, we show that there exists M3i ≥ 1 such that if α > M3i, then H3i ≤ 1/8 for i = 1, 2. If this is true,
To find M31, for x ∈ U and y ∈ X1(x), since
4 as desired.
For x ∈ Q and z ∈ Q∗, by Q∗ ⊂ 10√nQ, we have x − z ≤ 10nlQ ≤ 10nd(x, Ω). Similarly, for y ∈ P, and
w ∈ P∗, we have y − w ≤ 10nd(y, Ω). If y ∈ X(x), that is, 132nx − y ≥ max{d(x, Ω), d(y, Ω)}, we further have
z − w ≤ x − z + x − y + y − w ≤ 2641nx − y.
Thus
ϕQ(x)ϕP(y)?Q∗ ?P∗
By Q ≤ γ1Q∗ and P ≤ γ1P∗ as given in Lemma 2.2 (ii), we have
H31(α) ≤ (2641n)2nγ2
H31(α) ≤ (2641n)2nZUZX1(x) XQ∈W XP∈W
1 XQ∈W XP∈W 1
QZU
QZU
1
ϕQ(x) dx
1
PZU
PZU
1
By Lemma 3.3 and (3.3) we have
ϕQ(x) dx
ϕP(y) dy!ZQ∗ZP∗
ϕP(y) dy ≤ (4nγ0)2.
Thus
φ u(z) − u(w)
α
!
dwdz
z − w2n
!
φ u(z) − u(w)
α
dydx
dwdz
z − w2n
.
H31(α) ≤ (2641n)2nγ2
1(4nγ0)2 XQ∈W XP∈W ZQ∗ZP∗
φ u(z) − u(w)
α
!
dwdz
z − w2n
.
10
Observing PQ∈W
TIAN LIANG AND YUAN ZHOU
χQ∗ ≤ γ2 as in Lemma 3.2 (iii), we arrive at
2(4nγ0)2ZUZU
H31(α) ≤ (2641n)2nγ2
1γ2
φ u(z) − u(w)
α
!
dwdz
z − w2n
.
Letting M31 = 8(2641n)2nγ2
To find M32, write
1γ2
2(4nγ0)2. If α > M31, by the convexity of φ again we have H31(α) ≤ 1/8.
Let x ∈ U and y ∈ X2(x) ∩ P for some P ∈ W . Since
H32(α) =ZU XP∈W ZP∩X2(x)
α
φ Eu(x) − Eu(y)
XQ∈W (cid:2)ϕQ(x) − ϕQ(y)(cid:3) = 0,
!
dy
x − y2n
dx.
we write
Eu(x) − Eu(y) = XQ∈W (cid:2)ϕQ(x) − ϕQ(y)(cid:3) uQ∗ = XQ∈W (cid:2)ϕQ(x) − ϕQ(y)(cid:3) [uQ∗ − uP∗].
Note that by Lemma 3.3,
∇ϕQ ≤
L
lQ
16 Q.
χ 17
One gets
Moreover, we have
eq4.x1
eq4.x1
(4.3)
x − y
lQ
Eu(x) − Eu(y) ≤ L XQ∈W
(cid:20)χ 17
Eu(x) − Eu(y) ≤ 2L XQ∈N(P)
16 Q(x) + χ 17
16 Q(y)(cid:21) uQ∗ − uP∗.
x − y
lQ
8 Q(x)uQ∗ − uP∗.
χ 9
Indeed, since y ∈ X2(x), that is, x − y ≤ 1
132n max{d(x, Ω), d(y, Ω)}, taking ¯y ∈ ¯Ω with y − ¯y = d(y, Ω) we have
d(x, Ω) ≤ x − ¯y ≤ x − y + y − ¯y ≤
132n
d(x, Ω) +
1
1 + 132n
132n
d(y, Ω),
which implies
Similarly, we have
Thus,
d(x, Ω) ≤
d(y, Ω) ≤
132n + 1
132n − 1
132n + 1
132n − 1
d(y, Ω),
d(x, Ω).
If y ∈ 17
x − y ≤
132n
16 Q, by y ∈ P we have Q ∈ N(P), and hence
d(z, Ω) ≤
d(y, Ω) ≤ d(y, Q) + max
z∈Q
1
132n + 1
132n − 1
d(x, Ω).
1
16
√nlQ + 4√nlQ ≤
65
16
√nlQ,
Thus
x − y ≤
1
132n
which implies that x ∈ 9
8 Q. Moreover, if x ∈ 17
conclude that
1
132n + 1
65
√nlQ ≤
132n − 1 ×
16 Q, similarly we have y ∈ 9
32√n
16
lQ,
8 Q, and hence Q ∈ N(P). We
16 Q(x) + χ 17
χ 17
16 Q(y) = 0
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
11
when Q < N(P), and
when Q ∈ N(P), This gives (4.3).
Note that by Lemma 3.1(iii), PQ∈W
16 Q(x) + χ 17
χ 17
16 Q(y) ≤ 2χ 9
8 Q(x)
χ 17
16 Q(x) ≤ γ0. From the convexity of φ and (4.3) it follows that
Therefore, we obtain
H32(α) ≤
8 Q(x)uQ∗ − uP∗
α/L
α/2Lγ0 ! .
uQ∗ − uP∗
1
α
χ 9
≤
2χ 9
φ Eu(x) − Eu(y)
x − y
lQ
8 Q(x)φ x − y
! ≤ φ XQ∈N(P)
γ0 XQ∈N(P)
γ0 ZU XP∈W ZP∩X2(x) XQ∈N(P)
γ0 ZU XP∈W XQ∈N(P)
8 Q and y ∈ P ∩ X2(x), by d(x, Ω) ≤ 4√nlQ we have
8 Q(x)φ x − y
φ x − y
8 Q(x)ZP∩X2(x)
χ 9
χ 9
lQ
lQ
lQ
1
1
=
α/2Lγ0 !
uQ∗ − uP∗
α/2Lγ0 !
uQ∗ − uP∗
dydx
x − y2n
dy
x − y2n
dx
Observe that for x ∈ 9
x − y ≤
By the assumption (1.2) for φ, we have
α/2Lγ0 !
uQ∗ − uP∗
φ x − y
ZP∩X2(x)
lQ
dy
x − y2n ≤ nωnZ lQ
1
132n
0
132n + 1
132n − 1
lQ
α/2Lγ0
tn+1
d(x, Ω) ≤ lQ.
α/2Lγ0 ! dt
φ t
uQ∗ − uP∗
α/2Lγ0 !nZ uQ∗ −uP∗
≤ nωn(lQ)−n uQ∗ − uP∗
α/2Lγ0 ! .
≤ nCφωn(lQ)−nφ uQ∗ − uP∗
4α/2Lγ0 ! dx
8 Q(x)φ uQ∗ − uP∗
8 Q(x) dx! φ uQ∗ − uP∗
α/2Lγ0 !
(lQ)−nχ 9
χ 9
0
φ (s)
ds
sn+1
1
γ0
1
γ0
≤ nCφ
ωnZU XP∈W XQ∈N(P)
1
QZU
ωn XP∈W XQ∈N(P)
α/2Lγ0 ! .
φ uQ∗ − uP∗
≤ Cφnωn4n XP∈W XQ∈N(P)
φ uQ∗ − uP∗
α/2Lγ0 ! ≤ ?Q∗ ?P∗
φ u(z) − u(w)
α/2Lγ0
! dz dw
Using the above inequality and (3.3), one has
H32(α) ≤ nCφ
For each P ∈ W and Q ∈ N(P), by Jessen's inequality
Note that by Lemma 3.2 (i), P∗ ⊂ 10√nP and Q∗ ⊂ 10√nQ. Thus for any z ∈ P∗ and w ∈ Q∗, by Q ∈ N(P),
we have
z − w ≤ 10√n(lQ + lP) ≤ 50n min{lQ, lP}.
Since Q ≤ γ1Q∗ and P ≤ γ1P∗ as given in Lemma 3.2 (ii), one gets
z − w2n ≤ (50n)2n(γ1)2Q∗P∗.
Letting M32 = 8Lγ0Cφnωn42n(50n)2n(γ1)2(γ2)2. If α > M32, we have H32(α) ≤ 1/8 as desired.
Case diam Ω < ∞.
To find M2, write
H2(α) =ZV (2)ZΩ
Recall that V (2) is defined by (3.4) in Section 3. It suffices to find M2i such that H2i ≤ 1/8 for i = 1, 2.
, we have N(Q) ∩ Wǫ0
P ∈ N(Q). Thus, for any x ∈ U \ V (2), by Lemma 3.3 and Lemma 3.1 we have
ϕP(x)uP∗ = uΩ.
+ZU\V (2)ZΩ
Regards of H22(α), observe that for any Q ∈ W \ W (2)
φ Eu(x) − u(y)
φ u(x) − u(y)
dydx
x − y2n
dydx
x − y2n
!
!
α
α
ǫ0
= H21(α) + H22(α).
= ∅, and hence P∗ = Ω for all
Thus
Eu(x) = XP∈W
H22(α) =ZU\V (2)ZΩ
ϕP(x)uP∗ = XP∈N(Q)
φ uΩ − u(y)
α
!
dydx
x − y2n
.
12
Therefore,
and hence
TIAN LIANG AND YUAN ZHOU
φ uQ∗ − uP∗
α/2Lγ0 ! ≤ (50n)2n(γ1)2ZQ∗ZP∗
H32(α) ≤ Cφnωn42n(50n)2n(γ1)2 XP∈W XQ∈N(P)ZQ∗ZP∗
WithPQ∈W χQ∗ ≤ γ2 as given in Lemma 2.2 (iii), we obtain
H32(α) ≤ Cφnωn42n(50n)2n(γ1)2(γ2)2ZΩZΩ
φ u(z) − u(w)
α/2Lγ0
! dz dw
z − w2n
φ u(z) − u(w)
α/2Lγ0
! dz dw
z − w2n
φ u(z) − u(w)
α/2Lγ0
! dz dw
z − w2n
.
By Jensen's inequality, one gets
α
! dz
φ u(z) − u(y)
H22(α) ≤ZU\V (2)ZΩ?Ω
=ZΩZU\V (2)
x − y2n ?Ω
=ZΩ" diam Ω2n
ZU\V (2)
Ω
dydx
x − y2n
φ u(z) − u(y)
! dzdy
φ u(z) − u(y)
x − y2n#ZΩ
dx
dx
α
α
!
dz
z − y2n
.
For any x ∈ U \ V (2) and y ∈ Ω, since there exists Q ∈ W \ Wǫ0 so that x ∈ Q, one always has
x − y ≥ d(x, Ω) ≥ lQ ≥
1
ǫ0
diam Ω.
Moreover, by the Ahlfors n-regular assumption, it holds that Ω ≥ CA(Ω) diam Ω2. Thus,
diam Ω2n
Ω
ZU\V (2)
dx
x − y2n ≤
1
CA(Ω) diam ΩnZx−y> 1
≤ diam ΩnnωnZ ∞
≤ ωn
CA(Ω)
ǫn
0 ,
diam Ω
1
ǫ0
1
ǫ0
dx
diam Ω
x − y2n
1
rn+1
dr
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
13
from which, we conclude that
1
0ZΩZΩ
ǫn
φ u(z) − u(y)
α
H22(α) ≤ ωn
0 , by the convexity of φ again, for α > M22 we have H22(α) ≤ 1/8.
CA(Ω)
.
! dzdy
y − z2n
Letting M22 = 8ωn
1
C A
(Ω)ǫn
Regards of H21(α), observe that
XQ∈W
ϕQ(x) = XQ∈W (3)
ǫ0
ϕQ(x) = 1
whenever x ∈ V (2). With aid of this and following, line by line, the argument to get (4.2) for H2(α) in the case
diam Ω = ∞, one has
H21(α) ≤ 2γ14nγ0(21n)2n XQ∈W (2)
ǫ0
ZQ∗ZΩ
φ u(z) − u(y)
α
! dzdy
y − z2n
.
Here we omit the details. Since
χQ∗ ≤ γ2 + (ǫ0 + 128√n)n,
XQ∈W (2)
ǫ0
we have
H21(α) ≤ 2γ14nγ0[γ2 + (ǫ0 + 64√n)n](21n)2nZΩZΩ
φ u(z) − u(y)
α
! dzdy
y − z2n
.
Set M21 = 16γ14nγ0[γ2 +(ǫ0 +64√n)n](21n)2n. By the convexity of φ again, if α > M21, we have H21(α) ≤ 1/8
as dsired.
To find M3, notice that
We write
U × U ⊂ [V (3) × V (3)] ∪ [V (2) × (U\V (3))] ∪ [(U\V (3)) × V (2)] ∪ [(U\V (2)) × (U\V (2))].
H3 =ZV (3)ZV (3)
φ Eu(x) − Eu(y)
φ Eu(x) − Eu(y)
+ 2ZV (2)ZU\V (3)
!
!
dydx
x − y2n
α
α
dydx
x − y2n
φ Eu(x) − Eu(y)
!
+ZU\V (2)ZU\V (2)
α
=: H31(α) + 2H32(α) + H33(α).
dydx
x − y2n
for all α > M3i and i = 1, 2.
by γ2 + (ǫ0 + 45 √n)n, we can show that if α ≥ M31, then H31(α) ≤ 1/8. Here we omit the details.
Since Eu(x) = Eu(y) = uΩ for x, y ∈ U \ V (2), we have H33(α) = 0. It suffices to find M3i such that H3i(α)
Regard of H31(α), similarly to H3(α) in the case diam Ω = ∞ and taking M31 as M3 there with γ2 replaced
For H32(α), note that for y ∈ U\V (2), we have Eu(y) = uΩ. Thus
φ Eu(x) − uΩ
!
.
By Jessen's inequality, one has
H32(α) = 2ZV (2)ZU\V (3)
H32(α) ≤ZV (2)ZU\V (3)
ZU\V (3)
x − y2n ?Ω
For any x ∈ V (2) and y ∈ U\V (3) note that x − y ≥ l(Q) ≥ 1
x − y2n ≤ ǫn
dy
dy
ǫ0
α
dydx
x − y2n
φ Eu(x) − u(z)
! dx dz.
diam Ω, where Q ∈ W (3)
0 ( diam Ω)−n.
α
ǫ0 \ W (2)
ǫ0 and y ∈ Q. Thus
14
TIAN LIANG AND YUAN ZHOU
Since Ω ≥ CA(Ω) diam Ω, one has
1
H32(α) ≤
CA(Ω)
0 ( diam Ω)−2nZV (2)ZΩ
ǫn
φ Eu(x) − u(z)
α
! dx dz.
Note that for any x ∈ V (2) there exists a Pi ∈ W (i)
l(P0) ≤ 1
ǫ0
ǫ0
diam Ω, by Lemma 3.1 we know that l(P2) ≤ 42 1
x − y ≤ dist (x, Ω) + diam Ω ≤ diam P2 + dist (P2, Ω) + diam Ω ≤ 44 1
diam Ω. Thus for y ∈ Ω, one has
ǫ0
√n diam Ω.
ǫ0
such that x ∈ P2 and Pi ∈ N(Pi−1) for i = 1, 2. Since
Therefore,
H32(α) ≤
If α > M32 = 8
CA(Ω)
CA(Ω) 48nǫ−n
1
0 (44 1
ǫn
ǫ0
√n)2nZV (2)ZΩ
φ Eu(x) − u(z)
α
! dx dz
x − z2n ≤
1
CA(Ω)
48nǫ−n
0 nnH21(α).
0 nn M21, we have H32(α) ≤ 1/8. This completes the proof of Theorem 1.1 (i).
Remark 4.1. We emphasis that the bounded overlaps of reflecting cubes Q∗ in Lemma 3.2 (iii) play central
roles in the proof of the boundedness of extension operator E : Bφ(Ω) to Bφ(Rn).
Similarly to [24, 5, 11] and the reference therein, one may define the extension operator eEu similarly to
Eu but replacing Q∗ in Eu with Q(x∗Q, lQ) ∩ Ω, where x∗Q is the nearest point in Ω of Q ∈ W . Note that
{Q(x∗Q, lQ) ∩ Ω, Q ∈ W } does not have bounded overlap property as in Lemma 3.2(iii) in general.
In the case φ(t) = t p with p > n, similarly to [24], one may prove that eE is bounded from Bn/p
pp (Rn). The point is prove that
pp (Ω) to
Bn/p
eEu(x) − eEu(y)p
x − y2n
≤ M u(z) − u(w)p
z − w2n
χΩ×Ω! (x, y)
where M is certain Hardy-Littlewood maixmal operator. See page 968 in the proof of [24, Theorem 1.1].
For general φ in Theorem 1.1, some appropriate estimates of of φ(eEu(x)−eEu(y)
x−y2n via certain maximal
functions are not available for us. We do not know if it is possible to obtain the boundedness of eE from Bφ(Ω)
to Bφ(Rn). Note the our proof of the boundedness of E does not work for eE since {Q(x∗Q, lQ)∩ Ω, Q ∈ W } does
not have the bounded overlap property.
α
)
1
We divide the proof into 3 steps.
5. Proof of Theorem 1.1 (ii)
Step 1. Since Ω is a Bφ-extension domain, there exists a bounded linear extension operator E : Bφ(Ω) →
Bφ(Rn). For any u ∈ Bφ(Ω), we have Eu ∈ Bφ(Rn) with Eu = u in Ω, kEuk Bφ(Rn) ≤ kEkkuk Bφ (Ω). By Lemma
2.3, we haveeu ∈ BMO(Rn) and kEukBMO(Rn ) ≤ φ−1(S 12)kEuk Bφ (Rn). From the John-Nirenberg inequality, it
follows that
Thus,
IMIM
(5.1)
inf
u − c
Eu − (Eu)B
exp
CJN (n)kEukBMO(Rn )! dx ≤ C(n)B
ZB
C(φ, n, Ω)kuk Bφ(Ω)! dx ≤ C(n)B
exp
c∈RZB∩Ω
ux,r,t(z) =
1
t−x−z
t−r
0
z ∈ B(x, r) ∩ Ω
z ∈ (B(x, t) \ B(x, r)∩) Ω
z ∈ Ω \ B(x, t)
for all balls B ⊂ Rn
for all balls B ⊂ Rn.
Step 2. For x ∈ Ω and 0 < r < t < diam Ω, set the function
We have the following.
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
15
l2.3 Lemma 5.1. Suppose that φ is a Young function satisfying (1.2). For x ∈ Ω and 0 < r < t < diam Ω, we have
ux,r,t ∈ Bφ(Ω) with
kux,r,tk Bφ(Ω) ≤ 8ωn[Cφ4n + 1]"φ−1
(t − r)n
B(x, t) ∩ Ω!#−1
.
α
! dzdw
z − w2n
φ ux,r,t(z)
B(x,t)∩Ω(cid:17)]−1, we have H1 ≤ 1
(t−r)n
2
dz
z − w2n
dw.
φ (s)
ds
sn+1
.
Proof of Lemma 5.1. Write
φ ux,r,t(z) − ux,r,t(w)
ZΩZΩ
=ZB(x,t)∩ΩZB(x,t)∩Ω
α
! dzdw
z − w2n
α
dz
Write
Observe that
=: H1(α) + H2(α).
! dzdw
z − w2n
φ ux,r,t(z) − ux,r,t(w)
+ZΩ\B(x,t)ZB(x,t)∩Ω
If suffices to find a constant M depending only on n such that for α = M[φ−1(cid:16)
and H2(α) ≤ 1
2 .
H1(α) ≤ZB(x,t)∩ΩZB(w,t−r)∩Ω
α(t − r)!
φ z − w
ZB(w,t−r)∩Ω
φ 1
α!
sn+1 ≤ nωn(t − r)−nα−nZ 1/α
dw +ZB(x,t)∩ΩZ(B(z,2t)\B(w,t−r))∩Ω
z − w2n
α(t − r)! ds
φ
α! .
z − w2n ≤ nCφωn(t − r)−nφ 1
φ z − w
α(t − r)!
z − w2n ≤ nωnZ t−r
α(t − r)!
φ z − w
ZB(z,t−r)∩Ω
Z(B(w,2t)\B(w,t−r))∩Ω
z − w2n ≤ZRn\B(w,t−r)
H1(α) ≤ ωn(nCφ + 1)B(x, t) ∩ Ω
(t − r)n
= ωn(t − r)−n
z − w2n
φ 1
α! .
Applying (1.2), we have
On the other hand,
dz
dz
dz
dz
s
0
0
(t−r)n
B(x,t)∩Ω(cid:17)]−1 and M ≥ 2(nCφ + 1)ωn, we have
(nCφ + 1)ωn
H1(α) ≤
M
≤ 1/2.
Thus
If α = M[φ−1(cid:16)
dz +ZB(x,r)∩ΩZΩ\B(x,t)
φ 1
α!
dw
z − w2n
dz.
Write
Note that Ω \ B(x, t) ⊂ Ω \ B(z, t − z − x), we have
dw
dw
z − w2n
H2(α) ≤Z(B(x,t)\B(x,r))∩Ω
ZΩ\B(x,t)
H2(α) ≤Z(B(x,t)\B(x,r))∩Ω
≤ 2ωnB(x, t) ∩ Ω
(t − r)n
φ t − z − x
α(t − r) !ZΩ\B(x,t)
z − w2n ≤ZRn\B(z,t−z−x)
α(t − r) ! ωn(t − z − x)−ndz +ZB(x,r)∩Ω
φ t − z − x
sup
φ(cid:18) s
α(cid:19) 1
α!
+ φ 1
s∈(0,1]
dw
sn
Hence,
z − w2n ≤ ωn(t − z − x)−n
φ 1
α! ωn(t − z − x)−ndz
16
TIAN LIANG AND YUAN ZHOU
Notice that
an hence
Therefore,
If α = M[φ−1(cid:16)
as desired.
sup
s∈(2− j−1,2− j]
sn ≤ 2nZ 2− j+1
α(cid:19) 1
α(cid:19) ds
φ(cid:18) s
φ(cid:18) s
sn+1 ≤ 2nα−nZ 2/α
α(cid:19) ds
φ(cid:18) s
φ (s)
2− j
0
sn+1
sn ≤ 2nZ 2
φ(cid:18) s
α(cid:19) 1
0
sup
s∈(0,1]
H2(α) ≤ 2(Cφ4n + 1)ωnB(x, t) ∩ Ω
(t − r)n
(t−r)n
B(x,t)∩Ω(cid:17)]−1 and M ≥ 8(Cφ4n + 1)ωn, we have
ds
sn+1 ≤ Cφ4nφ 1
α/2!
φ 2
α! .
2(Cφ4n + 1)ωn
H2(α) ≤
M/2
1
2
.
≤
(cid:3)
e5.w1
e5.w1
Step 3. Let x ∈ Ω and 0 < r < 2 diam Ω. Let b0 = 1 and b j ∈ (0, 1) for j ∈ N such that
(5.2)
Let u j = ux,b j+1r,b jr for j ≥ 1 be as in Lemma 5.1. By (5.1), we have
B(x, b jr) ∩ Ω = 2−1B(x, b j−1r) ∩ Ω = 2− jB(x, r) ∩ Ω.
c∈RZB(x,b j−1r)∩Ω
C(φ, n, Ω)kuk Bφ(Ω)! dy ≤ C(n)rn;
exp
u j − c
inf
For any c ∈ R, we know that u j − c ≥ 1/2 either on B(x, b j+1r ) ∩ Ω or on [B(x, b j−1r)(cid:31)B(x, b jr )] ∩ Ω, and
note that, by (5.2),
B(x, b j+1r ) ∩ Ω = [B(x, b j−1r)(cid:31)B(x, b jr )] ∩ Ω = 2− j−1B(x, r ) ∩ Ω.
Thus, for any j ≥ 1, we have
2− j−1B(x, r ) ∩ Ω exp
C(φ, n, Ω)ku jk Bφ(Ω)! ≤ C(n)rn
u j − c
that is,
Since
we have
and hence
u j − c
C(φ, n, Ω)ku jk Bφ(Ω) ≤ ln 2 j
B(x, b jr) ∩ Ω!#−1
ku jk Bφ(Ω) ≤ C(φ, n)"φ−1 (b jr − b j+1r)n
φ−1 2 j (b jr − b j+1r)n
C(φ, n, Ω)
1
C(n)rn
B(x, r ) ∩ Ω! .
= C(φ, n)"φ−1 2 j (b jr − b j+1r)n
B(x, r) ∩ Ω !#−1
,
C(n)rn
B(x, r) ∩ Ω ! ≤ ln 2 j
φ"C(φ, n, Ω) ln 2 j
B(x, r ) ∩ Ω! ,
B(x, r ) ∩ Ω!#
2C(n)rn
(b j − b j+1)n ≤ 2− jB(x, r) ∩ Ω
rn
By (1.3), for any δ > 0, we have φ(t) ≤ C(δ)eδt for all t ≥ 0. Taking δ0 = 1/2C(φ, n, Ω), that is,
C(φ, n, Ω)δ0 = 1/2, we obtain
(b j − b j+1)n ≤ C(δ0)[2C(n)]δ0C(φ,n,Ω) 2− jB(x, r) ∩ Ω
rn
!1−δ0C(φ,n,Ω)
ORLICZ-BESOV EXTENSION AND AHLFORS n-REGULAR DOMAINS
17
Thus
b1 =
If b1 ≥ 1/10, we get
as desired.
rn
2− j/2.
!1/2
≤ C(φ, n, Ω) B(x, r) ∩ Ω
∞Xj=1(cid:16)b j − b j+1(cid:17) ≤ C(φ, n, Ω) B(x, r) ∩ Ω
rn
!1/2n
B(x, r) ∩ Ω ≥ C(φ, n, Ω)rn
If b1 < 1/10, we can know that exists a point x′ ∈ B(x, r) ∩ Ω satisfying x − x′ = b1r + r/5. Let R = 2r/5,
then B(x, b1r) ⊂ B(x′, R) ⊂ B(x, r) and B(x, b1r) ∩ B(x′, R/2) = ∅. Thus
(cid:12)(cid:12)(cid:12)B(x′, R/2) ∩ Ω(cid:12)(cid:12)(cid:12) ≤
=
1
1
2(cid:16)(B(x, r)\B(x, b1r)) ∩ Ω +(cid:12)(cid:12)(cid:12)B(x′, R/2) ∩ Ω(cid:12)(cid:12)(cid:12)(cid:17)
2(cid:16)B(x, b1r) ∩ Ω +(cid:12)(cid:12)(cid:12)B(x′, R/2) ∩ Ω(cid:12)(cid:12)(cid:12)(cid:17)
2(cid:12)(cid:12)(cid:12)B(x′, R) ∩ Ω(cid:12)(cid:12)(cid:12) ,
1
≤
2 B(x′, R) ∩ Ω, then b′1 ≥ 1/2. Applying the result when b1 ≥ 1/10 to the
By this, if (cid:12)(cid:12)(cid:12)B(x′, b′1R) ∩ Ω(cid:12)(cid:12)(cid:12) = 1
B(x′, R/2) and b′1 ≥ 1/2, we get
as desired. This completes the proof of Theorem 1.1 (ii).
B(x, r) ∩ Ω ≥(cid:12)(cid:12)(cid:12)B(x′, R) ∩ Ω(cid:12)(cid:12)(cid:12) ≥ C(φ, n, Ω)Rn ≥ C(φ, n, Ω)rn,
References
bsk96
ds93
L
gkz13
hkt08
[1] S.M. Buckley and P. Koskela, Criterria for imbeddings of Sobolev-Poincar´e type, Internat. Math. Res. Notices 18(1996), 881-902.
[2] R. A. DeVore, R. C. Sharpley, Besov spaces on domains in Rd, Trans. Amer. Math. Soc. 335 (1993), 843-864.
[3] L. Grafakos, Classical and Modern Fourior Analysis, Pearson Education Inc. 2008.
[4] A. Gogatishvili, P. Koskela and Y. Zhou, Characterizations of Besov and Triebel-Lizorkin spaces on metric measure spaces,
Forum Math. 25(2013), 787-819.
[5] P. Hajłasz, P. Koskela and H. Tuominen, Sobolev imbeddings, extensions and measure density condition, J. Funct. Anal. 254
hkt08b
[6] P. Hajłasz, P. Koskela and H. Tuominen, Measure density and extendability of Sobolev functions, Rev. Mat. Iberoam. 24 (2008),
(2008), 1217-1234.
645-669.
j80
j81
jw78
jw84
k98
kyz11
[7] P. W. Jones, Extension theorems for BMO, Indiana Univ. Math. J. 29 (1980), 41-66.
[8] P. W. Jones, Quasiconformal mappings and extendability of functions in Sobolev spaces, Acta Math. 147 (1981), 71-88.
[9] A. Jonsson, H. Wallin, A Whitney extension theorem in L p and Besov spaces, Ann. Inst. Fourier (Grenoble) 28 (1978), 139-192.
[10] Jonsson A, Wallin H. Function spaces on subsets of Rn, J. Mathematical Reports, 1984(1).
[11] P. Koskela, Extensions and imbeddings, J. Funct. Anal. 159 (1998), 369-384.
[12] P. Koskela, D. Yang and Y. Zhou, Pointwise characterizations of Besov and Triebel-Lizorkin spaces and quasiconformal map-
pings, Adv. Math. 226 (2011), 3579-3621.
kzz
mcp
[13] P. Koskela, Y. Zhang and Y. Zhou, Morrey-Sobolev Extension Domains.27(2017),1413-1434.
[14] Matias Carrasco Piaggio, Orlicz spaces and the large scale geometry of Heintze groups,Mathematische Annalen 368.6( 2017),
433-481.
npv12
[15] Nzza, Eleonnra Di, G.Palatucci, and E. Valdinoci. Hatchhiker's guide to the fractionalSobolev spaces, Buletin Des Science
Matheematiques 1136.5(2012), 521-573 .
r99
[16] V. S. Rychkov, On restrictions and extensions of the Besov and Triebel-Lizorkin spaces with respect to Lipschitz domains, J.
London Math. Soc. (2) 60 (1999), 237-257.
s06
[17] P. Shvartsman, Local approximations and intrinsic characterizations of spaces of smooth functions on regular subsets of Rn,
Math. Nachr. 279 (2006), 1212-1241.
s07
[18] P.Shvartsman, On extensions of Sobolev functions defined on rgular subsets of metric measure spaces, Journal of Approximation
Theory. 2144 (2007), 139-161.
s10
[19] P. Shvartsman, On Sobolev extension domains in Rn, J. Funct. Anal. 258 (2010), 2205-2245.
18
TIAN LIANG AND YUAN ZHOU
s70
t02
t08
wxz
z14
[20] E. M. Stein, Singular integrals and differentiability properties of functions, Princeton Mathematical Series, No. 30, Princeton
University Press, Princeton, N.J. 1970
[21] H. Triebel, Function spaces in Lipschitz domains and on Lipschitz manifolds. Characteristic functions as pointwise multipliers,
Rev. Mat. Complut. 15 (2002), no. 2, 475-524.
[22] H. Triebel, Function spaces and wavelets on domains, EMS Tracts in Mathematics, 7. European Mathematical Society (EMS),
Zorich, 2008. x+256 pp.
[23] Z. Wang, J. Xiao and Y. Zhou, A solution to the Q-restriction-extension problem in uniform domains, submitted.
[24] Y. Zhou, Fractional Sobolev extension and imbedding, Trans. Amer. Math. Soc. 367 (2015), 959-979.
Department of Mathematics, Beihang University, Beijing 100191, P.R. China
E-mail address: [email protected], [email protected]
|
1306.3001 | 1 | 1306 | 2013-06-13T00:50:37 | On the spectra of fermionic second quantization operators | [
"math.FA",
"math-ph",
"math-ph"
] | We derive several formulae for the spectra of the second quantization operators in abstract fermionic Fock spaces. | math.FA | math |
On the spectra of fermionic second quantization operators
Shinichiro Futakuchi∗and Kouta Usui†
Department of Mathematics,
Hokkaido University, Sapporo 060-0810, Japan.
March 1, 2018
Abstract
We derive several formulae for the spectra of the second quantization operators in abstract fermionic Fock
spaces.
1 Introduction
Abstract theory of Fock spaces [1, 2, 3, 4] provides powerful mathematical tools when one analyzes models
of quantum field theory, the most promising physical theory which is expected to describe the fundamental
interactions of elementary particles. This results from the fact that quantum filed theory deals with a quantum
system with infinitely many degrees of freedom, including particles which may be created or annihilated,
and that Fock spaces are furnished with suitable structure to describe particle creation or annihilation. In
mathematical physics, two different types of Fock spaces, bosonic (or symmetric) Fock space and fermionic (or
antisymmetric) Fock spaces, are considered, reflecting the fact that there are two different sorts of elementary
particles in Nature - bosons and fermions - .
In mathematical analyses of quantum theories, one of the most important problems includes to determine
the spectra of various self-adjoint operators representing physical observables, especially, that of a Hamilto-
nian, which represents the total energy of the system under consideration. To each self-adjoint operator A
acting in an underlying one particle Hilbert space H, bosonic or fermionic second quantization is defined as
an operator which naturally "lifts" A up to the bosonic or fermionic Fock space over H, respectively. In a
bosonic Fock space, the spectra of second quantization operators were well investigated and useful formulae
for them have been available. However, as far as we know, the corresponding useful formulae in a fermionic
Fock space are still missing. The main motivation of the present work is to derive such formulae in fermionic
Fock spaces to fill the gap.
Let H be an infinite dimensional separable Hilbert space over C with inner product h·,·iH and norm k·kH
(we omit the subscript H if there will be no danger of confusion). For a linear operator T on H, we denote
its domain by D(T ). For a subspace D ⊂ D(T ), the symbol T ↾ D denotes the restriction of T to D. We
denote by ¯T the closure of T if T is closable. The spectrum (resp. the point spectrum) of T is denoted by
σ(T ) (resp. σp(T )). The symbol ⊗n H (resp. ∧nH) denotes the n-fold tensor product of H (resp. the n-fold
antisymmetric tensor product). Let Sn be the symmetric group of order n. The antisymmetrization operator
An on ⊗n H is defined to be
sgn(σ)Uσ,
An :=
1
n! Xσ∈Sn
where Uσ is a unitary operator on ⊗n H such that Uσ(ψ1 ⊗ ···⊗ ψn) = ψσ(1) ⊗···⊗ ψσ(n), ψj ∈ H, j =
1, . . . , n, and sgn(σ) is the signature of the permutation σ ∈ Sn. Then, An is an orthogonal projection onto
∗Email: [email protected]
†Email: [email protected]
1
∧nH. The fermionic Fock space over H is defined by
⊕n=0∧nH :=nΨ = {Ψ(n)}∞
Ff (H) :=
∞
n=0(cid:12)(cid:12)(cid:12) Ψ(n) ∈ ∧nH,
∞Xn=0
kΨ(n)k2 < ∞o.
For a densely defined closable operator A on H and j = 1, . . . , n, we define a linear operator eAj on ⊗n H by
where I denotes the identity. For each n ∈ {0} ∪ N, a linear operator A(n) on ⊗n H is defined by
eAj := I ⊗···⊗ I ⊗
nXj=1 eAj ↾
A(0) := 0, A(n) :=
j-th
`
A ⊗ I ⊗ ···⊗ I,
n
⊗alg
D(A), n ≥ 1,
alg D(A) means the n-fold algebraic tensor product of D(A). Denote the reduced part of A(n) (resp.
where ⊗n
⊗n A) to ∧nH by A(n)
f
(resp. ∧nA). The infinite direct sum of these closed operators
dΓf (A) :=
∞
⊕n=0
A(n)
f
is called the first type fermionic second quantization of A, while the direct sum
Γf (A) :=
∞
⊕n=0∧nA
is the second type.
2 Main Results
For a linear operator T on H, σd(T ) denotes the discrete spectrum of T . We introduce the notation
t(λ;{λ1, . . . , λn}) := #{j λ = λj},
(2.1)
or, in words, t(λ;{λ1, . . . , λn}) represents how many λ's appear in the set {λ1, . . . , λn}. The main results of
the present paper are summarized in the following theorems:
Theorem 2.1. Let T be a self-adjoint operator on H. Then, the following (i), (ii) hold.
(i). The point spectrum of T (n)
is given by
f
σp(T (n)
f
) =n nXj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ) (j = 1, 2, . . . , n), t(λj ;{λ1, . . . , λn}) ≤ dim ker(T − λj)o.
(ii). If 0 /∈ σp(T ), then the point spectrum of ∧nT is given by
σp(∧nT ) =n nYj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ) (j = 1, 2, . . . , n), t(λj;{λ1, . . . , λn}) ≤ dim ker(T − λj)o.
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ) (j = 1, 2, . . . , n), t(λj ;{λ1, . . . , λn}) ≤ dim ker(T − λj )o.
If 0 ∈ σp(T ), then it is given by
σp(∧nT ) = {0} ∪n nYj=1
(2.2)
(2.3)
(2.4)
Theorem 2.2. Let T be a self-adjoint operator on H. Then, the following (i), (ii) hold.
2
(i). The spectrum of T (n)
f
is given by
σ(T (n)
f
) =n nXj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σ(T ) (j = 1, 2, . . . , n), if λj ∈ σd(T ), t(λj;{λ1, . . . , λn}) ≤ dim ker(T − λj)o.
(2.5)
(ii). The spectrum of ∧nT is given by
σ(∧nT ) =n nYj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σ(T ) (j = 1, 2, . . . , n), if λj ∈ σd(T ), t(λj ;{λ1, . . . , λn}) ≤ dim ker(T − λj )o.
(2.6)
3 Proof of the Theorems
Hp(T ) denotes the closed linear subspace spanned by the eigenvectors of a linear operator T on H.
Lemma 3.1. If T is a self-adjoint operator, then
(i).
(ii).
n
⊗ Hp(T ),
Hp(T (n)) =
⊗(Hp(T ) ∩ (ker T )⊥)i⊕ ker(⊗nT ).
n
⊗ T ) =h n
Hp(
(3.1)
(3.2)
Proof. (i) Let σc(T ) := σ(T ) \ σp(T ) and let Hc(T ) := Ran ET (σc(T )), where ET (·) is the one dimensional
spectral measure of T . Then, we obtain the direct sum decomposition T = Tp ⊕ Tc corresponding to the
decomposition H = Hp(T )⊕Hc(T ). Then, one can show that
T (n) =
⊕♯j =p or c
nXj=1
I♯1 ⊗ ··· ⊗ I♯j−1 ⊗ T♯j ⊗ I♯j+1 ⊗ ··· ⊗ I♯n
=
⊕♯j =p or c
S♯1,...,♯n ,
where I♯j is the identity operator in H♯j (T ), and
S♯1,...,♯n :=
nXj=1
I♯1 ⊗ ··· ⊗ I♯j−1 ⊗ T♯j ⊗ I♯j+1 ⊗ ··· ⊗ I♯n .
(3.3)
Note that Tc have no eigenvalues. Since σp(T ) and σc(T ) are disjoint, one finds that, if ♯j = c for some j,
then Hp(S♯1,...,♯n) = {0} by using Lemma A.1. Hence, in the direct sum decomposition in equation (3.3),
only the term with ♯j = p, for all j, is nontrivial, which implies
Hp(T (n)) =
⊕♯j =p or cHp(S♯1,...,♯n)
Since Sp,...,p is an operator in ⊗n Hp(T ), it immediately follows that Hp(Sp,...,p) ⊂ ⊗n Hp(T ).
Conversely, let ψ = ψ1 ⊗ ··· ⊗ ψn ∈ ⊗nHp(T ) with ψj ∈ ker(T − λj). Then, direct computation shows
= Hp(Sp,...,p).
T (n)ψ =
=
=
nXj=1
nXj=1
nXj=1
ψ1 ⊗ ··· ⊗ T ψj ⊗ ··· ⊗ ψn
ψ1 ⊗ ··· ⊗ λj ψj ⊗ ··· ⊗ ψn
λjψ.
3
(3.4)
Thus, we see ψ ∈ ker(T (n)− λ) with λ =Pj λj, and especially ψ ∈ Hp(T (n)). Since the closed linear subspace
spanned by such ψ's is ⊗nHp(T ) and Hp(T (n)) is closed, the converse inclusion follows. This proves (3.1).
sum decomposition of Hp(⊗n T ):
(ii) Let T0 and T1 be the reduced parts of T by ker T and (ker T )⊥ respectively. Then, we have a direct
n
Hp(
⊗ T ) =
⊕♭j =0 or 1Hp(T♭1 ⊗···⊗ T♭n ).
From this, we learn
n
n
n
Hp(
⊗ T ) = Hp(
⊗ T1)⊕ ker(
⊗ T ),
(3.5)
because, if ♭j = 0 for some j, then T♭1 ⊗···⊗ T♭n is a null operator.
zero,
Now, we will show, in general, that for self-adjoint operators A, B, whose point spectra do not contain
Hp(A ⊗ B) = Hp(A) ⊗ Hp(B).
(3.6)
In the same manner as in the proof of (i), we use the direct sum decompositions A = Ap⊕Ac and B = Bp⊕Bc
to obtain a decomposition of Hp(A ⊗ B):
Hp(A ⊗ B) = Hp(Ap ⊗ Bp) ⊕ Hp(Ap ⊗ Bc) ⊕ Hp(Ac ⊗ Bp) ⊕ Hp(Ac ⊗ Bc).
(3.7)
But, by Lemma A.1, we have Hp(Ac⊗Bc) = {0}. Moreover, by the same Lemma, we also have Hp(Ap⊗Bc) =
{0} and Hp(Ac ⊗ Bp) = {0}, because we have assumed that 0 6∈ σp(A) and 0 6∈ σp(B). Therefore, one finds
(3.8)
Hp(A ⊗ B) = Hp(Ap ⊗ Bp).
Since the operator Ap ⊗ Bp acts in Hp(A) ⊗ Hp(B), it is clear that Hp(Ap ⊗ Bp) ⊂ Hp(A) ⊗ Hp(B), which
means
Hp(A ⊗ B) ⊂ Hp(A) ⊗ Hp(B).
The converse inclusion follows from the similar discussion given in (i). Thus, we prove (3.6).
The above general discussion shows that Hp(⊗n T1) = ⊗n Hp(T1) since T1 does not have zero eigenvalue.
Substituting this equation in (3.5), we obtain (3.2).
Proof of Theorem 2.1. (i) For vectors ζ1, . . . , ζn ∈ H, we define the wedge product of these vectors by
ζ1 ∧ ··· ∧ ζn := √n!An(ζ1 ⊗···⊗ ζn).
Let {ξk}k be a complete orthonormal system (CONS) of Hp(T ) consisting of eigenvectors of T . Then, as is
well known, the family
forms a CONS of ∧nHp(T ), and each element is obviously an eigenvector of T (n). By the reducibility and
Lemma 3.1 (i), we have
Λ := {ξk1 ∧ ··· ∧ ξkn(cid:12)(cid:12) k1 < ··· < kn}
Hp(T (n)
f
) = AnHp(T (n)) = An
n
⊗Hp(T ) = ∧nHp(T ).
We claim
σp(T (n)
f
) =nλ ∈ R(cid:12)(cid:12)(cid:12) There exists η ∈ Λ such that T (n)
f
η = ληo .
The right hand side is clearly included by the left hand side. To prove the converse, let λ ∈ σp(T (n)
f
eigenvector ψ:
T (n)
f ψ = λψ.
4
(3.9)
(3.10)
(3.11)
) with an
Take the inner product with ξk1 ∧ ··· ∧ ξkn to obtain
λkj − λ
nXj=1
hξk1 ∧ ··· ∧ ξkn , ψi = 0,
(3.12)
for all k1 < ··· < kn, where λkj is an eigenvalue of T to which ξkj belongs. By equation (3.10), Λ is a CONS
of Hp(T (n)
), and thus, for at least one choice of (k1, . . . , kn), hξk1 ∧···∧ ξkn , ψi 6= 0. This and equation (3.12)
f
imply
λ =
nXj=1
λkj ,
j=1 λkj is an eigenvalue of T (n)
to which the eigenvector ξk1 ∧···∧ ξkn belongs,
λ is an element of the right hand side of (3.11). This proves (3.11), but the right hand side of (3.11) is exactly
the same set that appears in the right hand side of (2.2). Then, the proof is completed.
for such (k1, . . . , kn). SincePn
(ii) In the case where 0 /∈ σp(T ), we can prove (2.3) by using Lemma 3.1 (ii) in the same way as in the
Next, suppose 0 ∈ σp(T ). By the reducibility and Lemma 3.1 (ii), we have
proof of (2.2).
f
Hp(∧nT ) = (∧nHp(T1)) ⊕ An ker(⊗nT ).
Let {ξk}k be a CONS of Hp(T1) and
Λ1 := {ξk1 ∧ ··· ∧ ξkn k1 < ··· < kn},
then Λ1 forms a CONS of ∧nHp(T1). We claim
σp(∧nT ) \ {0} =nλ ∈ R \ {0}(cid:12)(cid:12)(cid:12) There exists η ∈ Λ1 such that ∧n T η = ληo .
Since the right hand side is clearly included by the left, it suffices to prove that, for each λ ∈ σp(∧nT ) \ {0},
there is an η ∈ Λ1 such that ∧nT η = λη. Let λ ∈ σp(∧nT )\{0}. Then, there exists a ψ ∈ Hp(∧nT ) satisfying
∧nT ψ = λψ.
But since λ 6= 0, we may assume ψ ∈ ∧nHp(T1) by (3.13). Taking the inner product with ξk1 ∧ ··· ∧ ξkn on
both sides, one obtains
nYj=1
λkj − λ
hξk1 ∧ ··· ∧ ξkn , ψi = 0,
(3.15)
for all k1 < ··· < kn, where λkj is a non-zero eigenvalue of T to which ξkj belongs. By noting that ψ ∈
∧nHp(T1), and Λ1 is a CONS of ∧nHp(T1), we learn that for at least one of (k1, . . . , kn)'s hξk1∧···∧ξkn , ψi 6= 0.
Thus, we have
(3.13)
(3.14)
λ =
λkj
nYj=1
for such (k1, . . . , kn)'s. But since Qn
belongs, we proved the claim (3.14).
j=1 λkj is an eigenvalue to which the eigenvector ξk1 ∧ ··· ∧ ξkn ∈ Λ1
The right hand side of (3.14) is rewritten as
n nYj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ) \ {0}, t(λj ;{λ1, . . . , λn}) ≤ dim ker(T − λj)o,
and therefore, we have proved
σp(∧nT ) \ {0} =n nYj=1
which implies (2.4).
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ) \ {0}, t(λj ;{λ1, . . . , λn}) ≤ dim ker(T − λj )o,
5
(3.16)
(3.17)
Proof of Theorem 2.2. (i) Let
Σn :=n nXj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σ(T ), if λj ∈ σd(T ), t(λj;{λ1, . . . , λn}) ≤ dim ker(T − λj)o
be the right hand side of (2.5) without closure.
First, we prove the left hand side includes the right. Let λ = Pn
j=1 λj , λj ∈ σ(T ), and if λj ∈ σd(T ),
t(λj;{λ1, . . . , λn}) ≤ dim ker(T −λj). Choose ε0 > 0 in such a way that λj 6= λk implies Uε0(λj )∩Uε0 (λk) = ∅,
where Uε(λ) is the ε-neighborhood of λ. Then, for all ε with 0 < ε ≤ ε0, there exists an orthonormal set
{ψj}n
j=1 ⊂ D(T ) such that ψj ∈ Ran ET (Uε(λj )), k(T − λj )ψk < ε. Then, we obtain
k(T (n)
f − λ)ψ1 ∧ ··· ∧ ψnk
nXj=1
≤k
ψ1 ∧ ··· ∧ (T − λj)ψj ∧ ··· ∧ ψnk
√n!ε.
≤
).
Hence, λ ∈ σ(T (n)
f
Next, in order to prove the converse inclusion, we will show that there exists a CONS {Ψk}∞
satisfying the following condition:
k=1 ⊂ Vn H
supp µΨk ⊂ Σn,
k = 1, 2, . . . ,
(3.18)
where µΨk (B) := kET (n)
Vn H = L({Ψk}k) ⊂ Ran ET (n)
f
f
(B)Ψkk2 for a one dimensional borel set B ∈ B1. Then, it immediately follows that
(Σn). Here, L({Ψk}k) denotes the closed linear supspace spanned by {Ψk}k.
f ⊂ Σn by definition of support of a
(Σn) = I, and this implies σ(T (n)
) = suppET (n)
f
Hence, we obtain ET (n)
spectral measure.
f
f
f
(σd(E(n)
Now, we shall show (3.18). Denote the essential spectrum of T by σess(T ) and introduce the notations
)). Let {ζk}k be a CONS of Hd(T ) consist-
Hd(T ) := Ran E(n)
ing of eigenvectors of T , and let {ηl}j be a CONS of Hess(T ). Then, the set
)) and Hess(T ) := Ran E(n)
Λ := {ζk1 ∧ ··· ∧ ζkN ∧ ηl1 ∧ ··· ∧ ηlN ′ (cid:12)(cid:12)(cid:12) k1 < ··· < kN , l1 < ··· < lN ′, N + N ′ = n}
forms a CONS of Vn H. Fix some Ψ = ζk1 ∧ ··· ∧ ζkN ∧ ηl1 ∧ ··· ∧ ηlN ′ ∈ Λ. In addition, let νki (B) :=
kET (B)ζkik2 (i = 1, . . . , N ), νlj (B) := kET (B)ηljk (j = 1, . . . , N ′) for B ∈ B1, and Bi := supp νki , BN +j :=
supp νlj . Let
(σess(E(n)
f
f
J := [σ∈Sn
JΣn := {(λ1, . . . , λn) ∈ Rn λ1 + ··· + λn ∈ Σn},
Bσ(1) × ··· × Bσ(n),
and En
T (·) be an n-dimensional spectral measure acting in ⊗nH defined by the relation
T (B1 × ··· × Bn) = ET (B1) ⊗ ··· ⊗ ET (Bn), B1, . . . , Bn ∈ B1.
En
Then, we see that J ⊂ JΣn by the construction of Λ. By direct computation, we have En
it follows that En
T (JΣn )Ψ = Ψ and one finds
T (J)Ψ = Ψ. Therefore,
ET (n)
f
(Σn)Ψ = ET (n) (Σn)AnΨ = En
T (JΣn )Ψ
= Ψ,
where Lemma A.2 is used to obtain the second equality. This shows that suppµΨ ⊂ Σn, and therefore, Λ is
a desired CONS satisfying (3.18), completing the proof of (2.5).
(ii) The proof is very similar to that of (i), and we omit it.
6
As a corollary of these theorems, we can derive the formulae for spectra of the fermonic second quantization
operators. In what follows, we will use the simpler notation t(λj ) in place of t(λj ;{λ1, . . . , λn}) for notational
simplicity.
Corollary 3.2. Let T be a self-adjoint operator. Then, the following (i), (ii) hold.
(i). The spectrum and the point spectrum of the first type fermionic second quantization operator of T are
given by
(ii). As to the second type fermionic second quantization operator of T , one has
λj(cid:12)(cid:12)(cid:12) λj ∈ σ(T ), j = 1, . . . , n, if λj ∈ σd(T ), t(λj) ≤ dim ker(T − λj )o(cid:17),
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ), j = 1, . . . , n, t(λj ) ≤ dim ker(T − λj)o(cid:17).
λj(cid:12)(cid:12)(cid:12) λj ∈ σ(T ), j = 1, . . . , n, if λj ∈ σd(T ), t(λj ) ≤ dim ker(T − λj)o(cid:17).
σ(dΓf (T )) = {0} ∪(cid:16) ∞[n=1n nXj=1
σp(dΓf (T )) = {0} ∪(cid:16) ∞[n=1n nXj=1
σ(Γf (T )) = {1} ∪(cid:16) ∞[n=1n nYj=1
If 0 /∈ σp(T ), then
σp(Γf (T )) = {1} ∪(cid:16) ∞[n=1n nXj=1
If 0 ∈ σp(T ), then
σp(Γf (T )) = {0} ∪ {1} ∪(cid:16) ∞[n=1n nXj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ), j = 1, . . . , n, t(λj ) ≤ dim ker(T − λj )o(cid:17).
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(T ), j = 1, . . . , n, t(λj ) ≤ dim ker(T − λj)o(cid:17).
4 Example - Kinetic energy of free fermions in a finite box -
Let F be the fermionic Fock space over H = l2(ΓL; C4), where
ΓL :=
2π
L
Z3, L > 0.
This Hilbert space H consists of quantum mechanical state vectors of one Dirac fermion in momentum
representation living in a finite volume box [−L/2, L/2]3 ⊂ R3. As a one particle Hamiltonian, we adopt a
multiplication operator by a function EM :
where p ∈ R3 is a spacial momentum of a Dirac particle and M ≥ 0 is a constant representing a bare mass
of a Dirac particle. On a spinor space C4, EM acts as a diagonal matrix.
EM (p) =pp2 + M 2, p ∈ ΓL,
The spectrum of EM is given as follows:
Lemma 4.1.
(i). The spectrum of EM is given by
(ii). The multiplicity of eigenvalue λ is given by
σ(EM ) = σd(EM ) =npp2 + M 2(cid:12)(cid:12)(cid:12) p ∈ ΓLo .
4π2 (λ2 − M 2)(cid:19) ,
dim ker(EM − λ) = 4r(cid:18) L2
where r(N ) denotes
r(N ) := #{n ∈ Z3 N = n3}.
7
(4.1)
(4.2)
Proof. Since (i) is well known, we will prove only (ii).
For each (p, l) ∈ ΓL × {1, 2, 3, 4}, let
δl
p(q, m) = δpqδlm, q ∈ ΓL, m = 1, 2, 3, 4.
Then, {δl
theory of multiplication operators, we have
p}p,l forms a CONS of H under natural identification H = l2(ΓL × {1, 2, 3, 4}). From a general
ker(EM − λ) = {ψ ∈ D(EM ) ψ(p) 6= 0 implies pp2 + M 2 = λ}.
This means that ψ is an eigenvector of EM if and only if it belongs to the linear subspace spanned by
Since the above vectors are linearly independent, we find
p ∈ H pp2 + M 2 = λ, l = 1, 2, 3, 4}.
{δl
dim ker(EM − λ) = 4 · #{p ∈ ΓL pp2 + M 2 = λ}
= 4 · r(cid:18) L2
4π2 (λ2 − M 2)(cid:19) .
(4.3)
(4.4)
From Lemma 4.1 and Corollary 3.2, we finally arrive at the formula for the second quantization operator
dΓf (EM ) acting in F :
Theorem 4.2.
σ(dΓf (EM )) = {0} ∪(cid:16) ∞[n=1n nXj=1
r 4π2
L2 Nj + M 2(cid:12)(cid:12)(cid:12) 1 ≤ r(Nj), t(Nj;{N1, . . . , Nn}) ≤ 4r(Nj), j = 1, . . . , no(cid:17).
(4.5)
The crucial difference from the bosonic case is the existence of the restriction t(Nj;{N1, . . . , Nn}) ≤ 4r(Nj)
reflecting Pauli's exclusion principle. It would be interesting to note that, if Dirac fermions are not contained
in a finite box but live in R3, then there is no restriction because the spectrum of EM in this case consists
only of essential spectra.
Acknowledgements
The authors are grateful to Prof. Asao Arai for valuable comments. This work is partially supported by
JSPS Research Fellowships for Young Scientists and by World Premier International Center Initiative (WPI
Program), MEXT, Japan.
A Appendix
We will collect well known facts in abstract Fock spaces used in this paper. Detailed proofs will be found in
[1].
Lemma A.1. Let Aj (j = 1,··· , n) be self-adjoint operators on a separable Hilbert space Hi. Then,
(i)
σp(cid:16) nXj=1 eAj(cid:17) =n nXj=1
(ii) If 0 /∈ σp(Aj) for all Aj , then
σp(cid:16) n
⊗j=1
Aj(cid:17) =n nYj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(Aj), j = 1, . . . , no.
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(Aj), j = 1, . . . , no.
8
If 0 ∈ σp(Aj ) for a Aj , then
σp(cid:16) n
⊗j=1
Aj(cid:17) = {0} ∪n nYj=1
λj(cid:12)(cid:12)(cid:12) λj ∈ σp(Aj ), j = 1, . . . , no.
(i)
Lemma A.2. Let Aj (j = 1,··· , n) be self-adjoint operators on a Hilbert space Hi. Let E := E eA1 ×···× E eAn
be the product measure and B1 be the Borel field of R. Then,
nXj=1
λj ∈ Jo(cid:17),
J ∈ B1
EΣ(J) := E(cid:16)n(λ1, . . . , λn) ∈ Rn(cid:12)(cid:12)(cid:12)
j=1 eAj.
E⊗(J) := E(cid:16)n(λ1, . . . , λn) ∈ Rn(cid:12)(cid:12)(cid:12)
is the spectral measure of Pn
(ii)
is the spectral measure of ⊗n
j=1 Aj.
nYj=1
λj ∈ Jo(cid:17),
J ∈ B1
References
[1] A. Arai: "Fock Spaces and Quantum Fields," Nihon-Hyouronsha, Tokyo, 2000. (in Japanese)
[2] O. Bratteli and D. W. Robinson: "Operator Algebras and Quantum Statistical Mechanics, Vol. 2 (Second
Edition)," Springer, Berlin, Heidelberg, 1997.
[3] M. Reed and B. Simon: "Methods of Modern Mathematical Physics I: Functional Analysis," Academic
Press, New York, 1972.
[4] M. Reed and B. Simon:
"Methods of Modern Mathematical Physics II: Fourier Analysis, Self-
Adjointness," Academic Press, New York, 1975.
9
|
1106.2227 | 2 | 1106 | 2011-11-19T15:25:38 | A "hidden" characterization of polyhedral convex sets | [
"math.FA",
"math.CO"
] | We prove that a closed convex subset $C$ of a complete linear metric space $X$ is polyhedral in its closed linear hull if and only if no infinite subset $A\subset X\backslash C$ can be hidden behind $C$ in the sense $[x,y]\cap C\not = \emptyset$ for any distinct points $x,y\in A$. | math.FA | math |
A "HIDDEN" CHARACTERIZATION OF
POLYHEDRAL CONVEX SETS
TARAS BANAKH AND IVAN HETMAN
Abstract. We prove that a closed convex subset C of a complete linear metric space X is polyhedral
in its closed linear hull if and only if no infinite subset A ⊂ X\C can be hidden behind C in the sense
that [x, y] ∩ C 6= ∅ for any distinct points x, y ∈ A.
1. Introduction
A convex subset C of a real linear topological space L is called polyhedral in L if it can be written
i=1 f−1
i (cid:0)(−∞, ai](cid:1) of closed half-spaces determined by some linear
This notion has also an algebraic version. We shall say that a convex subset C of a linear space
i=1 Hi for some convex subsets H1, . . . , Hn ⊂ D
In this paper polyhedral sets will be characterized with help of a combinatorial notion of a hidden
as a finite intersection C = Tn
continuous functionals f1, . . . , fn : L → R and some real numbers a1, . . . , an, see [1].
L is polyhedric in a convex set D ⊃ C of L if C = Tn
having convex complements D \ Hi, i ≤ n.
set.
We say that a subset A of a linear space L is hidden behind a set C ⊂ L if A ⊂ L \ C and for any
distinct points a, b ∈ A the closed segment [a, b] = {ta + (1 − t)b : t ∈ [0, 1]} meets the set C. In this
case we shall also say that the set C hides the set A.
The main result of this paper is the following "hidden" characterization of closed polyhedral
convex sets in complete linear metric spaces. This characterization has been applied in the paper [2]
devoted to a characterization of approximatively polyhedral convex sets in Banach spaces, which was
applied in the paper [3] devoted to recognizing the topological type of connected components of the
hyperspace of closed convex subsets of a Banach space. Some other characterizations of polyhedral
convex sets can be found in [12] and [7].
Theorem 1.1. For a closed convex subset C of a complete linear metric space X the following
conditions are equivalent:
(1) C is polyhedral in its closed linear hull lin(C);
(2) C is polyhedric in its affine hull aff(C);
(3) C hides no infinite subset A ⊂ X \ C.
The proof of this theorem is rather long and will be presented in Section 3. Now let us show that
the assumption of the completeness of the linear space X in Theorem 1.1 is essential. A suitable
counterexample will be constructed in the (non-complete) normed space
c00 = {(xn)n∈ω ∈ Rω : ∃n ∈ ω ∀m ≥ n xm = 0}
endowed with sup-norm kxk = supn∈ω xn, where x = (xn)n∈ω ∈ c00.
Example 1.2. The standard infinite-dimensional simplex
∆ = {(xn)n∈ω ∈ c00 ∩ [0, 1]ω : Xn∈ω
hides no infinite subset of c00 \ ∆ but is not polyhedral in c00.
xn = 1} ⊂ c00
2010 Mathematics Subject Classification. Primary 46A55, 52A07; Secondary 52B05, 52A37.
Key words and phrases. Polyhedral convex set; a hidden set; complete linear metric space.
1
2
T. BANAKH AND I. HETMAN
i=1 f−1
i
i=1 f−1
i
i=1 f−1
i
This contradiction shows that ∆ is not polyhedral in c00.
Proof. First we show that the simplex ∆ is not polyhedral in c00. Assuming the opposite, we would find
linear functionals f1, . . . , fn : c00 → R and real numbers a1, . . . , an such that ∆ =Tn
((−∞, ai]).
Consider the linear subspace X0 =Tn
(0) that has finite codimension in c00. It follows that for
each x0 ∈ ∆, we get x0 + X0 ⊂Tn
((−∞, ai]) = ∆, which implies that the set ∆ is unbounded.
Now assume that some infinite subset A ⊂ c00\∆ can be hidden behind the simplex ∆. Decompose
the space c00 into the union c00 = Σ< ∪ Σ1 ∪ Σ> of the sets
Σ< = {(xn)n∈ω ∈ c00 : Xn∈ω
Σ1 = {(xn)n∈ω ∈ c00 : Xn∈ω
Σ> = {(xn)n∈ω ∈ c00 : Xn∈ω
xn < 1},
xn = 1},
xn > 1}.
Observe that for any two points x, y ∈ Σ< the segment [x, y] does not intersect ∆. Consequently,
A∩ Σ< ≤ 1. By the same reason, A∩ Σ> ≤ 1. So, we lose no generality assuming that A ⊂ Σ1 \ ∆.
For each element a ∈ A let
supp+(a) = {n ∈ ω : xn > 0} and supp−(a) = {n ∈ ω : xn < 0}.
It is easy to see that each point a ∈ Σ1 \ ∆ has non-empty negative support supp−(a).
Fix any point b ∈ A. We claim that for any point a ∈ A \ {b} we get supp−(a) ⊂ supp+(b). In the
opposite case the set supp−(a) \ supp+(b) contains some number k ∈ ω and then [a, b[⊂ {(xn)n∈ω ∈
c00 : xk < 0} \ ∆ which is not possible as {a, b} is hidden behind ∆. Since (the power-set of) the set
supp+(b) is finite and A \ {b} is infinite, the Pigeonhole Principle yields two distinct points a, a′ ∈ A
such that supp−(a) = supp−(a′) ⊂ supp+(b). Now we see that for any k ∈ supp−(a) = supp−(a′), we
get [a, a′] ⊂ {(xn)n∈ω ∈ c00 : xn < 0} ⊂ c00 \ ∆, which contradicts the choice of A as a set hidden
behind ∆.
(cid:3)
2. Preliminaries
In this section we prove some lemmas which will be used in the proof of Theorem 1.1.
Lemma 2.1. Let T : X → Y be a linear continuous operator between linear topological spaces. If a
convex subset D ⊂ Y is polyhedral in its closed linear hull lin(D), then its preimage C = T −1(D) is
polyhedral in its closed linear hull lin(C).
Proof. Write the polyhedral set D as a finite intersection
D =
n
\i=1
f−1
i
((−∞, ai])
of closed half-spaces defined by linear continuous functionals f1, . . . , fn : lin(D) → R and real numbers
a1, . . . , an. The continuity of the operator T implies that T (lin(C)) ⊂ lin(D). Consequently, for every
i ≤ n the linear continuous functional gi = fi ◦ T : lin(C) → R is well-defined. Since C = T −1(D) =
Tn
i=1 g−1
An operator A : X → Y between two linear spaces is called affine if
((−∞, ai]), the set C is polyhedral in lin(C).
(cid:3)
i
A(tx + (1 − t)y) = tA(x) + (1 − t)A(y) for any x, y ∈ X and t ∈ R.
It is well-known that an operator A : X → Y is affine if and only if the operator B : X → Y ,
B : x 7→ A(x) − A(0), is linear. The following lemma trivially follows from the definition of a hidden
set.
A CHARACTERIZATION OF POLYHEDRAL CONVEX SETS
3
Lemma 2.2. Let T : X → Y be an affine operator between linear topological spaces, D ⊂ Y be a
convex set, C = T −1(D), and A ⊂ X \ C be a subset such that TA is injective. The set C hides the
set A if and only if the set D = T (C) hides the set T (A).
Let us recall that a convex subset C of a linear topological space X is called a convex body in X
if C has non-empty interior in X.
Lemma 2.3. Let C be an infinite-dimensional closed convex subset of a complete linear metric space
Y . If C is infinite-dimensional, then there is an injective continuous affine operator T : l2 → Y such
that T −1(C) is a closed convex body in l2.
We lose no generality assuming that the convex set C contains the neutral element 0 of Y . In this
Proof. By [9, 1.2.2], the topology of Y is generated by a complete invariant metric d such that the
F -norm kyk = d(y, 0) has the property ktyk ≤ kyk for all y ∈ Y and t ∈ [−1, 1].
case for any points yn ∈ C, n ∈ ω, and any non-negative real numbers tn, n ∈ ω, with Pn∈ω tn ≤ 1
we get Pn∈ω tnyn ∈ C whenever the series P∞n=0 tnyn converges in Y .
The set C is infinite dimensional and hence contains a linearly independent sequence (yn)∞n=1.
Multiplying each yn by a small positive real number, we can additionally assume that kynk ≤ 2−n.
4n yn converges in Y and its sum s0 = P∞n=1
It follows that series P∞n=1
1
4n yn belongs to the closed
convex set C as P∞n=1
3 ≤ 1.
2 be the linear hull of the standard orthonormal basis (en)n∈ω in the separable Hilbert space
l2. Define a linear operator S : l2 → Y letting S(en) = 1
4n yn for every n ∈ N. The convergence of the
seriesP∞n=1 kynk implies that the operator S is continuous and hence can be extended to a continuous
linear operator ¯S : l2 → Y . Let B1 = {x ∈ l2 : kxk < 1} denote the open unit ball in the Hilbert
space l2. We claim that ¯S(B1) + s0 ⊂ C. Indeed, for every x = (xn)∞n=1 ∈ B1 and every n ∈ N we get
xn ≤ 1 and hence
4n = 1
Let lf
1
1
1
4n +
xn
4n ≥
1
4n −
1
4n = 0.
Taking into account that
and 0 ∈ C, we conclude that
1
4n +
∞
Xn=1(cid:0)
xn
4n(cid:1) ≤
∞
Xn=1
2
4n =
2
3 ≤ 1
s0 + ¯S(x) =
∞
Xn=1
1
4n yn +
xn
4n yn ∈ C.
∞
Xn=1
Let H = ¯S−1(0) be the kernel of the operator ¯S and H⊥ be the orthogonal complement of H in
the Hilbert space l2. It follows that the affine operator T : H⊥ → Y , T : x 7→ ¯S(x) + s0, is injective
and the preimage T −1(C) contains the unit ball B1 ∩ H⊥ of the Hilbert space H⊥. So, T −1(C) is a
closed convex body in the Hilbert space H⊥. Since ¯S(l2) = ¯S(K⊥) ⊃ {yn}n∈ω, the Hilbert space H⊥
is infinite-dimensional and hence can be identified with l2.
(cid:3)
The following lemma is the most important and technically difficult ingredient of the proof of
Theorem 1.1.
Lemma 2.4. If a closed convex body ¯C in a separable Hilbert space X is not polyhedral, then ¯C hides
some infinite subset A ⊂ X \ ¯C.
Proof. Let h·,·i denote the inner product of the Hilbert space X. Each element y ∈ X determines a
functional y∗ : x 7→ hx, yi on X. By the Riesz Representation Theorem [4, 3.4], the operator y 7→ y∗ is a
linear isometry between the Hilbert space X and its dual Hilbert space X∗. By S = {x ∈ X : kxk = 1}
and S∗ = {x∗ ∈ X∗ : kx∗k = 1} we denote the unit spheres in the Hilbert spaces X and X∗,
respectively.
4
T. BANAKH AND I. HETMAN
Let C be the interior of the convex body ¯C and ∂C = ¯C \ C be the boundary of ¯C in the Hilbert
space X. A functional x∗ ∈ S∗ is called supporting to ¯C at a point x ∈ ∂C if x∗(x) = sup x∗(C).
The Hahn-Banach Theorem guarantees that for each point x ∈ ∂C there is a supporting functional
x∗ ∈ S∗ to C at x. If such a supporting functional is unique, then the point x is called smooth. By
the classical Mazur's Theorem [8, 1.20], the set Σ of smooth points is a dense Gδ in the boundary ∂C
of C. By σ : Σ → S∗ we shall denote the function assigning to each smooth point x ∈ Σ the unique
support functional σx ∈ S∗ to C at x. Let us observe that the function σ has closed graph
Γ := {(x, σx) : x ∈ Σ} = (Σ × S∗) ∩ {(x, x∗) ∈ C × S∗ : x∗(x) ≥ sup x∗(C)}
1
in the Polish space Σ× S∗. Let pr1 : Γ → Σ and pr2 : Γ → S∗ be the projections on the corresponding
factors. Observe that the projection pr1 is a bijective and continuous map between Polish spaces.
By the Lusin-Souslin Theorem [6, 15.1], it is a Borel isomorphism, which implies that the map
σ = pr2 ◦ pr−1
: Σ → S∗ is Borel measurable. By Theorem 8.38 [6], there is a dense Gδ-subset G ⊂ Σ
such that the restriction σG is continuous.
Claim 2.5. The image σ(G) is infinite.
Proof. Assume that σ(G) is finite and find functionals f1, . . . , fn ∈ S∗ such that σ(G) = {f1, . . . , fn}.
((−∞, max fi( ¯C)]) is not polyhedral, there is a point x ∈ X \ ¯C such that
Since the set ¯C (Tn
fi(x) ≤ max fi( ¯C) for all i ≤ n. Fix any point x0 ∈ C. Since C is open, fi(x0) < max fi( ¯C) for all
i ≤ n. Since x /∈ ¯C, the segment [x, x0] meets the boundary ∂C at some point y = (1− t)x + tx0 where
t ∈ (0, 1). For the point y we get fi(y) = (1 − t)fi(x) + tfi(x0) < max fi( ¯C). It follows that the set
(−∞, max fi( ¯C)) is an open neighborhood of the point y in X. Since the set G in dense
U =Tn
in ∂C, there is a point z ∈ G ∩ U . Consider the unique supporting functional σz to C at the point z.
The inclusion z ∈ U implies that σz ∈ σ(G) \ {f1, . . . , fn}, which is a desired contradiction.
i=1 f−1
(cid:3)
i=1 f−1
i
i
Depending on the cardinality of the set σ(G) we divide the further proof of Lemma 2.4 into two
Lemmas 2.6 and 2.8.
Lemma 2.6. If the image σ(G) is uncountable, then the set C hides some infinite subset of X.
Proof. The continuous map σG induces a closed equivalence relation E = {(x, y) ∈ G × G : σ(x) =
σ(y)} on the Polish space G. Since this equivalence relation has uncountably many equivalence classes,
the Silver's Theorem [10] yields a topological copy K ⊂ G of the Cantor cube {0, 1}ω such that K
has at most one-point intersection with each equivalence class. This is equivalent to saying that the
restriction σK is injective. The existence of such Cantor set K can be also derived from Feng's
Theorem [5] saying that the Open Coloring Axiom holds for analytic spaces.
For any x ∈ K let yx ∈ S be the unique vector such that σx(z) = hz, yxi for all z ∈ X. For a real
number ε ∈ [0, 1] consider the open subset Λ(x, ε) = {z ∈ K : [x + εyx, z] ∩ C 6= ∅} of the Cantor set
K.
Claim 2.7. For any x ∈ K the sets Λ(x, ε) have the following properties:
(1) Λ(x, ε) ⊃ Λ(x, δ) for any 0 < ε ≤ δ ≤ 1;
(2) Sε∈(0,1] Λ(x, ε) = K \ {x}.
Proof. 1. Fix any numbers 0 < ε ≤ δ ≤ 1 and a point z ∈ Λ(x, δ). By the definition of the set Λ(x, δ),
the segment [x + δyx, z] meets the open convex set C at some point c. Since the points x, z belong to
the convex set ¯C and the point c belongs to its interior C, the triangle
∆ = {tcc + txx + tzz : tc > 0, tx, tz ≥ 0,
tc + tx + ty = 1}
lies in the interior C of ¯C. Since the segment [x + εyx, z] intersects this triangle, it has non-empty
intersection with C.
2. Take any point z ∈ K \ {x}. Since σK is injective, the supporting functionals σx and σy are
distinct. Then the open segment ]x, z[ = [x, z]\{x, z} lies in the interior C of ¯C. In the opposite case,
Assume that for some n, the points xk, k < n, and real numbers εk, k < n, have been constructed.
Consider the intersection Tk<n Λ(xk, εk) and observe that it has positive measure:
Λ(xk, εk)(cid:1) = 1 − µ(cid:0)K \ \k<n
Λ(xk, εk)(cid:1) =
µ(cid:0) \k<n
= 1 − µ(cid:0) [k<n
= 1 −Xk<n
K \ Λ(xk, εk)(cid:1) ≥ 1 −Xk<n
(1 − µ(Λ(xk, εk)) > 1 −Xk<n
µ(K \ Λ(xk, εk)) =
2−k−1 > 0.
A CHARACTERIZATION OF POLYHEDRAL CONVEX SETS
5
2 x + 1
[x, z] ⊂ ∂C and for the midpoint 1
2 z there would exist a supporting functional x∗, which would
be supporting for each point of the segment [x, y]. This is impossible as the points x, z are smooth
and have unique and distinct supporting functionals. This contradiction proves that the segment [x, z]
meets the interior C of ¯C. Then for some positive ε > 0 the segment [x + εyx, z] also meets C, which
implies that z ∈ Λ(x, ε).
(cid:3)
Being homeomorphic to the Cantor cube, the space K carries an atomless σ-additive Borel
probability measure µ. Fix any point x0 ∈ K. Using Claim 2.7(2), find ε0 ∈ (0, 1] such that
µ(Λ(x0, ε0)) > 1 − 2−1. Next proceed by induction and construct a sequence of points (xn)n∈ω and a
sequence of positive real numbers (εn)n∈ω such that for every n ∈ N
(1) xn ∈Tk<n Λ(xk, εk);
(2) µ(Λ(xn, εn)) > 1 − 2−n−1;
(3) [xk + εkyxk, xn + εnyxn] ∩ C 6= ∅ for all k < n;
(4) xn + εnyxn /∈ ¯C.
So, this intersection is not empty and we can select a point xn satisfying the condition (1). For every
k < n the definition of the set Λ(xk, εk) ensures that the segment [xk + εkyxk, xn] meets the interior
C of the convex set ¯C. Consequently, there is ε′n > 0 such that for every εn ≤ ε′n and every k < n
the segment [xk + εkyxk, xn + εnyxn] still meets the open set C. Finally, using Claim 2.7(2), choose a
positive real εn ∈ (0, ε′n] such that µ(Λ(xn, ε′n)) > 1 − 2−n−1. Observe that
σxn(xn + εnyxn) = σxn(xn) + εnσxn(yxn) > max σxn( ¯C) + εn
The conditions (3) and (4) of the inductive construction guarantee that A = {xn + εnyxn}n∈ω is
and hence xn + εnyxn /∈ ¯C. This completes the inductive step.
a required infinite set, hidden behind the convex set ¯C.
Lemma 2.8. If the image σ(G) is countable, then the set ¯C hides some infinite subset of X.
Proof. Denote by F the set of functionals f ∈ σ(G) for which the set f−1(sup f (C))∩C has non-empty
interior in ∂C.
(cid:3)
Claim 2.9. The set F is infinite.
Proof. Assume that the set F is finite and write F = {f1, . . . , fn} for some functionals f1, . . . , fn ∈ S∗.
Since ¯C is not polyhedral,
n
f−1
i
¯C 6=
((−∞, max fi( ¯C)]).
\i=1
Repeating the argument from Claim 2.5, we can find a point y ∈ ∂C such that fi(y) < max fi( ¯C)
((−∞, max fi( ¯C)) is an open neighborhood of y in X. Since
for all i ≤ n. Then U = Tn
G ∩ U ⊂ Sf∈σ(G) f−1(max f ( ¯C)), the Baire Theorem guarantees that for some functional f ∈ σ(G)
the intersection f−1(max f ( ¯C)) ∩ G ∩ U has non-empty interior in G ∩ U . Since G ∩ U is dense in
U∩∂C, the intersection f−1(max f ( ¯C))∩U has non-empty interior in U∩∂C and in ∂C. Consequently,
f ∈ F . Since f−1(max f ( ¯C)) ∩ U 6= ∅, we conclude that f ∈ F \ {f1, . . . , fn}, which is a desired
contradiction.
i=1 f−1
i
(cid:3)
6
T. BANAKH AND I. HETMAN
By Claim 2.9, the set F ⊂ σ(G) ⊂ S∗ is infinite and hence contains an infinite discrete subspace
{fn}n∈ω. By the definition of F , for every n ∈ ω we can choose a point xn ∈ ∂C and a positive real
n (max fn( ¯C)). Here ¯B(xn, εn) = {x ∈ X : kx − xnk ≤ ε}
number εn such that ∂C ∩ ¯B(xn, εn) ⊂ f−1
denotes the closed εn-ball centered at xn. Moreover, since the subspace {fn}n∈ω of S∗ is discrete, we
can additionally assume that ¯B(fn, εn) ∩ ¯B(fm, εm) = ∅ for any distinct n, m ∈ ω. For every n ∈ ω
let yn ∈ S be the unique point such that fn(z) = hz, yni for all z ∈ X. The Riesz Representation
Theorem guarantees that
kyn − ymk = kfn − fmk ≥ εn + εm for all n 6= m.
We shall need the following elementary (but not trivial) geometric fact.
Claim 2.10. For any distinct numbers n, m ∈ ω and a positive real number δn ≤ 1
[xn + δnyn, xm] meets the open convex set C.
Proof. Assume conversely that [xn + δnyn, xm] ∩ C = ∅. Taking into account that f−1
¯B(xn, εn) ⊂ ¯C, we conclude that kxn − xmk ≥ εn. Now consider the unit vector
3 ε2
n the segment
n (fn(xn)) ∩
Since hxm, yni = fn(xm) ≤ max fn( ¯C) = fn(xn) = hxn, yni, we get hxm− xn, yni ≤ 0, which means
that the angle between the vectors yn and i is obtuse. Since [xn + δnyn, xm] ∩ C = ∅, the unit vector
yn is not equal to −i and hence the unit vector
i =
xm − xn
kxm − xnk
.
j =
yn − hi, yni · i
kyn − hi, yni · ik
is well-defined. Let α be the angle between the vectors yn and j. It follows that yn = − sin(α) i +
cos(α) j. Consider the vector y⊥n = cos(α) i + sin(α) j, which is orthogonal to the vector yn. Looking
at the following picture, we can see that the angle α is less than the angle β between the vectors y⊥n
and xm − (xn + δnyn).
j
✻
yn
❈❈❖
❈
❈
❈
❈
α
❈
q
εn
❈
δn
❈
❈
xn
q❳❳❳❳❳❳❳❳❳❳❳❳❳
✘✘✘✘✘✘✘✘✘✘✿ y⊥n
n (fn(xn)) ∩ ¯B(xn, εn) ⊂ ¯C and [xn + δnyn, xm] ∩ C = ∅, the angle β is less
β
✲
i
q
xm
q
Since xn + εny⊥n ∈ f−1
than arctan( δn
εn
). Then
kyn − jk = 2 sin(α/2) ≤ α ≤ β ≤ arctan(δn/εn) ≤
δn
εn ≤
1
3
εn.
between the vectors ym and j.
Next, we evaluate the distance kym − jk. It is clear that kym − jk = 2 sin(γ/2) where γ is the angle
Let us consider separately two possible cases.
1) The vector ym lies in the plane spanned by the vectors i and j. Since fn(x) = hx, ymi is a
supporting functional for C at the point xm, we get hxn − xm, ymi ≤ 0 and hence hi, ymi ≥ 0.
A CHARACTERIZATION OF POLYHEDRAL CONVEX SETS
7
On the other hand, [xn + δnyn, xm] ∩ C = ∅ and f−1
m (fm(xm)) ∩ B(xm, εm) ⊂ ¯C imply that
hxn+δnyn−xm, ymi ≥ 0 and hxn+δnj−xm, ymi ≥ 0. Consequently, γ < π/2 and ym = sin(γ)i+cos(γ)j.
It follows that
−kxn − xmk sin(γ) + δn cos(γ) = hxn − xm + δnj, ymi ≥ 0
kxn−xmk ≤ δn
εn and
δn
kym − jk = 2 sin(γ/2) ≤ γ ≤ tan(γ) ≤
δn
εn ≤
1
3
εn.
and hence tan(γ) ≤
Then
kfn − fmk = kyn − ymk ≤ kyn − jk + kj − ymk ≤
2
3
εn < εn + εm,
which contradicts the choice of the sequence (εk).
2) The vectors i, j, ym are linearly independent. Let k be a vector of unit length in X such that
k is orthogonal to i and j and ym = ai + bj + ck for some real numbers a, b, c. It follows that
n (fn(xn)) ∩ ¯B(xn, εn) ⊂ ¯C. Since fm is a supporting functional to C at the point xm,
xn ± εnk ∈ f−1
we get 0 ≥ hxn ± εnk − xm, ymi = −kxn − xmka ± εnc, which implies
c ≤ kxn − xmk
εn
a.
On the other hand, [xn + δnj, xm] ∩ C = ∅ implies 0 ≤ hxn + δnj − xm, ymi = −kxn − xmka + δnb and
a
b ≤
δn
kxn − xmk
.
Now we see that
kym − jk = 2 sin(γ/2) ≤ tan(γ) =
√a2 + c2
b
≤
≤ δns
a
bs1 + kxn − xmk2
ε2
n
≤
1
kxn − xmk2 +
1
n ≤
ε2
δn
kxn − xmks1 + kxn − xmk2
≤
√2
≤ δns 1
3
δn
εn ≤
= √2
1
ε2
n
ε2
n
ε2
n
+
εn.
Then kfn − fmk = kyn − ymk ≤ kyn − jk + kj − ymk ≤ 1
choice of the sequence (εk). This contradiction completes the proof of Claim 2.10.
3 εn +
√2
3 εn < εn + εm, which contradicts the
(cid:3)
3 ε2
n;
real number δn such that
Now we can continue to prove Lemma 2.8. By induction for every n ∈ ω we shall choose a positive
(1) δn ≤ 1
(2) [xk + δkyk, xn + δnyn] ∩ C 6= ∅ for any k < n.
To start the inductive construction put δ0 = 1
3 ε2
0. Assume that for some n ∈ ω we have constructed
positive real numbers δk, k < n, satisfying the conditions (1) -- (2). By Claim 2.10, for every k < n
the intersection [xk + δkyk, xn] ∩ C is not empty. Since the set C is open, we can choose a positive
δn ≤ 1
3 ε2
n so small that for every k < n the intersection [xk + δkyk, xn + δnyn] ∩ C still is not empty.
This completes the inductive construction.
It follows from (2) that the infinite set A = {xn + δnyn}n∈ω is hidden behind the convex set ¯C. (cid:3)
Lemmas 2.6 and 2.8 complete the proof of Lemma 2.4.
(cid:3)
8
T. BANAKH AND I. HETMAN
3. Proof of Theorem 1.1
i
i
(cid:3)
i=1 f−1
i
i=1 Hi, the set C is polyhedric in aff(C).
Lemma 3.1. If C is polyhedral in lin(C), then C is polyhedric in aff(C).
((−∞, ai]) and observe that its complement aff(C) \ Hi = aff(C) ∩ f−1
The implications (1) ⇒ (2) ⇒ (3) ⇒ (1) of Theorem 1.1 are proved in the following three lemmas.
Proof. If C is polyhedral in lin(C), then C =Tn
((−∞, ai]) for some linear functionals f1, . . . , fn :
lin(C) → R and some real numbers a1, . . . , an. For every i ≤ n consider the convex set Hi =
aff(C) ∩ f−1
((ai, +∞)) also is
convex. Since C =Tn
Lemma 3.2. If a convex subset C of a linear space X is polyhedric in aff(C), then C hides no infinite
subset A ⊂ X \ C.
Proof. Assume conversely that some infinite subset A ⊂ X \ C can be hidden behind C. First we
show that the set A \ aff(C) contains at most two distinct points. Assume conversely that there are
three pairwise distinct points a1, a2, a3 ∈ A \ aff(C). Let P = {t1a1 + t2a2 + t3a3 : t1 + t2 + t3 = 1}
be the affine subspace of X spanned by the points a1, a2, a3. The subspace P has dimension 1 or
2. The intersection P ∩ aff(C) is an affine subspace of P that intersects the open segments ]a1, a2[,
]a1, a3[ and ]a2, a3[ and hence coincides with P , which is not possible as a1, a2, a3 ∈ P \ aff(C). So,
A \ aff(C) ≤ 2 and we lose no generality assuming that A ⊂ aff(C).
Being polyhedric in aff(C), the set C can be written as a finite intersection C = Tn
i=1 Hi of
convex subsets H1, . . . , Hn ⊂ aff(C) having convex complements aff(C) \ Hi, i ≤ n. Since A \ C =
Sn
i=1 aff(C) \ Hi, by the Pigeonhole Principle, there is an index i ∈ {1, . . . , n} such that the convex
set aff(C) \ Hi contains two distinct points a, b ∈ A and hence contains the segment [a, b], which is
not possible as [a, b] meets the set C ⊂ Hi.
Lemma 3.3. If a closed convex subset C of a complete metric linear space X is not polyhedral in
lin(C), then some infinite A ⊂ X \ C can be hidden behind C.
Proof. Assume that C is not polyhedral in lin(C). It is easy to check that
Ker(C) = {x ∈ X : ∀c ∈ C ∀t ∈ R c + tx ∈ C}
(cid:3)
is a closed linear subspace of X and C = C + Ker(C). Let Y = X/Ker(C) be the quotient linear
metric space and Q : X → Y be the quotient operator. By [9, 2.3.1], the operator Q is open and
by [9, 1.4.10], Y is a complete linear metric space. Let D = Q(C). The equality C = C + Ker(C)
implies that C = Q−1(D) and Y \ D = Q(X \ C) is an open set. So, D is a closed convex set in Y .
By Lemma 2.1, the set D is not polyhedral in its closed linear hull lin(D).
If the linear space lin(D) is finite-dimensional, then it is isomorphic to a finite-dimensional Hilbert
space H. Let T : H → lin(D) be the corresponding isomorphism. Since D is not polyhedral in lin(D),
the preimage E = T −1(D) is not polyhedral in the Hilbert space H. Being finite-dimensional, the
closed convex set E is a convex body in its affine hull aff(E) ⊂ H. Then for every e0 ∈ E the convex
set E0 = E − e0 is a convex body in the linear subspace H0 = aff(E) − e0 of H. Since E is not
polyhedral in H, the shift E0 = E − e0 is not polyhedral in the Hilbert space H0. By Lemma 2.4 the
set E0 hides an infinite subset A0 ⊂ H0 \ E0. Then the set E hides the infinite set A0 + e0 and the set
T (E) = D hides the infinite set B = T (A0 + e0). Choose any subset A ⊂ X such that QA : A → B
is bijective. By Lemma 2.2 the infinite set A is hidden behind the convex set C = Q−1(D) and we
are done.
Next, assume that the linear space lin(D) is infinite-dimensional. Then the convex set D also is
infinite-dimensional. By Lemma 2.3, there is a continuous injective affine operator T : l2 → lin(D)
such that E = T −1(D) is a closed convex body in l2. Since Ker(D) = {0}, we get Ker(E) = {0}. This
implies that the set E is not polyhedral in the Hilbert l2. By Lemma 2.4, the convex set E hides some
infinite subset A0 ⊂ l2 \ E. Then the infinite set B = T (A0) ⊂ Y \ D is hidden behind the convex set
D. Choose any subset A ⊂ X such that QA : A → B is bijective. By Lemma 2.2 the infinite set A
is hidden behind the convex set C = Q−1(D) and we are done.
(cid:3)
A CHARACTERIZATION OF POLYHEDRAL CONVEX SETS
9
4. Open Problems
It would be interesting to know whether a relative version of Theorem 1.1 is true.
Problem 4.1. Let C ⊂ D be two closed convex subsets of a complete linear metric space. Is it true
that C hides no infinite subset A ⊂ D \ C if and only if C is polyhedric in D ∩ aff(C)?
In fact, the notions of polyhedric and hidden sets can be defined in a general context of convex
structures, see [11]. Let us recall that a convex structure on a set X is a family C of subsets of X such
that
• ∅, X ∈ C;
• for any subfamily A ⊂ C the intersection ∩A ∈ C;
• for any linearly ordered subfamily A ⊂ A we get ∪A ∈ C.
called the convex hull of the set A.
For a convex structure (X,C) and a subset A ⊂ C the intersection conv(A) = T{C ∈ C : A ⊂ C} is
We say that a subset C ⊂ X hides a subset A ⊂ X if conv({a, b}) ∩ C 6= ∅ for any two distinct
points a, b ∈ A.
A subset C is polyhedric in a subset D ⊃ C if C =Tn
i=1 Hi for some subsets H1, . . . , Hn ⊂ D such
that Hi, D \ Hi ∈ C for all i ≤ n.
Problem 4.2. Given a convex structure (X,C) (possibly with topology) characterize (closed) convex
sets C ∈ C that hide no infinite subset A ⊂ X \ C.
References
[1] A.D. Alexandrov, Convex polyhedra, Springer-Verlag, Berlin, 2005.
[2] T. Banakh, I. Hetman, A "hidden" characterization of approximatively polyhedral convex sets, preprint.
[3] T. Banakh, I. Hetman, K. Sakai, Topology of the hyperspace of closed convex subsets of a Banach space, preprint.
[4] J. Conway, A Course in Functional Analysis, Springer-Verlag, New York, 1990.
[5] Q. Feng, Homogeneity for open partitions of pairs of reals, Trans. Amer. Math. Soc. 339 (1993), 659 -- 684.
[6] A. Kechris, Classical Descriptive Set Theory, Springer-Verlag, New York, 1995.
[7] V. Klee, Some characterizations of convex polyhedra, Acta Math. 102 (1959) 79 -- 107.
[8] R. Phelps, Convex Functions, Monotone Operators and Differentiability, Springer-Verlag, Berlin, 1993.
[9] S. Rolewicz, Metric Linear Spaces, Reidel Publishing Co., Dordrecht, PWN, Warsaw, 1985.
[10] J. Silver, Counting the number of equivalence classes of Borel and coanalytic equivalence relations, Ann. Math.
Logic 18 (1980), 1 -- 28.
[11] M. van de Vel, Theory of Convex Structures, North-Holland Publ., Amsterdam, 1993.
[12] D. Walkup, R. Wets, A Lipschitzian characterization of convex polyhedra, Proc. Amer. Math. Soc. 23 (1969),
167 -- 173.
Ivan Franko National University of Lviv, Ukraine, and
Jan Kochanowski University, Kielce, Poland
E-mail address: [email protected]
Ivan Franko National University of Lviv
Universytetska 1, Lviv, 79000, Ukraine
E-mail address: [email protected]
|
1101.3219 | 1 | 1101 | 2011-01-17T14:08:17 | Weak Pullbacks of Topological Groupoids | [
"math.FA",
"math.CT"
] | We introduce the category HG, whose objects are topological groupoids endowed with compatible measure theoretic data: a Haar system and a measure on the unit space. We then define and study the notion of weak-pullback in the category of topological groupoids, and subsequently in HG. The category HG is the setting for topological groupoidification, which we present in separate papers, and in which the weak pullback is a key ingredient. | math.FA | math | WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
A. CENSOR AND D. GRANDINI
Abstract. We introduce the category HG, whose objects are topological groupoids en-
dowed with compatible measure theoretic data: a Haar system and a measure on the unit
space. We then define and study the notion of weak-pullback in the category of topolog-
ical groupoids, and subsequently in HG. The category HG is the setting for topological
groupoidification, which we present in separate papers, and in which the weak pullback is a
key ingredient.
1. Introduction
The leading actors in this paper are groupoids that we call Haar groupoids1. A Haar
groupoid is a topological groupoid endowed with certain compatible measure theoretic in-
gredients. More precisely, a Haar groupoid is a locally compact, second countable, Hausdorff
groupoid G, which admits a continuous left Haar system λ•, and is equipped with a non-zero
Radon measure µ(0) on its unit space G(0), such that µ(0) is quasi-invariant with respect to
λ•. Maps between Haar groupoids are continuous groupoid homomorphisms, which respect
the extra structure in an appropriate sense. One is naturally led to define a category, which
we denote by HG, the category of Haar groupoids. Section 2 introduces this category.
A general study of the category HG from a purely categorical perspective will be presented
in a separate paper. In this paper we focus on one specific categorical notion, namely the
weak pullback. We first construct the weak pullback of topological groupoids. The weak
pullback of the following given cospan diagram of topological groupoids and continuous
homomorphisms:
S
???????
p
G
T
~~~~~~~
q
is a topological groupoid P along with projections πS : P → S and πT : P → T , which
together give rise to the following diagram (which does not commute):
1
1
0
2
n
a
J
7
1
]
.
A
F
h
t
a
m
[
1
v
9
1
2
3
.
1
0
1
1
:
v
i
X
r
a
S
πS
???????
p
P
G
πT
@@@@@@@
~~~~~~~
q
T
Date: August 2, 2018.
2010 Mathematics Subject Classification. 22A22; 28A50.
Key words and phrases. Groupoid; Haar groupoid; Weak pullback; Haar system; quasi invariant measure;
disintegration.
1A discussion regarding terminology appears at the end of this introduction.
1
2
A. CENSOR AND D. GRANDINI
As a set, P is contained in the cartesian product S × G × T , from which it inherits its
topology. The elements of P are triples of the form (s, g, t), where p(s) and q(t) are not
equal to g, but rather in the same orbit of G via g. More precisely, denoting the range and
source maps of G by rG and dG respectively,
P := {(s, g, t) s ∈ S, g ∈ G, t ∈ T, rG(g) = rG(p(s)) and dG(g) = rG(q(t))}.
The groupoid structure of P is described in Section 3, followed by a discussion of its prop-
In the discrete groupoid setting, our notion of weak pullback reduces to the one
erties.
introduced by Baez et al.
in [2], which in turn generalizes the more familiar notion of a
pullback in the category of sets.
Upgrading the weak pullback from topological groupoids to the category HG requires non-
trivial measure theory and analysis. In Section 4 we construct a Haar system for P . Section
5 is then devoted to creating a quasi invariant measure on P (0). Finally, in Section 6, we
prove that with these additional ingredients, subject to a certain additional assumption, we
indeed obtain a weak pullback in HG.
This paper is part of a project we are currently working on, in which we are extending
groupoidification from the discrete setting to the realm of topology and measure theory.
Groupoidification is a form of categorification, introduced by John Baez and James Dolan. It
has been successfully applied to several structures, which include Feynman Diagrams, Hecke
Algebras and Hall Algebras. An excellent account of groupoidification and its triumphs to
date can be found in [2]. So far, the scope of groupoidification and its inverse process of
degroupoidification has been limited to purely algebraic structures and discrete groupoids.
The category HG provides the setting for our attempt at topological groupoidification, in
which the notion of the weak pullback plays a vital role. This line of research is pursued in
separate papers.
This paper relies heavily on general topological and measure theoretic techniques related
to Borel and continuous systems of measures and their mapping properties. A detailed study
of this necessary background theory appears in our paper [4], from which we quote many
definitions and results and to which we make frequent references throughout this text.
1.1. A note about terminology. Seeking a distinctive name for the groupoids we consider
in these notes and in our subsequent work on topological groupoidification, we opted to call
them "Haar groupoids". These groupoids bear close resemblance to measure groupoids with
Haar measures, as studied by Peter Hahn in [5], following Mackey [6] and Ramsay [10],
leading to the theory of groupoid von-Neumann algebras. Like the groupoids we consider,
measure groupoids carry a measure (or measure class), which admits a disintegration via the
range map, namely what is nowadays known as a Haar system. The main discrepancies are
that we require our groupoids to exhibit a nice topology (locally compact, Hausdorff) and
to be endowed with a continuous Haar system, whereas measure groupoids need only have
a Borel structure in general, and host Borel Haar systems.
Locally compact topological groupoids which may admit continuous Haar systems are as
well studied in the literature as measure groupoids, in particular as part of groupoid C ∗-
algebra theory as developed by Jean Renault in [11] (other standard references include [7]
and [8]). In many cases locally compact groupoids indeed exhibit the full structure of our
Haar groupoids, yet the literature does not single them out terminology-wise.
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
3
2. Preliminaries and the category HG
We begin by fixing notation. We shall denote the unit space of a groupoid G by G(0) and
the set of composable pairs by G(2). The range (or target) and domain (or source) maps of
G are denoted respectively by r and d, or by rG and dG when disambiguation is necessary.
We set Gu = {x ∈ G r(x) = u}, Gv = {x ∈ G d(x) = v} and Gu
v = Gu ∩ Gv, for all
u, v ∈ G(0). Thus Gu
u is the isotropy group at u.
We let G = G(0)/G = {[u] u ∈ G(0)} denote the orbit space of a groupoid G. The
orbit space G inherits a topology from G via G(0), defined by declaring W ⊆ G to be open
whenever q−1(W ) is open in G(0), where q : G(0) −→ G is the quotient map u 7→ [u].
Throughout this paper, we will assume our topological groupoids to be second countable,
locally compact and Hausdorff. Any such groupoid G is metrizable and normal, and satis-
fies that every locally finite measure is σ-finite. Moreover, G is a Polish space and hence
strongly Radon, i.e. every locally finite Borel measure is a Radon measure. For more on
Polish groupoids, we refer the reader to a paper by Ramsay [9]. In general, however, G does
not necessarily inherit these properties, a fact that will require occasional extra caution.
Haar systems for groupoids play a key role in this paper.
In the groupoid literature,
modulo minor discrepancies between various sources (see for example standard references
such as [7], [8], [11] and [1]), a continuous left Haar system is usually defined to be a family
λ = {λu : u ∈ G(0)} of positive (Radon) measures on G satisfying the following properties:
(1) supp(λu) = Gu for every u ∈ G(0);
(2) for any f ∈ Cc(G), the function u 7→R f dλu on G(0) is in Cc(G(0));
(3) for any x ∈ G and f ∈ Cc(G),R f (xy)dλd(x)(y) =R f (y)dλr(x)(y).
In this paper we shall use Definition 2.1 below as our definition of a Haar system. It is taken
from [4], where it is shown to be equivalent to the more common definition above. For the
convenience of the reader we include here a very brief summary of the notions from [4] that
lead to Definition 2.1, all of which we will use extensively throughout this paper. Henceforth,
as in [4], all topological spaces are assumed to be second countable and T1 in general, and
also locally compact and Hausdorff whenever dealing with continuous systems of measures.
Let π : X → Y be a Borel map. A system of measures ([4], Definition 2.2) on π is a family
of (positive, Borel) measures λ• = {λy}y∈Y such that:
(1) Each λy is a Borel measure on X;
(2) For every y, λy is concentrated on π−1(y).
We will denote a map π : X → Y admitting a system of measures λ• by the diagram
X
/ Y .
π
λ•
We will say that a system of measures λ• is positive on open sets ([4], Definition 2.3) if
λy(A) > 0 for every y ∈ Y and for every open set A ⊆ X such that A ∩ π−1(y) 6= ∅. A
system of measures λ• on a continuous map π : X → Y will be called a continuous system
of measures or CSM ([4], Definition 2.5) if for every non-negative continuous compactly
supported function 0 ≤ f ∈ Cc(X), the map y 7→RX f (x)dλy(x) is a continuous function on
Y . A system of measures λ• on a Borel map π : X → Y is called a Borel system of measures
or BSM ([4], Definition 2.6) if for every Borel subset E ⊆ X, the function λ•(E) : Y → [0, ∞]
given by y 7→ λy(E) is a Borel function. A system of measures λ• satisfying that every x ∈ X
/
4
A. CENSOR AND D. GRANDINI
has a neighborhood Ux such that λy(Ux) < ∞ for every y ∈ Y , will be called locally finite
([4], Definition 2.14), and locally bounded if there is a constant Cx > 0 such that λy(Ux) < Cx
for any y ∈ Y ([4], Definition 2.3). A detailed discussion of the mutual relations between
the above concepts appears in [4].
Let G be a topological groupoid. A system of measures λ• on the range map r : G → G(0)
is said to be a system of measures on G ([4], Definition 7.1). It is called left invariant ([4],
Definition 7.2) if for every x ∈ G and for every Borel subset E ⊆ G,
λd(x)(E) = λr(x)(cid:0)x · (E ∩ Gd(x))(cid:1) .
Definition 2.1. ([4], Definition 7.5) A continuous left Haar system for G is a system of
measures λ• on G which is continuous, left invariant and positive on open sets.
Playing side by side to the Haar system λ•, another leading actor in our work is a Radon
measure on the unit space G(0) of a groupoid G, which we denote by µ(0). The measure µ(0)
will be related to λ• via the notion of quasi invariance, which we spell out below. We usually
follow [7], where the reader can find much more about the important role of quasi invariant
measures in groupoid theory.
Definition 2.2. Let G be a groupoid admitting a Haar system λ• and a Radon measure µ(0)
on G(0). The induced measure µ on G is defined for any Borel set E ⊆ G by the formula:
µ(E) =ZG(0)
λu(E)dµ(0)(u).
Lemma 2.3. The induced measure µ is a Radon measure on G.
Proof. Since G is strongly Radon, it suffices to prove that µ is locally finite. The induced
measure µ is obtained as a composition of the system λ• with the measure µ(0). The Haar
system λ• is a CSM, hence a locally bounded BSM, by Lemma 2.11 and Proposition 2.23 of
[4]. In addition, the measure µ(0) is locally finite. Therefore, the conditions of Corollary 3.7
in [4] are met, and we conclude that µ is locally finite.
(cid:3)
The following simple observation will be useful in the sequel.
Lemma 2.4. For any Borel function f on G:
ZG
f (x)dµ(x) =ZG(0)(cid:18)ZG
f (x)dλu(x)(cid:19) dµ(0)(u).
Proof. For every Borel subset E ⊆ G, by Definition 2.2,
ZG
χE (x)dµ(x) = µ(E) =ZG(0)
λu(E)dµ(0)(u) =ZG(0)(cid:18)ZG
Generalizing from χE to any Borel function f is routine.
χE (x)dλu(x)(cid:19) dµ(0)(u).
(cid:3)
The image of µ under inversion is defined by
µ−1(E) := µ(E −1) = µ({x−1 x ∈ E})
for any Borel set E ⊆ G.
Remark 2.5. It is a standard exercise to show that for any Borel function f ,
ZG
f (x)dµ−1(x) =ZG
f (x)dµ(x−1).
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
5
Definition 2.6. Let G be a groupoid admitting a Haar system λ• and a Radon measure
µ(0) on G(0). The measure µ(0) is called quasi invariant if the induced measure µ satisfies
µ ∼ µ−1.
Here ∼ denotes equivalence of measures in the sense of being mutually absolutely continuous.
Remark 2.7. Let µ(0) be quasi invariant. The Radon-Nikodym derivative ∆ = dµ/dµ−1 is
called the modular function of µ. Although ∆ is determined only a.e., it can be chosen ([7],
Theorem 3.15) to be a homomorphism from G to R×
+, so we will assume this to be the case.
Recall that for any Borel function f ,
(1)
(2)
f (x)∆(x)dµ−1(x).
f (x−1)dµ(x),
Furthermore, ∆−1 = dµ−1/dµ satisfies the useful formula
ZG
f (x)dµ(x) =ZG
f (x)∆−1(x)dµ(x) =ZG
ZG
since RG f (x)∆−1(x)dµ(x) = RG f (x)dµ−1(x) = RG f (x)dµ(x−1) = RG f (x−1)dµ(x) by Re-
mark 2.5.
Definition 2.8. Let G be a topological groupoid, which satisfies the following assumptions:
(1) The topology of G is locally compact, second countable and Hausdorff.
(2) G admits a continuous left Haar system λ•.
(3) G(0) is equipped with a non-zero Radon measure µ(0) which is quasi-invariant with
respect to λ•.
Such a groupoid will be called a Haar groupoid.
We will denote a Haar groupoid by (G, λ•, µ(0)), or just by G when λ• and µ(0) are evident
from the context.
Definition 2.9. Let (G, λ•, µ(0)) and (H, η•, ν(0)) be Haar groupoids. Let p : G → H be
a continuous groupoid homomorphism which is also measure class preserving with respect
to the induced measures, i.e. p∗(µ) ∼ ν. We say that p is a homomorphism of Haar
groupoids.
In the above definition p∗ is the push-forward, defined for any Borel set E ⊂ H by p∗µ(E) =
µ(p−1(E)). A homomorphism of Haar groupoids is also measure class preserving on the unit
spaces, as we shall shortly see. We first need the following fact.
Lemma 2.10. Let (G, λ•, µ(0)) be a Haar groupoid. The range map r : G → G(0) satisfies
r∗(µ) ∼ µ(0).
Proof. Let E ⊆ G(0) be a Borel subset. We need to show that µ(r−1(E)) = 0 if and only if
µ(0)(E) = 0. By the definition of the induced measure, µ(r−1(E)) =RG(0) λu(r−1(E))dµ(0)(u) =
RG(0) χE (u)λu(G)dµ(0)(u), since λu(r−1(E)) = 0 if u /∈ E whereas λu(r−1(E)) = λu(G) if
u ∈ E. Since λ• is a Haar system, supp(λu) = Gu 6= ∅, and in particular λu(G) > 0 for every
u. It follows that µ(r−1(E)) = 0 if and only if χE (u) = 0 µ(0)-a.e., which is if and only if
µ(0)(E) = 0.
(cid:3)
While the proof we included above is elementary, we point out that Lemma 2.10 also follows
from the fact that by the definition of the induced measure µ, the Haar system λ• is a
6
A. CENSOR AND D. GRANDINI
disintegration of µ with respect to µ(0), which implies that r : G → G(0) is measure class
preserving. See Lemma 6.4 of [4].
Slightly abusing notation, we also denote the restriction of p to G(0) by p.
Proposition 2.11. Let (G, λ•, µ(0)) and (H, η•, ν(0)) be Haar groupoids, and let p : G → H
be a homomorphism of Haar groupoids. Then p∗(µ(0)) ∼ ν(0).
Proof. Consider the following commuting diagram:
G
p
H
rG
/ G(0)
p
rH
/ H (0)
Indeed, by Lemma 2.10 applied to H, ν(0)(E) = 0 ⇔ ν(r−1
Let E ⊆ H (0) be a Borel subset. We need to show that µ(0)(p−1(E)) = 0 if and only if
ν(0)(E) = 0.
H (E)) = 0 ⇔
µ(p−1(r−1
H (E))) = 0. At the same time, by Lemma 2.10 applied to G, we have that
µ(0)(p−1(E)) = 0 ⇔ µ(r−1
H (E)) =
r−1
G (p−1(E)), and it follows that ν(0)(E) = 0 ⇔ µ(0)(p−1(E)) = 0.
(cid:3)
G (p−1(E))) = 0. Since the diagram commutes, p−1(r−1
Having defined Haar groupoids and their appropriate maps, we are ready to define the
setting for this paper and its sequels.
Definition 2.12. We introduce the category HG, which has Haar groupoids as objects and
homomorphisms of Haar groupoids as morphisms.
3. The topological weak pullback
The purpose of this paper is to construct and study the weak pullback of Haar groupoids.
We start by constructing the weak pullback of topological groupoids. We shall leave it to the
reader to verify that in the case of discrete groupoids, our notion of weak pullback reduces to
the one in [2], which in turn generalizes the more familiar notion of pullback in the category
of sets. Examples 3.4 and 3.5 below illustrate that the weak pullback is a natural notion.
Definition 3.1. Given the following diagram of topological groupoids and continuous homo-
morphisms
S
???????
p
G
T
~~~~~~~
q
we define the weak pullback to be the topological groupoid
P = {(s, g, t) s ∈ S, g ∈ G, t ∈ T, rG(g) = rG(p(s)) and dG(g) = rG(q(t))}
together with the obvious projections πS : P → S and πT : P → T . We describe the groupoid
structure of P and its topology below.
/
/
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
7
The weak pullback groupoid P gives rise to the following diagram:
S
πS
???????
p
P
G
πT
@@@@@@@
~~~~~~~
q
T
Observe that even at the level of sets, this diagram does not commute. However, it is not
hard to see that the weak pullback does make the following diamond commute:
S
πS
>>>>>>>
π◦p
P
G
πT
@@@@@@@
π◦q
T
where π : G −→ G is the map g 7−→ [r(g)] = [d(g)].
Intuitively, we think of an element (s, g, t) in P as giving rise to the following picture in
G:
p(s)
q(t)
g
Composition of (s, g, t) and (σ, h, τ ) is then thought of as:
p(σ)
q(τ )
p(s)
q(t)
h
g
Formally, the composable pairs of P are
P (2) = {(s, g, t), (σ, h, τ ) rS(σ) = dS(s), rT (τ ) = dT (t) and h = p(s)−1gq(t)}.
The product is given by
and the inverse is given by
(s, g, t)(σ, h, τ ) = (sσ, g, tτ ),
(s, g, t)−1 = (s−1, p(s)−1gq(t), t−1).
Thus the range and source maps of P are
rP (s, g, t) = (rS(s), g, rT (t))
o
o
o
o
o
o
8
and
A. CENSOR AND D. GRANDINI
dP (s, g, t) = (dS(s), p(s)−1gq(t), dT (t)).
The unit space of P is
q(t) }.
The topology of P is induced from the Cartesian product S × G × T :
P (0) = {(s, g, t) s ∈ S(0), t ∈ T (0) and g ∈ Gp(s)
X ⊆ P is open ⇔ there exists an open set Z ⊆ S × G × T such that X = Z ∩ P.
The product and inverse of P are continuous with respect to this topology.
n=1, {Bm}∞
Remark 3.2. Let {An}∞
k=1 be countable bases for the topologies of
S, G and T respectively. Then B = {(An × Bm × Ck) ∩ P }∞
n,m,k=1 gives a countable basis B
for the topology of P , consisting of open sets of the form E = (A × B × C) ∩ P , which we
call elementary open sets. Moreover, all finite intersections of sets in B are also of the this
form.
m=1 and {Ck}∞
Lemma 3.3. The groupoid P is locally compact, Hausdorff and second countable.
Proof. The groupoid P is second countable by Remark 3.2, and it is Hausdorff as a subspace
of S × G × T . Let
b : S × G × T −→ G(0) × G(0) × G(0) × G(0)
be the continuous map given by
(σ, x, τ ) 7−→ (rG(p(σ)), rG(x), dG(x), rG(q(τ ))).
Observe that P = b−1(∆ × ∆), where ∆ is the diagonal of G(0) × G(0). Therefore, P is closed
in S × G × T , and therefore it is locally compact.
(cid:3)
The following examples show that the weak pullback of groupoids is a natural notion. A
more detailed study of these examples and many others will appear in a separate paper,
where we discuss the weak pullback in the context of topological and measure theoretic
degroupoidification.
Example 3.4. (weak pullback of open cover groupoids)
Let X, Y and Z be locally compact topological spaces, and let p : Y → X and q : Z → X
be continuous, open and surjective maps. Assume that U = {Uα}α∈A and W = {Wα}α∈A
are locally finite open covers of Y and Z, respectively (with the same indexing set A), and
assume that p(Uα) = q(Wα) for every α ∈ A, defining an open cover V = {Vα}α∈A of X,
where Vα = p(Uα). Consider the regular pullback diagram in the category Top of topological
spaces and continuous functions:
Y ∗Z
Y
πY
yyyyyyyy
"EEEEEEEEE
p
πZ
"EEEEEEEE
yyyyyyyyy
q
Z
X
where Y ∗Z = {(y, z) ∈ Y ×Z p(y) = q(z)}. All sets of the form (Uα×Wβ) ∩ Y ∗Z constitute
an open cover of the pullback space Y ∗Z, which we will denote by U ∗W.
"
"
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
9
Associated to an open cover U of a space Y is a groupoid GU = {(α, y, β) : y ∈ Uα ∩ Uβ}
(called an open cover groupoid, or Cech groupoid). A pair (α, y, β), (γ, y′, δ) is composable
if and only if β = γ and y = y′, in which case their product is (α, y, δ), and the inverse is
given by (α, y, β)−1 = (β, y, α). Let GU , GW and GV be the open cover groupoids associated
to the covers of Y , Z and X above, and let bp : GU → GV and bq : GW → GV be the induced
homomorphisms, given by bp(α, y, β) = (α, p(y), β) and bq(α, z, β) = (α, q(z), β). This gives
rise to a cospan diagram of groupoids, which can be completed to a weak pullback diagram:
!CCCCCCCC
}{{{{{{{{
bq
GW
GU
}
BBBBBBBB
bp
P
GV
We omit the technical but straightforward calculations which yield the upshot: the weak
pullback groupoid P is isomorphic to the open cover groupoid GU ∗W corresponding to the
cover U ∗W of the regular pullback space Y ∗Z.
Example 3.5. (weak pullback of transformation groupoids)
Let X, Y and Z be locally compact topological spaces, and let p : Y → X and q : Z → X
be continuous maps. Let Y ∗ Z be the regular pullback in the category Top, as in the
previous example. Let Γ and Λ be locally compact groups acting on Y and Z respectively,
and let Y ×Γ and Z ×Λ be the corresponding transformation groupoids. Recall that in a
transformation groupoid, say Y ×Γ, the elements (y, γ) and (y, γ) are composable if and only
if y = yγ, in which case (y, γ)(yγ, γ) = (y, γγ). The inverse, range and domain are given by
(y, γ)−1 = (yγ, γ −1), r(y, γ) = (y, e) and d(y, γ) = (yγ, e).
We view X as a transformation groupoid by endowing it with an action of the trivial
group, which amounts to regarding X as a cotrivial groupoid. Assume that the maps p and
q are equivariant with respect to the group actions, i.e. p(y · γ) = p(y) and q(z · λ) = q(z). In
this case p and q induce groupoid homomorphisms p : Y ×Γ → X and q : Z ×Λ → X given
by p(y, γ) = p(y) and q(z, λ) = q(z). This yields a cospan diagram of topological groupoids
which gives rise to the following weak pullback diagram:
Y ×Γ
πY
xxxxxxxxx
"FFFFFFFF
p
P
X
πZ
"FFFFFFFFF
xxxxxxxx
q
Z ×Λ
It is now not hard to verify that the weak pullback groupoid P can be identified with the
transformation groupoid (Y ∗Z)×(Γ×Λ) corresponding to the action of the group (Γ×Λ) on
the regular pullback space (Y ∗Z), given by (y, z) · (γ, λ) = (yγ, zλ).
Remark 3.6. In general, the weak pullback coincides with a regular pullback whenever the
groupoid G in Definition 3.1 is a cotrivial groupoid. This is the case in example 3.5 above.
}
!
}
"
"
10
A. CENSOR AND D. GRANDINI
The following observation will be essential in the sequel.
Lemma 3.7. For any u = (s, g, t) ∈ P (0), the fiber P u is a cartesian product of the form
P u = P (s,g,t) = Ss × {g} × T t.
Proof. We follow the definitions:
P (s,g,t) = {(σ, h, τ ) ∈ P rP (σ, h, τ ) = (s, g, t)}
= {(σ, h, τ ) ∈ P (rS(σ), h, rT (τ )) = (s, g, t)}
= {(σ, h, τ ) ∈ P rS(σ) = s, h = g, rT (τ ) = t}
= {(σ, h, τ ) ∈ P σ ∈ Ss, h = g, τ ∈ T t}.
Note that since (s, g, t) is an element of P (0), any σ ∈ Ss satisfies rG(p(σ)) = p(rS(σ)) =
p(s) = p(rS(s)) = rG(p(s)) = rG(g) and likewise any τ ∈ T t satisfies rG(q(τ )) = dG(g).
Therefore Ss × {g} × T t ⊆ P and thus
P (s,g,t) = {(σ, h, τ ) ∈ P σ ∈ Ss, h = g, τ ∈ T t} = Ss × {g} × T t.
(cid:3)
Proposition 3.8. The projections πS : P → S and πT : P → T are continuous groupoid
homomorphisms.
Proof. The proof is straightforward. For continuity, let A ⊆ S be an open subset. Then
π−1
S (A) is open in P since π−1
S (A) = {(s, g, t) ∈ P πS(s, g, t) ∈ A} = {(s, g, t) ∈ P s ∈
A} = (A × G × T ) ∩ P . Now take ((s, g, t), (σ, h, τ )) ∈ P (2). Then πS((s, g, t)(σ, h, τ )) =
πS(sσ, g, tτ ) = sσ = πS(s, g, t)πS(σ, h, τ ). Also, πS((s, g, t)−1) = πS(s−1, p(s)−1gq(t), t−1) =
s−1 = (πS(s, g, t))−1. Thus πS is a groupoid homomorphism. The proof for πT is similar. (cid:3)
4. A Haar system for the weak pullback
We now assume that S, G and T are Haar groupoids and that the maps p and q are
homomorphisms of Haar groupoids. In order to define the weak pullback of the following
diagram in the category HG, we let P be the weak pullback of the underlying diagram of
topological groupoids, as defined above.
P
λ•
(0)
S , µ
S
S
========
p
T
λ•
(0)
T , µ
T
>>>>>>>>
q
(0)
G
G λ•
G, µ
Our goal is to construct a Haar groupoid structure on P . We start by defining the Haar
system λ•
P . From Lemma 3.7 we know that the r-fibers of P are cartesian products of the
form P u = P (s,g,t) = Ss × {g} × T t. In light of this it is reasonable to propose the following
definition.
Definition 4.1. Let u = (s, g, t) ∈ P (0). Define
P = λ(s,g,t)
λu
P
:= λs
S × δg × λt
T .
We denote λ•
P = {λu
P }u∈P (0).
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
11
Theorem 4.2. The system λ•
P is a continuous left Haar system for P .
Proof. The proof will rely on the technology developed in [4]. We consider the following three
pullback diagrams in the category Top of topological spaces and continuous functions (i.e.
we temporarily forget the algebraic structures of the groupoids involved, and view them only
as topological spaces. Likewise all groupoid homomorphisms are regarded only as continuous
functions):
Diagram A
G(0) ∗ G(0)
/ G(0)
G(0)
u7→[u]
Diagram B
S(0) ∗ T (0)
v7→[v]
/ G
/ T (0)
t7→[q(t)]
Diagram C
S(0)
S ∗ T
s7→[p(s)]
/ G
/ T
τ 7→[q(r(τ ))]
S
σ7→[p(r(σ))]
/ G
Note that in order to lighten notation, we denote the pullback object, for example in Diagram
C, by S ∗ T in place of S ∗G T . By definition
S ∗ T = S ∗G T = {(σ, τ ) ∈ S × T [p(r(σ))] = [q(r(τ ))] in G}
and the maps to S and T are the obvious projections. The topology of S ∗ T is the restriction
of the product topology on S × T .
Using G(0) ∗ G(0), S(0) ∗ T (0) and S ∗ T , we can now construct two more pullback diagrams
(still in Top). Our identifications of the pullback objects in Diagrams D and E with P (0)
and P , respectively, are justified below. A moment's reflection reveals that the maps in these
diagrams are well defined.
Diagram D
P (0)
/ S(0) ∗ T (0)
(s,t)7→(p(s),q(t))
G
x7→(r(x),d(x))
/ G(0) ∗ G(0)
/
/
/
/
/
/
/
/
12
A. CENSOR AND D. GRANDINI
Diagram E
P
/ S ∗ T
(σ,τ )7→(p(r(σ)),q(r(τ )))
G
x7→(r(x),d(x))
/ G(0) ∗ G(0)
In Diagram D we identified the pullback object G ∗G(0)∗GG(0) (S(0) ∗G T (0)) with P (0). Indeed,
G ∗G(0)∗GG(0) (S(0) ∗G T (0)) = {(g, (s, t)) (rG(g), dG(g)) = (p(s), q(t))}
= {(g, (s, t)) rG(g) = p(s) and dG(g) = q(t)}
= {(g, (s, t)) g ∈ Gp(s)
q(t) }
which can obviously be identified, as sets, with our definition of P (0). Moreover, the topology
on the pullback is precisely that of P (0), namely the induced topology from S(0) ×G×T (0).
Similarly, in Diagram E we identified the pullback object G ∗G(0)∗GG(0) (S ∗G T ) with P .
Indeed,
G ∗G(0) ∗GG(0) (S ∗G T ) = {(g, (s, t)) (rG(g), dG(g)) = (p(rS(s)), q(rT (t)))}
= {(g, (s, t)) rG(g) = p(rS(s)) and dG(g) = q(rT (t))}
which can be identified with our definition of P , as sets as well as in Top.
Henceforth, we shall follow Section 5 of [4], where we studied fibred products of systems
of measures. Observe that the results we invoke at this point from [4] only require spaces
to be T1 and second countable. The spaces we consider all satisfy these hypotheses. Using
Diagram C as the front face and Diagram B as the back face, we construct the following
fibred product diagram:
rS ∗rT
S ∗ T
S(0) ∗ T (0)
T (0)
rT
λ•
T
[q]
T
S(0)
rS
λ•
S
[p◦rS]
S
[p]
G
[q◦rT ]
G
The connecting maps are the range maps rT and rS, and they are endowed respectively with
the Haar systems λ•
S, which are continuous systems of measures and therefore locally
finite (see Corollary 2.15 of [4]). It is immediate to see that the compatibility conditions on
the maps of the bottom and the right faces are satisfied. The map rS ∗ rT : S ∗ T → S(0) ∗ T (0)
T and λ•
/
/
/
/
/
/
/
/
/
/
?
?
?
?
?
?
?
?
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
13
is defined by (rS ∗ rT )(s, t) = (rS(s), rT (t)). By Definition 5.1 and Proposition 5.2 of [4], we
obtain a locally finite system of measures (λS ∗ λT )• on rS ∗ rT , where
(λS ∗ λT )(s,t) = λs
S × λt
T .
Moreover, by Proposition 5.5 of [4] it is positive on open sets.
With this at hand, we construct another fibred product diagram. We take Diagram E
as the front face and Diagram D as the back face, and use rS ∗ rT and id : G → G as
the connecting maps. The map rS ∗ rT is equipped with the above locally finite system of
measures (λS ∗ λT )•, whereas the identity map on G naturally admits the system δ• of Dirac
masses, which is trivially locally finite:
P (0)
rP
S(0) ∗ T (0)
rS ∗rT
(λS ∗λT )•
S ∗ T
p∗q
(p◦r)∗(q◦r)
G
(r,d)
G(0) ∗ G(0)
δ•
id
(r,d)
G(0) ∗ G(0)
P
G
It is again easy to see that the compatibility conditions on the maps of the bottom and the
right faces are satisfied. Note that in this last diagram we have identified the map from P
to P (0) with rP , the range map of P .
Resorting once again to Definition 5.1 and Proposition 5.2 of [4], we obtain a locally finite
system of measures (δ ∗ (λS ∗ λT ))• on rP : P → P (0), where
(δ ∗ (λS ∗ λT ))(g,s,t) = δg × (λS ∗ λT )(s,t) = δg × λs
S × λt
T .
We denote this system of measures on rP by λ•
P . Yielding to the original convention of
writing elements of P as (s, g, t) rather than (g, s, t), we write λ(s,g,t)
T . Our
construction of λ•
P as a fibred product of the systems δ• and (λS ∗ λT )•, which are locally
finite and positive on open sets, guarantees (by Propositions 5.2 and 5.5 of [4]) that λ•
P
inherits these properties.
S × δg × λt
= λs
P
Recall that as we have pointed out in the preliminaries, G need not be a Hausdorff space
in general. Moreover, S ∗ T , for example, need not be locally compact, as it is not necessarily
closed in S ×T . The assumption that all spaces are locally compact and Hausdorff is essential
in the CSM setting in [4]. For this reason we cannot simply use Proposition 5.4 of [4] to
deduce that as fibred products, (λS ∗ λT )• and subsequently λ•
P are CSMs. Thus, we present
a separate direct proof that λ•
P is a CSM in Proposition 4.3 below. Furthermore, at this
point we return to viewing P , G, S and T as groupoids, and in Proposition 4.4 we state and
prove that λ•
P is a continuous left Haar system for the
groupoid P .
(cid:3)
P is left invariant. We conclude that λ•
Proposition 4.3. The system λ•
P is a continuous system of measures.
/
/
/
/
/
/
/
/
?
?
?
?
?
?
?
?
14
A. CENSOR AND D. GRANDINI
Proof. From the definition of a CSM, in order to prove that λ•
we need to show that for any 0 ≤ f ∈ Cc(P ), the map (s, g, t) 7→RP f (σ, x, τ )dλ(s,g,t)
is a continuous function on P (0).
P is a CSM on rP : P → P (0),
(σ, x, τ )
Let 0 ≤ f ∈ Cc(P ). Recall from the proof of Lemma 3.3 that P is closed in S×G×T . By
Tietze's Extension Theorem, there exists a function F ∈ C(S ×G×T ) such that F P = f .
Since we can multiply F by a function ϕ ∈ Cc(S×G×T ) which satisfies ϕ = 1 on K = supp(f ),
we can assume, without loss of generality, that F ∈ Cc(S ×G×T ).
P
G, Y = T , Z = T (0) and γ • = λ•
We now resort to (symmetric versions of) Lemma 4.5 in [4]. First we take X = S ×
T , to deduce that the function F1 defined by (σ, x, t) 7→
T (τ ) is in Cc(S×G×T (0)). Next, taking X = S×T (0), Y = G, Z = G and γ • = δ•,
RT F (σ, x, τ )dλt
we get that the function F2 defined by (σ, g, t) 7→RG F1(σ, x, t)dδg(x) is in Cc(S ×G×T (0)).
function F3 defined by (s, g, t) 7→RS F2(σ, g, t)dλs
S, Lemma 4.5 of [4] implies that the
S(σ) is in Cc(S(0)×G×T (0)). Merging these
Finally, with X = G×T (0), Y = S, Z = S(0) and γ • = λ•
results, we can rewrite the function F3 by
(s, g, t) 7−→ZSZGZT
F (σ, x, τ ) dλt
T (τ )dδg(x)λs
S(σ).
T ) = r−1
Note that in the above integral rS(σ) = s and rT (τ ) = t, since supp(λs
S (s) and
T (t). Therefore, if we take (s, g, t) ∈ P (0), in which case p(s) = rG(g) and
supp(λt
q(t) = dG(g), we get that p(rS(σ)) = rG(g) and q(rT (τ )) = dG(g). In other words, when
restricting F3 to P (0), we are actually integrating over P . Recalling the definition of λ•
P
(σ, x, τ ),
which is continuous on P (0) as a restriction of a continuous function on S(0) ×G×T (0). (cid:3)
and that F P = f , we retrieve precisely the function (s, g, t) 7→RP f (σ, x, τ )dλ(s,g,t)
S) = r−1
P
Proposition 4.4. The system λ•
P is left invariant.
Proof. From the definition of left invariance, we need to show that
(3)
λdP (x)
P
(E) = λrP (x)
P
for every x ∈ P and for every Borel subset E ⊆ P .
(cid:0)x · (E ∩ P dP (x))(cid:1) ,
Assume first that E is a set of the form E = (A×B×C)∩P , where A ⊆ S, B ⊆ G and C ⊆ T .
Let x = (σ, y, τ ) ∈ P , so rP (x) = (rS(σ), y, rT (τ )) and dP (x) = (dS(σ), p(σ)−1yq(τ ), dT (τ )).
We will denote z = p(σ)−1yq(τ ). We calculate the left and right hand sides of (3) separately.
On the one hand we get:
λdP (x)
P
P
P
(E) = λdP (x)
= λdP (x)
= λdP (x)
= λdS (σ)
= λdS (σ)
P
S
S
P
since λdP (x)
(cid:0)(A × B × C) ∩ P dP (x)(cid:1)
(cid:0)(A × B × C) ∩ (SdS(σ) × {z} × T dT (τ ))(cid:1)
(cid:0)(A ∩ SdS (σ)) × (B ∩ {z}) × (C ∩ T dT (τ ))(cid:1)
(A ∩ SdS (σ)) · δz(B ∩ {z}) · λdT (τ )
(A) · δz(B) · λdT (τ )
(C)
T
T
(C ∩ T dT (τ ))
is concentrated on P dP (x)
by Lemma 3.7
On the other hand,
λrP (x)
P
(cid:0)x · (E ∩ P dP (x))(cid:1) = λrP (x)
= λrP (x)
P
P
(cid:0)(σ, y, τ ) ·(cid:0)(A × B × C) ∩ P dP (x)(cid:1)(cid:1)
(cid:0)(σ, y, τ ) ·(cid:0)(A ∩ SdS(σ)) × (B ∩ {z}) × (C ∩ T dT (τ ))(cid:1)(cid:1)
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
15
By the definition of P (2), note that (σ, y, τ ) ·(cid:0)(A ∩ SdS (σ)) × (B ∩ {z}) × (C ∩ T dT (τ ))(cid:1) can
be nonempty only when z = p(σ)−1yq(τ ) ∈ B, in which case the middle component of the
product is {y}. Hence
P
P
λrP (x)
= (λrP (x)
= (λrS (σ)
= (λdS (σ)
0
0
S
S
(∅)
(cid:0)σ · (A ∩ SdS (σ)) × {y} × τ · (C ∩ T dT (τ ))(cid:1)
(cid:0)σ · (A ∩ SdS (σ))(cid:1) · δy({y}) · λrT (τ )
T
(cid:0)τ · (C ∩ T dT (τ ))(cid:1)
z ∈ B
z /∈ B
z ∈ B
z /∈ B
(A) · λdT (τ )
T
(C)
z ∈ B
z /∈ B
by the left invariance of λ•
S and λ•
T
= λdS (σ)
S
(A) · δz(B) · λdT (τ )
T
(C)
Thus (3) holds for any set E of the form E = (A × B × C) ∩ P .
Fix x ∈ P , and for any Borel subset E of P define
µ(E) = λdP (x)
P
(E)
and
ν(E) = λrP (x)
P
(cid:0)x · (E ∩ P dP (x))(cid:1) .
We claim that µ and ν are both locally finite measures on P . Since λ•
locally finite BSM by Proposition 2.23 of [4]. Hence λu
u ∈ P (0), and in particular µ = λdP (x)
is a locally finite measure.
P is a CSM, it is a
P is a locally finite measure for any
We turn to ν. It is trivial that ν(∅) = 0. Let {Ei}∞
i=1 be a countable collection of disjoint
P
Borel subsets of P .
ν(
∞[i=1
Ei) = λrP (x)
P
(x · ((
∞[i=1
Ei) ∩ P dP (x))) = λrP (x)
P
(x · (
(Ei ∩ P dP (x)))) =
∞[i=1
= λrP (x)
P
(
x · (Ei ∩ P dP (x))) =
λrP (x)
P
∞Xi=1
(cid:0)x ·(cid:0)Ei ∩ P dP (x)(cid:1)(cid:1) =
∞Xi=1
ν(Ei).
∞[i=1
P
P
(cid:0)x · (Uy ∩ P dP (x))(cid:1) = λrP (x)
Therefore ν is countably additive, and hence a measure.
In order to prove that ν is lo-
cally finite we need to show that every y ∈ P admits an open neighborhood Uy such
In the case where y /∈ P dP (x), the open set Uy = P \ P dP (x) satisfies
that ν(Uy) < ∞.
ν(Uy) = λrP (x)
(∅) = 0 < ∞. Now assume that y ∈ P dP (x). In
this case the product z = xy is well defined, and since λrP (x)
is a locally finite measure, there
exists an open neighborhood Uz of z such that λrP (x)
(Uz) < ∞. The map P dP (x) → P de-
fined by w 7→ x · w is continuous, hence there exists an open neighborhood Uy of y such that
(Uz) < ∞.
Finally, let B be a countable basis for the topology of P consisting of elementary open
sets, as in Remark 3.2. As we have just shown, elementary open sets satisfy (3), hence µ
and ν agree on all finite intersections of sets in B. We can now invoke Lemma 2.24 of [4],
which states that if µ and ν are two locally finite measures on a space X, and there exists a
countable basis B for the topology of X such that µ(U1∩U2∩· · ·∩Un) = ν(U1∩U2∩· · ·∩Un) for
any {U1, U2, . . . , Un} ⊂ B, n ≥ 1, then µ(E) = ν(E) for any Borel subset E ⊆ X. Applying
Lemma 2.24 of [4] to µ, ν and B above completes the proof.
(cid:3)
x ·(cid:0)Uy ∩ P dP (x)(cid:1) ⊂ Uz. Consequently, ν(Uy) = λrP (x)
(cid:0)x · (Uy ∩ P dP (x))(cid:1) ≤ λrP (x)
P
P
P
P
16
A. CENSOR AND D. GRANDINI
5. A measure on the unit space of the weak pullback
We return to the weak pullback diagram. Our next task is to construct a measure µ(0)
P
on P (0), and for starters we will need to have certain systems of measures γ •
q on the
maps p and q, respectively. These systems of measures arise via a disintegration theorem,
as we explain below.
p and γ •
λ•
(0)
S , µ
S
S
========
p,γ •
p
P
λ•
P
>>>>>>>>
q,γ •
q
T
λ•
(0)
T , µ
T
G λ•
G, µ
(0)
G
µ(E) =ZY
Let (X, µ) and (Y, ν) be measure spaces, and let f : X → Y be a Borel map. A system of
measures γ • on f will be called a disintegration ([4], Definition 6.2) of µ with respect to ν if
γy(E)dν(y) for every Borel set E ⊆ X. A disintegration theorem gives sufficient
conditions which guarantee the existence of such a disintegration, and the version we will
use appears as Corollary 6.6 of [4]. It requires µ to be locally finite (and σ-finite), ν to be
σ-finite, and f : X → Y to be measure class preserving. Under these conditions there exists
a locally finite BSM γ • on f which is a disintegration of µ with respect to ν.
Each of the Haar groupoids S, G and T is equipped with a Radon (hence locally finite and
σ-finite) measure on its unit spaces, which is quasi-invariant with respect to its Haar system.
The maps p and q are homomorphisms of Haar groupoids, therefore p : S(0) → G(0) and
q : T (0) → G(0) are measure class preserving. These ingredients allow us to invoke Corollary
p on p : S(0) → G(0) which is a disintegration
6.6 of [4], and to obtain locally finite BSMs γ •
of µ(0)
T with
respect to µ(0)
G .
q on q : T (0) → G(0) which is a disintegration of µ(0)
S with respect to µ(0)
G , and γ •
The following requirement will be essential for our proof of Proposition 5.6 below, which
states that the measure µ(0)
P which we are constructing is locally finite.
Assumption 5.1. We will henceforth assume that the disintegration systems γ •
be taken to be locally bounded.
p and γ •
q can
Remark 5.2. By Lemma 2.11 of [4], a CSM is always locally bounded. Therefore, an
appropriate disintegration theorem that produces a system which is either a CSM or at least
locally bounded would have allowed us to remove Assumption 5.1.
Continuous (hence locally bounded) disintegrations are abundant: Examples include dis-
integrations of Lebesgue measures along maps from Rn to Rm, as well as fiber bundles that
admit a continuous disintegration of a measure on the total space with respect to a mea-
sure on the base space. Seda shows that more general constructions of fiber spaces also
host continuous disintegrations, see Theorem 3.2 of [12]. In our context, a Haar system is
of course a continuous disintegration of the induced measure with respect to the measure
on the unit space. A very general result (see Theorem 5.43 of [7], which is a corollary of
Theorem 3.3 of [3]) states that any continuous and open map f : X → Y between second
countable locally compact Hausdorff spaces, admits a continuous system of measures γ •. In
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
17
particular this implies that if ν is a measure on Y and we define the measure µ on X via γ •
γy(E)dν(y), then γ • is a continuous disintegration of µ with respect to ν.
by µ(E) =ZY
The next step is to construct a BSM on the projection πG : P (0) → G, using γ •
p and γ •
q .
Proposition 5.3. The projection πG : P (0) → G admits a locally finite BSM η•, given by
ηx = γr(x)
p × δx × γd(x)
q
.
Proof. We form the following fibred product diagram in the category Top, with Diagram B
as the front face and Diagram A as the back face. The connecting maps are p : S(0) → G(0)
and q : T (0) → G(0), equipped with the locally finite BSMs γ •
q constructed above.
The compatibility conditions on the maps of the bottom and the right faces are easily seen
to be satisfied.
p and γ •
G(0) ∗ G(0)
G(0)
p∗q
q
γ •
q
S(0) ∗ T (0)
T (0)
G(0)
p
γ •
p
S(0)
[p]
G
[q]
G
We point out that the results we use from [4] throughout this proof do not require spaces
to be locally compact and Hausdorff. By Proposition 5.2 in [4], we obtain from the above
diagram the locally finite BSM (γp ∗ γq)• on p ∗ q : S(0) ∗ T (0) → G(0) ∗ G(0), where
(γp ∗ γq)(u,v) = γu
p × γv
q .
Next, we consider the following pullback diagram in Top (this was Diagram D in the
proof of Theorem 4.2). We equip the map p ∗ q with the BSM (γp ∗ γq)•:
P (0)
/ S(0) ∗ T (0)
πG
G
p∗q
(γp ∗γq)•
(r,d)
/ G(0) ∗ G(0)
We follow Section 4 of [4], where we studied lifting of systems of measures. By Definition
4.1, Remark 4.2 and Proposition 4.4 of [4], we can lift the locally finite BSM (γp ∗ γq)•
and obtain a locally finite BSM ((r, d)∗(γp ∗ γq))• on the projection πG : P (0) → G. We
denote η• = ((r, d)∗(γp ∗ γq))•, and from the definition of lifting it follows that for x ∈ G,
ηx = δx × (γp ∗ γq)(r(x),d(x)) = δx × γr(x)
p × δx × γd(x)
.
This completes the proof.
(cid:3)
, which we rewrite as ηx = γr(x)
p × γd(x)
q
q
/
/
/
/
/
/
/
/
?
?
?
?
?
?
?
?
/
/
18
A. CENSOR AND D. GRANDINI
Lemma 5.4. Let E ⊆ P (0) be a set of the form E = (A × B × C) ∩ P (0), where A ⊆ S(0),
B ⊆ G and C ⊆ T (0). For any x ∈ G,
ηx(E) = γr(x)
p
(A)δx(B)γd(x)
q
(C).
q
)(cid:0)(A × B × C) ∩ P (0)(cid:1) . Clearly if x /∈ B then ηx(E) = 0. If x ∈ B then, since δx
)(cid:0)(A × {x} × C) ∩ P (0)(cid:1) .
Proof. From the definition of η• in Proposition 5.3 above, we have that ηx(E) = (γr(x)
δx × γd(x)
is concentrated on {x}, we can write ηx(E) = (γr(x)
A point (s, x, t) ∈ P (0) whose G component is x, satisfies s ∈ p−1(r(x)) and t ∈ q−1(d(x)),
(A ∩ p−1(r(x))) · δx ({x}) · γd(x)
hence for x ∈ B we have ηx(E) = γr(x)
(C ∩ q−1(d(x))) . Since
supp(γr(x)
) = q−1(d(x)), it follows that for x ∈ B, ηx(E) =
γr(x)
(C).
(cid:3)
(C). We conclude that for any x ∈ G, ηx(E) = γr(x)
) = p−1(r(x)) and supp(γd(x)
(A)δx ({x}) γd(x)
p × δx × γd(x)
(A)δx(B)γd(x)
p ×
p
p
p
p
q
q
q
q
q
We can now cook up a measure µ(0)
P on P (0). The ingredients will be the induced measure
µG from Definition 2.2, as well as η• which we have just constructed.
Definition 5.5. Let B ⊆ P (0) be a Borel subset. Define:
µ(0)
P (B) :=ZG
ηx(B)dµG(x).
Since the open sets of the same form as E constitute a basis for the topology of P (0), we
conclude that µ(0)
(cid:3)
P is locally finite.
Note that an alternative proof of Proposition 5.6 is obtained by arguing that the system
η• is locally bounded (modulo Assumption 5.1), and then applying Corollary 3.7 of [4].
Proposition 5.7. The measure µ(0)
and γ •
q .
P
is independent of the choice of the disintegrations γ •
p
In fact, the measure µ(0)
P can be written as
µ(0)
P = µG ◦ [(r, d)∗(γp ∗ γq)],
as it was obtained by lifting the fibred product of the disintegrations γp and γq to πG : P (0) →
G and then composing with the induced measure of G.
In order for P to be a Haar groupoid, µ(0)
P must be a Radon measure, and in particular
locally finite. This is guaranteed modulo our standing Assumption 5.1.
Proposition 5.6. µ(0)
P is a Radon measure on P (0).
Proof. It suffices to show that µ(0)
P is locally finite. Let A ⊆ S(0), B ⊆ G and C ⊆ T (0) be
open subsets with compact closures and consider the set E = (A × B × C) ∩ P (0), which is
an open subset of P (0). Using the definition of µ(0)
P above along with Lemma 5.4, we get
It thus follows from Assumption 5.1 that
µ(0)
P (E) =ZG
ηx(E)dµG(x) =ZG
P (E) ≤(cid:18)sup
µ(0)
s
γr(x)
p
(C)dµG(x) =ZB
q(C)(cid:19) · µG(B) < ∞.
γt
γs
p(A)(cid:19) ·(cid:18)sup
t
γr(x)
p
(A)δx(B)γd(x)
q
(A)γd(x)
q
(C)dµG(x).
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
19
p and eγq
r(x)(A)eγq
d(x)(C)dλu
G(x)(cid:19) dµ(0)
G (u)
• be two other disintegrations on p and q respectively, and let eµ(0)
µ(0)
G -almost every u in G(0).
Let A ⊆ S(0), B ⊆ G and C ⊆ T (0) be open and let E = (A × B × C) ∩ P (0) be the
corresponding open subset of P (0). By the calculation in the proof of Proposition 5.6 above,
µ(0)
d(x)(C)dµG(x).
Using Lemma 2.4 and the fact that supp(λu
P be
q for
G) = r−1(u) we get
(A)γd(x)
u = γu
u = γu
p
q
γu
G (u)
G (u)
d(x)(C)dλu
d(x)(C)dλu
• and eγq
Proof. Let eγp
the corresponding measure on P (0). By Corollary 6.6 in [4], eγp
P (E) = RB γr(x)
P (E) = RB eγp
P (E) = ZB eγp
eµ(0)
r(x)(A)eγq
= ZG(0) eγp
= ZG(0)
= ZG(0)(cid:18)ZB
(C)dµG(x), and likewise eµ(0)
d(x)(C)dµG(x) = ZG(0)(cid:18)ZB eγp
r(x)(A)eγq
u(A)(cid:18)ZB eγq
G(x)(cid:19) dµ(0)
p (A)(cid:18)ZB eγq
G(x)(cid:19) dµ(0)
G(x)(cid:19) dµ(0)
(A)eγq
= ZG(0)(cid:18)ZB
= ZG(0) eγq
= ZG(0)
= ZG(0)(cid:18)ZB
= ZG(0)(cid:18)ZB
G(x)(cid:19) dµ(0)
G(x)(cid:19) dµ(0)
G(x)(cid:19) dµ(0)
G(x)(cid:19) dµ(0)
(A)eγq
u(C)(cid:18)ZB
q (C)(cid:18)ZB
G(x)(cid:19) dµ(0)
r(x)(C)∆−1
d(x)(C)dλu
G (x)dλu
G (x)dλu
(A)∆−1
G (x)dλu
(C)∆−1
G (x)dλu
(A)γd(x)
(A)∆−1
(C)dλu
γd(x)
p
(A)γr(x)
q
G (u)
γr(x)
p
q
γd(x)
p
γd(x)
p
γd(x)
p
γr(x)
p
γu
Justification for the next step is based on formula (2) of Remark 2.7. The remaining calcu-
lation retraces the previous arguments.
G (u)
G (u)
G (u)
G (u)
G (u) = µ(0)
P (E)
Thus, eµ(0)
P and eµ(0)
eµ(0)
P (E) = µ(0)
P (E) for any open set of the form E = (A × B × C) ∩ P (0). These
sets constitute a countable basis B(0) for the topology of P (0), in analogy to Remark 3.2.
Therefore, since µ(0)
P is locally finite as well. Moreover,
µ(0)
P agree on finite intersections of sets in B(0) as these sets are also in B(0), so
we can now use Lemma 2.24 of [4], as in the proof of Proposition 4.4, and conclude that
P = µ(0)
P .
(cid:3)
P is locally finite, it follows that eµ(0)
The following is a simple observation, whose proof is analogous to the proof of Lemma 2.4,
and thus omitted.
Lemma 5.8. For any Borel function f on P (0):
ZP (0)
f (u)dµ(0)
P (u) =ZG(cid:18)ZP (0)
f (u)dηy(u)(cid:19) dµG(y).
20
A. CENSOR AND D. GRANDINI
In §3 of [4] we defined the composition (β ◦ α)• of BSMs X
p
α•
/ Y
q
β •
/ Z ,
which is characterized by
(4)
ZX
f (x)d(β ◦ α)z(x) =ZY(cid:18)ZX
f (x)dαy(x)(cid:19) dβz(y).
This will be essential for proving the following lemma.
Lemma 5.9. For any Borel function f (y, σ) on G ∗ S,
ZS(0)ZSZG
f (y, σ)dλp(rS(σ))
G
(y)dλs
S(σ)dγu
p (s) =ZGZS(0)ZS
f (y, σ)dλs
S(σ)dγrG(y)
p
(s)dλu
G(y).
Proof. Consider the composition (γp ◦ λS)• of the BSMs S
rS
λ•
S
/ S(0)
p
γ •
p
/ G(0) .
We use this as the right edge in the pull-back diagram below. Following §4 of [4], we lift the
BSM λ•
G to obtain a BSM ((p ◦ rS)∗λG)• on πS : G ∗ S → S, and we lift the BSM (γp ◦ λS)•
to obtain a BSM (r∗
G(γp ◦ λS))• on πG : G ∗ S → G.
G ∗ S
πS
((p◦rS )∗λG)•
/ S
πG (r∗
G(γp ◦λS ))•
p◦rS (γp ◦λS )•
G
rG
λ•
G
/ G(0)
By the definition of lifting,
((p ◦ rS)∗λG)σ = λp(rS(σ))
G
× δσ,
σ ∈ S
and
(rG
∗(γp ◦ λS))y = δy × (γp ◦ λS)rG(y),
y ∈ G.
The above diagram gives rise to two compositions: G ∗ S
πS
((p◦rS )∗λG)•
/ S
p◦rS
(γp ◦λS )•
/ G(0) and
G ∗ S
πG
(r∗
G(γp ◦λS ))•
/ G
rG
λ•
G
/ G(0) . However, proposition 4.8 of [4] states that the above
diagram is a commutative diagram of BSMs, and explicitly,
[(γp ◦ λS) ◦ ((p ◦ rS)∗λG)]• = [λG ◦ (rG
∗(γp ◦ λS)]•,
as BSMs on G ∗ S → G(0). The above equality implies that for any Borel function f (y, σ)
on G ∗ S,
ZG∗S
f (y, σ)d((γp ◦ λS) ◦ ((p ◦ rS)∗λG))u(y, σ) =ZG∗S
f (y, σ)d(λG ◦ (rG
∗(γp ◦ λS))u(y, σ).
/
/
/
/
/
/
/
/
/
/
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
21
We expand the left and the right hand sides of the above equality separately, using repeatedly
the characterization (4) of composition of BSMs above:
f (y, σ)d((γp ◦ λS) ◦ ((p ◦ rS)∗λG))u(y, σ)
f (y, σ)d((p ◦ rS)∗λG)eσ(y, σ)(cid:19) d(γp ◦ λS)u(eσ)
f (y, σ)d((p ◦ rS)∗λG)eσ(y, σ)(cid:19) dλs
S(eσ)dγu
× δeσ)(y, σ)(cid:19) dλs
S(eσ)dγu
(y)dδeσ(σ)(cid:19) dλs
S(eσ)dγu
f (y, σ)d(λp(rS(eσ))
f (y, σ)dλp(rS(eσ))
f (y, σ)dλp(rS(σ))
(y)dλs
S(σ)dγu
p (s)
p (s)
G
G
G
p (s)
p (s)
LHS = ZG∗S
= ZS(cid:18)ZG∗S
= ZS(0)ZS(cid:18)ZG∗S
= ZS(0)ZS(cid:18)ZG∗S
= ZS(0)ZS(cid:18)ZG∗S
= ZS(0)ZSZG
RHS = ZG∗S
f (y, σ)d(λG ◦ (rG
∗(γp ◦ λS))u(y, σ)
= ZG(cid:18)ZG∗S
= ZG(cid:18)ZG∗S
= ZG(cid:18)ZG∗S
= ZG(cid:18)ZS
= ZGZS(0)ZS
f (y, σ)d(rG
∗(γp ◦ λS))ey(y, σ)(cid:19) dλu
G(ey)
f (y, σ)d(δey × (γp ◦ λS)rG(ey))(y, σ)(cid:19) dλu
G(ey)
f (y, σ)dδey(y)d(γp ◦ λS)rG(ey)(σ)(cid:19) dλu
G(ey)
f (y, σ)d(γp ◦ λS)rG(y)(σ)(cid:19) dλu
G(y)
f (y, σ)dλs
S(σ)dγrG(y)
p
(s)dλu
G(y)
Since the above expressions are equal, this yields the desired formula.
(cid:3)
Lemma 5.10. Let f (σ, x, τ ) be a Borel function on P . Then
f (σ, y, τ )dλt
T (τ )dγd(y)
q
(t)dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u).
f (σ, x, τ )dµP (σ, x, τ ) =
ZP
ZG(0)ZGZS(0)ZSZT (0)ZT
f (σ, x, τ )dµP (σ, x, τ ) =
Proof.ZP
= ZP (0)ZP
= ZGZP (0)ZP
f (σ, x, τ )dλ(s,g,t)
P
(σ, x, τ )dµ(0)
P (s, g, t)
(by Lemma 2.4)
f (σ, x, τ )dλ(s,g,t)
P
(σ, x, τ )dηy(s, g, t)dµG(y)
(by Lemma 5.8)
22
A. CENSOR AND D. GRANDINI
Rewriting ηy by Proposition 5.3, and then rewriting λ(s,g,t)
by Definition 4.1, we get
P
f (σ, x, τ )dλ(s,g,t)
P
(σ, x, τ )dγr(y)
p
(s)dδy(g)dγd(y)
q
(t)dµG(y)
f (σ, x, τ )dλs
S(σ)dδg(x)dλt
T (τ )dγr(y)
p
(s)dδy(g)dγd(y)
q
(t)dµG(y)
f (σ, y, τ )dλs
S(σ)dλt
T (τ )dγr(y)
p
(s)dγd(y)
q
(t)dµG(y)
Using Lemma 2.4 again, followed by Fubini's theorem, we have
= ZGZZZS(0)×G×T (0)ZP
= ZGZZZS(0)×G×T (0)ZZZS×G×T
= ZGZZS(0)×T (0)ZZS×T
= ZG(0)ZGZZS(0)×T (0)ZZS×T
= ZG(0)ZGZS(0)ZT (0)ZSZT
f (σ, y, τ )dλs
S(σ)dλt
T (τ )dγr(y)
p
(s)dγd(y)
q
(t)dλu
G(y)dµ(0)
G (u)
f (σ, y, τ )dλt
T (τ )dλs
S(σ)dγd(y)
q
(t)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
We now invoke Proposition 5.6 from [4], which asserts that for locally finite BSMs, fibred
products commute with compositions. We apply this theorem to the following diagram (it is
straightforward to verify that the conditions for the proposition indeed hold. In particular,
λ•
S and λ•
T are locally bounded). We obtain that (γq ∗ γp) ◦ (λT ∗ λS) = (γq ◦ λT ) ∗ (γp ◦ λS).
T ∗ S
rT ∗rS
(λT ∗λS )•
T (0) ∗ S(0)
q∗p
(γq ∗γp)•
G(0) ∗ G(0)
}zzzzzzzzzzzzz
yyyyyyyyyyyy
rS
λ•
S
S(0)
}
G(0)
T
rT
λ•
T
}zzzzzzzzzzzzz
id
/ G
/ T (0)
yyyyyyyyyyyy
id
/ G
/ G(0)
~
p
γ •
p
q
γ •
q
S
G
Therefore, returning to our main calculation, we get
= ZG(0)ZGZS(0)ZSZT (0)ZT
This completes the proof.
f (σ, y, τ )dλt
T (τ )dγd(y)
q
(t)dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
(cid:3)
Proposition 5.11. The measure µ(0)
P is quasi-invariant with respect to λ•
P .
Borel set E ⊆ P by µP (E) =RP (0) λv
Proof. By definition 2.6, we need to show that µP and µ−1
P are mutually absolutely con-
tinuous. We recall from Definition 2.2 that µP is the induced measure, defined for any
P is its image under inversion, i.e.
P (v), and µ−1
P (E)dµ(0)
µ−1
P (E) = µP (E −1). We will prove:
Claim: There exists a function Λ : P → R satisfying Λ(α) > 0 µP -a.e., such that for any
Borel set E ⊆ P , µ−1
χE (α)Λ(α)dµP (α).
P (E) =ZP
/
/
}
/
/
}
/
/
/
/
}
/
/
~
/
/
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
23
It will then follow that µP ∼ µ−1
the modular function of µP .
P , since µP (E) =RP χE (α)dµP (α). In fact, ∆ = Λ−1 will be
We first prove the claim for elementary open subsets of the form E = (A × B × C) ∩ P ,
where A ⊆ S, B ⊆ G and C ⊆ T . Note that the characteristic function χE is the restriction
of the product χA · χB · χC to P .
We denote α = (σ, x, τ ) ∈ P and v = (s, g, t) ∈ P (0). By Lemma 5.10:
µ−1
P (E) = µP (E −1) = ZP
χ
E−1 (σ, x, τ )dµP (σ, x, τ )
= ZG(0)ZGZS(0)ZSZT (0)ZT
= ZG(0)ZGZS(0)ZSZT (0)ZT
= ZG(0)ZGZS(0)ZSZT (0)ZT
Using Lemma 5.9, we obtain
E−1 (σ, y, τ )dλt
χ
T (τ )dγd(y)
q
(t)
dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
χE (σ−1, p(σ)−1yq(τ ), τ −1)dλt
T (τ )dγd(y)
q
(t)
dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
χA(σ−1)χB (p(σ)−1yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t)
dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
χA(σ−1)χB (p(σ)−1yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t)
= ZG(0)ZS(0)ZSZGZT (0)ZT
= ZG(0)ZS(0)ZS
χA(σ−1)ZGZT (0)ZT
dλp(r(σ))
G
(y)dλs
S(σ)dγu
p (s)dµ(0)
G (u)
χB (p(σ)−1yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t)
dλp(r(σ))
G
(y)dλs
S(σ)dγu
p (s)dµ(0)
G (u)
Let f1 be a function on G, defined by the formula
f1(y) =ZT (0)ZT
χB(yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t).
From Lemma 7.3 of [4] we know that a system of measures λ• on a groupoid G is left invariant
if and only if for any x ∈ G and every non-negative Borel function f on G,
(5)
Z f (xy)dλd(x)(y) =Z f (y)dλr(x)(y).
This implies, using x = p(σ)−1 and the above f1, that
ZG
f1(p(σ)−1y)dλp(r(σ))
G
=ZG
f1(y)dλp(d(σ))
G
.
24
A. CENSOR AND D. GRANDINI
Therefore, returning to our main calculation and noting that d(p(σ)−1y) = d(y), we have
µ−1
P (E) = ZG(0)ZS(0)ZS
χA(σ−1)ZGZT (0)ZT
χB (yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t)
dλp(d(σ))
G
(y)dλs
S(σ)dγu
p (s)dµ(0)
G (u)
p is a disintegration of µ(0)
S with respect to µ(0)
G , followed by Lemma 2.4,
χB (yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t)dλp(d(σ))
G
(y)dλs
S(σ)dµ(0)
S (s)
χB (yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t)dλp(d(σ))
G
(y)dµS(σ)
χA(σ−1)χB (yq(τ ))χC (τ −1)dλt
T (τ )dγd(y)
q
(t)dλp(d(σ))
G
(y)dµS(σ)
The measure µ(0)
replace σ−1 by σ at the price of inserting ∆−1
S
S (σ):
is quasi-invariant. Therefore, formula (2) of Remark 2.7 permits us to
χA(σ)χB(yq(τ ))χC (τ −1)∆−1
S (σ)dλt
T (τ )dγd(y)
q
(t)dλp(r(σ))
G
(y)dµS(σ)
Re-expanding dµS and then using Lemma 5.9 again, followed by Lemma 2.4, we have
Using the fact that γ •
we get
χA(σ−1)ZGZT (0)ZT
= ZS(0)ZS
= ZS
= ZSZGZT (0)ZT
χA(σ−1)ZGZT (0)ZT
= ZSZGZT (0)ZT
= ZG(0)ZS(0)ZSZGZT (0)ZT
= ZG(0)ZGZS(0)ZSZT (0)ZT
= ZGZS(0)ZSZT (0)ZT
χA(σ)χB (yq(τ ))χC (τ −1)∆−1
S (σ)dλt
T (τ )dγd(y)
q
(t)
dλp(r(σ))
G
(y)dλs
S(σ)dγu
p (s)dµ(0)
G (u)
χA(σ)χB (yq(τ ))χC (τ −1)∆−1
S (σ)dλt
T (τ )dγd(y)
q
(t)
dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
χA(σ)χB (yq(τ ))χC (τ −1)∆−1
S (σ)dλt
T (τ )dγd(y)
q
(t)
dλs
S(σ)dγr(y)
p
(s)dµG(y)
We now use the quasi-invariance of µ(0)
G and formula (2) of Remark 2.7 to write
= ZGZS(0)ZSZT (0)ZT
χA(σ)χB (y−1q(τ ))χC (τ −1)∆−1
S (σ)∆−1
G (y)dλt
T (τ )dγr(y)
q
(t)
dλs
S(σ)dγd(y)
p
(s)dµG(y)
Next, we apply the characterization (4) preceding Lemma 5.9 above to the compositions
S
rS
λ•
S
/ S(0)
p
γ •
p
/ G(0) and T
rT
λ•
T
/ T (0)
q
γ •
q
/ G(0) . We obtain
= ZGZSZT
χA(σ)χB (y−1q(τ ))χC (τ −1)∆−1
S (σ)∆−1
G (y)d(γq ◦ λT )r(y)(τ )
d(γp ◦ λS)d(y)(σ)dµG(y)
/
/
/
/
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
25
We can now use Fubini's theorem, after which we re-expand the compositions as well as µG:
χA(σ)χB(y−1q(τ ))χC (τ −1)∆−1
S (σ)∆−1
G (y)d(γp ◦ λS)d(y)(σ)
= ZGZTZS
= ZG(0)ZGZT (0)ZTZS(0)ZS
χA(σ)χB (y−1q(τ ))χC (τ −1)∆−1
S (σ)∆−1
G (y)dλs
S(σ)dγd(y)
p
(s)
d(γq ◦ λT )r(y)(τ )dµG(y)
dλt
T (τ )dγr(y)
q
(t)dλu
G(y)dµ(0)
G (u)
By Lemma 5.9 with T, t, τ and q in place of S, s, σ and p, we get
= ZG(0)ZT (0)ZTZGZS(0)ZS
= ZG(0)ZT (0)ZTZGZS(0)ZS
χA(σ)χB (y−1q(τ ))χC (τ −1)∆−1
S (σ)∆−1
G (y)dλs
S(σ)dγd(y)
p
(s)
dλq(r(τ ))
G
(y)dλt
T (τ )dγu
q (t)dµ(0)
G (u)
χA(σ)χ−1
B
(q(τ )−1y)χC (τ −1)∆−1
S (σ)∆−1
G (y)dλs
S(σ)dγd(y)
p
(s)
dλq(r(τ ))
G
(y)dλt
T (τ )dγu
q (t)dµ(0)
G (u)
Let f2 be a function on G, defined by the formula
f2(y) =ZS(0)ZS
χA(σ)χ−1
B
(y)χC (τ −1)∆−1
S (σ)∆−1
G (q(τ ))∆−1
G (y)dλs
S(σ)dγd(y)
p
(s).
Using x = q(τ )−1 and f2 in Equation (5) above, we obtain that
ZG
f2(q(τ )−1y)dλq(r(τ ))
G
=ZG
f2(y)dλq(d(τ ))
G
.
G (q(τ ))∆−1
Recall that we take ∆G to be a groupoid homomorphism (see Remark 2.7). Therefore,
∆−1
G (y). Hence, noting also that
d(q(τ )−1y) = d(y), the left hand side of the above equality gives precisely the last line of our
main calculation. From the right hand side we then get
G (q(τ )−1y) = ∆−1
G (q(τ )−1)∆−1
G (y) = ∆−1
G (q(τ ))∆−1
µ−1
P (E) = ZG(0)ZT (0)ZTZGZS(0)ZS
χA(σ)χ
B−1 (y)χC (τ −1)∆−1
S (σ)∆−1
G (q(τ ))∆−1
G (y)
dλs
S(σ)dγd(y)
p
(s)dλq(d(τ ))
G
(y)dλt
T (τ )dγu
q (t)dµ(0)
G (u)
From the fact that γ •
we get
q is a disintegration of µ(0)
T with respect to µ(0)
G , followed by Lemma 2.4,
= ZT (0)ZTZGZS(0)ZS
= ZTZGZS(0)ZS
χA(σ)χ
B−1 (y)χC (τ −1)∆−1
S (σ)∆−1
G (q(τ ))∆−1
G (y)
dλs
S(σ)dγd(y)
p
(s)dλq(d(τ ))
G
(y)dλt
T (τ )dµ(0)
T (t)
χA(σ)χ
B−1 (y)χC (τ −1)∆−1
S (σ)∆−1
G (q(τ ))∆−1
G (y)
dλs
S(σ)dγd(y)
p
(s)dλq(d(τ ))
G
(y)dµT (τ )
26
A. CENSOR AND D. GRANDINI
Using the quasi-invariance of µ(0)
T and formula (2) of Remark 2.7 gives
χA(σ)χ
B−1 (y)χC (τ )∆−1
S (σ)∆−1
G (q(τ )−1)∆−1
G (y)∆−1
T (τ )
dλs
S(σ)dγd(y)
p
(s)dλq(r(τ ))
G
(y)dµT (τ )
= ZTZGZS(0)ZS
Re-expanding dµT we get:
= ZG(0)ZT (0)ZTZGZS(0)ZS
χA(σ)χ
B−1 (y)χC (τ )∆−1
S (σ)∆−1
G (q(τ )−1)∆−1
G (y)∆−1
T (τ )
q (t)dµ(0)
We invoke Lemma 5.9 once again, with T, t, τ and q in place of S, s, σ and p. We obtain
(s)dλq(r(τ ))
S(σ)dγd(y)
T (τ )dγu
(y)dλt
dλs
G
p
G (u)
= ZG(0)ZGZT (0)ZTZS(0)ZS
χA(σ)χ
B−1 (y)χC (τ )∆−1
S (σ)∆−1
G (q(τ )−1)∆−1
G (y)∆−1
T (τ )
dλs
S(σ)dγd(y)
p
(s)dλt
T (τ )dγr(y)
q
(t)dλu
G(y)dµ(0)
G (u)
By Lemma 2.4 this equals
= ZGZT (0)ZTZS(0)ZS
χA(σ)χ
B−1 (y)χC (τ )∆−1
S (σ)∆−1
G (q(τ )−1)∆−1
G (y)∆−1
T (τ )
dλs
S(σ)dγd(y)
p
(s)dλt
T (τ )dγr(y)
q
(t)dµG(y)
We once again now use the quasi-invariance of µ(0)
G and formula (2) of Remark 2.7 to write
= ZGZT (0)ZTZS(0)ZS
χA(σ)χB (y)χC (τ )∆−1
S (σ)∆−1
G (q(τ )−1)∆−1
G (y−1)∆−1
G (y)∆−1
T (τ )
dλs
S(σ)dγr(y)
p
(s)dλt
T (τ )dγd(y)
q
(t)dµG(y)
Returning to χE and using Lemma 2.4, we get
= ZG(0)ZGZT (0)ZTZS(0)ZS
χE (σ, y, τ )∆−1
S (σ)∆−1
G (q(τ )−1)∆−1
T (τ )dλs
S(σ)dγr(y)
p
(s)
dλt
T (τ )dγd(y)
q
(t)dλu
G(y)dµ(0)
G (u)
As we argued earlier in this calculation, we can change the order of integration:
= ZG(0)ZGZS(0)ZSZT (0)ZT
Finally, we define Λ(σ, y, τ ) = ∆−1
µ−1
P (E) = ZG(0)ZGZS(0)ZSZT (0)ZT
By Lemma 5.10 this equals ZP
χE (σ, y, τ )∆−1
S (σ)∆−1
G (q(τ )−1)∆−1
T (τ )dλt
T (τ )dγd(y)
q
(t)
dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
S (σ)∆−1
G (q(τ )−1)∆−1
T (τ ). We get:
χE (σ, y, τ )Λ(σ, y, τ )dλt
T (τ )dγd(y)
q
(t)
dλs
S(σ)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
elementary open set. In order to complete the proof, we need to show that the claim holds
χE (σ, x, τ )Λ(σ, x, τ )dµP (σ, x, τ ), proving the claim for any
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
27
for any Borel set E ⊆ P . For this, we will invoke Lemma 2.24 of [4], as in the proof of
Proposition 4.4. For any Borel subset E, we define
µ(E) = µ−1
P (E)
and
ν(E) =ZP
χE (α)Λ(α)dµP (α).
P is locally finite and λ•
As in Lemma 2.3, since µ(0)
P is a continuous Haar system, the induced
measure µP is locally finite, hence so is the measure µ. Thus ν is locally finite as well, since
µ(E) = ν(E) for any elementary open set E, and these sets constitute a basis B for the
topology of P by Remark 3.2. Finally, µ and ν agree on finite intersections of sets in B as
these are themselves elementary open sets, so Lemma 2.24 of [4] implies that µ(E) = ν(E)
for all Borel sets. The proof is complete.
(cid:3)
Remark 5.12. In particular, it follows from the above calculation that the modular function
of µP is given by ∆P (σ, x, τ ) = ∆S(σ)∆T (τ )/∆G(q(τ )).
6. The weak pullback of Haar groupoids
We return to the weak pullback diagram, which we have now completed:
P
λ•
(0)
S , µ
S
S
πS
========
p
T
λ•
(0)
T , µ
T
λ•
(0)
P , µ
P
πT
========
q
G λ•
G, µ
(0)
G
P , µ(0)
In order for (P, λ•
P ) to indeed be the weak pullback in the category HG, it must be a
Haar groupoid in the sense of Definition 2.8, and the maps πS : P → S and πT : P → T
need to be homomorphisms of Haar groupoids in the sense of Definition 2.9. The first fact
is an immediate corollary of Theorem 4.2 and Proposition 5.11. The second fact is proved
below.
Corollary 6.1. The groupoid (P, λ•
P , µ(0)
P ) is a Haar groupoid.
Proposition 6.2. The maps πS : P → S and πT : P → T are homomorphisms of Haar
groupoids.
Proof. By lemma 3.8, the maps πS and πT are continuous groupoid homomorphisms.
It
remains to show that they are measure class preserving with respect to the induced measures.
We prove first that (πS)∗(µP ) ∼ µS.
Let Σ ⊆ S be a Borel subset. Using the definition of µP , we have
(πS)∗(µP )(Σ) = µP (π−1
S (Σ)) =ZP (0)
λ(s,g,t)
P
(π−1
S (Σ))dµ(0)
P (s, g, t).
Observe that π−1
λs
S ×δg ×λt
S (Σ) = {(σ, x, τ ) ∈ P σ ∈ Σ} = (Σ × G × T ) ∩ P . Substituting λ(s,g,t)
=
T according to Definition 4.1, and noting that systems of measures are concentrated
P
28
A. CENSOR AND D. GRANDINI
on fibers, we get:
λ(s,g,t)
P
(π−1
S (Σ)) = λ(s,g,t)
(cid:0)(Σ × G × T ) ∩ P (s,g,t)(cid:1)
T )(cid:0)(Σ × G × T ) ∩ (Ss × {g} × T t)(cid:1)
T (T )
P
= (λs
= λs
= λs
S × δg × λt
S(Σ) · δg({g}) · λt
S(Σ)λt
T (T )
(by Lemma 3.7)
Therefore, using Lemma 5.10 and then rewriting ηy by Proposition 5.3, we have
(πS)∗(µP )(Σ) = ZGZP (0)
λs
S(Σ)λt
T (T )dηy(s, g, t)dµG(y)
S(Σ)λt
λs
T (T )dγr(y)
p
(s)dδy(g)dγd(y)
q
(t)dµG(y)
S(Σ)λt
λs
T (T )dγr(y)
p
(s)dγd(y)
q
(t)dµG(y)
We use Fubini's theorem, as well as Lemma 2.4, to obtain
Furthermore, the fact that λu
G is supported on Gu dictates that r(y) = u, hence we get
λs
S(Σ)λt
T (T )dγd(y)
q
(t)dγr(y)
p
(s)dλu
G(y)dµ(0)
G (u)
= ZGZZZS(0)×G×T (0)
= ZGZZS(0)×T (0)
= ZG(0)ZGZS(0)ZT (0)
= ZG(0)ZGZS(0)
= ZG(0)ZS(0)
λs
λs
S(Σ)ZT (0)
S(Σ)ZGZT (0)
h1(u) =ZGZT (0)
We now define a function h1 on G(0) by
T (T )dγd(y)
λt
q
T (T )dγd(y)
λt
q
(t)dγu
p (s)dλu
G(y)dµ(0)
G (u)
(t)dλu
G(y)dγu
p (s)dµ(0)
G (u)
T (T )dγd(y)
λt
q
(t)dλu
G(y).
Since λt
main calculation, we have:
T (T ) > 0 for any t, the function h1(u) is strictly positive on G(0). Returning to our
(πS)∗(µP )(Σ) = ZG(0)ZS(0)
= ZG(0)ZS(0)
λs
S(Σ)h1(u)dγu
p (s)dµ(0)
G (u)
λs
S(Σ)h1(p(s))dγu
p (s)dµ(0)
G (u)
p is concentrated on p−1(u). Finally, γ •
p is a disintegration of µ(0)
S with respect to µ(0)
G ,
since γu
hence
(πS)∗(µP )(Σ) = ZS(0)
S(Σ)h1(p(s))dµ(0)
λs
S (s)
S(Σ)dµ(0)
λs
S (s). It follows that µS(Σ) = 0 if and only if
On the other hand, µS(Σ) = ZS(0)
(πS)∗(µP )(Σ) = 0.
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
29
We turn to πT . Proving that (πT )∗(µP ) ∼ µT will require a detour via the quasi-invariance
of µ(0)
G . Let Ω ⊆ T be a Borel subset. Tracing the line of arguments above, we have
where
λ(s,g,t)
P
(π−1
T (Ω)) = λ(s,g,t)
(πT )∗(µP )(Ω) = µP (π−1
T (Ω))dµ(0)
P (s, g, t),
P
(π−1
λ(s,g,t)
T (Ω)) =ZP (0)
(cid:0)(S × G × Ω) ∩ P (s,g,t)(cid:1)
P
= (λs
= λs
S × δg × λt
S(S)λt
T (Ω)
T )(cid:0)(S × G × Ω) ∩ (Ss × {g} × T t)(cid:1)
Therefore,
(πT )∗(µP )(Ω) = ZGZP (0)
λs
S(S)λt
T (Ω)dηy(s, g, t)dµG(y)
λs
S(S)λt
T (Ω)dγr(y)
p
(s)dδy(g)dγd(y)
q
(t)dµG(y)
λs
S(S)λt
T (Ω)dγr(y)
p
(s)dγd(y)
q
(t)dµG(y)
= ZGZZZS(0)×G×T (0)
= ZGZT (0)ZS(0)
= ZGZT (0)ZS(0)
Using the quasi-invariance of µ(0)
G and formula (2) of Remark 2.7, we get
λs
S(S)λt
T (Ω)∆−1
G (y)dγd(y)
p
(s)dγr(y)
q
(t)dµG(y)
Replacing γr(y)
q
by γu
q as before, and using Lemma 2.4 and Fubini's theorem, we get
= ZG(0)ZGZT (0)
= ZG(0)ZT (0)
λt
λt
T (Ω)ZS(0)
T (Ω)ZGZS(0)
h2(u) =ZGZS(0)
The function h2 on G(0) defined by
S(S)∆−1
λs
G (y)dγd(y)
p
S(S)∆−1
λs
G (y)dγd(y)
p
(s)dγu
q (t)dλu
G(y)dµ(0)
G (u)
(s)dλu
G(y)dγu
q (t)dµ(0)
G (u)
S(S)∆−1
λs
G (y)dγd(y)
p
(s)dλu
G(y)
is positive since λs
our main calculation, we have:
S(S) > 0 for any s and the modular function ∆G is positive. Returning to
(πT )∗(µP )(Ω) = ZG(0)ZT (0)
= ZG(0)ZT (0)
λt
T (Ω)h2(u)dγu
q (t)dµ(0)
G (u)
λt
T (Ω)h2(q(t))dγu
q (t)dµ(0)
G (u)
q is concentrated on q−1(u). Finally, γ •
q is a disintegration of µ(0)
T with respect to µ(0)
G ,
since γu
hence
(πT )∗(µP )(Ω) = ZT (0)
T (Ω)h2(q(t))dµ(0)
λt
T (t)
30
A. CENSOR AND D. GRANDINI
On the other hand, µT (Ω) = ZT (0)
T (Ω)dµ(0)
λt
(πT )∗(µP )(Ω) = 0. This completes the proof.
T (t). It follows that µT (Ω) = 0 if and only if
(cid:3)
Recall our standing Assumption 5.1, by which the maps p and q (restricted to the unit
spaces) admit disintegrations which are locally bounded. As we show in the following propo-
sition, the map πS will automatically inherit this property. However, in order to guarantee
that the map πT admits a disintegration which is locally bounded, we will need another
assumption.
Assumption 6.3. We will assume that the modular function ∆G is locally bounded on G,
in the sense that for every point x ∈ G there exist a neighborhood Ux and positive constants
cx and Cx such that cx < ∆G(y) < Cx for every y ∈ Ux.
Note that ∆−1
G is locally bounded whenever ∆G is locally bounded.
Remark 6.4. If we assume that ∆S and ∆T are also locally bounded in the above sense,
then Remark 5.12 implies that ∆P is locally bounded as well.
Proposition 6.5. The maps πS : P (0) → S(0) and πT : P (0) → T (0) admit disintegrations
which are locally bounded.
Proof. We start with the map πS. We shall use Proposition 6.8 from [4], which provides a
necessary and sufficient condition for admitting a disintegration which is locally bounded:
for any compact set K ⊆ P (0) there must exist a constant CK such that for all Borel sets
Σ ⊆ S(0), µ(0)
S (Σ)) ≤ CK · µ(0)
P (K ∩ π−1
S (Σ).
open subsets with compact closures in S, G and T respectively, such that S =S∞
S∞
n=1 Bn, and T =S∞
Let K ⊆ P (0) be compact. Consider three increasing sequences {An}, {Bn} and {Cn} of
n=1 An, G =
n=1 Cn (such sequences exist in any locally compact second countable
space). The elementary open sets En = (An × Bn × Cn) ∩ P (0) determine an increasing open
cover of P (0) and in particular of K. Since K is compact, K ⊆ Ei for some i. Denoting
K1 = Ai, K2 = Bi and K3 = Ci, we have K ⊆ (K1 × K2 × K3) ∩ P (0) where K1 ⊆ S, K2 ⊆ G
and K3 ⊆ T are each compact.
For any Borel set Σ ⊆ S(0),
µ(0)
P (K ∩ π−1
S (Σ)) ≤ µ(0)
P (((K1 ∩ Σ) × K2 × K3) ∩ P (0))
P ((K1×K2×K3) ∩ π−1
S (Σ)) = µ(0)
γr(x)
p
(K1 ∩ Σ)γd(x)
q
(K3)dµG(x)
= ZK2
where the last equality follows from a calculation as in the proof of Proposition 5.6. Ex-
panding µG we get
= ZG(0)ZK2
≤ ZG(0)ZK2
γr(x)
p
(K1 ∩ Σ)γd(x)
q
(K3)dλu
G(x)dµ(0)
G (u)
γr(x)
p
(Σ)γd(x)
q
(K3)dλu
G(x)dµ(0)
G (u)
WEAK PULLBACKS OF TOPOLOGICAL GROUPOIDS
31
Next, we note that r(x) = u since λu
G is supported on r−1(u), and then rewrite γu
p (Σ):
= ZG(0)ZK2
= ZG(0)ZK2ZS(0)
p (Σ)γd(x)
γu
q
(K3)dλu
G(x)dµ(0)
G (u)
χΣ(s)γd(x)
q
(K3)dγu
p (s)dλu
G(x)dµ(0)
G (u)
We use Fubini's Theorem and note that p(s) = u since γu
which we can collapse the outer two integrals, since γp is a disintegration:
p is supported on p−1(u), after
χΣ(s)γd(x)
q
(K3)dλp(s)
G (x)dγu
p (s)dµ(0)
G (u)
= ZG(0)ZS(0)ZK2
= ZS(0)ZK2
q (K3)(cid:19) ·(cid:18)sup
γu
v
where C =(cid:18)sup
u
χΣ(s)γd(x)
q
(K3)dλp(s)
G (x)dµ(0)
S (s) ≤ C · µ(0)
S (Σ),
λv
G(K2)(cid:19). Both suprema exist since γ •
q and λ•
G are locally
bounded, hence bounded on compact sets.
We turn to the map πT . The proof will be analogous, but will require the use of the
function ∆−1
µ(0)
P (K ∩ π−1
G , which is locally bounded by Assumption 6.3. Let Ω ⊆ T (0).
T (Ω)) ≤ µ(0)
P ((K1×K2×K3) ∩ π−1
T (Ω)) = µ(0)
P ((K1 × K2 × (K3 ∩ Ω)) ∩ P (0))
γr(x)
p
(K1)γd(x)
q
(K3 ∩ Ω)dµG(x)
γd(x)
p
(K1)γr(x)
q
(K3 ∩ Ω)∆−1
G (x) dµG(x)
= ZK2
= ZK −1
≤ ZT (0)ZK −1
2
2
Skipping intermediate calculations which mimic the πS case, we get
χΩ(t)γd(x)
p
(K1)∆−1
G (x)dλq(t)
G (x)dµ(0)
T (t) ≤ D · µ(0)
T (Ω)
where D =(cid:18)sup
u
p (K1)(cid:19) · sup
x∈K −1
2
γu
∆−1
G (x)! ·(cid:18)sup
v
G(K −1
λv
2 )(cid:19). All suprema exist since γ •
p
and λ•
G are bounded on compact sets, and ∆−1
G is locally bounded.
(cid:3)
We thank John Baez, Christopher Walker and most of all Paul Muhly for inspiring dis-
cussions and useful remarks.
Acknowledgments
References
[1] Claire Anantharaman-Delaroche and Jean Renault, Amenable groupoids, Monographies de L'Enseigne-
ment Math´ematique, volume 36, Geneva, 2000.
[2] John C. Baez, Alexander E. Hoffnung and Christopher D. Walker, Higher-dimensional algebra VII:
groupoidification, 2009, preprint arXiv:0908.4305v2 [math.QA].
[3] ´Etienne Blanchard, D´eformations de C ∗-alg`ebres de Hopf, Bull. Soc. Math. France 124 (1996), pp.
141 -- 215.
[4] Aviv Censor and Daniele Grandini, Borel and continuous systems of measures, 2010, preprint
arXiv:1004.3750v1 [math.FA].
[5] Peter Hahn, Haar measure for measure groupoids, Trans. Amer. Math. Soc. vol. 242 (1978), pp. 1 -- 33.
32
A. CENSOR AND D. GRANDINI
[6] George W. Mackey, Ergodic theory, group theory, and differential geometry, Proc. Nat. Acad. Sci.
U.S.A. 50 (1963), pp. 1184 -- 1191
[7] Paul S. Muhly, Coordinates in operator algebras, to appear in CBMS lecture notes series.
[8] Alan L. T. Paterson, Groupoids, inverse semigroups, and their operator algebras, Progress in Mathe-
matics, volume 170, Birkhauser, Boston, 1999.
[9] Arlan B. Ramsay, Polish groupoids, in Descriptive set theory and dynamical systems, London Math.
Soc. Lecture Note Series, volume 277, pp. 259 -- 271, Cambridge University Press, Cambridge, 2000.
[10] Arlan B. Ramsay, Virtual groups and group actions, Advances in Math. 6 (1971), pp. 253 -- 322.
[11] Jean Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, volume 793,
Springer, Berlin, 1980.
[12] Anthony K. Seda, On measures in fibre spaces, Cahiers de Topologie est G´eom´etrie Differentielle
Cat´egoriques, vol.21 no.3 (1980) pp. 247 -- 276.
Aviv Censor, School of Mathematical Sciences, Tel-Aviv University, Tel-Aviv 69978, Is-
rael
E-mail address: [email protected]
Daniele Grandini, Department of Mathematics, University of California at Riverside,
Riverside, CA 92521, U.S.A.
E-mail address: [email protected]
|
1912.06890 | 1 | 1912 | 2019-12-14T17:40:35 | Topological algebras of statistical $\tau$-bounded operators on ordered topological vector spaces | [
"math.FA"
] | In this paper, we introduce statistical bounded set on topological vector space. Also, we consider three classes of bounded operators from topological vector spaces to ordered topological vector spaces. Moreover, we give relations between them and order bounded operators. We give algebraic properties of these operators concerning to the uniform convergence topology. | math.FA | math |
Topological algebras of statistical τ -bounded
operators on ordered topological vector spaces
April 25, 2020
Abdullah Aydın∗ and Muhammed C¸ ınar
Department of Mathematics, Mu¸s Alparslan University, Mu¸s, Turkey
Abstract
In this paper, we introduce statistical bounded set on topological vector space. Also,
we consider three classes of bounded operators from topological vector spaces to ordered
topological vector spaces. Moreover, we give relations between them and order bounded
operators. We give algebraic properties of these operators concerning to the uniform
convergence topology.
1
Introductory Facts
Bounded operators and the statistical convergence are natural and efficient tools in the
theory of functional analysis. A vector lattice that was introduced by F. Riesz [9] is
an ordered vector space that has many applications in measure theory, operator theory,
and applications in economics; see for example [1, 2, 3, 9, 11]. On the other hand, the
statistical convergence is a generalization of the ordinary convergence of a real sequence;
see for example [4, 6, 8]. Studies related to this paper are done by Troitsky [10] where
the bb- and nb-bounded operators were defined between topological vector spaces and by
Aydın [5] in which the ob-bounded operator was defined from vector lattices to locally
solid vector lattice and by Hejazian et al.
[12], where nb- and bb-bounded operators
were defined on topological vector spaces. Also, Albayrak and Pehlivan introduced the
statistical τ bounded on locally solid vector lattice in [6] and Aydın defined st-uτ -closed
set in [4]. Our aim in this paper is to introduce and study some classes of bounded
operators from topological vector spaces to topological ordered vector spaces.
Now, let give some basic notations and terminologies that will be used in this paper.
A neighborhood of an element x in a topological vector space E is a subset of E con-
taining an open set that contains x. Neighborhoods of zero will often be referred to as
zero neighborhoods. Every linear topology τ on a vector space E has a base N of zero
neighborhoods satisfying the following four properties; for each V ∈ N , we have λV ⊆ V
for all scalar λ ≤ 1; for any V1, V2 ∈ N there is another V ∈ N such that V ⊆ V1 ∩ V2;
for each V ∈ N there exists another U ∈ N with U + U ⊆ V ; for any scalar λ and each
V ∈ N , the set λV is also in N ; for much more detail see [1, 2, 10]. In this article, unless
Keywords: statistical bounded set, statistically bounded operator, st-bo-operator, st-bb-operator
2010 AMS Mathematics Subject Classification: 47B65, 46A40, 46H35
∗Corresponding author: [email protected]
1
otherwise, when we mention a zero neighborhood, it means that it always belongs to a
base that holds the above properties. Let E be a real-valued vector space. If there is
an order relation "≤" on E, i.e., it is antisymmetric, reflexive and transitive then E is
called ordered vector space whenever the following conditions hold: for every x, y ∈ E
such that x ≤ y, we have x + z ≤ y + z and αx ≤ αy for all z ∈ E and α ∈ R. An ordered
vector space E is called Riesz space or vector lattice if, for any two vectors x, y ∈ E, the
infimum x ∧ y = inf{x, y} and the supremum x ∨ y = sup{x, y} exist in E. Let E be a
vector lattice. Then, for any x ∈ E, the positive part of x is x+ := x ∨ 0, the negative
part of x is x− := (−x) ∨ 0 and absolute value of x is x := x ∨ (−x). A vector lattice
is called order complete if every nonempty bounded above subset has a supremum (or,
equivalently, whenever every nonempty bounded below subset has an infimum). A vector
lattice is order complete iff 0 ≤ xn ↑≤ x implies the existence of sup xn. For a positive el-
ement a in a vector lattice E, the set [−a, a] = {x ∈ E : −a ≤ x ≤ a} is an order interval.
Also, if any subset in E is included in an order interval then it is called order bounded
set. An order bounded operator between vector lattices E and F sends order bounded
subsets to order bounded subsets, abbreviated as Lb(E; F ); for more detail information
on these notions see [1, 2, 3, 5, 9, 10, 11].
Recall that a subset A of a vector lattice E is called solid if, for each x ∈ A and y ∈ E,
y ≤ x implies y ∈ A. A solid vector subspace of a vector lattice is referred to as an
ideal. An order closed ideal is called a band. Let E be vector lattice and τ be a linear
topology on it. Then (E, τ ) is called a locally solid vector lattice (or, locally solid Riesz
space) if τ has a base which consists of solid sets, for more details on these notions see
[1, 2, 11]. A vector lattice E is called Archimedean whenever 1
n x ↓ 0 holds in E for each
x ∈ E+. In this article, unless otherwise, all vector lattices are assumed to be real and
Archimedean.
Consider a subset K of the set N of all natural numbers. Let's define a new set
Kn = {k ∈ K : k ≤ n}. Then we denote Kn for the cardinality of the set Kn. If the
Kn/n exists then µ(K) is called the asymptotic density of the set
limit of µ(K) = lim
n→∞
K. Let X be a topological space and (xn) be a sequence in X. Then (xn) is said to
be statistically convergent to x ∈ X whenever, for each neighborhood U of x, we have
µ(cid:0){n ∈ N : xn /∈ U }(cid:1) = 0; see for example [6, 8]. Similarly, a sequence (xn) in a locally
solid Riesz space (E, τ ) is said to be statistically τ -convergent to x ∈ E if it is provided
(cid:12) = 0 holds. Let
that, for every τ -neighborhood U of zero,
(xn) be a sequence in a locally solid Riesz space (E, τ ). If there exists some scalar λ > 0
such that µ(cid:0){n ∈ N : λxn /∈ U }(cid:1) = 0 holds for every τ -neighborhood U of zero then we
say that (xn) is statistically τ -bounded; see [6].
n(cid:12)
(cid:12){k ≤ n : (xk − x) /∈ U }(cid:12)
1
lim
n→∞
2 Boundedness
In this section, we give the notion statistical bounded set and introduce st-bo-, st-bb- and
st-bs-bounded operators. The following notion could be known, but since we do not have
a reference for which we give that without reference.
Definition 2.1. Let (X, τ ) be a topological vector space. A subset B ⊆ X is called
statistical bounded (or shortly, st-bounded) set in X if, for every zero neighborhood U ,
there is a scalar λ > 0 such that
lim
n→∞
1
n
(cid:12)
(cid:12){b ∈ B : λb /∈ U }(cid:12)
(cid:12) = 0.
It is clear that subset of an st-bounded set is st-bounded. On the other hand, a
subset B in a topological vector space (E, τ ) is called topological bounded (or shortly,
τ -bounded) if, for every zero neighborhood U in E, there is a positive scalar λ such that
B ⊆ λU . So, we have the following useful fact.
2
Lemma 2.2. Every τ -bounded set in topological vector spaces is st-bounded.
Let X and Y be topological vector spaces. An operator T : X → Y is said to be
bb-bounded if it maps every bounded set into a bounded set; see [10]. Motivated by this
definition and by the ob-bounded operator in [5] and the nb- and bb-bounded operators in
[12], we can give the following notions.
Definition 2.3. Let E be a topological vector space, F be an ordered topological vector
space and T : E → F be an operator. Then T is said to be
(1) st-bo-bounded if it maps every statistically bounded set into order bounded set,
(2) st-bb-bounded if it maps each statistically bounded set into a topological bounded set,
(3) st-bs-bounded if it maps statistically bounded sets into statistically bounded sets.
Consider the following theorems which are two classical results about order bounded
subsets in a locally solid vector lattice; see [1, Theorem 2.19.(i)] and [7, Theorem 2.2.],
respectively.
Theorem 2.4. Let (E, τ ) be a locally solid vector lattice. Then every order bounded subset
in E is τ -bounded.
Theorem 2.5. Let (E, τ ) be an ordered topological vector space that has order bounded
τ -neighborhood of zero. Then every τ -bounded subset is order bounded.
We continue with the following several basic results that follow directly from their
basic definitions and Lemma 2.2, Theorem 2.4 and Theorem 2.5, so its proof is omitted.
Remark 2.6. Let E be a topological vector space and F be an ordered topological vector
space. Then we have that
(i) every ordered bounded set in a locally solid vector lattice is st-bounded,
(ii) the st-bb-boundedness implies st-bs-boundedness,
(iii) the st-bo-boundedness implies st-bb-boundedness if F is a locally solid vector lattice,
(iv) the st-bo-boundedness implies st-bs-boundedness if F is a locally solid vector lattice,
(v) the st-bb-boundedness implies st-bo-boundedness whenever F has order bounded τ -
neighborhood of zero,
(vi) every ordered bounded operator is st-bo-bounded if E is a locally solid vector lattice,
(vii) ordered boundedness implies st-bo-boundedness implies st-bb-boundedness implies st-
bs-boundedness if E = F is a locally solid vector lattice, so we have Lb(E) ⊆
Bst−bo(E) ⊆ Bst−bb(E) ⊆ Bst−bs(E),
(vii) if E = F is a locally solid vector lattice with an order bounded τ -neighborhood of
zero then the st-bo-boundedness and the st-bb are coinciding, so we have Lb(E) ⊆
Bst−bo(E) = Bst−bb(E) ⊆ Bst−bs(E).
We denote Bst−bo(E, F ) the class of all st-bo-bounded operators from a topological
vector space E to an ordered topological vector space F . Also, we denote Bst−bb(E, F )
and Bst−bs(E, F ) the class of all st-bb-bounded and the class of all st-bb-bounded operators
from E to F , respectively. Now, we can allocate the sets Bst−bo(E, F ), Bst−bb(E, F ) and
Bst−bs(E, F ) to the topology of uniform convergence on st-bounded sets. That is, a net
(Tα) in Bst−bo(E, F ) (respectively, in Bst−bb(E, F ) or Bst−bs(E, F )) converges to zero on
an st-bounded set B with this topology if, for each u ∈ F+(respectively, for every zero
neighborhood U or for each st-bounded zero neighborhood V of F ), there is an index α0
such that T (B) ⊆ [−u, u] (respectively, T (B) ⊆ U or T (B) ⊆ V ) for all α ≥ α0. We
should note that these classes of operators are not equal, in general. To see that let's
consider the following example [7, Example 2.3.].
3
Example 2.7. Let E = F be the ordered topological vector space R2 with the lexicographic
ordering and the usual topology. Then the identity operator I on E is st-bo-bounded, but
it fails to be st-bb-bounded. Indeed, consider two elements x = (−1, 0) and y = (1, 0) in
E. Then the order interval [x, y] is clearly an order bounded in E. It can be seen that
this interval includes uncountably many infinite vertical rays. Hence, the order interval
[x, y] is not topological bounded set in E because E is the usual topology and a topological
bounded set cannot contain any infinite vertical ray. Thus, st-bo-boundedness does not
imply st-bb-boundedness in general.
Also, we can consider the following example.
Example 2.8. Let E = F be c0, the space of all null sequences with the usual order
and norm topology. Then the identity operator I on E is st-bb-bounded, but it fails to be
st-bo-bounded. Indeed, consider the unit ball B(0, 1) centered at zero with radius one of
c0. Let's take the sequence (xn) defined by xn = (1, 1, 1, . . . , 1, 0, 0, 0 . . . ). Then it can be
seen that (xn) is st-bounded because topological bounded. But it is not order bounded in
c0. Thus, st-bb-boundedness does not imply st-bo-boundedness in general.
3 Main Results
We give some results about st-bo-, st-bb- and st-bs-bounded operators. In ordered vector
spaces, the summation of finite order bounded sets is order bounded, and the summation
of finite topological bounded sets is bounded in topological vector spaces. Thus, the
summation of st-bo-bounded operators is also st-bo-bounded and the summation of st-bb-
bounded operators is also st-bb-bounded. For the st-bs-bounded operators, we have the
following proposition.
Proposition 3.1. The summation of st-bs-bounded operators S and T from a topological
vector space E to a locally solid vector space F is st-bs-bounded.
Proof. It is enough to show that the summation of finite statistical bounded sets in lo-
cally solid vector lattice is st-bounded. Consider two st-bounded sets B1 and B2 in
a locally solid vector lattice F . For any zero neighborhood U , there exists another
zero neighborhood W such that W + W ⊆ U . Hence, we have some positive scalars
λw1, λw2 > 0 such that µ(Kw1) = µ(Kw2) = 0 for sets Kw1 = {x ∈ B1 : λw1x /∈ W } and
Kw2 = {y ∈ B2 : λw2y /∈ W }. Let choose a positive scalar λ = min{λw1, λw2, 1}. Then
we have λx ≤ λw1x and λy ≤ λw2y for x ∈ B1 and y ∈ B2, respectively. Since W is
solid and absorbing, we have λx, λy ∈ W . Then we have
λx + λy ∈ W + W ⊆ U.
Thus we get µ(cid:0){x + y : x ∈ B1, y ∈ B2 and λ(x + y) /∈ U } = 0. Consequently, B1 + B2 is
a statistically bounded set.
Firstly, we show the continuity of addition and scalar multiplication with the uniform
convergence topology.
Theorem 3.2. The operations of addition and scalar multiplication are continuous in
Bst−bo(E, F ), Bst−bb(E, F ) and Bst−bs(E, F ).
Proof. The continuity of addition and scalar multiplication in Bst−bs(E, F ) can be shown
in the same way in the proof [5, Theorem 2.1.]. So, it is enough to show continuity for
st-bo-operators.
Suppose the nets (Tα) and (Sα) are nets of st-bo-bounded operators that converges
to zero uniformly on st-bounded sets. Fix an arbitrary st-bounded set B in E. Then,
for any f ∈ F+, there are some λ1, λ2 > 0 such that Tα(B) ⊆ [−f, f ] for all α ≥ α1
4
and Sα(B) ⊆ [−f, f ] for each α ≥ α2. So, there is another index α0 such that α0 ≥ α1
and α0 ≥ α1 because the index set of nets is directed. Hence, Tα(B) ⊆ [−f, f ] and
Sα(B) ⊆ [−f, f ] for all α ≥ α0. Then we have
(Tα + Sα)(B) ⊆ [−f, f ] + [−f, f ] = [−2f, 2f ]
for each α ≥ α0. Now, we show convergence for the scalar multiplication. Consider the
st-bounded set B in E. Take a sequence of reals (λn) and assume it converges to zero.
Since (Tα) is uniform convergent to zero on st-bounded sets, for every positive vector
u ∈ F+, there exists an index α0 such that Tα(B) ⊆ [−u, u] for all α ≥ α0. We know that
for enough large n, we have λn ≤ 1, and so λn[−u, u] ⊆ [−u, u]. Then, for all α ≥ α0
and sufficiently large n, we have
λnT (B) = T (λnB) ⊆ λn[−u, u] ⊆ [−u, u].
Therefore, we get the desired result.
We investigate the lattice operations are continuous with the topology of uniform
convergence in the following result.
Theorem 3.3. Let E be a topological vector space and F be an order complete locally solid
vector lattice. Then the lattice operations with the uniform convergence are continuous in
Bst−bo(E, F ), Bst−bb(E, F ) and Bst−bs(E, F ) on statistically bounded sets.
Proof. We show the continuity of the lattice operations in Bst−bs(E, F ). The other cases
can be shown similarly. Suppose the nets (Sα) and (Tα) of st-bs-bounded operators
converge to the linear operators S and T uniformly on st-bounded sets, respectively. By
applying [2, Theorem 1.8], we have
(cid:0)S ∨ T (cid:1)(x) = sup{Sy + T z : y + z = x and y, z ∈ E+}
for every x ∈ E+.
Let's fix an st-bounded set B in E. So, consider another set A = {y ∈ E+ : ∃z ∈
E+, x = y + z for some x ∈ B+}. Then A is also st-bounded.
Indeed, for any zero
neighborhood U , there exists a positive scalar λ such that µ(KB) = 0 for KB = {x ∈
B : λx /∈ U } since B is st-bounded. Consider the set KA = {y ∈ A : λy /∈ U }. For a
vector y ∈ A, we have x ∈ B+ and z ∈ E+ such that y + z = x, and so λy ≤ λx. Then,
by using the solidness of U , we get y ∈ KA whenever x ∈ KB. Thus, one can see that
the carnality of KA can not be bigger than two times the carnality of KB since we have
z ∈ KA if y ∈ KA. As a result, we get µ(KA) = 0 because of µ(KB) = 0. Therefore,
A is statistically bounded set in E. Thus, we have that Sα → S and Tα → T converge
uniformly on A. Take a fixed x ∈ B+ and some elements y, z ∈ E+ such that x = y+z. So,
by the formula sup(M ) − sup(N ) ≤ sup(M − N ) for any set M, N , we have the following
inequality
(cid:0)Sα ∨ Tα(cid:1)(x) − (cid:0)S ∨ T (cid:1)(x) = sup{Sαy + Tαz : y + z = x and y, z ∈ E+}
− sup{Sy + T z : y + z = x and y, z ∈ E+}
≤ sup{(Sα − S)y + (Tα − T )z : y + z = x and y, z ∈ E+}.
Now, consider any st-bounded zero neighborhood V in F . Then there is another zero
neighborhood W such that W + W ⊆ V . One can see that W is also st-bounded. Hence,
since Sα → S and Tα → T converge uniformly on A and indexed set is directed, we have
some index α0 such that (Sα − S)(A) ⊆ W and (Tα − T )(A) ⊆ W for all α ≥ α0. Thus,
we have
(cid:0)Sα ∧ Tα(cid:1)(x) − (cid:0)S ∧ T (cid:1)(x) ≤ (Sα − S)(x) + (Tα − T )(x) ∈ W + W ⊆ V.
for all α ≥ α0. Therefore, Sα ∧ Tα − S ∧ T is st-bs-bounded.
5
It is clear that the product of st-bs-bounded operators is continuous with the topology
of uniform convergence on st-bounded sets because every subset of an st-bounded set is
st-bounded. For the other cases, we give the next theorems.
Theorem 3.4. Let E be a topological vector space, F be a locally solid vector lattice and
T : E → F be an operator. Then the product of st-bo-bounded operators is continuous in
Bst−bo(E, F ) with the topology of uniform convergence on st-bounded sets.
Proof. Suppose the nets of st-bo-bounded operators (Tα) and (Sα) converge to zero uni-
formly on st-bounded sets. Let's take an st-bounded set B. Then, for fixed positive
vector u ∈ E+, there is an index α1 such that Tα(B) ⊆ [−u, u] for all α ≥ α1. Since every
order bounded set in E is topological bounded, we get that [−u, u] is τ -bounded, and so
st-bounded; see Theorem 2.4. Thus, Tα(B) is also st-bounded for each α ≥ α1. Then
Sα(Tα(B)) is order bounded for each α ≥ α1. So, for a given positive u, we have another
index α2 ≥ α1 such that
for all α ≥ α2. Therefore, we get the desired result.
Sα(cid:0)Tα(B)(cid:1) ⊆ [−u, u]
Proposition 3.5. Let E be a topological vector space, F be an ordered topological vector
space. If a net (Tα) of st-bs-bounded operators order converges uniformly on some st-
bounded sets to na operator T then T is st-bs-bounded.
Proof. Suppose B is st-bounded set in E. Then, for each st-bounded zero neighborhood
U in F , there exists an index α0 such that (T − Tα)(B) ⊆ U for each α ≥ α0. So, we get
T (B) ⊆ Tα0(B) + U.
Since Tα0 is st-bs-bounded, we have Tα0(B) is also st-bounded. Thus, Tα0(B) + U is
also st-bounded because sum of st-bounded sets is st-bounded. As a result, T (B) is
st-bounded, and so we get the desired result.
Proposition 3.6. Let E be a topological vector space, F be an ordered topological vector
space. If a net (Tα) of st-bo-bounded operators order converges uniformly on some st-
bounded sets to an operator T then T is st-bo-bounded.
Proof. Let B be st-bounded set in E. Then, for each positive vector u in F , there exists
an index α0 such that (T − Tα)(B) ⊆ [−u, u] for each α ≥ α0. So, we get
T (B) ⊆ Tα0(B) + [−u, u].
Since Tα0 is st-bo-bounded, we have Tα0(B) is also order bounded set in F . Thus, Tα0(B)+
[−u, u] is also order bounded because the sum of order bounded sets is order bounded.
As a result, T (B) is order bounded set in F .
Theorem 3.7. Let E be a topological vector space, F be an ordered topological complete
vector space. If every st-bounded set in E is absorbing then Bst−bs(E, F ) is complete with
the topology of uniform convergence on st-bounded sets.
Proof. Let (Tα) be a Cauchy sequence in Bst−bs(E, F ) with respect to uniform convergence
on st-bounded sets. consider an arbitrary st-bounded set B in E. Hence, for an st-
bounded zero neighborhood U in F , there exists an α0 such that (Tα − Tβ)(B) ⊆ U for
each α, β ≥ α0. Next, take arbitrary vector x in E. Then since B is absorbing, there is
some positive scalar λ such that λx ∈ B. So, we get (Tα(x) − Tβ)(x)) ⊂ 1
λ U for every
α, β ≥ α0. So, (Tα(x)) is a Cauchy net in F . Then, by choosing an operator T from E to
F as T (x) is the limit of (Tα(x)) and by applying Proposition 3.5, T is st-bs-bounded.
6
Theorem 3.8. Let E be a topological vector space, F be an ordered topological complete
vector space. If every st-bounded set in E is absorbing then Bst−bo(E, F ) is complete with
the topology of uniform convergence on st-bounded sets.
Proof. Suppose (Tα) is a Cauchy sequence in Bst−bo(E, F ) with respect to uniform con-
vergence on st-bounded sets. Fix an st-bounded set B in E. Thus, for every positive
vector u ∈ F+, there is an index α0 such that (Tα − Tβ)(B) ⊆ [−u, u] for each α, β ≥ α0.
Now, let's take an element x ∈ E. Then we have some positive scalar λ such that
λx ∈ B because B is absorbing. So, we get (Tα(x) − Tβ)(x)) ⊂ [ −u
λ ] for every α, β ≥ α0.
Thus, we have that (Tα(x)) is a Cauchy net in F . So, one can choose an operator T from
E to F as T (x) is the limit of (Tα(x)). Then, by applying Proposition 3.6, we get
T ∈ Bst−bo(E, F ).
λ , u
References
[1] C. D. Aliprantis and O. Burkinshaw, Locally solid Riesz spaces with applications to
economics, Mathematical Surveys and Monographs, American Mathematical Society,
105, 2003.
[2] C. D. Aliprantis and O. Burkinshaw, Positive operators, Academic Press, London,
2006.
[3] A. Aydn, Multiplicative order convergence in f-algebras, to appear Hacettepe Math.
Stat., doi.org/10.15672/hujms.512073.
[4] A. Aydın, The statistically unbounded τ -convergence on locally solid Riesz spaces,
submitted.
[5] A. Aydın, Topological algebras of bounded operators with locally solid Riesz spaces,
Erzincan Uni. Sci. Tech., 11(3), 543-549, 2018.
[6] H. Albayrak and S. Pehlivan, Statistical convergence and statistical continuity on
locally solid Riesz, Topology App., Vol.159, 1887-1893, 2012.
[7] L. Hong, On Order Bounded Subsets of Locally Solid Riesz Spaces, Quaest. Math.,
39(3), 2016.
[8] G. D. Maio and L. D. R. Kocinac, Statistical convergence in topology, Topology
Appl. 156, 2845, 2008.
[9] F. Riesz, Sur la dcomposition des oprations fonctionelles linaires, in: Atti del Congr.
Internaz. dei Mat., 3, Bologna, 143148, 1928.
[10] V. G. Troitsky, Spectral radii of bounded operators on topological vector spaces, P.
American Math. 11(3), 1-35, 2001.
[11] A. C. Zaanen, Riesz spaces II. Amsterdam, The Netherlands: North-Holland Pub-
lishing, 1983.
[12] S. Hejazian, M. Mirzavaziri and O. Zabeti, Bounded operators on topological vector
spaces and their spectral radii, Filomat, 26(6), 1283-1290, 2012.
7
|
1906.06494 | 1 | 1906 | 2019-06-15T08:18:32 | Finite Reflection Groups: Invariant functions and functions of the Invariants in finite class of differentiability | [
"math.FA"
] | Let $W$ be a finite reflection group. A $W$-invariant function of class~$C^{\infty}$ may be expressed as a functions of class $C^{\infty}$ of the basic invariants. In finite class of differentiability, the situation is not this simple. Let~$h$ be the greatest Coxeter number of the irreducible components of $W$ and $P$ be~the Chevalley mapping, if $f$ is an invariant function of class $C^{hr}$, and $F$ is the function of invariants associated by $f=F\circ P$, then $F$ is of class $C^r$. However if~$F$ is of class $C^r$, in general $f=F\circ P$ is of class $C^r$ and not of class $C^{hr}$. Here we determine the space of $W$-invariant functions that may be written as functions of class $C^r$ of the polynomial invariants and the subspace of functions $F$ of class $C^r$ of the invariants such that the invariant function $f=F\circ P$ is of class $C^{hr}$. | math.FA | math |
FINITE REFLECTION GROUPS:
INVARIANT FUNCTIONS AND
FUNCTIONS OF THE INVARIANTS
IN FINITE CLASS OF DIFFERENTIABILITY
G´ERARD P. BARBANC¸ ON
Abstract. Let W be a finite reflection group. A W -invariant function of
class C∞ may be expressed as a functions of class C∞ of the basic invariants.
In finite class of differentiability, the situation is not this simple. Let h be the
greatest Coxeter number of the irreducible components of W and P be the
Chevalley mapping, if f is an invariant function of class C hr , and F is the
function of invariants associated by f = F ◦ P , then F is of class C r . However
if F is of class C r , in general f = F ◦ P is of class C r and not of class C hr .
Here we determine the space of W -invariant functions that may be written as
functions of class C r of the polynomial invariants and the subspace of functions
F of class C r of the invariants such that the invariant function f = F ◦ P is
of class C hr .
1. Introduction
Let W be a finite reflection group acting orthogonally on Rn. The algebra
of invariant polynomials is generated by n algebraically independent W -invariant
homogeneous polynomials and the degrees of these polynomials are uniquely deter-
mined [3], [4]. Let p1, . . . , pn be a set of basic invariants, we define the "Chevalley"
mapping
P : Rn −→ Rn, P (x) = (cid:0)p1(x), ..., pn(x)(cid:1).
Glaeser's theorem [7] shows that W -invariant functions of class C∞ may be ex-
pressed as functions of class C∞ of the basic invariants, and that the subalgebra
P ∗(C∞(Rn)) of composite mappings of the form F ◦ P with F of class C∞, is closed
in C∞(Rn).
In finite class of differentiability, the situation is not this simple. Let h be the
highest degree of the coordinate polynomials in P , equal to the greatest Coxeter
number of the irreducible components of W . We have:
Proposition 1.1 (see [1]). There exists a linear and continuous mapping
Chr(Rn)W ∋ f −→ F ∈ Cr(Rn)
such that f = F ◦ P .
A general counter-example shows that this result is the best possible.
However if F is of class Cr, in general f = F ◦ P is of class Cr and not of
class Chr. So the following questions naturally arise:
1991 Mathematics Subject Classification. Primary:20F55, 58B10, Secondary:37B,57R.
Key words and phrases. Whitney functions, Finite reflection groups, Finite class of
differentiability.
1
2
G´ERARD P. BARBANC¸ ON
1) What is the space of W -invariant functions that may be written as functions
of class Cr of the polynomial invariants, and
2) What is the subspace of functions F of class Cr of the invariants such that
the invariant function f = F ◦ P is of class Chr? To answer this second question
we will complete the results given in [1] and determine the subspace of E r(P (Rn))
isomorphic to Chr(Rn)W .
2. The Chevalley mapping
When the reflection group W is reducible, it is the direct product of its irreducible
components, say
W = W 0 × W 1 × · · · × W s
and we may write Rn = Rn0 ⊕ Rn1 ⊕ · · · ⊕ Rns as an orthogonal direct sum.
The first component W 0 is the identity on the W -invariant subspace Rn0, and W i
is an irreducible finite Coxeter group acting on Rni for i = 1, . . . , s. We choose
coordinates that fit with this orthogonal direct sum: if w = w0w1 · · · ws ∈ W with
wi ∈ W i, 0 ≤ i ≤ s, we have for all x ∈ Rn
w(x) = w(x0, x1, . . . , xs) = (cid:0)x0, w1(x1), . . . , ws(xs)(cid:1).
The direct product of the identity P 0 acting on Rn0 and Chevalley mappings P i
associated with W i acting on Rni , 1 ≤ i ≤ s, is a Chevalley map associated with W
acting on Rn.
For an irreducible W (or for an irreducible component) we may assume that the
degrees of the polynomials p1, . . . , pn are in increasing order
k1 = 2 ≤ · · · ≤ kn = h,
where h is the Coxeter number of W . In the reducible case, if there is an invariant
subspace Rn0 , the corresponding pis say P 0
j , 1 ≤ j ≤ n0 are of degree k0 = 1.
Then for each i = 1, . . . , s the P i
1 are
ni is the Coxeter number of W i. We denote with h the maximal
equal to 2, and ki
Coxeter number (or highest degree of the coordinate polynomials) of the irreducible
components.
j, 1 ≤ j ≤ ni. The ki
j are of degree ki
Let R be the set of reflections different from identity in W . The number of these
reflections is
R# = d =
n
X
i=1
(ki − 1).
For each τ ∈ R, we consider a linear form λτ , the kernel of which is the reflecting
hyperplane
Hτ = (cid:8)x ∈ Rn ; τ (x) = x(cid:9).
The jacobian matrix JP of P , is block diagonal. Let In0 , JP 1 , . . . , JP ns be the
diagonal blocks. The n0×n0 minor in the upper left corner is 1 but the determinants
JP 1, . . . , JP ns , all vanish on Rn0 . The Jacobian determinant of P is
JP = c Y
τ ∈R
λτ
for some constant c 6= 0. The critical set is the union of the Hτ when τ runs
through R. The invariant subspace is the intersection Tτ ∈R Hτ of the reflection
hyperplanes.
FINITE REFLECTION GROUPS: INVARIANT FUNCTIONS
3
A Weyl chamber C is a connected component of the regular set. The other
connected components are obtained by the action of W and the regular set is
Sw∈W w(C). There is a stratification of Rn by the regular set, the reflecting hyper-
planes Hτ and their intersections. The mapping P is neither injective nor surjective
but it induces an analytic diffeomorphism of C onto the interior of P (Rn) and an
homeomorphism that carries the stratification from the fundamental domain C
onto P (Rn). Finally let us recall that the Chevalley mapping is proper and sepa-
rates the orbits.
3. Whitney functions and r-regular jets of order m
A general study of Whitney functions may be found in [9], the notations of which
will be used freely.
A jet of order m ∈ N, on a locally closed set E ⊂ Rn is a collection
A = (ak)k∈Nn,k≤m
of real valued functions ak continuous on E. At each point x ∈ E the jet A deter-
mines a polynomial of degree m, Ax(X) and we will sometimes speak of continuous
polynomial fields instead of jets. As a function, Ax(X) acts on vectors x′−x tangent
to Rn and we write
Ax : x′ 7−→ Ax(x′) = X
k∈Nn
k≤m
1
k!
ak(x)(x′ − x)k.
The space J m(E) of jets of order m on E is naturally provided with the Fr´echet
topology induced by the family of semi-norms
Am
Ks = sup
x∈Ks
k≤m
(cid:12)(cid:12)ak(x)(cid:12)(cid:12),
where Ks runs through an exhaustive countable collection of compact sets of E.
By formal derivation of A of order q ∈ Nn, q ≤ m, we get jets of the form
(aq+k)k≤m−q, inducing polynomials
(DqA)x(x′) = aq(x) + X
k>q
1
(k − q)!
k≤m
ak(x)(x′ − x)k−q.
For 0 ≤ q ≤ r ≤ m, we put
(RxA)q(x′) = (DqA)x′ (x′) − (DqA)x(x′).
Definition 3.1. Let A be a jet of order m on E. For r ≤ m, we say that A is
r-regular on E, if for all compact set K ⊂ E, for (x, x′) ∈ K 2, and for all q ∈ Nn
with q ≤ r, it satisfies the Whitney conditions
(W r
q )
(RxA)q(x′) = o(cid:0)x′ − xr−q(cid:1),
Remark 3.2. If m > r there is no need to consider the truncated field Ar =
(ak)k∈Nn,k≤r instead of A in the conditions (W r
q ). Actually (RxAr)q(x′) and
(RxA)q(x′) differ by a sum of terms [ ak(x)
(k−q)! ](x′−x)k−q, with ak uniformly continuous
on K and k − q > r − q.
4
G´ERARD P. BARBANC¸ ON
The space of r-regular jets of order m on E is naturally provided with the Fr´echet
topology defined by the family of semi-norms
kAkr,m
Ks
= Am
Ks + sup
(x,x′)∈K 2
s
x6=x′, k≤r
(cid:16) (RxA)k(x′)
x − x′r−k (cid:17)
when Ks, s ∈ N runs through an exhaustive family of compact sets in E. This
Fr´echet space will be denoted by E r,m(E).
If r = m, E r(E) is the space of Whitney fields of order r or Whitney functions
of class Cr on E. When E is locally closed we have:
Theorem 3.3 (see [11]). The restriction mapping of the space E r(Rn) to the space
E r(E) is surjective. There is a linear section, continuous when the spaces are
provided with their natural Fr´echet topologies.
When A ∈ E r(E) and q ≤ r, the formal derivation DqA and the regular deriv-
ative ∂qA/∂xq define the same polynomial
(DqA)x(x′) = (cid:16) ∂qA
∂xq (cid:17)x
(x′)
and they may be identified.
In general the semi-norms k.kr
Ks are not equivalent on E r(E). However
if E is an open set O, the semi-norms are equivalent. In this case E r(O) is the
space of fields of Taylor polynomials of functions in Cr(O) and the two spaces
E r(O) and Cr(O) may be identified.
Ks and .r
Definition 3.4 (see [10]). We say that a compact set K ⊂ Rn connected by
rectifiable arcs, is Whitney 1-regular if the geodesic distance in K is equivalent to the
Euclidian distance. That is, there is a constant kK > 0 such that for all (x, x′) ∈ K 2,
there is a rectifiable arc from x to x′ in K with length ℓ(x, x′) ≤ kK x − x′.
Proposition 3.5 (see [10]). If the compact set K is 1-regular, the norms k.kr
K
and .r
K are equivalent on E r(K).
Proposition 3.6 (see [2]). The image by the Chevalley mapping P of any closed
ball centered at the origin is Whitney 1-regular.
This entails that any compact in P (Rn) is Whitney 1-regular. Hence on E r(P (Rn))
the families of semi-norms k.kr
K and .r
K are equivalent.
4. Functions of class Cr of the invariants and invariant functions
We consider the homomorphism induced by P :
P ∗ : E r(cid:0)P (Rn)(cid:1) −→ E r(Rn),
P ∗(F ) = F ◦ P.
For any (a, x) ∈ Rn × Rn, by the Taylor formula for F between P (a) and P (x),
either by using an extension of F to E r(Rn) given by Whitney extension theorem
or by using a Taylor integral remainder along an integrable path satisfying the
inequality given by the 1-regularity property of P (Rn), we have:
F(cid:2)P (x)(cid:3) = F(cid:2)P (a)(cid:3) + X
1≤β≤r
1
β!
DβF(cid:2)P (a)(cid:3)(cid:0)P (x) − P (a)(cid:1)β
+ o(cid:0)P (x) − P (a)r(cid:1).
FINITE REFLECTION GROUPS: INVARIANT FUNCTIONS
5
Expanding P (x) − P (a) by the polynomial Taylor formula, we get a polynomial
in x − a of degree h. Hence for f = F ◦ P ,
f (x) = f (a) + X
1≤α≤hr
1
α!
fα(a)(x − a)α + o(cid:0)P (x) − P (a)r(cid:1).
On a compact K ⊂ Rn containing [a, x], there exists a constant CK such that
P (x) − P (a)r ≤ CKx − ar.
When α ≤ r the fα are continuous derivatives. When r < α ≤ hr the fα are
obtained by the composition process with β ≤ r and they are continuous, so f
belongs to E r,hr(Rn).
Lemma 4.1. Let F be of class Cr and f = F ◦ P . For 0 ≤ β ≤ r, Fβ ◦ P (x) is a
linear combination of some fα(x), 0 ≤ α ≤ hr.
Proof. We have f0(x) = F0(P (x)). By induction, let us first assume r = 1. We
identify
fx(x′) = f0(x) + X
1≤α≤h
1
α!
fα(x)(x′
1 − x1)α1 · · · (x′
ℓ − xℓ)αℓ
with
fx(x′) = F0(cid:0)P (x)(cid:1) +
ℓ
X
1
Fi ◦ P (x)(cid:16)
ki
X
α=1
1
α!
∂αpi
∂xα (x)(x′ − x)α(cid:17),
where Fi stands for Fβ, βi = 1, βj = 0 if i 6= j.
For α = ki, we get
fα = Fi ◦ P
∂kipi
∂xα + X
s>i
Fs ◦ P
∂kips
∂xα ·
In particular when α = h, we have fα = Fn ◦ P (∂hpn/∂xα) and since pn is of
degree h there is a ∂hpn/∂xα which is not 0.
Hence the result for Fn. Solving the equations in succession gives the result for
the Fi ◦ P , i = 1 . . . n.
For more explicit computations, observe that if W is reducible, it would be
sufficient to study each irreducible component Wi in each Rni and gather the results
at the end. For an irreducible component we may use the polynomial invariants
given in [8]. Disregarding Dn for a while, for the other groups the ki are distinct
and there is an invariant set of linear forms {L1, . . . , Lv} such that their symmetric
functions Pv
j=1[Lj(x)]ki with
the kis determined in [4]. At least one of the Lj contains a monomial in X1,
bringing in pi(X) a monomial in X ki
that cannot be canceled since the ki are all
1
even, with two exceptions: An and I2p.
j are W -invariant and we may take pi = Pv
j=1 Lk
⊲ For I2p we may choose
p1(X) = X 2
1 + X 2
2
and p2(X) =
p
X
i=1
(X1 cos 2jθ + X2 sin 2jθ)p
⊲ For An we may take Li = Xi, i = 1, . . . , n + 1 and there is no possible cance-
1 is PP
j=1(cos 2jθ)p 6= 0.
in which the coefficient of X p
lation either.
6
G´ERARD P. BARBANC¸ ON
Finally for Dn if we choose as basic invariant polynomials
pj =
n
X
i=1
x2j
i
(j = 1, . . . , n − 1) and q(x) = x1x2 · · · xn,
we may use the above method when 1 ≤ j ≤ n−1, and consider ∂nq/∂x1 · · · ∂xn = 1
to get the continuity of (∂F/∂q) ◦ P .
Anyway, one has ∂kn pn/∂xkn
1 = kn!cn for some constant cn 6= 0, while for
1 = 0, since the greatest exponent of x1 in pj(x) is kj < kn. The
j < n, ∂kn pj/∂xkn
identification shows that
cn Fn ◦ P (x) =
1
kn!
fkn,0,...,0(x) with cn 6= 0.
Assuming that when s > i, the Fs ◦ P are linear combinations of the fα, α ≤ h,
since pi(x) contains a monomial in xki
∂kipi
∂xki
1
1 , we have
= ki! ci
for some constant ci 6= 0, and ∂ki pj/∂xki
gives
1 = 0 for j < i. The identification now
fki,0,...,0 = ci ki! Fi ◦ P + X
s>i
Fs ◦ P
∂kips
∂xki
1
·
With the induction assumption we have the result for Fi ◦ P and by decreasing
induction for all the Fj ◦ P, j = 1, . . . , n.
Let us assume that the statement of the lemma is true when r ≤ k. When
β = k, Fβ ◦ P is a linear combination of some fαs with α ≤ hk. By using the
basis step of the induction for the function g = G ◦ P with G = Fβ, we get that
the Gi ◦ P are linear combinations of gα, α ≤ h. The induction assumption for g
shows that for γ = k + 1, Fγ ◦ P is a linear combination of fα, α ≤ h(k + 1).
This achieves the induction and the proof of the lemma.
(cid:3)
From the lemma we get that for any compact K ⊂ Rn there is a constant CK
depending only on K and W such that
F r
P (K) ≤ CKf hr
K .
K f hr
Any compact in P (Rn) being Whitney 1-regular, there is a constant C1
kF kr
P (K) ≤ C1
The linear injective mapping P ∗ is surjective on its image J . This image J is the
space of invariant functions associated with fonctions of class Cr of the invariants.
By the above inequality, the inverse
K , and of course kF kr
P (K) ≤ C1
K such that
Kkf kr,hr
K .
(P ∗)(−1) : J −→ E r(cid:0)P (Rn)(cid:1)
is continuous. Hence by the Banach theorem it is an isomorphism of Fr´echet spaces.
The image J ⊂ E r,hr(Rn) of P ∗ is closed in both E r(Rn) and E r,hr(Rn).
The map
is also an isomorphism of Fr´echet spaces and for each compact K ⊂ Rn, there is a
constant CK such that kf kr,hr
K ≤ kF kr
P (K).
P ∗ : E r(cid:0)P (Rn)(cid:1) −→ J
FINITE REFLECTION GROUPS: INVARIANT FUNCTIONS
7
Conclusion: the space of invariant functions given by a function of class Cr
isomorphic to the Fr´echet
of the invariants is a closed subspace of E r,hr(Rn),
space E r(P (Rn) .
5. Invariant functions of class Chr and functions of the invariants
There exists a linear and continuous mapping (see [1]):
L : Chr(Rn)W −→ E r(cid:0)P (Rn)(cid:1) L(f ) = F, f = F ◦ P.
Thanks to the Whitney extension theorem, the target might be Cr(Rn) as well.
The map L is a version of P ∗(−1), we gave it another name because the source and
target are different from those of P ∗(−1) in the previous section.
Clearly L is injective and surjective on its image.
Let C be a Weyl chamber. Its image P (C) is the interior of P (Rn) where F is
of class Chr. The loss of differentiability happens on the border of P (Rn).
Let S be a stratum of C, intersection of p reflecting hyperplanes. On P (S) the
loss of differentiability is governed by WS, the isotropy subgroup of S, and F is of
class Chr/hS , where hS is the highest degree of the WS-invariants. Moreover, we
have:
Proposition 5.1. If f = F ◦ P is of class Chr, the derivatives of F of order
β = (β1, . . . , βn) with Pn
i=1 βiki ≤ hr, are continuous on P (Rn).
Proof. Let us begin with two remarks. First it is sufficient to study the problem
for each W i acting on Rni, so we may and will assume W to be irreducible. Then
since P is proper if ∂βF/∂xβ ◦ P is continuous, so is ∂βF/∂xβ.
The first derivatives of F with respect to pj are solutions of the system:
(cid:16) ∂f
∂x(cid:17) = (cid:16)(cid:16) ∂pi
∂xj
(cid:17)1≤i,j≤n(cid:17)(cid:16) ∂F
∂p
◦ P(cid:17).
The solutions are given by:
c(cid:16) Y
τ ∈R
λτ(cid:17) ∂F
∂pj
◦ P =
n
X
i=1
(−1)i+jMi,j
∂f
∂xi
,
j = 1, . . . , n.
where Mi,j is the Jacobian of the polynomial mapping:
(z1, ..., zi−1, zi+1, ..., zn; zi) 7→ (p1(z), ..., pj−1(z), pj+1(z), ..., pn(z); zi)
This mapping is invariant by the subgroup Wi of W , that leaves invariant the
ith coordinate axis. Wi is a reflection group, product of the identity on Rei and
the reflection group generated by the set Ri ⊂ R of those of the hyperplanes
generating W containing Rei.
Actually Mi,j is an homogeneous polynomial of degree sj = Pn
At the origin there is a loss of differentiability of 1 + d − sj = kj units. In the
which is sj − 1 flat at the origin and divisible by Qτ ∈Ri
neighborhood of a stratum S intersection of p reflection hyperplanes, we write
i=1,i6=j(ki − 1)
λτ .
with Q invertible and V a Chevalley mapping for the isotropy subgroup WS of S.
We have
n
P = Q ◦ V
Mi,j =
X
l=1
Vl,jQi,l.
8
G´ERARD P. BARBANC¸ ON
The Vl,j and accordingly the Mi,j are (sj
p is the degree of
the Vl,j. Since p is the number of reflections in WS, there is a loss of differentiability
j of the jth polynomial of V which is less than
of 1 + p − sj
or equal to kj . So overall the derivation of F with respect to pj brings a loss of
differentiability of kj units.
p − 1)-flat on S, where sj
p equal to the degree k′
As a consequence a derivation of order β brings a loss of differentiability of
i=1 βiki units and if this sum is less than hr, the derivative will be continuous
(cid:3)
Pn
on P (Rn).
In the neighborhood of S, the loss of differentiability entailed by a β-derivation
i are the degrees of the invariant polynomials of WS.
i=1 βiki, more derivatives will be continuous in the
is Pn
Since this sum is less than Pn
i where the k′
i=1 βik′
neighborhood of S.
If Pn
i=1 βiki = hr, the derivative ∂βF/∂xβ is continuous at the origin.
Its
derivative with respect to pj is not continuous but its product by xkj tends to 0
with x.
When computing ∂α(F ◦ P )/∂xα, with α ≤ hr by the Faa di Bruno formula,
the derivative ∂βF/∂pβ ◦ P is multiplied by (we use the notations of [7]) :
εβ = α!X
Eβ
n
Y
j=1
Y
Aα
1
µiqi (cid:16) 1
qi!
∂qipj
∂xqi (cid:17)µiqi
with
and
Eβ = n(µ1q1 , . . . , µnqn ) ; µiqi ∈ N, 1 ≤ qi ≤ s,
s
X
qi=1
µiqi = βi, 1 ≤ i ≤ no
Aα = nQ = (qi
j) ; 1 ≤ i, j ≤ n, 1 ≤ qi ≤ s, 1 ≤ i ≤ n,
n
s
X
1
X
qi
µiqi qi = αo.
Remark 5.2. The pj are homogeneous polynomials of degree kj. Up to a multi-
plicative constant, the terms Qn
(∂qipj/∂xqi )µiqi are equal to xP kj βj −α.
This exponent is strictly positive when P βjkj > hr and we have
−→ 0, when x → 0.
j=1 QAα
∂βF
xP kj βj −α(cid:12)(cid:12)(cid:12)
∂pβ ◦ P (x)(cid:12)(cid:12)(cid:12)
In the neighborhood of a stratum S, we could show that εβ(x) is equivalent to a
power of the distance from x to S.
So, when f ∈ Chr(Rn)W , the function L(f ) = F is in a subspace of E r(P (Rn)),
the functions of which are of class Crh on the interior of P (Rn), in E rh/hS (P (S)) on
the strata P (S) of the border of P (Rn). Additionally the behavior of the derivatives
of F that are lost on the border is as follows.
When Pn
i=1 βiki > hr, if the derivative (∂βF/∂pβ) is lost on strata S = P (A)
of the border of P (Rn), the function equal to εβ · (∂βF/∂pβ) ◦ P on the interior C
of P (Rn) and equal to 0 on A is continuous.
To be more consistent, we put εβ = 1 when Pn
i=1 βiki ≤ hr and for any β we
will consider the functions εβ · ∂βF/∂pβ ◦ P that vanish on the strata A when
∂βF/∂pβ does not exist on P (A).
FINITE REFLECTION GROUPS: INVARIANT FUNCTIONS
9
If we define the semi-norms
F r,hr
P (Ks) =
max
εβ(x)
β≤hr, x∈Ks
P (Ks) = kF kr
the Faa di Bruno formula shows that f hr
regularity kf khr
Ks and F r,hr
kF khr
P (Ks) are also equivalent.
P (Ks). By Whitney
(cid:12)(cid:12)(cid:12)
∂βF
∂pβ ◦ P (x)(cid:12)(cid:12)(cid:12)
,
P (Ks) + F r,hr
Ks is equivalent to F r,hr
P (Ks),
From proposition 1.1, there is a constant BKs such that kF kr
So, we have kF khr
Ks for some constant B1
P (Ks) ≤ B1
Kskf khr
P (Ks) ≤ BKskf khr
Ks.
Ks. Hence the continu-
ity of L when the target has the topology induced by the semi-norms k.khr
P (Ks).
Since L is a bijection on its image, by the Banach theorem, we have:
Proposition 5.3. The map L is a Fr´echet isomorphism of Chr(Rn)W onto its im-
age in E r(P (Rn)) provided with the topology induced by the semi-norms kF khr
P (Ks).
References
[1] G. Barban¸con. -- Chevalley theorem in class Cr, Proc. Edinb. Math. Soc. 139
A (2009), pp. 743 -- 758.
[2] G. Barban¸con. -- Whitney regularity of the image of the Chevalley mapping,
Proc. Edinb. Math. Soc. 146 A (2016), pp. 895 -- 904
[3] C. Chevalley. -- Invariants of a finite group generated by reflections. Amer. J.
Math. 77 (1955), pp. 778 -- 782.
[4] H.S.M. Coxeter. -- The product of the generators of a finite group generated
by reflections. Duke Math. J. 18 (1951), pp. 765 -- 782.
[5] Hernandez Encinas & Munoz Masque. -- A short proof of the generalized Faa
di Bruno's formula. Applied Math. Letters 16 (2003), pp. 975 -- 979.
[6] G. Glaeser. -- ´Etude de quelques alg`ebres tayloriennes. J. Anal. Math. 6
(1958), pp. 1 -- 124.
[7] G. Glaeser. -- Fonctions compos´ees diff´erentiables. Ann. of Math. 77 (1963),
pp. 193 -- 209.
[8] M.L. Mehta. -- Basic set of invariant polynomials for finite reflection groups.
Commun. Alg. 16 (1988), pp. 1083 -- 1098.
[9] J.C. Tougeron. -- Id´eaux de fonctions diff´erentiables. Ergeb. Math. Grenzgeb.
71, Springer (1972).
[10] H. Whitney. -- Functions differentiable on the boundary of regions. Ann. of
Math. 35 (1934), pp. 482 -- 485.
[11] H. Whitney. -- Analytic extension of differentiable functions defined in closed
sets. Ann. of Math. 36 (1934), pp. 63 -- 89.
University of Texas at Austin
E-mail address: [email protected]
|
1104.4848 | 2 | 1104 | 2011-04-27T12:44:09 | Multiple solutions for equations involving bilinear, coercive and compact forms with applications to differential equations | [
"math.FA",
"math.AP"
] | The existence of multiple fixed points for the coercive, bilinear, compact forms defined in the cone in the Banach space. Multiple applications to the integral equations derived from BVPs for differential equations are provided. | math.FA | math |
Multiple solutions for equations involving
bilinear, coercive and compact forms with
applications to differential equations
Robert Sta´nczy
Instytut Matematyczny, Uniwersytet Wroc lawski,
pl. Grunwaldzki 2/4, 50 -- 384 Wroc law, Poland.
[email protected]
September 25, 2018
Abstract
The existence of multiple fixed points for the coercive, bilinear, com-
pact forms defined in the cone in the Banach space. Multiple applications
to the integral equations derived from BVPs for differential euations are
provided.
Key words and phrases: nonlinear nonlocal elliptic equation, multiple fixed
points, bilinear form.
2000 Mathematics Subject Classification: 35Q, 35J65, 82B05
1 Intoduction and motivation
It is well known that the quadratic equation
u = au2 + u0
(1)
can have either none, one or two solutions u ∈ IR, depending on the data a > 0
and u0 ∈ IR. For example, if we assume that
4au0 < 1
1
(2)
than the existence of two positive solutions of (1) is guaranteed.
In this paper we would like to show that this simple observation can be
generalized if we replace quadratic term ax2 with a bilinear form under suitable
conditions. More specifically, we shall consider the equation in the cone P in
the Banach space U with the norm · in the form
u = b(u, u) + u0
(3)
for some given element u0 ∈ P and bilinear, coercive and compact form b defined
on the product space P × P .
The assumption (2) guaranteeing the existence of two solutions for the
quadratic equation (1) has to be adequately rephrased for (3) as
4bu0 < 1
(4)
where b denotes the norm of the bilinear norm. However, the proof of this
result we shall postpone to the next section.
Let us, however notice that the Banach Fixed Point Theorem for local con-
traction was used extensively to prove the existence of at least one fixed point
(e.g. for the Navier-Stokes equations in [2] or for a Boltzmann equation) for the
bilinear model equation like (3). This fixed point was located in the neighbour-
hood of 0 thus making contraction approach feasible.
To prove the existence of two solutions we shall use another Fixed Point
Theorem due to Krasnosielski (cf. [3]) which allows to obtain more solutions if
the nonlinear opearator has the required property of "crossing" identity twice,
i.e. by the cone compression and the expansion on some appropriate convex
2
subsets of the cone.
A quadratic nonlinearity being the simplest possible in the nonlinear world
poses many questions about the global existence and uniqueness. There are
numerous models which posses such a structure, like the Navier Stokes euqation,
the Boltzmann equation, the quadratic reaction diffusion equation (cf. [5]), the
Smoluchowski coagulation equation or the system modelling chemotaxis and to
name but a few. The problem of uniqueness of solutions for these equations
attracted a lot of attention and only some partial results are known. In some
cases nonuniqueness occurs and the existence of two solutions can be proved.
Sometimes one of the solution is a trivial one and then the proof relies on
finding a nontrivial one. In these dificult but important models one encounters
another problem making our approach not feasible i.e. very common lack of
compactness, thus if we would like to make our apprach feasible we are forced
to consider some truncated baby model compatible with compact bilinear forms.
2 The main result
To prove the result announced in the previous section we shall use the following
theorem [3, Theorem 2.3.4] originating from the works of Krasnosielski, cf., e.g.,
[6].
Theorem 2.1 Let E be a Banach space, and let P ⊂ E be a cone in E. Let
Ω1 and Ω2 be two bounded open sets in E such that 0 ∈ Ω1 and Ω1 ⊂ Ω2. Let
operator A satisfy conditions
kAuk ≤ kuk for any u ∈ P ∩ ∂Ω1 and kAuk ≥ kuk for any u ∈ P ∩ ∂Ω2
3
or
kAuk ≥ kuk for any u ∈ P ∩ ∂Ω1 and kAuk ≤ kuk for any u ∈ P ∩ ∂Ω2
is satisfied. Then A has at least one fixed point in P ∩ (Ω2 \ Ω1).
Theorem 2.2 Assume that, for the given cone P ⊂ E, the bilinear and compact
form b : P × P → P satisfies the following coercivity condition
inf
u=1,u∈P
b(u, u) > 0.
(5)
Then for any u0 ∈ P as small as to satisfy (4) the equation (3) admits at least
two solutions in P.
Proof. Let us define the operator
T u = b(u, u) + u0
(6)
then we shall apply Krasnosielski Theorem once as a cone-compression in the
neighbouhodd of zero and secondly as a cone-expansion at infinity.
Notice that we have the following estimates
T u ≤ u0 + Bu2,
T u ≥ u0 − Bu2,
(7)
where constant B = b > 0 denotes the norm of the bilinear form b, i.e., the
least (smallest??) constant B satifying for any u, v ∈ the inequality
for any u, v ∈ E.
b(u, v) ≤ Buv .
4
Hence for sufficiently small ρ1 > 0, i.e., such that Bρ2
1 + ρ1 < u0 and any
u ∈ P and u = ρ1 one has
T u ≥ u0 − Bρ2
1 > ρ1 = u.
Moreover, if we assume that there exists ρ2 such that
u0 + Bρ2
2 < ρ2
then apparently for any u ∈ P and u = ρ2 one has
(8)
(9)
T u ≤ u0 + Bu2 < ρ2 = u.
(10)
But this can be accomplished if we assume Bρ2
1 + ρ1 < u0 < ρ2 − Bρ2
2.
Finally, for sufficiently large values of ρ3 > 0 and any u ∈ P and u = ρ3,
due to the coercivity assumption (5), one has
b(u, u) ≥ Cu2
implying
T u ≥ Cρ2
3 − u0 > ρ3 = u.
(11)
(12)
To be more specific ρ3 has to be so large that u0 < Cρ−
3 ρ3.
Combining (8) with (10) we get that part of the cone P between the values
ρ1 and ρ2 (in the · norm) is compressed while between the values ρ2 and ρ3 is
expanded yielding the desired two fixed points in each set. Note that it might
be necessary to distinguish between ρ2 used in both sets as to prevent both fixed
points to glue together.
The last the only assumption which should be made is to guarantee (10) to
hold which follows readily from (4).
5
3 Examples of applications to differential equa-
tions
Example 1.
Consider the following BVP for ODE for continuous postitive function f
− u′′(t) = (u(t))2 + f (t),
u(0) = u(1) = 0.
This problem can be formulated in a required form
where the function
u = b(u, u) + u0
u0(t) =Z 1
0
G(t, s)f (s) ds
(13)
(14)
(15)
(16)
for the symetric, Green function given by G(t, s) = t(1 − s), 0 ≤ t ≤ s ≤ 1, while
the bilinear form b is defined, for any u, v ∈ P , by
b(u, v) =Z 1
0
G(t, s)u(s)v(s),
and the cone P on which coercivity of B holds can be defined as
P =(cid:26)u ≥ 0 :
inf
t∈[a,b]
u(t) ≥ min{a, 1 − b}u∞(cid:27) .
For more applications to this kind of problems see [8, 7].
Example 2. Consider the following BVP for PDE, where Ω is an annulus in IRn,
∆u(x) = (u(x))2 + f (x) , x ∈ Ω,
u∂Ω = 0.
6
(17)
(18)
Then the problem can be formulated as
where
u = b(u, u) + u0
u0(x) =ZΩ
G(x, y)f (y) ds
(19)
(20)
for the appropriate, symmetric Green function G. The bilinear form B is defined
by B(u, v) =R 1
0 G(x, y)u(y)v(y) and the cone P on which coercivity of B holds,
can be expressed as P {u ≥ 0 : inf x∈D u(x) ≥ γu∞)} , for D ⊂ Ω and some
positive γ. For more applications to this kind of problems see [4].
Example 3. Consider for u = u(x, t) the boundary value problem
ut = ∆u + u2, u(x, 0) = f (x) .
Then one can formulate this problem as
where
u = b(u, u) + u0
u0 = S(t)f,
(21)
(22)
(23)
while the bilinear form b is defined by
b(u, v) =Z t
0
S(t − s)u(s)v(s) ds .
(24)
Note that for any t, s and the heat semigroup S(t) one has
CS(t − s)u(s)22 ≥ S(t − s)u(s)21 ≥ u(s)21 ≥ u(s)2
2 ,
(25)
7
hence after integration on (0, t) and taking sup norm with respect to t ∈ [0, T ],
the right hand side can be estimated by
Z T
0
u(s)2
2 ds ,
(26)
while taking squared integral with respect to t ∈ [0, T ] the right hand side can
be estimated by
Moreover
T Z T
0
u(s)2
2! ds .
S(t)f 2 ≤ Cf 2 ,
(27)
(28)
hence the second condition for sup norm in t follows. However one cannot
guarantee the condition for coercivity. Note that due to [5] there are some
results guaranteeing non uniqueness results. Our approach cannot yield this
second solution, though some nonuniqueness results are known for semilinear
parabolic equations [9].
Example 4. General case.
Since the crucial assumptions in both examples which guarantee that the
cone P is invariant is the following property of the Green function (or in general
the kernel of Hammerstein operator involving the bilinar form b(u, v)) G
inf
x∈U
G(x, y) ≥ γ sup
x∈V
G(x, y)
(29)
where γ > 0 is independent of a set V and its subset U such that U ⊂ V .
Note that this condition holds either if V is interval or annulus but we were
not able to prove it for arbitrary domain, e.g. for a ball in higher dimension.
8
If (29) holds without any problems one can prove that properly defined cone is
invariant under the action of the bilinear form, thus making it coercive. This
problem can be illustrated in the second example, since if we replace an annulus
with a ball the Green function is not bounded from above and also in the third
one for the heat semigroup where the norms cannot be compared. This is the
main obstacle in extending this kind of results to physically interesting models
like the Navier Stokes equation, the Boltzmann equation or the Smoluchowski
coagulation models sharing the property of quadratic nonlinearities.
Acknowledgments. This work has been partially supported by the Polish Ministry of
Science project N N201 418839.
References
[1] Non-unicit´e des solutions d'une ´quation d'´volution non-lin´eaire, Annales
Facult´e des Sciences Toulouse, 5 (1983), 287 -- 302
[2] M. Cannone and Y. Meyer LittlewoodPaley decomposition and the
NavierStokes equations, Methods Appl. Anal. 2 (1995) 307319.
[3] D. Guo and V. Lakshmikantham, Nonlinear Problems in Abstract
Cones, Academic Press, Orlando, FL, 1988.
[4] K. S. Ha and Y. H. Lee, Existence of multiple posiitve solutions of
singular boundary value problems, Nonlinear Analysis 28 (1997), 1429 --
1438.
9
[5] A. Haraux and F. B. Weissler, Non-unicit´e pour un probl´eme de
Cauchy semi-lin´eaire, Nonlinear partial differential equations and their ap-
plications, Coll´ege de France Seminar, Vol. III 428 220-233 428 (Paris,
1980/1981), Res. Notes in Math., 70, Pitman, Boston, Massachussets, Lon-
don, 1982.
[6] Krasnosielski, Topological methods in the theory of nonlinear integral
equations, translated by A. H. Armstrong, translation edited by J. Burlak,
A Pergamon Press Book The Macmillan Co., New York 1964.
[7] Y. H. Lee, An existence result of positive solutions for singular superlinear
boundary value problems and its applications, J. Korean Math. Soc. 34
(1997), 247 -- 255.
[8] B. Przeradzki and R. Stanczy, Positive solutions for sublinear elliptic
equations, Colloq. Math. 92 (2002), 141 -- 151.
[9] , F. B. Weissler Asymptotic analysis of an ordinary differential equa-
tion and nonuniqueness for a semilinear partial differential equation, Arch.
Rational Mech. Anal, 91 (1985), 231-245
10
|
1907.00418 | 1 | 1907 | 2019-06-30T17:02:37 | Compton scattering tomography in translational geometries | [
"math.FA"
] | Here we present new $L^2$ injectivity results for 2-D and 3-D Compton scattering tomography (CST) problems in translational geometries. The results are proven through the explicit inversion of a new toric section and apple Radon transform, which describe novel 2-D and 3-D acquisition geometries in CST. The geometry considered has potential applications in airport baggage screening and threat detection. We also present a generalization of our injectivity results in 3-D to Radon transforms which describe the integrals of the charge density over the surfaces of revolution of a class of $C^1$ curves. | math.FA | math |
COMPTON SCATTERING TOMOGRAPHY IN TRANSLATIONAL
GEOMETRIES
02/07/2019 01:11
JAMES WEBBER AND ERIC L. MILLER
Abstract. Here we present new L2 injectivity results for 2-D and 3-D Compton scat-
tering tomography (CST) problems in translational geometries. The results are proven
through the explicit inversion of a new toric section and apple Radon transform, which
describe novel 2-D and 3-D acquisition geometries in CST. The geometry considered
has potential applications in airport baggage screening and threat detection. We also
present a generalization of our injectivity results in 3-D to Radon transforms which de-
scribe the integrals of the charge density over the surfaces of revolution of a class of C 1
curves.
1. Introduction
Here we present novel inversion formulae for two and three dimensional scanning modal-
ities in CST, which have potential applications in airport baggage screening and threat
detection. We consider the scanning geometry illustrated in figure 1. The electron charge
density f (represented by a real valued function) is translated in the y direction (on a
conveyor belt) and illuminated by a line of monochromatic photon sources. The scat-
tered intensity is then collected by the detector array. The source and detector array
are assumed to be slightly offset (similarly to the machine geometry considered in [6])
in the y direction so that the detectors do not block any incoming photons. Given the
small offset we model the sources and detectors to lie in the same (xz) plane, as in [7],
for elegance of mathematical derivation.
The works of [8, 9, 10, 11, 12, 13, 14, 15] (2-D geometries) and [16, 17, 18, 19, 20] (3-D
geometries) respectively consider CST problems in a variety of two and three dimensional
scanning geometries. In electron density reconstruction in CST, the scattered intensity
is modelled as integrals of the electron density over circular arcs (in 2-D with collimated
detectors [9]), toric sections (in 2-D with uncollimated detectors [7, 15]) and spindle tori
(in 3-D with no collimation [17, 18]). In gamma ray source reconstruction using CST, the
scattered intensity is modelled as integrals of the source intensity over V-lines (or broken
rays, in 2-D [11, 13]) and cones (in 3-D [16, 19]). In contrast to these efforts, here we
consider the problem of reconstructing the electron density from a set of vertical (with
axis of revolution parallel to the z axis) toric sections (2-D) and apples (3-D, the apple
is the part of the spindle torus which corresponds to backscattered photons) which are
translated along the line (2-D) and in the plane (3-D) respectively. The nature of the
data acquisition is such that the previous results from the literature are insufficient so as
to provide an explicit solution. We aim to provide such explicit inversion results here in
this paper.
The inversion process we present relies on Paley-Weiner-Schwartz ideas (using analytic
continuation as in [6]) and the explicit inversion of 1-D Volterra operators in the Fourier
domain. In [16], the source intensity is reconstructed from a set of vertical cones translated
on the plane (the data dimensions are the cone opening angle, and a 2-D translation).
Here, after the data is transformed to the Fourier space, the inversion is carried out by
1
inverse Hankel transform of a set of 1-D integral equations with Bessel kernels. In section
4 we also discover Bessel function kernels in the Fourier domain, but in our case we use
the theory of Volterra integral operators [1] to invert the resulting set of 1-D integral
equations. In 3-D our geometry is akin to configuration (d) of [17, page 5], but we only
consider the vertical spindle tori. In [17] they provide microlocal inversion results (modulo
smoothing) for the proposed geometry. We provide injectivity results and exact explicit
inversion formulae for geometry (d) in this paper. In [15] injectivity results are proven
for a toric section transform with rotational invariance. Here, after an expansion to the
Fourier series, the solution was obtained using the theory of Cormack [4]. We consider
a toric section transform with translational invariance, whereby our solution is obtained
using the theory of [1].
In section 2 we state some definitions and preliminary results that will be needed to
prove our main theorems. In section 3 we explain the parameterization of the toric section
curves in the translation geometry and introduce a new toric section transform T . We
then prove the boundedness of T in L2 (this is a typical property of Radon transforms
as smoothing operators) before going on to prove our first main theorem which proves
the injectivity and explicit invertibility of T on the domain of compactly supported L2
functions.
In section 4 we consider the 3-D case, where we introduce a new apple Radon transform
A. As in the 2-D case in section 3 we prove the continuity of A in L2 before going on
to prove our second main theorem, which shows the injectivity and explicit invertibility
of A on the L2 domain. The proof uses similar ideas to the 2-D case, except in 3-D the
resulting kernels are the finite sum of Bessel functions, whereas in 2-D we see Volterra
operators with cosine kernels.
In section 5 we introduce the generalized Radon transform RA, which describes the
integrals of a density over the surfaces of revolution of C 1 curves, in the translational
geometry. Example C 1 curves would include the previously considered semicircles (for
apples) and straight lines (for cones [16]). We show the injectivity of RA on the domain
of compactly supported continuous functions in this case. Our proof uses the ideas of
Cormack [4], which require continuity in the target function.
2. Preliminary results and definitions
Here we state some definitions and preliminary results which will be used in our theo-
rems.
Definition 2.1. Let f ∈ L2
of angular frequency
(2.1)
0(Rn). Then we define the Fourier transform f of f in terms
f (ω) = (2π)−n/2ZRn
f (x)e−ix·ωdx.
We state the Plancharel theorem [5].
Theorem 2.2 (Plancharel theorem). Let f ∈ L2
(2.2)
kfkL2
0(Rn) = k fkL2(Rn).
0(Rn). Then f ∈ L2(Rn) and
From [3, page 22], we have the Paley-Weiner-Schwartz theorem.
Theorem 2.3 (Paley-Weiner-Schwartz). Let E′(Rn) be the set of distributions of compact
support in Rn and let f ∈ E′(Rn). Let U = supp(f ) and let HU be the support function
2
T (r) (2-D), A(r) (3-D)
z
1
1
1
O
R
r
h
s
d
x
f
{z = 2 − rm}
Figure 1. Vertical torus/toric section scanning geometry. We denote T (r)
to be the union of the two circles drawn above, and the apple A(r) to be
the surface of revolution of T (r) about z. s and d label the source and
detector lines respectively, r is the radius of the torus, R = √r2 − 1 is the
distance from the center of the torus to the centre of the torus tube and
h = 2 − r.
of U , defined as
(2.3)
HU (ω) = sup
x∈U
(x · ω).
Then the Fourier transform f is an entire analytic function. Additionally, there exists a
constant c > 0 and integer N such that
(2.4)
f (ω) ≤ c(1 + ω)N eHU (Im(ω)),
∀ω ∈ Cn,
where Im(ω) denotes the imaginary part of ω. Conversely, let F be an entire analytic
function, let H be the support function of a compact convex set K ⊂ Rn and suppose that
(2.5)
for some constants c and N . Then there exists a unique f ∈ E′(Rn) such that f = F and
supp(f ) ⊂ K.
F (ω) ≤ c(1 + ω)N eH(Im(ω)),
∀ω ∈ Cn
We now state some results on Volterra type integral equations from [1, page 10].
3
Definition 2.4. We define a Volterra equation of the second kind to be an equation of
the form
(2.6)
K(x, y)f (y)dy + f (x)
g(x) = λZ x
0
with real valued kernel K on a triangle T ′ = {0 < x < x′, 0 < y < x}. K is said to be an
L2 kernel if
(2.7)
kKk2
L2(T ′) =Z x′
0 Z x
0
K 2(x, y)dydx ≤ N 2
for some N > 0.
Theorem 2.5. Let f ∈ L2([0, x′]) and let K(x, y) be an L2 kernel on T ′ = {0 < x <
x′, 0 < y < x} for some x′ > 0. Then the Volterra integral equation of the second kind
(2.8)
K(x, y)f (y)dy + f (x)
has one and only one solution in L2([0, x′]), and the solution is given by the formula
0
g(x) = λZ x
f (x) = λZ x
0
H(x, y; λ)g(y)dy + g(y),
H(x, y; λ) =
λνKν+1(x, y)
∞Xν=0
and the iterated kernels Kν are defined by
and
for ν ≥ 1.
K1(x, y) = K(x, y)
Kν+1(x, y) =Z x
0
K(x, z)Kν (z, y)dz
Next we have Young's inequality [22, Theorem 0.3.1]
Theorem 2.6 (Young's inequality). Let X and Y be measurable spaces and let q, p, t ∈
Z+\{0} be such that 1
t − 1. If K : X × Y → R is such that there exists a C > 0
with
q = 1
p + 1
(2.9)
where
(2.10)
(2.11)
and
(2.12)
ZX K(x, y)tdx ≤ C t
ZY K(x, y)tdy ≤ C t.
Then for an integrable function f
(2.13)
ZX(cid:12)(cid:12)(cid:12)(cid:12)ZY
q
K(x, y)f (y)dy(cid:12)(cid:12)(cid:12)(cid:12)
dx ≤ C q(cid:18)ZY f (y)pdy(cid:19) q
p
4
.
3. Two dimensional Compton tomography under translations
Let f ∈ L2
0(Ω) be the electron charge density, where Ω ⊂ {2 − rm < z < 1} for some
rm > 1, is compact. Here rm controls the (fixed) depth of the scanning tunnel as is
depicted in figure 1. Let r = 2 − h and R = √r2 − 1 (as is illustrated in figure 1). Let
S1(r), . . . , S4(r) be the four semicircles, radius r, whose disjoint union ∪4
j=1Sj(r) = T (r)
is the toric section in figure 1, and such that the parameterization of each Sj(r) for
j = 1, 2, 3, 4 is given by the vertical axis coordinate
x1 = R +pr2 − (z − 2)2,
x3 = −R +pr2 − (z − 2)2,
(right hand circle)
(left hand circle)
for 2 − r < z < 1. We define the toric section transform T : L2
translational geometry as
0(Ω) → T (L2
0(Ω)) in the
x2 = R −pr2 − (z − 2)2
x4 = −R −pr2 − (z − 2)2,
4Xj=1ZTx0 (Sj (r))
f ds =
f dsj,
T f (x0, r) =ZTx0 (T (r))
where dsj denotes the arc length measure on Sj and Tx0(x, z) = (x + x0, z) denotes a
translation along the x axis of length x0.
Proposition 3.1. Let f ∈ L2
T f (x0, r) =Z r
1
r
0(Ω), and let f1 be defined as f1(x, z) = f (x, 2 − z). Then
√r2 − z2
2Xj=1
+ f1(−R + (−1)j√r2 − z2 + x0, z)dz.
f1(R + (−1)j√r2 − z2 + x0, z)
Proof. The arc length measure is given by
(3.1)
(3.2)
(3.3)
(3.4)
(3.5)
From which it follows that
T f (x0, r) =
dz.
r
=
f dsj
pr2 − (z − 2)2
dsj = dzs1 +(cid:18)dxj
dz(cid:19)2
4Xj=1ZTx0 (Sj(r))
=Z 1
pr2 − (z − 2)2
=Z r
2Xj=1
√r2 − z2
+ f1(−R + (−1)j√r2 − z2 + x0, z)dz,
r
h
1
r
f (xj + x0, z)dz
4Xj=1
f1(R + (−1)j√r2 − z2 + x0, z)
which completes the proof.
(cid:3)
We have the result which proves the boundedness of T in L2
0.
Proposition 3.2. Let rm > 1 and let Ω ⊂ {2− rm < z < 1} be compact. Then the image
T (L2
0(Ω)) ⊂ L2([1, rm] × R).
5
Proof. Taking the Fourier transform in the x0 variable of (3.5) yields
cT f (ω1, r) =ZR T f (x0, r)e−ix0ω1dx0
rK(r, z)
f1(ω1, z)dz,
=Z r
1
√r2 − z2
where ω1 is dual to x (or x0) and
(3.6)
(3.7)
K(r, z) =
(3.8)
where K ≤ 4 is bounded. We have
2Xj=1
exp(cid:16)−iω1(R + (−1)j√r2 − z2)(cid:17) + exp(cid:16)−iω1(−R + (−1)j√r2 − z2)(cid:17) ,
L2([1,rm]×R) = kcT fk2
kT fk2
1 (cid:18)Z r
=Z ∞
−∞Z rm
√r2 − z2
1 Z r
mZ ∞
−∞Z rm
≤ 16r2
1 (cid:18)Z r
Z ∞
−∞Z rm
f1(ω1, z)dz(cid:19)2
dz!2
f1(ω1, z)
√r2 − z2
f1(ω1, z)L(r, z)dz(cid:19)2
L2([1,rm]×R) by the Plancharel theorem
drdω1
rK(r, z)
(4rm)2
drdω1
drdω1
≤
2
1
1
1
,
where L(r, z) = (r − z)−1/2 is defined on the triangle H = {1 < r < rm, 1 < z < r}. Let
us extend the domain of L to [1, rm]2, in the sense that L(r, z) = 0 on [1, rm]2\H. Let
t = 1. Then we have
(r − z)−1/2dr = 2√rm − z ≤ 2√rm − 1 = C t
(3.9)
and
(3.10)
(3.11)
z
Z[1,rm] L(r, z)tdr =Z rm
Z[1,rm] L(r, z)tdz =Z r
1 (cid:18)Z r
f1(ω1, z)L(r, z)dz(cid:19)2
Z rm
1
1
Hence we have
kT fk2
L2([1,rm]×R) ≤
(3.12)
which completes the proof.
We now have our first main theorem
6
(r − z)−1/2dz = 2√r − 1 ≤ 2√rm − 1 = C t.
dr =Z[1,rm](cid:18)Z[1,rm]
f1(ω1, z)L(r, z)dz(cid:19)2
≤ C 2Z[1,rm] f1(ω1, z)2dz.
1 (cid:18)Z r
Z ∞
−∞Z rm
≤ 8(rmC)2Z ∞
k f1(·, ω1)k2
−∞
= 8(rmC)2kfk2
L2(R2) < ∞,
f1(ω1, z)L(r, z)dz(cid:19)2
L2([1,rm])dω1
(4rm)2
drdω1
2
1
dr
(cid:3)
Setting p = q = 2 and X = Y = [1, rm], it follows from Young's inequality that
(3.13)
(3.14)
K(r, z) =
Theorem 3.3. Let rm > 1 and let Ω ⊂ {2 − rm < z < 1} be compact. Then the toric
section transform T : L2
Proof. From Proposition 3.2 we have the representation of the toric section transform in
the Fourier domain
0(Ω) → L2([1, rm] × R) is injective.
cT f (ω1, r) =Z r
1
rK(r, z)
√r2 − z2
f1(ω1, z)dz,
where we now have the simplified form for K
2Xj=1
exp(cid:16)−iω1(R + (−1)j√r2 − z2)(cid:17) + exp(cid:16)−iω1(−R + (−1)j√r2 − z2)(cid:17)
2Xj=1
= (exp(−iω1R) + exp(iω1R))
= 4 cos(ω1R) cos(ω1√r2 − z2).
exp(cid:16)−iω1(−1)j√r2 − z2(cid:17)
After making the substitution z = z2 and letting f2(ω1, z) =
f1(ω1,√z)
2√z
, we have
(3.15)
For ω1 <
2√r2
π
m−1
(3.16)
dT1f (ω1, r) = cT f (ω1,√r)
4√r
cos(ω1√r − z)
= cos(ω1√r − 1)Z r
m, cos(ω1√r − 1) > 0 and we have
√r − z
1
and r < r2
f2(ω1, z)dz.
g(ω1, r) = dT1f (ω1, r)
cos(ω1√r − 1)
=Z r
√r − z
cos(ω1√r − z)
1
f2(ω1, z)dz,
a Volterra integral equation of the first kind with weakly singular kernel.
Applying the Abel transform to both sides removes the singularity
Z s
1
g(ω1, r)
√s − r
1
cos(ω1√r − z)
√r − z√s − r
cos(ω1√r − z)
√r − z√s − r
K1(s, z) f2(ω1, z)dz,
z
f2(ω1, z)dzdr
dr(cid:21) f2(ω1, z)dz
dr =Z s
1 Z r
1 (cid:20)Z s
=Z s
=Z s
cos(ω1√r − z)
√r − z√s − r
1
K1(s, z) =Z s
z
dr =Z 1
0
7
cos(ω1√u√s − z)
√u√1 − u
du,
(3.17)
where
(3.18)
d
ds
(3.19)
It follows that
(3.20)
g1(ω1, s) =
1
π
= −
= −
1Z 1
K1(s, z) = −ω2
1Z 1
= −ω2
0
0
u
sin(ω1√u√s − z)
ω1√u√s − z
√u√1 − u ·
√u
√1 − u · sinc(ω1√u√s − z)du.
du
1
d
ω2
1
g(ω1, r)
dsZ s
√s − r
dr
√u
π Z s
1 Z 1
√1 − u · sinc(ω1√u√s − z)du f2(ω1, z)dz + f2(ω1, s)
π Z s
K2(s, z) f2(ω1, z)dz + f2(ω1, s),
ω2
1
1
0
after making the substitution r = z +(s−z)u. Hence K1(s, s) = π and the first derivative
with respect to s is
π
π
2√rm−1 and z ∈ [1, r2
with K2 < π
2 , a bounded kernel. The Volterra equation of the second kind above can now
be solved explicitly by Theorem 2.5, and hence f2(ω1, z) is uniquely determined by T f
for ω1 <
m] (and hence for all z ∈ R due to the support restrictions
on f ). Hence we have a unique determination of the 2-D Fourier transform of f2 on the
2√rm−1, ω3 ∈ R}, where ω3 is dual to z. As f2 is compactly
open band B = {ω1 <
supported, its Fourier transform is analytic by the Paley-Weiner-Schwartz theorem, and
hence f2 is uniquely determined everywhere on the plane by analytic continuation from
B. It follows that T is injective.
(cid:3)
Remark 3.4. The inversion process presented above uses analytic continuation in the
Fourier domain to recover the density uniquely. The inversion is not stable however
(due to analytic continuation) and severely ill-posed, in the sense that the solution is
not bounded in any Sobolev space.
In [15] a toric section transform is considered in
a rotational geometry. A toric section transform may be written as the sum of two
circular Radon transforms (as in [15], or equivalently here as the sum of four semicircle
transforms). When the circle transforms are considered separately, the injectivity follows
from the results of [21], and the inversion is stable (as in [15]). When the sum is considered
however there exist image artefacts in the reconstruction (proven microlocally in [15] in
rotational geometries). In our case we require additional intuition to prove injectivity,
since we are considering the sum of two circle transforms with circle centers on a line. We
wonder if similar artefacts to those of [15] may be present also in translational geometries
for the toric section transform.
4. The three dimensional case
Let the apple A(r) = ∪2
j=1Aj(r) (as in figure 1) be written as the disjoint union of the
surfaces of revolution of two semicircles, which are parameterized by
and
(4.1) x1 = (R +pr2 − (z − 2)2) cos ϕ,
(4.2) x2 = (R −pr2 − (z − 2)2) cos ϕ,
y1 = (R +pr2 − (z − 2)2) sin ϕ (for A1(r))
y2 = (R −pr2 − (z − 2)2) sin ϕ (for A2(r))
for 2 − r < z < 1 and ϕ ∈ [0, 2π]. Together, the above parameterizations describe the
set of points on an apple surface (a spindle torus with the lemon part removed). Let
8
f ∈ L2
transform A : L2
0(Ω) → A(L2
0(Ω) for Ω compactly supported in {2 − rm < z < 1} ⊂ R3. We define the apple
0(Ω)) in the translational geometry as
(4.3)
Af (x0, y0, r) =ZTx0 ,y0 (A(r))
f dA =
2Xj=1ZTx0 ,y0 (Aj (r))
f dAj,
where dAj denotes the surface area measure on Aj(r) and Tx0,y0(x, y, z) = (x+x0, y+y0, z)
denotes a translation in the xy plane to the point (x0, y0). We now proceed in a similar
vein to the 2-D case.
Proposition 4.1. Let f ∈ L2
and let f1(x, y, z) = f (x, y, 2 − z). Then
(4.4)
0(Ω) for Ω compactly supported in {2 − rm < z < 1} ⊂ R3,
Af (x0, y0, r) =Z π
−πZ r
1
r
√r2 − z2
2Xj=1
ρjf1(ρj cos ϕ + x0, ρj sin ϕ + y0, z) ρj=R+(−1)j√r2−z2 dzdϕ,
Proof. Let dsj, for j = 1, 2, be the circular arc measures as in equation (3.4). Then the
surface measures dAj for j = 1, 2 are given by [17, page 4]
(4.5)
dAj = dsjρjdzdϕ =
It follows that
r
pr2 − (z − 2)2
(R + (−1)jpr2 − (z − 2)2)dzdϕ.
(4.6)
Af (x0, y0, r) =
=
f dAj
2Xj=1ZTx0 ,y0 (Aj (r))
2Xj=1Z π
−πZ 1
pr2 − (z − 2)2
=Z π
−πZ r
√r2 − z2
2Xj=1
r
r
h
1
(R + (−1)jpr2 − (z − 2)2)f (xj + x0, yj + y0, z)dzdϕ
ρjf1(ρj cos ϕ + x0, ρj sin ϕ + y0, z) ρj=R+(−1)j√r2−z2 dzdϕ
Proposition 4.2. Let rm > 1 and let Ω ⊂ {2 − rm < z < 1} ⊂ R3 be compact. Then the
image A(L2
Proof. Taking the Fourier transform in the x0 and y0 variables of (4.4) yields
0(Ω)) ⊂ L2([1, rm] × R2).
(cid:3)
(4.7)
cAf (ω1, ω2, r) =ZR2 Af (x0, y0, r)e−i(x0,y0)·(ω1,ω2)dx0dy0
rK(r, z)
f1(ω1, ω2, z)dz,
=Z r
1
√r2 − z2
where ω2 is dual to y (or y0) and
(4.8)
K(r, z) =
ρjZ π
−π
2Xj=1
exp(−iρj(ω1 cos ϕ + ω2 sin ϕ)) ρj=R+(−1)j√r2−z2 dϕ
9
and K < M is bounded. We have
(4.9)
kAfk2
L2([1,rm]×R2) by the Plancharel theorem
rK(r, z)
L2([1,rm]×R2) = kcAfk2
1 (cid:18)Z r
=Z ∞
−∞Z ∞
−∞Z rm
√r2 − z2
1 Z r
mZ ∞
−∞Z ∞
−∞Z rm
≤ M 2r2
1 (cid:18)Z r
Z ∞
−∞Z ∞
−∞Z rm
(Mrm)2
≤
2
1
1
1
f1(ω1, ω2, z)dz(cid:19)2
dz!2
f1(ω1, ω2, z)
√r2 − z2
f1(ω1, ω2, z)L(r, z)dz(cid:19)2
drdω1dω2
drdω1dω2
drdω1dω2
where L(r, z) = (r − z)1/2. From here the proof follows the same arguments as in the
proof of Proposition 3.2.
(cid:3)
We now have our second main theorem, which proves the injectivity of the apple
0(Ω).
transform on L2
Theorem 4.3. Let rm > 1, let 0 < δ < rm − 1 and let Ω ⊂ {2 − rm < z < 1 − δ} ⊂ R3
be compact and bounded away from {z = 1}. Then the apple transform A : L2
0(Ω) →
L2([1, rm] × R2) is injective.
Proof. Let
1 + ω2
Then from Proposition 4.2 we have
(ω1, ω2) =qω2
2(cos ϕω, sin ϕω) = ω(cos ϕω, sin ϕω)
where J0 is a Bessel function of the first kind of order zero. Letting f2(ω1, ω2, z) =
f1(ω1,ω2,√z)
2√z
, we have
(4.12)
g(ω1, ω2, r) = cAf (ω1, ω2,√r)
√r
K(√r,√z)
√r − z
=Z r
10
1
f2(ω1, ω2, z)dz.
(4.10)
where
K(r, z) =
(4.11)
=
=
cAf (ω1, ω2, r) =Z r
1
rK(r, z)
√r2 − z2
f1(ω1, ω2, z)dz,
−π
−π
exp(−iρj(ω1 cos ϕ + ω2 sin ϕ)) ρj=R+(−1)j√r2−z2 dϕ
exp(iρjω cos(ϕ − ϕω − π)) ρj=R+(−1)j√r2−z2 dϕ
ρjZ π
2Xj=1
ρjZ π
2Xj=1
ρjZ π
2Xj=1
2Xj=1
(R + (−1)j√r2 − z2)J0(ω(R + (−1)j√r2 − z2)),
exp(iρjω cos ϕ) ρj=R+(−1)j√r2−z2 dϕ,
−π
= 2π
(due to periodicity)
It follows that
Z s
1
g(ω1, ω2, r)
√s − r
K(√r,√z)
√r − z√s − r
K(√r,√z)
√r − z√s − r
f2(ω1, ω2, z)dzdr
dr(cid:21) f2(ω1, ω2, z)dz
K1(s, z) f2(ω1, ω2, z)dz,
1
dr =Z s
1 Z r
1 (cid:20)Z s
=Z s
=Z s
1
z
(4.13)
where
(4.14)
(4.15)
Hence
(4.16)
dr
z
K(√r,√z)
K1(s, z) =Z s
√r − z√s − r
0 P2
= 2πZ 1
=Z 1
√u√1 − u
H(s, z, u)
du,
0
j=1(R + (−1)j√s − z√u)J0(ω(R + (−1)j√s − z√u))
du
√u√1 − u
where R =p(z − 1) + (s − z)u. To calculate the derivatives of H we use the previous
expression for the Bessel function
J0(ν) =
eiν cos ϕdϕ.
1
2πZ π
−π
−π
H(s, z, u) = 2RZ π
+ 2i√u√s − zZ π
=Z π
−π
−π
eiωR cos ϕ cos(ω√u√s − z cos ϕ)dϕ
eiωR cos ϕ sin(ω√u√s − z cos ϕ)dϕ
h(s, z, u, ϕ)(h1(s, z, u, ϕ) + ih2(s, z, u, ϕ))dϕ,
where h(s, z, u, ϕ) = eiωR cos ϕ,
(4.17)
and
(4.18)
h1(s, z, u, ϕ) = 2R cos(ω√u√s − z cos ϕ)
h2(s, z, u, ϕ) = 2√u√s − z sin(ω√u√s − z cos ϕ).
The first partial derivative of h with respect to s is
iωu cos ϕ
h(s, z, u, ϕ) =
(4.19)
d
ds
eiωR cos ϕ,
2R
which is bounded on the support of f , since supp(f ) is bounded away from {z = 1} by
assumption (and hence R is bounded away from 0). We have
(4.20)
d
ds
h1(s, z, u, ϕ) =
=
u
R
u
R
cos(ω√u√s − z cos ϕ) − Rω√u cos ϕ sin(ω√u√s − z cos ϕ)
cos(ω√u√s − z cos ϕ) − Rω2u cos2 ϕ sinc(ω√u√s − z cos ϕ),
√s − z
11
and
(4.21)
d
ds
h2(s, z, u, ϕ) =
sin(ω√u√s − z cos ϕ) +
√u
√s − z
= ωu cos ϕ sinc(ω√u√s − z cos ϕ) + uω cos ϕ cos(ω√u√s − z cos ϕ).
It can now be seen from the above and by an application of the product rule, that
H1(s, z, u) = d
dsH(s, z, u) is bounded on the limits of integration, where f is supported.
That is for (1 + δ)2 < s < r2
u√s − zω cos ϕ cos(ω√u√s − z cos ϕ)
m, (1 + δ)2 < z < s and 0 < u < 1, for any fixed ω.
√s − z
Now (4.13) becomes
(4.22)
g1(ω1, ω2, r) =
d
dsZ s
=Z s
1
d
ds
1
g(ω1, ω2, r)
dr
√s − r
K1(s, z) f2(ω1, ω2, z)dz + 2√s − 1J0(ω√s − 1) f2(ω1, ω2, s).
and s < r2
m, J0(ω√s − 1) > 0
Let t0 denote the first root of J0. Then for ω < t0√r2
and we have
(4.23)
m−1
g1(ω1, ω2, s)
2√s − 1J0(ω√s − 1)
1
d
=
dr
g(ω1, ω2, r)
dsZ s
=Z s
2√s − 1J0(ω√s − 1)
√s − r
d
ds K1(s, z)
1
f2(ω1, ω2, z)dz + f2(ω1, ω2, s),
a Volterra equation of the second kind with d
kernel on {(1 + δ)2 < s < r2
we can reconstruct the 3-D Fourier transform of f2 uniquely on the open set
d
ds H(s,z,u)
√u√1−u du < M bounded
m, (1 + δ)2 < z < s}. As in the 2-D case, and by Theorem 2.5,
0
ds K1(s, z) ≤R 1
, ω3 ∈ R)
t0pr2
m − 1
B =(qω2
1 + ω2
2 <
and hence for all (ω1, ω2, ω3) ∈ R3 by analytic continuation and the Paley-Weiner-
Schwartz theorem. We conclude that A is injective.
Discussion 4.4. Theorems 3.3 and 4.3 explain how we can recover f on an open set in
the Fourier domain, by bounding the limits on cosine and Bessel kernels so as to obtain
an explicit solution via repeated application of Volterra operators. The set S of ω for
which f is known (without the need for analytic continuation) is bounded by the curves
(cid:3)
c1 =(ω1 = −
3 − 1) ,
t0pω2
c2 =(ω1 =
3 − 1) ,
t0pω2
c3 = {ω3 = 1}
in 2-D and by the surface of revolution of c1 (or c2 due to symmetry) about ω3 and
the plane {ω3 = 1} in 3-D (this is evident from the proofs of theorems 3.3 and 4.3).
See figure 2. While the entirety of the Fourier space is uniquely determined from toric
section (2-D) and apple (3-D) integral data, the reconstruction outside of S is not stable
using the proposed inversion formulae, due to analytic continuation. Further, we have
greater recovery of f in the Fourier space (without the need for analytic continuation) as
ω3 → 1 (and at a rate 1/ω3), and hence we can expect an increasingly stable recovery of
12
f as z → 1 (for points in the scanning region closer to the detector line and the photon
source).
ω3
source line
detector line
ω1 = − t0√ω2
3−1
O
ω1
ω1 = t0√ω2
3−1
Figure 2. The RTT geometry (with the same dimensions as in figure 1
and a visualisation of the set S = {ω1 > − t0√ω2
3−1}∩{ω3 <
1} (shaded in blue) of ω for which f (ω) is determined without the need
for analytic continuation. In 2-D t0 = π
2 . In 3-D t0 is the first root of J0
and the set of ω recovered without analytic continuation is the volume of
revolution of S about z.
3−1}∩{ω1 < t0√ω2
5. A generalization to the surfaces of revolution of C 1 curves
So far we have considered the surfaces of revolution of semicircles in translational ge-
ometries and proven injectivity and explicit invertibility results for classes of L2 functions
of compact support. Here we consider a more general class of surfaces, which are the sur-
faces of revolution of C 1 curves, and we prove injectivity results on the set of continuous
functions of compact support.
For some rm > 1, let ρj(r, z) ∈ C 1([1, rm]2), for j = 1, . . . , m, parameterize a finite set
of surfaces of revolution in cartesian coordinates
xj = ρj(r, z) cos ϕ,
yj = ρj(r, z) sin ϕ,
(5.1)
Let Ω ⊂ {2 − rm < z < 1} ⊂ R3 be compact. Then we define the generalized apple
transform RA : C0(Ω) → C([1, rm] × R2)
(5.2)
0 ≤ ϕ ≤ 2π.
1 < z < rm,
1 s1 +(cid:18)dρj
dz(cid:19)2
mXj=1Z π
−πZ r
RAf (x0, y0, r) =
ρj(r, z)f (ρj(r, z) cos ϕ+x0, ρj(r, z) sin ϕ+y0, 2−z)dzdϕ
We now have our third main theorem, which is a natural extension of the results of
sections 3 and 4.
13
Theorem 5.1. Let Ω ⊂ {2−rm < z < 1} ⊂ R3 be compact and let ρj(r, z) ∈ C 1([1, rm]2),
for j = 1, . . . , m be a finite set of C 1 curves. Then the generalized apple transform
RA : C0(Ω) → C([1, rm] × R2) in (5.2) is injective.
Proof. We have
(5.3)
K(r, z) f1(ω1, ω2, z)dz,
where f1(x, y, z) = f (x, y, 2 − z) as before, and
K(r, z) =
exp(−iρj(r, z)(ω1 cos ϕ + ω2 sin ϕ)dϕ
(5.4)
1 s1 +(cid:18)dρj
dz(cid:19)2
dRAf (ω1, ω2, r) =Z r
ρj(r, z)Z π
mXj=1
mXj=1
= 2π
−π
where J0 is a Bessel function of the first kind and ω = pω2
without loss of generality we can assume that ρj ≥ 0 for all j. Now let
1 + ω2
2. Let ρj < Mj and
(5.5)
K(r, z) f1(ω1, ω2, z)dz,
ρj(r, z)J0(ωρj(r, z)),
0 =Z r
1 s1 +(cid:18)dρj
dz(cid:19)2
for r ∈ [1, rm]. Let t0 be the first root of J0 and let ω < t0
M , where M = maxj Mj. Then
for such w, the integrand (5.5) is positive except for f1, and hence the only continuous
solution to (5.5) is f1(ω1, ω2, z) = 0 for z ∈ [1, rm] and ω < t0
M . Hence it follows that the
3-D Fourier transform of f1 is zero on the open set
, ω3 ∈ R(cid:27) .
B =(cid:26)qω2
The result follows by the Paley-Weiner-Schwartz theorem and analytic continuation. (cid:3)
1 + ω2
t0
M
2 <
Remark 5.2. The above proof uses the same arguments to that of Cormack [4, page
2724] in proofs of injectivity for Volterra integral equations, which suggest that Volterra
operators of the first kind with positive kernel, are injective on the domain of continuous
functions. We wonder if such ideas could be extended to prove injectivity in the L2 case.
6. Conclusions and further work
Here we have presented new injectivity results and explicit inversion formulae for CST
problems in translational geometries. We considered the problem of electron density
reconstruction from sets of toric section (in 2-D) and apple integral data (in 3-D). In
section 3 we introduced a new two dimensional toric section Radon transform T which
describes the integrals of an L2 density over the set of toric sections translated along a
line, whose central axis is vertical (parallel to the z axis). Here we provided boundedness
theorems for T in L2 and went on the prove the injectivity and explicit invertibility of T on
0. After proving the L2 injectivity in the 2-D case, we considered the 3-D case in section
L2
4, where we introduced a new apple Radon transform A in a three dimensional scanning
modality previously introduced in [17] (configuration (d) on page 5). The transformation
A takes an L2 density to its sinogram of integrals over vertical apples which are translated
in the xy plane. In a similar vein to the 2-D case, we showed the continuity of the apple
operator in L2 and then went on to prove the injectivity and explicit invertibility of A
in L2
0. The proofs of injectivity in the 2-D and 3-D cases followed similar ideas, solving
14
a set of 1-D Volterra operators to recover the density in an open subset of the Fourier
domain, and then extending to the entire Fourier space uniquely by analytic continuation
(which was possible using Paley-Weiner-Schwartz ideas as the densities considered were
of compact support). In the final part of the paper in section 5 we presented a generalized
apple Radon transform RA, which describes the integrals of a density over the surfaces of
revolution of C 1 curves in the translational geometry. Here we proved the injectivity of
RA on the space of continuous functions of compact support using the results of Cormack
[4].
The work of [15] characterizes microlocally the image artefacts in a reconstruction from
toric section integral data in rotational geometries. Further, an iterative reconstruction
scheme using TV is found to be effective in suppressing the discovered artefacts. In further
work we aim to consider the microlocal properties of T and A to investigate the existence
of image artefacts. The inversion formulae presented here rely on analytic continuation
in the Fourier domain, and hence the inversion process is severely ill-posed, and we would
expect a large amplification of the measurement error in the solution. However, this is
only one possible inversion approach, and is not proof that the problem of recovering f
from T f or Af is severely ill-posed. We aim to clarify this in future work. Following
this, we aim to devise a reconstruction method as in [15] which can offer a practically
useful image quality for the desired application in airport baggage screening.
Acknowledgements
We would like to thank Prof. Eric Todd Quinto for his helpful discussions, comments
and insight towards the results presented in this paper.
This material is based upon work supported by the U.S. Department of Homeland
Security, Science and Technology Directorate, Office of University Programs, under Grant
Award 2013-ST-061-ED0001. The views and conclusions contained in this document are
those of the authors and should not be interpreted as necessarily epresenting the official
policies, either expressed or implied, of the U.S. Department of Homeland Security.
References
[1] Tricomi, Francesco Giacomo. Integral equations. Vol. 5. Courier Corporation, 1985.
[2] Natterer, Frank. The mathematics of computerized tomography. Society for Industrial and Applied
Mathematics, 2001.
[3] Hormander, Lars. Linear partial differential operators. Vol. 116. Berlin: Springer-Verlag, 1969.
[4] Cormack, Allan Macleod. "Representation of a function by its line integrals, with some radiological
applications." Journal of applied physics 34, no. 9 (1963): 2722-2727.
[5] Plancherel, Michel, and Mittag Leffler. "Contribution tude de la reprsentation dune fonction arbi-
traire par des intgrales dfinies." Rendiconti del Circolo Matematico di Palermo (1884-1940) 30, no.
1 (1910): 289-335.
[6] Thompson, William Michael. "Source Firing Patterns and Reconstruction Algorithms for a Switched
Source, O set Detector CT Machine." PhD diss., The University of Manchester (United Kingdom),
2011.
[7] Webber, James."X-ray Compton scattering tomography." Inverse problems in science and engineer-
ing 24, no. 8 (2016): 1323-1346.
[8] Palamodov, V. P. "An analytic reconstruction for the Compton scattering tomography in a plane."
Inverse Problems 27, no. 12 (2011): 125004.
[9] Nguyen, M. K., and T. T. Truong. "Inversion of a new circular-arc Radon transform for Compton
scattering tomography." Inverse Problems 26, no. 6 (2010): 065005.
[10] Rigaud, Gal, Ma K. Nguyen, and Alfred K. Louis. "Novel numerical inversions of two circular-arc
Radon transforms in Compton scattering tomography." Inverse Problems in Science and Engineering
20, no. 6 (2012): 809-839.
15
[11] Truong, Tuong T., and Mai K. Nguyen. "New properties of the V-line Radon transform and their
imaging applications." Journal of Physics A: Mathematical and Theoretical 48, no. 40 (2015):
405204.
[12] Truong, T. T., and M. K. Nguyen. "Radon transforms on generalized Cormacks curves and a new
Compton scatter tomography modality." Inverse Problems 27, no. 12 (2011): 125001.
[13] Michael R. Walker II, Joseph A. O'Sullivan "The Broken Ray Transform: Additional Properties
and New Inversion Formula" arXiv:1904.00341
[14] Norton, Stephen J. "Compton scattering tomography." Journal of applied physics 76, no. 4 (1994):
2007-2015.
[15] Webber, James, and Eric Todd Quinto. "Microlocal analysis of a Compton tomography problem."
arXiv preprint arXiv:1902.09623 (2019).
[16] Truong, Tuong T., Ma Khuong Nguyen, and Habib Zaidi. "The mathematical foundations of 3D
Compton scatter emission imaging." International journal of biomedical imaging 2007 (2007).
[17] Webber, James W., and William RB Lionheart. "Three dimensional Compton scattering tomogra-
phy." Inverse Problems 34, no. 8 (2018): 084001.
[18] Rigaud, Gal, and Bernadette N. Hahn. "3D Compton scattering imaging and contour reconstruction
for a class of Radon transforms." Inverse Problems 34, no. 7 (2018): 075004.
[19] Jung, Chang-Yeol, and Sunghwan Moon. "Inversion formulas for cone transforms arising in appli-
cation of Compton cameras." Inverse Problems 31, no. 1 (2015): 015006.
[20] Webber, James, and Sean Holman. "Microlocal analysis of a spindle transform." arXiv preprint
arXiv:1706.03168 (2017).
[21] Agranovsky, Mark L., and Eric Todd Quinto. "Injectivity sets for the Radon transform over circles
and complete systems of radial functions." journal of functional analysis 139, no. 2 (1996): 383-414.
[22] Sogge, Christopher D. Fourier integrals in classical analysis. Vol. 210. Cambridge University Press,
2017.
Department of Electrical and Computer Engineering, Tufts University, Medford,
MA USA
E-mail address: [email protected]
Department of Electrical and Computer Engineering, Tufts University, Medford,
MA USA
E-mail address: [email protected]
16
|
1703.06771 | 1 | 1703 | 2017-03-20T14:24:38 | Riesz Wavelets, Tiling and Spectral Sets in LCA Groups | [
"math.FA",
"math.RT"
] | This paper is devoted to the study of geometry properties of wavelet and Riesz wavelet sets on locally compact abelian groups. The catalyst for our research is a result by Wang ([32], Theorem 1.1) in the Euclidean wavelet theory. Here, we extend the result to wavelet and Riesz wavelet collection of sets in infinite locally compact abelian groups. | math.FA | math | Riesz Wavelets, Tiling and Spectral Sets in LCA
7
1
0
2
r
a
M
0
2
]
Groups
Azita Mayeli
September 28, 2018
Abstract
.
A
F
h
t
a
m
[
1
v
1
7
7
6
0
.
3
0
7
1
:
v
i
X
r
a
This paper is devoted to the study of geometry properties of wavelet and Riesz
wavelet sets on locally compact abelian groups. The catalyst for our research is
a result by Wang ([32], Theorem 1.1) in the Euclidean wavelet theory. Here, we
extend the result to wavelet and Riesz wavelet collection of sets in infinite locally
compact abelian groups.
Keywords: LCA groups, wavelet set, Riesz wavelet collection of sets, spectral set, tiling
with multiplicity,
AMS 2000 Mathematics Subject Classification: Primary: 43A25, 43A46, 42C40; Sec-
ondary: 52C22
1
Introduction
Let ϕ ∈ L2(Rn), n ≥ 1. ϕ is called a wavelet for L2(Rn) if there exists countable sets
L ⊂ Rn and D ⊂ GL(n, R) such that the family
{ det D1/2ϕ(Dx − t) : D ∈ D, t ∈ L}
(1.1)
is an orthonormal basis for L2(Rn). In this case, the sets D and L are called dilation and
translation sets, respectively. Wavelets are important tools in approximation theory due
to their time and frequency localization property.
One type of wavelets with simplest structures are those which are supported on a Lebesgue
measurable set in Rn with finite and positive measure. The most basic one is the Haar
wavelet ϕ(x) = 1[0,1/2) − 1[1/2,1), constructed by Alfred Haar in 1910, is a one-dimensional
example of this kind and it serves as a prototypical wavelet. The dilation and trans-
lation sets associated to the Haar wavelet are given by D = {2m : m ∈ Z} (dyadic
dilation) and L = Z. Another type of simple wavelets are those whose Fourier transforms
1
2
are characteristic functions of a Lebesgue measurable domain in Rn with finite measure.
A typical example of these kind is Littlewood-Paley or Shannon wavelet whose Fourier
transform is 1[−1,−1/2)∪(1/2,1]. Shannon wavelet is a significant tool in sampling and recon-
struction of functions. Haar wavelet and Shannon wavelet are situated at the opposite
side of each other and "somehow" related. However, in this paper we shall only focus
on Shannon-type wavelets, i.e., those with Fourier transform supported in a domain with
finite measure.
The Fourier support of the Shannon wavelet I := [−1, −1/2) ∪ (1/2, 1] has this nice geom-
etry property that its dyadic dilations by {2n : n ∈ Z} as well as it integer translations
tile R. The sets with the similar tiling properties have been considered in the paper by
Dai and Larson [10] where the authors proved the following results: Given a measurable
set E ⊂ R with finite and positive measure, the function ϕ with ϕ = 1E is a wavelet
with dyadic dilation and integer translations if and only if E tiles R by dyadic dilations
as well as by integer translations. This result and characterization of wavelets in terms
of tilings properties of their Fourier supports was somehow generalized by Wang [32] in
n-dimensional Euclidean spaces, where the dilation set is any countable subset of invert-
ible matrices and the translation set is any countable subset in Rn. Before we recall his
result, we need several definitions.
Definition 1.1. A measurable subset Ω ⊂ Rn with finite and positive measure is called
wavelet set if for the function ϕ ∈ L2(Rn) with ϕ = 1Ω the family (1.1) is an orthonormal
basis for L2(Rn).
These wavelets are also known as MSF wavelets (minimally supported frequency wavelets)
or in older literatures as s-wavelets. The wavelet sets and minimally supported frequency
wavelets were introduced by Fang and Wang ([13]) and studied exclusively by Hernandez,
Wang and Weiss ([17, 18]). For a series of paper on wavelet sets see e.g., [2, 4, 11, 12] and
references contained therein.
Additive and multiplicative tilings. We say a set Ω tiles Rn additively with respect to a
countable set L ⊂ Rn if Xℓ∈L
1Ω(x − ℓ) = 1 a.e. x ∈ Rn.
Equivalently, Rn = ∪ℓ∈LΩ + ℓ, where for any distinct ℓ and ℓ′, the sets Ω + ℓ and Ω + ℓ′ are
disjoint in measure, i.e., Ω + ℓ ∩ Ω + ℓ′ = 0. Here, Ω denotes the Lebsgue measure. We
say Ω is a multiplicative tiling set for Rn with respect to a collection of n by n invertible
matrices A if {α(Ω) : α ∈ A} is a partition for Rn, i.e.,
Xα∈A
1α(Ω)(x) = 1 a.e. x ∈ Rn.
This is also equivalent to say that Rn = ∪α∈Aα(Ω) where for any α 6= α′ the two sets
α(Ω) and α′(Ω) are disjoint in measure.
The characterization of the wavelet sets in terms of translation and multiplicative tiling
was obtained by Wang (Theorem 1.1, [32]).
3
Theorem A. Let Ω ⊂ Rn with finite and positive Lebesgue measure, A ⊂ GL(n, R) and
L be a countable subset of Rn. For α ∈ A, let ατ denote the transpose matrix of α. If
the set {ατ (Ω) : α ∈ A} tiles Rn and the set of exponentials EL := {e2πil·x :
l ∈ L} is an
orthogonal basis for L2(Ω), then Ω is a wavelet set with respect to the dilation set A and
translation set L. The converse also holds, provided that 0 ∈ L.
Here, l · x denote the inner product in Rn and "x" is generic.
In his paper [32], Wang shows that the additional condition 0 ∈ L can not be dropped.
Indeed, Ω = [−1, −1/2] ∪ [1/2, 1] is a wavelet set with respect to the dilations M =
{±2n}n∈Z and translations T = 2Z + 1/6. However, Ω nor tiles R multiplicatively by M
neither the exponentials ET is dense in L2(Ω). The latter is true since, for example, the
non-zero function g(x) = e−πix/3 is orthogonal to all elements of ET with inner product in
L2(Ω).
In the classical context, given Ω and Λ, the (Ω, Λ) is called spectral pair if the exponentials
EΛ := {e2πiλ·x : λ ∈ Λ} from an orthogonal basis for L2(Ω). In this case, Ω is spectral and
Λ is spectra (or spectrum). It has been shown by Fuglede ([15]) that when Λ is a lattice
in Rn, then the existence of exponential orthogonal basis EΛ for L2(Ω) is equivalent to say
that Ω tiles Rn additively by the dual lattice Λ⊥. The relation between exponential basis
for spectral sets and their tiling property in general has been conjectured by Fuglede
in '74. The Fuglede Conjecture asserts that given a bounded set Ω ⊂ Rd of positive
Lebesgue measure, L2(Ω) has an orthogonal basis of exponentials EΛ = {e2πiλ·x}λ∈Λ for
some countable subset Λ ⊂ Rn if and only if Ω tiles Rd by additively by some translations
T . The Fuglede Conjecture led to considerable activity in the past three decades. For
example, in 2001, Iosevich, Katz, Tao ([20]) proved the conjecture for complex planar
domain. In [26] and [27], Laba established the conjecture for the union of two intervals
on the real line and provided a series of connections between the study of orthonormal
bases and interesting problems in algebraic number theory. However, in 2004, Tao ([31])
disproved the Fuglede conjecture for dimension 5 and higher, followed by [24] double check
the reference, where Kolountzakis and Matolcsi also disprove the Fuglede Conjecture for
dimensions 3 and higher. Yet the conjecture may still be true in several important special
cases. In spite of the disproof of its general validity, the conjecture has still generated
many interesting results. Most recently, it was proved that the Fuglede Conjecture holds
in some abelian finite groups ([22]). For the link between tiling property and existence of
general bases in locally compact abelian groups see [3].
k-additive tiling. Given an integer number k ≥ 1, we say Ω multi-tiles Rd additively by a
multiset L with multiplicity k (or is a k-additive tiling) when
Xℓ∈L
1Ω(x − ℓ) = k
a.e. x ∈ Rn.
By the additive multi-tiling property, almost every point in Rn can be covered by exactly
k-translations of Ω. In dimension d = 2, it is known that every k-additive tiling of R2 by
a convex polygon, which is not a parallelogram, is a k-additive tiling with a finite union
4
of two-dimensional lattices ([23]). When k = 1, then we have the tiling situation. The
multi-tiling property of a bounded measurable set is a sufficient condition for existence
of exponentials set EΛ which is a Riesz bases for L2(Ω). This result was proved by
Kolountzakis ([25], Theorem 1).
An extension of the sufficiency part of Theorem A has been recently proved by Fuhr and
Maus [14] where the authors replace the additive tiling by additive multi-tiling property,
thus the orthogonality by Riesz bases condition. Their result is the following.
Theorem B. ([14], Theorem 1.2) Let Ω ⊂ Rn be a bounded and measurable set with pos-
itive measure. Assume that Ω tiles Rn multiplicatively by a countable subset of GL(n, R)
and multi-tiles Rn additively by a lattice with multiplicity k (k ∈ N). Then Ω is a Riesz
wavelet set in Rn, i.e, the system (1.1) for ϕ with ϕ = 1Ω is a Riesz basis for L2(Rn).
The focus of this paper is to extend the results of Theorems A and B for infinite locally
compact topological abelian (LCA) groups (G, +) with identity e and equipped with a
Haar measure. Note that some LCA groups, like p-adic additive groups Qp, the field of p-
adic rational numbers, which do not posses any lattice. Therefore, the classical definition
of wavelet sets does not apply to such groups. It is also known that the finite abelian group
do not posses any wavelet set in the traditional sense ([21]). Therefore, we exclude such
cases here and assume that G is infinite and admits a lattice, i.e., a discrete subgroup
which is co-compact. For the notion of wavelet sets in Qp and in a non-commutative
setting we refer to [5] and [9], respectively.
For a characterization of Riesz wavelets generated by MRA see e.g., [6] and [16].
The purpose of this paper is to establish the necessary formalism to state and prove
Theorem A and Theorem B in the setting of locally compact abelian groups. We shall
also prove the converse of Theorem B, not only in Euclidean space but also in the general
context.
In the classical setting of Rn, the definition of a wavelet is associated to two type of unitary
operators: dilations and translations. The translation operator for the functions on an
LCA group can be defined by the group action analogous with the Euclidean setting. The
dilation of a function defined on Rn is given by the action of an invertible matrix A and
the associated normalization factor c = det An/2. Therefore the immediate feeling is
that the dilation operator acting on the functions on a LCA group must be defined by
using group automorphisms. The only obstacle here would be the normalization factor.
Obviously, for discrete groups, the normalization factor is c = 1 since the Haar measure
is the counting measure. Therefore we shall assume that our setting is not discrete. In
this case we overcome the problem by using modular function ∆, which is a positive
homomorphism defined on the class of automorphisms of G. We will define this function
in Section 4.
5
2 Statement of main results
Borrowing the notion of wavelet collection of sets for the Euclidean setting in ([4]), we
have the following definition.
Definition 2.1. Let G = (G, +) be an LCA group with identity e and the dual group G.
We equip G by a Haar measure. A collection of measurable subsets Ωi, 1 ≤ i ≤ m, of G is
called a (Riesz) wavelet collection of sets for G if there are countable subset A of Aut(G)
and countable subsets Λi ⊂ G (1 ≤ i ≤ m) such that
∪m
i=1{ψi,λ,a : 1 ≤ i ≤ m, λ ∈ Λi, a ∈ A}
is (Riesz) orthogonal basis for L2(G), where ψi,λ,a(x) = ∆(a)1/2ψi(a(x) − λ) and ψi = 1Ωi.
When m = 1, we may say Ω = Ω1 is a (Riesz) wavelet set.
The study of wavelet sets is an interplay between group theory, geometry, operator theory
and analysis. The following result is an extension form of Theorem A and Theorem 2.2
in [4]. In the sequel, we assume that Ω1, · · · , Ωm are m mutual disjoint and measurable
subsets of G with finite and non-zero Haar measure, and Ω := ∪m
i=1Ωi.
Theorem 2.2. Let Ω1, · · · , Ωm be given. Let A ⊂ Aut( G) and Λi ⊂ G, 1 ≤ i ≤ m, be
countable and non-empty subsets. Assume that for every 1 ≤ i ≤ m, the pair (Ωi, Λi) is
spectral and the set Ω tiles G multiplicatively by the automorphisms A. Then {Ω1, · · · , Ωm}
is a wavelet collection of sets for L2(G) with respect to the dilations A and translations
Λi. The converse also holds, provided that e ∈ ∩m
i=1Λi.
Theorem 2.2 along with a result of Fuglede [15] yields the following corollary, where the
spectral condition has been replaced by additive tiling property.
Corollary 2.3. The set {Ω1, · · · , Ωm} is a wavelet collection of sets for L2(G) with re-
i=1Ωi tiles G
spect to the translations by lattices Λi (1 ≤ i ≤ m) if and only if Ω := ∪m
multiplicatively and each Ωi tiles G additivity with Λ⊥
i .
The following result is an extension form of Theorem 2.2 to the Riesz bases.
Theorem 2.4. Let Ω1, · · · , Ωm be as above and countable sets Λi ⊂ G be given such that
for any 1 ≤ i ≤ m the pairs (Ωi, Λi) are Riesz spectral with Riesz constants Li and Ui.
Moreover, assume that Ω is a multiplicative tiling for G with respect to the automorphisms
A ⊂ Aut( G). Then Ω1, · · · , Ωm is a Riesz wavelet collection of sets for L2(G), i.e., the
family
∪m
i=1{ψi,λ,a : λ ∈ Λi, a ∈ A}
(2.1)
is a Riesz basis for L2(G) where ψi := 1Ωi. The Riesz constants in this case are given by
L := min{Li} and U := max{Ui}.
A combination of Theorem 2.4 and [1, Theorem 4.1] yields the following result as an
extension version of Theorem B for m ≥ 1.
6
Corollary 2.5. Let Ω1, · · · , Ωm be as above. Assume that each Ωi multi-tiles G additively
by some lattice Γi ≤ G with multiplicity ki ∈ N, ki > 1. Furthermore, assume that Ω =
∪m
i=1Ωi tiles G multiplicatively with respect to a subset A of Aut( G). Then {Ω1, · · · , Ωm}
is a Riesz wavelet collection of sets for G.
An inverse theorem for the results in Theorem 2.4 follows.
Theorem 2.6. Assume that Ω is a Riesz wavelet set with respect to the translations
by Λ and dilations by A. Then Ω is a multiplicative tiling for G with respect to the
automorphisms A if and only if ( α(Ω), α−1(−Λ)) is a Riesz spectral pair for all α ∈ A.
Outline of the paper. The paper is organized as follows. After Section 3 on some
open problems, in Section 4 we recall necessary backgrounds on the Fourier transform
on locally compact abelian group G and wavelet and Riesz bases in L2(G). Theorem 2.2
asserts that a collection of measurable sets with finite and positive measure in the dual
space of a LCA group G generate an orthogonal wavelet bases for L2(G) if and only if the
sets are spectral and admit multiplicative tiling property, provided that the intersection
of all translation sets contains e. In section 5 we give the proof of Theorem 2.2. In Section
6 we prove Theorems 2.4 and 2.6. These theorems are as an extension version of both
necessary and sufficient part of Theorem 2.2. We conclude the paper by presenting some
examples of wavelet and Riesz wavelet sets in Section 7.
3 Open problems
In the example constructed by Wang [32], we observed that the wavelet set [−1, −1/2] ∪
[1/2, 1] does tile the space R multiplicatively on {±2n}n∈Z. However, the set Ω has a
multiplicative tiling property with respect to M with multiplicity 2. In this relation we
make the following definition.
Definition 3.1. Let Ω ⊂ G. We say Ω tiles G multiplicatively with multiplicity k if there
is a countable subset A of automorphisms of G such that
Xα∈A
1α(Ω)(x) = k
a.e. x ∈ G.
The case k = 1 is the traditional multiplicative tiling.
Question 3.2. Suppose that Ω is a wavelet set in a locally compact abelian group G with
respect to a dilation set A ⊂ Aut(G) and translation set Λ ⊂ G, where e 6∈ Λ. Must Ω
tile G multiplicatively by some A′ ⊂ Aut(G), A′ 6= A?
Notice that, when e ∈ Λ, then according to Theorem 2.2 the wavelet set is a multiplicative
tiling and A′ = A. A version of Question 3.2 for a non-separable wavelet index is the
following.
Question 3.3. Let B := {(λ, a) : λ ∈ G, a ∈ Auto(G)} be countable. Assume that for
ψ with ψ = 1Ω, the set {ψλ,a : (λ, a) ∈ B} is an orthogonal basis for L2(G), i.e., Ω is
7
a wavelet set with respect to B. Assume that (e, a) ∈ B for some a ∈ G. Must Ω tile
additively and multiplicatively?
In relation to Corollary 2.5 we have the following question.
Question 3.4. Suppose that {Ωi}m
i=1 is a Riesz wavelet collection of sets for L2(G), which
is not orthogonal. Let Ω := ∪iΩ. Must {α(Ω)}a∈A be a multiplicative tiling for G with
multiplicity k ≥ 1?
Acknowledgements. The author would like to thank Alex Iosevich and Shahaf Nitzan
for helpful conversations. Support for this project was partially provided by PSC-CUNY
by PSC-CUNY Award B ♯ 69625-00 47, jointly funded by The Professional Staff Congress
and The City University of New York.
4 Notations and Preliminaries
Following [30], let (G, +) denote a locally compact abelian group with the group identity
e. We assume G is second countable and we equip G with a normalized Haar measure
mG which is non-zero and translation invariant. Let G denote the dual group of G and
m G the Haar measure on G. For any element ξ in the dual group G, we associate the
character χξ : G → C of G and denote the duality by χξ(x) := hξ, xi, x ∈ G. When G in
n-dimensional Euclidean space Rn, then cRn = Rn as well and hξ, xi = e2πiξ·x, ξ, x ∈ Rn,
is the canonical mapping.
By the properties of characters, we have
χξ(−x) = χ−ξ(x) = χξ(x).
(4.1)
For any p > 0, analogously to the Euclidean case, we define the space Lp(G) with respect
to the measure mG by
Lp(G) = {f : G → C, f mG-measurable and ZG
f (x)pdmG(x) < ∞}.
(4.2)
For any f ∈ L1(G) ∩ L2(G), let f denote the Fourier transform of f given by
f (ξ) =ZG
f (x)χξ(x)dmG(x) , ξ ∈ G
(4.3)
By the Plancherel theorem the definition of the Fourier transform can be extend uniquely
to the functions in L2(G) (See, for example, 1.6.1 in [30]), so that for any f, g ∈ L2(G)
the Parseval identity holds:
ZG
f (x)g(x)dmG(x) =Z G
f (ξ)g(ξ)dm G(ξ) .
(4.4)
8
We denote topological automorphisms of G onto itself by Aut(G). For α ∈ Aut(G), the
adjoint homomorphism α : G → G is given as following: For any character χξ, α(χξ) is a
character and its duality is given by
α(χξ)(x) := χξ(α(x)) = hξ, α(x)i x ∈ G.
It is easy to show that α is an automorphism (one-to-one homomorphism) of G onto G
and its inverse α−1 acts on G by
α−1(χξ)(x) := χξ(α−1(x)) = hξ, α−1(x)i.
This implies that α−1 = dα−1. By the definition of the adjoint homomorphism, the follow-
ing result immediately follows.
Lemma 4.1. Given any automorphism α : G → G, ξ ∈ G and x ∈ G
α(χξ)(x) = χ α(ξ)(x)
and
α−1(χξ)(x) = χ α−1(ξ)(x)
A subgroup Λ of G is called a lattice if it is discrete and co-compact, i.e., the quotient
group G/Λ is compact. G is second countable, therefore any lattice in G is countable ([29,
Section 12, Example 17]). The annihilator or dual lattice of a lattice Λ is given by
Λ⊥ = {ξ ∈ G : χξ(λ) = 1 ∀λ ∈ Λ}.
Then Λ⊥ is also a lattice in G. By the duality theorem [30, Lemma 2.1.3], Λ⊥ is topolog-
ically isomorphic to the dual of G/Λ, that is, Λ⊥ ∼= \(G/Λ).
Let α : G → G be an automorphism. Then
α(Λ⊥) = (α−1(Λ))⊥.
Given a countable set Λ in G and a non-zero measurable set Ω ⊂ G, we say Ω tiles G
additively by Λ with multiplicity k ∈ N if
Xλ∈Λ
1Ω(x − λ) = k a.e. x ∈ G .
(4.5)
If k = 1, then we simply say Ω tiles G additively by Λ.
Given Ω ⊂ G with non-zero measure and a countable set T ⊂ G, we say (Ω, T ) is a
spectral pair if the countable family ET := {χt : t ∈ T } is an orthogonal basis for L2(Ω).
The pair is called Riesz spectral pair if ET constitute a Riesz basis for L2(Ω). The set
Ω ⊂ G is multiplicative tiling if there is a set of automorphisms A of G such that
Xα∈A
1α(Ω)(x) = 1 a.e. x ∈ G.
9
For general definition of multiplicative tiling sets in a measure space (X, µ) see [12].
Modular function. Associated to a locally compact abelian group G with the Haar measure
mG, the modular function ∆ : Aut(G) → (0, ∞) is a continuous homomorphism such that
for any measurable set Ω ⊂ G there holds
mG(α(Ω)) = ∆(α)mG(Ω).
(4.6)
It is well-known that the modular function ∆ exists. For this, consult [19, 26.21]. (The
existence of such map for a larger class of continuous group homomorphisms of G onto
G has also been proved in [7, Theorem 6.2].) The relation (4.6) implies that for any
integrable function f on G with respect to the Haar measure mG the equation
ZG
f (α(x))dmG(x) = ∆(α)−1ZG
f (x)dmG(x)
(4.7)
holds for all automorphism α of G. For our purpose, we shall assume that ∆ is not
identical to one.
Dilations and translations. Let G be a LCA group. There are two unitary operators
associated to the definition of wavelets on the group: translation and dilation. For x ∈ G,
the translation operator τx is defined on function f : G → C and given by τxf (y) =
f (y − x). By the translation invariant property of Haar measure, τx is unitary restricted
to the functions in L2(G). For a fixed automorphism α ∈ Aut(G), we define δa on f
by
δαf (x) = ∆(α)1/2f (α(x)), ∀ x ∈ G.
When ∆(α) 6= 1, we call δα dilation operator. In the sequel we shall assume that the
operator δα is a dilation. It is easy to see that by (4.7) the operator δα is an unitary
operator. For a mapping f : G → C, λ ∈ Λ and α ∈ Aut(G) we denote by fα,λ the
"dilation and translation copy" of f with respect to α and λ, respectively, and define it
by
fα,λ(x) := δατλf (x) = ∆(α)1/2f (α(x) − λ).
For x ∈ G and g : G → C, the modulation operator Mx on g is given by
Mxg(ξ) = hξ, xig(ξ)
ξ ∈ G.
The following result is immediate by the group Fourier transform.
Lemma 4.2. Let α be an automorphism of G and λ ∈ G. Then for any f ∈ L1(G)∩L2(G)
bfα,λ(ξ) = \δατλf (ξ) = ∆(α)−1/2(M−λ
f )( α−1(ξ)),
ξ ∈ G.
Proof. The result obtains from the application of the Fourier transform on the dilation
10
and translation operators, as follows. Let ξ ∈ G. Then by (4.3), the following holds.
fα,λ(x)χξ(x)dmG(x)
(4.8)
bfα,λ(ξ) =ZG
f (α(x) − λ)χξ(x)dmG(x)
f (x)χξ(α−1(x + λ))dmG(x)
= ∆(α)1/2ZG
= ∆(α)−1/2ZG
= ∆(α)−1/2χξ(α−1(λ))ZG
= ∆(α)−1/2χξ(α−1(λ))ZG
f (x)χξ(α−1x)dmG(x)
f (x)χ α−1(ξ)(x)dmG(x)
= ∆(α)−1/2χξ(α−1(λ)) f ( α−1(ξ))
= ∆(α)−1/2(M−λ
f )( α−1(ξ)).
This completes the proof. Note that within the above lines we used the property χξ(α−1(x)) =
χ α−1(ξ)(x) proved in Lemma 4.1 and the equation χξ(x) = χξ(−x) by (4.1).
Assume that f ∈ L2(G) with f = 1Ω. As a result of the previous lemma one has
bfα,λ(ξ) = ∆(α)−1/2χ α−1(ξ)(−λ)1 α(Ω)(ξ)
ξ ∈ G.
(4.9)
Convention. In the sequel, we shall use the notation "exponential" ex(ξ) := χξ(x) for
x ∈ G and ξ ∈ G.
Riesz bases. A countable sequence {xn}n∈I in a Hilbert space H is a Riesz basis for H if
the sequence is of the form {Uen}n∈I for some orthonormal basis {en}n∈I for H and an
invertible and bounded linear operator U : H → H. This definition is equivalent to say
that the sequence {xn}n∈I is dense in H and there is finite and positive constants A, B
such that for any finite sequence {cn}n∈I we have
2
AXn∈I
cn2 ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xn∈I
cnxn(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
≤ BXn∈I
H
cn2.
(4.10)
It is obvious that if the upper and lower estimations in (4.10) hold for any finite sequence
{cn}n∈I, then they also hold for any sequence {cn}n∈I ∈ ℓ2(I). For more about Riesz
bases for any Hilbert space we refer to [8].
The following result is well-known, but we give a proof below to keep the presentation as
self-contained as possible.
Lemma 4.3. Let {xn}n∈I be a Riesz basis for a Hilbert space H. Assume that for any
finite sequence {cn} the identity holds:
kXn
cnxnk = (Xn
cn2)1/2
(4.11)
Then {xn} is an orthonormal basis.
Proof. We only need to show that for any m, n ∈ I, hxn, xmi = δm,n The identity (4.11)
implies that kxM k = 1. For the orthogonality of the vectors we proceed as follows: By
(4.11), it is immediate that kxn + xmk2 = 2 for any n 6= m. From the other hand
kxn + xmk2 = 2 + 2Rehxn, xmi. This implies that we must have Rehxn, xmi = 0. By
replacing xn by ixn, we also obtain Imhxn, xmi = 0. These imply that xn and xm are
orthogonal, and we are done.
11
Definition 4.4. Let H be a infinite-dimensional Hilbert spaces. An infinite collection
{xn} of vectors in H is w-linearly independent if for any sequence {cn} such thatPn cnxn
converges to zero in the norm of H is identically zero.
It is well-known that any w-linearly independent sequence is linearly independent. How-
ever, the converse does not always hold.
Lemma 4.5. A sequence {xn} in a Hilbert spaces is a Riesz basis for H if and only if
{xn} is w-linearly independent and there are constants 0 < A ≤ B < ∞ such that for any
vector x ∈ H,
Akxk2 ≤Xn
hx, xni2 ≤ Bkxk2.
(4.12)
The inequality in the preceding lemma is called the frame inequality. The w-linearly
independency of Riesz basis follows from the fact that any Riesz basis has a biorthogonal
sequence. For the proof we refer to [8].
5 Proof of Theorem 2.2
For the rest of this paper we fix the following notations unless it is stated otherwise: We
let A denote a countable subset of Aut(G), the group of automorphisms of G onto G with
∆(α) 6≡ 1, α ∈ Aut(G), Λ be a countable subset in G, and Ω be a subset of G with finite
and non-zero Haar measure.
The first result of this section shows that the spectral property of a set is preserved under
the action of automorphisms.
Lemma 5.1. Assume that (Ω, Λ) is a spectral pair and α ∈ A. Then the pair ( α(Ω), α−1(Λ−1))
is spectral, where Λ−1 = {−λ : λ ∈ Λ}.
Proof. Define Tα : L2(Ω) → L2( α(Ω)) by Tα(g)(ξ) = ∆( α)−1/2g( α−1(ξ)), ξ ∈ α(Ω). T is
linear and
Tα(eλ) = ∆( α)−1/2eα−1(−λ) ,
λ ∈ Λ
T is a unitary isomorphism and (Ω, Λ) is a spectral pair, then {∆( α)−1/2eα−1(−λ)}λ∈Λ is
an orthogonal basis for L2( α(Ω)). This completes the proof of the lemma.
12
Proof of Theorem 2.2. We shall "somehow" adjust the proof of Theorem A for our situ-
ation. Fix A ⊂ Aut(G) and Λi ⊂ G and let {Ωi}m
i=1 be a collection of disjoint subsets in
G such that (Ωi, Λi) are spectral pairs.
Let ϕi ∈ L2(G) be the inverse Fourier transform of 1Ωi, i.e., ϕi = 1Ωi. Define
W := ∪m
i=1{ϕi,α,λ : α ∈ A, λ ∈ Λi}.
(5.1)
Then by an application of the Fourier transform and the equation (4.9), W is an orthogonal
basis for L2(G) if and only if the system cW is an orthogonal basis for L2( G):
(5.2)
cW = ∪m
= ∪m
i=1{[ϕi,α,λ : α ∈ A, λ ∈ Λi}
i=1{∆(α)−1/2eα−1(−λ)1 α(Ωi) : α ∈ A, λ ∈ Λi}.
By the assumption, each pair (Ωi, Λi) is a spectral. Therefore, for any α ∈ A, by Lemma
5.1 the collection {eα−1(−λ)1 α(Ωi) : λ ∈ Λi} is an orthogonal basis for L2( α(Ωi)). From the
other hand, by multiplicative tiling property of {Ωi}m
i=1, for any α, β ∈ A and 1 ≤ i, j ≤ m
we have
m G( α(Ωi) ∩ β(Ωj)) = m G( α(Ωi))δα,βδi,j.
(5.3)
Thus the elements of cW are mutual orthogonal. The completeness of the system cW in
L2( G) holds by the A-multiplicative tiling of G and the decomposition
L2( G) =Mα∈A M1≤i≤m
L2( α(Ωi)).
(5.4)
Conversely, let {Ωi}m
i=1 be a wavelet collection of sets with respect to the automorphisms
A and translations Λi, 1 ≤ i ≤ m. Assume that e ∈ Λi for all i. Then the elements of
the collection {ϕi,α := ϕi,α,0 : α ∈ A, 1 ≤ i ≤ m} are mutual orthogonal and for any
(i, α1) 6= (j, α2) we have
0 =ZG
=Z G
=Z G
ϕi,α1(x)ϕj,α2(x)dmG(x)
(5.5)
ϕi,α1(ξ) ϕj,α2(ξ)dm G(ξ)
1 α1(Ωi)(ξ)1 α2(Ωj )(ξ)dm G(ξ)
= m G( α1(Ωi) ∩ α2(Ωj)).
This implies that for Ω = ∪m
i=1Ωi, the subsets α(Ω), α ∈ A are mutual disjoint and form
a covering for G. Indeed, assume that there is a subset W ⊂ G of positive and finite
Haar measure such that W ∩ α(Ω) = 0 for all α ∈ A. Then W ∩ α(Ωi) = 0 for all
1 ≤ i ≤ m. Then for the function f : G → C with f = 1W is hf, ϕi,α,λi = h1W , [ϕi,α,λi = 0
13
for all i, α and λ ∈ Λi. From the other hand, {Ωi}m
i=1 is a wavelet collection of sets. Then
f must be identical to zero. This contradicts the assumption that W > 0, hence Ω tiles
G multiplicatively by A.
For the rest, we prove that the pairs (Ωi, Λi) are spectral. To this end, according to
Lemma 5.1, it is sufficient to show that for any α ∈ A each pair ( α(Ωi), α−1(−Λi)) is a
spectral. Or, equivalently, the system
{∆(α)−1/2eα−1(−λ)1 α(Ωi) : λ ∈ Λi}
is orthogonal and complete in L2( α(Ωi)). The orthogonality holds from the assumption
that {Ωi}m
i=1 is a wavelet collection of sets. For the completeness, assume that g ∈
L2( α(Ωi)) such that hg, eα−1(−λ)iL2(α(Ωi)) = 0 for all λ ∈ Λi. This, along the mutual
disjointness of { α(Ωj) : α ∈ A, j}, yields
kgk2 =Xβ∈A
mXj=1 Xλ∈Λj
hg1 α(Ωi), ∆(β)−1/21 β(Ωj )eβ−1(−λ)i2 = 0,
which implies that g must be zero. Thus ( α(Ωi), α−1(Λ−1)) is a spectral pair and we are
done.
Remark. Note that the multiplicative tiling property of the set Ω for G is required for the
completeness of the wavelet system in L2(G). However, the results hold for a subspace of
L2(G) if we only assume that Ω is a multiplicative tiling for a subset of G.
Example 5.2. We shall use the multiplicative tiling property of the Shannon set to con-
struct a simple example of one dimensional orthogonal wavelet basis for a subspace of
L2(R). Let Ω1 := [−1, −1/2 − ǫ1) and Ω2 := (1/2 + ǫ2, 1] where 0 ≤ ǫi < 1/2 . Then
(Ωi, αiZ) is a spectral pair with αi = 2(1 − 2ǫi)−1. The set Ω = Ω1 ∪ Ω2 is a multiplicative
tiling for some subset U ⊂ R with respect to the dyadic dilations {2n : n ∈ Z}. This
along with the result of Theorem 2.2 implies that the collection cW (5.2) is an orthogonal
basis for L2(U).
6 Proof of Theorems 2.4 and 2.6
To prove Theorem 2.4, we need to prove an analogous result to Lemma 5.1 for Riesz
spectral pairs, as follows.
Lemma 6.1. Let (Ω, Λ) be a Riesz spectra with Riesz constants A and B. Then for any
α ∈ A, the pair ( α(Ω), α−1(Λ−1)) is a Riesz spectral with unified Riesz constants.
Proof. The proof is straight-forward and uses an application of the operator Tα in Lemma
5.1, and the fact the image of a Riesz basis under any unitary map is Riesz basis with the
unified constants.
Proof of Theorem 2.4. By unitary Fourier transform it is sufficient to show that
14
i=1{ ψi,α,λi} = ∪m
i=1{∆(α)−1/2eα−1(−λi)(ξ)1 α(Ωi) : α ∈ A, λi ∈ Λi}
cW := ∪m
is a Riesz basis for L2( G). By the assumptions, for any i, the pair (Ωi, Λi) is Riesz spectral.
Therefore by Lemma 6.1 for any α ∈ A the pair ( α(Ωi), α−1(−Λi)) is Riesz spectral with
the unified constants. Then the completeness of the system [WA,T in L2( G) holds by the
A multiplicative tiling assumption of Ω = ∪iΩi for G and the decomposition L2( G) =
Lα∈A L2( α(Ω)) = Lα∈A,1≤i≤m L2( α(Ωi)). To prove the Riesz sequence approximation
inequalities, we continue as follows: Let {cα,λi
: α ∈ A, λi ∈ Λi, i = 1, · · · , m} be a
finite collection of numbers. Then, due the disjointness of α(Ωi), i, and that the pairs
(α(Ωi), Λi) are Riesz spectral with upper bound Ui, we can write the following:
k Xα,i,λi
cα,λi
ψi,α,λik2
cα,λi
ψi,α,λk2
L2(α(Ωi))
cα,λi2
mXi=1
kXλi
UiXα Xλi
L2( G) =Xα
mXi=1
= U Xα,i,λi
≤
cα,λi2 ,
(6.1)
(6.2)
(6.3)
with U = max{Ui : 1 ≤ i ≤ m}. The lower estimate for Riesz sequence can be obtained
in the same fashion. This completes the proof of the theorem.
Note that the completeness of the Riesz system in the previous theorem can also be
obtained by a different approach: Let f ∈ L2(G) such that hf, ψi,α,λi = 0 for all λ, α and
1 ≤ i ≤ m. Fix α and i. Then hf 1α(Ωi), ψi,α,λi = 0 for all λ ∈ Λi. Since (Ωi, Λi) is a
Riesz spectral, and { ψi,α,λ} is complete in L2(α(Ωi)), then we must have f 1α(Ωi) = 0. But
f = ⊕i,αf 1α(Ωi), which implies that f must be zero.
Proof of Corollary 2.5. Assume that Ωi tiles G additively by some lattice Γi with multi-
plicity ki. Then by [1, Theorem 4.1], for each i, there is a countable set Λi ⊂ G such that
(Ωi, Λi) is a Riesz spectral pair in G. The rest of the proof is immediate from Theorem
2.4.
Example 6.2. Let Λ be a lattice such that, Ω1, Ω2, · · · , Ωk, be k mutual disjoint funda-
i=1Ωi is an additive tiling set for G with multiplicity
mental domains for Λ. Then Ω = ∪k
k with respect to translations Λ. Moreover, assume that Ω is a multiplicative tiling with
respect to a subset of automorphisms. Then Ω is a Riesz wavelet set. In this connection,
see Example 5.2 in Section 5.
Remark. Let (Ωi, Λ), 1 ≤ i ≤ m, be spectral pair, i.e., the exponentials EΛ form an
orthogonal basis for L2(Ωi).
i=1Ωi is a Riesz
If {Ωi} are mutual disjoint, then Ω = ∪m
spectral set with spectra Λ. To see this, note that for any finite {cλ}λ∈Λ we have
kXλ
cλeλk2
L2(Ω) = X1≤i≤m
kXλ
cλeλk2
L2(Ωi) = X1≤i≤mXλ
cλ2.
15
(6.4)
The completeness of EΛ in L2(Ωi) for disjoint sets Ωi, 1 ≤ i ≤ m, implies that EΛ is
complete in L2(Ω), too. This along the equation (6.4) and Lemma 4.11 completes the
proof.
Proof of Theorem 2.6. To prove the "if" part, let α 6= β and let M be any subset of
α(Ω) ∩ β(Ω). Put f := 1M . We show that M must have zero measure. Since f ∈ L2( G)
and also belongs to the spaces L2( α(Ω)) and L2( β(Ω)), then by the theorem's assumptions,
there are l2 sequences {cθ,λ : θ ∈ A, λ ∈ Λ}, {dα
λ : λ ∈ Λ} for which
we can write
λ : λ ∈ Λ} and {dβ
1M =Xθ,λ
cθ,λ
\δθτλψ =Xλ
dα
λ
\δατλψ =Xλ
dβ
λ
\δβτλψ.
Since the Riesz bases are w-linearly independent (Lemma 4.5), the preceding first and
second equalities imply that we must have cα,λ = dα
λ and cθ,λ = 0 for θ 6= α. With a
similar argument, by the first and third equalities we obtain cβ,λ = dβ
λ and cθ,λ = 0 for
θ 6= β. These conclude that dα
λ = 0 for all λ, hence 1M is the zero function and
M = 0. This implies the measure disjointness of α(Ω). To prove that { α(Ω) : α ∈ A} is
a cover for G, we shall use a contradiction approach and apply the frame inequality (4.12)
in Lemma 4.5 for the wavelet Riesz basis W = {δατλψ : α ∈ A, λ ∈ Λ}, respectively.
λ = dβ
To prove the "only if" part, notice that the disjointness of sets α(Ω) with the frame
inequality (4.12) imply that for any given α ∈ A, the exponentials EΛ are complete in
L2( α(Ω)). The Riesz inequality for EΛ is a direct implication of the Riesz inequality for
the wavelet system W.
Remark: Note that in Theorem 2.6, only the w-linearly independency of the system
{δατtψ} yields the disjointness of sets α(Ω).
7 Examples
Example 7.1. Let Ω = [−1, 1]d\[−1/2, 1/2]d. Fix a lattice Λ and write Ω as a finite union
of Ωi such that each Ωi is a tiling with respect the lattice Λ. For example, when d = 2, one
can take Λ = {(m/2, n/2) : m, n ∈ Z}. It is easy to see that {Ωi} is a wavelet collection
of sets. Indeed, each pair (Ωi, Λ⊥) is a spectral pair and Ω is a multiplicative tiling set for
Rd by the automorphisms A = {2nI : n ∈ Z}, where I is the d × d identity matrix. The
result now follows by Theorem 2.2. (For more examples of wavelet sets in Rd see e.g. [4].)
16
Example 7.2. Let Ωi are given as in Example 7.1. Assume that Ω := ∪iΩi is a multi-
plicative tiling with respect to the dyadic dilations {2nI : n ∈ Z}. Take Λ := 4−1Zd. Then
each pair (Ωi, Λ) tiles Rd additively with multiplicity 4, i.e., almost every point in Rd is
covered four times by Λ translations of each Ωi. Therefore Ωi is a Riesz spectral set with
spectra Λ. Theorem 2.4 implies that {Ωi} is a Riesz wavelet collection of sets in Rd.
Example 7.3. Let Ω1 be a symmetric polygon in R2 centered at origin. Take Ω2 = 2−1Ω1,
the dilation of Ω1 by scale 2−1. Define Ω = Ω1 \ Ω2. Then it is easy to check that Ω tiles
the plane mutliplicatively with respect to the dyadic dilations {2nI : n ∈ Z}. Assume that
Ωi, · · · , Ωm be m convex tiles of Ω which are pairwise disjoint. Furthermore, we assume
that each Ωi is a symmetric polygon with respect to a basis for R2. Since every convex
and symmetric polygon in R2 has a Riesz spectra ([28]), then each Ωi is a Riesz spectral
set with spectra Λi. Theorem 2.4 implies that {Ωi}m
1 is a Riesz wavelet collection of sets
for L2(R2).
The following example illustrates an approach to the construction of an orthogonal wavelet
basis on G induced by a wavelet set in a subgroup K.
Example 7.4. Let G be a locally compact abelian group which is topologically isomorphic
to Rn × D × K, where D discrete abelian group and K is compact. (Indeed, by Theorem
24.30 of [19], every LCA group has this form.) Assume that n is a non-negative integer
and D is a finite direct sum of finite abelian cyclic groups of prime power order, i.e., pr,
p prime. Let T := bK denote the Pontryagin dual group of K. T is discrete and a spectra
for K. Therefore, for given any spectral set S in Rd with spectra Λ, Ω := S × D × K is
spectral for G with spectra Λ × D × T .
Furthermore, assume that S tiles Rd multiplicatively with respect to a subset of automor-
phisms A ⊂ GL(R, d). For any α ∈ A define tα : G → G with tα(x, d, k) := (α(x), k).
Then tα, α ∈ A, is an automorphisms of G and {tα(S × D × K) : α ∈ A} is a mutual
disjoint tiling (a partition) for G. This implies that Ω is a wavelet set for L2(G).
References
[1] E. Agoraa, J. Antezanaa, C. Cabrelli, Multi-tiling sets, Riesz bases, and sampling
near the critical density in LCA groups, Advances in Mathematics, Volume 285, 5
November 2015, Pages 454 -- 477 [5, 14]
[2] L. W. Baggett, H.A. Medina, K. D. Merrill, Generalized multi-resolution analyses
and a construction procedure for all wavelet sets in Rn, J. Fourier Anal. Appl. 5
(1999), no. 6, 563 -- 573. [2]
17
[3] D. Barbieri, E. Hernandez, A. Mayeli, Tiling by lattices for locally compact
abelian groups, to appear in Comptes rendus Mathematique. DOI information:
10.1016/j.crma.2016.11.017 [3]
[4] J. J. Benedetto, R. L. Benedetto, The construction of wavelet sets, Wavelets and
Multiscale Analysis: Theory and Applications. Applied and Numerical Harmonic
Analysis pp 17 -- 56, 2011. [2, 5, 15]
[5] J. J. Benedetto and M. T. Leon, The construction of multiple dyadic minimally
supported frequency wavelets on Rd, AMS Contemporary Math., 247, (1999). [4]
[6] M. Bownik, Riesz wavelets and generalized multiresolution analysis, Appl. Comput.
Harmon. Anal. 14 (2003) 181 -- 194. [4]
[7] M. Bownik, K. Ross, The Structure of Translation-Invariant Spaces on Locally Com-
pact Abelian Groups, Journal of Fourier Analysis and Applications (2015), Volume
21, Issue 4, pp 849 -- 884 [9]
[8] O. Christensen, Frames and Bases, An introductory course, Birkhauser, Boston, 2008.
[10, 11]
[9] B. Currey, A. Mayeli, Gabor fields and wavelet sets for the Heisenberg group;
Monatsh. Math. (2011), 162:119 -- 142. [4]
[10] X. Dai, D. R. Larson, Wandering vectors for unitary systems and orthogonal wavelets,
Mem. Amer. Math. Soc. 134, no. 640 (1998). [2]
[11] X. Dai, D. R. Larson, D. M. Speegle, Wavelet sets in Rn, J. Fourier Anal. Appl.,
Volume 3, Number 4, 1997 [2]
[12] M. Dobrescu, G. ´Olafsson, Wavelet sets without groups, arXiv:0710.3508, preprint.
[2, 9]
[13] X. Fang and X.-H. Wang, Construction of minimally supported frequency wavelets,
J. Fourier Anal. Appl. 2 (1996), 315 -- 327. [2]
[14] H. Fuhr, Y. Maus, Wavelet Riesz bases associated to nonisotropic dilations, preprint.
Arxiv: 1510.01832 [4]
[15] B. Fuglede, Commuting self-adjoint partial differential operators and a group theoretic
problem, J. Funct. Anal. 16 (1974), 101 -- 121. [3, 5]
[16] B. Han, R-Q. J, Characterization of Riesz bases of wavelets generated from multires-
olution analysis, Appl. Comput. Harmon. Anal. 23 (2007) 321 -- 345. [4]
[17] E. Hernandez, X.-H. Wang and G. Weiss, Smoothing minimally supported frequency
wavelets. I, J. Fourier. Anal. Appl. 2 (1996), 329 -- 340. [2]
[18] E. Hernandez, X.-H. Wang and G. Weiss, Smoothing minimally supported frequency
wavelets. II, J. Fourier. Anal. Appl. 3 (1997), 23 -- 41. [2]
18
[19] E. Hewitt, K. A. Ross, Abstract harmonic analysis. Vol. I. Structure of topological
groups, integration theory, group representations. Second edition. Springer-Verlag,
Berlin-New York, 1979. [9, 16]
[20] A. Iosevich, N. Katz and T. Tao, The Fuglede spectral conjecture holds for convex
planar domains, Math. Res. Lett. 10 (2003), no. 5-6, 559-569. [3]
[21] A. Iosevich, C.K. Lai, A. Mayeli, Tight wavelet frame sets on finite vector spaces,
submitted. [4]
[22] A. Iosevich, A. Mayeli and J. Pakianathan, The Fuglede Conjecture holds in Zp × Zp,
to appear in Analysis and PDE. [3]
[23] . M. N. Kolountzakis, The study of translational tiling with Fourier Analysis, Lectures
given at the Workshop on Fourier Analysis and Convexity, Universita di Milano-
Bicocca, June 11 -- 22, (2001) [4]
[24] M. Kolountzakis and M. Matolcsi, Tiles with no spectra, Forum Math. 18 (2006), no.
3, 519-528. [3]
[25] M. N. Kolountzakis, Multiple lattice tiles and Riesz bases of exponentials, Proc. Amer.
Math. Soc. 143 (2015), no. 2, 741 -- 747. [4]
[26] I. Laba, Fuglede's conjecture for a union of two intervals, Proc. Amer. Math. Soc.
129 (2001), no. 10, 2965-2972. [3]
[27] I. Laba, The spectral set conjecture and multiplicative properties of roots of polyno-
mials, J. London Math. Soc. (2) 65 (2002), no. 3, 661-671. [3]
[28] Y. Lyubarskii and A. Rashkovskii, Complete interpolation sequences for Fourier
transforms supported by convex symmetric polygons, Ark. Mat. 38 (2000), no. 1,
139 -- 170. [16]
[29] L.S. Pontryagin, Topological Groups, Princeton Univ. Press (1946) (Translated from
Russian) [8]
[30] W. Rudin, Fourier Analysis on Groups, Interscience Tracts in Pure and Applied
Mathematics, No. 12, John Wiley and Sons, New York-London, 1962. [7, 8]
[31] T. Tao, Fuglede's conjecture is false in 5 and higher dimensions, Math. Res. Lett. 11
(2004), no. 2-3, 251-258. [3]
[32] Y. Wang, Wavelets, tiling, and spectral sets, Duke Math. J. 114 (2002), 43 -- 57.
[1, 2, 3, 6]
|
1705.04028 | 2 | 1705 | 2018-07-09T06:49:40 | Wave packet frames generated by hyponormal operators on $L^2(\mathbb{R})$ | [
"math.FA"
] | In this paper we study frame-like properties of a wave packet system by using hyponormal operators on $L^2(\mathbb{R})$. We present necessary and sufficient conditions in terms of relative hyponormality of operators for a system to be a wave packet frame in $L^2(\mathbb{R})$. A characterization of hyponormal operators by using tight wave packet frames is proved. This is different from a method proved by Djordjevi$\acute{c}$ by using the Moore-Penrose inverse of a bounded linear operator with a closed range. The linear combinations of wave packet frames generated by hyponormal operators are discussed. | math.FA | math |
WAVE PACKET FRAMES GENERATED BY HYPONORMAL
OPERATORS ON L2(R)
LALIT KUMAR VASHISHT
PRINCIPAL INVESTIGATOR
Abstract. In this paper we study frame-like properties of a wave packet sys-
tem by using hyponormal operators on L2(R). We present necessary and suffi-
cient conditions in terms of relative hyponormality of operators for a system to
be a wave packet frame in L2(R). A characterization of hyponormal operators
by using tight wave packet frames is proved. This is different from a method
proved by Djordjevi´c by using the Moore-Penrose inverse of a bounded linear
operator with a closed range. We extends some results by Kaushik, Singh and
Virender to wave packet frames generated by hyponormal operators .
1. Introduction
Frames in Hilbert spaces are a redundant system of vectors which provides a
series representation for each vector in the space. Duffin and Schaffer [11] in 1952,
introduced frames for Hilbert spaces, in the context of nonharmonic Fourier series.
Frames were revived by Daubechies, Grossmann and Meyer in [8]. For applications
of frames in various directions, see [3, 4]
Feichtinger and Werther [12] introduced a family of analysis and synthesis sys-
tems with frame-like properties for closed subspaces of a separable Hilbert space H
and call it an atomic system (or local atoms). The motivation for the atomic system
is based on examples arising in sampling theory. One of the important properties
of the atomic system is that it can generate a proper subspace even though they
do not belong to them.
Definition 1.1. [12] Let H be a Hilbert space and let H0 be a closed subspace of
H. A sequence {fk} ⊂ H is called a family of local atoms (or atomic system) for
H0, if
(i) there exists a real number B > 0 such that k{hf, fki}k2
ℓ2 ≤ Bkfk2 for all
(ii) there exists a sequence of linear functionals {ck} and a real number C > 0
f ∈ H,
such that
and
k{ck(f )}k2
ℓ2 ≤ Ckfk2 for all f ∈ H0
f =
∞
Xk=1
ck(f )fk for all f ∈ H0.
2010 Mathematics Subject Classification. 42C15, 42C30, 42B35, 47A05, 46B15.
Key words and phrases. Wave packet system, analysis operator, frame operator, hyponormal
operator, Hilbert space frame.
Lalit was supported by R & D Doctoral Research Programme, University of Delhi, Delhi-110007,
India. Grant No. : RC/2014/6820.
2
VASHISHT
Gavruta in [14] introduced and studied K-frames in Hilbert spaces to study
atomic systems with respect to a bounded linear operator K on Hilbert spaces.
Definition 1.2. [14] Let H be a Hilbert space and let K be a bounded linear
operator on H. A sequence {fk} ⊂ H is called a K-frame for H, if there exist
constants A, B > 0 such that
AkK ∗fk2 ≤
∞
Xk=1
hf, fki2 ≤ Bkfk2 for all f ∈ H.
(1.1)
The lower inequality in (1.1) is controlled by a bounded linear operator on H.
It is observed in [14] that K-frames are more general than standard frames in the
sense that the lower frame bound only holds for the elements in the range of K,
where K is a bounded linear operator on the underlying Hilbert space. Gavruta in
[14] characterize K-frames in Hilbert spaces by using bounded linear operators.
It would be interesting to control both lower and upper frame condition in (1.1)
by bounded linear operators on H. In this direction, we study frame-like properties
of an irregular wave packet system in L2(R), where both lower and upper frame
conditions are controlled by bounded linear operators on L2(R) (see Definition 3.1).
The wave packet system is a family of functions generated by combined action
of dilation, translation and modulation operators on L2(R). More precisely, we
consider a system of the form
{Daj TbkEcm ψ}j,k,m∈Z,
(1.2)
where ψ ∈ L2(R), {aj}j∈Z ⊂ R+, b 6= 0 and {cm}m∈Z ⊂ R and call it irregular
Weyl-Heisenberg wave packet system (or simply wave packet system) in L2(R). A
frame for L2(R) of the form {Daj TbkEcm ψ}j,k,m∈Z is called an irregular wave packet
frame (or wave packet frame). The wave packet system was introduced by Cordoba
and Fefferman [6] by applying certain collections of dilations, modulations and
translations to the Gaussian function in the study of some classes of singular integral
operators. Later, Labate et al. [20] adopted the same expression to describe, more
generally, any collection of functions which are obtained by applying the same
operations to a finite family of functions in L2(Rd). More precisely, Gabor systems,
wavelet systems and the Fourier transform of wavelet systems are special cases of
wave packet systems. Lacey and Thiele [21, 22] gave applications of wave packet
systems in boundedness of the Hilbert transforms. The wave packet systems have
been studied by several authors, see [7, 15, 17, 18, ?, ?].
1.1. Outline: This paper is organized as follows: In Section 2, we give basic defini-
tions and results which will be used throughout the paper. Section 3 is devoted to
the study of frame-like properties of irregular Weyl-Heisenberg wave packet systems.
We introduce Θ-irregular Weyl-Heisenberg wave packet frame (in short, Θ-IW H
wave packet frame) for L2(R), where Θ is a bounded linear operator on L2(R) (see
Definition 3.1). This type of wave packet frame can control both lower and upper
frame conditions by bounded linear operators on L2(R) . The Θ-IW H wave packet
frame (in the context of standard Hilbert frame) for a Hilbert space is a K-frame,
but converse is not true (see Example 3.2). Furthermore, the Θ-IW H wave packet
frame control both lower and upper frame conditions by bounded linear opera-
tors. Necessary and sufficient conditions for a certain system to be a Θ-IW H wave
packet frames for L2(R) by using hyponormality of operators on L2(R) have been
WAVE PACKET FRAMES
3
obtained. A characterization of hyponormal operator in terms of a special type of
tight wave packet frames for L2(R) is given. This is different from a method proved
by Djordjevi´c in [9] by using the Moore-Penrose inverse of a bounded linear oper-
ator with a closed range (see Theorem 3.7). The linear combinations of frames or
redundant building blocks are important in applied mathematics, we discuss linear
combinations of Θ-IW H wave packet frames for L2(R) in Section 4.
2. Preliminaries
In this section, we recall basic notations and definitions to make the paper self-
contained. Let H be a separable real (or complex) Hilbert space with inner product
h., .i linear in the first entry. A countable sequence {fk} ⊂ H is called a frame
(or Hilbert frame) for H, if there exist numbers 0 < ao ≤ bo < ∞ such that
∞
aokfk2 ≤
hf, fki2 ≤ bokfk2 for all f ∈ H.
(2.1)
The numbers ao and bo are called lower and upper frame bounds, respectively. They
are not unique. If it is possible to choose ao = bo, then the frame {fk} is called
Parseval frame (or tight frame).
The scalars
γ0 = inf{bo > 0 : bo satisfies (2.1)}
δ0 = sup{ao > 0 : ao satisfies (2.1)}
are called the optimal bounds or best bounds of the frame.
Associated with a frame {fk} for H, there are three bounded linear operators:
∞
Xk=1
synthesis operator V : ℓ2 → H,
analysis operator V ∗ : H → ℓ2,
frame operator S = V V ∗ : H → H, S(f ) =
V ({ck}) =
V ∗(f ) = {hf, fki}, f ∈ H,
Xk=1
∞
ckfk, {ck} ∈ ℓ2,
hf, fkifk, f ∈ H.
Xk=1
The frame operator S is a positive, self-adjoint and invertible operator on H.
This gives the reconstruction formula for all f ∈ H,
hS−1f, fkifk =
Xk=1
f = SS−1f =
Xk=1
∞
∞
hf, S−1fkifk ! .
The scalars {hS−1f, fki} are called frame coefficients of the vector f ∈ H. The
representation of f in the reconstruction formula need not be unique. This reflects
one of the important properties of frames in applied mathematics.
Let a, b ∈ R and c ∈ R\{0}. We consider operators Ta, Eb, Dc : L2(R) → L2(R)
given by
Translation by a ↔ Taf (t) = f (t − a),
Modulation by b ↔ Ebf (t) = e2πibtf (t),
Dilation by c ↔ Dcf (t) = c
2 f (ct).
1
A bounded linear operator T defined on H is said to be positive, if hT f, fi ≥ 0 for all
f ∈ H. In symbol we write T ≥ 0. If T1, T2 are bounded linear operator on H such
4
VASHISHT
that T1 − T2 ≥ 0, then we write T1 ≥ T2. A bounded linear operator T : H → H
is said to be hyponormal, if T ∗T − T T ∗ ≥ 0, or equivalently if kT ∗fk ≤ kT fk for
all f ∈ H. The characteristic function of any set E is denoted by χE. By R(T )
we denote the range of a bounded linear operator T from a normed space X into a
normed space Y .
Theorem 2.1. [10] Let H, H1, H2 be Hilbert spaces. Assume that T1 : H1 → H and
T2 : H2 → H be bounded linear operators. The following statement are equivalent:
(i) R(T1) ⊂ R(T2).
1 ≤ λ2T2T ∗
(ii) T1T ∗
(iii) There exists a bounded linear operator S : H1 → H2 such that T1 = T2S.
2 for some λ ≥ 0.
3. Wave Packet Frames in L2(R)
Definition 3.1. Let ψ ∈ L2(R), {aj}j∈Z ⊂ R+, {cm}m∈Z ⊂ R and b 6= 0 and let Θ
be a bounded linear operator on L2(R). A system {Daj TbkEcm ψ}j,k,m∈Z is called
a Θ-irregular Weyl-Heisenberg wave packet frame (in short, Θ-IW H wave packet
frame) for L2(R), if there exist constants 0 < α0 ≤ β0 < ∞ such that
hf, Daj TbkEcm ψi2 ≤ β0kΘfk2 for all f ∈ L2(R).
(3.1)
α0kΘ∗fk2 ≤ Xj,k,m∈Z
The scalars α0 and β0 are called lower and upper bounds of the Θ-IW H wave
packet frame {Daj TbkEcm ψ}j,k,m∈Z, respectively.
If upper inequality in (3.1) is
satisfied, then {Daj TbkEcm ψ}j,k,m∈Z is called a Bessel sequence in L2(R) with
Bessel bound β0. If Θ is the identity operator on L2(R), then Θ-IW H wave packet
frame for L2(R) is the standard IW H wave packet frame for L2(R).
If a countable sequence {fk} in a Hilbert space H satisfies the inequality (3.1),
i.e., if
∞
α0kΘ∗fk2 ≤
hf, fki2 ≤ β0kΘfk2 for all f ∈ H,
Xk=1
then we say that {fk} is a Θ-Hilbert frame for H.
3.1. Examples and comments: Every Θ-Hilbert frame for H is a K-frame for
H, but not conversely. More precisely, if {fk} is a Θ-Hilbert frame for H with
frame bounds α0 and β0. Then, {fk} is a K-frame for H with frame bounds α0
and β0 kΘk2. The following example shows that a K-frame for H need not be a
Θ-Hilbert frame for H.
Example 3.2. Let {χk} be the canonical orthonormal basis for the discrete signal
space H = L2(Ω, µ) (where Ω = N and µ is the counting measure) and let Θ be the
backward shift operator on H given by
Θ({ξ1, ξ2, ξ3, .......}) = {ξ2, ξ3, .......}, {ξj} ∈ H.
Then, its conjugate Θ∗ is the forward shift operator on H which is given by
Θ∗({ξ1, ξ2, ξ3, .......}) = {0, ξ1, ξ2, ξ3, .......}, {ξj} ∈ H.
Choose fk = χk for all k ∈ N.
We compute
kΘ∗fk2 = kfk2 =
∞
Xj=1
hf, fki2 for all f = {ξj} ∈ H.
Xk=1
WAVE PACKET FRAMES
5
Hence {fk} is a K-frame (with a choice K = Θ) for H with frame bounds A =
B = 1. But {fk} is not a Θ-Hilbert frame for H. Indeed, let ao and bo be positive
numbers such that
∞
aokΘ∗fk2 ≤
hf, fki2 ≤ bokΘfk2 for all f ∈ H.
(3.2)
Then, for fo = χ1 ∈ H, we obtain Θfo = 0. Therefore, by using upper inequality
in (3.2), we have fo = 0, a contradiction.
Remark 3.3. A Θ-Hilbert frame for H (Θ 6= I, the identity operator on H) need
not be a standard Hilbert frame for H and vice-versa. Indeed, let H be the discrete
signal space given in Example 3.2 with canonical orthonormal basis {χk}.
Choose fk = χk + χk+1, k ∈ N.
Define Θ : H → H by
Θ(f = {ξ1, ξ2, ξ3, .......}) = {ξ1, ξ1 + ξ2, ξ2 + ξ3, .......}, f = {ξj} ∈ H.
Then, Θ is a bounded linear operator on H and its conjugate operator Θ∗ is given
by
Θ∗({ξ1, ξ2, ξ3, .......}) = {ξ1 + ξ2, ξ2 + ξ3, .......}, {ξj} ∈ H.
One can verify that there exists a γ ∈ (0, 1) such that
γ kΘ∗fk2 ≤
hf, fki2 ≤ kΘfk2 for all f ∈ H.
∞
Xj=1
Hence F ≡ {fk} is a Θ-Hilbert frame for H. But F is not a standard Hilbert frame
for H (see Example 5.4.6 in [4], p. 98).
To show that a standard Hilbert frame for H need not be Θ-Hilbert frame for H.
Choose gk = χk, k ∈ N and let Θ be the backward shift operator on H. Then,
G = {gk} is a Hilbert frame for H, but not a Θ-Hilbert frame for H.
Regarding the existence of Θ-IW H wave packet frames for L2(R), we have fol-
lowing examples.
Example 3.4. Let a > 1 and b > 0 and cm = 0 for all m ∈ Z. Choose aj = aj for
all j ∈ Z. Then, there exist ψ ∈ L2(R) such that ψ = χE, where E is a compact
subset of R. Therefore,
{Daj TbkEcm ψ}j,k,m∈Z = {Daj Tbkψ}j,k∈Z
is an orthonormal basis for L2(R) (see Theorem 12.3 in [16] p. 357), hence a tight
IW H wave packet frame for L2(R) .
Let β ∈ R be arbitrary, but fixed.
Choose Θ = Eβ (the modulation operator on L2(R)) and dm = cm + β (m ∈ Z).
We compute
kΘ∗fk2 = αokE∗
hE∗
βf, Daj TbkEcm ψi2
βfk2 = Xj,k,m∈Z
= Xj,k,m∈Z
hf, Daj TbkEdm ψi2
= kΘfk2, for all f ∈ L2(R).
Hence {Daj TbkEdm ψ}j,k,m∈Z is a Θ-IW H wave packet frame for L2(R).
6
VASHISHT
Example 3.5. Let Θ : L2(R) → L2(R) be the multiplication operator given by
Θ(f ) = f.χ[0,1], f ∈ L2(R).
Then, Θ is a bounded linear operator on L2(R).
Choose b = 1, aj = 1, cm = 0 (j, m ∈ Z) and ψ = χ[0,1].
Then
{Daj TbkEcm ψ}j,k,m∈Z = {Tkψ}k∈Z = {χ[k,k+1]}k∈Z.
The system {Daj TbkEcm ψ}j,k,m∈Z is not a Θ-IW H wave packet frame for L2(R).
Indeed, let B be an upper Θ-IW H wave packet frame bound for {Daj TbkEcm ψ}j,k,m∈Z.
Let h ∈ L2(R) be a function given by
χ[0,1],
x ∈ [0, 1]
√B χ[2,3], x ∈ [2, 3]
otherwise.
h(x) =
0
We compute
Xj,k,m∈Zhh, Daj TbkEcm ψi2 =Xk∈Zhh, χ[k,k+1]i2
= hh, χ[0,1]i2 + hh, χ[2,3]i2
= 1 + B.
On the other hand, kΘhk2 = kh.χ[0,1]k2 = 1.
Therefore, Pj,k,m∈Z hh, Daj TbkEcm ψi2 = 1 + B > BkΘhk2. Hence B is not an
upper Θ-IW H wave packet frame bound for {Daj TbkEcm ψ}j,k,m∈Z, a contradiction.
3.2. Operators associated with Θ-IW H wave packet frames. Suppose that
F ≡ {Daj TbkEcm ψ}j,k,m∈Z is a Θ-IW H wave packet frame for L2(R). The operator
T : ℓ2(Z3) → L2(R) given by
T{cjkm}j,k,m∈Z = Xj,k,m∈Z
cjkmDaj TbkEcm ψ,
is called the pre-frame operator or synthesis operator associated with F and the
adjoint operator T ∗ : L2(R) → ℓ2(Z3) is given by
T ∗f = {hf, Daj TbkEcm ψi}j,k,m∈Z
is called the analysis operator associated with F . Composing T and T ∗, we obtain
the frame operator S : L2(R) → L2(R) given by
Sf = T T ∗f = Xj,k,m∈Zhf, Daj TbkEcm ψiDaj TbkEcm ψ.
(3.3)
Since F is a Θ-IW H wave packet Bessel sequence in L2(R), the series defining S
converges unconditionally for all f ∈ L2(R). Notice that, in general, frame operator
of the Θ- IW H wave packet frame F is not invertible on L2(R), but it is invertible
on a subspace R(Θ) ⊂ L2(R).
In fact, if R(Θ) is closed , then there exist a
pseudoinverse Θ† of Θ such that ΘΘ†f = f for all f ∈ R(Θ), i.e., ΘΘ†R(Θ) =
IR(Θ), so we have (cid:0)Θ†R(Θ)(cid:1)∗
R(Θ). Hence for any f ∈ R(Θ), we obtain
Θ∗ = I ∗
kfk =(cid:13)(cid:13)(cid:13)(cid:0)Θ†R(Θ)(cid:1)∗
Θ∗f(cid:13)(cid:13)(cid:13) ≤ kΘ†kkΘ∗fk.
WAVE PACKET FRAMES
7
Therefore, by using (3.3), we can write
hSf, fi ≥ AkΘ∗fk2 ≥ AkΘ†k−2kfk2 for all f ∈ R(Θ).
That is
AkΘ†k−2kfk2 ≤ kSfk2 ≤ Bkfk2 for all f ∈ R(Θ).
Thus, the operator S : R(Θ) → S(R(Θ)) is a homeomorphism. Furthermore, we
have
B−1kfk ≤ kS −1fk ≤ A−1kΘ†k2kfk for all f ∈ S(R(Θ)).
Next, we characterizes a system {Daj TbkEcm ψ}j,k,m∈Z ⊂ L2(R) as Θ-IW H wave
packet frame. Let T1 : H → H and T2 : H1 → H be bounded linear operators, where
H, H1 are Hilbert spaces. We say that the pair (T1, T2) is relatively hyponormal, if
λT ∗
1 T1 ≥ T2 T ∗
2 for some λ > 0.
In this case we say that T1 and T2 are relatively hyponormal. Aldroubi in [1]
characterized operators on a Hilbert space H, which can generate Hilbert frames
(as images of given frames) for H. Actually, Aldroubi considered operators which
are relative hyponormal with the identity operator on H. The following theorem
characterizes a certain system as a Θ-IW H wave packet frame for L2(R) in terms
of the relative hyponormality of operators.
Theorem 3.6. Let ψ ∈ L2(R), {aj}j∈Z ⊂ R+, {cm}m∈Z ⊂ R and b 6= 0 and let Θ
be a bounded linear operator on L2(R). Then, {Daj TbkEcm ψ}j,k,m∈Z is a Θ-IW H
wave packet frame for L2(R) if and only if there exist a bounded linear operator
Ξ : ℓ2(Z3) → L2(R) such that
(i) the pair (Θ, Ξ) is relative hyponormal, i.e., λΘ∗Θ ≥ Ξ Ξ∗ for some λ > 0,
(ii) Ξ(ej,k,m) = Daj TbkEcm ψ (j, k, m ∈ Z) and R(Θ) ⊂ R(Ξ),
where {ej,k,m}j,k,m∈Z is an orthonormal basis for ℓ2(Z3).
Proof. Suppose first that {Daj TbkEcm ψ}j,k,m∈Z is a Θ-IW H wave packet frame for
L2(R). Then, we can find positive constants a0, b0 such that
a0kΘ∗fk2 ≤ Xj,k,m∈Z
hf, Daj TbkEcm ψi2 ≤ b0kΘfk2 for all f ∈ L2(R).
(3.4)
Define W : L2(R) → ℓ2(Z3) by
W(f ) = Xj,k,m∈Z
hf, Daj TbkEcm ψiej,k,m.
Clearly, W is a well defined bounded linear operator on L2(R).
8
VASHISHT
We compute
hW ∗ej,k,m, hi = hej,k,m,Whi
=*ej,k,m, Xj,k,m∈Zhh, Daj TbkEcm ψiej,k,m+
= Xj,k,m∈Z
= hh, Daj TbkEcm ψi
= hDaj TbkEcm ψ, hi for all h ∈ L2(R).
hh, Daj TbkEcm ψihej,k,m, ej,k,mi
This gives
W ∗ej,k,m = Daj TbkEcm ψ (j, k, m ∈ Z).
By using (3.5) and lower frame inequality in (3.4), we obtain
(3.5)
a0kΘ∗fk2 ≤ Xj,k,m∈Zhf,W ∗ej,k,mi2 = kWfk2 for all f ∈ L2(R).
This gives a0ΘΘ∗ ≤ W ∗W.
Choose Ξ = W ∗. Then, by Theorem 2.1, we have R(Θ) ⊂ R(Ξ). The condition (ii)
in the result is proved.
To show λΘ∗Θ ≥ Ξ Ξ∗ (λ > 0), we consider upper frame inequality in (3.4):
b0kΘfk2 ≥ Xj,k,m∈Z
hf, Daj TbkEcm ψi2
= Xj,k,m∈Zhf,W ∗ej,k,mi2
= kWfk2 for all f ∈ L2(R).
the condition (i) in the result.
This gives b0Θ∗Θ ≥ W ∗W. That is, λΘ∗Θ ≥ Ξ Ξ∗ (λ = b0 > 0). This proves
Conversely, assume that both conditions (i) and (ii) given in the theorem hold.
We compute
hΞ∗f, hi =*Ξ∗f, Xj,k,m∈Z
aj,k,mej,k,m+
aj,k,mhf, Ξej,k,mi
= Xj,k,m∈Z
= Xj,k,m∈Z
= Xj,k,m∈Z
= Xj,k,m∈Z
=* Xj,k,m∈Z
aj,k,mhf, Daj TbkEcm ψi
hh, ej,k,mihf, Daj TbkEcm ψi
hej,k,m, hihf, Daj TbkEcm ψi
hf, Daj TbkEcm ψiej,k,m, h+ ,
WAVE PACKET FRAMES
9
for all f ∈ L2(R) and for all h ∈ ℓ2(Z3).
This gives
Ξ∗f = Xj,k,m∈Z
hf, Daj TbkEcm ψiej,k,m for all f ∈ L2(R).
(3.6)
Therefore, by using (3.6) and the condition (i), we have
hf, Daj TbkEcm ψi2 = kΞ∗fk2 ≤ λkΘfk2 for all f ∈ L2(R) (λ > 0).
Xj,k,m∈Z
By hypothesis R(Θ) ⊂ R(Ξ) (see condition (ii)). So, by Theorem 2.1, we can
find a positive constant β such that ΘΘ∗ ≤ β Ξ Ξ∗ (note that β is positive, since
otherwise Θ = O). Again by using the condition (ii), we have
(3.7)
1
βkΘ∗fk2 ≤ kΞ∗fk2 = Xj,k,m∈Z
= Xj,k,m∈Z
= Xj,k,m∈Z
hΞ∗f, ej,k,mi2
hf, Ξej,k,mi2
hf, Daj TbkEcm ψi2 for all f ∈ L2(R).
(3.8)
By using (3.7) and (3.8), we conclude that {Daj TbkEcm ψ}j,k,m∈Z is a Θ-IW H wave
packet frame for L2(R).
(cid:3)
Djordjevi´c in [9] characterized hyponormal operators by using the Moore-Penrose
inverse of a bounded linear operator with a closed range. There may be other
conditions for a bounded linear operator on a Hilbert space to be hyponormal.
Let H and K be Hilbert spaces and A : H → K be a bounded linear operator.
The Moore-Penrose inverse of A is denoted by A†, see [2]. Djordjevi´c proved the
following result by using the Moore-Penrose inverse of a bounded linear operator
with a closed range.
Theorem 3.7. [9] Let A and AA∗ + A∗A have closed ranges. Then the following
statements are equivalent:
(i) A is hyponormal
(ii) 2AA∗(AA∗ + A∗A)†AA∗ ≤ AA∗.
Thus, a bounded linear operator A defined on a Hilbert space is hyponormal if a
certain operator inequality (consisting of adjoint and Moore-Penrose inverse of A)
is satisfied. Frame can be used to characterizes a hyponormal operator on L2(R).
First we define a type of tight frame (or Parseval frame) in L2(R). In Definition 3.1,
if α0 = β0, then {Daj TbkEcm ψ}j,k,m∈Z is not a standard tight frame, in general.
This is the motivation for new type of tight frames in L2(R).
Definition 3.8. Let Θ 6= I (where I the identity operator on L2(R)). A Θ-Hilbert
frame {fn} ⊂ H for H with frame bounds α0 = β0 is called a (Θ, α0)-Hilbert tight
frame.
The following theorem characterizes hyponormal operators on L2(R) in terms of
(Θ, α0)- Hilbert tight frames for L2(R).
Theorem 3.9. A bounded linear operator Θ on L2(R) is hyponormal if and only
if there exists a (Θ, 1)-Hilbert tight frame for L2(R).
10
VASHISHT
Proof. Assume first that Θ is a hyponormal operator on L2(R). Let {Daj TbkEcm ψ}j,k,m∈Z
be a tight IW H wave packet frame for L2(R).
Then
Xj,k,m∈Z
hf, Daj TbkEcm ψi2 = kfk2 for all f ∈ L2(R).
(3.9)
Choose fn(n ∈ N) ↔ ϕj,k,m = Θ(Daj TbkEcm ψ), j, k, m ∈ Z.
Then, by using (3.9) and hyponormality of Θ, we compute
Xj,k,m∈Z
hf, Θ(Daj TbkEcm ψ)i2
hf, ϕj,k,mi2 = Xj,k,m∈Z
= Xj,k,m∈Z
= kΘ∗fk2
≤ kΘfk2 for all f ∈ L2(R).
hΘ∗f, Daj TbkEcm ψi2
(3.10)
For the lower frame inequality, we compute
kΘ∗fk2 = Xj,k,m∈ZhΘ∗f, Daj TbkEcm ψi2
hf, Θ(Daj TbkEcm ψ)i2
= Xj,k,m∈Z
= Xj,k,m∈Z
hf, ϕj,k,mi2 for all f ∈ L2(R).
(3.11)
By using (3.10) and (3.11) we have
kΘ∗fk2 ≤ Xj,k,m∈Z
hf, ϕj,k,mi2 ≤ kΘfk2 for all f ∈ L2(R).
Hence {ϕj,k,m}j,k,m∈Z is a (Θ, 1)-Hilbert tight frame for L2(R).
Then
For the reverse part, suppose that {fn} is a (Θ, 1)-Hilbert tight frame for L2(R).
kΘ∗fk2 ≤
hf, fni2 ≤ kΘfk2 for all f ∈ L2(R).
∞
Xn=1
This gives kΘ∗fk ≤ kΘfk for all f ∈ H. Hence Θ is a hyponormal operator on
L2(R).
(cid:3)
Favier and Zalik proved in [13] that the image of a Hilbert frame for H under a
linear homeomorphism is a Hilbert frame for H. They established relation between
optimal bounds of a given Hilbert frame and its image (as frame). This is not
true for Θ-IW H wave packet frame (see Example 3.12), in general. The problem
(regarding invariance behaviour as a frame under linear homeomorphism) for Θ-
IW H wave packet frames can be solved, provided the given linear homeomorphism
commutes with Θ∗. This is proved in the following theorem.
Theorem 3.10. Let F ≡ {Daj TbkEcm ψ}j,k,m∈Z be a Θ-IW H wave packet frame
for L2(R) and U be a linear homeomorphism on L2(R) such that U commutes with
Θ∗. Then, FU ≡ {U (Daj TbkEcm ψ)}j,k,m∈Z is a Θ-IW H wave packet frame for
L2(R). Furthermore, if A1 and B1 are optimal bounds of the frame F and the pair
WAVE PACKET FRAMES
11
(Θ, U ∗) is relatively hyponormal, then the optimal bounds A2 and B2 of the frame
FU satisfy the inequalities
A1kUk−2 ≤ A2 ≤ A1kU −1k2 ; γB1kΘk−2 ≤ B2 ≤ B1kUk2 (γ > 0).
(3.12)
Proof. We compute
Xj,k,m∈Z
hf, U (Daj TbkEcm ψ)i2 = Xj,k,m∈Z
hU ∗f, Daj TbkEcm ψi2
≤ B1kΘU ∗fk2
= B1kU ∗Θfk2
≤ B1kU ∗k2kΘfk2 for all f ∈ L2(R).
(3.13)
By using the fact that A1 is one of the choice for lower Θ-IW H wave packet frame
bound for {Daj TbkEcm ψ}j,k,m∈Z and U commutes with Θ∗, we compute
kΘ∗fk2 = kΘ∗(U U −1)fk2
= kU Θ∗(U −1f )k2
≤ kUk2kΘ∗(U −1f )k2
≤ kUk2
= kUk2
A1 Xj,k,m∈Z
A1 Xj,k,m∈Z
A1 Xj,k,m∈Z
= kUk2
hU −1f, Daj TbkEcm ψi2
hU U −1f, U (Daj TbkEcm ψ)i2
hf, U (Daj TbkEcm ψ)i2.
(3.14)
By using (3.13) and (3.14), we obtain
A1kUk−2kΘ∗fk2 ≤ Xj,k,m∈Zhf, U (Daj TbkEcm ψ)i2 ≤ B1kU ∗k2kΘfk2 for all f ∈ L2(R).
Hence {U (Daj TbkEcm ψ)}j,k,m∈Z is a Θ-IW H wave packet frame for L2(R) with
one of the choice of frame bounds A1kUk−2, B1kUk2.
Since A2 and B2 are best frame bounds for {U (Daj TbkEcm ψ)}j,k,m∈Z, we have
(3.15)
Again {U (Daj TbkEcm ψ)}j,k,m∈Z is a Θ-IW H wave packet frame for L2(R) with
A2, B2 as one of the choice of frame bounds. So, for all f ∈ L2(R), we have
A1kUk−2 ≤ A2, B2 ≤ B1kUk2.
A2kΘ∗fk2 ≤ Xj,k,m∈Zhf, U (Daj TbkEcm ψ)i2 ≤ B2kΘfk2.
kΘ∗fk2 = kU −1U Θ∗fk2 = kU −1Θ∗U fk2 ≤ kU −1k2kΘ∗U fk2.
For all f ∈ L2(R), we have
(3.16)
(3.17)
By using (3.16), (3.17) and relative hyponormality of the pair (Θ, U ∗) , we have
A2kU −1k−2kΘ∗fk2 ≤ A2kΘ∗U fk2
≤ Xj,k,m∈Z
hU f, U (Daj TbkEcm ψ)i2
= Xj,k,m∈Z
hf, Daj TbkEcm ψi2
12
VASHISHT
≤ B2kΘU fk2
≤ B2kΘk2kU fk2
≤ λB2kΘk2kΘfk2 for all f ∈ L2(R),
(3.18)
where λ is a positive constant which appears in the relative hyponormality of the
pair (Θ, U ∗).
Since A1 and B1 are the best Θ-IW H wave packet frame bounds for {Daj TbkEcm ψ}j,k,m∈Z,
by using (3.18), we have
A2kU −1k−2 ≤ A1, B1 ≤ λB2kΘk2.
(3.19)
The inequalities in (3.12) are obtained from (3.15) and (3.19). The result is proved.
(cid:3)
Remark 3.11. The condition that the linear homeomorphism U commutes with
Θ∗ in Theorem 3.10 cannot be relaxed. This is justified in the following example.
Example 3.12. Consider the multiplication operator Θ : L2(R) −→ L2(R) given
by
Θ(f ) = f.χ[0,1], f ∈ L2(R).
Then, Θ is a bounded linear self-adjoint operator on L2(R).
Choose aj = 1, cm = m for all j, m ∈ Z, b = 0 and ψ = χ[0,1). Then, {Daj TbkEcm ψ}j,k,m∈Z
is a Θ-IW H wave packet frame for L2(R). Indeed, for all f ∈ L2(R), we have
Xj,k,m∈Z
hf, Daj TbkEcm ψi2 = Xm∈Zhf, Emχ[0,1)i2
= Xm∈Zhf, Θ(Emχ[0,1))i2
= Xm∈Z
hΘ∗f, Emχ[0,1)i2
= kΘ∗fk2
= kΘfk2
Hence {Daj TbkEcm ψ}j,k,m∈Z is a Θ-IW H wave packet frame for L2(R).
Choose U◦ = T1, the translation operator on L2(R), i.e., U◦f (•) = f (• − 1).
Then, U◦ is a linear homeomorphism on L2(R). First we show that the operator
U◦ and Θ∗(= Θ) does not commutes. For this, we compute
and
Θ∗U◦(f )(γ) = U◦(f )(γ).χ[0,1)(γ)
= f (γ − 1).χ[0,1)(γ),
U◦Θ∗(f )(γ) = U◦(f.χ[0,1))(γ)
= f (γ − 1).χ[0,1)(γ − 1)
= f (γ − 1).χ[1,2)(γ).
(3.20)
(3.21)
By using (3.20) and (3.21), we conclude that the operators U◦ and Θ∗ does not
commutes.
WAVE PACKET FRAMES
13
Next, we show that the system FU◦ ≡ {U◦(Daj TbkEcm ψ)}j,k,m∈Z is not a
Θ-IW H wave packet frame for L2(R). Let a◦ and b◦ be a choice of frame bounds
for FU◦.
Then
∞
aokΘ∗fk2 ≤
hf, U◦(Daj TbkEcm ψ)i2 ≤ bokΘfk2 for all f ∈ H.
(3.22)
Xk=1
Choose f◦ = χ[0,1[ ∈ L2(R). Then, kΘ∗f◦k = 1.
Then, by using lower inequality in (3.22), we compute
a◦ = a◦kΘ∗f◦k2 ≤ Xj,k,m∈Z
◦ f◦, Daj TbkEcm ψi2
hU ∗
hf◦, U◦(Daj TbkEcm ψ)i2 = Xj,k,m∈Z
= Xm∈Z
◦ f◦, Emψi2
hU ∗
◦ f◦)k2
= kΘ(U ∗
= kΘ(χ[−1,0))k2
= kχ[−1,0).χ[0,1)k2
= 0,
a contradiction. Hence FU◦ is not a Θ-IW H wave packet frame for L2(R).
4. Linear Combinations of Θ-IW H Wave Packet Frames
Linear combination of frames (or redundant building blocks) is important in
applied mathematics. Aldroubi in [1] considered the following problem: given a
Hilbert frame {fk} for H, define a set of functions Φj by taking linear combinations
of the frame elements fk. What are the conditions on the coefficients in the linear
combinations, so that the new system {Φj} constitutes a frame for H ? More
precisely, Aldroubi considered a linear combination of the form
∞
Φj =
αj,kfk, (j ∈ N)
Xk=1
where αj,k are scalars. Aldroubi proved sufficient conditions on {αj,k} such that
{Φj} constitutes a frame for H. Christensen in [5] gave sufficient conditions which
are different from those proved by Aldroubi. In this section, we extend some results
by Kaushik et al. in [19] to Θ-IW H wave packet frames for L2(R).
Let {Daj TbkEcm ψ}j,k,m∈Z be a Θ-IW H wave packet frame for L2(R). First we
consider a linear combination of the form:
Φr,s,t = X(j,k,m)∈Ir,s,t
Ir,s,t = Z × Z × Z,
αj,k,mDaj TbkEcm ψ, (r, s, t ∈ Z),
Ir,s,tT Ir′,s′,t′ = ∅, (r, s, t) 6= (r′, s′, t′) for all
where Sr,s,t∈Z
r, s, t, r′, s′, t′ ∈ Z and αj,k,m are scalars. The system {Φr,s,t}r,s,t∈Z is not a Θ-
IW H wave packet frame for L2(R), in general. This type of combinations under
the WH-packet for Gabor system were studied by Kaushik et al. [19]. The following
theorem gives necessary and sufficient conditions for the system {Φr,s,t}r,s,t∈Z to
be a Θ-IW H wave packet frame for L2(R). This is an adaption of [19, Theorem
3.5].
(4.1)
14
VASHISHT
Theorem 4.1. Let Θ be a bounded linear operator on L2(R) such that Θ∗ is hy-
ponormal. Assume that {Daj TbkEcm ψ}j,k,m∈Z is a Θ-IW H wave packet frame
for L2(R) and {Φr,s,t}r,s,t∈Z ⊂ L2(R) be the sequence defined in (4.1). Let
T : ℓ2(Z3) → ℓ2(Z3) be a bounded linear operator such that
T ({hDaj TbkEcm ψ, fi}j,k,m∈Z) = {hΦr,s,t, fi}r,s,t∈Z, f ∈ L2(R).
Then, {Φr,s,t}r,s,t∈Z is a Θ-IW H wave packet frame for L2(R) if and only if there
exists a constant λ > 0 such that
Xr,s,t∈ZhΦr,s,t, fi2 ≥ λ Xj,k,m∈Z
hDaj TbkEcm ψ, fi2 for all f ∈ L2(R).
(4.2)
Proof. Assume first that {Φr,s,t}r,s,t∈Z is a Θ-IW H wave packet frame for L2(R)
with frame bounds A′, B′. Then, for any f ∈ L2(R), we have
hΦr,s,t, fi2 ≥ A′kΘ∗fk2.
(4.3)
Xr,s,t∈Z
If B is an upper Θ-IW H wave packet frame bound for {Daj TbkEcm ψ}j,k,m∈Z, then
Xj,k,m∈Z
hf, Daj TbkEcm ψi2 ≤ BkΘfk2, f ∈ L2(R).
i.e.
1
B Xj,k,m∈Z
hf, Daj TbkEcm ψi2 ≤ kΘfk2, f ∈ L2(R).
(4.4)
Choose λ = A′
B > 0. Then, by using hyponormality of Θ∗, (4.3) and (4.4), we have
Xr,s,t∈ZhΦr,s,t, fi2 ≥ A′kΘ∗fk2
≥ A′kΘfk2
≥ λ Xj,k,m∈Z
hf, Daj TbkEcm ψi2 for all f ∈ L2(R).
The inequality given in (4.2) is proved.
For the reverse part, since {Daj TbkEcm ψ}j,k,m∈Z is a Θ-IW H wave packet frame
for L2(R). There exist positive constants A, B such that
AkΘ∗fk2 ≤ Xj,k,m∈Z
hf, Daj TbkEcm ψi2 ≤ BkΘfk2 for all f ∈ L2(R).
(4.5)
By using (4.2) and (4.5), we have
Xr,s,t∈Z
hΦr,s,t, fi2 ≥ λ Xj,k,m∈Z
hDaj TbkEcm ψ, fi2
≥ λAkΘ∗fk2 for all f ∈ L2(R).
(4.6)
We compute
Xr,s,t∈ZhΦr,s,t, fi2 = k{hΦr,s,t, fi}r,s,t∈Zk2
ℓ2(Z3)
= kT ({hDaj TbkEcm ψ, fi}j,k,m∈Z)k2
ℓ2(Z3)
WAVE PACKET FRAMES
15
≤ kTk2 Xj,k,m∈Z
hDaj TbkEcm ψ, fi2
≤ kTk2BkΘfk2 for all f ∈ L2(R).
(4.7)
By using (4.6) and (4.7), we conclude that {Φr,s,t}r,s,t∈Z is a Θ-IW H wave packet
frame for L2(R).
(cid:3)
4.1. The case of finite sum: We now consider a linear combination of the form
αsDaj TbkEcm ψs(cid:27)j,k,m∈Z
Fp ≡ (cid:26) p
Ps=1
, where α1, α2, ..., αp are nonzero scalars,
ψs ∈ L2(R) and {Daj TbkEcm ψs}j,k,m∈Z is a Θ-IW H wave packet frame for L2(R)
for each s ∈ Λp = {1, 2, 3, .., p} . The finite sum Fp is not a Θ-IW H wave packet
frame for L2(R), in general. Kaushik, Singh, and Virender [19] showed that if some
scalar multiple of a series associated with a Gabor frame is dominated by the series
associated with the finite sum of Gabor frames, then the finite sum constitutes a
Gabor frame for the underlying space and vice-versa, see Theorem 4.2 of [19]. The
following theorem extend this result in the context of Θ-IW H wave packet frame
for L2(R).
Theorem 4.2. Assume that Θ : L2(R) → L2(R) is a bounded linear operator such
that Θ∗ is hyponormal. Let {Daj TbkEcm ψs} j,k,m∈Z
be a finite family of Θ-IW H
frames for L2(R). Then, Fp ≡ (cid:26) p
Ps=1
is a Θ-IW H wave
packet frame for L2(R) if and only if there exists µ > 0 and some ξ ∈ Λp such that
hDaj TbkEcm ψξ, fi2 ≤ Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
µ Xj,k,m∈Z
for any finite sequence of scalars {αs}.
Proof. Let Aξ, Bξ be frame bounds for Θ-IW H wave packet frame {Daj TbkEcm ψξ}j,k,m∈Z
for L2(R) (1 ≤ ξ ≤ p).
Then
αsDaj TbkEcm ψs(cid:27)j,k,m∈Z
αsDaj TbkEcm ψs, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
, f ∈ L2(R)
* p
Xs=1
s∈Λp
2
µAξkΘ∗fk2 ≤ µ Xj,k,m∈Z
hDaj TbkEcm ψξ, fi2
* p
≤ Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xs=1
αsDaj TbkEcm ψs, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Thus, the lower frame condition is satisfied for the finite system Fp.
2
, f ∈ L2(R).
(4.8)
For the upper frame condition, we compute
Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
* p
Xs=1
αsDaj TbkEcm ψs, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
2
2
p
= Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
αshDaj TbkEcm ψs, fi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xs=1
Xs=1(cid:12)(cid:12)αshDaj TbkEcm ψs, fi(cid:12)(cid:12)#2
≤ Xj,k,m∈Z" p
If Bξ is an upper frame bound for {Daj TbkEcm ψξ}j,k,m∈Z, then
AkΘ∗fk2 ≤ Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Bξ Xj,k,m∈Z
1
* p
Xs=1
2
αsDaj TbkEcm ψs, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
.
(4.10)
hf, Daj TbkEcm ψξi2 ≤ kΘfk2, f ∈ L2(R).
(4.11)
16
VASHISHT
p
αs2 Xj,k,m∈Z
Xs=1
≤
≤ p max
Xs=1
1≤s≤pαs2
p
hDaj TbkEcm ψs, fi2
Bs!kΘfk2, f ∈ L2(R).
(4.9)
By (4.8) and (4.9), we conclude that the finite sum Fp is a Θ-IW H wave packet
frame for L2(R).
Conversely, assume that the finite sum Fp is a Θ-IW H wave packet frame for L2(R)
with frame bounds A, B. Then, for all f ∈ L2(R), we have
Choose µ = A
Bξ
> 0. Then, using hyponormality of Θ∗, (4.10) and (4.11) we have
µ Xj,k,m∈Z
hDaj TbkEcm ψξ, fi2 ≤ AkΘfk2
≤ AkΘ∗fk2
* p
≤ Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xs=1
The theorem is proved.
2
αsDaj TbkEcm ψs, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
, f ∈ L2(R).
(cid:3)
Application: The following example gives an application of Theorem 4.2.
Example 4.3. Let Θ : L2(R) → L2(R) be the modulation operator. That is,
Θf (t) = e2πibtf (t), where b ∈ R is fixed. Then, Θ∗ is hyponormal on L2(R).
Choose ψ = χ[0,1[, aj = 1, b = 1, cm = m for all j, m ∈ Z and ψs = ψ for all s ∈ Λp.
Then, for any nonzero scalars α1, α2, ..., αp with Pp
s=1 αs 6= 0, we have
* p
Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
= Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xs=1
Xs=1
αs2 Xj,k,m∈Z
αsDaj TbkEcm ψs, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
αsDaj TbkEcm ψ, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
* p
Xs=1
=
2
2
p
hDaj TbkEcm ψξ, fi2, f ∈ L2(R),
where ψξ = χ[0,1).
Then
s=1 αs2 > 0.
Choose µ = Pp
µ Xj,k,m∈Z
By Theorem 4.2, the finite system Fp is a Θ-IW H wave packet frame for L2(R).
hDaj TbkEcm ψξ, fi2 = Xj,k,m∈Z(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
αsDaj TbkEcm ψs, f+(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
* p
Xs=1
2
for all f ∈ L2(R).
WAVE PACKET FRAMES
17
This work is jointly with A. K. Sah and Deepshikha
Lalit is supported by R & D Doctoral Research Programme, University of Delhi,
Delhi-110007, India. Grant No. : RC/2014/6820.
Acknowledgement
References
1. A. Aldroubi, Portraits of frames, Proc. Amer. Math. Soc., 123 (1995), no. 6, 1661 -- 1668.
2. A. Ben-Israel, T.N.E. Greville, Generalized Inverses: Theory and Applications, second ed.,
Springer (2003).
3. P. G. Casazza and G. Kutyniok, Finite frames: Theory and Applications, Birkhauser (2012).
4. O. Christensen, Introduction to frames and Riesz bases, Birkhauser (2003).
5. O. Christensen, Linear combinations of frames and frame packets, Zeit. Anal. Anwen., 20
(2001), no. 4, 805 -- 815.
6. A. Cordoba and C. Fefferman, Wave packets and Fourier integral operators, Comm. Partial
Differential Equations, 3 (11) (1978), 979 -- 1005
7. W. Czaja, G. Kutyniok and D. Speegle, The Geometry of sets of prameters of wave packets,
Appl. Comput. Harmon. Anal., 20 (2006), 108 -- 125.
8. I. Daubechies, A. Grossmann and Y. Meyer, Painless non-orthogonal expansions, J. Math.
Phys., 27 (1986), 1271 -- 1283.
9. D. S. Djordjevi´c, Characterizations of normal, hyponormal and EP operators, J. Math. Anal.
Appl., 329 (2007), 1181 -- 1190.
10. R. G. Douglas, On majorization, factorization and range inclusion of operators on Hilbert
space, Proc. Amer. Math. Soc., 72 (1966), no. 2, 413-415.
11. R. J. Duffin and A. C. Schaeffer, A class of nonharmonic Fourier series, Trans. Amer. Math.
Soc., 72 (1952), 341 -- 366.
12. H.G. Feichtinger, T. Werther, Atomic systems for subspaces, in: L. Zayed (Ed.), Proceedings
SampTA 2001, Orlando, FL (2001), 163-165.
13. S. J. Favier and R. A. Zalik , On the stability of frames and Riesz bases, Appl. Comput.
Harmon. Anal., 2 (1995), 160 -- 173.
14. L. Gavruta, Frames for operators, Appl. Compu. Harmon. Appl., 32 (2012), 139 -- 144.
15. K. Guo and D. Labate, Some remarks on the unified characerization of reproducing systems,
Collect. Math., 57 (3) (2006), 309-318.
16. C. Heil, A Basis Theory Primer, Expanded edition. Applied and Numerical Harmonic Anal-
ysis. Birkhauser/Springer, New York (2011).
17. E. Hern´andez, D. Labate and G. Weiss, A unified characterization of reproducing systems
generated by a finite family II, J. Geom. Anal., 12 (4) (2002), 615-662.
18. E. Hern´andez, D. Labate, G. Weiss and E. Wilson, Oversampling, quasi-affine frames and
wave packets, Appl. Comput. Harmon. Anal., 16 (2004,) 111 -- 147.
19. S. K. Kaushik, G. Singh and Virender, On W H packets in L2(R), Commun. Math. Appl. 3
(2012), 333 -- 344.
20. D. Labate, G. Weiss and E. Wilson, An approach to the study of wave packet systems, Con-
temp. Math., 345 (2004), 215 -- 235.
21. M. Lacey, C. Thiele, Lp estimates on the bilinear Hilbert transform for 2 < p < ∞, Ann.
Math., 146 (1997), 693-724.
22. M. Lacey, C. Thiele, On Calder´on's conjecture, Ann. Math., 149 (1999), 475 -- 496.
L. K. Vashisht, Department of Mathematics, University of Delhi, Delhi-110007, India
E-mail address:
[email protected]
|
1606.06844 | 1 | 1606 | 2016-06-22T08:15:55 | Perturbations of Admissibility, Exact Controllability, Exact Observability and Regularity | [
"math.FA",
"math.OC"
] | This paper is concerned with the notions of admissibility, exact controllability, exact observability and regularity of linear systems in the Banach space setting. It is proved that admissible controllability, exact controllability, admissible observation, exact observability and regularity are invariant under some regular perturbations of the generators, such results are generalizations of some previous references. Moreover, the related boundary linear systems and some illustrative examples are presented. | math.FA | math |
Perturbations of Admissibility, Exact Controllability,
Exact Observability and Regularity ∗†
Zhan-Dong Mei ‡ Ji-Gen Peng §
Abstract
This paper is concerned with the notions of admissibility, exact controllability,
exact observability and regularity of linear systems in the Banach space setting. It
is proved that admissible controllability, exact controllability, admissible observa-
tion, exact observability and regularity are invariant under some regular perturba-
tions of the generators, such results are generalizations of some previous references.
Moreover, the related boundary linear systems and some illustrative examples are
presented.
Key words: Admissibility; Exact controllability; Exactly observability; Regular
linear systems; Boundary linear systems.
∗This work was supported by the National Natural Science Foundation of China (grant nos. 11301412
and 11131006), Research Fund for the Doctoral Program of Higher Education of China (grant no.
20130201120053), Natural Science Foundation of Shaanxi Province (grant no. 2014JQ1017), Project
funded by China Postdoctoral Science Foundation (grant nos. 2014M550482 and 2015T81011). Part of
this work was done during the first author was visiting Prof. Bao-Zhu Guo at Academy of Mathematics
and Systems Science, The Chinese Academy of Sciences.
†2010 Mathematics Subject Classification. 34K30; 35F15; 47D06; 92D25.
‡Corresponding author, School of Mathematics and Statistics, Xi'an Jiaotong University, Xi'an 710049,
China; Email: [email protected]
§School of Mathematics and Statistics, Xi'an Jiaotong University, Xi'an 710049, China; Email: jg-
[email protected]
1
Introduction
In the theory of finite dimensional linear control system, the final state and output are
continuously depended on the initial state and input. Observe that such continuous de-
pendence is the essential property in system theory. Motivated by this, Salamon [34]
introduce the class of well-posed liner systems by continuous dependence in Hilbert space
setting. Later, Weiss [39, 40, 41] simplified Salamon's theory; he described well-posed lin-
ear system equivalently by using four algebraic equations (see the description in Section
2). In the functional analysis frame, the control operators and observation operators of
well-posed linear system may be unbounded, which allow ones to study partial differential
equations with boundary control and boundary observation. Over the last decades there
has been a growing interest in well-posedness of partial differential equations with control
and observation on the boundary, and it has been proved that many partial differential
equations can be formulated as well-posed linear systems [1, 11, 22, 24, 25, 34, 35]. Reg-
ular linear systems, introduced by Weiss [41], are among the well-posed systems whose
output function corresponding to a step input function and zero initial state is not very
discontinuous at zero (see the definition in Section 2). Many well-posed physical systems
are also regular, see [2, 5, 6, 7, 9, 10, 15, 17, 45]. Regular linear systems have a convenient
representation, similar to that of finite dimensional systems. Concretely, Weiss showed
in [41] that regular linear systems with unbounded control and observation operators can
be simply represented by
x(t) = Ax(t) + Bu(t), y(t) = CΛx(t) + Du(t),
where CΛ is the Λ-extension of the observation operator C with respect to system operator
A (see Section 2).
In this sense, an infinite-dimensional regular linear systems have
the characteristic of "finite-dimensional systems". Many references were concerned with
abstract control theory under the frame of regular linear systems.
In order to obtain a well-posed and regular linear system, the control and observation
operators should be admissible for the system operator (see, e.g., [34, 35, 41, 42, 43]).
Hence the the concepts of admissibility of control and observation operators have been
discussed by many references, most of which are interested in proving or disproving Weiss'
2
conjecture (see, e.g., [12, 19, 47, 48]). Here, we mention an important work due to Zwart
[48]; he proved that the Weiss conjecture almost holds in Hilbert spaces.
For admissible control and admissible observation system, one can consider the no-
tion of exact controllability and exact observability, because the enters into the study
of many other important concepts. For instance, exact controllability is closely re-
lated to stabilizability and optimizability, while exact controllability is closely related
to detectability and estimatability [4, 24, 44]. Exact controllability and exact observ-
ability have received considerable attention in the functional analysis frame (see e.g.
[13, 18, 20, 23, 31, 32, 33, 46]), where some necessary and/or sufficient conditions have
been given.
Generally, it is not an easy task to verify the admissibility, exact controllability,
exact observability and regularity for a specific linear system with boundary control and
/or boundary observation. Due to the difficulties of direct proving the admissibility and
regularity, perturbation method has been successfully used to study the the admissibility
and regularity. Weiss [40] discussed the admissibility of observation system under bounded
perturbation of the system operator, namely, C being admissible observable operator for
A implies that C is admissible for A + P , provided P is an bounded linear operator on the
state space. In [43], Weiss showed that the closed-loop system of well-posed linear system
preserves the admissibility, exact controllability and exact observability. Moreover, the
closed-loop system of regular linear system preserve the regularity. Hadd [14] proved
that both B and ∆A are p-admissible controllable operators for A imply that B is p-
admissible for (A + ∆A)X; ((A + ∆A)X , B) is exactly controllable provided (A, B) is
exactly p-controllable and ∆A is "small" enough. In their paper [16], Hadd showed that
C and ∆A being p-admissible observable operators for A implies C is p-admissible for
A + ∆A. Moreover, Tucsnak and Weiss [38] proved that if (A, C) is exactly observable
and ∆A is "small" enough, then (A + ∆A, C) is exactly observable. Later, Mei and
Peng [28, 29] weakened the condition of [14, 16, 38] that ∆A is p-admissible controllable
(observable) operator to ∆A being q-admissible controllable (observable) operator. Mei
and Peng [26] proved that the admissibility, exact controllability and exact observation
are preserved under cross perturbations, that is, (A, B, ∆A) is a regular linear system,
3
then B is admissible for A + ∆A and (A + ∆A, B) is exactly controllable provided (A, B)
is exactly controllable and ∆A is "small" enough; (A, ∆A, C) is a regular linear system,
then C A
Λ is admissible for (A−1 + ∆A)X and ((A−1 + ∆A)X, C A
Λ ) is exactly observable
provided (A, C) is exactly observable and ∆A is "small" enough.
In their paper [27],
Mei and Peng proved that (A, B, ∆A) and (A, B, C) generating regular linear systems
imply that (A + ∆A, B, C A
Λ ) generates a regular linear system; (A, ∆A, C) and (A, B, C)
generating regular linear systems imply that ((A−1 + ∆A)X , B, C) generates a regular
linear system.
The aim of this paper is to study some general perturbation theorems of admissibili-
ties, exact controllabilities, exact observations and regularities. Apart from the introduc-
tion, our arrangement is as follows. In Section 2, we introduce some basic notions and
properties related to regular linear systems and boundary systems; the notions of exact
controllability and exact observation are also be introduced. Section 3 is to give our main
results. Concretely, we obtain admissible controllability, admissible observation, exact
controllability, exact observation and regularity under some regular perturbations. More-
over, all the perturbation results are used to solve the corresponding boundary systems.
The systems governed by specific partial differential equations are presented to illustrate
our results.
2 Preliminaries
In this section, we recall some definitions related to regular linear systems and boundary
linear systems. As stated in the introduction, Weiss has showed that the continuous
dependence of state and output on the initial state and input can be simplified to four
algebraic equations. We adopted Weiss' definition for well-posed linear system [41].
Definition 2.1 A quadruple Σ = (T, Φ, Ψ, F ) is said to be a well-posed linear system on
(X, U, Y ), if the following four conditions are satisfied:
(i) T = {T (t)}t≥0 is a C0-semigroup generated by A on X;
(ii) Φ = {Φ(t)}t≥0 is a family of bounded linear operators, called input maps, from
4
Lp(R+, U) to X such that
Φ(t + τ )u = T (t)Φ(τ )u + Φ(t)u(· + τ ), ∀u ∈ Lp(R+, U), τ ≥ 0, t ≥ 0,
we call (T, Φ) an abstract linear control system;
(iii) Ψ = {Ψ(t)}t≥0 is a family of bounded linear operators, called output maps, from
X to Lp(R+, Y ) such that
(Ψ(t + τ )x)(s) = (Ψ(t)T (τ )x)(s − τ ), ∀x ∈ X, t + τ ≥ s ≥ τ ≥ 0, t ≥ 0,
we call (T, Ψ) an abstract linear observation system;
(iv) F = {F (t)}t≥0 is a family of bounded linear operators, called input-output map,
from Lp(R+, U) to Lp(R+, Y ) such that
(F (t+τ )u)(s) = (Ψ(t)Φ(τ )u+F (t)u(·+τ ))(s−τ ), ∀u ∈ Lp(R+, U), t+τ ≥ s ≥ τ ≥ 0, t ≥ 0.
By a representation theorem due to Salamon [35] (see also Weiss [39]), corresponding
to abstract linear control system (T, Φ), there is a unique control operator B ∈ L(U, X−1),
called admissible control operator (also p-admissible control operator), satisfying
Φ(t)u = Z t
0
T−1(t − s)Bu(s)ds ∈ X, ∀u ∈ Lp(R+, U), t ≥ 0.
Here T−1 is the extrapolation semigroup, which is the continuous extension of T to the
extrapolation space X A
−1 defined by the completion of X under the norm kR(λ0, A)·k with
R(λ0, A) being the resolvent of A and λ belonging the resolvent set of A. The generator of
T−1 is the continuous extension of A to X and is denoted by A−1. In this case, we also say
(A, B) generates an abstract linear control system and denote ΦA,B. Moreover, if ΦA,B(τ )
is surjective, we call (A, B) to be exactly controllable (also exactly p-controllable) at τ.
It follows from Salamon [35] or Weiss [40] that an abstract linear observation sys-
tem (T, Ψ) corresponds a unique operator, called admissible observation operator (also
p-admissible observation operator) C ∈ L(D(A), Y ) satisfying
Z t0
0
kCT (t)xkpdt ≤ c(t0)kxkp, ∀x ∈ D(A)
such that (Ψ(t)x)(τ ) = CT (τ )x, ∀x ∈ D(A), τ ≤ t.
In this case, we also say (A, C)
generates an abstract linear control system and denote ΨA,C. Moreover, (A, C) is called
5
to be exactly observable (also exactly p-observable) at τ , provided there exists a constant
k > 0 such that kΨA,C(τ )xk ≥ kkxk, x ∈ X.
Let Σ = (T, Φ, Ψ, F ) be well-posed linear system. For any x(0) ∈ X, u ∈ Lp
loc(R+, U),
y = Ψ(∞)x0 + F (∞)u, where Ψ(∞) : X → L2
Lp
x(t) = T (t)x(0) + Φ(t)u is the solution of equation x(t) = A−1x(t) + Bu(t). Define output
loc(R+, U) →
loc(R+, Y ) are the extended output map defined by the strong limit of Φ(τ ) and F (τ )
as τ → +∞, respectively (see [40, 41]). In the special case u = 0, it follows from [40,
loc(R+, Y ) and F (∞) : Lp
Theorem 4.5 and Proposition 4.7] that for any x(0) ∈ X, y(t) = C A
Λ T (t)x(0) a.e. t ≥ 0,
where C A
Λ defined by
C A
Λ x = lim
λ→∞
CλR(λ, A)x, x ∈ D(C A
Λ ) = {x ∈ X : this above limit exists in Y }
(2.1)
is called Λ-extension of C with respect to A. By [36], it follows that the output y(t) can
be expressed by
y(t) = C A
Λ [x(t) − (λ − A−1)−1B]u(t) + G(λ)u(t),
a.e. t ≥ 0, where G(λ) is the transform function. It is not hard to see that Definition 2.1
implies continuous dependence, that is, there exist positive function m and n on R+ such
that
kx(t)k + kykLp([0,t],Y ) ≤ m(t)kx(0)k + n(t)kukL([0,t],U ), t ≥ 0.
The well-posed linear system Σ is called a regular linear system if, there exists a
bounded operator D, called feedthrough operator, such that the limit
lim
s→0
1
s Z s
0
(F (t)u0)(σ)dσ = Dz
exists in Y for the constant input u0(t) = z, z ∈ U, t ≥ 0. Weiss showed that well-posed
linear system Σ is regular if and only if G(λ) strongly converges to D as λ → +∞, that
is,
The regular linear system is described by
lim
λ→+∞
G(λ)u = Du, u ∈ U.
x(t) = Ax(t) + Bu(t), y(t) = C A
Λ x(t) + Du(t).
6
Definition 2.2 [37, 43] An operator Γ ∈ L(Y, U) is called an admissible feedback operator
for Σ = (T, Φ, Ψ, F ) if I − F (t)Γ is invertible for some t ≥ 0 (hence any t ≥ 0).
Theorem 2.3 [43] Let (A, B, C, D) be the generator of regular linear system Σ = (T, Φ, Ψ, F )
on (X, U, Y ) with admissible feedback operator Γ ∈ L(Y, U). Suppose that I −DΓ is invert-
ible. Then the feedback system ΣΓ is a well-posed linear system generated by (AΓ, BΓ, C Γ):
AΓ = (A−1 + BΓ(I − DΓ)−1
lef tC A
Λ )X,
D(AΓ) := {z ∈ D(C A
Λ ) : (A−1 + BΓ(I − DΓ)−1C A
Λ )z ∈ X},
C Γ = (I − DΓ)−1C A
defined by J A,AΓ
Λ restricted to D(AΓ) and BΓ = J A,AΓ(I − DΓ)−1B, where J A,AΓ is
−1 : the limit
−1 ) with D(J A,AΓ) = {x ∈ X A
x = limλ→∞(λ − A−1)−1x (in X AΓ
limλ→∞(λ − A−1)−1x exists }.
In the rest of this section, we introduce some notions related to linear boundary
system described in the abstract frame as follows [25, 34].
z(t)
= Lz(t),
Gz(t) = u(t),
y(t)
= Kz(t),
(2.2)
where(cid:20)
L
G
K
(cid:21) is closed linear operators from D(L) to space X×U ×Y ; D(L) is continuously
embedded in X; G is surjection and Ker{G} := {z ∈ Z : Gz = 0} is dense in X; LKer{G}
generates a C0-semigroup on X. We denote system (2.2) by (L, G, K) for brief.
Denote A = LG, C = KD(A). By [3], D(L) can be decomposed to direct sum
D(L) = D(A)L Ker{λ − L} and the operator G is bijective from Ker{λ − L} onto U,
where λ is any component of resolvent set ρ(A) of A. Hence we can denote Dλ,L,G by the
solution operator from z to u of the following function
(λ − L)z = 0,
Gz = u,
7
that is z = Dλ,L,Gu. By [25, 34], it follows that boundary control system
is equivalent to system
z(t) = Lz(t),
Gz(t) = u(t),
z(t) = A−1z(t) + Bu(t)
in the sense of classical solution, where B is given by B = (λ − A−1)Dλ,L,G ∈ L(U, X−1).
Concretely, z(t) and u(t) satisfying Gz(t) = u(t) is equivalent to A−1z(t) + Bu(t) ∈ X;
if z(0) ∈ X, u ∈ W 2,p(R+, U) satisfying A−1z(0) + Bu(0) ∈ X, we have z(t) = Lz(t) =
A−1x(t)+Bu(t). Moreover the initial condition implies that y(t) = Kz(t) = C(x(t)−(λ−
A−1)−1Bu(t)) + K(λ− A−1)−1Bu(t). Hence C is the observation operator of the boundary
system (L, G, K) and the corresponding transform function is K(λ − A−1)−1B, λ ∈ ρ(A).
Boundary system (L, G, K) is well-posed if there exist positive function m and n on
R+ such that
kx(t)k + kykLp([0,t],Y ) ≤ m(t)kx(0)k + n(t)kukL([0,t],U ), t ≥ 0.
It is regular if it is well-posed and the strong limit of the transform function exists, that is,
limλ→+∞ KDλ,L,Gu exists for any u ∈ U. In this case, denote by K A,B the corresponding
feedthrough operator, which means K A,Bu := limλ→+∞ Dλ,L,Gu, u ∈ U. Then boundary
system (L, G, K) is regular with generator (A, B, C, KA,B). We also say that the generator
is (A, B, K, KA,B) and denote K A
Λ by C A
Λ = (KD(A))A
Λ.
The following two lemmas will be used in the next section.
Lemma 2.4 [30] Assume that the boundary control system
(BCS)
z(t)
= Lz(t)
Gz(t) = u(t)
is an abstract linear control system generated by (A, B). Then the boundary system
(L, G, Q) is a regular linear system on (X, U, Y ) if and only if (A, B, Q) generates a
regular linear system. In this case, for any z ∈ Z, we have
Qz = QA
Λz + QGz.
8
Lemma 2.5 [30] Assume that the boundary system (L, G, Q) is a regular linear system
generated by (A, B, Q, QA,B) on (X, U, X) with admissible feedback operator I. Then the
system
(OS)
z(t)
= Lz(t)
Gz(t) = Qz(t) + v(t)
is an abstract linear control system generated by (AI , BI).
3 Main Results
In this section, we shall obtain the admissible controllability, exact controllability, ad-
missible observation, exact observation and regularity under some regular perturbations.
Moreover, all the perturbation results are used to solve the corresponding boundary sys-
tems. The systems governed by specific partial differential equations are presented after
every perturbation result to illustrate our results.
We first consider the admissible controllability and exactly controllability under as-
sociated perturbation. To prove the robustness of exact controllability, we introduce the
following important lemma related to radius of surjectivity.
Lemma 3.1 [21, page 227] Let E and F be Banach spaces. Then, S(E, F ) := {Ξ ∈
L(E, F ) : Ξ is surjective} is an open set in L(E, F ), i.e., given Π ∈ S(E, F ), there
exists α > 0 such that
{Ξ ∈ L(E, F ) : kΠ − Ξk < α} ⊂ S(E, F ).
The constant α is called radius of surjectivity of Π.
Theorem 3.2 Let (A, B, C, D) generate a regular linear system with admissible feedback
operator I on (X, Y, Y ), I − D is invertible, and (A, ∆B, C, P ) generate a regular linear
system on (X, U, Y ). Then (AI , J A,AI
B(I − D)−1P + J A,AI ∆B, (I − D)−1C A
Λ ) generates
a regular linear system, and there holds
ΦAI ,J A,AI B(I−D)−1P +J A,AI ∆B = ΦA,B(I − FA,B,C,D)−1FA,∆B,C,P + ΦA,∆B,
9
where AI = (A+ B(I − D)−1C A
Λ )X. Moreover, if (A, ∆B) is exactly controllable at t0 > 0,
then there exists k0 > 0 such that (AI(k), J A,AI (k)B(I − kD)−1kP + J A,AI (k)∆B) is also
exactly controllable at t0 whenever k < k0, where AI(k) = (A + B(I − kD)−1kC A
Λ )X.
Proof. We consider the operators B := (B, ∆B) : Y × U → X−1, C =
C
0
: X1 →
Y × U, D =
D P
0
0
: Y × U → Y × U.
Since (A, B, C, D) and (A, ∆B, C, P ) generate regular linear systems, it is easy to
verify that (A, B, C, D) generates a regular linear system given by
ΣA, B, C :=
T
ΨA,C
0
(ΦA,B, ΦA,∆B)
FA,B,C,D FA,∆B,C,P
0
0
.
Observe that I is an admissible feedback operator for ΣA,B,C,D. We have that
IY ×U −
FA,B,C,D FA,∆B,C,P
0
0
=
I − FA,B,C,D −FA,∆B,C,P
0
I
is invertible and
(cid:18)IY ×U −
FA,B,C,D FA,∆B,C,P
0
0
(cid:19)−1
=
(I − FA,B,C,D)−1
(I − FA,B,C,D)−1FA,∆B,C,P
0
I
,
that is, IX×U is an admissible feedback operator for ΣA, B, C, D. It follows from Theorem
2.3 that AIX×U = (A−1 + B(I − D) C A
Λ )X = AI, BIX×U =
J A,AI B(I − D)−1 = (cid:16) J A,AI
X×U = (I −
B(I − D)−1P + J A,AI ∆B (cid:17), C I
Λ )X = (A−1 + B(I − D)−1C A
B(I − D)−1 J A,AI
10
D)−1 C A
(I − D)−1C A
Λ
Λ =
ΦAI ,J A,AI B(I−D)−1P +J A,AI ∆B
and
0
=ΦAI , BIX×U
0
I
=ΦA, B(I − FA, B, C, D)−1
0
I
=(ΦA,B, ΦA,∆B)
(I − FA,B,C,D)−1
(I − FA,B,C,D)−1FA,∆B,C,P
0
I
0
I
=ΦA,B(I − FA,B,C,D)−1FA,∆B,C,P + ΦA,∆B.
Moreover, it is not hard to see that
(AI , J A,AI
B(I −D)−1P +J A,AI
is a regular linear system.
∆B, (I −D)−1C A
Λ ) = (cid:18)AI , BIX×U
0
I
,
I
0
C IX×U )(cid:19)
Below we prove the robustness of exact controllability. Observe that I −kD is invert-
kDk (if kDk = 0,
ible for any k < 1
4.10], kI is admissible feedback for (A, B, C, D) whenever for k < 1
1
kDk = +∞). By [43, Proposition 3.12 and Proposition
kDk, which indicates
that I is admissible feedback for (A, B, kC, kD). Since (A, ∆B) is exactly controllable at
t0 > 0, ΦA,∆B(t0) is surjective. Let s0 be the radius of surjectivity of ΦA,∆B(t0). It follows
from the above proof that
kΦAI ,J A,AI (k)B(I−kD)−1kP +J A,AI (k)∆B(t0) − ΦA,∆B(t0)k
=kΦA,B(t0)(I − FA,B,kC,kD(t0))−1FA,∆B,kC,kP (t0)k
≤kkΦA,B(t0)(I − kFA,B,C,D(t0))−1kkFA,∆B,C,P (t0)k.
Let
k0 = min{
1
kDk
,
1
kFA,B,C,D(t0)k
,
kΦA,BkkFA,B,C,P (t0)k + s0kFA,B,C,D(t0)k
s0
}.
11
Then kΦAI ,J A,AI (k)B(I−kD)−1kP +J A,AI (k)∆B(t0)−ΦA,∆B(t0)k < s0 whenever k < k0. It follows
from Lemma 3.1 that ΦAI ,J A,AI (k)B(I−kD)−1kP +J A,AI (k)∆B(t0) is surjective. This implies that
(AI(k), J A,AI (k)B(I − kD)−1kP + J A,AI (k)∆B) is exactly controllability at t0. The proof
is therefore completed.
Remark 3.3 In the special case that Y = X, D = 0, P = 0 and C = I, the above
theorem says that both B and ∆B being admissible for A implies that J A,(A−1+B)X ∆B
is admissible for (A−1 + B)X, such result has been proved by Hadd [16], as mentioned
in the introduction section. If Y = X, D = 0, P = 0 and B = I, Theorem 3.2 tells
that (A, ∆B, C) generating a regular linear system implies that (A + C, J A,A+C∆B, C)
generates a regular linear system, particularly, J A,A+C∆B is admissible for A + C, such
result has been proved by Mei and Peng [27]. This means that our result is a generalization
of [16] and [27].
Theorem 3.4 Assume that the boundary system
z(t)
= Lz(t)
G1z(t) = u(t)
G2z(t) = 0
y(t)
= Kz(t)
is a regular linear
system generated by (A, B1, K, KA,B1) on (X, U, U) with I being admissible feedback oper-
ator. Suppose that
operator B2. Then
z(t)
= Lz(t)
G1z(t) = 0
G2z(t) = v(t)
y(t)
z(t)
= Kz(t)
= Lz(t)
is regular linear system on (X, V, U) with control
G1z(t) = Kz(t)
is an abstract linear control system generated by
G2z(t) = v(t)
(AI , J A,AI
B1(I − K A,B1)−1K A,B2 + J A,AI
B2). If
z(t)
= Lz(t)
G1z(t) = 0
is exactly control-
G2z(t) = v(t)
= Lz(t)
z(t)
lable at t0, there exists k0 > 0 such that
G1z(t) = kKz(t)
is exactly controllable at
G2z(t) = v(t)
12
t0 whenever k < k0.
Proof. By the assumption, G1 : D(L)T KerG2 → U and G2 : D(L)T KerG1 → V are
surjectives. This implies that for any u ∈ U and v ∈ V , there exist z1 ∈ D(L)T KerG2
and z2 ∈ D(L)T KerG1 such that G1z1 = u, G2z2 = v. Hence, z1 + z2 ∈ D(L) and
are surjective. Moreover, B1 = (λ −
(z1 + z2) =
., that is,
A−1)D1,λ and B1 = (λ − A−1)D2,λ are indicated by the assumption. Here D1,λu and D1,λu
G1
G2
G1
G2
u
v
are the solution of the equations
, respectively. Hence,
λz = Lz,
G1z = 0,
G2z = v
λz = Lz,
G1z = u,
G2z = 0
and
satisfies Dλ
the solution Dλ
u
v
of
λz = Lz,
G1z = u,
G2z = v
So the control operator of boundary system
u
v
= (cid:16) D1,λ D1,λ (cid:17)
u
v
.
z(t) = Lz
G1z = u(t)
G2z = v(t)
is B = (λ − A−1)Dλ =
(λ − A−1)(cid:16) D1,λ D2,λ (cid:17) = (cid:16) B1 B2 (cid:17) . Let Y (t) =
z(t). Since (A, B1, K) and
(A, B2, K) are regular linear system, we obtain that (cid:18)A, B,
(cid:19) is a regular linear
K
K
0
0
system. Then
λz = Lz,
G1z = u,
G2z = v
with output Y (t) is a regular linear system, the feedthrough
13
operator D0 is computed by
D0
p
q
K
0
(λ − A)−1B
p
q
λ→+∞
= lim
=
=
limλ→+∞ K(λ − A)−1B1p + limλ→+∞ K(λ − A)−1B2q
K A,B1 K A,B2
0
0
p
q
p
q
∈ U × V.
0
, ∀
,
it follows that
B
ator(cid:18)AI , J A,AI
Sine I is admissible feedback operator of regular linear system (A, B1, K, KA,B1), I is
admissible feedback operator for (cid:18)A, B,
K
0
z(t)
= Lz(t)
K A,B1 K A,B2
0
0
(cid:19). By Lemma 2.5,
G1z(t) = Kz(t) + u(t)
is an abstract linear control system with gener-
G2z(t) = v(t)
(I − K A,B1)−1
(I − K A,B1)−1K A,B2
0
I
z(t)
= Lz(t)
G1z(t) = Kz(t)
(cid:19). Hence
G2z(t) = v(t)
B1(I − K A,B1)−1K A,B2 +
is an abstract linear control system with generator (AI , J A,AI
J A,AI
B2). The rest result is obtained directly from Theorem 3.2. This completes the
proof.
Example 3.5 We consider Schrodinger equation equation with Dirichlet boundary control
and observation described by
wtt(x, t) = ∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = −
∂(∆−1w)
∂ν
, x ∈ Γ1, t ≥ 0,
w(x, t) = u(x, t), x ∈ Γ0, t ≥ 0,
(3.1)
where Ω ⊂ Rn, n ≥ 2 is an open bounded region with smooth C 3-boundary ∂Ω = Γ0 ∪ Γ1.
Γ0, Γ1 are disjoint parts of the boundary relatively open in ∂Ω, int(Γ1) 6= ∅ and int(Γ0) 6= ∅,
ν is the unit normal vector of Γ0 pointing towards the exterior of Ω, u is the input function
(or control) and y is the output function (or output).
14
Let H = H −1(Ω) be the state space and U = L2(∂Ω) be the control (input) or
observation (output) space. It has been proved in [5] that
wt(x, t) = −i∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = v(x, t), x ∈ Γ1, t ≥ 0,
w(x, t) = 0, x ∈ Γ0, t ≥ 0,
y(x, t) = −i
∂(−∆)−1w)
∂ν
, x ∈ Γ1, t ≥ 0
is regular linear systems with feedthrough operator zero and I being admissible feedback
operator. Similarly, one can obtain that
is regular linear systems with feedthrough operator zero. The combinations of [1] and [4]
implies that system
wt(x, t) = −i∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = 0, x ∈ Γ1, t ≥ 0,
w(x, t) = u(x, t), x ∈ Γ0, t ≥ 0,
y(x, t) = −i
∂((−∆)−1w)
∂ν
, x ∈ Γ1, t ≥ 0
wt(x, t) = −i∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = 0, x ∈ Γ1, t ≥ 0,
w(x, t) = u(x, t), x ∈ Γ0, t ≥ 0,
y(x, t) = −i
∂((−∆)−1w)
∂ν
, x ∈ Γ1, t ≥ 0
is exactly controllable at some t0 > 0. By Theorem 3.4, it follows that (3.1) is an abstract
linear control system. Moreover, there exists a constant k0 > 0 such that
wt(x, t) = −i∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = −ki
∂((−∆)−1w)
∂ν
, x ∈ Γ1, t ≥ 0,
w(x, t) = u(x, t), x ∈ Γ0, t ≥ 0,
is exactly controllable at t0 > 0 whenever k < k0.
Next, we are concerned with admissible observation and exactly observation under
some regularity perturbation.
15
Theorem 3.6 Let (A, B, C, D) generate a regular linear system with admissible feedback
operator I on (X, U, U), and (A, B, ∆C, P ) generate a regular linear system on (X, U, Y ).
Then (AI , BI , P (I − D)−1C A
Λ + ∆C A
Λ ) generates a regular linear system, and there holds
ΨAI ,P (I−D)−1C A
Λ +∆C A
Λ
= FA,B,∆C,P (I − FA,B,C,D)−1ΨA,C + ΨA,∆C,
where AI = (A + B(I − D)−1C A
Λ )X and BI = J A,AI
B. Moreover, if (A, ∆C) is exactly
observable at t0 > 0, then there exists k0 > 0 such that (AI(k), kP (I − kD)−1C A
is also exactly observable at t0 whenever k < k0, where AI(k) = (A + B(I − kD)−1kC A
Λ + ∆C A
Λ )
Λ )X.
Proof. Similar to the proof of [30, Lemma 4.3], let B = (B, 0), C =
C
∆C
,
D =
D 0
P 0
system and
, we obtain that (AI , BI , P (I −D)−1C A
Λ +∆C A
Λ ) generates a regular linear
ΨAI ,P (I−D)−1C A
Λ +∆C A
Λ
=(0, I)ΨAI , BIX×U
=(0, I)(I − FA, B, C, D)−1ΨA, C
=(0, I)
=FA,B,∆C,P (I − FA,B,C,D)−1ΨA,C + ΨA,∆C.
FA,B,∆C,P (I − FA,B,C,D)−1
(I − FA,B,C,D)−1
0
I
ΨA,C
ΨA,∆C
Below we prove the robustness of exactly observability. As stated in the proof of
Theorem 3.2, kI is admissible feedback for (A, B, C, D) whenever for k < 1
kDk, which
indicates that I is admissible feedback for (A, B, kC, kD). Since (A, ∆C) is exactly ob-
servable at t0, there exists a constant k0 > 0 such that kΨA,∆C(t0)xk ≥ k0kxk, x ∈ X. It
follows from the above proof that
kΨAI ,kP (I−kD)−1C A
Λ +∆C A
Λ
(t0)xk
≥kΨA,∆C(t0)xk − kΨAI ,kP (I−kD)−1C A
Λ +∆C A
Λ
(t0)x − ΨA,∆C(t0)xk
≥k0kxk − kFA,B,∆C,P (t0)(I − FA,B,kC,kD(t0))−1ΨA,kC(t0)xk
=k0kxk − kFA,B,∆C,P (t0)(I − kFA,B,C,D(t0))−1kΨA,C(t0)xk.
16
Let α0 ∈ (0, k0) and
θ0 = min{
1
kDk
,
1
kFA,B,C,D(t0)k
,
Then
(k0 − α0)kFA,B,C,D(t0)k + kFA,B,∆C,P (t0)kΨA,C(t0)kk
k0 − α0
}.
kΨAI ,kP (I−kD)−1C A
Λ +∆C A
Λ
(t0)xk > α0kxk,
whenever k < θ0. The proof is therefore completed.
Remark 3.7 In the special case that Y = X and B = I, the above theorem says that both
P and C being admissible for A implies that C is admissible for A + C, such result has
been proved by Hadd [16]. If Y = X and C = I, theorem tells that (A, B, P ) generating
a regular linear system implies that ((A−1 + B)X, J A,(A−1+B)X B, P A
Λ ) generates a regular
linear system, particularly, J A,(A−1+B)X B is admissible for A + P . This means that our
result is a generalization of [16].
Theorem 3.8 Assume that the boundary system (L, G, Q) is a regular linear system gen-
erated by (A, B, G, GA,B) on (X, U, U) with admissible feedback operator I. Suppose that
boundary system (L, G, K) is a regular linear system on (X, U, Y ). Then the system
z(t)
= Lz(t)
Gz(t) = Qz(t)
y(t)
= Kz(t)
(3.2)
is an abstract linear observation system generated by (AI , K). If, in addition, system
z(t)
= Lz(t)
Gz(t) = 0
is exactly observable at some t0 > 0, there exists a constant θ0 > 0
y(t)
= Kz(t)
such that system
z(t)
= Lz(t)
Gz(t) = kQz(t)
is exactly observable at t0 > 0 whenever k < θ0.
y(t)
= Kz(t)
Proof. Since boundary system (L, G, K) is a regular linear system with admissible feed-
back operator I, it follows from Lemma 2.4 that
Kz = K A
Λ z + K A,BGz, z ∈ Z,
(3.3)
17
and
Gz = Qz, z ∈ D(AI) ⊂ D(L).
The assumption (L, G, Q) is a regular linear system implies
Qz = QA
Λz + QA,BGz, z ∈ D(L).
(3.4)
(3.5)
Observe that I − QA,B is invertible. The combination of (3.4) and (3.5) implies that
Qz = (I − QA,B)−1QA
Λz, z ∈ D(AI), substituted which into 3.3 to get
Kz = K A
Λ z + K A,B(I − QA,B)−1QA
Λz, z ∈ D(AI).
By Theorem 3.6, (3.2) is an abstract linear observation system generated by (AI , K).
Furthermore, the rest result is obtained directly from Theorem 3.6. This completes the
proof.
Example 3.9 Consider the following one-dimensional Euler-Bernoulli beam equation
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = wt(1, t),
(3.6)
y(t) = wx(1, t).
It follows from [4] that
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = u(t),
(3.7)
y(t) = wt(1, t).
is a regular linear system with admissible feedback operator I and the corresponding feed-
back operator is zero. By Theorem 3.4, to obtain that (3.6) is an abstract linear observa-
tion system, we only have to prove that
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = u(t),
(3.8)
y(t) = wx(1, t),
.
is a regular linear systems. We divide the rest proof into three steps.
18
Step 1. Boundary observation system
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = 0,
y(t) = wx(1, t),
.
is an abstract linear observation system. To this end, we let
F (t) =
1
2 Z 1
0
[w2
t (x, t) + w2
xx(x, t)]dx.
It is not hard to see that
F (t) = 0 thereby F (t) = F (0), t ≥ 0. Set
ρ(t) = Z 1
0
x(x − 1)wt(x, t)wx(x, t)dx.
We obtain ρ(t) ≤ F (t) = F (0). Take the derivative with respect to the time on both sides
to get
ρ(t) =Z 1
= −Z 1
0
0
x(x − 1)wtt(x, t)wx(x, t)dx +Z 1
0
x(x − 1)wxxxx(x, t)wx(x, t)dx +
= − x(x − 1)wxxx(x, t)wx(x, t)1
(2x − 1)wxxx(x, t)wx(x, t)dx +
x(x − 1)wt(x, t)wxt(x, t)dx
1
2 Z 1
0
x(x − 1)
∂
∂x
w2
t (x, t)dx
x(x − 1)wxxx(x, t)wxx(x, t)dx
0
1
2 Z 1
0
x(x − 1)
∂
∂x
w2
t (x, t)dx
=
x(x − 1)
w2
xx(x, t)dx +
∂
∂x
x(x − 1)
∂
∂x
w2
t (x, t)dx
x=0 +Z 1
1
2 Z 1
x=0 −Z 1
0
0
0
+Z 1
2 Z 1
1
0
−Z 1
2 Z 1
0
1
0
= −
+ (2x − 1)wxx(x, t)wx(x, t)1
wxx(x, t)[2wx(x, t) + (2x − 1)wxx(x, t)]dx
=
1
2
x(x − 1)[w2
t (x, t) + w2
xx(x, t)]1
x=0 −
1
2 Z 1
0
(2x − 1)[w2
t (x, t) + 3w2
xx(x, t)]dx
2wxx(x, t)wx(x, t)dx
(2x − 1)[w2
t (x, t) + 3w2
xx(x, t)]dx − w2
x(1, t).
Integrate from 0 to T with respect to t to derive
Z T
0
w2
x(1, t)dt = Z T
0 (cid:20) −
1
2 Z 1
0
(2x − 1)[w2
t (x, t) + 3w2
xx(x, t)]dx(cid:21)dt + ρ(0) − ρ(T )
≤ (3T + 2)F (0).
19
Step 2. Boundary system (3.8) is a well-posed. We consider the boundary system
under the zero initial condition: w(x, 0) = wt(x, 0) = 0. Define F (t) and ρ as the same
in Step 1. By [8], it follows that
F (t) ≤ Cδ,T Z T
1+4T ) and Cδ,T = 1+δ+4T
1
0
2[1−(1+4T )δ] + 1
2δ .
u2(t)dt, ∀t ∈ [0, T ],
where δ ∈ (0,
Observe that ρ(t) ≤ F (t) holds. Take the derivative with respect to the time on both
sides to get
ρ(t) = −
1
2 Z 1
0
(2x − 1)[w2
t (x, t) + 3w2
xx(x, t)]dx − w2
x(1, t).
Integrate from 0 to T with respect to t to derive
(2x − 1)[w2
t (x, t) + 3w2
xx(x, t)]dx(cid:21)dt − ρ(T )
Z T
0
w2
0
1
2 Z 1
x(1, t)dt = Z T
0 (cid:20) −
≤ 3Z T
≤ (1 + 3T )Cδ,T Z T
0
0
F (t)dt + F (T )
u2(t)dt.
Step 3. Boundary system (3.8) is regular. Denote by w(x, s) the Laplace transform
of w(x, s) with respect to t, that is, w(x, s) = R ∞
transform u(s) of u(t) with respect to t is u(s) = R ∞
condition w(x, 0) = wt(x, 0) = 0, we get
0 w(x, t)e−stdt. Similarly, the Laplace
0 u(t)e−stdt. For the zero initial
s2 w(x, s) + wxxxx(x, s) = 0,
w(0, s) = wx(0, s) = wxx(1, s) = 0, wxxx(1, s) = u(s),
y(s) = wx(1, t),
x ∈ (0, 1)
.
Denote by H(s) the corresponding transform function. Since the system is well-posed.
Then we have that H(s) satisfies that y(s) = H(s)u(s) and it is bounded on some right
half plane. In order derive the regularity, we only need to show that the limit of transfer
function exists as s → +∞. So we can set s > 0 and t = p s
that
2. The first equation implies
w(x, s) = ach(tx)cos(tx) + bch(tx)sin(tx) + csh(tx)cos(tx) + dsh(tx)sin(tx)
20
with a, b, c and d being to be determined. Use w(0, s) = 0 to get a = 0. We obtain
wx(x, s) =t[bsh(tx)sin(tx) + (b + c)ch(tx)cos(tx) − csh(tx)sin(tx)
+ dch(tx)sin(tx) + dsh(tx)cos(tx)].
Use wx(0, s) = 0 to get b + c = 0. We obtain
wxx(x, s) =2t2[bch(tx)sin(tx) + bsh(tx)cos(tx) + dch(tx)cos(tx)]
and
wxxx(x, s) =2t3[2bch(tx)cos(tx) + dsh(tx)cos(tx) − dch(tx)sin(tx)].
Use wxx(1, s) = 0 and wxxx(1, s) = u(s) to get
b =
chtcost
2t3(ch2t+cos2t) u(s),
d = − chtsint+shtcost
2t3(ch2t+cos2t) u(s).
Hence we obtain that
Observe that
H(s) = −
shtchtsintcost − [chtsint + shtcost]2
2t2(ch2t + cos2t)
.
H(s) ≤
ch2t + (2cht)2
2t2ch2t
=
5
2t2 =
5
s
.
Hence H(s) → 0 as s → +∞. The regularity of (3.8) is therefore proved. This completes
the proof.
Example 3.10 Consider one-dimensional Euler-Bernoulli beam equation
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = wt(1, t),
(3.9)
y(t) = wxx(0, t).
Guo, Wang and Yang showed in [8] that
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = w(1, t),
(3.10)
y(t) = wxx(0, t).
21
is a well-posed linear system. From Step 3 of the above Example 3.9, we obtain that the
transform function of system (3.10) is presented on R+ by
H1(s) =
wxx(0, s)
u(s)
=
2t2d
u(s)
= −
chtsint + shtcost
t(ch2t + cos2t)
with t = p s
2. Then
H1(s) ≤
cht + sht
tch2t
≤
2
tcht
→ 0,
as s → +∞, that is, system (3.10) is regular. Observe that system (3.7) is regular. By
Theorem 3.8, system (3.9) is an abstract linear observation system.
Next, we show that system
is exactly observable. Let
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = 0,
(3.11)
y(t) = wxx(0, t),
F (t) =
1
2 Z 1
0
[w2
t (x, t) + w2
xx(x, t)]dx.
We have F (t) = 0 thereby F (t) = F (0), t ≥ 0. Set
ρ1(t) = Z 1
0
(x − 1)wt(x, t)wx(x, t)dx.
Obviously, ρ1(t) ≤ F (t) = F (0). We compute
ρ1(t) =
1
2
w2
xx(0, t) −
1
2 Z 1
0
Integrate from 0 to T with respect to t to get
(w2
t (x, t) + 3w2
xx(x, t))dx.
Z T
0
w2
xx(0, t)dt =
1
0 Z 1
2 Z t
0
(w2
t (x, t) + 3w2
xx(x, t))dxdt + ρ1(0) − ρ1(T )
≥(T − 2)E(0).
This indicates that system (3.11) is exactly observable at any T > 2. Then, by Theorem
3.8, system
wtt(x, t) + wxxxx(t, x) = 0,
x ∈ (0, 1)
w(0, t) = wx(0, t) = wxx(1, t) = 0, wxxx(1, t) = wt(1, t),
y(t) = wxx(0, t).
is exactly observable at T > 2 whenever k is small enough.
22
In the rest of this section, we are concern with the regularity under perturbations.
Theorem 3.11 Assume that (A, B, C, D) generates a regular linear system with admis-
sible feedback operator I. Suppose (A, ∆B, C), (A, B, ∆C) and (A, ∆B, ∆C) generates
regular linear systems. Then (AI , J A,AI ∆B, ∆C A
Λ ) generates a regular linear system.
Proof. By Theorem 3.2, it follows that (AI , J A,AI ∆B) generates an abstract linear control
system with
ΦAI ,J A,AI ∆B = ΦA,B(I − FA,B,C,D)−1FA,∆B,C + ΦA,∆B.
Theorem 3.6 implies that (AI , ∆C A
Λ ) generates an abstract linear observation system with
ΨAI ,∆C A
Λ
= FA,B,∆C(I − FA,B,C,D)−1ΨA,C + ΨA,∆C.
Since (A, B, ∆C) and (A, ∆B, ∆C) are regular linear system, we define F = FA,B,∆C(I −
FA,B,C,D)−1FA,∆B,C+FA,∆B,∆C. Then it is not hard to verify that (TAI , ΦAI ,J A,AI ∆B, ΨAI ,∆C A
is a regular linear system generated by (AI , J A,AI ∆B, ∆C A
Λ ). The proof is therefore com-
Λ
, F )
pleted.
Remark 3.12 In the special case that Y = X and C = I, the above theorem says
that both (A, B, ∆C) and (A, ∆B, ∆C) being regular linear system implies that ((A−1 +
B)X, J A,(A−1+B)X ∆B, ∆C A
Λ ) generates a regular linear system, such result has been proved
by Hadd [16]; If Y = X and B = I, the above theorem says that both (A, ∆B, C) and
(A, ∆B, ∆C) being regular linear system implies that (A + C, J A,+C∆B, ∆C) generates a
regular linear system. This means that our result is a generalization of [16].
Theorem 3.13 Assume that the boundary system
system generated by (A, B1, K, KA,B1) with I being admissible feedback operator. Suppose
z(t)
= Lz(t)
G1z(t) = u(t)
G2z(t) = 0
y(t)
= Kz(t)
is regular linear
z(t)
= Lz(t)
G1z(t) = u(t)
G2z(t) = 0
y(t)
= W z(t)
and
that
z(t)
= Lz(t)
G1z(t) = 0
G2z(t) = v(t)
y(t)
K
W
z(t)
=
23
are regular linear systems with
B2 being the control operator of the second system. Then
z(t)
= Lz(t)
G1z(t) = Kz(t)
G2z(t) = v(t)
y(t)
= W z(t)
(3.12)
is a regular linear system generated by (AI , J A,AI
B1(I−K A,B1)−1K A,B2+J A,AI
B2, W, W A,B1(I−
K A,B1)−1K A,B2 + W A,B2).
Proof. By Theorem 3.4, it follows that
z(t)
= Lz(t)
G1z(t) = Kz(t)
is an abstract linear control
G2z(t) = v(t)
system with generator (AI , J A,AI
B1(I − K A,B1)−1K A,B2 + J A,AI
B2). It follows from Theo-
rem 3.8 that
z(t)
= Lz(t)
G1z(t) = Kz(t)
G2z(t) = 0
y(t)
= W z(t)
is an abstract linear observation system with generator
(AI , W ) and the restriction of W to D(AI) is equal to W A
Λ + W A,B1(I − K A,B1)−1K A
Λ . By
Theorem 3.11, our assumptions imply that (AI , J A,AI
B2, W A
system. Combining this with the boundedness of operator J A,AI
Λ ) generates a regular linear
B1(I − K A,B1)−1K A,B2
implies that (AI , J A,AI
B1(I − K A,B1)−1K A,B2 + J A,AI
B2, W A
Λ ) generates a regular linear
system. By Theorem 3.2, we obtain that (AI , J A,AI
B1(I − K A,B1)−1K A,B2 + J A,AI
Λ ) generates a regular linear system. Since W A,B1 is bounded, (AI , J A,AI
B2, (I −
B1(I −
K A,B1)−1K A
K A,B1)−1K A,B2 + J A,AI
Therefore, (AI , J A,AI
B2, W A,B1(I − K A,B1)−1K A
B1(I − K A,B1)−1K A,B2 + J A,AI
Λ ) generates a regular linear system.
B2, W ) is a regular linear system.
Hence the regularity of system (3.12) is obtained by Lemma 2.4.
Next, we shall compute the feedthrough operator. By Theorem 3.2, for any enough
big Re(λ), we have
(λ − (AI)−1)−1(cid:18)J A,AI
=(λ − A−1)−1B1(I − GA,B1,K,KA,B1
B1(I − K A,B1)−1K A,B2 + J A,AI
B2(cid:19)
(λ) + (λ − A−1)−1B2,
(λ))−1GA,B2,K,KA,B2
24
and the transform function of (3.12) is given by
W (λ − A−1)−1B1(I − GA,B1,K,KA,B1
(λ))−1GA,B2,K,KA,B2
(λ) + W (λ − A−1)−1B2.
Observe that the assumption implies that the strong limit
lim
λ→+∞(cid:18)W (λ − A−1)−1B1(I − GA,B1,K,KA,B1
W (λ − A−1)−1B1(I − GA,B1,K,KA,B1
= lim
λ→+∞
(λ))−1GA,B2,K,KA,B2
(λ)
(λ))−1GA,B2,K,KA,B2
(λ) + W (λ − A−1)−1B2(cid:19)
+ lim
λ→+∞
W (λ − A−1)−1B2
=(cid:18)W A,B1(I − K A,B1)−1K A,B2 + W A,B2(cid:19)
hold. Therefore the feedthrough operator of (3.12) is W A,B1(I − K A,B1)−1K A,B2 + W A,B2.
The proof is therefore completed.
Example 3.14 Consider the following boundary system governed by wave equations
wtt(x, t) = ∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = −
∂(A−1w)
∂ν
, x ∈ Γ1, t ≥ 0,
w(x, t) = u(x, t), x ∈ Γ0, t ≥ 0,
y(x, t) = −
∂(A−1w)
∂ν
, x ∈ Γ0, t ≥ 0
(3.13)
where Ω ⊂ Rn, n ≥ 2 is an open bounded region with smooth C 3-boundary ∂Ω = Γ0 ∪ Γ1.
Γ0, Γ1 are disjoint parts of the boundary relatively open in ∂Ω, int(Γ1) 6= ∅ and int(Γ0) 6= ∅,
ν is the unit normal vector of Γ0 pointing towards the exterior of Ω, u is the input function
(or control) and y is the output function (or output).
Let H = L2(Ω) × H −1(Ω) be the state space and U = L2(∂Γ0), V = L2(∂Γ1) be the
control (input) or observation (output) space. Guo and Zhang [9] proved that system
wtt(x, t) = ∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = v(x, t), x ∈ Γ1, t ≥ 0,
w(x, t) = 0, x ∈ Γ0, t ≥ 0,
y(x, t) = −
∂(A−1w)
∂ν
25
, x ∈ Γ1, t ≥ 0
is a regular linear system with feedthrough operator I and with admissible feedback operator
I. Moreover,
is a regular linear system. By the same procedure, one can verify that
and
wtt(x, t) = ∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = 0, x ∈ Γ1, t ≥ 0,
w(x, t) = u(x, t), x ∈ Γ0, t ≥ 0,
y(x, t) = −
∂(A−1w)
∂ν
, x ∈ Γ1, t ≥ 0
wtt(x, t) = ∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = v(x, t), x ∈ Γ1, t ≥ 0,
w(x, t) = 0, x ∈ Γ0, t ≥ 0,
y(x, t) = −
∂(A−1w)
∂ν
, x ∈ Γ0, t ≥ 0
wtt(x, t) = ∆w(x, t), x ∈ Ω, t > 0,
w(x, t) = 0, x ∈ Γ1, t ≥ 0,
w(x, t) = u(x, t), x ∈ Γ0, t ≥ 0,
y(x, t) = −
∂(A−1w)
∂ν
, x ∈ Γ1, t ≥ 0
are regular linear systems. Then we claim by Theorem 3.13 that system (3.13) is regular
with feedthrough operator I.
References
[1] K. Ammari, Dirichlet boundary stabilization of the wave equation, Asymptot. Anal.,
30 (2002), pp. 117-130.
[2] S.G. Chai, B.Z. Guo, Feedthrough operator for linear elasticity system with boundary
control and observation, SIAM Journal on Control and Optimization, 2010, 48(6):
3708-3734.
[3] G. Greiner, Perturbing the boundary conditions of a generator. Houston J. Math.
1987; 13: 213-229.
26
[4] B.Z. Guo, Y.H. Luo, Controllability and stability of a second-order hyperbolic system
with collocated sensor/actuator, Systems & Control Letters, 2002, 46(1): 45-65.
[5] B.Z. Guo, Z.C. Shao, Regularity of a Schrodinger equation with Dirichlet control and
collocated observation. Syst. Control Lett., 54, 1135-1142, 2005.
[6] B. Z. Guo, Z. C. Shao, Regularity of an Euler-Bernoulli plate equation with Neumann
control and collocated observation, J. Dyn. Control Syst., 12 (2006), pp. 405-418.
[7] B.Z. Guo, Z.C. Shao, Well-posedness and regularity for non-uniform Schrodinger and
Euler-Bernoulli equations with boundary control and observation, Quarterly of Ap-
plied Mathematics, 70, 111-132, 2012.
[8] B.Z. Guo, J.M. Wang, K.Y. Yang, Dynamic stabilization of an EulerCBernoulli beam
under boundary control and non-collocated observation, Systems & Control Letters,
2008, 57(9): 740-749.
[9] B.Z. Guo, X. Zhang, The regularity of the wave equation with partial Dirichlet control
and colocated observation, SIAM J. Control Optim., 44 (2005), pp. 1598-1613.
[10] B. Z. Guo, Z. X. Zhang, Well-posedness and regularity for an Euler-Bernoulli plate
with variable coefficients and boundary control and observation, Math. Control Signals
Systems, 19 (2007), pp. 337-360.
[11] B.Z. Guo, Z.X. Zhang, Well-posedness of systems of linear elasticity with Dirichlet
boundary control and observation, SIAM J. Control Optim., 48 (2009), pp. 2139-2167.
[12] B. Haak, P.C. Kunstmann, Admissibility of unbounded operators and wellposedness
of linear systems in Banach spaces, Integral Equ. Oper. Theory 55 (4) (2006) 497-533.
[13] B.H. Haak, E.M. Ouhabaz, Exact observability, square functions and spectral the-
ory[J]. Journal of Functional Analysis, 2012, 262(6): 2903-2927.
[14] S. Hadd, Exact controllability of infinite dimensional systems persists under small
perturbations, J. Evolution Equations, 5 (2005) 545-555.
27
[15] S. Hadd, A. Idrissi, Regular linear systems governed by systems with state, input
and output delays, IMA J. Math. Control Inform., vol. 22, no. 4, pp. 423-439, 2005.
[16] S. Hadd, A. Idrissi, On the admissibility of observation for perturbed C0-semigroups
on Banach spaces. Systems Control Letter. 2006; 55: 1-7.
[17] S. Hadd, A. Idrissi, A. Rhandi, The regular linear systems associated to the shift
semigroups and application to control delay systems, Math. Control Signals Systems.
2006; 18: 272-291.
[18] B. Jacob, Exact observability of diagonal systems with a finite-dimensional output
operator, Systems Control Lett., 43 (2001) 101-109.
[19] B. Jacob, J.R. Partington, The Weiss conjecture on admissibility of observation op-
erators for contraction semigroups, Integral Equ. Oper. Theory, 40 (2) (2001) 231-241.
[20] B. Jacob, H. Zwart, On the Hautus test for exponentially stable C0-Groups. SIAM
journal on control and optimization, 2009, 48(3): 1275-1288.
[21] A.N. Kolmogorov, S.V. Fomin, Elementi di teoria delle funzioni e di analisi funzionale,
Mir, Moscow, 1980.
[22] I. Lasiecka, Exponential decay rates for the solutions of Euler-Bernoulli equations
with boundary dissipation occurring in the moments only, J. Differential Equations
95 (1992) 169-182.
[23] I. Lasiecka, R. Triggiani, Optimal regularity, exact controllability and uniform stabi-
lization of Shrodinger equations with Dirichlet control, Differential Integral Equations,
5 (1992) 521-535.
[24] I. Lasiecka, R. Triggiani, L2(Σ)-regularity of the boundary to boundary operator
B ∗ L for hyperbolic and Petrowski PDEs. Abstr. Appl. Anal. 19 (2003), 1061-1139.
[25] Malinen J., Staffans OJ. Conservative boundary control systems. J. Differential Equa-
tions. 2006; 231: 290-312.
28
[26] Z.D. Mei, J.G. Peng, On robustness of exact controllability and exact observabil-
ity under cross perturbations of the generator in Banach spaces. Proceedings of the
American Mathematical Society. 2010, 138:4455-4468.
[27] Z.D. Mei, J.G. Peng, On the perturbations of regular linear systems and linear sys-
tems with state and output delays. Integr. Equ. Oper. Theory., 68 (2010), 357-381.
[28] Z.D. Mei, J.G. Peng, On Invariance of p-Admissibility of Control and Observation
Operators to q-Type of Perturbations of Generator of C0-Semigroup, Systems & Con-
trol Letters, 59 (2010) 470-475.
[29] Z.D. Mei, J.G. Peng, Robustness of Exact p-Controllability and Exact p-Observability
to q-Type of Perturbations of the Generator, Asian Journal of Control, 2014, 16(4):
1164-1168.
[30] Z.D. Mei, J.G. Peng, A Class of Linear Boundary Systems with Delays in State,
Input and Boundary Output, arXiv preprint arXiv:1510.00705, 2015.
[31] J.R. Partington, S. Pott, Admissibility and exact observability of observation oper-
ators for semigroups, Irish Math. Soc. Bull., 55 (2005) 19-39.
[32] R. Rebarber, G. Weiss, Necessary conditions for exact controllability with a finite-
dimensional input space, Systems Control Lett., 40 (2000) 217-227.
[33] D.L. Russell, G. Weiss, A general necessary condition for exact observability, SIAM
J. Control Optim., 32 (1994) 1-23.
[34] D. Salamon, Infinite-dimensional linear system with unbounded control and observa-
tion: a functional analytic approach. Transactions on American Mathematical Society.
1987; 300: 383-431.
[35] D. Salamon, Realization theory in Hilbert space, Math. Systems Theory 21 (1989),
147-164.
29
[36] O.J. Staffans, G. Weiss, Transfer functions of regular linear systems. Part II: the
system operator and the Lax-Phillips semigroup. Trans. Amer. Math. Soc., 354 (2002)
3229-3262.
[37] O.J. Staffans, Well-Posed Linear Systems. Cambridge, U.K.: Cambridge Univ. Press,
2005.
[38] M. Tucsnak, G. Weiss, Observation and Control
for Operators Semigroups.
Birkhauser Verlag, Basel, 2009.
[39] G. Weiss, Admissibility of unbounded control operators. SIAM J. Control Optimiza-
tion. 1989; 27: 527-545.
[40] G. Weiss, Admissible observation operators for linear semigroups. Israel J. Mathe-
matics. 1989; 65: 17-43.
[41] G. Weiss, The representation of regular linear systems on Hilbert spaces. In: Kappel
F, Kunisch K, Schappacher W (eds) Control and estimation of distributed parameter
systems (Proceedings Vorau 1988), Birkhauser, pp 401-16.
[42] G. Weiss, Transfer functions of regular linear systems Part I: characterizations of
regularity. Transactions of the American Mathematical Society. 1994; 342(2): 827-
854.
[43] G. Weiss, Regular linear systems with feedback. Math. Control Signals Systems.
1994; 7: 23-57.
[44] G. Weiss, R. Rebarber, Optimizability and estimatability for infinite-dimensional
linear systems. SIAM Journal on Control and Optimization, 2000, 39(4): 1204-1232.
[45] C.I. Byrnes, D.S. Gilliam, V.I. Shubov, et al., Regular linear systems governed by a
boundary controlled heat equation, Journal of Dynamical and Control Systems, 2002,
8(3): 341-370.
[46] G.Q. Xu, C. Liu, S.P. Yung, Necessary conditions for the exact observability of
systems on Hilbert space, Systems Control Lett., 57 (3) (2008) 222-227.
30
[47] H. Zwart, B. Jacob, O. Staffans, Weak admissibility does not imply admissibility for
analytic semigroups, System Control Lett. 48 (3) (2003) 341-350.
[48] H. Zwart, Sufficient conditions for admissibility, Systems Control Letter, 54 (2005)
973-979.
31
|
1501.05869 | 2 | 1501 | 2016-05-19T20:50:18 | A Spectral Characterization of $\mathcal{AN}$ Operators | [
"math.FA"
] | We establish a spectral characterization theorem for the operators on complex Hilbert spaces of arbitrary dimensions that attain their norm on every closed subspace. The class of these operators is not closed under addition. Nevertheless, we prove that the intersection of these operators with the positive operators forms a proper cone in the real Banach space of hermitian operators. | math.FA | math |
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
SATISH K. PANDEY AND VERN I. PAULSEN
Abstract. We establish a spectral characterization theorem for the operators
on complex Hilbert spaces of arbitrary dimensions that attain their norm on
every closed subspace. The class of these operators is not closed under addition.
Nevertheless, we prove that the intersection of these operators with the positive
operators forms a proper cone in the real Banach space of hermitian operators.
1. Introduction
Throughout this paper H and K will denote complex Hilbert spaces and we write
B(H,K) for the set of all bounded, linear operators from H to K. We recall that
B(H,K) is a complex Banach space with respect to the operator norm
kTk = sup{kT xk : x ∈ H,kxk 6 1}.
Definition 1.1. An operator T ∈ B(H,K) is said to be an N operator or to
satisfy the property N if there is an element x in the unit sphere of H such that
kTk = kT xk.
Such operators achieve their norm and hence are known as norming operators.
A generalization of the property N leads to a new class of operators in B(H,K).
Definition 1.2. An operator T ∈ B(H,K) is said to be an AN operator or to
satisfy the property AN , if for every nontrivial closed subspace M of H, TM
satisfies the property N .
Alternatively, an operator T ∈ B(H,K) is said to be an AN operator if for ev-
ery nontrivial closed subspace M of H there is an element x ∈ M,kxk = 1 such
that kTMk = kTM(x)k. Since these operators, when restricted to any nontrivial
closed subspace of H, achieve their norm on that closed subspace, we say that these
operators are absolutely norming and hence the name AN operator. Needless to
say, every AN operator is an N operator.
The AN operators were introduced and studied in [1] and [4]. Carvajal and
Neves [1] proved a partial structure theorem [1, Theorem 3.25] for the class of
positive AN operators on complex Hilbert spaces that included an uncharacterized
"remainder" operator. This theorem motivated Ramesh [4] to attempt to obtain a
full characterization theorem [4, Theorem 2.3], without remainder, for positive AN
operators on separable complex Hilbert spaces.
In this paper, we present a counterexample to Ramesh's characterization theorem
[4, Theorem 2.3]. We then give a full spectral characterization of the class of positive
Date: January 5, 2015.
2010 Mathematics Subject Classification. Primary 47B07, 47A10, 47A75; Secondary 47L07,
47L25, 47B65, 46L05.
1
2
S. K. PANDEY AND V. I. PAULSEN
AN operators on complex Hilbert spaces of arbitrary dimensions, earlier results
needed to assume separability. The correct characterization requires more terms
than were used in [4] and [1]. Using this theorem, we prove a full characterization
theorem for the class of AN operators on complex Hilbert spaces of arbitrary
dimension.
We begin by giving noninductive proofs of some basic facts, which allows us to
remove the separability assumption. In section 3 and 4, we derive necessary and
sufficient conditions for a positive operator that satisfies the AN condition and con-
sequently establish a spectral characterization theorem (see 5.1) for these operators
in section 5. This theorem, together with the polar decomposition theorem, then
paves the way for our main result in section 6: the full spectral characterization
of the class of AN operators (see 6.4). The class of these operators is not closed
under addition. Nevertheless, we prove that the class of positive AN operators is
a proper cone in the real Banach space of hermitian operators.
We end this section by presenting a counterexample to [4, Theorem 2.3].
Example 1.3. Consider the operator
1
2
1
0
1
. . .
0
1
T =
∈ B(l2).
1
That T is positive operator on a separable Hilbert space is obvious. T is not
compact. The infimum of the eigenvalues of this operator, Ramesh's m(T ) = 1/2.
The operator T − m(T )I = diag(0, 1/2, 1/2, ...) is not compact. Consequently, T
is neither compact nor of the form K + m(T )I for some positive compact operator
K. Even more, there does not exist α ≥ 0 such that T = K + αI for some positive
compact operator K. Thus, if [4, Theorem 2.3] was correct, then T would not
satisfy AN .
However, we now prove that T satisfies the property AN . Suppose that M is
an arbitrary nontrivial closed subspace of H. If M is one dimensional, then TM
attains its norm at any vector in M with unit norm. If dim(M) ≥ 2 and M contains
two noncollinear vectors which are nonzero in the first entry, then there exists a
linear combination of these two vectors with 0 in the first entry. Letting x0 be the
normalization of this vector, we get 1 = kx0k = kT (x0)k ≤ kTMk ≤ kTk = 1 and
so we have equality throughout and T attains its norm on M.
Finally, if dim(M) ≥ 2 and it does not contain any two such vectors, then it
either has a single such vector and its scalar multiples or no such vector. Since
dim(M) ≥ 2, M has at least one vector linearly independent from all vectors with
nonzero first entry and that vector must have 0 in its first entry. If we normalize this
vector -- we call this vector x0 -- we get 1 = kx0k = kT (x0)k ≤ kTMk ≤ kTk = 1
and hence T attains its norm on M. This proves the assertion and serves to be a
counterexample to the characterization Theorem 2.3 of [4].
While some of our results have parallels in [1] and [4], our proofs are quite
different, since we do not assume separability and avoid representing operators by
ordered series indexed by N.
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
3
2. Preliminaries
Proposition 2.1. If T ∈ B(H,K) is a compact operator, then T satisfies the
property AN .
Proof. If T is a compact operator from H to K then the restriction of T to any
closed subspace M is a compact operator from M to K. So it will be sufficient to
prove that if T is a compact operator then T satisfies N .
Let B = {x ∈ H : kxk 6 1} be the closed unit ball of H. Since T is a compact
operator, T (B) is a compact subset of K in the norm topology [3, page 55]. Also,
k · kK : T (B) −→ [0,∞) is a continuous function on T (B). Consequently we have
sup{kT xkK : kxkH 6 1} = max{kT xkK : kxkH 6 1}.
It therfore implies that
there exists x0 ∈ B such that kTk = kT x0kK. This, together with kT x0kK 6
kTkkx0kH 6 kTk, implies that kx0kH = 1. This proves the proposition.
Proposition 2.2. [1, Proposition 2.3] Let T ∈ B(H) be a self-adjoint operator.
Then T satisfies N iff kTk or −kTk is an eigenvalue of T .
(cid:3)
This result leads us to the following theorem.
Theorem 2.3. Let T ∈ B(H) be a positive operator. Then T satisfies N iff kTk
is an eigenvalue of T .
Theorem 2.4. Let H and K be complex Hilbert spaces and T ∈ B(H,K). Then T
satisfies N iff T ∗T satisfies N .
Proof. First assume that T satisfies N . There exists x in the unit sphere of H such
that kT xk = kTk. Then
kT ∗Tk = kTk2 = hT x, T xi = hT ∗T x, xi ≤ kT ∗T xk ≤ kT ∗Tk,
and so we have equality throughout which implies that T ∗T satisfies N .
Conversely, if T ∗T satisfies N , then by Theorem 2.3 kT ∗Tk is an eigenvalue
of T ∗T . Suppose y ∈ H is the corresponding eigenvector of unit norm. Then
kT yk2 = hT ∗T y, yi = hkT ∗Tky, yi = kTk2, and the result follows.
Let T ∈ B(H,K), recall that every positive operator has a unique positive square
root and that T := √T ∗T .
Theorem 2.5. Let H and K be complex Hilbert spaces and T ∈ B(H,K). Then
the following statements are equivalent.
(1) T satisfies N .
(2) T ∗ satisfies N .
(3) kTk is an eigenvalue of T.
(4) kTk is an eigenvalue of T.
(5) T satisfies N .
(6) T ∗ satisfies N .
(7) T2 satisfies N .
(8) T ∗2 satisfies N .
(9) kTk is an eigenvalue of T ∗.
Proof. The equivalence of (1) and (7) follows from Theorem 2.4 as does the equiv-
alence of (5) and (7). Since kTk = kTk, by Theorem 2.3, (5) is equivalent to (3)
and (4).
(cid:3)
4
S. K. PANDEY AND V. I. PAULSEN
equivalence of (2), (6), (8) and (9).
Replacing T by T ∗ in these equivalences and using that kTk = kT ∗k, shows the
All that remains is to show the equivalence of (1) and (2). Assume that T
satisfies N . By equivalence of (1) and (4), kTk is an eigenvalue of T. Let z ∈ H
be an eigenvector of T of unit norm corresponding to the eigenvalue kTk. Since
T(z) = kTkz we have T ∗T z = T2(z) = T(T(z)) = kTk2z. Consequently,
kT ∗(T z)k = kTk2 = kTkkT ∗k, since kzk = 1. Notice that T ( z
kT k ) is in the unit
kT k )k = kT ∗k which means that T ∗ satisfies N . This
sphere of K and hence kT ∗(T z
proves that (1) implies (2). The backward implication follows if we replace T by
T ∗ in the proof and use T ∗∗ = T. This completes the proof.
(cid:3)
Remark 2.6. Later (see 6.3) we will give an example of an operator such that T
is AN but T ∗ is not AN .
3. Necessary conditions for positive AN Operators
The purpose of this section is to study the properties of positive AN operators.
Theorem 3.1. Let T ∈ B(H) be a positive AN operator. Then H has an orthonor-
mal basis consisting of eigenvectors of T .
Proof. Let B = {vα : α ∈ Λ} be the maximal orthonormal set of eigenvectors
of T . That B is nonempty is a trivial observation; for T , being a positive AN
operator, must have kTk as its eigenvalue. Considering w to be a unit eigenvector
corresponding to the eigenvalue kTk, we have T w = kTkw which implies that
w ∈ B.
To show that H has an orthonormal basis consisting entirely of eigenvectors of
T we define H0 := clos(span(B)) and show that H0 = H. It suffices to show that
0 = {0}; for then H0 = H⊥⊥
H⊥
We first claim that H⊥
0 is an invariant subspace of H under T . To see this, let
F denote the collection of finite subsets of Λ, that is, F = {F ⊆ Λ : F is finite}. If
v ∈ H0, then
0 = {0}⊥ = H.
v = Xα∈Λ
hv, vαi vα = lim
hv, vαi vα.
F ∈F Xα∈F
Since the above limit is the norm limit and T is bounded (norm continuous), it
follows that
F ∈F Xα∈F
hv, vαi βαvα ∈ H0,
hv, vαi βαvα = Xα∈Λ
hv, vαi vα(cid:17) = lim
F ∈F Xα∈F
T v = T(cid:16) lim
considering T vα = βαvα where βα ∈ C for every α ∈ Λ. This shows that H0 is
an invariant subspace of H under T . Since T = T ∗, we infer that H⊥
0 is also an
invariant subspace of H under T .
We complete the proof by showing H⊥
0 6= {0}, that is, H⊥
H⊥
operator, TH⊥
on H⊥
is an eigenvalue of TH⊥
the eigenvalue kTH⊥
kTH⊥
0 = {0}. Suppose, on the contrary, that
0 is a nontrivial closed subspace of H. Since T is a positive AN
is a positive operator
0 k
0 is invariant under T . Consequently, kTH⊥
corresponding to
(z) =
0 kz. But this means that
0 such that kzk = 1 and TH⊥
(z) = kTH⊥
satisfies the property N . Even more, TH⊥
. Let z be a unit eigenvector of TH⊥
0
0 kz, which implies that T z = TH⊥
0
0 which satisfies N because H⊥
0
0 k. Clearly then z ∈ H⊥
0
0
0
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
5
z /∈ H0 is an eigenvector of T which contradicts the maximality of the set B = {vα :
α ∈ Λ} of T and we conclude that H⊥
0 = {0}. This completes the proof.
(cid:3)
Let H, K be Hilbert spaces with w ∈ H and v ∈ K. v ⊗ w then denotes the
operator from H to K defined as: (v ⊗ w)x = hx, wi v for every x ∈ H.
Corollary 3.2. If T ∈ B(H) is a positive AN operator, then
T = Xα∈Λ
βαvα ⊗ vα,
where {vα : α ∈ Λ} is an orthonormal basis consisting entirely of eigenvectors of
T and for every α ∈ Λ, T vα = βαvα with βα > 0. Moreover, for every nonempty
subset Γ ⊆ Λ of Λ, we have sup{βα : α ∈ Γ} = max{βα : α ∈ Γ}.
Proof. That T = Pα∈Λ βαvα ⊗ vα is obvious. To prove the final claim we use
the method of contradiction and assume on the contrary that sup{βα : α ∈ Γ} 6=
max{βα : α ∈ Γ} for some nonempty subset Γ ⊆ Λ, that is, the supremum of the
set {βα : α ∈ Γ} (say β) is not achieved. In that case, for any x ∈ HΓ with kxk = 1,
kTHΓ(x)k2 = Xα∈Γ
βα2hx, vαi2 < Xα∈Γ
β2hx, vαi2 = β2 = kTHΓk2.
This implies that kTHΓ(x)k < kTHΓk for every x ∈ HΓ with kxk = 1 which
contradicts the fact that T is an AN operator. This proves the assertion.
(cid:3)
The spectral conditions given in the above corollary do not characterize AN
operators as the following example and result show.
Example 3.3. Let K1, K2 be positive compact operators that are not of finite rank
on the complex Hilbert space l2, and 0 6 a < b. Consider the operator
T =(cid:20)aI + K1
0
0
bI + K2(cid:21) ∈ B(l2 ⊕ l2).
Then the supremum of each subset of the spectrum is equal to the maximum of
that subset since the spectrum of T consists of the closure of the union of two
decreasing sequences, {an} and {bn}, with limn an = a and limn bn = b. However,
the spectrum of T has two limit points, and so by the following result T is not AN .
Thus, the spectral condition given by the above corollary does not characterize
positive AN operators.
Proposition 3.4. If T ∈ B(H) is a positive AN operator, then the spectrum σ(T )
of T has at most one limit point. Moreover, this unique limit point (if it exists) can
only be the limit of a decreasing sequence in the spectrum.
Proof. By the Corollary 3.2, we know that
T = Xα∈Λ
βαvα ⊗ vα,
where {vα : α ∈ Λ} is an orthonormal basis consisting entirely of eigenvectors of T
and for every α ∈ Λ, T vα = βαvα with βα > 0. All that remains is to show that
the spectrum σ(T ), which is precisely the closure of {βα}α∈Λ has at most one limit
point and this unique limit point (if it exists) can only be the limit of a decreasing
sequence in the spectrum.
First we show that whenever λ is a limit point of the spectrum σ(T ) of T , then
there exists a decreasing sequence (λn)n∈N ⊆ {βα : α ∈ Λ} such that λn ց λ. To
6
S. K. PANDEY AND V. I. PAULSEN
see this, it is sufficient to prove that there are at most only finitely many terms of the
sequence of (λn)n∈N that are strictly less than λ; for if there are infinitely many such
terms, then there exists an increasing subsequence (λnk ) such that λnk ր λ and for
each nk ∈ N, λnk < λ. But then if we define M0 := clos(span{vnk}), where vnk 's are
the eigenvectors corresponding to the eigenvalues λnk , then it is a trivial observation
that kTM0k = sup{λnk} = λ. However, for every x = Pnk αnk vnk ∈ M0 with
Pnk αnk2 = 1 so that kxk = 1,
kTM0(x)k2 = kXnk
αnk λnk vnkk2 =Xnk
αnk2λnk2 < λ2Xnk
αnk2 = λ2,
so that kTM0(x)k < λ 6 kTM0k. This contradicts the fact that T is an AN
operator. This proves our first claim.
We next prove, by the method of contradiction, that the spectrum σ(T ) of T
has at most one limit point. Suppose on the contrary that the spectrum σ(T ) =
clos({βα}α∈Λ) has two limit points a < b. By the discussion in the above paragraph,
there exist decreasing sequences (an)n∈N ⊆ {βα}α∈Λ and (bn)n∈N ⊆ {βα}α∈Λ such
that an ց a and bn ց b. Let us rename and denote by {fn} and {gn} the eigen-
vectors corresponding to the eigenvalues {an} and {bn} respectively. Without any
loss of generality we may assume that a1 < b so that an < bn for each n ∈ N. (For
if it happens otherwise then we can choose a natural number m such that am < b
and redefine the sequence (an)∞
n=1.) Also note that T fn = anfn and
T gn = bngn for each n ∈ N. Define
n=m by (an)∞
M := clos(cid:16)spanncnfn +p1 − c2
ngn : n ∈ No(cid:17) ,
where c2
n ∈ [0, 1] are yet to be determined. Needless to say that M is a closed
subspace of H and hence a Hilbert space in its own right. Moreover, it is a trivial
observation that the set {en : n ∈ N}, where en := cnfn +p1 − c2
ngn serves to be
an orthonormal basis of M. Then,
kTMk2 > sup{kT enk2} = sup{kT (cnfn +p1 − c2
ngn)k2 : n ∈ N}
= sup{kcnanfn +p1 − c2
na2
nbngnk2 : n ∈ N} = sup{c2
n + (1 − c2
n : n ∈ N}.
n)b2
At this point we define a sequence (γn)n∈N by
a1 − b
2n
γn := b +
; n ∈ N.
Then, (γn)n∈N is a strictly increasing sequence such that for every n ∈ N, a2
n <
n)b2
b2 and limn→∞ γn = sup{γn : n ∈ N} = b. Notice that c2
na2
n + (1− c2
n is a convex
n)b2
n + (1 − c2
na2
combination of a2
n ∈ [a2
n, and hence it follows that c2
n] for
each n ∈ N. In fact, by choosing the right value of c2
n + (1 − c2
na2
n ∈ [0, 1], c2
n can
give any point in the interval [a2
n]. Let us then choose a sequence (cn)n∈N such
that c2
n, b2
n)b2
n and b2
1 < γ 2
n, b2
n)b2
n + (1 − c2
na2
kTMk2 > sup{c2
n = γ 2
na2
n. This yields
n)b2
n + (1 − c2
n : n ∈ N} = sup{γ 2
n : n ∈ N} = b2.
However, any x ∈ M with kxk = 1 can be written as
∞
∞
Xn=1
αn(cnfn +p1 − c2
ngn), with
Xn=1αn2 = 1,
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
7
in which case,
kTM(x)k2 = kT xk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xn=1
=
∞
T ∞
Xn=1
n + (1 − c2
αn(cnfn +p1 − c2
Xn=1
n)b2
n) =
na2
∞
αn2(c2
ngn)!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
αn2γ 2
2
n <
∞
Xn=1
αn2b2 = b2.
This implies that for every element x ∈ M with kxk = 1, kTM(x)k < b 6 kTMk
which means that T does not satisfy the property AN . So we arrive at a contra-
diction. Hence, our hypothesis is wrong and we conclude that the spectrum of T
can have at most one limit point. This completes the proof.
(cid:3)
We now use this as a tool to prove the following result.
Corollary 3.5. If T ∈ B(H) is a positive AN operator, then the set {βα}α∈Λ of
distinct eigenvalues of T , that is, without counting multiplicities, is countable.
Proof. This corollary is a direct consequence of the fact that If E ⊆ R is an un-
countable subset, then E has at least two limit points. Since the set {βα}α∈Λ has at
most one limit point, by the contrapositive of the above fact, it is countable.
(cid:3)
Corollary 3.6. If T ∈ B(H) is a positive AN operator, then the set {βα}α∈Λ of
eigenvalues of T has at most one eigenvalue with infinite multiplicity.
Proof. To show that this set has at most one eigenvalue with infinite multiplicity,
we assume that it has two distinct eigenvalues β1 and β2 with infinite multiplicity,
and we deduce a contradiction from the assumption. Without loss of generality, we
assume that 0 ≤ β1 < β2. Now let (an)n∈N ⊆ {βα}α∈Λ and (bn)n∈N ⊆ {βα}α∈Λ be
two sequences such that for every n ∈ N, we have an = β1 and bn = β2. Clearly
then an −→ β1 and bn −→ β2. Let us, like in the previous proof, rename and denote
by {fn} and {gn} the eigenvectors corresponding to the eigenvalues {an} and {bn}
respectively where T fn = anfn = β1fn and T gn = bngn = β2gn for each n ∈ N. At
this point we define a sequence (γn)n∈N by
β1 − β2
2n
nβ2
n)β2
1 < γ 2
n < β2
γn := β2 +
1 + (1 − c2
; n ∈ N.
That (γn)n∈N is a strictly increasing sequence with β2
2 for every n ∈ N
such that limn→∞ γn = sup{γn : n ∈ N} = β2 is obvious. Let c2
n ∈ [0, 1] be
arbitrary, then since c2
1 and
n)β2
2, it follows that for each n ∈ N, we have c2
β2
1 , β2
2 ∈ [β2
2]. In fact,
by choosing the right value of c2
nβ2
2 gives any desired point
in the interval [β2
n < β2
for every n ∈ N, allows us to define the sequence (cn)n∈N concretely, which is as
follows:
for each n ∈ N, choose cn so that c2
We will use this so defined sequence (cn)n∈N as a tool to define a closed subspace
M of H by
2]. This observation, together with the fact that β2
2 is a convex linear combination of β2
n ∈ [0, 1], c2
1 + (1 − c2
n)β2
nβ2
1 + (1 − c2
1 + (1 − c2
2 = γ 2
n.
1 < γ 2
n)β2
1 , β2
nβ2
2
M := clos(cid:16)spanncnfn +p1 − c2
ngn : n ∈ No(cid:17) .
8
S. K. PANDEY AND V. I. PAULSEN
ngn. It now follows that
It is easy to see that the set {en : n ∈ N} serves to be an orthonormal basis of M,
where en := cnfn +p1 − c2
kTMk2 > sup{kT enk2} = sup{kcnβ1fn +p1 − c2
= sup{c2
However, any x ∈ M with kxk = 1 can be written as
Xn=1
2 : n ∈ N} = sup{γ 2
nβ2gnk2 : n ∈ N}
n : n ∈ N} = β2
2.
1 + (1 − c2
αn2 = 1.
Xn=1
n)β2
nβ2
∞
∞
In that case,
ngn) with
T ∞
Xn=1
αn(cnfn +p1 − c2
kTM(x)k2 = kT xk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xn=1
αn(cnβ1fn +p1 − c2
Xn=1
1 + (1 − c2
αn2(c2
Xn=1
Xn=1
αn2β2
αn2γ 2
nβ2
n <
=
=
∞
∞
∞
∞
2
ngn)!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2
αn(cnfn +p1 − c2
nβ2gn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n)β2
2 )
2 = β2
2.
This implies that for every element x ∈ M with kxk = 1, kTM(x)k < β2 6
kTMk which means that T does not satisfy the property AN . So we arrive at a
contradiction. Hence, our hypothesis was wrong and we conclude that the spectrum
of T can have at most one eigenvalue with infinite multiplicity. This completes the
proof.
(cid:3)
Corollary 3.7. Let T ∈ B(H) be a positive AN operator. If the spectrum σ(T ) =
clos{βα : α ∈ Λ} of T has both a limit point β and an eigenvalue β with infinite
multiplicity, then β = β.
Proof. To show that β = β, we assume that β 6= β, and we deduce a contradiction
from the assumption. We first consider the case when β < β. Because β is a limit
point of the spectrum, we know that there exists a decreasing sequence (an)n∈N ⊆
{βα}α∈Λ such that an ց β. Let (bn)n∈N ⊆ {βα}α∈Λ be the constant sequence
whose each term is β so that bn −→ β. Without any loss of generality we may
assume that a1 < β so that an < bn for each n ∈ N. Next we rename and denote
by {fn} and {gn} the eigenvectors corresponding to the eigenvalues {an} and {bn}
respectively where T fn = anfn and T gn = bngn = βgn for each n ∈ N.
As we did in the previous proof, we define a sequence (γn)n ∈ N by
γn := β +
β − β
2n
; n ∈ N.
n < β2 for
Observe that (γn)n∈N is a strictly increasing sequence with a2
every n ∈ N. It immediately follows then that limn→∞ γn = sup{γn : n ∈ N} = β.
n) β = γ 2
Thereafter for each n ∈ N, we choose cn so that c2
n.
n ∈ [0, 1] and c2
n+(1−c2
n < γ 2
na2
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
9
Finally, with the help of this sequence (cn)n∈N let us define a closed subspace M
of H by
M := clos(cid:16)spanncnfn +p1 − c2
ngn : n ∈ No(cid:17) .
We know that the set {en : n ∈ N}, where en := cnfn +p1 − c2
mal basis of M. It now follows, like the argument in the previous proof, that
kTMk2 > sup{kT enk2} = sup{kcnanfn +p1 − c2
Since each x ∈ M with kxk = 1 can be written as
n) β2 : n ∈ N} = sup{γ 2
βgnk2 : n ∈ N}
n + (1 − c2
= sup{c2
na2
n
n : n ∈ N} = β2.
ngn, is an orthonor-
∞
∞
αn2 = 1, it follows that
2
ngn)!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xn=1
ngn) with
T ∞
Xn=1
Xn=1
αn(cnfn +p1 − c2
kTM(x)k2 = kT xk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xn=1
αn(cnanfn +p1 − c2
Xn=1
αn2(c2
n + (1 − c2
Xn=1
Xn=1
αn2γ 2
na2
n <
=
=
∞
∞
∞
∞
2
αn(cnfn +p1 − c2
βgn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
n) β2)
n
αn2 β2 = β2.
This implies that for every element x ∈ M with kxk = 1, kTM(x)k < β 6
kTMk which means that T does not satisfy the property AN . So we arrive at a
contradiction. Hence, our hypothesis was wrong and we conclude that β = β.
To prove the assertion for the case when β < β, we follow the same line of
argument. Let (an)n∈N ⊆ {βα}α∈Λ be the decreasing sequence such that an ց β,
(bn)n∈N ⊆ {βα}α∈Λ be the constant sequence whose each term is β so that bn −→ β,
and rename and denote by {fn} and {gn} the eigenvectors corresponding to the
eigenvalues {an} and {bn} respectively where T fn = anfn and T gn = bngn = βgn
for each n ∈ N. We define the sequence (γn)n ∈ N a bit differently by
γn := β +
β − β
2n
; n ∈ N.
n < a2
It is now a trivial observation that (γn)n∈N is a strictly increasing sequence with
β2 < γ 2
n for every n ∈ N. Consequently, limn→∞ γn = sup{γn : n ∈ N} = β.
n)a2
Thereafter for each n ∈ N, we choose cn so that c2
n =
n. Finally, with the help of this sequence (cn)n∈N, we define a closed subspace M
γ 2
of H by
n ∈ [0, 1] and c2
β2+(1−c2
n
M := clos(cid:16)spanncngn +p1 − c2
nfn : n ∈ No(cid:17) .
10
S. K. PANDEY AND V. I. PAULSEN
That the set {en : n ∈ N}, where en := cngn +p1 − c2
of M can be easily verified. It now follows that
kT Mk2 > sup{kT enk2} = sup{kcn βgn +p1 − c2
Since each x ∈ M with kxk = 1 can be written as
β2 + (1 − c2
= sup{c2
n)a2
n
ngfn, is an orthonormal basis
nanfnk2 : n ∈ N}
n : n ∈ N} = sup{γ 2
n : n ∈ N} = β2.
nfn) with
αn2 = 1, it follows that
∞
Xn=1
αn(cngn +p1 − c2
kT M(x)k2 = kT xk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
=
∞
αn2(c2
∞
Xn=1
Xn=1
=
αn2γ 2
n <
∞
Xn=1
T ∞
Xn=1
αn(cngn +p1 − c2
β2 + (1 − c2
Xn=1
αn2β2 = β2.
n)a2
n)
∞
n
2
nfn)!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
This implies that for every element x ∈ M with kxk = 1, kT M(x)k < β 6 kT Mk
which contradicts the fact that T satisfies the property AN . Thus, we conclude
that β = β. This completes the proof.
(cid:3)
We finish this section by stating the final proposition in its full strength.
Theorem 3.8. If T ∈ B(H) is a positive AN operator, then
T = Xα∈Λ
βαvα ⊗ vα,
where {vα : α ∈ Λ} is an orthonormal basis consisting entirely of eigenvectors of T
and for every α ∈ Λ, T vα = βαvα with βα > 0 such that
(i) for every nonempty subset Γ ⊆ Λ of Λ, we have sup{βα : α ∈ Γ} = max{βα :
α ∈ Γ};
(ii) the spectrum σ(T ) = clos {βα : α ∈ Λ} of T has at most one limit point.
Moreover, this unique limit point (if it exists) can only be the limit of a decreasing
sequence in the spectrum;
(iii) the set {βα}α∈Λ of eigenvalues of T , without counting multiplicities, is
countable and has at most one eigenvalue with infinite multiplicity;
(iv) if the spectrum σ(T ) = clos{βα : α ∈ Λ} of T has both, a limit point β and
an eigenvalue β with infinite multiplicity, then β = β.
4. Sufficient Conditions For AN Operators
We now discuss the sufficient conditions for an operator (not necessarily positive)
to satisfy the AN condition. There is an important and useful criterion for an
operator T ∈ B(H,K) to satisfy the property AN , which depends on the following
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
11
facts: For a closed linear subspace M of a complex Hilbert space H let VM : M −→
H be the inclusion map from M to H defined as VM(x) = x for each x ∈ M. It is
then a trivial observation that the adjoint V ∗
M : H −→ M of VM is the orthogonal
projection of H on M (viewed as a map from H onto M), that is, V ∗
M : H −→ M
such that
V ∗
M (y) =(y
0
if y ∈ M,
if y ∈ M ⊥.
The criterion referred to is the following: T satisfies the property AN iff for every
closed linear subspace M of H, T VM satisfies the property N . To prove this
assertion we first observe that for any given nontrivial closed subspace M of H,
kT VMk = kTMk; for
kT VMk2 = sup{kT VM(x)k2 : kxk ≤ 1, x ∈ M}
= sup{kT xk2 : kxk ≤ 1, x ∈ M} = kTMk2.
We next assume that T satisfies the property AN and prove the forward implica-
tion. Let M be an arbitrary nontrivial closed subspace of H. Clearly then there
exists x0 ∈ M with kx0k = 1 such that kTMk = kT x0k.
It follows then that
there exists x0 ∈ H such that kT VMk = kTMk = kT x0k = kT VM(x0)k. Since
M is arbitrary, it follows that T VM satisfies the property N . We complete the
proof by showing that T is an AN operator if T VM satisfies the property N for
every nontrivial closed subspace M of H. Since T VM is an N operator, there
exists xM ∈ H(depending on M) with kxMk = 1 and kT VMk = kT VM(xM)k.
This means that for every M, kTMk = kT VMk = kT VM(xM)k = kT (VMxM)k =
kT xMk = kTM(xM)k where xM ∈ M and kxMk = 1. This essentially guarantees
that for every M, TM achieves its norm on unit sphere and hence satisfies the
property N .
We can summarize the result of the above discussion in the following lemma:
Lemma 4.1. For a closed linear subspace M of a complex Hilbert space H let
VM : M −→ H be the inclusion map from M to H defined as VM(x) = x for each
x ∈ M. An operator T ∈ B(H,K) satisfies the property AN if and only if for every
nontrivial closed linear subspace M of H, T VM satisfies the property N .
The following application illustrates the power of this result.
Proposition 4.2. If T ∈ B(H,K) is an isometry, then T satisfies the property
AN .
Proof. That an isometry satisfies the property N is obvious; for the operator norm
of an isometry is 1 and it attains its norm on any vector of unit length. For a closed
linear subspace M of the Hilbert space H let VM : M −→ H be the inclusion map
from M to H defined as VM(x) = x for each x ∈ M. To prove the assertion, it
suffices to show that for every nonzero closed linear subspace M, T VM is an N
operator. But T VM ∈ B(M,K) is an isometry and hence satisfies the property
N .
Lemma 4.3. Let T ∈ B(H) be a diagonalizable operator on the complex Hilbert
space H and B = {vα : α ∈ Λ} be an orthonormal basis of H corresponding to
which T is diagonalizable. If T achieves its norm on the unit sphere of H, then it
achieves it norm on some v0 ∈ B. Alternatively, if T satisfies the property N , then
there exists v0 ∈ B such that kTk = kT v0k.
(cid:3)
12
S. K. PANDEY AND V. I. PAULSEN
Proof. Let {λα : α ∈ Λ} be the set of eigenvalues of T corresponding to the the
eigenvectors {vα : α ∈ Λ}. From [2, Problem 61], we know that kTk = sup{λα :
α ∈ Λ}, so it suffices to prove thatkTk = max{λα : α ∈ Λ}.
To this end, by the way of contradiction, we assume the negation of the above
claim. It implies that for every α ∈ Λ, we have λα < kTk. However, for every
x ∈ H with kxk = 1, we have T x =Pα∈Λ λα hx, vαi vα so that
kTk2hx, vαi2 = kTk2kxk2 = kTk2;
kT xk2 = Xα∈Λ
λα2hx, vαi 2 < Xα∈Λ
which is a contradiction of the fact that T satisfies the property N . This proves
the claim.
Lemma 4.4. Let F ∈ B(H) be a self-adjoint finite-rank operator and α ≥ 0. Then
αI + F satisfies the property N .
Proof. Let the range of F be k-dimensional. Since F is self-adjoint, there exists
an orthonormal basis B = {vλ : λ ∈ Λ} of H corresponding to which the matrix
MB(F ) is a diagonal matrix with k nonzero real diagonal entries, say {β1, β2, ..., βk}.
Clearly then, MB(αI + F ) is also a diagonal matrix and
(cid:3)
kαI + Fk = sup{α + β1,α + β2, ...,α + βk, α}
= max{α + β1,α + β2, ...,α + βk, α}.
It is then a trivial observation that there exists v0 ∈ B such that kαI + Fk =
k(αI + F )v0k. This proves that αI + F achieves its norm on the unit sphere and
hence is an N operator.
(cid:3)
This lemma leads to the following propostion.
Proposition 4.5. If F ∈ B(H) is a self-adjoint finite-rank operator and α ≥ 0,
then αI + F satisfies the property AN .
Proof. For a closed linear subspace M of the Hilbert space H let VM : M −→ H
be the inclusion map from M to H defined as VM(x) = x for each x ∈ M.
Let us then define T := αI + F so that we have T ∗ = αI + F and T ∗T =
(αI + F )2 = α2I + 2αF + F 2 = βI + F where β = α2 ≥ 0 and F = 2αF + F 2 is
another self-adjoint finite-rank operator. We observe that
T is AN ⇐⇒ for every closed subspace M of H, T VM is N
⇐⇒ for every closed subspace M of H, (T VM)∗(T VM) is N
M(T ∗T )VM is N
⇐⇒ for every closed subspace M of H, V ∗
M(βI + F )VM is N .
⇐⇒ for every closed subspace M of H, V ∗
So, it suffices to show that for every closed subspace M of H, V ∗
N . But V ∗
M(βI + F )VM : M −→ M is an operator on M and
M(βI + F )VM = V ∗
V ∗
MβIVM + V ∗
M
F VM = βIM + FM,
M(βI + F )VM is
M(βI + F )VM is sum of a nonnegative scalar multiple of
which implies that V ∗
identity and a self-adjoint finite-rank operator on a Hilbert space M which, by the
previous lemma, does satisfy the property N and thus proves our assertion.
Lemma 4.6. For any positive compact operator K ∈ B(H) and α ≥ 0, αI + K
satisfies the property N .
(cid:3)
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
13
Proof. That K satisfies the property N is obvious, for K is compact. The positivity
of K ascertains that there is an orthonormal basis B = {vλ : λ ∈ Λ} of H, consisting
entirely of eigenvectors of K, corresponding to which K is diagonalizable; this
fact , together with the lemma 4.3 implies that there exists v0 ∈ B such that
kKk = β0 = max {βλ : λ ∈ Λ} = kKv0k, where K(vλ) = βλvλ for each λ ∈ Λ.
Since α ≥ 0, it readily follows that
kαI + Kk = sup{α + βλ : λ ∈ Λ} = α + sup{βλ : λ ∈ Λ}
= α + max {βλ : λ ∈ Λ} = α + β0 = k(αI + K)(v0)k.
αI + K therefore achieves its norm on unit sphere for each α ≥ 0. This completes
the proof.
(cid:3)
This lemma is a special case of what the following proposition states.
Proposition 4.7. For any positive compact operator K ∈ B(H) and α ≥ 0, αI +K
is AN .
Proof. Let us define T := αI + K so that we have T ∗ = αI + K and T ∗T =
(αI + K)2 = α2I + 2αK + K 2 = βI + K where β = α2 ≥ 0 and K = 2αK + K 2 is
another positive compact operator.
T is AN ⇐⇒ for every closed subspace M of H, T VM is N
⇐⇒ for every closed subspace M of H, (T VM)∗(T VM) is N
⇐⇒ for every closed subspace M of H, V ∗
M(T ∗T )VM is N
M(βI + K)VM is N .
⇐⇒ for every closed subspace M of H, V ∗
So, it suffices to show that for every closed subspace M of H, V ∗
N . But V ∗
M(βI + K)VM : M −→ M is an operator on M and
M(βI + K)VM is
M(βI + K)VM = V ∗
V ∗
MβIVM + V ∗
M
KVM = βIM + KM,
M(βI + K)VM is sum of a nonnegative scalar multiple of
which implies that V ∗
Identity and a positive compact operator on a Hilbert space M which does satisfy
the property N and hence proves our assertion.
Lemma 4.8. Let K ∈ B(H) be a positive compact operator and F ∈ B(H) be
a self-adjoint finite-rank operator. Then K + F can have at most finitely many
negative eigenvalues.
(cid:3)
Proof. Since F is a self-adjoint finite-rank operator, there is an orthonormal basis
B of H consisting of eigenvectors of F corresponding to which it is diagonalizable.
This allows us to write F as the difference of two positive finite-rank operators,
F+ and F− so that F = F+ − F−. Consider the set of all eigenvectors in B
corresponding to which F− has nonzero (positive) eigenvalues. Needless to say
that they are finite in numbers. Define H− to be the span of these eigenvectors.
It is trivial to observe that H− is a closed finite-dimensional subspace of H and
H = H− ⊕ H ⊥
− . We assume that the dimension of H− is k, that is, dim H− = k.
We claim that the total number of negative eigenvalues of K + F does not exceed
k. To prove this claim, we first observe that K + F can now be rewritten as
K + (F+ − F−) = (K + F+)− F− = K − F− where K = K + F+ is positive compact
operator on H. Also, K − F− is a self-adjoint compact operator and thus there
14
S. K. PANDEY AND V. I. PAULSEN
αivi!,
exists an orthonormal basis B of H consisting entirely of eigenvectors of K − F−
corresponding to which K − F− is diagonalizable. We next observe that
for any x ∈ H ⊥
− , D( K − F−)x, xE ≥ 0;
− and D Kx, xE ≥ 0 for each x ∈ H and
because F−(x) = 0 for every x ∈ H ⊥
hence for each x ∈ H ⊥
− . We are now ready to prove our claim. Consider the set
of all orthonormal eigenvectors in B corresponding to which K − F− has negative
eigenvalues. By the way of contradiction let us assume that the cardinality of
this set is strictly bigger than k. We fix some m > k and extract m eigenvectors
from this set. Let the set of these extracted eigenvectors be {v1, v2, v3, ..., vm} and
the corresponding eigenvalues be {λ1, λ2, λ3, ..., λm}. Since m > k, there exists
α1, α2, ..., αm not all zero such that PH− (Pm
i=1 αivi) = 0. Then
m
m
m
h( K − F−) m
Xi=1
m
=
H ⊥
αiλivi,
αjvji
Xj=1
This observation leads us directly to the following proposition.
Xj=1
Xi=1
αjvji = h
Xi=1
αi2λi < 0.
But this contradicts the fact thatPm
i=1 αivi ∈ H ⊥
− ; for we established that for any x ∈
− ,D( K − F−)x, xE ≥ 0. This proves our claim.
Proposition 4.9. Let K ∈ B(H) be a positive compact operator and F ∈ B(H) be
a self-adjoint finite-rank operator. Then for every α ≥ 0, αI + K + F satisfies the
property N .
Proof. The assertion is trivial if α = 0; for then K + F is a compact operator which
satisfies the property N . We assume that α > 0. Notice that K + F is a self-adjoint
compact operator on H and thus there exists an orthonormal basis B of H consisting
entirely of eigenvectors of K + F corresponding to which it is diagonalizable. From
the previous lemma, K +F can have at most finitely many negative eigenvalues. Let
{λ1, λ2, ..., λn} be the set of all negative eigenvalues of K + F with {v1, v2, ..., vn}
as the corresponding eigenvectors in basis B; and let {µβ : β ∈ Λ} be the set of all
remaining nonnegative eigenvalues of K +F with {wβ : β ∈ Λ} as the corresponding
eigenvectors in B. We have B := {v1, v2, ..., vn} ∪ {wβ : β ∈ Λ} and the matrix
MB(K + F ) of K + F with respect to B is given by
(cid:3)
K + F =
. . .
λ1
λn
. . .
. . .
. . .
0
...
...
...
. . .
...
...
...
0
. . .
µβ
. . .
. . .
. . .
.
. . .
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
15
Observing the fact that
kK + Fk = max{{λi}n
i=1 ∪ {µβ}β∈Λ},
we proceed to show that αI + K + F satisfies property N . To accomplish this we
distinguish cases:
Case I. If µ β = max{{λi}n
kK + Fk = µ β = k(K + F )(w β )k. Clearly then
i=1 ∪ {µβ}β∈Λ} for some β ∈ Λ. Needless to say that
α + µ β ≥ α + λi ≥ α + λi for each i ∈ {1, 2, ..., n}, and
α + µ β ≥ α + µβ for each β ∈ Λ.
It is now easy to convince ourselves that if w β be the eigenvector corresponding to
the eigenvalue µ β then kαI + K + Fk = kα + µ βk = k(αI + K + F )(w β )k which
implies that αI + K + F achieves its norm at w β.
Case II. If λm = max{{λi}n
it is important to observe that
i=1∪{µβ}β∈Λ} for some m ∈ {1, 2, ..., n}. In this case
indeed the matrix MB(K + F ) can be written as
sup{µβ : β ∈ Λ} = max{µβ : β ∈ Λ};
λ1
. . .
. . .
. . .
λn
. . .
0
...
...
...
. . .
...
...
...
0
. . .
. . .
. . .
0
+
. . .
0
. . .
. . .
0
...
...
...
. . .
...
...
...
0
. . .
. . .
. . .
. . .
µβ
. . .
,
K+F =
where the first matrix is compact. Consequently the second matrix is forced to be
compact which implies that sup{µβ : β ∈ Λ} = max{µβ : β ∈ Λ}. Let max{µβ :
β ∈ Λ} = µ β for some β ∈ Λ. It is then a trivial observation that sup{{α +
λi}n
i=1 ∪ {α + µβ}β∈Λ} = max{α + µ β,α + λ1, ...,α + λn} which ascertains that
the operator αI + K + F satisfies the property N . We conclude the proof by a note
that αI + K + F need not necessarily be positive for the proof to work.
(cid:3)
This result is the key to the theorem that follows. The following result could
be deduced from [1, Theorem 3.23] but there are some gaps in their proof of [1,
Lemma 3.7] which is essential to their proof of [1, Theorem 3.23]; so we provide an
independent proof.
Theorem 4.10. Let K ∈ B(H) be a positive compact operator and F ∈ B(H) be
a self-adjoint finite-rank operator. Then for every α ≥ 0, αI + K + F satisfies the
property AN .
16
S. K. PANDEY AND V. I. PAULSEN
Proof. Let M be an arbitrary nonempty closed linear subspace of the Hilbert space
H and VM : M −→ H be the inclusion map from M to H defined as VM(x) = x
for each x ∈ M.
Let us then define T := αI + K + F so that we have T ∗ = αI + K + F and
T ∗T = (αI +K +F )2 = (α2I)+(2αK +K 2)+(2αF +F K +KF +F 2) = βI + K + F
where β = α2 ≥ 0, K = 2αK + K 2 and F = 2αF + F K + KF + F 2 are respectively
positive compact and self-adjoint finite-rank operators. Observe that
T VM is N ⇐⇒ (T VM)∗(T VM) is N
⇐⇒ V ∗
M(T ∗T )VM is N ⇐⇒ V ∗
M(βI + K + F )VM is N .
M(βI + K + F )VM : M −→ M is an operator on M and
M(βI + K + F )VM is N ; for then, since M is arbitrary, it
It suffices to show that V ∗
immediately follows from lemma 4.1 that T is an AN operator. To this end, notice
that V ∗
F VM = βIM + KM + FM,
M(βI + K + F )VM = V ∗
V ∗
M(βI + K + F )VM is sum of a nonnegative scalar multiple
which implies that V ∗
of Identity, a positive compact operator and a self-adjoint finite-rank operator on
a Hilbert space M which, by the preceding proposition, satisfies the property N .
This proves the assertion.
MβIVM + V ∗
M
KVM + V ∗
M
(cid:3)
Remark 4.11. It is desirable at this stage to make an important remark: the
sum of two AN operators need not necessarily be an AN operator. An example
[1, Section 2, Page 182] appears in [1] which establishes that the sum of two N
operators need not necessarily be an N operator. In fact, one can show that the
operators they consider are not just N operators but AN .
In what follows, we
give an example of an operator T ∈ H which is AN but 2Re(T ) is not, which also
implies that sum of two AN operators need not be AN .
Example 4.12. Let {ei}i∈N be the canonical orthonormal basis of the Hilbert
space ℓ2(N), a ∈ (0, 1], and (ai)i∈N, (bi)i∈N be two sequences of real numbers such
that
0 < a1 < a2 < ... < a,
ai ր a, and a2
i + b2
i = 1.
2
∞
∞
i=1 hx, eii ei which implies that
Let T ∈ B(ℓ2(N)) defined as T ei = λiei for each i ∈ N, where λi = ai + ibi. Then
T ∗ei = λiei. It is easy to observe that both T and T ∗ are isometries. Indeed, if
x ∈ ℓ2(N), then x =P∞
kT xk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
By Proposition 4.2, we infer that T and T ∗ are AN operators. We now show
that T + T ∗ is not an AN operator. Since every AN operator is an N operator,
it suffices to show that T + T ∗ is not an N operator. To this end, notice that
kT + T ∗k > sup{kT eik : i ∈ N} = sup{λi + λi : i ∈ N} = sup{2ai : i ∈ N} = 2a.
But for every x ∈ ℓ2(N) with kxk = 1, we have
λi + λi2hx, eii2 =
= kxk2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
Xi=1
hx, eii λiei(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
hx, eii λiei(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
2ai2hx, eii2 < 4a2.
k(T + T ∗)xk2 =
= kT ∗xk2.
2
∞
Xi=1
∞
Xi=1
Consequently, for every x ∈ ℓ2(N) of unit length k(T + T ∗)xk < 2a 6 kT + T ∗k
which implies that T + T ∗ does not satisfy the property N .
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
17
5. Spectral Characterization Of Positive AN Operators
The final theorem of the preceding section just established -- -- that for every
α ≥ 0, αI + K + F satisfies the AN property where K and F are respectively
positive compact and self-adjoint finite-rank operators -- -- is the stronger version
of the backward implication of our spectral theorem for positive AN operators. If
the operator αI + K + F is also positive then the implication can be reversed and
the two conditions are equivalent. This is what the next theorem states.
Theorem 5.1 (Spectral Theorem For positive AN Operators). Let H be a complex
Hilbert space of arbitrary dimension and let P be a positive operator on H. Then P
is an AN operator if and only if P is of the form P = αI + K + F , where α ≥ 0, K
is a positive compact operator and F is self-adjoint finite-rank operator.
Proof. It suffices to prove the forward implication. We assume that P ∈ B(H)
is a positive AN operator. Theorem 3.8 asserts that there exists an orthonormal
basis B = {vλ : λ ∈ Λ} consisting entirely of eigenvectors of P and for every
λ ∈ Λ, T vλ = βλvλ with βλ ≥ 0. A moment's thought will convince the reader
that there are four mutually exclusive and exhaustive set of possibilities for the
spectrum σ(P ) = clos{βλ : λ ∈ Λ} of P .
Case 1. σ(P ) has neither a limit point nor an eigenvalue with infinite multiplicity.
The index set Λ is then finite; for if it is not then the set {βλ : λ ∈ Λ} (count-
ing multiplicities) of eigenvalues is also infinite. Since each eigenvalue in this set
can have at most finite multiplicity, it is obvious then that the set {βλ : λ ∈ Λ}
(without counting multiplicities) of distinct eigenvalues of P is infinite. More inter-
estingly, {βλ : λ ∈ Λ} is bounded above by the operator norm of P and below by
0. Since every infinite bounded subset of real numbers has a limit point, we arrive
at a contradiction and hence Λ is finite. This forces the Hilbert space H to be
finite dimensional. In that case P boils down to a positive (and hence self-adjoint)
finite-rank operator and we can safely assume that P = αI + K + F with α = 0,
K = 0 and F the operator in question.
Case 2. σ(P ) has no limit point but has one eigenvalue with infinite multiplicity.
Let β0 ∈ {βλ : λ ∈ Λ} be the eigenvalue with infinite multiplicity. Then the
set Γ := Λ \ {λ ∈ Λ : βλ = β0} is finite; for if it is not, then the set {βλ :
λ ∈ Γ} (counting multiplicities) is also infinite which in turn implies that the set
{βλ : λ ∈ Γ} (without counting multiplicities) is infinite because each eigenvalue
in this set can have at most finite multiplicity. Since {βλ : λ ∈ Γ} is bounded
and every infinite bounded subset of real numbers has a limit point, we arrive at a
contradiction.This implies that Γ is finite. Observe that for an arbitrary x ∈ H, we
have x = Pλ∈Λ hx, vλi vλ = Pλ∈Γ hx, vλi vλ +Pλ∈Λ\Γ hx, vλi vλ so that for every
18
x ∈ H,
S. K. PANDEY AND V. I. PAULSEN
hx, vλi P (vλ)
hx, vλi vλ
P x = Xλ∈Γ
= Xλ∈Γ
= Xλ∈Γ
= Xλ∈Γ
= Xλ∈Γ
hx, vλi β0vλ
hx, vλi P (vλ) + Xλ∈Λ\Γ
hx, vλi βλvλ + Xλ∈Λ\Γ
(βλ − β0)hx, vλi vλ + β0Xλ∈Λ
(βλ − β0)(vλ ⊗ vλ)(x) + β0Ix
(βλ − β0)(vλ ⊗ vλ) + β0I! (x).
is a self-adjoint finite-rank operator. It then readily follows that P = αI + K + F ,
To conclude this case it suffices to observe that β0 ≥ 0 andPλ∈Γ(βλ − β0)(vλ ⊗ vλ)
where α = β0, K = 0 and F =Pλ∈Γ(βλ − β0)(vλ ⊗ vλ).
Case 3. σ(P ) has no eigenvalue with infinite multiplicity but has a limit point.
The index set Λ is then countable; for if it is uncountable then the set {βλ : λ ∈ Λ}
(counting multiplicities) is also uncountable thereby rendering the set {βλ : λ ∈ Λ}
(without counting multiplicities) uncountable since each eigenvalue in this set has
finite multiplicity. Then this uncountable set must have at least two limit points;
and since this is impossible, we infer that Λ is countable and hence H is separable.
Having shown that Λ is countable, we can safely replace Λ by N. This essentially
redefines the spectrum σ(P ) = clos{βn : n ∈ N} of P .
Now let β ∈ σ(P ) be the unique limit point in the spectrum. We wish to
reorder the elements of {βn : n ∈ N} linearly in accordance with their size. To
accomplish this, we first notice that there are at most only finitely many terms
of the set {βn : n ∈ N} that are strictly less than β -- -- represent this set of
finite elements by {β1, β2, ..., βk} counting multiplicities. We next consider the set
{βn : βn > β}n∈N of all terms that are strictly bigger than β. We then inductively
define a nonincreasing sequence (βk+m)m∈N as
βk+1 := max{βn : βn > β}n∈N,
βk+2 := max{βn : βn > β}n∈N \ {βk+1},
...
βk+m := max{βn : βn > β}n∈N,\{βk+1, ..., βk+m−1},
...
This decreasing sequence is bounded below by β, so it converges to β; for if it
converges to any other point -- -- which, in that case, happens to be a limit point of
σ(P ) -- -- then that contradicts the existence of only one limit point in the spectrum.
Before we go further, it is worth establishing that the set {βn : βn > β}n∈N of all
eigenvalues of P has been exhausted in the process of constructing the sequence
(βk+m)m∈N, that is, each eigenvalue of P that is strictly bigger than β is a term
of the sequence (βk+m)m∈N. This is rather a trivial observation if we show that
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
19
whenever βn > β is an eigenvalue of P there exist only finitely many j's such that
βj > βn. Now suppose, on the contrary, that there are infinitely many such j's.
Then they form an infinite bounded set of real numbers with a limit point greater
than or equal to βn. But since βn > β, it contradicts the fact that β is the unique
limit point of the σ(T ).
This inductive method of constructing the decreasing sequence is exhaustive too
and as an immediate consequence we re-order the eigenvalues of P :
n=k+1 converges to β.
n=k+1 the eigenvectors correspond-
n=k+1 respectively. Observe that for an
n=k+1 hx, wni wn so that for
{βn}k
n=1 ∪ {βn}∞
n=1 and {βn}∞
n=k+1; where {βn}∞
n=1 and {wn}∞
Let us rename and denote by {vn}k
ing to the eigenvalues {βn}k
n=1 hx, vni vn +P∞
arbitrary x ∈ H, we have x = Pk
every x ∈ H,
Xn=k+1
hx, wni P (wn)
Xn=k+1
(βn − β)hx, vni vn +
hx, vni P (vn) +
hx, vni βnvn +
hx, wni βnwn
P x =
∞
∞
∞
k
k
=
Xn=1
Xn=1
Xn=1
= k
Xn=1
=
k
Xn=k+1
Xn=k+1
∞
(βn − β)hx, wni wn + βIx
(βn − β)(wn ⊗ wn) + βI! (x).
(βn − β)(vn ⊗ vn) +
itive compact operator.
n=k+1(βn − β)(wn ⊗ wn) and F =Pk
n=1(βn − β)(vn ⊗ vn)
n=k+1(βn − β)(wn ⊗ wn) is a pos-
It then readily follows that P = αI + K + F , where
To conclude this case it suffices to observe that β ≥ 0, Pk
is a self-adjoint finite-rank operator, and P∞
α = β, K =P∞
Case 4. σ(P ) has both a limit point and an eigenvalue with infinite multiplicity.
Let β ∈ {βλ : λ ∈ Λ} be the unique eigenvalue with infinite multiplicity which
compels it to be the unique limit point of the spectrum σ(P ) of P . That the set
Γ := Λ \ {λ : βλ = β} is countable is, at this stage, a trivial observation. This
leaves us with{βλ : λ ∈ Λ} = {βλ : λ ∈ Γ} ∪ {β}. Since {βλ : λ ∈ Γ} is countable,
by the argument in the previous case, we can reorder the eigenvalues of this set in
such a way that for some k ∈ N,
n=1(βn − β)(vn ⊗ vn).
n=1 ∪ {βn}∞
{βλ : λ ∈ Γ} = {βn}k
n=k+1 ∪ {β},
where, by the constructive method discussed previously, {βn}k
n=1 (counting mul-
tiplicities) is the set of all eigenvalues strictly less than β and {βn}∞
n=k+1 is a nonin-
n=1,{wn}∞
creasing sequence converging to β. We next rename and denote by {vn}k
n=1,{βn}∞
and {vλ}λ∈Λ\Γ the eigenvectors corresponding to the eigenvalues {βn}k
and {βλ}λ∈Λ\Γ respectively. Observe that for an arbitrary x ∈ H, we have x =
n=1 hx, vni vn +P∞
Pk
n=k+1 hx, wni wn +Pλ∈Λ\Γ hx, vλi vλ. This yields that for every
n=k+1,
n=k+1,
20
x ∈ H
P x =
=
=
S. K. PANDEY AND V. I. PAULSEN
k
∞
k
Xn=1
Xn=1
Xn=1
k
∞
hx, vni P (vn) +
Xn=k+1
Xn=k+1
(βn − β)hx, vni vn +
hx, vni βnvn +
hx, wni P (wn) + Xλ∈Λ\Γ
hx, wni βnwn + Xλ∈Λ\Γ
Xn=k+1
(βn − β)hx, wni wn
∞
+Xλ∈Γ
hx, vλi βvλ + Xλ∈Λ\Γ
hx, vλi βvλ
hx, vλi P (vλ)
hx, vλi βvλ
= k
Xn=1
(βn − β)(vn ⊗ vn) +
∞
Xn=k+1
(βn − β)(wn ⊗ wn) + βI! (x).
n=k+1(βn−
n=1(βn − β)(vn ⊗ vn).
We complete the proof by observing that in all the four possibilities, we get the
desired form.
(cid:3)
It then immediately follows that P = αI + K + F , where α = β, K =P∞
β)(wn ⊗ wn) and F =Pk
Example 4.12 establishes the fact that the class of AN operators is not closed
under addition. However, it is easy to see that it is closed under scalar multipli-
cation, that is, if T ∈ B(H,K) is AN and α ∈ C, then αT is also AN ; for if M
is an arbitrary nontrivial closed subspace of H, then kαT VMk = αkT VMk =
αkT VM(x0)k = kαT VM(x0)k, where x0 ∈ M,kx0k = 1, and kT VM(x0)k =
kT VMk.
If we consider the class B(H)AN + of positive AN operators, what can be said
about it in the similar vein? To answer this question, let T1, T2 ∈ B(H)AN + .
It is fairly obvious that T1 + T2 is positive. Moreover, by Theorem 5.1, T1 =
α1I + K1 + F1 and T2 = α2I + K2 + F2 where α1, α2 ≥ 0; K1, K2 are posi-
tive compact operators, and F1, F2 are self-adjoint finite-rank operators. Then
T1+T2 = (α1 +α2)I +(K1 +K2)+(F1+F2) and hence it is AN . Also, if c ∈ R, c ≥ 0,
then cT1 ∈ B(H)AN +. Finally, if T and −T are both in B(H)AN + , then hT x, xi ≥ 0
and hT x, xi ≤ 0 which implies that hT x, xi = 0 for each x ∈ H and so T = 0. These
observations, together with the fact that B(H)sa := {T ∈ B(H) : T = T ∗} is a real
Banach space, implies that B(H)AN + is a cone in B(H)sa, which is proper in the
sense that B(H)AN + ∩ (−B(H)AN + ) = {0}.
6. Spectral Characterization Of AN Operators
For any operator T ∈ B(H,K), we know that T ∗T ∈ B(H) and T ∗T ≥ 0.
Moreover, there exists a unique positive operator T := √T ∗T such that T2 =
T ∗T . We state the polar decomposition theorem, which is a standard theorem and
its proof is thus omitted.
Proposition 6.1 (Polar Decomposition Theorem). Let H,K be complex Hilbert
spaces. If T ∈ B(H,K), then there exists a unique partial isometry U : H −→ K
A SPECTRAL CHARACTERIZATION OF AN OPERATORS
21
with final space clos(ranT ) and initial space clos(ranT) such that T = UT and
T = U ∗T . If T is invertible, then U is unitary.
The following lemma is the key to the main theorem of this section.
Lemma 6.2. Let H and K be complex Hilbert spaces and let T ∈ B(H,K). Then
T is AN iff T is AN .
Proof. Let M be an arbitrary nontrivial closed subspace of H. For any x ∈ M
notice that
kTM(x)k = kT xk =phT x, T xi =phT ∗T x, xi
=phT2x, xi =phTx,Txi == kT(x)k = kTM(x)k,
which essentially guarantees that
kTMk = kTMk.
Since M is arbitrary, the assertion follows.
Example 6.3. Let V : l2 → l2 be an isometry onto a subspace M with infinite
codimension. By Proposition 4.2, V is AN . But V ∗ = V V ∗ = PM is the orthog-
onal projection onto M and since PM has two eigenvalues of infinite multiplicity,
PM is not AN by Corollary 3.6. Thus, V is AN but V ∗ is not AN .
(cid:3)
By the preceding lemma, the polar decomposition theorem and the spectral
theorem for positive AN operators, we can safely consider the following theorem
to be fully proved.
Theorem 6.4 (Spectral Theorem For AN Operators). Let H and K be complex
Hilbert spaces of arbitrary dimensions and let T ∈ B(H,K) such that T = U ∗T ,
where U is the unique partial isometry U : H −→ K with final space clos(ranT ) and
initial space clos(ranT). Then T is an AN operator if and only if U ∗T is of the
form U ∗T = αI + F + K, where α ≥ 0, K is a positive compact operator and F is
self-adjoint finite-rank operator.
References
[1] X. Carvajal and W. Neves, Operators that achieve the norm, Integral Equations and Operator
Theory 72 (2012), no. 2, 179 -- 195.
[2] P. Halmos, A hilbert space problem book, Springer-Verlag, New York, 1982.
[3] R. V. Kadison and J. R. Ringrose, Fundamentals of the theory of operator algebras. Special
Topics Volume III: Elementary theory - An exercise approach, Birkhauser Boston, Boston,
MA, 1991.
[4] G. Ramesh, Structure theorem for AN -operators, Journal of the Australian Mathematical
Society 96 (2014), no. 3, 386 -- 395.
Department of Pure Mathematics, University of Waterloo, Ontario, Canada N2L
3G1
E-mail address: [email protected]
Department of Pure Mathematics, University of Waterloo, Ontario, Canada N2L
3G1
E-mail address: [email protected]
|
1708.08116 | 1 | 1708 | 2017-08-27T17:33:03 | Bernstein type inequality in the shift-invariant spaces | [
"math.FA"
] | We are proving a Bernstein type inequality in the shift-invariant spaces of $L_2(R)$. | math.FA | math | НЕРАВЕНСТВА ТИПА БЕРНШТЕЙНА В ПОДПРОСТРАНСТВАХ
ИНВАРИАНТНЫХ ОТНОСИТЕЛЬНА СДВИГА
В.Ф. Бабенко*'**, А.А. Лигун, С.А. Спектор*.
Украина. *Днепропетровский национальный университет
** Институт прикладной математики и механики НАН Украины
Решение многих
задач математического анализа на классах
периодических функций базируется на неравенствах для тригонометрических
многочленов. Причем особое значение имеют неравенства, дающие оценку
нормы производной тригонометрического многочлена через норму самого
многочлена. Первое неравенство такого рода было получено в 1912 году
С.Н.Бернштейном.
Классическое неравенство Бернштейна [1,122]
,
для тригонометрического полинома
порядка
, имеет важные
приложения в теории аппроксимации и других областях математики.
Точные неравенства типа Бернштейна для любого полинома
известны в следующих случаях.
Неравенство
установлено А.А.Лигуном [2].
Неравенства
,
,
1
nnCCn,2,1ktn21nt11coscoskknnppn,1,2...1nkиp2,221Cnn1101,2qppqnq0(1,),гдеqpаqнаименьшеечетноечислобольшееилиравноеq доказаны С.М.Никольским [3]. Также, более детальный обзор точных
неравенств типа Бернштейна можно найти, например, в 3 главе работы [1].
Мы будем рассматривать вопрос о неравенствах типа Бернштейна для
подпространств пространства
и
, инвариантных относительно
сдвига.
Пусть задана функция
. Рассмотрим ортонормированную
систему функций
Очевидно,
[4, 266-271].
.
Пусть
- подпространство пространства
порожденное сдвигами
функции
. Любую функцию
можно представить в виде суммы
сходящегося в
ряда
.
Преобразование Фурье функции
зададим в виде
Тогда,
.
, где
.
Нами доказана следующая
Теорема 1. Зададим функцию
такую, что
.
Тогда, для любой функции
,
выполняется точное
неравенство
(1)
Доказательство. Воспользовавшись а равенство Парсеваля:
,
2
RL2n2RLRLx2h:hxhR212WRL2xfWRL2cfhf12ixRffxedxciffemfmiecxR2k222supfWk22kf2222Rsup22kf2212RRfxdxfd получим
Заметим, что
Таким образом, получим
.
.
Докажем теперь, что константа в правой части неравенства (1)
неулучшаема. Для этого нужно выбрать такое семейство функций
,
чтобы
Пусть
.
–
3
22222211i22kkkRffifd2212kiZRced2201222kiRZced22201222kiZZced222201222kfZmd222222011222iRRffdced2222200112.22ffRmmd22kf222201222kfZmd2222Rsup22kfnf,n222222222sup2kknRnnRff2n21sinn+ω2Φ(ω)=1(n+1)sinω2 ядро Фейера.
Положим
.
Определим последовательность
следующим образом:
Заметим, что
Рассмотрим
(2)
Теорема доказана.
.
Рассмотрим систему функций нецелочисленных сдвигов
.
Соответственно теперь
4
nnmnfnnfm2222n20011122ffmdd22222222ikknnnnffff2222knniR2nR2nk2dd2220220222knRnRdd2220220222knRnnRddR2R2k2n222sup22Rsup22k:,ahhxha -
подпространство пространства
порожденное нецелочисленными
сдвигами функции
.
Для любой функции
имеет место следующая
Теорема 2. Зададим функцию
такую, что
,
Тогда, для любой функции
,
выполняется неравенство
.
Доказательство. Учитывая то, что
,
(3)
,
получим
Заметим, что
Таким образом, получим
.
.
Точность неравенства (3) доказывается также, как в неравенстве (1).
5
2:xWfLRfcaRL2xfWxR22R22supkааWfk22kf2222R22supkfaagxfaxbW1ibagefaa22222221i122kkkkRffiafdaaa22212kkiZRaceda2222201222kkiRZaaced222-2k2012a2kfZmda2222201.2ffаmd22kf2222R22supakfa Теорема доказана.
Обобщим теперь неравенство (1) на многомерном случае.
Будем рассматривать подпространство
пространства
,
порожденное сдвигами фнкции
. Все свойства функций
и
в
многомерном случае остаются такими же как и в одномерном.
Теорема 3. Зададим функцию
такую, что
,
Тогда, для любой функции
,
выполняется
.
неравенство
. (4)
Доказательство. В данном случае имеет место обобщенное равенство
Парсеваля:
Так как
, то
, где
.
Оценим норму
.
Сделаем замену переменной таким образом, чтобы получить интеграл
на промежутке
. Тогда,
6
nW2nLR1n,...,xxf1n,...,xx1,...,nnR1,...,nnR112211,,,...,1sup22,...,2snnnkssnnsnfW1k,...,nnkk1212nkkknfxx11222112,,,...,1sup22,...,2snnnkssnnsf1112222112112,,snnnknkkknsnkksnRfffdxxnfW111,,,,nnnfcxx1n,...,n111222221,,1112221,,1,,ssnnnnsnknnkkisnskkssnRRnkisnsRRffdcedxxcedd0;2n1112222i,,,1120,222snnnnnknksskksnfcedxx где
Заметим, что
Таким образом, получим
.
Покажем теперь, что константа в правой части неравенства (4)
неулучшаема..
Выберем семейство функций
таких, что
Определим последовательность
следующим образом:
положим
где
Рассмотрим
,
.
7
11222,,,10,222,snnnnkissscednn112,,22nn1n12;0,2,i2nn1122dec2,,2fd22ecxxfnn1sn1n12,0,,2n1sk2ss2,i22knk1k11222112,,,...,1sup22,...,2snnnkssnnsf11222,,,10,222,snnnnkissscednWmf121222nkmkknnmfxxf112211,,,...,1sup22,...,2snnnkssnnsmfmmfu1:nmmmssu1222111222222222ssnknnkkmkksmsnnssmmnffxxff
Учитывая свойства (2) ядра Фейера
, получим:
Теорема доказана.
Литература
1. Н.П. Корнейчук, В.Ф. Бабенко, А.А. Лигун. Экстремальные свойства полиномов
и сплайнов. – Киев: Наукова Думка, 1992, С-304.
2. А.А. Лигун. О неравенствах между нормами производных периодических
функций // Мат. заметки. -1983.-33, №3.-С. 385-391.
3. С.М. Никольский.Неравенства для целых функций многих переменных // Тр.
мат. ин-та им. В.А. Стеклова АН СССРю-1951.-38.-С.244-278.
4. Кашин В.С., Саакян А.А. Ортогональне ряды. Изд. 2-е, доп.-М.: Изд-во АФЦ,
1999,-с.550.
8
212snnnksnsRnRdd210,220,2222snnnkssnsndd210,220,222.2snnnkssnsnddn12210,2122220,2222snnnnkkmssnkksnnmnfdxxfd221sup22snkssns |
1008.0108 | 1 | 1008 | 2010-07-28T16:04:16 | Global Saturation of Regularization Methods for Inverse Ill-Posed Problems | [
"math.FA"
] | In this article the concept of saturation of an arbitrary regularization method is formalized based upon the original idea of saturation for spectral regularization methods introduced by A. Neubauer in 1994. Necessary and sufficient conditions for a regularization method to have global saturation are provided. It is shown that for a method to have global saturation the total error must be optimal in two senses, namely as optimal order of convergence over a certain set which at the same time, must be optimal (in a very precise sense) with respect to the error. Finally, two converse results are proved and the theory is applied to find sufficient conditions which ensure the existence of global saturation for spectral methods with classical qualification of finite positive order and for methods with maximal qualification. Finally, several examples of regularization methods possessing global saturation are shown. | math.FA | math |
GLOBAL SATURATION OF REGULARIZATION METHODS FOR
INVERSE ILL-POSED PROBLEMS ∗
TERRY HERDMAN†, RUBEN D. SPIES(cid:12), ‡, AND KARINA G. TEMPERINI §
Abstract.
In this article the concept of saturation of an arbitrary regularization method is
formalized based upon the original idea of saturation for spectral regularization methods introduced
by Neubauer [6]. Necessary and sufficient conditions for a regularization method to have global
saturation are provided. It is shown that for a method to have global saturation the total error must
be optimal in two senses, namely as optimal order of convergence over a certain set which at the
same time, must be optimal (in a very precise sense) with respect to the error. Finally, two converse
results are proved and the theory is applied to find sufficient conditions which ensure the existence
of global saturation for spectral methods with classical qualification of finite positive order and for
methods with maximal qualification. Finally, several examples of regularization methods possessing
global saturation are shown.
Key words. Ill-posed, inverse problem, qualification, saturation.
AMS subject classifications. 47A52, 65J20
1. Introduction. Let X, Y be infinite dimensional Hilbert spaces and T : X →
Y a bounded linear operator such that R(T ) is not closed. It is well known that under
these conditions, the linear operator equation
T x = y
(1)
In that case, x†
.
= R(T ) ⊕ R(T )⊥.
is ill-posed, in the sense that T †, the Moore-Penrose generalized inverse of T , is not
bounded [1]. The Moore-Penrose generalized inverse is strongly related to the least
squares solutions of (1).
In fact this equation has a least squares solution if and
.
= T †y is the least squares
only if y ∈ D(T †)
solution of minimum norm and the set of all least-squares solutions of (1) is given by
x† + N (T ). If the problem is ill-posed then x† does not depend continuously on the
data y. Therefore, if instead of the exact data y, a noisy observation yδ is available,
with (cid:13)(cid:13)y − yδ(cid:13)(cid:13) ≤ δ, where δ > 0 is small, then it is possible that T †yδ does not
even exist and if it does, it will not necessarily be a good approximation of x†. This
instability becomes evident when trying to approximate x† by traditional numerical
methods and procedures. Thus, for instance, it is possible that the application of the
standard least squares approximating procedure on an increasing sequence of finite-
dimensional subspaces {Xn} of X whose union is dense in X, result in a sequence
{xn} of least squares solutions that does not converge to x† (see [8]) or, even worst,
that they diverge from x† with speed arbitrarily large (see [9]).
∗This work was supported in part by Consejo Nacional de Investigaciones Cient´ıficas y T´ecnicas,
CONICET, through PIP 2010-2012 Nro. 0219, by Universidad Nacional del Litoral, U.N.L., through
project CAI+D 2009-PI-62-315, and by the Air Force Office of Scientific Research, AFOSR, through
Grant FA9550-10-1-0018.
†Interdisciplinary Center for Applied Mathematics, ICAM, Virginia Tech, Blacksburg, VA 24061,
USA ([email protected]).
‡Instituto de Matem´atica Aplicada del Litoral, IMAL, CONICET-UNL, Guemes 3450, S3000GLN,
Santa Fe, Argentina and Departamento de Matem´atica, Facultad de Ingenier´ıa Qu´ımica, Universidad
Nacional del Litoral, Santa Fe, Argentina ((cid:12) : [email protected]).
§Instituto de Matem´atica Aplicada del Litoral, IMAL, CONICET-UNL, Guemes 3450, S3000GLN,
Santa Fe, Argentina, and Departamento de Matem´atica, Facultad de Humanidades y Ciencias, Uni-
versidad Nacional del Litoral, Santa Fe, Argentina ([email protected]).
1
2
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
.
0
1
0
gα(λ) = 1
λ .
Ill-posed problems must be first regularized if one wants to successfully attack the
task of numerically approximating their solutions. Regularizing an ill-posed problem
such as (1) essentially means approximating the operator T † by a parametric family
of bounded operators {Rα}, where α is a regularization parameter. If y ∈ D(T †),
then the best approximate solution x† of (1) can be written as x† =R kTk2+
λ dEλT ∗y
where {Eλ} is the spectral family associated to the operator T ∗T (see [1]). This is
mainly why many regularization methods are based on spectral theory and consist in
gα(λ) dEλT ∗ where {gα} is a family of functions appropriately
defining Rα
chosen such that for every λ ∈ (0,kTk2] there holds lim
α→0+
=R kTk2+
However, it is important to emphasize that no mathematical trick can make stable
a problem that is intrinsically unstable. In any case there is always loss of informa-
tion. All a regularization method can do is to recover the largest possible amount
of information about the solution of the problem, maintaining stability. It is often
said that the art of applying regularization methods consist always in maintaining an
adequate balance between accuracy and stability. In 1994, however, Neubauer ([6])
showed that certain spectral regularization methods "saturate", that is, they become
unable to continue extracting additional information about the exact solution even
upon increasing regularity assumptions on it. In his article, Neubauer introduced for
the first time the idea of the concept of "saturation" of regularization methods. This
idea referred to the best order of convergence that a method can achieve indepen-
dently of the smoothness assumptions on the exact solution and on the selection of
the parameter choice rule. Later on, in 1997, Neubauer ([7]) showed that this sat-
uration phenomenon occurs in particular in the classical Tikhonov-Phillips method.
Saturation is however a rather subtle and complex issue in the study of regularization
methods for inverse ill-posed problems and the concept has always escaped rigorous
formalization in a general context.
In 2001, Math´e and Pereverzev ([4]) used Hilbert scales to study the efficiency of
approximating solutions based on observations with noise (stochastic or determinis-
tic). In this context it is possible to quantify the degree of ill-posedness and to obtain
general conditions on projection methods so that they attain optimal order of conver-
gence. These concepts were later extended by the same authors ([5]) who studied the
optimal convergence problem in variable Hilbert scales. In their article they showed
that there is a close relationship between the optimal convergence of a method and
the "a-priori" regularity (in terms of source sets) for spectral methods possessing
qualification of finite order. In 2009 Herdman et al. ([2]) introduced an extension of
the concept of qualification and introduced three different levels: weak, strong and
optimal. It was shown that weak qualification extends the definition introduced by
Math´e and Pereverzev ([5]), in the sense that the functions associated to orders of
convergence and source sets need not be the same.
In 2004, Math´e ([3]) proposed general definitions of the concepts of qualification
and saturation for spectral regularization methods. However, the concept of satura-
tion defined by Math´e is not applicable to general regularization methods and it is
not fully compatible with the original idea of saturation proposed by Neubauer in
[6]. In particular, for instance, the definition of saturation given in [3] does not imply
uniqueness and therefore, neither a best global order of convergence.
In this article a general theory of global saturation for arbitrary regularization
methods is developed. It is shown that saturation involves two aspects: on one hand
(just like in Neubauer's original idea) the characterization of the best global order of
GLOBAL SATURATION OF REGULARIZATION METHODS
3
convergence of the method, and on the other hand, the description of the source set
on which such a best global order of convergence is achieved. Also, necessary and
sufficient conditions are found for a regularization method to have global saturation.
In particular, it is shown that for a method to have saturation, it is necessary that
the total error be optimal in two senses, namely as optimal order of convergence over
a certain set which at the same time, must satisfy a certain optimality condition with
respect to the error. Moreover, an explicit form for the global saturation is given
in terms of the family of regularization operators and the operator associated to the
problem. Lastly, sufficient conditions are provided for spectral methods with qualifi-
cation of positive finite order and for spectral methods with maximal qualification to
have global saturation.
The organization of the paper is as follows. In Section 2 convergence bounds for
regularization methods are defined and an appropriate framework for their comparison
is developed. In Section 3 the concept of global saturation is introduced, its relation
with the total error and with convergence bounds is shown and necessary and sufficient
conditions for the existence of global saturation are provided.
In Section 4, a few
converse results are proved which, together with the results of Section 3, are used to
derive sufficient conditions for the existence of global saturation for certain spectral
regularization methods.
2. Upper Bounds of Convergence for Regularization Methods. In this
section we define what we call upper bounds of convergence for regularization methods
and we develop ways of comparing them on the same as well as on different sets.
Although this section may seem a little lengthy and tedious at a first glance, it provides
a solid mathematical background on which all subsequent formalization and definitions
are based upon.
In sequel and for convenience of notation, unless otherwise specified, we shall
assume that all subsets of the Hilbert space X under consideration are not empty and
they do not contain x = 0. Also, without loss of generality we will assume that the
operator T is invertible (since in the context of inverse problems one always works
with the Moore-Penrose generalized inverse of T , the lack of injectivity is not really
a problem). Given M ⊂ X, we will denote with FM the collection of the following
functions: we will say that ψ ∈ FM if there exists a = a(ψ) > 0 such that ψ is defined
in M × (0, a), with values in (0,∞) and it satisfies the following conditions:
ψ(x, δ) = 0 for all x ∈ M , and
1.
2. ψ is continuous and increasing as a function of δ in (0, a) for each fixed x ∈ M .
Roughly speaking, the collection FM contains all possible δ-"orders of convergence"
on the set M .
lim
δ→0+
Definition 2.1. Let M ⊂ X and ψ, ψ ∈ FM .
i) We say that "ψ precedes ψ on M ", and we denote it ψ
(cid:22) ψ, if there exist a
constant r > 0 and p : M → (0,∞) such that ψ(x, δ) ≤ p(x) ψ(x, δ) for all x ∈ M and
for every δ ∈ (0, r).
M
ii) We say that "ψ, ψ are equivalent on M ", and we denote it ψ
M
≈ ψ, if ψ
M
(cid:22) ψ
M
≺ ψ if ψ
M
(cid:22) ψ
and ψ
M
(cid:22) ψ.
iii) We say that "ψ strictly precedes ψ on M " and we denote it ψ
and lim sup
δ→0+
The following observations follow immediately from these definitions.
= 0 for every x ∈ M.
ψ(x,δ)
ψ(x,δ)
4
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
• Given that ψ, ψ > 0, in iii) the condition lim sup
δ→0+
= 0, i.e., ψ(x, δ) = o( ψ(x, δ)) for δ → 0+.
ψ(x,δ)
ψ(x,δ)
lim
δ→0+
ψ(x,δ)
ψ(x,δ)
= 0 is equivalent to
M
(cid:22)" introduces a partial ordering in FM and "
M
≈" is an equivalence
• The relation "
relation in FM .
≈,
(cid:22) (
• If ψ
M
M
M ⊂ M .
M
≺) ψ then ψ
M
M
M
≈,
(cid:22) (
≺) ψ for every M ⊂ M . With (cid:14), ⊀ and ≈/ we
will denote the negation of the relations (cid:22), ≺ and ≈, respectively.
Lemma 2.2. Let M ⊂ X and ψ, ψ ∈ FM . If ψ
M
≺ ψ then ψ
M
(cid:14) ψ for every
Proof. For the contrareciprocal. Suppose there exists M ⊂ M such that ψ
M
(cid:22) ψ.
{x0}
(cid:22) ψ, that is, there exist constants 0 < p < ∞ and r > 0 such
Let x0 ∈ M , then ψ
that
sup
ψ(x0,δ)
ψ(x0,δ) ≤ p < ∞. Then,
δ ∈(0,r)
lim sup
δ→0+
ψ(x0, δ)
ψ(x0, δ) ≥ lim inf
δ→0+
= sup
ψ(x0, δ)
ψ(x0, δ) ≥ inf
δ∈(0,r)
ψ(x0, δ)!−1
ψ(x0, δ)
≥
δ∈(0,r)
1
p
ψ(x0, δ)
ψ(x0, δ)
> 0.
Therefore, ψ
{x0}
⊀ ψ, from which we deduce that ψ
M
⊀ ψ, since x0 ∈ M ⊂ M .
Definition 2.3. Let {Rα}α∈(0,α0) be a family of regularization operators for the
problem T x = y. We define the "total error of {Rα}α∈(0,α0) at x ∈ X for a noise
level δ" as
E tot
{Rα}
(x, δ)
.
= inf
α∈(0,α0)
sup
yδ∈Bδ (T x)(cid:13)(cid:13)Rαyδ − x(cid:13)(cid:13) ,
where Bδ(T x)
.
= {y ∈ Y : kT x − yk ≤ δ}.
Note that E tot
{Rα}
is the error in the sense of the largest possible discrepancy that
can be obtained for an observation within the noise level δ, with any choice of the
regularization parameter α.
Remark 2.4. Let a > 0, M ⊂ X and E tot
{Rα}
{Rα} ∈ FM . In fact, for each x ∈ M , E tot
{Rα}
: M × (0, a) → (0,∞) be the total
error of {Rα}. Then E tot
(x, δ) is increasing
as a function of δ, and given that {Rα} is a family of regularization operators, it
follows that E tot
(x, δ) is continuous as a function of δ for each fixed x ∈ M and
{Rα}
δ→0+E tot
lim
{Rα}
Definition 2.5. Let {Rα}α∈(0,α0) be a family of regularization operators for the
problem T x = y, M ⊂ X and ψ ∈ FM .
i) We say that ψ is an "upper bound of convergence for the total error of {Rα}α∈(0,α0)
(x, δ) = 0 for every x ∈ M.
on M " if E tot
{Rα}
M
(cid:22) ψ.
{Rα}α∈(0,α0) on M " if E tot
{Rα}
M
≺ ψ.
ii) We say that ψ is a "strict upper bound of convergence for the total error of
GLOBAL SATURATION OF REGULARIZATION METHODS
5
iii) We say that ψ is an "optimal upper bound of convergence for the total error
M
of {Rα}α∈(0,α0) on M " if E tot
{Rα}
E tot
{Rα}
ψ(x, δ)
(cid:22) ψ and
(x, δ)
lim sup
δ→0+
> 0
for every x ∈ M,
) and U opt
M (E tot
{Rα}
(x, δ) 6= o(ψ(x, δ)) when δ → 0+.
or equivalently, if for every x ∈ M E tot
{Rα}
M (E tot
), U str
{Rα}
We will denote with UM (E tot
{Rα}
{Rα} ∈ U opt
) the set of all func-
tions ψ ∈ FM that are, respectively, upper bounds, strict upper bounds and optimal
upper bounds of convergence for the total error of {Rα}α∈(0,α0) on M . In view of
Remark 2.4, it is clear that E tot
The observations below follow immediately from the previous definitions.
• If ψ ∈ FM , then ψ ∈ UM (E tot
{Rα}
δ → 0+ for every x ∈ M . Moreover, U str
of UM (E tot
{Rα}
of just one element).
M (E tot
{Rα}
) if (and only if) E tot
{Rα}
) and U opt
M (E tot
M (E tot
{Rα}
{Rα}
), although their union is not all of UM (E tot
{Rα}
), U opt
M (E tot
{Rα}
) are disjoint subsets
) (except when M consists
• If M ⊂ M , then UM (E tot
{Rα}
) for every M ⊂ X.
) ⊂ U M (E tot
{Rα}
(x, δ) = O(ψ(x, δ)) as
) ⊂ U opt
(E tot
{Rα}
) and
M
M
).
M
) ⊂ U str
• If ψ ∈ UM (E tot
{Rα}
• If ψ ∈ U opt
M (E tot
{Rα}
• If ψ ∈ U str
M (E tot
{Rα}
Definition 2.6. Let ψ, ψ ∈ FM . We say that "ψ and ψ are comparable on M "
(E tot
{Rα}
), ψ ∈ FM and ψ
), ψ ∈ UM (E tot
{Rα}
), ψ ∈ FM and ψ
(cid:22) ψ, then ψ ∈ UM (E tot
{Rα}
) and ψ
M
(cid:22) ψ, then ψ ∈ U str
).
(cid:22) ψ, then ψ ∈ U opt
M (E tot
).
{Rα}
M (E tot
{Rα}
).
M
U str
M (E tot
{Rα}
M
if they verify ψ
(cid:22) ψ (or both).
M
(cid:22) ψ or ψ
(cid:22)(cid:19)" if ψ∗
M
M
M
M
M
ψ∗
Definition 2.7. Let A ⊂ FM and ψ∗ ∈ A. We say that "ψ∗ is a minimal ele-
(cid:22) ψ for every ψ ∈ A comparable with ψ∗ on M . Equivalently,
(cid:22) ψ∗ implies
ment of (cid:18)A,
ψ∗ is minimal element of (cid:18)A,
(cid:22) ψ.
Lemma 2.8. Let A ⊂ FM , ψ, ψ∗ ∈ A and ψ, ψ∗ be comparable on M . If there
(cid:22)(cid:19) if for every ψ ∈ A, the condition ψ
(cid:22)(cid:19) .
≺ ψ∗ then ψ∗ is not a minimal element of (cid:18)A,
≺ ψ∗, then it follows from Lemma 2.2 that ψ∗
Proof. Let A ⊂ FM and ψ, ψ∗ ∈ A be comparable on M . Let us suppose that
M0
(cid:14) ψ.
(cid:14) ψ and since ψ, ψ∗ ∈ A are comparable on M , it follows from Definition
there exists M0 ⊂ M such that ψ
Thus ψ∗
2.7 that ψ∗ cannot be a minimal element of (cid:18)A,
exists M0 ⊂ M such that ψ
(cid:22)(cid:19) .
M0
M0
M
M
M
such that ψ0
Corollary 2.9. If ψ∗ ∈ UM (E tot
{Rα}
) and there exist x0 ∈ M and ψ0 ∈ U{x0}(E tot
{Rα}
)
{x0}≺ ψ∗ then ψ∗ is not a minimal element of (cid:18)UM (E tot
{Rα}
),
M
(cid:22)(cid:19) .
Proof. This corollary is an immediate consequence of the previous lemma with
6
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
A = UM (E tot
{Rα}
), M0 = {x0} and
.
ψ(x, δ)
=(cid:26) ψ0(x0, δ),
Note that this function ψ so defined is in UM (E tot
{Rα}
M (moreover ψ
ψ∗(x, δ),
M
if x = x0
if x 6= x0.
(cid:22) ψ∗).
M
),
{Rα}
(cid:22)(cid:19). More precisely, we have the following result.
Next we will show that the optimal upper bounds of convergence for the total error
of {Rα}α∈(0,α0) on M are characterized by being minimal elements of the partially
ordered set (cid:18)UM (E tot
a minimal element of (cid:18)UM (E tot
M (E tot
{Rα}
(cid:18)UM (E tot
Theorem 2.10. Let ψ ∈ UM (E tot
{Rα}
(cid:22)(cid:19).
M
(cid:22)(cid:19). Then there exists ψc ∈ UM (E tot
) and suppose that ψ is not a minimal element of
Proof. Let ψ ∈ U opt
). Then ψ ∈ U opt
) comparable with ψ on M for
) if and only if ψ is
M (E tot
{Rα}
{Rα}
{Rα}
{Rα}
),
),
M
which it is not true that ψ
M
(cid:22) ψc. Then, there exists x0 ∈ M such that
ψ(x0, δ)
ψc(x0, δ)
= ∞.
lim sup
δ→0+
Now, since ψ ∈ U opt
M (E tot
{Rα}
) and x0 ∈ M , we have that
(x0, δ)
E tot
{Rα}
ψ(x0, δ)
> 0.
lim sup
δ→0+
(2)
(3)
Thus
(x0, δ)
E tot
{Rα}
ψc(x0, δ)
lim sup
δ→0+
(x0, δ)
E tot
{Rα}
ψ(x0, δ)
ψ(x0, δ)
ψc(x0, δ)
= lim sup
δ→0+
). This contradicts the fact that ψc ∈ UM (E tot
{Rα}
= ∞
).
) and it is comparable with ψ∗ on
which implies that ψc /∈ U{x0}(E tot
{Rα}
Therefore, ψ must be a minimal element of (cid:18)UM (E tot
Conversely, assume that ψ ∈ UM (E tot
{Rα}
(E tot
{Rα}
x0 ∈ M such that ψ ∈ U str
{x0}
then implies that ψ is not a minimal element of (cid:18)UM (E tot
of(cid:18)UM (E tot
(cid:22)(cid:19) if and only if it is minimal of(cid:18)UM∗ (E tot
{Rα}
),
M
{Rα}
M
{Rα}
(cid:22)(cid:19).
),
) and ψ /∈ U opt
M (E tot
{Rα}
), which implies that E tot
{Rα}
M
),
{Rα}
). Then there exists
{x0}≺ ψ. Lemma 2.8
(cid:22)(cid:19).
(cid:22)(cid:19) for every M∗ ⊂
M∗
),
From the proof of Theorem 2.10 it follows immediately that ψ is a minimal element
M . Also, as a consequence of Theorem 2.10 one has that all optimal upper bounds
of convergence for the total error must be equivalent in the sense of Definition 2.1-ii.
More precisely we have the following
M (E tot
) and E tot
{Rα}
{Rα}
M
≈ E tot
{Rα}
Corollary 2.11. Let U opt
i) If ψ ∈ U opt
M (E tot
{Rα}
be as before
) then ψ
.
GLOBAL SATURATION OF REGULARIZATION METHODS
7
M (E tot
{Rα}
) then ψ
M
≈ ψ.
ii) If ψ, ψ ∈ U opt
Proof.
i) If ψ ∈ U opt
M
M (E tot
{Rα}
) then E tot
{Rα}
are comparable on M . Then, since E tot
minimal element of (cid:18)UM (E tot
(cid:22) ψ, from which it follows that E tot
{Rα}
{Rα} ∈ UM (E tot
{Rα}
(cid:22)(cid:19), we have that ψ
and ψ
) and by Theorem 2.10 ψ is a
M
(cid:22) E tot
{Rα}
ii) This is an immediate consequence of i) and the transitivity and reflexivity of
≈ E tot
{Rα}
. Hence, ψ
{Rα}
),
M
M
.
M
on M .
≈", because by i) every ψ ∈ U opt
the equivalence relation "
E tot
{Rα}
This result says that if ψ is an optimal upper bound of convergence on M for the
total error of a regularization method, then at every point of M , ψ tends to zero, as
the noise level tends to zero, exactly with the "same speed" with which the total error
does.
M (E tot
{Rα}
) is equivalent to
In order to introduce the concept of saturation in the next section, we will previ-
ously need a few more definitions and tools that will allow us to compare bounds of
convergence on different sets of X.
M, M
if ψ
Definition 2.12. Let M, M ⊂ X, ψ ∈ FM and ψ ∈ F M .
i) We say that "ψ on M precedes ψ on M ", and we denote it with ψ
(cid:22) ψ,
if there exist a constant d > 0 and a function k : M × M → (0,∞) such that
ψ(x, δ) ≤ k(x, x) ψ(x, δ) for every x ∈ M , for every x ∈ M and for every δ ∈ (0, d).
≈ ψ,
ii) We say that "ψ on M is equivalent to ψ on M ", and we denote it with ψ
M, M
(cid:22) ψ and ψ
iii) We say that "ψ on M strictly precedes ψ on M ", and we denote it with
≺ ψ, if ψ
Remark 2.13. In a certain sense, when M = M , the previous definitions gener-
alize (although they are slightly stronger than) the relations introduced in Definition
≺ ψ, although the converse, in general,
= 0 for every x ∈ M , x ∈ M .
(cid:22) ψ and lim sup
δ→0+
2.1. Note for instance that if ψ
is not true.
≺ ψ then ψ
M ,M
(cid:22) ψ.
ψ(x,δ)
ψ(x,δ)
M, M
M, M
M, M
M,M
ψ
M
It follows immediately from Definition 2.12 that if ψ
(cid:22) ψ then ψ
every M ⊂ M and for every N ⊂ N. The same happens for the relations "
≺ ".
"
Next, we also need to extend the notion of "comparability" given in Definition
(cid:22) ψ for
≈ " and
M,N
M,N
M,N
M , N
2.6, to this case.
Definition 2.14. Let M, M ⊂ X, ψ ∈ FM and ψ ∈ F M .
i) We say that "ψ on M is comparable with ψ on M " if ψ
ii) We say that "ψ is invariant over M " if ψ
M,M
≈ ψ.
Remark 2.15.
It is immediate that the condition ψ
M,M
M, M
(cid:22) ψ or ψ
M ,M
(cid:22) ψ.
M,M
≈ ψ is equivalent to
ψ
(cid:22) ψ.
This last notion of "invariance", which will play an important roll in the char-
acterization of saturation, roughly speaking establishes that if ψ is invariant over M
M,M
M,M
≈ ψ and
≈ ψ (i.e. ψ is also invariant over M ).
Lemma 2.16. Let M ⊂ X, ψ, ψ ∈ FM be such that ψ
i) ψ
ii) ψ
Proof.
Let M ⊂ X, ψ, ψ ∈ FM , x, x ∈ M and suppose that ψ
i) Since ψ
M
relation.
δ ∈ (0, d),
≈ ψ, there exist positive constants d, kx and kx such that for every
M,M
≈ ψ and ψ
M
≈ ψ.
8
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
then the orders of convergence of ψ as a function of δ when δ → 0+, in any two points
of M , are equivalent.
The following result is related to a certain transitivity property of this invariance
M
≈ ψ and ψ
M,M
≈ ψ. Then:
ψ(x, δ) ≤ kxψ(x, δ)
and ψ(x, δ) ≤ kx ψ(x, δ).
(4)
On the other hand, from the invariance of ψ over M it follows that there exist positive
constants d∗ y k∗x,x such that ψ(x, δ) ≤ k∗x,xψ(x, δ) for every δ ∈ (0, d∗), which together
with (4) implies that for every δ ∈ (0, min{d, d∗}),
ψ(x, δ) ≤ kxψ(x, δ) ≤ kxk∗x,xψ(x, δ) and ψ(x, δ) ≤ k∗x,xψ(x, δ) ≤ k∗x,xkx ψ(x, δ).
(5)
≈ ψ.
ii) From the first inequality in (5) and from the second inequality in (4) it follows
Since x, x ∈ M are arbitrary, it follows that ψ
(cid:22) ψ, that is, ψ
(cid:22) ψ and ψ
M,M
M,M
M,M
immediately that ψ
The following result is analogous to Lemma 2.2 for this case of comparison of
(cid:22) ψ and therefore by Remark 2.15, ψ is invariant over M .
M,M
convergence bounds on different sets.
Lemma 2.17. Let M, N ⊂ X, ψ ∈ FM and ψ ∈ FN . If ψ
M,N
≺ ψ then ∀ M ⊂ M ,
∀ N ⊂ N we have that ψ
N , M
(cid:14) ψ.
N , M
Proof. By the contrareciprocal. Suppose that there exist M ⊂ M and N ⊂ N
such that ψ
(cid:22) ψ. Then there exist a constant d > 0 and k : N × M → (0,∞)
such that ψ(x, δ) ≤ k(x, x) ψ(x, δ) for every x ∈ N , x ∈ M and δ ∈ (0, d). Let
x0 ∈ M and x0 ∈ N , then ψ(x0, δ) ≤ k(x0, x0) ψ(x0, δ) for every δ ∈ (0, d). Thus,
sup
δ ∈(0,d)
ψ(x0,δ)
ψ(x0,δ) ≤ k(x0, x0) < ∞. Then,
lim sup
δ→0+
ψ(x0, δ)
ψ(x0, δ) ≥ lim inf
δ→0+
= sup
ψ(x0, δ)
ψ(x0, δ) ≥ inf
δ∈(0,d)
ψ(x0, δ)!−1
ψ(x0, δ)
≥
δ∈(0,d)
ψ(x0, δ)
ψ(x0, δ)
1
k(x0, x0)
> 0.
Hence, ψ
{x0},{x0}
⊀
ψ, from which it follows that ψ
x0 ∈ N .
M,N
⊀ ψ, since x0 ∈ M and
3. Global Saturation. We will now proceed to formalize the concept of global
saturation.
Definition 3.1. Let MS ⊂ X and ψS ∈ UMS (E tot
{Rα}
). We say that ψS is a "global
saturation function of {Rα} over MS" if ψS satisfies the following three conditions:
GLOBAL SATURATION OF REGULARIZATION METHODS
9
S1. For every x∗ ∈ X, x∗ 6= 0, x ∈ MS, lim sup
δ→0+
S2. ψS is invariant over MS.
S3. There is no upper bound of convergence for the total error of {Rα} that is a
proper extension of ψS (in the variable x) and satisfies S1 and S2, that is, there exist
no M % MS and ψ ∈ U M (E tot
) such that ψ satisfies S1 and S2 with MS replaced
{Rα}
by M and ψS replaced by ψ.
(x∗,δ)
E tot
{Rα}
ψS (x,δ) > 0.
We shall refer to ψS and MS as the saturation function and the saturation set,
respectively.
M
ψ(x∗,δ)
Remark 3.2. Note that condition S1 implies that for every M ⊂ X and for every
ψS (x,δ) > 0 for every x∗ ∈ M , x ∈ MS (this is an immediate
)). Therefore, it
ψ ∈ UM (E tot
{Rα}
consequence of S1 and the fact that E tot
{Rα}
cannot happen that ψ
), lim sup
δ→0+
(cid:22) ψ ∀ ψ ∈ UM (E tot
{Rα}
≺ ψS. On the other hand, if ψ ∈ UM (E tot
{Rα}
(cid:22) ψ even if ψ on M is comparable to ψS on MS, because
ψS (xS,δ) = 0 for some x ∈ M and some xS ∈
ψ(x,δ)
(cid:14) ψ), and still have lim sup
ψS (xS,δ) > 0.
δ→0+
) then it is not
necessarily true that ψS
in this case it can happen that lim inf
δ→0+
MS (which obviously implies that ψS
MS ,M
MS ,M
M,MS
ψ(x,δ)
However, if ψ on M is comparable with ψS on MS and there exists lim
δ→0+
ψ(x,δ)
ψS(xS ,δ) for
every x ∈ M and for every xS ∈ MS, then it is in fact true that ψS
also that condition S1 can be replaced by
MS ,M
(cid:22) ψ. Note
ψ(x∗, δ)
ψS(x, δ)
lim sup
δ→0+
> 0 ∀ ψ ∈ U{x∗}(E tot
{Rα}),∀ x∗ ∈ X, x∗ 6= 0, x ∈ MS.
This conception of global saturation essentially establishes that in no point x∗ ∈
X, x∗ 6= 0, can exist an upper bound of convergence for the total error of the regular-
ization method that is "strictly better" than the saturation function ψS at any point
of the saturation set MS.
Next we show that any function satisfying condition S1, in particular any satura-
tion function, is always an optimal upper bound of convergence.
). If ψS satisfies the condition S1 on MS, then
Lemma 3.3. Let ψS ∈ UMS (E tot
{Rα}
ψS ∈ U opt
MS
(E tot
{Rα}
).
Proof. The condition S1 implies in particular that lim sup
δ→0+
x ∈ MS. Since also by definition ψS ∈ UMS (E tot
{Rα}
upper bound of convergence for the total error of {Rα}, i.e. ψS ∈ U opt
tion function and the total error on the saturation set.
An immediate consequence of this lemma is the equivalence between the satura-
(E tot
{Rα}
MS
).
(x,δ)
E tot
{Rα}
ψS (x,δ) > 0 for every
) it follows that ψS is an optimal
Corollary 3.4.
If ψS is a saturation function of {Rα} on MS then ψS
MS≈
E tot
{Rα}
. Moreover, we have the stronger equivalence ψS
MS ,MS≈ E tot
{Rα}
.
Proof. The first part of the corollary is an immediate consequence of the previous
lemma and of Corollary 2.11 i). The second part follows from the first and the fact
that ψS
MS ,MS≈ ψS, via Lemma 2.16 i).
10
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
Remark 3.5. A consequence of the first part of this corollary and of Lemma 2.16
MS ,MS≈ E tot
ii) is that if ψS is a saturation function of {Rα} on MS, then E tot
,
{Rα}
{Rα}
that is, the total error must be invariant over MS. We will shed more light on this
matter in Theorem 3.8.
Definition 3.6. Let M ⊂ X and ψ ∈ U X (E tot
{Rα}
for ψ", and we denote it with M ∈ O(ψ), if the following condition holds:
C2. For every x ∈ M , xc ∈ M c neither ψ {xc},{x}≺ ψ nor ψ {xc},{x}≈ ψ.
). We say that "M is optimal
That a set M be optimal for ψ essentially means that at any point of the com-
plement of M , the order of convergence of ψ as a function of δ, for δ → 0+, cannot
be better nor even equivalent to the order of convergence of ψ at any point outside
M ; that is, at any point outside of M , the order of convergence of ψ must be strictly
worse than itself at any point of M . However, we will see next that this optimality
condition imposes a very precise restriction. As we shall see later on (Theorem 3.8),
it is precisely this property of the total error, together with its invariance on the
set MS, what will allow us to characterize the regularization methods which do have
saturation.
Condition C2 is very precise and gives no room for maneuver. In fact, let ψ ∈
U X (E tot
{Rα}
C1.
), M ⊂ X and consider the following conditions:
ψ
M,M c
M c,M
M c,M
≈/
≈/ ψ.
{xc},{x}
{xc},{x}
ψ, from which it follows immediately that ψ
C3.
Then it follows that condition C2 (of optimal set) is strictly stronger than condi-
tion C3, and strictly weaker than condition C1. In fact, if M is optimal for ψ in the
sense of Definition 3.6, then for every x ∈ M , xc ∈ M c we have that ψ
⊀ ψ and
⊀ ψ and ψ
≈/ ψ, that
ψ
is, C3 holds. However, for condition C3 to hold it is sufficient that there exist x ∈ M
and xc ∈ M c such that ψ
ψ, which obviously does not imply
condition C2. On the other hand if C1 holds, then it follows from Lemma 2.17 that
for every x ∈ M , xc ∈ M c, there holds ψ
⊀ ψ and
ψ for every x ∈ M , xc ∈ M c, that is, condition C2 holds. However, C2
ψ
does not imply C1 since it can happen that M be optimal for ψ and that there exist
x ∈ M and xc ∈ M c such that ψ on {x} is not comparable with ψ on {xc}. This
implies in particular that ψ
(cid:14) ψ and therefore, ψ
(cid:14) ψ and therefore, ψ
⊀ ψ and ψ
{xc},{x}
{xc},{x}
{xc},{x}
{xc},{x}
{x},{xc}
{xc},{x}
M,M c
⊀ ψ.
≈/
≈/
In order to be able to characterize the regularization methods which do have
saturation, we will previously need the following result.
every x ∈ M , xc ∈ M c there holds E tot
{Rα}
for every x ∈ M , xc ∈ M c.
Lemma 3.7. Suppose that {Rα} has saturation function on M ⊂ X and for
E tot
{Rα}
Proof. Since {Rα} has saturation function on M , it follows from Remark 3.5
for every x ∈ M ,
that E tot
{Rα}
xc ∈ M c and that there exist x ∈ M , xc ∈ M c such that
is invariant over M . Suppose that E tot
{Rα}
. Then E tot
{Rα}
⊀ E tot
{Rα}
⊀ E tot
{Rα}
{xc},{x}
{xc},{x}
{xc},{x}
≈/
M c,M
≺ ψ.
⊀ ψ and ψ
ψ
M c,M
E tot
{Rα}
{xc},{x}≈ E tot
{Rα}
.
(6)
GLOBAL SATURATION OF REGULARIZATION METHODS
Then,
lim sup
δ→0+
E tot
{Rα}
E tot
{Rα}
(x, δ)
(xc, δ)
Define M
.
= M ∪ {xc} and
ψ(x, δ)
.
=(cid:26) ψ(x, δ),
E tot
{Rα}
(x, δ),
> 0.
if x ∈ M
if x = xc,
11
(7)
where ψ is a saturation function of {Rα} on M . We will show next that ψ is saturation
function on M . Clearly, ψ is upper bound of convergence for the total error on M ,
i.e., ψ ∈ U M (E tot
) and since ψ is saturation on M , it follows that ψ(x, δ) satisfies
{Rα}
condition S1 for all x ∈ M . We will now check that ψ(xc, δ) also satisfies S1. Since
x ∈ M it follows that
lim sup
δ→0+
E tot
{Rα}
E tot
{Rα}
(x∗, δ)
(x, δ)
> 0 ∀ x∗ ∈ X, x∗ 6= 0.
(8)
If x∗ ∈ M , the above inequality follows from the fact that E tot
{Rα}
{x∗},{x}
⊀ E tot
and if x∗ ∈ M c, it is a consequence of the fact that E tot
{Rα}
{Rα}
.
is invariant over M
Then, for every x∗ ∈ X, x∗ 6= 0 we have that
E tot
{Rα}
E tot
{Rα}
E tot
{Rα}
ψ(xc, δ)
= lim sup
δ→0+
lim sup
δ→0+
(x∗, δ)
(x∗, δ)
(x, δ)
(x, δ)
E tot
{Rα}
ψ(xc, δ)
> 0
by virtue of (7) and (8). Thus, ψ(x, δ) satisfies S1 for every x ∈ M .
We will now check that ψ satisfies S2 on M . Since ψ is saturation function of
{Rα} on M , and ψM = ψ we have that ψ is invariant over M . It remains to prove
that ψ {xc},M
≈ ψ. But this is an immediate consequence of
and the fact that ψ is invariant
≈ ψ, i.e. that E tot
{Rα}
{xc},M
M
(6), of Corollary 3.4 which implies that ψ
over M .
≈ E tot
{Rα}
.
≈/
Thus, we have shown that ψ is a proper extension of ψ satisfying S1 and S2 on
M , which then implies that ψ does not satisfy condition S3. This contradicts the fact
that ψ is saturation function of {Rα} on M . Therefore, for every x ∈ M , xc ∈ M c
there must hold that E tot
{Rα}
E tot
{Rα}
{xc},{x}
M (x, δ)
. In this case E tot
Theorem 3.8. (Necessary and sufficient condition for the existence of satura-
tion.) A regularization method {Rα} has saturation function if and only if there exists
M ⊂ X (M 6= {0}, M 6= ∅) such that E tot
is invariant over M and M is optimal
{Rα}
.
for E tot
= E tot
(x, δ) for x ∈ M and δ > 0 is saturation
{Rα}
{Rα}
function of {Rα} on M .
Remark 3.5 that E tot
{Rα}
Proof. Suppose that {Rα} has saturation function ψ on M . Then it follows from
Let us now check that M is optimal for E tot
. Let x ∈ M and xc ∈ M c. We will
{Rα}
) and x ∈ M , there exist
. Since ψ ∈ UM (E tot
{Rα}
first show that E tot
{Rα}
is invariant over M .
E tot
{Rα}
{xc},{x}
⊀
12
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
positive constants d and kx such that E tot
{Rα}
Then
(x, δ) ≤ kxψ(x, δ) for every δ ∈ (0, d).
lim sup
δ→0+
E tot
{Rα}
E tot
{Rα}
(xc, δ)
(x, δ) ≥ lim sup
δ→0+
(xc, δ)
E tot
{Rα}
kxψ(x, δ)
> 0,
{xc},{x}
⊀ E tot
{Rα}
where the last inequality follows from the fact that ψ satisfies condition S1 on M .
Therefore ∀ x ∈ M , ∀ xc ∈ M c, E tot
{Rα}
the fact that E tot
{Rα}
∀ xc ∈ M c, E tot
{Rα}
. This condition together with
is invariant over M implies, by virtue of Lemma 3.7, that ∀ x ∈ M ,
{xc},{x}
. We have thus shown that M is optimal for E tot
≈/
.
{Rα}
Conversely, suppose that there exists M ⊂ X (M 6= {0}, M 6= ∅) such that E tot
{Rα}
is invariant over M and M is optimal for E tot
(x, δ)
{Rα}
for x ∈ M and δ > 0. We will show that E tot
M is saturation function of {Rα} on M .
Clearly, E tot
M is invariant over M , it only
remains to be shown that E tot
) and since by hypothesis E tot
M satisfies conditions S1 and S3.
In order to prove S1, let x∗ ∈ X, x∗ 6= 0 and x ∈ M. If x∗ ∈ M , then the
M ∈ UM (E tot
{Rα}
and define E tot
.
= E tot
{Rα}
E tot
{Rα}
M (x, δ)
invariance of E tot
M over M implies that E tot
M
M and therefore
lim sup
δ→0+
(x∗, δ)
E tot
{Rα}
E tot
M (x, δ)
{x∗},{x}≈ E tot
M (x∗, δ)
E tot
E tot
M (x, δ)
= lim sup
δ→0+
> 0.
(9)
⊀
{x∗},{x}
E tot
M (by condition C2 ) because M is optimal for E tot
{Rα}
On the other hand, if x∗ ∈ M c, the previous limit is also positive due to the fact that
. Then, E tot
E tot
{Rα}
satisfies condition S1.
M does not satisfy condition S3, i.e. there exist M % M
and ψ ∈ U M (E tot
M satisfying conditions
{Rα}
S1 and S2 on M . Let x ∈ M \ M , then the invariance of ψ over M implies that
ψ {x},M
) such that ψ is a proper extension of E tot
Finally, suppose that E tot
M on M , it follows that
M
≈ ψ and since ψ coincides with E tot
ψ {x},M
≈ E tot
M .
(10)
M
M
) satisfies S1 on M , Lemma 3.3 implies that ψ ∈ U opt
Now since ψ ∈ U M (E tot
(E tot
{Rα}
{Rα}
{x}≈
≈ ψ. In particular, E tot
Then, by virtue of Corollary 2.11.i we have that E tot
{Rα}
{Rα}
{x},M
{x},M
ψ, which, together with (10) imply that E tot
M , that is, E tot
≈ E tot
≈ E tot
.
{Rα}
{Rα}
{Rα}
But since x ∈ M c, this equivalence contradicts the fact that M is optimal for E tot
.
{Rα}
Therefore, E tot
M must satisfy condition S3 and, as a consequence, it is saturation func-
tion of {Rα} on M .
Remark 3.9. From the previous theorem we conclude that a saturation function
of a regularization method is an optimal upper bound of convergence for the total error,
invariant and without proper extensions.
).
Note that a saturation function must be optimal in two senses. In fact, if ψ is
saturation function on M , then M is optimal for ψ and ψ is optimal (upper bound)
for the total error of {Rα} on M . Moreover, M and ψ (modulus M, M equivalence)
are uniquely determined. In fact, if the domain M is changed, then M is no longer
GLOBAL SATURATION OF REGULARIZATION METHODS
13
optimal for ψ and if the function ψ is changed, even at a single point of M , in such
a way that ψ is not invariant on M , then ψ it is no longer an optimal upper bound.
Suppose that at a point x0 ∈ M , we redefine ψ as ψ(x0, δ), where ψ ∈ FM . If ψ {x0}≺ ψ,
then ψ is no longer an upper bound for the total error of {Rα} on M and if ψ {x0}≺ ψ
then ψ is upper bound but it is not optimal. Thus for every M ⊂ M and for every
ψ ∈ F M , if ψ and ψ are comparable on M then ψ
≈ ψ must hold.
M, M
4. Saturation for Spectral Regularization Methods. The objective of this
section is to apply the theory previously developed to the case of spectral regulariza-
tion methods. Further, we show that this theory is consistent with previously existing
results about optimal convergence of spectral regularization methods.
Let {Eλ}λ∈IR be the spectral family associated to the linear selfadjoint operator
T ∗T and {gα}α∈(0,α0) a parametric family of functions gα : [0,kTk2] → IR for α ∈
(0, α0), and consider the following standing hypotheses:
H1. For every α ∈ (0, α0) the function gα is piecewise continuous on [0,kTk2].
H2. There exists a constant C > 0 (independent of α) such that λgα(λ) ≤ C
H3. For every λ ∈ (0,kTk2], there exists lim
α→0+
= O(cid:16) 1√α(cid:17) for α → 0+.
H4. Gα
If {gα}α∈(0,α0) satisfies hypotheses H1-H3, then (see [1], Theorem 4.1) the collec-
for every λ ∈ [0,kTk2].
.
= kgα(·)k∞
gα(λ) = 1
λ .
tion of operators {Rα}α∈(0,α0), where
Rα
.
=Z gα(λ) dEλ T ∗ = gα(T ∗T )T ∗,
(11)
is a family of regularization operators for T †. In this case we say that {Rα}α∈(0,α0)
is a family of spectral regularization operators for T x = y.
Next, we recall the classical definition of qualification for a family of spectral
regularization operators.
Definition 4.1. Let {Rα}α∈(0,α0) be the family of spectral regularization opera-
.
= 1− λgα(λ),
tors for T x = y generated by the family of functions {gα}α∈(0,α0), rα(λ)
0 < α < α0, 0 ≤ λ ≤ kTk2, and let us denote with I(gα) the set
Remark 4.2. Note that by virtue of H2, 0 ∈ I(gα) and the order µ0 of the
classical qualification of a regularization method is always nonnegative (it can be equal
to 0 or +∞).
4.1. Spectral Methods with Classical Qualification of Finite Positive
Order. We start by considering first the case of spectral methods for which 0 < µ0 <
∞. For these methods we will first show the existence of certain upper bounds of
convergence and then we will show that they saturate. We will also characterize their
saturation functions and saturation sets.
Lemma 4.3. Suppose that {gα}α∈(0,α0) satisfies the hypotheses H1-H4. If the
family of regularization operators {Rα}α∈(0,α0), with Rα defined as in (11), has clas-
.
= {µ ≥ 0 : ∃ k > 0 and λµ rα(λ) ≤ k αµ ∀ λ ∈ [0,kTk2], ∀ α ∈ (0, α0)}.
I(gα)
The order of the classical qualification of {Rα}α∈(0,α0) is defined to be µ0
µ and we say that {Rα}α∈(0,α0) has classical qualification of order µ0.
.
=
sup
µ∈I(gα)
14
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
2µ0
.
= δ
2µ0 +1 , for x ∈ Xµ0
.
sical qualification of order µ0, 0 < µ0 < +∞, then ψµ0 (x, δ)
=
R((T ∗T )µ0) \ {0} and δ > 0, is upper bound of convergence for the total error of
{Rα}α∈(0,α0) on Xµ0 , that is, ψµ0 ∈ UXµ0
Proof. Since {gα} satisfies hypothesis H4, we have that Gα = O(cid:16) 1√α(cid:17) when
α → 0+ and therefore Gα = o( 1
α ) when α → 0+. From this and from the fact that
{gα} satisfies hypothesis H1-H3 and {Rα} has classical qualification of order µ0,
0 < µ0 < +∞, it follows that (see [1], Corollary 4.4 and Remark 4.5 therein) there
exists an a-priori parameter choice rule α∗ : IR+ → (0, α0) such that the regularization
method (Rα, α∗) is of optimal order on Xµ0 , that is, for every x ∈ Xµ0 there exists
k(x) > 0 such that for every δ > 0,
(E tot
{Rα}
).
2µ0
2µ0 +1 .
sup
yδ∈Bδ(T x)(cid:13)(cid:13)Rα∗(δ)yδ − x(cid:13)(cid:13) ≤ k(x) δ
yδ∈Bδ(T x)(cid:13)(cid:13)Rαyδ − x(cid:13)(cid:13) ≤ k(x) δ
sup
inf
α∈(0,α0)
2µ0
2µ0 +1 ,
Then
that is, E tot
{Rα}
(x, δ) ≤ k(x)ψµ0 (δ). Thus E tot
{Rα}
Xµ0
(cid:22) ψµ0 and therefore ψµ0 ∈ UXµ0
(E tot
{Rα}
).
Theorem 4.4. (Saturation for families of spectral regularization operators with
classical qualification of finite positive order.)
Suppose further that:
Suppose that {gα}α>0 satisfies hypotheses H1-H4 and let rα(λ)
i) The spectrum of T ∗T has λ = 0 as accumulation point.
ii) There exist positive constants γ1, γ2, λ1, c1, with λ1 ≤ kTk2 and c1 > 1 such
.
= 1 − λgα(λ).
that
a) 0 ≤ rα(λ) ≤ 1, α > 0, 0 ≤ λ ≤ λ1;
b) rα(λ) ≥ γ1, 0 ≤ λ < α ≤ λ1;
c) rα(λ) is monotone increasing with respect to α for λ ∈ (0,kTk2];
d) gα(c1α) ≥ γ2
e) gα(λ) ≥ gα(λ), for 0 < α ≤ λ ≤ λ ≤ λ1.
α , 0 < c1α ≤ λ1 and
There exist constants γ, c > 0 such that:
iii) The family of regularization operators {Rα}α∈(0,α0) defined by (11), where
.
= min{λ1, λ1
iv)
c }, has classical qualification of order µ0, 0 < µ0 < +∞.
α0
α(cid:19)µ0
(cid:18) λ
rα(λ) ≥ γ,
for every 0 < cα ≤ λ ≤ kTk2 .
(12)
Then ψµ0 (x, δ)
.
= δ
2µ0
2µ0 +1 for x ∈ Xµ0
.
= R((T ∗T )µ0 )\{0} and δ > 0, is saturation
function of {Rα}α∈(0,α0) on Xµ0.
Note that the hypothesis i) is trivially satisfied if T is compact. To prove this
theorem we will need two previous lemmas. In the first one we show that under the
hypotheses of Theorem 4.4, for all α in a right neighborhood of zero one has that
0 ∈ ρ (rα(T ∗T )), i.e. zero belongs to the resolvent set of the operator rα(T ∗T ). More
precisely we have the following:
GLOBAL SATURATION OF REGULARIZATION METHODS
15
Lemma 4.5. Suppose that {gα}α>0 satisfies hypotheses H1-H4 and assume further
that hypotheses ii.b), ii.c), iii) and vi) of Theorem 4.4 hold. Then for every α ∈
(0, α0) the operator rα(T ∗T ) is invertible, where α0
.
= min{λ1, λ1
c }.
two cases.
Z α
0
< +∞.
Proof. It suffices to show that for every α ∈ (0, α0) and for every x ∈ X, the
α (λ) is integrable with respect to the measure dkEλxk2. Let α ∈ (0, α0)
function r−2
be arbitrary but fixed. Since α0 ≤ λ1, it follows from hypothesis ii.b) that rα(λ) ≥
γ1 > 0 for every λ ∈ [0, α). Then
1
r2
α(λ)
dkEλxk2 ≤ kxk2
It remains to prove thatR kTk2+
α(λ) dkEλxk2 < +∞. For that we shall consider
Case I: c ≤ 1. In this case, for every λ ∈ [α,kTk2] we have that λ ≥ α ≥ c α > 0
λ(cid:1)µ0 for every λ ∈ [α,kTk2]. Therefore
and from (12) it follows that rα(λ) ≥ γ(cid:0) α
Z kTk2+
(αµ0 γ)2 dkEλxk2 ≤ k(T ∗T )µ0xk2
< +∞.
Case II: c > 1. In this case, since cα < kTk2 we write
dkEλxk2 +Z kTk2+
Z kTk2+
dkEλxk2 ≤Z kTk2+
dkEλxk2 =Z c α
dkEλxk2 .
1
r2
α(λ)
1
r2
α(λ)
1
r2
α(λ)
1
r2
α(λ)
α
1
r2
α
α
α
α
(αµ0 γ)2
λ2µ0
c α
(14)
(13)
γ2
1
Like in the previous case, by virtue of (12), the second integral on the RHS of
(αµ0 γ)2 < +∞. For the first integral on the RHS of
(14) is bounded above by k(T ∗T )µ0 xk2
(14), by virtue of hypothesis ii.c) we have that
r2
α(λ) ≥ r2
α/c(λ),
∀ λ ∈ [α, cα]
(15)
c < α. On the other hand, again by using (12), and given that 0 < c( α
because α
we have that
c ) ≤ λ
(16)
α/c(cid:19)2µ0
(cid:18) λ
α/c(λ) ≥ γ2.
r2
α(λ) ≥ γ2(cid:0) α
c2µ0
α2µ0 γ2 λ2µ0 dkEλxk2 ≤
From (15) and (16) we conclude that r2
for the first integral on the RHS of (14) we have the estimate
α
1
r2
α(λ)
Z c α
Hence rα(T ∗T ) is an invertible operator for every α ∈ (0, α0).
dkEλxk2 ≤Z c α
α
c λ(cid:1)2µ0 for every λ ∈ [α, c α]. Thus,
c2µ0
α2µ0 γ2 k(T ∗T )µ0xk2 < ∞.
Lemma 4.6. Suppose that {gα}α>0 satisfies the hypotheses H1-H4 and sup-
pose further that hypotheses ii.b), ii.c), iii) and iv) of Theorem 4.4 hold. Let
ϕ : [0,kTk2] → IR+ be a continuous, strictly increasing function satisfying ϕ(0) = 0.
If for some x∗ ∈ X, x∗ 6= 0 we have that E tot
(x∗, δ) = o(ϕ(δ)) for δ → 0+, then
{Rα}
there exists an a-priori parameter choice rule α(δ) such that
sup
yδ∈Bδ (T x∗)(cid:13)(cid:13)R α(δ)yδ − x∗(cid:13)(cid:13) = o(ϕ(δ))
for δ → 0+.
16
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
The same remains true if we replace o(ϕ(δ)) by O(ϕ(δ)).
Proof. Let ϕ be as in the hypotheses and suppose that there exists x∗ ∈ X, x∗ 6= 0
such that E tot
{Rα}
inf
α∈(0,α0)
lim
δ→0+
(x∗, δ) = o(ϕ(δ)) for δ → 0+. Then by definition of E tot
{Rα}
yδ∈Bδ(T x∗)(cid:13)(cid:13)Rαyδ − x∗(cid:13)(cid:13)
yδ∈Bδ(T x∗)(cid:13)(cid:13)Rαyδ − x∗(cid:13)(cid:13)
ϕ(δ)
ϕ(δ)
sup
sup
,
= lim
δ→0+
inf
α∈(0,α0)
= 0.
(17)
For the sake of simplify we introduce the following notation:
f (α, δ)
.
=
sup
yδ∈Bδ (T x∗)(cid:13)(cid:13)Rαyδ − x∗(cid:13)(cid:13)
ϕ(δ)
and h(δ)
.
= inf
α∈(0,α0)
f (α, δ).
Then h(δ) > 0 for every δ ∈ (0,∞) and (17) can be written simply as lim
δ→0+
Next, for n ∈ IN we define
h(δ) = 0.
δn
.
= sup(cid:26)δ > 0 : h(δ) ≤
1
n(cid:27) .
Clearly, δn ↓ 0 and h(δ) = inf
Then, there exists αn = αn(δn) ∈ (0, α0) such that
f (α, δ) ≤ 1
α∈(0,α0)
n for every δ ∈ (0, δn] for every n ∈ IN.
f (αn, δ) ≤
2
n ∀ δ ∈ (0, δn], ∀ n ∈ IN.
(18)
n→+∞
lim
δ→0+
f (α(δ), δ) = lim
We then define α(δ)
it follows from (18) that
.
= αn for all δ ∈ (δn+1, δn] for every n ∈ IN. Then, since δn ↓ 0
f (αn, δn) = 0. We could choose
α as the parameter choice rule we are looking for. The problem is that we cannot
guarantee the existence of the limit of α(δ) for δ → 0+. However, we will see next
that α(δ) can be replaced by a function α : IR+ → (0, α0) such that lim
α(δ) = 0
δ→0+
(i.e, such that α(δ) is an admissible parameter choice rule) maintaining the condition
In fact, since {αn}n∈IN ⊂ (0, α0) is a bounded sequence of
lim
δ→0+
real numbers, it contains a convergent subsequence {αnk}k∈IN, with αnk → α∗ for
.
k → +∞, and some α∗ ∈ [0, α0]. We define α(δ)
= αnk for all δ ∈ (δnk+1 , δnk ], for
every k ∈ IN. Then,
f (α(δ), δ) = 0.
lim
δ→0+
α(δ) = lim
k→+∞
αnk = α∗.
(19)
Since {αnk}k∈IN and {δnk}k∈IN are subsequences of {αn}n∈IN and {δn}n∈IN,
lim
δ→0+
f (αnk , δnk ) = 0. Then, by definition of f ,
f (α(δ), δ) = lim
k→+∞
lim
δ→0+
that is,
sup
yδ∈Bδ (T x∗)(cid:13)(cid:13)R α(δ)yδ − x∗(cid:13)(cid:13)
ϕ(δ)
= 0,
sup
yδ∈Bδ(T x∗)(cid:13)(cid:13)R α(δ)yδ − x∗(cid:13)(cid:13) = o(ϕ(δ)), as δ → 0+.
(20)
GLOBAL SATURATION OF REGULARIZATION METHODS
17
It remains to be shown that α∗ = 0. If α∗ > 0, then it follows from (19) that there
exists δ0 > 0 such that α(δ) > α∗
2 for all δ ∈ (0, δ0). Hypothesis ii.c) of Theorem 4.4
for all λ ∈ (0,kTk2]. It
implies then that for every δ ∈ (0, δ0), (cid:12)(cid:12)r α(δ)(λ)(cid:12)(cid:12) ≥ (cid:12)(cid:12)(cid:12)
follows that for every δ ∈ (0, δ0),
(λ)(cid:12)(cid:12)(cid:12)
r α∗
2
2
(cid:13)(cid:13)r α(δ)(T ∗T )x∗(cid:13)(cid:13)
Then, for all δ ∈ (0, δ0),
0
=Z kTk2+
≥Z kTk2+
=(cid:13)(cid:13)(cid:13)
r α∗
0
2
2
.
(T ∗T )x∗(cid:13)(cid:13)(cid:13)
α(δ)(λ) dkEλx∗k2
r2
r2
α∗
2
(λ) dkEλx∗k2
sup
yδ∈Bδ(T x∗)(cid:13)(cid:13)R α(δ)yδ − x∗(cid:13)(cid:13) ≥(cid:13)(cid:13)R α(δ)T x∗ − x∗(cid:13)(cid:13) =(cid:13)(cid:13)(I − g α(δ)(T ∗T )T ∗T )x∗(cid:13)(cid:13)
Taking limit for δ → 0+ and using (20) we conclude that (cid:13)(cid:13)(cid:13)
= 0. But
since α∗
(T ∗T ) is invertible and therefore
x∗ = 0, which is a contradiction since x∗ was not zero to start with. Hence, α∗ must
be equal to zero, as wanted.
=(cid:13)(cid:13)r α(δ)(T ∗T )x∗(cid:13)(cid:13) ≥(cid:13)(cid:13)(cid:13)
2 < α0, it follows from Lemma 4.5 that r α∗
(T ∗T )x∗(cid:13)(cid:13)(cid:13)
(T ∗T )x∗(cid:13)(cid:13)(cid:13)
r α∗
r α∗
.
2
2
2
We proceed now to prove the second part of the Lemma. Suppose that there
(x∗, δ) = O(ϕ(δ)) as δ → 0+. Then there
exists x∗ ∈ X, x∗ 6= 0 such that E tot
{Rα}
f (α, δ) ≤ k for every δ ∈ (0, d),
exist positive constants k and d such that
where f (α, δ) is as previously defined. Let {δn}n∈IN ⊂ (0, d) be such that δn ↓ 0 and
αn = αn(δn) ∈ (0, α0) such that
α∈(0,α0)
inf
f (αn, δ) ≤ k + δn, ∀ δ ∈ (0, d), ∀ n ∈ IN.
.
= αn for all δ ∈
We define (just like we did it previously for the "o" case) α(δ)
(δn+1, δn] for every n ∈ IN. Since δn ↓ 0 it follows that f (α(δ), δ) ≤ k + δ1 for every
δ ∈ (0, d) and therefore
sup
(21)
yδ∈Bδ(T x∗)(cid:13)(cid:13)Rα(δ)yδ − x∗(cid:13)(cid:13) = O(ϕ(δ))
as δ → 0+.
Exactly in the same way as we proceeded before in the first part of the proof, by
defining the function α(δ) (from a convergent subsequence of {αn}n∈IN), equation (21)
is proved with α(δ) in place of α(δ). Finally, and also by proceeding in an analogous
way, it is shown that α(δ) converges to zero as δ → 0+, i.e. that α(δ) is an admissible
parameter choice rule. Since the steps are essentially the same we do not give details
here.
We are now ready to prove Theorem 4.4.
Proof of Theorem 4.4. We will show that ψµ0 (x, δ)
2µ0 +1 for x ∈ Xµ0 and
δ > 0, is saturation function of {Rα}α∈(0,α0) on Xµ0 .
). Next we will show
that ψµ0 satisfies condition S1 of saturation on Xµ0 (see Definition 3.1). Suppose that
First we note that by virtue of Lemma 4.3, ψµ0 ∈ UXµ0
(E tot
{Rα}
.
= δ
2µ0
18
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
(x∗,δ)
it is not true, i.e. suppose that there exist x∗ ∈ X, x∗ 6= 0 and x ∈ Xµ0 such that
2µ0 +1(cid:17) as δ → 0+ and from Lemma
lim sup
δ→0+
4.6 it follows that there exists an a-priori admissible parameter choice rule α(δ) such
that
E tot
ψµ0 (x,δ) = 0. Then E tot
{Rα}
{Rα}
(x∗, δ) = o(cid:16)δ
2µ0
sup
yδ∈Bδ (T x∗)(cid:13)(cid:13)Rα(δ)yδ − x∗(cid:13)(cid:13) = o(δ
2µ0
2µ0 +1 )
for δ → 0+.
Now note that hypothesis H4 implies that there exists a finite positive constant
β such that √λgα(λ) ≤ β√α , for every α ∈ (0, α0) and for every λ ∈ [0,kTk2]. Since
{gα} satisfies the hypotheses H1-H4 and i)-iv) hold, it follows from Theorem 3.1 of
[6] that x∗ = 0, which contradicts the fact that x∗ was different from zero. Hence,
ψµ0 satisfies condition S1 on Xµ0 . Since ψµ0 does not depend on x, we further have
that ψµ0 is (trivially) invariant over Xµ0, i.e., it satisfies condition S2.
It only remains to prove that ψµ0 satisfies condition S3, that is, that the set Xµ0
is optimal for ψµ0 . Suppose that is not the case. Then there must exist M % Xµ0
= ψµ0 and ψ satisfies S1 and S2 on M . Let
and ψ ∈ UM (E tot
{Rα}
x∗ ∈ M \ Xµ0 , x∗ 6= 0. Since ψ ∈ UM (E tot
{Rα}
) such that ψ Xµ0
) we have that
E tot
{Rα}
{x∗}
(cid:22) ψ.
Also, since ψ is invariant over M , we have that ψ
with ψµ0 on Xµ0 , it follows that ψ
E tot
{Rα}
then implies that there exists an a-priori admissible parameter choice rule α(δ) such
that
(cid:22)
ψµ0 and therefore E tot
{Rα}
2µ0 +1(cid:17) as δ → 0+. Lemma 4.6
ψµ0 . This, together with (22) implies that
(x∗, δ) = O(cid:16)δ
ψ, and since ψ coincides
{x∗},Xµ0
{x∗},Xµ0
{x∗},Xµ0
(22)
(cid:22)
2µ0
(cid:22)
sup
yδ∈Bδ(T x∗)(cid:13)(cid:13)Rα(δ)yδ − x∗(cid:13)(cid:13) = O(cid:16)δ
2µ0
2µ0 +1(cid:17) as δ → 0+.
Since µ0 < +∞ it follows that x∗ ∈ R((T ∗T )µ0) (see [6], Corollary 2.6) and since
x∗ 6= 0, we have that x∗ ∈ Xµ0 which contradicts that x∗ ∈ M \ Xµ0 . Thus, ψµ0
satisfies condition S3 and ψµ0 is saturation function of {Rα} on Xµ0, as we wanted
to prove.
4.2. Spectral Methods with Maximal Qualification. The concept of classi-
cal qualification is a special case of a more general definition of qualification introduced
by Math´e and Pereverzev ([5], [3]).
Definition 4.7. Let {Rα}α∈(0,α0) be a family of spectral regularization operators
.
for T x = y generated by the family of functions {gα}α∈(0,α0) and let rα(λ)
= 1 −
λgα(λ). A function ρ : (0,kTk2] → IR+ is said to be qualification of {Rα}α∈(0,α0) if ρ
is increasing and there exists a constant γ > 0 such that
sup
λ∈(0,kTk2]rα(λ) ρ(λ) ≤ γ ρ(α)
for every α ∈ (0, α0).
If, moreover, for every λ ∈ (0,kTk2] there exists a constant c
.
= c(λ) > 0 such that
inf
α∈(0,α0)
rα(λ)
ρ(α) ≥ c
GLOBAL SATURATION OF REGULARIZATION METHODS
19
Note then that the classical qualification of order µ corresponds to the case in
then ρ is said to be maximal qualification of {Rα}α∈(0,α0).
which the functions ρ are restricted to monomials ρ(t) = tµ for 0 ≤ µ < +∞.
These two definitions of qualification are closely related. For instance, if a spectral
regularization method {Rα} possesses classical qualification of order µ0 < ∞, then
any increasing function ρ : (0,kTk2] → IR+ satisfying αµ0 ≤ k ρ(α) for some constant
k > 0, for α in a neighborhood of α = 0, is also qualification of {Rα}. Also, if αµ0
and ρ(α) are two maximal qualifications then they are necessarily equivalent in the
sense that there exist constants k, k > 0 such that k αµ0 ≤ ρ(α) ≤ k αµ0 for every
α ∈ (0, α0). On the other hand, if a spectral regularization method {Rα} has classical
qualification of infinite order, then it does not necessarily have maximal qualification.
Next, we will show that under certain general hypotheses, it is also possible to
characterize the saturation of spectral regularization methods possessing maximal
qualification. For that we will previously need the following definition.
lim
t→0+
Definition 4.8. Let ρ : (0, a] → (0, +∞) be a continuous non-decreasing function
ρ(t) = 0 and β ∈ IR, β ≥ 0. We say that ρ is of local upper type β
s )βρ(s t) for every s ∈ (0, 1],
such that
if there exists a positive constant d such that ρ(t) ≤ d ( 1
t ∈ (0, a].
A function of finite upper type is also said to satisfy a ∆β condition.
Theorem 4.9. (Saturation for families of spectral regularization operators with
maximal qualification.)
Let T be a compact linear operator. Suppose that {gα}α∈(0,α0) satisfies hypotheses
H1-H4 and let {Rα}α∈(0,α0) be as defined by (11). Suppose further that the following
hypotheses are satisfied:
λn
λn+1 ≤ c
M1: There exist {λn}∞n=1 ⊂ σp(T T ∗) and c ≥ 1 such that λn ↓ 0 and
M2: There exist positive constants λ1 ≤ kTk2, γ1, γ2 and c1 > 1 such that
for every n ∈ IN.
a) 0 ≤ rα(λ) ≤ 1, α > 0, 0 ≤ λ ≤ λ1;
b) rα(λ) ≥ γ1, 0 ≤ λ < α ≤ λ1;
c) rα(λ) is monotone increasing as a function of α for each λ ∈ (0,kTk2];
d) gα(c1α) ≥ γ2
e) gα(λ) ≥ gα(λ), for 0 < α ≤ λ ≤ λ ≤ λ1.
α , 0 < c1α ≤ λ1 and
M3: There exists ρ : (0,kTk2] → (0, +∞), strictly increasing and of local upper
type β, for some β ≥ 0, such that ρ is maximal qualification of {Rα}α∈(0,α0) and there
exist positive constants a and k such that
ρ(λ)rα(λ)
ρ(α)
≥ a,
for all α, λ such that 0 < k α ≤ λ ≤ kTk2 .
.
M4: For every α ∈ (0, α0) the function λ → rα(λ)2, λ ∈ (0,kTk2] is convex.
= (ρ ◦ Θ−1)(δ) for x ∈ X ρ .
=
Let Θ(t)
R(ρ(T ∗T )) \ {0} and δ ∈ (0, Θ(α0)), is saturation function of {Rα}α∈(0,α0) on X ρ.
In order to prove this theorem we will previously need two converse results that
= √tρ(t) for t ∈ (0,kTk2]. Then ψ(x, δ)
.
we establish in the following two Lemmas.
Lemma 4.10. Let {Rα}α∈(0,α0) be a family of spectral regularization operators for
T x = y and ρ : (0,kTk2] → IR+ a strictly increasing continuous function satisfying
If for some x ∈ X, kRαT x − xk = O(ρ(α)) for
hypothesis M3 of Theorem 4.9.
α → 0+, then x ∈ R(ρ(T ∗T )).
20
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
Proof. From hypothesis M3 it follows that
kRαT x − xk2 =Z kTk2+
Since kRαT x − xk = O(ρ(α)) for α → 0+, it then follows that there are constants
C > 0 and α∗, 0 < α∗ ≤ α0 such that
α(λ) dkEλxk2 ≥ a2 ρ2(α)Z kTk2+
ρ−2(λ) dkEλxk2 .
(23)
r2
k α
0
k α
Z kTk2+
ρ−2(λ) dkEλxk2 ≤ kRαT x − xk2
Taking limit for α → 0+ we obtain that R kTk2+
it follows that w
a2ρ2(α)
0
ρ−1(λ) dEλ x ∈ X. Then,
.
0
=R kTk2+
ρ(T ∗T )w =Z kTk2+
0
ρ(λ)ρ−1(λ) dEλx = x
C2
a2
≤
for every α ∈ (0, α∗).
ρ−2(λ) dkEλxk2 < +∞, from which
and therefore x ∈ R(ρ(T ∗T )).
Lemma 4.11. Under the same hypotheses of Theorem 4.9, if for some x ∈ X we
have that
inf
sup
yδ∈Bδ (T x)
then x ∈ R(ρ(T ∗T )).
α∈(0,α0)(cid:13)(cid:13)Rαyδ − x(cid:13)(cid:13) = O(ρ(Θ−1(δ))) when δ → 0+,
, λ1
Proof. Without loss of generality we assume that α0 ≤ min{ λ1
k } and that
Let ¯λ ∈ σp(T T ∗) be such that 0 < c1 ¯λ ≤ λ1 (the compactness of T guarantees
x 6= 0 (if x = 0 the result is trivial).
the existence of such ¯λ), and define
(24)
c1
¯δ = ¯δ(¯λ)
.
=
¯λ1/2
γ2 kR¯λT x − xk .
Then, clearly the equation
kRαT x − xk2 =
(γ2 ¯δ)2
α
(25)
in the unknown α, has α = ¯λ as a solution. Moreover, from the hypothesis M2 c)
and given that x 6= 0, it follows that α = ¯λ is the unique solution of (25). Note also
that ¯δ → 0+ if (and only if) ¯λ → 0+.
Now, for δ > 0 define
where G¯λ
.
= Fc1 ¯λ − F¯λ and {Fλ} is the spectral family associated to T T ∗ and
¯y δ .
= T x − δG¯λz,
∀ δ > 0,
(26)
z
.
=(cid:26) kG¯λT xk−1 T x,
arbitrary withkG¯λzk = 1,
if G¯λT x 6= 0,
in other case.
Note that since ¯λ ∈ σp(T T ∗) and c1 > 1 it follows that G¯λ is not the null operator and
therefore the definition makes sense. Note also that (cid:13)(cid:13)¯y δ − T x(cid:13)(cid:13) = δ, which implies
that ¯y δ ∈ Bδ(T x).
GLOBAL SATURATION OF REGULARIZATION METHODS
21
Now, from (11), (26) and from the fact that gα(T ∗T )T ∗ = T ∗gα(T T ∗) it follows
that for every α ∈ (0, α0) and δ > 0,
hRαT x − x , Rα(¯y δ − T x)(cid:11) = hgα(T ∗T )T ∗T x − x,−gα(T ∗T )T ∗ δG¯λzi
= δ hgα(T ∗T )T ∗T x − x,−T ∗gα(T T ∗)G¯λzi
= δ hT gα(T ∗T )T ∗T x − T x,−gα(T T ∗)G¯λzi
= δ h(T T ∗gα(T T ∗) − I)T x,−gα(T T ∗)G¯λzi
= δ h−rα(T T ∗)T x,−gα(T T ∗)G¯λzi
= δZ kTk2+
rα(λ)gα(λ) dhFλT x, G¯λzi .
0
(27)
Now since c1¯λ ≤ λ1, it follows from hypothesis M2 a) that both gα(λ) and rα(λ)
are nonnegative for all λ ∈ [0, c1¯λ]. On the other hand, from the definitions of G¯λ
.
= hFλT x, G¯λzi for λ ∈ [0, c1¯λ] is
and z it follows immediately that the function h(λ)
real and non-decreasing and therefore
Z c1 ¯λ+
0
rα(λ)gα(λ) dhFλT x, G¯λzi ≥ 0.
(28)
On the other hand, since h(λ) = hT x, FλG¯λzi and FλG¯λ = G¯λ for every λ ≥ c1¯λ, it
follows that h(λ) is constant for every λ ≥ c1¯λ and therefore
Z kTk2+
c1 ¯λ+
rα(λ)gα(λ) dhFλT x, G¯λzi = 0.
(29)
From (28) and (29) we have that
Z kTk2+
0
rα(λ)gα(λ) dhFλT x, G¯λzi ≥ 0,
which, by virtue of (27), implies that
By using once again (11) and (26) together with (30) it then follows that for every
(cid:10)RαT x − x, Rα(¯y δ − T x)(cid:11) ≥ 0.
2
2
α ∈ (0, α0), for every ¯λ ∈ σp(T T ∗) such that c1¯λ ≤ λ1 and for every δ > 0,
(cid:13)(cid:13)Rα ¯y δ − x(cid:13)(cid:13)
+ 2(cid:10)RαT x − x, Rα(¯y δ − T x)(cid:11)
= kRαT x − xk2 +(cid:13)(cid:13)Rα(¯y δ − T x)(cid:13)(cid:13)
= kRαT x − xk2 + δ2 kgα(T ∗T )T ∗G¯λzk2 + 2(cid:10)RαT x − x, Rα(¯y δ − T x)(cid:11)
≥ kRαT x − xk2 + δ2 kgα(T ∗T )T ∗G¯λzk2
= kRαT x − xk2 + δ2Z kTk2 +
≥ kRαT x − xk2 + δ2Z c1 ¯λ
α(λ) dkFλG¯λzk2
λ g2
α(λ) dkFλG¯λzk2 .
λ g2
0
¯λ
We now consider two different possible cases.
Case I: α ≤ ¯λ. Since c1¯λ ≤ λ1 and c1 > 1, it follows from hypothesis M2 e) that
(32)
gα(λ) ≥ gα(c1¯λ) ≥ gα(λ1)
for every λ ∈ [¯λ, c1 ¯λ].
(30)
(31)
22
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
¯λ
λ g2
On the other hand, from hypothesis M2 a) it follows that rα(λ1) ≤ 1, which implies
that λ1 gα(λ1) ≥ 0 and therefore, gα(λ1) ≥ 0. It then follows from (32) that g2
α(λ) ≥
α(c1 ¯λ) for every λ ∈ [¯λ, c1 ¯λ]. Then,
g2
Z c1 ¯λ
dkFλG¯λzk2
α(λ) dkFλG¯λzk2 ≥ ¯λ g2
= ¯λ g2
α(c1 ¯λ)Z c1 ¯λ
(33)
where the last equality follows from the fact that R c1 ¯λ
dkFλG¯λzk2 = 1, which is a
consequence of the fact that R c1 ¯λ
− kF¯λG¯λzk2, from the
definition of G¯λ, from the fact that FλFµ = Fmin{λ,µ} for every λ, µ ∈ IR and the fact
that kG¯λzk = 1.
At the same time, the hypotheses M2 a) and M2 c) imply that gα(λ) is mono-
tone decreasing as a function of α for each λ ∈ [0, λ1]. Since α ≤ ¯λ and c1 ¯λ ≤ λ1, we
then have that
dkFλG¯λzk2 = (cid:13)(cid:13)Fc1 ¯λG¯λz(cid:13)(cid:13)
α(c1 ¯λ),
¯λ
¯λ
¯λ
2
and from hypothesis M2 d) we also have that
gα(c1 ¯λ) ≥ g¯λ(c1 ¯λ),
(34)
(35)
(36)
From (34) and (35) we conclude that
Substituting (36) into (33) we obtain
g2
g¯λ(c1 ¯λ) ≥ γ2/¯λ > 0.
α(c1 ¯λ) ≥(cid:0)γ2/¯λ(cid:1)2
α(λ) dkFλG¯λzk2 ≥ γ2
.
λ g2
2 /¯λ,
Z c1 ¯λ
¯λ
2
0
0
r2
¯λ(λ)
α(λ) ≥ r2
≥ (γ2 δ)2/¯λ.
Case II: α > ¯λ. In this case, it follows from hypothesis M2 c) that r2
which, together with (31) imply that if α ≤ ¯λ, then (cid:13)(cid:13)Rα ¯y δ − x(cid:13)(cid:13)
for every λ ∈ (0,kTk2]. Then,
kRαT x − xk2 =Z kTk2 +
α(λ) dkEλxk2 ≥Z kTk2 +
which, together with (31) imply that (cid:13)(cid:13)Rα ¯y δ − x(cid:13)(cid:13)
≥(cid:26) kR¯λT x − xk2 ,
≥ min{kR¯λT x − xk2 , (γ2 δ)2/¯λ},
¯λ(λ) dkEλxk2 = kR¯λT x − xk2 ,
r2
≥ kR¯λT x − xk2.
Summarizing the results obtained in cases I and II, we can write:
(37)
which is valid for every α ∈ (0, α0), ¯λ ∈ σp(T T ∗) such that c1¯λ ≤ λ1 and for every
δ > 0. Then
minnkR¯λT x − xk , γ2 δ/p¯λo =(cid:16)min{kR¯λT x − xk2 , (γ2 δ)2/¯λ}(cid:17)1/2
(cid:13)(cid:13)Rα ¯y δ − x(cid:13)(cid:13)
if α > ¯λ
if α ≤ ¯λ.
(γ2 δ)2/¯λ,
2
2
(by (37))
α∈(0,α0)(cid:13)(cid:13)Rα ¯y δ − x(cid:13)(cid:13)
≤ inf
≤ sup
= O(ρ(Θ−1(δ)))
α∈(0,α0)(cid:13)(cid:13)Rαyδ − x(cid:13)(cid:13)
yδ∈Bδ(T x)
for
inf
(since ¯y δ ∈ Bδ(T x))
δ → 0+ (by hypothesis).
GLOBAL SATURATION OF REGULARIZATION METHODS
23
Now, given that ¯λ = α(¯δ) solves equation (25), from the previous inequality we
have that
(cid:13)(cid:13)(cid:13)
Rα(¯δ)T x − x(cid:13)(cid:13)(cid:13)
which implies that
= γ2 ¯δ/p¯λ = O(ρ(Θ−1(¯δ)))
for
¯δ → 0+,
¯δ
ρ(Θ−1(¯δ))
= O(cid:18)qα(¯δ)(cid:19) for
¯δ → 0+.
(38)
(39)
(38) we then deduce that:
Since δ = Θ(Θ−1(δ)) it follows from the definition of Θ that δ =pΘ−1(δ) ρ(Θ−1(δ)).
Then, it follows from (39) that pΘ−1(¯δ) = O(pα(¯δ)) for ¯δ → 0+. From this and
(cid:13)(cid:13)(cid:13)
Rα(¯δ)T x − x(cid:13)(cid:13)(cid:13)
for ¯δ → 0+ ∀ α(¯δ) ∈ σp(T T ∗) : c1 α(¯δ) ≤ λ1.
= O(ρ(α(¯δ)))
(40)
Now, let α ∈ IR+ such that α ≤ max
j∈IN{λj : λj ≤ λ1
c1 }. Then, there exist n = n(α) ∈
N such that λn+1 < α ≤ λn ≤ λ1
that
c1
From hypothesis M2 c) and the fact that λn ∈ σp(T T ∗) and λn ≤ λ1
c1
it follows
. Note here that n → ∞ if (and only if) α → 0+.
0
kRαT x − xk2 =Z kTk2 +
≤Z kTk2 +
=(cid:13)(cid:13)Rλn
0
2
T x − x(cid:13)(cid:13)
α(λ) dkEλxk2
r2
r2
λn
(λ) dkEλxk2
= O(ρ2(λn)),
(41)
From hypothesis M1 we have that λn ≤ c λn+1 and since ρ is strictly increasing and
positive (by hypothesis M3 ) it follows that for all n big enough, more precisely for
all n such that c λn+1 ≤ kTk2,
(by virtue of (40)).
ρ2(λn) ≤ ρ2(c λn+1).
(42)
Now since c ≥ 1 and ρ is of local upper type β for some β ≥ 0 (hypothesis M3 ), there
exists a positive constant d such that
ρ(c λn+1) ≤ d cβρ(cid:18) 1
c
c λn+1(cid:19) = d cβρ(λn+1).
(43)
From (41), (42), (43) and from the fact that ρ(λn+1) < ρ(α) it follows that kRαT x − xk =
O(ρ(α)) for α → 0+. Therefore, Lemma 4.10 now implies that x ∈ R(ρ(T ∗T )). This
concludes the proof of the Lemma.
Remark 4.12. From the definition of qualification (Definition 4.7) it follows that
kRαT x − xk2 ≤ γ2 ρ2(α)Z +∞
0
ρ−2(λ) dkEλxk2 .
24
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
Therefore, in Lemma 4.11, the hypothesis M1 and the assumption that ρ be of lo-
cal upper type β for some β ≥ 0 can be substituted by the requirement that ρ(T ∗T )
be invertible, or equivalently, that ρ−2(λ) be integrable with respect to the measure
dkEλxk2 for every x ∈ X.
Proof of Theorem 4.9. First we note that from hypotheses M2 d) and M2 e) it
follows easily that
M5 :
sup
λ∈(0,kTk2]
√λ gα(λ) ≥
b
√α
for every α ∈ (0, α0),
c1
, λ1
k }. First we will prove that ψ(x, δ)
where b = γ2√c1. As in Lemma 4.11, without loss of generality we assume that
.
α0 ≤ min{ λ1
= (ρ ◦ Θ−1)(δ) for x ∈ X ρ and
δ ∈ (0, Θ(α0)), is an upper bound of convergence for the total error of {Rα}α∈(0,α0)
in X ρ, that is, we will show that ψ ∈ UX ρ(E tot
). For every r ≥ 0 we define the
{Rα}
source sets X ρ,r .
= {x ∈ X : x = ρ(T ∗T )ξ, kξk ≤ r}. Let x ∈ X ρ, then there exists
r ≥ 1 such that x ∈ X ρ,r. Since Θ is continuous and strictly increasing in (0, α0),
there exists a unique α ∈ (0, α0) such that x ∈ X ρ,r and Θ(α) = δ
yδ∈Bδ (T x)(cid:13)(cid:13)Rα yδ − x(cid:13)(cid:13)
yδ∈Bδ(T x)(cid:13)(cid:13)R α yδ − x(cid:13)(cid:13) .
yδ∈Bδ(T x)(cid:13)(cid:13)R α yδ − x(cid:13)(cid:13)
On the other hand, from hypotheses H1-H4, the fact that the function ρ is qualification
of {Rα}, the fact that ρ trivially covers ρ with constant equals to unity (see [5],
Definition 2) and given that Θ(α) = δ
r , it follows by virtue of Theorem 2 in [5], that
there exists a positive constant K, independent of δ such that
E tot
{Rα}(x, δ) = inf
α∈(0,α0)
≤ sup
r . Therefore,
≤ sup
x∈X ρ,r
sup
(44)
sup
sup
x∈X ρ,r
sup
yδ∈Bδ(T x)(cid:13)(cid:13)R α yδ − x(cid:13)(cid:13) ≤ K ρ(cid:18)Θ−1(cid:18) δ
From (44) and (45) it follows that for every δ ∈ (0, Θ(α0)),
r(cid:19)(cid:19) , for 0 < δ ≤ r Θ(kTk2).
(45)
E tot
{Rα}
(x, δ) ≤ K ρ(cid:18)Θ−1(cid:18) δ
r(cid:19)(cid:19) ≤ K ρ(Θ−1(δ)) = K ψ(x, δ),
where the last inequality follows from the fact that r ≥ 1 and both ρ and Θ−1 are
increasing functions. This proves that ψ ∈ UX ρ(E tot
{Rα}
Next we will see that ψ satisfies condition S1 of saturation on X ρ. From hy-
potheses H1-H4, M4 and M5 and the fact that ρ is maximal qualification of {Rα},
by virtue of Theorem 2.3 and Definition 2.2 in [3], it follows that for every x∗ ∈ X,
.
= a(x, x∗) and d = d(x, x∗) such
x∗ 6= 0 and x ∈ X ρ there exist positive constants a
that
).
(x∗, δ)
E tot
{Rα}
ψ(x, δ) ≥ a ∀ δ ∈ (0, d).
E tot
{Rα}
(x∗,δ)
Then, lim sup
δ→0+
condition S1 on X ρ.
ψ(x,δ) > 0 for every x∗ ∈ X, x∗ 6= 0 and x ∈ X ρ, that is, ψ satisfies
GLOBAL SATURATION OF REGULARIZATION METHODS
25
Also, since ψ does not depend on x, it is invariant over X ρ, i.e., ψ satisfies
condition S2 of saturation.
It remains to prove that ψ satisfies condition S3. Suppose not. Then, there exist
) such that ψ X ρ= ψ and ψ satisfies S1 and S2 over M .
) we have that
M % X ρ and ψ ∈ UM (E tot
{Rα}
Let x∗ ∈ M \ X ρ, x∗ 6= 0. Since ψ ∈ UM (E tot
{Rα}
{x∗}
(cid:22) ψ.
E tot
{Rα}
Since ψ is invariant over M and X ρ ⊂ M , it follows that ψ {x∗},X ρ
(cid:22)
ψ coincides with ψ on X ρ, it follows that ψ
(cid:22)
ψ. This together with (46)
(x∗, δ) = O(ρ(Θ−1(δ))) for δ → 0+.
(cid:22) ψ and therefore E tot
imply that E tot
{Rα}
{Rα}
Lemma 4.6 then implies that there exists an a-priori admissible parameter choice rule
α : IR+ → (0, α0) such that
{x∗},X ρ
{x∗},X ρ
(46)
ψ and since
Then,
sup
yδ∈Bδ(T x∗)(cid:13)(cid:13)R α(δ)yδ − x∗(cid:13)(cid:13) = O(ρ(Θ−1(δ)))
α∈(0,α0)(cid:13)(cid:13)Rαyδ − x∗(cid:13)(cid:13) = O(ρ(Θ−1(δ)))
sup
inf
yδ∈Bδ (T x∗)
for δ → 0+.
for δ → 0+.
Finally, Lemma 4.11 implies that x∗ ∈ R(ρ(T ∗T )) and since x∗ 6= 0, we have that
x∗ ∈ X ρ, which contradicts the fact that x∗ ∈ M \ X ρ. Hence, ψ satisfies condition
S3 and therefore, ψ is saturation function of {Rα} on X ρ.
Note that both Lemma 4.5 and Lemma 4.6 remain true if hypotheses iii) and iv)
of Theorem 4.4 are replaced by the requirement that there exists ρ : (0,kTk2] → (0,∞)
that is qualification of {Rα}α∈(0,α0) and satisfies the inequality in the hypothesis M3
of Theorem 4.9.
5. Examples. We close our investigation presenting a few examples of regular-
ization methods possessing global saturation. For the sake of brevity we shall not
provide much details here.
Example 1: The family of Tikhonov-Phillips regularization operators {Rα}α∈(0,α0)
is defined by (11) with gα(λ) = 1
λ+α . It is well known that this family of regular-
ization operators possesses classical qualification of order µ0 = 1. It can be easily
checked that the family {gα}α∈(0,α0) satisfies all hypotheses of the Theorem 4.4 with
.
.
= kTk2, γ1
constants C
2 , γ2
= 1. Therefore,
.
= R(T ∗T ) \ {0} and δ > 0 is global
the function ψ(x, δ) = δ
saturation of {Rα}α∈(0,α0) on X1.
.
= 3
3 defined for x ∈ X1
Example 2: Given k ∈ R+, for α, λ > 0 let
.
= 1, λ1
2 and c
.
= 1
.
= 1
.
= 2
2 , c1
5 , γ
2
λ
e− λ√α
e−√ λ
e−√ λ
λ
α
,
,
α
λ + αk
λk+1 ,
hk
α(λ)
.
=
for 0 < λ < α,
for α ≤ λ < 3α,
for λ ≥ 3α,
26
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
α(λ)
.
= 1
λ − αk√λ − hk
α(λ) for λ > 0, and for λ = 0 define gk
.
and define gk
=
α(λ) = 1√α . It can be easily verified that for any α0 > 0, {gα}α∈(0,α0) sat-
gk
lim
λ→0+
isfies the hypotheses H1 -H3 and therefore the corresponding collection of opera-
tors {Rα}α∈(0,α0) defined by (11) is a family of spectral regularization operators for
0. Also, it can easily be proved
T x = y. Hypothesis H2 is satisfied with C
= O(1) for α → 0+ and therefore {Rα}α∈(0,α0)
α(λ)
that for any λ > 0,
has classical qualification of order k.
.
= 1+kTk3 αk
λk1 − λgk
α(0)
αk
α(λ) is non-increasing. Thus, hypoth-
α(0) = 1√α , which implies
esis ii.e) of Theorem 4.4 holds and Gk
α
immediately that also hypothesis H4 is verified. From now on we shall assume k ≥ 1.
Now, for k ≥ 1 and α > 0, the function gk
= (cid:13)(cid:13)gk
α(·)(cid:13)(cid:13)∞
= gk
Defining
.
e− λ√α ,
e−√ λ
α ,
e−√ λ
.
sk
α(λ)
=
α(λ) = 1 − λgk
for 0 ≤ λ < α,
for α ≤ λ < 3α,
for λ ≥ 3α,
α(λ). Clearly, rk
it follows that rk
α(λ) > 0 for all λ ≥ 0.
α(λ) ≤ 1
Now let α0
for all λ ∈ [0, λ1] and for all α ∈ (0, α0), i.e., hypothesis ii.a) of Theorem 4.4 is
satisfied.
3 } and λ1
.
= min{ 1
α(λ) = αkλ
λ(cid:1)k
α +(cid:0) α
.
= min{1,kTk2}. It can be shown that rk
3 , kTk2
2 + sk
,
3
Also, for 0 ≤ λ < α ≤ λ1 we have that
rk
α(λ) = αkλ
3
2 + e− λ√α > e−1,
since λ√α < 1. Thus, hypothesis ii.b) of Theorem 4.4 is verified with γ1
.
= e−1. Since
α(λ) is monotone increasing with respect to α for all λ ≥ 0, hypothesis
(cid:12)(cid:12)rk
α(λ)(cid:12)(cid:12) = rk
ii.c) of Theorem 4.4 is also satisfied.
On the other hand we have that
αgk
α(2α) =
1 − e−√2
2
−
√2 α
3
2 +k ≥
1 − e−√2
2
−
√2 3− 3
2−k,
3 . Hence hypothesis ii.d) of Theorem 4.4 holds as well with constants
since α ≤ 1
= 1−e−√2
.
.
= 2 and γ2
c1
Finally, for λ ≥ 3α,
α(cid:19)k
(cid:18) λ
2
− √2 3− 3
α(λ)(cid:12)(cid:12) =(cid:18) λ
(cid:12)(cid:12)rk
2−k > 0 for all k ≥ 1.
α(cid:19)k(cid:18)e−√ λ
α + αkλ
3
2 +(cid:16) α
λ(cid:17)k(cid:19) ≥ 1,
2k
from which it follows that hypothesis iv) of Theorem 4.4 is satisfied with constants
.
.
= 1. Hence, Theorem 4.4 allows us to conclude that the function
c
= 3 and γ
.
= R((T ∗T )k) \ {0} and δ > 0 is global saturation of
2k+1 for x ∈ Xk
ψk(x, δ) = δ
{Rα}α∈(0,α0) on Xk.
Example 3: Given ε > 0, for λ > 0 and α ∈ (0, α0) with α0 < e−1, let
hε(λ)
.
=(α,
α1+ε,
for 0 ≤ λ < α,
for λ ≥ α,
GLOBAL SATURATION OF REGULARIZATION METHODS
27
and define
gε
α(λ)
.
=
1 + ln α
λ ln α − λ−εhε(λ)
.
It can be easily checked that {gε
α}α∈(0,α0) satisfies the hypotheses H1-H4. In partic-
.
= 1. Therefore {Rα}α∈(0,α0) with Rα as in
ular, hypothesis H2 is satisfied with C
(11) is a family of regularization operators for T x = y. Also it can be shown that for
every µ > 0,
λµ 1 − λgε
α(λ)
αµ
→ +∞ for α → 0+ for every λ > 0,
.
= −(ln α)−1 is strictly increasing and of local upper type β for β
which implies that {Rα}α∈(0,α0) has classical qualification of order µ0 = 0. Now, the
.
function ρ(α)
= 1
.
(moreover the constant d in Definition 4.8 can be taken to be d
= 1) and it can also
be proved that ρ is maximal qualification of {Rα}α∈(0,α0) and satisfies the inequality
in the hypothesis M3 of Theorem 4.9 with constants a
.
= 1 and k
.
= 1.
In this case we have that
rα(λ) =
hε(λ) + λ1+ε
hε(λ) − λ1+ε ln α
.
α(λ) > 0 for all λ ≥ 0. Also, it can be shown that rε
Clearly, rε
λ ∈ [0, λ1] and for all α ∈ (0, α0), where λ1
M2 a) of Theorem 4.9 is satisfied.
α(λ) ≤ 1 for all
.
= min{0.6,kTk2}. Thus, hypothesis
Now, for 0 ≤ λ < α ≤ λ1, we have that
rε
α(λ) =
α + λ1+ε
α − λ1+ε ln α ≥
1
1 − λ1+ε
α ln α
>
1
1 − αε ln α
,
(47)
since λ1+ε
α < α1+ε
α = αε. Since one can easily prove that
− αε ln α ≤ (3e)−1
for all α > 0,
(48)
α(λ) > (1 + 1
it follows from (47) and (48) that rε
implies that hypothesis M2 b) of Theorem 4.9 holds with γ1
α(λ) = rε
rε
M2 c) of Theorem 4.9 is also satisfied.
3e )−1 for 0 ≤ λ < α ≤ λ1, which
3e )−1. Since
α(λ) is monotone increasing with respect to α for all λ ≥ 0, hypothesis
α(λ) is non-
increasing for λ ∈ [α, λ1], which implies that hypothesis M2 e) of Theorem 4.9 is also
satisfied.
2 ln α−2−ε is a non-increasing function for
On the other hand, for ε ∈ (0, 1) and α ∈ (0, α0), the function gε
Assuming ε ∈ (0, 1), since sε(α)
.
= 1+ln α
.
= (1 + 1
α ∈ (0, α0) and 2α ≤ λ1 ≤ 0.6, we have that
1 + ln α
αgε
α(2α) =
2 ln α − 2−ε ≥
1 + ln 0.3
2 ln 0.3 − 2−ε .
Hence hypothesis M2 d) of Theorem 4.9 is satisfied with constants c1
sε(0.3).
Finally, for every α ∈ (0, α0) the mapping λ → rα(λ)2, λ ∈ (0,kTk2] is
letting
convex and therefore hypothesis M4 of Theorem 4.9 also holds. Hence,
.
= 2 and γ2
.
=
28
T. HERDMAN, R. D. SPIES and K. G. TEMPERINI
.
= √tρ(t) = −
Θ(t)
(ρ ◦ Θ−1)(δ) for x ∈ X ρ
√α0
δ ∈ (0, Θ(α0)) = (cid:16)0,−
√t
ln t for t ∈ (0,kTk2], by Theorem 4.9 we conclude that ψ(x, δ)
= R(ρ(T ∗T )) \ {0} = R(cid:0)−(ln(T ∗T ))−1(cid:1) \ {0} and
ln α0(cid:17), is global saturation function of {Rα}α∈(0,α0) on X ρ.
Example 4: For λ > 0 and α ∈ (0, α0), let gα(λ) be defined as:
.
=
.
gα(λ)
Thus
.
=(0,
e
λ
ln α
λ ,
for 0 ≤ λ < α,
for λ ≥ α.
rα(λ) =(1,
1 − e
λ
ln α ,
for 0 ≤ λ < α,
for λ ≥ α.
It can be immediately shown that {gα}α∈(0,α0) satisfies the hypotheses H1-H4 and
therefore {Rα}α∈(0,α0) with Rα as in (11) is a family of regularization operators for
T x = y. Also it can be easily checked that {Rα}α∈(0,α0) has classical qualification of
order µ0 = 0. Furthermore, it can be proved that the function ρ(α) defined by
ρ(α)
.
= αe
α
ln α
t
3
2
ln t
3
2 e
0 e
is maximal qualification of {Rα}α∈(0,α0) and all hypotheses of Theorem 4.9 are sat-
for t ∈ (0,kTk2], by Theorem 4.9
isfied. Hence, letting Θ(t)
.
= R(ρ(T ∗T )) \ {0} and δ ∈
we conclude that ψ(x, δ)
α0
ln(α0)(cid:17), is global saturation function of {Rα}α∈(0,α0) on X ρ.
= √tρ(t) = t
.
= (ρ ◦ Θ−1)(δ) for x ∈ X ρ .
(0, Θ(α0)) = (cid:16)0, α
6. Conclusions. In this article we have developed a general theory of global
saturation for arbitrary regularization methods for inverse ill-posed problems. This
concept of saturation formalizes the best global order of convergence that a method
can achieve independently of the smoothness assumptions on the exact solution and
on the selection of the parameter choice rule. Necessary and sufficient conditions for
a methods to have global saturation have been provided. It was shown that for a
method to have saturation the total error must be optimal in two senses, namely
as optimal order of convergence over a certain set which at the same time, must be
optimal with respect to the error. We have also proved two converse results and
applied the theory to derive sufficient conditions for the existence of global saturation
for spectral methods with classical qualification of finite positive order and for methods
with maximal qualification. Finally, examples of regularization methods possessing
global saturation were shown.
REFERENCES
1. Engl, H. W., Hanke, M., and Neubauer, A.: Regularization of inverse problems, volume 375
of Mathematics and its Applications. Kluwer Academic Publishers Group, Dordrecht, 1996.
2. T. Herdman, T., Spies, R. D., and Temperini, K. G.: Generalized Qualification and Qualifi-
cation Levels for Spectral Regularization Methods. J. Optim. Theory Appl., 141(3):547 -- 567,
2009.
3. Math´e, P.: Saturation of regularization methods for linear ill-posed problems in Hilbert
spaces. SIAM J. Numer. Anal., 42(3):968 -- 973 (electronic), 2004.
4. Math´e, P., and Pereverzev, S. V.: Optimal discretization of inverse problems in Hilbert
scales. Regularization and self-regularization of projection methods. SIAM J. Numer. Anal.,
38(6):1999 -- 2021 (electronic), 2001.
GLOBAL SATURATION OF REGULARIZATION METHODS
29
5. Math´e, P., and Pereverzev, S. V.: Geometry of linear ill-posed problems in variable Hilbert
scales. Inverse Problems, 19(3):789 -- 803, 2003.
6. Neubauer, A.: On converse and saturation results for regularization methods. In Beitrage
zur angewandten Analysis und Informatik, pages 262 -- 270. Shaker, Aachen, 1994.
7. Neubauer, A.: On converse and saturation results for Tikhonov regularization of linear
ill-posed problems. SIAM J. Numer. Anal., 34(2):517 -- 527, 1997.
8. Seidman, T. I.: Nonconvergence results for the application of least-squares estimation to
ill-posed problems. J. Optim. Theory Appl., 30(4):535 -- 547, 1980.
9. Spies, R. D., and Temperini, K. G: Arbitrary divergence speed of the least-squares method
in infinite-dimensional inverse ill-posed problems. Inverse Problems, 22(2):611 -- 626, 2006.
|
1706.00285 | 1 | 1706 | 2017-06-01T13:08:51 | A fresh approach to the Paley-Wiener theorem for Mellin transforms and the Mellin-Hardy spaces | [
"math.FA"
] | Here we give a new approach to the Paley--Wiener theorem in a Mellin analysis setting which avoids the use of the Riemann surface of the logarithm and analytical branches and is based on new concepts of "polar-analytic function" in the Mellin setting and Mellin--Bernstein spaces. A notion of Hardy spaces in the Mellin setting is also given along with applications to exponential sampling formulas of optical physics. | math.FA | math |
A fresh approach to the Paley–Wiener
theorem for Mellin transforms and the
Mellin–Hardy spaces
Carlo Bardaro, ∗
Paul L. Butzer, †
Gerhard Schmeisser§
Ilaria Mantellini, ‡
June 2, 2017
Abstract. Here we give a new approach to the Paley–Wiener theorem in a Mellin
analysis setting which avoids the use of the Riemann surface of the logarithm and
analytical branches and is based on new concepts of polar-analytic function in the
Mellin setting and Mellin–Bernstein spaces. A notion of Hardy spaces in the Mellin
setting is also given along with applications to exponential sampling formulas of
optical physics.
AMS Subject Classification. 44A05, 30D20, 26D10
KeyWords. Mellin transforms, Paley–Wiener spaces, Riemann surfaces, Paley–
Wiener theorem, Mellin–Bernstein spaces, polar-analytic functions
Introduction
1
The structure of the Paley–Wiener space of all continuous functions f ∈ L2(R)
having compactly supported Fourier transform is precisely described by the classical
Paley–Wiener theorem of Fourier analysis. This basic result characterizes the Paley–
Wiener spaces by the Bernstein spaces comprising all functions f ∈ L2(R) which
have an analytic extension to the whole complex plane and are of exponential type
(the type being connected with the bandwidth of the Fourier transform); see, e.g.,
∗Department of Mathematics and Computer Sciences, University of Perugia, via Vanvitelli 1,
†Lehrstuhl A fuer Mathematik, RWTH Aachen, Templergraben 55, Aachen, D-52056, Germany,
‡Department of Mathematics and Computer Sciences, University of Perugia, via Vanvitelli 1,
§Department Mathematik, FAU Erlangen-Nurnberg, Cauerstr. 11, 91058 Erlangen, Germany,
I-06123 Perugia, Italy, e-mail: [email protected]
I-06123 Perugia, Italy, e-mail: [email protected]
e-mail: [email protected]
email: [email protected]
1
[7], [27], [22], [23], [17]. Moreover, it has a wide range of applications, especially
in sampling theory and related fields; see [17], [20], [24]. Numerous variants of this
theorem have been proved by several authors [1], [2], [26]. In particular, analogues
of the Paley–Wiener theorem were obtained for other integral transforms. Recently,
in [5] we have proved a version for Mellin transforms, by introducing the notion
of a Mellin–Bernstein space, comprising all functions f ∈ X 2
c := {f : R+ → C :
(·)(·)c−1/2 ∈ L2(R+)} which have an analytic extension to the Riemann surface
of the (complex) logarithm and satisfy some exponential-type condition. We gave
two different approaches, one involving purely complex analysis arguments and the
other one using "real" arguments based on the statement of a Mellin extension of
the classical Bernstein inequality also proved in [5]. Later on, in [6] we applied our
Paley–Wiener theorem in the Mellin setting to the study of the so-called Mellin
distance of functions belonging to certain functional spaces (like Lipschitz spaces,
Mellin–Sobolev spaces, and so on) from the Mellin–Bernstein space. This leads
to precise estimates of the approximation error in certain basic formulae valid in
Mellin–Bernstein spaces such as the exponential sampling formula and the Mellin
reproducing kernel formula; see, e.g., [9], [10], [3], [4]. The results in [6] extend the
corresponding ones in Fourier analysis (see [12], [13]) to the Mellin frame.
This paper is concerned with equivalent formulations of the Paley–Wiener the-
orem in the Mellin setting, its content being fully different from our papers [5]
and [6]. It is a new and simpler approach as it avoids the use of an abstract Rie-
mann surface and analytic branches. It employs the helicoidal surface as a model
of the Riemann surface of the (complex) logarithm and a notion of analytic func-
tions on it. These considerations lead us to a concept of "polar analyticity" which
enables us to introduce in a simple way Bernstein classes and Hardy spaces in a
Mellin setting. Moreover, the proofs of the main theorems are notably different
from those of [5]. Concerning the notion of polar analyticity of a function f at a
point (r0, θ0) ∈ H := {(r, θ) ∈ R+ × R}, here discussed, we introduce it by means of
the limit (see Sec. 3, Definition 3)
lim
(r,θ)→(r0,θ0)
f (r, θ) − f (r0, θ0)
reiθ − r0eiθ0
.
Equivalently, polar-analytic functions can be described by the well-known Cauchy–
Riemann equations with respect to polar coordinates; see, e.g., [19], [14]. The
classical Hardy spaces are treated in many books; see, e.g., [22], [21], [15].
In Section 2, we give some basic notions concerning Mellin analysis and prelimi-
nary results, including the Paley–Wiener theorem in the Mellin setting stated in [5].
In Subsection 2.1, we illustrate the equivalent formulation in terms of the helicoidal
surface. In Sections 3 and 4, we study polar-analytic functions and Mellin–Bernstein
classes. The Paley–Wiener theorem is stated in Section 5. Sections 6 and 7 are de-
voted to the Hardy-type spaces in the Mellin frame and the study of the Mellin
distance of a function f belonging to a Hardy space from the Mellin–Bernstein
2
class. In the final section, we apply the results to estimates of the approximation
error in the exponential sampling formula.
2 Basic notions and preliminary results
Let C(R+) be the space of all continuous functions defined on R+, and C (r)(R+) be
the space of all functions in C(R+) with a derivative of order r in C(R+). Analo-
gously, by C∞(R+) we denote the space of all infinitely differentiable functions. By
loc(R+), we denote the space of all measurable functions which are integrable on
L1
every bounded interval in R+.
For 1 ≤ p < +∞, let Lp(R+) be the space of all Lebesgue measurable and
p-integrable complex-valued functions defined on R+ endowed with the usual norm
(cid:107)f(cid:107)p. Analogous notations hold for functions defined on R.
For p = 1 and c ∈ R, let us consider the space (see [8])
Xc = {f : R+ → C : f (·)(·)c−1 ∈ L1(R+)}
endowed with the norm
(cid:107)f(cid:107)Xc := (cid:107)f (·)(·)c−1(cid:107)1 =
(cid:90) +∞
f (u)uc−1du.
0
space with respect to the (invariant) measure µ(A) =(cid:82)
µ(R+), where Lp
More generally, let X p
f (·)(·)c−1/p ∈ Lp(R+) with 1 < p < ∞. In an equivalent form, X p
all functions f such that (·)cf (·) ∈ Lp
A ⊂ R+. Finally, for p = ∞, we define X∞
functions f : R+ → C such that (cid:107)f(cid:107)X∞
The Mellin translation operator τ c
c denote the space of all functions f : R+ → C such that
c is the space of
µ(R+) denotes the Lebesgue
A dt/t for any measurable set
c as the space comprising all measurable
:= supx>0 xcf (x) < ∞. For p = 2 see [11].
h, for h ∈ R+, c ∈ R, f : R+ → C, is denoted
c
by
hf )(x) := hcf (hx) (x ∈ R+).
(τ c
hf )(x) = hc(τhf )(x) and (cid:107)τ c
h, we have (τ c
hf(cid:107)Xc = (cid:107)f(cid:107)Xc.
Setting τh := τ 0
In the Mellin frame, the natural concept of a pointwise derivative of a function
f is given by the limit of the difference quotient involving the Mellin translation;
thus if f(cid:48) exists,
hf (x) − f (x)
τ c
h − 1
lim
h→1
= xf(cid:48)(x) + cf (x).
This gives the motivation for the following definition (see [8]): The pointwise Mellin
differential operator Θc, or the pointwise Mellin derivative Θcf of a function f :
R+ → C and c ∈ R, is defined by
Θcf (x) := xf(cid:48)(x) + cf (x)
(x ∈ R+)
3
provided that f(cid:48) exists a.e. on R+. The Mellin differential operator of order r ∈ N
is defined recursively by
Θ1
c := Θc,
c := Θc(Θr−1
Θr
c
).
For convenience, set Θr := Θr
operator. For instance, the first three Mellin derivatives are given by:
0 for c = 0 and Θ0
c := I with I denoting the identity
Θcf (x) = xf(cid:48)(x) + cf (x),
cf (x) = x2f(cid:48)(cid:48)(x) + (2c + 1)xf(cid:48)(x) + c2f (x),
Θ2
cf (x) = x3f(cid:48)(cid:48)(cid:48)(x) + (3c + 3)x2f(cid:48)(cid:48)(x)
Θ3
+(3c2 + 3c + 1)xf(cid:48)(x) + c3f (x).
The Mellin transform of a function f ∈ Xc is the linear and bounded operator
defined by (see, e.g., [18], [16], [8])
Mc[f ](s) ≡ [f ]∧
Mc(s) :=
us−1f (u)du (s = c + it, t ∈ R).
The inverse Mellin transform M−1
[g] of a function g ∈ L1({c} × iR), is defined by:
M−1
c
[g](x) :=
x−c
2π
g(c + it)x−itdt
(x ∈ R+),
where in general Lp({c} × iR), for p ≥ 1, will mean the space of all functions
g : c + iR → C with g(c + i·) ∈ Lp(R+).
We have the following preliminary results (see [8], [3]):
Lemma 1 (Inversion Theorem in Xc) If f ∈ Xc is such that Mc[f ] ∈ L1({c} ×
iR), then
(cid:90) +∞
0
c
(cid:90) +∞
−∞
M−1
c
[Mc[f ]](x) =
x−c
2π
(cid:90) +∞
−∞
[f ]∧
Mc(c + it)x−itdt = f (x)
(a.e. on R+).
The following lemma will enable us to work in a practical Hilbert space setting.
Lemma 2 If f ∈ Xc and Mc[f ] ∈ L1({c} × iR), then f ∈ X 2
c .
More generally, for 1 < p ≤ 2, the Mellin transform M p
c of f ∈ X p
c is given by
(see [11])
c [f ](s) ≡ [f ]∧
M p
M p
c
(s) = l.i.m.ρ→+∞
4
(cid:90) ρ
1/ρ
f (u)us−1du,
for s = c + it, in the sense that
(cid:13)(cid:13)(cid:13)(cid:13)M p
lim
ρ→∞
c [f ](c + it) −
(cid:90) ρ
1/ρ
(cid:13)(cid:13)(cid:13)(cid:13)Lp(cid:48)
f (u)us−1du
= 0,
({c}×iR)
where p(cid:48) is the conjugate exponent of p, that is, 1/p + 1/p(cid:48) = 1. In the following we
are interested in the case p = 2.
Analogously, we define the inverse Mellin transform of a function g ∈ X 2
c by
M 2,−1
c
[g](x) = l.i.m.ρ→+∞
1
2π
and for any f ∈ X 2
c , there holds
(cid:90) ρ
1/ρ
g(c + it)x−c−itdt,
M 2,−1
c
[M 2
c [f ]](x) = f (x)
(a.e. on R+);
see [11].
For functions in Xc∩X 2
c , we have the following important "consistency" property
of the Mellin transform (see [11]):
Lemma 3 If f ∈ Xc ∩ X 2
i.e., Mc[f ](c + it) = M 2
c , then the Mellin transforms Mc[f ] and M 2
c [f ] coincide,
c [f ](c + it) for almost all t ∈ R.
Moreover, the following Mellin version of the Plancherel Theorem holds (see [11,
Lemma 2.6]):
c from X 2
c onto L2({c} × iR) is bounded and norm pre-
Lemma 4 The operator M 2
serving, i.e., for f ∈ X 2
c , we have
(cid:107)f(cid:107)X 2
c =
1√
2π
(cid:107)M 2
c [f ](cid:107)L2({c}×iR).
2.1 The Mellin–Bernstein spaces via Riemann surfaces
We begin by introducing the following spaces:
M 2
c
c,T denote the space of all functions f ∈ X 2
(c + it) = 0 a.e. for t > T. Analogously, by B1
Definition 1 Let B2
[f ]∧
functions f ∈ Xc ∩ C(R+) such that [f ]∧
These spaces are called Paley–Wiener spaces. By Lemma 2, we have B1
and it is easily seen that B1
Mc(c + it) = 0 for all t > T.
c ∩ C(R+) such that
c,T we denote the space of all
c,T ⊂ X 2
c ,
c,T ⊂ B2
c,T .
In [5] we have seen that a Mellin bandlimited function cannot be extended to
the whole complex plane as an entire function, but we have proved that it has an
analytic extension on the Riemann surface Slog of the logarithm.
Consider a function g analytic on Slog. Then g can be split into branches gk (k ∈
Z) that are analytic on the slit complex plane Ω := C \ R+. Furthermore, they are
connected on the slit by the following properties:
5
(i) For x > 0 the limits
g+
k (x) := lim
ε→0+
gk(x + iε) and g−
k (x) := lim
ε→0+
gk(x − iε)
exist and
g−
k (x) = g+
k+1(x),
g+
0 (x) = g(x)
(x > 0).
(1)
(ii) For x > 0, let Ux be an open disk in the right half-plane with center at x.
Then ψk : Ux → C with
gk(z)
The Mellin–Bernstein space (cid:101)B2
is analytic.
ψk(z) :=
g−
k (z)
gk+1(z)
for
for
for
z ∈ Ux, (cid:61)z < 0,
z ∈ Ux ∩ R,
z ∈ Ux, (cid:61)z > 0
c,T comprises all functions f ∈ X 2
c for which g(x) :=
xcf (x) has an analytic extension on Slog with branches gk on Ω satisfying the fol-
lowing (additional) conditions:
(iii) There exists a constant C > 0 such that for all k ∈ Z and θ ∈ [0, 2π]
gk(reiθ) ≤ CeT2πk+θ
(r > 0).
(iv) For θ ∈ [0, 2π], we have
lim
r→0
gk(reiθ) = lim
r→∞ gk(reiθ) = 0
uniformly with respect to θ.
(2)
(3)
In (2) and (3) the value gk(reiθ) has to be defined as g+
when θ = 2π.
k (r) when θ = 0 and as g−
k (r)
Now the Paley–Wiener theorem for the Mellin transform can be stated as follows
(see [5]):
Theorem 1 (Paley–Wiener)
A simple and convenient model for Slog is the helicoidal surface in R3 defined by
c,T = B2
c,T .
E := {(x1, x2, x3) ∈ R3 : x1 = r cos θ, x2 = r sin θ, x3 = θ, r > 0, θ ∈ R}.
The subset obtained by setting θ = 0 on the right-hand side can be interpreted as
R+. Just as C is an extension of R, we shall see that E takes a corresponding role
for R+ in Mellin analysis. For α, β ∈ R with α < β, we consider the surface
6
(cid:101)B2
Figure 1: The helicoidal surface as a model of the Riemann surface of the logarithm.
Eα,β :=(cid:8)(x1, x2, x3) ∈ R3 : x1 = r cos θ, x2 = r sin θ, x3 = θ, r > 0, θ ∈ ]α, β[(cid:9)
and call it a segment of E. The projection of Eα,β into the (x1, x2)-plane may be
interpreted as a sector in C, given by
Sα,β := {x1 + ix2 ∈ C : x1 = r cos θ, x2 = r sin θ, r > 0, θ ∈ ]α, β[} .
When β − α ∈ ]0, 2π], then this projection is a bijection. Indeed, given z ∈ Sα,β,
there exists a unique θ ∈ ]α, β[, denoted by θ := argαz, such that z = zeiargαz and
then ((cid:60)z,(cid:61)z, argαz) is the pre-image of z on Eα,β. Now, equivalently to the usual
abstract approach (see [5]), analytic functions on E may be introduced as follows.
Definition 2 A function f : E → C is said to be analytic if for every segment Eα,β
with β − α ∈ ]0, 2π] the function z (cid:55)−→ f ((cid:60)z,(cid:61)z, argαz) is analytic on Sα,β.
As an example, the function L defined by
L(r cos θ, r sin θ, θ) := log r + iθ
(r > 0, θ ∈ R),
is analytic on E and coincides on R+ with the logarithm of real analysis.
Given α, the largest admissible β in the previous definition is α + 2π. For k ∈ Z,
all the sectors Sα+2kπ,α+2(k+1)π coincide with the complex plane slit along the ray
z = reiα(r > 0) and the analytic functions induced by f on these sectors are the
analytic branches of f. In this setting, the Paley–Wiener theorem for Mellin band-
limited functions in the previous subsection may be restated as follows.
Theorem 2 (Paley–Wiener) A function ϕ ∈ X 2
space B2
properties:
c belongs to the Paley–Wiener
c,T if and only if there exists a function f : E → C with the following
7
(i) f is analytic on E;
(ii) f (r, 0, 0) = ϕ(r)
(r > 0);
(iii) there exists a constant Cf such that
f (r cos θ, r sin θ, θ) ≤ Cf r−ceTθ
(θ ∈ R);
(iv) limr→0 rcf (r cos θ, r sin θ, θ) = limr→∞ rcf (r cos θ, r sin θ, θ) = 0 uniformly with
respect to θ on all compact subintervals of R.
The proof is essentially the same as that of Theorem 3 in [5], using the restrictions
of the function f to the sectors S2kπ,2(k+1)π as the analytic branches of f.
3 Analytic functions over the polar plane
In our previous approach with the helicoidal surface we have functions of three
variables x1, x2, x3 but we use them only as functions of the two variables r and θ.
Since there exists a bijection between the helicoidal surface and the right half-plane
understood as the set of all points (r, θ) with r > 0 and θ ∈ R, one may think
of considering functions defined on the right half-plane. However, these functions
will no longer be analytic in the classical sense. They are differentiable and satisfy
the Cauchy–Riemann equations transformed into polar coordinates. This approach
amounts to taking an analytic function, writing its variable in polar coordinates
z = reiθ and treating (r, θ) as if they were Cartesian coordinates.
Let H := {(r, θ) ∈ R+ × R} be the right half-plane and let D be a domain in H.
Definition 3 1 We say that f : D → C is polar-analytic on D if for any (r0, θ0) ∈ D
the limit
lim
(r,θ)→(r0,θ0)
=: (Dpolf )(r0, θ0)
f (r, θ) − f (r0, θ0)
reiθ − r0eiθ0
exists and is the same howsoever (r, θ) approaches (r0, θ0) within D.
For a polar-analytic function f we define the polar Mellin derivative as
Θcf (r, θ) := reiθ(Dpolf )(r, θ) + cf (r, θ).
1As far as the authors are aware, this modified notion of analyticity arising by treating the polar
coordinates as cartesian coordinates, has as yet not been presented. Our definition leads naturally
to the classical Cauchy-Riemann equations when written in their polar form, often treated in the
literature. Although other mathematicians may have come across this concept too, it seems that
it has not been used for practical purposes so far. In Mellin analysis it turns out to be very helpful
for an efficient approach, independent of Fourier analysis. In particular it leads to a precise and
simple analysis for functions defined over the Riemann surface of the complex logarithm, via the
helicoidal surface.
8
Remark 1 It can be verified that f = u + iv with u, v : D → R is polar-analytic
on D if and only if u and v have continuous partial derivatives on D that satisfy the
differential equations
∂u
∂θ
= −r
∂v
∂r
,
∂v
∂θ
= r
∂u
∂r
.
(4)
Note that these equations coincide with the Cauchy-Riemann equations of an ana-
lytic function g defined by g(z) := u(r, θ) + iv(r, θ) for z = reiθ. For the derivative
Dpol we easily find that
(Dpolf )(r, θ) = e−iθ
v(r, θ) − i
(cid:20) ∂
(cid:20) ∂
u(r, θ)
.
u(r, θ) + i
v(r, θ)
=
(cid:21)
(cid:21)
e−iθ
r
∂θ
∂r
∂
∂r
∂
∂θ
Also note that Dpol is the ordinary differentiation on R+. More precisely, if ϕ(·) :=
f (·, 0) then (Dpolf )(r, 0) = ϕ(cid:48)(r).
Moreover, for θ = 0 we obtain the known formula for ϕ
Θcϕ(r) = rϕ(cid:48)(r) + cϕ(r).
When g is an entire function, then f : (r, θ) (cid:55)→ g(reiθ) defines a function f on H
that is polar-analytic and 2π-periodic with respect to θ. The converse is also true.
However, there exist polar-analytic functions on H that are not 2π-periodic with
respect to θ. A simple example is the function L(r, θ) := log r + iθ, which is easily
seen to satisfy the differential equations (4).
A connection with the analytic functions in classical sense can be established by
a suitable substitution (see the proof of Theorem 3 below). Using this, we find that
every polar-analytic function f has a series expansion of type
∞(cid:88)
f (r, θ) =
an(log r + iθ)n,
convergent everywhere on H.
Now, for α, β ∈ R with α < β, we consider the set
n=o
Hα,β := {(r, θ) ∈ R+ × R : θ ∈]α, β[}
and call it a strip of H.
If f : H → C is polar-analytic but not 2π-periodic with respect to θ, then we
can associate with f a function g that is analytic on the Riemann surface Slog of
the logarithm. The restriction of f to a strip Hα+2kπ,α+2(k+1)π, where k ∈ Z, defines
an analytic function gk in the slit complex plane C \ {reiα : r > 0} by setting
gk(reiθ) := f (r, θ). The functions gk for k ∈ Z are the analytic branches of g.
We now study line integrals for polar-analytic functions. Here a piecewise contin-
uously differentiable curve will be called a regular curve. The following proposition
will be useful in what follows.
9
Proposition 1 Let f be a polar-analytic function on H and let (r1, θ1) and (r2, θ2)
be any two points in H. Then the line integral
f (r, θ)eiθ(dr + irdθ)
(5)
(cid:90)
γ
(cid:20)
∂
∂θ
has the same value for each regular curve γ in H that starts at (r1, θ1) and ends
(r2, θ2). In particular, the integral vanishes for closed regular curves.
Proof. Recalling equations (4), we easily verify that
(cid:21)
(cid:20)
=
∂
∂r
(cid:21)
f (r, θ)eiθ
f (r, θ)ireiθ
.
By a theorem of Schwartz, this implies that the integrand in (5) is an exact differ-
ential on H, that is, there exists a function F : H → C such that
∂F
∂r
(r, θ) = f (r, θ)eiθ
and
(r, θ) = f (r, θ)ireiθ,
∂F
∂θ
and so the integral in (5) is equal to F (r2, θ2) − F (r1, θ1).
2
4 The Mellin–Bernstein classes
In this section we introduce Mellin–Bernstein spaces in a somewhat different way
closer to the classical one of Fourier analysis.
Definition 4 For c ∈ R, T > 0 and p ∈ [1, +∞[ the Mellin–Bernstein space Bp
comprises all functions f : H → C with the following properties:
(i) f is polar-analytic on H;
(ii) f (·, 0) ∈ X p
c ;
c,T
(iii) there exists a positive constant Cf such that
f (r, θ) ≤ Cf r−ceTθ
((r, θ) ∈ H).
The above definition has two useful consequences, which are described in the fol-
lowing
Theorem 3 Let f ∈ Bp
(i) f (·, θ) ∈ X p
c,T with p ∈ [1, +∞]. Then the following statements hold:
c for all θ ∈ R and (cid:107)f (·, θ)(cid:107)X p
c ≤ eTθ(cid:107)f (·, 0)(cid:107)X p
c ;
(ii) limr→0 rcf (r, θ) = limr→+∞ rcf (r, θ) = 0 uniformly with respect to θ on all
compact subinterval of R.
10
Proof. The function g : x + iy → f (ex, y) is defined on C. Writing f = u + iv
with real-valued functions u and v, we know that the differential equations (4) hold.
Therefore g satisfies the Cauchy–Riemann equations on C, and so g is an entire
function. Consequently
F : x + iy (cid:55)−→ ec(x+iy)f (ex, y)
is also an entire function. Property (iii) of Definition 4 may now be rewritten in
terms of F as
F (x + iy) ≤ Cf eTy,
(6)
which shows that F is an entire function of exponential type T. Moreover, property
(ii) of Definition 4 shows by a substitution of variables that
F (x)pdx = (cid:107)f (·, 0)(cid:107)p
(cid:90) +∞
F (x + iy)pdx ≤ epTy(cid:90) +∞
−∞
X p
c
(cid:90) +∞
−∞
exists. Now using [7, Theorem 6.7.1], we have
F (x)pdx
(7)
−∞
for all y ∈ R and F (x) → 0 as x → ±∞. Therefore assertion (i) follows from (7) by
the substitution ex (cid:55)→ r and y (cid:55)→ θ.
Next we note that in view of (7), the convergence F (x) → 0 extends to F (x +
iy) → 0 as x → ±∞ pointwise for each y ∈ R. In connection with (6), a result in [7,
Theorem 1.4.9] guarantees that the convergence is even uniform with respect to y
on compact subintervals of R; also see [25, p. 170]. Writing this statement in terms
2
of f, we find that assertion (ii) is also true.
5 The Paley–Wiener theorem for B2
c,T
We are ready to state and prove an alternative version of the Paley–Wiener theorem,
in terms of the Mellin–Bernstein space B2
Theorem 4 (Paley-Wiener) A function ϕ ∈ X 2
space B2
Proof. First, suppose that ϕ ∈ B2
c,T if and only if there exists a function f ∈ B2
c belongs to the Paley–Wiener
c,T such that f (·, 0) = ϕ(·).
c,T . Then, by the inversion formula, we have
c,T .
(cid:90) T
−T
ϕ(r) =
1
2π
[ϕ]∧
M 2
c
(c + it)r−c−itdt
(r > 0).
11
This imples that
f (r, θ) :=
1
2π
[ϕ]∧
M 2
c
(c + it)(reiθ)−c−itdt.
(8)
(cid:90) T
−T
f (r, θ) ≤ r−ceTθ
(cid:90) T
2π
Cf :=
1
2π
−T
[ϕ]∧
M 2
c
(c + it)dt =
[ϕ]∧
M 2
c
(c + it)dt.
(cid:90) +∞
−∞
1
2π
[ϕ]∧
M 2
c
(c + it)dt.
Hence (iii) of Definition 4 holds with
(cid:90) T
−T
(cid:90) T
−T
We now extend ϕ to the (r, θ)-plane with r > 0 and θ ∈ R, by replacing r with reiθ
on the right-hand side. Denoting this extension by f (r, θ), we have
Concerning polar analyticity of f, for a fixed point (r0, θ0) ∈ H, we consider the
difference quotient
f (r, θ) − f (r0, θ0)
reiθ − r0eiθ0
=
1
2π
[ϕ]∧
M 2
c
(c + it)
(reiθ)−c−it − (r0eiθ0)−c−it
reiθ − r0eiθ0
dt.
The limit (r, θ) → (r0, θ0) carried out inside the integral leads to an ordinary dif-
ferentiation of z−c−it with respect to z at the point z0 := r0eiθ0. We have to justify
that the limit and the integration can be interchanged. In order to do that, let us
take a closed rectangle Q centered at the point (r0, θ0) such that Q ⊂ H. We note
that the function h : Q × [−T, T ] → C, defined by
(reiθ)−c−it − (r0eiθ0)−c−it
for (r, θ) (cid:54)= (r0, θ0),
h(r, θ, t) :=
reiθ − r0eiθ0
and
h(r0, θ0, t) := −(c + it)(r0eiθ0)−c−1−it,
is continuous and therefore its absolute value is bounded on Q× [−T, T ]. Hence the
desired interchange is guaranteed by Lebesgue's theorem of dominated convergence.
c,T . For any fixed t ∈ R, define
Now we prove the reverse implication. Let f ∈ B2
the function
g(r, θ) := e(c−1+it)(log r+iθ)f (r, θ),
which is polar-analytic in H.
Let us first assume that t > T. For R > 1, we define n := (cid:98)log R(cid:99), where for any
real number a the symbol (cid:98)a(cid:99) denotes the largest integer not exceeding a. We now
consider the regular curve γ which is the boundary of the rectangle in H defined by
[1/R, R] × [0, 2nπ]. Then, using Proposition 1, one has
(cid:90)
g(r, θ)eiθ(dr + irdθ) = 0.
γ
12
(cid:90) R
This easily implies that
(cid:90) R
(cid:90) 2nπ
1/R
+
i
R
1/R
e(c−1+it)(− log R+iθ)+iθf
0
=: I1 + I2 − I3,
where ϕ is the restriction of f to R+.
rc+it−1ϕ(r)dr =
e(c−1+it)(log r+2nπi)f (r, 2nπ)dr
(cid:18) 1
(cid:19)
, θ
R
dθ − iR
(cid:90) 2nπ
0
e(c−1+it)(log R+iθ)+iθf (R, θ)dθ
We now estimate the integrals on the right-hand side. Using (iii) of Definition 4,
we find that
I1 ≤
e(c−1) log re−2nπtf (r, 2πn)dr
(cid:90) R
(cid:90) R
1/R
= Cf e−(t−T )2nπ log R2.
Thus, according to our choice of n, we obtain I1 → 0 as R → +∞.
e−(t−T )2nπ dr
r
≤ Cf
1/R
As to I2, we easily see that
(cid:90) 2nπ
0
(cid:18) 1
(cid:12)(cid:12)(cid:12)(cid:12)f
, θ
R
1
Rc
I2 ≤
(cid:90) +∞
0
(cid:18) 1
, θ
R
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)e−tθdθ.
Now, since the integrand is dominated by Cf e−(t−T )θ, which is integrable as a func-
tion of θ, by (ii) of Theorem 3 and the theorem of dominated convergence, we obtain
I2 → 0 as R → +∞. The same result is obtained for I3. Altogether we have
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)e−tθdθ ≤
(cid:12)(cid:12)(cid:12)(cid:12)(cid:90) R
1/R
1
Rc
(cid:12)(cid:12)(cid:12)(cid:12)f
(cid:12)(cid:12)(cid:12)(cid:12) = 0.
lim
R→+∞
rc−1+itϕ(r)dr
This implies that [ϕ]∧
When t < −T , we choose n := −(cid:98)log R(cid:99) and proceed analogously using the
boundary of the rectangle [1/R, R] × [−2nπ, 0]. Hence ϕ(·) := f (·, 0) belongs to
B2
2
c,T .
(c + it) = 0.
M 2
c
6 A Hardy space in Mellin analysis
In this section we apply our approach in order to define in a simple way a Hardy
space in the Mellin frame. For c ∈ R and p ∈ [1, +∞[, we recall that the norm in
X p
c is defined by
(cid:107)ϕ(cid:107)X p
c =
ϕ(r)prcp−1dr
For a strip H−a,a with a > 0, we simply write Ha.
(cid:19)1/p
.
(cid:18)(cid:90) +∞
0
13
Definition 5 Let a, c, p ∈ R with a > 0 and p ≥ 1. The Mellin–Hardy space H p
comprises all functions f : Ha → C that satisfy the following conditions:
(i) f is polar-analytic on Ha;
(ii) f (·, θ) ∈ X p
c for each θ ∈ ] − a, a[;
c (Ha)
(iii) there holds
(cid:107)f(cid:107)H p
c (Ha) := sup
0<θ<a
(cid:18)(cid:107)f (·, θ)(cid:107)p
X p
c
(cid:19)1/p
< +∞.
+ (cid:107)f (·,−θ)(cid:107)p
2
X p
c
When a ∈ ]0, π] we can associate with each function f ∈ Ha a function g analytic on
the sector Sa := {z ∈ C : arg z < a} by defining g(reiθ) := f (r, θ). The collection
c (Sa), which may be identified
of all such functions constitutes a Hardy-type space H p
with H p
strip Sa := {z ∈ C : (cid:61)z < a}. The Hardy space H p
way that if g ∈ H p(Sa) and
In [13] the authors considered a Hardy space H p(Sa) of functions analytic on the
c (Ha) has been designed in such
c (Ha).
f (r, θ) := r−ce−icθg(log r + iθ),
then f ∈ H p
c (Ha) and conversely, if f ∈ H p
c (Ha) and
then g ∈ H p(Sa).
g(x + iy) := ec(x+iy)f (ex, y),
Using this correspondence, one can deduce the following propositions from the
analogous ones proved in [13].
Proposition 2 Let f ∈ H p
c (Ha) and let a1 ∈ ]0, a[. Then
(cid:18)
(cid:19)1/p(cid:107)f(cid:107)H p
c (Ha)
f (r, θ) ≤ r−c
4
π(a − a1)
for (r, θ) ∈ Ha1.
The next proposition is a Nikol'ski-type inequality for the Hardy space H p
Proposition 3 Let f ∈ H p
tn+1/tn > e2δ for all n ∈ Z with δ ∈ ]0, a[. Then
c (Ha) and let (tn)n∈Z be a sequence on R+ such that
c (Ha).
(cid:18)(cid:88)
n∈Z
n f (tn, 0)p
tcp
(cid:19)1/p ≤
(cid:18) 2
(cid:19)1/p(cid:107)f(cid:107)H p
πδ
c (Ha).
14
Proposition 4 Let f ∈ H p
c (Ha) and let a1 ∈ ]0, a[. Then
r→+∞ rcf (r, θ) = 0
lim
r→0+
uniformly for θ ∈ [−a1, a1].
rcf (r, θ) = lim
c (Ha) from the Mellin–Bernstein space Bp
The following theorem will be needed for estimating the distance of a function
f ∈ H p
c,σ (see the next section). Here we
will restrict ourselves to p ∈ {1, 2}.
Theorem 5 Let p ∈ {1, 2} and f ∈ H p
c (Ha). Then for α ∈ ]0, a[, we have
(a.e. f or t ∈ R),
(c + it)
M p
c
[f (·, 0)]∧
(c + it) = e−αt[f (·, εα)∧
where ε = 1 for t > 0 and ε = −1 for t < 0.
Proof. For f ∈ H p
M p
c
c (Ha) let us consider the function g defined by
g(r, θ) := e(c−1+it)(log r+iθ)f (r, θ),
which is polar-analytic on Ha. For sake of simple notation, we do not indicate the
dependence of g on c and t explicitely.
First we assume that t > 0. For ρ > 1, let γ be the positively oriented rectangular
curve in H with vertices at (1/ρ, 0), (ρ, 0), (1/ρ, α) and (ρ, α). By Proposition 1, we
have
g(r, θ)eiθ(dr + irdθ) = 0.
This equation may be rewritten in terms of ordinary integrals as
g(r, 0)dr =
g(r, α)eiαdr + I+
, t
− I+(ρ, t),
(cid:90)
γ
(cid:90) ρ
1/ρ
(cid:90) ρ
1/ρ
I+(r, t) =
where
and so
Now, for p = 2 the integrals
(cid:19)
(cid:18) 1
(cid:90) α
ρ
(cid:90) α
0
0
e−tθf (r, θ)dθ.
(cid:90) ρ
1/ρ
g(r, α)eiαdr
(cid:90) α
0
(cid:90) ρ
1/ρ
I+(r, t) ≤ rc
g(r, 0)dr
and
g(r, θ)ireiθdθ = irc+it
e−(t−ic)θf (r, θ)dθ,
converge in the L2-sense to the Mellin transforms
[f (·, 0)]∧
M p
c
(c + it) and e−αteiαc[f (·, α)]∧
M p
c
(c + it),
15
For t < 0, we use the rectangular curve γ with vertices at (1/ρ, 0), (ρ, 0), (1/ρ,−α)
respectively, while for p = 1 they converge uniformly.
and (ρ,−α), and proceed analogously. In any case, we obtain for all t ∈ R,
− I(ρ, t),
rc−1+itf (r, 0)dr = e−αteiεαc
rc−1+itf (r, εα)dr + I
(cid:18) 1
(cid:90) ρ
(cid:19)
(9)
, t
ρ
I(r, t) ≤ rc
e−θtf (r, θ)dθ.
Using Proposition 4, we can show that (cid:107)I(r,·)(cid:107)Lp(cid:48)
approaches 0 as r → 0 or r → +∞. Hence (9) implies the assertion.
(R) exists for p(cid:48) ∈ {2,∞} and
2
1/ρ
where
7 Estimates of the distance from Mellin–Bernstein
spaces
In [6] we introduced a notion of distance in terms of the Mellin transform. For c ∈ R
c be the linear space of all functions f : R+ → C that have
and q ∈ [1, +∞], let Gq
the representation
(cid:90) ρ
(cid:90) α
1/ρ
−α
(cid:90) ∞
−∞
f (x) =
1
2π
ψ(v)x−c−ivdv
(x > 0),
where ψ ∈ L1(R) ∩ Lq(R). We endowed this space with the norm
(cid:18)(cid:90)
(cid:19)1/q
[[f ]]q := (cid:107)ψ(cid:107)Lq(R) =
ψ(v)qdv
R
and the corresponding metric is defined by
distq(f, g) := [[f − g]]q,
f, g ∈ Gq
c.
c for any q ≥ 1, since its
c,σ, for p ∈ [1, 2], then f (·, 0) ∈ Gq
Note that if f ∈ Bp
Mellin transform has compact support, and therefore, from the uniqueness theorem
of Mellin transforms, one can take ψ(v) = [f (·, 0)]∧
In polar form, we can define the class (cid:101)Gq
Another important subspace is the Mellin inversion class, denoted by Mp
functions on H such that f (·, 0) ∈ Gq
c.
p ∈ [1, 2], consisting of all functions f ∈ X p
This class is contained in Gq
p.
c, with
∈ L1({c}× iR).
c for q ∈ [1, p(cid:48)] with p(cid:48) being the conjugate exponent of
For p ∈ {1, 2}, the Mellin–Hardy space defined above contains the space Bp
c,σ.
Moreover, using Proposition 2 and Theorem 5, we see that the restriction to the
positive real line of a function in the Hardy space is an element of Mp
c.
c ∩ C(R+) such that [f ]∧
c as the space of all polar-analytic
(c + iv).
M p
c
M p
c
16
For a subspace A ⊂ Gq
c we define
distq(f, A) := inf
g∈A
[[f − g]]q.
In [6] we obtained a representation formula for the distance of a function f ∈ Gq
from the Paley–Wiener space Bp
c,σ in the form
c
(cid:18)(cid:90)
(cid:19)1/q
distq(f, Bp
c,σ) =
ψ(v)qdv
v≥σ
(1 ≤ q < ∞)
and if ψ is also continuous, then
dist∞(f, Bp
c,σ) = sup
v≥σ
ψ(v).
Moreover, in the same paper, we estimated the distance in case of Mellin inversion
classes, Lipschitz spaces and Mellin-Sobolev spaces.
Now, we define the distance distq(f, Bp
c (Ha), with a > 0,
c,σ, by considering the restrictions to the positive real line of the
c,σ) of a function f ∈ H p
The following result gives estimates of the Mellin distance distq(f, Bp
c,σ) for a
from the space Bp
function f and the Mellin-Bernstein spaces Bp
function f ∈ H p
Theorem 6 Let f ∈ H p
c (Ha) with a > 0.
c,σ.
c (Ha) with p ∈ {1, 2}. Then for q ∈ [1,∞], we have
distq(f, B1
dist2(f, B2
c,σ) ≤ 2(cid:107)f(cid:107)H 1
√
c,σ) ≤ 2
c (Ha)
π(cid:107)f(cid:107)H 2
(cid:18) 2
(cid:19)1/q
c (Ha)e−aσ.
aq
e−aσ
Furthermore, for q ∈ [1, 2[,
distq(f, B2
√
c,σ) ≤ 2
π
(cid:18)2 − q
(cid:19)(1/q)−(1/2)(cid:107)f(cid:107)H 2
qa
c (Ha)e−aσ.
Proof. For p = 1, let us consider the function f (·, αε) as in Theorem 5. It is easy to
see that
[f (·, αε)]∧
Mc(c + it) ≤ (cid:107)f (·, αε)(cid:107)Xc ≤ 2(cid:107)f(cid:107)H 1
c (Ha).
Then, by Theorem 5,
[f (·, 0)]∧
(cid:18)(cid:90)
e−qαt(cid:12)(cid:12)[f (·,−α)]∧
(cid:18)
(cid:90) ∞
(cid:19)1/q
Mc(c + it)qdt
(cid:90) ∞
Mc(c + it)(cid:12)(cid:12)q dt +
(cid:19)1/q
t≥σ
σ
e−qαtdt
= 2(cid:107)f(cid:107)H 1
c (Ha)
c (Ha)
2
σ
distq(f, B1
c,σ) =
(cid:18)(cid:90) −σ
=
−∞
≤ 2(cid:107)f(cid:107)H 1
e−qαt(cid:12)(cid:12)[f (·, α)]∧
(cid:18) 2
(cid:19)1/q
e−ασ
αq
(cid:19)1/q
Mc(c + it)(cid:12)(cid:12)q dt
17
Lemma 4 to conclude that
dist2(f, B2
c,σ) =
=
(cid:18)(cid:90)
(cid:18)(cid:90)
≤ e−ασ
(c + it)
dt
(cid:12)(cid:12)(cid:12)2
t≥σ
t≥σ
M 2
c
(cid:12)(cid:12)(cid:12)[f (·, 0)]∧
e−2αt(cid:12)(cid:12)(cid:12)[f (·, εα)]∧
(cid:18)(cid:90) −σ
(cid:12)(cid:12)(cid:12)[f (·,−α)]∧
(cid:90) ∞
(cid:12)(cid:12)(cid:12)[f (·, α)]∧
(cid:16)(cid:107)f (·,−α)(cid:107)2
−∞
M 2
c
M 2
c
+
σ
(cid:19)1/2
(cid:12)(cid:12)(cid:12)2
(cid:12)(cid:12)(cid:12)2
(cid:12)(cid:12)(cid:12)2
(c + it)
M 2
c
(c + it)
dt
(c + it)
dt
(cid:19)1/2
dt
(cid:19)1/2
(cid:17)1/2
for every α ∈ ]0, a[. Now the assertion follows for α → a−.
Next, let p = 2. For q = 2 we may employ Theorem 5 in conjunction with
√
≤ e−ασ
2π
√
≤ e−ασ2
π (cid:107)f(cid:107)H 2
and the assertion follows for α → a−.
with the conjugate exponents µ = 2/(2 − q) and ν = 2/q. This leads us to
Finally, for q ∈ [1, 2[, we modify the previous proof by using Holder's inequality
+ (cid:107)f (·, α)(cid:107)2
c (Ha) ,
X 2
c
X 2
c
(cid:18)(cid:90)
(cid:18)(cid:90)
t≥σ
t≥σ
e−qαt(cid:12)(cid:12)(cid:12)[f (·, εα)]∧
(cid:19)1/(qµ)(cid:18)(cid:90)
M 2
c
e−qµαtdt
(cid:19)1/q
(cid:12)(cid:12)(cid:12)q
(cid:12)(cid:12)(cid:12)[f (·, αε)]∧
dt
M 2
c
(c + it)
t≥σ
(cid:12)(cid:12)(cid:12)qν
(c + it)
(cid:19)1/(qν)
dt
.
distq(f, B2
c,σ) =
≤
The first integral on the right-hand side can be easily calculated. The integrand of
the second integral has the exponent 2. Therefore, we can use Lemma 4 as in the
previous paragraph. Thus we obtain
distq(f, B2
√
c,σ) ≤ 2
π (cid:107)f(cid:107)H 2
c (Ha)
qα
and the assertion follows again by letting α → a−.
e−ασ,
2
(cid:18)2 − q
(cid:19)(1/q)−(1/2)
8 Mellin–Hardy spaces and sampling
In this section we apply the estimates obtained in the previuos section in order to
study the remainder of the exponential sampling formula for functions f belonging
to a Mellin–Hardy space. We recall here some notions concerning the exponential
sampling; see [10], [3], [24].
18
In the following, for c ∈ R, we denote by linc the function
linc(x) :=
x−c
2πi
xπi − x−πi
log x
x−c
2π
=
x−itdt
(x > 0, x (cid:54)= 1)
(cid:90) π
−π
with the continuous extension linc(1) = 1. Thus
linc(x) = x−csinc(log x)
(x > 0).
Here, as usual, the sinc function is defined by
sinc(t) :=
sin(πt)
πt
for t (cid:54)= 0,
sinc(0) = 1.
It is clear that linc (cid:54)∈ X 1
Mellin transform in X 2
b for any b ∈ R. However, it belongs to the space X 2
c -sense is given by
c and its
[linc]∧
M 2
c
(c + iv) = χ[−π,π](v),
where χA denotes the characteristic function of the set A.
For a function g ∈ B2
c,πT the following exponential sampling formula holds for
the restriction of g to the positive real axis (see [10], [8]):
g(x) =
g(ek/T )linc/T (e−kxT )
(x > 0).
(cid:88)
k∈Z
As an approximate version in the space M2
Proposition 5 Let g ∈ M2
c. Then there holds the error estimate
c, we have (see [11, Theorem 5.5]):
g(ek/T )linc/T (e−kxT )
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)g(x) − n(cid:88)
k=−n
≤ x−c
(cid:90)
[g]∧
M 2
c
(c + it)dt
(x ∈ R+, T > 0).
π
t>πT
Note that in general the series in Proposition 5 converges in the principal Cauchy
value sense. However, if f ∈ H 2
c (Ha), then by Proposition 3 one can deduce that the
series for g(·) := f (·, 0) is absolutely and uniformly convergent over any compact
subsets of R. In this case, introducing a remainder (RπT g)(x) by writing
n(cid:88)
k=−n
g(x) =
g(cid:0)ek/T(cid:1) linc/T
(cid:0)e−kxT(cid:1) + (RπT g) (x),
we have by Proposition 5 in conjunction with Theorem 4
(RπT g)(x) ≤ x−c
π
dist1(g, B2
c,πT )
(x > 0),
19
or equivalently,
(cid:107)RπT g(cid:107)X∞
c ≤ 1
π
dist1(g, B2
c,πT ).
Thus, as a consequence of Theorem 6, we obtain:
Corollary 1 Let f ∈ H 2
c (Ha). Then we have
c = O(e−aπT )
(cid:107)RπT f (·, 0)(cid:107)X∞
(T → +∞).
Other approximate formulae can be estimated analogously for Mellin–Hardy spaces,
as an example, the reproducing kernel formula and a sampling formulae for Mellin
derivatives; see [6].
9 Acknowledgments
Carlo Bardaro and Ilaria Mantellini have partially supported by the "Gruppo
Nazionale per l'Analisi Matematica e Applicazioni" (GNAMPA) of the "Istituto
Nazionale di Alta Matematica" (INDAM) as well as by the Department of Mathe-
matics and Computer Sciences of the University of Perugia.
References
[1] N.B. Andersen, On real Paley-Wiener theorems for certain integral transforms, J.
Math. Anal. Appl., 288, (2003), 124–135.
[2] N.B. Andersen and M. de Jeu, Real Paley-Wiener theorems and local spectral radius
formula, Trans. Amer. Math. Soc., 362, (2010), 3613–3640.
[3] C. Bardaro, P.L. Butzer and I. Mantellini, The exponential sampling theorem of
signal analysis and the reproducing kernel formula in the Mellin transform setting,
Sampl. Theory Signal Image Process., 13(1), (2014), 35–66.
[4] C. Bardaro, P.L. Butzer and I. Mantellini, The Mellin-Parseval formula and its in-
terconnections with the exponential sampling theorem of optical physics, Integral
Transforms and Special Functions, 27(1), (2016), 17–29.
[5] C. Bardaro, P.L. Butzer, I. Mantellini and G. Schmeisser, On the Paley-Wiener
theorem in the Mellin transform setting, J. Approx. Theory, 207, (2016), 60–75.
[6] C. Bardaro, P.L. Butzer, I. Mantellini and G. Schmeisser, Mellin analysis and its basic
associated metric. Applications to sampling theory, Analysis Math., 42(4), (2016),
297–321.
20
[7] R.P. Boas, Entire Functions, Academic Press, New York, 1954.
[8] P.L. Butzer and S. Jansche, A direct approach to the Mellin transform, J. Fourier
Anal. Appl., 3, (1997), 325–375.
[9] M. Bertero and E.R. Pike, Exponential sampling method for Laplace and other di-
lationally invariant transforms I. Singular-system analysis. II. Examples in photon
correction spectroscopy and Frauenhofer diffraction, Inverse Problems, 7, (1991), 1–
20; 21–41.
[10] P.L. Butzer and S. Jansche, The exponential sampling theorem of signal analysis,
Atti Sem. Mat. Fis. Univ. Modena, Suppl. Vol. 46, (1998), 99–122.
[11] P.L. Butzer and S. Jansche, A self-contained approach to Mellin transform analysis
for square integrable functions, applications, Integral Transforms Spec. Funct., 8,
(1999), 175–198.
[12] P.L. Butzer, G. Schmeisser and R.L. Stens, The classical and approximate sampling
theorems and their equivalence for entire functions of exponential type, J. Approx.
Theory, 179, (2014), 94–111.
[13] P.L. Butzer, G. Schmeisser and R.L. Stens, Basic relations valid for the Bernstein
σ and their extensions to larger functions spaces via a unified distance con-
spaces B2
cept, J. Fourier Anal. Appl., 19, (2013), 333–375.
[14] R.V. Churchill, J.W. Brown and R.F. Verhey, Complex Variables and Applications
(3. Ed.), McGraw-Hill, New York, 1974.
[15] G.B. Folland and E.M. Stein, Hardy Spaces on Homogeneous Groups, Mathematical
Notes 28, Princeton University Press, Princeton, N.J., 1982.
[16] H-J. Glaeske, A.P. Prudnikov and K.A. Skornik, Operational Calculus and Related
Topics, Chapman and Hall, CRC, Boca Raton, FL, 2006.
[17] J. R. Higgins, Sampling Theory in Fourier and Signal Analysis. Foundations, Oxford
Univ. Press, Oxford, 1996.
[18] R.G. Mamedov, The Mellin Transform and Approximation Theory, (in Russian),
"Elm", Baku, 1991.
[19] R. Nevanlinna and V. Paatero, Introduction to Complex Analysis, Addison-Wesley
Publ. Co., London, 1969.
[20] Q.I. Rahman and G. Schmeisser, Lp inequalities for entire functions of exponential
type, Trans. Amer. Math. Soc., 320(1), (1990), 91–103.
[21] M. Rosenblum and J. Rovnyak, Hardy Classes and Operator Theory, Oxford Math-
ematical Monographs, The Clarendon Press, Oxford, 1985.
21
[22] W. Rudin, Real and Complex Analysis (3. Ed.), McGraw-Hill, New York, 1986.
[23] W. Rudin, Functional Analysis (2. Ed.), McGraw-Hill, New York, 1991.
[24] G. Schmeisser, Quadrature over a semi-infinite interval and Mellin transform, in:
Y. Lyubarskii (ed.) "Proceedings of the 1999 International Workshop on Sampling
Theory and Applications", (ISBN 82-7151-0991), Norwegian University of Science
and Technology, Trondheim, 1999, pp. 203–208.
[25] E.C. Titchmarsh, Introduction to the Theory of Fourier Integrals (2. Ed.), Clarendon
Press, Oxford, 1948.
[26] V.K. Tuan, New type Paley–Wiener theorems for the modified multidimensional
Mellin transform, J. Fourier Anal. Appl., 4, (1998), 317–328.
[27] K. Yosida, Functional Analysis (6. Ed.), Springer-Verlag, Berlin, 1980.
22
|
1706.00623 | 1 | 1706 | 2017-06-02T10:53:21 | Structures on the way from classical to quantum spaces and their tensor products | [
"math.FA"
] | We study tensor products of two structures situated, in a sense, between normed spaces and (abstract) operator spaces. We call them Lambert and proto-Lambert spaces and pay more attention to the latter ones. The considered two tensor products lead to essentially different norms in the respective spaces. Moreover, the proto-Lambert tensor product is especially nice for spaces with the maximal proto-Lambert norm and in particular, for $L_1$-spaces. At the same time the Lambert tensor product is nice for Hilbert spaces with the minimal Lambert norm. | math.FA | math |
Structures on the way from classical to
quantum spaces and their tensor
products
A. Ya. Helemskii
In memoriam: Professor Charles Read
Abstract
We study tensor products of two structures situated, in a sense, between
normed spaces and (abstract) operator spaces. We call them Lambert and
proto-Lambert spaces and pay more attention to the latter ones. The con-
sidered two tensor products lead to essentially different norms in the re-
spective spaces. Moreover, the proto-Lambert tensor product is especially
nice for spaces with the maximal proto-Lambert norm and in particular,
for L1-spaces. At the same time the Lambert tensor product is nice for
Hilbert spaces with the minimal Lambert norm.
1.
Introduction
The subject of the present paper is a structure on a linear space that, in a
reasonable sense, is situated between the classical structure of a normed space and
the structure of an abstract operator, or quantum space. The latter structure was
discovered about 35 years ago; nowadays the theory of operator spaces, sometimes
called quantum functional analysis, is a well developed area of modern functional
analysis, presented in widely known textbooks [5, 17, 15, 1]. Leading idea of
that area was to investigate not just norm on a given linear space, say E, but
Keywords: proto-Lambert space, L-bounded operator, proto-Lambert tensor product, Lam-
bert space, Lambert tensor product.
Mathematics Subject Classification (2000): 46L07, 46M05.
1
2
A. Ya. Helemskii
a sequence of norms k·kn; n = 1, 2, . . . , each one on the space of n × n matrices
with entries in E, mutually related by certain natural conditions, the so-called
Ruan axioms.
The above-mentioned intermediate structure appeared in 2002 in the Ph.D
thesis of A. Lambert [11]; his superviser was G.Wittstock, one of the founding
fathers of operator space theory.
It was Lambert who suggested to consider,
for every n, not a norm on the matrix space Mn(E) but a norm on the column
space of length n, consisting of vectors from E. He called the resulting sequence of
norms operator-sequential norm on E, if it satisfied two natural axioms. Lambert
developed a beautiful and rich theory, in particular, clarifying (putting in proper
perspective) some aspects of quantum as well as classical functional analysis. As
one of the achievements of his theory, Lambert shows that for his spaces there
exists a concept of tensor product with good properties. One can say that his
tensor product is on the way from the projective tensor product of normed spaces
to the operator-projective tensor product of quantum spaces.
In the present paper we study some properties of the Lambert's tensor prod-
uct. But we also pay much attention to a certain natural generalization of Lam-
bert's "operator -- sequential space".
It is called a proto-Lambert space, and it
arises when we assume that only the first of Lambert's axioms for his spaces is
fulfilled. Our point is that one can obtain many good things, if he considers
proto -- Lambert, and not, generally speaking, Lambert spaces.
Note that these proto -- Lambert spaces are, after translation into an equivalent
language, an important particular case of the so -- called p -- multi -- normed spaces.
The latter were quite recently (just in time when this paper was under prepa-
ration) introduced and successfully studied by H.Dales, N.Laustsen, T.Oikhberg
and V.Troitsky in the memoir [3]. Proto -- Lambert spaces correspond to the case
of p = 2. Thus, they share the general properties of p -- multi -- normed spaces, in-
vestigated in the cited memoir. However, the properties, considered in our paper,
rely heavily on the specific advantages of that particular p, actually the same
advantages that distinguish ℓ2 among all ℓp.
Our presentation will be given in the frame -- work of the so -- called non --
coordinate ("index -- free") approach to the structures in question, similar to what
was done in [9] for operator spaces. Thus, it is different from the original approach
in [11]. Both ways of presentation have their own advantages (and drawbacks),
but in questions, revolving around tensor products, the non -- coordinate, "index --
free" presentation is, in our subjective opinion, more elegant and transparent.
Structures on the way from classical to quantum spaces
3
The contents of the paper are as follows.
The second section contains the definition of a proto -- Lambert (still not Lam-
bert) space and some examples, notably the space Lp(X, E); 1 ≤ p < ∞ of
relevant E -- valued measurable functions on X. (Running ahead, we note that
this space is a Lambert space only when p ≥ 2).
In Section 3, we consider classes of maps that reflect in a proper way the struc-
ture of a proto -- Lambert space: L -- bounded and L -- contractive linear and bilinear
operators. We prove that some classes of (bi)linear operators have the relevant
properties and, in particular, some bilinear operators, related to Lp -- spaces and
to "classical" projective tensor products of normed spaces, are completely con-
tractive.
In Section 4, we show that proto -- Lambert spaces have their own tensor prod-
uct
" ⊗pl ", possessing the universal property for the class of L -- bounded bilinear
operators between these spaces.
In Section 5, we concentrate on the case when one of the tensor factors is a
space with the so -- called maximal proto-Lambert norm, in particular, an L1(X) --
space, and give an explicit description of the resulting proto-Lambert tensor
product. As a corollary, we obtain a version, for proto-Lambert spaces, of the
Grothendieck's theorem on tensoring by L1 -- spaces in the "classical" context of
Banach spaces: cf., e.g., [6, §2,no2].
In Section 6, we pass from proto -- Lambert to Lambert spaces, adding in the
relevant definition the non-coordinate analogue of the second of the Lambert's
axioms. We introduce the respective version " ⊗l "of Lambert's "maximal tensor
product" of his Operatorfolgenraume and prove its existence.
We have seen in Section 5 that the proto-Lambert tensor product is especially
nice for L1 -- spaces. In Section 7 we show that the Lambert (without "proto -- ")
tensor product is nice for Hilbert spaces. Namely, if we equip both of Hilbert
spaces with the so -- called minimal Lambert norm, then their completed Lambert
tensor product is again a Hilbert space with the same structure.
In the last Section 8, we compare both tensor products," ⊗pl " and " ⊗l ", and
show that the first one provides, generally speaking, essentially greater norms.
In particular, for every n we display a certain element in the amplification of
the tensor square of a certain Lambert space. It turns out that the norm of this
element, provided by the proto -- Lambert tensor product, is n, whereas the norm,
provided by the Lambert tensor product, is √n.
4
A. Ya. Helemskii
2. Proto-Lambert spaces and their examples
To begin with, we choose an arbitrary, separable, infinite-dimensional Hilbert
space, denote it by H and fix it throughout the whole paper. The identity oper-
ator on H will be denoted by 1.
As usual, by B(E, F ) we denote the space of all bounded operators between
the normed spaces E and F , endowed with the operator norm. We write B(E)
instead of B(E, E), and also B instead of B(H).
If K, L are pre -- Hilbert spaces, x ∈ K, y ∈ L, we denote by x ◦ y : L → K the
rank 1 operator, taking z to hz, yix . Note that we have kx ◦ yk = kxkkyk.
The symbol ⊗ is used for the (algebraic) tensor product of linear spaces and for
elementary tensors. The symbols ⊗p and ⊗i denote the non -- completed projective
and injective tensor product of normed spaces, respectively, and the symbol ⊗hil
is used for the non-completed Hilbert tensor product of pre -- Hilbert spaces. The
symbols b⊗p, b⊗i and b⊗hil are used for the respective completed tensor products.
The complex -- conjugate space of a linear space E is denoted by Ecc.
In what follows we need the triple notion of the so-called amplification. First,
we amplify linear spaces, then linear operators and finally bilinear operators. Note
that these amplifications differ from the amplifications, serving in the theory of
quantum spaces (cf. [9]).
The amplification of a given linear space E is the tensor product H⊗E. Usu-
ally we briefly denote it by HE, and an elementary tensor, say ξ⊗x; ξ ∈ H, x ∈ E,
by ξx. Note that HE is a left module over the algebra B with the outer multi-
plication " · ", well defined by a·(ξx) := a(ξ)x.
Remark 2.1. In the non-coordinate presentation of the operator space theory
the amplification of E is F⊗E, where F is the space of finite rank bounded
operators on H. But F = H⊗H cc; so, passing to the "non-coordinate Lambert
theory", we replace the whole tensor product by its first factor. (One can observe
a similar transfer in the coordinate presentation: we replace Mn(E) = Cn⊗(Cn)cc
by Cn).
Note that the transfer from F⊗E to H⊗E was actually used in the "non-
coordinate" proof of the injective property of the Haagerup tensor product of op-
erator spaces [9, Section 7.3] (such a property was discovered by Paulsen/Smith [16]).
Definition 2.2. A semi-norm on HE is called proto-Lambert semi-norm or
briefly P L-semi-norm on E, if the left B-module HE is contractive, that is we
Structures on the way from classical to quantum spaces
5
always have the estimate ka·uk ≤ kakkuk. The space E, endowed by a P L-
semi-norm, is called semi-normed proto-Lambert space or briefly semi-normed
P L-space). If the semi-norm in question is actually a norm, we speak, naturally,
about a normed proto-Lambert ( = normed P L -- ) space, and in this case we often
omit the word "normed".
A semi-normed P L -- space E becomes semi-normed space (in the usual sense),
if for x ∈ E we set kxk := kξxk, where ξ ∈ H is an arbitrary vector with kξk = 1.
Clearly, the result does not depend on a choice of ξ. The obtained semi-normed
space is called underlying space of a given P L-space, and the latter is called a
P L -- quantization of a former. (We use such a term by analogy with quantizations
in operator space theory; see, e.g., [4], [5] or [9]). Obviously, for all ξ ∈ H and
x ∈ E we have kξxk = kξkkxk.
It is easy to verify that the space of scalars, C, has the only P L -- quantization,
given by the identification H C = H.
Proposition 2.3. Let E be a semi-normed P L -- space with a normed under-
lying space. Then the P L -- semi-norm on HE is a norm.
Proof. Take u ∈ HE; u 6= 0 and represent it asPn
k=1 ξkxk, where ξk are linearly
independent, kξ1k = 1 and x1 6= 0. Further, take η ∈ H with hξ1, ηi = 1 and
hξk, ηi = 0 for k > 1. Then (ξ1 ◦ η)·u = ξ1x1. Therefore we have 0 < kξ1x1k ≤
kξ1 ◦ ηkkuk = kξ1kkηkkuk, hence kuk > 0.
Example 2.4. Every normed space, say E, has, generally speaking, a lot of
P L -- quantizations. We distinguish two of them. The P L -- space, denoted by Emax,
respectively Emin, has the P L -- norm, obtained by the endowing HE with the norm
of H⊗pE, respectively of H⊗iE. We denote the norm on the former and on the
latter space by k·kmax and k·kmin, respectively; accordingly, the corresponding
P L-quantizations of E will be called maximal and minimal. Clearly, the P L-
norm of Emax is the greatest of all P L-norms of P L-quantizations of E. The
adjective "minimal" will be justified a little bit later.
Example 2.5. Let (X, µ) be a measure space, Lp(X); 1 ≤ p < ∞ the relevant
Banach space, F a normed P L-space (say, F := C in the simplest case). We want
to endow the "classical" space Lp(X, F ) (of relevant F -valued functions) with a
P L-norm.
As a preliminary step, consider the normed space Lp(X, HF ) of relevant HF -
valued measurable functions on X and observe that it is a left B-module with the
6
A. Ya. Helemskii
outer multiplication defined by [a·¯x](t) := a·[¯x(t)]; a ∈ B, ¯x ∈ Lp(X, HF ), t ∈ X.
A routine calculation shows that this module is contractive.
Now consider the operator α : H(Lp(X, F )) → Lp(X, HF ), well defined on
elementary tensors by taking ξx; x ∈ Lp(X, F ), ξ ∈ H to the HF -valued function
¯x(t) := ξ(x(t)), and introduce the semi-norm on H(Lp(X, F )), setting kuk :=
kα(u)k.
It is easy to veryfy that α is a B-module morphism. Thus there is
an isometric morphism of the module H(Lp(X, F )) into a contractive module.
It follows immediately that the former module is itself contractive, that is the
introduced semi-norm on H(Lp(X, F )) is a P L-semi-norm on Lp(X, F ). Further,
for ξ ∈ H;kξk = 1 and x ∈ Lp(X, F ) we have kξ[x(t)]k = kx(t)k for all t ∈ X.
Therefore for ξx ∈ H(Lp(X, F )) we have
kξxk =(cid:18)ZX k[ξx](t)kpdµ(t)(cid:19) 1
p
=(cid:18)ZX kx(t)kpdµ(t)(cid:19) 1
p
.
We see that the underlying semi-normed space of the constructed P L-space
is Lp(X, F ). Therefore Proposition 2.3 guarantees that the introduced P L-
seminorm on Lp(X, F ) is actually a norm.
Example 2.6. We want to introduce a P L -- quantization of the "classical"
tensor product E⊗pF of normed spaces, when one of tensor factors, say, to be
definite, F , is a P L -- space.
Consider the linear isomorphism β : H(E⊗F ) → E⊗p(HF ) : ξ(x⊗y) 7→
x⊗ξy and introduce a norm on H(E⊗F ) by setting kUk := kβ(U)k. The space
E⊗p(HF ), as a projective tensor product of a normed space and a contractive
B-module, has itself a standard structure of a contractive B-module. Since β is
a B-module morphism, the same is true with H(E⊗F ). Thus E⊗F becomes
a P L -- space, and we must show that its underlying normed space (E⊗F,k·k is
exactly E⊗pF .
Denote the norm on E⊗pF and on E⊗p(HF ) by k·kp. Take arbitrary u ∈
E⊗F . It is easy to check that the norm k·k on E⊗F is a cross-norm, we have
kξuk ≤ kukp. Therefore our task is to show that, for ξ ∈ H,kξk = 1, we have
kξuk ≥ kukp.
Identifying B-modules H(E⊗F ) and E⊗p(HF ) by means of β, represent ξu
k=1 xk⊗wk; xk ∈ E, wk ∈ HF . Set p := ξ ◦ ξ. Obviously, p·wk = ξyk for
k=1 kxkkkp·wkk =
k=1 xk⊗yk). It
as Pn
some yk ∈ F ; k = 1, ..., n. Therefore Pn
Pn
k=1 kxkkkykk. But we have ξu = p·(ξu) =Pn
k=1 kxkkkwkk ≥ Pn
k=1 xk⊗p·wk = ξ(Pn
Structures on the way from classical to quantum spaces
7
k=1 xk⊗yk. Consequently,Pn
k=1 kxkkkwkk ≥ kukp, and we are
From now on we denote the constructed P L -- quantization of E⊗pF again by
follows that u =Pn
done.
E⊗pF .
3. L-bounded linear and bilinear operators
Suppose we are given an operator ϕ : E → F between linear spaces. Denote, for
brevity, the operator 1⊗ϕ : HE → HF (taking ξx to ξϕ(x)) by ϕ∞ and call it
amplification of ϕ. Obviously, ϕ∞ is a morphism of left B-modules.
Definition 3.1. An operator ϕ : E → F between seminormed P L -- spaces
is called L -- bounded, L-contractive, L -- isometric, L -- isometric isomorphism, if ϕ∞
is bounded, contractive, isometric, isometric isomorphism, respectively. We set
kϕklb := kϕ∞k.
If ϕ is bounded, being considered between the respective underlying semi-
normed spaces, we say that it is (just) bounded, and denote its operator seminorm,
as usual, by kϕk. Clearly, every L -- bounded operator ϕ : E → F is bounded, and
kϕk ≤ kϕklb .
Some operators between P L -- spaces, bounded as operators between underly-
ing spaces, are "automatically" L -- bounded. Here is the first phenomenon of that
kind.
Proposition 3.2. Let E be a P L -- space. Then every bounded functional f :
E → C is L -- bounded, and kfklb := kfk.
Proof. Clearly, it is sufficient to show that for every u ∈ HE we have kf∞(u)k ≤
kfkkuk. Recall that kf∞(u)k = max{hf∞(u), ξi; ξ ∈ H,kξk = 1}. Presenting
u as a sum of elementary tensors, we see that, for every η ∈ H;kηk = 1 we
have hf∞(u), ξiη = f∞[(η ◦ ξ)·u] and also (η ◦ ξ)·u = ηxξ for some xξ ∈ E. It
follows that kxξk = k(η ◦ ξ)·uk ≤ kuk, hence hf∞(u), ξi = kf∞[(η ◦ ξ)·u]k =
kf∞(ηxξ)k = f (xξ) ≤ kfkkuk.
Thus for every P L -- space E and u ∈ HE we have kuk ≥ sup{kf∞(u)k}, where
supremum is taken over all f ∈ E ∗;kfk ≤ 1. But such a supremum is exactly
kukmin. This justifies the word "minimal" in Example 2.4.
8
A. Ya. Helemskii
A P L -- space is called complete (or Banach), if its underlying normed space
is complete. As in the "classical" context, for every P L -- space E there exists its
completion, which is defined as a pair (E, i : E → E), consisting of a complete
P L -- space and an L -- isometric operator, such that the same pair, considered for
respective underlying spaces and operators, is the "classical" completion of E as
of a normed space. The proof of the respective existence theorem repeats, with
obvious modifications, the simple argument given in [9, Chapter 4] for quantum
spaces. We only recall that the norm on HE is introduced with the help of
the natural embedding of HE into HE, the "classical" completion of HE. This
embedding is well defined by taking an elementary tensor ξx; ξ ∈ H, x ∈ E to
limn→∞ ξxn, where xn ∈ E converges to x; hence ξxn can be considered as a
converging sequence in HE. (Here, of course, we identify E with a subspace of
E and HE with a subspace of HE.)
It is easy to observe that the characteristic universal property of the "classical"
completion has its proto -- Lambert version. Namely, if (E, i) is the completion of
a P L -- space E, F a P L -- space and ϕ : E → F is an L -- bounded operator, then
there exists a unique L -- bounded operator ϕ : E → F , which is, in obvious sense,
the continuous extension of ϕ. Moreover, we have kϕklb = kϕklb.
Its proof is the same, up to obvious
Distinguish the following useful fact.
modifications, as of Proposition 4.8 in [9].
Proposition 3.3. Let ϕ : E → F be an L -- isometric isomorhism between
P L -- spaces. Then its continuous extension ϕ : E → F is also an L -- isometric
isomorhism.
We pass to bilinear operators. By virtue of Riesz/Fisher Theorem, we can
arbitrarily choose a unitary isomorphism ι : Hb⊗hilH → H and fix it throughout
the whole paper. Following [8], for ξ, η ∈ H we denote the vector ι(ξ⊗η) ∈ H by
⊗ b)ι−1 on H by a♦b; obviously,
ξ♦η, and for a, b ∈ B we denote the operator ι(a
the latter is well defined by the equality (a♦b)(ξ♦η) = a(ξ)♦b(η). Evidently, we
have
·
(3.1)
kξ♦ηk = kξkkηk
and
ka♦bk = kakkbk.
If E is a linear space, ξ ∈ H and u ∈ HE, we set ξ♦u := Tξ·u, where Tξ ∈ B
sends η to ξ♦η. Thus, this version of the operation ' ♦ ' is well defined on
elementary tensors by ξ♦ηx := (ξ♦η)x. Similarly, we introduce u♦η ∈ HE by
Structures on the way from classical to quantum spaces
9
ξx♦η := (ξ♦η)x. By (3.1), Tξ = kξkS, where S is an isometry. Therefore, if E
is a P L -- space, we have
(3.2)
kξ♦uk = kξkkuk
and similarly
ku♦ηk = kηkkuk.
Now let R : E × F → G be a bilinear operator between linear spaces. Its
amplification is the bilinear operator R∞ : HE× HF → HG, associated with the
4-linear operator H × E × H × F → HG : (ξ, x, η, y) 7→ (ξ♦η)R(x, y). In other
words, R∞ is well defined on elementary tensors by R∞(ξx, ηy) = (ξ♦η)R(x, y)
.
Definition 3.4. A bilinear operator R between P L-spaces is called L-
bounded, respectively, L-contractive, if its amplification is (just) bounded, re-
spectively, contractive. We put kRklb := kR∞k.
It is easy to see that an L-bounded bilinear operator, being considered between
respective underlying (semi-)normed spaces is just bounded, and kRk ≤ kRklb.
On the other hand, similarly to linear operators, sometimes the "classical" bound-
edness automatically implies the L -- boundedness.
Proposition 3.5. Let E, F be P L-spaces, f : E → C, and g : F → C bounded
functionals. Then the bilinear functional f × g : E × F → C : (x, y) 7→ f (x)g(y)
is L -- bounded and kf × gklb = kfkkgk.
Proof. Since kf × gk = kfkkgk, it suffices to show that kf × gklb ≤ kfkkgk. In-
deed, combining the obvious formula (f × g)∞(u, v) = f∞(u)♦g∞(v), Proposition
3.2 and (3.1), we have k(f × g)∞(u, v)k ≤ kfkkgkkukkvk.
Proposition 3.6. Let Lp(X) := Lp(X, C) and Lp(X, E) be the P L-spaces
from Example 2.5. Then the bilinear operator R : Lp(X)× E → Lp(X, E), taking
(z, x) to the E-valued function t 7→ z(t)x; t ∈ X, is L-contractive.
Proof. Recall the isometric operator α : H(Lp(X, E)) → Lp(X, HE) and distin-
guish its particular case α0 : H(Lp(X)) → Lp(X, H). Also consider the bilinear
operator S : Lp(X, H) × HE → Lp(X, HE), taking a pair (ω, u) to the HE --
valued function t 7→ ω(t)♦u; t ∈ X. With the help of (3.2), a routine calculation
gives kS(ω, u)k = kωkkuk.
Now consider the diagram
H(Lp(X)) × HE
α0×1HE
Lp(X, H) × HE
R∞
S
H(Lp(X, E))
,
α
/ Lp(X, HE)
/
/
/
10
A. Ya. Helemskii
It is easy to check on elementary tensors in the respective amplifications that
it is commutative. Therefore, for w ∈ H(Lp(X)) and u ∈ HE we have
kR∞(w, u)k = kα(R∞(w, u))k = kS(α0(w), u)k = kα0(w)kkuk = kwkkuk.
We shall denote the completion of the P L-space E⊗pF from Example 2.6 by
Proposition 3.7. Let E be a normed space, F a P L -- space, E⊗pF the re-
sulting P L -- space. Then the canonical bilinear operator ϑ : Emax × F → E⊗pF ,
Eb⊗pF . Clearly, it is the P L -- quantization of the "classical" Banach space Eb⊗pF .
considered between the respective P L -- spaces, is L -- contractive. Moreover, bϑ :
Emax × F → Eb⊗pF , that is ϑ, considered with Eb⊗pF as its range, is also L --
contractive.
Proof. Consider the trilinear operator T : E×H ×HF → E⊗p(HF ) : (x, ξ, v) 7→
x⊗(ξ♦v). It follows from (3.2) that T is contractive. Therefore the bilinear op-
erator S : (E⊗pH) × HF → E⊗pHF : (x⊗ξ, v) 7→ x⊗(ξ♦v), is also contractive.
Recall the isometric operator β : H(E⊗pF ) → E⊗pHF from Example 2.6
and distinguish its particular case, the "flip" β0 : HEmax → E⊗pH. Consider the
diagram
HEmax × HF
β0×1HF
(E⊗pH) × HF
ϑ∞
S
H(E⊗pF )
,
β
/ E⊗pHF
which is obviously commutative. Therefore a routine calculation shows that for
w ∈ HEmax and v ∈ HF we have kϑ∞(w, v)k ≤ kwkkuk, and we are done.
4. Proto-Lambert tensor product
We proceed to show that L-bounded bilinear operators between P L-spaces can
be linearized with the help of a specific tensor product " ⊗pl ", which seems to
be new. But before, since in this paper we shall come across several varieties
of a tensor product, it is convenient to give a general definition, embracing all
particular cases.
Let us fix, throughout this section, two arbitrary chosen P L -- spaces E and F .
Further, let ℧ be a subclass of the class of all normed P L-spaces.
/
/
/
Structures on the way from classical to quantum spaces
11
Definition 4.1. A pair (Θ, θ) that consists of Θ ∈ ℧ and an L-contractive
bilinear operator θ : E × F → Θ is called tensor product of E and F relative to
℧ if, for every G ∈ ℧ and every L -- bounded bilinear operator R : E × F → G,
there exists a unique L -- bounded operator R : Θ → G such that the diagram
E × F
θ
Θ
◗
◗
R
◗
◗
◗
◗
◗
◗
◗
R
(◗
/ G
is commutative, and moreover kRklb = kRklb.
Such a pair is unique in the following sense:
if (Θk, θk); k = 1, 2 are two
pairs, satisfying the given definition for a certain ℧, then there is a L -- isometric
isomorphism I : Θ1 → Θ2, such that Iθ1 = θ2. This fact is a particular case of
a general -- categorical observation concerning the uniqueness of an initial object
in a category; cf., e.g., [13], [7, Theorem 2.73]. However, the question about the
existence of such a pair depends on our luck with the choice of the class ℧.
Definition 4.2. The tensor product of E and F relative to the class of all
normed P L -- spaces is called non-completed P L -- tensor product of our spaces.
We shall prove the existence of such a pair, displaying its explicit construction.
First, we need a sort of "extended" version of the diamond multiplication,
this time between elements of amplifications of linear spaces. Namely, for u ∈
HE, v ∈ HF we consider the element u♦v := ϑ∞(u, v) ∈ H(E⊗F ), where
ϑ : E × F → E⊗F is the canonical bilinear operator.
In other words, this
"diamond operation" is well defined by ξx♦ηy := (ξ♦η)(x⊗y).
Proposition 4.3. Every U ∈ H(E⊗F ) can be represented asPn
for some natural n and ak ∈ B, uk ∈ HE, vk ∈ HF, k = 1, ..., n.
Proof. Evidently, it suffices to consider the simplest case, when U = ξ(x⊗y); ξ ∈
H, x ∈ E, y ∈ F . Take arbitrary non-zero η, ζ ∈ H; then we have ξ = a(η♦ζ),
for some a ∈ B. Consequently, U = a·(ηx♦ζy).
k=1 ak·(uk♦vk)
As a corollary, the operator B⊗HE⊗HF → H(E⊗F ), associated with the 3-
linear operator (a, u, v) 7→ a·(u♦v), is surjective. Thus H(E⊗F ) can be endowed
with the seminorm of the respective quotient space of B⊗pHE⊗pHF , denoted
by k·kpl. In other words, we have
(4.1)
kUkpl := inf{
kakkkukkkvkk},
nXk=1
(
/
12
A. Ya. Helemskii
where the infimum is taken over all possible representations of U as indicated in
Proposition 4.3.
Proposition 4.4. The seminormed B -- module (H(E⊗F ),k·kpl) is contractive.
Proof. Clearly, B⊗pHE⊗pHF is a contractive left B-module as a tensor product
of the left B-module B and the linear space HE⊗HF . Therefore H(E⊗F ) is
the image of a contractive left B-module with respect to a quotient map of semi-
normed spaces. Since the latter map is a module morphism, we easily obtain the
desired property.
Thus, k·kpl is a P L -- seminorm on E⊗F . Denote the respective P L -- space by
E⊗plF .
Observe the obvious estimate
(4.2)
ku♦vkpl ≤ kukkvk;
u ∈ HE, v ∈ HF.
Since u♦v = ϑ∞(u, v), we see that ϑ, considered with range E⊗plF , is L-
contractive.
Looking at the underlying spaces and using (3.1), we easily obtain that
(4.3)
kx⊗yk ≤ kxkyk;
x ∈ E, y ∈ F.
(In fact, in (4.2) and (4.3) we have the equality, but we shall not discuss it now).
Proposition 4.5. Let G be a P L-space, R : E × F → G an L -- bounded
bilinear operator, R : E⊗plF → G the associated linear operator. Then R is
L -- bounded, and kRklbk = kRklb.
Proof. Take U ∈ H(E⊗plF )) and represent it according to Proposition 4.3. We
remember that R∞ is a B-module morphism. Therefore, using the obvious equal-
ity R∞(u♦v) = R∞(u, v), we have that R∞(U) =Pn
nXk=1
kakkkR∞(uk, vk)k ≤ kRklb
kR∞(U)k ≤
nXk=1
k=1 ak·R∞(uk, vk), hence
kakkkukkkvkk.
From this, using 4.1, we obtain that kR∞(U)k ≤ kRklbkkUkpl. Thus our R is
L -- bounded, and kRklb ≤ kRklb. The converse inequality easily follows from
(4.2).
Proposition 4.6. (As a matter of fact), k·kpl is a norm.
Structures on the way from classical to quantum spaces
13
Proof. By Proposition 2.3, it is sufficient to show that, for a non-zero elementary
tensor ξw; w ∈ E⊗plF, ξ ∈ H;kξk = 1, w 6= 0 we have kξwkpl 6= 0. Since E and F
are normed spaces, then, as it is known, there exist bounded functionals f : E →
C, g : F → C such that (f⊗g)w 6= 0. Now consider in the previous proposition
R := f ×g : E×F → C. By virtue of Proposition 3.5, R is L -- bounded, hence the
operator (f⊗g)∞ is bounded. At the same time (f⊗g)∞(ξw) = [(f⊗g)(w)]ξ 6= 0,
and we are done.
Combining Propositions 4.5 and 4.6, we immediately obtain
Theorem 4.7. (Existence theorem). The pair (E⊗plF, ϑ) is a non-completed
P L-tensor product of E and F .
We can also speak about the "completed" version of Definition 4.2.
Definition 4.8. The tensor product of E and F relative to the class of all
complete P L -- spaces is called completed, or Banach P L -- tensor product of our
spaces.
Proposition 4.9. The Banach P L -- tensor product of P L -- spaces E and F
exists, and it is the pair (Eb⊗plF,bϑ), where Eb⊗plF is the completion of the P L --
space E⊗plF , and bϑ acts as ϑ, but with range Eb⊗plF .
Proof. This is an immediate corollary of the universal property of the completion.
5. Tensoring by maximal P L -- spaces and by L1(·)
In this section we show that for certain concrete tensor factors their P L -- tensor
product also becomes something concrete and transparent.
Theorem 5.1. Let E be a normed space, F a P L -- space, E⊗pF the P L --
space from Example 2.6. Then there exists an L -- isometric isomorphism I :
Emax⊗plF → E⊗pF , acting as the identity operator on the common underlying
linear space of our P L -- spaces. As a corollary (see Proposition 3.3), there exists
an L -- isometric isomorphism bI : Emaxb⊗plF → Eb⊗pF , which is the extension by
continuity of I.
14
A. Ya. Helemskii
Proof. Consider ϑ from Proposition 3.7. By Proposition 4.7, ϑ gives rise to the
L -- contractive operator I, acting as in the formulation. Therefore it is sufficient
to show that the operator I∞ does not decrease norms of elements.
Take U ∈ H(E⊗F ). Identifying the latter space with E⊗HF , we can repre-
k=1 xk⊗vk; xk ∈ E, vk ∈ HF . Choose e ∈ H;kek = 1 and denote by
S ∈ B the isometric operator ξ 7→ e♦ξ; ξ ∈ H. We easily see that
sent U asPn
U = S ∗·" nXk=1
exk♦vk# .
From this, by (4.1), we obtain the estimate kUkpl ≤Pn
have
k=1 kxkkkvkk. Hence, we
(5.1)
kxkkkvkk}.
kUkpl ≤ inf{
nXk=1
Now look at I∞(U). It is the samePn
where the infimum is taken over all representations of U in the indicated form.
k=1 xk⊗vk, only considered in the normed
space E⊗pHF . It follows that kI∞(U)k is exactly the infimum, indicated in (5.1).
Thus, kI∞(U)k ≥ kUkpl and we are done.
Remark 5.2. As an easy corollary of this theorem, we have, up to an L --
isometric isomorphism, that Emaxb⊗plFmax = [Eb⊗pF ]max for all normed spaces E
In particular, we have Hmaxb⊗plHmax = N (H)max, where N (H) is the
and F .
Banach space of trace class operators on H.
Now we want to apply this theorem to the description of P L -- tensor products
in the situation, when one of tensor factors is L1(X) from Example 2.5. As a
"classical" prototype of that description, we recall the following theorem, due to
Grothendieck.
Let (X, µ) be a measure space, F a Banach space. Then there exists an iso-
mentary tensor z⊗x; z ∈ L1(X), x ∈ F to the F -valued function t 7→ z(t)x; t ∈ X.
Theorem 5.3. Let (X, µ) be a measure space, F a complete P L -- space. Then
metric isomorphism GF : L1(X)b⊗pF → L1(X, F ), well defined by taking an ele-
there exists an L -- isometric isomorphism I : L1(X)b⊗plF → L1(X, F ), well de-
fined in the same way as GF in the Grothendieck theorem.
Proof. First we note that the P L-norm on L1(X), introduced in Example 2.5, co-
incides with the maximal P L-norm, introduced in Example 2.4. This is because
Structures on the way from classical to quantum spaces
15
the identity operator I : H⊗pL1(X) → H(L1(X)) participates in the commuta-
tive diagram
H⊗pL1(X)
f lip
L1(X)⊗pH
I
G0
H(L1(X))
,
α1
/ L1(X, H)
where G0 is the restriction of GH , α1 is a particular case of α from Example 2.5,
and these operators, as well as "flip", are isometric.
Combining this with Theorem 5.1, we come to the L -- isometric isomorphism
bI : L1(X)b⊗plF → L1(X)b⊗pF .
Now we show that the isometric isomorphism GF : L1(X)b⊗pF → L1(X, F ) is
L -- isometric with respect to the P L -- norm in the latter space, taken from Example
2.5. We see that GF is the extension by continuity (cf. Section 3) of its restriction
G0
F to L1(X)⊗pF , and this restriction maps the latter space onto a dense subspace
of L1(X, F ). Therefore, by virtue of Proposition 3.3, it is sufficient to show that
the operator G0
F )∞ is isometric. But the
latter participates in the commutative diagram
F is L -- isometric, or, equivalently, that (G0
H(L1(X)⊗pF )
β
L1(X)⊗pHF
(G 0
F )∞
G 0
HF
H(L1(X, F ))
,
α
/ L1(X, HF )
where α and (with L1(X) as E) β are operators from the Examples 2.5 and
2.6, respectively, and G0
HF is the restriction to the respective subspaces of the
isometric isomorphism, provided by the Grothendieck theorem (this time with
the completion of HF in the role of F ). Since α, β and G0
HF are isometric, then
(G0)∞ is also isometric.
After this, to end the proof, it remains to set I := GFbI.
Like in the "classical" context of Grothendieck theorem, we can distinguish a
transparent particular case, concerning integrable functions of two variables.
Proposition 5.4. Let (X, µ), (Y, ν) be two measure spaces, (X×Y, µ×ν) their
product measure space. Then L1(X)b⊗plL1(Y ) = L1(X × Y ) up to an L -- isometric
isomorphism. More precisely, there exists an L -- isometric isomorphism between
the indicated P L -- spaces, well defined by taking x⊗y; x ∈ L1(X), y ∈ L1(Y ) to the
function (s, t) 7→ x(s)y(t); (s, t) ∈ X × Y .
/
/
/
/
/
/
16
A. Ya. Helemskii
Proof. By virtue of Theorem 5.3, it is sufficient to show that the "classical"
isometric isomorphism I : L1(X, L1(Y )) → L1(X × Y ), taking the L1(Y )-valued
integrable function ¯x to the function I(¯x) : (s, t) 7→ [¯x(s)](t), is L -- isometric.
Since every u ∈ H(L1(X, L1(Y )) is (finite) sum of elementary tensors, a routine
calculation, using the construction of the P L -- norm on relevant Lp -- spaces and,
of course, Fubini Theorem, shows that indeed kuk = kI∞(u)k.
Remark 5.5. The results of this section can lead to the conjecture that
something similar, at least in formulations, can be said in the context of the
so-called proto-quantum spaces in the operator space theory (cf. [9, Ch. 2]). It
happened that to some extent it is indeed so, but proofs of crucial facts become
more complicated. However, we do not discuss it in the present paper.
6. Lambert spaces and the Lambert tensor prod-
uct
From now on we concentrate on a special type of P L-spaces, which is a non-
coordinate form of Operatorfolgenraume of Lambert.
If X is a left B-module and x ∈ X, we say that a projection P ∈ B is a
support of x, if P·x = x. A contractive seminormed left B-module Y is called
semi -- Ruan module, if it has the following "property (sR)": for x, y ∈ Y with
orthogonal supports we have
kx + yk2 ≤ kxk2 + kyk2,
hence for every xk ∈ x; k = 1, ..., n with pairwise orthogonal supports we have
k=1 xkk2 ≤Pn
kPn
k=1 kxkk2.
(Semi-Ruan modules were introduced and studied in [10]. Then, in more
general context and with more advanced results, they were investigated in [18].
However, earlier the same class of modules actually cropped up in [14]).
Definition 6.1. For a linear space E, a seminorm on HE is called Lambert
seminorm, or briefly, L -- seminorm, if the left B-module HE is a semi -- Ruan mod-
ule. In other words, L -- seminorm is a P L -- seminorm, satisfying the property (sR).
The linear space, endowed with an L -- seminorm, is called seminormed Lambert
space, or briefly seminormed L -- space. In a similar way, we use the term normed
Lambert space ( = normed L -- space), but in this case we usually omit the word
"normed".
Structures on the way from classical to quantum spaces
17
As to examples of P L -- spaces, it is easy to see that Emin is actually an L -- space
for all normed spaces E, whereas Emax is, generally speaking, not an L -- space.
The P L -- space Lp(X) is an L -- space if, and only if 2 ≤ p. (Of course, we suppose
here that our measure space is not a single atom). Finally, if K is a pre -- Hilbert
space, we can endow it with a so-called Hilbert L-norm, obtained after identifying
HK with H⊗hilK.
The following example will not be used in this paper. However, we mention
it because of its importance in the theory of L -- spaces.
·
Example 6.2. ("Concrete L -- space"). Suppose that E is given as a subspace
of B(K, L) for some Hilbert spaces K, L. Consider the operator γ : HE →
B(K, H
⊗ L), well defined by taking ξT : ξ ∈ H, T ∈ E to the operator x 7→
ξ⊗T (x); x ∈ K. Introduce a seminorm on HE, setting kuk := kγ(u)k. Then it is
not difficult to show that k·k is actually a norm, making E an L -- space.
As a matter of fact, this example is, in a sense, universal: every L -- space is
L -- isometrically isomorphic to some concrete, i.e. operator space. This assertion
can be rather quickly derived from the non -- coordinate version of the result of
Lambert [11, Folgerung 1.3.6] about the embedding of his spaces into products
of copies of ℓ2. However, details are outside the scope of this paper.
From now on we proceed to a special tensor product within the class of L --
spaces. As we shall see, its definition is parallel to that of the P L -- tensor product,
but the resulting object turns out to be a quite different thing.
Let us fix, for a time, two P L-spaces E and F .
Definition 6.3. The tensor product of E and F relative to the class of all
normed L -- spaces is called non-completed L -- tensor product of our spaces.
Remark 6.4. This tensor product is a "non -- coordinate' version of what Lam-
bert calls maximal tensor product. Indeed, in a sense, it is maximal within a rea-
sonable class of tensor products, and it plays in Lambert's theory a role similar
to the role of the operator-projective tensor product in the theory of quantum (
= abstract operator) spaces. See details in [11, 3.1.1].
We shall prove the existence of this kind of tensor product, displaying its
explicit construction. Such a construction and the crucial Proposition 6.7 can be
considered as the "non -- coordinate" version of what was done by Lambert.
18
A. Ya. Helemskii
Recall the diamond product in all its varieties. The following proposition
concerns all linear spaces E, F without any additional structure. First, note two
identities
(6.1)
a·u♦b·v = (a♦b)(u♦v); a·[(b·u)♦v] = a(b♦1)·(u♦v); a, b ∈ B, u ∈ HE, v ∈ HF
that can be easily checked on elementary tensors.
Proposition 6.5. Every U ∈ H(E⊗F ) can be represented as a·(Pn
k=1 uk♦vk),
a ∈ B, uk ∈ HE, vk ∈ HF, k = 1, ..., n, where uk have pairwise orthogonal sup-
ports.
Proof. Represent U as in Proposition 4.3. Choose isometric operators S1, ..., Sn ∈
B with pairwise orthogonal final projections. Set
Since S ∗
kSl = δl
k1, the identities (6.1) imply that
a :=
nXk=1
a·" nXk=1
nXk,l=1
[ak(S ∗
akS ∗
k♦1
and
u′
k := Sk·uk; k = 1, ..., n.
k♦vk# =
u′
nXk,l=1
[ak(S ∗
k♦1)]·[(Sl·ul)♦vl] =
k♦1)(Sl♦1)]·[ul♦vl] =
ak·(uk♦vk) = U.
nXk=1
Finally, the elements u′
k have orthogonal supports SkS ∗
k; k = 1, ..., n.
(6.2)
Now, for a given U ∈ H(E⊗F ), we introduce the number
kUkl = infkak nXk=1
2
kuk2kvk2! 1
,
where the infimum is taken over all possible representations of U in the form
indicated by Proposition 6.5. We distinguish the obvious
Proposition 6.6. For every U ∈ H(E⊗F ) and a ∈ B we have ka·Ukl ≤
kakkUkl.
Proposition 6.7. The function U 7→ kUkl is a seminorm on H(E⊗F ).
Structures on the way from classical to quantum spaces
19
k=1 u1
Proof. Let U = a·(Pn
k ∈
HE, respectively u2
l ∈ HE, have pairwise orthogonal supports Pk, respectively
Ql. Take arbitrary isometric operators S, T ∈ B with orthogonal final projections
and observe that
k), V = b·(Pm
l ), where the elements u1
k♦v1
l ♦v2
l=1 u2
U + V = (a(S ∗♦1) + b(T ∗♦1))· nXk=1
(S·u1
k)♦v1
k +
mXl=1
l! .
l )♦v2
(T·u2
The elements S·u1
l have
supports T QlT ∗; hence, taking together, these elements have pairwise orthogonal
supports. Therefore, by virtue of Proposition 6.6 and of (6.2), we have that
k have supports SPkS ∗, whereas the elements T·u2
k(T·u2
l k2kv2
2
l k2! 1
.
mXl=1
Combining this with the operator C ∗-property, we obtain that
1
kk2 +
k(S·u1
k)k2kv1
Further, obviously we can assume that
kU + V kl ≤ (kak2 + kbk2)
kU + V kl ≤ ka(S ∗♦1) + b(T ∗♦1)k nXk=1
2 nXk=1
mXl=1
kk2! 1
and kbk = mXl=1
kk2! 1
kU + V kl ≤ kak2 + kbk2 = kak nXk=1
kak = nXk=1
Therefore we have
k)k2kv1
kk2kv1
kk2kv2
kk2 +
ku1
ku1
ku1
2
2
2
l k2! 1
ku2
lk2kv2
.
2
l k2! 1
.
ku2
l k2kv2
+ kbk mXl=1
2
l k2! 1
ku2
l k2kv2
.
From this with the help of (6.2) we obtain that kU + V kl ≤ kUkl + kV kl.
The property of seminorms, concerning the scalar multiplication, is immedi-
ate.
Proposition 6.8. The module (H(E⊗F ),k·kl) has the property (sR).
Proof. Let U, V ∈ H(E⊗F ) have orthogonal supports P and Q. Choose their
l ♦v2
l ).
Take S, T as in Proposition 6.7; then, by a similar argument, we have
k=1 u1
l=1 u2
arbitrary suitable representation, say U = a·(Pn
kU + V kl ≤ ka(S ∗♦1) + b(T ∗♦1)k nXk=1
ku1
k♦v1
k), V = b·(Pm
mXl=1
l k2kv2
kk2 +
ku2
kk2kv1
2
l k2! 1
.
20
A. Ya. Helemskii
Evidently, we can assume that a = P a, b = Qb and kak = kbk = 1. Therefore, by
the operator C ∗-property, we have
ka(S ∗♦1) + b(T ∗♦1)k = kP aa∗P + Qbb∗Qk
1
2 = max{kak,kbk} = 1.
Consequently, by (6.2), we have that kU + V k2
l ≤ kUk2
l + kV k2
l .
Combining the last three propositions, we see that k·kl is an L -- seminorm on
E⊗F . We denote the resulting semi-normed L-space by E⊗lF .
Like in the "P L -- case" (cf. (4.2)), we have the obvious estimation
(6.3)
ku♦vkl ≤ kukkvk;
u ∈ HE, v ∈ HF.
Consequently, the canonical bilinear operator ϑ : E×F → E⊗lF is L-contractive.
Proposition 6.9. Let G be an L-space, R : E×F → G an L -- bounded bilinear
operator, R : E⊗plF → G the associated linear operator. Then R is L-bounded,
and kRklb = kRklb.
Proof. Take U ∈ H(E⊗lF )) and represent it as in Proposition 6.5. Since R∞
is a B-module morphism, and R∞(uk♦vk) = R∞(uk, vk) for all k, we have that
R∞(U) = a·(Pn
k=1 R∞(uk, vk)).
Now look at R∞(uk, vk) ∈ HG for some k. Obviously we have (Pk♦1)·R∞(uk, vk) =
R∞(Pk·uk, vk). This implies that the elements R∞(uk, vk) ∈ HG have pairwise
orthogonal supports, namely Pk♦1. Therefore, since G is an L-space, we have
2
kR∞(U)k ≤ kak nXk=1
lbkukk2kvkk2! 1
kRk2
kR∞(uk, vk)k2! 1
= kRklbkak nXk=1
2
≤
kukk2kvkk2! 1
2
.
kak nXk=1
From this, using (6.2), we obtain the estimate kR∞(U)k ≤ kRklbkkUklk.
Consequently, kRklb ≤ kRklb. The converse inequality easily follows from (6.3).
Proposition 6.10. (As a matter of fact), k·kl is a norm.
Proof. Since C is an L -- space, the argument in Proposition 4.6 works with obvious
modifications.
Structures on the way from classical to quantum spaces
21
Combining Propositions 6.6 -- 6.10, we immediately obtain
Theorem 6.11. (Existence theorem) The pair (E⊗lF, ϑ) is a non-completed
L -- tensor product of E and F .
The non-completed L -- tensor product has an obvious "completed" version.
The definition of the completed L -- tensor product of two P L -- spaces and the rele-
vant existence theorem repeat what was said about completed P L -- tensor prod-
uct, only we replace "P L" by "L" and the subscript "pl" by "l". Thus, the
completed L -- tensor product of two P L -- spaces E and F exists, and it is the pair
(Eb⊗lF,bϑ), where Eb⊗pF is the completion of the L -- space E⊗lF , and bϑ acts as
ϑ, but with range Eb⊗lF .
Remark 6.12. We do not discuss here the non-coordinate version of an-
other tensor product, the so -- called minimal, introduced in [11, 3.1.3]. Unlike
the maximal tensor product (cf. Remark 6.4), it corresponds, in a sense, to the
operator -- injective tensor product in the operator space theory.
7. The Lambert tensor product of Hilbert spaces
As we have seen before, the P L -- tensor product is especially good for maximal
P L -- spaces and L1-spaces with their specific P L -- norm. Here we shall show that,
in the same sense, the L -- tensor product is good for Hilbert spaces with the
minimal L -- norm, that is k·kmin of Example 2.4. Throughout this section, all
Hilbert and pre -- Hilbert spaces are supposed to be endowed with that L -- norm.
We shall use one of equivalent definitions of the minimal L -- norm.
It is a
particular case of the definition of the norm in the injective tensor product E⊗iF
of two normed spaces, expressed by means of an injective operator E⊗iF →
B(E ∗, F ) (see, e.g.,[2, pp. 62-63]). In the particular case of a pre-Hilbert space
K we obtain the following observation that we distinguish for the convenience of
references.
Proposition 7.1. There is an isometric operator I : HK → B(K cc, H), well
defined by ξx 7→ ξ ◦ x.
This, in its turn, implies
22
A. Ya. Helemskii
k=1 λkξkxk ∈ HK, where λk ∈ C and ξk, xk
are orthonormal systems in H and K respectively, we have kuk = max{λk; k =
1, ..., n}.
Proposition 7.2. (i) For u =Pn
(ii) Every u ∈ HK can be represented as Pn
operator between pre -- Hilbert spaces, has the formPn
orthonormal systems in H and K respectively, s1 ≥ s2 ≥ ... ≥ sn > 0.
Proof. (i) is immediate. To prove (ii), we recall that I(u), being a finite rank
k=1 skξk ◦ xk, where ξk, xk
and sk have the indicated properties. (E.g., the argument in [7, Section 3.4] works
with obvious modifications). It follows that u have the desired representation.
k=1 skξkxk, where ξk and xk are
Proposition 7.3. Let K and L be pre-Hilbert spaces. Then the canonical
bilinear operator ϑ : K × L → K⊗hilL : (x, y) 7→ x⊗y is L-contractive.
Proof. As we know, ϑ∞ : HK× HL → H(K⊗hilL) takes a pair (u, v) to u♦v. By
k=1 skξkxk with the mentioned properties,
l=1 s′
lηlyl with similar properties. Consequently,
Proposition 7.2(ii), u has the form Pn
and v has the formPm
mXl=1
nXk=1
u♦v =
sks′
l(ξk♦ηl)(xk⊗yl),
where the systems ξk♦ηl and xk⊗yl are orthonormal in H and K⊗hilL, respec-
tively. Therefore, by Proposition 7.2(i), ku♦vk = s1s′
1 = kukkvk.
Theorem 7.4. Let K and L be pre -- Hilbert spaces. Then we have K⊗lL =
phism, well defined by taking an elementary tensor x⊗y to the same x⊗y, but
K⊗hilL and Kb⊗lL = Kb⊗hilL. Both equalities are up to an L -- isometric isomor-
considered in K⊗hilL and Kb⊗hilL, respectively.
Proof. Since ϑ from the previous proposition has values in an L -- space, it gives
rise to the L -- contractive operator R : K⊗lL → K⊗hilL, which is the identity
map of the underlying linear spaces. Our task is to show that it is an L -- isometric
isomorphism.
Take U ∈ H(K⊗lL). Since it is the sum of several elementary tensors of the
l=1 ξkl(xk⊗yl),
where xk and yl are orthonormal systems in K and L, respectively, and ξkl ∈ H.
Applying to R∞(U) ∈ H(K⊗hilL) Proposition 7.1, we see that kR∞(U)k is the
l=1 ξkl ◦ (xk⊗yl) : (K⊗hilL)cc → H. Set
M := span{xk⊗yl} ⊂ (K⊗hilL)cc. Since dim M < ∞, the pre -- Hilbert space
form ξ(x⊗y), it easily follows that U can be represented asPn
norm of the operator S := Pn
k=1Pm
k=1Pm
Structures on the way from classical to quantum spaces
23
L)cc decomposes as M ⊕ M ⊥, and S takes M ⊥ to 0. Therefore kSk = kS0k,
(K ⊗·
where S0 is the restriction of S to M. Thus, we have
(7.1)
kR∞(U)k = kS0k.
Now return to our initial U. Choose arbitrary orthonormal systems ηk; k = 1, ..., n
l=1 ζlyl ∈ HL.
l=1(ηk♦ζl)(xk⊗yl). Consider the finite rank operator
k=1 ηkxk ∈ HK and v :=Pm
and ζl; l = 1, ..., m in H. Set u :=Pn
We see that u♦v =Pn
mXl=1
k=1Pm
nXk=1
T :=
ξkl ◦ (ηk♦ζl) : H → H.
An easy calculation shows that T·(u♦v) = U. Further, as a particular case of
Proposition 7.1, kuk = kvk = 1. Finally, if we set N := span{ηk♦ζl} and denote
by T0 the restriction of T to N, we obviously have kTk = kT0k. Combining this
with (6.2), we obtain that
(7.2)
kUkl ≤ kT0k.
The systems {xk⊗yl} and {ηk♦ζl} are orthonormal bases in M and N, respec-
tively, and S0(xk⊗yl) = ξkl = T0(ηk♦ζl). It follows that kT0k = kS0k. Combining
this with (7.1) and (7.2), and remembering that R∞ is contractive, we obtain
that kUkl = kR∞(U)k.
This gives the first of L -- isometric isomorphisms, claimed in the theorem.
The extension of the latter by continuity provides the second L -- isometric iso-
morphism.
8. Comparison of both tensor products
In conclusion, we want to compare P L -- and L -- tensor products. Since the class
of P L-spaces is larger than that of L-spaces, it immediately follows from the
definition of both tensor products in terms of their universal properties, that
k·kpl ≥ k·kl. We shall show that the first number is sometimes essentially greater
than the second number.
Endow the space ℓ2 with the Hilbert L -- norm (see above), and the space
ℓ1 = L1(N) with the P L-norm from Example 2.5.
Proposition 8.1. The bilinear operator M : ℓ2 × ℓ2 → ℓ1, acting as the
coordinate-wise multiplication, is L-contractive.
24
A. Ya. Helemskii
Proof. Our task is to show that the bilinear operator M∞ : Hℓ2 × Hℓ2 → Hℓ1
is contractive. Consider the isometric operators I : Hℓ2 = H⊗hilℓ2 → ℓ2(H)
and α : Hℓ1 → ℓ1(H); both are well defined by taking an elementary tensor
ξeλ;eλ = (. . . , λn, . . . ) to (. . . , λnξ, . . . ). Consider the diagram
I×I
Hℓ2 × Hℓ2
ℓ2(H) × ℓ2(H)
,
M∞
Hℓ1
α
S
/ ℓ1(H)
where S takes a pair (eξ := (. . . , ξn, . . . ),eη := (. . . , ηn, . . . )) to (. . . , ξn♦ηn, . . . ).
Since the Cauchy-Schwarz inequality implies that kS(eξ,eη)k ≤ keξkkeηk, the latter
sequence indeed belongs to ℓ1(H), and, moreover, S is contractive.
Now observe that our diagram, as one can easily verify on elementary tensors,
is commutative. Consequently, for u, v ∈ Hℓ2 we have kM∞(u, v)k ≤ kukkvk.
Have a look at the P L -- and L -- tensor square of the same Hilbert L -- space ℓ2.
Fix n ∈ N, denote by pm; m = 1, 2, ... sequences (. . . , 0, 1, 0, . . . )ℓ2 and choose an
arbitrary orthonormal system, say e1, e2, . . . , in H. In what follows, we set
V :=
nXk=1
ek(pk⊗pk) ∈ H(ℓ2⊗ℓ2).
It obviously can be presented as
(8.1)
V = S·
nXk=1
uk♦vk,
where uk = vk = ekpk, and S =Pn
Proposition 8.2. We have kV kpl = n.
k=1 ek ◦ (ek♦ek).
Proof. Consider the operator M : ℓ2⊗plℓ2 → ℓ1, associated with M from the pre-
vious proposition. Then it is L -- contractive together with the latter: in particular,
k=1 ek(pk); therefore,
since we are in Lℓ1, we have kM∞(V )k = n. Consequently, kV kpl ≥ n. On the
kM∞(V )k ≤ kV kpl. But we obviously have M∞(V ) = Pn
other hand, it follows from (8.1) that kV kpl ≤Pn
Proposition 8.3. (At the same time) we have kV kl = √n.
k=1 kSkkukkkvkk = n.
/
/
/
Structures on the way from classical to quantum spaces
25
Proof. Consider the bilinear operator N : ℓ2 × ℓ2 → ℓ2, acting as M, but with
the other range. Since the norm of an element in Hℓ1 can only decrease, if we
shall consider this element in Hℓ2, our N is L -- contractive together with M.
But ℓ2 (contrary to ℓ1!) is an L -- space; therefore the operator N : ℓ2⊗lℓ2 → ℓ2,
associated with N , is also L -- contractive. In particular, kN∞(V )k ≤ kV kl. Of
k=1 ek(pk) as in the previous proposition. However,
since now we are in Hℓ2 =
course, N∞(V ) is the samePn
H⊗hilℓ2, we have kN∞(V )k = √n, and hence kV kl ≥ √n.
we obtain, by (8.1), that kV kl ≤ kSk(Pn
k=1 kukk2kvkk2)
On the other hand, since the elements uk ∈ Hℓ2 form an orthonormal system,
1
2 = √n.
Nevertheless, despite P L -- and L -- tensor products of the same P L-spaces usu-
ally have essentially different norms, their underlying spaces coincide:
Proposition 8.4. Let E and F be P L -- spaces. Then the identity operator
on the linear space E⊗F is an isometric isomorphism, being considered as an
operator between underlying spaces of P L -- spaces E⊗plF and E⊗lF .
Proof. Since the latter operator is obviously contractive, our task is to show that
its inverse operator is also contractive. Denote by G be the underlying normed
space of E⊗plF , endowed with the minimal P L-norm (see Example 2.4). Since
ϑ : E × F → E⊗plF is L -- contractive, the same is true, if we consider ϑ with
range G. But, as we know, G is an L -- space. Therefore ϑ gives rise to the L --
contractive operator between E⊗lF and G, which is contractive as an operator
between the underlying normed spaces. But the latter is, of course, the desired
inverse operator.
This research was supported by the Russian Foundation for Basic Research
(grant No. 15-01-08392).
References
[1] D. P. Blecher, C. Le Merdy, Operator Algebras and their Modules (Claren-
don Press, Oxford, 2004).
[2] J. Cigler, V. Losert, P. Michor., Banach modules and functors on categories
of Banach spaces (Marcel Dekker, New York, 1979).
26
A. Ya. Helemskii
[3] H. G. Dales, N. J. Laustsen, T. Oikhberg, V. G. Troitsky, Multi-norms and
Banach lattices. Preprint.
[4] E. G. Effros, 'Advances in quantized functional analysis', Proc. ICM Berke-
ley, 1986.
[5] E. G. Effros, Z.-J. Ruan, Operator Spaces (Clarendon Press, Oxford, 2000).
[6] A. Grothendieck, Produits Tensoriels Topologiques et Espaces Nucleaires,
Mem. Amer. Math. Soc., No. 16, 1955.
[7] A. Ya. Helemskii, Lectures and exercises on functional analysis. American
Mathematical Society, Providence, R.I., 2005.
[8] A. Ya. Helemskii,
'Tensor products in quantum functional analysis:
non-coordinate approach'. In:
'Topological Algebras and Applications',
A. Mallios, M. Haralampidou, Eds. American Mathematical Society, Prov-
idence, R.I. (2007), 199-224.
[9] A. Ya. Helemskii, Quantum Functional Analysis (American Mathematical
Society, Providence, R.I., 2010).
[10] A. Ya. Helemskii,
'Extreme flatness of normed modules and Arveson-
Wittstock type theorems', J. Operator Theory, 64:1 (2010), 101-112.
[11] A. Lambert, Operatorfolgenraume. Dissertation. Saarbruken. 2002.
[12] A. Lambert, M. Neufang, V. Runde,
'Operator space structure and
amenability for Fig´a-Talamanca -- Herz algebras', J. Functional Analysis, 211
(2004), 245-269.
[13] S. Mac Lane, Categories for the Working Mathematician (Springer-Verlag,
Berlin, 1971).
[14] B. Magajna, 'The minimal operator module of a Banach module', Proc.
Edinburgh Math. Soc., No. 1, 42 (1999), 191 -- 208.
[15] V. I. Paulsen, Completely Bounded Maps and Operator Algebras (Cam.
Univ. Press, Cambridge, 2002).
[16] V. I. Paulsen, R. R. Smith, 'Multilinear maps and tensor norms on operator
systems', J. Funct. Anal., 73, (1987) 258-276.
Structures on the way from classical to quantum spaces
27
[17] J. Pisier, Introduction to Operator Space Theory (Cam. Univ. Press, Cam-
bridge, 2003).
[18] G. Wittstock, 'Injectivity of the module tensor product of semi-Ruan mod-
ules', J. Operator Theory, 65: 1 (2010), 87-113.
Moscow State (Lomonosov) University
Moscow, 111991, Russia
E-mail address: [email protected]
|
1910.07606 | 1 | 1910 | 2019-10-16T20:57:41 | Towards Generalized Riesz Systems Theory | [
"math.FA"
] | Pseudo-Hermitian Hamiltonians have recently become a field of wide investigation. Originally, the Generalized Riesz Systems (GRS) have been introduced as an auxiliary tool in this theory. In contrast, the current paper, GRSs are analysed in terms of basis theory. The relationship between semi-regular sequences and GRSs is provided. Various characterizations of GRSs are discussed. | math.FA | math |
Towards Generalized Riesz Systems Theory
A. Kamuda
AGH University of Science and Technology,
30-059 al. Mickiewicza 30, Krak´ow, Poland
e-mail: [email protected]
S. Kuzel
AGH University of Science and Technology,
30-059 al. Mickiewicza 30, Krak´ow, Poland
e-mail: [email protected]
Abstract
Pseudo-Hermitian Hamiltonians have recently become a field of wide investigation. Orig-
inally, the Generalized Riesz Systems (GRS) have been introduced as an auxiliary tool
in this theory. In contrast, the current paper, GRSs are analysed in terms of basis the-
ory. The relationship between semi-regular sequences and GRSs is provided. Various
characterizations of GRSs are discussed.
Keywords: -- Riesz basis; biorthogonal sequences; Krein space; PT-symmetric quantum
mechanics;
MSC classification: -- 46N50; 81Q12
I
Introduction
Theory of non self-adjoint operators attracts a steady interests in various fields of mathematics
and physics, see, e.g., [7] and the reference therein. This interest grew considerably due to
the recent progress in theoretical physics of PT -symmetric (pseudo-Hermitian) Hamiltonians
[10, 11, 24]. Studies of pseudo-Hermitian operators carried out in [22, 23, 27] show that, even
if the eigenvalues of a Hamiltonian are real, the Riesz basis property of its eigenstates is, in
many cases, lost.
Such kind of phenomenon is typical for PT -symmetric Hamiltonians and it gives rise to a
natural problem: How to define a suitable generalization of the Riesz bases concept which can
be useful in theory of pseudo-Hermitian Hamiltonians?
One of generalizations was proposed by Davies [13]: the concept of tame and wild sequences.
Each basis is a tame sequence. The tameness of eigenstates of non self-adjoint operators H
with a purely discrete real spectrum allows one to discover additional properties of H.
In
particular, a polynomially bounded behavior of the corresponding resolvent was established in
[13, Theorem 3]. However, in major part, eigenstates of pseudo-Hermitian Hamiltonians form
wild systems that are much more complicated for the investigation [13, 22, 23].
Another approach to the generalization of Riesz bases is based on the rigged Hilbert spaces
framework instead of the original Hilbert space [9].
In the present paper, we study generalized Riesz systems (GRS) which was originally intro-
duced in [15, 17] and then, slightly modified in [6, 8]. In order to explain the idea of definition
we note that vectors of a Riesz basis {φn} have the form φn = Ren, where R is a bounded and
boundedly invertible operator in a Hilbert space H and {en} is an orthonormal basis (ONB)
of H. Using the polar decomposition of R = R∗U = eQ/2U , where U is a unitary operator in
H, we arrive at the conclusion: a sequence {φn} is called a Riesz basis if there exists a bounded
self-adjoint operator Q in H and an ONB {en} such that φn = eQ/2en. This simple observation
leads to:
Definition 1 A sequence {φn} is called a generalized Riesz system (GRS) if there exists a self-
adjoint operator Q in H and an ONB {en} such that en ∈ D(eQ/2)∩D(e−Q/2) and φn = eQ/2en.
For a GRS {φn}, the dual GRS is determined by the formula {ψn = e−Q/2en}. Obviously,
{φn} and {ψn} are bi-orthogonal sequences.
Dual GRS's can be considered as a particular case of G-quasi bases introduced by Bagarello
in [4], and then analyzed in a series of papers, see, [7] and the references therein.
2
The main objective of the paper is to further development of the GRS theory. In contrast
to the standard approach [5, 6, 14, 15, 16, 17], where GRS were mainly used as auxiliary tools
for the definition and investigation of manifestly non self-adjoint Hamiltonians and relevant
physical operators, we consider GRS as a self-contained object of the basis theory [12, 19]. Our
studies are based on advanced methods of extension theory of symmetric operators, see [2, 3]
and Subsection II.1.
We say that a sequence of vectors {φn} of a Hilbert space H is semi-regular if {φn} is
minimal and complete in H. The minimality of {φn} yields the existence of a bi-orthogonal
sequence {ψn}, while the completeness of {φn} guarantees the uniqueness of {ψn}. The positive
symmetric operator S mapping {φn} onto {ψn}, see (II.4) plays an important role in our studies.
We show that a semi-regular sequence {φn} is a GRS if and only if the Friedrichs extension AF
of S is a positive operator (Theorem II.1). Another criterion of being GRS was established in
[15, Theorem 3.4] by methods based on the investigation of special operators associated with
{φn} and {en}.
Theorem II.1 allows one to explain the phenomenon of nonuniqueness of self-adjoint oper-
ators Q in the definition of GRS (Subsection II.3). Further we show that each semi-regular
sequence {φn} with the property of being Bessel sequence has to be a GRS and we character-
ize this important case in terms of Q (Theorem II.2). The Olevskii's result [25, Theorem 1]
allows one to establish a relationship between essential spectra of self-adjoint operators Q and
conditional bounded bases (Proposition II.3).
At the end of Section II we concentrate on an important particular case (which fits well the
specific of PT -symmetric Hamiltonians) where a semi-regular sequence {φn} is J-orthonormal
and we combine the GRS-related approach with Krein spaces based methods [7, Chap. 6 ],
[20]. Following [8], we define J-orthonormal sequences of the first/second type and discuss
advantages of first type sequences. In particular, eigenstates of the shifted harmonic oscillator
form a J-orthonormal sequence of the first type and it seems natural to suppose that eigenstates
of a PT -symmetric Hamiltonian with unbroken PT -symmetry [10, p. 41] form a first type
sequence. In Subsection III.2 we present a general method which allows one to construct of the
first/second type sequences.
Throughout the paper, D(A), R(A), and ker A denote the domain, the range, and the null-
space of a linear operator A, respectively. The symbol {vn} means the collection of vectors vn
parametrized by a set I of integers. Usually, I = N.
3
II General theory of GRS
II.1 Preliminaries
Here all necessary results of extension theory of symmetric operators are presented in a form
convenient for our exposition. The articles [1, 2, 3] and [26, Chap. 10] are recommended as
complementary reading on the subject.
Let H be a complex Hilbert space with inner product (·,·) linear in the first argument. An
operator A is called positive [nonnegative] if (Af, f ) > 0 [(Af, f ) ≥ 0] for non-zero f ∈ D(A).
Let A and B be nonnegative self-adjoint operators. We say that A is greater or equal B,
i.e., A ≥ B if
D(A1/2) ⊆ D(B1/2) and (cid:107)A1/2f(cid:107) ≥ (cid:107)B1/2f(cid:107),
f ∈ D(A1/2).
(II.1)
The next technical result follows from (II.1) (see [26, Corollary 10.12]).
Lemma II.1 If A ≥ B and B is positive, then A is also positive.
Let S be a nonnegative densely defined operator in H. M. Krein established that the set
of nonnegative self-adjoint extensions {A} of S can be ordered as follows [3, Theorem 3.5]:
AF ≥ A ≥ AK, where the greatest self-adjoint extension AF is called the Friedrichs extension,
while the smallest one AK is called the Krein-von Neumann extension.
The extensions AK and AF are examples of extremal extensions. We recall [2] that a
nonnegative self-adjoint extension A of S is called extremal if
(A(φ − f ), (φ − f )) = 0
for all φ ∈ D(A).
inf
f∈D(S)
(II.2)
If a nonnegative self-adjoint extension A of S is positive, we can set A = e−Q, where
Q = − ln A is a self-adjoint operator in H and define the new Hilbert space H−Q as the
completion of D(A) = D(e−Q) with respect to the new inner product
(f, g)−Q := (e−Qf, g) = (e−Q/2f, e−Q/2g),
f, g ∈ D(e−Q).
(II.3)
Rewriting (II.2) as
inf
f∈D(S)
(φ − f, φ − f )−Q = inf
f∈D(S)
(cid:107)φ − f(cid:107)2−Q = 0 for all φ ∈ D(e−Q)
we obtain
4
Lemma II.2 A positive self-adjoint extension A = e−Q of S is an extremal extension if and
only if D(S) is a dense set in H−Q.
A non-negative non-densely defined symmetric operator S admits self-adjoint extensions,
but not necessarily non-negative ones. The well-known Ando-Nishio result is [1, Theorem 1]:
Lemma II.3 A closed non-negative symmetric operator S admits a non-negative self-adjoint
extension if and only if it is positively closable, i.e., if the relations
n→∞(Sfn, fn) = 0
lim
and
lim
n→∞ Sfn = g
implies g = 0.
Lemma II.3 is evident for densely defined operators because each densely defined operator
S is positively closable [1, p. 67]. Another useful result follows from [1, Corollary 4]:
Lemma II.4 Let S be a closed densely defined positive operator. Then S−1 admits a non-
negative self-adjoint extension if and only if the Friedrich extension AF of S is positive.
II.2 Conditions of being GRS
Let {φn} be a GRS. In view of Definition 1, the sequence {ψn = e−Q/2en} is well defined and it
is a bi-orthogonal sequence for {φn = eQ/2en}. Obviously, {ψn} is a GRS which we call a dual
GRS.
The existence of a bi-orthogonal sequence means that each GRS {φn} has to be a minimal
sequence, i.e., φj /∈ span{φk}k(cid:54)=j [12, Lemma 3.3.1]. However, not each minimal sequence is a
GRS.
We say that a minimal sequence {φn} is semi-regular if {φn} is complete in H and regular
if its bi-orthogonal sequence {ψn} is also complete. For a semi-regular sequence {φn} the
corresponding bi-orthogonal sequence {ψn} is determined uniquely.
Let {φn} be a minimal sequence. Then there exists a bi-orthogonal sequence {ψn} and we
can consider an operator S defined initially on
Sφn = ψn
(II.4)
and extended on D(S) = span{φn} by the linearity. By the construction,
k(cid:88)
k(cid:88)
k(cid:88)
k(cid:88)
(Sf, f ) =
cncm(ψn, φm) =
cn2
for all
f =
cnφn ∈ D(S).
n=1
m=1
n=1
5
n=1
Therefore, S is a positive operator. For a semi-regular sequence {φn}, the operator S is densely
defined and the Friedrichs extension AF of S exists.
Theorem II.1 Let {φn} be a semi-regular sequence. The following are equivalent:
(i) {φn} is a GRS;
(ii) the Friedrichs extension AF of S is a positive operator;
(iii) the closure S of S is a positive operator and the relations
lim
n→∞(Sfn, fn) = 0
and
lim
n→∞ fn = g
(II.5)
imply that g = 0.
Proof -- (i) → (ii). If {φn} is a GRS, then e−Qφn = e−Q/2en = ψn. In view of (II.4), e−Q is a
positive self-adjoint extension of S and AF ≥ e−Q since the Friedrichs extension is the greatest
nonnegative self-adjoint extension of S. By virtue of Lemma II.1, AF is positive.
(ii) → (i). The positivity of AF means that AF = e−Q, where Q is a self-adjoint operator
in H. Denote en = e−Q/2φn. Due to (II.4), en = eQ/2ψn. Therefore, en ∈ D(eQ/2) ∩ D(e−Q/2)
and (en, em) = (e−Q/2φn, eQ/2ψm) = (φn, ψm) = δnm.
The orthonormal sequence {en} turns out to be an ONB if {en} is complete in H. Assume
that γ is orthogonal to {en} in H. Then there exists a sequence {fm} (fm ∈ D(e−Q)) such that
e−Q/2fm → γ in H (because e−Q/2D(e−Q) is a dense set in H). In this case, due to (II.3), {fm}
is a Cauchy sequence in H−Q and therefore, fm tends to some f ∈ H−Q. This means that
m→∞(e−Q/2fm, en) = lim
0 = (γ, en) = lim
(II.6)
By Lemma II.2, the set D(S) = span{φn} is dense in the Hilbert space H−Q. In view of
(II.6), f = 0 that means limm→∞ (cid:107)fm(cid:107)−Q = limm→∞ (cid:107)e−Q/2fm(cid:107) = 0 and therefore, γ = 0.
Hence, {en} is complete in H and {en} is an ONB of H.
m→∞(fm, φn)−Q = (f, φn)−Q.
The implication (ii) → (iii) is obvious.
(iii) → (ii). The operator S has the inverse S
−1
since S is positive. The operator S
−1
is
closed and (II.5) can be rewritten as follows:
−1
n→∞(gn, S
lim
gn) = 0
and
−1
lim
n→∞ S
gn = g
(gn = Sfn)
−1
This means that S
sion (Lemma II.3) Applying Lemma II.4 we complete the proof.
is positively closable. Hence, it admits a non-negative self-adjoint exten-
(cid:3)
6
Corollary II.1 A regular sequence is a GRS.
Proof -- Let {φn} be a regular sequence.
Its regularity means that R(S) is a dense set in
H. The latter means that each nonnegative self-adjoint extension A of S must be positive. In
particular, the Friedrichs extension AF is positive. By Theorem II.1, {φn} is a GRS.
(cid:3)
A shifting of the orthonormal Hermite functions en(x) in the complex plane gives rise to
regular sequences in L2(R).
In particular, eigenfunctions of the shifted harmonic oscillator
{φn(x) = en(x + ia)} form a regular sequence and φn = eQ/2en, where Q = 2ai d
dx is an un-
bounded self-adjoint operator in L2(R) [8, Subsection IV.1]. We refer [14, 17] for the relationship
between general regular sequences and some physical operators.
Remark II.1 Theorem II.1 can be generalized to the case of non-complete minimal sequence
{φn} such that its bi-orthogonal {ψn} is also non-complete. In this case, S is a non-densely
defined positive symmetric operator in H. We should suppose the existence of a positive self-
adjoint extension A of S. Similarly to the proof of Theorem II.1 we set A = e−Q and determine
orthonormal system {en} in H. By virtue of (II.6), the completeness of {en} in H is equivalent
to the completeness of {φn} in H−Q. This means that {en} is an ONB in H if {φn} is a complete
set in H−Q. Summing up: Let {φn} be a minimal sequence and let {ψn} be its bi-orthogonal
sequence. These sequences are GRS if and only if there exists a positive self-adjoint extension
A = e−Q of S such that {φn} is a complete set in the Hilbert space H−Q.
Another approach to the study of non-complete minimal sequences can be found in [6].
II.3 The uniqueness of Q in the definition of GRS
Let {φn} be a basis in H. Then {φn} is a regular sequence because its bi-orthogonal sequence
{ψn} has to be a basis [19, Corollary 5.22]. By Corollary II.1, {φn} is a GRS, i.e., φn = eQ/2en.
Moreover, by [8, Proposition II.9], the pair (Q,{en}) in Definition 1 is determined uniquely for
every basis {φn}. For this reason, a natural question arise: is the pair (Q,{en}) determined
uniquely for a given GRS {φn}?
The choice of the Friedrichs extension AF = e−Q of S in the proof of Theorem II.1 was
related to the fact that D(S) must be dense in the Hilbert space H−Q (that, in view of (II.6),
is equivalent to the completeness of {en} in H). Due to Lemma II.2, each positive extremal
extension A = e−Q can be used instead of AF in the proof of Theorem II.1. This observation
leads to the following result ([8, Proposition II.10]):
7
Proposition II.1 Let a semi-regular sequence {φn} be a GRS. Then a self-adjoint operator Q
and an ONB {en} are determined uniquely in the formula φn = eQ/2en if and only the symmetric
operator S in (II.4) has a unique positive extremal extension.
Remark II.2 The above mentioned unique positive extremal extension coincides with the Friedrichs
extension AF . Indeed, the existence of another positive extension A (cid:54)= AF means that AF is
also positive (Lemma II.1). Due to the uniqueness of positive extension, we get A = AF .
II.4 Bases and Bessel sequences
Various classes of GRS's can be easily characterized in terms of spectral properties of the
corresponding self-adjoint operators Q.
We recall that a semi-regular sequence {φn} is called a Riesz basis if there exists 0 < a ≤ b
a(f, f ) = a(cid:107)f(cid:107)2 ≤(cid:88)
such that1
a(cid:107)f(cid:107)2 ≤(cid:88)
(f, φn)2 ≤ b(cid:107)f(cid:107)2
for all
f ∈ H.
(II.7)
n
Proposition II.2 The following are equivalent:
(i) a sequence {φn} is a Riesz basis with bounds 0 < a ≤ b;
(ii) {φn} is a GRS, i.e., φn = eQ/2en, where Q is bounded self-adjoint operator such that
σ(Q) ⊂ [ln a, ln b].
Proof -- If {φn} is a Riesz basis, then φn = eQ/2en, where Q is a bounded operator (see Section
I). The substitution of φn = eQ/2en into (II.7) gives
(eQ/2f, en)2 ≤ (cid:107)eQ/2f(cid:107)2 = (eQf, f ) ≤ b(cid:107)f(cid:107)2 = b(f, f ).
Therefore, aI ≤ eQ ≤ bI that justifies (i) → (ii). The converse statement is obvious.
(cid:3)
n
Lemma II.5 Let Q be a self-adjoint operator such that σ(Q) ⊂ (−∞, ln b] and let {en} be an
arbitrary ONB of H. Then the sequence {φn = eQ/2en} is a GRS with the pair (Q,{en}).
1we refer [19, Theorems 7.13, 8.32] for equivalent definitions of Riesz basis
8
Proof -- Since σ(Q) ⊂ (−∞, ln b], the self-adjoint operator eQ/2 is bounded. Hence, the
If γ is orthogonal to {φn}, then 0 = (γ, eQ/2en) =
elements φn = eQ/2en are well-defined.
(eQ/2γ, en) implies that eQ/2γ = γ = 0. Therefore {φn} is semi-regular and its bi-orthogonal
sequence {ψn} is defined uniquely. According to Definition 1, it is sufficient to show that
en ∈ D(e−Q/2) and ψn = e−Q/2en. By virtue of the relation δmn = (φm, ψn) = (eQ/2em, ψn) =
(em, eQ/2ψn) we obtain that eQ/2ψn = en . The last relation means that en ∈ D(e−Q/2) and
(cid:3)
ψn = e−Q/2en.
A sequence {φn} is called a Bessel sequence if there exists b > 0 such that
(cid:88)
n
(f, φn)2 ≤ b(cid:107)f(cid:107)2
for all
f ∈ H.
(II.8)
Theorem II.2 The following are equivalent:
(i) a semi-regular sequence {φn} is a Bessel sequence;
(ii) {φn} is a GRS, i.e., φn = eQ/2en, where Q is a self-adjoint operator such that σ(Q) ⊂
Proof -- (i) → (ii). If {φn} is a Bessel sequence, then the synthesis operator R{cn} =(cid:80) cnφn
implies that {φn} is ω-independent [19, p. 156]. The latter means that the series (cid:80) cnφn
defines a bounded operator which maps l2(N) into H [19, p. 190]. The minimality of {φn}
(−∞, ln b].
converges and equal 0 only when cn = 0. Therefore, ker R = {0}.
√
R∗R) onto the closure of R(R) [21, Chapter VI, Subsect. 2.7]. We remark that R(
Let {δn} be the canonical basis of l2(N). Then Rδn = φn. Identifying {δn} with an ONB
{en} of H we obtain a bounded operator2 R in H such Ren = φn. The polar decomposition of
√
R is R = R∗U , where R∗ =
RR∗ and U is an isometric operator mapping the closure of
√
R(
R∗R)
and R(R) are dense sets in H, since, respectively, ker
R∗R = ker R = {0} and {φn} is a
complete set in H. Therefore, U is a unitary operator in H. Moreover, R∗ is a positive
bounded self-adjoint operator (since kerR∗ = ker R = {0}).
The positivity of R∗ leads to the formula R∗ = eQ/2, where Q is a self-adjoint operator
√
in H. Denote en = U en. Obviously, {en} is an ONB of H and
φn = Ren = R∗U en = eQ/2U en = eQ/2en.
2we keep the same notation R for the operator in H.
9
After the substitution of φn = eQ/2en into (II.8):
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(f, eQ/2en2 =
(eQ/2f, en2 = (cid:107)eQ/2f(cid:107)2 = (eQf, f ) ≤ b(f, f ).
n
n
The obtained inequality leads to the conclusion that σ(Q) ⊂ (−∞, ln b]. Applying Lemma II.5
we complete the proof of (i) → (ii).
(ii) → (i). In view of Lemma II.5, {φn} is a semi-regular sequence. The operator eQ is
bounded and (cid:107)eQ(cid:107) ≤ b (since σ(Q) ⊂ (−∞, ln b]). Hence,
(f, φn)2 =
(eQ/2f, en2 = (cid:107)eQ/2f(cid:107)2 = (eQf, f ) ≤ b(cid:107)f(cid:107)2
n
n
that completes the proof.
(cid:3)
A sequence {φn} is called bounded if 0 < a ≤ (cid:107)φn(cid:107) ≤ b for all n. A basis {φn} is called
conditional if its property of being basis depends on the permutation of elements φn.
Proposition II.3 Let Q be a self-adjoint operator in H such that σ(Q) ⊂ (−∞, ln b]. The
following are equivalent:
(i) there exists an ONB {en} of H such that the sequence {φn = eQ/2en} is a conditional
bounded basis;
(ii) there exists β < 0 such that each interval [(n + 1)β, nβ] (n = 0, 1, . . .) includes at least
one point of essential spectrum of Q.
Proof -- Applying [25, Theorem 1] to the positive bounded operator eQ/2 and taking into
account properties of an essential spectrum [26, Proposition 8.11] we arrive at the conclusion
that the item (i) is equivalent to the existence of 0 < q < 1 such that the essential spectrum
of eQ/2 has a non-zero interaction with each interval [qn+1, qn]. Since Q = 2 ln eQ/2, the later
(cid:3)
statement is equivalent to (ii) with β = 2 ln q.
Let H = L2(−π, π) and Q is an operator of multiplication by α lnx (0 < α) in H. Obvi-
ously, Q is self-adjoint, its spectrum coincides with (−∞, ln πα] and it is essential. By Propo-
sition II.3, there exists an ONB {en} of L2(−π, π) such that {φn = eQ/2en} is a conditional
bounded basis.
2, the
corresponding ONB can be chosen as {en = 1√
In view of the Babenko example [19, Example 5.13], for 0 < α < 1
einx}∞
−∞.
2π
10
J-orthonormal sequences and GRS
II.5
Let J be a bounded self-adjoint operator in a Hilbert space H such that J 2 = I. The Hilbert
space H equipped with the indefinite inner product [·,·] := (J·,·) is called a Krein space.
A sequence {φn} is called J-orthonormal if [φn, φm] = δnm.
Each J-orthonormal sequence {φn} is minimal since its bi-orthogonal one is determined as
ψn = [φn, φn]Jφn.
(II.9)
In view of (II.9), the positive symmetric operator S in (II.4) acts as Sφn = [φn, φn]Jφn.
Proposition II.4 Let {φn} be a complete J-orthonormal sequence. Then {φn} is a Bessel
sequence if and only if {φn} is a Riesz basis.
Proof -- Let us assume that {φn} is a Bessel sequence. Then {ψn} is also a Bessel sequence.
Indeed, substituting Jf instead of f into (II.8) and using (II.9), we obtain
(cid:88)
(Jf, φn)2 =
(cid:88)
(cid:88)
(f, Jφn)2 =
(f, ψn)2 ≤ b(cid:107)Jf(cid:107)2 = b(cid:107)f(cid:107)2.
n
n
n
By Theorem II.2, {φn} is a GRS and φn = eQ/2en, where σ(Q) ⊂ (−∞, ln b]. Since {ψn}
is also a Bessel sequence, applying Theorem II.2 again we obtain σ(−Q) ⊂ (−∞, ln b] or
σ(Q) ⊂ [− ln b,∞). Therefore, σ(Q) ⊂ [ln a, ln b], where a = 1/b. In view of Proposition II.2,
{φn} is a Riesz basis. The inverse statement is obvious.
(cid:3)
If {φn} is complete in H, then {ψn} in (II.9) is complete too. Therefore, {φn} is regular
and, by Corollary II.1, {φn} is a GRS. Thus, each complete J-orthonormal sequence is a GRS.
It follows from the proof of Corollary II.1 that each extremal extension A of S is positve
Therefore, the corresponding operator Q = − ln A in Definition 1 can be determined by every
extremal extension A. If Q is determined uniquely, then [8, Theorem III.3]:
JQ = −QJ.
(II.10)
However, if Q is not determined uniquely, not each Q = − ln A satisfies (II.10). In particular,
as follows from [20], the operator Q that corresponds to the Friedrichs extension AF does not
satisfy (II.10). Moreover, there exist complete J-orthonormal sequences for which no operators
Q satisfying (II.10) can be found.
We say that a complete J-orthonormal sequence {φn} is of the first type if there exists a
self-adjoint operator Q in Definition 1 such that (II.10) holds. Otherwise, {φn} is of the second
type.
11
J-orthonormal bases are examples of the first type sequences. The next statement was
proved in [8], where the notation "quasi-bases" was used for the first type sequences.
Proposition II.5 The following are equivalent:
(i) a complete J-orthonormal sequence {φn} is of the first type;
(ii) the sequence {φn} is regular and the corresponding pair (Q,{en}) in Definition 1 can be
chosen as follows: Q satisfies (II.10) and en are eigenfunctions of J, i.e., Jen = en or
Jen = −en.
In what follows, considering a first type sequence {φn = eQ/2en}, we assume that the pair
(Q,{en}) satisfies conditions (ii) of Proposition II.5. A detailed analysis of the first/second
type sequences can be found in [8]. We just mention that a first type sequence {ψn = eQ/2en}
generates a C-symmetry operator3 C = eQJ with the same operator Q. The latter allows one
to construct the new Hilbert space H−Q involving {φn} as ONB, directly as the completion of
D(C) with respect to "CPT -norm": (·,·)−Q = [C·,·] = (JeQJ·,·) = (e−Q·,·).
For a second type sequence, the inner product (·,·)−Q defined by (II.3) cannot be expressed
via [·,·] and one should apply much more efforts for the precise definition of (·,·)−Q.
III Examples
III.1 A semi-regular sequence which cannot be a GRS
Let {en}∞
n=0 be an ONB of H. Denote
1
nα e0,
1
nβ en +
φn =
α, β ∈ R, n = 1, 2, . . .
n=1 is bi-orthogonal to {φn}. It is easy to
The sequence {φn}∞
n=1 is minimal since {ψn = nβen}∞
see that {φn} is complete in H if and only if α− β ≤ 1
2. The last relation determines admissible
parameters α, β for which {φn} is a semi-regular sequence. During the subsection we suppose
that this inequality holds.
In view of (II.4), S( 1
nα e0) = nβen and the operator S can be described as:
nβ en + 1
Sf =
n2βcnen
for all
f =
cnen +
n=1
n=1
n=1
e0 ∈ D(S).
(III.1)
3the concept of C-symmetry is widely used in PT -symmetric quantum mechanics [7, 10]
12
k(cid:88)
k(cid:88)
(cid:32) k(cid:88)
(cid:33)
nβcn
nα
It follows from (III.1) that the non-negative self-adjoint operator
(cid:32) ∞(cid:88)
(cid:33)
∞(cid:88)
cnen
=
n2βcnen
with the domain D(A) =(cid:8)f =(cid:80)∞
Af = A
n=0
n=1
n=0 cnen : {cn}∞
n=1,{n2βcn}∞
n=1 ∈ (cid:96)2(N)(cid:9) is an extension of S.
(III.2)
Assume that β ≤ 0. Then the semi-regular sequence {φn} cannot be a GRS. Indeed, in this
case, the operator A is bounded. Therefore, A coincides with the closure S of S. In view of
(III.2), Se0 = Ae0 = 0. By Theorem II.1, {φn} cannot be a GRS.
Assume now that β > 0. Then A is an unbounded non-negative self-adjoint extension of S.
Hence, A is an extension of S. Using (III.1) and (III.2), we obtain
D(S) =
f =
cnen : {cn}∞
n=0
nβcn
nα converges and c0 =
n=1 ∈ (cid:96)2(N)
n=1, {n2βcn}∞
∞(cid:88)
nβcn
nα
n=1
(cid:41)
.
and
Since (Sf, f ) =(cid:80)∞
Let fm =(cid:80)∞
n=1 n2βcn2, the operator S is positive. Using item (iii) of Theorem II.1 we
2. To that end, it suffices to verify
show that the semi-regular sequence {φn} is a GRS for α > 1
the implication (II.5).
n=0 cm
n en be a sequence of elements fm ∈ D(S) satisfying (II.5). Then
∞(cid:88)
(cid:40)
∞(cid:88)
n=1
∞(cid:88)
m→∞(Sfm, fm) = lim
lim
m→∞
n2βcm
n 2 = (cid:107){nβcm
n }(cid:107)2
(cid:96)2(N) = 0
and, since {1/nα} ∈ (cid:96)2(N) for α > 1
2,
g = lim
m→∞ fm = lim
m→∞
cm
n en + lim
m→∞
∞(cid:88)
n=1
n=1
∞(cid:88)
n=1
nβcm
n
nα e0 = lim
m→∞({nβcm
n },{1/nα})(cid:96)2(N)e0 = 0
that justifies the implication (II.5).
J-orthonormal sequences of the first/second type
III.2
Let a sequence of real numbers {αk}∞
k=0 satisfy the conditions
0 ≤ α0 < α1 < α2 . . . ,
lim
n=0 be an ONB of H such that Jen = (−1)nen.
k→∞ αk = ∞
and let {en}∞
(III.3)
13
Each pair of orthonormal vectors {e2k, e2k+1}∞
k=0 can be identified with C2 assuming that
(cid:34)
(cid:35)
1
0
(cid:34)
(cid:35)
0
1
U e2k =
,
U e2k+1 =
.
(III.4)
(cid:34)
(cid:35)
The operator U is an isometric mapping of the space Hk = span{e2k, e2k+1} onto C2 and
k=0 ⊕Hk, the operator U can be extended
U J = σ3U , where σ3 =
to the isometric mapping of H onto the Hilbert space H of infinitely many copies of C2: H =
. Since H =(cid:80)∞
(cid:80)∞
k=0 ⊕C2. In the space H, we define self-adjoint operators
∞(cid:88)
1
0
0 −1
∞(cid:88)
⊕e−2αkσ1
Q/2 =
e
⊕eαkσ1,
Q = 2
⊕αkσ1,
e−Q
=
∞(cid:88)
k=0
⊕σ3
J =
(III.5)
where σ1 =
k=0
k=0
. Theirs unitary equivalent copies in H are:
∞(cid:88)
(cid:34)
k=0
0 1
1 0
(cid:35)
Q = U−1QU,
eQ/2 = U−1e
By the construction, Q anticommutes with J: JQ = −QJ.
e−Q = U−1e−Q
U,
Consider vectors {φn}∞
n=0 defined by the formulas:
Q/2U,
J = U−1JU.
(III.6)
φ2k+1 =
ck√
µ2k+1
∞(cid:88)
n=0
φ2k = cosh αke2k + sinh αke2k+1,
k = 0, 1, . . .
χn cosh αn
1 − µ2k+1 cosh2 αn
(cosh αne2n+1 + sinh αne2n),
(III.7)
where {χn}∞
µ5 . . . < 1 are roots of the equation
n=0 is a vector from (cid:96)2(N) such that χn (cid:54)= 0; the set of numbers 0 < µ1 < µ3 <
∞(cid:88)
n=0
χn cosh αn2
1 − µ cosh2 αn
χn2 cosh4 αn
(1 − µ2k+1 cosh2 αn)2
(cid:33)− 1
2
and
(cid:32) ∞(cid:88)
n=0
ck =
= 0
(III.8)
,
k = 0, 1 . . .
(III.9)
Theorem III.1 Let the sequences {αn} and {χn} satisfy the conditions above and let the se-
quence {χn cosh2 αn} do not belong to (cid:96)2(N). Then the vectors φn determined by (III.7) form a
14
complete J-orthonormal sequence {φn}∞
second type if {χn cosh αn} ∈ (cid:96)2(N).
n=0 of the first type if {χn cosh αn} (cid:54)∈ (cid:96)2(N) and of the
For the first type sequence {φn} the formula φn = eQ/2en holds where Q and eQ/2 are
determined by (III.6) and an ONB {en} has the form
e2k = e2k,
e2k+1 =
ck√
µ2k+1
χn cosh αn
1 − µ2k+1 cosh2 αn
e2n+1.
(III.10)
For the second type sequence such a choice of Q and {en} is impossible because the orthonor-
mal system (III.10) is not dense in H. A suitable operator Q can be chosen as Q = − ln AF ,
where AF is the Friedrichs extension of the symmetric operator S acting as Sφn = (−1)nJφn
on vectors φn and extended onto D(S) = span{φn} by the linearity.
∞(cid:88)
n=0
The proof of Theorem III.1 is given in Section IV.
Let us consider a particular case assuming that
tanh αn =
,
χn =
1
(n + 1) δ+1
2
, n ≥ 0,
n + 1
and 0 < δ ≤ 2 (the condition 0 < δ guarantees that {χn} ∈ (cid:96)2(N) while δ ≤ 2 ensures that
{χn cosh2 αn} (cid:54)∈ (cid:96)2(N)). Then the root equation (III.8) takes the form
(cid:114) n
∞(cid:88)
n=1
1
nδ ·
1
1 − nµ
= 0,
(III.11)
(cid:16)(cid:80)∞
n=1
ck =
n1−δ
(1−µ2k+1n)2
(cid:17)− 1
2 , and the sequence {φn}∞
n=0:
√
√
ne2n−1 +
∞(cid:88)
k + 1e2k +
ke2k+1,
φ2k =
√
1
1 − nµ2k+1
(
ck√
µ2k+1
1
nδ/2 ·
n=1
φ2k+1 =
k = 0, 1, 2 . . .
√
n − 1e2n−2),
turns out to be the first kind if 0 < δ ≤ 1 and the second kind if 1 < δ ≤ 2.
Figure 1 contains a numerical localization of the first 5 roots for of (III.11).
IV Appendix: the proof of Theorem III.1
We refer [8, Subsection III.1] and [20] for results of the Krein space theory which are necessary
for our exposition.
15
Figure 1: First five roots to the equation (III.11) for different parameter value.
IV.1 Preliminaries
An operator T defined by the formula
T e2k = tanh αke2k+1,
(IV.1)
n=0 and extended onto H by the linearity is a self-adjoint strong contraction
T e2k+1 = tanh αke2k,
on the ONB {en}∞
('strong' means that (cid:107)T f(cid:107) < (cid:107)f(cid:107) for non-zero f ). Moreover
k = 0, 1, 2 . . .
(IV.2)
since Jen = (−1)nen. The properties of T allow one to define J-orthogonal maximal positive
L+ and maximal negative L− subspaces of the Krein space (H, [·,·]) [20, Lemma 2.2]:
JT = −T J,
L+ = (I + T )H+,
L− = (I + T )H−,
(IV.3)
where H+ and H− are the closure (in H) of span{e2k}∞
k=0 and span{e2k+1}∞
k=0, respectively.
k=0 χke2k+1, where {χk} belongs to (cid:96)2(N). By the construction,
Consider a vector χ = (cid:80)∞
χ ∈ H− and
L0− = {(I + T )g, g ∈ M−}, where M− = {g ∈ H−, (g, χ) = 0}
(IV.4)
16
is a subspace of the maximal negative space L− defined by (IV.3).
The J-orthogonal sum L+ +L− is dense in H since L± are maximal subspaces. However, we
can not state that the set L+ +L0− remains dense in H since L0− is a proper subspace of L−.
Lemma IV.1 The sum L+ +L0− is dense in H if and only if(cid:8)χn cosh2 αn
(cid:9) (cid:54)∈ (cid:96)2(N).
Proof -- Assume that h ∈ H is orthogonal to L+ +L0−. Then Jh is orthogonal to L+ with
respect to the indefinite inner product [·,·]. Since L− is the J-orthogonal complement of L+,
the vector Jh belongs to L−. By (IV.3), Jh = (I + T )f , where f ∈ H− and h = J(I + T )f =
(I − T )Jf = −(I − T )f.
By the assumption, h is also orthogonal to L0−.
loss of generality we can assume that (I − T 2)f = χ. Here, f =(cid:80)∞
In view of (IV.4) this means that 0 =
(h, (I + T )g) = −((I − T )f, (I + T )g) = −((I − T 2)f, g) for all g ∈ M−. Therefore, without
k=0 fke2k+1 since f ∈ H−.
∞(cid:88)
In view of (IV.1),
∞(cid:88)
fk
(I − T 2)f =
(1 − tanh2 αk)fke2k+1 =
cosh2 αk
k=0
k=0
e2k+1 =
χke2k+1 = χ.
∞(cid:88)
k=0
Therefore, fk = χk cosh2 αk and we arrive at the conclusion that L+ +L0− is a non-dense set in
H if and only if {χk cosh2 αk}∞
(cid:3)
k=0 ∈ (cid:96)2(N).
IV.2 Complete J-orthonormal sequence {φn}
Lemma IV.2 If {χn cosh2 αn} (cid:54)∈ (cid:96)2(N), then the vectors {φn} defined by (III.7) form a com-
plete J-orthonormal sequence in H.
Proof -- The vectors {φ2k}∞
k=0 in (III.7) are J-orthonormal because [φ2k, φ2k(cid:48)] = 0 for k (cid:54)= k(cid:48)
and [φ2k, φ2k] = cosh2 αk − sinh2 αk = 1. Moreover, in view of of (IV.1), the vectors {φ2k} can
be presented as φ2k = cosh αk(I + T )e2k. This relation and (IV.3) imply that the closure of
span{φ2k}∞
k=0 coincides with L+.
By virtue of (IV.1),
(I − T 2)f =
∞(cid:88)
k=0
(1 − tanh2 αk)fke2k+1
for all
f =
∞(cid:88)
k=0
fke2k+1 ∈ H−.
This relation yields that I−T 2 is a compact operator in H−, since limk→∞(1−tanh2 αk) = 0, see
[18, problem 132]. Therefore, PM−(I − T 2)PM−, where PM− is an orthogonal projection in H−
17
onto the subspace M− defined in (IV.4), is a self-adjoint compact operator in M−. This implies
the existence of an ONB {γ2k+1} of M− which is formed by eigenfunctions of PM−(I − T 2)PM−.
Let {µ2k+1}∞
k=0 be the corresponding eigenvalues, i.e., PM−(I − T 2)γ2k+1 = µ2k+1γ2k+1. Since
I − T 2 is a positive contraction, we can state that 0 < µ2k+1 < 1 and limk→∞ µ2k+1 = 0.
Denote
φ2k+1 =
1√
µ2k+1
The vectors {φ2k+1}∞
k=0 are J-orthonormal because
[φ2k+1, φ2k(cid:48)+1] = −((I − T 2)γ2k+1, γ2k(cid:48)+1)
√
µ2k+1µ2k(cid:48)+1
= −
(cid:114) µ2k+1
µ2k(cid:48)+1
(γ2k+1, γ2k(cid:48)+1) = −δkk(cid:48).
(I + T )γ2k+1,
k = 0, 1 . . .
(IV.5)
Moreover, in view of (IV.4), φ2k+1 ∈ L0− and the closure of span{φ2k+1}∞
k=0 coincides with L0−.
Applying now Lemma IV.1 we arrive at the conclusion that, for the case {χn cosh2 αn} (cid:54)∈ (cid:96)2(N),
the J-orthonormal sequence {φn}∞
k=0 are defined by (III.7) and
(IV.5), respectively, is complete in H.
n=0, where {φ2k}∞
k=0 and {φ2k+1}∞
To finish the proof of Lemma IV.2 it suffices to show that the formulas (III.7) and (IV.5)
k=0. To do that, we describe the eigenvalues µ2k+1 and the
determine the same vectors {φ2k+1}∞
normalized eigenfunctions γ2k+1 of the equation
In view of (IV.4), the condition g ∈ M− means that
PM−(I − T 2)g = µg,
∞(cid:88)
gkχk = 0, where
g =
g ∈ M−.
(IV.6)
∞(cid:88)
k=0
gke2k+1.
(IV.7)
k=0 fke2k+1 be an arbitrary element of H−. Then
PM−f =
(fk − α[f ]χk)e2k+1,
α[f ] =
1
(cid:107)χ(cid:107)2
∞(cid:88)
k=0
fkχk.
(IV.8)
k=0
∞(cid:88)
k=0
Let f =(cid:80)∞
Using (IV.8) we rewrite (IV.6) as
∞(cid:88)
∞(cid:88)
PM−
(1 − tanh2 αk)gke2k+1 =
[(1 − tanh2 αk)gk − α[(I − T 2)g]χk]e2k+1 = µ
k=0
k=0
∞(cid:88)
k=0
gke2k+1
that implies
(1 − tanh2 αk − µ)gk = α[(I − T 2)g]χk,
k = 0, 1, . . .
(IV.9)
18
It is important that α[(I − T 2)g] (cid:54)= 0 in (IV.9). Indeed, if α[(I − T 2)g] = 0, then (1− tanh2 αk −
µ)gk = 0 for all k. Due to conditions imposed on αn in (III.3), there exists a unique k(cid:48) such
(cid:54)= 0 and µ = 1 − tanh2 αk(cid:48). This means that g = gk(cid:48)e2k(cid:48)+1 belongs to M−. The
that gk(cid:48)
(cid:54)= 0 (we recall that χn (cid:54)= 0 for all n by
last fact is impossible because 0 =< g, χ >= gk(cid:48)χk(cid:48)
the assumption). The obtained contradiction shows that α[(I − T 2)g] (cid:54)= 0. This means that
(1 − tanh2 αk − µ)gk (cid:54)= 0 and (IV.9) can be rewritten as
gk = α[(I − T 2)g]
The corresponding solution g(µ) = (cid:80)∞
χk cosh2 αk
1 − µ cosh2 αk
,
k = 0, 1, 2 . . .
k=0 gke2k+1 of (IV.6) must be in M−. By virtue of
(IV.7), g(µ) belongs to M− if and only if µ is the root of (III.8). The equation (III.8) has
infinitely many roots 0 < µ1 < µ3 < . . . < µ2k+1 . . . < 1 that coincide with eigenvalues of
PM−(I − T 2)PM−. The eigenfunctions of PM−(I − T 2)PM− corresponding to µ2k+1 have the
form
g(µ2k+1) = α[(I − T 2)g(µ2k+1)]
χn cosh2 αn
1 − µ2k+1 cosh2 αn
e2n+1,
k = 0, 1 . . .
Then
γ2k+1 =
g(µ2k+1)
(cid:107)g(µ2k+1)(cid:107) = ck
χn cosh2 αn
1 − µ2k+1 cosh2 αn
e2n+1,
k = 0, 1 . . . ,
where the normalizing factor ck is defined in (III.9). By the construction, {γ2k+1}∞
k=0 is an
ONB of M−. Substituting the obtained expression for γ2k+1 into (IV.5) and taking (IV.1) into
account, we obtain the vectors {φ2k+1} from (III.7).
(cid:3)
∞(cid:88)
n=0
∞(cid:88)
n=0
Lemma IV.3 If {χn cosh αn} (cid:54)∈ (cid:96)2(N), then the sequence {φn} is of the first type. The corre-
sponding operator Q in Definition 1 is defined by (III.6) while ONB {en} has the form (III.10).
Proof -- If {χn cosh αn} (cid:54)∈ (cid:96)2(N), then {χn cosh2 αn} (cid:54)∈ (cid:96)2(N) and, by Lemma IV.2, {φn} is a
complete J-orthonormal sequence. In view of (II.9), the operator S defined by (II.4) acts as
Sφn = (−1)nJφn. On the other hand, taking the relation e−2αkσ1 = cosh 2αkσ0−sinh 2αkσ1 into
account, we directly verify that e−Qφn = (−1)nJφn, where e−Q is defined by (III.6). Therefore,
e−Q is a positive self-adjoint extension of S. Denote en = e−Q/2φn. In view of (III.5),
cosh αk − sinh αk
− sinh αk
cosh αk
This expression, (III.4), and (III.7) allow one to calculate {en} precisely, as (III.10).
⊕[cosh αkσ0 − sinh αkσ1]U = U−1
e−Q/2 = U−1
∞(cid:88)
∞(cid:88)
(cid:34)
(cid:35)
U.
⊕
k=0
k=0
19
By analogy with the proof of Theorem II.1 we obtain that {en} is an orthonormal sequence in
H. Moreover, {en} is an ONB if and only if {φn} is complete in the Hilbert space H−Q. Below we
show that the completeness of {φn} in H−Q is equivalent to the condition {χn cosh αn} (cid:54)∈ (cid:96)2(N).
We begin with the remark that
e−Q = (I − T )(I + T )−1,
=
(IV.10)
where T is determined by (IV.1). Indeed, since the subspaces Hk = span{e2k, e2k+1} are invari-
ant with respect to T and U satisfies (III.4), we get that U T U−1Hk acts as the multiplication
(cid:34)
by tanh αkσ1 in C2 and U (I − T )(I + T )−1U−1Hk coincides with
This relation and the decomposition H =(cid:80)∞
cosh 2αk − sinh 2αk
− sinh 2αk
cosh 2αk
k=0 ⊕Hk justify (IV.10).
cosh αk − sinh αk
− sinh αk
cosh αk
= (cosh 2αkσ0−sinh 2αkσ1)2 = e−2αkσ1.
(cid:35)2
(cid:34)
(cid:35)
The formulas (IV.3) and (IV.10) lead to the conclusion that D(e−Q) = L+ +L−, where the
subspaces L± are orthogonal with respect to the inner product (II.3). Therefore, the space
H−Q has the decomposition
where the subspaces (cid:98)L± are the completions of linear manifolds L± in H−Q.
To prove the completeness of {φn} in H−Q we note that {φ2k} is a complete set in(cid:98)L+. This
fact can be justified as follows: due to the proof of Lemma IV.2, span{φ2k} is dense in the
subspace L+ of H. In view of (IV.2), (IV.3), and (IV.10),
H−Q =(cid:98)L+ +(cid:98)L−
(cid:107)f(cid:107)2−Q = (e−Qf, f ) = ((I − T )x+, (I + T )x+) = [f, f ] ≤ (cid:107)f(cid:107)2
for each f = (I + T )x+ ∈ L+. Therefore, each f ∈ L+ can be approximated by vectors from
span{φ2k} with respect to the norm (cid:107) · (cid:107)−Q. Since (cid:98)L+ is the completion of L+ in H−Q, the set
{φ2k} is complete in (cid:98)L+.
Similar arguments and the fact that span{φ2k+1} is dense in the subspace L0− lead to the
to (cid:107) · (cid:107)−Q. Therefore, in order to proof the completeness of {φ2k+1} in (cid:98)L− it suffices to find
conclusion that each vector f ∈ L0− can be approximated by vectors of span{φ2k+1} with respect
when L0− turns out to be dense in (cid:98)L− with respect to (cid:107) · (cid:107)−Q.
Let h ∈ (cid:98)L− be orthogonal to L0− in H−Q. Since L− is dense in (cid:98)L− we can approximate h
by a sequence {fn}, where fn ∈ L−. In view of (II.3), the sequence {e−Q/2fn} is fundamental
20
in H and, hence, limn→∞ e−Q/2fn = f ∈ H. Due to (IV.3), fn = (I + T )xn−, where xn− ∈ H−.
Moreover, e−Q/2 = [(I − T )(I + T )−1]1/2 in view of (IV.10). This means that
e−Q/2fn = [(I − T )(I + T )−1]1/2(I + T )xn− = (I − T 2)1/2xn−
and, since (I − T 2)1/2 leaves H± invariant, f = limn→∞ e−Q/2fn = limn→∞(I − T 2)1/2xn− = f ∈
H−. On the other hand, for each vector (I +T )g ∈ L0−, the relation e−Q/2(I +T )g = (I−T 2)1/2g
holds. After such kind of auxiliary work we obtain:
0 = (h, (I + T )g)−Q = lim
n→∞(fn, (I + T )g)−Q = lim
(f, (I − T 2)1/2g) = ((I − T 2)1/2f, g)
n→∞(e−Q/2fn, e−Q/2(I + T )g) =
for all
g ∈ M−.
Therefore, without loss of generality we can assume that (I − T 2)1/2f = χ. Reasoning by
analogy with the final part of the proof of Lemma IV.1, we obtain that L0− is dense in (cid:98)L− if
and only if {χn cosh αn} (cid:54)∈ (cid:96)2(N). This relation guarantees the completeness of {φn} in H−Q.
(cid:3)
n, where {e(cid:48)
IV.3 The proof of Theorem III.1
The implication '{χn cosh αn} (cid:54)∈ (cid:96)2(N) → {φn} is a first type sequence' was proved in Lemma
IV.3. Let us assume that {φn} is first type, i.e., there exists a self-adjoint operator Q(cid:48) anti-
commuting with J and such that φn = eQ(cid:48)/2e(cid:48)
. Since Q(cid:48) anti-commutes with J, the operator C satisfies the relation
C2f = f for f ∈ D(C) and JC = e−Q(cid:48)
In view of
[7, Theorem 6.2.3], there exists J-orthonormal maximal positive L(cid:48)
+ and maximal negative
− subspaces of the Krein space (H, [·,·]) which uniquely characterize C in the following way:
L(cid:48)
Cf+ = f+ and Cf− = f− for f± ∈ L(cid:48)
±.
is a positive self-adjoint operator in H.
n} is an ONB of H.
Denote C = Je−Q(cid:48)
Since e−Q(cid:48)
k=0 is dense in L+). This yields that L(cid:48)
is an extension of S, we obtain Cφn = Je−Q(cid:48)
φn = JSφn = (−1)nφn. Therefore,
the operator C acts as the identity operator on elements of the subspace L+ defined by (IV.3)
(since span{φ2k}∞
− = L− since
the maximal negative subspace L(cid:48)
− is determined uniquelly as J-orthogonal complement of
L(cid:48)
+ = L+. We obtain that the J-orthogonal sum L+ +L− determines two operators Je−Q(cid:48)
and
Je−Q. Applying again [7, Theorem 6.2.3], we conclude that Q(cid:48) = Q, where Q is determined
n = e−Q(cid:48)/2φn = e−Q/2φn = en, where {en} is determined by (III.10).
by (III.6). In this case, e(cid:48)
Therefore, {en} is an ONB of H that, in view of the proof of Lemma IV.3 is equivalent to
+ = L+ and, moreover L(cid:48)
21
the condition {χn cosh αn} (cid:54)∈ (cid:96)2(N). The inverse implication 'first type sequence {φn} →
{χn cosh αn} (cid:54)∈ (cid:96)2(N) ' is proved.
If {φn} is second type, the choice of Q as in (III.6) leads to the non-complete orthonormal
sequence (III.10) in H. Trying to keep the GRS's formula φn = eQ/2en we have to use Q =
− ln A, where A is a positive extremal extension of S (without loss of generality, we may assume
that A = AF ). In this case, the ONB {en} will be different from (III.10).
Acknowledgements
This work was partially supported by the Faculty of Applied Mathematics AGH UST statutory
tasks within subsidy of Ministry of Science and Higher Education.
References
[1] Ando, T., Nishio, K.: Positive self-adjoint operators of positive symmetric operators. To-
hoku Math. Journ. 22, pp.65 -- 75 (1970)
[2] Arlinskii, Y. M., Hassi, S., Sebesty´en, Z., de Snoo, H.S.V.: On the class of extremal
extensions of a nonnegative operator. In: Recent Advances in Operator Theory and Re-
lated Topics the Bela Szokefalvi-Nagy Memorial Volume, Operator Theory: Advances and
Applications, Vol. 127. Birkhauser, pp.41 -- 81 (2001)
[3] Arlinskii, Y. M., Tsekanovskii, E.: M. Krein's research on semi-bounded operators, its con-
temporary developments, and applications. Oper. Theory Adv. Appl. Vol 190. Birkhauser
Basel pp.65 -- 112 (2009)
[4] Bagarello, F.: More mathematics on pseudo-bosons. J. Math. Phys. 54, 063512 (2013)
[5] Bagarello, F., Bellomonte, G.: Hamiltonians defined by biorthogonal sets. J. Phys. A 50,
145203 (2017)
[6] Bagarello, F., Inoue, H., Trapani, C.: Biorthogonal vectors, sesquilinear forms and some
physical operators. J. Math. Phys. 59, 033506 (2018)
[7] Bagarello, F., Gazeau, J.-P., Szafraniec, F. H., Znojil, M. (eds.): Non-Selfadjoint Operators
in Quantum Physics. Mathematical Aspects. J. Wiley & Sons, Hoboken, New Jersey (2015)
22
[8] Bagarello, F., Kuzhel, S.: Generalized Riesz systems and orthonormal sequences in Krein
spaces. arXiv:1810.05218 (2018)
[9] Bellomonte, G., Trapani, C.: Riesz-like bases in rigged Hilbert spaces. Zeitschr. Anal.
Anwend. 35 pp.243 -- 265 (2016)
[10] C. M. Bender, C. M. et al: PT -Symmetry in Quantum and Classical Physics. World
Scientific, Singapore (2019) https://doi.org/10.1142/q0178
[11] Bender, C. M., Fring A., Gunther, U., Jones, H. (eds.): Special issue on quantum physics
with non-Hermitian operators. J. Phys. A 45, issue 44 (2012)
[12] Christensen, O.: An Introduction to Frames and Riesz Bases, Birkhauser Boston (2003)
[13] Davies, E. B.: Wild spectral behaviour of anharmonic oscillators. Bull. London Math. Soc.
32, pp.432 -- 438 (1999)
[14] Inoue, H.: General theory of regular biorthogonal pairs and its physical operators. J. Math.
Phys. 57, 083511 (2016)
[15] Inoue, H.: Semi-regular biorthogonal pairs and generalized Riesz bases. J. Math. Phys.
57, 113502 (2016)
[16] Inoue, H., Takakura, M.: Regular biorthogonal pairs and pseudo-bosonic operators. J.
Math. Phys. 57, 083503 (2016)
[17] Inoue, H., Takakura, M.: Non-self-adjoint hamiltonians defined by generalized Riesz bases.
J. Math. Phys. 57, 083505 (2016)
[18] Halmos, P.R.: A Hilbert Space Problem Book. Springer New York (1982)
[19] Heil, C.: A Basis Theory Primer. In: Applied and Numerical Harmonic Analysis.
Birkhauser Boston (2011)
[20] Kamuda, A., Kuzhel, S., Sudilovskaja, V.: On dual definite subspaces in Krein space.
Complex Anal. Oper. Theory (2019) 13: 1011. doi.org/10.1007/s11785-018-0838-x
[21] Kato, T.: Perturbation Theory for Linear Operators. Springer-Verlag, Berlin (1966)
[22] Krejcir´ık, D., Siegl, P., Tater, M., Viola, J.: Pseudospectra in non-Hermitian quantum
mechanics. J. Math. Phys. 56, 103513 (2015) doi.org/10.1063/1.4934378
23
[23] Mityagin, B., Siegl, P., Viola, J.: Differential operators admitting various rates of spectral
projection growth. J. Funct. Anal. 272, pp.3129 -- 3175 (2017)
[24] Mostafazadeh, A.: Pseudo-Hermitian representation of quantum mechanics. Int. J. Geom.
Methods Mod. Phys. 7, pp.1191 -- 1306 (2010)
[25] Olevskii, A.M.: On operators generating conditional bases in a Hilbert space. Math. Notes,
12, pp.476 -- 482 (1972)
[26] Schmudgen, K.: Unbounded Self-adjoint Operators on Hilbert Space. Springer Dordrecht
Heidelberg New York London, pp.435 (2012)
[27] Siegl, P., Krejcir´ık, D.: On the metric operator for the imaginary cubic oscillator. Physical
review D. 86, 121702 (2012) doi.org/10.1103/PhysRevD.86.121702
24
|
1201.5007 | 1 | 1201 | 2012-01-24T15:15:44 | On the Interplay of Regularity and Decay in Case of Radial Functions I. Inhomogeneous spaces | [
"math.FA"
] | We deal with decay and boundedness properties of radial functions belonging to Besov and Lizorkin-Triebel spaces. In detail we investigate the surprising interplay of regularity and decay. Our tools are atomic decompositions in combination with trace theorems. | math.FA | math |
On the Interplay of Regularity and Decay in Case
of Radial Functions I. Inhomogeneous spaces
Winfried Sickel∗, Leszek Skrzypczak and Jan Vybiral
September 18, 2018
Abstract
We deal with decay and boundedness properties of radial functions be-
longing to Besov and Lizorkin-Triebel spaces.
In detail we investigate the
surprising interplay of regularity and decay. Our tools are atomic decomposi-
tions in combination with trace theorems.
Keywords: Besov and Lizorkin-Triebel spaces, radial functions, radial sub-
spaces of Besov and Lizorkin-Triebel spaces, radial functions of bounded vari-
ation, decay near infinity.
MSC 2010 numbers: 46E35, 26B35.
Contents
1 Introduction
2 Main results
2
2.1 The characterization of the traces of radial subspaces . . . . . . . . .
2.1.1 Traces of radial subspaces with p = ∞ . . . . . . . . . . . . .
2.1.2 Traces of radial subspaces with p < ∞ . . . . . . . . . . . . .
2.1.3 Traces of radial subspaces of Sobolev spaces
2.1.4 The trace in S ′(R)
2.1.5 The trace in S ′(R) and weighted function spaces of Besov and
5
6
6
7
. . . . . . . . . . 10
. . . . . . . . . . . . . . . . . . . . . . . . 11
Lizorkin-Triebel type . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.6 The regularity of radial functions outside the origin . . . . . . 12
2.2 Decay and boundedness properties of radial functions . . . . . . . . . 14
2.2.1 The behaviour of radial functions near infinity . . . . . . . . . 15
2.2.2 The behaviour of radial functions near infinity -- borderline cases 17
2.2.3 The behaviour of radial functions near the origin . . . . . . . 18
∗Corresponding author
1
2.2.4 The behaviour of radial functions near the origin -- borderline
cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Traces of radial subspaces -- proofs
3.1
21
Interpolation of radial subspaces . . . . . . . . . . . . . . . . . . . . . 21
3.1.1 Complex interpolation of radial subspaces
. . . . . . . . . . . 21
3.1.2 Real Interpolation of radial subspaces . . . . . . . . . . . . . . 21
3.2 Proofs of the statements in Subsection 2.1.1 . . . . . . . . . . . . . . 22
3.3 Proofs of the assertions in Subsection 2.1.2 . . . . . . . . . . . . . . . 24
3.3.1 Proof of Lemma 1 . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3.2 Characterizations of radial subspaces by atoms . . . . . . . . . 24
3.3.3 Proof of Theorem 3 . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.4 Proof of Theorem 4 . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.5 Proof of Remark 4 . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.6 Proof of Theorem 5 . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Proofs of the statements in Subsection 2.1.3 . . . . . . . . . . . . . . 36
3.5 Proof of the statements in Subsection 2.1.4 . . . . . . . . . . . . . . . 36
3.5.1 Proof of Lemma 2 . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5.2 Proof of Theorem 8 . . . . . . . . . . . . . . . . . . . . . . . . 37
3.6 Proof of the assertions in Subsection 2.1.5 . . . . . . . . . . . . . . . 39
3.7 Proof of the assertions in Subsection 2.1.6 . . . . . . . . . . . . . . . 39
3.8 Test functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4 Decay properties of radial functions -- proofs
48
4.1 Proof of Theorem 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Traces of BV -functions and consequences for the decay . . . . . . . . 51
4.2.1 Proof of Theorem 12 . . . . . . . . . . . . . . . . . . . . . . . 51
4.2.2 Proof of Theorem 11 . . . . . . . . . . . . . . . . . . . . . . . 56
4.3 Proof of the assertions in Subsection 2.2.3 . . . . . . . . . . . . . . . 56
4.3.1 Proof of Lemma 3 . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.3.2 Proof of Theorem 13 . . . . . . . . . . . . . . . . . . . . . . . 57
4.3.3 Proof of Lemma 4 . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4 Proof of Theorem 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
1
Introduction
At the end of the seventies Strauss [39] was the first who observed that there is an
interplay between the regularity and decay properties of radial functions. We recall
his
2
Radial Lemma: Let d ≥ 2. Every radial function f ∈ H 1(Rd) is almost everywhere
equal to a function ef , continuous for x 6= 0, such that
2 k f H 1(Rd)k ,
1−d
ef (x) ≤ c x
(1)
where c depends only on d.
Strauss stated (1) with the extra condition x ≥ 1, but this restriction is not needed.
The Radial Lemma contains three different assertions:
(a) the existence of a representative of f , which is continuous outside the origin;
(b) the decay of f near infinity;
(c) the limited unboundedness near the origin.
These three properties do not extend to all functions in H 1(Rd), of course.
In
particular, H 1(Rd) 6⊂ L∞(Rd), d ≥ 2, and consequently, functions in H 1(Rd) can be
unbounded in the neigborhood of any fixed point x ∈ Rd. It will be our aim in this
paper to investigate the specific regularity and decay properties of radial functions
in a more general framework than Sobolev spaces.
In our opinion a discussion
of these properties in connection with fractional order of smoothness results in a
better understanding of the announced interplay of regularity on the one side and
local smoothness, decay at infinity and limited unboundedness near the origin on
the other side. In the literature there are several approaches to fractional order of
smoothness. Probably most popular are Bessel potential spaces H s
p(Rd), s ∈ R,
or Slobodeckij spaces W s
p (Rd) (s > 0, s 6∈ N). These scales would be enough to
explain the main interrelations. However, for some limiting cases these scales are
not sufficient. For that reason we shall discuss generalizations of the Radial Lemma
in the framework of Besov spaces Bs
p,q(Rd).
These scales essentially cover the Bessel potential and the Slobodeckij spaces since
p,q(Rd) and Lizorkin-Triebel spaces F s
• W m
p (Rd) = F m
p,2(Rd), m ∈ N0, 1 < p < ∞;
• H s
p(Rd) = F s
p,2(Rd), s ∈ R, 1 < p < ∞;
• W s
p (Rd) = F s
p,p(Rd) = Bs
p,p(Rd), s > 0, s 6∈ N, 1 ≤ p ≤ ∞,
where all identities have to be understood in the sense of equivalent norms, see, e.g.,
[41, 2.2.2] and the references given there.
All three phenomena (a)-(c) extend to a certain range of parameters which we shall
characterize exactly. For instance, decay near infinity will take place in spaces with
s ≥ 1/p (see Theorem 10) and limited unboundedness near the origin in the sense
of (1) will happen in spaces such that 1/p ≤ s ≤ d/p (see Theorem 13). For s = 1/p
(or s = d/p) always the microscopic parameter q comes into play. We will study the
3
above properties also for spaces with p < 1. To a certain extent this is motivated by
the fact, that the decay properties of radial functions near infinity are determined
by the parameter p and the decay rate increases when p decreases, see Theorem 10.
Our main tools here are the following. Based on the atomic decomposition theorem
for inhomogeneous Besov and Lizorkin-Triebel spaces, which we proved in [32], we
shall deduce a trace theorem for radial subspaces which is of interest on its own.
Then this trace theorem will be applied to derive the extra regularity properties
of radial functions. To derive the decay estimates and the assertions on controlled
unboundedness near zero we shall also employ the atomic decomposition technique.
With respect to the decay it makes a difference, whether one deals with inhomo-
geneous or homogeneous spaces of Besov and Lizorkin-Triebel type. Homogeneous
spaces (with a proper interpretation) are larger than their inhomogeneous counter-
parts (at least if s > d max(0, 1
p − 1)). Hence, the decay rate of the elements of
inhomogeneous spaces can be better than that one for homogeneous spaces. This
turns out to be true. However, here in this article we concentrate on inhomogeneous
spaces. Radial subspaces of homogeneous spaces will be subject to the continua-
tion of this paper, see [33]. In a further paper [34] we shall investigate a few more
properties of radial subspaces like complex interpolation and characterization by
differences.
The paper is organized as follows. In Section 2 we describe our main results. Here
Subsection 2.1 is devoted to the study of traces of radial subspaces. In particular, in
2.1.1, we state also trace assertions for radial subspaces of Holder-Zygmund classes.
In Subsection 2.2 the behaviour radial functions near infinity and at the origin is
investigated. Within the borderline cases in Subsection 2.2.2 the spaces BV (Rd)
show up. In this context we will also deal with the trace problem for the associated
radial subspaces. All proofs will be given in Sections 3 and 4. There also additional
material is collected, e.g., in Subsection 3.1 we deal with interpolation of radial
subspaces, in Subsection 3.3.2 we recall the characterization of radial subspaces by
atoms as given in [32], and finally, in Subsection 3.8 we discuss the regularity prop-
erties of some families of test functions.
Besov and Lizorkin-Triebel spaces are discussed at various places, we refer, e.g., to
the monographs [26, 29, 41, 42, 44] and to the fundamental paper [15]. We will not
give definitions here and refer for this to the quoted literature.
The present paper is a continuation of [32], [36] and [21].
Acknowledgement. The authors would like to thank the referee for a constructive
hint to simplify the proof of Theorem 1.
4
Notation
As usual, N denotes the natural numbers, N0 := N ∪ {0}, Z denotes the integers
and R the real numbers. If X and Y are two quasi-Banach spaces, then the symbol
X ֒→ Y indicates that the embedding is continuous. The set of all linear and
bounded operators T : X → Y , denoted by L(X, Y ), is equipped with the standard
quasi-norm. As usual, the symbol c denotes positive constants which depend only
on the fixed parameters s, p, q and probably on auxiliary functions, unless otherwise
stated; its value may vary from line to line. Sometimes we will use the symbols
∼ " instead of "≤" and "≥", respectively. The meaning of A <
" <
∼ B
is given by: there exists a constant c > 0 such that A ≤ c B. Similarly >
∼ is
defined. The symbol A ≍ B will be used as an abbreviation of A <
∼ A.
We shall use the following conventions throughout the paper:
∼ " and " >
∼ B <
• If E denotes a space of functions on Rd then by RE we mean the subset of
radial functions in E and we endow this subset with the same quasi-norm as
the original space.
• Inhomogeneous Besov and Lizorkin-Triebel spaces are denoted by Bs
p,q and
p,q, respectively. If there is no reason to distinguish between these two scales
F s
we will use the notation As
p,q. Similarly for the radial subspaces.
• If an equivalence class [f ] (equivalence with respect to coincidence almost
everywhere) contains a continuous representative then we call the class con-
tinuous and speak of values of f at any point (by taking the values of the
continuous representative).
• Throughout the paper ψ ∈ C ∞
0 (Rd) denotes a specific radial cut-off function,
i.e., ψ(x) = 1 if x ≤ 1 and ψ(x) = 0 if x ≥ 3/2.
2 Main results
This section consists of two parts. In Subsection 2.1 we concentrate on trace the-
orems which are the basis for the understanding of the higher regularity of radial
functions outside the origin. Subsection 2.2 is devoted to the study of decay and
boundedness properties of radial functions in dependence on their regularity. To
begin with we study the decay of radial functions near infinity. Special emphasize
is given to the limiting situation which arises for s = 1/p. Then we continue with
an investigation of the behaviour of radial functions near the origin. Also here we
investigate the limiting situations s = d/p and s = 1/p in some detail.
5
2.1 The characterization of the traces of radial subspaces
Let d ≥ 2. Let f : Rd → C be a locally integrable radial function. By using a
Lebesgue point argument its restriction
f0(t) := f (t, 0, . . . , 0) ,
t ∈ R .
is well defined a.e. on R. However, this restriction need not be locally integrable.
A simple example is given by the function
f (x) := ψ(x) x−1 ,
x ∈ Rd .
Furthermore, if we start with a measurable and even function g : R → C, s.t. g
is locally integrable on all intervals (a, b), 0 < a < b < ∞, then (again using a
Lebesgue point argument) the function
f (x) := g(x) ,
x ∈ Rd
is well-defined a.e. on Rd and is radial, of course. In what follows we shall study
properties of the associated operators
tr : f 7→ f0
and
ext : g 7→ f .
Both operators are defined pointwise only. Later on we shall have a short look onto
the existence of the trace in the distributional sense, see Subsection 2.1.4. Probably
it would be more natural to deal with functions defined on [0, ∞) in this context.
However, that would result in more complicated descriptions of the trace spaces.
So, our target spaces will be spaces of even functions defined on R.
2.1.1 Traces of radial subspaces with p = ∞
The first result is maybe well-known but we did not find a reference for it. Let
m ∈ N0. Then C m(Rd) denotes the collection of all functions f : Rd → C such that
all derivatives Dαf of order α ≤ m exist, are uniformly continuous and bounded.
We put
k f C m(Rd)k := Xα≤m
k Dαf L∞(Rd)k .
Theorem 1 Let d ≥ 2. For m ∈ N0 the mapping tr is a linear isomorphism of
RC m(Rd) onto RC m(R) with inverse ext .
Remark 1 If we replace uniformly continuous by continuous in the definition of
the spaces C m(Rd) Theorem 1 remains true with the same proof.
Using real interpolation it is not difficult to derive the following result for the spaces
of Holder-Zygmund type.
Theorem 2 Let s > 0 and let 0 < q ≤ ∞. Then the mapping tr is a linear
isomorphism of RBs
∞,q(R) with inverse ext .
∞,q(Rd) onto RBs
6
2.1.2 Traces of radial subspaces with p < ∞
Now we turn to the description of the trace classes of radial Besov and Lizorkin-
Triebel spaces with p < ∞. Again we start with an almost trivial result. We need
a further notation. By Lp(R, w) we denote the weighted Lebesgue space equipped
with the norm
k f Lp(R, w)k :=(cid:16)Z ∞
−∞
f (t)p w(t) dt(cid:17)1/p
with usual modification if p = ∞.
Lemma 1 Let d ≥ 2.
(i) Let 0 < p < ∞. Then tr : RLp(Rd) → RLp(R, td−1) is a linear isomorphism
with inverse ext .
(ii) Let p = ∞. Then tr
inverse ext .
: RL∞(Rd) → RL∞(R) is a linear isomorphism with
In particular this means, that whenever the Besov-Lizorkin-Triebel space As
p,q(Rd)
is contained in L1(Rd) + L∞(Rd), then tr is well-defined on its radial subspace. This
is in sharp contrast to the general theory of traces on these spaces. To guarantee
that tr is meaningful on As
p,q(Rd) one has to require
s >
d − 1
p
+ max(cid:16)0,
1
p
− 1(cid:17) ,
cf. e.g. [20], [14], [41, Rem. 2.7.2/4] or [12]. On the other hand we have
Bs
p,q(Rd) , F s
p,q(Rd) ֒→ L1(Rd) + L∞(Rd)
if s > d max(0, 1
p − 1), see, e.g., [35]. Since
d max(0,
1
p
− 1) <
d − 1
p
+ max(cid:16)0,
1
p
− 1(cid:17)
p,q(Rd).
we have the existence of tr with respect to RAs
than for As
Below we shall develop a description of the traces of the radial subspaces of Bs
and F s
appropriate notion of an atom and second, adapted sequence spaces.
p,q(Rd)
p,q(Rd) in terms of atoms. To explain this we need to introduce first an
p,q(Rd) for a wider range of parameters
Definition 1 Let L ≥ 0 be an integer. Let I be a set either of the form I = [−a, a]
or of the form I = [−b, −a] ∪ [a, b] for some 0 < a < b < ∞. An even function
g ∈ C L(R) is called an even L-atom centered at I if
and if either
b(n)(t) ≤ I−n ,
0 ≤ n ≤ L .
max
t∈R
supp g ⊂ [−
3a
2
,
3a
2
]
in case
I = [−a, a] ,
7
or
supp g ⊂ [−
3b − a
2
, −
3a − b
2
] ∪ [
3a − b
2
,
3b − a
2
]
in case
I = [−b, −a] ∪ [a, b] .
Definition 2 Let 0 < p < ∞, 0 < q ≤ ∞ and s ∈ R. Let
j,k(t) :=( 1
0
χ#
2−jk ≤ t ≤ 2−j(k + 1) ,
if
otherwise .
t ∈ R .
Then we define
bs
p,q,d :=(cid:26)s = (sj,k)j,k :
and
k s bs
p,q,dk =(cid:18) ∞Xj=0
2j(s− d
p )q(cid:18) ∞Xk=0
(1+k)d−1 sj,kp(cid:19)q/p(cid:19)1/q
< ∞(cid:27) .
f s
p,q,d
:= (cid:26)s = (sj,k)j,k :
∞Xk=0
p,q,dk = (cid:13)(cid:13)(cid:13)(cid:13)(cid:18) ∞Xj=0
2jsq
k s f s
respectively.
sj,kq χ#
j,k(·)(cid:19)1/q
Lp(R, td−1)(cid:13)(cid:13)(cid:13)(cid:13) < ∞(cid:27) ,
Remark 2 Observe bs
p,p,d = f s
p,p,d in the sense of equivalent quasi-norms.
Adapted to these sequence spaces we define now function spaces on R.
Definition 3 Let 0 < p < ∞, 0 < q ≤ ∞, s > 0 and L ∈ N0.
(i) Then T Bs
exists a decomposition
p,q(R, L, d) is the collection of all functions g : R → C such that there
g(t) =
sj,k gj,k(t)
(2)
∞Xj=0
∞Xk=0
(convergence in Lmax(1,p)(R, td−1)), where the sequence (sj,k)j,k belongs to bs
the functions gj,k are even L-atoms centered at either [−2−j, 2−j] if k = 0 or at
p,q,d and
[−2−j(k + 1), −2−jk] ∪ [2−jk, 2−j(k + 1)]
if k > 0. We put
k g T Bs
p,q(R, L, d)k := infnk (sj,k) bs
p,q,dk :
(2)
holdso .
(ii) Then T F s
exists a decomposition (2), where the sequence (sj,k)j,k belongs to f s
tions gj,k are as in (i). We put
p,q(R, L, d) is the collection of all functions g : R → C such that there
p,q,d and the func-
k g T F s
p,q(R, L, d)k := infnk (sj,k) f s
p,q,dk :
(2)
holdso .
8
We need a few further notation.
spaces quite often the following numbers occur:
In connection with Besov and Lizorkin-Triebel
σp(d) := d max(cid:16)0,
1
p
− 1(cid:17)
and
σp,q(d) := d max(cid:16)0,
1
p
− 1,
1
q
− 1(cid:17) .
(3)
For a real number s we denote by [s] the integer part, i.e. the largest integer m such
that m ≤ s.
Theorem 3 Let d ≥ 2, 0 < p < ∞ and 0 < q ≤ ∞.
(i) Suppose s > σp(d) and L ≥ [s] + 1. Then the mapping tr is a linear isomorphism
of RBs
(ii) Suppose s > σp,q(d) and L ≥ [s]+1. Then the mapping tr is a linear isomorphism
of RF s
p,q(R, L, d) with inverse ext .
p,q(Rd) onto T Bs
p,q(Rd) onto T F s
p,q(R, L, d) with inverse ext .
Remark 3 Let 0 < p ≤ 1 < q ≤ ∞. Then the spaces RBσp
p,q(Rd) contain sin-
gular distributions, see [35]. In particular, the Dirac delta distribution belongs to
d
p −d
p,∞ (Rd), see, e.g., [29, Rem. 2.2.4/3]. Hence, our pointwise defined mapping tr
RB
is not meaningful on those spaces, or, with other words, Theorem 3 does not extend
to values s < σp(d).
Outside the origin radial distributions are more regular. We shall discuss several
examples for this claim.
Theorem 4 Let d ≥ 2, 0 < p < ∞ and 0 < q ≤ ∞. Suppose s > max(0, 1
Let f ∈ RAs
p,q(Rd) s.t. 0 6∈ supp f . Then f is a regular distribution in S ′(Rd).
p − 1).
Remark 4 There is a nice and simple example which explains the sharpness of the
restrictions in Thm. 4. We consider the singular distribution f defined by
ϕ 7→ Zx=1
ϕ(x) dx ,
ϕ ∈ S(Rd) .
By using the wavelet characterization of Besov spaces, it is not difficult to prove
that the spherical mean distribution f belongs to the spaces B
1
p −1
p,∞ (Rd) for all p.
Theorem 5 Let d ≥ 2, 0 < p < ∞ and 0 < q ≤ ∞. Suppose s > max(0, 1
Let f ∈ RAs
p,q(Rd) s.t. 0 6∈ supp f . Then f0 = tr f belongs to As
p,q(R).
p − 1).
Remark 5 As mentioned above
As
p,q(R) ֒→ L1(R) + L∞(R)
if
s > σp(1) = max(cid:16)0,
1
p
− 1(cid:17) ,
which shows again that we deal with regular distributions. However, in Thm. 5
some additional regularity is proved.
9
2.1.3 Traces of radial subspaces of Sobolev spaces
Clearly, one can expect that the description of the traces of radial Sobolev spaces
can be given in more elementary terms. We discuss a few examples without having
the complete theory.
Theorem 6 Let d ≥ 2 and 1 ≤ p < ∞.
(i) The mapping tr is a linear isomorphism (with inverse ext ) of RW 1
the closure of RC ∞
0 (R) with respect to the norm
p (Rd) onto
k g Lp(R, td−1)k + k g′ Lp(R, td−1)k .
(ii) The mapping tr is a linear isomorphism (with inverse ext ) of RW 2
the closure of RC ∞
0 (R) with respect to the norm
p (Rd) onto
k g Lp(R, td−1)k + k g′ Lp(R, td−1)k + k g′/r Lp(R, td−1)k + k g′′ Lp(R, td−1)k .
Remark 6 Both statements have elementary proofs, see (11) for (i). However, the
complete extension to higher order Sobolev spaces is open.
There are several ways to define Sobolev spaces on Rd. For instance, if 1 < p < ∞
we have
f ∈ W 2m
p
(Rd) ⇐⇒ f ∈ Lp(Rd)
and ∆mf ∈ Lp(Rd) .
(4)
Such an equivalence does not extend to p = 1 or p = ∞ if d ≥ 2, see [37, pp. 135/160].
Recall that the Laplace operator ∆ applied to a radial function yields a radial
function. In particular we have
∆f (x) = Drf0(r) := f ′′
0 (r) +
d − 1
r
f ′
0(r) ,
r = x ,
in case that f is radial and tr f = f0. Obviously, if f ∈ RC ∞
0 (Rd), then
k f Lp(Rd)k + k ∆mf Lp(Rd)k
(5)
(6)
= (cid:16) πd/2
Γ(d/2)(cid:17)1/p(cid:16)k f0 Lp(R, td−1)k + k Dm
r f0 Lp(R, td−1)k(cid:17) .
This proves the next characterization.
Theorem 7 Let 1 < p < ∞ and m ∈ N. Then the mapping tr yields a linear
isomorphism (with inverse ext ) of RW 2m
0 (R) with
respect to the norm
(Rd) onto the closure of RC ∞
p
k f0 Lp(R, td−1)k + k Dm
r f0 Lp(R, td−1)k .
Remark 7 By means of Hardy-type inequalities one can simplify the terms
k Dm
r f0 Lp(R, td−1)k to some extent, see Theorem 6(ii) for a comparison. We do
not go into detail.
10
2.1.4 The trace in S ′(R)
Many times applications of traces are connected with boundary value problems. In
such a context the continuity of tr considered as a mapping into S′ is essential.
Again we consider the simple situation of the Lp-spaces first.
Lemma 2 Let d ≥ 2 and let 0 < p < ∞. Then RLp(R, td−1) ⊂ S′(R) if and only
if d < p.
From the known embedding relations of RAs
p,q(Rd) into Lu-spaces one obtains
one half of the proof of the following general result.
Theorem 8 Let d ≥ 2, 0 < p < ∞, and 0 < q ≤ ∞.
(a) Let s > σp(d) and L ≥ [s] + 1. Then the following assertions are equivalent:
(i) The mapping tr maps RBs
(ii) The mapping tr : RBs
(iii) We have T Bs
(iv) We have either s > d( 1
p,q(R, L, d) ֒→ S′(R).
p,q(Rd) into S′(R).
p,q(Rd) → S′(R) is continuous.
p − 1
d) or s = d( 1
p − 1
d ) and q ≤ 1.
(b) Let s > σp,q(d) and L ≥ [s] + 1. Then following assertions are equivalent:
(i) The mapping tr maps RF s
(ii) The mapping tr : RF s
(iii) We have T F s
(iv) We have either s > d( 1
p,q(R, L, d) ֒→ S′(R).
p,q(Rd) into S′(R).
p,q(Rd) → S′(R) is continuous.
p − 1
d) or s = d( 1
p − 1
d ) and 0 < p ≤ 1.
2.1.5 The trace in S ′(R) and weighted function spaces of Besov and
Lizorkin-Triebel type
Weighted function spaces of Besov and Lizorkin-Triebel type, denoted by Bs
p,q(R, w)
and F s
p,q(R, w), respectively, are a well-developed subject in the literature, we refer
to [5, 6, 30]. Fourier analytic definitions as well as characterizations by atoms
are given under various restrictions on the weights, see e.g.
[4, 5, 6, 16, 18, 31].
In this subsection we are interested in these spaces with respect to the weights
wd−1(t) := td−1, t ∈ R, d ≥ 2. Of course, these weights belong to the Muckenhoupt
class A∞, more exactly wd−1 ∈ Ar for any r > d, see [38].
Theorem 9 Let d ≥ 2, 0 < p < ∞, and 0 < q ≤ ∞.
(i) Suppose s > σp(d) and let L ≥ [s] + 1. If T Bs
8), then T Bs
(ii) Suppose s > σp,q(d) and let L ≥ [s] + 1. If T F s
8), then T F s
p,q(R, L, d) = RBs
p,q(R, L, d) = RF s
p,q(R, wd−1) in the sense of equivalent quasi-norms.
p,q(R, L, d) ֒→ S′(R) (see Theorem
p,q(R, wd−1) in the sense of equivalent quasi-norms.
p,q(R, L, d) ֒→ S′(R) (see Theorem
11
✻
s
s = d ( 1
p − 1
d )
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
T As
p,q(R, L, d)
= RAs
p,q(R, wd−1)
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
.....
......
.....
.....
......
.....
.....
......
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . .
.....
.....
......
.....
.....
......
.....
.....
......
.....
. . . . . . . .
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
T As
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
......
.....
.....
....
•
1
d
.....
......
.....
.....
.....
.....
......
.....
.....
.....
.....
......
.....
.....
.....
.....
......
.....
p,q(R, L, d)
6⊂ S ′(R)
•
1
......
.....
.....
.....
.....
.....
.....
....
✲
1
p
•
0
s = d ( 1
p − 1)
.
.....
.....
......
.....
.....
.....
.....
......
.....
.....
.....
.....
......
.....
.....
.....
......
.....
.....
.....
.....
......
.....
.....
.....
Fig. 1: Existence of the trace
Remark 8 We add some statements concerning the regularity of the most promi-
: ϕ → ϕ(0), ϕ ∈ S(Rd). This tempered
nent singular distribution, namely δ
distribution has the following regularity properties:
• First we deal with the situation on Rd. We have δ ∈ RB
d
p −d
p,∞ (Rd) (but δ 6∈
d
p −d
p,q (Rd) for q < ∞ and δ 6∈ RF
d
p −d
p,∞ (Rd)), see, e.g., [29, Rem. 2.2.4/3].
RB
• Now we turn to the situation on R. By using more or less the same arguments
as on Rd one can show δ ∈ B
d
p −1
p,∞ (R, wd−1) (but δ 6∈ B
d
p −1
p,q (R, wd−1) for any
q < ∞ and δ 6∈ F
d
p −1
p,∞ (R, wd−1)).
2.1.6 The regularity of radial functions outside the origin
Let f be a radial function such that supp f ⊂ {x ∈ Rd :
Then the following inequality is obvious:
x ≥ τ } for some τ > 0.
k f0 Lp(R)k ≤ τ −(d−1)/p(cid:16)Γ(d/2)
πd/2 (cid:17)1/p
k f Lp(Rd)k .
An extension to first or second order Sobolev spaces can be done by using Theorem 6.
However, an extension to all spaces under consideration here is less obvious. Partly
it could be done by interpolation, see Proposition 1, but we prefer a different way
(not to exclude p < 1). We shall compare the atomic decompositions in Theorem 3
with the known atomic and wavelet characterizations of Bs
p,q(R) and F s
p,q(R).
Corollary 1 Let τ > 0. Let d ≥ 2, 0 < p < ∞, and 0 < q ≤ ∞.
(i) We suppose s > σp(d). If f ∈ RBs
p,q(Rd) such that
supp f ⊂ {x ∈ Rd :
x ≥ τ }
(7)
12
then its trace f0 belongs to Bs
depending on f and τ ) such that
p,q(R). Furthermore, there exists a constant c (not
k f0 Bs
p,q(R)k ≤ c τ −(d−1)/p k f Bs
p,q(Rd)k
(8)
holds for all such functions f and all τ > 0.
(ii) We suppose s > σp,q(d). If f ∈ RF s
belongs to F s
τ ) such that
p,q(Rd) such that (7) holds, then its trace f0
p,q(R). Furthermore, there exists a constant c (not depending on f and
k f0 F s
p,q(R)k ≤ c τ −(d−1)/p k f F s
p,q(Rd)k
(9)
holds for all such functions f all τ > 0.
We wish to mention that Corollary 1 has a partial inverse.
Corollary 2 Let d ≥ 2, 0 < p < ∞, 0 < q ≤ ∞ and 0 < a < b < ∞.
(i) We suppose s > σp(d). If g ∈ RBs
p,q(R) such that
supp g ⊂ {x ∈ R :
a ≤ x ≤ b}
(10)
then the radial function f := ext g belongs to RBs
constants A, B such that
p,q(Rd) and there exist positive
A k g Bs
p,q(R)k ≤ k f Bs
p,q(Rd)k ≤ B k g Bs
p,q(R)k .
(ii) We suppose s > σp,q(d). If g ∈ RF s
function f := ext g belongs to RF s
that
p,q(R) such that (10) holds, then the radial
p,q(Rd) and there exist positive constants A, B such
A k g F s
p,q(R)k ≤ k f F s
p,q(Rd)k ≤ B k g F s
p,q(R)k .
For our next result we need Holder-Zygmund spaces. Recall, that C s(Rd) =
∞,∞(Rd) in the sense of equivalent norms if s 6∈ N0. Of course, also the spaces
∞,∞(Rd) with s ∈ N allow a characterization by differences. We refer to [41, 2.2.2,
Bs
Bs
2.5.7] and [42, 3.5.3]. We shall use the abbreviation
Z s(Rd) = Bs
∞,∞(Rd) ,
s > 0 .
Taking into account the well-known embedding relations for Besov as well as for
Lizorkin-Triebel spaces, defined on R, Thm. 5 implies in particular:
Corollary 3 Let d ≥ 2, 0 < p < ∞, 0 < q ≤ ∞, and s > 1/p. Let ϕ be a smooth
radial function, uniformly bounded together with all its derivatives, and such that
0 6∈ supp ϕ. If f ∈ RAs
p,q(Rd), then ϕ f ∈ Z s−1/p(Rd).
Remark 9 P.L. Lions [23] has proved the counterpart of Corollary 3 for first order
Sobolev spaces. We also dealt in [32] with these problems.
13
Finally, for later use, we would like to know when the radial functions are con-
tinuous out of the origin.
Corollary 4 Let τ > 0. Let d ≥ 2, 0 < p < ∞, and 0 < q ≤ ∞.
(i) If either s > 1/p or s = 1/p and q ≤ 1 then f ∈ RBs
continuous on the set x ≥ τ .
(ii) If either s > 1/p or s = 1/p and p ≤ 1 then f ∈ RF s
continuous on the set x ≥ τ .
p,q(Rd) is uniformly
p,q(Rd) is uniformly
By looking at the restrictions in Cor. 4 we introduce the following set of param-
eters.
Definition 4 (i) We say (s, p, q) belongs to the set U(B) if (s, p, q) satisfies the
restrictions in part (i) of Cor. 4.
(ii) The triple (s, p, q) belongs to the set U(F ) if (s, p, q) satisfies the restrictions in
part (ii) of Cor. 4.
Remark 10 (a) The abbreviation (s, p, q) ∈ U(A) will be used with the obvious
meaning.
(b) Let 1 ≤ p = p0 < ∞ be fixed. Then there is always a largest space in the set
{Bs
p0,q(Rd) : (s, p0, q) ∈ U(B)} ∪ {F s
p0,q(Rd) : (s, p0, q) ∈ U(F )} .
This space is given either by F 1
p0,1 (Rd) if 1 < p0 < ∞.
If p0 < 1, then obviously B1/p0
p0,∞(Rd) is
the largest Lizorkin-Triebel space in the above family. However, these spaces are
incomparable.
p0,1 (Rd) is the largest Besov space and F 1/p0
1,∞(Rd) if p0 = 1 or by B1/p0
2.2 Decay and boundedness properties of radial functions
We deal with improvements of Strauss' Radial Lemma. Decay can only be expected
if we measure smoothness in function spaces built on Lp(Rd) with p < ∞.
It is instructive to have a short look onto the case of first order Sobolev spaces. Let
f = g(r(x)) ∈ RC ∞
0 (Rd). Then
∂f
∂xi
(x) = g′(r)
xi
r
,
r = x > 0 ,
i = 1, . . . , d .
Hence
∇f (x) Lp(Rd) = cd g′ Lp(R, td−1) ,
(11)
where 1 ≤ p < ∞. Next we apply the identity
g(r) = −Z ∞
r
g′(t)dt
14
and obtain
g(r) ≤Z ∞
r
g′(t)dt ≤ r−(d−1)Z ∞
r
td−1g′(t)dt.
This extends to all functions in RW 1
way we have proved the inequality
xd−1 f (x) = rd−1 g(r) ≤
1 (Rd) by a density argument. On this elementary
1
cd Zx>r
∇f (x) dx ≤
1
cd
k ∇f (x) k1 .
(12)
This inequality can be interpreted in several ways:
• The possible unboundedness in the origin is limited.
• There is some decay, uniformly in f , if x tends to +∞.
• We have limx→∞ xd−1 f (x) = 0 for all f ∈ RW 1
1 (Rd).
• It makes sense to switch to homogeneous function spaces, since in (12) only
the norm of the homogeneous Sobolev space occurs (for this, see [7] and [33]).
We shall show that all these phenomena will occur also in the general context of
radial subspaces of Besov and Lizorkin-Triebel spaces.
2.2.1 The behaviour of radial functions near infinity
Suppose (s, p, q) ∈ U(A). Then f ∈ RAs
p,q(Rd) is uniformly continuous near infinity
and belongs to Lp(Rd). This implies limx→∞ f (x) = 0. However, much more is
true.
Theorem 10 Let d ≥ 2, 0 < p < ∞, and 0 < q ≤ ∞.
(i) Suppose (s, p, q) ∈ U(A). Then there exists a constant c s.t.
x(d−1)/p f (x) ≤ c k f As
p,q(Rd)k
holds for all x ≥ 1 and all f ∈ RAs
(ii) Suppose (s, p, q) ∈ U(A). Then
p,q(Rd).
d−1
p
x
f (x) = 0
lim
x→∞
(13)
(14)
p,q(Rd).
holds for all f ∈ RAs
(iii) Suppose (s, p, q) ∈ U(A). Then there exists a constant c > 0 such that for all x,
p,q(Rd)k = 1,
x > 1, there exists a smooth radial function f ∈ RAs
s.t.
p,q(Rd), k f RAs
d−1
p
x
f (x) ≥ c .
(15)
(iv) Suppose (s, p, q) 6∈ U(A) and 1
the F -case. Then, for all sequences (xj)∞
exists a radial function f ∈ RAs
any neighborhood of xj, j ∈ N.
p > σp(d). We assume also that 1
p > σq(d) in
j=1 ⊂ Rd \ {0} s.t. limj→∞ xj = ∞, there
p,q(Rd)k = 1, s.t. f is unbounded in
p,q(Rd), k f RAs
15
1−d
p , x ∈ Rd, does not belong to Lp(Rd).
Remark 11 (i) Increasing s (for fixed p) is not improving the decay rate. In the
case of Banach spaces, i.e., p, q ≥ 1, the additional assumptions in point (iv) are
always fullfiled. Hence, the largest spaces, guaranteeing the decay rate (d − 1)/p,
are spaces with s = 1/p, see Remark 10. The dependence of the decay on the pa-
rameters s and p is illustrated in Fig. 1 below.
(ii) Observe that in (iii) the function depends on x. There is no function in
RAs
p,q(Rd) such that (15) holds for all x, x ≥ 1, simultaneously. The naive con-
struction f (x) := (1 − ψ(x)) x
(iii) If one switches from inhomogeneous spaces to the larger homogeneous spaces
of Besov and Lizorkin-Triebel type, then the decay rate becomes smaller. It will
depend also on s, see [7] and [33] for details.
(iv) Of course, formula (13) generalizes the estimate (1). Also Coleman, Glazer and
Martin [8] have dealt with (1). P.L. Lions [23] proved a p-version of the Radial
Lemma.
(v) In case A = B the theorem has been proved in [32]. For A = F and 1 < p < ∞
p,∞(Rd), valid for
the result follows from the embedding Bs
all s and all q. The novelity in Theorem. 10 consists in a discussion of the case
RF 1/p
(vi) Originally the Radial Lemma has been used to prove compactness of embed-
dings of radial Sobolev spaces into Lp-spaces, see [8], [23]. In the framework of radial
subspaces of Besov and of Lizorkin-Triebel spaces compactness of embeddings has
been investigated in [32]. There we have given a final answer, i.e., we proved an if,
and only if, assertion.
(vii) Compactness of embeddings of radial subspaces of homogeneous Besov and of
Lizorkin-Triebel spaces will be investigated in [33].
p,∞(Rd), 0 < p ≤ 1, see also Remark 10.
p,q(Rd) ֒→ Bs
p,1(Rd) ֒→ F s
✻
s
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
..
..
.
..
..
.
..
..
.
..
. . . . . . . .
. . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
decay near infinity
............................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
..............................................................................................................................................................................................................................................................................................................................................................................................
1 (Rd)
•
BV (Rd)
no decay
W 1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . .
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
singular
radial distributions
✲
1
p
..
..
..
.
•
1
•
0
.
..
..
..
..
..
..
.
..
..
..
..
..
..
..
..
..
..
.
..
..
..
..
..
..
..
..
..
.
..
..
..
..
..
..
..
..
..
..
.
..
..
..
s = d( 1
p − 1)
s = 1
p
s = 1
p − 1
Fig. 2: Decay at infinity
16
2.2.2 The behaviour of radial functions near infinity -- borderline cases
As indicated in Remark 10, within the scales of Besov and Lizorkin-Triebel spaces
the borderline cases for the decay rate (d − 1)/p are either F 1
1,∞(Rd) if p = 1 or
B1/p
p,1 (Rd) if 1 < p < ∞. Now we turn to spaces which do not belong to these scales
and where the elements of the radial subspaces have such a decay rate. Hence, we
are looking for spaces of radial functions with a simple norm which satisfy (13). The
Sobolev space RW 1
1 (Rd) is such a candidate for which (13) is already known, see
[23]. But this is not the end of the story. Also for the radial functions of bounded
variation such a decay estimate is true.
Theorem 11 Let d ≥ 2. Then there exists constant c s.t.
xd−1 f (x) ≤ c k f BV (Rd)k
holds for all x > 0 and all f ∈ RBV (Rd). Also
lim
x→∞
xd−1 f (x) = 0
is true for all f ∈ RBV (Rd).
(16)
(17)
Remark 12 (i) Both assertions, (16) and (17), require an interpretation since, in
contrast to the classical definition of BV (R), the spaces BV (Rd), d ≥ 2, are spaces
of equivalence classes, see Subsection 4.2. Nevertheless, in every equivalence class
[f ] ∈ BV (Rd), there is a representative f ∈ [f ], such that
f (x) ≤ lim sup
f (y)
y→x
(simply take f (x) := f (x) in every Lebesgue point x of f and f (x) := 0 otherwise).
Hence, (16) and (17) have to be interpreted as follows: whenever we work with
values of the equivalence class [f ] then we mean the function values of the above
representative f .
(ii) Notice that F 1
(iii) Observe, as in case of the Radial Lemma, that (16) holds for x 6= 0.
1,∞(Rd) and BV (Rd) are incomparable.
As a preparation for Theorem 11 we shall characterize the traces of radial el-
ements in BV (Rd). This seems to be of independent interest. For this reason we
are forced to introduce weighted spaces of functions of bounded variation on the
positive half axis. We put R+ := (0, ∞). As usual, ν denotes the total variation of
the measure ν, see, e.g., [28, Chapt. 6].
Definition 5 (i) A function ϕ ∈ C([0, ∞)) belongs to C 1
ously differentiable on R+, has compact support, satisfies ϕ(0) = 0 and lim
t→0+
c ([0, ∞)) if it is continu-
ϕ′(t) =
17
ϕ(t)
exists and is finite.
ϕ′(0) = lim
t→0+
(ii) A function g ∈ L1(R+, td−1) is said to belong to BV (R+, td−1) if there is a signed
Radon measure ν on R+ such that
t
Z ∞
0
and
is finite.
ϕ(t) td−1 dν(t) ,
g(t) [ϕ(s)sd−1]′(t) dt = −Z ∞
k g BV (R+, td−1)k := k g L1(R+, td−1)k + Z ∞
0
0
∀ϕ ∈ C 1
c ([0, ∞))
(18)
rd−1 dν(r)
(19)
By using these new spaces we can prove the following trace theorem. With a
slight abuse of notation ext denotes here the radial extension of a function defined
on [0, ∞) (instead of R).
Theorem 12 Let g be a measurable function on R+. Then ext g ∈ BV (Rd) if, and
only if g ∈ BV (R+, td−1) and
k ext g BV (Rd)k ≍ k g BV (R+, td−1) k .
Spaces with 1 < p < ∞
For 1 < p < ∞ one could use interpolation between p = 1 and p = ∞ to obtain
spaces with the decay rate (d−1)/p. The largest spaces with this respect are obtained
by the real method. Let Mp(Rd) := (RC(Rd), RBV (Rd))Θ,∞, Θ = 1/p. Then (13)
holds for all elements f ∈ Mp(Rd). The disadvantage of these classes Mp(Rd) lies in
the fact that elementary descriptions of Mp(Rd) are not known. However, at least
some embeddings are known. From
RB1/p
p,1 (Rd) = [RB0
∞,1(Rd), RB1
1,1,(Rd)]Θ ֒→ (RC(Rd), RBV (Rd))Θ,∞ ,
Θ = 1/p ,
(combine Proposition 1 with [1, Thm. 4.7.1]), we get back Theorem 10 (i), but only
in case 1 < p < ∞.
2.2.3 The behaviour of radial functions near the origin
At first we mention that the embedding relations with respect to L∞(Rd) do not
change when we switch from As
p,q(Rd) to its radial subspace RAs
p,q(Rd).
Lemma 3 (i) The embedding RBs
d/p or s = d/p and q ≤ 1.
(ii) The embedding RF s
s = d/p and p ≤ 1.
p,q(Rd) ֒→ L∞(Rd) holds if and only if either s >
p,q(Rd) ֒→ L∞(Rd) holds if and only if either s > d/p or
18
The explicit counterexamples will be given in Lemma 8 below. Hence, unbound-
edness can only happen in case s ≤ d/p.
Theorem 13 Let d ≥ 2, 0 < p < ∞ and 0 < q ≤ ∞.
(i) Suppose (s, p, q) ∈ U(A) and s < d
p . Then there exists a constant c s.t.
d
x
p −s f (x) ≤ c k f RAs
p,q(Rd)k
(20)
holds for all 0 < x ≤ 1 and all f ∈ RAs
(ii) Let σp(d) < s < d/p. There exists a constant c > 0 such that for all x, 0 < x <
1, there exists a smooth radial function f ∈ RAs
p,q(Rd).
p,q(Rd)k = 1, s.t.
p,q(Rd), k f RAs
x
d
p −s f (x) ≥ c .
(21)
In the next figure (Fig. 3) we summarize our knowledge.
s = d
p
. . . . . .
. . . . . . . . . . . .
✻
global boundedness
.......................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
no boundedness
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
controlled unboundedness
near the origin
s
•
0
s = 1
p
✲
1
p
Fig. 3: Unboundedness at the origin
Remark 13 (i) In case of RBs
p,∞(Rd) we have a function which realizes the extremal
behaviour for all x < 1 simultaneously. It is well-known, see e.g. [29, Lem. 2.3.1/1],
that the function
f (x) := ψ(x) x
d
p −s ,
x ∈ Rd ,
p,q(Rd), q < ∞. Since it is also not contained in RF s
p,∞(Rd), as long as s > σp(d). This function does not belong to
belongs to RBs
p,q(Rd), 0 < q ≤ ∞ we
RBs
conclude that in these cases there is no function, which realizes this upper bound
for all x simultaneously. In these cases the function f in (21) has to depend on x.
(ii) These estimates do not change by switching to the larger homogeneous spaces
R As
p,2(Rd) this
has been observed in a recent paper by Cho and Ozawa [7], see also Ni [25], Rother
[27] and Kuzin, Pohozaev [22, 8.1]. The general case is treated in [33].
(iii) Estimates as in (20) have been investigated also in [32]. However, there we
p,q(Rd) of Besov and Lizorkin-Triebel type. In case of R H s(Rd) = R F s
19
proved a much weaker result only. We had overlooked the s-dependence of the
power of x.
(iv) In the literature one can find various types of further inequalities for radial
functions. Many times preference is given to a homogeneous context, see the in-
equalities (1) and (16) as examples. Then one has to deal with the behaviour at
infinity and around the origin simultaneously. That would be not appropriate in the
context of inhomogeneous spaces. Inequalities like (1) and (16) will be investigated
systematically in [33]. However, let us refer to [39], [23], [25], [27], [22, 8.1] and [7]
for results in this direction. Sometimes also decay estimates are proved by replacing
on the right-hand side the norm in the space As
p,q(Rd)) by products of
norms, e.g., k f Lp(Rd)k1−Θ k f As
p,q(Rd)kΘ for some Θ ∈ (0, 1), see [23], [25], [27]
and [7]. Here we will not deal with those modifications (improvements).
p,q(Rd) ( As
Finally we have to investigate s ≤ 1/p and (s, p, q) 6∈ U(A).
Lemma 4 Let d ≥ 2, 0 < p < ∞ and 0 < q ≤ ∞. Suppose (s, p, q) 6∈ U(A) and
σp(d) < 1/p. Moreover let σq(d) < 1/p in the F -case. Then there exists a radial
function f ∈ RAs
p,q(Rd)k = 1, and a sequence (xj)j ⊂ Rd \ {0} s.t.
limj→∞ xj = 0 and f is unbounded in a neighborhood of all xj.
p,q(Rd), k f RAs
2.2.4 The behaviour of radial functions near the origin -- borderline cases
Now we turn to the remaining limiting situation. We shall show that there is also
controlled unboundedness near the origin if s = d/p and RAd/p
p,q (Rd) 6⊂ L∞(Rd).
Theorem 14 Let d ≥ 2, 0 < p < ∞, 0 < q ≤ ∞, and suppose s = d/p.
(i) Let 1 < q ≤ ∞. Then there exists constant c s.t.
(− log x)−1/q′
f (x) ≤ c k f Bd/p
p,q (Rd)k
holds for all 0 < x ≤ 1/2 and all f ∈ RBd/p
(ii) Let 1 < p < ∞. Then there exists constant c s.t.
p,q (Rd).
(− log x)−1/p′
f (x) ≤ c k f F d/p
p,q (Rd)k
holds for all 0 < x ≤ 1/2 and all f ∈ RF d/p
p,q (Rd).
(22)
(23)
Remark 14 Comparing Lemma 8 below and Theorem 14 we find the following. For
the case q = ∞ in Theorem 14(i) the function f1,0, see (60), realizes the extremal
behaviour. In all other cases there remains a gap of order log log to some power.
20
3 Traces of radial subspaces -- proofs
The main aim of this section is to prove Theorem 3. It expresses the fact that all
information about a radial function is contained in its trace onto a straight line
through the origin. However, a few things more will be done here. For later use one
subsection is devoted to the study of interpolation of radial subspaces (Subsection
3.1) and another one is devoted to the study of certain test functions (Subsection
3.8).
3.1
Interpolation of radial subspaces
We mention two different results here, one with respect to the complex method and
one with respect to the real method of interpolation.
3.1.1 Complex interpolation of radial subspaces
p,q(Rd) (F s
p,q(Rd)) are complemented subspaces of Bs
p,q(Rd)
In [36] one of the authors has proved that in case p, q ≥ 1 the spaces RBs
p,q(Rd)). By means of the
(RF s
method of retraction and coretraction, see, e.g., Theorem 1.2.4 in [40], this allows to
p,q(Rd)) to its
transfer the interpolation formulas for the original spaces Bs
radial subspaces. However, we prefer to quote a slightly more general result, proved
in [34], concerning the complex method.
It is based on the results on complex
interpolation for Lizorkin-Triebel spaces from [14] and uses the method of [24] for
an extension to the quasi-Banach space case.
p,q(Rd) (F s
Proposition 1 Let 0 < p0, p1 ≤ ∞, 0 < q0, q1 ≤ ∞, s0, s1 ∈ R, and 0 < Θ < 1.
Define s := (1 − Θ) s0 + Θ s1,
1
p
:=
1 − Θ
p0
+
Θ
p1
and
1
q
:=
1 − Θ
q0
+
Θ
q1
.
(i) Let max(p0, q0) < ∞. Then we have
(ii) Let p1 < ∞ and min(q0, q1) < ∞. Then we have
RBs
RF s
p,q(Rd) =hRBs0
p,q(Rd) =hRF s0
p0,q0(Rd), RBs1
p0,q0(Rd), RF s1
p1,q1(Rd)iΘ
p1,q1(Rd)iΘ
.
.
3.1.2 Real Interpolation of radial subspaces
For later use we also formulate a result with respect to the real method of interpo-
lation.
21
Proposition 2 Let d ≥ 1, 1 ≤ q, q0, q1 ≤ ∞, s0, s1 ∈ R, s0 6= s1, and 0 < Θ < 1.
(i) Let 1 ≤ p ≤ ∞. Then, with s := (1 − Θ) s0 + Θ s1, we have
(ii) Let 1 ≤ p < ∞. Then, with s := (1 − Θ) s0 + Θ s1, we have
RBs
RBs
p,q(Rd) =(cid:16)RBs0
p,q(Rd) =(cid:16)RF s0
p,q0(Rd), RBs1
p,q0(Rd), RF s1
p,q1(Rd)(cid:17)Θ,q
p,q1(Rd)(cid:17)Θ,q
.
.
p,q(Rd)) are complemented
Proof. As mentioned above, the spaces RBs
subspaces of Bs
p,q(Rd)), see [36]. Using the method of retraction and core-
traction, see [41, 1.2.4], the above statements are consequences of the corresponding
formulas without R, see e.g. [41, 2.4.2].
p,q(Rd) (RF s
p,q(Rd) (F s
3.2 Proofs of the statements in Subsection 2.1.1
Proof of Theorem 1
Step 1. Let m ∈ N0. If f ∈ RC m(Rd), then, of course, f0 = tr f ∈ RC m(R) and the
inequality
k tr f C m(R)k ≤ k f C m(Rd)k
(24)
follows immediately.
Step 2. Let g ∈ RC m(R) for some m ∈ N.
Substep 2.1. We first deal with the regularity of f := ext g. Here we make use of
the following elementary observation. Let U(x0) denote the neigborhoud of a point
x0 ∈ Rd. If a function v belongs to C m(U(x0) \ {x0}), such that v is continuous in
x0 and the limit
Dαv(x)
lim
x→x0
(25)
exists for all α, α ≤ m, and all j, j = 1, . . . , d, then v ∈ C m(U(x0)). This is a
consequence of the mean-value theorem. By means of this argument it remains to
prove the existence of these limits for Dαf . This turns out to be most simple using
the extra condition
g(0) = g′(0) = · · · = g(m)(0) = 0 .
This clearly implies for 0 ≤ ℓ ≤ m
g(ℓ)(t) = o(tm−ℓ)
if
t → 0 .
(26)
(27)
Furthermore, we shall use, for every n ∈ N0 there is a constant c > 0 such that the
function r = r(x) satisfies
Dβr(x) ≤ c r(x)1−β
(28)
22
for all β ∈ Nd
estimates (28), (27) we find
0, β ≤ n, and all x ∈ Rd \ {0}. Hence, using chain rule and the
(Dαf )(x) <
∼
<
∼
αXℓ=1
αXℓ=1
g(ℓ)(r) Xβ1+...+βℓ=α
Dβ1
r(x) . . . Dβℓ
r(x)
o(rm−ℓ) rℓ−α = o(rm−α),
r ↓ 0 ,
(29)
where α ≤ m. This proves the existence of the limit in (25) and therefore, Dαf ,
α ≤ m, is continuous everywhere.
Next, we wish to remove the restriction (26). Suppose that m is even. Hence
g′(0) = g′′′(0) = . . . = g(m−1)(0) = 0,
but g(0), g′′(0), . . . , g(m)(0) can be arbitrary. Let ψ0 = tr ψ. We introduce the
function
h(t) := g(t) − g1(t) ψ0(t) ,
t ∈ R ,
where
g1(t) := g(0) +
g′′(0)
2!
t2 + · · · +
g(m)(0)
(m)!
tm,
t ∈ R .
The extension of g1 ψ0 is a radial function with compact support and continuous
derivatives of arbitrary order. The function h satisfies (26) and this implies that
ext h has continuous derivatives up to order m in Rd. and therefore, also f has
continuous derivatives up to order m in Rd. The uniform continuity of the derivatives
of f in x ≥ 1 follows immediately from the chain rule, the uniform continuity in
x ≤ 1 is obvious. For odd m the proof is similar.
Substep 2.2. It remains to estimate the norm of f in RC m(Rd). Again we deal with
m being even first. Using the same notation as in Step 2.1 we find on the one hand
the obvious estimate
k ext g1 ψ0 C m(Rd) ≤
g(2j)(0)
(2j)!
k x2j ψ(x) C m(Rd)k
m/2Xj=0
∼ k g C m(R)k ,
<
and on the other hand we can estimate k ext h C m(Rd)k as in (29). This leads to
∼ k h C m(R)k
k ext h C m(Rd)k <
∼ k g C m(R)k + k g1 ψ0 C m(R)k
<
∼ k g C m(R)k .
<
By means of ext g = ext h + ext (g1 ψ0) ∈ RC (m)(Rd) this proves
k ext g C m(Rd)k <
∼ k g C m(R)k .
As above, for odd m the proof is similar.
23
Proof of Theorem 2
∞,q(Rd), Bs
∞,q(Rd), RBs
For tr ∈ L(Bs
∞,q(R)) we refer to [41, 2.7.2]. This immediately gives tr ∈
L(RBs
∞,q(R)). Concerning ext we argue by using real interpolation.
Observe, that ext ∈ L(RC m(R), RC m(Rd)) for all m ∈ N0, see Theorem 1. From
the interpolation property of the real interpolation method we derive
ext ∈ L(cid:16)(RC m(R), RC(R))Θ,q, (RC m(Rd), RC(Rd))Θ,q(cid:17)
for all Θ ∈ (0, 1) and all q ∈ (0, ∞]. Using Proposition 2, simple monotonicity
properties of the real method and
Bn
∞,1(Rd) ֒→ C n(Rd) ֒→ Bn
∞,∞(Rd) ,
n ∈ N0 ,
the claim follows.
3.3 Proofs of the assertions in Subsection 2.1.2
3.3.1 Proof of Lemma 1
Recall, for f ∈ RLp(R) we have
Using
0
πd/2
Γ(d/2) Z ∞
ε↓0Z ∞
ε
f0(r)p rd−1 dr .
f0(r)p rd−1 dr ,
ZRd
Z ∞
0
f (x)p dx = 2
f0(r)p rd−1 dr = lim
which implies the density of the test functions in Lp([0, ∞), rd−1), we can read this
formula also from the other side, it means
ZRd
ext g(x)p dx = 2
πd/2
Γ(d/2) Z ∞
0
g(r)p rd−1 dr
for all g ∈ Lp([0, ∞), rd−1). This proves (i). Part (ii) is obvious.
3.3.2 Characterizations of radial subspaces by atoms
As mentioned above our proof of the trace theorem relies on atomic decompositions
of radial distributions on Rd. We recall our characterizations of RAs
p,q(Rd) from [32],
see also [21].
In this paper we shall consider two different versions of atoms. They are not related
to each other. We hope that it will be always clear from the context with which type
of atoms we are working. For the following definition of an atom we refer to [14] or
[42, 3.2.2]. For an open set Q and r > 0 we put r Q = {x ∈ Rd : dist (x, Q) < r}.
Observe that Q is always a subset of r Q whatever r is.
24
Definition 6 Let s ∈ R and let 0 < p ≤ ∞. Let L and M be integers such that
L ≥ 0 and M ≥ −1. Let Q ⊂ Rd be an open connected set with diam Q = r.
(a) A smooth function a(x) is called an 1L-atom centered in Q if
supp a ⊂
r
2
Dαa(y) ≤ 1 ,
Q ,
α ≤ L .
sup
y∈Rd
(b) A smooth function a(x) is called an (s, p)L,M -atom centered in Q if
supp a ⊂
r
2
Q ,
Dαa(y) ≤ rs−α− d
p ,
α ≤ L ,
a(y) yα dy = 0 ,
α ≤ M .
sup
y∈Rd
ZRd
Remark 15 If M = −1, then the interpretation is that no moment condition is
required.
In [32] and [21] we constructed a regular sequence of coverings with certain special
properties which we now recall. Consider the annuli (balls if k = 0)
Pj,k :=(cid:26)x ∈ Rd :
k 2−j ≤ x < (k + 1) 2−j(cid:27) ,
j = 0, 1, . . . ,
k = 0, 1, . . . .
Then there is a sequence (Ωj)∞
j=0 = ((Ωj,k,ℓ)k,ℓ)∞
j=0 of coverings of Rd such that
(a) all Ωj,k,ℓ are balls with center in xj,k,ℓ s.t. xj,0,1 = 0 and xj,k,ℓ = 2−j(k + 1/2)
if k ≥ 1;
(b) diam Ωj,k,ℓ = 12 · 2−j for all k and all ℓ;
(c) Pj,k ⊂
Ωj,k,ℓ ,
j = 0, 1, . . . ,
k = 0, 1, . . . , where the numbers
C(d, k) satisfy the relations C(d, k) ≤ (2k + 1)d−1, C(d, 0) = 1.
C(d,k)Sℓ=1
∞Pk=0
C(d,k)Pℓ=1
(d) the sums
Xj,k,ℓ(x) are uniformly bounded in x ∈ Rd and j = 0, 1, . . .
(here Xj,k,ℓ denotes the characteristic function of Ωj,k,ℓ);
(e) Ωj,k,ℓ = {x ∈ Rd : 2j x ∈ Ω0,k,ℓ} for all j, k and ℓ;
(f) There exists a natural number K (independent of j and k) such that
{(x1, 0, . . . , 0) :
x1 ∈ R} ∩
diam (Ωj,k,ℓ)
2
Ωj,k,ℓ = ∅
if
ℓ > K (30)
(with an appropriate enumeration).
25
We collect some properties of related atomic decompositions. To do this it is con-
venient to introduce some sequence spaces.
Definition 7 Let 0 < q ≤ ∞.
(i) If 0 < p ≤ ∞, then we define
bp,q,d :=(cid:26)s = (sj,k)j,k :
< ∞(cid:27) .
(ii) By eχj,k we denote the characteristic function of the set Pj,k. If 0 < p < ∞ we
(1 + k)d−1 sj,kp!q/p
∞Xj=0 ∞Xk=0
define
1/q
k s bp,q,dk =
:= (cid:26)s = (sj,k)j,k :
∞Xk=0
sj,kq 2
fp,q,d
k s fp,q,dk = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:18) ∞Xj=0
jdq
p eχj,k(·)(cid:19)1/q
Lp(Rd)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
< ∞(cid:27) .
Remark 16 Observe bp,p,d = fp,p,d in the sense of equivalent quasi-norms.
Atoms have to satisfy moment and regularity conditions. With this respect we
suppose
L ≥ max(0, [s] + 1) ,
M ≥ max([σp(d) − s], −1)
in case of Besov spaces and
L ≥ max(0, [s] + 1) ,
M ≥ max([σp,q(d) − s], −1)
(31)
(32)
in case of Lizorkin-Triebel spaces. Under these restrictions the following assertions
are known to be true:
(i) Each f ∈ RBs
p,q(Rd) ( f ∈ RF s
p,q(Rd)) can be decomposed into
f =
∞Xj=0
∞Xk=0
C(d,k)Xℓ=1
sj,k aj,k,ℓ
( convergence in S ′(Rd) ),
(33)
where the functions aj,k,ℓ are (s, p)L,M -atoms with respect to Ωj,k,ℓ (j ≥ 1),
and the functions a0,k,ℓ are 1L-atoms with respect to Ω0,k,ℓ.
(ii) Any formal series P∞
sj,k aj,k,ℓ converges in S ′(Rd) with limit
p,q(Rd) if the sequence s = (sj,k)j,k belongs to bp,q,d and if the aj,k,ℓ are
in Bs
(s, p)L,M -atoms with respect to Ωj,k,ℓ (j ≥ 1), and the a0,k,ℓ are 1L-atoms with
respect to Ω0,k,ℓ. There exists a universal constant such that
k=0PC(d,k)
j=0P∞
ℓ=1
k
∞Xj=0
∞Xk=0
C(d,k)Xℓ=1
holds for all sequences s = (sj,k)j,k.
sj,k aj,k,ℓ Bs
p,q(Rd)k ≤ c k s bp,q,dk
(34)
26
(iii) There exists a constant c such that for any f ∈ RBs
p,q(Rd) there exists an
atomic decomposition as in (33) satisfying
k (sj,k)j,k bp,q,dk ≤ c k f Bs
p,q(Rd)k .
(35)
(iv) The infimum on the left-hand side in (34) with respect to all admissible rep-
resentations (33) yields an equivalent norm on RBs
p,q(Rd).
j=0P∞
k=0PC(d,k)
(v) Any formal series P∞
sj,k aj,k,ℓ converges in S ′(Rd) with limit
in F s
p,q(Rd) if the sequence s = (sj,k)j,k belongs to fp,q,d and if the functions
aj,k,ℓ are (s, p)L,M -atoms with respect to Ωj,k,ℓ (j ≥ 1), and the functions a0,k,ℓ
are 1L-atoms with respect to Ω0,k,ℓ. There exists a universal constant such
that
ℓ=1
sj,k aj,k,ℓ F s
p,q(Rd)k ≤ c k s fp,q,dk
(36)
k
∞Xj=0
∞Xk=0
C(d,k)Xℓ=1
holds for all sequences s = (sj,k)j,k.
(vi) There exists a constant c such that for any f ∈ RF s
p,q(Rd) there exists an
atomic decomposition as in (33) satisfying
k (sj,k)j,k fp,q,dk ≤ c k f F s
p,q(Rd)k .
(37)
(vii) The infimum on the left-hand side in (36) with respect to all admissible repre-
p,q(Rd). Such decompositions
sentations (33) yields an equivalent norm on RF s
as in (35) and (37) we shall call optimal.
Remark 17 For proofs of all these facts (even with respect to more general decom-
positions of Rd) we refer to [32] and [36]. A different approach to atomic decompo-
sitions of radial subspaces has been given by Epperson and Frazier [10].
3.3.3 Proof of Theorem 3
Step 1. Let f ∈ RBs
p,q(Rd). Then there exists an optimal atomic decomposition, i.e.
∞Xj=0
∞Xk=0
C(d,k)Xℓ=1
k f Bs
p,q(Rd)k ≍ k (sj,k)j,k bp,q,dk ,
f =
sj,k aj,k,ℓ
(38)
see (33) - (35). Since f is even we obtain
f (x) =
∞Xj=0
∞Xk=0
C(d,k)Xℓ=1
sj,k
27
aj,k,ℓ(x) + aj,k,ℓ(−x)
2
.
(39)
We define
gj,k,ℓ(t) := 2j(s−d/p)(cid:16) tr
aj,k,ℓ( · ) + aj,k,ℓ(− · )
2
(cid:17)(t) ,
t ∈ R ,
and dj,k := 2−j(s−d/p) sj,k. Of course, aj,k,ℓ( · ) + aj,k,ℓ(− · ) is not a radial function.
But it is an even and continuous. So, tr means simply the restriction to the x1-axis.
Clearly,
fN (x) :=
NXj=0
NXk=0
C(d,k)Xℓ=1
sj,k
aj,k,ℓ(x) + aj,k,ℓ(−x)
2
,
x ∈ Rd, N ∈ N ,
is an even (not necessarily radial) function in C L(Rd). By means of property (f) of
the particular coverings of Rd, stated in the previous subsection, we obtain
tr fN =
NXj=0
NXk=0
min(C(d,k),K)Xℓ=1
dj,k gj,k,ℓ
(here K is the natural number in (30)). Furthermore
max
0≤n≤L
sup
t∈R
(gj,k,ℓ)(n)(t) ≤ 12s−d/p 2jn .
Obviously
1/q
k (sj,k)j,k bp,q,dk =
∞Xj=0 ∞Xk=0
=
∞Xj=0
2j(s− d
(1 + k)d−1 sj,kp!q/p
p )q ∞Xk=0
(1 + k)d−1 dj,kp!q/p
1/q
.
This implies
k tr fN T Bs
p,q(R, L, d)k ≤ K c k f Bs
p,q(Rd)k
where c and K are independent of f and N.
Next we comment on the convergence of the sequences (fN )N and (tr fN )N . Of
course, fN converges in S ′(Rd) to f . For the investigation of the convergence of
(tr fN )N we choose s′ such that s > s′ > σp(d) and conclude with N > M
k tr fN
− tr fM T Bs′
p,p(R, L, d)k <
dj,k gj,k,ℓ(cid:12)(cid:12)(cid:12)T Bs′
p,p(R)(cid:13)(cid:13)(cid:13)
∼ (cid:13)(cid:13)(cid:13)
min(C(d,k),K)Xℓ=1
NXk=0
NXj=M +1
dj,k gj,k,ℓ(cid:12)(cid:12)(cid:12)T Bs′
p,p(R)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)
min(C(d,k),K)Xℓ=1
MXj=0
NXk=M +1
(1 + k)d−12j(s′−s) sj,kp(cid:17)1/p
∞Xk=0
(1 + k)d−12j(s′−s) sj,kp(cid:17)1/p
(cid:16) MXj=0
∞Xk=M +1
,
+
<
∼ (cid:16) ∞Xj=M +1
+
28
by taking into account the different normalization of the atoms in RBs′
p,p(Rd) and
in RBs
p,q(Rd), respectively. The right-hand side in the previous inequality tends
to zero if M tends to infinity since k (sj,k)j,k bp,q,dk < ∞. Lemma 1 in combi-
nation with Bs
p,1(Rd) →
Lmax(1,p)(R, td−1) as well as the existence of tr f ∈ Lmax(1,p)(R, td−1). Consequently
p,q(Rd) ⊂ Lmax(1,p)(Rd) implies the continuity of tr : RBs
lim
N→∞
tr fN = tr ( lim
N→∞
fN ) = tr f
with convergence in Lmax(1,p)(R, td−1). This proves that tr maps RBs
T Bs
Step 2. The same type of arguments proves that tr maps RF s
in particular the convergence analysis is the same. Furthermore, observe
p,q(R, L, d) if L satisfies (31). Observe, that M can be chosen −1 in (31).
p,q(Rd) into T F s
p,q(Rd) into
p,q(R),
(cid:13)(cid:13)(cid:13)(cid:13)(cid:18) ∞Xj=0
sj,kq χ#
2jsq
∞Xk=0
= cd(cid:13)(cid:13)(cid:13)(cid:13)(cid:18) ∞Xj=0
j,k(·)(cid:19)1/q
∞Xk=0
Lp(R, td−1)(cid:13)(cid:13)(cid:13)(cid:13)
sj,kqeχj,k(·)(cid:19)1/q
2jsq
Lp(Rd)(cid:13)(cid:13)(cid:13)(cid:13) .
This proves that tr maps RF s
use M = −1).
Step 3. Properties of ext . Let g be an even function with a decomposition as in (2)
and
p,q(R, L, d) if L satisfies (32) (again we
p,q(Rd) into T F s
k g T Bs
p,q(R, L, d)k ≍ k (sj,k) bs
p,q,dk .
We define
aj,k(x) := gj,k(x) ,
x ∈ Rd .
The functions aj,k are compactly supported, continuous, and radial. Obviously
supp aj,k ⊂ {x :
2−jk − 2−j−1 ≤ x ≤ 2−j(k + 1) + 2−j−1} ,
k ∈ N ,
and
supp aj,0 ⊂ {x :
x ≤ 3 · 2−j−1} .
From Theorem 1 we derive
Dαaj,k(x) ≤ k aj,k C α(Rd)k <
∼ k gj,k C α(R)k <
∼ 2jα ,
(40)
if α ≤ L. Here the constants behind <
continue with an investigation of the sequence
∼ do not depend on j, k and gj,k. We
hN (x) :=
NXj=0
∞Xk=0
sj,k aj,k(x) ,
x ∈ Rd , N ∈ N .
(41)
29
Related to our decomposition (Ωj,k,ℓ)j,k,ℓ of Rd, see Subsection 3.3.2, there is a se-
quence of decompositions of unity (ψj,k,ℓ)j,k,ℓ, i.e.
∞Xk=0
C(d,k)Xℓ=1
ψj,k,ℓ(x) = 1
for all x ∈ Rd ,
j = 0, 1, . . . ,
supp ψj,k,ℓ ⊂ Ωj,k,ℓ ,
Dαψj,k,ℓ ≤ CL 2jα
α ≤ L ,
(42)
(43)
(44)
see [32]. Hence
hN (x) =
=
=
where
ψj,m,ℓ(x)
sj,k aj,k(x)
C(d,m)Xℓ=1
∞Xm=0
NXj=0
∞Xk=0
sj,m+k aj,m+k(x)ψj,m,ℓ(x)
C(d,m)Xℓ=1
7Xk=−7
∞Xm=0
NXj=0
C(d,k+m)Xℓ=1
NXj=0
∞Xm=0
ej,m,ℓ(x) .
λj,m
λj,m := 2j(s− d
p ) max
−1≤k≤1
sj,m+k
ej,m,ℓ(x)
:= 2−j(s− d
p ) tj,m
sj,m+k aj,m+k(x) ψj,m,ℓ(x)
7Xk=−7
tj,m :=
1
if
max
k=−7, ... ,7
sj,m+k = 0,
( max
k=−7, ... ,7
sj,m+k)−1
otherwise .
We claim that the functions ej,m,ℓ are (s, p)L,−1-atoms (1L-atoms if j = 0) on Rd
related to the covering (Ωj,k,ℓ)j,k,ℓ (up to a universal constant). But this follows
immediately from (43), (44), and (40). Finally we show that the sequence λ =
(λj,m)j,m belongs to bp,q,d. The estimate
k λ bp,q,dk = (cid:16) NXj=0(cid:16) ∞Xm=0
<
∼ (cid:16) NXj=0
2j(s− d
(1 + m)d−1 λj,mp(cid:17)q/p(cid:17)1/q
p )q(cid:16) ∞Xm=0
(1 + m)d−1 sj,mp(cid:17)q/p(cid:17)1/q
is obvious. Hence ext maps T Bs
pair (L, −1) satisfies (31).
Step 4. The proof of the F-case is similar. Here we need that the pair (L, −1)
satisfies (32). The proof is complete.
p,q(Rd). Here we need that the
p,q(R, L, d) into RBs
30
3.3.4 Proof of Theorem 4
p,q(Rd) ֒→ RBs
p,q(Rd) ⊂ Lp(Rd). Hence f is a regular distribution.
Since RF s
p,∞(Rd) it will be enough to deal with radial Besov spaces.
Step 1. Let 1 ≤ p < ∞. Then σp(d) = σp(1) = 0. From s > 0 we derive
Bs
Step 2. Let 0 < p < 1. Since 0 6∈ supp f there exists some ε > 0 s.t. the ball with
radius ε and centre in the origin has an empty intersection with supp f . Let λ > 0.
Since f is a regular distribution if, and only if f (λ · ) is a regular distribution we
may assume ε = 2. Let ϕ ∈ RC ∞(Rd) be a function s.t. ϕ(x) = 1 if x ≥ 2 and
ϕ(x) = 0 if x ≤ 1. Again we shall work with an optimal atomic decomposition of
f ∈ RBs
p,q(Rd), see (38). Obviously
f = f ϕ =
∞Xj=0
∞Xk=0
C(d,k)Xℓ=1
sj,k (ϕ aj,k,ℓ) .
By checking the various support conditions we obtain
f =
3Xj=0
sj,0 (ϕ aj,0,1) +
∞Xj=0
∞Xk=max(1,2j −9)
C(d,k)Xℓ=1
sj,k (ϕ aj,k,ℓ) .
We consider the following splitting
f1(x)
:=
f2(x)
:=
3Xj=0
∞Xj=0
∞Xk=max(1,2j −9)
C(d,k)Xℓ=1
aj,0,1(x) + aj,0,1(−x)
sj,0 ϕ(x)
2
sj,k ϕ(x)
aj,k,ℓ(x) + aj,k,ℓ(−x)
2
.
Concerning the first part f1 we observe that tr f1 is a compactly supported even C L
function. Now we concentrate on f2. Let
sj,k ϕ(x)
aj,k,ℓ(x) + aj,k,ℓ(−x)
2
,
N ∈ N .
f2,N (x) :=
We put
2NXk=max(1,2j −9)
C(d,k)Xℓ=1
NXj=0
gj,k,ℓ(t) := tr(cid:16)ϕ( · )
aj,k,ℓ( · ) + aj,k,ℓ(− · )
2
(cid:17)(t) ,
t ∈ R ,
by using the same convention concerning tr as in Step 1 of the proof of Thm. 3.
Since
g(n)
j,k,ℓ(t) ≤ cϕ (12 · 2−j)s−n−d/p ,
0 ≤ n ≤ L ,
sup
t∈R
31
and supp gj,k,ℓ <
∼ 2−j we obtain for a natural number m
Z m+1
m
tr f2,N (t) dt <
∼
<
∼
NXj=0
2−j 2−j(s−d/p)
sj,kZ m+1
m
NXj=0
min(C(d,k),K)Xℓ=1
2NXk=max(1,2j −9)
2j (m+1)+6Xk=2jm−9
2−j 2−j(s−d/p) 2−j(d−1)/p(cid:16) 2j(m+1)+6Xk=2j m−9
sj,k
NXj=0
∼ m−(d−1)/p
<
Hence
gj,k,ℓ(t) dt
(1 + k)d−1sj,kp(cid:17)1/p
.
k tr fN L1(R)k = 2Z ∞
1
∼ k f Bs
<
p,q(Rd)k
tr fN (t) dt <
∼ k (sj,k)j,k bp,∞,dk
m−(d−1)/p
∞Xm=1
since s > 0 and 0 < p < 1. Let M ≤ N. Then the same type of argument yields
Z m+1
m
tr f2,N (t) − tr f2,M (t) dt
sj,kZ m+1
m
gj,k,ℓ(t) dt
sj,kZ m+1
m
gj,k,ℓ(t) dt
<
∼
min(C(d,k),K)Xℓ=1
2NXk=max(1,2j −9)
NXj=M +1
2NXk=max(2M ,2jm−9)
NXj=0
+
∼ 2−M s
<
sup
j=M +1,...
2j d−1
p
min(C(d,k),K)Xℓ=1
2j (m+1)+6Xk=2jm−9
2j (m+1)+6Xk=max(2M ,2jm−9)
sj,k
+
NXj=0
2−j 2−j(s−d/p)
sj,k
<
∼ m−(d−1)/p(cid:16)2−M s k (sj,k)j,k bp,∞,dk
2j (m+1)+6Xk=max(2M ,2jm−9)
+
(1 + k)d−1sj,kp(cid:17)1/p(cid:17) .
sup
j=0,1,...(cid:16)
2j (m+1)+6Xk=max(2M ,2jm−9)
(1 + k)d−1sj,kp(cid:17)1/p
= 0
Since
lim
M →∞
sup
j=0,1,...(cid:16)
for all m ∈ N we conclude
k tr f2,N − tr f2,M L1(R)k −→ 0
if M → ∞ .
32
Hence
∞Xj=0
∞Xk=max(1,2j −9)
min(C(d,k),K)Xℓ=1
sj,k gj,k,ℓ ∈ L1(R) .
Let θ ∈ Rd, θ = 1. We denote by Tr θ the restriction of a continuous function to
the line Θ := {t θ :
t ∈ R}. Now we repeat, what we have done with respect to
the x1-axis, for such a line. As the outcome we obtain
Tr θ(cid:16) 3Xj=0
sj,0 (ϕ aj,0,1) +
NXj=0
2NXk=max(1,2j −9)
C(d,k)Xℓ=1
sj,k (ϕ aj,k,ℓ)(cid:17) ,
N ∈ N ,
is a Cauchy sequence in L1(Θ) and the limit satisfies
(cid:13)(cid:13)(cid:13) Tr θ(cid:16) 3Xj=0
sj,0 (ϕ aj,0,1) +
∞Xj=0
∞Xk=max(1,2j −9)
∼ k f Bs
<
p,q(Rd)k
C(d,k)Xℓ=1
sj,k (ϕ aj,k,ℓ)(cid:17)(cid:12)(cid:12)(cid:12)L1(Θ)(cid:13)(cid:13)(cid:13)
with a constant independent of θ (of course, here, by a slight abuse of notation,
Tr θ denotes the continuous extension of the previously defined mapping). Using
spherical coordinates this yields
ZRd
f (x) dx = Zθ=1Z ∞
∼ k f Bs
<
0
p,q(Rd)k .
f (t θ) dt dθ
But this means f is a regular distribution.
Remark 18 We have proved a bit more than stated. Under the given restrictions
the pointwise trace tr f of a distribution f ∈ RBs
p,q(Rd), 0 6∈ supp f , makes sense
and belongs to L1(R).
3.3.5 Proof of Remark 4
1
p −1
p,∞ (Rd), see, e.g., [44,
We shall argue by using the wavelet characterization of B
Thm. 1.20]. Let φ denote an appropriate univariate scaling function and Ψ an
associated Daubechies wavelet of sufficiently high order. The tensor product ansatz
yields (d − 1) generators Ψ1, . . . , Ψ2d−1 for the wavelet basis in L2(Rd). Let Φ
denote the d-fold tensor product of the univariate scaling function. We shall use the
abbreviations
Φk(x) := Φ(x − k) ,
k ∈ Zd ,
and
Ψi,j,k(x) := 2jd/2 Ψi(2jx − k) ,
k ∈ Zd ,
j ∈ N0 ,
i = 1, . . . , 2d − 1 .
33
An equivalent norm in B
1
p −1
p,∞ (Rd) is given by
k f B
1
p −1
p,∞ (Rd)k =(cid:16)Xk∈Zd
hf, Φkip(cid:17)1/p
+ sup
j=0,1,...
2j( 1
p −1+d( 1
2 − 1
p ))(cid:16) 2d−1Xi=1 Xk∈Zd
hf, Ψi,j,kip(cid:17)1/p
.
Daubechies wavelets have compact support. This implies
supp Ψi,j,k ⊂ C {x ∈ Rd : 2−j(kℓ − 1) ≤ xℓ ≤ 2−j(kℓ + 1) , ℓ = 1, . . . d}
and
supp Φk ⊂ C {x ∈ Rd : (kℓ − 1) ≤ xℓ ≤ (kℓ + 1) , ℓ = 1, . . . d}
for an appropriate C > 1. By employing these relations we conclude that for fixed
j the cardinality of the set of those functions Ψi,j,k, which do not vanish identically
on x = 1 is <
∼ 2j(d−1). There is the general estimate
by using the information on the size of the support. Inserting this we find
2jd/2 Ψi(2jx − k) dx(cid:12)(cid:12)(cid:12) <
∼ 2jd/2 2−j(d−1) ,
hf, Ψi,j,ki =(cid:12)(cid:12)(cid:12)Zx=1
p ))(cid:16) 2d−1Xi=1 Xk∈Zd
hf, Ψi,j,kip(cid:17)1/p
2j( 1
p −1+d( 1
2 − 1
∼ 2j( 1
<
p −1+d( 1
2 − 1
p )) 2j(d−1)/p 2jd/2 2−j(d−1)
<
∼ 1 .
This proves the claim.
3.3.6 Proof of Theorem 5
p,q(Rd), 0 6∈ supp f , the trace tr f
From Thm. 4 we already know that for f ∈ RAs
makes sense and that tr f ∈ L1(R).
p,q(Rd). Since 0 6∈ supp f there exists some ε > 0 s.t. the
Step 1. Let f ∈ RBs
ball with radius ε and centre in the origin has an empty intersection with supp f .
Without loss of generality we assume ε < 1. Let ϕ ∈ RC ∞(Rd) be a function s.t.
ϕ(x) = 1 if x ≥ ε and ϕ(x) = 0 if x ≤ ε/2. Again we shall work with an optimal
atomic decomposition of f , see (38). It follows
f =
mXj=0
sj,0 (ϕ aj,0,1) +
∞Xj=0
∞Xk=kj
C(d,k)Xℓ=1
sj,k (ϕ aj,k,ℓ)
where
m := 1 + [log2(18 ε−1)]
and
kj := max(1, [2j−1 ε] − 10) .
34
As in the previous proof we introduce the splitting f = f1 + f2, where
f1(x) :=
sj,0 ϕ(x)
mXj=0
aj,0,1(x) + aj,0,1(−x)
2
.
Obviously, tr f1 is a compactly supported even C L function. Let
f2,N (x) :=
sj,k ϕ(x)
aj,k,ℓ(x) + aj,k,ℓ(−x)
2
,
N ∈ N .
As above we use the notation
NXj=0
2NXk=kj
C(d,k)Xℓ=1
gj,k,ℓ(t) := tr(cid:16)ϕ( · )
NXj=0
tr f2,N (t) =
aj,k,ℓ( · ) + aj,k,ℓ(− · )
2
(cid:17)(t) ,
t ∈ R .
2NXk=kj
min(C(d,k),K)Xℓ=1
sj,k gj,k,ℓ(t)
Hence
Since
sup
t∈R
g(n)
j,k,ℓ(t) ≤ cϕ (12 · 2−j)s−n−d/p = cϕ (12 · 2−j)−(d−1)/p (12 · 2−j)s−n−1/p
the functions 2−j(d−1)/p12(d−1)/p gj,k,ℓ/cϕ are (s, p)L,−1-atoms in the sense of Sub-
section 3.3.2 (in the one-dimensional context). Applying property (ii) from this
subsection we find
k tr f2,N Bs
p,q(R)k <
∼ (cid:16) NXj=0
∼ (cid:16) NXj=0 (cid:16) 2NXk=kj
2j(d−1)q/p(cid:16) 2NXk=kj
sj,kp(cid:17)q/p(cid:17)1/q
(1 + k)d−1sj,kp(cid:17)q/p(cid:17)1/q
<
∼ k f RBs
<
p,q(Rd)k .
Now we consider convergence of the sequence tr fN . Let σp(1) < s′ < s. Arguing
as before (but taking into account the different normalization of the atoms with
respect to Bs′
p,p(R)) we find
k tr f2,N − tr f2,M L1(R)k ≤ k tr f2,N − tr f2,M Bs′
p,p(R)k
<
∼ (cid:16) NXj=M +1
+(cid:16) MXj=0
2NXk=kj
2NXk=max(2M ,kj)
(1 + k)d−12j(s′−s)sj,kp(cid:17)1/p
(1 + k)d−12j(s′−s)sj,kp(cid:17)1/p
Since k(sj,k)j,k bp,q,dk < ∞ it follows that the right-hand side tends to 0 if M → ∞.
The uniform boundedness of (tr f2,N )N in Bs
p,q(R) in combination with the weak
35
∂2f
∂xi∂xj
0 (r) · x2
f ′′
(x) =
(x)(cid:19)2
dXi=1(cid:18) ∂2f
dXi,j=1(cid:18) ∂2f
(x)(cid:19)2
∂xi∂xj
∂xi∂xj
=
We fix j ∈ {1, 2, . . . , d} and sum up
Now we sum up with respect to j and find
0 (r) · xi·xj
f ′′
r2 − f ′
0(r) xi·xj
r3 ,
if
i 6= j ,
i
r2 − f ′
0(r) · r2−x2
i
r3
,
if
i = j.
0 (r)2
f ′′
r2
x2
j +
0(r)2
f ′
r4
· (r2 − x2
j ) .
= f ′′
0 (r)2 +
d − 1
r2
· f ′
0(r)2.
convergence of this sequence yields limN→∞ tr f2,N ∈ Bs
so-called Fatou property, see [3, 13]. Hence, tr f2 ∈ Bs
our knowledge about f1 the claim in case of Besov spaces follows.
Step 2. Let f ∈ RF s
the spaces F s
p,q(R) we refer to [13].
p,q(Rd). One can argue as in Step 1. For the Fatou property of
p,q(R) by means of the
p,q(R). In combination with
3.4 Proofs of the statements in Subsection 2.1.3
Proof of Theorem 6
0 (Rd) in RW 1
Step 1. The proof of Theorem 6(i) follows from formula (11) and the density of
RC ∞
Step 2. Let f ∈ RC ∞
Observe, that
0 (Rd). This is equivalent to tr f = f0 ∈ RC ∞
0 (R), see Thm. 1.
p (Rd).
Since the terms on the right-hand side are nonnegative this proves the claim for
smooth f . As above the density argument completes the proof.
Proof of Theorem 7
The formulas (4)-(6) have to be combined with the density of RC ∞
RW 2m
(Rd).
0 (Rd) in
p
3.5 Proof of the statements in Subsection 2.1.4
3.5.1 Proof of Lemma 2
Step 1. Necessity of p > d. Let ϕ ∈ C ∞
0 (R) be an even function s.t. ϕ(0) 6= 0 and
supp ϕ ⊂ [−1/2, 1/2]. Since d ≥ 2 the function g1(t) := ϕ(t) t−1, t ∈ R, belongs
36
to RLp(R, td−1) if p < d. Hence RLp(R, td−1) 6⊂ S′(R) if p < d. Let p = d and
take g2(t) := ϕ(t) t−1 (− log t)−α, t ∈ R \ {0}, for α > 0. In case α d > 1 we have
g2 ∈ RLd(R, td−1). However, if α < 1 then g2 6∈ S′(R). With 1/d < α < 1 the
claim follows.
Step 2. Sufficiency of p > d. Using Holder's inequality we find
Z 1
−1
g(t) dt ≤(cid:16)Z 1
−1
g(t)p td−1 dt(cid:17)1/p(cid:16)Z 1
−1
t− (d−1)p′
p
dt(cid:17)1/p′
.
The second factor on the right-hand side is finite if, and only if,
(d − 1)(p′ − 1) < 1
⇐⇒
d < p .
Complemented by the obvious inequality
Zt>1
g(t)p dt ≤Zt>1
g(t)p td−1dt
we conclude Lp(R, td−1) ֒→ L1(R) + Lp(R) ⊂ S′(R).
3.5.2 Proof of Theorem 8
Thm. 3 implies the equivalence of (ii) and (iii). Also (i) and (ii) are obviously
equivalent.
Step 1. We shall prove that (iv) implies (i) and (ii).
It will be enough to deal with the limiting case. Let
Substep 1.1. The B-case.
s = d( 1
d) > 0 (s > σp(d)) and q = 1. In addition we assume 1 ≤ p < d, where
the upper bound results from the previous restriction on s, see Fig. 1 in Subsection
2.1.5. For f ∈ RBs
p,q(Rd) we select an optimal atomic decomposition of the trace in
the sense of Theorem 3. Let ϕ ∈ S(R). Then
p − 1
(cid:12)(cid:12)(cid:12)Z ∞
−∞
≤ 4
∞Xj=0
≤ 4k ϕ L∞(R)k
sj,k bj,k(t) ϕ(t) dt(cid:12)(cid:12)(cid:12)
∞Xk=0
∞Xj=0
∞Xk=0
∞Xj=0
≤ 4k ϕ L∞(R)k(cid:16) ∞Xk=0
∞Xj=0
×
37
sj,k 2−j k bj,k L∞(R)k k ϕ L∞(R)k
2j(s−d/p)
sj,k
∞Xk=0
(1 + k)−(d−1) p′
2j(s−d/p)(cid:16) ∞Xk=0
p(cid:17)1/p′
(1 + k)d−1sj,kp(cid:17)1/p
.
Since
< ∞
p(cid:17)1/p′
(1 + k)−(d−1) p′
(cid:16) ∞Xk=0
sj,k bj,k(t) ϕ(t) dt(cid:12)(cid:12)(cid:12) ≤ c1 k ϕ L∞(R)k k (sj,k)j,k bs
≤ c2 k ϕ L∞(R)k k f Bs
p,1,dk
p,1(Rd)k ,
if 1 ≤ p < d, we obtain
(cid:12)(cid:12)(cid:12)Z ∞
−∞
∞Xj=0
∞Xk=0
see Theorem 3. This proves sufficiency for 1 ≤ p < d and s = d( 1
0 < p < 1. Then it is enough to apply the continuous embedding
p − 1
d ). Now, let
d
p −1
p,1 (Rd) ֒→ Bd−1
1,1 (Rd) ,
B
see e.g. [41, 2.7.1] or [35].
Substep 1.2. Now we turn to the same implication in case of the F -spaces. Also here
an embedding argument turns out to be sufficient. For 0 < p ≤ 1 and p < p1 < ∞
we have
d
p −1
p,∞ (Rd) ֒→ B
F
p1,1 (Rd) ,
d
p1
−1
see [19] or [35]. Now the claim follows from Substep 1.1.
Step 2. Since tr is an isomorphism of RAs
p,q(Rd) onto T As
Step 1 the implication (iv) =⇒ (iii).
Step 3. It remains to prove the implication (i) =⇒ (vi). We argue by contradiction.
Substep 3.1. The B-case. Let s = d
p − 1 and suppose q > 1. Oriented at our
investigations in Lemma 5 we will use as test functions
p,q(R, L, d) we deduce from
fα(x) := ϕ(x) x−1(− log x)−α ,
x ∈ Rd .
(45)
It is known, see e.g. [29, Lem. 2.3.1], that
fα ∈ B
d
p −1
p,q (Rd)
if, and only if,
q α > 1 .
Since tr fα 6∈ S′(R) if α < 1, we obtain that tr does not map into S′(R) as long as
1/q < α < 1.
Substep 3.2. The F -case. This time it holds
fα ∈ F
d
p −1
p,∞ (Rd)
if, and only if,
p α > 1 ,
see [29, Lem. 2.3.1]. Choosing 1/p < α < 1 we obtain that tr does not map
d
p −1
p,∞ (Rd) into S′(R).
F
38
3.6 Proof of the assertions in Subsection 2.1.5
Proof of Theorem 9
Comparing our atomic decomposition with that one for weighted spaces obtained
in [16] it is essentially a question of renormalization of the atoms. This is enough to
prove T As
p,q(R, wd−1). To see the converse one has to start with
the fact that f ∈ RAs
p,q(R, wd−1) is even. This allows to decompose f into sum of
atoms that are even as well, see (38) and (39) for this argument.
p,q(R, L, d) ֒→ RAs
Proof of Remark 8
The regularity of the δ distribution is calculated at several places, see e.g.
[29,
Remark 2.2.4/3]. The argument, used in this reference, comes from Fourier analysis
and transfers to the weighted case. For the Fourier analytic characterization of
As
p,q(R, wd−1) we refer to [5, 6] and [16].
3.7 Proof of the assertions in Subsection 2.1.6
Proof of Corollary 1
We shall only prove part (i). The proof for the Lizorkin-Triebel spaces is similar.
By our trace theorem we have
k f0 T Bs
p,q(R, L, d)k <
∼ k f Bs
p,q(Rd)k
if L > [s] + 1, cf. Theorem 3. Thus, it is sufficient to prove that
k f0 Bs
p,q(R)k <
∼ τ −(d−1)/p k f0 T Bs
p,q(R, L, d)k .
The trace f0 ∈ T Bs
p,q(R, L, d) can be represented in the form
f0(t) =
∞Xj=0
∞Xk=0
sj,k gj,k(t)
(46)
(47)
(convergence in Lmax(1,p)(R, td−1)), where the sequence (sj,k)j,k belongs to bs
p,q,d, cf.
(2). Let ϕ ∈ C ∞(R) be an even function such that ϕ(t) = 0 if t ≤ 1
2 and ϕ(t) = 1
if t ≥ 1. For any τ > 0 we define ϕτ (t) = ϕ(τ −1t). We will consider two cases:
τ ≥ 2 and 0 < τ < 2.
Case 1. Let τ ≥ 2. Under this assumption any function ϕτ gj,k is an even L-atom
centered at the same interval as gj,k itself (up to a general constant depending on
ϕ), see Definition 1. For any j ∈ N0 we define a nonnegative integer kj by
kj := max{k ∈ N0 :
2−j (k + 1) ≤ τ /2} .
39
Hence, ϕτ gj,k = 0 if k < kj. Furthermore, the functions 2−j(s−1/p) ϕτ gj,k, k ≥ 1,
restricted either to the positive or negative half axis, are (s, p)L,−1-atoms in the
sense of Definition 6 up to a universal constant c. The functions 2−j(s−1/p) ϕτ gj,0 are
(s, p)L,−1-atoms as well (again up to a universal constant). We obtain
f0(t) = ϕτ (t) f0(t) =
∞Xj=0
∞Xk=kj
sj,k ϕτ (t) gj,k(t)
and applying (34) (which is also valid for d = 1) we arrive at the estimate
k f0 Bs
p,q(R)k <
2j(s− 1
∼ (cid:18) ∞Xj=0
p (cid:18) ∞Xj=0
<
∼ τ
1−d
sj,kp(cid:19)q/p(cid:19)1/q
p )q(cid:18) ∞Xk=kj
p )q(cid:18) ∞Xk=0
(1 + k)d−1 sj,kp(cid:19)q/p(cid:19)1/q
2j(s− d
=
τ
1−d
p k s bs
p,q,dk
(48)
since kj ∼ 2j τ . Taking the infimum with respect to all atomic representations of f0
we have proved (46).
Case 2. Let 0 < τ < 2.
Step 1. We assume s < d/p. Then we define j0 ∈ N0 via the relation 2−j0 ≤ τ <
2−j0+1. Further, we put Kj := max(1, 2j−j0−1 − 1). Now we decompose f0 into four
sums
sj,0 ϕτ (t) gj,0(t) +
sj,k ϕτ (t) gj,k(t)
f0(t) = ϕτ (t) f0(t) =
+
j0+1Xj=0
∞Xj=j0
∞Xk=2j−j0+1+1
∞Xj=j0
2j−j0 +1Xk=Kj
j0−1Xj=0
∞Xk=1
sj,k gj,k(t) +
sj,k gj,k(t)
= f1(t) + . . . + f4(t) ,
with f4 = 0 if j0 = 0. Observe
supp fi ⊂ {t : t ≥ τ } ,
i = 3, 4 ,
whereas the supports of the functions ϕτ gj,k, occuring in the defintions of f1 and
f2, may have nontrivial intersections with the interval (τ /2, τ ). The function f1
belongs to C L and has compact support. The functions f2, f3, and f4 are supported
on {t : t ≥ τ /2}. Thus, the known convergence in Lmax(1,p)(R, td−1) implies the
convergence in S ′(R). As in Case 1 the functions 2−j(s−1/p) gj,k, k ≥ 1, restricted
either to the positive or negative half axis, are (s, p)L,−1-atoms in the sense of Defi-
nition 6. An easy calculation shows that also the functions 2−j(s−1/p) ϕτ gj,k, j ≥ j0,
40
are (s, p)L,−1-atoms (up to a universal constant). Hence we may employ (34) and
obtain
k f2 + f3 Bs
p,q(R)k <
2j(s− 1
∼ (cid:18) ∞Xj=j0
p (cid:18) ∞Xj=j0
<
∼ τ
1−d
sj,kp(cid:19)q/p(cid:19)1/q
(1 + k)d−1 sj,kp(cid:19)q/p(cid:19)1/q
as well as
k f4 Bs
p,q(R)k <
2j(s− d
p )q(cid:18) ∞Xk=Kj
p )q(cid:18) ∞Xk=Kj
sj,kp(cid:19)q/p(cid:19)1/q
p )q(cid:18) ∞Xk=0
p )q(cid:18) ∞Xk=0
p )q(cid:18) ∞Xk=1
2j(s− d
2j(s− d
∼ (cid:18) j0−1Xj=0
∼ 2j0
<
2j(s− 1
d−1
p (cid:18) ∞Xj=0
p (cid:18) ∞Xj=0
1−d
<
∼ τ
(1 + k)d−1 sj,kp(cid:19)q/p(cid:19)1/q
(1 + k)d−1 sj,kp(cid:19)q/p(cid:19)1/q
.
Now we turn to the estimate of f1. First we deal with the estimate of the quasi-
norm of the functions ϕτ gj,0. Let in addition s ≥ 1/p. Employing the Moser-type
estimate of Lemma 5.3.7/1 in [29] (applied with r = ∞) we obtain
k ϕτ gj,0 Bs
p,q(R)k
p,q(R)k k gj,0 L∞(R)k + k ϕτ L∞(R)k k gj,0 Bs
p,q(R)k
p,q(R)k + k ϕ L∞(R)k k gj,0 Bs
p,q(R)k
∼ k ϕτ Bs
<
∼ τ −(s−1/p) k ϕ Bs
<
∼ τ −(s−1/p) + 2j(s−1/p)
<
∼ 2j0(s−1/p) ,
<
(49)
since the functions 2−j(s−1/p) gj,0 are atoms and j ≤ j0 + 1. If s < 1/p, we argue by
using real interpolation. Because of
Bs
p,q(R) =(cid:16)Bs0
p,q0(R), Lp(R)(cid:17)Θ,q
,
s = (1 − Θ)s0 > σp(1) ,
see [9], an application of the interpolation inequality
k ϕτ gj,0 Bs
p,q(R)k <
∼ k ϕτ gj,0 Bs0
p,q(R)k1−Θ k ϕτ gj,0 Lp(R)kΘ
41
yields (49) for all s > σp(1). With r := min(1, p, q) and σp(d) < s < d/p we conclude
k f1 Bs
p,q(R)kr ≤
j0+1Xj=0
sj,0r k ϕτ bj,0 Bs
p,q(R)kr
j0+1Xj=0
j0+1Xj=0
(cid:16)
<
∼
2j0(s− 1
p )r
sj,0r .
r(1−d)
p
<
∼ τ
r(1−d)
p
r(1−d)
p
<
∼ τ
<
∼ τ
2r(j0−j)(s− d
p )2j(s− d
p )r sj,0r
sup
2j(s− d
j=0,... ,j0+1
k f1 T Bs
p,∞(R)kr .
p ) sj,0(cid:17)r
p,q(R, L, d) = RAs
This proves the claim for s < d/p.
Step 2. Let s ≥ d/p. As in Step 1 we define j0 ∈ N via the relation 2−j0 ≤ τ < 2−j0+1.
We shall use T As
p,q(R, wd−1), cf. Theorem 9. Alternatively one
could use interpolation, see Propositions 1, 2. The spaces RAs
p,q(R, wd−1) allow
a characterization by Daubechies wavelets, see [17] for Besov spaces and [18] for
Lizorkin-Triebel spaces. The same is true with respect to the ordinary spaces
As
[44, Thm. 1.20]. Let φ denote an appropriate scaling function
and Ψ an associated Daubechies wavelet of sufficiently high order. Let
p,q(R), see e.g.
φ0,ℓ(t) := φ(t − ℓ)
and
Ψj,ℓ(t) := 2j/2 Ψj,ℓ(2jt − ℓ) ,
ℓ ∈ Z ,
j ∈ N0 .
Since Ψ has compact support, say supp Ψ ⊂ [−2N , 2N ] for some N ∈ N, and
supp f0 ⊂ {t ∈ R : t ≥ τ } we find that
hf0, Ψj,ℓi = 0
if
j − j0 ≥ N
and
ℓ ≤ 2j−j0 − 2N .
Hence, f0 has a wavelet expansion given by
hf0, φ0,ℓi φ0,ℓ +
f0 = Xℓ∈Z
= f1 + f2 + f3 .
j0+N −1Xj=0 Xℓ∈Z
hf0, Ψj,ℓi Ψj,ℓ +
∞Xj=j0+N Xℓ>2j−j0 −2N
hf0, Ψj,ℓi Ψj,ℓ .
By the references given above it follows
k f1 Bs
k f2 Bs
k f3 Bs
p,q(R, wd−1)k ≍ (cid:18)Xℓ∈Z
p,q(R, wd−1)k ≍ (cid:18) j0+N −1Xj=0
p,q(R, wd−1)k ≍ (cid:18) ∞Xj=j0+N
hf0, φ0,ℓip(1 + ℓ)d−1(cid:19)1/p
2j(s+ 1
2 − d
2j(s+ 1
2 − d
p )q(cid:18)Xℓ∈Z
p )q(cid:18) Xℓ≥2j−j0 −2N
42
hf0, Ψj,ℓip(1 + ℓ)d−1(cid:19)q/p(cid:19)1/q
hf0, Ψj,ℓip(1 + ℓ)d−1(cid:19)q/p(cid:19)1/q
.
The quasi-norm in the unweighted spaces is obtained by deleting the factor
2−j(d−1)/p (1 + ℓ)d−1, see [44, Thm. 1.20]. This immediately implies
k f1 Bs
k f2 Bs
p,q(R)k <
p,q(R)k <
∼ k f1 Bs
p,q(R, wd−1)k ,
∼ 2(j0+N )(d−1)/p k f2 Bs
p,q(R, wd−1)k .
Moreover, we also obtain
k f3 Bs
p,q(R)k <
∼ 2(j0+N )(d−1)/p k f3 Bs
p,q(R, wd−1)k .
This proves (46) in case s ≥ d/p and 0 < τ < 2.
Proof of Corollary 2
We concentrate on the proof in case of Besov spaces. The proof for Lizorkin-Triebel
spaces is similar.
Step 1. We claim that g ∈ T Bs
proof of Corollary 1.
Under the given restrictions g ∈ RBs
p,q(R, L, d). We argue as in Case 2, Step 2 of the
p,q(R) has a wavelet expansion of the form
g = Xℓ≤c1
hg, φ0,ℓi φ0,ℓ +
∞Xj=0 Xℓ≤c12j
hg0, Ψj,ℓi Ψj,ℓ
with an appropriate constant c1. Since g is even we obtain
g = Xℓ≤c1
hg, φ0,ℓi
φ0,ℓ(t) + φ0,ℓ(−t)
2
+
∞Xj=0 Xℓ≤c12j
hg0, Ψj,ℓi
Ψj,ℓ(t) + Ψj,ℓ(−t)
2
.
The functions 2−j/2(Ψj,ℓ(t)+Ψj,ℓ(−t)) are even L-atoms (up to a universal constant)
centered at c2 Ij,ℓ, where
Ij,k := [−2−j(ℓ + 1), −2−jℓ] ∪ [2−jℓ, 2−j(ℓ + 1)]
(modification if ℓ = 0, see Definition 3). The constant c2 > 1 depends on the size of
the supports of the generators φ and Ψ. Without proof we mention that Theorem
3 remains true also for those more general decompositions. This implies
k g T Bs
p,q(R, L, d)k ≍(cid:16) Xℓ≤c1b
+ (cid:16) ∞Xj=0
hg, φ0,ℓip(cid:17)1/p
(1 + ℓ)d−1hg, φ0,ℓip(cid:17)1/p
2j(s−d/p)q(cid:16) Xℓ≤c12j
+(cid:16) ∞Xj=0
∼ (cid:16) Xℓ≤c1b
2j(s+ 1
<
(1 + ℓ)d−12j/2 hg, Ψj,ℓip(cid:17)q/p(cid:17)1/q
hg, Ψj,ℓip(cid:17)q/p(cid:17)1/q
p )q(cid:16) Xℓ≤c12j
2 − 1
∼ k g Bs
<
p,q(R)k ,
43
see e.g. [44, Thm. 1.20] for the last step. This proves the claim.
Step 2. Since g belongs to T Bs
f := ext g is an element of RBs
p,q(R, L, d) we derive by means of Theorem 3 that
p,q(Rd) and
k f RBs
p,q(R)k .
p,q(Rd)k <
∼ k g T Bs
p,q(R, L, d)k <
∼ k g Bs
Since supp f ⊂ {x : x ≥ a} Corollary 1 yields
k g Bs
p,q(R)k <
∼ a−(d−1)/p k f Bs
p,q(Rd)k ,
because of f0 = g. This completes the proof.
Remark 19 A closer look onto the proof shows that
a(d−1)/p k g As
p,q(R)k <
∼ k f RAs
p,q(Rd)k <
∼ b(d−1)/p k g As
p,q(R)k
with constants independent of g, a > 0 and b ≥ 1.
Proof of Corollaries 3, 4
p,q(Rd), see e.g. [29, 4.8]. Hence, with f also the product ϕ f belongs to RAs
Step 1. Proof of Cor. 3. The function ϕ is a pointwise multiplier for the spaces
p,q(Rd)
As
and we can apply Thm. 5 with respect to this product. Concerning the sharp
embedding relations for the spaces As
p,q(R) into Holder-Zygmund spaces we refer to
[35] and the references given there. This proves the assertion for ϕ0 f0. A further
application of Theorem 2 finishes the proof.
Observe, that we do not need the assumption s > σp,q(d) in case of Lizorkin-Triebel
spaces. We may argue with RF s
p,∞(Rd) first and use the elementary embedding
RF s
Step 2. Proof of Cor. 4. The arguments are as above. Concerning te embedding
relations of the spaces As
p,q(R) into the space of uniformly continuous and bounded
functions we also refer to [35] and the references given there.
p,q(Rd) ֒→ RF s
p,∞(Rd) afterwards.
Remark 20 A different proof of Cor. 4, restricted to Besov spaces, has been given
in [32].
3.8 Test functions
Using our previous results, in particular Corollary 2, we shall investigate the regu-
larity of certain families of radial test functions.
44
Lemma 5 Let 0 < α < min(1, 1/p). Let ϕ ∈ C ∞
supp ϕ ⊂ [−2, −1/2] ∪ [1/2, 2] and ϕ(1) 6= 0.
(i) The function
0 (R) be an even function such that
fα(x) := ϕ(x) x − 1 −α ,
x ∈ Rd ,
belongs to B
1
p −α
p,∞ (Rd) if
α <
1
p
− σp(d).
(50)
(51)
1
p −α
(ii) Suppose 1
p,q (Rd) for any q < ∞.
(iii) Under the same restriction as in (ii) we have that fα does not belong to
p − α > σp(1). Then fα does not belong to B
1
p −α
p,∞ (Rd).
F
[1/2, 2]. Then the regularity of
Proof. Step 1. Proof of (i). Let eϕ ∈ C ∞
gα(t) := eϕ(t) t − 1−α ,
0 (R) be a function such that supp eϕ ⊂
t ∈ R ,
is well understood, cf. e.g.
0 < α < min(1, 1/p). An application of Corollary 2 yields the claim.
Step 2. Proof of (ii) and (iii). It is also known, see again [29, Lem. 2.3.1/1], that
[29, Lem. 2.3.1/1]. One has gα ∈ B
1
p −α
p,∞ (R) as long as
gα 6∈ (B
1
p −α
p,q (R) ∪ F
1
p −α
p,∞ (R)) ,
0 < q < ∞ ,
0 < α < min(1, 1/p) .
These properties do not change when we "add" the reflection of gα to the left half
of the real axis. With other words, if we replace eϕ by ϕ itself we do not change the
regularity properties. Now we use Thm. 5.
Remark 21 Let δ > 0. Then also the regularity of functions like
fα,δ(x) := ϕ(x) x − 1−α (− log x − 1)−δ,
x ∈ Rd ,
(52)
can be checked in this way. With the help of the parameter δ one can see the
microscopic index q. We refer to [37, 5.6.9] or [29, Lem. 2.3.1/1] for details.
Lemma 6 Let α > 0.
(i) Then the function
Φα(x) := max(0, (1 − x2))α ,
x ∈ Rd ,
belongs to B
1
p +α
p,∞ (Rd) if
1
p
+ α > σp(d) .
(53)
(54)
1
p +α
(ii) Suppose 1
p,q (Rd) for any q < ∞.
(iii) Under the same restrictions as in (ii) we have that Φα does not belong to
p + α > σp(1). Then Φα does not belong to B
1
p +α
p,∞ (Rd).
F
45
Proof. Step 1. Proof of (i). First we investigate the one-dimensional case. Let
ψ1, ψ2 ∈ C ∞
0 (R) be such that
supp ψ1 ⊂ [−1/2, ∞),
supp ψ2 ⊂ (−∞, 1/2] and ψ1(t) + ψ2(t) = 1
for all t ∈ R. We put Φi,α := ψi Φα, i = 1, 2. Then Φ1,α behaves near 1 like
φα(t) :=( tα
0
if
if
t > 0 ,
t < 0 ,
near the origin. The regularity of φα is well understood, we refer to [29, Lem. 2.3.1].
As above the transfer to general dimensions d > 1 is done by Corollary 2.
If
Step 2. To prove the statements in (ii) and (iii) we argue by contradiction.
Φα belongs to RAs
p,q(Rd) for any smooth
radial ϕ. Choosing ϕ s.t. 0 6∈ supp ϕ, we my apply Thm. 5 to conclude that
tr (ϕ Φα) ∈ As
p,q(R). In the one-dimensional case
necessary and sufficient conditions are known, we refer again to [29, Lem. 2.3.1].
p,q(Rd), then also ϕ Φα belongs to RAs
p,q(R). But this implies Φ2,α ∈ As
Next we shall consider smooth functions supported in thin annuli.
Lemma 7 Let d ≥ 2, 0 < p, q ≤ ∞ and s > σp(d). Let ϕ ∈ C ∞
0 (R) be an even
function such that ϕ(1) = 1 and supp ϕ ⊂ [−2, −1/2] ∪ [1/2, 2]. Then the functions
fj,λ(y) := ϕ(2jy − λ) ,
y ∈ Rd ,
j ∈ N ,
λ > 0 .
have the following properties:
supp fj,λ ⊂ {y :
p,q(Rd)k ≍ 2j(s− d
(λ − 2) 2−j ≤ y ≤ (2 + λ) 2−j} ,
p ) λ(d−1)/p
k fj,λ RBs
(55)
(56)
with constants in ≍ independent of λ > 2 and j ∈ N.
Proof. Step 1. Estimate from above in (56).
It will be convenient to use the
atomic characterizations described in Subsection 3.3.2. Therefore we shall use the
decompositions of unity from (42)-(44). Thanks to the support restrictions for the
functions ψj,k,ℓ we obtain
fj,λ(y) =
X
max(0,λ−2−n0)≤k≤λ+2+n0
(ϕ(2jy − λ) ψj,k,ℓ(y))
C(d,k)Xℓ=1
where n0 is a fixed number (n0 ≥ 18 would be sufficient). The functions
aj,k,ℓ(y) := 2−j(s− d
p ) ϕ(2jy − λ) ψj,k,ℓ(y)
46
are (s, p)M,−1-atoms for any M (up to a universal constant). Hence
k fj,λ RBs
p,q(Rd)k <
max(0,λ−2−n0)≤k≤λ+2+n0
∼ 2j(s− d
<
p )λ(d−1)/p .
∼ (cid:16)
X
kd−12j(s− d
p )p(cid:17)1/p
Step 2. Estimate from below.
Substep 2.1. First we deal with p = ∞. By construction fj,λ(y) = 1 if y =
(1 + λ) 2−j. Furthermore, calculating the derivatives of fj,λ(y1, 0, . . . , 0), y1 ∈ R, it
is immediate that
k fj,λ C m(Rd)k ≍ 2jm
(57)
for all m ∈ N0. Now we argue by contradiction. We fix s > 1, q1 ∈ (0, ∞] and
assume that
k fj,λ Bs
∞,q1(Rd)k ≤ φ(j, λ) 2js ,
where φ : N × [1, ∞) → (0, 1) and limℓ→∞ φ(jℓ, λℓ) = 0 for some sequence (jℓ, λℓ)ℓ ⊂
N × [1, ∞). We choose Θ ∈ (0, 1) s.t. m = Θ s and q = 1. Real interpolation
between C(Rd) and Bs
∞,q1(Rd) yields
k fj,λ Bm
∞,1(Rd)k ≤ c 2jm (φ(j, λ))Θ ,
∞,1(Rd) ֒→ C m(Rd) leads to a contradiction with (57).
where c is independent of j and λ, see the proof of Thm. 2. The continuous
embedding Bm
Now let 0 < s < 1. We interpolate between Bs
above we could improve the estimate from above with respect to the space B1
Since B1
with p = ∞, 0 < q ≤ ∞, and s > 0.
Substep 2.2. Also obvious is the behaviour in Lp(Rd). For 0 < p ≤ ∞ we have
∞,q(Rd). By arguing as
∞,1(Rd).
∞,1(Rd) ֒→ C 1(Rd) this contradicts again (57). Hence the claim is proved
∞,q1(Rd) and B2
k fj,λ Lp(Rd)k ≍ 2−jd/p λ(d−1)/p .
A few more calculations yield
k fj,λ W 1
p (Rd)k ≍ 2j(1−d/p) λ(d−1)/p ,
(58)
(59)
as long as 1 ≤ p ≤ ∞.
Substep 2.3. Let p1 < ∞. We assume that for some fixed s1 > σp1(d) and q1 ∈ (0, ∞]
k fj,λ Bs1
p1,q1(Rd)k ≤ φ(j, λ) 2j(s1−d/p1) λ(d−1)/p1
p1,q1(Rd) and
holds, where φ is as above. Complex interpolation between Bs1
∞,q0(Rd), s2 > 0, yields an improvement of our estimate with respect to
Bs2
p,q(Rd), where p > p1 is at our disposal. For s2 large we can choose p > 1 s.t.
Bs
s = (1 − Θ)s2 + Θ s1 > 1. Now we need a further interpolation, this time real,
between Bs
p,1(Rd) in this way. But
B1
p,q(Rd) and Lp(Rd), improving the estimate for B1
p (Rd) and so we found a contradiction to (59).
p,1(Rd) ֒→ W 1
47
Remark 22 Obviously there is no q-dependence in Lemma 7. As an immediate
consequence of the elementary embeddings
Bs
p,min(p,q)(Rd) ֒→ F s
p,q(Rd) ֒→ Bs
p,max(p,q)(Rd) ,
see [41, ], and (56) we obtain
k fj,λ F s
p,q(Rd)k ≍ 2j(s− d
p ) λ(d−1)/p ,
λ > 2,
j ∈ N .
Some extremal functions in Ad/p
p,q (Rd) have been investigated by Bourdaud [2], for
p,p(Rd) see also Triebel [43]. We recall the result obtained in [2]. For (α, σ) ∈ R2
Bs
we define
fα,σ(x) := ψ(x)(cid:12)(cid:12)(cid:12) log x(cid:12)(cid:12)(cid:12)
α(cid:12)(cid:12)(cid:12) log log x (cid:12)(cid:12)(cid:12)
−σ
Furthermore we define a set Ut ⊂ R2 as follows:
,
x ∈ Rd .
(60)
Ut :=
(α = 0
and σ > 0) or α < 0
(α = 1 − 1/t and σ > 1/t) or α < 1 − 1/t
(α = 1
and σ ≥ 0) or α < 1
if
if
if
t = 1 ,
1 < t < ∞ ,
t = ∞ ,
Lemma 8 (i) Let 0 < p ≤ ∞ and 1 < q ≤ ∞. Then fα,σ belongs to RBd/p
and only if (α, σ) ∈ Uq.
(ii) Let 1 < p < ∞. Then fα,σ belongs to RF d/p
p,q (Rd) if, and only if (α, σ) ∈ Up.
p,q (Rd) if,
Remark 23 Let us mention that in [2] the result is stated for p ≥ 1 only. However,
the proof extends to p < 1 nearly without changes (in his argument which follows
formula (9) in [2] one has to choose k > d/(2p)).
4 Decay properties of radial functions -- proofs
4.1 Proof of Theorem 10
p,1 (Rd), 0 < p < ∞, and for RF 1/p
Step 1. Proof of (i). Following Remark 10 it will be enough to prove the decay
estimate (13) for RB1/p
p,∞(Rd), 0 < p ≤ 1. A proof in
case RB1/p
p,1 (Rd) has been given in [32]. So we are left with the proof for the Lizorkin-
Triebel spaces. We will follow the ideas of the proof of Cor. 4. Let f ∈ RF 1/p
p,∞(Rd).
Let
f =
∞Xj=0
sj,0 aj,0 +
∞Xj=0
∞Xk=1
48
Cd,kXℓ=1
sj,k aj,k,ℓ,
be an atomic decomposition such that ksj,kfp,∞,dk ≍ k f F 1/p
x > 1. Observe, that for all j ≥ 0 there exists kj ≥ 1 such that
p,∞(Rd)k. We fix x,
kj 2−j ≤ x < (kj + 1) 2−j .
Then the main part of f near (x, 0, . . . , 0) is given by the function
(61)
(62)
f M (y) =
sj,kj aj,kj,0(y) ,
y ∈ Rd ,
(in fact, f is a finite sum of functions of type
sj,kj+rj aj,kj+rj ,tj (y) ,
∞Xj=0
∞Xj=0
and rj and tj are uniformly bounded). For convenience we shall derive an estimate
of the main part f M only. Because of (61) and the normalization of the atoms we
obtain
f M (y) <
∼
sj,kj 2j d−1
p <
∼ x
∞Xj=0
since p ≤ 1. On the other hand
k s fp,∞,dk = k sup
j=0,1,...
≥ k sup
j=0,1,...
sup
k∈N0
sj,kj 2
Using Pj+1,kj+1 ⊂ Pj,kj we obtain the identity
jd
sj,kj 2
sup
j
∞Xj=0
jd
max
i=0,...,j
p eχj,kj (·) =
p eχj,kj (·) Lp(Rd)k ≍ (cid:18) ∞Xj=0
≤ (cid:18) ∞Xj=0
j (cid:17)1/p
sj,kjp kd−1
k sup
j=0,1,...
sj,kj 2
Obviously
(cid:16) ∞Xj=0
,
jd
jd
1−d
sj,k 2
sj,kjp kd−1
p (cid:16) ∞Xj=0
j (cid:17)1/p
p eχj,k(·) Lp(Rd)k
p eχj,kj (·) Lp(Rd)k .
si,ki(cid:0)eχj,kj (·) −eχj+1,kj+1(·)(cid:1) .
j (cid:19)1/p
i=0,...,j(cid:0)si,kip2id(cid:1) 2−jdkd−1
j (cid:19)1/p
i=0,...,j(cid:0)si,kip2id(cid:1) 2−jdkd−1
max
max
.
(63)
(64)
.
(65)
(66)
By the pairwise disjointness of the sets Pj,kj \ Pj+1,kj+1 this implies
Combining (63) - (66) we have proved (13) in case of Lizorkin-Triebel spaces.
Step 2. Proof of (ii). As in Step 1 it will be sufficient to deal with the limiting cases.
Substep 2.1. Let f ∈ RF 1/p
p,∞(Rd). Let Br(0) be the ball in Rd with center in the
origin and radius r. Then (63)-(66) yield
f M (x)
<
∼
<
∼
x
1−d
p k sup
j=0,1,...
x
1−d
p k sup
j=0,1,...
jd
sj,kj 2
sj,k 2
jd
p eχj,kj (·) Lp(Rd \ Br(0))k
p eχj,k(·) Lp(Rd \ Br(0))k
sup
k∈N0
49
where r = x − 18 > 0, see property (b) of the covering (Ωj,k,ℓ) in Subsection
3.3.2. In view of this inequality an application of Lebesgue's theorem on dominated
convergences proves (14).
Substep 2.2. Let f ∈ RB1/p
from Step 1. Since
p,1 (Rd). We argue as in Substep 2.1 by using the notation
lim
r→∞
∞Xj=0(cid:16)Xk≥r
sj,kp (1 + k)d−1(cid:17)1/p
= 0
we conclude from (63) that (14) holds in this case as well.
Step 3. Proof of (iii). We shall use the test functions constructed in Lemma 7. We
choose s0 > max{σp(d), s}. For simplicity we consider x = 2r with r ∈ N. We
choose λ s.t. x = (1 + λ)/2. Hence f1,λ(x) = 1. This implies
d−1
p
x
2−r(d−1)/p f1,λ(x) = 1 ,
and
k 2−r(d−1)/p f1,λ As
p,q(Rd)k <
∼ k 2−r(d−1)/p f1,λ Bso
p,q(Rd)k ≍ 1 ,
see Rem. 22, which proves the claim.
Step 4. Proof of (iv). It will be enough to study the case s = 1/p.
Substep 4.1 Let q > 1. According to Lemma 8(i) there exists a compactly supported
function g0 which belongs to RB1/p
p,q (R) and is unbounded near the origin. By mul-
tiplying with a smooth cut-off function if necessary we can make the support of this
functions as small as we want. For the given sequence (xj)j we define
g(t) :=
∞Xj=1
1
max(xj, j)α g0(t − xj) ,
t ∈ R ,
where we will choose α > 0 in dependence on p. The function g is unbounded
near xj and by means of the translation invariance of the Besov spaces B1/p
p,q (R) we
obtain
k g B1/p
p,q (R)kmin(1,p) ≤ k g0 B1/p
p,q (R)kmin(1,p)
∞Xj=1
j−α min(1,p) <
∼ k g0 B1/p
p,q (R)kmin(1,p) ,
if α · min(1, p) > 1. We employ Rem. 19 (Cor. 2) with respect to each summand.
This yields
k ext g B1/p
p,q (Rd)kmin(1,p)
max(xj, j)−α min(1,p) k ext (g0( · − xj)) B1/p
p,q (Rd)kmin(1,p)
∞Xj=1
≤
<
∼
<
∼
k g0 B1/p
p,q (R)kmin(1,p)
k g0 B1/p
p,q (R)kmin(1,p) ,
∞Xj=1(cid:16) max(xj, j)−α xj(d−1)/p(cid:17)min(1,p)
50
if (α − (d − 1)/p) min(1, p) > 1.
It will be enough to study the situation
Substep 4.2. We turn to the F -case.
s = 1/p and 1 < p < ∞. We argue as above using this time Lemma 8(ii). An
application of Rem. 19 (Cor. 2) yields the result but with the extra condition
1/p > σq(d).
4.2 Traces of BV -functions and consequences for the decay
We recall a definition of the space BV (Rd), d ≥ 2, which will be convenient for us,
see [11, 5.1] or [45, 5.1].
Definition 8 Let g ∈ L1(Rd). We say, that g ∈ BV (Rd) if for every i = 1, . . . , d
there is a signed Radon measure µi of finite total variation such that
ZRd
g(x)
∂
∂xi
φ(x)dx = −ZRd
φ(x)dµi(x),
φ ∈ C 1
c (Rd),
c (Rd) denotes the set of all continuously differentiable functions on Rd with
where C 1
compact support. The space BV (Rd) is equipped with the norm
k g BV (Rd)k = k g L1(Rd)k +
k µi Mk,
dXi=1
where k µi Mk is the total variation of µi.
4.2.1 Proof of Theorem 12
We need some preparations. Recall, the space C 1
c ([0, ∞)) has been defined in Defi-
nition 5. By ωd−1 we denote the surface area of the unit sphere in Rd and by σ the
(d − 1)-dimensional Hausdorff measure in Rd, i.e. ωd−1 := σ({x ∈ Rd : x = 1}). As
above r = r(x) := x.
Lemma 9 (i) If ϕ ∈ C 1
c ([0, ∞)), then all the functions
φi(x) :=
ϕi(t) :=
0
ϕ(r(x)) · xi
r(x)
if x = (x1, . . . , xd) 6= 0 ,
0
if x = 0 ,
1
ωd−1 td−1 Rx=t φ(x) · xi
r(x) dσ(x)
if
t > 0 ,
if
t = 0 ,
i = 1, . . . , d, belong to C 1
(ii) If φ ∈ C 1
c (Rd), then all functions
c (Rd).
i = 1, . . . , d, belong to C 1
c ([0, ∞)).
51
(67)
(68)
Proof. Step 1. Proof of (i). Under the given assumption we immediately get
φi ∈ C 1(Rd \ {0}) and supp φi is compact. Hence, we have to study the regularity
properties in the origin. Obviously, φi(0) = 0 and lim
x→0
φi(x) = 0. We claim that
∂φi
∂xj
(0) =
0
ϕ′(0)
if i = j,
otherwise.
Let e1, . . . ed denote the elements of the canonical basis of Rd. If i = j, then
lim
t→0
φi(tei)
t
= lim
t→0
ϕ(t)
t
= ϕ′(0) .
(0) = ϕ′(0). The cases i 6= j are obvious.
Hence
∂φi
∂xi
Next, we show, that the functions
are continuous in the origin. To begin with
we investigate the case i = j. Then
∂φi
∂xi
(cid:12)(cid:12)(cid:12)(cid:12)
(x) − ϕ′(0)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)ϕ′(r(x)) ·
=(cid:12)(cid:12)(cid:12)(cid:12)ϕ′(r(x)) ·
∂φi
∂xj
x2
i
r2(x)
x2
i
r2(x)
+ (ϕ(r(x)) − ϕ(0)) ·
+ ϕ′(θr(x)) ·
r2(x) − x2
i
r2(x)
r2(x) − x2
i
r3(x)
− ϕ′(0)(cid:12)(cid:12)(cid:12)(cid:12)
− ϕ′(0)(cid:12)(cid:12)(cid:12)(cid:12) ,
where we have used the Mean Value Theorem with a suitable 0 < θ = θ(x) < 1.
The continuity of ϕ′ implies, that the expression on the right-hand side tends to
zero if x → 0. If i 6= j, we write
∂φi
∂xj
(x) = ϕ′(r(x))
= ϕ′(r(x))
xixj
r2(x)
xixj
r2(x)
− ϕ(r(x))
xixj
r3(x)
+ (ϕ(0) − ϕ(r(x)))
xixj
r3(x)
=
xixj
r2(x)
[ϕ′(r(x)) − ϕ′(θr(x))] ,
(69)
with some 0 < θ < 1. Again the continuity of ϕ′ implies, that the expression in (69)
tends to zero as x → 0. Hence, φi ∈ C 1
Step 2. Proof of (ii). The regularity and support properties of ϕi on (0, ∞) are
obvious. Hence, we are left with the study of the behaviour near 0. We shall use
the identity
c (Rd).
xi
r(x)
dσ(x) = 0 ,
t > 0 .
In case t > 0 this leads to the estimate
Zx=t
ωd−1td−1(cid:16)(cid:12)(cid:12)(cid:12)(cid:12)Zx=t
1
≤ 0 + sup
x=t
φ(x) − φ(0).
ϕi(t) ≤
φ(0)
xi
r(x)
(φ(x) − φ(0))
xi
r(x)
dσ(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:17)
dσ(x)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12)Zx=t
52
Hence, ϕi(t) tends to 0 if t → 0+. Furthermore,
ϕi(t)
t
=
=
=
1
ωd−1 td Zx=t
ωd−1 td+1 Zx=t
1
1
ωd−1 td+1
(φ(x) − φ(0))
xi
r(x)
dσ(x)
(∇φ(ηxx) · x) xi dσ(x)
dXj=1 Zx=t(cid:18) ∂φ
∂xj
(ηxx) −
∂φ
∂xj
+
1
ωd−1 td+1
dXj=1
(0)(cid:19) xj xi dσ(x)
(0)Zx=t
∂φ
∂xj
xj xi dσ(x)
with some 0 < ηx < 1. The first term on the right-hand side tends always to zero
as t → 0+ (since
is continuous). From the second term those summands with
∂φ
∂xj
i 6= j are vanishing for all t. If i = j, then the integrand is homogeneous. We obtain,
by taking the limit with respect to t,
ϕ′
i(0) = lim
t→0+
ϕi(t)
t
=
1
ωd−1
∂φ
∂xi
(0)Zy=1
y2
i dσ(y) .
(70)
It remains to check the limit of ϕ′
i(t) if t tends to 0+. Observe
ϕ′
i(t) =
1
ωd−1
d
dtZy=1
φ(ty) yi dσ(y) =
1
ωd−1
dXj=1 Zy=1
∂φ
∂yj
(ty) yj yi dσ(y) .
yj yi dσ(y) = 0 if i 6= j those summands (with i 6= j) tend to 0 if t tends
SinceZy=1
to 0+. Hence
lim
t→0+
ϕ′
i(t) = lim
t→0+
1
ωd−1 Zy=1
∂φ
∂yi
(ty) y2
i dσ(y) = ϕ′(0) ,
see (70). The proof is complete.
Proof of Theorem 12.
Step 1. Let f (x) = g(x) ∈ BV (Rd). We claim that g ∈ BV (R+, td−1). Let
µ1, . . . , µd denote the corresponding signed Radon measures according to Definition
8. By means of
dν :=
xi
r(x)
dµi
dXi=1
we define the measure ν on Rd. Since g(r(x)) is radial we conclude µi({0}) = 0,
i = 1, . . . , d. Hence the measure dν is well defined.
In addition we introduce a
measure ν+ on R+ by
ωd−1 ZA
td−1 dν+(t) :=Z{x∈A}
dν(x),
53
for any Lebesgue measurable subset A ⊂ R+. We fix ϕ ∈ C 1
ϕ(r(x)) xi
c (Rd), cf. Lemma 9(i), we calculate
r(x) ∈ C 1
c ([0, ∞)). Since
ωd−1 Z ∞
0
g(t) [ϕ(s) sd−1]′(t) dt = ωd−1 Z ∞
0
td−1 g(t)(cid:20)ϕ′(t) +
r(x)(cid:21) dx
d − 1
+ ϕ(r(x)) ·
d − 1
t
ϕ(t)(cid:21) dt
r2(x) − x2
i
r3(x)
(cid:21) dx
=
x2
i
r2(x)
g(r(x))(cid:20)ϕ′(r(x)) + ϕ(r(x)) ·
= ZRd
g(r(x))(cid:20)ϕ′(r(x))
dXi=1 ZRd
dXi=1 ZRd
dXi=1 ZRd
= −ωd−1 Z ∞
∂xi (cid:20)ϕ(r(x)) ·
td−1 ϕ(t) dν+(t) .
dµi = −ZRd
xi
r(x)
ϕ(r(x))
g(r(x))
= −
=
∂
0
xi
r(x)(cid:21) dx
ϕ(r(x)) dν(x)
This proves (18). Moreover, (19) follows from
k g(r(x)) L1(Rd)k = ωd−1 k g L1(R, td−1)k
and
ωd−1 Z ∞
0
td−1 dν+(t) = ZRd
dXi=1 ZRd
≤
dν(x) ≤
dXi=1 ZRd
xi
r(x)
dµi
dµi ≤ k g(r(x)) BV (Rd)k .
Step 2. Let g be a function in BV (R+, td−1). We claim, that g(r(x)) ∈ BV (Rd).
Let ν+ be the signed Radon measure associated to g according to (18). We define
ν(A) :=Z ∞
0
σ({x : x = t} ∩ A) dν+(t)
for any Lebesgue measurable set A ⊂ Rd. Further we put µi := xi
Let χA denote the characteristic function of A. Then
r(x)ν, i = 1, . . . d.
and this identity can be extended to
ν(A) =ZRd
φ(x) dν(x) =Z ∞
0
ZRd
χA(x) dν(x) =Z ∞
0
hZx=t
χA(x) dσ(x)i dν+(t)
hZx=t
φ(x) dσ(x)i dν+(t) ,
φ ∈ L1(Rd) ,
54
by using some standard arguments. Next we want to show, that µi, i = 1, . . . , d,
are the weak derivatives of g(r(x)). Let φ ∈ C 1
c (Rd) and let ϕi be the associated
functions, see (68). According to Lemma 9 (ii) we know that ϕi ∈ C 1
c ([0, ∞)).
Using the normalized outer normal with respect to the surface {x : x = T }, which
is obviously given by
n(x) = (n1(x), . . . , nd(x)) =
1
r(x)
(x1, . . . , xd) ,
and the Gauss Theorem, we obtain
φ(x) ni(x) dσ(x)
xi
r(x)
φ(x)
∂φ
∂xi
−ϕi(T ) ωd−1 T d−1 = −Zx=T
= −Zx≤T
=Z ∞
hZx=t
(T ) =Zx=T
ωd−1 (cid:2)ϕi(t) td−1(cid:3)′
ZRd
dx = Z ∞
∂φ(x)
∂xi
g(r(x))
T
0
This formula justifies the identity
Hence
dσ(x) = −Zx=T
∂φ
∂xi
(x) dx =Zx≥T
(x) dσ(x)i dt.
∂φ
∂xi
(x) dx
∂φ
∂xi
(x) dσ(x) ,
T > 0.
g(t)Zx=t
∂φ(x)
∂xi
dσ(x) dt
g(t)[ϕi(s) sd−1]′(t)dt .
= ωd−1 Z ∞
0
Next we use g ∈ BV (R+, td−1). This implies
ZRd
g(r(x))
∂φ(x)
∂xi
0
dx = −ωd−1 Z ∞
= −Z ∞
0 Zx=t
= −ZRd
φ(x)
ϕi(t) td−1 dν+(t)
φ(x) ·
dσ(x) dν+(t)
xi
r(x)
xi
r(x)
dν(x) = −ZRd
φ(x) dµi(x) ,
which proves that the µi are the weak derivatives of g(r(x)).
It remains to prove the estimates for the related norms. The required estimate
follows easily by
ZRd
dµi = ZRd
xi
r(x)
dν(x) =Z ∞
0
hZx=t
xi
r(x)
dσ(x)i dν+(t)
≤ ωd−1 Z ∞
0
The proof is complete.
td−1 dν+(t) .
55
4.2.2 Proof of Theorem 11
Recall, that we will work with the particular representative f of the equivalence
class [f ], see Remark 12. For convenience we will drop the tilde. We shall apply
standard mollifiers. Let ϕ ∈ C ∞
0 (R) be a function such that ϕ ≥ 0, supp ϕ ⊂ [0, 1],
andR ϕ(t) dt = 1. For R > 0 and ε > 0 we define
ϕε(t) := ε−1 Z ∞
R
ϕ(cid:16)t − y
ε (cid:17) dy =Z t−R
−∞
ε
ϕ(z) dz
(71)
(which is nothing but the mollification of the characteristic function of the interval
(R, ∞)). In addition we need a cut-off function. Let η ∈ C ∞
0 (R) s.t. η(t) = 1 if
t ≤ 1 and η(t) = 0 if t ≥ 2. For M ≥ 1 we define ηM (t) := η(t/M), t ∈ R. It is
easily checked that the functions
φM,ε(t) := t1−d ϕε(t) ηM (t) ,
t ∈ R ,
ϕε(t) ηM (t) dν+(t) ,
(72)
belong to C 1
c ([0, ∞)). For g ∈ BV (R+, td−1) this implies
see (18). Since for M > R + ε we have
0
g(t) [φM,ε(s) sd−1]′(t) dt = −Z ∞
Z ∞
g(t) [φM,ε(s) sd−1]′(t) dt = Z R+ε
R
0
Z ∞
0
and
g(t) ε−1 ϕ(cid:16) t − R
ε (cid:17) dt
+M −1 Z ∞
M
g(t) ϕε(t) η′(t/M) dt
g(t) ϕε(t) η′(t/M) dt = 0
M −1 Z ∞
M
lim
M →∞
(g ∈ L1(R+, td−1)), we get
if R is a Lebesgue point of g. But
0
lim
ε↓0,M →∞Z ∞
ϕε(t) ηM (t) dν+(t)(cid:12)(cid:12)(cid:12) ≤Z ∞
R
(cid:12)(cid:12)(cid:12)Z ∞
0
g(t) [φM,ε(s) sd−1]′(t) dt = g(R) ,
ϕε(t) dν+(t) ≤ R1−d Z ∞
R
(73)
td−1 dν+(t) .
Combining (73), (72) with this estimate we have proved (16) and (17) simultane-
ously.
4.3 Proof of the assertions in Subsection 2.2.3
4.3.1 Proof of Lemma 3
Sufficiency of the conditions is obvious, see e.g.
examples investigated in Lemma 8.
[35]. Necessity follows from the
56
4.3.2 Proof of Theorem 13
We argue by using the atomic characterizations in Subsection 3.3.2.
Step 1. Proof of (i). For simplicity let x = 2−r, r ∈ N.
If y satisfies 2−r−1 ≤
y ≤ 2−r+1, then, using the support condition for atoms, we know that f allows an
optimal atomic decomposition such that
f (y) =
∞Xj=0
[2j−r]+n0Xk=max(0,[2j−r]−n0)
C(d,k)Xℓ=1
sj,k aj,k,ℓ(y) .
(74)
Here n0 is a general natural number depending on the decomposition Ω, but not on
r.
Substep 1.1. We assume first that 1
property (f) of the coverings (Ωj,k,ℓ)j,k,ℓ, and the inequality (35) we derive
p . From the L∞-estimate of the atoms,
p < s < d
f (y1, 0, . . . , 0) ≤
sj,k aj,k,ℓ(y1, 0, . . . , 0)
∞Xj=0
[2j−r]+n0Xk=max(0,[2j−r]−n0)
KXℓ=1
2j−r+n0Xk=2j−r−n0
∞Xj=r+n1+1
∞Xj=r+n1+1
2−j(s−d/p) +
n2Xk=0
p,∞(Rd)k(cid:16) r+n1Xj=0
sj,k 2−j(s−d/p) +
<
∼ (cid:16) r+n1Xj=0
<
∼
k f RBs
∼ 2r( d
<
p −s) k f RBs
p,∞(Rd)k ,
sj,k 2−j(s−d/p)(cid:17)
2−(j−r)(d−1)/p 2−j(s−d/p)(cid:17)
p,q(Rd) ֒→ Bs
p,∞(Rd) we obtain the inequality (20).
for appropriate natural numbers n1, n2 (independent of r). For the last two steps of
the estimate we used 1/p < s < d/p. Taking into account the elementary embedding
As
Substep 1.2. Now, let s = 1
p,1 (Rd) the inequality (20)
was proved in [32]. So it remains to consider the Lizorkin -Triebel spaces RF 1/p
p,∞(Rd)
with 0 < p ≤ 1. For simplicity we regard the main part of f ∈ RF 1/p
p,∞(Rd), cf. (74)
and compare with (62). Now kj = 1 if 0 ≤ j ≤ r + n1 and kj ∼ x2j if j > r + n1.
Hence, we obtain
p. For the Besov spaces RB1/p
∞Xj=0
f M (x, 0, . . . , 0) ≤
sj,kj aj,kj,0(y) <
<
∼ x
1−d
p (cid:16) r+n1Xj=0
∞Xj=r+n1+1
sj,kjp +
sj,kp kd−1
∼ (cid:16) ∞Xj=0
j (cid:17)1/p
sj,kjp 2j(d−1)(cid:17)1/p
<
∼ x
1−d
p k f RF 1/p
p,∞(Rd)k ,
where we used p ≤ 1 for the second step and (64)-(65) for the last one.
Step 2. Proof of (ii). By using elementary embeddings it will be enough to prove
(21) with RAs
p,q(Rd) and q small. Again we concentrate on x = 2−r,
p,q(Rd) = RBs
57
r ∈ N. Our test function is taken to be 2−r(s−d/p) fj,λ, see Lemma 7, where we choose
j := 2 + r and λ := 3. Then it follows
k 2−r(s−d/p) fj,λ Bs
p,q(Rd)k ≍ 1
and
2−r(s−d/p) fj,λ(x) = 2−r(s−d/p) = xs−d/p .
The proof is complete.
4.3.3 Proof of Lemma 4
The arguments are the same as in proof of Theorem 10(iv).
4.4 Proof of Theorem 14
We shall use the notation from the proof of Theorem 13, Step 1. Again we employ
the formula (74) and obtain
f (y1, 0, . . . , 0) <
∼ (cid:16) r+n1Xj=0
sj,k +
n2Xk=0
∞Xj=r+n1+1
2j−r +n0Xk=2j−r−n0
sj,k(cid:17)
since s = d/p.
Step 1. Proof of (i). We shall use the standard abbreviation q′ := q/(q − 1). Since
q > 1 we can use Holder's inequality and conclude
r+n1Xj=0
n2Xk=0
sj,k <
∼ r1/q′(cid:16) r+n1Xj=0
n2Xk=0
sj,kq(cid:17)1/q
as well as
∼ (− log x)1/q′
<
k f Bd/p
p,q (Rd)k
(75)
sj,k <
∼
∞Xj=r+n1+1
∼ (cid:16)
2j−r+n0Xk=2j−r−n0
∞Xj=r+n1+1(cid:16) Xk≥2j−r−n0
<
∼ k f Bd/p
<
p,q (Rd)k .
∞Xj=r+n1+1
(1 + k)d−1 sj,kp(cid:17)q/p(cid:17)1/q(cid:16)
2−(j−r)(d−1)/p(cid:16) Xk≥2j−r−n0
∞Xj=r+n1+1
(1 + k)d−1 sj,kp(cid:17)1/p
2−(j−r)q′(d−1)/p(cid:17)1/q′
(76)
The inequalities (75) and (76) yield (22).
Step 2. Proof of (ii). Let 1 < p < p0 < ∞. The Jawerth-Franke embedding
F d/p
p,1 (Rd) ֒→ Bd/p0
p0,p (Rd), see [19] or [35], combined with (22) proves (23).
58
References
[1] J. Bergh and J. Lofstrom, Interpolation spaces. An Introduction,
Springer Verlag, 1976.
[2] G. Bourdaud, A sharpness result for powers of Besov functions, J. Func-
tion Spaces Appl. 2 (2004), 267-277.
[3] G. Bourdaud and Y. Meyer, Fonctions qui op`erent sur les espaces de
Sobolev. J. Funct. Anal. 97 (1991), 351 -- 360.
[4] M. Bownik and Kwok-Pun Ho, Atomic and molecular decompositions
of anisotropic Triebel-Lizorkin spaces. Trans. Amer. Math. Soc. 358 (2006),
no. 4, 1469 -- 1510.
[5] H.-Q. Bui, Weighted Besov and Triebel spaces : Interpolation by the real
method. Hiroshima Math. J., 12 (1982), no. 3, 581 -- 605.
[6] H.-Q. Bui, Characterizations of weighted Besov and Triebel-Lizorkin
spaces via temperatures. J. Funct. Anal., 55 (1984), no. 1, 39 -- 62.
[7] Y. Cho and T. Ozawa, Sobolev inequalities with symmetry, Comm.
Contemp. Math. 11 (2009), 355-365.
[8] S. Coleman, V. Glazer and A. Martin, Action minima among solu-
tions to a class of Euclidean scalar field equations, Comm. in Math. Phys.
58 (1978), 211-221.
[9] R. DeVore and P. Sharpley, Besov spaces on domains in Rd, Trans.
AMS 335 (1993), 843-864.
[10] J. Epperson and M. Frazier, An almost orthogonal radial wavelet
expansion for radial distributions, J. Fourier Analysis Appl. 1 (1995), 311-
353.
[11] L.C. Evans and R.F. Gariepy, Measure theory and fine properties of
functions, CRC Press, Boca Raton, 1992.
[12] W. Farkas, J. Johnsen, and W. Sickel, Traces of anisotropic Besov --
Lizorkin -- Triebel spaces -- a complete treatment of the borderline cases,
Math. Bohemica 125 (2000), 1-37.
[13] J. Franke, On the space F s
p,q of Triebel-Lizorkin type: Pointwise multi-
pliers and spaces on domains. Math. Nachr. 125 (1986), 29-68.
59
[14] M. Frazier and B. Jawerth, Decomposition of Besov spaces. Indiana
Univ. Math. J. 34 (1985), 777-799.
[15] M. Frazier and B. Jawerth, A discrete transform and decomposition
of distribution spaces, J. Funct. Anal. 93 (1990), 34-170.
[16] D. Haroske and I. Piotrowska, Atomic decompositions of function
spaces with Muckenhoupt weights, and some relations to fractal geometry.
Math. Nachr. 281 (2008), 1476-1494.
[17] D. Haroske and L. Skrzypczak, Entropy and approximation numbers
of embeddings of function spaces with Muckenhoupt weights, I. Rev. Mat.
Complut., 21 (2008), 135 -- 177.
[18] M. Izuki and Y. Sawano, Wavelet bases in the weighted Besov and
Triebel-Lizorkin spaces with Aloc
p -weights, JAT 161 (2009), 656-673.
[19] B. Jawerth, Some observations on Besov and Lizorkin-Triebel spaces.
Math. Scand. 40 (1977), 94 -- 104.
[20] B. Jawerth, The trace of Sobolev and Besov spaces if 0 < p < 1, Studia
Math. 62 (1978), 65-71.
[21] T. Kuhn, H.-G. Leopold, W. Sickel, and L. Skrzypczak, Entropy
numbers of Sobolev embeddings of radial Besov spaces, JAT 121 (2003),
244-268.
[22] I. Kuzin, S. Pohozaev, Entire solutions of semilinear elliptic equations.
Birkhauser, Basel 1997.
[23] P.L. Lions, Sym´etrie et compacit´e dans les espaces de Sobolev, J. Funct.
Anal. 49 (1982), 315-334.
[24] O. Mendez and M. Mitrea, The Banach envelopes of Besov and
Triebel -- Lizorkin spaces and applications to partial differential equations,
J. Four. Anal. and Appl. 6(5) (2000), 503-531.
[25] W.M. Ni, A nonlinear Dirichlet problem on the unit ball and its applica-
tions, Indiana Univ. Math. J. 31 (1982), 801-807.
[26] J. Peetre, New thoughts on Besov spaces, Duke Univ. Press, Durham,
1976.
[27] W. Rother, Some existence results for the equation ∆U + K(x) U p = 0,
Comm. in PDE 15 (1990), 1461-1473.
60
[28] W. Rudin, Real and complex analysis, (third ed.) McGraw-Hill Book Co.,
New York, 1987.
[29] T. Runst and W. Sickel, Sobolev spaces of fractional order, Nemytskij
operators and nonlinear partial differential equations, de Gruyter, Berlin,
1996.
[30] V.S. Rychkov, Littlewood-Paley theory and function spaces with Aℓoc
p
weights. Math. Nachr. 224 (2001), 145-180.
[31] H.-J. Schmeisser and H. Triebel, Topics in Fourier analysis and func-
tion spaces. Wiley, Chichester, 1987.
[32] W. Sickel and L. Skrzypczak, Radial subspaces of Besov and
Lizorkin-Triebel spaces: extended Strauss lemma and compactness of em-
beddings, J. Fourier Anal. Appl. 106 (2000), 639-662.
[33] W. Sickel and L. Skrzypczak, On the interplay of regularity and
decay in case of radial functions II. Homogeneous spaces, J. Fourier Anal.
Appl. (to appear), DOI 10.1007/s00041-011-9205-2.
[34] W. Sickel, L. Skrzypczak, and J. Vybiral, Radial subspaces of
Besov- and Lizorkin-Triebel spaces - complex interpolation and character-
ization by differences, (work in progress), Jena, Poznan, Linz 2011.
[35] W. Sickel and H. Triebel, Holder inequalities and sharp embeddings
in function spaces of Bs
p,q and F s
p,q type, ZAA 14 (1995), 105-140.
[36] L. Skrzypczak, Rotation invariant subspaces of Besov and Triebel-
Lizorkin spaces: compactness of embeddings, smoothness and decay prop-
erties. Revista Mat. Iberoamericana 18, 267-299 (2002).
[37] E.M. Stein, Singular integrals and differentiability properties of func-
tions. Princeton Univ. Press, Princeton, 1970.
[38] E.M. Stein, Harmonic analysis, real-variable methods, orthogonality, and
oscillatory integrals. Princeton Univ. Press, Princeton, 1993.
[39] W.A. Strauss, Existence of solitary waves in higher dimensions. Comm.
in Math. Physics 55 (1977), 149-162.
[40] H. Triebel, Interpolation theory, function spaces, differential operators,
North Holland, Amsterdam, 1978.
[41] H. Triebel, Theory of function spaces. Birkhauser, Basel, 1983.
61
[42] H. Triebel, Theory of Function Spaces II. Birkhauser, Basel, 1992.
[43] H. Triebel, Approximation numbers and entropy numbers of embeddings
of fractional Besov-Sobolev spaces in Orlicz spaces. Proc. London Math.
Soc. 66 (1993), 589-618.
[44] H. Triebel, Function spaces and wavelets on domains, EMS Publishing
House, Zurich, 2008.
[45] W.P. Ziemer, Weakly differentiable functions, Springer, New York, 1989.
Leszek Skrzypczak
Adam Mikiewicz University Poznan
Faculty of Mathematics and Computer Science
Ul. Umultowska 87
61-614 Pozna´n, Poland
e -- mail: [email protected]
Winfried Sickel
Friedrich-Schiller-Universitat Jena
Mathematisches Institut
Ernst-Abbe-Platz 2
07743 Jena, Germany
e-mail: [email protected]
Jan Vybiral
Austrian Academy of Sciences
Johann Radon Institute
for Computational and Applied Mathematics
Altenberger Str. 69
A-4040 Linz, Austria
e-mail: [email protected]
62
|
1107.1512 | 1 | 1107 | 2011-07-07T20:21:28 | Borel structure of the spectrum of a closed operator | [
"math.FA",
"math.SP"
] | For a linear operator $T$ in a Banach space let $\sigma_p(T)$ denote the point spectrum of $T$, $\sigma_{p[n]}(T)$ for finite $n > 0$ be the set of all $\lambda \in \sigma_p(T)$ such that $\dim \ker (T - \lambda) = n$ and let $\sigma_{p[\infty]}(T)$ be the set of all $\lambda \in \sigma_p(T)$ for which $\ker (T - \lambda)$ is infinite-dimensional. It is shown that $\sigma_p(T)$ is $\mathcal{F}_{\sigma}$, $\sigma_{p[\infty]}(T)$ is $\mathcal{F}_{\sigma\delta}$ and for each finite $n$ the set $\sigma_{p[n]}(T)$ is the intersection of an $\mathcal{F}_{\sigma}$ and a $\mathcal{G}_{\delta}$ set provided $T$ is closable and the domain of $T$ is separable and weakly $\sigma$-compact. For closed densely defined operators in a separable Hilbert space $\mathcal{H}$ more detailed decomposition of the spectra is done and the algebra of all bounded linear operators on $\mathcal{H}$ is decomposed into Borel parts. In particular, it is shown that the set of all closed range operators on $\mathcal{H}$ is Borel. | math.FA | math |
BOREL STRUCTURE OF THE SPECTRUM
OF A CLOSED OPERATOR
PIOTR NIEMIEC
Abstract. For a linear operator T in a Banach space let σp(T )
denote the point spectrum of T , σp[n](T ) for finite n > 0 be the set
of all λ ∈ σp(T ) such that dim ker(T − λ) = n and let σp[∞](T ) be
the set of all λ ∈ σp(T ) for which ker(T − λ) is infinite-dimensional.
It is shown that σp(T ) is Fσ, σp[∞](T ) is Fσδ and for each finite n
the set σp[n](T ) is the intersection of an Fσ and a Gδ set provided T
is closable and the domain of T is separable and weakly σ-compact.
For closed densely defined operators in a separable Hilbert space
H more detailed decomposition of the spectra is done and the alge-
bra of all bounded linear operators on H is decomposed into Borel
parts. In particular, it is shown that the set of all closed range
operators on H is Borel.
2010 MSC: Primary 47A10, Secondary 54H05, 28A05.
Key words: spectrum, point spectrum, Hilbert space, reflexive Ba-
nach space, Borel set, closed operator, weak topology.
1. Introduction
Let H be a separable Hilbert space. As it is easily seen, any subset
of the complex plane is the point spectrum of an unbounded linear
operator acting on H. To see this, note that the dimension of H as a
vector space is equal to the power of the continuum; take a Hamel basis
{es}s∈R of H and any function ψ : R → C, define A : H → H by the rule
Aes = ψ(s)es, and observe that the point spectrum of A coincides with
the image of the function f (see also [7, Example 2]). So, non-Borel
subsets of C may be the point spectra of certain linear operators acting
on H. It is also worth while to mention that every bounded subset of
C is the point spectrum of a bounded normal (even diagonal) operator
on a nonseparable Hilbert space. However, all densely defined linear
operators in separable Hilbert spaces whose point spectra are non-Borel
which appear in all existing examples in the literature are nonclosed.
It was therefore quite natural to ask for such an example in the class
of closed operators.
In the recent paper we will show that such an
example does not exist. More generally, we will prove the following
result (see Theorem 3.5 and Theorem 4.2):
Let X be a (real or complex) Banach space and T be a
closable operator in X whose domain is separable and
weakly σ-compact. Then the point spectrum of T is Fσ
1
2
P. NIEMIEC
and for each n > 1 the set of all scalars λ such that the
kernel of T − λ is n-dimensional is Fσδ.
In particular, the above theorem applies to the point spectrum of a
closed operator in H, which solves the problem posed in [7] (see Ques-
tion after Example 2 there). Partial solution of this issue for certain
classes of closed operators in Hilbert spaces may be found e.g. in [13,
Corollary 7]. In Example 3.9 we give a simple example of a bounded
densely defined operator in H whose point spectrum coincides with
an arbitrarily given uncountable subset of the unit disc. The prob-
lem whether the point spectrum of a bounded operator in a separable
nonreflexive Banach space is Fσ we leave as open.
Since the celebrated paper by Cowen and Douglas [3] was published,
the point spectra of bounded linear operators are widely investigated
and those operators which have these spectra rich have attracted a
growing interest, see e.g. [9], [17], [6] or [7] and references there.
Notation. In this paper Banach spaces are real or complex and they
need not be separable, and K is the field of real or complex numbers.
By an operator in a Banach space X we mean any linear function from
a linear subspace of X into X and such an operator is on X if its
domain is the whole space. For an operator T in X, D(T ), N(T ), R(T )
and R(T ) denote the domain, the kernel, the range and the closure of
the range of T , respectively. For a scalar λ ∈ K, T − λ stands for the
operator given by the rules: D(T −λ) = D(T ) and (T −λ)(x) = T x−λx
(x ∈ D(T )). The point spectrum of T , σp(T ), is the set of all eigenvalues
of T ; that is, σp(T ) consists of all scalars λ ∈ K such that N(T − λ) is
nonzero. The operator T is closed iff its graph Γ(T ) := {(x, T x) : x ∈
D(T )} is a closed subset of X × X. T is closable iff the norm closure
of Γ(T ) in X × X is the graph of some operator in X. Whenever E
and F are Banach spaces, B(E, F ) stands for the space of all bounded
operators from the whole space E to F . A subset of a topological
space is said to be σ-compact if it is the union of a countable family of
compact subsets of the space. A weakly σ-compact subset of a Banach
space is a set which is σ-compact with respect to the weak topology
of the space. A Borel subset of a topological space is a member of
the σ-algebra generated by all open sets in the space. Adapting the
notation proposed by A.H. Stone [14], a subset of a topological space
which is the intersection of an open and a closed set is called by us (of
type) F ∩ G. Similarly, any set being the intersection of a Gδ and an
Fσ set will be said to be Fσ ∩ Gδ.
2. Arbitrary closed operator in a Banach space
In this and next sections X denotes a Banach space.
Recall that a Souslin set is a metrizable space which is the contin-
uous image of a separable complete metric space (see the beginning
BOREL STRUCTURE OF THE SPECTRUM
3
of Chapter XIII of [8] and Theorem XIII.1.1 there, or Appendix in
[15]). Souslin subsets of separable complete metric spaces are abso-
lutely measurable, i.e. whenever A is a Souslin subset of a separable
complete metric space Y and µ is a nonnegative σ-finite Borel measure
on Y , there are Borel subsets B and C of Y such that B ⊂ A ⊂ C and
µ(C \ B) = 0 (see e.g. [15, Theorem A.13]).
We begin with
2.1. Proposition. Let T be a closable operator in X whose domain is
the image of a separable Banach space under a bounded operator. The
point spectrum of T is a Souslin subset of K. In particular, σp(T ) is
absolutely measurable.
Proof. Let E be a separable Banach space and S ∈ B(E, X) be such
that S(E) = D(T ). Passing to the quotient of E by N(S), we may
assume that S is one-to-one. Since T is closable, the operator C =
T S : E → X is bounded (by the Closed Graph Theorem). Notice that
σp(T ) = {λ ∈ K : N((T −λ)S) = N(C −λS) 6= {0}}. Let W be the set
of all x ∈ E such that kxk = 1 and Cx = λ(x)Sx for some λ(x) ∈ K.
We have obtained a function λ : W → K with λ(W ) = σp(T ). So, to
finish the proof it suffices to show that W is closed and λ is continuous.
Suppose xn ∈ W and xn → x ∈ X as n → ∞. Then kxk = 1
and hence Sx 6= 0. We infer from this that the sequence (λ(xn))∞
n=1 is
bounded because λ(xn) = kCxnk
kSxk (n → ∞). Now if (λ(xnk))∞
k=1
is any subsequence which converges to some w ∈ K, then Cx = wSx.
This shows that x ∈ W and w = λ(x). So, W is closed and λ is
continuous.
(cid:3)
kSxnk → kCxk
Now since the domain of a closed operator is the image of its graph
under the projection onto the first factor, Proposition 2.1 applies to
closed operators in separable Banach spaces and gives
2.2. Corollary. The point spectrum of a closed operator in a separable
Banach space is Lebesgue measurable.
In the next section (Corollary 3.7) we shall prove that the point spec-
trum of a closed operator in a separable reflexive Banach space is Borel
(even Fσ). We do not know whether the assumption of reflexivity may
be relaxed in this.
Question 1. Is the point spectrum of a bounded operator on a sepa-
rable Banach space Borel?
3. Operators with weakly σ-compact domains
For an operator T in X, a subset K of X and a nonnegative real
constant M, let
ΛT (K, M) = {w ∈ K : N(T − w) ∩ K 6= ∅ and w 6 M}.
4
P. NIEMIEC
Notice that ΛT (K, M) consists of all w ∈ K with w 6 M provided
0 ∈ K. The main tool of this section is the following
3.1. Lemma. If T is a closable operator in X and K is a weakly
compact subset of D(T ), then the set ΛT (K, M) is compact for every
M > 0.
Proof. We may and do assume that 0 /∈ K. Let W be the set of all
x ∈ K for which there is (unique) λ(x) ∈ K such that T x = λ(x)x
and λ(x) 6 M. Thus we have obtained a function λ : W → K. Since
λ(W ) = ΛT (K, M), it suffices to show that W is weakly compact and
λ is continuous when W is equipped with the weak topology.
Let X = (xσ)σ∈Σ be a net in W which is weakly convergent to some
x ∈ K. We need to show that x ∈ W and limσ∈Σ λ(xσ) = λ(x).
If (xτ )τ ∈Σ′ is a subnet of X such that limτ ∈Σ′ λ(xτ ) = w ∈ K, then
(xτ , T xτ )τ ∈Σ′ is weakly convergent to (x, wx). Since the norm closure
of Γ(T ) is weakly closed (and is the graph of some operator), we infer
from this that T x = wx and thus x ∈ W and λ(x) = w. Since λ(W ) is
bounded and any convergent subnet of (λ(xσ))σ∈Σ has the same limit,
the latter net converges to λ(x) and we are done.
(cid:3)
With use of the foregoing result we now easily prove the following
3.2. Proposition. If T is a closable operator in X such that D(T )\{0}
is weakly σ-compact, then σp(T ) is Fσ.
Proof. Write D(T ) \ {0} = S∞
n=1 Kn with each Kn weakly compact,
note that λ ∈ σp(T ) iff λ ∈ ΛT (Kn, m) for some natural n and m and
apply Lemma 3.1.
(cid:3)
For applications of the above result, we need the next well known
fact. For reader convenience, we give its short proof.
3.3. Lemma. Every weakly compact subset of X which is separable in
the norm topology is weakly metrizable.
Proof. Let K be a separable weakly compact subset of X. Then
(K − K) \ {0} is separable as well and thus there is a sequence of
linear functionals fn : X → K of norm 1 such that for every nonzero
z ∈ K − K there is n with fn(z) 6= 0. Observe that then the family
{fn}n∈N separates the points of K. Finally, since K is weakly compact,
the formula K ∋ x 7→ (fn(x))n∈N ∈ ∆N with ∆ = {w ∈ K :
w 6 1}
defines a topological embedding of K (equipped with the weak topol-
ogy) into the compact metrizable space ∆N.
(cid:3)
Now we have
3.4. Proposition. If D is a linear subspace of X, then D\{0} is weakly
σ-compact iff D is separable and weakly σ-compact as well.
BOREL STRUCTURE OF THE SPECTRUM
5
Proof. First assume that D is separable and D = S∞
n=1 Kn with each
Kn weakly compact. By Lemma 3.3, Kn is weakly metrizable and thus
Ln := Kn\{0} is an Fσ-subset of Kn with respect to the weak topology.
So, D \ {0} = S∞
n=1 Ln and each Ln is weakly σ-compact.
n=1 and (εn)∞
fn(x) < εn, n = 1, 2, . . .}. In particular, T∞
Conversely, if D \ {0} is weakly σ-compact, so is D and {0} is weakly
Gδ in D. Thus, by the definition of the weak topology, there are
sequences (fn)∞
n=1 of continuous linear functionals on X
and of positive real numbers (respectively) such that {0} = D ∩ {x ∈
N(fn) ∩ D = {0}
X :
and therefore the function ψ : D ∋ x 7→ (fn(x))∞
n=1 ∈ KN is one-to-one.
What is more, ψ is continuous with respect to the weak topology on D
and the product one on KN. So, ψ restricted to any weakly compact
subset K of D is a topological embedding and therefore every such
K is weakly separable. We conclude from this that D itself is weakly
separable, being weakly σ-compact. Finally, if A is a countable weakly
dense subset of D and E is the norm closure of the linear span of A,
then E is separable and weakly closed. The latter implies that D ⊂ E
and we are done.
(cid:3)
n=1
It follows from Proposition 3.4 that Proposition 3.2 is equivalent to
3.5. Theorem. If T is a closable operator in X whose domain is sep-
arable and weakly σ-compact, then σp(T ) is Fσ.
By Baire's theorem, a Banach space is weakly σ-compact iff it is
reflexive. Thus Theorem 3.5 applies mainly to reflexive spaces. For
example:
3.6. Corollary. If T is a closable operator in X whose domain is sep-
arable and is the image of a reflexive Banach space under a bounded
operator, then σp(T ) is Fσ.
Now argument used in the note just before Corollary 2.2 shows that
3.7. Corollary. The point spectrum of a closed operator in a separable
reflexive Banach space is Fσ. In particular, the point spectrum of a
closed densely defined operator in a separable Hilbert space is Fσ.
The second statement of the above result answers in the affirmative
the question posed by Jung and Stochel in [7].
3.8. Remark. Corollary 3.7 for bounded operators may also be proved
in a different way. Namely, if T is a bounded operator on a separable
reflexive Banach space X (respectively a closed densely defined oper-
ator in a separable Hilbert space H), then, by the reflexivity of X,
N(T ) = {0} iff the range of the dual operator T ′ : X ′ → X ′ (respec-
tively iff the range of the adjoint operator T ∗) is dense in X ′ (in H).
So, σp(T ) consists of precisely those λ ∈ K for which R(Sλ) is not
dense in X ′ (in H) where Sλ = T ′ − λ (Sλ = T ∗ − ¯λ). Consequently, if
6
P. NIEMIEC
{yn}∞
n=1 is a dense sequence in X ′ (in H), then
σp(T ) = [
\
F (n, m, x)
n,m>1
x∈D(S0)
with F (n, m, x) := {λ ∈ K : kSλx − ynk > 1
m}. We infer from the
closedness of F (n, m, x)'s that indeed σp(T ) is Fσ. This elementary and
simpler proof does not work in case when T is only closable. Moreover,
this approach hides the crucial property of reflexivity for this issue
revealed by Lemma 3.1, namely the weak σ-compactness of the domain
of an operator.
3.9. Example. Let H be a complex separable infinite-dimensional Hil-
bert space and let S : H → H be the backward shift with respect
to the orthonormal basis E = (en)∞
n=0 of H; that is, Se0 = 0 and
Sen = en−1 for n > 1. Let D be the open unit disc in C and ψ : D ∋
z 7→ P∞
n=0 znen ∈ H. Recall that σp(S) = D, Sψ(z) = zψ(z) for each
z ∈ D and the image of ψ is linearly independent. What is more, the
linear span DF of ψ(F ) is dense in H for every uncountable subset F
of D. (To see the latter, note that if x is orthogonal to ψ(F ), then
P∞
n=0 xnzn = 0 for each z ∈ F where xn is the n-th Fourier coefficient
of x in the basis E. Thus xn = 0 for any n and x = 0 as well.) So, the
restriction SF of S to DF is a bounded densely defined operator in H
such that σp(SF ) = F . The example shows that in Theorem 3.5 the
domain of a closable operator has to have additional properties beside
the separability.
3.10. Example. The inclusion map of L∞[0, 1] into L2[0, 1], as the
dual operator of the inclusion map of L2[0, 1] into L2[0, 1], is continu-
ous with respect to the weak-* topology of the domain and the weak
one in L2[0, 1]. This implies that L∞[0, 1], considered as a subspace
of L2[0, 1], is weakly σ-compact. However, there is no closed opera-
tor in L2[0, 1] whose domain is L∞[0, 1] (compare with the remark on
page 257 in [5]). The example shows that Theorem 3.5 can have quite
natural applications also for nonclosed operators and that it is more
general than Corollary 3.6 which does not apply here, since there is
no bounded operator on a reflexive Banach space into L2[0, 1] whose
image is L∞[0, 1] (again, by the remark on page 257 in [5]).
4. Decomposition of the point spectrum
Let T be an operator in X. For n > 1 let σp[n](T ) be the set of all λ ∈
K such that dim N(T −λ) = n and let σp[∞](T ) = σp(T )\S∞
n=1 σp[n](T ).
Our aim is to show that all just defined sets are Borel provided T is
closable and has separable and weakly σ-compact domain. To do this,
we need
BOREL STRUCTURE OF THE SPECTRUM
7
4.1. Lemma. If V is a T2 topological vector space, then for each n > 2
the set F [n] of all (x1, . . . , xn) ∈ V n such that x1, . . . , xn are linearly
dependent is closed in the product topology of V n.
j=2 λjxτ (j) for some λ2, . . . , λn ∈ ∆. Notice that F [n] = Sn
λ 6 1}. For a permutation τ of {1, . . . , n}
Proof. Let ∆ = {λ ∈ K :
let Fτ be the set of all n-tuples (x1, . . . , xn) ∈ V n such that xτ (1) =
Pn
j=1 Fτj
where τj(1) = j, τj(j) = 1 and τj(k) = k for k 6= 1, j. So, it suffices to
show that Fτ is closed. But Fτ = p(Φ−1({0}) where p : ∆n−1×V n → V n
is the projection onto the second factor and
Φ : ∆n−1 × V n ∋ (λ2, . . . , λn; x1, . . . , xn) 7→ x1 −
n
X
j=2
λjxj ∈ V.
Since ∆n−1 is compact, p is a closed map and we are done.
(cid:3)
The main result of the section is the following
4.2. Theorem. If T is a closable operator in X whose domain is sep-
arable and weakly σ-compact, then σp[n](T ) is Fσ ∩ Gδ for finite n and
σp[∞](T ) is Fσδ.
Proof. First fix finite N > 0. Let F [1] = {0} ⊂ X and F [N] be as
in the statement of Lemma 4.1 for N > 1. By Lemma 4.1, F [N] is
weakly closed in X N . Write D(T ) = S∞
n=1 Kn with each Kn weakly
compact. By Lemma 3.3, all Kn's are weakly metrizable and hence
(QN
j=1 Knj for any n1, . . . , nN when each Kn
is equipped with the weak topology. So, D(T )N \ F [N] may be written
in the form
j=1 Knj ) \ F [N] is Fσ in QN
(4-1)
D(T )N \ F [N] =
∞
[
n=1
Ln
where each Ln is a weakly compact subset of X N . Put S : D(T )N ∋
(x1, . . . , xN ) 7→ (T x1, . . . , T xN ) ∈ X N and observe that S is a closable
operator in X N . Thanks to Lemma 3.1, the set ΛS(Ln, m) is compact
for all natural n and m. So, GN := Sn,m ΛS(Ln, m) is Fσ. But GN is
the set of all λ ∈ σp(T ) such that N(T − λ) is at leastN-dimensional,
by (4-1).
Now observe that for finite n, σp[n](T ) = Gn \ Gn+1 and thus σp[n](T )
(cid:3)
is Fσ ∩ Gδ. Finally, since σp[∞](T ) = T∞
n=1 Gn, σp[∞](T ) is Fσδ.
5. Decomposition of the spectrum: Hilbert space
From now on, H is a complex separable infinite-dimensional Hilbert
space and B(H) = B(H, H). Every subset of B(H) which is Borel
with respect to, respectively, the weak operator, the strong operator
or the norm topology is called by us, respectively, WOT-Borel, SOT-
Borel and shortly Borel. Similarly, if f : S → B(H) with S ⊂ B(H)
8
P. NIEMIEC
is such that S and the inverse image of any WOT-Borel (respectively
SOT-Borel; Borel) set is WOT-Borel (SOT-Borel; Borel) is said to be
WOT-Borel (SOT-Borel ; Borel ). A fundamental result in this topic
says that every WOT-Borel set is SOT-Borel and conversely (see e.g.
[4]). Thus the same property holds true for WOT-Borel and SOT-Borel
functions. Therefore we shall only speak on WOT-Borel and Borel sets
and functions. Whenever in the sequel the classes Fσ, Gδ, F ∩ G, etc.,
appear, they are understood with respect to the norm topology.
We are interested in the decomposition of B(H) into Borel sets each
of which collects the operators of a similar type. For n = 0, 1, 2, . . .
let Σn(H) be the set of all operators A ∈ B(H) such that dim R(A) =
n. Further, for n, m = 0, 1, 2, . . . , ∞ let, respectively, Σ1
n,m(H) and
Σ0
n,m(H) consist of all A ∈ B(H) such that dim N(A) = n, dim R(A)⊥ =
m, dim R(A) = ∞ and, respectively, R(A) is closed or not. Notice that
all just defined sets of operators form a countable decomposition of
B(H), denoted by Σ(H), into nonempty and pairwise disjoint sets. We
want to show that each of them is Borel in B(H). To do this, we need
5.1. Lemma. For each n = 0, 1, 2, . . . , ∞, let Σn,∗(H) be the set of all
operators A ∈ B(H) such that dim N(A) = n. Then Σ0,∗(H) is Gδ,
Σn,∗(H) is Fσ ∩ Gδ for finite n > 0 and Σ∞,∗(H) is Fσδ in B(H).
Proof. We mimic the proof of Theorem 4.2. For fixed finite N > 0
let F [N] be defined as there. Write HN \ F [N] = S∞
n=1 Ln with each
Ln weakly compact in HN . For T ∈ B(H) let T ×N ∈ B(HN ) denote
the operator HN ∋ (x1, . . . , xN ) 7→ (T x1, . . . , T xN ) ∈ HN . Since Ln
is bounded, the function ψ : B(H) × Ln ∋ (T, x) 7→ T ×N x ∈ HN is
continuous when B(H) is considered with the norm topology and Ln
and HN with the weak one. Hence ψ−1({0}) is closed in B(H) × Ln in
the product topology of these topologies. Finally, since Ln is weakly
compact, the set Fn := p(ψ−1({0})) is closed in the norm topology
of B(H) where p : B(H) × Ln → B(H) is the projection onto the first
factor. Thus GN := S∞
n=1 Fn is Fσ. Note that GN coincides with the
set of all A ∈ B(H) for which dim N(A) > N.
Now we have Σn,∗(H) = Gn \ Gn+1 for finite n > 0, Σ∞,∗(H) =
n=1 Gn and Σ0,∗(H) = B(H) \ G1 which clearly finishes the proof. (cid:3)
T∞
Now let CR(H) be the set of all closed range operators of B(H). Our
next purpose is to show that CR(H) is WOT-Borel (and hence Borel
as well). This is however not so simple. Firstly, the maps B(H) ∋
A 7→ A∗ ∈ B(H) and B(H) × B(H) ∋ (A, B) 7→ AB ∈ B(H) are
WOT-Borel (cf.
[4]), since the first of them is WOT-continuous and
the second is SOT-continuous on bounded sets. Secondly, closed range
operators may be characterized in the space B(H) by means of the
so-called Moore-Penrose inverse in the following way:
BOREL STRUCTURE OF THE SPECTRUM
9
5.2. Proposition. An operator A ∈ B(H) has closed range iff there
is B ∈ B(H) such that ABA = A, BAB = B and AB and BA are
selfadjoint. What is more, the operator B is uniquely determined by
these properties and if A has closed range, B = A† := (A(cid:12)(cid:12)R(A∗))−1P
where P is the orthogonal projection onto the range of A.
The operator A† appearing in the statement of Proposition 5.2 is the
above mentioned Moore-Penrose inverse of an operator A ∈ CR(H).
For the proof of Proposition 5.2, see e.g. [12].
In the proof of the next result we use the theorem on Souslin sets
which states that if Y and Z are separable complete metric spaces,
B is a Borel subset of Y and f : B → Z is a continuous one-to-one
function, then f (B) is a Borel subset of Z (see [8, Theorem XIII.1.9]
or [15, Corollary A.7], or [16, Theorem A.25] for more general result).
5.3. Theorem. The set CR(H) is WOT-Borel.
Proof. For r > 0 let Br be the closed ball in B(H) with center at 0 and
of radius r equipped with the weak operator topology. Let ψr : B2
r ∋
(T, S) 7→ (T ST − T, ST S − S, S∗T ∗ − T S, T ∗S∗ − ST ) ∈ B4
s where
s := 2(r + 1)3. By the note preceding the statement of the theorem, ψr
is WOT-Borel. So, the set Cr := ψ−1
r ({(0, 0, 0, 0)}) is WOT-Borel in Br
as well. By Proposition 5.2, Cr = {(A, A†) : A ∈ CR(H), A, A† ∈ Br}.
So, the projection pr of Cr onto the first factor is one-to-one. But pr is
WOT-continuous and Br and Bs are compact metrizable spaces. We
infer from this that Er := pr(Cr) is WOT-Borel in Bs. Finally, the
observation that Bs is WOT-Borel in B(H) and CR(H) = S∞
n=1 En
finishes the proof.
(cid:3)
Question 2. Which additive or multiplicative class, in the hierarchy
of WOT-Borel sets in B(H), the set CR(H) is of?
Now we are ready to prove the following
5.4. Theorem. For each k, n, m = 0, 1, 2, . . . , ∞ with k finite:
(a) Σk(H) is F ∩ G,
(b) Σ1
n,m(H) is F ∩ G (respectively open) provided m or n is finite
(respectively m = 0 or n = 0),
(c) Σ0
0,0(H) is Gδ; Σ0
n,m(H) is Fσ ∩ Gδ for finite n and m; Σ0
n,m(H) is
Fσδ if either n or m is infinite (and the other is finite),
(d) Σ1
∞,∞(H) and Σ0
∞,∞(H) are Borel.
Proof. Clearly, for each finite k the set of all finite rank operators A ∈
B(H) such that dim R(A) 6 k is closed and therefore Σk(H) is F ∩ G.
Further, for each n = 0, 1, 2 . . . , ∞ let Σn,∗(H) be as in Lemma 5.1 and
let Σ∗,n(H) = {A∗ : A ∈ Σn,∗(H)}. Observe that Σ∗,n(H) is of the
same Borel class as Σn,∗(H).
10
P. NIEMIEC
Suppose that n or m is finite. With no loss of generality, we may
assume that n 6 m. Put k = n − m ∈ Z ∪ {−∞}. The set F (k)
of all semi-Fredholm operators of index k is open in B(H) (see e.g.
Proposition XI.2.4 and Theorem XI.3.2 in [2]). It may also be shown
that for each integer l > 0 the set Fl(k) of all A ∈ F (k) such that
dim N(A) 6 l is open in B(H) as well (see e.g. [11, Proposition 5.3]).
Now the relation Σ1
n,m(H) = Fn(k) \ Fn−1(k) (with F−1(k) = ∅) shows
(b). Further, since Σ0
n,m(H) = Σn,∗(H) ∩ Σ∗,m(H) \ F (k), we infer from
Lemma 5.1 the assertion of (c).
Finally, Σ1
∞,∞(H) = Σ∞,∗(H) ∩ Σ∗,∞(H) ∩ CR(H) \ S∞
n=0 Σn(H). So,
this set is Borel by Theorem 5.3. Since Σ0
∞,∞(H) is the complement in
B(H) of the union of all other members of Σ(H), it is Borel as well. (cid:3)
Question 3. Which additive or multiplicative class, in the hierarchy
of Borel sets in B(H), the sets Σ1
∞,∞(H) and Σ0
∞,∞(H) are of?
Now let T be a closed densely defined operator in H. By σ(T ) we
denote the spectrum of T . That is, a complex number λ does not
belong to σ(T ) iff N(T − λ) = {0}, R(T − λ) = H and (T − λ)−1 is
bounded (the latter condition may be omitted by the Closed Graph
Theorem). We decompose the complex plane into parts corresponding
to the members of Σ(H):
• σf (T ) is the set of all z ∈ C such that R(T − z) is finite-
dimensional,
• for n, m = 0, 1, 2, . . . , ∞ let σ1
n,m(T ) be the sets
consisting of all z ∈ C for which dim N(T − z) = n, dim R(T −
z)⊥ = m, dim R(T − z) = ∞ and, respectively, R(T − z) is
closed or not.
n,m(T ) and σ0
Notice that C \ σ(T ) = σ1
0,0(T ) is the resolvent set of T and that the
above defined sets are pairwise disjoint and cover the complex plane.
The collection of all of them is denoted by Σ(T ). We say that the
sets σ1
n,m(H) and
Σ0
n,m(T ) correspond to, respectively, Σ1
n,m(H). What we want is to prove
n,m(T ) and σ0
5.5. Proposition. For every closed densely defined operator T in H,
Σ(T ) consists of Borel subsets of C. What is more, card σf (T ) 6 1 and
each member of Σ(T ) different from σf (T ) is of the same Borel class
as the corresponding member of Σ(H).
Proof. First observe that if z ∈ σf (T ), then D(T ) = H and T is
bounded (since N(T − z) is closed). We conclude from this that indeed
card σf (T ) 6 1. Further, since T is closed, there is an operator A ∈
B(H) such that N(A) = {0} and R(A) = D(T ) (e.g. A := (T + 1)−1;
see also [5, Theorem 1.1]). Put C = T A ∈ B(H). Notice that for each
z ∈ C, R(C − zA) = R(T − z) and N(C − zA) = A−1(N(T − z)),
and thus dim N(T − z) = dim N(C − zA). It is infered from this that
BOREL STRUCTURE OF THE SPECTRUM
11
n,m(T ) iff C − zA ∈ Σj
z ∈ σj
C ∋ z 7→ C − zA ∈ B(H) finishes the proof.
n,m(H). So, the continuity of the function
(cid:3)
We end the paper with notes concerning the so-called residual and
continuous spectra of an operator. In the literature there are at least
two different, nonequivalent definitions of them. For some mathemati-
cians the residual spectrum σr(T ) of a closed densely defined operator
T in a separable Hilbert space H consists of all λ ∈ σ(T ) \ σp(T )
such that R(T − λ) is closed (e.g.
[1]), for others it is the set of all
λ ∈ σ(T ) \ σp(T ) such that R(T − λ) 6= H (e.g.
[10]). Similarly, the
continuous spectrum σc(T ) of T for some people means the set of all
λ ∈ K such that R(T − λ) is nonclosed (e.g. [1]), for others it coincides
with σ(T ) \ (σp(T ) ∪ σr(T )) (e.g.
[10]). Somehow or other -- what-
ever definitions we choose among the above, both the residual and the
continuous spectra are Borel and are the unions of subfamilies of Σ(T ).
References
[1] M.S. Birman and M.Z. Solomjak, Spectral Theory of Self-Adjoint Operators in
Hilbert Space, D. Reidel Publishing Co., Dordrecht, 1987.
[2] J.B. Conway, A Course in Functional Analysis, Springer-Verlag, New York,
1985.
[3] M.J. Cowen and R.G. Douglas, Complex geometry and operator theory, Acta
Math. 141 (1978), 187 -- 261.
[4] E.G. Effros, Global structure in von Neumann algebras, Trans. Amer. Math.
Soc. 121 (1966), 434 -- 454.
[5] P.A. Fillmore and J.P. Williams, On operator ranges, Adv. in Math. 7 (1971),
254 -- 281.
[6] C. Jiang, Similarity classification of Cowen-Douglas operators, Canad. J.
Math. 56 (2004), 742 -- 775.
[7] I.B. Jung and J. Stochel, Subnormal operators whose adjoints have rich point
spectrum, J. Funct. Anal. 255 (2008), 1797 -- 1816.
[8] K. Kuratowski and A. Mostowski, Set Theory with an Introduction to Descrip-
tive Set Theory, PWN -- Polish Scientific Publishers, Warszawa, 1976.
[9] J.E. McCarthy, Boundary values and Cowen-Douglas curvature, J. Funct.
Anal. 137 (1996), 1 -- 18.
[10] W. Mlak, Hilbert Spaces and Operator Theory, PWN -- Polish Scientific Pub-
lishers and Kluwer Academic Publishers, Warszawa-Dordrecht, 1991.
[11] P. Niemiec, Norm closures of orbits of bounded operators, to appear.
[12] R. Penrose, A generalized inverse for matrices, Proc. Cambridge Phil. Soc. 51
(1955), 406 -- 413.
[13] J. Stochel and F.H. Szafraniec, On normal extensions of unbounded operators.
III. Spectral properties, Publ. Res. Inst. Math. Sci. 25 (1989), 105 -- 139.
[14] A.H. Stone, Absolute Fσ-spaces, Proc. Amer. Math. Soc. 13 (1962), 495 -- 499.
[15] M. Takesaki, Theory of Operator Algebras I (Encyclopaedia of Mathematical
Sciences, Volume 124), Springer-Verlag, Berlin-Heidelberg-New York, 2002.
[16] M. Takesaki, Theory of Operator Algebras III (Encyclopaedia of Mathematical
Sciences, Volume 127), Springer-Verlag, Berlin-Heidelberg-New York, 2003.
[17] K. Zhu, Operators in Cowen-Douglas classes, Illinois J. Math. 44 (2000), 767 --
783.
12
P. NIEMIEC
Piotr Niemiec, Jagiellonian University, Institute of Mathematics,
ul. Lojasiewicza 6, 30-348 Krak´ow, Poland
E-mail address: [email protected]
|
1209.1221 | 1 | 1209 | 2012-09-06T08:27:53 | On the spectrum of frequently hypercyclic operators | [
"math.FA",
"math.DS"
] | A bounded linear operator $T$ on a Banach space $X$ is called frequently hypercyclic if there exists $x\in X$ such that the lower density of the set $\{n\in\N:T^nx\in U\}$ is positive for any non-empty open subset $U$ of $X$. Bayart and Grivaux have raised a question whether there is a frequently hypercyclic operator on any separable infinite dimensional Banach space. We prove that the spectrum of a frequently hypercyclic operator has no isolated points. It follows that there are no frequently hypercyclic operators on all complex and on some real hereditarily indecomposable Banach spaces, which provides a negative answer to the above question. | math.FA | math |
On the spectrum of frequently hypercyclic operators
Stanislav Shkarin
Abstract
A bounded linear operator T on a Banach space X is called frequently hypercyclic if there
exists x ∈ X such that the lower density of the set {n ∈ N : T nx ∈ U } is positive for any
non-empty open subset U of X. Bayart and Grivaux have raised a question whether there
is a frequently hypercyclic operator on any separable infinite dimensional Banach space. We
prove that the spectrum of a frequently hypercyclic operator has no isolated points. It follows
that there are no frequently hypercyclic operators on all complex and on some real hereditarily
indecomposable Banach spaces, which provides a negative answer to the above question.
MSC: 47A16, 37A25
Keywords: Frequently hypercyclic operators, hereditarily indecomposable Banach spaces, quasinilpotent
operators
1
Introduction
All vector spaces in this article are assumed to be over the field K being either the field C of complex
numbers or the field R of real numbers. As usual, T = {z ∈ C : z = 1}, Z is the set of integers,
Z+ is the set of non-negative integers and N is the set of positive integers. For a Banach space X,
L(X) stands for the algebra of bounded linear operators on X and X ∗ is the space of continuous
linear functionals on X.
Recall that a continuous linear operator T on a topological vector space X is called hypercyclic
if there is x ∈ X such that the orbit {T nx : x ∈ Z+} is dense in X. We refer to surveys [17, 18]
for additional information on hypercyclicity. There are several stronger versions of the above
property of linear operators like hereditarily hypercyclic operators, topologically mixing operators
and operators satisfying the Kitai Criterion. Recently Bayart and Grivaux [5] have introduced
the concept of frequently hypercyclic operators. Recall that the upper and lower densities of a set
A ⊂ Z+ are defined by the formula
d(A) = lim
n→∞
{m ∈ A : m 6 n}
n
,
d(A) = lim
n→∞
{m ∈ A : m 6 n}
n
,
where B stands for the number of elements of a finite set B.
Definition 1. Let T be a continuous linear operator on a topological vector space X. Then x ∈ X
is called a frequently hypercyclic vector for T if the lower density of the set {n ∈ N : T nx ∈ U } is
positive for any non-empty open subset U of X. An operator T is called frequently hypercyclic if it
has a frequently hypercyclic vector.
We also say that x ∈ X is an u-frequently hypercyclic vector for T if the upper density of the
set {n ∈ N : T nx ∈ U } is positive for any non-empty open subset U of X. An operator T is
called u-frequently hypercyclic if it has an u-frequently hypercyclic vector. We denote the set of
u-frequently hypercyclic vectors for T by the symbol u(T ).
Partially supported by Plan Nacional I+D+I grant no. MTM2006-09060 and Junta de Andaluc´ıa FQM-260.
1
Clearly frequent hypercyclicity implies u-frequent hypercyclicity, which, in turn, implies hyper-
cyclicity. Bayart and Grivaux [5] have raised the following question.
Question 1. Does every separable infinite dimensional Banach space support a frequently hyper-
cyclic operator?
This question is partially motivated by the following chain of results. Ansari [1] and Bernal-
Gonz´ales [6] showed independently that for any separable infinite dimensional Banach space X
there is a hypercyclic operator T ∈ L(X). Grivaux [15] proved that there is a mixing operator on
any separable infinite dimensional Banach space. Finally, in [25] it is shown that on any separable
infinite dimensional Banach space, there is a bounded linear operator, satisfying the Kitai Criterion.
In this article we show that unlike for other strong forms of hypercyclicity, the answer to the above
question is negative.
Theorem 1.1. Let T be a bounded linear operator on a separable infinite dimensional Banach
kT nk1/n = 0. Then I + T is not u-frequently hypercyclic. In particular,
space X such that
I + T is not frequently hypercyclic.
lim
n→∞
In the case K = C, we use the above theorem to prove the following result.
Theorem 1.2. Let X be a complex Banach space and T ∈ L(X) be frequently hypercyclic. Then
the spectrum σ(T ) of T has no isolated points.
This result shows that the class of frequently hypercyclic operators is significantly smaller than
the class of hypercyclic operators. For instance, in [16] it is demonstrated that if X is a separable
Banach space and M is the operator norm closure of the set of finite rank nilpotent operators, then
the set {T ∈ M : I + T is hypercyclic} is a dense Gδ subset of M . Thus, in a sense, a 'generic'
operator with one-point spectrum {1} is hypercyclic.
We use Theorem 1.1 to obtain a negative answer to Question 1. It turns out that there are
no frequently hypercyclic operators on Banach spaces with few operators. Recall that a bounded
linear operator T on a Banach space X is called strictly singular if the restriction of T to any closed
infinite dimensional subspace of X is not an isomorphism onto its image. As is well-known [24],
the set S(X) of strictly singular operators is an ideal in the algebra L(X).
By ΦK we denote the class of infinite dimensional Banach spaces X over the field K for which
It is straightforward to see that X ∈ ΦK if and only if any
L(X)/S(X) is isomorphic to K.
T ∈ L(X) has the shape T = zI + S for z ∈ K and S ∈ S(X).
Theorem 1.3. Let X ∈ ΦK. Then there is no bounded u-frequently hypercyclic operator on X. In
particular, X does not support a frequently hypercyclic operator.
The above theorem gives a negative answer to Question 1 provided there are separable Banach
spaces in classes ΦC and ΦR. Recall that an infinite dimensional Banach X space is called hered-
itarily indecomposable or an HI space if for any infinite dimensional closed linear subspaces L and
M of X satisfying L ∩ M = {0} their sum L + M is not closed. Separable HI Banach spaces were
first constructed by Gowers and Maurey [14], see also the survey [4]. One of the most interesting
features of HI spaces X is [14, 12] that any complex HI space belongs to ΦC. It is also worth noting
that there are separable Banach spaces in ΦC that are not HI [3]. Anyway, we have the following
corollary.
Corollary 1.4. Let X be a separable complex HI space. Then there is no bounded u-frequently
hypercyclic operator on X. In particular, X does not support a frequently hypercyclic operator.
The case of real HI spaces is slightly more complicated. Namely, Ferenczi [12, 13] demonstrated
that if X is a real HI space, then L(X)/S(X) is isomorphic to either R, or C or the quaternion
2
division ring H. Moreover, in [13] it is shown that all three possibilities do occur on the level of
separable real HI spaces and is observed that the original real separable HI space X of Gowers and
Maurey [14] satisfies L(X)/S(X) ≃ R and therefore belongs to ΦR. This remark together with
Theorem 1.3 gives us the following corollary.
Corollary 1.5. There is an infinite dimensional separable real Banach space X, which does not
support a u-frequently hypercyclic operator. In particular, X does not support a frequently hyper-
cyclic operator.
2 Proof of Theorem 1.1
Theorem 2.1. Let T be a continuous linear operator on a real topological vector space X and
x ∈ X be such that there exists f ∈ X ∗ \ {0} satisfying
f ((T − I)nx)1/n = 0
for each x ∈ X.
lim
n→∞
(2.1)
Then x is not a u-frequently hypercyclic vector for T .
Proof. Assume that x ∈ u(T ). First, observe, that we can find u ∈ X such that f (u) < 0 and
f (T u) > 0. Indeed, otherwise, using continuity of f and T , we see that f (T u) 6 0 if f (u) 6 0.
Since f (−y) = −f (y) for any y ∈ X, we also have f (T u) > 0 if f (u) > 0. Hence T (H) ⊆ H, where
H = ker f . Thus, T has a closed invariant subspace of codimension 1. On the other hand, in [26]
it is proven that any hypercyclic operator on any real or complex topological vector space has no
closed invariant subspaces of finite codimension. This contradiction shows that there is u ∈ X such
that f (u) < 0 and f (T u) > 0. Since f and T are continuous, we can choose an open neighborhood
U of u such that
Since x ∈ u(T ), there exists a subset A of Z+ of positive upper density such that
f (v) < 0 and f (T v) > 0 for any v ∈ U .
f (T nx) ∈ U for any n ∈ A.
(2.2)
(2.3)
Consider now the series
F (z) = f (x) +
∞
X
n=1
f ((T − I)nx)
n!
z(z − 1) . . . (z − n + 1),
z ∈ C.
(2.4)
Condition (2.1) implies that the series in (2.4) is uniformly convergent on any compact subset of
C and therefore defines an entire function F . Moreover, F has exponential type 0, that is,
lim
R→∞
ln MF (R)
R
= 0, where MF (R) = max
z6R
F (z).
(2.5)
Indeed, by (2.1), for any ε ∈ (0, 1), there is m ∈ N such that f ((T − I)nx) < εn for n > m. Then
MF (R) 6 f (x) +
f ((T − I)nx)
n!
R(R + 1) . . . (R + n − 1) 6
∞
X
n=1
εn
n!
6 1 +
∞
X
n=1
R(R + 1) . . . (R + n − 1) + O(Rm) = (1 − ε)−R + O(Rm) =
= ea(ε)R + O(Rm) = O(ea(ε)R) as R → ∞, where a(ε) = − ln(1 − ε).
3
Since a(ε) → 0 as ε → 0, (2.5) follows. From the definition of F it also easily follows that
F (k) = f (T kx) for any k ∈ Z+.
(2.6)
Using (2.6), (2.3) and (2.2), we see that F (n) < 0 and F (n + 1) > 0 for any n ∈ A. Since F is real
on the real axis, for any n ∈ A, there is tn ∈ (n, n + 1) such that F (tn) = 0. Let now NF (R) be the
number of zeros of F in the set {z ∈ C : 0 < z < R}. The last observation yields NF (n + 1) > kn
for any n ∈ Z+, where kn = {m ∈ A : m 6 n} is the counting function of the set A. On the other
hand, by a corollary of Jensen's Theorem, see [23, p. 15], we have
NF (R) 6 ln MF (eR) + O(ln R) as R → ∞.
(2.7)
Hence
lim
n→∞
ln MF (e(n + 1))
n
> lim
n→∞
NF (n + 1)
n
> lim
n→∞
kn
n
= d(A) > 0,
which contradicts (2.5).
Corollary 2.2. Let X be a Banach space and T ∈ L(X) be such that there exists f ∈ X ∗ \ {0} for
which lim
n→∞
kT ∗nf k1/n = 0. Then I + T is not u-frequently hypercyclic.
Proof. If K = R, we just apply Theorem 2.1 to the operator I + T to see that the latter has no
u-frequently hypercyclic vectors. If K = C, we consider X as a real Banach space, T as an R-linear
operator and apply Theorem 2.1 to the operator I + T and the functional Re f .
Theorem 1.1 follows immediately from Corollary 2.2 since kT ∗nf k 6 kf kkT ∗nk = kf kkT nk for
any f ∈ X ∗ and any n ∈ Z+.
3 Rotations and powers of u-frequently hypercyclic operators
In order to reduce Theorem 1.3 and Theorem 1.2 to Theorem 1.1, we need to show that the
statement of Theorem 1.1 remains true if we replace I + T be zI + T for any z ∈ K with z = 1.
In order to do so, we need to demonstrate that the class of u-frequently hypercyclic operators is
stable under multiplication by scalars z with z = 1.
For hypercyclic operators acting on complex Banach spaces, this kind of stability was proved
by Le´on-Saavedra and Muller [22]. In the case K = R the fact that T is hypercyclic if and only
if −T is hypercyclic follows from the result of Ansari [2], who showed that hypercyclicity of T
implies hypercyclicity of its powers. In particular T 2 is hypercyclic if T is. Since (−T )2 = T 2,
the latter observation shows that T is hypercyclic if and only if −T is. We are going to prove
the same stability results for u-frequently hypercyclic operators. Concerning the powers, we shall
use a result of Bourdon and Feldman [10], who demonstrated that an orbit of a continuous linear
operator on a locally convex topological vector space is either everywhere dense or nowhere dense.
The only obstacle that did not allow the same construction to work for arbitrary topological vector
spaces was that the proof of the density of the range of p(T ) for any non-zero polynomial p and
a hypercyclic operator T used the duality argument. This obstacle was removed by Wengenroth
[26], which allowed him to extend the Bourdon and Feldman theorem to arbitrary topological
vector spaces, see also the survey [18]. Throughout this section X is a topological vector space and
T ∈ L(X).
Theorem A. Let x ∈ X be such that F T
x = X.
Although the above mentioned result of Le´on-Saavedra and Muller is formulated and proved
for operators acting on Banach spaces, the same proof with minor modifications (just replace
convergent sequences by convergent nets and use the result of Wengenroth to ensure density of the
ranges of polynomials of T instead of the duality argument) gives the following extension of their
result.
x = {T nx : n ∈ Z+} has nonempty interior. Then F T
4
Theorem B. Let z ∈ K, z = 1. Then the sets of hypercyclic vectors for T and zT coincide.
For shortness, we denote by Λ the set of all subsets of Z+ of positive upper density. We also fix
a base U of open neighborhoods of 0 in X.
Definition 2. For x ∈ X we denote
M T
x = (cid:8)y ∈ X : {n ∈ Z+ : y − T nx ∈ U } ∈ Λ for any U ∈ U(cid:9).
Clearly, x ∈ u(T ) if and only if M T
x = X.
If additionally X is complex, we denote
N T
x = (cid:8)y ∈ X : {n ∈ Z+ : Cn(T, x) ∩ (y + U ) 6= ∅} ∈ Λ for any U ∈ U(cid:9),
where Cn(T, x) = {zT nx : z ∈ T}.
(3.1)
(3.2)
We start by establishing few straightforward properties of the sets M T
x and N T
x defined in (3.1)
and (3.2). The following lemma is an elementary exercise. We leave the proof to the reader.
Lemma 3.1. For any x ∈ X, M T
x is closed in X, T (M T
each z ∈ K. If X is complex, then for any x ∈ X, N T
and N T
x for any z ∈ T.
zx = N zT
x = N T
x ) ⊆ M T
T x = M T
x is closed in X, T (N T
x , M T
x and M T
x ) ⊆ N T
zx = zM T
x , N T
x for
T x = N T
x
Lemma 3.2. If u, x, y ∈ X are such that y ∈ M T
x and x ∈ M T
u , then y ∈ M T
u .
Proof. Let U ∈ U . Pick V ∈ U such that V + V ⊂ U . Since y ∈ M T
that y − T mx ∈ V . Since x ∈ M T
any k ∈ A, we have
x , we can find m ∈ Z+ such
u , we have A = {k ∈ Z+ : x − T ku ∈ (T m)−1(V )} ∈ Λ. Then for
y − T k+mu = y − T mx + T m(x − T ku) ∈ V + T m((T m)−1(V )) ⊆ V + V ⊆ U.
Thus y − T nu ∈ U for any n ∈ m + A. Since m + A ∈ Λ, we have y ∈ M T
u .
Lemma 3.3. If X is complex and x ∈ N T
x , then x ∈ M T
x .
Proof. Clearly the set F = {µ ∈ T : µx ∈ M T
non-empty. For each U ∈ U consider
x } is closed since M T
x is. Let us verify that F is
F (U ) = (cid:8)z ∈ T : {n ∈ Z+ : zx − T nx ∈ U } ∈ Λ(cid:9)
and F (U ) be the closure of F (U ) in T. Clearly F = T{F (U ) : U ∈ U }. Let us show that the
sets F (U ) are non-empty. Fix U ∈ U . Choose V ∈ U and ε > 0 such that V + Dεx ⊆ U , where
Dε = {z ∈ C : z < ε}. Since x ∈ N T
x , there exist A ∈ Λ and a sequence {zm}m∈A of elements of T
such that zmx − T mx ∈ V for any m ∈ A. Let k ∈ N be such that k > ε−1 and let wj = e2πij/k for
0 6 j 6 k − 1. Then for any m ∈ A we can pick ν(m) ∈ {0, . . . , k − 1} such that zm − wν(m) < ε.
Clearly A is the union of Aj = {m ∈ A : ν(m) = j} for 0 6 j 6 k − 1. Since the union of Aj has
positive upper density, there exists j ∈ {0, . . . , k − 1} such that Aj ∈ Λ. For each m ∈ Aj, we have
wjx − T mx = zmx − T mx + (wj − zm)x ∈ V + Dεx ⊂ U.
Hence wj ∈ F (U ) and F (U ) is non-empty. From the definition of the set F (U ), we also have that
F (V ) ⊂ JεF (V ) ⊆ F (U ), where Jε = {eit : −ε 6 t 6 ε}. Thus for each U ∈ U there is V ∈ U such
that F (V ) ⊂ F (U ). It follows that
F = \
U ∈U
F (U ) = \
U ∈U
F (U ).
5
(3.3)
Finally, for any U1, . . . , Un ∈ U , we have
∅ 6= β(U1 ∩ . . . ∩ Un) ⊆ F (U1) ∩ . . . ∩ F (Un) ⊆ F (U1) ∩ . . . ∩ F (Un).
Thus {F (U ) : U ∈ U } is a family of closed subsets of the compact topological space T, any finite
subfamily of which has nonempty intersection. Hence it has non-empty intersection and (3.3)
implies that F 6= ∅.
Thus there is z ∈ T such that zx ∈ M T
Lemma 3.2, the inclusions zx ∈ M T
have {znx : n ∈ N} ⊂ M T
1 ∈ Qz for any z ∈ T, we see that 1 ∈ F . That is, x ∈ M T
x .
zx. According to
x . Proceeding inductively, we
x . That is, Qz = {zn : n ∈ N} ⊆ F . Since F is closed Qz ⊆ F . Since
x . By Lemma 3.1, then z2x ∈ M T
zx imply z2x ∈ M T
x and z2x ∈ M T
Lemma 3.4. If x ∈ X is such that x ∈ M T
x , then M T
x = F T
x , where F T
x = {T nx : n ∈ Z+}.
Proof. Clearly M T
Since M T
x is closed, we have F T
x ⊆ M T
x . Thus, F T
x = M T
x .
x ⊆ F T
x . Since x ∈ M T
x and T (M T
x ) ⊆ M T
x , we see that {T nx : n ∈ Z+} ⊆ M T
x .
Lemma 3.5. If x ∈ X is such that M T
y ∈ M T
y = M T
x .
x has non-empty interior, then there exists y ∈ X such that
Proof. Since M T
In particular, y = T kx ∈ M T
x has non-empty interior, there is k ∈ Z+ such that T kx is an interior point of M T
x .
x . By Lemma 3.1, M T
y = M T
T kx = M T
x . Thus y ∈ M T
x = M T
y .
3.1 Rotations
Proposition 3.6. Let X be a complex topological vector space, z ∈ T and T ∈ L(X). Then T is
u-frequently hypercyclic if and only if zT is u-frequently hypercyclic and u(T ) = u(zT ).
x = X, where the sets M T
x = N T
Proof. Let x ∈ u(T ). Then M T
x = N T
x are defined in (3.1) and
(3.2). On the other hand, clearly N zT
x . By
Lemma 3.3, x ∈ M zT
is the closure of {(zT )nx :
n ∈ Z+}. Since x is a hypercyclic vector for T , Theorem B implies that x is a hypercyclic
vector for zT and therefore F zT
x = X and x ∈ u(zT ). Thus u(T ) ⊆ u(zT ).
Applying this inclusion to T replaced by zT and z replaced by z−1, we obtain u(zT ) ⊆ u(T ). Thus
u(T ) = u(zT ).
x . Then by Lemma 3.4, M zT
x = X. Hence M zT
In particular, x ∈ N zT
x . Hence N zT
x , where F zT
x
x and N T
x = F zT
x = X.
3.2 Powers
Proposition 3.7. Let X be a topological vector space and T ∈ L(X). If x ∈ X is such that the set
M T
x , defined in (3.1), has non-empty interior, then x ∈ u(T ).
Proof. By Lemma 3.5, we can pick y ∈ X such that y ∈ M T
where F T
y
implies that F T
is the closure of {T ny : n ∈ Z+}. Thus, F T
y = X. Since F T
x , we have M T
y = M T
y = M T
y = M T
y = F T
y ,
y has non-empty interior and Theorem A
x . By Lemma 3.4, M T
x = X. That is, x ∈ u(T ).
Corollary 3.8. Let X be a topological vector space, T ∈ L(X) be an u-frequently hypercyclic
operator and n ∈ N. Then T n is also u-frequently hypercyclic. Moreover u(T ) = u(T n).
Proof. Let x ∈ u(T ). Then the set M T
x defined in (3.1) coincides with X. Since the union of finitely
many subsets of Z+ has positive upper density if and only if at least one of them has positive upper
density, one can easily see that
X = M T
x =
n−1
[
j=0
M T n
T jx.
6
T j x are closed. Hence there is j ∈ {0, . . . , n − 1} such that M T n
By Lemma 3.1, the sets M T n
T jx has
non-empty interior. By Proposition 3.7, M T n
T j x = X. That is, T jx ∈ u(T n). Exactly as for usual
hypercyclicity, one can easily verify that if y is an u-frequently hypercyclic vector for S ∈ L(X)
and R ∈ L(X) has dense range and commutes with S, then Ry ∈ u(S). Applying this observation
with S = T n, y = T jx and R = T n−j, we see that T nx ∈ u(T n). Since x and T nx have the
same orbits with respect to T n (up to one element added or removed), we obtain x ∈ u(T n). Thus
u(T ) ⊆ u(T n). The inclusion u(T n) ⊆ u(T ) is obvious.
Corollary 3.9. Let X be a real topological vector space and T ∈ L(X). Then T is u-frequently
hypercyclic if and only if −T is u-frequently hypercyclic and u(T ) = u(−T ).
Proof. By Corollary 3.8, u(T ) = u(T 2) = u((−T )2) = u(−T ).
4 Proof of Theorems 1.3 and 1.2
The above results allow us to strengthen Theorem 1.1.
Corollary 4.1. Let X be a separable infinite dimensional Banach space, z ∈ K and T ∈ L(X) be
such that z = 1 and there exists f ∈ X ∗ \ {0} for which lim
kT ∗nf k1/n = 0. Then zI + T is not
n→∞
u-frequently hypercyclic.
Proof. By Corollary 2.2, I +z−1T is not u-frequently hypercyclic. Using Proposition 3.6 in the case
K = C and Corollary 3.9 in the case K = R, we see that z(I + z−1T ) = zI + T is not u-frequently
hypercyclic.
Since kT ∗nf k 6 kf kkT ∗nk = kf kkT nk for any f ∈ X ∗ and any n ∈ Z+, we immediately obtain
the following slightly stronger form of Theorem 1.1.
Theorem 4.2. Let T be a bounded linear operator on a separable infinite dimensional Banach space
kT nk1/n = 0 and z ∈ K be such that z = 1. Then zI + T is not u-frequently
X such that
hypercyclic. In particular, zI + T is not frequently hypercyclic.
lim
n→∞
In order to prove Theorems 1.3 and 1.2, we need the following lemmas.
Lemma 4.3. Let X ∈ ΦK and T ∈ L(X). Assume also that T has no non-trivial closed invariant
subspaces of finite codimension. Then T = zI + S, where z ∈ K and lim
n→∞
kSnk1/n = 0.
Proof. Since X ∈ ΦK, there are z ∈ K and S ∈ S(X) such that T = zI + S. Let XC = X and
RC = R for any R ∈ L(X) if K = C and XC be the complexification of X and RC ∈ L(XC) be the
complexification (=the unique complex linear extension) of R ∈ L(X) if K = R.
C of TC is empty.
In any case TC = zI + SC and SC ∈ S(X). The fact that T has no non-trivial closed linear
subspaces of finite codimension is clearly equivalent to the statement that the point spectrum
σp(T ∗
C) of the dual T ∗
Assume that w ∈ σ(TC) \ {z}. Then w − z ∈ σ(SC) \ {0}. Since a non-zero element of the
spectrum of any strictly singular operator on a complex Banach space is a normal eigenvalue [24],
we see that w − z is a normal eigenvalue of SC and therefore w − z ∈ σp(S∗
C),
which contradicts the equality σp(T ∗
C) = ∅. Thus σ(TC) = {z}. Hence σ(SC) = {0}. By the
Ck1/n = 0. Since S is the restriction of SC to an R-linear invariant
spectral radius formula lim
n→∞
subspace, kSnk 6 kSn
C). Hence w ∈ σp(T ∗
kSn
Ck for any n ∈ N and therefore lim
n→∞
kSnk1/n = 0.
Lemma 4.4. Let X ∈ ΦK and T ∈ L(X) be a hypercyclic operator. Then T = zI + S, where
z ∈ K, z = 1 and lim
n→∞
kSnk1/n = 0.
7
Proof. As it is shown in [26], a hypercyclic operator on a real or complex topological vector space has
no non-trivial closed invariant subspaces of finite codimension. Hence, by Lemma 4.3, T = zI + S,
kSnk1/n = 0. It remains to demonstrate that z = 1. Using the equalities
where z ∈ K and lim
n→∞
T = zI + S and lim
n→∞
kSnk1/n = 0, one can easily verify that
kT nxk1/n = z
for any x ∈ X \ {0}.
lim
n→∞
(4.1)
Now if z > 1, then according to (4.1), kT nxk → ∞ as n → ∞ for each non-zero x ∈ X.
Similarly, if z < 1, then kT nxk → 0 as n → ∞ for any x ∈ X. Hence T can not be hypercyclic if
z 6= 1. Thus z = 1.
Theorem 1.3 follows immediately from Lemma 4.4 and Theorem 4.2.
Proof of Theorem 1.2. Let T be an u-frequently hypercyclic operator on a complex Banach space
X. Assume also that z is an isolated point of the spectrum σ(T ). Let x ∈ u(T ), P be the spectral
projection corresponding to the component {z} of the spectrum of T , Y = P (X) and S ∈ L(Y ) be
the restriction of T to the closed invariant subspace Y . Clearly σ(S) = {z}. Since SnP x = P T nx
for each n ∈ Z+, we see that P x ∈ u(S). Thus, S is an u-frequently hypercyclic operator on Y
satisfying σ(S) = {z}. If z < 1, then kSnyk → 0, while kSnyk → ∞ if z > 1 for any y ∈ Y
and S can not be hypercyclic. Hence z ∈ T. Since σ(S) = {z}, we have S = zI + R, where
lim
n→∞
kRnk1/n = 0. We have obtained a contradiction with Theorem 4.2.
5 Concluding remarks
In the proofs of Proposition 3.6 and Corollary 3.8, we use the following property of the family
1.
Λ of subsets of Z+ of positive upper density. Namely, if a finite union of sets belongs to Λ, then one
of them does. The family Λ0 of subsets of Z+ of positive lower density fails to have this property.
Thus our proof of stability of the class of u-frequently hypercyclic operators under powers and
under multiplication by unimodular scalars does not work for the class of frequently hypercyclic
operators. This leads to the following question.
Question 2. Let T be a frequently hypercyclic operator on a separable infinite dimensional
Banach space X. Is it true that the operators T n and zT for n ∈ N and z ∈ K with z = 1 are
frequently hypercyclic?
2. Proposition 3.6, Corollary 3.8 and Corollary 3.9 admit the following generalization. Let Ω be
any family of non-empty subsets of Z+ satisfying the following conditions
if A ∈ Ω and A ⊆ B ⊆ Z+, then B ∈ Ω;
if A ∈ Ω and k ∈ Z+, then A + k ∈ Ω and (A − k) ∩ Z+ ∈ Ω;
k
S
if k ∈ N and Aj for 1 6 j 6 k are subsets of Z+ such that
j=1
then there exists j ∈ {1, . . . , k} such that Aj ∈ Ω.
Aj ∈ Ω,
We say that x is an Ω-hypercyclic vector for a continuous linear operator T on a topological vector
space X if {n ∈ Z+ : T nx ∈ U } ∈ Ω for each non-empty open subset U of X and that T is
Ω-hypercyclic if it has an Ω-hypercyclic vector.
All proofs in Section 3 work without any changes (just replace Λ by Ω) for Ω-hypercyclicity
instead of u-frequent hypercyclicity. Thus, the class of Ω-hypercyclic operators is stable under
taking the powers and under multiplying by unimodular scalars. Note also that when Ω is the
set of all infinite subsets of Z+, then Ω-hypercyclicity coincides with the usual hypercyclicity.
The following example seems to be interesting. Let Ψ be the set of all decreasing sequences
8
ϕ = {ϕn}n∈Z+ of positive numbers such that
of subsets A of Z+ for which P
n∈A
easy to verify that an operator T is u-frequently hypercyclic if and only if it is Ωϕ-hypercyclic for
any ϕ ∈ Ψ. This naturally leads to the following question.
ϕn = ∞ satisfies all properties from the above display. It is also
∞
P
n=0
ϕn = ∞. Clearly, for each ϕ ∈ Ψ, the family Ωϕ
Question 3. For which ϕ ∈ Ψ, there is an Ωϕ-hypercyclic operator on any separable infinite
dimensional Banach space?
3. One can observe that all operators T proven to be frequently hypercyclic have a stronger
property. Namely they admit a vector x such that for any non-empty open set U the set {n ∈
Z+ : T nx ∈ U } contains a subset, which has positive density. Since there are sets A ⊂ Z+
of positive lower density which contain no subsets of positive density, the last property of x is,
formally speaking, stronger than to be a frequently hypercyclic vector. It might be interesting to
find out whether all frequently hypercyclic operators admit vectors having this stronger property.
4. Herrero [20] have characterized the operator norm closure HC(H) of the set HC(H) of hypercyclic
operators on a separable infinite dimensional complex Hilbert space H. Later Herrero and Wang
[21] have shown that any T ∈ HC(H) is a limit of a sequence Tn of elements of HC(H) such that the
differences T − Tn are compact. Using these results, one can easily prove the following proposition.
Proposition 5.1. A nonempty compact subset K of C is the spectrum of some bounded hypercyclic
operator T on a separable infinite dimensional complex Hilbert space H if and only if K ∪ T is
connected.
Proof. The 'only if' part is proved in [20]. Let K be a non-empty compact subset of C such that
K ∪ T is connected. We have to show that there is a hypercyclic operator T ∈ L(H) such that
σ(T ) = K. Let A be the set of isolated points of K. Exactly as in the proof of Theorem 1.2, one
can easily see that A ⊂ T. Clearly A is at most countable. Let S ∈ L(ℓ2) be a compact weighted
backward shift: Se0 = 0 and Sen = wnen−1 for n ∈ N, where w = {wn}n∈N is a sequence of
non-zero complex numbers tending to 0 as n → ∞. Consider the ℓ2-direct sum T0 of zI + S for
z ∈ A:
T0 = M
z∈A
(zI + S).
Then σ(T0) = A. From [16][Theorem 5.1] if follows that T0 is topologically mixing. Next, we take a
bounded normal operator S1 on a separable complex Hilbert space H1 such that σ(S1) = K \ A and
S1 has empty point spectrum. According to the mentioned results of Herrero [20] and Herrero and
Wang [21], we have that S1 ∈ HC(H1) and therefore there exists a compact operator K ∈ L(H1)
such that T1 = S1 + K is hypercyclic. For each z ∈ C \ σ(S1) = C \ (K \ A), the operator
S1 − zI is invertible and therefore by the Fredholm theorem, either T1 − zI is invertible or T ∗
1 − zI
has non-trivial kernel (here T ∗
1 is the dual operator, not the Hilbert space adjoint). The latter
case is not possible since the dual of any hypercyclic operator has empty point spectrum. Thus
σ(T1) ⊆ σ(S1). One the other hand, S1 has purely continuous spectrum and T1 is a compact
perturbation of S1. Hence σ(S1) ⊆ σ(T1). Thus σ(T1) = K \ A. Let T = T0 ⊕ T1. Then
σ(T ) = σ(T0) ∪ σ(T1) = A ∪ (K \ A) = K. Moreover T is hypercyclic as a direct sum of a
hypercyclic operator with a mixing operator.
Question 4. What compact subsets of C are the spectra of frequently hypercyclic operators
on ℓ2? What about u-frequently hypercyclic operators?
Acknowledgements. The author would like to thank the referee for helpful comments and
corrections.
9
References
[1] S. Ansari, Existence of hypercyclic operators on topological vector spaces, J. Funct. Anal. 148 (1997),
384 -- 390
[2] S. Ansari, Hypercyclic and cyclic vectors, J. Funct. Anal. 128 (1995), 374 -- 383
[3] S. Argyros and A. Manoussakis, An indecomposable and unconditionally saturated Banach space, Studia
Math. 159 (2003), 1 -- 32
[4] S. Argyros and A. Tolias, Methods in the theory of hereditarily indecomposable Banach spaces, Mem.
Amer. Math. Soc. 170 (2004), 1 -- 114
[5] F. Bayart and S. Grivaux, Frequently hypercyclic operators, Trans. Amer. Math. Soc. 358 (2006), 5083 --
5117
[6] L. Bernal-Gonz´alez, On hypercyclic operators on Banach spaces, Proc. Amer. Math. Soc. 127 (1999),
1003 -- 1010
[7] J. B´es and A. Peris, Hereditarily hypercyclic operators, J. Funct. Anal. 167 (1999), 94 -- 112
[8] J. Bonet and A. Peris, Hypercyclic operators on non-normable Fr´echet spaces, J. Funct. Anal. 159
(1998), 587 -- 595
[9] A. Bonilla and , K.-G. Grosse-Erdmann, On a theorem of Godefroy and Shapiro, Integral Equations
Operator Theory 56 (2006), 151 -- 162
[10] P. Bourdon and N. Feldman, Somewhere dense orbits are everywhere dense, Indiana Univ. Math. J. 52
(2003), 811 -- 819
[11] M. De La Rosa and C. Read, A hypercyclic operator whose direct sum is not hypercyclic, [preprint]
[12] V. Ferenczi, Operators on subspaces of hereditarily indecomposable Banach spaces, Bull. London Math.
Soc. 29 (1997), 338 -- 344
[13] V. Ferenczi, Uniqueness of complex structure and real hereditarily indecomposable Banach spaces, Ad-
vances Math. 213 (2007), 462 -- 488
[14] W. Gowers and B. Maurey, The unconditional basic sequence problem, J. Amer. Math. Soc. 6 (1993),
851 -- 874
[15] S. Grivaux, Hypercyclic operators, mixing operators and the bounded steps problem, J. Operator Theory
54 (2005), 147 -- 168
[16] S. Grivaux and S. Shkarin, Non-mixing hypercyclic operators, [preprint]
[17] K. Grosse-Erdmann, Universal families and hypercyclic operators, Bull. Amer. Math. Soc. 36 (1999),
345 -- 381
[18] K. Grosse-Erdmann, Recent developments in hypercyclicity, RACSAM Rev. R. Acad. Cienc. Exactas
Fis. Nat. Ser. A Mat. 97 (2003), 273 -- 286
[19] K.-G. Grosse-Erdmann and A. Peris, Frequently dense orbits, C. R. Math. Acad. Sci. Paris 341 (2005),
123 -- 128
[20] D. Herrero, Limits of hypercyclic and supercyclic operators, J. Funct. Anal. 99 (1991), 179 -- 190.
[21] D. Herrero and Z. Wang, Compact perturbations of hypercyclic and supercyclic operators, Indiana Univ.
Math. J. 39 (1990), 819 -- 829
[22] F. Le´on-Saavedra and Vladimir Muller, Rotations of hypercyclic and supercyclic operators, Integral
Equations and Operator Theory 50 (2004), 385 -- 391
[23] B. Levin, Distribution of zeros of entire functions, AMS Providence, Rhode Island, 1980
[24] A. Pietsch, Operator ideals, North-Holland, Amsterdam, 1980
[25] S. Shkarin, The Kitai Criterion and backward shifts, Proc. Amer. Math. Soc. [to appear]
[26] J. Wengenroth, Hypercyclic operators on non-locally convex spaces, Proc. Amer. Math. Soc. 131 (2003),
1759 -- 1761
Stanislav Shkarin
Queens's University Belfast
Department of Pure Mathematics
University road, Belfast, BT7 1NN, UK
E-mail address:
[email protected]
10
|
1610.04257 | 1 | 1610 | 2016-10-13T20:40:25 | Weakly Radon-Nikod\'ym Boolean algebras and independent sequences | [
"math.FA"
] | A compact space is said to be weakly Radon-Nikod\'{y}m (WRN) if it can be weak*-embedded into the dual of a Banach space not containing $\ell_1$. We investigate WRN Boolean algebras, i.e. algebras whose Stone space is WRN compact. We show that the class of WRN algebras and the class of minimally generated algebras are incomparable. In particular, we construct a minimally generated nonWRN Boolean algebra whose Stone space is a separable Rosenthal compactum, answering in this way a question of W. Marciszewski.
We also study questions of J. Rodr\'{i}guez and R. Haydon concerning measures and the existence of nontrivial convergent sequences on WRN compacta, obtaining partial results on some natural subclasses. | math.FA | math |
WEAKLY RADON-NIKODÝM BOOLEAN ALGEBRAS AND
INDEPENDENT SEQUENCES
ANTONIO AVILÉS, GONZALO MARTÍNEZ-CERVANTES, AND GRZEGORZ PLEBANEK
Abstract. A compact space is said to be weakly Radon-Nikodým (WRN) if it can be
weak*-embedded into the dual of a Banach space not containing ℓ1. We investigate WRN
Boolean algebras, i.e. algebras whose Stone space is WRN compact. We show that the
class of WRN algebras and the class of minimally generated algebras are incomparable.
In particular, we construct a minimally generated nonWRN Boolean algebra whose Stone
space is a separable Rosenthal compactum, answering in this way a question of W. Mar-
ciszewski.
We also study questions of J. Rodríguez and R. Haydon concerning measures and the
existence of nontrivial convergent sequences on WRN compacta, obtaining partial results
on some natural subclasses.
1. Introduction
A compact space is said to be Radon-Nikodým (RN) if it is homeomorphic to a weak∗-
compact subset of a dual Banach space with the Radon-Nikodým property. C. Stegall
proved that the dual X ∗ of a Banach space X has the Radon-Nikodým property if and
only if X is Asplund, i.e. every separable subspace of X has separable dual. The class
of RN compacta has been widely studied since it was introduced by I. Namioka in [18].
Some remarkable results concerning this class and its relation with other classes of compact
spaces such as Corson compacta and Eberlein compacta can be found in [19], [1], [23] and
[2]. In a similar way, E. Glasner and M. Megrelishvili defined in [8] the class of weakly
Radon-Nikodým compact spaces (WRN). These are compact spaces which can be weak∗
embedded into the dual of a Banach space not containing ℓ1. Every Eberlein compact
space is RN and every RN compact space is WRN. Glasner and Megrelishvili proved that
linearly ordered compact spaces are WRN. Thus, the double arrow is an example of a WRN
compact space which is not RN (see [18, Example 5.9]). On the other hand, it follows from
a result of Talagrand [24] that βω is not WRN (another proof of S. Todorcevic is included
2010 Mathematics Subject Classification. Primary 03G05,O6E15,28A60; Secondary 46B22, 46B50.
Key words and phrases. Boolean algebra, Stone space, independent sequence, weakly Radon-Nikodým
compact.
A. Avilés and G. Martínez-Cervantes were partially supported by the research project 19275/PI/14
funded by Fundación Séneca - Agencia de Ciencia y Tecnología de la Región de Murcia within the framework
of PCTIRM 2011-2014 and by Ministerio de Economía y Competitividad and FEDER (project MTM2014-
54182-P).
G. Plebanek was partially supported by NCN grant 2013/11/B/ST1/03596 (2014-2017).
1
2
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
in [9, Appendix]; cf. Corollary 3.4 below). Some other results about this class can be found
in [17].
In this paper we introduce a notion of a weakly Radon-Nikodým Boolean algebra (WRN)
-- an algebra A is WRN if its Stone space is weakly Radon-Nikodým compact. In fact
our basic definition is given in a purely Boolean language and says that A is WRN if A is
generated by a family that may be decomposed into countably many pieces, of which none
contains an infinite independent sequence (see Definition 3.1).
WRN algebras are, in a sense, small and share some properties with the class of minimally
generated algebras introduced by Koppelberg [15]. We show, however, that there is a WRN
algebra which is not minimally generated, see Example 4.4. Our example of a minimally
generated algebra which is not WRN is more involved, see section 5. The algebra A we
construct here has the additional property that its Stone space K is separable Rosenthal
compact. In this way we answer a question communicated to us by W. Marciszewski, see
Theorem 5.5.
By strengthening the property WRN we define two seemingly natural subclasses: we call
the first one uniformly weakly Radon-Nikodým algebras (UWRN), see Definition 6.1, and
the latter is denoted by SWRN. An algebra A is SWRN if it is generated by some family
containing no infinite independent sequence, see Definition 6.5. We show that the classes
UWRN and SWRN are incomparable, see Proposition 6.6 and Corollary 9.5.
The class of WRN algebras is hereditary. To show that the class of UWRN is also stable
under taking subalgebras we prove in section 7 a result on families containing only short
independent sequences; our Theorem 7.1 is quantitative in nature and is based on the
Sauer-Shelah lemma (cf. Appendix A).
One of themes under consideration is properties of finitely additive measures on Boolean
algebras. A family E in a Boolean algebra contains an infinite independent sequence if and
only if contains a sequence separated by some measure, see Lemma 2.2. Using this fact
we show in Theorem 7.2 that if E contains no infinite independent sequence then the same
property has the image of E under any Boolean polynomial. In section 8 we discuss two
problems on measures on WRN algebras we were not able to resolve. They are related to a
question posed by J. Rodríguez, if every Borel measure defined on a WRN compact space
is concentrated on a separable subspace. Rodríguez's question is connected with his result
on measures of convex WRN compacta, which we enclose in Appendix B.
Finally, we discuss a problem posed by R. Haydon, if every WRN compact space contains
a nontrivial converging sequence (Problem 9.1). We prove that if A is a UWRN algebra
then its Stone space is even sequentially compact (Theorem 9.3). This implies that an
algebra F considered by Haydon [11] is in SWRN but not in UWRN.
We wish to thank Witold Marciszewski and José Rodríguez for fruitful discussions; we
are grateful to J. Rodríguez for his consent to include Theorem B.2 here.
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
3
2. Preliminaries
We shall consider abstract Boolean algebras A, B, . . ., keeping the usual set-theoretic
notation. In particular, ac denotes the complement of a ∈ A, but we shall also write a1 = a
and a0 = ac when convenient. Recall that, an indexed family {ai : i ∈ I} in an algebra A
is independent if
\i∈I ′
ai ∩ \i∈I ′′
ac
i 6= 0,
for every pair I ′, I ′′ of finite disjoint subsets of I.
Given an algebra A and any G ⊆ A, we denote by hGi the algebra generated by G, the
smallest subalgebra of A containing G. If hGi = A then G is called a generating family.
For an algebra A, by ult(A) we denote its Stone space (of ultrafilters on A) and A ∋ a →
ba ∈ Clop(ult(A)) is the Stone isomorphism between A and the algebra of clopen subsets of
By a measure µ on an algebra A we mean a finitely additive nonnegative probability
its Stone space.
functional A → [0, 1].
Definition 2.1. Let µ be a measure on an algebra A. We say that the measure µ
(a) is nonatomic if for every ε > 0 there are n ≥ 1 and a finite partition {a1, . . . , an} of
1A such that µ(ai) < ε for every i ≤ n.
(b) has countable type if there is a countable subalgebra B of A such that for every a ∈ A
we have
inf{µ(a △ b) : b ∈ B} = 0.
(c) is strongly countably determined if there is a countable subalgebra C ⊆ A such that for
every a ∈ A we have
µ(a) = sup{µ(c) : c ∈ C, c ⊆ a}.
This kind of properties of measures on Boolean algebras are discussed in [4] and [16].
Clearly a strongly countably determined measure has countable type; recall that the reverse
implication does not hold in general.
We shall say that a family E (of sets or elements of some Boolean algebra) is ε-separated
by a (finitely additive) measure µ if µ(a △ b) ≥ ε for all distinct a, b ∈ E (so, in particular,
E should be contained in the domain of µ). Note that if we are interested in properties
of E defined in terms of finite Boolean operations then we can in fact assume that µ is
if E is contained in a Boolean algebra A then we can switch to the
countably additive:
family bE = {ba : a ∈ E} of clopen subsets of the Stone space ult(A) and consider bµ, where
bµ(ba) = µ(a) for a ∈ A.
Lemma 2.2 ([7]). If ε > 0 and E is an infinite family ε-separated by some measure µ then
E contains an infinite independent sequence.
4
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
Note that, in turn, if (an)n is an independent sequence in some Boolean algebra A then
there is a measure µ on A such that µ(an) = 1/2 for every n and an are (stochastically)
µ-independent. In particular, if n 6= k then µ(an △ ak) = 1/2 so an are separated by µ.
Lemma 2.2 was proved in [7] by a direct argument; it can be also easily derived from the
classical lemma due to Rosenthal, which we recall here in the following form.
Lemma 2.3 ([20]). Let (Ak)k be a sequence of subsets of some set T . Then
(i) either there is an infinite N ⊆ ω such that the sequence of functions (χAk)k∈N con-
verges pointwise,
(ii) or else there is an infinite N ⊆ ω such that the sequence (Ak)k∈N is independent.
3. Weakly Radon-Nikodým algebras
The following definition singles out the class of Boolean algebras that plays a central role
in our paper.
Definition 3.1. A Boolean algebra A is weakly Radon-Nikodým (WRN) if there is a
family G ⊆ A generating A such that G can be written as G = Sn Gn, where, for every n,
Gn contains no infinite independent sequence.
The class of WRN algebras is directly related to the class of weakly Radon-Nikodým com-
pacta. Glasner and Megrelishvili [8] define a compact space to be weakly Radon-Nikodým
(WRN) if it is homeomorphic to a weak* compact subset of the dual of a Banach space
not containing an isomorphic copy of ℓ1. Using the Davis-Figiel-Johnson-Pełczyński fac-
torization technique one can easily prove that a compact space K is WRN if and only if
the Banach space C(K) is weakly precompactly generated. The concept of weakly pre-
compactly generated Banach spaces was introduced by Haydon [11]. A Banach space X is
said to be weakly precompactly generated if X = span(W ) for some weakly precompact
set W ⊆ X, i.e.
for some set W such that every sequence in W has a weakly Cauchy
subsequence. We refer the reader to [17] for further discussion on WRN compacta and
related classes.
The following characterization of WRN Boolean algebras is a consequence of [17, Lemma
2.7 and Theorem 2.8].
Proposition 3.2. The following conditions are equivalent for a Boolean algebra A:
(i) A is WRN;
(ii) There is a decomposition A =Sn∈ω En such that, for every n, En contains no infinite
independent sequence;
(iii) the Stone space ult(A) of A is weakly Radon-Nikodým compact.
The implication (i) → (ii) in Proposition 3.2 does not seem to be obvious; we shall give
in section 7 its direct proof. Note that condition (ii) provides an equivalent definition of
WRN algebras that is sometimes more convenient; for instance it yields the following.
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
5
Corollary 3.3. If A is a WRN algebra then every subalgebra B ⊆ A is WRN too.
Note that the algebra A itself does not contain an infinite independent sequence if and
only if the space ult(A) is scattered, i.e. A is superatomic. Clearly every countable algebra
is WRN. More generally, if A is an interval algebra, that is A = hGi, where the family G
is linearly ordered then A is WRN. It follows from Corollary 3.3 that every subalgebra of
an interval algebra is WRN. This is also a consequence of a result due to Heindorf [12],
stating that an algebra A embeds into some interval algebra if and only if A = hGi, where
G has the property that any two elements of G are either comparable or disjoint.
To give some examples of algebras that are not WRN note another obvious consequence
of Proposition 3.2.
Corollary 3.4. If A is a WRN algebra then A contains no uncountable independent se-
quence.
Hence P (ω), P (ω)/fin, Clop(2ω1) are not WRN; moreover, no infinite complete algebra
can be weakly Radon-Nikodým.
4. Minimally generated algebras
The notion of minimal extensions of algebras was introduced by S. Koppelberg, see [15];
the basic facts we mention below can be found in [15] or [4].
If B is a subalgebra of a Boolean algebra A and x ∈ A \ B then B(x) denotes the
subalgebra of A generated by B ∪ {x}, that is
B(x) = {(b ∩ x) ∪ (b′ ∩ xc) : b, b′ ∈ B}.
B(x) is said to be a minimal extension of B if for any algebra C, if B ⊆ C ⊆ B(x) then
either C = B or C = B(x). We recall the following basic fact on minimal extensions, see
Proposition 3.1 in [15].
Proposition 4.1. Let A be a Boolean algebra, B ≤ A a subalgebra and x ∈ A. Then B(x)
is a minimal extension of B if and only if for every b ∈ B, x ∩ b or x ∩ bc is in B.
Proof. Suppose that the extension B ≤ B(x) is minimal. Take b ∈ B. If x ∩ b is not in B,
then B(x ∩ b) = B(x). Therefore, there exist a1, a2 ∈ B such that
x =(cid:0)a1 ∩ (x ∩ b)(cid:1) ∪(cid:0)a2 ∩ (x ∩ b)c(cid:1) = (a1 ∩ x ∩ b) ∪ (a2 ∩ (xc ∪ bc)).
Hence x ∩ bc = a2 ∩ bc ∈ B.
Suppose now that for every b ∈ B, either x ∩ b or x ∩ bc is in B. Consider any element
y ∈ B(x) \ B. Then there are disjoint a1, a2, a3 ∈ B such that
y = (a1 ∩ x) ∪ (a2 ∩ xc) ∪ a3.
Since a2 ∩ xc = a2 ∩ (a1 ∪ xc) = a2 ∩ ((ac
1 ∩ x)c), it follows that either a1 ∩ x or a2 ∩ xc is in B.
By symmetry, we can assume that a2 ∩ xc ∈ B. Then y can be written as y = (a1 ∩ x) ∪ c
6
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
where a1, c ∈ B are disjoint. It follows that a1 ∩ x /∈ B and a1 ∩ x = y ∩ cc ∈ B(y). Since
ac
1 ∩ x must be in B, we get x ∈ B(y) so B(x) = B(y), as required.
(cid:3)
Corollary 4.2. In the setting of Proposition 4.1, if for every finite B0 ≤ B there is a
finite subalgebra B1 with B0 ≤ B1 ≤ B such that B1 ≤ B1(x) is a minimal extension then
B(x) is a minimal extension of B.
A Boolean algebra A is minimally generated over B ≤ A if A can be written, for some
ordinal number γ, as a continuous union
A = [ξ<γ
Bξ,
where B0 = B, and Bξ+1 is a minimal extension of Bξ for every ξ < γ. Finally, A is said
to be minimally generated if it is minimally generated over the trivial algebra B = {0, 1}.
Note that if A is minimally generated then its Stone space ult(A) can be seen as a limit
ξ )ξ<η<κ, where K0 = {0, 1}, every Kξ is compact, and for every
: Kξ+1 → Kξ has the property that there is a unique xξ ∈ Kξ
of an inverse system (Kξ, πη
ξ < κ, the bonding map πξ+1
such that (πξ+1
ξ
ξ
)−1(xξ) = 2, and (πξ+1
)−1(x) = 1 for x 6= xξ.
ξ
Koppelberg [15] proved that a minimally generated algebra cannot contain an uncount-
able independent sequence. This was generalized by Borodulin-Nadzieja [4] as follows.
Theorem 4.3. Let A be a minimally generated algebra. Then every measure µ on A
has countable type. In particular, A contains no uncountable independent sequence and,
consequently, ult(A) cannot be mapped onto 2ω1.
Let us remark that Theorem 4.3 and Corollary 3.4 name a common feature of the classes
of minimally generated algebras and WRN algebras. Another shared property is that both
classes contain all superatomic algebras (that is algebras having scattered Stone spaces).
On the other hand, we show below that there is a WRN Boolean algebra which is not
minimally generated and, in the next section, we give an example of a minimally generated
algebra which is not weakly Radon-Nikodým.
Example 4.4. There exists a WRN Boolean algebra which is not minimally generated.
Proof. Let B be the algebra of subsets of [0, 1) generated by the chain {[0, t) : 0 < t < 1}.
Then B is an interval algebra and therefore is minimally generated, see Example 2.1 in [15]
(note that its Stone space is the familiar split interval).
Koppelberg [14, Example 1] proved, in particular, that the free product A = B ⊗ B is
not minimally generated. Such a free product is generated by
G =n[0, a) × [0, 1) : 0 < a < 1o ∪n[0, 1) × [0, a) : 0 < a < 1o.
Note that no three elements of G are independent so A is a WRN algebra.
(cid:3)
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
7
5. Example of a minimally generated algebra which is not WRN
We give in this section a construction of a Boolean algebra with the properties announced
by the title. We shall work in the Cantor set 2ω; let A0 = Clop(2ω). For every partial
function σ on ω into 2 we write
[σ] = {x ∈ 2w : x(i) = σ(i) for every i ∈ dom σ}.
Let T = {3n : n ∈ ω} and let S(T ) be the space of all permutations of T .
Let x ∈ 2ω, ϕ ∈ S(T ) be given. We shall define a certain set A(x, ϕ) ⊆ 2ω. First define
partial functions σn(x, ϕ) on ω as follows.
(i) σn(x, ϕ)(i) = x(i) if i ∈ 3n \ T ;
(ii) σn(x, ϕ)(ϕ(i)) = x(ϕ(i)) if i ∈ T and i < 3n;
(iii) σn(x, ϕ)(ϕ(3n)) = (x(ϕ(3n)) + 1) mod 2.
Note that every σn(x, ϕ) is defined on the set
(3n \ T ) ∪ ϕ({i ∈ T : i ≤ 3n}),
so the domain of σn(x, ϕ) is of size 3n + 1. We now set
A(x, ϕ) =[n
[σn(x, ϕ)] .
We shall say below that a sequence (An)n of subsets of 2ω converges to a point x ∈ 2ω if
every neighbourhood of x contains An for almost all n.
Claim 5.1. For any x and ϕ, ([σn(x, ϕ)])∞
converging to x.
n=1 is a sequence of disjoint clopen subsets of 2ω
Proof. If n < k then σn(x, ϕ)(ϕ(3n)) 6= σk(x, ϕ)(ϕ(3n)) so the clopen sets in question are
disjoint. If τ is any partial function with a finite domain I and x ∈ [τ ] then take n0 such
that
I ⊆ (3n0 \ T ) ∪ ϕ({i ∈ T : i < 3n0}).
Then σn(x, ϕ) extends τ so [σn(x, ϕ)] ⊆ [τ ] for every n ≥ n0.
(cid:3)
Let us now fix a Borel bijective map g : 2ω → S(T ) (recall that between any two
uncountable Polish spaces there is always a Borel isomorphism; the fact that g is Borel
will be needed for the proof of Theorem 5.5). For every x ∈ 2ω take Ax = A(x, g(x)). We
define the desired algebra A of subsets of 2ω as the one generated by A0 together with the
family {Ax : x ∈ 2ω}.
Claim 5.2. The algebra A is minimally generated.
Proof. Note that for any distinct x, y ∈ 2ω, [σn(x, g(x))] and [σn(y, g(y))] are sequences of
It follows that either x /∈ Ay and then
clopen sets converging to x and y, respectively.
Ax ∩ Ay is clopen or, x ∈ Ay and then Ax \ Ay is clopen. Therefore, A is minimally
8
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
generated over A0 by Proposition 4.1 and hence A is minimally generated (since A0 is
minimally generated because it is countable).
(cid:3)
Claim 5.3. The algebra A is not weakly Radon-Nikodým.
An contains an independent sequence.
Proof. Take any decomposition A =Sn<ω An; We shall prove that there is n < ω such that
Define Φn = {g(x) : Ax ∈ An} for every n < ω. Since Sn Φn = g(2ω) = S(T ) and S(T )
is a completely metrizable space (notice that S(T ) is a Gδ-set in T T , which is completely
metrizable), the Baire Category Theorem asserts that there exists n0 < ω and a partial
function ψ from ω to ω such that Φn0 ∩ [ψ] is dense in S(T ) ∩ [ψ]. We can assume that
the domain of ψ is {0, 3, ..., 3(p − 1)} for some p; fix also i0 ∈ ω such that the range of ψ is
included in {0, 3, ..., 3(i0 − 1)}.
Note that, by density, for any i ≥ i0 there is xi ∈ 2ω such that Axi ∈ An0, g(xi)(3p) = 3i
and g(xi)(3j) = ψ(3j) for every j < p. Passing to a subsequence of i's, we can additionally
assume that xi have constant values for all n < 3p. Then the following are satisfied:
(a) for every n < p and for every i, j,
[σn(xi, g(xi))] = [σn(xj, g(xj))] .
(b) there is a partial function σ from ω into 2 with the domain of size 3p such that for
every i ≥ i0 we have
[σp(xi, g(xi))] = [σ] ∩ C ǫi
i ,
where we write C ǫi
cylinder in 2ω.
i = {x ∈ 2ω : x(3i) = ǫi} for the corresponding one-dimensional
Let µ be the canonical product measure on 2ω. We shall prove that {Axi}i≥i0 is ε-
separated for some ε > 0.
Note that for every xi and every n < ω,
µ([σn(xi, g(xi))]) =
1
23n+1 .
Using (a) -- (b) above, for distinct i, j ≥ i0 we get
µ(Axi \ Axj ) ≥ µ(cid:16)(cid:16)([σ] ∩ C ǫi
23p+2 −Xn>p
23n+1 =
≥
1
1
i ) \ ([σ] ∩ C ǫj
j )(cid:17) \(cid:16)[n>p
[σn(xj, g(xj))](cid:17)(cid:17) ≥
1
23p+2 −
1
23p+4
1
1 − 2−3 =
5
7
1
23p+2 .
It follows that the sets Axi for i ≥ i0 are ε-separated with ε > 0 so by Lemma 2.2 there
(cid:3)
is an independent subsequence in An0 and we are done.
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
9
The following result summarises our considerations and gives another property of the
Boolean algebra we have constructed. Recall that a compact space K is Rosenthal compact
if K can be embedded into B1(X), the space of Baire-one functions on a Polish space X
equipped with the topology of pointwise convergence. We shall use the following result.
Theorem 5.4. [5, Corollary 4.9] Every separable compact space consisting of Borel func-
tions over a Polish space is Rosenthal compact.
Theorem 5.5. There is a minimally generated algebra A such that its Stone space K =
ult(A) is a separable Rosenthal compactum which is not weakly Radon-Nikodým.
Proof. By Claim 5.3 the algebra A is not WRN so K = ult(A) is not weakly Radon-
Nikodým compact. It follows easily from Claim 5.2 that A0 is a dense subalgebra of A.
Hence K has a countable π-base so is, in particular, separable. We prove below that K is
indeed Rosenthal compact.
Given an ultrafilter u ∈ K, let zu be the unique point in 2ω such that
\{C ∈ Clop(2ω) : C ∈ u} = {zu}.
Claim A. For every u ∈ K we have
{y ∈ 2ω : zu ∈ Ay} ⊆ {y ∈ 2ω : Ay ∈ u} ⊆ {y ∈ 2ω : zu ∈ Ay} ∪ {zu}.
The first inclusion is clear. To check the latter, note first that Ay = Ay ∪ {y} for every
y ∈ 2ω since Ay is the union of clopen sets converging to y. Hence if Ay ∈ u and zu /∈ Ay
then y = zu (otherwise, zy /∈ Ay which contradicts the definition of zu).
Claim B. For every u ∈ K, {y ∈ 2ω : Ay ∈ u} is a Borel subset of 2ω.
By Claim A, it is sufficient to check that for any z ∈ 2ω the set {y ∈ 2ω : z ∈ Ay} is
Borel. But
{y ∈ 2ω : z ∈ Ay} =[n
{y ∈ 2ω : z ∈ [σn(y, g(y))]},
and every set {y ∈ 2ω : z ∈ [σn(y, g(y))]} is plainly Borel because the function g is Borel.
Consider now the following mapping f : K → 2ω × 22ω
f (u) =(cid:0)zu, χ{y∈2ω:Ay∈u}(cid:1) .
Then f is injective since every ultrafilter u ∈ K is uniquely determined by the family of
generators of the algebra A that are in u. It is clear that f is continuous. It follows from
Claim B that K is homeomorphic to a pointwise-compact set of Borel functions on a Polish
space. Since K is separable, K is Rosenthal compact by Theorem 5.4.
(cid:3)
10
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
6. WRN algebras with stronger properties
In this section we introduce two subclasses of WRN algebras; they are defined by natural
conditions that are slightly stronger than that of Definition 3.1.
Definition 6.1. A Boolean algebra A is in the class I(n), where n ≥ 1, if A is generated
by a family G ⊆ A such that G contains no n + 1 independent elements.
Definition 6.2. A Boolean algebra A is uniformly weakly Radon-Nikodým (UWRN) if A
is generated by a family G =Sn∈ω Gn such that no Gn contains an independent sequence of
length n.
Note that every interval algebra is in I(1). In turn, the following holds.
Theorem 6.3. Every Boolean algebra from I(1) is minimally generated.
Proof. Take a Boolean algebra A ∈ I(1) and a family G generating A containing no inde-
pendent pairs of elements. We shall check the following.
Claim. For every finite J ⊆ G and every x ∈ G, the extension hJ i ≤ hJ ∪{x}i is minimal.
It is clear that Claim holds if J = {y}, since x, y are not independent. We argue by
induction on J .
Suppose that every extension hJ i ≤ hJ ∪ {x}i is minimal whenever J = n. Take
x ∈ G and J ⊆ G with J = n + 1. We prove that the extension hJ i ≤ hJ ∪ {x}i is
also minimal. Choose y ∈ J and set S = J \ {y}. We are going to prove that for every
z ∈ hJ i, z ∩ x or zc ∩ x is in hJ i. Since z ∈ hJ i = hS ∪ {y}i, we know that
z = (a ∩ y) ∪ (b ∩ yc) for some a, b ∈ hSi.
Since S = n, we know that a ∩ x or ac ∩ x is in hSi, and b ∩ x or bc ∩ x is in hSi.
Suppose that z ∩ x /∈ hJ i. Since
z ∩ x = (a ∩ y ∩ x) ∪ (b ∩ yc ∩ x),
then either a ∩ y ∩ x /∈ hJ i or b ∩ yc ∩ x /∈ hJ i.
Consider the case when a ∩ y ∩ x /∈ hJ i. Then a ∩ x /∈ hSi so ac ∩ x ∈ hSi. Moreover,
it follows that x ∩ y 6= 0 and y 6⊆ x. Since y and x are not independent, this leaves us two
possibilities: either x ⊆ y or x ∪ y = 1. Hence yc ∩ x = 0 or yc ∩ x = yc so yc ∩ x ∈ hJ i in
both cases. By easy calculation we get
zc = (a∩y)c ∩(b∩yc)c = (ac ∪yc)∩(bc ∪y) = (ac ∩bc)∪(ac ∩y)∪(bc ∩yc) = (ac ∩y)∪(bc ∩yc),
and it follows that
zc ∩ x =(cid:0)(ac ∩ x) ∩ y(cid:1) ∪(cid:0)(yc ∩ x) ∩ bc(cid:1) ∈ hJ i.
If b ∩ yc ∩ x /∈ hJ i then in a similar way we get bc ∩ x ∈ hSi and x ∩ y ∈ hJ i, giving
zc ∩ x ∈ hJ i. This finishes the proof of Claim.
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
Now we conclude the proof of the theorem applying Claim and Corollary 4.2.
11
(cid:3)
Note that Example 4.4 in fact gives the following.
Corollary 6.4. There exists a Boolean algebra in I(2) which is not minimally generated.
We can also strengthen the condition of Definition 3.1 in the following way.
Definition 6.5. Let us say that a Boolean algebra is strongly WRN (SWRN) if it is gen-
erated by a family containing no infinite independent sequence.
The classes UWRN and SWRN are incomparable, see Corollary 9.3 and the following
result.
Proposition 6.6. There exists a UWRN algebra which is not SWRN.
Proof. Let A be the algebra of clopen sets of a countable product of one point compactifi-
cations of a discrete set of cardinality ω1. Let F = {en
α : n < ω, α < ω1} be the canonical
generators of A which are independent except for the relation en
β = 0 whenever α 6= β.
Clearly the algebra A is UWRN. We prove below that it is not SWRN.
α ∩ en
α) be the set of all β < ω1 such that ek
Suppose that G is a system of generators. It is enough to check that the image of G under
some quotient contains an infinite independent sequence. Express each en
α as a Boolean
polynomial of generators from G and in turn each such generator as a Boolean polynomial
of generators from F . Let Fk(en
β appears in such
α. Notice that for every α < ω1, n < ω each set Fk(en
expression of en
α) is finite and,
moreover, Fk(en
α) = ∅ for all except finitely many k < ω. By passing, for each n, to an
uncountable subset An ⊆ ω1 (by this we mean, making a quotient that makes each en
α,
α 6∈ An vanish), we can suppose that for every n there is mn < ω such that Fk(en
α) = ∅
if k ≥ mn and Fk(en
α) < mn if k < mn. Moreover, we can also suppose that each family
{Fk(en
α) : α ∈ An} is a ∆-system. By removing all roots (that form just a countable set),
we can suppose that the family {Fk(en
α) : α ∈ An} is always pairwise disjoint. Now is
αn) ∩ Fq(em
easy to get αn ∈ An such that Fk(en
αm) = ∅ for all k, q, n, m with n 6= m. If we
make vanish all generators of F except the en
αn's, we will find that one of the generators
from G (call it gn) in the expression of en
αn is en
αn itself. We found an infinite independent
sequence.
(cid:3)
We give in the next section (see Corollary 7.3) a characterization of UWRN algebras.
7. Playing with independence
In this section we study the behaviour of families containing no infinite independent
sequences or independent sequences of size n for some n ∈ ω. As a consequence of the
results obtained in this section we get a characterization of UWRN algebras similar to the
one given in Proposition 3.3. The proof below is based on the Sauer-Shelah Lemma ([21],
[22]) which is recalled in the appendix together with a proof.
12
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
Theorem 7.1. Let E be a family in some Boolean algebra A such that E contains no
independent family of size n. Fix r ≥ 1 and set
I(n, r) := min{s ∈ ω :(cid:18)rs
0(cid:19) +(cid:18)rs
1(cid:19) + · · · +(cid:18) rs
n − 1(cid:19) < 2s}.
Then, for any Boolean polynomial p(x1, x2, . . . , xr) the family p(E) = {p(a1, . . . , ar) :
a1, . . . ar ∈ E} contains no independent sequence of length I(n, r).
Proof. Suppose p(E) contains an independent sequence of length I(n, r). Then there exist
b1 = p(a1,1, a1,2, . . . , a1,r),
b2 = p(a2,1, a2,2, . . . , a2,r),
. . .
bI(n,r) = p(aI(n,r),1, aI(n,r),2, . . . , aI(n,r),r)
such that b1, b2, . . . , bI(n,r) is an independent family. Without loss of generality, we may
suppose that E = {ai,j
: 1 ≤ i ≤ I(n, r), 1 ≤ j ≤ r}. Let us put, for convenience,
N = rI(n, r) and E = {a1, a2, . . . , aN }. Since hEi contains an independent family of size
I(n, r), it must contain at least 2I(n,r) atoms. Moreover, every atom of hEi has a unique
representation of the form af (1)
N , where f ∈ 2{1,2,...,N } and for each element
a ∈ hEi, we denote the complement of a as a0 and a as a1. Set
2 ∩· · ·∩af (N )
1 ∩af (2)
F = {f ∈ 2{1,2,...,N } : af (1)
1 ∩ af (2)
2 ∩ · · · ∩ af (N )
N
is an atom of hEi}.
We claim that F ≤ (cid:0)N
0(cid:1) +(cid:0)N
1(cid:1) + · · · +(cid:0) N
n−1(cid:1). If not, by the Sauer-Shelah Lemma (see
Lemma A.1) there exists a set S ⊆ {1, 2, . . . , N} with S = n such that
{f S : f ∈ F } = 2S.
But this means that {ai : i ∈ S} is an independent family, since for each f ∈ 2S, the
is nonempty because it contains an atom. This is in contradiction with
i
element Ti∈S af (i)
F ≤(cid:18)N
the hypothesis on E, so
0(cid:19) +(cid:18)N
1(cid:19) + · · · +(cid:18) N
n − 1(cid:19).
Since the number of atoms of hEi is exactly F , we conclude that
2I(n,r) ≤(cid:18)N
0(cid:19)+(cid:18)N
1(cid:19)+· · ·+(cid:18) N
n − 1(cid:19) =(cid:18)rI(n, r)
0 (cid:19)+(cid:18)rI(n, r)
n − 1 (cid:19),
1 (cid:19)+· · ·+(cid:18)rI(n, r)
in contradiction with the definition of I(n, r).
(cid:3)
Our Theorem 7.1 has the following counterpart.
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
13
Theorem 7.2. Let E be a family in some Boolean algebra A such that E contains no infinite
independent sequence. Let, for a fixed r, p(x1, x2, . . . , xr) be any Boolean polynomial. Then
the family
p(E) = {p(a1, . . . , ar) : a1, . . . , ar ∈ E}
contains no infinite independent sequence.
Proof. Consider first the polynomial p(x, y) = x∩y. Suppose that p(E) contains cn = an∩bn
with an, bn ∈ E such that the sequence (cn)n is independent. By the remark following
Lemma 2.2 there is a finitely additive probability measure µ on A such that µ(cn) = 1/2
and cn's are stochastically independent with respect to µ.
For k 6= n we have
1/4 = µ(ck \ cn) ≤ µ(ck \ an) + µ(ck \ bn),
so either µ(ck \ an) ≥ 1/8 or µ(ck \ bn) ≥ 1/8. Say that the pair {k, n} gets the colour a if
the first inequality holds and the colour b otherwise. By the Ramsey theorem there is an
infinite N ⊆ ω such that whenever k, n ∈ N are different then {k, n} has the same colour;
say that this is a.
It follows that for k, n ∈ N, k < n, we have
µ(ak △ an) ≥ µ(ak \ an) ≥ µ(ck \ an) ≥ 1/8,
so the family {an : n ∈ N} is 1/8-separated by µ. Applying Lemma 2.2 we get a contra-
diction.
We can assume that E is closed under taking complements. If we consider the polynomial
p′(x, y) = x ∪ y then p′(x, y) = (p(xc, yc))c so the result follows for p(E) by the argument
above.
The general case follows by induction on the complexity of the Boolean polynomial in
(cid:3)
question.
Using Theorems 7.1 we conclude the following analogue of Proposition 3.2.
Corollary 7.3. A Boolean algebra A is UWRN if and only if there is a decomposition
A =Sn∈ω En such that no En contains an independent sequence of length n.
Consequently, the class of UWRN algebras is stable under taking subalgebras.
Problem 7.4. Is the class SWRN hereditary?
8. Measures on UWRN algebras
The following result and Theorem 4.3 show another common feature of WRN algebras
and minimally generated ones.
Proposition 8.1. If µ is a measure on a WRN algebra A then µ has countable type.
14
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
Proof. Suppose otherwise; note that then there is ε > 0 and an uncountable family F such
that µ(a △ b) ≥ ε for any distinct a, b ∈ F .
Since A is WRN, we have a decomposition A =Sn En as in Proposition 3.2(ii). But then
F ∩ En is uncountable for some n and we arrive at a contradiction with Lemma 2.2.
(cid:3)
In fact Proposition 8.1 can be generalized to a result on Borel measures defined on WRN
compacta, see Proposition B.1.
Proposition 8.2. If µ is a nonatomic measure on A and A ∈ I(1) then µ is strongly
countably determined.
Proof. By the assumption, A = hGi where G contains no independent pair.
Fix ε > 0. There is a finite G0 ⊆ G such that B = hG0i has all atoms of measure < ε.
Take any g ∈ G and consider b0, b1 ∈ B, where b0 is the maximal element of B contained
in g, while b1 is the minimal element of B containing g.
Claim. b1 \ b0 is an atom of B.
Indeed, for any h ∈ G0, either h ⊆ g which implies h ⊆ b0, or h ∩ g = 0 which gives
h ∩ b1 = 0, or g ⊆ h which implies b1 ⊆ h, or else h ∪ g = 1 and in this case h ⊇ b1 \ b0. So
b1 \ b0 is split by no h ∈ G0 and hence it is an atom of B.
It follows from Claim that µ(b1 \ b0) < ε, so
µ(b0) = µ(b1) − µ(b1 \ b0) ≥ µ(g) − ε,
so b0 approximates g from inside; likewise, bc
1) ≤ ε. Now, taking a
countable G′ ⊆ G such that µ is nonatomic on C = hG′i, it follows that for every g ∈ G, we
have
1 ⊆ gc and µ(gc \ bc
µ(g) = sup{µ(c) : c ∈ C, c ⊆ g} and µ(gc) = sup{µ(c) : c ∈ C, c ⊆ gc}.
By a standard argument we conclude that µ(a) = sup{µ(c) : c ∈ C, c ⊆ a} for every a ∈ A,
so µ is strongly countably determined.
(cid:3)
Borodulin-Nadzieja [4, 4.11,4.12] proves that a measure on an algebra that is minimally
generated by a sequence of order type ω1 is strongly countably determined but this is no
longer true for arbitrary minimally generated algebras.
Problem 8.3. Is it true that for every n and every algebra A ∈ I(n), every nonatomic
measure on A is strongly countably determined?
Note that if the answer to the above problem is positive then every nonatomic measure
on an UWRN algebra is strongly countably determined. In turn, this would imply that if
K is a zerodimensional compact space with Clop(K) being a UWRN algebra then every
regular Borel measure on K is concentrated on a separable subspace of K.
Problem 8.4 (Rodríguez). Let K be a weakly Radon-Nikodým compact space. Is it true
that every regular Borel measure on K is concentrated on a separable subspace of K?
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
15
That question is motivated by a result due to J. Rodríguez, see Appendix B.
9. Nontrivial convergent sequences
As we noted above, typical spaces without nontrivial converging sequences, such as βω,
are not weakly Radon-Nikodým compact. On the other hand, R. Haydon proved that there
is a WRN compact space which is not sequentially compact. As far as we are concerned,
the following problem is open:
Problem 9.1. (Haydon, [11]) Does every WRN compact space contain a nontrivial con-
vergent sequence?
In this section we study Problem 9.1 for zero-dimensional compact spaces related to
UWRN and SWRN Boolean algebras.
We recall the construction of D. H. Fremlin used by Haydon [11]. This construction
provides an example of a SWRN Boolean algebra F such that ult(F) is not sequentially
compact.
Example 9.2. Let G be a family of subsets of ω maximal with respect to the condition
that for every A, B ∈ G there exists ε1, ε2 ∈ {0, 1} such that Aε1 ∩ Bε2 is finite. Let F be
the subalgebra of subsets of ω generated by G (note that G contains all finite subsets of
ω). It is clear that G does not contain an infinite independent sequence, so F is a SWRN
algebra.
Notice that ult(F) contains a natural copy of ω which consists of principal ultrafilters of
F. By the maximality of G every infinite A ⊆ ω is split into two infinite parts by some
G ∈ G. Consequently, the sequence of natural numbers ω ⊆ ult(F) does not contain a
converging subsequence, and therefore ult(F) is not sequentially compact.
Note that ult(F) \ ω is a compact space which is homeomorphic to the Stone space of the
quotient Boolean algebra A = F/f in. Then A is generated by G• = {G• : G ∈ G}. Since
no pair from G• is independent, A is in I(1). J. Bourgain proved that every sequence of
nonprincipal ultrafilters in ult(F) contains a convergent subsequence, cf. [10]. Thus, ult(A)
is sequentially compact.
We shall now generalize Bourgain's idea mentioned above.
Lemma 9.3. If A is a Boolean algebra in I(n) for some n < ω, then ult(A) is sequentially
compact.
Proof. Let G be a set of generators of A such that G contains no n+1 independent elements.
Take a sequence (xn)n in ult(A) and suppose that it admits no convergent subsequence.
For every k ∈ ω put Ak = {g ∈ G : g ∈ xk} and consider the sequence (Ak)k of subsets
of G.
Note that if N ⊆ ω is infinite then the subsequence (χAk )k∈N cannot converge pointwise;
indeed, otherwise, if χA = limk∈N χAk then it is easy to check that there is x ∈ ult(A) such
that x ⊇ A ∪ {gc : g ∈ G \ A} and limk∈N xk = x.
16
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
Hence, by Lemma 2.3, (Ak)k contains an independent subsequence. For the rest of the
proof we can assume that the whole sequence (Ak)k is independent. We shall check that
this is in contradiction with the assumption A being in I(n).
Let I = {1, 2, . . . , 2n+1}. Take any bijection ϕ between I and the finite product {0, 1}n+1
and let the functions rk : I → 2 be defined as ri = πi ◦ ϕ for i = 1, 2, . . . , n + 1, where πi
is the projection. Then the family (ri)i≤n+1 has the property that for every ε ∈ {0, 1}n+1
there is k ∈ I such that ε = (ri(k))i≤n+1.
Now for every i ≤ n + 1, using the independence of the sets Ak we can pick
gi ∈
2n+1\k=1
Ari(k)
k
,
where, for a set A, we mean A1 = A and A0 is its complement.
We claim that G0 = {gi : i ≤ n + 1} is an independent subfamily of G. Indeed, take any
ε ∈ {0, 1}n+1 and choose k ∈ I such that ε = (ri(k))i≤n+1. If ri(k) = ε(i) = 1 then gi ∈ Ak
which means that gi ∈ xk. If ri(k) = ε(i) = 0 then gi /∈ Ak, i.e. gi /∈ xk, which is equivalent
to gc
6= 0. This shows
that G0 is independent and we get a contradiction, as required.
(cid:3)
i ∈ xk for every i ≤ n + 1 so Ti≤n+1 gε(i)
i ∈ xk. It follows that gε(i)
i
Theorem 9.4. If A is an UWRN Boolean algebra, then ult(A) is sequentially compact.
Proof. Since A is UWRN we have A = hGi, and the decomposition G = Sn Gn as in
Definition 6.1. If we let An = hGni for every n we have an obvious embedding
ult(A) →
∞Yn=1
ult(An).
We conclude the proof applying Lemma 9.3 and the fact that the class of sequentially
compact spaces is stable under closed subspaces and countable products.
(cid:3)
Corollary 9.5. The Boolean algebra F from Example 9.2 is SWRN but not UWRN.
We remark that another example of a SWRN Boolean algebra which is not UWRN is
given by the well-known example of an Eberlein compact space which is not uniformly
Eberlein constructed by Benyamini and Starbird [3].
We finally remark that, although there are WRN algebras which are not SWRN, in
order to give an answer to Problem 9.1 for zero-dimensional compact spaces, it is enough
to consider SWRN algebras because we can use a similar argument as in the proof of
Theorem 9.4.
Appendix A. Sauer-Shelah Lemma
The following proof of the Sauer-Shelah Lemma ([21], [22]) is based on the proof contained
in Gil Kalai's blog [13].
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
17
Lemma A.1 (Sauer-Shelah). Let N, n be natural numbers with 1 ≤ n ≤ N and let T =
{1, 2, . . . , N}. Then for every family C ⊆ 2T with
C >(cid:18)N
0(cid:19) +(cid:18)N
1(cid:19) + · · · +(cid:18) N
n − 1(cid:19),
there exists a set S ⊆ T with S = n such that {f S : f ∈ C} = 2S.
Proof. We first prove the following stronger result:
Claim. For every family C ⊆ 2T there exists a family of sets F ⊆ P(T ) such that
F = C and
{f S : f ∈ C} = 2S for any S ∈ F .
We check the claim by induction on C. If C = 1 then take F = {∅}. Suppose C ≥ 2.
Without loss of generality, we may suppose that both the families
C0 = {f ∈ C : f (1) = 0} and C1 = {f ∈ C : f (1) = 1},
are nonempty. Put T ′ = T \ {1}. By induction, there exists F0 ⊆ P(T ′) with F0 = C0
such that
{f S : f ∈ C0} = 2S for any S ∈ F0.
Now take C ′
F1 = C ′
1 = {f T ′
1 such that
: f ∈ C1}. Again by induction, there exists F1 ⊆ P(T ′) with
{f S : f ∈ C ′
1} = 2S for any S ∈ F1.
Set
F = F0 ∪ F1 ∪ {S ∪ {1} : S ∈ F0 ∩ F1},
and note that
F = F0 + F1 = C0 + C ′
1 = C0 + C1 = C.
Therefore it is enough to prove that {f S : f ∈ C} = 2S for any S ∈ F , but this is a
consequence of the properties of F0 and F1. Thus the claim is proved.
Now the lemma follows from the fact that T has exactly (cid:0)N
0(cid:1) +(cid:0)N
of cardinality smaller than n, so by the assumption on C there exists a set S ⊆ T with
S ≥ n such that {f S : f ∈ C} = 2S.
(cid:3)
1(cid:1) + · · · +(cid:0) N
n−1(cid:1) subsets
18
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
Appendix B. On measures on WRN compacta
We enclose here two results on regular Borel measures defined on WRN compacta that
are proved in an unpublished note by José Rodríguez.
Proposition B.1. If K is WRN compact and if µ is a probability regular Borel measure
on K then µ has countable type (i.e. L1(µ) is separable).
Proof. As we mentioned in section 3, if K is WRN compact then C(K) is spanned by some
weakly precompact set W ⊆ C(K). If we consider the natural embedding C(K) ֒→ L1(µ)
then the image of W is norm-separable. Indeed, otherwise for some ε > 0 we could find
functions fn ∈ W such thatRK fn −fk dµ ≥ ε for n 6= k. But then (fn)n admits no weakly
Cauchy subsequence, a contradiction.
Since W is norm-separable in L1(µ), a standard argument gives that C(K) = span(W ) is
also norm-separable in L1(µ). But C(K) is dense in L1(µ) so L1(µ) is separable itself. (cid:3)
Let X be a Banach space and let K be a weak∗-compact subset of the dual unit ball
BX ∗. Let µ be a regular probability measure on K; denote by f : K → X ∗ the identity
function. Then for every B ∈ Borel(K) there is a vector ν(B) = RB f dµ ∈ X ∗ which is
the Gelfand integral of f on B, that is
(1)
hν(B), xi =ZB
x dµ
for every x ∈ X, see [6, page 53]. Here every x ∈ X is seen as a continuous function
K ∋ x∗ → x∗(x) on K. In other language, ν(B) is the barycenter of a measure 1/µ(B) · µB
which is the normalized restriction of µ to B.
Theorem B.2. Suppose that X is a Banach space not containing ℓ1 and that the set
K ⊆ BX ∗ is weak∗ compact and convex. Then for every probability regular Borel measure
µ on K there is a weak∗-closed and weak∗-separable set L ⊆ K such that µ(L) = 1.
Proof. Consider the set
S :=(cid:26) 1
µ(B)ZB
f dµ : B ∈ Borel(K), µ(B) > 0(cid:27) .
As above, we write ν(B) =RB f dµ for simplicity.
Claim. The set S is norm-separable.
The space L1(µ) is separable by Proposition B.1 so there is a countable family A of Borel
subsets of K of positive measure such that inf{µ(A △ B) : A ∈ A} = 0 for every Borel set
B ⊆ K. Note that
(2)
kν(B) − ν(A)k = sup
x∈BX(cid:12)(cid:12)(cid:12)(cid:12)
ZB
x dµ −ZA
x dµ(cid:12)(cid:12)(cid:12)(cid:12) ≤ µ(B △ A).
WRN ALGEBRAS AND INDEPENDENT SEQUENCES
19
Fix ε > 0 and a Borel set B of positive measure; take A ∈ A such that µ(B△A) < ε·µ(B)
and 1/µ(B) − 1/µ(A) < ε. Then, using (2) we get
1
µ(B)
ν(A) −
1
µ(A)
ν(A)(cid:13)(cid:13)(cid:13)(cid:13) ≤
(cid:13)(cid:13)(cid:13)(cid:13)
ν(A)(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13)
1
µ(B)
ν(B) −
1
µ(A)
1
µ(B)
ν(B) −
1
µ(B)
≤
1
µ(B)
kν(B) − ν(A)k + kν(A)k(cid:12)(cid:12)(cid:12)(cid:12)
1
µ(B)
−
and this verifies the claim.
ν(A)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13)
µ(A)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2ε.
1
Since S is norm-separable in X ∗, the weak∗- closed convex hull L := co(S)
is weak∗-
ω∗
separable. Let us check that L fulfills the required properties.
First note that L ⊆ K. To verify this it suffices to check that S ⊆ K. Take any
x∗ ∈ X ∗ \ K. By the Hahn-Banach theorem, there is x ∈ X such that x∗(x) > α :=
sup{y∗(x) : y∗ ∈ K}, therefore
(cid:28) 1
µ(B)ZB
f dµ, x(cid:29) (1)
=
1
µ(B)ZB
x dµ ≤ α < x∗(x),
for every B ∈ Borel(K) with µ(B) > 0. Hence x∗ 6∈ S.
It remains to prove that µ(L) = 1; we achieve it by checking that for every x∗ ∈ K \ L
there is a weak∗-open set U ⊆ K such that x∗ ∈ U and µ(U) = 0. Again, the Hahn-Banach
theorem ensures the existence of x ∈ X such that
x∗(x) > β := sup
y∗∈S
y∗(x)
(1)
= sup(cid:26) 1
µ(B)ZB
x dµ : B ∈ Borel(K), µ(B) > 0(cid:27) .
Fix β < γ < x∗(x). Then x∗ belongs to the w∗-open set U := {y∗ ∈ K : y∗(x) > γ}
and RU x dµ ≥ γµ(U). On the other hand, by the very definition of β we also have
βµ(U) ≥RU x dµ. Then βµ(U) ≥ γµ(U) and so µ(U) = 0.
Let us remark that the proof of Theorem B.2 actually shows the following assertion:
Given a convex weak∗-compact subset K of a dual Banach space, any regular Borel proba-
bility measure of countable type defined on K is concentrated on a weak∗-separable subset
of K.
(cid:3)
References
[1] A. D. Arvanitakis, Some remarks on Radon-Nikodým compact spaces, Fund. Math. 172 (2002), 41 -- 60.
[2] A. Avilés, P. Koszmider, A continuous image of a Radon-Nikodým compact space which is not Radon-
Nikodým, Duke Math. J., 162 (2013), no. 12, 2285 -- 2299.
[3] Y. Benyamini, T. Starbird, Embedding weakly compact sets into Hilbert space, Israel J. Math. 23
(1976), 137 -- 141.
[4] P. Borodulin-Nadzieja, Measures on minimally generated Boolean algebras, Topology Appl. 154 (2007),
3107 -- 3124.
[5] G. Debs, Descriptive aspects of Rosenthal compacta, Recent Progress in General Topology III, Springer
(2014), 205 -- 227.
[6] J. Diestel, J. J. Uhl, jr, Vector measures, Mathematical Surveys 15, Amer. Math. Soc., Providence,
R.I. (1977).
20
A. AVILÉS, G. MARTÍNEZ-CERVANTES, AND G. PLEBANEK
[7] D. H. Fremlin, G. Plebanek, Independence-precalibers of measure algebras. preprint (2004); available
on Plebanek's and Fremlin's webpages.
[8] E. Glasner, M. Megrelishvili, Representations of dynamical systems on Banach spaces not containing
l1, Trans. Am. Math. Soc., 364 (2012), no. 12, 6395 -- 6424.
[9] E. Glasner, M. Megrelishvili, Eventual nonsensitivity and tame dynamical systems, ArXiv e-prints,
arXiv:1405.2588, May 2014.
[10] J. Hagler, F. Sullivan, Smoothness and weak∗ sequential compactness, Proc. Amer. Math. Soc. 78
(1980), 497 -- 503.
[11] R. Haydon. Nonseparable Banach spaces, Functional analysis: surveys and recent results II, North-
Holland, Amsterdam (1981), 19 -- 30.
[12] L. Heindorf, On subalgebras of Boolean interval algebras, Proc. Amer. Math. Soc. 125 (1997), 2265 --
2274.
[13] G. Kalai, blog, see
https://gilkalai.wordpress.com/2008/09/28/extremal-combinatorics-iii-some-basic-theorems/.
[14] S. Koppelberg, Counterexamples in minimally generated Boolean algebras, Acta Univ. Carolin. Math.
Phys. 29 (1988), 27 -- 36.
[15] S. Koppelberg, Minimally generated Boolean algebras, Order 5 (1989), 393 -- 406.
[16] M. Krupski, G. Plebanek, A dichotomy for the convex spaces of probability measures, Topology Appl.
158 (2011), 2184 -- 2190.
[17] G. Martínez-Cervantes, On weakly Radon-Nikodým compact spaces, Israel J. Math. (to appear),
arXiv:1509.05324.
[18] I. Namioka, Radon-Nikodým compact spaces and fragmentability, Mathematika 34 (1987), no. 2, 258 --
281.
[19] J. Orihuela, W. Schachermayer, M. Valdivia, Every Radon-Nikodým Corson compact space is Eberlein
compact, Studia Math. 98 (1991), no. 2, 157 -- 174.
[20] H. Rosenthal, A characterization of Banach spaces containing l1, Proc. Nat. Acad. Sci. USA 71 (1974),
2411 -- 2413.
[21] N. Sauer, On the density of families of sets, J. Comb. Theory, 13 (1972), 145 -- 147.
[22] S. Shelah, A combinatorial problem; stability and order for models and theories in infinitary languages,
Pacific J. Math. 41 (1972), 247 -- 261.
[23] C. Stegall, Spaces of Lipschitz functions on Banach spaces, Functional analysis (Essen, 1991), Lecture
Notes in Pure and Appl. Math., vol. 150, Dekker, New York, 1994, 265 -- 278.
[24] M. Talagrand, Sur les espaces de Banach contenant ℓ1(τ ), Israel J. Math. 40 (1981), 324 -- 330.
Departamento de Matemáticas, Facultad de Matemáticas, Universidad de Murcia, 30100
Espinardo, Murcia, Spain
E-mail address: [email protected]
Departamento de Matemáticas, Facultad de Matemáticas, Universidad de Murcia, 30100
Espinardo, Murcia, Spain
E-mail address: [email protected]
Instytut Matematyczny, Uniwersytet Wrocławski, Pl. Grunwaldzki 2/4, 50-384 Wroc-
ław, Poland
E-mail address: [email protected]
|
1709.02522 | 1 | 1709 | 2017-09-08T03:34:27 | Remarks on strong embeddability for discrete metric spaces and groups | [
"math.FA",
"math.MG"
] | In this paper, we show that the strong embeddability has fibering permanence property and is preserved under the direct limit for the metric space. Moreover, we show the following result: let $G$ is a finitely generated group with a coarse quasi-action on a metric space $X$. If $X$ has finite asymptotic dimension and the quasi-stabilizers are strongly embeddable, then $G$ is also strongly embeddable. | math.FA | math |
Remarks on strong embeddability
for discrete metric spaces and groups∗
Guoqiang Li and Xianjin Wang
Abstract
In this paper, we show that the strong embeddability has fibering permanence
property and is preserved under the direct limit for the metric space. Moreover, we
show the following result: let G is a finitely generated group with a coarse quasi-action
on a metric space X. If X has finite asymptotic dimension and the quasi-stabilizers
are strongly embeddable, then G is also strongly embeddable.
1
Introduction
In [8], Gromov introduced the notion of coarse embeddability of metric spaces and sug-
gested that a discrete finitely generated group that coarsely embeds into a Hilbert space,
when equipped with a word length metric, would satisfy the Novikov conjecture [7, 8].
Subsequently, in [18], Yu proved the coarse Baum-Connes conjecture holds for bounded
geometry discrete metric spaces which are coarsely embeddable in Hilbert space. In the
same paper, Yu introduced a weak form of amenability that he called property A, which
ensures the existence of a coarse embedding into Hilbert space. In later years, property
A and coarse embeddability have been further studied [3, 5, 9, 13, 14]. In [5], the author
showed that property A is preserved under group extensions. Unlike property A, coarse
embeddability is not closed under group extensions [1]. In [10], Ji, Ogle and W. Ramsey
introduced the notion of strong embeddability which is stable under arbitary extensions.
Strong embeddability is an intermediate notion of strong coarse embeddability implied
by Property A and implying coarse embeddability. In [16], J. Xia and X. Wang studied
the permanence properties of strong embeddability. Moreover, they proved that a metric
space is strongly embeddable if and only if it has weak finite decomposition complexity
with respect to strong embeddability. In [17], J. Xia and X. Wang showed that a finitely
generated group acting on a finitely asymptotic dimension metric space by isometries
whose k-stabilizers are strongly embeddable is strongly embeddable. Now we extend this
result to conclude as follows.
∗The authors are supported by NSFC (No. 11231002).
1
Theorem 1.1. Assume that G is a finitely generated group with a coarse quasi-action on
a metric space X. If X has finite asymptotic dimension and there exists a base point such
that its all quasi-stabilizers are strongly embeddable, then G is strongly embeddable.
The coarse quasi-action (see Definition 2.7) is designed to describe situations where
elements of a group act on a metric space via coarse equivalence. And it is a significant
and useful generalization of actions by isometries and quasi-isometries.
2 Preliminaries
A discrete metric space X has bounded geometry if for all R > 0 there exists NR such
that B(x, R) ≤ NR for all x ∈ X, where · denotes the number of elements of the ball
B(x, R). X is called uniformly discrete if there exists a constant C > 0 such that for any
two distinct points x, y ∈ X we have d(x, y) ≥ C. Assume that all metric spaces in this
paper are uniformly discrete with bounded geometry. This class includes many interesting
examples, in particular, all countable and uniformly discrete groups. Let B be a Banach
space and B1 = {η ∈ B kηk = 1}. For any R, ǫ > 0, a map x 7→ ξx from X to B will be
said to have (R, ǫ) variation if d(x, y) ≤ R implies kξx − ξyk ≤ ǫ.
Definition 2.1 (see [10]). Let X be a metric space. Then X is strongly embeddable if
and only if for every R, ǫ > 0 there exists a Hilbert space valued map β : X → (l2(X))1
satisfying:
(1) β has (R, ǫ) variation;
(2) lim
S→∞
sup
x∈X Pw /∈B(x,S)
βx(w) 2= 0.
We will need to make use of a family of strongly embeddable metric spaces, with some
uniform control.
Definition 2.2 (see [10]). A family (Xi)i∈I of metric spaces is equi-strongly embeddable
if for every R, ǫ > 0 there exists a family of Hilbert space valued maps ξi : Xi → (l2(Xi))1
satisfying:
(1) for each i ∈ I, ξi has (R, ǫ) variation;
(2) lim
S→∞
sup
i∈I
sup
x∈Xi Pw /∈B(x,S)
ξi
x(w) 2= 0.
Let X be a set. A partition of unity on X is a collection of functions {φi}i∈I , with
φi : X → [0, 1], and such thatPi∈I φi(x) = 1 for every x ∈ X. A partition of unity {φi}i∈I
is said to subordinate to a cover U = {Ui}i∈I of X if each φi vanishes outside Ui.
The following conclusion will be useful to prove the main theorem.
2
Theorem 2.3 (see [17]). Say X is a metric space such that for any R, ǫ > 0 there exists
a partition of unity {φi}i∈I on X satisfying:
(1) for all x, y ∈ X, if d(x, y) ≤ R, then Pi∈I
φi(x) − φi(y) ≤ ǫ;
(2) {φi}i∈I is subordinated to an equi-strongly embeddable cover U = {Ui}i∈I of X.
Then X is strongly embeddable.
Let U = {Ui}i∈I be a cover of a metric space X. We say the multiplicity of U is k if
each point of X is contained in at most k elements of U. The R-multiplicity of U is the
maximum number of elements of U that meet a commom ball of radius R in X (R > 0).
The Lebesgue number of U is L if any ball of radius at most L is contained in one element
of U. If d(U, V ) > L for all U, V ∈ U with U 6= V then U is L-separated (L > 0). A cover
U of X is (k, L)-separated (k ≥ 0 and L > 0) if there is a partition of U into k + 1 families
U = U0 ∪ · · · ∪ Uk
such that each family Ui is L-separated. Note that a (k, 2L)-separated cover has L-
multiplicity ≤ k + 1. Let
UL = {U (L)U ∈ U}, U (L) = {x ∈ Xd(x, U ) ≤ L}.
Note then that: if a cover U of X has L-multiplicity ≤ k + 1 then the enlarged cover UL
has multiplicity ≤ k + 1 and lebesgue number L.
Definition 2.4 (see [15]). Let X be a metric space. X is said to have finite asymptotic
dimension if there exists k ≥ 0 such that for all L ≥ 0 there exists a uniformly bounded
cover of X of Lebesgue number at least L and multiplicity k + 1. The least possible such
k is the asymptotic dimension of X.
The following result is part of the folklore of the subject.
Lemma 2.5 (see [4]). Let X be a metric space and U = {Ui}i∈I be a cover of X with mul-
tiplicity k and lebesgue number L. Then there is a partition of unity {φi}i∈I subordinating
to the cover such that
φi(x) − φi(y) ≤
(2k + 2)(2k + 3)
L
d(x, y)
Xi∈I
for any x, y ∈ X.
Next we will recall the notion of coarse quasi-action which is a generalization of the
notion of quasi-action central to the fundamental problem of quasi-isometry classification
of finitely generated groups, see [6, 11, 12].
3
Definition 2.6 (see [13]). Let X and Y be metric space and f : X → Y be a map.
(1) The map f is called proper if the inverse image, under f , of each bounded subset
of Y , is a bounded subset of X.
(2) The map f is called bornologous if there exists a real positive non-decreasing
function ℓ : [0, +∞) → [0, +∞) such that
d(x, x′) ≤ r ⇒ d(f (x), f (x′)) ≤ ℓ(r).
(3) The map f is called coarse if it is both proper and bornologous.
We say that two maps f, f ′ : X → Y are close if d(f, f ′) = sup
x∈X
d(f (x), f ′(x)) is
bounded. X and Y are coarsely equivalent if there exist coarse maps f : X → Y and
g : Y → X such that f ◦ g and g ◦ f are close to the identity maps on Y and on X
respectively.
Definition 2.7 (see [2]). A coarse quasi-action of a group G on a metric space X is an
assignment of a coarse self-equivalence fg : X → X for each element g ∈ G such that the
following conditions are satisfied:
(1) all fg are coarse maps with respect to a uniform choice of the function ℓ;
(2) there exists a number A ≥ 0 such that d(fid, idX ) ≤ A;
(3) there exists a number B ≥ 0 such that d(fg ◦ fh, fgh) ≤ B for all elements g, h ∈ G.
It is immediate from the above that d(fg ◦ f−g, idX ) ≤ A + B for all g ∈ G.
3 Direct limits for the metric space
In this section, we will prove that the strong embeddability is closed under the direct limit
for the metric space. Note that this result is rather different form ([16], Theorem 4.3).
The group structure on the direct limit is essential to enable one to drop the boundedness
assumptions necessary for this result.
First, we have the following fact.
Lemma 3.1. Let {Xi}i∈I be a collection of equi-strongly embeddable metric spaces. If Yi
is the subspace of Xi for each i, then {Yi}i∈I is also equi-strongly embeddable.
Proof. For each i ∈ I, suppose di is the metric on Xi. Let pi : Xi → Yi be a function
defined by
pi(x) = (y,
x,
if x ∈ Xi \ Yi, di(x, y) ≤ 2 di(x, Yi);
if x ∈ Yi.
4
Let R > 0 and ǫ > 0, as in Definition 2.2, there exists a family of maps βi : Xi → (l2(Xi))1
satisfying:
(1) for each i ∈ I, βi has (R, ǫ) variation;
(2) lim
S→∞
sup
i∈I
sup
x∈Xi Pw /∈B(x,S)
βi
x(w) 2= 0.
Then, for each i ∈ I we define an isometry αi : ℓ2(Xi) → ℓ2(Yi × Xi) by
αi(ζ)(y, x) = (ζ(x),
0,
if y = pi(x);
if otherwise.
for each ζ ∈ ℓ2(Xi).
Define ξi : Yi → ℓ2(Yi × Xi) by ξi
y(t, s) = αi(βi
y)(t, s) for t ∈ Yi, s ∈ Xi. Note that for
any y ∈ Yi,
ξi
y(t, s)2
αi(βi
y)(t, s)2
kξi
yk2
ℓ2(Yi×Xi) = X(t,s)∈Yi×Xi
= X(t,s)∈Yi×Xi
y(s)2
= Xs∈Xi
yk2
=kβi
βi
ℓ2(Xi)
Similarly, for any y, y′ ∈ Yi,
=1.
kξi
y − ξi
y′k2
ℓ2(Yi×Xi) = kβi
y − βi
y′k2
ℓ2(Xi)
Now we define ηi : Yi → ℓ2(Yi) by
y(t) = kξi
ηi
y(t, ·)kℓ2(Xi)
for t ∈ Yi.
Note that for any y ∈ Yi,
kηi
yk2
ℓ2(Yi) = Xt∈Yi
ηi
y(t)2 = Xt∈Yi
kξi
y(t, ·)k2
ℓ2(Xi) = kξi
yk2
ℓ2(Yi×Xi) = 1.
Then for any y, y′ ∈ Yi and di(y, y′) ≤ R, we have
kηi
y − ηi
y′k2
ℓ2(Yi) = Xt∈Yi
ηi
y(t) − ηi
y′(t)2
5
kξi
y(t, ·)kℓ2(Xi) − kξi
y′(t, ·)kℓ2(Xi)2
kξi
y(t, ·) − ξi
y′(t, ·)k2
ℓ2(Xi)
= Xt∈Yi
≤ Xt∈Yi
=kξi
=kβi
y − ξi
y − βi
y′k2
y′k2
ℓ2(Xi)
ℓ2(Yi×Xi)
≤ǫ.
It follows that ηi has (R, ǫ) variation.
Moreover,
lim
S→∞
sup
i∈I
= lim
S→∞
sup
i∈I
= lim
S→∞
sup
i∈I
= lim
S→∞
sup
i∈I
= lim
S→∞
sup
i∈I
+ lim
S→∞
sup
i∈I
ℓ2(Xi)
αi(βi
y)(z, x)2
αi(βi
y)(z, x)2
αi(βi
y)(z, x)2
kξi
y(z, ·)k2
sup
sup
sup
sup
ηi
y(z) 2
y∈Yi Xz /∈B(y,S)
y∈Yi Xz /∈B(y,S)
y∈Yi Xz /∈B(y,S) Xx∈Xi
y∈Yi Xx∈Xi Xz /∈B(y,S)
y∈Yi Xx∈Yi Xz /∈B(y,S)
y∈Yi Xx∈Xi\Yi Xz /∈B(y,S)
y∈Yi Xx /∈B(y,S)
y∈Xi Xx /∈B(y,S)
sup
sup
sup
sup
βi
βi
αi(βi
y)(z, x)2
y(x)2 + lim
S→∞
sup
i∈I
y(x)2 + lim
S→∞
sup
i∈I
sup
y∈Yi Xx∈Xi\Yi
y∈Xi Xx /∈B(y,S)
sup
βi
y(x)2
βi
y(x)2
= lim
S→∞
sup
i∈I
≤ lim
S→∞
sup
i∈I
=0.
Then the claim follows.
Lemma 3.2. Let c > 0. Let {Xi}i∈I be a collection of metric spaces and Yi is a c-net of
Xi for each i ∈ I. If {Yi}i∈I is equi-strongly embeddable, then so is {Xi}i∈I .
Proof. This proof is analogous to ([16], Lemma 5.2), so we omit it.
Proposition 3.3. Let X1 ⊆ X2 ⊆ X3 ⊆ · · · be an increasing sequence of bounded metric
n=1 Xn. Assume also that any bounded subset of X is contained in
n=1 is equi-strongly embeddable, then X is strongly embeddable.
space, and let X = S∞
some Xn. If {Xn}∞
6
Proof. Take L > 0. For given Xnk , since each Xn is bounded, we can choose Xnk+1
¯B(x, 3L) ⊆ Xnk+1. Thus we obtain a subsequence {Xnk } of {Xn} such
satisfying Sx∈Xn
that
k
( [x∈Xnk
¯B(x, L)) \ (
[
x∈Xn
k+2 \Xn
k+1
¯B(x, L)) = ∅.
Set
for every k ≥ 1.
Uk =
[x∈Xn
k+1 \Xn
k
¯B(x, L).
Then we obtain a cover U = {Uk}∞
k=1 of multiplicity at most 2 and Lebesgue number
at least L. We claim that U is equi-strongly embeddable.
Indeed, {Xnk+1\Xnk } are
subspaces of the equi-strongly embeddable sequence {Xnk }. It follows from Lemma 3.1
that {Xnk+1\Xnk } is equi-strongly embeddable. Note that for every k, {Xnk+1\Xnk } is a
L-net of Uk. By lemma 3.2, {Uk}∞
k=1 is equi-strongly embeddable.
By choosing L large enough, the result follows from Lemma 2.5 and Theorem 2.3.
4 Fibering permanence
In this section, we show an important property of strong embeddability which is called
fibrering permanence. The main motivating examples are fibre bundles p : X → Y with
base space Y and total space X. Moreover, fibering permanence is somewhat more subtle
than the other permanence properties, and care must be taken to formulate it correctly
for our other basic properties.
Proposition 4.1. Let X and Y be two metric spaces and f : X → Y be an uniformly
expansive map. Assume that Y has Property A. If for every uniformly bounded cover
{Ui}i∈I of Y , the collection {f −1(Ui)}i∈I of subspaces of X is equi-strongly embeddable,
then X is strongly embeddable.
Proof. Let R, ǫ > 0 be given. Since f is uniformly expansive, there exists S > 0, such that
d(f (x), f (y)) ≤ S whenever d(x, y) ≤ R. Since Y has property A, by one of the equivalent
definitions of property A (see [15]), there exists an uniformly bounded cover U = {Ui}i∈I
of Y , together with a partition of unity {φi}i∈I subordinated to U, such that
for y, y′ ∈ Y , d(y, y′) ≤ S.
φi(y) − φi(y′) ≤ ǫ
Xi∈I
7
We define ϕi = φi ◦ f for each i ∈ I. Clearly, {ϕi}i∈I is a partition of unity on X
subordinated to {f−1(Ui)}i∈I and satisfying the assumptions of Theorem 2.3. Thus X is
strongly embeddable.
Corollary 4.2. Let X and Y be two metric spaces and f : X → Y be a Lipschitz map
of metric spaces. Suppose that a group G acts by isometries on both X and Y , that the
action on Y is transtitive and that f is G-equivariant. Suppose that Y has property A. If
there exists y0 ∈ Y satisfying for every n ∈ N the inverse image f −1(B(y0, n)) is strongly
embeddable, then X is strongly embeddable.
Proof. Let {Ui}i∈I be a uniformly bounded cover of Y . Note that the action of G on Y is
isometrical and transitive, there exists n ∈ N and gi ∈ G such that giUi ⊆ B(y0, n) for all
i ∈ I. Note also that gif −1(Ui) = f −1(giUi) ⊆ f −1(B(y0, n)), we have that the collection
{f −1(Ui)}i∈I is isometric to a collection of subspaces of f −1(B(y0, n)).
Continuing, note that f −1(B(y0, n)) is strongly embeddable, then we have that
{f −1(Ui)}i∈I is equi-strongly embeddable. Now the result immediately follows from The-
orem 4.1.
A metric space of finite asymptotic dimension has property A, also is strongly embed-
dable. Now we prove a natural generalization of this result, where uniform boundedness
of the cover is replaced by the appropriate uniform version of strong embeddability.
Theorem 4.3. Let X be a metric space. If for any σ > 0 there exists a (k, L)-separated
cover U of X with k2 + 1 ≤ Lσ and U is equi-strongly embeddable, then X is strongly
embedable.
Proof. Let R, ǫ > 0. Take a number σ, such that 0 < σ < 1/20R. Then for any integer
k ≥ 0, we have
k2 + 1 ≥ 2(2k + 2)(2k + 3)Rσ.
It follows from the assumption that there exists a (k, 2L)-separated cover U of X such that
U is equi-strongly embeddable and k2 + 1 ≤ 2Lσǫ. Note that the cover UL has multiplicity
≤ k + 1 and Lebesgue number L. Moreover, UL is equi-strongly embeddable, as it is
coarsely equivalent to UL. By Lemma 2.5, there is a partition of unity {φU(L)}U(L)∈UL
subordinated to UL such that for all x, y ∈ X
XU(L)∈UL
φU(L)(x) − φU(L)(y) ≤
(2k + 2)(2k + 3)
L
d(x, y)
≤
k2 + 1
2RLσ
d(x, y)
8
In particular, if d(x, y) ≤ R, we have that
φU(L)(x) − φU(L)(y) ≤
k2 + 1
2Lσ
≤ ǫ.
XU(L)∈UL
This shows that X satisfies the conditions in Theorem 2.3. Hence, X is strongly embed-
dable.
5 Groups acting on metric spaces
In this section, we are ready to complete the proof of our main result.
First, we recall the notion of T -quasi-stabilizer (T > 0). Let G be a finitely generated
group with a coarse quasi-action on a metric space X and x0 be a chosen base point in X.
The T -quasi-stabilizer WT (x0) is defined to be the subset of all elements g in G such that
d(gx0, x0) ≤ T . Moreover, we can view G as a metric space with a word length metric
(see [13]).
Theorem 5.1. Assume that G is a finitely generated group with a coarse quasi-action
on a metric space X. If X has finite asymptotic dimension and there exists a base point
x0 ∈ X such that WT (x0) is strongly embeddable for any T > 0. Then G is strongly
embeddable.
Proof. Since the orbit Gx0 is a subset of X, Gx0 has finite asymptotic dimension. Without
loss of generality, we can assume that the action of G on X is transitive. Let S be the
finite symmetric generating set in the definition of the word length metric for G and take
λ = max{d(sx0, x0) s ∈ S}.
Then there exists a map π : G → X given by π(g) = gx0 for all g ∈ G. If the action of
G on X is by isometries, π is λ-Lipschitz. In our case, we show that π is ℓ(λ)-Lipschitz.
Indeed, d(π(g), π(gs)) = d(gx0, gsx0) ≤ ℓ(d(x0, sx0)) ≤ ℓ(λ) for any g ∈ G and s ∈ S.
Suppose X has asymptotic dimension ≤ k. Let L > 0 be given. By definition 2.4,
there exists a uniformly bounded cover U = {Ui}i∈I of X with Lebesgue number L and
multiplicity k + 1 such that the L-neighbourhood V = {Vi}i∈I of U is also a cover of
multiplicity k + 1.
Since V is uniformly bounded, there exist T > 0 and xi ∈ X such that Vi ⊆ B(xi, T )
for each i ∈ I. On the other hand, we can take gi ∈ G such that xi = gix0.
By Definition 2.7, we have that d(g−1
i xi, x0) ≤ A + B, then
g−1
i
(B(xi, T )) ⊆ B(g−1
i xi, ℓ(T )) ⊆ B(x0, A + B + ℓ(T )).
9
It follows from the definition of π,
g−1
i π−1(B(xi, T )) ⊆π−1(B(x0, A + 2B + ℓ(T )))
=WA+2B+ℓ(T )(x0)
Then we get
g−1
i
(Vi) ⊆ WA+2B+ℓ(T )(x0).
Note that {π−1(Vi)}i∈I is isometric to a family of subspace of WA+2B+ℓ(T )(x0). Since
WA+2B+ℓ(T )(x0) is strongly embeddable, we have that {π−1(Vi)}i∈I is equi-strongly em-
beddable. By the same argument, {π−1(Ui)}i∈I is also equi-strongly embeddable and
covers G.
We will use Theorem 2.3 to complete the proof. Let R > 0 and ǫ > 0. Take L ≥
2ℓ(λ)R(2k+2)(2k+3)
ǫ
. Since U is a uniformly bounded cover of X with Lebesgue number L
and multiplicity k + 1, it follows from Lemma 2.5 that there exists a partition of unity
{φUi}ui∈U subordinated to the cover satisfying
φUi(x) − φUi(y) ≤
(2k + 2)(2k + 3)
L
d(x, y)
XUi∈U
for any x, y ∈ X.
Moreover, note that {π−1(Vi)}i∈I is equi-strongly embeddable, where there exists a
collection of maps
βi : π−1(Vi) → (ℓ2(π−1(Vi)))1
such that βi has (R, ǫ
4 ) variation for every i ∈ I.
Define for each i ∈ I a map ϕi : G → [0, 1] by setting
ϕi(g) = Xh∈π−1(Vi)
φUi(π(g))βi
g (h)2
for g ∈ G. We claim that {ϕi}i∈I is partition of unity on G satisfying the conditions of
Theorem 2.3. First, for any g ∈ G,
Xi∈I
ϕi(g) =Xi∈I Xh∈π−1(Vi)
φUi(π(g))βi
g (h)2
φUi(π(g)) Xh∈π−1(Vi)
βi
g(h)2
φUi(π(g))
=Xi∈I
=Xi∈I
=1.
10
Note also that each ϕi vanishes outside π−1(Ui), so {ϕi}i∈I is a partition of unity on
G and subordinates to {π−1(Ui)}i∈I . Second, for any g, g′ ∈ G with d(g, g′) ≤ R.
If
g ∈ π−1(Ui)}i∈I , then π(g′) belongs to the ℓ(λ)R-neighbourhood of Ui. The latter space is
a subspace of Vi as L is large enough. Then g′ ∈ π−1(Vi). We have the following estimates
ϕi(g) − ϕi(g′)
Xi∈I
=Xi∈I
Xh∈π−1(Vi)
=Xi∈I
Xh∈π−1(Vi)
≤Xi∈I Xh∈π−1(Vi)
≤Xi∈I Xh∈π−1(Vi)
φUi(π(g))βi
(φUi(π(g))βi
g (h)2 − Xh∈π−1(Vi)
g (h)2 − φUi(π(g′))βi
g′ (h)2)
φUi(π(g′))βi
g′ (h)2
φUi(π(g))βi
g (h)2 − φUi(π(g′))βi
g′ (h)2
φUi(π(g))βi
g (h)2 − φUi(π(g))βi
g′ (h)2
+ φUi(π(g))βi
g′ (h)2 − φUi(π(g′))βi
φUi(π(g)) βi
g(h)2 − βi
g′ (h)2
g′(h)2
φUi(π(g)) − φUi(π(g′))
≤Xi∈I Xh∈π−1(Vi)
+Xi∈I
≤Xi∈I
φUi(π(g)) Xh∈π−1(Vi)
βi
g(h) + βi
g′(h) · βi
g(h) − βi
g′(h)
+
(2k + 2)(2k + 3)
L
d(π(g), π(g′))
≤Xi∈I
φUi(π(g))[( Xh∈π−1(Vi)
βi
g(h) + βi
g′(h)2)
1
2 · ( Xh∈π−1(Vi)
βi
g(h) − βi
g′(h)2)
1
2 ]
+
(2k + 2)(2k + 3)
L
ℓ(λ)R
φUi(π(g)) k βi
g + βi
g′ k · k βi
g − βi
g′ k +
φUi(π(g)) · 2 ·
ǫ
4
+
ǫ
2
≤Xi∈I
≤Xi∈I
=ǫ.
ǫ
2
Thus {ϕi}i∈I has the properties in Theorem 2.3, whence G is strongly embeddable.
Since Theorem 2.3 also holds for coarse embeddablity and exactness (it is equivalent
to property A for the metric spaces with bounded geometry) [5], a similar proof shows
11
that the same hypotheses, with 'coarse embeddability' (resp. exactness) replacing 'strong
embeddability', imply that G is coarsely embeddable (resp. exact), as described next.
Theorem 5.2 (see [5]). Say X is a metric space such that for any R, ǫ > 0 there exists
a partition of unity {φi}i∈I on X satisfying:
(1) for all x, y ∈ X, if d(x, y) ≤ R, the Pi∈I
φi(x) − φi(y) ≤ ǫ;
(2) {φi}i∈I is subordinated to an equi-coarsely embeddable (resp. equi-exact) cover
U = {Ui}i∈I of X.
Then X is coarsely embeddable (resp. exact).
Theorem 5.3. Assume that G is a finitely generated group with a coarse quasi-action
on a metric space X. If X has finite asymptotic dimension and there exists a base point
x0 ∈ X such that WT (x0) is coarsely embeddable (resp. exact) for any T > 0. Then G is
coarsely embeddable (resp. exact).
References
[1] G. Arzhantseva and R. Tessera, Admitting a coarse embedding is not preserved under
group extensions, arXiv:1605.01192.
[2] S. Beckhardt and B. Goldfarb, Extension properties of asymptotic property C and
finite decomposition complexity, arXiv:1607.00445.
[3] G. Bell, Property A for groups acting on metric spaces, Topology Appl. 130(2003),
239C251.
[4] G. Bell, Asymptotic properties of groups acting on complexes, Proc. Amer. Math. Soc.
133(2005), no. 2, 387C396.
[5] M. Dadarlat and E. Guentner, Uniform embeddability of relatively hyperbolic groups,
J. Reine Angew. Math. 612(2007), 1-15.
[6] C. Drutu, Quasi-isometry rigidity of groups. in G´eom´etries `a courbure n´egative ou
nulle, groupes discrets et rigidit´es, in S´emin. Congr., 18, Soc. Math. France, Paris,
2009, 321-371.
[7] S. Ferry, A. Ranicki, and J. Rosenberg (eds.), Novikov conjectures, index theorems
and rigidity, London Mathematical Society Lecture Notes, no. 226, 227, Cambridge
University Press, 1995.
12
[8] M. Gromov, Asymptotic Invariants of Infnite Groups, Volume 2 of "Geometry Group
Theory, Sussex 1991", G. A. Niblo and M. A. Roller Eds, Cambridge Univ. Press,
1993.
[9] N. Higson and J. Roe, Amenable group actions and the Novikov conjecture, J. Reine
Angew. Math. 519(2000), 143C153.
[10] R. Ji, C. Ogle and B. W. Ramsey, Strong embeddablility and extensions of groups,
arXiv:1307.1935.
[11] M. Kapovich, Lectures on quasi-isometric rigidity, in Geometric Group Theory, vol-
ume 21 of Publications of IAS/Park City Summer Institute, Amer. Math. Soc., Prov-
idence, RI, 2014, 127-172.
[12] L. Mosher, M. Sageev and K. Whyte, Quasi-actions on trees I. Bounded valence,
Annals Math. 158(2003), 115-164.
[13] P. Nowak and G. Yu, Large Scale Geometry, European Mathematical Society, 2012.
[14] Jean-Louis Tu, Remarks on Yu's property A for disrete metric spaces and groups,
Bull. Soc. Math. France. 129(2001), no. 1, 115-139.
[15] R. Willett, Some notes on property A. EPFL Press, Lansanne, 2009.
[16] J. Xia and X. Wang, On strong embeddability and finite decomposition complexity,
Acta Math. Sin., Engl. Ser. 33(2017), 403-418.
[17] J. Xia and X. Wang, Strong embeddability for groups acting on metric spaces, In press,
2017.
[18] G. Yu, The coarse Baum-connes conjecture for spaces which admit a uniform embed-
ding into Hilbert space, Invent. Math. 139(2000), 201-240.
Guoqiang Li
College of Mathematics and Statistics,
Chongqing University (at Huxi Campus),
Chongqing 401331, P. R. China
E-mail: [email protected]
Xianjin Wang
College of Mathematics and Statistics,
Chongqing University (at Huxi Campus),
13
Chongqing 401331, P. R. China
E-mail: [email protected]
14
|
Subsets and Splits